diff --git a/.gitattributes b/.gitattributes index 7ff4242bc49e23d83f1f893fdaedb76ff03cac21..4463a9967840ca082c80a0c61e1a9ba1d95bde6d 100644 --- a/.gitattributes +++ b/.gitattributes @@ -2215,3 +2215,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/TAET_[[:space:]]Two-Stage[[:space:]]Adversarial[[:space:]]Equalization[[:space:]]Training[[:space:]]on[[:space:]]Long-Tailed[[:space:]]Distributions/4476cf03-3c08-4623-986a-a92479d57f0f_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/TAGA_[[:space:]]Self-supervised[[:space:]]Learning[[:space:]]for[[:space:]]Template-free[[:space:]]Animatable[[:space:]]Gaussian[[:space:]]Articulated[[:space:]]Model/e7cef16f-7c47-4c54-903b-2d640c6e7390_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/TAMT_[[:space:]]Temporal-Aware[[:space:]]Model[[:space:]]Tuning[[:space:]]for[[:space:]]Cross-Domain[[:space:]]Few-Shot[[:space:]]Action[[:space:]]Recognition/7caf9c86-26da-41ce-9aca-6e2c4dcc4ad4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TANGO_[[:space:]]Training-free[[:space:]]Embodied[[:space:]]AI[[:space:]]Agents[[:space:]]for[[:space:]]Open-world[[:space:]]Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TAPT_[[:space:]]Test-Time[[:space:]]Adversarial[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Robust[[:space:]]Inference[[:space:]]in[[:space:]]Vision-Language[[:space:]]Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TAROT_[[:space:]]Towards[[:space:]]Essentially[[:space:]]Domain-Invariant[[:space:]]Robustness[[:space:]]with[[:space:]]Theoretical[[:space:]]Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TASTE-Rob_[[:space:]]Advancing[[:space:]]Video[[:space:]]Generation[[:space:]]of[[:space:]]Task-Oriented[[:space:]]Hand-Object[[:space:]]Interaction[[:space:]]for[[:space:]]Generalizable[[:space:]]Robotic[[:space:]]Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TCFG_[[:space:]]Tangential[[:space:]]Damping[[:space:]]Classifier-free[[:space:]]Guidance/8113438a-9689-405a-b6bc-61aa320e1894_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TFCustom_[[:space:]]Customized[[:space:]]Image[[:space:]]Generation[[:space:]]with[[:space:]]Time-Aware[[:space:]]Frequency[[:space:]]Feature[[:space:]]Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TIDE_[[:space:]]Training[[:space:]]Locally[[:space:]]Interpretable[[:space:]]Domain[[:space:]]Generalization[[:space:]]Models[[:space:]]Enables[[:space:]]Test-time[[:space:]]Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TIMotion_[[:space:]]Temporal[[:space:]]and[[:space:]]Interactive[[:space:]]Framework[[:space:]]for[[:space:]]Efficient[[:space:]]Human-Human[[:space:]]Motion[[:space:]]Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TKG-DM_[[:space:]]Training-free[[:space:]]Chroma[[:space:]]Key[[:space:]]Content[[:space:]]Generation[[:space:]]Diffusion[[:space:]]Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TSAM_[[:space:]]Temporal[[:space:]]SAM[[:space:]]Augmented[[:space:]]with[[:space:]]Multimodal[[:space:]]Prompts[[:space:]]for[[:space:]]Referring[[:space:]]Audio-Visual[[:space:]]Segmentation/8535e16b-e9a4-4280-8556-88908a875920_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TSD-SR_[[:space:]]One-Step[[:space:]]Diffusion[[:space:]]with[[:space:]]Target[[:space:]]Score[[:space:]]Distillation[[:space:]]for[[:space:]]Real-World[[:space:]]Image[[:space:]]Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TSP-Mamba_[[:space:]]The[[:space:]]Travelling[[:space:]]Salesman[[:space:]]Problem[[:space:]]Meets[[:space:]]Mamba[[:space:]]for[[:space:]]Image[[:space:]]Super-resolution[[:space:]]and[[:space:]]Beyond/989548d5-80da-4477-a3a5-42129f121eff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TacoDepth_[[:space:]]Towards[[:space:]]Efficient[[:space:]]Radar-Camera[[:space:]]Depth[[:space:]]Estimation[[:space:]]with[[:space:]]One-stage[[:space:]]Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TailedCore_[[:space:]]Few-Shot[[:space:]]Sampling[[:space:]]for[[:space:]]Unsupervised[[:space:]]Long-Tail[[:space:]]Noisy[[:space:]]Anomaly[[:space:]]Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Take[[:space:]]the[[:space:]]Bull[[:space:]]by[[:space:]]the[[:space:]]Horns_[[:space:]]Learning[[:space:]]to[[:space:]]Segment[[:space:]]Hard[[:space:]]Samples/c1e057f3-058c-42b0-8d35-8e21af099503_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Taming[[:space:]]Teacher[[:space:]]Forcing[[:space:]]for[[:space:]]Masked[[:space:]]Autoregressive[[:space:]]Video[[:space:]]Generation/e2161c88-f390-4144-acf4-52130bcde3bd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Taming[[:space:]]Video[[:space:]]Diffusion[[:space:]]Prior[[:space:]]with[[:space:]]Scene-Grounding[[:space:]]Guidance[[:space:]]for[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]from[[:space:]]Sparse[[:space:]]Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TaoAvatar_[[:space:]]Real-Time[[:space:]]Lifelike[[:space:]]Full-Body[[:space:]]Talking[[:space:]]Avatars[[:space:]]for[[:space:]]Augmented[[:space:]]Reality[[:space:]]via[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Targeted[[:space:]]Forgetting[[:space:]]of[[:space:]]Image[[:space:]]Subgroups[[:space:]]in[[:space:]]CLIP[[:space:]]Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Tartan[[:space:]]IMU_[[:space:]]A[[:space:]]Light[[:space:]]Foundation[[:space:]]Model[[:space:]]for[[:space:]]Inertial[[:space:]]Positioning[[:space:]]in[[:space:]]Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Task[[:space:]]Preference[[:space:]]Optimization_[[:space:]]Improving[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Vision[[:space:]]Task[[:space:]]Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Task[[:space:]]Singular[[:space:]]Vectors_[[:space:]]Reducing[[:space:]]Task[[:space:]]Interference[[:space:]]in[[:space:]]Model[[:space:]]Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Task-Agnostic[[:space:]]Guided[[:space:]]Feature[[:space:]]Expansion[[:space:]]for[[:space:]]Class-Incremental[[:space:]]Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Task-Aware[[:space:]]Clustering[[:space:]]for[[:space:]]Prompting[[:space:]]Vision-Language[[:space:]]Models/aabdfb34-0949-4921-9337-b3ad9044174e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Task-Specific[[:space:]]Gradient[[:space:]]Adaptation[[:space:]]for[[:space:]]Few-Shot[[:space:]]One-Class[[:space:]]Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Task-aware[[:space:]]Cross-modal[[:space:]]Feature[[:space:]]Refinement[[:space:]]Transformer[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Visual[[:space:]]Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Task-driven[[:space:]]Image[[:space:]]Fusion[[:space:]]with[[:space:]]Learnable[[:space:]]Fusion[[:space:]]Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Taste[[:space:]]More,[[:space:]]Taste[[:space:]]Better_[[:space:]]Diverse[[:space:]]Data[[:space:]]and[[:space:]]Strong[[:space:]]Model[[:space:]]Boost[[:space:]]Semi-Supervised[[:space:]]Crowd[[:space:]]Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Taxonomy-Aware[[:space:]]Evaluation[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Teaching[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Regress[[:space:]]Accurate[[:space:]]Image[[:space:]]Quality[[:space:]]Scores[[:space:]]Using[[:space:]]Score[[:space:]]Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Teller_[[:space:]]Real-Time[[:space:]]Streaming[[:space:]]Audio-Driven[[:space:]]Portrait[[:space:]]Animation[[:space:]]with[[:space:]]Autoregressive[[:space:]]Motion[[:space:]]Generation/acf23ef4-3e0f-4da0-a811-56761959df29_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Temporal[[:space:]]Action[[:space:]]Detection[[:space:]]Model[[:space:]]Compression[[:space:]]by[[:space:]]Progressive[[:space:]]Block[[:space:]]Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Temporal[[:space:]]Alignment-Free[[:space:]]Video[[:space:]]Matching[[:space:]]for[[:space:]]Few-shot[[:space:]]Action[[:space:]]Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Temporal[[:space:]]Score[[:space:]]Analysis[[:space:]]for[[:space:]]Understanding[[:space:]]and[[:space:]]Correcting[[:space:]]Diffusion[[:space:]]Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Temporal[[:space:]]Separation[[:space:]]with[[:space:]]Entropy[[:space:]]Regularization[[:space:]]for[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]in[[:space:]]Spiking[[:space:]]Neural[[:space:]]Networks/511efbd9-e72e-4675-94ed-5687d3419119_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Temporally[[:space:]]Consistent[[:space:]]Object-Centric[[:space:]]Learning[[:space:]]by[[:space:]]Contrasting[[:space:]]Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TensoFlow_[[:space:]]Tensorial[[:space:]]Flow-based[[:space:]]Sampler[[:space:]]for[[:space:]]Inverse[[:space:]]Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Test-Time[[:space:]]Backdoor[[:space:]]Detection[[:space:]]for[[:space:]]Object[[:space:]]Detection[[:space:]]Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Test-Time[[:space:]]Domain[[:space:]]Generalization[[:space:]]via[[:space:]]Universe[[:space:]]Learning_[[:space:]]A[[:space:]]Multi-Graph[[:space:]]Matching[[:space:]]Approach[[:space:]]for[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Test-Time[[:space:]]Fine-Tuning[[:space:]]of[[:space:]]Image[[:space:]]Compression[[:space:]]Models[[:space:]]for[[:space:]]Multi-Task[[:space:]]Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Test-Time[[:space:]]Visual[[:space:]]In-Context[[:space:]]Tuning/1106c942-047b-40f5-b114-2de12258d183_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Test-time[[:space:]]Augmentation[[:space:]]Improves[[:space:]]Efficiency[[:space:]]in[[:space:]]Conformal[[:space:]]Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TexGarment_[[:space:]]Consistent[[:space:]]Garment[[:space:]]UV[[:space:]]Texture[[:space:]]Generation[[:space:]]via[[:space:]]Efficient[[:space:]]3D[[:space:]]Structure-Guided[[:space:]]Diffusion[[:space:]]Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TexGaussian_[[:space:]]Generating[[:space:]]High-quality[[:space:]]PBR[[:space:]]Material[[:space:]]via[[:space:]]Octree-based[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text[[:space:]]Augmented[[:space:]]Correlation[[:space:]]Transformer[[:space:]]For[[:space:]]Few-shot[[:space:]]Classification[[:space:]]&[[:space:]]Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text[[:space:]]Embedding[[:space:]]is[[:space:]]Not[[:space:]]All[[:space:]]You[[:space:]]Need_[[:space:]]Attention[[:space:]]Control[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Semantic[[:space:]]Alignment[[:space:]]with[[:space:]]Text[[:space:]]Self-Attention[[:space:]]Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text-Driven[[:space:]]Fashion[[:space:]]Image[[:space:]]Editing[[:space:]]with[[:space:]]Compositional[[:space:]]Concept[[:space:]]Learning[[:space:]]and[[:space:]]Counterfactual[[:space:]]Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text-guided[[:space:]]Sparse[[:space:]]Voxel[[:space:]]Pruning[[:space:]]for[[:space:]]Efficient[[:space:]]3D[[:space:]]Visual[[:space:]]Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Textured[[:space:]]Gaussians[[:space:]]for[[:space:]]Enhanced[[:space:]]3D[[:space:]]Scene[[:space:]]Appearance[[:space:]]Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Art[[:space:]]of[[:space:]]Deception_[[:space:]]Color[[:space:]]Visual[[:space:]]Illusions[[:space:]]and[[:space:]]Diffusion[[:space:]]Models/32f81902-9541-4c54-8edb-2ca63c430edb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Change[[:space:]]You[[:space:]]Want[[:space:]]To[[:space:]]Detect_[[:space:]]Semantic[[:space:]]Change[[:space:]]Detection[[:space:]]In[[:space:]]Earth[[:space:]]Observation[[:space:]]With[[:space:]]Hybrid[[:space:]]Data[[:space:]]Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Devil[[:space:]]is[[:space:]]in[[:space:]]Low-Level[[:space:]]Features[[:space:]]for[[:space:]]Cross-Domain[[:space:]]Few-Shot[[:space:]]Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Devil[[:space:]]is[[:space:]]in[[:space:]]Temporal[[:space:]]Token_[[:space:]]High[[:space:]]Quality[[:space:]]Video[[:space:]]Reasoning[[:space:]]Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Devil[[:space:]]is[[:space:]]in[[:space:]]the[[:space:]]Prompts_[[:space:]]Retrieval-Augmented[[:space:]]Prompt[[:space:]]Optimization[[:space:]]for[[:space:]]Text-to-Video[[:space:]]Generation/bf61a946-383a-4ffe-b29d-37540410b684_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Illusion[[:space:]]of[[:space:]]Unlearning_[[:space:]]The[[:space:]]Unstable[[:space:]]Nature[[:space:]]of[[:space:]]Machine[[:space:]]Unlearning[[:space:]]in[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Impact[[:space:]]Label[[:space:]]Noise[[:space:]]and[[:space:]]Choice[[:space:]]of[[:space:]]Threshold[[:space:]]has[[:space:]]on[[:space:]]Cross-Entropy[[:space:]]and[[:space:]]Soft-Dice[[:space:]]in[[:space:]]Image[[:space:]]Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Language[[:space:]]of[[:space:]]Motion_[[:space:]]Unifying[[:space:]]Verbal[[:space:]]and[[:space:]]Non-verbal[[:space:]]Language[[:space:]]of[[:space:]]3D[[:space:]]Human[[:space:]]Motion/5856b279-35bc-44f9-b231-826fe1b3d735_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]PanAf-FGBG[[:space:]]Dataset_[[:space:]]Understanding[[:space:]]the[[:space:]]Impact[[:space:]]of[[:space:]]Backgrounds[[:space:]]in[[:space:]]Wildlife[[:space:]]Behaviour[[:space:]]Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Photographer's[[:space:]]Eye_[[:space:]]Teaching[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]See,[[:space:]]and[[:space:]]Critique[[:space:]]Like[[:space:]]Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Power[[:space:]]of[[:space:]]Context_[[:space:]]How[[:space:]]Multimodality[[:space:]]Improves[[:space:]]Image[[:space:]]Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Scene[[:space:]]Language_[[:space:]]Representing[[:space:]]Scenes[[:space:]]with[[:space:]]Programs,[[:space:]]Words,[[:space:]]and[[:space:]]Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Theoretical[[:space:]]Insights[[:space:]]in[[:space:]]Model[[:space:]]Inversion[[:space:]]Robustness[[:space:]]and[[:space:]]Conditional[[:space:]]Entropy[[:space:]]Maximization[[:space:]]for[[:space:]]Collaborative[[:space:]]Inference[[:space:]]Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Theory-Inspired[[:space:]]Deep[[:space:]]Multi-View[[:space:]]Multi-Label[[:space:]]Learning[[:space:]]with[[:space:]]Incomplete[[:space:]]Views[[:space:]]and[[:space:]]Noisy[[:space:]]Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Thin-Shell-SfT_[[:space:]]Fine-Grained[[:space:]]Monocular[[:space:]]Non-rigid[[:space:]]3D[[:space:]]Surface[[:space:]]Tracking[[:space:]]with[[:space:]]Neural[[:space:]]Deformation[[:space:]]Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_content_list.json b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..efb2500c4c2fceac66c3d86af7ce6e9fc648baeb --- /dev/null +++ b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_content_list.json @@ -0,0 +1,1304 @@ +[ + { + "type": "text", + "text": "TANGO: Training-free Embodied AI Agents for Open-world Tasks", + "text_level": 1, + "bbox": [ + 158, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Filippo Ziliotto $^{1,2*}$", + "bbox": [ + 137, + 179, + 287, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tommaso Campari*", + "bbox": [ + 326, + 180, + 500, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luciano Serafini²", + "bbox": [ + 532, + 180, + 674, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lamberto Ballan1", + "bbox": [ + 714, + 180, + 856, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Padova, Italy", + "bbox": [ + 383, + 215, + 612, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fondazione Bruno Kessler (FBK), Trento, Italy", + "bbox": [ + 302, + 234, + 689, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) have demonstrated excellent capabilities in composing various modules together to create programs that can perform complex reasoning tasks on images. In this paper, we propose TANGO, an approach that extends the program composition via LLMs already observed for images, aiming to integrate those capabilities into embodied agents capable of observing and acting in the world. Specifically, by employing a simple PointGoal Navigation model combined with a memory-based exploration policy as a foundational primitive for guiding an agent through the world, we show how a single model can address diverse tasks without additional training. We task an LLM with composing the provided primitives to solve a specific task, using only a few in-context examples in the prompt. We evaluate our approach on three key Embodied AI tasks: Open-Set ObjectGoal Navigation, MultiModal Lifelong Navigation, and Open Embodied Question Answering, achieving state-of-the-art results without any specific fine-tuning in challenging zero-shot scenarios.", + "bbox": [ + 89, + 316, + 483, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 636, + 220, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) have gained significant attention in the field of AI due to their remarkable capability to generalize across unseen tasks [5, 24, 48, 49]. Systems like VisProg and ViperGPT [18, 47] demonstrated strong performance on a broad range of vision-and-language tasks by just relying on a few numbers of in-context examples to generate complex compositional programs without requiring any specific training. Nevertheless, the evaluation of these concepts to Embodied AI, particularly for navigation tasks, remains largely unexplored. This approach may be crucial in building agents capable to navigate and operate efficiently in unfamiliar environments.", + "bbox": [ + 89, + 662, + 482, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previously, methods such as Neural Module Networks (NMN) [2, 19] have demonstrated promising compositional", + "bbox": [ + 89, + 844, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f7b7e81ceb8828b89be8dd0cff489b6023a2caad0fdf20e17f343f77f9b4ce0.jpg", + "image_caption": [ + "Figure 1. We introduce TANGO, a modular neuro-symbolic system for compositional embodied visual navigation. Given a few examples of natural language instructions and the corresponding programs composed of action primitives, TANGO can generate executable programs, enabling the agent to perform multiple tasks within a 3D environment." + ], + "image_footnote": [], + "bbox": [ + 522, + 284, + 898, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "properties for high-level tasks like visual question answering, via end-to-end training of networks and specialized, differentiable neural modules. However, they required the combination of semantic parsers and predefined templates to learn to solve the compositional tasks. Alternative approaches leverage LLMs through predefined task modules or APIs for action execution [22, 23, 32, 52, 58]. While effective, these methods often lack a compositional structure to address multiple tasks or primarily focus on manipulation and robotic planning, with limited emphasis on navigation.", + "bbox": [ + 511, + 715, + 906, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we present TANGO (see Figure 1) a novel neuro-symbolic compositional approach which uti-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The authors contributed equally to the work.", + "bbox": [ + 107, + 887, + 351, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24603", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "izes primitives and employs a LLM as a planner to sequence these primitives within photorealistic 3D environments, where agents must perceive and act. This framework integrates high-level planning with low-level action execution, without the need for training. TANGO makes use of diverse modules designed for visual navigation and question-answering tasks, resulting in a system that seamlessly addresses both challenges while achieving SoA performance without requiring any prior adjustments. By providing a few in-context examples that show how to tackle multiple tasks, TANGO is capable of generalizing to the specific task at hand. This is facilitated by the LLM, which effectively combines the individual modules available within the TANGO system. Moreover, it extends [55] exploration policy by incorporating a memory mechanism in the form of a stored feature map that retains information about previously explored areas, supporting for efficient life-long navigation tasks.", + "bbox": [ + 89, + 90, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the flexibility of the proposed framework, TANGO has been tested on three popular benchmarks: namely, i) Open-Vocabulary ObjectGoal Navigation [56], ii) Multi-Modal Lifelong Navigation [26], and iii) Embodied Question Answering [15, 34]. Results match or surpass previous approaches, without requiring specialized training. Summing up, the contributions of this paper are as follows:", + "bbox": [ + 89, + 364, + 480, + 482 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Introduction of a neuro-symbolic compositional LLM-based framework for EAI leveraging specialized primitive modules.", + "- Demonstration of robust generalization capabilities across multiple tasks without the need for any specific training or fine-tuning.", + "- Extension of the exploration policy presented in [55] to multi-goal scenarios through the incorporation of a memory mechanism stored as a feature vector map.", + "- Achievement of state-of-the-art results, underscoring the effectiveness of the proposed framework." + ], + "bbox": [ + 91, + 484, + 482, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 667, + 240, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work takes inspiration from modular approaches for visual tasks, like the seminal Neural Module Networks [2, 19], as well as the recent VisProg [18] and ViperGPT [47] frameworks. The main contribution of this paper is to extend these ideas to embodied visual navigation, that usually require heavy end-to-end learning, although modular approaches have recently shown promising results [7, 17]. Therefore, we discuss prior work in the area of embodied AI, the use of language models in robotics, and program generation approaches for image recognition tasks.", + "bbox": [ + 89, + 695, + 482, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Embodied AI and Visual Navigation. The field of Embodied AI has recently undergone a paradigm shift, fuelled", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "by the emergence of highly efficient simulators [27, 36, 43-45]. These simulators enable processing numerous parallel simulations in photorealistic indoor environments, facilitating large-scale testing that was otherwise challenging in classical robotics. Alongside these simulators, a large variety of tasks (and benchmarks) has been proposed: PointGoal Navigation [50], where an agent navigates from point A to point B in an unknown environment; ObjectGoal Navigation [3], requiring the agent to locate and navigate to an object in the scene; Instance-Image Goal Navigation [28], akin to ObjectGoal Navigation, but in which the agent should find a specific object depicted in a given image; Multi-Modal Lifelong Navigation [26] involving navigating to a sequence of target objects that can be specified through labels, images, or textual descriptions; Embodied Question Answering [15, 34], in which an agent navigates an environment to gather information needed to answer a question, and Vision and Language Navigation (VLN) [1].", + "bbox": [ + 511, + 90, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various approaches have been proposed to tackle these tasks, primarily falling into three categories: end-to-end [6, 53, 54], modular [11-13, 29, 39, 40] and LLM-based [16, 30, 57-59]. However, the main limitation of these paradigms is their dependence on task-specific architectures. In the case of end-to-end approaches, a model is trained for a specific problem, often requiring days of training, such as in [50], where a policy for PointGoal Navigation was trained for 2.5 billion steps. To adapt the learned model to a new problem, adjustments and retraining are necessary, and this comes with a high cost. Modular approaches share a similar challenge, usually relying on policies designed for specific tasks, although common elements, such as exploration and navigation modules, are often shared [11, 12, 29, 39, 40]. Adapting these solutions to different tasks requires a manual adjustment of the approach to fit the new domain, often involving the addition or modification of modules.", + "bbox": [ + 511, + 364, + 903, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language models for Robotics In robotics, foundation models (LLMs) trained on vast internet-scale datasets [5, 24, 48, 49] have the potential to equip robots with real-world priors and advanced reasoning abilities without the need for extensive task-specific training. Early approaches equipped agents with learned language embeddings, requiring large amounts of training data [4, 32]. Recent studies, on the other hand, have explored zero-shot and few-shot solutions mainly focusing on robotic planning and manipulation tasks [22, 23, 32, 41, 46]. In the context of visual navigation, LLM-based approaches leverage the powerful LLMs priors and reasoning capabilities to guide navigation within the environment [16, 21, 30, 57-59]; however, despite their remarkable contributions, they often lack a modular design that allows for flexible integration and extensibility, as they are typically optimized for single tasks rather", + "bbox": [ + 511, + 659, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24604", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "than a comprehensive, adaptable framework.", + "bbox": [ + 89, + 90, + 387, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, our work deviates from these paradigms by relying on a series of diverse pre-trained modules that, thanks to a LLM, can be combined to potentially solve various tasks, compositionally. None of our modules are fine-tuned explicitly for the target problems, placing us in a zero-shot setting. Moreover, TANGO does not only rely on an exploration policy based entirely on LLM output, as in [16, 58, 59]. Instead, it combines a frontier exploration policy with LLM priors and memory guidance to enhance navigation performance. To achieve this, we use a pre-trained PointGoal policy as the base waypoint navigation module of TANGO and extend the exploration policy introduced in [55] by equipping the agent with a memory mechanism stored as a feature vector for each pixel of the map.", + "bbox": [ + 89, + 109, + 485, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modular vision and program composition for visual tasks. Neural Module Networks (NMN) [2, 19] introduced modular and compositional methodologies for visual question answering (VQA). NMNs integrate neural modules into an end-to-end differentiable network; the approach originally relied on pre-existing parsers [2], whereas more recent methods [19, 20, 25] have evolved to learn the layout generation model concurrently with the neural modules, employing reinforcement learning [51] and weak supervision. Stack-NMN [20] extends N2NMN by transitioning from discrete to soft layout generation, incorporating a weighted average of predictions from all modules at each step, determined by a layout generation network.", + "bbox": [ + 89, + 354, + 483, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, [18] introduced VisProg, a framework that offers two key advantages over NMNs. Firstly, it constructs high-level programs that invoke state-of-the-art neural models and Python functions at intermediate steps, diverging from the conventional approach of generating end-to-end neural networks. This design facilitates the integration of symbolic, non-differentiable modules. Secondly, it leverages the in-context learning ability of LLMs [5]. This is achieved by prompting the LLM with a natural language instruction (or a visual question or a statement to be verified), along with a few examples of similar instructions and their corresponding programs. Almost concurrently, ViperGPT [47] presented a similar approach to VisProg, but in this case the model directly generates Python code, instead of composing pre-defined modules and routines. In contrast, in our framework, the LLM is provided only with high-level knowledge of the functions of each primitive, which is encoded directly in the primitive's name. Therefore, the system it is built similarly to VisProg. It is noteworthy that both VisProg and ViperGPT were tested on images, while our model is specifically designed for embodied agents operating in an interactive environment, relying on sensor-based perception.", + "bbox": [ + 89, + 551, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neuro-symbolic approaches offer the possibility to address a broad range of diverse complex tasks efficiently. Systems like VisProg [18] operate on images and have demonstrated excellent results in zero-shot settings. Understanding how a similar approach works in an action-oriented scenario is key to enable robots to navigate autonomously in novel environments. Therefore, we develop this paradigm further into EAI, where the integration with action execution adds several challenges. Inspired by [18], given an input prompt, we rely on a LLM to generate synthetic pseudo-programs that can be executed by the embodied agent in the environment. Our TANGO framework is therefore able to generalize to new tasks, without any direct training on task-specific datasets, thus effectively mitigating the heavy computational training costs inherent to embodied navigation tasks. The procedure that enables generating these new programs is based on \"in-context examples\" that are fed to the LLM, alongside a natural language instruction. An outline of TANGO is shown in Figure 2.", + "bbox": [ + 511, + 114, + 906, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TANGO Interpreter. The first key component of the proposed framework is referred to as \"Program Interpreter\". It comprises visual recognition modules that can be used by the agent to extract the semantics of the scene, as well as to provide an understanding of the visual context.", + "bbox": [ + 511, + 414, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 1, users can ask questions or give specific tasks to the agent; the natural language prompt is then processed by the LLM (in our implementation GPT-4o [5]), which serves as a planner and outputs a step-by-step executable program. To ensure the LLM delivers a reasonable output, it is fed with 15 \"in-context examples\" across diverse tasks. This enables the LLM to make use of its reasoning capabilities effectively, identifying the most suitable planning for the required task. Moreover, the examples remain the same regardless of the task identified in the prompt; it is the LLM's responsibility to output the specific program target for the given question. Programs use a higher level of abstraction than previous modular attempts such as Neural Module Networks (NMN) [2, 19]. Each program is constructed as a sequence of primitives (e.g., detect, answer, match, etc.) that invoke corresponding TANGO modules. These modules are either powered by pre-trained state-of-the-art vision models or implemented as simple Python subroutines (e.g., count, is_found, eval, etc.), with additional navigation-specific modules designed to \"steer\" the agent's movements (e.g. navigate_to, return, turn, etc.). Figure 3 provides a comprehensive list of all the currently implemented modules.", + "bbox": [ + 511, + 492, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows an Embodied Question Answering (EQA) example for the user's question: \"on the left of the toilet there's a shower, is the shower curtain open or closed?\" The LLM, which acts as a planner, transforms the", + "bbox": [ + 511, + 839, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24605", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a79c71eb605a0bc6cae2e5569d8ee535c0a077ecb5e8eded071b2392ae2c8f12.jpg", + "image_caption": [ + "Figure 2. Overview of the program generation in TANGO. Given few \"in-context examples\", the LLM provide a detailed sequence of steps to be executed by the agent in the given environment. The LLM is instructed to comment its output to allow for explainability." + ], + "image_footnote": [], + "bbox": [ + 112, + 87, + 890, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "user's initial prompt into navigation subtasks by leveraging its reasoning abilities and powerful priors. This decomposition enables the model to break down complex queries into more manageable steps, guiding the agent's navigation process effectively. Each line of the generated program corresponds to a module serving a specific task. The agent must then execute the generated program. The LLM is also guided to comment on its steps directly within the generated pseudocode.", + "bbox": [ + 88, + 546, + 482, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All the given primitives are equipped with methods to: i) parse lines in order to extract input argument names and values, as well as the output variable name; ii) execute the module in the environment, and update the program state with the output variable name and value. The outputs for each step and the comments in the generated pseudocode can also be used to better understand what is happening under the hood and why the LLM generated a specific line, thus enabling a good interpretability of the system behavior.", + "bbox": [ + 89, + 686, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TANGO modules utilize open-source software and models, readily downloadable from the web. They can also be effortlessly updated with newer, more efficient models as they become available.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Navigation Module. In order to navigate the environment, we define a module employing a PointGoal navigation agent as our foundational module [40, 55]. This model achieved nearly perfect PointGoal performance, both in terms of success ( $\\sim 99\\%$ ) and efficiency (a forward pass that takes a fraction of a second) on standard datasets [9, 38]. The agent starts the exploration of the environment until it locates its target goal. Once the target is identified, the focus of exploration transitions to reaching the designated goal. This PointNav policy exclusively uses the egocentric depth image and the robot's relative distance and heading towards the desired goal point as input.", + "bbox": [ + 511, + 546, + 906, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exploration Policy. Exploration is performed using the policy outlined in [55]. This method builds occupancy maps based on depth observations to identify exploration frontiers. It further leverages RGB observations and a pretrained vision-language model (BLIP2 [31]) to generate a language-grounded \"value map\". This value map is then used to guide the exploration, facilitating an efficient search for instances of a given target. Notably, as in [55], the agent performs a $360^{\\circ}$ turn at the start of navigation to initialize frontiers.", + "bbox": [ + 511, + 734, + 908, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We extend this policy for sequential goals by incorpora", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24606", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/14ba6ce6596cd3c232b08380c74a7d43e3b9e114dab00fe5dddeee2a6d0dd781.jpg", + "image_caption": [ + "Figure 3. Overview of TANGO modules. Modules span a variety of inputs and outputs. Orange modules use Python subroutines, while blue modules use pre-trained computer vision models (similarly to [18]). The navigate_to and explore(scene modules, in green, both implement our foundational PointNav module; however, only explorescene integrates the memory mechanism." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 482, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing a memory mechanism, represented as a “feature map” in which each pixel of the value map is encoded as a vector and updated at each step. This feature map then updates the language-grounded original value map when a new target is specified, by calculating the cosine similarity between each pixel's vector and the new target embeddings (either from text or image). We sample the highest value in the updated value map; if this value exceeds a predefined threshold, indicating the agent has previously encountered and “rememberers” the target's location, the agent navigates to it, supporting lifelong navigation (see Section 2-Supp).", + "bbox": [ + 88, + 292, + 482, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Navigation Tasks with TANGO", + "text_level": 1, + "bbox": [ + 89, + 468, + 366, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TANGO leverages pre-trained multimodal vision-language and vision-only models as foundational components to extract semantic information from the scene, making it well-suited for several Embodied AI tasks.", + "bbox": [ + 89, + 492, + 482, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(Open-set) ObjectGoal Navigation [3, 56], in particular, emerges as a suitable testbed to evaluate the efficacy of our method. A key module to tackle this problem involves utilizing an object detector (Owlv2 in our implementation [35] or DETR [8] for target objects that fall within the COCO classes [33]) to identify objects within the image. Subsequently, navigation towards the detected objects is facilitated by waypoints, leveraging depth distance calculation to guide the agent effectively. For simplicity, we use the center of the bounding box to determine the target waypoint.", + "bbox": [ + 88, + 551, + 482, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TANGO also integrates a specialized module for matching target objectives with ongoing visual observations in image-based scenarios, resulting in accurate navigation towards an instance image. To evaluate our agent's performance in this context, we use real-time matching SuperGlue network [42], optimized for indoor environments (i.e. using hyperparameters recommended in the respective paper).", + "bbox": [ + 89, + 704, + 482, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TANGO being task-agnostic, can process any type of navigation target, regardless of the order in which it is presented. In this context, the GOAT benchmark [26] is well-suited for evaluating our system, as it operates in scenarios where targets are specified through images, text, or descriptive phrases, provided in random sequence. It also heavily", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "relies on memory to find previously seen targets when sequential goals are given.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An agent should also be able to answer user queries like \"can you check if the kitchen table is clean?\" Hence, EQA serves as an excellent benchmark to assess TANGO's capabilities and robustness. EQA consists of three phases: understanding the semantic structure of the prompt, locating the target object(s), and analyzing the visual semantics to generate an accurate response. To this end, TANGO integrates a specific answer module, relying on BLIP2 [31] as its core foundation.", + "bbox": [ + 511, + 121, + 903, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 272, + 645, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TANGO provides a flexible framework that can be applied to various embodied navigation problems. We evaluate our approach on three popular tasks that require a wide range of capabilities, including efficient environment exploration, path planning, scene and context understanding, and image similarity comparison.", + "bbox": [ + 511, + 297, + 906, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 398, + 702, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Agent Configuration. Prior research on visual navigation commonly uses different agent configurations depending on the considered task [15, 26, 56]. The configuration typically employed for (OPEN-SET) OBJNAV closely resembles that of a LoCoBot, with an agent height of $0.88m$ , a radius of $0.18m$ , and a single $640 \\times 480$ RGB sensor with a $79^{\\circ}$ hfov, positioned $0.88m$ above the ground.", + "bbox": [ + 511, + 422, + 905, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the context of GOAT Benchmark, some of the default settings differ, featuring a camera situated $1.31m$ above the ground. The agent has a height of $1.41m$ and is given $360 \\times 640$ RGB images. Both the above tasks' configurations use a step size of $0.25m$ and a left and right turning angle of $30^{\\circ}$ . Lastly, the settings for EQA mirror those of PointGoal Navigation as specified in the Habitat-Lab configuration files. All other settings remain consistent with those of [15].", + "bbox": [ + 511, + 529, + 905, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Evaluation Protocol. We assess the TANGO performance across task-specific datasets using the Habitat simulator [44]. As the system is not trained in any way (except for the PointGoal Navigation model, which has been pre-trained on the training set of HM3D, never used in our tests), each scene within episodes is novel to the agent. Therefore, the entire validation set can be categorized as \"unseen\". We now describe in detail the tasks and datasets used:", + "bbox": [ + 511, + 686, + 905, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- OPEN VOCABULARY OBJECT NAVIGATION (OVON) [56]: a large-scale benchmark featuring over 15,000 annotated household objects across 379 categories, derived from real-world 3D scans [38]. The agent is initialized at a random location within a scene and", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24607", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tasked with navigating to a target object category within a time limit of 500 steps. We evaluate our approach across all episodes included in the \"val unseen\" set.", + "bbox": [ + 102, + 90, + 482, + 135 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MULTI-MODAL LIFELONG NAVIGATION (GOAT-BENCH) [26]: an agent is tasked with sequentially navigating to five to ten target objects identified by category names, descriptions, or images. Each target represents a subtask executed within an open-vocabulary framework that spans over 312 categories. The agent is required to reach a goal within a specified time constraint and is assigned new targets upon the completion of each subtask. We evaluate our approach across all episodes included in the \"val unseen\" set.", + "- EMBodied Question ANSWERING (OPENEQA) [34]: containing over 1,600 question-answer pairs sourced from more than 180 real-world environments and scans [14, 38]. It is divided into two task categories: Episodic Memory-EQA (EM-EQA) and Active-EQA (A-EQA). Our focus is on the latter, in which the agent must autonomously navigate and answer questions within a time constraint of 500 steps. Questions span over seven categories, including world knowledge, attribute recognition, spatial reasoning, and object localization. Examples of questions from the episodes include: \"Is the microwave door propped open?\" and \"What is left of the kitchen pass-through?\". We evaluate our approach across all episodes belonging to the A-EQA task category." + ], + "bbox": [ + 89, + 136, + 482, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. In all tasks, we use the standard metrics as in prior works [3, 15, 34, 56]. Namely:", + "bbox": [ + 89, + 531, + 482, + 561 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Success Rate (SR): measures the ratio of episodes where the agent successfully reached its target (Open ObjNav, GOAT-Bench).", + "- Success weighted by Path Length (SPL): measures the optimality of the path taken by the agent w.r.t. the optimal path (Open ObjNav, GOAT-Bench).", + "- Distance to Goal (DTG): measures the average distance to goal of the agent at the end of the episode (Open ObjNav, GOAT-Bench).", + "- Score (LLM-Match): an LLM compares the ground-truth $GT_{i}$ answers with model output $A_{i}$ given a question $Q_{i}$ and assigns a score $\\sigma_{i}$ on a scale of 1 to 5. On this scale, 1 indicates an incorrect response, 5 represents a correct response, and intermediate values reflect varying levels of similarity. For example, the answer to the question \"What color is the bed\" could be correctly answered either as \"white\" or \"the bed is white\". The final results aggregation is as follows:" + ], + "bbox": [ + 89, + 561, + 482, + 833 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS = \\frac {1}{N} \\sum_ {i} ^ {N} \\frac {\\sigma_ {i} - 1}{4} \\times 100 \\% \\tag{1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 838, + 482, + 880 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Answer Accuracy: the average accuracy of the answers", + "bbox": [ + 89, + 886, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "provided by the agent for the EQA task. This metric, used in [15], is presented with the related results in the supplementary material.", + "bbox": [ + 526, + 90, + 903, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implemented Modules. TANGO contains descriptive and easily understandable names for the modules, arguments, and variables to facilitate the LLM comprehension of the input and output. Figure 3 showcases a list of all available modules. Each module exclusively outputs predefined object variables to the next one, enabling the possibility to monitor the progression of the agent at each step. Therefore, dramatically enhancing the explainability of failure cases, as it allows for a deep examination of the agent's decision-making process through its interactions with the environment. Figure 6 provides a general overview of the failure analysis in EQA task. As presented in Section 3, \"navigate_to\" and \"explore_structure\" modules are specifically engineered to navigate the environment, leveraging inputs from depth and RGB sensors as well as the PointGoal GPS + compass sensors. If this sensor is not available for a specific task, it can be derived from the current agent pose and the exploration goal location. The agent is provided with the actions: Forward, Move Left, Move Right, Look Up, Look Down and Stop.", + "bbox": [ + 511, + 191, + 906, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the \"detect\" module, we employ Owlv2 [35] object detector for general classes and use DETR [8] for categories within the COCO classes [33]. Notably, in synthetic 3D scenes, we noticed numerous false positives. To address the problem, bounding boxes produced by the detector are forwarded to a \"classify\" module. Within this module, each detection undergoes classification to differentiate between categories of similar instances. For example, the class \"chair\" encompasses diverse subcategories such as \"armchair\", \"couch\", and \"other\". Subclasses are outputted under the hood directly by the LLM requiring no human annotation. We utilize a CLIP-based classifier in its default configuration [37].", + "bbox": [ + 511, + 501, + 905, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Embodied Question Answering (EQA) task uses an \"answer\" module based on BLIP2 [31], capable of performing various multi-modal tasks, including Visual Question Answering, Image-Text retrieval, and Image Captioning. Notably, the module is also used when goals are specified through images, to capture the class of the input image before starting the navigation. Furthermore, for image-based tasks we exploit a \"match\" module to evaluate the similarity between the agent's observation and the instance image target. This module performs feature matching between the two, as described in Section 3.1, to aid in navigation. Hence, it assists the agent in determining whether what it perceives as the current target is indeed the correct goal.", + "bbox": [ + 511, + 704, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24608", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1dd0c7143bf53da833fd5a99cffd8672f2a01db8f06fed4fc24d4b29efbfbae9.jpg", + "image_caption": [ + "Figure 4. Examples from OpenEQA [34]. The top section illustrates a successful episode where TANGO is able to understand the input query, correctly specifying the sequential targets. The lower section illustrates a failure caused by overly general directions from the LLM, which TANGO struggled to resolve." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 363 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e82169919ca102b03c10664b9d1e54747cf934698a6e99e5adc42f63784f442.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVAL UNSEEN
SR (↑)SPL (↑)
RL [56]18.6 ±0.37.5 ±0.2
BCRL [56]8.0 ±0.22.8 ±0.1
DAgRL [56]18.3 ±0.37.9 ±0.1
*VLFM [55, 56]35.219.6
DAgRL+OD [56]37.1 ±0.219.9 ±0.3
TANGO (ours)35.5 ±0.319.5 ±0.3
", + "bbox": [ + 142, + 473, + 431, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 718, + 292, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Open-set ObjectGoal Navigation We evaluated our method in the standard Open-Set OBJNAV setting, with targets sampled directly from objects in the scene. As shown in Table 1, our approach achieves performance on par with state-of-the-art methods (rows 4-5) on the validation-unseen split both in SR (35.5%) and SPL (19.5%). Instead of focusing solely on numerical results, these findings highlight TANGO's flexibility in solving navigation tasks in unfamiliar environments, underscoring the potential of our approach within the OBJNAV framework.", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9535f5e8c1a3f3950ba4241f50dbfb7900c7cbf6cb18aedc17f2ee7c27b07bae.jpg", + "table_caption": [ + "Table 1. HM3D-OVON. Performance comparison of different methods on VAL UNSEEN split. * deterministic method." + ], + "table_footnote": [], + "table_body": "
MethodScore (↑)
Human Agent [34]85.1 ± 1.1
Blind LLMs [34]35.5 ± 1.7
Socratic LLMs w/ Frame Captions [34]38.1 ± 1.8
Socratic LLMs w/ Scene-Graph Captions [34]34.4 ± 1.8
TANGO (ours)37.2 ± 1.8
", + "bbox": [ + 514, + 88, + 915, + 218 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. OpenEQA results. Comparison of TANGO with the SoA on the OpenEQA dataset. All methods involve zero-shot approaches.", + "bbox": [ + 511, + 227, + 906, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Embodied Question Answering. In EQA, the agent is queried with a natural language question and must autonomously explore the environment and gather information in order to answer accordingly. In recent literature, this task shifted from being regarded as a purely classification problem aimed at determining the most suitable answer from a set of pre-defined possibilities (i.e., class labels) to an open-vocabulary benchmark where the agent must answer with natural language [15, 34]. We compare TANGO results, using the score from Eq. 1, against other zero-shot approaches. Table 2 shows that our method ranks second and closely matches the performance of the leading SoA method, with a gap of less than $1\\%$ . Notably, these approaches [34] utilize LLMs to interpret scene objects step-by-step (rows 3-4), but they do not leverage LLMs to effectively guide navigation towards specific targets. Hence, our approach demonstrates that SoA results can still be achieved by pre-planning a path that prioritizes relevant areas for a given question. Moreover, the results indicate that human agents achieve a significantly higher score of $85\\%$ , which remains well above the performance achieved by using large language models (LLMs), particularly in open-world settings. This gap underscores the challenges LLMs face in handling the complexity and variability of open-world scenarios compared to human agents.", + "bbox": [ + 511, + 277, + 906, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GOAT. We assess our method within the GOAT-Bench setting, where the agent is spawned randomly and tasked with sequential targets. Each goal can be specified either by its category name, a description, or an image (e.g. the input could be \"gas boiler\", \"the gas boiler on the corner of the room. The gas boiler is located on the left of the washing machine and freezer\" or an image of the \"gas boiler\"). Table 3 compares our approach to the methods in [10]. TANGO significantly outperforms other state-of-the-art techniques (rows 1-4), achieving a $+2.6\\%$ gain in the success metric. Additionally, it ranks second in navigation efficiency, with only a slight $-0.7\\%$ difference from the top-performing method. We primarily attribute these results to the LLM's ability to accurately transform target descriptions into a sequence of selected goals, rather than relying", + "bbox": [ + 511, + 672, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24609", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2aa5ccd4f5751c9c77f8fc242b2aa039d28f92a6ee83c3cf9fe4ef15b2a3a4ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVAL UNSEEN
SR (↑)SPL (↑)
SenseAct-NN Skill Chain [26]29.511.3
SenseAct-NN Monolithic [26]12.36.8
Modular GOAT [26]24.917.2
Modular Clip on Wheels [26]12.310.4
TANGO (ours)32.116.5
", + "bbox": [ + 119, + 88, + 455, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. GOAT-Bench. Performance comparison of different methods on the VAL UNSEEN split.", + "bbox": [ + 89, + 247, + 482, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33565f800cdb158f4546f0bafdf71bb56ab35c441bdbeae367a5ef776003ccef.jpg", + "image_caption": [ + "Figure 5. Multi-Modal Lifelong Navigation Success Example. (top) RGB observation of the target during STOP action (step $t_{i+steps}$ ). (middle) Value map for the specific target recomputed from the memory map (step $t_{i+steps}$ ). (bottom) Memory map after target changes (step $t_i$ )." + ], + "image_footnote": [], + "bbox": [ + 94, + 291, + 486, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "solely on potentially \"noisy\" text embeddings. Moreover, the use of a memory mechanism is essential for the task (Fig. 5); however, if the memorized target object is incorrect, the path efficiency can significantly decrease.", + "bbox": [ + 89, + 529, + 483, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. TANGO Failure analysis", + "text_level": 1, + "bbox": [ + 89, + 604, + 316, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to the high explainability of our system, we conducted an analysis of the failures encountered during EQA episodes (Fig. 6). We extracted a significant subsample for the task and manually classified the instances where the model failed. We considered scores below 3 as failures, i.e. incorrect answers. We observed that the main cause of failure can be attributed to the \"detect\" module (stopped at wrong object or Ignored goal object in the image). Furthermore, the exploration policy appears to perform well given the targets produced by the LLM, as it fails in only $\\sim 10\\%$ of the cases (labeled as: didn't see target goal in the image). The LLM generates incorrect pseudo-code around $18\\%$ of the time, with $6.9\\%$ failures leading to ambiguous or incorrect targets, and $11.2\\%$ due to incorrect primitive ordering or usage which instead leads directly to episode failure. Notably, only $11.2\\%$ of errors are actually due to the LLM generating incorrect code, while the remaining is attributable to prompt-related issues. A comprehensive fail-", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/669b7466ee6c54f43c6f6686209f35f2169febe03ef95c60344ba09a232f1c75.jpg", + "image_caption": [ + "Figure 6. Causes of system failure in EQA tasks. The analysis was manually conducted on a significant sample of failed episodes, facilitated by the explainability of TANGO. Answers were deemed as \"incorrect\" if they received a score below three." + ], + "image_footnote": [], + "bbox": [ + 529, + 90, + 893, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ure analysis is provided in the supplementary material.", + "bbox": [ + 511, + 375, + 874, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 412, + 632, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted a systematic analysis of our compositional approach, offering valuable insights into the future of modular, neuro-symbolic systems. Moreover, TANGO highlights the versatility of simple PointGoal Navigation agents equipped with specific task-oriented modules, yielding promising results in zero-shot scenarios across all considered tasks. TANGO effectively handles diverse multi-modal prompts, following instructions to complete tasks, underscoring the potential of LLMs in robotic navigation. Its integrated memory supports lifelong navigation, suggesting improved capabilities with increased exploration. However, it may struggle with overly complex or ambiguous prompts, which can limit the LLM's ability to identify the correct primitives or targets. Future work could extend this system to tasks such as Visual Language Navigation (VLN) and explore the use of open-source LLMs. Additionally, refining the memory mechanism—particularly through improved sampling strategies for identifying high-value pixels linked to objects' memory—could further enhance the system's effectiveness.", + "bbox": [ + 511, + 441, + 906, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 766, + 671, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank the University of Padua, Department of Mathematics \"Tullio Levi-Civita\", for providing the computational resources. TC and LS were supported by the PNRR project Future AI Research (FAIR - PE00000013) under the NRRP MUR program, funded by NextGenerationEU. We also thank Elena Izzo and Luca Parolari for their valuable comments on early drafts of this paper.", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24610", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[2] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 3", + "[3] Dhruv Batra, Aaron Gokaslan, Aniruddha Kembhavi, Oleksandr Maksymets, Roozbeh Mottaghi, Manolis Savva, Alexander Toshev, and Erik Wijmans. ObjectNav revisited: On evaluation of embodied agents navigating to objects. arXiv preprint arXiv:2006.13171, 2020. 2, 5, 6", + "[4] Zhenshan Bing, Alexander Koch, Xiangtong Yao, Kai Huang, and Alois Knoll. Meta-reinforcement learning via language instructions. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5985-5991. IEEE, 2023. 2", + "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Proc. of Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 3", + "[6] Tommaso Campari, Paolo Eccher, Luciano Serafini, and Lamberto Ballan. Exploiting scene-specific features for object goal navigation. In Proc. of the European Conference on Computer Vision Workshops (ECCVW), 2020. 2", + "[7] Tommaso Campari, Leonardo Lamanna, Paolo Traverso, Luciano Serafini, and Lamberto Ballan. Online learning of reusable abstract models for object goal navigation. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5, 6", + "[9] Angel X. Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. In Proc. of the International Conference on 3D Vision (3DV), 2017. 4", + "[10] Matthew Chang, Theophile Gervet, Mukul Khanna, Sriram Yenamandra, Dhruv Shah, So Yeon Min, Kavit Shah, Chris Paxton, Saurabh Gupta, Dhruv Batra, et al. Goat: Go to any thing. arXiv preprint arXiv:2311.06430, 2023. 7", + "[11] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to ex" + ], + "bbox": [ + 94, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "plore using active neural SLAM. In Proc. of the International Conference on Learning Representations (ICLR), 2019. 2", + "[12] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. In Proc. of Advances in Neural Information Processing Systems (NeurIPS), 2020. 2", + "[13] Devendra Singh Chaplot, Ruslan Salakhutdinov, Abhinav Gupta, and Saurabh Gupta. Neural topological slam for visual navigation. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[14] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Niessner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6", + "[15] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 5, 6, 7", + "[16] Vishnu Sashank Dorbala, James F Mullen Jr, and Dinesh Manocha. Can an embodied agent find your \"cat-shaped mug\"? llm-based zero-shot object navigation. IEEE Robotics and Automation Letters, 2023. 2, 3", + "[17] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79), 2023. 2", + "[18] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5", + "[19] Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 1, 2, 3", + "[20] Ronghang Hu, Jacob Andreas, Trevor Darrell, and Kate Saenko. Explainable neural computation via stack neural module networks. In Proc. of the European Conference on Computer Vision (ECCV), 2018. 3", + "[21] C. et al. Huang. Visual language maps for robot navigation. In 2023 IEEE Int. Conf. Robot. Autom. (ICRA), pages 10608-10615, 2023. 2", + "[22] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1, 2", + "[23] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 1, 2", + "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7B. arXiv preprint arXiv:2310.06825, 2023. 1, 2" + ], + "bbox": [ + 517, + 93, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24611", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C. Lawrence Zitnick, and Ross B. Girshick. Inferring and executing programs for visual reasoning. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 3", + "[26] Mukul Khanna, Ram Ramrakhya, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. Goat-bench: A benchmark for multi-modal lifelong navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16373-16383, 2024. 2, 5, 6, 8", + "[27] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 2", + "[28] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances. arXiv preprint arXiv:2211.15876, 2022. 2", + "[29] Jacob Krantz, Theophile Gervet, Karmesh Yadav, Austin Wang, Chris Paxton, Roozbeh Mottaghi, Dhruv Batra, Jitendra Malik, Stefan Lee, and Devendra Singh Chaplot. Navigating to objects specified by images. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[30] Dingbang Li, Wenzhou Chen, and Xin Lin. Tina: Think, interaction, and action framework for zero-shot vision language navigation. arXiv preprint arXiv:2403.08833, 2024. 2", + "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 4, 5, 6", + "[32] J. et al. Liang. Code as policies: Language model programs for embodied control. In 2023 IEEE Int. Conf. Robot. Autom. (ICRA), pages 9493-9500, 2023. 1, 2", + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6", + "[34] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. Openeqa: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2, 6, 7", + "[35] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36, 2024. 5, 6", + "[36] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai," + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Alexander William Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondrus, Theophile Gervet, Vincent-Pierre Berges, John M. Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars and robots, 2023. 2", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[38] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X. Chang, et al. Habitat-matterport 3D dataset (HM3D): 1000 large-scale 3D environments for embodied AI. arXiv preprint arXiv:2109.08238, 2021. 4, 5, 6", + "[39] Santhosh K. Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[40] Sonia Raychaudhuri, Tommaso Campari, Unnat Jain, Manolis Savva, and Angel X Chang. Mopa: Modular object navigation with pointgoal agents. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5763-5773, 2024. 2, 4", + "[41] G. et al. Sarch. Open-ended instructable embodied agents with memory-augmented llms. arXiv:2310.15127, 2023. 2", + "[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks, 2020. 5", + "[43] Manolis Savva, Angel X. Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. Minos: Multimodal indoor simulator for navigation in complex environments. arXiv preprint arXiv:1712.03931, 2017. 2", + "[44] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied AI research. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5", + "[45] Bokui Shen, Fei Xia, Chengshu Li, Roberto Martín-Martín, Linxi Fan, Guanzhi Wang, Claudia Pérez-D'Arpino, Shyamal Buch, Sanjana Srivastava, Lyne Tchapmi, et al. iGibson 1.0: a simulation environment for interactive tasks in large realistic scenes. In Proc. of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 2", + "[46] C.H. et al. Song. Llm-planner: Few-shot grounded planning for embodied agents. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pages 2998-3009, 2023. 2", + "[47] Dídac Surís, Sachit Menon, and Carl Vondrick. ViperGPT: Visual Inference via Python Execution for Reasoning. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24612", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2", + "[49] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2", + "[50] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. DD-PPO: Learning near-perfect pointgoal navigators from 2.5 billion frames. In Proc. of the International Conference on Learning Representations (ICLR), 2019. 2", + "[51] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 3", + "[52] Zhenyu Wu, Ziwei Wang, Xiuwei Xu, Jiwen Lu, and Haibin Yan. Embodied task planning with large language models. arXiv preprint arXiv:2307.01848, 2023. 1", + "[53] Joel Ye, Dhruv Batra, Erik Wijmans, and Abhishek Das. Auxiliary Tasks Speed Up Learning PointGoal Navigation. In Proc. of the International Conference on Robot Learning (CoRL), 2020. 2", + "[54] Joel Ye, Dhruv Batra, Abhishek Das, and Erik Wijmans. Auxiliary Tasks and Exploration Enable ObjectNav. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[55] Naoki Yokoyama, Sehoon Ha, Dhruv Batra, Jiuguang Wang, and Bernadette Bucher. Vlfm: Vision-language frontier maps for zero-shot semantic navigation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 42-48. IEEE, 2024. 2, 3, 4, 7", + "[56] Naoki Yokoyama, Ram Ramrakhya, Abhishek Das, Dhruv Batra, and Sehoon Ha. Hm3d-ovon: A dataset and benchmark for open-vocabulary object goal navigation. arXiv preprint arXiv:2409.14296, 2024. 2, 5, 6, 7", + "[57] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target navigation. in 2023 IEEE. In RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3554-3560. 2", + "[58] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1, 3", + "[59] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 2, 3" + ], + "bbox": [ + 91, + 92, + 480, + 810 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24613", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_model.json b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f105d2545b697f9bcedcd8a01ab3d12cc64babfe --- /dev/null +++ b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_model.json @@ -0,0 +1,2048 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "TANGO: Training-free Embodied AI Agents for Open-world Tasks" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.18, + 0.289, + 0.2 + ], + "angle": 0, + "content": "Filippo Ziliotto\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.181, + 0.501, + 0.2 + ], + "angle": 0, + "content": "Tommaso Campari*" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.181, + 0.676, + 0.198 + ], + "angle": 0, + "content": "Luciano Serafini²" + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.181, + 0.857, + 0.198 + ], + "angle": 0, + "content": "Lamberto Ballan1" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.216, + 0.613, + 0.234 + ], + "angle": 0, + "content": "1 University of Padova, Italy" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.235, + 0.69, + 0.253 + ], + "angle": 0, + "content": "Fondazione Bruno Kessler (FBK), Trento, Italy" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.607 + ], + "angle": 0, + "content": "Large Language Models (LLMs) have demonstrated excellent capabilities in composing various modules together to create programs that can perform complex reasoning tasks on images. In this paper, we propose TANGO, an approach that extends the program composition via LLMs already observed for images, aiming to integrate those capabilities into embodied agents capable of observing and acting in the world. Specifically, by employing a simple PointGoal Navigation model combined with a memory-based exploration policy as a foundational primitive for guiding an agent through the world, we show how a single model can address diverse tasks without additional training. We task an LLM with composing the provided primitives to solve a specific task, using only a few in-context examples in the prompt. We evaluate our approach on three key Embodied AI tasks: Open-Set ObjectGoal Navigation, MultiModal Lifelong Navigation, and Open Embodied Question Answering, achieving state-of-the-art results without any specific fine-tuning in challenging zero-shot scenarios." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.637, + 0.222, + 0.652 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.844 + ], + "angle": 0, + "content": "Large Language Models (LLMs) have gained significant attention in the field of AI due to their remarkable capability to generalize across unseen tasks [5, 24, 48, 49]. Systems like VisProg and ViperGPT [18, 47] demonstrated strong performance on a broad range of vision-and-language tasks by just relying on a few numbers of in-context examples to generate complex compositional programs without requiring any specific training. Nevertheless, the evaluation of these concepts to Embodied AI, particularly for navigation tasks, remains largely unexplored. This approach may be crucial in building agents capable to navigate and operate efficiently in unfamiliar environments." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.845, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Previously, methods such as Neural Module Networks (NMN) [2, 19] have demonstrated promising compositional" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.285, + 0.9, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.606, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Figure 1. We introduce TANGO, a modular neuro-symbolic system for compositional embodied visual navigation. Given a few examples of natural language instructions and the corresponding programs composed of action primitives, TANGO can generate executable programs, enabling the agent to perform multiple tasks within a 3D environment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.907, + 0.868 + ], + "angle": 0, + "content": "properties for high-level tasks like visual question answering, via end-to-end training of networks and specialized, differentiable neural modules. However, they required the combination of semantic parsers and predefined templates to learn to solve the compositional tasks. Alternative approaches leverage LLMs through predefined task modules or APIs for action execution [22, 23, 32, 52, 58]. While effective, these methods often lack a compositional structure to address multiple tasks or primarily focus on manipulation and robotic planning, with limited emphasis on navigation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this work, we present TANGO (see Figure 1) a novel neuro-symbolic compositional approach which uti-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.352, + 0.901 + ], + "angle": 0, + "content": "*The authors contributed equally to the work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24603" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.363 + ], + "angle": 0, + "content": "izes primitives and employs a LLM as a planner to sequence these primitives within photorealistic 3D environments, where agents must perceive and act. This framework integrates high-level planning with low-level action execution, without the need for training. TANGO makes use of diverse modules designed for visual navigation and question-answering tasks, resulting in a system that seamlessly addresses both challenges while achieving SoA performance without requiring any prior adjustments. By providing a few in-context examples that show how to tackle multiple tasks, TANGO is capable of generalizing to the specific task at hand. This is facilitated by the LLM, which effectively combines the individual modules available within the TANGO system. Moreover, it extends [55] exploration policy by incorporating a memory mechanism in the form of a stored feature map that retains information about previously explored areas, supporting for efficient life-long navigation tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.482, + 0.483 + ], + "angle": 0, + "content": "To demonstrate the flexibility of the proposed framework, TANGO has been tested on three popular benchmarks: namely, i) Open-Vocabulary ObjectGoal Navigation [56], ii) Multi-Modal Lifelong Navigation [26], and iii) Embodied Question Answering [15, 34]. Results match or surpass previous approaches, without requiring specialized training. Summing up, the contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.486, + 0.483, + 0.529 + ], + "angle": 0, + "content": "- Introduction of a neuro-symbolic compositional LLM-based framework for EAI leveraging specialized primitive modules." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.532, + 0.483, + 0.576 + ], + "angle": 0, + "content": "- Demonstration of robust generalization capabilities across multiple tasks without the need for any specific training or fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.578, + 0.483, + 0.621 + ], + "angle": 0, + "content": "- Extension of the exploration policy presented in [55] to multi-goal scenarios through the incorporation of a memory mechanism stored as a feature vector map." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.623, + 0.483, + 0.652 + ], + "angle": 0, + "content": "- Achievement of state-of-the-art results, underscoring the effectiveness of the proposed framework." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.486, + 0.483, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.669, + 0.241, + 0.685 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.483, + 0.846 + ], + "angle": 0, + "content": "Our work takes inspiration from modular approaches for visual tasks, like the seminal Neural Module Networks [2, 19], as well as the recent VisProg [18] and ViperGPT [47] frameworks. The main contribution of this paper is to extend these ideas to embodied visual navigation, that usually require heavy end-to-end learning, although modular approaches have recently shown promising results [7, 17]. Therefore, we discuss prior work in the area of embodied AI, the use of language models in robotics, and program generation approaches for image recognition tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Embodied AI and Visual Navigation. The field of Embodied AI has recently undergone a paradigm shift, fuelled" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.363 + ], + "angle": 0, + "content": "by the emergence of highly efficient simulators [27, 36, 43-45]. These simulators enable processing numerous parallel simulations in photorealistic indoor environments, facilitating large-scale testing that was otherwise challenging in classical robotics. Alongside these simulators, a large variety of tasks (and benchmarks) has been proposed: PointGoal Navigation [50], where an agent navigates from point A to point B in an unknown environment; ObjectGoal Navigation [3], requiring the agent to locate and navigate to an object in the scene; Instance-Image Goal Navigation [28], akin to ObjectGoal Navigation, but in which the agent should find a specific object depicted in a given image; Multi-Modal Lifelong Navigation [26] involving navigating to a sequence of target objects that can be specified through labels, images, or textual descriptions; Embodied Question Answering [15, 34], in which an agent navigates an environment to gather information needed to answer a question, and Vision and Language Navigation (VLN) [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.905, + 0.635 + ], + "angle": 0, + "content": "Various approaches have been proposed to tackle these tasks, primarily falling into three categories: end-to-end [6, 53, 54], modular [11-13, 29, 39, 40] and LLM-based [16, 30, 57-59]. However, the main limitation of these paradigms is their dependence on task-specific architectures. In the case of end-to-end approaches, a model is trained for a specific problem, often requiring days of training, such as in [50], where a policy for PointGoal Navigation was trained for 2.5 billion steps. To adapt the learned model to a new problem, adjustments and retraining are necessary, and this comes with a high cost. Modular approaches share a similar challenge, usually relying on policies designed for specific tasks, although common elements, such as exploration and navigation modules, are often shared [11, 12, 29, 39, 40]. Adapting these solutions to different tasks requires a manual adjustment of the approach to fit the new domain, often involving the addition or modification of modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Language models for Robotics In robotics, foundation models (LLMs) trained on vast internet-scale datasets [5, 24, 48, 49] have the potential to equip robots with real-world priors and advanced reasoning abilities without the need for extensive task-specific training. Early approaches equipped agents with learned language embeddings, requiring large amounts of training data [4, 32]. Recent studies, on the other hand, have explored zero-shot and few-shot solutions mainly focusing on robotic planning and manipulation tasks [22, 23, 32, 41, 46]. In the context of visual navigation, LLM-based approaches leverage the powerful LLMs priors and reasoning capabilities to guide navigation within the environment [16, 21, 30, 57-59]; however, despite their remarkable contributions, they often lack a modular design that allows for flexible integration and extensibility, as they are typically optimized for single tasks rather" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24604" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.388, + 0.106 + ], + "angle": 0, + "content": "than a comprehensive, adaptable framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.486, + 0.322 + ], + "angle": 0, + "content": "In contrast, our work deviates from these paradigms by relying on a series of diverse pre-trained modules that, thanks to a LLM, can be combined to potentially solve various tasks, compositionally. None of our modules are fine-tuned explicitly for the target problems, placing us in a zero-shot setting. Moreover, TANGO does not only rely on an exploration policy based entirely on LLM output, as in [16, 58, 59]. Instead, it combines a frontier exploration policy with LLM priors and memory guidance to enhance navigation performance. To achieve this, we use a pre-trained PointGoal policy as the base waypoint navigation module of TANGO and extend the exploration policy introduced in [55] by equipping the agent with a memory mechanism stored as a feature vector for each pixel of the map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.485, + 0.552 + ], + "angle": 0, + "content": "Modular vision and program composition for visual tasks. Neural Module Networks (NMN) [2, 19] introduced modular and compositional methodologies for visual question answering (VQA). NMNs integrate neural modules into an end-to-end differentiable network; the approach originally relied on pre-existing parsers [2], whereas more recent methods [19, 20, 25] have evolved to learn the layout generation model concurrently with the neural modules, employing reinforcement learning [51] and weak supervision. Stack-NMN [20] extends N2NMN by transitioning from discrete to soft layout generation, incorporating a weighted average of predictions from all modules at each step, determined by a layout generation network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Recently, [18] introduced VisProg, a framework that offers two key advantages over NMNs. Firstly, it constructs high-level programs that invoke state-of-the-art neural models and Python functions at intermediate steps, diverging from the conventional approach of generating end-to-end neural networks. This design facilitates the integration of symbolic, non-differentiable modules. Secondly, it leverages the in-context learning ability of LLMs [5]. This is achieved by prompting the LLM with a natural language instruction (or a visual question or a statement to be verified), along with a few examples of similar instructions and their corresponding programs. Almost concurrently, ViperGPT [47] presented a similar approach to VisProg, but in this case the model directly generates Python code, instead of composing pre-defined modules and routines. In contrast, in our framework, the LLM is provided only with high-level knowledge of the functions of each primitive, which is encoded directly in the primitive's name. Therefore, the system it is built similarly to VisProg. It is noteworthy that both VisProg and ViperGPT were tested on images, while our model is specifically designed for embodied agents operating in an interactive environment, relying on sensor-based perception." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.606, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.908, + 0.405 + ], + "angle": 0, + "content": "Neuro-symbolic approaches offer the possibility to address a broad range of diverse complex tasks efficiently. Systems like VisProg [18] operate on images and have demonstrated excellent results in zero-shot settings. Understanding how a similar approach works in an action-oriented scenario is key to enable robots to navigate autonomously in novel environments. Therefore, we develop this paradigm further into EAI, where the integration with action execution adds several challenges. Inspired by [18], given an input prompt, we rely on a LLM to generate synthetic pseudo-programs that can be executed by the embodied agent in the environment. Our TANGO framework is therefore able to generalize to new tasks, without any direct training on task-specific datasets, thus effectively mitigating the heavy computational training costs inherent to embodied navigation tasks. The procedure that enables generating these new programs is based on \"in-context examples\" that are fed to the LLM, alongside a natural language instruction. An outline of TANGO is shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.415, + 0.907, + 0.492 + ], + "angle": 0, + "content": "TANGO Interpreter. The first key component of the proposed framework is referred to as \"Program Interpreter\". It comprises visual recognition modules that can be used by the agent to extract the semantics of the scene, as well as to provide an understanding of the visual context." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.493, + 0.909, + 0.84 + ], + "angle": 0, + "content": "As shown in Figure 1, users can ask questions or give specific tasks to the agent; the natural language prompt is then processed by the LLM (in our implementation GPT-4o [5]), which serves as a planner and outputs a step-by-step executable program. To ensure the LLM delivers a reasonable output, it is fed with 15 \"in-context examples\" across diverse tasks. This enables the LLM to make use of its reasoning capabilities effectively, identifying the most suitable planning for the required task. Moreover, the examples remain the same regardless of the task identified in the prompt; it is the LLM's responsibility to output the specific program target for the given question. Programs use a higher level of abstraction than previous modular attempts such as Neural Module Networks (NMN) [2, 19]. Each program is constructed as a sequence of primitives (e.g., detect, answer, match, etc.) that invoke corresponding TANGO modules. These modules are either powered by pre-trained state-of-the-art vision models or implemented as simple Python subroutines (e.g., count, is_found, eval, etc.), with additional navigation-specific modules designed to \"steer\" the agent's movements (e.g. navigate_to, return, turn, etc.). Figure 3 provides a comprehensive list of all the currently implemented modules." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Figure 2 shows an Embodied Question Answering (EQA) example for the user's question: \"on the left of the toilet there's a shower, is the shower curtain open or closed?\" The LLM, which acts as a planner, transforms the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24605" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.088, + 0.891, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.492, + 0.908, + 0.522 + ], + "angle": 0, + "content": "Figure 2. Overview of the program generation in TANGO. Given few \"in-context examples\", the LLM provide a detailed sequence of steps to be executed by the agent in the given environment. The LLM is instructed to comment its output to allow for explainability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.547, + 0.483, + 0.684 + ], + "angle": 0, + "content": "user's initial prompt into navigation subtasks by leveraging its reasoning abilities and powerful priors. This decomposition enables the model to break down complex queries into more manageable steps, guiding the agent's navigation process effectively. Each line of the generated program corresponds to a module serving a specific task. The agent must then execute the generated program. The LLM is also guided to comment on its steps directly within the generated pseudocode." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.837 + ], + "angle": 0, + "content": "All the given primitives are equipped with methods to: i) parse lines in order to extract input argument names and values, as well as the output variable name; ii) execute the module in the environment, and update the program state with the output variable name and value. The outputs for each step and the comments in the generated pseudocode can also be used to better understand what is happening under the hood and why the LLM generated a specific line, thus enabling a good interpretability of the system behavior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "TANGO modules utilize open-source software and models, readily downloadable from the web. They can also be effortlessly updated with newer, more efficient models as they become available." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.73 + ], + "angle": 0, + "content": "Navigation Module. In order to navigate the environment, we define a module employing a PointGoal navigation agent as our foundational module [40, 55]. This model achieved nearly perfect PointGoal performance, both in terms of success (\\(\\sim 99\\%\\)) and efficiency (a forward pass that takes a fraction of a second) on standard datasets [9, 38]. The agent starts the exploration of the environment until it locates its target goal. Once the target is identified, the focus of exploration transitions to reaching the designated goal. This PointNav policy exclusively uses the egocentric depth image and the robot's relative distance and heading towards the desired goal point as input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Exploration Policy. Exploration is performed using the policy outlined in [55]. This method builds occupancy maps based on depth observations to identify exploration frontiers. It further leverages RGB observations and a pretrained vision-language model (BLIP2 [31]) to generate a language-grounded \"value map\". This value map is then used to guide the exploration, facilitating an efficient search for instances of a given target. Notably, as in [55], the agent performs a \\(360^{\\circ}\\) turn at the start of navigation to initialize frontiers." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We extend this policy for sequential goals by incorpora" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.483, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.268 + ], + "angle": 0, + "content": "Figure 3. Overview of TANGO modules. Modules span a variety of inputs and outputs. Orange modules use Python subroutines, while blue modules use pre-trained computer vision models (similarly to [18]). The navigate_to and explore(scene modules, in green, both implement our foundational PointNav module; however, only explorescene integrates the memory mechanism." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.294, + 0.483, + 0.461 + ], + "angle": 0, + "content": "ing a memory mechanism, represented as a “feature map” in which each pixel of the value map is encoded as a vector and updated at each step. This feature map then updates the language-grounded original value map when a new target is specified, by calculating the cosine similarity between each pixel's vector and the new target embeddings (either from text or image). We sample the highest value in the updated value map; if this value exceeds a predefined threshold, indicating the agent has previously encountered and “rememberers” the target's location, the agent navigates to it, supporting lifelong navigation (see Section 2-Supp)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.367, + 0.485 + ], + "angle": 0, + "content": "3.1. Navigation Tasks with TANGO" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.553 + ], + "angle": 0, + "content": "TANGO leverages pre-trained multimodal vision-language and vision-only models as foundational components to extract semantic information from the scene, making it well-suited for several Embodied AI tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.706 + ], + "angle": 0, + "content": "(Open-set) ObjectGoal Navigation [3, 56], in particular, emerges as a suitable testbed to evaluate the efficacy of our method. A key module to tackle this problem involves utilizing an object detector (Owlv2 in our implementation [35] or DETR [8] for target objects that fall within the COCO classes [33]) to identify objects within the image. Subsequently, navigation towards the detected objects is facilitated by waypoints, leveraging depth distance calculation to guide the agent effectively. For simplicity, we use the center of the bounding box to determine the target waypoint." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.811 + ], + "angle": 0, + "content": "TANGO also integrates a specialized module for matching target objectives with ongoing visual observations in image-based scenarios, resulting in accurate navigation towards an instance image. To evaluate our agent's performance in this context, we use real-time matching SuperGlue network [42], optimized for indoor environments (i.e. using hyperparameters recommended in the respective paper)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "TANGO being task-agnostic, can process any type of navigation target, regardless of the order in which it is presented. In this context, the GOAT benchmark [26] is well-suited for evaluating our system, as it operates in scenarios where targets are specified through images, text, or descriptive phrases, provided in random sequence. It also heavily" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "relies on memory to find previously seen targets when sequential goals are given." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.258 + ], + "angle": 0, + "content": "An agent should also be able to answer user queries like \"can you check if the kitchen table is clean?\" Hence, EQA serves as an excellent benchmark to assess TANGO's capabilities and robustness. EQA consists of three phases: understanding the semantic structure of the prompt, locating the target object(s), and analyzing the visual semantics to generate an accurate response. To this end, TANGO integrates a specific answer module, relying on BLIP2 [31] as its core foundation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.273, + 0.646, + 0.29 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.39 + ], + "angle": 0, + "content": "TANGO provides a flexible framework that can be applied to various embodied navigation problems. We evaluate our approach on three popular tasks that require a wide range of capabilities, including efficient environment exploration, path planning, scene and context understanding, and image similarity comparison." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.4, + 0.704, + 0.417 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.529 + ], + "angle": 0, + "content": "Agent Configuration. Prior research on visual navigation commonly uses different agent configurations depending on the considered task [15, 26, 56]. The configuration typically employed for (OPEN-SET) OBJNAV closely resembles that of a LoCoBot, with an agent height of \\(0.88m\\), a radius of \\(0.18m\\), and a single \\(640 \\times 480\\) RGB sensor with a \\(79^{\\circ}\\) hfov, positioned \\(0.88m\\) above the ground." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.906, + 0.665 + ], + "angle": 0, + "content": "In the context of GOAT Benchmark, some of the default settings differ, featuring a camera situated \\(1.31m\\) above the ground. The agent has a height of \\(1.41m\\) and is given \\(360 \\times 640\\) RGB images. Both the above tasks' configurations use a step size of \\(0.25m\\) and a left and right turning angle of \\(30^{\\circ}\\). Lastly, the settings for EQA mirror those of PointGoal Navigation as specified in the Habitat-Lab configuration files. All other settings remain consistent with those of [15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.821 + ], + "angle": 0, + "content": "Datasets and Evaluation Protocol. We assess the TANGO performance across task-specific datasets using the Habitat simulator [44]. As the system is not trained in any way (except for the PointGoal Navigation model, which has been pre-trained on the training set of HM3D, never used in our tests), each scene within episodes is novel to the agent. Therefore, the entire validation set can be categorized as \"unseen\". We now describe in detail the tasks and datasets used:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "- OPEN VOCABULARY OBJECT NAVIGATION (OVON) [56]: a large-scale benchmark featuring over 15,000 annotated household objects across 379 categories, derived from real-world 3D scans [38]. The agent is initialized at a random location within a scene and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24607" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.092, + 0.483, + 0.136 + ], + "angle": 0, + "content": "tasked with navigating to a target object category within a time limit of 500 steps. We evaluate our approach across all episodes included in the \"val unseen\" set." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.137, + 0.483, + 0.287 + ], + "angle": 0, + "content": "- MULTI-MODAL LIFELONG NAVIGATION (GOAT-BENCH) [26]: an agent is tasked with sequentially navigating to five to ten target objects identified by category names, descriptions, or images. Each target represents a subtask executed within an open-vocabulary framework that spans over 312 categories. The agent is required to reach a goal within a specified time constraint and is assigned new targets upon the completion of each subtask. We evaluate our approach across all episodes included in the \"val unseen\" set." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.288, + 0.483, + 0.515 + ], + "angle": 0, + "content": "- EMBodied Question ANSWERING (OPENEQA) [34]: containing over 1,600 question-answer pairs sourced from more than 180 real-world environments and scans [14, 38]. It is divided into two task categories: Episodic Memory-EQA (EM-EQA) and Active-EQA (A-EQA). Our focus is on the latter, in which the agent must autonomously navigate and answer questions within a time constraint of 500 steps. Questions span over seven categories, including world knowledge, attribute recognition, spatial reasoning, and object localization. Examples of questions from the episodes include: \"Is the microwave door propped open?\" and \"What is left of the kitchen pass-through?\". We evaluate our approach across all episodes belonging to the A-EQA task category." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.137, + 0.483, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.483, + 0.562 + ], + "angle": 0, + "content": "Evaluation Metrics. In all tasks, we use the standard metrics as in prior works [3, 15, 34, 56]. Namely:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.562, + 0.483, + 0.607 + ], + "angle": 0, + "content": "- Success Rate (SR): measures the ratio of episodes where the agent successfully reached its target (Open ObjNav, GOAT-Bench)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.608, + 0.483, + 0.652 + ], + "angle": 0, + "content": "- Success weighted by Path Length (SPL): measures the optimality of the path taken by the agent w.r.t. the optimal path (Open ObjNav, GOAT-Bench)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.653, + 0.483, + 0.697 + ], + "angle": 0, + "content": "- Distance to Goal (DTG): measures the average distance to goal of the agent at the end of the episode (Open ObjNav, GOAT-Bench)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.698, + 0.483, + 0.834 + ], + "angle": 0, + "content": "- Score (LLM-Match): an LLM compares the ground-truth \\(GT_{i}\\) answers with model output \\(A_{i}\\) given a question \\(Q_{i}\\) and assigns a score \\(\\sigma_{i}\\) on a scale of 1 to 5. On this scale, 1 indicates an incorrect response, 5 represents a correct response, and intermediate values reflect varying levels of similarity. For example, the answer to the question \"What color is the bed\" could be correctly answered either as \"white\" or \"the bed is white\". The final results aggregation is as follows:" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.562, + 0.483, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.839, + 0.483, + 0.881 + ], + "angle": 0, + "content": "\\[\nS = \\frac {1}{N} \\sum_ {i} ^ {N} \\frac {\\sigma_ {i} - 1}{4} \\times 100 \\% \\tag{1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Answer Accuracy: the average accuracy of the answers" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "provided by the agent for the EQA task. This metric, used in [15], is presented with the related results in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.193, + 0.907, + 0.495 + ], + "angle": 0, + "content": "Implemented Modules. TANGO contains descriptive and easily understandable names for the modules, arguments, and variables to facilitate the LLM comprehension of the input and output. Figure 3 showcases a list of all available modules. Each module exclusively outputs predefined object variables to the next one, enabling the possibility to monitor the progression of the agent at each step. Therefore, dramatically enhancing the explainability of failure cases, as it allows for a deep examination of the agent's decision-making process through its interactions with the environment. Figure 6 provides a general overview of the failure analysis in EQA task. As presented in Section 3, \"navigate_to\" and \"explore_structure\" modules are specifically engineered to navigate the environment, leveraging inputs from depth and RGB sensors as well as the PointGoal GPS + compass sensors. If this sensor is not available for a specific task, it can be derived from the current agent pose and the exploration goal location. The agent is provided with the actions: Forward, Move Left, Move Right, Look Up, Look Down and Stop." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.906, + 0.699 + ], + "angle": 0, + "content": "For the \"detect\" module, we employ Owlv2 [35] object detector for general classes and use DETR [8] for categories within the COCO classes [33]. Notably, in synthetic 3D scenes, we noticed numerous false positives. To address the problem, bounding boxes produced by the detector are forwarded to a \"classify\" module. Within this module, each detection undergoes classification to differentiate between categories of similar instances. For example, the class \"chair\" encompasses diverse subcategories such as \"armchair\", \"couch\", and \"other\". Subclasses are outputted under the hood directly by the LLM requiring no human annotation. We utilize a CLIP-based classifier in its default configuration [37]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Embodied Question Answering (EQA) task uses an \"answer\" module based on BLIP2 [31], capable of performing various multi-modal tasks, including Visual Question Answering, Image-Text retrieval, and Image Captioning. Notably, the module is also used when goals are specified through images, to capture the class of the input image before starting the navigation. Furthermore, for image-based tasks we exploit a \"match\" module to evaluate the similarity between the agent's observation and the instance image target. This module performs feature matching between the two, as described in Section 3.1, to aid in navigation. Hence, it assists the agent in determining whether what it perceives as the current target is indeed the correct goal." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.482, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.375, + 0.483, + 0.445 + ], + "angle": 0, + "content": "Figure 4. Examples from OpenEQA [34]. The top section illustrates a successful episode where TANGO is able to understand the input query, correctly specifying the sequential targets. The lower section illustrates a failure caused by overly general directions from the LLM, which TANGO struggled to resolve." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.474, + 0.432, + 0.639 + ], + "angle": 0, + "content": "
MethodVAL UNSEEN
SR (↑)SPL (↑)
RL [56]18.6 ±0.37.5 ±0.2
BCRL [56]8.0 ±0.22.8 ±0.1
DAgRL [56]18.3 ±0.37.9 ±0.1
*VLFM [55, 56]35.219.6
DAgRL+OD [56]37.1 ±0.219.9 ±0.3
TANGO (ours)35.5 ±0.319.5 ±0.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.677 + ], + "angle": 0, + "content": "Table 1. HM3D-OVON. Performance comparison of different methods on VAL UNSEEN split. * deterministic method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.294, + 0.736 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Open-set ObjectGoal Navigation We evaluated our method in the standard Open-Set OBJNAV setting, with targets sampled directly from objects in the scene. As shown in Table 1, our approach achieves performance on par with state-of-the-art methods (rows 4-5) on the validation-unseen split both in SR (35.5%) and SPL (19.5%). Instead of focusing solely on numerical results, these findings highlight TANGO's flexibility in solving navigation tasks in unfamiliar environments, underscoring the potential of our approach within the OBJNAV framework." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.916, + 0.219 + ], + "angle": 0, + "content": "
MethodScore (↑)
Human Agent [34]85.1 ± 1.1
Blind LLMs [34]35.5 ± 1.7
Socratic LLMs w/ Frame Captions [34]38.1 ± 1.8
Socratic LLMs w/ Scene-Graph Captions [34]34.4 ± 1.8
TANGO (ours)37.2 ± 1.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.27 + ], + "angle": 0, + "content": "Table 2. OpenEQA results. Comparison of TANGO with the SoA on the OpenEQA dataset. All methods involve zero-shot approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.908, + 0.657 + ], + "angle": 0, + "content": "Embodied Question Answering. In EQA, the agent is queried with a natural language question and must autonomously explore the environment and gather information in order to answer accordingly. In recent literature, this task shifted from being regarded as a purely classification problem aimed at determining the most suitable answer from a set of pre-defined possibilities (i.e., class labels) to an open-vocabulary benchmark where the agent must answer with natural language [15, 34]. We compare TANGO results, using the score from Eq. 1, against other zero-shot approaches. Table 2 shows that our method ranks second and closely matches the performance of the leading SoA method, with a gap of less than \\(1\\%\\). Notably, these approaches [34] utilize LLMs to interpret scene objects step-by-step (rows 3-4), but they do not leverage LLMs to effectively guide navigation towards specific targets. Hence, our approach demonstrates that SoA results can still be achieved by pre-planning a path that prioritizes relevant areas for a given question. Moreover, the results indicate that human agents achieve a significantly higher score of \\(85\\%\\), which remains well above the performance achieved by using large language models (LLMs), particularly in open-world settings. This gap underscores the challenges LLMs face in handling the complexity and variability of open-world scenarios compared to human agents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.903 + ], + "angle": 0, + "content": "GOAT. We assess our method within the GOAT-Bench setting, where the agent is spawned randomly and tasked with sequential targets. Each goal can be specified either by its category name, a description, or an image (e.g. the input could be \"gas boiler\", \"the gas boiler on the corner of the room. The gas boiler is located on the left of the washing machine and freezer\" or an image of the \"gas boiler\"). Table 3 compares our approach to the methods in [10]. TANGO significantly outperforms other state-of-the-art techniques (rows 1-4), achieving a \\(+2.6\\%\\) gain in the success metric. Additionally, it ranks second in navigation efficiency, with only a slight \\(-0.7\\%\\) difference from the top-performing method. We primarily attribute these results to the LLM's ability to accurately transform target descriptions into a sequence of selected goals, rather than relying" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24609" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.457, + 0.237 + ], + "angle": 0, + "content": "
MethodVAL UNSEEN
SR (↑)SPL (↑)
SenseAct-NN Skill Chain [26]29.511.3
SenseAct-NN Monolithic [26]12.36.8
Modular GOAT [26]24.917.2
Modular Clip on Wheels [26]12.310.4
TANGO (ours)32.116.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.248, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Table 3. GOAT-Bench. Performance comparison of different methods on the VAL UNSEEN split." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.292, + 0.487, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.45, + 0.483, + 0.52 + ], + "angle": 0, + "content": "Figure 5. Multi-Modal Lifelong Navigation Success Example. (top) RGB observation of the target during STOP action (step \\( t_{i+steps} \\)). (middle) Value map for the specific target recomputed from the memory map (step \\( t_{i+steps} \\)). (bottom) Memory map after target changes (step \\( t_i \\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.592 + ], + "angle": 0, + "content": "solely on potentially \"noisy\" text embeddings. Moreover, the use of a memory mechanism is essential for the task (Fig. 5); however, if the memorized target object is incorrect, the path efficiency can significantly decrease." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.317, + 0.62 + ], + "angle": 0, + "content": "4.3. TANGO Failure analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Due to the high explainability of our system, we conducted an analysis of the failures encountered during EQA episodes (Fig. 6). We extracted a significant subsample for the task and manually classified the instances where the model failed. We considered scores below 3 as failures, i.e. incorrect answers. We observed that the main cause of failure can be attributed to the \"detect\" module (stopped at wrong object or Ignored goal object in the image). Furthermore, the exploration policy appears to perform well given the targets produced by the LLM, as it fails in only \\(\\sim 10\\%\\) of the cases (labeled as: didn't see target goal in the image). The LLM generates incorrect pseudo-code around \\(18\\%\\) of the time, with \\(6.9\\%\\) failures leading to ambiguous or incorrect targets, and \\(11.2\\%\\) due to incorrect primitive ordering or usage which instead leads directly to episode failure. Notably, only \\(11.2\\%\\) of errors are actually due to the LLM generating incorrect code, while the remaining is attributable to prompt-related issues. A comprehensive fail-" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.092, + 0.895, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.286, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 6. Causes of system failure in EQA tasks. The analysis was manually conducted on a significant sample of failed episodes, facilitated by the explainability of TANGO. Answers were deemed as \"incorrect\" if they received a score below three." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.376, + 0.875, + 0.391 + ], + "angle": 0, + "content": "ure analysis is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.414, + 0.633, + 0.429 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.743 + ], + "angle": 0, + "content": "We conducted a systematic analysis of our compositional approach, offering valuable insights into the future of modular, neuro-symbolic systems. Moreover, TANGO highlights the versatility of simple PointGoal Navigation agents equipped with specific task-oriented modules, yielding promising results in zero-shot scenarios across all considered tasks. TANGO effectively handles diverse multi-modal prompts, following instructions to complete tasks, underscoring the potential of LLMs in robotic navigation. Its integrated memory supports lifelong navigation, suggesting improved capabilities with increased exploration. However, it may struggle with overly complex or ambiguous prompts, which can limit the LLM's ability to identify the correct primitives or targets. Future work could extend this system to tasks such as Visual Language Navigation (VLN) and explore the use of open-source LLMs. Additionally, refining the memory mechanism—particularly through improved sampling strategies for identifying high-value pixels linked to objects' memory—could further enhance the system's effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.767, + 0.673, + 0.784 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We thank the University of Padua, Department of Mathematics \"Tullio Levi-Civita\", for providing the computational resources. TC and LS were supported by the PNRR project Future AI Research (FAIR - PE00000013) under the NRRP MUR program, funded by NextGenerationEU. We also thank Elena Izzo and Luca Parolari for their valuable comments on early drafts of this paper." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24610" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Dhruv Batra, Aaron Gokaslan, Aniruddha Kembhavi, Oleksandr Maksymets, Roozbeh Mottaghi, Manolis Savva, Alexander Toshev, and Erik Wijmans. ObjectNav revisited: On evaluation of embodied agents navigating to objects. arXiv preprint arXiv:2006.13171, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Zhenshan Bing, Alexander Koch, Xiangtong Yao, Kai Huang, and Alois Knoll. Meta-reinforcement learning via language instructions. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5985-5991. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Proc. of Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[6] Tommaso Campari, Paolo Eccher, Luciano Serafini, and Lamberto Ballan. Exploiting scene-specific features for object goal navigation. In Proc. of the European Conference on Computer Vision Workshops (ECCVW), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[7] Tommaso Campari, Leonardo Lamanna, Paolo Traverso, Luciano Serafini, and Lamberto Ballan. Online learning of reusable abstract models for object goal navigation. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[9] Angel X. Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. In Proc. of the International Conference on 3D Vision (3DV), 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[10] Matthew Chang, Theophile Gervet, Mukul Khanna, Sriram Yenamandra, Dhruv Shah, So Yeon Min, Kavit Shah, Chris Paxton, Saurabh Gupta, Dhruv Batra, et al. Goat: Go to any thing. arXiv preprint arXiv:2311.06430, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to ex" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.906, + 0.12 + ], + "angle": 0, + "content": "plore using active neural SLAM. In Proc. of the International Conference on Learning Representations (ICLR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.906, + 0.189 + ], + "angle": 0, + "content": "[12] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. In Proc. of Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[13] Devendra Singh Chaplot, Ruslan Salakhutdinov, Abhinav Gupta, and Saurabh Gupta. Neural topological slam for visual navigation. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[14] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Niessner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[15] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[16] Vishnu Sashank Dorbala, James F Mullen Jr, and Dinesh Manocha. Can an embodied agent find your \"cat-shaped mug\"? llm-based zero-shot object navigation. IEEE Robotics and Automation Letters, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[17] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.471, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[18] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[19] Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[20] Ronghang Hu, Jacob Andreas, Trevor Darrell, and Kate Saenko. Explainable neural computation via stack neural module networks. In Proc. of the European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.652, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[21] C. et al. Huang. Visual language maps for robot navigation. In 2023 IEEE Int. Conf. Robot. Autom. (ICRA), pages 10608-10615, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[22] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[23] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7B. arXiv preprint arXiv:2310.06825, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24611" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[25] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C. Lawrence Zitnick, and Ross B. Girshick. Inferring and executing programs for visual reasoning. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[26] Mukul Khanna, Ram Ramrakhya, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. Goat-bench: A benchmark for multi-modal lifelong navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16373-16383, 2024. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[27] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[28] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances. arXiv preprint arXiv:2211.15876, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[29] Jacob Krantz, Theophile Gervet, Karmesh Yadav, Austin Wang, Chris Paxton, Roozbeh Mottaghi, Dhruv Batra, Jitendra Malik, Stefan Lee, and Devendra Singh Chaplot. Navigating to objects specified by images. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[30] Dingbang Li, Wenzhou Chen, and Xin Lin. Tina: Think, interaction, and action framework for zero-shot vision language navigation. arXiv preprint arXiv:2403.08833, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[32] J. et al. Liang. Code as policies: Language model programs for embodied control. In 2023 IEEE Int. Conf. Robot. Autom. (ICRA), pages 9493-9500, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[34] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. Openeqa: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.188 + ], + "angle": 0, + "content": "Alexander William Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondrus, Theophile Gervet, Vincent-Pierre Berges, John M. Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars and robots, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[38] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X. Chang, et al. Habitat-matterport 3D dataset (HM3D): 1000 large-scale 3D environments for embodied AI. arXiv preprint arXiv:2109.08238, 2021. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[39] Santhosh K. Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[40] Sonia Raychaudhuri, Tommaso Campari, Unnat Jain, Manolis Savva, and Angel X Chang. Mopa: Modular object navigation with pointgoal agents. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5763-5773, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[41] G. et al. Sarch. Open-ended instructable embodied agents with memory-augmented llms. arXiv:2310.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[43] Manolis Savva, Angel X. Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. Minos: Multimodal indoor simulator for navigation in complex environments. arXiv preprint arXiv:1712.03931, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[44] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied AI research. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[45] Bokui Shen, Fei Xia, Chengshu Li, Roberto Martín-Martín, Linxi Fan, Guanzhi Wang, Claudia Pérez-D'Arpino, Shyamal Buch, Sanjana Srivastava, Lyne Tchapmi, et al. iGibson 1.0: a simulation environment for interactive tasks in large realistic scenes. In Proc. of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[46] C.H. et al. Song. Llm-planner: Few-shot grounded planning for embodied agents. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pages 2998-3009, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Dídac Surís, Sachit Menon, and Carl Vondrick. ViperGPT: Visual Inference via Python Execution for Reasoning. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24612" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[48] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[50] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. DD-PPO: Learning near-perfect pointgoal navigators from 2.5 billion frames. In Proc. of the International Conference on Learning Representations (ICLR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[51] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[52] Zhenyu Wu, Ziwei Wang, Xiuwei Xu, Jiwen Lu, and Haibin Yan. Embodied task planning with large language models. arXiv preprint arXiv:2307.01848, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[53] Joel Ye, Dhruv Batra, Erik Wijmans, and Abhishek Das. Auxiliary Tasks Speed Up Learning PointGoal Navigation. In Proc. of the International Conference on Robot Learning (CoRL), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[54] Joel Ye, Dhruv Batra, Abhishek Das, and Erik Wijmans. Auxiliary Tasks and Exploration Enable ObjectNav. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[55] Naoki Yokoyama, Sehoon Ha, Dhruv Batra, Jiuguang Wang, and Bernadette Bucher. Vlfm: Vision-language frontier maps for zero-shot semantic navigation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 42-48. IEEE, 2024. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[56] Naoki Yokoyama, Ram Ramrakhya, Abhishek Das, Dhruv Batra, and Sehoon Ha. Hm3d-ovon: A dataset and benchmark for open-vocabulary object goal navigation. arXiv preprint arXiv:2409.14296, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[57] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target navigation. in 2023 IEEE. In RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3554-3560. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.742 + ], + "angle": 0, + "content": "[58] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.743, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[59] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24613" + } + ] +] \ No newline at end of file diff --git a/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_origin.pdf b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e6e1a5786e602c55cde7f2fcbafb7a5ff9de9d3 --- /dev/null +++ b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/0a8d6e42-42fa-4911-84af-bd3526112049_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10ed65aa92a200a952600f07c70abd90d0342cde87b72a7bf6ed34c3e4c31cff +size 5785844 diff --git a/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/full.md b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f6f0f62049cacd0f7fc979398f025fb32decaebf --- /dev/null +++ b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/full.md @@ -0,0 +1,256 @@ +# TANGO: Training-free Embodied AI Agents for Open-world Tasks + +Filippo Ziliotto $^{1,2*}$ + +Tommaso Campari* + +Luciano Serafini² + +Lamberto Ballan1 + +1 University of Padova, Italy + +Fondazione Bruno Kessler (FBK), Trento, Italy + +# Abstract + +Large Language Models (LLMs) have demonstrated excellent capabilities in composing various modules together to create programs that can perform complex reasoning tasks on images. In this paper, we propose TANGO, an approach that extends the program composition via LLMs already observed for images, aiming to integrate those capabilities into embodied agents capable of observing and acting in the world. Specifically, by employing a simple PointGoal Navigation model combined with a memory-based exploration policy as a foundational primitive for guiding an agent through the world, we show how a single model can address diverse tasks without additional training. We task an LLM with composing the provided primitives to solve a specific task, using only a few in-context examples in the prompt. We evaluate our approach on three key Embodied AI tasks: Open-Set ObjectGoal Navigation, MultiModal Lifelong Navigation, and Open Embodied Question Answering, achieving state-of-the-art results without any specific fine-tuning in challenging zero-shot scenarios. + +# 1. Introduction + +Large Language Models (LLMs) have gained significant attention in the field of AI due to their remarkable capability to generalize across unseen tasks [5, 24, 48, 49]. Systems like VisProg and ViperGPT [18, 47] demonstrated strong performance on a broad range of vision-and-language tasks by just relying on a few numbers of in-context examples to generate complex compositional programs without requiring any specific training. Nevertheless, the evaluation of these concepts to Embodied AI, particularly for navigation tasks, remains largely unexplored. This approach may be crucial in building agents capable to navigate and operate efficiently in unfamiliar environments. + +Previously, methods such as Neural Module Networks (NMN) [2, 19] have demonstrated promising compositional + +![](images/8f7b7e81ceb8828b89be8dd0cff489b6023a2caad0fdf20e17f343f77f9b4ce0.jpg) +Figure 1. We introduce TANGO, a modular neuro-symbolic system for compositional embodied visual navigation. Given a few examples of natural language instructions and the corresponding programs composed of action primitives, TANGO can generate executable programs, enabling the agent to perform multiple tasks within a 3D environment. + +properties for high-level tasks like visual question answering, via end-to-end training of networks and specialized, differentiable neural modules. However, they required the combination of semantic parsers and predefined templates to learn to solve the compositional tasks. Alternative approaches leverage LLMs through predefined task modules or APIs for action execution [22, 23, 32, 52, 58]. While effective, these methods often lack a compositional structure to address multiple tasks or primarily focus on manipulation and robotic planning, with limited emphasis on navigation. + +In this work, we present TANGO (see Figure 1) a novel neuro-symbolic compositional approach which uti- + +izes primitives and employs a LLM as a planner to sequence these primitives within photorealistic 3D environments, where agents must perceive and act. This framework integrates high-level planning with low-level action execution, without the need for training. TANGO makes use of diverse modules designed for visual navigation and question-answering tasks, resulting in a system that seamlessly addresses both challenges while achieving SoA performance without requiring any prior adjustments. By providing a few in-context examples that show how to tackle multiple tasks, TANGO is capable of generalizing to the specific task at hand. This is facilitated by the LLM, which effectively combines the individual modules available within the TANGO system. Moreover, it extends [55] exploration policy by incorporating a memory mechanism in the form of a stored feature map that retains information about previously explored areas, supporting for efficient life-long navigation tasks. + +To demonstrate the flexibility of the proposed framework, TANGO has been tested on three popular benchmarks: namely, i) Open-Vocabulary ObjectGoal Navigation [56], ii) Multi-Modal Lifelong Navigation [26], and iii) Embodied Question Answering [15, 34]. Results match or surpass previous approaches, without requiring specialized training. Summing up, the contributions of this paper are as follows: + +- Introduction of a neuro-symbolic compositional LLM-based framework for EAI leveraging specialized primitive modules. +- Demonstration of robust generalization capabilities across multiple tasks without the need for any specific training or fine-tuning. +- Extension of the exploration policy presented in [55] to multi-goal scenarios through the incorporation of a memory mechanism stored as a feature vector map. +- Achievement of state-of-the-art results, underscoring the effectiveness of the proposed framework. + +# 2. Related Works + +Our work takes inspiration from modular approaches for visual tasks, like the seminal Neural Module Networks [2, 19], as well as the recent VisProg [18] and ViperGPT [47] frameworks. The main contribution of this paper is to extend these ideas to embodied visual navigation, that usually require heavy end-to-end learning, although modular approaches have recently shown promising results [7, 17]. Therefore, we discuss prior work in the area of embodied AI, the use of language models in robotics, and program generation approaches for image recognition tasks. + +Embodied AI and Visual Navigation. The field of Embodied AI has recently undergone a paradigm shift, fuelled + +by the emergence of highly efficient simulators [27, 36, 43-45]. These simulators enable processing numerous parallel simulations in photorealistic indoor environments, facilitating large-scale testing that was otherwise challenging in classical robotics. Alongside these simulators, a large variety of tasks (and benchmarks) has been proposed: PointGoal Navigation [50], where an agent navigates from point A to point B in an unknown environment; ObjectGoal Navigation [3], requiring the agent to locate and navigate to an object in the scene; Instance-Image Goal Navigation [28], akin to ObjectGoal Navigation, but in which the agent should find a specific object depicted in a given image; Multi-Modal Lifelong Navigation [26] involving navigating to a sequence of target objects that can be specified through labels, images, or textual descriptions; Embodied Question Answering [15, 34], in which an agent navigates an environment to gather information needed to answer a question, and Vision and Language Navigation (VLN) [1]. + +Various approaches have been proposed to tackle these tasks, primarily falling into three categories: end-to-end [6, 53, 54], modular [11-13, 29, 39, 40] and LLM-based [16, 30, 57-59]. However, the main limitation of these paradigms is their dependence on task-specific architectures. In the case of end-to-end approaches, a model is trained for a specific problem, often requiring days of training, such as in [50], where a policy for PointGoal Navigation was trained for 2.5 billion steps. To adapt the learned model to a new problem, adjustments and retraining are necessary, and this comes with a high cost. Modular approaches share a similar challenge, usually relying on policies designed for specific tasks, although common elements, such as exploration and navigation modules, are often shared [11, 12, 29, 39, 40]. Adapting these solutions to different tasks requires a manual adjustment of the approach to fit the new domain, often involving the addition or modification of modules. + +Language models for Robotics In robotics, foundation models (LLMs) trained on vast internet-scale datasets [5, 24, 48, 49] have the potential to equip robots with real-world priors and advanced reasoning abilities without the need for extensive task-specific training. Early approaches equipped agents with learned language embeddings, requiring large amounts of training data [4, 32]. Recent studies, on the other hand, have explored zero-shot and few-shot solutions mainly focusing on robotic planning and manipulation tasks [22, 23, 32, 41, 46]. In the context of visual navigation, LLM-based approaches leverage the powerful LLMs priors and reasoning capabilities to guide navigation within the environment [16, 21, 30, 57-59]; however, despite their remarkable contributions, they often lack a modular design that allows for flexible integration and extensibility, as they are typically optimized for single tasks rather + +than a comprehensive, adaptable framework. + +In contrast, our work deviates from these paradigms by relying on a series of diverse pre-trained modules that, thanks to a LLM, can be combined to potentially solve various tasks, compositionally. None of our modules are fine-tuned explicitly for the target problems, placing us in a zero-shot setting. Moreover, TANGO does not only rely on an exploration policy based entirely on LLM output, as in [16, 58, 59]. Instead, it combines a frontier exploration policy with LLM priors and memory guidance to enhance navigation performance. To achieve this, we use a pre-trained PointGoal policy as the base waypoint navigation module of TANGO and extend the exploration policy introduced in [55] by equipping the agent with a memory mechanism stored as a feature vector for each pixel of the map. + +Modular vision and program composition for visual tasks. Neural Module Networks (NMN) [2, 19] introduced modular and compositional methodologies for visual question answering (VQA). NMNs integrate neural modules into an end-to-end differentiable network; the approach originally relied on pre-existing parsers [2], whereas more recent methods [19, 20, 25] have evolved to learn the layout generation model concurrently with the neural modules, employing reinforcement learning [51] and weak supervision. Stack-NMN [20] extends N2NMN by transitioning from discrete to soft layout generation, incorporating a weighted average of predictions from all modules at each step, determined by a layout generation network. + +Recently, [18] introduced VisProg, a framework that offers two key advantages over NMNs. Firstly, it constructs high-level programs that invoke state-of-the-art neural models and Python functions at intermediate steps, diverging from the conventional approach of generating end-to-end neural networks. This design facilitates the integration of symbolic, non-differentiable modules. Secondly, it leverages the in-context learning ability of LLMs [5]. This is achieved by prompting the LLM with a natural language instruction (or a visual question or a statement to be verified), along with a few examples of similar instructions and their corresponding programs. Almost concurrently, ViperGPT [47] presented a similar approach to VisProg, but in this case the model directly generates Python code, instead of composing pre-defined modules and routines. In contrast, in our framework, the LLM is provided only with high-level knowledge of the functions of each primitive, which is encoded directly in the primitive's name. Therefore, the system it is built similarly to VisProg. It is noteworthy that both VisProg and ViperGPT were tested on images, while our model is specifically designed for embodied agents operating in an interactive environment, relying on sensor-based perception. + +# 3. Method + +Neuro-symbolic approaches offer the possibility to address a broad range of diverse complex tasks efficiently. Systems like VisProg [18] operate on images and have demonstrated excellent results in zero-shot settings. Understanding how a similar approach works in an action-oriented scenario is key to enable robots to navigate autonomously in novel environments. Therefore, we develop this paradigm further into EAI, where the integration with action execution adds several challenges. Inspired by [18], given an input prompt, we rely on a LLM to generate synthetic pseudo-programs that can be executed by the embodied agent in the environment. Our TANGO framework is therefore able to generalize to new tasks, without any direct training on task-specific datasets, thus effectively mitigating the heavy computational training costs inherent to embodied navigation tasks. The procedure that enables generating these new programs is based on "in-context examples" that are fed to the LLM, alongside a natural language instruction. An outline of TANGO is shown in Figure 2. + +TANGO Interpreter. The first key component of the proposed framework is referred to as "Program Interpreter". It comprises visual recognition modules that can be used by the agent to extract the semantics of the scene, as well as to provide an understanding of the visual context. + +As shown in Figure 1, users can ask questions or give specific tasks to the agent; the natural language prompt is then processed by the LLM (in our implementation GPT-4o [5]), which serves as a planner and outputs a step-by-step executable program. To ensure the LLM delivers a reasonable output, it is fed with 15 "in-context examples" across diverse tasks. This enables the LLM to make use of its reasoning capabilities effectively, identifying the most suitable planning for the required task. Moreover, the examples remain the same regardless of the task identified in the prompt; it is the LLM's responsibility to output the specific program target for the given question. Programs use a higher level of abstraction than previous modular attempts such as Neural Module Networks (NMN) [2, 19]. Each program is constructed as a sequence of primitives (e.g., detect, answer, match, etc.) that invoke corresponding TANGO modules. These modules are either powered by pre-trained state-of-the-art vision models or implemented as simple Python subroutines (e.g., count, is_found, eval, etc.), with additional navigation-specific modules designed to "steer" the agent's movements (e.g. navigate_to, return, turn, etc.). Figure 3 provides a comprehensive list of all the currently implemented modules. + +Figure 2 shows an Embodied Question Answering (EQA) example for the user's question: "on the left of the toilet there's a shower, is the shower curtain open or closed?" The LLM, which acts as a planner, transforms the + +![](images/a79c71eb605a0bc6cae2e5569d8ee535c0a077ecb5e8eded071b2392ae2c8f12.jpg) +Figure 2. Overview of the program generation in TANGO. Given few "in-context examples", the LLM provide a detailed sequence of steps to be executed by the agent in the given environment. The LLM is instructed to comment its output to allow for explainability. + +user's initial prompt into navigation subtasks by leveraging its reasoning abilities and powerful priors. This decomposition enables the model to break down complex queries into more manageable steps, guiding the agent's navigation process effectively. Each line of the generated program corresponds to a module serving a specific task. The agent must then execute the generated program. The LLM is also guided to comment on its steps directly within the generated pseudocode. + +All the given primitives are equipped with methods to: i) parse lines in order to extract input argument names and values, as well as the output variable name; ii) execute the module in the environment, and update the program state with the output variable name and value. The outputs for each step and the comments in the generated pseudocode can also be used to better understand what is happening under the hood and why the LLM generated a specific line, thus enabling a good interpretability of the system behavior. + +TANGO modules utilize open-source software and models, readily downloadable from the web. They can also be effortlessly updated with newer, more efficient models as they become available. + +Navigation Module. In order to navigate the environment, we define a module employing a PointGoal navigation agent as our foundational module [40, 55]. This model achieved nearly perfect PointGoal performance, both in terms of success ( $\sim 99\%$ ) and efficiency (a forward pass that takes a fraction of a second) on standard datasets [9, 38]. The agent starts the exploration of the environment until it locates its target goal. Once the target is identified, the focus of exploration transitions to reaching the designated goal. This PointNav policy exclusively uses the egocentric depth image and the robot's relative distance and heading towards the desired goal point as input. + +Exploration Policy. Exploration is performed using the policy outlined in [55]. This method builds occupancy maps based on depth observations to identify exploration frontiers. It further leverages RGB observations and a pretrained vision-language model (BLIP2 [31]) to generate a language-grounded "value map". This value map is then used to guide the exploration, facilitating an efficient search for instances of a given target. Notably, as in [55], the agent performs a $360^{\circ}$ turn at the start of navigation to initialize frontiers. + +We extend this policy for sequential goals by incorpora + +![](images/14ba6ce6596cd3c232b08380c74a7d43e3b9e114dab00fe5dddeee2a6d0dd781.jpg) +Figure 3. Overview of TANGO modules. Modules span a variety of inputs and outputs. Orange modules use Python subroutines, while blue modules use pre-trained computer vision models (similarly to [18]). The navigate_to and explore(scene modules, in green, both implement our foundational PointNav module; however, only explorescene integrates the memory mechanism. + +ing a memory mechanism, represented as a “feature map” in which each pixel of the value map is encoded as a vector and updated at each step. This feature map then updates the language-grounded original value map when a new target is specified, by calculating the cosine similarity between each pixel's vector and the new target embeddings (either from text or image). We sample the highest value in the updated value map; if this value exceeds a predefined threshold, indicating the agent has previously encountered and “rememberers” the target's location, the agent navigates to it, supporting lifelong navigation (see Section 2-Supp). + +# 3.1. Navigation Tasks with TANGO + +TANGO leverages pre-trained multimodal vision-language and vision-only models as foundational components to extract semantic information from the scene, making it well-suited for several Embodied AI tasks. + +(Open-set) ObjectGoal Navigation [3, 56], in particular, emerges as a suitable testbed to evaluate the efficacy of our method. A key module to tackle this problem involves utilizing an object detector (Owlv2 in our implementation [35] or DETR [8] for target objects that fall within the COCO classes [33]) to identify objects within the image. Subsequently, navigation towards the detected objects is facilitated by waypoints, leveraging depth distance calculation to guide the agent effectively. For simplicity, we use the center of the bounding box to determine the target waypoint. + +TANGO also integrates a specialized module for matching target objectives with ongoing visual observations in image-based scenarios, resulting in accurate navigation towards an instance image. To evaluate our agent's performance in this context, we use real-time matching SuperGlue network [42], optimized for indoor environments (i.e. using hyperparameters recommended in the respective paper). + +TANGO being task-agnostic, can process any type of navigation target, regardless of the order in which it is presented. In this context, the GOAT benchmark [26] is well-suited for evaluating our system, as it operates in scenarios where targets are specified through images, text, or descriptive phrases, provided in random sequence. It also heavily + +relies on memory to find previously seen targets when sequential goals are given. + +An agent should also be able to answer user queries like "can you check if the kitchen table is clean?" Hence, EQA serves as an excellent benchmark to assess TANGO's capabilities and robustness. EQA consists of three phases: understanding the semantic structure of the prompt, locating the target object(s), and analyzing the visual semantics to generate an accurate response. To this end, TANGO integrates a specific answer module, relying on BLIP2 [31] as its core foundation. + +# 4. Experiments + +TANGO provides a flexible framework that can be applied to various embodied navigation problems. We evaluate our approach on three popular tasks that require a wide range of capabilities, including efficient environment exploration, path planning, scene and context understanding, and image similarity comparison. + +# 4.1. Experimental Setup + +Agent Configuration. Prior research on visual navigation commonly uses different agent configurations depending on the considered task [15, 26, 56]. The configuration typically employed for (OPEN-SET) OBJNAV closely resembles that of a LoCoBot, with an agent height of $0.88m$ , a radius of $0.18m$ , and a single $640 \times 480$ RGB sensor with a $79^{\circ}$ hfov, positioned $0.88m$ above the ground. + +In the context of GOAT Benchmark, some of the default settings differ, featuring a camera situated $1.31m$ above the ground. The agent has a height of $1.41m$ and is given $360 \times 640$ RGB images. Both the above tasks' configurations use a step size of $0.25m$ and a left and right turning angle of $30^{\circ}$ . Lastly, the settings for EQA mirror those of PointGoal Navigation as specified in the Habitat-Lab configuration files. All other settings remain consistent with those of [15]. + +Datasets and Evaluation Protocol. We assess the TANGO performance across task-specific datasets using the Habitat simulator [44]. As the system is not trained in any way (except for the PointGoal Navigation model, which has been pre-trained on the training set of HM3D, never used in our tests), each scene within episodes is novel to the agent. Therefore, the entire validation set can be categorized as "unseen". We now describe in detail the tasks and datasets used: + +- OPEN VOCABULARY OBJECT NAVIGATION (OVON) [56]: a large-scale benchmark featuring over 15,000 annotated household objects across 379 categories, derived from real-world 3D scans [38]. The agent is initialized at a random location within a scene and + +tasked with navigating to a target object category within a time limit of 500 steps. We evaluate our approach across all episodes included in the "val unseen" set. + +- MULTI-MODAL LIFELONG NAVIGATION (GOAT-BENCH) [26]: an agent is tasked with sequentially navigating to five to ten target objects identified by category names, descriptions, or images. Each target represents a subtask executed within an open-vocabulary framework that spans over 312 categories. The agent is required to reach a goal within a specified time constraint and is assigned new targets upon the completion of each subtask. We evaluate our approach across all episodes included in the "val unseen" set. +- EMBodied Question ANSWERING (OPENEQA) [34]: containing over 1,600 question-answer pairs sourced from more than 180 real-world environments and scans [14, 38]. It is divided into two task categories: Episodic Memory-EQA (EM-EQA) and Active-EQA (A-EQA). Our focus is on the latter, in which the agent must autonomously navigate and answer questions within a time constraint of 500 steps. Questions span over seven categories, including world knowledge, attribute recognition, spatial reasoning, and object localization. Examples of questions from the episodes include: "Is the microwave door propped open?" and "What is left of the kitchen pass-through?". We evaluate our approach across all episodes belonging to the A-EQA task category. + +Evaluation Metrics. In all tasks, we use the standard metrics as in prior works [3, 15, 34, 56]. Namely: + +- Success Rate (SR): measures the ratio of episodes where the agent successfully reached its target (Open ObjNav, GOAT-Bench). +- Success weighted by Path Length (SPL): measures the optimality of the path taken by the agent w.r.t. the optimal path (Open ObjNav, GOAT-Bench). +- Distance to Goal (DTG): measures the average distance to goal of the agent at the end of the episode (Open ObjNav, GOAT-Bench). +- Score (LLM-Match): an LLM compares the ground-truth $GT_{i}$ answers with model output $A_{i}$ given a question $Q_{i}$ and assigns a score $\sigma_{i}$ on a scale of 1 to 5. On this scale, 1 indicates an incorrect response, 5 represents a correct response, and intermediate values reflect varying levels of similarity. For example, the answer to the question "What color is the bed" could be correctly answered either as "white" or "the bed is white". The final results aggregation is as follows: + +$$ +S = \frac {1}{N} \sum_ {i} ^ {N} \frac {\sigma_ {i} - 1}{4} \times 100 \% \tag{1} +$$ + +- Answer Accuracy: the average accuracy of the answers + +provided by the agent for the EQA task. This metric, used in [15], is presented with the related results in the supplementary material. + +Implemented Modules. TANGO contains descriptive and easily understandable names for the modules, arguments, and variables to facilitate the LLM comprehension of the input and output. Figure 3 showcases a list of all available modules. Each module exclusively outputs predefined object variables to the next one, enabling the possibility to monitor the progression of the agent at each step. Therefore, dramatically enhancing the explainability of failure cases, as it allows for a deep examination of the agent's decision-making process through its interactions with the environment. Figure 6 provides a general overview of the failure analysis in EQA task. As presented in Section 3, "navigate_to" and "explore_structure" modules are specifically engineered to navigate the environment, leveraging inputs from depth and RGB sensors as well as the PointGoal GPS + compass sensors. If this sensor is not available for a specific task, it can be derived from the current agent pose and the exploration goal location. The agent is provided with the actions: Forward, Move Left, Move Right, Look Up, Look Down and Stop. + +For the "detect" module, we employ Owlv2 [35] object detector for general classes and use DETR [8] for categories within the COCO classes [33]. Notably, in synthetic 3D scenes, we noticed numerous false positives. To address the problem, bounding boxes produced by the detector are forwarded to a "classify" module. Within this module, each detection undergoes classification to differentiate between categories of similar instances. For example, the class "chair" encompasses diverse subcategories such as "armchair", "couch", and "other". Subclasses are outputted under the hood directly by the LLM requiring no human annotation. We utilize a CLIP-based classifier in its default configuration [37]. + +Embodied Question Answering (EQA) task uses an "answer" module based on BLIP2 [31], capable of performing various multi-modal tasks, including Visual Question Answering, Image-Text retrieval, and Image Captioning. Notably, the module is also used when goals are specified through images, to capture the class of the input image before starting the navigation. Furthermore, for image-based tasks we exploit a "match" module to evaluate the similarity between the agent's observation and the instance image target. This module performs feature matching between the two, as described in Section 3.1, to aid in navigation. Hence, it assists the agent in determining whether what it perceives as the current target is indeed the correct goal. + +![](images/1dd0c7143bf53da833fd5a99cffd8672f2a01db8f06fed4fc24d4b29efbfbae9.jpg) +Figure 4. Examples from OpenEQA [34]. The top section illustrates a successful episode where TANGO is able to understand the input query, correctly specifying the sequential targets. The lower section illustrates a failure caused by overly general directions from the LLM, which TANGO struggled to resolve. + +
MethodVAL UNSEEN
SR (↑)SPL (↑)
RL [56]18.6 ±0.37.5 ±0.2
BCRL [56]8.0 ±0.22.8 ±0.1
DAgRL [56]18.3 ±0.37.9 ±0.1
*VLFM [55, 56]35.219.6
DAgRL+OD [56]37.1 ±0.219.9 ±0.3
TANGO (ours)35.5 ±0.319.5 ±0.3
+ +# 4.2. Experimental Results + +Open-set ObjectGoal Navigation We evaluated our method in the standard Open-Set OBJNAV setting, with targets sampled directly from objects in the scene. As shown in Table 1, our approach achieves performance on par with state-of-the-art methods (rows 4-5) on the validation-unseen split both in SR (35.5%) and SPL (19.5%). Instead of focusing solely on numerical results, these findings highlight TANGO's flexibility in solving navigation tasks in unfamiliar environments, underscoring the potential of our approach within the OBJNAV framework. + +Table 1. HM3D-OVON. Performance comparison of different methods on VAL UNSEEN split. * deterministic method. + +
MethodScore (↑)
Human Agent [34]85.1 ± 1.1
Blind LLMs [34]35.5 ± 1.7
Socratic LLMs w/ Frame Captions [34]38.1 ± 1.8
Socratic LLMs w/ Scene-Graph Captions [34]34.4 ± 1.8
TANGO (ours)37.2 ± 1.8
+ +Table 2. OpenEQA results. Comparison of TANGO with the SoA on the OpenEQA dataset. All methods involve zero-shot approaches. + +Embodied Question Answering. In EQA, the agent is queried with a natural language question and must autonomously explore the environment and gather information in order to answer accordingly. In recent literature, this task shifted from being regarded as a purely classification problem aimed at determining the most suitable answer from a set of pre-defined possibilities (i.e., class labels) to an open-vocabulary benchmark where the agent must answer with natural language [15, 34]. We compare TANGO results, using the score from Eq. 1, against other zero-shot approaches. Table 2 shows that our method ranks second and closely matches the performance of the leading SoA method, with a gap of less than $1\%$ . Notably, these approaches [34] utilize LLMs to interpret scene objects step-by-step (rows 3-4), but they do not leverage LLMs to effectively guide navigation towards specific targets. Hence, our approach demonstrates that SoA results can still be achieved by pre-planning a path that prioritizes relevant areas for a given question. Moreover, the results indicate that human agents achieve a significantly higher score of $85\%$ , which remains well above the performance achieved by using large language models (LLMs), particularly in open-world settings. This gap underscores the challenges LLMs face in handling the complexity and variability of open-world scenarios compared to human agents. + +GOAT. We assess our method within the GOAT-Bench setting, where the agent is spawned randomly and tasked with sequential targets. Each goal can be specified either by its category name, a description, or an image (e.g. the input could be "gas boiler", "the gas boiler on the corner of the room. The gas boiler is located on the left of the washing machine and freezer" or an image of the "gas boiler"). Table 3 compares our approach to the methods in [10]. TANGO significantly outperforms other state-of-the-art techniques (rows 1-4), achieving a $+2.6\%$ gain in the success metric. Additionally, it ranks second in navigation efficiency, with only a slight $-0.7\%$ difference from the top-performing method. We primarily attribute these results to the LLM's ability to accurately transform target descriptions into a sequence of selected goals, rather than relying + +
MethodVAL UNSEEN
SR (↑)SPL (↑)
SenseAct-NN Skill Chain [26]29.511.3
SenseAct-NN Monolithic [26]12.36.8
Modular GOAT [26]24.917.2
Modular Clip on Wheels [26]12.310.4
TANGO (ours)32.116.5
+ +Table 3. GOAT-Bench. Performance comparison of different methods on the VAL UNSEEN split. + +![](images/33565f800cdb158f4546f0bafdf71bb56ab35c441bdbeae367a5ef776003ccef.jpg) +Figure 5. Multi-Modal Lifelong Navigation Success Example. (top) RGB observation of the target during STOP action (step $t_{i+steps}$ ). (middle) Value map for the specific target recomputed from the memory map (step $t_{i+steps}$ ). (bottom) Memory map after target changes (step $t_i$ ). + +solely on potentially "noisy" text embeddings. Moreover, the use of a memory mechanism is essential for the task (Fig. 5); however, if the memorized target object is incorrect, the path efficiency can significantly decrease. + +# 4.3. TANGO Failure analysis + +Due to the high explainability of our system, we conducted an analysis of the failures encountered during EQA episodes (Fig. 6). We extracted a significant subsample for the task and manually classified the instances where the model failed. We considered scores below 3 as failures, i.e. incorrect answers. We observed that the main cause of failure can be attributed to the "detect" module (stopped at wrong object or Ignored goal object in the image). Furthermore, the exploration policy appears to perform well given the targets produced by the LLM, as it fails in only $\sim 10\%$ of the cases (labeled as: didn't see target goal in the image). The LLM generates incorrect pseudo-code around $18\%$ of the time, with $6.9\%$ failures leading to ambiguous or incorrect targets, and $11.2\%$ due to incorrect primitive ordering or usage which instead leads directly to episode failure. Notably, only $11.2\%$ of errors are actually due to the LLM generating incorrect code, while the remaining is attributable to prompt-related issues. A comprehensive fail- + +![](images/669b7466ee6c54f43c6f6686209f35f2169febe03ef95c60344ba09a232f1c75.jpg) +Figure 6. Causes of system failure in EQA tasks. The analysis was manually conducted on a significant sample of failed episodes, facilitated by the explainability of TANGO. Answers were deemed as "incorrect" if they received a score below three. + +ure analysis is provided in the supplementary material. + +# 5. Conclusion + +We conducted a systematic analysis of our compositional approach, offering valuable insights into the future of modular, neuro-symbolic systems. Moreover, TANGO highlights the versatility of simple PointGoal Navigation agents equipped with specific task-oriented modules, yielding promising results in zero-shot scenarios across all considered tasks. TANGO effectively handles diverse multi-modal prompts, following instructions to complete tasks, underscoring the potential of LLMs in robotic navigation. Its integrated memory supports lifelong navigation, suggesting improved capabilities with increased exploration. However, it may struggle with overly complex or ambiguous prompts, which can limit the LLM's ability to identify the correct primitives or targets. Future work could extend this system to tasks such as Visual Language Navigation (VLN) and explore the use of open-source LLMs. Additionally, refining the memory mechanism—particularly through improved sampling strategies for identifying high-value pixels linked to objects' memory—could further enhance the system's effectiveness. + +# Acknowledgments + +We thank the University of Padua, Department of Mathematics "Tullio Levi-Civita", for providing the computational resources. TC and LS were supported by the PNRR project Future AI Research (FAIR - PE00000013) under the NRRP MUR program, funded by NextGenerationEU. We also thank Elena Izzo and Luca Parolari for their valuable comments on early drafts of this paper. + +# References + +[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[2] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 3 +[3] Dhruv Batra, Aaron Gokaslan, Aniruddha Kembhavi, Oleksandr Maksymets, Roozbeh Mottaghi, Manolis Savva, Alexander Toshev, and Erik Wijmans. ObjectNav revisited: On evaluation of embodied agents navigating to objects. arXiv preprint arXiv:2006.13171, 2020. 2, 5, 6 +[4] Zhenshan Bing, Alexander Koch, Xiangtong Yao, Kai Huang, and Alois Knoll. Meta-reinforcement learning via language instructions. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5985-5991. IEEE, 2023. 2 +[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Proc. of Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 3 +[6] Tommaso Campari, Paolo Eccher, Luciano Serafini, and Lamberto Ballan. Exploiting scene-specific features for object goal navigation. In Proc. of the European Conference on Computer Vision Workshops (ECCVW), 2020. 2 +[7] Tommaso Campari, Leonardo Lamanna, Paolo Traverso, Luciano Serafini, and Lamberto Ballan. Online learning of reusable abstract models for object goal navigation. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5, 6 +[9] Angel X. Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. In Proc. of the International Conference on 3D Vision (3DV), 2017. 4 +[10] Matthew Chang, Theophile Gervet, Mukul Khanna, Sriram Yenamandra, Dhruv Shah, So Yeon Min, Kavit Shah, Chris Paxton, Saurabh Gupta, Dhruv Batra, et al. Goat: Go to any thing. arXiv preprint arXiv:2311.06430, 2023. 7 +[11] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to ex + +plore using active neural SLAM. In Proc. of the International Conference on Learning Representations (ICLR), 2019. 2 +[12] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. In Proc. of Advances in Neural Information Processing Systems (NeurIPS), 2020. 2 +[13] Devendra Singh Chaplot, Ruslan Salakhutdinov, Abhinav Gupta, and Saurabh Gupta. Neural topological slam for visual navigation. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[14] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Niessner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6 +[15] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 5, 6, 7 +[16] Vishnu Sashank Dorbala, James F Mullen Jr, and Dinesh Manocha. Can an embodied agent find your "cat-shaped mug"? llm-based zero-shot object navigation. IEEE Robotics and Automation Letters, 2023. 2, 3 +[17] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79), 2023. 2 +[18] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5 +[19] Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 1, 2, 3 +[20] Ronghang Hu, Jacob Andreas, Trevor Darrell, and Kate Saenko. Explainable neural computation via stack neural module networks. In Proc. of the European Conference on Computer Vision (ECCV), 2018. 3 +[21] C. et al. Huang. Visual language maps for robot navigation. In 2023 IEEE Int. Conf. Robot. Autom. (ICRA), pages 10608-10615, 2023. 2 +[22] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1, 2 +[23] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 1, 2 +[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7B. arXiv preprint arXiv:2310.06825, 2023. 1, 2 + +[25] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C. Lawrence Zitnick, and Ross B. Girshick. Inferring and executing programs for visual reasoning. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 3 +[26] Mukul Khanna, Ram Ramrakhya, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. Goat-bench: A benchmark for multi-modal lifelong navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16373-16383, 2024. 2, 5, 6, 8 +[27] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 2 +[28] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances. arXiv preprint arXiv:2211.15876, 2022. 2 +[29] Jacob Krantz, Theophile Gervet, Karmesh Yadav, Austin Wang, Chris Paxton, Roozbeh Mottaghi, Dhruv Batra, Jitendra Malik, Stefan Lee, and Devendra Singh Chaplot. Navigating to objects specified by images. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[30] Dingbang Li, Wenzhou Chen, and Xin Lin. Tina: Think, interaction, and action framework for zero-shot vision language navigation. arXiv preprint arXiv:2403.08833, 2024. 2 +[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 4, 5, 6 +[32] J. et al. Liang. Code as policies: Language model programs for embodied control. In 2023 IEEE Int. Conf. Robot. Autom. (ICRA), pages 9493-9500, 2023. 1, 2 +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6 +[34] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. Openeqa: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2, 6, 7 +[35] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36, 2024. 5, 6 +[36] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, + +Alexander William Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondrus, Theophile Gervet, Vincent-Pierre Berges, John M. Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars and robots, 2023. 2 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[38] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X. Chang, et al. Habitat-matterport 3D dataset (HM3D): 1000 large-scale 3D environments for embodied AI. arXiv preprint arXiv:2109.08238, 2021. 4, 5, 6 +[39] Santhosh K. Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[40] Sonia Raychaudhuri, Tommaso Campari, Unnat Jain, Manolis Savva, and Angel X Chang. Mopa: Modular object navigation with pointgoal agents. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5763-5773, 2024. 2, 4 +[41] G. et al. Sarch. Open-ended instructable embodied agents with memory-augmented llms. arXiv:2310.15127, 2023. 2 +[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks, 2020. 5 +[43] Manolis Savva, Angel X. Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. Minos: Multimodal indoor simulator for navigation in complex environments. arXiv preprint arXiv:1712.03931, 2017. 2 +[44] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied AI research. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5 +[45] Bokui Shen, Fei Xia, Chengshu Li, Roberto Martín-Martín, Linxi Fan, Guanzhi Wang, Claudia Pérez-D'Arpino, Shyamal Buch, Sanjana Srivastava, Lyne Tchapmi, et al. iGibson 1.0: a simulation environment for interactive tasks in large realistic scenes. In Proc. of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 2 +[46] C.H. et al. Song. Llm-planner: Few-shot grounded planning for embodied agents. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pages 2998-3009, 2023. 2 +[47] Dídac Surís, Sachit Menon, and Carl Vondrick. ViperGPT: Visual Inference via Python Execution for Reasoning. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3 + +[48] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2 +[49] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2 +[50] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. DD-PPO: Learning near-perfect pointgoal navigators from 2.5 billion frames. In Proc. of the International Conference on Learning Representations (ICLR), 2019. 2 +[51] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 3 +[52] Zhenyu Wu, Ziwei Wang, Xiuwei Xu, Jiwen Lu, and Haibin Yan. Embodied task planning with large language models. arXiv preprint arXiv:2307.01848, 2023. 1 +[53] Joel Ye, Dhruv Batra, Erik Wijmans, and Abhishek Das. Auxiliary Tasks Speed Up Learning PointGoal Navigation. In Proc. of the International Conference on Robot Learning (CoRL), 2020. 2 +[54] Joel Ye, Dhruv Batra, Abhishek Das, and Erik Wijmans. Auxiliary Tasks and Exploration Enable ObjectNav. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[55] Naoki Yokoyama, Sehoon Ha, Dhruv Batra, Jiuguang Wang, and Bernadette Bucher. Vlfm: Vision-language frontier maps for zero-shot semantic navigation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 42-48. IEEE, 2024. 2, 3, 4, 7 +[56] Naoki Yokoyama, Ram Ramrakhya, Abhishek Das, Dhruv Batra, and Sehoon Ha. Hm3d-ovon: A dataset and benchmark for open-vocabulary object goal navigation. arXiv preprint arXiv:2409.14296, 2024. 2, 5, 6, 7 +[57] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target navigation. in 2023 IEEE. In RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3554-3560. 2 +[58] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1, 3 +[59] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 2, 3 \ No newline at end of file diff --git a/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/images.zip b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2ffe5a84dabd95f229c0ecb465f5b815c06351f5 --- /dev/null +++ b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf19f5d01928ce85c294719449e14ad6a260114f62b895bd7a16d0e80f4bb45a +size 470722 diff --git a/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/layout.json b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c1ca62ac2f1e7bad85a57c880864d3d1ebaef8db --- /dev/null +++ b/2025/TANGO_ Training-free Embodied AI Agents for Open-world Tasks/layout.json @@ -0,0 +1,6729 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 512, + 121 + ], + "type": "text", + "content": "TANGO: Training-free Embodied AI Agents for Open-world Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 142, + 176, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 142, + 176, + 158 + ], + "spans": [ + { + "bbox": [ + 84, + 142, + 176, + 158 + ], + "type": "text", + "content": "Filippo Ziliotto" + }, + { + "bbox": [ + 84, + 142, + 176, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 143, + 306, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 143, + 306, + 158 + ], + "spans": [ + { + "bbox": [ + 200, + 143, + 306, + 158 + ], + "type": "text", + "content": "Tommaso Campari*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 143, + 413, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 143, + 413, + 156 + ], + "spans": [ + { + "bbox": [ + 326, + 143, + 413, + 156 + ], + "type": "text", + "content": "Luciano Serafini²" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 437, + 143, + 524, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 143, + 524, + 156 + ], + "spans": [ + { + "bbox": [ + 437, + 143, + 524, + 156 + ], + "type": "text", + "content": "Lamberto Ballan1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 171, + 375, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 171, + 375, + 185 + ], + "spans": [ + { + "bbox": [ + 235, + 171, + 375, + 185 + ], + "type": "text", + "content": "1 University of Padova, Italy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 185, + 186, + 422, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 186, + 422, + 200 + ], + "spans": [ + { + "bbox": [ + 185, + 186, + 422, + 200 + ], + "type": "text", + "content": "Fondazione Bruno Kessler (FBK), Trento, Italy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 251, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 480 + ], + "type": "text", + "content": "Large Language Models (LLMs) have demonstrated excellent capabilities in composing various modules together to create programs that can perform complex reasoning tasks on images. In this paper, we propose TANGO, an approach that extends the program composition via LLMs already observed for images, aiming to integrate those capabilities into embodied agents capable of observing and acting in the world. Specifically, by employing a simple PointGoal Navigation model combined with a memory-based exploration policy as a foundational primitive for guiding an agent through the world, we show how a single model can address diverse tasks without additional training. We task an LLM with composing the provided primitives to solve a specific task, using only a few in-context examples in the prompt. We evaluate our approach on three key Embodied AI tasks: Open-Set ObjectGoal Navigation, MultiModal Lifelong Navigation, and Open Embodied Question Answering, achieving state-of-the-art results without any specific fine-tuning in challenging zero-shot scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 135, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 135, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 135, + 516 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 525, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 668 + ], + "type": "text", + "content": "Large Language Models (LLMs) have gained significant attention in the field of AI due to their remarkable capability to generalize across unseen tasks [5, 24, 48, 49]. Systems like VisProg and ViperGPT [18, 47] demonstrated strong performance on a broad range of vision-and-language tasks by just relying on a few numbers of in-context examples to generate complex compositional programs without requiring any specific training. Nevertheless, the evaluation of these concepts to Embodied AI, particularly for navigation tasks, remains largely unexplored. This approach may be crucial in building agents capable to navigate and operate efficiently in unfamiliar environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "type": "text", + "content": "Previously, methods such as Neural Module Networks (NMN) [2, 19] have demonstrated promising compositional" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 320, + 225, + 550, + 469 + ], + "blocks": [ + { + "bbox": [ + 320, + 225, + 550, + 469 + ], + "lines": [ + { + "bbox": [ + 320, + 225, + 550, + 469 + ], + "spans": [ + { + "bbox": [ + 320, + 225, + 550, + 469 + ], + "type": "image", + "image_path": "8f7b7e81ceb8828b89be8dd0cff489b6023a2caad0fdf20e17f343f77f9b4ce0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 479, + 555, + 545 + ], + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 545 + ], + "type": "text", + "content": "Figure 1. We introduce TANGO, a modular neuro-symbolic system for compositional embodied visual navigation. Given a few examples of natural language instructions and the corresponding programs composed of action primitives, TANGO can generate executable programs, enabling the agent to perform multiple tasks within a 3D environment." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "content": "properties for high-level tasks like visual question answering, via end-to-end training of networks and specialized, differentiable neural modules. However, they required the combination of semantic parsers and predefined templates to learn to solve the compositional tasks. Alternative approaches leverage LLMs through predefined task modules or APIs for action execution [22, 23, 32, 52, 58]. While effective, these methods often lack a compositional structure to address multiple tasks or primarily focus on manipulation and robotic planning, with limited emphasis on navigation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "In this work, we present TANGO (see Figure 1) a novel neuro-symbolic compositional approach which uti-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 215, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 215, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 215, + 713 + ], + "type": "text", + "content": "*The authors contributed equally to the work." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24603" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "content": "izes primitives and employs a LLM as a planner to sequence these primitives within photorealistic 3D environments, where agents must perceive and act. This framework integrates high-level planning with low-level action execution, without the need for training. TANGO makes use of diverse modules designed for visual navigation and question-answering tasks, resulting in a system that seamlessly addresses both challenges while achieving SoA performance without requiring any prior adjustments. By providing a few in-context examples that show how to tackle multiple tasks, TANGO is capable of generalizing to the specific task at hand. This is facilitated by the LLM, which effectively combines the individual modules available within the TANGO system. Moreover, it extends [55] exploration policy by incorporating a memory mechanism in the form of a stored feature map that retains information about previously explored areas, supporting for efficient life-long navigation tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 294, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 294, + 382 + ], + "type": "text", + "content": "To demonstrate the flexibility of the proposed framework, TANGO has been tested on three popular benchmarks: namely, i) Open-Vocabulary ObjectGoal Navigation [56], ii) Multi-Modal Lifelong Navigation [26], and iii) Embodied Question Answering [15, 34]. Results match or surpass previous approaches, without requiring specialized training. Summing up, the contributions of this paper are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 384, + 295, + 516 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 384, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 295, + 418 + ], + "type": "text", + "content": "- Introduction of a neuro-symbolic compositional LLM-based framework for EAI leveraging specialized primitive modules." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 421, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 295, + 456 + ], + "type": "text", + "content": "- Demonstration of robust generalization capabilities across multiple tasks without the need for any specific training or fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 457, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 457, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 457, + 295, + 491 + ], + "type": "text", + "content": "- Extension of the exploration policy presented in [55] to multi-goal scenarios through the incorporation of a memory mechanism stored as a feature vector map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 493, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 295, + 516 + ], + "type": "text", + "content": "- Achievement of state-of-the-art results, underscoring the effectiveness of the proposed framework." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 529, + 147, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 147, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 147, + 542 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 551, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 295, + 670 + ], + "type": "text", + "content": "Our work takes inspiration from modular approaches for visual tasks, like the seminal Neural Module Networks [2, 19], as well as the recent VisProg [18] and ViperGPT [47] frameworks. The main contribution of this paper is to extend these ideas to embodied visual navigation, that usually require heavy end-to-end learning, although modular approaches have recently shown promising results [7, 17]. Therefore, we discuss prior work in the area of embodied AI, the use of language models in robotics, and program generation approaches for image recognition tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Embodied AI and Visual Navigation. The field of Embodied AI has recently undergone a paradigm shift, fuelled" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": "by the emergence of highly efficient simulators [27, 36, 43-45]. These simulators enable processing numerous parallel simulations in photorealistic indoor environments, facilitating large-scale testing that was otherwise challenging in classical robotics. Alongside these simulators, a large variety of tasks (and benchmarks) has been proposed: PointGoal Navigation [50], where an agent navigates from point A to point B in an unknown environment; ObjectGoal Navigation [3], requiring the agent to locate and navigate to an object in the scene; Instance-Image Goal Navigation [28], akin to ObjectGoal Navigation, but in which the agent should find a specific object depicted in a given image; Multi-Modal Lifelong Navigation [26] involving navigating to a sequence of target objects that can be specified through labels, images, or textual descriptions; Embodied Question Answering [15, 34], in which an agent navigates an environment to gather information needed to answer a question, and Vision and Language Navigation (VLN) [1]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 289, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 502 + ], + "type": "text", + "content": "Various approaches have been proposed to tackle these tasks, primarily falling into three categories: end-to-end [6, 53, 54], modular [11-13, 29, 39, 40] and LLM-based [16, 30, 57-59]. However, the main limitation of these paradigms is their dependence on task-specific architectures. In the case of end-to-end approaches, a model is trained for a specific problem, often requiring days of training, such as in [50], where a policy for PointGoal Navigation was trained for 2.5 billion steps. To adapt the learned model to a new problem, adjustments and retraining are necessary, and this comes with a high cost. Modular approaches share a similar challenge, usually relying on policies designed for specific tasks, although common elements, such as exploration and navigation modules, are often shared [11, 12, 29, 39, 40]. Adapting these solutions to different tasks requires a manual adjustment of the approach to fit the new domain, often involving the addition or modification of modules." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "content": "Language models for Robotics In robotics, foundation models (LLMs) trained on vast internet-scale datasets [5, 24, 48, 49] have the potential to equip robots with real-world priors and advanced reasoning abilities without the need for extensive task-specific training. Early approaches equipped agents with learned language embeddings, requiring large amounts of training data [4, 32]. Recent studies, on the other hand, have explored zero-shot and few-shot solutions mainly focusing on robotic planning and manipulation tasks [22, 23, 32, 41, 46]. In the context of visual navigation, LLM-based approaches leverage the powerful LLMs priors and reasoning capabilities to guide navigation within the environment [16, 21, 30, 57-59]; however, despite their remarkable contributions, they often lack a modular design that allows for flexible integration and extensibility, as they are typically optimized for single tasks rather" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24604" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 237, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 237, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 237, + 83 + ], + "type": "text", + "content": "than a comprehensive, adaptable framework." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 297, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 297, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 297, + 255 + ], + "type": "text", + "content": "In contrast, our work deviates from these paradigms by relying on a series of diverse pre-trained modules that, thanks to a LLM, can be combined to potentially solve various tasks, compositionally. None of our modules are fine-tuned explicitly for the target problems, placing us in a zero-shot setting. Moreover, TANGO does not only rely on an exploration policy based entirely on LLM output, as in [16, 58, 59]. Instead, it combines a frontier exploration policy with LLM priors and memory guidance to enhance navigation performance. To achieve this, we use a pre-trained PointGoal policy as the base waypoint navigation module of TANGO and extend the exploration policy introduced in [55] by equipping the agent with a memory mechanism stored as a feature vector for each pixel of the map." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 437 + ], + "type": "text", + "content": "Modular vision and program composition for visual tasks. Neural Module Networks (NMN) [2, 19] introduced modular and compositional methodologies for visual question answering (VQA). NMNs integrate neural modules into an end-to-end differentiable network; the approach originally relied on pre-existing parsers [2], whereas more recent methods [19, 20, 25] have evolved to learn the layout generation model concurrently with the neural modules, employing reinforcement learning [51] and weak supervision. Stack-NMN [20] extends N2NMN by transitioning from discrete to soft layout generation, incorporating a weighted average of predictions from all modules at each step, determined by a layout generation network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 297, + 715 + ], + "type": "text", + "content": "Recently, [18] introduced VisProg, a framework that offers two key advantages over NMNs. Firstly, it constructs high-level programs that invoke state-of-the-art neural models and Python functions at intermediate steps, diverging from the conventional approach of generating end-to-end neural networks. This design facilitates the integration of symbolic, non-differentiable modules. Secondly, it leverages the in-context learning ability of LLMs [5]. This is achieved by prompting the LLM with a natural language instruction (or a visual question or a statement to be verified), along with a few examples of similar instructions and their corresponding programs. Almost concurrently, ViperGPT [47] presented a similar approach to VisProg, but in this case the model directly generates Python code, instead of composing pre-defined modules and routines. In contrast, in our framework, the LLM is provided only with high-level knowledge of the functions of each primitive, which is encoded directly in the primitive's name. Therefore, the system it is built similarly to VisProg. It is noteworthy that both VisProg and ViperGPT were tested on images, while our model is specifically designed for embodied agents operating in an interactive environment, relying on sensor-based perception." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 320 + ], + "type": "text", + "content": "Neuro-symbolic approaches offer the possibility to address a broad range of diverse complex tasks efficiently. Systems like VisProg [18] operate on images and have demonstrated excellent results in zero-shot settings. Understanding how a similar approach works in an action-oriented scenario is key to enable robots to navigate autonomously in novel environments. Therefore, we develop this paradigm further into EAI, where the integration with action execution adds several challenges. Inspired by [18], given an input prompt, we rely on a LLM to generate synthetic pseudo-programs that can be executed by the embodied agent in the environment. Our TANGO framework is therefore able to generalize to new tasks, without any direct training on task-specific datasets, thus effectively mitigating the heavy computational training costs inherent to embodied navigation tasks. The procedure that enables generating these new programs is based on \"in-context examples\" that are fed to the LLM, alongside a natural language instruction. An outline of TANGO is shown in Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 328, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 389 + ], + "type": "text", + "content": "TANGO Interpreter. The first key component of the proposed framework is referred to as \"Program Interpreter\". It comprises visual recognition modules that can be used by the agent to extract the semantics of the scene, as well as to provide an understanding of the visual context." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 390, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 556, + 665 + ], + "type": "text", + "content": "As shown in Figure 1, users can ask questions or give specific tasks to the agent; the natural language prompt is then processed by the LLM (in our implementation GPT-4o [5]), which serves as a planner and outputs a step-by-step executable program. To ensure the LLM delivers a reasonable output, it is fed with 15 \"in-context examples\" across diverse tasks. This enables the LLM to make use of its reasoning capabilities effectively, identifying the most suitable planning for the required task. Moreover, the examples remain the same regardless of the task identified in the prompt; it is the LLM's responsibility to output the specific program target for the given question. Programs use a higher level of abstraction than previous modular attempts such as Neural Module Networks (NMN) [2, 19]. Each program is constructed as a sequence of primitives (e.g., detect, answer, match, etc.) that invoke corresponding TANGO modules. These modules are either powered by pre-trained state-of-the-art vision models or implemented as simple Python subroutines (e.g., count, is_found, eval, etc.), with additional navigation-specific modules designed to \"steer\" the agent's movements (e.g. navigate_to, return, turn, etc.). Figure 3 provides a comprehensive list of all the currently implemented modules." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 557, + 715 + ], + "type": "text", + "content": "Figure 2 shows an Embodied Question Answering (EQA) example for the user's question: \"on the left of the toilet there's a shower, is the shower curtain open or closed?\" The LLM, which acts as a planner, transforms the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24605" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 545, + 379 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 545, + 379 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 545, + 379 + ], + "type": "image", + "image_path": "a79c71eb605a0bc6cae2e5569d8ee535c0a077ecb5e8eded071b2392ae2c8f12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 389, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 54, + 389, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 555, + 413 + ], + "type": "text", + "content": "Figure 2. Overview of the program generation in TANGO. Given few \"in-context examples\", the LLM provide a detailed sequence of steps to be executed by the agent in the given environment. The LLM is instructed to comment its output to allow for explainability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 433, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 433, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 295, + 541 + ], + "type": "text", + "content": "user's initial prompt into navigation subtasks by leveraging its reasoning abilities and powerful priors. This decomposition enables the model to break down complex queries into more manageable steps, guiding the agent's navigation process effectively. Each line of the generated program corresponds to a module serving a specific task. The agent must then execute the generated program. The LLM is also guided to comment on its steps directly within the generated pseudocode." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 662 + ], + "type": "text", + "content": "All the given primitives are equipped with methods to: i) parse lines in order to extract input argument names and values, as well as the output variable name; ii) execute the module in the environment, and update the program state with the output variable name and value. The outputs for each step and the comments in the generated pseudocode can also be used to better understand what is happening under the hood and why the LLM generated a specific line, thus enabling a good interpretability of the system behavior." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "TANGO modules utilize open-source software and models, readily downloadable from the web. They can also be effortlessly updated with newer, more efficient models as they become available." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 433, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 578 + ], + "type": "text", + "content": "Navigation Module. In order to navigate the environment, we define a module employing a PointGoal navigation agent as our foundational module [40, 55]. This model achieved nearly perfect PointGoal performance, both in terms of success (" + }, + { + "bbox": [ + 313, + 433, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\sim 99\\%" + }, + { + "bbox": [ + 313, + 433, + 555, + 578 + ], + "type": "text", + "content": ") and efficiency (a forward pass that takes a fraction of a second) on standard datasets [9, 38]. The agent starts the exploration of the environment until it locates its target goal. Once the target is identified, the focus of exploration transitions to reaching the designated goal. This PointNav policy exclusively uses the egocentric depth image and the robot's relative distance and heading towards the desired goal point as input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "text", + "content": "Exploration Policy. Exploration is performed using the policy outlined in [55]. This method builds occupancy maps based on depth observations to identify exploration frontiers. It further leverages RGB observations and a pretrained vision-language model (BLIP2 [31]) to generate a language-grounded \"value map\". This value map is then used to guide the exploration, facilitating an efficient search for instances of a given target. Notably, as in [55], the agent performs a " + }, + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "text", + "content": " turn at the start of navigation to initialize frontiers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "We extend this policy for sequential goals by incorpora" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24606" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 295, + 140 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 295, + 140 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 295, + 140 + ], + "type": "image", + "image_path": "14ba6ce6596cd3c232b08380c74a7d43e3b9e114dab00fe5dddeee2a6d0dd781.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 144, + 295, + 212 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 212 + ], + "type": "text", + "content": "Figure 3. Overview of TANGO modules. Modules span a variety of inputs and outputs. Orange modules use Python subroutines, while blue modules use pre-trained computer vision models (similarly to [18]). The navigate_to and explore(scene modules, in green, both implement our foundational PointNav module; however, only explorescene integrates the memory mechanism." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 232, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 232, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 295, + 365 + ], + "type": "text", + "content": "ing a memory mechanism, represented as a “feature map” in which each pixel of the value map is encoded as a vector and updated at each step. This feature map then updates the language-grounded original value map when a new target is specified, by calculating the cosine similarity between each pixel's vector and the new target embeddings (either from text or image). We sample the highest value in the updated value map; if this value exceeds a predefined threshold, indicating the agent has previously encountered and “rememberers” the target's location, the agent navigates to it, supporting lifelong navigation (see Section 2-Supp)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 371, + 224, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 224, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 224, + 384 + ], + "type": "text", + "content": "3.1. Navigation Tasks with TANGO" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 437 + ], + "type": "text", + "content": "TANGO leverages pre-trained multimodal vision-language and vision-only models as foundational components to extract semantic information from the scene, making it well-suited for several Embodied AI tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 437, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 559 + ], + "type": "text", + "content": "(Open-set) ObjectGoal Navigation [3, 56], in particular, emerges as a suitable testbed to evaluate the efficacy of our method. A key module to tackle this problem involves utilizing an object detector (Owlv2 in our implementation [35] or DETR [8] for target objects that fall within the COCO classes [33]) to identify objects within the image. Subsequently, navigation towards the detected objects is facilitated by waypoints, leveraging depth distance calculation to guide the agent effectively. For simplicity, we use the center of the bounding box to determine the target waypoint." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "text", + "content": "TANGO also integrates a specialized module for matching target objectives with ongoing visual observations in image-based scenarios, resulting in accurate navigation towards an instance image. To evaluate our agent's performance in this context, we use real-time matching SuperGlue network [42], optimized for indoor environments (i.e. using hyperparameters recommended in the respective paper)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "TANGO being task-agnostic, can process any type of navigation target, regardless of the order in which it is presented. In this context, the GOAT benchmark [26] is well-suited for evaluating our system, as it operates in scenarios where targets are specified through images, text, or descriptive phrases, provided in random sequence. It also heavily" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "relies on memory to find previously seen targets when sequential goals are given." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 96, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 204 + ], + "type": "text", + "content": "An agent should also be able to answer user queries like \"can you check if the kitchen table is clean?\" Hence, EQA serves as an excellent benchmark to assess TANGO's capabilities and robustness. EQA consists of three phases: understanding the semantic structure of the prompt, locating the target object(s), and analyzing the visual semantics to generate an accurate response. To this end, TANGO integrates a specific answer module, relying on BLIP2 [31] as its core foundation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 216, + 395, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 395, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 395, + 229 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "content": "TANGO provides a flexible framework that can be applied to various embodied navigation problems. We evaluate our approach on three popular tasks that require a wide range of capabilities, including efficient environment exploration, path planning, scene and context understanding, and image similarity comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 316, + 430, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 430, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 430, + 330 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": "Agent Configuration. Prior research on visual navigation commonly uses different agent configurations depending on the considered task [15, 26, 56]. The configuration typically employed for (OPEN-SET) OBJNAV closely resembles that of a LoCoBot, with an agent height of " + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "inline_equation", + "content": "0.88m" + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": ", a radius of " + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "inline_equation", + "content": "0.18m" + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": ", and a single " + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": " RGB sensor with a " + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "inline_equation", + "content": "79^{\\circ}" + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": " hfov, positioned " + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "inline_equation", + "content": "0.88m" + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": " above the ground." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "text", + "content": "In the context of GOAT Benchmark, some of the default settings differ, featuring a camera situated " + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "inline_equation", + "content": "1.31m" + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "text", + "content": " above the ground. The agent has a height of " + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "inline_equation", + "content": "1.41m" + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "text", + "content": " and is given " + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "inline_equation", + "content": "360 \\times 640" + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "text", + "content": " RGB images. Both the above tasks' configurations use a step size of " + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "inline_equation", + "content": "0.25m" + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "text", + "content": " and a left and right turning angle of " + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 419, + 554, + 526 + ], + "type": "text", + "content": ". Lastly, the settings for EQA mirror those of PointGoal Navigation as specified in the Habitat-Lab configuration files. All other settings remain consistent with those of [15]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 554, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 650 + ], + "type": "text", + "content": "Datasets and Evaluation Protocol. We assess the TANGO performance across task-specific datasets using the Habitat simulator [44]. As the system is not trained in any way (except for the PointGoal Navigation model, which has been pre-trained on the training set of HM3D, never used in our tests), each scene within episodes is novel to the agent. Therefore, the entire validation set can be categorized as \"unseen\". We now describe in detail the tasks and datasets used:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "- OPEN VOCABULARY OBJECT NAVIGATION (OVON) [56]: a large-scale benchmark featuring over 15,000 annotated household objects across 379 categories, derived from real-world 3D scans [38]. The agent is initialized at a random location within a scene and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24607" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 107 + ], + "type": "text", + "content": "tasked with navigating to a target object category within a time limit of 500 steps. We evaluate our approach across all episodes included in the \"val unseen\" set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 407 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 108, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 227 + ], + "type": "text", + "content": "- MULTI-MODAL LIFELONG NAVIGATION (GOAT-BENCH) [26]: an agent is tasked with sequentially navigating to five to ten target objects identified by category names, descriptions, or images. Each target represents a subtask executed within an open-vocabulary framework that spans over 312 categories. The agent is required to reach a goal within a specified time constraint and is assigned new targets upon the completion of each subtask. We evaluate our approach across all episodes included in the \"val unseen\" set." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 407 + ], + "type": "text", + "content": "- EMBodied Question ANSWERING (OPENEQA) [34]: containing over 1,600 question-answer pairs sourced from more than 180 real-world environments and scans [14, 38]. It is divided into two task categories: Episodic Memory-EQA (EM-EQA) and Active-EQA (A-EQA). Our focus is on the latter, in which the agent must autonomously navigate and answer questions within a time constraint of 500 steps. Questions span over seven categories, including world knowledge, attribute recognition, spatial reasoning, and object localization. Examples of questions from the episodes include: \"Is the microwave door propped open?\" and \"What is left of the kitchen pass-through?\". We evaluate our approach across all episodes belonging to the A-EQA task category." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 421, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 295, + 445 + ], + "type": "text", + "content": "Evaluation Metrics. In all tasks, we use the standard metrics as in prior works [3, 15, 34, 56]. Namely:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 445, + 295, + 660 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 445, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 480 + ], + "type": "text", + "content": "- Success Rate (SR): measures the ratio of episodes where the agent successfully reached its target (Open ObjNav, GOAT-Bench)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "text", + "content": "- Success weighted by Path Length (SPL): measures the optimality of the path taken by the agent w.r.t. the optimal path (Open ObjNav, GOAT-Bench)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 552 + ], + "type": "text", + "content": "- Distance to Goal (DTG): measures the average distance to goal of the agent at the end of the episode (Open ObjNav, GOAT-Bench)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "text", + "content": "- Score (LLM-Match): an LLM compares the ground-truth " + }, + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "inline_equation", + "content": "GT_{i}" + }, + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "text", + "content": " answers with model output " + }, + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "text", + "content": " given a question " + }, + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "text", + "content": " and assigns a score " + }, + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 55, + 552, + 295, + 660 + ], + "type": "text", + "content": " on a scale of 1 to 5. On this scale, 1 indicates an incorrect response, 5 represents a correct response, and intermediate values reflect varying levels of similarity. For example, the answer to the question \"What color is the bed\" could be correctly answered either as \"white\" or \"the bed is white\". The final results aggregation is as follows:" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 121, + 664, + 295, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 664, + 295, + 697 + ], + "spans": [ + { + "bbox": [ + 121, + 664, + 295, + 697 + ], + "type": "interline_equation", + "content": "S = \\frac {1}{N} \\sum_ {i} ^ {N} \\frac {\\sigma_ {i} - 1}{4} \\times 100 \\% \\tag{1}", + "image_path": "cb12e2b494a9fdf15ef2c0062ed5f386447e413297cf1966ce9821f0320ad7bd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 702, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 295, + 713 + ], + "type": "text", + "content": "- Answer Accuracy: the average accuracy of the answers" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "type": "text", + "content": "provided by the agent for the EQA task. This metric, used in [15], is presented with the related results in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 152, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 392 + ], + "type": "text", + "content": "Implemented Modules. TANGO contains descriptive and easily understandable names for the modules, arguments, and variables to facilitate the LLM comprehension of the input and output. Figure 3 showcases a list of all available modules. Each module exclusively outputs predefined object variables to the next one, enabling the possibility to monitor the progression of the agent at each step. Therefore, dramatically enhancing the explainability of failure cases, as it allows for a deep examination of the agent's decision-making process through its interactions with the environment. Figure 6 provides a general overview of the failure analysis in EQA task. As presented in Section 3, \"navigate_to\" and \"explore_structure\" modules are specifically engineered to navigate the environment, leveraging inputs from depth and RGB sensors as well as the PointGoal GPS + compass sensors. If this sensor is not available for a specific task, it can be derived from the current agent pose and the exploration goal location. The agent is provided with the actions: Forward, Move Left, Move Right, Look Up, Look Down and Stop." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 397, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 553 + ], + "type": "text", + "content": "For the \"detect\" module, we employ Owlv2 [35] object detector for general classes and use DETR [8] for categories within the COCO classes [33]. Notably, in synthetic 3D scenes, we noticed numerous false positives. To address the problem, bounding boxes produced by the detector are forwarded to a \"classify\" module. Within this module, each detection undergoes classification to differentiate between categories of similar instances. For example, the class \"chair\" encompasses diverse subcategories such as \"armchair\", \"couch\", and \"other\". Subclasses are outputted under the hood directly by the LLM requiring no human annotation. We utilize a CLIP-based classifier in its default configuration [37]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": "Embodied Question Answering (EQA) task uses an \"answer\" module based on BLIP2 [31], capable of performing various multi-modal tasks, including Visual Question Answering, Image-Text retrieval, and Image Captioning. Notably, the module is also used when goals are specified through images, to capture the class of the input image before starting the navigation. Furthermore, for image-based tasks we exploit a \"match\" module to evaluate the similarity between the agent's observation and the instance image target. This module performs feature matching between the two, as described in Section 3.1, to aid in navigation. Hence, it assists the agent in determining whether what it perceives as the current target is indeed the correct goal." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24608" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 288 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 288 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 288 + ], + "type": "image", + "image_path": "1dd0c7143bf53da833fd5a99cffd8672f2a01db8f06fed4fc24d4b29efbfbae9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 297, + 295, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 352 + ], + "type": "text", + "content": "Figure 4. Examples from OpenEQA [34]. The top section illustrates a successful episode where TANGO is able to understand the input query, correctly specifying the sequential targets. The lower section illustrates a failure caused by overly general directions from the LLM, which TANGO struggled to resolve." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 87, + 375, + 264, + 506 + ], + "blocks": [ + { + "bbox": [ + 87, + 375, + 264, + 506 + ], + "lines": [ + { + "bbox": [ + 87, + 375, + 264, + 506 + ], + "spans": [ + { + "bbox": [ + 87, + 375, + 264, + 506 + ], + "type": "table", + "html": "
MethodVAL UNSEEN
SR (↑)SPL (↑)
RL [56]18.6 ±0.37.5 ±0.2
BCRL [56]8.0 ±0.22.8 ±0.1
DAgRL [56]18.3 ±0.37.9 ±0.1
*VLFM [55, 56]35.219.6
DAgRL+OD [56]37.1 ±0.219.9 ±0.3
TANGO (ours)35.5 ±0.319.5 ±0.3
", + "image_path": "6e82169919ca102b03c10664b9d1e54747cf934698a6e99e5adc42f63784f442.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 569, + 179, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 179, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 179, + 582 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Open-set ObjectGoal Navigation We evaluated our method in the standard Open-Set OBJNAV setting, with targets sampled directly from objects in the scene. As shown in Table 1, our approach achieves performance on par with state-of-the-art methods (rows 4-5) on the validation-unseen split both in SR (35.5%) and SPL (19.5%). Instead of focusing solely on numerical results, these findings highlight TANGO's flexibility in solving navigation tasks in unfamiliar environments, underscoring the potential of our approach within the OBJNAV framework." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 560, + 173 + ], + "blocks": [ + { + "bbox": [ + 55, + 513, + 295, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 536 + ], + "type": "text", + "content": "Table 1. HM3D-OVON. Performance comparison of different methods on VAL UNSEEN split. * deterministic method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 560, + 173 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 560, + 173 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 560, + 173 + ], + "type": "table", + "html": "
MethodScore (↑)
Human Agent [34]85.1 ± 1.1
Blind LLMs [34]35.5 ± 1.7
Socratic LLMs w/ Frame Captions [34]38.1 ± 1.8
Socratic LLMs w/ Scene-Graph Captions [34]34.4 ± 1.8
TANGO (ours)37.2 ± 1.8
", + "image_path": "9535f5e8c1a3f3950ba4241f50dbfb7900c7cbf6cb18aedc17f2ee7c27b07bae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "type": "text", + "content": "Table 2. OpenEQA results. Comparison of TANGO with the SoA on the OpenEQA dataset. All methods involve zero-shot approaches." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 220, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 520 + ], + "type": "text", + "content": "Embodied Question Answering. In EQA, the agent is queried with a natural language question and must autonomously explore the environment and gather information in order to answer accordingly. In recent literature, this task shifted from being regarded as a purely classification problem aimed at determining the most suitable answer from a set of pre-defined possibilities (i.e., class labels) to an open-vocabulary benchmark where the agent must answer with natural language [15, 34]. We compare TANGO results, using the score from Eq. 1, against other zero-shot approaches. Table 2 shows that our method ranks second and closely matches the performance of the leading SoA method, with a gap of less than " + }, + { + "bbox": [ + 313, + 220, + 555, + 520 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 220, + 555, + 520 + ], + "type": "text", + "content": ". Notably, these approaches [34] utilize LLMs to interpret scene objects step-by-step (rows 3-4), but they do not leverage LLMs to effectively guide navigation towards specific targets. Hence, our approach demonstrates that SoA results can still be achieved by pre-planning a path that prioritizes relevant areas for a given question. Moreover, the results indicate that human agents achieve a significantly higher score of " + }, + { + "bbox": [ + 313, + 220, + 555, + 520 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 220, + 555, + 520 + ], + "type": "text", + "content": ", which remains well above the performance achieved by using large language models (LLMs), particularly in open-world settings. This gap underscores the challenges LLMs face in handling the complexity and variability of open-world scenarios compared to human agents." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": "GOAT. We assess our method within the GOAT-Bench setting, where the agent is spawned randomly and tasked with sequential targets. Each goal can be specified either by its category name, a description, or an image (e.g. the input could be \"gas boiler\", \"the gas boiler on the corner of the room. The gas boiler is located on the left of the washing machine and freezer\" or an image of the \"gas boiler\"). Table 3 compares our approach to the methods in [10]. TANGO significantly outperforms other state-of-the-art techniques (rows 1-4), achieving a " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "+2.6\\%" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " gain in the success metric. Additionally, it ranks second in navigation efficiency, with only a slight " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "-0.7\\%" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " difference from the top-performing method. We primarily attribute these results to the LLM's ability to accurately transform target descriptions into a sequence of selected goals, rather than relying" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24609" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 279, + 187 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 279, + 187 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 279, + 187 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 279, + 187 + ], + "type": "table", + "html": "
MethodVAL UNSEEN
SR (↑)SPL (↑)
SenseAct-NN Skill Chain [26]29.511.3
SenseAct-NN Monolithic [26]12.36.8
Modular GOAT [26]24.917.2
Modular Clip on Wheels [26]12.310.4
TANGO (ours)32.116.5
", + "image_path": "2aa5ccd4f5751c9c77f8fc242b2aa039d28f92a6ee83c3cf9fe4ef15b2a3a4ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 196, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 196, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 295, + 218 + ], + "type": "text", + "content": "Table 3. GOAT-Bench. Performance comparison of different methods on the VAL UNSEEN split." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 231, + 298, + 345 + ], + "blocks": [ + { + "bbox": [ + 58, + 231, + 298, + 345 + ], + "lines": [ + { + "bbox": [ + 58, + 231, + 298, + 345 + ], + "spans": [ + { + "bbox": [ + 58, + 231, + 298, + 345 + ], + "type": "image", + "image_path": "33565f800cdb158f4546f0bafdf71bb56ab35c441bdbeae367a5ef776003ccef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 356, + 295, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 411 + ], + "type": "text", + "content": "Figure 5. Multi-Modal Lifelong Navigation Success Example. (top) RGB observation of the target during STOP action (step " + }, + { + "bbox": [ + 55, + 356, + 295, + 411 + ], + "type": "inline_equation", + "content": "t_{i+steps}" + }, + { + "bbox": [ + 55, + 356, + 295, + 411 + ], + "type": "text", + "content": "). (middle) Value map for the specific target recomputed from the memory map (step " + }, + { + "bbox": [ + 55, + 356, + 295, + 411 + ], + "type": "inline_equation", + "content": "t_{i+steps}" + }, + { + "bbox": [ + 55, + 356, + 295, + 411 + ], + "type": "text", + "content": "). (bottom) Memory map after target changes (step " + }, + { + "bbox": [ + 55, + 356, + 295, + 411 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 356, + 295, + 411 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": "solely on potentially \"noisy\" text embeddings. Moreover, the use of a memory mechanism is essential for the task (Fig. 5); however, if the memorized target object is incorrect, the path efficiency can significantly decrease." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 194, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 194, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 194, + 491 + ], + "type": "text", + "content": "4.3. TANGO Failure analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "Due to the high explainability of our system, we conducted an analysis of the failures encountered during EQA episodes (Fig. 6). We extracted a significant subsample for the task and manually classified the instances where the model failed. We considered scores below 3 as failures, i.e. incorrect answers. We observed that the main cause of failure can be attributed to the \"detect\" module (stopped at wrong object or Ignored goal object in the image). Furthermore, the exploration policy appears to perform well given the targets produced by the LLM, as it fails in only " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sim 10\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": " of the cases (labeled as: didn't see target goal in the image). The LLM generates incorrect pseudo-code around " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": " of the time, with " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": " failures leading to ambiguous or incorrect targets, and " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "11.2\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": " due to incorrect primitive ordering or usage which instead leads directly to episode failure. Notably, only " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "11.2\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": " of errors are actually due to the LLM generating incorrect code, while the remaining is attributable to prompt-related issues. A comprehensive fail-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 324, + 72, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 324, + 72, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 324, + 72, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 547, + 220 + ], + "type": "image", + "image_path": "669b7466ee6c54f43c6f6686209f35f2169febe03ef95c60344ba09a232f1c75.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "text", + "content": "Figure 6. Causes of system failure in EQA tasks. The analysis was manually conducted on a significant sample of failed episodes, facilitated by the explainability of TANGO. Answers were deemed as \"incorrect\" if they received a score below three." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 535, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 535, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 535, + 309 + ], + "type": "text", + "content": "ure analysis is provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 327, + 387, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 387, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 387, + 339 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 350, + 555, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 588 + ], + "type": "text", + "content": "We conducted a systematic analysis of our compositional approach, offering valuable insights into the future of modular, neuro-symbolic systems. Moreover, TANGO highlights the versatility of simple PointGoal Navigation agents equipped with specific task-oriented modules, yielding promising results in zero-shot scenarios across all considered tasks. TANGO effectively handles diverse multi-modal prompts, following instructions to complete tasks, underscoring the potential of LLMs in robotic navigation. Its integrated memory supports lifelong navigation, suggesting improved capabilities with increased exploration. However, it may struggle with overly complex or ambiguous prompts, which can limit the LLM's ability to identify the correct primitives or targets. Future work could extend this system to tasks such as Visual Language Navigation (VLN) and explore the use of open-source LLMs. Additionally, refining the memory mechanism—particularly through improved sampling strategies for identifying high-value pixels linked to objects' memory—could further enhance the system's effectiveness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 607, + 411, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 607, + 411, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 607, + 411, + 620 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "We thank the University of Padua, Department of Mathematics \"Tullio Levi-Civita\", for providing the computational resources. TC and LS were supported by the PNRR project Future AI Research (FAIR - PE00000013) under the NRRP MUR program, funded by NextGenerationEU. We also thank Elena Izzo and Luca Parolari for their valuable comments on early drafts of this paper." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24610" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "text", + "content": "[2] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 257 + ], + "type": "text", + "content": "[3] Dhruv Batra, Aaron Gokaslan, Aniruddha Kembhavi, Oleksandr Maksymets, Roozbeh Mottaghi, Manolis Savva, Alexander Toshev, and Erik Wijmans. ObjectNav revisited: On evaluation of embodied agents navigating to objects. arXiv preprint arXiv:2006.13171, 2020. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Zhenshan Bing, Alexander Koch, Xiangtong Yao, Kai Huang, and Alois Knoll. Meta-reinforcement learning via language instructions. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5985-5991. IEEE, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 295, + 434 + ], + "type": "text", + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Proc. of Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[6] Tommaso Campari, Paolo Eccher, Luciano Serafini, and Lamberto Ballan. Exploiting scene-specific features for object goal navigation. In Proc. of the European Conference on Computer Vision Workshops (ECCVW), 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 480, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 534 + ], + "type": "text", + "content": "[7] Tommaso Campari, Leonardo Lamanna, Paolo Traverso, Luciano Serafini, and Lamberto Ballan. Online learning of reusable abstract models for object goal navigation. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "type": "text", + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 591, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 295, + 646 + ], + "type": "text", + "content": "[9] Angel X. Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3D: Learning from RGB-D data in indoor environments. In Proc. of the International Conference on 3D Vision (3DV), 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 58, + 647, + 295, + 690 + ], + "type": "text", + "content": "[10] Matthew Chang, Theophile Gervet, Mukul Khanna, Sriram Yenamandra, Dhruv Shah, So Yeon Min, Kavit Shah, Chris Paxton, Saurabh Gupta, Dhruv Batra, et al. Goat: Go to any thing. arXiv preprint arXiv:2311.06430, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 692, + 295, + 713 + ], + "type": "text", + "content": "[11] Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to ex" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "text", + "content": "plore using active neural SLAM. In Proc. of the International Conference on Learning Representations (ICLR), 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 96, + 554, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 554, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 554, + 149 + ], + "type": "text", + "content": "[12] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. In Proc. of Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "text", + "content": "[13] Devendra Singh Chaplot, Ruslan Salakhutdinov, Abhinav Gupta, and Saurabh Gupta. Neural topological slam for visual navigation. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 195, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 249 + ], + "type": "text", + "content": "[14] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Niessner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "text", + "content": "[15] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 295, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 337 + ], + "type": "text", + "content": "[16] Vishnu Sashank Dorbala, James F Mullen Jr, and Dinesh Manocha. Can an embodied agent find your \"cat-shaped mug\"? llm-based zero-shot object navigation. IEEE Robotics and Automation Letters, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 338, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 371 + ], + "type": "text", + "content": "[17] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79), 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 373, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 373, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 373, + 553, + 415 + ], + "type": "text", + "content": "[18] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 416, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 554, + 470 + ], + "type": "text", + "content": "[19] Ronghang Hu, Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Kate Saenko. Learning to reason: End-to-end module networks for visual question answering. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[20] Ronghang Hu, Jacob Andreas, Trevor Darrell, and Kate Saenko. Explainable neural computation via stack neural module networks. In Proc. of the European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 516, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 516, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 553, + 547 + ], + "type": "text", + "content": "[21] C. et al. Huang. Visual language maps for robot navigation. In 2023 IEEE Int. Conf. Robot. Autom. (ICRA), pages 10608-10615, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "type": "text", + "content": "[22] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "type": "text", + "content": "[23] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7B. arXiv preprint arXiv:2310.06825, 2023. 1, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24611" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[25] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Judy Hoffman, Li Fei-Fei, C. Lawrence Zitnick, and Ross B. Girshick. Inferring and executing programs for visual reasoning. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 205 + ], + "type": "text", + "content": "[26] Mukul Khanna, Ram Ramrakhya, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. Goat-bench: A benchmark for multi-modal lifelong navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16373-16383, 2024. 2, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "type": "text", + "content": "[27] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "text", + "content": "[28] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances. arXiv preprint arXiv:2211.15876, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 374 + ], + "type": "text", + "content": "[29] Jacob Krantz, Theophile Gervet, Karmesh Yadav, Austin Wang, Chris Paxton, Roozbeh Mottaghi, Dhruv Batra, Jitendra Malik, Stefan Lee, and Devendra Singh Chaplot. Navigating to objects specified by images. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 418 + ], + "type": "text", + "content": "[30] Dingbang Li, Wenzhou Chen, and Xin Lin. Tina: Think, interaction, and action framework for zero-shot vision language navigation. arXiv preprint arXiv:2403.08833, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "type": "text", + "content": "[32] J. et al. Liang. Code as policies: Language model programs for embodied control. In 2023 IEEE Int. Conf. Robot. Autom. (ICRA), pages 9493-9500, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 655 + ], + "type": "text", + "content": "[34] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. Openeqa: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "text", + "content": "[35] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[36] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 148 + ], + "type": "text", + "content": "Alexander William Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondrus, Theophile Gervet, Vincent-Pierre Berges, John M. Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars and robots, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 152, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 216 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 219, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 283 + ], + "type": "text", + "content": "[38] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X. Chang, et al. Habitat-matterport 3D dataset (HM3D): 1000 large-scale 3D environments for embodied AI. arXiv preprint arXiv:2109.08238, 2021. 4, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "type": "text", + "content": "[39] Santhosh K. Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "text", + "content": "[40] Sonia Raychaudhuri, Tommaso Campari, Unnat Jain, Manolis Savva, and Angel X Chang. Mopa: Modular object navigation with pointgoal agents. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5763-5773, 2024. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 399, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 420 + ], + "type": "text", + "content": "[41] G. et al. Sarch. Open-ended instructable embodied agents with memory-augmented llms. arXiv:2310.15127, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "text", + "content": "[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks, 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "text", + "content": "[43] Manolis Savva, Angel X. Chang, Alexey Dosovitskiy, Thomas Funkhouser, and Vladlen Koltun. Minos: Multimodal indoor simulator for navigation in complex environments. arXiv preprint arXiv:1712.03931, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 501, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 564 + ], + "type": "text", + "content": "[44] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied AI research. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "type": "text", + "content": "[45] Bokui Shen, Fei Xia, Chengshu Li, Roberto Martín-Martín, Linxi Fan, Guanzhi Wang, Claudia Pérez-D'Arpino, Shyamal Buch, Sanjana Srivastava, Lyne Tchapmi, et al. iGibson 1.0: a simulation environment for interactive tasks in large realistic scenes. In Proc. of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[46] C.H. et al. Song. Llm-planner: Few-shot grounded planning for embodied agents. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), pages 2998-3009, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[47] Dídac Surís, Sachit Menon, and Carl Vondrick. ViperGPT: Visual Inference via Python Execution for Reasoning. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24612" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 642 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[48] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[49] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[50] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. DD-PPO: Learning near-perfect pointgoal navigators from 2.5 billion frames. In Proc. of the International Conference on Learning Representations (ICLR), 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "type": "text", + "content": "[51] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 308 + ], + "type": "text", + "content": "[52] Zhenyu Wu, Ziwei Wang, Xiuwei Xu, Jiwen Lu, and Haibin Yan. Embodied task planning with large language models. arXiv preprint arXiv:2307.01848, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "text", + "content": "[53] Joel Ye, Dhruv Batra, Erik Wijmans, and Abhishek Das. Auxiliary Tasks Speed Up Learning PointGoal Navigation. In Proc. of the International Conference on Robot Learning (CoRL), 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "type": "text", + "content": "[54] Joel Ye, Dhruv Batra, Abhishek Das, and Erik Wijmans. Auxiliary Tasks and Exploration Enable ObjectNav. In Proc. of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 453 + ], + "type": "text", + "content": "[55] Naoki Yokoyama, Sehoon Ha, Dhruv Batra, Jiuguang Wang, and Bernadette Bucher. Vlfm: Vision-language frontier maps for zero-shot semantic navigation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 42-48. IEEE, 2024. 2, 3, 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 498 + ], + "type": "text", + "content": "[56] Naoki Yokoyama, Ram Ramrakhya, Abhishek Das, Dhruv Batra, and Sehoon Ha. Hm3d-ovon: A dataset and benchmark for open-vocabulary object goal navigation. arXiv preprint arXiv:2409.14296, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 543 + ], + "type": "text", + "content": "[57] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target navigation. in 2023 IEEE. In RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3554-3560. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 587 + ], + "type": "text", + "content": "[58] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 588, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 294, + 642 + ], + "type": "text", + "content": "[59] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 2, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24613" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_content_list.json b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b21f84448bb90cf119b761cb0175c7de0337133a --- /dev/null +++ b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_content_list.json @@ -0,0 +1,1606 @@ +[ + { + "type": "text", + "text": "TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models", + "text_level": 1, + "bbox": [ + 148, + 128, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Wang $^{1}$ , Kai Chen $^{1}$ , Jiaming Zhang $^{2}$ , Jingjing Chen $^{1}$ , Xingjun Ma $^{1*}$", + "bbox": [ + 218, + 202, + 784, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University $^{2}$ Hong Kong University of Science and Technology", + "bbox": [ + 187, + 227, + 810, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 297, + 326, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large pre-trained Vision-Language Models (VLMs) such as CLIP have demonstrated excellent zero-shot generalizability across various downstream tasks. However, recent studies have shown that the inference performance of CLIP can be greatly degraded by small adversarial perturbations, especially its visual modality, posing significant safety threats. To mitigate this vulnerability, in this paper, we propose a novel defense method called Test-Time Adversarial Prompt Tuning (TAPT) to enhance the inference robustness of CLIP against visual adversarial attacks. TAPT is a test-time defense method that learns defensive bimodal (textual and visual) prompts to robustify the inference process of CLIP. Specifically, it is an unsupervised method that optimizes the defensive prompts for each test sample by minimizing a multi-view entropy and aligning adversarial-clean distributions. We evaluate the effectiveness of TAPT on 11 benchmark datasets, including ImageNet and 10 other zero-shot datasets, demonstrating that it enhances the zero-shot adversarial robustness of the original CLIP by at least $48.9\\%$ against AutoAttack (AA), while largely maintaining performance on clean examples. Moreover, TAPT outperforms existing adversarial prompt tuning methods across various backbones, achieving an average robustness improvement of at least $36.6\\%$ . Code is available at https://github.com/xinwong/TAPT.", + "bbox": [ + 88, + 330, + 485, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 739, + 220, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) pre-trained on large-scale datasets of image-text pairs have emerged as powerful backbones for numerous applications, including computer vision [17, 34, 57], medical image analysis [15, 51], and robotics [3, 18, 39]. Despite these advancements, studies indicate that even minor adversarial perturbations in input images can significantly degrade the inference performance", + "bbox": [ + 89, + 767, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91b3058d60f0e678bbcb9234a9d0aa5a0096983e27997f8c47566b26fb78aa27.jpg", + "image_caption": [ + "Figure 1. Inference with different prompts. Top: inference with hand-crafted prompts fails to recognize the class 'cat'; Middle: Inference with fixed prompts tuned by APT methods cannot recognize all adversarial images; Bottom: Inference with test-time prompts optimized for each image produces accurate recognitions." + ], + "image_footnote": [], + "bbox": [ + 514, + 300, + 589, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff5ac9bcadb07e84dd9ecfd95ba3b4b02aa49f639f2b5544d97519878016feb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 300, + 699, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59d3b88b3eea202b9110e1f44829a5b35a28a2637a9a93dbba153e036e11800b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 300, + 790, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba39176cbb1ab492474f747efc647abce1341fc20270c8a38d681e6e2d193939.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 799, + 300, + 900, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of VLMs [9, 26, 44, 59, 63], thereby posing critical safety risks across a wide range of downstream applications.", + "bbox": [ + 511, + 460, + 905, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial training [12, 26, 52, 58] is a general defense strategy that augments training data with adversarial examples crafted to mislead the model. While this method has proven effective [8], it entails costly min-max training, which restricts its practicality—especially for large VLMs, where standard training alone can cost millions of dollars [38]. To address the efficiency limitations of adversarial training, recent research has introduced adversarial prompt tuning (APT) methods [23, 60, 67], which align learnable text prompts with adversarial image embeddings. These differentiable prompts [19, 64, 65] provide a more cost-effective alternative to adversarial training.", + "bbox": [ + 509, + 492, + 906, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its promising results, APT faces three fundamental limitations: (1) The distribution-dependent nature of learnable prompts restricts their generalization to out-of-distribution scenarios and novel tasks. (2) The requirement for task-specific annotated data presents significant challenges for zero-shot applications. (3) While APT methods improve robustness on specific tasks [42], they often compromise the model's overall generalization performance [33, 43]. Given that real-world applications involve an unbounded set of potential tasks, evaluating adversarial robustness solely on predefined downstream tasks proves insufficient. Therefore, achieving zero-shot adversarial robustness—maintaining performance against adversarial attacks on unseen tasks without task-specific training—remains an open challenge.", + "bbox": [ + 509, + 674, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: ", + "bbox": [ + 107, + 886, + 426, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19910", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key to addressing the zero-shot adversarial robustness challenge lies in adaptively identifying a robust prompt that consistently aligns an adversarial image embedding with its correct text embedding for each test sample. To this end, we propose a simple yet effective framework called Test-Time Adversarial Prompt Tuning (TAPT), which dynamically tunes a robust prompt on the fly based on only the provided test sample. Specifically, TAPT learns defensive prompts by minimizing two unsupervised losses: (1) multi-view entropy, which ensures consistent predictions across various augmented views of each test sample, and (2) adversarial-clean embedding alignment, which aligns the means and variances of test sample embeddings with precomputed adversarial-clean embeddings of a public dataset (e.g., ImageNet) to enhance inference robustness. Notably, TAPT operates during the inference phase without requiring any task-specific training set or annotations. The different inference schemes of training-time defense APT methods and our test-time defense TAPT are illustrated in Figure 1.", + "bbox": [ + 89, + 90, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluated TAPT against both white-box and black-box adversarial attacks across 11 benchmark datasets. TAPT demonstrated superior performance compared to vanilla CLIP (using hand-crafted prompts) and existing APT methods across three different prompt designs: visually-only prompts, visual-language (V-L) joint prompts, and V-L independent prompts. By dynamically adapting prompts at inference time, TAPT opens a new direction for safeguarding the inference process of pre-trained VLMs. Offering a flexible and efficient solution for zero-shot adversarial robustness while maintaining performance on clean samples, our TAPT method bridges the gap between the need for adversarial robustness and the performance challenges posed by open, real-world environments.", + "bbox": [ + 89, + 380, + 482, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are:", + "bbox": [ + 109, + 595, + 377, + 609 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel test-time defense method named Test-Time Adversarial Prompt Tuning (TAPT) to robustify the zero-shot inference process of pre-trained VLMs. TAPT adapts the prompt for each test image to achieve robust inference without the need for task-specific tuning for downstream datasets. To the best of our knowledge, TAPT is the first inference-time adversarial defense method for pre-trained VLMs.", + "- TAPT introduces an adversarial-clean alignment loss that aligns the distribution of a test sample with pre-computed adversarial-clean distributions from a public dataset (ImageNet), thereby improving adversarial robustness while maintaining accuracy on clean samples. It also leverages the advantage of APT by using pre-tuned prompts on ImageNet to further enhance zero-shot adversarial robustness.", + "- We conduct extensive experiments on 11 datasets, including ImageNet and 10 other zero-shot datasets. Our results demonstrate that TAPT significantly outperforms existing APT baselines against both white-box and black-box at" + ], + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tacks. Specifically, TAPT improves zero-shot adversarial robustness against AutoAttack by $36.6\\%$ with ViT-B/16, and by $38.0\\%$ with ViT-B/32, respectively.", + "bbox": [ + 526, + 90, + 903, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 154, + 653, + 169 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we briefly review related works on adversarial attacks and defenses for pre-trained VLMs, and test-time adaptation techniques proposed to improve generalization.", + "bbox": [ + 511, + 181, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Attacks on Pre-trained VLMs Adversarial attacks on pre-trained VLMs are broadly categorized as white-box or black-box attacks according to the threat model. In white-box attacks, the attacker has full access to the model parameters and can directly compute adversarial gradients [25, 26, 59, 68]. Black-box attacks, on the other hand, restrict the attacker to querying model outputs [10, 13, 24, 47, 54, 55, 62, 63]. Traditional single-modal attacks designed for vision models, such as PGD [26], DI [54], and AutoAttack [8], can be used directly to attack the image encoders of pre-trained VLMs. Several recent multi-modal attacks have targeted pre-trained VLMs by simultaneously exploiting vulnerabilities in both their image and text encoders. For example, Co-Attack [59] pioneered white-box multi-modal attacks, perturbing both image and text modalities concurrently. SGA [24] extended Co-Attack to the black-box setting, improving the transferability of multimodal adversarial examples. VLATTACK [55] further generates adversarial examples by fusing perturbations of images and texts from both single-modal and multi-modal.", + "bbox": [ + 511, + 231, + 903, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Defenses for Pre-trained VLMs Adversarial training/tuning is a widely used defense strategy for pretrained VLMs, with existing methods generally classified into adversarial contrastive tuning [28, 36, 49, 50, 52, 66] and adversarial prompt tuning [16, 23, 60, 67]. Adversarial contrastive tuning focuses on enhancing the adversarial robustness of the backbone model. For example, TeCoA [28] examines the impact of fine-tuning and visual prompt tuning on the zero-shot adversarial robustness of VLMs. FARE [36] improves CLIP's image encoder through unsupervised adversarial fine-tuning, enhancing adversarial robustness in models like LLaVA and OpenFlamingo without retraining. PMG-AFT [49] introduces an auxiliary branch to boost zero-shot adversarial robustness, while MMCoA [66] investigates VLM vulnerabilities to multimodal attacks. In contrast, AdvPT [60] and APT [23] offer an efficient approach to bolster the adversarial robustness of VLMs by tuning only the textual prompts without altering the model parameters. FAP [67] further refines APT by balancing cross-modal consistency between benign and adversarial inputs. While these methods are all training-time defense methods that need to pre-tune the prompt for a specific downstream task, in this work we propose a novel test-time defense method that optimizes the prompt on the", + "bbox": [ + 511, + 537, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19911", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ba494573aa9aef5ac1b53da337f113f33ce39e7dbe90673148520ae3bc4cd761.jpg", + "image_caption": [ + "(a) CLIP" + ], + "image_footnote": [], + "bbox": [ + 98, + 112, + 279, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8205d7f9b11531cbd2b06517e6ea99556776a2d3fc46bb4a821713d3db8b15cf.jpg", + "image_caption": [ + "(b) Visual Prompt" + ], + "image_footnote": [], + "bbox": [ + 300, + 114, + 483, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/850ff2d4bfd8ce3a6f0f49d728b2c3695fc8dee003c103d749a807da9917b52d.jpg", + "image_caption": [ + "(c) V-L Joint Prompt", + "Figure 2. An illustration of CLIP and different adversarial prompt tuning schemes. (a) The original CLIP [34]; (b) - (d) Adversarial prompt tuning with three distinctive prompt designs: Visual Prompt (b), V-L Joint Prompt (c), and V-L Independent Prompt (d)." + ], + "image_footnote": [], + "bbox": [ + 511, + 114, + 691, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef5799173a7b574bc2d110861ddf16496a380b893c6efc2c819ac86cf3176a1e.jpg", + "image_caption": [ + "(d) V-L Independent Prompt" + ], + "image_footnote": [], + "bbox": [ + 720, + 113, + 901, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fly during inference and is task-agnostic.", + "bbox": [ + 89, + 320, + 359, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Test-Time Adaptation Test-time adaptation (TTA) methods enhance the generalization performance of pre-trained models by adapting to individual test samples or batches, addressing distribution shifts between training (source) and testing (target) data. Early TTA methods [29, 37] address domain shift by updating batch normalization statistics based on test batch statistics. Building on this, TENT [46] improves adaptation by updating batch normalization layers to minimize the entropy of predicted probabilities for each test batch, while MEMO [61] extends this by minimizing entropy over multiple augmented input samples. Inspired by TENT, subsequent TTA methods such as CoTTA [48] and EATA [31] enable pre-trained models to adapt to continuously shifting test distributions. More recently, methods like TPT [40], PromptAlign [1], and MTA [56] have focused on tuning prompts exclusively at test time to ensure consistent predictions across various augmented views of a test sample. Kim et al. [20] finds that combining TTA with AGL [4] further improves out-of-distribution (OOD) accuracy prediction. In contrast to existing TTA methods which primarily emphasize performance on clean samples, in this work we introduce TAPT, a test-time defense, that specifically enhances zero-shot adversarial robustness against potential attacks.", + "bbox": [ + 91, + 339, + 483, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 715, + 263, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 741, + 230, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Threat Model We assume a white-box threat model in which the adversary has full knowledge of the target model's architecture and parameters, allowing direct perturbation of test images based on adversarial gradients prior to inference. The defender, or model owner, can deploy any defense strategies to protect against potential adversarial attacks. Specifically, we focus on securing CLIP zero-shot inference, where the defender lacks access to task-specific training data or annotations of the downstream application.", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP We denote the CLIP image encoder as $\\mathcal{I}$ , parameterized by $\\theta_{\\mathcal{I}}$ , and the text encoder as $\\mathcal{T}$ , parameterized by $\\theta_{\\mathcal{T}}$ . Considering a $K$ -class classification problem, where each image $\\mathbf{x}$ is associated with a class label in the format \"a photo of a \" For a clean sample $\\mathbf{x} \\in [0,1]^d$ and a target CLIP model comprising $\\{\\mathcal{I},\\mathcal{T}\\}$ , a white-box adversarial attack seeks to generate an adversarial example $\\mathbf{x}'$ that maximizes the model loss as follows:", + "bbox": [ + 511, + 319, + 906, + 441 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} = \\underset {\\| \\mathbf {x} ^ {\\prime} - \\mathbf {x} \\| _ {\\infty} \\leq \\epsilon} {\\arg \\max } \\mathcal {L} (\\mathcal {I} (\\mathbf {x} ^ {\\prime}), \\mathcal {T} (y)), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 460, + 903, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\cdot)$ represents the loss function, $\\mathbf{x}'$ is the adversarial example, and $\\epsilon$ denotes the perturbation budget.", + "bbox": [ + 511, + 500, + 905, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Prompt Tuning (APT) APT applies adversarial training during the prompt tuning process to enhance adversarial robustness, with recent APT methods primarily developed to defend the visual component of CLIP. Rather than relying on hand-crafted prompts like \"a photo of a \", APT learns robust prompts from training data to improve adversarial robustness on downstream tasks. As shown in Figure 2, APT can be extended to three distinctive prompt designs: (1) Visual-only (prompts applied solely to the vision branch), (2) Vision-Language (V-L) joint (shared prompts across both branches), and (3) V-L independent (separate prompts for each branch). Specifically, APT optimizes the prompts $\\pmb{P} = \\{P_v, P_t\\} \\in \\mathbb{R}^{L \\times D}$ , where $P_v$ and $P_t$ are the visual and textual prompts, respectively, $L$ denotes the number of prompt tokens, and $D$ is the embedding dimension.", + "bbox": [ + 511, + 536, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a downstream training dataset $\\mathcal{D}_{\\mathrm{train}} = \\{(\\mathbf{x},y)\\}$ , the learnable visual prompt $P_{v}$ is appended to the visual input tokens, forming the sequence $\\{\\mathbf{x},P_{v}\\} = \\{\\mathrm{CLS},e_1,e_2,\\dots ,e_M,P_v\\}$ . Similarly, the textual prompt $P_{t}$ is appended to the text input, forming $\\{y,P_t\\}$ . APT enhances adversarial robustness by aligning the embeddings of clean text with those of adversarial images through a min-max optimization. The corresponding optimization prob", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19912", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4cd60936454997b72dce9db9eef39a5935ded2dd1cdf3effd38679d7ee717cde.jpg", + "image_caption": [ + "Figure 3. An overview of our proposed TAPT method: Given an adversarial image, TAPT generates multiple augmented views of the image and retains only those views with low entropy in their averaged prediction probabilities. During inference, TAPT then optimizes the prompt by minimizing multi-view entropy across these selected views while aligning their embedding distribution with pre-computed adversarial-clean statistics from a public dataset (ImageNet)." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 903, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Test-Time Adversarial Prompt Tuning", + "bbox": [ + 91, + 316, + 431, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: Input: input image $\\mathbf{x}$ , image encoder $\\mathcal{I}$ , text encoder $\\mathcal{T}$ , augmentation function $\\mathcal{A}$ , entropy threshold $\\tau$ , precomputed $\\mathcal{D}_{\\mathrm{public}}$ statistics $\\{\\mu_{\\mathrm{adv}}, \\sigma_{\\mathrm{adv}}, \\mu_{\\mathrm{clean}}, \\sigma_{\\mathrm{clean}}\\}$", + "bbox": [ + 98, + 335, + 483, + 383 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2: Output: learnable prompts $P$", + "3: 1. Initialize adversarial prompts:", + "4: $\\pmb{P} \\gets \\mathrm{APT}(\\mathcal{D}_{\\mathrm{public}}, \\epsilon; \\mathcal{I}, \\mathcal{T})$", + "5: 2. Multi-view entropy-based sample selection:", + "6: Select the top $\\tau$ entropy from $\\mathcal{A}(\\mathbf{x})$ to form $\\mathcal{H}_{\\tau}(\\mathbf{x})$", + "7: 3. Compute multi-view entropy loss:" + ], + "bbox": [ + 99, + 383, + 470, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "8: $\\mathcal{L}_{\\mathrm{entropy}} = -\\sum_{i=1}^{K}\\tilde{p}(y_i|\\mathcal{H}_\\tau(\\mathbf{x}),\\boldsymbol{P})\\log \\tilde{p}(y_i|\\mathcal{H}_\\tau(\\mathbf{x}),\\boldsymbol{P})$", + "bbox": [ + 99, + 472, + 465, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "9: 4. Compute current embedding statistics:", + "bbox": [ + 99, + 503, + 411, + 516 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "10: for each layer $l$ in $\\mathcal{I}$ do", + "11: $\\mu_{l} = \\mathrm{mean}(I_{l}(\\hat{\\mathbf{x}},\\pmb {P}))$ for $\\hat{\\mathbf{x}}$ in $\\mathcal{H}_{\\tau}(\\mathbf{x})$", + "12: $\\sigma_l^2 = \\mathrm{std}(\\mathcal{I}_l(\\hat{\\mathbf{x}},\\pmb {P}))$ for $\\hat{\\mathbf{x}}$ in $\\mathcal{H}_{\\tau}(\\mathbf{x})$", + "13: end for" + ], + "bbox": [ + 96, + 517, + 398, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "14: 5. Adversarial-clean embedding alignment:", + "bbox": [ + 96, + 577, + 424, + 590 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "15: $\\mathcal{L}_{\\mathrm{adv}} = \\frac{1}{L}\\sum_{l = 1}^{L}\\left(\\| \\mu_{l} - \\mu_{\\mathrm{adv},l}\\|_{1} + \\left\\| \\sigma_{l}^{2} - \\sigma_{\\mathrm{adv},l}^{2}\\right\\|_{1}\\right)$", + "16: $\\mathcal{L}_{\\mathrm{clean}} = \\frac{1}{L}\\sum_{l = 1}^{L}\\left(\\| \\mu_{l} - \\mu_{\\mathrm{clean},l}\\|_{1} + \\left\\| \\sigma_{l}^{2} - \\sigma_{\\mathrm{clean},l}^{2}\\right\\|_{1}\\right)$", + "17: 6. Optimize prompts:", + "18: $\\mathcal{L}_{\\mathrm{TAPT}} = \\mathcal{L}_{\\mathrm{entropy}} + \\alpha \\mathcal{L}_{\\mathrm{adv}} + (1 - \\alpha)\\mathcal{L}_{\\mathrm{clean}}$", + "19: Optimize $\\pmb{P} \\gets$ Minimize $\\mathcal{L}_{\\mathrm{TAPT}}$" + ], + "bbox": [ + 96, + 590, + 464, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lem can be formulated as:", + "bbox": [ + 89, + 739, + 264, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {P}} {\\arg \\min } \\mathbb {E} _ {\\mathcal {D} _ {\\text {t r a i n}}} \\max _ {\\| \\mathbf {x} ^ {\\prime} - \\mathbf {x} \\| _ {\\infty} \\leq \\epsilon} \\mathcal {L} \\left(\\mathcal {I} \\left(\\mathbf {x} ^ {\\prime}, \\boldsymbol {P} _ {\\boldsymbol {v}}\\right), \\mathcal {T} (y, \\boldsymbol {P} _ {\\boldsymbol {t}})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 773, + 482, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{I}(\\mathbf{x}', P_v)$ and $\\mathcal{T}(y, P_t)$ denote the adversarial image embedding and text embedding, respectively. APT methods tune the prompts on the training dataset of each downstream task. At test time, the tuned prompts are fixed to perform inference for different test images.", + "bbox": [ + 89, + 823, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Test-Time Adversarial Prompt Tuning (TAPT)", + "text_level": 1, + "bbox": [ + 511, + 316, + 906, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Framework Overview As illustrated in Figure 3, TAPT comprises two main modules: (1) multi-view entropy-based sample selection, and (2) adversarial-clean embedding alignment. The defense procedure of TAPT operates as follows. Given a test image $\\mathbf{x} \\in \\mathcal{D}_{\\mathrm{test}}$ , TAPT begins by generating $M$ randomly augmented views $\\{\\mathcal{A}_1(\\mathbf{x}), \\mathcal{A}_2(\\mathbf{x}), \\dots, \\mathcal{A}_M(\\mathbf{x})\\}$ through random augmentations $\\mathcal{A}$ . The multi-view entropy-based sample selection module then chooses the top- $K$ views with the lowest entropy in their averaged prediction probabilities. Using these selected views, TAPT optimizes the prompt $P$ during inference by minimizing multi-view entropy and enforcing adversarial-clean alignment. The prompt is reset to its initial state before processing each new test sample or batch. The complete procedure of TAPT is outlined in Algorithm 1.", + "bbox": [ + 511, + 338, + 906, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-View Entropy-Based Sample Selection Following prior works [1, 40], we first discard ineffective augmented views $\\mathcal{A}(\\mathbf{x})$ (e.g., when a random crop removes essential image content) by applying a selection filter with a threshold $\\tau$ . This filter retains only augmented views with low prediction entropy (high-confidence predictions). Specifically, we define the set of selected augmented views as $\\mathcal{H}_{\\tau}(\\mathbf{x}) = \\{\\mathcal{A}_j(\\mathbf{x})|1\\leq j\\leq M,H(\\mathcal{A}_j(\\mathbf{x}))\\leq H_{\\tau}\\}$ , where $H_{\\tau}$ is the entropy threshold corresponding to the top $\\tau$ lowest entropy values among all $M$ augmentations. TAPT then optimizes prompts by minimizing the multi-view entropy of the averaged prediction probability over these selected views:", + "bbox": [ + 511, + 584, + 908, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e n t r o p y}} = - \\sum_ {i = 1} ^ {K} \\tilde {p} \\left(y _ {i} \\mid \\mathcal {H} _ {\\tau} (\\mathbf {x}), P\\right) \\log \\tilde {p} \\left(y _ {i} \\mid \\mathcal {H} _ {\\tau} (\\mathbf {x}), P\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 787, + 906, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{p}(y_i|\\mathcal{H}_{\\tau}(\\mathbf{x}),\\mathbf{P}) = \\frac{1}{|\\mathcal{H}_{\\tau}(\\mathbf{x})|}\\sum_{\\hat{\\mathbf{x}}\\in \\mathcal{H}_{\\tau}(\\mathbf{x})}p(y_i|\\hat{\\mathbf{x}},\\mathbf{P})$ denotes the average predicted probability for class $y_i$ over the selected augmented views $\\mathcal{H}_{\\tau}(\\mathbf{x})$ when using prompt $\\mathbf{P}$ . This optimization encourages the model to achieve consis", + "bbox": [ + 511, + 837, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19913", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tent predictions by adjusting the prompt specifically for the given instance.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial-Clean Embedding Alignment An adversarial test image $\\mathbf{x}$ can shift the image embeddings produced by the image encoder $\\mathcal{I}(\\mathbf{x},\\mathbf{P})$ compared to those from clean images, potentially misleading the model. To mitigate this vulnerability, we align the mean and variance of the test image embedding with pre-computed statistics from a public dataset $\\mathcal{D}_{\\mathrm{public}}$ . Ideally, this public dataset would originate from the original CLIP pre-training data. However, since this data is unavailable, we use ImageNet as a proxy, given CLIP's strong zero-shot performance on ImageNet with standard prompt tuning [5]. Specifically, we compute the mean and variance of the current embeddings for alignment as follows:", + "bbox": [ + 89, + 125, + 483, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {l} \\left(\\mathcal {H} _ {\\tau} (\\mathbf {x}); \\boldsymbol {P}\\right) = \\frac {1}{\\left| \\mathcal {H} _ {\\tau} (\\mathbf {x}) \\right|} \\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {H} _ {\\tau} (\\mathbf {x})} \\mathcal {I} _ {l} (\\hat {\\mathbf {x}}, \\boldsymbol {P}), (4) \\\\ \\sigma_ {l} ^ {2} \\left(\\mathcal {H} _ {\\tau} (\\mathbf {x}); \\boldsymbol {P}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {H} _ {\\tau} (\\mathbf {x})} \\left(\\mathcal {I} _ {l} \\left(\\hat {\\mathbf {x}} , \\boldsymbol {P}\\right) - \\mu_ {l} \\left(\\mathcal {H} _ {\\tau} (\\mathbf {x}); \\boldsymbol {P}\\right)\\right) ^ {2}}{\\left| \\mathcal {H} _ {\\tau} (\\mathbf {x}) \\right| - 1} (5) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 323, + 483, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_l(\\hat{\\mathbf{x}},P)$ represents the embedding vector at layer $l$ for the augmented input $\\hat{\\mathbf{x}}\\in \\mathcal{H}_{\\tau}(\\mathbf{x})$ given prompt $P$ . Here, $\\mu_l(\\mathcal{H}_{\\tau}(\\mathbf{x});P)$ and $\\sigma_l^2 (\\mathcal{H}_\\tau (\\mathbf{x});P)$ denote the mean and variance of the test sample embeddings at layer $l$ , respectively. We similarly pre-compute the mean and variance of embeddings from the public dataset using the robust prompt $P_{adv}$ (obtained via APT on the public data) and clean prompt $P_{clean}$ (obtained via standard prompt tuning on the public data). These offline statistics are represented by $\\mu_{\\mathrm{adv}}$ , $\\sigma_{\\mathrm{adv}}^2$ , $\\mu_{\\mathrm{clean}}$ , and $\\sigma_{\\mathrm{clean}}^2$ , respectively. We then align the mean and variance of the current embeddings with these pre-computed statistics as follows:", + "bbox": [ + 89, + 420, + 483, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a d v}} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\left(\\| \\mu_ {l} - \\mu_ {a d v, l} \\| _ {1} + \\| \\sigma_ {l} ^ {2} - \\sigma_ {a d v, l} ^ {2} \\| _ {1}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 607, + 482, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c l e a n}} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\left(\\| \\mu_ {l} - \\mu_ {\\text {c l e a n}, l} \\| _ {1} + \\| \\sigma_ {l} ^ {2} - \\sigma_ {\\text {c l e a n}, l} ^ {2} \\| _ {1}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 652, + 483, + 707 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T A P T}} = \\mathcal {L} _ {\\text {e n t r o p y}} + \\alpha \\mathcal {L} _ {\\mathrm {a d v}} + (1 - \\alpha) \\mathcal {L} _ {\\text {c l e a n}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 713, + 482, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter. Setting $\\alpha = 0$ aligns the prompt with the clean distribution, which may reduce adversarial robustness, while setting $\\alpha = 1$ aligns the prompt with the robust distribution, which may impact performance on clean samples. Our final objective combines the multiview entropy loss with the adversarial-clean alignment to optimize the prompt during inference for a given test sample. This approach enhances adversarial robustness while preserving accuracy on clean samples. Note that, to ensure inference efficiency, TAPT performs only a single step of prompt tuning for each inference.", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 114, + 702, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Models We experiment on 11 benchmark datasets (ImageNet val set and 10 other zero-shot test datasets): ImageNet [35], Caltech101 [11], DTD [7], EuroSAT [14], Pets [32], Aircraft [27], Food101 [6], Flowers [30], Cars [22], SUN397 [53], and UCF101 [41]. Our experiments focus on the CLIP model, specifically utilizing the ViT-B/16 and ViT-B/32 architectures. Following standard CLIP usage, we used hand-crafted prompts as textual inputs. For example, the prompt \"a photo of a , a type of pet\" was applied for the Pets dataset. A summary of these datasets and their corresponding hand-crafted prompts are provided in the Appendix.", + "bbox": [ + 511, + 138, + 906, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attack Configuration We evaluate the zero-shot adversarial robustness of CLIP against both white-box and black-box attacks. Specifically, we employ PGD-100 [26] for white-box attacks, DI [54] for black-box attacks, and the more powerful AutoAttack [8]. The hyperparameters for PGD-100 and DI were configured based on the TorchAttacks library [21]. Consistent with [28], we use perturbation budgets of $\\epsilon = 1/255$ , $2/255$ , and $4/255$ for both attacks.", + "bbox": [ + 511, + 325, + 906, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Defense Configuration For existing APT methods, we use the original configurations of APT methods [23, 60] and generate adversarial examples using the PGD-2 attack with step size $\\alpha = 1 / 255$ . We then develop more robust versions of APT methods with different prompt designs, including (1) adversarial visual prompt tuning (APT-V), (2) adversarial V-L joint prompt tuning (APT-VLJ), and (3) adversarial V-L independent prompt tuning (APT-VLI).", + "bbox": [ + 511, + 450, + 908, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details For our TAPT method, we initialize the defensive prompt using APT on ImageNet, training for 100 epochs with a batch size of 32 and a learning rate of 0.035. To generate augmented views for test-time finetuning, we create 63 variations of each test sample using random resized crops and horizontal flips, resulting in 64 images per sample, including the original. From these 64 images, we select the top $10\\%$ most confident predictions (with the lowest entropy) and compute the average entropy of their predicted probabilities. For adversarial-clean alignment, we pre-compute embedding statistics from the public dataset (ImageNet) using APT and standard PT, respectively. We then optimize the defensive prompts by minimizing a combined loss of multi-view entropy and adversarial-clean alignment using the AdamW optimizer, with a learning rate of $5 \\times 10^{-4}$ and an adversarial-clean scale factor of $\\alpha = 0.5$ , on a single NVIDIA A100 GPU.", + "bbox": [ + 511, + 575, + 908, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 513, + 847, + 653, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-Shot Adversarial Robustness We compare our TAPT method with existing APT methods across three", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19914", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3c34af431b95383fba014ad0e6cf2aa2622c7d502e80953bc6e0af411937298c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ImageNetCaltech101DTDEuroSATPetsAircraftFood101FlowersCarsSUN397UCF101Avg.
CLIPVanillaPGD 1.421.31.46.05.00.09.81.60.70.81.84.5
DI 6.125.88.70.311.10.59.53.03.44.94.77.1
AA 0.00.00.00.10.00.10.00.00.00.00.00.1
PGD 1.322.95.00.02.60.03.61.50.21.11.73.6
DI 6.437.211.10.79.90.17.59.33.98.65.89.1
AA 0.00.00.00.10.00.10.00.10.10.00.10.1
Visual OnlyAPT-VPGD 19.461.218.58.037.43.912.125.19.817.216.420.8
DI 29.267.523.512.347.15.021.830.218.127.122.027.6
AA 14.855.915.22.331.32.38.318.05.812.412.716.3
PGD 18.963.820.20.636.12.715.023.19.419.118.420.7
DI 26.968.622.05.146.04.823.726.513.926.424.626.2
AA 8.344.912.70.116.40.55.59.02.77.57.810.5
TAPT-VPGD 40.1 (20.7↑)69.7 (8.5↑)28.1 (9.6↑)23.9 (15.9↑)49.2 (11.8↑)11.5 (7.6↑)54.5 (42.4↑)41.1 (16.0↑)28.6 (18.8↑)40.3 (23.1↑)36.5 (20.1↑)38.5 (17.7↑)
DI 46.7 (17.5↑)75.8 (8.3↑)33.2 (9.7↑)33.0 (20.7↑)56.7 (9.6↑)12.7 (7.7↑)61.5 (39.7↑)46.2 (16.0↑)35.7 (17.6↑)45.5 (18.4↑)40.1 (18.1↑)44.3 (16.7↑)
AA 49.2 (34.4↑)75.7 (19.8↑)36.6 (21.4↑)36.9 (34.6↑)57.1 (25.8↑)19.4 (17.1↑)68.8 (60.5↑)52.9 (34.9↑)44.0 (38.2↑)48.9 (36.5↑)48.1 (35.4↑)48.9 (32.6↑)
PGD 42.2 (23.3↑)79.4 (15.6↑)32.2 (12.0↑)24.4 (23.8↑)62.4 (26.3↑)12.1 (9.4↑)53.2 (38.2↑)47.1 (24.0↑)33.5 (24.1↑)44.1 (25.0↑)44.5 (26.1↑)43.2 (22.5↑)
DI 46.5 (19.6↑)81.5 (12.9↑)33.6 (11.6↑)24.8 (19.7↑)66.0 (20.0↑)13.3 (8.5↑)57.7 (34.0↑)48.7 (22.2↑)37.5 (23.6↑)47.3 (20.9↑)47.7 (23.1↑)45.9 (19.7↑)
AA 44.1 (35.8↑)76.3 (31.4↑)33.5 (20.8↑)31.4 (31.3↑)54.9 (38.5↑)16.4 (15.9↑)53.6 (48.1↑)47.3 (38.3↑)42.4 (39.7↑)45.6 (38.1↑)45.0 (37.2↑)44.6 (34.1↑)
V-L JointAPT-VLJPGD 23.961.718.79.741.43.214.123.412.217.715.522.0
DI 34.972.025.010.552.84.224.632.918.928.524.829.9
AA 16.553.212.65.631.31.78.016.04.911.111.115.6
PGD 21.464.314.710.437.72.016.521.08.917.518.021.1
DI 30.469.517.011.947.92.926.726.117.326.526.427.5
AA 10.346.210.03.018.50.66.79.51.77.47.211.1
TAPT-VLJPGD 50.2 (26.3↑)81.0 (19.3↑)29.5 (10.8↑)13.5 (3.8↑)68.7 (27.3↑)5.6 (2.4↑)41.7 (27.6↑)39.3 (15.9↑)28.1 (15.9↑)39.8 (22.1↑)41.5 (26.0↑)39.9 (17.9↑)
DI 51.7 (16.8↑)82.3 (10.3↑)30.4 (5.4↑)14.5 (4.0↑)69.5 (16.7↑)5.8 (1.6↑)43.2 (18.6↑)40.2 (7.3↑)30.2 (11.3↑)41.3 (12.8↑)42.2 (17.4↑)41.0 (11.1↑)
AA 52.4 (35.9↑)81.5 (28.3↑)30.2 (17.6↑)15.4 (9.8↑)69.1 (37.8↑)5.8 (4.1↑)44.8 (36.8↑)40.0 (24.0↑)31.0 (26.1↑)42.1 (31.0↑)44.3 (33.2↑)41.5 (25.9↑)
PGD 52.1 (30.7↑)80.6 (16.3↑)27.1 (12.4↑)13.4 (3.0↑)72.0 (34.3↑)7.6 (5.6↑)52.5 (36.0↑)39.9 (18.0↑)32.3 (23.4↑)44.6 (27.1↑)45.3 (27.1↑)42.4 (21.3↑)
DI 52.4 (22.0↑)80.5 (11.0↑)27.1 (10.1↑)13.6 (1.7↑)71.6 (23.7↑)7.7 (4.8↑)52.0 (25.3↑)40.0 (13.9↑)33.4 (16.1↑)44.6 (18.1↑)45.5 (19.1↑)42.6 (15.1↑)
AA 52.4 (42.1↑)78.6 (32.4↑)27.4 (17.4↑)13.1 (10.1↑)70.0 (51.5↑)8.6 (8.0↑)53.0 (46.3↑)40.5 (31.0↑)34.7 (33.0↑)44.9 (37.5↑)44.8 (37.6↑)42.5 (31.4↑)
V-L IndependentAPT-VLJPGD 24.365.318.910.043.63.114.323.610.518.217.422.7
DI 35.167.719.310.649.84.323.627.716.226.421.527.5
AA 17.257.114.48.235.31.58.916.65.111.912.517.2
PGD 21.263.215.810.437.61.516.120.28.316.917.620.8
DI 28.969.520.310.947.02.824.824.314.823.823.626.4
AA 9.745.610.66.817.30.46.08.72.06.86.911.0
TAPT-VLJPGD 50.0 (25.7↑)79.0 (13.7↑)32.4 (13.5↑)36.2 (26.2↑)67.5 (51.4↑)13.1 (10.0↑)65.7 (51.4↑)49.8 (26.2↑)39.6 (29.1↑)48.3 (30.1↑)47.5 (30.1↑)48.1 (25.4↑)
DI 53.8 (18.7↑)80.5 (12.8↑)32.3 (13.0↑)39.6 (29.0↑)69.4 (19.6↑)13.3 (9.0↑)70.6 (47.0↑)51.1 (23.4↑)42.5 (26.3↑)50.3 (23.9↑)48.2 (26.7↑)50.1 (22.6↑)
AA 55.1 (37.9↑)80.3 (23.2↑)35.7 (21.3↑)44.4 (36.2↑)70.2 (34.9↑)16.0 (14.5↑)76.2 (67.3↑)55.1 (38.5↑)50.5 (45.4↑)53.6 (41.7↑)54.5 (42.0↑)53.8 (36.6↑)
PGD 48.2 (27.0↑)82.1 (18.9↑)30.7 (14.9↑)31.6 (21.2↑)68.1 (30.5↑)4.5 (3.0↑)64.7 (48.6↑)44.6 (24.4↑)41.3 (33.0↑)47.6 (30.7↑)49.1 (31.5↑)46.6 (25.8↑)
DI 50.0 (21.1↑)82.7 (13.2↑)31.5 (11.2↑)32.7 (21.8↑)69.4 (22.4↑)4.4 (1.6↑)65.6 (40.8↑)45.1 (20.8↑)42.8 (28.0↑)49.1 (25.3↑)49.9 (26.3↑)47.6 (21.2↑)
AA 49.7 (40.0↑)80.6 (35.0↑)33.3 (22.7↑)38.1 (31.3↑)68.4 (51.1↑)5.0 (4.6↑)67.5 (61.5↑)47.1 (38.4↑)48.1 (46.1↑)50.3 (43.5↑)50.4 (43.5↑)49.0 (38.0↑)
", + "bbox": [ + 96, + 88, + 905, + 531 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c854a7c2cabd55c0dc9d31f1db562a172ba3892b596681b619112ae97d5e4692.jpg", + "table_caption": [ + "Table 1. Zero-shot adversarial robustness (\\%) of different defense methods from ImageNet to downstream datasets, evaluated against PGD [26], DI [54], and AutoAttack (AA) [8] under perturbation budget $\\epsilon = 1/255$ . The baseline APT methods (APT-V, APT-VLJ, and APT-VLI) were tuned on ImageNet under a 16-shot setting and then assessed on the other 10 datasets. The green upward arrows ( $\\uparrow$ ) highlight the performance improvement of our TAPT over the baselines." + ], + "table_footnote": [], + "table_body": "
SourceTarget
ImageNetCaltech101DTDEuroSATPetsAircraftFood101FlowersCarsSUN397UCF101Avg.
APT-V60.689.736.926.784.816.863.255.752.158.254.454.5
TAPT-V66.992.844.641.088.523.885.767.465.762.965.564.1
APT-VLJ64.088.534.816.081.18.162.653.347.053.052.851.0
TAPT-VLJ66.588.536.716.780.67.464.650.448.754.755.451.8
APT-VLI63.888.634.117.280.711.661.151.845.953.552.751.0
TAPT-VLI65.890.439.248.382.516.385.860.362.560.064.461.4
", + "bbox": [ + 96, + 613, + 482, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Zero-shot clean accuracy (\\%) of different defense methods from ImageNet to downstream datasets, including APT baselines (APT-V, APT-VLJ, and APT-VLI) and our TAPT. The backbone is ViT-B/16. The best results are boldfaced.", + "bbox": [ + 89, + 758, + 482, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prompt designs: visual-only (V), V-L joint (VLJ), and V-L independent (VLI). Table 1 presents the zero-shot adversarial robustness results under PGD-100, DI, and AutoAttack attacks, with the 'Vanilla' showing results without any", + "bbox": [ + 89, + 839, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "defense. It is clear that both white-box and black-box attacks can drastically reduce accuracy (nearly $0\\%$ ) in the absence of defenses, using only imperceptible noise with $\\epsilon = 1/255$ . AutoAttack emerges as the most effective attack, achieving a nearly $100\\%$ attack success rate on average across all datasets. For defense, our TAPT method consistently achieves the best average performance under AutoAttack across all prompt designs (visual-only, V-L joint, and V-L independent) and ViT architectures (ViT-B/16 and ViT-B/32). With ViT-B/16, TAPT enhances robustness by $32.6\\%$ (visual-only), $25.9\\%$ (V-L joint), and $36.6\\%$ (V-L independent), respectively. Similar improvements are observed with ViT-B/32, where TAPT yields gains of $34.1\\%$ , $31.4\\%$ , and $38.0\\%$ for the respective prompt designs. TAPT also demonstrates superior performance against both PGD-100 and DI attacks.", + "bbox": [ + 511, + 616, + 906, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, TAPT with the V-L independent prompt design achieves the highest average zero-shot adversar", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19915", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0053048ad4208ae58b6acdec534134b644d35dd3cc5696393a3dd0230e47375.jpg", + "image_caption": [ + "(a) Average" + ], + "image_footnote": [], + "bbox": [ + 96, + 126, + 292, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc12d30225ac86133592118f2cc96980481e1a8ffd70b98a919a0afb2b5dfb76.jpg", + "image_caption": [ + "(b) ImageNet" + ], + "image_footnote": [], + "bbox": [ + 299, + 126, + 493, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb4309f20a4e9d6423f9bd32215cdf4cf7c9c389ddf7ca74d3b3e8f848e43641.jpg", + "image_caption": [ + "(c) Caltech101" + ], + "image_footnote": [], + "bbox": [ + 501, + 126, + 696, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3dbb3c3593dce09a3a5f5bba84ac34c896db565a54e78926d0f7f2a4eb568dce.jpg", + "image_caption": [ + "(d) DTD" + ], + "image_footnote": [], + "bbox": [ + 705, + 126, + 900, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29402cccd71b8c2711e7f5cb74e300bbde007063996479475683f8d5faeb4d60.jpg", + "image_caption": [ + "(e) EuroSAT" + ], + "image_footnote": [], + "bbox": [ + 96, + 263, + 292, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ee2fc34d5254e393b47cccdb77ecb0969e0788b08006a35193bbce88eaf063e.jpg", + "image_caption": [ + "(f) Pets" + ], + "image_footnote": [], + "bbox": [ + 299, + 263, + 493, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/805a0a6f130809b44e01b26abe5573e7294639eaf7d19c12d1a755eecc9c4487.jpg", + "image_caption": [ + "(g) Aircraft" + ], + "image_footnote": [], + "bbox": [ + 495, + 263, + 697, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef54c6080313b02051e842451b594de293ce24dcc633e8ae48530119009c9803.jpg", + "image_caption": [ + "(h) Food101" + ], + "image_footnote": [], + "bbox": [ + 705, + 263, + 900, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/564bf260419f0c0976f92fad5a56a618ef0cc2e62cb8dcb153cbe4c519fa2640.jpg", + "image_caption": [ + "(i) Flowers", + "Figure 4. Adversarial robustness $(\\%)$ of our TAPT method under different test-time adaptation steps (i.e., $\\{0,1,2,4\\}$ ). The results are reported against the PGD-100 attack on ViT-B/16 and ViT-B/32 architectures." + ], + "image_footnote": [ + "Visual Prompt ViT-B/16", + "Visual Prompt ViT-B/32" + ], + "bbox": [ + 96, + 402, + 292, + 531 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0423ae1f17d6566b6be988a2dfd4e87324ffe73e67d4083c77004e705c137db.jpg", + "image_caption": [ + "(j) Cars" + ], + "image_footnote": [ + "V-L Joint Prompt ViT-B/16", + "V-L Joint Prompt ViT-B/32" + ], + "bbox": [ + 299, + 402, + 493, + 531 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e77c93031090c10d8aa0a5c6781b4af1ee57e1bd3200f072035ebf4424fea127.jpg", + "image_caption": [ + "(k) SUN397" + ], + "image_footnote": [ + "V-L Independent Prompt ViT-B/16", + "V-L Independent Prompt ViT-B/32" + ], + "bbox": [ + 501, + 402, + 697, + 531 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98f2f37e69d8adef2cc1703a75e5e10947060189543fcd984ce70a9f280bb7d1.jpg", + "image_caption": [ + "(I)UCF101" + ], + "image_footnote": [], + "bbox": [ + 705, + 402, + 900, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ial robustness, surpassing both visual-only and V-L joint prompt designs. To summarize, our experiments reveal that: (1) incorporating textual prompts generally improves zero-shot adversarial robustness across various datasets; (2) the V-L joint prompt enhances robustness on the source domain (ImageNet) more effectively than other prompt designs; and (3) the V-L independent prompt design is more effective in enhancing zero-shot adversarial robustness than the V-L joint prompt design, likely due to the challenges in optimizing the interplay between visual and textual prompts.", + "bbox": [ + 88, + 614, + 485, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-Shot Clean Accuracy Table 2 presents the zero-shot clean accuracy of different defense methods from ImageNet to downstream datasets, comparing APT baselines (APT-V, APT-VLJ, and APT-VLI) with our TAPT method on the ViT-B/16 architecture. Note that the APT baselines were trained on ImageNet and subsequently tested on 10 downstream datasets, while for our TAPT, only the robust statistics were computed based on ImageNet. The", + "bbox": [ + 88, + 779, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "zero-shot clean accuracy should be compared on all 10 datasets. As shown, TAPT consistently achieves superior clean performance across all 11 datasets, demonstrating significantly stronger generalization capabilities. While a slight performance decrease is observed with TAPT's V-L joint prompt design on Pets, Aircraft, and Flowers, it remains competitive. Overall, TAPT outperforms APT across all three prompt designs (visual-only, V-L joint, and V-L independent), with average improvements of $9.6\\%$ , $0.8\\%$ , and $10.4\\%$ , respectively. These results underscore TAPT's ability to improve adversarial robustness without compromising too much clean accuracy.", + "bbox": [ + 511, + 616, + 906, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 815, + 678, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of TAPT Steps We first examine the effect of the number of test-time adaptation steps on the robust accuracy of TAPT. Figure 4 shows TAPT's robustness performance with varying steps (0, 1, 2, and 4) under the PGD-100 at", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19916", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tack. Notably, step $= 0$ represents the baseline performance without TAPT, reflecting the adversarial robustness achieved solely through APT. A clear trend emerges, showing that robust accuracy increases with additional TAPT steps across most datasets and prompt designs. While robustness gains generally stabilize after a few TAPT steps, the improvement from 0 to 1 step is consistently substantial, highlighting the effectiveness of even a single adaptation step. However, datasets vary in sensitivity to the number of TAPT steps; for example, performance stabilizes quickly on datasets like EuroSAT, whereas for datasets such as ImageNet and DTD, further improvements are observed with additional steps. The figure also illustrates that TAPT's benefits are consistent across different prompt designs. We further analyzed the per-sample time overhead of TAPT, finding that the additional inference time per image is 0.095s (visual-only), 0.166s (V-L joint), and 0.165s (V-L independent), respectively. This demonstrates that TAPT not only consistently enhances robust accuracy but also maintains a relatively low time cost.", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different Perturbation Budgets We further assess TAPT's robustness under varying attack strengths, defined by the perturbation budget $\\epsilon$ . Figure 5 displays zero-shot adversarial robustness across 11 datasets with $\\epsilon$ values of 1/255, 2/255, and 4/255, and TAPT steps of 1, 2, and 4. As expected, robust accuracy decreases as $\\epsilon$ increases, indicating stronger attacks. Nonetheless, TAPT consistently enhances robust accuracy across all datasets and $\\epsilon$ values.", + "bbox": [ + 89, + 397, + 480, + 517 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25b83c83242023549391d0dff3871670fda981f903e7c2d54cd327fe1d89fd0c.jpg", + "image_caption": [ + "Figure 5. Zero-shot adversarial robustness (y-axis) ofs TAPT under varying perturbation budgets $\\epsilon$ (1/255, 2/255, and 4/255) and TAPT steps (1, 2, and 4)." + ], + "image_footnote": [], + "bbox": [ + 94, + 523, + 478, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TAPT Reset Intervals Our TAPT method resets the prompt to its initial state before processing each test sample, ensuring that each test sample is handled independently. However, we also explored alternative strategies with varied reset intervals. As shown in Table 3, we experimented with reset intervals ranging from 1 to 32, as well as a strategy with no reset during the entire inference process (i.e., \"reset=all\"). Our findings indicate that continuous prompt adaptation without frequent resets can further improve robust accuracy, suggesting that accumulating information across multiple inputs can be beneficial. However, this con", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tinuous strategy introduces a vulnerability to potential poisoning attacks. To mitigate this risk and ensure reliable test-time defense, we recommend the per-sample reset strategy (\"reset=1\"), which prioritizes robustness against potential attacks over marginal gains from continued TAPT.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/91b0fa1d9324b279fb910e0c4f9d6559ba964d9b3c1d1dbac96d05c39eef1c26.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Reset IntervalImageNet10 Zero-Shot Datasets Avg.
reset=149.9247.85
reset=250.2047.91
reset=450.6947.91
reset=851.1447.53
reset=1651.4946.39
reset=3251.6244.50
reset=all0.483.79
", + "bbox": [ + 547, + 176, + 877, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Zero-shot adversarial robustness $(\\%)$ of TAPT with varying reset intervals on ImageNet and 10 other zero-shot datasets. \"reset $= N$ \" means the prompt is reset after every $N$ test samples.", + "bbox": [ + 511, + 297, + 903, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitation", + "text_level": 1, + "bbox": [ + 511, + 354, + 627, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a test-time defense method, TAPT has certain limitations that warrant further research. Our method primarily addresses attacks in the image modality by aligning adversarial image embeddings with pre-computed public data statistics on a public dataset. Future work could explore additional modality alignment and acceleration techniques to facilitate TAPT's deployment in industrial applications. Moreover, our current focus is limited to image recognition tasks. Extending TAPT to a broader range of tasks, such as visual reasoning and visual question answering in advanced models like GPT-4V [2] and Gemini [45], represents a promising direction for future research.", + "bbox": [ + 511, + 380, + 906, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 573, + 633, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced a novel test-time defense method, Test-Time Adversarial Prompt Tuning (TAPT), to enhance the inference robustness of pre-trained VLMs, such as CLIP. TAPT tunes defensive bimodal (textual and visual) prompts for each test sample by leveraging multi-view entropy minimization and adversarial-clean alignment, effectively safeguarding CLIP's zero-shot inference. It also utilizes pre-computed statistics from a public dataset (ImageNet) to defend a wide range of downstream tasks. Comprehensive evaluation across 11 benchmark datasets demonstrates that TAPT effectively enhances zero-shot adversarial robustness against both white-box and black-box attacks, while largely preserving clean accuracy. Compared to training-time adversarial prompt tuning (APT) methods, TAPT offers several advantages: (1) it is unsupervised, (2) enables sample-wise prompt adaptation, (3) delivers superior zero-shot adversarial robustness and clean accuracy, (4) is independent of downstream tasks, and (5) is lightweight. Future research could explore the scalability of test-time defense across different modalities.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19917", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work is supported by National Key R&D Program of China (Grant No. 2022ZD0160103) and National Natural Science Foundation of China (Grant No. 62276067). The computations in this research were performed using the CFFF platform of Fudan University.", + "bbox": [ + 89, + 90, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hussein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization. In NeurIPS, 2024. 3, 4", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. preprint arXiv:2303.08774, 2023. 8", + "[3] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. preprint arXiv:2204.01691, 2022. 1", + "[4] Christina Baek, Yiding Jiang, Aditi Raghunathan, and J Zico Kolter. Agreement-on-the-line: Predicting the performance of neural networks under distribution shift. In NeurIPS, 2022. 3", + "[5] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. preprint arXiv:2203.17274, 2022. 5", + "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5", + "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, 2014. 5", + "[8] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 1, 2, 5, 6", + "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1", + "[10] Hao Fang, Jiawei Kong, Wenbo Yu, Bin Chen, Jiawei Li, Shutao Xia, and Ke Xu. One perturbation is enough: On generating universal adversarial perturbations against vision-language pre-training models. preprint arXiv:2406.05491, 2024. 2", + "[11] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshops, 2004. 5", + "[12] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. In NeurIPS, 2020. 1", + "[13] Bangyan He, Xiaojun Jia, Siyuan Liang, Tianrui Lou, Yang Liu, and Xiaochun Cao. Sa-attack: Improving adversarial transferability of vision-language pre-training models via self-augmentation. preprint arXiv:2312.04913, 2023. 2" + ], + "bbox": [ + 93, + 212, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J-STARS, 2019. 5", + "[15] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine, 2023. 1", + "[16] Noor Hussein, Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Promptsmooth: Certifying robustness of medical vision-language models via prompt learning. In MICCAI, 2024. 2", + "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1", + "[18] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In CVPR, 2022. 1", + "[19] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 1", + "[20] Eungyeup Kim, Mingjie Sun, Christina Baek, Aditi Raghunathan, and J Zico Kolter. Test-time adaptation induces stronger accuracy and agreement-on-the-line. In NeurIPS, 2024. 3", + "[21] Hoki Kim. Torchattacks: A pytorch repository for adversarial attacks. preprint arXiv:2010.01950, 2020. 5", + "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops, 2013. 5", + "[23] Lin Li, Haoyan Guan, Jianing Qiu, and Michael Spratling. One prompt word is enough to boost adversarial robustness for pre-trained vision-language models. In CVPR, 2024. 1, 2, 5", + "[24] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In ICCV, 2023. 2", + "[25] Xingjun Ma, Linxi Jiang, Hanxun Huang, Zejia Weng, James Bailey, and Yu-Gang Jiang. Imbalanced gradients: a subtle cause of overestimated adversarial robustness. Machine Learning, 2024. 2", + "[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. *ICLR*, 2018. 1, 2, 5, 6", + "[27] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. preprint arXiv:1306.5151, 2013. 5", + "[28] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. In ICLR, 2023. 2, 5", + "[29] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. preprint arXiv:2006.10963, 2020. 3" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19918", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, 2008. 5", + "[31] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In ICML, 2022. 3", + "[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, 2012. 5", + "[33] Anibal Pedraza, Oscar Deniz, and Gloria Bueno. On the relationship between generalization and robustness to adversarial examples. Symmetry, 2021. 1", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 3", + "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5", + "[36] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust CLIP: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. In ICML, 2024. 2", + "[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. NeurIPS, 2020. 3", + "[38] Or Sharir, Barak Peleg, and Yoav Shoham. The cost of training nlp models: A concise overview. preprint arXiv:2004.08900, 2020. 1", + "[39] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, 2022. 1", + "[40] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 3, 4", + "[41] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. preprint arXiv:1212.0402, 2012. 5", + "[42] David Stutz, Matthias Hein, and Bernt Schiele. Disentangling adversarial robustness and generalization. In CVPR, 2019. 1", + "[43] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In ECCV, 2018. 1", + "[44] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2013. 1", + "[45] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv:2312.11805, 2023. 8" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 3", + "[47] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In IEEE S&P, 2024. 2", + "[48] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, 2022. 3", + "[49] Sibo Wang, Jie Zhang, Zheng Yuan, and Shiguang Shan. Pretrained model guided fine-tuning for zero-shot adversarial robustness. In CVPR, 2024. 2", + "[50] Xin Wang, Kai Chen, Xingjun Ma, Zhineng Chen, Jingjing Chen, and Yu-Gang Jiang. AdvQDet: Detecting query-based adversarial attacks with adversarial contrastive prompt tuning. In ACM MM, 2024. 2", + "[51] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In EMNLP, 2022. 1", + "[52] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In CVPR, 2024. 1, 2", + "[53] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 5", + "[54] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 2, 5, 6", + "[55] Ziyi Yin, Muchao Ye, Tianrong Zhang, Tianyu Du, Jinguo Zhu, Han Liu, Jinghui Chen, Ting Wang, and Fenglong Ma. Vlattack: Multimodal adversarial attacks on vision-language tasks via pre-trained models. In NeurIPS, 2023. 2", + "[56] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, 2024. 3", + "[57] Gengyuan Zhang, Jisen Ren, Jindong Gu, and Volker Tresp. Multi-event video-text retrieval. In ICCV, 2023. 1", + "[58] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 1", + "[59] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In ACM MM, 2022. 1, 2", + "[60] Jiaming Zhang, Xingjun Ma, Xin Wang, Lingyu Qiu, Jiaqi Wang, Yu-Gang Jiang, and Jitao Sang. Adversarial prompt tuning for vision-language models. In ECCV, 2024. 1, 2, 5", + "[61] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In NeurIPS, 2022. 3", + "[62] Peng-Fei Zhang, Zi Huang, and Guangdong Bai. Universal adversarial perturbations for vision-language pre-trained models. In ACM SIGIR, pages 862–871, 2024. 2", + "[63] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. In NeurIPS, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19919", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 1", + "[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 1", + "[66] Wanqi Zhou, Shuanghao Bai, Qibin Zhao, and Badong Chen. Revisiting the adversarial robustness of vision language models: a multimodal perspective. preprint arXiv:2404.19287, 2024. 2", + "[67] Yiwei Zhou, Xiaobo Xia, Zhiwei Lin, Bo Han, and Tongliang Liu. Few-shot adversarial prompt learning on vision-language models. In NeurIPS, 2024. 1, 2", + "[68] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic adversarial examples in multimodal contrastive learning. In ACM MM, 2023. 2" + ], + "bbox": [ + 91, + 90, + 480, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19920", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_model.json b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a36159256af9a94256b4c2744e9321f10ba327b4 --- /dev/null +++ b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.13, + 0.85, + 0.177 + ], + "angle": 0, + "content": "TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.203, + 0.785, + 0.224 + ], + "angle": 0, + "content": "Xin Wang\\(^{1}\\), Kai Chen\\(^{1}\\), Jiaming Zhang\\(^{2}\\), Jingjing Chen\\(^{1}\\), Xingjun Ma\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.228, + 0.812, + 0.266 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University \\(^{2}\\)Hong Kong University of Science and Technology" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.299, + 0.327, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.332, + 0.486, + 0.71 + ], + "angle": 0, + "content": "Large pre-trained Vision-Language Models (VLMs) such as CLIP have demonstrated excellent zero-shot generalizability across various downstream tasks. However, recent studies have shown that the inference performance of CLIP can be greatly degraded by small adversarial perturbations, especially its visual modality, posing significant safety threats. To mitigate this vulnerability, in this paper, we propose a novel defense method called Test-Time Adversarial Prompt Tuning (TAPT) to enhance the inference robustness of CLIP against visual adversarial attacks. TAPT is a test-time defense method that learns defensive bimodal (textual and visual) prompts to robustify the inference process of CLIP. Specifically, it is an unsupervised method that optimizes the defensive prompts for each test sample by minimizing a multi-view entropy and aligning adversarial-clean distributions. We evaluate the effectiveness of TAPT on 11 benchmark datasets, including ImageNet and 10 other zero-shot datasets, demonstrating that it enhances the zero-shot adversarial robustness of the original CLIP by at least \\(48.9\\%\\) against AutoAttack (AA), while largely maintaining performance on clean examples. Moreover, TAPT outperforms existing adversarial prompt tuning methods across various backbones, achieving an average robustness improvement of at least \\(36.6\\%\\). Code is available at https://github.com/xinwong/TAPT." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.222, + 0.758 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) pre-trained on large-scale datasets of image-text pairs have emerged as powerful backbones for numerous applications, including computer vision [17, 34, 57], medical image analysis [15, 51], and robotics [3, 18, 39]. Despite these advancements, studies indicate that even minor adversarial perturbations in input images can significantly degrade the inference performance" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.301, + 0.59, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.301, + 0.7, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.301, + 0.792, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.301, + 0.901, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.437 + ], + "angle": 0, + "content": "Figure 1. Inference with different prompts. Top: inference with hand-crafted prompts fails to recognize the class 'cat'; Middle: Inference with fixed prompts tuned by APT methods cannot recognize all adversarial images; Bottom: Inference with test-time prompts optimized for each image produces accurate recognitions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.493 + ], + "angle": 0, + "content": "of VLMs [9, 26, 44, 59, 63], thereby posing critical safety risks across a wide range of downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.493, + 0.907, + 0.674 + ], + "angle": 0, + "content": "Adversarial training [12, 26, 52, 58] is a general defense strategy that augments training data with adversarial examples crafted to mislead the model. While this method has proven effective [8], it entails costly min-max training, which restricts its practicality—especially for large VLMs, where standard training alone can cost millions of dollars [38]. To address the efficiency limitations of adversarial training, recent research has introduced adversarial prompt tuning (APT) methods [23, 60, 67], which align learnable text prompts with adversarial image embeddings. These differentiable prompts [19, 64, 65] provide a more cost-effective alternative to adversarial training." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Despite its promising results, APT faces three fundamental limitations: (1) The distribution-dependent nature of learnable prompts restricts their generalization to out-of-distribution scenarios and novel tasks. (2) The requirement for task-specific annotated data presents significant challenges for zero-shot applications. (3) While APT methods improve robustness on specific tasks [42], they often compromise the model's overall generalization performance [33, 43]. Given that real-world applications involve an unbounded set of potential tasks, evaluating adversarial robustness solely on predefined downstream tasks proves insufficient. Therefore, achieving zero-shot adversarial robustness—maintaining performance against adversarial attacks on unseen tasks without task-specific training—remains an open challenge." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.428, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: " + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19910" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.379 + ], + "angle": 0, + "content": "The key to addressing the zero-shot adversarial robustness challenge lies in adaptively identifying a robust prompt that consistently aligns an adversarial image embedding with its correct text embedding for each test sample. To this end, we propose a simple yet effective framework called Test-Time Adversarial Prompt Tuning (TAPT), which dynamically tunes a robust prompt on the fly based on only the provided test sample. Specifically, TAPT learns defensive prompts by minimizing two unsupervised losses: (1) multi-view entropy, which ensures consistent predictions across various augmented views of each test sample, and (2) adversarial-clean embedding alignment, which aligns the means and variances of test sample embeddings with precomputed adversarial-clean embeddings of a public dataset (e.g., ImageNet) to enhance inference robustness. Notably, TAPT operates during the inference phase without requiring any task-specific training set or annotations. The different inference schemes of training-time defense APT methods and our test-time defense TAPT are illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.593 + ], + "angle": 0, + "content": "We evaluated TAPT against both white-box and black-box adversarial attacks across 11 benchmark datasets. TAPT demonstrated superior performance compared to vanilla CLIP (using hand-crafted prompts) and existing APT methods across three different prompt designs: visually-only prompts, visual-language (V-L) joint prompts, and V-L independent prompts. By dynamically adapting prompts at inference time, TAPT opens a new direction for safeguarding the inference process of pre-trained VLMs. Offering a flexible and efficient solution for zero-shot adversarial robustness while maintaining performance on clean samples, our TAPT method bridges the gap between the need for adversarial robustness and the performance challenges posed by open, real-world environments." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.597, + 0.378, + 0.611 + ], + "angle": 0, + "content": "In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.614, + 0.483, + 0.735 + ], + "angle": 0, + "content": "- We propose a novel test-time defense method named Test-Time Adversarial Prompt Tuning (TAPT) to robustify the zero-shot inference process of pre-trained VLMs. TAPT adapts the prompt for each test image to achieve robust inference without the need for task-specific tuning for downstream datasets. To the best of our knowledge, TAPT is the first inference-time adversarial defense method for pre-trained VLMs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.736, + 0.483, + 0.84 + ], + "angle": 0, + "content": "- TAPT introduces an adversarial-clean alignment loss that aligns the distribution of a test sample with pre-computed adversarial-clean distributions from a public dataset (ImageNet), thereby improving adversarial robustness while maintaining accuracy on clean samples. It also leverages the advantage of APT by using pre-tuned prompts on ImageNet to further enhance zero-shot adversarial robustness." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We conduct extensive experiments on 11 datasets, including ImageNet and 10 other zero-shot datasets. Our results demonstrate that TAPT significantly outperforms existing APT baselines against both white-box and black-box at" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "tacks. Specifically, TAPT improves zero-shot adversarial robustness against AutoAttack by \\(36.6\\%\\) with ViT-B/16, and by \\(38.0\\%\\) with ViT-B/32, respectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.155, + 0.655, + 0.17 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.182, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Here, we briefly review related works on adversarial attacks and defenses for pre-trained VLMs, and test-time adaptation techniques proposed to improve generalization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.534 + ], + "angle": 0, + "content": "Adversarial Attacks on Pre-trained VLMs Adversarial attacks on pre-trained VLMs are broadly categorized as white-box or black-box attacks according to the threat model. In white-box attacks, the attacker has full access to the model parameters and can directly compute adversarial gradients [25, 26, 59, 68]. Black-box attacks, on the other hand, restrict the attacker to querying model outputs [10, 13, 24, 47, 54, 55, 62, 63]. Traditional single-modal attacks designed for vision models, such as PGD [26], DI [54], and AutoAttack [8], can be used directly to attack the image encoders of pre-trained VLMs. Several recent multi-modal attacks have targeted pre-trained VLMs by simultaneously exploiting vulnerabilities in both their image and text encoders. For example, Co-Attack [59] pioneered white-box multi-modal attacks, perturbing both image and text modalities concurrently. SGA [24] extended Co-Attack to the black-box setting, improving the transferability of multimodal adversarial examples. VLATTACK [55] further generates adversarial examples by fusing perturbations of images and texts from both single-modal and multi-modal." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.539, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Adversarial Defenses for Pre-trained VLMs Adversarial training/tuning is a widely used defense strategy for pretrained VLMs, with existing methods generally classified into adversarial contrastive tuning [28, 36, 49, 50, 52, 66] and adversarial prompt tuning [16, 23, 60, 67]. Adversarial contrastive tuning focuses on enhancing the adversarial robustness of the backbone model. For example, TeCoA [28] examines the impact of fine-tuning and visual prompt tuning on the zero-shot adversarial robustness of VLMs. FARE [36] improves CLIP's image encoder through unsupervised adversarial fine-tuning, enhancing adversarial robustness in models like LLaVA and OpenFlamingo without retraining. PMG-AFT [49] introduces an auxiliary branch to boost zero-shot adversarial robustness, while MMCoA [66] investigates VLM vulnerabilities to multimodal attacks. In contrast, AdvPT [60] and APT [23] offer an efficient approach to bolster the adversarial robustness of VLMs by tuning only the textual prompts without altering the model parameters. FAP [67] further refines APT by balancing cross-modal consistency between benign and adversarial inputs. While these methods are all training-time defense methods that need to pre-tune the prompt for a specific downstream task, in this work we propose a novel test-time defense method that optimizes the prompt on the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19911" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.093, + 0.213, + 0.105 + ], + "angle": 0, + "content": "(a) CLIP" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.113, + 0.28, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.093, + 0.438, + 0.107 + ], + "angle": 0, + "content": "(b) Visual Prompt" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.116, + 0.484, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.093, + 0.656, + 0.107 + ], + "angle": 0, + "content": "(c) V-L Joint Prompt" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.116, + 0.692, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.093, + 0.885, + 0.107 + ], + "angle": 0, + "content": "(d) V-L Independent Prompt" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.114, + 0.902, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 2. An illustration of CLIP and different adversarial prompt tuning schemes. (a) The original CLIP [34]; (b) - (d) Adversarial prompt tuning with three distinctive prompt designs: Visual Prompt (b), V-L Joint Prompt (c), and V-L Independent Prompt (d)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.361, + 0.336 + ], + "angle": 0, + "content": "fly during inference and is task-agnostic." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.34, + 0.485, + 0.702 + ], + "angle": 0, + "content": "Test-Time Adaptation Test-time adaptation (TTA) methods enhance the generalization performance of pre-trained models by adapting to individual test samples or batches, addressing distribution shifts between training (source) and testing (target) data. Early TTA methods [29, 37] address domain shift by updating batch normalization statistics based on test batch statistics. Building on this, TENT [46] improves adaptation by updating batch normalization layers to minimize the entropy of predicted probabilities for each test batch, while MEMO [61] extends this by minimizing entropy over multiple augmented input samples. Inspired by TENT, subsequent TTA methods such as CoTTA [48] and EATA [31] enable pre-trained models to adapt to continuously shifting test distributions. More recently, methods like TPT [40], PromptAlign [1], and MTA [56] have focused on tuning prompts exclusively at test time to ensure consistent predictions across various augmented views of a test sample. Kim et al. [20] finds that combining TTA with AGL [4] further improves out-of-distribution (OOD) accuracy prediction. In contrast to existing TTA methods which primarily emphasize performance on clean samples, in this work we introduce TAPT, a test-time defense, that specifically enhances zero-shot adversarial robustness against potential attacks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.717, + 0.264, + 0.734 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.232, + 0.757 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Threat Model We assume a white-box threat model in which the adversary has full knowledge of the target model's architecture and parameters, allowing direct perturbation of test images based on adversarial gradients prior to inference. The defender, or model owner, can deploy any defense strategies to protect against potential adversarial attacks. Specifically, we focus on securing CLIP zero-shot inference, where the defender lacks access to task-specific training data or annotations of the downstream application." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.907, + 0.442 + ], + "angle": 0, + "content": "CLIP We denote the CLIP image encoder as \\(\\mathcal{I}\\), parameterized by \\(\\theta_{\\mathcal{I}}\\), and the text encoder as \\(\\mathcal{T}\\), parameterized by \\(\\theta_{\\mathcal{T}}\\). Considering a \\(K\\)-class classification problem, where each image \\(\\mathbf{x}\\) is associated with a class label in the format \"a photo of a \" For a clean sample \\(\\mathbf{x} \\in [0,1]^d\\) and a target CLIP model comprising \\(\\{\\mathcal{I},\\mathcal{T}\\}\\), a white-box adversarial attack seeks to generate an adversarial example \\(\\mathbf{x}'\\) that maximizes the model loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.461, + 0.905, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} = \\underset {\\| \\mathbf {x} ^ {\\prime} - \\mathbf {x} \\| _ {\\infty} \\leq \\epsilon} {\\arg \\max } \\mathcal {L} (\\mathcal {I} (\\mathbf {x} ^ {\\prime}), \\mathcal {T} (y)), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.532 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\cdot)\\) represents the loss function, \\(\\mathbf{x}'\\) is the adversarial example, and \\(\\epsilon\\) denotes the perturbation budget." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Adversarial Prompt Tuning (APT) APT applies adversarial training during the prompt tuning process to enhance adversarial robustness, with recent APT methods primarily developed to defend the visual component of CLIP. Rather than relying on hand-crafted prompts like \"a photo of a \", APT learns robust prompts from training data to improve adversarial robustness on downstream tasks. As shown in Figure 2, APT can be extended to three distinctive prompt designs: (1) Visual-only (prompts applied solely to the vision branch), (2) Vision-Language (V-L) joint (shared prompts across both branches), and (3) V-L independent (separate prompts for each branch). Specifically, APT optimizes the prompts \\( \\pmb{P} = \\{P_v, P_t\\} \\in \\mathbb{R}^{L \\times D} \\), where \\( P_v \\) and \\( P_t \\) are the visual and textual prompts, respectively, \\( L \\) denotes the number of prompt tokens, and \\( D \\) is the embedding dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Given a downstream training dataset \\(\\mathcal{D}_{\\mathrm{train}} = \\{(\\mathbf{x},y)\\}\\), the learnable visual prompt \\(P_{v}\\) is appended to the visual input tokens, forming the sequence \\(\\{\\mathbf{x},P_{v}\\} = \\{\\mathrm{CLS},e_1,e_2,\\dots ,e_M,P_v\\}\\). Similarly, the textual prompt \\(P_{t}\\) is appended to the text input, forming \\(\\{y,P_t\\}\\). APT enhances adversarial robustness by aligning the embeddings of clean text with those of adversarial images through a min-max optimization. The corresponding optimization prob" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19912" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.905, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.244, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 3. An overview of our proposed TAPT method: Given an adversarial image, TAPT generates multiple augmented views of the image and retains only those views with low entropy in their averaged prediction probabilities. During inference, TAPT then optimizes the prompt by minimizing multi-view entropy across these selected views while aligning their embedding distribution with pre-computed adversarial-clean statistics from a public dataset (ImageNet)." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.318, + 0.433, + 0.334 + ], + "angle": 0, + "content": "Algorithm 1 Test-Time Adversarial Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.337, + 0.484, + 0.384 + ], + "angle": 0, + "content": "1: Input: input image \\(\\mathbf{x}\\), image encoder \\(\\mathcal{I}\\), text encoder \\(\\mathcal{T}\\), augmentation function \\(\\mathcal{A}\\), entropy threshold \\(\\tau\\), precomputed \\(\\mathcal{D}_{\\mathrm{public}}\\) statistics \\(\\{\\mu_{\\mathrm{adv}}, \\sigma_{\\mathrm{adv}}, \\mu_{\\mathrm{clean}}, \\sigma_{\\mathrm{clean}}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.384, + 0.321, + 0.398 + ], + "angle": 0, + "content": "2: Output: learnable prompts \\( P \\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.398, + 0.353, + 0.413 + ], + "angle": 0, + "content": "3: 1. Initialize adversarial prompts:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.413, + 0.316, + 0.429 + ], + "angle": 0, + "content": "4: \\(\\pmb{P} \\gets \\mathrm{APT}(\\mathcal{D}_{\\mathrm{public}}, \\epsilon; \\mathcal{I}, \\mathcal{T})\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.429, + 0.442, + 0.443 + ], + "angle": 0, + "content": "5: 2. Multi-view entropy-based sample selection:" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.443, + 0.471, + 0.459 + ], + "angle": 0, + "content": "6: Select the top \\(\\tau\\) entropy from \\(\\mathcal{A}(\\mathbf{x})\\) to form \\(\\mathcal{H}_{\\tau}(\\mathbf{x})\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.459, + 0.376, + 0.473 + ], + "angle": 0, + "content": "7: 3. Compute multi-view entropy loss:" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.384, + 0.471, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.473, + 0.466, + 0.504 + ], + "angle": 0, + "content": "8: \\(\\mathcal{L}_{\\mathrm{entropy}} = -\\sum_{i=1}^{K}\\tilde{p}(y_i|\\mathcal{H}_\\tau(\\mathbf{x}),\\boldsymbol{P})\\log \\tilde{p}(y_i|\\mathcal{H}_\\tau(\\mathbf{x}),\\boldsymbol{P})\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.504, + 0.412, + 0.517 + ], + "angle": 0, + "content": "9: 4. Compute current embedding statistics:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.518, + 0.294, + 0.532 + ], + "angle": 0, + "content": "10: for each layer \\(l\\) in \\(\\mathcal{I}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.532, + 0.399, + 0.548 + ], + "angle": 0, + "content": "11: \\(\\mu_{l} = \\mathrm{mean}(I_{l}(\\hat{\\mathbf{x}},\\pmb {P}))\\) for \\(\\hat{\\mathbf{x}}\\) in \\(\\mathcal{H}_{\\tau}(\\mathbf{x})\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.548, + 0.387, + 0.564 + ], + "angle": 0, + "content": "12: \\(\\sigma_l^2 = \\mathrm{std}(\\mathcal{I}_l(\\hat{\\mathbf{x}},\\pmb {P}))\\) for \\(\\hat{\\mathbf{x}}\\) in \\(\\mathcal{H}_{\\tau}(\\mathbf{x})\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.564, + 0.191, + 0.576 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.518, + 0.399, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.578, + 0.425, + 0.592 + ], + "angle": 0, + "content": "14: 5. Adversarial-clean embedding alignment:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.592, + 0.443, + 0.623 + ], + "angle": 0, + "content": "15: \\(\\mathcal{L}_{\\mathrm{adv}} = \\frac{1}{L}\\sum_{l = 1}^{L}\\left(\\| \\mu_{l} - \\mu_{\\mathrm{adv},l}\\|_{1} + \\left\\| \\sigma_{l}^{2} - \\sigma_{\\mathrm{adv},l}^{2}\\right\\|_{1}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.623, + 0.465, + 0.654 + ], + "angle": 0, + "content": "16: \\(\\mathcal{L}_{\\mathrm{clean}} = \\frac{1}{L}\\sum_{l = 1}^{L}\\left(\\| \\mu_{l} - \\mu_{\\mathrm{clean},l}\\|_{1} + \\left\\| \\sigma_{l}^{2} - \\sigma_{\\mathrm{clean},l}^{2}\\right\\|_{1}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.654, + 0.273, + 0.668 + ], + "angle": 0, + "content": "17: 6. Optimize prompts:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.668, + 0.414, + 0.684 + ], + "angle": 0, + "content": "18: \\(\\mathcal{L}_{\\mathrm{TAPT}} = \\mathcal{L}_{\\mathrm{entropy}} + \\alpha \\mathcal{L}_{\\mathrm{adv}} + (1 - \\alpha)\\mathcal{L}_{\\mathrm{clean}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.684, + 0.351, + 0.699 + ], + "angle": 0, + "content": "19: Optimize \\( \\pmb{P} \\gets \\) Minimize \\( \\mathcal{L}_{\\mathrm{TAPT}} \\)" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.592, + 0.465, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.74, + 0.266, + 0.754 + ], + "angle": 0, + "content": "lem can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.774, + 0.483, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {P}} {\\arg \\min } \\mathbb {E} _ {\\mathcal {D} _ {\\text {t r a i n}}} \\max _ {\\| \\mathbf {x} ^ {\\prime} - \\mathbf {x} \\| _ {\\infty} \\leq \\epsilon} \\mathcal {L} \\left(\\mathcal {I} \\left(\\mathbf {x} ^ {\\prime}, \\boldsymbol {P} _ {\\boldsymbol {v}}\\right), \\mathcal {T} (y, \\boldsymbol {P} _ {\\boldsymbol {t}})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.824, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}(\\mathbf{x}', P_v)\\) and \\(\\mathcal{T}(y, P_t)\\) denote the adversarial image embedding and text embedding, respectively. APT methods tune the prompts on the training dataset of each downstream task. At test time, the tuned prompts are fixed to perform inference for different test images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.318, + 0.907, + 0.335 + ], + "angle": 0, + "content": "3.2. Test-Time Adversarial Prompt Tuning (TAPT)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.583 + ], + "angle": 0, + "content": "Framework Overview As illustrated in Figure 3, TAPT comprises two main modules: (1) multi-view entropy-based sample selection, and (2) adversarial-clean embedding alignment. The defense procedure of TAPT operates as follows. Given a test image \\(\\mathbf{x} \\in \\mathcal{D}_{\\mathrm{test}}\\), TAPT begins by generating \\(M\\) randomly augmented views \\(\\{\\mathcal{A}_1(\\mathbf{x}), \\mathcal{A}_2(\\mathbf{x}), \\dots, \\mathcal{A}_M(\\mathbf{x})\\}\\) through random augmentations \\(\\mathcal{A}\\). The multi-view entropy-based sample selection module then chooses the top-\\(K\\) views with the lowest entropy in their averaged prediction probabilities. Using these selected views, TAPT optimizes the prompt \\(P\\) during inference by minimizing multi-view entropy and enforcing adversarial-clean alignment. The prompt is reset to its initial state before processing each new test sample or batch. The complete procedure of TAPT is outlined in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Multi-View Entropy-Based Sample Selection Following prior works [1, 40], we first discard ineffective augmented views \\(\\mathcal{A}(\\mathbf{x})\\) (e.g., when a random crop removes essential image content) by applying a selection filter with a threshold \\(\\tau\\). This filter retains only augmented views with low prediction entropy (high-confidence predictions). Specifically, we define the set of selected augmented views as \\(\\mathcal{H}_{\\tau}(\\mathbf{x}) = \\{\\mathcal{A}_j(\\mathbf{x})|1\\leq j\\leq M,H(\\mathcal{A}_j(\\mathbf{x}))\\leq H_{\\tau}\\}\\), where \\(H_{\\tau}\\) is the entropy threshold corresponding to the top \\(\\tau\\) lowest entropy values among all \\(M\\) augmentations. TAPT then optimizes prompts by minimizing the multi-view entropy of the averaged prediction probability over these selected views:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.788, + 0.907, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e n t r o p y}} = - \\sum_ {i = 1} ^ {K} \\tilde {p} \\left(y _ {i} \\mid \\mathcal {H} _ {\\tau} (\\mathbf {x}), P\\right) \\log \\tilde {p} \\left(y _ {i} \\mid \\mathcal {H} _ {\\tau} (\\mathbf {x}), P\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.838, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(\\tilde{p}(y_i|\\mathcal{H}_{\\tau}(\\mathbf{x}),\\mathbf{P}) = \\frac{1}{|\\mathcal{H}_{\\tau}(\\mathbf{x})|}\\sum_{\\hat{\\mathbf{x}}\\in \\mathcal{H}_{\\tau}(\\mathbf{x})}p(y_i|\\hat{\\mathbf{x}},\\mathbf{P})\\) denotes the average predicted probability for class \\(y_i\\) over the selected augmented views \\(\\mathcal{H}_{\\tau}(\\mathbf{x})\\) when using prompt \\(\\mathbf{P}\\). This optimization encourages the model to achieve consis" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19913" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "tent predictions by adjusting the prompt specifically for the given instance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.125, + 0.484, + 0.32 + ], + "angle": 0, + "content": "Adversarial-Clean Embedding Alignment An adversarial test image \\(\\mathbf{x}\\) can shift the image embeddings produced by the image encoder \\(\\mathcal{I}(\\mathbf{x},\\mathbf{P})\\) compared to those from clean images, potentially misleading the model. To mitigate this vulnerability, we align the mean and variance of the test image embedding with pre-computed statistics from a public dataset \\(\\mathcal{D}_{\\mathrm{public}}\\). Ideally, this public dataset would originate from the original CLIP pre-training data. However, since this data is unavailable, we use ImageNet as a proxy, given CLIP's strong zero-shot performance on ImageNet with standard prompt tuning [5]. Specifically, we compute the mean and variance of the current embeddings for alignment as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.324, + 0.484, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {l} \\left(\\mathcal {H} _ {\\tau} (\\mathbf {x}); \\boldsymbol {P}\\right) = \\frac {1}{\\left| \\mathcal {H} _ {\\tau} (\\mathbf {x}) \\right|} \\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {H} _ {\\tau} (\\mathbf {x})} \\mathcal {I} _ {l} (\\hat {\\mathbf {x}}, \\boldsymbol {P}), (4) \\\\ \\sigma_ {l} ^ {2} \\left(\\mathcal {H} _ {\\tau} (\\mathbf {x}); \\boldsymbol {P}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {H} _ {\\tau} (\\mathbf {x})} \\left(\\mathcal {I} _ {l} \\left(\\hat {\\mathbf {x}} , \\boldsymbol {P}\\right) - \\mu_ {l} \\left(\\mathcal {H} _ {\\tau} (\\mathbf {x}); \\boldsymbol {P}\\right)\\right) ^ {2}}{\\left| \\mathcal {H} _ {\\tau} (\\mathbf {x}) \\right| - 1} (5) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.484, + 0.603 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_l(\\hat{\\mathbf{x}},P)\\) represents the embedding vector at layer \\(l\\) for the augmented input \\(\\hat{\\mathbf{x}}\\in \\mathcal{H}_{\\tau}(\\mathbf{x})\\) given prompt \\(P\\). Here, \\(\\mu_l(\\mathcal{H}_{\\tau}(\\mathbf{x});P)\\) and \\(\\sigma_l^2 (\\mathcal{H}_\\tau (\\mathbf{x});P)\\) denote the mean and variance of the test sample embeddings at layer \\(l\\), respectively. We similarly pre-compute the mean and variance of embeddings from the public dataset using the robust prompt \\(P_{adv}\\) (obtained via APT on the public data) and clean prompt \\(P_{clean}\\) (obtained via standard prompt tuning on the public data). These offline statistics are represented by \\(\\mu_{\\mathrm{adv}}\\), \\(\\sigma_{\\mathrm{adv}}^2\\), \\(\\mu_{\\mathrm{clean}}\\), and \\(\\sigma_{\\mathrm{clean}}^2\\), respectively. We then align the mean and variance of the current embeddings with these pre-computed statistics as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.608, + 0.483, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a d v}} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\left(\\| \\mu_ {l} - \\mu_ {a d v, l} \\| _ {1} + \\| \\sigma_ {l} ^ {2} - \\sigma_ {a d v, l} ^ {2} \\| _ {1}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.653, + 0.484, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c l e a n}} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\left(\\| \\mu_ {l} - \\mu_ {\\text {c l e a n}, l} \\| _ {1} + \\| \\sigma_ {l} ^ {2} - \\sigma_ {\\text {c l e a n}, l} ^ {2} \\| _ {1}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.714, + 0.483, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T A P T}} = \\mathcal {L} _ {\\text {e n t r o p y}} + \\alpha \\mathcal {L} _ {\\mathrm {a d v}} + (1 - \\alpha) \\mathcal {L} _ {\\text {c l e a n}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter. Setting \\(\\alpha = 0\\) aligns the prompt with the clean distribution, which may reduce adversarial robustness, while setting \\(\\alpha = 1\\) aligns the prompt with the robust distribution, which may impact performance on clean samples. Our final objective combines the multiview entropy loss with the adversarial-clean alignment to optimize the prompt during inference for a given test sample. This approach enhances adversarial robustness while preserving accuracy on clean samples. Note that, to ensure inference efficiency, TAPT performs only a single step of prompt tuning for each inference." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.704, + 0.133 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.907, + 0.322 + ], + "angle": 0, + "content": "Datasets and Models We experiment on 11 benchmark datasets (ImageNet val set and 10 other zero-shot test datasets): ImageNet [35], Caltech101 [11], DTD [7], EuroSAT [14], Pets [32], Aircraft [27], Food101 [6], Flowers [30], Cars [22], SUN397 [53], and UCF101 [41]. Our experiments focus on the CLIP model, specifically utilizing the ViT-B/16 and ViT-B/32 architectures. Following standard CLIP usage, we used hand-crafted prompts as textual inputs. For example, the prompt \"a photo of a , a type of pet\" was applied for the Pets dataset. A summary of these datasets and their corresponding hand-crafted prompts are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.326, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Attack Configuration We evaluate the zero-shot adversarial robustness of CLIP against both white-box and black-box attacks. Specifically, we employ PGD-100 [26] for white-box attacks, DI [54] for black-box attacks, and the more powerful AutoAttack [8]. The hyperparameters for PGD-100 and DI were configured based on the TorchAttacks library [21]. Consistent with [28], we use perturbation budgets of \\(\\epsilon = 1/255\\), \\(2/255\\), and \\(4/255\\) for both attacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.909, + 0.573 + ], + "angle": 0, + "content": "Defense Configuration For existing APT methods, we use the original configurations of APT methods [23, 60] and generate adversarial examples using the PGD-2 attack with step size \\(\\alpha = 1 / 255\\). We then develop more robust versions of APT methods with different prompt designs, including (1) adversarial visual prompt tuning (APT-V), (2) adversarial V-L joint prompt tuning (APT-VLJ), and (3) adversarial V-L independent prompt tuning (APT-VLI)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.909, + 0.835 + ], + "angle": 0, + "content": "Implementation Details For our TAPT method, we initialize the defensive prompt using APT on ImageNet, training for 100 epochs with a batch size of 32 and a learning rate of 0.035. To generate augmented views for test-time finetuning, we create 63 variations of each test sample using random resized crops and horizontal flips, resulting in 64 images per sample, including the original. From these 64 images, we select the top \\(10\\%\\) most confident predictions (with the lowest entropy) and compute the average entropy of their predicted probabilities. For adversarial-clean alignment, we pre-compute embedding statistics from the public dataset (ImageNet) using APT and standard PT, respectively. We then optimize the defensive prompts by minimizing a combined loss of multi-view entropy and adversarial-clean alignment using the AdamW optimizer, with a learning rate of \\(5 \\times 10^{-4}\\) and an adversarial-clean scale factor of \\(\\alpha = 0.5\\), on a single NVIDIA A100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.848, + 0.654, + 0.862 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Zero-Shot Adversarial Robustness We compare our TAPT method with existing APT methods across three" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19914" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.906, + 0.532 + ], + "angle": 0, + "content": "
ImageNetCaltech101DTDEuroSATPetsAircraftFood101FlowersCarsSUN397UCF101Avg.
CLIPVanillaPGD 1.421.31.46.05.00.09.81.60.70.81.84.5
DI 6.125.88.70.311.10.59.53.03.44.94.77.1
AA 0.00.00.00.10.00.10.00.00.00.00.00.1
PGD 1.322.95.00.02.60.03.61.50.21.11.73.6
DI 6.437.211.10.79.90.17.59.33.98.65.89.1
AA 0.00.00.00.10.00.10.00.10.10.00.10.1
Visual OnlyAPT-VPGD 19.461.218.58.037.43.912.125.19.817.216.420.8
DI 29.267.523.512.347.15.021.830.218.127.122.027.6
AA 14.855.915.22.331.32.38.318.05.812.412.716.3
PGD 18.963.820.20.636.12.715.023.19.419.118.420.7
DI 26.968.622.05.146.04.823.726.513.926.424.626.2
AA 8.344.912.70.116.40.55.59.02.77.57.810.5
TAPT-VPGD 40.1 (20.7↑)69.7 (8.5↑)28.1 (9.6↑)23.9 (15.9↑)49.2 (11.8↑)11.5 (7.6↑)54.5 (42.4↑)41.1 (16.0↑)28.6 (18.8↑)40.3 (23.1↑)36.5 (20.1↑)38.5 (17.7↑)
DI 46.7 (17.5↑)75.8 (8.3↑)33.2 (9.7↑)33.0 (20.7↑)56.7 (9.6↑)12.7 (7.7↑)61.5 (39.7↑)46.2 (16.0↑)35.7 (17.6↑)45.5 (18.4↑)40.1 (18.1↑)44.3 (16.7↑)
AA 49.2 (34.4↑)75.7 (19.8↑)36.6 (21.4↑)36.9 (34.6↑)57.1 (25.8↑)19.4 (17.1↑)68.8 (60.5↑)52.9 (34.9↑)44.0 (38.2↑)48.9 (36.5↑)48.1 (35.4↑)48.9 (32.6↑)
PGD 42.2 (23.3↑)79.4 (15.6↑)32.2 (12.0↑)24.4 (23.8↑)62.4 (26.3↑)12.1 (9.4↑)53.2 (38.2↑)47.1 (24.0↑)33.5 (24.1↑)44.1 (25.0↑)44.5 (26.1↑)43.2 (22.5↑)
DI 46.5 (19.6↑)81.5 (12.9↑)33.6 (11.6↑)24.8 (19.7↑)66.0 (20.0↑)13.3 (8.5↑)57.7 (34.0↑)48.7 (22.2↑)37.5 (23.6↑)47.3 (20.9↑)47.7 (23.1↑)45.9 (19.7↑)
AA 44.1 (35.8↑)76.3 (31.4↑)33.5 (20.8↑)31.4 (31.3↑)54.9 (38.5↑)16.4 (15.9↑)53.6 (48.1↑)47.3 (38.3↑)42.4 (39.7↑)45.6 (38.1↑)45.0 (37.2↑)44.6 (34.1↑)
V-L JointAPT-VLJPGD 23.961.718.79.741.43.214.123.412.217.715.522.0
DI 34.972.025.010.552.84.224.632.918.928.524.829.9
AA 16.553.212.65.631.31.78.016.04.911.111.115.6
PGD 21.464.314.710.437.72.016.521.08.917.518.021.1
DI 30.469.517.011.947.92.926.726.117.326.526.427.5
AA 10.346.210.03.018.50.66.79.51.77.47.211.1
TAPT-VLJPGD 50.2 (26.3↑)81.0 (19.3↑)29.5 (10.8↑)13.5 (3.8↑)68.7 (27.3↑)5.6 (2.4↑)41.7 (27.6↑)39.3 (15.9↑)28.1 (15.9↑)39.8 (22.1↑)41.5 (26.0↑)39.9 (17.9↑)
DI 51.7 (16.8↑)82.3 (10.3↑)30.4 (5.4↑)14.5 (4.0↑)69.5 (16.7↑)5.8 (1.6↑)43.2 (18.6↑)40.2 (7.3↑)30.2 (11.3↑)41.3 (12.8↑)42.2 (17.4↑)41.0 (11.1↑)
AA 52.4 (35.9↑)81.5 (28.3↑)30.2 (17.6↑)15.4 (9.8↑)69.1 (37.8↑)5.8 (4.1↑)44.8 (36.8↑)40.0 (24.0↑)31.0 (26.1↑)42.1 (31.0↑)44.3 (33.2↑)41.5 (25.9↑)
PGD 52.1 (30.7↑)80.6 (16.3↑)27.1 (12.4↑)13.4 (3.0↑)72.0 (34.3↑)7.6 (5.6↑)52.5 (36.0↑)39.9 (18.0↑)32.3 (23.4↑)44.6 (27.1↑)45.3 (27.1↑)42.4 (21.3↑)
DI 52.4 (22.0↑)80.5 (11.0↑)27.1 (10.1↑)13.6 (1.7↑)71.6 (23.7↑)7.7 (4.8↑)52.0 (25.3↑)40.0 (13.9↑)33.4 (16.1↑)44.6 (18.1↑)45.5 (19.1↑)42.6 (15.1↑)
AA 52.4 (42.1↑)78.6 (32.4↑)27.4 (17.4↑)13.1 (10.1↑)70.0 (51.5↑)8.6 (8.0↑)53.0 (46.3↑)40.5 (31.0↑)34.7 (33.0↑)44.9 (37.5↑)44.8 (37.6↑)42.5 (31.4↑)
V-L IndependentAPT-VLJPGD 24.365.318.910.043.63.114.323.610.518.217.422.7
DI 35.167.719.310.649.84.323.627.716.226.421.527.5
AA 17.257.114.48.235.31.58.916.65.111.912.517.2
PGD 21.263.215.810.437.61.516.120.28.316.917.620.8
DI 28.969.520.310.947.02.824.824.314.823.823.626.4
AA 9.745.610.66.817.30.46.08.72.06.86.911.0
TAPT-VLJPGD 50.0 (25.7↑)79.0 (13.7↑)32.4 (13.5↑)36.2 (26.2↑)67.5 (51.4↑)13.1 (10.0↑)65.7 (51.4↑)49.8 (26.2↑)39.6 (29.1↑)48.3 (30.1↑)47.5 (30.1↑)48.1 (25.4↑)
DI 53.8 (18.7↑)80.5 (12.8↑)32.3 (13.0↑)39.6 (29.0↑)69.4 (19.6↑)13.3 (9.0↑)70.6 (47.0↑)51.1 (23.4↑)42.5 (26.3↑)50.3 (23.9↑)48.2 (26.7↑)50.1 (22.6↑)
AA 55.1 (37.9↑)80.3 (23.2↑)35.7 (21.3↑)44.4 (36.2↑)70.2 (34.9↑)16.0 (14.5↑)76.2 (67.3↑)55.1 (38.5↑)50.5 (45.4↑)53.6 (41.7↑)54.5 (42.0↑)53.8 (36.6↑)
PGD 48.2 (27.0↑)82.1 (18.9↑)30.7 (14.9↑)31.6 (21.2↑)68.1 (30.5↑)4.5 (3.0↑)64.7 (48.6↑)44.6 (24.4↑)41.3 (33.0↑)47.6 (30.7↑)49.1 (31.5↑)46.6 (25.8↑)
DI 50.0 (21.1↑)82.7 (13.2↑)31.5 (11.2↑)32.7 (21.8↑)69.4 (22.4↑)4.4 (1.6↑)65.6 (40.8↑)45.1 (20.8↑)42.8 (28.0↑)49.1 (25.3↑)49.9 (26.3↑)47.6 (21.2↑)
AA 49.7 (40.0↑)80.6 (35.0↑)33.3 (22.7↑)38.1 (31.3↑)68.4 (51.1↑)5.0 (4.6↑)67.5 (61.5↑)47.1 (38.4↑)48.1 (46.1↑)50.3 (43.5↑)50.4 (43.5↑)49.0 (38.0↑)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.544, + 0.907, + 0.601 + ], + "angle": 0, + "content": "Table 1. Zero-shot adversarial robustness (\\%) of different defense methods from ImageNet to downstream datasets, evaluated against PGD [26], DI [54], and AutoAttack (AA) [8] under perturbation budget \\(\\epsilon = 1/255\\). The baseline APT methods (APT-V, APT-VLJ, and APT-VLI) were tuned on ImageNet under a 16-shot setting and then assessed on the other 10 datasets. The green upward arrows (\\(\\uparrow\\)) highlight the performance improvement of our TAPT over the baselines." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.614, + 0.483, + 0.748 + ], + "angle": 0, + "content": "
SourceTarget
ImageNetCaltech101DTDEuroSATPetsAircraftFood101FlowersCarsSUN397UCF101Avg.
APT-V60.689.736.926.784.816.863.255.752.158.254.454.5
TAPT-V66.992.844.641.088.523.885.767.465.762.965.564.1
APT-VLJ64.088.534.816.081.18.162.653.347.053.052.851.0
TAPT-VLJ66.588.536.716.780.67.464.650.448.754.755.451.8
APT-VLI63.888.634.117.280.711.661.151.845.953.552.751.0
TAPT-VLI65.890.439.248.382.516.385.860.362.560.064.461.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Table 2. Zero-shot clean accuracy (\\%) of different defense methods from ImageNet to downstream datasets, including APT baselines (APT-V, APT-VLJ, and APT-VLI) and our TAPT. The backbone is ViT-B/16. The best results are boldfaced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.903 + ], + "angle": 0, + "content": "prompt designs: visual-only (V), V-L joint (VLJ), and V-L independent (VLI). Table 1 presents the zero-shot adversarial robustness results under PGD-100, DI, and AutoAttack attacks, with the 'Vanilla' showing results without any" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.907, + 0.858 + ], + "angle": 0, + "content": "defense. It is clear that both white-box and black-box attacks can drastically reduce accuracy (nearly \\(0\\%\\)) in the absence of defenses, using only imperceptible noise with \\(\\epsilon = 1/255\\). AutoAttack emerges as the most effective attack, achieving a nearly \\(100\\%\\) attack success rate on average across all datasets. For defense, our TAPT method consistently achieves the best average performance under AutoAttack across all prompt designs (visual-only, V-L joint, and V-L independent) and ViT architectures (ViT-B/16 and ViT-B/32). With ViT-B/16, TAPT enhances robustness by \\(32.6\\%\\) (visual-only), \\(25.9\\%\\) (V-L joint), and \\(36.6\\%\\) (V-L independent), respectively. Similar improvements are observed with ViT-B/32, where TAPT yields gains of \\(34.1\\%\\), \\(31.4\\%\\), and \\(38.0\\%\\) for the respective prompt designs. TAPT also demonstrates superior performance against both PGD-100 and DI attacks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Furthermore, TAPT with the V-L independent prompt design achieves the highest average zero-shot adversar" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19915" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.117, + 0.242, + 0.127 + ], + "angle": 0, + "content": "(a) Average" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.127, + 0.293, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.117, + 0.445, + 0.127 + ], + "angle": 0, + "content": "(b) ImageNet" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.127, + 0.495, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.117, + 0.65, + 0.127 + ], + "angle": 0, + "content": "(c) Caltech101" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.127, + 0.697, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.117, + 0.836, + 0.127 + ], + "angle": 0, + "content": "(d) DTD" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.127, + 0.901, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.256, + 0.238, + 0.265 + ], + "angle": 0, + "content": "(e) EuroSAT" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.265, + 0.293, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.256, + 0.429, + 0.265 + ], + "angle": 0, + "content": "(f) Pets" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.265, + 0.495, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.256, + 0.642, + 0.266 + ], + "angle": 0, + "content": "(g) Aircraft" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.265, + 0.698, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.256, + 0.846, + 0.266 + ], + "angle": 0, + "content": "(h) Food101" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.265, + 0.901, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.395, + 0.235, + 0.404 + ], + "angle": 0, + "content": "(i) Flowers" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.404, + 0.293, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.395, + 0.429, + 0.405 + ], + "angle": 0, + "content": "(j) Cars" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.404, + 0.495, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.395, + 0.642, + 0.404 + ], + "angle": 0, + "content": "(k) SUN397" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.404, + 0.698, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.395, + 0.843, + 0.404 + ], + "angle": 0, + "content": "(I)UCF101" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.404, + 0.901, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.261, + 0.535, + 0.391, + 0.543 + ], + "angle": 0, + "content": "Visual Prompt ViT-B/16" + }, + { + "type": "image_footnote", + "bbox": [ + 0.261, + 0.543, + 0.391, + 0.552 + ], + "angle": 0, + "content": "Visual Prompt ViT-B/32" + }, + { + "type": "image_footnote", + "bbox": [ + 0.404, + 0.535, + 0.545, + 0.543 + ], + "angle": 0, + "content": "V-L Joint Prompt ViT-B/16" + }, + { + "type": "image_footnote", + "bbox": [ + 0.404, + 0.543, + 0.545, + 0.552 + ], + "angle": 0, + "content": "V-L Joint Prompt ViT-B/32" + }, + { + "type": "image_footnote", + "bbox": [ + 0.561, + 0.535, + 0.737, + 0.543 + ], + "angle": 0, + "content": "V-L Independent Prompt ViT-B/16" + }, + { + "type": "image_footnote", + "bbox": [ + 0.561, + 0.543, + 0.737, + 0.552 + ], + "angle": 0, + "content": "V-L Independent Prompt ViT-B/32" + }, + { + "type": "list", + "bbox": [ + 0.261, + 0.535, + 0.737, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.571, + 0.907, + 0.599 + ], + "angle": 0, + "content": "Figure 4. Adversarial robustness \\((\\%)\\) of our TAPT method under different test-time adaptation steps (i.e., \\(\\{0,1,2,4\\}\\)). The results are reported against the PGD-100 attack on ViT-B/16 and ViT-B/32 architectures." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.616, + 0.486, + 0.769 + ], + "angle": 0, + "content": "ial robustness, surpassing both visual-only and V-L joint prompt designs. To summarize, our experiments reveal that: (1) incorporating textual prompts generally improves zero-shot adversarial robustness across various datasets; (2) the V-L joint prompt enhances robustness on the source domain (ImageNet) more effectively than other prompt designs; and (3) the V-L independent prompt design is more effective in enhancing zero-shot adversarial robustness than the V-L joint prompt design, likely due to the challenges in optimizing the interplay between visual and textual prompts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Zero-Shot Clean Accuracy Table 2 presents the zero-shot clean accuracy of different defense methods from ImageNet to downstream datasets, comparing APT baselines (APT-V, APT-VLJ, and APT-VLI) with our TAPT method on the ViT-B/16 architecture. Note that the APT baselines were trained on ImageNet and subsequently tested on 10 downstream datasets, while for our TAPT, only the robust statistics were computed based on ImageNet. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.907, + 0.799 + ], + "angle": 0, + "content": "zero-shot clean accuracy should be compared on all 10 datasets. As shown, TAPT consistently achieves superior clean performance across all 11 datasets, demonstrating significantly stronger generalization capabilities. While a slight performance decrease is observed with TAPT's V-L joint prompt design on Pets, Aircraft, and Flowers, it remains competitive. Overall, TAPT outperforms APT across all three prompt designs (visual-only, V-L joint, and V-L independent), with average improvements of \\(9.6\\%\\), \\(0.8\\%\\), and \\(10.4\\%\\), respectively. These results underscore TAPT's ability to improve adversarial robustness without compromising too much clean accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.679, + 0.83 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Number of TAPT Steps We first examine the effect of the number of test-time adaptation steps on the robust accuracy of TAPT. Figure 4 shows TAPT's robustness performance with varying steps (0, 1, 2, and 4) under the PGD-100 at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19916" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.394 + ], + "angle": 0, + "content": "tack. Notably, step \\(= 0\\) represents the baseline performance without TAPT, reflecting the adversarial robustness achieved solely through APT. A clear trend emerges, showing that robust accuracy increases with additional TAPT steps across most datasets and prompt designs. While robustness gains generally stabilize after a few TAPT steps, the improvement from 0 to 1 step is consistently substantial, highlighting the effectiveness of even a single adaptation step. However, datasets vary in sensitivity to the number of TAPT steps; for example, performance stabilizes quickly on datasets like EuroSAT, whereas for datasets such as ImageNet and DTD, further improvements are observed with additional steps. The figure also illustrates that TAPT's benefits are consistent across different prompt designs. We further analyzed the per-sample time overhead of TAPT, finding that the additional inference time per image is 0.095s (visual-only), 0.166s (V-L joint), and 0.165s (V-L independent), respectively. This demonstrates that TAPT not only consistently enhances robust accuracy but also maintains a relatively low time cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.482, + 0.518 + ], + "angle": 0, + "content": "Different Perturbation Budgets We further assess TAPT's robustness under varying attack strengths, defined by the perturbation budget \\(\\epsilon\\). Figure 5 displays zero-shot adversarial robustness across 11 datasets with \\(\\epsilon\\) values of 1/255, 2/255, and 4/255, and TAPT steps of 1, 2, and 4. As expected, robust accuracy decreases as \\(\\epsilon\\) increases, indicating stronger attacks. Nonetheless, TAPT consistently enhances robust accuracy across all datasets and \\(\\epsilon\\) values." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.524, + 0.48, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.727 + ], + "angle": 0, + "content": "Figure 5. Zero-shot adversarial robustness (y-axis) ofs TAPT under varying perturbation budgets \\(\\epsilon\\) (1/255, 2/255, and 4/255) and TAPT steps (1, 2, and 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "TAPT Reset Intervals Our TAPT method resets the prompt to its initial state before processing each test sample, ensuring that each test sample is handled independently. However, we also explored alternative strategies with varied reset intervals. As shown in Table 3, we experimented with reset intervals ranging from 1 to 32, as well as a strategy with no reset during the entire inference process (i.e., \"reset=all\"). Our findings indicate that continuous prompt adaptation without frequent resets can further improve robust accuracy, suggesting that accumulating information across multiple inputs can be beneficial. However, this con" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "tinuous strategy introduces a vulnerability to potential poisoning attacks. To mitigate this risk and ensure reliable test-time defense, we recommend the per-sample reset strategy (\"reset=1\"), which prioritizes robustness against potential attacks over marginal gains from continued TAPT." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.177, + 0.879, + 0.288 + ], + "angle": 0, + "content": "
Reset IntervalImageNet10 Zero-Shot Datasets Avg.
reset=149.9247.85
reset=250.2047.91
reset=450.6947.91
reset=851.1447.53
reset=1651.4946.39
reset=3251.6244.50
reset=all0.483.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.298, + 0.905, + 0.34 + ], + "angle": 0, + "content": "Table 3. Zero-shot adversarial robustness \\((\\%)\\) of TAPT with varying reset intervals on ImageNet and 10 other zero-shot datasets. \"reset \\(= N\\) \" means the prompt is reset after every \\(N\\) test samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.356, + 0.628, + 0.371 + ], + "angle": 0, + "content": "5. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.562 + ], + "angle": 0, + "content": "As a test-time defense method, TAPT has certain limitations that warrant further research. Our method primarily addresses attacks in the image modality by aligning adversarial image embeddings with pre-computed public data statistics on a public dataset. Future work could explore additional modality alignment and acceleration techniques to facilitate TAPT's deployment in industrial applications. Moreover, our current focus is limited to image recognition tasks. Extending TAPT to a broader range of tasks, such as visual reasoning and visual question answering in advanced models like GPT-4V [2] and Gemini [45], represents a promising direction for future research." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.634, + 0.589 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced a novel test-time defense method, Test-Time Adversarial Prompt Tuning (TAPT), to enhance the inference robustness of pre-trained VLMs, such as CLIP. TAPT tunes defensive bimodal (textual and visual) prompts for each test sample by leveraging multi-view entropy minimization and adversarial-clean alignment, effectively safeguarding CLIP's zero-shot inference. It also utilizes pre-computed statistics from a public dataset (ImageNet) to defend a wide range of downstream tasks. Comprehensive evaluation across 11 benchmark datasets demonstrates that TAPT effectively enhances zero-shot adversarial robustness against both white-box and black-box attacks, while largely preserving clean accuracy. Compared to training-time adversarial prompt tuning (APT) methods, TAPT offers several advantages: (1) it is unsupervised, (2) enables sample-wise prompt adaptation, (3) delivers superior zero-shot adversarial robustness and clean accuracy, (4) is independent of downstream tasks, and (5) is lightweight. Future research could explore the scalability of test-time defense across different modalities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19917" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.162 + ], + "angle": 0, + "content": "Acknowledgements This work is supported by National Key R&D Program of China (Grant No. 2022ZD0160103) and National Natural Science Foundation of China (Grant No. 62276067). The computations in this research were performed using the CFFF platform of Fudan University." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.213, + 0.482, + 0.282 + ], + "angle": 0, + "content": "[1] Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hussein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization. In NeurIPS, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. preprint arXiv:2303.08774, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[3] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. preprint arXiv:2204.01691, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.482, + 0.463 + ], + "angle": 0, + "content": "[4] Christina Baek, Yiding Jiang, Aditi Raghunathan, and J Zico Kolter. Agreement-on-the-line: Predicting the performance of neural networks under distribution shift. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[5] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. preprint arXiv:2203.17274, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.591 + ], + "angle": 0, + "content": "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[8] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[10] Hao Fang, Jiawei Kong, Wenbo Yu, Bin Chen, Jiawei Li, Shutao Xia, and Ke Xu. One perturbation is enough: On generating universal adversarial perturbations against vision-language pre-training models. preprint arXiv:2406.05491, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[11] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshops, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[12] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] Bangyan He, Xiaojun Jia, Siyuan Liang, Tianrui Lou, Yang Liu, and Xiaochun Cao. Sa-attack: Improving adversarial transferability of vision-language pre-training models via self-augmentation. preprint arXiv:2312.04913, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.213, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J-STARS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[15] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.906, + 0.258 + ], + "angle": 0, + "content": "[16] Noor Hussein, Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Promptsmooth: Certifying robustness of medical vision-language models via prompt learning. In MICCAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[18] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[19] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.906, + 0.453 + ], + "angle": 0, + "content": "[20] Eungyeup Kim, Mingjie Sun, Christina Baek, Aditi Raghunathan, and J Zico Kolter. Test-time adaptation induces stronger accuracy and agreement-on-the-line. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[21] Hoki Kim. Torchattacks: A pytorch repository for adversarial attacks. preprint arXiv:2010.01950, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.579 + ], + "angle": 0, + "content": "[23] Lin Li, Haoyan Guan, Jianing Qiu, and Michael Spratling. One prompt word is enough to boost adversarial robustness for pre-trained vision-language models. In CVPR, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[24] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[25] Xingjun Ma, Linxi Jiang, Hanxun Huang, Zejia Weng, James Bailey, and Yu-Gang Jiang. Imbalanced gradients: a subtle cause of overestimated adversarial robustness. Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.906, + 0.747 + ], + "angle": 0, + "content": "[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. *ICLR*, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[27] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.791, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[28] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. In ICLR, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[29] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. preprint arXiv:2006.10963, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19918" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[30] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[31] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.222 + ], + "angle": 0, + "content": "[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.482, + 0.266 + ], + "angle": 0, + "content": "[33] Anibal Pedraza, Oscar Deniz, and Gloria Bueno. On the relationship between generalization and robustness to adversarial examples. Symmetry, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.482, + 0.452 + ], + "angle": 0, + "content": "[36] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust CLIP: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.454, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[38] Or Sharir, Barak Peleg, and Yoav Shoham. The cost of training nlp models: A concise overview. preprint arXiv:2004.08900, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.555, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[39] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[40] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[41] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. preprint arXiv:1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[42] David Stutz, Matthias Hein, and Bernt Schiele. Disentangling adversarial robustness and generalization. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[43] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[44] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv:2312.11805, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[47] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In IEEE S&P, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[48] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[49] Sibo Wang, Jie Zhang, Zheng Yuan, and Shiguang Shan. Pretrained model guided fine-tuning for zero-shot adversarial robustness. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[50] Xin Wang, Kai Chen, Xingjun Ma, Zhineng Chen, Jingjing Chen, and Yu-Gang Jiang. AdvQDet: Detecting query-based adversarial attacks with adversarial contrastive prompt tuning. In ACM MM, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[51] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In EMNLP, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[52] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[53] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[54] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[55] Ziyi Yin, Muchao Ye, Tianrong Zhang, Tianyu Du, Jinguo Zhu, Han Liu, Jinghui Chen, Ting Wang, and Fenglong Ma. Vlattack: Multimodal adversarial attacks on vision-language tasks via pre-trained models. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[56] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[57] Gengyuan Zhang, Jisen Ren, Jindong Gu, and Volker Tresp. Multi-event video-text retrieval. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[58] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[59] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In ACM MM, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[60] Jiaming Zhang, Xingjun Ma, Xin Wang, Lingyu Qiu, Jiaqi Wang, Yu-Gang Jiang, and Jitao Sang. Adversarial prompt tuning for vision-language models. In ECCV, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[61] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[62] Peng-Fei Zhang, Zi Huang, and Guangdong Bai. Universal adversarial perturbations for vision-language pre-trained models. In ACM SIGIR, pages 862–871, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[63] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. In NeurIPS, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19919" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[66] Wanqi Zhou, Shuanghao Bai, Qibin Zhao, and Badong Chen. Revisiting the adversarial robustness of vision language models: a multimodal perspective. preprint arXiv:2404.19287, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[67] Yiwei Zhou, Xiaobo Xia, Zhiwei Lin, Bo Han, and Tongliang Liu. Few-shot adversarial prompt learning on vision-language models. In NeurIPS, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[68] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic adversarial examples in multimodal contrastive learning. In ACM MM, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19920" + } + ] +] \ No newline at end of file diff --git a/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_origin.pdf b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b50058f8a526fc679495e7919808805b25b5d406 --- /dev/null +++ b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/4adafa67-4896-4739-bdf7-1f5ba54cc07b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35946b6e2f9955669067e3a173b71259bc6dd7020970abc6eb6617ab546c62b3 +size 1559192 diff --git a/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/full.md b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..613219c85efc3c4e464a770b5bc3a6e433db288d --- /dev/null +++ b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/full.md @@ -0,0 +1,341 @@ +# TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models + +Xin Wang $^{1}$ , Kai Chen $^{1}$ , Jiaming Zhang $^{2}$ , Jingjing Chen $^{1}$ , Xingjun Ma $^{1*}$ + +$^{1}$ Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University $^{2}$ Hong Kong University of Science and Technology + +# Abstract + +Large pre-trained Vision-Language Models (VLMs) such as CLIP have demonstrated excellent zero-shot generalizability across various downstream tasks. However, recent studies have shown that the inference performance of CLIP can be greatly degraded by small adversarial perturbations, especially its visual modality, posing significant safety threats. To mitigate this vulnerability, in this paper, we propose a novel defense method called Test-Time Adversarial Prompt Tuning (TAPT) to enhance the inference robustness of CLIP against visual adversarial attacks. TAPT is a test-time defense method that learns defensive bimodal (textual and visual) prompts to robustify the inference process of CLIP. Specifically, it is an unsupervised method that optimizes the defensive prompts for each test sample by minimizing a multi-view entropy and aligning adversarial-clean distributions. We evaluate the effectiveness of TAPT on 11 benchmark datasets, including ImageNet and 10 other zero-shot datasets, demonstrating that it enhances the zero-shot adversarial robustness of the original CLIP by at least $48.9\%$ against AutoAttack (AA), while largely maintaining performance on clean examples. Moreover, TAPT outperforms existing adversarial prompt tuning methods across various backbones, achieving an average robustness improvement of at least $36.6\%$ . Code is available at https://github.com/xinwong/TAPT. + +# 1. Introduction + +Vision-Language Models (VLMs) pre-trained on large-scale datasets of image-text pairs have emerged as powerful backbones for numerous applications, including computer vision [17, 34, 57], medical image analysis [15, 51], and robotics [3, 18, 39]. Despite these advancements, studies indicate that even minor adversarial perturbations in input images can significantly degrade the inference performance + +![](images/91b3058d60f0e678bbcb9234a9d0aa5a0096983e27997f8c47566b26fb78aa27.jpg) +Figure 1. Inference with different prompts. Top: inference with hand-crafted prompts fails to recognize the class 'cat'; Middle: Inference with fixed prompts tuned by APT methods cannot recognize all adversarial images; Bottom: Inference with test-time prompts optimized for each image produces accurate recognitions. + +![](images/ff5ac9bcadb07e84dd9ecfd95ba3b4b02aa49f639f2b5544d97519878016feb7.jpg) + +![](images/59d3b88b3eea202b9110e1f44829a5b35a28a2637a9a93dbba153e036e11800b.jpg) + +![](images/ba39176cbb1ab492474f747efc647abce1341fc20270c8a38d681e6e2d193939.jpg) + +of VLMs [9, 26, 44, 59, 63], thereby posing critical safety risks across a wide range of downstream applications. + +Adversarial training [12, 26, 52, 58] is a general defense strategy that augments training data with adversarial examples crafted to mislead the model. While this method has proven effective [8], it entails costly min-max training, which restricts its practicality—especially for large VLMs, where standard training alone can cost millions of dollars [38]. To address the efficiency limitations of adversarial training, recent research has introduced adversarial prompt tuning (APT) methods [23, 60, 67], which align learnable text prompts with adversarial image embeddings. These differentiable prompts [19, 64, 65] provide a more cost-effective alternative to adversarial training. + +Despite its promising results, APT faces three fundamental limitations: (1) The distribution-dependent nature of learnable prompts restricts their generalization to out-of-distribution scenarios and novel tasks. (2) The requirement for task-specific annotated data presents significant challenges for zero-shot applications. (3) While APT methods improve robustness on specific tasks [42], they often compromise the model's overall generalization performance [33, 43]. Given that real-world applications involve an unbounded set of potential tasks, evaluating adversarial robustness solely on predefined downstream tasks proves insufficient. Therefore, achieving zero-shot adversarial robustness—maintaining performance against adversarial attacks on unseen tasks without task-specific training—remains an open challenge. + +The key to addressing the zero-shot adversarial robustness challenge lies in adaptively identifying a robust prompt that consistently aligns an adversarial image embedding with its correct text embedding for each test sample. To this end, we propose a simple yet effective framework called Test-Time Adversarial Prompt Tuning (TAPT), which dynamically tunes a robust prompt on the fly based on only the provided test sample. Specifically, TAPT learns defensive prompts by minimizing two unsupervised losses: (1) multi-view entropy, which ensures consistent predictions across various augmented views of each test sample, and (2) adversarial-clean embedding alignment, which aligns the means and variances of test sample embeddings with precomputed adversarial-clean embeddings of a public dataset (e.g., ImageNet) to enhance inference robustness. Notably, TAPT operates during the inference phase without requiring any task-specific training set or annotations. The different inference schemes of training-time defense APT methods and our test-time defense TAPT are illustrated in Figure 1. + +We evaluated TAPT against both white-box and black-box adversarial attacks across 11 benchmark datasets. TAPT demonstrated superior performance compared to vanilla CLIP (using hand-crafted prompts) and existing APT methods across three different prompt designs: visually-only prompts, visual-language (V-L) joint prompts, and V-L independent prompts. By dynamically adapting prompts at inference time, TAPT opens a new direction for safeguarding the inference process of pre-trained VLMs. Offering a flexible and efficient solution for zero-shot adversarial robustness while maintaining performance on clean samples, our TAPT method bridges the gap between the need for adversarial robustness and the performance challenges posed by open, real-world environments. + +In summary, our main contributions are: + +- We propose a novel test-time defense method named Test-Time Adversarial Prompt Tuning (TAPT) to robustify the zero-shot inference process of pre-trained VLMs. TAPT adapts the prompt for each test image to achieve robust inference without the need for task-specific tuning for downstream datasets. To the best of our knowledge, TAPT is the first inference-time adversarial defense method for pre-trained VLMs. +- TAPT introduces an adversarial-clean alignment loss that aligns the distribution of a test sample with pre-computed adversarial-clean distributions from a public dataset (ImageNet), thereby improving adversarial robustness while maintaining accuracy on clean samples. It also leverages the advantage of APT by using pre-tuned prompts on ImageNet to further enhance zero-shot adversarial robustness. +- We conduct extensive experiments on 11 datasets, including ImageNet and 10 other zero-shot datasets. Our results demonstrate that TAPT significantly outperforms existing APT baselines against both white-box and black-box at + +tacks. Specifically, TAPT improves zero-shot adversarial robustness against AutoAttack by $36.6\%$ with ViT-B/16, and by $38.0\%$ with ViT-B/32, respectively. + +# 2. Related Work + +Here, we briefly review related works on adversarial attacks and defenses for pre-trained VLMs, and test-time adaptation techniques proposed to improve generalization. + +Adversarial Attacks on Pre-trained VLMs Adversarial attacks on pre-trained VLMs are broadly categorized as white-box or black-box attacks according to the threat model. In white-box attacks, the attacker has full access to the model parameters and can directly compute adversarial gradients [25, 26, 59, 68]. Black-box attacks, on the other hand, restrict the attacker to querying model outputs [10, 13, 24, 47, 54, 55, 62, 63]. Traditional single-modal attacks designed for vision models, such as PGD [26], DI [54], and AutoAttack [8], can be used directly to attack the image encoders of pre-trained VLMs. Several recent multi-modal attacks have targeted pre-trained VLMs by simultaneously exploiting vulnerabilities in both their image and text encoders. For example, Co-Attack [59] pioneered white-box multi-modal attacks, perturbing both image and text modalities concurrently. SGA [24] extended Co-Attack to the black-box setting, improving the transferability of multimodal adversarial examples. VLATTACK [55] further generates adversarial examples by fusing perturbations of images and texts from both single-modal and multi-modal. + +Adversarial Defenses for Pre-trained VLMs Adversarial training/tuning is a widely used defense strategy for pretrained VLMs, with existing methods generally classified into adversarial contrastive tuning [28, 36, 49, 50, 52, 66] and adversarial prompt tuning [16, 23, 60, 67]. Adversarial contrastive tuning focuses on enhancing the adversarial robustness of the backbone model. For example, TeCoA [28] examines the impact of fine-tuning and visual prompt tuning on the zero-shot adversarial robustness of VLMs. FARE [36] improves CLIP's image encoder through unsupervised adversarial fine-tuning, enhancing adversarial robustness in models like LLaVA and OpenFlamingo without retraining. PMG-AFT [49] introduces an auxiliary branch to boost zero-shot adversarial robustness, while MMCoA [66] investigates VLM vulnerabilities to multimodal attacks. In contrast, AdvPT [60] and APT [23] offer an efficient approach to bolster the adversarial robustness of VLMs by tuning only the textual prompts without altering the model parameters. FAP [67] further refines APT by balancing cross-modal consistency between benign and adversarial inputs. While these methods are all training-time defense methods that need to pre-tune the prompt for a specific downstream task, in this work we propose a novel test-time defense method that optimizes the prompt on the + +![](images/ba494573aa9aef5ac1b53da337f113f33ce39e7dbe90673148520ae3bc4cd761.jpg) +(a) CLIP + +![](images/8205d7f9b11531cbd2b06517e6ea99556776a2d3fc46bb4a821713d3db8b15cf.jpg) +(b) Visual Prompt + +![](images/850ff2d4bfd8ce3a6f0f49d728b2c3695fc8dee003c103d749a807da9917b52d.jpg) +(c) V-L Joint Prompt +Figure 2. An illustration of CLIP and different adversarial prompt tuning schemes. (a) The original CLIP [34]; (b) - (d) Adversarial prompt tuning with three distinctive prompt designs: Visual Prompt (b), V-L Joint Prompt (c), and V-L Independent Prompt (d). + +![](images/ef5799173a7b574bc2d110861ddf16496a380b893c6efc2c819ac86cf3176a1e.jpg) +(d) V-L Independent Prompt + +fly during inference and is task-agnostic. + +Test-Time Adaptation Test-time adaptation (TTA) methods enhance the generalization performance of pre-trained models by adapting to individual test samples or batches, addressing distribution shifts between training (source) and testing (target) data. Early TTA methods [29, 37] address domain shift by updating batch normalization statistics based on test batch statistics. Building on this, TENT [46] improves adaptation by updating batch normalization layers to minimize the entropy of predicted probabilities for each test batch, while MEMO [61] extends this by minimizing entropy over multiple augmented input samples. Inspired by TENT, subsequent TTA methods such as CoTTA [48] and EATA [31] enable pre-trained models to adapt to continuously shifting test distributions. More recently, methods like TPT [40], PromptAlign [1], and MTA [56] have focused on tuning prompts exclusively at test time to ensure consistent predictions across various augmented views of a test sample. Kim et al. [20] finds that combining TTA with AGL [4] further improves out-of-distribution (OOD) accuracy prediction. In contrast to existing TTA methods which primarily emphasize performance on clean samples, in this work we introduce TAPT, a test-time defense, that specifically enhances zero-shot adversarial robustness against potential attacks. + +# 3. Proposed Method + +# 3.1. Preliminaries + +Threat Model We assume a white-box threat model in which the adversary has full knowledge of the target model's architecture and parameters, allowing direct perturbation of test images based on adversarial gradients prior to inference. The defender, or model owner, can deploy any defense strategies to protect against potential adversarial attacks. Specifically, we focus on securing CLIP zero-shot inference, where the defender lacks access to task-specific training data or annotations of the downstream application. + +CLIP We denote the CLIP image encoder as $\mathcal{I}$ , parameterized by $\theta_{\mathcal{I}}$ , and the text encoder as $\mathcal{T}$ , parameterized by $\theta_{\mathcal{T}}$ . Considering a $K$ -class classification problem, where each image $\mathbf{x}$ is associated with a class label in the format "a photo of a " For a clean sample $\mathbf{x} \in [0,1]^d$ and a target CLIP model comprising $\{\mathcal{I},\mathcal{T}\}$ , a white-box adversarial attack seeks to generate an adversarial example $\mathbf{x}'$ that maximizes the model loss as follows: + +$$ +\mathbf {x} ^ {\prime} = \underset {\| \mathbf {x} ^ {\prime} - \mathbf {x} \| _ {\infty} \leq \epsilon} {\arg \max } \mathcal {L} (\mathcal {I} (\mathbf {x} ^ {\prime}), \mathcal {T} (y)), \tag {1} +$$ + +where $\mathcal{L}(\cdot)$ represents the loss function, $\mathbf{x}'$ is the adversarial example, and $\epsilon$ denotes the perturbation budget. + +Adversarial Prompt Tuning (APT) APT applies adversarial training during the prompt tuning process to enhance adversarial robustness, with recent APT methods primarily developed to defend the visual component of CLIP. Rather than relying on hand-crafted prompts like "a photo of a ", APT learns robust prompts from training data to improve adversarial robustness on downstream tasks. As shown in Figure 2, APT can be extended to three distinctive prompt designs: (1) Visual-only (prompts applied solely to the vision branch), (2) Vision-Language (V-L) joint (shared prompts across both branches), and (3) V-L independent (separate prompts for each branch). Specifically, APT optimizes the prompts $\pmb{P} = \{P_v, P_t\} \in \mathbb{R}^{L \times D}$ , where $P_v$ and $P_t$ are the visual and textual prompts, respectively, $L$ denotes the number of prompt tokens, and $D$ is the embedding dimension. + +Given a downstream training dataset $\mathcal{D}_{\mathrm{train}} = \{(\mathbf{x},y)\}$ , the learnable visual prompt $P_{v}$ is appended to the visual input tokens, forming the sequence $\{\mathbf{x},P_{v}\} = \{\mathrm{CLS},e_1,e_2,\dots ,e_M,P_v\}$ . Similarly, the textual prompt $P_{t}$ is appended to the text input, forming $\{y,P_t\}$ . APT enhances adversarial robustness by aligning the embeddings of clean text with those of adversarial images through a min-max optimization. The corresponding optimization prob + +![](images/4cd60936454997b72dce9db9eef39a5935ded2dd1cdf3effd38679d7ee717cde.jpg) +Figure 3. An overview of our proposed TAPT method: Given an adversarial image, TAPT generates multiple augmented views of the image and retains only those views with low entropy in their averaged prediction probabilities. During inference, TAPT then optimizes the prompt by minimizing multi-view entropy across these selected views while aligning their embedding distribution with pre-computed adversarial-clean statistics from a public dataset (ImageNet). + +Algorithm 1 Test-Time Adversarial Prompt Tuning + +1: Input: input image $\mathbf{x}$ , image encoder $\mathcal{I}$ , text encoder $\mathcal{T}$ , augmentation function $\mathcal{A}$ , entropy threshold $\tau$ , precomputed $\mathcal{D}_{\mathrm{public}}$ statistics $\{\mu_{\mathrm{adv}}, \sigma_{\mathrm{adv}}, \mu_{\mathrm{clean}}, \sigma_{\mathrm{clean}}\}$ + +2: Output: learnable prompts $P$ +3: 1. Initialize adversarial prompts: +4: $\pmb{P} \gets \mathrm{APT}(\mathcal{D}_{\mathrm{public}}, \epsilon; \mathcal{I}, \mathcal{T})$ +5: 2. Multi-view entropy-based sample selection: +6: Select the top $\tau$ entropy from $\mathcal{A}(\mathbf{x})$ to form $\mathcal{H}_{\tau}(\mathbf{x})$ +7: 3. Compute multi-view entropy loss: + +8: $\mathcal{L}_{\mathrm{entropy}} = -\sum_{i=1}^{K}\tilde{p}(y_i|\mathcal{H}_\tau(\mathbf{x}),\boldsymbol{P})\log \tilde{p}(y_i|\mathcal{H}_\tau(\mathbf{x}),\boldsymbol{P})$ + +9: 4. Compute current embedding statistics: + +10: for each layer $l$ in $\mathcal{I}$ do +11: $\mu_{l} = \mathrm{mean}(I_{l}(\hat{\mathbf{x}},\pmb {P}))$ for $\hat{\mathbf{x}}$ in $\mathcal{H}_{\tau}(\mathbf{x})$ +12: $\sigma_l^2 = \mathrm{std}(\mathcal{I}_l(\hat{\mathbf{x}},\pmb {P}))$ for $\hat{\mathbf{x}}$ in $\mathcal{H}_{\tau}(\mathbf{x})$ +13: end for + +14: 5. Adversarial-clean embedding alignment: + +15: $\mathcal{L}_{\mathrm{adv}} = \frac{1}{L}\sum_{l = 1}^{L}\left(\| \mu_{l} - \mu_{\mathrm{adv},l}\|_{1} + \left\| \sigma_{l}^{2} - \sigma_{\mathrm{adv},l}^{2}\right\|_{1}\right)$ +16: $\mathcal{L}_{\mathrm{clean}} = \frac{1}{L}\sum_{l = 1}^{L}\left(\| \mu_{l} - \mu_{\mathrm{clean},l}\|_{1} + \left\| \sigma_{l}^{2} - \sigma_{\mathrm{clean},l}^{2}\right\|_{1}\right)$ +17: 6. Optimize prompts: +18: $\mathcal{L}_{\mathrm{TAPT}} = \mathcal{L}_{\mathrm{entropy}} + \alpha \mathcal{L}_{\mathrm{adv}} + (1 - \alpha)\mathcal{L}_{\mathrm{clean}}$ +19: Optimize $\pmb{P} \gets$ Minimize $\mathcal{L}_{\mathrm{TAPT}}$ + +lem can be formulated as: + +$$ +\underset {\boldsymbol {P}} {\arg \min } \mathbb {E} _ {\mathcal {D} _ {\text {t r a i n}}} \max _ {\| \mathbf {x} ^ {\prime} - \mathbf {x} \| _ {\infty} \leq \epsilon} \mathcal {L} \left(\mathcal {I} \left(\mathbf {x} ^ {\prime}, \boldsymbol {P} _ {\boldsymbol {v}}\right), \mathcal {T} (y, \boldsymbol {P} _ {\boldsymbol {t}})\right), \tag {2} +$$ + +where $\mathcal{I}(\mathbf{x}', P_v)$ and $\mathcal{T}(y, P_t)$ denote the adversarial image embedding and text embedding, respectively. APT methods tune the prompts on the training dataset of each downstream task. At test time, the tuned prompts are fixed to perform inference for different test images. + +# 3.2. Test-Time Adversarial Prompt Tuning (TAPT) + +Framework Overview As illustrated in Figure 3, TAPT comprises two main modules: (1) multi-view entropy-based sample selection, and (2) adversarial-clean embedding alignment. The defense procedure of TAPT operates as follows. Given a test image $\mathbf{x} \in \mathcal{D}_{\mathrm{test}}$ , TAPT begins by generating $M$ randomly augmented views $\{\mathcal{A}_1(\mathbf{x}), \mathcal{A}_2(\mathbf{x}), \dots, \mathcal{A}_M(\mathbf{x})\}$ through random augmentations $\mathcal{A}$ . The multi-view entropy-based sample selection module then chooses the top- $K$ views with the lowest entropy in their averaged prediction probabilities. Using these selected views, TAPT optimizes the prompt $P$ during inference by minimizing multi-view entropy and enforcing adversarial-clean alignment. The prompt is reset to its initial state before processing each new test sample or batch. The complete procedure of TAPT is outlined in Algorithm 1. + +Multi-View Entropy-Based Sample Selection Following prior works [1, 40], we first discard ineffective augmented views $\mathcal{A}(\mathbf{x})$ (e.g., when a random crop removes essential image content) by applying a selection filter with a threshold $\tau$ . This filter retains only augmented views with low prediction entropy (high-confidence predictions). Specifically, we define the set of selected augmented views as $\mathcal{H}_{\tau}(\mathbf{x}) = \{\mathcal{A}_j(\mathbf{x})|1\leq j\leq M,H(\mathcal{A}_j(\mathbf{x}))\leq H_{\tau}\}$ , where $H_{\tau}$ is the entropy threshold corresponding to the top $\tau$ lowest entropy values among all $M$ augmentations. TAPT then optimizes prompts by minimizing the multi-view entropy of the averaged prediction probability over these selected views: + +$$ +\mathcal {L} _ {\text {e n t r o p y}} = - \sum_ {i = 1} ^ {K} \tilde {p} \left(y _ {i} \mid \mathcal {H} _ {\tau} (\mathbf {x}), P\right) \log \tilde {p} \left(y _ {i} \mid \mathcal {H} _ {\tau} (\mathbf {x}), P\right), \tag {3} +$$ + +where $\tilde{p}(y_i|\mathcal{H}_{\tau}(\mathbf{x}),\mathbf{P}) = \frac{1}{|\mathcal{H}_{\tau}(\mathbf{x})|}\sum_{\hat{\mathbf{x}}\in \mathcal{H}_{\tau}(\mathbf{x})}p(y_i|\hat{\mathbf{x}},\mathbf{P})$ denotes the average predicted probability for class $y_i$ over the selected augmented views $\mathcal{H}_{\tau}(\mathbf{x})$ when using prompt $\mathbf{P}$ . This optimization encourages the model to achieve consis + +tent predictions by adjusting the prompt specifically for the given instance. + +Adversarial-Clean Embedding Alignment An adversarial test image $\mathbf{x}$ can shift the image embeddings produced by the image encoder $\mathcal{I}(\mathbf{x},\mathbf{P})$ compared to those from clean images, potentially misleading the model. To mitigate this vulnerability, we align the mean and variance of the test image embedding with pre-computed statistics from a public dataset $\mathcal{D}_{\mathrm{public}}$ . Ideally, this public dataset would originate from the original CLIP pre-training data. However, since this data is unavailable, we use ImageNet as a proxy, given CLIP's strong zero-shot performance on ImageNet with standard prompt tuning [5]. Specifically, we compute the mean and variance of the current embeddings for alignment as follows: + +$$ +\begin{array}{l} \mu_ {l} \left(\mathcal {H} _ {\tau} (\mathbf {x}); \boldsymbol {P}\right) = \frac {1}{\left| \mathcal {H} _ {\tau} (\mathbf {x}) \right|} \sum_ {\hat {\mathbf {x}} \in \mathcal {H} _ {\tau} (\mathbf {x})} \mathcal {I} _ {l} (\hat {\mathbf {x}}, \boldsymbol {P}), (4) \\ \sigma_ {l} ^ {2} \left(\mathcal {H} _ {\tau} (\mathbf {x}); \boldsymbol {P}\right) = \frac {\sum_ {\hat {\mathbf {x}} \in \mathcal {H} _ {\tau} (\mathbf {x})} \left(\mathcal {I} _ {l} \left(\hat {\mathbf {x}} , \boldsymbol {P}\right) - \mu_ {l} \left(\mathcal {H} _ {\tau} (\mathbf {x}); \boldsymbol {P}\right)\right) ^ {2}}{\left| \mathcal {H} _ {\tau} (\mathbf {x}) \right| - 1} (5) \\ \end{array} +$$ + +where $\mathcal{I}_l(\hat{\mathbf{x}},P)$ represents the embedding vector at layer $l$ for the augmented input $\hat{\mathbf{x}}\in \mathcal{H}_{\tau}(\mathbf{x})$ given prompt $P$ . Here, $\mu_l(\mathcal{H}_{\tau}(\mathbf{x});P)$ and $\sigma_l^2 (\mathcal{H}_\tau (\mathbf{x});P)$ denote the mean and variance of the test sample embeddings at layer $l$ , respectively. We similarly pre-compute the mean and variance of embeddings from the public dataset using the robust prompt $P_{adv}$ (obtained via APT on the public data) and clean prompt $P_{clean}$ (obtained via standard prompt tuning on the public data). These offline statistics are represented by $\mu_{\mathrm{adv}}$ , $\sigma_{\mathrm{adv}}^2$ , $\mu_{\mathrm{clean}}$ , and $\sigma_{\mathrm{clean}}^2$ , respectively. We then align the mean and variance of the current embeddings with these pre-computed statistics as follows: + +$$ +\mathcal {L} _ {\mathrm {a d v}} = \frac {1}{L} \sum_ {l = 1} ^ {L} \left(\| \mu_ {l} - \mu_ {a d v, l} \| _ {1} + \| \sigma_ {l} ^ {2} - \sigma_ {a d v, l} ^ {2} \| _ {1}\right), \tag {6} +$$ + +$$ +\mathcal {L} _ {\text {c l e a n}} = \frac {1}{L} \sum_ {l = 1} ^ {L} \left(\| \mu_ {l} - \mu_ {\text {c l e a n}, l} \| _ {1} + \| \sigma_ {l} ^ {2} - \sigma_ {\text {c l e a n}, l} ^ {2} \| _ {1}\right), \tag {7} +$$ + +$$ +\mathcal {L} _ {\mathrm {T A P T}} = \mathcal {L} _ {\text {e n t r o p y}} + \alpha \mathcal {L} _ {\mathrm {a d v}} + (1 - \alpha) \mathcal {L} _ {\text {c l e a n}}, \tag {8} +$$ + +where $\alpha$ is a hyperparameter. Setting $\alpha = 0$ aligns the prompt with the clean distribution, which may reduce adversarial robustness, while setting $\alpha = 1$ aligns the prompt with the robust distribution, which may impact performance on clean samples. Our final objective combines the multiview entropy loss with the adversarial-clean alignment to optimize the prompt during inference for a given test sample. This approach enhances adversarial robustness while preserving accuracy on clean samples. Note that, to ensure inference efficiency, TAPT performs only a single step of prompt tuning for each inference. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets and Models We experiment on 11 benchmark datasets (ImageNet val set and 10 other zero-shot test datasets): ImageNet [35], Caltech101 [11], DTD [7], EuroSAT [14], Pets [32], Aircraft [27], Food101 [6], Flowers [30], Cars [22], SUN397 [53], and UCF101 [41]. Our experiments focus on the CLIP model, specifically utilizing the ViT-B/16 and ViT-B/32 architectures. Following standard CLIP usage, we used hand-crafted prompts as textual inputs. For example, the prompt "a photo of a , a type of pet" was applied for the Pets dataset. A summary of these datasets and their corresponding hand-crafted prompts are provided in the Appendix. + +Attack Configuration We evaluate the zero-shot adversarial robustness of CLIP against both white-box and black-box attacks. Specifically, we employ PGD-100 [26] for white-box attacks, DI [54] for black-box attacks, and the more powerful AutoAttack [8]. The hyperparameters for PGD-100 and DI were configured based on the TorchAttacks library [21]. Consistent with [28], we use perturbation budgets of $\epsilon = 1/255$ , $2/255$ , and $4/255$ for both attacks. + +Defense Configuration For existing APT methods, we use the original configurations of APT methods [23, 60] and generate adversarial examples using the PGD-2 attack with step size $\alpha = 1 / 255$ . We then develop more robust versions of APT methods with different prompt designs, including (1) adversarial visual prompt tuning (APT-V), (2) adversarial V-L joint prompt tuning (APT-VLJ), and (3) adversarial V-L independent prompt tuning (APT-VLI). + +Implementation Details For our TAPT method, we initialize the defensive prompt using APT on ImageNet, training for 100 epochs with a batch size of 32 and a learning rate of 0.035. To generate augmented views for test-time finetuning, we create 63 variations of each test sample using random resized crops and horizontal flips, resulting in 64 images per sample, including the original. From these 64 images, we select the top $10\%$ most confident predictions (with the lowest entropy) and compute the average entropy of their predicted probabilities. For adversarial-clean alignment, we pre-compute embedding statistics from the public dataset (ImageNet) using APT and standard PT, respectively. We then optimize the defensive prompts by minimizing a combined loss of multi-view entropy and adversarial-clean alignment using the AdamW optimizer, with a learning rate of $5 \times 10^{-4}$ and an adversarial-clean scale factor of $\alpha = 0.5$ , on a single NVIDIA A100 GPU. + +# 4.2. Main Results + +Zero-Shot Adversarial Robustness We compare our TAPT method with existing APT methods across three + +
ImageNetCaltech101DTDEuroSATPetsAircraftFood101FlowersCarsSUN397UCF101Avg.
CLIPVanillaPGD 1.421.31.46.05.00.09.81.60.70.81.84.5
DI 6.125.88.70.311.10.59.53.03.44.94.77.1
AA 0.00.00.00.10.00.10.00.00.00.00.00.1
PGD 1.322.95.00.02.60.03.61.50.21.11.73.6
DI 6.437.211.10.79.90.17.59.33.98.65.89.1
AA 0.00.00.00.10.00.10.00.10.10.00.10.1
Visual OnlyAPT-VPGD 19.461.218.58.037.43.912.125.19.817.216.420.8
DI 29.267.523.512.347.15.021.830.218.127.122.027.6
AA 14.855.915.22.331.32.38.318.05.812.412.716.3
PGD 18.963.820.20.636.12.715.023.19.419.118.420.7
DI 26.968.622.05.146.04.823.726.513.926.424.626.2
AA 8.344.912.70.116.40.55.59.02.77.57.810.5
TAPT-VPGD 40.1 (20.7↑)69.7 (8.5↑)28.1 (9.6↑)23.9 (15.9↑)49.2 (11.8↑)11.5 (7.6↑)54.5 (42.4↑)41.1 (16.0↑)28.6 (18.8↑)40.3 (23.1↑)36.5 (20.1↑)38.5 (17.7↑)
DI 46.7 (17.5↑)75.8 (8.3↑)33.2 (9.7↑)33.0 (20.7↑)56.7 (9.6↑)12.7 (7.7↑)61.5 (39.7↑)46.2 (16.0↑)35.7 (17.6↑)45.5 (18.4↑)40.1 (18.1↑)44.3 (16.7↑)
AA 49.2 (34.4↑)75.7 (19.8↑)36.6 (21.4↑)36.9 (34.6↑)57.1 (25.8↑)19.4 (17.1↑)68.8 (60.5↑)52.9 (34.9↑)44.0 (38.2↑)48.9 (36.5↑)48.1 (35.4↑)48.9 (32.6↑)
PGD 42.2 (23.3↑)79.4 (15.6↑)32.2 (12.0↑)24.4 (23.8↑)62.4 (26.3↑)12.1 (9.4↑)53.2 (38.2↑)47.1 (24.0↑)33.5 (24.1↑)44.1 (25.0↑)44.5 (26.1↑)43.2 (22.5↑)
DI 46.5 (19.6↑)81.5 (12.9↑)33.6 (11.6↑)24.8 (19.7↑)66.0 (20.0↑)13.3 (8.5↑)57.7 (34.0↑)48.7 (22.2↑)37.5 (23.6↑)47.3 (20.9↑)47.7 (23.1↑)45.9 (19.7↑)
AA 44.1 (35.8↑)76.3 (31.4↑)33.5 (20.8↑)31.4 (31.3↑)54.9 (38.5↑)16.4 (15.9↑)53.6 (48.1↑)47.3 (38.3↑)42.4 (39.7↑)45.6 (38.1↑)45.0 (37.2↑)44.6 (34.1↑)
V-L JointAPT-VLJPGD 23.961.718.79.741.43.214.123.412.217.715.522.0
DI 34.972.025.010.552.84.224.632.918.928.524.829.9
AA 16.553.212.65.631.31.78.016.04.911.111.115.6
PGD 21.464.314.710.437.72.016.521.08.917.518.021.1
DI 30.469.517.011.947.92.926.726.117.326.526.427.5
AA 10.346.210.03.018.50.66.79.51.77.47.211.1
TAPT-VLJPGD 50.2 (26.3↑)81.0 (19.3↑)29.5 (10.8↑)13.5 (3.8↑)68.7 (27.3↑)5.6 (2.4↑)41.7 (27.6↑)39.3 (15.9↑)28.1 (15.9↑)39.8 (22.1↑)41.5 (26.0↑)39.9 (17.9↑)
DI 51.7 (16.8↑)82.3 (10.3↑)30.4 (5.4↑)14.5 (4.0↑)69.5 (16.7↑)5.8 (1.6↑)43.2 (18.6↑)40.2 (7.3↑)30.2 (11.3↑)41.3 (12.8↑)42.2 (17.4↑)41.0 (11.1↑)
AA 52.4 (35.9↑)81.5 (28.3↑)30.2 (17.6↑)15.4 (9.8↑)69.1 (37.8↑)5.8 (4.1↑)44.8 (36.8↑)40.0 (24.0↑)31.0 (26.1↑)42.1 (31.0↑)44.3 (33.2↑)41.5 (25.9↑)
PGD 52.1 (30.7↑)80.6 (16.3↑)27.1 (12.4↑)13.4 (3.0↑)72.0 (34.3↑)7.6 (5.6↑)52.5 (36.0↑)39.9 (18.0↑)32.3 (23.4↑)44.6 (27.1↑)45.3 (27.1↑)42.4 (21.3↑)
DI 52.4 (22.0↑)80.5 (11.0↑)27.1 (10.1↑)13.6 (1.7↑)71.6 (23.7↑)7.7 (4.8↑)52.0 (25.3↑)40.0 (13.9↑)33.4 (16.1↑)44.6 (18.1↑)45.5 (19.1↑)42.6 (15.1↑)
AA 52.4 (42.1↑)78.6 (32.4↑)27.4 (17.4↑)13.1 (10.1↑)70.0 (51.5↑)8.6 (8.0↑)53.0 (46.3↑)40.5 (31.0↑)34.7 (33.0↑)44.9 (37.5↑)44.8 (37.6↑)42.5 (31.4↑)
V-L IndependentAPT-VLJPGD 24.365.318.910.043.63.114.323.610.518.217.422.7
DI 35.167.719.310.649.84.323.627.716.226.421.527.5
AA 17.257.114.48.235.31.58.916.65.111.912.517.2
PGD 21.263.215.810.437.61.516.120.28.316.917.620.8
DI 28.969.520.310.947.02.824.824.314.823.823.626.4
AA 9.745.610.66.817.30.46.08.72.06.86.911.0
TAPT-VLJPGD 50.0 (25.7↑)79.0 (13.7↑)32.4 (13.5↑)36.2 (26.2↑)67.5 (51.4↑)13.1 (10.0↑)65.7 (51.4↑)49.8 (26.2↑)39.6 (29.1↑)48.3 (30.1↑)47.5 (30.1↑)48.1 (25.4↑)
DI 53.8 (18.7↑)80.5 (12.8↑)32.3 (13.0↑)39.6 (29.0↑)69.4 (19.6↑)13.3 (9.0↑)70.6 (47.0↑)51.1 (23.4↑)42.5 (26.3↑)50.3 (23.9↑)48.2 (26.7↑)50.1 (22.6↑)
AA 55.1 (37.9↑)80.3 (23.2↑)35.7 (21.3↑)44.4 (36.2↑)70.2 (34.9↑)16.0 (14.5↑)76.2 (67.3↑)55.1 (38.5↑)50.5 (45.4↑)53.6 (41.7↑)54.5 (42.0↑)53.8 (36.6↑)
PGD 48.2 (27.0↑)82.1 (18.9↑)30.7 (14.9↑)31.6 (21.2↑)68.1 (30.5↑)4.5 (3.0↑)64.7 (48.6↑)44.6 (24.4↑)41.3 (33.0↑)47.6 (30.7↑)49.1 (31.5↑)46.6 (25.8↑)
DI 50.0 (21.1↑)82.7 (13.2↑)31.5 (11.2↑)32.7 (21.8↑)69.4 (22.4↑)4.4 (1.6↑)65.6 (40.8↑)45.1 (20.8↑)42.8 (28.0↑)49.1 (25.3↑)49.9 (26.3↑)47.6 (21.2↑)
AA 49.7 (40.0↑)80.6 (35.0↑)33.3 (22.7↑)38.1 (31.3↑)68.4 (51.1↑)5.0 (4.6↑)67.5 (61.5↑)47.1 (38.4↑)48.1 (46.1↑)50.3 (43.5↑)50.4 (43.5↑)49.0 (38.0↑)
+ +Table 1. Zero-shot adversarial robustness (\%) of different defense methods from ImageNet to downstream datasets, evaluated against PGD [26], DI [54], and AutoAttack (AA) [8] under perturbation budget $\epsilon = 1/255$ . The baseline APT methods (APT-V, APT-VLJ, and APT-VLI) were tuned on ImageNet under a 16-shot setting and then assessed on the other 10 datasets. The green upward arrows ( $\uparrow$ ) highlight the performance improvement of our TAPT over the baselines. + +
SourceTarget
ImageNetCaltech101DTDEuroSATPetsAircraftFood101FlowersCarsSUN397UCF101Avg.
APT-V60.689.736.926.784.816.863.255.752.158.254.454.5
TAPT-V66.992.844.641.088.523.885.767.465.762.965.564.1
APT-VLJ64.088.534.816.081.18.162.653.347.053.052.851.0
TAPT-VLJ66.588.536.716.780.67.464.650.448.754.755.451.8
APT-VLI63.888.634.117.280.711.661.151.845.953.552.751.0
TAPT-VLI65.890.439.248.382.516.385.860.362.560.064.461.4
+ +Table 2. Zero-shot clean accuracy (\%) of different defense methods from ImageNet to downstream datasets, including APT baselines (APT-V, APT-VLJ, and APT-VLI) and our TAPT. The backbone is ViT-B/16. The best results are boldfaced. + +prompt designs: visual-only (V), V-L joint (VLJ), and V-L independent (VLI). Table 1 presents the zero-shot adversarial robustness results under PGD-100, DI, and AutoAttack attacks, with the 'Vanilla' showing results without any + +defense. It is clear that both white-box and black-box attacks can drastically reduce accuracy (nearly $0\%$ ) in the absence of defenses, using only imperceptible noise with $\epsilon = 1/255$ . AutoAttack emerges as the most effective attack, achieving a nearly $100\%$ attack success rate on average across all datasets. For defense, our TAPT method consistently achieves the best average performance under AutoAttack across all prompt designs (visual-only, V-L joint, and V-L independent) and ViT architectures (ViT-B/16 and ViT-B/32). With ViT-B/16, TAPT enhances robustness by $32.6\%$ (visual-only), $25.9\%$ (V-L joint), and $36.6\%$ (V-L independent), respectively. Similar improvements are observed with ViT-B/32, where TAPT yields gains of $34.1\%$ , $31.4\%$ , and $38.0\%$ for the respective prompt designs. TAPT also demonstrates superior performance against both PGD-100 and DI attacks. + +Furthermore, TAPT with the V-L independent prompt design achieves the highest average zero-shot adversar + +![](images/b0053048ad4208ae58b6acdec534134b644d35dd3cc5696393a3dd0230e47375.jpg) +(a) Average + +![](images/bc12d30225ac86133592118f2cc96980481e1a8ffd70b98a919a0afb2b5dfb76.jpg) +(b) ImageNet + +![](images/cb4309f20a4e9d6423f9bd32215cdf4cf7c9c389ddf7ca74d3b3e8f848e43641.jpg) +(c) Caltech101 + +![](images/3dbb3c3593dce09a3a5f5bba84ac34c896db565a54e78926d0f7f2a4eb568dce.jpg) +(d) DTD + +![](images/29402cccd71b8c2711e7f5cb74e300bbde007063996479475683f8d5faeb4d60.jpg) +(e) EuroSAT + +![](images/2ee2fc34d5254e393b47cccdb77ecb0969e0788b08006a35193bbce88eaf063e.jpg) +(f) Pets + +![](images/805a0a6f130809b44e01b26abe5573e7294639eaf7d19c12d1a755eecc9c4487.jpg) +(g) Aircraft + +![](images/ef54c6080313b02051e842451b594de293ce24dcc633e8ae48530119009c9803.jpg) +(h) Food101 + +(i) Flowers +Figure 4. Adversarial robustness $(\%)$ of our TAPT method under different test-time adaptation steps (i.e., $\{0,1,2,4\}$ ). The results are reported against the PGD-100 attack on ViT-B/16 and ViT-B/32 architectures. +![](images/564bf260419f0c0976f92fad5a56a618ef0cc2e62cb8dcb153cbe4c519fa2640.jpg) +Visual Prompt ViT-B/16 +Visual Prompt ViT-B/32 + +(j) Cars +![](images/c0423ae1f17d6566b6be988a2dfd4e87324ffe73e67d4083c77004e705c137db.jpg) +V-L Joint Prompt ViT-B/16 +V-L Joint Prompt ViT-B/32 + +(k) SUN397 +![](images/e77c93031090c10d8aa0a5c6781b4af1ee57e1bd3200f072035ebf4424fea127.jpg) +V-L Independent Prompt ViT-B/16 +V-L Independent Prompt ViT-B/32 + +![](images/98f2f37e69d8adef2cc1703a75e5e10947060189543fcd984ce70a9f280bb7d1.jpg) +(I)UCF101 + +ial robustness, surpassing both visual-only and V-L joint prompt designs. To summarize, our experiments reveal that: (1) incorporating textual prompts generally improves zero-shot adversarial robustness across various datasets; (2) the V-L joint prompt enhances robustness on the source domain (ImageNet) more effectively than other prompt designs; and (3) the V-L independent prompt design is more effective in enhancing zero-shot adversarial robustness than the V-L joint prompt design, likely due to the challenges in optimizing the interplay between visual and textual prompts. + +Zero-Shot Clean Accuracy Table 2 presents the zero-shot clean accuracy of different defense methods from ImageNet to downstream datasets, comparing APT baselines (APT-V, APT-VLJ, and APT-VLI) with our TAPT method on the ViT-B/16 architecture. Note that the APT baselines were trained on ImageNet and subsequently tested on 10 downstream datasets, while for our TAPT, only the robust statistics were computed based on ImageNet. The + +zero-shot clean accuracy should be compared on all 10 datasets. As shown, TAPT consistently achieves superior clean performance across all 11 datasets, demonstrating significantly stronger generalization capabilities. While a slight performance decrease is observed with TAPT's V-L joint prompt design on Pets, Aircraft, and Flowers, it remains competitive. Overall, TAPT outperforms APT across all three prompt designs (visual-only, V-L joint, and V-L independent), with average improvements of $9.6\%$ , $0.8\%$ , and $10.4\%$ , respectively. These results underscore TAPT's ability to improve adversarial robustness without compromising too much clean accuracy. + +# 4.3. Ablation Studies + +Number of TAPT Steps We first examine the effect of the number of test-time adaptation steps on the robust accuracy of TAPT. Figure 4 shows TAPT's robustness performance with varying steps (0, 1, 2, and 4) under the PGD-100 at + +tack. Notably, step $= 0$ represents the baseline performance without TAPT, reflecting the adversarial robustness achieved solely through APT. A clear trend emerges, showing that robust accuracy increases with additional TAPT steps across most datasets and prompt designs. While robustness gains generally stabilize after a few TAPT steps, the improvement from 0 to 1 step is consistently substantial, highlighting the effectiveness of even a single adaptation step. However, datasets vary in sensitivity to the number of TAPT steps; for example, performance stabilizes quickly on datasets like EuroSAT, whereas for datasets such as ImageNet and DTD, further improvements are observed with additional steps. The figure also illustrates that TAPT's benefits are consistent across different prompt designs. We further analyzed the per-sample time overhead of TAPT, finding that the additional inference time per image is 0.095s (visual-only), 0.166s (V-L joint), and 0.165s (V-L independent), respectively. This demonstrates that TAPT not only consistently enhances robust accuracy but also maintains a relatively low time cost. + +Different Perturbation Budgets We further assess TAPT's robustness under varying attack strengths, defined by the perturbation budget $\epsilon$ . Figure 5 displays zero-shot adversarial robustness across 11 datasets with $\epsilon$ values of 1/255, 2/255, and 4/255, and TAPT steps of 1, 2, and 4. As expected, robust accuracy decreases as $\epsilon$ increases, indicating stronger attacks. Nonetheless, TAPT consistently enhances robust accuracy across all datasets and $\epsilon$ values. + +![](images/25b83c83242023549391d0dff3871670fda981f903e7c2d54cd327fe1d89fd0c.jpg) +Figure 5. Zero-shot adversarial robustness (y-axis) ofs TAPT under varying perturbation budgets $\epsilon$ (1/255, 2/255, and 4/255) and TAPT steps (1, 2, and 4). + +TAPT Reset Intervals Our TAPT method resets the prompt to its initial state before processing each test sample, ensuring that each test sample is handled independently. However, we also explored alternative strategies with varied reset intervals. As shown in Table 3, we experimented with reset intervals ranging from 1 to 32, as well as a strategy with no reset during the entire inference process (i.e., "reset=all"). Our findings indicate that continuous prompt adaptation without frequent resets can further improve robust accuracy, suggesting that accumulating information across multiple inputs can be beneficial. However, this con + +tinuous strategy introduces a vulnerability to potential poisoning attacks. To mitigate this risk and ensure reliable test-time defense, we recommend the per-sample reset strategy ("reset=1"), which prioritizes robustness against potential attacks over marginal gains from continued TAPT. + +
Reset IntervalImageNet10 Zero-Shot Datasets Avg.
reset=149.9247.85
reset=250.2047.91
reset=450.6947.91
reset=851.1447.53
reset=1651.4946.39
reset=3251.6244.50
reset=all0.483.79
+ +Table 3. Zero-shot adversarial robustness $(\%)$ of TAPT with varying reset intervals on ImageNet and 10 other zero-shot datasets. "reset $= N$ " means the prompt is reset after every $N$ test samples. + +# 5. Limitation + +As a test-time defense method, TAPT has certain limitations that warrant further research. Our method primarily addresses attacks in the image modality by aligning adversarial image embeddings with pre-computed public data statistics on a public dataset. Future work could explore additional modality alignment and acceleration techniques to facilitate TAPT's deployment in industrial applications. Moreover, our current focus is limited to image recognition tasks. Extending TAPT to a broader range of tasks, such as visual reasoning and visual question answering in advanced models like GPT-4V [2] and Gemini [45], represents a promising direction for future research. + +# 6. Conclusion + +In this paper, we introduced a novel test-time defense method, Test-Time Adversarial Prompt Tuning (TAPT), to enhance the inference robustness of pre-trained VLMs, such as CLIP. TAPT tunes defensive bimodal (textual and visual) prompts for each test sample by leveraging multi-view entropy minimization and adversarial-clean alignment, effectively safeguarding CLIP's zero-shot inference. It also utilizes pre-computed statistics from a public dataset (ImageNet) to defend a wide range of downstream tasks. Comprehensive evaluation across 11 benchmark datasets demonstrates that TAPT effectively enhances zero-shot adversarial robustness against both white-box and black-box attacks, while largely preserving clean accuracy. Compared to training-time adversarial prompt tuning (APT) methods, TAPT offers several advantages: (1) it is unsupervised, (2) enables sample-wise prompt adaptation, (3) delivers superior zero-shot adversarial robustness and clean accuracy, (4) is independent of downstream tasks, and (5) is lightweight. Future research could explore the scalability of test-time defense across different modalities. + +Acknowledgements This work is supported by National Key R&D Program of China (Grant No. 2022ZD0160103) and National Natural Science Foundation of China (Grant No. 62276067). The computations in this research were performed using the CFFF platform of Fudan University. + +# References + +[1] Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hussein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization. In NeurIPS, 2024. 3, 4 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. preprint arXiv:2303.08774, 2023. 8 +[3] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. preprint arXiv:2204.01691, 2022. 1 +[4] Christina Baek, Yiding Jiang, Aditi Raghunathan, and J Zico Kolter. Agreement-on-the-line: Predicting the performance of neural networks under distribution shift. In NeurIPS, 2022. 3 +[5] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. preprint arXiv:2203.17274, 2022. 5 +[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5 +[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, 2014. 5 +[8] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 1, 2, 5, 6 +[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1 +[10] Hao Fang, Jiawei Kong, Wenbo Yu, Bin Chen, Jiawei Li, Shutao Xia, and Ke Xu. One perturbation is enough: On generating universal adversarial perturbations against vision-language pre-training models. preprint arXiv:2406.05491, 2024. 2 +[11] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshops, 2004. 5 +[12] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. In NeurIPS, 2020. 1 +[13] Bangyan He, Xiaojun Jia, Siyuan Liang, Tianrui Lou, Yang Liu, and Xiaochun Cao. Sa-attack: Improving adversarial transferability of vision-language pre-training models via self-augmentation. preprint arXiv:2312.04913, 2023. 2 + +[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J-STARS, 2019. 5 +[15] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine, 2023. 1 +[16] Noor Hussein, Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Promptsmooth: Certifying robustness of medical vision-language models via prompt learning. In MICCAI, 2024. 2 +[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1 +[18] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In CVPR, 2022. 1 +[19] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 1 +[20] Eungyeup Kim, Mingjie Sun, Christina Baek, Aditi Raghunathan, and J Zico Kolter. Test-time adaptation induces stronger accuracy and agreement-on-the-line. In NeurIPS, 2024. 3 +[21] Hoki Kim. Torchattacks: A pytorch repository for adversarial attacks. preprint arXiv:2010.01950, 2020. 5 +[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops, 2013. 5 +[23] Lin Li, Haoyan Guan, Jianing Qiu, and Michael Spratling. One prompt word is enough to boost adversarial robustness for pre-trained vision-language models. In CVPR, 2024. 1, 2, 5 +[24] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In ICCV, 2023. 2 +[25] Xingjun Ma, Linxi Jiang, Hanxun Huang, Zejia Weng, James Bailey, and Yu-Gang Jiang. Imbalanced gradients: a subtle cause of overestimated adversarial robustness. Machine Learning, 2024. 2 +[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. *ICLR*, 2018. 1, 2, 5, 6 +[27] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. preprint arXiv:1306.5151, 2013. 5 +[28] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. In ICLR, 2023. 2, 5 +[29] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. preprint arXiv:2006.10963, 2020. 3 + +[30] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, 2008. 5 +[31] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In ICML, 2022. 3 +[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, 2012. 5 +[33] Anibal Pedraza, Oscar Deniz, and Gloria Bueno. On the relationship between generalization and robustness to adversarial examples. Symmetry, 2021. 1 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 3 +[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5 +[36] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust CLIP: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. In ICML, 2024. 2 +[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. NeurIPS, 2020. 3 +[38] Or Sharir, Barak Peleg, and Yoav Shoham. The cost of training nlp models: A concise overview. preprint arXiv:2004.08900, 2020. 1 +[39] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, 2022. 1 +[40] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 3, 4 +[41] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. preprint arXiv:1212.0402, 2012. 5 +[42] David Stutz, Matthias Hein, and Bernt Schiele. Disentangling adversarial robustness and generalization. In CVPR, 2019. 1 +[43] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In ECCV, 2018. 1 +[44] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2013. 1 +[45] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv:2312.11805, 2023. 8 + +[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 3 +[47] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In IEEE S&P, 2024. 2 +[48] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, 2022. 3 +[49] Sibo Wang, Jie Zhang, Zheng Yuan, and Shiguang Shan. Pretrained model guided fine-tuning for zero-shot adversarial robustness. In CVPR, 2024. 2 +[50] Xin Wang, Kai Chen, Xingjun Ma, Zhineng Chen, Jingjing Chen, and Yu-Gang Jiang. AdvQDet: Detecting query-based adversarial attacks with adversarial contrastive prompt tuning. In ACM MM, 2024. 2 +[51] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In EMNLP, 2022. 1 +[52] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In CVPR, 2024. 1, 2 +[53] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 5 +[54] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 2, 5, 6 +[55] Ziyi Yin, Muchao Ye, Tianrong Zhang, Tianyu Du, Jinguo Zhu, Han Liu, Jinghui Chen, Ting Wang, and Fenglong Ma. Vlattack: Multimodal adversarial attacks on vision-language tasks via pre-trained models. In NeurIPS, 2023. 2 +[56] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, 2024. 3 +[57] Gengyuan Zhang, Jisen Ren, Jindong Gu, and Volker Tresp. Multi-event video-text retrieval. In ICCV, 2023. 1 +[58] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 1 +[59] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In ACM MM, 2022. 1, 2 +[60] Jiaming Zhang, Xingjun Ma, Xin Wang, Lingyu Qiu, Jiaqi Wang, Yu-Gang Jiang, and Jitao Sang. Adversarial prompt tuning for vision-language models. In ECCV, 2024. 1, 2, 5 +[61] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In NeurIPS, 2022. 3 +[62] Peng-Fei Zhang, Zi Huang, and Guangdong Bai. Universal adversarial perturbations for vision-language pre-trained models. In ACM SIGIR, pages 862–871, 2024. 2 +[63] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. In NeurIPS, 2024. 1, 2 + +[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 1 +[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 1 +[66] Wanqi Zhou, Shuanghao Bai, Qibin Zhao, and Badong Chen. Revisiting the adversarial robustness of vision language models: a multimodal perspective. preprint arXiv:2404.19287, 2024. 2 +[67] Yiwei Zhou, Xiaobo Xia, Zhiwei Lin, Bo Han, and Tongliang Liu. Few-shot adversarial prompt learning on vision-language models. In NeurIPS, 2024. 1, 2 +[68] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic adversarial examples in multimodal contrastive learning. In ACM MM, 2023. 2 \ No newline at end of file diff --git a/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/images.zip b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e3fe95e3bb028cfd8eebef7d83f953713d66be03 --- /dev/null +++ b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f4dc83e9894d30ffbe5b3aaf6c43ba277ff7bd68e8cf5fb97371fbfb97af108 +size 652096 diff --git a/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/layout.json b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..def8f3fab9d4d6313ae4e96e2693ec4987f441e4 --- /dev/null +++ b/2025/TAPT_ Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models/layout.json @@ -0,0 +1,10583 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "type": "text", + "content": "TAPT: Test-Time Adversarial Prompt Tuning for Robust Inference in Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "spans": [ + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "text", + "content": "Xin Wang" + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "text", + "content": ", Kai Chen" + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "text", + "content": ", Jiaming Zhang" + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "text", + "content": ", Jingjing Chen" + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "text", + "content": ", Xingjun Ma" + }, + { + "bbox": [ + 134, + 160, + 480, + 177 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 180, + 496, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 180, + 496, + 210 + ], + "spans": [ + { + "bbox": [ + 115, + 180, + 496, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 180, + 496, + 210 + ], + "type": "text", + "content": "Shanghai Key Lab of Intell. Info. Processing, School of CS, Fudan University " + }, + { + "bbox": [ + 115, + 180, + 496, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 180, + 496, + 210 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 236, + 200, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 236, + 200, + 249 + ], + "spans": [ + { + "bbox": [ + 151, + 236, + 200, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 262, + 297, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 262, + 297, + 562 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 297, + 562 + ], + "type": "text", + "content": "Large pre-trained Vision-Language Models (VLMs) such as CLIP have demonstrated excellent zero-shot generalizability across various downstream tasks. However, recent studies have shown that the inference performance of CLIP can be greatly degraded by small adversarial perturbations, especially its visual modality, posing significant safety threats. To mitigate this vulnerability, in this paper, we propose a novel defense method called Test-Time Adversarial Prompt Tuning (TAPT) to enhance the inference robustness of CLIP against visual adversarial attacks. TAPT is a test-time defense method that learns defensive bimodal (textual and visual) prompts to robustify the inference process of CLIP. Specifically, it is an unsupervised method that optimizes the defensive prompts for each test sample by minimizing a multi-view entropy and aligning adversarial-clean distributions. We evaluate the effectiveness of TAPT on 11 benchmark datasets, including ImageNet and 10 other zero-shot datasets, demonstrating that it enhances the zero-shot adversarial robustness of the original CLIP by at least " + }, + { + "bbox": [ + 54, + 262, + 297, + 562 + ], + "type": "inline_equation", + "content": "48.9\\%" + }, + { + "bbox": [ + 54, + 262, + 297, + 562 + ], + "type": "text", + "content": " against AutoAttack (AA), while largely maintaining performance on clean examples. Moreover, TAPT outperforms existing adversarial prompt tuning methods across various backbones, achieving an average robustness improvement of at least " + }, + { + "bbox": [ + 54, + 262, + 297, + 562 + ], + "type": "inline_equation", + "content": "36.6\\%" + }, + { + "bbox": [ + 54, + 262, + 297, + 562 + ], + "type": "text", + "content": ". Code is available at https://github.com/xinwong/TAPT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 135, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 135, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 135, + 600 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 608, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 296, + 693 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) pre-trained on large-scale datasets of image-text pairs have emerged as powerful backbones for numerous applications, including computer vision [17, 34, 57], medical image analysis [15, 51], and robotics [3, 18, 39]. Despite these advancements, studies indicate that even minor adversarial perturbations in input images can significantly degrade the inference performance" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 238, + 361, + 281 + ], + "blocks": [ + { + "bbox": [ + 315, + 238, + 361, + 281 + ], + "lines": [ + { + "bbox": [ + 315, + 238, + 361, + 281 + ], + "spans": [ + { + "bbox": [ + 315, + 238, + 361, + 281 + ], + "type": "image", + "image_path": "91b3058d60f0e678bbcb9234a9d0aa5a0096983e27997f8c47566b26fb78aa27.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 289, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 346 + ], + "type": "text", + "content": "Figure 1. Inference with different prompts. Top: inference with hand-crafted prompts fails to recognize the class 'cat'; Middle: Inference with fixed prompts tuned by APT methods cannot recognize all adversarial images; Bottom: Inference with test-time prompts optimized for each image produces accurate recognitions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 361, + 238, + 428, + 281 + ], + "blocks": [ + { + "bbox": [ + 361, + 238, + 428, + 281 + ], + "lines": [ + { + "bbox": [ + 361, + 238, + 428, + 281 + ], + "spans": [ + { + "bbox": [ + 361, + 238, + 428, + 281 + ], + "type": "image", + "image_path": "ff5ac9bcadb07e84dd9ecfd95ba3b4b02aa49f639f2b5544d97519878016feb7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 438, + 238, + 484, + 281 + ], + "blocks": [ + { + "bbox": [ + 438, + 238, + 484, + 281 + ], + "lines": [ + { + "bbox": [ + 438, + 238, + 484, + 281 + ], + "spans": [ + { + "bbox": [ + 438, + 238, + 484, + 281 + ], + "type": "image", + "image_path": "59d3b88b3eea202b9110e1f44829a5b35a28a2637a9a93dbba153e036e11800b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 489, + 238, + 551, + 280 + ], + "blocks": [ + { + "bbox": [ + 489, + 238, + 551, + 280 + ], + "lines": [ + { + "bbox": [ + 489, + 238, + 551, + 280 + ], + "spans": [ + { + "bbox": [ + 489, + 238, + 551, + 280 + ], + "type": "image", + "image_path": "ba39176cbb1ab492474f747efc647abce1341fc20270c8a38d681e6e2d193939.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 365, + 554, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 390 + ], + "type": "text", + "content": "of VLMs [9, 26, 44, 59, 63], thereby posing critical safety risks across a wide range of downstream applications." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 390, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 390, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 390, + 555, + 533 + ], + "type": "text", + "content": "Adversarial training [12, 26, 52, 58] is a general defense strategy that augments training data with adversarial examples crafted to mislead the model. While this method has proven effective [8], it entails costly min-max training, which restricts its practicality—especially for large VLMs, where standard training alone can cost millions of dollars [38]. To address the efficiency limitations of adversarial training, recent research has introduced adversarial prompt tuning (APT) methods [23, 60, 67], which align learnable text prompts with adversarial image embeddings. These differentiable prompts [19, 64, 65] provide a more cost-effective alternative to adversarial training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 556, + 714 + ], + "type": "text", + "content": "Despite its promising results, APT faces three fundamental limitations: (1) The distribution-dependent nature of learnable prompts restricts their generalization to out-of-distribution scenarios and novel tasks. (2) The requirement for task-specific annotated data presents significant challenges for zero-shot applications. (3) While APT methods improve robustness on specific tasks [42], they often compromise the model's overall generalization performance [33, 43]. Given that real-world applications involve an unbounded set of potential tasks, evaluating adversarial robustness solely on predefined downstream tasks proves insufficient. Therefore, achieving zero-shot adversarial robustness—maintaining performance against adversarial attacks on unseen tasks without task-specific training—remains an open challenge." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 261, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 261, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 261, + 713 + ], + "type": "text", + "content": "*Corresponding author: " + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19910" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "content": "The key to addressing the zero-shot adversarial robustness challenge lies in adaptively identifying a robust prompt that consistently aligns an adversarial image embedding with its correct text embedding for each test sample. To this end, we propose a simple yet effective framework called Test-Time Adversarial Prompt Tuning (TAPT), which dynamically tunes a robust prompt on the fly based on only the provided test sample. Specifically, TAPT learns defensive prompts by minimizing two unsupervised losses: (1) multi-view entropy, which ensures consistent predictions across various augmented views of each test sample, and (2) adversarial-clean embedding alignment, which aligns the means and variances of test sample embeddings with precomputed adversarial-clean embeddings of a public dataset (e.g., ImageNet) to enhance inference robustness. Notably, TAPT operates during the inference phase without requiring any task-specific training set or annotations. The different inference schemes of training-time defense APT methods and our test-time defense TAPT are illustrated in Figure 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 301, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 469 + ], + "type": "text", + "content": "We evaluated TAPT against both white-box and black-box adversarial attacks across 11 benchmark datasets. TAPT demonstrated superior performance compared to vanilla CLIP (using hand-crafted prompts) and existing APT methods across three different prompt designs: visually-only prompts, visual-language (V-L) joint prompts, and V-L independent prompts. By dynamically adapting prompts at inference time, TAPT opens a new direction for safeguarding the inference process of pre-trained VLMs. Offering a flexible and efficient solution for zero-shot adversarial robustness while maintaining performance on clean samples, our TAPT method bridges the gap between the need for adversarial robustness and the performance challenges posed by open, real-world environments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 472, + 231, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 472, + 231, + 483 + ], + "spans": [ + { + "bbox": [ + 67, + 472, + 231, + 483 + ], + "type": "text", + "content": "In summary, our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": "- We propose a novel test-time defense method named Test-Time Adversarial Prompt Tuning (TAPT) to robustify the zero-shot inference process of pre-trained VLMs. TAPT adapts the prompt for each test image to achieve robust inference without the need for task-specific tuning for downstream datasets. To the best of our knowledge, TAPT is the first inference-time adversarial defense method for pre-trained VLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": "- TAPT introduces an adversarial-clean alignment loss that aligns the distribution of a test sample with pre-computed adversarial-clean distributions from a public dataset (ImageNet), thereby improving adversarial robustness while maintaining accuracy on clean samples. It also leverages the advantage of APT by using pre-tuned prompts on ImageNet to further enhance zero-shot adversarial robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "- We conduct extensive experiments on 11 datasets, including ImageNet and 10 other zero-shot datasets. Our results demonstrate that TAPT significantly outperforms existing APT baselines against both white-box and black-box at" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "type": "text", + "content": "tacks. Specifically, TAPT improves zero-shot adversarial robustness against AutoAttack by " + }, + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "type": "inline_equation", + "content": "36.6\\%" + }, + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "type": "text", + "content": " with ViT-B/16, and by " + }, + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "type": "inline_equation", + "content": "38.0\\%" + }, + { + "bbox": [ + 322, + 72, + 553, + 109 + ], + "type": "text", + "content": " with ViT-B/32, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 122, + 400, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 122, + 400, + 134 + ], + "spans": [ + { + "bbox": [ + 314, + 122, + 400, + 134 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "content": "Here, we briefly review related works on adversarial attacks and defenses for pre-trained VLMs, and test-time adaptation techniques proposed to improve generalization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 183, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 422 + ], + "type": "text", + "content": "Adversarial Attacks on Pre-trained VLMs Adversarial attacks on pre-trained VLMs are broadly categorized as white-box or black-box attacks according to the threat model. In white-box attacks, the attacker has full access to the model parameters and can directly compute adversarial gradients [25, 26, 59, 68]. Black-box attacks, on the other hand, restrict the attacker to querying model outputs [10, 13, 24, 47, 54, 55, 62, 63]. Traditional single-modal attacks designed for vision models, such as PGD [26], DI [54], and AutoAttack [8], can be used directly to attack the image encoders of pre-trained VLMs. Several recent multi-modal attacks have targeted pre-trained VLMs by simultaneously exploiting vulnerabilities in both their image and text encoders. For example, Co-Attack [59] pioneered white-box multi-modal attacks, perturbing both image and text modalities concurrently. SGA [24] extended Co-Attack to the black-box setting, improving the transferability of multimodal adversarial examples. VLATTACK [55] further generates adversarial examples by fusing perturbations of images and texts from both single-modal and multi-modal." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 714 + ], + "type": "text", + "content": "Adversarial Defenses for Pre-trained VLMs Adversarial training/tuning is a widely used defense strategy for pretrained VLMs, with existing methods generally classified into adversarial contrastive tuning [28, 36, 49, 50, 52, 66] and adversarial prompt tuning [16, 23, 60, 67]. Adversarial contrastive tuning focuses on enhancing the adversarial robustness of the backbone model. For example, TeCoA [28] examines the impact of fine-tuning and visual prompt tuning on the zero-shot adversarial robustness of VLMs. FARE [36] improves CLIP's image encoder through unsupervised adversarial fine-tuning, enhancing adversarial robustness in models like LLaVA and OpenFlamingo without retraining. PMG-AFT [49] introduces an auxiliary branch to boost zero-shot adversarial robustness, while MMCoA [66] investigates VLM vulnerabilities to multimodal attacks. In contrast, AdvPT [60] and APT [23] offer an efficient approach to bolster the adversarial robustness of VLMs by tuning only the textual prompts without altering the model parameters. FAP [67] further refines APT by balancing cross-modal consistency between benign and adversarial inputs. While these methods are all training-time defense methods that need to pre-tune the prompt for a specific downstream task, in this work we propose a novel test-time defense method that optimizes the prompt on the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19911" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 89, + 171, + 205 + ], + "blocks": [ + { + "bbox": [ + 100, + 73, + 130, + 83 + ], + "lines": [ + { + "bbox": [ + 100, + 73, + 130, + 83 + ], + "spans": [ + { + "bbox": [ + 100, + 73, + 130, + 83 + ], + "type": "text", + "content": "(a) CLIP" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 89, + 171, + 205 + ], + "lines": [ + { + "bbox": [ + 60, + 89, + 171, + 205 + ], + "spans": [ + { + "bbox": [ + 60, + 89, + 171, + 205 + ], + "type": "image", + "image_path": "ba494573aa9aef5ac1b53da337f113f33ce39e7dbe90673148520ae3bc4cd761.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 184, + 91, + 296, + 205 + ], + "blocks": [ + { + "bbox": [ + 209, + 73, + 268, + 84 + ], + "lines": [ + { + "bbox": [ + 209, + 73, + 268, + 84 + ], + "spans": [ + { + "bbox": [ + 209, + 73, + 268, + 84 + ], + "type": "text", + "content": "(b) Visual Prompt" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 91, + 296, + 205 + ], + "lines": [ + { + "bbox": [ + 184, + 91, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 184, + 91, + 296, + 205 + ], + "type": "image", + "image_path": "8205d7f9b11531cbd2b06517e6ea99556776a2d3fc46bb4a821713d3db8b15cf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 313, + 91, + 423, + 205 + ], + "blocks": [ + { + "bbox": [ + 334, + 73, + 401, + 84 + ], + "lines": [ + { + "bbox": [ + 334, + 73, + 401, + 84 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 401, + 84 + ], + "type": "text", + "content": "(c) V-L Joint Prompt" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 91, + 423, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 91, + 423, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 423, + 205 + ], + "type": "image", + "image_path": "850ff2d4bfd8ce3a6f0f49d728b2c3695fc8dee003c103d749a807da9917b52d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "text", + "content": "Figure 2. An illustration of CLIP and different adversarial prompt tuning schemes. (a) The original CLIP [34]; (b) - (d) Adversarial prompt tuning with three distinctive prompt designs: Visual Prompt (b), V-L Joint Prompt (c), and V-L Independent Prompt (d)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 441, + 90, + 552, + 205 + ], + "blocks": [ + { + "bbox": [ + 450, + 73, + 541, + 84 + ], + "lines": [ + { + "bbox": [ + 450, + 73, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 450, + 73, + 541, + 84 + ], + "type": "text", + "content": "(d) V-L Independent Prompt" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 90, + 552, + 205 + ], + "lines": [ + { + "bbox": [ + 441, + 90, + 552, + 205 + ], + "spans": [ + { + "bbox": [ + 441, + 90, + 552, + 205 + ], + "type": "image", + "image_path": "ef5799173a7b574bc2d110861ddf16496a380b893c6efc2c819ac86cf3176a1e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 254, + 220, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 220, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 220, + 266 + ], + "type": "text", + "content": "fly during inference and is task-agnostic." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 269, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 296, + 555 + ], + "type": "text", + "content": "Test-Time Adaptation Test-time adaptation (TTA) methods enhance the generalization performance of pre-trained models by adapting to individual test samples or batches, addressing distribution shifts between training (source) and testing (target) data. Early TTA methods [29, 37] address domain shift by updating batch normalization statistics based on test batch statistics. Building on this, TENT [46] improves adaptation by updating batch normalization layers to minimize the entropy of predicted probabilities for each test batch, while MEMO [61] extends this by minimizing entropy over multiple augmented input samples. Inspired by TENT, subsequent TTA methods such as CoTTA [48] and EATA [31] enable pre-trained models to adapt to continuously shifting test distributions. More recently, methods like TPT [40], PromptAlign [1], and MTA [56] have focused on tuning prompts exclusively at test time to ensure consistent predictions across various augmented views of a test sample. Kim et al. [20] finds that combining TTA with AGL [4] further improves out-of-distribution (OOD) accuracy prediction. In contrast to existing TTA methods which primarily emphasize performance on clean samples, in this work we introduce TAPT, a test-time defense, that specifically enhances zero-shot adversarial robustness against potential attacks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 567, + 161, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 161, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 161, + 581 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 587, + 141, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 141, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 141, + 599 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Threat Model We assume a white-box threat model in which the adversary has full knowledge of the target model's architecture and parameters, allowing direct perturbation of test images based on adversarial gradients prior to inference. The defender, or model owner, can deploy any defense strategies to protect against potential adversarial attacks. Specifically, we focus on securing CLIP zero-shot inference, where the defender lacks access to task-specific training data or annotations of the downstream application." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": "CLIP We denote the CLIP image encoder as " + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{I}}" + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": ", and the text encoder as " + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": ". Considering a " + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": "-class classification problem, where each image " + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": " is associated with a class label in the format \"a photo of a \" For a clean sample " + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in [0,1]^d" + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": " and a target CLIP model comprising " + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I},\\mathcal{T}\\}" + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": ", a white-box adversarial attack seeks to generate an adversarial example " + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 313, + 253, + 555, + 350 + ], + "type": "text", + "content": " that maximizes the model loss as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 365, + 553, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 365, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 365, + 365, + 553, + 387 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} = \\underset {\\| \\mathbf {x} ^ {\\prime} - \\mathbf {x} \\| _ {\\infty} \\leq \\epsilon} {\\arg \\max } \\mathcal {L} (\\mathcal {I} (\\mathbf {x} ^ {\\prime}), \\mathcal {T} (y)), \\tag {1}", + "image_path": "908bb6d1c0f36c33b510936d654ae451ddd4e0167c93651cb21a19c72e973405.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 396, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 396, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 313, + 396, + 554, + 421 + ], + "type": "text", + "content": " represents the loss function, " + }, + { + "bbox": [ + 313, + 396, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 313, + 396, + 554, + 421 + ], + "type": "text", + "content": " is the adversarial example, and " + }, + { + "bbox": [ + 313, + 396, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 396, + 554, + 421 + ], + "type": "text", + "content": " denotes the perturbation budget." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": "Adversarial Prompt Tuning (APT) APT applies adversarial training during the prompt tuning process to enhance adversarial robustness, with recent APT methods primarily developed to defend the visual component of CLIP. Rather than relying on hand-crafted prompts like \"a photo of a \", APT learns robust prompts from training data to improve adversarial robustness on downstream tasks. As shown in Figure 2, APT can be extended to three distinctive prompt designs: (1) Visual-only (prompts applied solely to the vision branch), (2) Vision-Language (V-L) joint (shared prompts across both branches), and (3) V-L independent (separate prompts for each branch). Specifically, APT optimizes the prompts " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{P} = \\{P_v, P_t\\} \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "P_v" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "P_t" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": " are the visual and textual prompts, respectively, " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": " denotes the number of prompt tokens, and " + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 425, + 555, + 616 + ], + "type": "text", + "content": " is the embedding dimension." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Given a downstream training dataset " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{train}} = \\{(\\mathbf{x},y)\\}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ", the learnable visual prompt " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "P_{v}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " is appended to the visual input tokens, forming the sequence " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x},P_{v}\\} = \\{\\mathrm{CLS},e_1,e_2,\\dots ,e_M,P_v\\}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ". Similarly, the textual prompt " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " is appended to the text input, forming " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{y,P_t\\}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ". APT enhances adversarial robustness by aligning the embeddings of clean text with those of adversarial images through a min-max optimization. The corresponding optimization prob" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19912" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 553, + 186 + ], + "type": "image", + "image_path": "4cd60936454997b72dce9db9eef39a5935ded2dd1cdf3effd38679d7ee717cde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 193, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 193, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 193, + 555, + 240 + ], + "type": "text", + "content": "Figure 3. An overview of our proposed TAPT method: Given an adversarial image, TAPT generates multiple augmented views of the image and retains only those views with low entropy in their averaged prediction probabilities. During inference, TAPT then optimizes the prompt by minimizing multi-view entropy across these selected views while aligning their embedding distribution with pre-computed adversarial-clean statistics from a public dataset (ImageNet)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 251, + 264, + 264 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 264, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 264, + 264 + ], + "type": "text", + "content": "Algorithm 1 Test-Time Adversarial Prompt Tuning" + } + ] + } + ], + "index": 2, + "type": "text" + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "text", + "content": "1: Input: input image " + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "text", + "content": ", image encoder " + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "text", + "content": ", text encoder " + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "text", + "content": ", augmentation function " + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "text", + "content": ", entropy threshold " + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "text", + "content": ", precomputed " + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{public}}" + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "text", + "content": " statistics " + }, + { + "bbox": [ + 60, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\{\\mu_{\\mathrm{adv}}, \\sigma_{\\mathrm{adv}}, \\mu_{\\mathrm{clean}}, \\sigma_{\\mathrm{clean}}\\}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 304, + 288, + 374 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 304, + 196, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 304, + 196, + 315 + ], + "spans": [ + { + "bbox": [ + 61, + 304, + 196, + 315 + ], + "type": "text", + "content": "2: Output: learnable prompts " + }, + { + "bbox": [ + 61, + 304, + 196, + 315 + ], + "type": "inline_equation", + "content": "P" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 315, + 216, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 315, + 216, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 315, + 216, + 327 + ], + "type": "text", + "content": "3: 1. Initialize adversarial prompts:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 193, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 193, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 193, + 339 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 62, + 327, + 193, + 339 + ], + "type": "inline_equation", + "content": "\\pmb{P} \\gets \\mathrm{APT}(\\mathcal{D}_{\\mathrm{public}}, \\epsilon; \\mathcal{I}, \\mathcal{T})" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 339, + 270, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 339, + 270, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 339, + 270, + 350 + ], + "type": "text", + "content": "5: 2. Multi-view entropy-based sample selection:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 350, + 288, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 288, + 363 + ], + "type": "text", + "content": "6: Select the top " + }, + { + "bbox": [ + 61, + 350, + 288, + 363 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 61, + 350, + 288, + 363 + ], + "type": "text", + "content": " entropy from " + }, + { + "bbox": [ + 61, + 350, + 288, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\mathbf{x})" + }, + { + "bbox": [ + 61, + 350, + 288, + 363 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 61, + 350, + 288, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\tau}(\\mathbf{x})" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 363, + 230, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 230, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 230, + 374 + ], + "type": "text", + "content": "7: 3. Compute multi-view entropy loss:" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 61, + 374, + 285, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 374, + 285, + 399 + ], + "spans": [ + { + "bbox": [ + 61, + 374, + 285, + 399 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 61, + 374, + 285, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{entropy}} = -\\sum_{i=1}^{K}\\tilde{p}(y_i|\\mathcal{H}_\\tau(\\mathbf{x}),\\boldsymbol{P})\\log \\tilde{p}(y_i|\\mathcal{H}_\\tau(\\mathbf{x}),\\boldsymbol{P})" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 399, + 252, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 399, + 252, + 409 + ], + "spans": [ + { + "bbox": [ + 61, + 399, + 252, + 409 + ], + "type": "text", + "content": "9: 4. Compute current embedding statistics:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 410, + 244, + 456 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 59, + 410, + 179, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 410, + 179, + 421 + ], + "spans": [ + { + "bbox": [ + 59, + 410, + 179, + 421 + ], + "type": "text", + "content": "10: for each layer " + }, + { + "bbox": [ + 59, + 410, + 179, + 421 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 59, + 410, + 179, + 421 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 59, + 410, + 179, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 59, + 410, + 179, + 421 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 421, + 244, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 421, + 244, + 434 + ], + "spans": [ + { + "bbox": [ + 59, + 421, + 244, + 434 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 59, + 421, + 244, + 434 + ], + "type": "inline_equation", + "content": "\\mu_{l} = \\mathrm{mean}(I_{l}(\\hat{\\mathbf{x}},\\pmb {P}))" + }, + { + "bbox": [ + 59, + 421, + 244, + 434 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 59, + 421, + 244, + 434 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 59, + 421, + 244, + 434 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 59, + 421, + 244, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\tau}(\\mathbf{x})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 434, + 236, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 434, + 236, + 446 + ], + "spans": [ + { + "bbox": [ + 59, + 434, + 236, + 446 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 59, + 434, + 236, + 446 + ], + "type": "inline_equation", + "content": "\\sigma_l^2 = \\mathrm{std}(\\mathcal{I}_l(\\hat{\\mathbf{x}},\\pmb {P}))" + }, + { + "bbox": [ + 59, + 434, + 236, + 446 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 59, + 434, + 236, + 446 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 59, + 434, + 236, + 446 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 59, + 434, + 236, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\tau}(\\mathbf{x})" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 446, + 116, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 446, + 116, + 456 + ], + "spans": [ + { + "bbox": [ + 59, + 446, + 116, + 456 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 59, + 457, + 260, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 457, + 260, + 468 + ], + "spans": [ + { + "bbox": [ + 59, + 457, + 260, + 468 + ], + "type": "text", + "content": "14: 5. Adversarial-clean embedding alignment:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 59, + 468, + 284, + 553 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 59, + 468, + 271, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 468, + 271, + 493 + ], + "spans": [ + { + "bbox": [ + 59, + 468, + 271, + 493 + ], + "type": "text", + "content": "15: " + }, + { + "bbox": [ + 59, + 468, + 271, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}} = \\frac{1}{L}\\sum_{l = 1}^{L}\\left(\\| \\mu_{l} - \\mu_{\\mathrm{adv},l}\\|_{1} + \\left\\| \\sigma_{l}^{2} - \\sigma_{\\mathrm{adv},l}^{2}\\right\\|_{1}\\right)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 493, + 284, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 493, + 284, + 517 + ], + "spans": [ + { + "bbox": [ + 59, + 493, + 284, + 517 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 59, + 493, + 284, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{clean}} = \\frac{1}{L}\\sum_{l = 1}^{L}\\left(\\| \\mu_{l} - \\mu_{\\mathrm{clean},l}\\|_{1} + \\left\\| \\sigma_{l}^{2} - \\sigma_{\\mathrm{clean},l}^{2}\\right\\|_{1}\\right)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 59, + 517, + 167, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 517, + 167, + 529 + ], + "spans": [ + { + "bbox": [ + 59, + 517, + 167, + 529 + ], + "type": "text", + "content": "17: 6. Optimize prompts:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 529, + 253, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 529, + 253, + 541 + ], + "spans": [ + { + "bbox": [ + 59, + 529, + 253, + 541 + ], + "type": "text", + "content": "18: " + }, + { + "bbox": [ + 59, + 529, + 253, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TAPT}} = \\mathcal{L}_{\\mathrm{entropy}} + \\alpha \\mathcal{L}_{\\mathrm{adv}} + (1 - \\alpha)\\mathcal{L}_{\\mathrm{clean}}" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 59, + 541, + 214, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 541, + 214, + 553 + ], + "spans": [ + { + "bbox": [ + 59, + 541, + 214, + 553 + ], + "type": "text", + "content": "19: Optimize " + }, + { + "bbox": [ + 59, + 541, + 214, + 553 + ], + "type": "inline_equation", + "content": "\\pmb{P} \\gets" + }, + { + "bbox": [ + 59, + 541, + 214, + 553 + ], + "type": "text", + "content": " Minimize " + }, + { + "bbox": [ + 59, + 541, + 214, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TAPT}}" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 586, + 162, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 162, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 162, + 597 + ], + "type": "text", + "content": "lem can be formulated as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 66, + 613, + 295, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 613, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 66, + 613, + 295, + 635 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {P}} {\\arg \\min } \\mathbb {E} _ {\\mathcal {D} _ {\\text {t r a i n}}} \\max _ {\\| \\mathbf {x} ^ {\\prime} - \\mathbf {x} \\| _ {\\infty} \\leq \\epsilon} \\mathcal {L} \\left(\\mathcal {I} \\left(\\mathbf {x} ^ {\\prime}, \\boldsymbol {P} _ {\\boldsymbol {v}}\\right), \\mathcal {T} (y, \\boldsymbol {P} _ {\\boldsymbol {t}})\\right), \\tag {2}", + "image_path": "10aa8ae315b7468b3179f021483fa277c6dbac3a93f0dd55bcedd85d6e93ac0a.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 652, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 652, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\mathbf{x}', P_v)" + }, + { + "bbox": [ + 55, + 652, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 652, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(y, P_t)" + }, + { + "bbox": [ + 55, + 652, + 296, + 715 + ], + "type": "text", + "content": " denote the adversarial image embedding and text embedding, respectively. APT methods tune the prompts on the training dataset of each downstream task. At test time, the tuned prompts are fixed to perform inference for different test images." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 251, + 555, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 265 + ], + "type": "text", + "content": "3.2. Test-Time Adversarial Prompt Tuning (TAPT)" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "text", + "content": "Framework Overview As illustrated in Figure 3, TAPT comprises two main modules: (1) multi-view entropy-based sample selection, and (2) adversarial-clean embedding alignment. The defense procedure of TAPT operates as follows. Given a test image " + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{D}_{\\mathrm{test}}" + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "text", + "content": ", TAPT begins by generating " + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "text", + "content": " randomly augmented views " + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{A}_1(\\mathbf{x}), \\mathcal{A}_2(\\mathbf{x}), \\dots, \\mathcal{A}_M(\\mathbf{x})\\}" + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "text", + "content": " through random augmentations " + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "text", + "content": ". The multi-view entropy-based sample selection module then chooses the top-" + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "text", + "content": " views with the lowest entropy in their averaged prediction probabilities. Using these selected views, TAPT optimizes the prompt " + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 268, + 555, + 461 + ], + "type": "text", + "content": " during inference by minimizing multi-view entropy and enforcing adversarial-clean alignment. The prompt is reset to its initial state before processing each new test sample or batch. The complete procedure of TAPT is outlined in Algorithm 1." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "text", + "content": "Multi-View Entropy-Based Sample Selection Following prior works [1, 40], we first discard ineffective augmented views " + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "text", + "content": " (e.g., when a random crop removes essential image content) by applying a selection filter with a threshold " + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "text", + "content": ". This filter retains only augmented views with low prediction entropy (high-confidence predictions). Specifically, we define the set of selected augmented views as " + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\tau}(\\mathbf{x}) = \\{\\mathcal{A}_j(\\mathbf{x})|1\\leq j\\leq M,H(\\mathcal{A}_j(\\mathbf{x}))\\leq H_{\\tau}\\}" + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "inline_equation", + "content": "H_{\\tau}" + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "text", + "content": " is the entropy threshold corresponding to the top " + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "text", + "content": " lowest entropy values among all " + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 463, + 556, + 617 + ], + "type": "text", + "content": " augmentations. TAPT then optimizes prompts by minimizing the multi-view entropy of the averaged prediction probability over these selected views:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 319, + 624, + 555, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 624, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 319, + 624, + 555, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e n t r o p y}} = - \\sum_ {i = 1} ^ {K} \\tilde {p} \\left(y _ {i} \\mid \\mathcal {H} _ {\\tau} (\\mathbf {x}), P\\right) \\log \\tilde {p} \\left(y _ {i} \\mid \\mathcal {H} _ {\\tau} (\\mathbf {x}), P\\right), \\tag {3}", + "image_path": "65d6b32edffeba49de29715ac1d1390cd07430a553f6a7dd98cdea20b5d43a8e.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{p}(y_i|\\mathcal{H}_{\\tau}(\\mathbf{x}),\\mathbf{P}) = \\frac{1}{|\\mathcal{H}_{\\tau}(\\mathbf{x})|}\\sum_{\\hat{\\mathbf{x}}\\in \\mathcal{H}_{\\tau}(\\mathbf{x})}p(y_i|\\hat{\\mathbf{x}},\\mathbf{P})" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": " denotes the average predicted probability for class " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": " over the selected augmented views " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\tau}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": " when using prompt " + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 663, + 556, + 715 + ], + "type": "text", + "content": ". This optimization encourages the model to achieve consis" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19913" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "tent predictions by adjusting the prompt specifically for the given instance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 253 + ], + "type": "text", + "content": "Adversarial-Clean Embedding Alignment An adversarial test image " + }, + { + "bbox": [ + 55, + 99, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 99, + 296, + 253 + ], + "type": "text", + "content": " can shift the image embeddings produced by the image encoder " + }, + { + "bbox": [ + 55, + 99, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\mathbf{x},\\mathbf{P})" + }, + { + "bbox": [ + 55, + 99, + 296, + 253 + ], + "type": "text", + "content": " compared to those from clean images, potentially misleading the model. To mitigate this vulnerability, we align the mean and variance of the test image embedding with pre-computed statistics from a public dataset " + }, + { + "bbox": [ + 55, + 99, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{public}}" + }, + { + "bbox": [ + 55, + 99, + 296, + 253 + ], + "type": "text", + "content": ". Ideally, this public dataset would originate from the original CLIP pre-training data. However, since this data is unavailable, we use ImageNet as a proxy, given CLIP's strong zero-shot performance on ImageNet with standard prompt tuning [5]. Specifically, we compute the mean and variance of the current embeddings for alignment as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 256, + 296, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 296, + 329 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {l} \\left(\\mathcal {H} _ {\\tau} (\\mathbf {x}); \\boldsymbol {P}\\right) = \\frac {1}{\\left| \\mathcal {H} _ {\\tau} (\\mathbf {x}) \\right|} \\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {H} _ {\\tau} (\\mathbf {x})} \\mathcal {I} _ {l} (\\hat {\\mathbf {x}}, \\boldsymbol {P}), (4) \\\\ \\sigma_ {l} ^ {2} \\left(\\mathcal {H} _ {\\tau} (\\mathbf {x}); \\boldsymbol {P}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} \\in \\mathcal {H} _ {\\tau} (\\mathbf {x})} \\left(\\mathcal {I} _ {l} \\left(\\hat {\\mathbf {x}} , \\boldsymbol {P}\\right) - \\mu_ {l} \\left(\\mathcal {H} _ {\\tau} (\\mathbf {x}); \\boldsymbol {P}\\right)\\right) ^ {2}}{\\left| \\mathcal {H} _ {\\tau} (\\mathbf {x}) \\right| - 1} (5) \\\\ \\end{array}", + "image_path": "e5abe368f62591bc998d405c0d777c5ace0e18ebd795f38aa76152ac780171e8.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_l(\\hat{\\mathbf{x}},P)" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": " represents the embedding vector at layer " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": " for the augmented input " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}\\in \\mathcal{H}_{\\tau}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": " given prompt " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\mu_l(\\mathcal{H}_{\\tau}(\\mathbf{x});P)" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\sigma_l^2 (\\mathcal{H}_\\tau (\\mathbf{x});P)" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": " denote the mean and variance of the test sample embeddings at layer " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": ", respectively. We similarly pre-compute the mean and variance of embeddings from the public dataset using the robust prompt " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "P_{adv}" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": " (obtained via APT on the public data) and clean prompt " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "P_{clean}" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": " (obtained via standard prompt tuning on the public data). These offline statistics are represented by " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{adv}}^2" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{clean}}" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{clean}}^2" + }, + { + "bbox": [ + 55, + 333, + 296, + 477 + ], + "type": "text", + "content": ", respectively. We then align the mean and variance of the current embeddings with these pre-computed statistics as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 481, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 481, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 71, + 481, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a d v}} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\left(\\| \\mu_ {l} - \\mu_ {a d v, l} \\| _ {1} + \\| \\sigma_ {l} ^ {2} - \\sigma_ {a d v, l} ^ {2} \\| _ {1}\\right), \\tag {6}", + "image_path": "93ff829d82588c2b3ab7005885a7d649f336229ac635a8ab4e2ab3d09944412e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 517, + 296, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 517, + 296, + 560 + ], + "spans": [ + { + "bbox": [ + 67, + 517, + 296, + 560 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c l e a n}} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\left(\\| \\mu_ {l} - \\mu_ {\\text {c l e a n}, l} \\| _ {1} + \\| \\sigma_ {l} ^ {2} - \\sigma_ {\\text {c l e a n}, l} ^ {2} \\| _ {1}\\right), \\tag {7}", + "image_path": "8c814140dcca7393756812c38688ec8c15a6a1418b327cf69c5c984631bac509.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 565, + 295, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 565, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 66, + 565, + 295, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T A P T}} = \\mathcal {L} _ {\\text {e n t r o p y}} + \\alpha \\mathcal {L} _ {\\mathrm {a d v}} + (1 - \\alpha) \\mathcal {L} _ {\\text {c l e a n}}, \\tag {8}", + "image_path": "b8524371788e28ba5833f64cc77d5ab1d9856a94eac2441d9343bdd594757e3d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " is a hyperparameter. Setting " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " aligns the prompt with the clean distribution, which may reduce adversarial robustness, while setting " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " aligns the prompt with the robust distribution, which may impact performance on clean samples. Our final objective combines the multiview entropy loss with the adversarial-clean alignment to optimize the prompt during inference for a given test sample. This approach enhances adversarial robustness while preserving accuracy on clean samples. Note that, to ensure inference efficiency, TAPT performs only a single step of prompt tuning for each inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 91, + 430, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 430, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 430, + 105 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 110, + 555, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 255 + ], + "type": "text", + "content": "Datasets and Models We experiment on 11 benchmark datasets (ImageNet val set and 10 other zero-shot test datasets): ImageNet [35], Caltech101 [11], DTD [7], EuroSAT [14], Pets [32], Aircraft [27], Food101 [6], Flowers [30], Cars [22], SUN397 [53], and UCF101 [41]. Our experiments focus on the CLIP model, specifically utilizing the ViT-B/16 and ViT-B/32 architectures. Following standard CLIP usage, we used hand-crafted prompts as textual inputs. For example, the prompt \"a photo of a , a type of pet\" was applied for the Pets dataset. A summary of these datasets and their corresponding hand-crafted prompts are provided in the Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "content": "Attack Configuration We evaluate the zero-shot adversarial robustness of CLIP against both white-box and black-box attacks. Specifically, we employ PGD-100 [26] for white-box attacks, DI [54] for black-box attacks, and the more powerful AutoAttack [8]. The hyperparameters for PGD-100 and DI were configured based on the TorchAttacks library [21]. Consistent with [28], we use perturbation budgets of " + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\epsilon = 1/255" + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "inline_equation", + "content": "2/255" + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "inline_equation", + "content": "4/255" + }, + { + "bbox": [ + 313, + 258, + 555, + 354 + ], + "type": "text", + "content": " for both attacks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 357, + 556, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 556, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 556, + 453 + ], + "type": "text", + "content": "Defense Configuration For existing APT methods, we use the original configurations of APT methods [23, 60] and generate adversarial examples using the PGD-2 attack with step size " + }, + { + "bbox": [ + 313, + 357, + 556, + 453 + ], + "type": "inline_equation", + "content": "\\alpha = 1 / 255" + }, + { + "bbox": [ + 313, + 357, + 556, + 453 + ], + "type": "text", + "content": ". We then develop more robust versions of APT methods with different prompt designs, including (1) adversarial visual prompt tuning (APT-V), (2) adversarial V-L joint prompt tuning (APT-VLJ), and (3) adversarial V-L independent prompt tuning (APT-VLI)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 456, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 661 + ], + "type": "text", + "content": "Implementation Details For our TAPT method, we initialize the defensive prompt using APT on ImageNet, training for 100 epochs with a batch size of 32 and a learning rate of 0.035. To generate augmented views for test-time finetuning, we create 63 variations of each test sample using random resized crops and horizontal flips, resulting in 64 images per sample, including the original. From these 64 images, we select the top " + }, + { + "bbox": [ + 313, + 456, + 556, + 661 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 456, + 556, + 661 + ], + "type": "text", + "content": " most confident predictions (with the lowest entropy) and compute the average entropy of their predicted probabilities. For adversarial-clean alignment, we pre-compute embedding statistics from the public dataset (ImageNet) using APT and standard PT, respectively. We then optimize the defensive prompts by minimizing a combined loss of multi-view entropy and adversarial-clean alignment using the AdamW optimizer, with a learning rate of " + }, + { + "bbox": [ + 313, + 456, + 556, + 661 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 456, + 556, + 661 + ], + "type": "text", + "content": " and an adversarial-clean scale factor of " + }, + { + "bbox": [ + 313, + 456, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 313, + 456, + 556, + 661 + ], + "type": "text", + "content": ", on a single NVIDIA A100 GPU." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 671, + 400, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 671, + 400, + 682 + ], + "spans": [ + { + "bbox": [ + 314, + 671, + 400, + 682 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Zero-Shot Adversarial Robustness We compare our TAPT method with existing APT methods across three" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19914" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 554, + 421 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 554, + 421 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 554, + 421 + ], + "type": "table", + "html": "
ImageNetCaltech101DTDEuroSATPetsAircraftFood101FlowersCarsSUN397UCF101Avg.
CLIPVanillaPGD 1.421.31.46.05.00.09.81.60.70.81.84.5
DI 6.125.88.70.311.10.59.53.03.44.94.77.1
AA 0.00.00.00.10.00.10.00.00.00.00.00.1
PGD 1.322.95.00.02.60.03.61.50.21.11.73.6
DI 6.437.211.10.79.90.17.59.33.98.65.89.1
AA 0.00.00.00.10.00.10.00.10.10.00.10.1
Visual OnlyAPT-VPGD 19.461.218.58.037.43.912.125.19.817.216.420.8
DI 29.267.523.512.347.15.021.830.218.127.122.027.6
AA 14.855.915.22.331.32.38.318.05.812.412.716.3
PGD 18.963.820.20.636.12.715.023.19.419.118.420.7
DI 26.968.622.05.146.04.823.726.513.926.424.626.2
AA 8.344.912.70.116.40.55.59.02.77.57.810.5
TAPT-VPGD 40.1 (20.7↑)69.7 (8.5↑)28.1 (9.6↑)23.9 (15.9↑)49.2 (11.8↑)11.5 (7.6↑)54.5 (42.4↑)41.1 (16.0↑)28.6 (18.8↑)40.3 (23.1↑)36.5 (20.1↑)38.5 (17.7↑)
DI 46.7 (17.5↑)75.8 (8.3↑)33.2 (9.7↑)33.0 (20.7↑)56.7 (9.6↑)12.7 (7.7↑)61.5 (39.7↑)46.2 (16.0↑)35.7 (17.6↑)45.5 (18.4↑)40.1 (18.1↑)44.3 (16.7↑)
AA 49.2 (34.4↑)75.7 (19.8↑)36.6 (21.4↑)36.9 (34.6↑)57.1 (25.8↑)19.4 (17.1↑)68.8 (60.5↑)52.9 (34.9↑)44.0 (38.2↑)48.9 (36.5↑)48.1 (35.4↑)48.9 (32.6↑)
PGD 42.2 (23.3↑)79.4 (15.6↑)32.2 (12.0↑)24.4 (23.8↑)62.4 (26.3↑)12.1 (9.4↑)53.2 (38.2↑)47.1 (24.0↑)33.5 (24.1↑)44.1 (25.0↑)44.5 (26.1↑)43.2 (22.5↑)
DI 46.5 (19.6↑)81.5 (12.9↑)33.6 (11.6↑)24.8 (19.7↑)66.0 (20.0↑)13.3 (8.5↑)57.7 (34.0↑)48.7 (22.2↑)37.5 (23.6↑)47.3 (20.9↑)47.7 (23.1↑)45.9 (19.7↑)
AA 44.1 (35.8↑)76.3 (31.4↑)33.5 (20.8↑)31.4 (31.3↑)54.9 (38.5↑)16.4 (15.9↑)53.6 (48.1↑)47.3 (38.3↑)42.4 (39.7↑)45.6 (38.1↑)45.0 (37.2↑)44.6 (34.1↑)
V-L JointAPT-VLJPGD 23.961.718.79.741.43.214.123.412.217.715.522.0
DI 34.972.025.010.552.84.224.632.918.928.524.829.9
AA 16.553.212.65.631.31.78.016.04.911.111.115.6
PGD 21.464.314.710.437.72.016.521.08.917.518.021.1
DI 30.469.517.011.947.92.926.726.117.326.526.427.5
AA 10.346.210.03.018.50.66.79.51.77.47.211.1
TAPT-VLJPGD 50.2 (26.3↑)81.0 (19.3↑)29.5 (10.8↑)13.5 (3.8↑)68.7 (27.3↑)5.6 (2.4↑)41.7 (27.6↑)39.3 (15.9↑)28.1 (15.9↑)39.8 (22.1↑)41.5 (26.0↑)39.9 (17.9↑)
DI 51.7 (16.8↑)82.3 (10.3↑)30.4 (5.4↑)14.5 (4.0↑)69.5 (16.7↑)5.8 (1.6↑)43.2 (18.6↑)40.2 (7.3↑)30.2 (11.3↑)41.3 (12.8↑)42.2 (17.4↑)41.0 (11.1↑)
AA 52.4 (35.9↑)81.5 (28.3↑)30.2 (17.6↑)15.4 (9.8↑)69.1 (37.8↑)5.8 (4.1↑)44.8 (36.8↑)40.0 (24.0↑)31.0 (26.1↑)42.1 (31.0↑)44.3 (33.2↑)41.5 (25.9↑)
PGD 52.1 (30.7↑)80.6 (16.3↑)27.1 (12.4↑)13.4 (3.0↑)72.0 (34.3↑)7.6 (5.6↑)52.5 (36.0↑)39.9 (18.0↑)32.3 (23.4↑)44.6 (27.1↑)45.3 (27.1↑)42.4 (21.3↑)
DI 52.4 (22.0↑)80.5 (11.0↑)27.1 (10.1↑)13.6 (1.7↑)71.6 (23.7↑)7.7 (4.8↑)52.0 (25.3↑)40.0 (13.9↑)33.4 (16.1↑)44.6 (18.1↑)45.5 (19.1↑)42.6 (15.1↑)
AA 52.4 (42.1↑)78.6 (32.4↑)27.4 (17.4↑)13.1 (10.1↑)70.0 (51.5↑)8.6 (8.0↑)53.0 (46.3↑)40.5 (31.0↑)34.7 (33.0↑)44.9 (37.5↑)44.8 (37.6↑)42.5 (31.4↑)
V-L IndependentAPT-VLJPGD 24.365.318.910.043.63.114.323.610.518.217.422.7
DI 35.167.719.310.649.84.323.627.716.226.421.527.5
AA 17.257.114.48.235.31.58.916.65.111.912.517.2
PGD 21.263.215.810.437.61.516.120.28.316.917.620.8
DI 28.969.520.310.947.02.824.824.314.823.823.626.4
AA 9.745.610.66.817.30.46.08.72.06.86.911.0
TAPT-VLJPGD 50.0 (25.7↑)79.0 (13.7↑)32.4 (13.5↑)36.2 (26.2↑)67.5 (51.4↑)13.1 (10.0↑)65.7 (51.4↑)49.8 (26.2↑)39.6 (29.1↑)48.3 (30.1↑)47.5 (30.1↑)48.1 (25.4↑)
DI 53.8 (18.7↑)80.5 (12.8↑)32.3 (13.0↑)39.6 (29.0↑)69.4 (19.6↑)13.3 (9.0↑)70.6 (47.0↑)51.1 (23.4↑)42.5 (26.3↑)50.3 (23.9↑)48.2 (26.7↑)50.1 (22.6↑)
AA 55.1 (37.9↑)80.3 (23.2↑)35.7 (21.3↑)44.4 (36.2↑)70.2 (34.9↑)16.0 (14.5↑)76.2 (67.3↑)55.1 (38.5↑)50.5 (45.4↑)53.6 (41.7↑)54.5 (42.0↑)53.8 (36.6↑)
PGD 48.2 (27.0↑)82.1 (18.9↑)30.7 (14.9↑)31.6 (21.2↑)68.1 (30.5↑)4.5 (3.0↑)64.7 (48.6↑)44.6 (24.4↑)41.3 (33.0↑)47.6 (30.7↑)49.1 (31.5↑)46.6 (25.8↑)
DI 50.0 (21.1↑)82.7 (13.2↑)31.5 (11.2↑)32.7 (21.8↑)69.4 (22.4↑)4.4 (1.6↑)65.6 (40.8↑)45.1 (20.8↑)42.8 (28.0↑)49.1 (25.3↑)49.9 (26.3↑)47.6 (21.2↑)
AA 49.7 (40.0↑)80.6 (35.0↑)33.3 (22.7↑)38.1 (31.3↑)68.4 (51.1↑)5.0 (4.6↑)67.5 (61.5↑)47.1 (38.4↑)48.1 (46.1↑)50.3 (43.5↑)50.4 (43.5↑)49.0 (38.0↑)
", + "image_path": "3c34af431b95383fba014ad0e6cf2aa2622c7d502e80953bc6e0af411937298c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 486, + 295, + 592 + ], + "blocks": [ + { + "bbox": [ + 55, + 430, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 555, + 475 + ], + "type": "text", + "content": "Table 1. Zero-shot adversarial robustness (\\%) of different defense methods from ImageNet to downstream datasets, evaluated against PGD [26], DI [54], and AutoAttack (AA) [8] under perturbation budget " + }, + { + "bbox": [ + 55, + 430, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon = 1/255" + }, + { + "bbox": [ + 55, + 430, + 555, + 475 + ], + "type": "text", + "content": ". The baseline APT methods (APT-V, APT-VLJ, and APT-VLI) were tuned on ImageNet under a 16-shot setting and then assessed on the other 10 datasets. The green upward arrows (" + }, + { + "bbox": [ + 55, + 430, + 555, + 475 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 430, + 555, + 475 + ], + "type": "text", + "content": ") highlight the performance improvement of our TAPT over the baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 486, + 295, + 592 + ], + "lines": [ + { + "bbox": [ + 59, + 486, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 59, + 486, + 295, + 592 + ], + "type": "table", + "html": "
SourceTarget
ImageNetCaltech101DTDEuroSATPetsAircraftFood101FlowersCarsSUN397UCF101Avg.
APT-V60.689.736.926.784.816.863.255.752.158.254.454.5
TAPT-V66.992.844.641.088.523.885.767.465.762.965.564.1
APT-VLJ64.088.534.816.081.18.162.653.347.053.052.851.0
TAPT-VLJ66.588.536.716.780.67.464.650.448.754.755.451.8
APT-VLI63.888.634.117.280.711.661.151.845.953.552.751.0
TAPT-VLI65.890.439.248.382.516.385.860.362.560.064.461.4
", + "image_path": "c854a7c2cabd55c0dc9d31f1db562a172ba3892b596681b619112ae97d5e4692.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 601, + 295, + 644 + ], + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 644 + ], + "type": "text", + "content": "Table 2. Zero-shot clean accuracy (\\%) of different defense methods from ImageNet to downstream datasets, including APT baselines (APT-V, APT-VLJ, and APT-VLI) and our TAPT. The backbone is ViT-B/16. The best results are boldfaced." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "prompt designs: visual-only (V), V-L joint (VLJ), and V-L independent (VLI). Table 1 presents the zero-shot adversarial robustness results under PGD-100, DI, and AutoAttack attacks, with the 'Vanilla' showing results without any" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": "defense. It is clear that both white-box and black-box attacks can drastically reduce accuracy (nearly " + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": ") in the absence of defenses, using only imperceptible noise with " + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\epsilon = 1/255" + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": ". AutoAttack emerges as the most effective attack, achieving a nearly " + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": " attack success rate on average across all datasets. For defense, our TAPT method consistently achieves the best average performance under AutoAttack across all prompt designs (visual-only, V-L joint, and V-L independent) and ViT architectures (ViT-B/16 and ViT-B/32). With ViT-B/16, TAPT enhances robustness by " + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "inline_equation", + "content": "32.6\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": " (visual-only), " + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "inline_equation", + "content": "25.9\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": " (V-L joint), and " + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "inline_equation", + "content": "36.6\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": " (V-L independent), respectively. Similar improvements are observed with ViT-B/32, where TAPT yields gains of " + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "inline_equation", + "content": "34.1\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "inline_equation", + "content": "31.4\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "inline_equation", + "content": "38.0\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 679 + ], + "type": "text", + "content": " for the respective prompt designs. TAPT also demonstrates superior performance against both PGD-100 and DI attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Furthermore, TAPT with the V-L independent prompt design achieves the highest average zero-shot adversar" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19915" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 100, + 179, + 201 + ], + "blocks": [ + { + "bbox": [ + 106, + 92, + 148, + 100 + ], + "lines": [ + { + "bbox": [ + 106, + 92, + 148, + 100 + ], + "spans": [ + { + "bbox": [ + 106, + 92, + 148, + 100 + ], + "type": "text", + "content": "(a) Average" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 100, + 179, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 100, + 179, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 100, + 179, + 201 + ], + "type": "image", + "image_path": "b0053048ad4208ae58b6acdec534134b644d35dd3cc5696393a3dd0230e47375.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 183, + 100, + 302, + 201 + ], + "blocks": [ + { + "bbox": [ + 230, + 92, + 272, + 100 + ], + "lines": [ + { + "bbox": [ + 230, + 92, + 272, + 100 + ], + "spans": [ + { + "bbox": [ + 230, + 92, + 272, + 100 + ], + "type": "text", + "content": "(b) ImageNet" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 100, + 302, + 201 + ], + "lines": [ + { + "bbox": [ + 183, + 100, + 302, + 201 + ], + "spans": [ + { + "bbox": [ + 183, + 100, + 302, + 201 + ], + "type": "image", + "image_path": "bc12d30225ac86133592118f2cc96980481e1a8ffd70b98a919a0afb2b5dfb76.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 100, + 426, + 201 + ], + "blocks": [ + { + "bbox": [ + 351, + 92, + 397, + 100 + ], + "lines": [ + { + "bbox": [ + 351, + 92, + 397, + 100 + ], + "spans": [ + { + "bbox": [ + 351, + 92, + 397, + 100 + ], + "type": "text", + "content": "(c) Caltech101" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 100, + 426, + 201 + ], + "lines": [ + { + "bbox": [ + 307, + 100, + 426, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 100, + 426, + 201 + ], + "type": "image", + "image_path": "cb4309f20a4e9d6423f9bd32215cdf4cf7c9c389ddf7ca74d3b3e8f848e43641.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 432, + 100, + 551, + 201 + ], + "blocks": [ + { + "bbox": [ + 486, + 92, + 511, + 100 + ], + "lines": [ + { + "bbox": [ + 486, + 92, + 511, + 100 + ], + "spans": [ + { + "bbox": [ + 486, + 92, + 511, + 100 + ], + "type": "text", + "content": "(d) DTD" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 100, + 551, + 201 + ], + "lines": [ + { + "bbox": [ + 432, + 100, + 551, + 201 + ], + "spans": [ + { + "bbox": [ + 432, + 100, + 551, + 201 + ], + "type": "image", + "image_path": "3dbb3c3593dce09a3a5f5bba84ac34c896db565a54e78926d0f7f2a4eb568dce.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 209, + 179, + 311 + ], + "blocks": [ + { + "bbox": [ + 108, + 202, + 145, + 209 + ], + "lines": [ + { + "bbox": [ + 108, + 202, + 145, + 209 + ], + "spans": [ + { + "bbox": [ + 108, + 202, + 145, + 209 + ], + "type": "text", + "content": "(e) EuroSAT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 209, + 179, + 311 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 179, + 311 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 179, + 311 + ], + "type": "image", + "image_path": "29402cccd71b8c2711e7f5cb74e300bbde007063996479475683f8d5faeb4d60.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 183, + 209, + 302, + 311 + ], + "blocks": [ + { + "bbox": [ + 239, + 202, + 262, + 209 + ], + "lines": [ + { + "bbox": [ + 239, + 202, + 262, + 209 + ], + "spans": [ + { + "bbox": [ + 239, + 202, + 262, + 209 + ], + "type": "text", + "content": "(f) Pets" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 209, + 302, + 311 + ], + "lines": [ + { + "bbox": [ + 183, + 209, + 302, + 311 + ], + "spans": [ + { + "bbox": [ + 183, + 209, + 302, + 311 + ], + "type": "image", + "image_path": "2ee2fc34d5254e393b47cccdb77ecb0969e0788b08006a35193bbce88eaf063e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 303, + 209, + 427, + 311 + ], + "blocks": [ + { + "bbox": [ + 358, + 202, + 392, + 210 + ], + "lines": [ + { + "bbox": [ + 358, + 202, + 392, + 210 + ], + "spans": [ + { + "bbox": [ + 358, + 202, + 392, + 210 + ], + "type": "text", + "content": "(g) Aircraft" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 209, + 427, + 311 + ], + "lines": [ + { + "bbox": [ + 303, + 209, + 427, + 311 + ], + "spans": [ + { + "bbox": [ + 303, + 209, + 427, + 311 + ], + "type": "image", + "image_path": "805a0a6f130809b44e01b26abe5573e7294639eaf7d19c12d1a755eecc9c4487.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 432, + 209, + 551, + 311 + ], + "blocks": [ + { + "bbox": [ + 480, + 202, + 517, + 210 + ], + "lines": [ + { + "bbox": [ + 480, + 202, + 517, + 210 + ], + "spans": [ + { + "bbox": [ + 480, + 202, + 517, + 210 + ], + "type": "text", + "content": "(h) Food101" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 209, + 551, + 311 + ], + "lines": [ + { + "bbox": [ + 432, + 209, + 551, + 311 + ], + "spans": [ + { + "bbox": [ + 432, + 209, + 551, + 311 + ], + "type": "image", + "image_path": "ef54c6080313b02051e842451b594de293ce24dcc633e8ae48530119009c9803.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 59, + 319, + 179, + 421 + ], + "blocks": [ + { + "bbox": [ + 110, + 312, + 143, + 319 + ], + "lines": [ + { + "bbox": [ + 110, + 312, + 143, + 319 + ], + "spans": [ + { + "bbox": [ + 110, + 312, + 143, + 319 + ], + "type": "text", + "content": "(i) Flowers" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 319, + 179, + 421 + ], + "lines": [ + { + "bbox": [ + 59, + 319, + 179, + 421 + ], + "spans": [ + { + "bbox": [ + 59, + 319, + 179, + 421 + ], + "type": "image", + "image_path": "564bf260419f0c0976f92fad5a56a618ef0cc2e62cb8dcb153cbe4c519fa2640.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 423, + 239, + 430 + ], + "lines": [ + { + "bbox": [ + 159, + 423, + 239, + 430 + ], + "spans": [ + { + "bbox": [ + 159, + 423, + 239, + 430 + ], + "type": "text", + "content": "Visual Prompt ViT-B/16" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 159, + 430, + 239, + 437 + ], + "lines": [ + { + "bbox": [ + 159, + 430, + 239, + 437 + ], + "spans": [ + { + "bbox": [ + 159, + 430, + 239, + 437 + ], + "type": "text", + "content": "Visual Prompt ViT-B/32" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 452, + 555, + 474 + ], + "lines": [ + { + "bbox": [ + 55, + 452, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 555, + 474 + ], + "type": "text", + "content": "Figure 4. Adversarial robustness " + }, + { + "bbox": [ + 55, + 452, + 555, + 474 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 452, + 555, + 474 + ], + "type": "text", + "content": " of our TAPT method under different test-time adaptation steps (i.e., " + }, + { + "bbox": [ + 55, + 452, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\{0,1,2,4\\}" + }, + { + "bbox": [ + 55, + 452, + 555, + 474 + ], + "type": "text", + "content": "). The results are reported against the PGD-100 attack on ViT-B/16 and ViT-B/32 architectures." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 183, + 319, + 302, + 421 + ], + "blocks": [ + { + "bbox": [ + 239, + 312, + 262, + 320 + ], + "lines": [ + { + "bbox": [ + 239, + 312, + 262, + 320 + ], + "spans": [ + { + "bbox": [ + 239, + 312, + 262, + 320 + ], + "type": "text", + "content": "(j) Cars" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 319, + 302, + 421 + ], + "lines": [ + { + "bbox": [ + 183, + 319, + 302, + 421 + ], + "spans": [ + { + "bbox": [ + 183, + 319, + 302, + 421 + ], + "type": "image", + "image_path": "c0423ae1f17d6566b6be988a2dfd4e87324ffe73e67d4083c77004e705c137db.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 423, + 333, + 430 + ], + "lines": [ + { + "bbox": [ + 247, + 423, + 333, + 430 + ], + "spans": [ + { + "bbox": [ + 247, + 423, + 333, + 430 + ], + "type": "text", + "content": "V-L Joint Prompt ViT-B/16" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 247, + 430, + 333, + 437 + ], + "lines": [ + { + "bbox": [ + 247, + 430, + 333, + 437 + ], + "spans": [ + { + "bbox": [ + 247, + 430, + 333, + 437 + ], + "type": "text", + "content": "V-L Joint Prompt ViT-B/32" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 307, + 319, + 427, + 421 + ], + "blocks": [ + { + "bbox": [ + 357, + 312, + 392, + 319 + ], + "lines": [ + { + "bbox": [ + 357, + 312, + 392, + 319 + ], + "spans": [ + { + "bbox": [ + 357, + 312, + 392, + 319 + ], + "type": "text", + "content": "(k) SUN397" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 319, + 427, + 421 + ], + "lines": [ + { + "bbox": [ + 307, + 319, + 427, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 427, + 421 + ], + "type": "image", + "image_path": "e77c93031090c10d8aa0a5c6781b4af1ee57e1bd3200f072035ebf4424fea127.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 423, + 451, + 430 + ], + "lines": [ + { + "bbox": [ + 343, + 423, + 451, + 430 + ], + "spans": [ + { + "bbox": [ + 343, + 423, + 451, + 430 + ], + "type": "text", + "content": "V-L Independent Prompt ViT-B/16" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 343, + 430, + 451, + 437 + ], + "lines": [ + { + "bbox": [ + 343, + 430, + 451, + 437 + ], + "spans": [ + { + "bbox": [ + 343, + 430, + 451, + 437 + ], + "type": "text", + "content": "V-L Independent Prompt ViT-B/32" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 432, + 319, + 551, + 421 + ], + "blocks": [ + { + "bbox": [ + 481, + 312, + 515, + 319 + ], + "lines": [ + { + "bbox": [ + 481, + 312, + 515, + 319 + ], + "spans": [ + { + "bbox": [ + 481, + 312, + 515, + 319 + ], + "type": "text", + "content": "(I)UCF101" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 319, + 551, + 421 + ], + "lines": [ + { + "bbox": [ + 432, + 319, + 551, + 421 + ], + "spans": [ + { + "bbox": [ + 432, + 319, + 551, + 421 + ], + "type": "image", + "image_path": "98f2f37e69d8adef2cc1703a75e5e10947060189543fcd984ce70a9f280bb7d1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 487, + 297, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 297, + 609 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 297, + 609 + ], + "type": "text", + "content": "ial robustness, surpassing both visual-only and V-L joint prompt designs. To summarize, our experiments reveal that: (1) incorporating textual prompts generally improves zero-shot adversarial robustness across various datasets; (2) the V-L joint prompt enhances robustness on the source domain (ImageNet) more effectively than other prompt designs; and (3) the V-L independent prompt design is more effective in enhancing zero-shot adversarial robustness than the V-L joint prompt design, likely due to the challenges in optimizing the interplay between visual and textual prompts." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 54, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 296, + 714 + ], + "type": "text", + "content": "Zero-Shot Clean Accuracy Table 2 presents the zero-shot clean accuracy of different defense methods from ImageNet to downstream datasets, comparing APT baselines (APT-V, APT-VLJ, and APT-VLI) with our TAPT method on the ViT-B/16 architecture. Note that the APT baselines were trained on ImageNet and subsequently tested on 10 downstream datasets, while for our TAPT, only the robust statistics were computed based on ImageNet. The" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 488, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 555, + 632 + ], + "type": "text", + "content": "zero-shot clean accuracy should be compared on all 10 datasets. As shown, TAPT consistently achieves superior clean performance across all 11 datasets, demonstrating significantly stronger generalization capabilities. While a slight performance decrease is observed with TAPT's V-L joint prompt design on Pets, Aircraft, and Flowers, it remains competitive. Overall, TAPT outperforms APT across all three prompt designs (visual-only, V-L joint, and V-L independent), with average improvements of " + }, + { + "bbox": [ + 313, + 488, + 555, + 632 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 632 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 488, + 555, + 632 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 632 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 488, + 555, + 632 + ], + "type": "inline_equation", + "content": "10.4\\%" + }, + { + "bbox": [ + 313, + 488, + 555, + 632 + ], + "type": "text", + "content": ", respectively. These results underscore TAPT's ability to improve adversarial robustness without compromising too much clean accuracy." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 646, + 415, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 415, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 415, + 657 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Number of TAPT Steps We first examine the effect of the number of test-time adaptation steps on the robust accuracy of TAPT. Figure 4 shows TAPT's robustness performance with varying steps (0, 1, 2, and 4) under the PGD-100 at" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19916" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "tack. Notably, step " + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": " represents the baseline performance without TAPT, reflecting the adversarial robustness achieved solely through APT. A clear trend emerges, showing that robust accuracy increases with additional TAPT steps across most datasets and prompt designs. While robustness gains generally stabilize after a few TAPT steps, the improvement from 0 to 1 step is consistently substantial, highlighting the effectiveness of even a single adaptation step. However, datasets vary in sensitivity to the number of TAPT steps; for example, performance stabilizes quickly on datasets like EuroSAT, whereas for datasets such as ImageNet and DTD, further improvements are observed with additional steps. The figure also illustrates that TAPT's benefits are consistent across different prompt designs. We further analyzed the per-sample time overhead of TAPT, finding that the additional inference time per image is 0.095s (visual-only), 0.166s (V-L joint), and 0.165s (V-L independent), respectively. This demonstrates that TAPT not only consistently enhances robust accuracy but also maintains a relatively low time cost." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "text", + "content": "Different Perturbation Budgets We further assess TAPT's robustness under varying attack strengths, defined by the perturbation budget " + }, + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "text", + "content": ". Figure 5 displays zero-shot adversarial robustness across 11 datasets with " + }, + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "text", + "content": " values of 1/255, 2/255, and 4/255, and TAPT steps of 1, 2, and 4. As expected, robust accuracy decreases as " + }, + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "text", + "content": " increases, indicating stronger attacks. Nonetheless, TAPT consistently enhances robust accuracy across all datasets and " + }, + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 315, + 294, + 410 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 415, + 293, + 532 + ], + "blocks": [ + { + "bbox": [ + 58, + 415, + 293, + 532 + ], + "lines": [ + { + "bbox": [ + 58, + 415, + 293, + 532 + ], + "spans": [ + { + "bbox": [ + 58, + 415, + 293, + 532 + ], + "type": "image", + "image_path": "25b83c83242023549391d0dff3871670fda981f903e7c2d54cd327fe1d89fd0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 542, + 295, + 575 + ], + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 575 + ], + "type": "text", + "content": "Figure 5. Zero-shot adversarial robustness (y-axis) ofs TAPT under varying perturbation budgets " + }, + { + "bbox": [ + 55, + 542, + 295, + 575 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 542, + 295, + 575 + ], + "type": "text", + "content": " (1/255, 2/255, and 4/255) and TAPT steps (1, 2, and 4)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "TAPT Reset Intervals Our TAPT method resets the prompt to its initial state before processing each test sample, ensuring that each test sample is handled independently. However, we also explored alternative strategies with varied reset intervals. As shown in Table 3, we experimented with reset intervals ranging from 1 to 32, as well as a strategy with no reset during the entire inference process (i.e., \"reset=all\"). Our findings indicate that continuous prompt adaptation without frequent resets can further improve robust accuracy, suggesting that accumulating information across multiple inputs can be beneficial. However, this con" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "tinuous strategy introduces a vulnerability to potential poisoning attacks. To mitigate this risk and ensure reliable test-time defense, we recommend the per-sample reset strategy (\"reset=1\"), which prioritizes robustness against potential attacks over marginal gains from continued TAPT." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 335, + 140, + 537, + 228 + ], + "blocks": [ + { + "bbox": [ + 335, + 140, + 537, + 228 + ], + "lines": [ + { + "bbox": [ + 335, + 140, + 537, + 228 + ], + "spans": [ + { + "bbox": [ + 335, + 140, + 537, + 228 + ], + "type": "table", + "html": "
Reset IntervalImageNet10 Zero-Shot Datasets Avg.
reset=149.9247.85
reset=250.2047.91
reset=450.6947.91
reset=851.1447.53
reset=1651.4946.39
reset=3251.6244.50
reset=all0.483.79
", + "image_path": "91b0fa1d9324b279fb910e0c4f9d6559ba964d9b3c1d1dbac96d05c39eef1c26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 236, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 269 + ], + "type": "text", + "content": "Table 3. Zero-shot adversarial robustness " + }, + { + "bbox": [ + 313, + 236, + 553, + 269 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 236, + 553, + 269 + ], + "type": "text", + "content": " of TAPT with varying reset intervals on ImageNet and 10 other zero-shot datasets. \"reset " + }, + { + "bbox": [ + 313, + 236, + 553, + 269 + ], + "type": "inline_equation", + "content": "= N" + }, + { + "bbox": [ + 313, + 236, + 553, + 269 + ], + "type": "text", + "content": " \" means the prompt is reset after every " + }, + { + "bbox": [ + 313, + 236, + 553, + 269 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 236, + 553, + 269 + ], + "type": "text", + "content": " test samples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 281, + 384, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 384, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 384, + 293 + ], + "type": "text", + "content": "5. Limitation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 301, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 445 + ], + "type": "text", + "content": "As a test-time defense method, TAPT has certain limitations that warrant further research. Our method primarily addresses attacks in the image modality by aligning adversarial image embeddings with pre-computed public data statistics on a public dataset. Future work could explore additional modality alignment and acceleration techniques to facilitate TAPT's deployment in industrial applications. Moreover, our current focus is limited to image recognition tasks. Extending TAPT to a broader range of tasks, such as visual reasoning and visual question answering in advanced models like GPT-4V [2] and Gemini [45], represents a promising direction for future research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduced a novel test-time defense method, Test-Time Adversarial Prompt Tuning (TAPT), to enhance the inference robustness of pre-trained VLMs, such as CLIP. TAPT tunes defensive bimodal (textual and visual) prompts for each test sample by leveraging multi-view entropy minimization and adversarial-clean alignment, effectively safeguarding CLIP's zero-shot inference. It also utilizes pre-computed statistics from a public dataset (ImageNet) to defend a wide range of downstream tasks. Comprehensive evaluation across 11 benchmark datasets demonstrates that TAPT effectively enhances zero-shot adversarial robustness against both white-box and black-box attacks, while largely preserving clean accuracy. Compared to training-time adversarial prompt tuning (APT) methods, TAPT offers several advantages: (1) it is unsupervised, (2) enables sample-wise prompt adaptation, (3) delivers superior zero-shot adversarial robustness and clean accuracy, (4) is independent of downstream tasks, and (5) is lightweight. Future research could explore the scalability of test-time defense across different modalities." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19917" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 128 + ], + "type": "text", + "content": "Acknowledgements This work is supported by National Key R&D Program of China (Grant No. 2022ZD0160103) and National Natural Science Foundation of China (Grant No. 62276067). The computations in this research were performed using the CFFF platform of Fudan University." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 168, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 168, + 294, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 168, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 168, + 294, + 223 + ], + "type": "text", + "content": "[1] Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hussein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization. In NeurIPS, 2024. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. preprint arXiv:2303.08774, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 269, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 294, + 323 + ], + "type": "text", + "content": "[3] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. preprint arXiv:2204.01691, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 324, + 294, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 294, + 366 + ], + "type": "text", + "content": "[4] Christina Baek, Yiding Jiang, Aditi Raghunathan, and J Zico Kolter. Agreement-on-the-line: Predicting the performance of neural networks under distribution shift. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 401 + ], + "type": "text", + "content": "[5] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. preprint arXiv:2203.17274, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 434 + ], + "type": "text", + "content": "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 468 + ], + "type": "text", + "content": "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, 2014. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 294, + 502 + ], + "type": "text", + "content": "[8] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 503, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 294, + 534 + ], + "type": "text", + "content": "[9] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "text", + "content": "[10] Hao Fang, Jiawei Kong, Wenbo Yu, Bin Chen, Jiawei Li, Shutao Xia, and Ke Xu. One perturbation is enough: On generating universal adversarial perturbations against vision-language pre-training models. preprint arXiv:2406.05491, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "type": "text", + "content": "[11] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshops, 2004. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 669 + ], + "type": "text", + "content": "[12] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[13] Bangyan He, Xiaojun Jia, Siyuan Liang, Tianrui Lou, Yang Liu, and Xiaochun Cao. Sa-attack: Improving adversarial transferability of vision-language pre-training models via self-augmentation. preprint arXiv:2312.04913, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J-STARS, 2019. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "text", + "content": "[15] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature Medicine, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 554, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 554, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 554, + 204 + ], + "type": "text", + "content": "[16] Noor Hussein, Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Promptsmooth: Certifying robustness of medical vision-language models via prompt learning. In MICCAI, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 249 + ], + "type": "text", + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 251, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 554, + 282 + ], + "type": "text", + "content": "[18] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In CVPR, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 284, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 316 + ], + "type": "text", + "content": "[19] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 317, + 554, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 554, + 358 + ], + "type": "text", + "content": "[20] Eungyeup Kim, Mingjie Sun, Christina Baek, Aditi Raghunathan, and J Zico Kolter. Test-time adaptation induces stronger accuracy and agreement-on-the-line. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 381 + ], + "type": "text", + "content": "[21] Hoki Kim. Torchattacks: A pytorch repository for adversarial attacks. preprint arXiv:2010.01950, 2020. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "text", + "content": "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops, 2013. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 416, + 554, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 458 + ], + "type": "text", + "content": "[23] Lin Li, Haoyan Guan, Jianing Qiu, and Michael Spratling. One prompt word is enough to boost adversarial robustness for pre-trained vision-language models. In CVPR, 2024. 1, 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 460, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 503 + ], + "type": "text", + "content": "[24] Dong Lu, Zhiqiang Wang, Teng Wang, Weili Guan, Hongchang Gao, and Feng Zheng. Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 505, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 554, + 547 + ], + "type": "text", + "content": "[25] Xingjun Ma, Linxi Jiang, Hanxun Huang, Zejia Weng, James Bailey, and Yu-Gang Jiang. Imbalanced gradients: a subtle cause of overestimated adversarial robustness. Machine Learning, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 548, + 554, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 591 + ], + "type": "text", + "content": "[26] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. *ICLR*, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 593, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 554, + 624 + ], + "type": "text", + "content": "[27] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 626, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 626, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 626, + 554, + 657 + ], + "type": "text", + "content": "[28] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. In ICLR, 2023. 2, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "type": "text", + "content": "[29] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. preprint arXiv:2006.10963, 2020. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19918" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[30] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, 2008. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[31] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In ICML, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 175 + ], + "type": "text", + "content": "[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, 2012. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "type": "text", + "content": "[33] Anibal Pedraza, Oscar Deniz, and Gloria Bueno. On the relationship between generalization and robustness to adversarial examples. Symmetry, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 212, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 294, + 266 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 268, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 312 + ], + "type": "text", + "content": "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 314, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 294, + 357 + ], + "type": "text", + "content": "[36] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust CLIP: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. In ICML, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 359, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 359, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 359, + 294, + 403 + ], + "type": "text", + "content": "[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. NeurIPS, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "text", + "content": "[38] Or Sharir, Barak Peleg, and Yoav Shoham. The cost of training nlp models: A concise overview. preprint arXiv:2004.08900, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 439, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 439, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 439, + 294, + 472 + ], + "type": "text", + "content": "[39] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 474, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 518 + ], + "type": "text", + "content": "[40] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 520, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 542 + ], + "type": "text", + "content": "[41] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. preprint arXiv:1212.0402, 2012. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 544, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 575 + ], + "type": "text", + "content": "[42] David Stutz, Matthias Hein, and Bernt Schiele. Disentangling adversarial robustness and generalization. In CVPR, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "text", + "content": "[43] Dong Su, Huan Zhang, Hongge Chen, Jinfeng Yi, Pin-Yu Chen, and Yupeng Gao. Is robustness the cost of accuracy? a comprehensive study on the robustness of 18 deep image classification models. In ECCV, 2018. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "text", + "content": "[44] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2013. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[45] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv:2312.11805, 2023. 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[47] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In IEEE S&P, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 174 + ], + "type": "text", + "content": "[48] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "text", + "content": "[49] Sibo Wang, Jie Zhang, Zheng Yuan, and Shiguang Shan. Pretrained model guided fine-tuning for zero-shot adversarial robustness. In CVPR, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[50] Xin Wang, Kai Chen, Xingjun Ma, Zhineng Chen, Jingjing Chen, and Yu-Gang Jiang. AdvQDet: Detecting query-based adversarial attacks with adversarial contrastive prompt tuning. In ACM MM, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[51] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In EMNLP, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "type": "text", + "content": "[52] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "text", + "content": "[53] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[54] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[55] Ziyi Yin, Muchao Ye, Tianrong Zhang, Tianyu Du, Jinguo Zhu, Han Liu, Jinghui Chen, Ting Wang, and Fenglong Ma. Vlattack: Multimodal adversarial attacks on vision-language tasks via pre-trained models. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "text", + "content": "[56] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 467, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 487 + ], + "type": "text", + "content": "[57] Gengyuan Zhang, Jisen Ren, Jindong Gu, and Volker Tresp. Multi-event video-text retrieval. In ICCV, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 490, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 532 + ], + "type": "text", + "content": "[58] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "text", + "content": "[59] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In ACM MM, 2022. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "text", + "content": "[60] Jiaming Zhang, Xingjun Ma, Xin Wang, Lingyu Qiu, Jiaqi Wang, Yu-Gang Jiang, and Jitao Sang. Adversarial prompt tuning for vision-language models. In ECCV, 2024. 1, 2, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 602, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 633 + ], + "type": "text", + "content": "[61] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "text", + "content": "[62] Peng-Fei Zhang, Zi Huang, and Guangdong Bai. Universal adversarial perturbations for vision-language pre-trained models. In ACM SIGIR, pages 862–871, 2024. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[63] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. In NeurIPS, 2024. 1, 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19919" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 262 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 138 + ], + "type": "text", + "content": "[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "text", + "content": "[66] Wanqi Zhou, Shuanghao Bai, Qibin Zhao, and Badong Chen. Revisiting the adversarial robustness of vision language models: a multimodal perspective. preprint arXiv:2404.19287, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[67] Yiwei Zhou, Xiaobo Xia, Zhiwei Lin, Bo Han, and Tongliang Liu. Few-shot adversarial prompt learning on vision-language models. In NeurIPS, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[68] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic adversarial examples in multimodal contrastive learning. In ACM MM, 2023. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19920" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_content_list.json b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2477ec24214b84ed0a50ff464875511d019bc52f --- /dev/null +++ b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_content_list.json @@ -0,0 +1,2007 @@ +[ + { + "type": "text", + "text": "TAROT: Towards Essentially Domain-Invariant Robustness with Theoretical Justification", + "text_level": 1, + "bbox": [ + 197, + 128, + 800, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongyoon Yang $^{1*†}$ Jihu Lee $^{2*}$ Yongdai Kim $^{2†}$", + "bbox": [ + 303, + 200, + 692, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ AI Advanced Technology, SK Hynix", + "bbox": [ + 346, + 219, + 648, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Statistics, Seoul National University", + "bbox": [ + 292, + 236, + 705, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ydy0415@gmail.com, rieky0426@snu.ac.kr, ydkim903@snu.ac.kr", + "bbox": [ + 238, + 255, + 751, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robust domain adaptation against adversarial attacks is a critical research area that aims to develop models capable of maintaining consistent performance across diverse and challenging domains. In this paper, we derive a new generalization bound for robust risk on the target domain using a novel divergence measure specifically designed for robust domain adaptation. Building upon this, we propose a new algorithm named TAROT, which is designed to enhance both domain adaptability and robustness. Through extensive experiments, TAROT not only surpasses state-of-the-art methods in accuracy and robustness but also significantly enhances domain generalization and scalability by effectively learning domain-invariant features. In particular, TAROT achieves superior performance on the challenging DomainNet dataset, demonstrating its ability to learn domain-invariant representations that generalize well across different domains, including unseen ones. These results highlight the broader applicability of our approach in real-world domain adaptation scenarios.", + "bbox": [ + 89, + 335, + 485, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 607, + 220, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in machine learning have highlighted the importance of developing transferable models that can maintain performance across different domains. However, collecting labeled data from all potential domains is too resource-intensive for many practical applications. To address this limitation, Unsupervised Domain Adaptation (UDA) has emerged as a type of transfer learning that leverages labeled source-domain data to improve performance on a target domain without needing labels from that domain. UDA has proven pivotal in diverse machine learning tasks, largely because distributional discrepancies between source and target domains often lead to severe performance degradation [4, 16, 24, 31, 34].", + "bbox": [ + 89, + 632, + 485, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meanwhile, adversarial robustness has gained significant attention as a crucial element of secure and trustworthy AI. Pioneering works demonstrated the vulnerability of neural", + "bbox": [ + 89, + 820, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "networks to small, carefully designed perturbations called adversarial attacks [21, 32] and proposed adversarial training methods [21, 38, 40, 41, 44]. Recently, the adversarial robustness community has expanded its focus to multimodal tasks (e.g., text-to-image, image-to-text), recognizing that large-scale multi-modal models are also susceptible to adversarial manipulations in one or more input modalities [15, 19, 22, 29]. These cross-modal attacks can exploit alignment mechanisms to trick models into misclassification or erroneous content generation.", + "bbox": [ + 511, + 306, + 908, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given these two developments—domain adaptation and adversarial robustness—there is a growing interest in learning transferable models in UDA settings that remain resilient to adversarial perturbations in the target domain [3, 14, 20, 46]. A common strategy in existing robust UDA methods is to rely on pseudo-labels to handle unlabeled target data. However, in this paper, we propose a novel robust margin disparity metric for comparing source and target distributions, enabling us to derive a refined generalization bound on the robust risk in the target domain. From our bound, we derive two key insights: (1) using only pseudo-labels may be insufficient, but combining pseudo-labels with distributional alignment significantly enhances robust UDA performance; and (2) initializing robust UDA algorithms with a robustly pre-trained model (Robust-PT [3]) is notably more effective than starting from a non-robust pre-trained model.", + "bbox": [ + 511, + 458, + 908, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building upon these theoretical foundations, we propose TAROT (Transfer Adversarially RObust Training), a new algorithm that leverages our theoretical insights to achieve superior performance in robust UDA tasks. Through extensive experiments on various benchmark datasets, we demonstrate that TAROT outperforms existing algorithms. Furthermore, we empirically demonstrate that TAROT exhibits distinct advantages in terms of scalability compared to pseudo-labeling approaches, particularly with significant performance improvements not only on the target domain but also on source and unseen source domains. This work bridges the gap between theoretical understanding and practical implementation in robust transfer learning, offering both fundamental insights and actionable solutions for real-world applications.", + "bbox": [ + 511, + 696, + 910, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 887, + 240, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25780", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 90, + 413, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel robust margin disparity measure and derive a refined generalization bound for robust UDA.", + "- From the derived generalization bound, we propose TAROT (Transferring Adversarial RObust Training), which integrates pseudo-labeling and distributional alignment to enhance adversarial robustness in UDA.", + "- TAROT outperforms existing robust UDA methods across multiple benchmarks, demonstrating strong performance on the target, source, and unseen source domains, indicating that TAROT effectively facilitates domain-invariant robust representations.", + "- We provide theoretical insights and practical solutions for scalable and adversarially robust domain adaptation in real-world scenarios." + ], + "bbox": [ + 89, + 107, + 482, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 325, + 228, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we outline the core concepts and notations for our analysis, beginning with adversarial training and its risk formulation for robust learning. We then describe the unsupervised domain adaptation (UDA) framework, highlighting distributional notations, the formal UDA objective, and existing theoretical approaches like DANN [13] and MDD [45]. These elements provide the groundwork for our subsequent derivation of a robust UDA upper bound and the proposed learning algorithm.", + "bbox": [ + 89, + 352, + 483, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Adversarial Training", + "text_level": 1, + "bbox": [ + 89, + 496, + 290, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $\\mathcal{X} \\in \\mathbb{R}^d$ represent the input space, and let $\\mathcal{Y} = \\{1, \\ldots, C\\}$ be the label space, where $C$ represents the number of classes. A classifier can be represented using a score function $f: \\mathcal{X} \\to \\mathbb{R}^C$ , where $f_c$ denotes the $c$ -th component of $f$ . The classifier is then defined as: $h_f(\\pmb{x}) = \\operatorname*{argmax}_{c \\in \\mathcal{Y}} f_c(\\pmb{x})$ . Let $\\mathcal{B}_p(\\pmb{x}, \\varepsilon) = \\{\\pmb{x}' \\in \\mathcal{X}: \\| \\pmb{x} - \\pmb{x}' \\|_p \\leq \\varepsilon\\}$ denote an $\\varepsilon$ -ball centered at $\\pmb{x}$ .", + "bbox": [ + 89, + 518, + 483, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The population risk considering adversarial robustness is defined as follows:", + "bbox": [ + 89, + 628, + 482, + 657 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\mathrm {r o b}} (f) = \\mathbb {E} _ {(\\mathbf {X}, Y)} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\mathbb {1} \\left\\{Y \\neq h _ {f} \\left(\\mathbf {X} ^ {\\prime}\\right) \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 672, + 483, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ denotes the indicator function. The goal of adversarial training is to estimate the optimal $f$ , the minimizer of Eq. (1).", + "bbox": [ + 89, + 710, + 483, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unsupervised Domain Adaptation", + "text_level": 1, + "bbox": [ + 89, + 767, + 387, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are two distributions, the source distribution $S$ and the target distribution $\\mathcal{T}$ . The empirical versions of $S$ and $\\mathcal{T}$ are $\\widehat{S} = \\{(x_i^s,y_i^s)\\}_{i = 1}^{n_s}$ and $\\widehat{\\mathcal{T}} = \\{(x_j^t,y_j^t)\\}_{j = 1}^{n_t}$ , respectively. In unsupervised domain adaptation, $\\widehat{S}$ and only $\\widehat{\\mathcal{T}}_{\\mathbf{X}} := \\{x_j^t\\}_{j = 1}^{n_t}$ are observed. i.e. the labels $y_{j}^{t}$ can not be observed. Denote the distributions of the input on the source and target domains by $S_{\\mathbf{X}}$ and $\\mathcal{T}_{\\mathbf{X}}$ , respectively.", + "bbox": [ + 89, + 790, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For a given score function $f \\in \\mathcal{F}$ , the risk of a data distribution $\\mathcal{D}$ is defined as", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\mathcal {D}} (f) := \\mathbb {E} _ {(\\mathbf {X}, Y) \\sim \\mathcal {D}} \\mathbb {1} \\{Y \\neq h _ {f} (\\mathbf {X}) \\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 130, + 906, + 148 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of UDA is to minimize $\\mathcal{R}_{\\mathcal{T}}(f)$ using $\\widehat{S}$ and $\\widehat{\\mathcal{T}}_{\\mathbf{X}}$ which can be done by building an upper bound of $\\mathcal{R}_{\\mathcal{T}}(h)$ .", + "bbox": [ + 511, + 157, + 906, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The pivotal learning algorithms for UDA are DANN (Domain-Adversarial Neural Network) [13] and MDD (Margin Disparity Discrepancy) [45]. Ben-David et al. [7] suggested a distance $d_{\\mathcal{H}\\Delta \\mathcal{H}}(\\mathcal{S}_{\\mathbf{X}},\\mathcal{T}_{\\mathbf{X}})$ between $\\mathcal{S}_{\\mathbf{X}}$ and $\\mathcal{T}_{\\mathbf{X}}$ to give an upper bound of the robust risk, and Ganin et al. [13] utilized this upper bound to implement their learning algorithm DANN. DANN uses an adversarial neural network to minimize the distance between the source and target distributions, using gradient reversal layers (GRL). On the other hand, MDD [45] derives an upper bound of $\\mathcal{R}_{\\mathcal{T}}(f)$ as the sum of $\\mathcal{R}_S^{(\\rho)}(f)$ , a standard non-robust risk on the source domain, and $d_{f,\\mathcal{F}}^{(\\rho)}(\\mathcal{S}_{\\mathbf{X}},\\mathcal{T}_{\\mathbf{X}})$ , a distance between $\\mathcal{S}_{\\mathbf{X}}$ and $\\mathcal{T}_{\\mathbf{X}}$ , and develop the MDD algorithm based on this upper bound. The main distinction between the two upper bounds is that MDD utilizes a scoring function for multi-class classification, while DANN only employs a classifier for binary classification.", + "bbox": [ + 511, + 189, + 906, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objective of our work is to derive an upper bound of the robust risk for the target domain by utilizing the risk in the source domain and the robust divergence between the target domain $\\mathcal{T}$ and the source domain $S$ . In turn, inspired by the upper bound, we propose a novel robust UDA algorithm.", + "bbox": [ + 511, + 441, + 906, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Related Works", + "text_level": 1, + "bbox": [ + 513, + 541, + 663, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While a significant body of research on adversarial training has focused on supervised learning and semi-supervised learning tasks [21, 36, 40, 41, 44], there has been relatively less attention given to its application in domain adaptation tasks. Only a few studies have explored adversarial training in the context of domain adaptation [3, 20, 46], highlighting the need for further investigation in this area.", + "bbox": [ + 511, + 566, + 905, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RFA [3] leverages a pretrained ImageNet model to develop a regularization term. When the trainable model is initialized with a pretrained robust model, the regularization term encourages the outputs of each layer of the trainable model to resemble those of the pretrained model. This facilitates the learning of robust features. This approach can be applied to existing UDA algorithms such as DANN [13] and MDD [45] to make them adversarially robust.", + "bbox": [ + 511, + 667, + 906, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ARTUDA [20] adopts a self-supervised adversarial training approach, where the predictions of the current model are used as pseudo-labels to generate adversarial examples for the target domain. It also generates adversarial examples for the source domain and applies the distance metrics used in DANN. However, there is no theoretical justification of the complex loss function used in ARTUDA. One key challenge in ARTUDA is that pseudo-labels are generated by the same", + "bbox": [ + 511, + 784, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25781", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model being trained. As a result, due to the inherent trade-off between robustness and generalization, as the model learns more robust features, pseudo-labels are getting more inaccurate. This issue becomes particularly pronounced when the perturbation budget $(\\varepsilon)$ is large.", + "bbox": [ + 89, + 90, + 483, + 164 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SRoUDA [46] attempts to resolve this limitation by introducing a separate teacher model to assign pseudo-labels to the target data. By leveraging meta-learning, SRoUDA improves the accuracy of pseudo-labeling by the teacher model and dynamically performs adversarial training [21] on the target domain using these pseudo-labels. Despite the improvement, SRoUDA struggles to effectively capture domain-invariant features, as it does not explicitly consider the deviation of the distributions between the source and target domains. This leads to a decline in performance in the source as well as in the unseen domains, indicating a failure to capture what we call essentially domain-invariant robust features.", + "bbox": [ + 89, + 164, + 483, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 366, + 264, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the theoretical results for adversarially robust UDA. This is an extension of MDD [45] to its robust version, and the proofs are deferred to Sec. 8.", + "bbox": [ + 89, + 391, + 483, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Robust Margin Disparity Discrepancy", + "text_level": 1, + "bbox": [ + 89, + 445, + 418, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this subsection, we introduce a novel disparity measure and key propositions to establish the generalization bound.", + "bbox": [ + 89, + 468, + 482, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denote the classifier $h_f$ for a given score function $f$ as", + "bbox": [ + 109, + 496, + 473, + 512 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh_{f}:\\boldsymbol {x}\\mapsto \\operatorname *{argmax}_{y\\in \\mathcal{Y}}f_{y}(\\boldsymbol {x}),\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 523, + 370, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and denote the margin of $f$ at a labeled sample $(x, y)$ as", + "bbox": [ + 89, + 561, + 464, + 578 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {f} (\\boldsymbol {x}, y) = f _ {y} (\\boldsymbol {x}) - \\max _ {y ^ {\\prime} \\neq y} f _ {y ^ {\\prime}} (\\boldsymbol {x}).\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 588, + 401, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1. (Robust Disparity). For given $f, f' \\in \\mathcal{F}$ , the 0-1 robust disparity between $f$ and $f'$ on a distribution $\\mathcal{D}_{\\mathbf{X}}$ on $\\mathcal{X}$ is defined by", + "bbox": [ + 89, + 642, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i s p} _ {\\mathcal {D} _ {\\mathbf {X}}} ^ {r o b} \\left(f ^ {\\prime}, f\\right) := \\mathbb {E} _ {\\mathbf {X} \\sim \\mathcal {D} _ {\\mathbf {X}}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\mathbb {1} \\left\\{h _ {f ^ {\\prime}} (\\mathbf {X}) \\neq h _ {f} \\left(\\mathbf {X} ^ {\\prime}\\right) \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 696, + 483, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The robust disparity is measured based on the adversarial robustness of the second model $f$ , while the first model $f'$ serves as a pseudo-label generator. Note that if the prediction of $f'$ is consistent to true label, it is the robust risk as defined in Eq. (1).", + "bbox": [ + 89, + 747, + 482, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As an extension of the margin loss in the non-robust setting [45] to its robust counterpart, we define the robust margin loss as", + "bbox": [ + 89, + 820, + 483, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\mathcal {D}} ^ {\\mathrm {r o b}, (\\rho)} (f) := \\mathbb {E} _ {(\\mathbf {X}, Y) \\sim \\mathcal {D}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\Phi_ {\\rho} \\circ \\mathcal {M} _ {f} \\left(\\mathbf {X} ^ {\\prime}, Y\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 875, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 92, + 558, + 104 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {\\rho} (x) := \\left\\{ \\begin{array}{l l} 0 & \\rho \\leq x \\\\ 1 - x / \\rho & 0 \\leq x \\leq \\rho \\\\ 1 & x \\leq 0 \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 102, + 906, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $\\circ$ is composite operator of functions.", + "bbox": [ + 511, + 162, + 784, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2. (Robust Margin Disparity). For given $f, f' \\in \\mathcal{F}$ , the Robust Margin Disparity is defined as", + "bbox": [ + 511, + 186, + 906, + 217 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {d i s p} _ {\\mathcal {D} _ {\\mathbf {X}}} ^ {r o b, (\\rho)} \\left(f ^ {\\prime}, f\\right) \\\\ := \\mathbb {E} _ {\\mathbf {X} \\sim \\mathcal {D} _ {\\mathbf {X}}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {\\rho} (\\mathbf {X}, \\varepsilon)} \\Phi_ {\\rho} \\circ \\mathcal {M} _ {f} \\left(\\mathbf {X} ^ {\\prime}, h _ {f ^ {\\prime}} (\\mathbf {X})\\right) \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 227, + 906, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An important property of the robust margin disparity is that it is greater than the robust disparity, i.e. $\\mathrm{disp}_{\\mathcal{D}_{\\mathbf{X}}}^{\\mathrm{rob},(\\rho)}(f', f) \\geq \\mathrm{disp}_{\\mathcal{D}_{\\mathbf{X}}}^{\\mathrm{rob}}(f', f)$ .", + "bbox": [ + 511, + 286, + 906, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now use this disparity term to bound the robust risk on the target distribution with additional feasible terms.", + "bbox": [ + 511, + 333, + 903, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. Let $S$ and $\\mathcal{T}$ represent the distributions of the source and target domains, respectively. Similarly, let $S_{\\mathbf{X}}$ and $\\mathcal{T}_{\\mathbf{X}}$ denote the marginal distributions of the source and target domains over $\\mathbf{X}$ , respectively. For every score function $f \\in \\mathcal{F}$ , the following inequality holds:", + "bbox": [ + 511, + 371, + 906, + 445 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {R} _ {T} ^ {r o b} (f) \\leq \\\\ \\mathcal {R} _ {\\mathcal {S}} ^ {(\\rho)} (f) + \\left\\{\\operatorname {d i s p} _ {\\mathcal {T} _ {\\mathbf {X}}} ^ {r o b, (\\rho)} \\left(f ^ {*}, f\\right) - \\operatorname {d i s p} _ {\\mathcal {S} _ {\\mathbf {X}}} ^ {(\\rho)} \\left(f ^ {*}, f\\right) \\right\\} + \\lambda , \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 454, + 905, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f^{*} = \\operatorname *{argmin}_{f\\in \\mathcal{F}}\\{\\mathcal{R}_{\\mathcal{T}}^{(\\rho)}(f) + \\mathcal{R}_{\\mathcal{S}}^{(\\rho)}(f)\\}$ is ideal hypothesis and $\\lambda = \\lambda (\\mathcal{F},\\mathcal{S},\\mathcal{T},\\varepsilon ,\\rho) = \\mathcal{R}_{\\mathcal{T}}^{(\\rho)}(f^{*}) + \\mathcal{R}_{\\mathcal{S}}^{(\\rho)}(f^{*})$ is constant of $f$", + "bbox": [ + 511, + 513, + 906, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that the upper bound consists of standard margin risk of the source domain, not the robust margin risk. This enables the corresponding algorithm introduced in Sec. 4.4 computationally simpler because generation of adversarial samples is not required to calculate the standard margin risk on the source domain.", + "bbox": [ + 511, + 582, + 905, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the above proposition, we can naturally induce the following discrepancy between the distributions of two domains, which we call as a robust margin disparity discrepancy.", + "bbox": [ + 511, + 669, + 906, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3. (Robust Margin Disparity Discrepancy).", + "text_level": 1, + "bbox": [ + 511, + 737, + 893, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a score function $f \\in \\mathcal{F}$ , the Robust Margin Disparity Discrepancy induced by $f' \\in \\mathcal{F}$ of two distributions $S_{\\mathbf{X}}$ and $T_{\\mathbf{X}}$ is defined as follows:", + "bbox": [ + 511, + 751, + 906, + 796 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d _ {f, \\mathcal {F}} ^ {r o b, (\\rho)} (S _ {\\mathbf {X}}, \\mathcal {T} _ {\\mathbf {X}}) \\\\ := \\sup _ {f ^ {\\prime} \\in \\mathcal {F}} \\left\\{\\operatorname {d i s p} _ {\\mathcal {T} _ {\\mathbf {x}}} ^ {r o b, (\\rho)} \\left(f ^ {\\prime}, f\\right) - \\operatorname {d i s p} _ {\\mathcal {S} _ {\\mathbf {x}}} ^ {(\\rho)} \\left(f ^ {\\prime}, f\\right) \\right\\} \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 806, + 906, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This discrepancy measures the difference of the robust margin disparity on target domain and the standard margin", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25782", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "disparity on the source domain, where the objective $f$ is fixed. If there is significant gap between the adversarial target domain and standard source domain, the discrepancy becomes large. Now we introduce a non-asymptotic upper bound for the target robust risk:", + "bbox": [ + 89, + 90, + 483, + 162 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\mathcal {T}} ^ {\\mathrm {r o b}, (\\rho)} (f) \\leq \\mathcal {R} _ {\\mathcal {S}} ^ {(\\rho)} (f) + d _ {f, \\mathcal {F}} ^ {\\mathrm {r o b}, (\\rho)} \\left(\\mathcal {S} _ {\\mathbf {X}}, \\mathcal {T} _ {\\mathbf {X}}\\right) + \\lambda \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 171, + 483, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with the same $\\lambda$ in Proposition 1.", + "bbox": [ + 89, + 198, + 312, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Generalization bound", + "text_level": 1, + "bbox": [ + 89, + 220, + 294, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we provide a generalization bound based on Eq. (9). We use the notion of local Lipschitz constant of a function, following Yang et al. [42].", + "bbox": [ + 89, + 241, + 483, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 4 (Local Lipschitz Constant). Consider an arbitrary real function $f: \\mathbb{R}^d \\to \\mathbb{R}^C$ , and a $L_p$ -norm $\\| \\cdot \\|_p$ . Define a point-wise local Lipschitz constant as", + "bbox": [ + 89, + 292, + 483, + 337 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f} (\\boldsymbol {x}, \\varepsilon) := \\sup _ {\\boldsymbol {x} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\boldsymbol {x}, \\varepsilon)} \\frac {\\| f \\left(\\boldsymbol {x} ^ {\\prime}\\right) - f (\\boldsymbol {x}) \\| _ {1}}{\\| \\boldsymbol {x} ^ {\\prime} - \\boldsymbol {x} \\| _ {p}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 342, + 483, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for a point $\\pmb{x}$ . Then, we define a local Lipschitz constant on $\\mathcal{D}_{\\mathbf{X}}$ as", + "bbox": [ + 88, + 385, + 482, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f} \\left(\\mathcal {D} _ {\\mathbf {X}}, \\varepsilon\\right) := \\sup _ {x \\in \\mathcal {D} _ {\\mathbf {X}}} L _ {f} (x, \\varepsilon) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 414, + 482, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We call that $f$ is locally Lipschitz, if $L_{f}(\\mathcal{D}_{\\mathbf{X}},\\varepsilon)$ is bounded (i.e. $L_{f}(\\mathcal{D}_{\\mathbf{X}},\\varepsilon) < \\infty$ ).", + "bbox": [ + 89, + 443, + 483, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also define following auxiliary function spaces.", + "bbox": [ + 109, + 479, + 447, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 5. Consider a set of score functions $\\mathcal{F}$ and let $\\mathcal{H}$ be the class of the induced classifiers $h_f$ for each $f\\in \\mathcal{F}$ . We define the following two sets.", + "bbox": [ + 89, + 500, + 483, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Pi_ {\\mathcal {H}} \\mathcal {F} := \\{x \\mapsto f (x, h (x)) | h \\in \\mathcal {H}, f \\in \\mathcal {F} \\} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 551, + 482, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Pi_ {1} \\mathcal {F} := \\{x \\mapsto f (x, y) | y \\in \\mathcal {Y}, f \\in \\mathcal {F} \\} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 570, + 482, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is known that $\\Pi_{\\mathcal{H}}\\mathcal{F}$ takes a form of the space of inner products of vector fields from $\\mathcal{H}$ to $\\mathcal{F}$ [11, 45]. The notion $\\Pi_1\\mathcal{F}$ follows from Mehryar Mohri and Talwalkar [23].", + "bbox": [ + 89, + 594, + 482, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now derive a generalization bound for the robust risk on the target domain using Rademacher complexity, which quantifies the expressiveness of a given hypothesis space [23].", + "bbox": [ + 89, + 638, + 482, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 6. (Rademacher Complexity). Let $\\mathcal{F}$ be a family of functions mapping from $\\mathcal{Z} = \\mathcal{X} \\times \\mathcal{Y}$ to $[a, b]$ and $\\widehat{\\mathcal{D}} = \\{z_1, \\dots, z_n\\}$ are finite samples independently drawn from a distribution $\\mathcal{D}$ on $\\mathcal{Z}$ . Then, the empirical Rademacher complexity of $\\mathcal{F}$ with $\\widehat{\\mathcal{D}}$ is defined by", + "bbox": [ + 89, + 702, + 483, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathfrak {R}} _ {\\widehat {\\mathcal {D}}} (\\mathcal {F}) \\triangleq \\mathbb {E} _ {\\sigma} \\sup _ {f \\in \\mathcal {F}} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sigma_ {i} f (z _ {i}), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 786, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{i}$ s are independent random variables following the Rademacher distribution. In turn, the Rademacher complexity of $\\mathcal{F}$ is defined by", + "bbox": [ + 89, + 830, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathfrak {R} _ {n, \\mathcal {D}} (\\mathcal {F}) \\triangleq \\mathbb {E} _ {\\widehat {\\mathcal {D}} \\sim \\mathcal {D} ^ {n}} \\widehat {\\mathfrak {R}} _ {\\widehat {\\mathcal {D}}} (\\mathcal {F}) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 883, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now we provide our generalization bound, which replaces the terms in the upper bound (Eq. (9)) that depend on the population distribution by the terms depending on the empirical distributions.", + "bbox": [ + 511, + 90, + 906, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. (Generalization Bound on the Robust Risk of Target Distribution). For any $\\delta > 0$ , with probability $1 - 3\\delta$ , we have the following uniform generalization bound for any score function $f$ in $\\mathcal{F}$ :", + "bbox": [ + 511, + 159, + 906, + 219 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {R} _ {\\mathcal {T}} ^ {r o b} (f) \\\\ \\leq \\mathcal {R} _ {\\widehat {\\mathcal {S}}} ^ {(\\rho)} (f) + d _ {f, \\mathcal {F}} ^ {r o b, (\\rho)} \\left(\\widehat {\\mathcal {S}} _ {\\mathbf {X}}, \\widehat {\\mathcal {T}} _ {\\mathbf {X}}\\right) + \\lambda \\\\ + \\frac {2 C ^ {2}}{\\rho} \\mathfrak {R} _ {n, S} (\\Pi_ {1} \\mathcal {F}) + \\frac {2 C}{\\rho} \\mathfrak {R} _ {n, S} (\\Pi_ {\\mathcal {H}} \\mathcal {F}) + 2 \\sqrt {\\frac {\\log 2 / \\delta}{2 n}} \\\\ + \\frac {2 C}{\\rho} \\Re_ {m, \\mathcal {T}} \\left(\\Pi_ {\\mathcal {H}} \\mathcal {F}\\right) + \\sqrt {\\frac {\\log 2 / \\delta}{2 m}} + \\frac {2 \\varepsilon L _ {f} \\left(\\mathcal {T} _ {\\mathbf {X}}, \\varepsilon\\right)}{\\rho}, \\tag {16} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 229, + 905, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda = \\min_{f\\in \\mathcal{F}}\\{\\mathcal{R}_{\\mathcal{T}}^{(\\rho)}(f) + \\mathcal{R}_{\\mathcal{S}}^{(\\rho)}(f)\\}$ .", + "bbox": [ + 513, + 358, + 767, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Two insights", + "text_level": 1, + "bbox": [ + 513, + 392, + 648, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we present two key insights for robust UDA. First, we analyze the impact of the local Lipschitz constant on generalization performance, showing why initializing robust UDA algorithms with a robustly pre-trained (Robust-PT) model can substantially improve outcomes. Second, we highlight why pseudo labeling alone may be insufficient, especially in large or complex target domains where confirmation bias and the lack of explicit domain alignment can lead to persistent misclassifications.", + "bbox": [ + 511, + 412, + 906, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "About Local Lipschitz constant In Theorem 1, there is the local Lipschitz constant $L_{f}(\\mathcal{T}_{\\mathbf{X}},\\varepsilon)$ . As discovered in previous literature, the local Lipschitz constant is known to be deeply involved with a generalization performance of a model [6, 17, 30]. This fact leads to various researches that investigate a method to insure provable robustness of a model [9, 35, 39]. It is also known that adversarial training empirically reduces the local Lipschitz constant of a neural network [42, 43].", + "bbox": [ + 511, + 561, + 906, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the numerical studies, we have found that robust UDA approaches with a Robust-PT as an initial improve the prediction performance much compared to those without a Robust-PT as an initial. This phenomenon can be partly explained by the local Lipschitz term. That is, model which is not initialized with Robust-PT could have a large value of the local Lipschitz constant and thus could have a too large robust risk value on the target domain. Hence, using an initial solution with a small local Lipschitz constant is promising and a Robust-PT would be such one.", + "bbox": [ + 511, + 693, + 908, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pseudo labeling only may not be sufficient Pseudo labeling is a widely used technique in domain adaptation [20, 46], where the objective is to improve the performance of the", + "bbox": [ + 511, + 856, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model on the target domain by iteratively using the predictive labels of the currently trained model (or its variation) as labels for the target domain data when generating adversarial examples for the target data. There are at least two problems in those pseudo labeling approaches.", + "bbox": [ + 88, + 90, + 482, + 164 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first problem is a risk of confirmation bias. Since the model generates pseudo labels based on its own predictions, early-stage errors in the model could become self-reinforcing. When incorrect pseudo labels are assigned with high confidence, these mistakes persist through training cycles, amplifying errors and leading to poor generalization on the target domain. This phenomenon is widely studied in semi-supervised settings [1, 8]. Similarly in the UDA setting, as the training phase progresses, the model may increasingly rely on those incorrect predictions, making it difficult to recover especially in complex or noisy target domains.", + "bbox": [ + 88, + 164, + 482, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The risk of confirmation bias can be reduced by generating pseudo labels from a pretrained non-robust UDA model called a (fixed) teacher model. In this case, we can write the objective function as", + "bbox": [ + 89, + 323, + 483, + 381 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathbf {X} \\sim \\mathcal {T} _ {\\mathbf {X}}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\mathbb {1} \\left\\{h _ {f} \\left(\\mathbf {X} ^ {\\prime}\\right) \\neq h _ {f *} (\\mathbf {X}) \\right\\} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 391, + 483, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f^{*} = \\operatorname*{argmin}_{f\\in \\mathcal{F}}\\mathcal{R}_{\\widehat{\\mathcal{T}}}(\\boldsymbol {f})$ . In our notation, this term exactly coincides with $\\mathrm{disp}_{T_{\\mathbf{x}}}^{\\mathrm{rob}}(f^{*},f)$ . The following proposition justifies this approach.", + "bbox": [ + 89, + 426, + 483, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2. For every score function $f \\in \\mathcal{F}$ , the following inequality holds:", + "bbox": [ + 89, + 491, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\mathcal {T}} ^ {r o b} (f) \\leq \\operatorname {d i s p} _ {\\mathcal {T} _ {\\mathbf {X}}} ^ {r o b} \\left(f ^ {*}, f\\right) + \\mathcal {R} _ {\\mathcal {T}} \\left(f ^ {*}\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 529, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for arbitrary $f^{*}\\in \\mathcal{F}$", + "bbox": [ + 86, + 556, + 235, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It can be easily seen through the intermediate steps of Proposition 1. By letting $f^{*}$ as a non-robust UDA model, the pseudo labeling approach with the teacher model $f^{*}$ can be understood to minimize the upper bound in Proposition 2. Our empirical studies amply confirm this insight.", + "bbox": [ + 88, + 580, + 483, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo labeling approaches with a fixed teacher model, however, still has limitations. In particular, the scalability of pseudo labeling across diverse or large datasets can be challenging due to the absence of explicit domain alignment mechanisms. As the size or complexity of the target domain grows, the chance of incorrect pseudo labels increases, which could overwhelm the overall learning process. This may happen especially in scenarios where the source and target domains differ significantly. This makes it harder for the model to generalize to more extensive or varied datasets, particularly when the domain gap becomes larger due to increasing size or diversity of data.", + "bbox": [ + 88, + 654, + 482, + 827 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that our generalization bound Theorem 1 uses the robust margin disparity discrepancy between the source and target domains instead of the robust disparity only on the target domain in Proposition 2. That is, our generalization bound integrates pseudo labeling and domain alignment", + "bbox": [ + 88, + 827, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "together. In the following subsection, inspired by these insights, we introduce an algorithm designed to overcome the drawbacks of pseudo labeling approaches, tackling both confirmation bias and scalability while enhancing generalization abilities across diverse and complex target domains.", + "bbox": [ + 511, + 90, + 906, + 164 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Transferring Adversarily Robust Training (TAROT)" + ], + "code_body": "Input : $\\psi$ : feature extractor, $\\pi$ : head, $\\pi'$ : auxiliary head, $s$ : standard trained model, $\\mathcal{D}_{\\mathcal{S}} = \\{(x_{i}^{s},y_{i}^{s})\\}_{i = 1}^{m}$ : source domain dataset, $\\mathcal{D}_{\\mathcal{S}} = \\{(x_{i}^{s},y_{i}^{s})\\}_{i = 1}^{m}:\\mathcal{D}_{\\mathcal{T}} = \\{x_{j}^{t}\\}_{j = 1}^{m^{\\prime}}$ : target domain dataset, $\\eta_1,\\eta_2,\\lambda ,\\rho$ : learning rates in Eq. (20), $T$ : the number of epochs, $B$ : the number of batches, $K$ : batch size Output: adversarially robust network $\\pi \\circ \\psi$ \n1 $\\theta_{\\psi}\\gets \\theta_{\\mathrm{R}}$ (initialized by a robustly trained feature extractor) \n2 for $b = 1,\\dots ,B$ do \n3 for $k = 1,\\dots ,K$ do \n4 Generate $\\widehat{\\pmb{x}}_{b,k}^{\\mathrm{t,adv}}$ by PGD $(\\pi \\circ \\psi (\\pmb{x}_{b,k}^{t}),\\tilde{y}_{b,k}^{t})$ $\\tilde{y}_{b,k}^{t} = h_{s}(\\pmb{x}_{b,k}^{t})$ \n5 Update $(\\theta_{\\psi},\\theta_{\\pi})\\gets (\\theta_{\\psi},\\theta_{\\pi})-$ $\\eta_1\\frac{1}{K}\\nabla_{(\\pmb {\\theta}_\\psi ,\\pmb {\\theta}_\\pi)}\\ell_{\\mathrm{TAROT}}(\\pmb {\\theta};\\{(x_{b,k}^{\\mathrm{s}},y_{b,k}^{\\mathrm{s}})\\}_{k = 1}^{K},\\{(x_{b,k}^{\\mathrm{t}})\\}_{k = 1}^{K})$ \n6 Update $\\theta_{\\pi '}\\gets \\theta_{\\pi ' + }$ $\\eta_2\\frac{1}{K}\\nabla_{\\pmb{\\theta}_{\\pi '}}\\ell_{\\mathrm{TAROT}}(\\pmb {\\theta};\\{(x_{b,k}^{\\mathrm{s}},y_{b,k}^{\\mathrm{s}})\\}_{k = 1}^{K},\\{(x_{b,k}^{\\mathrm{t}})\\}_{k = 1}^{K},\\lambda)$ \n7 end \n8 end \n9 Return $\\pi \\circ \\psi$", + "bbox": [ + 519, + 213, + 924, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Learning Algorithm", + "text_level": 1, + "bbox": [ + 513, + 539, + 702, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by Theorem 1, we propose the robust domain adaptation algorithm, which we call Transferring Adversarially RObust Training (TAROT). Our basic idea is to find a optimal score function $f \\in \\mathcal{F}$ by solving a modified version of the following optimization problem:", + "bbox": [ + 511, + 560, + 906, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {f \\in \\mathcal {F}} \\mathcal {R} _ {\\widehat {\\mathcal {S}}} ^ {(\\rho)} (f) + d _ {f, \\mathcal {F}} ^ {\\operatorname {r o b}, (\\rho)} \\left(\\widehat {\\mathcal {S}} _ {\\mathbf {X}}, \\widehat {\\mathcal {T}} _ {\\mathbf {X}}\\right) + L _ {f} \\left(\\widehat {\\mathcal {T}} _ {\\mathbf {X}}, \\varepsilon\\right) \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 642, + 906, + 667 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- As is done in DANN (Domain-Adversarial Neural Network) [13] and MDD (Margin Disparity Discrepancy) [45], we consider a neural network structure that is composed of a feature extractor $\\psi$ and a head $\\pi$ , i.e. a score function $f \\in \\mathcal{F}$ can be represented as $\\pi \\circ \\psi$ . We only consider $\\pi$ with $\\psi$ being fixed, when calculating $d_{f,\\mathcal{F}}^{\\mathrm{rob},(\\rho)}$ . That is, we calculate the robust margin disparity discrepancy on the feature space of $\\psi$ .", + "- To reduce the first and second term, we follow the formulation of the MDD algorithm. Likewise, we utilize an auxiliary head $\\pi'$ and $\\gamma \\coloneqq \\exp \\rho$ designed to attain the margin. $\\pi'$ shares the same hypothesis space with $\\pi$ , and works as an adversarial network. The discrepancy term, which is non-differentiable, can be optimized through GRL [12]. The only difference is that, when to calculate $d_{f,\\mathcal{F}}^{\\mathrm{rob},(\\rho)}$ ," + ], + "bbox": [ + 511, + 676, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25784", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/174ce636fe189adfb9db2e7411e5b4fac1ad994f24129b9774bde2a0ab86f855.jpg", + "image_caption": [ + "Figure 1. Overview of TAROT algorithm." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 486, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "it requires calculating a robust disparity and a standard disparity rather than two standard disparities.", + "bbox": [ + 102, + 316, + 482, + 347 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To reduce the local Lipschitz constant, we minimize the cross entropy loss with adversarial examples for given pseudo labels attached by a teacher model learned by a nonrobust UDA approach. This is based on the previously mentioned phenomenon that adversarial training reduces the local Lipschitz constant [42, 43].", + "- When calculating the robust margin disparity discrepancy, instead of generating adversarial examples, we reuse those generated for the local Lipsshitz constant, making the algorithm computationally effective.", + "- As an initial model for $\\psi$ , we used a robustly pretrained models with from the repository provided in [28]." + ], + "bbox": [ + 89, + 347, + 483, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The objective function of TAROT is given as:", + "bbox": [ + 109, + 530, + 411, + 545 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\ell_ {\\text {T A R O T}} (\\boldsymbol {\\theta}; \\left\\{\\left(\\boldsymbol {x} _ {i} ^ {\\mathrm {s}}, y _ {i} ^ {\\mathrm {s}}\\right) \\right\\} _ {i = 1} ^ {m}, \\left\\{\\left(\\boldsymbol {x} _ {j} ^ {\\mathrm {t}}\\right) \\right\\} _ {j = m + 1} ^ {m + m ^ {\\prime}}, \\alpha , \\gamma) \\\\ = \\alpha \\left\\{ \\right.\\frac {1}{m} \\sum_ {i = 1} ^ {m} \\ell_ {\\mathrm {c e}} \\left(\\left(\\pi \\circ \\psi\\right)\\left(\\boldsymbol {x} _ {i}\\right), y _ {i}\\right)\\left. \\right)\\left. \\right. \\\\ + \\left(\\frac {1}{m ^ {\\prime}} \\sum_ {j = m + 1} ^ {m + m ^ {\\prime}} \\ell_ {\\mathrm {m o d - c e}} ^ {\\mathrm {r o b}} \\left(\\left(\\pi^ {\\prime} \\circ \\psi\\right) \\left(\\boldsymbol {x} _ {j}\\right), h _ {\\pi \\circ \\psi} \\left(\\boldsymbol {x} _ {j}\\right)\\right) \\right. \\tag {20} \\\\ \\left. \\left. - \\frac {\\gamma}{m} \\sum_ {j = 1} ^ {m} \\ell_ {\\mathrm {c e}} \\left(\\left(\\pi^ {\\prime} \\circ \\psi\\right) \\left(\\boldsymbol {x} _ {j}\\right), h _ {\\pi \\circ \\psi} \\left(\\boldsymbol {x} _ {j}\\right)\\right)\\right) \\right\\} \\\\ + \\frac {1}{m ^ {\\prime}} \\sum_ {j = m + 1} ^ {m + m ^ {\\prime}} \\ell_ {\\mathrm {c e}} ^ {\\mathrm {r o b}} ((\\pi \\circ \\psi) (\\boldsymbol {x} _ {j}), \\tilde {y} _ {j}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 564, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\ell_{\\mathrm{ce}}$ , $\\ell_{\\mathrm{mod - ce}}^{\\mathrm{rob}}$ [45] and $\\ell_{\\mathrm{ce}}^{\\mathrm{rob}}$ [21] as cross-entropy, modified cross-entropy with adversarial examples and cross-entropy with adversarial examples respectively. Also, $\\pmb{\\theta} = (\\pmb{\\theta}_{\\psi},\\pmb{\\theta}_{\\pi},\\pmb{\\theta}_{\\pi'})$ to represent the parameters corresponding to $\\psi, \\pi, \\pi'$ . $\\tilde{y}_j = h_s(\\pmb{x}_j)$ is a pseudo label generated by a standard trained model $s$ and $\\alpha$ is a trade-off parameter that balances transferability and target's robustness. TAROT algorithm is summarized in Algorithm 1 and Figure 1.", + "bbox": [ + 89, + 782, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 89, + 645, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare TAROT with other competitors for adversarially robust UDA. We use several benchmark datasets: Office-31 [27], Office-Home [37], VisDA2017 [25], and DomainNet [26]. Throughout these datasets, effectiveness of TAROT is evaluated across a wide range of domain shifts and robustness scenarios. Further experimental details including explanation for datasets and experimental setups are presented in Sec. 9. The code is available in the repository $\\ddagger$ .", + "bbox": [ + 511, + 114, + 906, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Performance Evaluation", + "text_level": 1, + "bbox": [ + 511, + 256, + 736, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Setup All of the results are averaged over three runs, with different teacher model initializations. Different perturbation budgets $(\\varepsilon)$ are chosen based on the difficulty of each benchmark dataset. For more challenging datasets, a smaller $\\varepsilon$ is used, and vice versa. Specifically, we use $\\varepsilon = 16 / 255$ for the Office-31 and Office-Home datasets, while $\\varepsilon = 8 / 255$ and $\\varepsilon = 4 / 255$ are used for VisDA2017 and DomainNet, respectively. The values of $\\alpha$ in Eq. (20) are set to 0.05, 0.1, 0.5, and 1.0 for Office31, OfficeHome, VisDA2017, and DomainNet, respectively. We empirically observed that the optimal $\\alpha$ values for each task vary much. The sensitivity analysis of $\\alpha$ is provided in Sec. 5.2.", + "bbox": [ + 511, + 277, + 906, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As baselines, along with ARTUDA [20], RFA [3], and SRoUDA [46], we consider the pseudo labeling (PL) algorithm which is implemented as a simple PGD-AT algorithm where target pseudo labels are obtained from a fixed non-robustly trained teachers learned by MDD [45]. All algorithms are initialized with Robust-PT as is done for TAROT. Without Robust-PT, training with larger perturbations such as $\\varepsilon = 16 / 255$ can lead to non-decreasing training loss in certain tasks, as discussed in Sec. 10.2.", + "bbox": [ + 511, + 452, + 908, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Setup We report standard accuracies on clean samples and robust accuracies against AutoAttack [10]. AutoAttack, which includes both white-box and black-box attacks, is a powerful evaluation tool that mitigates errors in robustness measurement caused by gradient obfuscation—a phenomenon where model gradients are distorted or masked, leading to overestimated robustness [2]. AutoAttack is widely recognized for accurately assessing adversarial robustness by bypassing these gradient masking issues.", + "bbox": [ + 511, + 585, + 908, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.1. Performance on Target Domains", + "text_level": 1, + "bbox": [ + 511, + 726, + 785, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tables 1, 2, 4 and 5 demonstrate that PL outperforms other algorithms, with the exception of TAROT. As seen in Figure 3, simply combining Robust-PT as an initial solution with pseudo labeling can yield significant performance gains, which confirms our first theoretical insight. (See Sec. 5.2.) This combination serves as a strong baseline in adversarially robust domain adaptation. It's noteworthy that ARTUDA [20], RFA [3], and SRoUDA [46] algorithms show remarkably poor performance on several tasks. Note that the robust", + "bbox": [ + 511, + 744, + 908, + 876 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "https://github.com/dyoony/TAROT", + "bbox": [ + 529, + 886, + 784, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/75fa8586d56bd814ae2cf20d2303ec4f784cb2933556e504f2bbce91817ad281.jpg", + "table_caption": [ + "Table 1. Performances of ARTUDA, RFA, SROUDA, PL and TAROT on Office31 ( $\\varepsilon = 16/255$ ). In each cell, the first number is the standard accuracy (\\%), while the second number is the robust accuracy (\\%) for AA. Bold numbers indicate the best performance." + ], + "table_footnote": [], + "table_body": "
MethodA → DA → WD → AD → WW → AW → DAvg.
ARTUDA23.49 / 11.6522.39 / 9.5635.00 / 16.5168.55 / 35.8536.81 / 19.7474.10 / 51.6143.39 / 24.15
RFA78.72 / 1.2171.20 / 1.2655.84 / 18.0396.73 / 6.4260.10 / 23.0099.40 / 6.0277.00 / 9.32
SRoUDA93.37 / 42.3778.87 / 65.1617.93 / 3.4496.00 / 44.2866.53 / 47.21100.00 / 62.4574.62 / 44.15
PL91.77 / 84.1491.82 / 91.7074.19 / 67.8797.99 / 95.9872.42 / 65.46100.00 / 92.5788.03 / 82.95
TAROT94.18 / 84.1495.47 / 93.4674.51 / 67.1397.99 / 96.2373.73 / 66.88100.00 / 91.7789.31 / 83.27
", + "bbox": [ + 145, + 121, + 849, + 196 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4abba62c153644c41c883861282521a3a8cbcf05245b183dbf2af94fb3b26c30.jpg", + "table_caption": [ + "Table 2. Performances of ARTUDA, RFA, SRoUDA, PL and TAROT on OfficeHome (ε = 16/255). In each cell, the first number is the standard accuracy (%), while the second number is the robust accuracy (% for AA. Bold numbers indicate the best performance." + ], + "table_footnote": [], + "table_body": "
MethodAr → ClAr → PrAr → RwCl → ArCl → PrCl → Rw
ARTUDA30.29 / 13.5626.18 / 4.755.88 / 0.003.42 / 0.0025.34 / 10.4820.98 / 5.53
RFA48.39 / 17.8954.67 / 8.7663.74 / 6.3443.22 / 3.1758.03 / 9.1557.24 / 5.69
SRoUDA51.84 / 37.8975.60 / 45.4674.25 / 40.2857.97 / 33.4267.76 / 54.7767.32 / 35.83
PL55.72 / 44.3173.53 / 58.4973.10 / 40.2657.68 / 31.1570.31 / 56.2168.14 / 37.37
TAROT55.72 / 46.5373.85 / 58.9377.78 / 42.6260.12 / 34.6973.08 / 58.1469.93 / 38.95
Pr → ArPr → ClPr → RwRw → ArRw → ClRw → PrAvg.
ARTUDA8.74 / 1.7338.92 / 14.5747.53 / 5.3030.12 / 2.1041.79 / 22.2945.19 / 13.9727.03 / 7.86
RFA42.03 / 2.5647.86 / 13.3163.44 / 5.2154.27 / 3.3854.27 / 16.7072.85 / 9.7855.00 / 8.49
SRoUDA59.21 / 31.4048.52 / 39.7775.40 / 41.2070.87 / 40.3853.40 / 42.4383.26 / 67.3857.97 / 33.42
PL57.15 / 29.5852.74 / 44.4074.82 / 40.7468.97 / 38.2857.03 / 45.1782.86 / 66.6266.00 / 44.38
TAROT59.62 / 32.7654.98 / 47.0178.66 / 42.9471.90 / 40.7959.45 / 49.7484.39 / 68.5368.29 / 46.80
", + "bbox": [ + 148, + 237, + 846, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "accuracies of a standard training algorithm (MDD) on the Office-31, Office-Home, and VisDA-2017 datasets are exactly 0.0, highlighting the importance of robust adaptation methods. This is mainly attributed to our use of larger perturbation budgets during training and evaluation compared to those used in the original papers. For reference, we provide results using smaller perturbation budgets $(4/255, 8/255)$ on Office-31 and OfficeHome datasets in the Appendix. Overall, TAROT outperforms existing algorithms across all benchmark datasets. Note that the robust accuracies of a standard training algorithm (MDD) on the Office-31, Office-Home, and VisDA-2017 datasets are exactly 0.0.", + "bbox": [ + 88, + 397, + 485, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.2. Essentially Domain-Invariant Robustness", + "text_level": 1, + "bbox": [ + 89, + 609, + 426, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we empirically highlight the limitations of the pseudo labeling approach, which focuses solely on improving target domain performance without effectively learning domain-invariant features. We select tasks with Real world (Rw) as the target domain on the OfficeHome dataset. We compare TAROT (our algorithm), which is designed to reduce the distribution gap between source and target domains, with PL, evaluating both the source domain accuracy and the performance on unseen domains. Table 3 shows that our method not only adapts well to the target domain but also learns more generalizable and robust features, leading to improved robustness across domains by significant margins. We present an upper bound for robust risk on the source domain and demonstrate that the components of this bound closely align with those of the target domain, offering a partial theoretical explanation for TAROT's robustness to adversarial attacks on the source domain. Additional results for other tasks are provided in Sec. 10.", + "bbox": [ + 91, + 638, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7a0d9aa3ba660fae36ef210edebc235b2ea86162e3a9b51fefa6975aa9c3ee2a.jpg", + "table_caption": [ + "Table 3. Performances of PL and TAROT on Source Domain and Unseen Domain, on OfficeHome ( $\\varepsilon = 8/255$ ). In each cell, the first number is the standard accuracy (\\%), while the second number corresponds to the robust accuracy (\\%) for AA. Bold numbers indicate the best performance. A $\\rightarrow$ B(C) indicates that A, B, and C represent the source, (unlabeled) target, and evaluation domains, respectively." + ], + "table_footnote": [], + "table_body": "
SourceUnseen
MethodAr → Rw(Ar)Ar → Rw(Cl)Ar → Rw(Pr)Avg.
ARTUDA18.83 / 2.647.70 / 0.896.28 / 0.4510.94 / 1.33
RFA99.62 / 47.0242.52 / 21.5351.57 / 21.5164.57 / 30.02
SRoUDA39.39 / 17.4339.54 / 28.6650.75 / 35.8642.23 / 27.32
PL41.78 / 19.4141.97 / 31.3952.42 / 37.2445.39 / 29.34
TAROT98.35 / 64.2447.86 / 35.1958.32 / 39.5468.18 / 46.32
Cl → Rw(Cl)Cl → Rw(Ar)Cl → Rw(Pr)Avg.
ARTUDA84.81 / 55.8014.34 / 3.2127.26 / 13.9942.14 / 24.33
RFA98.35 / 80.1040.37 / 8.8654.86 / 22.2564.53 / 37.07
SRoUDA48.29 / 35.5833.87 / 15.6248.19 / 33.0743.45 / 28.09
PL48.75 / 35.9536.30 / 16.6950.80 / 35.6645.28 / 29.43
TAROT93.65 / 84.3539.72 / 17.1855.46 / 37.6762.95 / 46.40
Pr → Rw(Pr)Pr → Rw(Ar)Pr → Rw(Cl)Avg.
ARTUDA98.29 / 60.1725.67 / 4.6236.24 / 16.2453.40 / 27.01
RFA99.75 / 68.5940.67 / 7.8741.76 / 19.7360.73 / 32.06
SRoUDA61.48 / 41.4034.82 / 16.0742.08 / 30.6846.13 / 29.38
PL61.05 / 43.3436.09 / 17.1042.11 / 30.9146.42 / 30.45
TAROT96.13 / 82.4339.14 / 16.1546.87 / 33.0160.71 / 43.86
", + "bbox": [ + 514, + 493, + 921, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 695, + 679, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies about: (i) effect of Robust-PT (ii) sensitivity analysis of $\\alpha$ . All exact values and additional experimental results are presented in Sec. 10.", + "bbox": [ + 511, + 717, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Robust-PT on Various $\\varepsilon$ Figures 3a and 3b show that the both algorithms perform significantly better when Robust-PT is applied, particularly as $\\varepsilon$ increases, indicating that Robust-PT is crucial for robustness against strong attack. Without Robust-PT, both TAROT and PL experience sharp declines in accuracy, especially as $\\varepsilon$ becomes larger. Figures 3c and 3d illustrate that TAROT uniformly outperforms PL across different perturbation budgets $(\\varepsilon)$ both on", + "bbox": [ + 511, + 784, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25786", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eaf3562e92faf3a94bd5d11ee4720993c2994ed0e9bbf1eea171f232e564b6a6.jpg", + "table_caption": [ + "Table 4. Performances of ARTUDA, RFA, SROUDA, PL and TAROT on DomainNet $(\\varepsilon = 4 / 255)$ . In each cell, the first number is the standard accuracy $(\\%)$ , and the second number is the robust accuracy $(\\%)$ for AA. Bold numbers indicate the best performance." + ], + "table_footnote": [], + "table_body": "
MethodC → IC → PC → RC → SAvg.
ARTUDA12.64 / 3.7631.15 / 7.9044.90 / 16.4939.90 / 17.5832.15 / 11.43
RFA9.42 / 3.4723.33 / 7.6037.54 / 17.7427.40 / 12.1724.42 / 10.24
SRoUDA9.27 / 5.6824.33 / 15.0742.91 / 28.3832.42 / 23.8227.23 / 18.24
PL10.59 / 6.6927.06 / 15.7343.57 / 28.6832.13 / 23.5228.34 / 18.66
TAROT14.65 / 8.1633.23 / 17.3650.02 / 30.4640.11 / 27.7734.50 / 20.94
I → CI → PI → RI → SAvg.
ARTUDA31.89 / 12.0720.38 / 2.5623.79 / 5.4734.61 / 5.9027.67 / 6.50
RFA15.00 / 8.8214.80 / 4.8019.15 / 8.7611.44 / 5.0515.10 / 6.86
SRoUDA25.95 / 14.1824.27 / 13.9336.32 / 22.0620.08 / 13.5226.65 / 15.92
PL17.45 / 13.1622.42 / 13.4132.50 / 21.9418.25 / 13.1422.65 / 15.41
TAROT29.36 / 20.2327.66 / 14.3039.14 / 23.3921.89 / 14.0829.51 / 18.00
P → CP → IP → RP → SAvg.
ARTUDA42.89 / 20.5814.36 / 3.2149.03 / 16.6234.02 / 13.1235.08 / 13.38
RFA26.27 / 14.859.45 / 3.0937.47 / 17.0419.67 / 8.5923.22 / 10.89
SRoUDA36.69 / 28.2610.17 / 5.8645.69 / 29.9530.95 / 19.1430.87 / 20.80
PL33.86 / 26.6911.18 / 7.2544.56 / 29.4729.76 / 20.6629.84 / 21.02
TAROT41.27 / 30.5114.82 / 7.6951.10 / 30.7637.22 / 24.7136.10 / 23.42
R → CR → IR → PR → SAvg.
ARTUDA52.94 / 28.1016.94 / 4.1741.95 / 12.0236.11 / 14.8636.98 / 14.79
RFA36.72 / 19.3012.36 / 3.0232.03 / 8.3422.73 / 8.1225.96 / 9.70
SRoUDA44.33 / 34.3310.10 / 5.6340.26 / 23.2629.10 / 18.5630.94 / 20.44
PL42.03 / 33.2711.85 / 7.1439.26 / 22.8029.05 / 19.6830.55 / 20.72
TAROT52.26 / 37.5818.86 / 9.4547.62 / 23.4339.38 / 25.2339.53 / 23.92
S → CS → IS → PS → RAvg.
ARTUDA54.24 / 30.5913.14 / 3.8934.87 / 7.7941.40 / 13.4735.91 / 13.94
RFA33.87 / 19.949.47 / 3.5427.29 / 7.8330.39 / 13.7025.26 / 11.25
SRoUDA46.62 / 36.658.50 / 5.2831.09 / 15.1442.95 / 22.3332.29 / 19.85
PL42.48 / 34.668.95 / 6.1831.48 / 18.7542.55 / 28.2431.37 / 21.96
TAROT53.35 / 41.3813.71 / 7.5239.77 / 20.4348.77 / 30.0238.90 / 24.84
", + "bbox": [ + 101, + 133, + 624, + 494 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b00a7bb923545dce8c7a68edd670233c2f49c9941898ad8705778d0b30dec4de.jpg", + "table_caption": [ + "Table 5. Performances of ARTUDA, RFA, SRoUDA, PL and TAROT on VisDA2017 $(\\varepsilon = 8 / 255)$ . Standard accuracy $(\\%)$ / Robust accuracy $(\\%)$ for AA." + ], + "table_footnote": [], + "table_body": "
MethodSyn. → RealSyn.
ARTUDA10.73 / 0.0061.43 / 5.70
RFA64.86 / 8.6599.41 / 33.55
SRoUDA24.44 / 0.2048.56 / 24.05
PL66.71 / 37.8340.98 / 24.09
TAROT67.12 / 37.7785.18 / 51.21
", + "bbox": [ + 648, + 143, + 893, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3b8f55ff8b96c4338004d1955886489d7ca9f863c8293509ef079c1dc666455.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 234, + 908, + 353 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96b420a4ef224c89539bcbab4a0c5d3cd7490df450632d87d181ee64d13b631c.jpg", + "image_caption": [ + "Figure 2. Sensitivity Analysis of $\\alpha$ . $\\alpha = 0$ corresponds to PL." + ], + "image_footnote": [], + "bbox": [ + 648, + 354, + 906, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the standard accuracies and robust accuracies against AA. Especially, that performance gap is increasing as $\\varepsilon$ increases. Figures 3a and 3b show that both algorithms perform significantly better when Robust-PT is applied, particularly as $\\varepsilon$ increases, indicating that Robust-PT is crucial for robustness against strong attacks. Figures 3c and 3d illustrate that TAROT uniformly outperforms the PL approach across different perturbation budgets $(\\varepsilon)$ , both in standard accuracies and robust accuracies against AA. Notably, the performance gap increases as $\\varepsilon$ increases.", + "bbox": [ + 89, + 513, + 483, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sensitivity Analysis of $\\alpha$ Figure 2 illustrates the sensitivity with respect to $\\alpha$ in Eq. (20). It shows how the performance of generalization and robustness changes for the target and unseen domains as the value of $\\alpha$ varies. As shown in the figure, each task has a different optimal $\\alpha$ , indicating that the optimal $\\alpha$ value for one task may not be optimal for another. The detailed numbers are provided in Sec. 10.3.", + "bbox": [ + 89, + 683, + 483, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions and Future Works", + "text_level": 1, + "bbox": [ + 89, + 801, + 375, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To our best knowledge, this is the first work for adversarially robust domain adaptation algorithms grounded with theoretical explanations. Compared to existing works, our algorithm shows essentially domain invariant robust features, resulting significant improved performance in source and", + "bbox": [ + 89, + 827, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2807fb685b1d7e9c8a842740ad4a1ccb52e9364942b8b28014dd38e90f2b0f1.jpg", + "image_caption": [ + "(a) PL w/ and w/o R-PT" + ], + "image_footnote": [], + "bbox": [ + 517, + 512, + 689, + 604 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aafed76dc8d0f073ede67c38a8a7cbfeed5c6cc481d8bdb7f4cb325c3594f27f.jpg", + "image_caption": [ + "(c) Rob Acc. w/ R-PT" + ], + "image_footnote": [], + "bbox": [ + 517, + 622, + 689, + 714 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f7c223d0d73c962e58b3e88a13ba6652ceada2cb71175daf27c0fba7d73bcfe.jpg", + "image_caption": [ + "(b) TAROT w/ and w/o R-PT" + ], + "image_footnote": [], + "bbox": [ + 730, + 512, + 901, + 603 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6378e763a5c4c74ae86c8697fc4c490b20a2b2ae6acb9f3f2eceefbee0b679f3.jpg", + "image_caption": [ + "(d) Stand Acc. w/ R-PT", + "Figure 3. Effect of Robust-PT across various methods and $\\varepsilon$ values. Each plot compares different configurations of PL w/ MDD and TAROT for standard and robust accuracy." + ], + "image_footnote": [], + "bbox": [ + 730, + 622, + 901, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "unseen domains. Our work not only serves as a cornerstone for understanding robust domain adaptation from a theoretical perspective, but we also believe TAROT can be served as a state-of-the-art algorithm in practical perspective by excelling researches of adversarial robustness for transfer learning such as domain adaptation/generalization and few-shot learning task.", + "bbox": [ + 511, + 799, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25787", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work is partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-II220184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics), National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(No. 2022R1A5A7083908) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) [NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)].", + "bbox": [ + 89, + 90, + 485, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 263, + 187, + 277 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eric Arazo, Diego Ortega, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2020. 5", + "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 6", + "[3] Muhammad Awais, Fengwei Zhou, Hang Xu, Lanqing Hong, Ping Luo, Sung-Ho Bae, and Zhenguo Li. Adversarial robustness for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8568-8577, 2021. 1, 2, 6, 7", + "[4] Francesco Barbato, Marco Toldo, Umberto Michieli, and Pietro Zanuttigh. Latent space regularization for unsupervised domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2835-2845, 2021. 1", + "[5] Peter L Bartlett, Michael I Jordan, and Jon D McAuliffe. Convexity, classification, and risk bounds. Journal of the American Statistical Association, 101(473):138-156, 2006. 2", + "[6] Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. Advances in neural information processing systems, 30, 2017. 4", + "[7] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79: 151-175, 2010. 2", + "[8] Mingcai Chen, Hao Cheng, Yuntao Du, Ming Xu, Wenyu Jiang, and Chongjun Wang. Two wrongs don't make a right: Combating confirmation bias in learning with label noise. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14765-14773, 2023. 5", + "[9] Moustapha Cisse, Piotr Bojanowski, Edouard Grave, Yann Dauphin, and Nicolas Usunier. Parseval networks: Improving robustness to adversarial examples. In International conference on machine learning, pages 854-863. PMLR, 2017. 4", + "[10] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 6", + "[11] Antonio Galbis and Manuel Maestre. Vector analysis versus vector calculus. Springer Science & Business Media, 2012. 4" + ], + "bbox": [ + 93, + 287, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 5", + "[13] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 2, 5", + "[14] Zhiqiang Gao, Shufei Zhang, Kaizhu Huang, Qiufeng Wang, Rui Zhang, and Chaoliang Zhong. Certifying better robust generalization for unsupervised domain adaptation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2399-2410, 2022. 1", + "[15] Chao Gong, Kai Chen, Zhipeng Wei, Jingjing Chen, and Yulong Jiang. Reliable and efficient concept erasure of text-to-image diffusion models. In In European Conference on Computer Vision (ECCV), 2024. 1", + "[16] Fredrik D. Johansson, David Sontag, and Rajesh Ranganath. Support and invertibility in domain-invariant representations. International Conference on Artificial Intelligence and Statistics, 2019. 1", + "[17] Matt Jordan and Alexandros G Dimakis. Exactly computing the local lipschitz constant of relu networks. Advances in Neural Information Processing Systems, 33:7344-7353, 2020. 4", + "[18] Bo Fu Jinguang Jiang, Baixu Chen and Mingsheng Long. Transfer-learning-library. https://github.com/thuml/Transfer-Learning-Library, 2020.5", + "[19] Changhoon Kim, Kyle Min, and Yezhou Yang. R.a.c.e.: Robust adversarial concept erasure for secure text-to-image diffusion model. In In European Conference on Computer Vision (ECCV), 2025. 1", + "[20] Shao-Yuan Lo and Vishal M. Patel. Exploring adversarially robust training for unsupervised domain adaptation. In Asian Conference on Computer Vision (ACCV), 2022. 1, 2, 4, 6, 7", + "[21] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations (ICLR), 2018. 1, 2, 3, 6", + "[22] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. In In International Conference on Learning Representations (ICLR), 2023. 1", + "[23] Afshin Rostamizadeh Mehryar Mohri and Ameet Talwalkar. Foundations of Machine Learning. MIT Press, 2012. 4, 1", + "[24] Farzeen Munir, Shoaib Azam, and Moongu Jeon. Sstn: Self-supervised domain adaptation thermal object detection for autonomous driving, 2021. 1", + "[25] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge, 2017. 6", + "[26] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 6", + "[27] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In In" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25788", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision (ECCV), pages 213-226. Springer, 2010. 6", + "[28] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? In Conference on Neural Information Processing Systems (NeurIPS), 2020. 6", + "[29] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. In In International Conference on Machine Learning (ICML), 2024. 1", + "[30] Zhouxing Shi, Yihan Wang, Huan Zhang, J Zico Kolter, and Cho-Jui Hsieh. Efficiently computing local lipschitz constants of neural networks via bound propagation. Advances in Neural Information Processing Systems, 35:2350-2364, 2022. 4", + "[31] Peeyush Singhal, Rahee Walambe, Sheela Ramanna, and Ketan Kotecha. Domain adaptation: Challenges, methods, datasets, and applications. IEEE Access, 11:6973-7020, 2023. 1", + "[32] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations (ICLR), 2014. 1", + "[33] Michel Talagrand. Upper and lower bounds for stochastic processes. Springer, 2014. 1", + "[34] Marvin Teichmann, Michael Weber, Marius Zoellner, Roberto Cipolla, and Raquel Urtasun. Multinet: Real-time joint semantic reasoning for autonomous driving, 2018. 1", + "[35] Thulasi Tholeti and Sheetal Kalyani. The robust way to stack and bag: the local lipschitz way. arXiv preprint arXiv:2206.00513, 2022. 4", + "[36] Jonathan Uesato, Jean-Baptiste Alayrac, Po-Sen Huang, Robert Stanforth, Alhussein Fawzi, and Pushmeet Kohli. Are labels required for improving adversarial robustness? In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2", + "[37] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018-5027, 2017. 6", + "[38] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International Conference on Learning Representataions(ICLR), 2020. 1", + "[39] Boxi Wu, Jinghui Chen, Deng Cai, Xiaofei He, and Quanquan Gu. Do wider neural networks really help adversarial robustness? Advances in Neural Information Processing Systems, 34:7054-7067, 2021. 4", + "[40] Dongyoon Yang, Insung Kong, and Yongdai Kim. Enhancing adversarial robustness in low-label regime via adaptively weighted regularization and knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4552-4561, 2023. 1, 2", + "[41] Dongyoon Yang, Insung Kong, and Yongdai Kim. Improving adversarial robustness by putting more regularizations on less robust samples. In International Conference on Machine Learning, pages 39331-39348. PMLR, 2023. 1, 2" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Yao-Yuan Yang, Cyrus Rashtchian, Hongyang Zhang, Ruslan Salakhutdinov, and Kamalika Chaudhuri. A closer look at accuracy vs. robustness. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4, 6, 1, 7", + "[43] Bohang Zhang, Du Jiang, Di He, and Liwei Wang. Rethinking lipschitz neural networks and certified robustness: A boolean function perspective. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 4, 6, 1", + "[44] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P Xing, Laurent El Ghaoui, and Michael I Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning (ICML), 2019. 1, 2", + "[45] Yuchen Zhang, Tianle Liu, Mingsheng Long, and Michael I. Jordan. Bridging theory and algorithm for domain adaptation. In International Conference on Machine Learning (ICML), 2019. 2, 3, 4, 5, 6, 1", + "[46] Wanqing Zhu, Jia-Li Yin, Bo-Hao Chen, and Xineng Liu. Srouda: meta self-training for robust unsupervised domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3852-3860, 2023. 1, 2, 3, 4, 6, 7" + ], + "bbox": [ + 514, + 92, + 906, + 387 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25789", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_model.json b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9165475771a7228826f2d50f91f51f33532038f3 --- /dev/null +++ b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_model.json @@ -0,0 +1,2618 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.13, + 0.802, + 0.173 + ], + "angle": 0, + "content": "TAROT: Towards Essentially Domain-Invariant Robustness with Theoretical Justification" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.201, + 0.694, + 0.22 + ], + "angle": 0, + "content": "Dongyoon Yang\\(^{1*†}\\) Jihu Lee\\(^{2*}\\) Yongdai Kim\\(^{2†}\\)" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.22, + 0.65, + 0.237 + ], + "angle": 0, + "content": "\\(^{1}\\)AI Advanced Technology, SK Hynix" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.237, + 0.706, + 0.254 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Statistics, Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.256, + 0.753, + 0.27 + ], + "angle": 0, + "content": "ydy0415@gmail.com, rieky0426@snu.ac.kr, ydkim903@snu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.486, + 0.598 + ], + "angle": 0, + "content": "Robust domain adaptation against adversarial attacks is a critical research area that aims to develop models capable of maintaining consistent performance across diverse and challenging domains. In this paper, we derive a new generalization bound for robust risk on the target domain using a novel divergence measure specifically designed for robust domain adaptation. Building upon this, we propose a new algorithm named TAROT, which is designed to enhance both domain adaptability and robustness. Through extensive experiments, TAROT not only surpasses state-of-the-art methods in accuracy and robustness but also significantly enhances domain generalization and scalability by effectively learning domain-invariant features. In particular, TAROT achieves superior performance on the challenging DomainNet dataset, demonstrating its ability to learn domain-invariant representations that generalize well across different domains, including unseen ones. These results highlight the broader applicability of our approach in real-world domain adaptation scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.608, + 0.222, + 0.624 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.486, + 0.821 + ], + "angle": 0, + "content": "Recent advances in machine learning have highlighted the importance of developing transferable models that can maintain performance across different domains. However, collecting labeled data from all potential domains is too resource-intensive for many practical applications. To address this limitation, Unsupervised Domain Adaptation (UDA) has emerged as a type of transfer learning that leverages labeled source-domain data to improve performance on a target domain without needing labels from that domain. UDA has proven pivotal in diverse machine learning tasks, largely because distributional discrepancies between source and target domains often lead to severe performance degradation [4, 16, 24, 31, 34]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Meanwhile, adversarial robustness has gained significant attention as a crucial element of secure and trustworthy AI. Pioneering works demonstrated the vulnerability of neural" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.909, + 0.452 + ], + "angle": 0, + "content": "networks to small, carefully designed perturbations called adversarial attacks [21, 32] and proposed adversarial training methods [21, 38, 40, 41, 44]. Recently, the adversarial robustness community has expanded its focus to multimodal tasks (e.g., text-to-image, image-to-text), recognizing that large-scale multi-modal models are also susceptible to adversarial manipulations in one or more input modalities [15, 19, 22, 29]. These cross-modal attacks can exploit alignment mechanisms to trick models into misclassification or erroneous content generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.909, + 0.692 + ], + "angle": 0, + "content": "Given these two developments—domain adaptation and adversarial robustness—there is a growing interest in learning transferable models in UDA settings that remain resilient to adversarial perturbations in the target domain [3, 14, 20, 46]. A common strategy in existing robust UDA methods is to rely on pseudo-labels to handle unlabeled target data. However, in this paper, we propose a novel robust margin disparity metric for comparing source and target distributions, enabling us to derive a refined generalization bound on the robust risk in the target domain. From our bound, we derive two key insights: (1) using only pseudo-labels may be insufficient, but combining pseudo-labels with distributional alignment significantly enhances robust UDA performance; and (2) initializing robust UDA algorithms with a robustly pre-trained model (Robust-PT [3]) is notably more effective than starting from a non-robust pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Building upon these theoretical foundations, we propose TAROT (Transfer Adversarially RObust Training), a new algorithm that leverages our theoretical insights to achieve superior performance in robust UDA tasks. Through extensive experiments on various benchmark datasets, we demonstrate that TAROT outperforms existing algorithms. Furthermore, we empirically demonstrate that TAROT exhibits distinct advantages in terms of scalability compared to pseudo-labeling approaches, particularly with significant performance improvements not only on the target domain but also on source and unseen source domains. This work bridges the gap between theoretical understanding and practical implementation in robust transfer learning, offering both fundamental insights and actionable solutions for real-world applications." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.241, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.414, + 0.106 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.108, + 0.482, + 0.136 + ], + "angle": 0, + "content": "- We introduce a novel robust margin disparity measure and derive a refined generalization bound for robust UDA." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.137, + 0.483, + 0.193 + ], + "angle": 0, + "content": "- From the derived generalization bound, we propose TAROT (Transferring Adversarial RObust Training), which integrates pseudo-labeling and distributional alignment to enhance adversarial robustness in UDA." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.194, + 0.483, + 0.267 + ], + "angle": 0, + "content": "- TAROT outperforms existing robust UDA methods across multiple benchmarks, demonstrating strong performance on the target, source, and unseen source domains, indicating that TAROT effectively facilitates domain-invariant robust representations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.267, + 0.483, + 0.31 + ], + "angle": 0, + "content": "- We provide theoretical insights and practical solutions for scalable and adversarially robust domain adaptation in real-world scenarios." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.108, + 0.483, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.327, + 0.229, + 0.343 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.485 + ], + "angle": 0, + "content": "In this section, we outline the core concepts and notations for our analysis, beginning with adversarial training and its risk formulation for robust learning. We then describe the unsupervised domain adaptation (UDA) framework, highlighting distributional notations, the formal UDA objective, and existing theoretical approaches like DANN [13] and MDD [45]. These elements provide the groundwork for our subsequent derivation of a robust UDA upper bound and the proposed learning algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.497, + 0.291, + 0.513 + ], + "angle": 0, + "content": "2.1. Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.484, + 0.628 + ], + "angle": 0, + "content": "Let \\(\\mathcal{X} \\in \\mathbb{R}^d\\) represent the input space, and let \\(\\mathcal{Y} = \\{1, \\ldots, C\\}\\) be the label space, where \\(C\\) represents the number of classes. A classifier can be represented using a score function \\(f: \\mathcal{X} \\to \\mathbb{R}^C\\), where \\(f_c\\) denotes the \\(c\\)-th component of \\(f\\). The classifier is then defined as: \\(h_f(\\pmb{x}) = \\operatorname*{argmax}_{c \\in \\mathcal{Y}} f_c(\\pmb{x})\\). Let \\(\\mathcal{B}_p(\\pmb{x}, \\varepsilon) = \\{\\pmb{x}' \\in \\mathcal{X}: \\| \\pmb{x} - \\pmb{x}' \\|_p \\leq \\varepsilon\\}\\) denote an \\(\\varepsilon\\)-ball centered at \\(\\pmb{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.658 + ], + "angle": 0, + "content": "The population risk considering adversarial robustness is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.673, + 0.484, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\mathrm {r o b}} (f) = \\mathbb {E} _ {(\\mathbf {X}, Y)} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\mathbb {1} \\left\\{Y \\neq h _ {f} \\left(\\mathbf {X} ^ {\\prime}\\right) \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.757 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) denotes the indicator function. The goal of adversarial training is to estimate the optimal \\(f\\), the minimizer of Eq. (1)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.768, + 0.388, + 0.785 + ], + "angle": 0, + "content": "2.2. Unsupervised Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.791, + 0.484, + 0.902 + ], + "angle": 0, + "content": "There are two distributions, the source distribution \\(S\\) and the target distribution \\(\\mathcal{T}\\). The empirical versions of \\(S\\) and \\(\\mathcal{T}\\) are \\(\\widehat{S} = \\{(x_i^s,y_i^s)\\}_{i = 1}^{n_s}\\) and \\(\\widehat{\\mathcal{T}} = \\{(x_j^t,y_j^t)\\}_{j = 1}^{n_t}\\), respectively. In unsupervised domain adaptation, \\(\\widehat{S}\\) and only \\(\\widehat{\\mathcal{T}}_{\\mathbf{X}} := \\{x_j^t\\}_{j = 1}^{n_t}\\) are observed. i.e. the labels \\(y_{j}^{t}\\) can not be observed. Denote the distributions of the input on the source and target domains by \\(S_{\\mathbf{X}}\\) and \\(\\mathcal{T}_{\\mathbf{X}}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "For a given score function \\( f \\in \\mathcal{F} \\), the risk of a data distribution \\( \\mathcal{D} \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.131, + 0.907, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\mathcal {D}} (f) := \\mathbb {E} _ {(\\mathbf {X}, Y) \\sim \\mathcal {D}} \\mathbb {1} \\{Y \\neq h _ {f} (\\mathbf {X}) \\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.159, + 0.907, + 0.19 + ], + "angle": 0, + "content": "The goal of UDA is to minimize \\(\\mathcal{R}_{\\mathcal{T}}(f)\\) using \\(\\widehat{S}\\) and \\(\\widehat{\\mathcal{T}}_{\\mathbf{X}}\\) which can be done by building an upper bound of \\(\\mathcal{R}_{\\mathcal{T}}(h)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.908, + 0.442 + ], + "angle": 0, + "content": "The pivotal learning algorithms for UDA are DANN (Domain-Adversarial Neural Network) [13] and MDD (Margin Disparity Discrepancy) [45]. Ben-David et al. [7] suggested a distance \\(d_{\\mathcal{H}\\Delta \\mathcal{H}}(\\mathcal{S}_{\\mathbf{X}},\\mathcal{T}_{\\mathbf{X}})\\) between \\(\\mathcal{S}_{\\mathbf{X}}\\) and \\(\\mathcal{T}_{\\mathbf{X}}\\) to give an upper bound of the robust risk, and Ganin et al. [13] utilized this upper bound to implement their learning algorithm DANN. DANN uses an adversarial neural network to minimize the distance between the source and target distributions, using gradient reversal layers (GRL). On the other hand, MDD [45] derives an upper bound of \\(\\mathcal{R}_{\\mathcal{T}}(f)\\) as the sum of \\(\\mathcal{R}_S^{(\\rho)}(f)\\), a standard non-robust risk on the source domain, and \\(d_{f,\\mathcal{F}}^{(\\rho)}(\\mathcal{S}_{\\mathbf{X}},\\mathcal{T}_{\\mathbf{X}})\\), a distance between \\(\\mathcal{S}_{\\mathbf{X}}\\) and \\(\\mathcal{T}_{\\mathbf{X}}\\), and develop the MDD algorithm based on this upper bound. The main distinction between the two upper bounds is that MDD utilizes a scoring function for multi-class classification, while DANN only employs a classifier for binary classification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.53 + ], + "angle": 0, + "content": "The objective of our work is to derive an upper bound of the robust risk for the target domain by utilizing the risk in the source domain and the robust divergence between the target domain \\(\\mathcal{T}\\) and the source domain \\(S\\). In turn, inspired by the upper bound, we propose a novel robust UDA algorithm." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.542, + 0.664, + 0.558 + ], + "angle": 0, + "content": "3. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.669 + ], + "angle": 0, + "content": "While a significant body of research on adversarial training has focused on supervised learning and semi-supervised learning tasks [21, 36, 40, 41, 44], there has been relatively less attention given to its application in domain adaptation tasks. Only a few studies have explored adversarial training in the context of domain adaptation [3, 20, 46], highlighting the need for further investigation in this area." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.907, + 0.785 + ], + "angle": 0, + "content": "RFA [3] leverages a pretrained ImageNet model to develop a regularization term. When the trainable model is initialized with a pretrained robust model, the regularization term encourages the outputs of each layer of the trainable model to resemble those of the pretrained model. This facilitates the learning of robust features. This approach can be applied to existing UDA algorithms such as DANN [13] and MDD [45] to make them adversarially robust." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.785, + 0.909, + 0.902 + ], + "angle": 0, + "content": "ARTUDA [20] adopts a self-supervised adversarial training approach, where the predictions of the current model are used as pseudo-labels to generate adversarial examples for the target domain. It also generates adversarial examples for the source domain and applies the distance metrics used in DANN. However, there is no theoretical justification of the complex loss function used in ARTUDA. One key challenge in ARTUDA is that pseudo-labels are generated by the same" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25781" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.165 + ], + "angle": 0, + "content": "model being trained. As a result, due to the inherent trade-off between robustness and generalization, as the model learns more robust features, pseudo-labels are getting more inaccurate. This issue becomes particularly pronounced when the perturbation budget \\((\\varepsilon)\\) is large." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.165, + 0.484, + 0.352 + ], + "angle": 0, + "content": "SRoUDA [46] attempts to resolve this limitation by introducing a separate teacher model to assign pseudo-labels to the target data. By leveraging meta-learning, SRoUDA improves the accuracy of pseudo-labeling by the teacher model and dynamically performs adversarial training [21] on the target domain using these pseudo-labels. Despite the improvement, SRoUDA struggles to effectively capture domain-invariant features, as it does not explicitly consider the deviation of the distributions between the source and target domains. This leads to a decline in performance in the source as well as in the unseen domains, indicating a failure to capture what we call essentially domain-invariant robust features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.367, + 0.265, + 0.385 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.437 + ], + "angle": 0, + "content": "In this section, we present the theoretical results for adversarially robust UDA. This is an extension of MDD [45] to its robust version, and the proofs are deferred to Sec. 8." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.446, + 0.419, + 0.463 + ], + "angle": 0, + "content": "4.1. Robust Margin Disparity Discrepancy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.497 + ], + "angle": 0, + "content": "In this subsection, we introduce a novel disparity measure and key propositions to establish the generalization bound." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.497, + 0.474, + 0.513 + ], + "angle": 0, + "content": "Denote the classifier \\( h_f \\) for a given score function \\( f \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.524, + 0.372, + 0.55 + ], + "angle": 0, + "content": "\\[\nh_{f}:\\boldsymbol {x}\\mapsto \\operatorname *{argmax}_{y\\in \\mathcal{Y}}f_{y}(\\boldsymbol {x}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.562, + 0.465, + 0.579 + ], + "angle": 0, + "content": "and denote the margin of \\( f \\) at a labeled sample \\( (x, y) \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.589, + 0.402, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {f} (\\boldsymbol {x}, y) = f _ {y} (\\boldsymbol {x}) - \\max _ {y ^ {\\prime} \\neq y} f _ {y ^ {\\prime}} (\\boldsymbol {x}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Definition 1. (Robust Disparity). For given \\( f, f' \\in \\mathcal{F} \\), the 0-1 robust disparity between \\( f \\) and \\( f' \\) on a distribution \\( \\mathcal{D}_{\\mathbf{X}} \\) on \\( \\mathcal{X} \\) is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.698, + 0.485, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i s p} _ {\\mathcal {D} _ {\\mathbf {X}}} ^ {r o b} \\left(f ^ {\\prime}, f\\right) := \\mathbb {E} _ {\\mathbf {X} \\sim \\mathcal {D} _ {\\mathbf {X}}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\mathbb {1} \\left\\{h _ {f ^ {\\prime}} (\\mathbf {X}) \\neq h _ {f} \\left(\\mathbf {X} ^ {\\prime}\\right) \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.82 + ], + "angle": 0, + "content": "The robust disparity is measured based on the adversarial robustness of the second model \\( f \\), while the first model \\( f' \\) serves as a pseudo-label generator. Note that if the prediction of \\( f' \\) is consistent to true label, it is the robust risk as defined in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.484, + 0.865 + ], + "angle": 0, + "content": "As an extension of the margin loss in the non-robust setting [45] to its robust counterpart, we define the robust margin loss as" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.876, + 0.484, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\mathcal {D}} ^ {\\mathrm {r o b}, (\\rho)} (f) := \\mathbb {E} _ {(\\mathbf {X}, Y) \\sim \\mathcal {D}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\Phi_ {\\rho} \\circ \\mathcal {M} _ {f} \\left(\\mathbf {X} ^ {\\prime}, Y\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.56, + 0.105 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.103, + 0.907, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {\\rho} (x) := \\left\\{ \\begin{array}{l l} 0 & \\rho \\leq x \\\\ 1 - x / \\rho & 0 \\leq x \\leq \\rho \\\\ 1 & x \\leq 0 \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.164, + 0.785, + 0.178 + ], + "angle": 0, + "content": "and \\( \\circ \\) is composite operator of functions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.187, + 0.907, + 0.218 + ], + "angle": 0, + "content": "Definition 2. (Robust Margin Disparity). For given \\( f, f' \\in \\mathcal{F} \\), the Robust Margin Disparity is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.228, + 0.907, + 0.275 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {d i s p} _ {\\mathcal {D} _ {\\mathbf {X}}} ^ {r o b, (\\rho)} \\left(f ^ {\\prime}, f\\right) \\\\ := \\mathbb {E} _ {\\mathbf {X} \\sim \\mathcal {D} _ {\\mathbf {X}}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {\\rho} (\\mathbf {X}, \\varepsilon)} \\Phi_ {\\rho} \\circ \\mathcal {M} _ {f} \\left(\\mathbf {X} ^ {\\prime}, h _ {f ^ {\\prime}} (\\mathbf {X})\\right) \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.287, + 0.908, + 0.335 + ], + "angle": 0, + "content": "An important property of the robust margin disparity is that it is greater than the robust disparity, i.e. \\(\\mathrm{disp}_{\\mathcal{D}_{\\mathbf{X}}}^{\\mathrm{rob},(\\rho)}(f', f) \\geq \\mathrm{disp}_{\\mathcal{D}_{\\mathbf{X}}}^{\\mathrm{rob}}(f', f)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.905, + 0.362 + ], + "angle": 0, + "content": "We now use this disparity term to bound the robust risk on the target distribution with additional feasible terms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.908, + 0.446 + ], + "angle": 0, + "content": "Proposition 1. Let \\( S \\) and \\( \\mathcal{T} \\) represent the distributions of the source and target domains, respectively. Similarly, let \\( S_{\\mathbf{X}} \\) and \\( \\mathcal{T}_{\\mathbf{X}} \\) denote the marginal distributions of the source and target domains over \\( \\mathbf{X} \\), respectively. For every score function \\( f \\in \\mathcal{F} \\), the following inequality holds:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.455, + 0.906, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {R} _ {T} ^ {r o b} (f) \\leq \\\\ \\mathcal {R} _ {\\mathcal {S}} ^ {(\\rho)} (f) + \\left\\{\\operatorname {d i s p} _ {\\mathcal {T} _ {\\mathbf {X}}} ^ {r o b, (\\rho)} \\left(f ^ {*}, f\\right) - \\operatorname {d i s p} _ {\\mathcal {S} _ {\\mathbf {X}}} ^ {(\\rho)} \\left(f ^ {*}, f\\right) \\right\\} + \\lambda , \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.514, + 0.908, + 0.574 + ], + "angle": 0, + "content": "where \\(f^{*} = \\operatorname *{argmin}_{f\\in \\mathcal{F}}\\{\\mathcal{R}_{\\mathcal{T}}^{(\\rho)}(f) + \\mathcal{R}_{\\mathcal{S}}^{(\\rho)}(f)\\}\\) is ideal hypothesis and \\(\\lambda = \\lambda (\\mathcal{F},\\mathcal{S},\\mathcal{T},\\varepsilon ,\\rho) = \\mathcal{R}_{\\mathcal{T}}^{(\\rho)}(f^{*}) + \\mathcal{R}_{\\mathcal{S}}^{(\\rho)}(f^{*})\\) is constant of \\(f\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.669 + ], + "angle": 0, + "content": "Note that the upper bound consists of standard margin risk of the source domain, not the robust margin risk. This enables the corresponding algorithm introduced in Sec. 4.4 computationally simpler because generation of adversarial samples is not required to calculate the standard margin risk on the source domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.908, + 0.729 + ], + "angle": 0, + "content": "From the above proposition, we can naturally induce the following discrepancy between the distributions of two domains, which we call as a robust margin disparity discrepancy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.738, + 0.895, + 0.752 + ], + "angle": 0, + "content": "Definition 3. (Robust Margin Disparity Discrepancy)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.752, + 0.907, + 0.797 + ], + "angle": 0, + "content": "Given a score function \\( f \\in \\mathcal{F} \\), the Robust Margin Disparity Discrepancy induced by \\( f' \\in \\mathcal{F} \\) of two distributions \\( S_{\\mathbf{X}} \\) and \\( T_{\\mathbf{X}} \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.807, + 0.907, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d _ {f, \\mathcal {F}} ^ {r o b, (\\rho)} (S _ {\\mathbf {X}}, \\mathcal {T} _ {\\mathbf {X}}) \\\\ := \\sup _ {f ^ {\\prime} \\in \\mathcal {F}} \\left\\{\\operatorname {d i s p} _ {\\mathcal {T} _ {\\mathbf {x}}} ^ {r o b, (\\rho)} \\left(f ^ {\\prime}, f\\right) - \\operatorname {d i s p} _ {\\mathcal {S} _ {\\mathbf {x}}} ^ {(\\rho)} \\left(f ^ {\\prime}, f\\right) \\right\\} \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "This discrepancy measures the difference of the robust margin disparity on target domain and the standard margin" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25782" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.164 + ], + "angle": 0, + "content": "disparity on the source domain, where the objective \\( f \\) is fixed. If there is significant gap between the adversarial target domain and standard source domain, the discrepancy becomes large. Now we introduce a non-asymptotic upper bound for the target robust risk:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.172, + 0.484, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\mathcal {T}} ^ {\\mathrm {r o b}, (\\rho)} (f) \\leq \\mathcal {R} _ {\\mathcal {S}} ^ {(\\rho)} (f) + d _ {f, \\mathcal {F}} ^ {\\mathrm {r o b}, (\\rho)} \\left(\\mathcal {S} _ {\\mathbf {X}}, \\mathcal {T} _ {\\mathbf {X}}\\right) + \\lambda \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.199, + 0.313, + 0.213 + ], + "angle": 0, + "content": "with the same \\(\\lambda\\) in Proposition 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.221, + 0.295, + 0.236 + ], + "angle": 0, + "content": "4.2. Generalization bound" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.485, + 0.287 + ], + "angle": 0, + "content": "In this subsection, we provide a generalization bound based on Eq. (9). We use the notion of local Lipschitz constant of a function, following Yang et al. [42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.485, + 0.338 + ], + "angle": 0, + "content": "Definition 4 (Local Lipschitz Constant). Consider an arbitrary real function \\( f: \\mathbb{R}^d \\to \\mathbb{R}^C \\), and a \\( L_p \\)-norm \\( \\| \\cdot \\|_p \\). Define a point-wise local Lipschitz constant as" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.343, + 0.484, + 0.379 + ], + "angle": 0, + "content": "\\[\nL _ {f} (\\boldsymbol {x}, \\varepsilon) := \\sup _ {\\boldsymbol {x} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\boldsymbol {x}, \\varepsilon)} \\frac {\\| f \\left(\\boldsymbol {x} ^ {\\prime}\\right) - f (\\boldsymbol {x}) \\| _ {1}}{\\| \\boldsymbol {x} ^ {\\prime} - \\boldsymbol {x} \\| _ {p}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.483, + 0.414 + ], + "angle": 0, + "content": "for a point \\( \\pmb{x} \\). Then, we define a local Lipschitz constant on \\( \\mathcal{D}_{\\mathbf{X}} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.415, + 0.483, + 0.439 + ], + "angle": 0, + "content": "\\[\nL _ {f} \\left(\\mathcal {D} _ {\\mathbf {X}}, \\varepsilon\\right) := \\sup _ {x \\in \\mathcal {D} _ {\\mathbf {X}}} L _ {f} (x, \\varepsilon) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.474 + ], + "angle": 0, + "content": "We call that \\( f \\) is locally Lipschitz, if \\( L_{f}(\\mathcal{D}_{\\mathbf{X}},\\varepsilon) \\) is bounded (i.e. \\( L_{f}(\\mathcal{D}_{\\mathbf{X}},\\varepsilon) < \\infty \\))." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.48, + 0.449, + 0.495 + ], + "angle": 0, + "content": "We also define following auxiliary function spaces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.485, + 0.545 + ], + "angle": 0, + "content": "Definition 5. Consider a set of score functions \\(\\mathcal{F}\\) and let \\(\\mathcal{H}\\) be the class of the induced classifiers \\(h_f\\) for each \\(f\\in \\mathcal{F}\\). We define the following two sets." + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.553, + 0.483, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\Pi_ {\\mathcal {H}} \\mathcal {F} := \\{x \\mapsto f (x, h (x)) | h \\in \\mathcal {H}, f \\in \\mathcal {F} \\} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.571, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\Pi_ {1} \\mathcal {F} := \\{x \\mapsto f (x, y) | y \\in \\mathcal {Y}, f \\in \\mathcal {F} \\} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.638 + ], + "angle": 0, + "content": "It is known that \\(\\Pi_{\\mathcal{H}}\\mathcal{F}\\) takes a form of the space of inner products of vector fields from \\(\\mathcal{H}\\) to \\(\\mathcal{F}\\) [11, 45]. The notion \\(\\Pi_1\\mathcal{F}\\) follows from Mehryar Mohri and Talwalkar [23]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.696 + ], + "angle": 0, + "content": "We now derive a generalization bound for the robust risk on the target domain using Rademacher complexity, which quantifies the expressiveness of a given hypothesis space [23]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.78 + ], + "angle": 0, + "content": "Definition 6. (Rademacher Complexity). Let \\(\\mathcal{F}\\) be a family of functions mapping from \\(\\mathcal{Z} = \\mathcal{X} \\times \\mathcal{Y}\\) to \\([a, b]\\) and \\(\\widehat{\\mathcal{D}} = \\{z_1, \\dots, z_n\\}\\) are finite samples independently drawn from a distribution \\(\\mathcal{D}\\) on \\(\\mathcal{Z}\\). Then, the empirical Rademacher complexity of \\(\\mathcal{F}\\) with \\(\\widehat{\\mathcal{D}}\\) is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.787, + 0.484, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathfrak {R}} _ {\\widehat {\\mathcal {D}}} (\\mathcal {F}) \\triangleq \\mathbb {E} _ {\\sigma} \\sup _ {f \\in \\mathcal {F}} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sigma_ {i} f (z _ {i}), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.832, + 0.485, + 0.876 + ], + "angle": 0, + "content": "where \\(\\sigma_{i}\\)s are independent random variables following the Rademacher distribution. In turn, the Rademacher complexity of \\(\\mathcal{F}\\) is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.884, + 0.484, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathfrak {R} _ {n, \\mathcal {D}} (\\mathcal {F}) \\triangleq \\mathbb {E} _ {\\widehat {\\mathcal {D}} \\sim \\mathcal {D} ^ {n}} \\widehat {\\mathfrak {R}} _ {\\widehat {\\mathcal {D}}} (\\mathcal {F}) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.15 + ], + "angle": 0, + "content": "Now we provide our generalization bound, which replaces the terms in the upper bound (Eq. (9)) that depend on the population distribution by the terms depending on the empirical distributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.16, + 0.908, + 0.22 + ], + "angle": 0, + "content": "Theorem 1. (Generalization Bound on the Robust Risk of Target Distribution). For any \\(\\delta > 0\\), with probability \\(1 - 3\\delta\\), we have the following uniform generalization bound for any score function \\(f\\) in \\(\\mathcal{F}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.23, + 0.906, + 0.359 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {R} _ {\\mathcal {T}} ^ {r o b} (f) \\\\ \\leq \\mathcal {R} _ {\\widehat {\\mathcal {S}}} ^ {(\\rho)} (f) + d _ {f, \\mathcal {F}} ^ {r o b, (\\rho)} \\left(\\widehat {\\mathcal {S}} _ {\\mathbf {X}}, \\widehat {\\mathcal {T}} _ {\\mathbf {X}}\\right) + \\lambda \\\\ + \\frac {2 C ^ {2}}{\\rho} \\mathfrak {R} _ {n, S} (\\Pi_ {1} \\mathcal {F}) + \\frac {2 C}{\\rho} \\mathfrak {R} _ {n, S} (\\Pi_ {\\mathcal {H}} \\mathcal {F}) + 2 \\sqrt {\\frac {\\log 2 / \\delta}{2 n}} \\\\ + \\frac {2 C}{\\rho} \\Re_ {m, \\mathcal {T}} \\left(\\Pi_ {\\mathcal {H}} \\mathcal {F}\\right) + \\sqrt {\\frac {\\log 2 / \\delta}{2 m}} + \\frac {2 \\varepsilon L _ {f} \\left(\\mathcal {T} _ {\\mathbf {X}}, \\varepsilon\\right)}{\\rho}, \\tag {16} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.359, + 0.768, + 0.384 + ], + "angle": 0, + "content": "where \\(\\lambda = \\min_{f\\in \\mathcal{F}}\\{\\mathcal{R}_{\\mathcal{T}}^{(\\rho)}(f) + \\mathcal{R}_{\\mathcal{S}}^{(\\rho)}(f)\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.393, + 0.649, + 0.409 + ], + "angle": 0, + "content": "4.3. Two insights" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.908, + 0.545 + ], + "angle": 0, + "content": "In this subsection, we present two key insights for robust UDA. First, we analyze the impact of the local Lipschitz constant on generalization performance, showing why initializing robust UDA algorithms with a robustly pre-trained (Robust-PT) model can substantially improve outcomes. Second, we highlight why pseudo labeling alone may be insufficient, especially in large or complex target domains where confirmation bias and the lack of explicit domain alignment can lead to persistent misclassifications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.693 + ], + "angle": 0, + "content": "About Local Lipschitz constant In Theorem 1, there is the local Lipschitz constant \\(L_{f}(\\mathcal{T}_{\\mathbf{X}},\\varepsilon)\\). As discovered in previous literature, the local Lipschitz constant is known to be deeply involved with a generalization performance of a model [6, 17, 30]. This fact leads to various researches that investigate a method to insure provable robustness of a model [9, 35, 39]. It is also known that adversarial training empirically reduces the local Lipschitz constant of a neural network [42, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.909, + 0.839 + ], + "angle": 0, + "content": "In the numerical studies, we have found that robust UDA approaches with a Robust-PT as an initial improve the prediction performance much compared to those without a Robust-PT as an initial. This phenomenon can be partly explained by the local Lipschitz term. That is, model which is not initialized with Robust-PT could have a large value of the local Lipschitz constant and thus could have a too large robust risk value on the target domain. Hence, using an initial solution with a small local Lipschitz constant is promising and a Robust-PT would be such one." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Pseudo labeling only may not be sufficient Pseudo labeling is a widely used technique in domain adaptation [20, 46], where the objective is to improve the performance of the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.165 + ], + "angle": 0, + "content": "model on the target domain by iteratively using the predictive labels of the currently trained model (or its variation) as labels for the target domain data when generating adversarial examples for the target data. There are at least two problems in those pseudo labeling approaches." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.165, + 0.483, + 0.324 + ], + "angle": 0, + "content": "The first problem is a risk of confirmation bias. Since the model generates pseudo labels based on its own predictions, early-stage errors in the model could become self-reinforcing. When incorrect pseudo labels are assigned with high confidence, these mistakes persist through training cycles, amplifying errors and leading to poor generalization on the target domain. This phenomenon is widely studied in semi-supervised settings [1, 8]. Similarly in the UDA setting, as the training phase progresses, the model may increasingly rely on those incorrect predictions, making it difficult to recover especially in complex or noisy target domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.382 + ], + "angle": 0, + "content": "The risk of confirmation bias can be reduced by generating pseudo labels from a pretrained non-robust UDA model called a (fixed) teacher model. In this case, we can write the objective function as" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.392, + 0.484, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathbf {X} \\sim \\mathcal {T} _ {\\mathbf {X}}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\mathbb {1} \\left\\{h _ {f} \\left(\\mathbf {X} ^ {\\prime}\\right) \\neq h _ {f *} (\\mathbf {X}) \\right\\} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.483 + ], + "angle": 0, + "content": "where \\( f^{*} = \\operatorname*{argmin}_{f\\in \\mathcal{F}}\\mathcal{R}_{\\widehat{\\mathcal{T}}}(\\boldsymbol {f}) \\). In our notation, this term exactly coincides with \\( \\mathrm{disp}_{T_{\\mathbf{x}}}^{\\mathrm{rob}}(f^{*},f) \\). The following proposition justifies this approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Proposition 2. For every score function \\( f \\in \\mathcal{F} \\), the following inequality holds:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.53, + 0.483, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\mathcal {T}} ^ {r o b} (f) \\leq \\operatorname {d i s p} _ {\\mathcal {T} _ {\\mathbf {X}}} ^ {r o b} \\left(f ^ {*}, f\\right) + \\mathcal {R} _ {\\mathcal {T}} \\left(f ^ {*}\\right) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.558, + 0.236, + 0.573 + ], + "angle": 0, + "content": "for arbitrary \\(f^{*}\\in \\mathcal{F}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.484, + 0.655 + ], + "angle": 0, + "content": "It can be easily seen through the intermediate steps of Proposition 1. By letting \\( f^{*} \\) as a non-robust UDA model, the pseudo labeling approach with the teacher model \\( f^{*} \\) can be understood to minimize the upper bound in Proposition 2. Our empirical studies amply confirm this insight." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.655, + 0.483, + 0.828 + ], + "angle": 0, + "content": "Pseudo labeling approaches with a fixed teacher model, however, still has limitations. In particular, the scalability of pseudo labeling across diverse or large datasets can be challenging due to the absence of explicit domain alignment mechanisms. As the size or complexity of the target domain grows, the chance of incorrect pseudo labels increases, which could overwhelm the overall learning process. This may happen especially in scenarios where the source and target domains differ significantly. This makes it harder for the model to generalize to more extensive or varied datasets, particularly when the domain gap becomes larger due to increasing size or diversity of data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.828, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Note that our generalization bound Theorem 1 uses the robust margin disparity discrepancy between the source and target domains instead of the robust disparity only on the target domain in Proposition 2. That is, our generalization bound integrates pseudo labeling and domain alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.165 + ], + "angle": 0, + "content": "together. In the following subsection, inspired by these insights, we introduce an algorithm designed to overcome the drawbacks of pseudo labeling approaches, tackling both confirmation bias and scalability while enhancing generalization abilities across diverse and complex target domains." + }, + { + "type": "code_caption", + "bbox": [ + 0.522, + 0.183, + 0.862, + 0.212 + ], + "angle": 0, + "content": "Algorithm 1: Transferring Adversarily Robust Training (TAROT)" + }, + { + "type": "algorithm", + "bbox": [ + 0.52, + 0.214, + 0.925, + 0.512 + ], + "angle": 0, + "content": "Input : \\(\\psi\\) : feature extractor, \\(\\pi\\) : head, \\(\\pi'\\) : auxiliary head, \\(s\\) : standard trained model, \\(\\mathcal{D}_{\\mathcal{S}} = \\{(x_{i}^{s},y_{i}^{s})\\}_{i = 1}^{m}\\) : source domain dataset, \\(\\mathcal{D}_{\\mathcal{S}} = \\{(x_{i}^{s},y_{i}^{s})\\}_{i = 1}^{m}:\\mathcal{D}_{\\mathcal{T}} = \\{x_{j}^{t}\\}_{j = 1}^{m^{\\prime}}\\) : target domain dataset, \\(\\eta_1,\\eta_2,\\lambda ,\\rho\\) : learning rates in Eq. (20), \\(T\\) : the number of epochs, \\(B\\) : the number of batches, \\(K\\) : batch size Output: adversarially robust network \\(\\pi \\circ \\psi\\) \n1 \\(\\theta_{\\psi}\\gets \\theta_{\\mathrm{R}}\\) (initialized by a robustly trained feature extractor) \n2 for \\(b = 1,\\dots ,B\\) do \n3 for \\(k = 1,\\dots ,K\\) do \n4 Generate \\(\\widehat{\\pmb{x}}_{b,k}^{\\mathrm{t,adv}}\\) by PGD \\((\\pi \\circ \\psi (\\pmb{x}_{b,k}^{t}),\\tilde{y}_{b,k}^{t})\\) \\(\\tilde{y}_{b,k}^{t} = h_{s}(\\pmb{x}_{b,k}^{t})\\) \n5 Update \\((\\theta_{\\psi},\\theta_{\\pi})\\gets (\\theta_{\\psi},\\theta_{\\pi})-\\) \\(\\eta_1\\frac{1}{K}\\nabla_{(\\pmb {\\theta}_\\psi ,\\pmb {\\theta}_\\pi)}\\ell_{\\mathrm{TAROT}}(\\pmb {\\theta};\\{(x_{b,k}^{\\mathrm{s}},y_{b,k}^{\\mathrm{s}})\\}_{k = 1}^{K},\\{(x_{b,k}^{\\mathrm{t}})\\}_{k = 1}^{K})\\) \n6 Update \\(\\theta_{\\pi '}\\gets \\theta_{\\pi ' + }\\) \\(\\eta_2\\frac{1}{K}\\nabla_{\\pmb{\\theta}_{\\pi '}}\\ell_{\\mathrm{TAROT}}(\\pmb {\\theta};\\{(x_{b,k}^{\\mathrm{s}},y_{b,k}^{\\mathrm{s}})\\}_{k = 1}^{K},\\{(x_{b,k}^{\\mathrm{t}})\\}_{k = 1}^{K},\\lambda)\\) \n7 end \n8 end \n9 Return \\(\\pi \\circ \\psi\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.54, + 0.704, + 0.556 + ], + "angle": 0, + "content": "4.4. Learning Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.908, + 0.634 + ], + "angle": 0, + "content": "Inspired by Theorem 1, we propose the robust domain adaptation algorithm, which we call Transferring Adversarially RObust Training (TAROT). Our basic idea is to find a optimal score function \\( f \\in \\mathcal{F} \\) by solving a modified version of the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.643, + 0.907, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\min _ {f \\in \\mathcal {F}} \\mathcal {R} _ {\\widehat {\\mathcal {S}}} ^ {(\\rho)} (f) + d _ {f, \\mathcal {F}} ^ {\\operatorname {r o b}, (\\rho)} \\left(\\widehat {\\mathcal {S}} _ {\\mathbf {X}}, \\widehat {\\mathcal {T}} _ {\\mathbf {X}}\\right) + L _ {f} \\left(\\widehat {\\mathcal {T}} _ {\\mathbf {X}}, \\varepsilon\\right) \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.678, + 0.907, + 0.796 + ], + "angle": 0, + "content": "- As is done in DANN (Domain-Adversarial Neural Network) [13] and MDD (Margin Disparity Discrepancy) [45], we consider a neural network structure that is composed of a feature extractor \\(\\psi\\) and a head \\(\\pi\\), i.e. a score function \\(f \\in \\mathcal{F}\\) can be represented as \\(\\pi \\circ \\psi\\). We only consider \\(\\pi\\) with \\(\\psi\\) being fixed, when calculating \\(d_{f,\\mathcal{F}}^{\\mathrm{rob},(\\rho)}\\). That is, we calculate the robust margin disparity discrepancy on the feature space of \\(\\psi\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.797, + 0.909, + 0.904 + ], + "angle": 0, + "content": "- To reduce the first and second term, we follow the formulation of the MDD algorithm. Likewise, we utilize an auxiliary head \\(\\pi'\\) and \\(\\gamma \\coloneqq \\exp \\rho\\) designed to attain the margin. \\(\\pi'\\) shares the same hypothesis space with \\(\\pi\\), and works as an adversarial network. The discrepancy term, which is non-differentiable, can be optimized through GRL [12]. The only difference is that, when to calculate \\(d_{f,\\mathcal{F}}^{\\mathrm{rob},(\\rho)}\\)," + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.678, + 0.909, + 0.904 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25784" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.488, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.27, + 0.411, + 0.284 + ], + "angle": 0, + "content": "Figure 1. Overview of TAROT algorithm." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.318, + 0.483, + 0.348 + ], + "angle": 0, + "content": "it requires calculating a robust disparity and a standard disparity rather than two standard disparities." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.348, + 0.483, + 0.435 + ], + "angle": 0, + "content": "- To reduce the local Lipschitz constant, we minimize the cross entropy loss with adversarial examples for given pseudo labels attached by a teacher model learned by a nonrobust UDA approach. This is based on the previously mentioned phenomenon that adversarial training reduces the local Lipschitz constant [42, 43]." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.435, + 0.484, + 0.492 + ], + "angle": 0, + "content": "- When calculating the robust margin disparity discrepancy, instead of generating adversarial examples, we reuse those generated for the local Lipsshitz constant, making the algorithm computationally effective." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.492, + 0.483, + 0.523 + ], + "angle": 0, + "content": "- As an initial model for \\(\\psi\\), we used a robustly pretrained models with from the repository provided in [28]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.348, + 0.484, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.531, + 0.412, + 0.546 + ], + "angle": 0, + "content": "The objective function of TAROT is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.565, + 0.483, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\ell_ {\\text {T A R O T}} (\\boldsymbol {\\theta}; \\left\\{\\left(\\boldsymbol {x} _ {i} ^ {\\mathrm {s}}, y _ {i} ^ {\\mathrm {s}}\\right) \\right\\} _ {i = 1} ^ {m}, \\left\\{\\left(\\boldsymbol {x} _ {j} ^ {\\mathrm {t}}\\right) \\right\\} _ {j = m + 1} ^ {m + m ^ {\\prime}}, \\alpha , \\gamma) \\\\ = \\alpha \\left\\{ \\right.\\frac {1}{m} \\sum_ {i = 1} ^ {m} \\ell_ {\\mathrm {c e}} \\left(\\left(\\pi \\circ \\psi\\right)\\left(\\boldsymbol {x} _ {i}\\right), y _ {i}\\right)\\left. \\right)\\left. \\right. \\\\ + \\left(\\frac {1}{m ^ {\\prime}} \\sum_ {j = m + 1} ^ {m + m ^ {\\prime}} \\ell_ {\\mathrm {m o d - c e}} ^ {\\mathrm {r o b}} \\left(\\left(\\pi^ {\\prime} \\circ \\psi\\right) \\left(\\boldsymbol {x} _ {j}\\right), h _ {\\pi \\circ \\psi} \\left(\\boldsymbol {x} _ {j}\\right)\\right) \\right. \\tag {20} \\\\ \\left. \\left. - \\frac {\\gamma}{m} \\sum_ {j = 1} ^ {m} \\ell_ {\\mathrm {c e}} \\left(\\left(\\pi^ {\\prime} \\circ \\psi\\right) \\left(\\boldsymbol {x} _ {j}\\right), h _ {\\pi \\circ \\psi} \\left(\\boldsymbol {x} _ {j}\\right)\\right)\\right) \\right\\} \\\\ + \\frac {1}{m ^ {\\prime}} \\sum_ {j = m + 1} ^ {m + m ^ {\\prime}} \\ell_ {\\mathrm {c e}} ^ {\\mathrm {r o b}} ((\\pi \\circ \\psi) (\\boldsymbol {x} _ {j}), \\tilde {y} _ {j}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.486, + 0.902 + ], + "angle": 0, + "content": "where \\(\\ell_{\\mathrm{ce}}\\), \\(\\ell_{\\mathrm{mod - ce}}^{\\mathrm{rob}}\\) [45] and \\(\\ell_{\\mathrm{ce}}^{\\mathrm{rob}}\\) [21] as cross-entropy, modified cross-entropy with adversarial examples and cross-entropy with adversarial examples respectively. Also, \\(\\pmb{\\theta} = (\\pmb{\\theta}_{\\psi},\\pmb{\\theta}_{\\pi},\\pmb{\\theta}_{\\pi'})\\) to represent the parameters corresponding to \\(\\psi, \\pi, \\pi'\\). \\(\\tilde{y}_j = h_s(\\pmb{x}_j)\\) is a pseudo label generated by a standard trained model \\(s\\) and \\(\\alpha\\) is a trade-off parameter that balances transferability and target's robustness. TAROT algorithm is summarized in Algorithm 1 and Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.646, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.908, + 0.248 + ], + "angle": 0, + "content": "In this section, we compare TAROT with other competitors for adversarially robust UDA. We use several benchmark datasets: Office-31 [27], Office-Home [37], VisDA2017 [25], and DomainNet [26]. Throughout these datasets, effectiveness of TAROT is evaluated across a wide range of domain shifts and robustness scenarios. Further experimental details including explanation for datasets and experimental setups are presented in Sec. 9. The code is available in the repository\\(\\ddagger\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.257, + 0.738, + 0.272 + ], + "angle": 0, + "content": "5.1. Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.907, + 0.452 + ], + "angle": 0, + "content": "Training Setup All of the results are averaged over three runs, with different teacher model initializations. Different perturbation budgets \\((\\varepsilon)\\) are chosen based on the difficulty of each benchmark dataset. For more challenging datasets, a smaller \\(\\varepsilon\\) is used, and vice versa. Specifically, we use \\(\\varepsilon = 16 / 255\\) for the Office-31 and Office-Home datasets, while \\(\\varepsilon = 8 / 255\\) and \\(\\varepsilon = 4 / 255\\) are used for VisDA2017 and DomainNet, respectively. The values of \\(\\alpha\\) in Eq. (20) are set to 0.05, 0.1, 0.5, and 1.0 for Office31, OfficeHome, VisDA2017, and DomainNet, respectively. We empirically observed that the optimal \\(\\alpha\\) values for each task vary much. The sensitivity analysis of \\(\\alpha\\) is provided in Sec. 5.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.909, + 0.584 + ], + "angle": 0, + "content": "As baselines, along with ARTUDA [20], RFA [3], and SRoUDA [46], we consider the pseudo labeling (PL) algorithm which is implemented as a simple PGD-AT algorithm where target pseudo labels are obtained from a fixed non-robustly trained teachers learned by MDD [45]. All algorithms are initialized with Robust-PT as is done for TAROT. Without Robust-PT, training with larger perturbations such as \\(\\varepsilon = 16 / 255\\) can lead to non-decreasing training loss in certain tasks, as discussed in Sec. 10.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Evaluation Setup We report standard accuracies on clean samples and robust accuracies against AutoAttack [10]. AutoAttack, which includes both white-box and black-box attacks, is a powerful evaluation tool that mitigates errors in robustness measurement caused by gradient obfuscation—a phenomenon where model gradients are distorted or masked, leading to overestimated robustness [2]. AutoAttack is widely recognized for accurately assessing adversarial robustness by bypassing these gradient masking issues." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.787, + 0.741 + ], + "angle": 0, + "content": "5.1.1. Performance on Target Domains" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.909, + 0.877 + ], + "angle": 0, + "content": "Tables 1, 2, 4 and 5 demonstrate that PL outperforms other algorithms, with the exception of TAROT. As seen in Figure 3, simply combining Robust-PT as an initial solution with pseudo labeling can yield significant performance gains, which confirms our first theoretical insight. (See Sec. 5.2.) This combination serves as a strong baseline in adversarially robust domain adaptation. It's noteworthy that ARTUDA [20], RFA [3], and SRoUDA [46] algorithms show remarkably poor performance on several tasks. Note that the robust" + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.887, + 0.785, + 0.901 + ], + "angle": 0, + "content": "https://github.com/dyoony/TAROT" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25785" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Performances of ARTUDA, RFA, SROUDA, PL and TAROT on Office31 (\\( \\varepsilon = 16/255 \\)). In each cell, the first number is the standard accuracy (\\%), while the second number is the robust accuracy (\\%) for AA. Bold numbers indicate the best performance." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.122, + 0.85, + 0.198 + ], + "angle": 0, + "content": "
MethodA → DA → WD → AD → WW → AW → DAvg.
ARTUDA23.49 / 11.6522.39 / 9.5635.00 / 16.5168.55 / 35.8536.81 / 19.7474.10 / 51.6143.39 / 24.15
RFA78.72 / 1.2171.20 / 1.2655.84 / 18.0396.73 / 6.4260.10 / 23.0099.40 / 6.0277.00 / 9.32
SRoUDA93.37 / 42.3778.87 / 65.1617.93 / 3.4496.00 / 44.2866.53 / 47.21100.00 / 62.4574.62 / 44.15
PL91.77 / 84.1491.82 / 91.7074.19 / 67.8797.99 / 95.9872.42 / 65.46100.00 / 92.5788.03 / 82.95
TAROT94.18 / 84.1495.47 / 93.4674.51 / 67.1397.99 / 96.2373.73 / 66.88100.00 / 91.7789.31 / 83.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.206, + 0.907, + 0.234 + ], + "angle": 0, + "content": "Table 2. Performances of ARTUDA, RFA, SRoUDA, PL and TAROT on OfficeHome (ε = 16/255). In each cell, the first number is the standard accuracy (%), while the second number is the robust accuracy (% for AA. Bold numbers indicate the best performance." + }, + { + "type": "table", + "bbox": [ + 0.149, + 0.238, + 0.847, + 0.386 + ], + "angle": 0, + "content": "
MethodAr → ClAr → PrAr → RwCl → ArCl → PrCl → Rw
ARTUDA30.29 / 13.5626.18 / 4.755.88 / 0.003.42 / 0.0025.34 / 10.4820.98 / 5.53
RFA48.39 / 17.8954.67 / 8.7663.74 / 6.3443.22 / 3.1758.03 / 9.1557.24 / 5.69
SRoUDA51.84 / 37.8975.60 / 45.4674.25 / 40.2857.97 / 33.4267.76 / 54.7767.32 / 35.83
PL55.72 / 44.3173.53 / 58.4973.10 / 40.2657.68 / 31.1570.31 / 56.2168.14 / 37.37
TAROT55.72 / 46.5373.85 / 58.9377.78 / 42.6260.12 / 34.6973.08 / 58.1469.93 / 38.95
Pr → ArPr → ClPr → RwRw → ArRw → ClRw → PrAvg.
ARTUDA8.74 / 1.7338.92 / 14.5747.53 / 5.3030.12 / 2.1041.79 / 22.2945.19 / 13.9727.03 / 7.86
RFA42.03 / 2.5647.86 / 13.3163.44 / 5.2154.27 / 3.3854.27 / 16.7072.85 / 9.7855.00 / 8.49
SRoUDA59.21 / 31.4048.52 / 39.7775.40 / 41.2070.87 / 40.3853.40 / 42.4383.26 / 67.3857.97 / 33.42
PL57.15 / 29.5852.74 / 44.4074.82 / 40.7468.97 / 38.2857.03 / 45.1782.86 / 66.6266.00 / 44.38
TAROT59.62 / 32.7654.98 / 47.0178.66 / 42.9471.90 / 40.7959.45 / 49.7484.39 / 68.5368.29 / 46.80
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.486, + 0.572 + ], + "angle": 0, + "content": "accuracies of a standard training algorithm (MDD) on the Office-31, Office-Home, and VisDA-2017 datasets are exactly 0.0, highlighting the importance of robust adaptation methods. This is mainly attributed to our use of larger perturbation budgets during training and evaluation compared to those used in the original papers. For reference, we provide results using smaller perturbation budgets \\((4/255, 8/255)\\) on Office-31 and OfficeHome datasets in the Appendix. Overall, TAROT outperforms existing algorithms across all benchmark datasets. Note that the robust accuracies of a standard training algorithm (MDD) on the Office-31, Office-Home, and VisDA-2017 datasets are exactly 0.0." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.611, + 0.427, + 0.626 + ], + "angle": 0, + "content": "5.1.2. Essentially Domain-Invariant Robustness" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.639, + 0.486, + 0.901 + ], + "angle": 0, + "content": "In this subsection, we empirically highlight the limitations of the pseudo labeling approach, which focuses solely on improving target domain performance without effectively learning domain-invariant features. We select tasks with Real world (Rw) as the target domain on the OfficeHome dataset. We compare TAROT (our algorithm), which is designed to reduce the distribution gap between source and target domains, with PL, evaluating both the source domain accuracy and the performance on unseen domains. Table 3 shows that our method not only adapts well to the target domain but also learns more generalizable and robust features, leading to improved robustness across domains by significant margins. We present an upper bound for robust risk on the source domain and demonstrate that the components of this bound closely align with those of the target domain, offering a partial theoretical explanation for TAROT's robustness to adversarial attacks on the source domain. Additional results for other tasks are provided in Sec. 10." + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.395, + 0.909, + 0.49 + ], + "angle": 0, + "content": "Table 3. Performances of PL and TAROT on Source Domain and Unseen Domain, on OfficeHome (\\(\\varepsilon = 8/255\\)). In each cell, the first number is the standard accuracy (\\%), while the second number corresponds to the robust accuracy (\\%) for AA. Bold numbers indicate the best performance. A \\(\\rightarrow\\) B(C) indicates that A, B, and C represent the source, (unlabeled) target, and evaluation domains, respectively." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.494, + 0.922, + 0.687 + ], + "angle": 0, + "content": "
SourceUnseen
MethodAr → Rw(Ar)Ar → Rw(Cl)Ar → Rw(Pr)Avg.
ARTUDA18.83 / 2.647.70 / 0.896.28 / 0.4510.94 / 1.33
RFA99.62 / 47.0242.52 / 21.5351.57 / 21.5164.57 / 30.02
SRoUDA39.39 / 17.4339.54 / 28.6650.75 / 35.8642.23 / 27.32
PL41.78 / 19.4141.97 / 31.3952.42 / 37.2445.39 / 29.34
TAROT98.35 / 64.2447.86 / 35.1958.32 / 39.5468.18 / 46.32
Cl → Rw(Cl)Cl → Rw(Ar)Cl → Rw(Pr)Avg.
ARTUDA84.81 / 55.8014.34 / 3.2127.26 / 13.9942.14 / 24.33
RFA98.35 / 80.1040.37 / 8.8654.86 / 22.2564.53 / 37.07
SRoUDA48.29 / 35.5833.87 / 15.6248.19 / 33.0743.45 / 28.09
PL48.75 / 35.9536.30 / 16.6950.80 / 35.6645.28 / 29.43
TAROT93.65 / 84.3539.72 / 17.1855.46 / 37.6762.95 / 46.40
Pr → Rw(Pr)Pr → Rw(Ar)Pr → Rw(Cl)Avg.
ARTUDA98.29 / 60.1725.67 / 4.6236.24 / 16.2453.40 / 27.01
RFA99.75 / 68.5940.67 / 7.8741.76 / 19.7360.73 / 32.06
SRoUDA61.48 / 41.4034.82 / 16.0742.08 / 30.6846.13 / 29.38
PL61.05 / 43.3436.09 / 17.1042.11 / 30.9146.42 / 30.45
TAROT96.13 / 82.4339.14 / 16.1546.87 / 33.0160.71 / 43.86
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.681, + 0.711 + ], + "angle": 0, + "content": "5.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.908, + 0.763 + ], + "angle": 0, + "content": "We conduct ablation studies about: (i) effect of Robust-PT (ii) sensitivity analysis of \\(\\alpha\\). All exact values and additional experimental results are presented in Sec. 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.785, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Effect of Robust-PT on Various \\(\\varepsilon\\) Figures 3a and 3b show that the both algorithms perform significantly better when Robust-PT is applied, particularly as \\(\\varepsilon\\) increases, indicating that Robust-PT is crucial for robustness against strong attack. Without Robust-PT, both TAROT and PL experience sharp declines in accuracy, especially as \\(\\varepsilon\\) becomes larger. Figures 3c and 3d illustrate that TAROT uniformly outperforms PL across different perturbation budgets \\((\\varepsilon)\\) both on" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25786" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.089, + 0.643, + 0.13 + ], + "angle": 0, + "content": "Table 4. Performances of ARTUDA, RFA, SROUDA, PL and TAROT on DomainNet \\((\\varepsilon = 4 / 255)\\). In each cell, the first number is the standard accuracy \\((\\%)\\), and the second number is the robust accuracy \\((\\%)\\) for AA. Bold numbers indicate the best performance." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.135, + 0.625, + 0.496 + ], + "angle": 0, + "content": "
MethodC → IC → PC → RC → SAvg.
ARTUDA12.64 / 3.7631.15 / 7.9044.90 / 16.4939.90 / 17.5832.15 / 11.43
RFA9.42 / 3.4723.33 / 7.6037.54 / 17.7427.40 / 12.1724.42 / 10.24
SRoUDA9.27 / 5.6824.33 / 15.0742.91 / 28.3832.42 / 23.8227.23 / 18.24
PL10.59 / 6.6927.06 / 15.7343.57 / 28.6832.13 / 23.5228.34 / 18.66
TAROT14.65 / 8.1633.23 / 17.3650.02 / 30.4640.11 / 27.7734.50 / 20.94
I → CI → PI → RI → SAvg.
ARTUDA31.89 / 12.0720.38 / 2.5623.79 / 5.4734.61 / 5.9027.67 / 6.50
RFA15.00 / 8.8214.80 / 4.8019.15 / 8.7611.44 / 5.0515.10 / 6.86
SRoUDA25.95 / 14.1824.27 / 13.9336.32 / 22.0620.08 / 13.5226.65 / 15.92
PL17.45 / 13.1622.42 / 13.4132.50 / 21.9418.25 / 13.1422.65 / 15.41
TAROT29.36 / 20.2327.66 / 14.3039.14 / 23.3921.89 / 14.0829.51 / 18.00
P → CP → IP → RP → SAvg.
ARTUDA42.89 / 20.5814.36 / 3.2149.03 / 16.6234.02 / 13.1235.08 / 13.38
RFA26.27 / 14.859.45 / 3.0937.47 / 17.0419.67 / 8.5923.22 / 10.89
SRoUDA36.69 / 28.2610.17 / 5.8645.69 / 29.9530.95 / 19.1430.87 / 20.80
PL33.86 / 26.6911.18 / 7.2544.56 / 29.4729.76 / 20.6629.84 / 21.02
TAROT41.27 / 30.5114.82 / 7.6951.10 / 30.7637.22 / 24.7136.10 / 23.42
R → CR → IR → PR → SAvg.
ARTUDA52.94 / 28.1016.94 / 4.1741.95 / 12.0236.11 / 14.8636.98 / 14.79
RFA36.72 / 19.3012.36 / 3.0232.03 / 8.3422.73 / 8.1225.96 / 9.70
SRoUDA44.33 / 34.3310.10 / 5.6340.26 / 23.2629.10 / 18.5630.94 / 20.44
PL42.03 / 33.2711.85 / 7.1439.26 / 22.8029.05 / 19.6830.55 / 20.72
TAROT52.26 / 37.5818.86 / 9.4547.62 / 23.4339.38 / 25.2339.53 / 23.92
S → CS → IS → PS → RAvg.
ARTUDA54.24 / 30.5913.14 / 3.8934.87 / 7.7941.40 / 13.4735.91 / 13.94
RFA33.87 / 19.949.47 / 3.5427.29 / 7.8330.39 / 13.7025.26 / 11.25
SRoUDA46.62 / 36.658.50 / 5.2831.09 / 15.1442.95 / 22.3332.29 / 19.85
PL42.48 / 34.668.95 / 6.1831.48 / 18.7542.55 / 28.2431.37 / 21.96
TAROT53.35 / 41.3813.71 / 7.5239.77 / 20.4348.77 / 30.0238.90 / 24.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.646, + 0.085, + 0.898, + 0.14 + ], + "angle": 0, + "content": "Table 5. Performances of ARTUDA, RFA, SRoUDA, PL and TAROT on VisDA2017 \\((\\varepsilon = 8 / 255)\\). Standard accuracy \\((\\%)\\) / Robust accuracy \\((\\%)\\) for AA." + }, + { + "type": "table", + "bbox": [ + 0.649, + 0.144, + 0.895, + 0.219 + ], + "angle": 0, + "content": "
MethodSyn. → RealSyn.
ARTUDA10.73 / 0.0061.43 / 5.70
RFA64.86 / 8.6599.41 / 33.55
SRoUDA24.44 / 0.2048.56 / 24.05
PL66.71 / 37.8340.98 / 24.09
TAROT67.12 / 37.7785.18 / 51.21
" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.236, + 0.91, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.356, + 0.908, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.473, + 0.896, + 0.499 + ], + "angle": 0, + "content": "Figure 2. Sensitivity Analysis of \\(\\alpha\\). \\(\\alpha = 0\\) corresponds to PL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.66 + ], + "angle": 0, + "content": "the standard accuracies and robust accuracies against AA. Especially, that performance gap is increasing as \\(\\varepsilon\\) increases. Figures 3a and 3b show that both algorithms perform significantly better when Robust-PT is applied, particularly as \\(\\varepsilon\\) increases, indicating that Robust-PT is crucial for robustness against strong attacks. Figures 3c and 3d illustrate that TAROT uniformly outperforms the PL approach across different perturbation budgets \\((\\varepsilon)\\), both in standard accuracies and robust accuracies against AA. Notably, the performance gap increases as \\(\\varepsilon\\) increases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.484, + 0.785 + ], + "angle": 0, + "content": "Sensitivity Analysis of \\(\\alpha\\) Figure 2 illustrates the sensitivity with respect to \\(\\alpha\\) in Eq. (20). It shows how the performance of generalization and robustness changes for the target and unseen domains as the value of \\(\\alpha\\) varies. As shown in the figure, each task has a different optimal \\(\\alpha\\), indicating that the optimal \\(\\alpha\\) value for one task may not be optimal for another. The detailed numbers are provided in Sec. 10.3." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.802, + 0.376, + 0.818 + ], + "angle": 0, + "content": "6. Conclusions and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.828, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To our best knowledge, this is the first work for adversarially robust domain adaptation algorithms grounded with theoretical explanations. Compared to existing works, our algorithm shows essentially domain invariant robust features, resulting significant improved performance in source and" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.513, + 0.691, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.609, + 0.669, + 0.621 + ], + "angle": 0, + "content": "(a) PL w/ and w/o R-PT" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.623, + 0.691, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.719, + 0.663, + 0.731 + ], + "angle": 0, + "content": "(c) Rob Acc. w/ R-PT" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.513, + 0.903, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.609, + 0.894, + 0.621 + ], + "angle": 0, + "content": "(b) TAROT w/ and w/o R-PT" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.623, + 0.903, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.719, + 0.88, + 0.731 + ], + "angle": 0, + "content": "(d) Stand Acc. w/ R-PT" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.743, + 0.908, + 0.784 + ], + "angle": 0, + "content": "Figure 3. Effect of Robust-PT across various methods and \\(\\varepsilon\\) values. Each plot compares different configurations of PL w/ MDD and TAROT for standard and robust accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.8, + 0.909, + 0.901 + ], + "angle": 0, + "content": "unseen domains. Our work not only serves as a cornerstone for understanding robust domain adaptation from a theoretical perspective, but we also believe TAROT can be served as a state-of-the-art algorithm in practical perspective by excelling researches of adversarial robustness for transfer learning such as domain adaptation/generalization and few-shot learning task." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25787" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.241 + ], + "angle": 0, + "content": "Acknowledgement This work is partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-II220184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics), National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(No. 2022R1A5A7083908) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) [NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)]." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.264, + 0.188, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[1] Eric Arazo, Diego Ortega, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.409, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[3] Muhammad Awais, Fengwei Zhou, Hang Xu, Lanqing Hong, Ping Luo, Sung-Ho Bae, and Zhenguo Li. Adversarial robustness for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8568-8577, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.484, + 0.542 + ], + "angle": 0, + "content": "[4] Francesco Barbato, Marco Toldo, Umberto Michieli, and Pietro Zanuttigh. Latent space regularization for unsupervised domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2835-2845, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.542, + 0.484, + 0.582 + ], + "angle": 0, + "content": "[5] Peter L Bartlett, Michael I Jordan, and Jon D McAuliffe. Convexity, classification, and risk bounds. Journal of the American Statistical Association, 101(473):138-156, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[6] Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[7] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79: 151-175, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.755 + ], + "angle": 0, + "content": "[8] Mingcai Chen, Hao Cheng, Yuntao Du, Ming Xu, Wenyu Jiang, and Chongjun Wang. Two wrongs don't make a right: Combating confirmation bias in learning with label noise. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14765-14773, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.755, + 0.484, + 0.82 + ], + "angle": 0, + "content": "[9] Moustapha Cisse, Piotr Bojanowski, Edouard Grave, Yann Dauphin, and Nicolas Usunier. Parseval networks: Improving robustness to adversarial examples. In International conference on machine learning, pages 854-863. PMLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.821, + 0.484, + 0.874 + ], + "angle": 0, + "content": "[10] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Antonio Galbis and Manuel Maestre. Vector analysis versus vector calculus. Springer Science & Business Media, 2012. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.288, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.134, + 0.907, + 0.2 + ], + "angle": 0, + "content": "[13] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.201, + 0.907, + 0.269 + ], + "angle": 0, + "content": "[14] Zhiqiang Gao, Shufei Zhang, Kaizhu Huang, Qiufeng Wang, Rui Zhang, and Chaoliang Zhong. Certifying better robust generalization for unsupervised domain adaptation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2399-2410, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.27, + 0.906, + 0.324 + ], + "angle": 0, + "content": "[15] Chao Gong, Kai Chen, Zhipeng Wei, Jingjing Chen, and Yulong Jiang. Reliable and efficient concept erasure of text-to-image diffusion models. In In European Conference on Computer Vision (ECCV), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.325, + 0.907, + 0.378 + ], + "angle": 0, + "content": "[16] Fredrik D. Johansson, David Sontag, and Rajesh Ranganath. Support and invertibility in domain-invariant representations. International Conference on Artificial Intelligence and Statistics, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[17] Matt Jordan and Alexandros G Dimakis. Exactly computing the local lipschitz constant of relu networks. Advances in Neural Information Processing Systems, 33:7344-7353, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[18] Bo Fu Jinguang Jiang, Baixu Chen and Mingsheng Long. Transfer-learning-library. https://github.com/thuml/Transfer-Learning-Library, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.529 + ], + "angle": 0, + "content": "[19] Changhoon Kim, Kyle Min, and Yezhou Yang. R.a.c.e.: Robust adversarial concept erasure for secure text-to-image diffusion model. In In European Conference on Computer Vision (ECCV), 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[20] Shao-Yuan Lo and Vishal M. Patel. Exploring adversarially robust training for unsupervised domain adaptation. In Asian Conference on Computer Vision (ACCV), 2022. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.906, + 0.638 + ], + "angle": 0, + "content": "[21] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations (ICLR), 2018. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.64, + 0.906, + 0.693 + ], + "angle": 0, + "content": "[22] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. In In International Conference on Learning Representations (ICLR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.694, + 0.906, + 0.722 + ], + "angle": 0, + "content": "[23] Afshin Rostamizadeh Mehryar Mohri and Ameet Talwalkar. Foundations of Machine Learning. MIT Press, 2012. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.723, + 0.906, + 0.763 + ], + "angle": 0, + "content": "[24] Farzeen Munir, Shoaib Azam, and Moongu Jeon. Sstn: Self-supervised domain adaptation thermal object detection for autonomous driving, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.764, + 0.906, + 0.804 + ], + "angle": 0, + "content": "[25] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[26] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[27] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25788" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "European Conference on Computer Vision (ECCV), pages 213-226. Springer, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.175 + ], + "angle": 0, + "content": "[28] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? In Conference on Neural Information Processing Systems (NeurIPS), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[29] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. In In International Conference on Machine Learning (ICML), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.245, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[30] Zhouxing Shi, Yihan Wang, Huan Zhang, J Zico Kolter, and Cho-Jui Hsieh. Efficiently computing local lipschitz constants of neural networks via bound propagation. Advances in Neural Information Processing Systems, 35:2350-2364, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.313, + 0.484, + 0.365 + ], + "angle": 0, + "content": "[31] Peeyush Singhal, Rahee Walambe, Sheela Ramanna, and Ketan Kotecha. Domain adaptation: Challenges, methods, datasets, and applications. IEEE Access, 11:6973-7020, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.484, + 0.421 + ], + "angle": 0, + "content": "[32] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations (ICLR), 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.423, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[33] Michel Talagrand. Upper and lower bounds for stochastic processes. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.451, + 0.484, + 0.491 + ], + "angle": 0, + "content": "[34] Marvin Teichmann, Michael Weber, Marius Zoellner, Roberto Cipolla, and Raquel Urtasun. Multinet: Real-time joint semantic reasoning for autonomous driving, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.493, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[35] Thulasi Tholeti and Sheetal Kalyani. The robust way to stack and bag: the local lipschitz way. arXiv preprint arXiv:2206.00513, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[36] Jonathan Uesato, Jean-Baptiste Alayrac, Po-Sen Huang, Robert Stanforth, Alhussein Fawzi, and Pushmeet Kohli. Are labels required for improving adversarial robustness? In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.484, + 0.668 + ], + "angle": 0, + "content": "[37] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018-5027, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.67, + 0.484, + 0.723 + ], + "angle": 0, + "content": "[38] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International Conference on Learning Representataions(ICLR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.725, + 0.484, + 0.778 + ], + "angle": 0, + "content": "[39] Boxi Wu, Jinghui Chen, Deng Cai, Xiaofei He, and Quanquan Gu. Do wider neural networks really help adversarial robustness? Advances in Neural Information Processing Systems, 34:7054-7067, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.779, + 0.484, + 0.846 + ], + "angle": 0, + "content": "[40] Dongyoon Yang, Insung Kong, and Yongdai Kim. Enhancing adversarial robustness in low-label regime via adaptively weighted regularization and knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4552-4561, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.848, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[41] Dongyoon Yang, Insung Kong, and Yongdai Kim. Improving adversarial robustness by putting more regularizations on less robust samples. In International Conference on Machine Learning, pages 39331-39348. PMLR, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.146 + ], + "angle": 0, + "content": "[42] Yao-Yuan Yang, Cyrus Rashtchian, Hongyang Zhang, Ruslan Salakhutdinov, and Kamalika Chaudhuri. A closer look at accuracy vs. robustness. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4, 6, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.148, + 0.906, + 0.2 + ], + "angle": 0, + "content": "[43] Bohang Zhang, Du Jiang, Di He, and Liwei Wang. Rethinking lipschitz neural networks and certified robustness: A boolean function perspective. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 4, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.203, + 0.908, + 0.268 + ], + "angle": 0, + "content": "[44] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P Xing, Laurent El Ghaoui, and Michael I Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning (ICML), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.27, + 0.908, + 0.322 + ], + "angle": 0, + "content": "[45] Yuchen Zhang, Tianle Liu, Mingsheng Long, and Michael I. Jordan. Bridging theory and algorithm for domain adaptation. In International Conference on Machine Learning (ICML), 2019. 2, 3, 4, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.324, + 0.908, + 0.388 + ], + "angle": 0, + "content": "[46] Wanqing Zhu, Jia-Li Yin, Bo-Hao Chen, and Xineng Liu. Srouda: meta self-training for robust unsupervised domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3852-3860, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.908, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25789" + } + ] +] \ No newline at end of file diff --git a/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_origin.pdf b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d56390e7b22faab307633c704e7960f3442718a0 --- /dev/null +++ b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/ec4c8b5a-9e2a-4f15-83e9-9d7b778961e8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1d9b416d6d10b79002cd50b498ae4c0fce64e49bfdfc3a06b8e64241897bdde +size 1159467 diff --git a/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/full.md b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c48978de3b9f6440d9835270a38aaa97feaed090 --- /dev/null +++ b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/full.md @@ -0,0 +1,423 @@ +# TAROT: Towards Essentially Domain-Invariant Robustness with Theoretical Justification + +Dongyoon Yang $^{1*†}$ Jihu Lee $^{2*}$ Yongdai Kim $^{2†}$ + +$^{1}$ AI Advanced Technology, SK Hynix + +$^{2}$ Department of Statistics, Seoul National University + +ydy0415@gmail.com, rieky0426@snu.ac.kr, ydkim903@snu.ac.kr + +# Abstract + +Robust domain adaptation against adversarial attacks is a critical research area that aims to develop models capable of maintaining consistent performance across diverse and challenging domains. In this paper, we derive a new generalization bound for robust risk on the target domain using a novel divergence measure specifically designed for robust domain adaptation. Building upon this, we propose a new algorithm named TAROT, which is designed to enhance both domain adaptability and robustness. Through extensive experiments, TAROT not only surpasses state-of-the-art methods in accuracy and robustness but also significantly enhances domain generalization and scalability by effectively learning domain-invariant features. In particular, TAROT achieves superior performance on the challenging DomainNet dataset, demonstrating its ability to learn domain-invariant representations that generalize well across different domains, including unseen ones. These results highlight the broader applicability of our approach in real-world domain adaptation scenarios. + +# 1. Introduction + +Recent advances in machine learning have highlighted the importance of developing transferable models that can maintain performance across different domains. However, collecting labeled data from all potential domains is too resource-intensive for many practical applications. To address this limitation, Unsupervised Domain Adaptation (UDA) has emerged as a type of transfer learning that leverages labeled source-domain data to improve performance on a target domain without needing labels from that domain. UDA has proven pivotal in diverse machine learning tasks, largely because distributional discrepancies between source and target domains often lead to severe performance degradation [4, 16, 24, 31, 34]. + +Meanwhile, adversarial robustness has gained significant attention as a crucial element of secure and trustworthy AI. Pioneering works demonstrated the vulnerability of neural + +networks to small, carefully designed perturbations called adversarial attacks [21, 32] and proposed adversarial training methods [21, 38, 40, 41, 44]. Recently, the adversarial robustness community has expanded its focus to multimodal tasks (e.g., text-to-image, image-to-text), recognizing that large-scale multi-modal models are also susceptible to adversarial manipulations in one or more input modalities [15, 19, 22, 29]. These cross-modal attacks can exploit alignment mechanisms to trick models into misclassification or erroneous content generation. + +Given these two developments—domain adaptation and adversarial robustness—there is a growing interest in learning transferable models in UDA settings that remain resilient to adversarial perturbations in the target domain [3, 14, 20, 46]. A common strategy in existing robust UDA methods is to rely on pseudo-labels to handle unlabeled target data. However, in this paper, we propose a novel robust margin disparity metric for comparing source and target distributions, enabling us to derive a refined generalization bound on the robust risk in the target domain. From our bound, we derive two key insights: (1) using only pseudo-labels may be insufficient, but combining pseudo-labels with distributional alignment significantly enhances robust UDA performance; and (2) initializing robust UDA algorithms with a robustly pre-trained model (Robust-PT [3]) is notably more effective than starting from a non-robust pre-trained model. + +Building upon these theoretical foundations, we propose TAROT (Transfer Adversarially RObust Training), a new algorithm that leverages our theoretical insights to achieve superior performance in robust UDA tasks. Through extensive experiments on various benchmark datasets, we demonstrate that TAROT outperforms existing algorithms. Furthermore, we empirically demonstrate that TAROT exhibits distinct advantages in terms of scalability compared to pseudo-labeling approaches, particularly with significant performance improvements not only on the target domain but also on source and unseen source domains. This work bridges the gap between theoretical understanding and practical implementation in robust transfer learning, offering both fundamental insights and actionable solutions for real-world applications. + +Our contributions are summarized as follows: + +- We introduce a novel robust margin disparity measure and derive a refined generalization bound for robust UDA. +- From the derived generalization bound, we propose TAROT (Transferring Adversarial RObust Training), which integrates pseudo-labeling and distributional alignment to enhance adversarial robustness in UDA. +- TAROT outperforms existing robust UDA methods across multiple benchmarks, demonstrating strong performance on the target, source, and unseen source domains, indicating that TAROT effectively facilitates domain-invariant robust representations. +- We provide theoretical insights and practical solutions for scalable and adversarially robust domain adaptation in real-world scenarios. + +# 2. Preliminaries + +In this section, we outline the core concepts and notations for our analysis, beginning with adversarial training and its risk formulation for robust learning. We then describe the unsupervised domain adaptation (UDA) framework, highlighting distributional notations, the formal UDA objective, and existing theoretical approaches like DANN [13] and MDD [45]. These elements provide the groundwork for our subsequent derivation of a robust UDA upper bound and the proposed learning algorithm. + +# 2.1. Adversarial Training + +Let $\mathcal{X} \in \mathbb{R}^d$ represent the input space, and let $\mathcal{Y} = \{1, \ldots, C\}$ be the label space, where $C$ represents the number of classes. A classifier can be represented using a score function $f: \mathcal{X} \to \mathbb{R}^C$ , where $f_c$ denotes the $c$ -th component of $f$ . The classifier is then defined as: $h_f(\pmb{x}) = \operatorname*{argmax}_{c \in \mathcal{Y}} f_c(\pmb{x})$ . Let $\mathcal{B}_p(\pmb{x}, \varepsilon) = \{\pmb{x}' \in \mathcal{X}: \| \pmb{x} - \pmb{x}' \|_p \leq \varepsilon\}$ denote an $\varepsilon$ -ball centered at $\pmb{x}$ . + +The population risk considering adversarial robustness is defined as follows: + +$$ +\mathcal {R} _ {\mathrm {r o b}} (f) = \mathbb {E} _ {(\mathbf {X}, Y)} \max _ {\mathbf {X} ^ {\prime} \in \mathcal {B} _ {p} (\mathbf {X}, \varepsilon)} \mathbb {1} \left\{Y \neq h _ {f} \left(\mathbf {X} ^ {\prime}\right) \right\}, \tag {1} +$$ + +where $\mathbb{1}(\cdot)$ denotes the indicator function. The goal of adversarial training is to estimate the optimal $f$ , the minimizer of Eq. (1). + +# 2.2. Unsupervised Domain Adaptation + +There are two distributions, the source distribution $S$ and the target distribution $\mathcal{T}$ . The empirical versions of $S$ and $\mathcal{T}$ are $\widehat{S} = \{(x_i^s,y_i^s)\}_{i = 1}^{n_s}$ and $\widehat{\mathcal{T}} = \{(x_j^t,y_j^t)\}_{j = 1}^{n_t}$ , respectively. In unsupervised domain adaptation, $\widehat{S}$ and only $\widehat{\mathcal{T}}_{\mathbf{X}} := \{x_j^t\}_{j = 1}^{n_t}$ are observed. i.e. the labels $y_{j}^{t}$ can not be observed. Denote the distributions of the input on the source and target domains by $S_{\mathbf{X}}$ and $\mathcal{T}_{\mathbf{X}}$ , respectively. + +For a given score function $f \in \mathcal{F}$ , the risk of a data distribution $\mathcal{D}$ is defined as + +$$ +\mathcal {R} _ {\mathcal {D}} (f) := \mathbb {E} _ {(\mathbf {X}, Y) \sim \mathcal {D}} \mathbb {1} \{Y \neq h _ {f} (\mathbf {X}) \}. \tag {2} +$$ + +The goal of UDA is to minimize $\mathcal{R}_{\mathcal{T}}(f)$ using $\widehat{S}$ and $\widehat{\mathcal{T}}_{\mathbf{X}}$ which can be done by building an upper bound of $\mathcal{R}_{\mathcal{T}}(h)$ . + +The pivotal learning algorithms for UDA are DANN (Domain-Adversarial Neural Network) [13] and MDD (Margin Disparity Discrepancy) [45]. Ben-David et al. [7] suggested a distance $d_{\mathcal{H}\Delta \mathcal{H}}(\mathcal{S}_{\mathbf{X}},\mathcal{T}_{\mathbf{X}})$ between $\mathcal{S}_{\mathbf{X}}$ and $\mathcal{T}_{\mathbf{X}}$ to give an upper bound of the robust risk, and Ganin et al. [13] utilized this upper bound to implement their learning algorithm DANN. DANN uses an adversarial neural network to minimize the distance between the source and target distributions, using gradient reversal layers (GRL). On the other hand, MDD [45] derives an upper bound of $\mathcal{R}_{\mathcal{T}}(f)$ as the sum of $\mathcal{R}_S^{(\rho)}(f)$ , a standard non-robust risk on the source domain, and $d_{f,\mathcal{F}}^{(\rho)}(\mathcal{S}_{\mathbf{X}},\mathcal{T}_{\mathbf{X}})$ , a distance between $\mathcal{S}_{\mathbf{X}}$ and $\mathcal{T}_{\mathbf{X}}$ , and develop the MDD algorithm based on this upper bound. The main distinction between the two upper bounds is that MDD utilizes a scoring function for multi-class classification, while DANN only employs a classifier for binary classification. + +The objective of our work is to derive an upper bound of the robust risk for the target domain by utilizing the risk in the source domain and the robust divergence between the target domain $\mathcal{T}$ and the source domain $S$ . In turn, inspired by the upper bound, we propose a novel robust UDA algorithm. + +# 3. Related Works + +While a significant body of research on adversarial training has focused on supervised learning and semi-supervised learning tasks [21, 36, 40, 41, 44], there has been relatively less attention given to its application in domain adaptation tasks. Only a few studies have explored adversarial training in the context of domain adaptation [3, 20, 46], highlighting the need for further investigation in this area. + +RFA [3] leverages a pretrained ImageNet model to develop a regularization term. When the trainable model is initialized with a pretrained robust model, the regularization term encourages the outputs of each layer of the trainable model to resemble those of the pretrained model. This facilitates the learning of robust features. This approach can be applied to existing UDA algorithms such as DANN [13] and MDD [45] to make them adversarially robust. + +ARTUDA [20] adopts a self-supervised adversarial training approach, where the predictions of the current model are used as pseudo-labels to generate adversarial examples for the target domain. It also generates adversarial examples for the source domain and applies the distance metrics used in DANN. However, there is no theoretical justification of the complex loss function used in ARTUDA. One key challenge in ARTUDA is that pseudo-labels are generated by the same + +model being trained. As a result, due to the inherent trade-off between robustness and generalization, as the model learns more robust features, pseudo-labels are getting more inaccurate. This issue becomes particularly pronounced when the perturbation budget $(\varepsilon)$ is large. + +SRoUDA [46] attempts to resolve this limitation by introducing a separate teacher model to assign pseudo-labels to the target data. By leveraging meta-learning, SRoUDA improves the accuracy of pseudo-labeling by the teacher model and dynamically performs adversarial training [21] on the target domain using these pseudo-labels. Despite the improvement, SRoUDA struggles to effectively capture domain-invariant features, as it does not explicitly consider the deviation of the distributions between the source and target domains. This leads to a decline in performance in the source as well as in the unseen domains, indicating a failure to capture what we call essentially domain-invariant robust features. + +# 4. Proposed Method + +In this section, we present the theoretical results for adversarially robust UDA. This is an extension of MDD [45] to its robust version, and the proofs are deferred to Sec. 8. + +# 4.1. Robust Margin Disparity Discrepancy + +In this subsection, we introduce a novel disparity measure and key propositions to establish the generalization bound. + +Denote the classifier $h_f$ for a given score function $f$ as + +$$ +h_{f}:\boldsymbol {x}\mapsto \operatorname *{argmax}_{y\in \mathcal{Y}}f_{y}(\boldsymbol {x}), +$$ + +and denote the margin of $f$ at a labeled sample $(x, y)$ as + +$$ +\mathcal {M} _ {f} (\boldsymbol {x}, y) = f _ {y} (\boldsymbol {x}) - \max _ {y ^ {\prime} \neq y} f _ {y ^ {\prime}} (\boldsymbol {x}). +$$ + +Definition 1. (Robust Disparity). For given $f, f' \in \mathcal{F}$ , the 0-1 robust disparity between $f$ and $f'$ on a distribution $\mathcal{D}_{\mathbf{X}}$ on $\mathcal{X}$ is defined by + +$$ +\operatorname {d i s p} _ {\mathcal {D} _ {\mathbf {X}}} ^ {r o b} \left(f ^ {\prime}, f\right) := \mathbb {E} _ {\mathbf {X} \sim \mathcal {D} _ {\mathbf {X}}} \max _ {\mathbf {X} ^ {\prime} \in \mathcal {B} _ {p} (\mathbf {X}, \varepsilon)} \mathbb {1} \left\{h _ {f ^ {\prime}} (\mathbf {X}) \neq h _ {f} \left(\mathbf {X} ^ {\prime}\right) \right\} \tag {3} +$$ + +The robust disparity is measured based on the adversarial robustness of the second model $f$ , while the first model $f'$ serves as a pseudo-label generator. Note that if the prediction of $f'$ is consistent to true label, it is the robust risk as defined in Eq. (1). + +As an extension of the margin loss in the non-robust setting [45] to its robust counterpart, we define the robust margin loss as + +$$ +\mathcal {R} _ {\mathcal {D}} ^ {\mathrm {r o b}, (\rho)} (f) := \mathbb {E} _ {(\mathbf {X}, Y) \sim \mathcal {D}} \max _ {\mathbf {X} ^ {\prime} \in \mathcal {B} _ {p} (\mathbf {X}, \varepsilon)} \Phi_ {\rho} \circ \mathcal {M} _ {f} \left(\mathbf {X} ^ {\prime}, Y\right), \tag {4} +$$ + +where + +$$ +\Phi_ {\rho} (x) := \left\{ \begin{array}{l l} 0 & \rho \leq x \\ 1 - x / \rho & 0 \leq x \leq \rho \\ 1 & x \leq 0 \end{array} \right. \tag {5} +$$ + +and $\circ$ is composite operator of functions. + +Definition 2. (Robust Margin Disparity). For given $f, f' \in \mathcal{F}$ , the Robust Margin Disparity is defined as + +$$ +\begin{array}{l} \operatorname {d i s p} _ {\mathcal {D} _ {\mathbf {X}}} ^ {r o b, (\rho)} \left(f ^ {\prime}, f\right) \\ := \mathbb {E} _ {\mathbf {X} \sim \mathcal {D} _ {\mathbf {X}}} \max _ {\mathbf {X} ^ {\prime} \in \mathcal {B} _ {\rho} (\mathbf {X}, \varepsilon)} \Phi_ {\rho} \circ \mathcal {M} _ {f} \left(\mathbf {X} ^ {\prime}, h _ {f ^ {\prime}} (\mathbf {X})\right) \tag {6} \\ \end{array} +$$ + +An important property of the robust margin disparity is that it is greater than the robust disparity, i.e. $\mathrm{disp}_{\mathcal{D}_{\mathbf{X}}}^{\mathrm{rob},(\rho)}(f', f) \geq \mathrm{disp}_{\mathcal{D}_{\mathbf{X}}}^{\mathrm{rob}}(f', f)$ . + +We now use this disparity term to bound the robust risk on the target distribution with additional feasible terms. + +Proposition 1. Let $S$ and $\mathcal{T}$ represent the distributions of the source and target domains, respectively. Similarly, let $S_{\mathbf{X}}$ and $\mathcal{T}_{\mathbf{X}}$ denote the marginal distributions of the source and target domains over $\mathbf{X}$ , respectively. For every score function $f \in \mathcal{F}$ , the following inequality holds: + +$$ +\begin{array}{l} \mathcal {R} _ {T} ^ {r o b} (f) \leq \\ \mathcal {R} _ {\mathcal {S}} ^ {(\rho)} (f) + \left\{\operatorname {d i s p} _ {\mathcal {T} _ {\mathbf {X}}} ^ {r o b, (\rho)} \left(f ^ {*}, f\right) - \operatorname {d i s p} _ {\mathcal {S} _ {\mathbf {X}}} ^ {(\rho)} \left(f ^ {*}, f\right) \right\} + \lambda , \tag {7} \\ \end{array} +$$ + +where $f^{*} = \operatorname *{argmin}_{f\in \mathcal{F}}\{\mathcal{R}_{\mathcal{T}}^{(\rho)}(f) + \mathcal{R}_{\mathcal{S}}^{(\rho)}(f)\}$ is ideal hypothesis and $\lambda = \lambda (\mathcal{F},\mathcal{S},\mathcal{T},\varepsilon ,\rho) = \mathcal{R}_{\mathcal{T}}^{(\rho)}(f^{*}) + \mathcal{R}_{\mathcal{S}}^{(\rho)}(f^{*})$ is constant of $f$ + +Note that the upper bound consists of standard margin risk of the source domain, not the robust margin risk. This enables the corresponding algorithm introduced in Sec. 4.4 computationally simpler because generation of adversarial samples is not required to calculate the standard margin risk on the source domain. + +From the above proposition, we can naturally induce the following discrepancy between the distributions of two domains, which we call as a robust margin disparity discrepancy. + +# Definition 3. (Robust Margin Disparity Discrepancy). + +Given a score function $f \in \mathcal{F}$ , the Robust Margin Disparity Discrepancy induced by $f' \in \mathcal{F}$ of two distributions $S_{\mathbf{X}}$ and $T_{\mathbf{X}}$ is defined as follows: + +$$ +\begin{array}{l} d _ {f, \mathcal {F}} ^ {r o b, (\rho)} (S _ {\mathbf {X}}, \mathcal {T} _ {\mathbf {X}}) \\ := \sup _ {f ^ {\prime} \in \mathcal {F}} \left\{\operatorname {d i s p} _ {\mathcal {T} _ {\mathbf {x}}} ^ {r o b, (\rho)} \left(f ^ {\prime}, f\right) - \operatorname {d i s p} _ {\mathcal {S} _ {\mathbf {x}}} ^ {(\rho)} \left(f ^ {\prime}, f\right) \right\} \tag {8} \\ \end{array} +$$ + +This discrepancy measures the difference of the robust margin disparity on target domain and the standard margin + +disparity on the source domain, where the objective $f$ is fixed. If there is significant gap between the adversarial target domain and standard source domain, the discrepancy becomes large. Now we introduce a non-asymptotic upper bound for the target robust risk: + +$$ +\mathcal {R} _ {\mathcal {T}} ^ {\mathrm {r o b}, (\rho)} (f) \leq \mathcal {R} _ {\mathcal {S}} ^ {(\rho)} (f) + d _ {f, \mathcal {F}} ^ {\mathrm {r o b}, (\rho)} \left(\mathcal {S} _ {\mathbf {X}}, \mathcal {T} _ {\mathbf {X}}\right) + \lambda \tag {9} +$$ + +with the same $\lambda$ in Proposition 1. + +# 4.2. Generalization bound + +In this subsection, we provide a generalization bound based on Eq. (9). We use the notion of local Lipschitz constant of a function, following Yang et al. [42]. + +Definition 4 (Local Lipschitz Constant). Consider an arbitrary real function $f: \mathbb{R}^d \to \mathbb{R}^C$ , and a $L_p$ -norm $\| \cdot \|_p$ . Define a point-wise local Lipschitz constant as + +$$ +L _ {f} (\boldsymbol {x}, \varepsilon) := \sup _ {\boldsymbol {x} ^ {\prime} \in \mathcal {B} _ {p} (\boldsymbol {x}, \varepsilon)} \frac {\| f \left(\boldsymbol {x} ^ {\prime}\right) - f (\boldsymbol {x}) \| _ {1}}{\| \boldsymbol {x} ^ {\prime} - \boldsymbol {x} \| _ {p}} \tag {10} +$$ + +for a point $\pmb{x}$ . Then, we define a local Lipschitz constant on $\mathcal{D}_{\mathbf{X}}$ as + +$$ +L _ {f} \left(\mathcal {D} _ {\mathbf {X}}, \varepsilon\right) := \sup _ {x \in \mathcal {D} _ {\mathbf {X}}} L _ {f} (x, \varepsilon) \tag {11} +$$ + +We call that $f$ is locally Lipschitz, if $L_{f}(\mathcal{D}_{\mathbf{X}},\varepsilon)$ is bounded (i.e. $L_{f}(\mathcal{D}_{\mathbf{X}},\varepsilon) < \infty$ ). + +We also define following auxiliary function spaces. + +Definition 5. Consider a set of score functions $\mathcal{F}$ and let $\mathcal{H}$ be the class of the induced classifiers $h_f$ for each $f\in \mathcal{F}$ . We define the following two sets. + +$$ +\Pi_ {\mathcal {H}} \mathcal {F} := \{x \mapsto f (x, h (x)) | h \in \mathcal {H}, f \in \mathcal {F} \} \tag {12} +$$ + +$$ +\Pi_ {1} \mathcal {F} := \{x \mapsto f (x, y) | y \in \mathcal {Y}, f \in \mathcal {F} \} \tag {13} +$$ + +It is known that $\Pi_{\mathcal{H}}\mathcal{F}$ takes a form of the space of inner products of vector fields from $\mathcal{H}$ to $\mathcal{F}$ [11, 45]. The notion $\Pi_1\mathcal{F}$ follows from Mehryar Mohri and Talwalkar [23]. + +We now derive a generalization bound for the robust risk on the target domain using Rademacher complexity, which quantifies the expressiveness of a given hypothesis space [23]. + +Definition 6. (Rademacher Complexity). Let $\mathcal{F}$ be a family of functions mapping from $\mathcal{Z} = \mathcal{X} \times \mathcal{Y}$ to $[a, b]$ and $\widehat{\mathcal{D}} = \{z_1, \dots, z_n\}$ are finite samples independently drawn from a distribution $\mathcal{D}$ on $\mathcal{Z}$ . Then, the empirical Rademacher complexity of $\mathcal{F}$ with $\widehat{\mathcal{D}}$ is defined by + +$$ +\widehat {\mathfrak {R}} _ {\widehat {\mathcal {D}}} (\mathcal {F}) \triangleq \mathbb {E} _ {\sigma} \sup _ {f \in \mathcal {F}} \frac {1}{n} \sum_ {i = 1} ^ {n} \sigma_ {i} f (z _ {i}), \tag {14} +$$ + +where $\sigma_{i}$ s are independent random variables following the Rademacher distribution. In turn, the Rademacher complexity of $\mathcal{F}$ is defined by + +$$ +\mathfrak {R} _ {n, \mathcal {D}} (\mathcal {F}) \triangleq \mathbb {E} _ {\widehat {\mathcal {D}} \sim \mathcal {D} ^ {n}} \widehat {\mathfrak {R}} _ {\widehat {\mathcal {D}}} (\mathcal {F}) \tag {15} +$$ + +Now we provide our generalization bound, which replaces the terms in the upper bound (Eq. (9)) that depend on the population distribution by the terms depending on the empirical distributions. + +Theorem 1. (Generalization Bound on the Robust Risk of Target Distribution). For any $\delta > 0$ , with probability $1 - 3\delta$ , we have the following uniform generalization bound for any score function $f$ in $\mathcal{F}$ : + +$$ +\begin{array}{l} \mathcal {R} _ {\mathcal {T}} ^ {r o b} (f) \\ \leq \mathcal {R} _ {\widehat {\mathcal {S}}} ^ {(\rho)} (f) + d _ {f, \mathcal {F}} ^ {r o b, (\rho)} \left(\widehat {\mathcal {S}} _ {\mathbf {X}}, \widehat {\mathcal {T}} _ {\mathbf {X}}\right) + \lambda \\ + \frac {2 C ^ {2}}{\rho} \mathfrak {R} _ {n, S} (\Pi_ {1} \mathcal {F}) + \frac {2 C}{\rho} \mathfrak {R} _ {n, S} (\Pi_ {\mathcal {H}} \mathcal {F}) + 2 \sqrt {\frac {\log 2 / \delta}{2 n}} \\ + \frac {2 C}{\rho} \Re_ {m, \mathcal {T}} \left(\Pi_ {\mathcal {H}} \mathcal {F}\right) + \sqrt {\frac {\log 2 / \delta}{2 m}} + \frac {2 \varepsilon L _ {f} \left(\mathcal {T} _ {\mathbf {X}}, \varepsilon\right)}{\rho}, \tag {16} \\ \end{array} +$$ + +where $\lambda = \min_{f\in \mathcal{F}}\{\mathcal{R}_{\mathcal{T}}^{(\rho)}(f) + \mathcal{R}_{\mathcal{S}}^{(\rho)}(f)\}$ . + +# 4.3. Two insights + +In this subsection, we present two key insights for robust UDA. First, we analyze the impact of the local Lipschitz constant on generalization performance, showing why initializing robust UDA algorithms with a robustly pre-trained (Robust-PT) model can substantially improve outcomes. Second, we highlight why pseudo labeling alone may be insufficient, especially in large or complex target domains where confirmation bias and the lack of explicit domain alignment can lead to persistent misclassifications. + +About Local Lipschitz constant In Theorem 1, there is the local Lipschitz constant $L_{f}(\mathcal{T}_{\mathbf{X}},\varepsilon)$ . As discovered in previous literature, the local Lipschitz constant is known to be deeply involved with a generalization performance of a model [6, 17, 30]. This fact leads to various researches that investigate a method to insure provable robustness of a model [9, 35, 39]. It is also known that adversarial training empirically reduces the local Lipschitz constant of a neural network [42, 43]. + +In the numerical studies, we have found that robust UDA approaches with a Robust-PT as an initial improve the prediction performance much compared to those without a Robust-PT as an initial. This phenomenon can be partly explained by the local Lipschitz term. That is, model which is not initialized with Robust-PT could have a large value of the local Lipschitz constant and thus could have a too large robust risk value on the target domain. Hence, using an initial solution with a small local Lipschitz constant is promising and a Robust-PT would be such one. + +Pseudo labeling only may not be sufficient Pseudo labeling is a widely used technique in domain adaptation [20, 46], where the objective is to improve the performance of the + +model on the target domain by iteratively using the predictive labels of the currently trained model (or its variation) as labels for the target domain data when generating adversarial examples for the target data. There are at least two problems in those pseudo labeling approaches. + +The first problem is a risk of confirmation bias. Since the model generates pseudo labels based on its own predictions, early-stage errors in the model could become self-reinforcing. When incorrect pseudo labels are assigned with high confidence, these mistakes persist through training cycles, amplifying errors and leading to poor generalization on the target domain. This phenomenon is widely studied in semi-supervised settings [1, 8]. Similarly in the UDA setting, as the training phase progresses, the model may increasingly rely on those incorrect predictions, making it difficult to recover especially in complex or noisy target domains. + +The risk of confirmation bias can be reduced by generating pseudo labels from a pretrained non-robust UDA model called a (fixed) teacher model. In this case, we can write the objective function as + +$$ +\mathbb {E} _ {\mathbf {X} \sim \mathcal {T} _ {\mathbf {X}}} \max _ {\mathbf {X} ^ {\prime} \in \mathcal {B} _ {p} (\mathbf {X}, \varepsilon)} \mathbb {1} \left\{h _ {f} \left(\mathbf {X} ^ {\prime}\right) \neq h _ {f *} (\mathbf {X}) \right\} \tag {17} +$$ + +where $f^{*} = \operatorname*{argmin}_{f\in \mathcal{F}}\mathcal{R}_{\widehat{\mathcal{T}}}(\boldsymbol {f})$ . In our notation, this term exactly coincides with $\mathrm{disp}_{T_{\mathbf{x}}}^{\mathrm{rob}}(f^{*},f)$ . The following proposition justifies this approach. + +Proposition 2. For every score function $f \in \mathcal{F}$ , the following inequality holds: + +$$ +\mathcal {R} _ {\mathcal {T}} ^ {r o b} (f) \leq \operatorname {d i s p} _ {\mathcal {T} _ {\mathbf {X}}} ^ {r o b} \left(f ^ {*}, f\right) + \mathcal {R} _ {\mathcal {T}} \left(f ^ {*}\right) \tag {18} +$$ + +for arbitrary $f^{*}\in \mathcal{F}$ + +It can be easily seen through the intermediate steps of Proposition 1. By letting $f^{*}$ as a non-robust UDA model, the pseudo labeling approach with the teacher model $f^{*}$ can be understood to minimize the upper bound in Proposition 2. Our empirical studies amply confirm this insight. + +Pseudo labeling approaches with a fixed teacher model, however, still has limitations. In particular, the scalability of pseudo labeling across diverse or large datasets can be challenging due to the absence of explicit domain alignment mechanisms. As the size or complexity of the target domain grows, the chance of incorrect pseudo labels increases, which could overwhelm the overall learning process. This may happen especially in scenarios where the source and target domains differ significantly. This makes it harder for the model to generalize to more extensive or varied datasets, particularly when the domain gap becomes larger due to increasing size or diversity of data. + +Note that our generalization bound Theorem 1 uses the robust margin disparity discrepancy between the source and target domains instead of the robust disparity only on the target domain in Proposition 2. That is, our generalization bound integrates pseudo labeling and domain alignment + +together. In the following subsection, inspired by these insights, we introduce an algorithm designed to overcome the drawbacks of pseudo labeling approaches, tackling both confirmation bias and scalability while enhancing generalization abilities across diverse and complex target domains. + +Algorithm 1: Transferring Adversarily Robust Training (TAROT) +Input : $\psi$ : feature extractor, $\pi$ : head, $\pi'$ : auxiliary head, $s$ : standard trained model, $\mathcal{D}_{\mathcal{S}} = \{(x_{i}^{s},y_{i}^{s})\}_{i = 1}^{m}$ : source domain dataset, $\mathcal{D}_{\mathcal{S}} = \{(x_{i}^{s},y_{i}^{s})\}_{i = 1}^{m}:\mathcal{D}_{\mathcal{T}} = \{x_{j}^{t}\}_{j = 1}^{m^{\prime}}$ : target domain dataset, $\eta_1,\eta_2,\lambda ,\rho$ : learning rates in Eq. (20), $T$ : the number of epochs, $B$ : the number of batches, $K$ : batch size Output: adversarially robust network $\pi \circ \psi$ +1 $\theta_{\psi}\gets \theta_{\mathrm{R}}$ (initialized by a robustly trained feature extractor) +2 for $b = 1,\dots ,B$ do +3 for $k = 1,\dots ,K$ do +4 Generate $\widehat{\pmb{x}}_{b,k}^{\mathrm{t,adv}}$ by PGD $(\pi \circ \psi (\pmb{x}_{b,k}^{t}),\tilde{y}_{b,k}^{t})$ $\tilde{y}_{b,k}^{t} = h_{s}(\pmb{x}_{b,k}^{t})$ +5 Update $(\theta_{\psi},\theta_{\pi})\gets (\theta_{\psi},\theta_{\pi})-$ $\eta_1\frac{1}{K}\nabla_{(\pmb {\theta}_\psi ,\pmb {\theta}_\pi)}\ell_{\mathrm{TAROT}}(\pmb {\theta};\{(x_{b,k}^{\mathrm{s}},y_{b,k}^{\mathrm{s}})\}_{k = 1}^{K},\{(x_{b,k}^{\mathrm{t}})\}_{k = 1}^{K})$ +6 Update $\theta_{\pi '}\gets \theta_{\pi ' + }$ $\eta_2\frac{1}{K}\nabla_{\pmb{\theta}_{\pi '}}\ell_{\mathrm{TAROT}}(\pmb {\theta};\{(x_{b,k}^{\mathrm{s}},y_{b,k}^{\mathrm{s}})\}_{k = 1}^{K},\{(x_{b,k}^{\mathrm{t}})\}_{k = 1}^{K},\lambda)$ +7 end +8 end +9 Return $\pi \circ \psi$ + +# 4.4. Learning Algorithm + +Inspired by Theorem 1, we propose the robust domain adaptation algorithm, which we call Transferring Adversarially RObust Training (TAROT). Our basic idea is to find a optimal score function $f \in \mathcal{F}$ by solving a modified version of the following optimization problem: + +$$ +\min _ {f \in \mathcal {F}} \mathcal {R} _ {\widehat {\mathcal {S}}} ^ {(\rho)} (f) + d _ {f, \mathcal {F}} ^ {\operatorname {r o b}, (\rho)} \left(\widehat {\mathcal {S}} _ {\mathbf {X}}, \widehat {\mathcal {T}} _ {\mathbf {X}}\right) + L _ {f} \left(\widehat {\mathcal {T}} _ {\mathbf {X}}, \varepsilon\right) \tag {19} +$$ + +- As is done in DANN (Domain-Adversarial Neural Network) [13] and MDD (Margin Disparity Discrepancy) [45], we consider a neural network structure that is composed of a feature extractor $\psi$ and a head $\pi$ , i.e. a score function $f \in \mathcal{F}$ can be represented as $\pi \circ \psi$ . We only consider $\pi$ with $\psi$ being fixed, when calculating $d_{f,\mathcal{F}}^{\mathrm{rob},(\rho)}$ . That is, we calculate the robust margin disparity discrepancy on the feature space of $\psi$ . +- To reduce the first and second term, we follow the formulation of the MDD algorithm. Likewise, we utilize an auxiliary head $\pi'$ and $\gamma \coloneqq \exp \rho$ designed to attain the margin. $\pi'$ shares the same hypothesis space with $\pi$ , and works as an adversarial network. The discrepancy term, which is non-differentiable, can be optimized through GRL [12]. The only difference is that, when to calculate $d_{f,\mathcal{F}}^{\mathrm{rob},(\rho)}$ , + +![](images/174ce636fe189adfb9db2e7411e5b4fac1ad994f24129b9774bde2a0ab86f855.jpg) +Figure 1. Overview of TAROT algorithm. + +it requires calculating a robust disparity and a standard disparity rather than two standard disparities. + +- To reduce the local Lipschitz constant, we minimize the cross entropy loss with adversarial examples for given pseudo labels attached by a teacher model learned by a nonrobust UDA approach. This is based on the previously mentioned phenomenon that adversarial training reduces the local Lipschitz constant [42, 43]. +- When calculating the robust margin disparity discrepancy, instead of generating adversarial examples, we reuse those generated for the local Lipsshitz constant, making the algorithm computationally effective. +- As an initial model for $\psi$ , we used a robustly pretrained models with from the repository provided in [28]. + +The objective function of TAROT is given as: + +$$ +\begin{array}{l} \ell_ {\text {T A R O T}} (\boldsymbol {\theta}; \left\{\left(\boldsymbol {x} _ {i} ^ {\mathrm {s}}, y _ {i} ^ {\mathrm {s}}\right) \right\} _ {i = 1} ^ {m}, \left\{\left(\boldsymbol {x} _ {j} ^ {\mathrm {t}}\right) \right\} _ {j = m + 1} ^ {m + m ^ {\prime}}, \alpha , \gamma) \\ = \alpha \left\{ \right.\frac {1}{m} \sum_ {i = 1} ^ {m} \ell_ {\mathrm {c e}} \left(\left(\pi \circ \psi\right)\left(\boldsymbol {x} _ {i}\right), y _ {i}\right)\left. \right)\left. \right. \\ + \left(\frac {1}{m ^ {\prime}} \sum_ {j = m + 1} ^ {m + m ^ {\prime}} \ell_ {\mathrm {m o d - c e}} ^ {\mathrm {r o b}} \left(\left(\pi^ {\prime} \circ \psi\right) \left(\boldsymbol {x} _ {j}\right), h _ {\pi \circ \psi} \left(\boldsymbol {x} _ {j}\right)\right) \right. \tag {20} \\ \left. \left. - \frac {\gamma}{m} \sum_ {j = 1} ^ {m} \ell_ {\mathrm {c e}} \left(\left(\pi^ {\prime} \circ \psi\right) \left(\boldsymbol {x} _ {j}\right), h _ {\pi \circ \psi} \left(\boldsymbol {x} _ {j}\right)\right)\right) \right\} \\ + \frac {1}{m ^ {\prime}} \sum_ {j = m + 1} ^ {m + m ^ {\prime}} \ell_ {\mathrm {c e}} ^ {\mathrm {r o b}} ((\pi \circ \psi) (\boldsymbol {x} _ {j}), \tilde {y} _ {j}) \\ \end{array} +$$ + +where $\ell_{\mathrm{ce}}$ , $\ell_{\mathrm{mod - ce}}^{\mathrm{rob}}$ [45] and $\ell_{\mathrm{ce}}^{\mathrm{rob}}$ [21] as cross-entropy, modified cross-entropy with adversarial examples and cross-entropy with adversarial examples respectively. Also, $\pmb{\theta} = (\pmb{\theta}_{\psi},\pmb{\theta}_{\pi},\pmb{\theta}_{\pi'})$ to represent the parameters corresponding to $\psi, \pi, \pi'$ . $\tilde{y}_j = h_s(\pmb{x}_j)$ is a pseudo label generated by a standard trained model $s$ and $\alpha$ is a trade-off parameter that balances transferability and target's robustness. TAROT algorithm is summarized in Algorithm 1 and Figure 1. + +# 5. Experiments + +In this section, we compare TAROT with other competitors for adversarially robust UDA. We use several benchmark datasets: Office-31 [27], Office-Home [37], VisDA2017 [25], and DomainNet [26]. Throughout these datasets, effectiveness of TAROT is evaluated across a wide range of domain shifts and robustness scenarios. Further experimental details including explanation for datasets and experimental setups are presented in Sec. 9. The code is available in the repository $\ddagger$ . + +# 5.1. Performance Evaluation + +Training Setup All of the results are averaged over three runs, with different teacher model initializations. Different perturbation budgets $(\varepsilon)$ are chosen based on the difficulty of each benchmark dataset. For more challenging datasets, a smaller $\varepsilon$ is used, and vice versa. Specifically, we use $\varepsilon = 16 / 255$ for the Office-31 and Office-Home datasets, while $\varepsilon = 8 / 255$ and $\varepsilon = 4 / 255$ are used for VisDA2017 and DomainNet, respectively. The values of $\alpha$ in Eq. (20) are set to 0.05, 0.1, 0.5, and 1.0 for Office31, OfficeHome, VisDA2017, and DomainNet, respectively. We empirically observed that the optimal $\alpha$ values for each task vary much. The sensitivity analysis of $\alpha$ is provided in Sec. 5.2. + +As baselines, along with ARTUDA [20], RFA [3], and SRoUDA [46], we consider the pseudo labeling (PL) algorithm which is implemented as a simple PGD-AT algorithm where target pseudo labels are obtained from a fixed non-robustly trained teachers learned by MDD [45]. All algorithms are initialized with Robust-PT as is done for TAROT. Without Robust-PT, training with larger perturbations such as $\varepsilon = 16 / 255$ can lead to non-decreasing training loss in certain tasks, as discussed in Sec. 10.2. + +Evaluation Setup We report standard accuracies on clean samples and robust accuracies against AutoAttack [10]. AutoAttack, which includes both white-box and black-box attacks, is a powerful evaluation tool that mitigates errors in robustness measurement caused by gradient obfuscation—a phenomenon where model gradients are distorted or masked, leading to overestimated robustness [2]. AutoAttack is widely recognized for accurately assessing adversarial robustness by bypassing these gradient masking issues. + +# 5.1.1. Performance on Target Domains + +Tables 1, 2, 4 and 5 demonstrate that PL outperforms other algorithms, with the exception of TAROT. As seen in Figure 3, simply combining Robust-PT as an initial solution with pseudo labeling can yield significant performance gains, which confirms our first theoretical insight. (See Sec. 5.2.) This combination serves as a strong baseline in adversarially robust domain adaptation. It's noteworthy that ARTUDA [20], RFA [3], and SRoUDA [46] algorithms show remarkably poor performance on several tasks. Note that the robust + +Table 1. Performances of ARTUDA, RFA, SROUDA, PL and TAROT on Office31 ( $\varepsilon = 16/255$ ). In each cell, the first number is the standard accuracy (\%), while the second number is the robust accuracy (\%) for AA. Bold numbers indicate the best performance. + +
MethodA → DA → WD → AD → WW → AW → DAvg.
ARTUDA23.49 / 11.6522.39 / 9.5635.00 / 16.5168.55 / 35.8536.81 / 19.7474.10 / 51.6143.39 / 24.15
RFA78.72 / 1.2171.20 / 1.2655.84 / 18.0396.73 / 6.4260.10 / 23.0099.40 / 6.0277.00 / 9.32
SRoUDA93.37 / 42.3778.87 / 65.1617.93 / 3.4496.00 / 44.2866.53 / 47.21100.00 / 62.4574.62 / 44.15
PL91.77 / 84.1491.82 / 91.7074.19 / 67.8797.99 / 95.9872.42 / 65.46100.00 / 92.5788.03 / 82.95
TAROT94.18 / 84.1495.47 / 93.4674.51 / 67.1397.99 / 96.2373.73 / 66.88100.00 / 91.7789.31 / 83.27
+ +Table 2. Performances of ARTUDA, RFA, SRoUDA, PL and TAROT on OfficeHome (ε = 16/255). In each cell, the first number is the standard accuracy (%), while the second number is the robust accuracy (% for AA. Bold numbers indicate the best performance. + +
MethodAr → ClAr → PrAr → RwCl → ArCl → PrCl → Rw
ARTUDA30.29 / 13.5626.18 / 4.755.88 / 0.003.42 / 0.0025.34 / 10.4820.98 / 5.53
RFA48.39 / 17.8954.67 / 8.7663.74 / 6.3443.22 / 3.1758.03 / 9.1557.24 / 5.69
SRoUDA51.84 / 37.8975.60 / 45.4674.25 / 40.2857.97 / 33.4267.76 / 54.7767.32 / 35.83
PL55.72 / 44.3173.53 / 58.4973.10 / 40.2657.68 / 31.1570.31 / 56.2168.14 / 37.37
TAROT55.72 / 46.5373.85 / 58.9377.78 / 42.6260.12 / 34.6973.08 / 58.1469.93 / 38.95
Pr → ArPr → ClPr → RwRw → ArRw → ClRw → PrAvg.
ARTUDA8.74 / 1.7338.92 / 14.5747.53 / 5.3030.12 / 2.1041.79 / 22.2945.19 / 13.9727.03 / 7.86
RFA42.03 / 2.5647.86 / 13.3163.44 / 5.2154.27 / 3.3854.27 / 16.7072.85 / 9.7855.00 / 8.49
SRoUDA59.21 / 31.4048.52 / 39.7775.40 / 41.2070.87 / 40.3853.40 / 42.4383.26 / 67.3857.97 / 33.42
PL57.15 / 29.5852.74 / 44.4074.82 / 40.7468.97 / 38.2857.03 / 45.1782.86 / 66.6266.00 / 44.38
TAROT59.62 / 32.7654.98 / 47.0178.66 / 42.9471.90 / 40.7959.45 / 49.7484.39 / 68.5368.29 / 46.80
+ +accuracies of a standard training algorithm (MDD) on the Office-31, Office-Home, and VisDA-2017 datasets are exactly 0.0, highlighting the importance of robust adaptation methods. This is mainly attributed to our use of larger perturbation budgets during training and evaluation compared to those used in the original papers. For reference, we provide results using smaller perturbation budgets $(4/255, 8/255)$ on Office-31 and OfficeHome datasets in the Appendix. Overall, TAROT outperforms existing algorithms across all benchmark datasets. Note that the robust accuracies of a standard training algorithm (MDD) on the Office-31, Office-Home, and VisDA-2017 datasets are exactly 0.0. + +# 5.1.2. Essentially Domain-Invariant Robustness + +In this subsection, we empirically highlight the limitations of the pseudo labeling approach, which focuses solely on improving target domain performance without effectively learning domain-invariant features. We select tasks with Real world (Rw) as the target domain on the OfficeHome dataset. We compare TAROT (our algorithm), which is designed to reduce the distribution gap between source and target domains, with PL, evaluating both the source domain accuracy and the performance on unseen domains. Table 3 shows that our method not only adapts well to the target domain but also learns more generalizable and robust features, leading to improved robustness across domains by significant margins. We present an upper bound for robust risk on the source domain and demonstrate that the components of this bound closely align with those of the target domain, offering a partial theoretical explanation for TAROT's robustness to adversarial attacks on the source domain. Additional results for other tasks are provided in Sec. 10. + +Table 3. Performances of PL and TAROT on Source Domain and Unseen Domain, on OfficeHome ( $\varepsilon = 8/255$ ). In each cell, the first number is the standard accuracy (\%), while the second number corresponds to the robust accuracy (\%) for AA. Bold numbers indicate the best performance. A $\rightarrow$ B(C) indicates that A, B, and C represent the source, (unlabeled) target, and evaluation domains, respectively. + +
SourceUnseen
MethodAr → Rw(Ar)Ar → Rw(Cl)Ar → Rw(Pr)Avg.
ARTUDA18.83 / 2.647.70 / 0.896.28 / 0.4510.94 / 1.33
RFA99.62 / 47.0242.52 / 21.5351.57 / 21.5164.57 / 30.02
SRoUDA39.39 / 17.4339.54 / 28.6650.75 / 35.8642.23 / 27.32
PL41.78 / 19.4141.97 / 31.3952.42 / 37.2445.39 / 29.34
TAROT98.35 / 64.2447.86 / 35.1958.32 / 39.5468.18 / 46.32
Cl → Rw(Cl)Cl → Rw(Ar)Cl → Rw(Pr)Avg.
ARTUDA84.81 / 55.8014.34 / 3.2127.26 / 13.9942.14 / 24.33
RFA98.35 / 80.1040.37 / 8.8654.86 / 22.2564.53 / 37.07
SRoUDA48.29 / 35.5833.87 / 15.6248.19 / 33.0743.45 / 28.09
PL48.75 / 35.9536.30 / 16.6950.80 / 35.6645.28 / 29.43
TAROT93.65 / 84.3539.72 / 17.1855.46 / 37.6762.95 / 46.40
Pr → Rw(Pr)Pr → Rw(Ar)Pr → Rw(Cl)Avg.
ARTUDA98.29 / 60.1725.67 / 4.6236.24 / 16.2453.40 / 27.01
RFA99.75 / 68.5940.67 / 7.8741.76 / 19.7360.73 / 32.06
SRoUDA61.48 / 41.4034.82 / 16.0742.08 / 30.6846.13 / 29.38
PL61.05 / 43.3436.09 / 17.1042.11 / 30.9146.42 / 30.45
TAROT96.13 / 82.4339.14 / 16.1546.87 / 33.0160.71 / 43.86
+ +# 5.2. Ablation Studies + +We conduct ablation studies about: (i) effect of Robust-PT (ii) sensitivity analysis of $\alpha$ . All exact values and additional experimental results are presented in Sec. 10. + +Effect of Robust-PT on Various $\varepsilon$ Figures 3a and 3b show that the both algorithms perform significantly better when Robust-PT is applied, particularly as $\varepsilon$ increases, indicating that Robust-PT is crucial for robustness against strong attack. Without Robust-PT, both TAROT and PL experience sharp declines in accuracy, especially as $\varepsilon$ becomes larger. Figures 3c and 3d illustrate that TAROT uniformly outperforms PL across different perturbation budgets $(\varepsilon)$ both on + +Table 4. Performances of ARTUDA, RFA, SROUDA, PL and TAROT on DomainNet $(\varepsilon = 4 / 255)$ . In each cell, the first number is the standard accuracy $(\%)$ , and the second number is the robust accuracy $(\%)$ for AA. Bold numbers indicate the best performance. + +
MethodC → IC → PC → RC → SAvg.
ARTUDA12.64 / 3.7631.15 / 7.9044.90 / 16.4939.90 / 17.5832.15 / 11.43
RFA9.42 / 3.4723.33 / 7.6037.54 / 17.7427.40 / 12.1724.42 / 10.24
SRoUDA9.27 / 5.6824.33 / 15.0742.91 / 28.3832.42 / 23.8227.23 / 18.24
PL10.59 / 6.6927.06 / 15.7343.57 / 28.6832.13 / 23.5228.34 / 18.66
TAROT14.65 / 8.1633.23 / 17.3650.02 / 30.4640.11 / 27.7734.50 / 20.94
I → CI → PI → RI → SAvg.
ARTUDA31.89 / 12.0720.38 / 2.5623.79 / 5.4734.61 / 5.9027.67 / 6.50
RFA15.00 / 8.8214.80 / 4.8019.15 / 8.7611.44 / 5.0515.10 / 6.86
SRoUDA25.95 / 14.1824.27 / 13.9336.32 / 22.0620.08 / 13.5226.65 / 15.92
PL17.45 / 13.1622.42 / 13.4132.50 / 21.9418.25 / 13.1422.65 / 15.41
TAROT29.36 / 20.2327.66 / 14.3039.14 / 23.3921.89 / 14.0829.51 / 18.00
P → CP → IP → RP → SAvg.
ARTUDA42.89 / 20.5814.36 / 3.2149.03 / 16.6234.02 / 13.1235.08 / 13.38
RFA26.27 / 14.859.45 / 3.0937.47 / 17.0419.67 / 8.5923.22 / 10.89
SRoUDA36.69 / 28.2610.17 / 5.8645.69 / 29.9530.95 / 19.1430.87 / 20.80
PL33.86 / 26.6911.18 / 7.2544.56 / 29.4729.76 / 20.6629.84 / 21.02
TAROT41.27 / 30.5114.82 / 7.6951.10 / 30.7637.22 / 24.7136.10 / 23.42
R → CR → IR → PR → SAvg.
ARTUDA52.94 / 28.1016.94 / 4.1741.95 / 12.0236.11 / 14.8636.98 / 14.79
RFA36.72 / 19.3012.36 / 3.0232.03 / 8.3422.73 / 8.1225.96 / 9.70
SRoUDA44.33 / 34.3310.10 / 5.6340.26 / 23.2629.10 / 18.5630.94 / 20.44
PL42.03 / 33.2711.85 / 7.1439.26 / 22.8029.05 / 19.6830.55 / 20.72
TAROT52.26 / 37.5818.86 / 9.4547.62 / 23.4339.38 / 25.2339.53 / 23.92
S → CS → IS → PS → RAvg.
ARTUDA54.24 / 30.5913.14 / 3.8934.87 / 7.7941.40 / 13.4735.91 / 13.94
RFA33.87 / 19.949.47 / 3.5427.29 / 7.8330.39 / 13.7025.26 / 11.25
SRoUDA46.62 / 36.658.50 / 5.2831.09 / 15.1442.95 / 22.3332.29 / 19.85
PL42.48 / 34.668.95 / 6.1831.48 / 18.7542.55 / 28.2431.37 / 21.96
TAROT53.35 / 41.3813.71 / 7.5239.77 / 20.4348.77 / 30.0238.90 / 24.84
+ +Table 5. Performances of ARTUDA, RFA, SRoUDA, PL and TAROT on VisDA2017 $(\varepsilon = 8 / 255)$ . Standard accuracy $(\%)$ / Robust accuracy $(\%)$ for AA. + +
MethodSyn. → RealSyn.
ARTUDA10.73 / 0.0061.43 / 5.70
RFA64.86 / 8.6599.41 / 33.55
SRoUDA24.44 / 0.2048.56 / 24.05
PL66.71 / 37.8340.98 / 24.09
TAROT67.12 / 37.7785.18 / 51.21
+ +![](images/f3b8f55ff8b96c4338004d1955886489d7ca9f863c8293509ef079c1dc666455.jpg) + +![](images/96b420a4ef224c89539bcbab4a0c5d3cd7490df450632d87d181ee64d13b631c.jpg) +Figure 2. Sensitivity Analysis of $\alpha$ . $\alpha = 0$ corresponds to PL. + +the standard accuracies and robust accuracies against AA. Especially, that performance gap is increasing as $\varepsilon$ increases. Figures 3a and 3b show that both algorithms perform significantly better when Robust-PT is applied, particularly as $\varepsilon$ increases, indicating that Robust-PT is crucial for robustness against strong attacks. Figures 3c and 3d illustrate that TAROT uniformly outperforms the PL approach across different perturbation budgets $(\varepsilon)$ , both in standard accuracies and robust accuracies against AA. Notably, the performance gap increases as $\varepsilon$ increases. + +Sensitivity Analysis of $\alpha$ Figure 2 illustrates the sensitivity with respect to $\alpha$ in Eq. (20). It shows how the performance of generalization and robustness changes for the target and unseen domains as the value of $\alpha$ varies. As shown in the figure, each task has a different optimal $\alpha$ , indicating that the optimal $\alpha$ value for one task may not be optimal for another. The detailed numbers are provided in Sec. 10.3. + +# 6. Conclusions and Future Works + +To our best knowledge, this is the first work for adversarially robust domain adaptation algorithms grounded with theoretical explanations. Compared to existing works, our algorithm shows essentially domain invariant robust features, resulting significant improved performance in source and + +![](images/d2807fb685b1d7e9c8a842740ad4a1ccb52e9364942b8b28014dd38e90f2b0f1.jpg) +(a) PL w/ and w/o R-PT + +![](images/aafed76dc8d0f073ede67c38a8a7cbfeed5c6cc481d8bdb7f4cb325c3594f27f.jpg) +(c) Rob Acc. w/ R-PT + +![](images/2f7c223d0d73c962e58b3e88a13ba6652ceada2cb71175daf27c0fba7d73bcfe.jpg) +(b) TAROT w/ and w/o R-PT + +![](images/6378e763a5c4c74ae86c8697fc4c490b20a2b2ae6acb9f3f2eceefbee0b679f3.jpg) +(d) Stand Acc. w/ R-PT +Figure 3. Effect of Robust-PT across various methods and $\varepsilon$ values. Each plot compares different configurations of PL w/ MDD and TAROT for standard and robust accuracy. + +unseen domains. Our work not only serves as a cornerstone for understanding robust domain adaptation from a theoretical perspective, but we also believe TAROT can be served as a state-of-the-art algorithm in practical perspective by excelling researches of adversarial robustness for transfer learning such as domain adaptation/generalization and few-shot learning task. + +Acknowledgement This work is partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-II220184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics), National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(No. 2022R1A5A7083908) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) [NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)]. + +# References + +[1] Eric Arazo, Diego Ortega, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2020. 5 +[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 6 +[3] Muhammad Awais, Fengwei Zhou, Hang Xu, Lanqing Hong, Ping Luo, Sung-Ho Bae, and Zhenguo Li. Adversarial robustness for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8568-8577, 2021. 1, 2, 6, 7 +[4] Francesco Barbato, Marco Toldo, Umberto Michieli, and Pietro Zanuttigh. Latent space regularization for unsupervised domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2835-2845, 2021. 1 +[5] Peter L Bartlett, Michael I Jordan, and Jon D McAuliffe. Convexity, classification, and risk bounds. Journal of the American Statistical Association, 101(473):138-156, 2006. 2 +[6] Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. Advances in neural information processing systems, 30, 2017. 4 +[7] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79: 151-175, 2010. 2 +[8] Mingcai Chen, Hao Cheng, Yuntao Du, Ming Xu, Wenyu Jiang, and Chongjun Wang. Two wrongs don't make a right: Combating confirmation bias in learning with label noise. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14765-14773, 2023. 5 +[9] Moustapha Cisse, Piotr Bojanowski, Edouard Grave, Yann Dauphin, and Nicolas Usunier. Parseval networks: Improving robustness to adversarial examples. In International conference on machine learning, pages 854-863. PMLR, 2017. 4 +[10] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 6 +[11] Antonio Galbis and Manuel Maestre. Vector analysis versus vector calculus. Springer Science & Business Media, 2012. 4 + +[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 5 +[13] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 2, 5 +[14] Zhiqiang Gao, Shufei Zhang, Kaizhu Huang, Qiufeng Wang, Rui Zhang, and Chaoliang Zhong. Certifying better robust generalization for unsupervised domain adaptation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2399-2410, 2022. 1 +[15] Chao Gong, Kai Chen, Zhipeng Wei, Jingjing Chen, and Yulong Jiang. Reliable and efficient concept erasure of text-to-image diffusion models. In In European Conference on Computer Vision (ECCV), 2024. 1 +[16] Fredrik D. Johansson, David Sontag, and Rajesh Ranganath. Support and invertibility in domain-invariant representations. International Conference on Artificial Intelligence and Statistics, 2019. 1 +[17] Matt Jordan and Alexandros G Dimakis. Exactly computing the local lipschitz constant of relu networks. Advances in Neural Information Processing Systems, 33:7344-7353, 2020. 4 +[18] Bo Fu Jinguang Jiang, Baixu Chen and Mingsheng Long. Transfer-learning-library. https://github.com/thuml/Transfer-Learning-Library, 2020.5 +[19] Changhoon Kim, Kyle Min, and Yezhou Yang. R.a.c.e.: Robust adversarial concept erasure for secure text-to-image diffusion model. In In European Conference on Computer Vision (ECCV), 2025. 1 +[20] Shao-Yuan Lo and Vishal M. Patel. Exploring adversarially robust training for unsupervised domain adaptation. In Asian Conference on Computer Vision (ACCV), 2022. 1, 2, 4, 6, 7 +[21] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations (ICLR), 2018. 1, 2, 3, 6 +[22] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. In In International Conference on Learning Representations (ICLR), 2023. 1 +[23] Afshin Rostamizadeh Mehryar Mohri and Ameet Talwalkar. Foundations of Machine Learning. MIT Press, 2012. 4, 1 +[24] Farzeen Munir, Shoaib Azam, and Moongu Jeon. Sstn: Self-supervised domain adaptation thermal object detection for autonomous driving, 2021. 1 +[25] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge, 2017. 6 +[26] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 6 +[27] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In In + +European Conference on Computer Vision (ECCV), pages 213-226. Springer, 2010. 6 +[28] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? In Conference on Neural Information Processing Systems (NeurIPS), 2020. 6 +[29] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. In In International Conference on Machine Learning (ICML), 2024. 1 +[30] Zhouxing Shi, Yihan Wang, Huan Zhang, J Zico Kolter, and Cho-Jui Hsieh. Efficiently computing local lipschitz constants of neural networks via bound propagation. Advances in Neural Information Processing Systems, 35:2350-2364, 2022. 4 +[31] Peeyush Singhal, Rahee Walambe, Sheela Ramanna, and Ketan Kotecha. Domain adaptation: Challenges, methods, datasets, and applications. IEEE Access, 11:6973-7020, 2023. 1 +[32] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations (ICLR), 2014. 1 +[33] Michel Talagrand. Upper and lower bounds for stochastic processes. Springer, 2014. 1 +[34] Marvin Teichmann, Michael Weber, Marius Zoellner, Roberto Cipolla, and Raquel Urtasun. Multinet: Real-time joint semantic reasoning for autonomous driving, 2018. 1 +[35] Thulasi Tholeti and Sheetal Kalyani. The robust way to stack and bag: the local lipschitz way. arXiv preprint arXiv:2206.00513, 2022. 4 +[36] Jonathan Uesato, Jean-Baptiste Alayrac, Po-Sen Huang, Robert Stanforth, Alhussein Fawzi, and Pushmeet Kohli. Are labels required for improving adversarial robustness? In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2 +[37] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018-5027, 2017. 6 +[38] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International Conference on Learning Representataions(ICLR), 2020. 1 +[39] Boxi Wu, Jinghui Chen, Deng Cai, Xiaofei He, and Quanquan Gu. Do wider neural networks really help adversarial robustness? Advances in Neural Information Processing Systems, 34:7054-7067, 2021. 4 +[40] Dongyoon Yang, Insung Kong, and Yongdai Kim. Enhancing adversarial robustness in low-label regime via adaptively weighted regularization and knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4552-4561, 2023. 1, 2 +[41] Dongyoon Yang, Insung Kong, and Yongdai Kim. Improving adversarial robustness by putting more regularizations on less robust samples. In International Conference on Machine Learning, pages 39331-39348. PMLR, 2023. 1, 2 + +[42] Yao-Yuan Yang, Cyrus Rashtchian, Hongyang Zhang, Ruslan Salakhutdinov, and Kamalika Chaudhuri. A closer look at accuracy vs. robustness. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4, 6, 1, 7 +[43] Bohang Zhang, Du Jiang, Di He, and Liwei Wang. Rethinking lipschitz neural networks and certified robustness: A boolean function perspective. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 4, 6, 1 +[44] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P Xing, Laurent El Ghaoui, and Michael I Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning (ICML), 2019. 1, 2 +[45] Yuchen Zhang, Tianle Liu, Mingsheng Long, and Michael I. Jordan. Bridging theory and algorithm for domain adaptation. In International Conference on Machine Learning (ICML), 2019. 2, 3, 4, 5, 6, 1 +[46] Wanqing Zhu, Jia-Li Yin, Bo-Hao Chen, and Xineng Liu. Srouda: meta self-training for robust unsupervised domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3852-3860, 2023. 1, 2, 3, 4, 6, 7 \ No newline at end of file diff --git a/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/images.zip b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d25badfc02ec17fa5b4c9685adf3c0678909b36d --- /dev/null +++ b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bfd953f2a249dd25f1b284d08a36a8e294032a1b98c8638deba99ca0eacabdf +size 709199 diff --git a/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/layout.json b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c8d8096b66d9c9ceb9f0272e4596143ae36ff726 --- /dev/null +++ b/2025/TAROT_ Towards Essentially Domain-Invariant Robustness with Theoretical Justification/layout.json @@ -0,0 +1,11785 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 102, + 490, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 102, + 490, + 137 + ], + "spans": [ + { + "bbox": [ + 121, + 102, + 490, + 137 + ], + "type": "text", + "content": "TAROT: Towards Essentially Domain-Invariant Robustness with Theoretical Justification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 159, + 424, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 159, + 424, + 174 + ], + "spans": [ + { + "bbox": [ + 186, + 159, + 424, + 174 + ], + "type": "text", + "content": "Dongyoon Yang" + }, + { + "bbox": [ + 186, + 159, + 424, + 174 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 186, + 159, + 424, + 174 + ], + "type": "text", + "content": " Jihu Lee" + }, + { + "bbox": [ + 186, + 159, + 424, + 174 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 186, + 159, + 424, + 174 + ], + "type": "text", + "content": " Yongdai Kim" + }, + { + "bbox": [ + 186, + 159, + 424, + 174 + ], + "type": "inline_equation", + "content": "^{2†}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 174, + 397, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 174, + 397, + 187 + ], + "spans": [ + { + "bbox": [ + 212, + 174, + 397, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 212, + 174, + 397, + 187 + ], + "type": "text", + "content": "AI Advanced Technology, SK Hynix" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 187, + 432, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 187, + 432, + 201 + ], + "spans": [ + { + "bbox": [ + 179, + 187, + 432, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 179, + 187, + 432, + 201 + ], + "type": "text", + "content": "Department of Statistics, Seoul National University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 202, + 460, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 202, + 460, + 213 + ], + "spans": [ + { + "bbox": [ + 146, + 202, + 460, + 213 + ], + "type": "text", + "content": "ydy0415@gmail.com, rieky0426@snu.ac.kr, ydkim903@snu.ac.kr" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 241, + 200, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 254 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 266, + 297, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 473 + ], + "type": "text", + "content": "Robust domain adaptation against adversarial attacks is a critical research area that aims to develop models capable of maintaining consistent performance across diverse and challenging domains. In this paper, we derive a new generalization bound for robust risk on the target domain using a novel divergence measure specifically designed for robust domain adaptation. Building upon this, we propose a new algorithm named TAROT, which is designed to enhance both domain adaptability and robustness. Through extensive experiments, TAROT not only surpasses state-of-the-art methods in accuracy and robustness but also significantly enhances domain generalization and scalability by effectively learning domain-invariant features. In particular, TAROT achieves superior performance on the challenging DomainNet dataset, demonstrating its ability to learn domain-invariant representations that generalize well across different domains, including unseen ones. These results highlight the broader applicability of our approach in real-world domain adaptation scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 481, + 135, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 135, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 135, + 494 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 501, + 297, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 297, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 297, + 650 + ], + "type": "text", + "content": "Recent advances in machine learning have highlighted the importance of developing transferable models that can maintain performance across different domains. However, collecting labeled data from all potential domains is too resource-intensive for many practical applications. To address this limitation, Unsupervised Domain Adaptation (UDA) has emerged as a type of transfer learning that leverages labeled source-domain data to improve performance on a target domain without needing labels from that domain. UDA has proven pivotal in diverse machine learning tasks, largely because distributional discrepancies between source and target domains often lead to severe performance degradation [4, 16, 24, 31, 34]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 650, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 296, + 685 + ], + "type": "text", + "content": "Meanwhile, adversarial robustness has gained significant attention as a crucial element of secure and trustworthy AI. Pioneering works demonstrated the vulnerability of neural" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 243, + 556, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 556, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 556, + 357 + ], + "type": "text", + "content": "networks to small, carefully designed perturbations called adversarial attacks [21, 32] and proposed adversarial training methods [21, 38, 40, 41, 44]. Recently, the adversarial robustness community has expanded its focus to multimodal tasks (e.g., text-to-image, image-to-text), recognizing that large-scale multi-modal models are also susceptible to adversarial manipulations in one or more input modalities [15, 19, 22, 29]. These cross-modal attacks can exploit alignment mechanisms to trick models into misclassification or erroneous content generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 363, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 556, + 548 + ], + "type": "text", + "content": "Given these two developments—domain adaptation and adversarial robustness—there is a growing interest in learning transferable models in UDA settings that remain resilient to adversarial perturbations in the target domain [3, 14, 20, 46]. A common strategy in existing robust UDA methods is to rely on pseudo-labels to handle unlabeled target data. However, in this paper, we propose a novel robust margin disparity metric for comparing source and target distributions, enabling us to derive a refined generalization bound on the robust risk in the target domain. From our bound, we derive two key insights: (1) using only pseudo-labels may be insufficient, but combining pseudo-labels with distributional alignment significantly enhances robust UDA performance; and (2) initializing robust UDA algorithms with a robustly pre-trained model (Robust-PT [3]) is notably more effective than starting from a non-robust pre-trained model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 552, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 557, + 713 + ], + "type": "text", + "content": "Building upon these theoretical foundations, we propose TAROT (Transfer Adversarially RObust Training), a new algorithm that leverages our theoretical insights to achieve superior performance in robust UDA tasks. Through extensive experiments on various benchmark datasets, we demonstrate that TAROT outperforms existing algorithms. Furthermore, we empirically demonstrate that TAROT exhibits distinct advantages in terms of scalability compared to pseudo-labeling approaches, particularly with significant performance improvements not only on the target domain but also on source and unseen source domains. This work bridges the gap between theoretical understanding and practical implementation in robust transfer learning, offering both fundamental insights and actionable solutions for real-world applications." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 147, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 147, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 147, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25780" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 253, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 253, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 253, + 83 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 295, + 245 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 85, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 294, + 107 + ], + "type": "text", + "content": "- We introduce a novel robust margin disparity measure and derive a refined generalization bound for robust UDA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 108, + 295, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 152 + ], + "type": "text", + "content": "- From the derived generalization bound, we propose TAROT (Transferring Adversarial RObust Training), which integrates pseudo-labeling and distributional alignment to enhance adversarial robustness in UDA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 153, + 295, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 295, + 211 + ], + "type": "text", + "content": "- TAROT outperforms existing robust UDA methods across multiple benchmarks, demonstrating strong performance on the target, source, and unseen source domains, indicating that TAROT effectively facilitates domain-invariant robust representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 211, + 295, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 245 + ], + "type": "text", + "content": "- We provide theoretical insights and practical solutions for scalable and adversarially robust domain adaptation in real-world scenarios." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 258, + 140, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 140, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 140, + 271 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 279, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 384 + ], + "type": "text", + "content": "In this section, we outline the core concepts and notations for our analysis, beginning with adversarial training and its risk formulation for robust learning. We then describe the unsupervised domain adaptation (UDA) framework, highlighting distributional notations, the formal UDA objective, and existing theoretical approaches like DANN [13] and MDD [45]. These elements provide the groundwork for our subsequent derivation of a robust UDA upper bound and the proposed learning algorithm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 393, + 178, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 178, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 178, + 406 + ], + "type": "text", + "content": "2.1. Adversarial Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": " represent the input space, and let " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{1, \\ldots, C\\}" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": " be the label space, where " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": " represents the number of classes. A classifier can be represented using a score function " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "f_c" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": "-th component of " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": ". The classifier is then defined as: " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "h_f(\\pmb{x}) = \\operatorname*{argmax}_{c \\in \\mathcal{Y}} f_c(\\pmb{x})" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_p(\\pmb{x}, \\varepsilon) = \\{\\pmb{x}' \\in \\mathcal{X}: \\| \\pmb{x} - \\pmb{x}' \\|_p \\leq \\varepsilon\\}" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": " denote an " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": "-ball centered at " + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 411, + 296, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "content": "The population risk considering adversarial robustness is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 533, + 296, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 533, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 77, + 533, + 296, + 553 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\mathrm {r o b}} (f) = \\mathbb {E} _ {(\\mathbf {X}, Y)} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\mathbb {1} \\left\\{Y \\neq h _ {f} \\left(\\mathbf {X} ^ {\\prime}\\right) \\right\\}, \\tag {1}", + "image_path": "eb3e21d903259d3e51aaba47e97af43c2061ac4dc1cba5a3545911c2c8c4b791.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": " denotes the indicator function. The goal of adversarial training is to estimate the optimal " + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": ", the minimizer of Eq. (1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 608, + 237, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 237, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 237, + 621 + ], + "type": "text", + "content": "2.2. Unsupervised Domain Adaptation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": "There are two distributions, the source distribution " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " and the target distribution " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": ". The empirical versions of " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{S} = \\{(x_i^s,y_i^s)\\}_{i = 1}^{n_s}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{T}} = \\{(x_j^t,y_j^t)\\}_{j = 1}^{n_t}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": ", respectively. In unsupervised domain adaptation, " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{S}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " and only " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{T}}_{\\mathbf{X}} := \\{x_j^t\\}_{j = 1}^{n_t}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " are observed. i.e. the labels " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_{j}^{t}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " can not be observed. Denote the distributions of the input on the source and target domains by " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 626, + 296, + 714 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "For a given score function " + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": ", the risk of a data distribution " + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 355, + 103, + 555, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 103, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 355, + 103, + 555, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\mathcal {D}} (f) := \\mathbb {E} _ {(\\mathbf {X}, Y) \\sim \\mathcal {D}} \\mathbb {1} \\{Y \\neq h _ {f} (\\mathbf {X}) \\}. \\tag {2}", + "image_path": "0ec67368881bd557c2a25e11ed200cde8ca30134534fb2e263fc0cd490fd66fa.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "text", + "content": "The goal of UDA is to minimize " + }, + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathcal{T}}(f)" + }, + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "inline_equation", + "content": "\\widehat{S}" + }, + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{T}}_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "text", + "content": " which can be done by building an upper bound of " + }, + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathcal{T}}(h)" + }, + { + "bbox": [ + 313, + 125, + 555, + 150 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "content": "The pivotal learning algorithms for UDA are DANN (Domain-Adversarial Neural Network) [13] and MDD (Margin Disparity Discrepancy) [45]. Ben-David et al. [7] suggested a distance " + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{H}\\Delta \\mathcal{H}}(\\mathcal{S}_{\\mathbf{X}},\\mathcal{T}_{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "content": " to give an upper bound of the robust risk, and Ganin et al. [13] utilized this upper bound to implement their learning algorithm DANN. DANN uses an adversarial neural network to minimize the distance between the source and target distributions, using gradient reversal layers (GRL). On the other hand, MDD [45] derives an upper bound of " + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathcal{T}}(f)" + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "content": " as the sum of " + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_S^{(\\rho)}(f)" + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "content": ", a standard non-robust risk on the source domain, and " + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "inline_equation", + "content": "d_{f,\\mathcal{F}}^{(\\rho)}(\\mathcal{S}_{\\mathbf{X}},\\mathcal{T}_{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "content": ", a distance between " + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 150, + 555, + 350 + ], + "type": "text", + "content": ", and develop the MDD algorithm based on this upper bound. The main distinction between the two upper bounds is that MDD utilizes a scoring function for multi-class classification, while DANN only employs a classifier for binary classification." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 350, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 419 + ], + "type": "text", + "content": "The objective of our work is to derive an upper bound of the robust risk for the target domain by utilizing the risk in the source domain and the robust divergence between the target domain " + }, + { + "bbox": [ + 313, + 350, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 350, + 555, + 419 + ], + "type": "text", + "content": " and the source domain " + }, + { + "bbox": [ + 313, + 350, + 555, + 419 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 350, + 555, + 419 + ], + "type": "text", + "content": ". In turn, inspired by the upper bound, we propose a novel robust UDA algorithm." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 429, + 406, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 429, + 406, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 429, + 406, + 441 + ], + "type": "text", + "content": "3. Related Works" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 449, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 529 + ], + "type": "text", + "content": "While a significant body of research on adversarial training has focused on supervised learning and semi-supervised learning tasks [21, 36, 40, 41, 44], there has been relatively less attention given to its application in domain adaptation tasks. Only a few studies have explored adversarial training in the context of domain adaptation [3, 20, 46], highlighting the need for further investigation in this area." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 529, + 555, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 621 + ], + "type": "text", + "content": "RFA [3] leverages a pretrained ImageNet model to develop a regularization term. When the trainable model is initialized with a pretrained robust model, the regularization term encourages the outputs of each layer of the trainable model to resemble those of the pretrained model. This facilitates the learning of robust features. This approach can be applied to existing UDA algorithms such as DANN [13] and MDD [45] to make them adversarially robust." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "type": "text", + "content": "ARTUDA [20] adopts a self-supervised adversarial training approach, where the predictions of the current model are used as pseudo-labels to generate adversarial examples for the target domain. It also generates adversarial examples for the source domain and applies the distance metrics used in DANN. However, there is no theoretical justification of the complex loss function used in ARTUDA. One key challenge in ARTUDA is that pseudo-labels are generated by the same" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25781" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 130 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 130 + ], + "type": "text", + "content": "model being trained. As a result, due to the inherent trade-off between robustness and generalization, as the model learns more robust features, pseudo-labels are getting more inaccurate. This issue becomes particularly pronounced when the perturbation budget " + }, + { + "bbox": [ + 55, + 72, + 296, + 130 + ], + "type": "inline_equation", + "content": "(\\varepsilon)" + }, + { + "bbox": [ + 55, + 72, + 296, + 130 + ], + "type": "text", + "content": " is large." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 130, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 130, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 296, + 278 + ], + "type": "text", + "content": "SRoUDA [46] attempts to resolve this limitation by introducing a separate teacher model to assign pseudo-labels to the target data. By leveraging meta-learning, SRoUDA improves the accuracy of pseudo-labeling by the teacher model and dynamically performs adversarial training [21] on the target domain using these pseudo-labels. Despite the improvement, SRoUDA struggles to effectively capture domain-invariant features, as it does not explicitly consider the deviation of the distributions between the source and target domains. This leads to a decline in performance in the source as well as in the unseen domains, indicating a failure to capture what we call essentially domain-invariant robust features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 290, + 162, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 162, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 162, + 304 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "text", + "content": "In this section, we present the theoretical results for adversarially robust UDA. This is an extension of MDD [45] to its robust version, and the proofs are deferred to Sec. 8." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 353, + 256, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 256, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 256, + 366 + ], + "type": "text", + "content": "4.1. Robust Margin Disparity Discrepancy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 371, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 393 + ], + "type": "text", + "content": "In this subsection, we introduce a novel disparity measure and key propositions to establish the generalization bound." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 393, + 290, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 393, + 290, + 406 + ], + "spans": [ + { + "bbox": [ + 67, + 393, + 290, + 406 + ], + "type": "text", + "content": "Denote the classifier " + }, + { + "bbox": [ + 67, + 393, + 290, + 406 + ], + "type": "inline_equation", + "content": "h_f" + }, + { + "bbox": [ + 67, + 393, + 290, + 406 + ], + "type": "text", + "content": " for a given score function " + }, + { + "bbox": [ + 67, + 393, + 290, + 406 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 67, + 393, + 290, + 406 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 415, + 227, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 415, + 227, + 435 + ], + "spans": [ + { + "bbox": [ + 123, + 415, + 227, + 435 + ], + "type": "interline_equation", + "content": "h_{f}:\\boldsymbol {x}\\mapsto \\operatorname *{argmax}_{y\\in \\mathcal{Y}}f_{y}(\\boldsymbol {x}),", + "image_path": "493ff12f85651d351331c5478c9e1bb6027788cef067caef56ec03149f26c98a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 445, + 284, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 284, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 284, + 458 + ], + "type": "text", + "content": "and denote the margin of " + }, + { + "bbox": [ + 55, + 445, + 284, + 458 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 445, + 284, + 458 + ], + "type": "text", + "content": " at a labeled sample " + }, + { + "bbox": [ + 55, + 445, + 284, + 458 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 445, + 284, + 458 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 466, + 246, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 466, + 246, + 485 + ], + "spans": [ + { + "bbox": [ + 104, + 466, + 246, + 485 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {f} (\\boldsymbol {x}, y) = f _ {y} (\\boldsymbol {x}) - \\max _ {y ^ {\\prime} \\neq y} f _ {y ^ {\\prime}} (\\boldsymbol {x}).", + "image_path": "3c847f449f36b99bb114b3467cd9ff68a1530fcc5540e045a4936f7a45867297.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": "Definition 1. (Robust Disparity). For given " + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "inline_equation", + "content": "f, f' \\in \\mathcal{F}" + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": ", the 0-1 robust disparity between " + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": " on a distribution " + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 509, + 296, + 544 + ], + "type": "text", + "content": " is defined by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 552, + 296, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 583 + ], + "type": "interline_equation", + "content": "\\operatorname {d i s p} _ {\\mathcal {D} _ {\\mathbf {X}}} ^ {r o b} \\left(f ^ {\\prime}, f\\right) := \\mathbb {E} _ {\\mathbf {X} \\sim \\mathcal {D} _ {\\mathbf {X}}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\mathbb {1} \\left\\{h _ {f ^ {\\prime}} (\\mathbf {X}) \\neq h _ {f} \\left(\\mathbf {X} ^ {\\prime}\\right) \\right\\} \\tag {3}", + "image_path": "d1096eb7f3c6a7815d1c0dc39f17d4055fe446922f1f266458e3d8ef8c6a1108.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 592, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 649 + ], + "type": "text", + "content": "The robust disparity is measured based on the adversarial robustness of the second model " + }, + { + "bbox": [ + 55, + 592, + 295, + 649 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 592, + 295, + 649 + ], + "type": "text", + "content": ", while the first model " + }, + { + "bbox": [ + 55, + 592, + 295, + 649 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 55, + 592, + 295, + 649 + ], + "type": "text", + "content": " serves as a pseudo-label generator. Note that if the prediction of " + }, + { + "bbox": [ + 55, + 592, + 295, + 649 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 55, + 592, + 295, + 649 + ], + "type": "text", + "content": " is consistent to true label, it is the robust risk as defined in Eq. (1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 650, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 296, + 685 + ], + "type": "text", + "content": "As an extension of the margin loss in the non-robust setting [45] to its robust counterpart, we define the robust margin loss as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 60, + 693, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 693, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 60, + 693, + 296, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\mathcal {D}} ^ {\\mathrm {r o b}, (\\rho)} (f) := \\mathbb {E} _ {(\\mathbf {X}, Y) \\sim \\mathcal {D}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\Phi_ {\\rho} \\circ \\mathcal {M} _ {f} \\left(\\mathbf {X} ^ {\\prime}, Y\\right), \\tag {4}", + "image_path": "2b6ecbcd972a700fa30a9b70485daf1f47ba2a03d223d63dce747e3d2a784dc9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 73, + 342, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 342, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 342, + 83 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 364, + 81, + 555, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 81, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 364, + 81, + 555, + 125 + ], + "type": "interline_equation", + "content": "\\Phi_ {\\rho} (x) := \\left\\{ \\begin{array}{l l} 0 & \\rho \\leq x \\\\ 1 - x / \\rho & 0 \\leq x \\leq \\rho \\\\ 1 & x \\leq 0 \\end{array} \\right. \\tag {5}", + "image_path": "a0a1aef187cf3e89ce01204ded8871ad985aa77b5df1f09c602797e74e800ca1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 129, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 480, + 140 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 129, + 480, + 140 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 313, + 129, + 480, + 140 + ], + "type": "text", + "content": " is composite operator of functions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 148, + 555, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 172 + ], + "type": "text", + "content": "Definition 2. (Robust Margin Disparity). For given " + }, + { + "bbox": [ + 313, + 148, + 555, + 172 + ], + "type": "inline_equation", + "content": "f, f' \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 148, + 555, + 172 + ], + "type": "text", + "content": ", the Robust Margin Disparity is defined as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 342, + 180, + 555, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 180, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 342, + 180, + 555, + 217 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {d i s p} _ {\\mathcal {D} _ {\\mathbf {X}}} ^ {r o b, (\\rho)} \\left(f ^ {\\prime}, f\\right) \\\\ := \\mathbb {E} _ {\\mathbf {X} \\sim \\mathcal {D} _ {\\mathbf {X}}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {\\rho} (\\mathbf {X}, \\varepsilon)} \\Phi_ {\\rho} \\circ \\mathcal {M} _ {f} \\left(\\mathbf {X} ^ {\\prime}, h _ {f ^ {\\prime}} (\\mathbf {X})\\right) \\tag {6} \\\\ \\end{array}", + "image_path": "5ad4fb98f7983df58dc9da6e1fb4aa5ce89407628cfb6bd82b7967ee4b762582.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 227, + 555, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 265 + ], + "type": "text", + "content": "An important property of the robust margin disparity is that it is greater than the robust disparity, i.e. " + }, + { + "bbox": [ + 313, + 227, + 555, + 265 + ], + "type": "inline_equation", + "content": "\\mathrm{disp}_{\\mathcal{D}_{\\mathbf{X}}}^{\\mathrm{rob},(\\rho)}(f', f) \\geq \\mathrm{disp}_{\\mathcal{D}_{\\mathbf{X}}}^{\\mathrm{rob}}(f', f)" + }, + { + "bbox": [ + 313, + 227, + 555, + 265 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 264, + 553, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 286 + ], + "type": "text", + "content": "We now use this disparity term to bound the robust risk on the target distribution with additional feasible terms." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "text", + "content": "Proposition 1. Let " + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "text", + "content": " represent the distributions of the source and target domains, respectively. Similarly, let " + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "inline_equation", + "content": "S_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "text", + "content": " denote the marginal distributions of the source and target domains over " + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "text", + "content": ", respectively. For every score function " + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 294, + 555, + 353 + ], + "type": "text", + "content": ", the following inequality holds:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 323, + 360, + 554, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 360, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 323, + 360, + 554, + 407 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {R} _ {T} ^ {r o b} (f) \\leq \\\\ \\mathcal {R} _ {\\mathcal {S}} ^ {(\\rho)} (f) + \\left\\{\\operatorname {d i s p} _ {\\mathcal {T} _ {\\mathbf {X}}} ^ {r o b, (\\rho)} \\left(f ^ {*}, f\\right) - \\operatorname {d i s p} _ {\\mathcal {S} _ {\\mathbf {X}}} ^ {(\\rho)} \\left(f ^ {*}, f\\right) \\right\\} + \\lambda , \\tag {7} \\\\ \\end{array}", + "image_path": "c087e4b1adab5e91fedd368d0fc9feebced619776fe6a8e4ff8f10a97f2313e0.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 407, + 555, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 555, + 454 + ], + "type": "inline_equation", + "content": "f^{*} = \\operatorname *{argmin}_{f\\in \\mathcal{F}}\\{\\mathcal{R}_{\\mathcal{T}}^{(\\rho)}(f) + \\mathcal{R}_{\\mathcal{S}}^{(\\rho)}(f)\\}" + }, + { + "bbox": [ + 313, + 407, + 555, + 454 + ], + "type": "text", + "content": " is ideal hypothesis and " + }, + { + "bbox": [ + 313, + 407, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\lambda = \\lambda (\\mathcal{F},\\mathcal{S},\\mathcal{T},\\varepsilon ,\\rho) = \\mathcal{R}_{\\mathcal{T}}^{(\\rho)}(f^{*}) + \\mathcal{R}_{\\mathcal{S}}^{(\\rho)}(f^{*})" + }, + { + "bbox": [ + 313, + 407, + 555, + 454 + ], + "type": "text", + "content": " is constant of " + }, + { + "bbox": [ + 313, + 407, + 555, + 454 + ], + "type": "inline_equation", + "content": "f" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 461, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 529 + ], + "type": "text", + "content": "Note that the upper bound consists of standard margin risk of the source domain, not the robust margin risk. This enables the corresponding algorithm introduced in Sec. 4.4 computationally simpler because generation of adversarial samples is not required to calculate the standard margin risk on the source domain." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 530, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 577 + ], + "type": "text", + "content": "From the above proposition, we can naturally induce the following discrepancy between the distributions of two domains, which we call as a robust margin disparity discrepancy." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 584, + 547, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 547, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 547, + 595 + ], + "type": "text", + "content": "Definition 3. (Robust Margin Disparity Discrepancy)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "text", + "content": "Given a score function " + }, + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "text", + "content": ", the Robust Margin Disparity Discrepancy induced by " + }, + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "inline_equation", + "content": "f' \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "text", + "content": " of two distributions " + }, + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "inline_equation", + "content": "S_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 595, + 555, + 631 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 340, + 639, + 555, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 639, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 340, + 639, + 555, + 681 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d _ {f, \\mathcal {F}} ^ {r o b, (\\rho)} (S _ {\\mathbf {X}}, \\mathcal {T} _ {\\mathbf {X}}) \\\\ := \\sup _ {f ^ {\\prime} \\in \\mathcal {F}} \\left\\{\\operatorname {d i s p} _ {\\mathcal {T} _ {\\mathbf {x}}} ^ {r o b, (\\rho)} \\left(f ^ {\\prime}, f\\right) - \\operatorname {d i s p} _ {\\mathcal {S} _ {\\mathbf {x}}} ^ {(\\rho)} \\left(f ^ {\\prime}, f\\right) \\right\\} \\tag {8} \\\\ \\end{array}", + "image_path": "c955c32b94d31bc6cc0cfbc77e4d26fb82b222a420fc069c0d77e118b22eeeb0.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "This discrepancy measures the difference of the robust margin disparity on target domain and the standard margin" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25782" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 129 + ], + "type": "text", + "content": "disparity on the source domain, where the objective " + }, + { + "bbox": [ + 55, + 72, + 296, + 129 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 72, + 296, + 129 + ], + "type": "text", + "content": " is fixed. If there is significant gap between the adversarial target domain and standard source domain, the discrepancy becomes large. Now we introduce a non-asymptotic upper bound for the target robust risk:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 136, + 296, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 136, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 80, + 136, + 296, + 152 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\mathcal {T}} ^ {\\mathrm {r o b}, (\\rho)} (f) \\leq \\mathcal {R} _ {\\mathcal {S}} ^ {(\\rho)} (f) + d _ {f, \\mathcal {F}} ^ {\\mathrm {r o b}, (\\rho)} \\left(\\mathcal {S} _ {\\mathbf {X}}, \\mathcal {T} _ {\\mathbf {X}}\\right) + \\lambda \\tag {9}", + "image_path": "ee0fcb8acadef68e8275f97480b84ab8a1b0b771614e330dfb92654102ac0d8e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 157, + 191, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 191, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 191, + 168 + ], + "type": "text", + "content": "with the same " + }, + { + "bbox": [ + 55, + 157, + 191, + 168 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 157, + 191, + 168 + ], + "type": "text", + "content": " in Proposition 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 175, + 180, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 180, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 180, + 186 + ], + "type": "text", + "content": "4.2. Generalization bound" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 191, + 296, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 227 + ], + "type": "text", + "content": "In this subsection, we provide a generalization bound based on Eq. (9). We use the notion of local Lipschitz constant of a function, following Yang et al. [42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 232, + 296, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 267 + ], + "type": "text", + "content": "Definition 4 (Local Lipschitz Constant). Consider an arbitrary real function " + }, + { + "bbox": [ + 55, + 232, + 296, + 267 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^d \\to \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 232, + 296, + 267 + ], + "type": "text", + "content": ", and a " + }, + { + "bbox": [ + 55, + 232, + 296, + 267 + ], + "type": "inline_equation", + "content": "L_p" + }, + { + "bbox": [ + 55, + 232, + 296, + 267 + ], + "type": "text", + "content": "-norm " + }, + { + "bbox": [ + 55, + 232, + 296, + 267 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_p" + }, + { + "bbox": [ + 55, + 232, + 296, + 267 + ], + "type": "text", + "content": ". Define a point-wise local Lipschitz constant as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 271, + 296, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 271, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 91, + 271, + 296, + 300 + ], + "type": "interline_equation", + "content": "L _ {f} (\\boldsymbol {x}, \\varepsilon) := \\sup _ {\\boldsymbol {x} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\boldsymbol {x}, \\varepsilon)} \\frac {\\| f \\left(\\boldsymbol {x} ^ {\\prime}\\right) - f (\\boldsymbol {x}) \\| _ {1}}{\\| \\boldsymbol {x} ^ {\\prime} - \\boldsymbol {x} \\| _ {p}} \\tag {10}", + "image_path": "a9ebd84925ff378f2c5d2e0b3fb97e1e77220ac0c26ee99da0848e2186d8f34b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 305, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 295, + 327 + ], + "type": "text", + "content": "for a point " + }, + { + "bbox": [ + 54, + 305, + 295, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 54, + 305, + 295, + 327 + ], + "type": "text", + "content": ". Then, we define a local Lipschitz constant on " + }, + { + "bbox": [ + 54, + 305, + 295, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{X}}" + }, + { + "bbox": [ + 54, + 305, + 295, + 327 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 328, + 295, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 328, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 115, + 328, + 295, + 347 + ], + "type": "interline_equation", + "content": "L _ {f} \\left(\\mathcal {D} _ {\\mathbf {X}}, \\varepsilon\\right) := \\sup _ {x \\in \\mathcal {D} _ {\\mathbf {X}}} L _ {f} (x, \\varepsilon) \\tag {11}", + "image_path": "d8fe47245f5b40059ed9a491a1db7fd16fcf40839a01ed32a02fea5a53aec4ea.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 351, + 296, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 375 + ], + "type": "text", + "content": "We call that " + }, + { + "bbox": [ + 55, + 351, + 296, + 375 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 351, + 296, + 375 + ], + "type": "text", + "content": " is locally Lipschitz, if " + }, + { + "bbox": [ + 55, + 351, + 296, + 375 + ], + "type": "inline_equation", + "content": "L_{f}(\\mathcal{D}_{\\mathbf{X}},\\varepsilon)" + }, + { + "bbox": [ + 55, + 351, + 296, + 375 + ], + "type": "text", + "content": " is bounded (i.e. " + }, + { + "bbox": [ + 55, + 351, + 296, + 375 + ], + "type": "inline_equation", + "content": "L_{f}(\\mathcal{D}_{\\mathbf{X}},\\varepsilon) < \\infty" + }, + { + "bbox": [ + 55, + 351, + 296, + 375 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 380, + 274, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 380, + 274, + 392 + ], + "spans": [ + { + "bbox": [ + 67, + 380, + 274, + 392 + ], + "type": "text", + "content": "We also define following auxiliary function spaces." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "text", + "content": "Definition 5. Consider a set of score functions " + }, + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "text", + "content": " be the class of the induced classifiers " + }, + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "inline_equation", + "content": "h_f" + }, + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "inline_equation", + "content": "f\\in \\mathcal{F}" + }, + { + "bbox": [ + 55, + 396, + 296, + 431 + ], + "type": "text", + "content": ". We define the following two sets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 437, + 295, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 437, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 86, + 437, + 295, + 450 + ], + "type": "interline_equation", + "content": "\\Pi_ {\\mathcal {H}} \\mathcal {F} := \\{x \\mapsto f (x, h (x)) | h \\in \\mathcal {H}, f \\in \\mathcal {F} \\} \\tag {12}", + "image_path": "c0d760738c3903191c8b49f89d8bfa1bd870c83cfa90732576435a3d5a5637f5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 91, + 452, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 452, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 91, + 452, + 295, + 464 + ], + "type": "interline_equation", + "content": "\\Pi_ {1} \\mathcal {F} := \\{x \\mapsto f (x, y) | y \\in \\mathcal {Y}, f \\in \\mathcal {F} \\} \\tag {13}", + "image_path": "2c90a21b383163a7bd5c57f13058e709ec876d2948cd041c83b510aa2d4444c9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "text", + "content": "It is known that " + }, + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "inline_equation", + "content": "\\Pi_{\\mathcal{H}}\\mathcal{F}" + }, + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "text", + "content": " takes a form of the space of inner products of vector fields from " + }, + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "text", + "content": " [11, 45]. The notion " + }, + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "inline_equation", + "content": "\\Pi_1\\mathcal{F}" + }, + { + "bbox": [ + 55, + 471, + 295, + 505 + ], + "type": "text", + "content": " follows from Mehryar Mohri and Talwalkar [23]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 506, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 551 + ], + "type": "text", + "content": "We now derive a generalization bound for the robust risk on the target domain using Rademacher complexity, which quantifies the expressiveness of a given hypothesis space [23]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "content": "Definition 6. (Rademacher Complexity). Let " + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "content": " be a family of functions mapping from " + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\mathcal{X} \\times \\mathcal{Y}" + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "inline_equation", + "content": "[a, b]" + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{D}} = \\{z_1, \\dots, z_n\\}" + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "content": " are finite samples independently drawn from a distribution " + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "content": ". Then, the empirical Rademacher complexity of " + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 556, + 296, + 617 + ], + "type": "text", + "content": " is defined by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 623, + 296, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 623, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 104, + 623, + 296, + 654 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathfrak {R}} _ {\\widehat {\\mathcal {D}}} (\\mathcal {F}) \\triangleq \\mathbb {E} _ {\\sigma} \\sup _ {f \\in \\mathcal {F}} \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\sigma_ {i} f (z _ {i}), \\tag {14}", + "image_path": "3dc84408160bbc1cacb9f28754804213c97383405002563ab329b85086babacc.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "content": "s are independent random variables following the Rademacher distribution. In turn, the Rademacher complexity of " + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "content": " is defined by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 116, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "\\mathfrak {R} _ {n, \\mathcal {D}} (\\mathcal {F}) \\triangleq \\mathbb {E} _ {\\widehat {\\mathcal {D}} \\sim \\mathcal {D} ^ {n}} \\widehat {\\mathfrak {R}} _ {\\widehat {\\mathcal {D}}} (\\mathcal {F}) \\tag {15}", + "image_path": "297e896266693dda2133ef2616c91c73caa9fcebcf363fc18ca45d6db62087d4.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 72, + 555, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 118 + ], + "type": "text", + "content": "Now we provide our generalization bound, which replaces the terms in the upper bound (Eq. (9)) that depend on the population distribution by the terms depending on the empirical distributions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "text", + "content": "Theorem 1. (Generalization Bound on the Robust Risk of Target Distribution). For any " + }, + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "inline_equation", + "content": "\\delta > 0" + }, + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "text", + "content": ", with probability " + }, + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "inline_equation", + "content": "1 - 3\\delta" + }, + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "text", + "content": ", we have the following uniform generalization bound for any score function " + }, + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 126, + 555, + 174 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 322, + 182, + 554, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 182, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 322, + 182, + 554, + 284 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {R} _ {\\mathcal {T}} ^ {r o b} (f) \\\\ \\leq \\mathcal {R} _ {\\widehat {\\mathcal {S}}} ^ {(\\rho)} (f) + d _ {f, \\mathcal {F}} ^ {r o b, (\\rho)} \\left(\\widehat {\\mathcal {S}} _ {\\mathbf {X}}, \\widehat {\\mathcal {T}} _ {\\mathbf {X}}\\right) + \\lambda \\\\ + \\frac {2 C ^ {2}}{\\rho} \\mathfrak {R} _ {n, S} (\\Pi_ {1} \\mathcal {F}) + \\frac {2 C}{\\rho} \\mathfrak {R} _ {n, S} (\\Pi_ {\\mathcal {H}} \\mathcal {F}) + 2 \\sqrt {\\frac {\\log 2 / \\delta}{2 n}} \\\\ + \\frac {2 C}{\\rho} \\Re_ {m, \\mathcal {T}} \\left(\\Pi_ {\\mathcal {H}} \\mathcal {F}\\right) + \\sqrt {\\frac {\\log 2 / \\delta}{2 m}} + \\frac {2 \\varepsilon L _ {f} \\left(\\mathcal {T} _ {\\mathbf {X}}, \\varepsilon\\right)}{\\rho}, \\tag {16} \\\\ \\end{array}", + "image_path": "b9bef114e33fbfbbfa059ae424f31efda62326d9ea4f93ef311fd956167e40b7.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 284, + 470, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 284, + 470, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 284, + 470, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 284, + 470, + 304 + ], + "type": "inline_equation", + "content": "\\lambda = \\min_{f\\in \\mathcal{F}}\\{\\mathcal{R}_{\\mathcal{T}}^{(\\rho)}(f) + \\mathcal{R}_{\\mathcal{S}}^{(\\rho)}(f)\\}" + }, + { + "bbox": [ + 314, + 284, + 470, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 311, + 397, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 311, + 397, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 311, + 397, + 323 + ], + "type": "text", + "content": "4.3. Two insights" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 327, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 431 + ], + "type": "text", + "content": "In this subsection, we present two key insights for robust UDA. First, we analyze the impact of the local Lipschitz constant on generalization performance, showing why initializing robust UDA algorithms with a robustly pre-trained (Robust-PT) model can substantially improve outcomes. Second, we highlight why pseudo labeling alone may be insufficient, especially in large or complex target domains where confirmation bias and the lack of explicit domain alignment can lead to persistent misclassifications." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 445, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 548 + ], + "type": "text", + "content": "About Local Lipschitz constant In Theorem 1, there is the local Lipschitz constant " + }, + { + "bbox": [ + 313, + 445, + 555, + 548 + ], + "type": "inline_equation", + "content": "L_{f}(\\mathcal{T}_{\\mathbf{X}},\\varepsilon)" + }, + { + "bbox": [ + 313, + 445, + 555, + 548 + ], + "type": "text", + "content": ". As discovered in previous literature, the local Lipschitz constant is known to be deeply involved with a generalization performance of a model [6, 17, 30]. This fact leads to various researches that investigate a method to insure provable robustness of a model [9, 35, 39]. It is also known that adversarial training empirically reduces the local Lipschitz constant of a neural network [42, 43]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 549, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 556, + 664 + ], + "type": "text", + "content": "In the numerical studies, we have found that robust UDA approaches with a Robust-PT as an initial improve the prediction performance much compared to those without a Robust-PT as an initial. This phenomenon can be partly explained by the local Lipschitz term. That is, model which is not initialized with Robust-PT could have a large value of the local Lipschitz constant and thus could have a too large robust risk value on the target domain. Hence, using an initial solution with a small local Lipschitz constant is promising and a Robust-PT would be such one." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 678, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 556, + 714 + ], + "type": "text", + "content": "Pseudo labeling only may not be sufficient Pseudo labeling is a widely used technique in domain adaptation [20, 46], where the objective is to improve the performance of the" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25783" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 130 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 130 + ], + "type": "text", + "content": "model on the target domain by iteratively using the predictive labels of the currently trained model (or its variation) as labels for the target domain data when generating adversarial examples for the target data. There are at least two problems in those pseudo labeling approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 130, + 295, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 130, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 54, + 130, + 295, + 256 + ], + "type": "text", + "content": "The first problem is a risk of confirmation bias. Since the model generates pseudo labels based on its own predictions, early-stage errors in the model could become self-reinforcing. When incorrect pseudo labels are assigned with high confidence, these mistakes persist through training cycles, amplifying errors and leading to poor generalization on the target domain. This phenomenon is widely studied in semi-supervised settings [1, 8]. Similarly in the UDA setting, as the training phase progresses, the model may increasingly rely on those incorrect predictions, making it difficult to recover especially in complex or noisy target domains." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 256, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 302 + ], + "type": "text", + "content": "The risk of confirmation bias can be reduced by generating pseudo labels from a pretrained non-robust UDA model called a (fixed) teacher model. In this case, we can write the objective function as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 310, + 296, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 310, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 89, + 310, + 296, + 330 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathbf {X} \\sim \\mathcal {T} _ {\\mathbf {X}}} \\max _ {\\mathbf {X} ^ {\\prime} \\in \\mathcal {B} _ {p} (\\mathbf {X}, \\varepsilon)} \\mathbb {1} \\left\\{h _ {f} \\left(\\mathbf {X} ^ {\\prime}\\right) \\neq h _ {f *} (\\mathbf {X}) \\right\\} \\tag {17}", + "image_path": "e56c74dbd0e2aac1956548f342eba3838bfc871c8d45b1f65c115ec99a143f7c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 382 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 338, + 296, + 382 + ], + "type": "inline_equation", + "content": "f^{*} = \\operatorname*{argmin}_{f\\in \\mathcal{F}}\\mathcal{R}_{\\widehat{\\mathcal{T}}}(\\boldsymbol {f})" + }, + { + "bbox": [ + 55, + 338, + 296, + 382 + ], + "type": "text", + "content": ". In our notation, this term exactly coincides with " + }, + { + "bbox": [ + 55, + 338, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\mathrm{disp}_{T_{\\mathbf{x}}}^{\\mathrm{rob}}(f^{*},f)" + }, + { + "bbox": [ + 55, + 338, + 296, + 382 + ], + "type": "text", + "content": ". The following proposition justifies this approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 389, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 413 + ], + "type": "text", + "content": "Proposition 2. For every score function " + }, + { + "bbox": [ + 55, + 389, + 296, + 413 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{F}" + }, + { + "bbox": [ + 55, + 389, + 296, + 413 + ], + "type": "text", + "content": ", the following inequality holds:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 419, + 295, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 419, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 99, + 419, + 295, + 435 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\mathcal {T}} ^ {r o b} (f) \\leq \\operatorname {d i s p} _ {\\mathcal {T} _ {\\mathbf {X}}} ^ {r o b} \\left(f ^ {*}, f\\right) + \\mathcal {R} _ {\\mathcal {T}} \\left(f ^ {*}\\right) \\tag {18}", + "image_path": "1cf6d858689be9b08b44a1db4da807db767b8cc2034be70ab5326583ca18cf05.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 441, + 144, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 144, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 144, + 453 + ], + "type": "text", + "content": "for arbitrary " + }, + { + "bbox": [ + 53, + 441, + 144, + 453 + ], + "type": "inline_equation", + "content": "f^{*}\\in \\mathcal{F}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 460, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 296, + 518 + ], + "type": "text", + "content": "It can be easily seen through the intermediate steps of Proposition 1. By letting " + }, + { + "bbox": [ + 54, + 460, + 296, + 518 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 54, + 460, + 296, + 518 + ], + "type": "text", + "content": " as a non-robust UDA model, the pseudo labeling approach with the teacher model " + }, + { + "bbox": [ + 54, + 460, + 296, + 518 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 54, + 460, + 296, + 518 + ], + "type": "text", + "content": " can be understood to minimize the upper bound in Proposition 2. Our empirical studies amply confirm this insight." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 518, + 295, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 518, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 54, + 518, + 295, + 655 + ], + "type": "text", + "content": "Pseudo labeling approaches with a fixed teacher model, however, still has limitations. In particular, the scalability of pseudo labeling across diverse or large datasets can be challenging due to the absence of explicit domain alignment mechanisms. As the size or complexity of the target domain grows, the chance of incorrect pseudo labels increases, which could overwhelm the overall learning process. This may happen especially in scenarios where the source and target domains differ significantly. This makes it harder for the model to generalize to more extensive or varied datasets, particularly when the domain gap becomes larger due to increasing size or diversity of data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 655, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 655, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 655, + 296, + 713 + ], + "type": "text", + "content": "Note that our generalization bound Theorem 1 uses the robust margin disparity discrepancy between the source and target domains instead of the robust disparity only on the target domain in Proposition 2. That is, our generalization bound integrates pseudo labeling and domain alignment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 130 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 130 + ], + "type": "text", + "content": "together. In the following subsection, inspired by these insights, we introduce an algorithm designed to overcome the drawbacks of pseudo labeling approaches, tackling both confirmation bias and scalability while enhancing generalization abilities across diverse and complex target domains." + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 318, + 169, + 566, + 405 + ], + "blocks": [ + { + "bbox": [ + 319, + 144, + 527, + 167 + ], + "lines": [ + { + "bbox": [ + 319, + 144, + 527, + 167 + ], + "spans": [ + { + "bbox": [ + 319, + 144, + 527, + 167 + ], + "type": "text", + "content": "Algorithm 1: Transferring Adversarily Robust Training (TAROT)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "lines": [ + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "spans": [ + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": "Input : " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : feature extractor, " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : head, " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\pi'" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : auxiliary head, " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : standard trained model, " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{S}} = \\{(x_{i}^{s},y_{i}^{s})\\}_{i = 1}^{m}" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : source domain dataset, " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{S}} = \\{(x_{i}^{s},y_{i}^{s})\\}_{i = 1}^{m}:\\mathcal{D}_{\\mathcal{T}} = \\{x_{j}^{t}\\}_{j = 1}^{m^{\\prime}}" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : target domain dataset, " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\eta_1,\\eta_2,\\lambda ,\\rho" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : learning rates in Eq. (20), " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : the number of epochs, " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : the number of batches, " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " : batch size Output: adversarially robust network " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\pi \\circ \\psi" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " \n1 " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\theta_{\\psi}\\gets \\theta_{\\mathrm{R}}" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " (initialized by a robustly trained feature extractor) \n2 for " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "b = 1,\\dots ,B" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " do \n3 for " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "k = 1,\\dots ,K" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " do \n4 Generate " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}}_{b,k}^{\\mathrm{t,adv}}" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " by PGD " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "(\\pi \\circ \\psi (\\pmb{x}_{b,k}^{t}),\\tilde{y}_{b,k}^{t})" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{b,k}^{t} = h_{s}(\\pmb{x}_{b,k}^{t})" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " \n5 Update " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "(\\theta_{\\psi},\\theta_{\\pi})\\gets (\\theta_{\\psi},\\theta_{\\pi})-" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\eta_1\\frac{1}{K}\\nabla_{(\\pmb {\\theta}_\\psi ,\\pmb {\\theta}_\\pi)}\\ell_{\\mathrm{TAROT}}(\\pmb {\\theta};\\{(x_{b,k}^{\\mathrm{s}},y_{b,k}^{\\mathrm{s}})\\}_{k = 1}^{K},\\{(x_{b,k}^{\\mathrm{t}})\\}_{k = 1}^{K})" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " \n6 Update " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\theta_{\\pi '}\\gets \\theta_{\\pi ' + }" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\eta_2\\frac{1}{K}\\nabla_{\\pmb{\\theta}_{\\pi '}}\\ell_{\\mathrm{TAROT}}(\\pmb {\\theta};\\{(x_{b,k}^{\\mathrm{s}},y_{b,k}^{\\mathrm{s}})\\}_{k = 1}^{K},\\{(x_{b,k}^{\\mathrm{t}})\\}_{k = 1}^{K},\\lambda)" + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "text", + "content": " \n7 end \n8 end \n9 Return " + }, + { + "bbox": [ + 318, + 169, + 566, + 405 + ], + "type": "inline_equation", + "content": "\\pi \\circ \\psi" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "algorithm" + }, + { + "bbox": [ + 314, + 427, + 430, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 427, + 430, + 440 + ], + "spans": [ + { + "bbox": [ + 314, + 427, + 430, + 440 + ], + "type": "text", + "content": "4.4. Learning Algorithm" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 444, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 502 + ], + "type": "text", + "content": "Inspired by Theorem 1, we propose the robust domain adaptation algorithm, which we call Transferring Adversarially RObust Training (TAROT). Our basic idea is to find a optimal score function " + }, + { + "bbox": [ + 313, + 444, + 555, + 502 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 444, + 555, + 502 + ], + "type": "text", + "content": " by solving a modified version of the following optimization problem:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 331, + 509, + 555, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 509, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 331, + 509, + 555, + 529 + ], + "type": "interline_equation", + "content": "\\min _ {f \\in \\mathcal {F}} \\mathcal {R} _ {\\widehat {\\mathcal {S}}} ^ {(\\rho)} (f) + d _ {f, \\mathcal {F}} ^ {\\operatorname {r o b}, (\\rho)} \\left(\\widehat {\\mathcal {S}} _ {\\mathbf {X}}, \\widehat {\\mathcal {T}} _ {\\mathbf {X}}\\right) + L _ {f} \\left(\\widehat {\\mathcal {T}} _ {\\mathbf {X}}, \\varepsilon\\right) \\tag {19}", + "image_path": "96730024763987bf7bf0c3a53b14b7c388f5c10255e4c13645722e3e85b0c0f9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 536, + 556, + 715 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "content": "- As is done in DANN (Domain-Adversarial Neural Network) [13] and MDD (Margin Disparity Discrepancy) [45], we consider a neural network structure that is composed of a feature extractor " + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "content": " and a head " + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "content": ", i.e. a score function " + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "content": " can be represented as " + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "inline_equation", + "content": "\\pi \\circ \\psi" + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "content": ". We only consider " + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "content": " being fixed, when calculating " + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "inline_equation", + "content": "d_{f,\\mathcal{F}}^{\\mathrm{rob},(\\rho)}" + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "content": ". That is, we calculate the robust margin disparity discrepancy on the feature space of " + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 536, + 555, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "text", + "content": "- To reduce the first and second term, we follow the formulation of the MDD algorithm. Likewise, we utilize an auxiliary head " + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi'" + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\gamma \\coloneqq \\exp \\rho" + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "text", + "content": " designed to attain the margin. " + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi'" + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "text", + "content": " shares the same hypothesis space with " + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "text", + "content": ", and works as an adversarial network. The discrepancy term, which is non-differentiable, can be optimized through GRL [12]. The only difference is that, when to calculate " + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "inline_equation", + "content": "d_{f,\\mathcal{F}}^{\\mathrm{rob},(\\rho)}" + }, + { + "bbox": [ + 313, + 631, + 556, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25784" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 298, + 207 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 298, + 207 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 298, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 298, + 207 + ], + "type": "image", + "image_path": "174ce636fe189adfb9db2e7411e5b4fac1ad994f24129b9774bde2a0ab86f855.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 213, + 251, + 224 + ], + "lines": [ + { + "bbox": [ + 99, + 213, + 251, + 224 + ], + "spans": [ + { + "bbox": [ + 99, + 213, + 251, + 224 + ], + "type": "text", + "content": "Figure 1. Overview of TAROT algorithm." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 251, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 251, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 63, + 251, + 295, + 275 + ], + "type": "text", + "content": "it requires calculating a robust disparity and a standard disparity rather than two standard disparities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 275, + 296, + 414 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 275, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 344 + ], + "type": "text", + "content": "- To reduce the local Lipschitz constant, we minimize the cross entropy loss with adversarial examples for given pseudo labels attached by a teacher model learned by a nonrobust UDA approach. This is based on the previously mentioned phenomenon that adversarial training reduces the local Lipschitz constant [42, 43]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 344, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 389 + ], + "type": "text", + "content": "- When calculating the robust margin disparity discrepancy, instead of generating adversarial examples, we reuse those generated for the local Lipsshitz constant, making the algorithm computationally effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 389, + 295, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 295, + 414 + ], + "type": "text", + "content": "- As an initial model for " + }, + { + "bbox": [ + 56, + 389, + 295, + 414 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 56, + 389, + 295, + 414 + ], + "type": "text", + "content": ", we used a robustly pretrained models with from the repository provided in [28]." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 420, + 252, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 420, + 252, + 432 + ], + "spans": [ + { + "bbox": [ + 67, + 420, + 252, + 432 + ], + "type": "text", + "content": "The objective function of TAROT is given as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 447, + 295, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 447, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 64, + 447, + 295, + 605 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\ell_ {\\text {T A R O T}} (\\boldsymbol {\\theta}; \\left\\{\\left(\\boldsymbol {x} _ {i} ^ {\\mathrm {s}}, y _ {i} ^ {\\mathrm {s}}\\right) \\right\\} _ {i = 1} ^ {m}, \\left\\{\\left(\\boldsymbol {x} _ {j} ^ {\\mathrm {t}}\\right) \\right\\} _ {j = m + 1} ^ {m + m ^ {\\prime}}, \\alpha , \\gamma) \\\\ = \\alpha \\left\\{ \\right.\\frac {1}{m} \\sum_ {i = 1} ^ {m} \\ell_ {\\mathrm {c e}} \\left(\\left(\\pi \\circ \\psi\\right)\\left(\\boldsymbol {x} _ {i}\\right), y _ {i}\\right)\\left. \\right)\\left. \\right. \\\\ + \\left(\\frac {1}{m ^ {\\prime}} \\sum_ {j = m + 1} ^ {m + m ^ {\\prime}} \\ell_ {\\mathrm {m o d - c e}} ^ {\\mathrm {r o b}} \\left(\\left(\\pi^ {\\prime} \\circ \\psi\\right) \\left(\\boldsymbol {x} _ {j}\\right), h _ {\\pi \\circ \\psi} \\left(\\boldsymbol {x} _ {j}\\right)\\right) \\right. \\tag {20} \\\\ \\left. \\left. - \\frac {\\gamma}{m} \\sum_ {j = 1} ^ {m} \\ell_ {\\mathrm {c e}} \\left(\\left(\\pi^ {\\prime} \\circ \\psi\\right) \\left(\\boldsymbol {x} _ {j}\\right), h _ {\\pi \\circ \\psi} \\left(\\boldsymbol {x} _ {j}\\right)\\right)\\right) \\right\\} \\\\ + \\frac {1}{m ^ {\\prime}} \\sum_ {j = m + 1} ^ {m + m ^ {\\prime}} \\ell_ {\\mathrm {c e}} ^ {\\mathrm {r o b}} ((\\pi \\circ \\psi) (\\boldsymbol {x} _ {j}), \\tilde {y} _ {j}) \\\\ \\end{array}", + "image_path": "a71799c71c79de14aa8018722d2966aa2923bdc9acfd888f4c9bc712ed79292d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{ce}}" + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{mod - ce}}^{\\mathrm{rob}}" + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "content": " [45] and " + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{ce}}^{\\mathrm{rob}}" + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "content": " [21] as cross-entropy, modified cross-entropy with adversarial examples and cross-entropy with adversarial examples respectively. Also, " + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} = (\\pmb{\\theta}_{\\psi},\\pmb{\\theta}_{\\pi},\\pmb{\\theta}_{\\pi'})" + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "content": " to represent the parameters corresponding to " + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\psi, \\pi, \\pi'" + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{y}_j = h_s(\\pmb{x}_j)" + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "content": " is a pseudo label generated by a standard trained model " + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 620, + 297, + 714 + ], + "type": "text", + "content": " is a trade-off parameter that balances transferability and target's robustness. TAROT algorithm is summarized in Algorithm 1 and Figure 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 91, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 196 + ], + "type": "text", + "content": "In this section, we compare TAROT with other competitors for adversarially robust UDA. We use several benchmark datasets: Office-31 [27], Office-Home [37], VisDA2017 [25], and DomainNet [26]. Throughout these datasets, effectiveness of TAROT is evaluated across a wide range of domain shifts and robustness scenarios. Further experimental details including explanation for datasets and experimental setups are presented in Sec. 9. The code is available in the repository" + }, + { + "bbox": [ + 313, + 91, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 313, + 91, + 555, + 196 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 203, + 451, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 451, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 451, + 215 + ], + "type": "text", + "content": "5.1. Performance Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "content": "Training Setup All of the results are averaged over three runs, with different teacher model initializations. Different perturbation budgets " + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "inline_equation", + "content": "(\\varepsilon)" + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "content": " are chosen based on the difficulty of each benchmark dataset. For more challenging datasets, a smaller " + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "content": " is used, and vice versa. Specifically, we use " + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\varepsilon = 16 / 255" + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "content": " for the Office-31 and Office-Home datasets, while " + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\varepsilon = 8 / 255" + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\varepsilon = 4 / 255" + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "content": " are used for VisDA2017 and DomainNet, respectively. The values of " + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "content": " in Eq. (20) are set to 0.05, 0.1, 0.5, and 1.0 for Office31, OfficeHome, VisDA2017, and DomainNet, respectively. We empirically observed that the optimal " + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "content": " values for each task vary much. The sensitivity analysis of " + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 220, + 555, + 357 + ], + "type": "text", + "content": " is provided in Sec. 5.2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 358, + 556, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 556, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 556, + 462 + ], + "type": "text", + "content": "As baselines, along with ARTUDA [20], RFA [3], and SRoUDA [46], we consider the pseudo labeling (PL) algorithm which is implemented as a simple PGD-AT algorithm where target pseudo labels are obtained from a fixed non-robustly trained teachers learned by MDD [45]. All algorithms are initialized with Robust-PT as is done for TAROT. Without Robust-PT, training with larger perturbations such as " + }, + { + "bbox": [ + 313, + 358, + 556, + 462 + ], + "type": "inline_equation", + "content": "\\varepsilon = 16 / 255" + }, + { + "bbox": [ + 313, + 358, + 556, + 462 + ], + "type": "text", + "content": " can lead to non-decreasing training loss in certain tasks, as discussed in Sec. 10.2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 464, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 568 + ], + "type": "text", + "content": "Evaluation Setup We report standard accuracies on clean samples and robust accuracies against AutoAttack [10]. AutoAttack, which includes both white-box and black-box attacks, is a powerful evaluation tool that mitigates errors in robustness measurement caused by gradient obfuscation—a phenomenon where model gradients are distorted or masked, leading to overestimated robustness [2]. AutoAttack is widely recognized for accurately assessing adversarial robustness by bypassing these gradient masking issues." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 575, + 481, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 481, + 586 + ], + "type": "text", + "content": "5.1.1. Performance on Target Domains" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 590, + 556, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 556, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 556, + 694 + ], + "type": "text", + "content": "Tables 1, 2, 4 and 5 demonstrate that PL outperforms other algorithms, with the exception of TAROT. As seen in Figure 3, simply combining Robust-PT as an initial solution with pseudo labeling can yield significant performance gains, which confirms our first theoretical insight. (See Sec. 5.2.) This combination serves as a strong baseline in adversarially robust domain adaptation. It's noteworthy that ARTUDA [20], RFA [3], and SRoUDA [46] algorithms show remarkably poor performance on several tasks. Note that the robust" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 480, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 480, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 480, + 713 + ], + "type": "text", + "content": "https://github.com/dyoony/TAROT" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25785" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 96, + 520, + 156 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Performances of ARTUDA, RFA, SROUDA, PL and TAROT on Office31 (" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\varepsilon = 16/255" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "). In each cell, the first number is the standard accuracy (\\%), while the second number is the robust accuracy (\\%) for AA. Bold numbers indicate the best performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 96, + 520, + 156 + ], + "lines": [ + { + "bbox": [ + 89, + 96, + 520, + 156 + ], + "spans": [ + { + "bbox": [ + 89, + 96, + 520, + 156 + ], + "type": "table", + "html": "
MethodA → DA → WD → AD → WW → AW → DAvg.
ARTUDA23.49 / 11.6522.39 / 9.5635.00 / 16.5168.55 / 35.8536.81 / 19.7474.10 / 51.6143.39 / 24.15
RFA78.72 / 1.2171.20 / 1.2655.84 / 18.0396.73 / 6.4260.10 / 23.0099.40 / 6.0277.00 / 9.32
SRoUDA93.37 / 42.3778.87 / 65.1617.93 / 3.4496.00 / 44.2866.53 / 47.21100.00 / 62.4574.62 / 44.15
PL91.77 / 84.1491.82 / 91.7074.19 / 67.8797.99 / 95.9872.42 / 65.46100.00 / 92.5788.03 / 82.95
TAROT94.18 / 84.1495.47 / 93.4674.51 / 67.1397.99 / 96.2373.73 / 66.88100.00 / 91.7789.31 / 83.27
", + "image_path": "75fa8586d56bd814ae2cf20d2303ec4f784cb2933556e504f2bbce91817ad281.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 91, + 188, + 518, + 305 + ], + "blocks": [ + { + "bbox": [ + 55, + 163, + 555, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 555, + 185 + ], + "type": "text", + "content": "Table 2. Performances of ARTUDA, RFA, SRoUDA, PL and TAROT on OfficeHome (ε = 16/255). In each cell, the first number is the standard accuracy (%), while the second number is the robust accuracy (% for AA. Bold numbers indicate the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 188, + 518, + 305 + ], + "lines": [ + { + "bbox": [ + 91, + 188, + 518, + 305 + ], + "spans": [ + { + "bbox": [ + 91, + 188, + 518, + 305 + ], + "type": "table", + "html": "
MethodAr → ClAr → PrAr → RwCl → ArCl → PrCl → Rw
ARTUDA30.29 / 13.5626.18 / 4.755.88 / 0.003.42 / 0.0025.34 / 10.4820.98 / 5.53
RFA48.39 / 17.8954.67 / 8.7663.74 / 6.3443.22 / 3.1758.03 / 9.1557.24 / 5.69
SRoUDA51.84 / 37.8975.60 / 45.4674.25 / 40.2857.97 / 33.4267.76 / 54.7767.32 / 35.83
PL55.72 / 44.3173.53 / 58.4973.10 / 40.2657.68 / 31.1570.31 / 56.2168.14 / 37.37
TAROT55.72 / 46.5373.85 / 58.9377.78 / 42.6260.12 / 34.6973.08 / 58.1469.93 / 38.95
Pr → ArPr → ClPr → RwRw → ArRw → ClRw → PrAvg.
ARTUDA8.74 / 1.7338.92 / 14.5747.53 / 5.3030.12 / 2.1041.79 / 22.2945.19 / 13.9727.03 / 7.86
RFA42.03 / 2.5647.86 / 13.3163.44 / 5.2154.27 / 3.3854.27 / 16.7072.85 / 9.7855.00 / 8.49
SRoUDA59.21 / 31.4048.52 / 39.7775.40 / 41.2070.87 / 40.3853.40 / 42.4383.26 / 67.3857.97 / 33.42
PL57.15 / 29.5852.74 / 44.4074.82 / 40.7468.97 / 38.2857.03 / 45.1782.86 / 66.6266.00 / 44.38
TAROT59.62 / 32.7654.98 / 47.0178.66 / 42.9471.90 / 40.7959.45 / 49.7484.39 / 68.5368.29 / 46.80
", + "image_path": "4abba62c153644c41c883861282521a3a8cbcf05245b183dbf2af94fb3b26c30.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 315, + 297, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 297, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 297, + 453 + ], + "type": "text", + "content": "accuracies of a standard training algorithm (MDD) on the Office-31, Office-Home, and VisDA-2017 datasets are exactly 0.0, highlighting the importance of robust adaptation methods. This is mainly attributed to our use of larger perturbation budgets during training and evaluation compared to those used in the original papers. For reference, we provide results using smaller perturbation budgets " + }, + { + "bbox": [ + 54, + 315, + 297, + 453 + ], + "type": "inline_equation", + "content": "(4/255, 8/255)" + }, + { + "bbox": [ + 54, + 315, + 297, + 453 + ], + "type": "text", + "content": " on Office-31 and OfficeHome datasets in the Appendix. Overall, TAROT outperforms existing algorithms across all benchmark datasets. Note that the robust accuracies of a standard training algorithm (MDD) on the Office-31, Office-Home, and VisDA-2017 datasets are exactly 0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 261, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 261, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 261, + 495 + ], + "type": "text", + "content": "5.1.2. Essentially Domain-Invariant Robustness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 506, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 506, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 506, + 297, + 713 + ], + "type": "text", + "content": "In this subsection, we empirically highlight the limitations of the pseudo labeling approach, which focuses solely on improving target domain performance without effectively learning domain-invariant features. We select tasks with Real world (Rw) as the target domain on the OfficeHome dataset. We compare TAROT (our algorithm), which is designed to reduce the distribution gap between source and target domains, with PL, evaluating both the source domain accuracy and the performance on unseen domains. Table 3 shows that our method not only adapts well to the target domain but also learns more generalizable and robust features, leading to improved robustness across domains by significant margins. We present an upper bound for robust risk on the source domain and demonstrate that the components of this bound closely align with those of the target domain, offering a partial theoretical explanation for TAROT's robustness to adversarial attacks on the source domain. Additional results for other tasks are provided in Sec. 10." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 391, + 564, + 544 + ], + "blocks": [ + { + "bbox": [ + 313, + 312, + 556, + 388 + ], + "lines": [ + { + "bbox": [ + 313, + 312, + 556, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 556, + 388 + ], + "type": "text", + "content": "Table 3. Performances of PL and TAROT on Source Domain and Unseen Domain, on OfficeHome (" + }, + { + "bbox": [ + 313, + 312, + 556, + 388 + ], + "type": "inline_equation", + "content": "\\varepsilon = 8/255" + }, + { + "bbox": [ + 313, + 312, + 556, + 388 + ], + "type": "text", + "content": "). In each cell, the first number is the standard accuracy (\\%), while the second number corresponds to the robust accuracy (\\%) for AA. Bold numbers indicate the best performance. A " + }, + { + "bbox": [ + 313, + 312, + 556, + 388 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 312, + 556, + 388 + ], + "type": "text", + "content": " B(C) indicates that A, B, and C represent the source, (unlabeled) target, and evaluation domains, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 391, + 564, + 544 + ], + "lines": [ + { + "bbox": [ + 315, + 391, + 564, + 544 + ], + "spans": [ + { + "bbox": [ + 315, + 391, + 564, + 544 + ], + "type": "table", + "html": "
SourceUnseen
MethodAr → Rw(Ar)Ar → Rw(Cl)Ar → Rw(Pr)Avg.
ARTUDA18.83 / 2.647.70 / 0.896.28 / 0.4510.94 / 1.33
RFA99.62 / 47.0242.52 / 21.5351.57 / 21.5164.57 / 30.02
SRoUDA39.39 / 17.4339.54 / 28.6650.75 / 35.8642.23 / 27.32
PL41.78 / 19.4141.97 / 31.3952.42 / 37.2445.39 / 29.34
TAROT98.35 / 64.2447.86 / 35.1958.32 / 39.5468.18 / 46.32
Cl → Rw(Cl)Cl → Rw(Ar)Cl → Rw(Pr)Avg.
ARTUDA84.81 / 55.8014.34 / 3.2127.26 / 13.9942.14 / 24.33
RFA98.35 / 80.1040.37 / 8.8654.86 / 22.2564.53 / 37.07
SRoUDA48.29 / 35.5833.87 / 15.6248.19 / 33.0743.45 / 28.09
PL48.75 / 35.9536.30 / 16.6950.80 / 35.6645.28 / 29.43
TAROT93.65 / 84.3539.72 / 17.1855.46 / 37.6762.95 / 46.40
Pr → Rw(Pr)Pr → Rw(Ar)Pr → Rw(Cl)Avg.
ARTUDA98.29 / 60.1725.67 / 4.6236.24 / 16.2453.40 / 27.01
RFA99.75 / 68.5940.67 / 7.8741.76 / 19.7360.73 / 32.06
SRoUDA61.48 / 41.4034.82 / 16.0742.08 / 30.6846.13 / 29.38
PL61.05 / 43.3436.09 / 17.1042.11 / 30.9146.42 / 30.45
TAROT96.13 / 82.4339.14 / 16.1546.87 / 33.0160.71 / 43.86
", + "image_path": "7a0d9aa3ba660fae36ef210edebc235b2ea86162e3a9b51fefa6975aa9c3ee2a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 551, + 416, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 416, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 416, + 563 + ], + "type": "text", + "content": "5.2. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 568, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 604 + ], + "type": "text", + "content": "We conduct ablation studies about: (i) effect of Robust-PT (ii) sensitivity analysis of " + }, + { + "bbox": [ + 313, + 568, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 568, + 555, + 604 + ], + "type": "text", + "content": ". All exact values and additional experimental results are presented in Sec. 10." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "text", + "content": "Effect of Robust-PT on Various " + }, + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "text", + "content": " Figures 3a and 3b show that the both algorithms perform significantly better when Robust-PT is applied, particularly as " + }, + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "text", + "content": " increases, indicating that Robust-PT is crucial for robustness against strong attack. Without Robust-PT, both TAROT and PL experience sharp declines in accuracy, especially as " + }, + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "text", + "content": " becomes larger. Figures 3c and 3d illustrate that TAROT uniformly outperforms PL across different perturbation budgets " + }, + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "inline_equation", + "content": "(\\varepsilon)" + }, + { + "bbox": [ + 313, + 621, + 557, + 715 + ], + "type": "text", + "content": " both on" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25786" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 106, + 382, + 392 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 393, + 102 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 393, + 102 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 393, + 102 + ], + "type": "text", + "content": "Table 4. Performances of ARTUDA, RFA, SROUDA, PL and TAROT on DomainNet " + }, + { + "bbox": [ + 61, + 70, + 393, + 102 + ], + "type": "inline_equation", + "content": "(\\varepsilon = 4 / 255)" + }, + { + "bbox": [ + 61, + 70, + 393, + 102 + ], + "type": "text", + "content": ". In each cell, the first number is the standard accuracy " + }, + { + "bbox": [ + 61, + 70, + 393, + 102 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 61, + 70, + 393, + 102 + ], + "type": "text", + "content": ", and the second number is the robust accuracy " + }, + { + "bbox": [ + 61, + 70, + 393, + 102 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 61, + 70, + 393, + 102 + ], + "type": "text", + "content": " for AA. Bold numbers indicate the best performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 106, + 382, + 392 + ], + "lines": [ + { + "bbox": [ + 62, + 106, + 382, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 106, + 382, + 392 + ], + "type": "table", + "html": "
MethodC → IC → PC → RC → SAvg.
ARTUDA12.64 / 3.7631.15 / 7.9044.90 / 16.4939.90 / 17.5832.15 / 11.43
RFA9.42 / 3.4723.33 / 7.6037.54 / 17.7427.40 / 12.1724.42 / 10.24
SRoUDA9.27 / 5.6824.33 / 15.0742.91 / 28.3832.42 / 23.8227.23 / 18.24
PL10.59 / 6.6927.06 / 15.7343.57 / 28.6832.13 / 23.5228.34 / 18.66
TAROT14.65 / 8.1633.23 / 17.3650.02 / 30.4640.11 / 27.7734.50 / 20.94
I → CI → PI → RI → SAvg.
ARTUDA31.89 / 12.0720.38 / 2.5623.79 / 5.4734.61 / 5.9027.67 / 6.50
RFA15.00 / 8.8214.80 / 4.8019.15 / 8.7611.44 / 5.0515.10 / 6.86
SRoUDA25.95 / 14.1824.27 / 13.9336.32 / 22.0620.08 / 13.5226.65 / 15.92
PL17.45 / 13.1622.42 / 13.4132.50 / 21.9418.25 / 13.1422.65 / 15.41
TAROT29.36 / 20.2327.66 / 14.3039.14 / 23.3921.89 / 14.0829.51 / 18.00
P → CP → IP → RP → SAvg.
ARTUDA42.89 / 20.5814.36 / 3.2149.03 / 16.6234.02 / 13.1235.08 / 13.38
RFA26.27 / 14.859.45 / 3.0937.47 / 17.0419.67 / 8.5923.22 / 10.89
SRoUDA36.69 / 28.2610.17 / 5.8645.69 / 29.9530.95 / 19.1430.87 / 20.80
PL33.86 / 26.6911.18 / 7.2544.56 / 29.4729.76 / 20.6629.84 / 21.02
TAROT41.27 / 30.5114.82 / 7.6951.10 / 30.7637.22 / 24.7136.10 / 23.42
R → CR → IR → PR → SAvg.
ARTUDA52.94 / 28.1016.94 / 4.1741.95 / 12.0236.11 / 14.8636.98 / 14.79
RFA36.72 / 19.3012.36 / 3.0232.03 / 8.3422.73 / 8.1225.96 / 9.70
SRoUDA44.33 / 34.3310.10 / 5.6340.26 / 23.2629.10 / 18.5630.94 / 20.44
PL42.03 / 33.2711.85 / 7.1439.26 / 22.8029.05 / 19.6830.55 / 20.72
TAROT52.26 / 37.5818.86 / 9.4547.62 / 23.4339.38 / 25.2339.53 / 23.92
S → CS → IS → PS → RAvg.
ARTUDA54.24 / 30.5913.14 / 3.8934.87 / 7.7941.40 / 13.4735.91 / 13.94
RFA33.87 / 19.949.47 / 3.5427.29 / 7.8330.39 / 13.7025.26 / 11.25
SRoUDA46.62 / 36.658.50 / 5.2831.09 / 15.1442.95 / 22.3332.29 / 19.85
PL42.48 / 34.668.95 / 6.1831.48 / 18.7542.55 / 28.2431.37 / 21.96
TAROT53.35 / 41.3813.71 / 7.5239.77 / 20.4348.77 / 30.0238.90 / 24.84
", + "image_path": "eaf3562e92faf3a94bd5d11ee4720993c2994ed0e9bbf1eea171f232e564b6a6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 397, + 114, + 547, + 173 + ], + "blocks": [ + { + "bbox": [ + 395, + 67, + 549, + 110 + ], + "lines": [ + { + "bbox": [ + 395, + 67, + 549, + 110 + ], + "spans": [ + { + "bbox": [ + 395, + 67, + 549, + 110 + ], + "type": "text", + "content": "Table 5. Performances of ARTUDA, RFA, SRoUDA, PL and TAROT on VisDA2017 " + }, + { + "bbox": [ + 395, + 67, + 549, + 110 + ], + "type": "inline_equation", + "content": "(\\varepsilon = 8 / 255)" + }, + { + "bbox": [ + 395, + 67, + 549, + 110 + ], + "type": "text", + "content": ". Standard accuracy " + }, + { + "bbox": [ + 395, + 67, + 549, + 110 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 395, + 67, + 549, + 110 + ], + "type": "text", + "content": " / Robust accuracy " + }, + { + "bbox": [ + 395, + 67, + 549, + 110 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 395, + 67, + 549, + 110 + ], + "type": "text", + "content": " for AA." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 397, + 114, + 547, + 173 + ], + "lines": [ + { + "bbox": [ + 397, + 114, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 397, + 114, + 547, + 173 + ], + "type": "table", + "html": "
MethodSyn. → RealSyn.
ARTUDA10.73 / 0.0061.43 / 5.70
RFA64.86 / 8.6599.41 / 33.55
SRoUDA24.44 / 0.2048.56 / 24.05
PL66.71 / 37.8340.98 / 24.09
TAROT67.12 / 37.7785.18 / 51.21
", + "image_path": "b00a7bb923545dce8c7a68edd670233c2f49c9941898ad8705778d0b30dec4de.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 397, + 186, + 556, + 280 + ], + "blocks": [ + { + "bbox": [ + 397, + 186, + 556, + 280 + ], + "lines": [ + { + "bbox": [ + 397, + 186, + 556, + 280 + ], + "spans": [ + { + "bbox": [ + 397, + 186, + 556, + 280 + ], + "type": "image", + "image_path": "f3b8f55ff8b96c4338004d1955886489d7ca9f863c8293509ef079c1dc666455.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 397, + 281, + 555, + 373 + ], + "blocks": [ + { + "bbox": [ + 397, + 281, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 397, + 281, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 397, + 281, + 555, + 373 + ], + "type": "image", + "image_path": "96b420a4ef224c89539bcbab4a0c5d3cd7490df450632d87d181ee64d13b631c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 374, + 548, + 395 + ], + "lines": [ + { + "bbox": [ + 395, + 374, + 548, + 395 + ], + "spans": [ + { + "bbox": [ + 395, + 374, + 548, + 395 + ], + "type": "text", + "content": "Figure 2. Sensitivity Analysis of " + }, + { + "bbox": [ + 395, + 374, + 548, + 395 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 395, + 374, + 548, + 395 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 395, + 374, + 548, + 395 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 395, + 374, + 548, + 395 + ], + "type": "text", + "content": " corresponds to PL." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "text", + "content": "the standard accuracies and robust accuracies against AA. Especially, that performance gap is increasing as " + }, + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "text", + "content": " increases. Figures 3a and 3b show that both algorithms perform significantly better when Robust-PT is applied, particularly as " + }, + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "text", + "content": " increases, indicating that Robust-PT is crucial for robustness against strong attacks. Figures 3c and 3d illustrate that TAROT uniformly outperforms the PL approach across different perturbation budgets " + }, + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "inline_equation", + "content": "(\\varepsilon)" + }, + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "text", + "content": ", both in standard accuracies and robust accuracies against AA. Notably, the performance gap increases as " + }, + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 407, + 296, + 522 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "text", + "content": "Sensitivity Analysis of " + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "text", + "content": " Figure 2 illustrates the sensitivity with respect to " + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "text", + "content": " in Eq. (20). It shows how the performance of generalization and robustness changes for the target and unseen domains as the value of " + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "text", + "content": " varies. As shown in the figure, each task has a different optimal " + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "text", + "content": ", indicating that the optimal " + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 541, + 296, + 621 + ], + "type": "text", + "content": " value for one task may not be optimal for another. The detailed numbers are provided in Sec. 10.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 230, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 230, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 230, + 647 + ], + "type": "text", + "content": "6. Conclusions and Future Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 655, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 296, + 713 + ], + "type": "text", + "content": "To our best knowledge, this is the first work for adversarially robust domain adaptation algorithms grounded with theoretical explanations. Compared to existing works, our algorithm shows essentially domain invariant robust features, resulting significant improved performance in source and" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 406, + 422, + 479 + ], + "blocks": [ + { + "bbox": [ + 317, + 406, + 422, + 479 + ], + "lines": [ + { + "bbox": [ + 317, + 406, + 422, + 479 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 422, + 479 + ], + "type": "image", + "image_path": "d2807fb685b1d7e9c8a842740ad4a1ccb52e9364942b8b28014dd38e90f2b0f1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 482, + 409, + 491 + ], + "lines": [ + { + "bbox": [ + 331, + 482, + 409, + 491 + ], + "spans": [ + { + "bbox": [ + 331, + 482, + 409, + 491 + ], + "type": "text", + "content": "(a) PL w/ and w/o R-PT" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 493, + 422, + 566 + ], + "blocks": [ + { + "bbox": [ + 317, + 493, + 422, + 566 + ], + "lines": [ + { + "bbox": [ + 317, + 493, + 422, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 422, + 566 + ], + "type": "image", + "image_path": "aafed76dc8d0f073ede67c38a8a7cbfeed5c6cc481d8bdb7f4cb325c3594f27f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 569, + 405, + 578 + ], + "lines": [ + { + "bbox": [ + 333, + 569, + 405, + 578 + ], + "spans": [ + { + "bbox": [ + 333, + 569, + 405, + 578 + ], + "type": "text", + "content": "(c) Rob Acc. w/ R-PT" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 447, + 406, + 552, + 478 + ], + "blocks": [ + { + "bbox": [ + 447, + 406, + 552, + 478 + ], + "lines": [ + { + "bbox": [ + 447, + 406, + 552, + 478 + ], + "spans": [ + { + "bbox": [ + 447, + 406, + 552, + 478 + ], + "type": "image", + "image_path": "2f7c223d0d73c962e58b3e88a13ba6652ceada2cb71175daf27c0fba7d73bcfe.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 482, + 547, + 491 + ], + "lines": [ + { + "bbox": [ + 452, + 482, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 452, + 482, + 547, + 491 + ], + "type": "text", + "content": "(b) TAROT w/ and w/o R-PT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 447, + 493, + 552, + 565 + ], + "blocks": [ + { + "bbox": [ + 447, + 493, + 552, + 565 + ], + "lines": [ + { + "bbox": [ + 447, + 493, + 552, + 565 + ], + "spans": [ + { + "bbox": [ + 447, + 493, + 552, + 565 + ], + "type": "image", + "image_path": "6378e763a5c4c74ae86c8697fc4c490b20a2b2ae6acb9f3f2eceefbee0b679f3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 569, + 538, + 578 + ], + "lines": [ + { + "bbox": [ + 460, + 569, + 538, + 578 + ], + "spans": [ + { + "bbox": [ + 460, + 569, + 538, + 578 + ], + "type": "text", + "content": "(d) Stand Acc. w/ R-PT" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 588, + 555, + 620 + ], + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 620 + ], + "type": "text", + "content": "Figure 3. Effect of Robust-PT across various methods and " + }, + { + "bbox": [ + 313, + 588, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 588, + 555, + 620 + ], + "type": "text", + "content": " values. Each plot compares different configurations of PL w/ MDD and TAROT for standard and robust accuracy." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 633, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 556, + 713 + ], + "type": "text", + "content": "unseen domains. Our work not only serves as a cornerstone for understanding robust domain adaptation from a theoretical perspective, but we also believe TAROT can be served as a state-of-the-art algorithm in practical perspective by excelling researches of adversarial robustness for transfer learning such as domain adaptation/generalization and few-shot learning task." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25787" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 190 + ], + "type": "text", + "content": "Acknowledgement This work is partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-II220184, Development and Study of AI Technologies to Inexpensively Conform to Evolving Policy on Ethics), National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(No. 2022R1A5A7083908) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) [NO.RS-2021-II211343, Artificial Intelligence Graduate School Program (Seoul National University)]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 228, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 228, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 280 + ], + "type": "text", + "content": "[1] Eric Arazo, Diego Ortega, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 281, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 296, + 323 + ], + "type": "text", + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International conference on machine learning, pages 274-283. PMLR, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 323, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 296, + 376 + ], + "type": "text", + "content": "[3] Muhammad Awais, Fengwei Zhou, Hang Xu, Lanqing Hong, Ping Luo, Sung-Ho Bae, and Zhenguo Li. Adversarial robustness for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8568-8577, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 376, + 296, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 429 + ], + "type": "text", + "content": "[4] Francesco Barbato, Marco Toldo, Umberto Michieli, and Pietro Zanuttigh. Latent space regularization for unsupervised domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2835-2845, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 429, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 429, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 429, + 296, + 460 + ], + "type": "text", + "content": "[5] Peter L Bartlett, Michael I Jordan, and Jon D McAuliffe. Convexity, classification, and risk bounds. Journal of the American Statistical Association, 101(473):138-156, 2006. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 501 + ], + "type": "text", + "content": "[6] Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 544 + ], + "type": "text", + "content": "[7] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79: 151-175, 2010. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 544, + 296, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 597 + ], + "type": "text", + "content": "[8] Mingcai Chen, Hao Cheng, Yuntao Du, Ming Xu, Wenyu Jiang, and Chongjun Wang. Two wrongs don't make a right: Combating confirmation bias in learning with label noise. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14765-14773, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 597, + 296, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 597, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 62, + 597, + 296, + 649 + ], + "type": "text", + "content": "[9] Moustapha Cisse, Piotr Bojanowski, Edouard Grave, Yann Dauphin, and Nicolas Usunier. Parseval networks: Improving robustness to adversarial examples. In International conference on machine learning, pages 854-863. PMLR, 2017. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 650, + 296, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 650, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 57, + 650, + 296, + 692 + ], + "type": "text", + "content": "[10] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[11] Antonio Galbis and Manuel Maestre. Vector analysis versus vector calculus. Springer Science & Business Media, 2012. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "text", + "content": "[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 106, + 555, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 555, + 158 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 555, + 158 + ], + "type": "text", + "content": "[13] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 159, + 555, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 159, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 317, + 159, + 555, + 213 + ], + "type": "text", + "content": "[14] Zhiqiang Gao, Shufei Zhang, Kaizhu Huang, Qiufeng Wang, Rui Zhang, and Chaoliang Zhong. Certifying better robust generalization for unsupervised domain adaptation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2399-2410, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 213, + 554, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 213, + 554, + 256 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 554, + 256 + ], + "type": "text", + "content": "[15] Chao Gong, Kai Chen, Zhipeng Wei, Jingjing Chen, and Yulong Jiang. Reliable and efficient concept erasure of text-to-image diffusion models. In In European Conference on Computer Vision (ECCV), 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 257, + 555, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 257, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 257, + 555, + 299 + ], + "type": "text", + "content": "[16] Fredrik D. Johansson, David Sontag, and Rajesh Ranganath. Support and invertibility in domain-invariant representations. International Conference on Artificial Intelligence and Statistics, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 300, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 554, + 342 + ], + "type": "text", + "content": "[17] Matt Jordan and Alexandros G Dimakis. Exactly computing the local lipschitz constant of relu networks. Advances in Neural Information Processing Systems, 33:7344-7353, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 376 + ], + "type": "text", + "content": "[18] Bo Fu Jinguang Jiang, Baixu Chen and Mingsheng Long. Transfer-learning-library. https://github.com/thuml/Transfer-Learning-Library, 2020.5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 554, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 418 + ], + "type": "text", + "content": "[19] Changhoon Kim, Kyle Min, and Yezhou Yang. R.a.c.e.: Robust adversarial concept erasure for secure text-to-image diffusion model. In In European Conference on Computer Vision (ECCV), 2025. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 419, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 554, + 453 + ], + "type": "text", + "content": "[20] Shao-Yuan Lo and Vishal M. Patel. Exploring adversarially robust training for unsupervised domain adaptation. In Asian Conference on Computer Vision (ACCV), 2022. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 453, + 554, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 554, + 505 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 554, + 505 + ], + "type": "text", + "content": "[21] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations (ICLR), 2018. 1, 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 506, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 506, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 506, + 554, + 548 + ], + "type": "text", + "content": "[22] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. In In International Conference on Learning Representations (ICLR), 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 549, + 554, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 549, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 316, + 549, + 554, + 571 + ], + "type": "text", + "content": "[23] Afshin Rostamizadeh Mehryar Mohri and Ameet Talwalkar. Foundations of Machine Learning. MIT Press, 2012. 4, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 572, + 554, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 572, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 316, + 572, + 554, + 604 + ], + "type": "text", + "content": "[24] Farzeen Munir, Shoaib Azam, and Moongu Jeon. Sstn: Self-supervised domain adaptation thermal object detection for autonomous driving, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 605, + 554, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 605, + 554, + 636 + ], + "spans": [ + { + "bbox": [ + 316, + 605, + 554, + 636 + ], + "type": "text", + "content": "[25] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge, 2017. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 637, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 554, + 689 + ], + "type": "text", + "content": "[26] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "type": "text", + "content": "[27] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In In" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25788" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision (ECCV), pages 213-226. Springer, 2010. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 138 + ], + "type": "text", + "content": "[28] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? In Conference on Neural Information Processing Systems (NeurIPS), 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 192 + ], + "type": "text", + "content": "[29] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. In In International Conference on Machine Learning (ICML), 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 245 + ], + "type": "text", + "content": "[30] Zhouxing Shi, Yihan Wang, Huan Zhang, J Zico Kolter, and Cho-Jui Hsieh. Efficiently computing local lipschitz constants of neural networks via bound propagation. Advances in Neural Information Processing Systems, 35:2350-2364, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 247, + 296, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 247, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 296, + 289 + ], + "type": "text", + "content": "[31] Peeyush Singhal, Rahee Walambe, Sheela Ramanna, and Ketan Kotecha. Domain adaptation: Challenges, methods, datasets, and applications. IEEE Access, 11:6973-7020, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 291, + 296, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 296, + 333 + ], + "type": "text", + "content": "[32] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations (ICLR), 2014. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 335, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 296, + 356 + ], + "type": "text", + "content": "[33] Michel Talagrand. Upper and lower bounds for stochastic processes. Springer, 2014. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 357, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 357, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 296, + 388 + ], + "type": "text", + "content": "[34] Marvin Teichmann, Michael Weber, Marius Zoellner, Roberto Cipolla, and Raquel Urtasun. Multinet: Real-time joint semantic reasoning for autonomous driving, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 390, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 390, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 390, + 296, + 421 + ], + "type": "text", + "content": "[35] Thulasi Tholeti and Sheetal Kalyani. The robust way to stack and bag: the local lipschitz way. arXiv preprint arXiv:2206.00513, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 474 + ], + "type": "text", + "content": "[36] Jonathan Uesato, Jean-Baptiste Alayrac, Po-Sen Huang, Robert Stanforth, Alhussein Fawzi, and Pushmeet Kohli. Are labels required for improving adversarial robustness? In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 476, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 296, + 529 + ], + "type": "text", + "content": "[37] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018-5027, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 530, + 296, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 530, + 296, + 572 + ], + "spans": [ + { + "bbox": [ + 56, + 530, + 296, + 572 + ], + "type": "text", + "content": "[38] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International Conference on Learning Representataions(ICLR), 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 574, + 296, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 574, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 574, + 296, + 616 + ], + "type": "text", + "content": "[39] Boxi Wu, Jinghui Chen, Deng Cai, Xiaofei He, and Quanquan Gu. Do wider neural networks really help adversarial robustness? Advances in Neural Information Processing Systems, 34:7054-7067, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 616, + 296, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 296, + 670 + ], + "type": "text", + "content": "[40] Dongyoon Yang, Insung Kong, and Yongdai Kim. Enhancing adversarial robustness in low-label regime via adaptively weighted regularization and knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4552-4561, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 671, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 671, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 671, + 296, + 713 + ], + "type": "text", + "content": "[41] Dongyoon Yang, Insung Kong, and Yongdai Kim. Improving adversarial robustness by putting more regularizations on less robust samples. In International Conference on Machine Learning, pages 39331-39348. PMLR, 2023. 1, 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 307 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 73, + 554, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 554, + 115 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 554, + 115 + ], + "type": "text", + "content": "[42] Yao-Yuan Yang, Cyrus Rashtchian, Hongyang Zhang, Ruslan Salakhutdinov, and Kamalika Chaudhuri. A closer look at accuracy vs. robustness. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 4, 6, 1, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 117, + 554, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 117, + 554, + 158 + ], + "spans": [ + { + "bbox": [ + 316, + 117, + 554, + 158 + ], + "type": "text", + "content": "[43] Bohang Zhang, Du Jiang, Di He, and Liwei Wang. Rethinking lipschitz neural networks and certified robustness: A boolean function perspective. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 4, 6, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 160, + 555, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 160, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 316, + 160, + 555, + 212 + ], + "type": "text", + "content": "[44] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P Xing, Laurent El Ghaoui, and Michael I Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning (ICML), 2019. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 213, + 555, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 213, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 555, + 255 + ], + "type": "text", + "content": "[45] Yuchen Zhang, Tianle Liu, Mingsheng Long, and Michael I. Jordan. Bridging theory and algorithm for domain adaptation. In International Conference on Machine Learning (ICML), 2019. 2, 3, 4, 5, 6, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 256, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 256, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 555, + 307 + ], + "type": "text", + "content": "[46] Wanqing Zhu, Jia-Li Yin, Bo-Hao Chen, and Xineng Liu. Srouda: meta self-training for robust unsupervised domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3852-3860, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25789" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_content_list.json b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ec8c634f8a55591758dc8c4c44a8b2b414db41b3 --- /dev/null +++ b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_content_list.json @@ -0,0 +1,1686 @@ +[ + { + "type": "text", + "text": "TASTE-Rob: Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation", + "text_level": 1, + "bbox": [ + 124, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongxiang Zhao $^{1*}$ Xingchen Liu $^{1*†}$ Mutian Xu $^{1}$ Yiming Hao $^{1}$ Weikai Chen $\\ddagger$ Xiaoguang Han $^{1,2§}$", + "bbox": [ + 104, + 202, + 942, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ SSE, CUHKSZ $^{2}$ FNii, CUHKSZ https://taste-rob.github.io", + "bbox": [ + 351, + 237, + 676, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address key limitations in existing datasets and models for task-oriented hand-object interaction video generation, a critical approach of generating video demonstrations for robotic imitation learning. Current datasets, such as Ego4D [16], often suffer from inconsistent view perspectives and misaligned interactions, leading to reduced video quality and limiting their applicability for precise imitation learning tasks. Towards this end, we introduce TASTE-Rob — a pioneering large-scale dataset of 100,856 ego-centric hand-object interaction videos. Each video is meticulously aligned with language instructions and recorded from a consistent camera viewpoint to ensure interaction clarity. By fine-tuning a Video Diffusion Model (VDM) on TASTE-Rob, we achieve realistic object interactions, though we observed occasional inconsistencies in hand grasping postures. To enhance realism, we introduce a three-stage pose-refinement pipeline that improves hand posture accuracy in generated videos. Our curated dataset, coupled with the specialized pose-refinement framework, provides notable performance gains in generating high-quality, task-oriented hand-object interaction videos, resulting in achieving superior generalizable robotic manipulation. To foster further advancements in the field, TASTE-Rob dataset and source code will be made publicly available on our website https://taste-rob.github.io.", + "bbox": [ + 88, + 339, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 744, + 223, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robotic manipulation plays an essential role in daily life, assisting with tasks like object handling, liquid pouring, surface cleaning, and drawer operation. Powered by the success of Imitation Learning (IL) [8-10, 15, 20, 22, 42, 48,", + "bbox": [ + 89, + 768, + 482, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "use a spatula to scoop the tomato onto a plate", + "bbox": [ + 576, + 308, + 843, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97dcc98e83ad6eb25463cbe23ae489411f2275715c320780124030a1392cf437.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 321, + 625, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a39ee47f7ff47b4d9b324be820685ca9bd66ce0e13ff4ca2603a7cea23bd4210.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 321, + 901, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "use left hand bring the pink glass to the table, and right hand put the white remote control to the Yellow Book", + "bbox": [ + 540, + 376, + 880, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/def49f50066aa83f2e7ab2a4dd36560e2bda336a9be2b614573ceced85f3c404.jpg", + "image_caption": [ + "Environment" + ], + "image_footnote": [], + "bbox": [ + 517, + 404, + 625, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/825e6e5a3009dbeb3fa6eed0f13abf8de78cf685f146f8a35d25906e50b06293.jpg", + "image_caption": [ + "Generated Videos", + "Figure 1. TASTE-Rob generates high-quality task-oriented hand-object interaction videos, enabling robust generalization in robotic manipulation." + ], + "image_footnote": [], + "bbox": [ + 625, + 404, + 901, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "56, 57, 64], current robots can imitate actions from video demonstrations but require a nearly identical environment in the recorded video, limiting their generalization to novel scenes where factors like the position or movement direction of the manipulated object differ.", + "bbox": [ + 511, + 534, + 906, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve the issue of limited generalization, recent research combines IL with generative models [2, 3, 5-7, 12, 25, 47, 58], like video diffusion models (VDMs), to create adaptable demonstrations and enable more generalizable robotic manipulation [12, 25, 47]. Specifically, these works develop VDMs to generate robot-object interaction videos as demonstrations for IL. While robot-object interaction VDMs enhance generalization performance, they are constrained by limited robot datasets [4, 8, 21, 24, 51]. A scalable alternative is generating hand-object interaction (HOI) videos, which leverage extensive human manipulation video datasets. Although hand-object interaction differs from robotic motion, advanced policy models [1, 5, 9, 22, 48, 57, 65] can bridge this gap effectively, achieving remarkable performance through IL. However, even if existing powerful VDMs are able to generate remarkable human videos, as shown in Figure 7, they struggle with generating high-quality, task-oriented hand-object interaction videos with accurate task understanding.", + "bbox": [ + 511, + 613, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 838, + 218, + 849 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Work done as a visiting researcher at CUHKSZ.", + "bbox": [ + 109, + 851, + 369, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This paper solely reflects the author's personal research and is not associated with the author's affiliated institution.", + "bbox": [ + 91, + 864, + 482, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\S}$ Corresponding author: hanxiaoguang@cuhk.edu.cn.", + "bbox": [ + 109, + 887, + 390, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27683", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To enhance video generation quality, Ego4D [16], a large-scale data containing 83,647 ego-centric hand-object interaction videos, has proven useful in robotic manipulation tasks [23, 32, 53]. However, Ego4D presents two notable limitations for our purpose, as shown in Appendix 8. Firstly, the camera perspective in each clip is inconsistent, whereas video demonstrations for IL are typically recorded from a fixed viewpoint. Secondly, since these clips are extracted from longer videos that encompass multiple tasks, some clips contain partial interactions that overlap with others, resulting in misalignment between video clips and language instructions. Both limitations negatively impact the quality of video generation.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these above limitations, we introduce TASTE-Rob, a pioneering large-scale dataset designed to enhance the generation quality of task-oriented hand-object interaction videos. TASTE-Rob contains 100,856 egocentric videos with well-aligned language instructions. Unlike Ego4D, TASTE-Rob is crafted to ensure clarity and consistency - each video captures a single, complete interaction corresponding to its instruction and is recorded independently with a fixed camera viewpoint, mirroring the conditions needed for reliable imitation learning demonstrations. By resolving these key issues, TASTE-Rob provides a robust foundation for advancing research in video generation, imitation learning, and related fields.", + "bbox": [ + 89, + 289, + 480, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After fine-tuning a VDM on our curated dataset, we achieve high-quality video generation of task-oriented hand-object interactions. While the model demonstrates proficiency in identifying target objects and determining appropriate coarse interactions, it exhibits inconsistencies in hand poses, with the grip changing unnaturally during manipulation. This limitation presents a critical challenge for robotic imitation learning where precise and stable hand-object interactions are essential. To address this issue, we further propose a dedicated three-stage pipeline for pose refinement. First, we generate a coarse manipulation video conditioned solely on language instructions and environment images. Second, we refine the hand pose sequences using a Motion Diffusion Model (MDM) [45], ensuring realistic grip poses. Finally, we regenerate the videos by incorporating the revised pose sequences as an additional conditioning factor. This pipeline enhances grasping realism in generated videos, enabling robots to execute novel tasks with greater accuracy and adaptability in unseen scenarios. Extensive experiments demonstrate that our proposed TASTE-Rob dataset, combined with the pose-refinement pipeline, achieves significant performance gains, surpassing state-of-the-art methods in the quality of generated hand-object interaction videos and achieving accurate robotic manipulation.", + "bbox": [ + 89, + 488, + 482, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 868, + 434, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce TASTE-Rob, a large-scale dataset contain", + "bbox": [ + 89, + 885, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing 100,856 ego-centric hand-object interaction videos, each specially tailored for training high-quality video generation model for robot imitation learning.", + "bbox": [ + 526, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a three-stage pose-refinement pipeline that significantly enhances the fidelity of the hand posture in the generated videos.", + "- We set a new state of the art in generating task-oriented hand-object interaction videos, which further boosts the performance of robot manipulation." + ], + "bbox": [ + 511, + 136, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 242, + 661, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imitation Learning from video demonstrations. IL enables robots to learn from expert [8, 10, 15, 42, 64] or human [9, 20, 22, 48, 56, 57] video demonstrations by training policies to directly replicate expert actions through supervised learning. While these methods require video demonstrations from environments identical to the robot execution setup, they exhibit limited generalization to novel scenes and tasks. Alternatively, existing works address this challenge by generating demonstrations, such as videos [5, 23, 32, 53], optical flows [52, 58], and images [7], successfully achieving better generalization. Among existing approaches, video demonstration generation represents the most straightforward solution. Yet, robot manipulation video generation [23, 32, 53] is constrained by limited robot datasets, while human hand video generation [5] is hindered by the absence of effective HOI video generative models. In this work, we explore enhancing HOI video generation quality to facilitate robot manipulation.", + "bbox": [ + 511, + 268, + 906, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ego-Centric HOI Video datasets. To enhance video quality, we propose a specialized HOI video generation model, which requires a large-scale, high-quality egocentric dataset. Existing ego-centric HOI datasets have various limitations: insufficient resolution [27, 28, 30, 43], limited scale [13, 27, 28, 30, 35, 43], and restricted scene diversity [11, 13, 28]. While Ego4D [16] provides diverse, large-scale HOI videos, its varied camera perspectives and imprecise action-language alignment pose challenges for IL video demonstrations. Therefore, we introduce TASTE-Rob, containing 100,856 videos specifically designed for our task.", + "bbox": [ + 511, + 561, + 905, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models for Video Generation. Based on the success of VDMs [18], researchers have achieved high-quality video generation with various conditioning inputs: text [33, 41, 61, 62] and images [26, 38, 54, 55, 60]. Imaged-to-video generation aligns well with our task. While these models excel at producing realistic colors and temporal coherence, they struggle with significant motion variations. Motion diffusion models [19, 36, 45, 50, 59] address this limitation by leveraging high-quality human body models [39] to generate motion control parameters. In this work,", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27684", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/184ac5c94f17f029dbb81993dcf89e3c2eddd42b8e3713b28f99b7bf814d9cc1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Orientation(°)0-9090-180180-270270-360
Proportion(%)33.8247.79.598.89
", + "bbox": [ + 98, + 88, + 475, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Palm orientation distribution of grasping poses. We analyze the palm orientation distribution of grasping poses across TASTE-Rob. Specifically, the degree of orientation denotes the angle between the palm normal direction and the positive $x$ -axis of the frame plane.", + "bbox": [ + 89, + 145, + 482, + 214 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/de7d7110624790b5927a665de9760e6f7f1d59431ea1cb2b5cf21655e50f50c0.jpg", + "image_caption": [ + "Figure 2. Distribution of environment and HOI task. In TASTE-Rob, we collect ego-centric task-oriented HOI videos with diverse tasks under diverse environments." + ], + "image_footnote": [], + "bbox": [ + 112, + 215, + 264, + 304 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ee66407e6383468d8fd75c3fcf3259ead2108771721cf1dfcff342d576d26706.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 217, + 460, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we combine the advantages of VDM and MDM through our pose-refinement pipeline to generate HOI videos with realistic grasping poses.", + "bbox": [ + 89, + 363, + 483, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Building TASTE-Rob Dataset", + "text_level": 1, + "bbox": [ + 89, + 421, + 364, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We collect a massive and diverse ego-centric task-oriented HOI video dataset TASTE-Rob, which includes 100,856 pairs of video and their corresponding language task instruction. To serve for HOI video generation, TASTE-Rob is required to achieve following goals: 1) Each video is recorded with a static camera perspective and contains a single action that closely aligns with the task instruction. 2) It encompasses diversified environments and tasks. 3) It features a variety of hand poses across different HOI scenarios.", + "bbox": [ + 89, + 446, + 483, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Collection Strategy and Camera Setting", + "text_level": 1, + "bbox": [ + 89, + 590, + 431, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve the first goal, we utilize multiple cameras equipped with a wide-angle lens capable of capturing 1080p ego-centric videos. We make the following improvements during each recording process.", + "bbox": [ + 89, + 613, + 482, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, since our data collection aims to generate task-oriented HOI videos for demonstrations in IL, where demonstrations are typically recorded from fixed camera viewpoints for effective robot imitation learning, we ensure that no camera perspective variation occurs during recording. In addition, as shown in Figure 12, we align the camera perspective specifically to match the head-mounted camera setup of Ego4D [16], ensuring consistency with the egocentric perspective.", + "bbox": [ + 89, + 674, + 482, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our second objective is to ensure precise alignment between language task instructions and video actions in TASTE-Rob, a crucial aspect for maintaining the action integrity in generated HOI videos. Unlike Ego4D [16], which captures extended recordings of daily activities via head-mounted cameras and later segments them into shorter clips,", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2167fe86bb638a720d08d3f5e8a321fe2a209f504cb689a89475e6be114d7fe9.jpg", + "image_caption": [ + "Figure 3. Inter-finger angle distribution of human hands. We analyze the distribution of inter-finger angles, specifically between the thumb-index and index-middle finger pairs, across TASTE-Rob. Specifically, higher degrees mean higher inter-finger angles." + ], + "image_footnote": [], + "bbox": [ + 555, + 89, + 864, + 247 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2dc4741158b7f212cb04befe3fccc471af10e82c5a2bda02673b2feedb72f462.jpg", + "image_caption": [ + "Figure 4. Finger curvature distribution of human hands. We analyze the distribution of finger curvatures, specifically that of thumb, index finger and middle finger, across TASTE-Rob. Specifically, higher degrees mean higher curvature." + ], + "image_footnote": [], + "bbox": [ + 553, + 323, + 864, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we employ a more controlled collection protocol: 1) Each video is strictly limited to under eight seconds in duration and captures a single action. 2) Collectors follow a structured recording process: they press the \"start recording\" button, execute the specified HOI task according to the provided instruction, and stop recording upon task completion. This approach ensures a precise correspondence between actions and task instructions.", + "bbox": [ + 511, + 564, + 906, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Diversity", + "text_level": 1, + "bbox": [ + 513, + 708, + 661, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distribution of environment and task. To promote broad generalization, videos in TASTE-Rob are recorded across diverse environments and cover a wide range of HOI tasks. As shown in Figure 2, environments include locations such as kitchens, bedrooms, dining tables, office tables, etc. Collectors are required to interact with various commonly used objects, and perform tasks including picking, placing, pushing, pouring, etc. To further ensure task diversity, we consider different patterns of hand usage. Specifically, TASTE-Rob comprises 75,389 single-hand task videos and 25,467 double-hand task videos.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27685", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/8a408fa3747df624d22523129383b91ace99a2bacd496922c679f771965b2270.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetLanguage InstructionsEnvironmentData Statistics
Perfect Action AlignmentDeterminerCameraSettingFramesClipsResolution
Disney [13]XXDynamicGarden11M88001080p
ADL [35]XXDynamicDiverse1M436960p
Charades-Ego [43]XXDynamicDiverse5.8M7860720p
HOI4D [29]✓XStaticDiverse4M24661080p
UT Ego [27, 30]XXDynamicDiverse0.9M80320p
EGTEA Gaze+ [28]✓XDynamicKitchen2.5M10,000480p
EPIC-KITCHENS [11]✓XDynamicKitchen6.8M35,0001080p
Ego4D [16]XXDynamicDiverse19.2M83,6471080p
Ours✓✓StaticDiverse9M100,8561080p
", + "bbox": [ + 120, + 88, + 885, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Dataset comparison with existing ego-centric HOI video datasets. TASTE-Rob, specifically designed for generating task-oriented HOI videos, provides a diverse collection of ego-centric HOI videos with several distinctive features: static camera perspectives, precise language-action alignment, and rich object descriptors (Please see Appendix 8 for details). In the above table, we compare these key characteristics: Perfect Action Alignment (ensuring all actions in the video strictly correspond to language task descriptions), Descriptive Determiners (whether object descriptions exist), Camera Setting (fixed or dynamic camera perspective), and Environment Setting (recording locations). By the way, Even if Ego4D [16] offers an impressive 3000-hours video data, it is designed for various general tasks, and the portion applicable to HOI video generation is significantly smaller, as presented in the above table, which is only 83,647 clips. Cells highlighted in $\\square$ denotes the dataset with best feature in each column.", + "bbox": [ + 89, + 294, + 906, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distribution of Grasping Hands. To ensure a wide variety of hand poses, we consider two primary factors: diverse palm orientations (global pose) and varied grasping poses (detailed pose). To demonstrate the diversity of hand poses, we utilize HaMeR [34] to extract hand pose parameters and analyze the distribution based on these parameters (see further details in Appendix 9).", + "bbox": [ + 89, + 414, + 480, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Table 1, we analyze the distribution of palm orientations during HOI interaction across TASTE-Rob. The analysis reveals the following insights: 1) Hand poses with palms facing down $(0^{\\circ} - 180^{\\circ})$ are the most common, as this orientation is practical for grasping objects. 2) Hand poses with palms facing left $(90^{\\circ} - 270^{\\circ})$ occur slightly more frequently than those facing right, likely because all collectors are right-handed and naturally prefer using their right hand for object manipulation.", + "bbox": [ + 88, + 520, + 482, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also provide the analysis of hand grasping pose distribution in Figures 3 and 4. Given the dominant roles of the thumb, index finger, and middle finger in HOI, we focus on examining both the inter-finger angles between these fingers and their individual curvature distributions. The analysis in Figure 3 shows a broad distribution of inter-finger angles, indicating diverse hand orientations. Figure 4 reveals two key findings: 1) The curvature distributions of the index and middle fingers exhibit similar patterns, reflecting their synchronized bending during HOI actions. 2) Our dataset captures a broad spectrum of grasping poses, driven by the variety of objects being manipulated.", + "bbox": [ + 89, + 657, + 482, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Comparisons with Other Datasets", + "text_level": 1, + "bbox": [ + 89, + 848, + 387, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Table 2, we provide comparison between TASTE-Rob and existing ego-centric HOI video datasets.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TASTE-Rob is the first video dataset specifically designed for task-oriented HOI video generation, and it can also serve as a valuable resource for demonstrations in IL. Given that IL video demonstrations are recorded from a fixed camera perspective and include only a single action that aligns with the task instruction, we collect HOI videos under the same settings, distinguishing TASTE-Rob from other datasets. In addition, to improve comprehension of target objects, we incorporate diverse determiners of objects in language task instructions (see Appendix 8 for further details). With TASTE-Rob, we are able to generate high-quality HOI video demonstrations to achieve IL.", + "bbox": [ + 511, + 414, + 906, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 614, + 604, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an environment image and a task description, the generated task-oriented HOI videos are required to satisfy: 1) Accurate Task Understanding: correctly identifying which object to manipulate and how to manipulate it. 2) Feasible HOI: maintaining consistent hand grasping poses throughout the manipulation.", + "bbox": [ + 511, + 642, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in **Stage I** area of Figure 5, while videos generated by a single VDM $(\\hat{v}_c)$ demonstrate accurate task understanding, exhibit limited fidelity in maintaining consistent grasping poses. To fulfill both requirements, we propose a three-stage pose-refinement pipeline, outlined in Figure 5: **Stage I:** employing a learnable Image-to-Video (I2V) diffusion model to generate a coarse HOI video that meets \"Accurate Task Understanding\". **Stage II:** extracting a hand pose sequence from this coarse video and refining it using a learnable Motion Diffusion Model (MDM) [45]. **Stage III:** using the refined hand pose sequence to generate a high", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27686", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3e9f25d98a58220e4b327640bd04b9bb0f9cab81d6ca270f5af51c238685cba3.jpg", + "image_caption": [ + "Figure 5. Overview of our proposed three-stage pose-refinement pipeline. Stage I: our learnable I2V latent diffusion model $\\mathcal{V}$ generates a coarse HOI video $\\hat{\\pmb{v}}_c$ . Stage II: extracting hand pose sequence $\\pmb{p}_c$ from $\\hat{\\pmb{v}}_c$ , and our learnable MDM $\\mathcal{M}$ refines it into $\\hat{\\pmb{p}}$ . Besides, we also obtain the corresponding visualized hand pose images $\\pmb{s}$ of $\\hat{\\pmb{p}}$ . Stage III: incorporating this additional pose condition into $\\mathcal{V}$ to generate a fine HOI video $\\hat{\\pmb{v}}$ ." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 459, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "fidelity HOI video that satisfies both requirements.", + "bbox": [ + 89, + 353, + 426, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Preliminary: Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 378, + 364, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion models [17, 44] work by learning to reverse a noising process. They firstly transform data samples $\\boldsymbol{x}_0 \\sim p_{data}(\\boldsymbol{x})$ into Gaussian noise $\\boldsymbol{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ , then learn to reconstruct the original data by denoising $\\boldsymbol{x}_T$ with multiple steps. With the following objective:", + "bbox": [ + 89, + 402, + 483, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {t, \\boldsymbol {x}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 491, + 482, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the model $\\epsilon_{\\theta}(\\pmb{x}_t,t)$ learns to predict single-step noise at each timestep $t$ , where $\\epsilon$ is the sampled ground truth noise and $\\theta$ indicates the learnable network parameters. After training, the model generates $\\hat{\\pmb{x}}_0$ through a $T$ -step denoising process from a random Gaussian noise $\\pmb{x}_T$ .", + "bbox": [ + 89, + 526, + 483, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Generation. In this work, we explore HOI video generation based on DynamiCrafter [54], a powerful I2V latent diffusion model. Suppose $\\mathcal{T}$ , $i\\in \\mathbb{R}^{3\\times H\\times W}$ , and $\\pmb {v}\\in \\mathbb{R}^{L\\times 3\\times H\\times W}$ denote task language descriptions, environment images, and ground truth video frames, respectively. DynamiCrafter learns the denoising process in a compact latent space: $\\pmb{v}$ is encoded into a compact latent space through an encoder $\\mathcal{E}$ , yielding latent representations $\\pmb {z} = \\mathcal{E}(\\pmb {v})\\in \\mathbb{R}^{L\\times C\\times h\\times w}$ , and decoded via the decoder $\\mathcal{D}$ . Within this latent space, the model performs both forward diffusion and guided denoising processes conditioned on $\\mathcal{T}$ and $\\pmb{i}$ . Specifically, $\\pmb{z}_0\\sim p_{data}(\\pmb {z})$ and $\\pmb{z}_{T}\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})$ . The optimization objective 1 is updated as:", + "bbox": [ + 89, + 625, + 483, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {V}} \\mathbb {E} _ {t, \\boldsymbol {z}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, \\mathcal {T}, \\boldsymbol {i}, t, f p s\\right) \\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 834, + 482, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $fps$ is the FPS control introduced in [54] and $\\theta_V$ denotes the learnable parameters of DynamiCrafter.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4bfb0fb6f03d3da2d37e2abc1c33535b4aae39434300760d07b57e1239788b64.jpg", + "image_caption": [ + "Figure 6. overview of our learnable Image-to-Hand-Params MDM. Based on MDM [45], we extend the framework by incorporating environment information through an additional image condition branch. The model takes a noised hand pose sequence $\\pmb{p}_t^{1:L_p}$ , $t$ itself, a CLIP [37] text embedding $\\pmb{c}_{txt}$ of $\\mathcal{T}$ and a CLIP image embedding $\\pmb{c}_{img}$ of $\\pmb{i}$ as input." + ], + "image_footnote": [], + "bbox": [ + 555, + 87, + 866, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Sequence Generation. MDM [45], utilizing a unique transformer encoder-only architecture, demonstrates remarkable performance in generating human motion sequences. Instead of predicting single-step noise, MDM directly generates clean motion sequences. Let $\\pmb{p}$ denotes motion sequences and $\\mathcal{M}$ denotes MDM network. Here, $\\pmb{p}_0 \\sim p_{data}(\\pmb{p})$ and $\\pmb{p}_T \\sim \\mathcal{N}(\\pmb{0},\\mathbf{I})$ . The optimization objective 1 is modified as:", + "bbox": [ + 511, + 333, + 906, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {M}} \\mathbb {E} _ {t, \\boldsymbol {p}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\boldsymbol {p} _ {0} - \\mathcal {M} (\\boldsymbol {p} _ {t}, \\mathcal {T}, t) \\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 460, + 906, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{M}$ denotes the trainable parameters of MDM. After training, MDM generates the final clean motion sequence $\\hat{\\pmb{p}}_0$ through a $T$ -step denoising process. However, during each step, instead of directly generating $\\pmb{p}_{t-1}$ through single-step denoising, MDM firstly predicts the clean motion sequence $\\hat{\\pmb{p}}_{0,t}$ from $\\pmb{p}_t$ with:", + "bbox": [ + 511, + 489, + 905, + 582 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} _ {0, t} = \\mathcal {M} (\\boldsymbol {p} _ {t}, \\mathcal {T}, t), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 589, + 906, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and then reintroduces noises to obtain $\\pmb{p}_{t-1}$ . Through repeating this process for $T$ times, MDM generates the final clean motion sequence $\\hat{\\pmb{p}}_{0,0}$ , denoted as $\\hat{\\pmb{p}}_0$ .", + "bbox": [ + 511, + 612, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Stage I: A Coarse Action Planner", + "text_level": 1, + "bbox": [ + 511, + 664, + 807, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our learnable coarse action planner is designed to generate a coarse HOI video $\\hat{\\pmb{v}}_c\\in \\mathbb{R}^{L\\times 3\\times H\\times W}$ conditioned on task description $\\mathcal{T}$ and environment image $\\pmb{i}$ . Specifically, we fine-tune DynamiCrafter [54] on TASTE-Rob to serve as the coarse action planner, denoted as $\\nu$ .", + "bbox": [ + 511, + 686, + 905, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. During fine-tuning, we adopt the training strategy similar to that of DynamiCrafter [54]. To leverage DynamiCrafter's robust temporal capabilities while adapting it to our specific HOI video generation, we only finetune its image context projector and spatial layers in its denoising U-Net. The training objective remains consistent with Equation 2, with trainable parameters $\\theta_V$ representing parameters of the image context projector and spatial layers.", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27687", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hand-Object Interacting Inconsistency Problem. Our generated coarse HOI videos exhibit accurate task understanding, such as identifying the object to manipulate and determining its target position. However, as shown in $\\pmb{p}_c$ of Figure 5, the grasping poses exhibit temporal inconsistencies during manipulation, indicating a lack of motion coherence. Specifically, these inconsistencies refer to undesirable variations in grasping poses over time, manifesting as unnatural changes in hand posture when the relative position between the hands and the grasped object should ideally remain stable. As shown in $\\pmb{p}_c$ of Figure 5, the green hand pose demonstrates a pinching gesture, which is inconsistent with the grasping posture of the yellow hand pose and unsuitable for the target object being manipulated.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Stage II: Revising Hand Pose Sequences", + "text_level": 1, + "bbox": [ + 89, + 311, + 433, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address HOI inconsistencies, we train an Image-to-Hand-Params MDM $\\mathcal{M}$ . This model refines hand pose sequences $\\pmb{p}_c\\in \\mathbb{R}^{L_p\\times N_h\\times 2}$ extracted from the coarse video $\\hat{\\boldsymbol{v}}_c$ . Specifically, we define $\\pmb{p}_c$ as the normalized coordinates of hand key-point sequences, where $L_{p}$ denotes the length of sequence and $N_{h}$ is the number of hand key-points.", + "bbox": [ + 89, + 334, + 483, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. Our learnable $\\mathcal{M}$ is designed to take the task description $\\mathcal{T}$ and environment image $i$ as input to predict refined human hand pose sequences. To achieve this, we extend the original MDM [45] framework by incorporating environmental information through an additional image branch. As shown in Figure 6, similar to the text condition branch of [45], the added image branch integrates CLIP [37] class features of environment image $i$ . Consequently, the optimization objective 3 is updated as follows:", + "bbox": [ + 89, + 445, + 483, + 583 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {M}} \\mathbb {E} _ {t, \\boldsymbol {p}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\boldsymbol {p} _ {0} - \\mathcal {M} (\\boldsymbol {p} _ {t}, \\mathcal {T}, \\boldsymbol {i}, t) \\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 595, + 482, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta_{M}$ denotes the parameters of $\\mathcal{M}$ and $\\epsilon$ denotes the sampled ground truth noise.", + "bbox": [ + 89, + 626, + 482, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When we directly generate final clean pose sequence $\\hat{\\pmb{p}}$ through a $T$ -step denoising process, $\\hat{\\pmb{p}}$ achieves physically plausible hand motions but exhibits limited spatial awareness. Conversely, $\\pmb{p}_c$ exhibits remarkable spatial awareness. To address this limitation, we refine $\\pmb{p}_c$ using $\\mathcal{M}$ rather than generating from Gaussian noise. Specifically, we initialize the denoising process in Equation 4 of $\\mathcal{M}$ with $\\pmb{p}_{0,N_{rv}}$ , setting $\\pmb{p}_c$ as $\\pmb{p}_{0,N_{rv}}$ . Through an $N_{rv}$ -step denoising, we refine $\\pmb{p}_c$ to obtain the final clean hand pose sequence $\\hat{\\pmb{p}}$ which satisfies both spatial accuracy and motion feasibility.", + "bbox": [ + 89, + 657, + 483, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Stage III: Regenerating with Refined Pose", + "text_level": 1, + "bbox": [ + 89, + 818, + 449, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the refined hand pose sequences, we generate the fine-grained HOI videos with the additional pose condition $\\hat{p}$ . Inspired by ToonCrafter [55], we train a frame-independent pose encoder $S$ to control the hand pose in the generated", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "videos. We design $S$ as a frame-wise adapter that adjusts the intermediate features of each frame independently, conditioned on $\\hat{p}$ : $F_{inject}^{i} = S(s^{i},z^{i},t)$ , where $s^i$ is the visualized image sequences of $\\hat{p}$ , shown as Figure 5, and $F_{inject}^{i}$ is processed using a method similar to ControlNet [62].", + "bbox": [ + 511, + 90, + 905, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. During training, we adopt a strategy similar to ToonCrafter [55], where all parameters of $\\mathcal{V}$ are frozen, and only the parameters of $S$ , denoted as $\\eta$ , are trained with the following optimization objective:", + "bbox": [ + 511, + 186, + 906, + 247 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\eta} \\mathbb {E} _ {t, \\boldsymbol {s}, \\mathcal {E} (\\boldsymbol {v}), \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} ^ {S} (\\boldsymbol {z} _ {t}; \\mathcal {T}, \\boldsymbol {i}, \\boldsymbol {s}, t, f p s) \\| _ {2} ^ {2}, (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 256, + 906, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}^{\\mathcal{S}}$ denotes the combination of $\\epsilon_{\\theta}$ and $\\mathcal{S}$ , and $s$ represents the visualized images of hand pose sequences.", + "bbox": [ + 511, + 292, + 905, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we consider our generated fine HOI videos $\\hat{v}$ as the video demonstrations for IL and enable robot manipulation using policy model of Im2Flow2Act [58]. As illustrated in Figure 10, we present the imitation learning results with our generated HOI videos as demonstrations, confirming their effectiveness for enabling robot manipulation.", + "bbox": [ + 511, + 323, + 905, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 426, + 648, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation", + "text_level": 1, + "bbox": [ + 511, + 450, + 671, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, we train our model on the training set of TASTE-Rob dataset. **Stage I:** We fine-tune our coarse action planner based on DynamamiCrafter for 30K steps, with a batch size of 16 and a learning rate of $5 \\times 10^{-5}$ . **Stage II:** We train our MDM for 100K steps, with a batch size of 64 and a learning rate of $1 \\times 10^{-4}$ . **Stage III:** We fine-tune the pose encoder based on SD for 30K steps, with a batch size of 32 and a learning rate of $5 \\times 10^{-5}$ . During inference, we generate videos with a 50-step denoising process and refine pose sequences with $N_{rv}$ as 10 (See more details in Appendix 12).", + "bbox": [ + 511, + 473, + 906, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation", + "text_level": 1, + "bbox": [ + 511, + 648, + 635, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate on two datasets: TASTE-Rob-Test and Mujoco simulation dataset. TASTE-Rob-Test: We split TASTE-Rob dataset into training and test sets, with the test set denoted as TASTE-Rob-Test. For fair evaluation, we select $2\\%$ of samples from each action category as test samples. Mujoco simulation dataset: we randomly select 50 tasks from robot simulation platform provided by Im2Flow2Act [58] for robot manipulation.", + "bbox": [ + 511, + 670, + 905, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, as mentioned in Section 3 and Table 2, other ego-centric video datasets feature different settings from our task, making them unsuitable as evaluation datasets.", + "bbox": [ + 511, + 792, + 905, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We evaluate the performance on three aspects: video generation, grasping pose consistency and robot manipulation. Video Generation: To evaluate the quality of", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27688", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generated videos in both the spatial and temporal domains, we report Fréchet Video Distance (FVD) [46], Kernel Video Distance (KVD) [46] and Perceptual Input Conformity (PIC) [54] as used in DynamiCrafter [54]. Grasping Pose Consistency: We evaluate with three metrics - Grasping Pose Variance (GPV, computed by $\\frac{1}{L'}\\sum_{i=1}^{N}(\\theta_i^p - \\bar{\\theta}^p)^2$ ), Grasp Type Classification Error (GTCE, calculating the classification error of grasping types in generated videos) and Hand Movement Direction Accuracy (HMDA, Measuring the directional consistency of hand movements and their alignment with the intended task trajectory), and details of these metrics are illustrated in Appendix 10. Robot Manipulation: We utilize videos generated on the Mujoco simulation dataset as demonstrations for IL, and deploy the policy model of Im2Flow2Act [58] for robot manipulation. The performance is evaluated by comparing the success rates.", + "bbox": [ + 89, + 90, + 483, + 347 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae7e4363c7697e0f3f20276c03449758023f07c02052f4e6165cbf5ae6df0674.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGeneration Quality
KVD↓FVD↓PIC↑
consistI2V0.2465.770.85
DynamiCrafter0.2162.360.79
Open-Sora Plan0.1850.190.84
CogVideoX0.1648.720.85
TASTE-Rob0.039.430.90
", + "bbox": [ + 142, + 351, + 431, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Baselines and Comparisons.", + "text_level": 1, + "bbox": [ + 89, + 542, + 341, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We select four existing powerful I2V diffusion models - DynamamiCrafter [54], consistI2V [38], Open-Sora Plan [26] and CogVideoX [60], as baselines, and conduct comparative experiments among these baselines and our method. We provide qualitative comparisons of video generation performance on TASTE-Rob-Test and wild environment in Figure 7, and quantitative comparisons on TASTE-Rob-Test in Table 3, which demonstrates superior video quality and better generalization capability of our method. According to above experiments, all existing powerful general VDMs fail to achieve manipulation tasks well, making them unsuitable for generating HOI video demonstrations. Given that the other two evaluation aspects focus on measuring the fine-grained details of generated videos, we omit further comparisons with these baseline methods.", + "bbox": [ + 89, + 565, + 483, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 801, + 243, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TASTE-Rob Dataset. We evaluate the performance improvements achieved by using our TASTE-Rob dataset. To be fair, we compare video generation performance between our coarse action planner (Coarse-TASTE-Rob) and Ego4D-Gen, a comparison baseline fine-tuned from Dy", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e6f345b330f5acb110c29fa48d55ec8985bba41e7033447e3133a39d0c73d5c3.jpg", + "table_caption": [ + "Table 3. Quantitative comparison between TASTE-Rob and baselines. All video generation metrics prove the better generation performance of TASTE-Rob." + ], + "table_footnote": [], + "table_body": "
MethodGeneration Quality
KVD↓FVD↓PIC↑
Ego4D-Gen0.1852.170.77
Coarse-TASTE-Rob0.0410.850.88
", + "bbox": [ + 547, + 88, + 870, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "namiCrafter [54] on 83,647 ego-centric HOI videos from Ego4D [16]. The only difference between them is the training datasets they use. As shown in Figure 8 and Table 4, our coarse action planner achieves superior performance.", + "bbox": [ + 511, + 226, + 906, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b504a9921437d9cee8680c0c21e26ce243d5ae62d3d7cdce3c47643269dfe3ea.jpg", + "table_caption": [ + "Table 4. Quantitative comparison between Ego4D-Gen and Coarse-TASTE-Rob. All video generation metrics prove the better generation performance of Coarse-TASTE-Rob." + ], + "table_footnote": [], + "table_body": "
MetricCoarse-TASTE-RobTASTE-Rob
KVD↓0.040.03
FVD↓10.859.43
PIC↑0.880.90
GPV↓0.280.24
GTCE↓67.8%9.7%
HMDA↓26.4°11.3°
success rate↑84%96%
", + "bbox": [ + 535, + 301, + 880, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparison between TASTE-Rob and Coarse-TASTE-Rob. All metrics prove the better generation performance, grasping pose consistency and more precise robot manipulation of TASTE-Rob.", + "bbox": [ + 511, + 455, + 906, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pose-Refinement Pipeline. We compare Coarse-TASTE-Rob and TASTE-Rob across all three evaluation aspects, as shown in Table 5. At a coarse level, both the video generation metrics in Table 5 and the visual comparisons in Figure 9 demonstrate that our proposed pose-refinement pipeline achieves better HOI video generation. At a fine level, both the grasping pose consistency metrics in Table 5 and Figure 9 further validate that our proposed pose-refinement pipeline effectively addresses the pose inconsistency issue and enables more precise robot manipulation. Besides, as shown in Figure 10 and success rate metrics in Table 5, our proposed pose-refinement pipeline helps the policy model achieve more accurate target placement, leading to improved success rates. In addition, we only do this ablation on our dataset, because videos generated by fine-tuning on existing HOI datasets (e.g., Ego4D [16]) are of low quality so that it's difficult to extract reasonable poses for further refinement.", + "bbox": [ + 511, + 531, + 908, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Denoising Steps of Pose Refinement. We also discuss the selection of $N_{rv}$ during pose refinement. When $N_{rv} = 0$ (no refinement), hand pose sequences exhibit inconsistency issues, while $N_{rv} = 50$ (generating pose sequences from Gaussian noise) results in limited spatial awareness.", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27689", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e23b6c71d74b1c7f2b864d781ca157ac4791aac1653c5be8b741c640fed5f47.jpg", + "image_caption": [ + "Figure 7. Comparisons of video generation performance. TASTE-Rob generates videos with high-quality and accurate action, while other existing powerful general VDMs fail." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 890, + 310 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/334fb23f5d8a6a1a60dd3228b86918901c669c611c0b20987d11d7c6e6ef205f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 375, + 464, + 532 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a224e58c0dfb441bc57fdc33c95f02f94e97a2353396486f75ca237866027d4d.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison between Ego4D-Gen and Coarse-TASTE-Rob. Our TASTE-Rob dataset significantly improves the generalization of video generation models in generating HOI videos with accurate actions.", + "Figure 9. Qualitative comparison between TASTE-Rob and Coarse-TASTE-Rob. With our proposed pose-refinement pipeline, we can generate HOI videos with more accurate and feasible grasping poses." + ], + "image_footnote": [], + "bbox": [ + 109, + 587, + 464, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As $N_{rv}$ increases, we observe improved pose consistency but degraded spatial awareness. As shown in Table 6, when $N_{rv}$ increases, video generation quality first improves and then deteriorates. Based on this trade-off, we finally select the value of $N_{rv}$ as 10 (see more details in Appendix 11).", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/498ccf4bb16d5238faea1a92b42639920f4303fd6ab737ae1ce65fc6400bb325.jpg", + "image_caption": [ + "Figure 10. Comparison of Robot Manipulation Performance. TASTE-Rob generates higher-quality HOI videos, leading to more precise robot manipulation." + ], + "image_footnote": [], + "bbox": [ + 532, + 375, + 883, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 631, + 761, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Summary. In this work, we collect a diverse, large-scale ego-centric HOI dataset and develop a three-stage pose-refinement pipeline to generate high-quality task-oriented HOI videos. We successfully enhance HOI video generation capabilities, thereby improving robot manipulation generalization to novel scenes through imitation learning.", + "bbox": [ + 511, + 656, + 906, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our approach achieves accurate manipulation on unseen scenes and enhances the grasping poses in generated videos. However, the generation quality degrades when objects undergo significant transformations, such as rotation and opening. In our future work, we plan to address this limitation.", + "bbox": [ + 511, + 766, + 908, + 856 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27690", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work was supported in part by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, Guangdong Provincial Outstanding Youth Fund(No. 2023B1515020055), NSFC-61931024, and Shenzhen Science and Technology Program No. JCYJ20220530143604010. It is also partly supported by the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by Key Area R&D Program of Guangdong Province (Grant No. 2018B030338001), by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055).", + "bbox": [ + 89, + 114, + 485, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 369, + 187, + 383 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. 2022. 1", + "[2] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans, 2023. 1", + "[3] Homanga Bharadhwaj, Abhinav Gupta, and Shubham Tulsiani. Visual affordance prediction for guiding robot exploration. IEEE International Conference on Robotics and Automation (ICRA), 2023. 1", + "[4] Homanga Bharadhwaj, Jay Vakil, Mohit Sharma, Abhinav Gupta, Shubham Tulsiani, and Vikash Kumar. Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking, 2023. 1", + "[5] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation, 2024. 1, 2", + "[6] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables generalizable robot manipulation. In European Conference on Computer Vision (ECCV), 2024.", + "[7] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models, 2023. 1, 2", + "[8] Anthony Brohan, Noah Brown, Justice Carbajal, and et al. Rt-1: Robotics transformer for real-world control at scale. In arXiv preprint arXiv:2212.06817, 2022. 1, 2", + "[9] Elliot Chane-Sane, Cordelia Schmid, and Ivan Laptev. Learning video-conditioned policies for unseen manipulation tasks. In ICRA, 2023. 1, 2", + "[10] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In" + ], + "bbox": [ + 94, + 393, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of Robotics: Science and Systems (RSS), 2023. 1, 2", + "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2, 4", + "[12] Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B. Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation, 2023. 1", + "[13] Alircza Fathi, Jessica K. Hodgins, and James M. Rehg. Social interactions: A first-person perspective. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1226-1233, 2012. 2, 4", + "[14] Thomas Feix, Javier Romero, Heinz-Bodo Schmiedmayer, Aaron M. Dollar, and Danica Kragic. The grasp taxonomy of human grasp types. IEEE Transactions on Human-Machine Systems, 46(1):66-77, 2016. 4", + "[15] Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. Conference on Robot Learning (CoRL), 2021. 1, 2", + "[16] Kristen Grauman, Andrew Westbury, and et al. Ego4d: Around the world in 3,000 hours of egocentric video, 2022. 1, 2, 3, 4, 7", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 5", + "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models, 2022. 2", + "[19] Vincent Tao Hu, Wenzhe Yin, Pingchuan Ma, Yunlu Chen, Basura Fernando, Yuki M Asano, Efstratios Gavves, Pascal Mettes, Bjorn Ommer, and Cees G. M. Snoek. Motion flow matching for human motion synthesis and editing, 2023. 2", + "[20] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, Igor Gilitschenski, Yonatan Bisk, and Debidatta Dwibedi. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers, 2024. 1, 2", + "[21] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. BC-z: Zero-shot task generalization with robotic imitation learning. In 5th Annual Conference on Robot Learning, 2021. 1", + "[22] Ananth Jonnavittula, Sagar Parekh, and Dylan P. Losey. View: Visual imitation learning with waypoints, 2024. 1, 2", + "[23] Siddharth Karamcheti, Suraj Nair, Annie S. Chen, Thomas Kollar, Chelsea Finn, Dorsa Sadigh, and Percy Liang. Language-driven representation learning for robotics. In Robotics: Science and Systems (RSS), 2023. 2", + "[24] Alexander Khazatsky, Karl Pertsch, and et al. Droid: A large-scale in-the-wild robot manipulation dataset, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27691", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B. Tenenbaum. Learning to act from actionless videos through dense correspondences, 2023. 1", + "[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2, 7", + "[27] Yong Jae Lee, Joydeep Ghosh, and Kristen Grauman. Discovering important people and objects for egocentric video summarization. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1346-1353, 2012. 2, 4", + "[28] Yin Li, Miao Liu, and James M. Rehg. In the eye of the beholder: Gaze and actions in first person video, 2020. 2, 4", + "[29] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21013-21022, 2022. 4", + "[30] Zheng Lu and Kristen Grauman. Story-driven summarization for egocentric video. In 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 2714-2721, 2013. 2, 4", + "[31] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines, 2019. 4, 5", + "[32] Yecheng Jason Ma, William Liang, Vaidehi Som, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Jayaraman. Liv: Language-image representations and rewards for robotic control, 2023. 2", + "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models, 2023. 2", + "[34] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3D with transformers. In CVPR, 2024. 4, 2, 3", + "[35] Hamed Pirsiavash and Deva Ramanan. Detecting activities of daily living in first-person camera views. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2847-2854, 2012. 2, 4", + "[36] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 2", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 6", + "[38] Weiming Ren, Huan Yang, Ge Zhang, Cong Wei, Xinrun Du, Wenhao Huang, and Wenhu Chen. Consisti2v: Enhancing visual consistency for image-to-video generation, 2024. 2, 7", + "[39] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "together. ACM Transactions on Graphics, 36(6):1-17, 2017. 2", + "[40] Dandan Shan, Jiaqi Geng, Michelle Shu, and David Fouhey. Understanding human hands in contact at internet scale. In CVPR, 2020. 3", + "[41] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning, 2023. 2", + "[42] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Proceedings of the 6th Conference on Robot Learning (CoRL), 2022. 1, 2", + "[43] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos, 2018. 2, 4", + "[44] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. 5", + "[45] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 5, 6", + "[46] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. In DGS@ICLR, 2019. 7", + "[47] Boyang Wang, Nikhil Sridhar, Chao Feng, Mark Van der Merwe, Adam Fishman, Nima Fazeli, and Jeong Joon Park. This&that: Language-gesture controlled video generation for robot planning, 2024. 1", + "[48] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play, 2023. 1, 2", + "[49] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability, 2023. 4", + "[50] Yin Wang, Zhiying Leng, Frederick W. B. Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 21978-21987, 2023. 2", + "[51] Zhiqiang Wang, Hao Zheng, Yunshuang Nie, Wenjun Xu, Qingwei Wang, Hua Ye, Zhe Li, Kaidong Zhang, Xuewen Cheng, Wanxi Dong, Chang Cai, Liang Lin, Feng Zheng, and Xiaodan Liang. All robots in one: A new standard and unified dataset for versatile, general-purpose embodied agents, 2024. 1", + "[52] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning, 2024. 2", + "[53] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. In International Conference on Learning Representations, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27692", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors, 2023. 2, 5, 7, 4", + "[55] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. ACM Transactions on Graphics (TOG), 43(6):1-11, 2024. 2, 6, 5", + "[56] Haoyu Xiong, Quanzhou Li, Yun-Chun Chen, Homanga Bharadhwaj, Samarth Sinha, and Animesh Garg. Learning by watching: Physical imitation of manipulation skills from human videos, 2021. 1, 2", + "[57] Mengda Xu, Zhenjia Xu, Cheng Chi, Manuela Veloso, and Shuran Song. XSkill: Cross embodiment skill discovery. In 7th Annual Conference on Robot Learning, 2023. 1, 2", + "[58] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. In 8th Annual Conference on Robot Learning, 2024. 1, 2, 6, 7", + "[59] Zhao Yang, Bing Su, and Ji-Rong Wen. Synthesizing long-term human motions with diffusion models via coherent sampling, 2023. 2", + "[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer, 2024. 2, 7, 3, 4", + "[61] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models, 2023. 2", + "[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 6", + "[63] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models, 2023. 4", + "[64] Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware, 2023. 1, 2", + "[65] Yifeng Zhu, Arisrei Lim, Peter Stone, and Yuke Zhu. Vision-based manipulation from single human video with open-world object graphs, 2024. 1" + ], + "bbox": [ + 91, + 92, + 480, + 714 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27693", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_model.json b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b17f5875122894ba75f796c34473c4717e180ea5 --- /dev/null +++ b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "TASTE-Rob: Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.203, + 0.943, + 0.224 + ], + "angle": 0, + "content": "Hongxiang Zhao\\(^{1*}\\) Xingchen Liu\\(^{1*†}\\) Mutian Xu\\(^{1}\\) Yiming Hao\\(^{1}\\) Weikai Chen\\(\\ddagger\\) Xiaoguang Han\\(^{1,2§}\\)" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.238, + 0.677, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)SSE, CUHKSZ \\(^{2}\\)FNii, CUHKSZ https://taste-rob.github.io" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.486, + 0.718 + ], + "angle": 0, + "content": "We address key limitations in existing datasets and models for task-oriented hand-object interaction video generation, a critical approach of generating video demonstrations for robotic imitation learning. Current datasets, such as Ego4D [16], often suffer from inconsistent view perspectives and misaligned interactions, leading to reduced video quality and limiting their applicability for precise imitation learning tasks. Towards this end, we introduce TASTE-Rob — a pioneering large-scale dataset of 100,856 ego-centric hand-object interaction videos. Each video is meticulously aligned with language instructions and recorded from a consistent camera viewpoint to ensure interaction clarity. By fine-tuning a Video Diffusion Model (VDM) on TASTE-Rob, we achieve realistic object interactions, though we observed occasional inconsistencies in hand grasping postures. To enhance realism, we introduce a three-stage pose-refinement pipeline that improves hand posture accuracy in generated videos. Our curated dataset, coupled with the specialized pose-refinement framework, provides notable performance gains in generating high-quality, task-oriented hand-object interaction videos, resulting in achieving superior generalizable robotic manipulation. To foster further advancements in the field, TASTE-Rob dataset and source code will be made publicly available on our website https://taste-rob.github.io." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.224, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.832 + ], + "angle": 0, + "content": "Robotic manipulation plays an essential role in daily life, assisting with tasks like object handling, liquid pouring, surface cleaning, and drawer operation. Powered by the success of Imitation Learning (IL) [8-10, 15, 20, 22, 42, 48," + }, + { + "type": "text", + "bbox": [ + 0.578, + 0.309, + 0.844, + 0.322 + ], + "angle": 0, + "content": "use a spatula to scoop the tomato onto a plate" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.323, + 0.626, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.323, + 0.903, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.377, + 0.881, + 0.404 + ], + "angle": 0, + "content": "use left hand bring the pink glass to the table, and right hand put the white remote control to the Yellow Book" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.405, + 0.626, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.459, + 0.613, + 0.47 + ], + "angle": 0, + "content": "Environment" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.405, + 0.902, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.459, + 0.814, + 0.47 + ], + "angle": 0, + "content": "Generated Videos" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Figure 1. TASTE-Rob generates high-quality task-oriented hand-object interaction videos, enabling robust generalization in robotic manipulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.61 + ], + "angle": 0, + "content": "56, 57, 64], current robots can imitate actions from video demonstrations but require a nearly identical environment in the recorded video, limiting their generalization to novel scenes where factors like the position or movement direction of the manipulated object differ." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To solve the issue of limited generalization, recent research combines IL with generative models [2, 3, 5-7, 12, 25, 47, 58], like video diffusion models (VDMs), to create adaptable demonstrations and enable more generalizable robotic manipulation [12, 25, 47]. Specifically, these works develop VDMs to generate robot-object interaction videos as demonstrations for IL. While robot-object interaction VDMs enhance generalization performance, they are constrained by limited robot datasets [4, 8, 21, 24, 51]. A scalable alternative is generating hand-object interaction (HOI) videos, which leverage extensive human manipulation video datasets. Although hand-object interaction differs from robotic motion, advanced policy models [1, 5, 9, 22, 48, 57, 65] can bridge this gap effectively, achieving remarkable performance through IL. However, even if existing powerful VDMs are able to generate remarkable human videos, as shown in Figure 7, they struggle with generating high-quality, task-oriented hand-object interaction videos with accurate task understanding." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.839, + 0.22, + 0.851 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.852, + 0.37, + 0.864 + ], + "angle": 0, + "content": "†Work done as a visiting researcher at CUHKSZ." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.865, + 0.483, + 0.887 + ], + "angle": 0, + "content": "This paper solely reflects the author's personal research and is not associated with the author's affiliated institution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.392, + 0.9 + ], + "angle": 0, + "content": "\\(^{\\S}\\)Corresponding author: hanxiaoguang@cuhk.edu.cn." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.839, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "To enhance video generation quality, Ego4D [16], a large-scale data containing 83,647 ego-centric hand-object interaction videos, has proven useful in robotic manipulation tasks [23, 32, 53]. However, Ego4D presents two notable limitations for our purpose, as shown in Appendix 8. Firstly, the camera perspective in each clip is inconsistent, whereas video demonstrations for IL are typically recorded from a fixed viewpoint. Secondly, since these clips are extracted from longer videos that encompass multiple tasks, some clips contain partial interactions that overlap with others, resulting in misalignment between video clips and language instructions. Both limitations negatively impact the quality of video generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.482, + 0.487 + ], + "angle": 0, + "content": "To address these above limitations, we introduce TASTE-Rob, a pioneering large-scale dataset designed to enhance the generation quality of task-oriented hand-object interaction videos. TASTE-Rob contains 100,856 egocentric videos with well-aligned language instructions. Unlike Ego4D, TASTE-Rob is crafted to ensure clarity and consistency - each video captures a single, complete interaction corresponding to its instruction and is recorded independently with a fixed camera viewpoint, mirroring the conditions needed for reliable imitation learning demonstrations. By resolving these key issues, TASTE-Rob provides a robust foundation for advancing research in video generation, imitation learning, and related fields." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.866 + ], + "angle": 0, + "content": "After fine-tuning a VDM on our curated dataset, we achieve high-quality video generation of task-oriented hand-object interactions. While the model demonstrates proficiency in identifying target objects and determining appropriate coarse interactions, it exhibits inconsistencies in hand poses, with the grip changing unnaturally during manipulation. This limitation presents a critical challenge for robotic imitation learning where precise and stable hand-object interactions are essential. To address this issue, we further propose a dedicated three-stage pipeline for pose refinement. First, we generate a coarse manipulation video conditioned solely on language instructions and environment images. Second, we refine the hand pose sequences using a Motion Diffusion Model (MDM) [45], ensuring realistic grip poses. Finally, we regenerate the videos by incorporating the revised pose sequences as an additional conditioning factor. This pipeline enhances grasping realism in generated videos, enabling robots to execute novel tasks with greater accuracy and adaptability in unseen scenarios. Extensive experiments demonstrate that our proposed TASTE-Rob dataset, combined with the pose-refinement pipeline, achieves significant performance gains, surpassing state-of-the-art methods in the quality of generated hand-object interaction videos and achieving accurate robotic manipulation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.869, + 0.435, + 0.883 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- We introduce TASTE-Rob, a large-scale dataset contain" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "ing 100,856 ego-centric hand-object interaction videos, each specially tailored for training high-quality video generation model for robot imitation learning." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.181 + ], + "angle": 0, + "content": "- We develop a three-stage pose-refinement pipeline that significantly enhances the fidelity of the hand posture in the generated videos." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- We set a new state of the art in generating task-oriented hand-object interaction videos, which further boosts the performance of robot manipulation." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.243, + 0.663, + 0.259 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Imitation Learning from video demonstrations. IL enables robots to learn from expert [8, 10, 15, 42, 64] or human [9, 20, 22, 48, 56, 57] video demonstrations by training policies to directly replicate expert actions through supervised learning. While these methods require video demonstrations from environments identical to the robot execution setup, they exhibit limited generalization to novel scenes and tasks. Alternatively, existing works address this challenge by generating demonstrations, such as videos [5, 23, 32, 53], optical flows [52, 58], and images [7], successfully achieving better generalization. Among existing approaches, video demonstration generation represents the most straightforward solution. Yet, robot manipulation video generation [23, 32, 53] is constrained by limited robot datasets, while human hand video generation [5] is hindered by the absence of effective HOI video generative models. In this work, we explore enhancing HOI video generation quality to facilitate robot manipulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.906, + 0.73 + ], + "angle": 0, + "content": "Ego-Centric HOI Video datasets. To enhance video quality, we propose a specialized HOI video generation model, which requires a large-scale, high-quality egocentric dataset. Existing ego-centric HOI datasets have various limitations: insufficient resolution [27, 28, 30, 43], limited scale [13, 27, 28, 30, 35, 43], and restricted scene diversity [11, 13, 28]. While Ego4D [16] provides diverse, large-scale HOI videos, its varied camera perspectives and imprecise action-language alignment pose challenges for IL video demonstrations. Therefore, we introduce TASTE-Rob, containing 100,856 videos specifically designed for our task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Diffusion Models for Video Generation. Based on the success of VDMs [18], researchers have achieved high-quality video generation with various conditioning inputs: text [33, 41, 61, 62] and images [26, 38, 54, 55, 60]. Imaged-to-video generation aligns well with our task. While these models excel at producing realistic colors and temporal coherence, they struggle with significant motion variations. Motion diffusion models [19, 36, 45, 50, 59] address this limitation by leveraging high-quality human body models [39] to generate motion control parameters. In this work," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27684" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.477, + 0.135 + ], + "angle": 0, + "content": "
Orientation(°)0-9090-180180-270270-360
Proportion(%)33.8247.79.598.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.146, + 0.483, + 0.215 + ], + "angle": 0, + "content": "Table 1. Palm orientation distribution of grasping poses. We analyze the palm orientation distribution of grasping poses across TASTE-Rob. Specifically, the degree of orientation denotes the angle between the palm normal direction and the positive \\( x \\)-axis of the frame plane." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.217, + 0.265, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.218, + 0.461, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Distribution of environment and HOI task. In TASTE-Rob, we collect ego-centric task-oriented HOI videos with diverse tasks under diverse environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.41 + ], + "angle": 0, + "content": "we combine the advantages of VDM and MDM through our pose-refinement pipeline to generate HOI videos with realistic grasping poses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.365, + 0.439 + ], + "angle": 0, + "content": "3. Building TASTE-Rob Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.584 + ], + "angle": 0, + "content": "We collect a massive and diverse ego-centric task-oriented HOI video dataset TASTE-Rob, which includes 100,856 pairs of video and their corresponding language task instruction. To serve for HOI video generation, TASTE-Rob is required to achieve following goals: 1) Each video is recorded with a static camera perspective and contains a single action that closely aligns with the task instruction. 2) It encompasses diversified environments and tasks. 3) It features a variety of hand poses across different HOI scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.433, + 0.609 + ], + "angle": 0, + "content": "3.1. Collection Strategy and Camera Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.675 + ], + "angle": 0, + "content": "To achieve the first goal, we utilize multiple cameras equipped with a wide-angle lens capable of capturing 1080p ego-centric videos. We make the following improvements during each recording process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Firstly, since our data collection aims to generate task-oriented HOI videos for demonstrations in IL, where demonstrations are typically recorded from fixed camera viewpoints for effective robot imitation learning, we ensure that no camera perspective variation occurs during recording. In addition, as shown in Figure 12, we align the camera perspective specifically to match the head-mounted camera setup of Ego4D [16], ensuring consistency with the egocentric perspective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our second objective is to ensure precise alignment between language task instructions and video actions in TASTE-Rob, a crucial aspect for maintaining the action integrity in generated HOI videos. Unlike Ego4D [16], which captures extended recordings of daily activities via head-mounted cameras and later segments them into shorter clips," + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.09, + 0.865, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.26, + 0.906, + 0.316 + ], + "angle": 0, + "content": "Figure 3. Inter-finger angle distribution of human hands. We analyze the distribution of inter-finger angles, specifically between the thumb-index and index-middle finger pairs, across TASTE-Rob. Specifically, higher degrees mean higher inter-finger angles." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.324, + 0.866, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.492, + 0.906, + 0.548 + ], + "angle": 0, + "content": "Figure 4. Finger curvature distribution of human hands. We analyze the distribution of finger curvatures, specifically that of thumb, index finger and middle finger, across TASTE-Rob. Specifically, higher degrees mean higher curvature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.685 + ], + "angle": 0, + "content": "we employ a more controlled collection protocol: 1) Each video is strictly limited to under eight seconds in duration and captures a single action. 2) Collectors follow a structured recording process: they press the \"start recording\" button, execute the specified HOI task according to the provided instruction, and stop recording upon task completion. This approach ensures a precise correspondence between actions and task instructions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.709, + 0.662, + 0.725 + ], + "angle": 0, + "content": "3.2. Data Diversity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Distribution of environment and task. To promote broad generalization, videos in TASTE-Rob are recorded across diverse environments and cover a wide range of HOI tasks. As shown in Figure 2, environments include locations such as kitchens, bedrooms, dining tables, office tables, etc. Collectors are required to interact with various commonly used objects, and perform tasks including picking, placing, pushing, pouring, etc. To further ensure task diversity, we consider different patterns of hand usage. Specifically, TASTE-Rob comprises 75,389 single-hand task videos and 25,467 double-hand task videos." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27685" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.089, + 0.887, + 0.285 + ], + "angle": 0, + "content": "
DatasetLanguage InstructionsEnvironmentData Statistics
Perfect Action AlignmentDeterminerCameraSettingFramesClipsResolution
Disney [13]XXDynamicGarden11M88001080p
ADL [35]XXDynamicDiverse1M436960p
Charades-Ego [43]XXDynamicDiverse5.8M7860720p
HOI4D [29]✓XStaticDiverse4M24661080p
UT Ego [27, 30]XXDynamicDiverse0.9M80320p
EGTEA Gaze+ [28]✓XDynamicKitchen2.5M10,000480p
EPIC-KITCHENS [11]✓XDynamicKitchen6.8M35,0001080p
Ego4D [16]XXDynamicDiverse19.2M83,6471080p
Ours✓✓StaticDiverse9M100,8561080p
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.295, + 0.907, + 0.407 + ], + "angle": 0, + "content": "Table 2. Dataset comparison with existing ego-centric HOI video datasets. TASTE-Rob, specifically designed for generating task-oriented HOI videos, provides a diverse collection of ego-centric HOI videos with several distinctive features: static camera perspectives, precise language-action alignment, and rich object descriptors (Please see Appendix 8 for details). In the above table, we compare these key characteristics: Perfect Action Alignment (ensuring all actions in the video strictly correspond to language task descriptions), Descriptive Determiners (whether object descriptions exist), Camera Setting (fixed or dynamic camera perspective), and Environment Setting (recording locations). By the way, Even if Ego4D [16] offers an impressive 3000-hours video data, it is designed for various general tasks, and the portion applicable to HOI video generation is significantly smaller, as presented in the above table, which is only 83,647 clips. Cells highlighted in \\(\\square\\) denotes the dataset with best feature in each column." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.482, + 0.52 + ], + "angle": 0, + "content": "Distribution of Grasping Hands. To ensure a wide variety of hand poses, we consider two primary factors: diverse palm orientations (global pose) and varied grasping poses (detailed pose). To demonstrate the diversity of hand poses, we utilize HaMeR [34] to extract hand pose parameters and analyze the distribution based on these parameters (see further details in Appendix 9)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.521, + 0.483, + 0.656 + ], + "angle": 0, + "content": "As shown in Table 1, we analyze the distribution of palm orientations during HOI interaction across TASTE-Rob. The analysis reveals the following insights: 1) Hand poses with palms facing down \\((0^{\\circ} - 180^{\\circ})\\) are the most common, as this orientation is practical for grasping objects. 2) Hand poses with palms facing left \\((90^{\\circ} - 270^{\\circ})\\) occur slightly more frequently than those facing right, likely because all collectors are right-handed and naturally prefer using their right hand for object manipulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.839 + ], + "angle": 0, + "content": "We also provide the analysis of hand grasping pose distribution in Figures 3 and 4. Given the dominant roles of the thumb, index finger, and middle finger in HOI, we focus on examining both the inter-finger angles between these fingers and their individual curvature distributions. The analysis in Figure 3 shows a broad distribution of inter-finger angles, indicating diverse hand orientations. Figure 4 reveals two key findings: 1) The curvature distributions of the index and middle fingers exhibit similar patterns, reflecting their synchronized bending during HOI actions. 2) Our dataset captures a broad spectrum of grasping poses, driven by the variety of objects being manipulated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.388, + 0.865 + ], + "angle": 0, + "content": "3.3. Comparisons with Other Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 2, we provide comparison between TASTE-Rob and existing ego-centric HOI video datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.596 + ], + "angle": 0, + "content": "TASTE-Rob is the first video dataset specifically designed for task-oriented HOI video generation, and it can also serve as a valuable resource for demonstrations in IL. Given that IL video demonstrations are recorded from a fixed camera perspective and include only a single action that aligns with the task instruction, we collect HOI videos under the same settings, distinguishing TASTE-Rob from other datasets. In addition, to improve comprehension of target objects, we incorporate diverse determiners of objects in language task instructions (see Appendix 8 for further details). With TASTE-Rob, we are able to generate high-quality HOI video demonstrations to achieve IL." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.615, + 0.605, + 0.631 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Given an environment image and a task description, the generated task-oriented HOI videos are required to satisfy: 1) Accurate Task Understanding: correctly identifying which object to manipulate and how to manipulate it. 2) Feasible HOI: maintaining consistent hand grasping poses throughout the manipulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "As illustrated in **Stage I** area of Figure 5, while videos generated by a single VDM \\((\\hat{v}_c)\\) demonstrate accurate task understanding, exhibit limited fidelity in maintaining consistent grasping poses. To fulfill both requirements, we propose a three-stage pose-refinement pipeline, outlined in Figure 5: **Stage I:** employing a learnable Image-to-Video (I2V) diffusion model to generate a coarse HOI video that meets \"Accurate Task Understanding\". **Stage II:** extracting a hand pose sequence from this coarse video and refining it using a learnable Motion Diffusion Model (MDM) [45]. **Stage III:** using the refined hand pose sequence to generate a high" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27686" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.46, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.247, + 0.485, + 0.346 + ], + "angle": 0, + "content": "Figure 5. Overview of our proposed three-stage pose-refinement pipeline. Stage I: our learnable I2V latent diffusion model \\(\\mathcal{V}\\) generates a coarse HOI video \\(\\hat{\\pmb{v}}_c\\). Stage II: extracting hand pose sequence \\(\\pmb{p}_c\\) from \\(\\hat{\\pmb{v}}_c\\), and our learnable MDM \\(\\mathcal{M}\\) refines it into \\(\\hat{\\pmb{p}}\\). Besides, we also obtain the corresponding visualized hand pose images \\(\\pmb{s}\\) of \\(\\hat{\\pmb{p}}\\). Stage III: incorporating this additional pose condition into \\(\\mathcal{V}\\) to generate a fine HOI video \\(\\hat{\\pmb{v}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.354, + 0.427, + 0.37 + ], + "angle": 0, + "content": "fidelity HOI video that satisfies both requirements." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.38, + 0.366, + 0.397 + ], + "angle": 0, + "content": "4.1. Preliminary: Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.48 + ], + "angle": 0, + "content": "Diffusion models [17, 44] work by learning to reverse a noising process. They firstly transform data samples \\( \\boldsymbol{x}_0 \\sim p_{data}(\\boldsymbol{x}) \\) into Gaussian noise \\( \\boldsymbol{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I}) \\), then learn to reconstruct the original data by denoising \\( \\boldsymbol{x}_T \\) with multiple steps. With the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.492, + 0.483, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {t, \\boldsymbol {x}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.603 + ], + "angle": 0, + "content": "the model \\(\\epsilon_{\\theta}(\\pmb{x}_t,t)\\) learns to predict single-step noise at each timestep \\(t\\), where \\(\\epsilon\\) is the sampled ground truth noise and \\(\\theta\\) indicates the learnable network parameters. After training, the model generates \\(\\hat{\\pmb{x}}_0\\) through a \\(T\\)-step denoising process from a random Gaussian noise \\(\\pmb{x}_T\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Video Generation. In this work, we explore HOI video generation based on DynamiCrafter [54], a powerful I2V latent diffusion model. Suppose \\(\\mathcal{T}\\), \\(i\\in \\mathbb{R}^{3\\times H\\times W}\\), and \\(\\pmb {v}\\in \\mathbb{R}^{L\\times 3\\times H\\times W}\\) denote task language descriptions, environment images, and ground truth video frames, respectively. DynamiCrafter learns the denoising process in a compact latent space: \\(\\pmb{v}\\) is encoded into a compact latent space through an encoder \\(\\mathcal{E}\\), yielding latent representations \\(\\pmb {z} = \\mathcal{E}(\\pmb {v})\\in \\mathbb{R}^{L\\times C\\times h\\times w}\\), and decoded via the decoder \\(\\mathcal{D}\\). Within this latent space, the model performs both forward diffusion and guided denoising processes conditioned on \\(\\mathcal{T}\\) and \\(\\pmb{i}\\). Specifically, \\(\\pmb{z}_0\\sim p_{data}(\\pmb {z})\\) and \\(\\pmb{z}_{T}\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})\\). The optimization objective 1 is updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.835, + 0.483, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {V}} \\mathbb {E} _ {t, \\boldsymbol {z}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, \\mathcal {T}, \\boldsymbol {i}, t, f p s\\right) \\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "where \\( fps \\) is the FPS control introduced in [54] and \\( \\theta_V \\) denotes the learnable parameters of DynamiCrafter." + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.088, + 0.867, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.245, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Figure 6. overview of our learnable Image-to-Hand-Params MDM. Based on MDM [45], we extend the framework by incorporating environment information through an additional image condition branch. The model takes a noised hand pose sequence \\(\\pmb{p}_t^{1:L_p}\\), \\(t\\) itself, a CLIP [37] text embedding \\(\\pmb{c}_{txt}\\) of \\(\\mathcal{T}\\) and a CLIP image embedding \\(\\pmb{c}_{img}\\) of \\(\\pmb{i}\\) as input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Motion Sequence Generation. MDM [45], utilizing a unique transformer encoder-only architecture, demonstrates remarkable performance in generating human motion sequences. Instead of predicting single-step noise, MDM directly generates clean motion sequences. Let \\( \\pmb{p} \\) denotes motion sequences and \\( \\mathcal{M} \\) denotes MDM network. Here, \\( \\pmb{p}_0 \\sim p_{data}(\\pmb{p}) \\) and \\( \\pmb{p}_T \\sim \\mathcal{N}(\\pmb{0},\\mathbf{I}) \\). The optimization objective 1 is modified as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.462, + 0.907, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {M}} \\mathbb {E} _ {t, \\boldsymbol {p}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\boldsymbol {p} _ {0} - \\mathcal {M} (\\boldsymbol {p} _ {t}, \\mathcal {T}, t) \\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.583 + ], + "angle": 0, + "content": "where \\(\\theta_{M}\\) denotes the trainable parameters of MDM. After training, MDM generates the final clean motion sequence \\(\\hat{\\pmb{p}}_0\\) through a \\(T\\)-step denoising process. However, during each step, instead of directly generating \\(\\pmb{p}_{t-1}\\) through single-step denoising, MDM firstly predicts the clean motion sequence \\(\\hat{\\pmb{p}}_{0,t}\\) from \\(\\pmb{p}_t\\) with:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.59, + 0.907, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} _ {0, t} = \\mathcal {M} (\\boldsymbol {p} _ {t}, \\mathcal {T}, t), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.659 + ], + "angle": 0, + "content": "and then reintroduces noises to obtain \\( \\pmb{p}_{t-1} \\). Through repeating this process for \\( T \\) times, MDM generates the final clean motion sequence \\( \\hat{\\pmb{p}}_{0,0} \\), denoted as \\( \\hat{\\pmb{p}}_0 \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.808, + 0.682 + ], + "angle": 0, + "content": "4.2. Stage I: A Coarse Action Planner" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Our learnable coarse action planner is designed to generate a coarse HOI video \\(\\hat{\\pmb{v}}_c\\in \\mathbb{R}^{L\\times 3\\times H\\times W}\\) conditioned on task description \\(\\mathcal{T}\\) and environment image \\(\\pmb{i}\\). Specifically, we fine-tune DynamiCrafter [54] on TASTE-Rob to serve as the coarse action planner, denoted as \\(\\nu\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Training. During fine-tuning, we adopt the training strategy similar to that of DynamiCrafter [54]. To leverage DynamiCrafter's robust temporal capabilities while adapting it to our specific HOI video generation, we only finetune its image context projector and spatial layers in its denoising U-Net. The training objective remains consistent with Equation 2, with trainable parameters \\(\\theta_V\\) representing parameters of the image context projector and spatial layers." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27687" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "Hand-Object Interacting Inconsistency Problem. Our generated coarse HOI videos exhibit accurate task understanding, such as identifying the object to manipulate and determining its target position. However, as shown in \\( \\pmb{p}_c \\) of Figure 5, the grasping poses exhibit temporal inconsistencies during manipulation, indicating a lack of motion coherence. Specifically, these inconsistencies refer to undesirable variations in grasping poses over time, manifesting as unnatural changes in hand posture when the relative position between the hands and the grasped object should ideally remain stable. As shown in \\( \\pmb{p}_c \\) of Figure 5, the green hand pose demonstrates a pinching gesture, which is inconsistent with the grasping posture of the yellow hand pose and unsuitable for the target object being manipulated." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.313, + 0.434, + 0.33 + ], + "angle": 0, + "content": "4.3. Stage II: Revising Hand Pose Sequences" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.429 + ], + "angle": 0, + "content": "To address HOI inconsistencies, we train an Image-to-Hand-Params MDM \\(\\mathcal{M}\\). This model refines hand pose sequences \\(\\pmb{p}_c\\in \\mathbb{R}^{L_p\\times N_h\\times 2}\\) extracted from the coarse video \\(\\hat{\\boldsymbol{v}}_c\\). Specifically, we define \\(\\pmb{p}_c\\) as the normalized coordinates of hand key-point sequences, where \\(L_{p}\\) denotes the length of sequence and \\(N_{h}\\) is the number of hand key-points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Training. Our learnable \\(\\mathcal{M}\\) is designed to take the task description \\(\\mathcal{T}\\) and environment image \\(i\\) as input to predict refined human hand pose sequences. To achieve this, we extend the original MDM [45] framework by incorporating environmental information through an additional image branch. As shown in Figure 6, similar to the text condition branch of [45], the added image branch integrates CLIP [37] class features of environment image \\(i\\). Consequently, the optimization objective 3 is updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.597, + 0.483, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {M}} \\mathbb {E} _ {t, \\boldsymbol {p}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\boldsymbol {p} _ {0} - \\mathcal {M} (\\boldsymbol {p} _ {t}, \\mathcal {T}, \\boldsymbol {i}, t) \\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.656 + ], + "angle": 0, + "content": "where \\(\\theta_{M}\\) denotes the parameters of \\(\\mathcal{M}\\) and \\(\\epsilon\\) denotes the sampled ground truth noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.809 + ], + "angle": 0, + "content": "When we directly generate final clean pose sequence \\(\\hat{\\pmb{p}}\\) through a \\(T\\)-step denoising process, \\(\\hat{\\pmb{p}}\\) achieves physically plausible hand motions but exhibits limited spatial awareness. Conversely, \\(\\pmb{p}_c\\) exhibits remarkable spatial awareness. To address this limitation, we refine \\(\\pmb{p}_c\\) using \\(\\mathcal{M}\\) rather than generating from Gaussian noise. Specifically, we initialize the denoising process in Equation 4 of \\(\\mathcal{M}\\) with \\(\\pmb{p}_{0,N_{rv}}\\), setting \\(\\pmb{p}_c\\) as \\(\\pmb{p}_{0,N_{rv}}\\). Through an \\(N_{rv}\\)-step denoising, we refine \\(\\pmb{p}_c\\) to obtain the final clean hand pose sequence \\(\\hat{\\pmb{p}}\\) which satisfies both spatial accuracy and motion feasibility." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.819, + 0.45, + 0.835 + ], + "angle": 0, + "content": "4.4. Stage III: Regenerating with Refined Pose" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "With the refined hand pose sequences, we generate the fine-grained HOI videos with the additional pose condition \\(\\hat{p}\\). Inspired by ToonCrafter [55], we train a frame-independent pose encoder \\(S\\) to control the hand pose in the generated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.17 + ], + "angle": 0, + "content": "videos. We design \\( S \\) as a frame-wise adapter that adjusts the intermediate features of each frame independently, conditioned on \\( \\hat{p} \\): \\( F_{inject}^{i} = S(s^{i},z^{i},t) \\), where \\( s^i \\) is the visualized image sequences of \\( \\hat{p} \\), shown as Figure 5, and \\( F_{inject}^{i} \\) is processed using a method similar to ControlNet [62]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.187, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Training. During training, we adopt a strategy similar to ToonCrafter [55], where all parameters of \\(\\mathcal{V}\\) are frozen, and only the parameters of \\(S\\), denoted as \\(\\eta\\), are trained with the following optimization objective:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.257, + 0.907, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\eta} \\mathbb {E} _ {t, \\boldsymbol {s}, \\mathcal {E} (\\boldsymbol {v}), \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} ^ {S} (\\boldsymbol {z} _ {t}; \\mathcal {T}, \\boldsymbol {i}, \\boldsymbol {s}, t, f p s) \\| _ {2} ^ {2}, (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.906, + 0.323 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}^{\\mathcal{S}}\\) denotes the combination of \\(\\epsilon_{\\theta}\\) and \\(\\mathcal{S}\\), and \\(s\\) represents the visualized images of hand pose sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.906, + 0.415 + ], + "angle": 0, + "content": "Finally, we consider our generated fine HOI videos \\(\\hat{v}\\) as the video demonstrations for IL and enable robot manipulation using policy model of Im2Flow2Act [58]. As illustrated in Figure 10, we present the imitation learning results with our generated HOI videos as demonstrations, confirming their effectiveness for enabling robot manipulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.428, + 0.649, + 0.445 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.452, + 0.673, + 0.468 + ], + "angle": 0, + "content": "5.1. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.641 + ], + "angle": 0, + "content": "During training, we train our model on the training set of TASTE-Rob dataset. **Stage I:** We fine-tune our coarse action planner based on DynamamiCrafter for 30K steps, with a batch size of 16 and a learning rate of \\(5 \\times 10^{-5}\\). **Stage II:** We train our MDM for 100K steps, with a batch size of 64 and a learning rate of \\(1 \\times 10^{-4}\\). **Stage III:** We fine-tune the pose encoder based on SD for 30K steps, with a batch size of 32 and a learning rate of \\(5 \\times 10^{-5}\\). During inference, we generate videos with a 50-step denoising process and refine pose sequences with \\(N_{rv}\\) as 10 (See more details in Appendix 12)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.65, + 0.636, + 0.664 + ], + "angle": 0, + "content": "5.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.906, + 0.792 + ], + "angle": 0, + "content": "Datasets. We evaluate on two datasets: TASTE-Rob-Test and Mujoco simulation dataset. TASTE-Rob-Test: We split TASTE-Rob dataset into training and test sets, with the test set denoted as TASTE-Rob-Test. For fair evaluation, we select \\(2\\%\\) of samples from each action category as test samples. Mujoco simulation dataset: we randomly select 50 tasks from robot simulation platform provided by Im2Flow2Act [58] for robot manipulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.906, + 0.838 + ], + "angle": 0, + "content": "In addition, as mentioned in Section 3 and Table 2, other ego-centric video datasets feature different settings from our task, making them unsuitable as evaluation datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Metrics. We evaluate the performance on three aspects: video generation, grasping pose consistency and robot manipulation. Video Generation: To evaluate the quality of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27688" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.348 + ], + "angle": 0, + "content": "generated videos in both the spatial and temporal domains, we report Fréchet Video Distance (FVD) [46], Kernel Video Distance (KVD) [46] and Perceptual Input Conformity (PIC) [54] as used in DynamiCrafter [54]. Grasping Pose Consistency: We evaluate with three metrics - Grasping Pose Variance (GPV, computed by \\(\\frac{1}{L'}\\sum_{i=1}^{N}(\\theta_i^p - \\bar{\\theta}^p)^2\\)), Grasp Type Classification Error (GTCE, calculating the classification error of grasping types in generated videos) and Hand Movement Direction Accuracy (HMDA, Measuring the directional consistency of hand movements and their alignment with the intended task trajectory), and details of these metrics are illustrated in Appendix 10. Robot Manipulation: We utilize videos generated on the Mujoco simulation dataset as demonstrations for IL, and deploy the policy model of Im2Flow2Act [58] for robot manipulation. The performance is evaluated by comparing the success rates." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.352, + 0.432, + 0.488 + ], + "angle": 0, + "content": "
MethodGeneration Quality
KVD↓FVD↓PIC↑
consistI2V0.2465.770.85
DynamiCrafter0.2162.360.79
Open-Sora Plan0.1850.190.84
CogVideoX0.1648.720.85
TASTE-Rob0.039.430.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison between TASTE-Rob and baselines. All video generation metrics prove the better generation performance of TASTE-Rob." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.544, + 0.342, + 0.56 + ], + "angle": 0, + "content": "5.3. Baselines and Comparisons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.793 + ], + "angle": 0, + "content": "We select four existing powerful I2V diffusion models - DynamamiCrafter [54], consistI2V [38], Open-Sora Plan [26] and CogVideoX [60], as baselines, and conduct comparative experiments among these baselines and our method. We provide qualitative comparisons of video generation performance on TASTE-Rob-Test and wild environment in Figure 7, and quantitative comparisons on TASTE-Rob-Test in Table 3, which demonstrates superior video quality and better generalization capability of our method. According to above experiments, all existing powerful general VDMs fail to achieve manipulation tasks well, making them unsuitable for generating HOI video demonstrations. Given that the other two evaluation aspects focus on measuring the fine-grained details of generated videos, we omit further comparisons with these baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.245, + 0.82 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "TASTE-Rob Dataset. We evaluate the performance improvements achieved by using our TASTE-Rob dataset. To be fair, we compare video generation performance between our coarse action planner (Coarse-TASTE-Rob) and Ego4D-Gen, a comparison baseline fine-tuned from Dy" + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.089, + 0.872, + 0.173 + ], + "angle": 0, + "content": "
MethodGeneration Quality
KVD↓FVD↓PIC↑
Ego4D-Gen0.1852.170.77
Coarse-TASTE-Rob0.0410.850.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.176, + 0.907, + 0.219 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison between Ego4D-Gen and Coarse-TASTE-Rob. All video generation metrics prove the better generation performance of Coarse-TASTE-Rob." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.907, + 0.289 + ], + "angle": 0, + "content": "namiCrafter [54] on 83,647 ego-centric HOI videos from Ego4D [16]. The only difference between them is the training datasets they use. As shown in Figure 8 and Table 4, our coarse action planner achieves superior performance." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.302, + 0.882, + 0.452 + ], + "angle": 0, + "content": "
MetricCoarse-TASTE-RobTASTE-Rob
KVD↓0.040.03
FVD↓10.859.43
PIC↑0.880.90
GPV↓0.280.24
GTCE↓67.8%9.7%
HMDA↓26.4°11.3°
success rate↑84%96%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.456, + 0.907, + 0.512 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison between TASTE-Rob and Coarse-TASTE-Rob. All metrics prove the better generation performance, grasping pose consistency and more precise robot manipulation of TASTE-Rob." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.909, + 0.803 + ], + "angle": 0, + "content": "Pose-Refinement Pipeline. We compare Coarse-TASTE-Rob and TASTE-Rob across all three evaluation aspects, as shown in Table 5. At a coarse level, both the video generation metrics in Table 5 and the visual comparisons in Figure 9 demonstrate that our proposed pose-refinement pipeline achieves better HOI video generation. At a fine level, both the grasping pose consistency metrics in Table 5 and Figure 9 further validate that our proposed pose-refinement pipeline effectively addresses the pose inconsistency issue and enables more precise robot manipulation. Besides, as shown in Figure 10 and success rate metrics in Table 5, our proposed pose-refinement pipeline helps the policy model achieve more accurate target placement, leading to improved success rates. In addition, we only do this ablation on our dataset, because videos generated by fine-tuning on existing HOI datasets (e.g., Ego4D [16]) are of low quality so that it's difficult to extract reasonable poses for further refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Denoising Steps of Pose Refinement. We also discuss the selection of \\( N_{rv} \\) during pose refinement. When \\( N_{rv} = 0 \\) (no refinement), hand pose sequences exhibit inconsistency issues, while \\( N_{rv} = 50 \\) (generating pose sequences from Gaussian noise) results in limited spatial awareness." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "27689" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.891, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.33, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 7. Comparisons of video generation performance. TASTE-Rob generates videos with high-quality and accurate action, while other existing powerful general VDMs fail." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.375, + 0.465, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.588 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison between Ego4D-Gen and Coarse-TASTE-Rob. Our TASTE-Rob dataset significantly improves the generalization of video generation models in generating HOI videos with accurate actions." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.588, + 0.465, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.817 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparison between TASTE-Rob and Coarse-TASTE-Rob. With our proposed pose-refinement pipeline, we can generate HOI videos with more accurate and feasible grasping poses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "As \\( N_{rv} \\) increases, we observe improved pose consistency but degraded spatial awareness. As shown in Table 6, when \\( N_{rv} \\) increases, video generation quality first improves and then deteriorates. Based on this trade-off, we finally select the value of \\( N_{rv} \\) as 10 (see more details in Appendix 11)." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.376, + 0.885, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.566, + 0.907, + 0.607 + ], + "angle": 0, + "content": "Figure 10. Comparison of Robot Manipulation Performance. TASTE-Rob generates higher-quality HOI videos, leading to more precise robot manipulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.762, + 0.648 + ], + "angle": 0, + "content": "6. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Summary. In this work, we collect a diverse, large-scale ego-centric HOI dataset and develop a three-stage pose-refinement pipeline to generate high-quality task-oriented HOI videos. We successfully enhance HOI video generation capabilities, thereby improving robot manipulation generalization to novel scenes through imitation learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.909, + 0.857 + ], + "angle": 0, + "content": "Limitations. Our approach achieves accurate manipulation on unseen scenes and enhances the grasping poses in generated videos. However, the generation quality degrades when objects undergo significant transformations, such as rotation and opening. In our future work, we plan to address this limitation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27690" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.356 + ], + "angle": 0, + "content": "The work was supported in part by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, Guangdong Provincial Outstanding Youth Fund(No. 2023B1515020055), NSFC-61931024, and Shenzhen Science and Technology Program No. JCYJ20220530143604010. It is also partly supported by the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by Key Area R&D Program of Guangdong Province (Grant No. 2018B030338001), by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.37, + 0.188, + 0.385 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.395, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[1] Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.424, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[2] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.52 + ], + "angle": 0, + "content": "[3] Homanga Bharadhwaj, Abhinav Gupta, and Shubham Tulsiani. Visual affordance prediction for guiding robot exploration. IEEE International Conference on Robotics and Automation (ICRA), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[4] Homanga Bharadhwaj, Jay Vakil, Mohit Sharma, Abhinav Gupta, Shubham Tulsiani, and Vikash Kumar. Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[5] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[6] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables generalizable robot manipulation. In European Conference on Computer Vision (ECCV), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[7] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[8] Anthony Brohan, Noah Brown, Justice Carbajal, and et al. Rt-1: Robotics transformer for real-world control at scale. In arXiv preprint arXiv:2212.06817, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[9] Elliot Chane-Sane, Cordelia Schmid, and Ivan Laptev. Learning video-conditioned policies for unseen manipulation tasks. In ICRA, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.395, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Proceedings of Robotics: Science and Systems (RSS), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.209, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[12] Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B. Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.267, + 0.905, + 0.321 + ], + "angle": 0, + "content": "[13] Alircza Fathi, Jessica K. Hodgins, and James M. Rehg. Social interactions: A first-person perspective. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1226-1233, 2012. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.324, + 0.905, + 0.378 + ], + "angle": 0, + "content": "[14] Thomas Feix, Javier Romero, Heinz-Bodo Schmiedmayer, Aaron M. Dollar, and Danica Kragic. The grasp taxonomy of human grasp types. IEEE Transactions on Human-Machine Systems, 46(1):66-77, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.382, + 0.904, + 0.436 + ], + "angle": 0, + "content": "[15] Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. Conference on Robot Learning (CoRL), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.44, + 0.904, + 0.479 + ], + "angle": 0, + "content": "[16] Kristen Grauman, Andrew Westbury, and et al. Ego4d: Around the world in 3,000 hours of egocentric video, 2022. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.483, + 0.904, + 0.511 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.513, + 0.904, + 0.553 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.557, + 0.904, + 0.612 + ], + "angle": 0, + "content": "[19] Vincent Tao Hu, Wenzhe Yin, Pingchuan Ma, Yunlu Chen, Basura Fernando, Yuki M Asano, Efstratios Gavves, Pascal Mettes, Bjorn Ommer, and Cees G. M. Snoek. Motion flow matching for human motion synthesis and editing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.615, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[20] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, Igor Gilitschenski, Yonatan Bisk, and Debidatta Dwibedi. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.7, + 0.905, + 0.768 + ], + "angle": 0, + "content": "[21] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. BC-z: Zero-shot task generalization with robotic imitation learning. In 5th Annual Conference on Robot Learning, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.772, + 0.905, + 0.812 + ], + "angle": 0, + "content": "[22] Ananth Jonnavittula, Sagar Parekh, and Dylan P. Losey. View: Visual imitation learning with waypoints, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[23] Siddharth Karamcheti, Suraj Nair, Annie S. Chen, Thomas Kollar, Chelsea Finn, Dorsa Sadigh, and Percy Liang. Language-driven representation learning for robotics. In Robotics: Science and Systems (RSS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Alexander Khazatsky, Karl Pertsch, and et al. Droid: A large-scale in-the-wild robot manipulation dataset, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27691" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B. Tenenbaum. Learning to act from actionless videos through dense correspondences, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.162 + ], + "angle": 0, + "content": "[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Yong Jae Lee, Joydeep Ghosh, and Kristen Grauman. Discovering important people and objects for egocentric video summarization. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1346-1353, 2012. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[28] Yin Li, Miao Liu, and James M. Rehg. In the eye of the beholder: Gaze and actions in first person video, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[29] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21013-21022, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[30] Zheng Lu and Kristen Grauman. Story-driven summarization for egocentric video. In 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 2714-2721, 2013. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[31] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.482, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[32] Yecheng Jason Ma, William Liang, Vaidehi Som, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Jayaraman. Liv: Language-image representations and rewards for robotic control, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[34] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3D with transformers. In CVPR, 2024. 4, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[35] Hamed Pirsiavash and Deva Ramanan. Detecting activities of daily living in first-person camera views. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2847-2854, 2012. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[36] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[38] Weiming Ren, Huan Yang, Ge Zhang, Cong Wei, Xinrun Du, Wenhao Huang, and Wenhu Chen. Consisti2v: Enhancing visual consistency for image-to-video generation, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.119 + ], + "angle": 0, + "content": "together. ACM Transactions on Graphics, 36(6):1-17, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[40] Dandan Shan, Jiaqi Geng, Michelle Shu, and David Fouhey. Understanding human hands in contact at internet scale. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[41] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[42] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Proceedings of the 6th Conference on Robot Learning (CoRL), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[43] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[44] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[45] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[46] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. In DGS@ICLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[47] Boyang Wang, Nikhil Sridhar, Chao Feng, Mark Van der Merwe, Adam Fishman, Nima Fazeli, and Jeong Joon Park. This&that: Language-gesture controlled video generation for robot planning, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[48] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[49] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[50] Yin Wang, Zhiying Leng, Frederick W. B. Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 21978-21987, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[51] Zhiqiang Wang, Hao Zheng, Yunshuang Nie, Wenjun Xu, Qingwei Wang, Hua Ye, Zhe Li, Kaidong Zhang, Xuewen Cheng, Wanxi Dong, Chang Cai, Liang Lin, Feng Zheng, and Xiaodan Liang. All robots in one: A new standard and unified dataset for versatile, general-purpose embodied agents, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[52] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. In International Conference on Learning Representations, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27692" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[54] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors, 2023. 2, 5, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[55] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. ACM Transactions on Graphics (TOG), 43(6):1-11, 2024. 2, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[56] Haoyu Xiong, Quanzhou Li, Yun-Chun Chen, Homanga Bharadhwaj, Samarth Sinha, and Animesh Garg. Learning by watching: Physical imitation of manipulation skills from human videos, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[57] Mengda Xu, Zhenjia Xu, Cheng Chi, Manuela Veloso, and Shuran Song. XSkill: Cross embodiment skill discovery. In 7th Annual Conference on Robot Learning, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[58] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. In 8th Annual Conference on Robot Learning, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[59] Zhao Yang, Bing Su, and Ji-Rong Wen. Synthesizing long-term human motions with diffusion models via coherent sampling, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer, 2024. 2, 7, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[61] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[63] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[64] Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[65] Yifeng Zhu, Arisrei Lim, Peter Stone, and Yuke Zhu. Vision-based manipulation from single human video with open-world object graphs, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27693" + } + ] +] \ No newline at end of file diff --git a/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_origin.pdf b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5b1cba516a2b049045de2ef48e8e0bd492304b9 --- /dev/null +++ b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/924d8cf7-7b9c-4d7b-995b-3d251b7a8e03_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c98c49adc0d08629138dc12642fe8d6fd24602158bffeddc850bb8ebbba1178c +size 5675650 diff --git a/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/full.md b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ef5ef74d48ec2d067ef7853d9f23345664ef97c5 --- /dev/null +++ b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/full.md @@ -0,0 +1,330 @@ +# TASTE-Rob: Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation + +Hongxiang Zhao $^{1*}$ Xingchen Liu $^{1*†}$ Mutian Xu $^{1}$ Yiming Hao $^{1}$ Weikai Chen $\ddagger$ Xiaoguang Han $^{1,2§}$ + +$^{1}$ SSE, CUHKSZ $^{2}$ FNii, CUHKSZ https://taste-rob.github.io + +# Abstract + +We address key limitations in existing datasets and models for task-oriented hand-object interaction video generation, a critical approach of generating video demonstrations for robotic imitation learning. Current datasets, such as Ego4D [16], often suffer from inconsistent view perspectives and misaligned interactions, leading to reduced video quality and limiting their applicability for precise imitation learning tasks. Towards this end, we introduce TASTE-Rob — a pioneering large-scale dataset of 100,856 ego-centric hand-object interaction videos. Each video is meticulously aligned with language instructions and recorded from a consistent camera viewpoint to ensure interaction clarity. By fine-tuning a Video Diffusion Model (VDM) on TASTE-Rob, we achieve realistic object interactions, though we observed occasional inconsistencies in hand grasping postures. To enhance realism, we introduce a three-stage pose-refinement pipeline that improves hand posture accuracy in generated videos. Our curated dataset, coupled with the specialized pose-refinement framework, provides notable performance gains in generating high-quality, task-oriented hand-object interaction videos, resulting in achieving superior generalizable robotic manipulation. To foster further advancements in the field, TASTE-Rob dataset and source code will be made publicly available on our website https://taste-rob.github.io. + +# 1. Introduction + +Robotic manipulation plays an essential role in daily life, assisting with tasks like object handling, liquid pouring, surface cleaning, and drawer operation. Powered by the success of Imitation Learning (IL) [8-10, 15, 20, 22, 42, 48, + +use a spatula to scoop the tomato onto a plate + +![](images/97dcc98e83ad6eb25463cbe23ae489411f2275715c320780124030a1392cf437.jpg) + +![](images/a39ee47f7ff47b4d9b324be820685ca9bd66ce0e13ff4ca2603a7cea23bd4210.jpg) + +use left hand bring the pink glass to the table, and right hand put the white remote control to the Yellow Book + +![](images/def49f50066aa83f2e7ab2a4dd36560e2bda336a9be2b614573ceced85f3c404.jpg) +Environment + +![](images/825e6e5a3009dbeb3fa6eed0f13abf8de78cf685f146f8a35d25906e50b06293.jpg) +Generated Videos +Figure 1. TASTE-Rob generates high-quality task-oriented hand-object interaction videos, enabling robust generalization in robotic manipulation. + +56, 57, 64], current robots can imitate actions from video demonstrations but require a nearly identical environment in the recorded video, limiting their generalization to novel scenes where factors like the position or movement direction of the manipulated object differ. + +To solve the issue of limited generalization, recent research combines IL with generative models [2, 3, 5-7, 12, 25, 47, 58], like video diffusion models (VDMs), to create adaptable demonstrations and enable more generalizable robotic manipulation [12, 25, 47]. Specifically, these works develop VDMs to generate robot-object interaction videos as demonstrations for IL. While robot-object interaction VDMs enhance generalization performance, they are constrained by limited robot datasets [4, 8, 21, 24, 51]. A scalable alternative is generating hand-object interaction (HOI) videos, which leverage extensive human manipulation video datasets. Although hand-object interaction differs from robotic motion, advanced policy models [1, 5, 9, 22, 48, 57, 65] can bridge this gap effectively, achieving remarkable performance through IL. However, even if existing powerful VDMs are able to generate remarkable human videos, as shown in Figure 7, they struggle with generating high-quality, task-oriented hand-object interaction videos with accurate task understanding. + +To enhance video generation quality, Ego4D [16], a large-scale data containing 83,647 ego-centric hand-object interaction videos, has proven useful in robotic manipulation tasks [23, 32, 53]. However, Ego4D presents two notable limitations for our purpose, as shown in Appendix 8. Firstly, the camera perspective in each clip is inconsistent, whereas video demonstrations for IL are typically recorded from a fixed viewpoint. Secondly, since these clips are extracted from longer videos that encompass multiple tasks, some clips contain partial interactions that overlap with others, resulting in misalignment between video clips and language instructions. Both limitations negatively impact the quality of video generation. + +To address these above limitations, we introduce TASTE-Rob, a pioneering large-scale dataset designed to enhance the generation quality of task-oriented hand-object interaction videos. TASTE-Rob contains 100,856 egocentric videos with well-aligned language instructions. Unlike Ego4D, TASTE-Rob is crafted to ensure clarity and consistency - each video captures a single, complete interaction corresponding to its instruction and is recorded independently with a fixed camera viewpoint, mirroring the conditions needed for reliable imitation learning demonstrations. By resolving these key issues, TASTE-Rob provides a robust foundation for advancing research in video generation, imitation learning, and related fields. + +After fine-tuning a VDM on our curated dataset, we achieve high-quality video generation of task-oriented hand-object interactions. While the model demonstrates proficiency in identifying target objects and determining appropriate coarse interactions, it exhibits inconsistencies in hand poses, with the grip changing unnaturally during manipulation. This limitation presents a critical challenge for robotic imitation learning where precise and stable hand-object interactions are essential. To address this issue, we further propose a dedicated three-stage pipeline for pose refinement. First, we generate a coarse manipulation video conditioned solely on language instructions and environment images. Second, we refine the hand pose sequences using a Motion Diffusion Model (MDM) [45], ensuring realistic grip poses. Finally, we regenerate the videos by incorporating the revised pose sequences as an additional conditioning factor. This pipeline enhances grasping realism in generated videos, enabling robots to execute novel tasks with greater accuracy and adaptability in unseen scenarios. Extensive experiments demonstrate that our proposed TASTE-Rob dataset, combined with the pose-refinement pipeline, achieves significant performance gains, surpassing state-of-the-art methods in the quality of generated hand-object interaction videos and achieving accurate robotic manipulation. + +Our contributions can be summarized as follows: + +- We introduce TASTE-Rob, a large-scale dataset contain + +ing 100,856 ego-centric hand-object interaction videos, each specially tailored for training high-quality video generation model for robot imitation learning. + +- We develop a three-stage pose-refinement pipeline that significantly enhances the fidelity of the hand posture in the generated videos. +- We set a new state of the art in generating task-oriented hand-object interaction videos, which further boosts the performance of robot manipulation. + +# 2. Related Works + +Imitation Learning from video demonstrations. IL enables robots to learn from expert [8, 10, 15, 42, 64] or human [9, 20, 22, 48, 56, 57] video demonstrations by training policies to directly replicate expert actions through supervised learning. While these methods require video demonstrations from environments identical to the robot execution setup, they exhibit limited generalization to novel scenes and tasks. Alternatively, existing works address this challenge by generating demonstrations, such as videos [5, 23, 32, 53], optical flows [52, 58], and images [7], successfully achieving better generalization. Among existing approaches, video demonstration generation represents the most straightforward solution. Yet, robot manipulation video generation [23, 32, 53] is constrained by limited robot datasets, while human hand video generation [5] is hindered by the absence of effective HOI video generative models. In this work, we explore enhancing HOI video generation quality to facilitate robot manipulation. + +Ego-Centric HOI Video datasets. To enhance video quality, we propose a specialized HOI video generation model, which requires a large-scale, high-quality egocentric dataset. Existing ego-centric HOI datasets have various limitations: insufficient resolution [27, 28, 30, 43], limited scale [13, 27, 28, 30, 35, 43], and restricted scene diversity [11, 13, 28]. While Ego4D [16] provides diverse, large-scale HOI videos, its varied camera perspectives and imprecise action-language alignment pose challenges for IL video demonstrations. Therefore, we introduce TASTE-Rob, containing 100,856 videos specifically designed for our task. + +Diffusion Models for Video Generation. Based on the success of VDMs [18], researchers have achieved high-quality video generation with various conditioning inputs: text [33, 41, 61, 62] and images [26, 38, 54, 55, 60]. Imaged-to-video generation aligns well with our task. While these models excel at producing realistic colors and temporal coherence, they struggle with significant motion variations. Motion diffusion models [19, 36, 45, 50, 59] address this limitation by leveraging high-quality human body models [39] to generate motion control parameters. In this work, + +
Orientation(°)0-9090-180180-270270-360
Proportion(%)33.8247.79.598.89
+ +Table 1. Palm orientation distribution of grasping poses. We analyze the palm orientation distribution of grasping poses across TASTE-Rob. Specifically, the degree of orientation denotes the angle between the palm normal direction and the positive $x$ -axis of the frame plane. + +![](images/de7d7110624790b5927a665de9760e6f7f1d59431ea1cb2b5cf21655e50f50c0.jpg) +Figure 2. Distribution of environment and HOI task. In TASTE-Rob, we collect ego-centric task-oriented HOI videos with diverse tasks under diverse environments. + +![](images/ee66407e6383468d8fd75c3fcf3259ead2108771721cf1dfcff342d576d26706.jpg) + +we combine the advantages of VDM and MDM through our pose-refinement pipeline to generate HOI videos with realistic grasping poses. + +# 3. Building TASTE-Rob Dataset + +We collect a massive and diverse ego-centric task-oriented HOI video dataset TASTE-Rob, which includes 100,856 pairs of video and their corresponding language task instruction. To serve for HOI video generation, TASTE-Rob is required to achieve following goals: 1) Each video is recorded with a static camera perspective and contains a single action that closely aligns with the task instruction. 2) It encompasses diversified environments and tasks. 3) It features a variety of hand poses across different HOI scenarios. + +# 3.1. Collection Strategy and Camera Setting + +To achieve the first goal, we utilize multiple cameras equipped with a wide-angle lens capable of capturing 1080p ego-centric videos. We make the following improvements during each recording process. + +Firstly, since our data collection aims to generate task-oriented HOI videos for demonstrations in IL, where demonstrations are typically recorded from fixed camera viewpoints for effective robot imitation learning, we ensure that no camera perspective variation occurs during recording. In addition, as shown in Figure 12, we align the camera perspective specifically to match the head-mounted camera setup of Ego4D [16], ensuring consistency with the egocentric perspective. + +Our second objective is to ensure precise alignment between language task instructions and video actions in TASTE-Rob, a crucial aspect for maintaining the action integrity in generated HOI videos. Unlike Ego4D [16], which captures extended recordings of daily activities via head-mounted cameras and later segments them into shorter clips, + +![](images/2167fe86bb638a720d08d3f5e8a321fe2a209f504cb689a89475e6be114d7fe9.jpg) +Figure 3. Inter-finger angle distribution of human hands. We analyze the distribution of inter-finger angles, specifically between the thumb-index and index-middle finger pairs, across TASTE-Rob. Specifically, higher degrees mean higher inter-finger angles. + +![](images/2dc4741158b7f212cb04befe3fccc471af10e82c5a2bda02673b2feedb72f462.jpg) +Figure 4. Finger curvature distribution of human hands. We analyze the distribution of finger curvatures, specifically that of thumb, index finger and middle finger, across TASTE-Rob. Specifically, higher degrees mean higher curvature. + +we employ a more controlled collection protocol: 1) Each video is strictly limited to under eight seconds in duration and captures a single action. 2) Collectors follow a structured recording process: they press the "start recording" button, execute the specified HOI task according to the provided instruction, and stop recording upon task completion. This approach ensures a precise correspondence between actions and task instructions. + +# 3.2. Data Diversity + +Distribution of environment and task. To promote broad generalization, videos in TASTE-Rob are recorded across diverse environments and cover a wide range of HOI tasks. As shown in Figure 2, environments include locations such as kitchens, bedrooms, dining tables, office tables, etc. Collectors are required to interact with various commonly used objects, and perform tasks including picking, placing, pushing, pouring, etc. To further ensure task diversity, we consider different patterns of hand usage. Specifically, TASTE-Rob comprises 75,389 single-hand task videos and 25,467 double-hand task videos. + +
DatasetLanguage InstructionsEnvironmentData Statistics
Perfect Action AlignmentDeterminerCameraSettingFramesClipsResolution
Disney [13]XXDynamicGarden11M88001080p
ADL [35]XXDynamicDiverse1M436960p
Charades-Ego [43]XXDynamicDiverse5.8M7860720p
HOI4D [29]✓XStaticDiverse4M24661080p
UT Ego [27, 30]XXDynamicDiverse0.9M80320p
EGTEA Gaze+ [28]✓XDynamicKitchen2.5M10,000480p
EPIC-KITCHENS [11]✓XDynamicKitchen6.8M35,0001080p
Ego4D [16]XXDynamicDiverse19.2M83,6471080p
Ours✓✓StaticDiverse9M100,8561080p
+ +Table 2. Dataset comparison with existing ego-centric HOI video datasets. TASTE-Rob, specifically designed for generating task-oriented HOI videos, provides a diverse collection of ego-centric HOI videos with several distinctive features: static camera perspectives, precise language-action alignment, and rich object descriptors (Please see Appendix 8 for details). In the above table, we compare these key characteristics: Perfect Action Alignment (ensuring all actions in the video strictly correspond to language task descriptions), Descriptive Determiners (whether object descriptions exist), Camera Setting (fixed or dynamic camera perspective), and Environment Setting (recording locations). By the way, Even if Ego4D [16] offers an impressive 3000-hours video data, it is designed for various general tasks, and the portion applicable to HOI video generation is significantly smaller, as presented in the above table, which is only 83,647 clips. Cells highlighted in $\square$ denotes the dataset with best feature in each column. + +Distribution of Grasping Hands. To ensure a wide variety of hand poses, we consider two primary factors: diverse palm orientations (global pose) and varied grasping poses (detailed pose). To demonstrate the diversity of hand poses, we utilize HaMeR [34] to extract hand pose parameters and analyze the distribution based on these parameters (see further details in Appendix 9). + +As shown in Table 1, we analyze the distribution of palm orientations during HOI interaction across TASTE-Rob. The analysis reveals the following insights: 1) Hand poses with palms facing down $(0^{\circ} - 180^{\circ})$ are the most common, as this orientation is practical for grasping objects. 2) Hand poses with palms facing left $(90^{\circ} - 270^{\circ})$ occur slightly more frequently than those facing right, likely because all collectors are right-handed and naturally prefer using their right hand for object manipulation. + +We also provide the analysis of hand grasping pose distribution in Figures 3 and 4. Given the dominant roles of the thumb, index finger, and middle finger in HOI, we focus on examining both the inter-finger angles between these fingers and their individual curvature distributions. The analysis in Figure 3 shows a broad distribution of inter-finger angles, indicating diverse hand orientations. Figure 4 reveals two key findings: 1) The curvature distributions of the index and middle fingers exhibit similar patterns, reflecting their synchronized bending during HOI actions. 2) Our dataset captures a broad spectrum of grasping poses, driven by the variety of objects being manipulated. + +# 3.3. Comparisons with Other Datasets + +As shown in Table 2, we provide comparison between TASTE-Rob and existing ego-centric HOI video datasets. + +TASTE-Rob is the first video dataset specifically designed for task-oriented HOI video generation, and it can also serve as a valuable resource for demonstrations in IL. Given that IL video demonstrations are recorded from a fixed camera perspective and include only a single action that aligns with the task instruction, we collect HOI videos under the same settings, distinguishing TASTE-Rob from other datasets. In addition, to improve comprehension of target objects, we incorporate diverse determiners of objects in language task instructions (see Appendix 8 for further details). With TASTE-Rob, we are able to generate high-quality HOI video demonstrations to achieve IL. + +# 4. Method + +Given an environment image and a task description, the generated task-oriented HOI videos are required to satisfy: 1) Accurate Task Understanding: correctly identifying which object to manipulate and how to manipulate it. 2) Feasible HOI: maintaining consistent hand grasping poses throughout the manipulation. + +As illustrated in **Stage I** area of Figure 5, while videos generated by a single VDM $(\hat{v}_c)$ demonstrate accurate task understanding, exhibit limited fidelity in maintaining consistent grasping poses. To fulfill both requirements, we propose a three-stage pose-refinement pipeline, outlined in Figure 5: **Stage I:** employing a learnable Image-to-Video (I2V) diffusion model to generate a coarse HOI video that meets "Accurate Task Understanding". **Stage II:** extracting a hand pose sequence from this coarse video and refining it using a learnable Motion Diffusion Model (MDM) [45]. **Stage III:** using the refined hand pose sequence to generate a high + +![](images/3e9f25d98a58220e4b327640bd04b9bb0f9cab81d6ca270f5af51c238685cba3.jpg) +Figure 5. Overview of our proposed three-stage pose-refinement pipeline. Stage I: our learnable I2V latent diffusion model $\mathcal{V}$ generates a coarse HOI video $\hat{\pmb{v}}_c$ . Stage II: extracting hand pose sequence $\pmb{p}_c$ from $\hat{\pmb{v}}_c$ , and our learnable MDM $\mathcal{M}$ refines it into $\hat{\pmb{p}}$ . Besides, we also obtain the corresponding visualized hand pose images $\pmb{s}$ of $\hat{\pmb{p}}$ . Stage III: incorporating this additional pose condition into $\mathcal{V}$ to generate a fine HOI video $\hat{\pmb{v}}$ . + +fidelity HOI video that satisfies both requirements. + +# 4.1. Preliminary: Diffusion Models + +Diffusion models [17, 44] work by learning to reverse a noising process. They firstly transform data samples $\boldsymbol{x}_0 \sim p_{data}(\boldsymbol{x})$ into Gaussian noise $\boldsymbol{x}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ , then learn to reconstruct the original data by denoising $\boldsymbol{x}_T$ with multiple steps. With the following objective: + +$$ +\min _ {\theta} \mathbb {E} _ {t, \boldsymbol {x}, \epsilon \sim \mathcal {N} (\boldsymbol {0}, \mathbf {I})} \| \epsilon - \epsilon_ {\theta} (\boldsymbol {x} _ {t}, t) \| _ {2} ^ {2}, \tag {1} +$$ + +the model $\epsilon_{\theta}(\pmb{x}_t,t)$ learns to predict single-step noise at each timestep $t$ , where $\epsilon$ is the sampled ground truth noise and $\theta$ indicates the learnable network parameters. After training, the model generates $\hat{\pmb{x}}_0$ through a $T$ -step denoising process from a random Gaussian noise $\pmb{x}_T$ . + +Video Generation. In this work, we explore HOI video generation based on DynamiCrafter [54], a powerful I2V latent diffusion model. Suppose $\mathcal{T}$ , $i\in \mathbb{R}^{3\times H\times W}$ , and $\pmb {v}\in \mathbb{R}^{L\times 3\times H\times W}$ denote task language descriptions, environment images, and ground truth video frames, respectively. DynamiCrafter learns the denoising process in a compact latent space: $\pmb{v}$ is encoded into a compact latent space through an encoder $\mathcal{E}$ , yielding latent representations $\pmb {z} = \mathcal{E}(\pmb {v})\in \mathbb{R}^{L\times C\times h\times w}$ , and decoded via the decoder $\mathcal{D}$ . Within this latent space, the model performs both forward diffusion and guided denoising processes conditioned on $\mathcal{T}$ and $\pmb{i}$ . Specifically, $\pmb{z}_0\sim p_{data}(\pmb {z})$ and $\pmb{z}_{T}\sim \mathcal{N}(\pmb {0},\pmb {I})$ . The optimization objective 1 is updated as: + +$$ +\min _ {\theta_ {V}} \mathbb {E} _ {t, \boldsymbol {z}, \epsilon \sim \mathcal {N} (\boldsymbol {0}, \mathbf {I})} \| \epsilon - \epsilon_ {\theta} \left(\boldsymbol {z} _ {t}, \mathcal {T}, \boldsymbol {i}, t, f p s\right) \| _ {2} ^ {2}, \tag {2} +$$ + +where $fps$ is the FPS control introduced in [54] and $\theta_V$ denotes the learnable parameters of DynamiCrafter. + +![](images/4bfb0fb6f03d3da2d37e2abc1c33535b4aae39434300760d07b57e1239788b64.jpg) +Figure 6. overview of our learnable Image-to-Hand-Params MDM. Based on MDM [45], we extend the framework by incorporating environment information through an additional image condition branch. The model takes a noised hand pose sequence $\pmb{p}_t^{1:L_p}$ , $t$ itself, a CLIP [37] text embedding $\pmb{c}_{txt}$ of $\mathcal{T}$ and a CLIP image embedding $\pmb{c}_{img}$ of $\pmb{i}$ as input. + +Motion Sequence Generation. MDM [45], utilizing a unique transformer encoder-only architecture, demonstrates remarkable performance in generating human motion sequences. Instead of predicting single-step noise, MDM directly generates clean motion sequences. Let $\pmb{p}$ denotes motion sequences and $\mathcal{M}$ denotes MDM network. Here, $\pmb{p}_0 \sim p_{data}(\pmb{p})$ and $\pmb{p}_T \sim \mathcal{N}(\pmb{0},\mathbf{I})$ . The optimization objective 1 is modified as: + +$$ +\min _ {\theta_ {M}} \mathbb {E} _ {t, \boldsymbol {p}, \epsilon \sim \mathcal {N} (\boldsymbol {0}, \mathbf {I})} \| \boldsymbol {p} _ {0} - \mathcal {M} (\boldsymbol {p} _ {t}, \mathcal {T}, t) \| _ {2} ^ {2}, \tag {3} +$$ + +where $\theta_{M}$ denotes the trainable parameters of MDM. After training, MDM generates the final clean motion sequence $\hat{\pmb{p}}_0$ through a $T$ -step denoising process. However, during each step, instead of directly generating $\pmb{p}_{t-1}$ through single-step denoising, MDM firstly predicts the clean motion sequence $\hat{\pmb{p}}_{0,t}$ from $\pmb{p}_t$ with: + +$$ +\hat {\boldsymbol {p}} _ {0, t} = \mathcal {M} (\boldsymbol {p} _ {t}, \mathcal {T}, t), \tag {4} +$$ + +and then reintroduces noises to obtain $\pmb{p}_{t-1}$ . Through repeating this process for $T$ times, MDM generates the final clean motion sequence $\hat{\pmb{p}}_{0,0}$ , denoted as $\hat{\pmb{p}}_0$ . + +# 4.2. Stage I: A Coarse Action Planner + +Our learnable coarse action planner is designed to generate a coarse HOI video $\hat{\pmb{v}}_c\in \mathbb{R}^{L\times 3\times H\times W}$ conditioned on task description $\mathcal{T}$ and environment image $\pmb{i}$ . Specifically, we fine-tune DynamiCrafter [54] on TASTE-Rob to serve as the coarse action planner, denoted as $\nu$ . + +Training. During fine-tuning, we adopt the training strategy similar to that of DynamiCrafter [54]. To leverage DynamiCrafter's robust temporal capabilities while adapting it to our specific HOI video generation, we only finetune its image context projector and spatial layers in its denoising U-Net. The training objective remains consistent with Equation 2, with trainable parameters $\theta_V$ representing parameters of the image context projector and spatial layers. + +Hand-Object Interacting Inconsistency Problem. Our generated coarse HOI videos exhibit accurate task understanding, such as identifying the object to manipulate and determining its target position. However, as shown in $\pmb{p}_c$ of Figure 5, the grasping poses exhibit temporal inconsistencies during manipulation, indicating a lack of motion coherence. Specifically, these inconsistencies refer to undesirable variations in grasping poses over time, manifesting as unnatural changes in hand posture when the relative position between the hands and the grasped object should ideally remain stable. As shown in $\pmb{p}_c$ of Figure 5, the green hand pose demonstrates a pinching gesture, which is inconsistent with the grasping posture of the yellow hand pose and unsuitable for the target object being manipulated. + +# 4.3. Stage II: Revising Hand Pose Sequences + +To address HOI inconsistencies, we train an Image-to-Hand-Params MDM $\mathcal{M}$ . This model refines hand pose sequences $\pmb{p}_c\in \mathbb{R}^{L_p\times N_h\times 2}$ extracted from the coarse video $\hat{\boldsymbol{v}}_c$ . Specifically, we define $\pmb{p}_c$ as the normalized coordinates of hand key-point sequences, where $L_{p}$ denotes the length of sequence and $N_{h}$ is the number of hand key-points. + +Training. Our learnable $\mathcal{M}$ is designed to take the task description $\mathcal{T}$ and environment image $i$ as input to predict refined human hand pose sequences. To achieve this, we extend the original MDM [45] framework by incorporating environmental information through an additional image branch. As shown in Figure 6, similar to the text condition branch of [45], the added image branch integrates CLIP [37] class features of environment image $i$ . Consequently, the optimization objective 3 is updated as follows: + +$$ +\min _ {\theta_ {M}} \mathbb {E} _ {t, \boldsymbol {p}, \epsilon \sim \mathcal {N} (\boldsymbol {0}, \mathbf {I})} \| \boldsymbol {p} _ {0} - \mathcal {M} (\boldsymbol {p} _ {t}, \mathcal {T}, \boldsymbol {i}, t) \| _ {2} ^ {2}, \tag {5} +$$ + +where $\theta_{M}$ denotes the parameters of $\mathcal{M}$ and $\epsilon$ denotes the sampled ground truth noise. + +When we directly generate final clean pose sequence $\hat{\pmb{p}}$ through a $T$ -step denoising process, $\hat{\pmb{p}}$ achieves physically plausible hand motions but exhibits limited spatial awareness. Conversely, $\pmb{p}_c$ exhibits remarkable spatial awareness. To address this limitation, we refine $\pmb{p}_c$ using $\mathcal{M}$ rather than generating from Gaussian noise. Specifically, we initialize the denoising process in Equation 4 of $\mathcal{M}$ with $\pmb{p}_{0,N_{rv}}$ , setting $\pmb{p}_c$ as $\pmb{p}_{0,N_{rv}}$ . Through an $N_{rv}$ -step denoising, we refine $\pmb{p}_c$ to obtain the final clean hand pose sequence $\hat{\pmb{p}}$ which satisfies both spatial accuracy and motion feasibility. + +# 4.4. Stage III: Regenerating with Refined Pose + +With the refined hand pose sequences, we generate the fine-grained HOI videos with the additional pose condition $\hat{p}$ . Inspired by ToonCrafter [55], we train a frame-independent pose encoder $S$ to control the hand pose in the generated + +videos. We design $S$ as a frame-wise adapter that adjusts the intermediate features of each frame independently, conditioned on $\hat{p}$ : $F_{inject}^{i} = S(s^{i},z^{i},t)$ , where $s^i$ is the visualized image sequences of $\hat{p}$ , shown as Figure 5, and $F_{inject}^{i}$ is processed using a method similar to ControlNet [62]. + +Training. During training, we adopt a strategy similar to ToonCrafter [55], where all parameters of $\mathcal{V}$ are frozen, and only the parameters of $S$ , denoted as $\eta$ , are trained with the following optimization objective: + +$$ +\min _ {\eta} \mathbb {E} _ {t, \boldsymbol {s}, \mathcal {E} (\boldsymbol {v}), \epsilon \sim \mathcal {N} (\boldsymbol {0}, \mathbf {I})} \| \epsilon - \epsilon_ {\theta} ^ {S} (\boldsymbol {z} _ {t}; \mathcal {T}, \boldsymbol {i}, \boldsymbol {s}, t, f p s) \| _ {2} ^ {2}, (6) +$$ + +where $\epsilon_{\theta}^{\mathcal{S}}$ denotes the combination of $\epsilon_{\theta}$ and $\mathcal{S}$ , and $s$ represents the visualized images of hand pose sequences. + +Finally, we consider our generated fine HOI videos $\hat{v}$ as the video demonstrations for IL and enable robot manipulation using policy model of Im2Flow2Act [58]. As illustrated in Figure 10, we present the imitation learning results with our generated HOI videos as demonstrations, confirming their effectiveness for enabling robot manipulation. + +# 5. Experiments + +# 5.1. Implementation + +During training, we train our model on the training set of TASTE-Rob dataset. **Stage I:** We fine-tune our coarse action planner based on DynamamiCrafter for 30K steps, with a batch size of 16 and a learning rate of $5 \times 10^{-5}$ . **Stage II:** We train our MDM for 100K steps, with a batch size of 64 and a learning rate of $1 \times 10^{-4}$ . **Stage III:** We fine-tune the pose encoder based on SD for 30K steps, with a batch size of 32 and a learning rate of $5 \times 10^{-5}$ . During inference, we generate videos with a 50-step denoising process and refine pose sequences with $N_{rv}$ as 10 (See more details in Appendix 12). + +# 5.2. Evaluation + +Datasets. We evaluate on two datasets: TASTE-Rob-Test and Mujoco simulation dataset. TASTE-Rob-Test: We split TASTE-Rob dataset into training and test sets, with the test set denoted as TASTE-Rob-Test. For fair evaluation, we select $2\%$ of samples from each action category as test samples. Mujoco simulation dataset: we randomly select 50 tasks from robot simulation platform provided by Im2Flow2Act [58] for robot manipulation. + +In addition, as mentioned in Section 3 and Table 2, other ego-centric video datasets feature different settings from our task, making them unsuitable as evaluation datasets. + +Metrics. We evaluate the performance on three aspects: video generation, grasping pose consistency and robot manipulation. Video Generation: To evaluate the quality of + +generated videos in both the spatial and temporal domains, we report Fréchet Video Distance (FVD) [46], Kernel Video Distance (KVD) [46] and Perceptual Input Conformity (PIC) [54] as used in DynamiCrafter [54]. Grasping Pose Consistency: We evaluate with three metrics - Grasping Pose Variance (GPV, computed by $\frac{1}{L'}\sum_{i=1}^{N}(\theta_i^p - \bar{\theta}^p)^2$ ), Grasp Type Classification Error (GTCE, calculating the classification error of grasping types in generated videos) and Hand Movement Direction Accuracy (HMDA, Measuring the directional consistency of hand movements and their alignment with the intended task trajectory), and details of these metrics are illustrated in Appendix 10. Robot Manipulation: We utilize videos generated on the Mujoco simulation dataset as demonstrations for IL, and deploy the policy model of Im2Flow2Act [58] for robot manipulation. The performance is evaluated by comparing the success rates. + +
MethodGeneration Quality
KVD↓FVD↓PIC↑
consistI2V0.2465.770.85
DynamiCrafter0.2162.360.79
Open-Sora Plan0.1850.190.84
CogVideoX0.1648.720.85
TASTE-Rob0.039.430.90
+ +# 5.3. Baselines and Comparisons. + +We select four existing powerful I2V diffusion models - DynamamiCrafter [54], consistI2V [38], Open-Sora Plan [26] and CogVideoX [60], as baselines, and conduct comparative experiments among these baselines and our method. We provide qualitative comparisons of video generation performance on TASTE-Rob-Test and wild environment in Figure 7, and quantitative comparisons on TASTE-Rob-Test in Table 3, which demonstrates superior video quality and better generalization capability of our method. According to above experiments, all existing powerful general VDMs fail to achieve manipulation tasks well, making them unsuitable for generating HOI video demonstrations. Given that the other two evaluation aspects focus on measuring the fine-grained details of generated videos, we omit further comparisons with these baseline methods. + +# 5.4. Ablation Study + +TASTE-Rob Dataset. We evaluate the performance improvements achieved by using our TASTE-Rob dataset. To be fair, we compare video generation performance between our coarse action planner (Coarse-TASTE-Rob) and Ego4D-Gen, a comparison baseline fine-tuned from Dy + +Table 3. Quantitative comparison between TASTE-Rob and baselines. All video generation metrics prove the better generation performance of TASTE-Rob. + +
MethodGeneration Quality
KVD↓FVD↓PIC↑
Ego4D-Gen0.1852.170.77
Coarse-TASTE-Rob0.0410.850.88
+ +namiCrafter [54] on 83,647 ego-centric HOI videos from Ego4D [16]. The only difference between them is the training datasets they use. As shown in Figure 8 and Table 4, our coarse action planner achieves superior performance. + +Table 4. Quantitative comparison between Ego4D-Gen and Coarse-TASTE-Rob. All video generation metrics prove the better generation performance of Coarse-TASTE-Rob. + +
MetricCoarse-TASTE-RobTASTE-Rob
KVD↓0.040.03
FVD↓10.859.43
PIC↑0.880.90
GPV↓0.280.24
GTCE↓67.8%9.7%
HMDA↓26.4°11.3°
success rate↑84%96%
+ +Table 5. Quantitative comparison between TASTE-Rob and Coarse-TASTE-Rob. All metrics prove the better generation performance, grasping pose consistency and more precise robot manipulation of TASTE-Rob. + +Pose-Refinement Pipeline. We compare Coarse-TASTE-Rob and TASTE-Rob across all three evaluation aspects, as shown in Table 5. At a coarse level, both the video generation metrics in Table 5 and the visual comparisons in Figure 9 demonstrate that our proposed pose-refinement pipeline achieves better HOI video generation. At a fine level, both the grasping pose consistency metrics in Table 5 and Figure 9 further validate that our proposed pose-refinement pipeline effectively addresses the pose inconsistency issue and enables more precise robot manipulation. Besides, as shown in Figure 10 and success rate metrics in Table 5, our proposed pose-refinement pipeline helps the policy model achieve more accurate target placement, leading to improved success rates. In addition, we only do this ablation on our dataset, because videos generated by fine-tuning on existing HOI datasets (e.g., Ego4D [16]) are of low quality so that it's difficult to extract reasonable poses for further refinement. + +Denoising Steps of Pose Refinement. We also discuss the selection of $N_{rv}$ during pose refinement. When $N_{rv} = 0$ (no refinement), hand pose sequences exhibit inconsistency issues, while $N_{rv} = 50$ (generating pose sequences from Gaussian noise) results in limited spatial awareness. + +![](images/8e23b6c71d74b1c7f2b864d781ca157ac4791aac1653c5be8b741c640fed5f47.jpg) +Figure 7. Comparisons of video generation performance. TASTE-Rob generates videos with high-quality and accurate action, while other existing powerful general VDMs fail. + +![](images/334fb23f5d8a6a1a60dd3228b86918901c669c611c0b20987d11d7c6e6ef205f.jpg) + +![](images/a224e58c0dfb441bc57fdc33c95f02f94e97a2353396486f75ca237866027d4d.jpg) +Figure 8. Qualitative comparison between Ego4D-Gen and Coarse-TASTE-Rob. Our TASTE-Rob dataset significantly improves the generalization of video generation models in generating HOI videos with accurate actions. +Figure 9. Qualitative comparison between TASTE-Rob and Coarse-TASTE-Rob. With our proposed pose-refinement pipeline, we can generate HOI videos with more accurate and feasible grasping poses. + +As $N_{rv}$ increases, we observe improved pose consistency but degraded spatial awareness. As shown in Table 6, when $N_{rv}$ increases, video generation quality first improves and then deteriorates. Based on this trade-off, we finally select the value of $N_{rv}$ as 10 (see more details in Appendix 11). + +![](images/498ccf4bb16d5238faea1a92b42639920f4303fd6ab737ae1ce65fc6400bb325.jpg) +Figure 10. Comparison of Robot Manipulation Performance. TASTE-Rob generates higher-quality HOI videos, leading to more precise robot manipulation. + +# 6. Discussion and Conclusion + +Summary. In this work, we collect a diverse, large-scale ego-centric HOI dataset and develop a three-stage pose-refinement pipeline to generate high-quality task-oriented HOI videos. We successfully enhance HOI video generation capabilities, thereby improving robot manipulation generalization to novel scenes through imitation learning. + +Limitations. Our approach achieves accurate manipulation on unseen scenes and enhances the grasping poses in generated videos. However, the generation quality degrades when objects undergo significant transformations, such as rotation and opening. In our future work, we plan to address this limitation. + +# 7. Acknowledgments + +The work was supported in part by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, Guangdong Provincial Outstanding Youth Fund(No. 2023B1515020055), NSFC-61931024, and Shenzhen Science and Technology Program No. JCYJ20220530143604010. It is also partly supported by the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by Key Area R&D Program of Guangdong Province (Grant No. 2018B030338001), by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). + +# References + +[1] Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. 2022. 1 +[2] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans, 2023. 1 +[3] Homanga Bharadhwaj, Abhinav Gupta, and Shubham Tulsiani. Visual affordance prediction for guiding robot exploration. IEEE International Conference on Robotics and Automation (ICRA), 2023. 1 +[4] Homanga Bharadhwaj, Jay Vakil, Mohit Sharma, Abhinav Gupta, Shubham Tulsiani, and Vikash Kumar. Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking, 2023. 1 +[5] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation, 2024. 1, 2 +[6] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables generalizable robot manipulation. In European Conference on Computer Vision (ECCV), 2024. +[7] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models, 2023. 1, 2 +[8] Anthony Brohan, Noah Brown, Justice Carbajal, and et al. Rt-1: Robotics transformer for real-world control at scale. In arXiv preprint arXiv:2212.06817, 2022. 1, 2 +[9] Elliot Chane-Sane, Cordelia Schmid, and Ivan Laptev. Learning video-conditioned policies for unseen manipulation tasks. In ICRA, 2023. 1, 2 +[10] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In + +Proceedings of Robotics: Science and Systems (RSS), 2023. 1, 2 +[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2, 4 +[12] Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B. Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation, 2023. 1 +[13] Alircza Fathi, Jessica K. Hodgins, and James M. Rehg. Social interactions: A first-person perspective. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1226-1233, 2012. 2, 4 +[14] Thomas Feix, Javier Romero, Heinz-Bodo Schmiedmayer, Aaron M. Dollar, and Danica Kragic. The grasp taxonomy of human grasp types. IEEE Transactions on Human-Machine Systems, 46(1):66-77, 2016. 4 +[15] Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. Conference on Robot Learning (CoRL), 2021. 1, 2 +[16] Kristen Grauman, Andrew Westbury, and et al. Ego4d: Around the world in 3,000 hours of egocentric video, 2022. 1, 2, 3, 4, 7 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 5 +[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models, 2022. 2 +[19] Vincent Tao Hu, Wenzhe Yin, Pingchuan Ma, Yunlu Chen, Basura Fernando, Yuki M Asano, Efstratios Gavves, Pascal Mettes, Bjorn Ommer, and Cees G. M. Snoek. Motion flow matching for human motion synthesis and editing, 2023. 2 +[20] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, Igor Gilitschenski, Yonatan Bisk, and Debidatta Dwibedi. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers, 2024. 1, 2 +[21] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. BC-z: Zero-shot task generalization with robotic imitation learning. In 5th Annual Conference on Robot Learning, 2021. 1 +[22] Ananth Jonnavittula, Sagar Parekh, and Dylan P. Losey. View: Visual imitation learning with waypoints, 2024. 1, 2 +[23] Siddharth Karamcheti, Suraj Nair, Annie S. Chen, Thomas Kollar, Chelsea Finn, Dorsa Sadigh, and Percy Liang. Language-driven representation learning for robotics. In Robotics: Science and Systems (RSS), 2023. 2 +[24] Alexander Khazatsky, Karl Pertsch, and et al. Droid: A large-scale in-the-wild robot manipulation dataset, 2024. 1 + +[25] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B. Tenenbaum. Learning to act from actionless videos through dense correspondences, 2023. 1 +[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2, 7 +[27] Yong Jae Lee, Joydeep Ghosh, and Kristen Grauman. Discovering important people and objects for egocentric video summarization. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1346-1353, 2012. 2, 4 +[28] Yin Li, Miao Liu, and James M. Rehg. In the eye of the beholder: Gaze and actions in first person video, 2020. 2, 4 +[29] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21013-21022, 2022. 4 +[30] Zheng Lu and Kristen Grauman. Story-driven summarization for egocentric video. In 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 2714-2721, 2013. 2, 4 +[31] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines, 2019. 4, 5 +[32] Yecheng Jason Ma, William Liang, Vaidehi Som, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Jayaraman. Liv: Language-image representations and rewards for robotic control, 2023. 2 +[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models, 2023. 2 +[34] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3D with transformers. In CVPR, 2024. 4, 2, 3 +[35] Hamed Pirsiavash and Deva Ramanan. Detecting activities of daily living in first-person camera views. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2847-2854, 2012. 2, 4 +[36] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 2 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 6 +[38] Weiming Ren, Huan Yang, Ge Zhang, Cong Wei, Xinrun Du, Wenhao Huang, and Wenhu Chen. Consisti2v: Enhancing visual consistency for image-to-video generation, 2024. 2, 7 +[39] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies + +together. ACM Transactions on Graphics, 36(6):1-17, 2017. 2 +[40] Dandan Shan, Jiaqi Geng, Michelle Shu, and David Fouhey. Understanding human hands in contact at internet scale. In CVPR, 2020. 3 +[41] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning, 2023. 2 +[42] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Proceedings of the 6th Conference on Robot Learning (CoRL), 2022. 1, 2 +[43] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos, 2018. 2, 4 +[44] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. 5 +[45] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 5, 6 +[46] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. In DGS@ICLR, 2019. 7 +[47] Boyang Wang, Nikhil Sridhar, Chao Feng, Mark Van der Merwe, Adam Fishman, Nima Fazeli, and Jeong Joon Park. This&that: Language-gesture controlled video generation for robot planning, 2024. 1 +[48] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play, 2023. 1, 2 +[49] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability, 2023. 4 +[50] Yin Wang, Zhiying Leng, Frederick W. B. Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 21978-21987, 2023. 2 +[51] Zhiqiang Wang, Hao Zheng, Yunshuang Nie, Wenjun Xu, Qingwei Wang, Hua Ye, Zhe Li, Kaidong Zhang, Xuewen Cheng, Wanxi Dong, Chang Cai, Liang Lin, Feng Zheng, and Xiaodan Liang. All robots in one: A new standard and unified dataset for versatile, general-purpose embodied agents, 2024. 1 +[52] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning, 2024. 2 +[53] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. In International Conference on Learning Representations, 2024. 2 + +[54] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors, 2023. 2, 5, 7, 4 +[55] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. ACM Transactions on Graphics (TOG), 43(6):1-11, 2024. 2, 6, 5 +[56] Haoyu Xiong, Quanzhou Li, Yun-Chun Chen, Homanga Bharadhwaj, Samarth Sinha, and Animesh Garg. Learning by watching: Physical imitation of manipulation skills from human videos, 2021. 1, 2 +[57] Mengda Xu, Zhenjia Xu, Cheng Chi, Manuela Veloso, and Shuran Song. XSkill: Cross embodiment skill discovery. In 7th Annual Conference on Robot Learning, 2023. 1, 2 +[58] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. In 8th Annual Conference on Robot Learning, 2024. 1, 2, 6, 7 +[59] Zhao Yang, Bing Su, and Ji-Rong Wen. Synthesizing long-term human motions with diffusion models via coherent sampling, 2023. 2 +[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer, 2024. 2, 7, 3, 4 +[61] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models, 2023. 2 +[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 6 +[63] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models, 2023. 4 +[64] Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware, 2023. 1, 2 +[65] Yifeng Zhu, Arisrei Lim, Peter Stone, and Yuke Zhu. Vision-based manipulation from single human video with open-world object graphs, 2024. 1 \ No newline at end of file diff --git a/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/images.zip b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e56908704d0bc2461666ce0beebb6475cd7eee39 --- /dev/null +++ b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e046f17c57c7e2c200a328496a009311c21fa3649d7c13f53ad6529f4cd5ca70 +size 609579 diff --git a/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/layout.json b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..166296149abd8031c29a4f463d07eb7ee4709c45 --- /dev/null +++ b/2025/TASTE-Rob_ Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation/layout.json @@ -0,0 +1,10042 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 76, + 102, + 536, + 140 + ], + "type": "text", + "content": "TASTE-Rob: Advancing Video Generation of Task-Oriented Hand-Object Interaction for Generalizable Robotic Manipulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "text", + "content": "Hongxiang Zhao" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "text", + "content": " Xingchen Liu" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "text", + "content": " Mutian Xu" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "text", + "content": " Yiming Hao" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "text", + "content": " Weikai Chen" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "text", + "content": " Xiaoguang Han" + }, + { + "bbox": [ + 64, + 160, + 577, + 177 + ], + "type": "inline_equation", + "content": "^{1,2§}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 215, + 188, + 414, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 188, + 414, + 217 + ], + "spans": [ + { + "bbox": [ + 215, + 188, + 414, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 215, + 188, + 414, + 217 + ], + "type": "text", + "content": "SSE, CUHKSZ " + }, + { + "bbox": [ + 215, + 188, + 414, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 215, + 188, + 414, + 217 + ], + "type": "text", + "content": "FNii, CUHKSZ https://taste-rob.github.io" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 269, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 297, + 568 + ], + "type": "text", + "content": "We address key limitations in existing datasets and models for task-oriented hand-object interaction video generation, a critical approach of generating video demonstrations for robotic imitation learning. Current datasets, such as Ego4D [16], often suffer from inconsistent view perspectives and misaligned interactions, leading to reduced video quality and limiting their applicability for precise imitation learning tasks. Towards this end, we introduce TASTE-Rob — a pioneering large-scale dataset of 100,856 ego-centric hand-object interaction videos. Each video is meticulously aligned with language instructions and recorded from a consistent camera viewpoint to ensure interaction clarity. By fine-tuning a Video Diffusion Model (VDM) on TASTE-Rob, we achieve realistic object interactions, though we observed occasional inconsistencies in hand grasping postures. To enhance realism, we introduce a three-stage pose-refinement pipeline that improves hand posture accuracy in generated videos. Our curated dataset, coupled with the specialized pose-refinement framework, provides notable performance gains in generating high-quality, task-oriented hand-object interaction videos, resulting in achieving superior generalizable robotic manipulation. To foster further advancements in the field, TASTE-Rob dataset and source code will be made publicly available on our website https://taste-rob.github.io." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 590, + 137, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 137, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 137, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 658 + ], + "type": "text", + "content": "Robotic manipulation plays an essential role in daily life, assisting with tasks like object handling, liquid pouring, surface cleaning, and drawer operation. Powered by the success of Imitation Learning (IL) [8-10, 15, 20, 22, 42, 48," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 244, + 516, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 244, + 516, + 255 + ], + "spans": [ + { + "bbox": [ + 353, + 244, + 516, + 255 + ], + "type": "text", + "content": "use a spatula to scoop the tomato onto a plate" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 255, + 383, + 297 + ], + "blocks": [ + { + "bbox": [ + 317, + 255, + 383, + 297 + ], + "lines": [ + { + "bbox": [ + 317, + 255, + 383, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 383, + 297 + ], + "type": "image", + "image_path": "97dcc98e83ad6eb25463cbe23ae489411f2275715c320780124030a1392cf437.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 383, + 255, + 552, + 297 + ], + "blocks": [ + { + "bbox": [ + 383, + 255, + 552, + 297 + ], + "lines": [ + { + "bbox": [ + 383, + 255, + 552, + 297 + ], + "spans": [ + { + "bbox": [ + 383, + 255, + 552, + 297 + ], + "type": "image", + "image_path": "a39ee47f7ff47b4d9b324be820685ca9bd66ce0e13ff4ca2603a7cea23bd4210.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 298, + 539, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 298, + 539, + 319 + ], + "spans": [ + { + "bbox": [ + 331, + 298, + 539, + 319 + ], + "type": "text", + "content": "use left hand bring the pink glass to the table, and right hand put the white remote control to the Yellow Book" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 320, + 383, + 361 + ], + "blocks": [ + { + "bbox": [ + 317, + 320, + 383, + 361 + ], + "lines": [ + { + "bbox": [ + 317, + 320, + 383, + 361 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 383, + 361 + ], + "type": "image", + "image_path": "def49f50066aa83f2e7ab2a4dd36560e2bda336a9be2b614573ceced85f3c404.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 363, + 375, + 372 + ], + "lines": [ + { + "bbox": [ + 329, + 363, + 375, + 372 + ], + "spans": [ + { + "bbox": [ + 329, + 363, + 375, + 372 + ], + "type": "text", + "content": "Environment" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 383, + 320, + 552, + 361 + ], + "blocks": [ + { + "bbox": [ + 383, + 320, + 552, + 361 + ], + "lines": [ + { + "bbox": [ + 383, + 320, + 552, + 361 + ], + "spans": [ + { + "bbox": [ + 383, + 320, + 552, + 361 + ], + "type": "image", + "image_path": "825e6e5a3009dbeb3fa6eed0f13abf8de78cf685f146f8a35d25906e50b06293.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 363, + 498, + 372 + ], + "lines": [ + { + "bbox": [ + 437, + 363, + 498, + 372 + ], + "spans": [ + { + "bbox": [ + 437, + 363, + 498, + 372 + ], + "type": "text", + "content": "Generated Videos" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 382, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 416 + ], + "type": "text", + "content": "Figure 1. TASTE-Rob generates high-quality task-oriented hand-object interaction videos, enabling robust generalization in robotic manipulation." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 423, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 483 + ], + "type": "text", + "content": "56, 57, 64], current robots can imitate actions from video demonstrations but require a nearly identical environment in the recorded video, limiting their generalization to novel scenes where factors like the position or movement direction of the manipulated object differ." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "content": "To solve the issue of limited generalization, recent research combines IL with generative models [2, 3, 5-7, 12, 25, 47, 58], like video diffusion models (VDMs), to create adaptable demonstrations and enable more generalizable robotic manipulation [12, 25, 47]. Specifically, these works develop VDMs to generate robot-object interaction videos as demonstrations for IL. While robot-object interaction VDMs enhance generalization performance, they are constrained by limited robot datasets [4, 8, 21, 24, 51]. A scalable alternative is generating hand-object interaction (HOI) videos, which leverage extensive human manipulation video datasets. Although hand-object interaction differs from robotic motion, advanced policy models [1, 5, 9, 22, 48, 57, 65] can bridge this gap effectively, achieving remarkable performance through IL. However, even if existing powerful VDMs are able to generate remarkable human videos, as shown in Figure 7, they struggle with generating high-quality, task-oriented hand-object interaction videos with accurate task understanding." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 664, + 134, + 673 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 664, + 134, + 673 + ], + "spans": [ + { + "bbox": [ + 66, + 664, + 134, + 673 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 67, + 674, + 226, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 674, + 226, + 684 + ], + "spans": [ + { + "bbox": [ + 67, + 674, + 226, + 684 + ], + "type": "text", + "content": "†Work done as a visiting researcher at CUHKSZ." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 56, + 685, + 295, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 685, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 56, + 685, + 295, + 702 + ], + "type": "text", + "content": "This paper solely reflects the author's personal research and is not associated with the author's affiliated institution." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 67, + 703, + 239, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 239, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 239, + 712 + ], + "type": "inline_equation", + "content": "^{\\S}" + }, + { + "bbox": [ + 67, + 703, + 239, + 712 + ], + "type": "text", + "content": "Corresponding author: hanxiaoguang@cuhk.edu.cn." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27683" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "To enhance video generation quality, Ego4D [16], a large-scale data containing 83,647 ego-centric hand-object interaction videos, has proven useful in robotic manipulation tasks [23, 32, 53]. However, Ego4D presents two notable limitations for our purpose, as shown in Appendix 8. Firstly, the camera perspective in each clip is inconsistent, whereas video demonstrations for IL are typically recorded from a fixed viewpoint. Secondly, since these clips are extracted from longer videos that encompass multiple tasks, some clips contain partial interactions that overlap with others, resulting in misalignment between video clips and language instructions. Both limitations negatively impact the quality of video generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 294, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 294, + 385 + ], + "type": "text", + "content": "To address these above limitations, we introduce TASTE-Rob, a pioneering large-scale dataset designed to enhance the generation quality of task-oriented hand-object interaction videos. TASTE-Rob contains 100,856 egocentric videos with well-aligned language instructions. Unlike Ego4D, TASTE-Rob is crafted to ensure clarity and consistency - each video captures a single, complete interaction corresponding to its instruction and is recorded independently with a fixed camera viewpoint, mirroring the conditions needed for reliable imitation learning demonstrations. By resolving these key issues, TASTE-Rob provides a robust foundation for advancing research in video generation, imitation learning, and related fields." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 387, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 685 + ], + "type": "text", + "content": "After fine-tuning a VDM on our curated dataset, we achieve high-quality video generation of task-oriented hand-object interactions. While the model demonstrates proficiency in identifying target objects and determining appropriate coarse interactions, it exhibits inconsistencies in hand poses, with the grip changing unnaturally during manipulation. This limitation presents a critical challenge for robotic imitation learning where precise and stable hand-object interactions are essential. To address this issue, we further propose a dedicated three-stage pipeline for pose refinement. First, we generate a coarse manipulation video conditioned solely on language instructions and environment images. Second, we refine the hand pose sequences using a Motion Diffusion Model (MDM) [45], ensuring realistic grip poses. Finally, we regenerate the videos by incorporating the revised pose sequences as an additional conditioning factor. This pipeline enhances grasping realism in generated videos, enabling robots to execute novel tasks with greater accuracy and adaptability in unseen scenarios. Extensive experiments demonstrate that our proposed TASTE-Rob dataset, combined with the pose-refinement pipeline, achieves significant performance gains, surpassing state-of-the-art methods in the quality of generated hand-object interaction videos and achieving accurate robotic manipulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 688, + 266, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 688, + 266, + 699 + ], + "spans": [ + { + "bbox": [ + 67, + 688, + 266, + 699 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "type": "text", + "content": "- We introduce TASTE-Rob, a large-scale dataset contain" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "content": "ing 100,856 ego-centric hand-object interaction videos, each specially tailored for training high-quality video generation model for robot imitation learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 108, + 553, + 179 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 108, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 143 + ], + "type": "text", + "content": "- We develop a three-stage pose-refinement pipeline that significantly enhances the fidelity of the hand posture in the generated videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "content": "- We set a new state of the art in generating task-oriented hand-object interaction videos, which further boosts the performance of robot manipulation." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 192, + 405, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 192, + 405, + 205 + ], + "spans": [ + { + "bbox": [ + 314, + 192, + 405, + 205 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 213, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 429 + ], + "type": "text", + "content": "Imitation Learning from video demonstrations. IL enables robots to learn from expert [8, 10, 15, 42, 64] or human [9, 20, 22, 48, 56, 57] video demonstrations by training policies to directly replicate expert actions through supervised learning. While these methods require video demonstrations from environments identical to the robot execution setup, they exhibit limited generalization to novel scenes and tasks. Alternatively, existing works address this challenge by generating demonstrations, such as videos [5, 23, 32, 53], optical flows [52, 58], and images [7], successfully achieving better generalization. Among existing approaches, video demonstration generation represents the most straightforward solution. Yet, robot manipulation video generation [23, 32, 53] is constrained by limited robot datasets, while human hand video generation [5] is hindered by the absence of effective HOI video generative models. In this work, we explore enhancing HOI video generation quality to facilitate robot manipulation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 445, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 578 + ], + "type": "text", + "content": "Ego-Centric HOI Video datasets. To enhance video quality, we propose a specialized HOI video generation model, which requires a large-scale, high-quality egocentric dataset. Existing ego-centric HOI datasets have various limitations: insufficient resolution [27, 28, 30, 43], limited scale [13, 27, 28, 30, 35, 43], and restricted scene diversity [11, 13, 28]. While Ego4D [16] provides diverse, large-scale HOI videos, its varied camera perspectives and imprecise action-language alignment pose challenges for IL video demonstrations. Therefore, we introduce TASTE-Rob, containing 100,856 videos specifically designed for our task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Diffusion Models for Video Generation. Based on the success of VDMs [18], researchers have achieved high-quality video generation with various conditioning inputs: text [33, 41, 61, 62] and images [26, 38, 54, 55, 60]. Imaged-to-video generation aligns well with our task. While these models excel at producing realistic colors and temporal coherence, they struggle with significant motion variations. Motion diffusion models [19, 36, 45, 50, 59] address this limitation by leveraging high-quality human body models [39] to generate motion control parameters. In this work," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27684" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 291, + 106 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 291, + 106 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 291, + 106 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 291, + 106 + ], + "type": "table", + "html": "
Orientation(°)0-9090-180180-270270-360
Proportion(%)33.8247.79.598.89
", + "image_path": "184ac5c94f17f029dbb81993dcf89e3c2eddd42b8e3713b28f99b7bf814d9cc1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 115, + 295, + 170 + ], + "lines": [ + { + "bbox": [ + 55, + 115, + 295, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 115, + 295, + 170 + ], + "type": "text", + "content": "Table 1. Palm orientation distribution of grasping poses. We analyze the palm orientation distribution of grasping poses across TASTE-Rob. Specifically, the degree of orientation denotes the angle between the palm normal direction and the positive " + }, + { + "bbox": [ + 55, + 115, + 295, + 170 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 115, + 295, + 170 + ], + "type": "text", + "content": "-axis of the frame plane." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 69, + 171, + 162, + 241 + ], + "blocks": [ + { + "bbox": [ + 69, + 171, + 162, + 241 + ], + "lines": [ + { + "bbox": [ + 69, + 171, + 162, + 241 + ], + "spans": [ + { + "bbox": [ + 69, + 171, + 162, + 241 + ], + "type": "image", + "image_path": "de7d7110624790b5927a665de9760e6f7f1d59431ea1cb2b5cf21655e50f50c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 282 + ], + "type": "text", + "content": "Figure 2. Distribution of environment and HOI task. In TASTE-Rob, we collect ego-centric task-oriented HOI videos with diverse tasks under diverse environments." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 172, + 282, + 240 + ], + "blocks": [ + { + "bbox": [ + 167, + 172, + 282, + 240 + ], + "lines": [ + { + "bbox": [ + 167, + 172, + 282, + 240 + ], + "spans": [ + { + "bbox": [ + 167, + 172, + 282, + 240 + ], + "type": "image", + "image_path": "ee66407e6383468d8fd75c3fcf3259ead2108771721cf1dfcff342d576d26706.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 288, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 324 + ], + "type": "text", + "content": "we combine the advantages of VDM and MDM through our pose-refinement pipeline to generate HOI videos with realistic grasping poses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 334, + 223, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 223, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 223, + 347 + ], + "type": "text", + "content": "3. Building TASTE-Rob Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "text", + "content": "We collect a massive and diverse ego-centric task-oriented HOI video dataset TASTE-Rob, which includes 100,856 pairs of video and their corresponding language task instruction. To serve for HOI video generation, TASTE-Rob is required to achieve following goals: 1) Each video is recorded with a static camera perspective and contains a single action that closely aligns with the task instruction. 2) It encompasses diversified environments and tasks. 3) It features a variety of hand poses across different HOI scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 468, + 264, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 264, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 264, + 482 + ], + "type": "text", + "content": "3.1. Collection Strategy and Camera Setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 486, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 534 + ], + "type": "text", + "content": "To achieve the first goal, we utilize multiple cameras equipped with a wide-angle lens capable of capturing 1080p ego-centric videos. We make the following improvements during each recording process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 534, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 641 + ], + "type": "text", + "content": "Firstly, since our data collection aims to generate task-oriented HOI videos for demonstrations in IL, where demonstrations are typically recorded from fixed camera viewpoints for effective robot imitation learning, we ensure that no camera perspective variation occurs during recording. In addition, as shown in Figure 12, we align the camera perspective specifically to match the head-mounted camera setup of Ego4D [16], ensuring consistency with the egocentric perspective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Our second objective is to ensure precise alignment between language task instructions and video actions in TASTE-Rob, a crucial aspect for maintaining the action integrity in generated HOI videos. Unlike Ego4D [16], which captures extended recordings of daily activities via head-mounted cameras and later segments them into shorter clips," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 340, + 71, + 529, + 196 + ], + "blocks": [ + { + "bbox": [ + 340, + 71, + 529, + 196 + ], + "lines": [ + { + "bbox": [ + 340, + 71, + 529, + 196 + ], + "spans": [ + { + "bbox": [ + 340, + 71, + 529, + 196 + ], + "type": "image", + "image_path": "2167fe86bb638a720d08d3f5e8a321fe2a209f504cb689a89475e6be114d7fe9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 554, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 554, + 250 + ], + "type": "text", + "content": "Figure 3. Inter-finger angle distribution of human hands. We analyze the distribution of inter-finger angles, specifically between the thumb-index and index-middle finger pairs, across TASTE-Rob. Specifically, higher degrees mean higher inter-finger angles." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 339, + 256, + 529, + 381 + ], + "blocks": [ + { + "bbox": [ + 339, + 256, + 529, + 381 + ], + "lines": [ + { + "bbox": [ + 339, + 256, + 529, + 381 + ], + "spans": [ + { + "bbox": [ + 339, + 256, + 529, + 381 + ], + "type": "image", + "image_path": "2dc4741158b7f212cb04befe3fccc471af10e82c5a2bda02673b2feedb72f462.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 389, + 554, + 434 + ], + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 434 + ], + "type": "text", + "content": "Figure 4. Finger curvature distribution of human hands. We analyze the distribution of finger curvatures, specifically that of thumb, index finger and middle finger, across TASTE-Rob. Specifically, higher degrees mean higher curvature." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "text", + "content": "we employ a more controlled collection protocol: 1) Each video is strictly limited to under eight seconds in duration and captures a single action. 2) Collectors follow a structured recording process: they press the \"start recording\" button, execute the specified HOI task according to the provided instruction, and stop recording upon task completion. This approach ensures a precise correspondence between actions and task instructions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 561, + 405, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 561, + 405, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 561, + 405, + 574 + ], + "type": "text", + "content": "3.2. Data Diversity" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Distribution of environment and task. To promote broad generalization, videos in TASTE-Rob are recorded across diverse environments and cover a wide range of HOI tasks. As shown in Figure 2, environments include locations such as kitchens, bedrooms, dining tables, office tables, etc. Collectors are required to interact with various commonly used objects, and perform tasks including picking, placing, pushing, pouring, etc. To further ensure task diversity, we consider different patterns of hand usage. Specifically, TASTE-Rob comprises 75,389 single-hand task videos and 25,467 double-hand task videos." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27685" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 542, + 225 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 542, + 225 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 542, + 225 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 542, + 225 + ], + "type": "table", + "html": "
DatasetLanguage InstructionsEnvironmentData Statistics
Perfect Action AlignmentDeterminerCameraSettingFramesClipsResolution
Disney [13]XXDynamicGarden11M88001080p
ADL [35]XXDynamicDiverse1M436960p
Charades-Ego [43]XXDynamicDiverse5.8M7860720p
HOI4D [29]✓XStaticDiverse4M24661080p
UT Ego [27, 30]XXDynamicDiverse0.9M80320p
EGTEA Gaze+ [28]✓XDynamicKitchen2.5M10,000480p
EPIC-KITCHENS [11]✓XDynamicKitchen6.8M35,0001080p
Ego4D [16]XXDynamicDiverse19.2M83,6471080p
Ours✓✓StaticDiverse9M100,8561080p
", + "image_path": "8a408fa3747df624d22523129383b91ace99a2bacd496922c679f771965b2270.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 233, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 555, + 322 + ], + "type": "text", + "content": "Table 2. Dataset comparison with existing ego-centric HOI video datasets. TASTE-Rob, specifically designed for generating task-oriented HOI videos, provides a diverse collection of ego-centric HOI videos with several distinctive features: static camera perspectives, precise language-action alignment, and rich object descriptors (Please see Appendix 8 for details). In the above table, we compare these key characteristics: Perfect Action Alignment (ensuring all actions in the video strictly correspond to language task descriptions), Descriptive Determiners (whether object descriptions exist), Camera Setting (fixed or dynamic camera perspective), and Environment Setting (recording locations). By the way, Even if Ego4D [16] offers an impressive 3000-hours video data, it is designed for various general tasks, and the portion applicable to HOI video generation is significantly smaller, as presented in the above table, which is only 83,647 clips. Cells highlighted in " + }, + { + "bbox": [ + 55, + 233, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 55, + 233, + 555, + 322 + ], + "type": "text", + "content": " denotes the dataset with best feature in each column." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 328, + 294, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 294, + 411 + ], + "type": "text", + "content": "Distribution of Grasping Hands. To ensure a wide variety of hand poses, we consider two primary factors: diverse palm orientations (global pose) and varied grasping poses (detailed pose). To demonstrate the diversity of hand poses, we utilize HaMeR [34] to extract hand pose parameters and analyze the distribution based on these parameters (see further details in Appendix 9)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 412, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 412, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 412, + 295, + 519 + ], + "type": "text", + "content": "As shown in Table 1, we analyze the distribution of palm orientations during HOI interaction across TASTE-Rob. The analysis reveals the following insights: 1) Hand poses with palms facing down " + }, + { + "bbox": [ + 54, + 412, + 295, + 519 + ], + "type": "inline_equation", + "content": "(0^{\\circ} - 180^{\\circ})" + }, + { + "bbox": [ + 54, + 412, + 295, + 519 + ], + "type": "text", + "content": " are the most common, as this orientation is practical for grasping objects. 2) Hand poses with palms facing left " + }, + { + "bbox": [ + 54, + 412, + 295, + 519 + ], + "type": "inline_equation", + "content": "(90^{\\circ} - 270^{\\circ})" + }, + { + "bbox": [ + 54, + 412, + 295, + 519 + ], + "type": "text", + "content": " occur slightly more frequently than those facing right, likely because all collectors are right-handed and naturally prefer using their right hand for object manipulation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "type": "text", + "content": "We also provide the analysis of hand grasping pose distribution in Figures 3 and 4. Given the dominant roles of the thumb, index finger, and middle finger in HOI, we focus on examining both the inter-finger angles between these fingers and their individual curvature distributions. The analysis in Figure 3 shows a broad distribution of inter-finger angles, indicating diverse hand orientations. Figure 4 reveals two key findings: 1) The curvature distributions of the index and middle fingers exhibit similar patterns, reflecting their synchronized bending during HOI actions. 2) Our dataset captures a broad spectrum of grasping poses, driven by the variety of objects being manipulated." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 237, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 237, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 237, + 685 + ], + "type": "text", + "content": "3.3. Comparisons with Other Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "As shown in Table 2, we provide comparison between TASTE-Rob and existing ego-centric HOI video datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "text", + "content": "TASTE-Rob is the first video dataset specifically designed for task-oriented HOI video generation, and it can also serve as a valuable resource for demonstrations in IL. Given that IL video demonstrations are recorded from a fixed camera perspective and include only a single action that aligns with the task instruction, we collect HOI videos under the same settings, distinguishing TASTE-Rob from other datasets. In addition, to improve comprehension of target objects, we incorporate diverse determiners of objects in language task instructions (see Appendix 8 for further details). With TASTE-Rob, we are able to generate high-quality HOI video demonstrations to achieve IL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 487, + 370, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 370, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 370, + 499 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "text", + "content": "Given an environment image and a task description, the generated task-oriented HOI videos are required to satisfy: 1) Accurate Task Understanding: correctly identifying which object to manipulate and how to manipulate it. 2) Feasible HOI: maintaining consistent hand grasping poses throughout the manipulation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "As illustrated in **Stage I** area of Figure 5, while videos generated by a single VDM " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "(\\hat{v}_c)" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " demonstrate accurate task understanding, exhibit limited fidelity in maintaining consistent grasping poses. To fulfill both requirements, we propose a three-stage pose-refinement pipeline, outlined in Figure 5: **Stage I:** employing a learnable Image-to-Video (I2V) diffusion model to generate a coarse HOI video that meets \"Accurate Task Understanding\". **Stage II:** extracting a hand pose sequence from this coarse video and refining it using a learnable Motion Diffusion Model (MDM) [45]. **Stage III:** using the refined hand pose sequence to generate a high" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27686" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 281, + 192 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 281, + 192 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 281, + 192 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 281, + 192 + ], + "type": "image", + "image_path": "3e9f25d98a58220e4b327640bd04b9bb0f9cab81d6ca270f5af51c238685cba3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": "Figure 5. Overview of our proposed three-stage pose-refinement pipeline. Stage I: our learnable I2V latent diffusion model " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": " generates a coarse HOI video " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}_c" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": ". Stage II: extracting hand pose sequence " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}_c" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": ", and our learnable MDM " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": " refines it into " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": ". Besides, we also obtain the corresponding visualized hand pose images " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{s}" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": ". Stage III: incorporating this additional pose condition into " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": " to generate a fine HOI video " + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}" + }, + { + "bbox": [ + 55, + 195, + 296, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 280, + 261, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 261, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 261, + 293 + ], + "type": "text", + "content": "fidelity HOI video that satisfies both requirements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 223, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 223, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 223, + 314 + ], + "type": "text", + "content": "4.1. Preliminary: Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 319, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 380 + ], + "type": "text", + "content": "Diffusion models [17, 44] work by learning to reverse a noising process. They firstly transform data samples " + }, + { + "bbox": [ + 55, + 319, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0 \\sim p_{data}(\\boldsymbol{x})" + }, + { + "bbox": [ + 55, + 319, + 296, + 380 + ], + "type": "text", + "content": " into Gaussian noise " + }, + { + "bbox": [ + 55, + 319, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 55, + 319, + 296, + 380 + ], + "type": "text", + "content": ", then learn to reconstruct the original data by denoising " + }, + { + "bbox": [ + 55, + 319, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_T" + }, + { + "bbox": [ + 55, + 319, + 296, + 380 + ], + "type": "text", + "content": " with multiple steps. With the following objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 389, + 295, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 389, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 104, + 389, + 295, + 407 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {t, \\boldsymbol {x}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "54ba1c7fe503f673be18b370a01a51b0eb9f0097c89fedd9a158ac367c422256.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": "the model " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " learns to predict single-step noise at each timestep " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " is the sampled ground truth noise and " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " indicates the learnable network parameters. After training, the model generates " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " through a " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": "-step denoising process from a random Gaussian noise " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\pmb{x}_T" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": "Video Generation. In this work, we explore HOI video generation based on DynamiCrafter [54], a powerful I2V latent diffusion model. Suppose " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "i\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\pmb {v}\\in \\mathbb{R}^{L\\times 3\\times H\\times W}" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": " denote task language descriptions, environment images, and ground truth video frames, respectively. DynamiCrafter learns the denoising process in a compact latent space: " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": " is encoded into a compact latent space through an encoder " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": ", yielding latent representations " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\pmb {z} = \\mathcal{E}(\\pmb {v})\\in \\mathbb{R}^{L\\times C\\times h\\times w}" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": ", and decoded via the decoder " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": ". Within this latent space, the model performs both forward diffusion and guided denoising processes conditioned on " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{i}" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{z}_0\\sim p_{data}(\\pmb {z})" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{T}\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})" + }, + { + "bbox": [ + 55, + 495, + 296, + 651 + ], + "type": "text", + "content": ". The optimization objective 1 is updated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 661, + 295, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 661, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 84, + 661, + 295, + 680 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {V}} \\mathbb {E} _ {t, \\boldsymbol {z}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {z} _ {t}, \\mathcal {T}, \\boldsymbol {i}, t, f p s\\right) \\| _ {2} ^ {2}, \\tag {2}", + "image_path": "14e5a1142a330d4ec75666107b7d999c23fe944b42da628eec926001591f6e26.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "fps" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " is the FPS control introduced in [54] and " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\theta_V" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " denotes the learnable parameters of DynamiCrafter." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 340, + 69, + 530, + 190 + ], + "blocks": [ + { + "bbox": [ + 340, + 69, + 530, + 190 + ], + "lines": [ + { + "bbox": [ + 340, + 69, + 530, + 190 + ], + "spans": [ + { + "bbox": [ + 340, + 69, + 530, + 190 + ], + "type": "image", + "image_path": "4bfb0fb6f03d3da2d37e2abc1c33535b4aae39434300760d07b57e1239788b64.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "text", + "content": "Figure 6. overview of our learnable Image-to-Hand-Params MDM. Based on MDM [45], we extend the framework by incorporating environment information through an additional image condition branch. The model takes a noised hand pose sequence " + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t^{1:L_p}" + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "text", + "content": " itself, a CLIP [37] text embedding " + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{txt}" + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "text", + "content": " and a CLIP image embedding " + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{img}" + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{i}" + }, + { + "bbox": [ + 313, + 194, + 555, + 262 + ], + "type": "text", + "content": " as input." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": "Motion Sequence Generation. MDM [45], utilizing a unique transformer encoder-only architecture, demonstrates remarkable performance in generating human motion sequences. Instead of predicting single-step noise, MDM directly generates clean motion sequences. Let " + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": " denotes motion sequences and " + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": " denotes MDM network. Here, " + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{p}_0 \\sim p_{data}(\\pmb{p})" + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{p}_T \\sim \\mathcal{N}(\\pmb{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": ". The optimization objective 1 is modified as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 352, + 365, + 555, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 365, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 352, + 365, + 555, + 385 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {M}} \\mathbb {E} _ {t, \\boldsymbol {p}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\boldsymbol {p} _ {0} - \\mathcal {M} (\\boldsymbol {p} _ {t}, \\mathcal {T}, t) \\| _ {2} ^ {2}, \\tag {3}", + "image_path": "b99a300390d7a0e7b71b9f8704a576a608510780837b7f0140341462811f7ec8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\theta_{M}" + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "text", + "content": " denotes the trainable parameters of MDM. After training, MDM generates the final clean motion sequence " + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_0" + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "text", + "content": " through a " + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "text", + "content": "-step denoising process. However, during each step, instead of directly generating " + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{t-1}" + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "text", + "content": " through single-step denoising, MDM firstly predicts the clean motion sequence " + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_{0,t}" + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t" + }, + { + "bbox": [ + 313, + 388, + 554, + 461 + ], + "type": "text", + "content": " with:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 391, + 467, + 555, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 467, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 391, + 467, + 555, + 480 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} _ {0, t} = \\mathcal {M} (\\boldsymbol {p} _ {t}, \\mathcal {T}, t), \\tag {4}", + "image_path": "7e8bbc75725bf29b90ce9d0f8af68a5a06849b62b2da4e2110dbf62de51b6a74.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": "and then reintroduces noises to obtain " + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{t-1}" + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": ". Through repeating this process for " + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": " times, MDM generates the final clean motion sequence " + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_{0,0}" + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_0" + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 526, + 494, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 494, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 494, + 540 + ], + "type": "text", + "content": "4.2. Stage I: A Coarse Action Planner" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "text", + "content": "Our learnable coarse action planner is designed to generate a coarse HOI video " + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}_c\\in \\mathbb{R}^{L\\times 3\\times H\\times W}" + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "text", + "content": " conditioned on task description " + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "text", + "content": " and environment image " + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\pmb{i}" + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "text", + "content": ". Specifically, we fine-tune DynamiCrafter [54] on TASTE-Rob to serve as the coarse action planner, denoted as " + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Training. During fine-tuning, we adopt the training strategy similar to that of DynamiCrafter [54]. To leverage DynamiCrafter's robust temporal capabilities while adapting it to our specific HOI video generation, we only finetune its image context projector and spatial layers in its denoising U-Net. The training objective remains consistent with Equation 2, with trainable parameters " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\theta_V" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " representing parameters of the image context projector and spatial layers." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27687" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "Hand-Object Interacting Inconsistency Problem. Our generated coarse HOI videos exhibit accurate task understanding, such as identifying the object to manipulate and determining its target position. However, as shown in " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " of Figure 5, the grasping poses exhibit temporal inconsistencies during manipulation, indicating a lack of motion coherence. Specifically, these inconsistencies refer to undesirable variations in grasping poses over time, manifesting as unnatural changes in hand posture when the relative position between the hands and the grasped object should ideally remain stable. As shown in " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " of Figure 5, the green hand pose demonstrates a pinching gesture, which is inconsistent with the grasping posture of the yellow hand pose and unsuitable for the target object being manipulated." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 265, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 265, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 265, + 261 + ], + "type": "text", + "content": "4.3. Stage II: Revising Hand Pose Sequences" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "content": "To address HOI inconsistencies, we train an Image-to-Hand-Params MDM " + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "content": ". This model refines hand pose sequences " + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c\\in \\mathbb{R}^{L_p\\times N_h\\times 2}" + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "content": " extracted from the coarse video " + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{v}}_c" + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "content": ". Specifically, we define " + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "content": " as the normalized coordinates of hand key-point sequences, where " + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "content": " denotes the length of sequence and " + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "inline_equation", + "content": "N_{h}" + }, + { + "bbox": [ + 55, + 265, + 296, + 339 + ], + "type": "text", + "content": " is the number of hand key-points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "text", + "content": "Training. Our learnable " + }, + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "text", + "content": " is designed to take the task description " + }, + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "text", + "content": " and environment image " + }, + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "text", + "content": " as input to predict refined human hand pose sequences. To achieve this, we extend the original MDM [45] framework by incorporating environmental information through an additional image branch. As shown in Figure 6, similar to the text condition branch of [45], the added image branch integrates CLIP [37] class features of environment image " + }, + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 353, + 296, + 462 + ], + "type": "text", + "content": ". Consequently, the optimization objective 3 is updated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 472, + 295, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 472, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 89, + 472, + 295, + 491 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {M}} \\mathbb {E} _ {t, \\boldsymbol {p}, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\boldsymbol {p} _ {0} - \\mathcal {M} (\\boldsymbol {p} _ {t}, \\mathcal {T}, \\boldsymbol {i}, t) \\| _ {2} ^ {2}, \\tag {5}", + "image_path": "ff5d28cdefa1002dff99827236b2e587788b769f8cdbf6443a03ec0dcf26b7eb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 496, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 496, + 295, + 519 + ], + "type": "inline_equation", + "content": "\\theta_{M}" + }, + { + "bbox": [ + 55, + 496, + 295, + 519 + ], + "type": "text", + "content": " denotes the parameters of " + }, + { + "bbox": [ + 55, + 496, + 295, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 496, + 295, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 496, + 295, + 519 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 496, + 295, + 519 + ], + "type": "text", + "content": " denotes the sampled ground truth noise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": "When we directly generate final clean pose sequence " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": " through a " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": "-step denoising process, " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": " achieves physically plausible hand motions but exhibits limited spatial awareness. Conversely, " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": " exhibits remarkable spatial awareness. To address this limitation, we refine " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": " rather than generating from Gaussian noise. Specifically, we initialize the denoising process in Equation 4 of " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{0,N_{rv}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": ", setting " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{0,N_{rv}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": ". Through an " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "N_{rv}" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": "-step denoising, we refine " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{p}_c" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": " to obtain the final clean hand pose sequence " + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": " which satisfies both spatial accuracy and motion feasibility." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 648, + 275, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 275, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 275, + 661 + ], + "type": "text", + "content": "4.4. Stage III: Regenerating with Refined Pose" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "With the refined hand pose sequences, we generate the fine-grained HOI videos with the additional pose condition " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". Inspired by ToonCrafter [55], we train a frame-independent pose encoder " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " to control the hand pose in the generated" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": "videos. We design " + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": " as a frame-wise adapter that adjusts the intermediate features of each frame independently, conditioned on " + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "inline_equation", + "content": "F_{inject}^{i} = S(s^{i},z^{i},t)" + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "inline_equation", + "content": "s^i" + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": " is the visualized image sequences of " + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": ", shown as Figure 5, and " + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "inline_equation", + "content": "F_{inject}^{i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": " is processed using a method similar to ControlNet [62]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 148, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 196 + ], + "type": "text", + "content": "Training. During training, we adopt a strategy similar to ToonCrafter [55], where all parameters of " + }, + { + "bbox": [ + 313, + 148, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 148, + 555, + 196 + ], + "type": "text", + "content": " are frozen, and only the parameters of " + }, + { + "bbox": [ + 313, + 148, + 555, + 196 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 148, + 555, + 196 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 148, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 148, + 555, + 196 + ], + "type": "text", + "content": ", are trained with the following optimization objective:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 323, + 203, + 555, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 203, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 323, + 203, + 555, + 223 + ], + "type": "interline_equation", + "content": "\\min _ {\\eta} \\mathbb {E} _ {t, \\boldsymbol {s}, \\mathcal {E} (\\boldsymbol {v}), \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} ^ {S} (\\boldsymbol {z} _ {t}; \\mathcal {T}, \\boldsymbol {i}, \\boldsymbol {s}, t, f p s) \\| _ {2} ^ {2}, (6)", + "image_path": "c9f1f820d36ae172e8382d2eb52dd18854c5252d6437ec4e4e212443df8ef2d3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\mathcal{S}}" + }, + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "text", + "content": " denotes the combination of " + }, + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 232, + 554, + 255 + ], + "type": "text", + "content": " represents the visualized images of hand pose sequences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 256, + 554, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 554, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 554, + 328 + ], + "type": "text", + "content": "Finally, we consider our generated fine HOI videos " + }, + { + "bbox": [ + 313, + 256, + 554, + 328 + ], + "type": "inline_equation", + "content": "\\hat{v}" + }, + { + "bbox": [ + 313, + 256, + 554, + 328 + ], + "type": "text", + "content": " as the video demonstrations for IL and enable robot manipulation using policy model of Im2Flow2Act [58]. As illustrated in Figure 10, we present the imitation learning results with our generated HOI videos as demonstrations, confirming their effectiveness for enabling robot manipulation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 338, + 397, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 397, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 397, + 352 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 357, + 411, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 411, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 411, + 370 + ], + "type": "text", + "content": "5.1. Implementation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "text", + "content": "During training, we train our model on the training set of TASTE-Rob dataset. **Stage I:** We fine-tune our coarse action planner based on DynamamiCrafter for 30K steps, with a batch size of 16 and a learning rate of " + }, + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "text", + "content": ". **Stage II:** We train our MDM for 100K steps, with a batch size of 64 and a learning rate of " + }, + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "text", + "content": ". **Stage III:** We fine-tune the pose encoder based on SD for 30K steps, with a batch size of 32 and a learning rate of " + }, + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "text", + "content": ". During inference, we generate videos with a 50-step denoising process and refine pose sequences with " + }, + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "inline_equation", + "content": "N_{rv}" + }, + { + "bbox": [ + 313, + 375, + 555, + 507 + ], + "type": "text", + "content": " as 10 (See more details in Appendix 12)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 514, + 389, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 389, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 389, + 525 + ], + "type": "text", + "content": "5.2. Evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 531, + 554, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 554, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 554, + 627 + ], + "type": "text", + "content": "Datasets. We evaluate on two datasets: TASTE-Rob-Test and Mujoco simulation dataset. TASTE-Rob-Test: We split TASTE-Rob dataset into training and test sets, with the test set denoted as TASTE-Rob-Test. For fair evaluation, we select " + }, + { + "bbox": [ + 313, + 531, + 554, + 627 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 531, + 554, + 627 + ], + "type": "text", + "content": " of samples from each action category as test samples. Mujoco simulation dataset: we randomly select 50 tasks from robot simulation platform provided by Im2Flow2Act [58] for robot manipulation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 628, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 663 + ], + "type": "text", + "content": "In addition, as mentioned in Section 3 and Table 2, other ego-centric video datasets feature different settings from our task, making them unsuitable as evaluation datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Metrics. We evaluate the performance on three aspects: video generation, grasping pose consistency and robot manipulation. Video Generation: To evaluate the quality of" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27688" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 275 + ], + "type": "text", + "content": "generated videos in both the spatial and temporal domains, we report Fréchet Video Distance (FVD) [46], Kernel Video Distance (KVD) [46] and Perceptual Input Conformity (PIC) [54] as used in DynamiCrafter [54]. Grasping Pose Consistency: We evaluate with three metrics - Grasping Pose Variance (GPV, computed by " + }, + { + "bbox": [ + 55, + 72, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\frac{1}{L'}\\sum_{i=1}^{N}(\\theta_i^p - \\bar{\\theta}^p)^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 275 + ], + "type": "text", + "content": "), Grasp Type Classification Error (GTCE, calculating the classification error of grasping types in generated videos) and Hand Movement Direction Accuracy (HMDA, Measuring the directional consistency of hand movements and their alignment with the intended task trajectory), and details of these metrics are illustrated in Appendix 10. Robot Manipulation: We utilize videos generated on the Mujoco simulation dataset as demonstrations for IL, and deploy the policy model of Im2Flow2Act [58] for robot manipulation. The performance is evaluated by comparing the success rates." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 87, + 278, + 264, + 386 + ], + "blocks": [ + { + "bbox": [ + 87, + 278, + 264, + 386 + ], + "lines": [ + { + "bbox": [ + 87, + 278, + 264, + 386 + ], + "spans": [ + { + "bbox": [ + 87, + 278, + 264, + 386 + ], + "type": "table", + "html": "
MethodGeneration Quality
KVD↓FVD↓PIC↑
consistI2V0.2465.770.85
DynamiCrafter0.2162.360.79
Open-Sora Plan0.1850.190.84
CogVideoX0.1648.720.85
TASTE-Rob0.039.430.90
", + "image_path": "ae7e4363c7697e0f3f20276c03449758023f07c02052f4e6165cbf5ae6df0674.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 430, + 209, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 209, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 209, + 443 + ], + "type": "text", + "content": "5.3. Baselines and Comparisons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": "We select four existing powerful I2V diffusion models - DynamamiCrafter [54], consistI2V [38], Open-Sora Plan [26] and CogVideoX [60], as baselines, and conduct comparative experiments among these baselines and our method. We provide qualitative comparisons of video generation performance on TASTE-Rob-Test and wild environment in Figure 7, and quantitative comparisons on TASTE-Rob-Test in Table 3, which demonstrates superior video quality and better generalization capability of our method. According to above experiments, all existing powerful general VDMs fail to achieve manipulation tasks well, making them unsuitable for generating HOI video demonstrations. Given that the other two evaluation aspects focus on measuring the fine-grained details of generated videos, we omit further comparisons with these baseline methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 635, + 149, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 149, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 149, + 649 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "TASTE-Rob Dataset. We evaluate the performance improvements achieved by using our TASTE-Rob dataset. To be fair, we compare video generation performance between our coarse action planner (Coarse-TASTE-Rob) and Ego4D-Gen, a comparison baseline fine-tuned from Dy" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 335, + 70, + 533, + 137 + ], + "blocks": [ + { + "bbox": [ + 55, + 389, + 295, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 422 + ], + "type": "text", + "content": "Table 3. Quantitative comparison between TASTE-Rob and baselines. All video generation metrics prove the better generation performance of TASTE-Rob." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 70, + 533, + 137 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 533, + 137 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 533, + 137 + ], + "type": "table", + "html": "
MethodGeneration Quality
KVD↓FVD↓PIC↑
Ego4D-Gen0.1852.170.77
Coarse-TASTE-Rob0.0410.850.88
", + "image_path": "e6f345b330f5acb110c29fa48d55ec8985bba41e7033447e3133a39d0c73d5c3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 179, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 555, + 228 + ], + "type": "text", + "content": "namiCrafter [54] on 83,647 ego-centric HOI videos from Ego4D [16]. The only difference between them is the training datasets they use. As shown in Figure 8 and Table 4, our coarse action planner achieves superior performance." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 239, + 539, + 357 + ], + "blocks": [ + { + "bbox": [ + 313, + 139, + 555, + 173 + ], + "lines": [ + { + "bbox": [ + 313, + 139, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 555, + 173 + ], + "type": "text", + "content": "Table 4. Quantitative comparison between Ego4D-Gen and Coarse-TASTE-Rob. All video generation metrics prove the better generation performance of Coarse-TASTE-Rob." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 239, + 539, + 357 + ], + "lines": [ + { + "bbox": [ + 328, + 239, + 539, + 357 + ], + "spans": [ + { + "bbox": [ + 328, + 239, + 539, + 357 + ], + "type": "table", + "html": "
MetricCoarse-TASTE-RobTASTE-Rob
KVD↓0.040.03
FVD↓10.859.43
PIC↑0.880.90
GPV↓0.280.24
GTCE↓67.8%9.7%
HMDA↓26.4°11.3°
success rate↑84%96%
", + "image_path": "b504a9921437d9cee8680c0c21e26ce243d5ae62d3d7cdce3c47643269dfe3ea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 555, + 405 + ], + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 405 + ], + "type": "text", + "content": "Table 5. Quantitative comparison between TASTE-Rob and Coarse-TASTE-Rob. All metrics prove the better generation performance, grasping pose consistency and more precise robot manipulation of TASTE-Rob." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 421, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 556, + 635 + ], + "type": "text", + "content": "Pose-Refinement Pipeline. We compare Coarse-TASTE-Rob and TASTE-Rob across all three evaluation aspects, as shown in Table 5. At a coarse level, both the video generation metrics in Table 5 and the visual comparisons in Figure 9 demonstrate that our proposed pose-refinement pipeline achieves better HOI video generation. At a fine level, both the grasping pose consistency metrics in Table 5 and Figure 9 further validate that our proposed pose-refinement pipeline effectively addresses the pose inconsistency issue and enables more precise robot manipulation. Besides, as shown in Figure 10 and success rate metrics in Table 5, our proposed pose-refinement pipeline helps the policy model achieve more accurate target placement, leading to improved success rates. In addition, we only do this ablation on our dataset, because videos generated by fine-tuning on existing HOI datasets (e.g., Ego4D [16]) are of low quality so that it's difficult to extract reasonable poses for further refinement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Denoising Steps of Pose Refinement. We also discuss the selection of " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{rv}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " during pose refinement. When " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{rv} = 0" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " (no refinement), hand pose sequences exhibit inconsistency issues, while " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "N_{rv} = 50" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " (generating pose sequences from Gaussian noise) results in limited spatial awareness." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27689" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 545, + 246 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 545, + 246 + ], + "type": "image", + "image_path": "8e23b6c71d74b1c7f2b864d781ca157ac4791aac1653c5be8b741c640fed5f47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 555, + 285 + ], + "type": "text", + "content": "Figure 7. Comparisons of video generation performance. TASTE-Rob generates videos with high-quality and accurate action, while other existing powerful general VDMs fail." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 297, + 284, + 422 + ], + "blocks": [ + { + "bbox": [ + 69, + 297, + 284, + 422 + ], + "lines": [ + { + "bbox": [ + 69, + 297, + 284, + 422 + ], + "spans": [ + { + "bbox": [ + 69, + 297, + 284, + 422 + ], + "type": "image", + "image_path": "334fb23f5d8a6a1a60dd3228b86918901c669c611c0b20987d11d7c6e6ef205f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 465, + 284, + 597 + ], + "blocks": [ + { + "bbox": [ + 55, + 423, + 295, + 465 + ], + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 465 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison between Ego4D-Gen and Coarse-TASTE-Rob. Our TASTE-Rob dataset significantly improves the generalization of video generation models in generating HOI videos with accurate actions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 465, + 284, + 597 + ], + "lines": [ + { + "bbox": [ + 67, + 465, + 284, + 597 + ], + "spans": [ + { + "bbox": [ + 67, + 465, + 284, + 597 + ], + "type": "image", + "image_path": "a224e58c0dfb441bc57fdc33c95f02f94e97a2353396486f75ca237866027d4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 603, + 295, + 647 + ], + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 647 + ], + "type": "text", + "content": "Figure 9. Qualitative comparison between TASTE-Rob and Coarse-TASTE-Rob. With our proposed pose-refinement pipeline, we can generate HOI videos with more accurate and feasible grasping poses." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_{rv}" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " increases, we observe improved pose consistency but degraded spatial awareness. As shown in Table 6, when " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_{rv}" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " increases, video generation quality first improves and then deteriorates. Based on this trade-off, we finally select the value of " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_{rv}" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " as 10 (see more details in Appendix 11)." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 297, + 541, + 447 + ], + "blocks": [ + { + "bbox": [ + 326, + 297, + 541, + 447 + ], + "lines": [ + { + "bbox": [ + 326, + 297, + 541, + 447 + ], + "spans": [ + { + "bbox": [ + 326, + 297, + 541, + 447 + ], + "type": "image", + "image_path": "498ccf4bb16d5238faea1a92b42639920f4303fd6ab737ae1ce65fc6400bb325.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 448, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 480 + ], + "type": "text", + "content": "Figure 10. Comparison of Robot Manipulation Performance. TASTE-Rob generates higher-quality HOI videos, leading to more precise robot manipulation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 500, + 466, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 466, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 466, + 513 + ], + "type": "text", + "content": "6. Discussion and Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 520, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 593 + ], + "type": "text", + "content": "Summary. In this work, we collect a diverse, large-scale ego-centric HOI dataset and develop a three-stage pose-refinement pipeline to generate high-quality task-oriented HOI videos. We successfully enhance HOI video generation capabilities, thereby improving robot manipulation generalization to novel scenes through imitation learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 607, + 556, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 556, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 556, + 678 + ], + "type": "text", + "content": "Limitations. Our approach achieves accurate manipulation on unseen scenes and enhances the grasping poses in generated videos. However, the generation quality degrades when objects undergo significant transformations, such as rotation and opening. In our future work, we plan to address this limitation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27690" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 281 + ], + "type": "text", + "content": "The work was supported in part by the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, Guangdong Provincial Outstanding Youth Fund(No. 2023B1515020055), NSFC-61931024, and Shenzhen Science and Technology Program No. JCYJ20220530143604010. It is also partly supported by the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by Key Area R&D Program of Guangdong Province (Grant No. 2018B030338001), by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 293, + 115, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 115, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 115, + 304 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 312, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 312, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 312, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 312, + 294, + 334 + ], + "type": "text", + "content": "[1] Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 335, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 335, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 61, + 335, + 295, + 368 + ], + "type": "text", + "content": "[2] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 369, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 411 + ], + "type": "text", + "content": "[3] Homanga Bharadhwaj, Abhinav Gupta, and Shubham Tulsiani. Visual affordance prediction for guiding robot exploration. IEEE International Conference on Robotics and Automation (ICRA), 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "text", + "content": "[4] Homanga Bharadhwaj, Jay Vakil, Mohit Sharma, Abhinav Gupta, Shubham Tulsiani, and Vikash Kumar. Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 458, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 512 + ], + "type": "text", + "content": "[5] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 514, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 567 + ], + "type": "text", + "content": "[6] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables generalizable robot manipulation. In European Conference on Computer Vision (ECCV), 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 569, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 294, + 612 + ], + "type": "text", + "content": "[7] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 294, + 646 + ], + "type": "text", + "content": "[8] Anthony Brohan, Noah Brown, Justice Carbajal, and et al. Rt-1: Robotics transformer for real-world control at scale. In arXiv preprint arXiv:2212.06817, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 678 + ], + "type": "text", + "content": "[9] Elliot Chane-Sane, Cordelia Schmid, and Ivan Laptev. Learning video-conditioned policies for unseen manipulation tasks. In ICRA, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 681, + 294, + 713 + ], + "type": "text", + "content": "[10] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Proceedings of Robotics: Science and Systems (RSS), 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 98, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 163 + ], + "type": "text", + "content": "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 165, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 165, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 553, + 208 + ], + "type": "text", + "content": "[12] Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B. Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 211, + 553, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 211, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 317, + 211, + 553, + 254 + ], + "type": "text", + "content": "[13] Alircza Fathi, Jessica K. Hodgins, and James M. Rehg. Social interactions: A first-person perspective. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1226-1233, 2012. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 256, + 553, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 256, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 256, + 553, + 299 + ], + "type": "text", + "content": "[14] Thomas Feix, Javier Romero, Heinz-Bodo Schmiedmayer, Aaron M. Dollar, and Danica Kragic. The grasp taxonomy of human grasp types. IEEE Transactions on Human-Machine Systems, 46(1):66-77, 2016. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 302, + 553, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 302, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 317, + 302, + 553, + 345 + ], + "type": "text", + "content": "[15] Pete Florence, Corey Lynch, Andy Zeng, Oscar Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. Conference on Robot Learning (CoRL), 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 348, + 553, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 348, + 553, + 379 + ], + "spans": [ + { + "bbox": [ + 317, + 348, + 553, + 379 + ], + "type": "text", + "content": "[16] Kristen Grauman, Andrew Westbury, and et al. Ego4d: Around the world in 3,000 hours of egocentric video, 2022. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 382, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 382, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 382, + 553, + 404 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 406, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 406, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 553, + 437 + ], + "type": "text", + "content": "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 441, + 553, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 441, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 317, + 441, + 553, + 484 + ], + "type": "text", + "content": "[19] Vincent Tao Hu, Wenzhe Yin, Pingchuan Ma, Yunlu Chen, Basura Fernando, Yuki M Asano, Efstratios Gavves, Pascal Mettes, Bjorn Ommer, and Cees G. M. Snoek. Motion flow matching for human motion synthesis and editing, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 487, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 487, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 487, + 553, + 552 + ], + "type": "text", + "content": "[20] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, Igor Gilitschenski, Yonatan Bisk, and Debidatta Dwibedi. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 554, + 553, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 554, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 317, + 554, + 553, + 608 + ], + "type": "text", + "content": "[21] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. BC-z: Zero-shot task generalization with robotic imitation learning. In 5th Annual Conference on Robot Learning, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 611, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 611, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 611, + 553, + 643 + ], + "type": "text", + "content": "[22] Ananth Jonnavittula, Sagar Parekh, and Dylan P. Losey. View: Visual imitation learning with waypoints, 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[23] Siddharth Karamcheti, Suraj Nair, Annie S. Chen, Thomas Kollar, Chelsea Finn, Dorsa Sadigh, and Percy Liang. Language-driven representation learning for robotics. In Robotics: Science and Systems (RSS), 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Alexander Khazatsky, Karl Pertsch, and et al. Droid: A large-scale in-the-wild robot manipulation dataset, 2024. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27691" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[25] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B. Tenenbaum. Learning to act from actionless videos through dense correspondences, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "text", + "content": "[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[27] Yong Jae Lee, Joydeep Ghosh, and Kristen Grauman. Discovering important people and objects for egocentric video summarization. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1346-1353, 2012. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 205 + ], + "type": "text", + "content": "[28] Yin Li, Miao Liu, and James M. Rehg. In the eye of the beholder: Gaze and actions in first person video, 2020. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 205, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 272 + ], + "type": "text", + "content": "[29] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21013-21022, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 315 + ], + "type": "text", + "content": "[30] Zheng Lu and Kristen Grauman. Story-driven summarization for egocentric video. In 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 2714-2721, 2013. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "text", + "content": "[31] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. Mediapipe: A framework for building perception pipelines, 2019. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 381, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 381, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 294, + 426 + ], + "type": "text", + "content": "[32] Yecheng Jason Ma, William Liang, Vaidehi Som, Vikash Kumar, Amy Zhang, Osbert Bastani, and Dinesh Jayaraman. Liv: Language-image representations and rewards for robotic control, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[34] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3D with transformers. In CVPR, 2024. 4, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "text", + "content": "[35] Hamed Pirsiavash and Deva Ramanan. Detecting activities of daily living in first-person camera views. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2847-2854, 2012. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[36] Sigal Raab, Inbal Leibovitch, Guy Tevet, Moab Arar, Amit H Bermano, and Daniel Cohen-Or. Single motion diffusion. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 658 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "text", + "content": "[38] Weiming Ren, Huan Yang, Ge Zhang, Cong Wei, Xinrun Du, Wenhao Huang, and Wenhu Chen. Consisti2v: Enhancing visual consistency for image-to-video generation, 2024. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "text", + "content": "together. ACM Transactions on Graphics, 36(6):1-17, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "text", + "content": "[40] Dandan Shan, Jiaqi Geng, Michelle Shu, and David Fouhey. Understanding human hands in contact at internet scale. In CVPR, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 128, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 161 + ], + "type": "text", + "content": "[41] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[42] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Proceedings of the 6th Conference on Robot Learning (CoRL), 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 205, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 249 + ], + "type": "text", + "content": "[43] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos, 2018. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "type": "text", + "content": "[44] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "text", + "content": "[45] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "text", + "content": "[46] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. In DGS@ICLR, 2019. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[47] Boyang Wang, Nikhil Sridhar, Chao Feng, Mark Van der Merwe, Adam Fishman, Nima Fazeli, and Jeong Joon Park. This&that: Language-gesture controlled video generation for robot planning, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "text", + "content": "[48] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 459, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 553, + 503 + ], + "type": "text", + "content": "[49] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability, 2023. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "type": "text", + "content": "[50] Yin Wang, Zhiying Leng, Frederick W. B. Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 21978-21987, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[51] Zhiqiang Wang, Hao Zheng, Yunshuang Nie, Wenjun Xu, Qingwei Wang, Hua Ye, Zhe Li, Kaidong Zhang, Xuewen Cheng, Wanxi Dong, Chang Cai, Liang Lin, Feng Zheng, and Xiaodan Liang. All robots in one: A new standard and unified dataset for versatile, general-purpose embodied agents, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "text", + "content": "[52] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[53] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. In International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27692" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 566 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[54] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors, 2023. 2, 5, 7, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[55] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. ACM Transactions on Graphics (TOG), 43(6):1-11, 2024. 2, 6, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[56] Haoyu Xiong, Quanzhou Li, Yun-Chun Chen, Homanga Bharadhwaj, Samarth Sinha, and Animesh Garg. Learning by watching: Physical imitation of manipulation skills from human videos, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "text", + "content": "[57] Mengda Xu, Zhenjia Xu, Cheng Chi, Manuela Veloso, and Shuran Song. XSkill: Cross embodiment skill discovery. In 7th Annual Conference on Robot Learning, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[58] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. In 8th Annual Conference on Robot Learning, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[59] Zhao Yang, Bing Su, and Ji-Rong Wen. Synthesizing long-term human motions with diffusion models via coherent sampling, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 385 + ], + "type": "text", + "content": "[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer, 2024. 2, 7, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 419 + ], + "type": "text", + "content": "[61] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "text", + "content": "[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[63] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models, 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 499, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 294, + 532 + ], + "type": "text", + "content": "[64] Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 533, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 566 + ], + "type": "text", + "content": "[65] Yifeng Zhu, Arisrei Lim, Peter Stone, and Yuke Zhu. Vision-based manipulation from single human video with open-world object graphs, 2024. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27693" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TCFG_ Tangential Damping Classifier-free Guidance/8113438a-9689-405a-b6bc-61aa320e1894_content_list.json b/2025/TCFG_ Tangential Damping Classifier-free Guidance/8113438a-9689-405a-b6bc-61aa320e1894_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..13ce1138a21a28f19dcc46d2c85f51d070b88835 --- /dev/null +++ b/2025/TCFG_ Tangential Damping Classifier-free Guidance/8113438a-9689-405a-b6bc-61aa320e1894_content_list.json @@ -0,0 +1,1883 @@ +[ + { + "type": "text", + "text": "TCFG: Tangential Damping Classifier-free Guidance", + "text_level": 1, + "bbox": [ + 230, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingi Kwon*", + "bbox": [ + 209, + 181, + 320, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yonsei University", + "bbox": [ + 191, + 199, + 331, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kwonmingi@yonsei.ac.kr", + "bbox": [ + 161, + 219, + 359, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shin seong Kim*", + "bbox": [ + 429, + 181, + 566, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yonsei University", + "bbox": [ + 424, + 199, + 566, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tltydl2@yonsei.ac.kr", + "bbox": [ + 406, + 219, + 584, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaeseok Jeong", + "bbox": [ + 674, + 181, + 789, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yonsei University", + "bbox": [ + 661, + 199, + 803, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jete jeong@yonsei.ac.kr", + "bbox": [ + 632, + 219, + 833, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Ting Hsiao", + "bbox": [ + 323, + 242, + 437, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Michigan", + "bbox": [ + 287, + 260, + 473, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hsiaoyt@umich.edu", + "bbox": [ + 305, + 280, + 457, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngjung Uh", + "bbox": [ + 563, + 242, + 684, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yonsei University", + "bbox": [ + 553, + 260, + 696, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yj.uh@yonsei.ac.kr", + "bbox": [ + 540, + 280, + 702, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 328, + 326, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have achieved remarkable success in text-to-image synthesis, largely attributed to the use of classifier-free guidance (CFG), which enables high-quality, condition-aligned image generation. CFG combines the conditional score (e.g., text-conditioned) with the unconditional score to control the output. However, the unconditional score is in charge of estimating the transition between manifolds of adjacent timesteps from $x_{t}$ to $x_{t-1}$ , which may inadvertently interfere with the trajectory toward the specific condition. In this work, we introduce a novel approach that leverages a geometric perspective on the unconditional score to enhance CFG performance when conditional scores are available. Specifically, we propose a method that filters the singular vectors of both conditional and unconditional scores using singular value decomposition. This filtering process aligns the unconditional score with the conditional score, thereby refining the sampling trajectory to stay closer to the manifold. Our approach improves image quality with negligible additional computation. We provide deeper insights into the score function behavior in diffusion models and present a practical technique for achieving more accurate and contextually coherent image synthesis. project page", + "bbox": [ + 89, + 361, + 483, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 739, + 220, + 756 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [12, 31] have shown remarkable progress in image generation [19, 27, 30]. In particular, the emergence of classifier-free guidance [6, 11] (CFG) has attracted significant attention because it allows us to provide desired guidance by leveraging the conditional estimated score directly within the diffusion model.", + "bbox": [ + 89, + 766, + 482, + 856 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The classifier-free guidance fundamentally computes the", + "bbox": [ + 109, + 858, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/706fdab5475f7a29b260c1b5cf8387caad4554e208c01e6cc676adeecc163e42.jpg", + "image_caption": [ + "(a) Classifier-free guidance", + "Figure 1. (a) Classifier-free guidance. When the unconditional score $s_{\\theta}(\\boldsymbol{z}_t)$ and the conditional score $s_{\\theta}(\\boldsymbol{z}_t, y)$ are misaligned, the result of CFG tends to fall off the manifold. (b) Our proposed method reduces the misalignment between the unconditional score $s_{\\theta}(\\boldsymbol{z}_t)$ and the conditional score $s_{\\theta}(\\boldsymbol{z}_t, y)$ , ensuring sampling aligns with the target manifold." + ], + "image_footnote": [], + "bbox": [ + 517, + 329, + 705, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/482be4850a43f1cf1bc6a4faede5132a6dabcb2bcda05e78d8b2a17eabdffb60.jpg", + "image_caption": [ + "(b) Ours" + ], + "image_footnote": [], + "bbox": [ + 715, + 329, + 903, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "final score by combining the unconditional and conditional estimated scores. This approach ensures a generation that aligns well with the given condition. Additionally, using an appropriate guidance scale has been shown to enhance image quality across various tasks, further driving improvements in applications like text-to-image generation.", + "bbox": [ + 511, + 597, + 906, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Let us say the guided score $\\tilde{s}_{\\theta}$ as $\\tilde{s}_{\\theta} = s_{\\theta}^{\\mathrm{uncond}} + \\omega_{\\mathrm{scale}}(s_{\\theta}^{\\mathrm{cond}} - s_{\\theta}^{\\mathrm{uncond}})$ . In text-to-image models, the text condition $(s_{\\theta}^{\\mathrm{cond}})$ is randomly replaced with a null condition $(s_{\\theta}^{\\mathrm{uncond}})$ during training (e.g., with a probability $p = 0.1$ ), enabling the null condition to act as a general estimator for any sample. It means that $s_{\\theta}^{\\mathrm{uncond}}$ is the score estimated from $z_{t}$ to $z_{t-1}$ for all samples in the sampling trajectory of diffusion models.", + "bbox": [ + 511, + 703, + 905, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The unconditional score for any sample has certainly enabled the successful use of classifier-free guidance. However, we argue that there can be a misalignment between the unconditional and conditional estimated scores (See Eq. (2) in Sec. 2), which hinders the approximation toward the", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 109, + 887, + 215, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2620", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "manifold by the given condition. Fig. 1 (a) conceptually illustrates the potential issue that arises when the manifold of the unconditional score differs from that of the conditional score. In this paper, we show that this misalignment can be resolved with a simple algorithm, which significantly reduces the tendency of CFG to generate off-manifold samples, as illustrated in Fig. 1 (b).", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach is based on the following insights. First, the score predicted by the diffusion model estimates the intrinsic dimension of the data manifold [32]. Additionally, this intrinsic dimension can be captured by the tangent space of the target manifold [3, 9]. Instead of directly estimating the intrinsic dimension, we focus on utilizing the tangential component inherent in the unconditional score during classifier-free guidance. By reducing its misalignment with the conditional score, we enhance the alignment and ultimately improve the quality of the generated outputs.", + "bbox": [ + 89, + 198, + 482, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we push the score $\\tilde{s}_{\\theta}$ toward the normal direction of the conditional manifold by eliminating the values of column vectors with small singular values using the orthogonal matrix $V$ obtained through the singular value decomposition of the conditional and unconditional scores.", + "bbox": [ + 89, + 351, + 483, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel sampling method that leverages the unconditional score within CFG. To support our approach, we first lay out the theoretical foundation in section Sec. 2 and Sec. 3, discussing the manifold hypothesis and its connection to diffusion models. In Sec. 4, we provide a comprehensive explanation of our proposed method. This is followed by a detailed analysis using a toy example in Sec. 5, and we demonstrate the practical applicability of our method on real-world text-to-image models in Sec. 6.", + "bbox": [ + 89, + 429, + 483, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments show a significant improvement in the MS-COCO Fréchet Inception Distance (FID) across various models that utilize classifier-free guidance, e.g., diffusion models (Stable Diffusion v1.5 [26] and SDXL [23]) and rectified flow (Stable Diffusion 3 [8]). Additionally, our method improves DiT [21] FID on ImageNet. Notably, our method helps mitigate the overexposure bias problem, leading to resulting images that better align with the underlying data distribution, as supported by improved quantitative metrics.", + "bbox": [ + 89, + 566, + 483, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 89, + 734, + 218, + 752 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models Diffusion models learn the score that reverses the forward noising process. This forward process from the real data distribution $p(\\pmb{x}_0)$ to a latent distribution $p(z_1) \\sim N(0, \\sigma_{\\max}^2 I)$ along timesteps $t \\in [0, 1]$ is defined by a Gaussian kernel: $z_t = x_0 + \\sigma(t)\\epsilon$ . The function $\\sigma(t)$ is a noise schedule where $\\sigma(0) = 0$ and $\\sigma(1) = \\sigma_{\\max}$ , determining the amount of noise to be added at each timestep $t$ to erase information from $x$ .", + "bbox": [ + 89, + 762, + 482, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A generative process is represented as its reverse with a", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stochastic differential equation (SDE):", + "bbox": [ + 513, + 90, + 769, + 106 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {d} \\boldsymbol {z} = - \\dot {\\sigma} (t) \\sigma (t) \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\mathrm {d} t \\\\ - \\beta (t) \\sigma (t) ^ {2} \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) d t + \\sqrt {2 \\beta (t)} \\sigma (t) d \\omega_ {t}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 114, + 893, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathrm{d}\\omega_{t}$ is a standard Wiener process. Alternatively, it can be expressed as an ordinary differential equation:", + "bbox": [ + 511, + 164, + 903, + 194 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} \\boldsymbol {z} = - \\dot {\\sigma} (t) \\sigma (t) \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\mathrm {d} t.\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 204, + 823, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models approximate the score function $\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{z}_t)$ with a neural network $s_\\theta (\\boldsymbol {z}_t,t)$ . They are trained to predict the clean data from the noisy $\\boldsymbol{z}_t$ . The trained model performs the reverse process using:", + "bbox": [ + 511, + 229, + 903, + 291 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\approx \\frac {\\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t} , t) - \\boldsymbol {z} _ {t}}{\\sigma (t) ^ {2}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 300, + 820, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classifier guidance (CG) and classifier-free guidance (CFG) For an arbitrary class label $y$ , CG defines the class-conditional sampling distribution $\\tilde{p}_{\\theta}(\\boldsymbol{z}_t \\mid y)$ as:", + "bbox": [ + 511, + 347, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) \\propto p _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) p _ {\\theta} (y \\mid \\boldsymbol {z} _ {t}) ^ {\\gamma},\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 402, + 826, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $p_{\\theta}(y \\mid z_t)$ is the classifier distribution and $\\gamma$ is a scaling parameter. [6] When $\\gamma > 0$ , it is known to reduce sample diversity but enhance quality. However, CG requires a classifier that can predict label $y$ from the noisy $z_t$ . CFG proposes a method to sample from the conditional distribution by expressing the classifier distribution $p_{\\theta}(y \\mid z_t)$ in terms of the conditional distribution $p_{\\theta}(z_t \\mid y)$ and the unconditional distribution $p_{\\theta}(z_t)$ :", + "bbox": [ + 511, + 429, + 903, + 550 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) \\propto p _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) ^ {1 + \\gamma} p _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) ^ {- \\gamma}.\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 559, + 833, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a result, the final score $\\nabla_{z_t}\\log \\tilde{p}_\\theta (\\boldsymbol {z}_t\\mid y)$ is approximated by:", + "bbox": [ + 511, + 585, + 903, + 616 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\boldsymbol {z} _ {t}} \\log \\tilde {p} _ {\\theta} (\\boldsymbol {z} _ {t} \\mid y) = (1 + \\gamma) \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}, y) - \\gamma \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}) \\\\ = \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}) + \\omega \\left(\\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}, y) - \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t})\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 626, + 888, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\omega = 1 + \\gamma$ . [11]", + "bbox": [ + 511, + 671, + 669, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In practice, both $s_{\\theta}(x_t,y)$ and $s_{\\theta}(z_t)$ are approximated by a single neural network that is jointly trained to estimate both the conditional and unconditional scores. Text-to-image models use the null condition $y_{\\mathrm{null}} = \\varnothing$ as a class label to train $s_{\\theta}(z_t)\\approx s_{\\theta}(z_t,y_{\\mathrm{null}})$ . This approach allows $s_{\\theta}(z_t,y) - s_{\\theta}(z_t)$ to provide guidance similar to the gradient of an implicit classifier. Hereafter, we will simply denote $s_{\\theta}(z_t,y_{\\mathrm{null}})$ as $s_{\\theta}(z_t)$ .", + "bbox": [ + 511, + 686, + 903, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models and data manifold encoding The manifold hypothesis suggests that high-dimensional data lies on or near a lower-dimensional manifold, making intrinsic dimension estimation essential for data representation [9]. This intrinsic dimension is often encoded in the", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "manifold's tangent spaces, which capture underlying degrees of freedom and align local structures to reveal the global geometry [3, 33].", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on these ideas, further studies have analyzed the approximation and generalization capabilities of diffusion models [20, 22, 24], and have also proven that their score functions can approximate the tangent space of the data manifold [32]. In particular, for a compact embedded sub-manifold $\\mathcal{M} \\subset \\mathbb{R}^n$ , it has been shown that for a sample $z_t \\in \\mathbb{R}^n$ sufficiently close to the target data, the score $\\nabla_{z_t} \\log p_t(z_t) (\\approx s_\\theta(z_t))$ and orthogonal projection $\\pi(z_t)$ onto data manifold $\\mathcal{M}_0$ satisfy a key relationship. For the projection $\\mathbf{N}_p$ onto the normal space and $\\mathbf{T}_p$ onto the tangent space of $\\mathcal{M}_0$ , the ratio of their magnitudes goes to zero as $t$ approaches 0 (i.e., gets closer to the target data). In other words, for samples $z_t$ close to the target data, the following equation holds:", + "bbox": [ + 89, + 137, + 483, + 348 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\left\\| \\mathbf {T} _ {p} \\nabla_ {\\mathbf {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\right\\|}{\\left\\| \\mathbf {N} _ {p} \\nabla_ {\\mathbf {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\right\\|} \\rightarrow 0, \\quad \\text {a s} t \\rightarrow 0, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 358, + 482, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_p$ and $\\mathbf{N}_p$ are the projection operators onto the tangent space $\\mathcal{T}_{\\pi (z_t)}\\mathcal{M}_0$ and the normal space $\\mathcal{N}_{\\pi (z_t)}\\mathcal{M}_0$ , respectively (for a detailed proof, see Theorem 4.1, Corollary 4.2 and Appendix D in [32]).", + "bbox": [ + 89, + 402, + 482, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This implies that, for samples sufficiently close to the target manifold $\\mathcal{M}_0$ , the cosine similarity between the score function and the normal vector $\\mathbf{n} = \\frac{\\pi(z_t) - z_t}{\\|\\pi(z_t) - z_t\\|}$ converges to 1 (i.e., $S_{\\mathrm{cos}}(\\mathbf{n},\\nabla_{z_t}\\log p_t(z_t))\\xrightarrow{t\\to 0}1$ ).", + "bbox": [ + 89, + 465, + 483, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This suggests that for a sample $z_{t}$ very close to the target, the score function $\\nabla_{\\mathbf{z}}\\log p_t(z_t)\\approx s_\\theta (z_t)$ becomes an element of the normal space of the target manifold (that is, $\\nabla_{\\mathbf{z}}\\log p_t(z_t)\\in \\mathcal{N}_{\\pi (z_t)}\\mathcal{M}_0\\approx \\mathcal{N}_{\\pi (z_0)}\\mathcal{M}_0$ for sufficiently small $t$ ). Leveraging this property, the estimated diffusion score can approximate the intrinsic dimension of the target data by utilizing the huge gap in the singular values of the sampling scores $S = \\left[s_{\\theta}\\left(\\boldsymbol {z}_t^{(1)},t\\right),\\dots ,s_{\\theta}\\left(\\boldsymbol {z}_t^{(4n)},t\\right)\\right]$ where the singular vectors corresponding to the higher singular values represent the normal components of $\\mathcal{M}_0$ , while those corresponding to the lower singular values represent the tangential components.[32]", + "bbox": [ + 89, + 535, + 483, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Intuition", + "text_level": 1, + "bbox": [ + 89, + 736, + 191, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we assume the mathematical concept behind our method and supporting experiments. Our approach refines CFG at each step by dropping the tangential component of the unconditional score, enhancing the quality of conditional generation. This adjustment allows the conditional score to guide the generated sample more directly", + "bbox": [ + 89, + 761, + 483, + 853 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2dc8ff2f56920b200ea0696cc91ea6a29dd819c8d4c64eb7681009255a20c7e6.jpg", + "image_caption": [ + "Figure 2. Singular values of the score function across all timesteps. We computed the singular values for all timesteps using a total of 17,000 samples from Stable Diffusion v1.5. For both the unconditional and the conditional scores, a significant drop in singular values was observed at indices close to 0 across all timesteps. This suggests the existence of an intermediate manifold." + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "toward the manifold specified by the condition, improving alignment.", + "bbox": [ + 511, + 349, + 903, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To support this, we provide empirical evidence suggesting that not only does the target data manifold $\\mathcal{M}_0$ exist but there is also a manifold $\\mathcal{M}_{t - \\epsilon}$ at each time step $t\\in (0,1)$ where $\\nabla_{\\mathbf{z}}\\log p_t(\\pmb {z}_t)\\in \\mathcal{N}_{\\pi_{t - \\epsilon}(\\pmb {z}_t)}\\mathcal{M}_{t - \\epsilon}$ .", + "bbox": [ + 511, + 380, + 905, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There exists an intermediate manifold $\\mathcal{M}_t$ We hypothesize the existence of a manifold $\\mathcal{M}_{(t - \\epsilon)}$ that contains $\\nabla_{\\pmb {z}_t}\\log p_t(\\pmb {z}_t)$ as elements of its normal space, not only for samples close to the target data but also for $t\\in (0,1)$ . Specifically, we assume the following:", + "bbox": [ + 511, + 458, + 905, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assumption 1. Suppose that the support of the data distribution $P_0$ is contained in a compact embedded submanifold $\\mathcal{M}_0\\subset \\mathbb{R}^d$ , and let $P_{t}$ be the distribution of latents at time $t$ diffused from $P_0$ . Then, under mild assumptions², $\\forall t\\in (0,1)$ , $\\exists t'\\in (t - \\epsilon ,t + \\epsilon)$ such that:", + "bbox": [ + 511, + 542, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\in \\mathcal {N} _ {\\pi_ {t ^ {\\prime}} (\\boldsymbol {z} _ {t})} \\mathcal {M} _ {t ^ {\\prime}},\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 630, + 813, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for sufficiently small $\\epsilon$ and orthogonal projection $\\pi_{t'}(\\mathbf{z}_t)$ onto manifold $\\mathcal{M}_{t'}$ . This hypothesis is indirectly supported by the clear gap in singular values arranged in descending order for a sufficient number of samples. This phenomenon occurs not only when $t$ goes to 0 (near the image manifold) but also consistently for all time step $t \\in (0,1)$ .", + "bbox": [ + 511, + 657, + 905, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To observe the gap, we compute 17,000 score samples across all timesteps on Stable Diffusion v1.5. Let $[\\sigma_1,\\dots ,\\sigma_D]$ represent the singular values from the SVD applied to $s_{\\theta}(z_{t})$ and $s_{\\theta}(z_{t},y)$ , with 17,000 samples collected per timestep, arranged in descending order. The corresponding singular vectors are denoted as", + "bbox": [ + 511, + 763, + 905, + 854 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Every compact embedded submanifold of $\\mathbb{R}^d$ has a tubular neighborhood, and for a given manifold $\\mathcal{M}$ , each point $\\pmb{z} \\in \\mathbb{R}^n$ within this tubular neighborhood has a unique orthogonal projection $\\pi$ onto $\\mathcal{M}$ [16].", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2) The distribution $P_0$ has a smooth density $p_0$ w.r.t the volume measure on the manifold. 2) The density $p_0$ is bounded away from zero on the manifold.", + "bbox": [ + 511, + 862, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2cb5021bbd32ce4dcf5b043fddcc1d4dbfe00ea27913e7103fae31a5322fbd8c.jpg", + "image_caption": [ + "Figure 3. Cosine similarity between singular vectors of unconditional and conditional scores. We computed the singular vectors $V$ at each timestep using a total of 17,000 samples from Stable Diffusion v1.5. We observe the similarity of significant singular vectors (i.e., those with indices close to 0) between unconditional and conditional scores are mostly high across all timesteps $T$ ." + ], + "image_footnote": [], + "bbox": [ + 130, + 89, + 444, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$[\\pmb{v}_1,\\dots ,\\pmb{v}_D]^T$ for $\\pmb{s}_{\\theta}(\\pmb {z}_t)$ and $[\\hat{\\pmb{v}}_1,\\dots ,\\hat{\\pmb{v}}_D]^T$ for $s_\\theta (z_t,y)$ respectively $(D\\approx^{3}17,000)$", + "bbox": [ + 89, + 297, + 483, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, both $s_{\\theta}(z_t)$ and $s_{\\theta}(z_t,y)$ have gaps between the highest singular values and the rest for all $t \\in [0,1]$ , not just for $0 + \\epsilon$ . Interpreting from the perspective that the score function $s_{\\theta}(z_t)$ becomes an element of the data manifold's normal space as $t$ approaches 0 [32]. Assuming the existence of an intermediate manifold $\\mathcal{M}_t$ for all $t \\in (0,1)$ , this suggests that the singular vectors associated with the largest singular values contain dominant components of $\\mathcal{NM}_t$ , while vectors associated with smaller singular values correspond to component of $\\mathcal{T}\\mathcal{M}_t$ .", + "bbox": [ + 89, + 344, + 483, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tangential misalignment between unconditional and conditional score We empirically justify the principle of modifying the unconditional score by dropping the components with low singular values and retaining only the components with high singular values.", + "bbox": [ + 89, + 513, + 483, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 shows that conditional and unconditional singular vectors $[\\pmb{v}_1, \\dots, \\pmb{v}_D]^T$ and $[\\hat{\\pmb{v}}_1, \\dots, \\hat{\\pmb{v}}_D]^T$ at corresponding indices are more similar when their singular values are high than the rest.", + "bbox": [ + 89, + 590, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, the cosine similarity of the singular vectors $\\pmb{v}_1$ and $\\hat{\\pmb{v}}_1$ associated with the highest singular value $\\sigma_1$ from $s_\\theta(z_t)$ and $s_\\theta(z_t, y)$ , respectively, is higher than the others.", + "bbox": [ + 89, + 651, + 483, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left[ S _ {\\cos} \\left(\\boldsymbol {v} _ {1}, \\hat {\\boldsymbol {v}} _ {1}\\right) > S _ {\\cos} \\left(\\boldsymbol {v} _ {j}, \\hat {\\boldsymbol {v}} _ {j}\\right) \\right] \\\\ \\approx \\left[ S _ {\\cos} \\left(\\mathbf {N} _ {p} \\nabla_ {z _ {t}} \\log p _ {t} \\left(z _ {t}, y\\right), \\mathbf {N} _ {p} \\nabla_ {z _ {t}} \\log p _ {t} \\left(z _ {t}\\right)\\right) \\right. \\tag {2} \\\\ \\left. > S _ {\\cos} \\left(\\mathbf {T} _ {p} \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} \\left(\\boldsymbol {z} _ {t}, y\\right), \\mathbf {T} _ {p} \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} \\left(\\boldsymbol {z} _ {t}\\right)\\right) \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 720, + 480, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for $1 < j \\leq D$ . The cosine similarity $S_{cos}$ between two vectors $\\mathbf{v}_i$ and $\\mathbf{v}_j$ is defined as $S_{\\mathrm{cos}}(\\mathbf{v}_i, \\mathbf{v}_j) = \\frac{\\mathbf{v}_i \\cdot \\mathbf{v}_j}{\\|\\mathbf{v}_i\\| \\|\\mathbf{v}_j\\|}$ .", + "bbox": [ + 89, + 786, + 482, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This indicates that the intermediate manifolds associated with $\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{z}_t)$ and $\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{z}_t,y)$ share similar normal components, while their tangent components are relatively less aligned.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These less-aligned components interfere with the generative process, making it harder to align with the target manifold. We modify the unconditional score $s_{\\theta}(z_t)$ at each timestep by removing its tangential components that are less aligned with the conditional score $s_{\\theta}(z_t,y)$ . We provide detailed methods in the following section.", + "bbox": [ + 511, + 152, + 905, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 511, + 255, + 612, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our main method proceeds as follows. At each step, we take the predicted unconditional score $s_{\\theta}(z_t)$ and the conditional score $s_{\\theta}(z_t,y)$ and concatenate them into a score matrix $A = [s_{\\theta}(z_t),s_{\\theta}(z_t,y)]$ . Next, we perform SVD on $A$ , obtaining singular values and corresponding singular vectors that consider both components $s(z_{t})$ and $s(z_{t},y)$ . This results in singular vectors $[\\pmb {v}_1,\\pmb {v}_2,\\dots ,\\pmb {v}_D]^T$ where $\\pmb{v}_{1}$ is the normal component of both $s(z_{t})$ and $s(z_{t},y)$ . We project the unconditional score onto $\\pmb{v}_{1}$ and drop the rest.", + "bbox": [ + 511, + 280, + 905, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {s}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) = \\boldsymbol {s} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) \\cdot \\boldsymbol {V} ^ {T} \\cdot \\left[ \\boldsymbol {v} _ {1}, \\boldsymbol {0} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 426, + 903, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consequently, the singular vectors associated with high singular values in the score matrix $A$ retain the well-aligned, normal components of $s_{\\theta}(z_t)$ and $s_{\\theta}(z_t,y)$ , while those with lower singular values represent misaligned tangential components, which we set to zero in Eq. (3) to drop these components from the unconditional score. Next, we update the score $\\hat{s}_{\\theta}(z_t,y)$ with classifier-free guidance (CFG):", + "bbox": [ + 511, + 454, + 905, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {z} _ {t}} \\log \\hat {p} _ {t} \\left(\\boldsymbol {z} _ {t} | y\\right) = \\hat {\\boldsymbol {s}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) + w \\left(\\boldsymbol {s} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, y\\right) - \\hat {\\boldsymbol {s}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 571, + 903, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Tangential damping classifier-free guidance (TCFG)", + "bbox": [ + 514, + 612, + 903, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inputs: $\\mathbf{s}_{\\theta}(\\mathbf{z}_t)$ and $\\mathbf{s}_{\\theta}(\\mathbf{z}_t|\\mathbf{y})$ : predicted unconditional and conditional scores, $t \\in (0,1)$ : time step, $y$ : condition, $w$ : CFG scale.", + "bbox": [ + 511, + 643, + 905, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: $z_0$", + "bbox": [ + 514, + 690, + 596, + 705 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $t \\in (0, 1)$ do", + "2: Get $s_{\\theta}$ from $z_{t}$", + "3: Make score matrix $\\mathbf{A} = \\left[s_{\\theta}\\left(z_{t}\\right), s_{\\theta}\\left(z_{t}, y\\right)\\right]$", + "4: $(\\sigma_{i})_{i = 1}^{d},(\\pmb{w}_{i})_{i = 1}^{d},(\\pmb{v}_{i})_{i = 1}^{d}\\gets \\mathrm{SVD}(\\pmb {A})$", + "5: $\\hat{s}_{\\theta}(\\pmb{z}_t) = s_{\\theta}(\\pmb{z}_t)\\cdot \\pmb{V}^T\\cdot [\\pmb{v}_1,\\pmb {0}]$", + "6: (Dropping $\\mathbf{T}\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{x}_t))$", + "7: $\\hat{s}_{\\theta}(\\boldsymbol{z}_t, y) = \\hat{s}_{\\theta}(\\boldsymbol{z}_t) + w(\\boldsymbol{s}_{\\theta}(\\boldsymbol{z}_t, y) - \\hat{s}_{\\theta}(\\boldsymbol{z}_t))$", + "8: Update $z_{t}$", + "9: end for", + "10: Output $z_0$" + ], + "bbox": [ + 519, + 707, + 861, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(\\sigma_{i})_{i = 1}^{d},(\\mathbf{w}_{i})_{i = 1}^{d},(\\mathbf{v}_{i})_{i = 1}^{d}$ denote singular values, left and right singular vectors respectively.", + "bbox": [ + 540, + 861, + 903, + 892 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3Approximately $4\\times n$ samples are sufficient to accurately estimate the intrinsic dimension of the target manifold $\\mathcal{M}_0$ [32]. However, our goal is to verify the existence of a manifold where $\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol {z}_t)$ is an element of the normal space. Therefore, it suffices to observe the presence of a large gap in the singular value spectrum, thus $N < D$ is enough.", + "bbox": [ + 89, + 838, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2623", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0080c29da4b011faefc3d7b8d9ae5d398c2b87e041f9109d754939c67993f165.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 147, + 93, + 287, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/adb4f0bcc1d9f2ef2b23e95c1a0c96e5997f0fec96dbeb52c59bc11932d8932f.jpg", + "image_caption": [ + "(b) Classifier-free" + ], + "image_footnote": [], + "bbox": [ + 290, + 93, + 428, + 184 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0c1ba6a9bc0e9751c7e789c548f4184c8b7f6c4d6f6bcd950d9cfb30a497644.jpg", + "image_caption": [ + "(a) Conditional scores only", + "(c) Calculate SVD on all samples", + "Figure 4. Sampling results on different methods with diffusion model trained on two moons dataset. Our proposed methods (c, d) demonstrate a closer match to the target distribution compared to using conditional scores only or CFG. In (c), SVD is computed across all samples, while in (d), SVD is calculated separately for each pair of conditional and unconditional scores." + ], + "image_footnote": [], + "bbox": [ + 147, + 202, + 287, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c805e93368939886f4821b153ed857109d6328029ad5c26cd336f4b5990b73d4.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 290, + 202, + 428, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide a detailed algorithm in Algorithm 1.", + "bbox": [ + 89, + 419, + 410, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike traditional CFG update methods, $\\nabla_{\\boldsymbol{z}_t}\\log \\hat{p_t} (\\boldsymbol {z}_t|\\boldsymbol {y})$ drops tangential component from the unconditional score at each step. It prevents accumulating misaligned components from the unconditional score $s_\\theta (z_t)$ using the direction of the manifold defined by the given condition $y$ over time evolution. This concept is further illustrated with a simple distribution in Sec. 5, where the toy example clarifies the benefits of our methods.", + "bbox": [ + 89, + 434, + 483, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Toy example", + "text_level": 1, + "bbox": [ + 89, + 570, + 220, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We empirically verify our method on a toy problem, generating the two moons dataset. Experiments consist of the generated samples with different guidances including the original classifier-free guidance (CFG) and ours, and the sampling trajectories following their respective score functions.", + "bbox": [ + 89, + 595, + 482, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The target data distribution $p(X_0)$ consists of samples distributed along two distinct curves (moons). We trained a conditional diffusion model using a small neural network that receives a binary label $y \\in \\{0,1\\}$ for the two moons or $y = \\varnothing$ denoting the null condition. For detailed settings, please refer to the Appendix.", + "bbox": [ + 89, + 686, + 482, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4 shows the generated samples using four different guiding strategies. (a) uses only the conditional score $s_{\\theta}(z_t, y)$ . (b) uses the CFG score. (c) and (d) employ our guidance score at each step with multiple samples and one sample, respectively, to compute singular value decomposition (SVD) of the unconditional score $s_{\\theta}^{i}(z_t)$ and conditional score $s_{\\theta}^{i}(z_t, y)$ .", + "bbox": [ + 89, + 779, + 482, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to the result, generated samples using our", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/29c3ce1ec7ae83c8421d95b085b9588d7688c0e8c2061cc3f7fc5791669d81c4.jpg", + "image_caption": [ + "Trajectory of sample A", + "Figure 5. Visualization of the sampling trajectory. In CFG (orange path), the unconditional scores (red arrows) include components that point towards directions other than the target distribution, making the final destination deviate from the target distribution. Whereas, our method (green path) removes the inconsistent tangent components in unconditional scores and eventually reaches the target distribution." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 686, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7721ebb674467b5145386389dceb2a9fdae583bbf25c194c6a0157d1a31c6571.jpg", + "image_caption": [ + "Trajectory of sample B" + ], + "image_footnote": [], + "bbox": [ + 691, + 90, + 901, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "strategies lie closer to the target compared to those generated using only the conditional score or CFG. CFG, while potentially bringing samples closer to the target than merely using conditional scores, may face challenges due to the misalignment of tangent components between unconditional score and conditional scores.", + "bbox": [ + 511, + 354, + 903, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, our guidance score can reduce the tangent component of the unconditional score at each step. This helps samples converge more effectively towards the target data, which suggests that the tangent components of the unconditional score might hinder alignment with the target data manifold under the given condition, and our method helps in mitigating this misalignment.", + "bbox": [ + 511, + 446, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further validate this hypothesis by examining the trajectories of generated samples. Fig. 5 visualizes the trajectories induced by our score $\\nabla_{\\boldsymbol{z}_t}\\log \\hat{p}_t(\\boldsymbol{z}_t|\\boldsymbol{y})$ compared to the original CFG score $\\nabla_{\\boldsymbol{z}_t}\\log \\tilde{p}_{\\theta}(\\boldsymbol{z}_t|\\boldsymbol{y})$ . As shown, in the orange CFG trajectory, the direction of unconditional scores changes frequently. This results in difficulties for the blue conditional score to maintain an orthogonal direction relative to the target manifold near the target distribution. In contrast, our method consistently adjusts the score to predict in a direction closer to orthogonal with respect to the target manifold, particularly as the samples converge toward the target data. Our method removes the tangential component of the unconditional score with respect to the manifold of the conditional score. This results in a direction that leans either to the right or to the left.", + "bbox": [ + 511, + 554, + 905, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, the similar results between (c) and (d) in Fig. 4 suggest that computing SVD for only a single sample is sufficient to yield nearly the same result.", + "bbox": [ + 511, + 781, + 903, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 511, + 844, + 643, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we demonstrate that our method is applicable to high-dimensional diffusion models. We employ rep", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ccba861f583ce3a189899f29ff2ee16bf07bfa88c176fde0aa7a76496423a511.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FID ↓CLIPScore ↑
SD v1.5original13.260.31
+ ours13.120.31
SDXLoriginal13.360.32
+ ours12.650.32
SD v3original16.660.32
+ ours13.740.32
", + "bbox": [ + 169, + 88, + 405, + 170 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6631c95025030ada415861ff4e4a4ef3904f72a2eecda090b8caa6751e291427.jpg", + "table_caption": [ + "Table 1. Zero-shot FID and CLIPScore measured on MSCOCO 30k. Our method consistently improves FID across all models—Stable Diffusion v1.5, SDXL, and SD v3—while maintaining a nearly identical CLIPScore." + ], + "table_footnote": [], + "table_body": "
FID ↓sFID ↓Precision ↑Recall ↑IS ↑
DiT32.6717.920.900.13271.1
DiT+ours29.513.270.900.19270.0
", + "bbox": [ + 91, + 242, + 480, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Evaluation metrics measured on ImageNet 50k using DiT. Our method achieves better performance in FID, sFID, Precision, and Recall while showing a slight decrease in Inception Score.", + "bbox": [ + 89, + 297, + 480, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "representative diffusion models such as Stable Diffusion v1.5 [27] and SDXL [23], and showed that it functions identically on SD v3 [7], which is based on Rectified Flow. Additionally, we conducted experiments on DiT [21], which is trained on ImageNet [5].", + "bbox": [ + 89, + 344, + 482, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental details For the text-to-image models, we used zero-shot FID [10] and CLIPScore [25] on the MSCOCO 2014 validation set [17] consisting of 30,000 images under the commonly used text-to-image evaluation protocols. [7, 23, 27] For DiT, we evaluated using 50,000 images under the same settings as ADM [6]. All models used the official pretrained weights, and sampling was performed using the same latent codes. We used the best CFG scales as the default value of each repository. Our method does not increase the inference time of all baselines.", + "bbox": [ + 88, + 448, + 482, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative evaluation Tab. 1 presents the FID and CLIP Scores for SD1.5, SDXL, and SD3. Our method achieved better FID scores while maintaining the same CLIP Scores across all three models. Notably, the decrease in FID is larger for SDXL compared to SD1.5, and even larger for SD3 compared to SDXL. We speculate that this is because SD3, known as a better model publicly, has a relatively clearer manifold. Furthermore, the results on SD3 demonstrate that our method is applicable not only to diffusion models but also to all CFG-based score functions, including those based on Rectified Flow. Fig. 6 also shows FID-CLIP curves on SDXL, demonstrating that FID improves even as the CFG scale changes.", + "bbox": [ + 88, + 627, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 shows the results on the DiT model. Except for a slight decrease in Inception Score, our method exhibits relatively superior performance in FID, sFID, and Recall. This indicates that our method can be equally applied to both text-to-image generation and class-conditioned generation.", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/952ac4373792e71419fd0a50bfed71ff1329eff7e94bc493d97eca39aad03288.jpg", + "image_caption": [ + "Figure 6. FID-CLIP curves on SDXL with 50 sampling steps." + ], + "image_footnote": [], + "bbox": [ + 573, + 89, + 844, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative evaluation Our method drops the tangential component from the unconditional score while retaining the normal component. This reduces misalignment with the conditional score, thereby improving image quality as shown in Fig. 7. Specifically, the changes introduced by our approach transform \"strange\" objects or scenes into more \"plausible\" images. This indirectly demonstrates that the misalignment of the unconditional score in the conventional CFG was causing the \"strange\" aspects in the final outputs.", + "bbox": [ + 511, + 207, + 903, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For example, our method converts physically impossible or unusual combinations of objects (SD3), uncommon appearances or characteristics (SDXL), and ambiguous shapes or forms (SD1.5) into \"normal\" results.", + "bbox": [ + 511, + 343, + 903, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 8 presents the results obtained from DiT. We observed that our method causes relatively more changes in the images generated by DiT. We speculate that this is because DiT is trained on ImageNet dataset with class labels. The results qualitatively show that when using our method, DiT generates images that are more detailed, have better structure, and appear more natural.", + "bbox": [ + 511, + 405, + 903, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, in both text-to-image and class-conditioned image generation, we observed a reduction in the overexposure bias problem. We attribute this improvement to the mitigation of misalignment between the unconditional score and the conditional score.", + "bbox": [ + 511, + 512, + 903, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What happened to the unconditional score? In this paragraph, we qualitatively demonstrate that the misalignment with the conditional score is reduced when we drop the tangential component from the unconditional score and retain the normal component. We compared the results sampled using CFG with those sampled using the null condition (i.e., unconditional) when generating images from the same random noise (i.e., latent variables) in SDXL. In our method, we used the text condition to compute $\\hat{s}$ but used only $\\hat{s}$ for denoising; that is, we set $\\omega = 0$ . Although this approach does not perfectly explain our method, we can indirectly infer its role by observing how the modified null condition changes.", + "bbox": [ + 511, + 612, + 903, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 9 shows that images sampled using the original null condition generate different objects such as trees, snowy mountain landscapes, and women. In contrast, images generated using our modified null condition $\\hat{s}$ show that the tree part takes the form of a feather, the snowy mountain landscape changes into a woman, and the woman transforms", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e5b21f654b98856ed13f4c4741391164a1bb465d129ab605085b553a340d29ed.jpg", + "image_caption": [ + "SD3", + "SD3 + ours", + "Figure 7. Qualitative evaluation of text-to-image models. Our method prevents overexposure, enhancing the shapes and details of objects." + ], + "image_footnote": [], + "bbox": [ + 98, + 107, + 362, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/273e6e2ef619d99e4436f3ec5c960d1c30da51ece63a5cee1b960d4cb90916bc.jpg", + "image_caption": [ + "SDXL", + "SDXL + ours" + ], + "image_footnote": [], + "bbox": [ + 366, + 107, + 630, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4f3e9c0787e8993a4d886f81698aecd5a281a054ae02c87683cee05ec36ba7d.jpg", + "image_caption": [ + "SD1.5", + "SD1.5 + ours" + ], + "image_footnote": [], + "bbox": [ + 635, + 107, + 900, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90389e3e52e478f073e557f056826834da79adebf095074e7a2a441af34ae4fd.jpg", + "image_caption": [ + "DiT", + "Figure 8. Qualitative evaluation of DiT Our method mitigates overexposure and enhances object shapes and details in DiT models trained on ImageNet." + ], + "image_footnote": [], + "bbox": [ + 96, + 474, + 285, + 693 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b54d34de12e13b69d130a5ca60a9efcf1c33cef56c68e8ae8359e10a430f68c1.jpg", + "image_caption": [ + "DiT $^+$ ours" + ], + "image_footnote": [], + "bbox": [ + 289, + 474, + 478, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "into a shape resembling a glove. These changes align with the objects we aim to generate: a feather, a woman, and a baseball glove. We observe that these changes due to the null condition help eliminate unwanted structures or artifacts in the generated images. In other words, we demonstrate that the misalignment of the null condition is reduced, and we claim that this improvement aids in image generation.", + "bbox": [ + 89, + 780, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a43183187866b9cb58cc54fe7e11d5e55d97cd3390bf524e310d4f84f68f81ac.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Quantitative comparison with existing baselines. The evaluation was conducted on ${30}\\mathrm{k}$ images from the MS-COCO dataset using the official code; SD v1.4 for SAG,SD v1.5 for PAG and SDXL for CFG++." + ], + "table_body": "
SAGSAG+TCFGPAGPAG+TCFGCFGPPCFGPP+TCFG
FID13.5311.4814.4511.8713.9713.44
CLIP Score0.310.300.310.310.320.32
", + "bbox": [ + 514, + 458, + 901, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Related work", + "text_level": 1, + "bbox": [ + 513, + 547, + 651, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Calssifier-free guidance Experimental methods to enhance the performance of Classifier-Free Guidance (CFG) have been studied. SAG [13] proposed a method to improve CFG by using intermediate self-attention maps. PAG [2] suggested computing CFG by transforming self-attention maps into identity matrices. ICG [29] enhanced CFG by utilizing random text embeddings. Recently, CFG++ [4] demonstrated better performance by modifying the CFG computation method. Our proposed approach modifies the unconditional score based on the conditional score and can be used alongside these existing works; please refer to Tab. 3 and the appendix for more results.", + "bbox": [ + 511, + 574, + 906, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Manifold hypothesis and diffusion There are also several studies that have utilized the manifold hypothesis properties of the score function estimated by diffusion models to address various inherent challenges associated with diffusion processes. One approach introduces the manifold memorization hypothesis to understand model memorization through the relationship between data and model manifold dimensionalities [28]. Another extends memorization", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2626", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/256d319284fa09a96fd3418b3acd5ca0fa7702c1be6860fc8d62d9a903fad4d3.jpg", + "image_caption": [ + "Figure 9. TCFG reduces misalignments between unconditional and conditional generation. Starting from the same random noise $z_{1}$ , when SDXL samples images with only the unconditional score, it produces random images such as trees, snowy mountain landscapes, and women. In contrast, our modified unconditional score, projected on dominant (conditional), generates images that somewhat match the desired text prompts. This is because our method reduces misalignment with the conditional score by dropping the tangential components of the unconditional score. Once the misalignment decreases, the quality of the final images (unconditional + conditional score) improves: The base of the feather has a more natural structure, the human arm appears more natural, and the extra string on the left side of the baseball glove is removed." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 906, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82d53a3b354e00e8625ecb192fcca134ccaf1375e60e7c705329562695dbb71f.jpg", + "image_caption": [ + "Figure 10. Limitations Our method occasionally struggles to fix severely wrong regions in the baseline samples." + ], + "image_footnote": [], + "bbox": [ + 155, + 417, + 418, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "theory to diffusion models [1], showing that high-variance subspaces are selectively lost due to memorization effects. Separately, different researchers proposed an approach for detecting synthetic images generated by diffusion models, achieving high accuracy across diverse datasets [18].", + "bbox": [ + 89, + 674, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Discussion and conclusion", + "text_level": 1, + "bbox": [ + 89, + 768, + 333, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work experimentally analyzes the issues arising in the standard CFG method, where the tangential component of the unconditional score does not align well with that of the conditional score. By using SVD to drop the tangential component in the unconditional score, we effectively improve text-to-image generation quality. Additionally, our CFG method is easily applicable, has low computational", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cost, and enhances image quality. We leverage the ability of the diffusion model's score function to encode the intrinsic dimension of the target data, demonstrating the misalignment between the conditional and unconditional scores to improve sampling quality. This is the first attempt to utilize this misalignment to enhance sampling.", + "bbox": [ + 511, + 416, + 906, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite these advantages, several unresolved issues remain. First, it is uncertain whether the misalignment of tangential component and the alignment of normal component between the predicted unconditional score $s_{\\theta}(z_t)$ and the conditional score $s_{\\theta}(z_t,y)$ in the CFG setting, at a given timestep $t$ , would similarly apply to the features derived from a separately trained classifier and the null condition score in the classifier guidance setting. Second, while our task leverages the capability of diffusion models to estimate intrinsic dimensions for enhancing conditional sampling methods, we present only experimental observations regarding the existence of an intermediate manifold for $t \\in [0,1]$ , without theoretical proof. Further exploration and rigorous analysis of these aspects are left as future work. Third, additional investigation is needed to adapt our approach effectively in the context of diffusion distillation using CFG scale as an input [14, 15], which we also identify as a promising direction for future research.", + "bbox": [ + 511, + 522, + 908, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, although our method successfully demonstrated on-manifold image generation, we observed that when the original image exhibits significant abnormalities, substantial changes may occasionally cause the structure to break down. Fig. 10 illustrates such examples. Nevertheless, it is evident that our method transforms \"strange\" images into more \"normal\" ones.", + "bbox": [ + 511, + 795, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Beatrice Achilli, Enrico Ventura, Gianluigi Silvestri, Bao Pham, Gabriel Raya, Dmitry Krotov, Carlo Lucibello, and Luca Ambrogioni. Losing dimensions: Geometric memorization in generative diffusion. arXiv preprint arXiv:2410.08727, 2024. 8", + "[2] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 7, 2", + "[3] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 2, 3", + "[4] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 7, 2", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2, 6", + "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 6", + "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2", + "[9] Charles Fefferman, Sanjoy Mitter, and Hariharan Narayanan. Testing the manifold hypothesis. Journal of the American Mathematical Society, 29(4):983-1049, 2016. 2", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1", + "[13] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 7, 1" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yi-Ting Hsiao, Siavash Khodadadeh, Kevin Duarte, Wei-An Lin, Hui Qu, Mingi Kwon, and Ratheesh Kalarot. Plug-and-play diffusion distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13743-13752, 2024. 8", + "[15] Black Forest Labs. FLUX. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-15. 8", + "[16] John M Lee. Introduction to Riemannian manifolds. Springer, 2018. 3", + "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[18] Peter Lorenz, Ricard L Durall, and Janis Keuper. Detecting images generated by deep diffusion models using their local intrinsic dimensionality. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 448-459, 2023. 8", + "[19] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[20] Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion models are minimax optimal distribution estimators. In International Conference on Machine Learning, pages 26517-26582. PMLR, 2023. 3", + "[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 6", + "[22] Jakiw Pidstrigach. Score-based generative models detect manifolds. Advances in Neural Information Processing Systems, 35:35852-35865, 2022. 3", + "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6", + "[24] Phillip Pope, Chen Zhu, Ahmed Abdelkader, Micah Goldblum, and Tom Goldstein. The intrinsic dimension of images and its impact on learning. arXiv preprint arXiv:2104.08894, 2021. 3", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2628", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 6", + "[28] Brendan Leigh Ross, Hamidreza Kamkari, Tongzi Wu, Rasa Hosseinzadeh, Zhaoyan Liu, George Stein, Jesse C Cresswell, and Gabriel Loaiza-Ganem. A geometric framework for understanding memorization in generative models. arXiv preprint arXiv:2411.00113, 2024. 7", + "[29] Seyedmorteza Sadat, Manuel Kansy, Otmar Hilliges, and Romann M Weber. No training, no problem: Rethinking classifier-free guidance for diffusion models. arXiv preprint arXiv:2407.02687, 2024. 7", + "[30] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1", + "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1", + "[32] Jan Pawel Stanczuk, Georgios Batzolis, Teo Deveney, and Carola-Bibiane Schonlieb. Diffusion models encode the intrinsic dimension of data manifolds. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4", + "[33] Zhenyue Zhang and Hongyuan Zha. Principal manifolds and nonlinear dimensionality reduction via tangent space alignment. SIAM journal on scientific computing, 26(1):313-338, 2004. 3" + ], + "bbox": [ + 91, + 92, + 485, + 527 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2629", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TCFG_ Tangential Damping Classifier-free Guidance/8113438a-9689-405a-b6bc-61aa320e1894_model.json b/2025/TCFG_ Tangential Damping Classifier-free Guidance/8113438a-9689-405a-b6bc-61aa320e1894_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8217bea306e074634acfe486fab52a76d7bd5850 --- /dev/null +++ b/2025/TCFG_ Tangential Damping Classifier-free Guidance/8113438a-9689-405a-b6bc-61aa320e1894_model.json @@ -0,0 +1,2519 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.232, + 0.131, + 0.768, + 0.154 + ], + "angle": 0, + "content": "TCFG: Tangential Damping Classifier-free Guidance" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.182, + 0.321, + 0.199 + ], + "angle": 0, + "content": "Mingi Kwon*" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.2, + 0.333, + 0.217 + ], + "angle": 0, + "content": "Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.22, + 0.361, + 0.233 + ], + "angle": 0, + "content": "kwonmingi@yonsei.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.182, + 0.567, + 0.199 + ], + "angle": 0, + "content": "Shin seong Kim*" + }, + { + "type": "text", + "bbox": [ + 0.426, + 0.2, + 0.568, + 0.217 + ], + "angle": 0, + "content": "Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.22, + 0.586, + 0.234 + ], + "angle": 0, + "content": "tltydl2@yonsei.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.675, + 0.183, + 0.79, + 0.199 + ], + "angle": 0, + "content": "Jaeseok Jeong" + }, + { + "type": "text", + "bbox": [ + 0.662, + 0.2, + 0.804, + 0.217 + ], + "angle": 0, + "content": "Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.22, + 0.834, + 0.234 + ], + "angle": 0, + "content": "jete jeong@yonsei.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.243, + 0.439, + 0.259 + ], + "angle": 0, + "content": "Yi Ting Hsiao" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.261, + 0.475, + 0.278 + ], + "angle": 0, + "content": "University of Michigan" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.281, + 0.458, + 0.294 + ], + "angle": 0, + "content": "hsiaoyt@umich.edu" + }, + { + "type": "text", + "bbox": [ + 0.565, + 0.243, + 0.686, + 0.26 + ], + "angle": 0, + "content": "Youngjung Uh" + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.261, + 0.697, + 0.278 + ], + "angle": 0, + "content": "Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.281, + 0.704, + 0.294 + ], + "angle": 0, + "content": "yj.uh@yonsei.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.329, + 0.327, + 0.345 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.485, + 0.711 + ], + "angle": 0, + "content": "Diffusion models have achieved remarkable success in text-to-image synthesis, largely attributed to the use of classifier-free guidance (CFG), which enables high-quality, condition-aligned image generation. CFG combines the conditional score (e.g., text-conditioned) with the unconditional score to control the output. However, the unconditional score is in charge of estimating the transition between manifolds of adjacent timesteps from \\( x_{t} \\) to \\( x_{t-1} \\), which may inadvertently interfere with the trajectory toward the specific condition. In this work, we introduce a novel approach that leverages a geometric perspective on the unconditional score to enhance CFG performance when conditional scores are available. Specifically, we propose a method that filters the singular vectors of both conditional and unconditional scores using singular value decomposition. This filtering process aligns the unconditional score with the conditional score, thereby refining the sampling trajectory to stay closer to the manifold. Our approach improves image quality with negligible additional computation. We provide deeper insights into the score function behavior in diffusion models and present a practical technique for achieving more accurate and contextually coherent image synthesis. project page" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.741, + 0.222, + 0.757 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.767, + 0.483, + 0.857 + ], + "angle": 0, + "content": "Diffusion models [12, 31] have shown remarkable progress in image generation [19, 27, 30]. In particular, the emergence of classifier-free guidance [6, 11] (CFG) has attracted significant attention because it allows us to provide desired guidance by leveraging the conditional estimated score directly within the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.859, + 0.483, + 0.875 + ], + "angle": 0, + "content": "The classifier-free guidance fundamentally computes the" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.33, + 0.707, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.473, + 0.688, + 0.487 + ], + "angle": 0, + "content": "(a) Classifier-free guidance" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.33, + 0.904, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.473, + 0.855, + 0.486 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.582 + ], + "angle": 0, + "content": "Figure 1. (a) Classifier-free guidance. When the unconditional score \\( s_{\\theta}(\\boldsymbol{z}_t) \\) and the conditional score \\( s_{\\theta}(\\boldsymbol{z}_t, y) \\) are misaligned, the result of CFG tends to fall off the manifold. (b) Our proposed method reduces the misalignment between the unconditional score \\( s_{\\theta}(\\boldsymbol{z}_t) \\) and the conditional score \\( s_{\\theta}(\\boldsymbol{z}_t, y) \\), ensuring sampling aligns with the target manifold." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.689 + ], + "angle": 0, + "content": "final score by combining the unconditional and conditional estimated scores. This approach ensures a generation that aligns well with the given condition. Additionally, using an appropriate guidance scale has been shown to enhance image quality across various tasks, further driving improvements in applications like text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Let us say the guided score \\(\\tilde{s}_{\\theta}\\) as \\(\\tilde{s}_{\\theta} = s_{\\theta}^{\\mathrm{uncond}} + \\omega_{\\mathrm{scale}}(s_{\\theta}^{\\mathrm{cond}} - s_{\\theta}^{\\mathrm{uncond}})\\). In text-to-image models, the text condition \\((s_{\\theta}^{\\mathrm{cond}})\\) is randomly replaced with a null condition \\((s_{\\theta}^{\\mathrm{uncond}})\\) during training (e.g., with a probability \\(p = 0.1\\)), enabling the null condition to act as a general estimator for any sample. It means that \\(s_{\\theta}^{\\mathrm{uncond}}\\) is the score estimated from \\(z_{t}\\) to \\(z_{t-1}\\) for all samples in the sampling trajectory of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The unconditional score for any sample has certainly enabled the successful use of classifier-free guidance. However, we argue that there can be a misalignment between the unconditional and conditional estimated scores (See Eq. (2) in Sec. 2), which hinders the approximation toward the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.217, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2620" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "manifold by the given condition. Fig. 1 (a) conceptually illustrates the potential issue that arises when the manifold of the unconditional score differs from that of the conditional score. In this paper, we show that this misalignment can be resolved with a simple algorithm, which significantly reduces the tendency of CFG to generate off-manifold samples, as illustrated in Fig. 1 (b)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.351 + ], + "angle": 0, + "content": "Our approach is based on the following insights. First, the score predicted by the diffusion model estimates the intrinsic dimension of the data manifold [32]. Additionally, this intrinsic dimension can be captured by the tangent space of the target manifold [3, 9]. Instead of directly estimating the intrinsic dimension, we focus on utilizing the tangential component inherent in the unconditional score during classifier-free guidance. By reducing its misalignment with the conditional score, we enhance the alignment and ultimately improve the quality of the generated outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.427 + ], + "angle": 0, + "content": "Specifically, we push the score \\(\\tilde{s}_{\\theta}\\) toward the normal direction of the conditional manifold by eliminating the values of column vectors with small singular values using the orthogonal matrix \\(V\\) obtained through the singular value decomposition of the conditional and unconditional scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.565 + ], + "angle": 0, + "content": "In this paper, we propose a novel sampling method that leverages the unconditional score within CFG. To support our approach, we first lay out the theoretical foundation in section Sec. 2 and Sec. 3, discussing the manifold hypothesis and its connection to diffusion models. In Sec. 4, we provide a comprehensive explanation of our proposed method. This is followed by a detailed analysis using a toy example in Sec. 5, and we demonstrate the practical applicability of our method on real-world text-to-image models in Sec. 6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Our experiments show a significant improvement in the MS-COCO Fréchet Inception Distance (FID) across various models that utilize classifier-free guidance, e.g., diffusion models (Stable Diffusion v1.5 [26] and SDXL [23]) and rectified flow (Stable Diffusion 3 [8]). Additionally, our method improves DiT [21] FID on ImageNet. Notably, our method helps mitigate the overexposure bias problem, leading to resulting images that better align with the underlying data distribution, as supported by improved quantitative metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.736, + 0.22, + 0.753 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.883 + ], + "angle": 0, + "content": "Diffusion models Diffusion models learn the score that reverses the forward noising process. This forward process from the real data distribution \\( p(\\pmb{x}_0) \\) to a latent distribution \\( p(z_1) \\sim N(0, \\sigma_{\\max}^2 I) \\) along timesteps \\( t \\in [0, 1] \\) is defined by a Gaussian kernel: \\( z_t = x_0 + \\sigma(t)\\epsilon \\). The function \\( \\sigma(t) \\) is a noise schedule where \\( \\sigma(0) = 0 \\) and \\( \\sigma(1) = \\sigma_{\\max} \\), determining the amount of noise to be added at each timestep \\( t \\) to erase information from \\( x \\)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "A generative process is represented as its reverse with a" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.771, + 0.107 + ], + "angle": 0, + "content": "stochastic differential equation (SDE):" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.116, + 0.895, + 0.155 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {d} \\boldsymbol {z} = - \\dot {\\sigma} (t) \\sigma (t) \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\mathrm {d} t \\\\ - \\beta (t) \\sigma (t) ^ {2} \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) d t + \\sqrt {2 \\beta (t)} \\sigma (t) d \\omega_ {t}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.165, + 0.905, + 0.195 + ], + "angle": 0, + "content": "where \\(\\mathrm{d}\\omega_{t}\\) is a standard Wiener process. Alternatively, it can be expressed as an ordinary differential equation:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.205, + 0.825, + 0.221 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} \\boldsymbol {z} = - \\dot {\\sigma} (t) \\sigma (t) \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\mathrm {d} t.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.231, + 0.905, + 0.292 + ], + "angle": 0, + "content": "Diffusion models approximate the score function \\(\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{z}_t)\\) with a neural network \\(s_\\theta (\\boldsymbol {z}_t,t)\\). They are trained to predict the clean data from the noisy \\(\\boldsymbol{z}_t\\). The trained model performs the reverse process using:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.301, + 0.821, + 0.333 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\approx \\frac {\\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t} , t) - \\boldsymbol {z} _ {t}}{\\sigma (t) ^ {2}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.905, + 0.394 + ], + "angle": 0, + "content": "Classifier guidance (CG) and classifier-free guidance (CFG) For an arbitrary class label \\(y\\), CG defines the class-conditional sampling distribution \\(\\tilde{p}_{\\theta}(\\boldsymbol{z}_t \\mid y)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.404, + 0.828, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) \\propto p _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) p _ {\\theta} (y \\mid \\boldsymbol {z} _ {t}) ^ {\\gamma},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.905, + 0.551 + ], + "angle": 0, + "content": "where \\(p_{\\theta}(y \\mid z_t)\\) is the classifier distribution and \\(\\gamma\\) is a scaling parameter. [6] When \\(\\gamma > 0\\), it is known to reduce sample diversity but enhance quality. However, CG requires a classifier that can predict label \\(y\\) from the noisy \\(z_t\\). CFG proposes a method to sample from the conditional distribution by expressing the classifier distribution \\(p_{\\theta}(y \\mid z_t)\\) in terms of the conditional distribution \\(p_{\\theta}(z_t \\mid y)\\) and the unconditional distribution \\(p_{\\theta}(z_t)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.56, + 0.834, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) \\propto p _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) ^ {1 + \\gamma} p _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) ^ {- \\gamma}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.586, + 0.904, + 0.617 + ], + "angle": 0, + "content": "As a result, the final score \\(\\nabla_{z_t}\\log \\tilde{p}_\\theta (\\boldsymbol {z}_t\\mid y)\\) is approximated by:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.627, + 0.889, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\boldsymbol {z} _ {t}} \\log \\tilde {p} _ {\\theta} (\\boldsymbol {z} _ {t} \\mid y) = (1 + \\gamma) \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}, y) - \\gamma \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}) \\\\ = \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}) + \\omega \\left(\\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}, y) - \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t})\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.672, + 0.671, + 0.687 + ], + "angle": 0, + "content": "where \\(\\omega = 1 + \\gamma\\). [11]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.905, + 0.808 + ], + "angle": 0, + "content": "In practice, both \\( s_{\\theta}(x_t,y) \\) and \\( s_{\\theta}(z_t) \\) are approximated by a single neural network that is jointly trained to estimate both the conditional and unconditional scores. Text-to-image models use the null condition \\( y_{\\mathrm{null}} = \\varnothing \\) as a class label to train \\( s_{\\theta}(z_t)\\approx s_{\\theta}(z_t,y_{\\mathrm{null}}) \\). This approach allows \\( s_{\\theta}(z_t,y) - s_{\\theta}(z_t) \\) to provide guidance similar to the gradient of an implicit classifier. Hereafter, we will simply denote \\( s_{\\theta}(z_t,y_{\\mathrm{null}}) \\) as \\( s_{\\theta}(z_t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Diffusion models and data manifold encoding The manifold hypothesis suggests that high-dimensional data lies on or near a lower-dimensional manifold, making intrinsic dimension estimation essential for data representation [9]. This intrinsic dimension is often encoded in the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2621" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "manifold's tangent spaces, which capture underlying degrees of freedom and align local structures to reveal the global geometry [3, 33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.349 + ], + "angle": 0, + "content": "Building on these ideas, further studies have analyzed the approximation and generalization capabilities of diffusion models [20, 22, 24], and have also proven that their score functions can approximate the tangent space of the data manifold [32]. In particular, for a compact embedded sub-manifold \\(\\mathcal{M} \\subset \\mathbb{R}^n\\), it has been shown that for a sample \\(z_t \\in \\mathbb{R}^n\\) sufficiently close to the target data, the score \\(\\nabla_{z_t} \\log p_t(z_t) (\\approx s_\\theta(z_t))\\) and orthogonal projection \\(\\pi(z_t)\\) onto data manifold \\(\\mathcal{M}_0\\) satisfy a key relationship. For the projection \\(\\mathbf{N}_p\\) onto the normal space and \\(\\mathbf{T}_p\\) onto the tangent space of \\(\\mathcal{M}_0\\), the ratio of their magnitudes goes to zero as \\(t\\) approaches 0 (i.e., gets closer to the target data). In other words, for samples \\(z_t\\) close to the target data, the following equation holds:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.359, + 0.483, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\frac {\\left\\| \\mathbf {T} _ {p} \\nabla_ {\\mathbf {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\right\\|}{\\left\\| \\mathbf {N} _ {p} \\nabla_ {\\mathbf {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\right\\|} \\rightarrow 0, \\quad \\text {a s} t \\rightarrow 0, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.465 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_p\\) and \\(\\mathbf{N}_p\\) are the projection operators onto the tangent space \\(\\mathcal{T}_{\\pi (z_t)}\\mathcal{M}_0\\) and the normal space \\(\\mathcal{N}_{\\pi (z_t)}\\mathcal{M}_0\\), respectively (for a detailed proof, see Theorem 4.1, Corollary 4.2 and Appendix D in [32])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.484, + 0.535 + ], + "angle": 0, + "content": "This implies that, for samples sufficiently close to the target manifold \\(\\mathcal{M}_0\\), the cosine similarity between the score function and the normal vector \\(\\mathbf{n} = \\frac{\\pi(z_t) - z_t}{\\|\\pi(z_t) - z_t\\|}\\) converges to 1 (i.e., \\(S_{\\mathrm{cos}}(\\mathbf{n},\\nabla_{z_t}\\log p_t(z_t))\\xrightarrow{t\\to 0}1\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.722 + ], + "angle": 0, + "content": "This suggests that for a sample \\( z_{t} \\) very close to the target, the score function \\( \\nabla_{\\mathbf{z}}\\log p_t(z_t)\\approx s_\\theta (z_t) \\) becomes an element of the normal space of the target manifold (that is, \\( \\nabla_{\\mathbf{z}}\\log p_t(z_t)\\in \\mathcal{N}_{\\pi (z_t)}\\mathcal{M}_0\\approx \\mathcal{N}_{\\pi (z_0)}\\mathcal{M}_0 \\) for sufficiently small \\( t \\)). Leveraging this property, the estimated diffusion score can approximate the intrinsic dimension of the target data by utilizing the huge gap in the singular values of the sampling scores \\( S = \\left[s_{\\theta}\\left(\\boldsymbol {z}_t^{(1)},t\\right),\\dots ,s_{\\theta}\\left(\\boldsymbol {z}_t^{(4n)},t\\right)\\right] \\) where the singular vectors corresponding to the higher singular values represent the normal components of \\( \\mathcal{M}_0 \\), while those corresponding to the lower singular values represent the tangential components.[32]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.737, + 0.192, + 0.751 + ], + "angle": 0, + "content": "3. Intuition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.484, + 0.854 + ], + "angle": 0, + "content": "In this section, we assume the mathematical concept behind our method and supporting experiments. Our approach refines CFG at each step by dropping the tangential component of the unconditional score, enhancing the quality of conditional generation. This adjustment allows the conditional score to guide the generated sample more directly" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.239, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Figure 2. Singular values of the score function across all timesteps. We computed the singular values for all timesteps using a total of 17,000 samples from Stable Diffusion v1.5. For both the unconditional and the conditional scores, a significant drop in singular values was observed at indices close to 0 across all timesteps. This suggests the existence of an intermediate manifold." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.905, + 0.38 + ], + "angle": 0, + "content": "toward the manifold specified by the condition, improving alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.906, + 0.443 + ], + "angle": 0, + "content": "To support this, we provide empirical evidence suggesting that not only does the target data manifold \\(\\mathcal{M}_0\\) exist but there is also a manifold \\(\\mathcal{M}_{t - \\epsilon}\\) at each time step \\(t\\in (0,1)\\) where \\(\\nabla_{\\mathbf{z}}\\log p_t(\\pmb {z}_t)\\in \\mathcal{N}_{\\pi_{t - \\epsilon}(\\pmb {z}_t)}\\mathcal{M}_{t - \\epsilon}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.906, + 0.535 + ], + "angle": 0, + "content": "There exists an intermediate manifold \\(\\mathcal{M}_t\\) We hypothesize the existence of a manifold \\(\\mathcal{M}_{(t - \\epsilon)}\\) that contains \\(\\nabla_{\\pmb {z}_t}\\log p_t(\\pmb {z}_t)\\) as elements of its normal space, not only for samples close to the target data but also for \\(t\\in (0,1)\\). Specifically, we assume the following:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.906, + 0.62 + ], + "angle": 0, + "content": "Assumption 1. Suppose that the support of the data distribution \\(P_0\\) is contained in a compact embedded submanifold \\(\\mathcal{M}_0\\subset \\mathbb{R}^d\\), and let \\(P_{t}\\) be the distribution of latents at time \\(t\\) diffused from \\(P_0\\). Then, under mild assumptions², \\(\\forall t\\in (0,1)\\), \\(\\exists t'\\in (t - \\epsilon ,t + \\epsilon)\\) such that:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.631, + 0.815, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\in \\mathcal {N} _ {\\pi_ {t ^ {\\prime}} (\\boldsymbol {z} _ {t})} \\mathcal {M} _ {t ^ {\\prime}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.75 + ], + "angle": 0, + "content": "for sufficiently small \\(\\epsilon\\) and orthogonal projection \\(\\pi_{t'}(\\mathbf{z}_t)\\) onto manifold \\(\\mathcal{M}_{t'}\\). This hypothesis is indirectly supported by the clear gap in singular values arranged in descending order for a sufficient number of samples. This phenomenon occurs not only when \\(t\\) goes to 0 (near the image manifold) but also consistently for all time step \\(t \\in (0,1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.855 + ], + "angle": 0, + "content": "To observe the gap, we compute 17,000 score samples across all timesteps on Stable Diffusion v1.5. Let \\([\\sigma_1,\\dots ,\\sigma_D]\\) represent the singular values from the SVD applied to \\(s_{\\theta}(z_{t})\\) and \\(s_{\\theta}(z_{t},y)\\), with 17,000 samples collected per timestep, arranged in descending order. The corresponding singular vectors are denoted as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Every compact embedded submanifold of \\( \\mathbb{R}^d \\) has a tubular neighborhood, and for a given manifold \\( \\mathcal{M} \\), each point \\( \\pmb{z} \\in \\mathbb{R}^n \\) within this tubular neighborhood has a unique orthogonal projection \\( \\pi \\) onto \\( \\mathcal{M} \\) [16]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.906, + 0.9 + ], + "angle": 0, + "content": "2) The distribution \\( P_0 \\) has a smooth density \\( p_0 \\) w.r.t the volume measure on the manifold. 2) The density \\( p_0 \\) is bounded away from zero on the manifold." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.09, + 0.445, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.201, + 0.483, + 0.286 + ], + "angle": 0, + "content": "Figure 3. Cosine similarity between singular vectors of unconditional and conditional scores. We computed the singular vectors \\( V \\) at each timestep using a total of 17,000 samples from Stable Diffusion v1.5. We observe the similarity of significant singular vectors (i.e., those with indices close to 0) between unconditional and conditional scores are mostly high across all timesteps \\( T \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.484, + 0.33 + ], + "angle": 0, + "content": "\\([\\pmb{v}_1,\\dots ,\\pmb{v}_D]^T\\) for \\(\\pmb{s}_{\\theta}(\\pmb {z}_t)\\) and \\([\\hat{\\pmb{v}}_1,\\dots ,\\hat{\\pmb{v}}_D]^T\\) for \\(s_\\theta (z_t,y)\\) respectively \\((D\\approx^{3}17,000)\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.496 + ], + "angle": 0, + "content": "As shown in Fig. 2, both \\( s_{\\theta}(z_t) \\) and \\( s_{\\theta}(z_t,y) \\) have gaps between the highest singular values and the rest for all \\( t \\in [0,1] \\), not just for \\( 0 + \\epsilon \\). Interpreting from the perspective that the score function \\( s_{\\theta}(z_t) \\) becomes an element of the data manifold's normal space as \\( t \\) approaches 0 [32]. Assuming the existence of an intermediate manifold \\( \\mathcal{M}_t \\) for all \\( t \\in (0,1) \\), this suggests that the singular vectors associated with the largest singular values contain dominant components of \\( \\mathcal{NM}_t \\), while vectors associated with smaller singular values correspond to component of \\( \\mathcal{T}\\mathcal{M}_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Tangential misalignment between unconditional and conditional score We empirically justify the principle of modifying the unconditional score by dropping the components with low singular values and retaining only the components with high singular values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.65 + ], + "angle": 0, + "content": "Fig. 3 shows that conditional and unconditional singular vectors \\([\\pmb{v}_1, \\dots, \\pmb{v}_D]^T\\) and \\([\\hat{\\pmb{v}}_1, \\dots, \\hat{\\pmb{v}}_D]^T\\) at corresponding indices are more similar when their singular values are high than the rest." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.484, + 0.711 + ], + "angle": 0, + "content": "More specifically, the cosine similarity of the singular vectors \\(\\pmb{v}_1\\) and \\(\\hat{\\pmb{v}}_1\\) associated with the highest singular value \\(\\sigma_1\\) from \\(s_\\theta(z_t)\\) and \\(s_\\theta(z_t, y)\\), respectively, is higher than the others." + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.722, + 0.482, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left[ S _ {\\cos} \\left(\\boldsymbol {v} _ {1}, \\hat {\\boldsymbol {v}} _ {1}\\right) > S _ {\\cos} \\left(\\boldsymbol {v} _ {j}, \\hat {\\boldsymbol {v}} _ {j}\\right) \\right] \\\\ \\approx \\left[ S _ {\\cos} \\left(\\mathbf {N} _ {p} \\nabla_ {z _ {t}} \\log p _ {t} \\left(z _ {t}, y\\right), \\mathbf {N} _ {p} \\nabla_ {z _ {t}} \\log p _ {t} \\left(z _ {t}\\right)\\right) \\right. \\tag {2} \\\\ \\left. > S _ {\\cos} \\left(\\mathbf {T} _ {p} \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} \\left(\\boldsymbol {z} _ {t}, y\\right), \\mathbf {T} _ {p} \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} \\left(\\boldsymbol {z} _ {t}\\right)\\right) \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.483, + 0.823 + ], + "angle": 0, + "content": "for \\(1 < j \\leq D\\). The cosine similarity \\(S_{cos}\\) between two vectors \\(\\mathbf{v}_i\\) and \\(\\mathbf{v}_j\\) is defined as \\(S_{\\mathrm{cos}}(\\mathbf{v}_i, \\mathbf{v}_j) = \\frac{\\mathbf{v}_i \\cdot \\mathbf{v}_j}{\\|\\mathbf{v}_i\\| \\|\\mathbf{v}_j\\|}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "This indicates that the intermediate manifolds associated with \\(\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{z}_t)\\) and \\(\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{z}_t,y)\\) share similar normal components, while their tangent components are relatively less aligned." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.906, + 0.243 + ], + "angle": 0, + "content": "These less-aligned components interfere with the generative process, making it harder to align with the target manifold. We modify the unconditional score \\( s_{\\theta}(z_t) \\) at each timestep by removing its tangential components that are less aligned with the conditional score \\( s_{\\theta}(z_t,y) \\). We provide detailed methods in the following section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.256, + 0.613, + 0.272 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.906, + 0.418 + ], + "angle": 0, + "content": "Our main method proceeds as follows. At each step, we take the predicted unconditional score \\( s_{\\theta}(z_t) \\) and the conditional score \\( s_{\\theta}(z_t,y) \\) and concatenate them into a score matrix \\( A = [s_{\\theta}(z_t),s_{\\theta}(z_t,y)] \\). Next, we perform SVD on \\( A \\), obtaining singular values and corresponding singular vectors that consider both components \\( s(z_{t}) \\) and \\( s(z_{t},y) \\). This results in singular vectors \\( [\\pmb {v}_1,\\pmb {v}_2,\\dots ,\\pmb {v}_D]^T \\) where \\( \\pmb{v}_{1} \\) is the normal component of both \\( s(z_{t}) \\) and \\( s(z_{t},y) \\). We project the unconditional score onto \\( \\pmb{v}_{1} \\) and drop the rest." + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.427, + 0.905, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {s}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) = \\boldsymbol {s} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) \\cdot \\boldsymbol {V} ^ {T} \\cdot \\left[ \\boldsymbol {v} _ {1}, \\boldsymbol {0} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.906, + 0.561 + ], + "angle": 0, + "content": "Consequently, the singular vectors associated with high singular values in the score matrix \\( A \\) retain the well-aligned, normal components of \\( s_{\\theta}(z_t) \\) and \\( s_{\\theta}(z_t,y) \\), while those with lower singular values represent misaligned tangential components, which we set to zero in Eq. (3) to drop these components from the unconditional score. Next, we update the score \\( \\hat{s}_{\\theta}(z_t,y) \\) with classifier-free guidance (CFG):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.572, + 0.905, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {z} _ {t}} \\log \\hat {p} _ {t} \\left(\\boldsymbol {z} _ {t} | y\\right) = \\hat {\\boldsymbol {s}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) + w \\left(\\boldsymbol {s} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, y\\right) - \\hat {\\boldsymbol {s}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.613, + 0.905, + 0.643 + ], + "angle": 0, + "content": "Algorithm 1 Tangential damping classifier-free guidance (TCFG)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Inputs: \\(\\mathbf{s}_{\\theta}(\\mathbf{z}_t)\\) and \\(\\mathbf{s}_{\\theta}(\\mathbf{z}_t|\\mathbf{y})\\): predicted unconditional and conditional scores, \\(t \\in (0,1)\\): time step, \\(y\\): condition, \\(w\\): CFG scale." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.691, + 0.597, + 0.706 + ], + "angle": 0, + "content": "Output: \\( z_0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.708, + 0.656, + 0.723 + ], + "angle": 0, + "content": "1: for \\( t \\in (0, 1) \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.724, + 0.668, + 0.738 + ], + "angle": 0, + "content": "2: Get \\( s_{\\theta} \\) from \\( z_{t} \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.739, + 0.851, + 0.754 + ], + "angle": 0, + "content": "3: Make score matrix \\( \\mathbf{A} = \\left[s_{\\theta}\\left(z_{t}\\right), s_{\\theta}\\left(z_{t}, y\\right)\\right] \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.754, + 0.821, + 0.769 + ], + "angle": 0, + "content": "4: \\((\\sigma_{i})_{i = 1}^{d},(\\pmb{w}_{i})_{i = 1}^{d},(\\pmb{v}_{i})_{i = 1}^{d}\\gets \\mathrm{SVD}(\\pmb {A})\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.77, + 0.769, + 0.784 + ], + "angle": 0, + "content": "5: \\(\\hat{s}_{\\theta}(\\pmb{z}_t) = s_{\\theta}(\\pmb{z}_t)\\cdot \\pmb{V}^T\\cdot [\\pmb{v}_1,\\pmb {0}]\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.785, + 0.754, + 0.799 + ], + "angle": 0, + "content": "6: (Dropping \\(\\mathbf{T}\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{x}_t))\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.799, + 0.862, + 0.814 + ], + "angle": 0, + "content": "7: \\(\\hat{s}_{\\theta}(\\boldsymbol{z}_t, y) = \\hat{s}_{\\theta}(\\boldsymbol{z}_t) + w(\\boldsymbol{s}_{\\theta}(\\boldsymbol{z}_t, y) - \\hat{s}_{\\theta}(\\boldsymbol{z}_t))\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.815, + 0.636, + 0.829 + ], + "angle": 0, + "content": "8: Update \\(z_{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.83, + 0.596, + 0.842 + ], + "angle": 0, + "content": "9: end for" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.845, + 0.617, + 0.859 + ], + "angle": 0, + "content": "10: Output \\(z_0\\)" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.708, + 0.862, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.862, + 0.905, + 0.893 + ], + "angle": 0, + "content": "\\((\\sigma_{i})_{i = 1}^{d},(\\mathbf{w}_{i})_{i = 1}^{d},(\\mathbf{v}_{i})_{i = 1}^{d}\\) denote singular values, left and right singular vectors respectively." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.839, + 0.484, + 0.901 + ], + "angle": 0, + "content": "3Approximately \\(4\\times n\\) samples are sufficient to accurately estimate the intrinsic dimension of the target manifold \\(\\mathcal{M}_0\\) [32]. However, our goal is to verify the existence of a manifold where \\(\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol {z}_t)\\) is an element of the normal space. Therefore, it suffices to observe the presence of a large gap in the singular value spectrum, thus \\(N < D\\) is enough." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2623" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.094, + 0.289, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.189, + 0.278, + 0.201 + ], + "angle": 0, + "content": "(a) Conditional scores only" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.094, + 0.429, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.19, + 0.396, + 0.2 + ], + "angle": 0, + "content": "(b) Classifier-free" + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.203, + 0.289, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.297, + 0.288, + 0.308 + ], + "angle": 0, + "content": "(c) Calculate SVD on all samples" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.203, + 0.429, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.297, + 0.376, + 0.307 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.404 + ], + "angle": 0, + "content": "Figure 4. Sampling results on different methods with diffusion model trained on two moons dataset. Our proposed methods (c, d) demonstrate a closer match to the target distribution compared to using conditional scores only or CFG. In (c), SVD is computed across all samples, while in (d), SVD is calculated separately for each pair of conditional and unconditional scores." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.42, + 0.411, + 0.434 + ], + "angle": 0, + "content": "We provide a detailed algorithm in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.555 + ], + "angle": 0, + "content": "Unlike traditional CFG update methods, \\(\\nabla_{\\boldsymbol{z}_t}\\log \\hat{p_t} (\\boldsymbol {z}_t|\\boldsymbol {y})\\) drops tangential component from the unconditional score at each step. It prevents accumulating misaligned components from the unconditional score \\(s_\\theta (z_t)\\) using the direction of the manifold defined by the given condition \\(y\\) over time evolution. This concept is further illustrated with a simple distribution in Sec. 5, where the toy example clarifies the benefits of our methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.571, + 0.221, + 0.589 + ], + "angle": 0, + "content": "5. Toy example" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.687 + ], + "angle": 0, + "content": "We empirically verify our method on a toy problem, generating the two moons dataset. Experiments consist of the generated samples with different guidances including the original classifier-free guidance (CFG) and ours, and the sampling trajectories following their respective score functions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.779 + ], + "angle": 0, + "content": "The target data distribution \\( p(X_0) \\) consists of samples distributed along two distinct curves (moons). We trained a conditional diffusion model using a small neural network that receives a binary label \\( y \\in \\{0,1\\} \\) for the two moons or \\( y = \\varnothing \\) denoting the null condition. For detailed settings, please refer to the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Fig. 4 shows the generated samples using four different guiding strategies. (a) uses only the conditional score \\( s_{\\theta}(z_t, y) \\). (b) uses the CFG score. (c) and (d) employ our guidance score at each step with multiple samples and one sample, respectively, to compute singular value decomposition (SVD) of the unconditional score \\( s_{\\theta}^{i}(z_t) \\) and conditional score \\( s_{\\theta}^{i}(z_t, y) \\)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "According to the result, generated samples using our" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.687, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.218, + 0.653, + 0.228 + ], + "angle": 0, + "content": "Trajectory of sample A" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.091, + 0.903, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.218, + 0.852, + 0.228 + ], + "angle": 0, + "content": "Trajectory of sample B" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.242, + 0.905, + 0.339 + ], + "angle": 0, + "content": "Figure 5. Visualization of the sampling trajectory. In CFG (orange path), the unconditional scores (red arrows) include components that point towards directions other than the target distribution, making the final destination deviate from the target distribution. Whereas, our method (green path) removes the inconsistent tangent components in unconditional scores and eventually reaches the target distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.905, + 0.445 + ], + "angle": 0, + "content": "strategies lie closer to the target compared to those generated using only the conditional score or CFG. CFG, while potentially bringing samples closer to the target than merely using conditional scores, may face challenges due to the misalignment of tangent components between unconditional score and conditional scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.553 + ], + "angle": 0, + "content": "In contrast, our guidance score can reduce the tangent component of the unconditional score at each step. This helps samples converge more effectively towards the target data, which suggests that the tangent components of the unconditional score might hinder alignment with the target data manifold under the given condition, and our method helps in mitigating this misalignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.906, + 0.78 + ], + "angle": 0, + "content": "We further validate this hypothesis by examining the trajectories of generated samples. Fig. 5 visualizes the trajectories induced by our score \\(\\nabla_{\\boldsymbol{z}_t}\\log \\hat{p}_t(\\boldsymbol{z}_t|\\boldsymbol{y})\\) compared to the original CFG score \\(\\nabla_{\\boldsymbol{z}_t}\\log \\tilde{p}_{\\theta}(\\boldsymbol{z}_t|\\boldsymbol{y})\\). As shown, in the orange CFG trajectory, the direction of unconditional scores changes frequently. This results in difficulties for the blue conditional score to maintain an orthogonal direction relative to the target manifold near the target distribution. In contrast, our method consistently adjusts the score to predict in a direction closer to orthogonal with respect to the target manifold, particularly as the samples converge toward the target data. Our method removes the tangential component of the unconditional score with respect to the manifold of the conditional score. This results in a direction that leans either to the right or to the left." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.905, + 0.827 + ], + "angle": 0, + "content": "Additionally, the similar results between (c) and (d) in Fig. 4 suggest that computing SVD for only a single sample is sufficient to yield nearly the same result." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.645, + 0.862 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section, we demonstrate that our method is applicable to high-dimensional diffusion models. We employ rep" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2624" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.17, + 0.089, + 0.406, + 0.171 + ], + "angle": 0, + "content": "
FID ↓CLIPScore ↑
SD v1.5original13.260.31
+ ours13.120.31
SDXLoriginal13.360.32
+ ours12.650.32
SD v3original16.660.32
+ ours13.740.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.239 + ], + "angle": 0, + "content": "Table 1. Zero-shot FID and CLIPScore measured on MSCOCO 30k. Our method consistently improves FID across all models—Stable Diffusion v1.5, SDXL, and SD v3—while maintaining a nearly identical CLIPScore." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.243, + 0.482, + 0.288 + ], + "angle": 0, + "content": "
FID ↓sFID ↓Precision ↑Recall ↑IS ↑
DiT32.6717.920.900.13271.1
DiT+ours29.513.270.900.19270.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.298, + 0.482, + 0.34 + ], + "angle": 0, + "content": "Table 2. Evaluation metrics measured on ImageNet 50k using DiT. Our method achieves better performance in FID, sFID, Precision, and Recall while showing a slight decrease in Inception Score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.483, + 0.421 + ], + "angle": 0, + "content": "representative diffusion models such as Stable Diffusion v1.5 [27] and SDXL [23], and showed that it functions identically on SD v3 [7], which is based on Rectified Flow. Additionally, we conducted experiments on DiT [21], which is trained on ImageNet [5]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.449, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Experimental details For the text-to-image models, we used zero-shot FID [10] and CLIPScore [25] on the MSCOCO 2014 validation set [17] consisting of 30,000 images under the commonly used text-to-image evaluation protocols. [7, 23, 27] For DiT, we evaluated using 50,000 images under the same settings as ADM [6]. All models used the official pretrained weights, and sampling was performed using the same latent codes. We used the best CFG scales as the default value of each repository. Our method does not increase the inference time of all baselines." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Quantitative evaluation Tab. 1 presents the FID and CLIP Scores for SD1.5, SDXL, and SD3. Our method achieved better FID scores while maintaining the same CLIP Scores across all three models. Notably, the decrease in FID is larger for SDXL compared to SD1.5, and even larger for SD3 compared to SDXL. We speculate that this is because SD3, known as a better model publicly, has a relatively clearer manifold. Furthermore, the results on SD3 demonstrate that our method is applicable not only to diffusion models but also to all CFG-based score functions, including those based on Rectified Flow. Fig. 6 also shows FID-CLIP curves on SDXL, demonstrating that FID improves even as the CFG scale changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Tab. 2 shows the results on the DiT model. Except for a slight decrease in Inception Score, our method exhibits relatively superior performance in FID, sFID, and Recall. This indicates that our method can be equally applied to both text-to-image generation and class-conditioned generation." + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.09, + 0.845, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.19, + 0.892, + 0.204 + ], + "angle": 0, + "content": "Figure 6. FID-CLIP curves on SDXL with 50 sampling steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.208, + 0.905, + 0.344 + ], + "angle": 0, + "content": "Qualitative evaluation Our method drops the tangential component from the unconditional score while retaining the normal component. This reduces misalignment with the conditional score, thereby improving image quality as shown in Fig. 7. Specifically, the changes introduced by our approach transform \"strange\" objects or scenes into more \"plausible\" images. This indirectly demonstrates that the misalignment of the unconditional score in the conventional CFG was causing the \"strange\" aspects in the final outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.905, + 0.404 + ], + "angle": 0, + "content": "For example, our method converts physically impossible or unusual combinations of objects (SD3), uncommon appearances or characteristics (SDXL), and ambiguous shapes or forms (SD1.5) into \"normal\" results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.905, + 0.512 + ], + "angle": 0, + "content": "Fig. 8 presents the results obtained from DiT. We observed that our method causes relatively more changes in the images generated by DiT. We speculate that this is because DiT is trained on ImageNet dataset with class labels. The results qualitatively show that when using our method, DiT generates images that are more detailed, have better structure, and appear more natural." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.905, + 0.587 + ], + "angle": 0, + "content": "Notably, in both text-to-image and class-conditioned image generation, we observed a reduction in the overexposure bias problem. We attribute this improvement to the mitigation of misalignment between the unconditional score and the conditional score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.905, + 0.809 + ], + "angle": 0, + "content": "What happened to the unconditional score? In this paragraph, we qualitatively demonstrate that the misalignment with the conditional score is reduced when we drop the tangential component from the unconditional score and retain the normal component. We compared the results sampled using CFG with those sampled using the null condition (i.e., unconditional) when generating images from the same random noise (i.e., latent variables) in SDXL. In our method, we used the text condition to compute \\(\\hat{s}\\) but used only \\(\\hat{s}\\) for denoising; that is, we set \\(\\omega = 0\\). Although this approach does not perfectly explain our method, we can indirectly infer its role by observing how the modified null condition changes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Fig. 9 shows that images sampled using the original null condition generate different objects such as trees, snowy mountain landscapes, and women. In contrast, images generated using our modified null condition \\(\\hat{s}\\) show that the tree part takes the form of a feather, the snowy mountain landscape changes into a woman, and the woman transforms" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2625" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.094, + 0.181, + 0.106 + ], + "angle": 0, + "content": "SD3" + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.094, + 0.338, + 0.106 + ], + "angle": 0, + "content": "SD3 + ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.094, + 0.456, + 0.106 + ], + "angle": 0, + "content": "SDXL" + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.094, + 0.611, + 0.106 + ], + "angle": 0, + "content": "SDXL + ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.094, + 0.719, + 0.106 + ], + "angle": 0, + "content": "SD1.5" + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.094, + 0.873, + 0.106 + ], + "angle": 0, + "content": "SD1.5 + ours" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.108, + 0.364, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.108, + 0.632, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.108, + 0.901, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.431, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Figure 7. Qualitative evaluation of text-to-image models. Our method prevents overexposure, enhancing the shapes and details of objects." + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.464, + 0.203, + 0.473 + ], + "angle": 0, + "content": "DiT" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.476, + 0.287, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.464, + 0.416, + 0.474 + ], + "angle": 0, + "content": "DiT \\(^+\\) ours" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.476, + 0.48, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.754 + ], + "angle": 0, + "content": "Figure 8. Qualitative evaluation of DiT Our method mitigates overexposure and enhances object shapes and details in DiT models trained on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.9 + ], + "angle": 0, + "content": "into a shape resembling a glove. These changes align with the objects we aim to generate: a feather, a woman, and a baseball glove. We observe that these changes due to the null condition help eliminate unwanted structures or artifacts in the generated images. In other words, we demonstrate that the misalignment of the null condition is reduced, and we claim that this improvement aids in image generation." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.459, + 0.903, + 0.489 + ], + "angle": 0, + "content": "
SAGSAG+TCFGPAGPAG+TCFGCFGPPCFGPP+TCFG
FID13.5311.4814.4511.8713.9713.44
CLIP Score0.310.300.310.310.320.32
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.491, + 0.906, + 0.545 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison with existing baselines. The evaluation was conducted on \\( {30}\\mathrm{k} \\) images from the MS-COCO dataset using the official code; SD v1.4 for SAG,SD v1.5 for PAG and SDXL for CFG++." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.549, + 0.652, + 0.563 + ], + "angle": 0, + "content": "7. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Calssifier-free guidance Experimental methods to enhance the performance of Classifier-Free Guidance (CFG) have been studied. SAG [13] proposed a method to improve CFG by using intermediate self-attention maps. PAG [2] suggested computing CFG by transforming self-attention maps into identity matrices. ICG [29] enhanced CFG by utilizing random text embeddings. Recently, CFG++ [4] demonstrated better performance by modifying the CFG computation method. Our proposed approach modifies the unconditional score based on the conditional score and can be used alongside these existing works; please refer to Tab. 3 and the appendix for more results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Manifold hypothesis and diffusion There are also several studies that have utilized the manifold hypothesis properties of the score function estimated by diffusion models to address various inherent challenges associated with diffusion processes. One approach introduces the manifold memorization hypothesis to understand model memorization through the relationship between data and model manifold dimensionalities [28]. Another extends memorization" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2626" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.907, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.907, + 0.414 + ], + "angle": 0, + "content": "Figure 9. TCFG reduces misalignments between unconditional and conditional generation. Starting from the same random noise \\( z_{1} \\), when SDXL samples images with only the unconditional score, it produces random images such as trees, snowy mountain landscapes, and women. In contrast, our modified unconditional score, projected on dominant (conditional), generates images that somewhat match the desired text prompts. This is because our method reduces misalignment with the conditional score by dropping the tangential components of the unconditional score. Once the misalignment decreases, the quality of the final images (unconditional + conditional score) improves: The base of the feather has a more natural structure, the human arm appears more natural, and the extra string on the left side of the baseball glove is removed." + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.418, + 0.419, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Figure 10. Limitations Our method occasionally struggles to fix severely wrong regions in the baseline samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.751 + ], + "angle": 0, + "content": "theory to diffusion models [1], showing that high-variance subspaces are selectively lost due to memorization effects. Separately, different researchers proposed an approach for detecting synthetic images generated by diffusion models, achieving high accuracy across diverse datasets [18]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.334, + 0.784 + ], + "angle": 0, + "content": "8. Discussion and conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our work experimentally analyzes the issues arising in the standard CFG method, where the tangential component of the unconditional score does not align well with that of the conditional score. By using SVD to drop the tangential component in the unconditional score, we effectively improve text-to-image generation quality. Additionally, our CFG method is easily applicable, has low computational" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.507 + ], + "angle": 0, + "content": "cost, and enhances image quality. We leverage the ability of the diffusion model's score function to encode the intrinsic dimension of the target data, demonstrating the misalignment between the conditional and unconditional scores to improve sampling quality. This is the first attempt to utilize this misalignment to enhance sampling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Despite these advantages, several unresolved issues remain. First, it is uncertain whether the misalignment of tangential component and the alignment of normal component between the predicted unconditional score \\( s_{\\theta}(z_t) \\) and the conditional score \\( s_{\\theta}(z_t,y) \\) in the CFG setting, at a given timestep \\( t \\), would similarly apply to the features derived from a separately trained classifier and the null condition score in the classifier guidance setting. Second, while our task leverages the capability of diffusion models to estimate intrinsic dimensions for enhancing conditional sampling methods, we present only experimental observations regarding the existence of an intermediate manifold for \\( t \\in [0,1] \\), without theoretical proof. Further exploration and rigorous analysis of these aspects are left as future work. Third, additional investigation is needed to adapt our approach effectively in the context of diffusion distillation using CFG scale as an input [14, 15], which we also identify as a promising direction for future research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Finally, although our method successfully demonstrated on-manifold image generation, we observed that when the original image exhibits significant abnormalities, substantial changes may occasionally cause the structure to break down. Fig. 10 illustrates such examples. Nevertheless, it is evident that our method transforms \"strange\" images into more \"normal\" ones." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2627" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Beatrice Achilli, Enrico Ventura, Gianluigi Silvestri, Bao Pham, Gabriel Raya, Dmitry Krotov, Carlo Lucibello, and Luca Ambrogioni. Losing dimensions: Geometric memorization in generative diffusion. arXiv preprint arXiv:2410.08727, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[4] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.483, + 0.429 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[9] Charles Fefferman, Sanjoy Mitter, and Hariharan Narayanan. Testing the manifold hypothesis. Journal of the American Mathematical Society, 29(4):983-1049, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.755 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 7, 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[14] Yi-Ting Hsiao, Siavash Khodadadeh, Kevin Duarte, Wei-An Lin, Hui Qu, Mingi Kwon, and Ratheesh Kalarot. Plug-and-play diffusion distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13743-13752, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Black Forest Labs. FLUX. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-15. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[16] John M Lee. Introduction to Riemannian manifolds. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.237, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.322, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[18] Peter Lorenz, Ricard L Durall, and Janis Keuper. Detecting images generated by deep diffusion models using their local intrinsic dimensionality. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 448-459, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[19] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[20] Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion models are minimax optimal distribution estimators. In International Conference on Machine Learning, pages 26517-26582. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[22] Jakiw Pidstrigach. Score-based generative models detect manifolds. Advances in Neural Information Processing Systems, 35:35852-35865, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.906, + 0.688 + ], + "angle": 0, + "content": "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.691, + 0.906, + 0.745 + ], + "angle": 0, + "content": "[24] Phillip Pope, Chen Zhu, Ahmed Abdelkader, Micah Goldblum, and Tom Goldstein. The intrinsic dimension of images and its impact on learning. arXiv preprint arXiv:2104.08894, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2628" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[28] Brendan Leigh Ross, Hamidreza Kamkari, Tongzi Wu, Rasa Hosseinzadeh, Zhaoyan Liu, George Stein, Jesse C Cresswell, and Gabriel Loaiza-Ganem. A geometric framework for understanding memorization in generative models. arXiv preprint arXiv:2411.00113, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[29] Seyedmorteza Sadat, Manuel Kansy, Otmar Hilliges, and Romann M Weber. No training, no problem: Rethinking classifier-free guidance for diffusion models. arXiv preprint arXiv:2407.02687, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.483, + 0.373 + ], + "angle": 0, + "content": "[30] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[32] Jan Pawel Stanczuk, Georgios Batzolis, Teo Deveney, and Carola-Bibiane Schonlieb. Diffusion models encode the intrinsic dimension of data manifolds. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[33] Zhenyue Zhang and Hongyuan Zha. Principal manifolds and nonlinear dimensionality reduction via tangent space alignment. SIAM journal on scientific computing, 26(1):313-338, 2004. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2629" + } + ] +] \ No newline at end of file diff --git a/2025/TCFG_ Tangential Damping Classifier-free Guidance/8113438a-9689-405a-b6bc-61aa320e1894_origin.pdf b/2025/TCFG_ Tangential Damping Classifier-free Guidance/8113438a-9689-405a-b6bc-61aa320e1894_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..02aead455e77fc06f824897b913925278bf5c736 --- /dev/null +++ b/2025/TCFG_ Tangential Damping Classifier-free Guidance/8113438a-9689-405a-b6bc-61aa320e1894_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2fdd630b3fcd61e1dda203f3a367ab020657542c640da66c9ae605c7f405233 +size 8986196 diff --git a/2025/TCFG_ Tangential Damping Classifier-free Guidance/full.md b/2025/TCFG_ Tangential Damping Classifier-free Guidance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c1e21191fdab4e2b23c86b9bb31f88843e992e53 --- /dev/null +++ b/2025/TCFG_ Tangential Damping Classifier-free Guidance/full.md @@ -0,0 +1,370 @@ +# TCFG: Tangential Damping Classifier-free Guidance + +Mingi Kwon* + +Yonsei University + +kwonmingi@yonsei.ac.kr + +Shin seong Kim* + +Yonsei University + +tltydl2@yonsei.ac.kr + +Jaeseok Jeong + +Yonsei University + +jete jeong@yonsei.ac.kr + +Yi Ting Hsiao + +University of Michigan + +hsiaoyt@umich.edu + +Youngjung Uh + +Yonsei University + +yj.uh@yonsei.ac.kr + +# Abstract + +Diffusion models have achieved remarkable success in text-to-image synthesis, largely attributed to the use of classifier-free guidance (CFG), which enables high-quality, condition-aligned image generation. CFG combines the conditional score (e.g., text-conditioned) with the unconditional score to control the output. However, the unconditional score is in charge of estimating the transition between manifolds of adjacent timesteps from $x_{t}$ to $x_{t-1}$ , which may inadvertently interfere with the trajectory toward the specific condition. In this work, we introduce a novel approach that leverages a geometric perspective on the unconditional score to enhance CFG performance when conditional scores are available. Specifically, we propose a method that filters the singular vectors of both conditional and unconditional scores using singular value decomposition. This filtering process aligns the unconditional score with the conditional score, thereby refining the sampling trajectory to stay closer to the manifold. Our approach improves image quality with negligible additional computation. We provide deeper insights into the score function behavior in diffusion models and present a practical technique for achieving more accurate and contextually coherent image synthesis. project page + +# 1. Introduction + +Diffusion models [12, 31] have shown remarkable progress in image generation [19, 27, 30]. In particular, the emergence of classifier-free guidance [6, 11] (CFG) has attracted significant attention because it allows us to provide desired guidance by leveraging the conditional estimated score directly within the diffusion model. + +The classifier-free guidance fundamentally computes the + +![](images/706fdab5475f7a29b260c1b5cf8387caad4554e208c01e6cc676adeecc163e42.jpg) +(a) Classifier-free guidance +Figure 1. (a) Classifier-free guidance. When the unconditional score $s_{\theta}(\boldsymbol{z}_t)$ and the conditional score $s_{\theta}(\boldsymbol{z}_t, y)$ are misaligned, the result of CFG tends to fall off the manifold. (b) Our proposed method reduces the misalignment between the unconditional score $s_{\theta}(\boldsymbol{z}_t)$ and the conditional score $s_{\theta}(\boldsymbol{z}_t, y)$ , ensuring sampling aligns with the target manifold. + +![](images/482be4850a43f1cf1bc6a4faede5132a6dabcb2bcda05e78d8b2a17eabdffb60.jpg) +(b) Ours + +final score by combining the unconditional and conditional estimated scores. This approach ensures a generation that aligns well with the given condition. Additionally, using an appropriate guidance scale has been shown to enhance image quality across various tasks, further driving improvements in applications like text-to-image generation. + +Let us say the guided score $\tilde{s}_{\theta}$ as $\tilde{s}_{\theta} = s_{\theta}^{\mathrm{uncond}} + \omega_{\mathrm{scale}}(s_{\theta}^{\mathrm{cond}} - s_{\theta}^{\mathrm{uncond}})$ . In text-to-image models, the text condition $(s_{\theta}^{\mathrm{cond}})$ is randomly replaced with a null condition $(s_{\theta}^{\mathrm{uncond}})$ during training (e.g., with a probability $p = 0.1$ ), enabling the null condition to act as a general estimator for any sample. It means that $s_{\theta}^{\mathrm{uncond}}$ is the score estimated from $z_{t}$ to $z_{t-1}$ for all samples in the sampling trajectory of diffusion models. + +The unconditional score for any sample has certainly enabled the successful use of classifier-free guidance. However, we argue that there can be a misalignment between the unconditional and conditional estimated scores (See Eq. (2) in Sec. 2), which hinders the approximation toward the + +manifold by the given condition. Fig. 1 (a) conceptually illustrates the potential issue that arises when the manifold of the unconditional score differs from that of the conditional score. In this paper, we show that this misalignment can be resolved with a simple algorithm, which significantly reduces the tendency of CFG to generate off-manifold samples, as illustrated in Fig. 1 (b). + +Our approach is based on the following insights. First, the score predicted by the diffusion model estimates the intrinsic dimension of the data manifold [32]. Additionally, this intrinsic dimension can be captured by the tangent space of the target manifold [3, 9]. Instead of directly estimating the intrinsic dimension, we focus on utilizing the tangential component inherent in the unconditional score during classifier-free guidance. By reducing its misalignment with the conditional score, we enhance the alignment and ultimately improve the quality of the generated outputs. + +Specifically, we push the score $\tilde{s}_{\theta}$ toward the normal direction of the conditional manifold by eliminating the values of column vectors with small singular values using the orthogonal matrix $V$ obtained through the singular value decomposition of the conditional and unconditional scores. + +In this paper, we propose a novel sampling method that leverages the unconditional score within CFG. To support our approach, we first lay out the theoretical foundation in section Sec. 2 and Sec. 3, discussing the manifold hypothesis and its connection to diffusion models. In Sec. 4, we provide a comprehensive explanation of our proposed method. This is followed by a detailed analysis using a toy example in Sec. 5, and we demonstrate the practical applicability of our method on real-world text-to-image models in Sec. 6. + +Our experiments show a significant improvement in the MS-COCO Fréchet Inception Distance (FID) across various models that utilize classifier-free guidance, e.g., diffusion models (Stable Diffusion v1.5 [26] and SDXL [23]) and rectified flow (Stable Diffusion 3 [8]). Additionally, our method improves DiT [21] FID on ImageNet. Notably, our method helps mitigate the overexposure bias problem, leading to resulting images that better align with the underlying data distribution, as supported by improved quantitative metrics. + +# 2. Background + +Diffusion models Diffusion models learn the score that reverses the forward noising process. This forward process from the real data distribution $p(\pmb{x}_0)$ to a latent distribution $p(z_1) \sim N(0, \sigma_{\max}^2 I)$ along timesteps $t \in [0, 1]$ is defined by a Gaussian kernel: $z_t = x_0 + \sigma(t)\epsilon$ . The function $\sigma(t)$ is a noise schedule where $\sigma(0) = 0$ and $\sigma(1) = \sigma_{\max}$ , determining the amount of noise to be added at each timestep $t$ to erase information from $x$ . + +A generative process is represented as its reverse with a + +stochastic differential equation (SDE): + +$$ +\begin{array}{l} \mathrm {d} \boldsymbol {z} = - \dot {\sigma} (t) \sigma (t) \nabla_ {\boldsymbol {z} _ {t}} \log p _ {t} (\boldsymbol {z} _ {t}) \mathrm {d} t \\ - \beta (t) \sigma (t) ^ {2} \nabla_ {\boldsymbol {z} _ {t}} \log p _ {t} (\boldsymbol {z} _ {t}) d t + \sqrt {2 \beta (t)} \sigma (t) d \omega_ {t}, \\ \end{array} +$$ + +where $\mathrm{d}\omega_{t}$ is a standard Wiener process. Alternatively, it can be expressed as an ordinary differential equation: + +$$ +\mathrm {d} \boldsymbol {z} = - \dot {\sigma} (t) \sigma (t) \nabla_ {\boldsymbol {z} _ {t}} \log p _ {t} (\boldsymbol {z} _ {t}) \mathrm {d} t. +$$ + +Diffusion models approximate the score function $\nabla_{\boldsymbol{z}_t}\log p_t(\boldsymbol{z}_t)$ with a neural network $s_\theta (\boldsymbol {z}_t,t)$ . They are trained to predict the clean data from the noisy $\boldsymbol{z}_t$ . The trained model performs the reverse process using: + +$$ +\nabla_ {\boldsymbol {z} _ {t}} \log p _ {t} (\boldsymbol {z} _ {t}) \approx \frac {\boldsymbol {s} _ {\theta} (\boldsymbol {z} _ {t} , t) - \boldsymbol {z} _ {t}}{\sigma (t) ^ {2}}. +$$ + +Classifier guidance (CG) and classifier-free guidance (CFG) For an arbitrary class label $y$ , CG defines the class-conditional sampling distribution $\tilde{p}_{\theta}(\boldsymbol{z}_t \mid y)$ as: + +$$ +\tilde {p} _ {\theta} \left(\boldsymbol {z} _ {t} \mid y\right) \propto p _ {\theta} \left(\boldsymbol {z} _ {t} \mid y\right) p _ {\theta} (y \mid \boldsymbol {z} _ {t}) ^ {\gamma}, +$$ + +where $p_{\theta}(y \mid z_t)$ is the classifier distribution and $\gamma$ is a scaling parameter. [6] When $\gamma > 0$ , it is known to reduce sample diversity but enhance quality. However, CG requires a classifier that can predict label $y$ from the noisy $z_t$ . CFG proposes a method to sample from the conditional distribution by expressing the classifier distribution $p_{\theta}(y \mid z_t)$ in terms of the conditional distribution $p_{\theta}(z_t \mid y)$ and the unconditional distribution $p_{\theta}(z_t)$ : + +$$ +\tilde {p} _ {\theta} \left(\boldsymbol {z} _ {t} \mid y\right) \propto p _ {\theta} \left(\boldsymbol {z} _ {t} \mid y\right) ^ {1 + \gamma} p _ {\theta} \left(\boldsymbol {z} _ {t}\right) ^ {- \gamma}. +$$ + +As a result, the final score $\nabla_{z_t}\log \tilde{p}_\theta (\boldsymbol {z}_t\mid y)$ is approximated by: + +$$ +\begin{array}{l} \nabla_ {\boldsymbol {z} _ {t}} \log \tilde {p} _ {\theta} (\boldsymbol {z} _ {t} \mid y) = (1 + \gamma) \boldsymbol {s} _ {\theta} (\boldsymbol {z} _ {t}, y) - \gamma \boldsymbol {s} _ {\theta} (\boldsymbol {z} _ {t}) \\ = \boldsymbol {s} _ {\theta} (\boldsymbol {z} _ {t}) + \omega \left(\boldsymbol {s} _ {\theta} (\boldsymbol {z} _ {t}, y) - \boldsymbol {s} _ {\theta} (\boldsymbol {z} _ {t})\right), \\ \end{array} +$$ + +where $\omega = 1 + \gamma$ . [11] + +In practice, both $s_{\theta}(x_t,y)$ and $s_{\theta}(z_t)$ are approximated by a single neural network that is jointly trained to estimate both the conditional and unconditional scores. Text-to-image models use the null condition $y_{\mathrm{null}} = \varnothing$ as a class label to train $s_{\theta}(z_t)\approx s_{\theta}(z_t,y_{\mathrm{null}})$ . This approach allows $s_{\theta}(z_t,y) - s_{\theta}(z_t)$ to provide guidance similar to the gradient of an implicit classifier. Hereafter, we will simply denote $s_{\theta}(z_t,y_{\mathrm{null}})$ as $s_{\theta}(z_t)$ . + +Diffusion models and data manifold encoding The manifold hypothesis suggests that high-dimensional data lies on or near a lower-dimensional manifold, making intrinsic dimension estimation essential for data representation [9]. This intrinsic dimension is often encoded in the + +manifold's tangent spaces, which capture underlying degrees of freedom and align local structures to reveal the global geometry [3, 33]. + +Building on these ideas, further studies have analyzed the approximation and generalization capabilities of diffusion models [20, 22, 24], and have also proven that their score functions can approximate the tangent space of the data manifold [32]. In particular, for a compact embedded sub-manifold $\mathcal{M} \subset \mathbb{R}^n$ , it has been shown that for a sample $z_t \in \mathbb{R}^n$ sufficiently close to the target data, the score $\nabla_{z_t} \log p_t(z_t) (\approx s_\theta(z_t))$ and orthogonal projection $\pi(z_t)$ onto data manifold $\mathcal{M}_0$ satisfy a key relationship. For the projection $\mathbf{N}_p$ onto the normal space and $\mathbf{T}_p$ onto the tangent space of $\mathcal{M}_0$ , the ratio of their magnitudes goes to zero as $t$ approaches 0 (i.e., gets closer to the target data). In other words, for samples $z_t$ close to the target data, the following equation holds: + +$$ +\frac {\left\| \mathbf {T} _ {p} \nabla_ {\mathbf {z} _ {t}} \log p _ {t} (\boldsymbol {z} _ {t}) \right\|}{\left\| \mathbf {N} _ {p} \nabla_ {\mathbf {z} _ {t}} \log p _ {t} (\boldsymbol {z} _ {t}) \right\|} \rightarrow 0, \quad \text {a s} t \rightarrow 0, \tag {1} +$$ + +where $\mathbf{T}_p$ and $\mathbf{N}_p$ are the projection operators onto the tangent space $\mathcal{T}_{\pi (z_t)}\mathcal{M}_0$ and the normal space $\mathcal{N}_{\pi (z_t)}\mathcal{M}_0$ , respectively (for a detailed proof, see Theorem 4.1, Corollary 4.2 and Appendix D in [32]). + +This implies that, for samples sufficiently close to the target manifold $\mathcal{M}_0$ , the cosine similarity between the score function and the normal vector $\mathbf{n} = \frac{\pi(z_t) - z_t}{\|\pi(z_t) - z_t\|}$ converges to 1 (i.e., $S_{\mathrm{cos}}(\mathbf{n},\nabla_{z_t}\log p_t(z_t))\xrightarrow{t\to 0}1$ ). + +This suggests that for a sample $z_{t}$ very close to the target, the score function $\nabla_{\mathbf{z}}\log p_t(z_t)\approx s_\theta (z_t)$ becomes an element of the normal space of the target manifold (that is, $\nabla_{\mathbf{z}}\log p_t(z_t)\in \mathcal{N}_{\pi (z_t)}\mathcal{M}_0\approx \mathcal{N}_{\pi (z_0)}\mathcal{M}_0$ for sufficiently small $t$ ). Leveraging this property, the estimated diffusion score can approximate the intrinsic dimension of the target data by utilizing the huge gap in the singular values of the sampling scores $S = \left[s_{\theta}\left(\boldsymbol {z}_t^{(1)},t\right),\dots ,s_{\theta}\left(\boldsymbol {z}_t^{(4n)},t\right)\right]$ where the singular vectors corresponding to the higher singular values represent the normal components of $\mathcal{M}_0$ , while those corresponding to the lower singular values represent the tangential components.[32] + +# 3. Intuition + +In this section, we assume the mathematical concept behind our method and supporting experiments. Our approach refines CFG at each step by dropping the tangential component of the unconditional score, enhancing the quality of conditional generation. This adjustment allows the conditional score to guide the generated sample more directly + +![](images/2dc8ff2f56920b200ea0696cc91ea6a29dd819c8d4c64eb7681009255a20c7e6.jpg) +Figure 2. Singular values of the score function across all timesteps. We computed the singular values for all timesteps using a total of 17,000 samples from Stable Diffusion v1.5. For both the unconditional and the conditional scores, a significant drop in singular values was observed at indices close to 0 across all timesteps. This suggests the existence of an intermediate manifold. + +toward the manifold specified by the condition, improving alignment. + +To support this, we provide empirical evidence suggesting that not only does the target data manifold $\mathcal{M}_0$ exist but there is also a manifold $\mathcal{M}_{t - \epsilon}$ at each time step $t\in (0,1)$ where $\nabla_{\mathbf{z}}\log p_t(\pmb {z}_t)\in \mathcal{N}_{\pi_{t - \epsilon}(\pmb {z}_t)}\mathcal{M}_{t - \epsilon}$ . + +There exists an intermediate manifold $\mathcal{M}_t$ We hypothesize the existence of a manifold $\mathcal{M}_{(t - \epsilon)}$ that contains $\nabla_{\pmb {z}_t}\log p_t(\pmb {z}_t)$ as elements of its normal space, not only for samples close to the target data but also for $t\in (0,1)$ . Specifically, we assume the following: + +Assumption 1. Suppose that the support of the data distribution $P_0$ is contained in a compact embedded submanifold $\mathcal{M}_0\subset \mathbb{R}^d$ , and let $P_{t}$ be the distribution of latents at time $t$ diffused from $P_0$ . Then, under mild assumptions², $\forall t\in (0,1)$ , $\exists t'\in (t - \epsilon ,t + \epsilon)$ such that: + +$$ +\nabla_ {\boldsymbol {z} _ {t}} \log p _ {t} (\boldsymbol {z} _ {t}) \in \mathcal {N} _ {\pi_ {t ^ {\prime}} (\boldsymbol {z} _ {t})} \mathcal {M} _ {t ^ {\prime}}, +$$ + +for sufficiently small $\epsilon$ and orthogonal projection $\pi_{t'}(\mathbf{z}_t)$ onto manifold $\mathcal{M}_{t'}$ . This hypothesis is indirectly supported by the clear gap in singular values arranged in descending order for a sufficient number of samples. This phenomenon occurs not only when $t$ goes to 0 (near the image manifold) but also consistently for all time step $t \in (0,1)$ . + +To observe the gap, we compute 17,000 score samples across all timesteps on Stable Diffusion v1.5. Let $[\sigma_1,\dots ,\sigma_D]$ represent the singular values from the SVD applied to $s_{\theta}(z_{t})$ and $s_{\theta}(z_{t},y)$ , with 17,000 samples collected per timestep, arranged in descending order. The corresponding singular vectors are denoted as + +![](images/2cb5021bbd32ce4dcf5b043fddcc1d4dbfe00ea27913e7103fae31a5322fbd8c.jpg) +Figure 3. Cosine similarity between singular vectors of unconditional and conditional scores. We computed the singular vectors $V$ at each timestep using a total of 17,000 samples from Stable Diffusion v1.5. We observe the similarity of significant singular vectors (i.e., those with indices close to 0) between unconditional and conditional scores are mostly high across all timesteps $T$ . + +$[\pmb{v}_1,\dots ,\pmb{v}_D]^T$ for $\pmb{s}_{\theta}(\pmb {z}_t)$ and $[\hat{\pmb{v}}_1,\dots ,\hat{\pmb{v}}_D]^T$ for $s_\theta (z_t,y)$ respectively $(D\approx^{3}17,000)$ + +As shown in Fig. 2, both $s_{\theta}(z_t)$ and $s_{\theta}(z_t,y)$ have gaps between the highest singular values and the rest for all $t \in [0,1]$ , not just for $0 + \epsilon$ . Interpreting from the perspective that the score function $s_{\theta}(z_t)$ becomes an element of the data manifold's normal space as $t$ approaches 0 [32]. Assuming the existence of an intermediate manifold $\mathcal{M}_t$ for all $t \in (0,1)$ , this suggests that the singular vectors associated with the largest singular values contain dominant components of $\mathcal{NM}_t$ , while vectors associated with smaller singular values correspond to component of $\mathcal{T}\mathcal{M}_t$ . + +Tangential misalignment between unconditional and conditional score We empirically justify the principle of modifying the unconditional score by dropping the components with low singular values and retaining only the components with high singular values. + +Fig. 3 shows that conditional and unconditional singular vectors $[\pmb{v}_1, \dots, \pmb{v}_D]^T$ and $[\hat{\pmb{v}}_1, \dots, \hat{\pmb{v}}_D]^T$ at corresponding indices are more similar when their singular values are high than the rest. + +More specifically, the cosine similarity of the singular vectors $\pmb{v}_1$ and $\hat{\pmb{v}}_1$ associated with the highest singular value $\sigma_1$ from $s_\theta(z_t)$ and $s_\theta(z_t, y)$ , respectively, is higher than the others. + +$$ +\begin{array}{l} \left[ S _ {\cos} \left(\boldsymbol {v} _ {1}, \hat {\boldsymbol {v}} _ {1}\right) > S _ {\cos} \left(\boldsymbol {v} _ {j}, \hat {\boldsymbol {v}} _ {j}\right) \right] \\ \approx \left[ S _ {\cos} \left(\mathbf {N} _ {p} \nabla_ {z _ {t}} \log p _ {t} \left(z _ {t}, y\right), \mathbf {N} _ {p} \nabla_ {z _ {t}} \log p _ {t} \left(z _ {t}\right)\right) \right. \tag {2} \\ \left. > S _ {\cos} \left(\mathbf {T} _ {p} \nabla_ {\boldsymbol {z} _ {t}} \log p _ {t} \left(\boldsymbol {z} _ {t}, y\right), \mathbf {T} _ {p} \nabla_ {\boldsymbol {z} _ {t}} \log p _ {t} \left(\boldsymbol {z} _ {t}\right)\right) \right] \\ \end{array} +$$ + +for $1 < j \leq D$ . The cosine similarity $S_{cos}$ between two vectors $\mathbf{v}_i$ and $\mathbf{v}_j$ is defined as $S_{\mathrm{cos}}(\mathbf{v}_i, \mathbf{v}_j) = \frac{\mathbf{v}_i \cdot \mathbf{v}_j}{\|\mathbf{v}_i\| \|\mathbf{v}_j\|}$ . + +This indicates that the intermediate manifolds associated with $\nabla_{\boldsymbol{z}_t}\log p_t(\boldsymbol{z}_t)$ and $\nabla_{\boldsymbol{z}_t}\log p_t(\boldsymbol{z}_t,y)$ share similar normal components, while their tangent components are relatively less aligned. + +These less-aligned components interfere with the generative process, making it harder to align with the target manifold. We modify the unconditional score $s_{\theta}(z_t)$ at each timestep by removing its tangential components that are less aligned with the conditional score $s_{\theta}(z_t,y)$ . We provide detailed methods in the following section. + +# 4. Methods + +Our main method proceeds as follows. At each step, we take the predicted unconditional score $s_{\theta}(z_t)$ and the conditional score $s_{\theta}(z_t,y)$ and concatenate them into a score matrix $A = [s_{\theta}(z_t),s_{\theta}(z_t,y)]$ . Next, we perform SVD on $A$ , obtaining singular values and corresponding singular vectors that consider both components $s(z_{t})$ and $s(z_{t},y)$ . This results in singular vectors $[\pmb {v}_1,\pmb {v}_2,\dots ,\pmb {v}_D]^T$ where $\pmb{v}_{1}$ is the normal component of both $s(z_{t})$ and $s(z_{t},y)$ . We project the unconditional score onto $\pmb{v}_{1}$ and drop the rest. + +$$ +\hat {\boldsymbol {s}} _ {\theta} \left(\boldsymbol {z} _ {t}\right) = \boldsymbol {s} _ {\theta} \left(\boldsymbol {z} _ {t}\right) \cdot \boldsymbol {V} ^ {T} \cdot \left[ \boldsymbol {v} _ {1}, \boldsymbol {0} \right]. \tag {3} +$$ + +Consequently, the singular vectors associated with high singular values in the score matrix $A$ retain the well-aligned, normal components of $s_{\theta}(z_t)$ and $s_{\theta}(z_t,y)$ , while those with lower singular values represent misaligned tangential components, which we set to zero in Eq. (3) to drop these components from the unconditional score. Next, we update the score $\hat{s}_{\theta}(z_t,y)$ with classifier-free guidance (CFG): + +$$ +\nabla_ {\boldsymbol {z} _ {t}} \log \hat {p} _ {t} \left(\boldsymbol {z} _ {t} | y\right) = \hat {\boldsymbol {s}} _ {\theta} \left(\boldsymbol {z} _ {t}\right) + w \left(\boldsymbol {s} _ {\theta} \left(\boldsymbol {z} _ {t}, y\right) - \hat {\boldsymbol {s}} _ {\theta} \left(\boldsymbol {z} _ {t}\right)\right). \tag {4} +$$ + +Algorithm 1 Tangential damping classifier-free guidance (TCFG) + +Inputs: $\mathbf{s}_{\theta}(\mathbf{z}_t)$ and $\mathbf{s}_{\theta}(\mathbf{z}_t|\mathbf{y})$ : predicted unconditional and conditional scores, $t \in (0,1)$ : time step, $y$ : condition, $w$ : CFG scale. + +Output: $z_0$ + +1: for $t \in (0, 1)$ do +2: Get $s_{\theta}$ from $z_{t}$ +3: Make score matrix $\mathbf{A} = \left[s_{\theta}\left(z_{t}\right), s_{\theta}\left(z_{t}, y\right)\right]$ +4: $(\sigma_{i})_{i = 1}^{d},(\pmb{w}_{i})_{i = 1}^{d},(\pmb{v}_{i})_{i = 1}^{d}\gets \mathrm{SVD}(\pmb {A})$ +5: $\hat{s}_{\theta}(\pmb{z}_t) = s_{\theta}(\pmb{z}_t)\cdot \pmb{V}^T\cdot [\pmb{v}_1,\pmb {0}]$ +6: (Dropping $\mathbf{T}\nabla_{\boldsymbol{z}_t}\log p_t(\boldsymbol{x}_t))$ +7: $\hat{s}_{\theta}(\boldsymbol{z}_t, y) = \hat{s}_{\theta}(\boldsymbol{z}_t) + w(\boldsymbol{s}_{\theta}(\boldsymbol{z}_t, y) - \hat{s}_{\theta}(\boldsymbol{z}_t))$ +8: Update $z_{t}$ +9: end for +10: Output $z_0$ + +$(\sigma_{i})_{i = 1}^{d},(\mathbf{w}_{i})_{i = 1}^{d},(\mathbf{v}_{i})_{i = 1}^{d}$ denote singular values, left and right singular vectors respectively. + +![](images/0080c29da4b011faefc3d7b8d9ae5d398c2b87e041f9109d754939c67993f165.jpg) + +![](images/adb4f0bcc1d9f2ef2b23e95c1a0c96e5997f0fec96dbeb52c59bc11932d8932f.jpg) +(b) Classifier-free + +![](images/b0c1ba6a9bc0e9751c7e789c548f4184c8b7f6c4d6f6bcd950d9cfb30a497644.jpg) +(a) Conditional scores only +(c) Calculate SVD on all samples +Figure 4. Sampling results on different methods with diffusion model trained on two moons dataset. Our proposed methods (c, d) demonstrate a closer match to the target distribution compared to using conditional scores only or CFG. In (c), SVD is computed across all samples, while in (d), SVD is calculated separately for each pair of conditional and unconditional scores. + +![](images/c805e93368939886f4821b153ed857109d6328029ad5c26cd336f4b5990b73d4.jpg) +(d) Ours + +We provide a detailed algorithm in Algorithm 1. + +Unlike traditional CFG update methods, $\nabla_{\boldsymbol{z}_t}\log \hat{p_t} (\boldsymbol {z}_t|\boldsymbol {y})$ drops tangential component from the unconditional score at each step. It prevents accumulating misaligned components from the unconditional score $s_\theta (z_t)$ using the direction of the manifold defined by the given condition $y$ over time evolution. This concept is further illustrated with a simple distribution in Sec. 5, where the toy example clarifies the benefits of our methods. + +# 5. Toy example + +We empirically verify our method on a toy problem, generating the two moons dataset. Experiments consist of the generated samples with different guidances including the original classifier-free guidance (CFG) and ours, and the sampling trajectories following their respective score functions. + +The target data distribution $p(X_0)$ consists of samples distributed along two distinct curves (moons). We trained a conditional diffusion model using a small neural network that receives a binary label $y \in \{0,1\}$ for the two moons or $y = \varnothing$ denoting the null condition. For detailed settings, please refer to the Appendix. + +Fig. 4 shows the generated samples using four different guiding strategies. (a) uses only the conditional score $s_{\theta}(z_t, y)$ . (b) uses the CFG score. (c) and (d) employ our guidance score at each step with multiple samples and one sample, respectively, to compute singular value decomposition (SVD) of the unconditional score $s_{\theta}^{i}(z_t)$ and conditional score $s_{\theta}^{i}(z_t, y)$ . + +According to the result, generated samples using our + +![](images/29c3ce1ec7ae83c8421d95b085b9588d7688c0e8c2061cc3f7fc5791669d81c4.jpg) +Trajectory of sample A +Figure 5. Visualization of the sampling trajectory. In CFG (orange path), the unconditional scores (red arrows) include components that point towards directions other than the target distribution, making the final destination deviate from the target distribution. Whereas, our method (green path) removes the inconsistent tangent components in unconditional scores and eventually reaches the target distribution. + +![](images/7721ebb674467b5145386389dceb2a9fdae583bbf25c194c6a0157d1a31c6571.jpg) +Trajectory of sample B + +strategies lie closer to the target compared to those generated using only the conditional score or CFG. CFG, while potentially bringing samples closer to the target than merely using conditional scores, may face challenges due to the misalignment of tangent components between unconditional score and conditional scores. + +In contrast, our guidance score can reduce the tangent component of the unconditional score at each step. This helps samples converge more effectively towards the target data, which suggests that the tangent components of the unconditional score might hinder alignment with the target data manifold under the given condition, and our method helps in mitigating this misalignment. + +We further validate this hypothesis by examining the trajectories of generated samples. Fig. 5 visualizes the trajectories induced by our score $\nabla_{\boldsymbol{z}_t}\log \hat{p}_t(\boldsymbol{z}_t|\boldsymbol{y})$ compared to the original CFG score $\nabla_{\boldsymbol{z}_t}\log \tilde{p}_{\theta}(\boldsymbol{z}_t|\boldsymbol{y})$ . As shown, in the orange CFG trajectory, the direction of unconditional scores changes frequently. This results in difficulties for the blue conditional score to maintain an orthogonal direction relative to the target manifold near the target distribution. In contrast, our method consistently adjusts the score to predict in a direction closer to orthogonal with respect to the target manifold, particularly as the samples converge toward the target data. Our method removes the tangential component of the unconditional score with respect to the manifold of the conditional score. This results in a direction that leans either to the right or to the left. + +Additionally, the similar results between (c) and (d) in Fig. 4 suggest that computing SVD for only a single sample is sufficient to yield nearly the same result. + +# 6. Experiments + +In this section, we demonstrate that our method is applicable to high-dimensional diffusion models. We employ rep + +
FID ↓CLIPScore ↑
SD v1.5original13.260.31
+ ours13.120.31
SDXLoriginal13.360.32
+ ours12.650.32
SD v3original16.660.32
+ ours13.740.32
+ +Table 1. Zero-shot FID and CLIPScore measured on MSCOCO 30k. Our method consistently improves FID across all models—Stable Diffusion v1.5, SDXL, and SD v3—while maintaining a nearly identical CLIPScore. + +
FID ↓sFID ↓Precision ↑Recall ↑IS ↑
DiT32.6717.920.900.13271.1
DiT+ours29.513.270.900.19270.0
+ +Table 2. Evaluation metrics measured on ImageNet 50k using DiT. Our method achieves better performance in FID, sFID, Precision, and Recall while showing a slight decrease in Inception Score. + +representative diffusion models such as Stable Diffusion v1.5 [27] and SDXL [23], and showed that it functions identically on SD v3 [7], which is based on Rectified Flow. Additionally, we conducted experiments on DiT [21], which is trained on ImageNet [5]. + +Experimental details For the text-to-image models, we used zero-shot FID [10] and CLIPScore [25] on the MSCOCO 2014 validation set [17] consisting of 30,000 images under the commonly used text-to-image evaluation protocols. [7, 23, 27] For DiT, we evaluated using 50,000 images under the same settings as ADM [6]. All models used the official pretrained weights, and sampling was performed using the same latent codes. We used the best CFG scales as the default value of each repository. Our method does not increase the inference time of all baselines. + +Quantitative evaluation Tab. 1 presents the FID and CLIP Scores for SD1.5, SDXL, and SD3. Our method achieved better FID scores while maintaining the same CLIP Scores across all three models. Notably, the decrease in FID is larger for SDXL compared to SD1.5, and even larger for SD3 compared to SDXL. We speculate that this is because SD3, known as a better model publicly, has a relatively clearer manifold. Furthermore, the results on SD3 demonstrate that our method is applicable not only to diffusion models but also to all CFG-based score functions, including those based on Rectified Flow. Fig. 6 also shows FID-CLIP curves on SDXL, demonstrating that FID improves even as the CFG scale changes. + +Tab. 2 shows the results on the DiT model. Except for a slight decrease in Inception Score, our method exhibits relatively superior performance in FID, sFID, and Recall. This indicates that our method can be equally applied to both text-to-image generation and class-conditioned generation. + +![](images/952ac4373792e71419fd0a50bfed71ff1329eff7e94bc493d97eca39aad03288.jpg) +Figure 6. FID-CLIP curves on SDXL with 50 sampling steps. + +Qualitative evaluation Our method drops the tangential component from the unconditional score while retaining the normal component. This reduces misalignment with the conditional score, thereby improving image quality as shown in Fig. 7. Specifically, the changes introduced by our approach transform "strange" objects or scenes into more "plausible" images. This indirectly demonstrates that the misalignment of the unconditional score in the conventional CFG was causing the "strange" aspects in the final outputs. + +For example, our method converts physically impossible or unusual combinations of objects (SD3), uncommon appearances or characteristics (SDXL), and ambiguous shapes or forms (SD1.5) into "normal" results. + +Fig. 8 presents the results obtained from DiT. We observed that our method causes relatively more changes in the images generated by DiT. We speculate that this is because DiT is trained on ImageNet dataset with class labels. The results qualitatively show that when using our method, DiT generates images that are more detailed, have better structure, and appear more natural. + +Notably, in both text-to-image and class-conditioned image generation, we observed a reduction in the overexposure bias problem. We attribute this improvement to the mitigation of misalignment between the unconditional score and the conditional score. + +What happened to the unconditional score? In this paragraph, we qualitatively demonstrate that the misalignment with the conditional score is reduced when we drop the tangential component from the unconditional score and retain the normal component. We compared the results sampled using CFG with those sampled using the null condition (i.e., unconditional) when generating images from the same random noise (i.e., latent variables) in SDXL. In our method, we used the text condition to compute $\hat{s}$ but used only $\hat{s}$ for denoising; that is, we set $\omega = 0$ . Although this approach does not perfectly explain our method, we can indirectly infer its role by observing how the modified null condition changes. + +Fig. 9 shows that images sampled using the original null condition generate different objects such as trees, snowy mountain landscapes, and women. In contrast, images generated using our modified null condition $\hat{s}$ show that the tree part takes the form of a feather, the snowy mountain landscape changes into a woman, and the woman transforms + +![](images/e5b21f654b98856ed13f4c4741391164a1bb465d129ab605085b553a340d29ed.jpg) +SD3 +SD3 + ours +Figure 7. Qualitative evaluation of text-to-image models. Our method prevents overexposure, enhancing the shapes and details of objects. + +![](images/273e6e2ef619d99e4436f3ec5c960d1c30da51ece63a5cee1b960d4cb90916bc.jpg) +SDXL +SDXL + ours + +![](images/f4f3e9c0787e8993a4d886f81698aecd5a281a054ae02c87683cee05ec36ba7d.jpg) +SD1.5 +SD1.5 + ours + +![](images/90389e3e52e478f073e557f056826834da79adebf095074e7a2a441af34ae4fd.jpg) +DiT +Figure 8. Qualitative evaluation of DiT Our method mitigates overexposure and enhances object shapes and details in DiT models trained on ImageNet. + +![](images/b54d34de12e13b69d130a5ca60a9efcf1c33cef56c68e8ae8359e10a430f68c1.jpg) +DiT $^+$ ours + +into a shape resembling a glove. These changes align with the objects we aim to generate: a feather, a woman, and a baseball glove. We observe that these changes due to the null condition help eliminate unwanted structures or artifacts in the generated images. In other words, we demonstrate that the misalignment of the null condition is reduced, and we claim that this improvement aids in image generation. + +
SAGSAG+TCFGPAGPAG+TCFGCFGPPCFGPP+TCFG
FID13.5311.4814.4511.8713.9713.44
CLIP Score0.310.300.310.310.320.32
+ +Table 3. Quantitative comparison with existing baselines. The evaluation was conducted on ${30}\mathrm{k}$ images from the MS-COCO dataset using the official code; SD v1.4 for SAG,SD v1.5 for PAG and SDXL for CFG++. + +# 7. Related work + +Calssifier-free guidance Experimental methods to enhance the performance of Classifier-Free Guidance (CFG) have been studied. SAG [13] proposed a method to improve CFG by using intermediate self-attention maps. PAG [2] suggested computing CFG by transforming self-attention maps into identity matrices. ICG [29] enhanced CFG by utilizing random text embeddings. Recently, CFG++ [4] demonstrated better performance by modifying the CFG computation method. Our proposed approach modifies the unconditional score based on the conditional score and can be used alongside these existing works; please refer to Tab. 3 and the appendix for more results. + +Manifold hypothesis and diffusion There are also several studies that have utilized the manifold hypothesis properties of the score function estimated by diffusion models to address various inherent challenges associated with diffusion processes. One approach introduces the manifold memorization hypothesis to understand model memorization through the relationship between data and model manifold dimensionalities [28]. Another extends memorization + +![](images/256d319284fa09a96fd3418b3acd5ca0fa7702c1be6860fc8d62d9a903fad4d3.jpg) +Figure 9. TCFG reduces misalignments between unconditional and conditional generation. Starting from the same random noise $z_{1}$ , when SDXL samples images with only the unconditional score, it produces random images such as trees, snowy mountain landscapes, and women. In contrast, our modified unconditional score, projected on dominant (conditional), generates images that somewhat match the desired text prompts. This is because our method reduces misalignment with the conditional score by dropping the tangential components of the unconditional score. Once the misalignment decreases, the quality of the final images (unconditional + conditional score) improves: The base of the feather has a more natural structure, the human arm appears more natural, and the extra string on the left side of the baseball glove is removed. + +![](images/82d53a3b354e00e8625ecb192fcca134ccaf1375e60e7c705329562695dbb71f.jpg) +Figure 10. Limitations Our method occasionally struggles to fix severely wrong regions in the baseline samples. + +theory to diffusion models [1], showing that high-variance subspaces are selectively lost due to memorization effects. Separately, different researchers proposed an approach for detecting synthetic images generated by diffusion models, achieving high accuracy across diverse datasets [18]. + +# 8. Discussion and conclusion + +Our work experimentally analyzes the issues arising in the standard CFG method, where the tangential component of the unconditional score does not align well with that of the conditional score. By using SVD to drop the tangential component in the unconditional score, we effectively improve text-to-image generation quality. Additionally, our CFG method is easily applicable, has low computational + +cost, and enhances image quality. We leverage the ability of the diffusion model's score function to encode the intrinsic dimension of the target data, demonstrating the misalignment between the conditional and unconditional scores to improve sampling quality. This is the first attempt to utilize this misalignment to enhance sampling. + +Despite these advantages, several unresolved issues remain. First, it is uncertain whether the misalignment of tangential component and the alignment of normal component between the predicted unconditional score $s_{\theta}(z_t)$ and the conditional score $s_{\theta}(z_t,y)$ in the CFG setting, at a given timestep $t$ , would similarly apply to the features derived from a separately trained classifier and the null condition score in the classifier guidance setting. Second, while our task leverages the capability of diffusion models to estimate intrinsic dimensions for enhancing conditional sampling methods, we present only experimental observations regarding the existence of an intermediate manifold for $t \in [0,1]$ , without theoretical proof. Further exploration and rigorous analysis of these aspects are left as future work. Third, additional investigation is needed to adapt our approach effectively in the context of diffusion distillation using CFG scale as an input [14, 15], which we also identify as a promising direction for future research. + +Finally, although our method successfully demonstrated on-manifold image generation, we observed that when the original image exhibits significant abnormalities, substantial changes may occasionally cause the structure to break down. Fig. 10 illustrates such examples. Nevertheless, it is evident that our method transforms "strange" images into more "normal" ones. + +# References + +[1] Beatrice Achilli, Enrico Ventura, Gianluigi Silvestri, Bao Pham, Gabriel Raya, Dmitry Krotov, Carlo Lucibello, and Luca Ambrogioni. Losing dimensions: Geometric memorization in generative diffusion. arXiv preprint arXiv:2410.08727, 2024. 8 +[2] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 7, 2 +[3] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 2, 3 +[4] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 7, 2 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2, 6 +[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 6 +[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2 +[9] Charles Fefferman, Sanjoy Mitter, and Hariharan Narayanan. Testing the manifold hypothesis. Journal of the American Mathematical Society, 29(4):983-1049, 2016. 2 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1 +[13] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 7, 1 + +[14] Yi-Ting Hsiao, Siavash Khodadadeh, Kevin Duarte, Wei-An Lin, Hui Qu, Mingi Kwon, and Ratheesh Kalarot. Plug-and-play diffusion distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13743-13752, 2024. 8 +[15] Black Forest Labs. FLUX. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-15. 8 +[16] John M Lee. Introduction to Riemannian manifolds. Springer, 2018. 3 +[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[18] Peter Lorenz, Ricard L Durall, and Janis Keuper. Detecting images generated by deep diffusion models using their local intrinsic dimensionality. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 448-459, 2023. 8 +[19] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[20] Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion models are minimax optimal distribution estimators. In International Conference on Machine Learning, pages 26517-26582. PMLR, 2023. 3 +[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 6 +[22] Jakiw Pidstrigach. Score-based generative models detect manifolds. Advances in Neural Information Processing Systems, 35:35852-35865, 2022. 3 +[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6 +[24] Phillip Pope, Chen Zhu, Ahmed Abdelkader, Micah Goldblum, and Tom Goldstein. The intrinsic dimension of images and its impact on learning. arXiv preprint arXiv:2104.08894, 2021. 3 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2 + +[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 6 +[28] Brendan Leigh Ross, Hamidreza Kamkari, Tongzi Wu, Rasa Hosseinzadeh, Zhaoyan Liu, George Stein, Jesse C Cresswell, and Gabriel Loaiza-Ganem. A geometric framework for understanding memorization in generative models. arXiv preprint arXiv:2411.00113, 2024. 7 +[29] Seyedmorteza Sadat, Manuel Kansy, Otmar Hilliges, and Romann M Weber. No training, no problem: Rethinking classifier-free guidance for diffusion models. arXiv preprint arXiv:2407.02687, 2024. 7 +[30] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1 +[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1 +[32] Jan Pawel Stanczuk, Georgios Batzolis, Teo Deveney, and Carola-Bibiane Schonlieb. Diffusion models encode the intrinsic dimension of data manifolds. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4 +[33] Zhenyue Zhang and Hongyuan Zha. Principal manifolds and nonlinear dimensionality reduction via tangent space alignment. SIAM journal on scientific computing, 26(1):313-338, 2004. 3 \ No newline at end of file diff --git a/2025/TCFG_ Tangential Damping Classifier-free Guidance/images.zip b/2025/TCFG_ Tangential Damping Classifier-free Guidance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..060ce5f744504a0a74c39713e3f640ad7f644cc3 --- /dev/null +++ b/2025/TCFG_ Tangential Damping Classifier-free Guidance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ab24eee843ecac3c6ec724bf70602c6d04b8437b97bcc42aad0f21e120e1046 +size 628652 diff --git a/2025/TCFG_ Tangential Damping Classifier-free Guidance/layout.json b/2025/TCFG_ Tangential Damping Classifier-free Guidance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e97d1c75d7d4d0dad5d4ced96eca19662d6f18e9 --- /dev/null +++ b/2025/TCFG_ Tangential Damping Classifier-free Guidance/layout.json @@ -0,0 +1,11537 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 141, + 103, + 470, + 121 + ], + "type": "text", + "content": "TCFG: Tangential Damping Classifier-free Guidance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 144, + 196, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 144, + 196, + 157 + ], + "spans": [ + { + "bbox": [ + 128, + 144, + 196, + 157 + ], + "type": "text", + "content": "Mingi Kwon*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 158, + 203, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 158, + 203, + 171 + ], + "spans": [ + { + "bbox": [ + 117, + 158, + 203, + 171 + ], + "type": "text", + "content": "Yonsei University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 174, + 220, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 174, + 220, + 184 + ], + "spans": [ + { + "bbox": [ + 99, + 174, + 220, + 184 + ], + "type": "text", + "content": "kwonmingi@yonsei.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 263, + 144, + 347, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 144, + 347, + 157 + ], + "spans": [ + { + "bbox": [ + 263, + 144, + 347, + 157 + ], + "type": "text", + "content": "Shin seong Kim*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 260, + 158, + 347, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 158, + 347, + 171 + ], + "spans": [ + { + "bbox": [ + 260, + 158, + 347, + 171 + ], + "type": "text", + "content": "Yonsei University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 249, + 174, + 358, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 174, + 358, + 185 + ], + "spans": [ + { + "bbox": [ + 249, + 174, + 358, + 185 + ], + "type": "text", + "content": "tltydl2@yonsei.ac.kr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 413, + 144, + 483, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 144, + 483, + 157 + ], + "spans": [ + { + "bbox": [ + 413, + 144, + 483, + 157 + ], + "type": "text", + "content": "Jaeseok Jeong" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 405, + 158, + 492, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 158, + 492, + 171 + ], + "spans": [ + { + "bbox": [ + 405, + 158, + 492, + 171 + ], + "type": "text", + "content": "Yonsei University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 387, + 174, + 510, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 174, + 510, + 185 + ], + "spans": [ + { + "bbox": [ + 387, + 174, + 510, + 185 + ], + "type": "text", + "content": "jete jeong@yonsei.ac.kr" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 198, + 192, + 268, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 192, + 268, + 205 + ], + "spans": [ + { + "bbox": [ + 198, + 192, + 268, + 205 + ], + "type": "text", + "content": "Yi Ting Hsiao" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 176, + 206, + 290, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 206, + 290, + 220 + ], + "spans": [ + { + "bbox": [ + 176, + 206, + 290, + 220 + ], + "type": "text", + "content": "University of Michigan" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 187, + 222, + 280, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 222, + 280, + 232 + ], + "spans": [ + { + "bbox": [ + 187, + 222, + 280, + 232 + ], + "type": "text", + "content": "hsiaoyt@umich.edu" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 192, + 419, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 192, + 419, + 205 + ], + "spans": [ + { + "bbox": [ + 345, + 192, + 419, + 205 + ], + "type": "text", + "content": "Youngjung Uh" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 339, + 206, + 426, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 206, + 426, + 220 + ], + "spans": [ + { + "bbox": [ + 339, + 206, + 426, + 220 + ], + "type": "text", + "content": "Yonsei University" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 331, + 222, + 430, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 222, + 430, + 232 + ], + "spans": [ + { + "bbox": [ + 331, + 222, + 430, + 232 + ], + "type": "text", + "content": "yj.uh@yonsei.ac.kr" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 151, + 260, + 200, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 260, + 200, + 273 + ], + "spans": [ + { + "bbox": [ + 151, + 260, + 200, + 273 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 286, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 296, + 563 + ], + "type": "text", + "content": "Diffusion models have achieved remarkable success in text-to-image synthesis, largely attributed to the use of classifier-free guidance (CFG), which enables high-quality, condition-aligned image generation. CFG combines the conditional score (e.g., text-conditioned) with the unconditional score to control the output. However, the unconditional score is in charge of estimating the transition between manifolds of adjacent timesteps from " + }, + { + "bbox": [ + 55, + 286, + 296, + 563 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 286, + 296, + 563 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 286, + 296, + 563 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 55, + 286, + 296, + 563 + ], + "type": "text", + "content": ", which may inadvertently interfere with the trajectory toward the specific condition. In this work, we introduce a novel approach that leverages a geometric perspective on the unconditional score to enhance CFG performance when conditional scores are available. Specifically, we propose a method that filters the singular vectors of both conditional and unconditional scores using singular value decomposition. This filtering process aligns the unconditional score with the conditional score, thereby refining the sampling trajectory to stay closer to the manifold. Our approach improves image quality with negligible additional computation. We provide deeper insights into the score function behavior in diffusion models and present a practical technique for achieving more accurate and contextually coherent image synthesis. project page" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 56, + 586, + 135, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 586, + 135, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 586, + 135, + 599 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 607, + 295, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 607, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 295, + 678 + ], + "type": "text", + "content": "Diffusion models [12, 31] have shown remarkable progress in image generation [19, 27, 30]. In particular, the emergence of classifier-free guidance [6, 11] (CFG) has attracted significant attention because it allows us to provide desired guidance by leveraging the conditional estimated score directly within the diffusion model." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 67, + 680, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 680, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 680, + 295, + 693 + ], + "type": "text", + "content": "The classifier-free guidance fundamentally computes the" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 317, + 261, + 432, + 371 + ], + "blocks": [ + { + "bbox": [ + 317, + 261, + 432, + 371 + ], + "lines": [ + { + "bbox": [ + 317, + 261, + 432, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 432, + 371 + ], + "type": "image", + "image_path": "706fdab5475f7a29b260c1b5cf8387caad4554e208c01e6cc676adeecc163e42.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 374, + 421, + 385 + ], + "lines": [ + { + "bbox": [ + 323, + 374, + 421, + 385 + ], + "spans": [ + { + "bbox": [ + 323, + 374, + 421, + 385 + ], + "type": "text", + "content": "(a) Classifier-free guidance" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. (a) Classifier-free guidance. When the unconditional score " + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "inline_equation", + "content": "s_{\\theta}(\\boldsymbol{z}_t)" + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "text", + "content": " and the conditional score " + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "inline_equation", + "content": "s_{\\theta}(\\boldsymbol{z}_t, y)" + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "text", + "content": " are misaligned, the result of CFG tends to fall off the manifold. (b) Our proposed method reduces the misalignment between the unconditional score " + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "inline_equation", + "content": "s_{\\theta}(\\boldsymbol{z}_t)" + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "text", + "content": " and the conditional score " + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "inline_equation", + "content": "s_{\\theta}(\\boldsymbol{z}_t, y)" + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "text", + "content": ", ensuring sampling aligns with the target manifold." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 438, + 261, + 553, + 371 + ], + "blocks": [ + { + "bbox": [ + 438, + 261, + 553, + 371 + ], + "lines": [ + { + "bbox": [ + 438, + 261, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 438, + 261, + 553, + 371 + ], + "type": "image", + "image_path": "482be4850a43f1cf1bc6a4faede5132a6dabcb2bcda05e78d8b2a17eabdffb60.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 374, + 523, + 384 + ], + "lines": [ + { + "bbox": [ + 492, + 374, + 523, + 384 + ], + "spans": [ + { + "bbox": [ + 492, + 374, + 523, + 384 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 473, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 545 + ], + "type": "text", + "content": "final score by combining the unconditional and conditional estimated scores. This approach ensures a generation that aligns well with the given condition. Additionally, using an appropriate guidance scale has been shown to enhance image quality across various tasks, further driving improvements in applications like text-to-image generation." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "content": "Let us say the guided score " + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "inline_equation", + "content": "\\tilde{s}_{\\theta}" + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "inline_equation", + "content": "\\tilde{s}_{\\theta} = s_{\\theta}^{\\mathrm{uncond}} + \\omega_{\\mathrm{scale}}(s_{\\theta}^{\\mathrm{cond}} - s_{\\theta}^{\\mathrm{uncond}})" + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "content": ". In text-to-image models, the text condition " + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "inline_equation", + "content": "(s_{\\theta}^{\\mathrm{cond}})" + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "content": " is randomly replaced with a null condition " + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "inline_equation", + "content": "(s_{\\theta}^{\\mathrm{uncond}})" + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "content": " during training (e.g., with a probability " + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "inline_equation", + "content": "p = 0.1" + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "content": "), enabling the null condition to act as a general estimator for any sample. It means that " + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "inline_equation", + "content": "s_{\\theta}^{\\mathrm{uncond}}" + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "content": " is the score estimated from " + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 313, + 557, + 554, + 652 + ], + "type": "text", + "content": " for all samples in the sampling trajectory of diffusion models." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "The unconditional score for any sample has certainly enabled the successful use of classifier-free guidance. However, we argue that there can be a misalignment between the unconditional and conditional estimated scores (See Eq. (2) in Sec. 2), which hinders the approximation toward the" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 132, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 132, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 132, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2620" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "manifold by the given condition. Fig. 1 (a) conceptually illustrates the potential issue that arises when the manifold of the unconditional score differs from that of the conditional score. In this paper, we show that this misalignment can be resolved with a simple algorithm, which significantly reduces the tendency of CFG to generate off-manifold samples, as illustrated in Fig. 1 (b)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 277 + ], + "type": "text", + "content": "Our approach is based on the following insights. First, the score predicted by the diffusion model estimates the intrinsic dimension of the data manifold [32]. Additionally, this intrinsic dimension can be captured by the tangent space of the target manifold [3, 9]. Instead of directly estimating the intrinsic dimension, we focus on utilizing the tangential component inherent in the unconditional score during classifier-free guidance. By reducing its misalignment with the conditional score, we enhance the alignment and ultimately improve the quality of the generated outputs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "text", + "content": "Specifically, we push the score " + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\tilde{s}_{\\theta}" + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "text", + "content": " toward the normal direction of the conditional manifold by eliminating the values of column vectors with small singular values using the orthogonal matrix " + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "text", + "content": " obtained through the singular value decomposition of the conditional and unconditional scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 447 + ], + "type": "text", + "content": "In this paper, we propose a novel sampling method that leverages the unconditional score within CFG. To support our approach, we first lay out the theoretical foundation in section Sec. 2 and Sec. 3, discussing the manifold hypothesis and its connection to diffusion models. In Sec. 4, we provide a comprehensive explanation of our proposed method. This is followed by a detailed analysis using a toy example in Sec. 5, and we demonstrate the practical applicability of our method on real-world text-to-image models in Sec. 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 567 + ], + "type": "text", + "content": "Our experiments show a significant improvement in the MS-COCO Fréchet Inception Distance (FID) across various models that utilize classifier-free guidance, e.g., diffusion models (Stable Diffusion v1.5 [26] and SDXL [23]) and rectified flow (Stable Diffusion 3 [8]). Additionally, our method improves DiT [21] FID on ImageNet. Notably, our method helps mitigate the overexposure bias problem, leading to resulting images that better align with the underlying data distribution, as supported by improved quantitative metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 134, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 134, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 134, + 596 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": "Diffusion models Diffusion models learn the score that reverses the forward noising process. This forward process from the real data distribution " + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_0)" + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": " to a latent distribution " + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "inline_equation", + "content": "p(z_1) \\sim N(0, \\sigma_{\\max}^2 I)" + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": " along timesteps " + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "inline_equation", + "content": "t \\in [0, 1]" + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": " is defined by a Gaussian kernel: " + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "inline_equation", + "content": "z_t = x_0 + \\sigma(t)\\epsilon" + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\sigma(t)" + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": " is a noise schedule where " + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\sigma(0) = 0" + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\sigma(1) = \\sigma_{\\max}" + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": ", determining the amount of noise to be added at each timestep " + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": " to erase information from " + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 604, + 295, + 699 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "A generative process is represented as its reverse with a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 471, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 471, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 471, + 84 + ], + "type": "text", + "content": "stochastic differential equation (SDE):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 91, + 547, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 547, + 122 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 547, + 122 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {d} \\boldsymbol {z} = - \\dot {\\sigma} (t) \\sigma (t) \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\mathrm {d} t \\\\ - \\beta (t) \\sigma (t) ^ {2} \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) d t + \\sqrt {2 \\beta (t)} \\sigma (t) d \\omega_ {t}, \\\\ \\end{array}", + "image_path": "6a7991aff9c4fd0d60b733f2f7e8942df1ae768dc1fdbe2e05ce7ce195a4b4e3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 130, + 553, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 553, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 553, + 154 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 130, + 553, + 154 + ], + "type": "inline_equation", + "content": "\\mathrm{d}\\omega_{t}" + }, + { + "bbox": [ + 313, + 130, + 553, + 154 + ], + "type": "text", + "content": " is a standard Wiener process. Alternatively, it can be expressed as an ordinary differential equation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 362, + 162, + 504, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 162, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 362, + 162, + 504, + 175 + ], + "type": "interline_equation", + "content": "\\mathrm {d} \\boldsymbol {z} = - \\dot {\\sigma} (t) \\sigma (t) \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\mathrm {d} t.", + "image_path": "1f47e97b5598e46de6100e55fc55b7dffa47c8e50b72665d9a91ecea50cc7fea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "content": "Diffusion models approximate the score function " + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{z}_t)" + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "content": " with a neural network " + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "inline_equation", + "content": "s_\\theta (\\boldsymbol {z}_t,t)" + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "content": ". They are trained to predict the clean data from the noisy " + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_t" + }, + { + "bbox": [ + 313, + 182, + 553, + 231 + ], + "type": "text", + "content": ". The trained model performs the reverse process using:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 238, + 502, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 238, + 502, + 263 + ], + "spans": [ + { + "bbox": [ + 365, + 238, + 502, + 263 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\approx \\frac {\\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t} , t) - \\boldsymbol {z} _ {t}}{\\sigma (t) ^ {2}}.", + "image_path": "1f2a88d3459b73a8bcddd053730aff7a5545f7cd7f5567c254567551d9f2eb6f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 275, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 553, + 312 + ], + "type": "text", + "content": "Classifier guidance (CG) and classifier-free guidance (CFG) For an arbitrary class label " + }, + { + "bbox": [ + 313, + 275, + 553, + 312 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 275, + 553, + 312 + ], + "type": "text", + "content": ", CG defines the class-conditional sampling distribution " + }, + { + "bbox": [ + 313, + 275, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\tilde{p}_{\\theta}(\\boldsymbol{z}_t \\mid y)" + }, + { + "bbox": [ + 313, + 275, + 553, + 312 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 360, + 319, + 506, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 319, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 360, + 319, + 506, + 332 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) \\propto p _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) p _ {\\theta} (y \\mid \\boldsymbol {z} _ {t}) ^ {\\gamma},", + "image_path": "68e2197d1cbfd5cdc34db9b82505fbe0e6db21dadf17d0a52996bf640f73cb95.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y \\mid z_t)" + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": " is the classifier distribution and " + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": " is a scaling parameter. [6] When " + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "inline_equation", + "content": "\\gamma > 0" + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": ", it is known to reduce sample diversity but enhance quality. However, CG requires a classifier that can predict label " + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": " from the noisy " + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": ". CFG proposes a method to sample from the conditional distribution by expressing the classifier distribution " + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y \\mid z_t)" + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": " in terms of the conditional distribution " + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "inline_equation", + "content": "p_{\\theta}(z_t \\mid y)" + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": " and the unconditional distribution " + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "inline_equation", + "content": "p_{\\theta}(z_t)" + }, + { + "bbox": [ + 313, + 340, + 553, + 436 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 443, + 510, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 443, + 510, + 456 + ], + "spans": [ + { + "bbox": [ + 356, + 443, + 510, + 456 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) \\propto p _ {\\theta} \\left(\\boldsymbol {z} _ {t} \\mid y\\right) ^ {1 + \\gamma} p _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) ^ {- \\gamma}.", + "image_path": "b20f03f60d1903ea4a67f03fa0e6869c3ee958fba9192e039d662fb32375e3ad.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 464, + 553, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 553, + 488 + ], + "type": "text", + "content": "As a result, the final score " + }, + { + "bbox": [ + 313, + 464, + 553, + 488 + ], + "type": "inline_equation", + "content": "\\nabla_{z_t}\\log \\tilde{p}_\\theta (\\boldsymbol {z}_t\\mid y)" + }, + { + "bbox": [ + 313, + 464, + 553, + 488 + ], + "type": "text", + "content": " is approximated by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 323, + 496, + 544, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 496, + 544, + 524 + ], + "spans": [ + { + "bbox": [ + 323, + 496, + 544, + 524 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\boldsymbol {z} _ {t}} \\log \\tilde {p} _ {\\theta} (\\boldsymbol {z} _ {t} \\mid y) = (1 + \\gamma) \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}, y) - \\gamma \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}) \\\\ = \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}) + \\omega \\left(\\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t}, y) - \\boldsymbol {s} _ {\\theta} (\\boldsymbol {z} _ {t})\\right), \\\\ \\end{array}", + "image_path": "a4ec20392b2669ccc0e5a38955bbe589ebe9a591409bfbc3135a46dd29b5864b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 532, + 410, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 410, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 410, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 410, + 544 + ], + "type": "inline_equation", + "content": "\\omega = 1 + \\gamma" + }, + { + "bbox": [ + 313, + 532, + 410, + 544 + ], + "type": "text", + "content": ". [11]" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "content": "In practice, both " + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "inline_equation", + "content": "s_{\\theta}(x_t,y)" + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t)" + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "content": " are approximated by a single neural network that is jointly trained to estimate both the conditional and unconditional scores. Text-to-image models use the null condition " + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{null}} = \\varnothing" + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "content": " as a class label to train " + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t)\\approx s_{\\theta}(z_t,y_{\\mathrm{null}})" + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "content": ". This approach allows " + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t,y) - s_{\\theta}(z_t)" + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "content": " to provide guidance similar to the gradient of an implicit classifier. Hereafter, we will simply denote " + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t,y_{\\mathrm{null}})" + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t)" + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Diffusion models and data manifold encoding The manifold hypothesis suggests that high-dimensional data lies on or near a lower-dimensional manifold, making intrinsic dimension estimation essential for data representation [9]. This intrinsic dimension is often encoded in the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2621" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "manifold's tangent spaces, which capture underlying degrees of freedom and align local structures to reveal the global geometry [3, 33]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": "Building on these ideas, further studies have analyzed the approximation and generalization capabilities of diffusion models [20, 22, 24], and have also proven that their score functions can approximate the tangent space of the data manifold [32]. In particular, for a compact embedded sub-manifold " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\subset \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": ", it has been shown that for a sample " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "z_t \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": " sufficiently close to the target data, the score " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\nabla_{z_t} \\log p_t(z_t) (\\approx s_\\theta(z_t))" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": " and orthogonal projection " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\pi(z_t)" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": " onto data manifold " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": " satisfy a key relationship. For the projection " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_p" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": " onto the normal space and " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_p" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": " onto the tangent space of " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": ", the ratio of their magnitudes goes to zero as " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": " approaches 0 (i.e., gets closer to the target data). In other words, for samples " + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 55, + 109, + 296, + 276 + ], + "type": "text", + "content": " close to the target data, the following equation holds:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 96, + 284, + 295, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 284, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 96, + 284, + 295, + 312 + ], + "type": "interline_equation", + "content": "\\frac {\\left\\| \\mathbf {T} _ {p} \\nabla_ {\\mathbf {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\right\\|}{\\left\\| \\mathbf {N} _ {p} \\nabla_ {\\mathbf {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\right\\|} \\rightarrow 0, \\quad \\text {a s} t \\rightarrow 0, \\tag {1}", + "image_path": "98b769d3bf2abdc8c875f1ccc47138219d298323f67d9a4859a38e71e26c5264.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_p" + }, + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_p" + }, + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "text", + "content": " are the projection operators onto the tangent space " + }, + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\pi (z_t)}\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "text", + "content": " and the normal space " + }, + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\pi (z_t)}\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 319, + 295, + 368 + ], + "type": "text", + "content": ", respectively (for a detailed proof, see Theorem 4.1, Corollary 4.2 and Appendix D in [32])." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 369, + 296, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 423 + ], + "type": "text", + "content": "This implies that, for samples sufficiently close to the target manifold " + }, + { + "bbox": [ + 55, + 369, + 296, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 369, + 296, + 423 + ], + "type": "text", + "content": ", the cosine similarity between the score function and the normal vector " + }, + { + "bbox": [ + 55, + 369, + 296, + 423 + ], + "type": "inline_equation", + "content": "\\mathbf{n} = \\frac{\\pi(z_t) - z_t}{\\|\\pi(z_t) - z_t\\|}" + }, + { + "bbox": [ + 55, + 369, + 296, + 423 + ], + "type": "text", + "content": " converges to 1 (i.e., " + }, + { + "bbox": [ + 55, + 369, + 296, + 423 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{cos}}(\\mathbf{n},\\nabla_{z_t}\\log p_t(z_t))\\xrightarrow{t\\to 0}1" + }, + { + "bbox": [ + 55, + 369, + 296, + 423 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "text", + "content": "This suggests that for a sample " + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "text", + "content": " very close to the target, the score function " + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{z}}\\log p_t(z_t)\\approx s_\\theta (z_t)" + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "text", + "content": " becomes an element of the normal space of the target manifold (that is, " + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{z}}\\log p_t(z_t)\\in \\mathcal{N}_{\\pi (z_t)}\\mathcal{M}_0\\approx \\mathcal{N}_{\\pi (z_0)}\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "text", + "content": " for sufficiently small " + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "text", + "content": "). Leveraging this property, the estimated diffusion score can approximate the intrinsic dimension of the target data by utilizing the huge gap in the singular values of the sampling scores " + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "inline_equation", + "content": "S = \\left[s_{\\theta}\\left(\\boldsymbol {z}_t^{(1)},t\\right),\\dots ,s_{\\theta}\\left(\\boldsymbol {z}_t^{(4n)},t\\right)\\right]" + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "text", + "content": " where the singular vectors corresponding to the higher singular values represent the normal components of " + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 424, + 296, + 571 + ], + "type": "text", + "content": ", while those corresponding to the lower singular values represent the tangential components.[32]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 583, + 117, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 117, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 117, + 594 + ], + "type": "text", + "content": "3. Intuition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 603, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 296, + 676 + ], + "type": "text", + "content": "In this section, we assume the mathematical concept behind our method and supporting experiments. Our approach refines CFG at each step by dropping the tangential component of the unconditional score, enhancing the quality of conditional generation. This adjustment allows the conditional score to guide the generated sample more directly" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 553, + 180 + ], + "type": "image", + "image_path": "2dc8ff2f56920b200ea0696cc91ea6a29dd819c8d4c64eb7681009255a20c7e6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. Singular values of the score function across all timesteps. We computed the singular values for all timesteps using a total of 17,000 samples from Stable Diffusion v1.5. For both the unconditional and the conditional scores, a significant drop in singular values was observed at indices close to 0 across all timesteps. This suggests the existence of an intermediate manifold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "type": "text", + "content": "toward the manifold specified by the condition, improving alignment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "text", + "content": "To support this, we provide empirical evidence suggesting that not only does the target data manifold " + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "text", + "content": " exist but there is also a manifold " + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t - \\epsilon}" + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "inline_equation", + "content": "t\\in (0,1)" + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{z}}\\log p_t(\\pmb {z}_t)\\in \\mathcal{N}_{\\pi_{t - \\epsilon}(\\pmb {z}_t)}\\mathcal{M}_{t - \\epsilon}" + }, + { + "bbox": [ + 313, + 301, + 554, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": "There exists an intermediate manifold " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": " We hypothesize the existence of a manifold " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{(t - \\epsilon)}" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": " that contains " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb {z}_t}\\log p_t(\\pmb {z}_t)" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": " as elements of its normal space, not only for samples close to the target data but also for " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "t\\in (0,1)" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": ". Specifically, we assume the following:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "text", + "content": "Assumption 1. Suppose that the support of the data distribution " + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "text", + "content": " is contained in a compact embedded submanifold " + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0\\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "text", + "content": " be the distribution of latents at time " + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "text", + "content": " diffused from " + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "text", + "content": ". Then, under mild assumptions², " + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\forall t\\in (0,1)" + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\exists t'\\in (t - \\epsilon ,t + \\epsilon)" + }, + { + "bbox": [ + 313, + 430, + 554, + 491 + ], + "type": "text", + "content": " such that:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 499, + 498, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 499, + 498, + 514 + ], + "spans": [ + { + "bbox": [ + 370, + 499, + 498, + 514 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} (\\boldsymbol {z} _ {t}) \\in \\mathcal {N} _ {\\pi_ {t ^ {\\prime}} (\\boldsymbol {z} _ {t})} \\mathcal {M} _ {t ^ {\\prime}},", + "image_path": "b67b67d6095a963f41acc0ff3d9dc88063b1ebcb7d2740dd02e43e1c01186c09.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": "for sufficiently small " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": " and orthogonal projection " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\pi_{t'}(\\mathbf{z}_t)" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": " onto manifold " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t'}" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": ". This hypothesis is indirectly supported by the clear gap in singular values arranged in descending order for a sufficient number of samples. This phenomenon occurs not only when " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": " goes to 0 (near the image manifold) but also consistently for all time step " + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "inline_equation", + "content": "t \\in (0,1)" + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": "To observe the gap, we compute 17,000 score samples across all timesteps on Stable Diffusion v1.5. Let " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "[\\sigma_1,\\dots ,\\sigma_D]" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " represent the singular values from the SVD applied to " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_{t})" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_{t},y)" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": ", with 17,000 samples collected per timestep, arranged in descending order. The corresponding singular vectors are denoted as" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "1Every compact embedded submanifold of " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": " has a tubular neighborhood, and for a given manifold " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": ", each point " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": " within this tubular neighborhood has a unique orthogonal projection " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": " [16]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 683, + 554, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 554, + 712 + ], + "type": "text", + "content": "2) The distribution " + }, + { + "bbox": [ + 313, + 683, + 554, + 712 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 313, + 683, + 554, + 712 + ], + "type": "text", + "content": " has a smooth density " + }, + { + "bbox": [ + 313, + 683, + 554, + 712 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 683, + 554, + 712 + ], + "type": "text", + "content": " w.r.t the volume measure on the manifold. 2) The density " + }, + { + "bbox": [ + 313, + 683, + 554, + 712 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 683, + 554, + 712 + ], + "type": "text", + "content": " is bounded away from zero on the manifold." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2622" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 272, + 150 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 272, + 150 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 272, + 150 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 272, + 150 + ], + "type": "image", + "image_path": "2cb5021bbd32ce4dcf5b043fddcc1d4dbfe00ea27913e7103fae31a5322fbd8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 159, + 295, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 226 + ], + "type": "text", + "content": "Figure 3. Cosine similarity between singular vectors of unconditional and conditional scores. We computed the singular vectors " + }, + { + "bbox": [ + 55, + 159, + 295, + 226 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 159, + 295, + 226 + ], + "type": "text", + "content": " at each timestep using a total of 17,000 samples from Stable Diffusion v1.5. We observe the similarity of significant singular vectors (i.e., those with indices close to 0) between unconditional and conditional scores are mostly high across all timesteps " + }, + { + "bbox": [ + 55, + 159, + 295, + 226 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 159, + 295, + 226 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "inline_equation", + "content": "[\\pmb{v}_1,\\dots ,\\pmb{v}_D]^T" + }, + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "inline_equation", + "content": "\\pmb{s}_{\\theta}(\\pmb {z}_t)" + }, + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "inline_equation", + "content": "[\\hat{\\pmb{v}}_1,\\dots ,\\hat{\\pmb{v}}_D]^T" + }, + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "inline_equation", + "content": "s_\\theta (z_t,y)" + }, + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "text", + "content": " respectively " + }, + { + "bbox": [ + 55, + 236, + 296, + 261 + ], + "type": "inline_equation", + "content": "(D\\approx^{3}17,000)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": "As shown in Fig. 2, both " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t)" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t,y)" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": " have gaps between the highest singular values and the rest for all " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": ", not just for " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "0 + \\epsilon" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": ". Interpreting from the perspective that the score function " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t)" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": " becomes an element of the data manifold's normal space as " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": " approaches 0 [32]. Assuming the existence of an intermediate manifold " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "t \\in (0,1)" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": ", this suggests that the singular vectors associated with the largest singular values contain dominant components of " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{NM}_t" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": ", while vectors associated with smaller singular values correspond to component of " + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\mathcal{M}_t" + }, + { + "bbox": [ + 55, + 273, + 296, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "type": "text", + "content": "Tangential misalignment between unconditional and conditional score We empirically justify the principle of modifying the unconditional score by dropping the components with low singular values and retaining only the components with high singular values." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 514 + ], + "type": "text", + "content": "Fig. 3 shows that conditional and unconditional singular vectors " + }, + { + "bbox": [ + 55, + 468, + 296, + 514 + ], + "type": "inline_equation", + "content": "[\\pmb{v}_1, \\dots, \\pmb{v}_D]^T" + }, + { + "bbox": [ + 55, + 468, + 296, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 468, + 296, + 514 + ], + "type": "inline_equation", + "content": "[\\hat{\\pmb{v}}_1, \\dots, \\hat{\\pmb{v}}_D]^T" + }, + { + "bbox": [ + 55, + 468, + 296, + 514 + ], + "type": "text", + "content": " at corresponding indices are more similar when their singular values are high than the rest." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "text", + "content": "More specifically, the cosine similarity of the singular vectors " + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\pmb{v}_1" + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{v}}_1" + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "text", + "content": " associated with the highest singular value " + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\sigma_1" + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "inline_equation", + "content": "s_\\theta(z_t)" + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "inline_equation", + "content": "s_\\theta(z_t, y)" + }, + { + "bbox": [ + 55, + 516, + 296, + 563 + ], + "type": "text", + "content": ", respectively, is higher than the others." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 571, + 294, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 571, + 294, + 614 + ], + "spans": [ + { + "bbox": [ + 71, + 571, + 294, + 614 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left[ S _ {\\cos} \\left(\\boldsymbol {v} _ {1}, \\hat {\\boldsymbol {v}} _ {1}\\right) > S _ {\\cos} \\left(\\boldsymbol {v} _ {j}, \\hat {\\boldsymbol {v}} _ {j}\\right) \\right] \\\\ \\approx \\left[ S _ {\\cos} \\left(\\mathbf {N} _ {p} \\nabla_ {z _ {t}} \\log p _ {t} \\left(z _ {t}, y\\right), \\mathbf {N} _ {p} \\nabla_ {z _ {t}} \\log p _ {t} \\left(z _ {t}\\right)\\right) \\right. \\tag {2} \\\\ \\left. > S _ {\\cos} \\left(\\mathbf {T} _ {p} \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} \\left(\\boldsymbol {z} _ {t}, y\\right), \\mathbf {T} _ {p} \\nabla_ {\\boldsymbol {z} _ {t}} \\log p _ {t} \\left(\\boldsymbol {z} _ {t}\\right)\\right) \\right] \\\\ \\end{array}", + "image_path": "c09a424b50bd9faa44e6137f47818c09a9dee16559e658fcf1af3a2f328db387.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "inline_equation", + "content": "1 < j \\leq D" + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "text", + "content": ". The cosine similarity " + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "inline_equation", + "content": "S_{cos}" + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "text", + "content": " between two vectors " + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j" + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{cos}}(\\mathbf{v}_i, \\mathbf{v}_j) = \\frac{\\mathbf{v}_i \\cdot \\mathbf{v}_j}{\\|\\mathbf{v}_i\\| \\|\\mathbf{v}_j\\|}" + }, + { + "bbox": [ + 55, + 623, + 295, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "This indicates that the intermediate manifolds associated with " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{z}_t)" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{z}_t,y)" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " share similar normal components, while their tangent components are relatively less aligned." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 121, + 554, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 192 + ], + "type": "text", + "content": "These less-aligned components interfere with the generative process, making it harder to align with the target manifold. We modify the unconditional score " + }, + { + "bbox": [ + 313, + 121, + 554, + 192 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t)" + }, + { + "bbox": [ + 313, + 121, + 554, + 192 + ], + "type": "text", + "content": " at each timestep by removing its tangential components that are less aligned with the conditional score " + }, + { + "bbox": [ + 313, + 121, + 554, + 192 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t,y)" + }, + { + "bbox": [ + 313, + 121, + 554, + 192 + ], + "type": "text", + "content": ". We provide detailed methods in the following section." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 202, + 375, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 375, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 375, + 215 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": "Our main method proceeds as follows. At each step, we take the predicted unconditional score " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t)" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": " and the conditional score " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t,y)" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": " and concatenate them into a score matrix " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "A = [s_{\\theta}(z_t),s_{\\theta}(z_t,y)]" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": ". Next, we perform SVD on " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": ", obtaining singular values and corresponding singular vectors that consider both components " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "s(z_{t})" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "s(z_{t},y)" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": ". This results in singular vectors " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "[\\pmb {v}_1,\\pmb {v}_2,\\dots ,\\pmb {v}_D]^T" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{1}" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": " is the normal component of both " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "s(z_{t})" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "s(z_{t},y)" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": ". We project the unconditional score onto " + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{1}" + }, + { + "bbox": [ + 313, + 222, + 554, + 331 + ], + "type": "text", + "content": " and drop the rest." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 338, + 553, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 338, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 370, + 338, + 553, + 353 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {s}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) = \\boldsymbol {s} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) \\cdot \\boldsymbol {V} ^ {T} \\cdot \\left[ \\boldsymbol {v} _ {1}, \\boldsymbol {0} \\right]. \\tag {3}", + "image_path": "9b310f42f823eea108b292e627779cd6312ec2ae8b1fa36077465c5f20beddbb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": "Consequently, the singular vectors associated with high singular values in the score matrix " + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": " retain the well-aligned, normal components of " + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t)" + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t,y)" + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": ", while those with lower singular values represent misaligned tangential components, which we set to zero in Eq. (3) to drop these components from the unconditional score. Next, we update the score " + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\hat{s}_{\\theta}(z_t,y)" + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": " with classifier-free guidance (CFG):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 453, + 553, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 453, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 322, + 453, + 553, + 465 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {z} _ {t}} \\log \\hat {p} _ {t} \\left(\\boldsymbol {z} _ {t} | y\\right) = \\hat {\\boldsymbol {s}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right) + w \\left(\\boldsymbol {s} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, y\\right) - \\hat {\\boldsymbol {s}} _ {\\theta} \\left(\\boldsymbol {z} _ {t}\\right)\\right). \\tag {4}", + "image_path": "c7bb39618a0e51eabbe7f42911a1ad438d7115f0085d7befec89a6b94b56b787.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 485, + 553, + 509 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 485, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 315, + 485, + 553, + 509 + ], + "type": "text", + "content": "Algorithm 1 Tangential damping classifier-free guidance (TCFG)" + } + ] + } + ], + "index": 16, + "type": "text" + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "text", + "content": "Inputs: " + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\theta}(\\mathbf{z}_t)" + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\theta}(\\mathbf{z}_t|\\mathbf{y})" + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "text", + "content": ": predicted unconditional and conditional scores, " + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "inline_equation", + "content": "t \\in (0,1)" + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "text", + "content": ": time step, " + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "text", + "content": ": condition, " + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 510, + 554, + 545 + ], + "type": "text", + "content": ": CFG scale." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 547, + 365, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 547, + 365, + 559 + ], + "spans": [ + { + "bbox": [ + 315, + 547, + 365, + 559 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 315, + 547, + 365, + 559 + ], + "type": "inline_equation", + "content": "z_0" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 560, + 527, + 680 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 321, + 560, + 401, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 560, + 401, + 572 + ], + "spans": [ + { + "bbox": [ + 321, + 560, + 401, + 572 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 321, + 560, + 401, + 572 + ], + "type": "inline_equation", + "content": "t \\in (0, 1)" + }, + { + "bbox": [ + 321, + 560, + 401, + 572 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 573, + 408, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 573, + 408, + 584 + ], + "spans": [ + { + "bbox": [ + 321, + 573, + 408, + 584 + ], + "type": "text", + "content": "2: Get " + }, + { + "bbox": [ + 321, + 573, + 408, + 584 + ], + "type": "inline_equation", + "content": "s_{\\theta}" + }, + { + "bbox": [ + 321, + 573, + 408, + 584 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 321, + 573, + 408, + 584 + ], + "type": "inline_equation", + "content": "z_{t}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 585, + 520, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 585, + 520, + 597 + ], + "spans": [ + { + "bbox": [ + 321, + 585, + 520, + 597 + ], + "type": "text", + "content": "3: Make score matrix " + }, + { + "bbox": [ + 321, + 585, + 520, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = \\left[s_{\\theta}\\left(z_{t}\\right), s_{\\theta}\\left(z_{t}, y\\right)\\right]" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 597, + 502, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 597, + 502, + 609 + ], + "spans": [ + { + "bbox": [ + 321, + 597, + 502, + 609 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 321, + 597, + 502, + 609 + ], + "type": "inline_equation", + "content": "(\\sigma_{i})_{i = 1}^{d},(\\pmb{w}_{i})_{i = 1}^{d},(\\pmb{v}_{i})_{i = 1}^{d}\\gets \\mathrm{SVD}(\\pmb {A})" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 609, + 470, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 609, + 470, + 620 + ], + "spans": [ + { + "bbox": [ + 321, + 609, + 470, + 620 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 321, + 609, + 470, + 620 + ], + "type": "inline_equation", + "content": "\\hat{s}_{\\theta}(\\pmb{z}_t) = s_{\\theta}(\\pmb{z}_t)\\cdot \\pmb{V}^T\\cdot [\\pmb{v}_1,\\pmb {0}]" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 621, + 461, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 621, + 461, + 632 + ], + "spans": [ + { + "bbox": [ + 321, + 621, + 461, + 632 + ], + "type": "text", + "content": "6: (Dropping " + }, + { + "bbox": [ + 321, + 621, + 461, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{T}\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol{x}_t))" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 632, + 527, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 632, + 527, + 644 + ], + "spans": [ + { + "bbox": [ + 321, + 632, + 527, + 644 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 321, + 632, + 527, + 644 + ], + "type": "inline_equation", + "content": "\\hat{s}_{\\theta}(\\boldsymbol{z}_t, y) = \\hat{s}_{\\theta}(\\boldsymbol{z}_t) + w(\\boldsymbol{s}_{\\theta}(\\boldsymbol{z}_t, y) - \\hat{s}_{\\theta}(\\boldsymbol{z}_t))" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 645, + 389, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 645, + 389, + 656 + ], + "spans": [ + { + "bbox": [ + 321, + 645, + 389, + 656 + ], + "type": "text", + "content": "8: Update " + }, + { + "bbox": [ + 321, + 645, + 389, + 656 + ], + "type": "inline_equation", + "content": "z_{t}" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 657, + 364, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 657, + 364, + 666 + ], + "spans": [ + { + "bbox": [ + 318, + 657, + 364, + 666 + ], + "type": "text", + "content": "9: end for" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 669, + 377, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 669, + 377, + 680 + ], + "spans": [ + { + "bbox": [ + 318, + 669, + 377, + 680 + ], + "type": "text", + "content": "10: Output " + }, + { + "bbox": [ + 318, + 669, + 377, + 680 + ], + "type": "inline_equation", + "content": "z_0" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 331, + 682, + 553, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 682, + 553, + 707 + ], + "spans": [ + { + "bbox": [ + 331, + 682, + 553, + 707 + ], + "type": "inline_equation", + "content": "(\\sigma_{i})_{i = 1}^{d},(\\mathbf{w}_{i})_{i = 1}^{d},(\\mathbf{v}_{i})_{i = 1}^{d}" + }, + { + "bbox": [ + 331, + 682, + 553, + 707 + ], + "type": "text", + "content": " denote singular values, left and right singular vectors respectively." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "text", + "content": "3Approximately " + }, + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "inline_equation", + "content": "4\\times n" + }, + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "text", + "content": " samples are sufficient to accurately estimate the intrinsic dimension of the target manifold " + }, + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "text", + "content": " [32]. However, our goal is to verify the existence of a manifold where " + }, + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{z}_t}\\log p_t(\\boldsymbol {z}_t)" + }, + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "text", + "content": " is an element of the normal space. Therefore, it suffices to observe the presence of a large gap in the singular value spectrum, thus " + }, + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "inline_equation", + "content": "N < D" + }, + { + "bbox": [ + 55, + 664, + 296, + 713 + ], + "type": "text", + "content": " is enough." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2623" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 74, + 176, + 145 + ], + "blocks": [ + { + "bbox": [ + 90, + 74, + 176, + 145 + ], + "lines": [ + { + "bbox": [ + 90, + 74, + 176, + 145 + ], + "spans": [ + { + "bbox": [ + 90, + 74, + 176, + 145 + ], + "type": "image", + "image_path": "0080c29da4b011faefc3d7b8d9ae5d398c2b87e041f9109d754939c67993f165.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 74, + 262, + 146 + ], + "blocks": [ + { + "bbox": [ + 178, + 74, + 262, + 146 + ], + "lines": [ + { + "bbox": [ + 178, + 74, + 262, + 146 + ], + "spans": [ + { + "bbox": [ + 178, + 74, + 262, + 146 + ], + "type": "image", + "image_path": "adb4f0bcc1d9f2ef2b23e95c1a0c96e5997f0fec96dbeb52c59bc11932d8932f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 150, + 242, + 158 + ], + "lines": [ + { + "bbox": [ + 194, + 150, + 242, + 158 + ], + "spans": [ + { + "bbox": [ + 194, + 150, + 242, + 158 + ], + "type": "text", + "content": "(b) Classifier-free" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 90, + 160, + 176, + 232 + ], + "blocks": [ + { + "bbox": [ + 97, + 149, + 170, + 159 + ], + "lines": [ + { + "bbox": [ + 97, + 149, + 170, + 159 + ], + "spans": [ + { + "bbox": [ + 97, + 149, + 170, + 159 + ], + "type": "text", + "content": "(a) Conditional scores only" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 160, + 176, + 232 + ], + "lines": [ + { + "bbox": [ + 90, + 160, + 176, + 232 + ], + "spans": [ + { + "bbox": [ + 90, + 160, + 176, + 232 + ], + "type": "image", + "image_path": "b0c1ba6a9bc0e9751c7e789c548f4184c8b7f6c4d6f6bcd950d9cfb30a497644.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 235, + 176, + 243 + ], + "lines": [ + { + "bbox": [ + 88, + 235, + 176, + 243 + ], + "spans": [ + { + "bbox": [ + 88, + 235, + 176, + 243 + ], + "type": "text", + "content": "(c) Calculate SVD on all samples" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 253, + 296, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 319 + ], + "type": "text", + "content": "Figure 4. Sampling results on different methods with diffusion model trained on two moons dataset. Our proposed methods (c, d) demonstrate a closer match to the target distribution compared to using conditional scores only or CFG. In (c), SVD is computed across all samples, while in (d), SVD is calculated separately for each pair of conditional and unconditional scores." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 160, + 262, + 232 + ], + "blocks": [ + { + "bbox": [ + 178, + 160, + 262, + 232 + ], + "lines": [ + { + "bbox": [ + 178, + 160, + 262, + 232 + ], + "spans": [ + { + "bbox": [ + 178, + 160, + 262, + 232 + ], + "type": "image", + "image_path": "c805e93368939886f4821b153ed857109d6328029ad5c26cd336f4b5990b73d4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 235, + 230, + 243 + ], + "lines": [ + { + "bbox": [ + 206, + 235, + 230, + 243 + ], + "spans": [ + { + "bbox": [ + 206, + 235, + 230, + 243 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 332, + 251, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 251, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 251, + 343 + ], + "type": "text", + "content": "We provide a detailed algorithm in Algorithm 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 344, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 439 + ], + "type": "text", + "content": "Unlike traditional CFG update methods, " + }, + { + "bbox": [ + 55, + 344, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{z}_t}\\log \\hat{p_t} (\\boldsymbol {z}_t|\\boldsymbol {y})" + }, + { + "bbox": [ + 55, + 344, + 296, + 439 + ], + "type": "text", + "content": " drops tangential component from the unconditional score at each step. It prevents accumulating misaligned components from the unconditional score " + }, + { + "bbox": [ + 55, + 344, + 296, + 439 + ], + "type": "inline_equation", + "content": "s_\\theta (z_t)" + }, + { + "bbox": [ + 55, + 344, + 296, + 439 + ], + "type": "text", + "content": " using the direction of the manifold defined by the given condition " + }, + { + "bbox": [ + 55, + 344, + 296, + 439 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 344, + 296, + 439 + ], + "type": "text", + "content": " over time evolution. This concept is further illustrated with a simple distribution in Sec. 5, where the toy example clarifies the benefits of our methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 452, + 135, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 135, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 135, + 466 + ], + "type": "text", + "content": "5. Toy example" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "We empirically verify our method on a toy problem, generating the two moons dataset. Experiments consist of the generated samples with different guidances including the original classifier-free guidance (CFG) and ours, and the sampling trajectories following their respective score functions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "content": "The target data distribution " + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "inline_equation", + "content": "p(X_0)" + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "content": " consists of samples distributed along two distinct curves (moons). We trained a conditional diffusion model using a small neural network that receives a binary label " + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "content": " for the two moons or " + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "inline_equation", + "content": "y = \\varnothing" + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "content": " denoting the null condition. For detailed settings, please refer to the Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": "Fig. 4 shows the generated samples using four different guiding strategies. (a) uses only the conditional score " + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t, y)" + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": ". (b) uses the CFG score. (c) and (d) employ our guidance score at each step with multiple samples and one sample, respectively, to compute singular value decomposition (SVD) of the unconditional score " + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "inline_equation", + "content": "s_{\\theta}^{i}(z_t)" + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": " and conditional score " + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "inline_equation", + "content": "s_{\\theta}^{i}(z_t, y)" + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "According to the result, generated samples using our" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 420, + 171 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 420, + 171 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 420, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 420, + 171 + ], + "type": "image", + "image_path": "29c3ce1ec7ae83c8421d95b085b9588d7688c0e8c2061cc3f7fc5791669d81c4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 172, + 399, + 180 + ], + "lines": [ + { + "bbox": [ + 336, + 172, + 399, + 180 + ], + "spans": [ + { + "bbox": [ + 336, + 172, + 399, + 180 + ], + "type": "text", + "content": "Trajectory of sample A" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 191, + 553, + 268 + ], + "lines": [ + { + "bbox": [ + 313, + 191, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 553, + 268 + ], + "type": "text", + "content": "Figure 5. Visualization of the sampling trajectory. In CFG (orange path), the unconditional scores (red arrows) include components that point towards directions other than the target distribution, making the final destination deviate from the target distribution. Whereas, our method (green path) removes the inconsistent tangent components in unconditional scores and eventually reaches the target distribution." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 423, + 72, + 552, + 171 + ], + "blocks": [ + { + "bbox": [ + 423, + 72, + 552, + 171 + ], + "lines": [ + { + "bbox": [ + 423, + 72, + 552, + 171 + ], + "spans": [ + { + "bbox": [ + 423, + 72, + 552, + 171 + ], + "type": "image", + "image_path": "7721ebb674467b5145386389dceb2a9fdae583bbf25c194c6a0157d1a31c6571.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 172, + 521, + 180 + ], + "lines": [ + { + "bbox": [ + 459, + 172, + 521, + 180 + ], + "spans": [ + { + "bbox": [ + 459, + 172, + 521, + 180 + ], + "type": "text", + "content": "Trajectory of sample B" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 281, + 553, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 352 + ], + "type": "text", + "content": "strategies lie closer to the target compared to those generated using only the conditional score or CFG. CFG, while potentially bringing samples closer to the target than merely using conditional scores, may face challenges due to the misalignment of tangent components between unconditional score and conditional scores." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 354, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 437 + ], + "type": "text", + "content": "In contrast, our guidance score can reduce the tangent component of the unconditional score at each step. This helps samples converge more effectively towards the target data, which suggests that the tangent components of the unconditional score might hinder alignment with the target data manifold under the given condition, and our method helps in mitigating this misalignment." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 439, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 554, + 617 + ], + "type": "text", + "content": "We further validate this hypothesis by examining the trajectories of generated samples. Fig. 5 visualizes the trajectories induced by our score " + }, + { + "bbox": [ + 313, + 439, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{z}_t}\\log \\hat{p}_t(\\boldsymbol{z}_t|\\boldsymbol{y})" + }, + { + "bbox": [ + 313, + 439, + 554, + 617 + ], + "type": "text", + "content": " compared to the original CFG score " + }, + { + "bbox": [ + 313, + 439, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\nabla_{\\boldsymbol{z}_t}\\log \\tilde{p}_{\\theta}(\\boldsymbol{z}_t|\\boldsymbol{y})" + }, + { + "bbox": [ + 313, + 439, + 554, + 617 + ], + "type": "text", + "content": ". As shown, in the orange CFG trajectory, the direction of unconditional scores changes frequently. This results in difficulties for the blue conditional score to maintain an orthogonal direction relative to the target manifold near the target distribution. In contrast, our method consistently adjusts the score to predict in a direction closer to orthogonal with respect to the target manifold, particularly as the samples converge toward the target data. Our method removes the tangential component of the unconditional score with respect to the manifold of the conditional score. This results in a direction that leans either to the right or to the left." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 619, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 553, + 654 + ], + "type": "text", + "content": "Additionally, the similar results between (c) and (d) in Fig. 4 suggest that computing SVD for only a single sample is sufficient to yield nearly the same result." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 669, + 394, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 394, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 394, + 682 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "In this section, we demonstrate that our method is applicable to high-dimensional diffusion models. We employ rep" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2624" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 70, + 248, + 135 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 248, + 135 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 248, + 135 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 248, + 135 + ], + "type": "table", + "html": "
FID ↓CLIPScore ↑
SD v1.5original13.260.31
+ ours13.120.31
SDXLoriginal13.360.32
+ ours12.650.32
SD v3original16.660.32
+ ours13.740.32
", + "image_path": "ccba861f583ce3a189899f29ff2ee16bf07bfa88c176fde0aa7a76496423a511.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 192, + 294, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 144, + 295, + 189 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 189 + ], + "type": "text", + "content": "Table 1. Zero-shot FID and CLIPScore measured on MSCOCO 30k. Our method consistently improves FID across all models—Stable Diffusion v1.5, SDXL, and SD v3—while maintaining a nearly identical CLIPScore." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 192, + 294, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 192, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 192, + 294, + 228 + ], + "type": "table", + "html": "
FID ↓sFID ↓Precision ↑Recall ↑IS ↑
DiT32.6717.920.900.13271.1
DiT+ours29.513.270.900.19270.0
", + "image_path": "6631c95025030ada415861ff4e4a4ef3904f72a2eecda090b8caa6751e291427.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 236, + 294, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 294, + 269 + ], + "type": "text", + "content": "Table 2. Evaluation metrics measured on ImageNet 50k using DiT. Our method achieves better performance in FID, sFID, Precision, and Recall while showing a slight decrease in Inception Score." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 273, + 295, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 295, + 333 + ], + "type": "text", + "content": "representative diffusion models such as Stable Diffusion v1.5 [27] and SDXL [23], and showed that it functions identically on SD v3 [7], which is based on Rectified Flow. Additionally, we conducted experiments on DiT [21], which is trained on ImageNet [5]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 355, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 355, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 355, + 295, + 474 + ], + "type": "text", + "content": "Experimental details For the text-to-image models, we used zero-shot FID [10] and CLIPScore [25] on the MSCOCO 2014 validation set [17] consisting of 30,000 images under the commonly used text-to-image evaluation protocols. [7, 23, 27] For DiT, we evaluated using 50,000 images under the same settings as ADM [6]. All models used the official pretrained weights, and sampling was performed using the same latent codes. We used the best CFG scales as the default value of each repository. Our method does not increase the inference time of all baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 497, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 295, + 652 + ], + "type": "text", + "content": "Quantitative evaluation Tab. 1 presents the FID and CLIP Scores for SD1.5, SDXL, and SD3. Our method achieved better FID scores while maintaining the same CLIP Scores across all three models. Notably, the decrease in FID is larger for SDXL compared to SD1.5, and even larger for SD3 compared to SDXL. We speculate that this is because SD3, known as a better model publicly, has a relatively clearer manifold. Furthermore, the results on SD3 demonstrate that our method is applicable not only to diffusion models but also to all CFG-based score functions, including those based on Rectified Flow. Fig. 6 also shows FID-CLIP curves on SDXL, demonstrating that FID improves even as the CFG scale changes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Tab. 2 shows the results on the DiT model. Except for a slight decrease in Inception Score, our method exhibits relatively superior performance in FID, sFID, and Recall. This indicates that our method can be equally applied to both text-to-image generation and class-conditioned generation." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 351, + 71, + 517, + 149 + ], + "blocks": [ + { + "bbox": [ + 351, + 71, + 517, + 149 + ], + "lines": [ + { + "bbox": [ + 351, + 71, + 517, + 149 + ], + "spans": [ + { + "bbox": [ + 351, + 71, + 517, + 149 + ], + "type": "image", + "image_path": "952ac4373792e71419fd0a50bfed71ff1329eff7e94bc493d97eca39aad03288.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 150, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 321, + 150, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 321, + 150, + 545, + 161 + ], + "type": "text", + "content": "Figure 6. FID-CLIP curves on SDXL with 50 sampling steps." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 164, + 553, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 553, + 272 + ], + "type": "text", + "content": "Qualitative evaluation Our method drops the tangential component from the unconditional score while retaining the normal component. This reduces misalignment with the conditional score, thereby improving image quality as shown in Fig. 7. Specifically, the changes introduced by our approach transform \"strange\" objects or scenes into more \"plausible\" images. This indirectly demonstrates that the misalignment of the unconditional score in the conventional CFG was causing the \"strange\" aspects in the final outputs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 272, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 553, + 319 + ], + "type": "text", + "content": "For example, our method converts physically impossible or unusual combinations of objects (SD3), uncommon appearances or characteristics (SDXL), and ambiguous shapes or forms (SD1.5) into \"normal\" results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 321, + 553, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 553, + 405 + ], + "type": "text", + "content": "Fig. 8 presents the results obtained from DiT. We observed that our method causes relatively more changes in the images generated by DiT. We speculate that this is because DiT is trained on ImageNet dataset with class labels. The results qualitatively show that when using our method, DiT generates images that are more detailed, have better structure, and appear more natural." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 406, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 553, + 464 + ], + "type": "text", + "content": "Notably, in both text-to-image and class-conditioned image generation, we observed a reduction in the overexposure bias problem. We attribute this improvement to the mitigation of misalignment between the unconditional score and the conditional score." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 485, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 640 + ], + "type": "text", + "content": "What happened to the unconditional score? In this paragraph, we qualitatively demonstrate that the misalignment with the conditional score is reduced when we drop the tangential component from the unconditional score and retain the normal component. We compared the results sampled using CFG with those sampled using the null condition (i.e., unconditional) when generating images from the same random noise (i.e., latent variables) in SDXL. In our method, we used the text condition to compute " + }, + { + "bbox": [ + 313, + 485, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 313, + 485, + 553, + 640 + ], + "type": "text", + "content": " but used only " + }, + { + "bbox": [ + 313, + 485, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 313, + 485, + 553, + 640 + ], + "type": "text", + "content": " for denoising; that is, we set " + }, + { + "bbox": [ + 313, + 485, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\omega = 0" + }, + { + "bbox": [ + 313, + 485, + 553, + 640 + ], + "type": "text", + "content": ". Although this approach does not perfectly explain our method, we can indirectly infer its role by observing how the modified null condition changes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "Fig. 9 shows that images sampled using the original null condition generate different objects such as trees, snowy mountain landscapes, and women. In contrast, images generated using our modified null condition " + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": " show that the tree part takes the form of a feather, the snowy mountain landscape changes into a woman, and the woman transforms" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2625" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 85, + 222, + 330 + ], + "blocks": [ + { + "bbox": [ + 91, + 74, + 110, + 83 + ], + "lines": [ + { + "bbox": [ + 91, + 74, + 110, + 83 + ], + "spans": [ + { + "bbox": [ + 91, + 74, + 110, + 83 + ], + "type": "text", + "content": "SD3" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 74, + 206, + 83 + ], + "lines": [ + { + "bbox": [ + 159, + 74, + 206, + 83 + ], + "spans": [ + { + "bbox": [ + 159, + 74, + 206, + 83 + ], + "type": "text", + "content": "SD3 + ours" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 85, + 222, + 330 + ], + "lines": [ + { + "bbox": [ + 60, + 85, + 222, + 330 + ], + "spans": [ + { + "bbox": [ + 60, + 85, + 222, + 330 + ], + "type": "image", + "image_path": "e5b21f654b98856ed13f4c4741391164a1bb465d129ab605085b553a340d29ed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 341, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 55, + 341, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 555, + 363 + ], + "type": "text", + "content": "Figure 7. Qualitative evaluation of text-to-image models. Our method prevents overexposure, enhancing the shapes and details of objects." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 224, + 85, + 386, + 330 + ], + "blocks": [ + { + "bbox": [ + 252, + 74, + 279, + 83 + ], + "lines": [ + { + "bbox": [ + 252, + 74, + 279, + 83 + ], + "spans": [ + { + "bbox": [ + 252, + 74, + 279, + 83 + ], + "type": "text", + "content": "SDXL" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 74, + 373, + 83 + ], + "lines": [ + { + "bbox": [ + 319, + 74, + 373, + 83 + ], + "spans": [ + { + "bbox": [ + 319, + 74, + 373, + 83 + ], + "type": "text", + "content": "SDXL + ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 224, + 85, + 386, + 330 + ], + "lines": [ + { + "bbox": [ + 224, + 85, + 386, + 330 + ], + "spans": [ + { + "bbox": [ + 224, + 85, + 386, + 330 + ], + "type": "image", + "image_path": "273e6e2ef619d99e4436f3ec5c960d1c30da51ece63a5cee1b960d4cb90916bc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 389, + 85, + 551, + 330 + ], + "blocks": [ + { + "bbox": [ + 413, + 74, + 440, + 83 + ], + "lines": [ + { + "bbox": [ + 413, + 74, + 440, + 83 + ], + "spans": [ + { + "bbox": [ + 413, + 74, + 440, + 83 + ], + "type": "text", + "content": "SD1.5" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 480, + 74, + 534, + 83 + ], + "lines": [ + { + "bbox": [ + 480, + 74, + 534, + 83 + ], + "spans": [ + { + "bbox": [ + 480, + 74, + 534, + 83 + ], + "type": "text", + "content": "SD1.5 + ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 85, + 551, + 330 + ], + "lines": [ + { + "bbox": [ + 389, + 85, + 551, + 330 + ], + "spans": [ + { + "bbox": [ + 389, + 85, + 551, + 330 + ], + "type": "image", + "image_path": "f4f3e9c0787e8993a4d886f81698aecd5a281a054ae02c87683cee05ec36ba7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 376, + 175, + 549 + ], + "blocks": [ + { + "bbox": [ + 110, + 367, + 124, + 374 + ], + "lines": [ + { + "bbox": [ + 110, + 367, + 124, + 374 + ], + "spans": [ + { + "bbox": [ + 110, + 367, + 124, + 374 + ], + "type": "text", + "content": "DiT" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 376, + 175, + 549 + ], + "lines": [ + { + "bbox": [ + 59, + 376, + 175, + 549 + ], + "spans": [ + { + "bbox": [ + 59, + 376, + 175, + 549 + ], + "type": "image", + "image_path": "90389e3e52e478f073e557f056826834da79adebf095074e7a2a441af34ae4fd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 563, + 295, + 597 + ], + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 597 + ], + "type": "text", + "content": "Figure 8. Qualitative evaluation of DiT Our method mitigates overexposure and enhances object shapes and details in DiT models trained on ImageNet." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 177, + 376, + 293, + 549 + ], + "blocks": [ + { + "bbox": [ + 217, + 367, + 254, + 375 + ], + "lines": [ + { + "bbox": [ + 217, + 367, + 254, + 375 + ], + "spans": [ + { + "bbox": [ + 217, + 367, + 254, + 375 + ], + "type": "text", + "content": "DiT " + }, + { + "bbox": [ + 217, + 367, + 254, + 375 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 217, + 367, + 254, + 375 + ], + "type": "text", + "content": " ours" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 376, + 293, + 549 + ], + "lines": [ + { + "bbox": [ + 177, + 376, + 293, + 549 + ], + "spans": [ + { + "bbox": [ + 177, + 376, + 293, + 549 + ], + "type": "image", + "image_path": "b54d34de12e13b69d130a5ca60a9efcf1c33cef56c68e8ae8359e10a430f68c1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 618, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 712 + ], + "type": "text", + "content": "into a shape resembling a glove. These changes align with the objects we aim to generate: a feather, a woman, and a baseball glove. We observe that these changes due to the null condition help eliminate unwanted structures or artifacts in the generated images. In other words, we demonstrate that the misalignment of the null condition is reduced, and we claim that this improvement aids in image generation." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 315, + 363, + 552, + 387 + ], + "blocks": [ + { + "bbox": [ + 315, + 363, + 552, + 387 + ], + "lines": [ + { + "bbox": [ + 315, + 363, + 552, + 387 + ], + "spans": [ + { + "bbox": [ + 315, + 363, + 552, + 387 + ], + "type": "table", + "html": "
SAGSAG+TCFGPAGPAG+TCFGCFGPPCFGPP+TCFG
FID13.5311.4814.4511.8713.9713.44
CLIP Score0.310.300.310.310.320.32
", + "image_path": "a43183187866b9cb58cc54fe7e11d5e55d97cd3390bf524e310d4f84f68f81ac.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 388, + 554, + 431 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 431 + ], + "type": "text", + "content": "Table 3. Quantitative comparison with existing baselines. The evaluation was conducted on " + }, + { + "bbox": [ + 313, + 388, + 554, + 431 + ], + "type": "inline_equation", + "content": "{30}\\mathrm{k}" + }, + { + "bbox": [ + 313, + 388, + 554, + 431 + ], + "type": "text", + "content": " images from the MS-COCO dataset using the official code; SD v1.4 for SAG,SD v1.5 for PAG and SDXL for CFG++." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 434, + 399, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 399, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 399, + 445 + ], + "type": "text", + "content": "7. Related work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 455, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 598 + ], + "type": "text", + "content": "Calssifier-free guidance Experimental methods to enhance the performance of Classifier-Free Guidance (CFG) have been studied. SAG [13] proposed a method to improve CFG by using intermediate self-attention maps. PAG [2] suggested computing CFG by transforming self-attention maps into identity matrices. ICG [29] enhanced CFG by utilizing random text embeddings. Recently, CFG++ [4] demonstrated better performance by modifying the CFG computation method. Our proposed approach modifies the unconditional score based on the conditional score and can be used alongside these existing works; please refer to Tab. 3 and the appendix for more results." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Manifold hypothesis and diffusion There are also several studies that have utilized the manifold hypothesis properties of the score function estimated by diffusion models to address various inherent challenges associated with diffusion processes. One approach introduces the manifold memorization hypothesis to understand model memorization through the relationship between data and model manifold dimensionalities [28]. Another extends memorization" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2626" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 555, + 240 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 240 + ], + "type": "image", + "image_path": "256d319284fa09a96fd3418b3acd5ca0fa7702c1be6860fc8d62d9a903fad4d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 327 + ], + "type": "text", + "content": "Figure 9. TCFG reduces misalignments between unconditional and conditional generation. Starting from the same random noise " + }, + { + "bbox": [ + 55, + 251, + 555, + 327 + ], + "type": "inline_equation", + "content": "z_{1}" + }, + { + "bbox": [ + 55, + 251, + 555, + 327 + ], + "type": "text", + "content": ", when SDXL samples images with only the unconditional score, it produces random images such as trees, snowy mountain landscapes, and women. In contrast, our modified unconditional score, projected on dominant (conditional), generates images that somewhat match the desired text prompts. This is because our method reduces misalignment with the conditional score by dropping the tangential components of the unconditional score. Once the misalignment decreases, the quality of the final images (unconditional + conditional score) improves: The base of the feather has a more natural structure, the human arm appears more natural, and the extra string on the left side of the baseball glove is removed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 95, + 331, + 256, + 499 + ], + "blocks": [ + { + "bbox": [ + 95, + 331, + 256, + 499 + ], + "lines": [ + { + "bbox": [ + 95, + 331, + 256, + 499 + ], + "spans": [ + { + "bbox": [ + 95, + 331, + 256, + 499 + ], + "type": "image", + "image_path": "82d53a3b354e00e8625ecb192fcca134ccaf1375e60e7c705329562695dbb71f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 509, + 295, + 531 + ], + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 531 + ], + "type": "text", + "content": "Figure 10. Limitations Our method occasionally struggles to fix severely wrong regions in the baseline samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": "theory to diffusion models [1], showing that high-variance subspaces are selectively lost due to memorization effects. Separately, different researchers proposed an approach for detecting synthetic images generated by diffusion models, achieving high accuracy across diverse datasets [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 609, + 204, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 204, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 204, + 620 + ], + "type": "text", + "content": "8. Discussion and conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Our work experimentally analyzes the issues arising in the standard CFG method, where the tangential component of the unconditional score does not align well with that of the conditional score. By using SVD to drop the tangential component in the unconditional score, we effectively improve text-to-image generation quality. Additionally, our CFG method is easily applicable, has low computational" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 330, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 401 + ], + "type": "text", + "content": "cost, and enhances image quality. We leverage the ability of the diffusion model's score function to encode the intrinsic dimension of the target data, demonstrating the misalignment between the conditional and unconditional scores to improve sampling quality. This is the first attempt to utilize this misalignment to enhance sampling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "text", + "content": "Despite these advantages, several unresolved issues remain. First, it is uncertain whether the misalignment of tangential component and the alignment of normal component between the predicted unconditional score " + }, + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t)" + }, + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "text", + "content": " and the conditional score " + }, + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_t,y)" + }, + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "text", + "content": " in the CFG setting, at a given timestep " + }, + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "text", + "content": ", would similarly apply to the features derived from a separately trained classifier and the null condition score in the classifier guidance setting. Second, while our task leverages the capability of diffusion models to estimate intrinsic dimensions for enhancing conditional sampling methods, we present only experimental observations regarding the existence of an intermediate manifold for " + }, + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 313, + 414, + 556, + 628 + ], + "type": "text", + "content": ", without theoretical proof. Further exploration and rigorous analysis of these aspects are left as future work. Third, additional investigation is needed to adapt our approach effectively in the context of diffusion distillation using CFG scale as an input [14, 15], which we also identify as a promising direction for future research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 712 + ], + "type": "text", + "content": "Finally, although our method successfully demonstrated on-manifold image generation, we observed that when the original image exhibits significant abnormalities, substantial changes may occasionally cause the structure to break down. Fig. 10 illustrates such examples. Nevertheless, it is evident that our method transforms \"strange\" images into more \"normal\" ones." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2627" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Beatrice Achilli, Enrico Ventura, Gianluigi Silvestri, Bao Pham, Gabriel Raya, Dmitry Krotov, Carlo Lucibello, and Luca Ambrogioni. Losing dimensions: Geometric memorization in generative diffusion. arXiv preprint arXiv:2410.08727, 2024. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 7, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 247 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8):1798-1828, 2013. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "type": "text", + "content": "[4] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 7, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 295, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 295, + 339 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 373 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 441 + ], + "type": "text", + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 295, + 508 + ], + "type": "text", + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 542 + ], + "type": "text", + "content": "[9] Charles Fefferman, Sanjoy Mitter, and Hariharan Narayanan. Testing the manifold hypothesis. Journal of the American Mathematical Society, 29(4):983-1049, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 597 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 601, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 623 + ], + "type": "text", + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[13] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 7, 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[14] Yi-Ting Hsiao, Siavash Khodadadeh, Kevin Duarte, Wei-An Lin, Hui Qu, Mingi Kwon, and Ratheesh Kalarot. Plug-and-play diffusion distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13743-13752, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "text", + "content": "[15] Black Forest Labs. FLUX. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-15. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 186 + ], + "type": "text", + "content": "[16] John M Lee. Introduction to Riemannian manifolds. Springer, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 187, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 187, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 187, + 554, + 252 + ], + "type": "text", + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 255, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 554, + 308 + ], + "type": "text", + "content": "[18] Peter Lorenz, Ricard L Durall, and Janis Keuper. Detecting images generated by deep diffusion models using their local intrinsic dimensionality. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 448-459, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 310, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 554, + 364 + ], + "type": "text", + "content": "[19] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 366, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 409 + ], + "type": "text", + "content": "[20] Kazusato Oko, Shunta Akiyama, and Taiji Suzuki. Diffusion models are minimax optimal distribution estimators. In International Conference on Machine Learning, pages 26517-26582. PMLR, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "text", + "content": "[22] Jakiw Pidstrigach. Score-based generative models detect manifolds. Advances in Neural Information Processing Systems, 35:35852-35865, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 491, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 554, + 544 + ], + "type": "text", + "content": "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 547, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 547, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 547, + 554, + 590 + ], + "type": "text", + "content": "[24] Phillip Pope, Chen Zhu, Ahmed Abdelkader, Micah Goldblum, and Tom Goldstein. The intrinsic dimension of images and its impact on learning. arXiv preprint arXiv:2104.08894, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 592, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 554, + 657 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2628" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 418 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[27] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "text", + "content": "[28] Brendan Leigh Ross, Hamidreza Kamkari, Tongzi Wu, Rasa Hosseinzadeh, Zhaoyan Liu, George Stein, Jesse C Cresswell, and Gabriel Loaiza-Ganem. A geometric framework for understanding memorization in generative models. arXiv preprint arXiv:2411.00113, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "text", + "content": "[29] Seyedmorteza Sadat, Manuel Kansy, Otmar Hilliges, and Romann M Weber. No training, no problem: Rethinking classifier-free guidance for diffusion models. arXiv preprint arXiv:2407.02687, 2024. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 295, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 295, + 295 + ], + "type": "text", + "content": "[30] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 328 + ], + "type": "text", + "content": "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "text", + "content": "[32] Jan Pawel Stanczuk, Georgios Batzolis, Teo Deveney, and Carola-Bibiane Schonlieb. Diffusion models encode the intrinsic dimension of data manifolds. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 418 + ], + "type": "text", + "content": "[33] Zhenyue Zhang and Hongyuan Zha. Principal manifolds and nonlinear dimensionality reduction via tangent space alignment. SIAM journal on scientific computing, 26(1):313-338, 2004. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2629" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_content_list.json b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c7dd839bd6dc9ad454bcfde9c552ad3691e23ec --- /dev/null +++ b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_content_list.json @@ -0,0 +1,1661 @@ +[ + { + "type": "text", + "text": "TFCustom: Customized Image Generation with Time-Aware Frequency Feature Guidance", + "text_level": 1, + "bbox": [ + 135, + 128, + 861, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mushui Liu $^{1*}$", + "bbox": [ + 169, + 202, + 281, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dong She $^{2*}$", + "bbox": [ + 303, + 203, + 398, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingxuan Pang1", + "bbox": [ + 419, + 203, + 544, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qihan Huang", + "bbox": [ + 566, + 203, + 681, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wanggui $\\mathrm{He}^3$", + "bbox": [ + 316, + 220, + 423, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanlei Hou", + "bbox": [ + 446, + 220, + 553, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siming $\\mathrm{Fu}^{1*^{\\dagger}}$", + "bbox": [ + 575, + 220, + 681, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Zhejiang University", + "bbox": [ + 204, + 241, + 370, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Science and Technology of China", + "bbox": [ + 411, + 241, + 794, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Alibaba Group", + "bbox": [ + 315, + 260, + 442, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Middle East Centre, RIPED", + "bbox": [ + 482, + 260, + 715, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e91af3c05738d38aed746639800e5c8f4d99887f7f080396d5bc09a0abf66b9.jpg", + "image_caption": [ + "Single-Subject Driven Generation" + ], + "image_footnote": [], + "bbox": [ + 99, + 311, + 903, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8cfcd605e8f120ce513bf1026caf78b0d8718f0a06fd6a6aa95f82fe86a5cfc2.jpg", + "image_caption": [ + "Multiple-Subject Driven Generation", + "Figure 1. Our proposed TFCustom demonstrates capabilities in single-subject and multi-subject driven generation tasks." + ], + "image_footnote": [], + "bbox": [ + 124, + 522, + 893, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 729, + 326, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Subject-driven image personalization has seen notable advancements, especially with the ReferenceNet paradigm, which excels in integrating reference image features for creative and commercial applications. However, current ReferenceNet implementations mainly function as latent-level feature extractors, limiting their potential. This restricts", + "bbox": [ + 88, + 763, + 485, + 856 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the delivery of suitable features to the denoising backbone across timesteps, resulting in suboptimal image consistency. In this paper, we revisit reference feature extraction and propose TFCustom, a framework that focuses on reference image features at different temporal and frequency levels. We introduce synchronized ReferenceNet to extract reference features while optimizing noise injection and denoising. We also propose a time-aware frequency refinement module that uses high- and low-frequency filters with time embeddings to adaptively select reference feature injection. Additionally, we introduce a reward-based loss to improve", + "bbox": [ + 511, + 732, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 325, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions.", + "bbox": [ + 114, + 875, + 230, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 114, + 887, + 240, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2714", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the similarity between reference objects and generated images. Experimental results show that TFCustom outperforms existing methods in single-object and multi-object reference generation, with significant improvements in textual details.", + "bbox": [ + 88, + 90, + 480, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 195, + 222, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the development of deep learning [14, 19, 24, 46, 52] and text-to-image models [8, 15, 23, 32, 34], customization generation which aims to synthesize text-driven images for given subjects, has recently revolutionized content creation. It has seen a shift from test-time fine-tuning methods (e.g., DreamBooth [35], Custom Diffusion [18]) to fine-tuning-free techniques (e.g., IP-Adapter [53], Subject-Diffusion [26], MS-Diffusion [48]). The latter approaches remove the need for fine-tuning during inference, significantly reducing computational costs while maintaining flexibility, and making them more scalable and practical for real-world use cases.", + "bbox": [ + 91, + 220, + 483, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These fine-tuning-free methods generally fall into two categories: image encoder-based [4, 5, 58] feature injection and reference network (ReferenceNet-based) [6, 12, 51] feature injection. Image encoder methods, such as IPAdapter [53] and ELITE [49], utilize pre-trained encoders like CLIP [33] to extract high-level features, which are then injected into the generation process via mechanisms such as decoupled cross-attention. While these methods are computationally efficient and straightforward, they tend to rely on high-level, abstract features, which may not capture the fine-grained details necessary for precise control over reference object characteristics. The increasing demand for greater consistency in reference images has brought significant attention to the concept of ReferenceNet [16] in the field. ReferenceNet-based methods [17, 47, 50, 57] introduce an additional diffusion network within the traditional Diffusion Net denoising process. In this setup, reference images are encoded by VAE and then used as inputs for ReferenceNet. This novel architecture enables the correlation of features across layers in both diffusion networks, which helps in making better use of the reference images.", + "bbox": [ + 91, + 402, + 483, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although ReferenceNet-based methodologies have made significant advancements, several challenges remain. Currently, ReferenceNet predominantly operates as a latent-level feature extractor, thereby not fully realizing its potential. While ReferenceNet maintains a consistent intensity of feature conditioning over time, the denoising backbone requires the injection of varying intensities and granularities of reference information at different timesteps. Specifically, during the earlier stages of denoising, the backbone tends to emphasize contour information, whereas, in the later stages, it focuses more on the refinement of details [45]. This limitation hinders the ability", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to provide the denoising backbone with the appropriate features necessary for optimal performance across varying timesteps, ultimately leading to subpar image consistency.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce TFCustom, a novel finetuning-free personalized generation framework that incorporates temporal awareness for reference feature injection. Unlike previous approaches, TFCustom considers the alignment of reference image features with generated image features across different time steps of the diffusion process. Specifically, we first apply noise to the reference image at multiple time steps, with ReferenceNet extracting features at various scales and granularities. The training of ReferenceNet follows a diffusion loss function, ensuring proper alignment with the denoising process. Additionally, we introduce a time-aware frequency module that enables the controlled injection of high- and low-frequency reference features. This module allows low-frequency features to define the primary subject at early time steps, while high-frequency features are progressively injected to refine finer image details. To further improve the model's ability to handle multi-object reference generation, we propose a novel reward-based loss function that enhances the alignment between the reference and generated objects. This loss function computes object ID similarity during the denoising process, minimizing conflicts between reference objects and improving the overall coherence of multi-object scenes. The experimental results demonstrate our method consistently outperforms the state-of-the-art approaches on all the benchmarks.", + "bbox": [ + 511, + 137, + 906, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions are as follows:", + "bbox": [ + 532, + 529, + 805, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel customized image generation framework, which includes synchronized noise injection and denoising within ReferenceNet, along with a time-aware frequency feature refinement module for extracting fine-grained image details.", + "- We propose a reward-based loss function that alleviates multi-object conflicts and enhances object identity consistency in generated images.", + "- Our model achieves state-of-the-art performance in single-object reference generation on DreamBooth and MS-Bench datasets. Additionally, we demonstrate superior qualitative results in terms of texture and textual detail generation compared to existing methods." + ], + "bbox": [ + 511, + 544, + 903, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 753, + 653, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Subject Driven Text-to-Image Generation. Despite significant advancements in image generation [1, 8, 15, 23, 34], the field still falls short of meeting the requirements for personalized generation [38-40, 42, 48, 54]. This gap has consequently drawn significant attention to subject-driven text-to-image generation methods [4, 26, 43]. Textual Inversion [9] and DreamBooth [35] use a textual identifier to bind the visual concept through fine-tuning model textual", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2715", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "embedding or UNet model. CustomDiffusion [18] further extends multi-subject generation by multiple identifiers. IP-Adapter [53] injects the visual concept by decoupling the cross-attention between text and images in a zero-shot manner. JeDi [55] learns the joint distribution of multiple text-image pairs that share a common subject to facilitate personalized generation. Subject-Diffusion [26] trains multiple adapters inside the model with an attention control mechanism for multiple-subject personalization. SSR-Encoder [58] and MS-Diffusion [48] develop the query architecture to encode the subject images via mask or layout control for multiple-subject personalization. This paper explores the zero-shot multiple-objects driven generation via synchronized ReferenceNet optimization.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ReferenceNet Based Personalized Generation. ReferenceNet [16, 41, 57] has shown remarkable capabilities in maintaining image consistency for personalized generation tasks. AnimateAnyone [16] first utilizes ReferenceNet for personalized human animation, employing a pre-trained generative model to extract reference features. FlashFace [57] extended this approach by combining ReferenceNet with ControlNet [56], enabling identity-preserving portrait generation that accommodates multiple reference images. Part2Whole [17] uses ReferenceNet to handle multi-part human body generation, while OOTD [50] and Stable-Garment [47] integrate it into virtual try-on tasks, aligning clothing features precisely with the target human body. These methods highlight ReferenceNet's versatility in image and video generation. To further improve performance, we introduce a synchronized ReferenceNet that refines reference features at each time step, enabling more precise feature injection and enhancing generation accuracy.", + "bbox": [ + 91, + 303, + 482, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frequency Information in Image Generation. Frequency information plays a crucial role in image restoration tasks [7, 22, 27, 36]. Recent advancements in text-to-image generation have leveraged frequency-based conditions, like canny and HED images, to control generation, as seen in ControlNet [56] and T2I-Adapter [28]. FCDiffusion [11] and FBSdiff [10] employ DCT-transformed image features to enhance texture detail in image-to-image generation. In this work, our method adaptively injects reference image features via learnable high-pass and low-pass filters in a time-aware fashion, enriching each denoising step with frequency-specific information for more refined generation.", + "bbox": [ + 89, + 575, + 482, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 768, + 181, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed TFCustom addresses these issues through three components: Sec. 3.2 Synchronized Reference Guidance, Sec. 3.3 Time-Aware Frequency Feature Refinement, and Sec. 3.4 Reward Model Optimization, as illustrated in Fig. 2. First, the reference image is incrementally noised over $T$ steps before feature extraction via the ReferenceNet, aligning it with the denoising network's noisy", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "latent features. Next, a time-aware frequency module injects frequency-specific information from the reference image into the noisy latent representation. Finally, a reward model optimizes this process, enhancing the quality and consistency of the generated reference-based images.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 176, + 656, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. Denoising Diffusion Probabilistic Models (DDPM) are a class of generative models that leverage a diffusion process to generate data. By progressively adding Gaussian noise to the data over a specified number of time steps, a sequence of increasingly noisy samples is produced. The forward process can be described as follows:", + "bbox": [ + 511, + 199, + 903, + 290 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {T} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {T}; \\sqrt {\\alpha_ {T}} x _ {0}, (1 - \\alpha_ {T}) \\mathbf {I}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 301, + 903, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} = \\prod_{s=1}^{t}(1 - \\beta_{s})$ . Here, $\\beta_{t}$ represents a variance schedule that regulates the noise level at each time step $t$ , and $N$ denotes a Gaussian distribution. In the reverse diffusion process, the model learns to denoise the noisy samples step by step, effectively generating new data from pure noise:", + "bbox": [ + 511, + 329, + 903, + 417 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\sigma_ {\\theta} ^ {2} (t) \\mathbf {I}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 431, + 903, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}$ and $\\sigma_{\\theta}$ are learned parameters that depend on the current noisy sample $x_{t}$ and the time step $t$ . The optimization goal is the variational lower bound on the log-likelihood of the data, which can be simplified to:", + "bbox": [ + 511, + 460, + 903, + 521 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL = \\mathbb {E} _ {t, x _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 531, + 903, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Synchronized ReferenceNet Guidance", + "text_level": 1, + "bbox": [ + 511, + 560, + 841, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The synchronized ReferenceNet is initially transferred from the denoising network. During training, at each time step $t$ , we apply a noise process to the reference image $x_{\\mathrm{ref}}$ , adding noise for $t$ steps, resulting in the noise reference $x_{\\mathrm{ref}}^{t}$ as formulated in Eq. (1). This step ensures that the reference features are temporally aligned with the noisy latent features processed by the denoising network. Once the noised reference $x_{\\mathrm{ref}}^{t}$ is generated, it passes through the synchronized ReferenceNet, which extracts the relevant features $\\mathbf{F}_{\\mathrm{ref}}$ and estimates the corresponding noise pattern. The extracted reference features and predicted noise are then optimized according to the objective $\\mathcal{L}_{\\mathrm{diff}}^{\\mathrm{ref}}$ defined in Eq. (3). This objective encourages the synchronized ReferenceNet to provide refined guidance that is consistent with the noisy latent representations, enabling the denoising network to more effectively incorporate reference information for improved reconstruction quality.", + "bbox": [ + 511, + 583, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given that the object information hierarchically learned by the denoising network varies significantly across different timesteps, the early timesteps tend to focus on learning macro-level information such as contours, while the", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2716", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aba4401df3cc440eee8b9b53f84da2a06be24a94f4a47ae75ac008714405975c.jpg", + "image_caption": [ + "(a) Overall Framework" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 517, + 294 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/daf168b010f9eff67fd0d14eecfb9a12313c2dfce082135fe173d71c4b50b574.jpg", + "image_caption": [ + "(b) Time-Aware Frequency Feature Refinement", + "Figure 2. (a) Overall Framework. Our TFCustom includes three key components: Synchronized Reference Net, Time-Aware Frequency Feature Refinement (TA-FFR), and Reward Model Optimization for subject-driven generation. (b) TA-FFR Module. This module extracts different frequency features in a time-aware manner, adaptively injecting reference features for enhanced generation quality." + ], + "image_footnote": [], + "bbox": [ + 522, + 98, + 901, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "later timesteps progressively refine texture details. However, the synchronized timestep information obtained from $\\mathcal{L}_{\\mathrm{diff}}^{\\mathrm{ref}}$ encompasses the entire reference object's information at timestep $t$ , which can be overly complex for the denoising network. To address this issue, we have implemented a Time-Aware Frequency Feature Refinement module (which will be described detailed in Sec. 3.3) to refine the reference information, thereby enhancing the learning process of the denoising network, as outlined in Eq. 7.", + "bbox": [ + 88, + 401, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the enhanced reference features $\\mathbf{F}_{\\mathrm{enh}}$ , we inject them using a Shared Self-Attention mechanism [16, 17]. Specifically, in each self-attention layer, given the features of $N$ reference images $\\mathbf{F}_{\\mathrm{enh}}^{1:N}$ and the feature map $\\mathbf{F}^0$ from the denoising net, we concatenate these feature maps side-by-side as input to the self-attention layer, represented as $[\\mathbf{F}_{\\mathrm{enh}}^0 | \\mathbf{F}_{\\mathrm{enh}}^1| \\dots | \\mathbf{F}_{\\mathrm{enh}}^N]$ . This setup enables each location $s$ on $\\mathbf{F}^0$ to attend to all locations on itself and the reference feature maps, computed as follows:", + "bbox": [ + 89, + 537, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\mathbf {F}} _ {s} ^ {0} = \\operatorname {S o f t M a x} \\left(\\frac {\\mathbf {F} _ {s} ^ {0} W _ {q} \\cdot \\left([ \\mathbf {F} _ {\\mathrm {e n h}} ^ {0} , \\mathbf {F} _ {\\mathrm {e n h}} ^ {1} , \\dots , \\mathbf {F} _ {\\mathrm {e n h}} ^ {N} ]\\right) ^ {\\top} W _ {k}}{\\sqrt {d}}\\right) \\\\ \\cdot \\left(\\left[ \\mathbf {F} _ {\\text {e n h}} ^ {0}, \\mathbf {F} _ {\\text {e n h}} ^ {1}, \\dots , \\mathbf {F} _ {\\text {e n h}} ^ {N} \\right]\\right) W _ {v}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 685, + 480, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{q}, W_{k}, W_{v}$ are corresponding projection weight matrices. And $d$ is the dimensionality of the key vectors.", + "bbox": [ + 89, + 760, + 482, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Time-Aware Frequency Feature Refinement", + "text_level": 1, + "bbox": [ + 89, + 800, + 465, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by previous work [45], which illustrates that neural network learns hierarchically—starting with low-frequency features that define shape and structure, followed by high-frequency details for realism, we propose that reference image feature injection should follow a similar coarse-to-fine", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "approach. This would better align with the denoising stages of diffusion.", + "bbox": [ + 511, + 401, + 903, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, we introduce a Time-Aware Frequency Feature Refinement (TA-FFR), as shown in Fig. 2. This module separates reference image features into high-frequency and low-frequency components, each processed independently to capture unique information. The high-frequency path applies a Krisch operator for edge and detail preservation, while the low-frequency path uses a Gaussian operator to retain foundational structures. These operators are initialized as $5 \\times 5$ convolution kernels $\\mathbf{H}_{\\mathrm{conv}}$ is Krisch operator, and $\\mathbf{L}_{\\mathrm{conv}}$ is Gaussian operator. Both operators are scaled by learnable parameters to adjust the frequency extraction dynamically:", + "bbox": [ + 511, + 431, + 906, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {\\text {o u t}} = \\operatorname {C o n v} _ {\\mathbf {H}} \\left(\\mathbf {F} _ {\\text {r e f}}\\right) \\cdot \\mathbf {W} _ {\\mathrm {H}}, \\quad \\mathbf {L} _ {\\text {o u t}} = \\operatorname {C o n v} _ {\\mathbf {L}} \\left(\\mathbf {F} _ {\\text {r e f}}\\right) \\cdot \\mathbf {W} _ {\\mathrm {L}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 627, + 903, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{\\mathrm{ref}}$ denotes reference image features, and $\\mathbf{W}_{\\mathrm{H}}$ and $\\mathbf{W}_{\\mathrm{L}}$ are scaling factors for high and low frequencies.", + "bbox": [ + 511, + 657, + 903, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each frequency component then undergoes Time-Aware Attention (TA-Attention) and Time-Aware Feed-Forward Network (TA-FFN), which incorporate time-step information using a time embedding $\\mathbf{t}_{\\mathrm{emb}}$ via Adaptive Layer Normalization (AdaLN) proposed in DiT [31]:", + "bbox": [ + 511, + 689, + 905, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {F} _ {\\text {h i g h}} = \\mathrm {T A - F F N} (\\mathrm {T A - A t t e n t i o n} (\\mathrm {A d a L N} (\\mathbf {H} _ {\\text {o u t}}, \\mathbf {t} _ {\\text {e m b}}))) \\\\ \\mathbf {F} _ {\\text {l o w}} = \\mathrm {T A - F F N} (\\mathrm {T A - A t t e n t i o n} (\\mathrm {A d a L N} (\\mathbf {L} _ {\\text {o u t}}, \\mathbf {t} _ {\\text {e m b}}))), \\end{array} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 777, + 903, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{\\mathrm{high}}$ and $\\mathbf{F}_{\\mathrm{low}}$ represent high-frequency and low-frequency features. Then the final enhanced feature is obtained by combining these outputs:", + "bbox": [ + 511, + 825, + 903, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\text {e n h}} = \\mathbf {F} _ {\\text {h i g h}} + \\mathbf {F} _ {\\text {l o w}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2717", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6da9dee83c5f758366ecffe6e56115382415310607805ccd8f498c10f8fd1b43.jpg", + "image_caption": [ + "Figure 3. Reward Model Optimization Process. \"G\" and \"D\" indicate the network's forward pass and backward optimization." + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 473, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This design ensures a coarse-to-fine feature injection, enhancing the model's ability to integrate temporal and spatial details, thereby improving the quality of the generated images.", + "bbox": [ + 89, + 267, + 483, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Reward Model Optimization", + "text_level": 1, + "bbox": [ + 89, + 335, + 346, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model can simulate the noise levels of both the reference image and the target image. Based on the following equation, we can perform a one-step prediction of the source image, allowing us to naturally determine the reward model score based on the noise level by:", + "bbox": [ + 89, + 359, + 483, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {0} ^ {\\prime} \\approx \\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} \\left(x _ {t} , x _ {\\text {r e f}} , c _ {t} , t\\right)}{\\sqrt {\\alpha_ {t}}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 446, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_{t}$ refers to the textual condition. This equation's calculation error increases with timestep $t$ . Therefore, during the training phase, we only apply the reward loss for timestep $t < T_0$ :", + "bbox": [ + 89, + 489, + 483, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e w a r d}} = \\operatorname {D i s} \\left(\\mathrm {D I N O} _ {v 2} \\left(x _ {0}\\right), \\mathrm {D I N O} _ {v 2} \\left(x _ {0} ^ {\\prime}\\right)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 561, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Dis is cosine similarity metric and $\\mathrm{DINO}_{v2}$ is the DINO v2 model [2] to compute the similarity between the target domain's denoised output and the features generated by the reference net. It is worth noticing that in our synchronized noise injection scheme, the output of the reference net corresponds to the noise level at timestep $t$ , which provides meaningful information. Thus, we take advantage of this synchronized noise injection feature and introduce a diffusion loss, $L_{\\mathrm{ref}}$ , for the reference net, which supervises the feature extraction capability of the reference branch. The overall loss function can be formulated as:", + "bbox": [ + 89, + 589, + 483, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d i f f}} ^ {\\text {d e n o i s i n g}} + \\lambda_ {1} \\mathcal {L} _ {\\text {d i f f}} ^ {\\text {r e f}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e w a r d}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 763, + 482, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ is hyper-parameters to balance.", + "bbox": [ + 89, + 794, + 413, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 821, + 223, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 847, + 267, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. The training dataset comprises 2.3 million (M) in-house image pairs. We assess the performance on", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/18da077c8cbdcce7e8c8c0a947c787a9767adf19cfeee0c4b3a36c50dc1ad1bc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTypeCLIP-IDINOCLIP-T
BLIP-Diffusion [20]No-FT77.959.430.0
IP-Adapter [53]No-FT81.061.329.2
Emu2 [44]No-FT76.556.327.3
Kosmos-G [29]No-FT82.261.825.0
λ-ECLIPSE [30]No-FT78.361.330.7
SSR-Encoder [58]No-FT82.161.230.8
MS-Diffusion [48]No-FT79.267.132.1
TFCustomNo-FT84.171.433.9
Textual Inversion [9]FT78.056.925.5
DreamBooth [35]FT80.366.830.5
BLIP-Diffusion [20]FT80.567.030.2
λ-ECLIPSE [30]FT79.668.230.4
MS-Diffusion [48]FT80.570.231.3
TFCustomFT85.974.135.1
", + "bbox": [ + 519, + 89, + 903, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison for single-subject models. FT denotes fine-tuned models on DreamBench [35]. The best and second-best results are shown in bold and underline, respectively.", + "bbox": [ + 511, + 320, + 903, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "single-subject tasks using the DreamBench dataset [35] and on multi-subject tasks using the MS-Bench dataset [48]. DreamBench encompasses 30 subjects and 25 prompts, while MS-Bench includes 40 subjects and 1148 combinations, with each combination being associated with up to 6 prompts, resulting in a total of 4488 unique test samples. Furthermore, we collect an additional 20 real-world cases featuring complex textures, characters, and patterns for further qualitative evaluation.", + "bbox": [ + 511, + 371, + 906, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. During training, we employ the AdamW optimizer with a learning rate of 1e-4 and train the model for 100K steps on 32 NVIDIA A100 80GB GPUs, using a batch size of 4 per GPU. The $\\lambda_{1}$ and $\\lambda_{2}$ in Eq. (10) are set to 0.4 and 0.6. To enable classifier-free guidance, we apply a probability of 0.05 to independently drop text and images, and 0.05 to drop both simultaneously. For inference, we adapt the DDIM sampler with 50 steps and set the guidance scale to 7.5.", + "bbox": [ + 511, + 507, + 905, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate performance based on both image and text fidelity metrics. For image fidelity, we compute the cosine similarity between generated images and subject images within the CLIP (denoted as CLIP-I) and DINO feature spaces. Text fidelity is measured by calculating the cosine similarity between generated images and text prompts in the CLIP space (referred to as CLIP-T). In the context of multi-subject personalization, average fidelity scores may not adequately reflect the representation of individual subjects. Therefore, we also employ the multi-subject DINO (M-DINO) metric [48] to assess whether each subject was accurately recreated.", + "bbox": [ + 511, + 643, + 905, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Competitors. We compare our model with BLIP-Diffusion [20], IP-Adapter [53], Kosmos-G [29], Emu2 [44], $\\lambda$ -ECLIPSE [30], SSR-Encoder [58], and MS-Diffusion [48] for single concept personalization, and with MS-Diffusion [48], SSR-Encoder [58] and $\\lambda$ -ECLIPSE [30] for multi-", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2718", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1a3d9c0a701ce79bac4229a27066b907a91fe1c3258c227750f24ea313020ae3.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with existing state-of-the-art customization methods. TFCustom produces higher-quality results with better subject similarity and text controllability compared to existing methods." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 903, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "subject personalization.", + "bbox": [ + 89, + 486, + 250, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Comparison", + "text_level": 1, + "bbox": [ + 89, + 512, + 331, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison of Single-Subject. The results are presented in Tab. 1. In the zero-shot evaluation (without fine-tuning on DreamBench), our proposed TFCustom achieves the highest scores across all metrics, with CLIP-I, DINO, and CLIP-T values of 84.1, 71.4, and 33.9, respectively. These results surpass the second-best methods by margins of 1.9 in CLIP-I (compared to 82.2 by Kosmos-G), 4.3 in DINO (compared to 67.1 by MS-Diffusion), and 1.8 in CLIP-T (compared to 32.1 by MS-Diffusion). Notably, our method even surpasses models fine-tuned at test time, showcasing robust zero-shot capabilities. TFCustom's superior performance can be attributed to the synchronized reference network, which provides multi-scale features for highly detailed image generation, and the time-aware frequency module, which harmonizes semantic information with reference image features. Additionally, the reward model enhances object similarity by leveraging referential cues. When fine-tuned on DreamBench, TFCustom shows further improvements, achieving top scores in CLIP-I, DINO, and CLIP-T metrics with values of 85.9, 74.1, and 35.1, respectively. This highlights TFCustom's adaptability in personalization scenarios, demonstrating strong transferability across various contexts.", + "bbox": [ + 89, + 537, + 483, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison of Multi-Subject. We also com", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aba44638a10e6e1667bdb7689af7c0ce19978c26621e7eaf6352dda30c6ca079.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCLIP-IDINOM-DINOCLIP-T
λ-ECLIPSE [30]72.441.99.431.6
SSR-Encoder [58]72.542.510.730.3
MS-Diffusion [48]69.842.510.834.1
TFCustom74.343.812.936.9
", + "bbox": [ + 516, + 483, + 903, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison for multi-subject models. The best and second-best results are shown in **bold** and **underline**, respectively.", + "bbox": [ + 511, + 592, + 906, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pare multi-subject personalization performance in Tab. 2. The results show that TFCustom achieves state-of-the-art performance across CLIP-I, DINO, M-DINO, and CLIP-T metrics, with scores of 74.3, 43.8, 12.9, and 36.9, respectively, significantly surpassing other reference image generation methods. Compared to MS-Diffusion, TFCustom demonstrates superior text-following capability, as evidenced by its higher CLIP-T scores, outperforming methods like MS-Diffusion and SSR-Encoder. Additionally, TFCustom excels in the M-DINO metric, highlighting its ability to preserve multi-object details in generated images.", + "bbox": [ + 511, + 643, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 4 presents qualitative results from the DreamBench dataset, demonstrating our model's remarkable capacity for generating high-fidelity single-subject images with excellent detail preservation. Furthermore, Fig. 5 illustrates qualitative outcomes from the MS-Bench dataset. Our approach adeptly maintains the textual", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2719", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73b07ca0f3e9828453e715165cdb21d75b6bc9e2352b9bd439a40c8e6865e820.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of multiple-subject customization. TFCustom achieves enhanced text controllability and subject fidelity compared to existing methods. Customized words are highlighted." + ], + "image_footnote": [], + "bbox": [ + 99, + 94, + 321, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb6189be6e80ccdb606571c4f142db0f8b8395649ad793ee4a554c2631625881.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 94, + 388, + 464 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/029a911d6c107bf7d0ef98d912ddf90ea56dfb771895d36f410c3f2e7c2f772b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 93, + 519, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7855f3ac4fdcc4665fe333aac8b565c9447429107cf7fe32c763585eda93be7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 93, + 647, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dbfc8cb3305f785a86b4522e1d20a0ee49a16368b5840c736b133562ccb82480.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 93, + 772, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08ce89588935041d5eb88e49f4b0a254ff032f750dfa1f3f56dba3fc9b13802e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 93, + 895, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38a68a7550eec924bf69f25bf1a1cc6cbc323ac5bd999bbb236a662102882eba.jpg", + "image_caption": [ + "Figure 6. Visualizations in customized human generation." + ], + "image_footnote": [], + "bbox": [ + 93, + 523, + 480, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "interactions among multiple subjects within the generated images, ensuring that each subject remains distinct and easily recognizable. This effectively mitigates issues related to subject omission or conflicts.", + "bbox": [ + 89, + 686, + 482, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Complex Scenarios. We evaluate the customized human generation in Fig. 6. The result shows that TFCustom is still capable of generating high-quality results for customized human generation. Additionally, we evaluate our model in real-world generation scenarios, as shown in Fig. 7. Our model is capable of generating images that accurately capture the fine-grained details of the reference images, including text, textures, and complex patterns. These showcase its strong capability in reference image feature transfer and high-quality image generation.", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f02264ee28ba1057f86c5f483d8e8f3245fdac29beac3d66501323ecc25d9ad0.jpg", + "image_caption": [ + "Figure 7. Visualizations in complex scenarios with intricate patterns and text." + ], + "image_footnote": [], + "bbox": [ + 519, + 518, + 903, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 751, + 666, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Design. We conduct ablation experiments to assess each component of our method, as shown in Tab. 3. $\\mathcal{L}_{\\mathrm{ref}}$ represents synchronized denoising, $\\mathcal{L}_{\\mathrm{reward}}$ involves the reward model, and Frequency denotes the time-aware frequency separation module. Each component significantly enhances performance. Synchronized denoising $\\mathcal{L}_{\\mathrm{ref}}$ sets a strong baseline, executing scores of 83.2, 72.4, and 32.9 for CLIP-T, CLIP-I, and DINO. The Frequency module improves the", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2720", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce7bad882a8ffa5d28fad480e0fa58f6b9ecd3b50fb8d20a22567373ae40056e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LrefLrewardFreq.CLIP-IDINOCLIP-T
XXX81.371.532.7
✓XX83.272.432.9
X✓✓84.373.734.1
✓✓✓85.974.135.1
", + "bbox": [ + 91, + 88, + 357, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34473d79f5dca5ad4d08116b2c6050793d3fa00376e28ce4fd08e95c5090dcd8.jpg", + "table_caption": [ + "Table 3. Impact $(\\%)$ of different components. Freq. denotes TA-FFR module." + ], + "table_footnote": [], + "table_body": "
Data ScaleCLIP-IDINOCLIP-T
0.5 M80.169.830.1
1 M84.072.832.3
1.5 M84.973.734.8
2.3 M85.974.135.1
", + "bbox": [ + 375, + 88, + 604, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0f1720135307f9934c7bf27b842146eb7575d98e193cb1b5b735ae12fcbd24b7.jpg", + "table_caption": [ + "Table 4. Impact (%) of data scale." + ], + "table_footnote": [], + "table_body": "
ModelCLIP-IDINOCLIP-T
DeepLabv3-MBv284.172.734.6
FCN-R10184.372.934.6
UperNet-R5085.473.834.7
DINOv285.974.135.1
", + "bbox": [ + 620, + 88, + 905, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Impact (%) of different reward models.", + "bbox": [ + 619, + 191, + 870, + 217 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1dda87ed4019b9a602ad08633b78d9d46d46942db787c7e519fa4710ecf523db.jpg", + "image_caption": [ + "Figure 8. The visualization of various ablation modules indicates that each proposed component within our methodology contributes positively to the overall results." + ], + "image_footnote": [], + "bbox": [ + 94, + 234, + 480, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performance across metrics (e.g., CLIP-I improves from 71.5 to 72.4), effectively enhancing temporal coherence and structural detail. The reward model $\\mathcal{L}_{\\text {reward }}$ further boosts results by an average of 1.3 points (e.g., CLIP-I increases from 72.4 to 73.7), capturing nuanced information critical for quality restoration. With all components combined, we achieve the highest scores, demonstrating the complementary strengths of each component. Fig. 8 further examines the effect of each component, demonstrating the guidance of each component.", + "bbox": [ + 88, + 429, + 482, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reward Model. Following [21], we examine the effectiveness of different reward models, as shown in Tab. 5. DINOv2 [2] achieves the highest scores across all metrics, demonstrating superior performance in text relevance and image fidelity. FCN [25] with ResNet-101 [13] backbone(FCN-R101), DeepLabv3 [3] with MobileNetv2 [37] (DeepLabv3-MBv2) also perform well, though slightly lower, confirming that the reward model enhances feature retention and alignment across architectures, with DINOv2 as the optimal choice.", + "bbox": [ + 89, + 580, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 89, + 741, + 256, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Map Visualization. Figure 9 visualizes the cross-attention map between reference and latent features, aggregated across all layers to a uniform scale and averaging. Distinct reference images produce unique feature maps, as shown in the top row. Our method demonstrates consistent object perception across timesteps, a capability lacking in previous ReferenceNet features. Additionally, the incorporation of frequency loss and the reward model enables our approach to capture finer details, resulting in enhanced im", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age quality and more precise object representation.", + "bbox": [ + 513, + 234, + 849, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/412320aa4e44b559cf3122db14918d6f4b8ac25a391e7cba05dbb4afcfac5d87.jpg", + "image_caption": [ + "Figure 9. Cross-Attention Feature Maps. Visualization of the attention maps between reference features and latent features." + ], + "image_footnote": [], + "bbox": [ + 521, + 263, + 901, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Data. Tab. 4 demonstrates the impact of training data size on model performance. As the amount of data increases, we observe consistent improvements across all evaluation metrics (CLIP-T, CLIP-I, and DINO). Specifically, using 0.5M data samples yields moderate results, but expanding to 1M and 1.5M significantly enhances model accuracy, indicating the value of larger datasets in refining feature representation. The best performance is achieved with 2.3M samples, highlighting that increased training data directly contributes to more robust and detailed outputs.", + "bbox": [ + 511, + 482, + 906, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 648, + 638, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we propose TFCustom, a novel model framework that emphasizes the importance of reference image features across different temporal steps and frequency levels. Our approach uses synchronized ReferenceNet to extract these features while optimizing noise injection and denoising processes. Additionally, we introduce a time-aware frequency refinement module that employs high- and low-frequency filters, along with time embeddings, to dynamically adjust the intensity of reference feature injection. To improve alignment between reference objects and generated images, we propose a reward model loss that strengthens their correspondence. Experimental results demonstrate the effectiveness of our framework, achieving state-of-the-art performance and significant improvements in texture and textual detail generation compared to existing methods.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2721", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2:3, 2023. 2", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5, 8", + "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation, 2017. 8", + "[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. CoRR, abs/2307.09481, 2023. 2", + "[5] Xi Chen, Zhiheng Liu, Mengting Chen, Yutong Feng, Yu Liu, Yujun Shen, and Hengshuang Zhao. Livephoto: Real image animation with text-guided motion control, 2023. 2", + "[6] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for authentic virtual try-on in the wild, 2024. 2", + "[7] Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. Adair: Adaptive all-in-one image restoration via frequency mining and modulation. arXiv preprint arXiv:2403.14614, 2024. 3", + "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2", + "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 5", + "[10] Xiang Gao and Jiaying Liu. Fbsdiff: Plug-and-play frequency band substitution of diffusion features for highly controllable text-driven image translation. In ACM MM, pages 4101-4109, 2024. 3", + "[11] Xiang Gao, Zhengbo Xu, Junhan Zhao, and Jiaying Liu. Frequency-controlled diffusion model for versatile text-guided image-to-image translation. In AAAI, pages 1824-1832, 2024. 3", + "[12] Yuming Gu, You Xie, Hongyi Xu, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis, 2024. 2", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 8", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2", + "[15] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive mod" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "els for fine-grained text-to-image synthesis. In AAAI, 2025. 2", + "[16] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation, 2024. 2, 3, 4", + "[17] Zehuan Huang, Hongxing Fan, Lipeng Wang, and Lu Sheng. From parts to whole: A unified reference framework for controllable human image generation, 2024. 2, 3, 4", + "[18] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pages 1931-1941, 2023. 2, 3", + "[19] Bozheng Li, Mushui Liu, Gaoang Wang, and Yunlong Yu. Frame order matters: A temporal sequence-aware model for few-shot action recognition. In AAAI, 2025. 2", + "[20] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 5", + "[21] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In ECCV, 2024.8", + "[22] Chengxu Liu, Xuan Wang, Shuai Li, Yuzhi Wang, and Xueming Qian. Fsi: Frequency and spatial interactive learning for image restoration in under-display cameras. In ICCV, pages 12537-12546, 2023. 3", + "[23] Mushui Liu, Yuhang Ma, Yang Zhen, Jun Dan, Yunlong Yu, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. In AAAI, 2025. 2", + "[24] Mushui Liu, Fangtai Wu, Bozheng Li, Ziqian Lu, Yunlong Yu, and Xi Li. Envisioning class entity reasoning by large language models for few-shot learning. In AAAI, 2025. 2", + "[25] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation, 2015. 8", + "[26] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2, 3", + "[27] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In AAAI, pages 1905–1913, 2023. 3", + "[28] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296–4304, 2024. 3", + "[29] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 5", + "[30] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. lambda-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 5, 6", + "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4172-4182, 2023. 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2722", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952, 2023. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685, 2022. 2", + "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2, 5", + "[36] Donghun Ryou, Inju Ha, Hyewon Yoo, Dongwan Kim, and Bohyung Han. Robust image denoising through adversarial frequency mixup. In CVPR, pages 2723-2732, 2024. 3", + "[37] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks, 2019. 8", + "[38] D She, Mushui Liu, Jingxuan Pang, Jin Wang, Zhen Yang, Wanggui He, Guanghao Zhang, Yi Wang, Qihan Huang, Haobin Tang, et al. Customvideox: 3d reference attention driven dynamic adaptation for zero-shot customized video diffusion transformers. arXiv preprint arXiv:2502.06527, 2025.2", + "[39] Fei Shen and Jinhui Tang. Imagpose: A unified conditional framework for pose-guided person generation. In NeurIPS, 2024.", + "[40] Fei Shen, Hu Ye, Jun Zhang, Cong Wang, Xiao Han, and Wei Yang. Advancing pose-guided image synthesis with progressive conditional diffusion models. In ICLR, 2023. 2", + "[41] Fei Shen, Xin Jiang, Xin He, Hu Ye, Cong Wang, Xiaoyu Du, Zechao Li, and Jinghui Tang. Imagdressing-v1: Customizable virtual dressing. In AAAI, 2025. 3", + "[42] Fei Shen, Cong Wang, Junyao Gao, Qin Guo, Jisheng Dang, Jinhui Tang, and Tat-Seng Chua. Long-term talkingface generation via motion-prior conditional diffusion model. arXiv preprint arXiv:2502.09533, 2025. 2", + "[43] Fei Shen, Hu Ye, Sibo Liu, Jun Zhang, Cong Wang, Xiao Han, and Wei Yang. Boosting consistency in story visualization with rich-contextual conditional diffusion models. In AAAI, 2025. 2", + "[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 5", + "[45] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimen" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2, 4", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2", + "[47] Rui Wang, Hailong Guo, Jiaming Liu, Huaxia Li, Haibo Zhao, Xu Tang, Yao Hu, Hao Tang, and Peipei Li. Stablegar-ment: Garment-centric generation via stable diffusion, 2024. 2,3", + "[48] X. Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance, 2024. 2, 3, 5, 6", + "[49] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15897-15907, 2023. 2", + "[50] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on, 2024. 2, 3", + "[51] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model, 2023. 2", + "[52] Zhen Yang, Guibao Shen, Liang Hou, Mushui Liu, Luozhou Wang, Xin Tao, Pengfei Wan, Di Zhang, and YingCong Chen. Rectifiedhr: Enable efficient high-resolution image generation via energy rectification. arXiv preprint arXiv:2503.02537, 2025. 2", + "[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3, 5", + "[54] Jiacheng Ying, Mushui Liu, Zhe Wu, Running Zhang, Zhu Yu, Siming Fu, Si-Yuan Cao, Chao Wu, Yunlong Yu, and Hui-Liang Shen. Restorerid: Towards tuning-free face restoration with id preservation. arXiv preprint arXiv:2411.14125, 2024. 2", + "[55] Yu Zeng, Vishal M. Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In CVPR 2024, pages 6786-6795, 2024. 3", + "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3813-3824, 2023. 3", + "[57] Shilong Zhang, Lianghua Huang, Xi Chen, Yifei Zhang, ZhiFan Wu, Yutong Feng, Wei Wang, Yujun Shen, Yu Liu, and Ping Luo. Flashface: Human image personalization with high-fidelity identity preservation, 2024. 2, 3", + "[58] Yuxuan Zhang, Jiaming Liu, Yiren Song, Rui Wang, Hao Tang, Jinpeng Yu, Huaxia Li, Xu Tang, Yao Hu, Han Pan, and Zhongliang Jing. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR, pages 8069–8078, 2024. 2, 3, 5, 6" + ], + "bbox": [ + 516, + 92, + 903, + 854 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2723", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_model.json b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9bc5816b36a7772ce0406c940d5277c3998e269 --- /dev/null +++ b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_model.json @@ -0,0 +1,2376 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.326, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.13, + 0.862, + 0.175 + ], + "angle": 0, + "content": "TFCustom: Customized Image Generation with Time-Aware Frequency Feature Guidance" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.203, + 0.282, + 0.221 + ], + "angle": 0, + "content": "Mushui Liu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.204, + 0.399, + 0.222 + ], + "angle": 0, + "content": "Dong She\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.204, + 0.545, + 0.222 + ], + "angle": 0, + "content": "Jingxuan Pang1" + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.204, + 0.682, + 0.222 + ], + "angle": 0, + "content": "Qihan Huang" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.222, + 0.424, + 0.24 + ], + "angle": 0, + "content": "Wanggui \\(\\mathrm{He}^3\\)" + }, + { + "type": "text", + "bbox": [ + 0.447, + 0.222, + 0.555, + 0.24 + ], + "angle": 0, + "content": "Yuanlei Hou" + }, + { + "type": "text", + "bbox": [ + 0.576, + 0.222, + 0.682, + 0.24 + ], + "angle": 0, + "content": "Siming \\(\\mathrm{Fu}^{1*^{\\dagger}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.242, + 0.372, + 0.261 + ], + "angle": 0, + "content": "1Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.242, + 0.795, + 0.261 + ], + "angle": 0, + "content": "\\(^{2}\\)University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.261, + 0.443, + 0.278 + ], + "angle": 0, + "content": "3Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.483, + 0.261, + 0.717, + 0.277 + ], + "angle": 0, + "content": "4 Middle East Centre, RIPED" + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.298, + 0.608, + 0.311 + ], + "angle": 0, + "content": "Single-Subject Driven Generation" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.313, + 0.905, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.508, + 0.622, + 0.521 + ], + "angle": 0, + "content": "Multiple-Subject Driven Generation" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.523, + 0.895, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.693, + 0.853, + 0.708 + ], + "angle": 0, + "content": "Figure 1. Our proposed TFCustom demonstrates capabilities in single-subject and multi-subject driven generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.731, + 0.327, + 0.747 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.486, + 0.857 + ], + "angle": 0, + "content": "Subject-driven image personalization has seen notable advancements, especially with the ReferenceNet paradigm, which excels in integrating reference image features for creative and commercial applications. However, current ReferenceNet implementations mainly function as latent-level feature extractors, limiting their potential. This restricts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.901 + ], + "angle": 0, + "content": "the delivery of suitable features to the denoising backbone across timesteps, resulting in suboptimal image consistency. In this paper, we revisit reference feature extraction and propose TFCustom, a framework that focuses on reference image features at different temporal and frequency levels. We introduce synchronized ReferenceNet to extract reference features while optimizing noise injection and denoising. We also propose a time-aware frequency refinement module that uses high- and low-frequency filters with time embeddings to adaptively select reference feature injection. Additionally, we introduce a reward-based loss to improve" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.875, + 0.231, + 0.888 + ], + "angle": 0, + "content": "*Equal contributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.888, + 0.241, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.875, + 0.241, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2714" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "the similarity between reference objects and generated images. Experimental results show that TFCustom outperforms existing methods in single-object and multi-object reference generation, with significant improvements in textual details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.196, + 0.223, + 0.212 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.221, + 0.485, + 0.401 + ], + "angle": 0, + "content": "With the development of deep learning [14, 19, 24, 46, 52] and text-to-image models [8, 15, 23, 32, 34], customization generation which aims to synthesize text-driven images for given subjects, has recently revolutionized content creation. It has seen a shift from test-time fine-tuning methods (e.g., DreamBooth [35], Custom Diffusion [18]) to fine-tuning-free techniques (e.g., IP-Adapter [53], Subject-Diffusion [26], MS-Diffusion [48]). The latter approaches remove the need for fine-tuning during inference, significantly reducing computational costs while maintaining flexibility, and making them more scalable and practical for real-world use cases." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.403, + 0.484, + 0.719 + ], + "angle": 0, + "content": "These fine-tuning-free methods generally fall into two categories: image encoder-based [4, 5, 58] feature injection and reference network (ReferenceNet-based) [6, 12, 51] feature injection. Image encoder methods, such as IPAdapter [53] and ELITE [49], utilize pre-trained encoders like CLIP [33] to extract high-level features, which are then injected into the generation process via mechanisms such as decoupled cross-attention. While these methods are computationally efficient and straightforward, they tend to rely on high-level, abstract features, which may not capture the fine-grained details necessary for precise control over reference object characteristics. The increasing demand for greater consistency in reference images has brought significant attention to the concept of ReferenceNet [16] in the field. ReferenceNet-based methods [17, 47, 50, 57] introduce an additional diffusion network within the traditional Diffusion Net denoising process. In this setup, reference images are encoded by VAE and then used as inputs for ReferenceNet. This novel architecture enables the correlation of features across layers in both diffusion networks, which helps in making better use of the reference images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Although ReferenceNet-based methodologies have made significant advancements, several challenges remain. Currently, ReferenceNet predominantly operates as a latent-level feature extractor, thereby not fully realizing its potential. While ReferenceNet maintains a consistent intensity of feature conditioning over time, the denoising backbone requires the injection of varying intensities and granularities of reference information at different timesteps. Specifically, during the earlier stages of denoising, the backbone tends to emphasize contour information, whereas, in the later stages, it focuses more on the refinement of details [45]. This limitation hinders the ability" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "to provide the denoising backbone with the appropriate features necessary for optimal performance across varying timesteps, ultimately leading to subpar image consistency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.529 + ], + "angle": 0, + "content": "In this work, we introduce TFCustom, a novel finetuning-free personalized generation framework that incorporates temporal awareness for reference feature injection. Unlike previous approaches, TFCustom considers the alignment of reference image features with generated image features across different time steps of the diffusion process. Specifically, we first apply noise to the reference image at multiple time steps, with ReferenceNet extracting features at various scales and granularities. The training of ReferenceNet follows a diffusion loss function, ensuring proper alignment with the denoising process. Additionally, we introduce a time-aware frequency module that enables the controlled injection of high- and low-frequency reference features. This module allows low-frequency features to define the primary subject at early time steps, while high-frequency features are progressively injected to refine finer image details. To further improve the model's ability to handle multi-object reference generation, we propose a novel reward-based loss function that enhances the alignment between the reference and generated objects. This loss function computes object ID similarity during the denoising process, minimizing conflicts between reference objects and improving the overall coherence of multi-object scenes. The experimental results demonstrate our method consistently outperforms the state-of-the-art approaches on all the benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.53, + 0.807, + 0.544 + ], + "angle": 0, + "content": "Overall, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.545, + 0.905, + 0.619 + ], + "angle": 0, + "content": "- We propose a novel customized image generation framework, which includes synchronized noise injection and denoising within ReferenceNet, along with a time-aware frequency feature refinement module for extracting fine-grained image details." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.62, + 0.905, + 0.664 + ], + "angle": 0, + "content": "- We propose a reward-based loss function that alleviates multi-object conflicts and enhances object identity consistency in generated images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.665, + 0.905, + 0.74 + ], + "angle": 0, + "content": "- Our model achieves state-of-the-art performance in single-object reference generation on DreamBooth and MS-Bench datasets. Additionally, we demonstrate superior qualitative results in terms of texture and textual detail generation compared to existing methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.545, + 0.905, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.755, + 0.655, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Subject Driven Text-to-Image Generation. Despite significant advancements in image generation [1, 8, 15, 23, 34], the field still falls short of meeting the requirements for personalized generation [38-40, 42, 48, 54]. This gap has consequently drawn significant attention to subject-driven text-to-image generation methods [4, 26, 43]. Textual Inversion [9] and DreamBooth [35] use a textual identifier to bind the visual concept through fine-tuning model textual" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2715" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "embedding or UNet model. CustomDiffusion [18] further extends multi-subject generation by multiple identifiers. IP-Adapter [53] injects the visual concept by decoupling the cross-attention between text and images in a zero-shot manner. JeDi [55] learns the joint distribution of multiple text-image pairs that share a common subject to facilitate personalized generation. Subject-Diffusion [26] trains multiple adapters inside the model with an attention control mechanism for multiple-subject personalization. SSR-Encoder [58] and MS-Diffusion [48] develop the query architecture to encode the subject images via mask or layout control for multiple-subject personalization. This paper explores the zero-shot multiple-objects driven generation via synchronized ReferenceNet optimization." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.304, + 0.483, + 0.575 + ], + "angle": 0, + "content": "ReferenceNet Based Personalized Generation. ReferenceNet [16, 41, 57] has shown remarkable capabilities in maintaining image consistency for personalized generation tasks. AnimateAnyone [16] first utilizes ReferenceNet for personalized human animation, employing a pre-trained generative model to extract reference features. FlashFace [57] extended this approach by combining ReferenceNet with ControlNet [56], enabling identity-preserving portrait generation that accommodates multiple reference images. Part2Whole [17] uses ReferenceNet to handle multi-part human body generation, while OOTD [50] and Stable-Garment [47] integrate it into virtual try-on tasks, aligning clothing features precisely with the target human body. These methods highlight ReferenceNet's versatility in image and video generation. To further improve performance, we introduce a synchronized ReferenceNet that refines reference features at each time step, enabling more precise feature injection and enhancing generation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.756 + ], + "angle": 0, + "content": "Frequency Information in Image Generation. Frequency information plays a crucial role in image restoration tasks [7, 22, 27, 36]. Recent advancements in text-to-image generation have leveraged frequency-based conditions, like canny and HED images, to control generation, as seen in ControlNet [56] and T2I-Adapter [28]. FCDiffusion [11] and FBSdiff [10] employ DCT-transformed image features to enhance texture detail in image-to-image generation. In this work, our method adaptively injects reference image features via learnable high-pass and low-pass filters in a time-aware fashion, enriching each denoising step with frequency-specific information for more refined generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.182, + 0.785 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Our proposed TFCustom addresses these issues through three components: Sec. 3.2 Synchronized Reference Guidance, Sec. 3.3 Time-Aware Frequency Feature Refinement, and Sec. 3.4 Reward Model Optimization, as illustrated in Fig. 2. First, the reference image is incrementally noised over \\(T\\) steps before feature extraction via the ReferenceNet, aligning it with the denoising network's noisy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "latent features. Next, a time-aware frequency module injects frequency-specific information from the reference image into the noisy latent representation. Finally, a reward model optimizes this process, enhancing the quality and consistency of the generated reference-based images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.177, + 0.658, + 0.191 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.905, + 0.291 + ], + "angle": 0, + "content": "Diffusion models. Denoising Diffusion Probabilistic Models (DDPM) are a class of generative models that leverage a diffusion process to generate data. By progressively adding Gaussian noise to the data over a specified number of time steps, a sequence of increasingly noisy samples is produced. The forward process can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.302, + 0.905, + 0.319 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {T} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {T}; \\sqrt {\\alpha_ {T}} x _ {0}, (1 - \\alpha_ {T}) \\mathbf {I}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.905, + 0.419 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} = \\prod_{s=1}^{t}(1 - \\beta_{s})\\). Here, \\(\\beta_{t}\\) represents a variance schedule that regulates the noise level at each time step \\(t\\), and \\(N\\) denotes a Gaussian distribution. In the reverse diffusion process, the model learns to denoise the noisy samples step by step, effectively generating new data from pure noise:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.432, + 0.905, + 0.451 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\sigma_ {\\theta} ^ {2} (t) \\mathbf {I}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.905, + 0.522 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}\\) and \\(\\sigma_{\\theta}\\) are learned parameters that depend on the current noisy sample \\(x_{t}\\) and the time step \\(t\\). The optimization goal is the variational lower bound on the log-likelihood of the data, which can be simplified to:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.532, + 0.905, + 0.551 + ], + "angle": 0, + "content": "\\[\nL = \\mathbb {E} _ {t, x _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2} \\right] \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.842, + 0.577 + ], + "angle": 0, + "content": "3.2. Synchronized ReferenceNet Guidance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.84 + ], + "angle": 0, + "content": "The synchronized ReferenceNet is initially transferred from the denoising network. During training, at each time step \\( t \\), we apply a noise process to the reference image \\( x_{\\mathrm{ref}} \\), adding noise for \\( t \\) steps, resulting in the noise reference \\( x_{\\mathrm{ref}}^{t} \\) as formulated in Eq. (1). This step ensures that the reference features are temporally aligned with the noisy latent features processed by the denoising network. Once the noised reference \\( x_{\\mathrm{ref}}^{t} \\) is generated, it passes through the synchronized ReferenceNet, which extracts the relevant features \\( \\mathbf{F}_{\\mathrm{ref}} \\) and estimates the corresponding noise pattern. The extracted reference features and predicted noise are then optimized according to the objective \\( \\mathcal{L}_{\\mathrm{diff}}^{\\mathrm{ref}} \\) defined in Eq. (3). This objective encourages the synchronized ReferenceNet to provide refined guidance that is consistent with the noisy latent representations, enabling the denoising network to more effectively incorporate reference information for improved reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Given that the object information hierarchically learned by the denoising network varies significantly across different timesteps, the early timesteps tend to focus on learning macro-level information such as contours, while the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2716" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.518, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.302, + 0.376, + 0.317 + ], + "angle": 0, + "content": "(a) Overall Framework" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.099, + 0.903, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.302, + 0.895, + 0.318 + ], + "angle": 0, + "content": "(b) Time-Aware Frequency Feature Refinement" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.333, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 2. (a) Overall Framework. Our TFCustom includes three key components: Synchronized Reference Net, Time-Aware Frequency Feature Refinement (TA-FFR), and Reward Model Optimization for subject-driven generation. (b) TA-FFR Module. This module extracts different frequency features in a time-aware manner, adaptively injecting reference features for enhanced generation quality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.484, + 0.538 + ], + "angle": 0, + "content": "later timesteps progressively refine texture details. However, the synchronized timestep information obtained from \\(\\mathcal{L}_{\\mathrm{diff}}^{\\mathrm{ref}}\\) encompasses the entire reference object's information at timestep \\(t\\), which can be overly complex for the denoising network. To address this issue, we have implemented a Time-Aware Frequency Feature Refinement module (which will be described detailed in Sec. 3.3) to refine the reference information, thereby enhancing the learning process of the denoising network, as outlined in Eq. 7." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.675 + ], + "angle": 0, + "content": "With the enhanced reference features \\(\\mathbf{F}_{\\mathrm{enh}}\\), we inject them using a Shared Self-Attention mechanism [16, 17]. Specifically, in each self-attention layer, given the features of \\(N\\) reference images \\(\\mathbf{F}_{\\mathrm{enh}}^{1:N}\\) and the feature map \\(\\mathbf{F}^0\\) from the denoising net, we concatenate these feature maps side-by-side as input to the self-attention layer, represented as \\([\\mathbf{F}_{\\mathrm{enh}}^0 | \\mathbf{F}_{\\mathrm{enh}}^1| \\dots | \\mathbf{F}_{\\mathrm{enh}}^N]\\). This setup enables each location \\(s\\) on \\(\\mathbf{F}^0\\) to attend to all locations on itself and the reference feature maps, computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.686, + 0.482, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\mathbf {F}} _ {s} ^ {0} = \\operatorname {S o f t M a x} \\left(\\frac {\\mathbf {F} _ {s} ^ {0} W _ {q} \\cdot \\left([ \\mathbf {F} _ {\\mathrm {e n h}} ^ {0} , \\mathbf {F} _ {\\mathrm {e n h}} ^ {1} , \\dots , \\mathbf {F} _ {\\mathrm {e n h}} ^ {N} ]\\right) ^ {\\top} W _ {k}}{\\sqrt {d}}\\right) \\\\ \\cdot \\left(\\left[ \\mathbf {F} _ {\\text {e n h}} ^ {0}, \\mathbf {F} _ {\\text {e n h}} ^ {1}, \\dots , \\mathbf {F} _ {\\text {e n h}} ^ {N} \\right]\\right) W _ {v}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.792 + ], + "angle": 0, + "content": "where \\( W_{q}, W_{k}, W_{v} \\) are corresponding projection weight matrices. And \\( d \\) is the dimensionality of the key vectors." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.801, + 0.466, + 0.818 + ], + "angle": 0, + "content": "3.3. Time-Aware Frequency Feature Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Inspired by previous work [45], which illustrates that neural network learns hierarchically—starting with low-frequency features that define shape and structure, followed by high-frequency details for realism, we propose that reference image feature injection should follow a similar coarse-to-fine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.905, + 0.431 + ], + "angle": 0, + "content": "approach. This would better align with the denoising stages of diffusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.614 + ], + "angle": 0, + "content": "To achieve this, we introduce a Time-Aware Frequency Feature Refinement (TA-FFR), as shown in Fig. 2. This module separates reference image features into high-frequency and low-frequency components, each processed independently to capture unique information. The high-frequency path applies a Krisch operator for edge and detail preservation, while the low-frequency path uses a Gaussian operator to retain foundational structures. These operators are initialized as \\(5 \\times 5\\) convolution kernels \\(\\mathbf{H}_{\\mathrm{conv}}\\) is Krisch operator, and \\(\\mathbf{L}_{\\mathrm{conv}}\\) is Gaussian operator. Both operators are scaled by learnable parameters to adjust the frequency extraction dynamically:" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.628, + 0.905, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {\\text {o u t}} = \\operatorname {C o n v} _ {\\mathbf {H}} \\left(\\mathbf {F} _ {\\text {r e f}}\\right) \\cdot \\mathbf {W} _ {\\mathrm {H}}, \\quad \\mathbf {L} _ {\\text {o u t}} = \\operatorname {C o n v} _ {\\mathbf {L}} \\left(\\mathbf {F} _ {\\text {r e f}}\\right) \\cdot \\mathbf {W} _ {\\mathrm {L}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.905, + 0.689 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{\\mathrm{ref}}\\) denotes reference image features, and \\(\\mathbf{W}_{\\mathrm{H}}\\) and \\(\\mathbf{W}_{\\mathrm{L}}\\) are scaling factors for high and low frequencies." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.906, + 0.766 + ], + "angle": 0, + "content": "Each frequency component then undergoes Time-Aware Attention (TA-Attention) and Time-Aware Feed-Forward Network (TA-FFN), which incorporate time-step information using a time embedding \\(\\mathbf{t}_{\\mathrm{emb}}\\) via Adaptive Layer Normalization (AdaLN) proposed in DiT [31]:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.778, + 0.905, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {F} _ {\\text {h i g h}} = \\mathrm {T A - F F N} (\\mathrm {T A - A t t e n t i o n} (\\mathrm {A d a L N} (\\mathbf {H} _ {\\text {o u t}}, \\mathbf {t} _ {\\text {e m b}}))) \\\\ \\mathbf {F} _ {\\text {l o w}} = \\mathrm {T A - F F N} (\\mathrm {T A - A t t e n t i o n} (\\mathrm {A d a L N} (\\mathbf {L} _ {\\text {o u t}}, \\mathbf {t} _ {\\text {e m b}}))), \\end{array} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.872 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{\\mathrm{high}}\\) and \\(\\mathbf{F}_{\\mathrm{low}}\\) represent high-frequency and low-frequency features. Then the final enhanced feature is obtained by combining these outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\text {e n h}} = \\mathbf {F} _ {\\text {h i g h}} + \\mathbf {F} _ {\\text {l o w}}. \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2717" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.094, + 0.475, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.243 + ], + "angle": 0, + "content": "Figure 3. Reward Model Optimization Process. \"G\" and \"D\" indicate the network's forward pass and backward optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.484, + 0.329 + ], + "angle": 0, + "content": "This design ensures a coarse-to-fine feature injection, enhancing the model's ability to integrate temporal and spatial details, thereby improving the quality of the generated images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.337, + 0.348, + 0.354 + ], + "angle": 0, + "content": "3.4. Reward Model Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.437 + ], + "angle": 0, + "content": "Our model can simulate the noise levels of both the reference image and the target image. Based on the following equation, we can perform a one-step prediction of the source image, allowing us to naturally determine the reward model score based on the noise level by:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.447, + 0.483, + 0.481 + ], + "angle": 0, + "content": "\\[\nx _ {0} ^ {\\prime} \\approx \\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} \\left(x _ {t} , x _ {\\text {r e f}} , c _ {t} , t\\right)}{\\sqrt {\\alpha_ {t}}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.551 + ], + "angle": 0, + "content": "where \\( c_{t} \\) refers to the textual condition. This equation's calculation error increases with timestep \\( t \\). Therefore, during the training phase, we only apply the reward loss for timestep \\( t < T_0 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.562, + 0.483, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e w a r d}} = \\operatorname {D i s} \\left(\\mathrm {D I N O} _ {v 2} \\left(x _ {0}\\right), \\mathrm {D I N O} _ {v 2} \\left(x _ {0} ^ {\\prime}\\right)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.755 + ], + "angle": 0, + "content": "where Dis is cosine similarity metric and \\(\\mathrm{DINO}_{v2}\\) is the DINO v2 model [2] to compute the similarity between the target domain's denoised output and the features generated by the reference net. It is worth noticing that in our synchronized noise injection scheme, the output of the reference net corresponds to the noise level at timestep \\(t\\), which provides meaningful information. Thus, we take advantage of this synchronized noise injection feature and introduce a diffusion loss, \\(L_{\\mathrm{ref}}\\), for the reference net, which supervises the feature extraction capability of the reference branch. The overall loss function can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.765, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d i f f}} ^ {\\text {d e n o i s i n g}} + \\lambda_ {1} \\mathcal {L} _ {\\text {d i f f}} ^ {\\text {r e f}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e w a r d}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.414, + 0.81 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) is hyper-parameters to balance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.823, + 0.224, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.268, + 0.865 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Datasets. The training dataset comprises 2.3 million (M) in-house image pairs. We assess the performance on" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.09, + 0.905, + 0.31 + ], + "angle": 0, + "content": "
MethodTypeCLIP-IDINOCLIP-T
BLIP-Diffusion [20]No-FT77.959.430.0
IP-Adapter [53]No-FT81.061.329.2
Emu2 [44]No-FT76.556.327.3
Kosmos-G [29]No-FT82.261.825.0
λ-ECLIPSE [30]No-FT78.361.330.7
SSR-Encoder [58]No-FT82.161.230.8
MS-Diffusion [48]No-FT79.267.132.1
TFCustomNo-FT84.171.433.9
Textual Inversion [9]FT78.056.925.5
DreamBooth [35]FT80.366.830.5
BLIP-Diffusion [20]FT80.567.030.2
λ-ECLIPSE [30]FT79.668.230.4
MS-Diffusion [48]FT80.570.231.3
TFCustomFT85.974.135.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.321, + 0.905, + 0.363 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison for single-subject models. FT denotes fine-tuned models on DreamBench [35]. The best and second-best results are shown in bold and underline, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.508 + ], + "angle": 0, + "content": "single-subject tasks using the DreamBench dataset [35] and on multi-subject tasks using the MS-Bench dataset [48]. DreamBench encompasses 30 subjects and 25 prompts, while MS-Bench includes 40 subjects and 1148 combinations, with each combination being associated with up to 6 prompts, resulting in a total of 4488 unique test samples. Furthermore, we collect an additional 20 real-world cases featuring complex textures, characters, and patterns for further qualitative evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Implementation Details. During training, we employ the AdamW optimizer with a learning rate of 1e-4 and train the model for 100K steps on 32 NVIDIA A100 80GB GPUs, using a batch size of 4 per GPU. The \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) in Eq. (10) are set to 0.4 and 0.6. To enable classifier-free guidance, we apply a probability of 0.05 to independently drop text and images, and 0.05 to drop both simultaneously. For inference, we adapt the DDIM sampler with 50 steps and set the guidance scale to 7.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate performance based on both image and text fidelity metrics. For image fidelity, we compute the cosine similarity between generated images and subject images within the CLIP (denoted as CLIP-I) and DINO feature spaces. Text fidelity is measured by calculating the cosine similarity between generated images and text prompts in the CLIP space (referred to as CLIP-T). In the context of multi-subject personalization, average fidelity scores may not adequately reflect the representation of individual subjects. Therefore, we also employ the multi-subject DINO (M-DINO) metric [48] to assess whether each subject was accurately recreated." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Competitors. We compare our model with BLIP-Diffusion [20], IP-Adapter [53], Kosmos-G [29], Emu2 [44], \\(\\lambda\\)-ECLIPSE [30], SSR-Encoder [58], and MS-Diffusion [48] for single concept personalization, and with MS-Diffusion [48], SSR-Encoder [58] and \\(\\lambda\\)-ECLIPSE [30] for multi-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2718" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.905, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.431, + 0.908, + 0.462 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with existing state-of-the-art customization methods. TFCustom produces higher-quality results with better subject similarity and text controllability compared to existing methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.487, + 0.251, + 0.503 + ], + "angle": 0, + "content": "subject personalization." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.513, + 0.333, + 0.53 + ], + "angle": 0, + "content": "4.2. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.884 + ], + "angle": 0, + "content": "Quantitative Comparison of Single-Subject. The results are presented in Tab. 1. In the zero-shot evaluation (without fine-tuning on DreamBench), our proposed TFCustom achieves the highest scores across all metrics, with CLIP-I, DINO, and CLIP-T values of 84.1, 71.4, and 33.9, respectively. These results surpass the second-best methods by margins of 1.9 in CLIP-I (compared to 82.2 by Kosmos-G), 4.3 in DINO (compared to 67.1 by MS-Diffusion), and 1.8 in CLIP-T (compared to 32.1 by MS-Diffusion). Notably, our method even surpasses models fine-tuned at test time, showcasing robust zero-shot capabilities. TFCustom's superior performance can be attributed to the synchronized reference network, which provides multi-scale features for highly detailed image generation, and the time-aware frequency module, which harmonizes semantic information with reference image features. Additionally, the reward model enhances object similarity by leveraging referential cues. When fine-tuned on DreamBench, TFCustom shows further improvements, achieving top scores in CLIP-I, DINO, and CLIP-T metrics with values of 85.9, 74.1, and 35.1, respectively. This highlights TFCustom's adaptability in personalization scenarios, demonstrating strong transferability across various contexts." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison of Multi-Subject. We also com" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.582 + ], + "angle": 0, + "content": "
MethodCLIP-IDINOM-DINOCLIP-T
λ-ECLIPSE [30]72.441.99.431.6
SSR-Encoder [58]72.542.510.730.3
MS-Diffusion [48]69.842.510.834.1
TFCustom74.343.812.936.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.593, + 0.907, + 0.635 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison for multi-subject models. The best and second-best results are shown in **bold** and **underline**, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.81 + ], + "angle": 0, + "content": "pare multi-subject personalization performance in Tab. 2. The results show that TFCustom achieves state-of-the-art performance across CLIP-I, DINO, M-DINO, and CLIP-T metrics, with scores of 74.3, 43.8, 12.9, and 36.9, respectively, significantly surpassing other reference image generation methods. Compared to MS-Diffusion, TFCustom demonstrates superior text-following capability, as evidenced by its higher CLIP-T scores, outperforming methods like MS-Diffusion and SSR-Encoder. Additionally, TFCustom excels in the M-DINO metric, highlighting its ability to preserve multi-object details in generated images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 4 presents qualitative results from the DreamBench dataset, demonstrating our model's remarkable capacity for generating high-fidelity single-subject images with excellent detail preservation. Furthermore, Fig. 5 illustrates qualitative outcomes from the MS-Bench dataset. Our approach adeptly maintains the textual" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2719" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.095, + 0.322, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.095, + 0.39, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.094, + 0.521, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.094, + 0.648, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.094, + 0.774, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.094, + 0.897, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.477, + 0.908, + 0.506 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of multiple-subject customization. TFCustom achieves enhanced text controllability and subject fidelity compared to existing methods. Customized words are highlighted." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.524, + 0.481, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.661, + 0.458, + 0.675 + ], + "angle": 0, + "content": "Figure 6. Visualizations in customized human generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.748 + ], + "angle": 0, + "content": "interactions among multiple subjects within the generated images, ensuring that each subject remains distinct and easily recognizable. This effectively mitigates issues related to subject omission or conflicts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Complex Scenarios. We evaluate the customized human generation in Fig. 6. The result shows that TFCustom is still capable of generating high-quality results for customized human generation. Additionally, we evaluate our model in real-world generation scenarios, as shown in Fig. 7. Our model is capable of generating images that accurately capture the fine-grained details of the reference images, including text, textures, and complex patterns. These showcase its strong capability in reference image feature transfer and high-quality image generation." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.52, + 0.905, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.7, + 0.907, + 0.727 + ], + "angle": 0, + "content": "Figure 7. Visualizations in complex scenarios with intricate patterns and text." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.752, + 0.667, + 0.769 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Model Design. We conduct ablation experiments to assess each component of our method, as shown in Tab. 3. \\(\\mathcal{L}_{\\mathrm{ref}}\\) represents synchronized denoising, \\(\\mathcal{L}_{\\mathrm{reward}}\\) involves the reward model, and Frequency denotes the time-aware frequency separation module. Each component significantly enhances performance. Synchronized denoising \\(\\mathcal{L}_{\\mathrm{ref}}\\) sets a strong baseline, executing scores of 83.2, 72.4, and 32.9 for CLIP-T, CLIP-I, and DINO. The Frequency module improves the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2720" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.359, + 0.182 + ], + "angle": 0, + "content": "
LrefLrewardFreq.CLIP-IDINOCLIP-T
XXX81.371.532.7
✓XX83.272.432.9
X✓✓84.373.734.1
✓✓✓85.974.135.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.192, + 0.34, + 0.22 + ], + "angle": 0, + "content": "Table 3. Impact \\((\\%)\\) of different components. Freq. denotes TA-FFR module." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.089, + 0.605, + 0.182 + ], + "angle": 0, + "content": "
Data ScaleCLIP-IDINOCLIP-T
0.5 M80.169.830.1
1 M84.072.832.3
1.5 M84.973.734.8
2.3 M85.974.135.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.387, + 0.192, + 0.591, + 0.206 + ], + "angle": 0, + "content": "Table 4. Impact (%) of data scale." + }, + { + "type": "table", + "bbox": [ + 0.622, + 0.089, + 0.906, + 0.182 + ], + "angle": 0, + "content": "
ModelCLIP-IDINOCLIP-T
DeepLabv3-MBv284.172.734.6
FCN-R10184.372.934.6
UperNet-R5085.473.834.7
DINOv285.974.135.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.62, + 0.192, + 0.871, + 0.218 + ], + "angle": 0, + "content": "Table 5. Impact (%) of different reward models." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.236, + 0.482, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.421 + ], + "angle": 0, + "content": "Figure 8. The visualization of various ablation modules indicates that each proposed component within our methodology contributes positively to the overall results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.43, + 0.483, + 0.58 + ], + "angle": 0, + "content": "performance across metrics (e.g., CLIP-I improves from 71.5 to 72.4), effectively enhancing temporal coherence and structural detail. The reward model \\(\\mathcal{L}_{\\text {reward }}\\) further boosts results by an average of 1.3 points (e.g., CLIP-I increases from 72.4 to 73.7), capturing nuanced information critical for quality restoration. With all components combined, we achieve the highest scores, demonstrating the complementary strengths of each component. Fig. 8 further examines the effect of each component, demonstrating the guidance of each component." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Reward Model. Following [21], we examine the effectiveness of different reward models, as shown in Tab. 5. DINOv2 [2] achieves the highest scores across all metrics, demonstrating superior performance in text relevance and image fidelity. FCN [25] with ResNet-101 [13] backbone(FCN-R101), DeepLabv3 [3] with MobileNetv2 [37] (DeepLabv3-MBv2) also perform well, though slightly lower, confirming that the reward model enhances feature retention and alignment across architectures, with DINOv2 as the optimal choice." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.258, + 0.758 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Feature Map Visualization. Figure 9 visualizes the cross-attention map between reference and latent features, aggregated across all layers to a uniform scale and averaging. Distinct reference images produce unique feature maps, as shown in the top row. Our method demonstrates consistent object perception across timesteps, a capability lacking in previous ReferenceNet features. Additionally, the incorporation of frequency loss and the reward model enables our approach to capture finer details, resulting in enhanced im" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.236, + 0.851, + 0.252 + ], + "angle": 0, + "content": "age quality and more precise object representation." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.264, + 0.903, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.445, + 0.906, + 0.473 + ], + "angle": 0, + "content": "Figure 9. Cross-Attention Feature Maps. Visualization of the attention maps between reference features and latent features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.635 + ], + "angle": 0, + "content": "Training Data. Tab. 4 demonstrates the impact of training data size on model performance. As the amount of data increases, we observe consistent improvements across all evaluation metrics (CLIP-T, CLIP-I, and DINO). Specifically, using 0.5M data samples yields moderate results, but expanding to 1M and 1.5M significantly enhances model accuracy, indicating the value of larger datasets in refining feature representation. The best performance is achieved with 2.3M samples, highlighting that increased training data directly contributes to more robust and detailed outputs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.64, + 0.665 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this study, we propose TFCustom, a novel model framework that emphasizes the importance of reference image features across different temporal steps and frequency levels. Our approach uses synchronized ReferenceNet to extract these features while optimizing noise injection and denoising processes. Additionally, we introduce a time-aware frequency refinement module that employs high- and low-frequency filters, along with time embeddings, to dynamically adjust the intensity of reference feature injection. To improve alignment between reference objects and generated images, we propose a reward model loss that strengthens their correspondence. Experimental results demonstrate the effectiveness of our framework, achieving state-of-the-art performance and significant improvements in texture and textual detail generation compared to existing methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2721" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2:3, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. CoRR, abs/2307.09481, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[5] Xi Chen, Zhiheng Liu, Mengting Chen, Yutong Feng, Yu Liu, Yujun Shen, and Hengshuang Zhao. Livephoto: Real image animation with text-guided motion control, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.483, + 0.414 + ], + "angle": 0, + "content": "[6] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for authentic virtual try-on in the wild, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[7] Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. Adair: Adaptive all-in-one image restoration via frequency mining and modulation. arXiv preprint arXiv:2403.14614, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.616, + 0.483, + 0.67 + ], + "angle": 0, + "content": "[10] Xiang Gao and Jiaying Liu. Fbsdiff: Plug-and-play frequency band substitution of diffusion features for highly controllable text-driven image translation. In ACM MM, pages 4101-4109, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[11] Xiang Gao, Zhengbo Xu, Junhan Zhao, and Jiaying Liu. Frequency-controlled diffusion model for versatile text-guided image-to-image translation. In AAAI, pages 1824-1832, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.483, + 0.784 + ], + "angle": 0, + "content": "[12] Yuming Gu, You Xie, Hongyi Xu, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.787, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive mod" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "els for fine-grained text-to-image synthesis. In AAAI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[16] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[17] Zehuan Huang, Hongxing Fan, Lipeng Wang, and Lu Sheng. From parts to whole: A unified reference framework for controllable human image generation, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[18] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pages 1931-1941, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[19] Bozheng Li, Mushui Liu, Gaoang Wang, and Yunlong Yu. Frame order matters: A temporal sequence-aware model for few-shot action recognition. In AAAI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[20] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[21] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In ECCV, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[22] Chengxu Liu, Xuan Wang, Shuai Li, Yuzhi Wang, and Xueming Qian. Fsi: Frequency and spatial interactive learning for image restoration in under-display cameras. In ICCV, pages 12537-12546, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[23] Mushui Liu, Yuhang Ma, Yang Zhen, Jun Dan, Yunlong Yu, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. In AAAI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[24] Mushui Liu, Fangtai Wu, Bozheng Li, Ziqian Lu, Yunlong Yu, and Xi Li. Envisioning class entity reasoning by large language models for few-shot learning. In AAAI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[25] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[26] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[27] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In AAAI, pages 1905–1913, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[28] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296–4304, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[29] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[30] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. lambda-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4172-4182, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2722" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[36] Donghun Ryou, Inju Ha, Hyewon Yoo, Dongwan Kim, and Bohyung Han. Robust image denoising through adversarial frequency mixup. In CVPR, pages 2723-2732, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[37] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[38] D She, Mushui Liu, Jingxuan Pang, Jin Wang, Zhen Yang, Wanggui He, Guanghao Zhang, Yi Wang, Qihan Huang, Haobin Tang, et al. Customvideox: 3d reference attention driven dynamic adaptation for zero-shot customized video diffusion transformers. arXiv preprint arXiv:2502.06527, 2025.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[39] Fei Shen and Jinhui Tang. Imagpose: A unified conditional framework for pose-guided person generation. In NeurIPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[40] Fei Shen, Hu Ye, Jun Zhang, Cong Wang, Xiao Han, and Wei Yang. Advancing pose-guided image synthesis with progressive conditional diffusion models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[41] Fei Shen, Xin Jiang, Xin He, Hu Ye, Cong Wang, Xiaoyu Du, Zechao Li, and Jinghui Tang. Imagdressing-v1: Customizable virtual dressing. In AAAI, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[42] Fei Shen, Cong Wang, Junyao Gao, Qin Guo, Jisheng Dang, Jinhui Tang, and Tat-Seng Chua. Long-term talkingface generation via motion-prior conditional diffusion model. arXiv preprint arXiv:2502.09533, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[43] Fei Shen, Hu Ye, Sibo Liu, Jun Zhang, Cong Wang, Xiao Han, and Wei Yang. Boosting consistency in story visualization with rich-contextual conditional diffusion models. In AAAI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimen" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "sional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[47] Rui Wang, Hailong Guo, Jiaming Liu, Huaxia Li, Haibo Zhao, Xu Tang, Yao Hu, Hao Tang, and Peipei Li. Stablegar-ment: Garment-centric generation via stable diffusion, 2024. 2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[48] X. Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[49] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15897-15907, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[50] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[51] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[52] Zhen Yang, Guibao Shen, Liang Hou, Mushui Liu, Luozhou Wang, Xin Tao, Pengfei Wan, Di Zhang, and YingCong Chen. Rectifiedhr: Enable efficient high-resolution image generation via energy rectification. arXiv preprint arXiv:2503.02537, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[54] Jiacheng Ying, Mushui Liu, Zhe Wu, Running Zhang, Zhu Yu, Siming Fu, Si-Yuan Cao, Chao Wu, Yunlong Yu, and Hui-Liang Shen. Restorerid: Towards tuning-free face restoration with id preservation. arXiv preprint arXiv:2411.14125, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[55] Yu Zeng, Vishal M. Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In CVPR 2024, pages 6786-6795, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3813-3824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[57] Shilong Zhang, Lianghua Huang, Xi Chen, Yifei Zhang, ZhiFan Wu, Yutong Feng, Wei Wang, Yujun Shen, Yu Liu, and Ping Luo. Flashface: Human image personalization with high-fidelity identity preservation, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[58] Yuxuan Zhang, Jiaming Liu, Yiren Song, Rui Wang, Hao Tang, Jinpeng Yu, Huaxia Li, Xu Tang, Yao Hu, Han Pan, and Zhongliang Jing. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR, pages 8069–8078, 2024. 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "2723" + } + ] +] \ No newline at end of file diff --git a/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_origin.pdf b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2325ba7c327596dd99b2a7f1b1ebe79fe1313596 --- /dev/null +++ b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/1a56ac6e-2cb9-45e4-af1e-434f2df3991f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcdbd3baae545d348fd361afda29d71445d268af03c082a1495b6104c3c3cc69 +size 6070650 diff --git a/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/full.md b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6dd28619edf64046d3585e977eb2b7cb6dfeedee --- /dev/null +++ b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/full.md @@ -0,0 +1,332 @@ +# TFCustom: Customized Image Generation with Time-Aware Frequency Feature Guidance + +Mushui Liu $^{1*}$ + +Dong She $^{2*}$ + +Jingxuan Pang1 + +Qihan Huang + +Wanggui $\mathrm{He}^3$ + +Yuanlei Hou + +Siming $\mathrm{Fu}^{1*^{\dagger}}$ + +1Zhejiang University + +$^{2}$ University of Science and Technology of China + +3Alibaba Group + +4 Middle East Centre, RIPED + +![](images/5e91af3c05738d38aed746639800e5c8f4d99887f7f080396d5bc09a0abf66b9.jpg) +Single-Subject Driven Generation + +![](images/8cfcd605e8f120ce513bf1026caf78b0d8718f0a06fd6a6aa95f82fe86a5cfc2.jpg) +Multiple-Subject Driven Generation +Figure 1. Our proposed TFCustom demonstrates capabilities in single-subject and multi-subject driven generation tasks. + +# Abstract + +Subject-driven image personalization has seen notable advancements, especially with the ReferenceNet paradigm, which excels in integrating reference image features for creative and commercial applications. However, current ReferenceNet implementations mainly function as latent-level feature extractors, limiting their potential. This restricts + +the delivery of suitable features to the denoising backbone across timesteps, resulting in suboptimal image consistency. In this paper, we revisit reference feature extraction and propose TFCustom, a framework that focuses on reference image features at different temporal and frequency levels. We introduce synchronized ReferenceNet to extract reference features while optimizing noise injection and denoising. We also propose a time-aware frequency refinement module that uses high- and low-frequency filters with time embeddings to adaptively select reference feature injection. Additionally, we introduce a reward-based loss to improve + +the similarity between reference objects and generated images. Experimental results show that TFCustom outperforms existing methods in single-object and multi-object reference generation, with significant improvements in textual details. + +# 1. Introduction + +With the development of deep learning [14, 19, 24, 46, 52] and text-to-image models [8, 15, 23, 32, 34], customization generation which aims to synthesize text-driven images for given subjects, has recently revolutionized content creation. It has seen a shift from test-time fine-tuning methods (e.g., DreamBooth [35], Custom Diffusion [18]) to fine-tuning-free techniques (e.g., IP-Adapter [53], Subject-Diffusion [26], MS-Diffusion [48]). The latter approaches remove the need for fine-tuning during inference, significantly reducing computational costs while maintaining flexibility, and making them more scalable and practical for real-world use cases. + +These fine-tuning-free methods generally fall into two categories: image encoder-based [4, 5, 58] feature injection and reference network (ReferenceNet-based) [6, 12, 51] feature injection. Image encoder methods, such as IPAdapter [53] and ELITE [49], utilize pre-trained encoders like CLIP [33] to extract high-level features, which are then injected into the generation process via mechanisms such as decoupled cross-attention. While these methods are computationally efficient and straightforward, they tend to rely on high-level, abstract features, which may not capture the fine-grained details necessary for precise control over reference object characteristics. The increasing demand for greater consistency in reference images has brought significant attention to the concept of ReferenceNet [16] in the field. ReferenceNet-based methods [17, 47, 50, 57] introduce an additional diffusion network within the traditional Diffusion Net denoising process. In this setup, reference images are encoded by VAE and then used as inputs for ReferenceNet. This novel architecture enables the correlation of features across layers in both diffusion networks, which helps in making better use of the reference images. + +Although ReferenceNet-based methodologies have made significant advancements, several challenges remain. Currently, ReferenceNet predominantly operates as a latent-level feature extractor, thereby not fully realizing its potential. While ReferenceNet maintains a consistent intensity of feature conditioning over time, the denoising backbone requires the injection of varying intensities and granularities of reference information at different timesteps. Specifically, during the earlier stages of denoising, the backbone tends to emphasize contour information, whereas, in the later stages, it focuses more on the refinement of details [45]. This limitation hinders the ability + +to provide the denoising backbone with the appropriate features necessary for optimal performance across varying timesteps, ultimately leading to subpar image consistency. + +In this work, we introduce TFCustom, a novel finetuning-free personalized generation framework that incorporates temporal awareness for reference feature injection. Unlike previous approaches, TFCustom considers the alignment of reference image features with generated image features across different time steps of the diffusion process. Specifically, we first apply noise to the reference image at multiple time steps, with ReferenceNet extracting features at various scales and granularities. The training of ReferenceNet follows a diffusion loss function, ensuring proper alignment with the denoising process. Additionally, we introduce a time-aware frequency module that enables the controlled injection of high- and low-frequency reference features. This module allows low-frequency features to define the primary subject at early time steps, while high-frequency features are progressively injected to refine finer image details. To further improve the model's ability to handle multi-object reference generation, we propose a novel reward-based loss function that enhances the alignment between the reference and generated objects. This loss function computes object ID similarity during the denoising process, minimizing conflicts between reference objects and improving the overall coherence of multi-object scenes. The experimental results demonstrate our method consistently outperforms the state-of-the-art approaches on all the benchmarks. + +Overall, our contributions are as follows: + +- We propose a novel customized image generation framework, which includes synchronized noise injection and denoising within ReferenceNet, along with a time-aware frequency feature refinement module for extracting fine-grained image details. +- We propose a reward-based loss function that alleviates multi-object conflicts and enhances object identity consistency in generated images. +- Our model achieves state-of-the-art performance in single-object reference generation on DreamBooth and MS-Bench datasets. Additionally, we demonstrate superior qualitative results in terms of texture and textual detail generation compared to existing methods. + +# 2. Related Work + +Subject Driven Text-to-Image Generation. Despite significant advancements in image generation [1, 8, 15, 23, 34], the field still falls short of meeting the requirements for personalized generation [38-40, 42, 48, 54]. This gap has consequently drawn significant attention to subject-driven text-to-image generation methods [4, 26, 43]. Textual Inversion [9] and DreamBooth [35] use a textual identifier to bind the visual concept through fine-tuning model textual + +embedding or UNet model. CustomDiffusion [18] further extends multi-subject generation by multiple identifiers. IP-Adapter [53] injects the visual concept by decoupling the cross-attention between text and images in a zero-shot manner. JeDi [55] learns the joint distribution of multiple text-image pairs that share a common subject to facilitate personalized generation. Subject-Diffusion [26] trains multiple adapters inside the model with an attention control mechanism for multiple-subject personalization. SSR-Encoder [58] and MS-Diffusion [48] develop the query architecture to encode the subject images via mask or layout control for multiple-subject personalization. This paper explores the zero-shot multiple-objects driven generation via synchronized ReferenceNet optimization. + +ReferenceNet Based Personalized Generation. ReferenceNet [16, 41, 57] has shown remarkable capabilities in maintaining image consistency for personalized generation tasks. AnimateAnyone [16] first utilizes ReferenceNet for personalized human animation, employing a pre-trained generative model to extract reference features. FlashFace [57] extended this approach by combining ReferenceNet with ControlNet [56], enabling identity-preserving portrait generation that accommodates multiple reference images. Part2Whole [17] uses ReferenceNet to handle multi-part human body generation, while OOTD [50] and Stable-Garment [47] integrate it into virtual try-on tasks, aligning clothing features precisely with the target human body. These methods highlight ReferenceNet's versatility in image and video generation. To further improve performance, we introduce a synchronized ReferenceNet that refines reference features at each time step, enabling more precise feature injection and enhancing generation accuracy. + +Frequency Information in Image Generation. Frequency information plays a crucial role in image restoration tasks [7, 22, 27, 36]. Recent advancements in text-to-image generation have leveraged frequency-based conditions, like canny and HED images, to control generation, as seen in ControlNet [56] and T2I-Adapter [28]. FCDiffusion [11] and FBSdiff [10] employ DCT-transformed image features to enhance texture detail in image-to-image generation. In this work, our method adaptively injects reference image features via learnable high-pass and low-pass filters in a time-aware fashion, enriching each denoising step with frequency-specific information for more refined generation. + +# 3. Method + +Our proposed TFCustom addresses these issues through three components: Sec. 3.2 Synchronized Reference Guidance, Sec. 3.3 Time-Aware Frequency Feature Refinement, and Sec. 3.4 Reward Model Optimization, as illustrated in Fig. 2. First, the reference image is incrementally noised over $T$ steps before feature extraction via the ReferenceNet, aligning it with the denoising network's noisy + +latent features. Next, a time-aware frequency module injects frequency-specific information from the reference image into the noisy latent representation. Finally, a reward model optimizes this process, enhancing the quality and consistency of the generated reference-based images. + +# 3.1. Preliminaries + +Diffusion models. Denoising Diffusion Probabilistic Models (DDPM) are a class of generative models that leverage a diffusion process to generate data. By progressively adding Gaussian noise to the data over a specified number of time steps, a sequence of increasingly noisy samples is produced. The forward process can be described as follows: + +$$ +q \left(x _ {T} \mid x _ {0}\right) = \mathcal {N} \left(x _ {T}; \sqrt {\alpha_ {T}} x _ {0}, (1 - \alpha_ {T}) \mathbf {I}\right) \tag {1} +$$ + +where $\alpha_{t} = \prod_{s=1}^{t}(1 - \beta_{s})$ . Here, $\beta_{t}$ represents a variance schedule that regulates the noise level at each time step $t$ , and $N$ denotes a Gaussian distribution. In the reverse diffusion process, the model learns to denoise the noisy samples step by step, effectively generating new data from pure noise: + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) = \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, t\right), \sigma_ {\theta} ^ {2} (t) \mathbf {I}\right) \tag {2} +$$ + +where $\mu_{\theta}$ and $\sigma_{\theta}$ are learned parameters that depend on the current noisy sample $x_{t}$ and the time step $t$ . The optimization goal is the variational lower bound on the log-likelihood of the data, which can be simplified to: + +$$ +L = \mathbb {E} _ {t, x _ {0}, \epsilon} \left[ \| \epsilon - \epsilon_ {\theta} (x _ {t}, t) \| ^ {2} \right] \tag {3} +$$ + +# 3.2. Synchronized ReferenceNet Guidance + +The synchronized ReferenceNet is initially transferred from the denoising network. During training, at each time step $t$ , we apply a noise process to the reference image $x_{\mathrm{ref}}$ , adding noise for $t$ steps, resulting in the noise reference $x_{\mathrm{ref}}^{t}$ as formulated in Eq. (1). This step ensures that the reference features are temporally aligned with the noisy latent features processed by the denoising network. Once the noised reference $x_{\mathrm{ref}}^{t}$ is generated, it passes through the synchronized ReferenceNet, which extracts the relevant features $\mathbf{F}_{\mathrm{ref}}$ and estimates the corresponding noise pattern. The extracted reference features and predicted noise are then optimized according to the objective $\mathcal{L}_{\mathrm{diff}}^{\mathrm{ref}}$ defined in Eq. (3). This objective encourages the synchronized ReferenceNet to provide refined guidance that is consistent with the noisy latent representations, enabling the denoising network to more effectively incorporate reference information for improved reconstruction quality. + +Given that the object information hierarchically learned by the denoising network varies significantly across different timesteps, the early timesteps tend to focus on learning macro-level information such as contours, while the + +![](images/aba4401df3cc440eee8b9b53f84da2a06be24a94f4a47ae75ac008714405975c.jpg) +(a) Overall Framework + +![](images/daf168b010f9eff67fd0d14eecfb9a12313c2dfce082135fe173d71c4b50b574.jpg) +(b) Time-Aware Frequency Feature Refinement +Figure 2. (a) Overall Framework. Our TFCustom includes three key components: Synchronized Reference Net, Time-Aware Frequency Feature Refinement (TA-FFR), and Reward Model Optimization for subject-driven generation. (b) TA-FFR Module. This module extracts different frequency features in a time-aware manner, adaptively injecting reference features for enhanced generation quality. + +later timesteps progressively refine texture details. However, the synchronized timestep information obtained from $\mathcal{L}_{\mathrm{diff}}^{\mathrm{ref}}$ encompasses the entire reference object's information at timestep $t$ , which can be overly complex for the denoising network. To address this issue, we have implemented a Time-Aware Frequency Feature Refinement module (which will be described detailed in Sec. 3.3) to refine the reference information, thereby enhancing the learning process of the denoising network, as outlined in Eq. 7. + +With the enhanced reference features $\mathbf{F}_{\mathrm{enh}}$ , we inject them using a Shared Self-Attention mechanism [16, 17]. Specifically, in each self-attention layer, given the features of $N$ reference images $\mathbf{F}_{\mathrm{enh}}^{1:N}$ and the feature map $\mathbf{F}^0$ from the denoising net, we concatenate these feature maps side-by-side as input to the self-attention layer, represented as $[\mathbf{F}_{\mathrm{enh}}^0 | \mathbf{F}_{\mathrm{enh}}^1| \dots | \mathbf{F}_{\mathrm{enh}}^N]$ . This setup enables each location $s$ on $\mathbf{F}^0$ to attend to all locations on itself and the reference feature maps, computed as follows: + +$$ +\begin{array}{l} \tilde {\mathbf {F}} _ {s} ^ {0} = \operatorname {S o f t M a x} \left(\frac {\mathbf {F} _ {s} ^ {0} W _ {q} \cdot \left([ \mathbf {F} _ {\mathrm {e n h}} ^ {0} , \mathbf {F} _ {\mathrm {e n h}} ^ {1} , \dots , \mathbf {F} _ {\mathrm {e n h}} ^ {N} ]\right) ^ {\top} W _ {k}}{\sqrt {d}}\right) \\ \cdot \left(\left[ \mathbf {F} _ {\text {e n h}} ^ {0}, \mathbf {F} _ {\text {e n h}} ^ {1}, \dots , \mathbf {F} _ {\text {e n h}} ^ {N} \right]\right) W _ {v}, \tag {4} \\ \end{array} +$$ + +where $W_{q}, W_{k}, W_{v}$ are corresponding projection weight matrices. And $d$ is the dimensionality of the key vectors. + +# 3.3. Time-Aware Frequency Feature Refinement + +Inspired by previous work [45], which illustrates that neural network learns hierarchically—starting with low-frequency features that define shape and structure, followed by high-frequency details for realism, we propose that reference image feature injection should follow a similar coarse-to-fine + +approach. This would better align with the denoising stages of diffusion. + +To achieve this, we introduce a Time-Aware Frequency Feature Refinement (TA-FFR), as shown in Fig. 2. This module separates reference image features into high-frequency and low-frequency components, each processed independently to capture unique information. The high-frequency path applies a Krisch operator for edge and detail preservation, while the low-frequency path uses a Gaussian operator to retain foundational structures. These operators are initialized as $5 \times 5$ convolution kernels $\mathbf{H}_{\mathrm{conv}}$ is Krisch operator, and $\mathbf{L}_{\mathrm{conv}}$ is Gaussian operator. Both operators are scaled by learnable parameters to adjust the frequency extraction dynamically: + +$$ +\mathbf {H} _ {\text {o u t}} = \operatorname {C o n v} _ {\mathbf {H}} \left(\mathbf {F} _ {\text {r e f}}\right) \cdot \mathbf {W} _ {\mathrm {H}}, \quad \mathbf {L} _ {\text {o u t}} = \operatorname {C o n v} _ {\mathbf {L}} \left(\mathbf {F} _ {\text {r e f}}\right) \cdot \mathbf {W} _ {\mathrm {L}}, \tag {5} +$$ + +where $\mathbf{F}_{\mathrm{ref}}$ denotes reference image features, and $\mathbf{W}_{\mathrm{H}}$ and $\mathbf{W}_{\mathrm{L}}$ are scaling factors for high and low frequencies. + +Each frequency component then undergoes Time-Aware Attention (TA-Attention) and Time-Aware Feed-Forward Network (TA-FFN), which incorporate time-step information using a time embedding $\mathbf{t}_{\mathrm{emb}}$ via Adaptive Layer Normalization (AdaLN) proposed in DiT [31]: + +$$ +\begin{array}{l} \mathbf {F} _ {\text {h i g h}} = \mathrm {T A - F F N} (\mathrm {T A - A t t e n t i o n} (\mathrm {A d a L N} (\mathbf {H} _ {\text {o u t}}, \mathbf {t} _ {\text {e m b}}))) \\ \mathbf {F} _ {\text {l o w}} = \mathrm {T A - F F N} (\mathrm {T A - A t t e n t i o n} (\mathrm {A d a L N} (\mathbf {L} _ {\text {o u t}}, \mathbf {t} _ {\text {e m b}}))), \end{array} \tag {6} +$$ + +where $\mathbf{F}_{\mathrm{high}}$ and $\mathbf{F}_{\mathrm{low}}$ represent high-frequency and low-frequency features. Then the final enhanced feature is obtained by combining these outputs: + +$$ +\mathbf {F} _ {\text {e n h}} = \mathbf {F} _ {\text {h i g h}} + \mathbf {F} _ {\text {l o w}}. \tag {7} +$$ + +![](images/6da9dee83c5f758366ecffe6e56115382415310607805ccd8f498c10f8fd1b43.jpg) +Figure 3. Reward Model Optimization Process. "G" and "D" indicate the network's forward pass and backward optimization. + +This design ensures a coarse-to-fine feature injection, enhancing the model's ability to integrate temporal and spatial details, thereby improving the quality of the generated images. + +# 3.4. Reward Model Optimization + +Our model can simulate the noise levels of both the reference image and the target image. Based on the following equation, we can perform a one-step prediction of the source image, allowing us to naturally determine the reward model score based on the noise level by: + +$$ +x _ {0} ^ {\prime} \approx \frac {x _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta} \left(x _ {t} , x _ {\text {r e f}} , c _ {t} , t\right)}{\sqrt {\alpha_ {t}}}. \tag {8} +$$ + +where $c_{t}$ refers to the textual condition. This equation's calculation error increases with timestep $t$ . Therefore, during the training phase, we only apply the reward loss for timestep $t < T_0$ : + +$$ +\mathcal {L} _ {\text {r e w a r d}} = \operatorname {D i s} \left(\mathrm {D I N O} _ {v 2} \left(x _ {0}\right), \mathrm {D I N O} _ {v 2} \left(x _ {0} ^ {\prime}\right)\right) \tag {9} +$$ + +where Dis is cosine similarity metric and $\mathrm{DINO}_{v2}$ is the DINO v2 model [2] to compute the similarity between the target domain's denoised output and the features generated by the reference net. It is worth noticing that in our synchronized noise injection scheme, the output of the reference net corresponds to the noise level at timestep $t$ , which provides meaningful information. Thus, we take advantage of this synchronized noise injection feature and introduce a diffusion loss, $L_{\mathrm{ref}}$ , for the reference net, which supervises the feature extraction capability of the reference branch. The overall loss function can be formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d i f f}} ^ {\text {d e n o i s i n g}} + \lambda_ {1} \mathcal {L} _ {\text {d i f f}} ^ {\text {r e f}} + \lambda_ {2} \mathcal {L} _ {\text {r e w a r d}}, \tag {10} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ is hyper-parameters to balance. + +# 4. Experiments + +# 4.1. Experiment Setup + +Datasets. The training dataset comprises 2.3 million (M) in-house image pairs. We assess the performance on + +
MethodTypeCLIP-IDINOCLIP-T
BLIP-Diffusion [20]No-FT77.959.430.0
IP-Adapter [53]No-FT81.061.329.2
Emu2 [44]No-FT76.556.327.3
Kosmos-G [29]No-FT82.261.825.0
λ-ECLIPSE [30]No-FT78.361.330.7
SSR-Encoder [58]No-FT82.161.230.8
MS-Diffusion [48]No-FT79.267.132.1
TFCustomNo-FT84.171.433.9
Textual Inversion [9]FT78.056.925.5
DreamBooth [35]FT80.366.830.5
BLIP-Diffusion [20]FT80.567.030.2
λ-ECLIPSE [30]FT79.668.230.4
MS-Diffusion [48]FT80.570.231.3
TFCustomFT85.974.135.1
+ +Table 1. Quantitative comparison for single-subject models. FT denotes fine-tuned models on DreamBench [35]. The best and second-best results are shown in bold and underline, respectively. + +single-subject tasks using the DreamBench dataset [35] and on multi-subject tasks using the MS-Bench dataset [48]. DreamBench encompasses 30 subjects and 25 prompts, while MS-Bench includes 40 subjects and 1148 combinations, with each combination being associated with up to 6 prompts, resulting in a total of 4488 unique test samples. Furthermore, we collect an additional 20 real-world cases featuring complex textures, characters, and patterns for further qualitative evaluation. + +Implementation Details. During training, we employ the AdamW optimizer with a learning rate of 1e-4 and train the model for 100K steps on 32 NVIDIA A100 80GB GPUs, using a batch size of 4 per GPU. The $\lambda_{1}$ and $\lambda_{2}$ in Eq. (10) are set to 0.4 and 0.6. To enable classifier-free guidance, we apply a probability of 0.05 to independently drop text and images, and 0.05 to drop both simultaneously. For inference, we adapt the DDIM sampler with 50 steps and set the guidance scale to 7.5. + +Evaluation Metrics. We evaluate performance based on both image and text fidelity metrics. For image fidelity, we compute the cosine similarity between generated images and subject images within the CLIP (denoted as CLIP-I) and DINO feature spaces. Text fidelity is measured by calculating the cosine similarity between generated images and text prompts in the CLIP space (referred to as CLIP-T). In the context of multi-subject personalization, average fidelity scores may not adequately reflect the representation of individual subjects. Therefore, we also employ the multi-subject DINO (M-DINO) metric [48] to assess whether each subject was accurately recreated. + +Competitors. We compare our model with BLIP-Diffusion [20], IP-Adapter [53], Kosmos-G [29], Emu2 [44], $\lambda$ -ECLIPSE [30], SSR-Encoder [58], and MS-Diffusion [48] for single concept personalization, and with MS-Diffusion [48], SSR-Encoder [58] and $\lambda$ -ECLIPSE [30] for multi- + +![](images/1a3d9c0a701ce79bac4229a27066b907a91fe1c3258c227750f24ea313020ae3.jpg) +Figure 4. Qualitative comparison with existing state-of-the-art customization methods. TFCustom produces higher-quality results with better subject similarity and text controllability compared to existing methods. + +subject personalization. + +# 4.2. Performance Comparison + +Quantitative Comparison of Single-Subject. The results are presented in Tab. 1. In the zero-shot evaluation (without fine-tuning on DreamBench), our proposed TFCustom achieves the highest scores across all metrics, with CLIP-I, DINO, and CLIP-T values of 84.1, 71.4, and 33.9, respectively. These results surpass the second-best methods by margins of 1.9 in CLIP-I (compared to 82.2 by Kosmos-G), 4.3 in DINO (compared to 67.1 by MS-Diffusion), and 1.8 in CLIP-T (compared to 32.1 by MS-Diffusion). Notably, our method even surpasses models fine-tuned at test time, showcasing robust zero-shot capabilities. TFCustom's superior performance can be attributed to the synchronized reference network, which provides multi-scale features for highly detailed image generation, and the time-aware frequency module, which harmonizes semantic information with reference image features. Additionally, the reward model enhances object similarity by leveraging referential cues. When fine-tuned on DreamBench, TFCustom shows further improvements, achieving top scores in CLIP-I, DINO, and CLIP-T metrics with values of 85.9, 74.1, and 35.1, respectively. This highlights TFCustom's adaptability in personalization scenarios, demonstrating strong transferability across various contexts. + +Quantitative Comparison of Multi-Subject. We also com + +
MethodCLIP-IDINOM-DINOCLIP-T
λ-ECLIPSE [30]72.441.99.431.6
SSR-Encoder [58]72.542.510.730.3
MS-Diffusion [48]69.842.510.834.1
TFCustom74.343.812.936.9
+ +Table 2. Quantitative comparison for multi-subject models. The best and second-best results are shown in **bold** and **underline**, respectively. + +pare multi-subject personalization performance in Tab. 2. The results show that TFCustom achieves state-of-the-art performance across CLIP-I, DINO, M-DINO, and CLIP-T metrics, with scores of 74.3, 43.8, 12.9, and 36.9, respectively, significantly surpassing other reference image generation methods. Compared to MS-Diffusion, TFCustom demonstrates superior text-following capability, as evidenced by its higher CLIP-T scores, outperforming methods like MS-Diffusion and SSR-Encoder. Additionally, TFCustom excels in the M-DINO metric, highlighting its ability to preserve multi-object details in generated images. + +Qualitative Results. Fig. 4 presents qualitative results from the DreamBench dataset, demonstrating our model's remarkable capacity for generating high-fidelity single-subject images with excellent detail preservation. Furthermore, Fig. 5 illustrates qualitative outcomes from the MS-Bench dataset. Our approach adeptly maintains the textual + +![](images/73b07ca0f3e9828453e715165cdb21d75b6bc9e2352b9bd439a40c8e6865e820.jpg) +Figure 5. Qualitative comparison of multiple-subject customization. TFCustom achieves enhanced text controllability and subject fidelity compared to existing methods. Customized words are highlighted. + +![](images/cb6189be6e80ccdb606571c4f142db0f8b8395649ad793ee4a554c2631625881.jpg) + +![](images/029a911d6c107bf7d0ef98d912ddf90ea56dfb771895d36f410c3f2e7c2f772b.jpg) + +![](images/e7855f3ac4fdcc4665fe333aac8b565c9447429107cf7fe32c763585eda93be7.jpg) + +![](images/dbfc8cb3305f785a86b4522e1d20a0ee49a16368b5840c736b133562ccb82480.jpg) + +![](images/08ce89588935041d5eb88e49f4b0a254ff032f750dfa1f3f56dba3fc9b13802e.jpg) + +![](images/38a68a7550eec924bf69f25bf1a1cc6cbc323ac5bd999bbb236a662102882eba.jpg) +Figure 6. Visualizations in customized human generation. + +interactions among multiple subjects within the generated images, ensuring that each subject remains distinct and easily recognizable. This effectively mitigates issues related to subject omission or conflicts. + +Complex Scenarios. We evaluate the customized human generation in Fig. 6. The result shows that TFCustom is still capable of generating high-quality results for customized human generation. Additionally, we evaluate our model in real-world generation scenarios, as shown in Fig. 7. Our model is capable of generating images that accurately capture the fine-grained details of the reference images, including text, textures, and complex patterns. These showcase its strong capability in reference image feature transfer and high-quality image generation. + +![](images/f02264ee28ba1057f86c5f483d8e8f3245fdac29beac3d66501323ecc25d9ad0.jpg) +Figure 7. Visualizations in complex scenarios with intricate patterns and text. + +# 4.3. Ablation Study + +Model Design. We conduct ablation experiments to assess each component of our method, as shown in Tab. 3. $\mathcal{L}_{\mathrm{ref}}$ represents synchronized denoising, $\mathcal{L}_{\mathrm{reward}}$ involves the reward model, and Frequency denotes the time-aware frequency separation module. Each component significantly enhances performance. Synchronized denoising $\mathcal{L}_{\mathrm{ref}}$ sets a strong baseline, executing scores of 83.2, 72.4, and 32.9 for CLIP-T, CLIP-I, and DINO. The Frequency module improves the + +
LrefLrewardFreq.CLIP-IDINOCLIP-T
XXX81.371.532.7
✓XX83.272.432.9
X✓✓84.373.734.1
✓✓✓85.974.135.1
+ +Table 3. Impact $(\%)$ of different components. Freq. denotes TA-FFR module. + +
Data ScaleCLIP-IDINOCLIP-T
0.5 M80.169.830.1
1 M84.072.832.3
1.5 M84.973.734.8
2.3 M85.974.135.1
+ +Table 4. Impact (%) of data scale. + +
ModelCLIP-IDINOCLIP-T
DeepLabv3-MBv284.172.734.6
FCN-R10184.372.934.6
UperNet-R5085.473.834.7
DINOv285.974.135.1
+ +Table 5. Impact (%) of different reward models. + +![](images/1dda87ed4019b9a602ad08633b78d9d46d46942db787c7e519fa4710ecf523db.jpg) +Figure 8. The visualization of various ablation modules indicates that each proposed component within our methodology contributes positively to the overall results. + +performance across metrics (e.g., CLIP-I improves from 71.5 to 72.4), effectively enhancing temporal coherence and structural detail. The reward model $\mathcal{L}_{\text {reward }}$ further boosts results by an average of 1.3 points (e.g., CLIP-I increases from 72.4 to 73.7), capturing nuanced information critical for quality restoration. With all components combined, we achieve the highest scores, demonstrating the complementary strengths of each component. Fig. 8 further examines the effect of each component, demonstrating the guidance of each component. + +Reward Model. Following [21], we examine the effectiveness of different reward models, as shown in Tab. 5. DINOv2 [2] achieves the highest scores across all metrics, demonstrating superior performance in text relevance and image fidelity. FCN [25] with ResNet-101 [13] backbone(FCN-R101), DeepLabv3 [3] with MobileNetv2 [37] (DeepLabv3-MBv2) also perform well, though slightly lower, confirming that the reward model enhances feature retention and alignment across architectures, with DINOv2 as the optimal choice. + +# 4.4. Further Analysis + +Feature Map Visualization. Figure 9 visualizes the cross-attention map between reference and latent features, aggregated across all layers to a uniform scale and averaging. Distinct reference images produce unique feature maps, as shown in the top row. Our method demonstrates consistent object perception across timesteps, a capability lacking in previous ReferenceNet features. Additionally, the incorporation of frequency loss and the reward model enables our approach to capture finer details, resulting in enhanced im + +age quality and more precise object representation. + +![](images/412320aa4e44b559cf3122db14918d6f4b8ac25a391e7cba05dbb4afcfac5d87.jpg) +Figure 9. Cross-Attention Feature Maps. Visualization of the attention maps between reference features and latent features. + +Training Data. Tab. 4 demonstrates the impact of training data size on model performance. As the amount of data increases, we observe consistent improvements across all evaluation metrics (CLIP-T, CLIP-I, and DINO). Specifically, using 0.5M data samples yields moderate results, but expanding to 1M and 1.5M significantly enhances model accuracy, indicating the value of larger datasets in refining feature representation. The best performance is achieved with 2.3M samples, highlighting that increased training data directly contributes to more robust and detailed outputs. + +# 5. Conclusions + +In this study, we propose TFCustom, a novel model framework that emphasizes the importance of reference image features across different temporal steps and frequency levels. Our approach uses synchronized ReferenceNet to extract these features while optimizing noise injection and denoising processes. Additionally, we introduce a time-aware frequency refinement module that employs high- and low-frequency filters, along with time embeddings, to dynamically adjust the intensity of reference feature injection. To improve alignment between reference objects and generated images, we propose a reward model loss that strengthens their correspondence. Experimental results demonstrate the effectiveness of our framework, achieving state-of-the-art performance and significant improvements in texture and textual detail generation compared to existing methods. + +# References + +[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2:3, 2023. 2 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5, 8 +[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation, 2017. 8 +[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. CoRR, abs/2307.09481, 2023. 2 +[5] Xi Chen, Zhiheng Liu, Mengting Chen, Yutong Feng, Yu Liu, Yujun Shen, and Hengshuang Zhao. Livephoto: Real image animation with text-guided motion control, 2023. 2 +[6] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for authentic virtual try-on in the wild, 2024. 2 +[7] Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. Adair: Adaptive all-in-one image restoration via frequency mining and modulation. arXiv preprint arXiv:2403.14614, 2024. 3 +[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2 +[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 5 +[10] Xiang Gao and Jiaying Liu. Fbsdiff: Plug-and-play frequency band substitution of diffusion features for highly controllable text-driven image translation. In ACM MM, pages 4101-4109, 2024. 3 +[11] Xiang Gao, Zhengbo Xu, Junhan Zhao, and Jiaying Liu. Frequency-controlled diffusion model for versatile text-guided image-to-image translation. In AAAI, pages 1824-1832, 2024. 3 +[12] Yuming Gu, You Xie, Hongyi Xu, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis, 2024. 2 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 8 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2 +[15] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive mod + +els for fine-grained text-to-image synthesis. In AAAI, 2025. 2 +[16] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation, 2024. 2, 3, 4 +[17] Zehuan Huang, Hongxing Fan, Lipeng Wang, and Lu Sheng. From parts to whole: A unified reference framework for controllable human image generation, 2024. 2, 3, 4 +[18] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pages 1931-1941, 2023. 2, 3 +[19] Bozheng Li, Mushui Liu, Gaoang Wang, and Yunlong Yu. Frame order matters: A temporal sequence-aware model for few-shot action recognition. In AAAI, 2025. 2 +[20] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 5 +[21] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In ECCV, 2024.8 +[22] Chengxu Liu, Xuan Wang, Shuai Li, Yuzhi Wang, and Xueming Qian. Fsi: Frequency and spatial interactive learning for image restoration in under-display cameras. In ICCV, pages 12537-12546, 2023. 3 +[23] Mushui Liu, Yuhang Ma, Yang Zhen, Jun Dan, Yunlong Yu, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. In AAAI, 2025. 2 +[24] Mushui Liu, Fangtai Wu, Bozheng Li, Ziqian Lu, Yunlong Yu, and Xi Li. Envisioning class entity reasoning by large language models for few-shot learning. In AAAI, 2025. 2 +[25] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation, 2015. 8 +[26] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2, 3 +[27] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In AAAI, pages 1905–1913, 2023. 3 +[28] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296–4304, 2024. 3 +[29] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 5 +[30] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. lambda-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 5, 6 +[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4172-4182, 2023. 4 + +[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952, 2023. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685, 2022. 2 +[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2, 5 +[36] Donghun Ryou, Inju Ha, Hyewon Yoo, Dongwan Kim, and Bohyung Han. Robust image denoising through adversarial frequency mixup. In CVPR, pages 2723-2732, 2024. 3 +[37] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks, 2019. 8 +[38] D She, Mushui Liu, Jingxuan Pang, Jin Wang, Zhen Yang, Wanggui He, Guanghao Zhang, Yi Wang, Qihan Huang, Haobin Tang, et al. Customvideox: 3d reference attention driven dynamic adaptation for zero-shot customized video diffusion transformers. arXiv preprint arXiv:2502.06527, 2025.2 +[39] Fei Shen and Jinhui Tang. Imagpose: A unified conditional framework for pose-guided person generation. In NeurIPS, 2024. +[40] Fei Shen, Hu Ye, Jun Zhang, Cong Wang, Xiao Han, and Wei Yang. Advancing pose-guided image synthesis with progressive conditional diffusion models. In ICLR, 2023. 2 +[41] Fei Shen, Xin Jiang, Xin He, Hu Ye, Cong Wang, Xiaoyu Du, Zechao Li, and Jinghui Tang. Imagdressing-v1: Customizable virtual dressing. In AAAI, 2025. 3 +[42] Fei Shen, Cong Wang, Junyao Gao, Qin Guo, Jisheng Dang, Jinhui Tang, and Tat-Seng Chua. Long-term talkingface generation via motion-prior conditional diffusion model. arXiv preprint arXiv:2502.09533, 2025. 2 +[43] Fei Shen, Hu Ye, Sibo Liu, Jun Zhang, Cong Wang, Xiao Han, and Wei Yang. Boosting consistency in story visualization with rich-contextual conditional diffusion models. In AAAI, 2025. 2 +[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 5 +[45] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimen + +sional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2, 4 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2 +[47] Rui Wang, Hailong Guo, Jiaming Liu, Huaxia Li, Haibo Zhao, Xu Tang, Yao Hu, Hao Tang, and Peipei Li. Stablegar-ment: Garment-centric generation via stable diffusion, 2024. 2,3 +[48] X. Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance, 2024. 2, 3, 5, 6 +[49] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15897-15907, 2023. 2 +[50] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on, 2024. 2, 3 +[51] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model, 2023. 2 +[52] Zhen Yang, Guibao Shen, Liang Hou, Mushui Liu, Luozhou Wang, Xin Tao, Pengfei Wan, Di Zhang, and YingCong Chen. Rectifiedhr: Enable efficient high-resolution image generation via energy rectification. arXiv preprint arXiv:2503.02537, 2025. 2 +[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3, 5 +[54] Jiacheng Ying, Mushui Liu, Zhe Wu, Running Zhang, Zhu Yu, Siming Fu, Si-Yuan Cao, Chao Wu, Yunlong Yu, and Hui-Liang Shen. Restorerid: Towards tuning-free face restoration with id preservation. arXiv preprint arXiv:2411.14125, 2024. 2 +[55] Yu Zeng, Vishal M. Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In CVPR 2024, pages 6786-6795, 2024. 3 +[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3813-3824, 2023. 3 +[57] Shilong Zhang, Lianghua Huang, Xi Chen, Yifei Zhang, ZhiFan Wu, Yutong Feng, Wei Wang, Yujun Shen, Yu Liu, and Ping Luo. Flashface: Human image personalization with high-fidelity identity preservation, 2024. 2, 3 +[58] Yuxuan Zhang, Jiaming Liu, Yiren Song, Rui Wang, Hao Tang, Jinpeng Yu, Huaxia Li, Xu Tang, Yao Hu, Han Pan, and Zhongliang Jing. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR, pages 8069–8078, 2024. 2, 3, 5, 6 \ No newline at end of file diff --git a/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/images.zip b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..947748096923a4d0c4400e132a420ceca9415f29 --- /dev/null +++ b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:439eff1318c499b651603fc19420d385007cd294fb60010493c41e40cd35236e +size 1005937 diff --git a/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/layout.json b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d084dcaa88d6b62f056c810abd56aaf584dab5bb --- /dev/null +++ b/2025/TFCustom_ Customized Image Generation with Time-Aware Frequency Feature Guidance/layout.json @@ -0,0 +1,8427 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 102, + 527, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 527, + 138 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 527, + 138 + ], + "type": "text", + "content": "TFCustom: Customized Image Generation with Time-Aware Frequency Feature Guidance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 160, + 172, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 160, + 172, + 175 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 172, + 175 + ], + "type": "text", + "content": "Mushui Liu" + }, + { + "bbox": [ + 104, + 160, + 172, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 186, + 161, + 244, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 161, + 244, + 175 + ], + "spans": [ + { + "bbox": [ + 186, + 161, + 244, + 175 + ], + "type": "text", + "content": "Dong She" + }, + { + "bbox": [ + 186, + 161, + 244, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 257, + 161, + 333, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 161, + 333, + 175 + ], + "spans": [ + { + "bbox": [ + 257, + 161, + 333, + 175 + ], + "type": "text", + "content": "Jingxuan Pang1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 347, + 161, + 417, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 161, + 417, + 175 + ], + "spans": [ + { + "bbox": [ + 347, + 161, + 417, + 175 + ], + "type": "text", + "content": "Qihan Huang" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 194, + 175, + 259, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 175, + 259, + 190 + ], + "spans": [ + { + "bbox": [ + 194, + 175, + 259, + 190 + ], + "type": "text", + "content": "Wanggui " + }, + { + "bbox": [ + 194, + 175, + 259, + 190 + ], + "type": "inline_equation", + "content": "\\mathrm{He}^3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 273, + 175, + 339, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 175, + 339, + 190 + ], + "spans": [ + { + "bbox": [ + 273, + 175, + 339, + 190 + ], + "type": "text", + "content": "Yuanlei Hou" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 352, + 175, + 417, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 175, + 417, + 190 + ], + "spans": [ + { + "bbox": [ + 352, + 175, + 417, + 190 + ], + "type": "text", + "content": "Siming " + }, + { + "bbox": [ + 352, + 175, + 417, + 190 + ], + "type": "inline_equation", + "content": "\\mathrm{Fu}^{1*^{\\dagger}}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 125, + 191, + 227, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 191, + 227, + 206 + ], + "spans": [ + { + "bbox": [ + 125, + 191, + 227, + 206 + ], + "type": "text", + "content": "1Zhejiang University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 252, + 191, + 486, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 191, + 486, + 206 + ], + "spans": [ + { + "bbox": [ + 252, + 191, + 486, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 252, + 191, + 486, + 206 + ], + "type": "text", + "content": "University of Science and Technology of China" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 193, + 206, + 271, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 206, + 271, + 220 + ], + "spans": [ + { + "bbox": [ + 193, + 206, + 271, + 220 + ], + "type": "text", + "content": "3Alibaba Group" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 206, + 438, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 206, + 438, + 219 + ], + "spans": [ + { + "bbox": [ + 295, + 206, + 438, + 219 + ], + "type": "text", + "content": "4 Middle East Centre, RIPED" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 61, + 247, + 553, + 398 + ], + "blocks": [ + { + "bbox": [ + 255, + 236, + 372, + 246 + ], + "lines": [ + { + "bbox": [ + 255, + 236, + 372, + 246 + ], + "spans": [ + { + "bbox": [ + 255, + 236, + 372, + 246 + ], + "type": "text", + "content": "Single-Subject Driven Generation" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 247, + 553, + 398 + ], + "lines": [ + { + "bbox": [ + 61, + 247, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 553, + 398 + ], + "type": "image", + "image_path": "5e91af3c05738d38aed746639800e5c8f4d99887f7f080396d5bc09a0abf66b9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 76, + 414, + 547, + 539 + ], + "blocks": [ + { + "bbox": [ + 252, + 402, + 380, + 412 + ], + "lines": [ + { + "bbox": [ + 252, + 402, + 380, + 412 + ], + "spans": [ + { + "bbox": [ + 252, + 402, + 380, + 412 + ], + "type": "text", + "content": "Multiple-Subject Driven Generation" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 414, + 547, + 539 + ], + "lines": [ + { + "bbox": [ + 76, + 414, + 547, + 539 + ], + "spans": [ + { + "bbox": [ + 76, + 414, + 547, + 539 + ], + "type": "image", + "image_path": "8cfcd605e8f120ce513bf1026caf78b0d8718f0a06fd6a6aa95f82fe86a5cfc2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 548, + 522, + 560 + ], + "lines": [ + { + "bbox": [ + 87, + 548, + 522, + 560 + ], + "spans": [ + { + "bbox": [ + 87, + 548, + 522, + 560 + ], + "type": "text", + "content": "Figure 1. Our proposed TFCustom demonstrates capabilities in single-subject and multi-subject driven generation tasks." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 152, + 578, + 200, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 578, + 200, + 591 + ], + "spans": [ + { + "bbox": [ + 152, + 578, + 200, + 591 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 54, + 605, + 297, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 297, + 678 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 297, + 678 + ], + "type": "text", + "content": "Subject-driven image personalization has seen notable advancements, especially with the ReferenceNet paradigm, which excels in integrating reference image features for creative and commercial applications. However, current ReferenceNet implementations mainly function as latent-level feature extractors, limiting their potential. This restricts" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 580, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 713 + ], + "type": "text", + "content": "the delivery of suitable features to the denoising backbone across timesteps, resulting in suboptimal image consistency. In this paper, we revisit reference feature extraction and propose TFCustom, a framework that focuses on reference image features at different temporal and frequency levels. We introduce synchronized ReferenceNet to extract reference features while optimizing noise injection and denoising. We also propose a time-aware frequency refinement module that uses high- and low-frequency filters with time embeddings to adaptively select reference feature injection. Additionally, we introduce a reward-based loss to improve" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 693, + 141, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 693, + 141, + 703 + ], + "spans": [ + { + "bbox": [ + 70, + 693, + 141, + 703 + ], + "type": "text", + "content": "*Equal contributions." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 70, + 703, + 147, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 147, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 147, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2714" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 131 + ], + "type": "text", + "content": "the similarity between reference objects and generated images. Experimental results show that TFCustom outperforms existing methods in single-object and multi-object reference generation, with significant improvements in textual details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 155, + 136, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 136, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 136, + 167 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 317 + ], + "type": "text", + "content": "With the development of deep learning [14, 19, 24, 46, 52] and text-to-image models [8, 15, 23, 32, 34], customization generation which aims to synthesize text-driven images for given subjects, has recently revolutionized content creation. It has seen a shift from test-time fine-tuning methods (e.g., DreamBooth [35], Custom Diffusion [18]) to fine-tuning-free techniques (e.g., IP-Adapter [53], Subject-Diffusion [26], MS-Diffusion [48]). The latter approaches remove the need for fine-tuning during inference, significantly reducing computational costs while maintaining flexibility, and making them more scalable and practical for real-world use cases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 319, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 569 + ], + "type": "text", + "content": "These fine-tuning-free methods generally fall into two categories: image encoder-based [4, 5, 58] feature injection and reference network (ReferenceNet-based) [6, 12, 51] feature injection. Image encoder methods, such as IPAdapter [53] and ELITE [49], utilize pre-trained encoders like CLIP [33] to extract high-level features, which are then injected into the generation process via mechanisms such as decoupled cross-attention. While these methods are computationally efficient and straightforward, they tend to rely on high-level, abstract features, which may not capture the fine-grained details necessary for precise control over reference object characteristics. The increasing demand for greater consistency in reference images has brought significant attention to the concept of ReferenceNet [16] in the field. ReferenceNet-based methods [17, 47, 50, 57] introduce an additional diffusion network within the traditional Diffusion Net denoising process. In this setup, reference images are encoded by VAE and then used as inputs for ReferenceNet. This novel architecture enables the correlation of features across layers in both diffusion networks, which helps in making better use of the reference images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Although ReferenceNet-based methodologies have made significant advancements, several challenges remain. Currently, ReferenceNet predominantly operates as a latent-level feature extractor, thereby not fully realizing its potential. While ReferenceNet maintains a consistent intensity of feature conditioning over time, the denoising backbone requires the injection of varying intensities and granularities of reference information at different timesteps. Specifically, during the earlier stages of denoising, the backbone tends to emphasize contour information, whereas, in the later stages, it focuses more on the refinement of details [45]. This limitation hinders the ability" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "to provide the denoising backbone with the appropriate features necessary for optimal performance across varying timesteps, ultimately leading to subpar image consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 109, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 418 + ], + "type": "text", + "content": "In this work, we introduce TFCustom, a novel finetuning-free personalized generation framework that incorporates temporal awareness for reference feature injection. Unlike previous approaches, TFCustom considers the alignment of reference image features with generated image features across different time steps of the diffusion process. Specifically, we first apply noise to the reference image at multiple time steps, with ReferenceNet extracting features at various scales and granularities. The training of ReferenceNet follows a diffusion loss function, ensuring proper alignment with the denoising process. Additionally, we introduce a time-aware frequency module that enables the controlled injection of high- and low-frequency reference features. This module allows low-frequency features to define the primary subject at early time steps, while high-frequency features are progressively injected to refine finer image details. To further improve the model's ability to handle multi-object reference generation, we propose a novel reward-based loss function that enhances the alignment between the reference and generated objects. This loss function computes object ID similarity during the denoising process, minimizing conflicts between reference objects and improving the overall coherence of multi-object scenes. The experimental results demonstrate our method consistently outperforms the state-of-the-art approaches on all the benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 419, + 493, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 419, + 493, + 430 + ], + "spans": [ + { + "bbox": [ + 326, + 419, + 493, + 430 + ], + "type": "text", + "content": "Overall, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 431, + 553, + 586 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 431, + 553, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 553, + 490 + ], + "type": "text", + "content": "- We propose a novel customized image generation framework, which includes synchronized noise injection and denoising within ReferenceNet, along with a time-aware frequency feature refinement module for extracting fine-grained image details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "text", + "content": "- We propose a reward-based loss function that alleviates multi-object conflicts and enhances object identity consistency in generated images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 526, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 553, + 586 + ], + "type": "text", + "content": "- Our model achieves state-of-the-art performance in single-object reference generation on DreamBooth and MS-Bench datasets. Additionally, we demonstrate superior qualitative results in terms of texture and textual detail generation compared to existing methods." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 400, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Subject Driven Text-to-Image Generation. Despite significant advancements in image generation [1, 8, 15, 23, 34], the field still falls short of meeting the requirements for personalized generation [38-40, 42, 48, 54]. This gap has consequently drawn significant attention to subject-driven text-to-image generation methods [4, 26, 43]. Textual Inversion [9] and DreamBooth [35] use a textual identifier to bind the visual concept through fine-tuning model textual" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2715" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "embedding or UNet model. CustomDiffusion [18] further extends multi-subject generation by multiple identifiers. IP-Adapter [53] injects the visual concept by decoupling the cross-attention between text and images in a zero-shot manner. JeDi [55] learns the joint distribution of multiple text-image pairs that share a common subject to facilitate personalized generation. Subject-Diffusion [26] trains multiple adapters inside the model with an attention control mechanism for multiple-subject personalization. SSR-Encoder [58] and MS-Diffusion [48] develop the query architecture to encode the subject images via mask or layout control for multiple-subject personalization. This paper explores the zero-shot multiple-objects driven generation via synchronized ReferenceNet optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 240, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 295, + 455 + ], + "type": "text", + "content": "ReferenceNet Based Personalized Generation. ReferenceNet [16, 41, 57] has shown remarkable capabilities in maintaining image consistency for personalized generation tasks. AnimateAnyone [16] first utilizes ReferenceNet for personalized human animation, employing a pre-trained generative model to extract reference features. FlashFace [57] extended this approach by combining ReferenceNet with ControlNet [56], enabling identity-preserving portrait generation that accommodates multiple reference images. Part2Whole [17] uses ReferenceNet to handle multi-part human body generation, while OOTD [50] and Stable-Garment [47] integrate it into virtual try-on tasks, aligning clothing features precisely with the target human body. These methods highlight ReferenceNet's versatility in image and video generation. To further improve performance, we introduce a synchronized ReferenceNet that refines reference features at each time step, enabling more precise feature injection and enhancing generation accuracy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "content": "Frequency Information in Image Generation. Frequency information plays a crucial role in image restoration tasks [7, 22, 27, 36]. Recent advancements in text-to-image generation have leveraged frequency-based conditions, like canny and HED images, to control generation, as seen in ControlNet [56] and T2I-Adapter [28]. FCDiffusion [11] and FBSdiff [10] employ DCT-transformed image features to enhance texture detail in image-to-image generation. In this work, our method adaptively injects reference image features via learnable high-pass and low-pass filters in a time-aware fashion, enriching each denoising step with frequency-specific information for more refined generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 609, + 111, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 111, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 111, + 621 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "Our proposed TFCustom addresses these issues through three components: Sec. 3.2 Synchronized Reference Guidance, Sec. 3.3 Time-Aware Frequency Feature Refinement, and Sec. 3.4 Reward Model Optimization, as illustrated in Fig. 2. First, the reference image is incrementally noised over " + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": " steps before feature extraction via the ReferenceNet, aligning it with the denoising network's noisy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "latent features. Next, a time-aware frequency module injects frequency-specific information from the reference image into the noisy latent representation. Finally, a reward model optimizes this process, enhancing the quality and consistency of the generated reference-based images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 140, + 402, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 402, + 151 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 402, + 151 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 158, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 553, + 230 + ], + "type": "text", + "content": "Diffusion models. Denoising Diffusion Probabilistic Models (DDPM) are a class of generative models that leverage a diffusion process to generate data. By progressively adding Gaussian noise to the data over a specified number of time steps, a sequence of increasingly noisy samples is produced. The forward process can be described as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 239, + 553, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 239, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 351, + 239, + 553, + 252 + ], + "type": "interline_equation", + "content": "q \\left(x _ {T} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {T}; \\sqrt {\\alpha_ {T}} x _ {0}, (1 - \\alpha_ {T}) \\mathbf {I}\\right) \\tag {1}", + "image_path": "6ce4a18f271ea56ee46d483c22da597c442a988726b0c1f1a12fcac8fe226bed.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\prod_{s=1}^{t}(1 - \\beta_{s})" + }, + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "text", + "content": " represents a variance schedule that regulates the noise level at each time step " + }, + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 261, + 553, + 331 + ], + "type": "text", + "content": " denotes a Gaussian distribution. In the reverse diffusion process, the model learns to denoise the noisy samples step by step, effectively generating new data from pure noise:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 347, + 342, + 553, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 342, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 347, + 342, + 553, + 357 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\sigma_ {\\theta} ^ {2} (t) \\mathbf {I}\\right) \\tag {2}", + "image_path": "0572d830a89604438c99c14b7be7b88cc1421fde16b9d33606c3a7f5cd305882.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "inline_equation", + "content": "\\sigma_{\\theta}" + }, + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "text", + "content": " are learned parameters that depend on the current noisy sample " + }, + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "text", + "content": " and the time step " + }, + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 365, + 553, + 413 + ], + "type": "text", + "content": ". The optimization goal is the variational lower bound on the log-likelihood of the data, which can be simplified to:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 421, + 553, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 421, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 372, + 421, + 553, + 436 + ], + "type": "interline_equation", + "content": "L = \\mathbb {E} _ {t, x _ {0}, \\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2} \\right] \\tag {3}", + "image_path": "f93bd871cd1d754441de238d3edf8c27762bc9fb5a57e65a421472807cba8015.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 444, + 515, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 515, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 515, + 456 + ], + "type": "text", + "content": "3.2. Synchronized ReferenceNet Guidance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "text", + "content": "The synchronized ReferenceNet is initially transferred from the denoising network. During training, at each time step " + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "text", + "content": ", we apply a noise process to the reference image " + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "text", + "content": ", adding noise for " + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "text", + "content": " steps, resulting in the noise reference " + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{ref}}^{t}" + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "text", + "content": " as formulated in Eq. (1). This step ensures that the reference features are temporally aligned with the noisy latent features processed by the denoising network. Once the noised reference " + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{ref}}^{t}" + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "text", + "content": " is generated, it passes through the synchronized ReferenceNet, which extracts the relevant features " + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "text", + "content": " and estimates the corresponding noise pattern. The extracted reference features and predicted noise are then optimized according to the objective " + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 462, + 554, + 665 + ], + "type": "text", + "content": " defined in Eq. (3). This objective encourages the synchronized ReferenceNet to provide refined guidance that is consistent with the noisy latent representations, enabling the denoising network to more effectively incorporate reference information for improved reconstruction quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Given that the object information hierarchically learned by the denoising network varies significantly across different timesteps, the early timesteps tend to focus on learning macro-level information such as contours, while the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2716" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 317, + 233 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 317, + 233 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 317, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 317, + 233 + ], + "type": "image", + "image_path": "aba4401df3cc440eee8b9b53f84da2a06be24a94f4a47ae75ac008714405975c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 239, + 230, + 251 + ], + "lines": [ + { + "bbox": [ + 132, + 239, + 230, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 230, + 251 + ], + "type": "text", + "content": "(a) Overall Framework" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 320, + 78, + 552, + 220 + ], + "blocks": [ + { + "bbox": [ + 320, + 78, + 552, + 220 + ], + "lines": [ + { + "bbox": [ + 320, + 78, + 552, + 220 + ], + "spans": [ + { + "bbox": [ + 320, + 78, + 552, + 220 + ], + "type": "image", + "image_path": "daf168b010f9eff67fd0d14eecfb9a12313c2dfce082135fe173d71c4b50b574.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 239, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 343, + 239, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 343, + 239, + 547, + 251 + ], + "type": "text", + "content": "(b) Time-Aware Frequency Feature Refinement" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 263, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 263, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 555, + 299 + ], + "type": "text", + "content": "Figure 2. (a) Overall Framework. Our TFCustom includes three key components: Synchronized Reference Net, Time-Aware Frequency Feature Refinement (TA-FFR), and Reward Model Optimization for subject-driven generation. (b) TA-FFR Module. This module extracts different frequency features in a time-aware manner, adaptively injecting reference features for enhanced generation quality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 318, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 296, + 426 + ], + "type": "text", + "content": "later timesteps progressively refine texture details. However, the synchronized timestep information obtained from " + }, + { + "bbox": [ + 54, + 318, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 54, + 318, + 296, + 426 + ], + "type": "text", + "content": " encompasses the entire reference object's information at timestep " + }, + { + "bbox": [ + 54, + 318, + 296, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 318, + 296, + 426 + ], + "type": "text", + "content": ", which can be overly complex for the denoising network. To address this issue, we have implemented a Time-Aware Frequency Feature Refinement module (which will be described detailed in Sec. 3.3) to refine the reference information, thereby enhancing the learning process of the denoising network, as outlined in Eq. 7." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "text", + "content": "With the enhanced reference features " + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{enh}}" + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "text", + "content": ", we inject them using a Shared Self-Attention mechanism [16, 17]. Specifically, in each self-attention layer, given the features of " + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "text", + "content": " reference images " + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{enh}}^{1:N}" + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "text", + "content": " and the feature map " + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^0" + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "text", + "content": " from the denoising net, we concatenate these feature maps side-by-side as input to the self-attention layer, represented as " + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "inline_equation", + "content": "[\\mathbf{F}_{\\mathrm{enh}}^0 | \\mathbf{F}_{\\mathrm{enh}}^1| \\dots | \\mathbf{F}_{\\mathrm{enh}}^N]" + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "text", + "content": ". This setup enables each location " + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^0" + }, + { + "bbox": [ + 55, + 426, + 296, + 534 + ], + "type": "text", + "content": " to attend to all locations on itself and the reference feature maps, computed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 543, + 294, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 543, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 58, + 543, + 294, + 602 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\mathbf {F}} _ {s} ^ {0} = \\operatorname {S o f t M a x} \\left(\\frac {\\mathbf {F} _ {s} ^ {0} W _ {q} \\cdot \\left([ \\mathbf {F} _ {\\mathrm {e n h}} ^ {0} , \\mathbf {F} _ {\\mathrm {e n h}} ^ {1} , \\dots , \\mathbf {F} _ {\\mathrm {e n h}} ^ {N} ]\\right) ^ {\\top} W _ {k}}{\\sqrt {d}}\\right) \\\\ \\cdot \\left(\\left[ \\mathbf {F} _ {\\text {e n h}} ^ {0}, \\mathbf {F} _ {\\text {e n h}} ^ {1}, \\dots , \\mathbf {F} _ {\\text {e n h}} ^ {N} \\right]\\right) W _ {v}, \\tag {4} \\\\ \\end{array}", + "image_path": "9a83639d27a0cfb10f7bb67d646ab39cceb0e7396da0584b16b1d31217b215e8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "inline_equation", + "content": "W_{q}, W_{k}, W_{v}" + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "content": " are corresponding projection weight matrices. And " + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 602, + 295, + 627 + ], + "type": "text", + "content": " is the dimensionality of the key vectors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 634, + 285, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 285, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 285, + 647 + ], + "type": "text", + "content": "3.3. Time-Aware Frequency Feature Refinement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Inspired by previous work [45], which illustrates that neural network learns hierarchically—starting with low-frequency features that define shape and structure, followed by high-frequency details for realism, we propose that reference image feature injection should follow a similar coarse-to-fine" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 318, + 553, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 553, + 341 + ], + "type": "text", + "content": "approach. This would better align with the denoising stages of diffusion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "content": "To achieve this, we introduce a Time-Aware Frequency Feature Refinement (TA-FFR), as shown in Fig. 2. This module separates reference image features into high-frequency and low-frequency components, each processed independently to capture unique information. The high-frequency path applies a Krisch operator for edge and detail preservation, while the low-frequency path uses a Gaussian operator to retain foundational structures. These operators are initialized as " + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "content": " convolution kernels " + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{conv}}" + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "content": " is Krisch operator, and " + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{conv}}" + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "content": " is Gaussian operator. Both operators are scaled by learnable parameters to adjust the frequency extraction dynamically:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 497, + 553, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 497, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 497, + 553, + 521 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {\\text {o u t}} = \\operatorname {C o n v} _ {\\mathbf {H}} \\left(\\mathbf {F} _ {\\text {r e f}}\\right) \\cdot \\mathbf {W} _ {\\mathrm {H}}, \\quad \\mathbf {L} _ {\\text {o u t}} = \\operatorname {C o n v} _ {\\mathbf {L}} \\left(\\mathbf {F} _ {\\text {r e f}}\\right) \\cdot \\mathbf {W} _ {\\mathrm {L}}, \\tag {5}", + "image_path": "0dc1aae8686c81ab9d5ea81212125530f71a1c724ba7b12eaf077a0bef903d48.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " denotes reference image features, and " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{H}}" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " are scaling factors for high and low frequencies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 546, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 606 + ], + "type": "text", + "content": "Each frequency component then undergoes Time-Aware Attention (TA-Attention) and Time-Aware Feed-Forward Network (TA-FFN), which incorporate time-step information using a time embedding " + }, + { + "bbox": [ + 313, + 546, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{emb}}" + }, + { + "bbox": [ + 313, + 546, + 554, + 606 + ], + "type": "text", + "content": " via Adaptive Layer Normalization (AdaLN) proposed in DiT [31]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 616, + 553, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 616, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 319, + 616, + 553, + 645 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {F} _ {\\text {h i g h}} = \\mathrm {T A - F F N} (\\mathrm {T A - A t t e n t i o n} (\\mathrm {A d a L N} (\\mathbf {H} _ {\\text {o u t}}, \\mathbf {t} _ {\\text {e m b}}))) \\\\ \\mathbf {F} _ {\\text {l o w}} = \\mathrm {T A - F F N} (\\mathrm {T A - A t t e n t i o n} (\\mathrm {A d a L N} (\\mathbf {L} _ {\\text {o u t}}, \\mathbf {t} _ {\\text {e m b}}))), \\end{array} \\tag {6}", + "image_path": "d81cd577e50f5a2eb87c9db6633e1c8c61242d5b02d1aeea37258e3bce0f2864.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{high}}" + }, + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{low}}" + }, + { + "bbox": [ + 313, + 654, + 553, + 690 + ], + "type": "text", + "content": " represent high-frequency and low-frequency features. Then the final enhanced feature is obtained by combining these outputs:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 390, + 701, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 390, + 701, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\text {e n h}} = \\mathbf {F} _ {\\text {h i g h}} + \\mathbf {F} _ {\\text {l o w}}. \\tag {7}", + "image_path": "d719d23a702fb5e56b2876b4234bded87d64af2865fb1c2a77357ccd17983dbf.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2717" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 290, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 290, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 290, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 290, + 159 + ], + "type": "image", + "image_path": "6da9dee83c5f758366ecffe6e56115382415310607805ccd8f498c10f8fd1b43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 168, + 295, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 192 + ], + "type": "text", + "content": "Figure 3. Reward Model Optimization Process. \"G\" and \"D\" indicate the network's forward pass and backward optimization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 212, + 296, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 260 + ], + "type": "text", + "content": "This design ensures a coarse-to-fine feature injection, enhancing the model's ability to integrate temporal and spatial details, thereby improving the quality of the generated images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 212, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 212, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 212, + 280 + ], + "type": "text", + "content": "3.4. Reward Model Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": "Our model can simulate the noise levels of both the reference image and the target image. Based on the following equation, we can perform a one-step prediction of the source image, allowing us to naturally determine the reward model score based on the noise level by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 354, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 354, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 98, + 354, + 295, + 380 + ], + "type": "interline_equation", + "content": "x _ {0} ^ {\\prime} \\approx \\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} \\left(x _ {t} , x _ {\\text {r e f}} , c _ {t} , t\\right)}{\\sqrt {\\alpha_ {t}}}. \\tag {8}", + "image_path": "31ea26edd0f728b3e1e141035d5dfe7924ae2b1d28ac7af6ec500cb65552e334.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 388, + 296, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 388, + 296, + 436 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 55, + 388, + 296, + 436 + ], + "type": "text", + "content": " refers to the textual condition. This equation's calculation error increases with timestep " + }, + { + "bbox": [ + 55, + 388, + 296, + 436 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 388, + 296, + 436 + ], + "type": "text", + "content": ". Therefore, during the training phase, we only apply the reward loss for timestep " + }, + { + "bbox": [ + 55, + 388, + 296, + 436 + ], + "type": "inline_equation", + "content": "t < T_0" + }, + { + "bbox": [ + 55, + 388, + 296, + 436 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 445, + 295, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 445, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 93, + 445, + 295, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e w a r d}} = \\operatorname {D i s} \\left(\\mathrm {D I N O} _ {v 2} \\left(x _ {0}\\right), \\mathrm {D I N O} _ {v 2} \\left(x _ {0} ^ {\\prime}\\right)\\right) \\tag {9}", + "image_path": "163f05ea563ffecd6a369c1188230596bce849cfb344a4d87f1fa76a209dbbc3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 467, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 597 + ], + "type": "text", + "content": "where Dis is cosine similarity metric and " + }, + { + "bbox": [ + 55, + 467, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\mathrm{DINO}_{v2}" + }, + { + "bbox": [ + 55, + 467, + 296, + 597 + ], + "type": "text", + "content": " is the DINO v2 model [2] to compute the similarity between the target domain's denoised output and the features generated by the reference net. It is worth noticing that in our synchronized noise injection scheme, the output of the reference net corresponds to the noise level at timestep " + }, + { + "bbox": [ + 55, + 467, + 296, + 597 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 467, + 296, + 597 + ], + "type": "text", + "content": ", which provides meaningful information. Thus, we take advantage of this synchronized noise injection feature and introduce a diffusion loss, " + }, + { + "bbox": [ + 55, + 467, + 296, + 597 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 467, + 296, + 597 + ], + "type": "text", + "content": ", for the reference net, which supervises the feature extraction capability of the reference branch. The overall loss function can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 605, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 605, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 100, + 605, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d i f f}} ^ {\\text {d e n o i s i n g}} + \\lambda_ {1} \\mathcal {L} _ {\\text {d i f f}} ^ {\\text {r e f}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e w a r d}}, \\tag {10}", + "image_path": "94eb75ac0d839fd8ae5da655d35168b96b89b1a044b67071eaea4e0e87412c6f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 253, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 253, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 253, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 253, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 629, + 253, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 253, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 629, + 253, + 641 + ], + "type": "text", + "content": " is hyper-parameters to balance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 671, + 164, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 164, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 164, + 685 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Datasets. The training dataset comprises 2.3 million (M) in-house image pairs. We assess the performance on" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 553, + 245 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 553, + 245 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 553, + 245 + ], + "type": "table", + "html": "
MethodTypeCLIP-IDINOCLIP-T
BLIP-Diffusion [20]No-FT77.959.430.0
IP-Adapter [53]No-FT81.061.329.2
Emu2 [44]No-FT76.556.327.3
Kosmos-G [29]No-FT82.261.825.0
λ-ECLIPSE [30]No-FT78.361.330.7
SSR-Encoder [58]No-FT82.161.230.8
MS-Diffusion [48]No-FT79.267.132.1
TFCustomNo-FT84.171.433.9
Textual Inversion [9]FT78.056.925.5
DreamBooth [35]FT80.366.830.5
BLIP-Diffusion [20]FT80.567.030.2
λ-ECLIPSE [30]FT79.668.230.4
MS-Diffusion [48]FT80.570.231.3
TFCustomFT85.974.135.1
", + "image_path": "18da077c8cbdcce7e8c8c0a947c787a9767adf19cfeee0c4b3a36c50dc1ad1bc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "lines": [ + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "type": "text", + "content": "Table 1. Quantitative comparison for single-subject models. FT denotes fine-tuned models on DreamBench [35]. The best and second-best results are shown in bold and underline, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 294, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 402 + ], + "type": "text", + "content": "single-subject tasks using the DreamBench dataset [35] and on multi-subject tasks using the MS-Bench dataset [48]. DreamBench encompasses 30 subjects and 25 prompts, while MS-Bench includes 40 subjects and 1148 combinations, with each combination being associated with up to 6 prompts, resulting in a total of 4488 unique test samples. Furthermore, we collect an additional 20 real-world cases featuring complex textures, characters, and patterns for further qualitative evaluation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "type": "text", + "content": "Implementation Details. During training, we employ the AdamW optimizer with a learning rate of 1e-4 and train the model for 100K steps on 32 NVIDIA A100 80GB GPUs, using a batch size of 4 per GPU. The " + }, + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "type": "text", + "content": " in Eq. (10) are set to 0.4 and 0.6. To enable classifier-free guidance, we apply a probability of 0.05 to independently drop text and images, and 0.05 to drop both simultaneously. For inference, we adapt the DDIM sampler with 50 steps and set the guidance scale to 7.5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 653 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate performance based on both image and text fidelity metrics. For image fidelity, we compute the cosine similarity between generated images and subject images within the CLIP (denoted as CLIP-I) and DINO feature spaces. Text fidelity is measured by calculating the cosine similarity between generated images and text prompts in the CLIP space (referred to as CLIP-T). In the context of multi-subject personalization, average fidelity scores may not adequately reflect the representation of individual subjects. Therefore, we also employ the multi-subject DINO (M-DINO) metric [48] to assess whether each subject was accurately recreated." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Competitors. We compare our model with BLIP-Diffusion [20], IP-Adapter [53], Kosmos-G [29], Emu2 [44], " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "-ECLIPSE [30], SSR-Encoder [58], and MS-Diffusion [48] for single concept personalization, and with MS-Diffusion [48], SSR-Encoder [58] and " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "-ECLIPSE [30] for multi-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2718" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 553, + 333 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 553, + 333 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 553, + 333 + ], + "type": "image", + "image_path": "1a3d9c0a701ce79bac4229a27066b907a91fe1c3258c227750f24ea313020ae3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 341, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 341, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 555, + 365 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with existing state-of-the-art customization methods. TFCustom produces higher-quality results with better subject similarity and text controllability compared to existing methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 385, + 153, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 153, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 153, + 398 + ], + "type": "text", + "content": "subject personalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 406, + 203, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 203, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 203, + 419 + ], + "type": "text", + "content": "4.2. Performance Comparison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 700 + ], + "type": "text", + "content": "Quantitative Comparison of Single-Subject. The results are presented in Tab. 1. In the zero-shot evaluation (without fine-tuning on DreamBench), our proposed TFCustom achieves the highest scores across all metrics, with CLIP-I, DINO, and CLIP-T values of 84.1, 71.4, and 33.9, respectively. These results surpass the second-best methods by margins of 1.9 in CLIP-I (compared to 82.2 by Kosmos-G), 4.3 in DINO (compared to 67.1 by MS-Diffusion), and 1.8 in CLIP-T (compared to 32.1 by MS-Diffusion). Notably, our method even surpasses models fine-tuned at test time, showcasing robust zero-shot capabilities. TFCustom's superior performance can be attributed to the synchronized reference network, which provides multi-scale features for highly detailed image generation, and the time-aware frequency module, which harmonizes semantic information with reference image features. Additionally, the reward model enhances object similarity by leveraging referential cues. When fine-tuned on DreamBench, TFCustom shows further improvements, achieving top scores in CLIP-I, DINO, and CLIP-T metrics with values of 85.9, 74.1, and 35.1, respectively. This highlights TFCustom's adaptability in personalization scenarios, demonstrating strong transferability across various contexts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Quantitative Comparison of Multi-Subject. We also com" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 383, + 553, + 460 + ], + "blocks": [ + { + "bbox": [ + 316, + 383, + 553, + 460 + ], + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 460 + ], + "type": "table", + "html": "
MethodCLIP-IDINOM-DINOCLIP-T
λ-ECLIPSE [30]72.441.99.431.6
SSR-Encoder [58]72.542.510.730.3
MS-Diffusion [48]69.842.510.834.1
TFCustom74.343.812.936.9
", + "image_path": "aba44638a10e6e1667bdb7689af7c0ce19978c26621e7eaf6352dda30c6ca079.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 469, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 469, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 555, + 502 + ], + "type": "text", + "content": "Table 2. Quantitative comparison for multi-subject models. The best and second-best results are shown in **bold** and **underline**, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 510, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 641 + ], + "type": "text", + "content": "pare multi-subject personalization performance in Tab. 2. The results show that TFCustom achieves state-of-the-art performance across CLIP-I, DINO, M-DINO, and CLIP-T metrics, with scores of 74.3, 43.8, 12.9, and 36.9, respectively, significantly surpassing other reference image generation methods. Compared to MS-Diffusion, TFCustom demonstrates superior text-following capability, as evidenced by its higher CLIP-T scores, outperforming methods like MS-Diffusion and SSR-Encoder. Additionally, TFCustom excels in the M-DINO metric, highlighting its ability to preserve multi-object details in generated images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Qualitative Results. Fig. 4 presents qualitative results from the DreamBench dataset, demonstrating our model's remarkable capacity for generating high-fidelity single-subject images with excellent detail preservation. Furthermore, Fig. 5 illustrates qualitative outcomes from the MS-Bench dataset. Our approach adeptly maintains the textual" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2719" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 75, + 197, + 366 + ], + "blocks": [ + { + "bbox": [ + 61, + 75, + 197, + 366 + ], + "lines": [ + { + "bbox": [ + 61, + 75, + 197, + 366 + ], + "spans": [ + { + "bbox": [ + 61, + 75, + 197, + 366 + ], + "type": "image", + "image_path": "73b07ca0f3e9828453e715165cdb21d75b6bc9e2352b9bd439a40c8e6865e820.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 377, + 555, + 400 + ], + "lines": [ + { + "bbox": [ + 55, + 377, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 555, + 400 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of multiple-subject customization. TFCustom achieves enhanced text controllability and subject fidelity compared to existing methods. Customized words are highlighted." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 75, + 238, + 368 + ], + "blocks": [ + { + "bbox": [ + 205, + 75, + 238, + 368 + ], + "lines": [ + { + "bbox": [ + 205, + 75, + 238, + 368 + ], + "spans": [ + { + "bbox": [ + 205, + 75, + 238, + 368 + ], + "type": "image", + "image_path": "cb6189be6e80ccdb606571c4f142db0f8b8395649ad793ee4a554c2631625881.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 249, + 74, + 318, + 366 + ], + "blocks": [ + { + "bbox": [ + 249, + 74, + 318, + 366 + ], + "lines": [ + { + "bbox": [ + 249, + 74, + 318, + 366 + ], + "spans": [ + { + "bbox": [ + 249, + 74, + 318, + 366 + ], + "type": "image", + "image_path": "029a911d6c107bf7d0ef98d912ddf90ea56dfb771895d36f410c3f2e7c2f772b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 327, + 74, + 396, + 366 + ], + "blocks": [ + { + "bbox": [ + 327, + 74, + 396, + 366 + ], + "lines": [ + { + "bbox": [ + 327, + 74, + 396, + 366 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 396, + 366 + ], + "type": "image", + "image_path": "e7855f3ac4fdcc4665fe333aac8b565c9447429107cf7fe32c763585eda93be7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 405, + 74, + 473, + 366 + ], + "blocks": [ + { + "bbox": [ + 405, + 74, + 473, + 366 + ], + "lines": [ + { + "bbox": [ + 405, + 74, + 473, + 366 + ], + "spans": [ + { + "bbox": [ + 405, + 74, + 473, + 366 + ], + "type": "image", + "image_path": "dbfc8cb3305f785a86b4522e1d20a0ee49a16368b5840c736b133562ccb82480.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 481, + 74, + 548, + 366 + ], + "blocks": [ + { + "bbox": [ + 481, + 74, + 548, + 366 + ], + "lines": [ + { + "bbox": [ + 481, + 74, + 548, + 366 + ], + "spans": [ + { + "bbox": [ + 481, + 74, + 548, + 366 + ], + "type": "image", + "image_path": "08ce89588935041d5eb88e49f4b0a254ff032f750dfa1f3f56dba3fc9b13802e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 415, + 294, + 513 + ], + "blocks": [ + { + "bbox": [ + 57, + 415, + 294, + 513 + ], + "lines": [ + { + "bbox": [ + 57, + 415, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 294, + 513 + ], + "type": "image", + "image_path": "38a68a7550eec924bf69f25bf1a1cc6cbc323ac5bd999bbb236a662102882eba.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 523, + 280, + 534 + ], + "lines": [ + { + "bbox": [ + 70, + 523, + 280, + 534 + ], + "spans": [ + { + "bbox": [ + 70, + 523, + 280, + 534 + ], + "type": "text", + "content": "Figure 6. Visualizations in customized human generation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": "interactions among multiple subjects within the generated images, ensuring that each subject remains distinct and easily recognizable. This effectively mitigates issues related to subject omission or conflicts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "Complex Scenarios. We evaluate the customized human generation in Fig. 6. The result shows that TFCustom is still capable of generating high-quality results for customized human generation. Additionally, we evaluate our model in real-world generation scenarios, as shown in Fig. 7. Our model is capable of generating images that accurately capture the fine-grained details of the reference images, including text, textures, and complex patterns. These showcase its strong capability in reference image feature transfer and high-quality image generation." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 411, + 553, + 544 + ], + "blocks": [ + { + "bbox": [ + 318, + 411, + 553, + 544 + ], + "lines": [ + { + "bbox": [ + 318, + 411, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 318, + 411, + 553, + 544 + ], + "type": "image", + "image_path": "f02264ee28ba1057f86c5f483d8e8f3245fdac29beac3d66501323ecc25d9ad0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 554, + 555, + 575 + ], + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 575 + ], + "type": "text", + "content": "Figure 7. Visualizations in complex scenarios with intricate patterns and text." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 595, + 408, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 408, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 408, + 609 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Model Design. We conduct ablation experiments to assess each component of our method, as shown in Tab. 3. " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " represents synchronized denoising, " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reward}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " involves the reward model, and Frequency denotes the time-aware frequency separation module. Each component significantly enhances performance. Synchronized denoising " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " sets a strong baseline, executing scores of 83.2, 72.4, and 32.9 for CLIP-T, CLIP-I, and DINO. The Frequency module improves the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2720" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 219, + 144 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 219, + 144 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 219, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 219, + 144 + ], + "type": "table", + "html": "
LrefLrewardFreq.CLIP-IDINOCLIP-T
XXX81.371.532.7
✓XX83.272.432.9
X✓✓84.373.734.1
✓✓✓85.974.135.1
", + "image_path": "ce7bad882a8ffa5d28fad480e0fa58f6b9ecd3b50fb8d20a22567373ae40056e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 230, + 70, + 370, + 144 + ], + "blocks": [ + { + "bbox": [ + 55, + 152, + 208, + 174 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 208, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 208, + 174 + ], + "type": "text", + "content": "Table 3. Impact " + }, + { + "bbox": [ + 55, + 152, + 208, + 174 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 152, + 208, + 174 + ], + "type": "text", + "content": " of different components. Freq. denotes TA-FFR module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 230, + 70, + 370, + 144 + ], + "lines": [ + { + "bbox": [ + 230, + 70, + 370, + 144 + ], + "spans": [ + { + "bbox": [ + 230, + 70, + 370, + 144 + ], + "type": "table", + "html": "
Data ScaleCLIP-IDINOCLIP-T
0.5 M80.169.830.1
1 M84.072.832.3
1.5 M84.973.734.8
2.3 M85.974.135.1
", + "image_path": "34473d79f5dca5ad4d08116b2c6050793d3fa00376e28ce4fd08e95c5090dcd8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 380, + 70, + 554, + 144 + ], + "blocks": [ + { + "bbox": [ + 236, + 152, + 361, + 163 + ], + "lines": [ + { + "bbox": [ + 236, + 152, + 361, + 163 + ], + "spans": [ + { + "bbox": [ + 236, + 152, + 361, + 163 + ], + "type": "text", + "content": "Table 4. Impact (%) of data scale." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 380, + 70, + 554, + 144 + ], + "lines": [ + { + "bbox": [ + 380, + 70, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 380, + 70, + 554, + 144 + ], + "type": "table", + "html": "
ModelCLIP-IDINOCLIP-T
DeepLabv3-MBv284.172.734.6
FCN-R10184.372.934.6
UperNet-R5085.473.834.7
DINOv285.974.135.1
", + "image_path": "0f1720135307f9934c7bf27b842146eb7575d98e193cb1b5b735ae12fcbd24b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 379, + 152, + 533, + 172 + ], + "lines": [ + { + "bbox": [ + 379, + 152, + 533, + 172 + ], + "spans": [ + { + "bbox": [ + 379, + 152, + 533, + 172 + ], + "type": "text", + "content": "Table 5. Impact (%) of different reward models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 186, + 294, + 290 + ], + "blocks": [ + { + "bbox": [ + 58, + 186, + 294, + 290 + ], + "lines": [ + { + "bbox": [ + 58, + 186, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 58, + 186, + 294, + 290 + ], + "type": "image", + "image_path": "1dda87ed4019b9a602ad08633b78d9d46d46942db787c7e519fa4710ecf523db.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 300, + 295, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 333 + ], + "type": "text", + "content": "Figure 8. The visualization of various ablation modules indicates that each proposed component within our methodology contributes positively to the overall results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 340, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 340, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 295, + 459 + ], + "type": "text", + "content": "performance across metrics (e.g., CLIP-I improves from 71.5 to 72.4), effectively enhancing temporal coherence and structural detail. The reward model " + }, + { + "bbox": [ + 54, + 340, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {reward }}" + }, + { + "bbox": [ + 54, + 340, + 295, + 459 + ], + "type": "text", + "content": " further boosts results by an average of 1.3 points (e.g., CLIP-I increases from 72.4 to 73.7), capturing nuanced information critical for quality restoration. With all components combined, we achieve the highest scores, demonstrating the complementary strengths of each component. Fig. 8 further examines the effect of each component, demonstrating the guidance of each component." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "type": "text", + "content": "Reward Model. Following [21], we examine the effectiveness of different reward models, as shown in Tab. 5. DINOv2 [2] achieves the highest scores across all metrics, demonstrating superior performance in text relevance and image fidelity. FCN [25] with ResNet-101 [13] backbone(FCN-R101), DeepLabv3 [3] with MobileNetv2 [37] (DeepLabv3-MBv2) also perform well, though slightly lower, confirming that the reward model enhances feature retention and alignment across architectures, with DINOv2 as the optimal choice." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 587, + 157, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 157, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 157, + 600 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Feature Map Visualization. Figure 9 visualizes the cross-attention map between reference and latent features, aggregated across all layers to a uniform scale and averaging. Distinct reference images produce unique feature maps, as shown in the top row. Our method demonstrates consistent object perception across timesteps, a capability lacking in previous ReferenceNet features. Additionally, the incorporation of frequency loss and the reward model enables our approach to capture finer details, resulting in enhanced im" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 186, + 520, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 186, + 520, + 199 + ], + "spans": [ + { + "bbox": [ + 314, + 186, + 520, + 199 + ], + "type": "text", + "content": "age quality and more precise object representation." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 209, + 552, + 342 + ], + "blocks": [ + { + "bbox": [ + 319, + 209, + 552, + 342 + ], + "lines": [ + { + "bbox": [ + 319, + 209, + 552, + 342 + ], + "spans": [ + { + "bbox": [ + 319, + 209, + 552, + 342 + ], + "type": "image", + "image_path": "412320aa4e44b559cf3122db14918d6f4b8ac25a391e7cba05dbb4afcfac5d87.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 352, + 554, + 374 + ], + "lines": [ + { + "bbox": [ + 313, + 352, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 554, + 374 + ], + "type": "text", + "content": "Figure 9. Cross-Attention Feature Maps. Visualization of the attention maps between reference features and latent features." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 382, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 502 + ], + "type": "text", + "content": "Training Data. Tab. 4 demonstrates the impact of training data size on model performance. As the amount of data increases, we observe consistent improvements across all evaluation metrics (CLIP-T, CLIP-I, and DINO). Specifically, using 0.5M data samples yields moderate results, but expanding to 1M and 1.5M significantly enhances model accuracy, indicating the value of larger datasets in refining feature representation. The best performance is achieved with 2.3M samples, highlighting that increased training data directly contributes to more robust and detailed outputs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 514, + 391, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 391, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 391, + 526 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "In this study, we propose TFCustom, a novel model framework that emphasizes the importance of reference image features across different temporal steps and frequency levels. Our approach uses synchronized ReferenceNet to extract these features while optimizing noise injection and denoising processes. Additionally, we introduce a time-aware frequency refinement module that employs high- and low-frequency filters, along with time embeddings, to dynamically adjust the intensity of reference feature injection. To improve alignment between reference objects and generated images, we propose a reward model loss that strengthens their correspondence. Experimental results demonstrate the effectiveness of our framework, achieving state-of-the-art performance and significant improvements in texture and textual detail generation compared to existing methods." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2721" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2:3, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 225 + ], + "type": "text", + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation, 2017. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 295, + 258 + ], + "type": "text", + "content": "[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. CoRR, abs/2307.09481, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 294 + ], + "type": "text", + "content": "[5] Xi Chen, Zhiheng Liu, Mengting Chen, Yutong Feng, Yu Liu, Yujun Shen, and Hengshuang Zhao. Livephoto: Real image animation with text-guided motion control, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 295, + 327 + ], + "type": "text", + "content": "[6] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for authentic virtual try-on in the wild, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 330, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 295, + 373 + ], + "type": "text", + "content": "[7] Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. Adair: Adaptive all-in-one image restoration via frequency mining and modulation. arXiv preprint arXiv:2403.14614, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 375, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 441 + ], + "type": "text", + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 442, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 295, + 485 + ], + "type": "text", + "content": "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 487, + 295, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 295, + 530 + ], + "type": "text", + "content": "[10] Xiang Gao and Jiaying Liu. Fbsdiff: Plug-and-play frequency band substitution of diffusion features for highly controllable text-driven image translation. In ACM MM, pages 4101-4109, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 575 + ], + "type": "text", + "content": "[11] Xiang Gao, Zhengbo Xu, Junhan Zhao, and Jiaying Liu. Frequency-controlled diffusion model for versatile text-guided image-to-image translation. In AAAI, pages 1824-1832, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 295, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 620 + ], + "type": "text", + "content": "[12] Yuming Gu, You Xie, Hongyi Xu, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 623, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 623, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 623, + 295, + 645 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[15] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive mod" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "els for fine-grained text-to-image synthesis. In AAAI, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[16] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. Animate anyone: Consistent and controllable image-to-video synthesis for character animation, 2024. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "text", + "content": "[17] Zehuan Huang, Hongxing Fan, Lipeng Wang, and Lu Sheng. From parts to whole: A unified reference framework for controllable human image generation, 2024. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[18] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pages 1931-1941, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 252 + ], + "type": "text", + "content": "[19] Bozheng Li, Mushui Liu, Gaoang Wang, and Yunlong Yu. Frame order matters: A temporal sequence-aware model for few-shot action recognition. In AAAI, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 254, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 286 + ], + "type": "text", + "content": "[20] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 288, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 330 + ], + "type": "text", + "content": "[21] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In ECCV, 2024.8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "text", + "content": "[22] Chengxu Liu, Xuan Wang, Shuai Li, Yuzhi Wang, and Xueming Qian. Fsi: Frequency and spatial interactive learning for image restoration in under-display cameras. In ICCV, pages 12537-12546, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "text", + "content": "[23] Mushui Liu, Yuhang Ma, Yang Zhen, Jun Dan, Yunlong Yu, Zeng Zhao, Zhipeng Hu, Bai Liu, and Changjie Fan. Llm4gen: Leveraging semantic representation of llms for text-to-image generation. In AAAI, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "text", + "content": "[24] Mushui Liu, Fangtai Wu, Bozheng Li, Ziqian Lu, Yunlong Yu, and Xi Li. Envisioning class entity reasoning by large language models for few-shot learning. In AAAI, 2025. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 456, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 478 + ], + "type": "text", + "content": "[25] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation, 2015. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "type": "text", + "content": "[26] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[27] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In AAAI, pages 1905–1913, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 601 + ], + "type": "text", + "content": "[28] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296–4304, 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 602, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 645 + ], + "type": "text", + "content": "[29] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[30] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. lambda-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 5, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4172-4182, 2023. 4" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2722" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 284 + ], + "type": "text", + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[36] Donghun Ryou, Inju Ha, Hyewon Yoo, Dongwan Kim, and Bohyung Han. Robust image denoising through adversarial frequency mixup. In CVPR, pages 2723-2732, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 352 + ], + "type": "text", + "content": "[37] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks, 2019. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 418 + ], + "type": "text", + "content": "[38] D She, Mushui Liu, Jingxuan Pang, Jin Wang, Zhen Yang, Wanggui He, Guanghao Zhang, Yi Wang, Qihan Huang, Haobin Tang, et al. Customvideox: 3d reference attention driven dynamic adaptation for zero-shot customized video diffusion transformers. arXiv preprint arXiv:2502.06527, 2025.2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 453 + ], + "type": "text", + "content": "[39] Fei Shen and Jinhui Tang. Imagpose: A unified conditional framework for pose-guided person generation. In NeurIPS, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "text", + "content": "[40] Fei Shen, Hu Ye, Jun Zhang, Cong Wang, Xiao Han, and Wei Yang. Advancing pose-guided image synthesis with progressive conditional diffusion models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "text", + "content": "[41] Fei Shen, Xin Jiang, Xin He, Hu Ye, Cong Wang, Xiaoyu Du, Zechao Li, and Jinghui Tang. Imagdressing-v1: Customizable virtual dressing. In AAAI, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 524, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 294, + 567 + ], + "type": "text", + "content": "[42] Fei Shen, Cong Wang, Junyao Gao, Qin Guo, Jisheng Dang, Jinhui Tang, and Tat-Seng Chua. Long-term talkingface generation via motion-prior conditional diffusion model. arXiv preprint arXiv:2502.09533, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 569, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 611 + ], + "type": "text", + "content": "[43] Fei Shen, Hu Ye, Sibo Liu, Jun Zhang, Cong Wang, Xiao Han, and Wei Yang. Boosting consistency in story visualization with rich-contextual conditional diffusion models. In AAAI, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 666 + ], + "type": "text", + "content": "[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[45] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimen" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 677 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "sional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "text", + "content": "[47] Rui Wang, Hailong Guo, Jiaming Liu, Huaxia Li, Haibo Zhao, Xu Tang, Yao Hu, Hao Tang, and Peipei Li. Stablegar-ment: Garment-centric generation via stable diffusion, 2024. 2,3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[48] X. Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[49] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV, pages 15897-15907, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[50] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "text", + "content": "[51] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 333, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 386 + ], + "type": "text", + "content": "[52] Zhen Yang, Guibao Shen, Liang Hou, Mushui Liu, Luozhou Wang, Xin Tao, Pengfei Wan, Di Zhang, and YingCong Chen. Rectifiedhr: Enable efficient high-resolution image generation via energy rectification. arXiv preprint arXiv:2503.02537, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 433, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 486 + ], + "type": "text", + "content": "[54] Jiacheng Ying, Mushui Liu, Zhe Wu, Running Zhang, Zhu Yu, Siming Fu, Si-Yuan Cao, Chao Wu, Yunlong Yu, and Hui-Liang Shen. Restorerid: Towards tuning-free face restoration with id preservation. arXiv preprint arXiv:2411.14125, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 489, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 542 + ], + "type": "text", + "content": "[55] Yu Zeng, Vishal M. Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In CVPR 2024, pages 6786-6795, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "text", + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3813-3824, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "text", + "content": "[57] Shilong Zhang, Lianghua Huang, Xi Chen, Yifei Zhang, ZhiFan Wu, Yutong Feng, Wei Wang, Yujun Shen, Yu Liu, and Ping Luo. Flashface: Human image personalization with high-fidelity identity preservation, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "text", + "content": "[58] Yuxuan Zhang, Jiaming Liu, Yiren Song, Rui Wang, Hao Tang, Jinpeng Yu, Huaxia Li, Xu Tang, Yao Hu, Han Pan, and Zhongliang Jing. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR, pages 8069–8078, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2723" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_content_list.json b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e5ba198bf5742128a401611d31314b4b46f5aee6 --- /dev/null +++ b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_content_list.json @@ -0,0 +1,1881 @@ +[ + { + "type": "text", + "text": "TIDE: Training Locally Interpretable Domain Generalization Models Enables Test-time Correction", + "text_level": 1, + "bbox": [ + 148, + 128, + 849, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aishwarya Agarwal1,2* Srikrishna Karanam2† Vineet Gandhi1‡", + "bbox": [ + 238, + 202, + 764, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ CVIT, Kohli Centre for Intelligent Systems, IIIT Hyderabad, India", + "bbox": [ + 297, + 229, + 699, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Adobe Research, Bengaluru, India", + "bbox": [ + 390, + 247, + 602, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 297, + 326, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider the problem of single-source domain generalization. Existing methods typically rely on extensive augmentations to synthetically cover diverse domains during training. However, they struggle with semantic shifts (e.g., background and viewpoint changes), as they often learn global features instead of local concepts that tend to be domain invariant. To address this gap, we propose an approach that compels models to leverage such local concepts during prediction. Given no suitable dataset with per-class concepts and localization maps exists, we first develop a novel pipeline to generate annotations by exploiting the rich features of diffusion and large-language models. Our next innovation is TIDE, a novel training scheme with a concept saliency alignment loss that ensures model focus on the right per-concept regions and a local concept contrastive loss that promotes learning domain-invariant concept representations. This not only gives a robust model but also can be visually interpreted using the predicted concept saliency maps. Given these maps at test time, our final contribution is a new correction algorithm that uses the corresponding local concept representations to iteratively refine the prediction until it aligns with prototypical concept representations that we store at the end of model training. We evaluate our approach extensively on four standard DG benchmark datasets and substantially outperform the current state-of-the-art (12% improvement on average) while also demonstrating that our predictions can be visually interpreted.", + "bbox": [ + 88, + 330, + 485, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 767, + 220, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enhancing deep neural networks to generalize to out-of-distribution samples remains a core challenge in machine learning and computer vision research, as real-world test data often diverges significantly from the training distribu", + "bbox": [ + 89, + 792, + 483, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69e775bd3f4808f78e554448d4f737adea66d3e269f1010adacc017ec20251c5.jpg", + "image_caption": [ + "Figure 1. Samples from VLCS (left) and PACS dataset (right) across domain shifts, corresponding to bird and person class. First row displays GradCAM maps [43] for ABA's class predictions. We observe that model attention of ABA [5] falters across domain shifts. Second and third row display the concept specific GradCAM maps from TIDE. We posit that accurate concept learning and localization facilitates DG." + ], + "image_footnote": [], + "bbox": [ + 517, + 300, + 915, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion [31]. The challenges compound when obtaining labeled samples from the target domain is expensive or unfeasible, hindering application of semi-supervised learning or domain adaptation [25, 32, 49]. The problem of Domain Generalization (DG) [30, 44, 54, 61, 63] represents a promising avenue for developing techniques that capture domain-invariant patterns and improve performance on out-of-distribution samples. In this paper, we focus on Single Source Domain Generalization (SSDG), where a model is trained on data from a single domain and aims to generalize well to unseen domains [40, 51]. It represents the most strict form of DG, as the model must extract domain-invariant features from a single, often limited perspective, without exposure to the variation present across multiple domains.", + "bbox": [ + 511, + 613, + 906, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most previous work on SSDG [5, 7, 8, 19, 20, 22, 59, 63] relies extensively on data augmentations to support the learning of domain-generalized features. The premise is that constructing an extensive repertoire of augmentations synthesizes instances encompassing a wide spectrum of", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\* aishwarya.agarwal@research.iit.ac.in, aishagar@adobe.com", + "bbox": [ + 107, + 864, + 403, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† skaranam@adobe.com", + "bbox": [ + 109, + 876, + 223, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ VGandhi@iiit.ac.in", + "bbox": [ + 109, + 887, + 204, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30210", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "human-recognizable domains. However, accounting for all conceivable real-world augmentations presents an immense challenge. These models have shown reasonable success in addressing variations in style and texture [45, 64]; however, their performance remains modest when faced with more semantic domain shifts, such as changes in background and viewpoint e.g., in the VLCS dataset [41, 48].", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consider the results of the state-of-the-art ABA [5] model on the two examples in Figure 1. In the first example (left), domain shifts manifest as variations in background and viewpoint. In the second example (right), the training data comprises photos of human faces, whereas the test data consists of skeletal sketches. In both cases, the domain shifts extend beyond style or texture variations, with ABA failing to correctly classify the samples. Class-level saliency maps [43] reveal that this misclassification stems from inadequate focus on critical local concepts under domain shifts, such as beak and feathers for birds, or eyes and lips for persons. Such local concepts are integral to class definition and remain invariant across domains, and a robust DG model must adeptly capture these stable features. We posit that prior efforts learn global features per class; and if the model fails to learn the correct feature set in the training domain, its generalization performance is compromised as noted from Figure 1 above. The inconsistency in concept localization further exacerbates the interpretability and explainability of these models.", + "bbox": [ + 91, + 200, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We adopt an alternative approach, wherein, rather than attempting to encompass all potential augmentations, we compel the model to leverage essential class-specific concepts for classification. Our key idea is to force the model to attend to these concepts during training. The primary hurdle in this path is lack of annotated data specifying relevant class-specific concepts along with their corresponding localizations. To this end, our first contribution is a novel pipeline that harnesses large language models (LLMs) and diffusion models (DMs) to identify key concepts and generate concept-level saliency maps in a scalable, automated fashion. We demonstrate that DMs, via cross-attention layers, can generate concept-level saliency maps for generated images with notable granularity. Given these maps for a single synthesized image within a class, we leverage the rich feature space of DMs to transfer them to images across diverse domains in DG benchmarks [47].", + "bbox": [ + 91, + 505, + 482, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We subsequently introduce TIDE, our second contribution which employs cross-entropy losses over class and concept labels along with a novel pair of loss functions: a concept saliency alignment loss that ensures the model attends to the correct regions for each local concept, and a local concept contrastive loss that promotes learning domain invariant features for these regions. As shown in Figure 1, TIDE consistently attends to local concepts such as beak and eyes in the training domain as well as across substan", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tial domain shifts. This precise localization significantly bolsters SSDG performance while enabling the generation of visual explanations for model decisions.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Marking a major leap in model interpretability and performance gains, our third contribution is to demonstrate how our model's concept-level localization can be effectively leveraged for performance verification and test-time correction. To this end, we introduce the notion of local concept signatures: prototypical features derived by pooling concept-level features across training samples, guided by corresponding saliency maps. If the concept features associated with the predicted class label do not align with their signatures, it signals the use of incorrect features for prediction. Consequently, we employ iterative refinement through concept saliency masking until concept predictions align with their corresponding signatures. More formally, our paper makes following contributions:", + "bbox": [ + 511, + 136, + 906, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel synthetic annotation pipeline to automatically identify key per-class local concepts and their localization maps, and transfer them to real images found in DG benchmarks.", + "- We propose a novel approach, TIDE, to train locally interpretable models with a novel concept saliency alignment loss that ensures accurate concept localization and a novel local concept contrastive loss that ensures domain-invariant features.", + "- We show TIDE enables correcting model prediction at test time using the predicted concept saliency maps as part of a novel iterative attention refinement strategy.", + "- We report extensive results on four standard DG benchmarks, outperforming the current state-of-the-art by a significant $12\\%$ on average." + ], + "bbox": [ + 513, + 349, + 903, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 588, + 661, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi Source domain generalization (MSDG) methods assume access to multiple source domains and domain-specific knowledge during training and have proposed ways to learn multi-source domain-invariant features [10, 15, 34, 35, 53], utilize domain labels to capture domain shifts [11, 35, 56], and design target-specific augmentation strategies [14, 16, 23]. However, these assumptions are not practical for real-world applications [44, 57] and we instead focus on the more challenging SSDG setting, where only one source domain and no prior target knowledge is available.", + "bbox": [ + 511, + 613, + 903, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most SSDG approaches have used augmentations to improve DG [5, 7, 8, 19, 20, 22, 59, 63]. While [59] also uses diffusion models (DMs) [42] for data augmentation, our approach differs by leveraging the rich feature space of DMs primarily for offline saliency map annotation instead of augmentation. More recent work in SSDG [4, 28, 29] has also turned to approaches based on domain-specific prompting and causal inference instead of data augmentations. However, these methods learn global features, limiting invari", + "bbox": [ + 511, + 765, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30211", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ba52611ef9468bac83a7dd8e044023b54f8e7859b587eb0effe60c809fe0d82a.jpg", + "image_caption": [ + "Figure 2. The TIDE pipeline: Left—Training on a single domain with cross-entropy losses for class $(\\mathcal{L}_{\\mathrm{c}})$ and concept labels $(\\mathcal{L}_{\\mathrm{k}})$ , alongside Concept Saliency Alignment $(\\mathcal{L}_{\\mathrm{CSA}})$ and Local Concept Contrastive losses $(\\mathcal{L}_{\\mathrm{LCC}})$ . Right—Test-time correction strategy applied in TIDE." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 571, + 327 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/02e40050d826f8975000ebd2ba02dd9c246c68c531b47cd493f08fb1304c7707.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 88, + 913, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ance across unseen domains. Finally, these methods also do not provide any signals to interpret model predictions.", + "bbox": [ + 88, + 393, + 482, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, while methods like Concept Bottle-neck Models (CBMs) [24, 33] also learn local concepts, they are restricted to a predefined set of concepts and more crucially, cannot ground them to image regions that these concepts represent. Recent methods [37, 58] have scaled CBMs to a large number of classes and also proposed to ways to learn multi-scale representations [52], but they are also unable to connect these concepts to image pixels. Our SSDG method addresses these challenges by producing local concept saliency maps, enhancing both DG performance and model prediction interpretability.", + "bbox": [ + 88, + 425, + 482, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 603, + 225, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed framework is depicted in Figure 2. Training comprises three components: class prediction, concept prediction with saliency alignment, and localized contrastive training. During inference, class and concept predictions are integrated with a test-time correction mechanism.", + "bbox": [ + 88, + 628, + 482, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the concept alignment phase, we employ concept-level saliency maps as ground truth to enforce focus on interpretable features, using a saliency alignment loss. Domain invariance is further promoted by localized contrastive training, where we train the model to cluster similar concepts (e.g., eyes across augmentations) while separating unrelated concepts (e.g., strings). Finally, a test-time correction mechanism iteratively refines attention by suppressing irrelevant regions, leveraging concept-signatures to redirect focus and improve classification accuracy. We proceed with a concise review of the key notations, followed by an in-depth exposition of our approach and the pipeline for generating ground truth concept-level annotations.", + "bbox": [ + 88, + 704, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assume the source data is observed from a single domain, denoted as $D = \\{x_{i},y_{i}^{c},y_{i}^{k}\\}_{i = 1}^{N}$ , where $x_{i}$ represents the $i$ -th image, $y_{i}^{c}$ the class label, $y_{i}^{k}$ the concept label, and $N$ the total sample count in the source domain. The shared CNN backbone is used to obtain the backbone features $F_{x}\\in \\mathbb{R}^{W\\times H\\times C}$ . The automatically generated ground truth concept-level saliency maps are denoted by $G_{x}^{k}$ . The GradCAM [43] saliency maps corresponding to class and concepts labels are denoted by $S_{x}^{c}$ and $S_{x}^{k}$ respectively.", + "bbox": [ + 511, + 393, + 906, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generating Concept-level Annotations", + "text_level": 1, + "bbox": [ + 511, + 542, + 843, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, we aim our model to identify and spatially attend to stable, discriminative regions. However, existing DG datasets lack fine-grained, concept-level annotations for such regions. To address this, we propose a novel pipeline that uses LLMs and DMs to automate scalable concept-level saliency map generation.", + "bbox": [ + 511, + 566, + 906, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our primary insight is that DMs can be harnessed to generate high-quality, concept-level saliency maps for synthesized images. Extracting cross-attention maps [1, 3] from a DM, given the prompt with concepts and corresponding synthesized image, yields highly granular concept-level saliency maps. Figure 3 demonstrates how DMs emphasize specific regions for distinct concepts, capturing fine-grained attention to features such as a cat's whiskers or a snowman's hands. With this technique serving as an efficient tool for yielding concept maps for synthesized images, the ensuing questions are: (i) how to automate the identification of pertinent concepts for each class across datasets, and (ii) how to transfer these concept maps to real-world images to generate ground-truth annotations.", + "bbox": [ + 511, + 657, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, to identify the key concepts associated with each class, we use GPT-3.5 [2], which we prompt to generate a", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30212", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6dee8d9d5cd5668554aad2e8dcca46d94b0fb597610b3ea26b16a7c15d8f1f03.jpg", + "image_caption": [ + "Prompt: A photo of a cat with its whiskers, ears, eyes" + ], + "image_footnote": [], + "bbox": [ + 106, + 109, + 196, + 176 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/775a048924f7abfb9f892c1dd2e4fb4897c56941137a006f6ef23b8a143c0ec1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 109, + 290, + 179 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0714bdac2ba4c5c47b5cdd01d830ea1ee5c52d117e8ee1c2a836c6efdd60de2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 109, + 380, + 179 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dfeeb94d1db27c4cd2d431223fb422eadf00e0b2602bbe1e19901e13eee3c9fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 109, + 472, + 179 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96d19e966bb870c786630b06fb675b87a2a88b9330d83bf68cc860e1c16b5304.jpg", + "image_caption": [ + "Prompt: A photo of a snowman with its nose, eyes, hands", + "Generated Image", + "Figure 3. The first column displays the image generated from the given prompt, while the subsequent three columns show the cross-attention maps corresponding to each concept in the prompt." + ], + "image_footnote": [], + "bbox": [ + 106, + 198, + 197, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da34b905aaa9a8c9d19e53c33627ff1dde704fe8d15c0194268a84b9ea363bac.jpg", + "image_caption": [ + "Extracted Cross Attention Maps" + ], + "image_footnote": [], + "bbox": [ + 199, + 198, + 290, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4013b509dcba52308e124b6a435d13726b3de40a64bbe458bdc5a901dea49aef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 198, + 380, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b580fdf126801013a7ade642429355bba544e9980ec0d010e213e7714fadbee3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 198, + 472, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "list of distinctive, stable features (prompt in the supplement) for each class. For instance, GPT-3.5 outputs concepts such as whiskers, ears, and eyes for a cat. We generate a prompt leveraging these concepts, which is then used to synthesize an exemplar image for each class. We further derive the corresponding concept-level attention maps as outlined in the preceding paragraph.", + "bbox": [ + 89, + 366, + 483, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a single synthesized exemplar image for each class and the concept-level saliency maps, we turn to the task of transferring these saliency annotations to real-world images. DM's feature space is particularly well-suited for this, as it captures detailed, semantically-rich representations that allow us to match synthetic concept regions to real images across domains. Leveraging this, we use the Diffusion Feature Transfer (DIFT) method [47], which computes pixel-level correspondences by comparing cosine similarities between DM features. Through this approach, we establish a region-to-region correspondence between synthetic saliency maps (e.g., mouth of a dog) and similar regions in real-world images. This enables us to generate comprehensive, consistent concept-level annotations across domains as shown in Figure 4.", + "bbox": [ + 89, + 474, + 483, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We obtain these concept-level annotations for widely-used benchmarks, including VLCS [13], PACS [27], DomainNet [38], and Office-Home [50]. For an image $x$ , the binarized concept-level maps $(G_x^k)$ are henceforth referred to as ground-truth saliency maps (GT-maps) in this paper. Having established a process for generating saliency maps, the next challenge is identifying the subset of concepts that the model actually relies on for making predictions.", + "bbox": [ + 89, + 700, + 483, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1 Concepts that matter", + "text_level": 1, + "bbox": [ + 89, + 845, + 292, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We restrict our method to essential concepts, filtering out those that do not contribute meaningfully to classification.", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d29618652d5205159ff10bac791cd7efd31973a16d30ee4a603223deba797526.jpg", + "image_caption": [ + "Prompt: A photo of a dog with its ear, mouth", + "PACS - photo", + "PACS - cartoon", + "PACS - sketch" + ], + "image_footnote": [], + "bbox": [ + 560, + 104, + 849, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c4fd72ee987295925f35d21eb0276da075ac6daff6707f332d502e2949c671d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 185, + 627, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac16b1af529de362cdc7f19c2931db27632963bf5dd3b7c00a189aef91fbe3c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 186, + 714, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bbfdcee532654e85acc9ba737890fcdaff6c6ee34032e52ad2f177d07cbe6c6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 186, + 792, + 246 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0174a03f238e6b72fd49c175772434e24b00fed9a966a9b37d62cc100e8fe34.jpg", + "image_caption": [ + "PACS - painting" + ], + "image_footnote": [], + "bbox": [ + 805, + 185, + 890, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ca02fb103336e15dc0db5d2712f1727c598bc38530be48221d2c73021b5ccc9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 251, + 627, + 315 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2ccd2e6a534ce7fb0dfd0e77432ba5f1959acd83830a4b60d7f77b9b33f89ca7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 251, + 712, + 315 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6c8e2f311ae85750fa4a8e75986439e9f999e7c7caf82d8750245b52e6008117.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 251, + 802, + 315 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e0c294a9a6a384faf3dd8902748b24a326c7d80b70053bda84b09b0f24efa158.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 251, + 888, + 315 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec4b65144e3a1a5f2260927e76428cfde59845b800810e3e89f71740f168cd5f.jpg", + "image_caption": [ + "Figure 4. The first row presents the prompt, corresponding synthesized image and attention maps for the concepts ear and mouth. Below, we demonstrate that using diffusion features correspondences these concept saliency maps from a single exemplar can be automatically transferred on dog images across domains." + ], + "image_footnote": [], + "bbox": [ + 531, + 316, + 627, + 381 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/57f698054fc670394c3daaf84b776d1c645e33b332999a0825c0d83fcc7ca8b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 316, + 712, + 381 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e784e8684e0a9adea66fdc7a28ee34d59cfe64b8ce704ea733eee31c78f993b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 316, + 802, + 381 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/625ec29125ed796b939c09db44c76f063b1c54e446a2f046a295a978a76017e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 316, + 890, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To do this, we train a ResNet-18 [18] classifier on the source domain [26], compute GradCAM [43] saliency maps for each class label, and use them to identify regions the model focuses on when making predictions.", + "bbox": [ + 511, + 489, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compute the overlap between the GradCAM maps and GT-maps for each known concept in the dataset. Given image $x$ , its saliency map $S_{x}^{c}$ for class $c$ and the GT-map $G_{x}^{k}$ for concept $k$ , we define the overlap $O_{c}^{k}(x)$ as:", + "bbox": [ + 511, + 551, + 905, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO _ {c} ^ {k} (x) = \\frac {\\sum_ {i , j} \\min \\left(S _ {x} ^ {c} (i , j) , G _ {x} ^ {k} (i , j)\\right)}{\\sum_ {i , j} G _ {x} ^ {k} (i , j)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 625, + 905, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(i,j)$ are matrix indices. This measures how much of the model's attention for a given class aligns with the regions corresponding to concept $k$ . We compute this overlap for all concepts and images in the training set, and for each class $c$ , we define the set of important concepts $\\mathcal{K}_c$ as those that consistently exhibit high overlap with the saliency maps:", + "bbox": [ + 511, + 671, + 906, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} _ {c} = \\left\\{k \\left| \\frac {1}{N _ {c}} \\sum_ {x \\in \\mathcal {D} _ {c}} O _ {c} ^ {k} (x) > \\tau \\right. \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 790, + 905, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{c}$ is the number of training examples in class $c$ , $\\mathcal{D}_c$ is the set of images in class $c$ , and $\\tau$ is a threshold that determines the importance of the concept. This procedure functions as a concept discovery module, identifying which", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30213", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "local concepts are critical for predicting each class.", + "bbox": [ + 89, + 90, + 428, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. TIDE", + "text_level": 1, + "bbox": [ + 89, + 130, + 171, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, TIDE utilizes cross-entropy losses $\\mathcal{L}_c$ and $\\mathcal{L}_k$ for class and concept labels, respectively, integrated with novel concept saliency alignment and local concept contrastive losses. We detail these loss terms below.", + "bbox": [ + 89, + 154, + 483, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Concept saliency alignment loss", + "text_level": 1, + "bbox": [ + 89, + 233, + 366, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each image $x$ , we predict important concepts $\\mathcal{K}_c$ and enforce alignment between the saliency maps $S_x^k$ for predicted concepts and the GT-maps $G_x^k$ . This is encouraged by our proposed concept saliency alignment (CSA) loss:", + "bbox": [ + 89, + 257, + 483, + 318 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C S A}} = \\frac {1}{| \\mathcal {K} _ {c} |} \\sum_ {k \\in \\mathcal {K} _ {c}} \\| S _ {x} ^ {k} - G _ {x} ^ {k} \\| _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 328, + 480, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aligning the model's attention with GT-maps enables class-specific reasoning, thereby elucidating the rationale behind predictions by linking them to relevant local features.", + "bbox": [ + 89, + 371, + 483, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Local concept contrastive loss", + "text_level": 1, + "bbox": [ + 89, + 450, + 352, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the CSA loss facilitates explicit localization of concepts, it is equally essential for these concept-level features to exhibit invariance across domain shifts. To achieve domain invariance, we propose a local concept contrastive (LCC) loss, employing a triplet strategy to cluster similar concepts (e.g., eyes) across domains while distinguishing unrelated ones (e.g., feathers and ears).", + "bbox": [ + 89, + 474, + 483, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $x$ be an anchor image containing concept $k$ , a positive image $x^{+}$ (an augmentation of $x$ that retains concept $k$ ), and a negative image $x^{-}$ (containing a different concept $k'$ ). For each image, we compute a concept-specific feature vector $f_{x}^{k} \\in \\mathbb{R}^{C}$ , emphasizing the concept's relevant regions using the GT-map $G_{x}^{k}$ . Each element $f_{x}^{k}(l)$ is computed as:", + "bbox": [ + 89, + 580, + 483, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {x} ^ {k} (l) = \\sum_ {i, j} G _ {x} ^ {k} \\cdot F _ {x} (i, j, l). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 681, + 480, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cdot$ is an element-wise multiplication operation, $G_{x}^{k}$ is a matrix representing the GT-map, $F_{x}$ represents convolutional feature maps from the last convolution layer, and $(i,j,l)$ denotes the $l^{th}$ channel's $(i,j)$ element. These vectors focus on the concepts of interest (concept $k$ for $x$ and $x^{+}$ , $k^{\\prime}$ for $x^{-}$ ), in contrast to the global features used in prior works. The LCC loss is then defined as:", + "bbox": [ + 89, + 723, + 483, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L C C}} = \\max \\left(0, d \\left(f _ {x} ^ {k}, f _ {x ^ {+}} ^ {k}\\right) - d \\left(f _ {x} ^ {k}, f _ {x ^ {-}} ^ {k ^ {\\prime}}\\right) + \\alpha\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 854, + 480, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(.)$ is the euclidean distance and $\\alpha$ is the margin.", + "bbox": [ + 109, + 885, + 480, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Iterative Test-time Correction" + ], + "code_body": "Input: Test image $x$ , initial class prediction $c^{(0)}$ , signatures $p^k$ for each concept $k$ \nParameter: Threshold $\\delta$ , max iterations $T$ \nOutput: Corrected class prediction, $c_{\\mathrm{final}}$ \n1: $S_x^{c(0),0} \\leftarrow \\operatorname{GradCAM}(x, c^{(0)})$ {Initialize saliency map for iteration $t = 0$ } \n2: $x_{\\mathrm{masked}}^{0} \\leftarrow x$ \n3: for each $k \\in \\mathcal{K}_{c(0)}$ do \n4: $f_x^{k,0} \\leftarrow \\sum_{i,j} S_x^k \\cdot F_x(i, j, :)$ \n5: $d(f_x^{k,0}, p^k) \\leftarrow 1 - \\frac{f_x^{k,0} \\cdot p^k}{\\|f_x^{k,0}\\| \\|p^k\\|}$ \n6: if $d(f_x^{k,0}, p^k) > \\delta$ then \n7: Correction Phase: \n8: for $t \\leftarrow 1$ to $T$ do \n9: $M^{c(t-1),t} \\leftarrow \\text{Binarize}(S_x^{c(t-1),t-1})$ \n10: $x_{\\mathrm{masked}}^t \\leftarrow x_{\\mathrm{masked}}^{t-1} \\cdot M^{c(t-1),t}$ \n11: $c^{(t)} \\leftarrow \\text{PredictClass}(x_{\\mathrm{masked}}^t)$ \n12: $S_{x_{\\mathrm{masked}}}^{c(t),t} \\leftarrow \\text{GradCAM}(x_{\\mathrm{masked}}^t, c^{(t)})$ \n13: for each $k \\in \\mathcal{K}_{c(t)}$ do \n14: $f_{x_{\\mathrm{masked}}}^{k,t} \\leftarrow \\sum_{i,j} S_{x_{\\mathrm{masked}}}^{k,t} \\cdot F_x(i,j,:)$ \n15: $d(f_{x_{\\mathrm{masked}}}^{k,t}, p^k) \\leftarrow 1 - \\frac{f_{x_{\\mathrm{masked}}}^{k,t} \\cdot p^k}{\\|f_{x_{\\mathrm{masked}}}^{k,t}\\| \\|p^k\\|}$ \n16: if $d(f_{x_{\\mathrm{masked}}}^{k,t}, p^k) \\leq \\delta$ then \n17: Return $c_{\\mathrm{final}} \\leftarrow c^{(t)}$ \n18: end if \n19: end for \n20: end for \n21: end if \n22: Return $c_{\\mathrm{final}} \\leftarrow c^{(0)}$ \n23: end for", + "bbox": [ + 513, + 107, + 903, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Test-time Correction", + "text_level": 1, + "bbox": [ + 513, + 508, + 707, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We establish that our localized, interpretable approach facilitates correction of misclassifications through concept-level feature verification. In this section, we first introduce concept signatures and detail the proposed correction strategy.", + "bbox": [ + 511, + 532, + 905, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Local concept signatures", + "text_level": 1, + "bbox": [ + 511, + 613, + 740, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each concept $k$ , we define a concept-signature $p^k \\in \\mathbb{R}^C$ , as its representative vector. We derive $p^k$ by averaging the concept-specific feature vectors $f_x^k$ across all training samples $x \\in \\mathcal{D}$ containing concept $k$ (denoted as $\\mathcal{D}^k$ ).", + "bbox": [ + 511, + 636, + 905, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np ^ {k} = \\frac {1}{\\left| \\mathcal {D} ^ {k} \\right|} \\sum_ {x \\in \\mathcal {D} ^ {k}} f _ {x} ^ {k}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 708, + 903, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These vectors act as reference points, helping the model recognize when its attention is aligned with the right concepts during prediction, even when encountering new, unseen domains.", + "bbox": [ + 511, + 751, + 905, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Detecting and Correcting misclassifications", + "text_level": 1, + "bbox": [ + 511, + 830, + 867, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider the example of test-time correction in Figure 2, where the model misclassifies a person as a guitar. The predicted class guitar, involves concepts like strings and", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30214", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "knobs, but these are absent in the image. As a result, the model erroneously focuses on irrelevant features (e.g., the person's legs or the background), as reflected in the Grad-CAM saliency maps.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address misclassifications, we employ a two-step approach: possible mistake detection followed by correction. The process is outlined in Algorithm 1. First, the model extracts concept-level saliency maps $(S_x^k)$ , for all the concepts corresponding to the predicted class, which are then used to compute concept-level features $f_{x}^{k}$ (step 4). In Step 5 and 6, we compare the features to the stored concept signatures; a deviation exceeding $\\delta$ signals a misclassification.", + "bbox": [ + 89, + 152, + 480, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once a misalignment is detected, the model enters an iterative refinement phase. The features prominent for the current class level predictions are masked through the corresponding saliency map $S_{x}^{c}$ (step 9 and 10). The masking process is cumulative, in effect utilizing all masks from the first to the $t^{th}$ iteration. The masked features are then used for subsequent predictions (step 11). This process continues until the concepts corresponding to the predicted class aligns with concept-signatures or the maximum iteration count is reached (steps 8-16 in the algorithm). If alignment is achieved, the predicted class in that iteration is confirmed as the final output.", + "bbox": [ + 89, + 272, + 480, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 467, + 222, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We conduct experiments on four widely used datasets - PACS [27], VLCS [13], OfficeHome [50], and DomainNet [38], within the DomainBed [17] evaluation benchmarks. PACS contains 9,991 images across 7 categories in four domains: 'sketch', 'photo', 'clipart', and 'painting'. VLCS consists of 10,729 images over 5 categories and 4 domains. The domains are from VOC2007 (V), LabelMe (L), Caltech101 (C), and SUN09 (S), with domain shifts primarily driven by background and viewpoint variations. OfficeHome, with 15,500 images across 65 categories, emphasizes indoor classes across 'product', 'real', 'clipart', and 'art' domains. For DomainNet, we follow prior work [39, 46] and use a subset of the 40 most common classes across 'sketch', 'real', 'clipart', and 'painting'.", + "bbox": [ + 89, + 492, + 482, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Setup: We adhere to the SSDG paradigm, training on one source domain and testing across three target domains, i.e., the model is trained independently four times, with the averaged test accuracies across target domains reported in each case. Training utilizes solely source-domain GT-maps, excluding any target-domain prior knowledge, data or annotations. Minimal augmentations (quantization, blurring, and canny edge), were used to introduce slight perturbations to create the triplets. We use SDv2.1 [42] for generating exemplar images and computing cross-attention maps [1, 3]. To ensure a fair", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "comparison, we adopt a ResNet-18 backbone throughout all the experiments. We use the Adam optimizer with an initial learning rate of $1 \\times 10^{-4}$ and a warm-up schedule over the first 1000 steps, after which the rate remains constant. Margin value $\\alpha$ is set to 1.0. The batch size is set to 32. We empirically set $\\delta = 0.1$ and cap test-time correction at $T = 10$ iterations.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared Approaches: We compare our method with ERM [17] baseline and existing approaches that utilize augmentation-based techniques (NJPP [59], AugMix [19], MixStyle [63], CutMix [60], RandAugment [7]), self-supervised and domain adaptation methods (RSC [22], pAdaIn [36], L2D [55], RSC+ASR [12]), uncertainty modeling (DSU [29], DSU-MAD [41]), attention and meta-learning methods (ACVC [8], P-RC [6], Meta-Casual [4]), and prompt-based learning (PromptD [28]). The number of methods evaluated differs by dataset. The discrepancy stems from PACS being the dominant benchmark in prior SSDG works, resulting in a larger number of methods evaluated on it. For VLCS, OfficeHome, and DomainNet, we rely on reported results from respective papers, or compute them ourselves (where code was available) to ensure a fair comparison.", + "bbox": [ + 511, + 212, + 903, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 511, + 464, + 651, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 provide a comparison of average classification accuracy for our approach against existing methods across the PACS, VLCS, OfficeHome, and DomainNet datasets. Each column in these tables represents a source domain used for training, with the numerical values indicating the average accuracy on the three target domains. The last column presents the average of the four columns.", + "bbox": [ + 511, + 488, + 903, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TIDE decisively outperforms existing approaches across all datasets. It achieves average accuracy gains of $8.33\\%$ , $13.37\\%$ , $16.16\\%$ , and $8.84\\%$ over the second-best approach on PACS, VLCS, OfficeHome, and DomainNet, respectively. Table 1b details the performance on VLCS, where domain shifts primarily stem from background and viewpoint variations, with scenes spanning urban to rural and favoring non-standard viewpoints. Data augmentation based methods, which focus on style variation, yield limited gains on VLCS relative to PACS. Nonetheless, TIDE secures substantial improvements, notably achieving a $25.34\\%$ gain over DSU-MAD when Caltech101 (C) serves as the source domain. Another noteworthy observation is that TIDE maintains strong performance across varying class counts, from VLCS's 5 to OfficeHome's 65, underscoring the scalability of the proposed approach.", + "bbox": [ + 511, + 595, + 903, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablations", + "text_level": 1, + "bbox": [ + 511, + 845, + 624, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Components of TIDE: Ablation experiments on the PACS dataset are conducted to assess the contribution of each loss", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30215", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/70f773775a09014de7ca7460dae570c44e486fe41818dd64492095b06357c354.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method VenueArtCartoonSketchPhotoAverage
ERM65.3864.2034.1533.6549.35
Augmix [19] ICLR'2166.5470.1652.4838.3057.12
RSC [22] ECCV'2073.4075.9056.2041.6061.03
Mixstyle [63] ICLR'2167.6070.3834.5737.4452.00
pAdalin [36] CVPR'2164.9665.2432.0433.6649.72
RSC+ASR [12] CVPR'2176.7079.3061.6054.6068.30
L2D [55] ICCV'2176.9177.8853.6652.2965.93
DSU [29] ICLR'2271.5474.5147.7542.1058.73
ACVC [8] CVPR'2273.6877.3955.3048.0563.10
DSU-MAD [41] CVPR'2372.4174.4749.6044.1560.66
P-RC [6] CVPR'2376.9878.5462.8957.1168.88
Meta-Casual [4] CVPR'2377.1380.1462.5559.6069.86
ABA [5] ICCV'2375.6977.3654.1259.0466.30
PromptD [28] CVPR'2478.7782.6962.9460.0971.87
TIDE86.2486.3773.1174.3680.02
", + "bbox": [ + 101, + 89, + 472, + 271 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dc38cae5c7522c4ca67180491611d7c388fd255d352ee869f59db766707eebab.jpg", + "table_caption": [ + "(a) PACS" + ], + "table_footnote": [], + "table_body": "
MethodVLCSAverage
ERM76.7258.8644.9557.7159.06
Augmix [19] ICLR'1975.2559.5245.9057.4359.03
pAdalin [36] CVPR'2176.0365.2143.1757.9460.34
Mixstyle [63] ICLR'2175.7361.2944.6656.5759.06
ACVC [8] CVPR'2276.1561.2347.4360.1861.75
DSU [29] ICLR'2276.9369.2046.5458.3662.11
DSU-MAD [41] CVPR'2376.9970.8544.7862.2363.71
TIDE82.6286.1370.1272.4477.08
", + "bbox": [ + 109, + 292, + 464, + 405 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d1129382925a184262ce403b7168c40a09e6226e8556bf3a78481bdd20afd982.jpg", + "table_caption": [ + "(b) VLCS" + ], + "table_footnote": [], + "table_body": "
MethodArtClipartProductRealAverage
ERM57.4350.8348.958.6853.96
MixUp [21] ICLR'1850.4143.1941.2451.8946.93
CutMix [60] ICCV'1949.1746.1541.253.6447.04
Augmix [19] ICLR'1956.8654.1252.0260.1256.03
RandAugment [7] CVPRW'2058.0755.3252.0260.8256.56
CutOut [9] arXiv:170854.3650.7947.6858.2452.77
RSC [22] ECCV'2053.5148.9847.1658.352.73
MEADA [62] NIPS'2057.053.248.8159.2154.80
PixMix [20] CVPR'2253.7752.6848.9158.6853.51
L2D [55] ICCV'2152.7948.9747.7558.3151.71
ACVC [8] CVPR'2254.351.3247.6956.2552.89
NJPP [59] ICML'2460.7254.9552.4761.2657.85
TIDE72.3275.1375.2273.3774.01
", + "bbox": [ + 94, + 425, + 478, + 589 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/50a487e0e0675ddab2705667d1b39c1a76b7ee11a926cc9185d3f3fb8085e069.jpg", + "table_caption": [ + "(c) OfficeHome" + ], + "table_footnote": [], + "table_body": "
MethodClipartPaintingRealSketchAverage
ERM68.7366.1268.5169.4468.2
MixUp [21] ICLR'1870.3164.3469.2168.8268.17
CutMix [60] ICCV'1971.5263.8467.1369.4167.98
Augmix [19] ICLR'1972.3762.9169.8471.2269.09
RandAugment [7] CVPRW'2069.7165.5168.3666.9367.63
CutOut [9] arXiv:170870.8664.4869.9271.5569.20
RSC [22] ECCV'2068.2567.9170.7666.1868.28
PixMix [20] CVPR'2272.1263.5171.3465.4668.12
NJPP [59] ICML'2476.1469.2476.6171.2173.3
TIDE82.4280.3784.1581.6182.14
", + "bbox": [ + 91, + 609, + 483, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "term on the overall performance. In Table 2, we present classification accuracy by incrementally adding components to the training pipeline, demonstrating the contribu", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aba6a76f1b428820aa02fa98d3b3301c3ab352803dc6442aef586924e2d1885d.jpg", + "table_caption": [ + "(d) DomainNet", + "Table 1. SSDG classification accuracies $(\\%)$ on PACS, VLCS, OfficeHome and DomainNet datasets, with ResNet-18 as backbone. Each column title indicates the source domain, and the numerical values represent the average performance in the target domains." + ], + "table_footnote": [], + "table_body": "
MethodArtCartoonSketchPhotoAverage
Lc65.3864.2034.1533.6549.35
Lc + Lk65.4764.1235.4433.8149.71
Lc + Lk + LCSA66.1164.2335.1434.2749.93
Lc + Lk + LCSA + LLC80.2882.9166.1265.3773.67
+ test-time correction86.2486.3773.1174.3680.02
", + "bbox": [ + 514, + 89, + 900, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study $(\\%)$ on PACS.", + "bbox": [ + 594, + 167, + 823, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/699c6c35be7a518f1f679236ea394c8cf9de202589adb6523fefe6cad8ba8829.jpg", + "image_caption": [ + "Figure 5. t-SNE visualizations to demonstrate impact of $\\mathcal{L}_{LCC}$ . Each column represents a test domain (Sketch, Cartoon, Painting), with the top row showing t-SNE plots without $\\mathcal{L}_{LCC}$ applied and the bottom one with it. Please zoom in for optimal viewing." + ], + "image_footnote": [], + "bbox": [ + 519, + 195, + 903, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion of each element to the model's performance. We observe that the introduction of the concept classification loss $\\mathcal{L}_k$ and concept saliency alignment loss $\\mathcal{L}_{CSA}$ does not significantly affect the classifier's accuracy on the test domain. However, these components are integral to our approach, enabling test-time correction and, in turn, enhancing both performance and model interpretability. We observe that the introduction of the $\\mathcal{L}_{LCC}$ loss leads to a substantial increase in test domain accuracy, clearly demonstrating its effectiveness in fostering domain invariance. Finally, we compare results before and after test-time correction, highlighting that even prior to correction, our method achieves state-of-the-art performance across all source domains on PACS, underscoring the strength of our approach.", + "bbox": [ + 511, + 507, + 906, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Error analysis of test-time correction: For this analysis, we train TIDE on the photo domain of PACS and test it on the sketch domain. The model gives an initial accuracy of $74.79\\%$ , which improves to $82.29\\%$ post test-time correction. In $72.2\\%$ of test samples, TIDE does not invoke test-time correction, with class predictions correct in $93.8\\%$ of these cases, demonstrating substantial reliability. The supplement provides examples of cases where the model misclassifications are not picked-up in the signature matching step of test-time correction. In the remaining $27.8\\%$ of cases where correction is initiated, $52.5\\%$ successfully converge to the correct classification, significantly contributing", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30216", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/beb08761c29002c92cfe81769a494f16d1a29a5c849c806991c0f9e7b5e0d0ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 156, + 90, + 325, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/613dd7df96c68fcd30a263262e6d4f50a52bb68486e394a4c858294a92297ac6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 90, + 496, + 171 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f4944b2c14e8b7d3f2cbed3b0366ab4b9cacb1de3ef1aafdaf31225282b744e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 90, + 669, + 171 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4c57f73735f26000c245e6994b65cd667fa41c932c183826dd4f89808fe1caf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 90, + 846, + 171 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/808f78ef24d3036fe78e68aad451302f96ebcec21b4656b09c56456f7e7aed9c.jpg", + "image_caption": [ + "Figure 6. Illustrative examples of concept level GradCAM maps corresponding to TIDE's predictions, across the four studied datasets. The concept names are displayed beneath the maps, with the target domain and predicted class indicated on the left. More results in supplement." + ], + "image_footnote": [], + "bbox": [ + 156, + 174, + 325, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0db9325e759eba3d4f641f5b03044feae90ae56d5b10cebdd0aacac945a4edc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 174, + 496, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/045e375c81bf497dad55db2b8779e6d65bdbf75438dc3c6a669be30b8eda9de1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 172, + 669, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b6e65d78a9c46f055e0f6746510e9efe88099346f2f130dc2ae4f02d894a600.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 172, + 844, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/70d048ca715ecbc489e4a59851919a23d6e45cd3b26264d1fedc0ff1b06b6dcb.jpg", + "image_caption": [ + "Figure 7. The top row shows initial class predictions and GradCAM maps for the concepts, while the bottom row presents the results after test-time correction." + ], + "image_footnote": [], + "bbox": [ + 122, + 315, + 454, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to the final post-correction accuracy of $82.29\\%$ .", + "bbox": [ + 89, + 566, + 406, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "t-SNE plots: To demonstrate the impact of $\\mathcal{L}_{LCC}$ , we present t-SNE visualizations in Figure 5, showing the distribution of concept-specific vectors (as computed in Equation 4). We individually plot the source domain (photo) along with sketch, cartoon, and painting domains as targets. Each case includes two plots: one without $\\mathcal{L}_{LCC}$ (top row) and one with it (bottom row). We observe that with $\\mathcal{L}_{LCC}$ , concept samples (e.g., mouth) align closely across domains, while distinct separation occurs between different concepts (e.g., mouth and horns). Without $\\mathcal{L}_{LCC}$ , cluster separation and alignment across domains are weaker, highlighting its role in improving intra-concept compactness and inter-concept separability.", + "bbox": [ + 89, + 580, + 483, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 786, + 272, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concept Localization: We train the SSDG model on photo domain for PACS, OfficeHome, and DomainNet and on Caltech101 for VLCS. The predicted class, concepts and corresponding saliency maps are shown in Figure 6. For each target class (e.g., mushroom, kettle, guitar), the model reliably highlights key concept-specific regions (e.g.,", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "stem, handle, strings) essential for classification. The model effectively isolates key features across diverse contexts, such as the legs and seat of a chair in complex zoomed-out scenes, the beak and feathers of camouflaged birds, the eyes and whiskers in deformed sketches of cats, and varying feature sizes, with the wick occupying a small area and the mushroom cap spanning a larger one.", + "bbox": [ + 511, + 315, + 906, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Test-time Correction: The Figure 7 visually illustrates the efficacy of TIDE's test time correction abilities, by comparing the initial and corrected results. The model initially mis-classifies the images as dog and elephant, resulting in poorly aligned concept-level saliency maps for corresponding concepts i.e. ears, snout, tusks and trunk. TIDE detects and rectifies such errors during inference, leading to accurate classification accompanied by precise concept localization. This qualitative evaluation reinforces the robustness of our approach in refining predictions and generating reliable, class-specific concept maps.", + "bbox": [ + 511, + 436, + 908, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 618, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we considered the problem of single-source domain generalization and observed that the current state-of-the-art methods fail in cases of semantic domain shifts primarily due to the global nature of their learned features. To alleviate this issue, we proposed TIDE, a new approach that not only learns local concept representations but also produces localization maps for these concepts. With these maps, we showed we could visually interpret model decisions while also enabling correction of these decisions at test time using our iterative attention refinement strategy. Extensive experimentation on standard benchmarks demonstrated substantial and consistent performance gains over the current state-of-the-art. Future work will rigorously explore methods to generate confidence scores grounded in TIDE's concept verification strategy and would explore the application of TIDE on other (e.g. fine-grained) classification datasets.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30217", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In International Conference on Computer Vision (ICCV), 2023. 3, 6", + "[2] Tom B Brown. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 2020. 3", + "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 3, 6", + "[4] Jin Chen, Zhi Gao, Xinxiao Wu, and Jiebo Luo. Meta-causal learning for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7", + "[5] Sheng Cheng, Tejas Gokhale, and Yezhou Yang. Adversarial bayesian augmentation for single-source domain generalization. In International Conference on Computer Vision (ICCV), 2023. 1, 2, 7", + "[6] Seokeon Choi, Debasmit Das, Sungha Choi, Seunghan Yang, Hyunsin Park, and Sungrack Yun. Progressive random convolutions for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6, 7", + "[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020. 1, 2, 6, 7", + "[8] Ilke Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 7", + "[9] Terrance Devries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. ArXiv, abs/1708.04552, 2017. 7", + "[10] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In European Conference on Computer Vision (ECCV), 2020. 2", + "[11] Antonio D'Innocente and Barbara Caputo. Domain generalization with domain-specific aggregation modules. In The German Conference on Pattern Recognition, 2019. 2", + "[12] Xinjie Fan, Qifei Wang, Junjie Ke, Feng Yang, Boqing Gong, and Mingyuan Zhou. Adversarily adaptive normalization for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6, 7", + "[13] Chen Fang, Ye Xu, and Daniel N Rockmore. Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 4, 6", + "[14] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel." + ], + "bbox": [ + 91, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations (ICLR), 2019. 2", + "[15] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In International Conference on Computer Vision (ICCV), 2015. 2", + "[16] Tejas Gokhale, Rushil Anirudh, Bhavya Kailkhura, Jayaraman J Thiagarajan, Chitta Baral, and Yezhou Yang. Attribute-guided adversarial training for robustness to natural perturbations. In Association for the Advancement of Artificial Intelligence (AAAI), 2021. 2", + "[17] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. International Conference on Learning Representations (ICLR), 2020. 6", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4", + "[19] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. International Conference on Learning Representations (ICLR), 2020. 1, 2, 6, 7", + "[20] Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. Pixmix: Dreamlike pictures comprehensively improve safety measures. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7", + "[21] Yann N. Dauphin David Lopez-Paz Hongyi Zhang, Moustapha Cisse. mixup: Beyond empirical risk minimization. International Conference on Learning Representations (ICLR), 2018. 7", + "[22] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 6, 7", + "[23] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in Neural Information Processing Systems (NeurIPS), 2015. 2", + "[24] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning (ICML), 2020. 3", + "[25] Wouter M Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 43(3):766-785, 2019. 1", + "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2012. 4", + "[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In International Conference on Computer Vision (ICCV), 2017. 4, 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30218", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Deng Li, Aiming Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 6, 7", + "[29] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. International Conference on Learning Representations (ICLR), 2022. 2, 6, 7", + "[30] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In European Conference on Computer Vision (ECCV), 2018. 1", + "[31] Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. ArXiv, abs/2108.13624, 2021. 1", + "[32] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning (ICML), 2015. 1", + "[33] Andrei Margeloiu, Matthew Ashman, Umang Bhatt, Yanzhi Chen, Mateja Jamnik, and Adrian Weller. Do concept bottleneck models learn as intended? International Conference on Learning Representations (ICLR), 2021. 3", + "[34] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In International Conference on Computer Vision (ICCV), 2017. 2", + "[35] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning (ICML), 2013. 2", + "[36] Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6, 7", + "[37] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. International Conference on Learning Representations (ICLR), 2023. 3", + "[38] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In International Conference on Computer Vision (ICCV), 2019. 4, 6", + "[39] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation. In European Conference on Computer Vision (ECCV), 2024. 6", + "[40] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1", + "[41] Sanqing Qu, Yingwei Pan, Guang Chen, Ting Yao, Changjun Jiang, and Tao Mei. Modality-agnostic debiasing for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In International Conference on Computer Vision (ICCV), 2017. 1, 2, 3, 4", + "[44] Sarath Sivaprasad, Akshay Goindani, Mario Fritz, and Vineet Gandhi. Class-wise domain generalization: A novel framework for evaluating distributional shift. In NeurIPS Workshop on Distribution Shifts: Connecting Methods and Applications, 2022. 1, 2", + "[45] Nathan Somavarapu, Chih-Yao Ma, and Zsolt Kira. Frustratingly simple domain generalization via image stylization. ArXiv, abs/2006.11207, 2020. 2", + "[46] Shuhan Tan, Xingchao Peng, and Kate Saenko. Class-imbalanced domain adaptation: An empirical odyssey. In European Conference on Computer Vision Workshops (ECCVW), 2020. 6", + "[47] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 4", + "[48] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2", + "[49] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1", + "[50] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 6", + "[51] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in Neural Information Processing Systems (NeurIPS), 2018. 1", + "[52] Bor-Shiun Wang, Chien-Yi Wang, and Wei-Chen Chiu. Mcpnet: An interpretable classifier via multi-level concept prototypes. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[53] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. International Conference on Learning Representations (ICLR), 2019. 2", + "[54] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 1", + "[55] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In International Conference on Computer Vision (ICCV), 2021. 6, 7", + "[56] Zehao Xiao, Jiayi Shen, Xiantong Zhen, Ling Shao, and Cees Snoek. A bit more bayesian: Domain-invariant learning with uncertainty. In International Conference on Machine Learning (ICML), 2021. 2" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30219", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain expansion approach for single-source domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[58] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[59] Jianhao Yuan, Francesco Pinto, Adam Davies, and Philip Torr. Not just pretty pictures: Toward interventional data augmentation using text-to-image generators. International Conference on Machine Learning (ICML), 2024. 1, 2, 6, 7", + "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision (ICCV), 2019. 6, 7", + "[61] Hanlin Zhang, Yi-Fan Zhang, Weiyang Liu, Adrian Weller, Bernhard Schölkopf, and Eric P Xing. Towards principled disentanglement for domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[62] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 7", + "[63] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. International Conference on Learning Representations (ICLR), 2021. 1, 2, 6, 7", + "[64] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation. International Journal of Computer Vision (IJCV), 132(3):822-836, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 585 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "30220", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_model.json b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..21d51b528081a0dbad5adb759f33b111d332b9f0 --- /dev/null +++ b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_model.json @@ -0,0 +1,2686 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.13, + 0.85, + 0.175 + ], + "angle": 0, + "content": "TIDE: Training Locally Interpretable Domain Generalization Models Enables Test-time Correction" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.203, + 0.765, + 0.223 + ], + "angle": 0, + "content": "Aishwarya Agarwal1,2* Srikrishna Karanam2† Vineet Gandhi1‡" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.23, + 0.7, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\)CVIT, Kohli Centre for Intelligent Systems, IIIT Hyderabad, India" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.248, + 0.604, + 0.264 + ], + "angle": 0, + "content": "\\(^{2}\\)Adobe Research, Bengaluru, India" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.299, + 0.327, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.332, + 0.486, + 0.74 + ], + "angle": 0, + "content": "We consider the problem of single-source domain generalization. Existing methods typically rely on extensive augmentations to synthetically cover diverse domains during training. However, they struggle with semantic shifts (e.g., background and viewpoint changes), as they often learn global features instead of local concepts that tend to be domain invariant. To address this gap, we propose an approach that compels models to leverage such local concepts during prediction. Given no suitable dataset with per-class concepts and localization maps exists, we first develop a novel pipeline to generate annotations by exploiting the rich features of diffusion and large-language models. Our next innovation is TIDE, a novel training scheme with a concept saliency alignment loss that ensures model focus on the right per-concept regions and a local concept contrastive loss that promotes learning domain-invariant concept representations. This not only gives a robust model but also can be visually interpreted using the predicted concept saliency maps. Given these maps at test time, our final contribution is a new correction algorithm that uses the corresponding local concept representations to iteratively refine the prediction until it aligns with prototypical concept representations that we store at the end of model training. We evaluate our approach extensively on four standard DG benchmark datasets and substantially outperform the current state-of-the-art (12% improvement on average) while also demonstrating that our predictions can be visually interpreted." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.768, + 0.222, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.484, + 0.855 + ], + "angle": 0, + "content": "Enhancing deep neural networks to generalize to out-of-distribution samples remains a core challenge in machine learning and computer vision research, as real-world test data often diverges significantly from the training distribu" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.301, + 0.916, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.587 + ], + "angle": 0, + "content": "Figure 1. Samples from VLCS (left) and PACS dataset (right) across domain shifts, corresponding to bird and person class. First row displays GradCAM maps [43] for ABA's class predictions. We observe that model attention of ABA [5] falters across domain shifts. Second and third row display the concept specific GradCAM maps from TIDE. We posit that accurate concept learning and localization facilitates DG." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.825 + ], + "angle": 0, + "content": "tion [31]. The challenges compound when obtaining labeled samples from the target domain is expensive or unfeasible, hindering application of semi-supervised learning or domain adaptation [25, 32, 49]. The problem of Domain Generalization (DG) [30, 44, 54, 61, 63] represents a promising avenue for developing techniques that capture domain-invariant patterns and improve performance on out-of-distribution samples. In this paper, we focus on Single Source Domain Generalization (SSDG), where a model is trained on data from a single domain and aims to generalize well to unseen domains [40, 51]. It represents the most strict form of DG, as the model must extract domain-invariant features from a single, often limited perspective, without exposure to the variation present across multiple domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Most previous work on SSDG [5, 7, 8, 19, 20, 22, 59, 63] relies extensively on data augmentations to support the learning of domain-generalized features. The premise is that constructing an extensive repertoire of augmentations synthesizes instances encompassing a wide spectrum of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.865, + 0.405, + 0.877 + ], + "angle": 0, + "content": "\\* aishwarya.agarwal@research.iit.ac.in, aishagar@adobe.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.877, + 0.225, + 0.888 + ], + "angle": 0, + "content": "† skaranam@adobe.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.205, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\)VGandhi@iiit.ac.in" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.865, + 0.405, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30210" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "human-recognizable domains. However, accounting for all conceivable real-world augmentations presents an immense challenge. These models have shown reasonable success in addressing variations in style and texture [45, 64]; however, their performance remains modest when faced with more semantic domain shifts, such as changes in background and viewpoint e.g., in the VLCS dataset [41, 48]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Consider the results of the state-of-the-art ABA [5] model on the two examples in Figure 1. In the first example (left), domain shifts manifest as variations in background and viewpoint. In the second example (right), the training data comprises photos of human faces, whereas the test data consists of skeletal sketches. In both cases, the domain shifts extend beyond style or texture variations, with ABA failing to correctly classify the samples. Class-level saliency maps [43] reveal that this misclassification stems from inadequate focus on critical local concepts under domain shifts, such as beak and feathers for birds, or eyes and lips for persons. Such local concepts are integral to class definition and remain invariant across domains, and a robust DG model must adeptly capture these stable features. We posit that prior efforts learn global features per class; and if the model fails to learn the correct feature set in the training domain, its generalization performance is compromised as noted from Figure 1 above. The inconsistency in concept localization further exacerbates the interpretability and explainability of these models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.506, + 0.483, + 0.762 + ], + "angle": 0, + "content": "We adopt an alternative approach, wherein, rather than attempting to encompass all potential augmentations, we compel the model to leverage essential class-specific concepts for classification. Our key idea is to force the model to attend to these concepts during training. The primary hurdle in this path is lack of annotated data specifying relevant class-specific concepts along with their corresponding localizations. To this end, our first contribution is a novel pipeline that harnesses large language models (LLMs) and diffusion models (DMs) to identify key concepts and generate concept-level saliency maps in a scalable, automated fashion. We demonstrate that DMs, via cross-attention layers, can generate concept-level saliency maps for generated images with notable granularity. Given these maps for a single synthesized image within a class, we leverage the rich feature space of DMs to transfer them to images across diverse domains in DG benchmarks [47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We subsequently introduce TIDE, our second contribution which employs cross-entropy losses over class and concept labels along with a novel pair of loss functions: a concept saliency alignment loss that ensures the model attends to the correct regions for each local concept, and a local concept contrastive loss that promotes learning domain invariant features for these regions. As shown in Figure 1, TIDE consistently attends to local concepts such as beak and eyes in the training domain as well as across substan" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "tial domain shifts. This precise localization significantly bolsters SSDG performance while enabling the generation of visual explanations for model decisions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Marking a major leap in model interpretability and performance gains, our third contribution is to demonstrate how our model's concept-level localization can be effectively leveraged for performance verification and test-time correction. To this end, we introduce the notion of local concept signatures: prototypical features derived by pooling concept-level features across training samples, guided by corresponding saliency maps. If the concept features associated with the predicted class label do not align with their signatures, it signals the use of incorrect features for prediction. Consequently, we employ iterative refinement through concept saliency masking until concept predictions align with their corresponding signatures. More formally, our paper makes following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.35, + 0.905, + 0.409 + ], + "angle": 0, + "content": "- We propose a novel synthetic annotation pipeline to automatically identify key per-class local concepts and their localization maps, and transfer them to real images found in DG benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.411, + 0.905, + 0.484 + ], + "angle": 0, + "content": "- We propose a novel approach, TIDE, to train locally interpretable models with a novel concept saliency alignment loss that ensures accurate concept localization and a novel local concept contrastive loss that ensures domain-invariant features." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.486, + 0.905, + 0.531 + ], + "angle": 0, + "content": "- We show TIDE enables correcting model prediction at test time using the predicted concept saliency maps as part of a novel iterative attention refinement strategy." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.532, + 0.905, + 0.576 + ], + "angle": 0, + "content": "- We report extensive results on four standard DG benchmarks, outperforming the current state-of-the-art by a significant \\(12\\%\\) on average." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.35, + 0.905, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.589, + 0.663, + 0.605 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.765 + ], + "angle": 0, + "content": "Multi Source domain generalization (MSDG) methods assume access to multiple source domains and domain-specific knowledge during training and have proposed ways to learn multi-source domain-invariant features [10, 15, 34, 35, 53], utilize domain labels to capture domain shifts [11, 35, 56], and design target-specific augmentation strategies [14, 16, 23]. However, these assumptions are not practical for real-world applications [44, 57] and we instead focus on the more challenging SSDG setting, where only one source domain and no prior target knowledge is available." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Most SSDG approaches have used augmentations to improve DG [5, 7, 8, 19, 20, 22, 59, 63]. While [59] also uses diffusion models (DMs) [42] for data augmentation, our approach differs by leveraging the rich feature space of DMs primarily for offline saliency map annotation instead of augmentation. More recent work in SSDG [4, 28, 29] has also turned to approaches based on domain-specific prompting and causal inference instead of data augmentations. However, these methods learn global features, limiting invari" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30211" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.573, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.089, + 0.915, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.34, + 0.908, + 0.37 + ], + "angle": 0, + "content": "Figure 2. The TIDE pipeline: Left—Training on a single domain with cross-entropy losses for class \\((\\mathcal{L}_{\\mathrm{c}})\\) and concept labels \\((\\mathcal{L}_{\\mathrm{k}})\\), alongside Concept Saliency Alignment \\((\\mathcal{L}_{\\mathrm{CSA}})\\) and Local Concept Contrastive losses \\((\\mathcal{L}_{\\mathrm{LCC}})\\). Right—Test-time correction strategy applied in TIDE." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.483, + 0.426 + ], + "angle": 0, + "content": "ance across unseen domains. Finally, these methods also do not provide any signals to interpret model predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.426, + 0.483, + 0.593 + ], + "angle": 0, + "content": "On the other hand, while methods like Concept Bottle-neck Models (CBMs) [24, 33] also learn local concepts, they are restricted to a predefined set of concepts and more crucially, cannot ground them to image regions that these concepts represent. Recent methods [37, 58] have scaled CBMs to a large number of classes and also proposed to ways to learn multi-scale representations [52], but they are also unable to connect these concepts to image pixels. Our SSDG method addresses these challenges by producing local concept saliency maps, enhancing both DG performance and model prediction interpretability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.226, + 0.621 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.703 + ], + "angle": 0, + "content": "The proposed framework is depicted in Figure 2. Training comprises three components: class prediction, concept prediction with saliency alignment, and localized contrastive training. During inference, class and concept predictions are integrated with a test-time correction mechanism." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In the concept alignment phase, we employ concept-level saliency maps as ground truth to enforce focus on interpretable features, using a saliency alignment loss. Domain invariance is further promoted by localized contrastive training, where we train the model to cluster similar concepts (e.g., eyes across augmentations) while separating unrelated concepts (e.g., strings). Finally, a test-time correction mechanism iteratively refines attention by suppressing irrelevant regions, leveraging concept-signatures to redirect focus and improve classification accuracy. We proceed with a concise review of the key notations, followed by an in-depth exposition of our approach and the pipeline for generating ground truth concept-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.533 + ], + "angle": 0, + "content": "We assume the source data is observed from a single domain, denoted as \\( D = \\{x_{i},y_{i}^{c},y_{i}^{k}\\}_{i = 1}^{N} \\), where \\( x_{i} \\) represents the \\( i \\)-th image, \\( y_{i}^{c} \\) the class label, \\( y_{i}^{k} \\) the concept label, and \\( N \\) the total sample count in the source domain. The shared CNN backbone is used to obtain the backbone features \\( F_{x}\\in \\mathbb{R}^{W\\times H\\times C} \\). The automatically generated ground truth concept-level saliency maps are denoted by \\( G_{x}^{k} \\). The GradCAM [43] saliency maps corresponding to class and concepts labels are denoted by \\( S_{x}^{c} \\) and \\( S_{x}^{k} \\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.844, + 0.56 + ], + "angle": 0, + "content": "3.1. Generating Concept-level Annotations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.657 + ], + "angle": 0, + "content": "During training, we aim our model to identify and spatially attend to stable, discriminative regions. However, existing DG datasets lack fine-grained, concept-level annotations for such regions. To address this, we propose a novel pipeline that uses LLMs and DMs to automate scalable concept-level saliency map generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Our primary insight is that DMs can be harnessed to generate high-quality, concept-level saliency maps for synthesized images. Extracting cross-attention maps [1, 3] from a DM, given the prompt with concepts and corresponding synthesized image, yields highly granular concept-level saliency maps. Figure 3 demonstrates how DMs emphasize specific regions for distinct concepts, capturing fine-grained attention to features such as a cat's whiskers or a snowman's hands. With this technique serving as an efficient tool for yielding concept maps for synthesized images, the ensuing questions are: (i) how to automate the identification of pertinent concepts for each class across datasets, and (ii) how to transfer these concept maps to real-world images to generate ground-truth annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "First, to identify the key concepts associated with each class, we use GPT-3.5 [2], which we prompt to generate a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "30212" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.094, + 0.437, + 0.106 + ], + "angle": 0, + "content": "Prompt: A photo of a cat with its whiskers, ears, eyes" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.11, + 0.197, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.11, + 0.291, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.11, + 0.382, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.11, + 0.473, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.184, + 0.443, + 0.197 + ], + "angle": 0, + "content": "Prompt: A photo of a snowman with its nose, eyes, hands" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.199, + 0.198, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.27, + 0.195, + 0.281 + ], + "angle": 0, + "content": "Generated Image" + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.199, + 0.292, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.27, + 0.413, + 0.281 + ], + "angle": 0, + "content": "Extracted Cross Attention Maps" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.199, + 0.382, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.199, + 0.473, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.297, + 0.483, + 0.34 + ], + "angle": 0, + "content": "Figure 3. The first column displays the image generated from the given prompt, while the subsequent three columns show the cross-attention maps corresponding to each concept in the prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.474 + ], + "angle": 0, + "content": "list of distinctive, stable features (prompt in the supplement) for each class. For instance, GPT-3.5 outputs concepts such as whiskers, ears, and eyes for a cat. We generate a prompt leveraging these concepts, which is then used to synthesize an exemplar image for each class. We further derive the corresponding concept-level attention maps as outlined in the preceding paragraph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.7 + ], + "angle": 0, + "content": "Given a single synthesized exemplar image for each class and the concept-level saliency maps, we turn to the task of transferring these saliency annotations to real-world images. DM's feature space is particularly well-suited for this, as it captures detailed, semantically-rich representations that allow us to match synthetic concept regions to real images across domains. Leveraging this, we use the Diffusion Feature Transfer (DIFT) method [47], which computes pixel-level correspondences by comparing cosine similarities between DM features. Through this approach, we establish a region-to-region correspondence between synthetic saliency maps (e.g., mouth of a dog) and similar regions in real-world images. This enables us to generate comprehensive, consistent concept-level annotations across domains as shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.484, + 0.822 + ], + "angle": 0, + "content": "We obtain these concept-level annotations for widely-used benchmarks, including VLCS [13], PACS [27], DomainNet [38], and Office-Home [50]. For an image \\( x \\), the binarized concept-level maps \\( (G_x^k) \\) are henceforth referred to as ground-truth saliency maps (GT-maps) in this paper. Having established a process for generating saliency maps, the next challenge is identifying the subset of concepts that the model actually relies on for making predictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.846, + 0.294, + 0.861 + ], + "angle": 0, + "content": "3.1.1 Concepts that matter" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We restrict our method to essential concepts, filtering out those that do not contribute meaningfully to classification." + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.093, + 0.824, + 0.105 + ], + "angle": 0, + "content": "Prompt: A photo of a dog with its ear, mouth" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.106, + 0.85, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.175, + 0.623, + 0.185 + ], + "angle": 0, + "content": "PACS - photo" + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.175, + 0.712, + 0.186 + ], + "angle": 0, + "content": "PACS - cartoon" + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.175, + 0.796, + 0.186 + ], + "angle": 0, + "content": "PACS - sketch" + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.175, + 0.889, + 0.186 + ], + "angle": 0, + "content": "PACS - painting" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.186, + 0.628, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.187, + 0.715, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.187, + 0.793, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.186, + 0.891, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.252, + 0.628, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.252, + 0.714, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.252, + 0.803, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.252, + 0.89, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.317, + 0.628, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.317, + 0.714, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.317, + 0.803, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.317, + 0.891, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.394, + 0.906, + 0.464 + ], + "angle": 0, + "content": "Figure 4. The first row presents the prompt, corresponding synthesized image and attention maps for the concepts ear and mouth. Below, we demonstrate that using diffusion features correspondences these concept saliency maps from a single exemplar can be automatically transferred on dog images across domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.552 + ], + "angle": 0, + "content": "To do this, we train a ResNet-18 [18] classifier on the source domain [26], compute GradCAM [43] saliency maps for each class label, and use them to identify regions the model focuses on when making predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.614 + ], + "angle": 0, + "content": "We compute the overlap between the GradCAM maps and GT-maps for each known concept in the dataset. Given image \\( x \\), its saliency map \\( S_{x}^{c} \\) for class \\( c \\) and the GT-map \\( G_{x}^{k} \\) for concept \\( k \\), we define the overlap \\( O_{c}^{k}(x) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.626, + 0.906, + 0.665 + ], + "angle": 0, + "content": "\\[\nO _ {c} ^ {k} (x) = \\frac {\\sum_ {i , j} \\min \\left(S _ {x} ^ {c} (i , j) , G _ {x} ^ {k} (i , j)\\right)}{\\sum_ {i , j} G _ {x} ^ {k} (i , j)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.779 + ], + "angle": 0, + "content": "where \\((i,j)\\) are matrix indices. This measures how much of the model's attention for a given class aligns with the regions corresponding to concept \\(k\\). We compute this overlap for all concepts and images in the training set, and for each class \\(c\\), we define the set of important concepts \\(\\mathcal{K}_c\\) as those that consistently exhibit high overlap with the saliency maps:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.791, + 0.906, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} _ {c} = \\left\\{k \\left| \\frac {1}{N _ {c}} \\sum_ {x \\in \\mathcal {D} _ {c}} O _ {c} ^ {k} (x) > \\tau \\right. \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(N_{c}\\) is the number of training examples in class \\(c\\), \\(\\mathcal{D}_c\\) is the set of images in class \\(c\\), and \\(\\tau\\) is a threshold that determines the importance of the concept. This procedure functions as a concept discovery module, identifying which" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30213" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.429, + 0.107 + ], + "angle": 0, + "content": "local concepts are critical for predicting each class." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.173, + 0.145 + ], + "angle": 0, + "content": "3.2. TIDE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.484, + 0.215 + ], + "angle": 0, + "content": "As shown in Figure 2, TIDE utilizes cross-entropy losses \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_k\\) for class and concept labels, respectively, integrated with novel concept saliency alignment and local concept contrastive losses. We detail these loss terms below." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.234, + 0.367, + 0.249 + ], + "angle": 0, + "content": "3.2.1 Concept saliency alignment loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.319 + ], + "angle": 0, + "content": "For each image \\(x\\), we predict important concepts \\(\\mathcal{K}_c\\) and enforce alignment between the saliency maps \\(S_x^k\\) for predicted concepts and the GT-maps \\(G_x^k\\). This is encouraged by our proposed concept saliency alignment (CSA) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.329, + 0.482, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C S A}} = \\frac {1}{| \\mathcal {K} _ {c} |} \\sum_ {k \\in \\mathcal {K} _ {c}} \\| S _ {x} ^ {k} - G _ {x} ^ {k} \\| _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.484, + 0.431 + ], + "angle": 0, + "content": "Aligning the model's attention with GT-maps enables class-specific reasoning, thereby elucidating the rationale behind predictions by linking them to relevant local features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.451, + 0.353, + 0.466 + ], + "angle": 0, + "content": "3.2.2 Local concept contrastive loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.58 + ], + "angle": 0, + "content": "While the CSA loss facilitates explicit localization of concepts, it is equally essential for these concept-level features to exhibit invariance across domain shifts. To achieve domain invariance, we propose a local concept contrastive (LCC) loss, employing a triplet strategy to cluster similar concepts (e.g., eyes) across domains while distinguishing unrelated ones (e.g., feathers and ears)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.673 + ], + "angle": 0, + "content": "Let \\( x \\) be an anchor image containing concept \\( k \\), a positive image \\( x^{+} \\) (an augmentation of \\( x \\) that retains concept \\( k \\)), and a negative image \\( x^{-} \\) (containing a different concept \\( k' \\)). For each image, we compute a concept-specific feature vector \\( f_{x}^{k} \\in \\mathbb{R}^{C} \\), emphasizing the concept's relevant regions using the GT-map \\( G_{x}^{k} \\). Each element \\( f_{x}^{k}(l) \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.683, + 0.482, + 0.716 + ], + "angle": 0, + "content": "\\[\nf _ {x} ^ {k} (l) = \\sum_ {i, j} G _ {x} ^ {k} \\cdot F _ {x} (i, j, l). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.484, + 0.83 + ], + "angle": 0, + "content": "where \\(\\cdot\\) is an element-wise multiplication operation, \\(G_{x}^{k}\\) is a matrix representing the GT-map, \\(F_{x}\\) represents convolutional feature maps from the last convolution layer, and \\((i,j,l)\\) denotes the \\(l^{th}\\) channel's \\((i,j)\\) element. These vectors focus on the concepts of interest (concept \\(k\\) for \\(x\\) and \\(x^{+}\\), \\(k^{\\prime}\\) for \\(x^{-}\\)), in contrast to the global features used in prior works. The LCC loss is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.856, + 0.482, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L C C}} = \\max \\left(0, d \\left(f _ {x} ^ {k}, f _ {x ^ {+}} ^ {k}\\right) - d \\left(f _ {x} ^ {k}, f _ {x ^ {-}} ^ {k ^ {\\prime}}\\right) + \\alpha\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.902 + ], + "angle": 0, + "content": "where \\(d(.)\\) is the euclidean distance and \\(\\alpha\\) is the margin." + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.092, + 0.8, + 0.105 + ], + "angle": 0, + "content": "Algorithm 1 Iterative Test-time Correction" + }, + { + "type": "algorithm", + "bbox": [ + 0.514, + 0.108, + 0.905, + 0.484 + ], + "angle": 0, + "content": "Input: Test image \\(x\\), initial class prediction \\(c^{(0)}\\), signatures \\(p^k\\) for each concept \\(k\\) \nParameter: Threshold \\(\\delta\\), max iterations \\(T\\) \nOutput: Corrected class prediction, \\(c_{\\mathrm{final}}\\) \n1: \\(S_x^{c(0),0} \\leftarrow \\operatorname{GradCAM}(x, c^{(0)})\\) {Initialize saliency map for iteration \\(t = 0\\)} \n2: \\(x_{\\mathrm{masked}}^{0} \\leftarrow x\\) \n3: for each \\(k \\in \\mathcal{K}_{c(0)}\\) do \n4: \\(f_x^{k,0} \\leftarrow \\sum_{i,j} S_x^k \\cdot F_x(i, j, :)\\) \n5: \\(d(f_x^{k,0}, p^k) \\leftarrow 1 - \\frac{f_x^{k,0} \\cdot p^k}{\\|f_x^{k,0}\\| \\|p^k\\|}\\) \n6: if \\(d(f_x^{k,0}, p^k) > \\delta\\) then \n7: Correction Phase: \n8: for \\(t \\leftarrow 1\\) to \\(T\\) do \n9: \\(M^{c(t-1),t} \\leftarrow \\text{Binarize}(S_x^{c(t-1),t-1})\\) \n10: \\(x_{\\mathrm{masked}}^t \\leftarrow x_{\\mathrm{masked}}^{t-1} \\cdot M^{c(t-1),t}\\) \n11: \\(c^{(t)} \\leftarrow \\text{PredictClass}(x_{\\mathrm{masked}}^t)\\) \n12: \\(S_{x_{\\mathrm{masked}}}^{c(t),t} \\leftarrow \\text{GradCAM}(x_{\\mathrm{masked}}^t, c^{(t)})\\) \n13: for each \\(k \\in \\mathcal{K}_{c(t)}\\) do \n14: \\(f_{x_{\\mathrm{masked}}}^{k,t} \\leftarrow \\sum_{i,j} S_{x_{\\mathrm{masked}}}^{k,t} \\cdot F_x(i,j,:)\\) \n15: \\(d(f_{x_{\\mathrm{masked}}}^{k,t}, p^k) \\leftarrow 1 - \\frac{f_{x_{\\mathrm{masked}}}^{k,t} \\cdot p^k}{\\|f_{x_{\\mathrm{masked}}}^{k,t}\\| \\|p^k\\|}\\) \n16: if \\(d(f_{x_{\\mathrm{masked}}}^{k,t}, p^k) \\leq \\delta\\) then \n17: Return \\(c_{\\mathrm{final}} \\leftarrow c^{(t)}\\) \n18: end if \n19: end for \n20: end for \n21: end if \n22: Return \\(c_{\\mathrm{final}} \\leftarrow c^{(0)}\\) \n23: end for" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.51, + 0.709, + 0.524 + ], + "angle": 0, + "content": "3.3. Test-time Correction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.595 + ], + "angle": 0, + "content": "We establish that our localized, interpretable approach facilitates correction of misclassifications through concept-level feature verification. In this section, we first introduce concept signatures and detail the proposed correction strategy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.614, + 0.741, + 0.629 + ], + "angle": 0, + "content": "3.3.1 Local concept signatures" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.637, + 0.906, + 0.698 + ], + "angle": 0, + "content": "For each concept \\( k \\), we define a concept-signature \\( p^k \\in \\mathbb{R}^C \\), as its representative vector. We derive \\( p^k \\) by averaging the concept-specific feature vectors \\( f_x^k \\) across all training samples \\( x \\in \\mathcal{D} \\) containing concept \\( k \\) (denoted as \\( \\mathcal{D}^k \\))." + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.709, + 0.905, + 0.746 + ], + "angle": 0, + "content": "\\[\np ^ {k} = \\frac {1}{\\left| \\mathcal {D} ^ {k} \\right|} \\sum_ {x \\in \\mathcal {D} ^ {k}} f _ {x} ^ {k}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.752, + 0.906, + 0.811 + ], + "angle": 0, + "content": "These vectors act as reference points, helping the model recognize when its attention is aligned with the right concepts during prediction, even when encountering new, unseen domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.868, + 0.847 + ], + "angle": 0, + "content": "3.3.2 Detecting and Correcting misclassifications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Consider the example of test-time correction in Figure 2, where the model misclassifies a person as a guitar. The predicted class guitar, involves concepts like strings and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30214" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "knobs, but these are absent in the image. As a result, the model erroneously focuses on irrelevant features (e.g., the person's legs or the background), as reflected in the Grad-CAM saliency maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.272 + ], + "angle": 0, + "content": "To address misclassifications, we employ a two-step approach: possible mistake detection followed by correction. The process is outlined in Algorithm 1. First, the model extracts concept-level saliency maps \\((S_x^k)\\), for all the concepts corresponding to the predicted class, which are then used to compute concept-level features \\(f_{x}^{k}\\) (step 4). In Step 5 and 6, we compare the features to the stored concept signatures; a deviation exceeding \\(\\delta\\) signals a misclassification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.455 + ], + "angle": 0, + "content": "Once a misalignment is detected, the model enters an iterative refinement phase. The features prominent for the current class level predictions are masked through the corresponding saliency map \\( S_{x}^{c} \\) (step 9 and 10). The masking process is cumulative, in effect utilizing all masks from the first to the \\( t^{th} \\) iteration. The masked features are then used for subsequent predictions (step 11). This process continues until the concepts corresponding to the predicted class aligns with concept-signatures or the maximum iteration count is reached (steps 8-16 in the algorithm). If alignment is achieved, the predicted class in that iteration is confirmed as the final output." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.223, + 0.485 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Datasets: We conduct experiments on four widely used datasets - PACS [27], VLCS [13], OfficeHome [50], and DomainNet [38], within the DomainBed [17] evaluation benchmarks. PACS contains 9,991 images across 7 categories in four domains: 'sketch', 'photo', 'clipart', and 'painting'. VLCS consists of 10,729 images over 5 categories and 4 domains. The domains are from VOC2007 (V), LabelMe (L), Caltech101 (C), and SUN09 (S), with domain shifts primarily driven by background and viewpoint variations. OfficeHome, with 15,500 images across 65 categories, emphasizes indoor classes across 'product', 'real', 'clipart', and 'art' domains. For DomainNet, we follow prior work [39, 46] and use a subset of the 40 most common classes across 'sketch', 'real', 'clipart', and 'painting'." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Experimental Setup: We adhere to the SSDG paradigm, training on one source domain and testing across three target domains, i.e., the model is trained independently four times, with the averaged test accuracies across target domains reported in each case. Training utilizes solely source-domain GT-maps, excluding any target-domain prior knowledge, data or annotations. Minimal augmentations (quantization, blurring, and canny edge), were used to introduce slight perturbations to create the triplets. We use SDv2.1 [42] for generating exemplar images and computing cross-attention maps [1, 3]. To ensure a fair" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "comparison, we adopt a ResNet-18 backbone throughout all the experiments. We use the Adam optimizer with an initial learning rate of \\( 1 \\times 10^{-4} \\) and a warm-up schedule over the first 1000 steps, after which the rate remains constant. Margin value \\( \\alpha \\) is set to 1.0. The batch size is set to 32. We empirically set \\( \\delta = 0.1 \\) and cap test-time correction at \\( T = 10 \\) iterations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.455 + ], + "angle": 0, + "content": "Compared Approaches: We compare our method with ERM [17] baseline and existing approaches that utilize augmentation-based techniques (NJPP [59], AugMix [19], MixStyle [63], CutMix [60], RandAugment [7]), self-supervised and domain adaptation methods (RSC [22], pAdaIn [36], L2D [55], RSC+ASR [12]), uncertainty modeling (DSU [29], DSU-MAD [41]), attention and meta-learning methods (ACVC [8], P-RC [6], Meta-Casual [4]), and prompt-based learning (PromptD [28]). The number of methods evaluated differs by dataset. The discrepancy stems from PACS being the dominant benchmark in prior SSDG works, resulting in a larger number of methods evaluated on it. For VLCS, OfficeHome, and DomainNet, we rely on reported results from respective papers, or compute them ourselves (where code was available) to ensure a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.652, + 0.48 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.905, + 0.595 + ], + "angle": 0, + "content": "Table 1 provide a comparison of average classification accuracy for our approach against existing methods across the PACS, VLCS, OfficeHome, and DomainNet datasets. Each column in these tables represents a source domain used for training, with the numerical values indicating the average accuracy on the three target domains. The last column presents the average of the four columns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.905, + 0.837 + ], + "angle": 0, + "content": "TIDE decisively outperforms existing approaches across all datasets. It achieves average accuracy gains of \\(8.33\\%\\), \\(13.37\\%\\), \\(16.16\\%\\), and \\(8.84\\%\\) over the second-best approach on PACS, VLCS, OfficeHome, and DomainNet, respectively. Table 1b details the performance on VLCS, where domain shifts primarily stem from background and viewpoint variations, with scenes spanning urban to rural and favoring non-standard viewpoints. Data augmentation based methods, which focus on style variation, yield limited gains on VLCS relative to PACS. Nonetheless, TIDE secures substantial improvements, notably achieving a \\(25.34\\%\\) gain over DSU-MAD when Caltech101 (C) serves as the source domain. Another noteworthy observation is that TIDE maintains strong performance across varying class counts, from VLCS's 5 to OfficeHome's 65, underscoring the scalability of the proposed approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.625, + 0.862 + ], + "angle": 0, + "content": "4.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Components of TIDE: Ablation experiments on the PACS dataset are conducted to assess the contribution of each loss" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30215" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.09, + 0.473, + 0.272 + ], + "angle": 0, + "content": "
Method VenueArtCartoonSketchPhotoAverage
ERM65.3864.2034.1533.6549.35
Augmix [19] ICLR'2166.5470.1652.4838.3057.12
RSC [22] ECCV'2073.4075.9056.2041.6061.03
Mixstyle [63] ICLR'2167.6070.3834.5737.4452.00
pAdalin [36] CVPR'2164.9665.2432.0433.6649.72
RSC+ASR [12] CVPR'2176.7079.3061.6054.6068.30
L2D [55] ICCV'2176.9177.8853.6652.2965.93
DSU [29] ICLR'2271.5474.5147.7542.1058.73
ACVC [8] CVPR'2273.6877.3955.3048.0563.10
DSU-MAD [41] CVPR'2372.4174.4749.6044.1560.66
P-RC [6] CVPR'2376.9878.5462.8957.1168.88
Meta-Casual [4] CVPR'2377.1380.1462.5559.6069.86
ABA [5] ICCV'2375.6977.3654.1259.0466.30
PromptD [28] CVPR'2478.7782.6962.9460.0971.87
TIDE86.2486.3773.1174.3680.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.262, + 0.275, + 0.313, + 0.286 + ], + "angle": 0, + "content": "(a) PACS" + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.293, + 0.465, + 0.406 + ], + "angle": 0, + "content": "
MethodVLCSAverage
ERM76.7258.8644.9557.7159.06
Augmix [19] ICLR'1975.2559.5245.9057.4359.03
pAdalin [36] CVPR'2176.0365.2143.1757.9460.34
Mixstyle [63] ICLR'2175.7361.2944.6656.5759.06
ACVC [8] CVPR'2276.1561.2347.4360.1861.75
DSU [29] ICLR'2276.9369.2046.5458.3662.11
DSU-MAD [41] CVPR'2376.9970.8544.7862.2363.71
TIDE82.6286.1370.1272.4477.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.261, + 0.408, + 0.315, + 0.42 + ], + "angle": 0, + "content": "(b) VLCS" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.426, + 0.48, + 0.59 + ], + "angle": 0, + "content": "
MethodArtClipartProductRealAverage
ERM57.4350.8348.958.6853.96
MixUp [21] ICLR'1850.4143.1941.2451.8946.93
CutMix [60] ICCV'1949.1746.1541.253.6447.04
Augmix [19] ICLR'1956.8654.1252.0260.1256.03
RandAugment [7] CVPRW'2058.0755.3252.0260.8256.56
CutOut [9] arXiv:170854.3650.7947.6858.2452.77
RSC [22] ECCV'2053.5148.9847.1658.352.73
MEADA [62] NIPS'2057.053.248.8159.2154.80
PixMix [20] CVPR'2253.7752.6848.9158.6853.51
L2D [55] ICCV'2152.7948.9747.7558.3151.71
ACVC [8] CVPR'2254.351.3247.6956.2552.89
NJPP [59] ICML'2460.7254.9552.4761.2657.85
TIDE72.3275.1375.2273.3774.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.245, + 0.592, + 0.331, + 0.604 + ], + "angle": 0, + "content": "(c) OfficeHome" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.611, + 0.485, + 0.744 + ], + "angle": 0, + "content": "
MethodClipartPaintingRealSketchAverage
ERM68.7366.1268.5169.4468.2
MixUp [21] ICLR'1870.3164.3469.2168.8268.17
CutMix [60] ICCV'1971.5263.8467.1369.4167.98
Augmix [19] ICLR'1972.3762.9169.8471.2269.09
RandAugment [7] CVPRW'2069.7165.5168.3666.9367.63
CutOut [9] arXiv:170870.8664.4869.9271.5569.20
RSC [22] ECCV'2068.2567.9170.7666.1868.28
PixMix [20] CVPR'2272.1263.5171.3465.4668.12
NJPP [59] ICML'2476.1469.2476.6171.2173.3
TIDE82.4280.3784.1581.6182.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.246, + 0.746, + 0.33, + 0.758 + ], + "angle": 0, + "content": "(d) DomainNet" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.827 + ], + "angle": 0, + "content": "Table 1. SSDG classification accuracies \\((\\%)\\) on PACS, VLCS, OfficeHome and DomainNet datasets, with ResNet-18 as backbone. Each column title indicates the source domain, and the numerical values represent the average performance in the target domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "term on the overall performance. In Table 2, we present classification accuracy by incrementally adding components to the training pipeline, demonstrating the contribu" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.09, + 0.901, + 0.166 + ], + "angle": 0, + "content": "
MethodArtCartoonSketchPhotoAverage
Lc65.3864.2034.1533.6549.35
Lc + Lk65.4764.1235.4433.8149.71
Lc + Lk + LCSA66.1164.2335.1434.2749.93
Lc + Lk + LCSA + LLC80.2882.9166.1265.3773.67
+ test-time correction86.2486.3773.1174.3680.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.595, + 0.169, + 0.824, + 0.182 + ], + "angle": 0, + "content": "Table 2. Ablation study \\((\\%)\\) on PACS." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.196, + 0.905, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.425, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 5. t-SNE visualizations to demonstrate impact of \\(\\mathcal{L}_{LCC}\\). Each column represents a test domain (Sketch, Cartoon, Painting), with the top row showing t-SNE plots without \\(\\mathcal{L}_{LCC}\\) applied and the bottom one with it. Please zoom in for optimal viewing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.72 + ], + "angle": 0, + "content": "tion of each element to the model's performance. We observe that the introduction of the concept classification loss \\(\\mathcal{L}_k\\) and concept saliency alignment loss \\(\\mathcal{L}_{CSA}\\) does not significantly affect the classifier's accuracy on the test domain. However, these components are integral to our approach, enabling test-time correction and, in turn, enhancing both performance and model interpretability. We observe that the introduction of the \\(\\mathcal{L}_{LCC}\\) loss leads to a substantial increase in test domain accuracy, clearly demonstrating its effectiveness in fostering domain invariance. Finally, we compare results before and after test-time correction, highlighting that even prior to correction, our method achieves state-of-the-art performance across all source domains on PACS, underscoring the strength of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Error analysis of test-time correction: For this analysis, we train TIDE on the photo domain of PACS and test it on the sketch domain. The model gives an initial accuracy of \\(74.79\\%\\), which improves to \\(82.29\\%\\) post test-time correction. In \\(72.2\\%\\) of test samples, TIDE does not invoke test-time correction, with class predictions correct in \\(93.8\\%\\) of these cases, demonstrating substantial reliability. The supplement provides examples of cases where the model misclassifications are not picked-up in the signature matching step of test-time correction. In the remaining \\(27.8\\%\\) of cases where correction is initiated, \\(52.5\\%\\) successfully converge to the correct classification, significantly contributing" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30216" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.158, + 0.092, + 0.326, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.091, + 0.498, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.092, + 0.67, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.092, + 0.847, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.175, + 0.326, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.175, + 0.498, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.174, + 0.671, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.174, + 0.846, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.908, + 0.292 + ], + "angle": 0, + "content": "Figure 6. Illustrative examples of concept level GradCAM maps corresponding to TIDE's predictions, across the four studied datasets. The concept names are displayed beneath the maps, with the target domain and predicted class indicated on the left. More results in supplement." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.316, + 0.455, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Figure 7. The top row shows initial class predictions and GradCAM maps for the concepts, while the bottom row presents the results after test-time correction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.407, + 0.582 + ], + "angle": 0, + "content": "to the final post-correction accuracy of \\(82.29\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.78 + ], + "angle": 0, + "content": "t-SNE plots: To demonstrate the impact of \\(\\mathcal{L}_{LCC}\\), we present t-SNE visualizations in Figure 5, showing the distribution of concept-specific vectors (as computed in Equation 4). We individually plot the source domain (photo) along with sketch, cartoon, and painting domains as targets. Each case includes two plots: one without \\(\\mathcal{L}_{LCC}\\) (top row) and one with it (bottom row). We observe that with \\(\\mathcal{L}_{LCC}\\), concept samples (e.g., mouth) align closely across domains, while distinct separation occurs between different concepts (e.g., mouth and horns). Without \\(\\mathcal{L}_{LCC}\\), cluster separation and alignment across domains are weaker, highlighting its role in improving intra-concept compactness and inter-concept separability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.274, + 0.803 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Concept Localization: We train the SSDG model on photo domain for PACS, OfficeHome, and DomainNet and on Caltech101 for VLCS. The predicted class, concepts and corresponding saliency maps are shown in Figure 6. For each target class (e.g., mushroom, kettle, guitar), the model reliably highlights key concept-specific regions (e.g.," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.907, + 0.436 + ], + "angle": 0, + "content": "stem, handle, strings) essential for classification. The model effectively isolates key features across diverse contexts, such as the legs and seat of a chair in complex zoomed-out scenes, the beak and feathers of camouflaged birds, the eyes and whiskers in deformed sketches of cats, and varying feature sizes, with the wick occupying a small area and the mushroom cap spanning a larger one." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.909, + 0.605 + ], + "angle": 0, + "content": "Test-time Correction: The Figure 7 visually illustrates the efficacy of TIDE's test time correction abilities, by comparing the initial and corrected results. The model initially mis-classifies the images as dog and elephant, resulting in poorly aligned concept-level saliency maps for corresponding concepts i.e. ears, snout, tusks and trunk. TIDE detects and rectifies such errors during inference, leading to accurate classification accompanied by precise concept localization. This qualitative evaluation reinforces the robustness of our approach in refining predictions and generating reliable, class-specific concept maps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.634, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we considered the problem of single-source domain generalization and observed that the current state-of-the-art methods fail in cases of semantic domain shifts primarily due to the global nature of their learned features. To alleviate this issue, we proposed TIDE, a new approach that not only learns local concept representations but also produces localization maps for these concepts. With these maps, we showed we could visually interpret model decisions while also enabling correction of these decisions at test time using our iterative attention refinement strategy. Extensive experimentation on standard benchmarks demonstrated substantial and consistent performance gains over the current state-of-the-art. Future work will rigorously explore methods to generate confidence scores grounded in TIDE's concept verification strategy and would explore the application of TIDE on other (e.g. fine-grained) classification datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30217" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In International Conference on Computer Vision (ICCV), 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.185, + 0.484, + 0.225 + ], + "angle": 0, + "content": "[2] Tom B Brown. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.284, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[4] Jin Chen, Zhi Gao, Xinxiao Wu, and Jiebo Luo. Meta-causal learning for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.34, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[5] Sheng Cheng, Tejas Gokhale, and Yezhou Yang. Adversarial bayesian augmentation for single-source domain generalization. In International Conference on Computer Vision (ICCV), 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.397, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[6] Seokeon Choi, Debasmit Das, Sungha Choi, Seunghan Yang, Hyunsin Park, and Sungrack Yun. Progressive random convolutions for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.466, + 0.484, + 0.535 + ], + "angle": 0, + "content": "[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.536, + 0.484, + 0.604 + ], + "angle": 0, + "content": "[8] Ilke Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.606, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[9] Terrance Devries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. ArXiv, abs/1708.04552, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[10] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[11] Antonio D'Innocente and Barbara Caputo. Domain generalization with domain-specific aggregation modules. In The German Conference on Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[12] Xinjie Fan, Qifei Wang, Junjie Ke, Feng Yang, Boqing Gong, and Mingyuan Zhou. Adversarily adaptive normalization for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[13] Chen Fang, Ye Xu, and Daniel N Rockmore. Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations (ICLR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[15] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In International Conference on Computer Vision (ICCV), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.908, + 0.276 + ], + "angle": 0, + "content": "[16] Tejas Gokhale, Rushil Anirudh, Bhavya Kailkhura, Jayaraman J Thiagarajan, Chitta Baral, and Yezhou Yang. Attribute-guided adversarial training for robustness to natural perturbations. In Association for the Advancement of Artificial Intelligence (AAAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[17] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. International Conference on Learning Representations (ICLR), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.373 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.446 + ], + "angle": 0, + "content": "[19] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. International Conference on Learning Representations (ICLR), 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[20] Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. Pixmix: Dreamlike pictures comprehensively improve safety measures. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[21] Yann N. Dauphin David Lopez-Paz Hongyi Zhang, Moustapha Cisse. mixup: Beyond empirical risk minimization. International Conference on Learning Representations (ICLR), 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[22] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.674 + ], + "angle": 0, + "content": "[23] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in Neural Information Processing Systems (NeurIPS), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[24] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning (ICML), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[25] Wouter M Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 43(3):766-785, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In International Conference on Computer Vision (ICCV), 2017. 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30218" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[28] Deng Li, Aiming Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[29] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. International Conference on Learning Representations (ICLR), 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[30] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In European Conference on Computer Vision (ECCV), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[31] Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. ArXiv, abs/2108.13624, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[32] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning (ICML), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[33] Andrei Margeloiu, Matthew Ashman, Umang Bhatt, Yanzhi Chen, Mateja Jamnik, and Adrian Weller. Do concept bottleneck models learn as intended? International Conference on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[34] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[35] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning (ICML), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[36] Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[37] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. International Conference on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[38] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In International Conference on Computer Vision (ICCV), 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[39] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation. In European Conference on Computer Vision (ECCV), 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.79 + ], + "angle": 0, + "content": "[40] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.791, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[41] Sanqing Qu, Yingwei Pan, Guang Chen, Ting Yao, Changjun Jiang, and Tao Mei. Modality-agnostic debiasing for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In International Conference on Computer Vision (ICCV), 2017. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[44] Sarath Sivaprasad, Akshay Goindani, Mario Fritz, and Vineet Gandhi. Class-wise domain generalization: A novel framework for evaluating distributional shift. In NeurIPS Workshop on Distribution Shifts: Connecting Methods and Applications, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[45] Nathan Somavarapu, Chih-Yao Ma, and Zsolt Kira. Frustratingly simple domain generalization via image stylization. ArXiv, abs/2006.11207, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[46] Shuhan Tan, Xingchao Peng, and Kate Saenko. Class-imbalanced domain adaptation: An empirical odyssey. In European Conference on Computer Vision Workshops (ECCVW), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[47] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.428 + ], + "angle": 0, + "content": "[48] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[49] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[50] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[51] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in Neural Information Processing Systems (NeurIPS), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.906, + 0.664 + ], + "angle": 0, + "content": "[52] Bor-Shiun Wang, Chien-Yi Wang, and Wei-Chen Chiu. Mcpnet: An interpretable classifier via multi-level concept prototypes. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.719 + ], + "angle": 0, + "content": "[53] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. International Conference on Learning Representations (ICLR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[54] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[55] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In International Conference on Computer Vision (ICCV), 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[56] Zehao Xiao, Jiayi Shen, Xiantong Zhen, Ling Shao, and Cees Snoek. A bit more bayesian: Domain-invariant learning with uncertainty. In International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30219" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[57] Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain expansion approach for single-source domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[58] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[59] Jianhao Yuan, Francesco Pinto, Adam Davies, and Philip Torr. Not just pretty pictures: Toward interventional data augmentation using text-to-image generators. International Conference on Machine Learning (ICML), 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision (ICCV), 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[61] Hanlin Zhang, Yi-Fan Zhang, Weiyang Liu, Adrian Weller, Bernhard Schölkopf, and Eric P Xing. Towards principled disentanglement for domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[62] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[63] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. International Conference on Learning Representations (ICLR), 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[64] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation. International Journal of Computer Vision (IJCV), 132(3):822-836, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30220" + } + ] +] \ No newline at end of file diff --git a/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_origin.pdf b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..78fadbc4294519d8f1a5891724e0b78fd1a98820 --- /dev/null +++ b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/e47fec81-867b-4ede-a8aa-c8b31b2fa2b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6dfc769f732f74cfbcbf22a2f66a46d965333c7ef2bc8c715dc90908ee00bbb4 +size 5827328 diff --git a/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/full.md b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5591bd937f96dabd4cadd4e498107de69892fb83 --- /dev/null +++ b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/full.md @@ -0,0 +1,393 @@ +# TIDE: Training Locally Interpretable Domain Generalization Models Enables Test-time Correction + +Aishwarya Agarwal1,2* Srikrishna Karanam2† Vineet Gandhi1‡ + +$^{1}$ CVIT, Kohli Centre for Intelligent Systems, IIIT Hyderabad, India + +$^{2}$ Adobe Research, Bengaluru, India + +# Abstract + +We consider the problem of single-source domain generalization. Existing methods typically rely on extensive augmentations to synthetically cover diverse domains during training. However, they struggle with semantic shifts (e.g., background and viewpoint changes), as they often learn global features instead of local concepts that tend to be domain invariant. To address this gap, we propose an approach that compels models to leverage such local concepts during prediction. Given no suitable dataset with per-class concepts and localization maps exists, we first develop a novel pipeline to generate annotations by exploiting the rich features of diffusion and large-language models. Our next innovation is TIDE, a novel training scheme with a concept saliency alignment loss that ensures model focus on the right per-concept regions and a local concept contrastive loss that promotes learning domain-invariant concept representations. This not only gives a robust model but also can be visually interpreted using the predicted concept saliency maps. Given these maps at test time, our final contribution is a new correction algorithm that uses the corresponding local concept representations to iteratively refine the prediction until it aligns with prototypical concept representations that we store at the end of model training. We evaluate our approach extensively on four standard DG benchmark datasets and substantially outperform the current state-of-the-art (12% improvement on average) while also demonstrating that our predictions can be visually interpreted. + +# 1. Introduction + +Enhancing deep neural networks to generalize to out-of-distribution samples remains a core challenge in machine learning and computer vision research, as real-world test data often diverges significantly from the training distribu + +![](images/69e775bd3f4808f78e554448d4f737adea66d3e269f1010adacc017ec20251c5.jpg) +Figure 1. Samples from VLCS (left) and PACS dataset (right) across domain shifts, corresponding to bird and person class. First row displays GradCAM maps [43] for ABA's class predictions. We observe that model attention of ABA [5] falters across domain shifts. Second and third row display the concept specific GradCAM maps from TIDE. We posit that accurate concept learning and localization facilitates DG. + +tion [31]. The challenges compound when obtaining labeled samples from the target domain is expensive or unfeasible, hindering application of semi-supervised learning or domain adaptation [25, 32, 49]. The problem of Domain Generalization (DG) [30, 44, 54, 61, 63] represents a promising avenue for developing techniques that capture domain-invariant patterns and improve performance on out-of-distribution samples. In this paper, we focus on Single Source Domain Generalization (SSDG), where a model is trained on data from a single domain and aims to generalize well to unseen domains [40, 51]. It represents the most strict form of DG, as the model must extract domain-invariant features from a single, often limited perspective, without exposure to the variation present across multiple domains. + +Most previous work on SSDG [5, 7, 8, 19, 20, 22, 59, 63] relies extensively on data augmentations to support the learning of domain-generalized features. The premise is that constructing an extensive repertoire of augmentations synthesizes instances encompassing a wide spectrum of + +human-recognizable domains. However, accounting for all conceivable real-world augmentations presents an immense challenge. These models have shown reasonable success in addressing variations in style and texture [45, 64]; however, their performance remains modest when faced with more semantic domain shifts, such as changes in background and viewpoint e.g., in the VLCS dataset [41, 48]. + +Consider the results of the state-of-the-art ABA [5] model on the two examples in Figure 1. In the first example (left), domain shifts manifest as variations in background and viewpoint. In the second example (right), the training data comprises photos of human faces, whereas the test data consists of skeletal sketches. In both cases, the domain shifts extend beyond style or texture variations, with ABA failing to correctly classify the samples. Class-level saliency maps [43] reveal that this misclassification stems from inadequate focus on critical local concepts under domain shifts, such as beak and feathers for birds, or eyes and lips for persons. Such local concepts are integral to class definition and remain invariant across domains, and a robust DG model must adeptly capture these stable features. We posit that prior efforts learn global features per class; and if the model fails to learn the correct feature set in the training domain, its generalization performance is compromised as noted from Figure 1 above. The inconsistency in concept localization further exacerbates the interpretability and explainability of these models. + +We adopt an alternative approach, wherein, rather than attempting to encompass all potential augmentations, we compel the model to leverage essential class-specific concepts for classification. Our key idea is to force the model to attend to these concepts during training. The primary hurdle in this path is lack of annotated data specifying relevant class-specific concepts along with their corresponding localizations. To this end, our first contribution is a novel pipeline that harnesses large language models (LLMs) and diffusion models (DMs) to identify key concepts and generate concept-level saliency maps in a scalable, automated fashion. We demonstrate that DMs, via cross-attention layers, can generate concept-level saliency maps for generated images with notable granularity. Given these maps for a single synthesized image within a class, we leverage the rich feature space of DMs to transfer them to images across diverse domains in DG benchmarks [47]. + +We subsequently introduce TIDE, our second contribution which employs cross-entropy losses over class and concept labels along with a novel pair of loss functions: a concept saliency alignment loss that ensures the model attends to the correct regions for each local concept, and a local concept contrastive loss that promotes learning domain invariant features for these regions. As shown in Figure 1, TIDE consistently attends to local concepts such as beak and eyes in the training domain as well as across substan + +tial domain shifts. This precise localization significantly bolsters SSDG performance while enabling the generation of visual explanations for model decisions. + +Marking a major leap in model interpretability and performance gains, our third contribution is to demonstrate how our model's concept-level localization can be effectively leveraged for performance verification and test-time correction. To this end, we introduce the notion of local concept signatures: prototypical features derived by pooling concept-level features across training samples, guided by corresponding saliency maps. If the concept features associated with the predicted class label do not align with their signatures, it signals the use of incorrect features for prediction. Consequently, we employ iterative refinement through concept saliency masking until concept predictions align with their corresponding signatures. More formally, our paper makes following contributions: + +- We propose a novel synthetic annotation pipeline to automatically identify key per-class local concepts and their localization maps, and transfer them to real images found in DG benchmarks. +- We propose a novel approach, TIDE, to train locally interpretable models with a novel concept saliency alignment loss that ensures accurate concept localization and a novel local concept contrastive loss that ensures domain-invariant features. +- We show TIDE enables correcting model prediction at test time using the predicted concept saliency maps as part of a novel iterative attention refinement strategy. +- We report extensive results on four standard DG benchmarks, outperforming the current state-of-the-art by a significant $12\%$ on average. + +# 2. Related Works + +Multi Source domain generalization (MSDG) methods assume access to multiple source domains and domain-specific knowledge during training and have proposed ways to learn multi-source domain-invariant features [10, 15, 34, 35, 53], utilize domain labels to capture domain shifts [11, 35, 56], and design target-specific augmentation strategies [14, 16, 23]. However, these assumptions are not practical for real-world applications [44, 57] and we instead focus on the more challenging SSDG setting, where only one source domain and no prior target knowledge is available. + +Most SSDG approaches have used augmentations to improve DG [5, 7, 8, 19, 20, 22, 59, 63]. While [59] also uses diffusion models (DMs) [42] for data augmentation, our approach differs by leveraging the rich feature space of DMs primarily for offline saliency map annotation instead of augmentation. More recent work in SSDG [4, 28, 29] has also turned to approaches based on domain-specific prompting and causal inference instead of data augmentations. However, these methods learn global features, limiting invari + +![](images/ba52611ef9468bac83a7dd8e044023b54f8e7859b587eb0effe60c809fe0d82a.jpg) +Figure 2. The TIDE pipeline: Left—Training on a single domain with cross-entropy losses for class $(\mathcal{L}_{\mathrm{c}})$ and concept labels $(\mathcal{L}_{\mathrm{k}})$ , alongside Concept Saliency Alignment $(\mathcal{L}_{\mathrm{CSA}})$ and Local Concept Contrastive losses $(\mathcal{L}_{\mathrm{LCC}})$ . Right—Test-time correction strategy applied in TIDE. + +![](images/02e40050d826f8975000ebd2ba02dd9c246c68c531b47cd493f08fb1304c7707.jpg) + +ance across unseen domains. Finally, these methods also do not provide any signals to interpret model predictions. + +On the other hand, while methods like Concept Bottle-neck Models (CBMs) [24, 33] also learn local concepts, they are restricted to a predefined set of concepts and more crucially, cannot ground them to image regions that these concepts represent. Recent methods [37, 58] have scaled CBMs to a large number of classes and also proposed to ways to learn multi-scale representations [52], but they are also unable to connect these concepts to image pixels. Our SSDG method addresses these challenges by producing local concept saliency maps, enhancing both DG performance and model prediction interpretability. + +# 3. Methodology + +The proposed framework is depicted in Figure 2. Training comprises three components: class prediction, concept prediction with saliency alignment, and localized contrastive training. During inference, class and concept predictions are integrated with a test-time correction mechanism. + +In the concept alignment phase, we employ concept-level saliency maps as ground truth to enforce focus on interpretable features, using a saliency alignment loss. Domain invariance is further promoted by localized contrastive training, where we train the model to cluster similar concepts (e.g., eyes across augmentations) while separating unrelated concepts (e.g., strings). Finally, a test-time correction mechanism iteratively refines attention by suppressing irrelevant regions, leveraging concept-signatures to redirect focus and improve classification accuracy. We proceed with a concise review of the key notations, followed by an in-depth exposition of our approach and the pipeline for generating ground truth concept-level annotations. + +We assume the source data is observed from a single domain, denoted as $D = \{x_{i},y_{i}^{c},y_{i}^{k}\}_{i = 1}^{N}$ , where $x_{i}$ represents the $i$ -th image, $y_{i}^{c}$ the class label, $y_{i}^{k}$ the concept label, and $N$ the total sample count in the source domain. The shared CNN backbone is used to obtain the backbone features $F_{x}\in \mathbb{R}^{W\times H\times C}$ . The automatically generated ground truth concept-level saliency maps are denoted by $G_{x}^{k}$ . The GradCAM [43] saliency maps corresponding to class and concepts labels are denoted by $S_{x}^{c}$ and $S_{x}^{k}$ respectively. + +# 3.1. Generating Concept-level Annotations + +During training, we aim our model to identify and spatially attend to stable, discriminative regions. However, existing DG datasets lack fine-grained, concept-level annotations for such regions. To address this, we propose a novel pipeline that uses LLMs and DMs to automate scalable concept-level saliency map generation. + +Our primary insight is that DMs can be harnessed to generate high-quality, concept-level saliency maps for synthesized images. Extracting cross-attention maps [1, 3] from a DM, given the prompt with concepts and corresponding synthesized image, yields highly granular concept-level saliency maps. Figure 3 demonstrates how DMs emphasize specific regions for distinct concepts, capturing fine-grained attention to features such as a cat's whiskers or a snowman's hands. With this technique serving as an efficient tool for yielding concept maps for synthesized images, the ensuing questions are: (i) how to automate the identification of pertinent concepts for each class across datasets, and (ii) how to transfer these concept maps to real-world images to generate ground-truth annotations. + +First, to identify the key concepts associated with each class, we use GPT-3.5 [2], which we prompt to generate a + +![](images/6dee8d9d5cd5668554aad2e8dcca46d94b0fb597610b3ea26b16a7c15d8f1f03.jpg) +Prompt: A photo of a cat with its whiskers, ears, eyes + +![](images/775a048924f7abfb9f892c1dd2e4fb4897c56941137a006f6ef23b8a143c0ec1.jpg) + +![](images/0714bdac2ba4c5c47b5cdd01d830ea1ee5c52d117e8ee1c2a836c6efdd60de2e.jpg) + +![](images/dfeeb94d1db27c4cd2d431223fb422eadf00e0b2602bbe1e19901e13eee3c9fb.jpg) + +![](images/96d19e966bb870c786630b06fb675b87a2a88b9330d83bf68cc860e1c16b5304.jpg) +Prompt: A photo of a snowman with its nose, eyes, hands +Generated Image +Figure 3. The first column displays the image generated from the given prompt, while the subsequent three columns show the cross-attention maps corresponding to each concept in the prompt. + +![](images/da34b905aaa9a8c9d19e53c33627ff1dde704fe8d15c0194268a84b9ea363bac.jpg) +Extracted Cross Attention Maps + +![](images/4013b509dcba52308e124b6a435d13726b3de40a64bbe458bdc5a901dea49aef.jpg) + +![](images/b580fdf126801013a7ade642429355bba544e9980ec0d010e213e7714fadbee3.jpg) + +list of distinctive, stable features (prompt in the supplement) for each class. For instance, GPT-3.5 outputs concepts such as whiskers, ears, and eyes for a cat. We generate a prompt leveraging these concepts, which is then used to synthesize an exemplar image for each class. We further derive the corresponding concept-level attention maps as outlined in the preceding paragraph. + +Given a single synthesized exemplar image for each class and the concept-level saliency maps, we turn to the task of transferring these saliency annotations to real-world images. DM's feature space is particularly well-suited for this, as it captures detailed, semantically-rich representations that allow us to match synthetic concept regions to real images across domains. Leveraging this, we use the Diffusion Feature Transfer (DIFT) method [47], which computes pixel-level correspondences by comparing cosine similarities between DM features. Through this approach, we establish a region-to-region correspondence between synthetic saliency maps (e.g., mouth of a dog) and similar regions in real-world images. This enables us to generate comprehensive, consistent concept-level annotations across domains as shown in Figure 4. + +We obtain these concept-level annotations for widely-used benchmarks, including VLCS [13], PACS [27], DomainNet [38], and Office-Home [50]. For an image $x$ , the binarized concept-level maps $(G_x^k)$ are henceforth referred to as ground-truth saliency maps (GT-maps) in this paper. Having established a process for generating saliency maps, the next challenge is identifying the subset of concepts that the model actually relies on for making predictions. + +# 3.1.1 Concepts that matter + +We restrict our method to essential concepts, filtering out those that do not contribute meaningfully to classification. + +![](images/d29618652d5205159ff10bac791cd7efd31973a16d30ee4a603223deba797526.jpg) +Prompt: A photo of a dog with its ear, mouth +PACS - photo +PACS - cartoon +PACS - sketch + +![](images/c4fd72ee987295925f35d21eb0276da075ac6daff6707f332d502e2949c671d9.jpg) + +![](images/ac16b1af529de362cdc7f19c2931db27632963bf5dd3b7c00a189aef91fbe3c5.jpg) + +![](images/bbfdcee532654e85acc9ba737890fcdaff6c6ee34032e52ad2f177d07cbe6c6f.jpg) + +![](images/f0174a03f238e6b72fd49c175772434e24b00fed9a966a9b37d62cc100e8fe34.jpg) +PACS - painting + +![](images/ca02fb103336e15dc0db5d2712f1727c598bc38530be48221d2c73021b5ccc9a.jpg) + +![](images/2ccd2e6a534ce7fb0dfd0e77432ba5f1959acd83830a4b60d7f77b9b33f89ca7.jpg) + +![](images/6c8e2f311ae85750fa4a8e75986439e9f999e7c7caf82d8750245b52e6008117.jpg) + +![](images/e0c294a9a6a384faf3dd8902748b24a326c7d80b70053bda84b09b0f24efa158.jpg) + +![](images/ec4b65144e3a1a5f2260927e76428cfde59845b800810e3e89f71740f168cd5f.jpg) +Figure 4. The first row presents the prompt, corresponding synthesized image and attention maps for the concepts ear and mouth. Below, we demonstrate that using diffusion features correspondences these concept saliency maps from a single exemplar can be automatically transferred on dog images across domains. + +![](images/57f698054fc670394c3daaf84b776d1c645e33b332999a0825c0d83fcc7ca8b7.jpg) + +![](images/5e784e8684e0a9adea66fdc7a28ee34d59cfe64b8ce704ea733eee31c78f993b.jpg) + +![](images/625ec29125ed796b939c09db44c76f063b1c54e446a2f046a295a978a76017e3.jpg) + +To do this, we train a ResNet-18 [18] classifier on the source domain [26], compute GradCAM [43] saliency maps for each class label, and use them to identify regions the model focuses on when making predictions. + +We compute the overlap between the GradCAM maps and GT-maps for each known concept in the dataset. Given image $x$ , its saliency map $S_{x}^{c}$ for class $c$ and the GT-map $G_{x}^{k}$ for concept $k$ , we define the overlap $O_{c}^{k}(x)$ as: + +$$ +O _ {c} ^ {k} (x) = \frac {\sum_ {i , j} \min \left(S _ {x} ^ {c} (i , j) , G _ {x} ^ {k} (i , j)\right)}{\sum_ {i , j} G _ {x} ^ {k} (i , j)}. \tag {1} +$$ + +where $(i,j)$ are matrix indices. This measures how much of the model's attention for a given class aligns with the regions corresponding to concept $k$ . We compute this overlap for all concepts and images in the training set, and for each class $c$ , we define the set of important concepts $\mathcal{K}_c$ as those that consistently exhibit high overlap with the saliency maps: + +$$ +\mathcal {K} _ {c} = \left\{k \left| \frac {1}{N _ {c}} \sum_ {x \in \mathcal {D} _ {c}} O _ {c} ^ {k} (x) > \tau \right. \right\}, \tag {2} +$$ + +where $N_{c}$ is the number of training examples in class $c$ , $\mathcal{D}_c$ is the set of images in class $c$ , and $\tau$ is a threshold that determines the importance of the concept. This procedure functions as a concept discovery module, identifying which + +local concepts are critical for predicting each class. + +# 3.2. TIDE + +As shown in Figure 2, TIDE utilizes cross-entropy losses $\mathcal{L}_c$ and $\mathcal{L}_k$ for class and concept labels, respectively, integrated with novel concept saliency alignment and local concept contrastive losses. We detail these loss terms below. + +# 3.2.1 Concept saliency alignment loss + +For each image $x$ , we predict important concepts $\mathcal{K}_c$ and enforce alignment between the saliency maps $S_x^k$ for predicted concepts and the GT-maps $G_x^k$ . This is encouraged by our proposed concept saliency alignment (CSA) loss: + +$$ +\mathcal {L} _ {\mathrm {C S A}} = \frac {1}{| \mathcal {K} _ {c} |} \sum_ {k \in \mathcal {K} _ {c}} \| S _ {x} ^ {k} - G _ {x} ^ {k} \| _ {2} ^ {2}. \tag {3} +$$ + +Aligning the model's attention with GT-maps enables class-specific reasoning, thereby elucidating the rationale behind predictions by linking them to relevant local features. + +# 3.2.2 Local concept contrastive loss + +While the CSA loss facilitates explicit localization of concepts, it is equally essential for these concept-level features to exhibit invariance across domain shifts. To achieve domain invariance, we propose a local concept contrastive (LCC) loss, employing a triplet strategy to cluster similar concepts (e.g., eyes) across domains while distinguishing unrelated ones (e.g., feathers and ears). + +Let $x$ be an anchor image containing concept $k$ , a positive image $x^{+}$ (an augmentation of $x$ that retains concept $k$ ), and a negative image $x^{-}$ (containing a different concept $k'$ ). For each image, we compute a concept-specific feature vector $f_{x}^{k} \in \mathbb{R}^{C}$ , emphasizing the concept's relevant regions using the GT-map $G_{x}^{k}$ . Each element $f_{x}^{k}(l)$ is computed as: + +$$ +f _ {x} ^ {k} (l) = \sum_ {i, j} G _ {x} ^ {k} \cdot F _ {x} (i, j, l). \tag {4} +$$ + +where $\cdot$ is an element-wise multiplication operation, $G_{x}^{k}$ is a matrix representing the GT-map, $F_{x}$ represents convolutional feature maps from the last convolution layer, and $(i,j,l)$ denotes the $l^{th}$ channel's $(i,j)$ element. These vectors focus on the concepts of interest (concept $k$ for $x$ and $x^{+}$ , $k^{\prime}$ for $x^{-}$ ), in contrast to the global features used in prior works. The LCC loss is then defined as: + +$$ +\mathcal {L} _ {\mathrm {L C C}} = \max \left(0, d \left(f _ {x} ^ {k}, f _ {x ^ {+}} ^ {k}\right) - d \left(f _ {x} ^ {k}, f _ {x ^ {-}} ^ {k ^ {\prime}}\right) + \alpha\right), \tag {5} +$$ + +where $d(.)$ is the euclidean distance and $\alpha$ is the margin. + +Algorithm 1 Iterative Test-time Correction +Input: Test image $x$ , initial class prediction $c^{(0)}$ , signatures $p^k$ for each concept $k$ +Parameter: Threshold $\delta$ , max iterations $T$ +Output: Corrected class prediction, $c_{\mathrm{final}}$ +1: $S_x^{c(0),0} \leftarrow \operatorname{GradCAM}(x, c^{(0)})$ {Initialize saliency map for iteration $t = 0$ } +2: $x_{\mathrm{masked}}^{0} \leftarrow x$ +3: for each $k \in \mathcal{K}_{c(0)}$ do +4: $f_x^{k,0} \leftarrow \sum_{i,j} S_x^k \cdot F_x(i, j, :)$ +5: $d(f_x^{k,0}, p^k) \leftarrow 1 - \frac{f_x^{k,0} \cdot p^k}{\|f_x^{k,0}\| \|p^k\|}$ +6: if $d(f_x^{k,0}, p^k) > \delta$ then +7: Correction Phase: +8: for $t \leftarrow 1$ to $T$ do +9: $M^{c(t-1),t} \leftarrow \text{Binarize}(S_x^{c(t-1),t-1})$ +10: $x_{\mathrm{masked}}^t \leftarrow x_{\mathrm{masked}}^{t-1} \cdot M^{c(t-1),t}$ +11: $c^{(t)} \leftarrow \text{PredictClass}(x_{\mathrm{masked}}^t)$ +12: $S_{x_{\mathrm{masked}}}^{c(t),t} \leftarrow \text{GradCAM}(x_{\mathrm{masked}}^t, c^{(t)})$ +13: for each $k \in \mathcal{K}_{c(t)}$ do +14: $f_{x_{\mathrm{masked}}}^{k,t} \leftarrow \sum_{i,j} S_{x_{\mathrm{masked}}}^{k,t} \cdot F_x(i,j,:)$ +15: $d(f_{x_{\mathrm{masked}}}^{k,t}, p^k) \leftarrow 1 - \frac{f_{x_{\mathrm{masked}}}^{k,t} \cdot p^k}{\|f_{x_{\mathrm{masked}}}^{k,t}\| \|p^k\|}$ +16: if $d(f_{x_{\mathrm{masked}}}^{k,t}, p^k) \leq \delta$ then +17: Return $c_{\mathrm{final}} \leftarrow c^{(t)}$ +18: end if +19: end for +20: end for +21: end if +22: Return $c_{\mathrm{final}} \leftarrow c^{(0)}$ +23: end for + +# 3.3. Test-time Correction + +We establish that our localized, interpretable approach facilitates correction of misclassifications through concept-level feature verification. In this section, we first introduce concept signatures and detail the proposed correction strategy. + +# 3.3.1 Local concept signatures + +For each concept $k$ , we define a concept-signature $p^k \in \mathbb{R}^C$ , as its representative vector. We derive $p^k$ by averaging the concept-specific feature vectors $f_x^k$ across all training samples $x \in \mathcal{D}$ containing concept $k$ (denoted as $\mathcal{D}^k$ ). + +$$ +p ^ {k} = \frac {1}{\left| \mathcal {D} ^ {k} \right|} \sum_ {x \in \mathcal {D} ^ {k}} f _ {x} ^ {k}. \tag {6} +$$ + +These vectors act as reference points, helping the model recognize when its attention is aligned with the right concepts during prediction, even when encountering new, unseen domains. + +# 3.3.2 Detecting and Correcting misclassifications + +Consider the example of test-time correction in Figure 2, where the model misclassifies a person as a guitar. The predicted class guitar, involves concepts like strings and + +knobs, but these are absent in the image. As a result, the model erroneously focuses on irrelevant features (e.g., the person's legs or the background), as reflected in the Grad-CAM saliency maps. + +To address misclassifications, we employ a two-step approach: possible mistake detection followed by correction. The process is outlined in Algorithm 1. First, the model extracts concept-level saliency maps $(S_x^k)$ , for all the concepts corresponding to the predicted class, which are then used to compute concept-level features $f_{x}^{k}$ (step 4). In Step 5 and 6, we compare the features to the stored concept signatures; a deviation exceeding $\delta$ signals a misclassification. + +Once a misalignment is detected, the model enters an iterative refinement phase. The features prominent for the current class level predictions are masked through the corresponding saliency map $S_{x}^{c}$ (step 9 and 10). The masking process is cumulative, in effect utilizing all masks from the first to the $t^{th}$ iteration. The masked features are then used for subsequent predictions (step 11). This process continues until the concepts corresponding to the predicted class aligns with concept-signatures or the maximum iteration count is reached (steps 8-16 in the algorithm). If alignment is achieved, the predicted class in that iteration is confirmed as the final output. + +# 4. Experiments + +Datasets: We conduct experiments on four widely used datasets - PACS [27], VLCS [13], OfficeHome [50], and DomainNet [38], within the DomainBed [17] evaluation benchmarks. PACS contains 9,991 images across 7 categories in four domains: 'sketch', 'photo', 'clipart', and 'painting'. VLCS consists of 10,729 images over 5 categories and 4 domains. The domains are from VOC2007 (V), LabelMe (L), Caltech101 (C), and SUN09 (S), with domain shifts primarily driven by background and viewpoint variations. OfficeHome, with 15,500 images across 65 categories, emphasizes indoor classes across 'product', 'real', 'clipart', and 'art' domains. For DomainNet, we follow prior work [39, 46] and use a subset of the 40 most common classes across 'sketch', 'real', 'clipart', and 'painting'. + +Experimental Setup: We adhere to the SSDG paradigm, training on one source domain and testing across three target domains, i.e., the model is trained independently four times, with the averaged test accuracies across target domains reported in each case. Training utilizes solely source-domain GT-maps, excluding any target-domain prior knowledge, data or annotations. Minimal augmentations (quantization, blurring, and canny edge), were used to introduce slight perturbations to create the triplets. We use SDv2.1 [42] for generating exemplar images and computing cross-attention maps [1, 3]. To ensure a fair + +comparison, we adopt a ResNet-18 backbone throughout all the experiments. We use the Adam optimizer with an initial learning rate of $1 \times 10^{-4}$ and a warm-up schedule over the first 1000 steps, after which the rate remains constant. Margin value $\alpha$ is set to 1.0. The batch size is set to 32. We empirically set $\delta = 0.1$ and cap test-time correction at $T = 10$ iterations. + +Compared Approaches: We compare our method with ERM [17] baseline and existing approaches that utilize augmentation-based techniques (NJPP [59], AugMix [19], MixStyle [63], CutMix [60], RandAugment [7]), self-supervised and domain adaptation methods (RSC [22], pAdaIn [36], L2D [55], RSC+ASR [12]), uncertainty modeling (DSU [29], DSU-MAD [41]), attention and meta-learning methods (ACVC [8], P-RC [6], Meta-Casual [4]), and prompt-based learning (PromptD [28]). The number of methods evaluated differs by dataset. The discrepancy stems from PACS being the dominant benchmark in prior SSDG works, resulting in a larger number of methods evaluated on it. For VLCS, OfficeHome, and DomainNet, we rely on reported results from respective papers, or compute them ourselves (where code was available) to ensure a fair comparison. + +# 4.1. Main Results + +Table 1 provide a comparison of average classification accuracy for our approach against existing methods across the PACS, VLCS, OfficeHome, and DomainNet datasets. Each column in these tables represents a source domain used for training, with the numerical values indicating the average accuracy on the three target domains. The last column presents the average of the four columns. + +TIDE decisively outperforms existing approaches across all datasets. It achieves average accuracy gains of $8.33\%$ , $13.37\%$ , $16.16\%$ , and $8.84\%$ over the second-best approach on PACS, VLCS, OfficeHome, and DomainNet, respectively. Table 1b details the performance on VLCS, where domain shifts primarily stem from background and viewpoint variations, with scenes spanning urban to rural and favoring non-standard viewpoints. Data augmentation based methods, which focus on style variation, yield limited gains on VLCS relative to PACS. Nonetheless, TIDE secures substantial improvements, notably achieving a $25.34\%$ gain over DSU-MAD when Caltech101 (C) serves as the source domain. Another noteworthy observation is that TIDE maintains strong performance across varying class counts, from VLCS's 5 to OfficeHome's 65, underscoring the scalability of the proposed approach. + +# 4.2. Ablations + +Components of TIDE: Ablation experiments on the PACS dataset are conducted to assess the contribution of each loss + +
Method VenueArtCartoonSketchPhotoAverage
ERM65.3864.2034.1533.6549.35
Augmix [19] ICLR'2166.5470.1652.4838.3057.12
RSC [22] ECCV'2073.4075.9056.2041.6061.03
Mixstyle [63] ICLR'2167.6070.3834.5737.4452.00
pAdalin [36] CVPR'2164.9665.2432.0433.6649.72
RSC+ASR [12] CVPR'2176.7079.3061.6054.6068.30
L2D [55] ICCV'2176.9177.8853.6652.2965.93
DSU [29] ICLR'2271.5474.5147.7542.1058.73
ACVC [8] CVPR'2273.6877.3955.3048.0563.10
DSU-MAD [41] CVPR'2372.4174.4749.6044.1560.66
P-RC [6] CVPR'2376.9878.5462.8957.1168.88
Meta-Casual [4] CVPR'2377.1380.1462.5559.6069.86
ABA [5] ICCV'2375.6977.3654.1259.0466.30
PromptD [28] CVPR'2478.7782.6962.9460.0971.87
TIDE86.2486.3773.1174.3680.02
+ +(a) PACS + +
MethodVLCSAverage
ERM76.7258.8644.9557.7159.06
Augmix [19] ICLR'1975.2559.5245.9057.4359.03
pAdalin [36] CVPR'2176.0365.2143.1757.9460.34
Mixstyle [63] ICLR'2175.7361.2944.6656.5759.06
ACVC [8] CVPR'2276.1561.2347.4360.1861.75
DSU [29] ICLR'2276.9369.2046.5458.3662.11
DSU-MAD [41] CVPR'2376.9970.8544.7862.2363.71
TIDE82.6286.1370.1272.4477.08
+ +(b) VLCS + +
MethodArtClipartProductRealAverage
ERM57.4350.8348.958.6853.96
MixUp [21] ICLR'1850.4143.1941.2451.8946.93
CutMix [60] ICCV'1949.1746.1541.253.6447.04
Augmix [19] ICLR'1956.8654.1252.0260.1256.03
RandAugment [7] CVPRW'2058.0755.3252.0260.8256.56
CutOut [9] arXiv:170854.3650.7947.6858.2452.77
RSC [22] ECCV'2053.5148.9847.1658.352.73
MEADA [62] NIPS'2057.053.248.8159.2154.80
PixMix [20] CVPR'2253.7752.6848.9158.6853.51
L2D [55] ICCV'2152.7948.9747.7558.3151.71
ACVC [8] CVPR'2254.351.3247.6956.2552.89
NJPP [59] ICML'2460.7254.9552.4761.2657.85
TIDE72.3275.1375.2273.3774.01
+ +(c) OfficeHome + +
MethodClipartPaintingRealSketchAverage
ERM68.7366.1268.5169.4468.2
MixUp [21] ICLR'1870.3164.3469.2168.8268.17
CutMix [60] ICCV'1971.5263.8467.1369.4167.98
Augmix [19] ICLR'1972.3762.9169.8471.2269.09
RandAugment [7] CVPRW'2069.7165.5168.3666.9367.63
CutOut [9] arXiv:170870.8664.4869.9271.5569.20
RSC [22] ECCV'2068.2567.9170.7666.1868.28
PixMix [20] CVPR'2272.1263.5171.3465.4668.12
NJPP [59] ICML'2476.1469.2476.6171.2173.3
TIDE82.4280.3784.1581.6182.14
+ +term on the overall performance. In Table 2, we present classification accuracy by incrementally adding components to the training pipeline, demonstrating the contribu + +(d) DomainNet +Table 1. SSDG classification accuracies $(\%)$ on PACS, VLCS, OfficeHome and DomainNet datasets, with ResNet-18 as backbone. Each column title indicates the source domain, and the numerical values represent the average performance in the target domains. + +
MethodArtCartoonSketchPhotoAverage
Lc65.3864.2034.1533.6549.35
Lc + Lk65.4764.1235.4433.8149.71
Lc + Lk + LCSA66.1164.2335.1434.2749.93
Lc + Lk + LCSA + LLC80.2882.9166.1265.3773.67
+ test-time correction86.2486.3773.1174.3680.02
+ +Table 2. Ablation study $(\%)$ on PACS. + +![](images/699c6c35be7a518f1f679236ea394c8cf9de202589adb6523fefe6cad8ba8829.jpg) +Figure 5. t-SNE visualizations to demonstrate impact of $\mathcal{L}_{LCC}$ . Each column represents a test domain (Sketch, Cartoon, Painting), with the top row showing t-SNE plots without $\mathcal{L}_{LCC}$ applied and the bottom one with it. Please zoom in for optimal viewing. + +tion of each element to the model's performance. We observe that the introduction of the concept classification loss $\mathcal{L}_k$ and concept saliency alignment loss $\mathcal{L}_{CSA}$ does not significantly affect the classifier's accuracy on the test domain. However, these components are integral to our approach, enabling test-time correction and, in turn, enhancing both performance and model interpretability. We observe that the introduction of the $\mathcal{L}_{LCC}$ loss leads to a substantial increase in test domain accuracy, clearly demonstrating its effectiveness in fostering domain invariance. Finally, we compare results before and after test-time correction, highlighting that even prior to correction, our method achieves state-of-the-art performance across all source domains on PACS, underscoring the strength of our approach. + +Error analysis of test-time correction: For this analysis, we train TIDE on the photo domain of PACS and test it on the sketch domain. The model gives an initial accuracy of $74.79\%$ , which improves to $82.29\%$ post test-time correction. In $72.2\%$ of test samples, TIDE does not invoke test-time correction, with class predictions correct in $93.8\%$ of these cases, demonstrating substantial reliability. The supplement provides examples of cases where the model misclassifications are not picked-up in the signature matching step of test-time correction. In the remaining $27.8\%$ of cases where correction is initiated, $52.5\%$ successfully converge to the correct classification, significantly contributing + +![](images/beb08761c29002c92cfe81769a494f16d1a29a5c849c806991c0f9e7b5e0d0ce.jpg) + +![](images/613dd7df96c68fcd30a263262e6d4f50a52bb68486e394a4c858294a92297ac6.jpg) + +![](images/9f4944b2c14e8b7d3f2cbed3b0366ab4b9cacb1de3ef1aafdaf31225282b744e.jpg) + +![](images/f4c57f73735f26000c245e6994b65cd667fa41c932c183826dd4f89808fe1caf.jpg) + +![](images/808f78ef24d3036fe78e68aad451302f96ebcec21b4656b09c56456f7e7aed9c.jpg) +Figure 6. Illustrative examples of concept level GradCAM maps corresponding to TIDE's predictions, across the four studied datasets. The concept names are displayed beneath the maps, with the target domain and predicted class indicated on the left. More results in supplement. + +![](images/0db9325e759eba3d4f641f5b03044feae90ae56d5b10cebdd0aacac945a4edc7.jpg) + +![](images/045e375c81bf497dad55db2b8779e6d65bdbf75438dc3c6a669be30b8eda9de1.jpg) + +![](images/5b6e65d78a9c46f055e0f6746510e9efe88099346f2f130dc2ae4f02d894a600.jpg) + +![](images/70d048ca715ecbc489e4a59851919a23d6e45cd3b26264d1fedc0ff1b06b6dcb.jpg) +Figure 7. The top row shows initial class predictions and GradCAM maps for the concepts, while the bottom row presents the results after test-time correction. + +to the final post-correction accuracy of $82.29\%$ . + +t-SNE plots: To demonstrate the impact of $\mathcal{L}_{LCC}$ , we present t-SNE visualizations in Figure 5, showing the distribution of concept-specific vectors (as computed in Equation 4). We individually plot the source domain (photo) along with sketch, cartoon, and painting domains as targets. Each case includes two plots: one without $\mathcal{L}_{LCC}$ (top row) and one with it (bottom row). We observe that with $\mathcal{L}_{LCC}$ , concept samples (e.g., mouth) align closely across domains, while distinct separation occurs between different concepts (e.g., mouth and horns). Without $\mathcal{L}_{LCC}$ , cluster separation and alignment across domains are weaker, highlighting its role in improving intra-concept compactness and inter-concept separability. + +# 4.3. Qualitative Results + +Concept Localization: We train the SSDG model on photo domain for PACS, OfficeHome, and DomainNet and on Caltech101 for VLCS. The predicted class, concepts and corresponding saliency maps are shown in Figure 6. For each target class (e.g., mushroom, kettle, guitar), the model reliably highlights key concept-specific regions (e.g., + +stem, handle, strings) essential for classification. The model effectively isolates key features across diverse contexts, such as the legs and seat of a chair in complex zoomed-out scenes, the beak and feathers of camouflaged birds, the eyes and whiskers in deformed sketches of cats, and varying feature sizes, with the wick occupying a small area and the mushroom cap spanning a larger one. + +Test-time Correction: The Figure 7 visually illustrates the efficacy of TIDE's test time correction abilities, by comparing the initial and corrected results. The model initially mis-classifies the images as dog and elephant, resulting in poorly aligned concept-level saliency maps for corresponding concepts i.e. ears, snout, tusks and trunk. TIDE detects and rectifies such errors during inference, leading to accurate classification accompanied by precise concept localization. This qualitative evaluation reinforces the robustness of our approach in refining predictions and generating reliable, class-specific concept maps. + +# 5. Conclusion + +In this work, we considered the problem of single-source domain generalization and observed that the current state-of-the-art methods fail in cases of semantic domain shifts primarily due to the global nature of their learned features. To alleviate this issue, we proposed TIDE, a new approach that not only learns local concept representations but also produces localization maps for these concepts. With these maps, we showed we could visually interpret model decisions while also enabling correction of these decisions at test time using our iterative attention refinement strategy. Extensive experimentation on standard benchmarks demonstrated substantial and consistent performance gains over the current state-of-the-art. Future work will rigorously explore methods to generate confidence scores grounded in TIDE's concept verification strategy and would explore the application of TIDE on other (e.g. fine-grained) classification datasets. + +# References + +[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In International Conference on Computer Vision (ICCV), 2023. 3, 6 +[2] Tom B Brown. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 2020. 3 +[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 3, 6 +[4] Jin Chen, Zhi Gao, Xinxiao Wu, and Jiebo Luo. Meta-causal learning for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7 +[5] Sheng Cheng, Tejas Gokhale, and Yezhou Yang. Adversarial bayesian augmentation for single-source domain generalization. In International Conference on Computer Vision (ICCV), 2023. 1, 2, 7 +[6] Seokeon Choi, Debasmit Das, Sungha Choi, Seunghan Yang, Hyunsin Park, and Sungrack Yun. Progressive random convolutions for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6, 7 +[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020. 1, 2, 6, 7 +[8] Ilke Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 7 +[9] Terrance Devries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. ArXiv, abs/1708.04552, 2017. 7 +[10] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In European Conference on Computer Vision (ECCV), 2020. 2 +[11] Antonio D'Innocente and Barbara Caputo. Domain generalization with domain-specific aggregation modules. In The German Conference on Pattern Recognition, 2019. 2 +[12] Xinjie Fan, Qifei Wang, Junjie Ke, Feng Yang, Boqing Gong, and Mingyuan Zhou. Adversarily adaptive normalization for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6, 7 +[13] Chen Fang, Ye Xu, and Daniel N Rockmore. Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 4, 6 +[14] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. + +Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations (ICLR), 2019. 2 +[15] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In International Conference on Computer Vision (ICCV), 2015. 2 +[16] Tejas Gokhale, Rushil Anirudh, Bhavya Kailkhura, Jayaraman J Thiagarajan, Chitta Baral, and Yezhou Yang. Attribute-guided adversarial training for robustness to natural perturbations. In Association for the Advancement of Artificial Intelligence (AAAI), 2021. 2 +[17] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. International Conference on Learning Representations (ICLR), 2020. 6 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4 +[19] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. International Conference on Learning Representations (ICLR), 2020. 1, 2, 6, 7 +[20] Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. Pixmix: Dreamlike pictures comprehensively improve safety measures. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7 +[21] Yann N. Dauphin David Lopez-Paz Hongyi Zhang, Moustapha Cisse. mixup: Beyond empirical risk minimization. International Conference on Learning Representations (ICLR), 2018. 7 +[22] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 6, 7 +[23] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in Neural Information Processing Systems (NeurIPS), 2015. 2 +[24] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning (ICML), 2020. 3 +[25] Wouter M Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 43(3):766-785, 2019. 1 +[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2012. 4 +[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In International Conference on Computer Vision (ICCV), 2017. 4, 6 + +[28] Deng Li, Aiming Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 6, 7 +[29] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. International Conference on Learning Representations (ICLR), 2022. 2, 6, 7 +[30] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In European Conference on Computer Vision (ECCV), 2018. 1 +[31] Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. ArXiv, abs/2108.13624, 2021. 1 +[32] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning (ICML), 2015. 1 +[33] Andrei Margeloiu, Matthew Ashman, Umang Bhatt, Yanzhi Chen, Mateja Jamnik, and Adrian Weller. Do concept bottleneck models learn as intended? International Conference on Learning Representations (ICLR), 2021. 3 +[34] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In International Conference on Computer Vision (ICCV), 2017. 2 +[35] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning (ICML), 2013. 2 +[36] Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6, 7 +[37] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. International Conference on Learning Representations (ICLR), 2023. 3 +[38] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In International Conference on Computer Vision (ICCV), 2019. 4, 6 +[39] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation. In European Conference on Computer Vision (ECCV), 2024. 6 +[40] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1 +[41] Sanqing Qu, Yingwei Pan, Guang Chen, Ting Yao, Changjun Jiang, and Tao Mei. Modality-agnostic debiasing for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6 + +[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In International Conference on Computer Vision (ICCV), 2017. 1, 2, 3, 4 +[44] Sarath Sivaprasad, Akshay Goindani, Mario Fritz, and Vineet Gandhi. Class-wise domain generalization: A novel framework for evaluating distributional shift. In NeurIPS Workshop on Distribution Shifts: Connecting Methods and Applications, 2022. 1, 2 +[45] Nathan Somavarapu, Chih-Yao Ma, and Zsolt Kira. Frustratingly simple domain generalization via image stylization. ArXiv, abs/2006.11207, 2020. 2 +[46] Shuhan Tan, Xingchao Peng, and Kate Saenko. Class-imbalanced domain adaptation: An empirical odyssey. In European Conference on Computer Vision Workshops (ECCVW), 2020. 6 +[47] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 4 +[48] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2 +[49] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1 +[50] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 6 +[51] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in Neural Information Processing Systems (NeurIPS), 2018. 1 +[52] Bor-Shiun Wang, Chien-Yi Wang, and Wei-Chen Chiu. Mcpnet: An interpretable classifier via multi-level concept prototypes. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[53] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. International Conference on Learning Representations (ICLR), 2019. 2 +[54] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 1 +[55] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In International Conference on Computer Vision (ICCV), 2021. 6, 7 +[56] Zehao Xiao, Jiayi Shen, Xiantong Zhen, Ling Shao, and Cees Snoek. A bit more bayesian: Domain-invariant learning with uncertainty. In International Conference on Machine Learning (ICML), 2021. 2 + +[57] Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain expansion approach for single-source domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[58] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[59] Jianhao Yuan, Francesco Pinto, Adam Davies, and Philip Torr. Not just pretty pictures: Toward interventional data augmentation using text-to-image generators. International Conference on Machine Learning (ICML), 2024. 1, 2, 6, 7 +[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision (ICCV), 2019. 6, 7 +[61] Hanlin Zhang, Yi-Fan Zhang, Weiyang Liu, Adrian Weller, Bernhard Schölkopf, and Eric P Xing. Towards principled disentanglement for domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[62] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 7 +[63] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. International Conference on Learning Representations (ICLR), 2021. 1, 2, 6, 7 +[64] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation. International Journal of Computer Vision (IJCV), 132(3):822-836, 2024. 2 \ No newline at end of file diff --git a/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/images.zip b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e8ef3421e8f5eae944516777324dbcdb887c1206 --- /dev/null +++ b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4886c29aaca506aef45a4559e7eebafc950076a10aaca8bf5dd10912981ceec1 +size 649424 diff --git a/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/layout.json b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..41cf912e510777362f9437d763f81893a685a488 --- /dev/null +++ b/2025/TIDE_ Training Locally Interpretable Domain Generalization Models Enables Test-time Correction/layout.json @@ -0,0 +1,11008 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "type": "text", + "content": "TIDE: Training Locally Interpretable Domain Generalization Models Enables Test-time Correction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 160, + 468, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 160, + 468, + 176 + ], + "spans": [ + { + "bbox": [ + 146, + 160, + 468, + 176 + ], + "type": "text", + "content": "Aishwarya Agarwal1,2* Srikrishna Karanam2† Vineet Gandhi1‡" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 182, + 428, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 182, + 428, + 194 + ], + "spans": [ + { + "bbox": [ + 182, + 182, + 428, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 182, + 182, + 428, + 194 + ], + "type": "text", + "content": "CVIT, Kohli Centre for Intelligent Systems, IIIT Hyderabad, India" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 239, + 196, + 369, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 196, + 369, + 209 + ], + "spans": [ + { + "bbox": [ + 239, + 196, + 369, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 239, + 196, + 369, + 209 + ], + "type": "text", + "content": "Adobe Research, Bengaluru, India" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 236, + 200, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 236, + 200, + 249 + ], + "spans": [ + { + "bbox": [ + 151, + 236, + 200, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 262, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 262, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 297, + 586 + ], + "type": "text", + "content": "We consider the problem of single-source domain generalization. Existing methods typically rely on extensive augmentations to synthetically cover diverse domains during training. However, they struggle with semantic shifts (e.g., background and viewpoint changes), as they often learn global features instead of local concepts that tend to be domain invariant. To address this gap, we propose an approach that compels models to leverage such local concepts during prediction. Given no suitable dataset with per-class concepts and localization maps exists, we first develop a novel pipeline to generate annotations by exploiting the rich features of diffusion and large-language models. Our next innovation is TIDE, a novel training scheme with a concept saliency alignment loss that ensures model focus on the right per-concept regions and a local concept contrastive loss that promotes learning domain-invariant concept representations. This not only gives a robust model but also can be visually interpreted using the predicted concept saliency maps. Given these maps at test time, our final contribution is a new correction algorithm that uses the corresponding local concept representations to iteratively refine the prediction until it aligns with prototypical concept representations that we store at the end of model training. We evaluate our approach extensively on four standard DG benchmark datasets and substantially outperform the current state-of-the-art (12% improvement on average) while also demonstrating that our predictions can be visually interpreted." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 608, + 135, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 608, + 135, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 608, + 135, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 677 + ], + "type": "text", + "content": "Enhancing deep neural networks to generalize to out-of-distribution samples remains a core challenge in machine learning and computer vision research, as real-world test data often diverges significantly from the training distribu" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 238, + 560, + 376 + ], + "blocks": [ + { + "bbox": [ + 317, + 238, + 560, + 376 + ], + "lines": [ + { + "bbox": [ + 317, + 238, + 560, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 238, + 560, + 376 + ], + "type": "image", + "image_path": "69e775bd3f4808f78e554448d4f737adea66d3e269f1010adacc017ec20251c5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 387, + 556, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 464 + ], + "type": "text", + "content": "Figure 1. Samples from VLCS (left) and PACS dataset (right) across domain shifts, corresponding to bird and person class. First row displays GradCAM maps [43] for ABA's class predictions. We observe that model attention of ABA [5] falters across domain shifts. Second and third row display the concept specific GradCAM maps from TIDE. We posit that accurate concept learning and localization facilitates DG." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 653 + ], + "type": "text", + "content": "tion [31]. The challenges compound when obtaining labeled samples from the target domain is expensive or unfeasible, hindering application of semi-supervised learning or domain adaptation [25, 32, 49]. The problem of Domain Generalization (DG) [30, 44, 54, 61, 63] represents a promising avenue for developing techniques that capture domain-invariant patterns and improve performance on out-of-distribution samples. In this paper, we focus on Single Source Domain Generalization (SSDG), where a model is trained on data from a single domain and aims to generalize well to unseen domains [40, 51]. It represents the most strict form of DG, as the model must extract domain-invariant features from a single, often limited perspective, without exposure to the variation present across multiple domains." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Most previous work on SSDG [5, 7, 8, 19, 20, 22, 59, 63] relies extensively on data augmentations to support the learning of domain-generalized features. The premise is that constructing an extensive repertoire of augmentations synthesizes instances encompassing a wide spectrum of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 685, + 247, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 685, + 247, + 694 + ], + "spans": [ + { + "bbox": [ + 66, + 685, + 247, + 694 + ], + "type": "text", + "content": "\\* aishwarya.agarwal@research.iit.ac.in, aishagar@adobe.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 694, + 137, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 694, + 137, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 694, + 137, + 703 + ], + "type": "text", + "content": "† skaranam@adobe.com" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 125, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 125, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 125, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 703, + 125, + 712 + ], + "type": "text", + "content": "VGandhi@iiit.ac.in" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30210" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "human-recognizable domains. However, accounting for all conceivable real-world augmentations presents an immense challenge. These models have shown reasonable success in addressing variations in style and texture [45, 64]; however, their performance remains modest when faced with more semantic domain shifts, such as changes in background and viewpoint e.g., in the VLCS dataset [41, 48]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 295, + 397 + ], + "type": "text", + "content": "Consider the results of the state-of-the-art ABA [5] model on the two examples in Figure 1. In the first example (left), domain shifts manifest as variations in background and viewpoint. In the second example (right), the training data comprises photos of human faces, whereas the test data consists of skeletal sketches. In both cases, the domain shifts extend beyond style or texture variations, with ABA failing to correctly classify the samples. Class-level saliency maps [43] reveal that this misclassification stems from inadequate focus on critical local concepts under domain shifts, such as beak and feathers for birds, or eyes and lips for persons. Such local concepts are integral to class definition and remain invariant across domains, and a robust DG model must adeptly capture these stable features. We posit that prior efforts learn global features per class; and if the model fails to learn the correct feature set in the training domain, its generalization performance is compromised as noted from Figure 1 above. The inconsistency in concept localization further exacerbates the interpretability and explainability of these models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 400, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 295, + 603 + ], + "type": "text", + "content": "We adopt an alternative approach, wherein, rather than attempting to encompass all potential augmentations, we compel the model to leverage essential class-specific concepts for classification. Our key idea is to force the model to attend to these concepts during training. The primary hurdle in this path is lack of annotated data specifying relevant class-specific concepts along with their corresponding localizations. To this end, our first contribution is a novel pipeline that harnesses large language models (LLMs) and diffusion models (DMs) to identify key concepts and generate concept-level saliency maps in a scalable, automated fashion. We demonstrate that DMs, via cross-attention layers, can generate concept-level saliency maps for generated images with notable granularity. Given these maps for a single synthesized image within a class, we leverage the rich feature space of DMs to transfer them to images across diverse domains in DG benchmarks [47]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "We subsequently introduce TIDE, our second contribution which employs cross-entropy losses over class and concept labels along with a novel pair of loss functions: a concept saliency alignment loss that ensures the model attends to the correct regions for each local concept, and a local concept contrastive loss that promotes learning domain invariant features for these regions. As shown in Figure 1, TIDE consistently attends to local concepts such as beak and eyes in the training domain as well as across substan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "tial domain shifts. This precise localization significantly bolsters SSDG performance while enabling the generation of visual explanations for model decisions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 108, + 555, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 275 + ], + "type": "text", + "content": "Marking a major leap in model interpretability and performance gains, our third contribution is to demonstrate how our model's concept-level localization can be effectively leveraged for performance verification and test-time correction. To this end, we introduce the notion of local concept signatures: prototypical features derived by pooling concept-level features across training samples, guided by corresponding saliency maps. If the concept features associated with the predicted class label do not align with their signatures, it signals the use of incorrect features for prediction. Consequently, we employ iterative refinement through concept saliency masking until concept predictions align with their corresponding signatures. More formally, our paper makes following contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 277, + 553, + 456 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 277, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 277, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 553, + 323 + ], + "type": "text", + "content": "- We propose a novel synthetic annotation pipeline to automatically identify key per-class local concepts and their localization maps, and transfer them to real images found in DG benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 325, + 553, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 325, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 314, + 325, + 553, + 383 + ], + "type": "text", + "content": "- We propose a novel approach, TIDE, to train locally interpretable models with a novel concept saliency alignment loss that ensures accurate concept localization and a novel local concept contrastive loss that ensures domain-invariant features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 384, + 553, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 553, + 420 + ], + "type": "text", + "content": "- We show TIDE enables correcting model prediction at test time using the predicted concept saliency maps as part of a novel iterative attention refinement strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 421, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 553, + 456 + ], + "type": "text", + "content": "- We report extensive results on four standard DG benchmarks, outperforming the current state-of-the-art by a significant " + }, + { + "bbox": [ + 314, + 421, + 553, + 456 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 314, + 421, + 553, + 456 + ], + "type": "text", + "content": " on average." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 466, + 405, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 466, + 405, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 466, + 405, + 479 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "type": "text", + "content": "Multi Source domain generalization (MSDG) methods assume access to multiple source domains and domain-specific knowledge during training and have proposed ways to learn multi-source domain-invariant features [10, 15, 34, 35, 53], utilize domain labels to capture domain shifts [11, 35, 56], and design target-specific augmentation strategies [14, 16, 23]. However, these assumptions are not practical for real-world applications [44, 57] and we instead focus on the more challenging SSDG setting, where only one source domain and no prior target knowledge is available." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": "Most SSDG approaches have used augmentations to improve DG [5, 7, 8, 19, 20, 22, 59, 63]. While [59] also uses diffusion models (DMs) [42] for data augmentation, our approach differs by leveraging the rich feature space of DMs primarily for offline saliency map annotation instead of augmentation. More recent work in SSDG [4, 28, 29] has also turned to approaches based on domain-specific prompting and causal inference instead of data augmentations. However, these methods learn global features, limiting invari" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30211" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 350, + 259 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 350, + 259 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 350, + 259 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 350, + 259 + ], + "type": "image", + "image_path": "ba52611ef9468bac83a7dd8e044023b54f8e7859b587eb0effe60c809fe0d82a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "type": "text", + "content": "Figure 2. The TIDE pipeline: Left—Training on a single domain with cross-entropy losses for class " + }, + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{c}})" + }, + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "type": "text", + "content": " and concept labels " + }, + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{k}})" + }, + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "type": "text", + "content": ", alongside Concept Saliency Alignment " + }, + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{CSA}})" + }, + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "type": "text", + "content": " and Local Concept Contrastive losses " + }, + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{LCC}})" + }, + { + "bbox": [ + 54, + 269, + 555, + 293 + ], + "type": "text", + "content": ". Right—Test-time correction strategy applied in TIDE." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 352, + 70, + 559, + 260 + ], + "blocks": [ + { + "bbox": [ + 352, + 70, + 559, + 260 + ], + "lines": [ + { + "bbox": [ + 352, + 70, + 559, + 260 + ], + "spans": [ + { + "bbox": [ + 352, + 70, + 559, + 260 + ], + "type": "image", + "image_path": "02e40050d826f8975000ebd2ba02dd9c246c68c531b47cd493f08fb1304c7707.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 312, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 295, + 337 + ], + "type": "text", + "content": "ance across unseen domains. Finally, these methods also do not provide any signals to interpret model predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 337, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 337, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 295, + 469 + ], + "type": "text", + "content": "On the other hand, while methods like Concept Bottle-neck Models (CBMs) [24, 33] also learn local concepts, they are restricted to a predefined set of concepts and more crucially, cannot ground them to image regions that these concepts represent. Recent methods [37, 58] have scaled CBMs to a large number of classes and also proposed to ways to learn multi-scale representations [52], but they are also unable to connect these concepts to image pixels. Our SSDG method addresses these challenges by producing local concept saliency maps, enhancing both DG performance and model prediction interpretability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 478, + 138, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 138, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 138, + 491 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 498, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 556 + ], + "type": "text", + "content": "The proposed framework is depicted in Figure 2. Training comprises three components: class prediction, concept prediction with saliency alignment, and localized contrastive training. During inference, class and concept predictions are integrated with a test-time correction mechanism." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": "In the concept alignment phase, we employ concept-level saliency maps as ground truth to enforce focus on interpretable features, using a saliency alignment loss. Domain invariance is further promoted by localized contrastive training, where we train the model to cluster similar concepts (e.g., eyes across augmentations) while separating unrelated concepts (e.g., strings). Finally, a test-time correction mechanism iteratively refines attention by suppressing irrelevant regions, leveraging concept-signatures to redirect focus and improve classification accuracy. We proceed with a concise review of the key notations, followed by an in-depth exposition of our approach and the pipeline for generating ground truth concept-level annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": "We assume the source data is observed from a single domain, denoted as " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "D = \\{x_{i},y_{i}^{c},y_{i}^{k}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": "-th image, " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "y_{i}^{c}" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": " the class label, " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "y_{i}^{k}" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": " the concept label, and " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": " the total sample count in the source domain. The shared CNN backbone is used to obtain the backbone features " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "F_{x}\\in \\mathbb{R}^{W\\times H\\times C}" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": ". The automatically generated ground truth concept-level saliency maps are denoted by " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "G_{x}^{k}" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": ". The GradCAM [43] saliency maps corresponding to class and concepts labels are denoted by " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "S_{x}^{c}" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "inline_equation", + "content": "S_{x}^{k}" + }, + { + "bbox": [ + 313, + 312, + 555, + 422 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 516, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 516, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 516, + 443 + ], + "type": "text", + "content": "3.1. Generating Concept-level Annotations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 520 + ], + "type": "text", + "content": "During training, we aim our model to identify and spatially attend to stable, discriminative regions. However, existing DG datasets lack fine-grained, concept-level annotations for such regions. To address this, we propose a novel pipeline that uses LLMs and DMs to automate scalable concept-level saliency map generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 689 + ], + "type": "text", + "content": "Our primary insight is that DMs can be harnessed to generate high-quality, concept-level saliency maps for synthesized images. Extracting cross-attention maps [1, 3] from a DM, given the prompt with concepts and corresponding synthesized image, yields highly granular concept-level saliency maps. Figure 3 demonstrates how DMs emphasize specific regions for distinct concepts, capturing fine-grained attention to features such as a cat's whiskers or a snowman's hands. With this technique serving as an efficient tool for yielding concept maps for synthesized images, the ensuing questions are: (i) how to automate the identification of pertinent concepts for each class across datasets, and (ii) how to transfer these concept maps to real-world images to generate ground-truth annotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "First, to identify the key concepts associated with each class, we use GPT-3.5 [2], which we prompt to generate a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30212" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 87, + 120, + 140 + ], + "blocks": [ + { + "bbox": [ + 94, + 74, + 267, + 83 + ], + "lines": [ + { + "bbox": [ + 94, + 74, + 267, + 83 + ], + "spans": [ + { + "bbox": [ + 94, + 74, + 267, + 83 + ], + "type": "text", + "content": "Prompt: A photo of a cat with its whiskers, ears, eyes" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 87, + 120, + 140 + ], + "lines": [ + { + "bbox": [ + 65, + 87, + 120, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 87, + 120, + 140 + ], + "type": "image", + "image_path": "6dee8d9d5cd5668554aad2e8dcca46d94b0fb597610b3ea26b16a7c15d8f1f03.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 121, + 87, + 178, + 142 + ], + "blocks": [ + { + "bbox": [ + 121, + 87, + 178, + 142 + ], + "lines": [ + { + "bbox": [ + 121, + 87, + 178, + 142 + ], + "spans": [ + { + "bbox": [ + 121, + 87, + 178, + 142 + ], + "type": "image", + "image_path": "775a048924f7abfb9f892c1dd2e4fb4897c56941137a006f6ef23b8a143c0ec1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 87, + 233, + 142 + ], + "blocks": [ + { + "bbox": [ + 178, + 87, + 233, + 142 + ], + "lines": [ + { + "bbox": [ + 178, + 87, + 233, + 142 + ], + "spans": [ + { + "bbox": [ + 178, + 87, + 233, + 142 + ], + "type": "image", + "image_path": "0714bdac2ba4c5c47b5cdd01d830ea1ee5c52d117e8ee1c2a836c6efdd60de2e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 233, + 87, + 289, + 142 + ], + "blocks": [ + { + "bbox": [ + 233, + 87, + 289, + 142 + ], + "lines": [ + { + "bbox": [ + 233, + 87, + 289, + 142 + ], + "spans": [ + { + "bbox": [ + 233, + 87, + 289, + 142 + ], + "type": "image", + "image_path": "dfeeb94d1db27c4cd2d431223fb422eadf00e0b2602bbe1e19901e13eee3c9fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 65, + 157, + 121, + 213 + ], + "blocks": [ + { + "bbox": [ + 86, + 145, + 271, + 156 + ], + "lines": [ + { + "bbox": [ + 86, + 145, + 271, + 156 + ], + "spans": [ + { + "bbox": [ + 86, + 145, + 271, + 156 + ], + "type": "text", + "content": "Prompt: A photo of a snowman with its nose, eyes, hands" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 157, + 121, + 213 + ], + "lines": [ + { + "bbox": [ + 65, + 157, + 121, + 213 + ], + "spans": [ + { + "bbox": [ + 65, + 157, + 121, + 213 + ], + "type": "image", + "image_path": "96d19e966bb870c786630b06fb675b87a2a88b9330d83bf68cc860e1c16b5304.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 213, + 119, + 222 + ], + "lines": [ + { + "bbox": [ + 67, + 213, + 119, + 222 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 119, + 222 + ], + "type": "text", + "content": "Generated Image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 235, + 295, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 295, + 269 + ], + "type": "text", + "content": "Figure 3. The first column displays the image generated from the given prompt, while the subsequent three columns show the cross-attention maps corresponding to each concept in the prompt." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 122, + 157, + 178, + 213 + ], + "blocks": [ + { + "bbox": [ + 122, + 157, + 178, + 213 + ], + "lines": [ + { + "bbox": [ + 122, + 157, + 178, + 213 + ], + "spans": [ + { + "bbox": [ + 122, + 157, + 178, + 213 + ], + "type": "image", + "image_path": "da34b905aaa9a8c9d19e53c33627ff1dde704fe8d15c0194268a84b9ea363bac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 213, + 252, + 222 + ], + "lines": [ + { + "bbox": [ + 157, + 213, + 252, + 222 + ], + "spans": [ + { + "bbox": [ + 157, + 213, + 252, + 222 + ], + "type": "text", + "content": "Extracted Cross Attention Maps" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 178, + 157, + 233, + 213 + ], + "blocks": [ + { + "bbox": [ + 178, + 157, + 233, + 213 + ], + "lines": [ + { + "bbox": [ + 178, + 157, + 233, + 213 + ], + "spans": [ + { + "bbox": [ + 178, + 157, + 233, + 213 + ], + "type": "image", + "image_path": "4013b509dcba52308e124b6a435d13726b3de40a64bbe458bdc5a901dea49aef.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 233, + 157, + 289, + 213 + ], + "blocks": [ + { + "bbox": [ + 233, + 157, + 289, + 213 + ], + "lines": [ + { + "bbox": [ + 233, + 157, + 289, + 213 + ], + "spans": [ + { + "bbox": [ + 233, + 157, + 289, + 213 + ], + "type": "image", + "image_path": "b580fdf126801013a7ade642429355bba544e9980ec0d010e213e7714fadbee3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 290, + 296, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 375 + ], + "type": "text", + "content": "list of distinctive, stable features (prompt in the supplement) for each class. For instance, GPT-3.5 outputs concepts such as whiskers, ears, and eyes for a cat. We generate a prompt leveraging these concepts, which is then used to synthesize an exemplar image for each class. We further derive the corresponding concept-level attention maps as outlined in the preceding paragraph." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 376, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 554 + ], + "type": "text", + "content": "Given a single synthesized exemplar image for each class and the concept-level saliency maps, we turn to the task of transferring these saliency annotations to real-world images. DM's feature space is particularly well-suited for this, as it captures detailed, semantically-rich representations that allow us to match synthetic concept regions to real images across domains. Leveraging this, we use the Diffusion Feature Transfer (DIFT) method [47], which computes pixel-level correspondences by comparing cosine similarities between DM features. Through this approach, we establish a region-to-region correspondence between synthetic saliency maps (e.g., mouth of a dog) and similar regions in real-world images. This enables us to generate comprehensive, consistent concept-level annotations across domains as shown in Figure 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "content": "We obtain these concept-level annotations for widely-used benchmarks, including VLCS [13], PACS [27], DomainNet [38], and Office-Home [50]. For an image " + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "content": ", the binarized concept-level maps " + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "inline_equation", + "content": "(G_x^k)" + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "content": " are henceforth referred to as ground-truth saliency maps (GT-maps) in this paper. Having established a process for generating saliency maps, the next challenge is identifying the subset of concepts that the model actually relies on for making predictions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 670, + 179, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 179, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 179, + 681 + ], + "type": "text", + "content": "3.1.1 Concepts that matter" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We restrict our method to essential concepts, filtering out those that do not contribute meaningfully to classification." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 343, + 83, + 520, + 137 + ], + "blocks": [ + { + "bbox": [ + 374, + 73, + 504, + 83 + ], + "lines": [ + { + "bbox": [ + 374, + 73, + 504, + 83 + ], + "spans": [ + { + "bbox": [ + 374, + 73, + 504, + 83 + ], + "type": "text", + "content": "Prompt: A photo of a dog with its ear, mouth" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 83, + 520, + 137 + ], + "lines": [ + { + "bbox": [ + 343, + 83, + 520, + 137 + ], + "spans": [ + { + "bbox": [ + 343, + 83, + 520, + 137 + ], + "type": "image", + "image_path": "d29618652d5205159ff10bac791cd7efd31973a16d30ee4a603223deba797526.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 138, + 381, + 146 + ], + "lines": [ + { + "bbox": [ + 337, + 138, + 381, + 146 + ], + "spans": [ + { + "bbox": [ + 337, + 138, + 381, + 146 + ], + "type": "text", + "content": "PACS - photo" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 138, + 435, + 147 + ], + "lines": [ + { + "bbox": [ + 386, + 138, + 435, + 147 + ], + "spans": [ + { + "bbox": [ + 386, + 138, + 435, + 147 + ], + "type": "text", + "content": "PACS - cartoon" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 138, + 487, + 147 + ], + "lines": [ + { + "bbox": [ + 442, + 138, + 487, + 147 + ], + "spans": [ + { + "bbox": [ + 442, + 138, + 487, + 147 + ], + "type": "text", + "content": "PACS - sketch" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 331, + 147, + 384, + 198 + ], + "blocks": [ + { + "bbox": [ + 331, + 147, + 384, + 198 + ], + "lines": [ + { + "bbox": [ + 331, + 147, + 384, + 198 + ], + "spans": [ + { + "bbox": [ + 331, + 147, + 384, + 198 + ], + "type": "image", + "image_path": "c4fd72ee987295925f35d21eb0276da075ac6daff6707f332d502e2949c671d9.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 386, + 148, + 437, + 198 + ], + "blocks": [ + { + "bbox": [ + 386, + 148, + 437, + 198 + ], + "lines": [ + { + "bbox": [ + 386, + 148, + 437, + 198 + ], + "spans": [ + { + "bbox": [ + 386, + 148, + 437, + 198 + ], + "type": "image", + "image_path": "ac16b1af529de362cdc7f19c2931db27632963bf5dd3b7c00a189aef91fbe3c5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 447, + 148, + 485, + 195 + ], + "blocks": [ + { + "bbox": [ + 447, + 148, + 485, + 195 + ], + "lines": [ + { + "bbox": [ + 447, + 148, + 485, + 195 + ], + "spans": [ + { + "bbox": [ + 447, + 148, + 485, + 195 + ], + "type": "image", + "image_path": "bbfdcee532654e85acc9ba737890fcdaff6c6ee34032e52ad2f177d07cbe6c6f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 493, + 147, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 493, + 138, + 544, + 147 + ], + "lines": [ + { + "bbox": [ + 493, + 138, + 544, + 147 + ], + "spans": [ + { + "bbox": [ + 493, + 138, + 544, + 147 + ], + "type": "text", + "content": "PACS - painting" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 147, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 493, + 147, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 493, + 147, + 545, + 198 + ], + "type": "image", + "image_path": "f0174a03f238e6b72fd49c175772434e24b00fed9a966a9b37d62cc100e8fe34.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 326, + 199, + 384, + 250 + ], + "blocks": [ + { + "bbox": [ + 326, + 199, + 384, + 250 + ], + "lines": [ + { + "bbox": [ + 326, + 199, + 384, + 250 + ], + "spans": [ + { + "bbox": [ + 326, + 199, + 384, + 250 + ], + "type": "image", + "image_path": "ca02fb103336e15dc0db5d2712f1727c598bc38530be48221d2c73021b5ccc9a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 384, + 199, + 436, + 250 + ], + "blocks": [ + { + "bbox": [ + 384, + 199, + 436, + 250 + ], + "lines": [ + { + "bbox": [ + 384, + 199, + 436, + 250 + ], + "spans": [ + { + "bbox": [ + 384, + 199, + 436, + 250 + ], + "type": "image", + "image_path": "2ccd2e6a534ce7fb0dfd0e77432ba5f1959acd83830a4b60d7f77b9b33f89ca7.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 439, + 199, + 491, + 250 + ], + "blocks": [ + { + "bbox": [ + 439, + 199, + 491, + 250 + ], + "lines": [ + { + "bbox": [ + 439, + 199, + 491, + 250 + ], + "spans": [ + { + "bbox": [ + 439, + 199, + 491, + 250 + ], + "type": "image", + "image_path": "6c8e2f311ae85750fa4a8e75986439e9f999e7c7caf82d8750245b52e6008117.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 492, + 199, + 544, + 250 + ], + "blocks": [ + { + "bbox": [ + 492, + 199, + 544, + 250 + ], + "lines": [ + { + "bbox": [ + 492, + 199, + 544, + 250 + ], + "spans": [ + { + "bbox": [ + 492, + 199, + 544, + 250 + ], + "type": "image", + "image_path": "e0c294a9a6a384faf3dd8902748b24a326c7d80b70053bda84b09b0f24efa158.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 325, + 251, + 384, + 302 + ], + "blocks": [ + { + "bbox": [ + 325, + 251, + 384, + 302 + ], + "lines": [ + { + "bbox": [ + 325, + 251, + 384, + 302 + ], + "spans": [ + { + "bbox": [ + 325, + 251, + 384, + 302 + ], + "type": "image", + "image_path": "ec4b65144e3a1a5f2260927e76428cfde59845b800810e3e89f71740f168cd5f.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 312, + 554, + 367 + ], + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 367 + ], + "type": "text", + "content": "Figure 4. The first row presents the prompt, corresponding synthesized image and attention maps for the concepts ear and mouth. Below, we demonstrate that using diffusion features correspondences these concept saliency maps from a single exemplar can be automatically transferred on dog images across domains." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 384, + 251, + 436, + 302 + ], + "blocks": [ + { + "bbox": [ + 384, + 251, + 436, + 302 + ], + "lines": [ + { + "bbox": [ + 384, + 251, + 436, + 302 + ], + "spans": [ + { + "bbox": [ + 384, + 251, + 436, + 302 + ], + "type": "image", + "image_path": "57f698054fc670394c3daaf84b776d1c645e33b332999a0825c0d83fcc7ca8b7.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 439, + 251, + 491, + 302 + ], + "blocks": [ + { + "bbox": [ + 439, + 251, + 491, + 302 + ], + "lines": [ + { + "bbox": [ + 439, + 251, + 491, + 302 + ], + "spans": [ + { + "bbox": [ + 439, + 251, + 491, + 302 + ], + "type": "image", + "image_path": "5e784e8684e0a9adea66fdc7a28ee34d59cfe64b8ce704ea733eee31c78f993b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 492, + 251, + 545, + 302 + ], + "blocks": [ + { + "bbox": [ + 492, + 251, + 545, + 302 + ], + "lines": [ + { + "bbox": [ + 492, + 251, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 492, + 251, + 545, + 302 + ], + "type": "image", + "image_path": "625ec29125ed796b939c09db44c76f063b1c54e446a2f046a295a978a76017e3.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "content": "To do this, we train a ResNet-18 [18] classifier on the source domain [26], compute GradCAM [43] saliency maps for each class label, and use them to identify regions the model focuses on when making predictions." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "content": "We compute the overlap between the GradCAM maps and GT-maps for each known concept in the dataset. Given image " + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "content": ", its saliency map " + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "inline_equation", + "content": "S_{x}^{c}" + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "content": " for class " + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "content": " and the GT-map " + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "inline_equation", + "content": "G_{x}^{k}" + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "content": " for concept " + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "content": ", we define the overlap " + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "inline_equation", + "content": "O_{c}^{k}(x)" + }, + { + "bbox": [ + 313, + 437, + 554, + 486 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 352, + 495, + 554, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 495, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 352, + 495, + 554, + 526 + ], + "type": "interline_equation", + "content": "O _ {c} ^ {k} (x) = \\frac {\\sum_ {i , j} \\min \\left(S _ {x} ^ {c} (i , j) , G _ {x} ^ {k} (i , j)\\right)}{\\sum_ {i , j} G _ {x} ^ {k} (i , j)}. \\tag {1}", + "image_path": "0823865b1017d60d0425a68e1dc1c07cb40c80461a1f8df8f9e40930913198bb.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "text", + "content": " are matrix indices. This measures how much of the model's attention for a given class aligns with the regions corresponding to concept " + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "text", + "content": ". We compute this overlap for all concepts and images in the training set, and for each class " + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "text", + "content": ", we define the set of important concepts " + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_c" + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "text", + "content": " as those that consistently exhibit high overlap with the saliency maps:" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 361, + 626, + 554, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 626, + 554, + 659 + ], + "spans": [ + { + "bbox": [ + 361, + 626, + 554, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {K} _ {c} = \\left\\{k \\left| \\frac {1}{N _ {c}} \\sum_ {x \\in \\mathcal {D} _ {c}} O _ {c} ^ {k} (x) > \\tau \\right. \\right\\}, \\tag {2}", + "image_path": "64eea1555e4a85c4ad692084092e802db0ca92263fd0e96a5d906908e2035b68.jpg" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the number of training examples in class " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the set of images in class " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is a threshold that determines the importance of the concept. This procedure functions as a concept discovery module, identifying which" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30213" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 262, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 262, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 262, + 84 + ], + "type": "text", + "content": "local concepts are critical for predicting each class." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 105, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 105, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 105, + 114 + ], + "type": "text", + "content": "3.2. TIDE" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 122, + 296, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 296, + 170 + ], + "type": "text", + "content": "As shown in Figure 2, TIDE utilizes cross-entropy losses " + }, + { + "bbox": [ + 55, + 122, + 296, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 122, + 296, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 122, + 296, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 55, + 122, + 296, + 170 + ], + "type": "text", + "content": " for class and concept labels, respectively, integrated with novel concept saliency alignment and local concept contrastive losses. We detail these loss terms below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 185, + 224, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 185, + 224, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 224, + 197 + ], + "type": "text", + "content": "3.2.1 Concept saliency alignment loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "text", + "content": "For each image " + }, + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "text", + "content": ", we predict important concepts " + }, + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_c" + }, + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "text", + "content": " and enforce alignment between the saliency maps " + }, + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "inline_equation", + "content": "S_x^k" + }, + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "text", + "content": " for predicted concepts and the GT-maps " + }, + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "inline_equation", + "content": "G_x^k" + }, + { + "bbox": [ + 55, + 204, + 296, + 252 + ], + "type": "text", + "content": ". This is encouraged by our proposed concept saliency alignment (CSA) loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 260, + 294, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 260, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 108, + 260, + 294, + 289 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C S A}} = \\frac {1}{| \\mathcal {K} _ {c} |} \\sum_ {k \\in \\mathcal {K} _ {c}} \\| S _ {x} ^ {k} - G _ {x} ^ {k} \\| _ {2} ^ {2}. \\tag {3}", + "image_path": "09d30b784a4e9495e09da1dc71ed61ff2c3d5d43bcc35ca00d12bb2a365cb22c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 294, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 341 + ], + "type": "text", + "content": "Aligning the model's attention with GT-maps enables class-specific reasoning, thereby elucidating the rationale behind predictions by linking them to relevant local features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 357, + 216, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 216, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 216, + 369 + ], + "type": "text", + "content": "3.2.2 Local concept contrastive loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 376, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 459 + ], + "type": "text", + "content": "While the CSA loss facilitates explicit localization of concepts, it is equally essential for these concept-level features to exhibit invariance across domain shifts. To achieve domain invariance, we propose a local concept contrastive (LCC) loss, employing a triplet strategy to cluster similar concepts (e.g., eyes) across domains while distinguishing unrelated ones (e.g., feathers and ears)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": " be an anchor image containing concept " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": ", a positive image " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "x^{+}" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": " (an augmentation of " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": " that retains concept " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": "), and a negative image " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "x^{-}" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": " (containing a different concept " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "k'" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": "). For each image, we compute a concept-specific feature vector " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "f_{x}^{k} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": ", emphasizing the concept's relevant regions using the GT-map " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "G_{x}^{k}" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": ". Each element " + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "inline_equation", + "content": "f_{x}^{k}(l)" + }, + { + "bbox": [ + 55, + 460, + 296, + 533 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 540, + 294, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 540, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 116, + 540, + 294, + 567 + ], + "type": "interline_equation", + "content": "f _ {x} ^ {k} (l) = \\sum_ {i, j} G _ {x} ^ {k} \\cdot F _ {x} (i, j, l). \\tag {4}", + "image_path": "0e4174d8b22543ad7ad43221fbbfc7ad82895215f46c9d4866edbc770f425f76.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": " is an element-wise multiplication operation, " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "G_{x}^{k}" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": " is a matrix representing the GT-map, " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "F_{x}" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": " represents convolutional feature maps from the last convolution layer, and " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "(i,j,l)" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": " channel's " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": " element. These vectors focus on the concepts of interest (concept " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "x^{+}" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "k^{\\prime}" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "inline_equation", + "content": "x^{-}" + }, + { + "bbox": [ + 55, + 573, + 296, + 657 + ], + "type": "text", + "content": "), in contrast to the global features used in prior works. The LCC loss is then defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 72, + 677, + 294, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 677, + 294, + 693 + ], + "spans": [ + { + "bbox": [ + 72, + 677, + 294, + 693 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L C C}} = \\max \\left(0, d \\left(f _ {x} ^ {k}, f _ {x ^ {+}} ^ {k}\\right) - d \\left(f _ {x} ^ {k}, f _ {x ^ {-}} ^ {k ^ {\\prime}}\\right) + \\alpha\\right), \\tag {5}", + "image_path": "2161bfb7d3d3ae5c7079dcb50e7a9ed5827745d767ed3120314938358a17785e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "inline_equation", + "content": "d(.)" + }, + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "text", + "content": " is the euclidean distance and " + }, + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "text", + "content": " is the margin." + } + ] + } + ], + "index": 13 + }, + { + "type": "code", + "bbox": [ + 314, + 85, + 553, + 383 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 489, + 83 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 489, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 489, + 83 + ], + "type": "text", + "content": "Algorithm 1 Iterative Test-time Correction" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "lines": [ + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": "Input: Test image " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": ", initial class prediction " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "c^{(0)}" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": ", signatures " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "p^k" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " for each concept " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \nParameter: Threshold " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": ", max iterations " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \nOutput: Corrected class prediction, " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{final}}" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "S_x^{c(0),0} \\leftarrow \\operatorname{GradCAM}(x, c^{(0)})" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " {Initialize saliency map for iteration " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": "} \n2: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{masked}}^{0} \\leftarrow x" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n3: for each " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "k \\in \\mathcal{K}_{c(0)}" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "f_x^{k,0} \\leftarrow \\sum_{i,j} S_x^k \\cdot F_x(i, j, :)" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "d(f_x^{k,0}, p^k) \\leftarrow 1 - \\frac{f_x^{k,0} \\cdot p^k}{\\|f_x^{k,0}\\| \\|p^k\\|}" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n6: if " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "d(f_x^{k,0}, p^k) > \\delta" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " then \n7: Correction Phase: \n8: for " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "t \\leftarrow 1" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " do \n9: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "M^{c(t-1),t} \\leftarrow \\text{Binarize}(S_x^{c(t-1),t-1})" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{masked}}^t \\leftarrow x_{\\mathrm{masked}}^{t-1} \\cdot M^{c(t-1),t}" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "c^{(t)} \\leftarrow \\text{PredictClass}(x_{\\mathrm{masked}}^t)" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n12: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "S_{x_{\\mathrm{masked}}}^{c(t),t} \\leftarrow \\text{GradCAM}(x_{\\mathrm{masked}}^t, c^{(t)})" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n13: for each " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "k \\in \\mathcal{K}_{c(t)}" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " do \n14: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "f_{x_{\\mathrm{masked}}}^{k,t} \\leftarrow \\sum_{i,j} S_{x_{\\mathrm{masked}}}^{k,t} \\cdot F_x(i,j,:)" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n15: " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "d(f_{x_{\\mathrm{masked}}}^{k,t}, p^k) \\leftarrow 1 - \\frac{f_{x_{\\mathrm{masked}}}^{k,t} \\cdot p^k}{\\|f_{x_{\\mathrm{masked}}}^{k,t}\\| \\|p^k\\|}" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n16: if " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "d(f_{x_{\\mathrm{masked}}}^{k,t}, p^k) \\leq \\delta" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " then \n17: Return " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{final}} \\leftarrow c^{(t)}" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n18: end if \n19: end for \n20: end for \n21: end if \n22: Return " + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{final}} \\leftarrow c^{(0)}" + }, + { + "bbox": [ + 314, + 85, + 553, + 383 + ], + "type": "text", + "content": " \n23: end for" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_body" + } + ], + "index": 15, + "sub_type": "algorithm" + }, + { + "bbox": [ + 314, + 403, + 433, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 403, + 433, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 403, + 433, + 415 + ], + "type": "text", + "content": "3.3. Test-time Correction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 422, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 471 + ], + "type": "text", + "content": "We establish that our localized, interpretable approach facilitates correction of misclassifications through concept-level feature verification. In this section, we first introduce concept signatures and detail the proposed correction strategy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 486, + 453, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 453, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 453, + 498 + ], + "type": "text", + "content": "3.3.1 Local concept signatures" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "text", + "content": "For each concept " + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "text", + "content": ", we define a concept-signature " + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "inline_equation", + "content": "p^k \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "text", + "content": ", as its representative vector. We derive " + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "inline_equation", + "content": "p^k" + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "text", + "content": " by averaging the concept-specific feature vectors " + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "inline_equation", + "content": "f_x^k" + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "text", + "content": " across all training samples " + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{D}" + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "text", + "content": " containing concept " + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "text", + "content": " (denoted as " + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^k" + }, + { + "bbox": [ + 313, + 504, + 554, + 552 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 391, + 561, + 553, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 561, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 391, + 561, + 553, + 590 + ], + "type": "interline_equation", + "content": "p ^ {k} = \\frac {1}{\\left| \\mathcal {D} ^ {k} \\right|} \\sum_ {x \\in \\mathcal {D} ^ {k}} f _ {x} ^ {k}. \\tag {6}", + "image_path": "8085ad35bffd0672bf8842d32e6371f5b9a785520d763addba2c6f91b62fd4ff.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 595, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 554, + 642 + ], + "type": "text", + "content": "These vectors act as reference points, helping the model recognize when its attention is aligned with the right concepts during prediction, even when encountering new, unseen domains." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 658, + 531, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 531, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 531, + 670 + ], + "type": "text", + "content": "3.3.2 Detecting and Correcting misclassifications" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Consider the example of test-time correction in Figure 2, where the model misclassifies a person as a guitar. The predicted class guitar, involves concepts like strings and" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30214" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "knobs, but these are absent in the image. As a result, the model erroneously focuses on irrelevant features (e.g., the person's legs or the background), as reflected in the Grad-CAM saliency maps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 215 + ], + "type": "text", + "content": "To address misclassifications, we employ a two-step approach: possible mistake detection followed by correction. The process is outlined in Algorithm 1. First, the model extracts concept-level saliency maps " + }, + { + "bbox": [ + 55, + 121, + 294, + 215 + ], + "type": "inline_equation", + "content": "(S_x^k)" + }, + { + "bbox": [ + 55, + 121, + 294, + 215 + ], + "type": "text", + "content": ", for all the concepts corresponding to the predicted class, which are then used to compute concept-level features " + }, + { + "bbox": [ + 55, + 121, + 294, + 215 + ], + "type": "inline_equation", + "content": "f_{x}^{k}" + }, + { + "bbox": [ + 55, + 121, + 294, + 215 + ], + "type": "text", + "content": " (step 4). In Step 5 and 6, we compare the features to the stored concept signatures; a deviation exceeding " + }, + { + "bbox": [ + 55, + 121, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 121, + 294, + 215 + ], + "type": "text", + "content": " signals a misclassification." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 294, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 360 + ], + "type": "text", + "content": "Once a misalignment is detected, the model enters an iterative refinement phase. The features prominent for the current class level predictions are masked through the corresponding saliency map " + }, + { + "bbox": [ + 55, + 216, + 294, + 360 + ], + "type": "inline_equation", + "content": "S_{x}^{c}" + }, + { + "bbox": [ + 55, + 216, + 294, + 360 + ], + "type": "text", + "content": " (step 9 and 10). The masking process is cumulative, in effect utilizing all masks from the first to the " + }, + { + "bbox": [ + 55, + 216, + 294, + 360 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 55, + 216, + 294, + 360 + ], + "type": "text", + "content": " iteration. The masked features are then used for subsequent predictions (step 11). This process continues until the concepts corresponding to the predicted class aligns with concept-signatures or the maximum iteration count is reached (steps 8-16 in the algorithm). If alignment is achieved, the predicted class in that iteration is confirmed as the final output." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 136, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 136, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 136, + 384 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 570 + ], + "type": "text", + "content": "Datasets: We conduct experiments on four widely used datasets - PACS [27], VLCS [13], OfficeHome [50], and DomainNet [38], within the DomainBed [17] evaluation benchmarks. PACS contains 9,991 images across 7 categories in four domains: 'sketch', 'photo', 'clipart', and 'painting'. VLCS consists of 10,729 images over 5 categories and 4 domains. The domains are from VOC2007 (V), LabelMe (L), Caltech101 (C), and SUN09 (S), with domain shifts primarily driven by background and viewpoint variations. OfficeHome, with 15,500 images across 65 categories, emphasizes indoor classes across 'product', 'real', 'clipart', and 'art' domains. For DomainNet, we follow prior work [39, 46] and use a subset of the 40 most common classes across 'sketch', 'real', 'clipart', and 'painting'." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Experimental Setup: We adhere to the SSDG paradigm, training on one source domain and testing across three target domains, i.e., the model is trained independently four times, with the averaged test accuracies across target domains reported in each case. Training utilizes solely source-domain GT-maps, excluding any target-domain prior knowledge, data or annotations. Minimal augmentations (quantization, blurring, and canny edge), were used to introduce slight perturbations to create the triplets. We use SDv2.1 [42] for generating exemplar images and computing cross-attention maps [1, 3]. To ensure a fair" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "comparison, we adopt a ResNet-18 backbone throughout all the experiments. We use the Adam optimizer with an initial learning rate of " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " and a warm-up schedule over the first 1000 steps, after which the rate remains constant. Margin value " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " is set to 1.0. The batch size is set to 32. We empirically set " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\delta = 0.1" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " and cap test-time correction at " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "T = 10" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " iterations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 168, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 360 + ], + "type": "text", + "content": "Compared Approaches: We compare our method with ERM [17] baseline and existing approaches that utilize augmentation-based techniques (NJPP [59], AugMix [19], MixStyle [63], CutMix [60], RandAugment [7]), self-supervised and domain adaptation methods (RSC [22], pAdaIn [36], L2D [55], RSC+ASR [12]), uncertainty modeling (DSU [29], DSU-MAD [41]), attention and meta-learning methods (ACVC [8], P-RC [6], Meta-Casual [4]), and prompt-based learning (PromptD [28]). The number of methods evaluated differs by dataset. The discrepancy stems from PACS being the dominant benchmark in prior SSDG works, resulting in a larger number of methods evaluated on it. For VLCS, OfficeHome, and DomainNet, we rely on reported results from respective papers, or compute them ourselves (where code was available) to ensure a fair comparison." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 368, + 399, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 399, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 399, + 380 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 387, + 553, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 553, + 471 + ], + "type": "text", + "content": "Table 1 provide a comparison of average classification accuracy for our approach against existing methods across the PACS, VLCS, OfficeHome, and DomainNet datasets. Each column in these tables represents a source domain used for training, with the numerical values indicating the average accuracy on the three target domains. The last column presents the average of the four columns." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "text", + "content": "TIDE decisively outperforms existing approaches across all datasets. It achieves average accuracy gains of " + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "inline_equation", + "content": "8.33\\%" + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "inline_equation", + "content": "13.37\\%" + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "inline_equation", + "content": "16.16\\%" + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "inline_equation", + "content": "8.84\\%" + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "text", + "content": " over the second-best approach on PACS, VLCS, OfficeHome, and DomainNet, respectively. Table 1b details the performance on VLCS, where domain shifts primarily stem from background and viewpoint variations, with scenes spanning urban to rural and favoring non-standard viewpoints. Data augmentation based methods, which focus on style variation, yield limited gains on VLCS relative to PACS. Nonetheless, TIDE secures substantial improvements, notably achieving a " + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "inline_equation", + "content": "25.34\\%" + }, + { + "bbox": [ + 313, + 472, + 553, + 662 + ], + "type": "text", + "content": " gain over DSU-MAD when Caltech101 (C) serves as the source domain. Another noteworthy observation is that TIDE maintains strong performance across varying class counts, from VLCS's 5 to OfficeHome's 65, underscoring the scalability of the proposed approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 670, + 382, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 382, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 382, + 682 + ], + "type": "text", + "content": "4.2. Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "Components of TIDE: Ablation experiments on the PACS dataset are conducted to assess the contribution of each loss" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30215" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 71, + 289, + 215 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 289, + 215 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 289, + 215 + ], + "type": "table", + "html": "
Method VenueArtCartoonSketchPhotoAverage
ERM65.3864.2034.1533.6549.35
Augmix [19] ICLR'2166.5470.1652.4838.3057.12
RSC [22] ECCV'2073.4075.9056.2041.6061.03
Mixstyle [63] ICLR'2167.6070.3834.5737.4452.00
pAdalin [36] CVPR'2164.9665.2432.0433.6649.72
RSC+ASR [12] CVPR'2176.7079.3061.6054.6068.30
L2D [55] ICCV'2176.9177.8853.6652.2965.93
DSU [29] ICLR'2271.5474.5147.7542.1058.73
ACVC [8] CVPR'2273.6877.3955.3048.0563.10
DSU-MAD [41] CVPR'2372.4174.4749.6044.1560.66
P-RC [6] CVPR'2376.9878.5462.8957.1168.88
Meta-Casual [4] CVPR'2377.1380.1462.5559.6069.86
ABA [5] ICCV'2375.6977.3654.1259.0466.30
PromptD [28] CVPR'2478.7782.6962.9460.0971.87
TIDE86.2486.3773.1174.3680.02
", + "image_path": "70f773775a09014de7ca7460dae570c44e486fe41818dd64492095b06357c354.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 232, + 284, + 321 + ], + "blocks": [ + { + "bbox": [ + 160, + 217, + 191, + 226 + ], + "lines": [ + { + "bbox": [ + 160, + 217, + 191, + 226 + ], + "spans": [ + { + "bbox": [ + 160, + 217, + 191, + 226 + ], + "type": "text", + "content": "(a) PACS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 232, + 284, + 321 + ], + "lines": [ + { + "bbox": [ + 67, + 232, + 284, + 321 + ], + "spans": [ + { + "bbox": [ + 67, + 232, + 284, + 321 + ], + "type": "table", + "html": "
MethodVLCSAverage
ERM76.7258.8644.9557.7159.06
Augmix [19] ICLR'1975.2559.5245.9057.4359.03
pAdalin [36] CVPR'2176.0365.2143.1757.9460.34
Mixstyle [63] ICLR'2175.7361.2944.6656.5759.06
ACVC [8] CVPR'2276.1561.2347.4360.1861.75
DSU [29] ICLR'2276.9369.2046.5458.3662.11
DSU-MAD [41] CVPR'2376.9970.8544.7862.2363.71
TIDE82.6286.1370.1272.4477.08
", + "image_path": "dc38cae5c7522c4ca67180491611d7c388fd255d352ee869f59db766707eebab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 337, + 293, + 467 + ], + "blocks": [ + { + "bbox": [ + 159, + 323, + 192, + 332 + ], + "lines": [ + { + "bbox": [ + 159, + 323, + 192, + 332 + ], + "spans": [ + { + "bbox": [ + 159, + 323, + 192, + 332 + ], + "type": "text", + "content": "(b) VLCS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 337, + 293, + 467 + ], + "lines": [ + { + "bbox": [ + 58, + 337, + 293, + 467 + ], + "spans": [ + { + "bbox": [ + 58, + 337, + 293, + 467 + ], + "type": "table", + "html": "
MethodArtClipartProductRealAverage
ERM57.4350.8348.958.6853.96
MixUp [21] ICLR'1850.4143.1941.2451.8946.93
CutMix [60] ICCV'1949.1746.1541.253.6447.04
Augmix [19] ICLR'1956.8654.1252.0260.1256.03
RandAugment [7] CVPRW'2058.0755.3252.0260.8256.56
CutOut [9] arXiv:170854.3650.7947.6858.2452.77
RSC [22] ECCV'2053.5148.9847.1658.352.73
MEADA [62] NIPS'2057.053.248.8159.2154.80
PixMix [20] CVPR'2253.7752.6848.9158.6853.51
L2D [55] ICCV'2152.7948.9747.7558.3151.71
ACVC [8] CVPR'2254.351.3247.6956.2552.89
NJPP [59] ICML'2460.7254.9552.4761.2657.85
TIDE72.3275.1375.2273.3774.01
", + "image_path": "d1129382925a184262ce403b7168c40a09e6226e8556bf3a78481bdd20afd982.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 483, + 296, + 589 + ], + "blocks": [ + { + "bbox": [ + 149, + 468, + 202, + 478 + ], + "lines": [ + { + "bbox": [ + 149, + 468, + 202, + 478 + ], + "spans": [ + { + "bbox": [ + 149, + 468, + 202, + 478 + ], + "type": "text", + "content": "(c) OfficeHome" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 483, + 296, + 589 + ], + "lines": [ + { + "bbox": [ + 56, + 483, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 296, + 589 + ], + "type": "table", + "html": "
MethodClipartPaintingRealSketchAverage
ERM68.7366.1268.5169.4468.2
MixUp [21] ICLR'1870.3164.3469.2168.8268.17
CutMix [60] ICCV'1971.5263.8467.1369.4167.98
Augmix [19] ICLR'1972.3762.9169.8471.2269.09
RandAugment [7] CVPRW'2069.7165.5168.3666.9367.63
CutOut [9] arXiv:170870.8664.4869.9271.5569.20
RSC [22] ECCV'2068.2567.9170.7666.1868.28
PixMix [20] CVPR'2272.1263.5171.3465.4668.12
NJPP [59] ICML'2476.1469.2476.6171.2173.3
TIDE82.4280.3784.1581.6182.14
", + "image_path": "50a487e0e0675ddab2705667d1b39c1a76b7ee11a926cc9185d3f3fb8085e069.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "term on the overall performance. In Table 2, we present classification accuracy by incrementally adding components to the training pipeline, demonstrating the contribu" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 551, + 131 + ], + "blocks": [ + { + "bbox": [ + 150, + 590, + 201, + 600 + ], + "lines": [ + { + "bbox": [ + 150, + 590, + 201, + 600 + ], + "spans": [ + { + "bbox": [ + 150, + 590, + 201, + 600 + ], + "type": "text", + "content": "(d) DomainNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 609, + 295, + 654 + ], + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 654 + ], + "type": "text", + "content": "Table 1. SSDG classification accuracies " + }, + { + "bbox": [ + 55, + 609, + 295, + 654 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 609, + 295, + 654 + ], + "type": "text", + "content": " on PACS, VLCS, OfficeHome and DomainNet datasets, with ResNet-18 as backbone. Each column title indicates the source domain, and the numerical values represent the average performance in the target domains." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 71, + 551, + 131 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 551, + 131 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 551, + 131 + ], + "type": "table", + "html": "
MethodArtCartoonSketchPhotoAverage
Lc65.3864.2034.1533.6549.35
Lc + Lk65.4764.1235.4433.8149.71
Lc + Lk + LCSA66.1164.2335.1434.2749.93
Lc + Lk + LCSA + LLC80.2882.9166.1265.3773.67
+ test-time correction86.2486.3773.1174.3680.02
", + "image_path": "aba6a76f1b428820aa02fa98d3b3301c3ab352803dc6442aef586924e2d1885d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 133, + 504, + 144 + ], + "lines": [ + { + "bbox": [ + 364, + 133, + 504, + 144 + ], + "spans": [ + { + "bbox": [ + 364, + 133, + 504, + 144 + ], + "type": "text", + "content": "Table 2. Ablation study " + }, + { + "bbox": [ + 364, + 133, + 504, + 144 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 364, + 133, + 504, + 144 + ], + "type": "text", + "content": " on PACS." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 155, + 553, + 324 + ], + "blocks": [ + { + "bbox": [ + 318, + 155, + 553, + 324 + ], + "lines": [ + { + "bbox": [ + 318, + 155, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 318, + 155, + 553, + 324 + ], + "type": "image", + "image_path": "699c6c35be7a518f1f679236ea394c8cf9de202589adb6523fefe6cad8ba8829.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "text", + "content": "Figure 5. t-SNE visualizations to demonstrate impact of " + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LCC}" + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "text", + "content": ". Each column represents a test domain (Sketch, Cartoon, Painting), with the top row showing t-SNE plots without " + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LCC}" + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "text", + "content": " applied and the bottom one with it. Please zoom in for optimal viewing." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 402, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 570 + ], + "type": "text", + "content": "tion of each element to the model's performance. We observe that the introduction of the concept classification loss " + }, + { + "bbox": [ + 313, + 402, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 313, + 402, + 555, + 570 + ], + "type": "text", + "content": " and concept saliency alignment loss " + }, + { + "bbox": [ + 313, + 402, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CSA}" + }, + { + "bbox": [ + 313, + 402, + 555, + 570 + ], + "type": "text", + "content": " does not significantly affect the classifier's accuracy on the test domain. However, these components are integral to our approach, enabling test-time correction and, in turn, enhancing both performance and model interpretability. We observe that the introduction of the " + }, + { + "bbox": [ + 313, + 402, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LCC}" + }, + { + "bbox": [ + 313, + 402, + 555, + 570 + ], + "type": "text", + "content": " loss leads to a substantial increase in test domain accuracy, clearly demonstrating its effectiveness in fostering domain invariance. Finally, we compare results before and after test-time correction, highlighting that even prior to correction, our method achieves state-of-the-art performance across all source domains on PACS, underscoring the strength of our approach." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Error analysis of test-time correction: For this analysis, we train TIDE on the photo domain of PACS and test it on the sketch domain. The model gives an initial accuracy of " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "74.79\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ", which improves to " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "82.29\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " post test-time correction. In " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "72.2\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " of test samples, TIDE does not invoke test-time correction, with class predictions correct in " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "93.8\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " of these cases, demonstrating substantial reliability. The supplement provides examples of cases where the model misclassifications are not picked-up in the signature matching step of test-time correction. In the remaining " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "27.8\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " of cases where correction is initiated, " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "52.5\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " successfully converge to the correct classification, significantly contributing" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30216" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 72, + 199, + 137 + ], + "blocks": [ + { + "bbox": [ + 96, + 72, + 199, + 137 + ], + "lines": [ + { + "bbox": [ + 96, + 72, + 199, + 137 + ], + "spans": [ + { + "bbox": [ + 96, + 72, + 199, + 137 + ], + "type": "image", + "image_path": "beb08761c29002c92cfe81769a494f16d1a29a5c849c806991c0f9e7b5e0d0ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 202, + 72, + 304, + 136 + ], + "blocks": [ + { + "bbox": [ + 202, + 72, + 304, + 136 + ], + "lines": [ + { + "bbox": [ + 202, + 72, + 304, + 136 + ], + "spans": [ + { + "bbox": [ + 202, + 72, + 304, + 136 + ], + "type": "image", + "image_path": "613dd7df96c68fcd30a263262e6d4f50a52bb68486e394a4c858294a92297ac6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 308, + 72, + 410, + 136 + ], + "blocks": [ + { + "bbox": [ + 308, + 72, + 410, + 136 + ], + "lines": [ + { + "bbox": [ + 308, + 72, + 410, + 136 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 410, + 136 + ], + "type": "image", + "image_path": "9f4944b2c14e8b7d3f2cbed3b0366ab4b9cacb1de3ef1aafdaf31225282b744e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 416, + 72, + 518, + 136 + ], + "blocks": [ + { + "bbox": [ + 416, + 72, + 518, + 136 + ], + "lines": [ + { + "bbox": [ + 416, + 72, + 518, + 136 + ], + "spans": [ + { + "bbox": [ + 416, + 72, + 518, + 136 + ], + "type": "image", + "image_path": "f4c57f73735f26000c245e6994b65cd667fa41c932c183826dd4f89808fe1caf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 96, + 138, + 199, + 195 + ], + "blocks": [ + { + "bbox": [ + 96, + 138, + 199, + 195 + ], + "lines": [ + { + "bbox": [ + 96, + 138, + 199, + 195 + ], + "spans": [ + { + "bbox": [ + 96, + 138, + 199, + 195 + ], + "type": "image", + "image_path": "808f78ef24d3036fe78e68aad451302f96ebcec21b4656b09c56456f7e7aed9c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "type": "text", + "content": "Figure 6. Illustrative examples of concept level GradCAM maps corresponding to TIDE's predictions, across the four studied datasets. The concept names are displayed beneath the maps, with the target domain and predicted class indicated on the left. More results in supplement." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 203, + 138, + 304, + 195 + ], + "blocks": [ + { + "bbox": [ + 203, + 138, + 304, + 195 + ], + "lines": [ + { + "bbox": [ + 203, + 138, + 304, + 195 + ], + "spans": [ + { + "bbox": [ + 203, + 138, + 304, + 195 + ], + "type": "image", + "image_path": "0db9325e759eba3d4f641f5b03044feae90ae56d5b10cebdd0aacac945a4edc7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 137, + 410, + 195 + ], + "blocks": [ + { + "bbox": [ + 307, + 137, + 410, + 195 + ], + "lines": [ + { + "bbox": [ + 307, + 137, + 410, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 410, + 195 + ], + "type": "image", + "image_path": "045e375c81bf497dad55db2b8779e6d65bdbf75438dc3c6a669be30b8eda9de1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 416, + 137, + 517, + 195 + ], + "blocks": [ + { + "bbox": [ + 416, + 137, + 517, + 195 + ], + "lines": [ + { + "bbox": [ + 416, + 137, + 517, + 195 + ], + "spans": [ + { + "bbox": [ + 416, + 137, + 517, + 195 + ], + "type": "image", + "image_path": "5b6e65d78a9c46f055e0f6746510e9efe88099346f2f130dc2ae4f02d894a600.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 75, + 250, + 278, + 384 + ], + "blocks": [ + { + "bbox": [ + 75, + 250, + 278, + 384 + ], + "lines": [ + { + "bbox": [ + 75, + 250, + 278, + 384 + ], + "spans": [ + { + "bbox": [ + 75, + 250, + 278, + 384 + ], + "type": "image", + "image_path": "70d048ca715ecbc489e4a59851919a23d6e45cd3b26264d1fedc0ff1b06b6dcb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 396, + 295, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 429 + ], + "type": "text", + "content": "Figure 7. The top row shows initial class predictions and GradCAM maps for the concepts, while the bottom row presents the results after test-time correction." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 449, + 249, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 249, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 249, + 460 + ], + "type": "text", + "content": "to the final post-correction accuracy of " + }, + { + "bbox": [ + 55, + 449, + 249, + 460 + ], + "type": "inline_equation", + "content": "82.29\\%" + }, + { + "bbox": [ + 55, + 449, + 249, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "text", + "content": "t-SNE plots: To demonstrate the impact of " + }, + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LCC}" + }, + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "text", + "content": ", we present t-SNE visualizations in Figure 5, showing the distribution of concept-specific vectors (as computed in Equation 4). We individually plot the source domain (photo) along with sketch, cartoon, and painting domains as targets. Each case includes two plots: one without " + }, + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LCC}" + }, + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "text", + "content": " (top row) and one with it (bottom row). We observe that with " + }, + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LCC}" + }, + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "text", + "content": ", concept samples (e.g., mouth) align closely across domains, while distinct separation occurs between different concepts (e.g., mouth and horns). Without " + }, + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LCC}" + }, + { + "bbox": [ + 55, + 460, + 296, + 617 + ], + "type": "text", + "content": ", cluster separation and alignment across domains are weaker, highlighting its role in improving intra-concept compactness and inter-concept separability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 623, + 167, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 167, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 167, + 635 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Concept Localization: We train the SSDG model on photo domain for PACS, OfficeHome, and DomainNet and on Caltech101 for VLCS. The predicted class, concepts and corresponding saliency maps are shown in Figure 6. For each target class (e.g., mushroom, kettle, guitar), the model reliably highlights key concept-specific regions (e.g.," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 250, + 555, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 345 + ], + "type": "text", + "content": "stem, handle, strings) essential for classification. The model effectively isolates key features across diverse contexts, such as the legs and seat of a chair in complex zoomed-out scenes, the beak and feathers of camouflaged birds, the eyes and whiskers in deformed sketches of cats, and varying feature sizes, with the wick occupying a small area and the mushroom cap spanning a larger one." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 346, + 556, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 556, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 556, + 479 + ], + "type": "text", + "content": "Test-time Correction: The Figure 7 visually illustrates the efficacy of TIDE's test time correction abilities, by comparing the initial and corrected results. The model initially mis-classifies the images as dog and elephant, resulting in poorly aligned concept-level saliency maps for corresponding concepts i.e. ears, snout, tusks and trunk. TIDE detects and rectifies such errors during inference, leading to accurate classification accompanied by precise concept localization. This qualitative evaluation reinforces the robustness of our approach in refining predictions and generating reliable, class-specific concept maps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 490, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 388, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "In this work, we considered the problem of single-source domain generalization and observed that the current state-of-the-art methods fail in cases of semantic domain shifts primarily due to the global nature of their learned features. To alleviate this issue, we proposed TIDE, a new approach that not only learns local concept representations but also produces localization maps for these concepts. With these maps, we showed we could visually interpret model decisions while also enabling correction of these decisions at test time using our iterative attention refinement strategy. Extensive experimentation on standard benchmarks demonstrated substantial and consistent performance gains over the current state-of-the-art. Future work will rigorously explore methods to generate confidence scores grounded in TIDE's concept verification strategy and would explore the application of TIDE on other (e.g. fine-grained) classification datasets." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30217" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In International Conference on Computer Vision (ICCV), 2023. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 178 + ], + "type": "text", + "content": "[2] Tom B Brown. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 180, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 223 + ], + "type": "text", + "content": "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 224, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 295, + 267 + ], + "type": "text", + "content": "[4] Jin Chen, Zhi Gao, Xinxiao Wu, and Jiebo Luo. Meta-causal learning for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 269, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 269, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 295, + 312 + ], + "type": "text", + "content": "[5] Sheng Cheng, Tejas Gokhale, and Yezhou Yang. Adversarial bayesian augmentation for single-source domain generalization. In International Conference on Computer Vision (ICCV), 2023. 1, 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 314, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 314, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 61, + 314, + 296, + 367 + ], + "type": "text", + "content": "[6] Seokeon Choi, Debasmit Das, Sungha Choi, Seunghan Yang, Hyunsin Park, and Sungrack Yun. Progressive random convolutions for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 369, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 369, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 61, + 369, + 296, + 423 + ], + "type": "text", + "content": "[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 424, + 296, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 424, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 61, + 424, + 296, + 478 + ], + "type": "text", + "content": "[8] Ilke Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 479, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 479, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 61, + 479, + 296, + 512 + ], + "type": "text", + "content": "[9] Terrance Devries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. ArXiv, abs/1708.04552, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 567 + ], + "type": "text", + "content": "[10] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "text", + "content": "[11] Antonio D'Innocente and Barbara Caputo. Domain generalization with domain-specific aggregation modules. In The German Conference on Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "text", + "content": "[12] Xinjie Fan, Qifei Wang, Junjie Ke, Feng Yang, Boqing Gong, and Mingyuan Zhou. Adversarily adaptive normalization for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "text", + "content": "[13] Chen Fang, Ye Xu, and Daniel N Rockmore. Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In Conference on Computer Vision and Pattern Recognition (CVPR), 2013. 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 116 + ], + "type": "text", + "content": "Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations (ICLR), 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[15] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In International Conference on Computer Vision (ICCV), 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 218 + ], + "type": "text", + "content": "[16] Tejas Gokhale, Rushil Anirudh, Bhavya Kailkhura, Jayaraman J Thiagarajan, Chitta Baral, and Yezhou Yang. Attribute-guided adversarial training for robustness to natural perturbations. In Association for the Advancement of Artificial Intelligence (AAAI), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 252 + ], + "type": "text", + "content": "[17] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. International Conference on Learning Representations (ICLR), 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 554, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 295 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 298, + 554, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 353 + ], + "type": "text", + "content": "[19] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. International Conference on Learning Representations (ICLR), 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "text", + "content": "[20] Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. Pixmix: Dreamlike pictures comprehensively improve safety measures. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 453 + ], + "type": "text", + "content": "[21] Yann N. Dauphin David Lopez-Paz Hongyi Zhang, Moustapha Cisse. mixup: Beyond empirical risk minimization. International Conference on Learning Representations (ICLR), 2018. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 498 + ], + "type": "text", + "content": "[22] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 500, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 533 + ], + "type": "text", + "content": "[23] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in Neural Information Processing Systems (NeurIPS), 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "type": "text", + "content": "[24] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International Conference on Machine Learning (ICML), 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 580, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 554, + 622 + ], + "type": "text", + "content": "[25] Wouter M Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 43(3):766-785, 2019. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2012. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In International Conference on Computer Vision (ICCV), 2017. 4, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30218" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[28] Deng Li, Aiming Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "text", + "content": "[29] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. International Conference on Learning Representations (ICLR), 2022. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[30] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In European Conference on Computer Vision (ECCV), 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "text", + "content": "[31] Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. ArXiv, abs/2108.13624, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[32] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning (ICML), 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "text", + "content": "[33] Andrei Margeloiu, Matthew Ashman, Umang Bhatt, Yanzhi Chen, Mateja Jamnik, and Adrian Weller. Do concept bottleneck models learn as intended? International Conference on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[34] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "text", + "content": "[35] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning (ICML), 2013. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "type": "text", + "content": "[36] Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[37] Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. International Conference on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "text", + "content": "[38] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In International Conference on Computer Vision (ICCV), 2019. 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[39] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation. In European Conference on Computer Vision (ECCV), 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 294, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 625 + ], + "type": "text", + "content": "[40] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 626, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 626, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 626, + 294, + 669 + ], + "type": "text", + "content": "[41] Sanqing Qu, Yingwei Pan, Guang Chen, Ting Yao, Changjun Jiang, and Tao Mei. Modality-agnostic debiasing for single domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In International Conference on Computer Vision (ICCV), 2017. 1, 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "text", + "content": "[44] Sarath Sivaprasad, Akshay Goindani, Mario Fritz, and Vineet Gandhi. Class-wise domain generalization: A novel framework for evaluating distributional shift. In NeurIPS Workshop on Distribution Shifts: Connecting Methods and Applications, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "text", + "content": "[45] Nathan Somavarapu, Chih-Yao Ma, and Zsolt Kira. Frustratingly simple domain generalization via image stylization. ArXiv, abs/2006.11207, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[46] Shuhan Tan, Xingchao Peng, and Kate Saenko. Class-imbalanced domain adaptation: An empirical odyssey. In European Conference on Computer Vision Workshops (ECCVW), 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[47] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 338 + ], + "type": "text", + "content": "[48] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 339, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 381 + ], + "type": "text", + "content": "[49] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[50] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 427, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 554, + 480 + ], + "type": "text", + "content": "[51] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in Neural Information Processing Systems (NeurIPS), 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 482, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 554, + 525 + ], + "type": "text", + "content": "[52] Bor-Shiun Wang, Chien-Yi Wang, and Wei-Chen Chiu. Mcpnet: An interpretable classifier via multi-level concept prototypes. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 526, + 554, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 569 + ], + "type": "text", + "content": "[53] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. International Conference on Learning Representations (ICLR), 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "type": "text", + "content": "[54] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "type": "text", + "content": "[55] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In International Conference on Computer Vision (ICCV), 2021. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[56] Zehao Xiao, Jiayi Shen, Xiantong Zhen, Ling Shao, and Cees Snoek. A bit more bayesian: Domain-invariant learning with uncertainty. In International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30219" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 464 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[57] Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain expansion approach for single-source domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[58] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[59] Jianhao Yuan, Francesco Pinto, Adam Davies, and Philip Torr. Not just pretty pictures: Toward interventional data augmentation using text-to-image generators. International Conference on Machine Learning (ICML), 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 284 + ], + "type": "text", + "content": "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision (ICCV), 2019. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "text", + "content": "[61] Hanlin Zhang, Yi-Fan Zhang, Weiyang Liu, Adrian Weller, Bernhard Schölkopf, and Eric P Xing. Towards principled disentanglement for domain generalization. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 385 + ], + "type": "text", + "content": "[62] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 419 + ], + "type": "text", + "content": "[63] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. International Conference on Learning Representations (ICLR), 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "type": "text", + "content": "[64] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation. International Journal of Computer Vision (IJCV), 132(3):822-836, 2024. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30220" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_content_list.json b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac7ec9527d66336984e4dd66a0f3f1bd13ae6bba --- /dev/null +++ b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation", + "text_level": 1, + "bbox": [ + 233, + 130, + 767, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yabiao Wang $^{1,2*}$ ; Shuo Wang $^{2*}$ ; Jiangning Zhang $^{2}$ , Ke Fan $^{3}$ ,", + "bbox": [ + 243, + 202, + 751, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiafu Wu², Zhucun Xue¹, Yong Lii¹", + "bbox": [ + 346, + 220, + 669, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University $^{2}$ Youtu Lab, Tencent $^{3}$ Shanghai Jiao Tong University", + "bbox": [ + 191, + 238, + 795, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://aigc-explorer.github.io/TIMotion-page/", + "bbox": [ + 359, + 258, + 635, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human-human motion generation is essential for understanding humans as social beings. Current methods fall into two main categories: single-person-based methods and separate modeling-based methods. To delve into this field, we abstract the overall generation process into a general framework MetaMotion, which consists of two phases: temporal modeling and interaction mixing. For temporal modeling, the single-person-based methods concatenate two people into a single one directly, while the separate modeling-based methods skip the modeling of interaction sequences. The inadequate modeling described above resulted in sub-optimal performance and redundant model parameters. In this paper, we introduce TIMotion (Temporal and Interactive Modeling), an efficient and effective framework for human-human motion generation. Specifically, we first propose Causal Interactive Injection to model two separate sequences as a causal sequence leveraging the temporal and causal properties. Then we present Role-Evolving Scanning to adjust to the change in the active and passive roles throughout the interaction. Finally, to generate smoother and more rational motion, we design Localized Pattern Amplification to capture short-term motion patterns. Extensive experiments on InterHuman and InterX demonstrate that our method achieves superior performance.", + "bbox": [ + 88, + 340, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 748, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of generative computer vision, human motion generation has significant implications for computer animation [18, 20], game development [3, 33], and robotic control [26, 36, 37]. In recent years, there have been remarkable advancements in human motion generation, driven by various user-specified conditions such as action cat", + "bbox": [ + 89, + 773, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "egories [9, 22], speeches [2, 12], and natural language prompts [11, 24]. Among these, many approaches leveraging large language models [1] and diffusion models [43] have yielded impressive results in generating realistic and diverse motions, benefiting from their powerful modeling capabilities. Despite this progress, most existing methods are designed primarily for single-person scenarios, thereby neglecting a crucial element of human motion: the complex and dynamic interactions between individuals. We address the challenge of generating two-person motion by fully utilizing the temporal and interactive dynamics between the individuals.", + "bbox": [ + 511, + 310, + 905, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To better explore human-human motion generation, we abstract a general framework MetaMotion, as shown on the left of Fig. 1, which consists of two phases: temporal modeling and interaction mixing. Previous approaches have prioritized interaction mixing over temporal modeling and can be divided into two primary categories: single-person-based methods and separate modeling-based methods. As shown in Fig. 1(a), the single-person-based methods (e.g. MDM [32]) concatenate two individuals into one, which is then fed into the existing single-person motion generation module, e.g. DiT. And the separate modeling-based methods (e.g. InterGen [16]), as shown in Fig. 1(b), model two individuals individually and then extract motion information from both themselves and each other, using self-attention and cross-attention mechanisms, respectively. Following the general logic of MetaMotion, we introduce the Temporal and Interactive Framework, as depicted in Fig. 1(c), which models the human-human causal interactions. In fact, this effective temporal modeling method can simplify the design of the interaction mixing module and reduce the number of learnable parameters.", + "bbox": [ + 511, + 492, + 905, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we design an effective temporal modeling method and introduce TIMotion, a temporal and interactive framework for efficient human-human motion generation. Specifically, we first propose Causal Interactive Injection to model the two single-person motion sequences as a causal interaction sequence, according to the temporal causal prop", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions.", + "bbox": [ + 107, + 875, + 223, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7169", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65d1e8c2bf254674ac4334cb4e4fde95ace27ca0138c7dafb08f435d1208c839.jpg", + "image_caption": [ + "Figure 1. MetaMotion and performance of MetaMotion-based models on InterHuman validation set. We abstract the MetaMotion concept that illustrates the intrinsic properties of human-human motion generation in the interaction process. (a) and (b) show the two types of methods currently, and (c) shows our method TIMotion, LPA refers to the Localized Pattern Amplification. In (d) we compare the performance of the different methods on the InterHuman dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 225, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3edd481d018504674e702784255bbbc37ba3c4f3189d61dc25f5263210cbdc10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 90, + 359, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bfa135e68db0c805d3881674d479a4765e50034e4134544eae71aacc54e32b93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 90, + 506, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ddcd6e045956c034cc60cd9cfff99c25241ed92e6ff86d54bace07c42c929f1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 90, + 660, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/17f2bbbbf96e6814896f9996f6cdc15d7c91ba2a66c2180533536671203201a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 97, + 893, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erties of motion sequences. Second, since active and passive roles are not static in interactions, we propose Role-Evolving Mixing to make two humans act in both active and passive roles. Then the network can dynamically adjust the roles of the two humans based on the text's semantics and the motion's context. Finally, we propose Localized Pattern Amplification, which captures short-term motion patterns for each individual separately, resulting in smoother and more logical motion generation. Our proposed framework can be well adapted to different interaction-mixing modules, including Transformer, RWKV, and Mamba.", + "bbox": [ + 88, + 340, + 480, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 508, + 434, + 522 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conceptualize human-human motion generation within a general framework MetaMotion and design an innovative method. Our proposed framework, TIMotion, is versatile enough to integrate with various interaction-mixing modules (e.g. Transformer, RWKV, Mamba) and reduces the number of parameters of these modules.", + "- To utilize the temporal and causal properties, we propose Causal Interactive Injection to model two separate motion sequences as a unified causal sequence. In addition, we introduce Role-Evolving Scanning to accommodate shifts between the active and passive roles during interactions. Moreover, we also design Localized Pattern Amplification to capture short-term motion patterns effectively.", + "- We perform extensive experiments on the benchmark human-human motion generation datasets: InterHuman and Inter-X. The results demonstrate the effectiveness and generalizability of our proposed methods." + ], + "bbox": [ + 89, + 527, + 482, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 799, + 232, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-Person Human Motion Generation. Creating human motion is vital for applications such as 3D modeling and robot manipulation. The primary approach, known as the Text-to-Motion task, involves learning a unified latent space for both language and motion.", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoencoders have been widely adopted in human motion generation. MotionCLIP [30] effectively integrates semantic knowledge from CLIP [25] into the human motion manifold. TEMOS [23] and T2M [10] combine a Transformer-based VAE with a text encoder to generate motion sequences based on text descriptions. AttT2M [44] and TM2D [7] integrate a spatial-temporal body-part encoder into VQ-VAE [34] to improve the learning of discrete latent space. T2M-GPT [40] redefines text-driven motion generation as a next-index prediction task and proposes a framework based on VQ-VAE and Generative Pretrained Transformer (GPT) for motion generation.", + "bbox": [ + 511, + 340, + 903, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, diffusion-based generative modeling has been gaining attention. MotionDiffuse [42] exhibits several desirable properties, such as probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [31] aims to predict motion directly and incorporates a geometric loss to boost the model's performance. Instead of employing a diffusion model to link raw motion sequences with conditional inputs, MLD [5] further utilizes the latent diffusion model to reduce the training and inference costs substantially. ReMoDiffuse [41] introduces an improvement mechanism based on dataset retrieval to enhance the denoising process of Diffusion.", + "bbox": [ + 511, + 523, + 905, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Person Human Motion Generation. As multiperson motion synthesis involves the interactive dynamics between multiple individuals, it is more challenging than single-person motion generation. Early work typically relied on motion graphs and momentum-based inverse kinematics to model human joints. Recently, ComMDM [27] finetunes a pre-trained text-to-motion diffusion model using a small scale of two human motions. RIG [29] converts the text of asymmetric interactions into both active and passive voice to maintain consistent textual context for each individual. To complete the interaction process by", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7170", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ed7e7f476381049f9a6319e9729cb8b989ff3e56c379faa7a78c6932a30a3d6f.jpg", + "image_caption": [ + "Figure 2. The overall framework of our TIMotion. We contribute three primary technical designs. First, we propose Causal Interactive Injection to utilize the temporal properties of motion sequences. Then we present Role-Evolving Mixing to adjust to the ever-evolving roles during interaction. Finally, we design Localized Pattern Amplification to capture short-term motion patterns." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 911, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "manipulating the joints of two individuals to specific positions, InterControl [38] employs a Large Language Model to generate movement plans for the two individuals by designing prompts. InterGen [16] introduces a large-scale, text-annotated two-person motion dataset. Building on this dataset, it proposes a diffusion model with shared weights and multiple regularization losses. FreeMotion [6] proposes to decouple the process of conditional motion generation and support the number-free motion synthesis.", + "bbox": [ + 88, + 392, + 483, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 559, + 181, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motion Diffusion Model", + "text_level": 1, + "bbox": [ + 89, + 589, + 312, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current methods usually use diffusion models [15] for motion generation. For the Interhuman dataset, a noncanonical representation [16] is typically used. The representation is formulated as: $x^{i} = [\\mathbf{j}_{g}^{p},\\mathbf{j}_{g}^{v},\\mathbf{j}^{r},\\mathbf{c}^{f}]$ , where the $i$ -th motion state $x^{i}$ is defined as a collection of global joint positions $\\mathbf{j}_g^p\\in \\mathbb{R}^{3N_j}$ , velocities $\\mathbf{j}_g^v\\in \\mathbb{R}^{3N_j}$ in the world frame, 6D representation of local rotations $\\mathbf{j}^r\\in \\mathbb{R}^{6N_j}$ in the root frame, and binary foot-ground contact $\\mathbf{c}^f\\in \\mathbb{R}^4$ . Our goal is to train a model parameterized by $\\theta$ to approximate the human interactive motion data distribution $p(\\mathbf{x_0})$ .", + "bbox": [ + 89, + 617, + 483, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Process. Following a specified schedule $\\beta_{t} \\in (0,1)$ , models incrementally degrade input data $\\mathbf{x}_0 \\sim p(\\mathbf{x}_0)$ , ultimately transforming the data distribution into an isotropic Gaussian over $T$ steps. Each diffusion transition can be considered as", + "bbox": [ + 89, + 777, + 483, + 853 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 872, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the full diffusion process can be written as", + "bbox": [ + 511, + 392, + 839, + 407 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {1 \\leq t \\leq T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 419, + 906, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising Process. In the denoising process, models are trained to reverse the diffusion procedure, enabling them to transform random noise into real data distribution during inference. The denoising process can be written as", + "bbox": [ + 511, + 464, + 906, + 523 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right) \\\\ = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon\\right), \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} \\beta_ {t}\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 536, + 905, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , $\\alpha_{t} = 1 - \\beta_{t}$ , $\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}$ and $\\theta$ denotes parameters of the network learning to denoise. The training objective is to maximize the likelihood of observed data $p_{\\theta}(\\mathbf{x}_0) = \\int p_{\\theta}(\\mathbf{x}_{0:T})d\\mathbf{x}_{1:T}$ , by maximizing its evidence lower bound (ELBO), which effectively aligns the true denoising model $q(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t)$ with the parameterized $p_{\\theta}(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t)$ . During training, the goal of the denoising network $\\epsilon_{\\theta}(.)$ is to reconstruct $\\mathbf{x}_0$ given any noised input $\\mathbf{x}_t$ , by predicting the added noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ via minimizing the noise prediction error", + "bbox": [ + 511, + 593, + 906, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\right\\| ^ {2} \\right]. (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 756, + 906, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To condition the model on additional context information $\\mathbf{c}$ , e.g., text, we inject $\\mathbf{c}$ into $\\epsilon_{\\theta}(.)$ by replacing $\\mu_{\\theta}(\\mathbf{x}_t,t)$ and $\\Sigma_{\\theta}(\\mathbf{x}_t,t)$ with $\\mu_{\\theta}(\\mathbf{x}_t,t,\\mathbf{c})$ and $\\Sigma_{\\theta}(\\mathbf{x}_t,t,\\mathbf{c})$ .", + "bbox": [ + 511, + 792, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. MetaMotion", + "text_level": 1, + "bbox": [ + 511, + 847, + 648, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present the core concept \"MetaMotion\" for human-human motion generation at first. As shown in Fig. 1,", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7171", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we abstract the process of human-human motion generation into two phases: temporal modeling and interaction mixing.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, two single-person sequences go through the temporal modeling module to get the input sequence $X$ . Then, the input sequence is fed to the interaction-mixing module and this process can be expressed as", + "bbox": [ + 89, + 121, + 483, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY = \\operatorname {I n t e r a c t i o n M i x i n g} (X), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 195, + 482, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where InteractionMixing is usually the transformer structure including self-attention and cross-attention. Notably, InteractionMixing can also be some emerging structures, e.g., Mamba [8], RWKV [21].", + "bbox": [ + 89, + 224, + 483, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. TIMotion", + "text_level": 1, + "bbox": [ + 89, + 295, + 202, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For human-human motion generation, most current methods are based on extensions of single-person motion generation. Some design-specific approach (e.g. InterGen [16]) considers the interaction between two persons in the interaction mixing module. However, current methods ignore the importance of temporal modeling, resulting in suboptimal performance in generating long-sequence motions and multi-person interactions.", + "bbox": [ + 89, + 318, + 483, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we propose TIMotion, as illustrated in Fig. 2, an effective temporal modeling approach that can be applied to different interaction-mixing structures, enabling better handling the human-human interactions. TIMotion consists of three key technical designs: (1) Causal Interactive Injection, which identifies the causal relationships between human-human interactions; (2) Role-Evolving Scanning, which adapts to shifts between active and passive roles during interactions; (3) Localized Pattern Amplification, designed to handle short-term sequences better.", + "bbox": [ + 89, + 439, + 483, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Causal Interactive Injection", + "text_level": 1, + "bbox": [ + 89, + 599, + 331, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perception of ego-motion and interaction between two persons are two important elements of human-human motion generation. Considering the causal properties of motion, we propose Causal Interactive Injection, a temporal modeling approach to simultaneously achieve both perceptions of ego-motion and interactions with each other.", + "bbox": [ + 89, + 619, + 483, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we denote two single-person motion sequences as $\\{x_{a}, x_{b}\\}$ , where $x_{a} = \\{x_{a}^{j}\\}_{j=1}^{L}$ and $x_{b} = \\{x_{b}^{j}\\}_{j=1}^{L}$ are respective sequences of motion, and $L$ is the length of the sequence. Since the motions of two individuals at the current time step are jointly determined by their motions at previous time steps, we model the two single-person motion sequences as a causal interaction sequence $x_{cii} = \\{x_{k}^{j/2}\\}_{j=1}^{2L}$ , the symbol // denotes division followed by rounding up and $k$ can be acquired as follows:", + "bbox": [ + 89, + 710, + 483, + 852 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk = \\left\\{ \\begin{array}{ll}a, & j\\% 2 = 1.\\\\ b, & j\\% 2 = 0. \\end{array} \\right. \\tag{6}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 863, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we can inject them into the interaction-mixing module and separate the motion embeddings of the two individuals from the outputs according to Eq. (6).", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/71ab8ac14325f66d143f13be212b975d25864f6c50b3729e69f7a1131078f7cb.jpg", + "image_caption": [ + "frame 1" + ], + "image_footnote": [], + "bbox": [ + 558, + 154, + 619, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/11df8a091534d89a05f9cb811d89f472c2bbac40da657fc1570fdb8a58579d31.jpg", + "image_caption": [ + "frame 2" + ], + "image_footnote": [], + "bbox": [ + 673, + 156, + 728, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9b954790ee49f4cf025bb6d156b0637f7cb4fe8b9632dd2b45299d4ba358eef3.jpg", + "image_caption": [ + "frame 3" + ], + "image_footnote": [], + "bbox": [ + 792, + 157, + 859, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17ae2a3f4f8a4aa1cca1d86633b47b7de07a43f1f2917c4a4fb1e34175ac2a3a.jpg", + "image_caption": [ + "frame 4" + ], + "image_footnote": [ + "The first person suddenly pushes the other one. The second person takes a step back, then forcefully pushes the first person. The first person was knocked down." + ], + "bbox": [ + 557, + 248, + 633, + 305 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/044b4c106d984b6aa885e10c9fbed0fbb8d7cac7983c0784599d5287a6a4e7d8.jpg", + "image_caption": [ + "frame 5" + ], + "image_footnote": [], + "bbox": [ + 676, + 244, + 728, + 304 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/00a1a7dfbfafc56057fad7854d5b9304b4e4c42f4c648a79c5fade3f01befab9.jpg", + "image_caption": [ + "frame 6", + "Figure 3. Illustration of changing active and passive roles. The first person acts as the active role in the early stages, and as time progresses, the other person becomes the active role of the motion." + ], + "image_footnote": [], + "bbox": [ + 785, + 244, + 869, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Role-Evolving Scanning", + "text_level": 1, + "bbox": [ + 511, + 441, + 728, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As noted in [4], human interactions inherently follow a certain order, e.g., \"shaking hands\" typically begins with one person extending a hand first, which means that interactive motion can be categorized as active or passive. Some methods [29] split the text description into active and passive voices. However, the \"active\" and \"passive\" roles are not static in a text description. As shown in Fig. 3, these roles constantly swap between characters as the interaction progresses. To minimize redundant text preprocessing and adapt to the ongoing shifts in roles, we design Role-Evolving Scanning, which is both efficient and effective.", + "bbox": [ + 511, + 460, + 906, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the causal interaction sequence $x$ defined in Causal Interactive Injection (Sec. 3.3.1), it is obvious that $x_{a}$ and $x_{b}$ represent the active sequence and passive sequence, respectively. However, the above assumption of active and passive motions is not always in line with the real order. To cope with the changing roles, we re-model the interactive motion sequences as a symmetric causal interaction sequence $x_{sym\\_ cii} = \\{x_{k'}^{j // 2}\\}_{j=1}^{2L}$ , where the symbol $//$ is defined as division followed by rounding up and $k'$ is obtained by exchanging $a$ and $b$ in Eq. (6).", + "bbox": [ + 511, + 628, + 905, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that the causal interaction sequence $x_{cii} \\in \\mathbb{R}^{2L \\times C}$ and the symmetric causal interaction sequence $x_{sym\\_ cii} \\in \\mathbb{R}^{2L \\times C}$ , where $L$ is the length of the sequence and $C$ is the dimension of motion embedding, we can acquire the final interaction sequence $X \\in \\mathbb{R}^{2L \\times 2C}$ through Role-Evolving Scanning as:", + "bbox": [ + 511, + 780, + 905, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX = \\operatorname {C o n c a t} \\left(x _ {c i i}, x _ {s y m. c i i}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 885, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7172", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then the sequence $X$ is fed to the interaction-mixing module. After obtaining the output $Y \\in \\mathbb{R}^{2L \\times 2C}$ through Eq. (5), we first divide it at the channel level into causal interaction embeddings and symmetric causal interaction embeddings. According to $x_{cii}$ and $x_{sym\\_ cii}$ , we can obtain the two individuals' splitting embeddings twice. Then we merge the embeddings of the two individuals to obtain the final global motion embeddings $y_a^g \\in \\mathbb{R}^{L \\times C}$ and $y_b^g \\in \\mathbb{R}^{L \\times C}$ . The overall process is as follows:", + "bbox": [ + 89, + 90, + 483, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} y _ {a 1}, y _ {b 1} = \\operatorname {S p l i t} (Y [ :, : C ]), \\\\ y _ {b 2}, y _ {a 2} = \\operatorname {S p l i t} (Y [:, C: ]), \\tag {8} \\\\ y _ {a} ^ {g} = y _ {a 1} \\oplus y _ {a 2}, \\\\ y _ {b} ^ {g} = y _ {b 1} \\oplus y _ {b 2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 243, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes element-wise sum.", + "bbox": [ + 109, + 333, + 352, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By utilizing Role-Evolving Mixing to make two humans act as both active and passive roles, the network can dynamically adjust the roles of the two humans based on the semantics of the text and the context of the motion.", + "bbox": [ + 89, + 349, + 483, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Localized Pattern Amplification", + "text_level": 1, + "bbox": [ + 89, + 424, + 359, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformers and RNNs excel at global modeling and capturing long-range dependencies but tend to overlook local semantic information [13]. Additionally, Causal Interactive Injection and Role-Evolving Mixing mainly model the overall motion based on the causality of the interactions, but neglect to focus on the localized motion patterns of the single person. To address this issue, we propose Localized Pattern Amplification, which summarizes short-term motion patterns for each person individually while generating smoother and more rational motion.", + "bbox": [ + 89, + 444, + 483, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we utilize 1-D convolution layers and the residual structure to realize Localized Pattern Amplification. Given that the condition embedding $e$ and two single-person motion sequences $x_{a}$ and $x_{b}$ , the local motion embedding for $x_{a}$ can be expressed as:", + "bbox": [ + 89, + 597, + 483, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {a} ^ {l} = \\operatorname {C o n v} _ {3} (\\operatorname {A d a L N} (x _ {a}, e)), \\\\ y _ {a} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {A d a L N} \\left(x _ {a} ^ {l}, e\\right)\\right), \\tag {9} \\\\ y _ {a} ^ {l} = x _ {a} + y _ {a}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 686, + 482, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv_k}$ denotes the 1-D convolution with kernel size $k$ and AdaLN is the adaptive layer normalization used in [42]. The process for $x_{b}$ is the same as for $x_{a}$ and they share network weights.", + "bbox": [ + 89, + 762, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once we obtain the outputs $\\{y_a^g, y_b^g\\}$ through Eq. (8) and the output $\\{y_a^l, y_b^l\\}$ from the convolution block, global embeddings and local embeddings are aggregated through concatenation along the channel dimension, followed by a linear layer to restore the original channels, resulting in the", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "final outputs $\\{y_a^{final}, y_b^{final}\\}$ :", + "bbox": [ + 513, + 88, + 715, + 108 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} y _ {a} ^ {f i n a l} = \\operatorname {L i n e a r} (\\operatorname {C o n c a t} (y _ {a} ^ {g}, y _ {a} ^ {l})), \\\\ y _ {b} ^ {\\text {f i n a l}} = \\operatorname {L i n e a r} \\left(\\operatorname {C o n c a t} \\left(y _ {b} ^ {g}, y _ {b} ^ {l}\\right)\\right). \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 114, + 903, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the final outputs are fed into the next encoder block or the final decoder layer.", + "bbox": [ + 511, + 162, + 906, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.4. Objective Function", + "text_level": 1, + "bbox": [ + 511, + 200, + 692, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the same loss functions as InterGen [16], including the diffusion loss $\\mathcal{L}_{simple}$ , the foot contact loss $\\mathcal{L}_{foot}$ , joint velocity loss $\\mathcal{L}_{vel}$ , the bone length loss $\\mathcal{L}_{BL}$ , the masked joint distance map loss $\\mathcal{L}_{DM}$ , and the relative orientation loss $\\mathcal{L}_{RO}$ . For more details about the losses, we refer readers to InterGen. Finally, the overall loss is defined as:", + "bbox": [ + 511, + 219, + 906, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {m o t i o n}} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {v e l} \\mathcal {L} _ {v e l} + \\lambda_ {f o o t} \\mathcal {L} _ {f o o t} \\\\ + \\lambda_ {B L} \\mathcal {L} _ {B L} + \\lambda_ {D M} \\mathcal {L} _ {D M} + \\lambda_ {R O} \\mathcal {L} _ {R O}, \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 319, + 903, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the hyper-parameters $\\lambda_{vel},\\lambda_{foot},\\lambda_{BL},\\lambda_{DM},\\lambda_{RO}$ are the same as InterGen.", + "bbox": [ + 511, + 363, + 905, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 405, + 645, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 430, + 702, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We assess our proposed framework using the InterHuman [16] and Inter-X [39] dataset. InterHuman is the first dataset to incorporate text annotations for two-person motions. This dataset includes 6,022 motions spanning various categories of human actions and is labeled with 16,756 unique descriptions made up of 5,656 distinct words. InterX is the largest human-human interaction dataset with diverse interaction patterns. It includes about 11k interaction sequences and more than 8.1M frames.", + "bbox": [ + 511, + 452, + 906, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We employ the same evaluation metrics as InterGen [16], which are as follows: (1) Frechet Inception Distance (FID). (2) $R$ -Precision. (3) Diversity. (4) Multimodality (MModality). (5) Multi-modal distance (MM Dist). Detailed explanations and calculations of the metrics can be found in the supplementary materials.", + "bbox": [ + 511, + 588, + 905, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1. Implementation Details.", + "text_level": 1, + "bbox": [ + 511, + 685, + 725, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a frozen CLIP-ViT-L/14 model as the text encoder. The dimension of motion embedding is set to 512. During training, the number of diffusion timesteps is set to 1000, and we employ the DDIM [28] sampling strategy with 50 timesteps and $\\eta = 0$ . The cosine noise level schedule [19] and classifier-free guidance [14] are adopted, with $10\\%$ of random CLIP embeddings set to zero during training and a guidance coefficient of 3.5 during sampling. All the models are trained using the AdamW [17] optimizer with betas of (0.9, 0.999), a weight decay of $2 \\times 10^{-5}$ , a maximum learning rate of $10^{-4}$ , and a cosine learning rate schedule with 10 linear warm-up epochs. To balance the contributions of different loss terms, we set $\\lambda_{vel} = 30$ , $\\lambda_{foot} = 30$ , $\\lambda_{BL} = 10$ ,", + "bbox": [ + 511, + 704, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7173", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a012227edeffb418684371cfc6da68b7c90eca3c93d1b44c39af8922c95d81cc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsR Precision↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.452±.0080.610±.0090.701±.0080.273±.0073.755±.0087.948±.064-
TEMOS [23]0.224±.0100.316±.0130.450±.01817.375±.0435.342±.0156.939±.0710.535±.014
T2M [10]0.238±.0120.325±.0100.464±.01413.769±.0724.731±.0137.046±.0221.387±.076
MDM [31]0.153±.0120.260±.0090.339±.0129.167±.0566.125±.0187.602±.0452.355±.080
ComMDM* [27]0.067±.0130.125±.0180.184±.01538.643±.09813.211±.0133.520±.0580.217±.018
ComMDM [27]0.223±.0090.334±.0080.466±.0107.069±.0545.212±.0217.244±.0381.822±.052
RIG [29]0.285±.0100.409±.0140.521±.0136.775±.0694.876±.0187.311±.0432.096±.065
InterGen [16]0.371±.0100.515±.0120.624±.0105.918±.0795.108±.0147.387±.0292.141±.063
TIMotion+Transformer(ours)0.491±.0050.648±.0040.724±.0045.433±.0803.775±.0018.032±.0300.952±.032
TIMotion+Mamba(ours)0.431±.0040.586±.0050.668±.0046.142±.0593.800±.0017.793±.0240.837±.036
TIMotion+RWKV(ours)0.501±.0050.656±.0060.734±.0064.702±.0693.769±.0017.943±.0341.005±.020
", + "bbox": [ + 96, + 88, + 903, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/61d4f1926ed2b0905ed701b435d42d30e704508d719705ca7e0b05727a9653ca.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation on the InterHuman [16] test set. We run all the evaluations 20 times. ± indicates a 95% confidence interval. Bold indicates the best result, while underline refers to the second best. ComMDM* indicates the ComMDM model fine-tuned in the original few-shot setting with 10 training samples and ComMDM indicates fine-tuned on the entire InterHuman training set." + ], + "table_footnote": [], + "table_body": "
Interaction MixingTemporal ModelingR Precision↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Transformer [35]single-person extension0.395±.0040.535±.0040.615±.0048.028±.0993.820±.0017.687±.0220.852±.024
separate modeling0.371±.0100.515±.0120.624±.0105.918±.0795.108±.0147.387±.0292.141±.063
TIMotion (Ours)0.491±.0050.648±.0040.724±.0045.433±.0803.775±.0018.032±.0300.952±.032
Mamba [8]single-person extension0.368±.0050.513±.0050.597±.0067.232±.0813.825±.0017.864±.0240.870±.020
separate modeling0.420±.0050.569±.0040.650±.0037.221±.0973.803±.0017.932±.0350.855±.020
TIMotion (Ours)0.431±.0040.586±.0050.668±.0046.142±.0593.800±.0017.793±.0240.837±.036
RWKV [21]single-person extension0.425±.0040.576±.0040.656±.0039.181±.0963.801±.0017.679±.0220.846±.026
separate modeling0.465±.0050.603±.0050.689±.0045.943±.0793.790±.0017.787±.0320.859±.036
TIMotion (Ours)0.501±.0050.656±.0060.734±.0064.702±.0693.769±.0017.943±.0341.005±.020
", + "bbox": [ + 133, + 342, + 864, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of different temporal modeling approaches on different interaction mixing structures. Our proposed causal interactive modeling is able to adapt to different interaction mixing architectures and outperforms the other two ways.", + "bbox": [ + 89, + 488, + 906, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\lambda_{DM} = 3, \\lambda_{RO} = 0.01$ and $\\lambda_{reg} = 1$ in all experiments, as same as InterGen. We train our diffusion denoisers with a batch size of 256 for 1500 epochs on 8 Nvidia L40S GPUs.", + "bbox": [ + 89, + 544, + 482, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-arts", + "text_level": 1, + "bbox": [ + 89, + 606, + 397, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Quantitative Results.", + "text_level": 1, + "bbox": [ + 89, + 631, + 282, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following established practices [16], each experiment is conducted 20 times, and the reported metric values represent the mean with a $95\\%$ statistical confidence interval. The results on InterHuman are shown in Tab. 1. In comparison to existing methods, our TIMotion achieves competitive results on three different interaction mixing architectures: transformer, mamba, and RWKV. Equipped with RWKV, TIMotion achieves 4.702 FID and 0.501 Top1 R precision, setting new state-of-the-art (SoTA) for the competitive InterHuman benchmark. The results on Inter-X [39] can be found in the supplementary materials.", + "bbox": [ + 89, + 652, + 482, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Qualitative Comparisons.", + "text_level": 1, + "bbox": [ + 89, + 833, + 313, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 4, we qualitatively compare the InterGen and our TIMotion. It can be seen that sequences generated by TIMotion are more consistent with the description.", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Effectiveness of TIMotion", + "text_level": 1, + "bbox": [ + 511, + 542, + 748, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of our proposed causal interactive modeling (TIMotion), we compare different temporal modeling approaches on three different structures, respectively. As shown in Tab. 2, single-person extension performs poorly, indicating that it does not model two-person interactions well. In contrast, separate modeling leverages the ability of the interaction mixing module to model the two-person interactions but achieves sub-optimal performance on the Transformer and mamba architectures. Our proposed approach is able to adapt to different interaction mixing architectures and outperforms the other two approaches.", + "bbox": [ + 511, + 568, + 906, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 511, + 768, + 769, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we conduct comprehensive ablation studies to investigate the effectiveness of the key components in TIMotion, thereby providing a deeper insight into our approach. All experiments are performed on the InterHuman dataset. Unless otherwise stated, we use RWKV as the interaction mixing structure for the following ablations. More theoretical analyses are in the supplementary materials.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7174", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/32f69a0128b9b6a9993a070ccf80c9be2b31ef8e8fab21a13d4d1b9d72f5ba4e.jpg", + "image_caption": [ + "[1] two individuals are lifting and transporting an object." + ], + "image_footnote": [], + "bbox": [ + 102, + 99, + 349, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/753a0bc2dbc51e6d9e0121a88cd27339c507bb7d5a0b897a290ec09baa4a2580.jpg", + "image_caption": [ + "[2] the other person pushes one person to the ground with both hands." + ], + "image_footnote": [], + "bbox": [ + 442, + 92, + 643, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cedfa3d746622368f0209122e2c7d9f47c33b5f692f50ca0dd848f8910ef4044.jpg", + "image_caption": [ + "[3] one wraps both arms around the back of the other person, while the other person reciprocates the action", + "Figure 4. Qualitative comparison with Intergen on human-human motion generation. Darker color indicates later frames. The sequences generated by TIMotion are more consistent with the text description." + ], + "image_footnote": [], + "bbox": [ + 743, + 90, + 864, + 309 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/addc9f83f0cf3ff8e94950f2b8c4357c501327b39026b2b01c6ab921c7719755.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CIIRESLPAR Precision Top 1↑FID ↓Params (M)
0.371±.0105.918±.079182
✓0.478±.0055.410±.069144
✓✓0.494±.0065.019±.077115
✓✓✓0.501±.0054.702±.069127
", + "bbox": [ + 109, + 401, + 465, + 497 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/34052abac03c8281759500561884a9ff6748dcc40c572492cdaf6439ef7c7edc.jpg", + "table_caption": [ + "Table 3. Ablation studies on the effectiveness of each component in TIMotion. \"CII\" denotes Causal Interactive Injection, \"RES\" denotes Role-Evolving Scanning, and \"LPA\" denotes Localized Pattern Amplification." + ], + "table_footnote": [], + "table_body": "
Kernel SizeNormR Precision Top 1↑FID ↓
k=3,1BN0.494±.0055.443±.093
k=3,1LN0.488±.0055.339±.080
k=3,1AdaLN0.501±.0054.702±.069
k=3,3AdaLN0.484±.0057.120±.090
k=5,1AdaLN0.497±.0055.102±.064
", + "bbox": [ + 555, + 401, + 864, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation studies on LPA. \"BN\" denotes batch normalization, \"LN\" denotes layer normalization and \"AdaLN\" denotes adaptive layer normalization. \"k=3,1\" means that the first kernel size of the convolution is 3 and the second kernel size is 1.", + "bbox": [ + 511, + 517, + 906, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Main Ablations", + "text_level": 1, + "bbox": [ + 89, + 592, + 245, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To understand the contribution of each component to the final performance, we incrementally add the proposed modules based on the baseline InterGen [16] and present the results in Tab. 3. Initially, we replace self-attention and cross-attention in the transformer with Causal Interactive Injection (CII), which effectively improves both R-Precision and FID while reducing the number of parameters. Next, we apply Role-Evolving Scanning (RES) and yield gains in both R-Precision and FID. It is worth noting that to keep the input feature dimension of RWKV constant, we reduce the dimension of the motion embedding to half of the original, thus also effectively reducing the number of parameters. Finally, when all three methods are applied together, the R-Precision achieves 0.501 and FID achieves 4.702.", + "bbox": [ + 88, + 612, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2. Design of Localized Pattern Amplification.", + "text_level": 1, + "bbox": [ + 89, + 835, + 431, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablations on the design of Localized Pattern Amplification (LPA), and the results are shown in Tab. 4. First, we explore the effect of different normal-", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "izations on model performance. Using batch normalization (BN) and layer normalization (LN) does not bring gains for text-to-motion tasks. In contrast, AdaLN, which integrates text information into the extraction process of the local motion pattern, proves to be the optimal normalization method. Additionally, we examine how the kernel size of the convolutional layers impacts the model's performance and select the setting of $k = 3,1$ .", + "bbox": [ + 511, + 618, + 906, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further demonstrate the effectiveness of LPA, we visualize the spectrum of TIMotion w/o and w/ LPA in Fig. 6, respectively. The magnitude has been normalized. As can be seen in Fig. 6, adding LPA reduces the high-frequency component of the feature, making the motion less likely to change drastically and making it smoother. Moreover, we randomly sampled 200 instances, and w/ LPA, the average proportion of the amplitude of high-frequency components is 0.3729, whereas w/o LPA, the average proportion of the amplitude of high-frequency components is 0.9063.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7175", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3175bf58ee1d5d1fc35fa25bd41e1d2b1066aa8d0504bd347afe7e8be989f26d.jpg", + "image_caption": [ + "Figure 5. Qualitative results on the motion in-between task. The first and last frames are fixed. Darker colors indicate later frames. Our method achieves smooth and natural transitions between the conditioned motions." + ], + "image_footnote": [], + "bbox": [ + 138, + 90, + 864, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dde4ba133cb0e6743496320a791b5007b3f85ee05c92f4833b169e11b14b4302.jpg", + "image_caption": [ + "Figure 6. Spectrum of motion features. (a) and (b) show the spectrum of TIMotion w/o and w/ LPA, respectively. The horizontal axis denotes the frequency and the vertical axis represents the normalized magnitude. TIMotion w/ LPA contains fewer high-frequency components and therefore generates smoother motion." + ], + "image_footnote": [], + "bbox": [ + 138, + 290, + 434, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Computational Complexity and Editability.", + "text_level": 1, + "bbox": [ + 89, + 666, + 459, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.1. Computational Complexity.", + "text_level": 1, + "bbox": [ + 89, + 688, + 330, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 5, we compare our approach TIMotion with the SOTA method InterGen [16] in terms of computational complexity. TIMotion requires fewer parameters and FLOPs than InterGen but outperforms it on the comprehensive metric FID and R precision. Notably, using a similar transformer architecture to InterGen, TIMotion's average inference time per sample is only 0.632 s while InterGen requires 1.991 s.", + "bbox": [ + 89, + 707, + 483, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.2. Editability.", + "text_level": 1, + "bbox": [ + 89, + 835, + 215, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 6 illustrates that TIMotion surpasses InterGen in the task of motion in-between editing. We perform experiments on the test set of InterHuman and evaluate by generat", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89d931d3c183c4bc13700cf5614ef3b780aa795091936965531d3f8e4f24b091.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsR Precision Top 1↑FID ↓Params (M)Flops (G)Inference time (s)
InterGen (Transformer)0.371±.0105.918±.07918280.51.991
TIMotion+Transformer0.491±.0055.433±.0806540.40.632
TIMotion+RWKV0.501±.0054.702±.06912758.01.733
", + "bbox": [ + 532, + 289, + 887, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing $80\\%$ of the sequences based on the first and last $10\\%$ of the sequences. The quantitative results are shown in Fig. 5 and our method achieves smooth and natural transitions between the conditioned motions.", + "bbox": [ + 511, + 400, + 906, + 460 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/67eeb12d159b501fe1a57cdf1ebc7ebd40ce471f78c99ac4ad0c5dbb3230f814.jpg", + "table_caption": [ + "Table 5. Comparison of computational complexity." + ], + "table_footnote": [], + "table_body": "
MethodsR Precision Top 1↑FID ↓MM Dist↓Diversity→
InterGen0.461±.0064.700±.0663.780±.0017.682±.029
TIMotion0.516±.0063.590±.0493.760±.0017.795±.031
", + "bbox": [ + 532, + 474, + 887, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Evaluation of motion in-between editing task.", + "bbox": [ + 537, + 544, + 880, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 601, + 633, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we abstract the overall human-human motion generation process into a general framework MetaMotion, which consists of two phases: temporal modeling and interaction mixing. We find that the current methods lead to sub-optimal results and redundancy of model parameters due to inadequate modeling. Based on this, we introduce TIMotion (Temporal and Interactive Modeling), an efficient and effective approach for human-human motion generation. Specifically, we first propose Causal Interactive Injection to model two separate sequences as a causal sequence leveraging the temporal and causal properties. Additionally, we proposed Role-Evolving Mixing to adapt to the dynamic roles throughout interactions and designed Localized Pattern Amplification to capture short-term motion patterns for generating smoother and more rational motion. Extensive experiments on the InterHuman and Inter-X datasets demonstrate that TIMotion significantly outperforms existing methods, achieving state-of-the-art results.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7176", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 1", + "[3] Erik Bethke. Game development and production. Wordware Publishing, Inc., 2003. 1", + "[4] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 4", + "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 2", + "[6] Ke Fan, Junshu Tang, Weijian Cao, Ran Yi, Moran Li, Jingyu Gong, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Freemotion: A unified framework for number-free text-to-motion synthesis. In European Conference on Computer Vision, pages 93-109. Springer, 2024. 3", + "[7] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. Tm2d: Bimodality driven 3d dance generation via music-text integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9942-9952, 2023. 2", + "[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4, 6", + "[9] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1", + "[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 6", + "[11] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 1", + "[12] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In ACM SIGGRAPH 2022 conference proceedings, pages 1-9, 2022. 1" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 5", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[16] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 1, 3, 4, 5, 6, 7, 8", + "[17] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[18] Nadia Magnenat-Thalmann, Daniel Thalmann, Nadia Magnenat-Thalmann, and Daniel Thalmann. Computer animation. Springer, 1985. 1", + "[19] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 5", + "[20] Rick Parent. Computer animation: algorithms and techniques. Newnes, 2012. 1", + "[21] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 4, 6", + "[22] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1", + "[23] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2, 6", + "[24] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1546-1555, 2024. 1", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[26] George Saridis. Intelligent robotic control. IEEE Transactions on Automatic Control, 28(5):547-557, 1983. 1", + "[27] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2, 6", + "[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7177", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 2, 4, 6", + "[30] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2", + "[31] G Tevet, S Raab, B Gordon, Y Shafir, D Cohen-Or, and AH Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 6", + "[32] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1", + "[33] Jay Urbain. Introduction to game development. Cell, 414: 745-5102, 2010. 1", + "[34] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2", + "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 6", + "[36] Shuo Wang, Xinhai Zhao, Hai-Ming Xu, Zehui Chen, Dameng Yu, Jiahao Chang, Zhen Yang, and Feng Zhao. Towards domain generalization for multi-view 3d object detection in bird-eye-view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13333-13342, 2023. 1", + "[37] Shuo Wang, Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd-map construction. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 1", + "[38] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 3", + "[39] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter- $\\mathbf{x}$ : Towards versatile human-human interaction analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22260-22271, 2024. 5, 6", + "[40] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 2", + "[41] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2", + "[42] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 5", + "[43] Chen Zhao, Weiling Cai, Chenyu Dong, and Chengwei Hu. Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8281-8291, 2024. 1", + "[44] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 509-519, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 273 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7178", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_model.json b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ca650378f25fd0288f5960a0d93c2dabcd06ff5 --- /dev/null +++ b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.768, + 0.175 + ], + "angle": 0, + "content": "TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.203, + 0.753, + 0.221 + ], + "angle": 0, + "content": "Yabiao Wang\\(^{1,2*}\\); Shuo Wang\\(^{2*}\\); Jiangning Zhang\\(^{2}\\), Ke Fan\\(^{3}\\)," + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.221, + 0.671, + 0.239 + ], + "angle": 0, + "content": "Jiafu Wu², Zhucun Xue¹, Yong Lii¹" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.239, + 0.796, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University \\(^{2}\\)Youtu Lab, Tencent \\(^{3}\\)Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.26, + 0.637, + 0.275 + ], + "angle": 0, + "content": "https://aigc-explorer.github.io/TIMotion-page/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Human-human motion generation is essential for understanding humans as social beings. Current methods fall into two main categories: single-person-based methods and separate modeling-based methods. To delve into this field, we abstract the overall generation process into a general framework MetaMotion, which consists of two phases: temporal modeling and interaction mixing. For temporal modeling, the single-person-based methods concatenate two people into a single one directly, while the separate modeling-based methods skip the modeling of interaction sequences. The inadequate modeling described above resulted in sub-optimal performance and redundant model parameters. In this paper, we introduce TIMotion (Temporal and Interactive Modeling), an efficient and effective framework for human-human motion generation. Specifically, we first propose Causal Interactive Injection to model two separate sequences as a causal sequence leveraging the temporal and causal properties. Then we present Role-Evolving Scanning to adjust to the change in the active and passive roles throughout the interaction. Finally, to generate smoother and more rational motion, we design Localized Pattern Amplification to capture short-term motion patterns. Extensive experiments on InterHuman and InterX demonstrate that our method achieves superior performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.749, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.483, + 0.866 + ], + "angle": 0, + "content": "In the field of generative computer vision, human motion generation has significant implications for computer animation [18, 20], game development [3, 33], and robotic control [26, 36, 37]. In recent years, there have been remarkable advancements in human motion generation, driven by various user-specified conditions such as action cat" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.906, + 0.49 + ], + "angle": 0, + "content": "egories [9, 22], speeches [2, 12], and natural language prompts [11, 24]. Among these, many approaches leveraging large language models [1] and diffusion models [43] have yielded impressive results in generating realistic and diverse motions, benefiting from their powerful modeling capabilities. Despite this progress, most existing methods are designed primarily for single-person scenarios, thereby neglecting a crucial element of human motion: the complex and dynamic interactions between individuals. We address the challenge of generating two-person motion by fully utilizing the temporal and interactive dynamics between the individuals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.809 + ], + "angle": 0, + "content": "To better explore human-human motion generation, we abstract a general framework MetaMotion, as shown on the left of Fig. 1, which consists of two phases: temporal modeling and interaction mixing. Previous approaches have prioritized interaction mixing over temporal modeling and can be divided into two primary categories: single-person-based methods and separate modeling-based methods. As shown in Fig. 1(a), the single-person-based methods (e.g. MDM [32]) concatenate two individuals into one, which is then fed into the existing single-person motion generation module, e.g. DiT. And the separate modeling-based methods (e.g. InterGen [16]), as shown in Fig. 1(b), model two individuals individually and then extract motion information from both themselves and each other, using self-attention and cross-attention mechanisms, respectively. Following the general logic of MetaMotion, we introduce the Temporal and Interactive Framework, as depicted in Fig. 1(c), which models the human-human causal interactions. In fact, this effective temporal modeling method can simplify the design of the interaction mixing module and reduce the number of learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we design an effective temporal modeling method and introduce TIMotion, a temporal and interactive framework for efficient human-human motion generation. Specifically, we first propose Causal Interactive Injection to model the two single-person motion sequences as a causal interaction sequence, according to the temporal causal prop" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.225, + 0.888 + ], + "angle": 0, + "content": "*Equal contributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.226, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.091, + 0.361, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.091, + 0.508, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.091, + 0.661, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.098, + 0.895, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 1. MetaMotion and performance of MetaMotion-based models on InterHuman validation set. We abstract the MetaMotion concept that illustrates the intrinsic properties of human-human motion generation in the interaction process. (a) and (b) show the two types of methods currently, and (c) shows our method TIMotion, LPA refers to the Localized Pattern Amplification. In (d) we compare the performance of the different methods on the InterHuman dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.482, + 0.508 + ], + "angle": 0, + "content": "erties of motion sequences. Second, since active and passive roles are not static in interactions, we propose Role-Evolving Mixing to make two humans act in both active and passive roles. Then the network can dynamically adjust the roles of the two humans based on the text's semantics and the motion's context. Finally, we propose Localized Pattern Amplification, which captures short-term motion patterns for each individual separately, resulting in smoother and more logical motion generation. Our proposed framework can be well adapted to different interaction-mixing modules, including Transformer, RWKV, and Mamba." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.509, + 0.435, + 0.523 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.528, + 0.483, + 0.618 + ], + "angle": 0, + "content": "- We conceptualize human-human motion generation within a general framework MetaMotion and design an innovative method. Our proposed framework, TIMotion, is versatile enough to integrate with various interaction-mixing modules (e.g. Transformer, RWKV, Mamba) and reduces the number of parameters of these modules." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.619, + 0.483, + 0.724 + ], + "angle": 0, + "content": "- To utilize the temporal and causal properties, we propose Causal Interactive Injection to model two separate motion sequences as a unified causal sequence. In addition, we introduce Role-Evolving Scanning to accommodate shifts between the active and passive roles during interactions. Moreover, we also design Localized Pattern Amplification to capture short-term motion patterns effectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.725, + 0.483, + 0.784 + ], + "angle": 0, + "content": "- We perform extensive experiments on the benchmark human-human motion generation datasets: InterHuman and Inter-X. The results demonstrate the effectiveness and generalizability of our proposed methods." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.528, + 0.483, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.233, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Single-Person Human Motion Generation. Creating human motion is vital for applications such as 3D modeling and robot manipulation. The primary approach, known as the Text-to-Motion task, involves learning a unified latent space for both language and motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.524 + ], + "angle": 0, + "content": "Autoencoders have been widely adopted in human motion generation. MotionCLIP [30] effectively integrates semantic knowledge from CLIP [25] into the human motion manifold. TEMOS [23] and T2M [10] combine a Transformer-based VAE with a text encoder to generate motion sequences based on text descriptions. AttT2M [44] and TM2D [7] integrate a spatial-temporal body-part encoder into VQ-VAE [34] to improve the learning of discrete latent space. T2M-GPT [40] redefines text-driven motion generation as a next-index prediction task and proposes a framework based on VQ-VAE and Generative Pretrained Transformer (GPT) for motion generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.906, + 0.707 + ], + "angle": 0, + "content": "Recently, diffusion-based generative modeling has been gaining attention. MotionDiffuse [42] exhibits several desirable properties, such as probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [31] aims to predict motion directly and incorporates a geometric loss to boost the model's performance. Instead of employing a diffusion model to link raw motion sequences with conditional inputs, MLD [5] further utilizes the latent diffusion model to reduce the training and inference costs substantially. ReMoDiffuse [41] introduces an improvement mechanism based on dataset retrieval to enhance the denoising process of Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Multi-Person Human Motion Generation. As multiperson motion synthesis involves the interactive dynamics between multiple individuals, it is more challenging than single-person motion generation. Early work typically relied on motion graphs and momentum-based inverse kinematics to model human joints. Recently, ComMDM [27] finetunes a pre-trained text-to-motion diffusion model using a small scale of two human motions. RIG [29] converts the text of asymmetric interactions into both active and passive voice to maintain consistent textual context for each individual. To complete the interaction process by" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.912, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.324, + 0.908, + 0.368 + ], + "angle": 0, + "content": "Figure 2. The overall framework of our TIMotion. We contribute three primary technical designs. First, we propose Causal Interactive Injection to utilize the temporal properties of motion sequences. Then we present Role-Evolving Mixing to adjust to the ever-evolving roles during interaction. Finally, we design Localized Pattern Amplification to capture short-term motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.393, + 0.484, + 0.53 + ], + "angle": 0, + "content": "manipulating the joints of two individuals to specific positions, InterControl [38] employs a Large Language Model to generate movement plans for the two individuals by designing prompts. InterGen [16] introduces a large-scale, text-annotated two-person motion dataset. Building on this dataset, it proposes a diffusion model with shared weights and multiple regularization losses. FreeMotion [6] proposes to decouple the process of conditional motion generation and support the number-free motion synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.182, + 0.575 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.59, + 0.313, + 0.605 + ], + "angle": 0, + "content": "3.1. Motion Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Current methods usually use diffusion models [15] for motion generation. For the Interhuman dataset, a noncanonical representation [16] is typically used. The representation is formulated as: \\( x^{i} = [\\mathbf{j}_{g}^{p},\\mathbf{j}_{g}^{v},\\mathbf{j}^{r},\\mathbf{c}^{f}] \\), where the \\( i \\)-th motion state \\( x^{i} \\) is defined as a collection of global joint positions \\( \\mathbf{j}_g^p\\in \\mathbb{R}^{3N_j} \\), velocities \\( \\mathbf{j}_g^v\\in \\mathbb{R}^{3N_j} \\) in the world frame, 6D representation of local rotations \\( \\mathbf{j}^r\\in \\mathbb{R}^{6N_j} \\) in the root frame, and binary foot-ground contact \\( \\mathbf{c}^f\\in \\mathbb{R}^4 \\). Our goal is to train a model parameterized by \\( \\theta \\) to approximate the human interactive motion data distribution \\( p(\\mathbf{x_0}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Diffusion Process. Following a specified schedule \\(\\beta_{t} \\in (0,1)\\), models incrementally degrade input data \\(\\mathbf{x}_0 \\sim p(\\mathbf{x}_0)\\), ultimately transforming the data distribution into an isotropic Gaussian over \\(T\\) steps. Each diffusion transition can be considered as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.873, + 0.484, + 0.899 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.393, + 0.841, + 0.409 + ], + "angle": 0, + "content": "where the full diffusion process can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.42, + 0.907, + 0.455 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {1 \\leq t \\leq T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Denoising Process. In the denoising process, models are trained to reverse the diffusion procedure, enabling them to transform random noise into real data distribution during inference. The denoising process can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.537, + 0.906, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right) \\\\ = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon\\right), \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} \\beta_ {t}\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.746 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), \\(\\alpha_{t} = 1 - \\beta_{t}\\), \\(\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}\\) and \\(\\theta\\) denotes parameters of the network learning to denoise. The training objective is to maximize the likelihood of observed data \\(p_{\\theta}(\\mathbf{x}_0) = \\int p_{\\theta}(\\mathbf{x}_{0:T})d\\mathbf{x}_{1:T}\\), by maximizing its evidence lower bound (ELBO), which effectively aligns the true denoising model \\(q(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t)\\) with the parameterized \\(p_{\\theta}(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t)\\). During training, the goal of the denoising network \\(\\epsilon_{\\theta}(.)\\) is to reconstruct \\(\\mathbf{x}_0\\) given any noised input \\(\\mathbf{x}_t\\), by predicting the added noise \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) via minimizing the noise prediction error" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.757, + 0.907, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\right\\| ^ {2} \\right]. (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.793, + 0.908, + 0.84 + ], + "angle": 0, + "content": "To condition the model on additional context information \\(\\mathbf{c}\\), e.g., text, we inject \\(\\mathbf{c}\\) into \\(\\epsilon_{\\theta}(.)\\) by replacing \\(\\mu_{\\theta}(\\mathbf{x}_t,t)\\) and \\(\\Sigma_{\\theta}(\\mathbf{x}_t,t)\\) with \\(\\mu_{\\theta}(\\mathbf{x}_t,t,\\mathbf{c})\\) and \\(\\Sigma_{\\theta}(\\mathbf{x}_t,t,\\mathbf{c})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.649, + 0.863 + ], + "angle": 0, + "content": "3.2. MetaMotion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "We present the core concept \"MetaMotion\" for human-human motion generation at first. As shown in Fig. 1," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7171" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "we abstract the process of human-human motion generation into two phases: temporal modeling and interaction mixing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.184 + ], + "angle": 0, + "content": "Specifically, two single-person sequences go through the temporal modeling module to get the input sequence \\( X \\). Then, the input sequence is fed to the interaction-mixing module and this process can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.196, + 0.483, + 0.213 + ], + "angle": 0, + "content": "\\[\nY = \\operatorname {I n t e r a c t i o n M i x i n g} (X), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.225, + 0.484, + 0.287 + ], + "angle": 0, + "content": "where InteractionMixing is usually the transformer structure including self-attention and cross-attention. Notably, InteractionMixing can also be some emerging structures, e.g., Mamba [8], RWKV [21]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.296, + 0.204, + 0.311 + ], + "angle": 0, + "content": "3.3. TIMotion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.439 + ], + "angle": 0, + "content": "For human-human motion generation, most current methods are based on extensions of single-person motion generation. Some design-specific approach (e.g. InterGen [16]) considers the interaction between two persons in the interaction mixing module. However, current methods ignore the importance of temporal modeling, resulting in suboptimal performance in generating long-sequence motions and multi-person interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.592 + ], + "angle": 0, + "content": "In this work, we propose TIMotion, as illustrated in Fig. 2, an effective temporal modeling approach that can be applied to different interaction-mixing structures, enabling better handling the human-human interactions. TIMotion consists of three key technical designs: (1) Causal Interactive Injection, which identifies the causal relationships between human-human interactions; (2) Role-Evolving Scanning, which adapts to shifts between active and passive roles during interactions; (3) Localized Pattern Amplification, designed to handle short-term sequences better." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.601, + 0.332, + 0.616 + ], + "angle": 0, + "content": "3.3.1. Causal Interactive Injection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.71 + ], + "angle": 0, + "content": "Perception of ego-motion and interaction between two persons are two important elements of human-human motion generation. Considering the causal properties of motion, we propose Causal Interactive Injection, a temporal modeling approach to simultaneously achieve both perceptions of ego-motion and interactions with each other." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.484, + 0.853 + ], + "angle": 0, + "content": "Specifically, we denote two single-person motion sequences as \\(\\{x_{a}, x_{b}\\}\\), where \\(x_{a} = \\{x_{a}^{j}\\}_{j=1}^{L}\\) and \\(x_{b} = \\{x_{b}^{j}\\}_{j=1}^{L}\\) are respective sequences of motion, and \\(L\\) is the length of the sequence. Since the motions of two individuals at the current time step are jointly determined by their motions at previous time steps, we model the two single-person motion sequences as a causal interaction sequence \\(x_{cii} = \\{x_{k}^{j/2}\\}_{j=1}^{2L}\\), the symbol // denotes division followed by rounding up and \\(k\\) can be acquired as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.864, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\nk = \\left\\{ \\begin{array}{ll}a, & j\\% 2 = 1.\\\\ b, & j\\% 2 = 0. \\end{array} \\right. \\tag{6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "Then we can inject them into the interaction-mixing module and separate the motion embeddings of the two individuals from the outputs according to Eq. (6)." + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.155, + 0.62, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.22, + 0.615, + 0.229 + ], + "angle": 0, + "content": "frame 1" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.157, + 0.73, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.22, + 0.719, + 0.229 + ], + "angle": 0, + "content": "frame 2" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.159, + 0.861, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.812, + 0.22, + 0.849, + 0.229 + ], + "angle": 0, + "content": "frame 3" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.249, + 0.635, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.307, + 0.616, + 0.316 + ], + "angle": 0, + "content": "frame 4" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.245, + 0.729, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.308, + 0.721, + 0.316 + ], + "angle": 0, + "content": "frame 5" + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.245, + 0.87, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.307, + 0.851, + 0.316 + ], + "angle": 0, + "content": "frame 6" + }, + { + "type": "image_footnote", + "bbox": [ + 0.571, + 0.332, + 0.866, + 0.365 + ], + "angle": 0, + "content": "The first person suddenly pushes the other one. The second person takes a step back, then forcefully pushes the first person. The first person was knocked down." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.419 + ], + "angle": 0, + "content": "Figure 3. Illustration of changing active and passive roles. The first person acts as the active role in the early stages, and as time progresses, the other person becomes the active role of the motion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.443, + 0.729, + 0.459 + ], + "angle": 0, + "content": "3.3.2. Role-Evolving Scanning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.628 + ], + "angle": 0, + "content": "As noted in [4], human interactions inherently follow a certain order, e.g., \"shaking hands\" typically begins with one person extending a hand first, which means that interactive motion can be categorized as active or passive. Some methods [29] split the text description into active and passive voices. However, the \"active\" and \"passive\" roles are not static in a text description. As shown in Fig. 3, these roles constantly swap between characters as the interaction progresses. To minimize redundant text preprocessing and adapt to the ongoing shifts in roles, we design Role-Evolving Scanning, which is both efficient and effective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.781 + ], + "angle": 0, + "content": "For the causal interaction sequence \\( x \\) defined in Causal Interactive Injection (Sec. 3.3.1), it is obvious that \\( x_{a} \\) and \\( x_{b} \\) represent the active sequence and passive sequence, respectively. However, the above assumption of active and passive motions is not always in line with the real order. To cope with the changing roles, we re-model the interactive motion sequences as a symmetric causal interaction sequence \\( x_{sym\\_ cii} = \\{x_{k'}^{j // 2}\\}_{j=1}^{2L} \\), where the symbol \\( // \\) is defined as division followed by rounding up and \\( k' \\) is obtained by exchanging \\( a \\) and \\( b \\) in Eq. (6)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.873 + ], + "angle": 0, + "content": "Given that the causal interaction sequence \\( x_{cii} \\in \\mathbb{R}^{2L \\times C} \\) and the symmetric causal interaction sequence \\( x_{sym\\_ cii} \\in \\mathbb{R}^{2L \\times C} \\), where \\( L \\) is the length of the sequence and \\( C \\) is the dimension of motion embedding, we can acquire the final interaction sequence \\( X \\in \\mathbb{R}^{2L \\times 2C} \\) through Role-Evolving Scanning as:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.886, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\nX = \\operatorname {C o n c a t} \\left(x _ {c i i}, x _ {s y m. c i i}\\right). \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7172" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.23 + ], + "angle": 0, + "content": "Then the sequence \\(X\\) is fed to the interaction-mixing module. After obtaining the output \\(Y \\in \\mathbb{R}^{2L \\times 2C}\\) through Eq. (5), we first divide it at the channel level into causal interaction embeddings and symmetric causal interaction embeddings. According to \\(x_{cii}\\) and \\(x_{sym\\_ cii}\\), we can obtain the two individuals' splitting embeddings twice. Then we merge the embeddings of the two individuals to obtain the final global motion embeddings \\(y_a^g \\in \\mathbb{R}^{L \\times C}\\) and \\(y_b^g \\in \\mathbb{R}^{L \\times C}\\). The overall process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.244, + 0.483, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} y _ {a 1}, y _ {b 1} = \\operatorname {S p l i t} (Y [ :, : C ]), \\\\ y _ {b 2}, y _ {a 2} = \\operatorname {S p l i t} (Y [:, C: ]), \\tag {8} \\\\ y _ {a} ^ {g} = y _ {a 1} \\oplus y _ {a 2}, \\\\ y _ {b} ^ {g} = y _ {b 1} \\oplus y _ {b 2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.334, + 0.353, + 0.348 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes element-wise sum." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.485, + 0.411 + ], + "angle": 0, + "content": "By utilizing Role-Evolving Mixing to make two humans act as both active and passive roles, the network can dynamically adjust the roles of the two humans based on the semantics of the text and the context of the motion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.425, + 0.36, + 0.44 + ], + "angle": 0, + "content": "3.3.3. Localized Pattern Amplification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Transformers and RNNs excel at global modeling and capturing long-range dependencies but tend to overlook local semantic information [13]. Additionally, Causal Interactive Injection and Role-Evolving Mixing mainly model the overall motion based on the causality of the interactions, but neglect to focus on the localized motion patterns of the single person. To address this issue, we propose Localized Pattern Amplification, which summarizes short-term motion patterns for each person individually while generating smoother and more rational motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.675 + ], + "angle": 0, + "content": "Specifically, we utilize 1-D convolution layers and the residual structure to realize Localized Pattern Amplification. Given that the condition embedding \\( e \\) and two single-person motion sequences \\( x_{a} \\) and \\( x_{b} \\), the local motion embedding for \\( x_{a} \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.688, + 0.483, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {a} ^ {l} = \\operatorname {C o n v} _ {3} (\\operatorname {A d a L N} (x _ {a}, e)), \\\\ y _ {a} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {A d a L N} \\left(x _ {a} ^ {l}, e\\right)\\right), \\tag {9} \\\\ y _ {a} ^ {l} = x _ {a} + y _ {a}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.825 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv_k}\\) denotes the 1-D convolution with kernel size \\(k\\) and AdaLN is the adaptive layer normalization used in [42]. The process for \\(x_{b}\\) is the same as for \\(x_{a}\\) and they share network weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Once we obtain the outputs \\(\\{y_a^g, y_b^g\\}\\) through Eq. (8) and the output \\(\\{y_a^l, y_b^l\\}\\) from the convolution block, global embeddings and local embeddings are aggregated through concatenation along the channel dimension, followed by a linear layer to restore the original channels, resulting in the" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.089, + 0.716, + 0.109 + ], + "angle": 0, + "content": "final outputs \\(\\{y_a^{final}, y_b^{final}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.115, + 0.905, + 0.157 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} y _ {a} ^ {f i n a l} = \\operatorname {L i n e a r} (\\operatorname {C o n c a t} (y _ {a} ^ {g}, y _ {a} ^ {l})), \\\\ y _ {b} ^ {\\text {f i n a l}} = \\operatorname {L i n e a r} \\left(\\operatorname {C o n c a t} \\left(y _ {b} ^ {g}, y _ {b} ^ {l}\\right)\\right). \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.164, + 0.907, + 0.195 + ], + "angle": 0, + "content": "Then the final outputs are fed into the next encoder block or the final decoder layer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.201, + 0.694, + 0.216 + ], + "angle": 0, + "content": "3.3.4. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.22, + 0.907, + 0.31 + ], + "angle": 0, + "content": "We use the same loss functions as InterGen [16], including the diffusion loss \\(\\mathcal{L}_{simple}\\), the foot contact loss \\(\\mathcal{L}_{foot}\\), joint velocity loss \\(\\mathcal{L}_{vel}\\), the bone length loss \\(\\mathcal{L}_{BL}\\), the masked joint distance map loss \\(\\mathcal{L}_{DM}\\), and the relative orientation loss \\(\\mathcal{L}_{RO}\\). For more details about the losses, we refer readers to InterGen. Finally, the overall loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.32, + 0.905, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {m o t i o n}} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {v e l} \\mathcal {L} _ {v e l} + \\lambda_ {f o o t} \\mathcal {L} _ {f o o t} \\\\ + \\lambda_ {B L} \\mathcal {L} _ {B L} + \\lambda_ {D M} \\mathcal {L} _ {D M} + \\lambda_ {R O} \\mathcal {L} _ {R O}, \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.906, + 0.393 + ], + "angle": 0, + "content": "where the hyper-parameters \\(\\lambda_{vel},\\lambda_{foot},\\lambda_{BL},\\lambda_{DM},\\lambda_{RO}\\) are the same as InterGen." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.406, + 0.646, + 0.423 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.431, + 0.704, + 0.448 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.589 + ], + "angle": 0, + "content": "Datasets. We assess our proposed framework using the InterHuman [16] and Inter-X [39] dataset. InterHuman is the first dataset to incorporate text annotations for two-person motions. This dataset includes 6,022 motions spanning various categories of human actions and is labeled with 16,756 unique descriptions made up of 5,656 distinct words. InterX is the largest human-human interaction dataset with diverse interaction patterns. It includes about 11k interaction sequences and more than 8.1M frames." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.589, + 0.906, + 0.68 + ], + "angle": 0, + "content": "Metrics. We employ the same evaluation metrics as InterGen [16], which are as follows: (1) Frechet Inception Distance (FID). (2) \\(R\\)-Precision. (3) Diversity. (4) Multimodality (MModality). (5) Multi-modal distance (MM Dist). Detailed explanations and calculations of the metrics can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.686, + 0.727, + 0.701 + ], + "angle": 0, + "content": "4.1.1. Implementation Details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We use a frozen CLIP-ViT-L/14 model as the text encoder. The dimension of motion embedding is set to 512. During training, the number of diffusion timesteps is set to 1000, and we employ the DDIM [28] sampling strategy with 50 timesteps and \\(\\eta = 0\\). The cosine noise level schedule [19] and classifier-free guidance [14] are adopted, with \\(10\\%\\) of random CLIP embeddings set to zero during training and a guidance coefficient of 3.5 during sampling. All the models are trained using the AdamW [17] optimizer with betas of (0.9, 0.999), a weight decay of \\(2 \\times 10^{-5}\\), a maximum learning rate of \\(10^{-4}\\), and a cosine learning rate schedule with 10 linear warm-up epochs. To balance the contributions of different loss terms, we set \\(\\lambda_{vel} = 30\\), \\(\\lambda_{foot} = 30\\), \\(\\lambda_{BL} = 10\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7173" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.904, + 0.278 + ], + "angle": 0, + "content": "
MethodsR Precision↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.452±.0080.610±.0090.701±.0080.273±.0073.755±.0087.948±.064-
TEMOS [23]0.224±.0100.316±.0130.450±.01817.375±.0435.342±.0156.939±.0710.535±.014
T2M [10]0.238±.0120.325±.0100.464±.01413.769±.0724.731±.0137.046±.0221.387±.076
MDM [31]0.153±.0120.260±.0090.339±.0129.167±.0566.125±.0187.602±.0452.355±.080
ComMDM* [27]0.067±.0130.125±.0180.184±.01538.643±.09813.211±.0133.520±.0580.217±.018
ComMDM [27]0.223±.0090.334±.0080.466±.0107.069±.0545.212±.0217.244±.0381.822±.052
RIG [29]0.285±.0100.409±.0140.521±.0136.775±.0694.876±.0187.311±.0432.096±.065
InterGen [16]0.371±.0100.515±.0120.624±.0105.918±.0795.108±.0147.387±.0292.141±.063
TIMotion+Transformer(ours)0.491±.0050.648±.0040.724±.0045.433±.0803.775±.0018.032±.0300.952±.032
TIMotion+Mamba(ours)0.431±.0040.586±.0050.668±.0046.142±.0593.800±.0017.793±.0240.837±.036
TIMotion+RWKV(ours)0.501±.0050.656±.0060.734±.0064.702±.0693.769±.0017.943±.0341.005±.020
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.288, + 0.907, + 0.332 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on the InterHuman [16] test set. We run all the evaluations 20 times. ± indicates a 95% confidence interval. Bold indicates the best result, while underline refers to the second best. ComMDM* indicates the ComMDM model fine-tuned in the original few-shot setting with 10 training samples and ComMDM indicates fine-tuned on the entire InterHuman training set." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.343, + 0.866, + 0.48 + ], + "angle": 0, + "content": "
Interaction MixingTemporal ModelingR Precision↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Transformer [35]single-person extension0.395±.0040.535±.0040.615±.0048.028±.0993.820±.0017.687±.0220.852±.024
separate modeling0.371±.0100.515±.0120.624±.0105.918±.0795.108±.0147.387±.0292.141±.063
TIMotion (Ours)0.491±.0050.648±.0040.724±.0045.433±.0803.775±.0018.032±.0300.952±.032
Mamba [8]single-person extension0.368±.0050.513±.0050.597±.0067.232±.0813.825±.0017.864±.0240.870±.020
separate modeling0.420±.0050.569±.0040.650±.0037.221±.0973.803±.0017.932±.0350.855±.020
TIMotion (Ours)0.431±.0040.586±.0050.668±.0046.142±.0593.800±.0017.793±.0240.837±.036
RWKV [21]single-person extension0.425±.0040.576±.0040.656±.0039.181±.0963.801±.0017.679±.0220.846±.026
separate modeling0.465±.0050.603±.0050.689±.0045.943±.0793.790±.0017.787±.0320.859±.036
TIMotion (Ours)0.501±.0050.656±.0060.734±.0064.702±.0693.769±.0017.943±.0341.005±.020
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.489, + 0.907, + 0.518 + ], + "angle": 0, + "content": "Table 2. Comparison of different temporal modeling approaches on different interaction mixing structures. Our proposed causal interactive modeling is able to adapt to different interaction mixing architectures and outperforms the other two ways." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.59 + ], + "angle": 0, + "content": "\\(\\lambda_{DM} = 3, \\lambda_{RO} = 0.01\\) and \\(\\lambda_{reg} = 1\\) in all experiments, as same as InterGen. We train our diffusion denoisers with a batch size of 256 for 1500 epochs on 8 Nvidia L40S GPUs." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.607, + 0.398, + 0.623 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-arts" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.632, + 0.284, + 0.646 + ], + "angle": 0, + "content": "4.2.1. Quantitative Results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.82 + ], + "angle": 0, + "content": "Following established practices [16], each experiment is conducted 20 times, and the reported metric values represent the mean with a \\(95\\%\\) statistical confidence interval. The results on InterHuman are shown in Tab. 1. In comparison to existing methods, our TIMotion achieves competitive results on three different interaction mixing architectures: transformer, mamba, and RWKV. Equipped with RWKV, TIMotion achieves 4.702 FID and 0.501 Top1 R precision, setting new state-of-the-art (SoTA) for the competitive InterHuman benchmark. The results on Inter-X [39] can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.315, + 0.85 + ], + "angle": 0, + "content": "4.2.2. Qualitative Comparisons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In Fig. 4, we qualitatively compare the InterGen and our TIMotion. It can be seen that sequences generated by TIMotion are more consistent with the description." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.749, + 0.558 + ], + "angle": 0, + "content": "4.3. Effectiveness of TIMotion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.751 + ], + "angle": 0, + "content": "To validate the effectiveness of our proposed causal interactive modeling (TIMotion), we compare different temporal modeling approaches on three different structures, respectively. As shown in Tab. 2, single-person extension performs poorly, indicating that it does not model two-person interactions well. In contrast, separate modeling leverages the ability of the interaction mixing module to model the two-person interactions but achieves sub-optimal performance on the Transformer and mamba architectures. Our proposed approach is able to adapt to different interaction mixing architectures and outperforms the other two approaches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.771, + 0.787 + ], + "angle": 0, + "content": "4.4. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this section we conduct comprehensive ablation studies to investigate the effectiveness of the key components in TIMotion, thereby providing a deeper insight into our approach. All experiments are performed on the InterHuman dataset. Unless otherwise stated, we use RWKV as the interaction mixing structure for the following ablations. More theoretical analyses are in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7174" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.1, + 0.35, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.31, + 0.334, + 0.33 + ], + "angle": 0, + "content": "[1] two individuals are lifting and transporting an object." + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.093, + 0.645, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.31, + 0.644, + 0.331 + ], + "angle": 0, + "content": "[2] the other person pushes one person to the ground with both hands." + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.092, + 0.866, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.311, + 0.898, + 0.34 + ], + "angle": 0, + "content": "[3] one wraps both arms around the back of the other person, while the other person reciprocates the action" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.35, + 0.908, + 0.38 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with Intergen on human-human motion generation. Darker color indicates later frames. The sequences generated by TIMotion are more consistent with the text description." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.402, + 0.466, + 0.498 + ], + "angle": 0, + "content": "
CIIRESLPAR Precision Top 1↑FID ↓Params (M)
0.371±.0105.918±.079182
✓0.478±.0055.410±.069144
✓✓0.494±.0065.019±.077115
✓✓✓0.501±.0054.702±.069127
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.565 + ], + "angle": 0, + "content": "Table 3. Ablation studies on the effectiveness of each component in TIMotion. \"CII\" denotes Causal Interactive Injection, \"RES\" denotes Role-Evolving Scanning, and \"LPA\" denotes Localized Pattern Amplification." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.402, + 0.865, + 0.508 + ], + "angle": 0, + "content": "
Kernel SizeNormR Precision Top 1↑FID ↓
k=3,1BN0.494±.0055.443±.093
k=3,1LN0.488±.0055.339±.080
k=3,1AdaLN0.501±.0054.702±.069
k=3,3AdaLN0.484±.0057.120±.090
k=5,1AdaLN0.497±.0055.102±.064
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.518, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Table 4. Ablation studies on LPA. \"BN\" denotes batch normalization, \"LN\" denotes layer normalization and \"AdaLN\" denotes adaptive layer normalization. \"k=3,1\" means that the first kernel size of the convolution is 3 and the second kernel size is 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.593, + 0.246, + 0.607 + ], + "angle": 0, + "content": "4.4.1. Main Ablations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.483, + 0.825 + ], + "angle": 0, + "content": "To understand the contribution of each component to the final performance, we incrementally add the proposed modules based on the baseline InterGen [16] and present the results in Tab. 3. Initially, we replace self-attention and cross-attention in the transformer with Causal Interactive Injection (CII), which effectively improves both R-Precision and FID while reducing the number of parameters. Next, we apply Role-Evolving Scanning (RES) and yield gains in both R-Precision and FID. It is worth noting that to keep the input feature dimension of RWKV constant, we reduce the dimension of the motion embedding to half of the original, thus also effectively reducing the number of parameters. Finally, when all three methods are applied together, the R-Precision achieves 0.501 and FID achieves 4.702." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.836, + 0.433, + 0.851 + ], + "angle": 0, + "content": "4.4.2. Design of Localized Pattern Amplification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we conduct ablations on the design of Localized Pattern Amplification (LPA), and the results are shown in Tab. 4. First, we explore the effect of different normal-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.742 + ], + "angle": 0, + "content": "izations on model performance. Using batch normalization (BN) and layer normalization (LN) does not bring gains for text-to-motion tasks. In contrast, AdaLN, which integrates text information into the extraction process of the local motion pattern, proves to be the optimal normalization method. Additionally, we examine how the kernel size of the convolutional layers impacts the model's performance and select the setting of \\( k = 3,1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To further demonstrate the effectiveness of LPA, we visualize the spectrum of TIMotion w/o and w/ LPA in Fig. 6, respectively. The magnitude has been normalized. As can be seen in Fig. 6, adding LPA reduces the high-frequency component of the feature, making the motion less likely to change drastically and making it smoother. Moreover, we randomly sampled 200 instances, and w/ LPA, the average proportion of the amplitude of high-frequency components is 0.3729, whereas w/o LPA, the average proportion of the amplitude of high-frequency components is 0.9063." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7175" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.092, + 0.865, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.238, + 0.905, + 0.265 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on the motion in-between task. The first and last frames are fixed. Darker colors indicate later frames. Our method achieves smooth and natural transitions between the conditioned motions." + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.291, + 0.436, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.571, + 0.483, + 0.641 + ], + "angle": 0, + "content": "Figure 6. Spectrum of motion features. (a) and (b) show the spectrum of TIMotion w/o and w/ LPA, respectively. The horizontal axis denotes the frequency and the vertical axis represents the normalized magnitude. TIMotion w/ LPA contains fewer high-frequency components and therefore generates smoother motion." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.667, + 0.46, + 0.683 + ], + "angle": 0, + "content": "4.5. Computational Complexity and Editability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.689, + 0.331, + 0.705 + ], + "angle": 0, + "content": "4.5.1. Computational Complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.484, + 0.829 + ], + "angle": 0, + "content": "In Tab. 5, we compare our approach TIMotion with the SOTA method InterGen [16] in terms of computational complexity. TIMotion requires fewer parameters and FLOPs than InterGen but outperforms it on the comprehensive metric FID and R precision. Notably, using a similar transformer architecture to InterGen, TIMotion's average inference time per sample is only 0.632 s while InterGen requires 1.991 s." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.837, + 0.216, + 0.852 + ], + "angle": 0, + "content": "4.5.2. Editability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Tab. 6 illustrates that TIMotion surpasses InterGen in the task of motion in-between editing. We perform experiments on the test set of InterHuman and evaluate by generat" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.29, + 0.888, + 0.345 + ], + "angle": 0, + "content": "
MethodsR Precision Top 1↑FID ↓Params (M)Flops (G)Inference time (s)
InterGen (Transformer)0.371±.0105.918±.07918280.51.991
TIMotion+Transformer0.491±.0055.433±.0806540.40.632
TIMotion+RWKV0.501±.0054.702±.06912758.01.733
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.355, + 0.869, + 0.37 + ], + "angle": 0, + "content": "Table 5. Comparison of computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.401, + 0.907, + 0.461 + ], + "angle": 0, + "content": "ing \\(80\\%\\) of the sequences based on the first and last \\(10\\%\\) of the sequences. The quantitative results are shown in Fig. 5 and our method achieves smooth and natural transitions between the conditioned motions." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.476, + 0.888, + 0.535 + ], + "angle": 0, + "content": "
MethodsR Precision Top 1↑FID ↓MM Dist↓Diversity→
InterGen0.461±.0064.700±.0663.780±.0017.682±.029
TIMotion0.516±.0063.590±.0493.760±.0017.795±.031
" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.545, + 0.881, + 0.56 + ], + "angle": 0, + "content": "Table 6. Evaluation of motion in-between editing task." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.602, + 0.634, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we abstract the overall human-human motion generation process into a general framework MetaMotion, which consists of two phases: temporal modeling and interaction mixing. We find that the current methods lead to sub-optimal results and redundancy of model parameters due to inadequate modeling. Based on this, we introduce TIMotion (Temporal and Interactive Modeling), an efficient and effective approach for human-human motion generation. Specifically, we first propose Causal Interactive Injection to model two separate sequences as a causal sequence leveraging the temporal and causal properties. Additionally, we proposed Role-Evolving Mixing to adapt to the dynamic roles throughout interactions and designed Localized Pattern Amplification to capture short-term motion patterns for generating smoother and more rational motion. Extensive experiments on the InterHuman and Inter-X datasets demonstrate that TIMotion significantly outperforms existing methods, achieving state-of-the-art results." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7176" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[3] Erik Bethke. Game development and production. Wordware Publishing, Inc., 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[4] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.483, + 0.429 + ], + "angle": 0, + "content": "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[6] Ke Fan, Junshu Tang, Weijian Cao, Ran Yi, Moran Li, Jingyu Gong, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Freemotion: A unified framework for number-free text-to-motion synthesis. In European Conference on Computer Vision, pages 93-109. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[7] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. Tm2d: Bimodality driven 3d dance generation via music-text integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9942-9952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[9] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[11] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In ACM SIGGRAPH 2022 conference proceedings, pages 1-9, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.907, + 0.293 + ], + "angle": 0, + "content": "[16] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.905, + 0.323 + ], + "angle": 0, + "content": "[17] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.326, + 0.905, + 0.367 + ], + "angle": 0, + "content": "[18] Nadia Magnenat-Thalmann, Daniel Thalmann, Nadia Magnenat-Thalmann, and Daniel Thalmann. Computer animation. Springer, 1985. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.369, + 0.905, + 0.423 + ], + "angle": 0, + "content": "[19] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.904, + 0.454 + ], + "angle": 0, + "content": "[20] Rick Parent. Computer animation: algorithms and techniques. Newnes, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[21] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.528, + 0.905, + 0.583 + ], + "angle": 0, + "content": "[22] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.585, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[23] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.643, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[24] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1546-1555, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.701, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.904, + 0.813 + ], + "angle": 0, + "content": "[26] George Saridis. Intelligent robotic control. IEEE Transactions on Automatic Control, 28(5):547-557, 1983. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[27] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7177" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[29] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[30] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[31] G Tevet, S Raab, B Gordon, Y Shafir, D Cohen-Or, and AH Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[32] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[33] Jay Urbain. Introduction to game development. Cell, 414: 745-5102, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.483, + 0.363 + ], + "angle": 0, + "content": "[34] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.365, + 0.483, + 0.42 + ], + "angle": 0, + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.422, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[36] Shuo Wang, Xinhai Zhao, Hai-Ming Xu, Zehui Chen, Dameng Yu, Jiahao Chang, Zhen Yang, and Feng Zhao. Towards domain generalization for multi-view 3d object detection in bird-eye-view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13333-13342, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.507, + 0.483, + 0.576 + ], + "angle": 0, + "content": "[37] Shuo Wang, Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd-map construction. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.578, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[38] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.621, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[39] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-\\(\\mathbf{x}\\): Towards versatile human-human interaction analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22260-22271, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[40] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[41] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[42] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[43] Chen Zhao, Weiling Cai, Chenyu Dong, and Chengwei Hu. Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8281-8291, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[44] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 509-519, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7178" + } + ] +] \ No newline at end of file diff --git a/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_origin.pdf b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ef79275d1f687dcfda7e92766ae38fe6c0d94786 --- /dev/null +++ b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/50cb4aae-7f81-4aeb-985d-8b68faf24e85_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4649c3b38780c440a3c22d55b86eb1cc8fc5ac27b84db8c786bcc5caf3f4fc0a +size 2885432 diff --git a/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/full.md b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..116e819cf01861e783abf482d468ede10c0230e7 --- /dev/null +++ b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/full.md @@ -0,0 +1,346 @@ +# TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation + +Yabiao Wang $^{1,2*}$ ; Shuo Wang $^{2*}$ ; Jiangning Zhang $^{2}$ , Ke Fan $^{3}$ , + +Jiafu Wu², Zhucun Xue¹, Yong Lii¹ + +$^{1}$ Zhejiang University $^{2}$ Youtu Lab, Tencent $^{3}$ Shanghai Jiao Tong University + +https://aigc-explorer.github.io/TIMotion-page/ + +# Abstract + +Human-human motion generation is essential for understanding humans as social beings. Current methods fall into two main categories: single-person-based methods and separate modeling-based methods. To delve into this field, we abstract the overall generation process into a general framework MetaMotion, which consists of two phases: temporal modeling and interaction mixing. For temporal modeling, the single-person-based methods concatenate two people into a single one directly, while the separate modeling-based methods skip the modeling of interaction sequences. The inadequate modeling described above resulted in sub-optimal performance and redundant model parameters. In this paper, we introduce TIMotion (Temporal and Interactive Modeling), an efficient and effective framework for human-human motion generation. Specifically, we first propose Causal Interactive Injection to model two separate sequences as a causal sequence leveraging the temporal and causal properties. Then we present Role-Evolving Scanning to adjust to the change in the active and passive roles throughout the interaction. Finally, to generate smoother and more rational motion, we design Localized Pattern Amplification to capture short-term motion patterns. Extensive experiments on InterHuman and InterX demonstrate that our method achieves superior performance. + +# 1. Introduction + +In the field of generative computer vision, human motion generation has significant implications for computer animation [18, 20], game development [3, 33], and robotic control [26, 36, 37]. In recent years, there have been remarkable advancements in human motion generation, driven by various user-specified conditions such as action cat + +egories [9, 22], speeches [2, 12], and natural language prompts [11, 24]. Among these, many approaches leveraging large language models [1] and diffusion models [43] have yielded impressive results in generating realistic and diverse motions, benefiting from their powerful modeling capabilities. Despite this progress, most existing methods are designed primarily for single-person scenarios, thereby neglecting a crucial element of human motion: the complex and dynamic interactions between individuals. We address the challenge of generating two-person motion by fully utilizing the temporal and interactive dynamics between the individuals. + +To better explore human-human motion generation, we abstract a general framework MetaMotion, as shown on the left of Fig. 1, which consists of two phases: temporal modeling and interaction mixing. Previous approaches have prioritized interaction mixing over temporal modeling and can be divided into two primary categories: single-person-based methods and separate modeling-based methods. As shown in Fig. 1(a), the single-person-based methods (e.g. MDM [32]) concatenate two individuals into one, which is then fed into the existing single-person motion generation module, e.g. DiT. And the separate modeling-based methods (e.g. InterGen [16]), as shown in Fig. 1(b), model two individuals individually and then extract motion information from both themselves and each other, using self-attention and cross-attention mechanisms, respectively. Following the general logic of MetaMotion, we introduce the Temporal and Interactive Framework, as depicted in Fig. 1(c), which models the human-human causal interactions. In fact, this effective temporal modeling method can simplify the design of the interaction mixing module and reduce the number of learnable parameters. + +In this paper, we design an effective temporal modeling method and introduce TIMotion, a temporal and interactive framework for efficient human-human motion generation. Specifically, we first propose Causal Interactive Injection to model the two single-person motion sequences as a causal interaction sequence, according to the temporal causal prop + +![](images/65d1e8c2bf254674ac4334cb4e4fde95ace27ca0138c7dafb08f435d1208c839.jpg) +Figure 1. MetaMotion and performance of MetaMotion-based models on InterHuman validation set. We abstract the MetaMotion concept that illustrates the intrinsic properties of human-human motion generation in the interaction process. (a) and (b) show the two types of methods currently, and (c) shows our method TIMotion, LPA refers to the Localized Pattern Amplification. In (d) we compare the performance of the different methods on the InterHuman dataset. + +![](images/3edd481d018504674e702784255bbbc37ba3c4f3189d61dc25f5263210cbdc10.jpg) + +![](images/bfa135e68db0c805d3881674d479a4765e50034e4134544eae71aacc54e32b93.jpg) + +![](images/ddcd6e045956c034cc60cd9cfff99c25241ed92e6ff86d54bace07c42c929f1a.jpg) + +![](images/17f2bbbbf96e6814896f9996f6cdc15d7c91ba2a66c2180533536671203201a7.jpg) + +erties of motion sequences. Second, since active and passive roles are not static in interactions, we propose Role-Evolving Mixing to make two humans act in both active and passive roles. Then the network can dynamically adjust the roles of the two humans based on the text's semantics and the motion's context. Finally, we propose Localized Pattern Amplification, which captures short-term motion patterns for each individual separately, resulting in smoother and more logical motion generation. Our proposed framework can be well adapted to different interaction-mixing modules, including Transformer, RWKV, and Mamba. + +Our contributions can be summarized as follows: + +- We conceptualize human-human motion generation within a general framework MetaMotion and design an innovative method. Our proposed framework, TIMotion, is versatile enough to integrate with various interaction-mixing modules (e.g. Transformer, RWKV, Mamba) and reduces the number of parameters of these modules. +- To utilize the temporal and causal properties, we propose Causal Interactive Injection to model two separate motion sequences as a unified causal sequence. In addition, we introduce Role-Evolving Scanning to accommodate shifts between the active and passive roles during interactions. Moreover, we also design Localized Pattern Amplification to capture short-term motion patterns effectively. +- We perform extensive experiments on the benchmark human-human motion generation datasets: InterHuman and Inter-X. The results demonstrate the effectiveness and generalizability of our proposed methods. + +# 2. Related Work + +Single-Person Human Motion Generation. Creating human motion is vital for applications such as 3D modeling and robot manipulation. The primary approach, known as the Text-to-Motion task, involves learning a unified latent space for both language and motion. + +Autoencoders have been widely adopted in human motion generation. MotionCLIP [30] effectively integrates semantic knowledge from CLIP [25] into the human motion manifold. TEMOS [23] and T2M [10] combine a Transformer-based VAE with a text encoder to generate motion sequences based on text descriptions. AttT2M [44] and TM2D [7] integrate a spatial-temporal body-part encoder into VQ-VAE [34] to improve the learning of discrete latent space. T2M-GPT [40] redefines text-driven motion generation as a next-index prediction task and proposes a framework based on VQ-VAE and Generative Pretrained Transformer (GPT) for motion generation. + +Recently, diffusion-based generative modeling has been gaining attention. MotionDiffuse [42] exhibits several desirable properties, such as probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [31] aims to predict motion directly and incorporates a geometric loss to boost the model's performance. Instead of employing a diffusion model to link raw motion sequences with conditional inputs, MLD [5] further utilizes the latent diffusion model to reduce the training and inference costs substantially. ReMoDiffuse [41] introduces an improvement mechanism based on dataset retrieval to enhance the denoising process of Diffusion. + +Multi-Person Human Motion Generation. As multiperson motion synthesis involves the interactive dynamics between multiple individuals, it is more challenging than single-person motion generation. Early work typically relied on motion graphs and momentum-based inverse kinematics to model human joints. Recently, ComMDM [27] finetunes a pre-trained text-to-motion diffusion model using a small scale of two human motions. RIG [29] converts the text of asymmetric interactions into both active and passive voice to maintain consistent textual context for each individual. To complete the interaction process by + +![](images/ed7e7f476381049f9a6319e9729cb8b989ff3e56c379faa7a78c6932a30a3d6f.jpg) +Figure 2. The overall framework of our TIMotion. We contribute three primary technical designs. First, we propose Causal Interactive Injection to utilize the temporal properties of motion sequences. Then we present Role-Evolving Mixing to adjust to the ever-evolving roles during interaction. Finally, we design Localized Pattern Amplification to capture short-term motion patterns. + +manipulating the joints of two individuals to specific positions, InterControl [38] employs a Large Language Model to generate movement plans for the two individuals by designing prompts. InterGen [16] introduces a large-scale, text-annotated two-person motion dataset. Building on this dataset, it proposes a diffusion model with shared weights and multiple regularization losses. FreeMotion [6] proposes to decouple the process of conditional motion generation and support the number-free motion synthesis. + +# 3. Method + +# 3.1. Motion Diffusion Model + +Current methods usually use diffusion models [15] for motion generation. For the Interhuman dataset, a noncanonical representation [16] is typically used. The representation is formulated as: $x^{i} = [\mathbf{j}_{g}^{p},\mathbf{j}_{g}^{v},\mathbf{j}^{r},\mathbf{c}^{f}]$ , where the $i$ -th motion state $x^{i}$ is defined as a collection of global joint positions $\mathbf{j}_g^p\in \mathbb{R}^{3N_j}$ , velocities $\mathbf{j}_g^v\in \mathbb{R}^{3N_j}$ in the world frame, 6D representation of local rotations $\mathbf{j}^r\in \mathbb{R}^{6N_j}$ in the root frame, and binary foot-ground contact $\mathbf{c}^f\in \mathbb{R}^4$ . Our goal is to train a model parameterized by $\theta$ to approximate the human interactive motion data distribution $p(\mathbf{x_0})$ . + +Diffusion Process. Following a specified schedule $\beta_{t} \in (0,1)$ , models incrementally degrade input data $\mathbf{x}_0 \sim p(\mathbf{x}_0)$ , ultimately transforming the data distribution into an isotropic Gaussian over $T$ steps. Each diffusion transition can be considered as + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {x} _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where the full diffusion process can be written as + +$$ +q \left(\mathbf {x} _ {1: T} \mid \mathbf {x} _ {0}\right) = \prod_ {1 \leq t \leq T} q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right). \tag {2} +$$ + +Denoising Process. In the denoising process, models are trained to reverse the diffusion procedure, enabling them to transform random noise into real data distribution during inference. The denoising process can be written as + +$$ +\begin{array}{l} p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right) = \mathcal {N} \left(\mathbf {x} _ {t - 1}; \mu_ {\theta} \left(\mathbf {x} _ {t}, t\right), \Sigma_ {\theta} \left(\mathbf {x} _ {t}, t\right)\right) \\ = \mathcal {N} \left(\mathbf {x} _ {t - 1}; \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {x} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon\right), \frac {1 - \bar {\alpha} _ {t - 1}}{1 - \bar {\alpha} _ {t}} \beta_ {t}\right), \tag {3} \\ \end{array} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , $\alpha_{t} = 1 - \beta_{t}$ , $\bar{\alpha}_{t} = \prod_{i = 1}^{t}\alpha_{i}$ and $\theta$ denotes parameters of the network learning to denoise. The training objective is to maximize the likelihood of observed data $p_{\theta}(\mathbf{x}_0) = \int p_{\theta}(\mathbf{x}_{0:T})d\mathbf{x}_{1:T}$ , by maximizing its evidence lower bound (ELBO), which effectively aligns the true denoising model $q(\mathbf{x}_{t - 1}\mid \mathbf{x}_t)$ with the parameterized $p_{\theta}(\mathbf{x}_{t - 1}\mid \mathbf{x}_t)$ . During training, the goal of the denoising network $\epsilon_{\theta}(.)$ is to reconstruct $\mathbf{x}_0$ given any noised input $\mathbf{x}_t$ , by predicting the added noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ via minimizing the noise prediction error + +$$ +\mathcal {L} _ {t} = \mathbb {E} _ {\mathbf {x} _ {0}, \epsilon} \left[ \left\| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t\right) \right\| ^ {2} \right]. (4) +$$ + +To condition the model on additional context information $\mathbf{c}$ , e.g., text, we inject $\mathbf{c}$ into $\epsilon_{\theta}(.)$ by replacing $\mu_{\theta}(\mathbf{x}_t,t)$ and $\Sigma_{\theta}(\mathbf{x}_t,t)$ with $\mu_{\theta}(\mathbf{x}_t,t,\mathbf{c})$ and $\Sigma_{\theta}(\mathbf{x}_t,t,\mathbf{c})$ . + +# 3.2. MetaMotion + +We present the core concept "MetaMotion" for human-human motion generation at first. As shown in Fig. 1, + +we abstract the process of human-human motion generation into two phases: temporal modeling and interaction mixing. + +Specifically, two single-person sequences go through the temporal modeling module to get the input sequence $X$ . Then, the input sequence is fed to the interaction-mixing module and this process can be expressed as + +$$ +Y = \operatorname {I n t e r a c t i o n M i x i n g} (X), \tag {5} +$$ + +where InteractionMixing is usually the transformer structure including self-attention and cross-attention. Notably, InteractionMixing can also be some emerging structures, e.g., Mamba [8], RWKV [21]. + +# 3.3. TIMotion + +For human-human motion generation, most current methods are based on extensions of single-person motion generation. Some design-specific approach (e.g. InterGen [16]) considers the interaction between two persons in the interaction mixing module. However, current methods ignore the importance of temporal modeling, resulting in suboptimal performance in generating long-sequence motions and multi-person interactions. + +In this work, we propose TIMotion, as illustrated in Fig. 2, an effective temporal modeling approach that can be applied to different interaction-mixing structures, enabling better handling the human-human interactions. TIMotion consists of three key technical designs: (1) Causal Interactive Injection, which identifies the causal relationships between human-human interactions; (2) Role-Evolving Scanning, which adapts to shifts between active and passive roles during interactions; (3) Localized Pattern Amplification, designed to handle short-term sequences better. + +# 3.3.1. Causal Interactive Injection + +Perception of ego-motion and interaction between two persons are two important elements of human-human motion generation. Considering the causal properties of motion, we propose Causal Interactive Injection, a temporal modeling approach to simultaneously achieve both perceptions of ego-motion and interactions with each other. + +Specifically, we denote two single-person motion sequences as $\{x_{a}, x_{b}\}$ , where $x_{a} = \{x_{a}^{j}\}_{j=1}^{L}$ and $x_{b} = \{x_{b}^{j}\}_{j=1}^{L}$ are respective sequences of motion, and $L$ is the length of the sequence. Since the motions of two individuals at the current time step are jointly determined by their motions at previous time steps, we model the two single-person motion sequences as a causal interaction sequence $x_{cii} = \{x_{k}^{j/2}\}_{j=1}^{2L}$ , the symbol // denotes division followed by rounding up and $k$ can be acquired as follows: + +$$ +k = \left\{ \begin{array}{ll}a, & j\% 2 = 1.\\ b, & j\% 2 = 0. \end{array} \right. \tag{6} +$$ + +Then we can inject them into the interaction-mixing module and separate the motion embeddings of the two individuals from the outputs according to Eq. (6). + +![](images/71ab8ac14325f66d143f13be212b975d25864f6c50b3729e69f7a1131078f7cb.jpg) +frame 1 + +![](images/11df8a091534d89a05f9cb811d89f472c2bbac40da657fc1570fdb8a58579d31.jpg) +frame 2 + +![](images/9b954790ee49f4cf025bb6d156b0637f7cb4fe8b9632dd2b45299d4ba358eef3.jpg) +frame 3 + +frame 4 +![](images/17ae2a3f4f8a4aa1cca1d86633b47b7de07a43f1f2917c4a4fb1e34175ac2a3a.jpg) +The first person suddenly pushes the other one. The second person takes a step back, then forcefully pushes the first person. The first person was knocked down. + +![](images/044b4c106d984b6aa885e10c9fbed0fbb8d7cac7983c0784599d5287a6a4e7d8.jpg) +frame 5 + +![](images/00a1a7dfbfafc56057fad7854d5b9304b4e4c42f4c648a79c5fade3f01befab9.jpg) +frame 6 +Figure 3. Illustration of changing active and passive roles. The first person acts as the active role in the early stages, and as time progresses, the other person becomes the active role of the motion. + +# 3.3.2. Role-Evolving Scanning + +As noted in [4], human interactions inherently follow a certain order, e.g., "shaking hands" typically begins with one person extending a hand first, which means that interactive motion can be categorized as active or passive. Some methods [29] split the text description into active and passive voices. However, the "active" and "passive" roles are not static in a text description. As shown in Fig. 3, these roles constantly swap between characters as the interaction progresses. To minimize redundant text preprocessing and adapt to the ongoing shifts in roles, we design Role-Evolving Scanning, which is both efficient and effective. + +For the causal interaction sequence $x$ defined in Causal Interactive Injection (Sec. 3.3.1), it is obvious that $x_{a}$ and $x_{b}$ represent the active sequence and passive sequence, respectively. However, the above assumption of active and passive motions is not always in line with the real order. To cope with the changing roles, we re-model the interactive motion sequences as a symmetric causal interaction sequence $x_{sym\_ cii} = \{x_{k'}^{j // 2}\}_{j=1}^{2L}$ , where the symbol $//$ is defined as division followed by rounding up and $k'$ is obtained by exchanging $a$ and $b$ in Eq. (6). + +Given that the causal interaction sequence $x_{cii} \in \mathbb{R}^{2L \times C}$ and the symmetric causal interaction sequence $x_{sym\_ cii} \in \mathbb{R}^{2L \times C}$ , where $L$ is the length of the sequence and $C$ is the dimension of motion embedding, we can acquire the final interaction sequence $X \in \mathbb{R}^{2L \times 2C}$ through Role-Evolving Scanning as: + +$$ +X = \operatorname {C o n c a t} \left(x _ {c i i}, x _ {s y m. c i i}\right). \tag {7} +$$ + +Then the sequence $X$ is fed to the interaction-mixing module. After obtaining the output $Y \in \mathbb{R}^{2L \times 2C}$ through Eq. (5), we first divide it at the channel level into causal interaction embeddings and symmetric causal interaction embeddings. According to $x_{cii}$ and $x_{sym\_ cii}$ , we can obtain the two individuals' splitting embeddings twice. Then we merge the embeddings of the two individuals to obtain the final global motion embeddings $y_a^g \in \mathbb{R}^{L \times C}$ and $y_b^g \in \mathbb{R}^{L \times C}$ . The overall process is as follows: + +$$ +\begin{array}{l} y _ {a 1}, y _ {b 1} = \operatorname {S p l i t} (Y [ :, : C ]), \\ y _ {b 2}, y _ {a 2} = \operatorname {S p l i t} (Y [:, C: ]), \tag {8} \\ y _ {a} ^ {g} = y _ {a 1} \oplus y _ {a 2}, \\ y _ {b} ^ {g} = y _ {b 1} \oplus y _ {b 2}, \\ \end{array} +$$ + +where $\oplus$ denotes element-wise sum. + +By utilizing Role-Evolving Mixing to make two humans act as both active and passive roles, the network can dynamically adjust the roles of the two humans based on the semantics of the text and the context of the motion. + +# 3.3.3. Localized Pattern Amplification + +Transformers and RNNs excel at global modeling and capturing long-range dependencies but tend to overlook local semantic information [13]. Additionally, Causal Interactive Injection and Role-Evolving Mixing mainly model the overall motion based on the causality of the interactions, but neglect to focus on the localized motion patterns of the single person. To address this issue, we propose Localized Pattern Amplification, which summarizes short-term motion patterns for each person individually while generating smoother and more rational motion. + +Specifically, we utilize 1-D convolution layers and the residual structure to realize Localized Pattern Amplification. Given that the condition embedding $e$ and two single-person motion sequences $x_{a}$ and $x_{b}$ , the local motion embedding for $x_{a}$ can be expressed as: + +$$ +\begin{array}{l} x _ {a} ^ {l} = \operatorname {C o n v} _ {3} (\operatorname {A d a L N} (x _ {a}, e)), \\ y _ {a} = \operatorname {C o n v} _ {1} \left(\operatorname {A d a L N} \left(x _ {a} ^ {l}, e\right)\right), \tag {9} \\ y _ {a} ^ {l} = x _ {a} + y _ {a}, \\ \end{array} +$$ + +where $\mathrm{Conv_k}$ denotes the 1-D convolution with kernel size $k$ and AdaLN is the adaptive layer normalization used in [42]. The process for $x_{b}$ is the same as for $x_{a}$ and they share network weights. + +Once we obtain the outputs $\{y_a^g, y_b^g\}$ through Eq. (8) and the output $\{y_a^l, y_b^l\}$ from the convolution block, global embeddings and local embeddings are aggregated through concatenation along the channel dimension, followed by a linear layer to restore the original channels, resulting in the + +final outputs $\{y_a^{final}, y_b^{final}\}$ : + +$$ +\begin{array}{l} y _ {a} ^ {f i n a l} = \operatorname {L i n e a r} (\operatorname {C o n c a t} (y _ {a} ^ {g}, y _ {a} ^ {l})), \\ y _ {b} ^ {\text {f i n a l}} = \operatorname {L i n e a r} \left(\operatorname {C o n c a t} \left(y _ {b} ^ {g}, y _ {b} ^ {l}\right)\right). \tag {10} \\ \end{array} +$$ + +Then the final outputs are fed into the next encoder block or the final decoder layer. + +# 3.3.4. Objective Function + +We use the same loss functions as InterGen [16], including the diffusion loss $\mathcal{L}_{simple}$ , the foot contact loss $\mathcal{L}_{foot}$ , joint velocity loss $\mathcal{L}_{vel}$ , the bone length loss $\mathcal{L}_{BL}$ , the masked joint distance map loss $\mathcal{L}_{DM}$ , and the relative orientation loss $\mathcal{L}_{RO}$ . For more details about the losses, we refer readers to InterGen. Finally, the overall loss is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {m o t i o n}} = \mathcal {L} _ {\text {s i m p l e}} + \lambda_ {v e l} \mathcal {L} _ {v e l} + \lambda_ {f o o t} \mathcal {L} _ {f o o t} \\ + \lambda_ {B L} \mathcal {L} _ {B L} + \lambda_ {D M} \mathcal {L} _ {D M} + \lambda_ {R O} \mathcal {L} _ {R O}, \tag {11} \\ \end{array} +$$ + +where the hyper-parameters $\lambda_{vel},\lambda_{foot},\lambda_{BL},\lambda_{DM},\lambda_{RO}$ are the same as InterGen. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We assess our proposed framework using the InterHuman [16] and Inter-X [39] dataset. InterHuman is the first dataset to incorporate text annotations for two-person motions. This dataset includes 6,022 motions spanning various categories of human actions and is labeled with 16,756 unique descriptions made up of 5,656 distinct words. InterX is the largest human-human interaction dataset with diverse interaction patterns. It includes about 11k interaction sequences and more than 8.1M frames. + +Metrics. We employ the same evaluation metrics as InterGen [16], which are as follows: (1) Frechet Inception Distance (FID). (2) $R$ -Precision. (3) Diversity. (4) Multimodality (MModality). (5) Multi-modal distance (MM Dist). Detailed explanations and calculations of the metrics can be found in the supplementary materials. + +# 4.1.1. Implementation Details. + +We use a frozen CLIP-ViT-L/14 model as the text encoder. The dimension of motion embedding is set to 512. During training, the number of diffusion timesteps is set to 1000, and we employ the DDIM [28] sampling strategy with 50 timesteps and $\eta = 0$ . The cosine noise level schedule [19] and classifier-free guidance [14] are adopted, with $10\%$ of random CLIP embeddings set to zero during training and a guidance coefficient of 3.5 during sampling. All the models are trained using the AdamW [17] optimizer with betas of (0.9, 0.999), a weight decay of $2 \times 10^{-5}$ , a maximum learning rate of $10^{-4}$ , and a cosine learning rate schedule with 10 linear warm-up epochs. To balance the contributions of different loss terms, we set $\lambda_{vel} = 30$ , $\lambda_{foot} = 30$ , $\lambda_{BL} = 10$ , + +
MethodsR Precision↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.452±.0080.610±.0090.701±.0080.273±.0073.755±.0087.948±.064-
TEMOS [23]0.224±.0100.316±.0130.450±.01817.375±.0435.342±.0156.939±.0710.535±.014
T2M [10]0.238±.0120.325±.0100.464±.01413.769±.0724.731±.0137.046±.0221.387±.076
MDM [31]0.153±.0120.260±.0090.339±.0129.167±.0566.125±.0187.602±.0452.355±.080
ComMDM* [27]0.067±.0130.125±.0180.184±.01538.643±.09813.211±.0133.520±.0580.217±.018
ComMDM [27]0.223±.0090.334±.0080.466±.0107.069±.0545.212±.0217.244±.0381.822±.052
RIG [29]0.285±.0100.409±.0140.521±.0136.775±.0694.876±.0187.311±.0432.096±.065
InterGen [16]0.371±.0100.515±.0120.624±.0105.918±.0795.108±.0147.387±.0292.141±.063
TIMotion+Transformer(ours)0.491±.0050.648±.0040.724±.0045.433±.0803.775±.0018.032±.0300.952±.032
TIMotion+Mamba(ours)0.431±.0040.586±.0050.668±.0046.142±.0593.800±.0017.793±.0240.837±.036
TIMotion+RWKV(ours)0.501±.0050.656±.0060.734±.0064.702±.0693.769±.0017.943±.0341.005±.020
+ +Table 1. Quantitative evaluation on the InterHuman [16] test set. We run all the evaluations 20 times. ± indicates a 95% confidence interval. Bold indicates the best result, while underline refers to the second best. ComMDM* indicates the ComMDM model fine-tuned in the original few-shot setting with 10 training samples and ComMDM indicates fine-tuned on the entire InterHuman training set. + +
Interaction MixingTemporal ModelingR Precision↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Transformer [35]single-person extension0.395±.0040.535±.0040.615±.0048.028±.0993.820±.0017.687±.0220.852±.024
separate modeling0.371±.0100.515±.0120.624±.0105.918±.0795.108±.0147.387±.0292.141±.063
TIMotion (Ours)0.491±.0050.648±.0040.724±.0045.433±.0803.775±.0018.032±.0300.952±.032
Mamba [8]single-person extension0.368±.0050.513±.0050.597±.0067.232±.0813.825±.0017.864±.0240.870±.020
separate modeling0.420±.0050.569±.0040.650±.0037.221±.0973.803±.0017.932±.0350.855±.020
TIMotion (Ours)0.431±.0040.586±.0050.668±.0046.142±.0593.800±.0017.793±.0240.837±.036
RWKV [21]single-person extension0.425±.0040.576±.0040.656±.0039.181±.0963.801±.0017.679±.0220.846±.026
separate modeling0.465±.0050.603±.0050.689±.0045.943±.0793.790±.0017.787±.0320.859±.036
TIMotion (Ours)0.501±.0050.656±.0060.734±.0064.702±.0693.769±.0017.943±.0341.005±.020
+ +Table 2. Comparison of different temporal modeling approaches on different interaction mixing structures. Our proposed causal interactive modeling is able to adapt to different interaction mixing architectures and outperforms the other two ways. + +$\lambda_{DM} = 3, \lambda_{RO} = 0.01$ and $\lambda_{reg} = 1$ in all experiments, as same as InterGen. We train our diffusion denoisers with a batch size of 256 for 1500 epochs on 8 Nvidia L40S GPUs. + +# 4.2. Comparisons with State-of-the-arts + +# 4.2.1. Quantitative Results. + +Following established practices [16], each experiment is conducted 20 times, and the reported metric values represent the mean with a $95\%$ statistical confidence interval. The results on InterHuman are shown in Tab. 1. In comparison to existing methods, our TIMotion achieves competitive results on three different interaction mixing architectures: transformer, mamba, and RWKV. Equipped with RWKV, TIMotion achieves 4.702 FID and 0.501 Top1 R precision, setting new state-of-the-art (SoTA) for the competitive InterHuman benchmark. The results on Inter-X [39] can be found in the supplementary materials. + +# 4.2.2. Qualitative Comparisons. + +In Fig. 4, we qualitatively compare the InterGen and our TIMotion. It can be seen that sequences generated by TIMotion are more consistent with the description. + +# 4.3. Effectiveness of TIMotion + +To validate the effectiveness of our proposed causal interactive modeling (TIMotion), we compare different temporal modeling approaches on three different structures, respectively. As shown in Tab. 2, single-person extension performs poorly, indicating that it does not model two-person interactions well. In contrast, separate modeling leverages the ability of the interaction mixing module to model the two-person interactions but achieves sub-optimal performance on the Transformer and mamba architectures. Our proposed approach is able to adapt to different interaction mixing architectures and outperforms the other two approaches. + +# 4.4. Ablation Study and Analysis + +In this section we conduct comprehensive ablation studies to investigate the effectiveness of the key components in TIMotion, thereby providing a deeper insight into our approach. All experiments are performed on the InterHuman dataset. Unless otherwise stated, we use RWKV as the interaction mixing structure for the following ablations. More theoretical analyses are in the supplementary materials. + +![](images/32f69a0128b9b6a9993a070ccf80c9be2b31ef8e8fab21a13d4d1b9d72f5ba4e.jpg) +[1] two individuals are lifting and transporting an object. + +![](images/753a0bc2dbc51e6d9e0121a88cd27339c507bb7d5a0b897a290ec09baa4a2580.jpg) +[2] the other person pushes one person to the ground with both hands. + +![](images/cedfa3d746622368f0209122e2c7d9f47c33b5f692f50ca0dd848f8910ef4044.jpg) +[3] one wraps both arms around the back of the other person, while the other person reciprocates the action +Figure 4. Qualitative comparison with Intergen on human-human motion generation. Darker color indicates later frames. The sequences generated by TIMotion are more consistent with the text description. + +
CIIRESLPAR Precision Top 1↑FID ↓Params (M)
0.371±.0105.918±.079182
✓0.478±.0055.410±.069144
✓✓0.494±.0065.019±.077115
✓✓✓0.501±.0054.702±.069127
+ +Table 3. Ablation studies on the effectiveness of each component in TIMotion. "CII" denotes Causal Interactive Injection, "RES" denotes Role-Evolving Scanning, and "LPA" denotes Localized Pattern Amplification. + +
Kernel SizeNormR Precision Top 1↑FID ↓
k=3,1BN0.494±.0055.443±.093
k=3,1LN0.488±.0055.339±.080
k=3,1AdaLN0.501±.0054.702±.069
k=3,3AdaLN0.484±.0057.120±.090
k=5,1AdaLN0.497±.0055.102±.064
+ +Table 4. Ablation studies on LPA. "BN" denotes batch normalization, "LN" denotes layer normalization and "AdaLN" denotes adaptive layer normalization. "k=3,1" means that the first kernel size of the convolution is 3 and the second kernel size is 1. + +# 4.4.1. Main Ablations + +To understand the contribution of each component to the final performance, we incrementally add the proposed modules based on the baseline InterGen [16] and present the results in Tab. 3. Initially, we replace self-attention and cross-attention in the transformer with Causal Interactive Injection (CII), which effectively improves both R-Precision and FID while reducing the number of parameters. Next, we apply Role-Evolving Scanning (RES) and yield gains in both R-Precision and FID. It is worth noting that to keep the input feature dimension of RWKV constant, we reduce the dimension of the motion embedding to half of the original, thus also effectively reducing the number of parameters. Finally, when all three methods are applied together, the R-Precision achieves 0.501 and FID achieves 4.702. + +# 4.4.2. Design of Localized Pattern Amplification. + +In this section, we conduct ablations on the design of Localized Pattern Amplification (LPA), and the results are shown in Tab. 4. First, we explore the effect of different normal- + +izations on model performance. Using batch normalization (BN) and layer normalization (LN) does not bring gains for text-to-motion tasks. In contrast, AdaLN, which integrates text information into the extraction process of the local motion pattern, proves to be the optimal normalization method. Additionally, we examine how the kernel size of the convolutional layers impacts the model's performance and select the setting of $k = 3,1$ . + +To further demonstrate the effectiveness of LPA, we visualize the spectrum of TIMotion w/o and w/ LPA in Fig. 6, respectively. The magnitude has been normalized. As can be seen in Fig. 6, adding LPA reduces the high-frequency component of the feature, making the motion less likely to change drastically and making it smoother. Moreover, we randomly sampled 200 instances, and w/ LPA, the average proportion of the amplitude of high-frequency components is 0.3729, whereas w/o LPA, the average proportion of the amplitude of high-frequency components is 0.9063. + +![](images/3175bf58ee1d5d1fc35fa25bd41e1d2b1066aa8d0504bd347afe7e8be989f26d.jpg) +Figure 5. Qualitative results on the motion in-between task. The first and last frames are fixed. Darker colors indicate later frames. Our method achieves smooth and natural transitions between the conditioned motions. + +![](images/dde4ba133cb0e6743496320a791b5007b3f85ee05c92f4833b169e11b14b4302.jpg) +Figure 6. Spectrum of motion features. (a) and (b) show the spectrum of TIMotion w/o and w/ LPA, respectively. The horizontal axis denotes the frequency and the vertical axis represents the normalized magnitude. TIMotion w/ LPA contains fewer high-frequency components and therefore generates smoother motion. + +# 4.5. Computational Complexity and Editability. + +# 4.5.1. Computational Complexity. + +In Tab. 5, we compare our approach TIMotion with the SOTA method InterGen [16] in terms of computational complexity. TIMotion requires fewer parameters and FLOPs than InterGen but outperforms it on the comprehensive metric FID and R precision. Notably, using a similar transformer architecture to InterGen, TIMotion's average inference time per sample is only 0.632 s while InterGen requires 1.991 s. + +# 4.5.2. Editability. + +Tab. 6 illustrates that TIMotion surpasses InterGen in the task of motion in-between editing. We perform experiments on the test set of InterHuman and evaluate by generat + +
MethodsR Precision Top 1↑FID ↓Params (M)Flops (G)Inference time (s)
InterGen (Transformer)0.371±.0105.918±.07918280.51.991
TIMotion+Transformer0.491±.0055.433±.0806540.40.632
TIMotion+RWKV0.501±.0054.702±.06912758.01.733
+ +ing $80\%$ of the sequences based on the first and last $10\%$ of the sequences. The quantitative results are shown in Fig. 5 and our method achieves smooth and natural transitions between the conditioned motions. + +Table 5. Comparison of computational complexity. + +
MethodsR Precision Top 1↑FID ↓MM Dist↓Diversity→
InterGen0.461±.0064.700±.0663.780±.0017.682±.029
TIMotion0.516±.0063.590±.0493.760±.0017.795±.031
+ +Table 6. Evaluation of motion in-between editing task. + +# 5. Conclusion + +In this paper, we abstract the overall human-human motion generation process into a general framework MetaMotion, which consists of two phases: temporal modeling and interaction mixing. We find that the current methods lead to sub-optimal results and redundancy of model parameters due to inadequate modeling. Based on this, we introduce TIMotion (Temporal and Interactive Modeling), an efficient and effective approach for human-human motion generation. Specifically, we first propose Causal Interactive Injection to model two separate sequences as a causal sequence leveraging the temporal and causal properties. Additionally, we proposed Role-Evolving Mixing to adapt to the dynamic roles throughout interactions and designed Localized Pattern Amplification to capture short-term motion patterns for generating smoother and more rational motion. Extensive experiments on the InterHuman and Inter-X datasets demonstrate that TIMotion significantly outperforms existing methods, achieving state-of-the-art results. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 1 +[3] Erik Bethke. Game development and production. Wordware Publishing, Inc., 2003. 1 +[4] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 4 +[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 2 +[6] Ke Fan, Junshu Tang, Weijian Cao, Ran Yi, Moran Li, Jingyu Gong, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Freemotion: A unified framework for number-free text-to-motion synthesis. In European Conference on Computer Vision, pages 93-109. Springer, 2024. 3 +[7] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. Tm2d: Bimodality driven 3d dance generation via music-text integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9942-9952, 2023. 2 +[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4, 6 +[9] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1 +[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 6 +[11] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 1 +[12] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In ACM SIGGRAPH 2022 conference proceedings, pages 1-9, 2022. 1 + +[13] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 5 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[16] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 1, 3, 4, 5, 6, 7, 8 +[17] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[18] Nadia Magnenat-Thalmann, Daniel Thalmann, Nadia Magnenat-Thalmann, and Daniel Thalmann. Computer animation. Springer, 1985. 1 +[19] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 5 +[20] Rick Parent. Computer animation: algorithms and techniques. Newnes, 2012. 1 +[21] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 4, 6 +[22] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1 +[23] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2, 6 +[24] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1546-1555, 2024. 1 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[26] George Saridis. Intelligent robotic control. IEEE Transactions on Automatic Control, 28(5):547-557, 1983. 1 +[27] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2, 6 +[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 + +[29] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 2, 4, 6 +[30] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2 +[31] G Tevet, S Raab, B Gordon, Y Shafir, D Cohen-Or, and AH Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 6 +[32] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1 +[33] Jay Urbain. Introduction to game development. Cell, 414: 745-5102, 2010. 1 +[34] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2 +[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 6 +[36] Shuo Wang, Xinhai Zhao, Hai-Ming Xu, Zehui Chen, Dameng Yu, Jiahao Chang, Zhen Yang, and Feng Zhao. Towards domain generalization for multi-view 3d object detection in bird-eye-view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13333-13342, 2023. 1 +[37] Shuo Wang, Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd-map construction. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 1 +[38] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 3 +[39] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter- $\mathbf{x}$ : Towards versatile human-human interaction analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22260-22271, 2024. 5, 6 +[40] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 2 +[41] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2 +[42] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion + +model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 5 +[43] Chen Zhao, Weiling Cai, Chenyu Dong, and Chengwei Hu. Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8281-8291, 2024. 1 +[44] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 509-519, 2023. 2 \ No newline at end of file diff --git a/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/images.zip b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..70e9436a6ecdc4a7c9a5ac09a9e8bd2559382972 --- /dev/null +++ b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed45c07320674a7d11819d069d48835e9687eac4809d727ebdb17104c3957028 +size 587285 diff --git a/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/layout.json b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7d2eeb12006a0aee8221a8e07eb14629aac0d195 --- /dev/null +++ b/2025/TIMotion_ Temporal and Interactive Framework for Efficient Human-Human Motion Generation/layout.json @@ -0,0 +1,9197 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 470, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 470, + 138 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 470, + 138 + ], + "type": "text", + "content": "TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "spans": [ + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "text", + "content": "Yabiao Wang" + }, + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "text", + "content": "; Shuo Wang" + }, + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "text", + "content": "; Jiangning Zhang" + }, + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Ke Fan" + }, + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 160, + 460, + 175 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 175, + 410, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 175, + 410, + 189 + ], + "spans": [ + { + "bbox": [ + 212, + 175, + 410, + 189 + ], + "type": "text", + "content": "Jiafu Wu², Zhucun Xue¹, Yong Lii¹" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 189, + 487, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 189, + 487, + 204 + ], + "spans": [ + { + "bbox": [ + 117, + 189, + 487, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 189, + 487, + 204 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 117, + 189, + 487, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 117, + 189, + 487, + 204 + ], + "type": "text", + "content": "Youtu Lab, Tencent " + }, + { + "bbox": [ + 117, + 189, + 487, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 117, + 189, + 487, + 204 + ], + "type": "text", + "content": "Shanghai Jiao Tong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 220, + 205, + 389, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 205, + 389, + 217 + ], + "spans": [ + { + "bbox": [ + 220, + 205, + 389, + 217 + ], + "type": "text", + "content": "https://aigc-explorer.github.io/TIMotion-page/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 270, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 568 + ], + "type": "text", + "content": "Human-human motion generation is essential for understanding humans as social beings. Current methods fall into two main categories: single-person-based methods and separate modeling-based methods. To delve into this field, we abstract the overall generation process into a general framework MetaMotion, which consists of two phases: temporal modeling and interaction mixing. For temporal modeling, the single-person-based methods concatenate two people into a single one directly, while the separate modeling-based methods skip the modeling of interaction sequences. The inadequate modeling described above resulted in sub-optimal performance and redundant model parameters. In this paper, we introduce TIMotion (Temporal and Interactive Modeling), an efficient and effective framework for human-human motion generation. Specifically, we first propose Causal Interactive Injection to model two separate sequences as a causal sequence leveraging the temporal and causal properties. Then we present Role-Evolving Scanning to adjust to the change in the active and passive roles throughout the interaction. Finally, to generate smoother and more rational motion, we design Localized Pattern Amplification to capture short-term motion patterns. Extensive experiments on InterHuman and InterX demonstrate that our method achieves superior performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 593, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "content": "In the field of generative computer vision, human motion generation has significant implications for computer animation [18, 20], game development [3, 33], and robotic control [26, 36, 37]. In recent years, there have been remarkable advancements in human motion generation, driven by various user-specified conditions such as action cat" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 246, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 554, + 388 + ], + "type": "text", + "content": "egories [9, 22], speeches [2, 12], and natural language prompts [11, 24]. Among these, many approaches leveraging large language models [1] and diffusion models [43] have yielded impressive results in generating realistic and diverse motions, benefiting from their powerful modeling capabilities. Despite this progress, most existing methods are designed primarily for single-person scenarios, thereby neglecting a crucial element of human motion: the complex and dynamic interactions between individuals. We address the challenge of generating two-person motion by fully utilizing the temporal and interactive dynamics between the individuals." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 390, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 640 + ], + "type": "text", + "content": "To better explore human-human motion generation, we abstract a general framework MetaMotion, as shown on the left of Fig. 1, which consists of two phases: temporal modeling and interaction mixing. Previous approaches have prioritized interaction mixing over temporal modeling and can be divided into two primary categories: single-person-based methods and separate modeling-based methods. As shown in Fig. 1(a), the single-person-based methods (e.g. MDM [32]) concatenate two individuals into one, which is then fed into the existing single-person motion generation module, e.g. DiT. And the separate modeling-based methods (e.g. InterGen [16]), as shown in Fig. 1(b), model two individuals individually and then extract motion information from both themselves and each other, using self-attention and cross-attention mechanisms, respectively. Following the general logic of MetaMotion, we introduce the Temporal and Interactive Framework, as depicted in Fig. 1(c), which models the human-human causal interactions. In fact, this effective temporal modeling method can simplify the design of the interaction mixing module and reduce the number of learnable parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we design an effective temporal modeling method and introduce TIMotion, a temporal and interactive framework for efficient human-human motion generation. Specifically, we first propose Causal Interactive Injection to model the two single-person motion sequences as a causal interaction sequence, according to the temporal causal prop" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 137, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 137, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 137, + 703 + ], + "type": "text", + "content": "*Equal contributions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7169" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 138, + 198 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 138, + 198 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 138, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 138, + 198 + ], + "type": "image", + "image_path": "65d1e8c2bf254674ac4334cb4e4fde95ace27ca0138c7dafb08f435d1208c839.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 205, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 555, + 251 + ], + "type": "text", + "content": "Figure 1. MetaMotion and performance of MetaMotion-based models on InterHuman validation set. We abstract the MetaMotion concept that illustrates the intrinsic properties of human-human motion generation in the interaction process. (a) and (b) show the two types of methods currently, and (c) shows our method TIMotion, LPA refers to the Localized Pattern Amplification. In (d) we compare the performance of the different methods on the InterHuman dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 72, + 220, + 198 + ], + "blocks": [ + { + "bbox": [ + 139, + 72, + 220, + 198 + ], + "lines": [ + { + "bbox": [ + 139, + 72, + 220, + 198 + ], + "spans": [ + { + "bbox": [ + 139, + 72, + 220, + 198 + ], + "type": "image", + "image_path": "3edd481d018504674e702784255bbbc37ba3c4f3189d61dc25f5263210cbdc10.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 310, + 198 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 310, + 198 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 310, + 198 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 310, + 198 + ], + "type": "image", + "image_path": "bfa135e68db0c805d3881674d479a4765e50034e4134544eae71aacc54e32b93.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 404, + 198 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 404, + 198 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 404, + 198 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 404, + 198 + ], + "type": "image", + "image_path": "ddcd6e045956c034cc60cd9cfff99c25241ed92e6ff86d54bace07c42c929f1a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 405, + 77, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 405, + 77, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 405, + 77, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 405, + 77, + 547, + 198 + ], + "type": "image", + "image_path": "17f2bbbbf96e6814896f9996f6cdc15d7c91ba2a66c2180533536671203201a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 294, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 294, + 402 + ], + "type": "text", + "content": "erties of motion sequences. Second, since active and passive roles are not static in interactions, we propose Role-Evolving Mixing to make two humans act in both active and passive roles. Then the network can dynamically adjust the roles of the two humans based on the text's semantics and the motion's context. Finally, we propose Localized Pattern Amplification, which captures short-term motion patterns for each individual separately, resulting in smoother and more logical motion generation. Our proposed framework can be well adapted to different interaction-mixing modules, including Transformer, RWKV, and Mamba." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 403, + 266, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 403, + 266, + 414 + ], + "spans": [ + { + "bbox": [ + 67, + 403, + 266, + 414 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 418, + 295, + 620 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 418, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 489 + ], + "type": "text", + "content": "- We conceptualize human-human motion generation within a general framework MetaMotion and design an innovative method. Our proposed framework, TIMotion, is versatile enough to integrate with various interaction-mixing modules (e.g. Transformer, RWKV, Mamba) and reduces the number of parameters of these modules." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 490, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 295, + 573 + ], + "type": "text", + "content": "- To utilize the temporal and causal properties, we propose Causal Interactive Injection to model two separate motion sequences as a unified causal sequence. In addition, we introduce Role-Evolving Scanning to accommodate shifts between the active and passive roles during interactions. Moreover, we also design Localized Pattern Amplification to capture short-term motion patterns effectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 574, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 620 + ], + "type": "text", + "content": "- We perform extensive experiments on the benchmark human-human motion generation datasets: InterHuman and Inter-X. The results demonstrate the effectiveness and generalizability of our proposed methods." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Single-Person Human Motion Generation. Creating human motion is vital for applications such as 3D modeling and robot manipulation. The primary approach, known as the Text-to-Motion task, involves learning a unified latent space for both language and motion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 270, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 415 + ], + "type": "text", + "content": "Autoencoders have been widely adopted in human motion generation. MotionCLIP [30] effectively integrates semantic knowledge from CLIP [25] into the human motion manifold. TEMOS [23] and T2M [10] combine a Transformer-based VAE with a text encoder to generate motion sequences based on text descriptions. AttT2M [44] and TM2D [7] integrate a spatial-temporal body-part encoder into VQ-VAE [34] to improve the learning of discrete latent space. T2M-GPT [40] redefines text-driven motion generation as a next-index prediction task and proposes a framework based on VQ-VAE and Generative Pretrained Transformer (GPT) for motion generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 415, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 559 + ], + "type": "text", + "content": "Recently, diffusion-based generative modeling has been gaining attention. MotionDiffuse [42] exhibits several desirable properties, such as probabilistic mapping, realistic synthesis, and multi-level manipulation. MDM [31] aims to predict motion directly and incorporates a geometric loss to boost the model's performance. Instead of employing a diffusion model to link raw motion sequences with conditional inputs, MLD [5] further utilizes the latent diffusion model to reduce the training and inference costs substantially. ReMoDiffuse [41] introduces an improvement mechanism based on dataset retrieval to enhance the denoising process of Diffusion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "Multi-Person Human Motion Generation. As multiperson motion synthesis involves the interactive dynamics between multiple individuals, it is more challenging than single-person motion generation. Early work typically relied on motion graphs and momentum-based inverse kinematics to model human joints. Recently, ComMDM [27] finetunes a pre-trained text-to-motion diffusion model using a small scale of two human motions. RIG [29] converts the text of asymmetric interactions into both active and passive voice to maintain consistent textual context for each individual. To complete the interaction process by" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7170" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 558, + 249 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 558, + 249 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 558, + 249 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 558, + 249 + ], + "type": "image", + "image_path": "ed7e7f476381049f9a6319e9729cb8b989ff3e56c379faa7a78c6932a30a3d6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 256, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 54, + 256, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 555, + 291 + ], + "type": "text", + "content": "Figure 2. The overall framework of our TIMotion. We contribute three primary technical designs. First, we propose Causal Interactive Injection to utilize the temporal properties of motion sequences. Then we present Role-Evolving Mixing to adjust to the ever-evolving roles during interaction. Finally, we design Localized Pattern Amplification to capture short-term motion patterns." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 311, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 311, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 296, + 419 + ], + "type": "text", + "content": "manipulating the joints of two individuals to specific positions, InterControl [38] employs a Large Language Model to generate movement plans for the two individuals by designing prompts. InterGen [16] introduces a large-scale, text-annotated two-person motion dataset. Building on this dataset, it proposes a diffusion model with shared weights and multiple regularization losses. FreeMotion [6] proposes to decouple the process of conditional motion generation and support the number-free motion synthesis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 443, + 111, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 111, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 111, + 455 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 191, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 191, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 191, + 479 + ], + "type": "text", + "content": "3.1. Motion Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": "Current methods usually use diffusion models [15] for motion generation. For the Interhuman dataset, a noncanonical representation [16] is typically used. The representation is formulated as: " + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "inline_equation", + "content": "x^{i} = [\\mathbf{j}_{g}^{p},\\mathbf{j}_{g}^{v},\\mathbf{j}^{r},\\mathbf{c}^{f}]" + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": "-th motion state " + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": " is defined as a collection of global joint positions " + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{j}_g^p\\in \\mathbb{R}^{3N_j}" + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": ", velocities " + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{j}_g^v\\in \\mathbb{R}^{3N_j}" + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": " in the world frame, 6D representation of local rotations " + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{j}^r\\in \\mathbb{R}^{6N_j}" + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": " in the root frame, and binary foot-ground contact " + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^f\\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": ". Our goal is to train a model parameterized by " + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": " to approximate the human interactive motion data distribution " + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x_0})" + }, + { + "bbox": [ + 55, + 489, + 296, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": "Diffusion Process. Following a specified schedule " + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\beta_{t} \\in (0,1)" + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": ", models incrementally degrade input data " + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": ", ultimately transforming the data distribution into an isotropic Gaussian over " + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 616, + 296, + 676 + ], + "type": "text", + "content": " steps. Each diffusion transition can be considered as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 691, + 296, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 691, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 83, + 691, + 296, + 712 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {x} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "5cbc969d38251229dd6441c2c77b18349e7a270d395a3c6daaf0e50f7d456ea8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 311, + 514, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 514, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 514, + 323 + ], + "type": "text", + "content": "where the full diffusion process can be written as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 358, + 332, + 555, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 332, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 358, + 332, + 555, + 360 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right) = \\prod_ {1 \\leq t \\leq T} q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right). \\tag {2}", + "image_path": "4bc124264b8c525a386abbf80c6b2f2ccaf26f1a601c25f0d342dea1c8f2615a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 368, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 415 + ], + "type": "text", + "content": "Denoising Process. In the denoising process, models are trained to reverse the diffusion procedure, enabling them to transform random noise into real data distribution during inference. The denoising process can be written as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 332, + 425, + 554, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 425, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 332, + 425, + 554, + 460 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right) \\\\ = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon\\right), \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} \\beta_ {t}\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "8a1aada92e7749222d7344119ec8d7116555f0f3ea3532fbd9e273b59d439d19.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": " denotes parameters of the network learning to denoise. The training objective is to maximize the likelihood of observed data " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0) = \\int p_{\\theta}(\\mathbf{x}_{0:T})d\\mathbf{x}_{1:T}" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": ", by maximizing its evidence lower bound (ELBO), which effectively aligns the true denoising model " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": " with the parameterized " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": ". During training, the goal of the denoising network " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(.)" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": " is to reconstruct " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": " given any noised input " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": ", by predicting the added noise " + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 470, + 555, + 590 + ], + "type": "text", + "content": " via minimizing the noise prediction error" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 599, + 555, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 599, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 326, + 599, + 555, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\right\\| ^ {2} \\right]. (4)", + "image_path": "edba708a56f1c60e2ec3749d42525c5be57e433fb38367dacf3472807dec3ed0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "text", + "content": "To condition the model on additional context information " + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "text", + "content": ", e.g., text, we inject " + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(.)" + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "text", + "content": " by replacing " + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\mathbf{x}_t,t,\\mathbf{c})" + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}(\\mathbf{x}_t,t,\\mathbf{c})" + }, + { + "bbox": [ + 313, + 628, + 555, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 671, + 397, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 397, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 397, + 683 + ], + "type": "text", + "content": "3.2. MetaMotion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "We present the core concept \"MetaMotion\" for human-human motion generation at first. As shown in Fig. 1," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7171" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "we abstract the process of human-human motion generation into two phases: temporal modeling and interaction mixing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 145 + ], + "type": "text", + "content": "Specifically, two single-person sequences go through the temporal modeling module to get the input sequence " + }, + { + "bbox": [ + 55, + 96, + 296, + 145 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 96, + 296, + 145 + ], + "type": "text", + "content": ". Then, the input sequence is fed to the interaction-mixing module and this process can be expressed as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 155, + 295, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 155, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 155, + 295, + 168 + ], + "type": "interline_equation", + "content": "Y = \\operatorname {I n t e r a c t i o n M i x i n g} (X), \\tag {5}", + "image_path": "ca2c509bfb5b0098a8c8459f3564bf34c2487c969f18e14deab8a8a4c59d1160.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 178, + 296, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 227 + ], + "type": "text", + "content": "where InteractionMixing is usually the transformer structure including self-attention and cross-attention. Notably, InteractionMixing can also be some emerging structures, e.g., Mamba [8], RWKV [21]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 234, + 124, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 124, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 124, + 246 + ], + "type": "text", + "content": "3.3. TIMotion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 252, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 347 + ], + "type": "text", + "content": "For human-human motion generation, most current methods are based on extensions of single-person motion generation. Some design-specific approach (e.g. InterGen [16]) considers the interaction between two persons in the interaction mixing module. However, current methods ignore the importance of temporal modeling, resulting in suboptimal performance in generating long-sequence motions and multi-person interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 348, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 468 + ], + "type": "text", + "content": "In this work, we propose TIMotion, as illustrated in Fig. 2, an effective temporal modeling approach that can be applied to different interaction-mixing structures, enabling better handling the human-human interactions. TIMotion consists of three key technical designs: (1) Causal Interactive Injection, which identifies the causal relationships between human-human interactions; (2) Role-Evolving Scanning, which adapts to shifts between active and passive roles during interactions; (3) Localized Pattern Amplification, designed to handle short-term sequences better." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 203, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 203, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 203, + 487 + ], + "type": "text", + "content": "3.3.1. Causal Interactive Injection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 562 + ], + "type": "text", + "content": "Perception of ego-motion and interaction between two persons are two important elements of human-human motion generation. Considering the causal properties of motion, we propose Causal Interactive Injection, a temporal modeling approach to simultaneously achieve both perceptions of ego-motion and interactions with each other." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "text", + "content": "Specifically, we denote two single-person motion sequences as " + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "inline_equation", + "content": "\\{x_{a}, x_{b}\\}" + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "inline_equation", + "content": "x_{a} = \\{x_{a}^{j}\\}_{j=1}^{L}" + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "inline_equation", + "content": "x_{b} = \\{x_{b}^{j}\\}_{j=1}^{L}" + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "text", + "content": " are respective sequences of motion, and " + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "text", + "content": " is the length of the sequence. Since the motions of two individuals at the current time step are jointly determined by their motions at previous time steps, we model the two single-person motion sequences as a causal interaction sequence " + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "inline_equation", + "content": "x_{cii} = \\{x_{k}^{j/2}\\}_{j=1}^{2L}" + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "text", + "content": ", the symbol // denotes division followed by rounding up and " + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 563, + 296, + 675 + ], + "type": "text", + "content": " can be acquired as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 684, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 684, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 127, + 684, + 295, + 717 + ], + "type": "interline_equation", + "content": "k = \\left\\{ \\begin{array}{ll}a, & j\\% 2 = 1.\\\\ b, & j\\% 2 = 0. \\end{array} \\right. \\tag{6}", + "image_path": "4539d26fb21f92034100e15ac355bc614870695f052d56f9b29a1307b1760b13.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "Then we can inject them into the interaction-mixing module and separate the motion embeddings of the two individuals from the outputs according to Eq. (6)." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 342, + 122, + 379, + 169 + ], + "blocks": [ + { + "bbox": [ + 342, + 122, + 379, + 169 + ], + "lines": [ + { + "bbox": [ + 342, + 122, + 379, + 169 + ], + "spans": [ + { + "bbox": [ + 342, + 122, + 379, + 169 + ], + "type": "image", + "image_path": "71ab8ac14325f66d143f13be212b975d25864f6c50b3729e69f7a1131078f7cb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 174, + 376, + 181 + ], + "lines": [ + { + "bbox": [ + 353, + 174, + 376, + 181 + ], + "spans": [ + { + "bbox": [ + 353, + 174, + 376, + 181 + ], + "type": "text", + "content": "frame 1" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 412, + 124, + 446, + 169 + ], + "blocks": [ + { + "bbox": [ + 412, + 124, + 446, + 169 + ], + "lines": [ + { + "bbox": [ + 412, + 124, + 446, + 169 + ], + "spans": [ + { + "bbox": [ + 412, + 124, + 446, + 169 + ], + "type": "image", + "image_path": "11df8a091534d89a05f9cb811d89f472c2bbac40da657fc1570fdb8a58579d31.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 174, + 440, + 181 + ], + "lines": [ + { + "bbox": [ + 417, + 174, + 440, + 181 + ], + "spans": [ + { + "bbox": [ + 417, + 174, + 440, + 181 + ], + "type": "text", + "content": "frame 2" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 485, + 125, + 526, + 171 + ], + "blocks": [ + { + "bbox": [ + 485, + 125, + 526, + 171 + ], + "lines": [ + { + "bbox": [ + 485, + 125, + 526, + 171 + ], + "spans": [ + { + "bbox": [ + 485, + 125, + 526, + 171 + ], + "type": "image", + "image_path": "9b954790ee49f4cf025bb6d156b0637f7cb4fe8b9632dd2b45299d4ba358eef3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 174, + 519, + 181 + ], + "lines": [ + { + "bbox": [ + 496, + 174, + 519, + 181 + ], + "spans": [ + { + "bbox": [ + 496, + 174, + 519, + 181 + ], + "type": "text", + "content": "frame 3" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 341, + 197, + 388, + 242 + ], + "blocks": [ + { + "bbox": [ + 341, + 197, + 388, + 242 + ], + "lines": [ + { + "bbox": [ + 341, + 197, + 388, + 242 + ], + "spans": [ + { + "bbox": [ + 341, + 197, + 388, + 242 + ], + "type": "image", + "image_path": "17ae2a3f4f8a4aa1cca1d86633b47b7de07a43f1f2917c4a4fb1e34175ac2a3a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 243, + 376, + 250 + ], + "lines": [ + { + "bbox": [ + 354, + 243, + 376, + 250 + ], + "spans": [ + { + "bbox": [ + 354, + 243, + 376, + 250 + ], + "type": "text", + "content": "frame 4" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 262, + 529, + 289 + ], + "lines": [ + { + "bbox": [ + 349, + 262, + 529, + 289 + ], + "spans": [ + { + "bbox": [ + 349, + 262, + 529, + 289 + ], + "type": "text", + "content": "The first person suddenly pushes the other one. The second person takes a step back, then forcefully pushes the first person. The first person was knocked down." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 414, + 194, + 446, + 241 + ], + "blocks": [ + { + "bbox": [ + 414, + 194, + 446, + 241 + ], + "lines": [ + { + "bbox": [ + 414, + 194, + 446, + 241 + ], + "spans": [ + { + "bbox": [ + 414, + 194, + 446, + 241 + ], + "type": "image", + "image_path": "044b4c106d984b6aa885e10c9fbed0fbb8d7cac7983c0784599d5287a6a4e7d8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 418, + 243, + 441, + 250 + ], + "lines": [ + { + "bbox": [ + 418, + 243, + 441, + 250 + ], + "spans": [ + { + "bbox": [ + 418, + 243, + 441, + 250 + ], + "type": "text", + "content": "frame 5" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 481, + 194, + 532, + 240 + ], + "blocks": [ + { + "bbox": [ + 481, + 194, + 532, + 240 + ], + "lines": [ + { + "bbox": [ + 481, + 194, + 532, + 240 + ], + "spans": [ + { + "bbox": [ + 481, + 194, + 532, + 240 + ], + "type": "image", + "image_path": "00a1a7dfbfafc56057fad7854d5b9304b4e4c42f4c648a79c5fade3f01befab9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 243, + 520, + 250 + ], + "lines": [ + { + "bbox": [ + 498, + 243, + 520, + 250 + ], + "spans": [ + { + "bbox": [ + 498, + 243, + 520, + 250 + ], + "type": "text", + "content": "frame 6" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 297, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 331 + ], + "type": "text", + "content": "Figure 3. Illustration of changing active and passive roles. The first person acts as the active role in the early stages, and as time progresses, the other person becomes the active role of the motion." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 350, + 446, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 446, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 446, + 363 + ], + "type": "text", + "content": "3.3.2. Role-Evolving Scanning" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 365, + 555, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 497 + ], + "type": "text", + "content": "As noted in [4], human interactions inherently follow a certain order, e.g., \"shaking hands\" typically begins with one person extending a hand first, which means that interactive motion can be categorized as active or passive. Some methods [29] split the text description into active and passive voices. However, the \"active\" and \"passive\" roles are not static in a text description. As shown in Fig. 3, these roles constantly swap between characters as the interaction progresses. To minimize redundant text preprocessing and adapt to the ongoing shifts in roles, we design Role-Evolving Scanning, which is both efficient and effective." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": "For the causal interaction sequence " + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": " defined in Causal Interactive Injection (Sec. 3.3.1), it is obvious that " + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": " represent the active sequence and passive sequence, respectively. However, the above assumption of active and passive motions is not always in line with the real order. To cope with the changing roles, we re-model the interactive motion sequences as a symmetric causal interaction sequence " + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "inline_equation", + "content": "x_{sym\\_ cii} = \\{x_{k'}^{j // 2}\\}_{j=1}^{2L}" + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": ", where the symbol " + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "inline_equation", + "content": "//" + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": " is defined as division followed by rounding up and " + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "inline_equation", + "content": "k'" + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": " is obtained by exchanging " + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": " in Eq. (6)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "text", + "content": "Given that the causal interaction sequence " + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "inline_equation", + "content": "x_{cii} \\in \\mathbb{R}^{2L \\times C}" + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "text", + "content": " and the symmetric causal interaction sequence " + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "inline_equation", + "content": "x_{sym\\_ cii} \\in \\mathbb{R}^{2L \\times C}" + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "text", + "content": " is the length of the sequence and " + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "text", + "content": " is the dimension of motion embedding, we can acquire the final interaction sequence " + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{2L \\times 2C}" + }, + { + "bbox": [ + 313, + 618, + 554, + 691 + ], + "type": "text", + "content": " through Role-Evolving Scanning as:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 373, + 701, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 701, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 373, + 701, + 553, + 715 + ], + "type": "interline_equation", + "content": "X = \\operatorname {C o n c a t} \\left(x _ {c i i}, x _ {s y m. c i i}\\right). \\tag {7}", + "image_path": "04b73d2bdcf2897ba7d91c698d71aadf4c356b6b439ed6a823d0caa2aabff99b.jpg" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7172" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": "Then the sequence " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": " is fed to the interaction-mixing module. After obtaining the output " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^{2L \\times 2C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": " through Eq. (5), we first divide it at the channel level into causal interaction embeddings and symmetric causal interaction embeddings. According to " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "x_{cii}" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "x_{sym\\_ cii}" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": ", we can obtain the two individuals' splitting embeddings twice. Then we merge the embeddings of the two individuals to obtain the final global motion embeddings " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "y_a^g \\in \\mathbb{R}^{L \\times C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "y_b^g \\in \\mathbb{R}^{L \\times C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": ". The overall process is as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 193, + 295, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 193, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 116, + 193, + 295, + 251 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} y _ {a 1}, y _ {b 1} = \\operatorname {S p l i t} (Y [ :, : C ]), \\\\ y _ {b 2}, y _ {a 2} = \\operatorname {S p l i t} (Y [:, C: ]), \\tag {8} \\\\ y _ {a} ^ {g} = y _ {a 1} \\oplus y _ {a 2}, \\\\ y _ {b} ^ {g} = y _ {b 1} \\oplus y _ {b 2}, \\\\ \\end{array}", + "image_path": "ce0aa94ca3566d1e20f07aecfc3e13f3d2c58e84822e6d845bcbd071f5045b8a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 264, + 216, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 264, + 216, + 275 + ], + "spans": [ + { + "bbox": [ + 67, + 264, + 216, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 264, + 216, + 275 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 67, + 264, + 216, + 275 + ], + "type": "text", + "content": " denotes element-wise sum." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 296, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 325 + ], + "type": "text", + "content": "By utilizing Role-Evolving Mixing to make two humans act as both active and passive roles, the network can dynamically adjust the roles of the two humans based on the semantics of the text and the context of the motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 336, + 220, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 220, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 220, + 348 + ], + "type": "text", + "content": "3.3.3. Localized Pattern Amplification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 352, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 471 + ], + "type": "text", + "content": "Transformers and RNNs excel at global modeling and capturing long-range dependencies but tend to overlook local semantic information [13]. Additionally, Causal Interactive Injection and Role-Evolving Mixing mainly model the overall motion based on the causality of the interactions, but neglect to focus on the localized motion patterns of the single person. To address this issue, we propose Localized Pattern Amplification, which summarizes short-term motion patterns for each person individually while generating smoother and more rational motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "text", + "content": "Specifically, we utilize 1-D convolution layers and the residual structure to realize Localized Pattern Amplification. Given that the condition embedding " + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "text", + "content": " and two single-person motion sequences " + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "text", + "content": ", the local motion embedding for " + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 544, + 295, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 544, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 111, + 544, + 295, + 593 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {a} ^ {l} = \\operatorname {C o n v} _ {3} (\\operatorname {A d a L N} (x _ {a}, e)), \\\\ y _ {a} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {A d a L N} \\left(x _ {a} ^ {l}, e\\right)\\right), \\tag {9} \\\\ y _ {a} ^ {l} = x _ {a} + y _ {a}, \\\\ \\end{array}", + "image_path": "b917ca0696416d7202e14d5240e5b1599d28e5952a2464ff54120fc122c21e7d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv_k}" + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "content": " denotes the 1-D convolution with kernel size " + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "content": " and AdaLN is the adaptive layer normalization used in [42]. The process for " + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "content": " is the same as for " + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "content": " and they share network weights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Once we obtain the outputs " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{y_a^g, y_b^g\\}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " through Eq. (8) and the output " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{y_a^l, y_b^l\\}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " from the convolution block, global embeddings and local embeddings are aggregated through concatenation along the channel dimension, followed by a linear layer to restore the original channels, resulting in the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 70, + 438, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 70, + 438, + 86 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 438, + 86 + ], + "type": "text", + "content": "final outputs " + }, + { + "bbox": [ + 314, + 70, + 438, + 86 + ], + "type": "inline_equation", + "content": "\\{y_a^{final}, y_b^{final}\\}" + }, + { + "bbox": [ + 314, + 70, + 438, + 86 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 361, + 91, + 553, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 91, + 553, + 124 + ], + "spans": [ + { + "bbox": [ + 361, + 91, + 553, + 124 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} y _ {a} ^ {f i n a l} = \\operatorname {L i n e a r} (\\operatorname {C o n c a t} (y _ {a} ^ {g}, y _ {a} ^ {l})), \\\\ y _ {b} ^ {\\text {f i n a l}} = \\operatorname {L i n e a r} \\left(\\operatorname {C o n c a t} \\left(y _ {b} ^ {g}, y _ {b} ^ {l}\\right)\\right). \\tag {10} \\\\ \\end{array}", + "image_path": "7bf4f64832a6eac9562ce000187b5943ce1ac58599570b40f2718ae93c42fd83.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 129, + 555, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 555, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 555, + 154 + ], + "type": "text", + "content": "Then the final outputs are fed into the next encoder block or the final decoder layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 159, + 424, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 424, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 424, + 171 + ], + "type": "text", + "content": "3.3.4. Objective Function" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "text", + "content": "We use the same loss functions as InterGen [16], including the diffusion loss " + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{simple}" + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "text", + "content": ", the foot contact loss " + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{foot}" + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "text", + "content": ", joint velocity loss " + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vel}" + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "text", + "content": ", the bone length loss " + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BL}" + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "text", + "content": ", the masked joint distance map loss " + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DM}" + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "text", + "content": ", and the relative orientation loss " + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{RO}" + }, + { + "bbox": [ + 313, + 174, + 555, + 245 + ], + "type": "text", + "content": ". For more details about the losses, we refer readers to InterGen. Finally, the overall loss is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 253, + 553, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 253, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 332, + 253, + 553, + 281 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {m o t i o n}} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {v e l} \\mathcal {L} _ {v e l} + \\lambda_ {f o o t} \\mathcal {L} _ {f o o t} \\\\ + \\lambda_ {B L} \\mathcal {L} _ {B L} + \\lambda_ {D M} \\mathcal {L} _ {D M} + \\lambda_ {R O} \\mathcal {L} _ {R O}, \\tag {11} \\\\ \\end{array}", + "image_path": "01aed2304fb64767d8b18655f3a8b0fc29231e91fb09f9202cd398a0c1454a57.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 288, + 554, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 311 + ], + "type": "text", + "content": "where the hyper-parameters " + }, + { + "bbox": [ + 313, + 288, + 554, + 311 + ], + "type": "inline_equation", + "content": "\\lambda_{vel},\\lambda_{foot},\\lambda_{BL},\\lambda_{DM},\\lambda_{RO}" + }, + { + "bbox": [ + 313, + 288, + 554, + 311 + ], + "type": "text", + "content": " are the same as InterGen." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 321, + 395, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 321, + 395, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 321, + 395, + 335 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 341, + 430, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 341, + 430, + 354 + ], + "spans": [ + { + "bbox": [ + 314, + 341, + 430, + 354 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 358, + 555, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 466 + ], + "type": "text", + "content": "Datasets. We assess our proposed framework using the InterHuman [16] and Inter-X [39] dataset. InterHuman is the first dataset to incorporate text annotations for two-person motions. This dataset includes 6,022 motions spanning various categories of human actions and is labeled with 16,756 unique descriptions made up of 5,656 distinct words. InterX is the largest human-human interaction dataset with diverse interaction patterns. It includes about 11k interaction sequences and more than 8.1M frames." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 466, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 554, + 538 + ], + "type": "text", + "content": "Metrics. We employ the same evaluation metrics as InterGen [16], which are as follows: (1) Frechet Inception Distance (FID). (2) " + }, + { + "bbox": [ + 313, + 466, + 554, + 538 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 466, + 554, + 538 + ], + "type": "text", + "content": "-Precision. (3) Diversity. (4) Multimodality (MModality). (5) Multi-modal distance (MM Dist). Detailed explanations and calculations of the metrics can be found in the supplementary materials." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 543, + 444, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 444, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 444, + 555 + ], + "type": "text", + "content": "4.1.1. Implementation Details." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": "We use a frozen CLIP-ViT-L/14 model as the text encoder. The dimension of motion embedding is set to 512. During training, the number of diffusion timesteps is set to 1000, and we employ the DDIM [28] sampling strategy with 50 timesteps and " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\eta = 0" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": ". The cosine noise level schedule [19] and classifier-free guidance [14] are adopted, with " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": " of random CLIP embeddings set to zero during training and a guidance coefficient of 3.5 during sampling. All the models are trained using the AdamW [17] optimizer with betas of (0.9, 0.999), a weight decay of " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": ", a maximum learning rate of " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": ", and a cosine learning rate schedule with 10 linear warm-up epochs. To balance the contributions of different loss terms, we set " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{vel} = 30" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{foot} = 30" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{BL} = 10" + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7173" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 220 + ], + "type": "table", + "html": "
MethodsR Precision↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Real0.452±.0080.610±.0090.701±.0080.273±.0073.755±.0087.948±.064-
TEMOS [23]0.224±.0100.316±.0130.450±.01817.375±.0435.342±.0156.939±.0710.535±.014
T2M [10]0.238±.0120.325±.0100.464±.01413.769±.0724.731±.0137.046±.0221.387±.076
MDM [31]0.153±.0120.260±.0090.339±.0129.167±.0566.125±.0187.602±.0452.355±.080
ComMDM* [27]0.067±.0130.125±.0180.184±.01538.643±.09813.211±.0133.520±.0580.217±.018
ComMDM [27]0.223±.0090.334±.0080.466±.0107.069±.0545.212±.0217.244±.0381.822±.052
RIG [29]0.285±.0100.409±.0140.521±.0136.775±.0694.876±.0187.311±.0432.096±.065
InterGen [16]0.371±.0100.515±.0120.624±.0105.918±.0795.108±.0147.387±.0292.141±.063
TIMotion+Transformer(ours)0.491±.0050.648±.0040.724±.0045.433±.0803.775±.0018.032±.0300.952±.032
TIMotion+Mamba(ours)0.431±.0040.586±.0050.668±.0046.142±.0593.800±.0017.793±.0240.837±.036
TIMotion+RWKV(ours)0.501±.0050.656±.0060.734±.0064.702±.0693.769±.0017.943±.0341.005±.020
", + "image_path": "a012227edeffb418684371cfc6da68b7c90eca3c93d1b44c39af8922c95d81cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 271, + 529, + 380 + ], + "blocks": [ + { + "bbox": [ + 55, + 228, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 262 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on the InterHuman [16] test set. We run all the evaluations 20 times. ± indicates a 95% confidence interval. Bold indicates the best result, while underline refers to the second best. ComMDM* indicates the ComMDM model fine-tuned in the original few-shot setting with 10 training samples and ComMDM indicates fine-tuned on the entire InterHuman training set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 271, + 529, + 380 + ], + "lines": [ + { + "bbox": [ + 82, + 271, + 529, + 380 + ], + "spans": [ + { + "bbox": [ + 82, + 271, + 529, + 380 + ], + "type": "table", + "html": "
Interaction MixingTemporal ModelingR Precision↑FID↓MM Dist↓Diversity→MModality↑
Top 1Top 2Top 3
Transformer [35]single-person extension0.395±.0040.535±.0040.615±.0048.028±.0993.820±.0017.687±.0220.852±.024
separate modeling0.371±.0100.515±.0120.624±.0105.918±.0795.108±.0147.387±.0292.141±.063
TIMotion (Ours)0.491±.0050.648±.0040.724±.0045.433±.0803.775±.0018.032±.0300.952±.032
Mamba [8]single-person extension0.368±.0050.513±.0050.597±.0067.232±.0813.825±.0017.864±.0240.870±.020
separate modeling0.420±.0050.569±.0040.650±.0037.221±.0973.803±.0017.932±.0350.855±.020
TIMotion (Ours)0.431±.0040.586±.0050.668±.0046.142±.0593.800±.0017.793±.0240.837±.036
RWKV [21]single-person extension0.425±.0040.576±.0040.656±.0039.181±.0963.801±.0017.679±.0220.846±.026
separate modeling0.465±.0050.603±.0050.689±.0045.943±.0793.790±.0017.787±.0320.859±.036
TIMotion (Ours)0.501±.0050.656±.0060.734±.0064.702±.0693.769±.0017.943±.0341.005±.020
", + "image_path": "61d4f1926ed2b0905ed701b435d42d30e704508d719705ca7e0b05727a9653ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 555, + 410 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 555, + 410 + ], + "type": "text", + "content": "Table 2. Comparison of different temporal modeling approaches on different interaction mixing structures. Our proposed causal interactive modeling is able to adapt to different interaction mixing architectures and outperforms the other two ways." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 431, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 467 + ], + "type": "inline_equation", + "content": "\\lambda_{DM} = 3, \\lambda_{RO} = 0.01" + }, + { + "bbox": [ + 55, + 431, + 295, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 431, + 295, + 467 + ], + "type": "inline_equation", + "content": "\\lambda_{reg} = 1" + }, + { + "bbox": [ + 55, + 431, + 295, + 467 + ], + "type": "text", + "content": " in all experiments, as same as InterGen. We train our diffusion denoisers with a batch size of 256 for 1500 epochs on 8 Nvidia L40S GPUs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 243, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 243, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 243, + 493 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-arts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 500, + 173, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 173, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 173, + 511 + ], + "type": "text", + "content": "4.2.1. Quantitative Results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "type": "text", + "content": "Following established practices [16], each experiment is conducted 20 times, and the reported metric values represent the mean with a " + }, + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "type": "text", + "content": " statistical confidence interval. The results on InterHuman are shown in Tab. 1. In comparison to existing methods, our TIMotion achieves competitive results on three different interaction mixing architectures: transformer, mamba, and RWKV. Equipped with RWKV, TIMotion achieves 4.702 FID and 0.501 Top1 R precision, setting new state-of-the-art (SoTA) for the competitive InterHuman benchmark. The results on Inter-X [39] can be found in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 192, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 192, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 192, + 673 + ], + "type": "text", + "content": "4.2.2. Qualitative Comparisons." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "In Fig. 4, we qualitatively compare the InterGen and our TIMotion. It can be seen that sequences generated by TIMotion are more consistent with the description." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 430, + 458, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 458, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 458, + 441 + ], + "type": "text", + "content": "4.3. Effectiveness of TIMotion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 594 + ], + "type": "text", + "content": "To validate the effectiveness of our proposed causal interactive modeling (TIMotion), we compare different temporal modeling approaches on three different structures, respectively. As shown in Tab. 2, single-person extension performs poorly, indicating that it does not model two-person interactions well. In contrast, separate modeling leverages the ability of the interaction mixing module to model the two-person interactions but achieves sub-optimal performance on the Transformer and mamba architectures. Our proposed approach is able to adapt to different interaction mixing architectures and outperforms the other two approaches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 609, + 471, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 471, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 471, + 623 + ], + "type": "text", + "content": "4.4. Ablation Study and Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "In this section we conduct comprehensive ablation studies to investigate the effectiveness of the key components in TIMotion, thereby providing a deeper insight into our approach. All experiments are performed on the InterHuman dataset. Unless otherwise stated, we use RWKV as the interaction mixing structure for the following ablations. More theoretical analyses are in the supplementary materials." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7174" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 79, + 214, + 243 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 214, + 243 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 214, + 243 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 214, + 243 + ], + "type": "image", + "image_path": "32f69a0128b9b6a9993a070ccf80c9be2b31ef8e8fab21a13d4d1b9d72f5ba4e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 245, + 204, + 261 + ], + "lines": [ + { + "bbox": [ + 116, + 245, + 204, + 261 + ], + "spans": [ + { + "bbox": [ + 116, + 245, + 204, + 261 + ], + "type": "text", + "content": "[1] two individuals are lifting and transporting an object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 271, + 73, + 394, + 244 + ], + "blocks": [ + { + "bbox": [ + 271, + 73, + 394, + 244 + ], + "lines": [ + { + "bbox": [ + 271, + 73, + 394, + 244 + ], + "spans": [ + { + "bbox": [ + 271, + 73, + 394, + 244 + ], + "type": "image", + "image_path": "753a0bc2dbc51e6d9e0121a88cd27339c507bb7d5a0b897a290ec09baa4a2580.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 245, + 394, + 262 + ], + "lines": [ + { + "bbox": [ + 271, + 245, + 394, + 262 + ], + "spans": [ + { + "bbox": [ + 271, + 245, + 394, + 262 + ], + "type": "text", + "content": "[2] the other person pushes one person to the ground with both hands." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 455, + 72, + 529, + 245 + ], + "blocks": [ + { + "bbox": [ + 455, + 72, + 529, + 245 + ], + "lines": [ + { + "bbox": [ + 455, + 72, + 529, + 245 + ], + "spans": [ + { + "bbox": [ + 455, + 72, + 529, + 245 + ], + "type": "image", + "image_path": "cedfa3d746622368f0209122e2c7d9f47c33b5f692f50ca0dd848f8910ef4044.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 246, + 549, + 269 + ], + "lines": [ + { + "bbox": [ + 452, + 246, + 549, + 269 + ], + "spans": [ + { + "bbox": [ + 452, + 246, + 549, + 269 + ], + "type": "text", + "content": "[3] one wraps both arms around the back of the other person, while the other person reciprocates the action" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with Intergen on human-human motion generation. Darker color indicates later frames. The sequences generated by TIMotion are more consistent with the text description." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 67, + 318, + 285, + 394 + ], + "blocks": [ + { + "bbox": [ + 67, + 318, + 285, + 394 + ], + "lines": [ + { + "bbox": [ + 67, + 318, + 285, + 394 + ], + "spans": [ + { + "bbox": [ + 67, + 318, + 285, + 394 + ], + "type": "table", + "html": "
CIIRESLPAR Precision Top 1↑FID ↓Params (M)
0.371±.0105.918±.079182
✓0.478±.0055.410±.069144
✓✓0.494±.0065.019±.077115
✓✓✓0.501±.0054.702±.069127
", + "image_path": "addc9f83f0cf3ff8e94950f2b8c4357c501327b39026b2b01c6ab921c7719755.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 340, + 318, + 529, + 402 + ], + "blocks": [ + { + "bbox": [ + 55, + 403, + 295, + 447 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 447 + ], + "type": "text", + "content": "Table 3. Ablation studies on the effectiveness of each component in TIMotion. \"CII\" denotes Causal Interactive Injection, \"RES\" denotes Role-Evolving Scanning, and \"LPA\" denotes Localized Pattern Amplification." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 318, + 529, + 402 + ], + "lines": [ + { + "bbox": [ + 340, + 318, + 529, + 402 + ], + "spans": [ + { + "bbox": [ + 340, + 318, + 529, + 402 + ], + "type": "table", + "html": "
Kernel SizeNormR Precision Top 1↑FID ↓
k=3,1BN0.494±.0055.443±.093
k=3,1LN0.488±.0055.339±.080
k=3,1AdaLN0.501±.0054.702±.069
k=3,3AdaLN0.484±.0057.120±.090
k=5,1AdaLN0.497±.0055.102±.064
", + "image_path": "34052abac03c8281759500561884a9ff6748dcc40c572492cdaf6439ef7c7edc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 410, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 453 + ], + "type": "text", + "content": "Table 4. Ablation studies on LPA. \"BN\" denotes batch normalization, \"LN\" denotes layer normalization and \"AdaLN\" denotes adaptive layer normalization. \"k=3,1\" means that the first kernel size of the convolution is 3 and the second kernel size is 1." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 469, + 150, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 150, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 150, + 480 + ], + "type": "text", + "content": "4.4.1. Main Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 485, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 295, + 653 + ], + "type": "text", + "content": "To understand the contribution of each component to the final performance, we incrementally add the proposed modules based on the baseline InterGen [16] and present the results in Tab. 3. Initially, we replace self-attention and cross-attention in the transformer with Causal Interactive Injection (CII), which effectively improves both R-Precision and FID while reducing the number of parameters. Next, we apply Role-Evolving Scanning (RES) and yield gains in both R-Precision and FID. It is worth noting that to keep the input feature dimension of RWKV constant, we reduce the dimension of the motion embedding to half of the original, thus also effectively reducing the number of parameters. Finally, when all three methods are applied together, the R-Precision achieves 0.501 and FID achieves 4.702." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 662, + 264, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 264, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 264, + 673 + ], + "type": "text", + "content": "4.4.2. Design of Localized Pattern Amplification." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In this section, we conduct ablations on the design of Localized Pattern Amplification (LPA), and the results are shown in Tab. 4. First, we explore the effect of different normal-" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 490, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 587 + ], + "type": "text", + "content": "izations on model performance. Using batch normalization (BN) and layer normalization (LN) does not bring gains for text-to-motion tasks. In contrast, AdaLN, which integrates text information into the extraction process of the local motion pattern, proves to be the optimal normalization method. Additionally, we examine how the kernel size of the convolutional layers impacts the model's performance and select the setting of " + }, + { + "bbox": [ + 313, + 490, + 555, + 587 + ], + "type": "inline_equation", + "content": "k = 3,1" + }, + { + "bbox": [ + 313, + 490, + 555, + 587 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "To further demonstrate the effectiveness of LPA, we visualize the spectrum of TIMotion w/o and w/ LPA in Fig. 6, respectively. The magnitude has been normalized. As can be seen in Fig. 6, adding LPA reduces the high-frequency component of the feature, making the motion less likely to change drastically and making it smoother. Moreover, we randomly sampled 200 instances, and w/ LPA, the average proportion of the amplitude of high-frequency components is 0.3729, whereas w/o LPA, the average proportion of the amplitude of high-frequency components is 0.9063." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7175" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 72, + 529, + 179 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 529, + 179 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 529, + 179 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 529, + 179 + ], + "type": "image", + "image_path": "3175bf58ee1d5d1fc35fa25bd41e1d2b1066aa8d0504bd347afe7e8be989f26d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 188, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 553, + 209 + ], + "type": "text", + "content": "Figure 5. Qualitative results on the motion in-between task. The first and last frames are fixed. Darker colors indicate later frames. Our method achieves smooth and natural transitions between the conditioned motions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 85, + 230, + 266, + 442 + ], + "blocks": [ + { + "bbox": [ + 85, + 230, + 266, + 442 + ], + "lines": [ + { + "bbox": [ + 85, + 230, + 266, + 442 + ], + "spans": [ + { + "bbox": [ + 85, + 230, + 266, + 442 + ], + "type": "image", + "image_path": "dde4ba133cb0e6743496320a791b5007b3f85ee05c92f4833b169e11b14b4302.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 452, + 295, + 507 + ], + "lines": [ + { + "bbox": [ + 55, + 452, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 295, + 507 + ], + "type": "text", + "content": "Figure 6. Spectrum of motion features. (a) and (b) show the spectrum of TIMotion w/o and w/ LPA, respectively. The horizontal axis denotes the frequency and the vertical axis represents the normalized magnitude. TIMotion w/ LPA contains fewer high-frequency components and therefore generates smoother motion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 528, + 281, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 281, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 281, + 540 + ], + "type": "text", + "content": "4.5. Computational Complexity and Editability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 202, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 202, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 202, + 558 + ], + "type": "text", + "content": "4.5.1. Computational Complexity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 560, + 296, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 656 + ], + "type": "text", + "content": "In Tab. 5, we compare our approach TIMotion with the SOTA method InterGen [16] in terms of computational complexity. TIMotion requires fewer parameters and FLOPs than InterGen but outperforms it on the comprehensive metric FID and R precision. Notably, using a similar transformer architecture to InterGen, TIMotion's average inference time per sample is only 0.632 s while InterGen requires 1.991 s." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 662, + 132, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 132, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 132, + 674 + ], + "type": "text", + "content": "4.5.2. Editability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Tab. 6 illustrates that TIMotion surpasses InterGen in the task of motion in-between editing. We perform experiments on the test set of InterHuman and evaluate by generat" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 326, + 229, + 543, + 273 + ], + "blocks": [ + { + "bbox": [ + 326, + 229, + 543, + 273 + ], + "lines": [ + { + "bbox": [ + 326, + 229, + 543, + 273 + ], + "spans": [ + { + "bbox": [ + 326, + 229, + 543, + 273 + ], + "type": "table", + "html": "
MethodsR Precision Top 1↑FID ↓Params (M)Flops (G)Inference time (s)
InterGen (Transformer)0.371±.0105.918±.07918280.51.991
TIMotion+Transformer0.491±.0055.433±.0806540.40.632
TIMotion+RWKV0.501±.0054.702±.06912758.01.733
", + "image_path": "89d931d3c183c4bc13700cf5614ef3b780aa795091936965531d3f8e4f24b091.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 317, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 365 + ], + "type": "text", + "content": "ing " + }, + { + "bbox": [ + 313, + 317, + 555, + 365 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 317, + 555, + 365 + ], + "type": "text", + "content": " of the sequences based on the first and last " + }, + { + "bbox": [ + 313, + 317, + 555, + 365 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 317, + 555, + 365 + ], + "type": "text", + "content": " of the sequences. The quantitative results are shown in Fig. 5 and our method achieves smooth and natural transitions between the conditioned motions." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 326, + 376, + 543, + 423 + ], + "blocks": [ + { + "bbox": [ + 336, + 281, + 531, + 293 + ], + "lines": [ + { + "bbox": [ + 336, + 281, + 531, + 293 + ], + "spans": [ + { + "bbox": [ + 336, + 281, + 531, + 293 + ], + "type": "text", + "content": "Table 5. Comparison of computational complexity." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 376, + 543, + 423 + ], + "lines": [ + { + "bbox": [ + 326, + 376, + 543, + 423 + ], + "spans": [ + { + "bbox": [ + 326, + 376, + 543, + 423 + ], + "type": "table", + "html": "
MethodsR Precision Top 1↑FID ↓MM Dist↓Diversity→
InterGen0.461±.0064.700±.0663.780±.0017.682±.029
TIMotion0.516±.0063.590±.0493.760±.0017.795±.031
", + "image_path": "67eeb12d159b501fe1a57cdf1ebc7ebd40ce471f78c99ac4ad0c5dbb3230f814.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 329, + 431, + 539, + 443 + ], + "lines": [ + { + "bbox": [ + 329, + 431, + 539, + 443 + ], + "spans": [ + { + "bbox": [ + 329, + 431, + 539, + 443 + ], + "type": "text", + "content": "Table 6. Evaluation of motion in-between editing task." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 476, + 388, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 476, + 388, + 488 + ], + "spans": [ + { + "bbox": [ + 314, + 476, + 388, + 488 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we abstract the overall human-human motion generation process into a general framework MetaMotion, which consists of two phases: temporal modeling and interaction mixing. We find that the current methods lead to sub-optimal results and redundancy of model parameters due to inadequate modeling. Based on this, we introduce TIMotion (Temporal and Interactive Modeling), an efficient and effective approach for human-human motion generation. Specifically, we first propose Causal Interactive Injection to model two separate sequences as a causal sequence leveraging the temporal and causal properties. Additionally, we proposed Role-Evolving Mixing to adapt to the dynamic roles throughout interactions and designed Localized Pattern Amplification to capture short-term motion patterns for generating smoother and more rational motion. Extensive experiments on the InterHuman and Inter-X datasets demonstrate that TIMotion significantly outperforms existing methods, achieving state-of-the-art results." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7176" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "type": "text", + "content": "[2] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 215 + ], + "type": "text", + "content": "[3] Erik Bethke. Game development and production. Wordware Publishing, Inc., 2003. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 296, + 282 + ], + "type": "text", + "content": "[4] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 339 + ], + "type": "text", + "content": "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 396 + ], + "type": "text", + "content": "[6] Ke Fan, Junshu Tang, Weijian Cao, Ran Yi, Moran Li, Jingyu Gong, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Freemotion: A unified framework for number-free text-to-motion synthesis. In European Conference on Computer Vision, pages 93-109. Springer, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 453 + ], + "type": "text", + "content": "[7] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. Tm2d: Bimodality driven 3d dance generation via music-text integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9942-9952, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 454, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 295, + 486 + ], + "type": "text", + "content": "[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 544 + ], + "type": "text", + "content": "[9] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 545, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 295, + 600 + ], + "type": "text", + "content": "[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "text", + "content": "[11] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[12] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In ACM SIGGRAPH 2022 conference proceedings, pages 1-9, 2022. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 152 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 153, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 186 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 188, + 555, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 555, + 232 + ], + "type": "text", + "content": "[16] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 234, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 553, + 255 + ], + "type": "text", + "content": "[17] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 258, + 553, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 258, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 553, + 290 + ], + "type": "text", + "content": "[18] Nadia Magnenat-Thalmann, Daniel Thalmann, Nadia Magnenat-Thalmann, and Daniel Thalmann. Computer animation. Springer, 1985. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 292, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 292, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 292, + 553, + 335 + ], + "type": "text", + "content": "[19] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 337, + 553, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 553, + 359 + ], + "type": "text", + "content": "[20] Rick Parent. Computer animation: algorithms and techniques. Newnes, 2012. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "text", + "content": "[21] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 418, + 553, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 418, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 317, + 418, + 553, + 461 + ], + "type": "text", + "content": "[22] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 463, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 463, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 463, + 553, + 507 + ], + "type": "text", + "content": "[23] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 509, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 509, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 509, + 553, + 553 + ], + "type": "text", + "content": "[24] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1546-1555, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 555, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 620 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 622, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 643 + ], + "type": "text", + "content": "[26] George Saridis. Intelligent robotic control. IEEE Transactions on Automatic Control, 28(5):547-557, 1983. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "type": "text", + "content": "[27] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7177" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 117 + ], + "type": "text", + "content": "[29] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 2, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "text", + "content": "[30] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 196 + ], + "type": "text", + "content": "[31] G Tevet, S Raab, B Gordon, Y Shafir, D Cohen-Or, and AH Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 229 + ], + "type": "text", + "content": "[32] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 253 + ], + "type": "text", + "content": "[33] Jay Urbain. Introduction to game development. Cell, 414: 745-5102, 2010. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 255, + 295, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 287 + ], + "type": "text", + "content": "[34] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 289, + 295, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 332 + ], + "type": "text", + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 334, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 334, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 334, + 295, + 399 + ], + "type": "text", + "content": "[36] Shuo Wang, Xinhai Zhao, Hai-Ming Xu, Zehui Chen, Dameng Yu, Jiahao Chang, Zhen Yang, and Feng Zhao. Towards domain generalization for multi-view 3d object detection in bird-eye-view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13333-13342, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 401, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 401, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 295, + 456 + ], + "type": "text", + "content": "[37] Shuo Wang, Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd-map construction. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 457, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 457, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 457, + 295, + 490 + ], + "type": "text", + "content": "[38] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 491, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 295, + 556 + ], + "type": "text", + "content": "[39] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-" + }, + { + "bbox": [ + 56, + 491, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 56, + 491, + 295, + 556 + ], + "type": "text", + "content": ": Towards versatile human-human interaction analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22260-22271, 2024. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 558, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 295, + 623 + ], + "type": "text", + "content": "[40] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "text", + "content": "[41] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[42] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 217 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[43] Chen Zhao, Weiling Cai, Chenyu Dong, and Chengwei Hu. Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8281-8291, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[44] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 509-519, 2023. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7178" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_content_list.json b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e0e511a75caaccf5de824d0cb922d71fdfbc778 --- /dev/null +++ b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_content_list.json @@ -0,0 +1,1520 @@ +[ + { + "type": "text", + "text": "TKG-DM: Training-free Chroma Key Content Generation Diffusion Model", + "text_level": 1, + "bbox": [ + 119, + 130, + 879, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ryugo Morita $^{1,2}$ , Stanislav Frolov $^{2}$ , Brian Bernhard Moser $^{2}$ , Takahiro Shirakawa, Ko Watanabe $^{2}$ , Andreas Dengel $^{2}$ , Jinjia Zhou $^{1}$", + "bbox": [ + 176, + 179, + 820, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Faculty of Science and Engineering, Hosei University, Tokyo, Japan", + "bbox": [ + 225, + 215, + 769, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ RPTU Kaiserslautern-Landau & DFKI GmbH, Kaiserslautern, Germany", + "bbox": [ + 210, + 234, + 785, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have enabled the generation of high-quality images with a strong focus on realism and textual fidelity. Yet, large-scale text-to-image models, such as Stable Diffusion, struggle to generate images where foreground objects are placed over a chroma key background, limiting their ability to separate foreground and background elements without fine-tuning. To address this limitation, we present a novel Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which optimizes the initial random noise to produce images with foreground objects on a specifiable color background. Our proposed method is the first to explore the manipulation of the color aspects in initial noise for controlled background generation, enabling precise separation of foreground and background without fine-tuning. Extensive experiments demonstrate that our training-free method outperforms existing methods in both qualitative and quantitative evaluations, matching or surpassing fine-tuned models. Finally, we successfully extend it to other tasks (e.g., consistency models and text-to-video), highlighting its transformative potential across various generative applications where independent control of foreground and background is crucial.", + "bbox": [ + 88, + 316, + 485, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 679, + 220, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have disrupted the landscape of image generation by producing high-quality images with high realism and textual fidelity [8, 11, 31, 38, 44]. Research has expanded into text-to-image generation conditioned on inputs like edges [49, 56], layouts [54, 57], and Virtual TryON (VTON) [10, 52]. However, real-world applications such as advertising, design, and game development often require precise control over foreground and background elements. The ability to generate images with transparent or chroma key backgrounds is essential for seamlessly integrating foreground objects into various scenes [55]. Traditional models like Stable Diffusion [37] and DeepFloyd [38] struggle to generate images where foreground and", + "bbox": [ + 89, + 703, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79214e5b59c17d378523b65970c5a5415ba9b9db82cf686a57878db630c01f03.jpg", + "image_caption": [ + "Figure 1. TKG-DM provides a training-free generation of foreground objects on a chroma key background, enabling independent control of foreground and background elements. Additionally, our method extends to conditional text-to-image and video generation tasks, making it suitable for creating foreground assets for various applications." + ], + "image_footnote": [], + "bbox": [ + 521, + 284, + 898, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "background elements are separated without fine-tuning [5]. Specifically, these models cannot produce foreground objects over specifiable backgrounds or generate independent foreground and background layers, limiting their applicability in workflows requiring such precise control.", + "bbox": [ + 509, + 613, + 906, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works like MAGICK [5] and LayerDiffuse [55] have significantly addressed this limitation. MAGICK introduces a prompt engineering-based approach for chroma key image generation. However, due to limited chroma key accuracy, it relies on post-processing and manual efforts to improve quality, including the release of foreground alpha image datasets. LayerDiffuse proposes a fine-tuning method for layer-wise generation using a dataset with 1 million images. Unfortunately, this dataset is not publicly available due to licensing restrictions, necessitating further resource-intensive dataset curation and fine-tuning.", + "bbox": [ + 509, + 688, + 908, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome these challenges, we present the Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which optimizes the initial noise in the dif", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13031", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fusion process to produce images with foreground objects on specifiable color backgrounds without fine-tuning. Our method is the first to explore manipulating the color aspects of initial noise for controlled background generation, enabling precise separation of foreground and background elements. TKG-DM offers high flexibility and provides precise control over the background color, layout, size, and number of foreground objects. Our approach expands possibilities in generative content by maintaining the independence of the foreground and chroma key background. It seamlessly extends to applications like conditional text-to-image generation, consistency models, and text-to-video generation.", + "bbox": [ + 89, + 90, + 480, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our extensive experiments demonstrate that TKG-DM improves FID and mask-FID scores by $33.7\\%$ and $35.9\\%$ , respectively. Thus, our training-free model rivals fine-tuned models, offering an efficient and versatile solution for various visual content creation tasks that require precise foreground and background control. Our contributions are summarized as follows:", + "bbox": [ + 89, + 287, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce TKG-DM, a training-free diffusion model that eliminates the need for fine-tuning and datasets for controllable foreground and background generation.", + "- TKG-DM provides precise control over background color, as well as the size, position and number of foreground elements, enabling flexible monochromatic chroma key generation.", + "- Our method is highly versatile, seamlessly extending to conditional text-to-image generation, consistency models, and text-to-video applications, thus enabling the creation of diverse chroma key content.", + "- TKG-DM outperforms existing models, delivering results that match or surpass fine-tuned alternatives while maintaining lower computational costs." + ], + "bbox": [ + 89, + 398, + 482, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 626, + 238, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 651, + 258, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [19, 33] generate realistic, high-quality images by iteratively refining Gaussian noise to approximate a target distribution [11, 45]. Building upon this, Latent Diffusion Models (LDMs) [37] improve efficiency by exploiting the low-dimensional latent spaces while preserving image quality. In particular, text-to-image models [32, 36-38] align text and image features through crossattention to achieve diverse, high-quality outputs, with classifier-free guidance [18] and attention control [6, 8] further enhancing text fidelity. Beyond image generation, diffusion models find use in inpainting [24, 53], image editing [4, 15], and layout generation [7, 21]. Their versatility extends across multiple modalities, including audio generation [20, 22], text-to-video [14, 50], and 3D object generation [35, 48].", + "bbox": [ + 89, + 672, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Initial Noise", + "text_level": 1, + "bbox": [ + 513, + 90, + 643, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models typically start with Gaussian noise and iteratively refine it to generate high-quality images that align closely with a given target distribution. In general, the choice and optimization of this initial noise are crucial, as they significantly impact image quality and alignment [40, 41, 51]. More specifically, noise optimization methods, such as human preference models or attention-based scores, can enhance fidelity and text alignment without additional training [12, 13]. Moreover, optimal noise seeds and targeted noise regions significantly impact image quality and object placement [1, 51]. In contrast, noise manipulation enables precise layout control, supporting layout-aware image generation and editing [25, 26, 43]. Building on these insights, our work introduces a novel method for shifting initial noise to control chromatic information. As a result, our method enables accurate chroma key backgrounds while preserving high-quality foreground content.", + "bbox": [ + 511, + 112, + 903, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Foreground and Background Separation in Image Generation", + "text_level": 1, + "bbox": [ + 511, + 381, + 903, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Foreground-background separation is crucial in advertising, design, and game development, allowing creators to manipulate individual image elements independently for tasks such as background replacement, foreground adjustments, and complex compositions [3, 9, 28, 29].", + "bbox": [ + 511, + 420, + 903, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to this challenge, MAGICK [5] employs a prompt engineering-based model to solve this task. However, as noted in this work: \"Unfortunately, all the publicly available methods we tested were incapable of consistently creating keyable images.\" To overcome this limitation, MAGICK combines multiple models and applies manual effort to release a foreground alpha image dataset. Similarly, LayerDiffuse [55] introduces a fine-tuning approach using a 1 million image dataset, which cannot be publicly shared due to licensing restrictions. Although these methods improve foreground-background separation, they require substantial data collection and computational resources, making them less efficient for real-time applications. In contrast, our training-free approach provides precise control over foreground and background elements without the need for extensive datasets and fine-tuning.", + "bbox": [ + 511, + 496, + 905, + 737 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 753, + 648, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 2, TKG-DM extends Stable Diffusion [37] by manipulating the initial Gaussian noise $\\mathbf{z}_T \\in \\mathbb{R}^{h \\times w \\times 4}$ through channel mean shift $F_c$ . This transformation produces the init color noise $\\mathbf{z}_T^* = F_c(z_T)$ , which guides the vanilla Stable Diffusion to generate a uniform color image $\\mathbf{x}_0^*$ without any text prompt. To generate an image where the foreground aligns with the input text prompt $p$ and the background has a specified color (e.g., green for chroma keying),", + "bbox": [ + 511, + 779, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13032", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1d26b4b49b5cc033332a1ab279c70e5e01db6c43c7bf27b21532ebd8c0e6d18e.jpg", + "image_caption": [ + "Figure 2. Starting with random noise $\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , init color noise $\\mathbf{z}_T^*$ is generated by applying channel mean shift $F_{c}$ . This produces a single-colored image $\\mathbf{x}_0^*$ without a prompt. By combining normal noise with init color noise via a Gaussian mask, TKG-DM generates a chroma key image $\\mathbf{x}_0^{key}$ with the specified foreground (e.g., \"the cat runs in the park\") over a uniform background, effectively separating the foreground from the monochromatic background." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 903, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we combine the initial noise $\\mathbf{z}_T$ and the init color noise $\\mathbf{z}_T^*$ using a Gaussian mask. This combined noise $\\mathbf{z}_T^{key}$ is input into the vanilla Stable Diffusion, generating the chroma key image $\\mathbf{x}_0^{key}$ .", + "bbox": [ + 89, + 325, + 483, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Channel Mean Shift", + "text_level": 1, + "bbox": [ + 89, + 398, + 282, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the relationship between Stable Diffusion's latent space and generated image color [47], we introduce a novel initial noise optimization technique, channel mean shift. It adjusts the mean of each channel in $z_{T}$ while keeping its standard deviation constant, enabling control over the generated image's color. To determine the optimal shift $\\Delta_{c}$ for each channel $c \\in \\{1,2,3,4\\}$ to achieve a target positive ratio in that channel. The values of channel $c$ in the noise tensor are denoted as $\\mathbf{z}_T^{(c)}$ . The initial positive ratio for channel $c$ is defined as:", + "bbox": [ + 89, + 419, + 483, + 570 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {I n i t i a l R a t i o} _ {c} = \\frac {\\sum_ {i , j} \\mathbb {1} \\left(\\mathbf {z} _ {T} ^ {(c)} (i , j) > 0\\right)}{\\text {T o t a l P i x e l s} _ {c}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 579, + 480, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is the indicator function, and TotalPikels $c$ is the total number of elements in $\\mathbf{z}_T^{(c)}$ . Given a target shift $\\mathrm{TargetShift}_{c}$ , the target positive ratio is:", + "bbox": [ + 89, + 627, + 483, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {T a r g e t R a t i o} _ {c} = \\text {I n i t i a l R a t i o} _ {c} + \\text {T a r g e t S h i f t} _ {c}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 686, + 482, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve TargetRatio, we iteratively adjust the mean shift $\\Delta_{c}$ for each channel $\\mathbf{z}_T^{(c)}$ . We initialize the shift with $\\Delta_c^{\\mathrm{init}} = 0$ and incrementally adjust $\\Delta_{c}$ until the positive ratio meets or exceeds TargetRatio. Once the target ratio is reached, we record the final shift as $\\Delta_c^{\\mathrm{final}} = \\Delta_c$ . The noise tensor obtained through this method is called init color noise $\\mathbf{z}_T^* = F_c(\\mathbf{z}_T) = \\mathbf{z}_T + \\Delta_c^{\\mathrm{final}}$ .", + "bbox": [ + 89, + 715, + 483, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Init Noise Selection", + "text_level": 1, + "bbox": [ + 89, + 832, + 274, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate the foreground object on the chroma key background, we apply an init noise selection strategy that selectively combines the initial noise $\\mathbf{z}_T$ and the init color noise", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{z}_T^*$ using a 2D Gaussian mask $\\mathbf{A}(i,j)$ . This mask creates a gradual transition by preserving the original noise in the foreground region and applying the color-shifted noise to the background region. Formally, the masked initial noise is computed as:", + "bbox": [ + 511, + 325, + 905, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {T} ^ {\\text {k e y}} (i, j) = \\mathbf {A} (i, j) \\cdot \\mathbf {z} _ {T} (i, j) + (1 - \\mathbf {A} (i, j)) \\cdot \\mathbf {z} _ {T} ^ {*} (i, j), (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 409, + 906, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{A}(i,j) = e^{-\\frac{(i - \\mu_i)^2 + (j - \\mu_j)^2}{2\\sigma^2}}$ controls the blend between the original noise and init color noise and $\\mu_{i}$ and $\\mu_{j}$ specify the center and $\\sigma$ the spread.", + "bbox": [ + 511, + 436, + 905, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adjusting the Gaussian parameters $\\mu_{i},\\mu_{j}$ , and $\\sigma$ controls the position and size of the foreground: shifting $\\mu_{i}$ and $\\mu_{j}$ adjusts its position while increasing $\\sigma$ enlarges the foreground element. Interestingly, TKG-DM also generates multiple foreground elements by applying multiple Gaussian masks, each independently configured to produce varied positions and sizes for flexible compositions.", + "bbox": [ + 511, + 491, + 905, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Mechanism of TKG-DM", + "text_level": 1, + "bbox": [ + 511, + 608, + 743, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we explain how TKG-DM specifies the background color and generates the foreground and background separately to generate chroma key content.", + "bbox": [ + 511, + 633, + 905, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Color Specify via Channel Mean Shift", + "text_level": 1, + "bbox": [ + 511, + 688, + 841, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by previous research [47], we control the chroma key background color by applying channel mean shift to specific channels of the initial noise $\\mathbf{z}_T\\in \\mathbb{R}^{h\\times w\\times 4}$ . Specifically, we adjust the mean of each channel $\\mathbf{z}_T^{(c)}$ , where $c\\in \\{1,2,3,4\\}$ , to influence the color composition of the generated images. In this experiment, we set $\\mathrm{TargetShift}_{c} = \\pm 7\\%$ , adjusting the positive ratio of each channel by adding or subtracting $7\\%$ relative to InitialRatio $c$ .", + "bbox": [ + 511, + 710, + 905, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3a, positive shifts in $\\mathbf{z}_T^{(2)}$ and $\\mathbf{z}_T^{(3)}$ intensify cyan and yellow tones, respectively, while negative shifts emphasize red and blue-purple hues. Shifts in $\\mathbf{z}_T^{(1)}$ and $\\mathbf{z}_T^{(4)}$ primarily affect luminance and shades of white and", + "bbox": [ + 511, + 832, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13033", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/423389d57a47f764fede64b680e78c16d7faa220ae1db65cc67015198f9000d3.jpg", + "image_caption": [ + "(a) Influence of Individual Noise Channel" + ], + "image_footnote": [], + "bbox": [ + 96, + 111, + 558, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b8d8b277e9060f3772d34bad1c878571c430aac3a92463e1de7c33b95bce926a.jpg", + "image_caption": [ + "(b) Additive and Subtractive Color Effects", + "Figure 3. Fig. 3a illustrates the relationship between positive/negative channel mean shift with initial noise and color variations in the generated images. Different channel shifts across noise channels result in changes in the hue of the generated image. Fig. 3b shows how simultaneous shifts across multiple channels facilitate additive and subtractive color mixing, providing intuitive and flexible color control." + ], + "image_footnote": [], + "bbox": [ + 648, + 112, + 903, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "black, respectively. These effects are confirmed by comparing generated images conditioned on specific text prompts.", + "bbox": [ + 89, + 401, + 480, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When we apply simultaneous shifts to multiple channels, we observe intuitive color mixing that aligns with additive and subtractive color theories. For example, positive shifts in both $\\mathbf{z}_T^{(2)}$ and $\\mathbf{z}_T^{(3)}$ enhance green tones, while combining red and yellow shifts with reduced luminance yields orange hues (see Fig. 3b). This channel-based manipulation enables flexible and natural color control in the image generation process.", + "bbox": [ + 89, + 433, + 482, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Content Separation via Attention Mechanism", + "text_level": 1, + "bbox": [ + 89, + 570, + 475, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4, self-attention and cross-attention mechanisms play distinct roles in generating the foreground and background elements, which is crucial for achieving effective foreground-background separation in TKG-DM.", + "bbox": [ + 89, + 594, + 482, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the foreground, self-attention ensures internal consistency and coherence within the object, while cross-attention aligns the generated content with the text prompt. Due to the inherent bias in the training dataset [42], where foreground objects are more prominently described in captions, cross-attention forms a strong link between the foreground and the text prompt.", + "bbox": [ + 89, + 656, + 482, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the background, the init color noise introduced by channel mean shift dominates the generation process. The self-attention mechanism synergizes with this modified initial noise to guide the background toward the specified chroma key color. Since background elements are often less detailed or vaguely described in dataset captions, the cross-attention has a weaker influence on the background, allowing the init color noise to take precedence.", + "bbox": [ + 89, + 763, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By exploiting this bias and manipulating the initial noise,", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7d698306951b410eb418d669d47aeb9ac86b3589bfdcb9e0c2ff32141b8efcbd.jpg", + "image_caption": [ + "Figure 4. Visualization of self- and cross-attention maps. The self-attention maps, segmented using a foreground segmentation model [58], show the regions attended to during generation. The cross-attention maps illustrate how the model attends to relevant regions based on the foreground and background prompts. This analysis shows how self-attention and cross-attention interact to influence foreground and background content generation." + ], + "image_footnote": [], + "bbox": [ + 519, + 402, + 901, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TKG-DM effectively decouples the background from the text prompt. This results in a uniform chroma key background and enables the isolated generation of foreground content without interference from undesired background details.", + "bbox": [ + 511, + 824, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13034", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ca613cf104e441ab8fcf022a19c10afbef3b02dbebaee9712508b1164381b14a.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison in SD1.5. Existing methods fail to produce an accurate chroma key background. In contrast, ours produces a highly accurate chroma key background while generating high-quality foregrounds without Green Background Prompt (GBP)." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 416, + 225, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 441, + 282, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt SD1.5 and SDXL in TKG-DM to generate images at $512 \\times 512$ and $1024 \\times 1024$ resolutions, respectively, with DDIMScheduler, a guidance scale of 7.5, and 50 denoising steps. To obtain the init color noise for a green background, we apply channel mean shift with channels 2 and 3 in the positive direction for $\\mathrm{TargetShift}_c = +7\\%$ . Gaussian mask parameters are set to $(\\sigma, \\mu_i, \\mu_j) = (0.5, w/2, h/2)$ .", + "bbox": [ + 89, + 464, + 483, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method against the naive usage of SD1.5 [37], SDXL [34], and DeepFloyd [38] using a Green Background Prompt (GBP) (=“isolated on a solid green background”) following the setup of MAGICK [5], as well as against fine-tuned models, namely GreenBack LoRA from CivitAI1 and LayerDiffuse [55].", + "bbox": [ + 89, + 570, + 483, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Dataset and Metrics", + "text_level": 1, + "bbox": [ + 89, + 671, + 284, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We collect 3,000 images from the MAGICK dataset [5], tailored for foreground image generation from text prompts. The MAGICK dataset is constructed by first generating monochromatic backgrounds using Deepfloyd [38] with GBP, followed by quality enhancement using SDEdit [27] and SDXL's img2img [34]. Foreground images with alpha values are curated using pixel-, deep learning-, and human-based methods. Our experiment utilizes alpha images, alpha masks, and corresponding text prompts.", + "bbox": [ + 89, + 694, + 482, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We construct ground truth images for evaluation by overlaying the alpha images onto a lime green background images $(\\mathrm{RGB} = 50,205,50)$ . We use the Fréchet Inception", + "bbox": [ + 89, + 830, + 483, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distance (FID) to assess foreground quality [17]. Additionally, we introduce a novel m-FID metric, which compares the mask extracted from generated chroma key content using BiRefNet [58] to the ground truth mask, providing a comprehensive assessment of foreground accuracy. For semantic alignment, we employ CLIP-Sentence (CLIP-S) and CLIP-Image (CLIP-I) [16]: CLIP-S measures alignment with the text, while CLIP-I assesses visual similarity to the ground truth.", + "bbox": [ + 511, + 417, + 906, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Qualitative Result", + "text_level": 1, + "bbox": [ + 511, + 561, + 691, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 5 and Fig. 6 show the qualitative results of our model and existing methods with SD1.5 and SDXL, respectively. They demonstrate that our model generates high-precision chroma key images without prompt engineering.", + "bbox": [ + 511, + 583, + 905, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the Green Background Prompt (GBP), SD1.5 struggles to generate a clean green background, while SDXL, while slightly better, generates unstable light green tints on the background and often produces multiple foreground objects, complicating foreground-background separation. Additionally, the GBP's \"green\" element unintentionally adds green tones to the foreground, such as on clothing or in a cat's eyes (Fig. 6, first and third columns). DeepFloyd performs the worst foreground quality and text alignment.", + "bbox": [ + 511, + 643, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the fine-tuned model, while LoRA (SD1.5) shows minor improvements, it still struggles with chroma key accuracy. LayerDiffuse produces well-separated foregrounds but occasionally loses details, like precise numbers or letters, due to dataset limitations in fine-tuning. Mask generation also occasionally fails, resulting in uncut images.", + "bbox": [ + 511, + 779, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, existing models struggle to generate accurate chroma key images when background information is in-", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://civitai.com/", + "bbox": [ + 109, + 886, + 274, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13035", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3fc01e815aaaf3c9bbdeaacf239995bcdd64f21a4e5786f0ceeca22212e59f95.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison in SDXL. Existing training-free methods fail to isolate foreground objects from the background. While LayerDiffuse, a fine-tuned model, successfully generates the transparent foregrounds on the lime green background, it may struggle with accurate text generation and background handling. In contrast, our model efficiently generates highly accurate chroma key images with foreground content placed over the background without fine-tuning and needing Green Background Prompt (GBP)." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 424 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd667350a80b006c543a8a1b73025b44ac0ecc6b74c151b7795daeb354d56a3f.jpg", + "image_caption": [ + "couple of petite chihuahua dogs", + "panorama with wine, cheese, grapes and olives", + "Figure 7. Qualitative various color result in SDXL. By adjusting the channel mean shift, TKG-DM controls the background color while generating a high-quality foreground." + ], + "image_footnote": [], + "bbox": [ + 96, + 527, + 477, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cluded in the text prompt (Fig. 6, last columns). In contrast, our training-free model consistently produces high-quality chroma key backgrounds with only the foreground content, ensuring precise separation.", + "bbox": [ + 89, + 772, + 482, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, as shown in Fig. 7, TKG-DM generates the various color background images through channel mean shift without prompt engineering. Additional results are provided in the supplementary material.", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4d2846bd5f732871a0c8b55526112bff0372d5c87accfa0e3a899bfcb9e0db47.jpg", + "table_caption": [ + "Table 1. Quantitative results. It indicates that our training-free model outperforms the existing model with Green Background Prompt (GBP) and rivals the fine-tuned model, LayerDiffuse [55]." + ], + "table_footnote": [], + "table_body": "
SD1.5
FID ↓m-FID ↓CLIP-I↑CLIP-S ↑
SD1.5 (GBP) [37]85.0063.540.7100.256
GB LoRA (GBP)60.2949.030.7040.243
Ours56.3240.750.7370.261
", + "bbox": [ + 514, + 569, + 901, + 630 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7ac3720b4e98344e4527e3baf2c29010dbaedc572e787a628b6166e3904075db.jpg", + "table_caption": [ + "SDXL" + ], + "table_footnote": [], + "table_body": "
DeepFloyd (GBP) [38]31.5720.310.7810.270
SDXL (GBP) [34]45.3239.170.7590.272
LayerDiffuse [55]29.3429.820.7780.276
Ours41.8131.430.7630.273
", + "bbox": [ + 516, + 638, + 901, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image Quality / Prompt Alignment", + "bbox": [ + 620, + 705, + 795, + 717 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2352ee960c758a4572cd6fa78f955c93f88a06a59ff42824f177d7b5978c9a2.jpg", + "image_caption": [ + "Figure 8. Results from our user study show the preference rates for foreground objects generated by our model and existing models based on their image quality and alignment with the text prompt. GBP means a model with a Green Background Prompt." + ], + "image_footnote": [], + "bbox": [ + 526, + 720, + 893, + 790 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13036", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abb121252802cb663d8cda58dc83b1de92b91edb53d751f26a85d735a95d52ee.jpg", + "image_caption": [ + "Figure 9. Qualitative comparison in ControlNET. Existing methods typically generate images with issues such as color erosion in the foreground, uneven backgrounds, or unintended elements from the conditioning input. In contrast, our model generates images with the foreground isolated from the background, maintaining a consistent chroma key background without the Green Background Prompt (GBP)." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Quantitative Result", + "text_level": 1, + "bbox": [ + 89, + 433, + 277, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 1, our method consistently outperforms other training-free models across all metrics, achieving better image realism and mask accuracy, as indicated by lower FID and m-FID scores. It also aligns well with input text prompts and has closer visual similarity to the ground truth, as indicated by higher CLIP-S and CLIP-I scores.", + "bbox": [ + 89, + 455, + 482, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the SDXL setting, LayerDiffuse achieves the best FID, while our model remains competitive across all other metrics. The FID advantage for LayerDiffuse likely results from its setup, which overlays foregrounds on a lime green background matching the ground truth.", + "bbox": [ + 89, + 547, + 483, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DeepFloyd's high FID, m-FID, and CLIP-I scores reflect its similarity to the ground truth based on DeepFloyd's outputs. However, this alignment gives it an inherent advantage, making it unsuitable as a fair benchmark for image quality. Its lower CLIP-S score further indicates weaker text alignment compared to other models.", + "bbox": [ + 89, + 623, + 483, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, these results underscore our model's ability to generate high-quality, text-aligned foregrounds without fine-tuning, offering an efficient chroma key content generation solution.", + "bbox": [ + 89, + 714, + 483, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. User Study", + "text_level": 1, + "bbox": [ + 89, + 786, + 212, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides objective metrics, we conducted a user study to assess subjective human preferences for prompt adherence and image quality, comparing our method against SDXL with Green Background Prompt (GBP) and LayerDiffuse. We generated 30 image pairs per method, totaling 60 images. 100 participants ranked images via two alternative", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce46d58c0eca4604600b989c5c9424a6b00fd34efa293234579ac44d23d824e7.jpg", + "table_caption": [ + "Table 2. Quantitative results comparing our method with existing methods for ControlNet [56]. GBP means a model with a Green Background Prompt." + ], + "table_footnote": [], + "table_body": "
ControlNet [56]
FID ↓m-FID ↓CLIP-I ↑CLIP-S ↑
SDXL (GBP) [34]22.0418.620.8190.279
Ours17.0917.220.8340.284
", + "bbox": [ + 514, + 484, + 906, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "forced-choice [30, 39] and based on foreground quality and text alignment. To ensure focus on the foreground, we used a foreground extraction model [58] and manually refined areas as needed. As shown in Fig. 8, our training-free approach improves object realism and text alignment over the existing model.", + "bbox": [ + 511, + 566, + 906, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Applications", + "text_level": 1, + "bbox": [ + 511, + 674, + 643, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the flexibility of TKG-DM in various tasks, implemented with the SD1.5 and SDXL, as presented in Table 2, Fig. 9 and Fig. 10. Further details and additional results are provided in the supplementary material.", + "bbox": [ + 511, + 700, + 905, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Application to ControlNet. As depicted in Fig. 9 and Table 2, our method integrates seamlessly with ControlNet [56], enabling precise generation control using conditioning inputs like canny edges. By applying a Gaussian mask to the init color noise, we preserve normal initial noise in the foreground regions, resulting in distinct foregrounds on uniform chroma key backgrounds.", + "bbox": [ + 511, + 762, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to the standard ControlNet, our method demonstrates superior qualitative and quantitative perfor", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13037", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2db3be692d6d8520c9b5b1e474d0dc6d001ad8d3716b3fb5ffe954963a1e15be.jpg", + "image_caption": [ + "Figure 10. Application results of TKG-DM across various tasks. TKG-DM effectively supports layout-aware text-to-image generation, consistency models and text-to-video. Each row demonstrates its versatility in adapting to different domains, from realistic object and character placements to animation." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 480, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance. Existing methods often face issues like color erosion in the foreground, uneven backgrounds, or unintended elements from the conditioning input. By isolating the foreground from background influence and maintaining a consistent chroma key background, our approach is highly effective for compositing tasks without prompt engineering.", + "bbox": [ + 89, + 455, + 482, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Layout-Aware Text-to-Image Generation. As discussed in Section 3.2, our model enables control over foreground objects' location and size by adjusting the Gaussian mask's center and spread. This control allows users to place objects precisely within specific areas of the generated image, making it practical for layout-aware tasks. Additionally, using multiple Gaussian masks enables the generation of multiple foreground objects within a single scene (Fig. 10, top-right). Unlike existing models that lack targeted control over object placement and size, our approach integrates layout information seamlessly without fine-tuning.", + "bbox": [ + 89, + 549, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Application to Consistency Models. Consistency Models [46] are generative models that produce high-quality outputs in a few steps, unlike traditional diffusion models. Operating in latent space, Consistency Models demonstrate the robustness of TKG-DM to generative models that handle images in the latent diffusion-based model, enabling significantly faster generation while maintaining high output quality without modifications.", + "bbox": [ + 89, + 717, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Application to Text-to-Video. Our TKG-DM extends seamlessly to text-to-video tasks, ensuring consistent foreground content across frames. We align foreground objects with the input prompt by applying TKG-DM with", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d4fd1c893556ffe0540ddac62225f74bf9bb022a2c0ded0d3286191af0c7707.jpg", + "image_caption": [ + "Figure 11. Failure Cases. TKG-DM fails to generate a foreground object when the size is too small or when the text prompt lacks object information, resulting in only the background being generated." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AnimateDiff [14] on each frame while maintaining a uniform background for chroma key video generation. This approach, the first to generate foreground video with controlled chroma key backgrounds, facilitates efficient and flexible video editing workflows involving background removal or replacement.", + "bbox": [ + 511, + 273, + 906, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitation and Future Work", + "text_level": 1, + "bbox": [ + 511, + 376, + 779, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 11, TKG-DM primarily focuses on generating foreground content (e.g., objects and people), which limits its ability to generate backgrounds like landscapes. Additionally, if the size parameter is too small, ignore generating the foreground object [26, 43].", + "bbox": [ + 511, + 402, + 905, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future enhancements could expand TKG-DM's capabilities to include controlled background generation, allowing separate and adjustable generation of both foreground and background content. This would greatly benefit applications in video and 3D content creation where isolated and reusable background assets are essential.", + "bbox": [ + 511, + 478, + 905, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 580, + 633, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presents the first exploration of initial noise color manipulation in diffusion models, demonstrating its effectiveness in controlling background chromaticity. Leveraging this insight, we introduce the Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which generates prompt-based foregrounds on a chroma key background without fine-tuning. TKG-DM extends seamlessly to tasks like consistency models and text-to-video, highlighting its versatility beyond static images. Experimental results show that our method matches or surpasses fine-tuned models in both qualitative and quantitative evaluations without additional training. This approach not only advances theoretical understanding but also provides practical benefits for real-world applications where isolating foreground content is essential.", + "bbox": [ + 511, + 604, + 906, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 845, + 681, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by BMBF project SustainML (Grant 101070408) and the BMDV NITD (45KI22B021).", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13038", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Boqing Gong, Cho-Jui Hsieh, and Minhao Cheng. The crystal ball hypothesis in diffusion models: Anticipating object positions from initial noise. arXiv preprint arXiv:2406.01970, 2024. 2", + "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation, 2023. 2", + "[3] Yaniv Benny and Lior Wolf. Onegan: Simultaneous unsupervised learning of conditional image generation, foreground segmentation, and fine-grained clustering. In ECCV, pages 514-530. Springer, 2020. 2", + "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 2", + "[5] Ryan D Burgert, Brian L Price, Jason Kuen, Yijun Li, and Michael S Ryoo. Magick: A large-scale captioned dataset from matting generated images using chroma keying. In CVPR, pages 22595-22604, 2024. 1, 2, 5", + "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 2", + "[7] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In CVPR, pages 18349-18358, 2023. 2", + "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 1, 2", + "[9] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In CVPR, pages 8724-8733, 2024. 2", + "[10] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 1", + "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 1, 2", + "[12] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. arXiv preprint arXiv:2406.04312, 2024. 2", + "[13] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, pages 9380-9389, 2024. 2", + "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 8", + "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 5", + "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2", + "[20] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In ICML, pages 13916-13932. PMLR, 2023. 2", + "[21] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Layoutdm: Discrete diffusion model for controllable layout generation. In CVPR, pages 10167-10176, 2023. 2", + "[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2", + "[23] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2", + "[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2", + "[25] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Guided image synthesis via initial image editing in diffusion model. arXiv preprint arXiv:2305.03382, 2023. 2", + "[26] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Semantic-driven initial image construction for guided image synthesis in diffusion model. arXiv preprint arXiv:2312.08872, 2023. 2, 8", + "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 5", + "[28] Ryugo Morita, Zhiqiang Zhang, Man M Ho, and Jinjia Zhou. Interactive image manipulation with complex text instructions. In WACV, pages 1053-1062, 2023. 2", + "[29] Ryugo Morita, Zhiqiang Zhang, and Jinjia Zhou. Batinet: Background-aware text to image synthesis and manipulation network. In ICIP, pages 765-769. IEEE, 2023. 2", + "[30] Brian B Moser, Stanislav Frolov, Federico Raue, Sebastian Palacio, and Andreas Dengel. Yoda: You only diffuse areas. an area-masked diffusion approach for image superresolution. arXiv preprint arXiv:2308.07977, 2023. 7", + "[31] Brian B Moser, Arundhati S Shanbhag, Federico Raue, Stanislav Frolov, Sebastian Palacio, and Andreas Dengel. Diffusion models, image super-resolution, and everything: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13039", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171. PMLR, 2021. 2", + "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 5, 6, 7", + "[35] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2", + "[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. Pmlr, 2021. 2", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 5, 6", + "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1, 2, 5, 6", + "[39] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 45(4): 4713-4726, 2022. 7", + "[40] Dvir Samuel, Rami Ben-Ari, Nir Darshan, Haggai Maron, and Gal Chechik. Norm-guided latent space exploration for text-to-image generation. NeurIPS, 36, 2024. 2", + "[41] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, pages 4695-4703, 2024. 2", + "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 4", + "[43] Takahiro Shirakawa and Seiichi Uchida. Noisecollapse: A layout-aware text-to-image diffusion model based on noise cropping and merging. In CVPR, pages 8921-8930, 2024. 2, 8", + "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015. 1" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2", + "[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023.8", + "[47] Timothy Alexis Vass. Explaining the sdxl latent space. https://huggingface.co/blog/TimothyAlexisVass/explaining-the-sdxl-latent-space, 2024.3", + "[48] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2", + "[49] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 1", + "[50] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 2", + "[51] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv preprint arXiv:2405.14828, 2024. 2", + "[52] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024. 1", + "[53] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, pages 18381-18391, 2023. 2", + "[54] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In CVPR, pages 14246-14255, 2023. 1", + "[55] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 1, 2, 5, 6", + "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 1, 7", + "[57] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, pages 22490-22499, 2023. 1", + "[58] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. arXiv preprint arXiv:2401.03407, 2024. 4, 5, 7" + ], + "bbox": [ + 514, + 90, + 905, + 854 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13040", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_model.json b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7bc88fc3b29fb920a010bce8317c3c8cfccd74e9 --- /dev/null +++ b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.88, + 0.152 + ], + "angle": 0, + "content": "TKG-DM: Training-free Chroma Key Content Generation Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.18, + 0.821, + 0.217 + ], + "angle": 0, + "content": "Ryugo Morita\\(^{1,2}\\), Stanislav Frolov\\(^{2}\\), Brian Bernhard Moser\\(^{2}\\), Takahiro Shirakawa, Ko Watanabe\\(^{2}\\), Andreas Dengel\\(^{2}\\), Jinjia Zhou\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.217, + 0.77, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Faculty of Science and Engineering, Hosei University, Tokyo, Japan" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.235, + 0.787, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)RPTU Kaiserslautern-Landau & DFKI GmbH, Kaiserslautern, Germany" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.651 + ], + "angle": 0, + "content": "Diffusion models have enabled the generation of high-quality images with a strong focus on realism and textual fidelity. Yet, large-scale text-to-image models, such as Stable Diffusion, struggle to generate images where foreground objects are placed over a chroma key background, limiting their ability to separate foreground and background elements without fine-tuning. To address this limitation, we present a novel Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which optimizes the initial random noise to produce images with foreground objects on a specifiable color background. Our proposed method is the first to explore the manipulation of the color aspects in initial noise for controlled background generation, enabling precise separation of foreground and background without fine-tuning. Extensive experiments demonstrate that our training-free method outperforms existing methods in both qualitative and quantitative evaluations, matching or surpassing fine-tuned models. Finally, we successfully extend it to other tasks (e.g., consistency models and text-to-video), highlighting its transformative potential across various generative applications where independent control of foreground and background is crucial." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.222, + 0.696 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Diffusion models have disrupted the landscape of image generation by producing high-quality images with high realism and textual fidelity [8, 11, 31, 38, 44]. Research has expanded into text-to-image generation conditioned on inputs like edges [49, 56], layouts [54, 57], and Virtual TryON (VTON) [10, 52]. However, real-world applications such as advertising, design, and game development often require precise control over foreground and background elements. The ability to generate images with transparent or chroma key backgrounds is essential for seamlessly integrating foreground objects into various scenes [55]. Traditional models like Stable Diffusion [37] and DeepFloyd [38] struggle to generate images where foreground and" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.285, + 0.9, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Figure 1. TKG-DM provides a training-free generation of foreground objects on a chroma key background, enabling independent control of foreground and background elements. Additionally, our method extends to conditional text-to-image and video generation tasks, making it suitable for creating foreground assets for various applications." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.614, + 0.907, + 0.689 + ], + "angle": 0, + "content": "background elements are separated without fine-tuning [5]. Specifically, these models cannot produce foreground objects over specifiable backgrounds or generate independent foreground and background layers, limiting their applicability in workflows requiring such precise control." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.909, + 0.856 + ], + "angle": 0, + "content": "Recent works like MAGICK [5] and LayerDiffuse [55] have significantly addressed this limitation. MAGICK introduces a prompt engineering-based approach for chroma key image generation. However, due to limited chroma key accuracy, it relies on post-processing and manual efforts to improve quality, including the release of foreground alpha image datasets. LayerDiffuse proposes a fine-tuning method for layer-wise generation using a dataset with 1 million images. Unfortunately, this dataset is not publicly available due to licensing restrictions, necessitating further resource-intensive dataset curation and fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To overcome these challenges, we present the Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which optimizes the initial noise in the dif" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13031" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.287 + ], + "angle": 0, + "content": "fusion process to produce images with foreground objects on specifiable color backgrounds without fine-tuning. Our method is the first to explore manipulating the color aspects of initial noise for controlled background generation, enabling precise separation of foreground and background elements. TKG-DM offers high flexibility and provides precise control over the background color, layout, size, and number of foreground objects. Our approach expands possibilities in generative content by maintaining the independence of the foreground and chroma key background. It seamlessly extends to applications like conditional text-to-image generation, consistency models, and text-to-video generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.482, + 0.395 + ], + "angle": 0, + "content": "Our extensive experiments demonstrate that TKG-DM improves FID and mask-FID scores by \\(33.7\\%\\) and \\(35.9\\%\\), respectively. Thus, our training-free model rivals fine-tuned models, offering an efficient and versatile solution for various visual content creation tasks that require precise foreground and background control. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.399, + 0.483, + 0.444 + ], + "angle": 0, + "content": "- We introduce TKG-DM, a training-free diffusion model that eliminates the need for fine-tuning and datasets for controllable foreground and background generation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.444, + 0.483, + 0.503 + ], + "angle": 0, + "content": "- TKG-DM provides precise control over background color, as well as the size, position and number of foreground elements, enabling flexible monochromatic chroma key generation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.505, + 0.483, + 0.564 + ], + "angle": 0, + "content": "- Our method is highly versatile, seamlessly extending to conditional text-to-image generation, consistency models, and text-to-video applications, thus enabling the creation of diverse chroma key content." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.565, + 0.483, + 0.61 + ], + "angle": 0, + "content": "- TKG-DM outperforms existing models, delivering results that match or surpass fine-tuned alternatives while maintaining lower computational costs." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.399, + 0.483, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.627, + 0.24, + 0.642 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.259, + 0.667 + ], + "angle": 0, + "content": "2.1. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Diffusion models [19, 33] generate realistic, high-quality images by iteratively refining Gaussian noise to approximate a target distribution [11, 45]. Building upon this, Latent Diffusion Models (LDMs) [37] improve efficiency by exploiting the low-dimensional latent spaces while preserving image quality. In particular, text-to-image models [32, 36-38] align text and image features through crossattention to achieve diverse, high-quality outputs, with classifier-free guidance [18] and attention control [6, 8] further enhancing text fidelity. Beyond image generation, diffusion models find use in inpainting [24, 53], image editing [4, 15], and layout generation [7, 21]. Their versatility extends across multiple modalities, including audio generation [20, 22], text-to-video [14, 50], and 3D object generation [35, 48]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.644, + 0.105 + ], + "angle": 0, + "content": "2.2. Initial Noise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.905, + 0.372 + ], + "angle": 0, + "content": "Diffusion models typically start with Gaussian noise and iteratively refine it to generate high-quality images that align closely with a given target distribution. In general, the choice and optimization of this initial noise are crucial, as they significantly impact image quality and alignment [40, 41, 51]. More specifically, noise optimization methods, such as human preference models or attention-based scores, can enhance fidelity and text alignment without additional training [12, 13]. Moreover, optimal noise seeds and targeted noise regions significantly impact image quality and object placement [1, 51]. In contrast, noise manipulation enables precise layout control, supporting layout-aware image generation and editing [25, 26, 43]. Building on these insights, our work introduces a novel method for shifting initial noise to control chromatic information. As a result, our method enables accurate chroma key backgrounds while preserving high-quality foreground content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.382, + 0.905, + 0.414 + ], + "angle": 0, + "content": "2.3. Foreground and Background Separation in Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.905, + 0.496 + ], + "angle": 0, + "content": "Foreground-background separation is crucial in advertising, design, and game development, allowing creators to manipulate individual image elements independently for tasks such as background replacement, foreground adjustments, and complex compositions [3, 9, 28, 29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.738 + ], + "angle": 0, + "content": "In response to this challenge, MAGICK [5] employs a prompt engineering-based model to solve this task. However, as noted in this work: \"Unfortunately, all the publicly available methods we tested were incapable of consistently creating keyable images.\" To overcome this limitation, MAGICK combines multiple models and applies manual effort to release a foreground alpha image dataset. Similarly, LayerDiffuse [55] introduces a fine-tuning approach using a 1 million image dataset, which cannot be publicly shared due to licensing restrictions. Although these methods improve foreground-background separation, they require substantial data collection and computational resources, making them less efficient for real-time applications. In contrast, our training-free approach provides precise control over foreground and background elements without the need for extensive datasets and fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.649, + 0.772 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 2, TKG-DM extends Stable Diffusion [37] by manipulating the initial Gaussian noise \\(\\mathbf{z}_T \\in \\mathbb{R}^{h \\times w \\times 4}\\) through channel mean shift \\(F_c\\). This transformation produces the init color noise \\(\\mathbf{z}_T^* = F_c(z_T)\\), which guides the vanilla Stable Diffusion to generate a uniform color image \\(\\mathbf{x}_0^*\\) without any text prompt. To generate an image where the foreground aligns with the input text prompt \\(p\\) and the background has a specified color (e.g., green for chroma keying)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13032" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.905, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.244, + 0.908, + 0.302 + ], + "angle": 0, + "content": "Figure 2. Starting with random noise \\(\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), init color noise \\(\\mathbf{z}_T^*\\) is generated by applying channel mean shift \\(F_{c}\\). This produces a single-colored image \\(\\mathbf{x}_0^*\\) without a prompt. By combining normal noise with init color noise via a Gaussian mask, TKG-DM generates a chroma key image \\(\\mathbf{x}_0^{key}\\) with the specified foreground (e.g., \"the cat runs in the park\") over a uniform background, effectively separating the foreground from the monochromatic background." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.391 + ], + "angle": 0, + "content": "we combine the initial noise \\(\\mathbf{z}_T\\) and the init color noise \\(\\mathbf{z}_T^*\\) using a Gaussian mask. This combined noise \\(\\mathbf{z}_T^{key}\\) is input into the vanilla Stable Diffusion, generating the chroma key image \\(\\mathbf{x}_0^{key}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.399, + 0.283, + 0.413 + ], + "angle": 0, + "content": "3.1. Channel Mean Shift" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.484, + 0.571 + ], + "angle": 0, + "content": "Inspired by the relationship between Stable Diffusion's latent space and generated image color [47], we introduce a novel initial noise optimization technique, channel mean shift. It adjusts the mean of each channel in \\( z_{T} \\) while keeping its standard deviation constant, enabling control over the generated image's color. To determine the optimal shift \\( \\Delta_{c} \\) for each channel \\( c \\in \\{1,2,3,4\\} \\) to achieve a target positive ratio in that channel. The values of channel \\( c \\) in the noise tensor are denoted as \\( \\mathbf{z}_T^{(c)} \\). The initial positive ratio for channel \\( c \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.58, + 0.482, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\text {I n i t i a l R a t i o} _ {c} = \\frac {\\sum_ {i , j} \\mathbb {1} \\left(\\mathbf {z} _ {T} ^ {(c)} (i , j) > 0\\right)}{\\text {T o t a l P i x e l s} _ {c}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.677 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is the indicator function, and TotalPikels \\(c\\) is the total number of elements in \\(\\mathbf{z}_T^{(c)}\\). Given a target shift \\(\\mathrm{TargetShift}_{c}\\), the target positive ratio is:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.688, + 0.483, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\text {T a r g e t R a t i o} _ {c} = \\text {I n i t i a l R a t i o} _ {c} + \\text {T a r g e t S h i f t} _ {c}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.484, + 0.826 + ], + "angle": 0, + "content": "To achieve TargetRatio, we iteratively adjust the mean shift \\(\\Delta_{c}\\) for each channel \\(\\mathbf{z}_T^{(c)}\\). We initialize the shift with \\(\\Delta_c^{\\mathrm{init}} = 0\\) and incrementally adjust \\(\\Delta_{c}\\) until the positive ratio meets or exceeds TargetRatio. Once the target ratio is reached, we record the final shift as \\(\\Delta_c^{\\mathrm{final}} = \\Delta_c\\). The noise tensor obtained through this method is called init color noise \\(\\mathbf{z}_T^* = F_c(\\mathbf{z}_T) = \\mathbf{z}_T + \\Delta_c^{\\mathrm{final}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.276, + 0.849 + ], + "angle": 0, + "content": "3.2. Init Noise Selection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To generate the foreground object on the chroma key background, we apply an init noise selection strategy that selectively combines the initial noise \\(\\mathbf{z}_T\\) and the init color noise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.906, + 0.403 + ], + "angle": 0, + "content": "\\(\\mathbf{z}_T^*\\) using a 2D Gaussian mask \\(\\mathbf{A}(i,j)\\). This mask creates a gradual transition by preserving the original noise in the foreground region and applying the color-shifted noise to the background region. Formally, the masked initial noise is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.41, + 0.907, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {T} ^ {\\text {k e y}} (i, j) = \\mathbf {A} (i, j) \\cdot \\mathbf {z} _ {T} (i, j) + (1 - \\mathbf {A} (i, j)) \\cdot \\mathbf {z} _ {T} ^ {*} (i, j), (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.906, + 0.492 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}(i,j) = e^{-\\frac{(i - \\mu_i)^2 + (j - \\mu_j)^2}{2\\sigma^2}}\\) controls the blend between the original noise and init color noise and \\(\\mu_{i}\\) and \\(\\mu_{j}\\) specify the center and \\(\\sigma\\) the spread." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.598 + ], + "angle": 0, + "content": "Adjusting the Gaussian parameters \\(\\mu_{i},\\mu_{j}\\), and \\(\\sigma\\) controls the position and size of the foreground: shifting \\(\\mu_{i}\\) and \\(\\mu_{j}\\) adjusts its position while increasing \\(\\sigma\\) enlarges the foreground element. Interestingly, TKG-DM also generates multiple foreground elements by applying multiple Gaussian masks, each independently configured to produce varied positions and sizes for flexible compositions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.609, + 0.745, + 0.625 + ], + "angle": 0, + "content": "4. Mechanism of TKG-DM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.906, + 0.681 + ], + "angle": 0, + "content": "In this section, we explain how TKG-DM specifies the background color and generates the foreground and background separately to generate chroma key content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.689, + 0.842, + 0.705 + ], + "angle": 0, + "content": "4.1. Color Specify via Channel Mean Shift" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.906, + 0.833 + ], + "angle": 0, + "content": "Inspired by previous research [47], we control the chroma key background color by applying channel mean shift to specific channels of the initial noise \\(\\mathbf{z}_T\\in \\mathbb{R}^{h\\times w\\times 4}\\). Specifically, we adjust the mean of each channel \\(\\mathbf{z}_T^{(c)}\\), where \\(c\\in \\{1,2,3,4\\}\\), to influence the color composition of the generated images. In this experiment, we set \\(\\mathrm{TargetShift}_{c} = \\pm 7\\%\\), adjusting the positive ratio of each channel by adding or subtracting \\(7\\%\\) relative to InitialRatio \\(c\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.906, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 3a, positive shifts in \\(\\mathbf{z}_T^{(2)}\\) and \\(\\mathbf{z}_T^{(3)}\\) intensify cyan and yellow tones, respectively, while negative shifts emphasize red and blue-purple hues. Shifts in \\(\\mathbf{z}_T^{(1)}\\) and \\(\\mathbf{z}_T^{(4)}\\) primarily affect luminance and shades of white and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.112, + 0.56, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.309, + 0.451, + 0.32 + ], + "angle": 0, + "content": "(a) Influence of Individual Noise Channel" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.113, + 0.904, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.309, + 0.868, + 0.32 + ], + "angle": 0, + "content": "(b) Additive and Subtractive Color Effects" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.333, + 0.907, + 0.377 + ], + "angle": 0, + "content": "Figure 3. Fig. 3a illustrates the relationship between positive/negative channel mean shift with initial noise and color variations in the generated images. Different channel shifts across noise channels result in changes in the hue of the generated image. Fig. 3b shows how simultaneous shifts across multiple channels facilitate additive and subtractive color mixing, providing intuitive and flexible color control." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.482, + 0.432 + ], + "angle": 0, + "content": "black, respectively. These effects are confirmed by comparing generated images conditioned on specific text prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.483, + 0.557 + ], + "angle": 0, + "content": "When we apply simultaneous shifts to multiple channels, we observe intuitive color mixing that aligns with additive and subtractive color theories. For example, positive shifts in both \\(\\mathbf{z}_T^{(2)}\\) and \\(\\mathbf{z}_T^{(3)}\\) enhance green tones, while combining red and yellow shifts with reduced luminance yields orange hues (see Fig. 3b). This channel-based manipulation enables flexible and natural color control in the image generation process." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.571, + 0.476, + 0.587 + ], + "angle": 0, + "content": "4.2. Content Separation via Attention Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.656 + ], + "angle": 0, + "content": "As illustrated in Fig. 4, self-attention and cross-attention mechanisms play distinct roles in generating the foreground and background elements, which is crucial for achieving effective foreground-background separation in TKG-DM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.762 + ], + "angle": 0, + "content": "For the foreground, self-attention ensures internal consistency and coherence within the object, while cross-attention aligns the generated content with the text prompt. Due to the inherent bias in the training dataset [42], where foreground objects are more prominently described in captions, cross-attention forms a strong link between the foreground and the text prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.884 + ], + "angle": 0, + "content": "For the background, the init color noise introduced by channel mean shift dominates the generation process. The self-attention mechanism synergizes with this modified initial noise to guide the background toward the specified chroma key color. Since background elements are often less detailed or vaguely described in dataset captions, the cross-attention has a weaker influence on the background, allowing the init color noise to take precedence." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "By exploiting this bias and manipulating the initial noise," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.403, + 0.902, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.69, + 0.907, + 0.788 + ], + "angle": 0, + "content": "Figure 4. Visualization of self- and cross-attention maps. The self-attention maps, segmented using a foreground segmentation model [58], show the regions attended to during generation. The cross-attention maps illustrate how the model attends to relevant regions based on the foreground and background prompts. This analysis shows how self-attention and cross-attention interact to influence foreground and background content generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.9 + ], + "angle": 0, + "content": "TKG-DM effectively decouples the background from the text prompt. This results in a uniform chroma key background and enables the isolated generation of foreground content without interference from undesired background details." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13034" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.364, + 0.908, + 0.395 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison in SD1.5. Existing methods fail to produce an accurate chroma key background. In contrast, ours produces a highly accurate chroma key background while generating high-quality foregrounds without Green Background Prompt (GBP)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.417, + 0.226, + 0.435 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.442, + 0.283, + 0.459 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.571 + ], + "angle": 0, + "content": "We adopt SD1.5 and SDXL in TKG-DM to generate images at \\(512 \\times 512\\) and \\(1024 \\times 1024\\) resolutions, respectively, with DDIMScheduler, a guidance scale of 7.5, and 50 denoising steps. To obtain the init color noise for a green background, we apply channel mean shift with channels 2 and 3 in the positive direction for \\(\\mathrm{TargetShift}_c = +7\\%\\). Gaussian mask parameters are set to \\((\\sigma, \\mu_i, \\mu_j) = (0.5, w/2, h/2)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.484, + 0.662 + ], + "angle": 0, + "content": "We compare our method against the naive usage of SD1.5 [37], SDXL [34], and DeepFloyd [38] using a Green Background Prompt (GBP) (=“isolated on a solid green background”) following the setup of MAGICK [5], as well as against fine-tuned models, namely GreenBack LoRA from CivitAI1 and LayerDiffuse [55]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.285, + 0.687 + ], + "angle": 0, + "content": "5.2. Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.831 + ], + "angle": 0, + "content": "We collect 3,000 images from the MAGICK dataset [5], tailored for foreground image generation from text prompts. The MAGICK dataset is constructed by first generating monochromatic backgrounds using Deepfloyd [38] with GBP, followed by quality enhancement using SDEdit [27] and SDXL's img2img [34]. Foreground images with alpha values are curated using pixel-, deep learning-, and human-based methods. Our experiment utilizes alpha images, alpha masks, and corresponding text prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.832, + 0.484, + 0.878 + ], + "angle": 0, + "content": "We construct ground truth images for evaluation by overlaying the alpha images onto a lime green background images \\((\\mathrm{RGB} = 50,205,50)\\). We use the Fréchet Inception" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Distance (FID) to assess foreground quality [17]. Additionally, we introduce a novel m-FID metric, which compares the mask extracted from generated chroma key content using BiRefNet [58] to the ground truth mask, providing a comprehensive assessment of foreground accuracy. For semantic alignment, we employ CLIP-Sentence (CLIP-S) and CLIP-Image (CLIP-I) [16]: CLIP-S measures alignment with the text, while CLIP-I assesses visual similarity to the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.692, + 0.579 + ], + "angle": 0, + "content": "5.3. Qualitative Result" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.645 + ], + "angle": 0, + "content": "Fig. 5 and Fig. 6 show the qualitative results of our model and existing methods with SD1.5 and SDXL, respectively. They demonstrate that our model generates high-precision chroma key images without prompt engineering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.78 + ], + "angle": 0, + "content": "With the Green Background Prompt (GBP), SD1.5 struggles to generate a clean green background, while SDXL, while slightly better, generates unstable light green tints on the background and often produces multiple foreground objects, complicating foreground-background separation. Additionally, the GBP's \"green\" element unintentionally adds green tones to the foreground, such as on clothing or in a cat's eyes (Fig. 6, first and third columns). DeepFloyd performs the worst foreground quality and text alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.871 + ], + "angle": 0, + "content": "In the fine-tuned model, while LoRA (SD1.5) shows minor improvements, it still struggles with chroma key accuracy. LayerDiffuse produces well-separated foregrounds but occasionally loses details, like precise numbers or letters, due to dataset limitations in fine-tuning. Mask generation also occasionally fails, resulting in uncut images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Moreover, existing models struggle to generate accurate chroma key images when background information is in-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.276, + 0.9 + ], + "angle": 0, + "content": "1https://civitai.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13035" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.437, + 0.908, + 0.496 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison in SDXL. Existing training-free methods fail to isolate foreground objects from the background. While LayerDiffuse, a fine-tuned model, successfully generates the transparent foregrounds on the lime green background, it may struggle with accurate text generation and background handling. In contrast, our model efficiently generates highly accurate chroma key images with foreground content placed over the background without fine-tuning and needing Green Background Prompt (GBP)." + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.519, + 0.262, + 0.529 + ], + "angle": 0, + "content": "couple of petite chihuahua dogs" + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.519, + 0.478, + 0.529 + ], + "angle": 0, + "content": "panorama with wine, cheese, grapes and olives" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.529, + 0.478, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.689, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Figure 7. Qualitative various color result in SDXL. By adjusting the channel mean shift, TKG-DM controls the background color while generating a high-quality foreground." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.835 + ], + "angle": 0, + "content": "cluded in the text prompt (Fig. 6, last columns). In contrast, our training-free model consistently produces high-quality chroma key backgrounds with only the foreground content, ensuring precise separation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Furthermore, as shown in Fig. 7, TKG-DM generates the various color background images through channel mean shift without prompt engineering. Additional results are provided in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.518, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Table 1. Quantitative results. It indicates that our training-free model outperforms the existing model with Green Background Prompt (GBP) and rivals the fine-tuned model, LayerDiffuse [55]." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.57, + 0.902, + 0.631 + ], + "angle": 0, + "content": "
SD1.5
FID ↓m-FID ↓CLIP-I↑CLIP-S ↑
SD1.5 (GBP) [37]85.0063.540.7100.256
GB LoRA (GBP)60.2949.030.7040.243
Ours56.3240.750.7370.261
" + }, + { + "type": "table_caption", + "bbox": [ + 0.692, + 0.631, + 0.725, + 0.64 + ], + "angle": 0, + "content": "SDXL" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.64, + 0.902, + 0.692 + ], + "angle": 0, + "content": "
DeepFloyd (GBP) [38]31.5720.310.7810.270
SDXL (GBP) [34]45.3239.170.7590.272
LayerDiffuse [55]29.3429.820.7780.276
Ours41.8131.430.7630.273
" + }, + { + "type": "table_caption", + "bbox": [ + 0.622, + 0.707, + 0.796, + 0.718 + ], + "angle": 0, + "content": "Image Quality / Prompt Alignment" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.721, + 0.895, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.807, + 0.907, + 0.864 + ], + "angle": 0, + "content": "Figure 8. Results from our user study show the preference rates for foreground objects generated by our model and existing models based on their image quality and alignment with the text prompt. GBP means a model with a Green Background Prompt." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13036" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.366, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparison in ControlNET. Existing methods typically generate images with issues such as color erosion in the foreground, uneven backgrounds, or unintended elements from the conditioning input. In contrast, our model generates images with the foreground isolated from the background, maintaining a consistent chroma key background without the Green Background Prompt (GBP)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.434, + 0.279, + 0.45 + ], + "angle": 0, + "content": "5.4. Quantitative Result" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.547 + ], + "angle": 0, + "content": "As shown in Table 1, our method consistently outperforms other training-free models across all metrics, achieving better image realism and mask accuracy, as indicated by lower FID and m-FID scores. It also aligns well with input text prompts and has closer visual similarity to the ground truth, as indicated by higher CLIP-S and CLIP-I scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.623 + ], + "angle": 0, + "content": "In the SDXL setting, LayerDiffuse achieves the best FID, while our model remains competitive across all other metrics. The FID advantage for LayerDiffuse likely results from its setup, which overlays foregrounds on a lime green background matching the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.715 + ], + "angle": 0, + "content": "DeepFloyd's high FID, m-FID, and CLIP-I scores reflect its similarity to the ground truth based on DeepFloyd's outputs. However, this alignment gives it an inherent advantage, making it unsuitable as a fair benchmark for image quality. Its lower CLIP-S score further indicates weaker text alignment compared to other models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.484, + 0.775 + ], + "angle": 0, + "content": "Overall, these results underscore our model's ability to generate high-quality, text-aligned foregrounds without fine-tuning, offering an efficient chroma key content generation solution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.214, + 0.804 + ], + "angle": 0, + "content": "5.5. User Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Besides objective metrics, we conducted a user study to assess subjective human preferences for prompt adherence and image quality, comparing our method against SDXL with Green Background Prompt (GBP) and LayerDiffuse. We generated 30 image pairs per method, totaling 60 images. 100 participants ranked images via two alternative" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.433, + 0.907, + 0.474 + ], + "angle": 0, + "content": "Table 2. Quantitative results comparing our method with existing methods for ControlNet [56]. GBP means a model with a Green Background Prompt." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.485, + 0.907, + 0.54 + ], + "angle": 0, + "content": "
ControlNet [56]
FID ↓m-FID ↓CLIP-I ↑CLIP-S ↑
SDXL (GBP) [34]22.0418.620.8190.279
Ours17.0917.220.8340.284
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.658 + ], + "angle": 0, + "content": "forced-choice [30, 39] and based on foreground quality and text alignment. To ensure focus on the foreground, we used a foreground extraction model [58] and manually refined areas as needed. As shown in Fig. 8, our training-free approach improves object realism and text alignment over the existing model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.675, + 0.644, + 0.693 + ], + "angle": 0, + "content": "6. Applications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.906, + 0.763 + ], + "angle": 0, + "content": "We demonstrate the flexibility of TKG-DM in various tasks, implemented with the SD1.5 and SDXL, as presented in Table 2, Fig. 9 and Fig. 10. Further details and additional results are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Application to ControlNet. As depicted in Fig. 9 and Table 2, our method integrates seamlessly with ControlNet [56], enabling precise generation control using conditioning inputs like canny edges. By applying a Gaussian mask to the init color noise, we preserve normal initial noise in the foreground regions, resulting in distinct foregrounds on uniform chroma key backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Compared to the standard ControlNet, our method demonstrates superior qualitative and quantitative perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13037" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.482, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.357, + 0.485, + 0.427 + ], + "angle": 0, + "content": "Figure 10. Application results of TKG-DM across various tasks. TKG-DM effectively supports layout-aware text-to-image generation, consistency models and text-to-video. Each row demonstrates its versatility in adapting to different domains, from realistic object and character placements to animation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.548 + ], + "angle": 0, + "content": "mance. Existing methods often face issues like color erosion in the foreground, uneven backgrounds, or unintended elements from the conditioning input. By isolating the foreground from background influence and maintaining a consistent chroma key background, our approach is highly effective for compositing tasks without prompt engineering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Layout-Aware Text-to-Image Generation. As discussed in Section 3.2, our model enables control over foreground objects' location and size by adjusting the Gaussian mask's center and spread. This control allows users to place objects precisely within specific areas of the generated image, making it practical for layout-aware tasks. Additionally, using multiple Gaussian masks enables the generation of multiple foreground objects within a single scene (Fig. 10, top-right). Unlike existing models that lack targeted control over object placement and size, our approach integrates layout information seamlessly without fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Application to Consistency Models. Consistency Models [46] are generative models that produce high-quality outputs in a few steps, unlike traditional diffusion models. Operating in latent space, Consistency Models demonstrate the robustness of TKG-DM to generative models that handle images in the latent diffusion-based model, enabling significantly faster generation while maintaining high output quality without modifications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Application to Text-to-Video. Our TKG-DM extends seamlessly to text-to-video tasks, ensuring consistent foreground content across frames. We align foreground objects with the input prompt by applying TKG-DM with" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.208, + 0.907, + 0.251 + ], + "angle": 0, + "content": "Figure 11. Failure Cases. TKG-DM fails to generate a foreground object when the size is too small or when the text prompt lacks object information, resulting in only the background being generated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.366 + ], + "angle": 0, + "content": "AnimateDiff [14] on each frame while maintaining a uniform background for chroma key video generation. This approach, the first to generate foreground video with controlled chroma key backgrounds, facilitates efficient and flexible video editing workflows involving background removal or replacement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.377, + 0.78, + 0.393 + ], + "angle": 0, + "content": "7. Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.478 + ], + "angle": 0, + "content": "As shown in Fig. 11, TKG-DM primarily focuses on generating foreground content (e.g., objects and people), which limits its ability to generate backgrounds like landscapes. Additionally, if the size parameter is too small, ignore generating the foreground object [26, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.906, + 0.569 + ], + "angle": 0, + "content": "Future enhancements could expand TKG-DM's capabilities to include controlled background generation, allowing separate and adjustable generation of both foreground and background content. This would greatly benefit applications in video and 3D content creation where isolated and reusable background assets are essential." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.581, + 0.634, + 0.597 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.907, + 0.833 + ], + "angle": 0, + "content": "This work presents the first exploration of initial noise color manipulation in diffusion models, demonstrating its effectiveness in controlling background chromaticity. Leveraging this insight, we introduce the Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which generates prompt-based foregrounds on a chroma key background without fine-tuning. TKG-DM extends seamlessly to tasks like consistency models and text-to-video, highlighting its versatility beyond static images. Experimental results show that our method matches or surpasses fine-tuned models in both qualitative and quantitative evaluations without additional training. This approach not only advances theoretical understanding but also provides practical benefits for real-world applications where isolating foreground content is essential." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.846, + 0.682, + 0.863 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This research was supported by BMBF project SustainML (Grant 101070408) and the BMDV NITD (45KI22B021)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13038" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Boqing Gong, Cho-Jui Hsieh, and Minhao Cheng. The crystal ball hypothesis in diffusion models: Anticipating object positions from initial noise. arXiv preprint arXiv:2406.01970, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Yaniv Benny and Lior Wolf. Onegan: Simultaneous unsupervised learning of conditional image generation, foreground segmentation, and fine-grained clustering. In ECCV, pages 514-530. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[5] Ryan D Burgert, Brian L Price, Jason Kuen, Yijun Li, and Michael S Ryoo. Magick: A large-scale captioned dataset from matting generated images using chroma keying. In CVPR, pages 22595-22604, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[7] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In CVPR, pages 18349-18358, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[9] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In CVPR, pages 8724-8733, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.578, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[10] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[12] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. arXiv preprint arXiv:2406.04312, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[13] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, pages 9380-9389, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.273, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[20] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In ICML, pages 13916-13932. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[21] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Layoutdm: Discrete diffusion model for controllable layout generation. In CVPR, pages 10167-10176, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[23] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[25] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Guided image synthesis via initial image editing in diffusion model. arXiv preprint arXiv:2305.03382, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[26] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Semantic-driven initial image construction for guided image synthesis in diffusion model. arXiv preprint arXiv:2312.08872, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.691, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[28] Ryugo Morita, Zhiqiang Zhang, Man M Ho, and Jinjia Zhou. Interactive image manipulation with complex text instructions. In WACV, pages 1053-1062, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[29] Ryugo Morita, Zhiqiang Zhang, and Jinjia Zhou. Batinet: Background-aware text to image synthesis and manipulation network. In ICIP, pages 765-769. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[30] Brian B Moser, Stanislav Frolov, Federico Raue, Sebastian Palacio, and Andreas Dengel. Yoda: You only diffuse areas. an area-masked diffusion approach for image superresolution. arXiv preprint arXiv:2308.07977, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[31] Brian B Moser, Arundhati S Shanbhag, Federico Raue, Stanislav Frolov, Sebastian Palacio, and Andreas Dengel. Diffusion models, image super-resolution, and everything: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13039" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.355 + ], + "angle": 0, + "content": "[35] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. Pmlr, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 45(4): 4713-4726, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[40] Dvir Samuel, Rami Ben-Ari, Nir Darshan, Haggai Maron, and Gal Chechik. Norm-guided latent space exploration for text-to-image generation. NeurIPS, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[41] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, pages 4695-4703, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[43] Takahiro Shirakawa and Seiichi Uchida. Noisecollapse: A layout-aware text-to-image diffusion model based on noise cropping and merging. In CVPR, pages 8921-8930, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[47] Timothy Alexis Vass. Explaining the sdxl latent space. https://huggingface.co/blog/TimothyAlexisVass/explaining-the-sdxl-latent-space, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[48] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.359 + ], + "angle": 0, + "content": "[49] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[50] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.487 + ], + "angle": 0, + "content": "[51] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv preprint arXiv:2405.14828, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.906, + 0.543 + ], + "angle": 0, + "content": "[52] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.906, + 0.601 + ], + "angle": 0, + "content": "[53] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, pages 18381-18391, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[54] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In CVPR, pages 14246-14255, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.906, + 0.699 + ], + "angle": 0, + "content": "[55] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.906, + 0.799 + ], + "angle": 0, + "content": "[57] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, pages 22490-22499, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[58] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. arXiv preprint arXiv:2401.03407, 2024. 4, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13040" + } + ] +] \ No newline at end of file diff --git a/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_origin.pdf b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..263f139a5f9af139fe2c9ce32884ce6c1d83975a --- /dev/null +++ b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/95b0d4d9-07cb-43c6-ad69-b00fd2844ccb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58686982fa4f47f106e3d57042a0c2f988f67a9670d6009165422cf8701c1163 +size 2193533 diff --git a/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/full.md b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..50f9839eb3a56675e7306f6245a363fc9b704ee2 --- /dev/null +++ b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/full.md @@ -0,0 +1,303 @@ +# TKG-DM: Training-free Chroma Key Content Generation Diffusion Model + +Ryugo Morita $^{1,2}$ , Stanislav Frolov $^{2}$ , Brian Bernhard Moser $^{2}$ , Takahiro Shirakawa, Ko Watanabe $^{2}$ , Andreas Dengel $^{2}$ , Jinjia Zhou $^{1}$ + +$^{1}$ Faculty of Science and Engineering, Hosei University, Tokyo, Japan + +$^{2}$ RPTU Kaiserslautern-Landau & DFKI GmbH, Kaiserslautern, Germany + +# Abstract + +Diffusion models have enabled the generation of high-quality images with a strong focus on realism and textual fidelity. Yet, large-scale text-to-image models, such as Stable Diffusion, struggle to generate images where foreground objects are placed over a chroma key background, limiting their ability to separate foreground and background elements without fine-tuning. To address this limitation, we present a novel Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which optimizes the initial random noise to produce images with foreground objects on a specifiable color background. Our proposed method is the first to explore the manipulation of the color aspects in initial noise for controlled background generation, enabling precise separation of foreground and background without fine-tuning. Extensive experiments demonstrate that our training-free method outperforms existing methods in both qualitative and quantitative evaluations, matching or surpassing fine-tuned models. Finally, we successfully extend it to other tasks (e.g., consistency models and text-to-video), highlighting its transformative potential across various generative applications where independent control of foreground and background is crucial. + +# 1. Introduction + +Diffusion models have disrupted the landscape of image generation by producing high-quality images with high realism and textual fidelity [8, 11, 31, 38, 44]. Research has expanded into text-to-image generation conditioned on inputs like edges [49, 56], layouts [54, 57], and Virtual TryON (VTON) [10, 52]. However, real-world applications such as advertising, design, and game development often require precise control over foreground and background elements. The ability to generate images with transparent or chroma key backgrounds is essential for seamlessly integrating foreground objects into various scenes [55]. Traditional models like Stable Diffusion [37] and DeepFloyd [38] struggle to generate images where foreground and + +![](images/79214e5b59c17d378523b65970c5a5415ba9b9db82cf686a57878db630c01f03.jpg) +Figure 1. TKG-DM provides a training-free generation of foreground objects on a chroma key background, enabling independent control of foreground and background elements. Additionally, our method extends to conditional text-to-image and video generation tasks, making it suitable for creating foreground assets for various applications. + +background elements are separated without fine-tuning [5]. Specifically, these models cannot produce foreground objects over specifiable backgrounds or generate independent foreground and background layers, limiting their applicability in workflows requiring such precise control. + +Recent works like MAGICK [5] and LayerDiffuse [55] have significantly addressed this limitation. MAGICK introduces a prompt engineering-based approach for chroma key image generation. However, due to limited chroma key accuracy, it relies on post-processing and manual efforts to improve quality, including the release of foreground alpha image datasets. LayerDiffuse proposes a fine-tuning method for layer-wise generation using a dataset with 1 million images. Unfortunately, this dataset is not publicly available due to licensing restrictions, necessitating further resource-intensive dataset curation and fine-tuning. + +To overcome these challenges, we present the Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which optimizes the initial noise in the dif + +fusion process to produce images with foreground objects on specifiable color backgrounds without fine-tuning. Our method is the first to explore manipulating the color aspects of initial noise for controlled background generation, enabling precise separation of foreground and background elements. TKG-DM offers high flexibility and provides precise control over the background color, layout, size, and number of foreground objects. Our approach expands possibilities in generative content by maintaining the independence of the foreground and chroma key background. It seamlessly extends to applications like conditional text-to-image generation, consistency models, and text-to-video generation. + +Our extensive experiments demonstrate that TKG-DM improves FID and mask-FID scores by $33.7\%$ and $35.9\%$ , respectively. Thus, our training-free model rivals fine-tuned models, offering an efficient and versatile solution for various visual content creation tasks that require precise foreground and background control. Our contributions are summarized as follows: + +- We introduce TKG-DM, a training-free diffusion model that eliminates the need for fine-tuning and datasets for controllable foreground and background generation. +- TKG-DM provides precise control over background color, as well as the size, position and number of foreground elements, enabling flexible monochromatic chroma key generation. +- Our method is highly versatile, seamlessly extending to conditional text-to-image generation, consistency models, and text-to-video applications, thus enabling the creation of diverse chroma key content. +- TKG-DM outperforms existing models, delivering results that match or surpass fine-tuned alternatives while maintaining lower computational costs. + +# 2. Related Works + +# 2.1. Diffusion Models + +Diffusion models [19, 33] generate realistic, high-quality images by iteratively refining Gaussian noise to approximate a target distribution [11, 45]. Building upon this, Latent Diffusion Models (LDMs) [37] improve efficiency by exploiting the low-dimensional latent spaces while preserving image quality. In particular, text-to-image models [32, 36-38] align text and image features through crossattention to achieve diverse, high-quality outputs, with classifier-free guidance [18] and attention control [6, 8] further enhancing text fidelity. Beyond image generation, diffusion models find use in inpainting [24, 53], image editing [4, 15], and layout generation [7, 21]. Their versatility extends across multiple modalities, including audio generation [20, 22], text-to-video [14, 50], and 3D object generation [35, 48]. + +# 2.2. Initial Noise + +Diffusion models typically start with Gaussian noise and iteratively refine it to generate high-quality images that align closely with a given target distribution. In general, the choice and optimization of this initial noise are crucial, as they significantly impact image quality and alignment [40, 41, 51]. More specifically, noise optimization methods, such as human preference models or attention-based scores, can enhance fidelity and text alignment without additional training [12, 13]. Moreover, optimal noise seeds and targeted noise regions significantly impact image quality and object placement [1, 51]. In contrast, noise manipulation enables precise layout control, supporting layout-aware image generation and editing [25, 26, 43]. Building on these insights, our work introduces a novel method for shifting initial noise to control chromatic information. As a result, our method enables accurate chroma key backgrounds while preserving high-quality foreground content. + +# 2.3. Foreground and Background Separation in Image Generation + +Foreground-background separation is crucial in advertising, design, and game development, allowing creators to manipulate individual image elements independently for tasks such as background replacement, foreground adjustments, and complex compositions [3, 9, 28, 29]. + +In response to this challenge, MAGICK [5] employs a prompt engineering-based model to solve this task. However, as noted in this work: "Unfortunately, all the publicly available methods we tested were incapable of consistently creating keyable images." To overcome this limitation, MAGICK combines multiple models and applies manual effort to release a foreground alpha image dataset. Similarly, LayerDiffuse [55] introduces a fine-tuning approach using a 1 million image dataset, which cannot be publicly shared due to licensing restrictions. Although these methods improve foreground-background separation, they require substantial data collection and computational resources, making them less efficient for real-time applications. In contrast, our training-free approach provides precise control over foreground and background elements without the need for extensive datasets and fine-tuning. + +# 3. Methodology + +As shown in Fig. 2, TKG-DM extends Stable Diffusion [37] by manipulating the initial Gaussian noise $\mathbf{z}_T \in \mathbb{R}^{h \times w \times 4}$ through channel mean shift $F_c$ . This transformation produces the init color noise $\mathbf{z}_T^* = F_c(z_T)$ , which guides the vanilla Stable Diffusion to generate a uniform color image $\mathbf{x}_0^*$ without any text prompt. To generate an image where the foreground aligns with the input text prompt $p$ and the background has a specified color (e.g., green for chroma keying), + +![](images/1d26b4b49b5cc033332a1ab279c70e5e01db6c43c7bf27b21532ebd8c0e6d18e.jpg) +Figure 2. Starting with random noise $\mathbf{z}_T \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , init color noise $\mathbf{z}_T^*$ is generated by applying channel mean shift $F_{c}$ . This produces a single-colored image $\mathbf{x}_0^*$ without a prompt. By combining normal noise with init color noise via a Gaussian mask, TKG-DM generates a chroma key image $\mathbf{x}_0^{key}$ with the specified foreground (e.g., "the cat runs in the park") over a uniform background, effectively separating the foreground from the monochromatic background. + +we combine the initial noise $\mathbf{z}_T$ and the init color noise $\mathbf{z}_T^*$ using a Gaussian mask. This combined noise $\mathbf{z}_T^{key}$ is input into the vanilla Stable Diffusion, generating the chroma key image $\mathbf{x}_0^{key}$ . + +# 3.1. Channel Mean Shift + +Inspired by the relationship between Stable Diffusion's latent space and generated image color [47], we introduce a novel initial noise optimization technique, channel mean shift. It adjusts the mean of each channel in $z_{T}$ while keeping its standard deviation constant, enabling control over the generated image's color. To determine the optimal shift $\Delta_{c}$ for each channel $c \in \{1,2,3,4\}$ to achieve a target positive ratio in that channel. The values of channel $c$ in the noise tensor are denoted as $\mathbf{z}_T^{(c)}$ . The initial positive ratio for channel $c$ is defined as: + +$$ +\text {I n i t i a l R a t i o} _ {c} = \frac {\sum_ {i , j} \mathbb {1} \left(\mathbf {z} _ {T} ^ {(c)} (i , j) > 0\right)}{\text {T o t a l P i x e l s} _ {c}}, \tag {1} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function, and TotalPikels $c$ is the total number of elements in $\mathbf{z}_T^{(c)}$ . Given a target shift $\mathrm{TargetShift}_{c}$ , the target positive ratio is: + +$$ +\text {T a r g e t R a t i o} _ {c} = \text {I n i t i a l R a t i o} _ {c} + \text {T a r g e t S h i f t} _ {c}. \tag {2} +$$ + +To achieve TargetRatio, we iteratively adjust the mean shift $\Delta_{c}$ for each channel $\mathbf{z}_T^{(c)}$ . We initialize the shift with $\Delta_c^{\mathrm{init}} = 0$ and incrementally adjust $\Delta_{c}$ until the positive ratio meets or exceeds TargetRatio. Once the target ratio is reached, we record the final shift as $\Delta_c^{\mathrm{final}} = \Delta_c$ . The noise tensor obtained through this method is called init color noise $\mathbf{z}_T^* = F_c(\mathbf{z}_T) = \mathbf{z}_T + \Delta_c^{\mathrm{final}}$ . + +# 3.2. Init Noise Selection + +To generate the foreground object on the chroma key background, we apply an init noise selection strategy that selectively combines the initial noise $\mathbf{z}_T$ and the init color noise + +$\mathbf{z}_T^*$ using a 2D Gaussian mask $\mathbf{A}(i,j)$ . This mask creates a gradual transition by preserving the original noise in the foreground region and applying the color-shifted noise to the background region. Formally, the masked initial noise is computed as: + +$$ +\mathbf {z} _ {T} ^ {\text {k e y}} (i, j) = \mathbf {A} (i, j) \cdot \mathbf {z} _ {T} (i, j) + (1 - \mathbf {A} (i, j)) \cdot \mathbf {z} _ {T} ^ {*} (i, j), (3) +$$ + +where $\mathbf{A}(i,j) = e^{-\frac{(i - \mu_i)^2 + (j - \mu_j)^2}{2\sigma^2}}$ controls the blend between the original noise and init color noise and $\mu_{i}$ and $\mu_{j}$ specify the center and $\sigma$ the spread. + +Adjusting the Gaussian parameters $\mu_{i},\mu_{j}$ , and $\sigma$ controls the position and size of the foreground: shifting $\mu_{i}$ and $\mu_{j}$ adjusts its position while increasing $\sigma$ enlarges the foreground element. Interestingly, TKG-DM also generates multiple foreground elements by applying multiple Gaussian masks, each independently configured to produce varied positions and sizes for flexible compositions. + +# 4. Mechanism of TKG-DM + +In this section, we explain how TKG-DM specifies the background color and generates the foreground and background separately to generate chroma key content. + +# 4.1. Color Specify via Channel Mean Shift + +Inspired by previous research [47], we control the chroma key background color by applying channel mean shift to specific channels of the initial noise $\mathbf{z}_T\in \mathbb{R}^{h\times w\times 4}$ . Specifically, we adjust the mean of each channel $\mathbf{z}_T^{(c)}$ , where $c\in \{1,2,3,4\}$ , to influence the color composition of the generated images. In this experiment, we set $\mathrm{TargetShift}_{c} = \pm 7\%$ , adjusting the positive ratio of each channel by adding or subtracting $7\%$ relative to InitialRatio $c$ . + +As shown in Fig. 3a, positive shifts in $\mathbf{z}_T^{(2)}$ and $\mathbf{z}_T^{(3)}$ intensify cyan and yellow tones, respectively, while negative shifts emphasize red and blue-purple hues. Shifts in $\mathbf{z}_T^{(1)}$ and $\mathbf{z}_T^{(4)}$ primarily affect luminance and shades of white and + +![](images/423389d57a47f764fede64b680e78c16d7faa220ae1db65cc67015198f9000d3.jpg) +(a) Influence of Individual Noise Channel + +![](images/b8d8b277e9060f3772d34bad1c878571c430aac3a92463e1de7c33b95bce926a.jpg) +(b) Additive and Subtractive Color Effects +Figure 3. Fig. 3a illustrates the relationship between positive/negative channel mean shift with initial noise and color variations in the generated images. Different channel shifts across noise channels result in changes in the hue of the generated image. Fig. 3b shows how simultaneous shifts across multiple channels facilitate additive and subtractive color mixing, providing intuitive and flexible color control. + +black, respectively. These effects are confirmed by comparing generated images conditioned on specific text prompts. + +When we apply simultaneous shifts to multiple channels, we observe intuitive color mixing that aligns with additive and subtractive color theories. For example, positive shifts in both $\mathbf{z}_T^{(2)}$ and $\mathbf{z}_T^{(3)}$ enhance green tones, while combining red and yellow shifts with reduced luminance yields orange hues (see Fig. 3b). This channel-based manipulation enables flexible and natural color control in the image generation process. + +# 4.2. Content Separation via Attention Mechanism + +As illustrated in Fig. 4, self-attention and cross-attention mechanisms play distinct roles in generating the foreground and background elements, which is crucial for achieving effective foreground-background separation in TKG-DM. + +For the foreground, self-attention ensures internal consistency and coherence within the object, while cross-attention aligns the generated content with the text prompt. Due to the inherent bias in the training dataset [42], where foreground objects are more prominently described in captions, cross-attention forms a strong link between the foreground and the text prompt. + +For the background, the init color noise introduced by channel mean shift dominates the generation process. The self-attention mechanism synergizes with this modified initial noise to guide the background toward the specified chroma key color. Since background elements are often less detailed or vaguely described in dataset captions, the cross-attention has a weaker influence on the background, allowing the init color noise to take precedence. + +By exploiting this bias and manipulating the initial noise, + +![](images/7d698306951b410eb418d669d47aeb9ac86b3589bfdcb9e0c2ff32141b8efcbd.jpg) +Figure 4. Visualization of self- and cross-attention maps. The self-attention maps, segmented using a foreground segmentation model [58], show the regions attended to during generation. The cross-attention maps illustrate how the model attends to relevant regions based on the foreground and background prompts. This analysis shows how self-attention and cross-attention interact to influence foreground and background content generation. + +TKG-DM effectively decouples the background from the text prompt. This results in a uniform chroma key background and enables the isolated generation of foreground content without interference from undesired background details. + +![](images/ca613cf104e441ab8fcf022a19c10afbef3b02dbebaee9712508b1164381b14a.jpg) +Figure 5. Qualitative comparison in SD1.5. Existing methods fail to produce an accurate chroma key background. In contrast, ours produces a highly accurate chroma key background while generating high-quality foregrounds without Green Background Prompt (GBP). + +# 5. Experiments + +# 5.1. Experimental Setup + +We adopt SD1.5 and SDXL in TKG-DM to generate images at $512 \times 512$ and $1024 \times 1024$ resolutions, respectively, with DDIMScheduler, a guidance scale of 7.5, and 50 denoising steps. To obtain the init color noise for a green background, we apply channel mean shift with channels 2 and 3 in the positive direction for $\mathrm{TargetShift}_c = +7\%$ . Gaussian mask parameters are set to $(\sigma, \mu_i, \mu_j) = (0.5, w/2, h/2)$ . + +We compare our method against the naive usage of SD1.5 [37], SDXL [34], and DeepFloyd [38] using a Green Background Prompt (GBP) (=“isolated on a solid green background”) following the setup of MAGICK [5], as well as against fine-tuned models, namely GreenBack LoRA from CivitAI1 and LayerDiffuse [55]. + +# 5.2. Dataset and Metrics + +We collect 3,000 images from the MAGICK dataset [5], tailored for foreground image generation from text prompts. The MAGICK dataset is constructed by first generating monochromatic backgrounds using Deepfloyd [38] with GBP, followed by quality enhancement using SDEdit [27] and SDXL's img2img [34]. Foreground images with alpha values are curated using pixel-, deep learning-, and human-based methods. Our experiment utilizes alpha images, alpha masks, and corresponding text prompts. + +We construct ground truth images for evaluation by overlaying the alpha images onto a lime green background images $(\mathrm{RGB} = 50,205,50)$ . We use the Fréchet Inception + +Distance (FID) to assess foreground quality [17]. Additionally, we introduce a novel m-FID metric, which compares the mask extracted from generated chroma key content using BiRefNet [58] to the ground truth mask, providing a comprehensive assessment of foreground accuracy. For semantic alignment, we employ CLIP-Sentence (CLIP-S) and CLIP-Image (CLIP-I) [16]: CLIP-S measures alignment with the text, while CLIP-I assesses visual similarity to the ground truth. + +# 5.3. Qualitative Result + +Fig. 5 and Fig. 6 show the qualitative results of our model and existing methods with SD1.5 and SDXL, respectively. They demonstrate that our model generates high-precision chroma key images without prompt engineering. + +With the Green Background Prompt (GBP), SD1.5 struggles to generate a clean green background, while SDXL, while slightly better, generates unstable light green tints on the background and often produces multiple foreground objects, complicating foreground-background separation. Additionally, the GBP's "green" element unintentionally adds green tones to the foreground, such as on clothing or in a cat's eyes (Fig. 6, first and third columns). DeepFloyd performs the worst foreground quality and text alignment. + +In the fine-tuned model, while LoRA (SD1.5) shows minor improvements, it still struggles with chroma key accuracy. LayerDiffuse produces well-separated foregrounds but occasionally loses details, like precise numbers or letters, due to dataset limitations in fine-tuning. Mask generation also occasionally fails, resulting in uncut images. + +Moreover, existing models struggle to generate accurate chroma key images when background information is in- + +![](images/3fc01e815aaaf3c9bbdeaacf239995bcdd64f21a4e5786f0ceeca22212e59f95.jpg) +Figure 6. Qualitative comparison in SDXL. Existing training-free methods fail to isolate foreground objects from the background. While LayerDiffuse, a fine-tuned model, successfully generates the transparent foregrounds on the lime green background, it may struggle with accurate text generation and background handling. In contrast, our model efficiently generates highly accurate chroma key images with foreground content placed over the background without fine-tuning and needing Green Background Prompt (GBP). + +![](images/dd667350a80b006c543a8a1b73025b44ac0ecc6b74c151b7795daeb354d56a3f.jpg) +couple of petite chihuahua dogs +panorama with wine, cheese, grapes and olives +Figure 7. Qualitative various color result in SDXL. By adjusting the channel mean shift, TKG-DM controls the background color while generating a high-quality foreground. + +cluded in the text prompt (Fig. 6, last columns). In contrast, our training-free model consistently produces high-quality chroma key backgrounds with only the foreground content, ensuring precise separation. + +Furthermore, as shown in Fig. 7, TKG-DM generates the various color background images through channel mean shift without prompt engineering. Additional results are provided in the supplementary material. + +Table 1. Quantitative results. It indicates that our training-free model outperforms the existing model with Green Background Prompt (GBP) and rivals the fine-tuned model, LayerDiffuse [55]. + +
SD1.5
FID ↓m-FID ↓CLIP-I↑CLIP-S ↑
SD1.5 (GBP) [37]85.0063.540.7100.256
GB LoRA (GBP)60.2949.030.7040.243
Ours56.3240.750.7370.261
+ +SDXL + +
DeepFloyd (GBP) [38]31.5720.310.7810.270
SDXL (GBP) [34]45.3239.170.7590.272
LayerDiffuse [55]29.3429.820.7780.276
Ours41.8131.430.7630.273
+ +Image Quality / Prompt Alignment + +![](images/b2352ee960c758a4572cd6fa78f955c93f88a06a59ff42824f177d7b5978c9a2.jpg) +Figure 8. Results from our user study show the preference rates for foreground objects generated by our model and existing models based on their image quality and alignment with the text prompt. GBP means a model with a Green Background Prompt. + +![](images/abb121252802cb663d8cda58dc83b1de92b91edb53d751f26a85d735a95d52ee.jpg) +Figure 9. Qualitative comparison in ControlNET. Existing methods typically generate images with issues such as color erosion in the foreground, uneven backgrounds, or unintended elements from the conditioning input. In contrast, our model generates images with the foreground isolated from the background, maintaining a consistent chroma key background without the Green Background Prompt (GBP). + +# 5.4. Quantitative Result + +As shown in Table 1, our method consistently outperforms other training-free models across all metrics, achieving better image realism and mask accuracy, as indicated by lower FID and m-FID scores. It also aligns well with input text prompts and has closer visual similarity to the ground truth, as indicated by higher CLIP-S and CLIP-I scores. + +In the SDXL setting, LayerDiffuse achieves the best FID, while our model remains competitive across all other metrics. The FID advantage for LayerDiffuse likely results from its setup, which overlays foregrounds on a lime green background matching the ground truth. + +DeepFloyd's high FID, m-FID, and CLIP-I scores reflect its similarity to the ground truth based on DeepFloyd's outputs. However, this alignment gives it an inherent advantage, making it unsuitable as a fair benchmark for image quality. Its lower CLIP-S score further indicates weaker text alignment compared to other models. + +Overall, these results underscore our model's ability to generate high-quality, text-aligned foregrounds without fine-tuning, offering an efficient chroma key content generation solution. + +# 5.5. User Study + +Besides objective metrics, we conducted a user study to assess subjective human preferences for prompt adherence and image quality, comparing our method against SDXL with Green Background Prompt (GBP) and LayerDiffuse. We generated 30 image pairs per method, totaling 60 images. 100 participants ranked images via two alternative + +Table 2. Quantitative results comparing our method with existing methods for ControlNet [56]. GBP means a model with a Green Background Prompt. + +
ControlNet [56]
FID ↓m-FID ↓CLIP-I ↑CLIP-S ↑
SDXL (GBP) [34]22.0418.620.8190.279
Ours17.0917.220.8340.284
+ +forced-choice [30, 39] and based on foreground quality and text alignment. To ensure focus on the foreground, we used a foreground extraction model [58] and manually refined areas as needed. As shown in Fig. 8, our training-free approach improves object realism and text alignment over the existing model. + +# 6. Applications + +We demonstrate the flexibility of TKG-DM in various tasks, implemented with the SD1.5 and SDXL, as presented in Table 2, Fig. 9 and Fig. 10. Further details and additional results are provided in the supplementary material. + +Application to ControlNet. As depicted in Fig. 9 and Table 2, our method integrates seamlessly with ControlNet [56], enabling precise generation control using conditioning inputs like canny edges. By applying a Gaussian mask to the init color noise, we preserve normal initial noise in the foreground regions, resulting in distinct foregrounds on uniform chroma key backgrounds. + +Compared to the standard ControlNet, our method demonstrates superior qualitative and quantitative perfor + +![](images/2db3be692d6d8520c9b5b1e474d0dc6d001ad8d3716b3fb5ffe954963a1e15be.jpg) +Figure 10. Application results of TKG-DM across various tasks. TKG-DM effectively supports layout-aware text-to-image generation, consistency models and text-to-video. Each row demonstrates its versatility in adapting to different domains, from realistic object and character placements to animation. + +mance. Existing methods often face issues like color erosion in the foreground, uneven backgrounds, or unintended elements from the conditioning input. By isolating the foreground from background influence and maintaining a consistent chroma key background, our approach is highly effective for compositing tasks without prompt engineering. + +Layout-Aware Text-to-Image Generation. As discussed in Section 3.2, our model enables control over foreground objects' location and size by adjusting the Gaussian mask's center and spread. This control allows users to place objects precisely within specific areas of the generated image, making it practical for layout-aware tasks. Additionally, using multiple Gaussian masks enables the generation of multiple foreground objects within a single scene (Fig. 10, top-right). Unlike existing models that lack targeted control over object placement and size, our approach integrates layout information seamlessly without fine-tuning. + +Application to Consistency Models. Consistency Models [46] are generative models that produce high-quality outputs in a few steps, unlike traditional diffusion models. Operating in latent space, Consistency Models demonstrate the robustness of TKG-DM to generative models that handle images in the latent diffusion-based model, enabling significantly faster generation while maintaining high output quality without modifications. + +Application to Text-to-Video. Our TKG-DM extends seamlessly to text-to-video tasks, ensuring consistent foreground content across frames. We align foreground objects with the input prompt by applying TKG-DM with + +![](images/0d4fd1c893556ffe0540ddac62225f74bf9bb022a2c0ded0d3286191af0c7707.jpg) +Figure 11. Failure Cases. TKG-DM fails to generate a foreground object when the size is too small or when the text prompt lacks object information, resulting in only the background being generated. + +AnimateDiff [14] on each frame while maintaining a uniform background for chroma key video generation. This approach, the first to generate foreground video with controlled chroma key backgrounds, facilitates efficient and flexible video editing workflows involving background removal or replacement. + +# 7. Limitation and Future Work + +As shown in Fig. 11, TKG-DM primarily focuses on generating foreground content (e.g., objects and people), which limits its ability to generate backgrounds like landscapes. Additionally, if the size parameter is too small, ignore generating the foreground object [26, 43]. + +Future enhancements could expand TKG-DM's capabilities to include controlled background generation, allowing separate and adjustable generation of both foreground and background content. This would greatly benefit applications in video and 3D content creation where isolated and reusable background assets are essential. + +# 8. Conclusion + +This work presents the first exploration of initial noise color manipulation in diffusion models, demonstrating its effectiveness in controlling background chromaticity. Leveraging this insight, we introduce the Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which generates prompt-based foregrounds on a chroma key background without fine-tuning. TKG-DM extends seamlessly to tasks like consistency models and text-to-video, highlighting its versatility beyond static images. Experimental results show that our method matches or surpasses fine-tuned models in both qualitative and quantitative evaluations without additional training. This approach not only advances theoretical understanding but also provides practical benefits for real-world applications where isolating foreground content is essential. + +# Acknowledgements + +This research was supported by BMBF project SustainML (Grant 101070408) and the BMDV NITD (45KI22B021). + +# References + +[1] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Boqing Gong, Cho-Jui Hsieh, and Minhao Cheng. The crystal ball hypothesis in diffusion models: Anticipating object positions from initial noise. arXiv preprint arXiv:2406.01970, 2024. 2 +[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation, 2023. 2 +[3] Yaniv Benny and Lior Wolf. Onegan: Simultaneous unsupervised learning of conditional image generation, foreground segmentation, and fine-grained clustering. In ECCV, pages 514-530. Springer, 2020. 2 +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 2 +[5] Ryan D Burgert, Brian L Price, Jason Kuen, Yijun Li, and Michael S Ryoo. Magick: A large-scale captioned dataset from matting generated images using chroma keying. In CVPR, pages 22595-22604, 2024. 1, 2, 5 +[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 2 +[7] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In CVPR, pages 18349-18358, 2023. 2 +[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 1, 2 +[9] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In CVPR, pages 8724-8733, 2024. 2 +[10] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 1 +[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 1, 2 +[12] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. arXiv preprint arXiv:2406.04312, 2024. 2 +[13] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, pages 9380-9389, 2024. 2 +[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 8 +[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 + +[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 5 +[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2 +[20] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In ICML, pages 13916-13932. PMLR, 2023. 2 +[21] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Layoutdm: Discrete diffusion model for controllable layout generation. In CVPR, pages 10167-10176, 2023. 2 +[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2 +[23] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2 +[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2 +[25] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Guided image synthesis via initial image editing in diffusion model. arXiv preprint arXiv:2305.03382, 2023. 2 +[26] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Semantic-driven initial image construction for guided image synthesis in diffusion model. arXiv preprint arXiv:2312.08872, 2023. 2, 8 +[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 5 +[28] Ryugo Morita, Zhiqiang Zhang, Man M Ho, and Jinjia Zhou. Interactive image manipulation with complex text instructions. In WACV, pages 1053-1062, 2023. 2 +[29] Ryugo Morita, Zhiqiang Zhang, and Jinjia Zhou. Batinet: Background-aware text to image synthesis and manipulation network. In ICIP, pages 765-769. IEEE, 2023. 2 +[30] Brian B Moser, Stanislav Frolov, Federico Raue, Sebastian Palacio, and Andreas Dengel. Yoda: You only diffuse areas. an area-masked diffusion approach for image superresolution. arXiv preprint arXiv:2308.07977, 2023. 7 +[31] Brian B Moser, Arundhati S Shanbhag, Federico Raue, Stanislav Frolov, Sebastian Palacio, and Andreas Dengel. Diffusion models, image super-resolution, and everything: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1 + +[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171. PMLR, 2021. 2 +[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 5, 6, 7 +[35] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2 +[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. Pmlr, 2021. 2 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 5, 6 +[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1, 2, 5, 6 +[39] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 45(4): 4713-4726, 2022. 7 +[40] Dvir Samuel, Rami Ben-Ari, Nir Darshan, Haggai Maron, and Gal Chechik. Norm-guided latent space exploration for text-to-image generation. NeurIPS, 36, 2024. 2 +[41] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, pages 4695-4703, 2024. 2 +[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 4 +[43] Takahiro Shirakawa and Seiichi Uchida. Noisecollapse: A layout-aware text-to-image diffusion model based on noise cropping and merging. In CVPR, pages 8921-8930, 2024. 2, 8 +[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015. 1 + +[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 +[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023.8 +[47] Timothy Alexis Vass. Explaining the sdxl latent space. https://huggingface.co/blog/TimothyAlexisVass/explaining-the-sdxl-latent-space, 2024.3 +[48] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2 +[49] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 1 +[50] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 2 +[51] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv preprint arXiv:2405.14828, 2024. 2 +[52] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024. 1 +[53] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, pages 18381-18391, 2023. 2 +[54] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In CVPR, pages 14246-14255, 2023. 1 +[55] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 1, 2, 5, 6 +[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 1, 7 +[57] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, pages 22490-22499, 2023. 1 +[58] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. arXiv preprint arXiv:2401.03407, 2024. 4, 5, 7 \ No newline at end of file diff --git a/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/images.zip b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0ded5d7385b790ba18684f4da32f2d3fbafdc152 --- /dev/null +++ b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d987223b698b2c0708aabf4ae70737a3c71806628c48190288c8e7ce4540418 +size 832377 diff --git a/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/layout.json b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3c44c3e2267192bff6d7cb95cc88314fecfea958 --- /dev/null +++ b/2025/TKG-DM_ Training-free Chroma Key Content Generation Diffusion Model/layout.json @@ -0,0 +1,8246 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 538, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 538, + 120 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 538, + 120 + ], + "type": "text", + "content": "TKG-DM: Training-free Chroma Key Content Generation Diffusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "spans": [ + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "text", + "content": "Ryugo Morita" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "text", + "content": ", Stanislav Frolov" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "text", + "content": ", Brian Bernhard Moser" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "text", + "content": ", Takahiro Shirakawa, Ko Watanabe" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "text", + "content": ", Andreas Dengel" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "text", + "content": ", Jinjia Zhou" + }, + { + "bbox": [ + 108, + 142, + 502, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "type": "text", + "content": "Faculty of Science and Engineering, Hosei University, Tokyo, Japan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 186, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 186, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 129, + 186, + 481, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 186, + 481, + 200 + ], + "type": "text", + "content": "RPTU Kaiserslautern-Landau & DFKI GmbH, Kaiserslautern, Germany" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 251, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 515 + ], + "type": "text", + "content": "Diffusion models have enabled the generation of high-quality images with a strong focus on realism and textual fidelity. Yet, large-scale text-to-image models, such as Stable Diffusion, struggle to generate images where foreground objects are placed over a chroma key background, limiting their ability to separate foreground and background elements without fine-tuning. To address this limitation, we present a novel Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which optimizes the initial random noise to produce images with foreground objects on a specifiable color background. Our proposed method is the first to explore the manipulation of the color aspects in initial noise for controlled background generation, enabling precise separation of foreground and background without fine-tuning. Extensive experiments demonstrate that our training-free method outperforms existing methods in both qualitative and quantitative evaluations, matching or surpassing fine-tuned models. Finally, we successfully extend it to other tasks (e.g., consistency models and text-to-video), highlighting its transformative potential across various generative applications where independent control of foreground and background is crucial." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 538, + 135, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 135, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 135, + 551 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": "Diffusion models have disrupted the landscape of image generation by producing high-quality images with high realism and textual fidelity [8, 11, 31, 38, 44]. Research has expanded into text-to-image generation conditioned on inputs like edges [49, 56], layouts [54, 57], and Virtual TryON (VTON) [10, 52]. However, real-world applications such as advertising, design, and game development often require precise control over foreground and background elements. The ability to generate images with transparent or chroma key backgrounds is essential for seamlessly integrating foreground objects into various scenes [55]. Traditional models like Stable Diffusion [37] and DeepFloyd [38] struggle to generate images where foreground and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 225, + 550, + 390 + ], + "blocks": [ + { + "bbox": [ + 319, + 225, + 550, + 390 + ], + "lines": [ + { + "bbox": [ + 319, + 225, + 550, + 390 + ], + "spans": [ + { + "bbox": [ + 319, + 225, + 550, + 390 + ], + "type": "image", + "image_path": "79214e5b59c17d378523b65970c5a5415ba9b9db82cf686a57878db630c01f03.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 398, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. TKG-DM provides a training-free generation of foreground objects on a chroma key background, enabling independent control of foreground and background elements. Additionally, our method extends to conditional text-to-image and video generation tasks, making it suitable for creating foreground assets for various applications." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 486, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 486, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 312, + 486, + 555, + 545 + ], + "type": "text", + "content": "background elements are separated without fine-tuning [5]. Specifically, these models cannot produce foreground objects over specifiable backgrounds or generate independent foreground and background layers, limiting their applicability in workflows requiring such precise control." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 545, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 556, + 677 + ], + "type": "text", + "content": "Recent works like MAGICK [5] and LayerDiffuse [55] have significantly addressed this limitation. MAGICK introduces a prompt engineering-based approach for chroma key image generation. However, due to limited chroma key accuracy, it relies on post-processing and manual efforts to improve quality, including the release of foreground alpha image datasets. LayerDiffuse proposes a fine-tuning method for layer-wise generation using a dataset with 1 million images. Unfortunately, this dataset is not publicly available due to licensing restrictions, necessitating further resource-intensive dataset curation and fine-tuning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "To overcome these challenges, we present the Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which optimizes the initial noise in the dif" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13031" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "type": "text", + "content": "fusion process to produce images with foreground objects on specifiable color backgrounds without fine-tuning. Our method is the first to explore manipulating the color aspects of initial noise for controlled background generation, enabling precise separation of foreground and background elements. TKG-DM offers high flexibility and provides precise control over the background color, layout, size, and number of foreground objects. Our approach expands possibilities in generative content by maintaining the independence of the foreground and chroma key background. It seamlessly extends to applications like conditional text-to-image generation, consistency models, and text-to-video generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "type": "text", + "content": "Our extensive experiments demonstrate that TKG-DM improves FID and mask-FID scores by " + }, + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "type": "inline_equation", + "content": "33.7\\%" + }, + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "type": "inline_equation", + "content": "35.9\\%" + }, + { + "bbox": [ + 55, + 228, + 294, + 312 + ], + "type": "text", + "content": ", respectively. Thus, our training-free model rivals fine-tuned models, offering an efficient and versatile solution for various visual content creation tasks that require precise foreground and background control. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 483 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "text", + "content": "- We introduce TKG-DM, a training-free diffusion model that eliminates the need for fine-tuning and datasets for controllable foreground and background generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 351, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 351, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 351, + 295, + 398 + ], + "type": "text", + "content": "- TKG-DM provides precise control over background color, as well as the size, position and number of foreground elements, enabling flexible monochromatic chroma key generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 399, + 295, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 446 + ], + "type": "text", + "content": "- Our method is highly versatile, seamlessly extending to conditional text-to-image generation, consistency models, and text-to-video applications, thus enabling the creation of diverse chroma key content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 447, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 447, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 447, + 295, + 483 + ], + "type": "text", + "content": "- TKG-DM outperforms existing models, delivering results that match or surpass fine-tuned alternatives while maintaining lower computational costs." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 496, + 146, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 146, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 146, + 508 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 516, + 158, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 158, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 158, + 528 + ], + "type": "text", + "content": "2.1. Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "Diffusion models [19, 33] generate realistic, high-quality images by iteratively refining Gaussian noise to approximate a target distribution [11, 45]. Building upon this, Latent Diffusion Models (LDMs) [37] improve efficiency by exploiting the low-dimensional latent spaces while preserving image quality. In particular, text-to-image models [32, 36-38] align text and image features through crossattention to achieve diverse, high-quality outputs, with classifier-free guidance [18] and attention control [6, 8] further enhancing text fidelity. Beyond image generation, diffusion models find use in inpainting [24, 53], image editing [4, 15], and layout generation [7, 21]. Their versatility extends across multiple modalities, including audio generation [20, 22], text-to-video [14, 50], and 3D object generation [35, 48]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 394, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 394, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 394, + 83 + ], + "type": "text", + "content": "2.2. Initial Noise" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 89, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 294 + ], + "type": "text", + "content": "Diffusion models typically start with Gaussian noise and iteratively refine it to generate high-quality images that align closely with a given target distribution. In general, the choice and optimization of this initial noise are crucial, as they significantly impact image quality and alignment [40, 41, 51]. More specifically, noise optimization methods, such as human preference models or attention-based scores, can enhance fidelity and text alignment without additional training [12, 13]. Moreover, optimal noise seeds and targeted noise regions significantly impact image quality and object placement [1, 51]. In contrast, noise manipulation enables precise layout control, supporting layout-aware image generation and editing [25, 26, 43]. Building on these insights, our work introduces a novel method for shifting initial noise to control chromatic information. As a result, our method enables accurate chroma key backgrounds while preserving high-quality foreground content." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 302, + 553, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 553, + 327 + ], + "type": "text", + "content": "2.3. Foreground and Background Separation in Image Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 553, + 392 + ], + "type": "text", + "content": "Foreground-background separation is crucial in advertising, design, and game development, allowing creators to manipulate individual image elements independently for tasks such as background replacement, foreground adjustments, and complex compositions [3, 9, 28, 29]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 393, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 584 + ], + "type": "text", + "content": "In response to this challenge, MAGICK [5] employs a prompt engineering-based model to solve this task. However, as noted in this work: \"Unfortunately, all the publicly available methods we tested were incapable of consistently creating keyable images.\" To overcome this limitation, MAGICK combines multiple models and applies manual effort to release a foreground alpha image dataset. Similarly, LayerDiffuse [55] introduces a fine-tuning approach using a 1 million image dataset, which cannot be publicly shared due to licensing restrictions. Although these methods improve foreground-background separation, they require substantial data collection and computational resources, making them less efficient for real-time applications. In contrast, our training-free approach provides precise control over foreground and background elements without the need for extensive datasets and fine-tuning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 597, + 397, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 397, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 397, + 611 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "text", + "content": "As shown in Fig. 2, TKG-DM extends Stable Diffusion [37] by manipulating the initial Gaussian noise " + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T \\in \\mathbb{R}^{h \\times w \\times 4}" + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "text", + "content": " through channel mean shift " + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "inline_equation", + "content": "F_c" + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "text", + "content": ". This transformation produces the init color noise " + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^* = F_c(z_T)" + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "text", + "content": ", which guides the vanilla Stable Diffusion to generate a uniform color image " + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^*" + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "text", + "content": " without any text prompt. To generate an image where the foreground aligns with the input text prompt " + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 617, + 553, + 715 + ], + "type": "text", + "content": " and the background has a specified color (e.g., green for chroma keying)," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13032" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 553, + 186 + ], + "type": "image", + "image_path": "1d26b4b49b5cc033332a1ab279c70e5e01db6c43c7bf27b21532ebd8c0e6d18e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "text", + "content": "Figure 2. Starting with random noise " + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "text", + "content": ", init color noise " + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^*" + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "text", + "content": " is generated by applying channel mean shift " + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "text", + "content": ". This produces a single-colored image " + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^*" + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "text", + "content": " without a prompt. By combining normal noise with init color noise via a Gaussian mask, TKG-DM generates a chroma key image " + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{key}" + }, + { + "bbox": [ + 54, + 193, + 555, + 239 + ], + "type": "text", + "content": " with the specified foreground (e.g., \"the cat runs in the park\") over a uniform background, effectively separating the foreground from the monochromatic background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "text", + "content": "we combine the initial noise " + }, + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T" + }, + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "text", + "content": " and the init color noise " + }, + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^*" + }, + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "text", + "content": " using a Gaussian mask. This combined noise " + }, + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{key}" + }, + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "text", + "content": " is input into the vanilla Stable Diffusion, generating the chroma key image " + }, + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{key}" + }, + { + "bbox": [ + 55, + 258, + 296, + 309 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 173, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 173, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 173, + 327 + ], + "type": "text", + "content": "3.1. Channel Mean Shift" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "text", + "content": "Inspired by the relationship between Stable Diffusion's latent space and generated image color [47], we introduce a novel initial noise optimization technique, channel mean shift. It adjusts the mean of each channel in " + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "text", + "content": " while keeping its standard deviation constant, enabling control over the generated image's color. To determine the optimal shift " + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\Delta_{c}" + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "text", + "content": " for each channel " + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "inline_equation", + "content": "c \\in \\{1,2,3,4\\}" + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "text", + "content": " to achieve a target positive ratio in that channel. The values of channel " + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "text", + "content": " in the noise tensor are denoted as " + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(c)}" + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "text", + "content": ". The initial positive ratio for channel " + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 332, + 296, + 452 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 459, + 294, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 459, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 94, + 459, + 294, + 488 + ], + "type": "interline_equation", + "content": "\\text {I n i t i a l R a t i o} _ {c} = \\frac {\\sum_ {i , j} \\mathbb {1} \\left(\\mathbf {z} _ {T} ^ {(c)} (i , j) > 0\\right)}{\\text {T o t a l P i x e l s} _ {c}}, \\tag {1}", + "image_path": "e7dd2986c206113e83ad80d9ae7f492b29c469389a90a065a8025def35bbc55e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "text", + "content": " is the indicator function, and TotalPikels " + }, + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "text", + "content": " is the total number of elements in " + }, + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(c)}" + }, + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "text", + "content": ". Given a target shift " + }, + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{TargetShift}_{c}" + }, + { + "bbox": [ + 55, + 497, + 296, + 536 + ], + "type": "text", + "content": ", the target positive ratio is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 544, + 295, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 544, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 85, + 544, + 295, + 559 + ], + "type": "interline_equation", + "content": "\\text {T a r g e t R a t i o} _ {c} = \\text {I n i t i a l R a t i o} _ {c} + \\text {T a r g e t S h i f t} _ {c}. \\tag {2}", + "image_path": "ac2705afad037f45a6f967337c43b11a968e2b1b836db988ec4cd31850933f6e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "text", + "content": "To achieve TargetRatio, we iteratively adjust the mean shift " + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\Delta_{c}" + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "text", + "content": " for each channel " + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(c)}" + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "text", + "content": ". We initialize the shift with " + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\Delta_c^{\\mathrm{init}} = 0" + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "text", + "content": " and incrementally adjust " + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\Delta_{c}" + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "text", + "content": " until the positive ratio meets or exceeds TargetRatio. Once the target ratio is reached, we record the final shift as " + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\Delta_c^{\\mathrm{final}} = \\Delta_c" + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "text", + "content": ". The noise tensor obtained through this method is called init color noise " + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^* = F_c(\\mathbf{z}_T) = \\mathbf{z}_T + \\Delta_c^{\\mathrm{final}}" + }, + { + "bbox": [ + 55, + 567, + 296, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 659, + 168, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 168, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 168, + 672 + ], + "type": "text", + "content": "3.2. Init Noise Selection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To generate the foreground object on the chroma key background, we apply an init noise selection strategy that selectively combines the initial noise " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and the init color noise" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 258, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^*" + }, + { + "bbox": [ + 313, + 258, + 554, + 319 + ], + "type": "text", + "content": " using a 2D Gaussian mask " + }, + { + "bbox": [ + 313, + 258, + 554, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(i,j)" + }, + { + "bbox": [ + 313, + 258, + 554, + 319 + ], + "type": "text", + "content": ". This mask creates a gradual transition by preserving the original noise in the foreground region and applying the color-shifted noise to the background region. Formally, the masked initial noise is computed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 324, + 555, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 324, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 318, + 324, + 555, + 340 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {T} ^ {\\text {k e y}} (i, j) = \\mathbf {A} (i, j) \\cdot \\mathbf {z} _ {T} (i, j) + (1 - \\mathbf {A} (i, j)) \\cdot \\mathbf {z} _ {T} ^ {*} (i, j), (3)", + "image_path": "79002211f992ba14e067d6c480e24276245594f261ca8077df6a7f0f74299ce1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(i,j) = e^{-\\frac{(i - \\mu_i)^2 + (j - \\mu_j)^2}{2\\sigma^2}}" + }, + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "text", + "content": " controls the blend between the original noise and init color noise and " + }, + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mu_{j}" + }, + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "text", + "content": " specify the center and " + }, + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 346, + 554, + 389 + ], + "type": "text", + "content": " the spread." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "content": "Adjusting the Gaussian parameters " + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\mu_{i},\\mu_{j}" + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "content": " controls the position and size of the foreground: shifting " + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\mu_{j}" + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "content": " adjusts its position while increasing " + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "content": " enlarges the foreground element. Interestingly, TKG-DM also generates multiple foreground elements by applying multiple Gaussian masks, each independently configured to produce varied positions and sizes for flexible compositions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 482, + 455, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 455, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 455, + 495 + ], + "type": "text", + "content": "4. Mechanism of TKG-DM" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 502, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 539 + ], + "type": "text", + "content": "In this section, we explain how TKG-DM specifies the background color and generates the foreground and background separately to generate chroma key content." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 515, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 515, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 515, + 558 + ], + "type": "text", + "content": "4.1. Color Specify via Channel Mean Shift" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "text", + "content": "Inspired by previous research [47], we control the chroma key background color by applying channel mean shift to specific channels of the initial noise " + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T\\in \\mathbb{R}^{h\\times w\\times 4}" + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "text", + "content": ". Specifically, we adjust the mean of each channel " + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(c)}" + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "inline_equation", + "content": "c\\in \\{1,2,3,4\\}" + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "text", + "content": ", to influence the color composition of the generated images. In this experiment, we set " + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{TargetShift}_{c} = \\pm 7\\%" + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "text", + "content": ", adjusting the positive ratio of each channel by adding or subtracting " + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "text", + "content": " relative to InitialRatio " + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 563, + 554, + 659 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "text", + "content": "As shown in Fig. 3a, positive shifts in " + }, + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(2)}" + }, + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(3)}" + }, + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "text", + "content": " intensify cyan and yellow tones, respectively, while negative shifts emphasize red and blue-purple hues. Shifts in " + }, + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(1)}" + }, + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(4)}" + }, + { + "bbox": [ + 313, + 659, + 554, + 715 + ], + "type": "text", + "content": " primarily affect luminance and shades of white and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13033" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 88, + 342, + 243 + ], + "blocks": [ + { + "bbox": [ + 59, + 88, + 342, + 243 + ], + "lines": [ + { + "bbox": [ + 59, + 88, + 342, + 243 + ], + "spans": [ + { + "bbox": [ + 59, + 88, + 342, + 243 + ], + "type": "image", + "image_path": "423389d57a47f764fede64b680e78c16d7faa220ae1db65cc67015198f9000d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 244, + 276, + 253 + ], + "lines": [ + { + "bbox": [ + 140, + 244, + 276, + 253 + ], + "spans": [ + { + "bbox": [ + 140, + 244, + 276, + 253 + ], + "type": "text", + "content": "(a) Influence of Individual Noise Channel" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 397, + 89, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 397, + 89, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 397, + 89, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 397, + 89, + 553, + 243 + ], + "type": "image", + "image_path": "b8d8b277e9060f3772d34bad1c878571c430aac3a92463e1de7c33b95bce926a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 244, + 531, + 253 + ], + "lines": [ + { + "bbox": [ + 394, + 244, + 531, + 253 + ], + "spans": [ + { + "bbox": [ + 394, + 244, + 531, + 253 + ], + "type": "text", + "content": "(b) Additive and Subtractive Color Effects" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 263, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 555, + 298 + ], + "type": "text", + "content": "Figure 3. Fig. 3a illustrates the relationship between positive/negative channel mean shift with initial noise and color variations in the generated images. Different channel shifts across noise channels result in changes in the hue of the generated image. Fig. 3b shows how simultaneous shifts across multiple channels facilitate additive and subtractive color mixing, providing intuitive and flexible color control." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 318, + 294, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 294, + 342 + ], + "type": "text", + "content": "black, respectively. These effects are confirmed by comparing generated images conditioned on specific text prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 343, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 295, + 441 + ], + "type": "text", + "content": "When we apply simultaneous shifts to multiple channels, we observe intuitive color mixing that aligns with additive and subtractive color theories. For example, positive shifts in both " + }, + { + "bbox": [ + 55, + 343, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(2)}" + }, + { + "bbox": [ + 55, + 343, + 295, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 343, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T^{(3)}" + }, + { + "bbox": [ + 55, + 343, + 295, + 441 + ], + "type": "text", + "content": " enhance green tones, while combining red and yellow shifts with reduced luminance yields orange hues (see Fig. 3b). This channel-based manipulation enables flexible and natural color control in the image generation process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 452, + 291, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 291, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 291, + 464 + ], + "type": "text", + "content": "4.2. Content Separation via Attention Mechanism" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 471, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 519 + ], + "type": "text", + "content": "As illustrated in Fig. 4, self-attention and cross-attention mechanisms play distinct roles in generating the foreground and background elements, which is crucial for achieving effective foreground-background separation in TKG-DM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 520, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 603 + ], + "type": "text", + "content": "For the foreground, self-attention ensures internal consistency and coherence within the object, while cross-attention aligns the generated content with the text prompt. Due to the inherent bias in the training dataset [42], where foreground objects are more prominently described in captions, cross-attention forms a strong link between the foreground and the text prompt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": "For the background, the init color noise introduced by channel mean shift dominates the generation process. The self-attention mechanism synergizes with this modified initial noise to guide the background toward the specified chroma key color. Since background elements are often less detailed or vaguely described in dataset captions, the cross-attention has a weaker influence on the background, allowing the init color noise to take precedence." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "By exploiting this bias and manipulating the initial noise," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 319, + 552, + 536 + ], + "blocks": [ + { + "bbox": [ + 318, + 319, + 552, + 536 + ], + "lines": [ + { + "bbox": [ + 318, + 319, + 552, + 536 + ], + "spans": [ + { + "bbox": [ + 318, + 319, + 552, + 536 + ], + "type": "image", + "image_path": "7d698306951b410eb418d669d47aeb9ac86b3589bfdcb9e0c2ff32141b8efcbd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 546, + 555, + 624 + ], + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 624 + ], + "type": "text", + "content": "Figure 4. Visualization of self- and cross-attention maps. The self-attention maps, segmented using a foreground segmentation model [58], show the regions attended to during generation. The cross-attention maps illustrate how the model attends to relevant regions based on the foreground and background prompts. This analysis shows how self-attention and cross-attention interact to influence foreground and background content generation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": "TKG-DM effectively decouples the background from the text prompt. This results in a uniform chroma key background and enables the isolated generation of foreground content without interference from undesired background details." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13034" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 281 + ], + "type": "image", + "image_path": "ca613cf104e441ab8fcf022a19c10afbef3b02dbebaee9712508b1164381b14a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 288, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 555, + 312 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison in SD1.5. Existing methods fail to produce an accurate chroma key background. In contrast, ours produces a highly accurate chroma key background while generating high-quality foregrounds without Green Background Prompt (GBP)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 330, + 138, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 138, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 138, + 344 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 173, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 173, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 173, + 363 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "text", + "content": "We adopt SD1.5 and SDXL in TKG-DM to generate images at " + }, + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "text", + "content": " resolutions, respectively, with DDIMScheduler, a guidance scale of 7.5, and 50 denoising steps. To obtain the init color noise for a green background, we apply channel mean shift with channels 2 and 3 in the positive direction for " + }, + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\mathrm{TargetShift}_c = +7\\%" + }, + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "text", + "content": ". Gaussian mask parameters are set to " + }, + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "inline_equation", + "content": "(\\sigma, \\mu_i, \\mu_j) = (0.5, w/2, h/2)" + }, + { + "bbox": [ + 55, + 368, + 296, + 452 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 452, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 296, + 524 + ], + "type": "text", + "content": "We compare our method against the naive usage of SD1.5 [37], SDXL [34], and DeepFloyd [38] using a Green Background Prompt (GBP) (=“isolated on a solid green background”) following the setup of MAGICK [5], as well as against fine-tuned models, namely GreenBack LoRA from CivitAI1 and LayerDiffuse [55]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 174, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 174, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 174, + 544 + ], + "type": "text", + "content": "5.2. Dataset and Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 550, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 658 + ], + "type": "text", + "content": "We collect 3,000 images from the MAGICK dataset [5], tailored for foreground image generation from text prompts. The MAGICK dataset is constructed by first generating monochromatic backgrounds using Deepfloyd [38] with GBP, followed by quality enhancement using SDEdit [27] and SDXL's img2img [34]. Foreground images with alpha values are curated using pixel-, deep learning-, and human-based methods. Our experiment utilizes alpha images, alpha masks, and corresponding text prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 658, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 695 + ], + "type": "text", + "content": "We construct ground truth images for evaluation by overlaying the alpha images onto a lime green background images " + }, + { + "bbox": [ + 55, + 658, + 296, + 695 + ], + "type": "inline_equation", + "content": "(\\mathrm{RGB} = 50,205,50)" + }, + { + "bbox": [ + 55, + 658, + 296, + 695 + ], + "type": "text", + "content": ". We use the Fréchet Inception" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 555, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 439 + ], + "type": "text", + "content": "Distance (FID) to assess foreground quality [17]. Additionally, we introduce a novel m-FID metric, which compares the mask extracted from generated chroma key content using BiRefNet [58] to the ground truth mask, providing a comprehensive assessment of foreground accuracy. For semantic alignment, we employ CLIP-Sentence (CLIP-S) and CLIP-Image (CLIP-I) [16]: CLIP-S measures alignment with the text, while CLIP-I assesses visual similarity to the ground truth." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 423, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 423, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 423, + 458 + ], + "type": "text", + "content": "5.3. Qualitative Result" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 510 + ], + "type": "text", + "content": "Fig. 5 and Fig. 6 show the qualitative results of our model and existing methods with SD1.5 and SDXL, respectively. They demonstrate that our model generates high-precision chroma key images without prompt engineering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": "With the Green Background Prompt (GBP), SD1.5 struggles to generate a clean green background, while SDXL, while slightly better, generates unstable light green tints on the background and often produces multiple foreground objects, complicating foreground-background separation. Additionally, the GBP's \"green\" element unintentionally adds green tones to the foreground, such as on clothing or in a cat's eyes (Fig. 6, first and third columns). DeepFloyd performs the worst foreground quality and text alignment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "content": "In the fine-tuned model, while LoRA (SD1.5) shows minor improvements, it still struggles with chroma key accuracy. LayerDiffuse produces well-separated foregrounds but occasionally loses details, like precise numbers or letters, due to dataset limitations in fine-tuning. Mask generation also occasionally fails, resulting in uncut images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Moreover, existing models struggle to generate accurate chroma key images when background information is in-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 702, + 168, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 168, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 168, + 712 + ], + "type": "text", + "content": "1https://civitai.com/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13035" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 336 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 336 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 336 + ], + "type": "image", + "image_path": "3fc01e815aaaf3c9bbdeaacf239995bcdd64f21a4e5786f0ceeca22212e59f95.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 346, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 555, + 392 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison in SDXL. Existing training-free methods fail to isolate foreground objects from the background. While LayerDiffuse, a fine-tuned model, successfully generates the transparent foregrounds on the lime green background, it may struggle with accurate text generation and background handling. In contrast, our model efficiently generates highly accurate chroma key images with foreground content placed over the background without fine-tuning and needing Green Background Prompt (GBP)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 418, + 292, + 537 + ], + "blocks": [ + { + "bbox": [ + 75, + 411, + 160, + 418 + ], + "lines": [ + { + "bbox": [ + 75, + 411, + 160, + 418 + ], + "spans": [ + { + "bbox": [ + 75, + 411, + 160, + 418 + ], + "type": "text", + "content": "couple of petite chihuahua dogs" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 172, + 411, + 292, + 418 + ], + "lines": [ + { + "bbox": [ + 172, + 411, + 292, + 418 + ], + "spans": [ + { + "bbox": [ + 172, + 411, + 292, + 418 + ], + "type": "text", + "content": "panorama with wine, cheese, grapes and olives" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 418, + 292, + 537 + ], + "lines": [ + { + "bbox": [ + 59, + 418, + 292, + 537 + ], + "spans": [ + { + "bbox": [ + 59, + 418, + 292, + 537 + ], + "type": "image", + "image_path": "dd667350a80b006c543a8a1b73025b44ac0ecc6b74c151b7795daeb354d56a3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 545, + 295, + 579 + ], + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 579 + ], + "type": "text", + "content": "Figure 7. Qualitative various color result in SDXL. By adjusting the channel mean shift, TKG-DM controls the background color while generating a high-quality foreground." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 661 + ], + "type": "text", + "content": "cluded in the text prompt (Fig. 6, last columns). In contrast, our training-free model consistently produces high-quality chroma key backgrounds with only the foreground content, ensuring precise separation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Furthermore, as shown in Fig. 7, TKG-DM generates the various color background images through channel mean shift without prompt engineering. Additional results are provided in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 451, + 552, + 499 + ], + "blocks": [ + { + "bbox": [ + 313, + 410, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 444 + ], + "type": "text", + "content": "Table 1. Quantitative results. It indicates that our training-free model outperforms the existing model with Green Background Prompt (GBP) and rivals the fine-tuned model, LayerDiffuse [55]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 451, + 552, + 499 + ], + "lines": [ + { + "bbox": [ + 315, + 451, + 552, + 499 + ], + "spans": [ + { + "bbox": [ + 315, + 451, + 552, + 499 + ], + "type": "table", + "html": "
SD1.5
FID ↓m-FID ↓CLIP-I↑CLIP-S ↑
SD1.5 (GBP) [37]85.0063.540.7100.256
GB LoRA (GBP)60.2949.030.7040.243
Ours56.3240.750.7370.261
", + "image_path": "4d2846bd5f732871a0c8b55526112bff0372d5c87accfa0e3a899bfcb9e0db47.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 506, + 552, + 548 + ], + "blocks": [ + { + "bbox": [ + 423, + 499, + 443, + 506 + ], + "lines": [ + { + "bbox": [ + 423, + 499, + 443, + 506 + ], + "spans": [ + { + "bbox": [ + 423, + 499, + 443, + 506 + ], + "type": "text", + "content": "SDXL" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 506, + 552, + 548 + ], + "lines": [ + { + "bbox": [ + 316, + 506, + 552, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 506, + 552, + 548 + ], + "type": "table", + "html": "
DeepFloyd (GBP) [38]31.5720.310.7810.270
SDXL (GBP) [34]45.3239.170.7590.272
LayerDiffuse [55]29.3429.820.7780.276
Ours41.8131.430.7630.273
", + "image_path": "7ac3720b4e98344e4527e3baf2c29010dbaedc572e787a628b6166e3904075db.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 559, + 487, + 568 + ], + "lines": [ + { + "bbox": [ + 380, + 559, + 487, + 568 + ], + "spans": [ + { + "bbox": [ + 380, + 559, + 487, + 568 + ], + "type": "text", + "content": "Image Quality / Prompt Alignment" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 322, + 571, + 547, + 626 + ], + "blocks": [ + { + "bbox": [ + 322, + 571, + 547, + 626 + ], + "lines": [ + { + "bbox": [ + 322, + 571, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 322, + 571, + 547, + 626 + ], + "type": "image", + "image_path": "b2352ee960c758a4572cd6fa78f955c93f88a06a59ff42824f177d7b5978c9a2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 639, + 555, + 684 + ], + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 684 + ], + "type": "text", + "content": "Figure 8. Results from our user study show the preference rates for foreground objects generated by our model and existing models based on their image quality and alignment with the text prompt. GBP means a model with a Green Background Prompt." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13036" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 281 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 281 + ], + "type": "image", + "image_path": "abb121252802cb663d8cda58dc83b1de92b91edb53d751f26a85d735a95d52ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 289, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 54, + 289, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 555, + 324 + ], + "type": "text", + "content": "Figure 9. Qualitative comparison in ControlNET. Existing methods typically generate images with issues such as color erosion in the foreground, uneven backgrounds, or unintended elements from the conditioning input. In contrast, our model generates images with the foreground isolated from the background, maintaining a consistent chroma key background without the Green Background Prompt (GBP)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 343, + 170, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 170, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 170, + 356 + ], + "type": "text", + "content": "5.4. Quantitative Result" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 433 + ], + "type": "text", + "content": "As shown in Table 1, our method consistently outperforms other training-free models across all metrics, achieving better image realism and mask accuracy, as indicated by lower FID and m-FID scores. It also aligns well with input text prompts and has closer visual similarity to the ground truth, as indicated by higher CLIP-S and CLIP-I scores." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 493 + ], + "type": "text", + "content": "In the SDXL setting, LayerDiffuse achieves the best FID, while our model remains competitive across all other metrics. The FID advantage for LayerDiffuse likely results from its setup, which overlays foregrounds on a lime green background matching the ground truth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 494, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 566 + ], + "type": "text", + "content": "DeepFloyd's high FID, m-FID, and CLIP-I scores reflect its similarity to the ground truth based on DeepFloyd's outputs. However, this alignment gives it an inherent advantage, making it unsuitable as a fair benchmark for image quality. Its lower CLIP-S score further indicates weaker text alignment compared to other models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 566, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 296, + 613 + ], + "type": "text", + "content": "Overall, these results underscore our model's ability to generate high-quality, text-aligned foregrounds without fine-tuning, offering an efficient chroma key content generation solution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 623, + 130, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 130, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 130, + 636 + ], + "type": "text", + "content": "5.5. User Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Besides objective metrics, we conducted a user study to assess subjective human preferences for prompt adherence and image quality, comparing our method against SDXL with Green Background Prompt (GBP) and LayerDiffuse. We generated 30 image pairs per method, totaling 60 images. 100 participants ranked images via two alternative" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 384, + 555, + 427 + ], + "blocks": [ + { + "bbox": [ + 313, + 342, + 555, + 375 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 375 + ], + "type": "text", + "content": "Table 2. Quantitative results comparing our method with existing methods for ControlNet [56]. GBP means a model with a Green Background Prompt." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 384, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 315, + 384, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 315, + 384, + 555, + 427 + ], + "type": "table", + "html": "
ControlNet [56]
FID ↓m-FID ↓CLIP-I ↑CLIP-S ↑
SDXL (GBP) [34]22.0418.620.8190.279
Ours17.0917.220.8340.284
", + "image_path": "ce46d58c0eca4604600b989c5c9424a6b00fd34efa293234579ac44d23d824e7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "content": "forced-choice [30, 39] and based on foreground quality and text alignment. To ensure focus on the foreground, we used a foreground extraction model [58] and manually refined areas as needed. As shown in Fig. 8, our training-free approach improves object realism and text alignment over the existing model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 394, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 394, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 394, + 548 + ], + "type": "text", + "content": "6. Applications" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 555, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 604 + ], + "type": "text", + "content": "We demonstrate the flexibility of TKG-DM in various tasks, implemented with the SD1.5 and SDXL, as presented in Table 2, Fig. 9 and Fig. 10. Further details and additional results are provided in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 604, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 689 + ], + "type": "text", + "content": "Application to ControlNet. As depicted in Fig. 9 and Table 2, our method integrates seamlessly with ControlNet [56], enabling precise generation control using conditioning inputs like canny edges. By applying a Gaussian mask to the init color noise, we preserve normal initial noise in the foreground regions, resulting in distinct foregrounds on uniform chroma key backgrounds." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Compared to the standard ControlNet, our method demonstrates superior qualitative and quantitative perfor" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13037" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 294, + 274 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 294, + 274 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 294, + 274 + ], + "type": "image", + "image_path": "2db3be692d6d8520c9b5b1e474d0dc6d001ad8d3716b3fb5ffe954963a1e15be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 282, + 296, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 338 + ], + "type": "text", + "content": "Figure 10. Application results of TKG-DM across various tasks. TKG-DM effectively supports layout-aware text-to-image generation, consistency models and text-to-video. Each row demonstrates its versatility in adapting to different domains, from realistic object and character placements to animation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 361, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 434 + ], + "type": "text", + "content": "mance. Existing methods often face issues like color erosion in the foreground, uneven backgrounds, or unintended elements from the conditioning input. By isolating the foreground from background influence and maintaining a consistent chroma key background, our approach is highly effective for compositing tasks without prompt engineering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 567 + ], + "type": "text", + "content": "Layout-Aware Text-to-Image Generation. As discussed in Section 3.2, our model enables control over foreground objects' location and size by adjusting the Gaussian mask's center and spread. This control allows users to place objects precisely within specific areas of the generated image, making it practical for layout-aware tasks. Additionally, using multiple Gaussian masks enables the generation of multiple foreground objects within a single scene (Fig. 10, top-right). Unlike existing models that lack targeted control over object placement and size, our approach integrates layout information seamlessly without fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 568, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 664 + ], + "type": "text", + "content": "Application to Consistency Models. Consistency Models [46] are generative models that produce high-quality outputs in a few steps, unlike traditional diffusion models. Operating in latent space, Consistency Models demonstrate the robustness of TKG-DM to generative models that handle images in the latent diffusion-based model, enabling significantly faster generation while maintaining high output quality without modifications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Application to Text-to-Video. Our TKG-DM extends seamlessly to text-to-video tasks, ensuring consistent foreground content across frames. We align foreground objects with the input prompt by applying TKG-DM with" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 156 + ], + "type": "image", + "image_path": "0d4fd1c893556ffe0540ddac62225f74bf9bb022a2c0ded0d3286191af0c7707.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 164, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 198 + ], + "type": "text", + "content": "Figure 11. Failure Cases. TKG-DM fails to generate a foreground object when the size is too small or when the text prompt lacks object information, resulting in only the background being generated." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 217, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 289 + ], + "type": "text", + "content": "AnimateDiff [14] on each frame while maintaining a uniform background for chroma key video generation. This approach, the first to generate foreground video with controlled chroma key backgrounds, facilitates efficient and flexible video editing workflows involving background removal or replacement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 298, + 477, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 477, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 477, + 311 + ], + "type": "text", + "content": "7. Limitation and Future Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "content": "As shown in Fig. 11, TKG-DM primarily focuses on generating foreground content (e.g., objects and people), which limits its ability to generate backgrounds like landscapes. Additionally, if the size parameter is too small, ignore generating the foreground object [26, 43]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 379, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 554, + 450 + ], + "type": "text", + "content": "Future enhancements could expand TKG-DM's capabilities to include controlled background generation, allowing separate and adjustable generation of both foreground and background content. This would greatly benefit applications in video and 3D content creation where isolated and reusable background assets are essential." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 460, + 388, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 388, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 388, + 472 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 479, + 555, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 659 + ], + "type": "text", + "content": "This work presents the first exploration of initial noise color manipulation in diffusion models, demonstrating its effectiveness in controlling background chromaticity. Leveraging this insight, we introduce the Training-Free Chroma Key Content Generation Diffusion Model (TKG-DM), which generates prompt-based foregrounds on a chroma key background without fine-tuning. TKG-DM extends seamlessly to tasks like consistency models and text-to-video, highlighting its versatility beyond static images. Experimental results show that our method matches or surpasses fine-tuned models in both qualitative and quantitative evaluations without additional training. This approach not only advances theoretical understanding but also provides practical benefits for real-world applications where isolating foreground content is essential." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 670, + 417, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 417, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 417, + 683 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "This research was supported by BMBF project SustainML (Grant 101070408) and the BMDV NITD (45KI22B021)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13038" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Boqing Gong, Cho-Jui Hsieh, and Minhao Cheng. The crystal ball hypothesis in diffusion models: Anticipating object positions from initial noise. arXiv preprint arXiv:2406.01970, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 168 + ], + "type": "text", + "content": "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 213 + ], + "type": "text", + "content": "[3] Yaniv Benny and Lior Wolf. Onegan: Simultaneous unsupervised learning of conditional image generation, foreground segmentation, and fine-grained clustering. In ECCV, pages 514-530. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 246 + ], + "type": "text", + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 246, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 295, + 291 + ], + "type": "text", + "content": "[5] Ryan D Burgert, Brian L Price, Jason Kuen, Yijun Li, and Michael S Ryoo. Magick: A large-scale captioned dataset from matting generated images using chroma keying. In CVPR, pages 22595-22604, 2024. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 335 + ], + "type": "text", + "content": "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, pages 22560-22570, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 369 + ], + "type": "text", + "content": "[7] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In CVPR, pages 18349-18358, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "type": "text", + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 413, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 295, + 456 + ], + "type": "text", + "content": "[9] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In CVPR, pages 8724-8733, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 457, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 457, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 457, + 295, + 491 + ], + "type": "text", + "content": "[10] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for virtual try-on. arXiv preprint arXiv:2403.05139, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 491, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 295, + 524 + ], + "type": "text", + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "text", + "content": "[12] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. arXiv preprint arXiv:2406.04312, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 569, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 613 + ], + "type": "text", + "content": "[13] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, pages 9380-9389, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 669 + ], + "type": "text", + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 162, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 183 + ], + "type": "text", + "content": "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 183, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 215 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 216, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 216, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 216, + 553, + 271 + ], + "type": "text", + "content": "[20] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In ICML, pages 13916-13932. PMLR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 271, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 315 + ], + "type": "text", + "content": "[21] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Layoutdm: Discrete diffusion model for controllable layout generation. In CVPR, pages 10167-10176, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 316, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 349 + ], + "type": "text", + "content": "[22] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 349, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 382 + ], + "type": "text", + "content": "[23] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, pages 11461-11471, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "text", + "content": "[25] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Guided image synthesis via initial image editing in diffusion model. arXiv preprint arXiv:2305.03382, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 459, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 553, + 503 + ], + "type": "text", + "content": "[26] Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Semantic-driven initial image construction for guided image synthesis in diffusion model. arXiv preprint arXiv:2312.08872, 2023. 2, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 503, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 547 + ], + "type": "text", + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 547, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 547, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 547, + 553, + 581 + ], + "type": "text", + "content": "[28] Ryugo Morita, Zhiqiang Zhang, Man M Ho, and Jinjia Zhou. Interactive image manipulation with complex text instructions. In WACV, pages 1053-1062, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "text", + "content": "[29] Ryugo Morita, Zhiqiang Zhang, and Jinjia Zhou. Batinet: Background-aware text to image synthesis and manipulation network. In ICIP, pages 765-769. IEEE, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[30] Brian B Moser, Stanislav Frolov, Federico Raue, Sebastian Palacio, and Andreas Dengel. Yoda: You only diffuse areas. an area-masked diffusion approach for image superresolution. arXiv preprint arXiv:2308.07977, 2023. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[31] Brian B Moser, Arundhati S Shanbhag, Federico Raue, Stanislav Frolov, Sebastian Palacio, and Andreas Dengel. Diffusion models, image super-resolution, and everything: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13039" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 160 + ], + "type": "text", + "content": "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 216 + ], + "type": "text", + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "text", + "content": "[35] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "text", + "content": "[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. Pmlr, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 435 + ], + "type": "text", + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "text", + "content": "[39] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE TPAMI, 45(4): 4713-4726, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "text", + "content": "[40] Dvir Samuel, Rami Ben-Ari, Nir Darshan, Haggai Maron, and Gal Chechik. Norm-guided latent space exploration for text-to-image generation. NeurIPS, 36, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 557 + ], + "type": "text", + "content": "[41] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, pages 4695-4703, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "text", + "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "text", + "content": "[43] Takahiro Shirakawa and Seiichi Uchida. Noisecollapse: A layout-aware text-to-image diffusion model based on noise cropping and merging. In CVPR, pages 8921-8930, 2024. 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 677 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 106 + ], + "type": "text", + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "type": "text", + "content": "[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023.8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 184 + ], + "type": "text", + "content": "[47] Timothy Alexis Vass. Explaining the sdxl latent space. https://huggingface.co/blog/TimothyAlexisVass/explaining-the-sdxl-latent-space, 2024.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 240 + ], + "type": "text", + "content": "[48] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 554, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 284 + ], + "type": "text", + "content": "[49] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 554, + 342 + ], + "type": "text", + "content": "[50] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 554, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 385 + ], + "type": "text", + "content": "[51] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv preprint arXiv:2405.14828, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 386, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 554, + 430 + ], + "type": "text", + "content": "[52] Yuhao Xu, Tao Gu, Weifeng Chen, and Chengcai Chen. Oot-diffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. arXiv preprint arXiv:2403.01779, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 431, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 554, + 475 + ], + "type": "text", + "content": "[53] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In CVPR, pages 18381-18391, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 476, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 554, + 520 + ], + "type": "text", + "content": "[54] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In CVPR, pages 14246-14255, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 521, + 554, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 554, + 553 + ], + "type": "text", + "content": "[55] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 555, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 588 + ], + "type": "text", + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 1, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 589, + 554, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 554, + 632 + ], + "type": "text", + "content": "[57] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, pages 22490-22499, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 634, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 554, + 677 + ], + "type": "text", + "content": "[58] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. arXiv preprint arXiv:2401.03407, 2024. 4, 5, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13040" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/8535e16b-e9a4-4280-8556-88908a875920_content_list.json b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/8535e16b-e9a4-4280-8556-88908a875920_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8fb2dd97f9daf8849f2b69ef6bf75b4fba0a3419 --- /dev/null +++ b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/8535e16b-e9a4-4280-8556-88908a875920_content_list.json @@ -0,0 +1,1810 @@ +[ + { + "type": "text", + "text": "TSAM: Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation", + "text_level": 1, + "bbox": [ + 116, + 128, + 880, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abduljalil Radman & Jorma Laaksonen \nDepartment of Computer Science, Aalto University, Finland \n{abduljalil.saif, jorma.laaksonen}@aalto.fi", + "bbox": [ + 259, + 203, + 736, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring audio-visual segmentation (Ref-AVS) aims to segment objects within audio-visual scenes using multimodal cues embedded in text expressions. While the Segment Anything Model (SAM) has revolutionized visual segmentation, its applicability to Ref-AVS, where multimodal cues act as novel prompts, remains unexplored. SAM's limitation to single-frame segmentation also hinders its ability to capture essential temporal context needed for multi-frame audio-visual segmentation. To address this gap, we propose TSAM, a novel extension of SAM designed to leverage multimodal cues for precise segmentation in dynamic audio-visual scenes. TSAM enhances SAM's image encoder with a temporal modeling branch, enabling spatio-temporal learning and deep multimodal fusion across video frames, while retaining SAM's pre-trained knowledge. Additionally, TSAM replaces SAM's user-interactive prompting mechanism with sparse and dense data-driven prompts, enabling more effective integration of audio-visual inputs and reference text expressions. Extensive experiments on the Ref-AVS dataset demonstrate TSAM's superiority over state-of-the-art methods. The results illustrate its effectiveness in segmenting objects in dynamic audio-visual scenes using text-based multimodal cues and its strong generalization to unseen objects. Project webpage: https://abdurad.github.io/TSAM/.", + "bbox": [ + 89, + 325, + 485, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real-world scenarios, people often rely on multimodal information, such as audio-visual and visual-text, to gather spatiotemporal and semantic insights for distinguishing objects of interest [10, 39]. This multimodal approach is equally valuable for machines, enabling applications like stage cameras to track a performing musician or movie cameras to focus on a superhero confronting a yelling villain. Relying on a single sensory input, such as vision, is often insufficient for accurate object segmentation in complex,", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dynamic audio-visual scenes [16, 36]. In the literature, researchers have addressed this need for multimodal integration in object segmentation through three main approaches, each focused on a specific modality cue.", + "bbox": [ + 511, + 292, + 903, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, video object segmentation (VOS) [6, 30] relies on a mask annotation in the initial frame to guide segmentation of objects across subsequent frames. While VOS allows for flexible segmentation, it is heavily dependent on labor-intensive frame-level annotations, making it challenging to scale. Second, referring video object segmentation (Ref-VOS) [5, 9, 13] replaces the need for mask annotations with text expressions to identify targeted objects. This approach introduces a more intuitive, text-based method for segmentation, but struggles with tracking and segmenting objects in more dynamic audio-visual environments. Third, audio-visual segmentation (AVS) [21, 24, 33, 35, 39] leverages audio cues to segment sounding objects in video, effectively handling complex audio-visual scenarios, but remaining limited in scope as it does not generalize to inaudible objects. Each of these approaches, VOS, Ref-VOS, and AVS, contributes valuable segmentation capabilities within its modality focus, but remains challenged in more naturalistic audio-visual scenarios [36].", + "bbox": [ + 511, + 354, + 906, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Wang et al. [36] introduced referring audio-visual segmentation (Ref-AVS), a task that segments objects in video by interpreting text expressions enriched with both audio and visual cues. They also presented Expression Enhancing with Multimodal Cues (EEMC), a simple method that integrates reference text expressions with dynamic audio-visual cues to segment targeted objects in scenes. By unifying these multimodal cues, EEMC enables robust segmentation in dynamic scenes, overcoming limitations of prior approaches that focused on narrower combinations of modalities.", + "bbox": [ + 511, + 642, + 908, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose that enhancing Ref-AVS can be achieved by leveraging prior knowledge from large-scale visual foundation models, such as the Segment Anything Model (SAM) [15], for segmentation. SAM is a flexible segmentation model that utilizes an image encoder to generate high-dimensional image embeddings. It incorpo", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23947", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfe22308fe89c6ec3d6e561d0015f84a726a9d3007c683c041f3cb692977268a.jpg", + "image_caption": [ + "Figure 1. Overview of TSAM. Left: TSAM integrates multimodal cues from diverse inputs, leveraging the prompting modules to guide segmentation. Right: The temporal branch fuses multimodal cues and learns temporal context for segmentation in audio-visual scenes. The adapter input, $\\dot{F}_a$ , is the output from the cached memory." + ], + "image_footnote": [], + "bbox": [ + 106, + 97, + 591, + 571 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/538387209496b06f899b05fbfb6b2020e5b163402fdf06ef08238f0f192a83cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 172, + 893, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rates user-provided visual prompts, which can be sparse (e.g., points, boxes) or dense (e.g., masks), processed by a dedicated prompt encoder. These embeddings are then passed through a lightweight, transformer-based mask decoder to generate accurate object masks based on the provided prompts. Although SAM has been applied to tasks such as AVS [24, 29, 33, 35], and Ref-VOS [19], its potential for segmenting objects in the visual domain using multimodal cues embedded in text expressions remains unexplored. This makes its application to the Ref-AVS task both novel and challenging as it requires the integration of diverse multimodal cues to guide segmentation, overcoming the limitations of SAM's traditional reliance on explicit user-interactive prompts.", + "bbox": [ + 88, + 659, + 482, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main challenge in adapting SAM for the Ref-AVS", + "bbox": [ + 109, + 869, + 482, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "task, which requires nuanced multimodal understanding without direct user guidance, arises from SAM's lack of temporal awareness in its image encoder, limiting its ability to capture complex interactions across audio, visual, and text modalities. Additionally, SAM relies on user-provided visual prompts for explicit spatial information of targeted objects, while its capacity to process text and audio modalities remains limited.", + "bbox": [ + 511, + 659, + 906, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper introduces TSAM (Fig. 1), a novel SAM-based architecture for the Ref-AVS task, designed to enhance segmentation by leveraging multimodal cues from audio, visual, and text inputs. TSAM incorporates an auxiliary temporal modeling branch in parallel with SAM's image encoder, preserving SAM's zero-shot capabilities while adapting it for temporally varying video content. More-", + "bbox": [ + 511, + 779, + 908, + 885 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23948", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "over, TSAM uses text cues to guide the selection of relevant audio cues, which are then used as sparse prompts, enriched with key visual cues, to query SAM's mask decoder. Since multimodal prompts lack explicit spatial information for the target objects, TSAM incorporates a lightweight adapter module into the temporal branch to enhance SAM's mask decoder prompts, facilitating deeper multimodal interactions during training. Through a cross-attention mechanism, TSAM synthesizes data-driven multimodal reference cues as dense prompts for SAM. Extensive experiments show that TSAM outperforms existing state-of-the-art methods in segmenting target objects in the Ref-AVS task. The main contributions of this work are:", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose TSAM, an end-to-end architecture built on the Segment Anything Model (SAM), which redefines traditional visual prompts as data-driven multimodal prompts for object segmentation in Ref-AVS tasks.", + "2. We introduce a temporal modeling branch that enhances SAM's image encoder with temporal awareness, enabling the capture of intricate spatio-temporal interactions across multimodal cues.", + "3. Through comprehensive experiments, we demonstrate that the proposed TSAM achieves superior performance compared to state-of-the-art methods in Ref-AVS. Furthermore, extensive ablation studies provide a detailed evaluation of TSAM's architecture, emphasizing the pivotal role of multimodal cue integration and temporal modeling in enhancing segmentation performance." + ], + "bbox": [ + 89, + 289, + 482, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 532, + 227, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Referring video object segmentation (Ref-VOS)", + "text_level": 1, + "bbox": [ + 89, + 559, + 482, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Referring video object segmentation (Ref-VOS) involves segmenting target objects within video frames based on textual descriptions. Gavrilyuk et al. [9] introduced an encoder-decoder model based on I3D [2], that generates dynamic convolutional filters from text and applies them to visual representations at different resolutions to produce segmentation masks for target objects. URVOS [32] enhanced object tracking by performing segmentation on a key frame and then propagating the mask across the video. This approach combines cross-modal attention to prevent drift and memory attention to maintain temporal consistency across frames. MTTR [1] formulated Ref-VOS as a sequence prediction task, utilizing multimodal transformers to process video and text jointly. ReferFormer [37] used Video-Swin (V-Swin) [26] as the visual backbone, linking textual expressions directly to relevant frame regions. In this approach, language-conditioned object queries serve as inputs, helping the model identify the referred objects and apply dynamic kernels as convolution filters to generate segmentation masks. Although Ref-VOS enhances user accessibility by replacing mask annotations with natural lan", + "bbox": [ + 89, + 583, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "guage expressions, it still struggles to accurately identify objects in dynamic audio-visual scenes [36].", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "R2-VOS [18] extended the Ref-VOS task to handle unpaired visual-text inputs, jointly modeling video segmentation and text reconstruction. By introducing a structural text-to-text cycle constraint, R2-VOS improves multimodal alignment across diverse text inputs, overcoming the limitations of prior point-wise methods. RefSAM [19] utilized text prompts with SAM [15], using a lightweight cross-modal MLP to project the referring expression's text embedding into prompt embeddings. RefSAM enables multimodal reference segmentation by integrating textual and visual modalities, but it does not fully exploit the simultaneous interaction of multimodal cues or support the temporal dynamics required for video segmentation tasks.", + "bbox": [ + 511, + 122, + 903, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Audio-visual segmentation (AVS)", + "text_level": 1, + "bbox": [ + 511, + 332, + 803, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-visual segmentation (AVS), introduced by Zhou et al. [39], is a challenging task that seeks to achieve pixel-level segmentation of sound-emitting objects in video, guided by the audio modality. AVSBench [39] used a temporal pixel-wise audio-visual interaction module to inject audio semantics as guidance for the visual segmentation process. Similarly, AVSegFormer [8], a transformer-based model built on the PVT-v2 backbone [34], used cross-attention to fuse spatial-temporal visual cues with audio cues. By decoding these combined audio-visual cues and employing audio queries, the model enhanced its ability to focus on sounding objects within the audio-visual scene. Other transformer-based models proposed in [14, 17, 23] integrated audio and visual cues across both temporal and spatial dimensions for the AVS task. Alternative approaches [20, 22] disentangled the audio input into audio queries, which interacts with the visual input to decode segmentation masks. These approaches emphasize implicit modality fusion strategies, training models to fit discrete samples in the dataset. However, their performance often falls short due to limitations in dataset size and diversity.", + "bbox": [ + 511, + 354, + 906, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternatively, a Explicit Conditional Multimodal Variational Auto-Encoder (ECMVAE) [28] was proposed to extract semantically correlated cues from audio-visual inputs for segmenting audible objects in video frames. In another approach, the AVS Bidirectional Generation (AVSBG) framework [11] was introduced, correlating visual cues of objects with their corresponding audio cues through a visual-to-audio projection and implicit volumetric motion estimation. Recent advancements in AVS [3, 27, 38] also integrated audio-visual contrastive learning methods to select informative samples and improve training efficiency and model performance. However, these methods often fail to fully leverage available information due to limited prior segmentation knowledge in their backbone architectures, hin", + "bbox": [ + 511, + 688, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23949", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dering their ability to effectively mine informative samples.", + "bbox": [ + 89, + 90, + 480, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the other hand, SAM-based methods showed improvements in the AVS task. AV-SAM [29] was an early attempt to apply SAM to AVS, utilizing pixel-wise audio-visual fusion to prompt SAM for sounding object segmentation. In [24], two SAM-based approaches for AVS were introduced: Audio-Prompted SAM (AP-SAM) and SAM Adapter (SAMA-AVS). AP-SAM encoded audio as a prompt for SAM's mask decoder, while SAMA-AVS incorporated an adapter module to inject the audio information during the training process, but did not fully leverage SAM's promptable nature. Recently, a Generalizable Audio-Visual Segmentation (GAVS) model [35] was introduced, targeting zero-shot and few-shot scenarios. GAVS proposed a semantic-aware audio prompt to bridge the semantic gap between visual and auditory modalities through contrastive learning. All of these methods take a single image and its corresponding audio as input, fusing multimodal inputs in the spatial domain, which limits their ability to capture temporal relationships frame by frame.", + "bbox": [ + 89, + 107, + 482, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a straightforward approach, Seon et al. [33] extended SAM for AVS by integrating a Spatio-Temporal, Bidirectional Audio-Visual Attention (ST-BAVA) module between SAM's image encoder and mask decoder. While this approach aims to model temporal cues for audible object segmentation, it oversimplifies the interaction between audio and visual modalities, treating them in a relatively shallow manner. The model relies on basic feature fusion instead of a more meaningful interaction, limiting its ability to fully leverage the potential of multimodal cues for AVS tasks.", + "bbox": [ + 89, + 395, + 482, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Referring audio-visual segmentation (Ref-AVS)", + "text_level": 1, + "bbox": [ + 89, + 559, + 482, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Referring audio-visual segmentation (Ref-AVS) is a novel pixel-level segmentation task in which the model predicts, for each pixel, whether it corresponds to an object specified by multimodal cues in a text expression. Wang et al. [36] introduced the Ref-AVS task and developed a dataset with mask annotations for segmenting objects in video frames guided by multimodal expressions, creating the first benchmark for supervised Ref-AVS. They also proposed the Expression Enhancing with Multimodal Cues (EEMC) model, based on the Mask2Former (M2F) [4] segmentation decoder, which adaptively combines audio, visual, and text cues to capture spatio-temporal correspondences between target objects and text expressions.", + "bbox": [ + 89, + 582, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike existing SAM-based models for tasks like AVS and Ref-VOS [19, 24, 29, 33, 35], which either fail to address deep temporal multimodal alignment [19, 33, 35] or fail to leverage SAM's flexible prompt-based nature [24], our proposed TSAM, shown in Fig. 1, uniquely supports temporal modeling for SAM. It addresses the challenge of segmenting temporally varying objects and enabling a more cohesive and robust integration of multimodal cues for ob", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ject segmentation in dynamic audio-visual scenes. Additionally, TSAM introduces a dual-pathway framework with sparse and dense prompting modules that replace SAM's visual prompts with data-driven multimodal prompts, fostering richer and more dynamic prompts.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. TSAM for Ref-AVS", + "text_level": 1, + "bbox": [ + 511, + 180, + 702, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the proposed TSAM method (Fig. 1) for the Ref-AVS task. We begin by presenting the problem formulation, followed by an overview of the foundational components of SAM. Next, we describe the temporal modeling branch and the integration of multimodal prompts in TSAM. Finally, we present the training objective that guides TSAM's training.", + "bbox": [ + 511, + 205, + 903, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Problem formulation", + "text_level": 1, + "bbox": [ + 511, + 323, + 710, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the Ref-AVS task, the input includes a sequence of video frames, corresponding audio, and a reference text expression specifying the objects to be segmented in the video. The goal is to segment target objects within each video frame, using multimodal cues embedded in the text expression. The segmentation output is represented as binary masks, with each mask highlighting the target objects within an individual video frame.", + "bbox": [ + 511, + 345, + 903, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visual input includes $n$ frames sampled at 1-second intervals, where each frame $\\in \\mathbb{R}^{3 \\times 1024 \\times 1024}$ represents RGB values at a resolution of $1024 \\times 1024$ pixels. The audio input, temporally aligned with the video frames, is encoded offline using VGGish [12], resulting in representations $F_{a} \\in \\mathbb{R}^{n \\times 128}$ . Similarly, the text expression is preextracted using RoBERTa [25] to generate a text representation $F_{t} \\in \\mathbb{R}^{l \\times 768}$ , where $l$ denotes the length of the expression.", + "bbox": [ + 511, + 465, + 905, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable effective multimodal prompting within SAM, TSAM first projects both audio and text representations into the SAM visual embedding space with dimensionality $d_{emb}$ . This results in audio cues $F_{a} \\in \\mathbb{R}^{n \\times 1 \\times d_{emb}}$ and text cues $F_{t} \\in \\mathbb{R}^{n \\times l \\times d_{emb}}$ , with the text cues $F_{t}$ temporally expanded along $n$ . These projected cues are then fused across the temporal dimension in TSAM's temporal modality fusion layer (TMFL), as shown in Fig. 1, producing expression-related multimodal cues as follows:", + "bbox": [ + 511, + 602, + 903, + 738 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {a} ^ {\\prime}, F _ {t} ^ {\\prime} = \\operatorname {S A} (\\operatorname {C o n c a t} (F _ {a}, F _ {t})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 750, + 903, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Concat}(\\cdot)$ denotes concatenation, and $\\mathrm{SA}(\\cdot)$ applies temporal self-attention to enable modality fusion over the temporal domain. Following the approach in [36], TSAM integrates a basic cached memory (CM), which stores $\\hat{F}_a$ , an accumulated summary of $F_a'$ over time. This memory captures the mean modality cues across the temporal domain and maintains a running summary of the evolving audio context. Concurrently, the attention-enhanced text cues", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23950", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$F_{t}^{\\prime}$ are added to the text cues $F_{t}$ to create the unified multimodal text cues $\\hat{F}_{t} \\in \\mathbb{R}^{n \\times l \\times d_{emb}}$ , resulting in a cohesive and enriched representation of audio and text cues.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. SAM as a foundational baseline", + "text_level": 1, + "bbox": [ + 89, + 143, + 369, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SAM [15] is the first large-scale pretrained foundational model for guided image segmentation. Trained on a massive dataset consisting of 1 billion masks across 11 million images, SAM has demonstrated remarkable effectiveness and versatility in image segmentation tasks. The core components of SAM are outlined here, with further details available in [15]. SAM's architecture is composed of three key modules:", + "bbox": [ + 89, + 165, + 482, + 286 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) Image encoder. A pre-trained Vision Transformer (ViT) [7], minimally adapted to process high-resolution images, extracts spatial embeddings.", + "(b) Prompt encoder. SAM's prompt encoder handles various user prompts, including sparse (points, boxes, text) and dense (masks) formats, providing regional information about the target objects for segmentation.", + "(c) Mask decoder. The mask decoder consists of two layers. First, the dense prompt embedding is integrated with the image embedding. Then, cross-attention is applied between the image embedding and the sparse prompt embeddings to update both representations. In each attention layer, the original prompt token is re-integrated into both the query and key tokens. Finally, the dot product of the upsampled image embedding and the MLP-mapped sparse prompt embeddings produces the final segmentation mask." + ], + "bbox": [ + 89, + 286, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Temporal modeling branch", + "text_level": 1, + "bbox": [ + 89, + 535, + 336, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SAM image encoder does not account for the temporal dynamics inherent in video frames, limiting its capacity to handle the complex interplay between audio and visual cues in tasks such as object segmentation. To address this, the proposed TSAM enhances SAM's architecture by incorporating a dedicated temporal modeling branch at the final $M$ blocks of the SAM image encoder, as shown in Fig. 1. While the SAM image encoder remains frozen during training, the newly introduced temporal blocks are the only trainable. These temporal blocks mirror the structure of the corresponding SAM encoder blocks, initialized with the same pre-trained weights.", + "bbox": [ + 89, + 556, + 482, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each video frame is processed independently through the SAM image encoder. At each temporal modeling block, TSAM integrates a straightforward cached memory and adapter module, which tracks and captures the evolving temporal and multimodal cues. Specifically, the $m$ -th $(m = 1,2,\\dots,M)$ temporal block receives input $x_{m}$ , combining the output $y_{m-1}$ of the previous temporal block, the output $y_{(N-M)+(m-1)}^{\\mathrm{SAM}}$ of the previous SAM encoder block, and the textual enriched audio cues $\\hat{F}_{a}$ via the adapter module:", + "bbox": [ + 89, + 738, + 482, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {m} = y _ {m - 1} + \\mathrm {C M} _ {m} \\left(y _ {(N - M) + (m - 1)} ^ {\\mathrm {S A M}}\\right) + \\mathrm {A M} _ {m} \\left(\\hat {F} _ {a}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 883, + 482, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{CM}_m(\\cdot)$ denotes the cached memory applied to the SAM block output, and $\\mathrm{AM}_m(\\cdot)$ represents the adapter module applied to the audio cues $F_{a}$ . The adapter module, a bottleneck block shown in Fig. 1, integrates audio cues that have been refined by text cues, facilitating deeper interaction among the audio, visual, and text modalities within the temporal modeling branch.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the overall video cues, $F_{v} = \\{f_{v}^{i}\\}_{i = 1}^{n}$ , where each $f_{v}^{i}$ represents a cue for frame $i$ , are derived by integrating the output from the temporal modeling branch $Y_{M}$ with the output from SAM's image encoder $Y_{N}^{SAM}$ :", + "bbox": [ + 511, + 196, + 903, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {v} = Y _ {M} + Y _ {N} ^ {S A M}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 267, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each video frame visual cues $f_{v}^{i}\\in \\mathbb{R}^{d_{emb}\\times 64\\times 64}$ , with $64\\times 64$ representing the spatial dimensions.", + "bbox": [ + 511, + 297, + 903, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Prompting SAM with multimodal cues", + "text_level": 1, + "bbox": [ + 511, + 338, + 844, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TSAM introduces a dual-pathway framework with dedicated sparse and dense prompting modules (SPM and DPM, respectively) to enhance SAM's segmentation capabilities. The SPM learns global context by leveraging sparse queries derived from multimodal cues, guiding TSAM toward high-level object localization. Meanwhile, the DPM operates at a finer granularity, providing dense prompts aligned with the frame's spatial features, highlighting its reliance on the outputs of SAM's image encoder, as shown in Fig. 1.", + "bbox": [ + 511, + 359, + 905, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sparse prompting module (SPM). To effectively guide segmentation within TSAM using the text expression, SPM first employs a language-guided query selection mechanism to select the $k$ most relevant audio cues, associated with the input text expression. The top $k$ query indices, denoted as $A_{k}$ , are selected as:", + "bbox": [ + 511, + 496, + 905, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {k} = \\operatorname {T o p} _ {k} \\left(\\max \\left(\\hat {F} _ {a} \\hat {F} _ {t} ^ {\\top}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 597, + 903, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Top}_k$ picks the top $k$ indices, maxf computes the maximum along the cues' dimension, and $(\\cdot)^{\\top}$ indicates matrix transposition. The selected audio cues are then fused with the visual cues through a cross-attention mechanism and transformed using a feed forward network-mapped sparse prompt embedding. This process generates contextually enriched prompts for SAM's mask decoder.", + "bbox": [ + 511, + 628, + 905, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dense prompting module (DPM). The DPM in TSAM generates spatially comprehensive embeddings by integrating visual, audio, and textual cues through cross-attention mechanisms, as illustrated in Fig. 1. First, an audio cross-attention layer aligns the audio cues with the visual cues, facilitating effective fusion of both modalities. Next, a text cross-attention layer incorporates the text cues, further refining the combined audio-visual cues to focus on the target object for segmentation. Finally, a feed forward network refines the multimodal embeddings, preparing them for input into SAM's mask decoder.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23951", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training objective", + "text_level": 1, + "bbox": [ + 89, + 90, + 267, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ binary cross-entropy $\\mathcal{L}_{\\mathrm{BCE}}$ and intersection over union $\\mathcal{L}_{\\mathrm{IoU}}$ losses to measure the discrepancy between TSAM's predicted mask and the ground truth mask during training. The total loss $\\mathcal{L}_{\\mathrm{total}}$ for model optimization is defined as a weighted sum of $\\mathcal{L}_{\\mathrm{BCE}}$ and $\\mathcal{L}_{\\mathrm{IoU}}$ losses, as follows:", + "bbox": [ + 89, + 112, + 483, + 189 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {B C E}} + \\lambda \\cdot \\mathcal {L} _ {\\mathrm {I o U}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 195, + 482, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is set to 1.0 to balance the two loss components.", + "bbox": [ + 89, + 218, + 467, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 244, + 223, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 268, + 186, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated TSAM using the Ref-AVS dataset [36], which contains a diverse collection of 20,000 text expressions and pixel-level annotations across 4,000 10-second videos. The dataset features a wide range of audible objects, including 20 categories of musical instruments, 8 animals, 15 machines, and 5 humans, as well as a smaller set of 3 categories of static, inaudible objects.", + "bbox": [ + 89, + 291, + 483, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The dataset is divided into three sets: a training set of 2,908 videos, a validation set of 276 videos, and a test set of 818 videos. The test set is further divided into three subsets: 292 Seen videos, which includes object categories present in the training set; 269 Unseen videos, which consists of object categories not seen during training but belonging to super-categories represented in the training data, assessing the model's ability to generalize to new objects; and 257 Null videos, where the text expression refers to an object that either does not exist or is not visible in the given context, with empty true masks, requiring the model to avoid segmenting any object.", + "bbox": [ + 89, + 397, + 483, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation metrics", + "text_level": 1, + "bbox": [ + 89, + 585, + 272, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [36], we used the Jaccard index $\\mathcal{I}$ and F-score $\\mathcal{F}$ as performance metrics. For the Null test set, we employed:", + "bbox": [ + 89, + 607, + 483, + 638 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS = \\sqrt {\\frac {\\text {p r e d i c t e d m a s k a r e a}}{\\text {b a c k g r o u n d a r e a}}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 643, + 482, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a metric to assess expression guidance in the model. A higher $S$ value indicates a larger mask relative to the background, suggesting less accurate expression guidance.", + "bbox": [ + 89, + 690, + 482, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 742, + 305, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We initialized TSAM for the Ref-AVS task using the pretrained ViT-B variant of SAM ( $N = 12$ , $d_{emb} = 256$ ). The temporal branch consists of $M = 4$ blocks, and the selected audio queries is $k = 5$ . Training was performed with PyTorch on eight AMD GPUs in a distributed setup. The AdamW optimizer was used, with an initial learning rate of $1 \\cdot 10^{-4}$ and a batch size of 1. The model was trained for 15 epochs with periodic evaluations, and the best-performing model on the validation subset was selected for testing.", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1651684acccf96378f73b9d61138ced4eda9cfbc92e6a24cd6d80453a640f1a7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Performance comparison on the Ref-AVS dataset. ${}^{ \\dagger }$ denotes text integration in AVS from [36], ${}^{ \\ddagger }$ indicates audio integration in Ref-VOS from [36], and * marks our implementation with text added to AVS." + ], + "table_body": "
MethodTaskVisual BackboneSeenUnseenNull S(↓)
J(%)FJ(%)F
AVSBench [39]AVS†PVT-v223.200.51132.360.5470.208
AVSegFormer [8]AVS†PVT-v233.470.47036.050.5010.171
GAVS [35]AVS†SAM28.930.49829.820.4970.190
SAMA [24]AVS*SAM39.220.56247.500.5660.130
ReferFormer [37]Ref-VOS‡V-Swin31.310.50130.400.4880.176
R2-VOS [18]Ref-VOS‡V-Swin25.010.41027.930.4980.183
EEMC [36]Ref-AVSM2F34.200.51349.540.6480.007
TSAM (Ours)Ref-AVSSAM43.430.56854.580.6640.017
", + "bbox": [ + 516, + 88, + 905, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Results and comparison", + "text_level": 1, + "bbox": [ + 511, + 300, + 733, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compared the proposed TSAM method with the EEMC method [36] and existing methods from the AVS and Ref-VOS domains, including AVSBench [39], AVSeg-Former [8], GAVS [35], SAMA [24], ReferFormer [37], and R2-VOS [18], using the Ref-AVS dataset [36]. For consistency and fairness, we reported the published results for all methods in [36] and re-evaluated the SAMA method by integrating text with both audio and visual inputs to align with our multimodal setup.", + "bbox": [ + 511, + 323, + 906, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results, presented in Tab. 1, demonstrate that even when augmented with text (for AVS methods) or audio (for Ref-VOS methods), these methods do not yield satisfactory performance in the Ref-AVS task. Interestingly, methods that utilize prior segmentation visual backbones, such as SAM and M2F (Mask2Former), tend to outperform those based on visual backbones like PVT-v2 and V-Swin.", + "bbox": [ + 511, + 460, + 905, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although GAVS and SAMA are SAM-based methods, they perform worse than EEMC. SAMA fails to leverage SAM's flexible, promptable nature, integrating additional modality cues superficially by layering them on SAM's frozen image encoder. This underutilizes the potential of multimodal cues, limiting SAMA's segmentation capabilities. GAVS, on the other hand, prompts SAM with multimodal cues but does not address the necessary alignment and interplay among these modalities, leading to a lack of cohesive multimodal fusion. These results confirm that simple modality fusion alone is insufficient for handling the complexity of multimodal integration in the Ref-AVS task.", + "bbox": [ + 511, + 566, + 905, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast, our TSAM leverages language-guided audio queries as sparse prompts, integrates multimodal dense prompts via cross-attention, and incorporates a temporal modeling branch and adapter module to enhance SAM's ability to perform accurate, contextually guided segmentation of dynamic objects in complex audio-visual scenes. Compared to the best-performing method, EEMC, on the Ref-AVS dataset, our TSAM achieves superior results: a $\\mathcal{I} = 9.23\\%$ and $\\mathcal{F} = 5.5\\%$ improvements on the Seen test set, and a $\\mathcal{J} = 5.04\\%$ and $\\mathcal{F} = 1.6\\%$ improvements", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23952", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on the Unseen test set. Although TSAM falls slightly behind EEMC on the Null test set by $S = 1\\%$ , likely due to SAM's limitation of always producing a segmentation mask even when no target object is present, TSAM maintains a strong advantage on Unseen test set. This is primarily due to TSAM's ability to generalize to unseen object categories, preserving SAM's pretrained knowledge while ensuring robust segmentation in novel audio-visual scenes, thanks to the added temporal modeling branch.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Qualitative results", + "text_level": 1, + "bbox": [ + 89, + 239, + 269, + 255 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluated our TSAM method qualitatively on both the Seen and Unseen test sets, as shown in Fig. 2 and Fig. 3, respectively. Fig. 2 demonstrates TSAM's consistent ability to produce high-quality masks for targeted objects, even when they are inaudible, as seen in the middle example. This highlights TSAM's proficiency in processing complex multimodal instructions through its prompting modules, enabling accurate segmentation based on textual cues. On the Unseen test set, Fig. 3, TSAM showcases a remarkable capacity to generalize to novel objects, effectively aligning audio-visual and textual inputs in novel scenes to accurately segment the targeted object of interest. This underscores TSAM's ability to generalize to new contexts, interpreting diverse multimodal instructions and segmenting unseen objects. TSAM's temporal branch enhances understanding of novel objects in dynamic scenes, while SAM's pre-trained knowledge is preserved for core segmentation capabilities. Additional qualitative results and comparisons are provided in the supplementary materials.", + "bbox": [ + 89, + 261, + 483, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 560, + 241, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyzed the impact of omitting key components of TSAM, including the temporal branch (TB), temporal modality fusion layer (TMFL), dense prompting module (DPM), sparse prompting module (SPM), cached memory (CM), and adapter module (AM), as shown in Tab. 2. $\\mathrm{CM(a + v)}$ denotes cached memory for both audio and visual modalities, while $\\mathrm{CM(v)}$ denotes cached memory on the visual modality only, see Fig. 1. Additionally, Tab. 2 presents the effect of the intersection over union $(\\mathcal{L}_{\\mathrm{IoU}})$ loss, defined in Eq. (5), on TSAM's training.", + "bbox": [ + 89, + 582, + 483, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 2, we observe a significant performance degradation in setting $②$ when the temporal branch (TB) is removed, with a drop of $7.24\\% \\mathrm{pt}$ in $\\mathcal{I}$ (from $49.01\\%$ to $41.77\\%$ ) and $3.4\\% \\mathrm{pt}$ in $\\mathcal{F}$ (from 0.616 to 0.582). This performance loss is significantly higher than that in other settings, emphasizing the critical role of the temporal modeling branch in TSAM's ability to generalize across video frames and maintain coherence in dynamic scenes, which is essential for object segmentation in complex Ref-AVS tasks. Settings $④$ and $⑤$ demonstrate the importance of prompting SAM with multimodal cues through SPM and", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa954cd612c52b82f28f8af8f5ed46650d45481d70b18fa1b33bc8b64547e88b.jpg", + "image_caption": [ + "Expression: The object making a sound by being played by the woman." + ], + "image_footnote": [], + "bbox": [ + 527, + 97, + 893, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c65021cc090792bd5d70e6eceda7a8215590934a7f7891e1e0c68f310a1780a.jpg", + "image_caption": [ + "Expression: The object behind the sounding women." + ], + "image_footnote": [], + "bbox": [ + 527, + 224, + 893, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f621832f3b34724703f53198d475c12598545d490505d2b379d746a86485808a.jpg", + "image_caption": [ + "Expression: The sounding object on the legs of the man on the left.", + "Figure 2. Segmentation masks for referred objects in the Ref-AVS dataset (Seen test set). Top row: ground truth with target object in green. Bottom row: our results with target object in violet." + ], + "image_footnote": [], + "bbox": [ + 527, + 352, + 893, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DPM. SPM and DPM play complementary roles in TSAM, and their joint contribution is pivotal for its overall performance. Omitting these prompts in setting ⑥ causes a notable decrease in segmentation performance, reinforcing our claim that the rich semantic cues from multimodal prompts are essential for improving segmentation performance.", + "bbox": [ + 511, + 566, + 906, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings ⑦ and ⑧ evaluate the effect of cached memory on both audio and visual (CM(a+v)) and visual-only", + "bbox": [ + 511, + 657, + 906, + 689 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/acd1b89c38cee2b7d774925c1e8d5742c23139b3c793bce86b835889db314762.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingSeenUnseenMix(S+U)Null S(↓)
J(%)FJ(%)FJ(%)F
① TSAM43.430.56854.580.66449.010.6160.017
② - TB33.050.50750.480.65741.770.5820.505
③ - TMFL40.350.57945.540.62742.950.6030.018
④ - DPM42.720.58049.100.64745.910.6140.018
⑤ - SPM43.040.58049.750.65246.400.6160.018
⑥ - SPM+DPM42.600.60240.580.60441.590.6030.018
⑦ - CM(a+v)42.070.54449.110.65945.590.6020.018
⑧ - CM(v)42.750.54951.180.66046.970.6050.018
⑨ - AM43.130.60040.790.61741.960.6090.017
⑩ - LloU38.290.56442.150.63140.220.5980.008
", + "bbox": [ + 522, + 713, + 898, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on TSAM components and $\\mathcal{L}_{IotU}$ loss. 'Mix' refers to the average performance on the Seen (S) and Unseen (U) test sets.", + "bbox": [ + 511, + 854, + 906, + 895 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23953", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c7b03e93af18185a0fe6dd4255c0ac7d28bf6f51c7d5b7a95d071667d9020cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 98, + 179, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c32c6da14c5f666204c3bce2e306ac792fd2a4447757fc814cc378d195f212f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 99, + 251, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab6589e26733cbfeaae20a626a2a74d187d883e0c5e54b190b5f7c2c1d943635.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 99, + 321, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33b9aada7beee949c280a65aecc91433c8f2ab8ae0a6fb02b82281f41c6a5656.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 99, + 393, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec4babeb3115fcbfabfa9d094834d8d459fbae6a8e5e9e78adc07846b62c487f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 99, + 467, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f7aa8fb6caedfc59e0cbe6ae5563e445b31bf15ba86bcc908104324d29bb6ea.jpg", + "image_caption": [ + "Expression: The object emits a sound following the man." + ], + "image_footnote": [], + "bbox": [ + 107, + 156, + 179, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4b75ef3ccb70fdbee0ab4814ac3c9cd31a33ff2b33258d6748beddc9fc6aca9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 156, + 250, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d731b9618978db5dd580d34b032fc26a57a82a18d332e9f4f426b35697da581f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 156, + 321, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f50b84d51ba7aa0130b6eaa201fef6876282ec07ba89cc474f2474f97694d335.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 156, + 393, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d55b504018cf187ef7fdc6f3c982cedd632cff44f7196b73e28b714465ec3fa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 156, + 467, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9316f29dbac323452a8d1bb6bd3e6d1813fb47e3af7606f8a9e827774a24b331.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 224, + 178, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48043a4ff06e3f3f07bf9d2953df3d509b9912f82e5c4731f33253cdf01c8edd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 224, + 250, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/30bc5a0f57d31fb99726c3c45b75b9c35e5b7fcb6d7657b9170df88376eec400.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 224, + 321, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8e4d33d4f566bf13e25228874f46641604ed59d23ecf6ebfd02dc0e056c45bf7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 224, + 393, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af52a2d194c050723d4f408cbbf3acf1456a634d45a71fb9ce65a55daa9ed37c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 224, + 467, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e3e38f98b194bd14b2595f889eb55708dda7d1f7a32f28211186040b99e8083.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 282, + 178, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd6e51ee8b5e70966ad066467c91dd8fbe1b9aa429e31a2245850149d4ea7ad9.jpg", + "image_caption": [ + "Expression: The truck moving." + ], + "image_footnote": [], + "bbox": [ + 179, + 282, + 250, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf93de1c2b3311e4075fddaf854f346fd63dbcde11c8e943aaaad9db0ec79b37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 282, + 321, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db8d19084ca67bf8a8dbc5d6e43b92d69f777e97105ea05911c363cdeccbc8d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 282, + 393, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f2f00566eae7bd3f77a6c3f89ac282c3fe454808c6ce58e561669b2c036bc63a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 282, + 467, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/502711f764fe10c94a05052468a5334985de8dbd10a6d8209d779ac0abfd6216.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 353, + 178, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3742fa8f2bacc2e2eb029dd50cc1b09e7cc40c35ae6260f3e1a9ea04d2d9b7e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 353, + 251, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d1f5e30f311036dcb309521287ff1fe269f079980caac3b4ca969dca6feb071.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 353, + 321, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93d6713fe4177877f852d01d0599507ada46ffd6198384b8e79f4b569f716521.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 353, + 393, + 409 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb02d1af3940e3208c0dae0699e861bd468293a71fd77741c54be93c8cc18c05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 353, + 467, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61a7bfed7db84710848a2012b9f6bbdec506c9bd816a560b5abd206bc8093198.jpg", + "image_caption": [ + "Expression: The tuba being played by a man.", + "Figure 3. Segmentation masks for referred objects in the Ref-AVS dataset (Unseen test set). Top row: ground truth with target object in green. Bottom row: our results with target object in violet." + ], + "image_footnote": [], + "bbox": [ + 107, + 410, + 178, + 464 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98fe3b13283bed1cb72c8b625aa0d65e5a9a4b48ff2a9c54e58b3486d52fd2fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 410, + 250, + 464 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0034bca49334dcff8e003e01bb79eeb4561a08c890f881f67da96789d64d4d3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 410, + 321, + 464 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44433495be3ce2ebba79580956a614f87c40deea78409b0b20cfbb67204582bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 410, + 393, + 464 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0bb20ecef774b8cc3526fa1af37b624dd489ba20885f941ee2a27506ef3fed29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 410, + 467, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$(\\mathrm{CM}(\\mathrm{v}))$ modalities, demonstrate a performance degradation in TSAM when cached memory is not applied across modalities $(\\mathrm{CM}(\\mathrm{a} + \\mathrm{v}))$ and when only the visual modality's memory $(\\mathrm{CM}(\\mathrm{v}))$ is omitted. The latter specifically emphasizes the importance of maintaining shared memory across video frames for effective temporal alignment and multimodal integration, as evidenced by the results.", + "bbox": [ + 89, + 566, + 482, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results from setting ⑨ highlight the adapter module's importance. Its omission causes a significant performance drop, underscoring its crucial role in TSAM. We also observe that the components of TSAM, particularly in settings ③ - ⑨, have a substantial impact on performance, especially on the Unseen test set. This set challenges the model to segment novel objects, demonstrating TSAM's ability to generalize beyond seen categories, an essential feature for robust performance in open-world scenarios.", + "bbox": [ + 89, + 672, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, setting ⑩ highlights the distinct contribution of $\\mathcal{L}_{\\mathrm{IoU}}$ to the overall loss function $\\mathcal{L}_{\\mathrm{total}}$ . Notably, the model performs best on both Seen and Unseen test sets when $\\mathcal{L}_{\\mathrm{IoU}}$ is included in the total loss. Omitting $\\mathcal{L}_{\\mathrm{IoU}}$ results in the highest performance on the Null test set. Further ablation studies of $\\lambda$ (Eq. (5)) are in the supplementary materials.", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b9248c9e9e5c7af0b14a881491539e6bfb21eb46e06c6d2b14e2f29804746d00.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TSAMSeenUnseenMix(S+U)Null S(↓)
J(%)FJ(%)FJ(%)F
k=343.580.57950.430.65547.010.6170.018
k=543.430.56854.580.66449.010.6160.017
k=743.240.57346.950.63045.100.6010.018
M=242.040.56653.510.65147.760.6090.023
M=443.430.56854.580.66449.010.6160.017
M=643.270.57549.860.64046.570.6080.020
", + "bbox": [ + 524, + 88, + 898, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Effect of audio queries $k$ and temporal branch depth $M$ on TSAM's performance.", + "bbox": [ + 511, + 202, + 903, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further evaluated the impact of varying the number of audio queries $k$ , defined in Eq. (4), and the temporal branch depth $M$ , see Sec. 3.3, on TSAM's performance, as shown in Tab. 3. Among the configurations $k = 3$ , $k = 5$ , and $k = 7$ , we observe that $k = 5$ performs best. This balance optimizes the relevance of audio cues without overloading the model with irrelevant queries to SAM's mask decoder.", + "bbox": [ + 511, + 257, + 903, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarly, we find that $M = 4$ yields optimal results for the temporal branch. A shallower setup ( $M = 2$ ) likely lacks the temporal depth required for effective generalization and multimodal integration, while a deeper branch ( $M = 6$ ) may introduce excessive complexity, impairing the model's generalization to unseen objects.", + "bbox": [ + 511, + 363, + 903, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 467, + 638, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presented TSAM, a novel extension of the Segment Anything Model (SAM) tailored for the referring audio-visual segmentation (Ref-AVS) task. TSAM enhances SAM's segmentation capabilities by effectively integrating multimodal cues through data-driven dense and sparse prompting modules while modeling temporal dependencies via a dedicated temporal branch. This approach preserves SAM's foundational knowledge while improving performance, particularly with novel objects. Extensive experiments on the Ref-AVS dataset demonstrate TSAM's strong generalization ability, showcasing its effectiveness in segmenting objects from textual cues and its robustness in unseen scenarios. These results highlight TSAM's potential to advance cross-modal understanding in multimodal tasks. Limitations and future work. Although TSAM outperforms existing Ref-AVS methods, a direct comparison with the recent SAM2 [31] remains to be explored. SAM2, designed for video segmentation with a memory module for temporal visual representations, lacks multimodal support. Future work will adapt SAM2 for the Ref-AVS task.", + "bbox": [ + 511, + 492, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the Research Council of Finland through the Finnish Center for Artificial Intelligence flagship (grant 345552) and the USSEE project (grant 345791). We also acknowledge CSC-IT Center for Science, Finland, for awarding access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking.", + "bbox": [ + 511, + 795, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23954", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In CVPR, pages 4985-4995, 2022. 3", + "[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the Kinetics dataset. In CVPR, pages 6299-6308, 2017. 3", + "[3] Yuanhong Chen, Yuyuan Liu, Hu Wang, Fengbei Liu, Chong Wang, Helen Frazer, and Gustavo Carneiro. Unraveling instance associations: A closer look for audio-visual segmentation. In CVPR, pages 26497–26507, 2024. 3", + "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 4", + "[5] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 1", + "[6] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 1", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5", + "[8] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In AAAI, pages 12155–12163, 2024. 3, 6", + "[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In CVPR, pages 5958-5966, 2018. 1, 3", + "[10] Yuxin Guo, Siyang Sun, Shuailei Ma, Kecheng Zheng, Xiaoyi Bao, Shijie Ma, Wei Zou, and Yun Zheng. Cross-MAE: Cross-modality masked autoencoders for region-aware audio-visual pre-training. In CVPR, pages 26721–26731, 2024. 1", + "[11] Dawei Hao, Yuxin Mao, Bowen He, Xiaodong Han, Yuchao Dai, and Yiran Zhong. Improving audio-visual segmentation with bidirectional generation. In AAAI, pages 2067–2075, 2024. 3", + "[12] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. CNN architectures for large-scale audio classification. In ICASSP, pages 131-135, 2017. 4", + "[13] Xiao Hu, Basavaraj Hampiholi, Heiko Neumann, and Jochen Lang. Temporal context enhanced referring video object segmentation. In WACV, pages 5574-5583, 2024. 1", + "[14] Shaofei Huang, Han Li, Yuqing Wang, Hongji Zhu, Jiao Dai, Jizhong Han, Wenge Rong, and Si Liu. Discovering sounding objects by audio queries for audio visual segmentation. In IJCAI, pages 875-883, 2023. 3" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 1, 3, 5", + "[16] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, pages 19108-19118, 2022. 1", + "[17] Kexin Li, Zongxin Yang, Lei Chen, Yi Yang, and Jun Xiao. Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. In ACM MM, pages 1485–1494, 2023. 3", + "[18] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. In ICCV, pages 22236-22245, 2023. 3, 6", + "[19] Yonglin Li, Jing Zhang, Xiao Teng, Long Lan, and Xinwang Liu. RefSAM: Efficiently adapting segmenting anything model for referring video object segmentation. arXiv preprint arxiv.org/abs/2307.00997, 2024. 2, 3, 4", + "[20] Yuhang Ling, Yuxi Li, Zhenye Gan, Jiangning Zhang, Ming-min Chi, and Yabiao Wang. TransAVS: End-to-end audiovisual segmentation with transformer. In ICASSP, pages 7845-7849, 2024. 3", + "[21] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In ACM MM, pages 7590–7598, 2023. 1", + "[22] Chen Liu, Peike Li, Hu Zhang, Lincheng Li, Zi Huang, Dadong Wang, and Xin Yu. BAVS: Bootstrapping audiovisual segmentation by integrating foundation knowledge. IEEE TMM, 26:10015-10028, 2024. 3", + "[23] Jinxiang Liu, Chen Ju, Chaofan Ma, Yanfeng Wang, Yu Wang, and Ya Zhang. Audio-aware query-enhanced transformer for audio-visual segmentation. arXiv preprint arXiv:2307.13236, 2023. 3", + "[24] Jinxiang Liu, Yu Wang, Chen Ju, Chaofan Ma, Ya Zhang, and Weidi Xie. Annotation-free audio-visual segmentation. In WACV, pages 5604–5614, 2024. 1, 2, 4, 6", + "[25] Y Liu, M Ott, N Goyal, J Du, M Joshi, D Chen, O Levy, M Lewis, L Zettlemoyer, and V Stoyanov. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4", + "[26] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In CVPR, pages 3202-3211, 2022. 3", + "[27] Yuxin Mao, Jing Zhang, Mochu Xiang, Yunqiu Lv, Yi ran Zhong, and Yuchao Dai. Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579, 2023. 3", + "[28] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In ICCV, pages 954–965, 2023. 3", + "[29] Shentong Mo and Yapeng Tian. AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836, 2023. 2, 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23955", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Federico Perazzi, Anna Khoreva, Rodrigo Benenson, Bernt Schiele, and Alexander Sorkine-Hornung. Learning video object segmentation from static images. In CVPR, pages 2663-2672, 2017. 1", + "[31] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. In ICLR, 2025. 8", + "[32] Seonguk Seo, Joon-Young Lee, and Bohyung Han. URVOS: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223, 2020. 3", + "[33] Juhyeong Seon, Woobin Im, Sebin Lee, Jumin Lee, and Sung-Eui Yoon. Extending segment anything model into auditory and temporal dimensions for audio-visual segmentation. In ICIP, pages 2480–2486, 2024. 1, 2, 4", + "[34] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. PVT v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3", + "[35] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. In AAAI, pages 5669–5677, 2024. 1, 2, 4, 6", + "[36] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-AVS: Refer and segment objects in audio-visual scenes. In ECCV, pages 196–213, 2024. 1, 3, 4, 6", + "[37] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 3, 6", + "[38] Jiarui Yu, Haoran Li, Yanbin Hao, Jinmeng Wu, Tong Xu, Shuo Wang, and Xiangnan He. How can contrastive pretraining benefit audio-visual segmentation? A study from supervised and zero-shot perspectives. In BMVC, pages 367–374, 2023. 3", + "[39] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In ECCV, pages 386–403, 2022. 1, 3, 6" + ], + "bbox": [ + 91, + 92, + 480, + 684 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23956", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/8535e16b-e9a4-4280-8556-88908a875920_model.json b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/8535e16b-e9a4-4280-8556-88908a875920_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bd21aea9544e2d487750507dab6f654d1d429a63 --- /dev/null +++ b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/8535e16b-e9a4-4280-8556-88908a875920_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.13, + 0.882, + 0.175 + ], + "angle": 0, + "content": "TSAM: Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.204, + 0.737, + 0.257 + ], + "angle": 0, + "content": "Abduljalil Radman & Jorma Laaksonen \nDepartment of Computer Science, Aalto University, Finland \n{abduljalil.saif, jorma.laaksonen}@aalto.fi" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.486, + 0.705 + ], + "angle": 0, + "content": "Referring audio-visual segmentation (Ref-AVS) aims to segment objects within audio-visual scenes using multimodal cues embedded in text expressions. While the Segment Anything Model (SAM) has revolutionized visual segmentation, its applicability to Ref-AVS, where multimodal cues act as novel prompts, remains unexplored. SAM's limitation to single-frame segmentation also hinders its ability to capture essential temporal context needed for multi-frame audio-visual segmentation. To address this gap, we propose TSAM, a novel extension of SAM designed to leverage multimodal cues for precise segmentation in dynamic audio-visual scenes. TSAM enhances SAM's image encoder with a temporal modeling branch, enabling spatio-temporal learning and deep multimodal fusion across video frames, while retaining SAM's pre-trained knowledge. Additionally, TSAM replaces SAM's user-interactive prompting mechanism with sparse and dense data-driven prompts, enabling more effective integration of audio-visual inputs and reference text expressions. Extensive experiments on the Ref-AVS dataset demonstrate TSAM's superiority over state-of-the-art methods. The results illustrate its effectiveness in segmenting objects in dynamic audio-visual scenes using text-based multimodal cues and its strong generalization to unseen objects. Project webpage: https://abdurad.github.io/TSAM/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In real-world scenarios, people often rely on multimodal information, such as audio-visual and visual-text, to gather spatiotemporal and semantic insights for distinguishing objects of interest [10, 39]. This multimodal approach is equally valuable for machines, enabling applications like stage cameras to track a performing musician or movie cameras to focus on a superhero confronting a yelling villain. Relying on a single sensory input, such as vision, is often insufficient for accurate object segmentation in complex," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.905, + 0.354 + ], + "angle": 0, + "content": "dynamic audio-visual scenes [16, 36]. In the literature, researchers have addressed this need for multimodal integration in object segmentation through three main approaches, each focused on a specific modality cue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.908, + 0.642 + ], + "angle": 0, + "content": "First, video object segmentation (VOS) [6, 30] relies on a mask annotation in the initial frame to guide segmentation of objects across subsequent frames. While VOS allows for flexible segmentation, it is heavily dependent on labor-intensive frame-level annotations, making it challenging to scale. Second, referring video object segmentation (Ref-VOS) [5, 9, 13] replaces the need for mask annotations with text expressions to identify targeted objects. This approach introduces a more intuitive, text-based method for segmentation, but struggles with tracking and segmenting objects in more dynamic audio-visual environments. Third, audio-visual segmentation (AVS) [21, 24, 33, 35, 39] leverages audio cues to segment sounding objects in video, effectively handling complex audio-visual scenarios, but remaining limited in scope as it does not generalize to inaudible objects. Each of these approaches, VOS, Ref-VOS, and AVS, contributes valuable segmentation capabilities within its modality focus, but remains challenged in more naturalistic audio-visual scenarios [36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Recently, Wang et al. [36] introduced referring audio-visual segmentation (Ref-AVS), a task that segments objects in video by interpreting text expressions enriched with both audio and visual cues. They also presented Expression Enhancing with Multimodal Cues (EEMC), a simple method that integrates reference text expressions with dynamic audio-visual cues to segment targeted objects in scenes. By unifying these multimodal cues, EEMC enables robust segmentation in dynamic scenes, overcoming limitations of prior approaches that focused on narrower combinations of modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We propose that enhancing Ref-AVS can be achieved by leveraging prior knowledge from large-scale visual foundation models, such as the Segment Anything Model (SAM) [15], for segmentation. SAM is a flexible segmentation model that utilizes an image encoder to generate high-dimensional image embeddings. It incorpo" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23947" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.098, + 0.592, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.173, + 0.895, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.59, + 0.908, + 0.634 + ], + "angle": 0, + "content": "Figure 1. Overview of TSAM. Left: TSAM integrates multimodal cues from diverse inputs, leveraging the prompting modules to guide segmentation. Right: The temporal branch fuses multimodal cues and learns temporal context for segmentation in audio-visual scenes. The adapter input, \\(\\dot{F}_a\\), is the output from the cached memory." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.483, + 0.87 + ], + "angle": 0, + "content": "rates user-provided visual prompts, which can be sparse (e.g., points, boxes) or dense (e.g., masks), processed by a dedicated prompt encoder. These embeddings are then passed through a lightweight, transformer-based mask decoder to generate accurate object masks based on the provided prompts. Although SAM has been applied to tasks such as AVS [24, 29, 33, 35], and Ref-VOS [19], its potential for segmenting objects in the visual domain using multimodal cues embedded in text expressions remains unexplored. This makes its application to the Ref-AVS task both novel and challenging as it requires the integration of diverse multimodal cues to guide segmentation, overcoming the limitations of SAM's traditional reliance on explicit user-interactive prompts." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.871, + 0.483, + 0.886 + ], + "angle": 0, + "content": "The main challenge in adapting SAM for the Ref-AVS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.779 + ], + "angle": 0, + "content": "task, which requires nuanced multimodal understanding without direct user guidance, arises from SAM's lack of temporal awareness in its image encoder, limiting its ability to capture complex interactions across audio, visual, and text modalities. Additionally, SAM relies on user-provided visual prompts for explicit spatial information of targeted objects, while its capacity to process text and audio modalities remains limited." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.886 + ], + "angle": 0, + "content": "This paper introduces TSAM (Fig. 1), a novel SAM-based architecture for the Ref-AVS task, designed to enhance segmentation by leveraging multimodal cues from audio, visual, and text inputs. TSAM incorporates an auxiliary temporal modeling branch in parallel with SAM's image encoder, preserving SAM's zero-shot capabilities while adapting it for temporally varying video content. More-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23948" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "over, TSAM uses text cues to guide the selection of relevant audio cues, which are then used as sparse prompts, enriched with key visual cues, to query SAM's mask decoder. Since multimodal prompts lack explicit spatial information for the target objects, TSAM incorporates a lightweight adapter module into the temporal branch to enhance SAM's mask decoder prompts, facilitating deeper multimodal interactions during training. Through a cross-attention mechanism, TSAM synthesizes data-driven multimodal reference cues as dense prompts for SAM. Extensive experiments show that TSAM outperforms existing state-of-the-art methods in segmenting target objects in the Ref-AVS task. The main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.29, + 0.483, + 0.35 + ], + "angle": 0, + "content": "1. We propose TSAM, an end-to-end architecture built on the Segment Anything Model (SAM), which redefines traditional visual prompts as data-driven multimodal prompts for object segmentation in Ref-AVS tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.35, + 0.483, + 0.409 + ], + "angle": 0, + "content": "2. We introduce a temporal modeling branch that enhances SAM's image encoder with temporal awareness, enabling the capture of intricate spatio-temporal interactions across multimodal cues." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.41, + 0.483, + 0.516 + ], + "angle": 0, + "content": "3. Through comprehensive experiments, we demonstrate that the proposed TSAM achieves superior performance compared to state-of-the-art methods in Ref-AVS. Furthermore, extensive ablation studies provide a detailed evaluation of TSAM's architecture, emphasizing the pivotal role of multimodal cue integration and temporal modeling in enhancing segmentation performance." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.29, + 0.483, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.534, + 0.228, + 0.549 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.483, + 0.577 + ], + "angle": 0, + "content": "2.1. Referring video object segmentation (Ref-VOS)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Referring video object segmentation (Ref-VOS) involves segmenting target objects within video frames based on textual descriptions. Gavrilyuk et al. [9] introduced an encoder-decoder model based on I3D [2], that generates dynamic convolutional filters from text and applies them to visual representations at different resolutions to produce segmentation masks for target objects. URVOS [32] enhanced object tracking by performing segmentation on a key frame and then propagating the mask across the video. This approach combines cross-modal attention to prevent drift and memory attention to maintain temporal consistency across frames. MTTR [1] formulated Ref-VOS as a sequence prediction task, utilizing multimodal transformers to process video and text jointly. ReferFormer [37] used Video-Swin (V-Swin) [26] as the visual backbone, linking textual expressions directly to relevant frame regions. In this approach, language-conditioned object queries serve as inputs, helping the model identify the referred objects and apply dynamic kernels as convolution filters to generate segmentation masks. Although Ref-VOS enhances user accessibility by replacing mask annotations with natural lan" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "guage expressions, it still struggles to accurately identify objects in dynamic audio-visual scenes [36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.905, + 0.32 + ], + "angle": 0, + "content": "R2-VOS [18] extended the Ref-VOS task to handle unpaired visual-text inputs, jointly modeling video segmentation and text reconstruction. By introducing a structural text-to-text cycle constraint, R2-VOS improves multimodal alignment across diverse text inputs, overcoming the limitations of prior point-wise methods. RefSAM [19] utilized text prompts with SAM [15], using a lightweight cross-modal MLP to project the referring expression's text embedding into prompt embeddings. RefSAM enables multimodal reference segmentation by integrating textual and visual modalities, but it does not fully exploit the simultaneous interaction of multimodal cues or support the temporal dynamics required for video segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.333, + 0.804, + 0.349 + ], + "angle": 0, + "content": "2.2. Audio-visual segmentation (AVS)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.688 + ], + "angle": 0, + "content": "Audio-visual segmentation (AVS), introduced by Zhou et al. [39], is a challenging task that seeks to achieve pixel-level segmentation of sound-emitting objects in video, guided by the audio modality. AVSBench [39] used a temporal pixel-wise audio-visual interaction module to inject audio semantics as guidance for the visual segmentation process. Similarly, AVSegFormer [8], a transformer-based model built on the PVT-v2 backbone [34], used cross-attention to fuse spatial-temporal visual cues with audio cues. By decoding these combined audio-visual cues and employing audio queries, the model enhanced its ability to focus on sounding objects within the audio-visual scene. Other transformer-based models proposed in [14, 17, 23] integrated audio and visual cues across both temporal and spatial dimensions for the AVS task. Alternative approaches [20, 22] disentangled the audio input into audio queries, which interacts with the visual input to decode segmentation masks. These approaches emphasize implicit modality fusion strategies, training models to fit discrete samples in the dataset. However, their performance often falls short due to limitations in dataset size and diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Alternatively, a Explicit Conditional Multimodal Variational Auto-Encoder (ECMVAE) [28] was proposed to extract semantically correlated cues from audio-visual inputs for segmenting audible objects in video frames. In another approach, the AVS Bidirectional Generation (AVSBG) framework [11] was introduced, correlating visual cues of objects with their corresponding audio cues through a visual-to-audio projection and implicit volumetric motion estimation. Recent advancements in AVS [3, 27, 38] also integrated audio-visual contrastive learning methods to select informative samples and improve training efficiency and model performance. However, these methods often fail to fully leverage available information due to limited prior segmentation knowledge in their backbone architectures, hin" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23949" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.107 + ], + "angle": 0, + "content": "dering their ability to effectively mine informative samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.483, + 0.395 + ], + "angle": 0, + "content": "On the other hand, SAM-based methods showed improvements in the AVS task. AV-SAM [29] was an early attempt to apply SAM to AVS, utilizing pixel-wise audio-visual fusion to prompt SAM for sounding object segmentation. In [24], two SAM-based approaches for AVS were introduced: Audio-Prompted SAM (AP-SAM) and SAM Adapter (SAMA-AVS). AP-SAM encoded audio as a prompt for SAM's mask decoder, while SAMA-AVS incorporated an adapter module to inject the audio information during the training process, but did not fully leverage SAM's promptable nature. Recently, a Generalizable Audio-Visual Segmentation (GAVS) model [35] was introduced, targeting zero-shot and few-shot scenarios. GAVS proposed a semantic-aware audio prompt to bridge the semantic gap between visual and auditory modalities through contrastive learning. All of these methods take a single image and its corresponding audio as input, fusing multimodal inputs in the spatial domain, which limits their ability to capture temporal relationships frame by frame." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.547 + ], + "angle": 0, + "content": "In a straightforward approach, Seon et al. [33] extended SAM for AVS by integrating a Spatio-Temporal, Bidirectional Audio-Visual Attention (ST-BAVA) module between SAM's image encoder and mask decoder. While this approach aims to model temporal cues for audible object segmentation, it oversimplifies the interaction between audio and visual modalities, treating them in a relatively shallow manner. The model relies on basic feature fusion instead of a more meaningful interaction, limiting its ability to fully leverage the potential of multimodal cues for AVS tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.575 + ], + "angle": 0, + "content": "2.3. Referring audio-visual segmentation (Ref-AVS)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Referring audio-visual segmentation (Ref-AVS) is a novel pixel-level segmentation task in which the model predicts, for each pixel, whether it corresponds to an object specified by multimodal cues in a text expression. Wang et al. [36] introduced the Ref-AVS task and developed a dataset with mask annotations for segmenting objects in video frames guided by multimodal expressions, creating the first benchmark for supervised Ref-AVS. They also proposed the Expression Enhancing with Multimodal Cues (EEMC) model, based on the Mask2Former (M2F) [4] segmentation decoder, which adaptively combines audio, visual, and text cues to capture spatio-temporal correspondences between target objects and text expressions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Unlike existing SAM-based models for tasks like AVS and Ref-VOS [19, 24, 29, 33, 35], which either fail to address deep temporal multimodal alignment [19, 33, 35] or fail to leverage SAM's flexible prompt-based nature [24], our proposed TSAM, shown in Fig. 1, uniquely supports temporal modeling for SAM. It addresses the challenge of segmenting temporally varying objects and enabling a more cohesive and robust integration of multimodal cues for ob" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "ject segmentation in dynamic audio-visual scenes. Additionally, TSAM introduces a dual-pathway framework with sparse and dense prompting modules that replace SAM's visual prompts with data-driven multimodal prompts, fostering richer and more dynamic prompts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.181, + 0.703, + 0.198 + ], + "angle": 0, + "content": "3. TSAM for Ref-AVS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.905, + 0.313 + ], + "angle": 0, + "content": "In this section, we introduce the proposed TSAM method (Fig. 1) for the Ref-AVS task. We begin by presenting the problem formulation, followed by an overview of the foundational components of SAM. Next, we describe the temporal modeling branch and the integration of multimodal prompts in TSAM. Finally, we present the training objective that guides TSAM's training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.324, + 0.712, + 0.339 + ], + "angle": 0, + "content": "3.1. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.905, + 0.466 + ], + "angle": 0, + "content": "In the Ref-AVS task, the input includes a sequence of video frames, corresponding audio, and a reference text expression specifying the objects to be segmented in the video. The goal is to segment target objects within each video frame, using multimodal cues embedded in the text expression. The segmentation output is represented as binary masks, with each mask highlighting the target objects within an individual video frame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.906, + 0.602 + ], + "angle": 0, + "content": "The visual input includes \\( n \\) frames sampled at 1-second intervals, where each frame \\( \\in \\mathbb{R}^{3 \\times 1024 \\times 1024} \\) represents RGB values at a resolution of \\( 1024 \\times 1024 \\) pixels. The audio input, temporally aligned with the video frames, is encoded offline using VGGish [12], resulting in representations \\( F_{a} \\in \\mathbb{R}^{n \\times 128} \\). Similarly, the text expression is preextracted using RoBERTa [25] to generate a text representation \\( F_{t} \\in \\mathbb{R}^{l \\times 768} \\), where \\( l \\) denotes the length of the expression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.905, + 0.739 + ], + "angle": 0, + "content": "To enable effective multimodal prompting within SAM, TSAM first projects both audio and text representations into the SAM visual embedding space with dimensionality \\( d_{emb} \\). This results in audio cues \\( F_{a} \\in \\mathbb{R}^{n \\times 1 \\times d_{emb}} \\) and text cues \\( F_{t} \\in \\mathbb{R}^{n \\times l \\times d_{emb}} \\), with the text cues \\( F_{t} \\) temporally expanded along \\( n \\). These projected cues are then fused across the temporal dimension in TSAM's temporal modality fusion layer (TMFL), as shown in Fig. 1, producing expression-related multimodal cues as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.751, + 0.905, + 0.768 + ], + "angle": 0, + "content": "\\[\nF _ {a} ^ {\\prime}, F _ {t} ^ {\\prime} = \\operatorname {S A} (\\operatorname {C o n c a t} (F _ {a}, F _ {t})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\operatorname{Concat}(\\cdot)\\) denotes concatenation, and \\(\\mathrm{SA}(\\cdot)\\) applies temporal self-attention to enable modality fusion over the temporal domain. Following the approach in [36], TSAM integrates a basic cached memory (CM), which stores \\(\\hat{F}_a\\), an accumulated summary of \\(F_a'\\) over time. This memory captures the mean modality cues across the temporal domain and maintains a running summary of the evolving audio context. Concurrently, the attention-enhanced text cues" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23950" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.138 + ], + "angle": 0, + "content": "\\(F_{t}^{\\prime}\\) are added to the text cues \\(F_{t}\\) to create the unified multimodal text cues \\(\\hat{F}_{t} \\in \\mathbb{R}^{n \\times l \\times d_{emb}}\\), resulting in a cohesive and enriched representation of audio and text cues." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.144, + 0.37, + 0.159 + ], + "angle": 0, + "content": "3.2. SAM as a foundational baseline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.166, + 0.483, + 0.287 + ], + "angle": 0, + "content": "SAM [15] is the first large-scale pretrained foundational model for guided image segmentation. Trained on a massive dataset consisting of 1 billion masks across 11 million images, SAM has demonstrated remarkable effectiveness and versatility in image segmentation tasks. The core components of SAM are outlined here, with further details available in [15]. SAM's architecture is composed of three key modules:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.483, + 0.332 + ], + "angle": 0, + "content": "(a) Image encoder. A pre-trained Vision Transformer (ViT) [7], minimally adapted to process high-resolution images, extracts spatial embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.483, + 0.392 + ], + "angle": 0, + "content": "(b) Prompt encoder. SAM's prompt encoder handles various user prompts, including sparse (points, boxes, text) and dense (masks) formats, providing regional information about the target objects for segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.529 + ], + "angle": 0, + "content": "(c) Mask decoder. The mask decoder consists of two layers. First, the dense prompt embedding is integrated with the image embedding. Then, cross-attention is applied between the image embedding and the sparse prompt embeddings to update both representations. In each attention layer, the original prompt token is re-integrated into both the query and key tokens. Finally, the dot product of the upsampled image embedding and the MLP-mapped sparse prompt embeddings produces the final segmentation mask." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.287, + 0.484, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.536, + 0.338, + 0.552 + ], + "angle": 0, + "content": "3.3. Temporal modeling branch" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.738 + ], + "angle": 0, + "content": "The SAM image encoder does not account for the temporal dynamics inherent in video frames, limiting its capacity to handle the complex interplay between audio and visual cues in tasks such as object segmentation. To address this, the proposed TSAM enhances SAM's architecture by incorporating a dedicated temporal modeling branch at the final \\(M\\) blocks of the SAM image encoder, as shown in Fig. 1. While the SAM image encoder remains frozen during training, the newly introduced temporal blocks are the only trainable. These temporal blocks mirror the structure of the corresponding SAM encoder blocks, initialized with the same pre-trained weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Each video frame is processed independently through the SAM image encoder. At each temporal modeling block, TSAM integrates a straightforward cached memory and adapter module, which tracks and captures the evolving temporal and multimodal cues. Specifically, the \\(m\\)-th \\((m = 1,2,\\dots,M)\\) temporal block receives input \\(x_{m}\\), combining the output \\(y_{m-1}\\) of the previous temporal block, the output \\(y_{(N-M)+(m-1)}^{\\mathrm{SAM}}\\) of the previous SAM encoder block, and the textual enriched audio cues \\(\\hat{F}_{a}\\) via the adapter module:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.884, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\nx _ {m} = y _ {m - 1} + \\mathrm {C M} _ {m} \\left(y _ {(N - M) + (m - 1)} ^ {\\mathrm {S A M}}\\right) + \\mathrm {A M} _ {m} \\left(\\hat {F} _ {a}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "where \\(\\mathrm{CM}_m(\\cdot)\\) denotes the cached memory applied to the SAM block output, and \\(\\mathrm{AM}_m(\\cdot)\\) represents the adapter module applied to the audio cues \\(F_{a}\\). The adapter module, a bottleneck block shown in Fig. 1, integrates audio cues that have been refined by text cues, facilitating deeper interaction among the audio, visual, and text modalities within the temporal modeling branch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.905, + 0.26 + ], + "angle": 0, + "content": "Finally, the overall video cues, \\( F_{v} = \\{f_{v}^{i}\\}_{i = 1}^{n} \\), where each \\( f_{v}^{i} \\) represents a cue for frame \\( i \\), are derived by integrating the output from the temporal modeling branch \\( Y_{M} \\) with the output from SAM's image encoder \\( Y_{N}^{SAM} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.268, + 0.905, + 0.288 + ], + "angle": 0, + "content": "\\[\nF _ {v} = Y _ {M} + Y _ {N} ^ {S A M}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.298, + 0.905, + 0.33 + ], + "angle": 0, + "content": "Each video frame visual cues \\( f_{v}^{i}\\in \\mathbb{R}^{d_{emb}\\times 64\\times 64} \\), with \\( 64\\times 64 \\) representing the spatial dimensions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.339, + 0.846, + 0.355 + ], + "angle": 0, + "content": "3.4. Prompting SAM with multimodal cues" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.906, + 0.497 + ], + "angle": 0, + "content": "TSAM introduces a dual-pathway framework with dedicated sparse and dense prompting modules (SPM and DPM, respectively) to enhance SAM's segmentation capabilities. The SPM learns global context by leveraging sparse queries derived from multimodal cues, guiding TSAM toward high-level object localization. Meanwhile, the DPM operates at a finer granularity, providing dense prompts aligned with the frame's spatial features, highlighting its reliance on the outputs of SAM's image encoder, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.587 + ], + "angle": 0, + "content": "Sparse prompting module (SPM). To effectively guide segmentation within TSAM using the text expression, SPM first employs a language-guided query selection mechanism to select the \\(k\\) most relevant audio cues, associated with the input text expression. The top \\(k\\) query indices, denoted as \\(A_{k}\\), are selected as:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.598, + 0.905, + 0.625 + ], + "angle": 0, + "content": "\\[\nA _ {k} = \\operatorname {T o p} _ {k} \\left(\\max \\left(\\hat {F} _ {a} \\hat {F} _ {t} ^ {\\top}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.734 + ], + "angle": 0, + "content": "where \\(\\mathrm{Top}_k\\) picks the top \\(k\\) indices, maxf computes the maximum along the cues' dimension, and \\((\\cdot)^{\\top}\\) indicates matrix transposition. The selected audio cues are then fused with the visual cues through a cross-attention mechanism and transformed using a feed forward network-mapped sparse prompt embedding. This process generates contextually enriched prompts for SAM's mask decoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Dense prompting module (DPM). The DPM in TSAM generates spatially comprehensive embeddings by integrating visual, audio, and textual cues through cross-attention mechanisms, as illustrated in Fig. 1. First, an audio cross-attention layer aligns the audio cues with the visual cues, facilitating effective fusion of both modalities. Next, a text cross-attention layer incorporates the text cues, further refining the combined audio-visual cues to focus on the target object for segmentation. Finally, a feed forward network refines the multimodal embeddings, preparing them for input into SAM's mask decoder." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23951" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.268, + 0.108 + ], + "angle": 0, + "content": "3.5. Training objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.19 + ], + "angle": 0, + "content": "We employ binary cross-entropy \\(\\mathcal{L}_{\\mathrm{BCE}}\\) and intersection over union \\(\\mathcal{L}_{\\mathrm{IoU}}\\) losses to measure the discrepancy between TSAM's predicted mask and the ground truth mask during training. The total loss \\(\\mathcal{L}_{\\mathrm{total}}\\) for model optimization is defined as a weighted sum of \\(\\mathcal{L}_{\\mathrm{BCE}}\\) and \\(\\mathcal{L}_{\\mathrm{IoU}}\\) losses, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.196, + 0.483, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {B C E}} + \\lambda \\cdot \\mathcal {L} _ {\\mathrm {I o U}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.468, + 0.235 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is set to 1.0 to balance the two loss components." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.245, + 0.224, + 0.263 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.27, + 0.187, + 0.285 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.398 + ], + "angle": 0, + "content": "We evaluated TSAM using the Ref-AVS dataset [36], which contains a diverse collection of 20,000 text expressions and pixel-level annotations across 4,000 10-second videos. The dataset features a wide range of audible objects, including 20 categories of musical instruments, 8 animals, 15 machines, and 5 humans, as well as a smaller set of 3 categories of static, inaudible objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.581 + ], + "angle": 0, + "content": "The dataset is divided into three sets: a training set of 2,908 videos, a validation set of 276 videos, and a test set of 818 videos. The test set is further divided into three subsets: 292 Seen videos, which includes object categories present in the training set; 269 Unseen videos, which consists of object categories not seen during training but belonging to super-categories represented in the training data, assessing the model's ability to generalize to new objects; and 257 Null videos, where the text expression refers to an object that either does not exist or is not visible in the given context, with empty true masks, requiring the model to avoid segmenting any object." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.274, + 0.601 + ], + "angle": 0, + "content": "4.2. Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Following [36], we used the Jaccard index \\(\\mathcal{I}\\) and F-score \\(\\mathcal{F}\\) as performance metrics. For the Null test set, we employed:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.645, + 0.483, + 0.686 + ], + "angle": 0, + "content": "\\[\nS = \\sqrt {\\frac {\\text {p r e d i c t e d m a s k a r e a}}{\\text {b a c k g r o u n d a r e a}}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.483, + 0.737 + ], + "angle": 0, + "content": "a metric to assess expression guidance in the model. A higher \\(S\\) value indicates a larger mask relative to the background, suggesting less accurate expression guidance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.306, + 0.759 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We initialized TSAM for the Ref-AVS task using the pretrained ViT-B variant of SAM (\\(N = 12\\), \\(d_{emb} = 256\\)). The temporal branch consists of \\(M = 4\\) blocks, and the selected audio queries is \\(k = 5\\). Training was performed with PyTorch on eight AMD GPUs in a distributed setup. The AdamW optimizer was used, with an initial learning rate of \\(1 \\cdot 10^{-4}\\) and a batch size of 1. The model was trained for 15 epochs with periodic evaluations, and the best-performing model on the validation subset was selected for testing." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.906, + 0.206 + ], + "angle": 0, + "content": "
MethodTaskVisual BackboneSeenUnseenNull S(↓)
J(%)FJ(%)F
AVSBench [39]AVS†PVT-v223.200.51132.360.5470.208
AVSegFormer [8]AVS†PVT-v233.470.47036.050.5010.171
GAVS [35]AVS†SAM28.930.49829.820.4970.190
SAMA [24]AVS*SAM39.220.56247.500.5660.130
ReferFormer [37]Ref-VOS‡V-Swin31.310.50130.400.4880.176
R2-VOS [18]Ref-VOS‡V-Swin25.010.41027.930.4980.183
EEMC [36]Ref-AVSM2F34.200.51349.540.6480.007
TSAM (Ours)Ref-AVSSAM43.430.56854.580.6640.017
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.216, + 0.907, + 0.272 + ], + "angle": 0, + "content": "Table 1. Performance comparison on the Ref-AVS dataset. \\( {}^{ \\dagger } \\) denotes text integration in AVS from [36], \\( {}^{ \\ddagger } \\) indicates audio integration in Ref-VOS from [36], and * marks our implementation with text added to AVS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.301, + 0.734, + 0.318 + ], + "angle": 0, + "content": "4.4. Results and comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.46 + ], + "angle": 0, + "content": "We compared the proposed TSAM method with the EEMC method [36] and existing methods from the AVS and Ref-VOS domains, including AVSBench [39], AVSeg-Former [8], GAVS [35], SAMA [24], ReferFormer [37], and R2-VOS [18], using the Ref-AVS dataset [36]. For consistency and fairness, we reported the published results for all methods in [36] and re-evaluated the SAMA method by integrating text with both audio and visual inputs to align with our multimodal setup." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.906, + 0.567 + ], + "angle": 0, + "content": "The results, presented in Tab. 1, demonstrate that even when augmented with text (for AVS methods) or audio (for Ref-VOS methods), these methods do not yield satisfactory performance in the Ref-AVS task. Interestingly, methods that utilize prior segmentation visual backbones, such as SAM and M2F (Mask2Former), tend to outperform those based on visual backbones like PVT-v2 and V-Swin." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Although GAVS and SAMA are SAM-based methods, they perform worse than EEMC. SAMA fails to leverage SAM's flexible, promptable nature, integrating additional modality cues superficially by layering them on SAM's frozen image encoder. This underutilizes the potential of multimodal cues, limiting SAMA's segmentation capabilities. GAVS, on the other hand, prompts SAM with multimodal cues but does not address the necessary alignment and interplay among these modalities, leading to a lack of cohesive multimodal fusion. These results confirm that simple modality fusion alone is insufficient for handling the complexity of multimodal integration in the Ref-AVS task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In contrast, our TSAM leverages language-guided audio queries as sparse prompts, integrates multimodal dense prompts via cross-attention, and incorporates a temporal modeling branch and adapter module to enhance SAM's ability to perform accurate, contextually guided segmentation of dynamic objects in complex audio-visual scenes. Compared to the best-performing method, EEMC, on the Ref-AVS dataset, our TSAM achieves superior results: a \\(\\mathcal{I} = 9.23\\%\\) and \\(\\mathcal{F} = 5.5\\%\\) improvements on the Seen test set, and a \\(\\mathcal{J} = 5.04\\%\\) and \\(\\mathcal{F} = 1.6\\%\\) improvements" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23952" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "on the Unseen test set. Although TSAM falls slightly behind EEMC on the Null test set by \\( S = 1\\% \\), likely due to SAM's limitation of always producing a segmentation mask even when no target object is present, TSAM maintains a strong advantage on Unseen test set. This is primarily due to TSAM's ability to generalize to unseen object categories, preserving SAM's pretrained knowledge while ensuring robust segmentation in novel audio-visual scenes, thanks to the added temporal modeling branch." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.24, + 0.27, + 0.256 + ], + "angle": 0, + "content": "4.5. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.485, + 0.55 + ], + "angle": 0, + "content": "We further evaluated our TSAM method qualitatively on both the Seen and Unseen test sets, as shown in Fig. 2 and Fig. 3, respectively. Fig. 2 demonstrates TSAM's consistent ability to produce high-quality masks for targeted objects, even when they are inaudible, as seen in the middle example. This highlights TSAM's proficiency in processing complex multimodal instructions through its prompting modules, enabling accurate segmentation based on textual cues. On the Unseen test set, Fig. 3, TSAM showcases a remarkable capacity to generalize to novel objects, effectively aligning audio-visual and textual inputs in novel scenes to accurately segment the targeted object of interest. This underscores TSAM's ability to generalize to new contexts, interpreting diverse multimodal instructions and segmenting unseen objects. TSAM's temporal branch enhances understanding of novel objects in dynamic scenes, while SAM's pre-trained knowledge is preserved for core segmentation capabilities. Additional qualitative results and comparisons are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.561, + 0.242, + 0.577 + ], + "angle": 0, + "content": "4.6. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.734 + ], + "angle": 0, + "content": "We analyzed the impact of omitting key components of TSAM, including the temporal branch (TB), temporal modality fusion layer (TMFL), dense prompting module (DPM), sparse prompting module (SPM), cached memory (CM), and adapter module (AM), as shown in Tab. 2. \\(\\mathrm{CM(a + v)}\\) denotes cached memory for both audio and visual modalities, while \\(\\mathrm{CM(v)}\\) denotes cached memory on the visual modality only, see Fig. 1. Additionally, Tab. 2 presents the effect of the intersection over union \\((\\mathcal{L}_{\\mathrm{IoU}})\\) loss, defined in Eq. (5), on TSAM's training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in Tab. 2, we observe a significant performance degradation in setting \\(②\\) when the temporal branch (TB) is removed, with a drop of \\(7.24\\% \\mathrm{pt}\\) in \\(\\mathcal{I}\\) (from \\(49.01\\%\\) to \\(41.77\\%\\)) and \\(3.4\\% \\mathrm{pt}\\) in \\(\\mathcal{F}\\) (from 0.616 to 0.582). This performance loss is significantly higher than that in other settings, emphasizing the critical role of the temporal modeling branch in TSAM's ability to generalize across video frames and maintain coherence in dynamic scenes, which is essential for object segmentation in complex Ref-AVS tasks. Settings \\(④\\) and \\(⑤\\) demonstrate the importance of prompting SAM with multimodal cues through SPM and" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.098, + 0.895, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.212, + 0.875, + 0.221 + ], + "angle": 0, + "content": "Expression: The object making a sound by being played by the woman." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.226, + 0.895, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.339, + 0.828, + 0.348 + ], + "angle": 0, + "content": "Expression: The object behind the sounding women." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.353, + 0.895, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.465, + 0.863, + 0.475 + ], + "angle": 0, + "content": "Expression: The sounding object on the legs of the man on the left." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.496, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 2. Segmentation masks for referred objects in the Ref-AVS dataset (Seen test set). Top row: ground truth with target object in green. Bottom row: our results with target object in violet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.658 + ], + "angle": 0, + "content": "DPM. SPM and DPM play complementary roles in TSAM, and their joint contribution is pivotal for its overall performance. Omitting these prompts in setting ⑥ causes a notable decrease in segmentation performance, reinforcing our claim that the rich semantic cues from multimodal prompts are essential for improving segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.907, + 0.69 + ], + "angle": 0, + "content": "Settings ⑦ and ⑧ evaluate the effect of cached memory on both audio and visual (CM(a+v)) and visual-only" + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.714, + 0.899, + 0.844 + ], + "angle": 0, + "content": "
SettingSeenUnseenMix(S+U)Null S(↓)
J(%)FJ(%)FJ(%)F
① TSAM43.430.56854.580.66449.010.6160.017
② - TB33.050.50750.480.65741.770.5820.505
③ - TMFL40.350.57945.540.62742.950.6030.018
④ - DPM42.720.58049.100.64745.910.6140.018
⑤ - SPM43.040.58049.750.65246.400.6160.018
⑥ - SPM+DPM42.600.60240.580.60441.590.6030.018
⑦ - CM(a+v)42.070.54449.110.65945.590.6020.018
⑧ - CM(v)42.750.54951.180.66046.970.6050.018
⑨ - AM43.130.60040.790.61741.960.6090.017
⑩ - LloU38.290.56442.150.63140.220.5980.008
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.855, + 0.907, + 0.896 + ], + "angle": 0, + "content": "Table 2. Ablation study on TSAM components and \\(\\mathcal{L}_{IotU}\\) loss. 'Mix' refers to the average performance on the Seen (S) and Unseen (U) test sets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23953" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.099, + 0.18, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.1, + 0.252, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.1, + 0.323, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.1, + 0.395, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.1, + 0.468, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.157, + 0.18, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.157, + 0.251, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.157, + 0.323, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.157, + 0.395, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.157, + 0.468, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.211, + 0.414, + 0.221 + ], + "angle": 0, + "content": "Expression: The object emits a sound following the man." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.226, + 0.179, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.226, + 0.251, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.226, + 0.323, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.226, + 0.395, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.226, + 0.468, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.283, + 0.179, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.283, + 0.251, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.283, + 0.323, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.283, + 0.395, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.283, + 0.468, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.339, + 0.353, + 0.348 + ], + "angle": 0, + "content": "Expression: The truck moving." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.354, + 0.179, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.354, + 0.252, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.354, + 0.323, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.354, + 0.395, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.354, + 0.468, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.411, + 0.179, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.411, + 0.251, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.411, + 0.323, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.411, + 0.395, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.411, + 0.468, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.466, + 0.389, + 0.475 + ], + "angle": 0, + "content": "Expression: The tuba being played by a man." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.496, + 0.483, + 0.54 + ], + "angle": 0, + "content": "Figure 3. Segmentation masks for referred objects in the Ref-AVS dataset (Unseen test set). Top row: ground truth with target object in green. Bottom row: our results with target object in violet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.672 + ], + "angle": 0, + "content": "\\((\\mathrm{CM}(\\mathrm{v}))\\) modalities, demonstrate a performance degradation in TSAM when cached memory is not applied across modalities \\((\\mathrm{CM}(\\mathrm{a} + \\mathrm{v}))\\) and when only the visual modality's memory \\((\\mathrm{CM}(\\mathrm{v}))\\) is omitted. The latter specifically emphasizes the importance of maintaining shared memory across video frames for effective temporal alignment and multimodal integration, as evidenced by the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.809 + ], + "angle": 0, + "content": "The results from setting ⑨ highlight the adapter module's importance. Its omission causes a significant performance drop, underscoring its crucial role in TSAM. We also observe that the components of TSAM, particularly in settings ③ - ⑨, have a substantial impact on performance, especially on the Unseen test set. This set challenges the model to segment novel objects, demonstrating TSAM's ability to generalize beyond seen categories, an essential feature for robust performance in open-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Finally, setting ⑩ highlights the distinct contribution of \\(\\mathcal{L}_{\\mathrm{IoU}}\\) to the overall loss function \\(\\mathcal{L}_{\\mathrm{total}}\\). Notably, the model performs best on both Seen and Unseen test sets when \\(\\mathcal{L}_{\\mathrm{IoU}}\\) is included in the total loss. Omitting \\(\\mathcal{L}_{\\mathrm{IoU}}\\) results in the highest performance on the Null test set. Further ablation studies of \\(\\lambda\\) (Eq. (5)) are in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.089, + 0.9, + 0.193 + ], + "angle": 0, + "content": "
TSAMSeenUnseenMix(S+U)Null S(↓)
J(%)FJ(%)FJ(%)F
k=343.580.57950.430.65547.010.6170.018
k=543.430.56854.580.66449.010.6160.017
k=743.240.57346.950.63045.100.6010.018
M=242.040.56653.510.65147.760.6090.023
M=443.430.56854.580.66449.010.6160.017
M=643.270.57549.860.64046.570.6080.020
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.203, + 0.905, + 0.232 + ], + "angle": 0, + "content": "Table 3. Effect of audio queries \\( k \\) and temporal branch depth \\( M \\) on TSAM's performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.363 + ], + "angle": 0, + "content": "We further evaluated the impact of varying the number of audio queries \\( k \\), defined in Eq. (4), and the temporal branch depth \\( M \\), see Sec. 3.3, on TSAM's performance, as shown in Tab. 3. Among the configurations \\( k = 3 \\), \\( k = 5 \\), and \\( k = 7 \\), we observe that \\( k = 5 \\) performs best. This balance optimizes the relevance of audio cues without overloading the model with irrelevant queries to SAM's mask decoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.905, + 0.456 + ], + "angle": 0, + "content": "Similarly, we find that \\( M = 4 \\) yields optimal results for the temporal branch. A shallower setup (\\( M = 2 \\)) likely lacks the temporal depth required for effective generalization and multimodal integration, while a deeper branch (\\( M = 6 \\)) may introduce excessive complexity, impairing the model's generalization to unseen objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.64, + 0.484 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.794 + ], + "angle": 0, + "content": "This paper presented TSAM, a novel extension of the Segment Anything Model (SAM) tailored for the referring audio-visual segmentation (Ref-AVS) task. TSAM enhances SAM's segmentation capabilities by effectively integrating multimodal cues through data-driven dense and sparse prompting modules while modeling temporal dependencies via a dedicated temporal branch. This approach preserves SAM's foundational knowledge while improving performance, particularly with novel objects. Extensive experiments on the Ref-AVS dataset demonstrate TSAM's strong generalization ability, showcasing its effectiveness in segmenting objects from textual cues and its robustness in unseen scenarios. These results highlight TSAM's potential to advance cross-modal understanding in multimodal tasks. Limitations and future work. Although TSAM outperforms existing Ref-AVS methods, a direct comparison with the recent SAM2 [31] remains to be explored. SAM2, designed for video segmentation with a memory module for temporal visual representations, lacks multimodal support. Future work will adapt SAM2 for the Ref-AVS task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the Research Council of Finland through the Finnish Center for Artificial Intelligence flagship (grant 345552) and the USSEE project (grant 345791). We also acknowledge CSC-IT Center for Science, Finland, for awarding access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23954" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In CVPR, pages 4985-4995, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the Kinetics dataset. In CVPR, pages 6299-6308, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.483, + 0.258 + ], + "angle": 0, + "content": "[3] Yuanhong Chen, Yuyuan Liu, Hu Wang, Fengbei Liu, Chong Wang, Helen Frazer, and Gustavo Carneiro. Unraveling instance associations: A closer look for audio-visual segmentation. In CVPR, pages 26497–26507, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[5] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[6] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[8] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In AAAI, pages 12155–12163, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In CVPR, pages 5958-5966, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[10] Yuxin Guo, Siyang Sun, Shuailei Ma, Kecheng Zheng, Xiaoyi Bao, Shijie Ma, Wei Zou, and Yun Zheng. Cross-MAE: Cross-modality masked autoencoders for region-aware audio-visual pre-training. In CVPR, pages 26721–26731, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[11] Dawei Hao, Yuxin Mao, Bowen He, Xiaodong Han, Yuchao Dai, and Yiran Zhong. Improving audio-visual segmentation with bidirectional generation. In AAAI, pages 2067–2075, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[12] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. CNN architectures for large-scale audio classification. In ICASSP, pages 131-135, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[13] Xiao Hu, Basavaraj Hampiholi, Heiko Neumann, and Jochen Lang. Temporal context enhanced referring video object segmentation. In WACV, pages 5574-5583, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Shaofei Huang, Han Li, Yuqing Wang, Hongji Zhu, Jiao Dai, Jizhong Han, Wenge Rong, and Si Liu. Discovering sounding objects by audio queries for audio visual segmentation. In IJCAI, pages 875-883, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[16] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, pages 19108-19118, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[17] Kexin Li, Zongxin Yang, Lei Chen, Yi Yang, and Jun Xiao. Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. In ACM MM, pages 1485–1494, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[18] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. In ICCV, pages 22236-22245, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[19] Yonglin Li, Jing Zhang, Xiao Teng, Long Lan, and Xinwang Liu. RefSAM: Efficiently adapting segmenting anything model for referring video object segmentation. arXiv preprint arxiv.org/abs/2307.00997, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[20] Yuhang Ling, Yuxi Li, Zhenye Gan, Jiangning Zhang, Ming-min Chi, and Yabiao Wang. TransAVS: End-to-end audiovisual segmentation with transformer. In ICASSP, pages 7845-7849, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[21] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In ACM MM, pages 7590–7598, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[22] Chen Liu, Peike Li, Hu Zhang, Lincheng Li, Zi Huang, Dadong Wang, and Xin Yu. BAVS: Bootstrapping audiovisual segmentation by integrating foundation knowledge. IEEE TMM, 26:10015-10028, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[23] Jinxiang Liu, Chen Ju, Chaofan Ma, Yanfeng Wang, Yu Wang, and Ya Zhang. Audio-aware query-enhanced transformer for audio-visual segmentation. arXiv preprint arXiv:2307.13236, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[24] Jinxiang Liu, Yu Wang, Chen Ju, Chaofan Ma, Ya Zhang, and Weidi Xie. Annotation-free audio-visual segmentation. In WACV, pages 5604–5614, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[25] Y Liu, M Ott, N Goyal, J Du, M Joshi, D Chen, O Levy, M Lewis, L Zettlemoyer, and V Stoyanov. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[26] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In CVPR, pages 3202-3211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[27] Yuxin Mao, Jing Zhang, Mochu Xiang, Yunqiu Lv, Yi ran Zhong, and Yuchao Dai. Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[28] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In ICCV, pages 954–965, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Shentong Mo and Yapeng Tian. AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836, 2023. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23955" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[30] Federico Perazzi, Anna Khoreva, Rodrigo Benenson, Bernt Schiele, and Alexander Sorkine-Hornung. Learning video object segmentation from static images. In CVPR, pages 2663-2672, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[31] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. In ICLR, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[32] Seonguk Seo, Joon-Young Lee, and Bohyung Han. URVOS: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[33] Juhyeong Seon, Woobin Im, Sebin Lee, Jumin Lee, and Sung-Eui Yoon. Extending segment anything model into auditory and temporal dimensions for audio-visual segmentation. In ICIP, pages 2480–2486, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[34] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. PVT v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[35] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. In AAAI, pages 5669–5677, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[36] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-AVS: Refer and segment objects in audio-visual scenes. In ECCV, pages 196–213, 2024. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[37] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[38] Jiarui Yu, Haoran Li, Yanbin Hao, Jinmeng Wu, Tong Xu, Shuo Wang, and Xiangnan He. How can contrastive pretraining benefit audio-visual segmentation? A study from supervised and zero-shot perspectives. In BMVC, pages 367–374, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[39] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In ECCV, pages 386–403, 2022. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23956" + } + ] +] \ No newline at end of file diff --git a/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/8535e16b-e9a4-4280-8556-88908a875920_origin.pdf b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/8535e16b-e9a4-4280-8556-88908a875920_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b38e52b2ad39c144ceec247eea4e2ca1f967e29d --- /dev/null +++ b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/8535e16b-e9a4-4280-8556-88908a875920_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4f2f3ff0397dfe3f90d6ab50f474d8a71768f74d3fc3d8a81e0fedc0995c0fa +size 1335451 diff --git a/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/full.md b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e60f41318895acdcdabb81569f7d614a1c534787 --- /dev/null +++ b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/full.md @@ -0,0 +1,336 @@ +# TSAM: Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation + +Abduljalil Radman & Jorma Laaksonen +Department of Computer Science, Aalto University, Finland +{abduljalil.saif, jorma.laaksonen}@aalto.fi + +# Abstract + +Referring audio-visual segmentation (Ref-AVS) aims to segment objects within audio-visual scenes using multimodal cues embedded in text expressions. While the Segment Anything Model (SAM) has revolutionized visual segmentation, its applicability to Ref-AVS, where multimodal cues act as novel prompts, remains unexplored. SAM's limitation to single-frame segmentation also hinders its ability to capture essential temporal context needed for multi-frame audio-visual segmentation. To address this gap, we propose TSAM, a novel extension of SAM designed to leverage multimodal cues for precise segmentation in dynamic audio-visual scenes. TSAM enhances SAM's image encoder with a temporal modeling branch, enabling spatio-temporal learning and deep multimodal fusion across video frames, while retaining SAM's pre-trained knowledge. Additionally, TSAM replaces SAM's user-interactive prompting mechanism with sparse and dense data-driven prompts, enabling more effective integration of audio-visual inputs and reference text expressions. Extensive experiments on the Ref-AVS dataset demonstrate TSAM's superiority over state-of-the-art methods. The results illustrate its effectiveness in segmenting objects in dynamic audio-visual scenes using text-based multimodal cues and its strong generalization to unseen objects. Project webpage: https://abdurad.github.io/TSAM/. + +# 1. Introduction + +In real-world scenarios, people often rely on multimodal information, such as audio-visual and visual-text, to gather spatiotemporal and semantic insights for distinguishing objects of interest [10, 39]. This multimodal approach is equally valuable for machines, enabling applications like stage cameras to track a performing musician or movie cameras to focus on a superhero confronting a yelling villain. Relying on a single sensory input, such as vision, is often insufficient for accurate object segmentation in complex, + +dynamic audio-visual scenes [16, 36]. In the literature, researchers have addressed this need for multimodal integration in object segmentation through three main approaches, each focused on a specific modality cue. + +First, video object segmentation (VOS) [6, 30] relies on a mask annotation in the initial frame to guide segmentation of objects across subsequent frames. While VOS allows for flexible segmentation, it is heavily dependent on labor-intensive frame-level annotations, making it challenging to scale. Second, referring video object segmentation (Ref-VOS) [5, 9, 13] replaces the need for mask annotations with text expressions to identify targeted objects. This approach introduces a more intuitive, text-based method for segmentation, but struggles with tracking and segmenting objects in more dynamic audio-visual environments. Third, audio-visual segmentation (AVS) [21, 24, 33, 35, 39] leverages audio cues to segment sounding objects in video, effectively handling complex audio-visual scenarios, but remaining limited in scope as it does not generalize to inaudible objects. Each of these approaches, VOS, Ref-VOS, and AVS, contributes valuable segmentation capabilities within its modality focus, but remains challenged in more naturalistic audio-visual scenarios [36]. + +Recently, Wang et al. [36] introduced referring audio-visual segmentation (Ref-AVS), a task that segments objects in video by interpreting text expressions enriched with both audio and visual cues. They also presented Expression Enhancing with Multimodal Cues (EEMC), a simple method that integrates reference text expressions with dynamic audio-visual cues to segment targeted objects in scenes. By unifying these multimodal cues, EEMC enables robust segmentation in dynamic scenes, overcoming limitations of prior approaches that focused on narrower combinations of modalities. + +We propose that enhancing Ref-AVS can be achieved by leveraging prior knowledge from large-scale visual foundation models, such as the Segment Anything Model (SAM) [15], for segmentation. SAM is a flexible segmentation model that utilizes an image encoder to generate high-dimensional image embeddings. It incorpo + +![](images/bfe22308fe89c6ec3d6e561d0015f84a726a9d3007c683c041f3cb692977268a.jpg) +Figure 1. Overview of TSAM. Left: TSAM integrates multimodal cues from diverse inputs, leveraging the prompting modules to guide segmentation. Right: The temporal branch fuses multimodal cues and learns temporal context for segmentation in audio-visual scenes. The adapter input, $\dot{F}_a$ , is the output from the cached memory. + +![](images/538387209496b06f899b05fbfb6b2020e5b163402fdf06ef08238f0f192a83cc.jpg) + +rates user-provided visual prompts, which can be sparse (e.g., points, boxes) or dense (e.g., masks), processed by a dedicated prompt encoder. These embeddings are then passed through a lightweight, transformer-based mask decoder to generate accurate object masks based on the provided prompts. Although SAM has been applied to tasks such as AVS [24, 29, 33, 35], and Ref-VOS [19], its potential for segmenting objects in the visual domain using multimodal cues embedded in text expressions remains unexplored. This makes its application to the Ref-AVS task both novel and challenging as it requires the integration of diverse multimodal cues to guide segmentation, overcoming the limitations of SAM's traditional reliance on explicit user-interactive prompts. + +The main challenge in adapting SAM for the Ref-AVS + +task, which requires nuanced multimodal understanding without direct user guidance, arises from SAM's lack of temporal awareness in its image encoder, limiting its ability to capture complex interactions across audio, visual, and text modalities. Additionally, SAM relies on user-provided visual prompts for explicit spatial information of targeted objects, while its capacity to process text and audio modalities remains limited. + +This paper introduces TSAM (Fig. 1), a novel SAM-based architecture for the Ref-AVS task, designed to enhance segmentation by leveraging multimodal cues from audio, visual, and text inputs. TSAM incorporates an auxiliary temporal modeling branch in parallel with SAM's image encoder, preserving SAM's zero-shot capabilities while adapting it for temporally varying video content. More- + +over, TSAM uses text cues to guide the selection of relevant audio cues, which are then used as sparse prompts, enriched with key visual cues, to query SAM's mask decoder. Since multimodal prompts lack explicit spatial information for the target objects, TSAM incorporates a lightweight adapter module into the temporal branch to enhance SAM's mask decoder prompts, facilitating deeper multimodal interactions during training. Through a cross-attention mechanism, TSAM synthesizes data-driven multimodal reference cues as dense prompts for SAM. Extensive experiments show that TSAM outperforms existing state-of-the-art methods in segmenting target objects in the Ref-AVS task. The main contributions of this work are: + +1. We propose TSAM, an end-to-end architecture built on the Segment Anything Model (SAM), which redefines traditional visual prompts as data-driven multimodal prompts for object segmentation in Ref-AVS tasks. +2. We introduce a temporal modeling branch that enhances SAM's image encoder with temporal awareness, enabling the capture of intricate spatio-temporal interactions across multimodal cues. +3. Through comprehensive experiments, we demonstrate that the proposed TSAM achieves superior performance compared to state-of-the-art methods in Ref-AVS. Furthermore, extensive ablation studies provide a detailed evaluation of TSAM's architecture, emphasizing the pivotal role of multimodal cue integration and temporal modeling in enhancing segmentation performance. + +# 2. Related work + +# 2.1. Referring video object segmentation (Ref-VOS) + +Referring video object segmentation (Ref-VOS) involves segmenting target objects within video frames based on textual descriptions. Gavrilyuk et al. [9] introduced an encoder-decoder model based on I3D [2], that generates dynamic convolutional filters from text and applies them to visual representations at different resolutions to produce segmentation masks for target objects. URVOS [32] enhanced object tracking by performing segmentation on a key frame and then propagating the mask across the video. This approach combines cross-modal attention to prevent drift and memory attention to maintain temporal consistency across frames. MTTR [1] formulated Ref-VOS as a sequence prediction task, utilizing multimodal transformers to process video and text jointly. ReferFormer [37] used Video-Swin (V-Swin) [26] as the visual backbone, linking textual expressions directly to relevant frame regions. In this approach, language-conditioned object queries serve as inputs, helping the model identify the referred objects and apply dynamic kernels as convolution filters to generate segmentation masks. Although Ref-VOS enhances user accessibility by replacing mask annotations with natural lan + +guage expressions, it still struggles to accurately identify objects in dynamic audio-visual scenes [36]. + +R2-VOS [18] extended the Ref-VOS task to handle unpaired visual-text inputs, jointly modeling video segmentation and text reconstruction. By introducing a structural text-to-text cycle constraint, R2-VOS improves multimodal alignment across diverse text inputs, overcoming the limitations of prior point-wise methods. RefSAM [19] utilized text prompts with SAM [15], using a lightweight cross-modal MLP to project the referring expression's text embedding into prompt embeddings. RefSAM enables multimodal reference segmentation by integrating textual and visual modalities, but it does not fully exploit the simultaneous interaction of multimodal cues or support the temporal dynamics required for video segmentation tasks. + +# 2.2. Audio-visual segmentation (AVS) + +Audio-visual segmentation (AVS), introduced by Zhou et al. [39], is a challenging task that seeks to achieve pixel-level segmentation of sound-emitting objects in video, guided by the audio modality. AVSBench [39] used a temporal pixel-wise audio-visual interaction module to inject audio semantics as guidance for the visual segmentation process. Similarly, AVSegFormer [8], a transformer-based model built on the PVT-v2 backbone [34], used cross-attention to fuse spatial-temporal visual cues with audio cues. By decoding these combined audio-visual cues and employing audio queries, the model enhanced its ability to focus on sounding objects within the audio-visual scene. Other transformer-based models proposed in [14, 17, 23] integrated audio and visual cues across both temporal and spatial dimensions for the AVS task. Alternative approaches [20, 22] disentangled the audio input into audio queries, which interacts with the visual input to decode segmentation masks. These approaches emphasize implicit modality fusion strategies, training models to fit discrete samples in the dataset. However, their performance often falls short due to limitations in dataset size and diversity. + +Alternatively, a Explicit Conditional Multimodal Variational Auto-Encoder (ECMVAE) [28] was proposed to extract semantically correlated cues from audio-visual inputs for segmenting audible objects in video frames. In another approach, the AVS Bidirectional Generation (AVSBG) framework [11] was introduced, correlating visual cues of objects with their corresponding audio cues through a visual-to-audio projection and implicit volumetric motion estimation. Recent advancements in AVS [3, 27, 38] also integrated audio-visual contrastive learning methods to select informative samples and improve training efficiency and model performance. However, these methods often fail to fully leverage available information due to limited prior segmentation knowledge in their backbone architectures, hin + +dering their ability to effectively mine informative samples. + +On the other hand, SAM-based methods showed improvements in the AVS task. AV-SAM [29] was an early attempt to apply SAM to AVS, utilizing pixel-wise audio-visual fusion to prompt SAM for sounding object segmentation. In [24], two SAM-based approaches for AVS were introduced: Audio-Prompted SAM (AP-SAM) and SAM Adapter (SAMA-AVS). AP-SAM encoded audio as a prompt for SAM's mask decoder, while SAMA-AVS incorporated an adapter module to inject the audio information during the training process, but did not fully leverage SAM's promptable nature. Recently, a Generalizable Audio-Visual Segmentation (GAVS) model [35] was introduced, targeting zero-shot and few-shot scenarios. GAVS proposed a semantic-aware audio prompt to bridge the semantic gap between visual and auditory modalities through contrastive learning. All of these methods take a single image and its corresponding audio as input, fusing multimodal inputs in the spatial domain, which limits their ability to capture temporal relationships frame by frame. + +In a straightforward approach, Seon et al. [33] extended SAM for AVS by integrating a Spatio-Temporal, Bidirectional Audio-Visual Attention (ST-BAVA) module between SAM's image encoder and mask decoder. While this approach aims to model temporal cues for audible object segmentation, it oversimplifies the interaction between audio and visual modalities, treating them in a relatively shallow manner. The model relies on basic feature fusion instead of a more meaningful interaction, limiting its ability to fully leverage the potential of multimodal cues for AVS tasks. + +# 2.3. Referring audio-visual segmentation (Ref-AVS) + +Referring audio-visual segmentation (Ref-AVS) is a novel pixel-level segmentation task in which the model predicts, for each pixel, whether it corresponds to an object specified by multimodal cues in a text expression. Wang et al. [36] introduced the Ref-AVS task and developed a dataset with mask annotations for segmenting objects in video frames guided by multimodal expressions, creating the first benchmark for supervised Ref-AVS. They also proposed the Expression Enhancing with Multimodal Cues (EEMC) model, based on the Mask2Former (M2F) [4] segmentation decoder, which adaptively combines audio, visual, and text cues to capture spatio-temporal correspondences between target objects and text expressions. + +Unlike existing SAM-based models for tasks like AVS and Ref-VOS [19, 24, 29, 33, 35], which either fail to address deep temporal multimodal alignment [19, 33, 35] or fail to leverage SAM's flexible prompt-based nature [24], our proposed TSAM, shown in Fig. 1, uniquely supports temporal modeling for SAM. It addresses the challenge of segmenting temporally varying objects and enabling a more cohesive and robust integration of multimodal cues for ob + +ject segmentation in dynamic audio-visual scenes. Additionally, TSAM introduces a dual-pathway framework with sparse and dense prompting modules that replace SAM's visual prompts with data-driven multimodal prompts, fostering richer and more dynamic prompts. + +# 3. TSAM for Ref-AVS + +In this section, we introduce the proposed TSAM method (Fig. 1) for the Ref-AVS task. We begin by presenting the problem formulation, followed by an overview of the foundational components of SAM. Next, we describe the temporal modeling branch and the integration of multimodal prompts in TSAM. Finally, we present the training objective that guides TSAM's training. + +# 3.1. Problem formulation + +In the Ref-AVS task, the input includes a sequence of video frames, corresponding audio, and a reference text expression specifying the objects to be segmented in the video. The goal is to segment target objects within each video frame, using multimodal cues embedded in the text expression. The segmentation output is represented as binary masks, with each mask highlighting the target objects within an individual video frame. + +The visual input includes $n$ frames sampled at 1-second intervals, where each frame $\in \mathbb{R}^{3 \times 1024 \times 1024}$ represents RGB values at a resolution of $1024 \times 1024$ pixels. The audio input, temporally aligned with the video frames, is encoded offline using VGGish [12], resulting in representations $F_{a} \in \mathbb{R}^{n \times 128}$ . Similarly, the text expression is preextracted using RoBERTa [25] to generate a text representation $F_{t} \in \mathbb{R}^{l \times 768}$ , where $l$ denotes the length of the expression. + +To enable effective multimodal prompting within SAM, TSAM first projects both audio and text representations into the SAM visual embedding space with dimensionality $d_{emb}$ . This results in audio cues $F_{a} \in \mathbb{R}^{n \times 1 \times d_{emb}}$ and text cues $F_{t} \in \mathbb{R}^{n \times l \times d_{emb}}$ , with the text cues $F_{t}$ temporally expanded along $n$ . These projected cues are then fused across the temporal dimension in TSAM's temporal modality fusion layer (TMFL), as shown in Fig. 1, producing expression-related multimodal cues as follows: + +$$ +F _ {a} ^ {\prime}, F _ {t} ^ {\prime} = \operatorname {S A} (\operatorname {C o n c a t} (F _ {a}, F _ {t})), \tag {1} +$$ + +where $\operatorname{Concat}(\cdot)$ denotes concatenation, and $\mathrm{SA}(\cdot)$ applies temporal self-attention to enable modality fusion over the temporal domain. Following the approach in [36], TSAM integrates a basic cached memory (CM), which stores $\hat{F}_a$ , an accumulated summary of $F_a'$ over time. This memory captures the mean modality cues across the temporal domain and maintains a running summary of the evolving audio context. Concurrently, the attention-enhanced text cues + +$F_{t}^{\prime}$ are added to the text cues $F_{t}$ to create the unified multimodal text cues $\hat{F}_{t} \in \mathbb{R}^{n \times l \times d_{emb}}$ , resulting in a cohesive and enriched representation of audio and text cues. + +# 3.2. SAM as a foundational baseline + +SAM [15] is the first large-scale pretrained foundational model for guided image segmentation. Trained on a massive dataset consisting of 1 billion masks across 11 million images, SAM has demonstrated remarkable effectiveness and versatility in image segmentation tasks. The core components of SAM are outlined here, with further details available in [15]. SAM's architecture is composed of three key modules: + +(a) Image encoder. A pre-trained Vision Transformer (ViT) [7], minimally adapted to process high-resolution images, extracts spatial embeddings. +(b) Prompt encoder. SAM's prompt encoder handles various user prompts, including sparse (points, boxes, text) and dense (masks) formats, providing regional information about the target objects for segmentation. +(c) Mask decoder. The mask decoder consists of two layers. First, the dense prompt embedding is integrated with the image embedding. Then, cross-attention is applied between the image embedding and the sparse prompt embeddings to update both representations. In each attention layer, the original prompt token is re-integrated into both the query and key tokens. Finally, the dot product of the upsampled image embedding and the MLP-mapped sparse prompt embeddings produces the final segmentation mask. + +# 3.3. Temporal modeling branch + +The SAM image encoder does not account for the temporal dynamics inherent in video frames, limiting its capacity to handle the complex interplay between audio and visual cues in tasks such as object segmentation. To address this, the proposed TSAM enhances SAM's architecture by incorporating a dedicated temporal modeling branch at the final $M$ blocks of the SAM image encoder, as shown in Fig. 1. While the SAM image encoder remains frozen during training, the newly introduced temporal blocks are the only trainable. These temporal blocks mirror the structure of the corresponding SAM encoder blocks, initialized with the same pre-trained weights. + +Each video frame is processed independently through the SAM image encoder. At each temporal modeling block, TSAM integrates a straightforward cached memory and adapter module, which tracks and captures the evolving temporal and multimodal cues. Specifically, the $m$ -th $(m = 1,2,\dots,M)$ temporal block receives input $x_{m}$ , combining the output $y_{m-1}$ of the previous temporal block, the output $y_{(N-M)+(m-1)}^{\mathrm{SAM}}$ of the previous SAM encoder block, and the textual enriched audio cues $\hat{F}_{a}$ via the adapter module: + +$$ +x _ {m} = y _ {m - 1} + \mathrm {C M} _ {m} \left(y _ {(N - M) + (m - 1)} ^ {\mathrm {S A M}}\right) + \mathrm {A M} _ {m} \left(\hat {F} _ {a}\right), \tag {2} +$$ + +where $\mathrm{CM}_m(\cdot)$ denotes the cached memory applied to the SAM block output, and $\mathrm{AM}_m(\cdot)$ represents the adapter module applied to the audio cues $F_{a}$ . The adapter module, a bottleneck block shown in Fig. 1, integrates audio cues that have been refined by text cues, facilitating deeper interaction among the audio, visual, and text modalities within the temporal modeling branch. + +Finally, the overall video cues, $F_{v} = \{f_{v}^{i}\}_{i = 1}^{n}$ , where each $f_{v}^{i}$ represents a cue for frame $i$ , are derived by integrating the output from the temporal modeling branch $Y_{M}$ with the output from SAM's image encoder $Y_{N}^{SAM}$ : + +$$ +F _ {v} = Y _ {M} + Y _ {N} ^ {S A M}. \tag {3} +$$ + +Each video frame visual cues $f_{v}^{i}\in \mathbb{R}^{d_{emb}\times 64\times 64}$ , with $64\times 64$ representing the spatial dimensions. + +# 3.4. Prompting SAM with multimodal cues + +TSAM introduces a dual-pathway framework with dedicated sparse and dense prompting modules (SPM and DPM, respectively) to enhance SAM's segmentation capabilities. The SPM learns global context by leveraging sparse queries derived from multimodal cues, guiding TSAM toward high-level object localization. Meanwhile, the DPM operates at a finer granularity, providing dense prompts aligned with the frame's spatial features, highlighting its reliance on the outputs of SAM's image encoder, as shown in Fig. 1. + +Sparse prompting module (SPM). To effectively guide segmentation within TSAM using the text expression, SPM first employs a language-guided query selection mechanism to select the $k$ most relevant audio cues, associated with the input text expression. The top $k$ query indices, denoted as $A_{k}$ , are selected as: + +$$ +A _ {k} = \operatorname {T o p} _ {k} \left(\max \left(\hat {F} _ {a} \hat {F} _ {t} ^ {\top}\right)\right), \tag {4} +$$ + +where $\mathrm{Top}_k$ picks the top $k$ indices, maxf computes the maximum along the cues' dimension, and $(\cdot)^{\top}$ indicates matrix transposition. The selected audio cues are then fused with the visual cues through a cross-attention mechanism and transformed using a feed forward network-mapped sparse prompt embedding. This process generates contextually enriched prompts for SAM's mask decoder. + +Dense prompting module (DPM). The DPM in TSAM generates spatially comprehensive embeddings by integrating visual, audio, and textual cues through cross-attention mechanisms, as illustrated in Fig. 1. First, an audio cross-attention layer aligns the audio cues with the visual cues, facilitating effective fusion of both modalities. Next, a text cross-attention layer incorporates the text cues, further refining the combined audio-visual cues to focus on the target object for segmentation. Finally, a feed forward network refines the multimodal embeddings, preparing them for input into SAM's mask decoder. + +# 3.5. Training objective + +We employ binary cross-entropy $\mathcal{L}_{\mathrm{BCE}}$ and intersection over union $\mathcal{L}_{\mathrm{IoU}}$ losses to measure the discrepancy between TSAM's predicted mask and the ground truth mask during training. The total loss $\mathcal{L}_{\mathrm{total}}$ for model optimization is defined as a weighted sum of $\mathcal{L}_{\mathrm{BCE}}$ and $\mathcal{L}_{\mathrm{IoU}}$ losses, as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {B C E}} + \lambda \cdot \mathcal {L} _ {\mathrm {I o U}}, \tag {5} +$$ + +where $\lambda$ is set to 1.0 to balance the two loss components. + +# 4. Experiments + +# 4.1. Dataset + +We evaluated TSAM using the Ref-AVS dataset [36], which contains a diverse collection of 20,000 text expressions and pixel-level annotations across 4,000 10-second videos. The dataset features a wide range of audible objects, including 20 categories of musical instruments, 8 animals, 15 machines, and 5 humans, as well as a smaller set of 3 categories of static, inaudible objects. + +The dataset is divided into three sets: a training set of 2,908 videos, a validation set of 276 videos, and a test set of 818 videos. The test set is further divided into three subsets: 292 Seen videos, which includes object categories present in the training set; 269 Unseen videos, which consists of object categories not seen during training but belonging to super-categories represented in the training data, assessing the model's ability to generalize to new objects; and 257 Null videos, where the text expression refers to an object that either does not exist or is not visible in the given context, with empty true masks, requiring the model to avoid segmenting any object. + +# 4.2. Evaluation metrics + +Following [36], we used the Jaccard index $\mathcal{I}$ and F-score $\mathcal{F}$ as performance metrics. For the Null test set, we employed: + +$$ +S = \sqrt {\frac {\text {p r e d i c t e d m a s k a r e a}}{\text {b a c k g r o u n d a r e a}}}, \tag {6} +$$ + +a metric to assess expression guidance in the model. A higher $S$ value indicates a larger mask relative to the background, suggesting less accurate expression guidance. + +# 4.3. Implementation details + +We initialized TSAM for the Ref-AVS task using the pretrained ViT-B variant of SAM ( $N = 12$ , $d_{emb} = 256$ ). The temporal branch consists of $M = 4$ blocks, and the selected audio queries is $k = 5$ . Training was performed with PyTorch on eight AMD GPUs in a distributed setup. The AdamW optimizer was used, with an initial learning rate of $1 \cdot 10^{-4}$ and a batch size of 1. The model was trained for 15 epochs with periodic evaluations, and the best-performing model on the validation subset was selected for testing. + +
MethodTaskVisual BackboneSeenUnseenNull S(↓)
J(%)FJ(%)F
AVSBench [39]AVS†PVT-v223.200.51132.360.5470.208
AVSegFormer [8]AVS†PVT-v233.470.47036.050.5010.171
GAVS [35]AVS†SAM28.930.49829.820.4970.190
SAMA [24]AVS*SAM39.220.56247.500.5660.130
ReferFormer [37]Ref-VOS‡V-Swin31.310.50130.400.4880.176
R2-VOS [18]Ref-VOS‡V-Swin25.010.41027.930.4980.183
EEMC [36]Ref-AVSM2F34.200.51349.540.6480.007
TSAM (Ours)Ref-AVSSAM43.430.56854.580.6640.017
+ +Table 1. Performance comparison on the Ref-AVS dataset. ${}^{ \dagger }$ denotes text integration in AVS from [36], ${}^{ \ddagger }$ indicates audio integration in Ref-VOS from [36], and * marks our implementation with text added to AVS. + +# 4.4. Results and comparison + +We compared the proposed TSAM method with the EEMC method [36] and existing methods from the AVS and Ref-VOS domains, including AVSBench [39], AVSeg-Former [8], GAVS [35], SAMA [24], ReferFormer [37], and R2-VOS [18], using the Ref-AVS dataset [36]. For consistency and fairness, we reported the published results for all methods in [36] and re-evaluated the SAMA method by integrating text with both audio and visual inputs to align with our multimodal setup. + +The results, presented in Tab. 1, demonstrate that even when augmented with text (for AVS methods) or audio (for Ref-VOS methods), these methods do not yield satisfactory performance in the Ref-AVS task. Interestingly, methods that utilize prior segmentation visual backbones, such as SAM and M2F (Mask2Former), tend to outperform those based on visual backbones like PVT-v2 and V-Swin. + +Although GAVS and SAMA are SAM-based methods, they perform worse than EEMC. SAMA fails to leverage SAM's flexible, promptable nature, integrating additional modality cues superficially by layering them on SAM's frozen image encoder. This underutilizes the potential of multimodal cues, limiting SAMA's segmentation capabilities. GAVS, on the other hand, prompts SAM with multimodal cues but does not address the necessary alignment and interplay among these modalities, leading to a lack of cohesive multimodal fusion. These results confirm that simple modality fusion alone is insufficient for handling the complexity of multimodal integration in the Ref-AVS task. + +In contrast, our TSAM leverages language-guided audio queries as sparse prompts, integrates multimodal dense prompts via cross-attention, and incorporates a temporal modeling branch and adapter module to enhance SAM's ability to perform accurate, contextually guided segmentation of dynamic objects in complex audio-visual scenes. Compared to the best-performing method, EEMC, on the Ref-AVS dataset, our TSAM achieves superior results: a $\mathcal{I} = 9.23\%$ and $\mathcal{F} = 5.5\%$ improvements on the Seen test set, and a $\mathcal{J} = 5.04\%$ and $\mathcal{F} = 1.6\%$ improvements + +on the Unseen test set. Although TSAM falls slightly behind EEMC on the Null test set by $S = 1\%$ , likely due to SAM's limitation of always producing a segmentation mask even when no target object is present, TSAM maintains a strong advantage on Unseen test set. This is primarily due to TSAM's ability to generalize to unseen object categories, preserving SAM's pretrained knowledge while ensuring robust segmentation in novel audio-visual scenes, thanks to the added temporal modeling branch. + +# 4.5. Qualitative results + +We further evaluated our TSAM method qualitatively on both the Seen and Unseen test sets, as shown in Fig. 2 and Fig. 3, respectively. Fig. 2 demonstrates TSAM's consistent ability to produce high-quality masks for targeted objects, even when they are inaudible, as seen in the middle example. This highlights TSAM's proficiency in processing complex multimodal instructions through its prompting modules, enabling accurate segmentation based on textual cues. On the Unseen test set, Fig. 3, TSAM showcases a remarkable capacity to generalize to novel objects, effectively aligning audio-visual and textual inputs in novel scenes to accurately segment the targeted object of interest. This underscores TSAM's ability to generalize to new contexts, interpreting diverse multimodal instructions and segmenting unseen objects. TSAM's temporal branch enhances understanding of novel objects in dynamic scenes, while SAM's pre-trained knowledge is preserved for core segmentation capabilities. Additional qualitative results and comparisons are provided in the supplementary materials. + +# 4.6. Ablation study + +We analyzed the impact of omitting key components of TSAM, including the temporal branch (TB), temporal modality fusion layer (TMFL), dense prompting module (DPM), sparse prompting module (SPM), cached memory (CM), and adapter module (AM), as shown in Tab. 2. $\mathrm{CM(a + v)}$ denotes cached memory for both audio and visual modalities, while $\mathrm{CM(v)}$ denotes cached memory on the visual modality only, see Fig. 1. Additionally, Tab. 2 presents the effect of the intersection over union $(\mathcal{L}_{\mathrm{IoU}})$ loss, defined in Eq. (5), on TSAM's training. + +As shown in Tab. 2, we observe a significant performance degradation in setting $②$ when the temporal branch (TB) is removed, with a drop of $7.24\% \mathrm{pt}$ in $\mathcal{I}$ (from $49.01\%$ to $41.77\%$ ) and $3.4\% \mathrm{pt}$ in $\mathcal{F}$ (from 0.616 to 0.582). This performance loss is significantly higher than that in other settings, emphasizing the critical role of the temporal modeling branch in TSAM's ability to generalize across video frames and maintain coherence in dynamic scenes, which is essential for object segmentation in complex Ref-AVS tasks. Settings $④$ and $⑤$ demonstrate the importance of prompting SAM with multimodal cues through SPM and + +![](images/fa954cd612c52b82f28f8af8f5ed46650d45481d70b18fa1b33bc8b64547e88b.jpg) +Expression: The object making a sound by being played by the woman. + +![](images/5c65021cc090792bd5d70e6eceda7a8215590934a7f7891e1e0c68f310a1780a.jpg) +Expression: The object behind the sounding women. + +![](images/f621832f3b34724703f53198d475c12598545d490505d2b379d746a86485808a.jpg) +Expression: The sounding object on the legs of the man on the left. +Figure 2. Segmentation masks for referred objects in the Ref-AVS dataset (Seen test set). Top row: ground truth with target object in green. Bottom row: our results with target object in violet. + +DPM. SPM and DPM play complementary roles in TSAM, and their joint contribution is pivotal for its overall performance. Omitting these prompts in setting ⑥ causes a notable decrease in segmentation performance, reinforcing our claim that the rich semantic cues from multimodal prompts are essential for improving segmentation performance. + +Settings ⑦ and ⑧ evaluate the effect of cached memory on both audio and visual (CM(a+v)) and visual-only + +
SettingSeenUnseenMix(S+U)Null S(↓)
J(%)FJ(%)FJ(%)F
① TSAM43.430.56854.580.66449.010.6160.017
② - TB33.050.50750.480.65741.770.5820.505
③ - TMFL40.350.57945.540.62742.950.6030.018
④ - DPM42.720.58049.100.64745.910.6140.018
⑤ - SPM43.040.58049.750.65246.400.6160.018
⑥ - SPM+DPM42.600.60240.580.60441.590.6030.018
⑦ - CM(a+v)42.070.54449.110.65945.590.6020.018
⑧ - CM(v)42.750.54951.180.66046.970.6050.018
⑨ - AM43.130.60040.790.61741.960.6090.017
⑩ - LloU38.290.56442.150.63140.220.5980.008
+ +Table 2. Ablation study on TSAM components and $\mathcal{L}_{IotU}$ loss. 'Mix' refers to the average performance on the Seen (S) and Unseen (U) test sets. + +![](images/8c7b03e93af18185a0fe6dd4255c0ac7d28bf6f51c7d5b7a95d071667d9020cd.jpg) + +![](images/c32c6da14c5f666204c3bce2e306ac792fd2a4447757fc814cc378d195f212f7.jpg) + +![](images/ab6589e26733cbfeaae20a626a2a74d187d883e0c5e54b190b5f7c2c1d943635.jpg) + +![](images/33b9aada7beee949c280a65aecc91433c8f2ab8ae0a6fb02b82281f41c6a5656.jpg) + +![](images/ec4babeb3115fcbfabfa9d094834d8d459fbae6a8e5e9e78adc07846b62c487f.jpg) + +![](images/5f7aa8fb6caedfc59e0cbe6ae5563e445b31bf15ba86bcc908104324d29bb6ea.jpg) +Expression: The object emits a sound following the man. + +![](images/c4b75ef3ccb70fdbee0ab4814ac3c9cd31a33ff2b33258d6748beddc9fc6aca9.jpg) + +![](images/d731b9618978db5dd580d34b032fc26a57a82a18d332e9f4f426b35697da581f.jpg) + +![](images/f50b84d51ba7aa0130b6eaa201fef6876282ec07ba89cc474f2474f97694d335.jpg) + +![](images/d55b504018cf187ef7fdc6f3c982cedd632cff44f7196b73e28b714465ec3fa9.jpg) + +![](images/9316f29dbac323452a8d1bb6bd3e6d1813fb47e3af7606f8a9e827774a24b331.jpg) + +![](images/48043a4ff06e3f3f07bf9d2953df3d509b9912f82e5c4731f33253cdf01c8edd.jpg) + +![](images/30bc5a0f57d31fb99726c3c45b75b9c35e5b7fcb6d7657b9170df88376eec400.jpg) + +![](images/8e4d33d4f566bf13e25228874f46641604ed59d23ecf6ebfd02dc0e056c45bf7.jpg) + +![](images/af52a2d194c050723d4f408cbbf3acf1456a634d45a71fb9ce65a55daa9ed37c.jpg) + +![](images/7e3e38f98b194bd14b2595f889eb55708dda7d1f7a32f28211186040b99e8083.jpg) + +![](images/dd6e51ee8b5e70966ad066467c91dd8fbe1b9aa429e31a2245850149d4ea7ad9.jpg) +Expression: The truck moving. + +![](images/bf93de1c2b3311e4075fddaf854f346fd63dbcde11c8e943aaaad9db0ec79b37.jpg) + +![](images/db8d19084ca67bf8a8dbc5d6e43b92d69f777e97105ea05911c363cdeccbc8d5.jpg) + +![](images/f2f00566eae7bd3f77a6c3f89ac282c3fe454808c6ce58e561669b2c036bc63a.jpg) + +![](images/502711f764fe10c94a05052468a5334985de8dbd10a6d8209d779ac0abfd6216.jpg) + +![](images/3742fa8f2bacc2e2eb029dd50cc1b09e7cc40c35ae6260f3e1a9ea04d2d9b7e6.jpg) + +![](images/6d1f5e30f311036dcb309521287ff1fe269f079980caac3b4ca969dca6feb071.jpg) + +![](images/93d6713fe4177877f852d01d0599507ada46ffd6198384b8e79f4b569f716521.jpg) + +![](images/fb02d1af3940e3208c0dae0699e861bd468293a71fd77741c54be93c8cc18c05.jpg) + +![](images/61a7bfed7db84710848a2012b9f6bbdec506c9bd816a560b5abd206bc8093198.jpg) +Expression: The tuba being played by a man. +Figure 3. Segmentation masks for referred objects in the Ref-AVS dataset (Unseen test set). Top row: ground truth with target object in green. Bottom row: our results with target object in violet. + +![](images/98fe3b13283bed1cb72c8b625aa0d65e5a9a4b48ff2a9c54e58b3486d52fd2fd.jpg) + +![](images/0034bca49334dcff8e003e01bb79eeb4561a08c890f881f67da96789d64d4d3b.jpg) + +![](images/44433495be3ce2ebba79580956a614f87c40deea78409b0b20cfbb67204582bd.jpg) + +![](images/0bb20ecef774b8cc3526fa1af37b624dd489ba20885f941ee2a27506ef3fed29.jpg) + +$(\mathrm{CM}(\mathrm{v}))$ modalities, demonstrate a performance degradation in TSAM when cached memory is not applied across modalities $(\mathrm{CM}(\mathrm{a} + \mathrm{v}))$ and when only the visual modality's memory $(\mathrm{CM}(\mathrm{v}))$ is omitted. The latter specifically emphasizes the importance of maintaining shared memory across video frames for effective temporal alignment and multimodal integration, as evidenced by the results. + +The results from setting ⑨ highlight the adapter module's importance. Its omission causes a significant performance drop, underscoring its crucial role in TSAM. We also observe that the components of TSAM, particularly in settings ③ - ⑨, have a substantial impact on performance, especially on the Unseen test set. This set challenges the model to segment novel objects, demonstrating TSAM's ability to generalize beyond seen categories, an essential feature for robust performance in open-world scenarios. + +Finally, setting ⑩ highlights the distinct contribution of $\mathcal{L}_{\mathrm{IoU}}$ to the overall loss function $\mathcal{L}_{\mathrm{total}}$ . Notably, the model performs best on both Seen and Unseen test sets when $\mathcal{L}_{\mathrm{IoU}}$ is included in the total loss. Omitting $\mathcal{L}_{\mathrm{IoU}}$ results in the highest performance on the Null test set. Further ablation studies of $\lambda$ (Eq. (5)) are in the supplementary materials. + +
TSAMSeenUnseenMix(S+U)Null S(↓)
J(%)FJ(%)FJ(%)F
k=343.580.57950.430.65547.010.6170.018
k=543.430.56854.580.66449.010.6160.017
k=743.240.57346.950.63045.100.6010.018
M=242.040.56653.510.65147.760.6090.023
M=443.430.56854.580.66449.010.6160.017
M=643.270.57549.860.64046.570.6080.020
+ +Table 3. Effect of audio queries $k$ and temporal branch depth $M$ on TSAM's performance. + +We further evaluated the impact of varying the number of audio queries $k$ , defined in Eq. (4), and the temporal branch depth $M$ , see Sec. 3.3, on TSAM's performance, as shown in Tab. 3. Among the configurations $k = 3$ , $k = 5$ , and $k = 7$ , we observe that $k = 5$ performs best. This balance optimizes the relevance of audio cues without overloading the model with irrelevant queries to SAM's mask decoder. + +Similarly, we find that $M = 4$ yields optimal results for the temporal branch. A shallower setup ( $M = 2$ ) likely lacks the temporal depth required for effective generalization and multimodal integration, while a deeper branch ( $M = 6$ ) may introduce excessive complexity, impairing the model's generalization to unseen objects. + +# 5. Conclusions + +This paper presented TSAM, a novel extension of the Segment Anything Model (SAM) tailored for the referring audio-visual segmentation (Ref-AVS) task. TSAM enhances SAM's segmentation capabilities by effectively integrating multimodal cues through data-driven dense and sparse prompting modules while modeling temporal dependencies via a dedicated temporal branch. This approach preserves SAM's foundational knowledge while improving performance, particularly with novel objects. Extensive experiments on the Ref-AVS dataset demonstrate TSAM's strong generalization ability, showcasing its effectiveness in segmenting objects from textual cues and its robustness in unseen scenarios. These results highlight TSAM's potential to advance cross-modal understanding in multimodal tasks. Limitations and future work. Although TSAM outperforms existing Ref-AVS methods, a direct comparison with the recent SAM2 [31] remains to be explored. SAM2, designed for video segmentation with a memory module for temporal visual representations, lacks multimodal support. Future work will adapt SAM2 for the Ref-AVS task. + +Acknowledgments. This work was supported by the Research Council of Finland through the Finnish Center for Artificial Intelligence flagship (grant 345552) and the USSEE project (grant 345791). We also acknowledge CSC-IT Center for Science, Finland, for awarding access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking. + +# References + +[1] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In CVPR, pages 4985-4995, 2022. 3 +[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the Kinetics dataset. In CVPR, pages 6299-6308, 2017. 3 +[3] Yuanhong Chen, Yuyuan Liu, Hu Wang, Fengbei Liu, Chong Wang, Helen Frazer, and Gustavo Carneiro. Unraveling instance associations: A closer look for audio-visual segmentation. In CVPR, pages 26497–26507, 2024. 3 +[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 4 +[5] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 1 +[6] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 1 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5 +[8] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In AAAI, pages 12155–12163, 2024. 3, 6 +[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In CVPR, pages 5958-5966, 2018. 1, 3 +[10] Yuxin Guo, Siyang Sun, Shuailei Ma, Kecheng Zheng, Xiaoyi Bao, Shijie Ma, Wei Zou, and Yun Zheng. Cross-MAE: Cross-modality masked autoencoders for region-aware audio-visual pre-training. In CVPR, pages 26721–26731, 2024. 1 +[11] Dawei Hao, Yuxin Mao, Bowen He, Xiaodong Han, Yuchao Dai, and Yiran Zhong. Improving audio-visual segmentation with bidirectional generation. In AAAI, pages 2067–2075, 2024. 3 +[12] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. CNN architectures for large-scale audio classification. In ICASSP, pages 131-135, 2017. 4 +[13] Xiao Hu, Basavaraj Hampiholi, Heiko Neumann, and Jochen Lang. Temporal context enhanced referring video object segmentation. In WACV, pages 5574-5583, 2024. 1 +[14] Shaofei Huang, Han Li, Yuqing Wang, Hongji Zhu, Jiao Dai, Jizhong Han, Wenge Rong, and Si Liu. Discovering sounding objects by audio queries for audio visual segmentation. In IJCAI, pages 875-883, 2023. 3 + +[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 1, 3, 5 +[16] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, pages 19108-19118, 2022. 1 +[17] Kexin Li, Zongxin Yang, Lei Chen, Yi Yang, and Jun Xiao. Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. In ACM MM, pages 1485–1494, 2023. 3 +[18] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. In ICCV, pages 22236-22245, 2023. 3, 6 +[19] Yonglin Li, Jing Zhang, Xiao Teng, Long Lan, and Xinwang Liu. RefSAM: Efficiently adapting segmenting anything model for referring video object segmentation. arXiv preprint arxiv.org/abs/2307.00997, 2024. 2, 3, 4 +[20] Yuhang Ling, Yuxi Li, Zhenye Gan, Jiangning Zhang, Ming-min Chi, and Yabiao Wang. TransAVS: End-to-end audiovisual segmentation with transformer. In ICASSP, pages 7845-7849, 2024. 3 +[21] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In ACM MM, pages 7590–7598, 2023. 1 +[22] Chen Liu, Peike Li, Hu Zhang, Lincheng Li, Zi Huang, Dadong Wang, and Xin Yu. BAVS: Bootstrapping audiovisual segmentation by integrating foundation knowledge. IEEE TMM, 26:10015-10028, 2024. 3 +[23] Jinxiang Liu, Chen Ju, Chaofan Ma, Yanfeng Wang, Yu Wang, and Ya Zhang. Audio-aware query-enhanced transformer for audio-visual segmentation. arXiv preprint arXiv:2307.13236, 2023. 3 +[24] Jinxiang Liu, Yu Wang, Chen Ju, Chaofan Ma, Ya Zhang, and Weidi Xie. Annotation-free audio-visual segmentation. In WACV, pages 5604–5614, 2024. 1, 2, 4, 6 +[25] Y Liu, M Ott, N Goyal, J Du, M Joshi, D Chen, O Levy, M Lewis, L Zettlemoyer, and V Stoyanov. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4 +[26] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In CVPR, pages 3202-3211, 2022. 3 +[27] Yuxin Mao, Jing Zhang, Mochu Xiang, Yunqiu Lv, Yi ran Zhong, and Yuchao Dai. Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579, 2023. 3 +[28] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In ICCV, pages 954–965, 2023. 3 +[29] Shentong Mo and Yapeng Tian. AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836, 2023. 2, 4 + +[30] Federico Perazzi, Anna Khoreva, Rodrigo Benenson, Bernt Schiele, and Alexander Sorkine-Hornung. Learning video object segmentation from static images. In CVPR, pages 2663-2672, 2017. 1 +[31] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. In ICLR, 2025. 8 +[32] Seonguk Seo, Joon-Young Lee, and Bohyung Han. URVOS: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223, 2020. 3 +[33] Juhyeong Seon, Woobin Im, Sebin Lee, Jumin Lee, and Sung-Eui Yoon. Extending segment anything model into auditory and temporal dimensions for audio-visual segmentation. In ICIP, pages 2480–2486, 2024. 1, 2, 4 +[34] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. PVT v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3 +[35] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. In AAAI, pages 5669–5677, 2024. 1, 2, 4, 6 +[36] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-AVS: Refer and segment objects in audio-visual scenes. In ECCV, pages 196–213, 2024. 1, 3, 4, 6 +[37] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 3, 6 +[38] Jiarui Yu, Haoran Li, Yanbin Hao, Jinmeng Wu, Tong Xu, Shuo Wang, and Xiangnan He. How can contrastive pretraining benefit audio-visual segmentation? A study from supervised and zero-shot perspectives. In BMVC, pages 367–374, 2023. 3 +[39] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In ECCV, pages 386–403, 2022. 1, 3, 6 \ No newline at end of file diff --git a/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/images.zip b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6686efc2d74f6844ffb0cfdfbd4d76e019b68826 --- /dev/null +++ b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dadb5c5a65dcd3e849521433e36f520b6d6a103d3ecd01ea1a78553a77ee9e76 +size 500784 diff --git a/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/layout.json b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3c3f8db0f934b821eae4199e48ca631999b5a6f4 --- /dev/null +++ b/2025/TSAM_ Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation/layout.json @@ -0,0 +1,9065 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 102, + 539, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 102, + 539, + 138 + ], + "spans": [ + { + "bbox": [ + 71, + 102, + 539, + 138 + ], + "type": "text", + "content": "TSAM: Temporal SAM Augmented with Multimodal Prompts for Referring Audio-Visual Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 161, + 451, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 161, + 451, + 203 + ], + "spans": [ + { + "bbox": [ + 159, + 161, + 451, + 203 + ], + "type": "text", + "content": "Abduljalil Radman & Jorma Laaksonen \nDepartment of Computer Science, Aalto University, Finland \n{abduljalil.saif, jorma.laaksonen}@aalto.fi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 258, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 558 + ], + "type": "text", + "content": "Referring audio-visual segmentation (Ref-AVS) aims to segment objects within audio-visual scenes using multimodal cues embedded in text expressions. While the Segment Anything Model (SAM) has revolutionized visual segmentation, its applicability to Ref-AVS, where multimodal cues act as novel prompts, remains unexplored. SAM's limitation to single-frame segmentation also hinders its ability to capture essential temporal context needed for multi-frame audio-visual segmentation. To address this gap, we propose TSAM, a novel extension of SAM designed to leverage multimodal cues for precise segmentation in dynamic audio-visual scenes. TSAM enhances SAM's image encoder with a temporal modeling branch, enabling spatio-temporal learning and deep multimodal fusion across video frames, while retaining SAM's pre-trained knowledge. Additionally, TSAM replaces SAM's user-interactive prompting mechanism with sparse and dense data-driven prompts, enabling more effective integration of audio-visual inputs and reference text expressions. Extensive experiments on the Ref-AVS dataset demonstrate TSAM's superiority over state-of-the-art methods. The results illustrate its effectiveness in segmenting objects in dynamic audio-visual scenes using text-based multimodal cues and its strong generalization to unseen objects. Project webpage: https://abdurad.github.io/TSAM/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "In real-world scenarios, people often rely on multimodal information, such as audio-visual and visual-text, to gather spatiotemporal and semantic insights for distinguishing objects of interest [10, 39]. This multimodal approach is equally valuable for machines, enabling applications like stage cameras to track a performing musician or movie cameras to focus on a superhero confronting a yelling villain. Relying on a single sensory input, such as vision, is often insufficient for accurate object segmentation in complex," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 553, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 280 + ], + "type": "text", + "content": "dynamic audio-visual scenes [16, 36]. In the literature, researchers have addressed this need for multimodal integration in object segmentation through three main approaches, each focused on a specific modality cue." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 508 + ], + "type": "text", + "content": "First, video object segmentation (VOS) [6, 30] relies on a mask annotation in the initial frame to guide segmentation of objects across subsequent frames. While VOS allows for flexible segmentation, it is heavily dependent on labor-intensive frame-level annotations, making it challenging to scale. Second, referring video object segmentation (Ref-VOS) [5, 9, 13] replaces the need for mask annotations with text expressions to identify targeted objects. This approach introduces a more intuitive, text-based method for segmentation, but struggles with tracking and segmenting objects in more dynamic audio-visual environments. Third, audio-visual segmentation (AVS) [21, 24, 33, 35, 39] leverages audio cues to segment sounding objects in video, effectively handling complex audio-visual scenarios, but remaining limited in scope as it does not generalize to inaudible objects. Each of these approaches, VOS, Ref-VOS, and AVS, contributes valuable segmentation capabilities within its modality focus, but remains challenged in more naturalistic audio-visual scenarios [36]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 509, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 640 + ], + "type": "text", + "content": "Recently, Wang et al. [36] introduced referring audio-visual segmentation (Ref-AVS), a task that segments objects in video by interpreting text expressions enriched with both audio and visual cues. They also presented Expression Enhancing with Multimodal Cues (EEMC), a simple method that integrates reference text expressions with dynamic audio-visual cues to segment targeted objects in scenes. By unifying these multimodal cues, EEMC enables robust segmentation in dynamic scenes, overcoming limitations of prior approaches that focused on narrower combinations of modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "We propose that enhancing Ref-AVS can be achieved by leveraging prior knowledge from large-scale visual foundation models, such as the Segment Anything Model (SAM) [15], for segmentation. SAM is a flexible segmentation model that utilizes an image encoder to generate high-dimensional image embeddings. It incorpo" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23947" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 77, + 362, + 453 + ], + "blocks": [ + { + "bbox": [ + 65, + 77, + 362, + 453 + ], + "lines": [ + { + "bbox": [ + 65, + 77, + 362, + 453 + ], + "spans": [ + { + "bbox": [ + 65, + 77, + 362, + 453 + ], + "type": "image", + "image_path": "bfe22308fe89c6ec3d6e561d0015f84a726a9d3007c683c041f3cb692977268a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 467, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 54, + 467, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 467, + 555, + 502 + ], + "type": "text", + "content": "Figure 1. Overview of TSAM. Left: TSAM integrates multimodal cues from diverse inputs, leveraging the prompting modules to guide segmentation. Right: The temporal branch fuses multimodal cues and learns temporal context for segmentation in audio-visual scenes. The adapter input, " + }, + { + "bbox": [ + 54, + 467, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\dot{F}_a" + }, + { + "bbox": [ + 54, + 467, + 555, + 502 + ], + "type": "text", + "content": ", is the output from the cached memory." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 378, + 137, + 547, + 439 + ], + "blocks": [ + { + "bbox": [ + 378, + 137, + 547, + 439 + ], + "lines": [ + { + "bbox": [ + 378, + 137, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 378, + 137, + 547, + 439 + ], + "type": "image", + "image_path": "538387209496b06f899b05fbfb6b2020e5b163402fdf06ef08238f0f192a83cc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 522, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 295, + 689 + ], + "type": "text", + "content": "rates user-provided visual prompts, which can be sparse (e.g., points, boxes) or dense (e.g., masks), processed by a dedicated prompt encoder. These embeddings are then passed through a lightweight, transformer-based mask decoder to generate accurate object masks based on the provided prompts. Although SAM has been applied to tasks such as AVS [24, 29, 33, 35], and Ref-VOS [19], its potential for segmenting objects in the visual domain using multimodal cues embedded in text expressions remains unexplored. This makes its application to the Ref-AVS task both novel and challenging as it requires the integration of diverse multimodal cues to guide segmentation, overcoming the limitations of SAM's traditional reliance on explicit user-interactive prompts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 689, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 689, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 67, + 689, + 295, + 701 + ], + "type": "text", + "content": "The main challenge in adapting SAM for the Ref-AVS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 522, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 616 + ], + "type": "text", + "content": "task, which requires nuanced multimodal understanding without direct user guidance, arises from SAM's lack of temporal awareness in its image encoder, limiting its ability to capture complex interactions across audio, visual, and text modalities. Additionally, SAM relies on user-provided visual prompts for explicit spatial information of targeted objects, while its capacity to process text and audio modalities remains limited." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 617, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 701 + ], + "type": "text", + "content": "This paper introduces TSAM (Fig. 1), a novel SAM-based architecture for the Ref-AVS task, designed to enhance segmentation by leveraging multimodal cues from audio, visual, and text inputs. TSAM incorporates an auxiliary temporal modeling branch in parallel with SAM's image encoder, preserving SAM's zero-shot capabilities while adapting it for temporally varying video content. More-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23948" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "over, TSAM uses text cues to guide the selection of relevant audio cues, which are then used as sparse prompts, enriched with key visual cues, to query SAM's mask decoder. Since multimodal prompts lack explicit spatial information for the target objects, TSAM incorporates a lightweight adapter module into the temporal branch to enhance SAM's mask decoder prompts, facilitating deeper multimodal interactions during training. Through a cross-attention mechanism, TSAM synthesizes data-driven multimodal reference cues as dense prompts for SAM. Extensive experiments show that TSAM outperforms existing state-of-the-art methods in segmenting target objects in the Ref-AVS task. The main contributions of this work are:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 295, + 408 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 229, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 295, + 277 + ], + "type": "text", + "content": "1. We propose TSAM, an end-to-end architecture built on the Segment Anything Model (SAM), which redefines traditional visual prompts as data-driven multimodal prompts for object segmentation in Ref-AVS tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 323 + ], + "type": "text", + "content": "2. We introduce a temporal modeling branch that enhances SAM's image encoder with temporal awareness, enabling the capture of intricate spatio-temporal interactions across multimodal cues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 295, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 408 + ], + "type": "text", + "content": "3. Through comprehensive experiments, we demonstrate that the proposed TSAM achieves superior performance compared to state-of-the-art methods in Ref-AVS. Furthermore, extensive ablation studies provide a detailed evaluation of TSAM's architecture, emphasizing the pivotal role of multimodal cue integration and temporal modeling in enhancing segmentation performance." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 422, + 139, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 139, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 139, + 434 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 443, + 295, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 456 + ], + "type": "text", + "content": "2.1. Referring video object segmentation (Ref-VOS)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": "Referring video object segmentation (Ref-VOS) involves segmenting target objects within video frames based on textual descriptions. Gavrilyuk et al. [9] introduced an encoder-decoder model based on I3D [2], that generates dynamic convolutional filters from text and applies them to visual representations at different resolutions to produce segmentation masks for target objects. URVOS [32] enhanced object tracking by performing segmentation on a key frame and then propagating the mask across the video. This approach combines cross-modal attention to prevent drift and memory attention to maintain temporal consistency across frames. MTTR [1] formulated Ref-VOS as a sequence prediction task, utilizing multimodal transformers to process video and text jointly. ReferFormer [37] used Video-Swin (V-Swin) [26] as the visual backbone, linking textual expressions directly to relevant frame regions. In this approach, language-conditioned object queries serve as inputs, helping the model identify the referred objects and apply dynamic kernels as convolution filters to generate segmentation masks. Although Ref-VOS enhances user accessibility by replacing mask annotations with natural lan" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "guage expressions, it still struggles to accurately identify objects in dynamic audio-visual scenes [36]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 97, + 553, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 253 + ], + "type": "text", + "content": "R2-VOS [18] extended the Ref-VOS task to handle unpaired visual-text inputs, jointly modeling video segmentation and text reconstruction. By introducing a structural text-to-text cycle constraint, R2-VOS improves multimodal alignment across diverse text inputs, overcoming the limitations of prior point-wise methods. RefSAM [19] utilized text prompts with SAM [15], using a lightweight cross-modal MLP to project the referring expression's text embedding into prompt embeddings. RefSAM enables multimodal reference segmentation by integrating textual and visual modalities, but it does not fully exploit the simultaneous interaction of multimodal cues or support the temporal dynamics required for video segmentation tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 263, + 492, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 492, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 492, + 276 + ], + "type": "text", + "content": "2.2. Audio-visual segmentation (AVS)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 544 + ], + "type": "text", + "content": "Audio-visual segmentation (AVS), introduced by Zhou et al. [39], is a challenging task that seeks to achieve pixel-level segmentation of sound-emitting objects in video, guided by the audio modality. AVSBench [39] used a temporal pixel-wise audio-visual interaction module to inject audio semantics as guidance for the visual segmentation process. Similarly, AVSegFormer [8], a transformer-based model built on the PVT-v2 backbone [34], used cross-attention to fuse spatial-temporal visual cues with audio cues. By decoding these combined audio-visual cues and employing audio queries, the model enhanced its ability to focus on sounding objects within the audio-visual scene. Other transformer-based models proposed in [14, 17, 23] integrated audio and visual cues across both temporal and spatial dimensions for the AVS task. Alternative approaches [20, 22] disentangled the audio input into audio queries, which interacts with the visual input to decode segmentation masks. These approaches emphasize implicit modality fusion strategies, training models to fit discrete samples in the dataset. However, their performance often falls short due to limitations in dataset size and diversity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": "Alternatively, a Explicit Conditional Multimodal Variational Auto-Encoder (ECMVAE) [28] was proposed to extract semantically correlated cues from audio-visual inputs for segmenting audible objects in video frames. In another approach, the AVS Bidirectional Generation (AVSBG) framework [11] was introduced, correlating visual cues of objects with their corresponding audio cues through a visual-to-audio projection and implicit volumetric motion estimation. Recent advancements in AVS [3, 27, 38] also integrated audio-visual contrastive learning methods to select informative samples and improve training efficiency and model performance. However, these methods often fail to fully leverage available information due to limited prior segmentation knowledge in their backbone architectures, hin" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23949" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "type": "text", + "content": "dering their ability to effectively mine informative samples." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 295, + 312 + ], + "type": "text", + "content": "On the other hand, SAM-based methods showed improvements in the AVS task. AV-SAM [29] was an early attempt to apply SAM to AVS, utilizing pixel-wise audio-visual fusion to prompt SAM for sounding object segmentation. In [24], two SAM-based approaches for AVS were introduced: Audio-Prompted SAM (AP-SAM) and SAM Adapter (SAMA-AVS). AP-SAM encoded audio as a prompt for SAM's mask decoder, while SAMA-AVS incorporated an adapter module to inject the audio information during the training process, but did not fully leverage SAM's promptable nature. Recently, a Generalizable Audio-Visual Segmentation (GAVS) model [35] was introduced, targeting zero-shot and few-shot scenarios. GAVS proposed a semantic-aware audio prompt to bridge the semantic gap between visual and auditory modalities through contrastive learning. All of these methods take a single image and its corresponding audio as input, fusing multimodal inputs in the spatial domain, which limits their ability to capture temporal relationships frame by frame." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 313, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 433 + ], + "type": "text", + "content": "In a straightforward approach, Seon et al. [33] extended SAM for AVS by integrating a Spatio-Temporal, Bidirectional Audio-Visual Attention (ST-BAVA) module between SAM's image encoder and mask decoder. While this approach aims to model temporal cues for audible object segmentation, it oversimplifies the interaction between audio and visual modalities, treating them in a relatively shallow manner. The model relies on basic feature fusion instead of a more meaningful interaction, limiting its ability to fully leverage the potential of multimodal cues for AVS tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 443, + 295, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 455 + ], + "type": "text", + "content": "2.3. Referring audio-visual segmentation (Ref-AVS)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "type": "text", + "content": "Referring audio-visual segmentation (Ref-AVS) is a novel pixel-level segmentation task in which the model predicts, for each pixel, whether it corresponds to an object specified by multimodal cues in a text expression. Wang et al. [36] introduced the Ref-AVS task and developed a dataset with mask annotations for segmenting objects in video frames guided by multimodal expressions, creating the first benchmark for supervised Ref-AVS. They also proposed the Expression Enhancing with Multimodal Cues (EEMC) model, based on the Mask2Former (M2F) [4] segmentation decoder, which adaptively combines audio, visual, and text cues to capture spatio-temporal correspondences between target objects and text expressions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Unlike existing SAM-based models for tasks like AVS and Ref-VOS [19, 24, 29, 33, 35], which either fail to address deep temporal multimodal alignment [19, 33, 35] or fail to leverage SAM's flexible prompt-based nature [24], our proposed TSAM, shown in Fig. 1, uniquely supports temporal modeling for SAM. It addresses the challenge of segmenting temporally varying objects and enabling a more cohesive and robust integration of multimodal cues for ob" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "ject segmentation in dynamic audio-visual scenes. Additionally, TSAM introduces a dual-pathway framework with sparse and dense prompting modules that replace SAM's visual prompts with data-driven multimodal prompts, fostering richer and more dynamic prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 143, + 430, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 430, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 430, + 156 + ], + "type": "text", + "content": "3. TSAM for Ref-AVS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 163, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 247 + ], + "type": "text", + "content": "In this section, we introduce the proposed TSAM method (Fig. 1) for the Ref-AVS task. We begin by presenting the problem formulation, followed by an overview of the foundational components of SAM. Next, we describe the temporal modeling branch and the integration of multimodal prompts in TSAM. Finally, we present the training objective that guides TSAM's training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 256, + 435, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 435, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 435, + 268 + ], + "type": "text", + "content": "3.1. Problem formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 274, + 553, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 553, + 369 + ], + "type": "text", + "content": "In the Ref-AVS task, the input includes a sequence of video frames, corresponding audio, and a reference text expression specifying the objects to be segmented in the video. The goal is to segment target objects within each video frame, using multimodal cues embedded in the text expression. The segmentation output is represented as binary masks, with each mask highlighting the target objects within an individual video frame." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": "The visual input includes " + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": " frames sampled at 1-second intervals, where each frame " + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{3 \\times 1024 \\times 1024}" + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": " represents RGB values at a resolution of " + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": " pixels. The audio input, temporally aligned with the video frames, is encoded offline using VGGish [12], resulting in representations " + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "inline_equation", + "content": "F_{a} \\in \\mathbb{R}^{n \\times 128}" + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": ". Similarly, the text expression is preextracted using RoBERTa [25] to generate a text representation " + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "inline_equation", + "content": "F_{t} \\in \\mathbb{R}^{l \\times 768}" + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 369, + 554, + 476 + ], + "type": "text", + "content": " denotes the length of the expression." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "text", + "content": "To enable effective multimodal prompting within SAM, TSAM first projects both audio and text representations into the SAM visual embedding space with dimensionality " + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "inline_equation", + "content": "d_{emb}" + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "text", + "content": ". This results in audio cues " + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "inline_equation", + "content": "F_{a} \\in \\mathbb{R}^{n \\times 1 \\times d_{emb}}" + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "text", + "content": " and text cues " + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "inline_equation", + "content": "F_{t} \\in \\mathbb{R}^{n \\times l \\times d_{emb}}" + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "text", + "content": ", with the text cues " + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "text", + "content": " temporally expanded along " + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 477, + 553, + 585 + ], + "type": "text", + "content": ". These projected cues are then fused across the temporal dimension in TSAM's temporal modality fusion layer (TMFL), as shown in Fig. 1, producing expression-related multimodal cues as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 594, + 553, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 594, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 370, + 594, + 553, + 608 + ], + "type": "interline_equation", + "content": "F _ {a} ^ {\\prime}, F _ {t} ^ {\\prime} = \\operatorname {S A} (\\operatorname {C o n c a t} (F _ {a}, F _ {t})), \\tag {1}", + "image_path": "e5b960ce20f6dc3db4c5e1ccb56822b2737ff1b5a8b52b5b6019b3d2aa4099f8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\operatorname{Concat}(\\cdot)" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " denotes concatenation, and " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SA}(\\cdot)" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " applies temporal self-attention to enable modality fusion over the temporal domain. Following the approach in [36], TSAM integrates a basic cached memory (CM), which stores " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{F}_a" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ", an accumulated summary of " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_a'" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " over time. This memory captures the mean modality cues across the temporal domain and maintains a running summary of the evolving audio context. Concurrently, the attention-enhanced text cues" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23950" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "F_{t}^{\\prime}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " are added to the text cues " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " to create the unified multimodal text cues " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\hat{F}_{t} \\in \\mathbb{R}^{n \\times l \\times d_{emb}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ", resulting in a cohesive and enriched representation of audio and text cues." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 114, + 226, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 226, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 226, + 125 + ], + "type": "text", + "content": "3.2. SAM as a foundational baseline" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 131, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 295, + 227 + ], + "type": "text", + "content": "SAM [15] is the first large-scale pretrained foundational model for guided image segmentation. Trained on a massive dataset consisting of 1 billion masks across 11 million images, SAM has demonstrated remarkable effectiveness and versatility in image segmentation tasks. The core components of SAM are outlined here, with further details available in [15]. SAM's architecture is composed of three key modules:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 227, + 296, + 418 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 227, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 295, + 262 + ], + "type": "text", + "content": "(a) Image encoder. A pre-trained Vision Transformer (ViT) [7], minimally adapted to process high-resolution images, extracts spatial embeddings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 262, + 295, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 310 + ], + "type": "text", + "content": "(b) Prompt encoder. SAM's prompt encoder handles various user prompts, including sparse (points, boxes, text) and dense (masks) formats, providing regional information about the target objects for segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 311, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 418 + ], + "type": "text", + "content": "(c) Mask decoder. The mask decoder consists of two layers. First, the dense prompt embedding is integrated with the image embedding. Then, cross-attention is applied between the image embedding and the sparse prompt embeddings to update both representations. In each attention layer, the original prompt token is re-integrated into both the query and key tokens. Finally, the dot product of the upsampled image embedding and the MLP-mapped sparse prompt embeddings produces the final segmentation mask." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 424, + 206, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 206, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 206, + 437 + ], + "type": "text", + "content": "3.3. Temporal modeling branch" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 441, + 295, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 584 + ], + "type": "text", + "content": "The SAM image encoder does not account for the temporal dynamics inherent in video frames, limiting its capacity to handle the complex interplay between audio and visual cues in tasks such as object segmentation. To address this, the proposed TSAM enhances SAM's architecture by incorporating a dedicated temporal modeling branch at the final " + }, + { + "bbox": [ + 55, + 441, + 295, + 584 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 441, + 295, + 584 + ], + "type": "text", + "content": " blocks of the SAM image encoder, as shown in Fig. 1. While the SAM image encoder remains frozen during training, the newly introduced temporal blocks are the only trainable. These temporal blocks mirror the structure of the corresponding SAM encoder blocks, initialized with the same pre-trained weights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "text", + "content": "Each video frame is processed independently through the SAM image encoder. At each temporal modeling block, TSAM integrates a straightforward cached memory and adapter module, which tracks and captures the evolving temporal and multimodal cues. Specifically, the " + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "inline_equation", + "content": "(m = 1,2,\\dots,M)" + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "text", + "content": " temporal block receives input " + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "text", + "content": ", combining the output " + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "inline_equation", + "content": "y_{m-1}" + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "text", + "content": " of the previous temporal block, the output " + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "inline_equation", + "content": "y_{(N-M)+(m-1)}^{\\mathrm{SAM}}" + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "text", + "content": " of the previous SAM encoder block, and the textual enriched audio cues " + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\hat{F}_{a}" + }, + { + "bbox": [ + 55, + 585, + 295, + 696 + ], + "type": "text", + "content": " via the adapter module:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 65, + 700, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 700, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 65, + 700, + 295, + 716 + ], + "type": "interline_equation", + "content": "x _ {m} = y _ {m - 1} + \\mathrm {C M} _ {m} \\left(y _ {(N - M) + (m - 1)} ^ {\\mathrm {S A M}}\\right) + \\mathrm {A M} _ {m} \\left(\\hat {F} _ {a}\\right), \\tag {2}", + "image_path": "f6bb8eaa6c00497e29dd88d5dd22798d43d2098b1ba0208ddb4ee6a5f0e1ffc6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{CM}_m(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " denotes the cached memory applied to the SAM block output, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{AM}_m(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " represents the adapter module applied to the audio cues " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ". The adapter module, a bottleneck block shown in Fig. 1, integrates audio cues that have been refined by text cues, facilitating deeper interaction among the audio, visual, and text modalities within the temporal modeling branch." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "text", + "content": "Finally, the overall video cues, " + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "inline_equation", + "content": "F_{v} = \\{f_{v}^{i}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "inline_equation", + "content": "f_{v}^{i}" + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "text", + "content": " represents a cue for frame " + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "text", + "content": ", are derived by integrating the output from the temporal modeling branch " + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "inline_equation", + "content": "Y_{M}" + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "text", + "content": " with the output from SAM's image encoder " + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "inline_equation", + "content": "Y_{N}^{SAM}" + }, + { + "bbox": [ + 313, + 156, + 553, + 205 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 392, + 212, + 553, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 212, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 392, + 212, + 553, + 228 + ], + "type": "interline_equation", + "content": "F _ {v} = Y _ {M} + Y _ {N} ^ {S A M}. \\tag {3}", + "image_path": "dd0a8e865a8bec227078886dbb7bd217770b08b04ce4d78cb58e209e557a51f4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 236, + 553, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 261 + ], + "type": "text", + "content": "Each video frame visual cues " + }, + { + "bbox": [ + 313, + 236, + 553, + 261 + ], + "type": "inline_equation", + "content": "f_{v}^{i}\\in \\mathbb{R}^{d_{emb}\\times 64\\times 64}" + }, + { + "bbox": [ + 313, + 236, + 553, + 261 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 236, + 553, + 261 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 313, + 236, + 553, + 261 + ], + "type": "text", + "content": " representing the spatial dimensions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 268, + 517, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 517, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 517, + 281 + ], + "type": "text", + "content": "3.4. Prompting SAM with multimodal cues" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 285, + 554, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 393 + ], + "type": "text", + "content": "TSAM introduces a dual-pathway framework with dedicated sparse and dense prompting modules (SPM and DPM, respectively) to enhance SAM's segmentation capabilities. The SPM learns global context by leveraging sparse queries derived from multimodal cues, guiding TSAM toward high-level object localization. Meanwhile, the DPM operates at a finer granularity, providing dense prompts aligned with the frame's spatial features, highlighting its reliance on the outputs of SAM's image encoder, as shown in Fig. 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 393, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 464 + ], + "type": "text", + "content": "Sparse prompting module (SPM). To effectively guide segmentation within TSAM using the text expression, SPM first employs a language-guided query selection mechanism to select the " + }, + { + "bbox": [ + 313, + 393, + 554, + 464 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 393, + 554, + 464 + ], + "type": "text", + "content": " most relevant audio cues, associated with the input text expression. The top " + }, + { + "bbox": [ + 313, + 393, + 554, + 464 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 393, + 554, + 464 + ], + "type": "text", + "content": " query indices, denoted as " + }, + { + "bbox": [ + 313, + 393, + 554, + 464 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 313, + 393, + 554, + 464 + ], + "type": "text", + "content": ", are selected as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 373, + 473, + 553, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 473, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 373, + 473, + 553, + 495 + ], + "type": "interline_equation", + "content": "A _ {k} = \\operatorname {T o p} _ {k} \\left(\\max \\left(\\hat {F} _ {a} \\hat {F} _ {t} ^ {\\top}\\right)\\right), \\tag {4}", + "image_path": "06fbc1225c08ca166a9d9ad4f7055e6062cf637444d2ac89509ce2f076a940af.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 498, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 498, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{Top}_k" + }, + { + "bbox": [ + 313, + 498, + 554, + 581 + ], + "type": "text", + "content": " picks the top " + }, + { + "bbox": [ + 313, + 498, + 554, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 498, + 554, + 581 + ], + "type": "text", + "content": " indices, maxf computes the maximum along the cues' dimension, and " + }, + { + "bbox": [ + 313, + 498, + 554, + 581 + ], + "type": "inline_equation", + "content": "(\\cdot)^{\\top}" + }, + { + "bbox": [ + 313, + 498, + 554, + 581 + ], + "type": "text", + "content": " indicates matrix transposition. The selected audio cues are then fused with the visual cues through a cross-attention mechanism and transformed using a feed forward network-mapped sparse prompt embedding. This process generates contextually enriched prompts for SAM's mask decoder." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Dense prompting module (DPM). The DPM in TSAM generates spatially comprehensive embeddings by integrating visual, audio, and textual cues through cross-attention mechanisms, as illustrated in Fig. 1. First, an audio cross-attention layer aligns the audio cues with the visual cues, facilitating effective fusion of both modalities. Next, a text cross-attention layer incorporates the text cues, further refining the combined audio-visual cues to focus on the target object for segmentation. Finally, a feed forward network refines the multimodal embeddings, preparing them for input into SAM's mask decoder." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23951" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 164, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 164, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 164, + 85 + ], + "type": "text", + "content": "3.5. Training objective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": "We employ binary cross-entropy " + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": " and intersection over union " + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{IoU}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": " losses to measure the discrepancy between TSAM's predicted mask and the ground truth mask during training. The total loss " + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": " for model optimization is defined as a weighted sum of " + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{IoU}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": " losses, as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 123, + 155, + 295, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 155, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 123, + 155, + 295, + 168 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {B C E}} + \\lambda \\cdot \\mathcal {L} _ {\\mathrm {I o U}}, \\tag {5}", + "image_path": "ca394ca2e2e6f19928f91b0d0d73e40e2d01760b201b86ea7a71d4929eb1ab3a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 173, + 286, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 286, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 173, + 286, + 186 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 173, + 286, + 186 + ], + "type": "text", + "content": " is set to 1.0 to balance the two loss components." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 194, + 137, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 137, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 137, + 208 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 213, + 114, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 114, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 114, + 225 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 231, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 315 + ], + "type": "text", + "content": "We evaluated TSAM using the Ref-AVS dataset [36], which contains a diverse collection of 20,000 text expressions and pixel-level annotations across 4,000 10-second videos. The dataset features a wide range of audible objects, including 20 categories of musical instruments, 8 animals, 15 machines, and 5 humans, as well as a smaller set of 3 categories of static, inaudible objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 315, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 460 + ], + "type": "text", + "content": "The dataset is divided into three sets: a training set of 2,908 videos, a validation set of 276 videos, and a test set of 818 videos. The test set is further divided into three subsets: 292 Seen videos, which includes object categories present in the training set; 269 Unseen videos, which consists of object categories not seen during training but belonging to super-categories represented in the training data, assessing the model's ability to generalize to new objects; and 257 Null videos, where the text expression refers to an object that either does not exist or is not visible in the given context, with empty true masks, requiring the model to avoid segmenting any object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 464, + 167, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 167, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 167, + 475 + ], + "type": "text", + "content": "4.2. Evaluation metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 481, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 506 + ], + "type": "text", + "content": "Following [36], we used the Jaccard index " + }, + { + "bbox": [ + 55, + 481, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 481, + 296, + 506 + ], + "type": "text", + "content": " and F-score " + }, + { + "bbox": [ + 55, + 481, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 481, + 296, + 506 + ], + "type": "text", + "content": " as performance metrics. For the Null test set, we employed:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 510, + 295, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 510, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 295, + 543 + ], + "type": "interline_equation", + "content": "S = \\sqrt {\\frac {\\text {p r e d i c t e d m a s k a r e a}}{\\text {b a c k g r o u n d a r e a}}}, \\tag {6}", + "image_path": "6b498c2059a53a66e9568d553db834d622c6e853f06024be6cefc8fdbb49e09d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 547, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 295, + 583 + ], + "type": "text", + "content": "a metric to assess expression guidance in the model. A higher " + }, + { + "bbox": [ + 55, + 547, + 295, + 583 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 547, + 295, + 583 + ], + "type": "text", + "content": " value indicates a larger mask relative to the background, suggesting less accurate expression guidance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "We initialized TSAM for the Ref-AVS task using the pretrained ViT-B variant of SAM (" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "N = 12" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "d_{emb} = 256" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "). The temporal branch consists of " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " blocks, and the selected audio queries is " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". Training was performed with PyTorch on eight AMD GPUs in a distributed setup. The AdamW optimizer was used, with an initial learning rate of " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "1 \\cdot 10^{-4}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " and a batch size of 1. The model was trained for 15 epochs with periodic evaluations, and the best-performing model on the validation subset was selected for testing." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 554, + 163 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 554, + 163 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 554, + 163 + ], + "type": "table", + "html": "
MethodTaskVisual BackboneSeenUnseenNull S(↓)
J(%)FJ(%)F
AVSBench [39]AVS†PVT-v223.200.51132.360.5470.208
AVSegFormer [8]AVS†PVT-v233.470.47036.050.5010.171
GAVS [35]AVS†SAM28.930.49829.820.4970.190
SAMA [24]AVS*SAM39.220.56247.500.5660.130
ReferFormer [37]Ref-VOS‡V-Swin31.310.50130.400.4880.176
R2-VOS [18]Ref-VOS‡V-Swin25.010.41027.930.4980.183
EEMC [36]Ref-AVSM2F34.200.51349.540.6480.007
TSAM (Ours)Ref-AVSSAM43.430.56854.580.6640.017
", + "image_path": "1651684acccf96378f73b9d61138ced4eda9cfbc92e6a24cd6d80453a640f1a7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 171, + 555, + 215 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 215 + ], + "type": "text", + "content": "Table 1. Performance comparison on the Ref-AVS dataset. " + }, + { + "bbox": [ + 313, + 171, + 555, + 215 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 313, + 171, + 555, + 215 + ], + "type": "text", + "content": " denotes text integration in AVS from [36], " + }, + { + "bbox": [ + 313, + 171, + 555, + 215 + ], + "type": "inline_equation", + "content": "{}^{ \\ddagger }" + }, + { + "bbox": [ + 313, + 171, + 555, + 215 + ], + "type": "text", + "content": " indicates audio integration in Ref-VOS from [36], and * marks our implementation with text added to AVS." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 238, + 449, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 449, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 449, + 251 + ], + "type": "text", + "content": "4.4. Results and comparison" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "content": "We compared the proposed TSAM method with the EEMC method [36] and existing methods from the AVS and Ref-VOS domains, including AVSBench [39], AVSeg-Former [8], GAVS [35], SAMA [24], ReferFormer [37], and R2-VOS [18], using the Ref-AVS dataset [36]. For consistency and fairness, we reported the published results for all methods in [36] and re-evaluated the SAMA method by integrating text with both audio and visual inputs to align with our multimodal setup." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 365, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 449 + ], + "type": "text", + "content": "The results, presented in Tab. 1, demonstrate that even when augmented with text (for AVS methods) or audio (for Ref-VOS methods), these methods do not yield satisfactory performance in the Ref-AVS task. Interestingly, methods that utilize prior segmentation visual backbones, such as SAM and M2F (Mask2Former), tend to outperform those based on visual backbones like PVT-v2 and V-Swin." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 449, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 593 + ], + "type": "text", + "content": "Although GAVS and SAMA are SAM-based methods, they perform worse than EEMC. SAMA fails to leverage SAM's flexible, promptable nature, integrating additional modality cues superficially by layering them on SAM's frozen image encoder. This underutilizes the potential of multimodal cues, limiting SAMA's segmentation capabilities. GAVS, on the other hand, prompts SAM with multimodal cues but does not address the necessary alignment and interplay among these modalities, leading to a lack of cohesive multimodal fusion. These results confirm that simple modality fusion alone is insufficient for handling the complexity of multimodal integration in the Ref-AVS task." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "In contrast, our TSAM leverages language-guided audio queries as sparse prompts, integrates multimodal dense prompts via cross-attention, and incorporates a temporal modeling branch and adapter module to enhance SAM's ability to perform accurate, contextually guided segmentation of dynamic objects in complex audio-visual scenes. Compared to the best-performing method, EEMC, on the Ref-AVS dataset, our TSAM achieves superior results: a " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = 9.23\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = 5.5\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " improvements on the Seen test set, and a " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{J} = 5.04\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = 1.6\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " improvements" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23952" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "on the Unseen test set. Although TSAM falls slightly behind EEMC on the Null test set by " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "S = 1\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", likely due to SAM's limitation of always producing a segmentation mask even when no target object is present, TSAM maintains a strong advantage on Unseen test set. This is primarily due to TSAM's ability to generalize to unseen object categories, preserving SAM's pretrained knowledge while ensuring robust segmentation in novel audio-visual scenes, thanks to the added temporal modeling branch." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 165, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 165, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 165, + 202 + ], + "type": "text", + "content": "4.5. Qualitative results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 207, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 435 + ], + "type": "text", + "content": "We further evaluated our TSAM method qualitatively on both the Seen and Unseen test sets, as shown in Fig. 2 and Fig. 3, respectively. Fig. 2 demonstrates TSAM's consistent ability to produce high-quality masks for targeted objects, even when they are inaudible, as seen in the middle example. This highlights TSAM's proficiency in processing complex multimodal instructions through its prompting modules, enabling accurate segmentation based on textual cues. On the Unseen test set, Fig. 3, TSAM showcases a remarkable capacity to generalize to novel objects, effectively aligning audio-visual and textual inputs in novel scenes to accurately segment the targeted object of interest. This underscores TSAM's ability to generalize to new contexts, interpreting diverse multimodal instructions and segmenting unseen objects. TSAM's temporal branch enhances understanding of novel objects in dynamic scenes, while SAM's pre-trained knowledge is preserved for core segmentation capabilities. Additional qualitative results and comparisons are provided in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 444, + 148, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 148, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 148, + 456 + ], + "type": "text", + "content": "4.6. Ablation study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 581 + ], + "type": "text", + "content": "We analyzed the impact of omitting key components of TSAM, including the temporal branch (TB), temporal modality fusion layer (TMFL), dense prompting module (DPM), sparse prompting module (SPM), cached memory (CM), and adapter module (AM), as shown in Tab. 2. " + }, + { + "bbox": [ + 55, + 461, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{CM(a + v)}" + }, + { + "bbox": [ + 55, + 461, + 296, + 581 + ], + "type": "text", + "content": " denotes cached memory for both audio and visual modalities, while " + }, + { + "bbox": [ + 55, + 461, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{CM(v)}" + }, + { + "bbox": [ + 55, + 461, + 296, + 581 + ], + "type": "text", + "content": " denotes cached memory on the visual modality only, see Fig. 1. Additionally, Tab. 2 presents the effect of the intersection over union " + }, + { + "bbox": [ + 55, + 461, + 296, + 581 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{IoU}})" + }, + { + "bbox": [ + 55, + 461, + 296, + 581 + ], + "type": "text", + "content": " loss, defined in Eq. (5), on TSAM's training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "As shown in Tab. 2, we observe a significant performance degradation in setting " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " when the temporal branch (TB) is removed, with a drop of " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "7.24\\% \\mathrm{pt}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "49.01\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "41.77\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "3.4\\% \\mathrm{pt}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " (from 0.616 to 0.582). This performance loss is significantly higher than that in other settings, emphasizing the critical role of the temporal modeling branch in TSAM's ability to generalize across video frames and maintain coherence in dynamic scenes, which is essential for object segmentation in complex Ref-AVS tasks. Settings " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "⑤" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " demonstrate the importance of prompting SAM with multimodal cues through SPM and" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 323, + 77, + 547, + 167 + ], + "blocks": [ + { + "bbox": [ + 323, + 77, + 547, + 167 + ], + "lines": [ + { + "bbox": [ + 323, + 77, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 323, + 77, + 547, + 167 + ], + "type": "image", + "image_path": "fa954cd612c52b82f28f8af8f5ed46650d45481d70b18fa1b33bc8b64547e88b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 167, + 535, + 175 + ], + "lines": [ + { + "bbox": [ + 328, + 167, + 535, + 175 + ], + "spans": [ + { + "bbox": [ + 328, + 167, + 535, + 175 + ], + "type": "text", + "content": "Expression: The object making a sound by being played by the woman." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 323, + 178, + 547, + 269 + ], + "blocks": [ + { + "bbox": [ + 323, + 178, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 323, + 178, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 323, + 178, + 547, + 269 + ], + "type": "image", + "image_path": "5c65021cc090792bd5d70e6eceda7a8215590934a7f7891e1e0c68f310a1780a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 268, + 506, + 275 + ], + "lines": [ + { + "bbox": [ + 354, + 268, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 354, + 268, + 506, + 275 + ], + "type": "text", + "content": "Expression: The object behind the sounding women." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 279, + 547, + 368 + ], + "blocks": [ + { + "bbox": [ + 323, + 279, + 547, + 368 + ], + "lines": [ + { + "bbox": [ + 323, + 279, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 323, + 279, + 547, + 368 + ], + "type": "image", + "image_path": "f621832f3b34724703f53198d475c12598545d490505d2b379d746a86485808a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 368, + 528, + 376 + ], + "lines": [ + { + "bbox": [ + 333, + 368, + 528, + 376 + ], + "spans": [ + { + "bbox": [ + 333, + 368, + 528, + 376 + ], + "type": "text", + "content": "Expression: The sounding object on the legs of the man on the left." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 392, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 427 + ], + "type": "text", + "content": "Figure 2. Segmentation masks for referred objects in the Ref-AVS dataset (Seen test set). Top row: ground truth with target object in green. Bottom row: our results with target object in violet." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "content": "DPM. SPM and DPM play complementary roles in TSAM, and their joint contribution is pivotal for its overall performance. Omitting these prompts in setting ⑥ causes a notable decrease in segmentation performance, reinforcing our claim that the rich semantic cues from multimodal prompts are essential for improving segmentation performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 546 + ], + "type": "text", + "content": "Settings ⑦ and ⑧ evaluate the effect of cached memory on both audio and visual (CM(a+v)) and visual-only" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 320, + 565, + 550, + 668 + ], + "blocks": [ + { + "bbox": [ + 320, + 565, + 550, + 668 + ], + "lines": [ + { + "bbox": [ + 320, + 565, + 550, + 668 + ], + "spans": [ + { + "bbox": [ + 320, + 565, + 550, + 668 + ], + "type": "table", + "html": "
SettingSeenUnseenMix(S+U)Null S(↓)
J(%)FJ(%)FJ(%)F
① TSAM43.430.56854.580.66449.010.6160.017
② - TB33.050.50750.480.65741.770.5820.505
③ - TMFL40.350.57945.540.62742.950.6030.018
④ - DPM42.720.58049.100.64745.910.6140.018
⑤ - SPM43.040.58049.750.65246.400.6160.018
⑥ - SPM+DPM42.600.60240.580.60441.590.6030.018
⑦ - CM(a+v)42.070.54449.110.65945.590.6020.018
⑧ - CM(v)42.750.54951.180.66046.970.6050.018
⑨ - AM43.130.60040.790.61741.960.6090.017
⑩ - LloU38.290.56442.150.63140.220.5980.008
", + "image_path": "acd1b89c38cee2b7d774925c1e8d5742c23139b3c793bce86b835889db314762.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 709 + ], + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 709 + ], + "type": "text", + "content": "Table 2. Ablation study on TSAM components and " + }, + { + "bbox": [ + 313, + 677, + 555, + 709 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IotU}" + }, + { + "bbox": [ + 313, + 677, + 555, + 709 + ], + "type": "text", + "content": " loss. 'Mix' refers to the average performance on the Seen (S) and Unseen (U) test sets." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23953" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 78, + 110, + 123 + ], + "blocks": [ + { + "bbox": [ + 65, + 78, + 110, + 123 + ], + "lines": [ + { + "bbox": [ + 65, + 78, + 110, + 123 + ], + "spans": [ + { + "bbox": [ + 65, + 78, + 110, + 123 + ], + "type": "image", + "image_path": "8c7b03e93af18185a0fe6dd4255c0ac7d28bf6f51c7d5b7a95d071667d9020cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 79, + 154, + 123 + ], + "blocks": [ + { + "bbox": [ + 110, + 79, + 154, + 123 + ], + "lines": [ + { + "bbox": [ + 110, + 79, + 154, + 123 + ], + "spans": [ + { + "bbox": [ + 110, + 79, + 154, + 123 + ], + "type": "image", + "image_path": "c32c6da14c5f666204c3bce2e306ac792fd2a4447757fc814cc378d195f212f7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 154, + 79, + 197, + 123 + ], + "blocks": [ + { + "bbox": [ + 154, + 79, + 197, + 123 + ], + "lines": [ + { + "bbox": [ + 154, + 79, + 197, + 123 + ], + "spans": [ + { + "bbox": [ + 154, + 79, + 197, + 123 + ], + "type": "image", + "image_path": "ab6589e26733cbfeaae20a626a2a74d187d883e0c5e54b190b5f7c2c1d943635.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 198, + 79, + 241, + 123 + ], + "blocks": [ + { + "bbox": [ + 198, + 79, + 241, + 123 + ], + "lines": [ + { + "bbox": [ + 198, + 79, + 241, + 123 + ], + "spans": [ + { + "bbox": [ + 198, + 79, + 241, + 123 + ], + "type": "image", + "image_path": "33b9aada7beee949c280a65aecc91433c8f2ab8ae0a6fb02b82281f41c6a5656.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 242, + 79, + 286, + 123 + ], + "blocks": [ + { + "bbox": [ + 242, + 79, + 286, + 123 + ], + "lines": [ + { + "bbox": [ + 242, + 79, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 242, + 79, + 286, + 123 + ], + "type": "image", + "image_path": "ec4babeb3115fcbfabfa9d094834d8d459fbae6a8e5e9e78adc07846b62c487f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 66, + 124, + 110, + 167 + ], + "blocks": [ + { + "bbox": [ + 66, + 124, + 110, + 167 + ], + "lines": [ + { + "bbox": [ + 66, + 124, + 110, + 167 + ], + "spans": [ + { + "bbox": [ + 66, + 124, + 110, + 167 + ], + "type": "image", + "image_path": "5f7aa8fb6caedfc59e0cbe6ae5563e445b31bf15ba86bcc908104324d29bb6ea.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 167, + 253, + 175 + ], + "lines": [ + { + "bbox": [ + 89, + 167, + 253, + 175 + ], + "spans": [ + { + "bbox": [ + 89, + 167, + 253, + 175 + ], + "type": "text", + "content": "Expression: The object emits a sound following the man." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 110, + 124, + 153, + 167 + ], + "blocks": [ + { + "bbox": [ + 110, + 124, + 153, + 167 + ], + "lines": [ + { + "bbox": [ + 110, + 124, + 153, + 167 + ], + "spans": [ + { + "bbox": [ + 110, + 124, + 153, + 167 + ], + "type": "image", + "image_path": "c4b75ef3ccb70fdbee0ab4814ac3c9cd31a33ff2b33258d6748beddc9fc6aca9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 154, + 124, + 197, + 167 + ], + "blocks": [ + { + "bbox": [ + 154, + 124, + 197, + 167 + ], + "lines": [ + { + "bbox": [ + 154, + 124, + 197, + 167 + ], + "spans": [ + { + "bbox": [ + 154, + 124, + 197, + 167 + ], + "type": "image", + "image_path": "d731b9618978db5dd580d34b032fc26a57a82a18d332e9f4f426b35697da581f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 198, + 124, + 241, + 167 + ], + "blocks": [ + { + "bbox": [ + 198, + 124, + 241, + 167 + ], + "lines": [ + { + "bbox": [ + 198, + 124, + 241, + 167 + ], + "spans": [ + { + "bbox": [ + 198, + 124, + 241, + 167 + ], + "type": "image", + "image_path": "f50b84d51ba7aa0130b6eaa201fef6876282ec07ba89cc474f2474f97694d335.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 242, + 124, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 242, + 124, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 242, + 124, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 242, + 124, + 286, + 167 + ], + "type": "image", + "image_path": "d55b504018cf187ef7fdc6f3c982cedd632cff44f7196b73e28b714465ec3fa9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 66, + 178, + 109, + 223 + ], + "blocks": [ + { + "bbox": [ + 66, + 178, + 109, + 223 + ], + "lines": [ + { + "bbox": [ + 66, + 178, + 109, + 223 + ], + "spans": [ + { + "bbox": [ + 66, + 178, + 109, + 223 + ], + "type": "image", + "image_path": "9316f29dbac323452a8d1bb6bd3e6d1813fb47e3af7606f8a9e827774a24b331.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 110, + 178, + 153, + 223 + ], + "blocks": [ + { + "bbox": [ + 110, + 178, + 153, + 223 + ], + "lines": [ + { + "bbox": [ + 110, + 178, + 153, + 223 + ], + "spans": [ + { + "bbox": [ + 110, + 178, + 153, + 223 + ], + "type": "image", + "image_path": "48043a4ff06e3f3f07bf9d2953df3d509b9912f82e5c4731f33253cdf01c8edd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 154, + 178, + 197, + 223 + ], + "blocks": [ + { + "bbox": [ + 154, + 178, + 197, + 223 + ], + "lines": [ + { + "bbox": [ + 154, + 178, + 197, + 223 + ], + "spans": [ + { + "bbox": [ + 154, + 178, + 197, + 223 + ], + "type": "image", + "image_path": "30bc5a0f57d31fb99726c3c45b75b9c35e5b7fcb6d7657b9170df88376eec400.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 198, + 178, + 241, + 223 + ], + "blocks": [ + { + "bbox": [ + 198, + 178, + 241, + 223 + ], + "lines": [ + { + "bbox": [ + 198, + 178, + 241, + 223 + ], + "spans": [ + { + "bbox": [ + 198, + 178, + 241, + 223 + ], + "type": "image", + "image_path": "8e4d33d4f566bf13e25228874f46641604ed59d23ecf6ebfd02dc0e056c45bf7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 242, + 178, + 286, + 223 + ], + "blocks": [ + { + "bbox": [ + 242, + 178, + 286, + 223 + ], + "lines": [ + { + "bbox": [ + 242, + 178, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 242, + 178, + 286, + 223 + ], + "type": "image", + "image_path": "af52a2d194c050723d4f408cbbf3acf1456a634d45a71fb9ce65a55daa9ed37c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 66, + 224, + 109, + 267 + ], + "blocks": [ + { + "bbox": [ + 66, + 224, + 109, + 267 + ], + "lines": [ + { + "bbox": [ + 66, + 224, + 109, + 267 + ], + "spans": [ + { + "bbox": [ + 66, + 224, + 109, + 267 + ], + "type": "image", + "image_path": "7e3e38f98b194bd14b2595f889eb55708dda7d1f7a32f28211186040b99e8083.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 110, + 224, + 153, + 267 + ], + "blocks": [ + { + "bbox": [ + 110, + 224, + 153, + 267 + ], + "lines": [ + { + "bbox": [ + 110, + 224, + 153, + 267 + ], + "spans": [ + { + "bbox": [ + 110, + 224, + 153, + 267 + ], + "type": "image", + "image_path": "dd6e51ee8b5e70966ad066467c91dd8fbe1b9aa429e31a2245850149d4ea7ad9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 268, + 216, + 275 + ], + "lines": [ + { + "bbox": [ + 127, + 268, + 216, + 275 + ], + "spans": [ + { + "bbox": [ + 127, + 268, + 216, + 275 + ], + "type": "text", + "content": "Expression: The truck moving." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 154, + 224, + 197, + 267 + ], + "blocks": [ + { + "bbox": [ + 154, + 224, + 197, + 267 + ], + "lines": [ + { + "bbox": [ + 154, + 224, + 197, + 267 + ], + "spans": [ + { + "bbox": [ + 154, + 224, + 197, + 267 + ], + "type": "image", + "image_path": "bf93de1c2b3311e4075fddaf854f346fd63dbcde11c8e943aaaad9db0ec79b37.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 198, + 224, + 241, + 267 + ], + "blocks": [ + { + "bbox": [ + 198, + 224, + 241, + 267 + ], + "lines": [ + { + "bbox": [ + 198, + 224, + 241, + 267 + ], + "spans": [ + { + "bbox": [ + 198, + 224, + 241, + 267 + ], + "type": "image", + "image_path": "db8d19084ca67bf8a8dbc5d6e43b92d69f777e97105ea05911c363cdeccbc8d5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 242, + 224, + 286, + 267 + ], + "blocks": [ + { + "bbox": [ + 242, + 224, + 286, + 267 + ], + "lines": [ + { + "bbox": [ + 242, + 224, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 242, + 224, + 286, + 267 + ], + "type": "image", + "image_path": "f2f00566eae7bd3f77a6c3f89ac282c3fe454808c6ce58e561669b2c036bc63a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 66, + 280, + 109, + 323 + ], + "blocks": [ + { + "bbox": [ + 66, + 280, + 109, + 323 + ], + "lines": [ + { + "bbox": [ + 66, + 280, + 109, + 323 + ], + "spans": [ + { + "bbox": [ + 66, + 280, + 109, + 323 + ], + "type": "image", + "image_path": "502711f764fe10c94a05052468a5334985de8dbd10a6d8209d779ac0abfd6216.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 110, + 280, + 154, + 323 + ], + "blocks": [ + { + "bbox": [ + 110, + 280, + 154, + 323 + ], + "lines": [ + { + "bbox": [ + 110, + 280, + 154, + 323 + ], + "spans": [ + { + "bbox": [ + 110, + 280, + 154, + 323 + ], + "type": "image", + "image_path": "3742fa8f2bacc2e2eb029dd50cc1b09e7cc40c35ae6260f3e1a9ea04d2d9b7e6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 154, + 280, + 197, + 323 + ], + "blocks": [ + { + "bbox": [ + 154, + 280, + 197, + 323 + ], + "lines": [ + { + "bbox": [ + 154, + 280, + 197, + 323 + ], + "spans": [ + { + "bbox": [ + 154, + 280, + 197, + 323 + ], + "type": "image", + "image_path": "6d1f5e30f311036dcb309521287ff1fe269f079980caac3b4ca969dca6feb071.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 198, + 280, + 241, + 324 + ], + "blocks": [ + { + "bbox": [ + 198, + 280, + 241, + 324 + ], + "lines": [ + { + "bbox": [ + 198, + 280, + 241, + 324 + ], + "spans": [ + { + "bbox": [ + 198, + 280, + 241, + 324 + ], + "type": "image", + "image_path": "93d6713fe4177877f852d01d0599507ada46ffd6198384b8e79f4b569f716521.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 242, + 280, + 286, + 323 + ], + "blocks": [ + { + "bbox": [ + 242, + 280, + 286, + 323 + ], + "lines": [ + { + "bbox": [ + 242, + 280, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 242, + 280, + 286, + 323 + ], + "type": "image", + "image_path": "fb02d1af3940e3208c0dae0699e861bd468293a71fd77741c54be93c8cc18c05.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 66, + 325, + 109, + 368 + ], + "blocks": [ + { + "bbox": [ + 66, + 325, + 109, + 368 + ], + "lines": [ + { + "bbox": [ + 66, + 325, + 109, + 368 + ], + "spans": [ + { + "bbox": [ + 66, + 325, + 109, + 368 + ], + "type": "image", + "image_path": "61a7bfed7db84710848a2012b9f6bbdec506c9bd816a560b5abd206bc8093198.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 369, + 238, + 376 + ], + "lines": [ + { + "bbox": [ + 107, + 369, + 238, + 376 + ], + "spans": [ + { + "bbox": [ + 107, + 369, + 238, + 376 + ], + "type": "text", + "content": "Expression: The tuba being played by a man." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 392, + 295, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 427 + ], + "type": "text", + "content": "Figure 3. Segmentation masks for referred objects in the Ref-AVS dataset (Unseen test set). Top row: ground truth with target object in green. Bottom row: our results with target object in violet." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 110, + 325, + 153, + 368 + ], + "blocks": [ + { + "bbox": [ + 110, + 325, + 153, + 368 + ], + "lines": [ + { + "bbox": [ + 110, + 325, + 153, + 368 + ], + "spans": [ + { + "bbox": [ + 110, + 325, + 153, + 368 + ], + "type": "image", + "image_path": "98fe3b13283bed1cb72c8b625aa0d65e5a9a4b48ff2a9c54e58b3486d52fd2fd.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 154, + 325, + 197, + 368 + ], + "blocks": [ + { + "bbox": [ + 154, + 325, + 197, + 368 + ], + "lines": [ + { + "bbox": [ + 154, + 325, + 197, + 368 + ], + "spans": [ + { + "bbox": [ + 154, + 325, + 197, + 368 + ], + "type": "image", + "image_path": "0034bca49334dcff8e003e01bb79eeb4561a08c890f881f67da96789d64d4d3b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 198, + 325, + 241, + 368 + ], + "blocks": [ + { + "bbox": [ + 198, + 325, + 241, + 368 + ], + "lines": [ + { + "bbox": [ + 198, + 325, + 241, + 368 + ], + "spans": [ + { + "bbox": [ + 198, + 325, + 241, + 368 + ], + "type": "image", + "image_path": "44433495be3ce2ebba79580956a614f87c40deea78409b0b20cfbb67204582bd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 242, + 325, + 286, + 368 + ], + "blocks": [ + { + "bbox": [ + 242, + 325, + 286, + 368 + ], + "lines": [ + { + "bbox": [ + 242, + 325, + 286, + 368 + ], + "spans": [ + { + "bbox": [ + 242, + 325, + 286, + 368 + ], + "type": "image", + "image_path": "0bb20ecef774b8cc3526fa1af37b624dd489ba20885f941ee2a27506ef3fed29.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 449, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 532 + ], + "type": "inline_equation", + "content": "(\\mathrm{CM}(\\mathrm{v}))" + }, + { + "bbox": [ + 55, + 449, + 295, + 532 + ], + "type": "text", + "content": " modalities, demonstrate a performance degradation in TSAM when cached memory is not applied across modalities " + }, + { + "bbox": [ + 55, + 449, + 295, + 532 + ], + "type": "inline_equation", + "content": "(\\mathrm{CM}(\\mathrm{a} + \\mathrm{v}))" + }, + { + "bbox": [ + 55, + 449, + 295, + 532 + ], + "type": "text", + "content": " and when only the visual modality's memory " + }, + { + "bbox": [ + 55, + 449, + 295, + 532 + ], + "type": "inline_equation", + "content": "(\\mathrm{CM}(\\mathrm{v}))" + }, + { + "bbox": [ + 55, + 449, + 295, + 532 + ], + "type": "text", + "content": " is omitted. The latter specifically emphasizes the importance of maintaining shared memory across video frames for effective temporal alignment and multimodal integration, as evidenced by the results." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "content": "The results from setting ⑨ highlight the adapter module's importance. Its omission causes a significant performance drop, underscoring its crucial role in TSAM. We also observe that the components of TSAM, particularly in settings ③ - ⑨, have a substantial impact on performance, especially on the Unseen test set. This set challenges the model to segment novel objects, demonstrating TSAM's ability to generalize beyond seen categories, an essential feature for robust performance in open-world scenarios." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Finally, setting ⑩ highlights the distinct contribution of " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{IoU}}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " to the overall loss function " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": ". Notably, the model performs best on both Seen and Unseen test sets when " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{IoU}}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " is included in the total loss. Omitting " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{IoU}}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " results in the highest performance on the Null test set. Further ablation studies of " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " (Eq. (5)) are in the supplementary materials." + } + ] + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 321, + 70, + 550, + 152 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 550, + 152 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 550, + 152 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 550, + 152 + ], + "type": "table", + "html": "
TSAMSeenUnseenMix(S+U)Null S(↓)
J(%)FJ(%)FJ(%)F
k=343.580.57950.430.65547.010.6170.018
k=543.430.56854.580.66449.010.6160.017
k=743.240.57346.950.63045.100.6010.018
M=242.040.56653.510.65147.760.6090.023
M=443.430.56854.580.66449.010.6160.017
M=643.270.57549.860.64046.570.6080.020
", + "image_path": "b9248c9e9e5c7af0b14a881491539e6bfb21eb46e06c6d2b14e2f29804746d00.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 160, + 553, + 183 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 553, + 183 + ], + "type": "text", + "content": "Table 3. Effect of audio queries " + }, + { + "bbox": [ + 313, + 160, + 553, + 183 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 160, + 553, + 183 + ], + "type": "text", + "content": " and temporal branch depth " + }, + { + "bbox": [ + 313, + 160, + 553, + 183 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 160, + 553, + 183 + ], + "type": "text", + "content": " on TSAM's performance." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "content": "We further evaluated the impact of varying the number of audio queries " + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "content": ", defined in Eq. (4), and the temporal branch depth " + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "content": ", see Sec. 3.3, on TSAM's performance, as shown in Tab. 3. Among the configurations " + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "inline_equation", + "content": "k = 7" + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "content": ", we observe that " + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "content": " performs best. This balance optimizes the relevance of audio cues without overloading the model with irrelevant queries to SAM's mask decoder." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "text", + "content": "Similarly, we find that " + }, + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "text", + "content": " yields optimal results for the temporal branch. A shallower setup (" + }, + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "text", + "content": ") likely lacks the temporal depth required for effective generalization and multimodal integration, while a deeper branch (" + }, + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "inline_equation", + "content": "M = 6" + }, + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "text", + "content": ") may introduce excessive complexity, impairing the model's generalization to unseen objects." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 370, + 391, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 391, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 391, + 383 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 390, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 628 + ], + "type": "text", + "content": "This paper presented TSAM, a novel extension of the Segment Anything Model (SAM) tailored for the referring audio-visual segmentation (Ref-AVS) task. TSAM enhances SAM's segmentation capabilities by effectively integrating multimodal cues through data-driven dense and sparse prompting modules while modeling temporal dependencies via a dedicated temporal branch. This approach preserves SAM's foundational knowledge while improving performance, particularly with novel objects. Extensive experiments on the Ref-AVS dataset demonstrate TSAM's strong generalization ability, showcasing its effectiveness in segmenting objects from textual cues and its robustness in unseen scenarios. These results highlight TSAM's potential to advance cross-modal understanding in multimodal tasks. Limitations and future work. Although TSAM outperforms existing Ref-AVS methods, a direct comparison with the recent SAM2 [31] remains to be explored. SAM2, designed for video segmentation with a memory module for temporal visual representations, lacks multimodal support. Future work will adapt SAM2 for the Ref-AVS task." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the Research Council of Finland through the Finnish Center for Artificial Intelligence flagship (grant 345552) and the USSEE project (grant 345791). We also acknowledge CSC-IT Center for Science, Finland, for awarding access to the LUMI supercomputer, owned by the EuroHPC Joint Undertaking." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23954" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In CVPR, pages 4985-4995, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "text", + "content": "[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the Kinetics dataset. In CVPR, pages 6299-6308, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 295, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 204 + ], + "type": "text", + "content": "[3] Yuanhong Chen, Yuyuan Liu, Hu Wang, Fengbei Liu, Chong Wang, Helen Frazer, and Gustavo Carneiro. Unraveling instance associations: A closer look for audio-visual segmentation. In CVPR, pages 26497–26507, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 294, + 249 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 294 + ], + "type": "text", + "content": "[5] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 340 + ], + "type": "text", + "content": "[6] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 294, + 407 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "type": "text", + "content": "[8] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In AAAI, pages 12155–12163, 2024. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 294, + 475 + ], + "type": "text", + "content": "[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In CVPR, pages 5958-5966, 2018. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 477, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 295, + 531 + ], + "type": "text", + "content": "[10] Yuxin Guo, Siyang Sun, Shuailei Ma, Kecheng Zheng, Xiaoyi Bao, Shijie Ma, Wei Zou, and Yun Zheng. Cross-MAE: Cross-modality masked autoencoders for region-aware audio-visual pre-training. In CVPR, pages 26721–26731, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "type": "text", + "content": "[11] Dawei Hao, Yuxin Mao, Bowen He, Xiaodong Han, Yuchao Dai, and Yiran Zhong. Improving audio-visual segmentation with bidirectional generation. In AAAI, pages 2067–2075, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "text", + "content": "[12] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. CNN architectures for large-scale audio classification. In ICASSP, pages 131-135, 2017. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "type": "text", + "content": "[13] Xiao Hu, Basavaraj Hampiholi, Heiko Neumann, and Jochen Lang. Temporal context enhanced referring video object segmentation. In WACV, pages 5574-5583, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[14] Shaofei Huang, Han Li, Yuqing Wang, Hongji Zhu, Jiao Dai, Jizhong Han, Wenge Rong, and Si Liu. Discovering sounding objects by audio queries for audio visual segmentation. In IJCAI, pages 875-883, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 1, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[16] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, pages 19108-19118, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[17] Kexin Li, Zongxin Yang, Lei Chen, Yi Yang, and Jun Xiao. Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. In ACM MM, pages 1485–1494, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 553, + 251 + ], + "type": "text", + "content": "[18] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. In ICCV, pages 22236-22245, 2023. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 297 + ], + "type": "text", + "content": "[19] Yonglin Li, Jing Zhang, Xiao Teng, Long Lan, and Xinwang Liu. RefSAM: Efficiently adapting segmenting anything model for referring video object segmentation. arXiv preprint arxiv.org/abs/2307.00997, 2024. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 298, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 341 + ], + "type": "text", + "content": "[20] Yuhang Ling, Yuxi Li, Zhenye Gan, Jiangning Zhang, Ming-min Chi, and Yabiao Wang. TransAVS: End-to-end audiovisual segmentation with transformer. In ICASSP, pages 7845-7849, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[21] Chen Liu, Peike Patrick Li, Xingqun Qi, Hu Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Audio-visual segmentation by exploring cross-modal mutual semantics. In ACM MM, pages 7590–7598, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[22] Chen Liu, Peike Li, Hu Zhang, Lincheng Li, Zi Huang, Dadong Wang, and Xin Yu. BAVS: Bootstrapping audiovisual segmentation by integrating foundation knowledge. IEEE TMM, 26:10015-10028, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 433, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 475 + ], + "type": "text", + "content": "[23] Jinxiang Liu, Chen Ju, Chaofan Ma, Yanfeng Wang, Yu Wang, and Ya Zhang. Audio-aware query-enhanced transformer for audio-visual segmentation. arXiv preprint arXiv:2307.13236, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[24] Jinxiang Liu, Yu Wang, Chen Ju, Chaofan Ma, Ya Zhang, and Weidi Xie. Annotation-free audio-visual segmentation. In WACV, pages 5604–5614, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "text", + "content": "[25] Y Liu, M Ott, N Goyal, J Du, M Joshi, D Chen, O Levy, M Lewis, L Zettlemoyer, and V Stoyanov. RoBERTa: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "text", + "content": "[26] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In CVPR, pages 3202-3211, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "type": "text", + "content": "[27] Yuxin Mao, Jing Zhang, Mochu Xiang, Yunqiu Lv, Yi ran Zhong, and Yuchao Dai. Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[28] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In ICCV, pages 954–965, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] Shentong Mo and Yapeng Tian. AV-SAM: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836, 2023. 2, 4" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23955" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 542 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[30] Federico Perazzi, Anna Khoreva, Rodrigo Benenson, Bernt Schiele, and Alexander Sorkine-Hornung. Learning video object segmentation from static images. In CVPR, pages 2663-2672, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 193 + ], + "type": "text", + "content": "[31] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. In ICLR, 2025. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 228 + ], + "type": "text", + "content": "[32] Seonguk Seo, Joon-Young Lee, and Bohyung Han. URVOS: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 273 + ], + "type": "text", + "content": "[33] Juhyeong Seon, Woobin Im, Sebin Lee, Jumin Lee, and Sung-Eui Yoon. Extending segment anything model into auditory and temporal dimensions for audio-visual segmentation. In ICIP, pages 2480–2486, 2024. 1, 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "type": "text", + "content": "[34] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. PVT v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "text", + "content": "[35] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. In AAAI, pages 5669–5677, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "text", + "content": "[36] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-AVS: Refer and segment objects in audio-visual scenes. In ECCV, pages 196–213, 2024. 1, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 441 + ], + "type": "text", + "content": "[37] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 496 + ], + "type": "text", + "content": "[38] Jiarui Yu, Haoran Li, Yanbin Hao, Jinmeng Wu, Tong Xu, Shuo Wang, and Xiangnan He. How can contrastive pretraining benefit audio-visual segmentation? A study from supervised and zero-shot perspectives. In BMVC, pages 367–374, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "text", + "content": "[39] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In ECCV, pages 386–403, 2022. 1, 3, 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23956" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_content_list.json b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5e602904479dbc16a550b6b146729b29dfe8be62 --- /dev/null +++ b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_content_list.json @@ -0,0 +1,1797 @@ +[ + { + "type": "text", + "text": "TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution", + "text_level": 1, + "bbox": [ + 109, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linwei Dong $^{1,2*}$ Qingnan Fan $^{2*}$ Yihong Guo $^{1}$ Zhonghao Wang $^{3}$ Qi Zhang $^{2}$ Jinwei Chen $^{2}$ Yawei Luo $^{1\\dagger}$ Changqing Zou $^{1,4}$", + "bbox": [ + 179, + 200, + 759, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University $^{2}$ Vivo Mobile Communication Co. Ltd $^{3}$ University of Chinese Academy of Sciences $^{4}$ Zhejiang Lab", + "bbox": [ + 217, + 255, + 736, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-trained text-to-image diffusion models are increasingly applied to real-world image super-resolution (RealISR) task. Given the iterative refinement nature of diffusion models, most existing approaches are computationally expensive. While methods such as SinSR and OSEDiff have emerged to condense inference steps via distillation, their performance in image restoration or details recovery is not satisfied. To address this, we propose TSD-SR, a novel distillation framework specifically designed for real-world image super-resolution, aiming to construct an efficient and effective one-step model. We first introduce the Target Score Distillation, which leverages the priors of diffusion models and real image references to achieve more realistic image restoration. Secondly, we propose a Distribution-Aware Sampling Module to make detail-oriented gradients more readily accessible, addressing the challenge of recovering fine details. Extensive experiments demonstrate that our TSD-SR has superior restoration results (most of the metrics perform the best) and the fastest inference speed (e.g. 40 times faster than SeeSR) compared to the past Real-ISR approaches based on pre-trained diffusion priors.", + "bbox": [ + 75, + 358, + 473, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 702, + 209, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image super-resolution (ISR) [8, 9, 23, 26] aims to transform low-quality (LQ) images, which have experienced noise or blur, into clear high-quality (HQ) images. Differing from traditional ISR [6, 66], which assumes a known degradation process, real-world image super-resolution (RealISR) [46, 62] is designed to enhance real-world images that have suffered from complex and unknown degradations, thereby offering greater practical utility.", + "bbox": [ + 75, + 728, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative models, particularly Generative Adversarial Networks (GANs) [11, 33, 37] and Diffusion Models (DMs)", + "bbox": [ + 75, + 849, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35b3f4bb33ba96096d711b0cce07fd28f9b642fcdc5828012c89c2bc0af04fdb.jpg", + "image_caption": [ + "Figure 1. Performance and efficiency comparison among RealISR methods. TSD-SR stands out for achieving high-quality restoration with the fastest speed among diffusion-based models. In contrast, existing models prioritize either speed or restoration performance. The performance of each method is benchmarked on an A100 GPU with the DRealSR dataset." + ], + "image_footnote": [], + "bbox": [ + 513, + 349, + 857, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[17, 39, 41], have shown remarkable power in handling Real-ISR task. GAN-based methods leverage adversarial training, toggling between the generator and discriminator to produce realistic images. While GANs are capable of one-step inference, they are commonly hampered by issues like mode collapse and training instability [2]. Recently, Diffusion Models (DMs) have shown impressive performance in the realm of image generation [21, 47]. Their robust priors empower them to produce more realistic images with richer details than GAN-based methods [39, 41]. Some researchers [29, 53, 56, 60] have successfully leveraged pre-trained DMs for Real-ISR task. However, due to the iterative denoising nature of diffusion models [17], the Real-ISR process is computationally expensive.", + "bbox": [ + 496, + 654, + 893, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve an efficient and one-step network akin to GANs, several pioneering methods that condense the iter-", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding authors.", + "bbox": [ + 94, + 886, + 344, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23174", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ations of diffusion models through distillation [12, 15, 18, 57] have been proposed [48, 52, 54]. Among these works, OSEDiff [52] introduced the Variational Score Distillation (VSD) loss [50] to Real-ISR task, achieving state-of-the-art (SOTA) one-step performance by leveraging prior knowledge from pre-trained models. Despite these advancements, our investigation has revealed two critical limitations associated with VSD in Real-ISR applications. (1) Unreliable gradient direction. VSD relies on a Teacher Model to provide a \"true gradient direction.\" However, this guidance proves unreliable in scenarios where initial ISR outputs are suboptimal. (2) Insufficient detail recovery. The VSD loss exhibits notable variation across different timesteps, and the uniform sampling strategy for $t$ poses challenges in aligning the score function with detailed texture recovery requirements. These findings highlight the need for enhanced approaches to address these issues effectively.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel method called TSD-SR to distill a multi-step Text-to-Image (T2I) DMs [10, 38, 39] into an effective one-step diffusion model tailored for the Real-ISR task. Specifically, TSD-SR consists of two components: Target Score Distillation (TSD) and Distribution-Aware Sampling Module (DASM). TSD incorporates our new proposed Target Score Matching (TSM) loss to compensate for the limitation of VSD loss. This significant score loss leverages HQ data to provide a reliable optimization trajectory for the distillation process. It effectively reduces the visual artifacts caused by deviant predictions from the Teacher Model. DASM is crafted to bolster detail recovery by strategically sampling low-noise samples that are distribution-based during training. This approach effectively allocates more optimization to early timesteps in a single iteration, enhancing the recovery of details.", + "bbox": [ + 75, + 347, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments on popular benchmarks demonstrate that TSD-SR achieves superior restoration performance (most of the metrics perform the best) and high efficiency (the fastest inference speed, 40 times faster than SeeSR) compared to the state-of-the-art Real-ISR methods based on pre-trained DMs, while requiring only a single inference step.", + "bbox": [ + 75, + 589, + 472, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contribution can be summarized as threefold:", + "bbox": [ + 76, + 680, + 441, + 694 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel method called TSD-SR to achieve one-step DMs distillation for the Real-ISR task.", + "- We introduce Target Score Distillation (TSD) to provide a reliable gradient to enhance the realism of the Real-ISR model's outputs.", + "- We design a Distribution-Aware Sampling Module (DASM) specifically tailored to enhance the capability of detail restoration." + ], + "bbox": [ + 76, + 695, + 468, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 829, + 218, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GAN-based Real-ISR. Since SRGAN [26] first applied GAN to ISR, it has effectively enhanced visual quality by combining adversarial loss with perceptual loss [7,", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[65]. Subsequently, ESRGAN [45] introduced Residual-in-Residual Dense Block and a relativistic average discriminator, further improving detail restoration. Methods like BSRGAN [62] and Real-ESRGAN [46] simulate complex real-world degradation processes, achieving ISR under unknown degradation conditions, which enhances the model's generalization ability. Although GAN-based methods are capable of adding more realistic details to images, they suffer from training instability and mode collapse [2].", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-step Diffusion-based Real-ISR. Some researches [29, 44, 53, 56, 60] in recent years have utilized the powerful image priors in pre-trained T2I diffusion models [35, 39, 64] for Real-SR tasks and achieved promising results. For example, StableSR [44] balances fidelity and perceptual quality by fine-tuning the time-aware encoder and employing controllable feature wrapping. DiffBiR [29] first processes the LR image through a reconstruction network and then uses the Stable Diffusion (SD) model [39] to supplement the details. SeeSR [53] attempts to better stimulate the generative power of the SD model by extracting the semantic information in the image as a conditional guide. PASD [56] introduces a pixel-aware cross attention module to enable the diffusion model to perceive the local structure of the image at the pixel level, while using a degradation removal module to extract degradation insensitive features to guide the diffusion process along with high-level information from the image. SUPIR [60] achieves a generative and fidelity capability using negative cues [16] as well as restoration-guided sampling, while using a larger pretraining model with a larger dataset to enhance the model capability. However, all of these methods are limited by the multi-step denoising of the diffusion model, which requires 20-50 iterations in inference, resulting in an inference time that lags far behind that of GAN-based methods.", + "bbox": [ + 496, + 231, + 892, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One-step Diffusion-based Real-ISR. Recently, there has been a surge of interest within the academic community in one-step distillation techniques [31, 34, 40, 58, 59] for diffusion-based Real-ISR task. SinSR [48] leverages consistency preserving distillation to condense the inference steps of ResShift [61] into a single step, yet the generalization of ResShift and SinSR is constrained due to the absence of large-scale data training. AddSR [54] introduces the adversarial diffusion distillation (ADD) [40] to RealISR tasks, resulting in a comparatively effective four-step model. However, this method has a propensity to produce excessive and unnatural image details. OSEDiff [52] directly uses LQ images as the beginning of the diffusion process, and employs VSD loss [50] as a regularization technique to condense a multi-step pre-trained T2I model into a one-step Real-ISR model. However, due to the incorporation of alternating training strategies, OSEDiff may initially tend towards unreliable optimization directions, which may lead to the visual artifacts.", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23175", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/797b3e0fbfafedb1fb8afd5e16397e0d4c56c5df62a332467c77473a2ed8a2ad.jpg", + "image_caption": [ + "Figure 2. Pipeline overview. We train one-step Student Model $G_{\\theta}$ to map the low-quality image $x_{L}$ into a more realistic one. The noisy latent $\\hat{z}_{t}$ and $z_{t}$ sampled by DASM (Details can be found in Fig. 6.) will be fed into both the pre-trained Teacher and LoRA Model to produce the Variational Score loss. Subsequently, two Models' predictions on $z_{t}$ yield Target Score loss. Their weighted forms, namely TSD (red flow), in conjunction with pixel-space reconstruction loss (green flow), are leveraged to update the Student Model. After updating the Student Model, we employ the diffusion loss (blue flow) to alternately update the LoRA Model." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 890, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 494, + 210, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 523, + 217, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. ISR problem aims to reconstruct a HQ image $x_{H}$ from a LQ input $x_{L}$ by training a parameterized ISR model $G_{\\theta}$ on a dataset $\\mathcal{D} = \\{(x_{L},x_{H})_{i = 1}^{N}\\}$ , where $N$ represents the number of image pairs. It is mathematically formulated as minimizing the following objective:", + "bbox": [ + 75, + 550, + 468, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta^ {*} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x _ {L}, x _ {H}) \\sim \\mathcal {D}} \\left[ \\mathcal {L} _ {R e c} \\left(G _ {\\theta} \\left(x _ {L}\\right), x _ {H}\\right) \\right. \\tag {1} \\\\ \\left. \\left. + \\lambda \\mathcal {L} _ {R e g} \\left(q _ {\\theta} \\left(\\hat {x} _ {H}\\right), p \\left(x _ {H}\\right)\\right) \\right] \\right. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 630, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_{Rec}$ denotes the reconstruction loss, commonly measured using distance metrics like $L_{2}$ or LPIPS [65]. The regularization term $\\mathcal{L}_{Reg}$ enhances the realism and generalization of the ISR model's outputs. This objective can be understood as aligning the ISR output $\\hat{x}_H$ 's distribution, $q_{\\theta}(\\hat{x}_H)$ , with the true distribution $p(x_H)$ by minimizing the KL-divergence [25]:", + "bbox": [ + 75, + 676, + 468, + 782 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathcal {D} _ {\\mathrm {K L}} \\left(q _ {\\theta} \\left(\\hat {x} _ {H}\\right) \\| p \\left(x _ {H}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 786, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While several studies [45, 46, 62] have employed adversarial loss to optimize this objective, they often encounter issues like mode collapse and training instability. Recent work [52] achieved the state-of-the-art results using Variational Score Distillation (VSD) as the regularization loss to minimizing this objective, which inspires our research.", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational Score Distillation. Variational Score Distillation (VSD) [50] was initially introduced for text-to-3D generation by distilling a pre-trained text-to-image diffusion model to optimize a single 3D representation [36].", + "bbox": [ + 496, + 496, + 890, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the VSD framework, a pre-trained diffusion model, represented as $\\epsilon_{\\psi}$ , and its trainable (LoRA [19]) replica $\\epsilon_{\\phi}$ are used to regularize the generator network $G_{\\theta}$ . As outlined in ProlificDreamer [50], the gradient with respect to the generator parameters $\\theta$ is formulated as follows:", + "bbox": [ + 496, + 558, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {V S D}} (\\hat {\\boldsymbol {z}}, c _ {y}) \\\\ = \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\left(\\boldsymbol {\\epsilon} _ {\\psi} (\\hat {\\boldsymbol {z}} _ {t}; t, c _ {y}) - \\boldsymbol {\\epsilon} _ {\\phi} (\\hat {\\boldsymbol {z}} _ {t}; t, c _ {y})\\right) \\frac {\\partial \\hat {\\boldsymbol {z}}}{\\partial \\boldsymbol {\\theta}} \\right] \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 638, + 890, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{z}}_t = \\alpha_t\\hat{\\pmb{z}} +\\sigma_t\\pmb {\\epsilon}$ is the noisy input, $\\hat{z}$ is the latent outputted by the generator network $G_{\\theta}$ , $\\pmb{\\epsilon}$ is a gaussian noise, and $\\alpha_{t},\\sigma_{t}$ are the noise-data scaling constants. $c_{y}$ is a text embedding corresponding to a caption that describes the input image, and $w(t)$ is a time-varying weighting function.", + "bbox": [ + 496, + 698, + 890, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview of TSD-SR", + "text_level": 1, + "bbox": [ + 500, + 785, + 697, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, our goal is to distill a given pretrained T2I DMs into a fast one-step Student Model $G_{\\theta}$ with Teacher Model $\\epsilon_{\\psi}$ and trainable LoRA Model $\\epsilon_{\\phi}$ . We denote the latent outputs of the distilled model as $\\hat{z}_0$ , and HQ latent representations as $z_0$ . Both $\\hat{z}_0$ and $z_0$ will pass through our Distribution-Aware Sampling Module (DASM)", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23176", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29c9948cfc3145c1a6289ec20654a3b42a21fe596a103deb8ff174ce8ada520d.jpg", + "image_caption": [ + "Figure 3. A visual comparison of the gradient direction. We set the timestep $t$ to 100 and calculated the cosine similarity between the predicted directions from the Teacher Model and the true direction (towards the HQ data). The prediction direction for $z_{t}$ closely matches the true direction, but not for $\\hat{z}_{t}$ , suggesting that suboptimal samples may lead to directional deviations." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 467, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to obtain distribution-based samples $\\hat{z}_t$ and $z_t$ (Sec. 3.4). We train $G_{\\theta}$ by minimizing the two losses: a reconstruction loss in pixel space to compare the model outputs against the ground truth, and a regularization loss (from Target Score Distillation) to enhances the realism (Sec. 3.3). After updating the Student Model, we update the LoRA Model with diffusion loss. Finally, in Sec. 3.5, we present an overview of all the losses encountered during the training phase.", + "bbox": [ + 75, + 381, + 467, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Target Score Distillation", + "text_level": 1, + "bbox": [ + 76, + 513, + 299, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to [52], we introduce VSD loss into our work as a regularization loss to enhance the realism and generalization of the $G_{\\theta}$ outputs. Upon reviewing VSD Eq. (3), $\\epsilon_{\\phi}(\\hat{z}_t;t,c_y)$ represents the gradient direction of the optimization on $G_{\\theta}$ noisy outputs $\\hat{z}_t$ , whereas $\\epsilon_{\\psi}(\\hat{z}_t;t,c_y)$ corresponds to the gradient direction on noisy HQ latent representations $z_{t}$ . The overarching goal of model optimization is to harmonize the predictions from $\\hat{z}_t$ with those from $z_{t}$ Nonetheless, this strategy encounters hurdles, especially in the early training phase: the quality of synthetic latents $\\hat{z}_t$ is not high enough for Teacher Model to give a precise prediction. As illustrated in Fig. 3, the Teacher Model finds it challenging to accurately predict the optimization direction for low-quality synthetic latents, with the cosine similarity to the ideal optimization direction only 0.2 (0.88 on HQ latents). This problem can lead to severe visual artifacts, as evident in Fig. 4(a).", + "bbox": [ + 75, + 537, + 467, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A straightforward remedial measure is to employ a mean squared error (MSE) loss to align the synthetic latents with the ideal inputs of the Teacher Model, which are derived from HQ latents. However, as shown in Fig. 4(b), this approach has been observed to lead to over-smoothed results [13]. Our strategy, instead, is to align the predictions made by the Teacher Model on both synthetic and HQ", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10d97d354a5ad2a357932d8ac64204b165d462476582967da01ee481ff54b6b0.jpg", + "image_caption": [ + "(a) Naive" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 627, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3e6306ca47e5f62481283641e1ba077a5bec164d6a5e866fd6ece3772e3a3de6.jpg", + "image_caption": [ + "(b) MSE" + ], + "image_footnote": [], + "bbox": [ + 633, + 88, + 759, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1d4e80265ad8eb0f1c332342eac376b800792314140e9b65d297f2cc55340347.jpg", + "image_caption": [ + "(c) Ours", + "Figure 4. The visualization of different strategies. (a) The naive method introduces fake textures and fails to recover fine details. (b) MSE leads to over-smoothed generation result, lacking high-frequency information. (c) Our method offers the superior visual effects and fine textures." + ], + "image_footnote": [], + "bbox": [ + 764, + 88, + 890, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "latents, thereby encouraging greater consistency between them. The core idea is that for samples drawn from the same distribution, the real scores predicted by the Teacher Model should be close to each other. We refer to this approach as Target Score Matching (TSM):", + "bbox": [ + 498, + 337, + 890, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {T S M}} (\\hat {\\mathbf {z}}, \\mathbf {z}, c _ {y}) \\\\ = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\psi} \\left(\\hat {\\mathbf {z}} _ {t}; t, c _ {y}\\right) - \\epsilon_ {\\psi} \\left(\\mathbf {z} _ {t}; t, c _ {y}\\right)\\right) \\frac {\\partial \\hat {\\mathbf {z}}}{\\partial \\theta} \\right] \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 419, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the expectation of the gradient is computed across all diffusion timesteps $t \\in \\{1, \\dots, T\\}$ and $\\epsilon \\sim \\mathcal{N}(0, I)$ . Equation (4) encapsulates the optimization loss for our Target Score Matching. Upon examining it in conjunction with the Eq. (3), we notice that VSD utilizes the prediction residual between a Teacher and a LoRA Model to drive gradient backpropagation. Conversely, our TSM employs the synthetic and the HQ data to produce the gradients. By blending these two strategies with hyperparameter weighting $\\lambda$ and $1 - \\lambda$ , we can construct our ultimate optimization loss, effectively unifying the strengths of both approaches to optimize the training process Eq. (5).", + "bbox": [ + 498, + 477, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {T S D}} (\\hat {\\mathbf {z}}, \\mathbf {z}, c _ {y}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left[ \\epsilon_ {\\psi} \\left(\\hat {\\mathbf {z}} _ {t}; t, c _ {y}\\right) - \\right. \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 664, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\left. \\epsilon_ {\\psi} (\\mathbf {z} _ {t}; t, c _ {y}) + \\lambda \\left(\\epsilon_ {\\psi} \\left(\\mathbf {z} _ {t}; t, c _ {y}\\right) - \\epsilon_ {\\phi} \\left(\\hat {\\mathbf {z}} _ {t}; t, c _ {y}\\right)\\right) \\right] \\frac {\\partial \\hat {\\mathbf {z}}}{\\partial \\theta} \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 691, + 890, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w(t)$ is time-aware weighting function tailored for Real-ISR. Other symbols are in accordance with those previously mentioned. By introducing the prediction of pretrained diffusion model on HQ latents, we have circumvented the issue of the model falling into the visual artifacts or over-smoothed problem, as illustrated in Fig. 4(c).", + "bbox": [ + 500, + 726, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Distribution-Aware Sampling Module", + "text_level": 1, + "bbox": [ + 500, + 830, + 823, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the VSD-based framework, we need to match the score functions predicted by the Teacher Model with those by the LoRA Model at different timesteps $t \\sim \\{0,1,\\dots,T\\}$ .", + "bbox": [ + 500, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23177", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/34e2f5d7aa72822b49fb6fc97431cef0b4ac1ec61c8aa52772cf1cedcbe766e4.jpg", + "image_caption": [ + "(a) Predicted Residual at Different Timesteps in VSD" + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 357, + 250 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2e4c1def5eba71ccb05c319f4d676ebe73146dc846ffc704abe3c00a79af3db5.jpg", + "image_caption": [ + "(b) Stage 1" + ], + "image_footnote": [], + "bbox": [ + 367, + 89, + 467, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1d8cbadb7f05c3880a3daec4e9187ba99c160cdeb44642c7832c2188d7da78f0.jpg", + "image_caption": [ + "(c) Stage 2", + "Figure 5. (a) The prediction errors of the VSD loss at different timesteps. The error divergence is more pronounced in early timesteps than later. This phenomenon is observed throughout the optimization process. (b) the visualization of Stage 1 prediction error. (c) the visualization of Stage 2 prediction error." + ], + "image_footnote": [], + "bbox": [ + 367, + 180, + 467, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, for the Real-ISR problem, this matching ability at different $t$ is unbalanced, as illustrated in Fig. 5(a). This phenomenon could be attributed to our leveraging of the low-frequency (LF) priors from LQ data, but losing guidance from high-frequency (HF) details. The noisy samples $\\hat{z}_t$ possesses low-frequency priors (from LQ data). These priors are easily learned by the LoRA Model, leading to more similar predictions (Fig. 5(b)) during LF restoration (Stage 1). However, In the Stage 2, the lack of HF details in $\\hat{z}_t$ causes the LoRA Model to poorly perceive details, resulting in divergent predictions (Fig. 5(c)). Our goal is to reduce this divergence.", + "bbox": [ + 75, + 371, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing method, for each iteration, only matches the score function at a single timestep sample $\\hat{z}_t$ , with uniform sampling strategy of time $t$ . This leads to slow convergence or even difficulty in optimization during Stage 2, as the gradients of important timesteps are averaged out. To this end, we propose our Distribution-Aware Sampling Module (DASM). This module accumulates optimization gradients for earlier timestep samples in a single iteration, enabling the backpropagation of more gradients focused on detail optimization. As shown in Fig. 6, we first introduce noise into the synthetic latent representation $\\hat{z}_t = (1 - \\sigma_t)\\hat{z}_0 + \\sigma_t\\epsilon$ , where $\\sigma$ is a weight factor and $\\epsilon$ is a gaussian noise. Subsequently, we employ a LoRA Model to perform denoising, yielding the previous time noisy samples by Eq. (6):", + "bbox": [ + 75, + 556, + 468, + 767 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {\\boldsymbol {t} - 1} = \\hat {z} _ {\\boldsymbol {t}} + \\left(\\sigma_ {t - 1} - \\sigma_ {t}\\right) \\cdot \\epsilon_ {\\phi} \\left(\\hat {z} _ {\\boldsymbol {t}}; t, c _ {y}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 773, + 468, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma_{t - 1}$ and $\\sigma_t$ are from the flow matching scheduler, and here LoRA Model has fit $\\hat{z}_0$ 's distribution. Similarly, $z_{t - 1}$ can be obtained by denoising on the Teacher Model. Ultimately, in a single iteration, the gradients from noise samples along its trajectory can be accumulated to update the Student Model. Since the obtained samples follow the diffusion sampling trajectory and are directed towards early", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/40af5971cae9330009f7fe1a9334a81903a124f25e487820766a53e91c2386d4.jpg", + "image_caption": [ + "Figure 6. Illustration of DASM. Top: The naive method that add noise to samples. Bottom: Our DASM leverages the priors of diffusion models to sample noise that more closely aligns with the true sampling trajectory. The noisy samples can all serve as inputs to the subsequent network to generate gradient backpropagation." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 890, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "timestep, we effectively optimize the divergence in Stage 2.", + "bbox": [ + 498, + 279, + 888, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Objective", + "text_level": 1, + "bbox": [ + 500, + 303, + 679, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We summarize all the losses we used in our framework. Student Model $G_{\\theta}$ . We train our Student Model with the reconstruction loss $\\mathcal{L}_{Rec}$ and the regularization loss $\\mathcal{L}_{Reg}$ . For the reconstruction loss, we use LPIPS loss:", + "bbox": [ + 498, + 325, + 888, + 385 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {R e c}} \\left(G _ {\\theta} \\left(\\boldsymbol {x} _ {L}\\right), \\boldsymbol {x} _ {H}\\right) = \\mathcal {L} _ {\\text {L P I P S}} \\left(G _ {\\theta} \\left(\\boldsymbol {x} _ {L}\\right), \\boldsymbol {x} _ {H}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 390, + 890, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the regularization loss, we use our TSD loss Eq. (5). Therefore, the overall training objective for the Student Model $G_{\\theta}$ is:", + "bbox": [ + 498, + 409, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S t u}} = \\mathcal {L} _ {\\text {R e c}} + \\gamma \\mathcal {L} _ {\\text {R e g}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 458, + 890, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is a weighting scalar.", + "bbox": [ + 498, + 478, + 697, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LoRA Model $\\epsilon_{\\phi}$ . As stipulated by VSD, the fine-tuned replica $\\epsilon_{\\phi}$ must be trainable, with its training objective being:", + "bbox": [ + 498, + 492, + 890, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D i f f} (\\hat {\\boldsymbol {z}}, c _ {y}) = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left\\| \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\hat {\\boldsymbol {z}} _ {t}; t, c _ {y}\\right) - \\boldsymbol {\\epsilon} ^ {\\prime} \\right\\| ^ {2} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 542, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon^{\\prime}$ serves as the training target for the denoising network, representing Gaussian noise in the context of DDPM, and a gradient towards HQ for flow matching.", + "bbox": [ + 498, + 566, + 890, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 626, + 632, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 650, + 705, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Datasets. We adopt DIV2K [1], Flickr2K [42], LSDIR [27] and first 10K face images from FFHQ [20] for training. We use the same degradation pipeline as Real-ESRGAN [46] to synthesize LR-HR pairs.", + "bbox": [ + 498, + 672, + 890, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test Datasets. We evaluated our model on synthetic DIV2K-Val [1] dataset and two real-world datasets, RealSR [4] and DRealSR [51]. The real-world datasets comprised 128x128 LQ and 512x512 HQ image pairs.", + "bbox": [ + 498, + 734, + 890, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. For evaluating our method, we apply both full-reference and no-reference metrics. Full-reference metrics include PSNR and SSIM [49] (calculated on the Y channel in YCbCr space) for fidelity, LPIPS [65] and DISTS [7] for perceptual quality, and FID [14] for distribution comparison. No-reference metrics include NIQE [63], MANIQA[55], MUSIQ [22] and CLIPIQA [43].", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23178", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2546b7db10ec64dd974cd4a31954a4ef3925ad43b4284536dbf560b6961844e5.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with the state-of-the-art one-step methods across both synthetic and real-world benchmarks. The number of diffusion inference steps is indicated by 's'. The best results of each metric are highlighted in red." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
DRealSROSEDiff-1s27.920.78360.29680.2162135.516.447164.690.58980.6958
AddSR-1s27.770.77220.31960.2242150.186.932160.850.54900.6188
SinSR-1s28.380.74970.36690.2484172.726.960655.030.49040.6412
Ours-1s27.770.75590.29670.2136134.985.913166.620.58740.7344
RealSROSEDiff-1s25.150.73410.29200.2128123.485.647169.100.63260.6687
AddSR-1s24.790.70770.30910.2191132.055.544066.180.60980.5722
SinSR-1s26.270.73510.32170.2341137.596.296460.760.54180.6163
Ours-1s24.810.71720.27430.2104114.455.129871.190.63470.7160
DIV2K-ValOSEDiff-1s23.720.61090.29420.197526.344.708967.960.61310.6681
AddSR-1s23.260.59020.36230.212329.684.761063.390.56570.5734
SinSR-1s24.410.60180.32620.206835.555.998162.950.54300.6501
Ours-1s23.020.58080.26730.182129.164.324471.690.61920.7416
", + "bbox": [ + 76, + 119, + 893, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared Methods. We categorize the test models into two groups: single-step and multi-step inference. The single-step inference diffusion models include SinSR [48], AddSR [54], and OSEDiff [52]. The multi-step inference diffusion models comprise StableSR [44], ResShift [61], PASD [56], DiffBIR [29], SeeSR [53], SUPIR [60], and AddSR [54]. Specifically, for AddSR, we have conducted comparisons between its single-step and four-step models. GAN-based Real-ISR methods [5, 28, 46, 62] are detailed in the supplementary material.", + "bbox": [ + 75, + 345, + 472, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. All models mentioned are initialized from the Teacher Model (SD3 [10] in our work). We only train vae encoder and denoising network in the Student Model, frozen the vae decoder in order to preserve vae's prior [24]. We utilize the default prompt for the Student Model, and prompts extracted from HQ for the Teacher and LoRA models when training. We use the AdamW optimizer [30] with a learning rate of 5e-6 for the Student Model and 1e-6 for the LoRA Model, and set rank of LoRA to 64 for both two models. The training process took roughly 120h, utilizing 8 NVIDIA V100 GPUs with a batch size of 16.", + "bbox": [ + 75, + 498, + 470, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 76, + 679, + 387, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparisons. Tab. 1 show the quantitative comparison of our method with single-step diffusion models on three datasets. Ours achieves the best results on most evaluation metrics. SinSR and AddSR, as distilled versions of previous multi-step super-resolution methods, reduce inference steps but experience a corresponding decrease in performance metrics. OSEDiff introduces the VSD loss from 3D generation tasks into Real-ISR without fully addressing the substantial differences between the two domains. Therefore, the no-reference metrics for image quality are not very satisfied. In contrast, our proposed TSD-SR, tailored for Real-ISR, outperform other single-step models in terms of the vast majority of key metrics.", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 show the quantitative comparison with multi-step models. We can draw the following conclusions: (1) TSD-SR demonstrates significant advantages over competing methods in terms of LPIPS, DISTS, and NIQE metrics. Additionally, it achieves performance that surpasses most multi-step models on FID, MUSIQ, and CLIPIQA. (2) DiffBIR, SeeSR, PASD, and AddSR exhibit better performance on the MANIQA metric, which may be attributed to the fact that multi-step models have more denoising iterations to produce rich details. (3) ResShift stands out with the highest PSNR and SSIM scores, while StableSR also shows notable performance regarding DISTS and FID metrics. However, both two models underperform in terms of no-reference metrics.", + "bbox": [ + 496, + 345, + 893, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we explain the lower PSNR and SSIM metrics in our experiments. Several works [54, 60] have found that these reconstruction metrics are not well-suited for evaluating Real-ISR tasks. When a model recovers better details, it leads to lower reconstruction metrics, indicating a trade-off [3, 32, 67]. Refer to the supplementary material for detailed visual comparisons and analysis.", + "bbox": [ + 496, + 561, + 893, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. Fig. 7 present visual comparisons of different Real-ISR methods. As shown in multi-step methods' results, SeeSR uses degradation-aware semantic cues to leverage image generation priors, but it sometimes produces over-smooth textures. SUPIR demonstrates a notably robust generative capacity, however, the overproduction of extraneous details will lead to results that lack naturalism in image restoration (e.g. Adding excessive wrinkles at the corners of a young girl's eyes). Under more realistic degradation conditions, PASD finds it challenging to restore the corresponding content, revealing a deficiency in the model's robustness capabilities. Among single-step methods, SinSR introduces artifacts, possibly because its distilled pre-trained model is trained from scratch, lacking sufficient real-world priors, which results in unsatisfac", + "bbox": [ + 496, + 674, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23179", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/69eb288ba70f26d92a7c12a267e37d3eaa3d4ee3a86fb302b44eabe322651db4.jpg", + "table_caption": [ + "Table 2. Quantitative comparison with state-of-the-art multi-step methods across both synthetic and real-world benchmarks. The number of diffusion inference steps is indicated by 's'. The best and second best results of each metric are highlighted in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
DRealSRStableSR-200s28.040.74540.32790.2272144.156.599958.530.56030.6250
DiffBIR-50s25.930.65250.45180.2761177.046.232465.660.62960.6860
SeeSR-50s28.140.77120.31410.2297146.956.463264.740.60220.6893
SUPIR-50s25.090.64600.42430.2795169.487.391858.790.54710.6749
PASD-20s27.790.74950.35790.2524171.036.766163.230.59190.6242
ResShift-15s28.690.78740.35250.2541176.777.876252.400.47560.5413
AddSR-4s26.720.71240.39820.2711164.127.668966.330.62570.7226
Ours-1s27.770.75590.29670.2136134.985.913166.620.58740.7344
RealSRStableSR-200s24.620.70410.30700.2156128.545.781765.480.62230.6198
DiffBIR-50s24.240.66500.34690.2300134.565.493268.350.65440.6961
SeeSR-50s25.210.72160.30030.2218125.105.397869.690.64430.6671
SUPIR-50s23.650.66200.35410.2488130.386.109962.090.57800.6707
PASD-20s25.680.72730.31440.2304134.185.761668.330.63230.5783
ResShift-15s26.390.75670.31580.2432149.596.874660.220.54190.5496
AddSR-4s23.330.64000.39250.2626154.225.895971.490.68260.7225
Ours-1s24.810.71720.27430.2104114.455.129871.190.63470.7160
DIV2K-ValStableSR-200s23.270.57220.31110.204624.954.773765.780.61640.6753
DiffBIR-50s23.130.57170.34690.210833.934.605668.540.63600.7125
SeeSR-50s23.730.60570.31980.195325.814.832268.490.61980.6899
SUPIR-50s22.130.52790.39190.231231.405.676763.860.59030.7146
PASD-20s24.000.60410.37790.230539.124.858767.360.61210.6327
ResShift-15s24.710.62340.34730.225342.016.361560.630.52830.5962
AddSR-4s22.160.52800.40530.236035.415.258470.990.65960.7593
Ours-1s23.020.58080.26730.182129.164.324471.690.61920.7416
", + "bbox": [ + 76, + 121, + 893, + 474 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd8eefaf64f2bdfcd3807b590e714a80bc3907309157ce3c865b6a8e976c9f7c.jpg", + "table_caption": [ + "Table 3. Comparison of computational complexity across different DMs-based methods. The performance of each method is measured on an A100 GPU, using input images sized at $512 \\times 512$ pixels. We have disregarded the loading time for model weights and data." + ], + "table_footnote": [], + "table_body": "
StableSRDiffBIRSeeSRSUPIRPASDResShiftAddSRAddSRSinSROSEDiffOurs
Inference Step200505050201541111
Inference Time12.41517.96375.816716.87044.84410.75461.01990.50430.14240.16500.1362
", + "bbox": [ + 78, + 516, + 892, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tory image restoration capabilities. AddSR produces oversmoothed results when using its 1-step model. OSEDiff offers improved restoration effects compared to SinSR and AddSR, yet it may fall short in terms of authenticity and naturalness when it comes to detail recovery. In contrast, our method effectively generates rich textures and realistic details with enhanced sharpness and contrast. Additional visual comparisons and results are provided in the supplementary material.", + "bbox": [ + 75, + 597, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Complexity Comparisons We assess the complexity of the state-of-the-art DM-based Real-ISR models as detailed in Tab. 3, focusing on inference time. The performance of each method is benchmarked on an A100 GPU with input images sized at $512 \\times 512$ pixels. We have disregarded the loading time for model weights and data, and the main computation time includes: (1) text extraction time when a text extractor is used; (2) text encoder computation time if needed; (3) VAE encoding and decoding time; (4) denoising network execution time. It is evident that TSD-SR has a substantial advantage in inference time compared with", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "multi-step models. Specifically, TSD-SR is more than 120 times faster than SUPIR, 90 times faster than StableSR, about 50 times faster than DiffBIR, more than 40 times faster than SeeSR, more than 35 times faster than PASD, and 4 times faster than ResShift. When compared with existing one-step models, our model boasts the fastest inference times. This is because we directly denoise from LQ data and use a fixed prompt.", + "bbox": [ + 496, + 597, + 890, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. User Study", + "text_level": 1, + "bbox": [ + 500, + 726, + 622, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a user study comparing our method with three other diffusion-based one-step super-resolution methods. To make the evaluation more comprehensive, we selected images from five categories—human faces, buildings, animals, vegetation and characters. A total of 50 participants were engaged in the voting process. We guided participants to evaluate the best restoration results based on the similarity to HQ image, structural similarity to the LQ image, and realism of textures and details. The results shown in the Fig. 8 indicate that our method receives a $69.2\\%$ approval", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23180", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe86e8eee002cf877a96d586a8a3857aed4c08c99aa54c190f5b5fff477507e9.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 300, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e022902d11bee8f3cc7b6daf54e816099a84ecec0461b2c3391a98c4ab84827c.jpg", + "image_caption": [ + "Zoomed LR" + ], + "image_footnote": [], + "bbox": [ + 303, + 90, + 447, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29755036bfcbd3f83680a23116c0dcbf7eb9139d41d75118d907060c3e3ed45d.jpg", + "image_caption": [ + "SeeSR-50s" + ], + "image_footnote": [], + "bbox": [ + 452, + 90, + 593, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/724cc0fbdcf19a5d9301f4d3c96ad6eb56cdd9aff8cca5910905a46df375b324.jpg", + "image_caption": [ + "SUPIR-50s" + ], + "image_footnote": [], + "bbox": [ + 598, + 90, + 741, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ca7be1e537bf4cb4134a6edca8f7c880544686fe04fd2fba33bb0fb4f7cd7b5.jpg", + "image_caption": [ + "PASD-20s" + ], + "image_footnote": [], + "bbox": [ + 746, + 90, + 888, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db129735dea7555c006e0682c6e383ce8a935e78b4e0b39f1a0f6c61536d46c7.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 81, + 268, + 299, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb42b0e29977ac9962f67489b65fb3a25a986c136b11684fe6fc6e5575d6e4a1.jpg", + "image_caption": [ + "SinSR-1s" + ], + "image_footnote": [], + "bbox": [ + 303, + 178, + 447, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b71ce843efcf03611aca132fa8446e5810df1aec80e367e257c43cfb6b64ccf.jpg", + "image_caption": [ + "AddSR-1s" + ], + "image_footnote": [], + "bbox": [ + 452, + 178, + 593, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/22930d09d16fd6160e3365cabf60ad34209152b69b7b0717575dfeafeb046044.jpg", + "image_caption": [ + "OSEDiff-1s" + ], + "image_footnote": [], + "bbox": [ + 598, + 178, + 741, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6751e74dcad40df56fea2d31e3373edb9434e45be3d4c1e293811fad8ff594d8.jpg", + "image_caption": [ + "Ours-1s" + ], + "image_footnote": [], + "bbox": [ + 746, + 178, + 888, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a17b5e00b02004a98fb29cedd24f72e34b827b63b135b57b39ff18c0599f3c63.jpg", + "image_caption": [ + "Figure 7. Visual comparisons of different Real-ISR methods. Please zoom in for a better view.", + "Figure 8. The results of our user study. Left: Category-based user preference radar chart. Our model has won the most user favor across all categories. Right: User preference pie chart. Our approach has garnered a $69.2\\%$ user satisfaction rating." + ], + "image_footnote": [], + "bbox": [ + 81, + 436, + 299, + 571 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46fc5e3c8340a63b9244833eb0082e595aa016f960d93afd8bb38a3fdb0207d1.jpg", + "image_caption": [ + "Zoomed LR" + ], + "image_footnote": [], + "bbox": [ + 303, + 268, + 446, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb947c9f6cb1c0d1cd9ed3183fe61f2291afc5fd1fb4d3ca5dcf7ffe6d107431.jpg", + "image_caption": [ + "SeeSR-50s" + ], + "image_footnote": [], + "bbox": [ + 452, + 268, + 593, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f20b4635ecbab927ca453b53f34a847e718a86647422f9d4b767c74614a36ed4.jpg", + "image_caption": [ + "SUPIR-50s" + ], + "image_footnote": [], + "bbox": [ + 598, + 268, + 741, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66a01f6b27aa7445518e436fefca1e27781a12d9ba01655fba9656efcfadb050.jpg", + "image_caption": [ + "PASD-20s" + ], + "image_footnote": [], + "bbox": [ + 746, + 268, + 888, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd1065f8c6b713bb488db295345507232575856b31f4b66002404d7d1e1c1688.jpg", + "image_caption": [ + "SinSR-1s" + ], + "image_footnote": [], + "bbox": [ + 303, + 357, + 446, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f5629488e3008b2e0c386e0d231fff50a34cf4906df7c13eb762bf1cf6971a6.jpg", + "image_caption": [ + "AddSR-1s" + ], + "image_footnote": [], + "bbox": [ + 452, + 357, + 593, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5bf3f251f012bc7b8fc7a1fa65974609daf6e8ec40be3c4d6f82637b47642561.jpg", + "image_caption": [ + "OSEDiff-1s" + ], + "image_footnote": [], + "bbox": [ + 598, + 357, + 741, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/edb0c465d072a5e7f438e452f5771433c626607d9a238543b31eaeb4e87f9212.jpg", + "image_caption": [ + "Ours-1s" + ], + "image_footnote": [], + "bbox": [ + 746, + 357, + 888, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 81, + 479, + 267, + 571 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/273a2e90fc2f112a8ac481c8051671bfe3cdf262cd1b18e35987f4de04a0b16f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 481, + 467, + 571 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/28e0cb320ff0984c22d219331b5f2d1019b5cd9b8a0b1d890f7ac3a9804f54f1.jpg", + "table_caption": [ + "Table 4. Ablation study of Target Score Matching loss and Distribution-Aware Sampling Module." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodDISTS↓MUSIQ↑MANIQA↑CLIPIQA↑
DRealSRw/o TSM0.232763.900.57490.6958
w/o DASM0.231163.560.58120.7123
Full0.213666.620.58740.7344
RealSRw/o TSM0.239768.560.63380.6987
w/o DASM0.227369.050.62730.7031
Full0.210471.190.63470.7160
", + "bbox": [ + 78, + 691, + 462, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rate from users. We scored $57.60\\%$ in Animals, $70.00\\%$ in Buildings, $68.80\\%$ in Human Faces, $65.20\\%$ in Vegetation, and $84.40\\%$ in Characters, surpassing other models.", + "bbox": [ + 75, + 792, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 847, + 228, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness of TSM loss and DASM, we conduct ablation studies by removing them separately in", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our experiments. We select DISTS, MUSIQ, MANIQA and CLIPIQA metrics for comparison, as these are critical for image quality assessment. The results are presented in Tab. 4. We draw the following conclusions: (1) The absence of TSM loss and DASM affects the metrics DISTS, MUSIQ, MANIQA and CLIPIQA. (2) The lack of TSM leads to a significant decrease in DISTS, MUSIQ and CLIPIQA metrics, possibly due to the unreliable directions in VSD causing unrealistic generations. (3) The absence of DASM results in a decline in MUSIQ and CLIPIQA, possibly due to the suboptimal optimization of details.", + "bbox": [ + 496, + 470, + 890, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 500, + 651, + 748, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose TSD-SR, an effective and one-step model based on diffusion prior for Real-ISR. TSD-SR utilizes the TSD to enhance the realism of images generated by the distillation model. And it leverages DASM to sample distribution-based samples and accumulate their gradients to enhance the recovery of details. Our experiments have demonstrated that TSD-SR outperforms existing one-step Real-ISR models in both performance and inference speed.", + "bbox": [ + 496, + 676, + 890, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "limitations. Although our model boasts excellent inference speed and restoration performance, it still has a large number of model parameters compared to the past GAN/non-diffusion approaches. In the future, we plan to employ pruning or quantization methods to compress the model parameters, striving for a lightweight and efficient Real-ISR model.", + "bbox": [ + 496, + 797, + 890, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23181", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by National Natural Science Foundation of China (62293554, U2336212), Zhejiang Provincial Natural Science Foundation of China (LD24F020007), National Key R&D Program of China (SQ2023AAA01005), Ningbo Innovation \"Yongjiang 2035\" Key Research and Development Programme (2024Z292), and Young Elite Scientists Sponsorship Program by CAST (2023QNRC001).", + "bbox": [ + 76, + 90, + 472, + 203 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 233, + 173, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5", + "[2] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 1, 2", + "[3] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 6", + "[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 5", + "[5] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Real-world blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1329-1338, 2022. 6", + "[6] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1", + "[7] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 2, 5", + "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 1", + "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1", + "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis." + ], + "bbox": [ + 78, + 258, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Forty-first International Conference on Machine Learning, 2024. 2, 6", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[12] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 2", + "[13] Xiangyu He and Jian Cheng. Revisiting 11 loss in superresolution: a probabilistic view and beyond. arXiv preprint arXiv:2201.10084, 2022. 4", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[15] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[18] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2", + "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5", + "[21] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 1", + "[22] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5", + "[23] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 1", + "[24] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 6", + "[25] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1): 79-86, 1951. 3", + "[26] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23182", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition, pages 4681-4690, 2017. 1, 2", + "[27] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775-1787, 2023. 5", + "[28] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 6", + "[29] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 1, 2, 6", + "[30] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[31] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2", + "[32] Xiaotong Luo, Yuan Xie, Yanyun Qu, and Yun Fu. Skipdiff: Adaptive skip diffusion model for high-fidelity perceptual image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4017-4025, 2024. 6", + "[33] Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 1", + "[34] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7807-7816, 2024. 2", + "[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[36] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3", + "[37] Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 1", + "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2", + "[40] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1", + "[42] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5", + "[43] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5", + "[44] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1–21, 2024. 2, 6", + "[45] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 3", + "[46] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2, 3, 5, 6", + "[47] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490, 2022. 1", + "[48] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2, 6", + "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[51] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 5", + "[52] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 4, 6", + "[53] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 1, 2, 6", + "[54] Rui Xie, Ying Tai, Kai Zhang, Zhenyu Zhang, Jun Zhou, and Jian Yang. Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 2, 6", + "[55] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 5", + "[56] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 1, 2, 6", + "[57] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4133-4141, 2017. 2", + "[58] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2", + "[59] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2", + "[60] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 1, 2, 6", + "[61] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 2, 6", + "[62] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 1, 2, 3, 6", + "[63] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 5", + "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 3, 5" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. In European conference on computer vision, pages 649-667. Springer, 2022. 1", + "[67] Yuehan Zhang, Bo Ji, Jia Hao, and Angela Yao. Perception-distortion balanced admm optimization for single-image super-resolution. In European Conference on Computer Vision, pages 108-125. Springer, 2022. 6" + ], + "bbox": [ + 501, + 92, + 890, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23184", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_model.json b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c82fce07c80d7aa572ca720ee86e8a7ef6ff5a91 --- /dev/null +++ b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_model.json @@ -0,0 +1,2763 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.861, + 0.177 + ], + "angle": 0, + "content": "TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.181, + 0.202, + 0.761, + 0.242 + ], + "angle": 0, + "content": "Linwei Dong\\(^{1,2*}\\) Qingnan Fan\\(^{2*}\\) Yihong Guo\\(^{1}\\) Zhonghao Wang\\(^{3}\\) Qi Zhang\\(^{2}\\) Jinwei Chen\\(^{2}\\) Yawei Luo\\(^{1\\dagger}\\) Changqing Zou\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.256, + 0.738, + 0.293 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University \\(^{2}\\)Vivo Mobile Communication Co. Ltd \\(^{3}\\)University of Chinese Academy of Sciences \\(^{4}\\)Zhejiang Lab" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.678 + ], + "angle": 0, + "content": "Pre-trained text-to-image diffusion models are increasingly applied to real-world image super-resolution (RealISR) task. Given the iterative refinement nature of diffusion models, most existing approaches are computationally expensive. While methods such as SinSR and OSEDiff have emerged to condense inference steps via distillation, their performance in image restoration or details recovery is not satisfied. To address this, we propose TSD-SR, a novel distillation framework specifically designed for real-world image super-resolution, aiming to construct an efficient and effective one-step model. We first introduce the Target Score Distillation, which leverages the priors of diffusion models and real image references to achieve more realistic image restoration. Secondly, we propose a Distribution-Aware Sampling Module to make detail-oriented gradients more readily accessible, addressing the challenge of recovering fine details. Extensive experiments demonstrate that our TSD-SR has superior restoration results (most of the metrics perform the best) and the fastest inference speed (e.g. 40 times faster than SeeSR) compared to the past Real-ISR approaches based on pre-trained diffusion priors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.703, + 0.21, + 0.719 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.47, + 0.85 + ], + "angle": 0, + "content": "Image super-resolution (ISR) [8, 9, 23, 26] aims to transform low-quality (LQ) images, which have experienced noise or blur, into clear high-quality (HQ) images. Differing from traditional ISR [6, 66], which assumes a known degradation process, real-world image super-resolution (RealISR) [46, 62] is designed to enhance real-world images that have suffered from complex and unknown degradations, thereby offering greater practical utility." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.85, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Generative models, particularly Generative Adversarial Networks (GANs) [11, 33, 37] and Diffusion Models (DMs)" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.351, + 0.859, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.555, + 0.895, + 0.639 + ], + "angle": 0, + "content": "Figure 1. Performance and efficiency comparison among RealISR methods. TSD-SR stands out for achieving high-quality restoration with the fastest speed among diffusion-based models. In contrast, existing models prioritize either speed or restoration performance. The performance of each method is benchmarked on an A100 GPU with the DRealSR dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.868 + ], + "angle": 0, + "content": "[17, 39, 41], have shown remarkable power in handling Real-ISR task. GAN-based methods leverage adversarial training, toggling between the generator and discriminator to produce realistic images. While GANs are capable of one-step inference, they are commonly hampered by issues like mode collapse and training instability [2]. Recently, Diffusion Models (DMs) have shown impressive performance in the realm of image generation [21, 47]. Their robust priors empower them to produce more realistic images with richer details than GAN-based methods [39, 41]. Some researchers [29, 53, 56, 60] have successfully leveraged pre-trained DMs for Real-ISR task. However, due to the iterative denoising nature of diffusion models [17], the Real-ISR process is computationally expensive." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To achieve an efficient and one-step network akin to GANs, several pioneering methods that condense the iter-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.346, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23174" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.348 + ], + "angle": 0, + "content": "ations of diffusion models through distillation [12, 15, 18, 57] have been proposed [48, 52, 54]. Among these works, OSEDiff [52] introduced the Variational Score Distillation (VSD) loss [50] to Real-ISR task, achieving state-of-the-art (SOTA) one-step performance by leveraging prior knowledge from pre-trained models. Despite these advancements, our investigation has revealed two critical limitations associated with VSD in Real-ISR applications. (1) Unreliable gradient direction. VSD relies on a Teacher Model to provide a \"true gradient direction.\" However, this guidance proves unreliable in scenarios where initial ISR outputs are suboptimal. (2) Insufficient detail recovery. The VSD loss exhibits notable variation across different timesteps, and the uniform sampling strategy for \\(t\\) poses challenges in aligning the score function with detailed texture recovery requirements. These findings highlight the need for enhanced approaches to address these issues effectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.59 + ], + "angle": 0, + "content": "In this paper, we propose a novel method called TSD-SR to distill a multi-step Text-to-Image (T2I) DMs [10, 38, 39] into an effective one-step diffusion model tailored for the Real-ISR task. Specifically, TSD-SR consists of two components: Target Score Distillation (TSD) and Distribution-Aware Sampling Module (DASM). TSD incorporates our new proposed Target Score Matching (TSM) loss to compensate for the limitation of VSD loss. This significant score loss leverages HQ data to provide a reliable optimization trajectory for the distillation process. It effectively reduces the visual artifacts caused by deviant predictions from the Teacher Model. DASM is crafted to bolster detail recovery by strategically sampling low-noise samples that are distribution-based during training. This approach effectively allocates more optimization to early timesteps in a single iteration, enhancing the recovery of details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.473, + 0.681 + ], + "angle": 0, + "content": "Experiments on popular benchmarks demonstrate that TSD-SR achieves superior restoration performance (most of the metrics perform the best) and high efficiency (the fastest inference speed, 40 times faster than SeeSR) compared to the state-of-the-art Real-ISR methods based on pre-trained DMs, while requiring only a single inference step." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.681, + 0.442, + 0.695 + ], + "angle": 0, + "content": "Our main contribution can be summarized as threefold:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.696, + 0.469, + 0.725 + ], + "angle": 0, + "content": "- We propose a novel method called TSD-SR to achieve one-step DMs distillation for the Real-ISR task." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.726, + 0.469, + 0.771 + ], + "angle": 0, + "content": "- We introduce Target Score Distillation (TSD) to provide a reliable gradient to enhance the realism of the Real-ISR model's outputs." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.771, + 0.469, + 0.815 + ], + "angle": 0, + "content": "- We design a Distribution-Aware Sampling Module (DASM) specifically tailored to enhance the capability of detail restoration." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.696, + 0.469, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.219, + 0.846 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "GAN-based Real-ISR. Since SRGAN [26] first applied GAN to ISR, it has effectively enhanced visual quality by combining adversarial loss with perceptual loss [7," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[65]. Subsequently, ESRGAN [45] introduced Residual-in-Residual Dense Block and a relativistic average discriminator, further improving detail restoration. Methods like BSRGAN [62] and Real-ESRGAN [46] simulate complex real-world degradation processes, achieving ISR under unknown degradation conditions, which enhances the model's generalization ability. Although GAN-based methods are capable of adding more realistic details to images, they suffer from training instability and mode collapse [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.893, + 0.609 + ], + "angle": 0, + "content": "Multi-step Diffusion-based Real-ISR. Some researches [29, 44, 53, 56, 60] in recent years have utilized the powerful image priors in pre-trained T2I diffusion models [35, 39, 64] for Real-SR tasks and achieved promising results. For example, StableSR [44] balances fidelity and perceptual quality by fine-tuning the time-aware encoder and employing controllable feature wrapping. DiffBiR [29] first processes the LR image through a reconstruction network and then uses the Stable Diffusion (SD) model [39] to supplement the details. SeeSR [53] attempts to better stimulate the generative power of the SD model by extracting the semantic information in the image as a conditional guide. PASD [56] introduces a pixel-aware cross attention module to enable the diffusion model to perceive the local structure of the image at the pixel level, while using a degradation removal module to extract degradation insensitive features to guide the diffusion process along with high-level information from the image. SUPIR [60] achieves a generative and fidelity capability using negative cues [16] as well as restoration-guided sampling, while using a larger pretraining model with a larger dataset to enhance the model capability. However, all of these methods are limited by the multi-step denoising of the diffusion model, which requires 20-50 iterations in inference, resulting in an inference time that lags far behind that of GAN-based methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "One-step Diffusion-based Real-ISR. Recently, there has been a surge of interest within the academic community in one-step distillation techniques [31, 34, 40, 58, 59] for diffusion-based Real-ISR task. SinSR [48] leverages consistency preserving distillation to condense the inference steps of ResShift [61] into a single step, yet the generalization of ResShift and SinSR is constrained due to the absence of large-scale data training. AddSR [54] introduces the adversarial diffusion distillation (ADD) [40] to RealISR tasks, resulting in a comparatively effective four-step model. However, this method has a propensity to produce excessive and unnatural image details. OSEDiff [52] directly uses LQ images as the beginning of the diffusion process, and employs VSD loss [50] as a regularization technique to condense a multi-step pre-trained T2I model into a one-step Real-ISR model. However, due to the incorporation of alternating training strategies, OSEDiff may initially tend towards unreliable optimization directions, which may lead to the visual artifacts." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23175" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.419, + 0.893, + 0.492 + ], + "angle": 0, + "content": "Figure 2. Pipeline overview. We train one-step Student Model \\( G_{\\theta} \\) to map the low-quality image \\( x_{L} \\) into a more realistic one. The noisy latent \\( \\hat{z}_{t} \\) and \\( z_{t} \\) sampled by DASM (Details can be found in Fig. 6.) will be fed into both the pre-trained Teacher and LoRA Model to produce the Variational Score loss. Subsequently, two Models' predictions on \\( z_{t} \\) yield Target Score loss. Their weighted forms, namely TSD (red flow), in conjunction with pixel-space reconstruction loss (green flow), are leveraged to update the Student Model. After updating the Student Model, we employ the diffusion loss (blue flow) to alternately update the LoRA Model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.212, + 0.513 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.524, + 0.218, + 0.539 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Problem Formulation. ISR problem aims to reconstruct a HQ image \\( x_{H} \\) from a LQ input \\( x_{L} \\) by training a parameterized ISR model \\( G_{\\theta} \\) on a dataset \\( \\mathcal{D} = \\{(x_{L},x_{H})_{i = 1}^{N}\\} \\), where \\( N \\) represents the number of image pairs. It is mathematically formulated as minimizing the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.631, + 0.47, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta^ {*} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x _ {L}, x _ {H}) \\sim \\mathcal {D}} \\left[ \\mathcal {L} _ {R e c} \\left(G _ {\\theta} \\left(x _ {L}\\right), x _ {H}\\right) \\right. \\tag {1} \\\\ \\left. \\left. + \\lambda \\mathcal {L} _ {R e g} \\left(q _ {\\theta} \\left(\\hat {x} _ {H}\\right), p \\left(x _ {H}\\right)\\right) \\right] \\right. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.469, + 0.784 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_{Rec}\\) denotes the reconstruction loss, commonly measured using distance metrics like \\(L_{2}\\) or LPIPS [65]. The regularization term \\(\\mathcal{L}_{Reg}\\) enhances the realism and generalization of the ISR model's outputs. This objective can be understood as aligning the ISR output \\(\\hat{x}_H\\) 's distribution, \\(q_{\\theta}(\\hat{x}_H)\\), with the true distribution \\(p(x_H)\\) by minimizing the KL-divergence [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.787, + 0.469, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathcal {D} _ {\\mathrm {K L}} \\left(q _ {\\theta} \\left(\\hat {x} _ {H}\\right) \\| p \\left(x _ {H}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "While several studies [45, 46, 62] have employed adversarial loss to optimize this objective, they often encounter issues like mode collapse and training instability. Recent work [52] achieved the state-of-the-art results using Variational Score Distillation (VSD) as the regularization loss to minimizing this objective, which inspires our research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.558 + ], + "angle": 0, + "content": "Variational Score Distillation. Variational Score Distillation (VSD) [50] was initially introduced for text-to-3D generation by distilling a pre-trained text-to-image diffusion model to optimize a single 3D representation [36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.892, + 0.634 + ], + "angle": 0, + "content": "In the VSD framework, a pre-trained diffusion model, represented as \\(\\epsilon_{\\psi}\\), and its trainable (LoRA [19]) replica \\(\\epsilon_{\\phi}\\) are used to regularize the generator network \\(G_{\\theta}\\). As outlined in ProlificDreamer [50], the gradient with respect to the generator parameters \\(\\theta\\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.64, + 0.891, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {V S D}} (\\hat {\\boldsymbol {z}}, c _ {y}) \\\\ = \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\left(\\boldsymbol {\\epsilon} _ {\\psi} (\\hat {\\boldsymbol {z}} _ {t}; t, c _ {y}) - \\boldsymbol {\\epsilon} _ {\\phi} (\\hat {\\boldsymbol {z}} _ {t}; t, c _ {y})\\right) \\frac {\\partial \\hat {\\boldsymbol {z}}}{\\partial \\boldsymbol {\\theta}} \\right] \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.777 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{z}}_t = \\alpha_t\\hat{\\pmb{z}} +\\sigma_t\\pmb {\\epsilon}\\) is the noisy input, \\(\\hat{z}\\) is the latent outputted by the generator network \\(G_{\\theta}\\), \\(\\pmb{\\epsilon}\\) is a gaussian noise, and \\(\\alpha_{t},\\sigma_{t}\\) are the noise-data scaling constants. \\(c_{y}\\) is a text embedding corresponding to a caption that describes the input image, and \\(w(t)\\) is a time-varying weighting function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.698, + 0.801 + ], + "angle": 0, + "content": "3.2. Overview of TSD-SR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "As depicted in Fig. 2, our goal is to distill a given pretrained T2I DMs into a fast one-step Student Model \\( G_{\\theta} \\) with Teacher Model \\( \\epsilon_{\\psi} \\) and trainable LoRA Model \\( \\epsilon_{\\phi} \\). We denote the latent outputs of the distilled model as \\( \\hat{z}_0 \\), and HQ latent representations as \\( z_0 \\). Both \\( \\hat{z}_0 \\) and \\( z_0 \\) will pass through our Distribution-Aware Sampling Module (DASM)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23176" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.271, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Figure 3. A visual comparison of the gradient direction. We set the timestep \\( t \\) to 100 and calculated the cosine similarity between the predicted directions from the Teacher Model and the true direction (towards the HQ data). The prediction direction for \\( z_{t} \\) closely matches the true direction, but not for \\( \\hat{z}_{t} \\), suggesting that suboptimal samples may lead to directional deviations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.468, + 0.503 + ], + "angle": 0, + "content": "to obtain distribution-based samples \\(\\hat{z}_t\\) and \\(z_t\\) (Sec. 3.4). We train \\(G_{\\theta}\\) by minimizing the two losses: a reconstruction loss in pixel space to compare the model outputs against the ground truth, and a regularization loss (from Target Score Distillation) to enhances the realism (Sec. 3.3). After updating the Student Model, we update the LoRA Model with diffusion loss. Finally, in Sec. 3.5, we present an overview of all the losses encountered during the training phase." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.3, + 0.53 + ], + "angle": 0, + "content": "3.3. Target Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.468, + 0.794 + ], + "angle": 0, + "content": "Similar to [52], we introduce VSD loss into our work as a regularization loss to enhance the realism and generalization of the \\(G_{\\theta}\\) outputs. Upon reviewing VSD Eq. (3), \\(\\epsilon_{\\phi}(\\hat{z}_t;t,c_y)\\) represents the gradient direction of the optimization on \\(G_{\\theta}\\) noisy outputs \\(\\hat{z}_t\\) , whereas \\(\\epsilon_{\\psi}(\\hat{z}_t;t,c_y)\\) corresponds to the gradient direction on noisy HQ latent representations \\(z_{t}\\) . The overarching goal of model optimization is to harmonize the predictions from \\(\\hat{z}_t\\) with those from \\(z_{t}\\) Nonetheless, this strategy encounters hurdles, especially in the early training phase: the quality of synthetic latents \\(\\hat{z}_t\\) is not high enough for Teacher Model to give a precise prediction. As illustrated in Fig. 3, the Teacher Model finds it challenging to accurately predict the optimization direction for low-quality synthetic latents, with the cosine similarity to the ideal optimization direction only 0.2 (0.88 on HQ latents). This problem can lead to severe visual artifacts, as evident in Fig. 4(a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "A straightforward remedial measure is to employ a mean squared error (MSE) loss to align the synthetic latents with the ideal inputs of the Teacher Model, which are derived from HQ latents. However, as shown in Fig. 4(b), this approach has been observed to lead to over-smoothed results [13]. Our strategy, instead, is to align the predictions made by the Teacher Model on both synthetic and HQ" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.629, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.228, + 0.591, + 0.239 + ], + "angle": 0, + "content": "(a) Naive" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.089, + 0.761, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.228, + 0.721, + 0.239 + ], + "angle": 0, + "content": "(b) MSE" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.089, + 0.892, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.228, + 0.851, + 0.239 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.311 + ], + "angle": 0, + "content": "Figure 4. The visualization of different strategies. (a) The naive method introduces fake textures and fails to recover fine details. (b) MSE leads to over-smoothed generation result, lacking high-frequency information. (c) Our method offers the superior visual effects and fine textures." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.338, + 0.892, + 0.414 + ], + "angle": 0, + "content": "latents, thereby encouraging greater consistency between them. The core idea is that for samples drawn from the same distribution, the real scores predicted by the Teacher Model should be close to each other. We refer to this approach as Target Score Matching (TSM):" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.42, + 0.891, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {T S M}} (\\hat {\\mathbf {z}}, \\mathbf {z}, c _ {y}) \\\\ = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\psi} \\left(\\hat {\\mathbf {z}} _ {t}; t, c _ {y}\\right) - \\epsilon_ {\\psi} \\left(\\mathbf {z} _ {t}; t, c _ {y}\\right)\\right) \\frac {\\partial \\hat {\\mathbf {z}}}{\\partial \\theta} \\right] \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.478, + 0.892, + 0.659 + ], + "angle": 0, + "content": "where the expectation of the gradient is computed across all diffusion timesteps \\(t \\in \\{1, \\dots, T\\}\\) and \\(\\epsilon \\sim \\mathcal{N}(0, I)\\). Equation (4) encapsulates the optimization loss for our Target Score Matching. Upon examining it in conjunction with the Eq. (3), we notice that VSD utilizes the prediction residual between a Teacher and a LoRA Model to drive gradient backpropagation. Conversely, our TSM employs the synthetic and the HQ data to produce the gradients. By blending these two strategies with hyperparameter weighting \\(\\lambda\\) and \\(1 - \\lambda\\), we can construct our ultimate optimization loss, effectively unifying the strengths of both approaches to optimize the training process Eq. (5)." + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.665, + 0.892, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {T S D}} (\\hat {\\mathbf {z}}, \\mathbf {z}, c _ {y}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left[ \\epsilon_ {\\psi} \\left(\\hat {\\mathbf {z}} _ {t}; t, c _ {y}\\right) - \\right. \\right.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.692, + 0.892, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\left. \\left. \\epsilon_ {\\psi} (\\mathbf {z} _ {t}; t, c _ {y}) + \\lambda \\left(\\epsilon_ {\\psi} \\left(\\mathbf {z} _ {t}; t, c _ {y}\\right) - \\epsilon_ {\\phi} \\left(\\hat {\\mathbf {z}} _ {t}; t, c _ {y}\\right)\\right) \\right] \\frac {\\partial \\hat {\\mathbf {z}}}{\\partial \\theta} \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.727, + 0.892, + 0.818 + ], + "angle": 0, + "content": "where \\( w(t) \\) is time-aware weighting function tailored for Real-ISR. Other symbols are in accordance with those previously mentioned. By introducing the prediction of pretrained diffusion model on HQ latents, we have circumvented the issue of the model falling into the visual artifacts or over-smoothed problem, as illustrated in Fig. 4(c)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.825, + 0.848 + ], + "angle": 0, + "content": "3.4. Distribution-Aware Sampling Module" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In the VSD-based framework, we need to match the score functions predicted by the Teacher Model with those by the LoRA Model at different timesteps \\(t \\sim \\{0,1,\\dots,T\\}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23177" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.359, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.257, + 0.332, + 0.267 + ], + "angle": 0, + "content": "(a) Predicted Residual at Different Timesteps in VSD" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.09, + 0.468, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.168, + 0.438, + 0.177 + ], + "angle": 0, + "content": "(b) Stage 1" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.181, + 0.468, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.258, + 0.438, + 0.267 + ], + "angle": 0, + "content": "(c) Stage 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.47, + 0.34 + ], + "angle": 0, + "content": "Figure 5. (a) The prediction errors of the VSD loss at different timesteps. The error divergence is more pronounced in early timesteps than later. This phenomenon is observed throughout the optimization process. (b) the visualization of Stage 1 prediction error. (c) the visualization of Stage 2 prediction error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.553 + ], + "angle": 0, + "content": "However, for the Real-ISR problem, this matching ability at different \\( t \\) is unbalanced, as illustrated in Fig. 5(a). This phenomenon could be attributed to our leveraging of the low-frequency (LF) priors from LQ data, but losing guidance from high-frequency (HF) details. The noisy samples \\( \\hat{z}_t \\) possesses low-frequency priors (from LQ data). These priors are easily learned by the LoRA Model, leading to more similar predictions (Fig. 5(b)) during LF restoration (Stage 1). However, In the Stage 2, the lack of HF details in \\( \\hat{z}_t \\) causes the LoRA Model to poorly perceive details, resulting in divergent predictions (Fig. 5(c)). Our goal is to reduce this divergence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.469, + 0.768 + ], + "angle": 0, + "content": "Existing method, for each iteration, only matches the score function at a single timestep sample \\(\\hat{z}_t\\), with uniform sampling strategy of time \\(t\\). This leads to slow convergence or even difficulty in optimization during Stage 2, as the gradients of important timesteps are averaged out. To this end, we propose our Distribution-Aware Sampling Module (DASM). This module accumulates optimization gradients for earlier timestep samples in a single iteration, enabling the backpropagation of more gradients focused on detail optimization. As shown in Fig. 6, we first introduce noise into the synthetic latent representation \\(\\hat{z}_t = (1 - \\sigma_t)\\hat{z}_0 + \\sigma_t\\epsilon\\), where \\(\\sigma\\) is a weight factor and \\(\\epsilon\\) is a gaussian noise. Subsequently, we employ a LoRA Model to perform denoising, yielding the previous time noisy samples by Eq. (6):" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.774, + 0.469, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {\\boldsymbol {t} - 1} = \\hat {z} _ {\\boldsymbol {t}} + \\left(\\sigma_ {t - 1} - \\sigma_ {t}\\right) \\cdot \\epsilon_ {\\phi} \\left(\\hat {z} _ {\\boldsymbol {t}}; t, c _ {y}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\sigma_{t - 1}\\) and \\(\\sigma_t\\) are from the flow matching scheduler, and here LoRA Model has fit \\(\\hat{z}_0\\) 's distribution. Similarly, \\(z_{t - 1}\\) can be obtained by denoising on the Teacher Model. Ultimately, in a single iteration, the gradients from noise samples along its trajectory can be accumulated to update the Student Model. Since the obtained samples follow the diffusion sampling trajectory and are directed towards early" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.891, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.19, + 0.892, + 0.261 + ], + "angle": 0, + "content": "Figure 6. Illustration of DASM. Top: The naive method that add noise to samples. Bottom: Our DASM leverages the priors of diffusion models to sample noise that more closely aligns with the true sampling trajectory. The noisy samples can all serve as inputs to the subsequent network to generate gradient backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.28, + 0.89, + 0.295 + ], + "angle": 0, + "content": "timestep, we effectively optimize the divergence in Stage 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.304, + 0.681, + 0.32 + ], + "angle": 0, + "content": "3.5. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.327, + 0.89, + 0.386 + ], + "angle": 0, + "content": "We summarize all the losses we used in our framework. Student Model \\(G_{\\theta}\\). We train our Student Model with the reconstruction loss \\(\\mathcal{L}_{Rec}\\) and the regularization loss \\(\\mathcal{L}_{Reg}\\). For the reconstruction loss, we use LPIPS loss:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.391, + 0.891, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {R e c}} \\left(G _ {\\theta} \\left(\\boldsymbol {x} _ {L}\\right), \\boldsymbol {x} _ {H}\\right) = \\mathcal {L} _ {\\text {L P I P S}} \\left(G _ {\\theta} \\left(\\boldsymbol {x} _ {L}\\right), \\boldsymbol {x} _ {H}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.455 + ], + "angle": 0, + "content": "For the regularization loss, we use our TSD loss Eq. (5). Therefore, the overall training objective for the Student Model \\( G_{\\theta} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.459, + 0.891, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S t u}} = \\mathcal {L} _ {\\text {R e c}} + \\gamma \\mathcal {L} _ {\\text {R e g}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.479, + 0.699, + 0.492 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a weighting scalar." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.539 + ], + "angle": 0, + "content": "LoRA Model \\(\\epsilon_{\\phi}\\). As stipulated by VSD, the fine-tuned replica \\(\\epsilon_{\\phi}\\) must be trainable, with its training objective being:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.544, + 0.891, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D i f f} (\\hat {\\boldsymbol {z}}, c _ {y}) = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left\\| \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\hat {\\boldsymbol {z}} _ {t}; t, c _ {y}\\right) - \\boldsymbol {\\epsilon} ^ {\\prime} \\right\\| ^ {2} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.568, + 0.891, + 0.613 + ], + "angle": 0, + "content": "where \\(\\epsilon^{\\prime}\\) serves as the training target for the denoising network, representing Gaussian noise in the context of DDPM, and a gradient towards HQ for flow matching." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.627, + 0.633, + 0.643 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.706, + 0.667 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.674, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Training Datasets. We adopt DIV2K [1], Flickr2K [42], LSDIR [27] and first 10K face images from FFHQ [20] for training. We use the same degradation pipeline as Real-ESRGAN [46] to synthesize LR-HR pairs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.892, + 0.796 + ], + "angle": 0, + "content": "Test Datasets. We evaluated our model on synthetic DIV2K-Val [1] dataset and two real-world datasets, RealSR [4] and DRealSR [51]. The real-world datasets comprised 128x128 LQ and 512x512 HQ image pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. For evaluating our method, we apply both full-reference and no-reference metrics. Full-reference metrics include PSNR and SSIM [49] (calculated on the Y channel in YCbCr space) for fidelity, LPIPS [65] and DISTS [7] for perceptual quality, and FID [14] for distribution comparison. No-reference metrics include NIQE [63], MANIQA[55], MUSIQ [22] and CLIPIQA [43]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23178" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with the state-of-the-art one-step methods across both synthetic and real-world benchmarks. The number of diffusion inference steps is indicated by 's'. The best results of each metric are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.121, + 0.894, + 0.322 + ], + "angle": 0, + "content": "
DatasetsMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
DRealSROSEDiff-1s27.920.78360.29680.2162135.516.447164.690.58980.6958
AddSR-1s27.770.77220.31960.2242150.186.932160.850.54900.6188
SinSR-1s28.380.74970.36690.2484172.726.960655.030.49040.6412
Ours-1s27.770.75590.29670.2136134.985.913166.620.58740.7344
RealSROSEDiff-1s25.150.73410.29200.2128123.485.647169.100.63260.6687
AddSR-1s24.790.70770.30910.2191132.055.544066.180.60980.5722
SinSR-1s26.270.73510.32170.2341137.596.296460.760.54180.6163
Ours-1s24.810.71720.27430.2104114.455.129871.190.63470.7160
DIV2K-ValOSEDiff-1s23.720.61090.29420.197526.344.708967.960.61310.6681
AddSR-1s23.260.59020.36230.212329.684.761063.390.56570.5734
SinSR-1s24.410.60180.32620.206835.555.998162.950.54300.6501
Ours-1s23.020.58080.26730.182129.164.324471.690.61920.7416
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.473, + 0.497 + ], + "angle": 0, + "content": "Compared Methods. We categorize the test models into two groups: single-step and multi-step inference. The single-step inference diffusion models include SinSR [48], AddSR [54], and OSEDiff [52]. The multi-step inference diffusion models comprise StableSR [44], ResShift [61], PASD [56], DiffBIR [29], SeeSR [53], SUPIR [60], and AddSR [54]. Specifically, for AddSR, we have conducted comparisons between its single-step and four-step models. GAN-based Real-ISR methods [5, 28, 46, 62] are detailed in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.472, + 0.665 + ], + "angle": 0, + "content": "Implementation Details. All models mentioned are initialized from the Teacher Model (SD3 [10] in our work). We only train vae encoder and denoising network in the Student Model, frozen the vae decoder in order to preserve vae's prior [24]. We utilize the default prompt for the Student Model, and prompts extracted from HQ for the Teacher and LoRA models when training. We use the AdamW optimizer [30] with a learning rate of 5e-6 for the Student Model and 1e-6 for the LoRA Model, and set rank of LoRA to 64 for both two models. The training process took roughly 120h, utilizing 8 NVIDIA V100 GPUs with a batch size of 16." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.388, + 0.696 + ], + "angle": 0, + "content": "4.2. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Quantitative Comparisons. Tab. 1 show the quantitative comparison of our method with single-step diffusion models on three datasets. Ours achieves the best results on most evaluation metrics. SinSR and AddSR, as distilled versions of previous multi-step super-resolution methods, reduce inference steps but experience a corresponding decrease in performance metrics. OSEDiff introduces the VSD loss from 3D generation tasks into Real-ISR without fully addressing the substantial differences between the two domains. Therefore, the no-reference metrics for image quality are not very satisfied. In contrast, our proposed TSD-SR, tailored for Real-ISR, outperform other single-step models in terms of the vast majority of key metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.895, + 0.557 + ], + "angle": 0, + "content": "Tab. 2 show the quantitative comparison with multi-step models. We can draw the following conclusions: (1) TSD-SR demonstrates significant advantages over competing methods in terms of LPIPS, DISTS, and NIQE metrics. Additionally, it achieves performance that surpasses most multi-step models on FID, MUSIQ, and CLIPIQA. (2) DiffBIR, SeeSR, PASD, and AddSR exhibit better performance on the MANIQA metric, which may be attributed to the fact that multi-step models have more denoising iterations to produce rich details. (3) ResShift stands out with the highest PSNR and SSIM scores, while StableSR also shows notable performance regarding DISTS and FID metrics. However, both two models underperform in terms of no-reference metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.895, + 0.67 + ], + "angle": 0, + "content": "Finally, we explain the lower PSNR and SSIM metrics in our experiments. Several works [54, 60] have found that these reconstruction metrics are not well-suited for evaluating Real-ISR tasks. When a model recovers better details, it leads to lower reconstruction metrics, indicating a trade-off [3, 32, 67]. Refer to the supplementary material for detailed visual comparisons and analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Qualitative Comparisons. Fig. 7 present visual comparisons of different Real-ISR methods. As shown in multi-step methods' results, SeeSR uses degradation-aware semantic cues to leverage image generation priors, but it sometimes produces over-smooth textures. SUPIR demonstrates a notably robust generative capacity, however, the overproduction of extraneous details will lead to results that lack naturalism in image restoration (e.g. Adding excessive wrinkles at the corners of a young girl's eyes). Under more realistic degradation conditions, PASD finds it challenging to restore the corresponding content, revealing a deficiency in the model's robustness capabilities. Among single-step methods, SinSR introduces artifacts, possibly because its distilled pre-trained model is trained from scratch, lacking sufficient real-world priors, which results in unsatisfac" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23179" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with state-of-the-art multi-step methods across both synthetic and real-world benchmarks. The number of diffusion inference steps is indicated by 's'. The best and second best results of each metric are highlighted in red and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.122, + 0.894, + 0.476 + ], + "angle": 0, + "content": "
DatasetsMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
DRealSRStableSR-200s28.040.74540.32790.2272144.156.599958.530.56030.6250
DiffBIR-50s25.930.65250.45180.2761177.046.232465.660.62960.6860
SeeSR-50s28.140.77120.31410.2297146.956.463264.740.60220.6893
SUPIR-50s25.090.64600.42430.2795169.487.391858.790.54710.6749
PASD-20s27.790.74950.35790.2524171.036.766163.230.59190.6242
ResShift-15s28.690.78740.35250.2541176.777.876252.400.47560.5413
AddSR-4s26.720.71240.39820.2711164.127.668966.330.62570.7226
Ours-1s27.770.75590.29670.2136134.985.913166.620.58740.7344
RealSRStableSR-200s24.620.70410.30700.2156128.545.781765.480.62230.6198
DiffBIR-50s24.240.66500.34690.2300134.565.493268.350.65440.6961
SeeSR-50s25.210.72160.30030.2218125.105.397869.690.64430.6671
SUPIR-50s23.650.66200.35410.2488130.386.109962.090.57800.6707
PASD-20s25.680.72730.31440.2304134.185.761668.330.63230.5783
ResShift-15s26.390.75670.31580.2432149.596.874660.220.54190.5496
AddSR-4s23.330.64000.39250.2626154.225.895971.490.68260.7225
Ours-1s24.810.71720.27430.2104114.455.129871.190.63470.7160
DIV2K-ValStableSR-200s23.270.57220.31110.204624.954.773765.780.61640.6753
DiffBIR-50s23.130.57170.34690.210833.934.605668.540.63600.7125
SeeSR-50s23.730.60570.31980.195325.814.832268.490.61980.6899
SUPIR-50s22.130.52790.39190.231231.405.676763.860.59030.7146
PASD-20s24.000.60410.37790.230539.124.858767.360.61210.6327
ResShift-15s24.710.62340.34730.225342.016.361560.630.52830.5962
AddSR-4s22.160.52800.40530.236035.415.258470.990.65960.7593
Ours-1s23.020.58080.26730.182129.164.324471.690.61920.7416
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.487, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Table 3. Comparison of computational complexity across different DMs-based methods. The performance of each method is measured on an A100 GPU, using input images sized at \\(512 \\times 512\\) pixels. We have disregarded the loading time for model weights and data." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.517, + 0.893, + 0.574 + ], + "angle": 0, + "content": "
StableSRDiffBIRSeeSRSUPIRPASDResShiftAddSRAddSRSinSROSEDiffOurs
Inference Step200505050201541111
Inference Time12.41517.96375.816716.87044.84410.75461.01990.50430.14240.16500.1362
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.734 + ], + "angle": 0, + "content": "tory image restoration capabilities. AddSR produces oversmoothed results when using its 1-step model. OSEDiff offers improved restoration effects compared to SinSR and AddSR, yet it may fall short in terms of authenticity and naturalness when it comes to detail recovery. In contrast, our method effectively generates rich textures and realistic details with enhanced sharpness and contrast. Additional visual comparisons and results are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Complexity Comparisons We assess the complexity of the state-of-the-art DM-based Real-ISR models as detailed in Tab. 3, focusing on inference time. The performance of each method is benchmarked on an A100 GPU with input images sized at \\(512 \\times 512\\) pixels. We have disregarded the loading time for model weights and data, and the main computation time includes: (1) text extraction time when a text extractor is used; (2) text encoder computation time if needed; (3) VAE encoding and decoding time; (4) denoising network execution time. It is evident that TSD-SR has a substantial advantage in inference time compared with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.719 + ], + "angle": 0, + "content": "multi-step models. Specifically, TSD-SR is more than 120 times faster than SUPIR, 90 times faster than StableSR, about 50 times faster than DiffBIR, more than 40 times faster than SeeSR, more than 35 times faster than PASD, and 4 times faster than ResShift. When compared with existing one-step models, our model boasts the fastest inference times. This is because we directly denoise from LQ data and use a fixed prompt." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.623, + 0.743 + ], + "angle": 0, + "content": "4.3. User Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We conduct a user study comparing our method with three other diffusion-based one-step super-resolution methods. To make the evaluation more comprehensive, we selected images from five categories—human faces, buildings, animals, vegetation and characters. A total of 50 participants were engaged in the voting process. We guided participants to evaluate the best restoration results based on the similarity to HQ image, structural similarity to the LQ image, and realism of textures and details. The results shown in the Fig. 8 indicate that our method receives a \\(69.2\\%\\) approval" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23180" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.301, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.258, + 0.2, + 0.267 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.091, + 0.448, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.17, + 0.407, + 0.179 + ], + "angle": 0, + "content": "Zoomed LR" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.091, + 0.594, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.17, + 0.55, + 0.178 + ], + "angle": 0, + "content": "SeeSR-50s" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.091, + 0.742, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.17, + 0.699, + 0.178 + ], + "angle": 0, + "content": "SUPIR-50s" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.091, + 0.889, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.17, + 0.845, + 0.178 + ], + "angle": 0, + "content": "PASD-20s" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.269, + 0.3, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.437, + 0.2, + 0.445 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.179, + 0.448, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.258, + 0.399, + 0.267 + ], + "angle": 0, + "content": "SinSR-1s" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.179, + 0.594, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.259, + 0.548, + 0.267 + ], + "angle": 0, + "content": "AddSR-1s" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.179, + 0.742, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.258, + 0.699, + 0.267 + ], + "angle": 0, + "content": "OSEDiff-1s" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.179, + 0.889, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.259, + 0.838, + 0.267 + ], + "angle": 0, + "content": "Ours-1s" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.437, + 0.3, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.269, + 0.447, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.349, + 0.406, + 0.357 + ], + "angle": 0, + "content": "Zoomed LR" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.269, + 0.594, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.349, + 0.55, + 0.357 + ], + "angle": 0, + "content": "SeeSR-50s" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.269, + 0.742, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.349, + 0.699, + 0.357 + ], + "angle": 0, + "content": "SUPIR-50s" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.269, + 0.889, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.349, + 0.844, + 0.357 + ], + "angle": 0, + "content": "PASD-20s" + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.358, + 0.447, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.437, + 0.399, + 0.446 + ], + "angle": 0, + "content": "SinSR-1s" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.358, + 0.594, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.437, + 0.549, + 0.446 + ], + "angle": 0, + "content": "AddSR-1s" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.358, + 0.742, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.437, + 0.7, + 0.446 + ], + "angle": 0, + "content": "OSEDiff-1s" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.358, + 0.889, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.437, + 0.839, + 0.446 + ], + "angle": 0, + "content": "Ours-1s" + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.45, + 0.762, + 0.463 + ], + "angle": 0, + "content": "Figure 7. Visual comparisons of different Real-ISR methods. Please zoom in for a better view." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.481, + 0.268, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.482, + 0.468, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Figure 8. The results of our user study. Left: Category-based user preference radar chart. Our model has won the most user favor across all categories. Right: User preference pie chart. Our approach has garnered a \\(69.2\\%\\) user satisfaction rating." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.661, + 0.469, + 0.689 + ], + "angle": 0, + "content": "Table 4. Ablation study of Target Score Matching loss and Distribution-Aware Sampling Module." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.692, + 0.464, + 0.786 + ], + "angle": 0, + "content": "
DatasetsMethodDISTS↓MUSIQ↑MANIQA↑CLIPIQA↑
DRealSRw/o TSM0.232763.900.57490.6958
w/o DASM0.231163.560.58120.7123
Full0.213666.620.58740.7344
RealSRw/o TSM0.239768.560.63380.6987
w/o DASM0.227369.050.62730.7031
Full0.210471.190.63470.7160
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.469, + 0.839 + ], + "angle": 0, + "content": "rate from users. We scored \\(57.60\\%\\) in Animals, \\(70.00\\%\\) in Buildings, \\(68.80\\%\\) in Human Faces, \\(65.20\\%\\) in Vegetation, and \\(84.40\\%\\) in Characters, surpassing other models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.229, + 0.864 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "To validate the effectiveness of TSM loss and DASM, we conduct ablation studies by removing them separately in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.639 + ], + "angle": 0, + "content": "our experiments. We select DISTS, MUSIQ, MANIQA and CLIPIQA metrics for comparison, as these are critical for image quality assessment. The results are presented in Tab. 4. We draw the following conclusions: (1) The absence of TSM loss and DASM affects the metrics DISTS, MUSIQ, MANIQA and CLIPIQA. (2) The lack of TSM leads to a significant decrease in DISTS, MUSIQ and CLIPIQA metrics, possibly due to the unreliable directions in VSD causing unrealistic generations. (3) The absence of DASM results in a decline in MUSIQ and CLIPIQA, possibly due to the suboptimal optimization of details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.652, + 0.749, + 0.668 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.892, + 0.797 + ], + "angle": 0, + "content": "We propose TSD-SR, an effective and one-step model based on diffusion prior for Real-ISR. TSD-SR utilizes the TSD to enhance the realism of images generated by the distillation model. And it leverages DASM to sample distribution-based samples and accumulate their gradients to enhance the recovery of details. Our experiments have demonstrated that TSD-SR outperforms existing one-step Real-ISR models in both performance and inference speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.798, + 0.892, + 0.889 + ], + "angle": 0, + "content": "limitations. Although our model boasts excellent inference speed and restoration performance, it still has a large number of model parameters compared to the past GAN/non-diffusion approaches. In the future, we plan to employ pruning or quantization methods to compress the model parameters, striving for a lightweight and efficient Real-ISR model." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23181" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.204 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by National Natural Science Foundation of China (62293554, U2336212), Zhejiang Provincial Natural Science Foundation of China (LD24F020007), National Key R&D Program of China (SQ2023AAA01005), Ningbo Innovation \"Yongjiang 2035\" Key Research and Development Programme (2024Z292), and Young Elite Scientists Sponsorship Program by CAST (2023QNRC001)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.234, + 0.174, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.472, + 0.371 + ], + "angle": 0, + "content": "[2] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.373, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[3] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.418, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.489, + 0.47, + 0.57 + ], + "angle": 0, + "content": "[5] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Real-world blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1329-1338, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.574, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[6] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.646, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[7] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.703, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.788, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.259, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Forty-first International Conference on Machine Learning, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[12] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[13] Xiangyu He and Jian Cheng. Revisiting 11 loss in superresolution: a probabilistic view and beyond. arXiv preprint arXiv:2201.10084, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[15] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.443, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[18] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[21] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[22] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[23] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[24] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[25] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1): 79-86, 1951. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23182" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "computer vision and pattern recognition, pages 4681-4690, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.47, + 0.193 + ], + "angle": 0, + "content": "[27] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775-1787, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[28] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[29] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[30] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[31] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.411, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[32] Xiaotong Luo, Yuan Xie, Yanyun Qu, and Yun Fu. Skipdiff: Adaptive skip diffusion model for high-fidelity perceptual image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4017-4025, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.497 + ], + "angle": 0, + "content": "[33] Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.499, + 0.469, + 0.567 + ], + "angle": 0, + "content": "[34] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7807-7816, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.469, + 0.639 + ], + "angle": 0, + "content": "[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.642, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[36] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[37] Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[42] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[43] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[44] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1–21, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[45] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[46] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[47] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[48] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[51] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[52] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[53] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23183" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[54] Rui Xie, Ying Tai, Kai Zhang, Zhenyu Zhang, Jun Zhou, and Jian Yang. Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[55] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[56] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[57] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4133-4141, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[58] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[59] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[60] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[61] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[62] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[63] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[66] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. In European conference on computer vision, pages 649-667. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[67] Yuehan Zhang, Bo Ji, Jia Hao, and Angela Yao. Perception-distortion balanced admm optimization for single-image super-resolution. In European Conference on Computer Vision, pages 108-125. Springer, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23184" + } + ] +] \ No newline at end of file diff --git a/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_origin.pdf b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6f31a1b50f9710e1c60724f7c3a58a07e746be80 --- /dev/null +++ b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/f7583ef3-e58b-4f63-b84a-7f9f17f0f37a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:836c83c562ea391c5dca18d7f02f56717e6424c9661b43b7cd8c0ff2d2312c4d +size 4385507 diff --git a/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/full.md b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b7be66275151b36800fbdd0b36ca96118143ac2 --- /dev/null +++ b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/full.md @@ -0,0 +1,375 @@ +# TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution + +Linwei Dong $^{1,2*}$ Qingnan Fan $^{2*}$ Yihong Guo $^{1}$ Zhonghao Wang $^{3}$ Qi Zhang $^{2}$ Jinwei Chen $^{2}$ Yawei Luo $^{1\dagger}$ Changqing Zou $^{1,4}$ + +$^{1}$ Zhejiang University $^{2}$ Vivo Mobile Communication Co. Ltd $^{3}$ University of Chinese Academy of Sciences $^{4}$ Zhejiang Lab + +# Abstract + +Pre-trained text-to-image diffusion models are increasingly applied to real-world image super-resolution (RealISR) task. Given the iterative refinement nature of diffusion models, most existing approaches are computationally expensive. While methods such as SinSR and OSEDiff have emerged to condense inference steps via distillation, their performance in image restoration or details recovery is not satisfied. To address this, we propose TSD-SR, a novel distillation framework specifically designed for real-world image super-resolution, aiming to construct an efficient and effective one-step model. We first introduce the Target Score Distillation, which leverages the priors of diffusion models and real image references to achieve more realistic image restoration. Secondly, we propose a Distribution-Aware Sampling Module to make detail-oriented gradients more readily accessible, addressing the challenge of recovering fine details. Extensive experiments demonstrate that our TSD-SR has superior restoration results (most of the metrics perform the best) and the fastest inference speed (e.g. 40 times faster than SeeSR) compared to the past Real-ISR approaches based on pre-trained diffusion priors. + +# 1. Introduction + +Image super-resolution (ISR) [8, 9, 23, 26] aims to transform low-quality (LQ) images, which have experienced noise or blur, into clear high-quality (HQ) images. Differing from traditional ISR [6, 66], which assumes a known degradation process, real-world image super-resolution (RealISR) [46, 62] is designed to enhance real-world images that have suffered from complex and unknown degradations, thereby offering greater practical utility. + +Generative models, particularly Generative Adversarial Networks (GANs) [11, 33, 37] and Diffusion Models (DMs) + +![](images/35b3f4bb33ba96096d711b0cce07fd28f9b642fcdc5828012c89c2bc0af04fdb.jpg) +Figure 1. Performance and efficiency comparison among RealISR methods. TSD-SR stands out for achieving high-quality restoration with the fastest speed among diffusion-based models. In contrast, existing models prioritize either speed or restoration performance. The performance of each method is benchmarked on an A100 GPU with the DRealSR dataset. + +[17, 39, 41], have shown remarkable power in handling Real-ISR task. GAN-based methods leverage adversarial training, toggling between the generator and discriminator to produce realistic images. While GANs are capable of one-step inference, they are commonly hampered by issues like mode collapse and training instability [2]. Recently, Diffusion Models (DMs) have shown impressive performance in the realm of image generation [21, 47]. Their robust priors empower them to produce more realistic images with richer details than GAN-based methods [39, 41]. Some researchers [29, 53, 56, 60] have successfully leveraged pre-trained DMs for Real-ISR task. However, due to the iterative denoising nature of diffusion models [17], the Real-ISR process is computationally expensive. + +To achieve an efficient and one-step network akin to GANs, several pioneering methods that condense the iter- + +ations of diffusion models through distillation [12, 15, 18, 57] have been proposed [48, 52, 54]. Among these works, OSEDiff [52] introduced the Variational Score Distillation (VSD) loss [50] to Real-ISR task, achieving state-of-the-art (SOTA) one-step performance by leveraging prior knowledge from pre-trained models. Despite these advancements, our investigation has revealed two critical limitations associated with VSD in Real-ISR applications. (1) Unreliable gradient direction. VSD relies on a Teacher Model to provide a "true gradient direction." However, this guidance proves unreliable in scenarios where initial ISR outputs are suboptimal. (2) Insufficient detail recovery. The VSD loss exhibits notable variation across different timesteps, and the uniform sampling strategy for $t$ poses challenges in aligning the score function with detailed texture recovery requirements. These findings highlight the need for enhanced approaches to address these issues effectively. + +In this paper, we propose a novel method called TSD-SR to distill a multi-step Text-to-Image (T2I) DMs [10, 38, 39] into an effective one-step diffusion model tailored for the Real-ISR task. Specifically, TSD-SR consists of two components: Target Score Distillation (TSD) and Distribution-Aware Sampling Module (DASM). TSD incorporates our new proposed Target Score Matching (TSM) loss to compensate for the limitation of VSD loss. This significant score loss leverages HQ data to provide a reliable optimization trajectory for the distillation process. It effectively reduces the visual artifacts caused by deviant predictions from the Teacher Model. DASM is crafted to bolster detail recovery by strategically sampling low-noise samples that are distribution-based during training. This approach effectively allocates more optimization to early timesteps in a single iteration, enhancing the recovery of details. + +Experiments on popular benchmarks demonstrate that TSD-SR achieves superior restoration performance (most of the metrics perform the best) and high efficiency (the fastest inference speed, 40 times faster than SeeSR) compared to the state-of-the-art Real-ISR methods based on pre-trained DMs, while requiring only a single inference step. + +Our main contribution can be summarized as threefold: + +- We propose a novel method called TSD-SR to achieve one-step DMs distillation for the Real-ISR task. +- We introduce Target Score Distillation (TSD) to provide a reliable gradient to enhance the realism of the Real-ISR model's outputs. +- We design a Distribution-Aware Sampling Module (DASM) specifically tailored to enhance the capability of detail restoration. + +# 2. Related Work + +GAN-based Real-ISR. Since SRGAN [26] first applied GAN to ISR, it has effectively enhanced visual quality by combining adversarial loss with perceptual loss [7, + +[65]. Subsequently, ESRGAN [45] introduced Residual-in-Residual Dense Block and a relativistic average discriminator, further improving detail restoration. Methods like BSRGAN [62] and Real-ESRGAN [46] simulate complex real-world degradation processes, achieving ISR under unknown degradation conditions, which enhances the model's generalization ability. Although GAN-based methods are capable of adding more realistic details to images, they suffer from training instability and mode collapse [2]. + +Multi-step Diffusion-based Real-ISR. Some researches [29, 44, 53, 56, 60] in recent years have utilized the powerful image priors in pre-trained T2I diffusion models [35, 39, 64] for Real-SR tasks and achieved promising results. For example, StableSR [44] balances fidelity and perceptual quality by fine-tuning the time-aware encoder and employing controllable feature wrapping. DiffBiR [29] first processes the LR image through a reconstruction network and then uses the Stable Diffusion (SD) model [39] to supplement the details. SeeSR [53] attempts to better stimulate the generative power of the SD model by extracting the semantic information in the image as a conditional guide. PASD [56] introduces a pixel-aware cross attention module to enable the diffusion model to perceive the local structure of the image at the pixel level, while using a degradation removal module to extract degradation insensitive features to guide the diffusion process along with high-level information from the image. SUPIR [60] achieves a generative and fidelity capability using negative cues [16] as well as restoration-guided sampling, while using a larger pretraining model with a larger dataset to enhance the model capability. However, all of these methods are limited by the multi-step denoising of the diffusion model, which requires 20-50 iterations in inference, resulting in an inference time that lags far behind that of GAN-based methods. + +One-step Diffusion-based Real-ISR. Recently, there has been a surge of interest within the academic community in one-step distillation techniques [31, 34, 40, 58, 59] for diffusion-based Real-ISR task. SinSR [48] leverages consistency preserving distillation to condense the inference steps of ResShift [61] into a single step, yet the generalization of ResShift and SinSR is constrained due to the absence of large-scale data training. AddSR [54] introduces the adversarial diffusion distillation (ADD) [40] to RealISR tasks, resulting in a comparatively effective four-step model. However, this method has a propensity to produce excessive and unnatural image details. OSEDiff [52] directly uses LQ images as the beginning of the diffusion process, and employs VSD loss [50] as a regularization technique to condense a multi-step pre-trained T2I model into a one-step Real-ISR model. However, due to the incorporation of alternating training strategies, OSEDiff may initially tend towards unreliable optimization directions, which may lead to the visual artifacts. + +![](images/797b3e0fbfafedb1fb8afd5e16397e0d4c56c5df62a332467c77473a2ed8a2ad.jpg) +Figure 2. Pipeline overview. We train one-step Student Model $G_{\theta}$ to map the low-quality image $x_{L}$ into a more realistic one. The noisy latent $\hat{z}_{t}$ and $z_{t}$ sampled by DASM (Details can be found in Fig. 6.) will be fed into both the pre-trained Teacher and LoRA Model to produce the Variational Score loss. Subsequently, two Models' predictions on $z_{t}$ yield Target Score loss. Their weighted forms, namely TSD (red flow), in conjunction with pixel-space reconstruction loss (green flow), are leveraged to update the Student Model. After updating the Student Model, we employ the diffusion loss (blue flow) to alternately update the LoRA Model. + +# 3. Methodology + +# 3.1. Preliminaries + +Problem Formulation. ISR problem aims to reconstruct a HQ image $x_{H}$ from a LQ input $x_{L}$ by training a parameterized ISR model $G_{\theta}$ on a dataset $\mathcal{D} = \{(x_{L},x_{H})_{i = 1}^{N}\}$ , where $N$ represents the number of image pairs. It is mathematically formulated as minimizing the following objective: + +$$ +\begin{array}{l} \theta^ {*} = \arg \min _ {\theta} \mathbb {E} _ {(x _ {L}, x _ {H}) \sim \mathcal {D}} \left[ \mathcal {L} _ {R e c} \left(G _ {\theta} \left(x _ {L}\right), x _ {H}\right) \right. \tag {1} \\ \left. \left. + \lambda \mathcal {L} _ {R e g} \left(q _ {\theta} \left(\hat {x} _ {H}\right), p \left(x _ {H}\right)\right) \right] \right. \\ \end{array} +$$ + +Here, $\mathcal{L}_{Rec}$ denotes the reconstruction loss, commonly measured using distance metrics like $L_{2}$ or LPIPS [65]. The regularization term $\mathcal{L}_{Reg}$ enhances the realism and generalization of the ISR model's outputs. This objective can be understood as aligning the ISR output $\hat{x}_H$ 's distribution, $q_{\theta}(\hat{x}_H)$ , with the true distribution $p(x_H)$ by minimizing the KL-divergence [25]: + +$$ +\min _ {\theta} \mathcal {D} _ {\mathrm {K L}} \left(q _ {\theta} \left(\hat {x} _ {H}\right) \| p \left(x _ {H}\right)\right) \tag {2} +$$ + +While several studies [45, 46, 62] have employed adversarial loss to optimize this objective, they often encounter issues like mode collapse and training instability. Recent work [52] achieved the state-of-the-art results using Variational Score Distillation (VSD) as the regularization loss to minimizing this objective, which inspires our research. + +Variational Score Distillation. Variational Score Distillation (VSD) [50] was initially introduced for text-to-3D generation by distilling a pre-trained text-to-image diffusion model to optimize a single 3D representation [36]. + +In the VSD framework, a pre-trained diffusion model, represented as $\epsilon_{\psi}$ , and its trainable (LoRA [19]) replica $\epsilon_{\phi}$ are used to regularize the generator network $G_{\theta}$ . As outlined in ProlificDreamer [50], the gradient with respect to the generator parameters $\theta$ is formulated as follows: + +$$ +\begin{array}{l} \nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\mathrm {V S D}} (\hat {\boldsymbol {z}}, c _ {y}) \\ = \mathbb {E} _ {t, \epsilon} \left[ \omega (t) \left(\boldsymbol {\epsilon} _ {\psi} (\hat {\boldsymbol {z}} _ {t}; t, c _ {y}) - \boldsymbol {\epsilon} _ {\phi} (\hat {\boldsymbol {z}} _ {t}; t, c _ {y})\right) \frac {\partial \hat {\boldsymbol {z}}}{\partial \boldsymbol {\theta}} \right] \tag {3} \\ \end{array} +$$ + +where $\hat{\pmb{z}}_t = \alpha_t\hat{\pmb{z}} +\sigma_t\pmb {\epsilon}$ is the noisy input, $\hat{z}$ is the latent outputted by the generator network $G_{\theta}$ , $\pmb{\epsilon}$ is a gaussian noise, and $\alpha_{t},\sigma_{t}$ are the noise-data scaling constants. $c_{y}$ is a text embedding corresponding to a caption that describes the input image, and $w(t)$ is a time-varying weighting function. + +# 3.2. Overview of TSD-SR + +As depicted in Fig. 2, our goal is to distill a given pretrained T2I DMs into a fast one-step Student Model $G_{\theta}$ with Teacher Model $\epsilon_{\psi}$ and trainable LoRA Model $\epsilon_{\phi}$ . We denote the latent outputs of the distilled model as $\hat{z}_0$ , and HQ latent representations as $z_0$ . Both $\hat{z}_0$ and $z_0$ will pass through our Distribution-Aware Sampling Module (DASM) + +![](images/29c9948cfc3145c1a6289ec20654a3b42a21fe596a103deb8ff174ce8ada520d.jpg) +Figure 3. A visual comparison of the gradient direction. We set the timestep $t$ to 100 and calculated the cosine similarity between the predicted directions from the Teacher Model and the true direction (towards the HQ data). The prediction direction for $z_{t}$ closely matches the true direction, but not for $\hat{z}_{t}$ , suggesting that suboptimal samples may lead to directional deviations. + +to obtain distribution-based samples $\hat{z}_t$ and $z_t$ (Sec. 3.4). We train $G_{\theta}$ by minimizing the two losses: a reconstruction loss in pixel space to compare the model outputs against the ground truth, and a regularization loss (from Target Score Distillation) to enhances the realism (Sec. 3.3). After updating the Student Model, we update the LoRA Model with diffusion loss. Finally, in Sec. 3.5, we present an overview of all the losses encountered during the training phase. + +# 3.3. Target Score Distillation + +Similar to [52], we introduce VSD loss into our work as a regularization loss to enhance the realism and generalization of the $G_{\theta}$ outputs. Upon reviewing VSD Eq. (3), $\epsilon_{\phi}(\hat{z}_t;t,c_y)$ represents the gradient direction of the optimization on $G_{\theta}$ noisy outputs $\hat{z}_t$ , whereas $\epsilon_{\psi}(\hat{z}_t;t,c_y)$ corresponds to the gradient direction on noisy HQ latent representations $z_{t}$ . The overarching goal of model optimization is to harmonize the predictions from $\hat{z}_t$ with those from $z_{t}$ Nonetheless, this strategy encounters hurdles, especially in the early training phase: the quality of synthetic latents $\hat{z}_t$ is not high enough for Teacher Model to give a precise prediction. As illustrated in Fig. 3, the Teacher Model finds it challenging to accurately predict the optimization direction for low-quality synthetic latents, with the cosine similarity to the ideal optimization direction only 0.2 (0.88 on HQ latents). This problem can lead to severe visual artifacts, as evident in Fig. 4(a). + +A straightforward remedial measure is to employ a mean squared error (MSE) loss to align the synthetic latents with the ideal inputs of the Teacher Model, which are derived from HQ latents. However, as shown in Fig. 4(b), this approach has been observed to lead to over-smoothed results [13]. Our strategy, instead, is to align the predictions made by the Teacher Model on both synthetic and HQ + +![](images/10d97d354a5ad2a357932d8ac64204b165d462476582967da01ee481ff54b6b0.jpg) +(a) Naive + +![](images/3e6306ca47e5f62481283641e1ba077a5bec164d6a5e866fd6ece3772e3a3de6.jpg) +(b) MSE + +![](images/1d4e80265ad8eb0f1c332342eac376b800792314140e9b65d297f2cc55340347.jpg) +(c) Ours +Figure 4. The visualization of different strategies. (a) The naive method introduces fake textures and fails to recover fine details. (b) MSE leads to over-smoothed generation result, lacking high-frequency information. (c) Our method offers the superior visual effects and fine textures. + +latents, thereby encouraging greater consistency between them. The core idea is that for samples drawn from the same distribution, the real scores predicted by the Teacher Model should be close to each other. We refer to this approach as Target Score Matching (TSM): + +$$ +\begin{array}{l} \nabla_ {\boldsymbol {\theta}} \mathcal {L} _ {\mathrm {T S M}} (\hat {\mathbf {z}}, \mathbf {z}, c _ {y}) \\ = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\epsilon_ {\psi} \left(\hat {\mathbf {z}} _ {t}; t, c _ {y}\right) - \epsilon_ {\psi} \left(\mathbf {z} _ {t}; t, c _ {y}\right)\right) \frac {\partial \hat {\mathbf {z}}}{\partial \theta} \right] \tag {4} \\ \end{array} +$$ + +where the expectation of the gradient is computed across all diffusion timesteps $t \in \{1, \dots, T\}$ and $\epsilon \sim \mathcal{N}(0, I)$ . Equation (4) encapsulates the optimization loss for our Target Score Matching. Upon examining it in conjunction with the Eq. (3), we notice that VSD utilizes the prediction residual between a Teacher and a LoRA Model to drive gradient backpropagation. Conversely, our TSM employs the synthetic and the HQ data to produce the gradients. By blending these two strategies with hyperparameter weighting $\lambda$ and $1 - \lambda$ , we can construct our ultimate optimization loss, effectively unifying the strengths of both approaches to optimize the training process Eq. (5). + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {T S D}} (\hat {\mathbf {z}}, \mathbf {z}, c _ {y}) = \mathbb {E} _ {t, \epsilon} \left[ w (t) \left[ \epsilon_ {\psi} \left(\hat {\mathbf {z}} _ {t}; t, c _ {y}\right) - \right. \right. +$$ + +$$ +\left. \left. \epsilon_ {\psi} (\mathbf {z} _ {t}; t, c _ {y}) + \lambda \left(\epsilon_ {\psi} \left(\mathbf {z} _ {t}; t, c _ {y}\right) - \epsilon_ {\phi} \left(\hat {\mathbf {z}} _ {t}; t, c _ {y}\right)\right) \right] \frac {\partial \hat {\mathbf {z}}}{\partial \theta} \right] \tag {5} +$$ + +where $w(t)$ is time-aware weighting function tailored for Real-ISR. Other symbols are in accordance with those previously mentioned. By introducing the prediction of pretrained diffusion model on HQ latents, we have circumvented the issue of the model falling into the visual artifacts or over-smoothed problem, as illustrated in Fig. 4(c). + +# 3.4. Distribution-Aware Sampling Module + +In the VSD-based framework, we need to match the score functions predicted by the Teacher Model with those by the LoRA Model at different timesteps $t \sim \{0,1,\dots,T\}$ . + +![](images/34e2f5d7aa72822b49fb6fc97431cef0b4ac1ec61c8aa52772cf1cedcbe766e4.jpg) +(a) Predicted Residual at Different Timesteps in VSD + +![](images/2e4c1def5eba71ccb05c319f4d676ebe73146dc846ffc704abe3c00a79af3db5.jpg) +(b) Stage 1 + +![](images/1d8cbadb7f05c3880a3daec4e9187ba99c160cdeb44642c7832c2188d7da78f0.jpg) +(c) Stage 2 +Figure 5. (a) The prediction errors of the VSD loss at different timesteps. The error divergence is more pronounced in early timesteps than later. This phenomenon is observed throughout the optimization process. (b) the visualization of Stage 1 prediction error. (c) the visualization of Stage 2 prediction error. + +However, for the Real-ISR problem, this matching ability at different $t$ is unbalanced, as illustrated in Fig. 5(a). This phenomenon could be attributed to our leveraging of the low-frequency (LF) priors from LQ data, but losing guidance from high-frequency (HF) details. The noisy samples $\hat{z}_t$ possesses low-frequency priors (from LQ data). These priors are easily learned by the LoRA Model, leading to more similar predictions (Fig. 5(b)) during LF restoration (Stage 1). However, In the Stage 2, the lack of HF details in $\hat{z}_t$ causes the LoRA Model to poorly perceive details, resulting in divergent predictions (Fig. 5(c)). Our goal is to reduce this divergence. + +Existing method, for each iteration, only matches the score function at a single timestep sample $\hat{z}_t$ , with uniform sampling strategy of time $t$ . This leads to slow convergence or even difficulty in optimization during Stage 2, as the gradients of important timesteps are averaged out. To this end, we propose our Distribution-Aware Sampling Module (DASM). This module accumulates optimization gradients for earlier timestep samples in a single iteration, enabling the backpropagation of more gradients focused on detail optimization. As shown in Fig. 6, we first introduce noise into the synthetic latent representation $\hat{z}_t = (1 - \sigma_t)\hat{z}_0 + \sigma_t\epsilon$ , where $\sigma$ is a weight factor and $\epsilon$ is a gaussian noise. Subsequently, we employ a LoRA Model to perform denoising, yielding the previous time noisy samples by Eq. (6): + +$$ +\hat {z} _ {\boldsymbol {t} - 1} = \hat {z} _ {\boldsymbol {t}} + \left(\sigma_ {t - 1} - \sigma_ {t}\right) \cdot \epsilon_ {\phi} \left(\hat {z} _ {\boldsymbol {t}}; t, c _ {y}\right), \tag {6} +$$ + +where $\sigma_{t - 1}$ and $\sigma_t$ are from the flow matching scheduler, and here LoRA Model has fit $\hat{z}_0$ 's distribution. Similarly, $z_{t - 1}$ can be obtained by denoising on the Teacher Model. Ultimately, in a single iteration, the gradients from noise samples along its trajectory can be accumulated to update the Student Model. Since the obtained samples follow the diffusion sampling trajectory and are directed towards early + +![](images/40af5971cae9330009f7fe1a9334a81903a124f25e487820766a53e91c2386d4.jpg) +Figure 6. Illustration of DASM. Top: The naive method that add noise to samples. Bottom: Our DASM leverages the priors of diffusion models to sample noise that more closely aligns with the true sampling trajectory. The noisy samples can all serve as inputs to the subsequent network to generate gradient backpropagation. + +timestep, we effectively optimize the divergence in Stage 2. + +# 3.5. Training Objective + +We summarize all the losses we used in our framework. Student Model $G_{\theta}$ . We train our Student Model with the reconstruction loss $\mathcal{L}_{Rec}$ and the regularization loss $\mathcal{L}_{Reg}$ . For the reconstruction loss, we use LPIPS loss: + +$$ +\mathcal {L} _ {\text {R e c}} \left(G _ {\theta} \left(\boldsymbol {x} _ {L}\right), \boldsymbol {x} _ {H}\right) = \mathcal {L} _ {\text {L P I P S}} \left(G _ {\theta} \left(\boldsymbol {x} _ {L}\right), \boldsymbol {x} _ {H}\right). \tag {7} +$$ + +For the regularization loss, we use our TSD loss Eq. (5). Therefore, the overall training objective for the Student Model $G_{\theta}$ is: + +$$ +\mathcal {L} _ {\text {S t u}} = \mathcal {L} _ {\text {R e c}} + \gamma \mathcal {L} _ {\text {R e g}}, \tag {8} +$$ + +where $\gamma$ is a weighting scalar. + +LoRA Model $\epsilon_{\phi}$ . As stipulated by VSD, the fine-tuned replica $\epsilon_{\phi}$ must be trainable, with its training objective being: + +$$ +\mathcal {L} _ {D i f f} (\hat {\boldsymbol {z}}, c _ {y}) = \mathbb {E} _ {t, \epsilon} \left[ \left\| \boldsymbol {\epsilon} _ {\phi} \left(\hat {\boldsymbol {z}} _ {t}; t, c _ {y}\right) - \boldsymbol {\epsilon} ^ {\prime} \right\| ^ {2} \right], \tag {9} +$$ + +where $\epsilon^{\prime}$ serves as the training target for the denoising network, representing Gaussian noise in the context of DDPM, and a gradient towards HQ for flow matching. + +# 4. Experiments + +# 4.1. Experimental Settings + +Training Datasets. We adopt DIV2K [1], Flickr2K [42], LSDIR [27] and first 10K face images from FFHQ [20] for training. We use the same degradation pipeline as Real-ESRGAN [46] to synthesize LR-HR pairs. + +Test Datasets. We evaluated our model on synthetic DIV2K-Val [1] dataset and two real-world datasets, RealSR [4] and DRealSR [51]. The real-world datasets comprised 128x128 LQ and 512x512 HQ image pairs. + +Evaluation Metrics. For evaluating our method, we apply both full-reference and no-reference metrics. Full-reference metrics include PSNR and SSIM [49] (calculated on the Y channel in YCbCr space) for fidelity, LPIPS [65] and DISTS [7] for perceptual quality, and FID [14] for distribution comparison. No-reference metrics include NIQE [63], MANIQA[55], MUSIQ [22] and CLIPIQA [43]. + +Table 1. Quantitative comparison with the state-of-the-art one-step methods across both synthetic and real-world benchmarks. The number of diffusion inference steps is indicated by 's'. The best results of each metric are highlighted in red. + +
DatasetsMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
DRealSROSEDiff-1s27.920.78360.29680.2162135.516.447164.690.58980.6958
AddSR-1s27.770.77220.31960.2242150.186.932160.850.54900.6188
SinSR-1s28.380.74970.36690.2484172.726.960655.030.49040.6412
Ours-1s27.770.75590.29670.2136134.985.913166.620.58740.7344
RealSROSEDiff-1s25.150.73410.29200.2128123.485.647169.100.63260.6687
AddSR-1s24.790.70770.30910.2191132.055.544066.180.60980.5722
SinSR-1s26.270.73510.32170.2341137.596.296460.760.54180.6163
Ours-1s24.810.71720.27430.2104114.455.129871.190.63470.7160
DIV2K-ValOSEDiff-1s23.720.61090.29420.197526.344.708967.960.61310.6681
AddSR-1s23.260.59020.36230.212329.684.761063.390.56570.5734
SinSR-1s24.410.60180.32620.206835.555.998162.950.54300.6501
Ours-1s23.020.58080.26730.182129.164.324471.690.61920.7416
+ +Compared Methods. We categorize the test models into two groups: single-step and multi-step inference. The single-step inference diffusion models include SinSR [48], AddSR [54], and OSEDiff [52]. The multi-step inference diffusion models comprise StableSR [44], ResShift [61], PASD [56], DiffBIR [29], SeeSR [53], SUPIR [60], and AddSR [54]. Specifically, for AddSR, we have conducted comparisons between its single-step and four-step models. GAN-based Real-ISR methods [5, 28, 46, 62] are detailed in the supplementary material. + +Implementation Details. All models mentioned are initialized from the Teacher Model (SD3 [10] in our work). We only train vae encoder and denoising network in the Student Model, frozen the vae decoder in order to preserve vae's prior [24]. We utilize the default prompt for the Student Model, and prompts extracted from HQ for the Teacher and LoRA models when training. We use the AdamW optimizer [30] with a learning rate of 5e-6 for the Student Model and 1e-6 for the LoRA Model, and set rank of LoRA to 64 for both two models. The training process took roughly 120h, utilizing 8 NVIDIA V100 GPUs with a batch size of 16. + +# 4.2. Comparison with Existing Methods + +Quantitative Comparisons. Tab. 1 show the quantitative comparison of our method with single-step diffusion models on three datasets. Ours achieves the best results on most evaluation metrics. SinSR and AddSR, as distilled versions of previous multi-step super-resolution methods, reduce inference steps but experience a corresponding decrease in performance metrics. OSEDiff introduces the VSD loss from 3D generation tasks into Real-ISR without fully addressing the substantial differences between the two domains. Therefore, the no-reference metrics for image quality are not very satisfied. In contrast, our proposed TSD-SR, tailored for Real-ISR, outperform other single-step models in terms of the vast majority of key metrics. + +Tab. 2 show the quantitative comparison with multi-step models. We can draw the following conclusions: (1) TSD-SR demonstrates significant advantages over competing methods in terms of LPIPS, DISTS, and NIQE metrics. Additionally, it achieves performance that surpasses most multi-step models on FID, MUSIQ, and CLIPIQA. (2) DiffBIR, SeeSR, PASD, and AddSR exhibit better performance on the MANIQA metric, which may be attributed to the fact that multi-step models have more denoising iterations to produce rich details. (3) ResShift stands out with the highest PSNR and SSIM scores, while StableSR also shows notable performance regarding DISTS and FID metrics. However, both two models underperform in terms of no-reference metrics. + +Finally, we explain the lower PSNR and SSIM metrics in our experiments. Several works [54, 60] have found that these reconstruction metrics are not well-suited for evaluating Real-ISR tasks. When a model recovers better details, it leads to lower reconstruction metrics, indicating a trade-off [3, 32, 67]. Refer to the supplementary material for detailed visual comparisons and analysis. + +Qualitative Comparisons. Fig. 7 present visual comparisons of different Real-ISR methods. As shown in multi-step methods' results, SeeSR uses degradation-aware semantic cues to leverage image generation priors, but it sometimes produces over-smooth textures. SUPIR demonstrates a notably robust generative capacity, however, the overproduction of extraneous details will lead to results that lack naturalism in image restoration (e.g. Adding excessive wrinkles at the corners of a young girl's eyes). Under more realistic degradation conditions, PASD finds it challenging to restore the corresponding content, revealing a deficiency in the model's robustness capabilities. Among single-step methods, SinSR introduces artifacts, possibly because its distilled pre-trained model is trained from scratch, lacking sufficient real-world priors, which results in unsatisfac + +Table 2. Quantitative comparison with state-of-the-art multi-step methods across both synthetic and real-world benchmarks. The number of diffusion inference steps is indicated by 's'. The best and second best results of each metric are highlighted in red and blue, respectively. + +
DatasetsMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
DRealSRStableSR-200s28.040.74540.32790.2272144.156.599958.530.56030.6250
DiffBIR-50s25.930.65250.45180.2761177.046.232465.660.62960.6860
SeeSR-50s28.140.77120.31410.2297146.956.463264.740.60220.6893
SUPIR-50s25.090.64600.42430.2795169.487.391858.790.54710.6749
PASD-20s27.790.74950.35790.2524171.036.766163.230.59190.6242
ResShift-15s28.690.78740.35250.2541176.777.876252.400.47560.5413
AddSR-4s26.720.71240.39820.2711164.127.668966.330.62570.7226
Ours-1s27.770.75590.29670.2136134.985.913166.620.58740.7344
RealSRStableSR-200s24.620.70410.30700.2156128.545.781765.480.62230.6198
DiffBIR-50s24.240.66500.34690.2300134.565.493268.350.65440.6961
SeeSR-50s25.210.72160.30030.2218125.105.397869.690.64430.6671
SUPIR-50s23.650.66200.35410.2488130.386.109962.090.57800.6707
PASD-20s25.680.72730.31440.2304134.185.761668.330.63230.5783
ResShift-15s26.390.75670.31580.2432149.596.874660.220.54190.5496
AddSR-4s23.330.64000.39250.2626154.225.895971.490.68260.7225
Ours-1s24.810.71720.27430.2104114.455.129871.190.63470.7160
DIV2K-ValStableSR-200s23.270.57220.31110.204624.954.773765.780.61640.6753
DiffBIR-50s23.130.57170.34690.210833.934.605668.540.63600.7125
SeeSR-50s23.730.60570.31980.195325.814.832268.490.61980.6899
SUPIR-50s22.130.52790.39190.231231.405.676763.860.59030.7146
PASD-20s24.000.60410.37790.230539.124.858767.360.61210.6327
ResShift-15s24.710.62340.34730.225342.016.361560.630.52830.5962
AddSR-4s22.160.52800.40530.236035.415.258470.990.65960.7593
Ours-1s23.020.58080.26730.182129.164.324471.690.61920.7416
+ +Table 3. Comparison of computational complexity across different DMs-based methods. The performance of each method is measured on an A100 GPU, using input images sized at $512 \times 512$ pixels. We have disregarded the loading time for model weights and data. + +
StableSRDiffBIRSeeSRSUPIRPASDResShiftAddSRAddSRSinSROSEDiffOurs
Inference Step200505050201541111
Inference Time12.41517.96375.816716.87044.84410.75461.01990.50430.14240.16500.1362
+ +tory image restoration capabilities. AddSR produces oversmoothed results when using its 1-step model. OSEDiff offers improved restoration effects compared to SinSR and AddSR, yet it may fall short in terms of authenticity and naturalness when it comes to detail recovery. In contrast, our method effectively generates rich textures and realistic details with enhanced sharpness and contrast. Additional visual comparisons and results are provided in the supplementary material. + +Complexity Comparisons We assess the complexity of the state-of-the-art DM-based Real-ISR models as detailed in Tab. 3, focusing on inference time. The performance of each method is benchmarked on an A100 GPU with input images sized at $512 \times 512$ pixels. We have disregarded the loading time for model weights and data, and the main computation time includes: (1) text extraction time when a text extractor is used; (2) text encoder computation time if needed; (3) VAE encoding and decoding time; (4) denoising network execution time. It is evident that TSD-SR has a substantial advantage in inference time compared with + +multi-step models. Specifically, TSD-SR is more than 120 times faster than SUPIR, 90 times faster than StableSR, about 50 times faster than DiffBIR, more than 40 times faster than SeeSR, more than 35 times faster than PASD, and 4 times faster than ResShift. When compared with existing one-step models, our model boasts the fastest inference times. This is because we directly denoise from LQ data and use a fixed prompt. + +# 4.3. User Study + +We conduct a user study comparing our method with three other diffusion-based one-step super-resolution methods. To make the evaluation more comprehensive, we selected images from five categories—human faces, buildings, animals, vegetation and characters. A total of 50 participants were engaged in the voting process. We guided participants to evaluate the best restoration results based on the similarity to HQ image, structural similarity to the LQ image, and realism of textures and details. The results shown in the Fig. 8 indicate that our method receives a $69.2\%$ approval + +![](images/fe86e8eee002cf877a96d586a8a3857aed4c08c99aa54c190f5b5fff477507e9.jpg) +HR + +![](images/e022902d11bee8f3cc7b6daf54e816099a84ecec0461b2c3391a98c4ab84827c.jpg) +Zoomed LR + +![](images/29755036bfcbd3f83680a23116c0dcbf7eb9139d41d75118d907060c3e3ed45d.jpg) +SeeSR-50s + +![](images/724cc0fbdcf19a5d9301f4d3c96ad6eb56cdd9aff8cca5910905a46df375b324.jpg) +SUPIR-50s + +![](images/8ca7be1e537bf4cb4134a6edca8f7c880544686fe04fd2fba33bb0fb4f7cd7b5.jpg) +PASD-20s + +![](images/db129735dea7555c006e0682c6e383ce8a935e78b4e0b39f1a0f6c61536d46c7.jpg) +HR + +![](images/bb42b0e29977ac9962f67489b65fb3a25a986c136b11684fe6fc6e5575d6e4a1.jpg) +SinSR-1s + +![](images/2b71ce843efcf03611aca132fa8446e5810df1aec80e367e257c43cfb6b64ccf.jpg) +AddSR-1s + +![](images/22930d09d16fd6160e3365cabf60ad34209152b69b7b0717575dfeafeb046044.jpg) +OSEDiff-1s + +![](images/6751e74dcad40df56fea2d31e3373edb9434e45be3d4c1e293811fad8ff594d8.jpg) +Ours-1s + +![](images/a17b5e00b02004a98fb29cedd24f72e34b827b63b135b57b39ff18c0599f3c63.jpg) +Figure 7. Visual comparisons of different Real-ISR methods. Please zoom in for a better view. +Figure 8. The results of our user study. Left: Category-based user preference radar chart. Our model has won the most user favor across all categories. Right: User preference pie chart. Our approach has garnered a $69.2\%$ user satisfaction rating. + +![](images/46fc5e3c8340a63b9244833eb0082e595aa016f960d93afd8bb38a3fdb0207d1.jpg) +Zoomed LR + +![](images/eb947c9f6cb1c0d1cd9ed3183fe61f2291afc5fd1fb4d3ca5dcf7ffe6d107431.jpg) +SeeSR-50s + +![](images/f20b4635ecbab927ca453b53f34a847e718a86647422f9d4b767c74614a36ed4.jpg) +SUPIR-50s + +![](images/66a01f6b27aa7445518e436fefca1e27781a12d9ba01655fba9656efcfadb050.jpg) +PASD-20s + +![](images/fd1065f8c6b713bb488db295345507232575856b31f4b66002404d7d1e1c1688.jpg) +SinSR-1s + +![](images/8f5629488e3008b2e0c386e0d231fff50a34cf4906df7c13eb762bf1cf6971a6.jpg) +AddSR-1s + +![](images/5bf3f251f012bc7b8fc7a1fa65974609daf6e8ec40be3c4d6f82637b47642561.jpg) +OSEDiff-1s + +![](images/edb0c465d072a5e7f438e452f5771433c626607d9a238543b31eaeb4e87f9212.jpg) +Ours-1s + +![](images/273a2e90fc2f112a8ac481c8051671bfe3cdf262cd1b18e35987f4de04a0b16f.jpg) + +Table 4. Ablation study of Target Score Matching loss and Distribution-Aware Sampling Module. + +
DatasetsMethodDISTS↓MUSIQ↑MANIQA↑CLIPIQA↑
DRealSRw/o TSM0.232763.900.57490.6958
w/o DASM0.231163.560.58120.7123
Full0.213666.620.58740.7344
RealSRw/o TSM0.239768.560.63380.6987
w/o DASM0.227369.050.62730.7031
Full0.210471.190.63470.7160
+ +rate from users. We scored $57.60\%$ in Animals, $70.00\%$ in Buildings, $68.80\%$ in Human Faces, $65.20\%$ in Vegetation, and $84.40\%$ in Characters, surpassing other models. + +# 4.4. Ablation Study + +To validate the effectiveness of TSM loss and DASM, we conduct ablation studies by removing them separately in + +our experiments. We select DISTS, MUSIQ, MANIQA and CLIPIQA metrics for comparison, as these are critical for image quality assessment. The results are presented in Tab. 4. We draw the following conclusions: (1) The absence of TSM loss and DASM affects the metrics DISTS, MUSIQ, MANIQA and CLIPIQA. (2) The lack of TSM leads to a significant decrease in DISTS, MUSIQ and CLIPIQA metrics, possibly due to the unreliable directions in VSD causing unrealistic generations. (3) The absence of DASM results in a decline in MUSIQ and CLIPIQA, possibly due to the suboptimal optimization of details. + +# 5. Conclusion and Limitation + +We propose TSD-SR, an effective and one-step model based on diffusion prior for Real-ISR. TSD-SR utilizes the TSD to enhance the realism of images generated by the distillation model. And it leverages DASM to sample distribution-based samples and accumulate their gradients to enhance the recovery of details. Our experiments have demonstrated that TSD-SR outperforms existing one-step Real-ISR models in both performance and inference speed. + +limitations. Although our model boasts excellent inference speed and restoration performance, it still has a large number of model parameters compared to the past GAN/non-diffusion approaches. In the future, we plan to employ pruning or quantization methods to compress the model parameters, striving for a lightweight and efficient Real-ISR model. + +Acknowledgment. This work was supported by National Natural Science Foundation of China (62293554, U2336212), Zhejiang Provincial Natural Science Foundation of China (LD24F020007), National Key R&D Program of China (SQ2023AAA01005), Ningbo Innovation "Yongjiang 2035" Key Research and Development Programme (2024Z292), and Young Elite Scientists Sponsorship Program by CAST (2023QNRC001). + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5 +[2] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 1, 2 +[3] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 6 +[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 5 +[5] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Real-world blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1329-1338, 2022. 6 +[6] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1 +[7] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 2, 5 +[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 1 +[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1 +[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. + +In Forty-first International Conference on Machine Learning, 2024. 2, 6 +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[12] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 2 +[13] Xiangyu He and Jian Cheng. Revisiting 11 loss in superresolution: a probabilistic view and beyond. arXiv preprint arXiv:2201.10084, 2022. 4 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[15] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[18] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2 +[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5 +[21] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 1 +[22] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5 +[23] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 1 +[24] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 6 +[25] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1): 79-86, 1951. 3 +[26] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on + +computer vision and pattern recognition, pages 4681-4690, 2017. 1, 2 +[27] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775-1787, 2023. 5 +[28] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 6 +[29] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 1, 2, 6 +[30] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[31] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2 +[32] Xiaotong Luo, Yuan Xie, Yanyun Qu, and Yun Fu. Skipdiff: Adaptive skip diffusion model for high-fidelity perceptual image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4017-4025, 2024. 6 +[33] Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 1 +[34] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7807-7816, 2024. 2 +[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[36] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3 +[37] Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 1 +[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2 +[40] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2 + +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1 +[42] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5 +[43] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5 +[44] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1–21, 2024. 2, 6 +[45] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 3 +[46] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2, 3, 5, 6 +[47] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490, 2022. 1 +[48] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2, 6 +[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[51] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 5 +[52] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 4, 6 +[53] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 1, 2, 6 +[54] Rui Xie, Ying Tai, Kai Zhang, Zhenyu Zhang, Jun Zhou, and Jian Yang. Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 2, 6 +[55] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 5 +[56] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 1, 2, 6 +[57] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4133-4141, 2017. 2 +[58] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2 +[59] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2 +[60] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 1, 2, 6 +[61] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 2, 6 +[62] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 1, 2, 3, 6 +[63] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 5 +[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 3, 5 + +[66] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. In European conference on computer vision, pages 649-667. Springer, 2022. 1 +[67] Yuehan Zhang, Bo Ji, Jia Hao, and Angela Yao. Perception-distortion balanced admm optimization for single-image super-resolution. In European Conference on Computer Vision, pages 108-125. Springer, 2022. 6 \ No newline at end of file diff --git a/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/images.zip b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..735d694102e31ccb28a8677d5638430ebd47e52a --- /dev/null +++ b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d72a2d292a6987406e4d2181e3efaf3fd17ec4cbda041c261aa8c104732ca03c +size 847117 diff --git a/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/layout.json b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..175ddbefe11a31a1060723fc160798d7efacc0a5 --- /dev/null +++ b/2025/TSD-SR_ One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution/layout.json @@ -0,0 +1,10553 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "text", + "content": "TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "spans": [ + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "text", + "content": "Linwei Dong" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "text", + "content": " Qingnan Fan" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "text", + "content": " Yihong Guo" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "text", + "content": " Zhonghao Wang" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "text", + "content": " Qi Zhang" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "text", + "content": " Jinwei Chen" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "text", + "content": " Yawei Luo" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "text", + "content": " Changqing Zou" + }, + { + "bbox": [ + 110, + 159, + 465, + 191 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "type": "text", + "content": "Vivo Mobile Communication Co. Ltd " + }, + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences " + }, + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 202, + 451, + 232 + ], + "type": "text", + "content": "Zhejiang Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 284, + 290, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 536 + ], + "type": "text", + "content": "Pre-trained text-to-image diffusion models are increasingly applied to real-world image super-resolution (RealISR) task. Given the iterative refinement nature of diffusion models, most existing approaches are computationally expensive. While methods such as SinSR and OSEDiff have emerged to condense inference steps via distillation, their performance in image restoration or details recovery is not satisfied. To address this, we propose TSD-SR, a novel distillation framework specifically designed for real-world image super-resolution, aiming to construct an efficient and effective one-step model. We first introduce the Target Score Distillation, which leverages the priors of diffusion models and real image references to achieve more realistic image restoration. Secondly, we propose a Distribution-Aware Sampling Module to make detail-oriented gradients more readily accessible, addressing the challenge of recovering fine details. Extensive experiments demonstrate that our TSD-SR has superior restoration results (most of the metrics perform the best) and the fastest inference speed (e.g. 40 times faster than SeeSR) compared to the past Real-ISR approaches based on pre-trained diffusion priors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 556, + 128, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 128, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 128, + 569 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 577, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 287, + 673 + ], + "type": "text", + "content": "Image super-resolution (ISR) [8, 9, 23, 26] aims to transform low-quality (LQ) images, which have experienced noise or blur, into clear high-quality (HQ) images. Differing from traditional ISR [6, 66], which assumes a known degradation process, real-world image super-resolution (RealISR) [46, 62] is designed to enhance real-world images that have suffered from complex and unknown degradations, thereby offering greater practical utility." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 673, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 287, + 696 + ], + "type": "text", + "content": "Generative models, particularly Generative Adversarial Networks (GANs) [11, 33, 37] and Diffusion Models (DMs)" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 314, + 277, + 525, + 437 + ], + "blocks": [ + { + "bbox": [ + 314, + 277, + 525, + 437 + ], + "lines": [ + { + "bbox": [ + 314, + 277, + 525, + 437 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 525, + 437 + ], + "type": "image", + "image_path": "35b3f4bb33ba96096d711b0cce07fd28f9b642fcdc5828012c89c2bc0af04fdb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 439, + 547, + 506 + ], + "lines": [ + { + "bbox": [ + 304, + 439, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 547, + 506 + ], + "type": "text", + "content": "Figure 1. Performance and efficiency comparison among RealISR methods. TSD-SR stands out for achieving high-quality restoration with the fastest speed among diffusion-based models. In contrast, existing models prioritize either speed or restoration performance. The performance of each method is benchmarked on an A100 GPU with the DRealSR dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 518, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 687 + ], + "type": "text", + "content": "[17, 39, 41], have shown remarkable power in handling Real-ISR task. GAN-based methods leverage adversarial training, toggling between the generator and discriminator to produce realistic images. While GANs are capable of one-step inference, they are commonly hampered by issues like mode collapse and training instability [2]. Recently, Diffusion Models (DMs) have shown impressive performance in the realm of image generation [21, 47]. Their robust priors empower them to produce more realistic images with richer details than GAN-based methods [39, 41]. Some researchers [29, 53, 56, 60] have successfully leveraged pre-trained DMs for Real-ISR task. However, due to the iterative denoising nature of diffusion models [17], the Real-ISR process is computationally expensive." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "To achieve an efficient and one-step network akin to GANs, several pioneering methods that condense the iter-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 211, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 211, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 211, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23174" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "ations of diffusion models through distillation [12, 15, 18, 57] have been proposed [48, 52, 54]. Among these works, OSEDiff [52] introduced the Variational Score Distillation (VSD) loss [50] to Real-ISR task, achieving state-of-the-art (SOTA) one-step performance by leveraging prior knowledge from pre-trained models. Despite these advancements, our investigation has revealed two critical limitations associated with VSD in Real-ISR applications. (1) Unreliable gradient direction. VSD relies on a Teacher Model to provide a \"true gradient direction.\" However, this guidance proves unreliable in scenarios where initial ISR outputs are suboptimal. (2) Insufficient detail recovery. The VSD loss exhibits notable variation across different timesteps, and the uniform sampling strategy for " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " poses challenges in aligning the score function with detailed texture recovery requirements. These findings highlight the need for enhanced approaches to address these issues effectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 467 + ], + "type": "text", + "content": "In this paper, we propose a novel method called TSD-SR to distill a multi-step Text-to-Image (T2I) DMs [10, 38, 39] into an effective one-step diffusion model tailored for the Real-ISR task. Specifically, TSD-SR consists of two components: Target Score Distillation (TSD) and Distribution-Aware Sampling Module (DASM). TSD incorporates our new proposed Target Score Matching (TSM) loss to compensate for the limitation of VSD loss. This significant score loss leverages HQ data to provide a reliable optimization trajectory for the distillation process. It effectively reduces the visual artifacts caused by deviant predictions from the Teacher Model. DASM is crafted to bolster detail recovery by strategically sampling low-noise samples that are distribution-based during training. This approach effectively allocates more optimization to early timesteps in a single iteration, enhancing the recovery of details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 467, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 289, + 539 + ], + "type": "text", + "content": "Experiments on popular benchmarks demonstrate that TSD-SR achieves superior restoration performance (most of the metrics perform the best) and high efficiency (the fastest inference speed, 40 times faster than SeeSR) compared to the state-of-the-art Real-ISR methods based on pre-trained DMs, while requiring only a single inference step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 539, + 270, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 270, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 270, + 550 + ], + "type": "text", + "content": "Our main contribution can be summarized as threefold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 551, + 287, + 645 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 551, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 287, + 574 + ], + "type": "text", + "content": "- We propose a novel method called TSD-SR to achieve one-step DMs distillation for the Real-ISR task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 610 + ], + "type": "text", + "content": "- We introduce Target Score Distillation (TSD) to provide a reliable gradient to enhance the realism of the Real-ISR model's outputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 610, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 287, + 645 + ], + "type": "text", + "content": "- We design a Distribution-Aware Sampling Module (DASM) specifically tailored to enhance the capability of detail restoration." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 657, + 134, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 134, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 134, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "GAN-based Real-ISR. Since SRGAN [26] first applied GAN to ISR, it has effectively enhanced visual quality by combining adversarial loss with perceptual loss [7," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "[65]. Subsequently, ESRGAN [45] introduced Residual-in-Residual Dense Block and a relativistic average discriminator, further improving detail restoration. Methods like BSRGAN [62] and Real-ESRGAN [46] simulate complex real-world degradation processes, achieving ISR under unknown degradation conditions, which enhances the model's generalization ability. Although GAN-based methods are capable of adding more realistic details to images, they suffer from training instability and mode collapse [2]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 183, + 546, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 546, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 546, + 482 + ], + "type": "text", + "content": "Multi-step Diffusion-based Real-ISR. Some researches [29, 44, 53, 56, 60] in recent years have utilized the powerful image priors in pre-trained T2I diffusion models [35, 39, 64] for Real-SR tasks and achieved promising results. For example, StableSR [44] balances fidelity and perceptual quality by fine-tuning the time-aware encoder and employing controllable feature wrapping. DiffBiR [29] first processes the LR image through a reconstruction network and then uses the Stable Diffusion (SD) model [39] to supplement the details. SeeSR [53] attempts to better stimulate the generative power of the SD model by extracting the semantic information in the image as a conditional guide. PASD [56] introduces a pixel-aware cross attention module to enable the diffusion model to perceive the local structure of the image at the pixel level, while using a degradation removal module to extract degradation insensitive features to guide the diffusion process along with high-level information from the image. SUPIR [60] achieves a generative and fidelity capability using negative cues [16] as well as restoration-guided sampling, while using a larger pretraining model with a larger dataset to enhance the model capability. However, all of these methods are limited by the multi-step denoising of the diffusion model, which requires 20-50 iterations in inference, resulting in an inference time that lags far behind that of GAN-based methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "One-step Diffusion-based Real-ISR. Recently, there has been a surge of interest within the academic community in one-step distillation techniques [31, 34, 40, 58, 59] for diffusion-based Real-ISR task. SinSR [48] leverages consistency preserving distillation to condense the inference steps of ResShift [61] into a single step, yet the generalization of ResShift and SinSR is constrained due to the absence of large-scale data training. AddSR [54] introduces the adversarial diffusion distillation (ADD) [40] to RealISR tasks, resulting in a comparatively effective four-step model. However, this method has a propensity to produce excessive and unnatural image details. OSEDiff [52] directly uses LQ images as the beginning of the diffusion process, and employs VSD loss [50] as a regularization technique to condense a multi-step pre-trained T2I model into a one-step Real-ISR model. However, due to the incorporation of alternating training strategies, OSEDiff may initially tend towards unreliable optimization directions, which may lead to the visual artifacts." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23175" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 545, + 329 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 329 + ], + "type": "image", + "image_path": "797b3e0fbfafedb1fb8afd5e16397e0d4c56c5df62a332467c77473a2ed8a2ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "text", + "content": "Figure 2. Pipeline overview. We train one-step Student Model " + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "text", + "content": " to map the low-quality image " + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "inline_equation", + "content": "x_{L}" + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "text", + "content": " into a more realistic one. The noisy latent " + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t}" + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "text", + "content": " sampled by DASM (Details can be found in Fig. 6.) will be fed into both the pre-trained Teacher and LoRA Model to produce the Variational Score loss. Subsequently, two Models' predictions on " + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "text", + "content": " yield Target Score loss. Their weighted forms, namely TSD (red flow), in conjunction with pixel-space reconstruction loss (green flow), are leveraged to update the Student Model. After updating the Student Model, we employ the diffusion loss (blue flow) to alternately update the LoRA Model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 392, + 129, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 129, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 129, + 406 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 415, + 133, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 133, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 133, + 426 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "text", + "content": "Problem Formulation. ISR problem aims to reconstruct a HQ image " + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "text", + "content": " from a LQ input " + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "inline_equation", + "content": "x_{L}" + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "text", + "content": " by training a parameterized ISR model " + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "text", + "content": " on a dataset " + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_{L},x_{H})_{i = 1}^{N}\\}" + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 436, + 287, + 496 + ], + "type": "text", + "content": " represents the number of image pairs. It is mathematically formulated as minimizing the following objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 499, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 499, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 70, + 499, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta^ {*} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x _ {L}, x _ {H}) \\sim \\mathcal {D}} \\left[ \\mathcal {L} _ {R e c} \\left(G _ {\\theta} \\left(x _ {L}\\right), x _ {H}\\right) \\right. \\tag {1} \\\\ \\left. \\left. + \\lambda \\mathcal {L} _ {R e g} \\left(q _ {\\theta} \\left(\\hat {x} _ {H}\\right), p \\left(x _ {H}\\right)\\right) \\right] \\right. \\\\ \\end{array}", + "image_path": "ec1c6047d1d6189764a3ab9c8dd288b81947ea6cdaff408627eef1f7baa503cc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Rec}" + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "text", + "content": " denotes the reconstruction loss, commonly measured using distance metrics like " + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "text", + "content": " or LPIPS [65]. The regularization term " + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Reg}" + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "text", + "content": " enhances the realism and generalization of the ISR model's outputs. This objective can be understood as aligning the ISR output " + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\hat{x}_H" + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "text", + "content": " 's distribution, " + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\hat{x}_H)" + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "text", + "content": ", with the true distribution " + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "inline_equation", + "content": "p(x_H)" + }, + { + "bbox": [ + 46, + 536, + 287, + 620 + ], + "type": "text", + "content": " by minimizing the KL-divergence [25]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 623, + 287, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 623, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 111, + 623, + 287, + 640 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathcal {D} _ {\\mathrm {K L}} \\left(q _ {\\theta} \\left(\\hat {x} _ {H}\\right) \\| p \\left(x _ {H}\\right)\\right) \\tag {2}", + "image_path": "1f05f446a57ec72ff8d1313628d667e5635ba866704e85d8991c7a114436a40d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "While several studies [45, 46, 62] have employed adversarial loss to optimize this objective, they often encounter issues like mode collapse and training instability. Recent work [52] achieved the state-of-the-art results using Variational Score Distillation (VSD) as the regularization loss to minimizing this objective, which inspires our research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "text", + "content": "Variational Score Distillation. Variational Score Distillation (VSD) [50] was initially introduced for text-to-3D generation by distilling a pre-trained text-to-image diffusion model to optimize a single 3D representation [36]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "text", + "content": "In the VSD framework, a pre-trained diffusion model, represented as " + }, + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "text", + "content": ", and its trainable (LoRA [19]) replica " + }, + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "text", + "content": " are used to regularize the generator network " + }, + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "text", + "content": ". As outlined in ProlificDreamer [50], the gradient with respect to the generator parameters " + }, + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 442, + 545, + 502 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 506, + 545, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 506, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 320, + 506, + 545, + 549 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {V S D}} (\\hat {\\boldsymbol {z}}, c _ {y}) \\\\ = \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\left(\\boldsymbol {\\epsilon} _ {\\psi} (\\hat {\\boldsymbol {z}} _ {t}; t, c _ {y}) - \\boldsymbol {\\epsilon} _ {\\phi} (\\hat {\\boldsymbol {z}} _ {t}; t, c _ {y})\\right) \\frac {\\partial \\hat {\\boldsymbol {z}}}{\\partial \\boldsymbol {\\theta}} \\right] \\tag {3} \\\\ \\end{array}", + "image_path": "057cc254d5323f5c104ded605404db7648714adac577ee9b2784b95541738d12.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}_t = \\alpha_t\\hat{\\pmb{z}} +\\sigma_t\\pmb {\\epsilon}" + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "text", + "content": " is the noisy input, " + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "text", + "content": " is the latent outputted by the generator network " + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "text", + "content": " is a gaussian noise, and " + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\alpha_{t},\\sigma_{t}" + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "text", + "content": " are the noise-data scaling constants. " + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "inline_equation", + "content": "c_{y}" + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "text", + "content": " is a text embedding corresponding to a caption that describes the input image, and " + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 304, + 553, + 545, + 615 + ], + "type": "text", + "content": " is a time-varying weighting function." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 622, + 427, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 427, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 427, + 634 + ], + "type": "text", + "content": "3.2. Overview of TSD-SR" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "As depicted in Fig. 2, our goal is to distill a given pretrained T2I DMs into a fast one-step Student Model " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " with Teacher Model " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " and trainable LoRA Model " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": ". We denote the latent outputs of the distilled model as " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": ", and HQ latent representations as " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " will pass through our Distribution-Aware Sampling Module (DASM)" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23176" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 286, + 210 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 210 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 210 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 210 + ], + "type": "image", + "image_path": "29c9948cfc3145c1a6289ec20654a3b42a21fe596a103deb8ff174ce8ada520d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 214, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 280 + ], + "type": "text", + "content": "Figure 3. A visual comparison of the gradient direction. We set the timestep " + }, + { + "bbox": [ + 46, + 214, + 287, + 280 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 214, + 287, + 280 + ], + "type": "text", + "content": " to 100 and calculated the cosine similarity between the predicted directions from the Teacher Model and the true direction (towards the HQ data). The prediction direction for " + }, + { + "bbox": [ + 46, + 214, + 287, + 280 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 214, + 287, + 280 + ], + "type": "text", + "content": " closely matches the true direction, but not for " + }, + { + "bbox": [ + 46, + 214, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t}" + }, + { + "bbox": [ + 46, + 214, + 287, + 280 + ], + "type": "text", + "content": ", suggesting that suboptimal samples may lead to directional deviations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 302, + 286, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 286, + 398 + ], + "type": "text", + "content": "to obtain distribution-based samples " + }, + { + "bbox": [ + 46, + 302, + 286, + 398 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 46, + 302, + 286, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 302, + 286, + 398 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 46, + 302, + 286, + 398 + ], + "type": "text", + "content": " (Sec. 3.4). We train " + }, + { + "bbox": [ + 46, + 302, + 286, + 398 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 302, + 286, + 398 + ], + "type": "text", + "content": " by minimizing the two losses: a reconstruction loss in pixel space to compare the model outputs against the ground truth, and a regularization loss (from Target Score Distillation) to enhances the realism (Sec. 3.3). After updating the Student Model, we update the LoRA Model with diffusion loss. Finally, in Sec. 3.5, we present an overview of all the losses encountered during the training phase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 183, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 183, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 183, + 419 + ], + "type": "text", + "content": "3.3. Target Score Distillation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": "Similar to [52], we introduce VSD loss into our work as a regularization loss to enhance the realism and generalization of the " + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": " outputs. Upon reviewing VSD Eq. (3), " + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\hat{z}_t;t,c_y)" + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": " represents the gradient direction of the optimization on " + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": " noisy outputs " + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": " , whereas " + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}(\\hat{z}_t;t,c_y)" + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": " corresponds to the gradient direction on noisy HQ latent representations " + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": " . The overarching goal of model optimization is to harmonize the predictions from " + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": " with those from " + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": " Nonetheless, this strategy encounters hurdles, especially in the early training phase: the quality of synthetic latents " + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 46, + 426, + 286, + 628 + ], + "type": "text", + "content": " is not high enough for Teacher Model to give a precise prediction. As illustrated in Fig. 3, the Teacher Model finds it challenging to accurately predict the optimization direction for low-quality synthetic latents, with the cosine similarity to the ideal optimization direction only 0.2 (0.88 on HQ latents). This problem can lead to severe visual artifacts, as evident in Fig. 4(a)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "A straightforward remedial measure is to employ a mean squared error (MSE) loss to align the synthetic latents with the ideal inputs of the Teacher Model, which are derived from HQ latents. However, as shown in Fig. 4(b), this approach has been observed to lead to over-smoothed results [13]. Our strategy, instead, is to align the predictions made by the Teacher Model on both synthetic and HQ" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 384, + 178 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 384, + 178 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 384, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 384, + 178 + ], + "type": "image", + "image_path": "10d97d354a5ad2a357932d8ac64204b165d462476582967da01ee481ff54b6b0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 180, + 361, + 189 + ], + "lines": [ + { + "bbox": [ + 331, + 180, + 361, + 189 + ], + "spans": [ + { + "bbox": [ + 331, + 180, + 361, + 189 + ], + "type": "text", + "content": "(a) Naive" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 388, + 70, + 465, + 178 + ], + "blocks": [ + { + "bbox": [ + 388, + 70, + 465, + 178 + ], + "lines": [ + { + "bbox": [ + 388, + 70, + 465, + 178 + ], + "spans": [ + { + "bbox": [ + 388, + 70, + 465, + 178 + ], + "type": "image", + "image_path": "3e6306ca47e5f62481283641e1ba077a5bec164d6a5e866fd6ece3772e3a3de6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 180, + 441, + 189 + ], + "lines": [ + { + "bbox": [ + 411, + 180, + 441, + 189 + ], + "spans": [ + { + "bbox": [ + 411, + 180, + 441, + 189 + ], + "type": "text", + "content": "(b) MSE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 468, + 70, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 468, + 70, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 468, + 70, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 468, + 70, + 545, + 178 + ], + "type": "image", + "image_path": "1d4e80265ad8eb0f1c332342eac376b800792314140e9b65d297f2cc55340347.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 180, + 520, + 189 + ], + "lines": [ + { + "bbox": [ + 493, + 180, + 520, + 189 + ], + "spans": [ + { + "bbox": [ + 493, + 180, + 520, + 189 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 192, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 246 + ], + "type": "text", + "content": "Figure 4. The visualization of different strategies. (a) The naive method introduces fake textures and fails to recover fine details. (b) MSE leads to over-smoothed generation result, lacking high-frequency information. (c) Our method offers the superior visual effects and fine textures." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 267, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 267, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 267, + 545, + 327 + ], + "type": "text", + "content": "latents, thereby encouraging greater consistency between them. The core idea is that for samples drawn from the same distribution, the real scores predicted by the Teacher Model should be close to each other. We refer to this approach as Target Score Matching (TSM):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 332, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 331, + 332, + 545, + 374 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} _ {\\mathrm {T S M}} (\\hat {\\mathbf {z}}, \\mathbf {z}, c _ {y}) \\\\ = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\epsilon_ {\\psi} \\left(\\hat {\\mathbf {z}} _ {t}; t, c _ {y}\\right) - \\epsilon_ {\\psi} \\left(\\mathbf {z} _ {t}; t, c _ {y}\\right)\\right) \\frac {\\partial \\hat {\\mathbf {z}}}{\\partial \\theta} \\right] \\tag {4} \\\\ \\end{array}", + "image_path": "f9dab6e1262ed6119017b28fe96ccb6b2b759013c8accb9d36892e8972726454.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "text", + "content": "where the expectation of the gradient is computed across all diffusion timesteps " + }, + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "text", + "content": ". Equation (4) encapsulates the optimization loss for our Target Score Matching. Upon examining it in conjunction with the Eq. (3), we notice that VSD utilizes the prediction residual between a Teacher and a LoRA Model to drive gradient backpropagation. Conversely, our TSM employs the synthetic and the HQ data to produce the gradients. By blending these two strategies with hyperparameter weighting " + }, + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "inline_equation", + "content": "1 - \\lambda" + }, + { + "bbox": [ + 305, + 378, + 545, + 521 + ], + "type": "text", + "content": ", we can construct our ultimate optimization loss, effectively unifying the strengths of both approaches to optimize the training process Eq. (5)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 335, + 526, + 545, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 526, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 335, + 526, + 545, + 554 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {T S D}} (\\hat {\\mathbf {z}}, \\mathbf {z}, c _ {y}) = \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left[ \\epsilon_ {\\psi} \\left(\\hat {\\mathbf {z}} _ {t}; t, c _ {y}\\right) - \\right. \\right.", + "image_path": "2f7db3b7cbd93183712e04c457abbe6ded65fa08607e3270853412996f3e1112.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 548, + 545, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 548, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 318, + 548, + 545, + 572 + ], + "type": "interline_equation", + "content": "\\left. \\left. \\epsilon_ {\\psi} (\\mathbf {z} _ {t}; t, c _ {y}) + \\lambda \\left(\\epsilon_ {\\psi} \\left(\\mathbf {z} _ {t}; t, c _ {y}\\right) - \\epsilon_ {\\phi} \\left(\\hat {\\mathbf {z}} _ {t}; t, c _ {y}\\right)\\right) \\right] \\frac {\\partial \\hat {\\mathbf {z}}}{\\partial \\theta} \\right] \\tag {5}", + "image_path": "528e9325bb0d2c4658a1dc4235a3e728e87771b73fe75f79329b0736dd208f66.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 575, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 545, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 575, + 545, + 647 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 306, + 575, + 545, + 647 + ], + "type": "text", + "content": " is time-aware weighting function tailored for Real-ISR. Other symbols are in accordance with those previously mentioned. By introducing the prediction of pretrained diffusion model on HQ latents, we have circumvented the issue of the model falling into the visual artifacts or over-smoothed problem, as illustrated in Fig. 4(c)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 658, + 504, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 504, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 504, + 671 + ], + "type": "text", + "content": "3.4. Distribution-Aware Sampling Module" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": "In the VSD-based framework, we need to match the score functions predicted by the Teacher Model with those by the LoRA Model at different timesteps " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "t \\sim \\{0,1,\\dots,T\\}" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23177" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 219, + 198 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 219, + 198 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 219, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 219, + 198 + ], + "type": "image", + "image_path": "34e2f5d7aa72822b49fb6fc97431cef0b4ac1ec61c8aa52772cf1cedcbe766e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 203, + 203, + 211 + ], + "lines": [ + { + "bbox": [ + 77, + 203, + 203, + 211 + ], + "spans": [ + { + "bbox": [ + 77, + 203, + 203, + 211 + ], + "type": "text", + "content": "(a) Predicted Residual at Different Timesteps in VSD" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 71, + 286, + 131 + ], + "blocks": [ + { + "bbox": [ + 225, + 71, + 286, + 131 + ], + "lines": [ + { + "bbox": [ + 225, + 71, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 225, + 71, + 286, + 131 + ], + "type": "image", + "image_path": "2e4c1def5eba71ccb05c319f4d676ebe73146dc846ffc704abe3c00a79af3db5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 133, + 268, + 140 + ], + "lines": [ + { + "bbox": [ + 241, + 133, + 268, + 140 + ], + "spans": [ + { + "bbox": [ + 241, + 133, + 268, + 140 + ], + "type": "text", + "content": "(b) Stage 1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 143, + 286, + 203 + ], + "blocks": [ + { + "bbox": [ + 225, + 143, + 286, + 203 + ], + "lines": [ + { + "bbox": [ + 225, + 143, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 225, + 143, + 286, + 203 + ], + "type": "image", + "image_path": "1d8cbadb7f05c3880a3daec4e9187ba99c160cdeb44642c7832c2188d7da78f0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 204, + 268, + 211 + ], + "lines": [ + { + "bbox": [ + 241, + 204, + 268, + 211 + ], + "spans": [ + { + "bbox": [ + 241, + 204, + 268, + 211 + ], + "type": "text", + "content": "(c) Stage 2" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 213, + 287, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 269 + ], + "type": "text", + "content": "Figure 5. (a) The prediction errors of the VSD loss at different timesteps. The error divergence is more pronounced in early timesteps than later. This phenomenon is observed throughout the optimization process. (b) the visualization of Stage 1 prediction error. (c) the visualization of Stage 2 prediction error." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "text", + "content": "However, for the Real-ISR problem, this matching ability at different " + }, + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "text", + "content": " is unbalanced, as illustrated in Fig. 5(a). This phenomenon could be attributed to our leveraging of the low-frequency (LF) priors from LQ data, but losing guidance from high-frequency (HF) details. The noisy samples " + }, + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "text", + "content": " possesses low-frequency priors (from LQ data). These priors are easily learned by the LoRA Model, leading to more similar predictions (Fig. 5(b)) during LF restoration (Stage 1). However, In the Stage 2, the lack of HF details in " + }, + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 46, + 294, + 287, + 437 + ], + "type": "text", + "content": " causes the LoRA Model to poorly perceive details, resulting in divergent predictions (Fig. 5(c)). Our goal is to reduce this divergence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "text", + "content": "Existing method, for each iteration, only matches the score function at a single timestep sample " + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "text", + "content": ", with uniform sampling strategy of time " + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "text", + "content": ". This leads to slow convergence or even difficulty in optimization during Stage 2, as the gradients of important timesteps are averaged out. To this end, we propose our Distribution-Aware Sampling Module (DASM). This module accumulates optimization gradients for earlier timestep samples in a single iteration, enabling the backpropagation of more gradients focused on detail optimization. As shown in Fig. 6, we first introduce noise into the synthetic latent representation " + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\hat{z}_t = (1 - \\sigma_t)\\hat{z}_0 + \\sigma_t\\epsilon" + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "text", + "content": " is a weight factor and " + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 441, + 287, + 608 + ], + "type": "text", + "content": " is a gaussian noise. Subsequently, we employ a LoRA Model to perform denoising, yielding the previous time noisy samples by Eq. (6):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 613, + 287, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 613, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 83, + 613, + 287, + 625 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {\\boldsymbol {t} - 1} = \\hat {z} _ {\\boldsymbol {t}} + \\left(\\sigma_ {t - 1} - \\sigma_ {t}\\right) \\cdot \\epsilon_ {\\phi} \\left(\\hat {z} _ {\\boldsymbol {t}}; t, c _ {y}\\right), \\tag {6}", + "image_path": "f73a989da63a9591f2337bfee4a2eac4e9963d049356f6661eb90720c75b24e2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{t - 1}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " are from the flow matching scheduler, and here LoRA Model has fit " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " 's distribution. Similarly, " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " can be obtained by denoising on the Teacher Model. Ultimately, in a single iteration, the gradients from noise samples along its trajectory can be accumulated to update the Student Model. Since the obtained samples follow the diffusion sampling trajectory and are directed towards early" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 545, + 147 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 147 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 147 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 147 + ], + "type": "image", + "image_path": "40af5971cae9330009f7fe1a9334a81903a124f25e487820766a53e91c2386d4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 150, + 545, + 206 + ], + "lines": [ + { + "bbox": [ + 305, + 150, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 150, + 545, + 206 + ], + "type": "text", + "content": "Figure 6. Illustration of DASM. Top: The naive method that add noise to samples. Bottom: Our DASM leverages the priors of diffusion models to sample noise that more closely aligns with the true sampling trajectory. The noisy samples can all serve as inputs to the subsequent network to generate gradient backpropagation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 221, + 544, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 221, + 544, + 233 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 544, + 233 + ], + "type": "text", + "content": "timestep, we effectively optimize the divergence in Stage 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 240, + 416, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 416, + 253 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 416, + 253 + ], + "type": "text", + "content": "3.5. Training Objective" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 258, + 544, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 544, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 544, + 305 + ], + "type": "text", + "content": "We summarize all the losses we used in our framework. Student Model " + }, + { + "bbox": [ + 305, + 258, + 544, + 305 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 305, + 258, + 544, + 305 + ], + "type": "text", + "content": ". We train our Student Model with the reconstruction loss " + }, + { + "bbox": [ + 305, + 258, + 544, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Rec}" + }, + { + "bbox": [ + 305, + 258, + 544, + 305 + ], + "type": "text", + "content": " and the regularization loss " + }, + { + "bbox": [ + 305, + 258, + 544, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Reg}" + }, + { + "bbox": [ + 305, + 258, + 544, + 305 + ], + "type": "text", + "content": ". For the reconstruction loss, we use LPIPS loss:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 309, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 309, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 321, + 309, + 545, + 322 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {R e c}} \\left(G _ {\\theta} \\left(\\boldsymbol {x} _ {L}\\right), \\boldsymbol {x} _ {H}\\right) = \\mathcal {L} _ {\\text {L P I P S}} \\left(G _ {\\theta} \\left(\\boldsymbol {x} _ {L}\\right), \\boldsymbol {x} _ {H}\\right). \\tag {7}", + "image_path": "1c633deeec3d4f2bff84bc6c28dc918f11372efc6eb362144f7f0993a902e81d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 324, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 360 + ], + "type": "text", + "content": "For the regularization loss, we use our TSD loss Eq. (5). Therefore, the overall training objective for the Student Model " + }, + { + "bbox": [ + 305, + 324, + 545, + 360 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 305, + 324, + 545, + 360 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 376, + 363, + 545, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 363, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 376, + 363, + 545, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S t u}} = \\mathcal {L} _ {\\text {R e c}} + \\gamma \\mathcal {L} _ {\\text {R e g}}, \\tag {8}", + "image_path": "79d6c7405a0a8adc50c846759b2849d2487b589e216260f654a4f9d168e6571c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 379, + 427, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 379, + 427, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 427, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 379, + 427, + 389 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 379, + 427, + 389 + ], + "type": "text", + "content": " is a weighting scalar." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 390, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 390, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 390, + 545, + 426 + ], + "type": "text", + "content": "LoRA Model " + }, + { + "bbox": [ + 305, + 390, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 305, + 390, + 545, + 426 + ], + "type": "text", + "content": ". As stipulated by VSD, the fine-tuned replica " + }, + { + "bbox": [ + 305, + 390, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 305, + 390, + 545, + 426 + ], + "type": "text", + "content": " must be trainable, with its training objective being:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 336, + 430, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 430, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 336, + 430, + 545, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D i f f} (\\hat {\\boldsymbol {z}}, c _ {y}) = \\mathbb {E} _ {t, \\epsilon} \\left[ \\left\\| \\boldsymbol {\\epsilon} _ {\\phi} \\left(\\hat {\\boldsymbol {z}} _ {t}; t, c _ {y}\\right) - \\boldsymbol {\\epsilon} ^ {\\prime} \\right\\| ^ {2} \\right], \\tag {9}", + "image_path": "feabd14daa50f199007b921dc599564a42fa3b1377980f5f4ab33328e345ed6a.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\epsilon^{\\prime}" + }, + { + "bbox": [ + 305, + 449, + 545, + 485 + ], + "type": "text", + "content": " serves as the training target for the denoising network, representing Gaussian noise in the context of DDPM, and a gradient towards HQ for flow matching." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 496, + 387, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 387, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 387, + 509 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 515, + 432, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 432, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 432, + 528 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 533, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 582 + ], + "type": "text", + "content": "Training Datasets. We adopt DIV2K [1], Flickr2K [42], LSDIR [27] and first 10K face images from FFHQ [20] for training. We use the same degradation pipeline as Real-ESRGAN [46] to synthesize LR-HR pairs." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 582, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 630 + ], + "type": "text", + "content": "Test Datasets. We evaluated our model on synthetic DIV2K-Val [1] dataset and two real-world datasets, RealSR [4] and DRealSR [51]. The real-world datasets comprised 128x128 LQ and 512x512 HQ image pairs." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. For evaluating our method, we apply both full-reference and no-reference metrics. Full-reference metrics include PSNR and SSIM [49] (calculated on the Y channel in YCbCr space) for fidelity, LPIPS [65] and DISTS [7] for perceptual quality, and FID [14] for distribution comparison. No-reference metrics include NIQE [63], MANIQA[55], MUSIQ [22] and CLIPIQA [43]." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23178" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 95, + 547, + 255 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with the state-of-the-art one-step methods across both synthetic and real-world benchmarks. The number of diffusion inference steps is indicated by 's'. The best results of each metric are highlighted in red." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 95, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 95, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 95, + 547, + 255 + ], + "type": "table", + "html": "
DatasetsMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
DRealSROSEDiff-1s27.920.78360.29680.2162135.516.447164.690.58980.6958
AddSR-1s27.770.77220.31960.2242150.186.932160.850.54900.6188
SinSR-1s28.380.74970.36690.2484172.726.960655.030.49040.6412
Ours-1s27.770.75590.29670.2136134.985.913166.620.58740.7344
RealSROSEDiff-1s25.150.73410.29200.2128123.485.647169.100.63260.6687
AddSR-1s24.790.70770.30910.2191132.055.544066.180.60980.5722
SinSR-1s26.270.73510.32170.2341137.596.296460.760.54180.6163
Ours-1s24.810.71720.27430.2104114.455.129871.190.63470.7160
DIV2K-ValOSEDiff-1s23.720.61090.29420.197526.344.708967.960.61310.6681
AddSR-1s23.260.59020.36230.212329.684.761063.390.56570.5734
SinSR-1s24.410.60180.32620.206835.555.998162.950.54300.6501
Ours-1s23.020.58080.26730.182129.164.324471.690.61920.7416
", + "image_path": "2546b7db10ec64dd974cd4a31954a4ef3925ad43b4284536dbf560b6961844e5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 274, + 289, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 289, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 289, + 393 + ], + "type": "text", + "content": "Compared Methods. We categorize the test models into two groups: single-step and multi-step inference. The single-step inference diffusion models include SinSR [48], AddSR [54], and OSEDiff [52]. The multi-step inference diffusion models comprise StableSR [44], ResShift [61], PASD [56], DiffBIR [29], SeeSR [53], SUPIR [60], and AddSR [54]. Specifically, for AddSR, we have conducted comparisons between its single-step and four-step models. GAN-based Real-ISR methods [5, 28, 46, 62] are detailed in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 395, + 288, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 288, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 288, + 526 + ], + "type": "text", + "content": "Implementation Details. All models mentioned are initialized from the Teacher Model (SD3 [10] in our work). We only train vae encoder and denoising network in the Student Model, frozen the vae decoder in order to preserve vae's prior [24]. We utilize the default prompt for the Student Model, and prompts extracted from HQ for the Teacher and LoRA models when training. We use the AdamW optimizer [30] with a learning rate of 5e-6 for the Student Model and 1e-6 for the LoRA Model, and set rank of LoRA to 64 for both two models. The training process took roughly 120h, utilizing 8 NVIDIA V100 GPUs with a batch size of 16." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 538, + 237, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 237, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 237, + 551 + ], + "type": "text", + "content": "4.2. Comparison with Existing Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Quantitative Comparisons. Tab. 1 show the quantitative comparison of our method with single-step diffusion models on three datasets. Ours achieves the best results on most evaluation metrics. SinSR and AddSR, as distilled versions of previous multi-step super-resolution methods, reduce inference steps but experience a corresponding decrease in performance metrics. OSEDiff introduces the VSD loss from 3D generation tasks into Real-ISR without fully addressing the substantial differences between the two domains. Therefore, the no-reference metrics for image quality are not very satisfied. In contrast, our proposed TSD-SR, tailored for Real-ISR, outperform other single-step models in terms of the vast majority of key metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 274, + 547, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 547, + 441 + ], + "type": "text", + "content": "Tab. 2 show the quantitative comparison with multi-step models. We can draw the following conclusions: (1) TSD-SR demonstrates significant advantages over competing methods in terms of LPIPS, DISTS, and NIQE metrics. Additionally, it achieves performance that surpasses most multi-step models on FID, MUSIQ, and CLIPIQA. (2) DiffBIR, SeeSR, PASD, and AddSR exhibit better performance on the MANIQA metric, which may be attributed to the fact that multi-step models have more denoising iterations to produce rich details. (3) ResShift stands out with the highest PSNR and SSIM scores, while StableSR also shows notable performance regarding DISTS and FID metrics. However, both two models underperform in terms of no-reference metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 445, + 547, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 530 + ], + "type": "text", + "content": "Finally, we explain the lower PSNR and SSIM metrics in our experiments. Several works [54, 60] have found that these reconstruction metrics are not well-suited for evaluating Real-ISR tasks. When a model recovers better details, it leads to lower reconstruction metrics, indicating a trade-off [3, 32, 67]. Refer to the supplementary material for detailed visual comparisons and analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "content": "Qualitative Comparisons. Fig. 7 present visual comparisons of different Real-ISR methods. As shown in multi-step methods' results, SeeSR uses degradation-aware semantic cues to leverage image generation priors, but it sometimes produces over-smooth textures. SUPIR demonstrates a notably robust generative capacity, however, the overproduction of extraneous details will lead to results that lack naturalism in image restoration (e.g. Adding excessive wrinkles at the corners of a young girl's eyes). Under more realistic degradation conditions, PASD finds it challenging to restore the corresponding content, revealing a deficiency in the model's robustness capabilities. Among single-step methods, SinSR introduces artifacts, possibly because its distilled pre-trained model is trained from scratch, lacking sufficient real-world priors, which results in unsatisfac" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23179" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 96, + 547, + 376 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with state-of-the-art multi-step methods across both synthetic and real-world benchmarks. The number of diffusion inference steps is indicated by 's'. The best and second best results of each metric are highlighted in red and blue, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 96, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 47, + 96, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 547, + 376 + ], + "type": "table", + "html": "
DatasetsMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓FID ↓NIQE ↓MUSIQ ↑MANIQA ↑CLIPiQA ↑
DRealSRStableSR-200s28.040.74540.32790.2272144.156.599958.530.56030.6250
DiffBIR-50s25.930.65250.45180.2761177.046.232465.660.62960.6860
SeeSR-50s28.140.77120.31410.2297146.956.463264.740.60220.6893
SUPIR-50s25.090.64600.42430.2795169.487.391858.790.54710.6749
PASD-20s27.790.74950.35790.2524171.036.766163.230.59190.6242
ResShift-15s28.690.78740.35250.2541176.777.876252.400.47560.5413
AddSR-4s26.720.71240.39820.2711164.127.668966.330.62570.7226
Ours-1s27.770.75590.29670.2136134.985.913166.620.58740.7344
RealSRStableSR-200s24.620.70410.30700.2156128.545.781765.480.62230.6198
DiffBIR-50s24.240.66500.34690.2300134.565.493268.350.65440.6961
SeeSR-50s25.210.72160.30030.2218125.105.397869.690.64430.6671
SUPIR-50s23.650.66200.35410.2488130.386.109962.090.57800.6707
PASD-20s25.680.72730.31440.2304134.185.761668.330.63230.5783
ResShift-15s26.390.75670.31580.2432149.596.874660.220.54190.5496
AddSR-4s23.330.64000.39250.2626154.225.895971.490.68260.7225
Ours-1s24.810.71720.27430.2104114.455.129871.190.63470.7160
DIV2K-ValStableSR-200s23.270.57220.31110.204624.954.773765.780.61640.6753
DiffBIR-50s23.130.57170.34690.210833.934.605668.540.63600.7125
SeeSR-50s23.730.60570.31980.195325.814.832268.490.61980.6899
SUPIR-50s22.130.52790.39190.231231.405.676763.860.59030.7146
PASD-20s24.000.60410.37790.230539.124.858767.360.61210.6327
ResShift-15s24.710.62340.34730.225342.016.361560.630.52830.5962
AddSR-4s22.160.52800.40530.236035.415.258470.990.65960.7593
Ours-1s23.020.58080.26730.182129.164.324471.690.61920.7416
", + "image_path": "69eb288ba70f26d92a7c12a267e37d3eaa3d4ee3a86fb302b44eabe322651db4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 409, + 546, + 454 + ], + "blocks": [ + { + "bbox": [ + 47, + 385, + 545, + 408 + ], + "lines": [ + { + "bbox": [ + 47, + 385, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 545, + 408 + ], + "type": "text", + "content": "Table 3. Comparison of computational complexity across different DMs-based methods. The performance of each method is measured on an A100 GPU, using input images sized at " + }, + { + "bbox": [ + 47, + 385, + 545, + 408 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 47, + 385, + 545, + 408 + ], + "type": "text", + "content": " pixels. We have disregarded the loading time for model weights and data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 409, + 546, + 454 + ], + "lines": [ + { + "bbox": [ + 48, + 409, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 546, + 454 + ], + "type": "table", + "html": "
StableSRDiffBIRSeeSRSUPIRPASDResShiftAddSRAddSRSinSROSEDiffOurs
Inference Step200505050201541111
Inference Time12.41517.96375.816716.87044.84410.75461.01990.50430.14240.16500.1362
", + "image_path": "bd8eefaf64f2bdfcd3807b590e714a80bc3907309157ce3c865b6a8e976c9f7c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": "tory image restoration capabilities. AddSR produces oversmoothed results when using its 1-step model. OSEDiff offers improved restoration effects compared to SinSR and AddSR, yet it may fall short in terms of authenticity and naturalness when it comes to detail recovery. In contrast, our method effectively generates rich textures and realistic details with enhanced sharpness and contrast. Additional visual comparisons and results are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Complexity Comparisons We assess the complexity of the state-of-the-art DM-based Real-ISR models as detailed in Tab. 3, focusing on inference time. The performance of each method is benchmarked on an A100 GPU with input images sized at " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " pixels. We have disregarded the loading time for model weights and data, and the main computation time includes: (1) text extraction time when a text extractor is used; (2) text encoder computation time if needed; (3) VAE encoding and decoding time; (4) denoising network execution time. It is evident that TSD-SR has a substantial advantage in inference time compared with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 473, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 569 + ], + "type": "text", + "content": "multi-step models. Specifically, TSD-SR is more than 120 times faster than SUPIR, 90 times faster than StableSR, about 50 times faster than DiffBIR, more than 40 times faster than SeeSR, more than 35 times faster than PASD, and 4 times faster than ResShift. When compared with existing one-step models, our model boasts the fastest inference times. This is because we directly denoise from LQ data and use a fixed prompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 575, + 381, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 381, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 381, + 588 + ], + "type": "text", + "content": "4.3. User Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "We conduct a user study comparing our method with three other diffusion-based one-step super-resolution methods. To make the evaluation more comprehensive, we selected images from five categories—human faces, buildings, animals, vegetation and characters. A total of 50 participants were engaged in the voting process. We guided participants to evaluate the best restoration results based on the similarity to HQ image, structural similarity to the LQ image, and realism of textures and details. The results shown in the Fig. 8 indicate that our method receives a " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "69.2\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " approval" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23180" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 184, + 205 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 184, + 205 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 184, + 205 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 184, + 205 + ], + "type": "image", + "image_path": "fe86e8eee002cf877a96d586a8a3857aed4c08c99aa54c190f5b5fff477507e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 204, + 122, + 211 + ], + "lines": [ + { + "bbox": [ + 111, + 204, + 122, + 211 + ], + "spans": [ + { + "bbox": [ + 111, + 204, + 122, + 211 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 72, + 274, + 134 + ], + "blocks": [ + { + "bbox": [ + 186, + 72, + 274, + 134 + ], + "lines": [ + { + "bbox": [ + 186, + 72, + 274, + 134 + ], + "spans": [ + { + "bbox": [ + 186, + 72, + 274, + 134 + ], + "type": "image", + "image_path": "e022902d11bee8f3cc7b6daf54e816099a84ecec0461b2c3391a98c4ab84827c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 134, + 249, + 141 + ], + "lines": [ + { + "bbox": [ + 211, + 134, + 249, + 141 + ], + "spans": [ + { + "bbox": [ + 211, + 134, + 249, + 141 + ], + "type": "text", + "content": "Zoomed LR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 277, + 72, + 363, + 134 + ], + "blocks": [ + { + "bbox": [ + 277, + 72, + 363, + 134 + ], + "lines": [ + { + "bbox": [ + 277, + 72, + 363, + 134 + ], + "spans": [ + { + "bbox": [ + 277, + 72, + 363, + 134 + ], + "type": "image", + "image_path": "29755036bfcbd3f83680a23116c0dcbf7eb9139d41d75118d907060c3e3ed45d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 134, + 336, + 140 + ], + "lines": [ + { + "bbox": [ + 303, + 134, + 336, + 140 + ], + "spans": [ + { + "bbox": [ + 303, + 134, + 336, + 140 + ], + "type": "text", + "content": "SeeSR-50s" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 366, + 72, + 454, + 134 + ], + "blocks": [ + { + "bbox": [ + 366, + 72, + 454, + 134 + ], + "lines": [ + { + "bbox": [ + 366, + 72, + 454, + 134 + ], + "spans": [ + { + "bbox": [ + 366, + 72, + 454, + 134 + ], + "type": "image", + "image_path": "724cc0fbdcf19a5d9301f4d3c96ad6eb56cdd9aff8cca5910905a46df375b324.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 134, + 427, + 140 + ], + "lines": [ + { + "bbox": [ + 392, + 134, + 427, + 140 + ], + "spans": [ + { + "bbox": [ + 392, + 134, + 427, + 140 + ], + "type": "text", + "content": "SUPIR-50s" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 457, + 72, + 544, + 134 + ], + "blocks": [ + { + "bbox": [ + 457, + 72, + 544, + 134 + ], + "lines": [ + { + "bbox": [ + 457, + 72, + 544, + 134 + ], + "spans": [ + { + "bbox": [ + 457, + 72, + 544, + 134 + ], + "type": "image", + "image_path": "8ca7be1e537bf4cb4134a6edca8f7c880544686fe04fd2fba33bb0fb4f7cd7b5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 134, + 517, + 140 + ], + "lines": [ + { + "bbox": [ + 484, + 134, + 517, + 140 + ], + "spans": [ + { + "bbox": [ + 484, + 134, + 517, + 140 + ], + "type": "text", + "content": "PASD-20s" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 213, + 183, + 346 + ], + "blocks": [ + { + "bbox": [ + 50, + 213, + 183, + 346 + ], + "lines": [ + { + "bbox": [ + 50, + 213, + 183, + 346 + ], + "spans": [ + { + "bbox": [ + 50, + 213, + 183, + 346 + ], + "type": "image", + "image_path": "db129735dea7555c006e0682c6e383ce8a935e78b4e0b39f1a0f6c61536d46c7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 346, + 122, + 352 + ], + "lines": [ + { + "bbox": [ + 111, + 346, + 122, + 352 + ], + "spans": [ + { + "bbox": [ + 111, + 346, + 122, + 352 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 186, + 141, + 274, + 204 + ], + "blocks": [ + { + "bbox": [ + 186, + 141, + 274, + 204 + ], + "lines": [ + { + "bbox": [ + 186, + 141, + 274, + 204 + ], + "spans": [ + { + "bbox": [ + 186, + 141, + 274, + 204 + ], + "type": "image", + "image_path": "bb42b0e29977ac9962f67489b65fb3a25a986c136b11684fe6fc6e5575d6e4a1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 204, + 244, + 211 + ], + "lines": [ + { + "bbox": [ + 215, + 204, + 244, + 211 + ], + "spans": [ + { + "bbox": [ + 215, + 204, + 244, + 211 + ], + "type": "text", + "content": "SinSR-1s" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 277, + 141, + 363, + 204 + ], + "blocks": [ + { + "bbox": [ + 277, + 141, + 363, + 204 + ], + "lines": [ + { + "bbox": [ + 277, + 141, + 363, + 204 + ], + "spans": [ + { + "bbox": [ + 277, + 141, + 363, + 204 + ], + "type": "image", + "image_path": "2b71ce843efcf03611aca132fa8446e5810df1aec80e367e257c43cfb6b64ccf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 205, + 335, + 211 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 335, + 211 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 335, + 211 + ], + "type": "text", + "content": "AddSR-1s" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 366, + 141, + 454, + 204 + ], + "blocks": [ + { + "bbox": [ + 366, + 141, + 454, + 204 + ], + "lines": [ + { + "bbox": [ + 366, + 141, + 454, + 204 + ], + "spans": [ + { + "bbox": [ + 366, + 141, + 454, + 204 + ], + "type": "image", + "image_path": "22930d09d16fd6160e3365cabf60ad34209152b69b7b0717575dfeafeb046044.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 204, + 427, + 211 + ], + "lines": [ + { + "bbox": [ + 392, + 204, + 427, + 211 + ], + "spans": [ + { + "bbox": [ + 392, + 204, + 427, + 211 + ], + "type": "text", + "content": "OSEDiff-1s" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 457, + 141, + 544, + 204 + ], + "blocks": [ + { + "bbox": [ + 457, + 141, + 544, + 204 + ], + "lines": [ + { + "bbox": [ + 457, + 141, + 544, + 204 + ], + "spans": [ + { + "bbox": [ + 457, + 141, + 544, + 204 + ], + "type": "image", + "image_path": "6751e74dcad40df56fea2d31e3373edb9434e45be3d4c1e293811fad8ff594d8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 205, + 512, + 211 + ], + "lines": [ + { + "bbox": [ + 487, + 205, + 512, + 211 + ], + "spans": [ + { + "bbox": [ + 487, + 205, + 512, + 211 + ], + "type": "text", + "content": "Ours-1s" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 50, + 346, + 183, + 453 + ], + "blocks": [ + { + "bbox": [ + 50, + 346, + 183, + 453 + ], + "lines": [ + { + "bbox": [ + 50, + 346, + 183, + 453 + ], + "spans": [ + { + "bbox": [ + 50, + 346, + 183, + 453 + ], + "type": "image", + "image_path": "a17b5e00b02004a98fb29cedd24f72e34b827b63b135b57b39ff18c0599f3c63.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 356, + 466, + 366 + ], + "lines": [ + { + "bbox": [ + 125, + 356, + 466, + 366 + ], + "spans": [ + { + "bbox": [ + 125, + 356, + 466, + 366 + ], + "type": "text", + "content": "Figure 7. Visual comparisons of different Real-ISR methods. Please zoom in for a better view." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 472, + 287, + 518 + ], + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 518 + ], + "type": "text", + "content": "Figure 8. The results of our user study. Left: Category-based user preference radar chart. Our model has won the most user favor across all categories. Right: User preference pie chart. Our approach has garnered a " + }, + { + "bbox": [ + 46, + 472, + 287, + 518 + ], + "type": "inline_equation", + "content": "69.2\\%" + }, + { + "bbox": [ + 46, + 472, + 287, + 518 + ], + "type": "text", + "content": " user satisfaction rating." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 186, + 213, + 273, + 276 + ], + "blocks": [ + { + "bbox": [ + 186, + 213, + 273, + 276 + ], + "lines": [ + { + "bbox": [ + 186, + 213, + 273, + 276 + ], + "spans": [ + { + "bbox": [ + 186, + 213, + 273, + 276 + ], + "type": "image", + "image_path": "46fc5e3c8340a63b9244833eb0082e595aa016f960d93afd8bb38a3fdb0207d1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 276, + 248, + 282 + ], + "lines": [ + { + "bbox": [ + 211, + 276, + 248, + 282 + ], + "spans": [ + { + "bbox": [ + 211, + 276, + 248, + 282 + ], + "type": "text", + "content": "Zoomed LR" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 277, + 213, + 363, + 276 + ], + "blocks": [ + { + "bbox": [ + 277, + 213, + 363, + 276 + ], + "lines": [ + { + "bbox": [ + 277, + 213, + 363, + 276 + ], + "spans": [ + { + "bbox": [ + 277, + 213, + 363, + 276 + ], + "type": "image", + "image_path": "eb947c9f6cb1c0d1cd9ed3183fe61f2291afc5fd1fb4d3ca5dcf7ffe6d107431.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 276, + 336, + 282 + ], + "lines": [ + { + "bbox": [ + 303, + 276, + 336, + 282 + ], + "spans": [ + { + "bbox": [ + 303, + 276, + 336, + 282 + ], + "type": "text", + "content": "SeeSR-50s" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 366, + 213, + 454, + 276 + ], + "blocks": [ + { + "bbox": [ + 366, + 213, + 454, + 276 + ], + "lines": [ + { + "bbox": [ + 366, + 213, + 454, + 276 + ], + "spans": [ + { + "bbox": [ + 366, + 213, + 454, + 276 + ], + "type": "image", + "image_path": "f20b4635ecbab927ca453b53f34a847e718a86647422f9d4b767c74614a36ed4.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 276, + 427, + 282 + ], + "lines": [ + { + "bbox": [ + 392, + 276, + 427, + 282 + ], + "spans": [ + { + "bbox": [ + 392, + 276, + 427, + 282 + ], + "type": "text", + "content": "SUPIR-50s" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 457, + 213, + 544, + 276 + ], + "blocks": [ + { + "bbox": [ + 457, + 213, + 544, + 276 + ], + "lines": [ + { + "bbox": [ + 457, + 213, + 544, + 276 + ], + "spans": [ + { + "bbox": [ + 457, + 213, + 544, + 276 + ], + "type": "image", + "image_path": "66a01f6b27aa7445518e436fefca1e27781a12d9ba01655fba9656efcfadb050.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 276, + 516, + 282 + ], + "lines": [ + { + "bbox": [ + 484, + 276, + 516, + 282 + ], + "spans": [ + { + "bbox": [ + 484, + 276, + 516, + 282 + ], + "type": "text", + "content": "PASD-20s" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 186, + 283, + 273, + 346 + ], + "blocks": [ + { + "bbox": [ + 186, + 283, + 273, + 346 + ], + "lines": [ + { + "bbox": [ + 186, + 283, + 273, + 346 + ], + "spans": [ + { + "bbox": [ + 186, + 283, + 273, + 346 + ], + "type": "image", + "image_path": "fd1065f8c6b713bb488db295345507232575856b31f4b66002404d7d1e1c1688.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 346, + 244, + 353 + ], + "lines": [ + { + "bbox": [ + 215, + 346, + 244, + 353 + ], + "spans": [ + { + "bbox": [ + 215, + 346, + 244, + 353 + ], + "type": "text", + "content": "SinSR-1s" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 277, + 283, + 363, + 346 + ], + "blocks": [ + { + "bbox": [ + 277, + 283, + 363, + 346 + ], + "lines": [ + { + "bbox": [ + 277, + 283, + 363, + 346 + ], + "spans": [ + { + "bbox": [ + 277, + 283, + 363, + 346 + ], + "type": "image", + "image_path": "8f5629488e3008b2e0c386e0d231fff50a34cf4906df7c13eb762bf1cf6971a6.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 346, + 335, + 353 + ], + "lines": [ + { + "bbox": [ + 304, + 346, + 335, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 335, + 353 + ], + "type": "text", + "content": "AddSR-1s" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 366, + 283, + 454, + 346 + ], + "blocks": [ + { + "bbox": [ + 366, + 283, + 454, + 346 + ], + "lines": [ + { + "bbox": [ + 366, + 283, + 454, + 346 + ], + "spans": [ + { + "bbox": [ + 366, + 283, + 454, + 346 + ], + "type": "image", + "image_path": "5bf3f251f012bc7b8fc7a1fa65974609daf6e8ec40be3c4d6f82637b47642561.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 346, + 428, + 353 + ], + "lines": [ + { + "bbox": [ + 392, + 346, + 428, + 353 + ], + "spans": [ + { + "bbox": [ + 392, + 346, + 428, + 353 + ], + "type": "text", + "content": "OSEDiff-1s" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 457, + 283, + 544, + 346 + ], + "blocks": [ + { + "bbox": [ + 457, + 283, + 544, + 346 + ], + "lines": [ + { + "bbox": [ + 457, + 283, + 544, + 346 + ], + "spans": [ + { + "bbox": [ + 457, + 283, + 544, + 346 + ], + "type": "image", + "image_path": "edb0c465d072a5e7f438e452f5771433c626607d9a238543b31eaeb4e87f9212.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 346, + 513, + 353 + ], + "lines": [ + { + "bbox": [ + 488, + 346, + 513, + 353 + ], + "spans": [ + { + "bbox": [ + 488, + 346, + 513, + 353 + ], + "type": "text", + "content": "Ours-1s" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 50, + 380, + 164, + 453 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 380, + 164, + 453 + ], + "spans": [ + { + "bbox": [ + 50, + 380, + 164, + 453 + ], + "type": "image", + "image_path": "6cb71981c16bf60ca83c3de03c0873d502ef4e30313559d436698dd58c856777.jpg" + } + ] + } + ], + "index": 38, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 172, + 381, + 286, + 453 + ], + "blocks": [ + { + "bbox": [ + 172, + 381, + 286, + 453 + ], + "lines": [ + { + "bbox": [ + 172, + 381, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 172, + 381, + 286, + 453 + ], + "type": "image", + "image_path": "273a2e90fc2f112a8ac481c8051671bfe3cdf262cd1b18e35987f4de04a0b16f.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "table", + "bbox": [ + 48, + 548, + 283, + 622 + ], + "blocks": [ + { + "bbox": [ + 46, + 523, + 287, + 545 + ], + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 545 + ], + "type": "text", + "content": "Table 4. Ablation study of Target Score Matching loss and Distribution-Aware Sampling Module." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 548, + 283, + 622 + ], + "lines": [ + { + "bbox": [ + 48, + 548, + 283, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 283, + 622 + ], + "type": "table", + "html": "
DatasetsMethodDISTS↓MUSIQ↑MANIQA↑CLIPIQA↑
DRealSRw/o TSM0.232763.900.57490.6958
w/o DASM0.231163.560.58120.7123
Full0.213666.620.58740.7344
RealSRw/o TSM0.239768.560.63380.6987
w/o DASM0.227369.050.62730.7031
Full0.210471.190.63470.7160
", + "image_path": "28e0cb320ff0984c22d219331b5f2d1019b5cd9b8a0b1d890f7ac3a9804f54f1.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "table_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "content": "rate from users. We scored " + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "inline_equation", + "content": "57.60\\%" + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "content": " in Animals, " + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "inline_equation", + "content": "70.00\\%" + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "content": " in Buildings, " + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "inline_equation", + "content": "68.80\\%" + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "content": " in Human Faces, " + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "inline_equation", + "content": "65.20\\%" + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "content": " in Vegetation, and " + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "inline_equation", + "content": "84.40\\%" + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "content": " in Characters, surpassing other models." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 47, + 671, + 140, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 140, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 140, + 684 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "To validate the effectiveness of TSM loss and DASM, we conduct ablation studies by removing them separately in" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 373, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 506 + ], + "type": "text", + "content": "our experiments. We select DISTS, MUSIQ, MANIQA and CLIPIQA metrics for comparison, as these are critical for image quality assessment. The results are presented in Tab. 4. We draw the following conclusions: (1) The absence of TSM loss and DASM affects the metrics DISTS, MUSIQ, MANIQA and CLIPIQA. (2) The lack of TSM leads to a significant decrease in DISTS, MUSIQ and CLIPIQA metrics, possibly due to the unreliable directions in VSD causing unrealistic generations. (3) The absence of DASM results in a decline in MUSIQ and CLIPIQA, possibly due to the suboptimal optimization of details." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 306, + 516, + 458, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 458, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 458, + 529 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 536, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 631 + ], + "type": "text", + "content": "We propose TSD-SR, an effective and one-step model based on diffusion prior for Real-ISR. TSD-SR utilizes the TSD to enhance the realism of images generated by the distillation model. And it leverages DASM to sample distribution-based samples and accumulate their gradients to enhance the recovery of details. Our experiments have demonstrated that TSD-SR outperforms existing one-step Real-ISR models in both performance and inference speed." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 632, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 704 + ], + "type": "text", + "content": "limitations. Although our model boasts excellent inference speed and restoration performance, it still has a large number of model parameters compared to the past GAN/non-diffusion approaches. In the future, we plan to employ pruning or quantization methods to compress the model parameters, striving for a lightweight and efficient Real-ISR model." + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23181" + } + ] + } + ], + "index": 50 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 161 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by National Natural Science Foundation of China (62293554, U2336212), Zhejiang Provincial Natural Science Foundation of China (LD24F020007), National Key R&D Program of China (SQ2023AAA01005), Ningbo Innovation \"Yongjiang 2035\" Key Research and Development Programme (2024Z292), and Young Elite Scientists Sponsorship Program by CAST (2023QNRC001)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 185, + 106, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 106, + 197 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 106, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 205, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "type": "text", + "content": "[2] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 295, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 328 + ], + "type": "text", + "content": "[3] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 331, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 288, + 385 + ], + "type": "text", + "content": "[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 387, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 287, + 451 + ], + "type": "text", + "content": "[5] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Real-world blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1329-1338, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "text", + "content": "[6] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 511, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 511, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 511, + 287, + 554 + ], + "type": "text", + "content": "[7] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 556, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 556, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 53, + 556, + 287, + 621 + ], + "type": "text", + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, pages 184-199. Springer, 2014. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 667 + ], + "type": "text", + "content": "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Forty-first International Conference on Machine Learning, 2024. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "text", + "content": "[12] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 215 + ], + "type": "text", + "content": "[13] Xiangyu He and Jian Cheng. Revisiting 11 loss in superresolution: a probabilistic view and beyond. arXiv preprint arXiv:2201.10084, 2022. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 293 + ], + "type": "text", + "content": "[15] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 294, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 316 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "type": "text", + "content": "[18] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 472, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 513 + ], + "type": "text", + "content": "[21] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 559 + ], + "type": "text", + "content": "[22] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[23] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 624 + ], + "type": "text", + "content": "[24] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "text", + "content": "[25] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1): 79-86, 1951. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[26] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23182" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "computer vision and pattern recognition, pages 4681-4690, 2017. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 152 + ], + "type": "text", + "content": "[27] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775-1787, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 208 + ], + "type": "text", + "content": "[28] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 255 + ], + "type": "text", + "content": "[29] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 278 + ], + "type": "text", + "content": "[30] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "type": "text", + "content": "[31] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 325, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 325, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 325, + 287, + 369 + ], + "type": "text", + "content": "[32] Xiaotong Luo, Yuan Xie, Yanyun Qu, and Yun Fu. Skipdiff: Adaptive skip diffusion model for high-fidelity perceptual image super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4017-4025, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 393 + ], + "type": "text", + "content": "[33] Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 395, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 395, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 395, + 287, + 449 + ], + "type": "text", + "content": "[34] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7807-7816, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 452, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 506 + ], + "type": "text", + "content": "[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 508, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 287, + 541 + ], + "type": "text", + "content": "[36] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 543, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 576 + ], + "type": "text", + "content": "[37] Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[40] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[42] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "text", + "content": "[43] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "type": "text", + "content": "[44] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1–21, 2024. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "type": "text", + "content": "[45] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "type": "text", + "content": "[46] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "text", + "content": "[47] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 470 + ], + "type": "text", + "content": "[48] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "text", + "content": "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 514, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 568 + ], + "type": "text", + "content": "[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 635 + ], + "type": "text", + "content": "[51] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "type": "text", + "content": "[52] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "text", + "content": "[53] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23183" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[54] Rui Xie, Ying Tai, Kai Zhang, Zhenyu Zhang, Jun Zhou, and Jian Yang. Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 207 + ], + "type": "text", + "content": "[55] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[56] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[57] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4133-4141, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "text", + "content": "[58] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 409 + ], + "type": "text", + "content": "[59] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 476 + ], + "type": "text", + "content": "[60] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "text", + "content": "[61] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 577 + ], + "type": "text", + "content": "[62] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "text", + "content": "[63] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 3, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 162 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[66] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. In European conference on computer vision, pages 649-667. Springer, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[67] Yuehan Zhang, Bo Ji, Jia Hao, and Angela Yao. Perception-distortion balanced admm optimization for single-image super-resolution. In European Conference on Computer Vision, pages 108-125. Springer, 2022. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23184" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/989548d5-80da-4477-a3a5-42129f121eff_content_list.json b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/989548d5-80da-4477-a3a5-42129f121eff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..03469d5300f8e5691f893f24491a16f9812031a8 --- /dev/null +++ b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/989548d5-80da-4477-a3a5-42129f121eff_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "TSP-Mamba: The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond", + "text_level": 1, + "bbox": [ + 181, + 128, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kun Zhou1 Xinyu Lin2 Jiangbo Lu1*", + "bbox": [ + 334, + 202, + 668, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ SmartMore Corporation $^{2}$ SSE, CUHK-Shenzhen", + "bbox": [ + 300, + 220, + 696, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhoukun303808@gmail.com, linxinyu0327@gmail.com, jiangbo.lu@gmail.com", + "bbox": [ + 196, + 241, + 800, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 290, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Mamba-based frameworks have achieved substantial advancements across diverse computer vision and NLP tasks, particularly in their capacity for reasoning over long-range information with linear complexity. However, the fixed 2D-to-1D scanning pattern overlooks the local structures of an image, limiting its effectiveness in aggregating 2D spatial information. While stacking additional Mamba layers can partially address this issue, it increases parameter intensity and constrains real-time application. In this work, we reconsider the local optimal scanning path in Mamba, enhancing the rigid and uniform 1D scan through the local shortest path theory, thus creating a structure-aware Mamba suited for lightweight single-image super-resolution. Specifically, we draw inspiration from the Traveling Salesman Problem (TSP) to establish a local optimal scanning path for improved structural 2D information utilization. Here, local patch aggregation occurs in a content-adaptive manner with minimal propagation cost. TSP-Mamba demonstrates substantial improvements over existing Mamba-based and Transformer-based architectures. For example, TSP-Mamba surpasses MambaIR by up to 0.7dB in lightweight SISR, with comparable parameters and very slightly extra computational demands (1-2 GFlops for 720P images).", + "bbox": [ + 88, + 321, + 485, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 686, + 220, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning techniques have advanced single-image superresolution (SISR) tasks considerably, enhancing image reconstruction quality. Convolutional Neural Networks (CNNs) have been central to this progress, excelling at capturing local, content-dependent image patterns. However, their local focus limits their capacity to model long-range dependencies. Transformers, by contrast, are highly effective in capturing global information through their self-attention mechanism, enabling them to model non-local dependencies. Yet, their quadratic computational complexity poses significant challenges in resource-constrained environments.", + "bbox": [ + 89, + 712, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e262c5ed2eefaac09d41c89cb9e9e4b2baf6d109607d0b6a440bd20854e60d7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 290, + 874, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3e4d1f00c614483fc8b1c4c5aacbe50a01ed17603d9c6eed075c0a8b13057c5.jpg", + "image_caption": [ + "Figure 1. (1) Comparison between our proposed TSP-Mamba and SOTA SISR models on BSD100 [27] for $\\times 2$ setting. TSP-Mamba series ('S'- 'Small', 'T'- 'Tiny') demonstrates highly competitive results. (2) Comparison of Effective Receptive Field for several SOTA methods. Darker area denotes larger receptive fields. Our proposed TSP-Mamba achieves the best global receptive field. Notably, larger circles indicate more model parameters (the first figure)." + ], + "image_footnote": [], + "bbox": [ + 519, + 482, + 596, + 547 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8d49b319205b9334996d91f908dee7525683ef9ce8689d863218b0f02550dd55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 482, + 684, + 549 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/705f6488904d86c70e2030088ab569888459dbcdaab91e03f9da02a27051e02f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 481, + 774, + 549 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d6e74f302e9322baa43167dd5d8cd9bbb6b1aa82412d485873697cbf5181e2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 481, + 903, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Mamba method has recently gained attention for its capacity to model global signals with linear computational complexity. Its fixed scanning approach is well-suited for efficient 1D sequence modeling, achieving strong computational performance. However, this approach overlooks 2D structures within images, leading to inefficient aggregation of 2D information. As illustrated in Fig. 2, conventional vision-Mambas typically rely on partial neighboring propagation and aggregation along a fixed 1D scan path, which can disrupt the locality of 2D image information—referred to as \"local pixel forgetting\" in [14] and \"destroying local correlations\" in [34]—resulting in sub-optimal outcomes. As shown in Fig. 2(a), the local information aggregation with an image patch as highlighted in the red box, is less effective in Mamba-based scan due to the long-range (also partial)", + "bbox": [ + 511, + 672, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28134", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2993a053834e6b728e4db49b991a7170eb369744dce0a2cd5f092fb877fac4c0.jpg", + "image_caption": [ + "Figure 2. Illustration of vanilla Mamba and our proposed locally enhanced TSP-Mamba. For a highlighted $3 \\times 3$ image patch (red box): (a) In Vanilla Mamba, only a limited set of neighboring pixels (e.g., '1, 2, 3, 4') can be aggregated for pixel '5' through a mix of long- and short-range propagation within a single global scan (\"Scan-1\"). (b) In TSP-Mamba, the shortest traversal path is constructed, allowing patch-adaptive propagation in a single local scan." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 289, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cac057b1c8b8dd7a4979044468e19a9d2762c444cd2359c078615aeadc80b2a0.jpg", + "image_caption": [ + "(a) 4 scan types in MambaIR" + ], + "image_footnote": [], + "bbox": [ + 299, + 89, + 547, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/debd949819ca88cd310f0e27a873d8eabf6ee80cef3a4e6d8bf04dbd72174173.jpg", + "image_caption": [ + "(b) Shortest traversal path search in TSP-Mamba" + ], + "image_footnote": [], + "bbox": [ + 547, + 90, + 898, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "propagation within a single scan. A common solution to this problem is stacking multiple multidirectional Mamba layers to improve aggregation performance. However, this will substantially increase computational cost and model complexity, hindering practicality in resource-limited scenarios.", + "bbox": [ + 88, + 344, + 483, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose TSP-Mamba, which enhances the propagation and aggregation of local 2D information by searching for a locally adaptive, optimal propagation path for each image patch. Unlike the fixed (content-invariant) scan pipeline used in traditional Mamba, our method first identifies a patch-adaptive scan path with the lowest propagation cost, leveraging the shortest path principle. Here, each pixel in a local image patch is treated as a node in a graph, and we frame the problem as a simplified Traveling Salesman Problem (TSP) to find the minimum-cost traversal path, along which local information is transmitted and aggregated to enhance the use and extraction of local structures. Later on, the enhanced local information is effectively propagated into the global Mamba-style aggregation in a conventionally local-then-global manner.", + "bbox": [ + 88, + 422, + 483, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To implement this adaptive propagation, we design a TSP-LSTM block that modulates the aggregation between the current node and the previously accumulated information based on the propagation cost (i.e., the distance between the current and previous nodes). As demonstrated in Fig. 2(b), TSP-Mamba obtains a local shortest scan path for aggregation, thus leading to improved performance at a comparable and even reduced model complexity (shown in Fig. 1).", + "bbox": [ + 88, + 652, + 483, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 109, + 776, + 339, + 791 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Drawing inspiration from the shortest path principle, we propose a novel local structure-aware Mamba, TSP-Mamba, enabling content-adaptive propagation for efficient single-image super-resolution. We design an efficient TSP-LSTM to facilitate effective local aggregation by leveraging the locally optimal scan path.", + "- Extensive experiments show that the proposed TSP" + ], + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mamba framework achieves state-of-the-art performance across multiple lightweight SISR benchmarks, requiring only an additional 1-2 GFLOPs for 720P outputs (Tab. 1).", + "bbox": [ + 524, + 344, + 906, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Beyond lightweight SISR, we also investigate the scalability and generalization capabilities of TSP-Mamba for various vision tasks, including classic SISR, and strong generalization for depth map super-resolution.", + "bbox": [ + 513, + 388, + 906, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 462, + 661, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CNN-based Methods. Due to their low complexity and effective feature extraction capabilities, CNNs have been widely employed for SISR tasks. SRCNN [9] was a pioneering work that introduced deep convolutional neural network (CNN) architectures specifically designed for single image super-resolution (SISR). SRCNN features a simple three-layer structure consisting of patch extraction, nonlinear mapping, and reconstruction, and it demonstrated competitive performance against traditional non-deep learning methods, sparking the development of numerous lightweight CNN-based approaches in the SISR domain. Inspired by dictionary-learning models [15, 17, 30], LAPAR [20] learns linear coefficients associated with pre-defined basic upsampling kernels to create an optimal pixel-specific kernel, leading to superior super-resolution outcomes. LatticeNet [26] introduces a parameter-efficient convolutional lattice block to extract hierarchical contextual features.", + "bbox": [ + 511, + 489, + 908, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer-based Methods. On the other hand, Transformer-based techniques [4-6, 8, 21, 35, 37-39] have achieved impressive results in SISR but are still hindered by their high computational complexity. The cost of modeling non-local self-similarity scales quadratically with the size of the image. To address these efficiency concerns, many window-based Transformer frameworks have been developed, inspired by the success of the Swin Transformer [24]. For instance, SwinIR [23] introduces a residual window-based transformer block (RSTB) for general image restora", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28135", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ea5df24049ab98011ea6a26ff048f3347d51a9f973e5e056c5eaa04bf77deab7.jpg", + "image_caption": [ + "Figure 3. Illustration of our proposed TSP-Mamba, including a shallow feature extraction module ('Head'), a deep feature extraction module and a high-resolution reconstruction module ('Tail'). The key point is that the TSP-LSTM, employed for local path-adaptive aggregation, facilitates subsequent global scanning of 2D-SSM." + ], + "image_footnote": [], + "bbox": [ + 98, + 93, + 897, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion. DLGSA-1 [22] enhances token aggregation by using a global sparse attention mechanism. Omni-SR [33] presents spatial-channel attention for lightweight SISR, achieving better performance and efficiency trade-off.", + "bbox": [ + 89, + 377, + 482, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mamba-based Methods. Recently, Mamba-based frameworks [13] have demonstrated impressive capabilities in modeling long-sequence dependencies while maintaining linear computational complexity. At the heart of Mamba is a structured state space model that expands the effective receptive field, allowing the framework to capture long-range dependencies more efficiently than traditional methods like CNNs and Transformers. This capacity to model global dependencies with linear complexity has positioned Mamba as a compelling alternative, particularly in resource-constrained environments. Building on Mamba's success, numerous follow-up studies [2, 11, 12, 14, 32] have explored the potential of designing different Mamba-based frameworks for various vision tasks. For example, MambaIR [14] offers a simple yet effective Mamba-based baseline for image restoration, while RetinexMamba [2] leverages Mamba's long-range dependency modeling for low-light image enhancement. Although these approaches integrate Mamba's global sequential reasoning (1D) into vision problems, the fixed scan pipeline often overlooks the locality and content-adaptive nature of 2D images. In this paper, we aim to develop an efficient and content-adaptive local propagation and aggregation method to improve the performance of Mamba-based frameworks, particularly for resource-constrained SISR.", + "bbox": [ + 89, + 441, + 483, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 821, + 189, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: Mamba", + "text_level": 1, + "bbox": [ + 89, + 847, + 302, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin by presenting a concise overview of the Mamba pipeline, followed by a discussion of the limitations asso", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ciated with the rigid 2D-to-1D scan scheme for low-level vision tasks. Selective structured state space models (SSMs) convert a 1D sequence $x(t)\\in \\mathbb{R}$ to the response $y(t)\\in \\mathbb{R}$ through a latent state $h(t)\\in \\mathbb{R}^N$ . This process is expressed:", + "bbox": [ + 511, + 377, + 906, + 438 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t),\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 446, + 792, + 463 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny (t) = \\mathbf {C} h (t) + \\mathbf {D} x (t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 463, + 906, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ , $\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}$ , $\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}$ and $D \\in \\mathbb{R}$ are time-varying matrices and $N$ is the state number. In practical applications, the continuous evolution system can be discretized using a zero-order hold assumption. Accordingly, Mamba converts the continuous matrices $\\mathbf{A}$ , $\\mathbf{B}$ to their discrete forms $\\bar{\\mathbf{A}}$ , $\\bar{\\mathbf{B}}$ with a timescale parameter $\\Delta \\in \\mathbb{R}$ :", + "bbox": [ + 511, + 493, + 906, + 580 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {A}} = \\exp (\\Delta \\mathbf {A}),\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 585, + 718, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {B}} = (\\Delta \\mathbf {A}) ^ {- 1} (e ^ {\\mathbf {A}} - \\mathbf {I}) \\cdot \\Delta \\mathbf {B}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 602, + 905, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then obtain a discretized formulation:", + "bbox": [ + 516, + 635, + 792, + 648 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k} = \\bar {\\mathbf {A}} h _ {k - 1} + \\bar {\\mathbf {B}} x _ {k}, \\quad y _ {k} = \\mathbf {C} h _ {k} + \\mathbf {D} x _ {k}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 652, + 906, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we can see, since the overall calculation is a linear combination of the accumulated sum, it can be efficiently implemented as a global 1D convolution:", + "bbox": [ + 511, + 680, + 906, + 720 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {K}} = \\left(\\mathbf {C} \\bar {\\mathbf {B}}, \\mathbf {C} \\bar {\\mathbf {A}} \\bar {\\mathbf {B}}, \\dots , \\mathbf {C} \\bar {\\mathbf {A}} ^ {S - 1} \\bar {\\mathbf {B}}\\right), \\mathbf {y} = \\mathbf {x} * \\bar {\\mathbf {K}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 726, + 906, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\otimes$ is a convolution operation and the $\\bar{\\mathbf{K}}\\in \\mathbb{R}^S$ refers to a convolution kernel. $S$ is the sequence length.", + "bbox": [ + 511, + 755, + 906, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussion. In NLP tasks, it is natural to employ state space models (SSMs) to process 1D sequences for global information reasoning at a linear computational cost. However, in 2D image scenarios, the neighborhoods often face challenges such as disconnection or long-range propagation to fit within the SSM framework. Although some recent studies have investigated alternative 1D scan paths to supplement", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28136", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "information and improve restoration capability, capturing multiple local neighbors in a single 1D scan remains difficult, which can result in inefficient local aggregation for Mamba-based frameworks.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Preliminaries: TSP", + "text_level": 1, + "bbox": [ + 89, + 161, + 276, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $n$ cities and their pairwise distances, the goal is to determine the shortest route that visits each city exactly once. Mathematically, TSP is formulated to find a minimum-cost Hamiltonian cycle in a weighted complete graph, where nodes represent cities and edges denote travel costs (e.g., distance or time). Inspired by TSP's optimal traversal property, we propose a TSP-style local scan scheme to complement the global modeling of vision Mambas in this work.", + "bbox": [ + 89, + 181, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Framework Overview", + "text_level": 1, + "bbox": [ + 89, + 313, + 295, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the framework of our proposed TSP-Mamba. Following MambaIR, a recent SOTA baseline [14], TSP-Mamba comprises three main modules: (1) a shallow feature extraction module that transforms the RGB input into a feature space; (2) a deep feature extraction module consisting of multiple TSP-RSSGs (TSP-based Residual State-Space Groups). Each T-RSSG consists of several T-RSSBs (TSP-based Residual State-Space Block). In TSP-RSSB, the key component is the T-VSSM (TSP-based Vision State-Space Module), where we first use a TSP-LSTM layer to perform local path-adaptive aggregation and then pass the local-aggregation information into the 2D-SSM (2D Selective Scan Module in MambaIR[14]); and (3) a high-resolution reconstruction module that produces the final results.", + "bbox": [ + 89, + 335, + 483, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following section, we mainly focus on the explanation of our proposed TSP-LSTM, including the scanning patch construction for optimal local propagation (Sec.3.4.1), simplified TSP (Sec.3.4.2) and efficient implementation (Sec.3.4.3).", + "bbox": [ + 89, + 547, + 483, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. TSP-LSTM", + "text_level": 1, + "bbox": [ + 89, + 631, + 218, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1. Scanning Path Construction", + "text_level": 1, + "bbox": [ + 89, + 654, + 333, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an image feature patch $\\mathbf{P} \\in \\mathbb{R}^{s \\times s \\times C}$ , where $s$ is the patch size and $C$ is the number of channels, our objective is to establish a locally adaptive aggregation path that minimizes the information propagation costs.", + "bbox": [ + 89, + 672, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition. We formulate the scan path as a Traveling Salesman Problem (TSP) problem: let $V = \\{1,2,\\dots n\\}$ be the set of pixels (cities) in image patch $P$ and $E = \\{e_{i,j}(i,j):i,j\\in V\\}$ and $D_{e_{i,j}}$ be the cost measure associated with edge $e_{i,j}\\in E$ . Our goal is to find a minimal scan path such that each city (pixel) is visited exactly once.", + "bbox": [ + 89, + 751, + 483, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, we use the negative Cosine similarity as the cost measure of an edge $e_{i,j}$ in the set $E$ :", + "bbox": [ + 89, + 843, + 482, + 873 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {e _ {i j}} = 1. 0 - \\operatorname {C o s i n e} \\left(F _ {i}, F _ {j}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 883, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ec369dbb7a2b91e49e1749504e236b9af4a6ed0a9ed3283c7e27a22856a6effe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Size (s)TSPOurs
24!2
39!8
416!52
525!824
", + "bbox": [ + 529, + 205, + 666, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/adf2335ff54d627c8ad948b1e4b7b1f96624f5186315aef740dfcdfc39cea06a.jpg", + "image_caption": [ + "Figure 4. The numbers of scan path candidates of TSP and our simplified TSP scheme under different patch size $s$ .", + "Figure 5. Speedup ratio curve under different patch sizes. Our effective measure hashing strategy attains notable speedup $(\\log_{10})$ for efficient TSP searching." + ], + "image_footnote": [], + "bbox": [ + 678, + 90, + 903, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{i}, F_{j}$ are the corresponding feature vector of the pixel $i, j$ . More discussion is shown in Tab. 3 (A).", + "bbox": [ + 511, + 327, + 903, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.2. Simplified TSP", + "text_level": 1, + "bbox": [ + 511, + 363, + 665, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Traveling Salesman Problem (TSP) is a classic NP-hard problem. With a large $n$ , searching for an optimal scan path becomes time-consuming ( $O(n!)$ ). To simplify the problem, we adopt the following strategies in this work:", + "bbox": [ + 511, + 382, + 906, + 443 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We limit the overall node count $n$ to the pixels within a local image patch, thereby reducing overall complexity;", + "(2) We designate the first city '1' (the top-left pixel of the given patch) as the starting node;", + "(3) At each intermediate step, we limit next-step candidates to the four neighboring 2D locations1 to reduce the search space further." + ], + "bbox": [ + 511, + 443, + 903, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on these three simplifications mentioned, it largely reduced the number of total candidate scan paths (in Fig. 4). Thus, we can initially employ the backtracking algorithm [10] to identify all possible scan paths:", + "bbox": [ + 511, + 549, + 905, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} T ^ {\\{1, 2, \\dots , l, \\dots , L \\}} = \\text {B a c k T r a c k i n g} (1, n), \\\\ \\text {w h e r e} T ^ {l} = \\left\\{i _ {1} ^ {l}, i _ {2} ^ {l}, \\dots , i _ {n} ^ {l} \\right\\}, i _ {1} ^ {l} = 1, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 616, + 841, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l$ -th path $T^l$ is represented by the ordered pixel indices $(\\{i_1^l, i_2^l, \\dots, i_n^l\\})$ and $L$ is the count of all scan candidates. We calculate the cost of the $l$ -th scan path $T^l$ :", + "bbox": [ + 511, + 665, + 906, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD ^ {l} = \\sum_ {k = 1} ^ {n - 1} D _ {e _ {i _ {k} ^ {l}, i _ {k + 1} ^ {l}}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 719, + 906, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we choose the optimal scan path $T^{l^*}$ with the lowest measure cost:", + "bbox": [ + 511, + 756, + 903, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl ^ {*} = \\underset {l} {\\operatorname {a r g m i n}} D ^ {l} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 786, + 906, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.3. Effective Implementation", + "text_level": 1, + "bbox": [ + 511, + 815, + 736, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As indicated in Eq. 7, we have to calculate the $n - 1$ pairwise distances for each scan path and sum them up. We have two observations for effective path cost computation:", + "bbox": [ + 511, + 834, + 906, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "left $(-1,0)$ , right $(+1,0)$ , top $(0,-1)$ , and bottom $(0,+1)$", + "bbox": [ + 529, + 886, + 815, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28137", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Symmetry of Distance Measurement. As indicated in Eq. 5, for edges $e_{i,j}$ , $e_{j,i}$ , we have $D_{e_{i,j}} = D_{e_{j,i}}$ .", + "2. Highly overlapping edges for different scan paths. Different scene paths may share lots of the same edges." + ], + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on this fact, we propose to design a hash-based technique to store the unique edges in all scan paths into an edge list $\\{e_{k_0}, e_{k_1}, \\dots, e_{k_m}\\}$ ( $m$ is the length of the unique edge list) to avoid repeated distance calculation. Accordingly, we can effectively compute all unique edge distances using Eq. 5 at one time and obtain the associated edge distance list $\\{D_{e_{k_0}}, D_{e_{k_1}}, \\dots, D_{e_{k_m}}\\}$ . Finally, we can obtain the path propagation cost in Eq. 7 by accumulating all its edges.", + "bbox": [ + 89, + 152, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The speedup ratio between the direct per-path cost computation and our efficiency implementation is shown in Fig. 5, we can see our effective hash-based calculation attains a significant speedup effect. Additionally, it becomes more effective with an increased patch size.", + "bbox": [ + 89, + 273, + 483, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Summary. In this part, we propose a simplified TSP formulation to effectively search the shortest path for a given 2D image patch, in which all neighborhoods are fused in a relatively short-term propagation. Therefore, it is capable of effectively integrating local information to facilitate the later global SSM-style propagation.", + "bbox": [ + 89, + 368, + 483, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.4. TSP-LSTM", + "text_level": 1, + "bbox": [ + 89, + 467, + 218, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bidirectional Propagation. After the scan path optimization, we perform a bidirectional LSTM for mutual information propagation. Especially, given a 2D image patch $\\mathbf{P}$ , we use the above simplified TSP scheme to get the optimal scan chain $T^{l^{*}} = \\{i_{1} \\rightleftharpoons i_{2}, \\dots, \\rightleftharpoons i_{n}\\}^{2}$ , where “ $\\rightleftharpoons$ ” means we employ the bidirectional propagation scheme (as shown in Fig. 2) (b) for effective and balanced aggregation.", + "bbox": [ + 89, + 486, + 483, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For implementation, we flip the reordered sequence $\\{i_1,i_2,\\dots ,i_n\\}$ to obtain the inverse sequence and concatenate them for bidirectional aggregation.", + "bbox": [ + 89, + 593, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distance Embedding. Within an image patch, discontinuities may need to be addressed in the chain-style propagation. Specifically, the aggregation weight of the previously accumulated information should positively correlate with the distance between the previous and current nodes (pixels). To account for this, inspired by fast global image smoothing (FGS [29]), we incorporate distance embedding features into the sequences to explicitly model this effect. We calculate the pairwise distances as shown in Eq. 7 and embed them (as a one-channel feature) into the propagation sequence feature.", + "bbox": [ + 89, + 657, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, the propagated feature $\\mathbf{x}$ is passed through the 2DSSM module (in Eq. 4) for global Mamba-style aggregation.", + "bbox": [ + 89, + 809, + 483, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 89, + 643, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 513, + 114, + 609, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our TSP-Mamba model is implemented in PyTorch and integrates several common data augmentation techniques, such as random cropping, vertical/horizontal flipping, and rotation. For training, we use the Adam optimizer [18] with cosine annealing [25], starting with an initial learning rate of $4e - 4$ . The batch size is set to 32, with input images sized at $64 \\times 64$ . Training is performed over 600K iterations using four NVIDIA RTX 3090 GPUs.", + "bbox": [ + 511, + 137, + 906, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scalable Model Configuration. In general, we train multiple TSP-Mamba models with different configurations. TSP-Mamba is based on the MambaIR baseline, utilizing 4 T-RSSGs, each containing 6 T-RSSBs, with a channel size of 60, denoted as “[6,6,6,6]@C60”. Additionally, we investigate the scalability of TSP-Mamba to better understand the trade-off between performance and model complexity. Specially, we use “[2,4,2,6]@C40” $^{4}$ and “[4,6,4,6]@C48” $^{5}$ to train another two smaller SISR models under $\\times 2$ SISR setting. Training these TSP-Mamba models usually takes around 3-6 days.", + "bbox": [ + 511, + 258, + 908, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Datasets. In line with previous studies [7, 20, 23], we use the DIV2K [1] image dataset for training. We then perform extensive evaluations on several widely-used SISR benchmarks, including Set5 [3], Set14 [36], BSD100 [27], Urban100 [16], and Manga109 [28]. Our quantitative analysis is based on PSNR and SSIM metrics, and following standard practice, we report results specifically for the Y channel in the YCbCr color space.", + "bbox": [ + 511, + 426, + 908, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 511, + 558, + 805, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform extensive comparisons with a wide range of lightweight SISR models: LAPAR-A (NeurIPS20) [20], LatticeNet (ECCV20) [26], RLFN (CVPRW22) [19], SwinIR-light [23], DLGSA-1 (ICCV23) [22], Onmi-SR (CVPR23) [33], SRFormer (ICCV23) [39], ATDS (CVPR24) [37], MambaIR (ECCV24) [14], HiTSR [38] (ECCV24).", + "bbox": [ + 511, + 580, + 908, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Comparison. Tab. 1 shows the quantitative evaluation results, highlighting that our proposed TSP-Mamba consistently outperforms existing methods across all benchmarks. In particular, our model achieves significant improvements over MambaIR, with gains of up to $0.71\\mathrm{dB}$ , $0.58\\mathrm{dB}$ , and $0.69\\mathrm{dB}$ on the Manga109 dataset for the $\\times 2$ , $\\times 3$ , and $\\times 4$ scales, respectively, while maintaining similar parameter and FLOP requirements.", + "bbox": [ + 511, + 686, + 908, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, TSP-Mamba-Small beats all leading lightweight SISR methods with smaller model sizes and", + "bbox": [ + 511, + 808, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2For clarity, we drop the upper letter and denote $T^{l^*}$ as $T^{l^*} = \\{i_1,i_2,\\dots ,i_n\\}$", + "bbox": [ + 89, + 848, + 482, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3Each pixel can fuse all its neighboring pixels in the bidirectional short-range propagation scheme", + "bbox": [ + 89, + 875, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4TSP-Mamba-Tiny: 4 TSP-RSSGs contain 2,4,2,6 TSP-RSSBs and the channel size is 40", + "bbox": [ + 511, + 849, + 903, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "5TSP-Mamba-Small: 4 TSP-RSSGs contain 4,6,4,6 TSP-RSSBs and the channel size is 48", + "bbox": [ + 511, + 875, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28138", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5854e7f5efbfe332d90767046ec9cb6bb5025d7ed6de0e4dd6ac19e5a3e2d964.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with SOTA lightweight SISR methods on multiple benchmark datasets. The best and second-best results on the default training setting (DIV2K) are highlighted in red and blue, respectively. We follow most lightweight SISR methods and calculate all Flops with an output size of ${1280} \\times {720}$ ." + ], + "table_footnote": [], + "table_body": "
MethodScaleParameters (K)Flops (G)Set5 PSNR / SSIMSet14 PSNR / SSIMBSD100 PSNR / SSIMUrban100 PSNR / SSIMManga109 PSNR / SSIM
LAPAR-A54817138.01 / 0.960533.62 / 0.918332.19 / 0.899932.10 / 0.928338.67 / 9772
LatticeNet756169.538.15 / 0.961033.78 / 0.919332.25 / 0.900532.43 / 0.9302- / -
SwinIR-light910244.438.14 / 0.961133.86 / 0.920632.31 / 0.901232.76 / 0.934039.12 / 0.9783
DLGSA-light745170.038.20 / 0.961233.89 / 0.920332.30 / 0.901232.94 / 0.935539.29 / 0.9780
Omni-SR772194.538.22 / 0.961333.98 / 0.921032.36 / 0.902033.05 / 0.936339.28 / 0.9784
SRFormer ×285323638.23 / 0.961333.94 / 0.920932.36 / 0.901932.91 / 0.935339.28 / 0.9785
HiT-SR847226.538.26 / 0.961534.01 / 0.921432.37 / 0.902333.13 / 0.937239.47 / 0.9787
ATD-S75334938.29 / 0.961634.10 / 0.921732.39 / 0.902333.27 / 0.937539.52 / 0.9789
MambaIR136319838.16 / 0.961034.00 / 0.921232.34 / 0.901732.92 / 0.935639.31 / 0.9779
TSP-Mamba130519938.46 / 0.964834.42 / 0.924332.52 / 0.904233.57 / 0.940540.02 / 0.9808
MambaIR-Small80611138.09 / 0.960233.85 / 0.920132.31 / 0.901532.58 / 0.932139.17 / 0.9787
TSP-Mamba-Small76911238.22 / 0.964034.19 / 0.922732.43 / 0.903033.18 / 0.937439.49 / 0.9791
MambaIR-Tiny4435937.99 / 0.960233.72 / 0.910932.21 / 0.900232.22 / 0.929138.87 / 0.9772
TSP-Mamba-Tiny4185938.04 / 0.963133.90 / 0.920532.29 / 0.901032.55 / 0.931739.12 / 0.9784
LAPAR-A54411434.36 / 0.926730.34 / 0.842129.11 / 0.805428.15 / 0.852333.51 / 9441
LatticeNet76576.334.53 / 0.928130.39 / 0.842429.15 / 0.805928.33 / 0.8538- / -
SwinIR-light918110.834.62 / 0.928930.54 / 0.846329.20 / 0.808228.66 / 0.862433.98 / 0.9478
DLGSA-light75275.434.70 / 0.929530.58 / 0.846529.24 / 0.808928.83 / 0.865334.16 / 0.9483
Omni-SR78088.434.70 / 0.929430.57 / 0.846929.28 / 0.809428.84 / 0.865634.22 / 0.9487
SRFormer ×386110534.67 / 0.929630.57 / 0.846929.26 / 0.809928.81 / 0.865534.19 / 0.9489
HiT-SR85510234.75 / 0.930030.61 / 0.847529.29 / 0.810628.99 / 0.868734.53 / 0.9502
MambaIR13718934.72 / 0.929630.63 / 0.847529.29 / 0.809929.00 / 0.868934.39 / 0.9495
TSP-Mamba13138934.80 / 0.933230.93 / 0.851029.41 / 0.813929.43 / 0.874434.97 / 0.9522
MambaIR-Small8125034.65 / 0.931930.59 / 0.847129.26 / 0.810228.84 / 0.864434.02 / 0.9472
TSP-Mamba-Small7765034.65 / 0.932230.66 / 0.847829.30 / 0.811428.94 / 0.868034.41 / 0.9495
MambaIR-Tiny4482734.36 / 0.922330.26 / 0.842829.12 / 0.806328.24 / 0.853233.74 / 0.9452
TSP-Mamba-Tiny4232734.49 / 0.930930.46 / 0.845829.22 / 0.809428.70 / 0.862334.16 / 0.9483
LAPAR-A6599432.15 / 0.894428.61 / 0.781827.61 / 0.736626.14 / 0.787130.42 / 0.9074
LatticeNet77743.632.30 / 0.896228.68 / 0.783027.62 / 0.736726.25 / 0.7873- / -
SwinIR-light93063.632.44 / 0.897628.77 / 0.785827.69 / 0.740626.47 / 0.798030.92 / 0.9151
DLGSA-light76142.532.54 / 0.899328.84 / 0.787127.73 / 0.741526.66 / 0.803331.13 / 0.9161
Omni-SR79250.932.49 / 0.898828.78 / 0.785927.71 / 0.741526.64 / 0.801831.02 / 0.9151
SRFormer ×48736332.51 / 0.898828.82 / 0.787227.73 / 0.742226.67 / 0.803231.17 / 0.9165
HiT-SR8665832.55 / 0.899928.87 / 0.788027.75 / 0.743226.80 / 0.806931.26 / 0.9171
ATD-S7698732.63 / 0.899828.89 / 0.788627.79 / 0.744026.97 / 0.810731.48 / 0.9198
MambaIR13835132.51 / 0.899328.85 / 0.787627.75 / 0.742326.75 / 0.805131.26 / 0.9175
TSP-Mamba13245332.71 / 0.902429.05 / 0.792527.91 / 0.747627.29 / 0.817431.95 / 0.9235
MambaIR-Small8212932.28 / 0.900128.70 / 0.786327.72 / 0.741226.56 / 0.798431.12 / 0.9156
TSP-Mamba-Small7853032.48 / 0.902428.90 / 0.789327.79 / 0.744226.93 / 0.808731.42 / 0.9184
MambaIR-Tiny4561532.22 / 0.899128.63 / 0.784727.68 / 0.739826.43 / 0.794030.86 / 0.9124
TSP-Mamba-Tiny4311632.29 / 0.899728.79 / 0.786027.72 / 0.740526.61 / 0.799231.04 / 0.9140
MambaIR (×16)1772523.24 / 0.641922.29 / 0.545422.94 / 0.521820.46 / 0.514821.15 / 0.6628
TSP-Mamba (×16)1714523.33 / 0.651522.38 / 0.545222.81 / 0.516920.53 / 0.518421.20 / 0.6700
", + "bbox": [ + 112, + 142, + 883, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "less computational complexity. For instance, it surpasses MambaIR by 0.26dB on Urban $100 \\times 2$ , while only requiring $59\\%$ parameters and $57\\%$ Flops.", + "bbox": [ + 89, + 732, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference efficiency. Tab. 2 presents a detailed comparison of inference costs across different models, including CNN-based, Transformer-based, and Mamba-based frameworks. Overall, the CNN-based LatticeNet achieves the highest inference efficiency. Compared to other transformer-based and Mamba-based models, our proposed TSP-Mamba achieves competitive inference latency.", + "bbox": [ + 89, + 777, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. Fig. 6 showcases several visual", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "examples. It is clear that TSP-Mamba effectively reconstructs accurate image textures with fewer super-resolved artifacts. In contrast, other approaches struggle to restore clear image patterns.", + "bbox": [ + 511, + 732, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 813, + 678, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate our design, we conduct extensive experiments, reporting quantitative results on the Urban100 and Manga109 benchmarks at the $\\times 4$ setting. We primarily adopt the TSP-Mamba-Small architecture for fast analysis.", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28139", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6aaa5f94cacb147a41f5bc9cd65db9c36c5601911e861d706968ae19c8de3bc.jpg", + "image_caption": [ + "Figure 6. Qualitative results of different lightweight SISR methods under $\\times 4$ setting. Our predictions present more detailed structures." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 901, + 308 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/852424bb56f051db9dffb4869be4ed045e0278b507cefdcf252f5cc6e1e183f8.jpg", + "table_caption": [ + "Table 2. Inference cost comparison of different lightweight SISR models under $\\times 2$ setting. Inference time is tested with an output resolution of ${256} \\times {256}$ using an NVIDIA 3090 GPU. The top-3 results are highlighted with red,blue,underline." + ], + "table_footnote": [], + "table_body": "
MetricsLatticeNetDLGSA-lightSwinIR-lightSRFormerMambaIR / Small / TinyTSP-Mamba / Small / Tiny
Time (ms) ↓973416060 / 50 / 3872 / 58 / 39
Memory (MB) ↓146616245441572686 / 536 / 528732 / 614 / 582
", + "bbox": [ + 129, + 385, + 870, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different Measure Cost in TSP-LSTM. Apart from Eq. 5, we try the widely-used L1/L2 distance metrics for scan path searching. Tab. 3 (A) concludes that our used negative Cosine distance performs better than the L1/L2 distances.", + "bbox": [ + 89, + 452, + 485, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TSP-LSTM vs. Uniform-LSTM. TSP-LSTM aims to perform an adaptive scan path by constructing a minimal propagation path. To assess the impact of our adaptive scan path, we also train two uniform scan models (vertical, and horizontal). We compare the proposed TSP-LSTM with the other models (UV-LSTM, HV-LSTM). As shown in Tab. 3 (B), we can observe that TSP-LSTM significantly outperforms the uniform scan models, highlighting the importance of adaptive local propagation.", + "bbox": [ + 89, + 532, + 483, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different Patch Sizes. As mentioned in Sec. 3.4, we set the patch size to $4 \\times 4$ considering the local receptive field and computational costs. When the patch size is increased, it becomes prolonged to construct the corresponding TSP path. We investigate the influence of different patch sizes. Tab. 3 (C) suggests the reconstructed performance is positively related to the patch size but with larger latency.", + "bbox": [ + 89, + 690, + 486, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finer Design in TSP-LSTM. Lastly, we evaluate the impact of bidirectional propagation of TSP-LSTM. We train another model with only a forward/backward scan pipeline to do this. In Tab. 3 (D), it is clear that our bidirectional propagation scheme is more effective than its counterparts with only a single-direction scan strategy.", + "bbox": [ + 89, + 816, + 485, + 907 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different Aggregation Layers. In this paper, we present a TSP-LSTM block to get locally fused information. To verify the effectiveness of TSP-LSTM, we replace it with other designs for local compensation, such as convolution, RNN, and Mamba. The results are shown in Tab. 3 (E) below, which suggests that our TSP attains the best performance and second-best efficiency compared with other designs.", + "bbox": [ + 511, + 452, + 908, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Extensions", + "text_level": 1, + "bbox": [ + 511, + 568, + 635, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classic SISR. Though TSP-Mamba is trained for lightweight Mamba-based architecture, i.e., MambaIR, we additionally explore the potential improvement for the classic SISR setting. To do this, we follow MambaIR-L(large) to enlarge the channel size to 180 and use 6 TSP-RSSG blocks (each one contains 6 basic TSP-SSD layers) to obtain the TSP-Mamba-L model and train this classic SISR model using the same DF2K training dataset. The quantitative results are illustrated in Tab. 4. We can see that TSP-Mamba-L consistently outperforms other classic SISR methods significantly. For instance, it achieves 0.32dB, 0.37dB, and 0.39dB improvements over MambaIR.", + "bbox": [ + 511, + 590, + 908, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Depth Map SR. We also compare the TSP-Mamba model with our baseline model (MambaIR) on the depth SR task for generalization analysis. To do this, without training on any depth images, we directly assess the two models on the NYU-V2 [31] depth benchmark $^6$ under $\\times 4$ and $\\times 16$ settings.", + "bbox": [ + 511, + 790, + 908, + 867 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{6}$ NYU V2 consists of 449 testing depth images at a size of $480 \\times 640$ . We use bicubic interpolation to generate the low-resolution depth images.", + "bbox": [ + 511, + 875, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28140", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/180c20f5dc7c5c38f594c31d7d9f314079fc54ba6378c98b0d57aa18522c55cf.jpg", + "table_caption": [ + "Table 3. Ablation Studies. A: The effects of different distance measures, with our default choice being the negative cosine distance as indicated in Eq. 5. B: Quantitative evaluation of uniform scanning strategies (vertical and horizontal), abbreviated as 'UV' and 'HV', compared to our TSP-based scan. C: The influence of different local scan patch sizes (P2-P5 indicates local patch sizes ranging from 2 to 5). D: Impacts of distance embeddings and bidirectional propagation. M1 drops both of the two designs. M2 only adopts distance embedding. M3 only uses the bidirectional propagation. M4 employs both of the two designs. E: The effects of various local aggregation modules (M5-M8 employ Conv1D, RNN, Mamba, and LSTM, respectively)." + ], + "table_footnote": [], + "table_body": "
AnalysisA.B.C.D.E.
L1L2Neg. Cos.UVUHTSPP2P3P4P5M1M2M3M4M5M6M7M8
Urban10026.5526.5426.6126.3126.3826.6126.3726.4326.6126.6126.1626.3826.2526.6126.4926.4426.4726.61
0.74030.79820.79920.79070.79360.79920.79330.79430.79920.79940.78810.79390.79070.79920.79580.79500.79560.7992
Manga10931.0631.0731.0430.7830.8631.0430.7530.9131.0431.0230.5030.8030.6531.0430.9630.9530.9731.04
0.91400.91440.91400.91040.91260.91400.91130.91310.91400.91380.90650.91030.90840.91400.91360.91320.91340.9140
Param.(K)431431431431431431431431431431430431430431440436428431
FLOPs (G)16.216.216.215.415.416.215.215.216.220.314.915.116.116.215.515.716.916.2
", + "bbox": [ + 91, + 176, + 911, + 289 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/53c257f5ac2f0987a54507370015083859ac9cc4547d0bb3522c45708de0bf2e.jpg", + "table_caption": [ + "Table 4. Quantitative comparison with recent SOTA Classic SR models. All the models are trained on the DF2K dataset for fair comparison. We report the averaged PSNR/SSIM values over the five benchmarks (Set5, Set14, BSD100, Urban100, Manga109)." + ], + "table_footnote": [], + "table_body": "
SettingsMetricsSwinIRSRFormer-LHATMambaIR-LTSP-Mamba-L
×2PSNR35.8335.9436.1736.0536.37
SSIM0.94280.94350.94460.94380.9457
Param.11.8M10.4M20.6M20.4M20.6M
×3PSNR32.0532.1532.2932.1932.56
SSIM0.88720.88850.88810.89000.8918
Param.11.8M10.6M20.8M20.6M20.8M
×4PSNR29.8929.9730.1430.0430.43
SSIM0.83990.84160.84410.84140.8470
Param.11.8M10.5M20.8M20.6M20.8M
", + "bbox": [ + 91, + 375, + 488, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use the PSNR, SSIM, and RMSE (the root-mean-square error) metrics for quantitative evaluation. The quantitative results are shown in the left Fig. 7. We can see that TSP-Mamba consistently outperforms its baseline model on all the objective metrics. For instance, TSP-Mamba exhibits superior performance compared to MambaIR with a PSNR improvement of 0.96dB (0.92dB) for $\\times 4$ ( $\\times 16$ ) configurations. The visual examples in Fig. 7 illustrate that TSP-Mamba generates clearer structures, leading to higher accuracy when compared to the baseline model.", + "bbox": [ + 88, + 534, + 485, + 686 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fd2cc6d4a3b4c2f70f34edb1d6a5c2f917748c4482d730603b9b2b37a7e85cd2.jpg", + "table_caption": [ + "Table 5. The quantitative results on NYU-V2 [31] $(\\times 4,\\times 16)$" + ], + "table_footnote": [], + "table_body": "
SettingsMetricsMambaIRTSP-Mamba
×4 (×16)PSNR47.25(37.03)48.21(37.95)
SSIM0.994(0.967)0.995(0.972)
RMSE2.339(7.895)2.145(7.142)
", + "bbox": [ + 116, + 723, + 457, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although our proposed TSP-Mamba demonstrates promising and significant improvements across several single-image super-resolution tasks, it remains challenging to reliably obtain accurate distance measurements for degraded images, particularly in the presence of high-frequency degradations. Noise components can adversely affect the", + "bbox": [ + 89, + 809, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af97791f8b819af208afca2962c69b89b403f059287142cc566fc762f256ae35.jpg", + "image_caption": [ + "PSNR/SSIM/RMSE" + ], + "image_footnote": [], + "bbox": [ + 519, + 306, + 650, + 385 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e4c214db5bd53d74e17d402c067a02865771591b0e9ddb75a8a62cf4131dfe5.jpg", + "image_caption": [ + "37.45/0.9698/0.0330" + ], + "image_footnote": [], + "bbox": [ + 651, + 308, + 784, + 385 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/229faf48ef40d423b13934a6c600e935a67fc9937affb2b77117907d84867f6a.jpg", + "image_caption": [ + "38.79/0.9750/0.0282" + ], + "image_footnote": [], + "bbox": [ + 785, + 308, + 918, + 385 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5b7734f1e4a9fff80a96060cda4db81977001cb9f9cdc95e26365d21138db7b.jpg", + "image_caption": [ + "PSNR/SSIM/RMSE" + ], + "image_footnote": [], + "bbox": [ + 519, + 395, + 650, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85c542c92fe9da51ae09ec98329d9991ced9dae65d9775b22bb5f76cadcc7312.jpg", + "image_caption": [ + "34.80/0.9672/0.0673" + ], + "image_footnote": [], + "bbox": [ + 651, + 395, + 784, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a9a0ccec9c600fb1fb4290670466b06da9717a147006ff4d8bf9fac59abb18e.jpg", + "image_caption": [ + "36.45/0.9742/0.0555", + "Figure 7. Qualitative comparison between our baseline and the proposed TSP-Mamba for unseen depth-SR maps. We also show normalized error maps in the left corner for better visual comparison" + ], + "image_footnote": [], + "bbox": [ + 785, + 395, + 916, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "calculation of pair-wise distances between degraded pixels and the optimal scan path searching. In future work, we plan to address the issue of robust distance calculation under various image degradations, such as pyramid representation to build robust 1D TSP scan paths hierarchically.", + "bbox": [ + 511, + 542, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce a novel locally adaptive propagation method for lightweight SISR. To this end, we develop an efficient scan path search strategy to enhance local information within the globally Mamba-based architecture. Extensive experiments validate the effectiveness of our approach, with TSP-Mamba achieving state-of-the-art performance across multiple SISR benchmarks. Furthermore, TSP-Mamba demonstrates superior generalization on unseen data (such as depth-map SR), and scalability across both ultralightweight and larger model configurations.", + "bbox": [ + 511, + 659, + 908, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is partially supported by Shenzhen Science and Technology Program KQTD20210811090149095 and also the Pearl River Talent Recruitment Program 2019QN01X226.", + "bbox": [ + 511, + 830, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28141", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5", + "[2] Jiesong Bai, Yuhao Yin, and Qiyuan He. Retinexmamba: Retinex-based mamba for low-light image enhancement. arXiv preprint arXiv:2405.03349, 2024. 3", + "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5", + "[4] Ke Chen, Liangyan Li, Huan Liu, Yunzhe Li, Congling Tang, and Jun Chen. Swinfsr: Stereo image super-resolution using swinir and frequency domain knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764-1774, 2023. 2", + "[5] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023.", + "[6] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12312-12321, 2023. 2", + "[7] Haram Choi, Jeongmin Lee, and Jihoon Yang. N-gram in swim transformers for efficient lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2071–2081, 2023. 5", + "[8] Tao Dai, Jianping Wang, Hang Guo, Jinmin Li, Jinbao Wang, and Zexuan Zhu. Freqformer: frequency-aware transformer for lightweight image super-resolution. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 731-739, 2024. 2", + "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2", + "[10] Eugene C Freuder and Alan K Mackworth. Constraint satisfaction: An emerging paradigm. In Foundations of artificial intelligence, pages 13-27. Elsevier, 2006. 4", + "[11] Guanyiman Fu, Fengchao Xiong, Jianfeng Lu, and Jun Zhou. Ssumamba: Spatial-spectral selective state space model for hyperspectral image denoising. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3", + "[12] Hu Gao, Bowen Ma, Ying Zhang, Jingfan Yang, Jing Yang, and Depeng Dang. Learning enriched features via selective state spaces model for efficient image deblurring. In ACM Multimedia 2024. 3", + "[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3", + "[14] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. 1, 3, 4, 5", + "[15] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 2", + "[16] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5", + "[17] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3224-3232, 2018. 2", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[19] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 766-776, 2022. 5", + "[20] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. Lapar: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. Advances in Neural Information Processing Systems, 33:20343-20355, 2020. 2, 5", + "[21] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2", + "[22] Xiang Li, Jiangxin Dong, Jinhui Tang, and Jinshan Pan. Dlgstanet: Lightweight dynamic local and global self-attention networks for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12792–12801, 2023. 3, 5", + "[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 5", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2", + "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5", + "[26] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image superresolution with lattice block. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 272-289. Springer, 2020. 2, 5" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28142", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 1, 5", + "[28] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multi-media Tools and Applications, 76:21811-21838, 2017. 5", + "[29] Dongbo Min, Sunghwan Choi, Jiangbo Lu, Bumsub Ham, Kwanghoon Sohn, and Minh N Do. Fast global image smoothing based on weighted least squares. IEEE Transactions on Image Processing, 23(12):5638-5653, 2014. 5", + "[30] Yaniv Romano, John Isidoro, and Peyman Milanfar. Raisr: Rapid and accurate image super resolution. IEEE Transactions on Computational Imaging, 3(1):110-125, 2016. 2", + "[31] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 7, 8", + "[32] Yujin Tang, Peijie Dong, Zhenheng Tang, Xiaowen Chu, and Junwei Liang. Vmrnn: Integrating vision mamba and lstm for efficient and accurate spatiotemporal forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5663-5673, 2024. 3", + "[33] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Jinfan Liu. Omni aggregation networks for lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22378-22387, 2023. 3, 5", + "[34] Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. arXiv preprint arXiv:2407.21773, 2024. 1", + "[35] Jinsu Yoo, Taehoon Kim, Sihaeng Lee, Seung Hwan Kim, Honglak Lee, and Tae Hyun Kim. Enriched cnn-transformer feature aggregation networks for super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4956-4965, 2023. 2", + "[36] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5", + "[37] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2024. 2, 5", + "[38] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. In European Conference on Computer Vision, pages 483-500. Springer, 2024. 5" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[39] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12780–12791, 2023. 2, 5", + "bbox": [ + 516, + 90, + 906, + 161 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28143", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/989548d5-80da-4477-a3a5-42129f121eff_model.json b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/989548d5-80da-4477-a3a5-42129f121eff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..34dfd09a1fc50d722237b17c8313e2a14b5fa508 --- /dev/null +++ b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/989548d5-80da-4477-a3a5-42129f121eff_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.13, + 0.819, + 0.177 + ], + "angle": 0, + "content": "TSP-Mamba: The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.203, + 0.669, + 0.221 + ], + "angle": 0, + "content": "Kun Zhou1 Xinyu Lin2 Jiangbo Lu1*" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.221, + 0.697, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)SmartMore Corporation \\(^{2}\\)SSE, CUHK-Shenzhen" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.242, + 0.802, + 0.256 + ], + "angle": 0, + "content": "zhoukun303808@gmail.com, linxinyu0327@gmail.com, jiangbo.lu@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.291, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.322, + 0.486, + 0.671 + ], + "angle": 0, + "content": "Recently, Mamba-based frameworks have achieved substantial advancements across diverse computer vision and NLP tasks, particularly in their capacity for reasoning over long-range information with linear complexity. However, the fixed 2D-to-1D scanning pattern overlooks the local structures of an image, limiting its effectiveness in aggregating 2D spatial information. While stacking additional Mamba layers can partially address this issue, it increases parameter intensity and constrains real-time application. In this work, we reconsider the local optimal scanning path in Mamba, enhancing the rigid and uniform 1D scan through the local shortest path theory, thus creating a structure-aware Mamba suited for lightweight single-image super-resolution. Specifically, we draw inspiration from the Traveling Salesman Problem (TSP) to establish a local optimal scanning path for improved structural 2D information utilization. Here, local patch aggregation occurs in a content-adaptive manner with minimal propagation cost. TSP-Mamba demonstrates substantial improvements over existing Mamba-based and Transformer-based architectures. For example, TSP-Mamba surpasses MambaIR by up to 0.7dB in lightweight SISR, with comparable parameters and very slightly extra computational demands (1-2 GFlops for 720P images)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.688, + 0.222, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.486, + 0.88 + ], + "angle": 0, + "content": "Deep learning techniques have advanced single-image superresolution (SISR) tasks considerably, enhancing image reconstruction quality. Convolutional Neural Networks (CNNs) have been central to this progress, excelling at capturing local, content-dependent image patterns. However, their local focus limits their capacity to model long-range dependencies. Transformers, by contrast, are highly effective in capturing global information through their self-attention mechanism, enabling them to model non-local dependencies. Yet, their quadratic computational complexity poses significant challenges in resource-constrained environments." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.875, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.483, + 0.597, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.483, + 0.685, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.482, + 0.776, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.482, + 0.904, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.561, + 0.908, + 0.66 + ], + "angle": 0, + "content": "Figure 1. (1) Comparison between our proposed TSP-Mamba and SOTA SISR models on BSD100 [27] for \\(\\times 2\\) setting. TSP-Mamba series ('S'- 'Small', 'T'- 'Tiny') demonstrates highly competitive results. (2) Comparison of Effective Receptive Field for several SOTA methods. Darker area denotes larger receptive fields. Our proposed TSP-Mamba achieves the best global receptive field. Notably, larger circles indicate more model parameters (the first figure)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The Mamba method has recently gained attention for its capacity to model global signals with linear computational complexity. Its fixed scanning approach is well-suited for efficient 1D sequence modeling, achieving strong computational performance. However, this approach overlooks 2D structures within images, leading to inefficient aggregation of 2D information. As illustrated in Fig. 2, conventional vision-Mambas typically rely on partial neighboring propagation and aggregation along a fixed 1D scan path, which can disrupt the locality of 2D image information—referred to as \"local pixel forgetting\" in [14] and \"destroying local correlations\" in [34]—resulting in sub-optimal outcomes. As shown in Fig. 2(a), the local information aggregation with an image patch as highlighted in the red box, is less effective in Mamba-based scan due to the long-range (also partial)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28134" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.29, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.09, + 0.548, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.255, + 0.5, + 0.267 + ], + "angle": 0, + "content": "(a) 4 scan types in MambaIR" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.091, + 0.899, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.254, + 0.866, + 0.267 + ], + "angle": 0, + "content": "(b) Shortest traversal path search in TSP-Mamba" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.278, + 0.908, + 0.336 + ], + "angle": 0, + "content": "Figure 2. Illustration of vanilla Mamba and our proposed locally enhanced TSP-Mamba. For a highlighted \\(3 \\times 3\\) image patch (red box): (a) In Vanilla Mamba, only a limited set of neighboring pixels (e.g., '1, 2, 3, 4') can be aggregated for pixel '5' through a mix of long- and short-range propagation within a single global scan (\"Scan-1\"). (b) In TSP-Mamba, the shortest traversal path is constructed, allowing patch-adaptive propagation in a single local scan." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.345, + 0.484, + 0.42 + ], + "angle": 0, + "content": "propagation within a single scan. A common solution to this problem is stacking multiple multidirectional Mamba layers to improve aggregation performance. However, this will substantially increase computational cost and model complexity, hindering practicality in resource-limited scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.423, + 0.484, + 0.65 + ], + "angle": 0, + "content": "To address these limitations, we propose TSP-Mamba, which enhances the propagation and aggregation of local 2D information by searching for a locally adaptive, optimal propagation path for each image patch. Unlike the fixed (content-invariant) scan pipeline used in traditional Mamba, our method first identifies a patch-adaptive scan path with the lowest propagation cost, leveraging the shortest path principle. Here, each pixel in a local image patch is treated as a node in a graph, and we frame the problem as a simplified Traveling Salesman Problem (TSP) to find the minimum-cost traversal path, along which local information is transmitted and aggregated to enhance the use and extraction of local structures. Later on, the enhanced local information is effectively propagated into the global Mamba-style aggregation in a conventionally local-then-global manner." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.653, + 0.484, + 0.774 + ], + "angle": 0, + "content": "To implement this adaptive propagation, we design a TSP-LSTM block that modulates the aggregation between the current node and the previously accumulated information based on the propagation cost (i.e., the distance between the current and previous nodes). As demonstrated in Fig. 2(b), TSP-Mamba obtains a local shortest scan path for aggregation, thus leading to improved performance at a comparable and even reduced model complexity (shown in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.777, + 0.341, + 0.792 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.484, + 0.886 + ], + "angle": 0, + "content": "- Drawing inspiration from the shortest path principle, we propose a novel local structure-aware Mamba, TSP-Mamba, enabling content-adaptive propagation for efficient single-image super-resolution. We design an efficient TSP-LSTM to facilitate effective local aggregation by leveraging the locally optimal scan path." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- Extensive experiments show that the proposed TSP" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.345, + 0.907, + 0.389 + ], + "angle": 0, + "content": "Mamba framework achieves state-of-the-art performance across multiple lightweight SISR benchmarks, requiring only an additional 1-2 GFLOPs for 720P outputs (Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.39, + 0.907, + 0.451 + ], + "angle": 0, + "content": "- Beyond lightweight SISR, we also investigate the scalability and generalization capabilities of TSP-Mamba for various vision tasks, including classic SISR, and strong generalization for depth map super-resolution." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.463, + 0.663, + 0.478 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.909, + 0.747 + ], + "angle": 0, + "content": "CNN-based Methods. Due to their low complexity and effective feature extraction capabilities, CNNs have been widely employed for SISR tasks. SRCNN [9] was a pioneering work that introduced deep convolutional neural network (CNN) architectures specifically designed for single image super-resolution (SISR). SRCNN features a simple three-layer structure consisting of patch extraction, nonlinear mapping, and reconstruction, and it demonstrated competitive performance against traditional non-deep learning methods, sparking the development of numerous lightweight CNN-based approaches in the SISR domain. Inspired by dictionary-learning models [15, 17, 30], LAPAR [20] learns linear coefficients associated with pre-defined basic upsampling kernels to create an optimal pixel-specific kernel, leading to superior super-resolution outcomes. LatticeNet [26] introduces a parameter-efficient convolutional lattice block to extract hierarchical contextual features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Transformer-based Methods. On the other hand, Transformer-based techniques [4-6, 8, 21, 35, 37-39] have achieved impressive results in SISR but are still hindered by their high computational complexity. The cost of modeling non-local self-similarity scales quadratically with the size of the image. To address these efficiency concerns, many window-based Transformer frameworks have been developed, inspired by the success of the Swin Transformer [24]. For instance, SwinIR [23] introduces a residual window-based transformer block (RSTB) for general image restora" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28135" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.094, + 0.898, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.908, + 0.369 + ], + "angle": 0, + "content": "Figure 3. Illustration of our proposed TSP-Mamba, including a shallow feature extraction module ('Head'), a deep feature extraction module and a high-resolution reconstruction module ('Tail'). The key point is that the TSP-LSTM, employed for local path-adaptive aggregation, facilitates subsequent global scanning of 2D-SSM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.438 + ], + "angle": 0, + "content": "tion. DLGSA-1 [22] enhances token aggregation by using a global sparse attention mechanism. Omni-SR [33] presents spatial-channel attention for lightweight SISR, achieving better performance and efficiency trade-off." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.485, + 0.804 + ], + "angle": 0, + "content": "Mamba-based Methods. Recently, Mamba-based frameworks [13] have demonstrated impressive capabilities in modeling long-sequence dependencies while maintaining linear computational complexity. At the heart of Mamba is a structured state space model that expands the effective receptive field, allowing the framework to capture long-range dependencies more efficiently than traditional methods like CNNs and Transformers. This capacity to model global dependencies with linear complexity has positioned Mamba as a compelling alternative, particularly in resource-constrained environments. Building on Mamba's success, numerous follow-up studies [2, 11, 12, 14, 32] have explored the potential of designing different Mamba-based frameworks for various vision tasks. For example, MambaIR [14] offers a simple yet effective Mamba-based baseline for image restoration, while RetinexMamba [2] leverages Mamba's long-range dependency modeling for low-light image enhancement. Although these approaches integrate Mamba's global sequential reasoning (1D) into vision problems, the fixed scan pipeline often overlooks the locality and content-adaptive nature of 2D images. In this paper, we aim to develop an efficient and content-adaptive local propagation and aggregation method to improve the performance of Mamba-based frameworks, particularly for resource-constrained SISR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.191, + 0.836 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.303, + 0.862 + ], + "angle": 0, + "content": "3.1. Preliminaries: Mamba" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.901 + ], + "angle": 0, + "content": "We begin by presenting a concise overview of the Mamba pipeline, followed by a discussion of the limitations asso" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.908, + 0.439 + ], + "angle": 0, + "content": "ciated with the rigid 2D-to-1D scan scheme for low-level vision tasks. Selective structured state space models (SSMs) convert a 1D sequence \\(x(t)\\in \\mathbb{R}\\) to the response \\(y(t)\\in \\mathbb{R}\\) through a latent state \\(h(t)\\in \\mathbb{R}^N\\). This process is expressed:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.448, + 0.793, + 0.464 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.464, + 0.907, + 0.483 + ], + "angle": 0, + "content": "\\[\ny (t) = \\mathbf {C} h (t) + \\mathbf {D} x (t), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.908, + 0.581 + ], + "angle": 0, + "content": "where \\(\\mathbf{A} \\in \\mathbb{R}^{N \\times N}\\), \\(\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}\\), \\(\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}\\) and \\(D \\in \\mathbb{R}\\) are time-varying matrices and \\(N\\) is the state number. In practical applications, the continuous evolution system can be discretized using a zero-order hold assumption. Accordingly, Mamba converts the continuous matrices \\(\\mathbf{A}\\), \\(\\mathbf{B}\\) to their discrete forms \\(\\bar{\\mathbf{A}}\\), \\(\\bar{\\mathbf{B}}\\) with a timescale parameter \\(\\Delta \\in \\mathbb{R}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.587, + 0.72, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {A}} = \\exp (\\Delta \\mathbf {A}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.603, + 0.906, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {B}} = (\\Delta \\mathbf {A}) ^ {- 1} (e ^ {\\mathbf {A}} - \\mathbf {I}) \\cdot \\Delta \\mathbf {B}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.636, + 0.794, + 0.65 + ], + "angle": 0, + "content": "We then obtain a discretized formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.653, + 0.907, + 0.67 + ], + "angle": 0, + "content": "\\[\nh _ {k} = \\bar {\\mathbf {A}} h _ {k - 1} + \\bar {\\mathbf {B}} x _ {k}, \\quad y _ {k} = \\mathbf {C} h _ {k} + \\mathbf {D} x _ {k}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.908, + 0.721 + ], + "angle": 0, + "content": "As we can see, since the overall calculation is a linear combination of the accumulated sum, it can be efficiently implemented as a global 1D convolution:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.727, + 0.907, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {K}} = \\left(\\mathbf {C} \\bar {\\mathbf {B}}, \\mathbf {C} \\bar {\\mathbf {A}} \\bar {\\mathbf {B}}, \\dots , \\mathbf {C} \\bar {\\mathbf {A}} ^ {S - 1} \\bar {\\mathbf {B}}\\right), \\mathbf {y} = \\mathbf {x} * \\bar {\\mathbf {K}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.908, + 0.782 + ], + "angle": 0, + "content": "where \\(\\otimes\\) is a convolution operation and the \\(\\bar{\\mathbf{K}}\\in \\mathbb{R}^S\\) refers to a convolution kernel. \\(S\\) is the sequence length." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Discussion. In NLP tasks, it is natural to employ state space models (SSMs) to process 1D sequences for global information reasoning at a linear computational cost. However, in 2D image scenarios, the neighborhoods often face challenges such as disconnection or long-range propagation to fit within the SSM framework. Although some recent studies have investigated alternative 1D scan paths to supplement" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28136" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "information and improve restoration capability, capturing multiple local neighbors in a single 1D scan remains difficult, which can result in inefficient local aggregation for Mamba-based frameworks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.162, + 0.277, + 0.176 + ], + "angle": 0, + "content": "3.2. Preliminaries: TSP" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.485, + 0.304 + ], + "angle": 0, + "content": "Given \\( n \\) cities and their pairwise distances, the goal is to determine the shortest route that visits each city exactly once. Mathematically, TSP is formulated to find a minimum-cost Hamiltonian cycle in a weighted complete graph, where nodes represent cities and edges denote travel costs (e.g., distance or time). Inspired by TSP's optimal traversal property, we propose a TSP-style local scan scheme to complement the global modeling of vision Mambas in this work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.314, + 0.296, + 0.329 + ], + "angle": 0, + "content": "3.3. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.546 + ], + "angle": 0, + "content": "Fig. 3 illustrates the framework of our proposed TSP-Mamba. Following MambaIR, a recent SOTA baseline [14], TSP-Mamba comprises three main modules: (1) a shallow feature extraction module that transforms the RGB input into a feature space; (2) a deep feature extraction module consisting of multiple TSP-RSSGs (TSP-based Residual State-Space Groups). Each T-RSSG consists of several T-RSSBs (TSP-based Residual State-Space Block). In TSP-RSSB, the key component is the T-VSSM (TSP-based Vision State-Space Module), where we first use a TSP-LSTM layer to perform local path-adaptive aggregation and then pass the local-aggregation information into the 2D-SSM (2D Selective Scan Module in MambaIR[14]); and (3) a high-resolution reconstruction module that produces the final results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.623 + ], + "angle": 0, + "content": "In the following section, we mainly focus on the explanation of our proposed TSP-LSTM, including the scanning patch construction for optimal local propagation (Sec.3.4.1), simplified TSP (Sec.3.4.2) and efficient implementation (Sec.3.4.3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.632, + 0.219, + 0.646 + ], + "angle": 0, + "content": "3.4. TSP-LSTM" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.334, + 0.669 + ], + "angle": 0, + "content": "3.4.1. Scanning Path Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.484, + 0.735 + ], + "angle": 0, + "content": "Given an image feature patch \\(\\mathbf{P} \\in \\mathbb{R}^{s \\times s \\times C}\\), where \\(s\\) is the patch size and \\(C\\) is the number of channels, our objective is to establish a locally adaptive aggregation path that minimizes the information propagation costs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.484, + 0.843 + ], + "angle": 0, + "content": "Definition. We formulate the scan path as a Traveling Salesman Problem (TSP) problem: let \\( V = \\{1,2,\\dots n\\} \\) be the set of pixels (cities) in image patch \\( P \\) and \\( E = \\{e_{i,j}(i,j):i,j\\in V\\} \\) and \\( D_{e_{i,j}} \\) be the cost measure associated with edge \\( e_{i,j}\\in E \\). Our goal is to find a minimal scan path such that each city (pixel) is visited exactly once." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.844, + 0.483, + 0.874 + ], + "angle": 0, + "content": "In practice, we use the negative Cosine similarity as the cost measure of an edge \\( e_{i,j} \\) in the set \\( E \\):" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.885, + 0.484, + 0.904 + ], + "angle": 0, + "content": "\\[\nD _ {e _ {i j}} = 1. 0 - \\operatorname {C o s i n e} \\left(F _ {i}, F _ {j}\\right), \\tag {5}\n\\]" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.126, + 0.681, + 0.194 + ], + "angle": 0, + "content": "Figure 4. The numbers of scan path candidates of TSP and our simplified TSP scheme under different patch size \\(s\\)." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.206, + 0.667, + 0.283 + ], + "angle": 0, + "content": "
Size (s)TSPOurs
24!2
39!8
416!52
525!824
" + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.091, + 0.905, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.248, + 0.91, + 0.317 + ], + "angle": 0, + "content": "Figure 5. Speedup ratio curve under different patch sizes. Our effective measure hashing strategy attains notable speedup \\((\\log_{10})\\) for efficient TSP searching." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.905, + 0.358 + ], + "angle": 0, + "content": "where \\(F_{i}, F_{j}\\) are the corresponding feature vector of the pixel \\(i, j\\). More discussion is shown in Tab. 3 (A)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.364, + 0.666, + 0.379 + ], + "angle": 0, + "content": "3.4.2. Simplified TSP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.444 + ], + "angle": 0, + "content": "The Traveling Salesman Problem (TSP) is a classic NP-hard problem. With a large \\( n \\), searching for an optimal scan path becomes time-consuming (\\( O(n!) \\)). To simplify the problem, we adopt the following strategies in this work:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.444, + 0.905, + 0.473 + ], + "angle": 0, + "content": "(1) We limit the overall node count \\( n \\) to the pixels within a local image patch, thereby reducing overall complexity;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.474, + 0.905, + 0.504 + ], + "angle": 0, + "content": "(2) We designate the first city '1' (the top-left pixel of the given patch) as the starting node;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.905, + 0.548 + ], + "angle": 0, + "content": "(3) At each intermediate step, we limit next-step candidates to the four neighboring 2D locations1 to reduce the search space further." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.444, + 0.905, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.61 + ], + "angle": 0, + "content": "Based on these three simplifications mentioned, it largely reduced the number of total candidate scan paths (in Fig. 4). Thus, we can initially employ the backtracking algorithm [10] to identify all possible scan paths:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.617, + 0.842, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} T ^ {\\{1, 2, \\dots , l, \\dots , L \\}} = \\text {B a c k T r a c k i n g} (1, n), \\\\ \\text {w h e r e} T ^ {l} = \\left\\{i _ {1} ^ {l}, i _ {2} ^ {l}, \\dots , i _ {n} ^ {l} \\right\\}, i _ {1} ^ {l} = 1, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.908, + 0.713 + ], + "angle": 0, + "content": "where \\( l \\)-th path \\( T^l \\) is represented by the ordered pixel indices \\( (\\{i_1^l, i_2^l, \\dots, i_n^l\\}) \\) and \\( L \\) is the count of all scan candidates. We calculate the cost of the \\( l \\)-th scan path \\( T^l \\):" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.72, + 0.907, + 0.748 + ], + "angle": 0, + "content": "\\[\nD ^ {l} = \\sum_ {k = 1} ^ {n - 1} D _ {e _ {i _ {k} ^ {l}, i _ {k + 1} ^ {l}}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.757, + 0.905, + 0.786 + ], + "angle": 0, + "content": "Finally, we choose the optimal scan path \\( T^{l^*} \\) with the lowest measure cost:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.787, + 0.907, + 0.812 + ], + "angle": 0, + "content": "\\[\nl ^ {*} = \\underset {l} {\\operatorname {a r g m i n}} D ^ {l} \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.738, + 0.83 + ], + "angle": 0, + "content": "3.4.3. Effective Implementation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.835, + 0.908, + 0.88 + ], + "angle": 0, + "content": "As indicated in Eq. 7, we have to calculate the \\( n - 1 \\) pairwise distances for each scan path and sum them up. We have two observations for effective path cost computation:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.816, + 0.901 + ], + "angle": 0, + "content": "left \\((-1,0)\\), right \\((+1,0)\\), top \\((0,-1)\\), and bottom \\((0,+1)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28137" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "1. Symmetry of Distance Measurement. As indicated in Eq. 5, for edges \\( e_{i,j} \\), \\( e_{j,i} \\), we have \\( D_{e_{i,j}} = D_{e_{j,i}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.152 + ], + "angle": 0, + "content": "2. Highly overlapping edges for different scan paths. Different scene paths may share lots of the same edges." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.484, + 0.273 + ], + "angle": 0, + "content": "Based on this fact, we propose to design a hash-based technique to store the unique edges in all scan paths into an edge list \\(\\{e_{k_0}, e_{k_1}, \\dots, e_{k_m}\\}\\) (\\(m\\) is the length of the unique edge list) to avoid repeated distance calculation. Accordingly, we can effectively compute all unique edge distances using Eq. 5 at one time and obtain the associated edge distance list \\(\\{D_{e_{k_0}}, D_{e_{k_1}}, \\dots, D_{e_{k_m}}\\}\\). Finally, we can obtain the path propagation cost in Eq. 7 by accumulating all its edges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.484, + 0.349 + ], + "angle": 0, + "content": "The speedup ratio between the direct per-path cost computation and our efficiency implementation is shown in Fig. 5, we can see our effective hash-based calculation attains a significant speedup effect. Additionally, it becomes more effective with an increased patch size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.46 + ], + "angle": 0, + "content": "Summary. In this part, we propose a simplified TSP formulation to effectively search the shortest path for a given 2D image patch, in which all neighborhoods are fused in a relatively short-term propagation. Therefore, it is capable of effectively integrating local information to facilitate the later global SSM-style propagation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.219, + 0.482 + ], + "angle": 0, + "content": "3.4.4. TSP-LSTM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.593 + ], + "angle": 0, + "content": "Bidirectional Propagation. After the scan path optimization, we perform a bidirectional LSTM for mutual information propagation. Especially, given a 2D image patch \\(\\mathbf{P}\\), we use the above simplified TSP scheme to get the optimal scan chain \\(T^{l^{*}} = \\{i_{1} \\rightleftharpoons i_{2}, \\dots, \\rightleftharpoons i_{n}\\}^{2}\\), where “\\(\\rightleftharpoons\\)” means we employ the bidirectional propagation scheme (as shown in Fig. 2) (b) for effective and balanced aggregation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.594, + 0.484, + 0.639 + ], + "angle": 0, + "content": "For implementation, we flip the reordered sequence \\(\\{i_1,i_2,\\dots ,i_n\\}\\) to obtain the inverse sequence and concatenate them for bidirectional aggregation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Distance Embedding. Within an image patch, discontinuities may need to be addressed in the chain-style propagation. Specifically, the aggregation weight of the previously accumulated information should positively correlate with the distance between the previous and current nodes (pixels). To account for this, inspired by fast global image smoothing (FGS [29]), we incorporate distance embedding features into the sequences to explicitly model this effect. We calculate the pairwise distances as shown in Eq. 7 and embed them (as a one-channel feature) into the propagation sequence feature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Lastly, the propagated feature \\(\\mathbf{x}\\) is passed through the 2DSSM module (in Eq. 4) for global Mamba-style aggregation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.645, + 0.107 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.611, + 0.132 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Implementation Details. Our TSP-Mamba model is implemented in PyTorch and integrates several common data augmentation techniques, such as random cropping, vertical/horizontal flipping, and rotation. For training, we use the Adam optimizer [18] with cosine annealing [25], starting with an initial learning rate of \\(4e - 4\\). The batch size is set to 32, with input images sized at \\(64 \\times 64\\). Training is performed over 600K iterations using four NVIDIA RTX 3090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.26, + 0.909, + 0.426 + ], + "angle": 0, + "content": "Scalable Model Configuration. In general, we train multiple TSP-Mamba models with different configurations. TSP-Mamba is based on the MambaIR baseline, utilizing 4 T-RSSGs, each containing 6 T-RSSBs, with a channel size of 60, denoted as “[6,6,6,6]@C60”. Additionally, we investigate the scalability of TSP-Mamba to better understand the trade-off between performance and model complexity. Specially, we use “[2,4,2,6]@C40”\\(^{4}\\) and “[4,6,4,6]@C48”\\(^{5}\\) to train another two smaller SISR models under \\(\\times 2\\) SISR setting. Training these TSP-Mamba models usually takes around 3-6 days." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.427, + 0.909, + 0.547 + ], + "angle": 0, + "content": "Benchmark Datasets. In line with previous studies [7, 20, 23], we use the DIV2K [1] image dataset for training. We then perform extensive evaluations on several widely-used SISR benchmarks, including Set5 [3], Set14 [36], BSD100 [27], Urban100 [16], and Manga109 [28]. Our quantitative analysis is based on PSNR and SSIM metrics, and following standard practice, we report results specifically for the Y channel in the YCbCr color space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.806, + 0.575 + ], + "angle": 0, + "content": "4.2. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.581, + 0.909, + 0.686 + ], + "angle": 0, + "content": "We perform extensive comparisons with a wide range of lightweight SISR models: LAPAR-A (NeurIPS20) [20], LatticeNet (ECCV20) [26], RLFN (CVPRW22) [19], SwinIR-light [23], DLGSA-1 (ICCV23) [22], Onmi-SR (CVPR23) [33], SRFormer (ICCV23) [39], ATDS (CVPR24) [37], MambaIR (ECCV24) [14], HiTSR [38] (ECCV24)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.688, + 0.909, + 0.808 + ], + "angle": 0, + "content": "Quantitative Comparison. Tab. 1 shows the quantitative evaluation results, highlighting that our proposed TSP-Mamba consistently outperforms existing methods across all benchmarks. In particular, our model achieves significant improvements over MambaIR, with gains of up to \\(0.71\\mathrm{dB}\\), \\(0.58\\mathrm{dB}\\), and \\(0.69\\mathrm{dB}\\) on the Manga109 dataset for the \\(\\times 2\\), \\(\\times 3\\), and \\(\\times 4\\) scales, respectively, while maintaining similar parameter and FLOP requirements." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.809, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Moreover, TSP-Mamba-Small beats all leading lightweight SISR methods with smaller model sizes and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.876 + ], + "angle": 0, + "content": "2For clarity, we drop the upper letter and denote \\(T^{l^*}\\) as \\(T^{l^*} = \\{i_1,i_2,\\dots ,i_n\\}\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.876, + 0.484, + 0.901 + ], + "angle": 0, + "content": "3Each pixel can fuse all its neighboring pixels in the bidirectional short-range propagation scheme" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.849, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.851, + 0.905, + 0.875 + ], + "angle": 0, + "content": "4TSP-Mamba-Tiny: 4 TSP-RSSGs contain 2,4,2,6 TSP-RSSBs and the channel size is 40" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.876, + 0.905, + 0.899 + ], + "angle": 0, + "content": "5TSP-Mamba-Small: 4 TSP-RSSGs contain 4,6,4,6 TSP-RSSBs and the channel size is 48" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.851, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28138" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with SOTA lightweight SISR methods on multiple benchmark datasets. The best and second-best results on the default training setting (DIV2K) are highlighted in red and blue, respectively. We follow most lightweight SISR methods and calculate all Flops with an output size of \\( {1280} \\times {720} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.143, + 0.885, + 0.721 + ], + "angle": 0, + "content": "
MethodScaleParameters (K)Flops (G)Set5 PSNR / SSIMSet14 PSNR / SSIMBSD100 PSNR / SSIMUrban100 PSNR / SSIMManga109 PSNR / SSIM
LAPAR-A54817138.01 / 0.960533.62 / 0.918332.19 / 0.899932.10 / 0.928338.67 / 9772
LatticeNet756169.538.15 / 0.961033.78 / 0.919332.25 / 0.900532.43 / 0.9302- / -
SwinIR-light910244.438.14 / 0.961133.86 / 0.920632.31 / 0.901232.76 / 0.934039.12 / 0.9783
DLGSA-light745170.038.20 / 0.961233.89 / 0.920332.30 / 0.901232.94 / 0.935539.29 / 0.9780
Omni-SR772194.538.22 / 0.961333.98 / 0.921032.36 / 0.902033.05 / 0.936339.28 / 0.9784
SRFormer ×285323638.23 / 0.961333.94 / 0.920932.36 / 0.901932.91 / 0.935339.28 / 0.9785
HiT-SR847226.538.26 / 0.961534.01 / 0.921432.37 / 0.902333.13 / 0.937239.47 / 0.9787
ATD-S75334938.29 / 0.961634.10 / 0.921732.39 / 0.902333.27 / 0.937539.52 / 0.9789
MambaIR136319838.16 / 0.961034.00 / 0.921232.34 / 0.901732.92 / 0.935639.31 / 0.9779
TSP-Mamba130519938.46 / 0.964834.42 / 0.924332.52 / 0.904233.57 / 0.940540.02 / 0.9808
MambaIR-Small80611138.09 / 0.960233.85 / 0.920132.31 / 0.901532.58 / 0.932139.17 / 0.9787
TSP-Mamba-Small76911238.22 / 0.964034.19 / 0.922732.43 / 0.903033.18 / 0.937439.49 / 0.9791
MambaIR-Tiny4435937.99 / 0.960233.72 / 0.910932.21 / 0.900232.22 / 0.929138.87 / 0.9772
TSP-Mamba-Tiny4185938.04 / 0.963133.90 / 0.920532.29 / 0.901032.55 / 0.931739.12 / 0.9784
LAPAR-A54411434.36 / 0.926730.34 / 0.842129.11 / 0.805428.15 / 0.852333.51 / 9441
LatticeNet76576.334.53 / 0.928130.39 / 0.842429.15 / 0.805928.33 / 0.8538- / -
SwinIR-light918110.834.62 / 0.928930.54 / 0.846329.20 / 0.808228.66 / 0.862433.98 / 0.9478
DLGSA-light75275.434.70 / 0.929530.58 / 0.846529.24 / 0.808928.83 / 0.865334.16 / 0.9483
Omni-SR78088.434.70 / 0.929430.57 / 0.846929.28 / 0.809428.84 / 0.865634.22 / 0.9487
SRFormer ×386110534.67 / 0.929630.57 / 0.846929.26 / 0.809928.81 / 0.865534.19 / 0.9489
HiT-SR85510234.75 / 0.930030.61 / 0.847529.29 / 0.810628.99 / 0.868734.53 / 0.9502
MambaIR13718934.72 / 0.929630.63 / 0.847529.29 / 0.809929.00 / 0.868934.39 / 0.9495
TSP-Mamba13138934.80 / 0.933230.93 / 0.851029.41 / 0.813929.43 / 0.874434.97 / 0.9522
MambaIR-Small8125034.65 / 0.931930.59 / 0.847129.26 / 0.810228.84 / 0.864434.02 / 0.9472
TSP-Mamba-Small7765034.65 / 0.932230.66 / 0.847829.30 / 0.811428.94 / 0.868034.41 / 0.9495
MambaIR-Tiny4482734.36 / 0.922330.26 / 0.842829.12 / 0.806328.24 / 0.853233.74 / 0.9452
TSP-Mamba-Tiny4232734.49 / 0.930930.46 / 0.845829.22 / 0.809428.70 / 0.862334.16 / 0.9483
LAPAR-A6599432.15 / 0.894428.61 / 0.781827.61 / 0.736626.14 / 0.787130.42 / 0.9074
LatticeNet77743.632.30 / 0.896228.68 / 0.783027.62 / 0.736726.25 / 0.7873- / -
SwinIR-light93063.632.44 / 0.897628.77 / 0.785827.69 / 0.740626.47 / 0.798030.92 / 0.9151
DLGSA-light76142.532.54 / 0.899328.84 / 0.787127.73 / 0.741526.66 / 0.803331.13 / 0.9161
Omni-SR79250.932.49 / 0.898828.78 / 0.785927.71 / 0.741526.64 / 0.801831.02 / 0.9151
SRFormer ×48736332.51 / 0.898828.82 / 0.787227.73 / 0.742226.67 / 0.803231.17 / 0.9165
HiT-SR8665832.55 / 0.899928.87 / 0.788027.75 / 0.743226.80 / 0.806931.26 / 0.9171
ATD-S7698732.63 / 0.899828.89 / 0.788627.79 / 0.744026.97 / 0.810731.48 / 0.9198
MambaIR13835132.51 / 0.899328.85 / 0.787627.75 / 0.742326.75 / 0.805131.26 / 0.9175
TSP-Mamba13245332.71 / 0.902429.05 / 0.792527.91 / 0.747627.29 / 0.817431.95 / 0.9235
MambaIR-Small8212932.28 / 0.900128.70 / 0.786327.72 / 0.741226.56 / 0.798431.12 / 0.9156
TSP-Mamba-Small7853032.48 / 0.902428.90 / 0.789327.79 / 0.744226.93 / 0.808731.42 / 0.9184
MambaIR-Tiny4561532.22 / 0.899128.63 / 0.784727.68 / 0.739826.43 / 0.794030.86 / 0.9124
TSP-Mamba-Tiny4311632.29 / 0.899728.79 / 0.786027.72 / 0.740526.61 / 0.799231.04 / 0.9140
MambaIR (×16)1772523.24 / 0.641922.29 / 0.545422.94 / 0.521820.46 / 0.514821.15 / 0.6628
TSP-Mamba (×16)1714523.33 / 0.651522.38 / 0.545222.81 / 0.516920.53 / 0.518421.20 / 0.6700
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.778 + ], + "angle": 0, + "content": "less computational complexity. For instance, it surpasses MambaIR by 0.26dB on Urban \\(100 \\times 2\\), while only requiring \\(59\\%\\) parameters and \\(57\\%\\) Flops." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Inference efficiency. Tab. 2 presents a detailed comparison of inference costs across different models, including CNN-based, Transformer-based, and Mamba-based frameworks. Overall, the CNN-based LatticeNet achieves the highest inference efficiency. Compared to other transformer-based and Mamba-based models, our proposed TSP-Mamba achieves competitive inference latency." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Qualitative Comparison. Fig. 6 showcases several visual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.793 + ], + "angle": 0, + "content": "examples. It is clear that TSP-Mamba effectively reconstructs accurate image textures with fewer super-resolved artifacts. In contrast, other approaches struggle to restore clear image patterns." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.679, + 0.829 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To validate our design, we conduct extensive experiments, reporting quantitative results on the Urban100 and Manga109 benchmarks at the \\(\\times 4\\) setting. We primarily adopt the TSP-Mamba-Small architecture for fast analysis." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28139" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.902, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.325, + 0.895, + 0.34 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of different lightweight SISR methods under \\(\\times 4\\) setting. Our predictions present more detailed structures." + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.353, + 0.908, + 0.384 + ], + "angle": 0, + "content": "Table 2. Inference cost comparison of different lightweight SISR models under \\( \\times 2 \\) setting. Inference time is tested with an output resolution of \\( {256} \\times {256} \\) using an NVIDIA 3090 GPU. The top-3 results are highlighted with red,blue,underline." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.386, + 0.871, + 0.446 + ], + "angle": 0, + "content": "
MetricsLatticeNetDLGSA-lightSwinIR-lightSRFormerMambaIR / Small / TinyTSP-Mamba / Small / Tiny
Time (ms) ↓973416060 / 50 / 3872 / 58 / 39
Memory (MB) ↓146616245441572686 / 536 / 528732 / 614 / 582
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.486, + 0.514 + ], + "angle": 0, + "content": "Different Measure Cost in TSP-LSTM. Apart from Eq. 5, we try the widely-used L1/L2 distance metrics for scan path searching. Tab. 3 (A) concludes that our used negative Cosine distance performs better than the L1/L2 distances." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.673 + ], + "angle": 0, + "content": "TSP-LSTM vs. Uniform-LSTM. TSP-LSTM aims to perform an adaptive scan path by constructing a minimal propagation path. To assess the impact of our adaptive scan path, we also train two uniform scan models (vertical, and horizontal). We compare the proposed TSP-LSTM with the other models (UV-LSTM, HV-LSTM). As shown in Tab. 3 (B), we can observe that TSP-LSTM significantly outperforms the uniform scan models, highlighting the importance of adaptive local propagation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.487, + 0.798 + ], + "angle": 0, + "content": "Different Patch Sizes. As mentioned in Sec. 3.4, we set the patch size to \\(4 \\times 4\\) considering the local receptive field and computational costs. When the patch size is increased, it becomes prolonged to construct the corresponding TSP path. We investigate the influence of different patch sizes. Tab. 3 (C) suggests the reconstructed performance is positively related to the patch size but with larger latency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.486, + 0.909 + ], + "angle": 0, + "content": "Finer Design in TSP-LSTM. Lastly, we evaluate the impact of bidirectional propagation of TSP-LSTM. We train another model with only a forward/backward scan pipeline to do this. In Tab. 3 (D), it is clear that our bidirectional propagation scheme is more effective than its counterparts with only a single-direction scan strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.909, + 0.561 + ], + "angle": 0, + "content": "Different Aggregation Layers. In this paper, we present a TSP-LSTM block to get locally fused information. To verify the effectiveness of TSP-LSTM, we replace it with other designs for local compensation, such as convolution, RNN, and Mamba. The results are shown in Tab. 3 (E) below, which suggests that our TSP attains the best performance and second-best efficiency compared with other designs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.569, + 0.637, + 0.584 + ], + "angle": 0, + "content": "4.4. Extensions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.909, + 0.773 + ], + "angle": 0, + "content": "Classic SISR. Though TSP-Mamba is trained for lightweight Mamba-based architecture, i.e., MambaIR, we additionally explore the potential improvement for the classic SISR setting. To do this, we follow MambaIR-L(large) to enlarge the channel size to 180 and use 6 TSP-RSSG blocks (each one contains 6 basic TSP-SSD layers) to obtain the TSP-Mamba-L model and train this classic SISR model using the same DF2K training dataset. The quantitative results are illustrated in Tab. 4. We can see that TSP-Mamba-L consistently outperforms other classic SISR methods significantly. For instance, it achieves 0.32dB, 0.37dB, and 0.39dB improvements over MambaIR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.909, + 0.868 + ], + "angle": 0, + "content": "Depth Map SR. We also compare the TSP-Mamba model with our baseline model (MambaIR) on the depth SR task for generalization analysis. To do this, without training on any depth images, we directly assess the two models on the NYU-V2 [31] depth benchmark\\(^6\\) under \\(\\times 4\\) and \\(\\times 16\\) settings." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.903 + ], + "angle": 0, + "content": "\\(^{6}\\)NYU V2 consists of 449 testing depth images at a size of \\(480 \\times 640\\). We use bicubic interpolation to generate the low-resolution depth images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28140" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.174 + ], + "angle": 0, + "content": "Table 3. Ablation Studies. A: The effects of different distance measures, with our default choice being the negative cosine distance as indicated in Eq. 5. B: Quantitative evaluation of uniform scanning strategies (vertical and horizontal), abbreviated as 'UV' and 'HV', compared to our TSP-based scan. C: The influence of different local scan patch sizes (P2-P5 indicates local patch sizes ranging from 2 to 5). D: Impacts of distance embeddings and bidirectional propagation. M1 drops both of the two designs. M2 only adopts distance embedding. M3 only uses the bidirectional propagation. M4 employs both of the two designs. E: The effects of various local aggregation modules (M5-M8 employ Conv1D, RNN, Mamba, and LSTM, respectively)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.178, + 0.912, + 0.29 + ], + "angle": 0, + "content": "
AnalysisA.B.C.D.E.
L1L2Neg. Cos.UVUHTSPP2P3P4P5M1M2M3M4M5M6M7M8
Urban10026.5526.5426.6126.3126.3826.6126.3726.4326.6126.6126.1626.3826.2526.6126.4926.4426.4726.61
0.74030.79820.79920.79070.79360.79920.79330.79430.79920.79940.78810.79390.79070.79920.79580.79500.79560.7992
Manga10931.0631.0731.0430.7830.8631.0430.7530.9131.0431.0230.5030.8030.6531.0430.9630.9530.9731.04
0.91400.91440.91400.91040.91260.91400.91130.91310.91400.91380.90650.91030.90840.91400.91360.91320.91340.9140
Param.(K)431431431431431431431431431431430431430431440436428431
FLOPs (G)16.216.216.215.415.416.215.215.216.220.314.915.116.116.215.515.716.916.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.308, + 0.484, + 0.364 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison with recent SOTA Classic SR models. All the models are trained on the DF2K dataset for fair comparison. We report the averaged PSNR/SSIM values over the five benchmarks (Set5, Set14, BSD100, Urban100, Manga109)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.375, + 0.49, + 0.511 + ], + "angle": 0, + "content": "
SettingsMetricsSwinIRSRFormer-LHATMambaIR-LTSP-Mamba-L
×2PSNR35.8335.9436.1736.0536.37
SSIM0.94280.94350.94460.94380.9457
Param.11.8M10.4M20.6M20.4M20.6M
×3PSNR32.0532.1532.2932.1932.56
SSIM0.88720.88850.88810.89000.8918
Param.11.8M10.6M20.8M20.6M20.8M
×4PSNR29.8929.9730.1430.0430.43
SSIM0.83990.84160.84410.84140.8470
Param.11.8M10.5M20.8M20.6M20.8M
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.486, + 0.687 + ], + "angle": 0, + "content": "We use the PSNR, SSIM, and RMSE (the root-mean-square error) metrics for quantitative evaluation. The quantitative results are shown in the left Fig. 7. We can see that TSP-Mamba consistently outperforms its baseline model on all the objective metrics. For instance, TSP-Mamba exhibits superior performance compared to MambaIR with a PSNR improvement of 0.96dB (0.92dB) for \\(\\times 4\\) (\\(\\times 16\\)) configurations. The visual examples in Fig. 7 illustrate that TSP-Mamba generates clearer structures, leading to higher accuracy when compared to the baseline model." + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.698, + 0.471, + 0.713 + ], + "angle": 0, + "content": "Table 5. The quantitative results on NYU-V2 [31] \\((\\times 4,\\times 16)\\)" + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.724, + 0.458, + 0.796 + ], + "angle": 0, + "content": "
SettingsMetricsMambaIRTSP-Mamba
×4 (×16)PSNR47.25(37.03)48.21(37.95)
SSIM0.994(0.967)0.995(0.972)
RMSE2.339(7.895)2.145(7.142)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Limitations. Although our proposed TSP-Mamba demonstrates promising and significant improvements across several single-image super-resolution tasks, it remains challenging to reliably obtain accurate distance measurements for degraded images, particularly in the presence of high-frequency degradations. Noise components can adversely affect the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.308, + 0.651, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.387, + 0.631, + 0.395 + ], + "angle": 0, + "content": "PSNR/SSIM/RMSE" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.309, + 0.785, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.387, + 0.762, + 0.395 + ], + "angle": 0, + "content": "37.45/0.9698/0.0330" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.309, + 0.919, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.387, + 0.899, + 0.395 + ], + "angle": 0, + "content": "38.79/0.9750/0.0282" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.396, + 0.651, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.463, + 0.632, + 0.472 + ], + "angle": 0, + "content": "PSNR/SSIM/RMSE" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.396, + 0.785, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.463, + 0.762, + 0.472 + ], + "angle": 0, + "content": "34.80/0.9672/0.0673" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.396, + 0.918, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.463, + 0.898, + 0.472 + ], + "angle": 0, + "content": "36.45/0.9742/0.0555" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison between our baseline and the proposed TSP-Mamba for unseen depth-SR maps. We also show normalized error maps in the left corner for better visual comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.62 + ], + "angle": 0, + "content": "calculation of pair-wise distances between degraded pixels and the optimal scan path searching. In future work, we plan to address the issue of robust distance calculation under various image degradations, such as pyramid representation to build robust 1D TSP scan paths hierarchically." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.812 + ], + "angle": 0, + "content": "In this work, we introduce a novel locally adaptive propagation method for lightweight SISR. To this end, we develop an efficient scan path search strategy to enhance local information within the globally Mamba-based architecture. Extensive experiments validate the effectiveness of our approach, with TSP-Mamba achieving state-of-the-art performance across multiple SISR benchmarks. Furthermore, TSP-Mamba demonstrates superior generalization on unseen data (such as depth-map SR), and scalability across both ultralightweight and larger model configurations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work is partially supported by Shenzhen Science and Technology Program KQTD20210811090149095 and also the Pearl River Talent Recruitment Program 2019QN01X226." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28141" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.485, + 0.214 + ], + "angle": 0, + "content": "[2] Jiesong Bai, Yuhao Yin, and Qiyuan He. Retinexmamba: Retinex-based mamba for low-light image enhancement. arXiv preprint arXiv:2405.03349, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[4] Ke Chen, Liangyan Li, Huan Liu, Yunzhe Li, Congling Tang, and Jun Chen. Swinfsr: Stereo image super-resolution using swinir and frequency domain knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764-1774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[5] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.484, + 0.483 + ], + "angle": 0, + "content": "[6] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12312-12321, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.484, + 0.542 + ], + "angle": 0, + "content": "[7] Haram Choi, Jeongmin Lee, and Jihoon Yang. N-gram in swim transformers for efficient lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2071–2081, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[8] Tao Dai, Jianping Wang, Hang Guo, Jinmin Li, Jinbao Wang, and Zexuan Zhu. Freqformer: frequency-aware transformer for lightweight image super-resolution. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 731-739, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.672, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[10] Eugene C Freuder and Alan K Mackworth. Constraint satisfaction: An emerging paradigm. In Foundations of artificial intelligence, pages 13-27. Elsevier, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[11] Guanyiman Fu, Fengchao Xiong, Jianfeng Lu, and Jun Zhou. Ssumamba: Spatial-spectral selective state space model for hyperspectral image denoising. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.484, + 0.827 + ], + "angle": 0, + "content": "[12] Hu Gao, Bowen Ma, Ying Zhang, Jingfan Yang, Jing Yang, and Depeng Dang. Learning enriched features via selective state spaces model for efficient image deblurring. In ACM Multimedia 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.906, + 0.12 + ], + "angle": 0, + "content": "image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[15] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[16] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[17] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3224-3232, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.906, + 0.433 + ], + "angle": 0, + "content": "[19] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 766-776, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[20] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. Lapar: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. Advances in Neural Information Processing Systems, 33:20343-20355, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[21] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[22] Xiang Li, Jiangxin Dong, Jinhui Tang, and Jinshan Pan. Dlgstanet: Lightweight dynamic local and global self-attention networks for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12792–12801, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.907, + 0.701 + ], + "angle": 0, + "content": "[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image superresolution with lattice block. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 272-289. Springer, 2020. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28142" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[27] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.485, + 0.233 + ], + "angle": 0, + "content": "[28] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multi-media Tools and Applications, 76:21811-21838, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[29] Dongbo Min, Sunghwan Choi, Jiangbo Lu, Bumsub Ham, Kwanghoon Sohn, and Minh N Do. Fast global image smoothing based on weighted least squares. IEEE Transactions on Image Processing, 23(12):5638-5653, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.333 + ], + "angle": 0, + "content": "[30] Yaniv Romano, John Isidoro, and Peyman Milanfar. Raisr: Rapid and accurate image super resolution. IEEE Transactions on Computational Imaging, 3(1):110-125, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[31] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[32] Yujin Tang, Peijie Dong, Zhenheng Tang, Xiaowen Chu, and Junwei Liang. Vmrnn: Integrating vision mamba and lstm for efficient and accurate spatiotemporal forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5663-5673, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[33] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Jinfan Liu. Omni aggregation networks for lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22378-22387, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[34] Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. arXiv preprint arXiv:2407.21773, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.484, + 0.688 + ], + "angle": 0, + "content": "[35] Jinsu Yoo, Taehoon Kim, Sihaeng Lee, Seung Hwan Kim, Honglak Lee, and Tae Hyun Kim. Enriched cnn-transformer feature aggregation networks for super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4956-4965, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[36] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.842 + ], + "angle": 0, + "content": "[37] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[38] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. In European Conference on Computer Vision, pages 483-500. Springer, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[39] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12780–12791, 2023. 2, 5" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28143" + } + ] +] \ No newline at end of file diff --git a/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/989548d5-80da-4477-a3a5-42129f121eff_origin.pdf b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/989548d5-80da-4477-a3a5-42129f121eff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d773a7d11da374a3715da1b0be803be44bbd48ca --- /dev/null +++ b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/989548d5-80da-4477-a3a5-42129f121eff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30bd4ab7926610d5c85a295ca79ebddb82f1a960036fbdcdd94e6bf3dd2b43b3 +size 2297110 diff --git a/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/full.md b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/full.md new file mode 100644 index 0000000000000000000000000000000000000000..18717486acc7c791bb81eecc4af7d5cc79dcbc71 --- /dev/null +++ b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/full.md @@ -0,0 +1,336 @@ +# TSP-Mamba: The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond + +Kun Zhou1 Xinyu Lin2 Jiangbo Lu1* + +$^{1}$ SmartMore Corporation $^{2}$ SSE, CUHK-Shenzhen + +zhoukun303808@gmail.com, linxinyu0327@gmail.com, jiangbo.lu@gmail.com + +# Abstract + +Recently, Mamba-based frameworks have achieved substantial advancements across diverse computer vision and NLP tasks, particularly in their capacity for reasoning over long-range information with linear complexity. However, the fixed 2D-to-1D scanning pattern overlooks the local structures of an image, limiting its effectiveness in aggregating 2D spatial information. While stacking additional Mamba layers can partially address this issue, it increases parameter intensity and constrains real-time application. In this work, we reconsider the local optimal scanning path in Mamba, enhancing the rigid and uniform 1D scan through the local shortest path theory, thus creating a structure-aware Mamba suited for lightweight single-image super-resolution. Specifically, we draw inspiration from the Traveling Salesman Problem (TSP) to establish a local optimal scanning path for improved structural 2D information utilization. Here, local patch aggregation occurs in a content-adaptive manner with minimal propagation cost. TSP-Mamba demonstrates substantial improvements over existing Mamba-based and Transformer-based architectures. For example, TSP-Mamba surpasses MambaIR by up to 0.7dB in lightweight SISR, with comparable parameters and very slightly extra computational demands (1-2 GFlops for 720P images). + +# 1. Introduction + +Deep learning techniques have advanced single-image superresolution (SISR) tasks considerably, enhancing image reconstruction quality. Convolutional Neural Networks (CNNs) have been central to this progress, excelling at capturing local, content-dependent image patterns. However, their local focus limits their capacity to model long-range dependencies. Transformers, by contrast, are highly effective in capturing global information through their self-attention mechanism, enabling them to model non-local dependencies. Yet, their quadratic computational complexity poses significant challenges in resource-constrained environments. + +![](images/e262c5ed2eefaac09d41c89cb9e9e4b2baf6d109607d0b6a440bd20854e60d7e.jpg) + +![](images/e3e4d1f00c614483fc8b1c4c5aacbe50a01ed17603d9c6eed075c0a8b13057c5.jpg) +Figure 1. (1) Comparison between our proposed TSP-Mamba and SOTA SISR models on BSD100 [27] for $\times 2$ setting. TSP-Mamba series ('S'- 'Small', 'T'- 'Tiny') demonstrates highly competitive results. (2) Comparison of Effective Receptive Field for several SOTA methods. Darker area denotes larger receptive fields. Our proposed TSP-Mamba achieves the best global receptive field. Notably, larger circles indicate more model parameters (the first figure). + +![](images/8d49b319205b9334996d91f908dee7525683ef9ce8689d863218b0f02550dd55.jpg) + +![](images/705f6488904d86c70e2030088ab569888459dbcdaab91e03f9da02a27051e02f.jpg) + +![](images/4d6e74f302e9322baa43167dd5d8cd9bbb6b1aa82412d485873697cbf5181e2a.jpg) + +The Mamba method has recently gained attention for its capacity to model global signals with linear computational complexity. Its fixed scanning approach is well-suited for efficient 1D sequence modeling, achieving strong computational performance. However, this approach overlooks 2D structures within images, leading to inefficient aggregation of 2D information. As illustrated in Fig. 2, conventional vision-Mambas typically rely on partial neighboring propagation and aggregation along a fixed 1D scan path, which can disrupt the locality of 2D image information—referred to as "local pixel forgetting" in [14] and "destroying local correlations" in [34]—resulting in sub-optimal outcomes. As shown in Fig. 2(a), the local information aggregation with an image patch as highlighted in the red box, is less effective in Mamba-based scan due to the long-range (also partial) + +![](images/2993a053834e6b728e4db49b991a7170eb369744dce0a2cd5f092fb877fac4c0.jpg) +Figure 2. Illustration of vanilla Mamba and our proposed locally enhanced TSP-Mamba. For a highlighted $3 \times 3$ image patch (red box): (a) In Vanilla Mamba, only a limited set of neighboring pixels (e.g., '1, 2, 3, 4') can be aggregated for pixel '5' through a mix of long- and short-range propagation within a single global scan ("Scan-1"). (b) In TSP-Mamba, the shortest traversal path is constructed, allowing patch-adaptive propagation in a single local scan. + +![](images/cac057b1c8b8dd7a4979044468e19a9d2762c444cd2359c078615aeadc80b2a0.jpg) +(a) 4 scan types in MambaIR + +![](images/debd949819ca88cd310f0e27a873d8eabf6ee80cef3a4e6d8bf04dbd72174173.jpg) +(b) Shortest traversal path search in TSP-Mamba + +propagation within a single scan. A common solution to this problem is stacking multiple multidirectional Mamba layers to improve aggregation performance. However, this will substantially increase computational cost and model complexity, hindering practicality in resource-limited scenarios. + +To address these limitations, we propose TSP-Mamba, which enhances the propagation and aggregation of local 2D information by searching for a locally adaptive, optimal propagation path for each image patch. Unlike the fixed (content-invariant) scan pipeline used in traditional Mamba, our method first identifies a patch-adaptive scan path with the lowest propagation cost, leveraging the shortest path principle. Here, each pixel in a local image patch is treated as a node in a graph, and we frame the problem as a simplified Traveling Salesman Problem (TSP) to find the minimum-cost traversal path, along which local information is transmitted and aggregated to enhance the use and extraction of local structures. Later on, the enhanced local information is effectively propagated into the global Mamba-style aggregation in a conventionally local-then-global manner. + +To implement this adaptive propagation, we design a TSP-LSTM block that modulates the aggregation between the current node and the previously accumulated information based on the propagation cost (i.e., the distance between the current and previous nodes). As demonstrated in Fig. 2(b), TSP-Mamba obtains a local shortest scan path for aggregation, thus leading to improved performance at a comparable and even reduced model complexity (shown in Fig. 1). + +In summary, our contributions are: + +- Drawing inspiration from the shortest path principle, we propose a novel local structure-aware Mamba, TSP-Mamba, enabling content-adaptive propagation for efficient single-image super-resolution. We design an efficient TSP-LSTM to facilitate effective local aggregation by leveraging the locally optimal scan path. +- Extensive experiments show that the proposed TSP + +Mamba framework achieves state-of-the-art performance across multiple lightweight SISR benchmarks, requiring only an additional 1-2 GFLOPs for 720P outputs (Tab. 1). + +- Beyond lightweight SISR, we also investigate the scalability and generalization capabilities of TSP-Mamba for various vision tasks, including classic SISR, and strong generalization for depth map super-resolution. + +# 2. Related Works + +CNN-based Methods. Due to their low complexity and effective feature extraction capabilities, CNNs have been widely employed for SISR tasks. SRCNN [9] was a pioneering work that introduced deep convolutional neural network (CNN) architectures specifically designed for single image super-resolution (SISR). SRCNN features a simple three-layer structure consisting of patch extraction, nonlinear mapping, and reconstruction, and it demonstrated competitive performance against traditional non-deep learning methods, sparking the development of numerous lightweight CNN-based approaches in the SISR domain. Inspired by dictionary-learning models [15, 17, 30], LAPAR [20] learns linear coefficients associated with pre-defined basic upsampling kernels to create an optimal pixel-specific kernel, leading to superior super-resolution outcomes. LatticeNet [26] introduces a parameter-efficient convolutional lattice block to extract hierarchical contextual features. + +Transformer-based Methods. On the other hand, Transformer-based techniques [4-6, 8, 21, 35, 37-39] have achieved impressive results in SISR but are still hindered by their high computational complexity. The cost of modeling non-local self-similarity scales quadratically with the size of the image. To address these efficiency concerns, many window-based Transformer frameworks have been developed, inspired by the success of the Swin Transformer [24]. For instance, SwinIR [23] introduces a residual window-based transformer block (RSTB) for general image restora + +![](images/ea5df24049ab98011ea6a26ff048f3347d51a9f973e5e056c5eaa04bf77deab7.jpg) +Figure 3. Illustration of our proposed TSP-Mamba, including a shallow feature extraction module ('Head'), a deep feature extraction module and a high-resolution reconstruction module ('Tail'). The key point is that the TSP-LSTM, employed for local path-adaptive aggregation, facilitates subsequent global scanning of 2D-SSM. + +tion. DLGSA-1 [22] enhances token aggregation by using a global sparse attention mechanism. Omni-SR [33] presents spatial-channel attention for lightweight SISR, achieving better performance and efficiency trade-off. + +Mamba-based Methods. Recently, Mamba-based frameworks [13] have demonstrated impressive capabilities in modeling long-sequence dependencies while maintaining linear computational complexity. At the heart of Mamba is a structured state space model that expands the effective receptive field, allowing the framework to capture long-range dependencies more efficiently than traditional methods like CNNs and Transformers. This capacity to model global dependencies with linear complexity has positioned Mamba as a compelling alternative, particularly in resource-constrained environments. Building on Mamba's success, numerous follow-up studies [2, 11, 12, 14, 32] have explored the potential of designing different Mamba-based frameworks for various vision tasks. For example, MambaIR [14] offers a simple yet effective Mamba-based baseline for image restoration, while RetinexMamba [2] leverages Mamba's long-range dependency modeling for low-light image enhancement. Although these approaches integrate Mamba's global sequential reasoning (1D) into vision problems, the fixed scan pipeline often overlooks the locality and content-adaptive nature of 2D images. In this paper, we aim to develop an efficient and content-adaptive local propagation and aggregation method to improve the performance of Mamba-based frameworks, particularly for resource-constrained SISR. + +# 3. Methods + +# 3.1. Preliminaries: Mamba + +We begin by presenting a concise overview of the Mamba pipeline, followed by a discussion of the limitations asso + +ciated with the rigid 2D-to-1D scan scheme for low-level vision tasks. Selective structured state space models (SSMs) convert a 1D sequence $x(t)\in \mathbb{R}$ to the response $y(t)\in \mathbb{R}$ through a latent state $h(t)\in \mathbb{R}^N$ . This process is expressed: + +$$ +h ^ {\prime} (t) = \mathbf {A} h (t) + \mathbf {B} x (t), +$$ + +$$ +y (t) = \mathbf {C} h (t) + \mathbf {D} x (t), \tag {1} +$$ + +where $\mathbf{A} \in \mathbb{R}^{N \times N}$ , $\mathbf{B} \in \mathbb{R}^{N \times 1}$ , $\mathbf{C} \in \mathbb{R}^{1 \times N}$ and $D \in \mathbb{R}$ are time-varying matrices and $N$ is the state number. In practical applications, the continuous evolution system can be discretized using a zero-order hold assumption. Accordingly, Mamba converts the continuous matrices $\mathbf{A}$ , $\mathbf{B}$ to their discrete forms $\bar{\mathbf{A}}$ , $\bar{\mathbf{B}}$ with a timescale parameter $\Delta \in \mathbb{R}$ : + +$$ +\bar {\mathbf {A}} = \exp (\Delta \mathbf {A}), +$$ + +$$ +\bar {\mathbf {B}} = (\Delta \mathbf {A}) ^ {- 1} (e ^ {\mathbf {A}} - \mathbf {I}) \cdot \Delta \mathbf {B}. \tag {2} +$$ + +We then obtain a discretized formulation: + +$$ +h _ {k} = \bar {\mathbf {A}} h _ {k - 1} + \bar {\mathbf {B}} x _ {k}, \quad y _ {k} = \mathbf {C} h _ {k} + \mathbf {D} x _ {k}. \tag {3} +$$ + +As we can see, since the overall calculation is a linear combination of the accumulated sum, it can be efficiently implemented as a global 1D convolution: + +$$ +\bar {\mathbf {K}} = \left(\mathbf {C} \bar {\mathbf {B}}, \mathbf {C} \bar {\mathbf {A}} \bar {\mathbf {B}}, \dots , \mathbf {C} \bar {\mathbf {A}} ^ {S - 1} \bar {\mathbf {B}}\right), \mathbf {y} = \mathbf {x} * \bar {\mathbf {K}}, \tag {4} +$$ + +where $\otimes$ is a convolution operation and the $\bar{\mathbf{K}}\in \mathbb{R}^S$ refers to a convolution kernel. $S$ is the sequence length. + +Discussion. In NLP tasks, it is natural to employ state space models (SSMs) to process 1D sequences for global information reasoning at a linear computational cost. However, in 2D image scenarios, the neighborhoods often face challenges such as disconnection or long-range propagation to fit within the SSM framework. Although some recent studies have investigated alternative 1D scan paths to supplement + +information and improve restoration capability, capturing multiple local neighbors in a single 1D scan remains difficult, which can result in inefficient local aggregation for Mamba-based frameworks. + +# 3.2. Preliminaries: TSP + +Given $n$ cities and their pairwise distances, the goal is to determine the shortest route that visits each city exactly once. Mathematically, TSP is formulated to find a minimum-cost Hamiltonian cycle in a weighted complete graph, where nodes represent cities and edges denote travel costs (e.g., distance or time). Inspired by TSP's optimal traversal property, we propose a TSP-style local scan scheme to complement the global modeling of vision Mambas in this work. + +# 3.3. Framework Overview + +Fig. 3 illustrates the framework of our proposed TSP-Mamba. Following MambaIR, a recent SOTA baseline [14], TSP-Mamba comprises three main modules: (1) a shallow feature extraction module that transforms the RGB input into a feature space; (2) a deep feature extraction module consisting of multiple TSP-RSSGs (TSP-based Residual State-Space Groups). Each T-RSSG consists of several T-RSSBs (TSP-based Residual State-Space Block). In TSP-RSSB, the key component is the T-VSSM (TSP-based Vision State-Space Module), where we first use a TSP-LSTM layer to perform local path-adaptive aggregation and then pass the local-aggregation information into the 2D-SSM (2D Selective Scan Module in MambaIR[14]); and (3) a high-resolution reconstruction module that produces the final results. + +In the following section, we mainly focus on the explanation of our proposed TSP-LSTM, including the scanning patch construction for optimal local propagation (Sec.3.4.1), simplified TSP (Sec.3.4.2) and efficient implementation (Sec.3.4.3). + +# 3.4. TSP-LSTM + +# 3.4.1. Scanning Path Construction + +Given an image feature patch $\mathbf{P} \in \mathbb{R}^{s \times s \times C}$ , where $s$ is the patch size and $C$ is the number of channels, our objective is to establish a locally adaptive aggregation path that minimizes the information propagation costs. + +Definition. We formulate the scan path as a Traveling Salesman Problem (TSP) problem: let $V = \{1,2,\dots n\}$ be the set of pixels (cities) in image patch $P$ and $E = \{e_{i,j}(i,j):i,j\in V\}$ and $D_{e_{i,j}}$ be the cost measure associated with edge $e_{i,j}\in E$ . Our goal is to find a minimal scan path such that each city (pixel) is visited exactly once. + +In practice, we use the negative Cosine similarity as the cost measure of an edge $e_{i,j}$ in the set $E$ : + +$$ +D _ {e _ {i j}} = 1. 0 - \operatorname {C o s i n e} \left(F _ {i}, F _ {j}\right), \tag {5} +$$ + +
Size (s)TSPOurs
24!2
39!8
416!52
525!824
+ +![](images/adf2335ff54d627c8ad948b1e4b7b1f96624f5186315aef740dfcdfc39cea06a.jpg) +Figure 4. The numbers of scan path candidates of TSP and our simplified TSP scheme under different patch size $s$ . +Figure 5. Speedup ratio curve under different patch sizes. Our effective measure hashing strategy attains notable speedup $(\log_{10})$ for efficient TSP searching. + +where $F_{i}, F_{j}$ are the corresponding feature vector of the pixel $i, j$ . More discussion is shown in Tab. 3 (A). + +# 3.4.2. Simplified TSP + +The Traveling Salesman Problem (TSP) is a classic NP-hard problem. With a large $n$ , searching for an optimal scan path becomes time-consuming ( $O(n!)$ ). To simplify the problem, we adopt the following strategies in this work: + +(1) We limit the overall node count $n$ to the pixels within a local image patch, thereby reducing overall complexity; +(2) We designate the first city '1' (the top-left pixel of the given patch) as the starting node; +(3) At each intermediate step, we limit next-step candidates to the four neighboring 2D locations1 to reduce the search space further. + +Based on these three simplifications mentioned, it largely reduced the number of total candidate scan paths (in Fig. 4). Thus, we can initially employ the backtracking algorithm [10] to identify all possible scan paths: + +$$ +\begin{array}{l} T ^ {\{1, 2, \dots , l, \dots , L \}} = \text {B a c k T r a c k i n g} (1, n), \\ \text {w h e r e} T ^ {l} = \left\{i _ {1} ^ {l}, i _ {2} ^ {l}, \dots , i _ {n} ^ {l} \right\}, i _ {1} ^ {l} = 1, \\ \end{array} +$$ + +where $l$ -th path $T^l$ is represented by the ordered pixel indices $(\{i_1^l, i_2^l, \dots, i_n^l\})$ and $L$ is the count of all scan candidates. We calculate the cost of the $l$ -th scan path $T^l$ : + +$$ +D ^ {l} = \sum_ {k = 1} ^ {n - 1} D _ {e _ {i _ {k} ^ {l}, i _ {k + 1} ^ {l}}} \tag {7} +$$ + +Finally, we choose the optimal scan path $T^{l^*}$ with the lowest measure cost: + +$$ +l ^ {*} = \underset {l} {\operatorname {a r g m i n}} D ^ {l} \tag {8} +$$ + +# 3.4.3. Effective Implementation + +As indicated in Eq. 7, we have to calculate the $n - 1$ pairwise distances for each scan path and sum them up. We have two observations for effective path cost computation: + +1. Symmetry of Distance Measurement. As indicated in Eq. 5, for edges $e_{i,j}$ , $e_{j,i}$ , we have $D_{e_{i,j}} = D_{e_{j,i}}$ . +2. Highly overlapping edges for different scan paths. Different scene paths may share lots of the same edges. + +Based on this fact, we propose to design a hash-based technique to store the unique edges in all scan paths into an edge list $\{e_{k_0}, e_{k_1}, \dots, e_{k_m}\}$ ( $m$ is the length of the unique edge list) to avoid repeated distance calculation. Accordingly, we can effectively compute all unique edge distances using Eq. 5 at one time and obtain the associated edge distance list $\{D_{e_{k_0}}, D_{e_{k_1}}, \dots, D_{e_{k_m}}\}$ . Finally, we can obtain the path propagation cost in Eq. 7 by accumulating all its edges. + +The speedup ratio between the direct per-path cost computation and our efficiency implementation is shown in Fig. 5, we can see our effective hash-based calculation attains a significant speedup effect. Additionally, it becomes more effective with an increased patch size. + +Summary. In this part, we propose a simplified TSP formulation to effectively search the shortest path for a given 2D image patch, in which all neighborhoods are fused in a relatively short-term propagation. Therefore, it is capable of effectively integrating local information to facilitate the later global SSM-style propagation. + +# 3.4.4. TSP-LSTM + +Bidirectional Propagation. After the scan path optimization, we perform a bidirectional LSTM for mutual information propagation. Especially, given a 2D image patch $\mathbf{P}$ , we use the above simplified TSP scheme to get the optimal scan chain $T^{l^{*}} = \{i_{1} \rightleftharpoons i_{2}, \dots, \rightleftharpoons i_{n}\}^{2}$ , where “ $\rightleftharpoons$ ” means we employ the bidirectional propagation scheme (as shown in Fig. 2) (b) for effective and balanced aggregation. + +For implementation, we flip the reordered sequence $\{i_1,i_2,\dots ,i_n\}$ to obtain the inverse sequence and concatenate them for bidirectional aggregation. + +Distance Embedding. Within an image patch, discontinuities may need to be addressed in the chain-style propagation. Specifically, the aggregation weight of the previously accumulated information should positively correlate with the distance between the previous and current nodes (pixels). To account for this, inspired by fast global image smoothing (FGS [29]), we incorporate distance embedding features into the sequences to explicitly model this effect. We calculate the pairwise distances as shown in Eq. 7 and embed them (as a one-channel feature) into the propagation sequence feature. + +Lastly, the propagated feature $\mathbf{x}$ is passed through the 2DSSM module (in Eq. 4) for global Mamba-style aggregation. + +# 4. Experiments + +# 4.1. Settings + +Implementation Details. Our TSP-Mamba model is implemented in PyTorch and integrates several common data augmentation techniques, such as random cropping, vertical/horizontal flipping, and rotation. For training, we use the Adam optimizer [18] with cosine annealing [25], starting with an initial learning rate of $4e - 4$ . The batch size is set to 32, with input images sized at $64 \times 64$ . Training is performed over 600K iterations using four NVIDIA RTX 3090 GPUs. + +Scalable Model Configuration. In general, we train multiple TSP-Mamba models with different configurations. TSP-Mamba is based on the MambaIR baseline, utilizing 4 T-RSSGs, each containing 6 T-RSSBs, with a channel size of 60, denoted as “[6,6,6,6]@C60”. Additionally, we investigate the scalability of TSP-Mamba to better understand the trade-off between performance and model complexity. Specially, we use “[2,4,2,6]@C40” $^{4}$ and “[4,6,4,6]@C48” $^{5}$ to train another two smaller SISR models under $\times 2$ SISR setting. Training these TSP-Mamba models usually takes around 3-6 days. + +Benchmark Datasets. In line with previous studies [7, 20, 23], we use the DIV2K [1] image dataset for training. We then perform extensive evaluations on several widely-used SISR benchmarks, including Set5 [3], Set14 [36], BSD100 [27], Urban100 [16], and Manga109 [28]. Our quantitative analysis is based on PSNR and SSIM metrics, and following standard practice, we report results specifically for the Y channel in the YCbCr color space. + +# 4.2. Comparison with SOTA Methods + +We perform extensive comparisons with a wide range of lightweight SISR models: LAPAR-A (NeurIPS20) [20], LatticeNet (ECCV20) [26], RLFN (CVPRW22) [19], SwinIR-light [23], DLGSA-1 (ICCV23) [22], Onmi-SR (CVPR23) [33], SRFormer (ICCV23) [39], ATDS (CVPR24) [37], MambaIR (ECCV24) [14], HiTSR [38] (ECCV24). + +Quantitative Comparison. Tab. 1 shows the quantitative evaluation results, highlighting that our proposed TSP-Mamba consistently outperforms existing methods across all benchmarks. In particular, our model achieves significant improvements over MambaIR, with gains of up to $0.71\mathrm{dB}$ , $0.58\mathrm{dB}$ , and $0.69\mathrm{dB}$ on the Manga109 dataset for the $\times 2$ , $\times 3$ , and $\times 4$ scales, respectively, while maintaining similar parameter and FLOP requirements. + +Moreover, TSP-Mamba-Small beats all leading lightweight SISR methods with smaller model sizes and + +Table 1. Quantitative comparison with SOTA lightweight SISR methods on multiple benchmark datasets. The best and second-best results on the default training setting (DIV2K) are highlighted in red and blue, respectively. We follow most lightweight SISR methods and calculate all Flops with an output size of ${1280} \times {720}$ . + +
MethodScaleParameters (K)Flops (G)Set5 PSNR / SSIMSet14 PSNR / SSIMBSD100 PSNR / SSIMUrban100 PSNR / SSIMManga109 PSNR / SSIM
LAPAR-A54817138.01 / 0.960533.62 / 0.918332.19 / 0.899932.10 / 0.928338.67 / 9772
LatticeNet756169.538.15 / 0.961033.78 / 0.919332.25 / 0.900532.43 / 0.9302- / -
SwinIR-light910244.438.14 / 0.961133.86 / 0.920632.31 / 0.901232.76 / 0.934039.12 / 0.9783
DLGSA-light745170.038.20 / 0.961233.89 / 0.920332.30 / 0.901232.94 / 0.935539.29 / 0.9780
Omni-SR772194.538.22 / 0.961333.98 / 0.921032.36 / 0.902033.05 / 0.936339.28 / 0.9784
SRFormer ×285323638.23 / 0.961333.94 / 0.920932.36 / 0.901932.91 / 0.935339.28 / 0.9785
HiT-SR847226.538.26 / 0.961534.01 / 0.921432.37 / 0.902333.13 / 0.937239.47 / 0.9787
ATD-S75334938.29 / 0.961634.10 / 0.921732.39 / 0.902333.27 / 0.937539.52 / 0.9789
MambaIR136319838.16 / 0.961034.00 / 0.921232.34 / 0.901732.92 / 0.935639.31 / 0.9779
TSP-Mamba130519938.46 / 0.964834.42 / 0.924332.52 / 0.904233.57 / 0.940540.02 / 0.9808
MambaIR-Small80611138.09 / 0.960233.85 / 0.920132.31 / 0.901532.58 / 0.932139.17 / 0.9787
TSP-Mamba-Small76911238.22 / 0.964034.19 / 0.922732.43 / 0.903033.18 / 0.937439.49 / 0.9791
MambaIR-Tiny4435937.99 / 0.960233.72 / 0.910932.21 / 0.900232.22 / 0.929138.87 / 0.9772
TSP-Mamba-Tiny4185938.04 / 0.963133.90 / 0.920532.29 / 0.901032.55 / 0.931739.12 / 0.9784
LAPAR-A54411434.36 / 0.926730.34 / 0.842129.11 / 0.805428.15 / 0.852333.51 / 9441
LatticeNet76576.334.53 / 0.928130.39 / 0.842429.15 / 0.805928.33 / 0.8538- / -
SwinIR-light918110.834.62 / 0.928930.54 / 0.846329.20 / 0.808228.66 / 0.862433.98 / 0.9478
DLGSA-light75275.434.70 / 0.929530.58 / 0.846529.24 / 0.808928.83 / 0.865334.16 / 0.9483
Omni-SR78088.434.70 / 0.929430.57 / 0.846929.28 / 0.809428.84 / 0.865634.22 / 0.9487
SRFormer ×386110534.67 / 0.929630.57 / 0.846929.26 / 0.809928.81 / 0.865534.19 / 0.9489
HiT-SR85510234.75 / 0.930030.61 / 0.847529.29 / 0.810628.99 / 0.868734.53 / 0.9502
MambaIR13718934.72 / 0.929630.63 / 0.847529.29 / 0.809929.00 / 0.868934.39 / 0.9495
TSP-Mamba13138934.80 / 0.933230.93 / 0.851029.41 / 0.813929.43 / 0.874434.97 / 0.9522
MambaIR-Small8125034.65 / 0.931930.59 / 0.847129.26 / 0.810228.84 / 0.864434.02 / 0.9472
TSP-Mamba-Small7765034.65 / 0.932230.66 / 0.847829.30 / 0.811428.94 / 0.868034.41 / 0.9495
MambaIR-Tiny4482734.36 / 0.922330.26 / 0.842829.12 / 0.806328.24 / 0.853233.74 / 0.9452
TSP-Mamba-Tiny4232734.49 / 0.930930.46 / 0.845829.22 / 0.809428.70 / 0.862334.16 / 0.9483
LAPAR-A6599432.15 / 0.894428.61 / 0.781827.61 / 0.736626.14 / 0.787130.42 / 0.9074
LatticeNet77743.632.30 / 0.896228.68 / 0.783027.62 / 0.736726.25 / 0.7873- / -
SwinIR-light93063.632.44 / 0.897628.77 / 0.785827.69 / 0.740626.47 / 0.798030.92 / 0.9151
DLGSA-light76142.532.54 / 0.899328.84 / 0.787127.73 / 0.741526.66 / 0.803331.13 / 0.9161
Omni-SR79250.932.49 / 0.898828.78 / 0.785927.71 / 0.741526.64 / 0.801831.02 / 0.9151
SRFormer ×48736332.51 / 0.898828.82 / 0.787227.73 / 0.742226.67 / 0.803231.17 / 0.9165
HiT-SR8665832.55 / 0.899928.87 / 0.788027.75 / 0.743226.80 / 0.806931.26 / 0.9171
ATD-S7698732.63 / 0.899828.89 / 0.788627.79 / 0.744026.97 / 0.810731.48 / 0.9198
MambaIR13835132.51 / 0.899328.85 / 0.787627.75 / 0.742326.75 / 0.805131.26 / 0.9175
TSP-Mamba13245332.71 / 0.902429.05 / 0.792527.91 / 0.747627.29 / 0.817431.95 / 0.9235
MambaIR-Small8212932.28 / 0.900128.70 / 0.786327.72 / 0.741226.56 / 0.798431.12 / 0.9156
TSP-Mamba-Small7853032.48 / 0.902428.90 / 0.789327.79 / 0.744226.93 / 0.808731.42 / 0.9184
MambaIR-Tiny4561532.22 / 0.899128.63 / 0.784727.68 / 0.739826.43 / 0.794030.86 / 0.9124
TSP-Mamba-Tiny4311632.29 / 0.899728.79 / 0.786027.72 / 0.740526.61 / 0.799231.04 / 0.9140
MambaIR (×16)1772523.24 / 0.641922.29 / 0.545422.94 / 0.521820.46 / 0.514821.15 / 0.6628
TSP-Mamba (×16)1714523.33 / 0.651522.38 / 0.545222.81 / 0.516920.53 / 0.518421.20 / 0.6700
+ +less computational complexity. For instance, it surpasses MambaIR by 0.26dB on Urban $100 \times 2$ , while only requiring $59\%$ parameters and $57\%$ Flops. + +Inference efficiency. Tab. 2 presents a detailed comparison of inference costs across different models, including CNN-based, Transformer-based, and Mamba-based frameworks. Overall, the CNN-based LatticeNet achieves the highest inference efficiency. Compared to other transformer-based and Mamba-based models, our proposed TSP-Mamba achieves competitive inference latency. + +Qualitative Comparison. Fig. 6 showcases several visual + +examples. It is clear that TSP-Mamba effectively reconstructs accurate image textures with fewer super-resolved artifacts. In contrast, other approaches struggle to restore clear image patterns. + +# 4.3. Ablation Studies + +To validate our design, we conduct extensive experiments, reporting quantitative results on the Urban100 and Manga109 benchmarks at the $\times 4$ setting. We primarily adopt the TSP-Mamba-Small architecture for fast analysis. + +![](images/d6aaa5f94cacb147a41f5bc9cd65db9c36c5601911e861d706968ae19c8de3bc.jpg) +Figure 6. Qualitative results of different lightweight SISR methods under $\times 4$ setting. Our predictions present more detailed structures. + +Table 2. Inference cost comparison of different lightweight SISR models under $\times 2$ setting. Inference time is tested with an output resolution of ${256} \times {256}$ using an NVIDIA 3090 GPU. The top-3 results are highlighted with red,blue,underline. + +
MetricsLatticeNetDLGSA-lightSwinIR-lightSRFormerMambaIR / Small / TinyTSP-Mamba / Small / Tiny
Time (ms) ↓973416060 / 50 / 3872 / 58 / 39
Memory (MB) ↓146616245441572686 / 536 / 528732 / 614 / 582
+ +Different Measure Cost in TSP-LSTM. Apart from Eq. 5, we try the widely-used L1/L2 distance metrics for scan path searching. Tab. 3 (A) concludes that our used negative Cosine distance performs better than the L1/L2 distances. + +TSP-LSTM vs. Uniform-LSTM. TSP-LSTM aims to perform an adaptive scan path by constructing a minimal propagation path. To assess the impact of our adaptive scan path, we also train two uniform scan models (vertical, and horizontal). We compare the proposed TSP-LSTM with the other models (UV-LSTM, HV-LSTM). As shown in Tab. 3 (B), we can observe that TSP-LSTM significantly outperforms the uniform scan models, highlighting the importance of adaptive local propagation. + +Different Patch Sizes. As mentioned in Sec. 3.4, we set the patch size to $4 \times 4$ considering the local receptive field and computational costs. When the patch size is increased, it becomes prolonged to construct the corresponding TSP path. We investigate the influence of different patch sizes. Tab. 3 (C) suggests the reconstructed performance is positively related to the patch size but with larger latency. + +Finer Design in TSP-LSTM. Lastly, we evaluate the impact of bidirectional propagation of TSP-LSTM. We train another model with only a forward/backward scan pipeline to do this. In Tab. 3 (D), it is clear that our bidirectional propagation scheme is more effective than its counterparts with only a single-direction scan strategy. + +Different Aggregation Layers. In this paper, we present a TSP-LSTM block to get locally fused information. To verify the effectiveness of TSP-LSTM, we replace it with other designs for local compensation, such as convolution, RNN, and Mamba. The results are shown in Tab. 3 (E) below, which suggests that our TSP attains the best performance and second-best efficiency compared with other designs. + +# 4.4. Extensions + +Classic SISR. Though TSP-Mamba is trained for lightweight Mamba-based architecture, i.e., MambaIR, we additionally explore the potential improvement for the classic SISR setting. To do this, we follow MambaIR-L(large) to enlarge the channel size to 180 and use 6 TSP-RSSG blocks (each one contains 6 basic TSP-SSD layers) to obtain the TSP-Mamba-L model and train this classic SISR model using the same DF2K training dataset. The quantitative results are illustrated in Tab. 4. We can see that TSP-Mamba-L consistently outperforms other classic SISR methods significantly. For instance, it achieves 0.32dB, 0.37dB, and 0.39dB improvements over MambaIR. + +Depth Map SR. We also compare the TSP-Mamba model with our baseline model (MambaIR) on the depth SR task for generalization analysis. To do this, without training on any depth images, we directly assess the two models on the NYU-V2 [31] depth benchmark $^6$ under $\times 4$ and $\times 16$ settings. + +Table 3. Ablation Studies. A: The effects of different distance measures, with our default choice being the negative cosine distance as indicated in Eq. 5. B: Quantitative evaluation of uniform scanning strategies (vertical and horizontal), abbreviated as 'UV' and 'HV', compared to our TSP-based scan. C: The influence of different local scan patch sizes (P2-P5 indicates local patch sizes ranging from 2 to 5). D: Impacts of distance embeddings and bidirectional propagation. M1 drops both of the two designs. M2 only adopts distance embedding. M3 only uses the bidirectional propagation. M4 employs both of the two designs. E: The effects of various local aggregation modules (M5-M8 employ Conv1D, RNN, Mamba, and LSTM, respectively). + +
AnalysisA.B.C.D.E.
L1L2Neg. Cos.UVUHTSPP2P3P4P5M1M2M3M4M5M6M7M8
Urban10026.5526.5426.6126.3126.3826.6126.3726.4326.6126.6126.1626.3826.2526.6126.4926.4426.4726.61
0.74030.79820.79920.79070.79360.79920.79330.79430.79920.79940.78810.79390.79070.79920.79580.79500.79560.7992
Manga10931.0631.0731.0430.7830.8631.0430.7530.9131.0431.0230.5030.8030.6531.0430.9630.9530.9731.04
0.91400.91440.91400.91040.91260.91400.91130.91310.91400.91380.90650.91030.90840.91400.91360.91320.91340.9140
Param.(K)431431431431431431431431431431430431430431440436428431
FLOPs (G)16.216.216.215.415.416.215.215.216.220.314.915.116.116.215.515.716.916.2
+ +Table 4. Quantitative comparison with recent SOTA Classic SR models. All the models are trained on the DF2K dataset for fair comparison. We report the averaged PSNR/SSIM values over the five benchmarks (Set5, Set14, BSD100, Urban100, Manga109). + +
SettingsMetricsSwinIRSRFormer-LHATMambaIR-LTSP-Mamba-L
×2PSNR35.8335.9436.1736.0536.37
SSIM0.94280.94350.94460.94380.9457
Param.11.8M10.4M20.6M20.4M20.6M
×3PSNR32.0532.1532.2932.1932.56
SSIM0.88720.88850.88810.89000.8918
Param.11.8M10.6M20.8M20.6M20.8M
×4PSNR29.8929.9730.1430.0430.43
SSIM0.83990.84160.84410.84140.8470
Param.11.8M10.5M20.8M20.6M20.8M
+ +We use the PSNR, SSIM, and RMSE (the root-mean-square error) metrics for quantitative evaluation. The quantitative results are shown in the left Fig. 7. We can see that TSP-Mamba consistently outperforms its baseline model on all the objective metrics. For instance, TSP-Mamba exhibits superior performance compared to MambaIR with a PSNR improvement of 0.96dB (0.92dB) for $\times 4$ ( $\times 16$ ) configurations. The visual examples in Fig. 7 illustrate that TSP-Mamba generates clearer structures, leading to higher accuracy when compared to the baseline model. + +Table 5. The quantitative results on NYU-V2 [31] $(\times 4,\times 16)$ + +
SettingsMetricsMambaIRTSP-Mamba
×4 (×16)PSNR47.25(37.03)48.21(37.95)
SSIM0.994(0.967)0.995(0.972)
RMSE2.339(7.895)2.145(7.142)
+ +Limitations. Although our proposed TSP-Mamba demonstrates promising and significant improvements across several single-image super-resolution tasks, it remains challenging to reliably obtain accurate distance measurements for degraded images, particularly in the presence of high-frequency degradations. Noise components can adversely affect the + +![](images/af97791f8b819af208afca2962c69b89b403f059287142cc566fc762f256ae35.jpg) +PSNR/SSIM/RMSE + +![](images/6e4c214db5bd53d74e17d402c067a02865771591b0e9ddb75a8a62cf4131dfe5.jpg) +37.45/0.9698/0.0330 + +![](images/229faf48ef40d423b13934a6c600e935a67fc9937affb2b77117907d84867f6a.jpg) +38.79/0.9750/0.0282 + +![](images/d5b7734f1e4a9fff80a96060cda4db81977001cb9f9cdc95e26365d21138db7b.jpg) +PSNR/SSIM/RMSE + +![](images/85c542c92fe9da51ae09ec98329d9991ced9dae65d9775b22bb5f76cadcc7312.jpg) +34.80/0.9672/0.0673 + +![](images/6a9a0ccec9c600fb1fb4290670466b06da9717a147006ff4d8bf9fac59abb18e.jpg) +36.45/0.9742/0.0555 +Figure 7. Qualitative comparison between our baseline and the proposed TSP-Mamba for unseen depth-SR maps. We also show normalized error maps in the left corner for better visual comparison + +calculation of pair-wise distances between degraded pixels and the optimal scan path searching. In future work, we plan to address the issue of robust distance calculation under various image degradations, such as pyramid representation to build robust 1D TSP scan paths hierarchically. + +# 5. Conclusion + +In this work, we introduce a novel locally adaptive propagation method for lightweight SISR. To this end, we develop an efficient scan path search strategy to enhance local information within the globally Mamba-based architecture. Extensive experiments validate the effectiveness of our approach, with TSP-Mamba achieving state-of-the-art performance across multiple SISR benchmarks. Furthermore, TSP-Mamba demonstrates superior generalization on unseen data (such as depth-map SR), and scalability across both ultralightweight and larger model configurations. + +Acknowledgments. This work is partially supported by Shenzhen Science and Technology Program KQTD20210811090149095 and also the Pearl River Talent Recruitment Program 2019QN01X226. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5 +[2] Jiesong Bai, Yuhao Yin, and Qiyuan He. Retinexmamba: Retinex-based mamba for low-light image enhancement. arXiv preprint arXiv:2405.03349, 2024. 3 +[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5 +[4] Ke Chen, Liangyan Li, Huan Liu, Yunzhe Li, Congling Tang, and Jun Chen. Swinfsr: Stereo image super-resolution using swinir and frequency domain knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764-1774, 2023. 2 +[5] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. +[6] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12312-12321, 2023. 2 +[7] Haram Choi, Jeongmin Lee, and Jihoon Yang. N-gram in swim transformers for efficient lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2071–2081, 2023. 5 +[8] Tao Dai, Jianping Wang, Hang Guo, Jinmin Li, Jinbao Wang, and Zexuan Zhu. Freqformer: frequency-aware transformer for lightweight image super-resolution. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 731-739, 2024. 2 +[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2 +[10] Eugene C Freuder and Alan K Mackworth. Constraint satisfaction: An emerging paradigm. In Foundations of artificial intelligence, pages 13-27. Elsevier, 2006. 4 +[11] Guanyiman Fu, Fengchao Xiong, Jianfeng Lu, and Jun Zhou. Ssumamba: Spatial-spectral selective state space model for hyperspectral image denoising. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3 +[12] Hu Gao, Bowen Ma, Ying Zhang, Jingfan Yang, Jing Yang, and Depeng Dang. Learning enriched features via selective state spaces model for efficient image deblurring. In ACM Multimedia 2024. 3 +[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3 +[14] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for + +image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. 1, 3, 4, 5 +[15] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 2 +[16] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5 +[17] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3224-3232, 2018. 2 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[19] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 766-776, 2022. 5 +[20] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. Lapar: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. Advances in Neural Information Processing Systems, 33:20343-20355, 2020. 2, 5 +[21] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2 +[22] Xiang Li, Jiangxin Dong, Jinhui Tang, and Jinshan Pan. Dlgstanet: Lightweight dynamic local and global self-attention networks for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12792–12801, 2023. 3, 5 +[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 5 +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2 +[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5 +[26] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image superresolution with lattice block. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 272-289. Springer, 2020. 2, 5 + +[27] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 1, 5 +[28] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multi-media Tools and Applications, 76:21811-21838, 2017. 5 +[29] Dongbo Min, Sunghwan Choi, Jiangbo Lu, Bumsub Ham, Kwanghoon Sohn, and Minh N Do. Fast global image smoothing based on weighted least squares. IEEE Transactions on Image Processing, 23(12):5638-5653, 2014. 5 +[30] Yaniv Romano, John Isidoro, and Peyman Milanfar. Raisr: Rapid and accurate image super resolution. IEEE Transactions on Computational Imaging, 3(1):110-125, 2016. 2 +[31] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 7, 8 +[32] Yujin Tang, Peijie Dong, Zhenheng Tang, Xiaowen Chu, and Junwei Liang. Vmrnn: Integrating vision mamba and lstm for efficient and accurate spatiotemporal forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5663-5673, 2024. 3 +[33] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Jinfan Liu. Omni aggregation networks for lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22378-22387, 2023. 3, 5 +[34] Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. arXiv preprint arXiv:2407.21773, 2024. 1 +[35] Jinsu Yoo, Taehoon Kim, Sihaeng Lee, Seung Hwan Kim, Honglak Lee, and Tae Hyun Kim. Enriched cnn-transformer feature aggregation networks for super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4956-4965, 2023. 2 +[36] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5 +[37] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2024. 2, 5 +[38] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. In European Conference on Computer Vision, pages 483-500. Springer, 2024. 5 + +[39] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12780–12791, 2023. 2, 5 \ No newline at end of file diff --git a/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/images.zip b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ef63dca325f26ee8e32b91fa29c97053561b0496 --- /dev/null +++ b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c210a5a6fd42d02f6d33f295252c0831d50b3808205aad40ec13a5008235c18 +size 959262 diff --git a/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/layout.json b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fd7749789deda98eaeae0689ffc37f0c18dc10b7 --- /dev/null +++ b/2025/TSP-Mamba_ The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond/layout.json @@ -0,0 +1,9027 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 102, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 102, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 102, + 501, + 140 + ], + "type": "text", + "content": "TSP-Mamba: The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 160, + 409, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 160, + 409, + 175 + ], + "spans": [ + { + "bbox": [ + 205, + 160, + 409, + 175 + ], + "type": "text", + "content": "Kun Zhou1 Xinyu Lin2 Jiangbo Lu1*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 175, + 426, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 175, + 426, + 190 + ], + "spans": [ + { + "bbox": [ + 184, + 175, + 426, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 175, + 426, + 190 + ], + "type": "text", + "content": "SmartMore Corporation " + }, + { + "bbox": [ + 184, + 175, + 426, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 184, + 175, + 426, + 190 + ], + "type": "text", + "content": "SSE, CUHK-Shenzhen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 191, + 490, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 191, + 490, + 202 + ], + "spans": [ + { + "bbox": [ + 120, + 191, + 490, + 202 + ], + "type": "text", + "content": "zhoukun303808@gmail.com, linxinyu0327@gmail.com, jiangbo.lu@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 230, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 230, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 230, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 255, + 297, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 297, + 531 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 297, + 531 + ], + "type": "text", + "content": "Recently, Mamba-based frameworks have achieved substantial advancements across diverse computer vision and NLP tasks, particularly in their capacity for reasoning over long-range information with linear complexity. However, the fixed 2D-to-1D scanning pattern overlooks the local structures of an image, limiting its effectiveness in aggregating 2D spatial information. While stacking additional Mamba layers can partially address this issue, it increases parameter intensity and constrains real-time application. In this work, we reconsider the local optimal scanning path in Mamba, enhancing the rigid and uniform 1D scan through the local shortest path theory, thus creating a structure-aware Mamba suited for lightweight single-image super-resolution. Specifically, we draw inspiration from the Traveling Salesman Problem (TSP) to establish a local optimal scanning path for improved structural 2D information utilization. Here, local patch aggregation occurs in a content-adaptive manner with minimal propagation cost. TSP-Mamba demonstrates substantial improvements over existing Mamba-based and Transformer-based architectures. For example, TSP-Mamba surpasses MambaIR by up to 0.7dB in lightweight SISR, with comparable parameters and very slightly extra computational demands (1-2 GFlops for 720P images)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 544, + 135, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 135, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 135, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 564, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 297, + 696 + ], + "type": "text", + "content": "Deep learning techniques have advanced single-image superresolution (SISR) tasks considerably, enhancing image reconstruction quality. Convolutional Neural Networks (CNNs) have been central to this progress, excelling at capturing local, content-dependent image patterns. However, their local focus limits their capacity to model long-range dependencies. Transformers, by contrast, are highly effective in capturing global information through their self-attention mechanism, enabling them to model non-local dependencies. Yet, their quadratic computational complexity poses significant challenges in resource-constrained environments." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 535, + 378 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 535, + 378 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 535, + 378 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 535, + 378 + ], + "type": "image", + "image_path": "e262c5ed2eefaac09d41c89cb9e9e4b2baf6d109607d0b6a440bd20854e60d7e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 382, + 365, + 434 + ], + "blocks": [ + { + "bbox": [ + 318, + 382, + 365, + 434 + ], + "lines": [ + { + "bbox": [ + 318, + 382, + 365, + 434 + ], + "spans": [ + { + "bbox": [ + 318, + 382, + 365, + 434 + ], + "type": "image", + "image_path": "e3e4d1f00c614483fc8b1c4c5aacbe50a01ed17603d9c6eed075c0a8b13057c5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 444, + 555, + 522 + ], + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 522 + ], + "type": "text", + "content": "Figure 1. (1) Comparison between our proposed TSP-Mamba and SOTA SISR models on BSD100 [27] for " + }, + { + "bbox": [ + 313, + 444, + 555, + 522 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 444, + 555, + 522 + ], + "type": "text", + "content": " setting. TSP-Mamba series ('S'- 'Small', 'T'- 'Tiny') demonstrates highly competitive results. (2) Comparison of Effective Receptive Field for several SOTA methods. Darker area denotes larger receptive fields. Our proposed TSP-Mamba achieves the best global receptive field. Notably, larger circles indicate more model parameters (the first figure)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 372, + 382, + 419, + 435 + ], + "blocks": [ + { + "bbox": [ + 372, + 382, + 419, + 435 + ], + "lines": [ + { + "bbox": [ + 372, + 382, + 419, + 435 + ], + "spans": [ + { + "bbox": [ + 372, + 382, + 419, + 435 + ], + "type": "image", + "image_path": "8d49b319205b9334996d91f908dee7525683ef9ce8689d863218b0f02550dd55.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 428, + 381, + 474, + 435 + ], + "blocks": [ + { + "bbox": [ + 428, + 381, + 474, + 435 + ], + "lines": [ + { + "bbox": [ + 428, + 381, + 474, + 435 + ], + "spans": [ + { + "bbox": [ + 428, + 381, + 474, + 435 + ], + "type": "image", + "image_path": "705f6488904d86c70e2030088ab569888459dbcdaab91e03f9da02a27051e02f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 484, + 381, + 553, + 435 + ], + "blocks": [ + { + "bbox": [ + 484, + 381, + 553, + 435 + ], + "lines": [ + { + "bbox": [ + 484, + 381, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 484, + 381, + 553, + 435 + ], + "type": "image", + "image_path": "4d6e74f302e9322baa43167dd5d8cd9bbb6b1aa82412d485873697cbf5181e2a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": "The Mamba method has recently gained attention for its capacity to model global signals with linear computational complexity. Its fixed scanning approach is well-suited for efficient 1D sequence modeling, achieving strong computational performance. However, this approach overlooks 2D structures within images, leading to inefficient aggregation of 2D information. As illustrated in Fig. 2, conventional vision-Mambas typically rely on partial neighboring propagation and aggregation along a fixed 1D scan path, which can disrupt the locality of 2D image information—referred to as \"local pixel forgetting\" in [14] and \"destroying local correlations\" in [34]—resulting in sub-optimal outcomes. As shown in Fig. 2(a), the local information aggregation with an image patch as highlighted in the red box, is less effective in Mamba-based scan due to the long-range (also partial)" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28134" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 177, + 199 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 177, + 199 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 177, + 199 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 177, + 199 + ], + "type": "image", + "image_path": "2993a053834e6b728e4db49b991a7170eb369744dce0a2cd5f092fb877fac4c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 220, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 220, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. Illustration of vanilla Mamba and our proposed locally enhanced TSP-Mamba. For a highlighted " + }, + { + "bbox": [ + 54, + 220, + 555, + 266 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 220, + 555, + 266 + ], + "type": "text", + "content": " image patch (red box): (a) In Vanilla Mamba, only a limited set of neighboring pixels (e.g., '1, 2, 3, 4') can be aggregated for pixel '5' through a mix of long- and short-range propagation within a single global scan (\"Scan-1\"). (b) In TSP-Mamba, the shortest traversal path is constructed, allowing patch-adaptive propagation in a single local scan." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 183, + 71, + 335, + 199 + ], + "blocks": [ + { + "bbox": [ + 183, + 71, + 335, + 199 + ], + "lines": [ + { + "bbox": [ + 183, + 71, + 335, + 199 + ], + "spans": [ + { + "bbox": [ + 183, + 71, + 335, + 199 + ], + "type": "image", + "image_path": "cac057b1c8b8dd7a4979044468e19a9d2762c444cd2359c078615aeadc80b2a0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 201, + 306, + 211 + ], + "lines": [ + { + "bbox": [ + 212, + 201, + 306, + 211 + ], + "spans": [ + { + "bbox": [ + 212, + 201, + 306, + 211 + ], + "type": "text", + "content": "(a) 4 scan types in MambaIR" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 335, + 72, + 550, + 199 + ], + "blocks": [ + { + "bbox": [ + 335, + 72, + 550, + 199 + ], + "lines": [ + { + "bbox": [ + 335, + 72, + 550, + 199 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 550, + 199 + ], + "type": "image", + "image_path": "debd949819ca88cd310f0e27a873d8eabf6ee80cef3a4e6d8bf04dbd72174173.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 201, + 529, + 211 + ], + "lines": [ + { + "bbox": [ + 373, + 201, + 529, + 211 + ], + "spans": [ + { + "bbox": [ + 373, + 201, + 529, + 211 + ], + "type": "text", + "content": "(b) Shortest traversal path search in TSP-Mamba" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 273, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 296, + 332 + ], + "type": "text", + "content": "propagation within a single scan. A common solution to this problem is stacking multiple multidirectional Mamba layers to improve aggregation performance. However, this will substantially increase computational cost and model complexity, hindering practicality in resource-limited scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 335, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 335, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 335, + 296, + 514 + ], + "type": "text", + "content": "To address these limitations, we propose TSP-Mamba, which enhances the propagation and aggregation of local 2D information by searching for a locally adaptive, optimal propagation path for each image patch. Unlike the fixed (content-invariant) scan pipeline used in traditional Mamba, our method first identifies a patch-adaptive scan path with the lowest propagation cost, leveraging the shortest path principle. Here, each pixel in a local image patch is treated as a node in a graph, and we frame the problem as a simplified Traveling Salesman Problem (TSP) to find the minimum-cost traversal path, along which local information is transmitted and aggregated to enhance the use and extraction of local structures. Later on, the enhanced local information is effectively propagated into the global Mamba-style aggregation in a conventionally local-then-global manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 517, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 517, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 517, + 296, + 613 + ], + "type": "text", + "content": "To implement this adaptive propagation, we design a TSP-LSTM block that modulates the aggregation between the current node and the previously accumulated information based on the propagation cost (i.e., the distance between the current and previous nodes). As demonstrated in Fig. 2(b), TSP-Mamba obtains a local shortest scan path for aggregation, thus leading to improved performance at a comparable and even reduced model complexity (shown in Fig. 1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 615, + 208, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 615, + 208, + 627 + ], + "spans": [ + { + "bbox": [ + 67, + 615, + 208, + 627 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 629, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 701 + ], + "type": "text", + "content": "- Drawing inspiration from the shortest path principle, we propose a novel local structure-aware Mamba, TSP-Mamba, enabling content-adaptive propagation for efficient single-image super-resolution. We design an efficient TSP-LSTM to facilitate effective local aggregation by leveraging the locally optimal scan path." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "- Extensive experiments show that the proposed TSP" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 273, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 273, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 273, + 555, + 308 + ], + "type": "text", + "content": "Mamba framework achieves state-of-the-art performance across multiple lightweight SISR benchmarks, requiring only an additional 1-2 GFLOPs for 720P outputs (Tab. 1)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 308, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 308, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 555, + 357 + ], + "type": "text", + "content": "- Beyond lightweight SISR, we also investigate the scalability and generalization capabilities of TSP-Mamba for various vision tasks, including classic SISR, and strong generalization for depth map super-resolution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 366, + 405, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 366, + 405, + 378 + ], + "spans": [ + { + "bbox": [ + 314, + 366, + 405, + 378 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 388, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 591 + ], + "type": "text", + "content": "CNN-based Methods. Due to their low complexity and effective feature extraction capabilities, CNNs have been widely employed for SISR tasks. SRCNN [9] was a pioneering work that introduced deep convolutional neural network (CNN) architectures specifically designed for single image super-resolution (SISR). SRCNN features a simple three-layer structure consisting of patch extraction, nonlinear mapping, and reconstruction, and it demonstrated competitive performance against traditional non-deep learning methods, sparking the development of numerous lightweight CNN-based approaches in the SISR domain. Inspired by dictionary-learning models [15, 17, 30], LAPAR [20] learns linear coefficients associated with pre-defined basic upsampling kernels to create an optimal pixel-specific kernel, leading to superior super-resolution outcomes. LatticeNet [26] introduces a parameter-efficient convolutional lattice block to extract hierarchical contextual features." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Transformer-based Methods. On the other hand, Transformer-based techniques [4-6, 8, 21, 35, 37-39] have achieved impressive results in SISR but are still hindered by their high computational complexity. The cost of modeling non-local self-similarity scales quadratically with the size of the image. To address these efficiency concerns, many window-based Transformer frameworks have been developed, inspired by the success of the Swin Transformer [24]. For instance, SwinIR [23] introduces a residual window-based transformer block (RSTB) for general image restora" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28135" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 74, + 549, + 248 + ], + "blocks": [ + { + "bbox": [ + 60, + 74, + 549, + 248 + ], + "lines": [ + { + "bbox": [ + 60, + 74, + 549, + 248 + ], + "spans": [ + { + "bbox": [ + 60, + 74, + 549, + 248 + ], + "type": "image", + "image_path": "ea5df24049ab98011ea6a26ff048f3347d51a9f973e5e056c5eaa04bf77deab7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 258, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 292 + ], + "type": "text", + "content": "Figure 3. Illustration of our proposed TSP-Mamba, including a shallow feature extraction module ('Head'), a deep feature extraction module and a high-resolution reconstruction module ('Tail'). The key point is that the TSP-LSTM, employed for local path-adaptive aggregation, facilitates subsequent global scanning of 2D-SSM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 299, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 295, + 346 + ], + "type": "text", + "content": "tion. DLGSA-1 [22] enhances token aggregation by using a global sparse attention mechanism. Omni-SR [33] presents spatial-channel attention for lightweight SISR, achieving better performance and efficiency trade-off." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 636 + ], + "type": "text", + "content": "Mamba-based Methods. Recently, Mamba-based frameworks [13] have demonstrated impressive capabilities in modeling long-sequence dependencies while maintaining linear computational complexity. At the heart of Mamba is a structured state space model that expands the effective receptive field, allowing the framework to capture long-range dependencies more efficiently than traditional methods like CNNs and Transformers. This capacity to model global dependencies with linear complexity has positioned Mamba as a compelling alternative, particularly in resource-constrained environments. Building on Mamba's success, numerous follow-up studies [2, 11, 12, 14, 32] have explored the potential of designing different Mamba-based frameworks for various vision tasks. For example, MambaIR [14] offers a simple yet effective Mamba-based baseline for image restoration, while RetinexMamba [2] leverages Mamba's long-range dependency modeling for low-light image enhancement. Although these approaches integrate Mamba's global sequential reasoning (1D) into vision problems, the fixed scan pipeline often overlooks the locality and content-adaptive nature of 2D images. In this paper, we aim to develop an efficient and content-adaptive local propagation and aggregation method to improve the performance of Mamba-based frameworks, particularly for resource-constrained SISR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 651, + 116, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 116, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 116, + 662 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 185, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 185, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 185, + 682 + ], + "type": "text", + "content": "3.1. Preliminaries: Mamba" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We begin by presenting a concise overview of the Mamba pipeline, followed by a discussion of the limitations asso" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 299, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 347 + ], + "type": "text", + "content": "ciated with the rigid 2D-to-1D scan scheme for low-level vision tasks. Selective structured state space models (SSMs) convert a 1D sequence " + }, + { + "bbox": [ + 313, + 299, + 555, + 347 + ], + "type": "inline_equation", + "content": "x(t)\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 299, + 555, + 347 + ], + "type": "text", + "content": " to the response " + }, + { + "bbox": [ + 313, + 299, + 555, + 347 + ], + "type": "inline_equation", + "content": "y(t)\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 299, + 555, + 347 + ], + "type": "text", + "content": " through a latent state " + }, + { + "bbox": [ + 313, + 299, + 555, + 347 + ], + "type": "inline_equation", + "content": "h(t)\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 299, + 555, + 347 + ], + "type": "text", + "content": ". This process is expressed:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 382, + 354, + 485, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 354, + 485, + 367 + ], + "spans": [ + { + "bbox": [ + 382, + 354, + 485, + 367 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t),", + "image_path": "373c0b81a0fcb9ae708ab03cc66c3aa84d7e46914b8a64f117de3c2f38fa95fb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 383, + 367, + 555, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 367, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 383, + 367, + 555, + 382 + ], + "type": "interline_equation", + "content": "y (t) = \\mathbf {C} h (t) + \\mathbf {D} x (t), \\tag {1}", + "image_path": "fb401d86622436e50bfe5caa6968841477a82616006eed2451bb18feba5ff5b6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": " are time-varying matrices and " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": " is the state number. In practical applications, the continuous evolution system can be discretized using a zero-order hold assumption. Accordingly, Mamba converts the continuous matrices " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": " to their discrete forms " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{A}}" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{B}}" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": " with a timescale parameter " + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 391, + 555, + 460 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 464, + 440, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 464, + 440, + 478 + ], + "spans": [ + { + "bbox": [ + 373, + 464, + 440, + 478 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {A}} = \\exp (\\Delta \\mathbf {A}),", + "image_path": "e4f28908d028505d7a029dfaa608aa0216ad1b5395380e23c8fb1dadb7eb8eed.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 373, + 477, + 554, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 477, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 373, + 477, + 554, + 495 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {B}} = (\\Delta \\mathbf {A}) ^ {- 1} (e ^ {\\mathbf {A}} - \\mathbf {I}) \\cdot \\Delta \\mathbf {B}. \\tag {2}", + "image_path": "c1d374e5cb7b038032e1cf9d88c0dbd39649e1cddb19b62687c6d58f683d88f0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 503, + 485, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 485, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 485, + 514 + ], + "type": "text", + "content": "We then obtain a discretized formulation:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 517, + 555, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 517, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 344, + 517, + 555, + 530 + ], + "type": "interline_equation", + "content": "h _ {k} = \\bar {\\mathbf {A}} h _ {k - 1} + \\bar {\\mathbf {B}} x _ {k}, \\quad y _ {k} = \\mathbf {C} h _ {k} + \\mathbf {D} x _ {k}. \\tag {3}", + "image_path": "f49044036ec06531d0823af5adf69086c8e1061741a87c18cc0e93c2a81e0318.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 539, + 555, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 571 + ], + "type": "text", + "content": "As we can see, since the overall calculation is a linear combination of the accumulated sum, it can be efficiently implemented as a global 1D convolution:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 329, + 575, + 555, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 575, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 329, + 575, + 555, + 590 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {K}} = \\left(\\mathbf {C} \\bar {\\mathbf {B}}, \\mathbf {C} \\bar {\\mathbf {A}} \\bar {\\mathbf {B}}, \\dots , \\mathbf {C} \\bar {\\mathbf {A}} ^ {S - 1} \\bar {\\mathbf {B}}\\right), \\mathbf {y} = \\mathbf {x} * \\bar {\\mathbf {K}}, \\tag {4}", + "image_path": "479b23fc61cf53e1dd1cb268c47dd5842502097b83de53c6c244213a88687a66.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 598, + 555, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 555, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 598, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 598, + 555, + 619 + ], + "type": "text", + "content": " is a convolution operation and the " + }, + { + "bbox": [ + 313, + 598, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{K}}\\in \\mathbb{R}^S" + }, + { + "bbox": [ + 313, + 598, + 555, + 619 + ], + "type": "text", + "content": " refers to a convolution kernel. " + }, + { + "bbox": [ + 313, + 598, + 555, + 619 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 598, + 555, + 619 + ], + "type": "text", + "content": " is the sequence length." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "Discussion. In NLP tasks, it is natural to employ state space models (SSMs) to process 1D sequences for global information reasoning at a linear computational cost. However, in 2D image scenarios, the neighborhoods often face challenges such as disconnection or long-range propagation to fit within the SSM framework. Although some recent studies have investigated alternative 1D scan paths to supplement" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28136" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "information and improve restoration capability, capturing multiple local neighbors in a single 1D scan remains difficult, which can result in inefficient local aggregation for Mamba-based frameworks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 128, + 169, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 169, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 169, + 139 + ], + "type": "text", + "content": "3.2. Preliminaries: TSP" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 240 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 55, + 144, + 296, + 240 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 144, + 296, + 240 + ], + "type": "text", + "content": " cities and their pairwise distances, the goal is to determine the shortest route that visits each city exactly once. Mathematically, TSP is formulated to find a minimum-cost Hamiltonian cycle in a weighted complete graph, where nodes represent cities and edges denote travel costs (e.g., distance or time). Inspired by TSP's optimal traversal property, we propose a TSP-style local scan scheme to complement the global modeling of vision Mambas in this work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 248, + 181, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 181, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 181, + 260 + ], + "type": "text", + "content": "3.3. Framework Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 432 + ], + "type": "text", + "content": "Fig. 3 illustrates the framework of our proposed TSP-Mamba. Following MambaIR, a recent SOTA baseline [14], TSP-Mamba comprises three main modules: (1) a shallow feature extraction module that transforms the RGB input into a feature space; (2) a deep feature extraction module consisting of multiple TSP-RSSGs (TSP-based Residual State-Space Groups). Each T-RSSG consists of several T-RSSBs (TSP-based Residual State-Space Block). In TSP-RSSB, the key component is the T-VSSM (TSP-based Vision State-Space Module), where we first use a TSP-LSTM layer to perform local path-adaptive aggregation and then pass the local-aggregation information into the 2D-SSM (2D Selective Scan Module in MambaIR[14]); and (3) a high-resolution reconstruction module that produces the final results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 493 + ], + "type": "text", + "content": "In the following section, we mainly focus on the explanation of our proposed TSP-LSTM, including the scanning patch construction for optimal local propagation (Sec.3.4.1), simplified TSP (Sec.3.4.2) and efficient implementation (Sec.3.4.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 500, + 134, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 134, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 134, + 511 + ], + "type": "text", + "content": "3.4. TSP-LSTM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 518, + 204, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 204, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 204, + 529 + ], + "type": "text", + "content": "3.4.1. Scanning Path Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "content": "Given an image feature patch " + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{s \\times s \\times C}" + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "content": " is the patch size and " + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "content": " is the number of channels, our objective is to establish a locally adaptive aggregation path that minimizes the information propagation costs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "text", + "content": "Definition. We formulate the scan path as a Traveling Salesman Problem (TSP) problem: let " + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "inline_equation", + "content": "V = \\{1,2,\\dots n\\}" + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "text", + "content": " be the set of pixels (cities) in image patch " + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "inline_equation", + "content": "E = \\{e_{i,j}(i,j):i,j\\in V\\}" + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "inline_equation", + "content": "D_{e_{i,j}}" + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "text", + "content": " be the cost measure associated with edge " + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "inline_equation", + "content": "e_{i,j}\\in E" + }, + { + "bbox": [ + 55, + 595, + 296, + 667 + ], + "type": "text", + "content": ". Our goal is to find a minimal scan path such that each city (pixel) is visited exactly once." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "text", + "content": "In practice, we use the negative Cosine similarity as the cost measure of an edge " + }, + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "inline_equation", + "content": "e_{i,j}" + }, + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "text", + "content": " in the set " + }, + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 113, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "D _ {e _ {i j}} = 1. 0 - \\operatorname {C o s i n e} \\left(F _ {i}, F _ {j}\\right), \\tag {5}", + "image_path": "9d3fcd8a059705faf5557849369dbd9e7ae77ffd50c542babbf256fd2d631bbd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 324, + 163, + 408, + 224 + ], + "blocks": [ + { + "bbox": [ + 324, + 163, + 408, + 224 + ], + "lines": [ + { + "bbox": [ + 324, + 163, + 408, + 224 + ], + "spans": [ + { + "bbox": [ + 324, + 163, + 408, + 224 + ], + "type": "table", + "html": "
Size (s)TSPOurs
24!2
39!8
416!52
525!824
", + "image_path": "ec369dbb7a2b91e49e1749504e236b9af4a6ed0a9ed3283c7e27a22856a6effe.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 415, + 72, + 553, + 187 + ], + "blocks": [ + { + "bbox": [ + 314, + 99, + 416, + 153 + ], + "lines": [ + { + "bbox": [ + 314, + 99, + 416, + 153 + ], + "spans": [ + { + "bbox": [ + 314, + 99, + 416, + 153 + ], + "type": "text", + "content": "Figure 4. The numbers of scan path candidates of TSP and our simplified TSP scheme under different patch size " + }, + { + "bbox": [ + 314, + 99, + 416, + 153 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 314, + 99, + 416, + 153 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 415, + 72, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 415, + 72, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 415, + 72, + 553, + 187 + ], + "type": "image", + "image_path": "adf2335ff54d627c8ad948b1e4b7b1f96624f5186315aef740dfcdfc39cea06a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 196, + 556, + 251 + ], + "lines": [ + { + "bbox": [ + 413, + 196, + 556, + 251 + ], + "spans": [ + { + "bbox": [ + 413, + 196, + 556, + 251 + ], + "type": "text", + "content": "Figure 5. Speedup ratio curve under different patch sizes. Our effective measure hashing strategy attains notable speedup " + }, + { + "bbox": [ + 413, + 196, + 556, + 251 + ], + "type": "inline_equation", + "content": "(\\log_{10})" + }, + { + "bbox": [ + 413, + 196, + 556, + 251 + ], + "type": "text", + "content": " for efficient TSP searching." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "type": "inline_equation", + "content": "F_{i}, F_{j}" + }, + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "type": "text", + "content": " are the corresponding feature vector of the pixel " + }, + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "type": "text", + "content": ". More discussion is shown in Tab. 3 (A)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 288, + 407, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 407, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 407, + 300 + ], + "type": "text", + "content": "3.4.2. Simplified TSP" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 303, + 555, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 351 + ], + "type": "text", + "content": "The Traveling Salesman Problem (TSP) is a classic NP-hard problem. With a large " + }, + { + "bbox": [ + 313, + 303, + 555, + 351 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 303, + 555, + 351 + ], + "type": "text", + "content": ", searching for an optimal scan path becomes time-consuming (" + }, + { + "bbox": [ + 313, + 303, + 555, + 351 + ], + "type": "inline_equation", + "content": "O(n!)" + }, + { + "bbox": [ + 313, + 303, + 555, + 351 + ], + "type": "text", + "content": "). To simplify the problem, we adopt the following strategies in this work:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 351, + 553, + 434 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 313, + 351, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 553, + 374 + ], + "type": "text", + "content": "(1) We limit the overall node count " + }, + { + "bbox": [ + 313, + 351, + 553, + 374 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 351, + 553, + 374 + ], + "type": "text", + "content": " to the pixels within a local image patch, thereby reducing overall complexity;" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 375, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 553, + 399 + ], + "type": "text", + "content": "(2) We designate the first city '1' (the top-left pixel of the given patch) as the starting node;" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 399, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 434 + ], + "type": "text", + "content": "(3) At each intermediate step, we limit next-step candidates to the four neighboring 2D locations1 to reduce the search space further." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 435, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 483 + ], + "type": "text", + "content": "Based on these three simplifications mentioned, it largely reduced the number of total candidate scan paths (in Fig. 4). Thus, we can initially employ the backtracking algorithm [10] to identify all possible scan paths:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 352, + 488, + 515, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 488, + 515, + 520 + ], + "spans": [ + { + "bbox": [ + 352, + 488, + 515, + 520 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} T ^ {\\{1, 2, \\dots , l, \\dots , L \\}} = \\text {B a c k T r a c k i n g} (1, n), \\\\ \\text {w h e r e} T ^ {l} = \\left\\{i _ {1} ^ {l}, i _ {2} ^ {l}, \\dots , i _ {n} ^ {l} \\right\\}, i _ {1} ^ {l} = 1, \\\\ \\end{array}", + "image_path": "ee74d15a1fe82929b22365434370b61c9b608b15d7dcbc44e4651eabb714cd43.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "text", + "content": "-th path " + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "inline_equation", + "content": "T^l" + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "text", + "content": " is represented by the ordered pixel indices " + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "inline_equation", + "content": "(\\{i_1^l, i_2^l, \\dots, i_n^l\\})" + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "text", + "content": " is the count of all scan candidates. We calculate the cost of the " + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "text", + "content": "-th scan path " + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "inline_equation", + "content": "T^l" + }, + { + "bbox": [ + 313, + 527, + 555, + 564 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 386, + 570, + 555, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 570, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 386, + 570, + 555, + 592 + ], + "type": "interline_equation", + "content": "D ^ {l} = \\sum_ {k = 1} ^ {n - 1} D _ {e _ {i _ {k} ^ {l}, i _ {k + 1} ^ {l}}} \\tag {7}", + "image_path": "b7e78f487bbb79dd2934b027733a3214220d2de682012924ca63f213a903b4b9.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 599, + 553, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 553, + 622 + ], + "type": "text", + "content": "Finally, we choose the optimal scan path " + }, + { + "bbox": [ + 313, + 599, + 553, + 622 + ], + "type": "inline_equation", + "content": "T^{l^*}" + }, + { + "bbox": [ + 313, + 599, + 553, + 622 + ], + "type": "text", + "content": " with the lowest measure cost:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 400, + 623, + 555, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 623, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 400, + 623, + 555, + 643 + ], + "type": "interline_equation", + "content": "l ^ {*} = \\underset {l} {\\operatorname {a r g m i n}} D ^ {l} \\tag {8}", + "image_path": "ecf5839c543f8ce8b8840f27d8f833af398a14035d4249ea9507e87954d4fc79.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 646, + 451, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 451, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 451, + 657 + ], + "type": "text", + "content": "3.4.3. Effective Implementation" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 661, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 661, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 661, + 555, + 696 + ], + "type": "text", + "content": "As indicated in Eq. 7, we have to calculate the " + }, + { + "bbox": [ + 313, + 661, + 555, + 696 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 313, + 661, + 555, + 696 + ], + "type": "text", + "content": " pairwise distances for each scan path and sum them up. We have two observations for effective path cost computation:" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "type": "text", + "content": "left " + }, + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "type": "inline_equation", + "content": "(-1,0)" + }, + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "type": "text", + "content": ", right " + }, + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "type": "inline_equation", + "content": "(+1,0)" + }, + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "type": "text", + "content": ", top " + }, + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "type": "inline_equation", + "content": "(0,-1)" + }, + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "type": "text", + "content": ", and bottom " + }, + { + "bbox": [ + 324, + 702, + 499, + 713 + ], + "type": "inline_equation", + "content": "(0,+1)" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28137" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "1. Symmetry of Distance Measurement. As indicated in Eq. 5, for edges " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "e_{i,j}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "e_{j,i}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "D_{e_{i,j}} = D_{e_{j,i}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 120 + ], + "type": "text", + "content": "2. Highly overlapping edges for different scan paths. Different scene paths may share lots of the same edges." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 121, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 216 + ], + "type": "text", + "content": "Based on this fact, we propose to design a hash-based technique to store the unique edges in all scan paths into an edge list " + }, + { + "bbox": [ + 55, + 121, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\{e_{k_0}, e_{k_1}, \\dots, e_{k_m}\\}" + }, + { + "bbox": [ + 55, + 121, + 296, + 216 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 121, + 296, + 216 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 121, + 296, + 216 + ], + "type": "text", + "content": " is the length of the unique edge list) to avoid repeated distance calculation. Accordingly, we can effectively compute all unique edge distances using Eq. 5 at one time and obtain the associated edge distance list " + }, + { + "bbox": [ + 55, + 121, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\{D_{e_{k_0}}, D_{e_{k_1}}, \\dots, D_{e_{k_m}}\\}" + }, + { + "bbox": [ + 55, + 121, + 296, + 216 + ], + "type": "text", + "content": ". Finally, we can obtain the path propagation cost in Eq. 7 by accumulating all its edges." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 217, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 276 + ], + "type": "text", + "content": "The speedup ratio between the direct per-path cost computation and our efficiency implementation is shown in Fig. 5, we can see our effective hash-based calculation attains a significant speedup effect. Additionally, it becomes more effective with an increased patch size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "text", + "content": "Summary. In this part, we propose a simplified TSP formulation to effectively search the shortest path for a given 2D image patch, in which all neighborhoods are fused in a relatively short-term propagation. Therefore, it is capable of effectively integrating local information to facilitate the later global SSM-style propagation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 370, + 134, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 134, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 134, + 381 + ], + "type": "text", + "content": "3.4.4. TSP-LSTM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 385, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 469 + ], + "type": "text", + "content": "Bidirectional Propagation. After the scan path optimization, we perform a bidirectional LSTM for mutual information propagation. Especially, given a 2D image patch " + }, + { + "bbox": [ + 55, + 385, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 385, + 296, + 469 + ], + "type": "text", + "content": ", we use the above simplified TSP scheme to get the optimal scan chain " + }, + { + "bbox": [ + 55, + 385, + 296, + 469 + ], + "type": "inline_equation", + "content": "T^{l^{*}} = \\{i_{1} \\rightleftharpoons i_{2}, \\dots, \\rightleftharpoons i_{n}\\}^{2}" + }, + { + "bbox": [ + 55, + 385, + 296, + 469 + ], + "type": "text", + "content": ", where “" + }, + { + "bbox": [ + 55, + 385, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\rightleftharpoons" + }, + { + "bbox": [ + 55, + 385, + 296, + 469 + ], + "type": "text", + "content": "” means we employ the bidirectional propagation scheme (as shown in Fig. 2) (b) for effective and balanced aggregation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 470, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 506 + ], + "type": "text", + "content": "For implementation, we flip the reordered sequence " + }, + { + "bbox": [ + 55, + 470, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\{i_1,i_2,\\dots ,i_n\\}" + }, + { + "bbox": [ + 55, + 470, + 296, + 506 + ], + "type": "text", + "content": " to obtain the inverse sequence and concatenate them for bidirectional aggregation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": "Distance Embedding. Within an image patch, discontinuities may need to be addressed in the chain-style propagation. Specifically, the aggregation weight of the previously accumulated information should positively correlate with the distance between the previous and current nodes (pixels). To account for this, inspired by fast global image smoothing (FGS [29]), we incorporate distance embedding features into the sequences to explicitly model this effect. We calculate the pairwise distances as shown in Eq. 7 and embed them (as a one-channel feature) into the propagation sequence feature." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 666 + ], + "type": "text", + "content": "Lastly, the propagated feature " + }, + { + "bbox": [ + 55, + 641, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 641, + 296, + 666 + ], + "type": "text", + "content": " is passed through the 2DSSM module (in Eq. 4) for global Mamba-style aggregation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 394, + 84 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 91, + 373, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 373, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 373, + 104 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 109, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 205 + ], + "type": "text", + "content": "Implementation Details. Our TSP-Mamba model is implemented in PyTorch and integrates several common data augmentation techniques, such as random cropping, vertical/horizontal flipping, and rotation. For training, we use the Adam optimizer [18] with cosine annealing [25], starting with an initial learning rate of " + }, + { + "bbox": [ + 313, + 109, + 555, + 205 + ], + "type": "inline_equation", + "content": "4e - 4" + }, + { + "bbox": [ + 313, + 109, + 555, + 205 + ], + "type": "text", + "content": ". The batch size is set to 32, with input images sized at " + }, + { + "bbox": [ + 313, + 109, + 555, + 205 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 109, + 555, + 205 + ], + "type": "text", + "content": ". Training is performed over 600K iterations using four NVIDIA RTX 3090 GPUs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 205, + 556, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 556, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 556, + 337 + ], + "type": "text", + "content": "Scalable Model Configuration. In general, we train multiple TSP-Mamba models with different configurations. TSP-Mamba is based on the MambaIR baseline, utilizing 4 T-RSSGs, each containing 6 T-RSSBs, with a channel size of 60, denoted as “[6,6,6,6]@C60”. Additionally, we investigate the scalability of TSP-Mamba to better understand the trade-off between performance and model complexity. Specially, we use “[2,4,2,6]@C40”" + }, + { + "bbox": [ + 313, + 205, + 556, + 337 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 313, + 205, + 556, + 337 + ], + "type": "text", + "content": " and “[4,6,4,6]@C48”" + }, + { + "bbox": [ + 313, + 205, + 556, + 337 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 313, + 205, + 556, + 337 + ], + "type": "text", + "content": " to train another two smaller SISR models under " + }, + { + "bbox": [ + 313, + 205, + 556, + 337 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 205, + 556, + 337 + ], + "type": "text", + "content": " SISR setting. Training these TSP-Mamba models usually takes around 3-6 days." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 338, + 556, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 556, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 556, + 433 + ], + "type": "text", + "content": "Benchmark Datasets. In line with previous studies [7, 20, 23], we use the DIV2K [1] image dataset for training. We then perform extensive evaluations on several widely-used SISR benchmarks, including Set5 [3], Set14 [36], BSD100 [27], Urban100 [16], and Manga109 [28]. Our quantitative analysis is based on PSNR and SSIM metrics, and following standard practice, we report results specifically for the Y channel in the YCbCr color space." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 442, + 493, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 493, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 493, + 455 + ], + "type": "text", + "content": "4.2. Comparison with SOTA Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 460, + 556, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 543 + ], + "type": "text", + "content": "We perform extensive comparisons with a wide range of lightweight SISR models: LAPAR-A (NeurIPS20) [20], LatticeNet (ECCV20) [26], RLFN (CVPRW22) [19], SwinIR-light [23], DLGSA-1 (ICCV23) [22], Onmi-SR (CVPR23) [33], SRFormer (ICCV23) [39], ATDS (CVPR24) [37], MambaIR (ECCV24) [14], HiTSR [38] (ECCV24)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": "Quantitative Comparison. Tab. 1 shows the quantitative evaluation results, highlighting that our proposed TSP-Mamba consistently outperforms existing methods across all benchmarks. In particular, our model achieves significant improvements over MambaIR, with gains of up to " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "0.71\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "0.58\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "0.69\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": " on the Manga109 dataset for the " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": " scales, respectively, while maintaining similar parameter and FLOP requirements." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 640, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 554, + 665 + ], + "type": "text", + "content": "Moreover, TSP-Mamba-Small beats all leading lightweight SISR methods with smaller model sizes and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 672, + 295, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 693 + ], + "type": "text", + "content": "2For clarity, we drop the upper letter and denote " + }, + { + "bbox": [ + 55, + 672, + 295, + 693 + ], + "type": "inline_equation", + "content": "T^{l^*}" + }, + { + "bbox": [ + 55, + 672, + 295, + 693 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 672, + 295, + 693 + ], + "type": "inline_equation", + "content": "T^{l^*} = \\{i_1,i_2,\\dots ,i_n\\}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 693, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 296, + 713 + ], + "type": "text", + "content": "3Each pixel can fuse all its neighboring pixels in the bidirectional short-range propagation scheme" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 673, + 553, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 553, + 693 + ], + "type": "text", + "content": "4TSP-Mamba-Tiny: 4 TSP-RSSGs contain 2,4,2,6 TSP-RSSBs and the channel size is 40" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "text", + "content": "5TSP-Mamba-Small: 4 TSP-RSSGs contain 4,6,4,6 TSP-RSSBs and the channel size is 48" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28138" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 113, + 541, + 571 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with SOTA lightweight SISR methods on multiple benchmark datasets. The best and second-best results on the default training setting (DIV2K) are highlighted in red and blue, respectively. We follow most lightweight SISR methods and calculate all Flops with an output size of " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "{1280} \\times {720}" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 113, + 541, + 571 + ], + "lines": [ + { + "bbox": [ + 69, + 113, + 541, + 571 + ], + "spans": [ + { + "bbox": [ + 69, + 113, + 541, + 571 + ], + "type": "table", + "html": "
MethodScaleParameters (K)Flops (G)Set5 PSNR / SSIMSet14 PSNR / SSIMBSD100 PSNR / SSIMUrban100 PSNR / SSIMManga109 PSNR / SSIM
LAPAR-A54817138.01 / 0.960533.62 / 0.918332.19 / 0.899932.10 / 0.928338.67 / 9772
LatticeNet756169.538.15 / 0.961033.78 / 0.919332.25 / 0.900532.43 / 0.9302- / -
SwinIR-light910244.438.14 / 0.961133.86 / 0.920632.31 / 0.901232.76 / 0.934039.12 / 0.9783
DLGSA-light745170.038.20 / 0.961233.89 / 0.920332.30 / 0.901232.94 / 0.935539.29 / 0.9780
Omni-SR772194.538.22 / 0.961333.98 / 0.921032.36 / 0.902033.05 / 0.936339.28 / 0.9784
SRFormer ×285323638.23 / 0.961333.94 / 0.920932.36 / 0.901932.91 / 0.935339.28 / 0.9785
HiT-SR847226.538.26 / 0.961534.01 / 0.921432.37 / 0.902333.13 / 0.937239.47 / 0.9787
ATD-S75334938.29 / 0.961634.10 / 0.921732.39 / 0.902333.27 / 0.937539.52 / 0.9789
MambaIR136319838.16 / 0.961034.00 / 0.921232.34 / 0.901732.92 / 0.935639.31 / 0.9779
TSP-Mamba130519938.46 / 0.964834.42 / 0.924332.52 / 0.904233.57 / 0.940540.02 / 0.9808
MambaIR-Small80611138.09 / 0.960233.85 / 0.920132.31 / 0.901532.58 / 0.932139.17 / 0.9787
TSP-Mamba-Small76911238.22 / 0.964034.19 / 0.922732.43 / 0.903033.18 / 0.937439.49 / 0.9791
MambaIR-Tiny4435937.99 / 0.960233.72 / 0.910932.21 / 0.900232.22 / 0.929138.87 / 0.9772
TSP-Mamba-Tiny4185938.04 / 0.963133.90 / 0.920532.29 / 0.901032.55 / 0.931739.12 / 0.9784
LAPAR-A54411434.36 / 0.926730.34 / 0.842129.11 / 0.805428.15 / 0.852333.51 / 9441
LatticeNet76576.334.53 / 0.928130.39 / 0.842429.15 / 0.805928.33 / 0.8538- / -
SwinIR-light918110.834.62 / 0.928930.54 / 0.846329.20 / 0.808228.66 / 0.862433.98 / 0.9478
DLGSA-light75275.434.70 / 0.929530.58 / 0.846529.24 / 0.808928.83 / 0.865334.16 / 0.9483
Omni-SR78088.434.70 / 0.929430.57 / 0.846929.28 / 0.809428.84 / 0.865634.22 / 0.9487
SRFormer ×386110534.67 / 0.929630.57 / 0.846929.26 / 0.809928.81 / 0.865534.19 / 0.9489
HiT-SR85510234.75 / 0.930030.61 / 0.847529.29 / 0.810628.99 / 0.868734.53 / 0.9502
MambaIR13718934.72 / 0.929630.63 / 0.847529.29 / 0.809929.00 / 0.868934.39 / 0.9495
TSP-Mamba13138934.80 / 0.933230.93 / 0.851029.41 / 0.813929.43 / 0.874434.97 / 0.9522
MambaIR-Small8125034.65 / 0.931930.59 / 0.847129.26 / 0.810228.84 / 0.864434.02 / 0.9472
TSP-Mamba-Small7765034.65 / 0.932230.66 / 0.847829.30 / 0.811428.94 / 0.868034.41 / 0.9495
MambaIR-Tiny4482734.36 / 0.922330.26 / 0.842829.12 / 0.806328.24 / 0.853233.74 / 0.9452
TSP-Mamba-Tiny4232734.49 / 0.930930.46 / 0.845829.22 / 0.809428.70 / 0.862334.16 / 0.9483
LAPAR-A6599432.15 / 0.894428.61 / 0.781827.61 / 0.736626.14 / 0.787130.42 / 0.9074
LatticeNet77743.632.30 / 0.896228.68 / 0.783027.62 / 0.736726.25 / 0.7873- / -
SwinIR-light93063.632.44 / 0.897628.77 / 0.785827.69 / 0.740626.47 / 0.798030.92 / 0.9151
DLGSA-light76142.532.54 / 0.899328.84 / 0.787127.73 / 0.741526.66 / 0.803331.13 / 0.9161
Omni-SR79250.932.49 / 0.898828.78 / 0.785927.71 / 0.741526.64 / 0.801831.02 / 0.9151
SRFormer ×48736332.51 / 0.898828.82 / 0.787227.73 / 0.742226.67 / 0.803231.17 / 0.9165
HiT-SR8665832.55 / 0.899928.87 / 0.788027.75 / 0.743226.80 / 0.806931.26 / 0.9171
ATD-S7698732.63 / 0.899828.89 / 0.788627.79 / 0.744026.97 / 0.810731.48 / 0.9198
MambaIR13835132.51 / 0.899328.85 / 0.787627.75 / 0.742326.75 / 0.805131.26 / 0.9175
TSP-Mamba13245332.71 / 0.902429.05 / 0.792527.91 / 0.747627.29 / 0.817431.95 / 0.9235
MambaIR-Small8212932.28 / 0.900128.70 / 0.786327.72 / 0.741226.56 / 0.798431.12 / 0.9156
TSP-Mamba-Small7853032.48 / 0.902428.90 / 0.789327.79 / 0.744226.93 / 0.808731.42 / 0.9184
MambaIR-Tiny4561532.22 / 0.899128.63 / 0.784727.68 / 0.739826.43 / 0.794030.86 / 0.9124
TSP-Mamba-Tiny4311632.29 / 0.899728.79 / 0.786027.72 / 0.740526.61 / 0.799231.04 / 0.9140
MambaIR (×16)1772523.24 / 0.641922.29 / 0.545422.94 / 0.521820.46 / 0.514821.15 / 0.6628
TSP-Mamba (×16)1714523.33 / 0.651522.38 / 0.545222.81 / 0.516920.53 / 0.518421.20 / 0.6700
", + "image_path": "5854e7f5efbfe332d90767046ec9cb6bb5025d7ed6de0e4dd6ac19e5a3e2d964.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": "less computational complexity. For instance, it surpasses MambaIR by 0.26dB on Urban " + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "inline_equation", + "content": "100 \\times 2" + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": ", while only requiring " + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "inline_equation", + "content": "59\\%" + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": " parameters and " + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "inline_equation", + "content": "57\\%" + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": " Flops." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 616, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 700 + ], + "type": "text", + "content": "Inference efficiency. Tab. 2 presents a detailed comparison of inference costs across different models, including CNN-based, Transformer-based, and Mamba-based frameworks. Overall, the CNN-based LatticeNet achieves the highest inference efficiency. Compared to other transformer-based and Mamba-based models, our proposed TSP-Mamba achieves competitive inference latency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Qualitative Comparison. Fig. 6 showcases several visual" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "content": "examples. It is clear that TSP-Mamba effectively reconstructs accurate image textures with fewer super-resolved artifacts. In contrast, other approaches struggle to restore clear image patterns." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 644, + 415, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 415, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 415, + 656 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "To validate our design, we conduct extensive experiments, reporting quantitative results on the Urban100 and Manga109 benchmarks at the " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": " setting. We primarily adopt the TSP-Mamba-Small architecture for fast analysis." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28139" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 552, + 244 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 552, + 244 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 552, + 244 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 552, + 244 + ], + "type": "image", + "image_path": "d6aaa5f94cacb147a41f5bc9cd65db9c36c5601911e861d706968ae19c8de3bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 257, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 61, + 257, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 257, + 547, + 269 + ], + "type": "text", + "content": "Figure 6. Qualitative results of different lightweight SISR methods under " + }, + { + "bbox": [ + 61, + 257, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 61, + 257, + 547, + 269 + ], + "type": "text", + "content": " setting. Our predictions present more detailed structures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 79, + 305, + 533, + 353 + ], + "blocks": [ + { + "bbox": [ + 54, + 279, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 279, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 555, + 304 + ], + "type": "text", + "content": "Table 2. Inference cost comparison of different lightweight SISR models under " + }, + { + "bbox": [ + 54, + 279, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 54, + 279, + 555, + 304 + ], + "type": "text", + "content": " setting. Inference time is tested with an output resolution of " + }, + { + "bbox": [ + 54, + 279, + 555, + 304 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 54, + 279, + 555, + 304 + ], + "type": "text", + "content": " using an NVIDIA 3090 GPU. The top-3 results are highlighted with red,blue,underline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 305, + 533, + 353 + ], + "lines": [ + { + "bbox": [ + 79, + 305, + 533, + 353 + ], + "spans": [ + { + "bbox": [ + 79, + 305, + 533, + 353 + ], + "type": "table", + "html": "
MetricsLatticeNetDLGSA-lightSwinIR-lightSRFormerMambaIR / Small / TinyTSP-Mamba / Small / Tiny
Time (ms) ↓973416060 / 50 / 3872 / 58 / 39
Memory (MB) ↓146616245441572686 / 536 / 528732 / 614 / 582
", + "image_path": "852424bb56f051db9dffb4869be4ed045e0278b507cefdcf252f5cc6e1e183f8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 358, + 297, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 297, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 297, + 407 + ], + "type": "text", + "content": "Different Measure Cost in TSP-LSTM. Apart from Eq. 5, we try the widely-used L1/L2 distance metrics for scan path searching. Tab. 3 (A) concludes that our used negative Cosine distance performs better than the L1/L2 distances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 533 + ], + "type": "text", + "content": "TSP-LSTM vs. Uniform-LSTM. TSP-LSTM aims to perform an adaptive scan path by constructing a minimal propagation path. To assess the impact of our adaptive scan path, we also train two uniform scan models (vertical, and horizontal). We compare the proposed TSP-LSTM with the other models (UV-LSTM, HV-LSTM). As shown in Tab. 3 (B), we can observe that TSP-LSTM significantly outperforms the uniform scan models, highlighting the importance of adaptive local propagation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 547, + 298, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 298, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 298, + 632 + ], + "type": "text", + "content": "Different Patch Sizes. As mentioned in Sec. 3.4, we set the patch size to " + }, + { + "bbox": [ + 55, + 547, + 298, + 632 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 55, + 547, + 298, + 632 + ], + "type": "text", + "content": " considering the local receptive field and computational costs. When the patch size is increased, it becomes prolonged to construct the corresponding TSP path. We investigate the influence of different patch sizes. Tab. 3 (C) suggests the reconstructed performance is positively related to the patch size but with larger latency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 297, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 297, + 719 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 297, + 719 + ], + "type": "text", + "content": "Finer Design in TSP-LSTM. Lastly, we evaluate the impact of bidirectional propagation of TSP-LSTM. We train another model with only a forward/backward scan pipeline to do this. In Tab. 3 (D), it is clear that our bidirectional propagation scheme is more effective than its counterparts with only a single-direction scan strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 358, + 556, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 556, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 556, + 444 + ], + "type": "text", + "content": "Different Aggregation Layers. In this paper, we present a TSP-LSTM block to get locally fused information. To verify the effectiveness of TSP-LSTM, we replace it with other designs for local compensation, such as convolution, RNN, and Mamba. The results are shown in Tab. 3 (E) below, which suggests that our TSP attains the best performance and second-best efficiency compared with other designs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 450, + 389, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 389, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 389, + 462 + ], + "type": "text", + "content": "4.4. Extensions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 468, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 612 + ], + "type": "text", + "content": "Classic SISR. Though TSP-Mamba is trained for lightweight Mamba-based architecture, i.e., MambaIR, we additionally explore the potential improvement for the classic SISR setting. To do this, we follow MambaIR-L(large) to enlarge the channel size to 180 and use 6 TSP-RSSG blocks (each one contains 6 basic TSP-SSD layers) to obtain the TSP-Mamba-L model and train this classic SISR model using the same DF2K training dataset. The quantitative results are illustrated in Tab. 4. We can see that TSP-Mamba-L consistently outperforms other classic SISR methods significantly. For instance, it achieves 0.32dB, 0.37dB, and 0.39dB improvements over MambaIR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 626, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 556, + 687 + ], + "type": "text", + "content": "Depth Map SR. We also compare the TSP-Mamba model with our baseline model (MambaIR) on the depth SR task for generalization analysis. To do this, without training on any depth images, we directly assess the two models on the NYU-V2 [31] depth benchmark" + }, + { + "bbox": [ + 313, + 626, + 556, + 687 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 313, + 626, + 556, + 687 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 313, + 626, + 556, + 687 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 626, + 556, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 626, + 556, + 687 + ], + "type": "inline_equation", + "content": "\\times 16" + }, + { + "bbox": [ + 313, + 626, + 556, + 687 + ], + "type": "text", + "content": " settings." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "type": "text", + "content": "NYU V2 consists of 449 testing depth images at a size of " + }, + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "type": "text", + "content": ". We use bicubic interpolation to generate the low-resolution depth images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28140" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 140, + 558, + 229 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": "Table 3. Ablation Studies. A: The effects of different distance measures, with our default choice being the negative cosine distance as indicated in Eq. 5. B: Quantitative evaluation of uniform scanning strategies (vertical and horizontal), abbreviated as 'UV' and 'HV', compared to our TSP-based scan. C: The influence of different local scan patch sizes (P2-P5 indicates local patch sizes ranging from 2 to 5). D: Impacts of distance embeddings and bidirectional propagation. M1 drops both of the two designs. M2 only adopts distance embedding. M3 only uses the bidirectional propagation. M4 employs both of the two designs. E: The effects of various local aggregation modules (M5-M8 employ Conv1D, RNN, Mamba, and LSTM, respectively)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 140, + 558, + 229 + ], + "lines": [ + { + "bbox": [ + 56, + 140, + 558, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 558, + 229 + ], + "type": "table", + "html": "
AnalysisA.B.C.D.E.
L1L2Neg. Cos.UVUHTSPP2P3P4P5M1M2M3M4M5M6M7M8
Urban10026.5526.5426.6126.3126.3826.6126.3726.4326.6126.6126.1626.3826.2526.6126.4926.4426.4726.61
0.74030.79820.79920.79070.79360.79920.79330.79430.79920.79940.78810.79390.79070.79920.79580.79500.79560.7992
Manga10931.0631.0731.0430.7830.8631.0430.7530.9131.0431.0230.5030.8030.6531.0430.9630.9530.9731.04
0.91400.91440.91400.91040.91260.91400.91130.91310.91400.91380.90650.91030.90840.91400.91360.91320.91340.9140
Param.(K)431431431431431431431431431431430431430431440436428431
FLOPs (G)16.216.216.215.415.416.215.215.216.220.314.915.116.116.215.515.716.916.2
", + "image_path": "180c20f5dc7c5c38f594c31d7d9f314079fc54ba6378c98b0d57aa18522c55cf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 297, + 299, + 404 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 296, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 288 + ], + "type": "text", + "content": "Table 4. Quantitative comparison with recent SOTA Classic SR models. All the models are trained on the DF2K dataset for fair comparison. We report the averaged PSNR/SSIM values over the five benchmarks (Set5, Set14, BSD100, Urban100, Manga109)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 297, + 299, + 404 + ], + "lines": [ + { + "bbox": [ + 56, + 297, + 299, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 299, + 404 + ], + "type": "table", + "html": "
SettingsMetricsSwinIRSRFormer-LHATMambaIR-LTSP-Mamba-L
×2PSNR35.8335.9436.1736.0536.37
SSIM0.94280.94350.94460.94380.9457
Param.11.8M10.4M20.6M20.4M20.6M
×3PSNR32.0532.1532.2932.1932.56
SSIM0.88720.88850.88810.89000.8918
Param.11.8M10.6M20.8M20.6M20.8M
×4PSNR29.8929.9730.1430.0430.43
SSIM0.83990.84160.84410.84140.8470
Param.11.8M10.5M20.8M20.6M20.8M
", + "image_path": "53c257f5ac2f0987a54507370015083859ac9cc4547d0bb3522c45708de0bf2e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 423, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 297, + 544 + ], + "type": "text", + "content": "We use the PSNR, SSIM, and RMSE (the root-mean-square error) metrics for quantitative evaluation. The quantitative results are shown in the left Fig. 7. We can see that TSP-Mamba consistently outperforms its baseline model on all the objective metrics. For instance, TSP-Mamba exhibits superior performance compared to MambaIR with a PSNR improvement of 0.96dB (0.92dB) for " + }, + { + "bbox": [ + 54, + 423, + 297, + 544 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 54, + 423, + 297, + 544 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 54, + 423, + 297, + 544 + ], + "type": "inline_equation", + "content": "\\times 16" + }, + { + "bbox": [ + 54, + 423, + 297, + 544 + ], + "type": "text", + "content": ") configurations. The visual examples in Fig. 7 illustrate that TSP-Mamba generates clearer structures, leading to higher accuracy when compared to the baseline model." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 71, + 573, + 280, + 630 + ], + "blocks": [ + { + "bbox": [ + 64, + 552, + 288, + 564 + ], + "lines": [ + { + "bbox": [ + 64, + 552, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 64, + 552, + 288, + 564 + ], + "type": "text", + "content": "Table 5. The quantitative results on NYU-V2 [31] " + }, + { + "bbox": [ + 64, + 552, + 288, + 564 + ], + "type": "inline_equation", + "content": "(\\times 4,\\times 16)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 573, + 280, + 630 + ], + "lines": [ + { + "bbox": [ + 71, + 573, + 280, + 630 + ], + "spans": [ + { + "bbox": [ + 71, + 573, + 280, + 630 + ], + "type": "table", + "html": "
SettingsMetricsMambaIRTSP-Mamba
×4 (×16)PSNR47.25(37.03)48.21(37.95)
SSIM0.994(0.967)0.995(0.972)
RMSE2.339(7.895)2.145(7.142)
", + "image_path": "fd2cc6d4a3b4c2f70f34edb1d6a5c2f917748c4482d730603b9b2b37a7e85cd2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 713 + ], + "type": "text", + "content": "Limitations. Although our proposed TSP-Mamba demonstrates promising and significant improvements across several single-image super-resolution tasks, it remains challenging to reliably obtain accurate distance measurements for degraded images, particularly in the presence of high-frequency degradations. Noise components can adversely affect the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 243, + 398, + 305 + ], + "blocks": [ + { + "bbox": [ + 318, + 243, + 398, + 305 + ], + "lines": [ + { + "bbox": [ + 318, + 243, + 398, + 305 + ], + "spans": [ + { + "bbox": [ + 318, + 243, + 398, + 305 + ], + "type": "image", + "image_path": "af97791f8b819af208afca2962c69b89b403f059287142cc566fc762f256ae35.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 306, + 386, + 312 + ], + "lines": [ + { + "bbox": [ + 333, + 306, + 386, + 312 + ], + "spans": [ + { + "bbox": [ + 333, + 306, + 386, + 312 + ], + "type": "text", + "content": "PSNR/SSIM/RMSE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 399, + 244, + 480, + 305 + ], + "blocks": [ + { + "bbox": [ + 399, + 244, + 480, + 305 + ], + "lines": [ + { + "bbox": [ + 399, + 244, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 399, + 244, + 480, + 305 + ], + "type": "image", + "image_path": "6e4c214db5bd53d74e17d402c067a02865771591b0e9ddb75a8a62cf4131dfe5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 306, + 466, + 312 + ], + "lines": [ + { + "bbox": [ + 411, + 306, + 466, + 312 + ], + "spans": [ + { + "bbox": [ + 411, + 306, + 466, + 312 + ], + "type": "text", + "content": "37.45/0.9698/0.0330" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 481, + 244, + 562, + 305 + ], + "blocks": [ + { + "bbox": [ + 481, + 244, + 562, + 305 + ], + "lines": [ + { + "bbox": [ + 481, + 244, + 562, + 305 + ], + "spans": [ + { + "bbox": [ + 481, + 244, + 562, + 305 + ], + "type": "image", + "image_path": "229faf48ef40d423b13934a6c600e935a67fc9937affb2b77117907d84867f6a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 306, + 550, + 312 + ], + "lines": [ + { + "bbox": [ + 495, + 306, + 550, + 312 + ], + "spans": [ + { + "bbox": [ + 495, + 306, + 550, + 312 + ], + "type": "text", + "content": "38.79/0.9750/0.0282" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 313, + 398, + 365 + ], + "blocks": [ + { + "bbox": [ + 318, + 313, + 398, + 365 + ], + "lines": [ + { + "bbox": [ + 318, + 313, + 398, + 365 + ], + "spans": [ + { + "bbox": [ + 318, + 313, + 398, + 365 + ], + "type": "image", + "image_path": "d5b7734f1e4a9fff80a96060cda4db81977001cb9f9cdc95e26365d21138db7b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 366, + 386, + 373 + ], + "lines": [ + { + "bbox": [ + 333, + 366, + 386, + 373 + ], + "spans": [ + { + "bbox": [ + 333, + 366, + 386, + 373 + ], + "type": "text", + "content": "PSNR/SSIM/RMSE" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 399, + 313, + 480, + 365 + ], + "blocks": [ + { + "bbox": [ + 399, + 313, + 480, + 365 + ], + "lines": [ + { + "bbox": [ + 399, + 313, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 399, + 313, + 480, + 365 + ], + "type": "image", + "image_path": "85c542c92fe9da51ae09ec98329d9991ced9dae65d9775b22bb5f76cadcc7312.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 366, + 466, + 373 + ], + "lines": [ + { + "bbox": [ + 411, + 366, + 466, + 373 + ], + "spans": [ + { + "bbox": [ + 411, + 366, + 466, + 373 + ], + "type": "text", + "content": "34.80/0.9672/0.0673" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 481, + 313, + 561, + 365 + ], + "blocks": [ + { + "bbox": [ + 481, + 313, + 561, + 365 + ], + "lines": [ + { + "bbox": [ + 481, + 313, + 561, + 365 + ], + "spans": [ + { + "bbox": [ + 481, + 313, + 561, + 365 + ], + "type": "image", + "image_path": "6a9a0ccec9c600fb1fb4290670466b06da9717a147006ff4d8bf9fac59abb18e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 366, + 549, + 373 + ], + "lines": [ + { + "bbox": [ + 495, + 366, + 549, + 373 + ], + "spans": [ + { + "bbox": [ + 495, + 366, + 549, + 373 + ], + "type": "text", + "content": "36.45/0.9742/0.0555" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 382, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 415 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison between our baseline and the proposed TSP-Mamba for unseen depth-SR maps. We also show normalized error maps in the left corner for better visual comparison" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "type": "text", + "content": "calculation of pair-wise distances between degraded pixels and the optimal scan path searching. In future work, we plan to address the issue of robust distance calculation under various image degradations, such as pyramid representation to build robust 1D TSP scan paths hierarchically." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 522, + 556, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 643 + ], + "type": "text", + "content": "In this work, we introduce a novel locally adaptive propagation method for lightweight SISR. To this end, we develop an efficient scan path search strategy to enhance local information within the globally Mamba-based architecture. Extensive experiments validate the effectiveness of our approach, with TSP-Mamba achieving state-of-the-art performance across multiple SISR benchmarks. Furthermore, TSP-Mamba demonstrates superior generalization on unseen data (such as depth-map SR), and scalability across both ultralightweight and larger model configurations." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 658, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work is partially supported by Shenzhen Science and Technology Program KQTD20210811090149095 and also the Pearl River Talent Recruitment Program 2019QN01X226." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28141" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "type": "text", + "content": "[2] Jiesong Bai, Yuhao Yin, and Qiyuan He. Retinexmamba: Retinex-based mamba for low-light image enhancement. arXiv preprint arXiv:2405.03349, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 296, + 213 + ], + "type": "text", + "content": "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 296, + 270 + ], + "type": "text", + "content": "[4] Ke Chen, Liangyan Li, Huan Liu, Yunzhe Li, Congling Tang, and Jun Chen. Swinfsr: Stereo image super-resolution using swinir and frequency domain knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1764-1774, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 273, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 296, + 327 + ], + "type": "text", + "content": "[5] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 296, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 382 + ], + "type": "text", + "content": "[6] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12312-12321, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 385, + 296, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 296, + 429 + ], + "type": "text", + "content": "[7] Haram Choi, Jeongmin Lee, and Jihoon Yang. N-gram in swim transformers for efficient lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2071–2081, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 430, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 296, + 485 + ], + "type": "text", + "content": "[8] Tao Dai, Jianping Wang, Hang Guo, Jinmin Li, Jinbao Wang, and Zexuan Zhu. Freqformer: frequency-aware transformer for lightweight image super-resolution. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 731-739, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 529 + ], + "type": "text", + "content": "[9] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 532, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 532, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 532, + 296, + 564 + ], + "type": "text", + "content": "[10] Eugene C Freuder and Alan K Mackworth. Constraint satisfaction: An emerging paradigm. In Foundations of artificial intelligence, pages 13-27. Elsevier, 2006. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 567, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 296, + 609 + ], + "type": "text", + "content": "[11] Guanyiman Fu, Fengchao Xiong, Jianfeng Lu, and Jun Zhou. Ssumamba: Spatial-spectral selective state space model for hyperspectral image denoising. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 612, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 296, + 654 + ], + "type": "text", + "content": "[12] Hu Gao, Bowen Ma, Ying Zhang, Jingfan Yang, Jing Yang, and Depeng Dang. Learning enriched features via selective state spaces model for efficient image deblurring. In ACM Multimedia 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 657, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 296, + 689 + ], + "type": "text", + "content": "[13] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[14] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "text", + "content": "image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. 1, 3, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "type": "text", + "content": "[15] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 196 + ], + "type": "text", + "content": "[16] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 252 + ], + "type": "text", + "content": "[17] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3224-3232, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 285 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 288, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 554, + 342 + ], + "type": "text", + "content": "[19] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 766-776, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 344, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 555, + 398 + ], + "type": "text", + "content": "[20] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. Lapar: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. Advances in Neural Information Processing Systems, 33:20343-20355, 2020. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "text", + "content": "[21] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 555, + 499 + ], + "type": "text", + "content": "[22] Xiang Li, Jiangxin Dong, Jinhui Tang, and Jinshan Pan. Dlgstanet: Lightweight dynamic local and global self-attention networks for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12792–12801, 2023. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 501, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 555, + 555 + ], + "type": "text", + "content": "[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 557, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 555, + 612 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 645 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "type": "text", + "content": "[26] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image superresolution with lattice block. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 272-289. Springer, 2020. 2, 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28142" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[27] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 184 + ], + "type": "text", + "content": "[28] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multi-media Tools and Applications, 76:21811-21838, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "text", + "content": "[29] Dongbo Min, Sunghwan Choi, Jiangbo Lu, Bumsub Ham, Kwanghoon Sohn, and Minh N Do. Fast global image smoothing based on weighted least squares. IEEE Transactions on Image Processing, 23(12):5638-5653, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 296, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 263 + ], + "type": "text", + "content": "[30] Yaniv Romano, John Isidoro, and Peyman Milanfar. Raisr: Rapid and accurate image super resolution. IEEE Transactions on Computational Imaging, 3(1):110-125, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 296, + 330 + ], + "type": "text", + "content": "[31] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 387 + ], + "type": "text", + "content": "[32] Yujin Tang, Peijie Dong, Zhenheng Tang, Xiaowen Chu, and Junwei Liang. Vmrnn: Integrating vision mamba and lstm for efficient and accurate spatiotemporal forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5663-5673, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 442 + ], + "type": "text", + "content": "[33] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Jinfan Liu. Omni aggregation networks for lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22378-22387, 2023. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 487 + ], + "type": "text", + "content": "[34] Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. arXiv preprint arXiv:2407.21773, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 296, + 544 + ], + "type": "text", + "content": "[35] Jinsu Yoo, Taehoon Kim, Sihaeng Lee, Seung Hwan Kim, Honglak Lee, and Tae Hyun Kim. Enriched cnn-transformer feature aggregation networks for super-resolution. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4956-4965, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 296, + 601 + ], + "type": "text", + "content": "[36] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 666 + ], + "type": "text", + "content": "[37] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2024. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[38] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. In European Conference on Computer Vision, pages 483-500. Springer, 2024. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[39] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12780–12791, 2023. 2, 5" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28143" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_content_list.json b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2e1d8e2fd2411d123333445ffa4665d06616349f --- /dev/null +++ b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_content_list.json @@ -0,0 +1,1529 @@ +[ + { + "type": "text", + "text": "TacoDepth: Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion", + "text_level": 1, + "bbox": [ + 233, + 130, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiran Wang $^{1,2,*}$ Jiaqi Li $^{2,*}$ Chaoyi Hong $^{1,2}$ Ruibo Li $^{1}$ \nLiusheng Sun $^{3}$ Xiao Song $^{3}$ Zhe Wang $^{3}$ Zhiguo Cao $^{2}$ Guosheng Lin $^{1,\\dagger}$ $^{1}$ S-Lab, Nanyang Technological University", + "bbox": [ + 200, + 202, + 803, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of AIA, Huazhong University of Science and Technology $^{3}$ SenseTime Research", + "bbox": [ + 138, + 256, + 859, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangyiran,zgcao}@hust.edu.cn, gslin@ntu.edu.sg", + "bbox": [ + 290, + 276, + 705, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/RaymondWang987/TacoDepth", + "bbox": [ + 303, + 294, + 687, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 335, + 326, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Radar-Camera depth estimation aims to predict dense and accurate metric depth by fusing input images and Radar data. Model efficiency is crucial for this task in pursuit of real-time processing on autonomous vehicles and robotic platforms. However, due to the sparsity of Radar returns, the prevailing methods adopt multi-stage frameworks with intermediate quasi-dense depth, which are time-consuming and not robust. To address these challenges, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation model with one-stage fusion. Specifically, the graph-based Radar structure extractor and the pyramid-based Radar fusion module are designed to capture and integrate the graph structures of Radar point clouds, delivering superior model efficiency and robustness without relying on the intermediate depth results. Moreover, TacoDepth can be flexible for different inference modes, providing a better balance of speed and accuracy. Extensive experiments are conducted to demonstrate the efficacy of our method. Compared with the previous state-of-the-art approach, TacoDepth improves depth accuracy and processing speed by $12.8\\%$ and $91.8\\%$ . Our work provides a new perspective on efficient Radar-Camera depth estimation.", + "bbox": [ + 89, + 368, + 483, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Radar-Camera depth estimation [6, 12, 20, 21, 25, 26, 28-31, 37] is a vital task for 3D perception and motion planning in autonomous driving [4, 15, 22, 49, 53] and robotics [13, 21, 39]. The task aims to predict dense metric depth by integrating images with mmWave Radar data, perceiving metric scale in different environments. As a typical range sensor on driving and robotic platforms, Radar is widely equipped and used for its high affordability and low power consump", + "bbox": [ + 89, + 755, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ea893d73a9f562e8ff7a9996d340f22ad41f96b1382fcc6f7a46389e4b8e5f5.jpg", + "image_caption": [ + "Figure 1. Performance and efficiency. Circle area indicates inference time (ms). Smaller circles showcase faster speed. The X-axis and Y-axis represent MAE and RMSE metrics for depth errors on the nuScenes [4] dataset. Lower MAE and RMSE mean higher accuracy. Our TacoDepth outperforms prior arts by large margins." + ], + "image_footnote": [], + "bbox": [ + 517, + 335, + 916, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion. It is valuable to explore Radar-Camera fusion in various tasks [2, 5, 9, 16, 32, 37] for its superior all-weather reliability. Besides, for real-time processing on vehicles and robotics, improving efficiency is crucial for fusion models.", + "bbox": [ + 511, + 625, + 906, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The obstacle for fusing Radar in depth estimation lies in its sparsity and noises. Due to the large beam width, Radar point clouds are $1000\\mathrm{x}$ sparser with higher noise levels than LiDAR returns. To deal with the issue, the prevailing methods [21, 26, 28, 29, 31, 37, 38, 40] adopt complex multi-stage frameworks. Intermediate quasi-dense depth is predicted to bridge the gap between sparse Radar and final dense depth. For instance, two-stage methods [26, 28, 29, 31, 37, 38, 40] predict final depth based on intermediate results, as independent models without utilizing other depth predictors [3, 34, 35, 51]. RadarCam-Depth [21], as a plugin module, converts relative depth [34, 35] to metric scale in four stages by post-processing. These multi-stage methods are inefficient and time-consuming, as presented in Fig. 1.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution, † Corresponding author.", + "bbox": [ + 112, + 886, + 354, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10523", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meanwhile, despite being described as quasi-dense, the intermediate depth [21, 26, 28, 29, 31, 37, 40] remains sparse and noisy. As shown in Fig. 2, only few pixels exhibit valid depth. For some adverse lighting or weather conditions, e.g., glaring and nighttime scenarios, even no pixels are predicted with depth values. These defective intermediate results lead to disrupted structures, blurred details, or noticeable artifacts in final predictions, as shown in Fig. 4. With these observations, a question naturally arises: can models directly predict accurate and dense depth without the intermediate results, achieving one-stage Radar-Camera depth estimation with higher efficiency and robustness?", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer the above question, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation framework with one-stage fusion. To simplify multi-stage fusion [21, 26, 28, 29, 31, 37, 38, 40] into one stage, TacoDepth needs to extract and integrate Radar information more effectively. Prior arts simply extract features from Radar coordinates [21, 37] or projected depth [26, 31], sensitive to noises and outliers [11]. In contrast, we utilize a graph-based Radar structure extractor to capture geometric structures and topologies of point clouds, improving robustness to inaccurate points [11, 18, 36, 43]. After that, our pyramid-based Radar fusion module integrates image and Radar features hierarchically. Shallow layers capture image details and Radar coordinates, while deep layers fuse scene semantics and structures [10, 14, 19, 41, 45, 50]. In each layer, Radar-centered flash attention mechanism is applied to establish the cross-modal correlations. For one Radar point, lightweight and accelerated flash attention [8] is calculated within Radar-centered areas for efficiency. Finally, TacoDepth predicts accurate and dense depth by a decoder.", + "bbox": [ + 89, + 275, + 483, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we further analyze the characteristics of different inference strategies in prior arts [21, 26, 28, 29, 31, 37, 38, 40]. Independent models [26, 28, 29, 31, 37, 40] are more efficient without initial depth prediction and postprocessing, while the plug-in model [21] can be more accurate by leveraging cutting-edge relative depth predictors [3, 23, 34, 51]. Therefore, TacoDepth aims to be flexible for the two settings. An auxiliary and optional branch is added for initial depth, allowing TacoDepth to support both independent and plug-in processing. Our approach can better balance depth accuracy and model efficiency.", + "bbox": [ + 89, + 580, + 483, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct evaluations on nuScenes [4] and ZJU-4DRadarCam [21] datasets with different depth ranges, following prior arts [21, 31, 37]. As shown in Fig. 1, whether compared with independent [26, 28, 31, 37, 40] or plugin in [21] models, TacoDepth significantly outperforms prior arts [21, 26, 28, 31, 37, 40] in accuracy and efficiency. Besides, our approach achieves real-time processing of over 37 fps as an independent model and strong compatibility with various depth predictors [3, 34, 35, 51] as a plug-in model, further proving the efficiency and flexibility of TacoDepth.", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1e7180a4707dc7ac431aa404e436ae2c9b86da99bc267c0cfa0a28614cba35cc.jpg", + "image_caption": [ + "Figure 2. Intermediate quasi-dense depth maps [21, 26, 28, 29, 31, 37, 40] remain sparse and noisy. Pixels with valid depth values are visualized by the gray areas in the red rectangular boxes." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 903, + 169 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 511, + 234, + 874, + 248 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose TacoDepth, an efficient and accurate Radar-Camera depth estimation method with one-stage fusion.", + "- We design the pyramid-based Radar fusion module, fusing Radar and image features effectively and efficiently.", + "- TacoDepth is flexible for both independent and plug-in processing, balancing model efficiency and performance." + ], + "bbox": [ + 513, + 253, + 903, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 359, + 653, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MmWave Radar for Vehicles and Robotics. Autonomous driving [4, 15, 48, 53] and robotics [13, 21, 24, 39] rely on multiple sensors to perceive the environment, e.g., camera, LiDAR, and mmWave Radar. LiDAR and Radar provide the metric scale of different scenarios as a complement to cameras. Compared with LiDAR, Radar is cheaper and more widely-equipped with lower power consumption. The reflection properties of mmWave also provide Radar with farther sensing ranges and stronger robustness to all-weather conditions. The main drawback of Radar lies in its sparsity and noise. Due to the large beam width, Radar point clouds are $1000\\mathrm{x}$ sparser [37] than LiDAR. Besides, Radar captures coordinates and depth values with lower accuracy. For instance, 3D Radar suffers from height ambiguity [31] due to insufficient antenna elements along the elevation axis, while this problem can be alleviated by the emerging 4D Radar. Thus, integrating Radar data is valuable for various tasks [2, 5, 9, 16, 27, 32, 37] in driving and robotic systems.", + "bbox": [ + 511, + 386, + 906, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Radar-Camera Depth Estimation. The task aims to predict dense metric depth by merging images and Radar, facilitating 3D perception and motion planning of vehicles and robotics. Model efficiency is crucial for these real-time applications. However, to overcome the sparsity and noise of Radar, prior arts [21, 26, 28, 29, 31, 37, 38, 40] adopt inefficient multi-stage frameworks. The final depth is predicted based on intermediate quasi-dense depth. These methods can be divided into independent and plug-in models. Most independent models [26, 28, 29, 31, 37, 40] predict final depth in two stages without relying on other depth predictors [17, 33-35, 46, 47]. For example, Lin et al. [26] and DORN [28] use two CNNs [14] for quasi-dense and final depth. RC-PDA [31] and RC-DPT [29] convert intermediate results into multiple channels for varied confidence levels. Singh et al. [37] propose gated fusion to filter the inac", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10524", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7af03cc5a3f094eef113e03d9d84f96b8eeaef12d4f649907d8ef409a66c004b.jpg", + "image_caption": [ + "Figure 3. Overview of the TacoDepth. The graph-based Radar structure extractor captures graph structures of Radar point clouds through the node feature $N_{l}$ and edge feature $E_{l}$ in a certain layer $l$ . The pyramid-based Radar fusion module integrates image and Radar features in a pyramidal hierarchical manner. In each layer, to efficiently build cross-modal correspondences, the Radar-centered flash attention is calculated within Radar-centered areas based on horizontal coordinates, e.g., pixels in the orange area as queries, while Radar points in the red area as keys and values. Our TacoDepth achieves efficient and accurate Radar-Camera depth estimation in one stage. During inference, the model is flexible and supports both independent and plug-in processing, facilitating a better balance of model efficiency and accuracy." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 620, + 354 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8ece566f18f36fa5fc0ecaf50fc399d52b01522b368a268e9d4d6ec4fe6fae46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 88, + 883, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "curate quasi-dense depth. Besides, RadarCam-Depth [21] designs a plug-in module, aligning relative depth [34, 35] to metric scale in four stages, including initial depth, global scale alignment, quasi-dense depth, and scale learner. These methods [21, 26, 28, 29, 31, 37, 38, 40] are inefficient with complex stages. Defective intermediate depth also limits their robustness. In contrast, we propose TacoDepth model to achieve efficient, accurate, and flexible one-stage fusion.", + "bbox": [ + 88, + 467, + 482, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. TacoDepth", + "text_level": 1, + "bbox": [ + 89, + 601, + 207, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a detailed illustration of TacoDepth, an efficient Radar-Camera depth estimation framework with one-stage fusion. In Sec. 3.1, we provide an overall description to outline our approach. In Sec. 3.2, we elaborate on the detailed architecture of the pyramid-based Radar fusion module. TacoDepth can be flexible for independent and plug-in processing, which will be discussed in Sec. 3.3. Besides, the training procedure for TacoDepth is illustrated in Sec. 3.4.", + "bbox": [ + 88, + 627, + 482, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 757, + 200, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we present an overview of our TacoDepth framework. Given a single image $I \\in \\mathcal{R}^{H \\times W \\times 3}$ and a Radar point cloud $P \\in \\mathcal{R}^{K \\times 3}$ as inputs, our goal is to predict dense metric depth $D \\in \\mathcal{R}^{H \\times W}$ . $H$ and $W$ represent image height and weight. The point cloud $P$ contains $K$ points $p \\in \\mathcal{R}^3$ . Radar-Camera depth estimation can be divided into two primary subtasks: the extraction and the integration of Radar information. Accordingly, as depicted in Fig. 3,", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TacoDepth comprises a graph-based Radar structure extractor and a pyramid-based Radar fusion module to undertake these two important steps respectively.", + "bbox": [ + 511, + 467, + 906, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, for Radar feature extraction, previous methods [21, 37] simply extract features from the coordinates of individual Radar points, focusing solely on the point-level information. For example, Singh et al. [37] extract point features in 32,256 dimensions from the three-dimensional coordinates using a multi-layer perception (MLP), which inevitably introduces substantial redundancy and noise.", + "bbox": [ + 511, + 515, + 906, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve efficient Radar-Camera fusion in a single stage, TacoDepth needs to extract more comprehensive and effective Radar information, prompting us to reconsider the complementarity between camera images and Radar data. Compared to the coordinates, the geometric structures of point clouds, e.g., the distances or relations between Radar points, are potentially more informative for Radar-Camera fusion. The overall topologies are also more robust to inaccurate outliers [11]. To this end, we employ the graph-based Radar structure extractor with a lightweight GNN [11, 36, 43] architecture. Specifically, Radar points are regarded as the graph nodes. The adjacency matrix of all points represents the graph edges. In each layer, the node features of Radar points can be updated and aggregated along edges by PCA-GM [44]. The graph-based Radar structure extractor can capture point coordinates and graph structures from shallow layers to deep layers [11, 41, 45].", + "bbox": [ + 511, + 625, + 908, + 882 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After extracting the Radar structures, the subsequent step", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10525", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is to merge RGB and Radar features. Prior arts [21, 26, 28, 29, 31, 37, 38, 40] adopt multi-stage methods with intermediate quasi-dense depth, which are inefficient and not robust as illustrated in Sec. 1. In contrast, we propose our pyramid-based Radar fusion module for efficient one-stage fusion.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To be specific, features of the two modalities are integrated in a pyramidal hierarchical manner. Shallow layers can merge image details with Radar coordinates, whereas deeper layers could fuse scene semantics and geometric structures [10, 14, 19, 41, 45, 50]. In each layer, we leverage Radar-centered flash attention for cross-modal correlations and feature fusion. The queries are derived from image pixels, while keys and values originate from Radar points. For each point, we delineate Radar-centered areas based on the horizontal coordinates, as exemplified by the orange and red areas in Fig. 3, given that the projected horizontal coordinates of Radar points are generally more precise than the height dimension [31, 37]. Lightweight flash attention [8] is calculated within these areas to maintain efficiency, since external pixels and points can not be correlated. Finally, with the fused features from all layers, TacoDepth predicts accurate and dense metric depth by a depth decoder [37].", + "bbox": [ + 89, + 167, + 482, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, TacoDepth is adaptable for both independent and plug-in inference. An optional input branch is introduced to handle the initial relative depth. For the independent inference, our model achieves real-time processing of over 37 fps. For the plug-in manner, TacoDepth demonstrates compatibility with various cutting-edge depth predictors [3, 34, 35, 51], as proved in Sec. 4.5. Consequently, our approach can deliver a balanced trade-off between performance and efficiency for different users and applications.", + "bbox": [ + 89, + 426, + 482, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pyramid-based Radar Fusion", + "text_level": 1, + "bbox": [ + 89, + 574, + 354, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the Radar point cloud $P$ , our graph-based Radar structure extractor derives node feature $N_{l}$ and edge feature $E_{l}$ at a certain layer $l \\in \\{1,2,\\dots ,L\\}$ , where $L$ represents the layer number of the GNN module [11, 36, 43]. For the input image $I$ , the image encoder [14, 21, 37] extracts multi-scale features $\\{F_1,F_2,\\dots ,F_{2L}\\}$ . The pyramid-based Radar fusion module hierarchically integrates the multi-modal features. $N_{l}$ and $E_{l}$ will be fused with image features in adjacent layers. We merge the node feature $N_{l}$ with image feature $F_{2l - 1}$ , and the edge feature $E_{l}$ with $F_{2l}$ , respectively.", + "bbox": [ + 89, + 597, + 482, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The prerequisite for feature fusion is cross-modal correspondences between Radar points and RGB pixels. Due to the noisy elevation coordinates of 3D Radar, prior arts [21, 29, 31, 37, 40] try to build one-to-many mappings by the intermediate quasi-dense depth. They adopt explicit classification supervision with the depth discrepancies between Li-DAR and Radar as pseudo-labels. However, the quasi-dense depth prediction leads to inefficiency and erroneous final outcomes, as discussed in Sec. 1. Given the absence of reliable explicit labels and supervision, we hope our TacoDepth", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "can learn and establish the correspondences implicitly.", + "bbox": [ + 511, + 90, + 872, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, for efficient Radar-Camera correlations, we propose the Radar-centered flash attention in TacoDepth. We define Radar-centered areas based on the precise horizontal coordinates of Radar points and image pixels. For each point, the attention is computed only with image pixels in the area, since external pixels can not be correlated. Here we take $E_{l} \\in \\mathcal{R}^{K \\times K}$ and $F_{2l} \\in \\mathcal{R}^{h \\times w \\times C_l}$ as examples, where $h$ and $w$ are feature height and width. $K$ and $C_l$ refer to the point number and layer feature channel. For a Radar point $p$ with the horizontal coordinate $x_{p}$ , as showcased by the exemplary orange area in Fig. 3, we only retain image pixels $m$ whose coordinates $x_{m}$ are close to $x_{p}$ :", + "bbox": [ + 511, + 107, + 903, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {2 l} = F _ {2 l} \\left[ x _ {p} - a _ {l} < x _ {m} < x _ {p} + a _ {l} \\right], \\hat {F} _ {2 l} \\in \\mathcal {R} ^ {\\hat {M} \\times C _ {l}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 292, + 903, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a_{l}$ indicates the width of the Radar-centered area in layer $l$ . The notation $[\\cdot]$ depicts the fetching operation of pixels or points. $\\hat{F}_{2l}$ is the image feature of the remaining $\\hat{M}$ pixels to generate queries. For a random pixel $m$ in $\\hat{F}_{2l}$ , as represented by the red area in Fig. 3, the attention calculation is also restricted to adjacent Radar points:", + "bbox": [ + 511, + 318, + 903, + 409 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} _ {l} = E _ {l} \\left[ x _ {m} - a _ {l} < x _ {p} < x _ {m} + a _ {l} \\right], \\hat {E} _ {l} \\in \\mathcal {R} ^ {\\hat {K} \\times K}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 414, + 903, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The feature dimension maintains the original $K$ to preserve integral graph topologies. $\\hat{E}_l$ is the edge feature of the retained $\\hat{K}$ points for keys and values. In this way, the computational costs can be significantly decreased with the small subset of $\\hat{M}$ pixels and $\\hat{K}$ points. For the pixel $m$ , we generate queries from $\\hat{F}_{2l}[m] \\in \\mathcal{R}^{C_l}$ , while keys and values are derived from $\\hat{E}_l$ . Radar-centered attention is computed as:", + "bbox": [ + 511, + 439, + 903, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {2 l} ^ {\\prime} [ m ] = \\operatorname {s o f t m a x} \\frac {W _ {q} \\hat {F} _ {2 l} [ m ] \\left(W _ {k} \\hat {E} _ {l}\\right) ^ {T}}{\\sqrt {C _ {l}}} W _ {v} \\hat {E} _ {l}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 549, + 903, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which $W_{q}$ , $W_{k}$ , and $W_{v}$ are learnable linear projections. The fused Radar-Camera feature $F_{2l}^{\\prime} \\in \\mathcal{R}^{h \\times w \\times C_l}$ can be acquired for the current layer. The above calculation is incorporated into attention blocks with MLPs and residual connections. To further improve model efficiency, we employ lightweight and accelerated flash attention modules [7, 8] to replace the original attention computation [42].", + "bbox": [ + 511, + 598, + 903, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, the node feature $N_{l}$ can be merged with the image feature $F_{2l-1}$ by slightly adjusting the dimensions of linear projections. Ultimately, the acquired Radar-Camera features $\\{F_1', F_2', \\dots, F_{2L}'\\}$ are fed to a common depth decoder [21, 37] to predict accurate metric depth.", + "bbox": [ + 511, + 704, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a consequence, the proposed pyramid-based Radar fusion module enables efficient and effective Radar-Camera fusion in a single stage. The Radar-centered flash attention establishes cross-modal correlations with minimal computational costs. Without relying on the previous multi-stage frameworks and intermediate quasi-dense depth [21, 26, 28, 29, 31, 37, 38, 40], TacoDepth achieves higher efficiency and stronger robustness for Radar-Camera depth estimation.", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10526", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Flexible Inference", + "text_level": 1, + "bbox": [ + 89, + 90, + 266, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the varied inference manners in prior arts [21, 26, 28, 29, 31, 37, 38, 40], the plug-in module [21] could achieve higher accuracy using state-of-the-art relative depth predictors [34, 35, 52], while independent models [26, 28, 29, 31, 37, 38, 40] are capable of attaining faster inference speed without initial depth prediction and post-processing.", + "bbox": [ + 89, + 112, + 483, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, TacoDepth is designed to be flexible, supporting both independent and plug-in predictions. For the plug-in inference, an initial relative depth map is converted to metric scale by fusing camera and Radar information. If we denote the TacoDepth model as $\\mathcal{T}_{\\theta}$ , parameterized by $\\theta$ , the depth prediction process can then be formulated as follows:", + "bbox": [ + 89, + 204, + 483, + 294 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD = \\mathcal {T} _ {\\theta} (I, P \\mid D ^ {*}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 308, + 480, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the vertical bar $(|)$ separates the required and optional inputs. $D$ is the predicted depth of our model. $D^{*}$ is the optional input of initial relative depth from depth predictors [3, 34, 35, 51]. We introduce an auxiliary branch to process and fuse $D^{*}$ with image features, while keeping all other settings identical to the independent mode. Our plugin in manner demonstrates compatibility with various depth predictors, e.g., MiDaS [3, 34], DPT [35], and Depth-Anything-v2 [51]. In this way, TacoDepth can achieve a better balance of model efficiency and accuracy, offering superior flexibility for different users and applications.", + "bbox": [ + 89, + 330, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training the TacoDepth", + "text_level": 1, + "bbox": [ + 89, + 507, + 310, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the training phase, the training data [4, 21] is randomly divided into two equal portions in each epoch. For the auxiliary branch, one half of the data is provided with initial relative depth, while the other half adopts zero as input. This training strategy facilitates the simultaneous optimization of the independent inference and plug-in prediction.", + "bbox": [ + 89, + 529, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for the training loss, we employ the same depth ground truth and $L_{1}$ loss as the prior art [37]. Specifically, LiDAR-captured depth $D_{gt}$ and reprojection-accumulated depth $D_{acc}$ are utilized for supervision [21, 31, 37]. The overall loss function $\\ell_{L_1}$ can be expressed as follows:", + "bbox": [ + 89, + 619, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {L _ {1}} = \\frac {1}{| \\Omega_ {g t} |} \\sum_ {\\Omega_ {g t}} | D - D _ {g t} | + \\frac {\\lambda}{| \\Omega_ {a c c} |} \\sum_ {\\Omega_ {a c c}} | D - D _ {a c c} |, (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 705, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Omega_{gt}$ and $\\Omega_{acc}$ depict the masks of $D_{gt}$ and $D_{acc}$ with valid depth values. $\\lambda$ is the coefficient of the second term.", + "bbox": [ + 89, + 755, + 482, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 800, + 223, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prove the efficacy of our TacoDepth, we conduct experiments on the nuScenes [4] and ZJU-4DRadarCam [21] datasets following prior arts [20, 21, 26, 28, 29, 31, 37, 40]. In Sec. 4.1, we describe the experimental datasets and evaluation protocols. The implementation details of our model", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are illustrated in Sec. 4.2. In Sec. 4.3, we compare the TacoDepth with previous Radar-Camera depth estimation approaches to demonstrate our state-of-the-art performance. The inference speed and model efficiency of different methods are evaluated and discussed in Sec. 4.4. In Sec. 4.5, we perform ablation studies to expound on our specific designs.", + "bbox": [ + 511, + 90, + 905, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 511, + 193, + 812, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the scarcity of public datasets for Radar-Camera depth estimation, prior works [20, 26, 28, 29, 31, 37, 38, 40] only conduct experiments on nuScenes dataset [4] with 3D Radar. Recently, RadarCam-Depth [21] releases their ZJU-4DRadarCam dataset [21] with 4D Radar to alleviate this problem. We adopt these two datasets [4, 21] for experiments, proving our effectiveness on both 3D and 4D Radar. NuScenes Dataset. As an outdoor driving dataset, the nuScenes [4] is collected by vehicles equipped with cameras, 3D Radar, LiDAR, and IMU. It contains 1,000 scenes and around 40,000 synchronized Radar-Camera keyframes captured in Boston and Singapore. We follow the same train-test split as prior arts [21, 26, 29, 31, 37, 40], with 700 scenes for training, 150 for validation, and 150 for testing.", + "bbox": [ + 511, + 215, + 906, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ZJU-4DRadarCam Dataset. The dataset [21] is collected by a ground robot fitted with more advanced camera, 4D Radar, and LiDAR sensors, offering denser Radar data and LiDAR depth than the nuScenes [4] dataset. It comprises 33,409 Radar-Camera keyframes in total. We follow the official data split [21] with 29,312 frames for training and validation, and 4,097 samples for testing in our experiments.", + "bbox": [ + 511, + 428, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocols. We evaluate independent and plug-in models separately for fair comparisons on the two datasets [4, 21] with depth ranges of 50, 70, and 80 meters [21, 31, 37, 40]. On nuScenes [4], previous state-of-the-art methods [12, 20, 21, 26, 28, 31, 37, 40] are compared comprehensively to the best of our knowledge. For the recently published ZJU-4DRadarCam [21], since most prior works do not include experiments on the dataset, we can only follow the same comparative methods [21, 28, 37] and results as RadarCam-Depth [21]. Besides, we also evaluate plug-in models with different depth predictors [3, 35, 51] to showcase model flexibility on the ZJU-4DRadarCam [21].", + "bbox": [ + 511, + 534, + 905, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt the commonly-applied depth metrics MAE, RMSE, iMAE, iRMSE, $\\delta_{1}$ , and Rel as prior arts [21, 26, 31, 37]. See supplement for more details.", + "bbox": [ + 511, + 715, + 903, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 771, + 730, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Processing. We adhere to the data processing procedures of prior arts [21, 37]. For the nuScenes [4], the $D_{acc}$ in Eq. 5 is generated by accumulating the $D_{gt}$ of 160 nearby frames [21, 37]. For the ZJU-4DRadarCam dataset [21], the $D_{acc}$ is obtained via the linear interpolation [1, 21] of $D_{gt}$ . Model Architecture. In line with prior arts [21, 37], the ResNet-18 [14] is adopted as the image encoder. The graph", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10527", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b0138951cc71c5ede6cdc87185b52c3d8c5b18f227274b09d184c56c0ff1ab18.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodRadar FramesImagesTime(ms)↓0 - 50m0 - 70m0 - 80m
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
Indep-endentLin et al. [26] (IROS'20)31117.62034.94316.52294.75338.22371.05623.0
RC-PDA [31] (CVPR'21)5379.22225.04156.53326.16700.63713.67692.8
R4Dyn [12] (3DV'21)41------6434.0
DORN [28] (ICIP'21)5(×3)1392.51926.64124.82380.65252.72467.75554.3
Singh et al. [37] (CVPR'23)1194.21727.73746.82073.24590.72179.34898.7
CaFNet [40] (IROS'24)11103.91674.23674.52010.34493.12109.84765.6
Li et al. [20] (ECCV'24)1167.61524.53567.31822.94303.61927.04609.6
TacoDepth (Ours)1126.71423.63275.81712.63960.51833.44150.2
Plug-inRadarCam-Depth [21] (ICRA'24)11358.31286.12964.31587.93662.51689.73948.0
TacoDepth (Ours)1129.31046.82487.51347.13152.81492.43324.8
", + "bbox": [ + 135, + 88, + 864, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparisons with state-of-the-art methods on the nuScenes dataset [4] (in millimeters). Independent [12, 20, 26, 28, 31, 37, 40] and plug-in [21] models are evaluated separately, with maximum evaluation distances of 50, 70, and 80 meters as prior arts [21, 37]. Plug-in models are compared using the same depth predictor DPT-Hybrid [35] as RadarCam-Depth [21]. We report the average runtime of processing one $900 \\times 1600$ frame and 30 Radar points by different methods on one NVIDIA RTX A6000 GPU. The inference time of plug-in models does not contain initial depth prediction. Some methods [12, 26, 28, 31] require multiple frames and Radar scans for inference, while others [20, 21, 37, 40] only need a single frame as input. The best performance is in boldface. Second best is underlined.", + "bbox": [ + 89, + 276, + 906, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dbaae734e43abccef215ebd53b3aeb0d7c53b5feb682ca5416b2f07576172f44.jpg", + "image_caption": [ + "Figure 4. Visual results of independent models on nuScenes [4]. Both daytime and nighttime samples are presented. Prior arts [31, 37] exhibit disrupted structures and noticeable artifacts. TacoDepth produces accurate depth with finer details and more complete structures." + ], + "image_footnote": [], + "bbox": [ + 133, + 366, + 864, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "based Radar structure extractor contains $L = 3$ layers with the GNN [11, 36] architecture. The widths $a_{l}$ of Radar-centered areas in Eq. 1 and Eq. 2 are set to $\\{48, 32, 16\\}$ for the three layers. Refer to the supplement for more details.", + "bbox": [ + 89, + 672, + 482, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Recipe. The model is trained by Adam optimizer on two NVIDIA A6000 GPUs with a batch size of 12 for 50 epochs. The initial learning rate is $1e^{-4}$ and decreases by $1e^{-5}$ for every ten epochs. The input image size is $900 \\times 1600$ [21, 37] and $300 \\times 1280$ [21] on nuScenes [4] and ZJU-4DRadarCam [21], respectively. The $\\lambda$ in Eq. 5 is set to one. All other settings are consistent with prior works [21, 37].", + "bbox": [ + 89, + 734, + 482, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 848, + 464, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons with Independent Models. Independent models [12, 20, 26, 28, 29, 31, 37, 38, 40] predict dense", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "metric depth without using relative depth predictors [3, 34, 35, 51]. Most of them adopt two-stage methods with intermediate quasi-dense depth, which are inefficient and not robust. Our TacoDepth significantly surpasses these methods in efficiency and accuracy with one-stage fusion. For example, in Table 1, compared with the recent CaFNet [40], our model reduces RMSE, MAE, and inference time by over $10.9\\%$ , $13.1\\%$ , and $74.3\\%$ on nuScenes dataset [4]. Additionally, Li et al. [20] aim to improve the training paradigm under sparse supervision by their Disruption-Compensation framework. Without using such advanced training and supervision strategy, TacoDepth still outperforms Li et al. [20] by over $60.5\\%$ , $4.9\\%$ , and $8.1\\%$ for runtime, MAE, and RMSE, showing the efficacy of our proposed architectures.", + "bbox": [ + 511, + 672, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly, in Table 2, compared with Singh et al. [37],", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10528", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/95ed752ab587a53076e237bf0903093410c0e458cf6d9f0f770bae89856c2ca9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethod0 - 50m0 - 70m0 - 80m
MAE↓RMSE↓iMAE↓iRMSE↓MAE↓RMSE↓iMAE↓iRMSE↓MAE↓RMSE↓iMAE↓iRMSE↓
Indep-endentDORN [28] (ICIP'21)2210.24129.719.831.92402.24625.219.831.92447.64760.019.931.9
Singh et al. [37] (CVPR'23)1785.43704.618.135.31932.74137.118.035.21979.54309.317.935.1
TacoDepth (Ours)1120.12686.712.825.01181.82906.312.724.91201.12990.712.724.9
Plug-inRadarCam-Depth [21] (ICRA'24)1067.52817.410.522.91157.03117.710.422.91183.53229.010.422.8
TacoDepth (Ours)930.22477.39.320.8983.12779.69.320.91032.52850.39.420.9
", + "bbox": [ + 135, + 88, + 864, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparisons with state-of-the-art methods on ZJU-4DRadarCam [21] (in millimeters). Plug-in models are compared using the same relative depth predictor DPT-Hybrid [35]. Since many prior works do not include experiments on this new dataset, we adhere to the same compared methods [21, 28, 37] and metrics as RadarCam-Depth [21]. Best performance is in boldface. Second best is underlined.", + "bbox": [ + 89, + 190, + 906, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e46bdf0b07e2ce89f1b4d1888720194317b6101fe78e2c0d317541fad59cfe01.jpg", + "image_caption": [ + "Figure 5. Visual results of plug-in models [21] on the ZJU-4DRadarCam [21]. The same depth predictor DPT-Hybrid [35] is adopted for fair comparisons. Regions with obvious differences are highlighted in the rectangular boxes. Best view zoomed in on-screen for details." + ], + "image_footnote": [], + "bbox": [ + 133, + 243, + 864, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TacoDepth reduces depth errors by over $37.3\\%$ , $27.5\\%$ , $28.7\\%$ , and $29.1\\%$ for MAE, RMSE, iMAE, and iRMSE on ZJU-4DRadarCam [21]. The superior efficiency and accuracy prove the effectiveness of our one-stage model design.", + "bbox": [ + 88, + 619, + 482, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with Plug-in Models. Plug-in models [21] convert relative depth to metric depth by fusing the RGB and Radar data. Our method supports both independent and plug-in processing by the auxiliary input branch. For the plug-in mode, in Table 1, TacoDepth decreases the MAE and RMSE by over $11.7\\%$ and $13.9\\%$ than RadarCam-Depth [21] on nuScenes [4]. Owing to their complex four-stage framework [21], we even improve the speed by $91.8\\%$ . Similarly, in Table 2, TacoDepth reduces the MAE, RMSE, iMAE, and iRMSE by over $12.8\\%$ , $10.8\\%$ , $9.6\\%$ , and $8.3\\%$ on ZJU-4DaradarCam [21]. The results prove the generality and flexibility of TacoDepth, providing a better balance of model efficiency and accuracy for different users.", + "bbox": [ + 88, + 684, + 482, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To be mentioned, in the plug-in mode, TacoDepth can", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "be compatible with various relative depth predictors [3, 35, 51]. It can benefit from stronger depth models [3, 51] to achieve higher accuracy, which will be presented in Table 4.", + "bbox": [ + 511, + 619, + 906, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparisons. The visual results of independent models [31, 37] are showcased in Fig. 4. Both daytime and nighttime samples are compared. TacoDepth can predict accurate depth maps with more complete structures (e.g., the cars and buses) and more meticulous details (e.g., the pedestrian overpass and metal guardrail). The plug-in models [21] are compared in Fig. 5. Based on the same depth predictor DPT-Hybrid [35], our method produces noticeably better structures and details in the rectangular boxes. The improvements are attributed to our graph-based Radar structure extractor and pyramid-based Radar fusion module, which extract and fuse Radar structures effectively.", + "bbox": [ + 511, + 669, + 908, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Robustness. Multi-stage methods [26, 28, 31, 37, 40] predict quasi-dense depth as an intermediate step. However, as shown in Fig. 2, the intermediate results are still", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10529", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1acc8f52cdcd4886da6f9bce4bd0ca90d719228d7f971d3877c62a175b4b6595.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodParams (M)↓FLOPs (G)↓Time (ms)↓
Indep-endentDORN [28] (ICIP'21)183.38923.47392.5
Singh et al. [37] (CVPR'23)22.81502.0994.2
TacoDepth (Ours)13.47139.3026.7
Plug-inRadarCam-Depth [21] (ICRA'24)33.26619.02358.3
TacoDepth (Ours)14.25139.8729.3
", + "bbox": [ + 94, + 89, + 480, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0137d81d96af557e94b1acac99b67e072ca08b94850bcf4a2d64924a9d1324a4.jpg", + "table_caption": [ + "Table 3. Model efficiency. We report FLOPs, model parameters, and runtime of different methods. Since varied depth predictors can be applied, the results of plug-in models do not contain initial depth prediction. FLOPs and runtime are evaluated with one $900 \\times 1600$ image and 30 Radar points on one NVIDIA A6000 GPU." + ], + "table_footnote": [], + "table_body": "
InitialRadarCam-Depth [21]TacoDepth (Ours)
δ1↑Rel↓MAE↓δ1↑Rel↓MAE↓δ1↑Rel↓MAE↓
DPT-Hybrid [35]0.9030.092-0.9210.0871157.00.9320.076983.1
MiDaS-v3.1 [3]0.9170.081-0.9260.0861177.20.9440.071852.4
Depth-Anything-v2 [51]0.9320.069-0.9400.073924.70.9610.059730.2
", + "bbox": [ + 94, + 252, + 480, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sparse and noisy, especially on challenging nighttime and glaring scenes, leading to disrupted structures, blurred details, and obvious artifacts in their final depth predictions, as presented in Fig. 4. In contrast, we propose the one-stage TacoDepth without the quasi-dense depth. By effectively integrating the complementary image details and Radar structures, our model produces robust depth for both daytime and nighttime scenarios in Fig. 4. See supplement for more quantitative and visual results on nighttime scenes.", + "bbox": [ + 88, + 381, + 482, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Model Efficiency", + "text_level": 1, + "bbox": [ + 89, + 527, + 256, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the processing time of different methods in Table 1. Compared with the second-best independent model of Li et al. [20] and the plug-in model RadarCam-Depth [21], TacoDepth improves inference speed by $60.5\\%$ and $91.8\\%$ .", + "bbox": [ + 89, + 549, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further showcase our efficiency, we report model parameters and FLOPs in Table 3. Compared with the two-stage Singh et al. [37], the one-stage TacoDepth reduces the parameters and FLOPs by $40.9\\%$ and $72.2\\%$ . Compared with the four-stage plug-in model RadarCam-Depth [21], we decrease parameters and FLOPs by $57.1\\%$ and $77.4\\%$ .", + "bbox": [ + 89, + 611, + 482, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that, the differences between the plug-in and independent modes of our method arise from the auxiliary branch to fuse initial relative depth. TacoDepth achieves real-time processing of 37.5 fps by independent inference.", + "bbox": [ + 89, + 700, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 771, + 254, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Plug-in Inference. In Table 4, our TacoDepth outperforms the RadarCam-Depth [21] with three different relative depth predictors [3, 35, 51]. Using more advanced depth predictors, e.g., MiDaS-v3.1 [3] and Depth-Anything-v2 [51], can further boost our accuracy without extra effort. The results demonstrate the versatility and flexibility of our approach. Efficacy of TacoDepth. In Table 5, we ablate the graph", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/680ddb6326fdff00ac0a511809c0d8f784355409019629fa0bf2407b04b39111.jpg", + "table_caption": [ + "Table 4. Plug-in models with varied depth predictors. Our method outperforms RadarCam-Depth [21] with different depth predictors [3, 35, 51] on ZJU-4DRadarCam [21] (0-70 meters)." + ], + "table_footnote": [], + "table_body": "
MethodRGBRadarGEPFMAE↓RMSE↓
RGB Baseline✓✗✗✗2474.35402.1
Singh et al. [37] (CVPR'23)✓✓✗✗2073.24590.7
TacoDepth (w/o GE)✓✓✗✓1815.64189.8
TacoDepth (Ours)✓✓✓✓1712.63960.5
", + "bbox": [ + 517, + 90, + 903, + 161 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c1a82b7127a5e76c32c8c230479175c3fcf1812f1243fb4a5ec971c5b1ea0bf4.jpg", + "table_caption": [ + "Table 5. Efficacy of the TacoDepth. We ablate the effectiveness of graph-based Radar structure extractor (GE) and pyramid-based Radar fusion (PF) for Radar structure extraction and integration. We adopt the same baseline as Singh et al. [37]. TacoDepth $(w / o$ GE) uses the same MLP to extract Radar point features as prior art [37]. The results are evaluated on nuScenes [4] in 0-70 meters." + ], + "table_footnote": [], + "table_body": "
MetricL=1L=2L=3L=4
MAE↓1840.61787.81712.61729.3
RMSE↓4203.74046.93960.53984.2
", + "bbox": [ + 568, + 255, + 849, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Layer number L. Results are on nuScenes [4] in $0 - 70\\mathrm{m}$", + "bbox": [ + 513, + 309, + 903, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "based Radar structure extractor (GE) and pyramid-based Radar fusion module (PF). With the same MLP to extract Radar point features [37], TacoDepth $(w / o$ GE) reduces MAE and RMSE by $12.4\\%$ and $8.7\\%$ than Singh et al. [37], which are attributed to the PF. Replacing the MLP [37] with GE further decreases MAE and RMSE by $5.7\\%$ and $5.5\\%$ .", + "bbox": [ + 511, + 343, + 903, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We adopt the RGB baseline of Singh et al. [37], a common single-image depth U-Net. Overall, TacoDepth improves accuracy beyond the baseline by $30.8\\%$ and $26.7\\%$ (MAE and RMSE), proving the efficacy of our approach.", + "bbox": [ + 511, + 434, + 903, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Layer Number $L$ . In Table 6, we ablate the layer number $L$ of GE and PF. Compared with shallow layers $(L = 1,2)$ , deeper layers $(L = 3)$ better capture structures of point clouds [41, 45], producing higher accuracy. Due to Radar sparsity, more layers $(L = 4)$ does not yield improvements but decreases efficiency. We adopt $L = 3$ in experiments.", + "bbox": [ + 511, + 494, + 903, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 601, + 633, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation model with one-stage fusion. Different from the multi-stage methods with quasidense depth, TacoDepth utilizes the graph-based Radar structure extractor and pyramid-based Radar fusion module to capture and integrate Radar features, achieving superior efficiency and robustness. Moreover, TacoDepth is flexible for independent and plug-in processing, offering a better balance of speed and accuracy. Our work provides a new perspective on the efficient Radar-Camera depth estimation. Limitations. We currently provide one efficient implementation of our TacoDepth. In future work, more techniques can be explored for varied performances and applications.", + "bbox": [ + 511, + 627, + 906, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. The study is supported under RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 1 grant (RG14/22).", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10530", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] C Bradford Barber, David P Dobkin, and Hannu Huhdanpaa. The quickhull algorithm for convex hulls. ACM Transactions on Mathematical Software (TOMS), 22(4):469-483, 1996. 5", + "[2] Oded Bialer and Yuval Haitman. Radsimreal: Bridging the gap between synthetic and real data in radar object detection with simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15407-15416, 2024. 1, 2", + "[3] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 1, 2, 4, 5, 6, 7, 8", + "[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11618-11628, 2020. 1, 2, 5, 6, 7, 8", + "[5] Yujeong Chae, Hyeonseong Kim, and Kuk-Jin Yoon. Towards robust 3d object detection with lidar and 4d radar fusion in various weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15162-15172, 2024. 1, 2", + "[6] Feipeng Chen, Lihui Wang, Pinzhang Zhao, and Jian Wei. 3d grid-based vdbscan clustering and radar-monocular depth completion. IEEE Sensors Journal, 2024. 1", + "[7] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 4", + "[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. 35, 2022. 2, 4", + "[9] Fangqiang Ding, Andras Palffy, Dariu M. Gavrila, and Chris Xiaoxuan Lu. Hidden gems: 4d radar scene flow learning using cross-modal supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9340-9349, 2023. 1, 2", + "[10] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6824-6835, 2021. 2, 4", + "[11] Kexue Fu, Shaolei Liu, Xiaoyuan Luo, and Manning Wang. Robust point cloud registration framework based on deep graph matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8893-8902, 2021. 2, 3, 4, 6", + "[12] Stefano Gasperini, Patrick Koch, Vinzensz Dallabetta, Nassir Navab, Benjamin Busam, and Federico Tombari. R4dyn: Exploring radar for self-supervised monocular depth estimation of dynamic scenes. In International Conference on 3D Vision (3DV), pages 751-760, 2021. 1, 5, 6", + "[13] Kyle Harlow, Hyesu Jang, Timothy D. Barfoot, Ayoung Kim, and Christoffer Heckman. A new wave in robotics: Survey on recent mmwave radar applications in robotics. IEEE Transactions on Robotics, 40:4544-4560, 2024. 1, 2" + ], + "bbox": [ + 93, + 114, + 483, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 2, 4, 5", + "[15] Daniel Casado Herreraz, Matthias Zeller, Le Chang, Ignacio Vizzo, Michael Heidingsfeld, and Cyril Stachniss. Radar-only odometry and mapping for autonomous vehicles. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 539-548. IEEE, 2019. 1, 2", + "[16] Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Implicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24118-24129, 2024. 1, 2", + "[17] Andrey Ignatov, Grigory Malivenko, David Plowman, Samarth Shukla, and Radu Timofte. Fast and accurate single-image depth estimation on mobile devices, mobile ai 2021 challenge: Report. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2545-2557, 2021. 2", + "[18] Yinuo Jiang, Beitong Zhou, Xiaoyu Liu, Qingyi Li, and Cheng Cheng. Gtinet: Global topology-aware interactions for unsupervised point cloud registration. IEEE Transactions on Circuits and Systems for Video Technology, 34(7): 6363-6375, 2024. 2", + "[19] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, 2012. 2, 4", + "[20] Huadong Li, Minhao Jing, Jiajun Liang, Haoqiang Fan, and Renhe Ji. Sparse beats dense: Rethinking supervision in radar-camera depth completion. arXiv preprint arXiv:2312.00844, 2023. 1, 5, 6, 8", + "[21] Han Li, Yukai Ma, Yaqing Gu, Kewei Hu, Yong Liu, and Xingxing Zuo. Radarcam-depth: Radar-camera fusion for depth estimation with learned metric scale. arXiv preprint arXiv: 2401.04325, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[22] Jiaqi Li, Yiran Wang, Zihao Huang, Jinghong Zheng, Ke Xian, Zhiguo Cao, and Jianming Zhang. Diffusion-augmented depth prediction with sparse annotations. In Proceedings of the 31st ACM International Conference on Multimedia, page 2865-2876, New York, NY, USA, 2023. Association for Computing Machinery. 1", + "[23] Jiaqi Li, Yiran Wang, Jinghong Zheng, Zihao Huang, Ke Xian, Zhiguo Cao, and Jianming Zhang. Self-distilled depth refinement with noisy poisson fusion, 2024. 2", + "[24] Xingyi Li, Chaoyi Hong, Yiran Wang, Zhiguo Cao, Ke Xian, and Guosheng Lin. Symmernf: Learning to explore symmetry prior for single-view view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1726-1742, 2022. 2", + "[25] Zecheng Li, Yuying Song, Fuyuan Ai, Chunyi Song, and Zhiwei Xu. Semantic-guided depth completion from monocular images and 4d radar data. IEEE Transactions on Intelligent Vehicles, pages 1-13, 2024. 1" + ], + "bbox": [ + 516, + 92, + 905, + 881 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10531", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Juan-Ting Lin, Dengxin Dai, and Luc Van Gool. Depth estimation from monocular images and sparse radar data. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10233-10240. IEEE, 2020. 1, 2, 3, 4, 5, 6, 7", + "[27] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. pages 14928-14937, 2024. 2", + "[28] Chen-Chou Lo and Patrick Vandewalle. Depth estimation from monocular images and sparse radar using deep ordinal regression network. In IEEE International Conference on Image Processing (ICIP), pages 3343-3347, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[29] Chen-Chou Lo and Patrick Vandewalle. Rcdpt: Radar-camera fusion dense prediction transformer. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 3, 4, 5, 6", + "[30] Chen-Chou Lo and Patrick Vandewalle. Expanding sparse radar depth based on joint bilateral filter for radar-guided monocular depth estimation. Sensors, 24(6):1864, 2024.", + "[31] Yunfei Long, Daniel Morris, Xiaoming Liu, Marcos Castro, Punarjay Chakravarty, and Praveen Narayanan. Radar-camera pixel depth association for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12507-12516, 2021. 1, 2, 3, 4, 5, 6, 7", + "[32] Guohao Peng, Heshan Li, Yangyang Zhao, Jun Zhang, Zhenyu Wu, Pengyu Zheng, and Danwei Wang. Transloc4d: Transformer-based 4d radar place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17595-17605, 2024. 1, 2", + "[33] Pierluigi Zama Ramirez, Fabio Tosi, Luigi Di Stefano, Radu Timofte, Alex Costanzino, Matteo Poggi, Samuele Salti, Stefano Mattoccia, Yangyang Zhang, Cailin Wu, Zhuangda He, Shuangshuang Yin, Jiaxu Dong, Yangchenxu Liu, Hao Jiang, Jun Shi, Yong A, Yixiang Jin, Dingzhe Li, Bingxin Ke, Anton Obukhov, Tinafu Wang, Nando Metzger, Shengyu Huang, Konrad Schindler, Yachuan Huang, Jiaqi Li, Junrui Zhang, Yiran Wang, Zihao Huang, Tianqi Liu, Zhiguo Cao, Pengzhi Li, Jui-Lin Wang, Wenjie Zhu, Hui Geng, Yuxin Zhang, Long Lan, Kele Xu, Tao Sun, Qisheng Xu, Sourav Saini, Aashray Gupta, Sahaj K. Mistry, Aryan Shukla, Vinit Jakhetiya, Sunil Jaiswal, Yuejin Sun, Zhuofan Zheng, Yi Ning, Jen-Hao Cheng, Hou-I Liu, Hsiang-Wei Huang, Cheng-Yen Yang, Zhongyu Jiang, Yi-Hao Peng, Aishi Huang, and Jenq-Neng Hwang. Ntire 2024 challenge on hr depth from images of specular and transparent surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6499–6512, 2024. 2", + "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(03):1623-1637, 2020. 1, 2, 3, 4, 5, 6", + "[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi" + ], + "bbox": [ + 91, + 90, + 480, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12179-12188, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[36] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks and Learning Systems, 20(1):61-80, 2009. 2, 3, 4, 6", + "[37] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9275-9285, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[38] Tieshuai Song, Bin Yang, Jun Wang, Guidong He, Zhao Dong, and Fengjun Zhong. mmwave radar and image fusion for depth completion: a two-stage fusion network. In The 27th International Conference on Information Fusion (FUSION), 2024. 1, 2, 3, 4, 5, 6", + "[39] Christian Stetco, Barnaba Ubezio, Stephan Muhlbacher-Karrer, and Hubert Zangl. Radar sensors in collaborative robotics: Fast simulation and experimental validation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10452-10458. IEEE, 2020. 1, 2", + "[40] Huawei Sun, Hao Feng, Julius Ott, Lorenzo Servadei, and Robert Wille. Cafnet: A confidence-driven framework for radar camera depth estimation. arXiv preprint arXiv:2407.00697, 2024. 1, 2, 3, 4, 5, 6, 7", + "[41] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6411–6420, 2019. 2, 3, 4, 8", + "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, 2017. 4", + "[43] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3056-3065, 2019. 2, 3, 4", + "[44] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3056-3065, 2019. 3", + "[45] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5), 2019. 2, 3, 4, 8", + "[46] Yiran Wang, Xingyi Li, Min Shi, Ke Xian, and Zhiguo Cao. Knowledge distillation for fast and accurate monocular depth estimation on mobile devices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2457-2465, 2021. 2", + "[47] Yiran Wang, Zhiyu Pan, Xingyi Li, Zhiguo Cao, Ke Xian, and Jianming Zhang. Less is more: Consistent video depth estimation with masked frames modeling. In Proceedings" + ], + "bbox": [ + 516, + 92, + 906, + 896 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10532", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the 30th ACM International Conference on Multimedia, page 6347-6358, New York, NY, USA, 2022. Association for Computing Machinery. 2", + "[48] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9466-9476, 2023. 2", + "[49] Yiran Wang, Min Shi, Jiaqi Li, Chaoyi Hong, Zihao Huang, Juewen Peng, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Nvds+: Towards efficient and versatile neural stabilizer for video depth estimation. IEEE transactions on pattern analysis and machine intelligence, pages 1-18, 2024. 1", + "[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar," + ], + "bbox": [ + 91, + 90, + 482, + 301 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. 34:12077-12090, 2021. 2, 4", + "[51] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 4, 5, 6, 7, 8", + "[52] Junrui Zhang, Jiaqi Li, Yachuan Huang, Yiran Wang, Jinghong Zheng, Liao Shen, and Zhiguo Cao. Towards robust monocular depth estimation in non-lambertian surfaces, 2024. 5", + "[53] Yi Zhou, Lulu Liu, Haocheng Zhao, Miguel López-Benítez, Limin Yu, and Yutao Yue. Towards deep radar perception for autonomous driving: Datasets, methods, and challenges. Sensors, 22(11):4208, 2022. 1, 2" + ], + "bbox": [ + 516, + 90, + 903, + 303 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10533", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_model.json b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_model.json new file mode 100644 index 0000000000000000000000000000000000000000..30f2cde8901555e5997b23ca6b9e30aa86e7e230 --- /dev/null +++ b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_model.json @@ -0,0 +1,2147 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.768, + 0.177 + ], + "angle": 0, + "content": "TacoDepth: Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.203, + 0.805, + 0.257 + ], + "angle": 0, + "content": "Yiran Wang\\(^{1,2,*}\\) Jiaqi Li\\(^{2,*}\\) Chaoyi Hong\\(^{1,2}\\) Ruibo Li\\(^{1}\\) \nLiusheng Sun\\(^{3}\\) Xiao Song\\(^{3}\\) Zhe Wang\\(^{3}\\) Zhiguo Cao\\(^{2}\\) Guosheng Lin\\(^{1,\\dagger}\\) \n\\(^{1}\\)S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.257, + 0.86, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)School of AIA, Huazhong University of Science and Technology \\(^{3}\\)SenseTime Research" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.277, + 0.707, + 0.293 + ], + "angle": 0, + "content": "{wangyiran,zgcao}@hust.edu.cn, gslin@ntu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.295, + 0.689, + 0.31 + ], + "angle": 0, + "content": "https://github.com/RaymondWang987/TacoDepth" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.337, + 0.327, + 0.354 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.485, + 0.704 + ], + "angle": 0, + "content": "Radar-Camera depth estimation aims to predict dense and accurate metric depth by fusing input images and Radar data. Model efficiency is crucial for this task in pursuit of real-time processing on autonomous vehicles and robotic platforms. However, due to the sparsity of Radar returns, the prevailing methods adopt multi-stage frameworks with intermediate quasi-dense depth, which are time-consuming and not robust. To address these challenges, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation model with one-stage fusion. Specifically, the graph-based Radar structure extractor and the pyramid-based Radar fusion module are designed to capture and integrate the graph structures of Radar point clouds, delivering superior model efficiency and robustness without relying on the intermediate depth results. Moreover, TacoDepth can be flexible for different inference modes, providing a better balance of speed and accuracy. Extensive experiments are conducted to demonstrate the efficacy of our method. Compared with the previous state-of-the-art approach, TacoDepth improves depth accuracy and processing speed by \\(12.8\\%\\) and \\(91.8\\%\\). Our work provides a new perspective on efficient Radar-Camera depth estimation." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.731, + 0.222, + 0.746 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Radar-Camera depth estimation [6, 12, 20, 21, 25, 26, 28-31, 37] is a vital task for 3D perception and motion planning in autonomous driving [4, 15, 22, 49, 53] and robotics [13, 21, 39]. The task aims to predict dense metric depth by integrating images with mmWave Radar data, perceiving metric scale in different environments. As a typical range sensor on driving and robotic platforms, Radar is widely equipped and used for its high affordability and low power consump" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.337, + 0.917, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.528, + 0.908, + 0.6 + ], + "angle": 0, + "content": "Figure 1. Performance and efficiency. Circle area indicates inference time (ms). Smaller circles showcase faster speed. The X-axis and Y-axis represent MAE and RMSE metrics for depth errors on the nuScenes [4] dataset. Lower MAE and RMSE mean higher accuracy. Our TacoDepth outperforms prior arts by large margins." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.685 + ], + "angle": 0, + "content": "tion. It is valuable to explore Radar-Camera fusion in various tasks [2, 5, 9, 16, 32, 37] for its superior all-weather reliability. Besides, for real-time processing on vehicles and robotics, improving efficiency is crucial for fusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The obstacle for fusing Radar in depth estimation lies in its sparsity and noises. Due to the large beam width, Radar point clouds are \\(1000\\mathrm{x}\\) sparser with higher noise levels than LiDAR returns. To deal with the issue, the prevailing methods [21, 26, 28, 29, 31, 37, 38, 40] adopt complex multi-stage frameworks. Intermediate quasi-dense depth is predicted to bridge the gap between sparse Radar and final dense depth. For instance, two-stage methods [26, 28, 29, 31, 37, 38, 40] predict final depth based on intermediate results, as independent models without utilizing other depth predictors [3, 34, 35, 51]. RadarCam-Depth [21], as a plugin module, converts relative depth [34, 35] to metric scale in four stages by post-processing. These multi-stage methods are inefficient and time-consuming, as presented in Fig. 1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.356, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution, † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10523" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "Meanwhile, despite being described as quasi-dense, the intermediate depth [21, 26, 28, 29, 31, 37, 40] remains sparse and noisy. As shown in Fig. 2, only few pixels exhibit valid depth. For some adverse lighting or weather conditions, e.g., glaring and nighttime scenarios, even no pixels are predicted with depth values. These defective intermediate results lead to disrupted structures, blurred details, or noticeable artifacts in final predictions, as shown in Fig. 4. With these observations, a question naturally arises: can models directly predict accurate and dense depth without the intermediate results, achieving one-stage Radar-Camera depth estimation with higher efficiency and robustness?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.485, + 0.578 + ], + "angle": 0, + "content": "To answer the above question, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation framework with one-stage fusion. To simplify multi-stage fusion [21, 26, 28, 29, 31, 37, 38, 40] into one stage, TacoDepth needs to extract and integrate Radar information more effectively. Prior arts simply extract features from Radar coordinates [21, 37] or projected depth [26, 31], sensitive to noises and outliers [11]. In contrast, we utilize a graph-based Radar structure extractor to capture geometric structures and topologies of point clouds, improving robustness to inaccurate points [11, 18, 36, 43]. After that, our pyramid-based Radar fusion module integrates image and Radar features hierarchically. Shallow layers capture image details and Radar coordinates, while deep layers fuse scene semantics and structures [10, 14, 19, 41, 45, 50]. In each layer, Radar-centered flash attention mechanism is applied to establish the cross-modal correlations. For one Radar point, lightweight and accelerated flash attention [8] is calculated within Radar-centered areas for efficiency. Finally, TacoDepth predicts accurate and dense depth by a decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.748 + ], + "angle": 0, + "content": "Moreover, we further analyze the characteristics of different inference strategies in prior arts [21, 26, 28, 29, 31, 37, 38, 40]. Independent models [26, 28, 29, 31, 37, 40] are more efficient without initial depth prediction and postprocessing, while the plug-in model [21] can be more accurate by leveraging cutting-edge relative depth predictors [3, 23, 34, 51]. Therefore, TacoDepth aims to be flexible for the two settings. An auxiliary and optional branch is added for initial depth, allowing TacoDepth to support both independent and plug-in processing. Our approach can better balance depth accuracy and model efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.904 + ], + "angle": 0, + "content": "We conduct evaluations on nuScenes [4] and ZJU-4DRadarCam [21] datasets with different depth ranges, following prior arts [21, 31, 37]. As shown in Fig. 1, whether compared with independent [26, 28, 31, 37, 40] or plugin in [21] models, TacoDepth significantly outperforms prior arts [21, 26, 28, 31, 37, 40] in accuracy and efficiency. Besides, our approach achieves real-time processing of over 37 fps as an independent model and strong compatibility with various depth predictors [3, 34, 35, 51] as a plug-in model, further proving the efficiency and flexibility of TacoDepth." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.905, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.173, + 0.907, + 0.216 + ], + "angle": 0, + "content": "Figure 2. Intermediate quasi-dense depth maps [21, 26, 28, 29, 31, 37, 40] remain sparse and noisy. Pixels with valid depth values are visualized by the gray areas in the red rectangular boxes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.235, + 0.875, + 0.249 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.254, + 0.905, + 0.284 + ], + "angle": 0, + "content": "- We propose TacoDepth, an efficient and accurate Radar-Camera depth estimation method with one-stage fusion." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.285, + 0.905, + 0.314 + ], + "angle": 0, + "content": "- We design the pyramid-based Radar fusion module, fusing Radar and image features effectively and efficiently." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.315, + 0.905, + 0.345 + ], + "angle": 0, + "content": "- TacoDepth is flexible for both independent and plug-in processing, balancing model efficiency and performance." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.254, + 0.905, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.361, + 0.655, + 0.377 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.907, + 0.658 + ], + "angle": 0, + "content": "MmWave Radar for Vehicles and Robotics. Autonomous driving [4, 15, 48, 53] and robotics [13, 21, 24, 39] rely on multiple sensors to perceive the environment, e.g., camera, LiDAR, and mmWave Radar. LiDAR and Radar provide the metric scale of different scenarios as a complement to cameras. Compared with LiDAR, Radar is cheaper and more widely-equipped with lower power consumption. The reflection properties of mmWave also provide Radar with farther sensing ranges and stronger robustness to all-weather conditions. The main drawback of Radar lies in its sparsity and noise. Due to the large beam width, Radar point clouds are \\(1000\\mathrm{x}\\) sparser [37] than LiDAR. Besides, Radar captures coordinates and depth values with lower accuracy. For instance, 3D Radar suffers from height ambiguity [31] due to insufficient antenna elements along the elevation axis, while this problem can be alleviated by the emerging 4D Radar. Thus, integrating Radar data is valuable for various tasks [2, 5, 9, 16, 27, 32, 37] in driving and robotic systems." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Radar-Camera Depth Estimation. The task aims to predict dense metric depth by merging images and Radar, facilitating 3D perception and motion planning of vehicles and robotics. Model efficiency is crucial for these real-time applications. However, to overcome the sparsity and noise of Radar, prior arts [21, 26, 28, 29, 31, 37, 38, 40] adopt inefficient multi-stage frameworks. The final depth is predicted based on intermediate quasi-dense depth. These methods can be divided into independent and plug-in models. Most independent models [26, 28, 29, 31, 37, 40] predict final depth in two stages without relying on other depth predictors [17, 33-35, 46, 47]. For example, Lin et al. [26] and DORN [28] use two CNNs [14] for quasi-dense and final depth. RC-PDA [31] and RC-DPT [29] convert intermediate results into multiple channels for varied confidence levels. Singh et al. [37] propose gated fusion to filter the inac" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10524" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.088, + 0.621, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.885, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.358, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 3. Overview of the TacoDepth. The graph-based Radar structure extractor captures graph structures of Radar point clouds through the node feature \\( N_{l} \\) and edge feature \\( E_{l} \\) in a certain layer \\( l \\). The pyramid-based Radar fusion module integrates image and Radar features in a pyramidal hierarchical manner. In each layer, to efficiently build cross-modal correspondences, the Radar-centered flash attention is calculated within Radar-centered areas based on horizontal coordinates, e.g., pixels in the orange area as queries, while Radar points in the red area as keys and values. Our TacoDepth achieves efficient and accurate Radar-Camera depth estimation in one stage. During inference, the model is flexible and supports both independent and plug-in processing, facilitating a better balance of model efficiency and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.468, + 0.483, + 0.59 + ], + "angle": 0, + "content": "curate quasi-dense depth. Besides, RadarCam-Depth [21] designs a plug-in module, aligning relative depth [34, 35] to metric scale in four stages, including initial depth, global scale alignment, quasi-dense depth, and scale learner. These methods [21, 26, 28, 29, 31, 37, 38, 40] are inefficient with complex stages. Defective intermediate depth also limits their robustness. In contrast, we propose TacoDepth model to achieve efficient, accurate, and flexible one-stage fusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.602, + 0.208, + 0.62 + ], + "angle": 0, + "content": "3. TacoDepth" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.483, + 0.749 + ], + "angle": 0, + "content": "We present a detailed illustration of TacoDepth, an efficient Radar-Camera depth estimation framework with one-stage fusion. In Sec. 3.1, we provide an overall description to outline our approach. In Sec. 3.2, we elaborate on the detailed architecture of the pyramid-based Radar fusion module. TacoDepth can be flexible for independent and plug-in processing, which will be discussed in Sec. 3.3. Besides, the training procedure for TacoDepth is illustrated in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.201, + 0.772 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Here we present an overview of our TacoDepth framework. Given a single image \\(I \\in \\mathcal{R}^{H \\times W \\times 3}\\) and a Radar point cloud \\(P \\in \\mathcal{R}^{K \\times 3}\\) as inputs, our goal is to predict dense metric depth \\(D \\in \\mathcal{R}^{H \\times W}\\). \\(H\\) and \\(W\\) represent image height and weight. The point cloud \\(P\\) contains \\(K\\) points \\(p \\in \\mathcal{R}^3\\). Radar-Camera depth estimation can be divided into two primary subtasks: the extraction and the integration of Radar information. Accordingly, as depicted in Fig. 3," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.514 + ], + "angle": 0, + "content": "TacoDepth comprises a graph-based Radar structure extractor and a pyramid-based Radar fusion module to undertake these two important steps respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.908, + 0.623 + ], + "angle": 0, + "content": "Firstly, for Radar feature extraction, previous methods [21, 37] simply extract features from the coordinates of individual Radar points, focusing solely on the point-level information. For example, Singh et al. [37] extract point features in 32,256 dimensions from the three-dimensional coordinates using a multi-layer perception (MLP), which inevitably introduces substantial redundancy and noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.909, + 0.883 + ], + "angle": 0, + "content": "To achieve efficient Radar-Camera fusion in a single stage, TacoDepth needs to extract more comprehensive and effective Radar information, prompting us to reconsider the complementarity between camera images and Radar data. Compared to the coordinates, the geometric structures of point clouds, e.g., the distances or relations between Radar points, are potentially more informative for Radar-Camera fusion. The overall topologies are also more robust to inaccurate outliers [11]. To this end, we employ the graph-based Radar structure extractor with a lightweight GNN [11, 36, 43] architecture. Specifically, Radar points are regarded as the graph nodes. The adjacency matrix of all points represents the graph edges. In each layer, the node features of Radar points can be updated and aggregated along edges by PCA-GM [44]. The graph-based Radar structure extractor can capture point coordinates and graph structures from shallow layers to deep layers [11, 41, 45]." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "After extracting the Radar structures, the subsequent step" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10525" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "is to merge RGB and Radar features. Prior arts [21, 26, 28, 29, 31, 37, 38, 40] adopt multi-stage methods with intermediate quasi-dense depth, which are inefficient and not robust as illustrated in Sec. 1. In contrast, we propose our pyramid-based Radar fusion module for efficient one-stage fusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.426 + ], + "angle": 0, + "content": "To be specific, features of the two modalities are integrated in a pyramidal hierarchical manner. Shallow layers can merge image details with Radar coordinates, whereas deeper layers could fuse scene semantics and geometric structures [10, 14, 19, 41, 45, 50]. In each layer, we leverage Radar-centered flash attention for cross-modal correlations and feature fusion. The queries are derived from image pixels, while keys and values originate from Radar points. For each point, we delineate Radar-centered areas based on the horizontal coordinates, as exemplified by the orange and red areas in Fig. 3, given that the projected horizontal coordinates of Radar points are generally more precise than the height dimension [31, 37]. Lightweight flash attention [8] is calculated within these areas to maintain efficiency, since external pixels and points can not be correlated. Finally, with the fused features from all layers, TacoDepth predicts accurate and dense metric depth by a depth decoder [37]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.483, + 0.562 + ], + "angle": 0, + "content": "Furthermore, TacoDepth is adaptable for both independent and plug-in inference. An optional input branch is introduced to handle the initial relative depth. For the independent inference, our model achieves real-time processing of over 37 fps. For the plug-in manner, TacoDepth demonstrates compatibility with various cutting-edge depth predictors [3, 34, 35, 51], as proved in Sec. 4.5. Consequently, our approach can deliver a balanced trade-off between performance and efficiency for different users and applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.355, + 0.59 + ], + "angle": 0, + "content": "3.2. Pyramid-based Radar Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.749 + ], + "angle": 0, + "content": "For the Radar point cloud \\( P \\), our graph-based Radar structure extractor derives node feature \\( N_{l} \\) and edge feature \\( E_{l} \\) at a certain layer \\( l \\in \\{1,2,\\dots ,L\\} \\), where \\( L \\) represents the layer number of the GNN module [11, 36, 43]. For the input image \\( I \\), the image encoder [14, 21, 37] extracts multi-scale features \\( \\{F_1,F_2,\\dots ,F_{2L}\\} \\). The pyramid-based Radar fusion module hierarchically integrates the multi-modal features. \\( N_{l} \\) and \\( E_{l} \\) will be fused with image features in adjacent layers. We merge the node feature \\( N_{l} \\) with image feature \\( F_{2l - 1} \\), and the edge feature \\( E_{l} \\) with \\( F_{2l} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The prerequisite for feature fusion is cross-modal correspondences between Radar points and RGB pixels. Due to the noisy elevation coordinates of 3D Radar, prior arts [21, 29, 31, 37, 40] try to build one-to-many mappings by the intermediate quasi-dense depth. They adopt explicit classification supervision with the depth discrepancies between Li-DAR and Radar as pseudo-labels. However, the quasi-dense depth prediction leads to inefficiency and erroneous final outcomes, as discussed in Sec. 1. Given the absence of reliable explicit labels and supervision, we hope our TacoDepth" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.874, + 0.107 + ], + "angle": 0, + "content": "can learn and establish the correspondences implicitly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.905, + 0.289 + ], + "angle": 0, + "content": "Therefore, for efficient Radar-Camera correlations, we propose the Radar-centered flash attention in TacoDepth. We define Radar-centered areas based on the precise horizontal coordinates of Radar points and image pixels. For each point, the attention is computed only with image pixels in the area, since external pixels can not be correlated. Here we take \\( E_{l} \\in \\mathcal{R}^{K \\times K} \\) and \\( F_{2l} \\in \\mathcal{R}^{h \\times w \\times C_l} \\) as examples, where \\( h \\) and \\( w \\) are feature height and width. \\( K \\) and \\( C_l \\) refer to the point number and layer feature channel. For a Radar point \\( p \\) with the horizontal coordinate \\( x_{p} \\), as showcased by the exemplary orange area in Fig. 3, we only retain image pixels \\( m \\) whose coordinates \\( x_{m} \\) are close to \\( x_{p} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.294, + 0.905, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {2 l} = F _ {2 l} \\left[ x _ {p} - a _ {l} < x _ {m} < x _ {p} + a _ {l} \\right], \\hat {F} _ {2 l} \\in \\mathcal {R} ^ {\\hat {M} \\times C _ {l}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.319, + 0.905, + 0.41 + ], + "angle": 0, + "content": "where \\(a_{l}\\) indicates the width of the Radar-centered area in layer \\(l\\). The notation \\([\\cdot]\\) depicts the fetching operation of pixels or points. \\(\\hat{F}_{2l}\\) is the image feature of the remaining \\(\\hat{M}\\) pixels to generate queries. For a random pixel \\(m\\) in \\(\\hat{F}_{2l}\\), as represented by the red area in Fig. 3, the attention calculation is also restricted to adjacent Radar points:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.415, + 0.905, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\hat {E} _ {l} = E _ {l} \\left[ x _ {m} - a _ {l} < x _ {p} < x _ {m} + a _ {l} \\right], \\hat {E} _ {l} \\in \\mathcal {R} ^ {\\hat {K} \\times K}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.905, + 0.545 + ], + "angle": 0, + "content": "The feature dimension maintains the original \\(K\\) to preserve integral graph topologies. \\(\\hat{E}_l\\) is the edge feature of the retained \\(\\hat{K}\\) points for keys and values. In this way, the computational costs can be significantly decreased with the small subset of \\(\\hat{M}\\) pixels and \\(\\hat{K}\\) points. For the pixel \\(m\\), we generate queries from \\(\\hat{F}_{2l}[m] \\in \\mathcal{R}^{C_l}\\), while keys and values are derived from \\(\\hat{E}_l\\). Radar-centered attention is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.55, + 0.905, + 0.596 + ], + "angle": 0, + "content": "\\[\nF _ {2 l} ^ {\\prime} [ m ] = \\operatorname {s o f t m a x} \\frac {W _ {q} \\hat {F} _ {2 l} [ m ] \\left(W _ {k} \\hat {E} _ {l}\\right) ^ {T}}{\\sqrt {C _ {l}}} W _ {v} \\hat {E} _ {l}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.704 + ], + "angle": 0, + "content": "in which \\(W_{q}\\), \\(W_{k}\\), and \\(W_{v}\\) are learnable linear projections. The fused Radar-Camera feature \\(F_{2l}^{\\prime} \\in \\mathcal{R}^{h \\times w \\times C_l}\\) can be acquired for the current layer. The above calculation is incorporated into attention blocks with MLPs and residual connections. To further improve model efficiency, we employ lightweight and accelerated flash attention modules [7, 8] to replace the original attention computation [42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.78 + ], + "angle": 0, + "content": "Similarly, the node feature \\( N_{l} \\) can be merged with the image feature \\( F_{2l-1} \\) by slightly adjusting the dimensions of linear projections. Ultimately, the acquired Radar-Camera features \\( \\{F_1', F_2', \\dots, F_{2L}'\\} \\) are fed to a common depth decoder [21, 37] to predict accurate metric depth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As a consequence, the proposed pyramid-based Radar fusion module enables efficient and effective Radar-Camera fusion in a single stage. The Radar-centered flash attention establishes cross-modal correlations with minimal computational costs. Without relying on the previous multi-stage frameworks and intermediate quasi-dense depth [21, 26, 28, 29, 31, 37, 38, 40], TacoDepth achieves higher efficiency and stronger robustness for Radar-Camera depth estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10526" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.267, + 0.106 + ], + "angle": 0, + "content": "3.3. Flexible Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.205 + ], + "angle": 0, + "content": "Considering the varied inference manners in prior arts [21, 26, 28, 29, 31, 37, 38, 40], the plug-in module [21] could achieve higher accuracy using state-of-the-art relative depth predictors [34, 35, 52], while independent models [26, 28, 29, 31, 37, 38, 40] are capable of attaining faster inference speed without initial depth prediction and post-processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.205, + 0.484, + 0.295 + ], + "angle": 0, + "content": "Therefore, TacoDepth is designed to be flexible, supporting both independent and plug-in predictions. For the plug-in inference, an initial relative depth map is converted to metric scale by fusing camera and Radar information. If we denote the TacoDepth model as \\(\\mathcal{T}_{\\theta}\\), parameterized by \\(\\theta\\), the depth prediction process can then be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.309, + 0.482, + 0.327 + ], + "angle": 0, + "content": "\\[\nD = \\mathcal {T} _ {\\theta} (I, P \\mid D ^ {*}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.501 + ], + "angle": 0, + "content": "where the vertical bar \\((|)\\) separates the required and optional inputs. \\(D\\) is the predicted depth of our model. \\(D^{*}\\) is the optional input of initial relative depth from depth predictors [3, 34, 35, 51]. We introduce an auxiliary branch to process and fuse \\(D^{*}\\) with image features, while keeping all other settings identical to the independent mode. Our plugin in manner demonstrates compatibility with various depth predictors, e.g., MiDaS [3, 34], DPT [35], and Depth-Anything-v2 [51]. In this way, TacoDepth can achieve a better balance of model efficiency and accuracy, offering superior flexibility for different users and applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.508, + 0.312, + 0.525 + ], + "angle": 0, + "content": "3.4. Training the TacoDepth" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.621 + ], + "angle": 0, + "content": "In the training phase, the training data [4, 21] is randomly divided into two equal portions in each epoch. For the auxiliary branch, one half of the data is provided with initial relative depth, while the other half adopts zero as input. This training strategy facilitates the simultaneous optimization of the independent inference and plug-in prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.698 + ], + "angle": 0, + "content": "As for the training loss, we employ the same depth ground truth and \\(L_{1}\\) loss as the prior art [37]. Specifically, LiDAR-captured depth \\(D_{gt}\\) and reprojection-accumulated depth \\(D_{acc}\\) are utilized for supervision [21, 31, 37]. The overall loss function \\(\\ell_{L_1}\\) can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.707, + 0.484, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\ell_ {L _ {1}} = \\frac {1}{| \\Omega_ {g t} |} \\sum_ {\\Omega_ {g t}} | D - D _ {g t} | + \\frac {\\lambda}{| \\Omega_ {a c c} |} \\sum_ {\\Omega_ {a c c}} | D - D _ {a c c} |, (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.756, + 0.483, + 0.789 + ], + "angle": 0, + "content": "where \\(\\Omega_{gt}\\) and \\(\\Omega_{acc}\\) depict the masks of \\(D_{gt}\\) and \\(D_{acc}\\) with valid depth values. \\(\\lambda\\) is the coefficient of the second term." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.224, + 0.818 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To prove the efficacy of our TacoDepth, we conduct experiments on the nuScenes [4] and ZJU-4DRadarCam [21] datasets following prior arts [20, 21, 26, 28, 29, 31, 37, 40]. In Sec. 4.1, we describe the experimental datasets and evaluation protocols. The implementation details of our model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.184 + ], + "angle": 0, + "content": "are illustrated in Sec. 4.2. In Sec. 4.3, we compare the TacoDepth with previous Radar-Camera depth estimation approaches to demonstrate our state-of-the-art performance. The inference speed and model efficiency of different methods are evaluated and discussed in Sec. 4.4. In Sec. 4.5, we perform ablation studies to expound on our specific designs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.194, + 0.813, + 0.209 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.216, + 0.907, + 0.427 + ], + "angle": 0, + "content": "Due to the scarcity of public datasets for Radar-Camera depth estimation, prior works [20, 26, 28, 29, 31, 37, 38, 40] only conduct experiments on nuScenes dataset [4] with 3D Radar. Recently, RadarCam-Depth [21] releases their ZJU-4DRadarCam dataset [21] with 4D Radar to alleviate this problem. We adopt these two datasets [4, 21] for experiments, proving our effectiveness on both 3D and 4D Radar. NuScenes Dataset. As an outdoor driving dataset, the nuScenes [4] is collected by vehicles equipped with cameras, 3D Radar, LiDAR, and IMU. It contains 1,000 scenes and around 40,000 synchronized Radar-Camera keyframes captured in Boston and Singapore. We follow the same train-test split as prior arts [21, 26, 29, 31, 37, 40], with 700 scenes for training, 150 for validation, and 150 for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.534 + ], + "angle": 0, + "content": "ZJU-4DRadarCam Dataset. The dataset [21] is collected by a ground robot fitted with more advanced camera, 4D Radar, and LiDAR sensors, offering denser Radar data and LiDAR depth than the nuScenes [4] dataset. It comprises 33,409 Radar-Camera keyframes in total. We follow the official data split [21] with 29,312 frames for training and validation, and 4,097 samples for testing in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.716 + ], + "angle": 0, + "content": "Evaluation Protocols. We evaluate independent and plug-in models separately for fair comparisons on the two datasets [4, 21] with depth ranges of 50, 70, and 80 meters [21, 31, 37, 40]. On nuScenes [4], previous state-of-the-art methods [12, 20, 21, 26, 28, 31, 37, 40] are compared comprehensively to the best of our knowledge. For the recently published ZJU-4DRadarCam [21], since most prior works do not include experiments on the dataset, we can only follow the same comparative methods [21, 28, 37] and results as RadarCam-Depth [21]. Besides, we also evaluate plug-in models with different depth predictors [3, 35, 51] to showcase model flexibility on the ZJU-4DRadarCam [21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.905, + 0.763 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt the commonly-applied depth metrics MAE, RMSE, iMAE, iRMSE, \\(\\delta_{1}\\), and Rel as prior arts [21, 26, 31, 37]. See supplement for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.731, + 0.789 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Data Processing. We adhere to the data processing procedures of prior arts [21, 37]. For the nuScenes [4], the \\(D_{acc}\\) in Eq. 5 is generated by accumulating the \\(D_{gt}\\) of 160 nearby frames [21, 37]. For the ZJU-4DRadarCam dataset [21], the \\(D_{acc}\\) is obtained via the linear interpolation [1, 21] of \\(D_{gt}\\). Model Architecture. In line with prior arts [21, 37], the ResNet-18 [14] is adopted as the image encoder. The graph" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10527" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.089, + 0.865, + 0.273 + ], + "angle": 0, + "content": "
TypeMethodRadar FramesImagesTime(ms)↓0 - 50m0 - 70m0 - 80m
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
Indep-endentLin et al. [26] (IROS'20)31117.62034.94316.52294.75338.22371.05623.0
RC-PDA [31] (CVPR'21)5379.22225.04156.53326.16700.63713.67692.8
R4Dyn [12] (3DV'21)41------6434.0
DORN [28] (ICIP'21)5(×3)1392.51926.64124.82380.65252.72467.75554.3
Singh et al. [37] (CVPR'23)1194.21727.73746.82073.24590.72179.34898.7
CaFNet [40] (IROS'24)11103.91674.23674.52010.34493.12109.84765.6
Li et al. [20] (ECCV'24)1167.61524.53567.31822.94303.61927.04609.6
TacoDepth (Ours)1126.71423.63275.81712.63960.51833.44150.2
Plug-inRadarCam-Depth [21] (ICRA'24)11358.31286.12964.31587.93662.51689.73948.0
TacoDepth (Ours)1129.31046.82487.51347.13152.81492.43324.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.277, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Table 1. Comparisons with state-of-the-art methods on the nuScenes dataset [4] (in millimeters). Independent [12, 20, 26, 28, 31, 37, 40] and plug-in [21] models are evaluated separately, with maximum evaluation distances of 50, 70, and 80 meters as prior arts [21, 37]. Plug-in models are compared using the same depth predictor DPT-Hybrid [35] as RadarCam-Depth [21]. We report the average runtime of processing one \\(900 \\times 1600\\) frame and 30 Radar points by different methods on one NVIDIA RTX A6000 GPU. The inference time of plug-in models does not contain initial depth prediction. Some methods [12, 26, 28, 31] require multiple frames and Radar scans for inference, while others [20, 21, 37, 40] only need a single frame as input. The best performance is in boldface. Second best is underlined." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.367, + 0.865, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.619, + 0.908, + 0.649 + ], + "angle": 0, + "content": "Figure 4. Visual results of independent models on nuScenes [4]. Both daytime and nighttime samples are presented. Prior arts [31, 37] exhibit disrupted structures and noticeable artifacts. TacoDepth produces accurate depth with finer details and more complete structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.735 + ], + "angle": 0, + "content": "based Radar structure extractor contains \\( L = 3 \\) layers with the GNN [11, 36] architecture. The widths \\( a_{l} \\) of Radar-centered areas in Eq. 1 and Eq. 2 are set to \\( \\{48, 32, 16\\} \\) for the three layers. Refer to the supplement for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.842 + ], + "angle": 0, + "content": "Training Recipe. The model is trained by Adam optimizer on two NVIDIA A6000 GPUs with a batch size of 12 for 50 epochs. The initial learning rate is \\(1e^{-4}\\) and decreases by \\(1e^{-5}\\) for every ten epochs. The input image size is \\(900 \\times 1600\\) [21, 37] and \\(300 \\times 1280\\) [21] on nuScenes [4] and ZJU-4DRadarCam [21], respectively. The \\(\\lambda\\) in Eq. 5 is set to one. All other settings are consistent with prior works [21, 37]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.465, + 0.866 + ], + "angle": 0, + "content": "4.3. Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparisons with Independent Models. Independent models [12, 20, 26, 28, 29, 31, 37, 38, 40] predict dense" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.885 + ], + "angle": 0, + "content": "metric depth without using relative depth predictors [3, 34, 35, 51]. Most of them adopt two-stage methods with intermediate quasi-dense depth, which are inefficient and not robust. Our TacoDepth significantly surpasses these methods in efficiency and accuracy with one-stage fusion. For example, in Table 1, compared with the recent CaFNet [40], our model reduces RMSE, MAE, and inference time by over \\(10.9\\%\\), \\(13.1\\%\\), and \\(74.3\\%\\) on nuScenes dataset [4]. Additionally, Li et al. [20] aim to improve the training paradigm under sparse supervision by their Disruption-Compensation framework. Without using such advanced training and supervision strategy, TacoDepth still outperforms Li et al. [20] by over \\(60.5\\%\\), \\(4.9\\%\\), and \\(8.1\\%\\) for runtime, MAE, and RMSE, showing the efficacy of our proposed architectures." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Similarly, in Table 2, compared with Singh et al. [37]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10528" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.089, + 0.865, + 0.187 + ], + "angle": 0, + "content": "
TypeMethod0 - 50m0 - 70m0 - 80m
MAE↓RMSE↓iMAE↓iRMSE↓MAE↓RMSE↓iMAE↓iRMSE↓MAE↓RMSE↓iMAE↓iRMSE↓
Indep-endentDORN [28] (ICIP'21)2210.24129.719.831.92402.24625.219.831.92447.64760.019.931.9
Singh et al. [37] (CVPR'23)1785.43704.618.135.31932.74137.118.035.21979.54309.317.935.1
TacoDepth (Ours)1120.12686.712.825.01181.82906.312.724.91201.12990.712.724.9
Plug-inRadarCam-Depth [21] (ICRA'24)1067.52817.410.522.91157.03117.710.422.91183.53229.010.422.8
TacoDepth (Ours)930.22477.39.320.8983.12779.69.320.91032.52850.39.420.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.191, + 0.907, + 0.234 + ], + "angle": 0, + "content": "Table 2. Comparisons with state-of-the-art methods on ZJU-4DRadarCam [21] (in millimeters). Plug-in models are compared using the same relative depth predictor DPT-Hybrid [35]. Since many prior works do not include experiments on this new dataset, we adhere to the same compared methods [21, 28, 37] and metrics as RadarCam-Depth [21]. Best performance is in boldface. Second best is underlined." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.244, + 0.865, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.566, + 0.907, + 0.595 + ], + "angle": 0, + "content": "Figure 5. Visual results of plug-in models [21] on the ZJU-4DRadarCam [21]. The same depth predictor DPT-Hybrid [35] is adopted for fair comparisons. Regions with obvious differences are highlighted in the rectangular boxes. Best view zoomed in on-screen for details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.621, + 0.483, + 0.681 + ], + "angle": 0, + "content": "TacoDepth reduces depth errors by over \\(37.3\\%\\), \\(27.5\\%\\), \\(28.7\\%\\), and \\(29.1\\%\\) for MAE, RMSE, iMAE, and iRMSE on ZJU-4DRadarCam [21]. The superior efficiency and accuracy prove the effectiveness of our one-stage model design." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.685, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Comparisons with Plug-in Models. Plug-in models [21] convert relative depth to metric depth by fusing the RGB and Radar data. Our method supports both independent and plug-in processing by the auxiliary input branch. For the plug-in mode, in Table 1, TacoDepth decreases the MAE and RMSE by over \\(11.7\\%\\) and \\(13.9\\%\\) than RadarCam-Depth [21] on nuScenes [4]. Owing to their complex four-stage framework [21], we even improve the speed by \\(91.8\\%\\). Similarly, in Table 2, TacoDepth reduces the MAE, RMSE, iMAE, and iRMSE by over \\(12.8\\%\\), \\(10.8\\%\\), \\(9.6\\%\\), and \\(8.3\\%\\) on ZJU-4DaradarCam [21]. The results prove the generality and flexibility of TacoDepth, providing a better balance of model efficiency and accuracy for different users." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To be mentioned, in the plug-in mode, TacoDepth can" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.907, + 0.667 + ], + "angle": 0, + "content": "be compatible with various relative depth predictors [3, 35, 51]. It can benefit from stronger depth models [3, 51] to achieve higher accuracy, which will be presented in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.909, + 0.852 + ], + "angle": 0, + "content": "Qualitative Comparisons. The visual results of independent models [31, 37] are showcased in Fig. 4. Both daytime and nighttime samples are compared. TacoDepth can predict accurate depth maps with more complete structures (e.g., the cars and buses) and more meticulous details (e.g., the pedestrian overpass and metal guardrail). The plug-in models [21] are compared in Fig. 5. Based on the same depth predictor DPT-Hybrid [35], our method produces noticeably better structures and details in the rectangular boxes. The improvements are attributed to our graph-based Radar structure extractor and pyramid-based Radar fusion module, which extract and fuse Radar structures effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Model Robustness. Multi-stage methods [26, 28, 31, 37, 40] predict quasi-dense depth as an intermediate step. However, as shown in Fig. 2, the intermediate results are still" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10529" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.169 + ], + "angle": 0, + "content": "
TypeMethodParams (M)↓FLOPs (G)↓Time (ms)↓
Indep-endentDORN [28] (ICIP'21)183.38923.47392.5
Singh et al. [37] (CVPR'23)22.81502.0994.2
TacoDepth (Ours)13.47139.3026.7
Plug-inRadarCam-Depth [21] (ICRA'24)33.26619.02358.3
TacoDepth (Ours)14.25139.8729.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.177, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Table 3. Model efficiency. We report FLOPs, model parameters, and runtime of different methods. Since varied depth predictors can be applied, the results of plug-in models do not contain initial depth prediction. FLOPs and runtime are evaluated with one \\(900 \\times 1600\\) image and 30 Radar points on one NVIDIA A6000 GPU." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.253, + 0.481, + 0.314 + ], + "angle": 0, + "content": "
InitialRadarCam-Depth [21]TacoDepth (Ours)
δ1↑Rel↓MAE↓δ1↑Rel↓MAE↓δ1↑Rel↓MAE↓
DPT-Hybrid [35]0.9030.092-0.9210.0871157.00.9320.076983.1
MiDaS-v3.1 [3]0.9170.081-0.9260.0861177.20.9440.071852.4
Depth-Anything-v2 [51]0.9320.069-0.9400.073924.70.9610.059730.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.365 + ], + "angle": 0, + "content": "Table 4. Plug-in models with varied depth predictors. Our method outperforms RadarCam-Depth [21] with different depth predictors [3, 35, 51] on ZJU-4DRadarCam [21] (0-70 meters)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.382, + 0.483, + 0.518 + ], + "angle": 0, + "content": "sparse and noisy, especially on challenging nighttime and glaring scenes, leading to disrupted structures, blurred details, and obvious artifacts in their final depth predictions, as presented in Fig. 4. In contrast, we propose the one-stage TacoDepth without the quasi-dense depth. By effectively integrating the complementary image details and Radar structures, our model produces robust depth for both daytime and nighttime scenarios in Fig. 4. See supplement for more quantitative and visual results on nighttime scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.528, + 0.258, + 0.545 + ], + "angle": 0, + "content": "4.4. Model Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.611 + ], + "angle": 0, + "content": "We evaluate the processing time of different methods in Table 1. Compared with the second-best independent model of Li et al. [20] and the plug-in model RadarCam-Depth [21], TacoDepth improves inference speed by \\(60.5\\%\\) and \\(91.8\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.702 + ], + "angle": 0, + "content": "To further showcase our efficiency, we report model parameters and FLOPs in Table 3. Compared with the two-stage Singh et al. [37], the one-stage TacoDepth reduces the parameters and FLOPs by \\(40.9\\%\\) and \\(72.2\\%\\). Compared with the four-stage plug-in model RadarCam-Depth [21], we decrease parameters and FLOPs by \\(57.1\\%\\) and \\(77.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Note that, the differences between the plug-in and independent modes of our method arise from the auxiliary branch to fuse initial relative depth. TacoDepth achieves real-time processing of 37.5 fps by independent inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.255, + 0.788 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Plug-in Inference. In Table 4, our TacoDepth outperforms the RadarCam-Depth [21] with three different relative depth predictors [3, 35, 51]. Using more advanced depth predictors, e.g., MiDaS-v3.1 [3] and Depth-Anything-v2 [51], can further boost our accuracy without extra effort. The results demonstrate the versatility and flexibility of our approach. Efficacy of TacoDepth. In Table 5, we ablate the graph" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.162 + ], + "angle": 0, + "content": "
MethodRGBRadarGEPFMAE↓RMSE↓
RGB Baseline✓✗✗✗2474.35402.1
Singh et al. [37] (CVPR'23)✓✓✗✗2073.24590.7
TacoDepth (w/o GE)✓✓✗✓1815.64189.8
TacoDepth (Ours)✓✓✓✓1712.63960.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.166, + 0.907, + 0.25 + ], + "angle": 0, + "content": "Table 5. Efficacy of the TacoDepth. We ablate the effectiveness of graph-based Radar structure extractor (GE) and pyramid-based Radar fusion (PF) for Radar structure extraction and integration. We adopt the same baseline as Singh et al. [37]. TacoDepth \\((w / o\\) GE) uses the same MLP to extract Radar point features as prior art [37]. The results are evaluated on nuScenes [4] in 0-70 meters." + }, + { + "type": "table", + "bbox": [ + 0.57, + 0.256, + 0.851, + 0.307 + ], + "angle": 0, + "content": "
MetricL=1L=2L=3L=4
MAE↓1840.61787.81712.61729.3
RMSE↓4203.74046.93960.53984.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.31, + 0.905, + 0.325 + ], + "angle": 0, + "content": "Table 6. Layer number L. Results are on nuScenes [4] in \\(0 - 70\\mathrm{m}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.905, + 0.434 + ], + "angle": 0, + "content": "based Radar structure extractor (GE) and pyramid-based Radar fusion module (PF). With the same MLP to extract Radar point features [37], TacoDepth \\((w / o\\) GE) reduces MAE and RMSE by \\(12.4\\%\\) and \\(8.7\\%\\) than Singh et al. [37], which are attributed to the PF. Replacing the MLP [37] with GE further decreases MAE and RMSE by \\(5.7\\%\\) and \\(5.5\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.905, + 0.496 + ], + "angle": 0, + "content": "We adopt the RGB baseline of Singh et al. [37], a common single-image depth U-Net. Overall, TacoDepth improves accuracy beyond the baseline by \\(30.8\\%\\) and \\(26.7\\%\\) (MAE and RMSE), proving the efficacy of our approach." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.496, + 0.905, + 0.587 + ], + "angle": 0, + "content": "Layer Number \\( L \\). In Table 6, we ablate the layer number \\( L \\) of GE and PF. Compared with shallow layers \\( (L = 1,2) \\), deeper layers \\( (L = 3) \\) better capture structures of point clouds [41, 45], producing higher accuracy. Due to Radar sparsity, more layers \\( (L = 4) \\) does not yield improvements but decreases efficiency. We adopt \\( L = 3 \\) in experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.634, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.825 + ], + "angle": 0, + "content": "In this paper, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation model with one-stage fusion. Different from the multi-stage methods with quasidense depth, TacoDepth utilizes the graph-based Radar structure extractor and pyramid-based Radar fusion module to capture and integrate Radar features, achieving superior efficiency and robustness. Moreover, TacoDepth is flexible for independent and plug-in processing, offering a better balance of speed and accuracy. Our work provides a new perspective on the efficient Radar-Camera depth estimation. Limitations. We currently provide one efficient implementation of our TacoDepth. In future work, more techniques can be explored for varied performances and applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. The study is supported under RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 1 grant (RG14/22)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10530" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] C Bradford Barber, David P Dobkin, and Hannu Huhdanpaa. The quickhull algorithm for convex hulls. ACM Transactions on Mathematical Software (TOMS), 22(4):469-483, 1996. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.484, + 0.225 + ], + "angle": 0, + "content": "[2] Oded Bialer and Yuval Haitman. Radsimreal: Bridging the gap between synthetic and real data in radar object detection with simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15407-15416, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.226, + 0.483, + 0.279 + ], + "angle": 0, + "content": "[3] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.282, + 0.484, + 0.377 + ], + "angle": 0, + "content": "[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11618-11628, 2020. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[5] Yujeong Chae, Hyeonseong Kim, and Kuk-Jin Yoon. Towards robust 3d object detection with lidar and 4d radar fusion in various weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15162-15172, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[6] Feipeng Chen, Lihui Wang, Pinzhang Zhao, and Jian Wei. 3d grid-based vdbscan clustering and radar-monocular depth completion. IEEE Sensors Journal, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[7] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. 35, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[9] Fangqiang Ding, Andras Palffy, Dariu M. Gavrila, and Chris Xiaoxuan Lu. Hidden gems: 4d radar scene flow learning using cross-modal supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9340-9349, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.642, + 0.484, + 0.71 + ], + "angle": 0, + "content": "[10] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6824-6835, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.711, + 0.483, + 0.779 + ], + "angle": 0, + "content": "[11] Kexue Fu, Shaolei Liu, Xiaoyuan Luo, and Manning Wang. Robust point cloud registration framework based on deep graph matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8893-8902, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.78, + 0.483, + 0.848 + ], + "angle": 0, + "content": "[12] Stefano Gasperini, Patrick Koch, Vinzensz Dallabetta, Nassir Navab, Benjamin Busam, and Federico Tombari. R4dyn: Exploring radar for self-supervised monocular depth estimation of dynamic scenes. In International Conference on 3D Vision (3DV), pages 751-760, 2021. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.849, + 0.483, + 0.904 + ], + "angle": 0, + "content": "[13] Kyle Harlow, Hyesu Jang, Timothy D. Barfoot, Ayoung Kim, and Christoffer Heckman. A new wave in robotics: Survey on recent mmwave radar applications in robotics. IEEE Transactions on Robotics, 40:4544-4560, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.904 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[15] Daniel Casado Herreraz, Matthias Zeller, Le Chang, Ignacio Vizzo, Michael Heidingsfeld, and Cyril Stachniss. Radar-only odometry and mapping for autonomous vehicles. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 539-548. IEEE, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[16] Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Implicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24118-24129, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[17] Andrey Ignatov, Grigory Malivenko, David Plowman, Samarth Shukla, and Radu Timofte. Fast and accurate single-image depth estimation on mobile devices, mobile ai 2021 challenge: Report. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2545-2557, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[18] Yinuo Jiang, Beitong Zhou, Xiaoyu Liu, Qingyi Li, and Cheng Cheng. Gtinet: Global topology-aware interactions for unsupervised point cloud registration. IEEE Transactions on Circuits and Systems for Video Technology, 34(7): 6363-6375, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.904, + 0.513 + ], + "angle": 0, + "content": "[19] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, 2012. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[20] Huadong Li, Minhao Jing, Jiajun Liang, Haoqiang Fan, and Renhe Ji. Sparse beats dense: Rethinking supervision in radar-camera depth completion. arXiv preprint arXiv:2312.00844, 2023. 1, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[21] Han Li, Yukai Ma, Yaqing Gu, Kewei Hu, Yong Liu, and Xingxing Zuo. Radarcam-depth: Radar-camera fusion for depth estimation with learned metric scale. arXiv preprint arXiv: 2401.04325, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[22] Jiaqi Li, Yiran Wang, Zihao Huang, Jinghong Zheng, Ke Xian, Zhiguo Cao, and Jianming Zhang. Diffusion-augmented depth prediction with sparse annotations. In Proceedings of the 31st ACM International Conference on Multimedia, page 2865-2876, New York, NY, USA, 2023. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.713, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[23] Jiaqi Li, Yiran Wang, Jinghong Zheng, Zihao Huang, Ke Xian, Zhiguo Cao, and Jianming Zhang. Self-distilled depth refinement with noisy poisson fusion, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.756, + 0.905, + 0.824 + ], + "angle": 0, + "content": "[24] Xingyi Li, Chaoyi Hong, Yiran Wang, Zhiguo Cao, Ke Xian, and Guosheng Lin. Symmernf: Learning to explore symmetry prior for single-view view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1726-1742, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.826, + 0.905, + 0.882 + ], + "angle": 0, + "content": "[25] Zecheng Li, Yuying Song, Fuyuan Ai, Chunyi Song, and Zhiwei Xu. Semantic-guided depth completion from monocular images and 4d radar data. IEEE Transactions on Intelligent Vehicles, pages 1-13, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10531" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] Juan-Ting Lin, Dengxin Dai, and Luc Van Gool. Depth estimation from monocular images and sparse radar data. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10233-10240. IEEE, 2020. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[27] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. pages 14928-14937, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[28] Chen-Chou Lo and Patrick Vandewalle. Depth estimation from monocular images and sparse radar using deep ordinal regression network. In IEEE International Conference on Image Processing (ICIP), pages 3343-3347, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[29] Chen-Chou Lo and Patrick Vandewalle. Rcdpt: Radar-camera fusion dense prediction transformer. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[30] Chen-Chou Lo and Patrick Vandewalle. Expanding sparse radar depth based on joint bilateral filter for radar-guided monocular depth estimation. Sensors, 24(6):1864, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[31] Yunfei Long, Daniel Morris, Xiaoming Liu, Marcos Castro, Punarjay Chakravarty, and Praveen Narayanan. Radar-camera pixel depth association for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12507-12516, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.555 + ], + "angle": 0, + "content": "[32] Guohao Peng, Heshan Li, Yangyang Zhao, Jun Zhang, Zhenyu Wu, Pengyu Zheng, and Danwei Wang. Transloc4d: Transformer-based 4d radar place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17595-17605, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.558, + 0.482, + 0.821 + ], + "angle": 0, + "content": "[33] Pierluigi Zama Ramirez, Fabio Tosi, Luigi Di Stefano, Radu Timofte, Alex Costanzino, Matteo Poggi, Samuele Salti, Stefano Mattoccia, Yangyang Zhang, Cailin Wu, Zhuangda He, Shuangshuang Yin, Jiaxu Dong, Yangchenxu Liu, Hao Jiang, Jun Shi, Yong A, Yixiang Jin, Dingzhe Li, Bingxin Ke, Anton Obukhov, Tinafu Wang, Nando Metzger, Shengyu Huang, Konrad Schindler, Yachuan Huang, Jiaqi Li, Junrui Zhang, Yiran Wang, Zihao Huang, Tianqi Liu, Zhiguo Cao, Pengzhi Li, Jui-Lin Wang, Wenjie Zhu, Hui Geng, Yuxin Zhang, Long Lan, Kele Xu, Tao Sun, Qisheng Xu, Sourav Saini, Aashray Gupta, Sahaj K. Mistry, Aryan Shukla, Vinit Jakhetiya, Sunil Jaiswal, Yuejin Sun, Zhuofan Zheng, Yi Ning, Jen-Hao Cheng, Hou-I Liu, Hsiang-Wei Huang, Cheng-Yen Yang, Zhongyu Jiang, Yi-Hao Peng, Aishi Huang, and Jenq-Neng Hwang. Ntire 2024 challenge on hr depth from images of specular and transparent surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6499–6512, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.823, + 0.482, + 0.891 + ], + "angle": 0, + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(03):1623-1637, 2020. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.892, + 0.482, + 0.904 + ], + "angle": 0, + "content": "[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.904 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "sion transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12179-12188, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[36] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks and Learning Systems, 20(1):61-80, 2009. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[37] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9275-9285, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.345 + ], + "angle": 0, + "content": "[38] Tieshuai Song, Bin Yang, Jun Wang, Guidong He, Zhao Dong, and Fengjun Zhong. mmwave radar and image fusion for depth completion: a two-stage fusion network. In The 27th International Conference on Information Fusion (FUSION), 2024. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[39] Christian Stetco, Barnaba Ubezio, Stephan Muhlbacher-Karrer, and Hubert Zangl. Radar sensors in collaborative robotics: Fast simulation and experimental validation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10452-10458. IEEE, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.906, + 0.471 + ], + "angle": 0, + "content": "[40] Huawei Sun, Hao Feng, Julius Ott, Lorenzo Servadei, and Robert Wille. Cafnet: A confidence-driven framework for radar camera depth estimation. arXiv preprint arXiv:2407.00697, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.906, + 0.556 + ], + "angle": 0, + "content": "[41] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6411–6420, 2019. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.906, + 0.613 + ], + "angle": 0, + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.906, + 0.67 + ], + "angle": 0, + "content": "[43] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3056-3065, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.906, + 0.727 + ], + "angle": 0, + "content": "[44] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3056-3065, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.906, + 0.783 + ], + "angle": 0, + "content": "[45] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5), 2019. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.908, + 0.854 + ], + "angle": 0, + "content": "[46] Yiran Wang, Xingyi Li, Min Shi, Ke Xian, and Zhiguo Cao. Knowledge distillation for fast and accurate monocular depth estimation on mobile devices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2457-2465, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.856, + 0.906, + 0.897 + ], + "angle": 0, + "content": "[47] Yiran Wang, Zhiyu Pan, Xingyi Li, Zhiguo Cao, Ke Xian, and Jianming Zhang. Less is more: Consistent video depth estimation with masked frames modeling. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10532" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "of the 30th ACM International Conference on Multimedia, page 6347-6358, New York, NY, USA, 2022. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[48] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9466-9476, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[49] Yiran Wang, Min Shi, Jiaqi Li, Chaoyi Hong, Zihao Huang, Juewen Peng, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Nvds+: Towards efficient and versatile neural stabilizer for video depth estimation. IEEE transactions on pattern analysis and machine intelligence, pages 1-18, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.288, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. 34:12077-12090, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[51] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[52] Junrui Zhang, Jiaqi Li, Yachuan Huang, Yiran Wang, Jinghong Zheng, Liao Shen, and Zhiguo Cao. Towards robust monocular depth estimation in non-lambertian surfaces, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[53] Yi Zhou, Lulu Liu, Haocheng Zhao, Miguel López-Benítez, Limin Yu, and Yutao Yue. Towards deep radar perception for autonomous driving: Datasets, methods, and challenges. Sensors, 22(11):4208, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10533" + } + ] +] \ No newline at end of file diff --git a/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_origin.pdf b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ca73b8424acb6b8d8bb7e4eb0fdb9777d9e0268b --- /dev/null +++ b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/60893af7-fdb2-4c8a-88de-cc24df777172_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d01130d545bf02d77fac18412efeabcdaa708045c80bcdf49cde664c364a832 +size 3522096 diff --git a/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/full.md b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac1ec6f1877c8874ff2596e76e9ce7796d76dec3 --- /dev/null +++ b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/full.md @@ -0,0 +1,297 @@ +# TacoDepth: Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion + +Yiran Wang $^{1,2,*}$ Jiaqi Li $^{2,*}$ Chaoyi Hong $^{1,2}$ Ruibo Li $^{1}$ +Liusheng Sun $^{3}$ Xiao Song $^{3}$ Zhe Wang $^{3}$ Zhiguo Cao $^{2}$ Guosheng Lin $^{1,\dagger}$ $^{1}$ S-Lab, Nanyang Technological University + +$^{2}$ School of AIA, Huazhong University of Science and Technology $^{3}$ SenseTime Research + +{wangyiran,zgcao}@hust.edu.cn, gslin@ntu.edu.sg + +https://github.com/RaymondWang987/TacoDepth + +# Abstract + +Radar-Camera depth estimation aims to predict dense and accurate metric depth by fusing input images and Radar data. Model efficiency is crucial for this task in pursuit of real-time processing on autonomous vehicles and robotic platforms. However, due to the sparsity of Radar returns, the prevailing methods adopt multi-stage frameworks with intermediate quasi-dense depth, which are time-consuming and not robust. To address these challenges, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation model with one-stage fusion. Specifically, the graph-based Radar structure extractor and the pyramid-based Radar fusion module are designed to capture and integrate the graph structures of Radar point clouds, delivering superior model efficiency and robustness without relying on the intermediate depth results. Moreover, TacoDepth can be flexible for different inference modes, providing a better balance of speed and accuracy. Extensive experiments are conducted to demonstrate the efficacy of our method. Compared with the previous state-of-the-art approach, TacoDepth improves depth accuracy and processing speed by $12.8\%$ and $91.8\%$ . Our work provides a new perspective on efficient Radar-Camera depth estimation. + +# 1. Introduction + +Radar-Camera depth estimation [6, 12, 20, 21, 25, 26, 28-31, 37] is a vital task for 3D perception and motion planning in autonomous driving [4, 15, 22, 49, 53] and robotics [13, 21, 39]. The task aims to predict dense metric depth by integrating images with mmWave Radar data, perceiving metric scale in different environments. As a typical range sensor on driving and robotic platforms, Radar is widely equipped and used for its high affordability and low power consump + +![](images/5ea893d73a9f562e8ff7a9996d340f22ad41f96b1382fcc6f7a46389e4b8e5f5.jpg) +Figure 1. Performance and efficiency. Circle area indicates inference time (ms). Smaller circles showcase faster speed. The X-axis and Y-axis represent MAE and RMSE metrics for depth errors on the nuScenes [4] dataset. Lower MAE and RMSE mean higher accuracy. Our TacoDepth outperforms prior arts by large margins. + +tion. It is valuable to explore Radar-Camera fusion in various tasks [2, 5, 9, 16, 32, 37] for its superior all-weather reliability. Besides, for real-time processing on vehicles and robotics, improving efficiency is crucial for fusion models. + +The obstacle for fusing Radar in depth estimation lies in its sparsity and noises. Due to the large beam width, Radar point clouds are $1000\mathrm{x}$ sparser with higher noise levels than LiDAR returns. To deal with the issue, the prevailing methods [21, 26, 28, 29, 31, 37, 38, 40] adopt complex multi-stage frameworks. Intermediate quasi-dense depth is predicted to bridge the gap between sparse Radar and final dense depth. For instance, two-stage methods [26, 28, 29, 31, 37, 38, 40] predict final depth based on intermediate results, as independent models without utilizing other depth predictors [3, 34, 35, 51]. RadarCam-Depth [21], as a plugin module, converts relative depth [34, 35] to metric scale in four stages by post-processing. These multi-stage methods are inefficient and time-consuming, as presented in Fig. 1. + +Meanwhile, despite being described as quasi-dense, the intermediate depth [21, 26, 28, 29, 31, 37, 40] remains sparse and noisy. As shown in Fig. 2, only few pixels exhibit valid depth. For some adverse lighting or weather conditions, e.g., glaring and nighttime scenarios, even no pixels are predicted with depth values. These defective intermediate results lead to disrupted structures, blurred details, or noticeable artifacts in final predictions, as shown in Fig. 4. With these observations, a question naturally arises: can models directly predict accurate and dense depth without the intermediate results, achieving one-stage Radar-Camera depth estimation with higher efficiency and robustness? + +To answer the above question, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation framework with one-stage fusion. To simplify multi-stage fusion [21, 26, 28, 29, 31, 37, 38, 40] into one stage, TacoDepth needs to extract and integrate Radar information more effectively. Prior arts simply extract features from Radar coordinates [21, 37] or projected depth [26, 31], sensitive to noises and outliers [11]. In contrast, we utilize a graph-based Radar structure extractor to capture geometric structures and topologies of point clouds, improving robustness to inaccurate points [11, 18, 36, 43]. After that, our pyramid-based Radar fusion module integrates image and Radar features hierarchically. Shallow layers capture image details and Radar coordinates, while deep layers fuse scene semantics and structures [10, 14, 19, 41, 45, 50]. In each layer, Radar-centered flash attention mechanism is applied to establish the cross-modal correlations. For one Radar point, lightweight and accelerated flash attention [8] is calculated within Radar-centered areas for efficiency. Finally, TacoDepth predicts accurate and dense depth by a decoder. + +Moreover, we further analyze the characteristics of different inference strategies in prior arts [21, 26, 28, 29, 31, 37, 38, 40]. Independent models [26, 28, 29, 31, 37, 40] are more efficient without initial depth prediction and postprocessing, while the plug-in model [21] can be more accurate by leveraging cutting-edge relative depth predictors [3, 23, 34, 51]. Therefore, TacoDepth aims to be flexible for the two settings. An auxiliary and optional branch is added for initial depth, allowing TacoDepth to support both independent and plug-in processing. Our approach can better balance depth accuracy and model efficiency. + +We conduct evaluations on nuScenes [4] and ZJU-4DRadarCam [21] datasets with different depth ranges, following prior arts [21, 31, 37]. As shown in Fig. 1, whether compared with independent [26, 28, 31, 37, 40] or plugin in [21] models, TacoDepth significantly outperforms prior arts [21, 26, 28, 31, 37, 40] in accuracy and efficiency. Besides, our approach achieves real-time processing of over 37 fps as an independent model and strong compatibility with various depth predictors [3, 34, 35, 51] as a plug-in model, further proving the efficiency and flexibility of TacoDepth. + +![](images/1e7180a4707dc7ac431aa404e436ae2c9b86da99bc267c0cfa0a28614cba35cc.jpg) +Figure 2. Intermediate quasi-dense depth maps [21, 26, 28, 29, 31, 37, 40] remain sparse and noisy. Pixels with valid depth values are visualized by the gray areas in the red rectangular boxes. + +Our main contributions can be summarized as follows: + +- We propose TacoDepth, an efficient and accurate Radar-Camera depth estimation method with one-stage fusion. +- We design the pyramid-based Radar fusion module, fusing Radar and image features effectively and efficiently. +- TacoDepth is flexible for both independent and plug-in processing, balancing model efficiency and performance. + +# 2. Related Work + +MmWave Radar for Vehicles and Robotics. Autonomous driving [4, 15, 48, 53] and robotics [13, 21, 24, 39] rely on multiple sensors to perceive the environment, e.g., camera, LiDAR, and mmWave Radar. LiDAR and Radar provide the metric scale of different scenarios as a complement to cameras. Compared with LiDAR, Radar is cheaper and more widely-equipped with lower power consumption. The reflection properties of mmWave also provide Radar with farther sensing ranges and stronger robustness to all-weather conditions. The main drawback of Radar lies in its sparsity and noise. Due to the large beam width, Radar point clouds are $1000\mathrm{x}$ sparser [37] than LiDAR. Besides, Radar captures coordinates and depth values with lower accuracy. For instance, 3D Radar suffers from height ambiguity [31] due to insufficient antenna elements along the elevation axis, while this problem can be alleviated by the emerging 4D Radar. Thus, integrating Radar data is valuable for various tasks [2, 5, 9, 16, 27, 32, 37] in driving and robotic systems. + +Radar-Camera Depth Estimation. The task aims to predict dense metric depth by merging images and Radar, facilitating 3D perception and motion planning of vehicles and robotics. Model efficiency is crucial for these real-time applications. However, to overcome the sparsity and noise of Radar, prior arts [21, 26, 28, 29, 31, 37, 38, 40] adopt inefficient multi-stage frameworks. The final depth is predicted based on intermediate quasi-dense depth. These methods can be divided into independent and plug-in models. Most independent models [26, 28, 29, 31, 37, 40] predict final depth in two stages without relying on other depth predictors [17, 33-35, 46, 47]. For example, Lin et al. [26] and DORN [28] use two CNNs [14] for quasi-dense and final depth. RC-PDA [31] and RC-DPT [29] convert intermediate results into multiple channels for varied confidence levels. Singh et al. [37] propose gated fusion to filter the inac + +![](images/7af03cc5a3f094eef113e03d9d84f96b8eeaef12d4f649907d8ef409a66c004b.jpg) +Figure 3. Overview of the TacoDepth. The graph-based Radar structure extractor captures graph structures of Radar point clouds through the node feature $N_{l}$ and edge feature $E_{l}$ in a certain layer $l$ . The pyramid-based Radar fusion module integrates image and Radar features in a pyramidal hierarchical manner. In each layer, to efficiently build cross-modal correspondences, the Radar-centered flash attention is calculated within Radar-centered areas based on horizontal coordinates, e.g., pixels in the orange area as queries, while Radar points in the red area as keys and values. Our TacoDepth achieves efficient and accurate Radar-Camera depth estimation in one stage. During inference, the model is flexible and supports both independent and plug-in processing, facilitating a better balance of model efficiency and accuracy. + +![](images/8ece566f18f36fa5fc0ecaf50fc399d52b01522b368a268e9d4d6ec4fe6fae46.jpg) + +curate quasi-dense depth. Besides, RadarCam-Depth [21] designs a plug-in module, aligning relative depth [34, 35] to metric scale in four stages, including initial depth, global scale alignment, quasi-dense depth, and scale learner. These methods [21, 26, 28, 29, 31, 37, 38, 40] are inefficient with complex stages. Defective intermediate depth also limits their robustness. In contrast, we propose TacoDepth model to achieve efficient, accurate, and flexible one-stage fusion. + +# 3. TacoDepth + +We present a detailed illustration of TacoDepth, an efficient Radar-Camera depth estimation framework with one-stage fusion. In Sec. 3.1, we provide an overall description to outline our approach. In Sec. 3.2, we elaborate on the detailed architecture of the pyramid-based Radar fusion module. TacoDepth can be flexible for independent and plug-in processing, which will be discussed in Sec. 3.3. Besides, the training procedure for TacoDepth is illustrated in Sec. 3.4. + +# 3.1. Overview + +Here we present an overview of our TacoDepth framework. Given a single image $I \in \mathcal{R}^{H \times W \times 3}$ and a Radar point cloud $P \in \mathcal{R}^{K \times 3}$ as inputs, our goal is to predict dense metric depth $D \in \mathcal{R}^{H \times W}$ . $H$ and $W$ represent image height and weight. The point cloud $P$ contains $K$ points $p \in \mathcal{R}^3$ . Radar-Camera depth estimation can be divided into two primary subtasks: the extraction and the integration of Radar information. Accordingly, as depicted in Fig. 3, + +TacoDepth comprises a graph-based Radar structure extractor and a pyramid-based Radar fusion module to undertake these two important steps respectively. + +Firstly, for Radar feature extraction, previous methods [21, 37] simply extract features from the coordinates of individual Radar points, focusing solely on the point-level information. For example, Singh et al. [37] extract point features in 32,256 dimensions from the three-dimensional coordinates using a multi-layer perception (MLP), which inevitably introduces substantial redundancy and noise. + +To achieve efficient Radar-Camera fusion in a single stage, TacoDepth needs to extract more comprehensive and effective Radar information, prompting us to reconsider the complementarity between camera images and Radar data. Compared to the coordinates, the geometric structures of point clouds, e.g., the distances or relations between Radar points, are potentially more informative for Radar-Camera fusion. The overall topologies are also more robust to inaccurate outliers [11]. To this end, we employ the graph-based Radar structure extractor with a lightweight GNN [11, 36, 43] architecture. Specifically, Radar points are regarded as the graph nodes. The adjacency matrix of all points represents the graph edges. In each layer, the node features of Radar points can be updated and aggregated along edges by PCA-GM [44]. The graph-based Radar structure extractor can capture point coordinates and graph structures from shallow layers to deep layers [11, 41, 45]. + +After extracting the Radar structures, the subsequent step + +is to merge RGB and Radar features. Prior arts [21, 26, 28, 29, 31, 37, 38, 40] adopt multi-stage methods with intermediate quasi-dense depth, which are inefficient and not robust as illustrated in Sec. 1. In contrast, we propose our pyramid-based Radar fusion module for efficient one-stage fusion. + +To be specific, features of the two modalities are integrated in a pyramidal hierarchical manner. Shallow layers can merge image details with Radar coordinates, whereas deeper layers could fuse scene semantics and geometric structures [10, 14, 19, 41, 45, 50]. In each layer, we leverage Radar-centered flash attention for cross-modal correlations and feature fusion. The queries are derived from image pixels, while keys and values originate from Radar points. For each point, we delineate Radar-centered areas based on the horizontal coordinates, as exemplified by the orange and red areas in Fig. 3, given that the projected horizontal coordinates of Radar points are generally more precise than the height dimension [31, 37]. Lightweight flash attention [8] is calculated within these areas to maintain efficiency, since external pixels and points can not be correlated. Finally, with the fused features from all layers, TacoDepth predicts accurate and dense metric depth by a depth decoder [37]. + +Furthermore, TacoDepth is adaptable for both independent and plug-in inference. An optional input branch is introduced to handle the initial relative depth. For the independent inference, our model achieves real-time processing of over 37 fps. For the plug-in manner, TacoDepth demonstrates compatibility with various cutting-edge depth predictors [3, 34, 35, 51], as proved in Sec. 4.5. Consequently, our approach can deliver a balanced trade-off between performance and efficiency for different users and applications. + +# 3.2. Pyramid-based Radar Fusion + +For the Radar point cloud $P$ , our graph-based Radar structure extractor derives node feature $N_{l}$ and edge feature $E_{l}$ at a certain layer $l \in \{1,2,\dots ,L\}$ , where $L$ represents the layer number of the GNN module [11, 36, 43]. For the input image $I$ , the image encoder [14, 21, 37] extracts multi-scale features $\{F_1,F_2,\dots ,F_{2L}\}$ . The pyramid-based Radar fusion module hierarchically integrates the multi-modal features. $N_{l}$ and $E_{l}$ will be fused with image features in adjacent layers. We merge the node feature $N_{l}$ with image feature $F_{2l - 1}$ , and the edge feature $E_{l}$ with $F_{2l}$ , respectively. + +The prerequisite for feature fusion is cross-modal correspondences between Radar points and RGB pixels. Due to the noisy elevation coordinates of 3D Radar, prior arts [21, 29, 31, 37, 40] try to build one-to-many mappings by the intermediate quasi-dense depth. They adopt explicit classification supervision with the depth discrepancies between Li-DAR and Radar as pseudo-labels. However, the quasi-dense depth prediction leads to inefficiency and erroneous final outcomes, as discussed in Sec. 1. Given the absence of reliable explicit labels and supervision, we hope our TacoDepth + +can learn and establish the correspondences implicitly. + +Therefore, for efficient Radar-Camera correlations, we propose the Radar-centered flash attention in TacoDepth. We define Radar-centered areas based on the precise horizontal coordinates of Radar points and image pixels. For each point, the attention is computed only with image pixels in the area, since external pixels can not be correlated. Here we take $E_{l} \in \mathcal{R}^{K \times K}$ and $F_{2l} \in \mathcal{R}^{h \times w \times C_l}$ as examples, where $h$ and $w$ are feature height and width. $K$ and $C_l$ refer to the point number and layer feature channel. For a Radar point $p$ with the horizontal coordinate $x_{p}$ , as showcased by the exemplary orange area in Fig. 3, we only retain image pixels $m$ whose coordinates $x_{m}$ are close to $x_{p}$ : + +$$ +\hat {F} _ {2 l} = F _ {2 l} \left[ x _ {p} - a _ {l} < x _ {m} < x _ {p} + a _ {l} \right], \hat {F} _ {2 l} \in \mathcal {R} ^ {\hat {M} \times C _ {l}}, \tag {1} +$$ + +where $a_{l}$ indicates the width of the Radar-centered area in layer $l$ . The notation $[\cdot]$ depicts the fetching operation of pixels or points. $\hat{F}_{2l}$ is the image feature of the remaining $\hat{M}$ pixels to generate queries. For a random pixel $m$ in $\hat{F}_{2l}$ , as represented by the red area in Fig. 3, the attention calculation is also restricted to adjacent Radar points: + +$$ +\hat {E} _ {l} = E _ {l} \left[ x _ {m} - a _ {l} < x _ {p} < x _ {m} + a _ {l} \right], \hat {E} _ {l} \in \mathcal {R} ^ {\hat {K} \times K}. \tag {2} +$$ + +The feature dimension maintains the original $K$ to preserve integral graph topologies. $\hat{E}_l$ is the edge feature of the retained $\hat{K}$ points for keys and values. In this way, the computational costs can be significantly decreased with the small subset of $\hat{M}$ pixels and $\hat{K}$ points. For the pixel $m$ , we generate queries from $\hat{F}_{2l}[m] \in \mathcal{R}^{C_l}$ , while keys and values are derived from $\hat{E}_l$ . Radar-centered attention is computed as: + +$$ +F _ {2 l} ^ {\prime} [ m ] = \operatorname {s o f t m a x} \frac {W _ {q} \hat {F} _ {2 l} [ m ] \left(W _ {k} \hat {E} _ {l}\right) ^ {T}}{\sqrt {C _ {l}}} W _ {v} \hat {E} _ {l}, \tag {3} +$$ + +in which $W_{q}$ , $W_{k}$ , and $W_{v}$ are learnable linear projections. The fused Radar-Camera feature $F_{2l}^{\prime} \in \mathcal{R}^{h \times w \times C_l}$ can be acquired for the current layer. The above calculation is incorporated into attention blocks with MLPs and residual connections. To further improve model efficiency, we employ lightweight and accelerated flash attention modules [7, 8] to replace the original attention computation [42]. + +Similarly, the node feature $N_{l}$ can be merged with the image feature $F_{2l-1}$ by slightly adjusting the dimensions of linear projections. Ultimately, the acquired Radar-Camera features $\{F_1', F_2', \dots, F_{2L}'\}$ are fed to a common depth decoder [21, 37] to predict accurate metric depth. + +As a consequence, the proposed pyramid-based Radar fusion module enables efficient and effective Radar-Camera fusion in a single stage. The Radar-centered flash attention establishes cross-modal correlations with minimal computational costs. Without relying on the previous multi-stage frameworks and intermediate quasi-dense depth [21, 26, 28, 29, 31, 37, 38, 40], TacoDepth achieves higher efficiency and stronger robustness for Radar-Camera depth estimation. + +# 3.3. Flexible Inference + +Considering the varied inference manners in prior arts [21, 26, 28, 29, 31, 37, 38, 40], the plug-in module [21] could achieve higher accuracy using state-of-the-art relative depth predictors [34, 35, 52], while independent models [26, 28, 29, 31, 37, 38, 40] are capable of attaining faster inference speed without initial depth prediction and post-processing. + +Therefore, TacoDepth is designed to be flexible, supporting both independent and plug-in predictions. For the plug-in inference, an initial relative depth map is converted to metric scale by fusing camera and Radar information. If we denote the TacoDepth model as $\mathcal{T}_{\theta}$ , parameterized by $\theta$ , the depth prediction process can then be formulated as follows: + +$$ +D = \mathcal {T} _ {\theta} (I, P \mid D ^ {*}), \tag {4} +$$ + +where the vertical bar $(|)$ separates the required and optional inputs. $D$ is the predicted depth of our model. $D^{*}$ is the optional input of initial relative depth from depth predictors [3, 34, 35, 51]. We introduce an auxiliary branch to process and fuse $D^{*}$ with image features, while keeping all other settings identical to the independent mode. Our plugin in manner demonstrates compatibility with various depth predictors, e.g., MiDaS [3, 34], DPT [35], and Depth-Anything-v2 [51]. In this way, TacoDepth can achieve a better balance of model efficiency and accuracy, offering superior flexibility for different users and applications. + +# 3.4. Training the TacoDepth + +In the training phase, the training data [4, 21] is randomly divided into two equal portions in each epoch. For the auxiliary branch, one half of the data is provided with initial relative depth, while the other half adopts zero as input. This training strategy facilitates the simultaneous optimization of the independent inference and plug-in prediction. + +As for the training loss, we employ the same depth ground truth and $L_{1}$ loss as the prior art [37]. Specifically, LiDAR-captured depth $D_{gt}$ and reprojection-accumulated depth $D_{acc}$ are utilized for supervision [21, 31, 37]. The overall loss function $\ell_{L_1}$ can be expressed as follows: + +$$ +\ell_ {L _ {1}} = \frac {1}{| \Omega_ {g t} |} \sum_ {\Omega_ {g t}} | D - D _ {g t} | + \frac {\lambda}{| \Omega_ {a c c} |} \sum_ {\Omega_ {a c c}} | D - D _ {a c c} |, (5) +$$ + +where $\Omega_{gt}$ and $\Omega_{acc}$ depict the masks of $D_{gt}$ and $D_{acc}$ with valid depth values. $\lambda$ is the coefficient of the second term. + +# 4. Experiments + +To prove the efficacy of our TacoDepth, we conduct experiments on the nuScenes [4] and ZJU-4DRadarCam [21] datasets following prior arts [20, 21, 26, 28, 29, 31, 37, 40]. In Sec. 4.1, we describe the experimental datasets and evaluation protocols. The implementation details of our model + +are illustrated in Sec. 4.2. In Sec. 4.3, we compare the TacoDepth with previous Radar-Camera depth estimation approaches to demonstrate our state-of-the-art performance. The inference speed and model efficiency of different methods are evaluated and discussed in Sec. 4.4. In Sec. 4.5, we perform ablation studies to expound on our specific designs. + +# 4.1. Datasets and Evaluation Protocols + +Due to the scarcity of public datasets for Radar-Camera depth estimation, prior works [20, 26, 28, 29, 31, 37, 38, 40] only conduct experiments on nuScenes dataset [4] with 3D Radar. Recently, RadarCam-Depth [21] releases their ZJU-4DRadarCam dataset [21] with 4D Radar to alleviate this problem. We adopt these two datasets [4, 21] for experiments, proving our effectiveness on both 3D and 4D Radar. NuScenes Dataset. As an outdoor driving dataset, the nuScenes [4] is collected by vehicles equipped with cameras, 3D Radar, LiDAR, and IMU. It contains 1,000 scenes and around 40,000 synchronized Radar-Camera keyframes captured in Boston and Singapore. We follow the same train-test split as prior arts [21, 26, 29, 31, 37, 40], with 700 scenes for training, 150 for validation, and 150 for testing. + +ZJU-4DRadarCam Dataset. The dataset [21] is collected by a ground robot fitted with more advanced camera, 4D Radar, and LiDAR sensors, offering denser Radar data and LiDAR depth than the nuScenes [4] dataset. It comprises 33,409 Radar-Camera keyframes in total. We follow the official data split [21] with 29,312 frames for training and validation, and 4,097 samples for testing in our experiments. + +Evaluation Protocols. We evaluate independent and plug-in models separately for fair comparisons on the two datasets [4, 21] with depth ranges of 50, 70, and 80 meters [21, 31, 37, 40]. On nuScenes [4], previous state-of-the-art methods [12, 20, 21, 26, 28, 31, 37, 40] are compared comprehensively to the best of our knowledge. For the recently published ZJU-4DRadarCam [21], since most prior works do not include experiments on the dataset, we can only follow the same comparative methods [21, 28, 37] and results as RadarCam-Depth [21]. Besides, we also evaluate plug-in models with different depth predictors [3, 35, 51] to showcase model flexibility on the ZJU-4DRadarCam [21]. + +Evaluation Metrics. We adopt the commonly-applied depth metrics MAE, RMSE, iMAE, iRMSE, $\delta_{1}$ , and Rel as prior arts [21, 26, 31, 37]. See supplement for more details. + +# 4.2. Implementation Details + +Data Processing. We adhere to the data processing procedures of prior arts [21, 37]. For the nuScenes [4], the $D_{acc}$ in Eq. 5 is generated by accumulating the $D_{gt}$ of 160 nearby frames [21, 37]. For the ZJU-4DRadarCam dataset [21], the $D_{acc}$ is obtained via the linear interpolation [1, 21] of $D_{gt}$ . Model Architecture. In line with prior arts [21, 37], the ResNet-18 [14] is adopted as the image encoder. The graph + +
TypeMethodRadar FramesImagesTime(ms)↓0 - 50m0 - 70m0 - 80m
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
Indep-endentLin et al. [26] (IROS'20)31117.62034.94316.52294.75338.22371.05623.0
RC-PDA [31] (CVPR'21)5379.22225.04156.53326.16700.63713.67692.8
R4Dyn [12] (3DV'21)41------6434.0
DORN [28] (ICIP'21)5(×3)1392.51926.64124.82380.65252.72467.75554.3
Singh et al. [37] (CVPR'23)1194.21727.73746.82073.24590.72179.34898.7
CaFNet [40] (IROS'24)11103.91674.23674.52010.34493.12109.84765.6
Li et al. [20] (ECCV'24)1167.61524.53567.31822.94303.61927.04609.6
TacoDepth (Ours)1126.71423.63275.81712.63960.51833.44150.2
Plug-inRadarCam-Depth [21] (ICRA'24)11358.31286.12964.31587.93662.51689.73948.0
TacoDepth (Ours)1129.31046.82487.51347.13152.81492.43324.8
+ +Table 1. Comparisons with state-of-the-art methods on the nuScenes dataset [4] (in millimeters). Independent [12, 20, 26, 28, 31, 37, 40] and plug-in [21] models are evaluated separately, with maximum evaluation distances of 50, 70, and 80 meters as prior arts [21, 37]. Plug-in models are compared using the same depth predictor DPT-Hybrid [35] as RadarCam-Depth [21]. We report the average runtime of processing one $900 \times 1600$ frame and 30 Radar points by different methods on one NVIDIA RTX A6000 GPU. The inference time of plug-in models does not contain initial depth prediction. Some methods [12, 26, 28, 31] require multiple frames and Radar scans for inference, while others [20, 21, 37, 40] only need a single frame as input. The best performance is in boldface. Second best is underlined. + +![](images/dbaae734e43abccef215ebd53b3aeb0d7c53b5feb682ca5416b2f07576172f44.jpg) +Figure 4. Visual results of independent models on nuScenes [4]. Both daytime and nighttime samples are presented. Prior arts [31, 37] exhibit disrupted structures and noticeable artifacts. TacoDepth produces accurate depth with finer details and more complete structures. + +based Radar structure extractor contains $L = 3$ layers with the GNN [11, 36] architecture. The widths $a_{l}$ of Radar-centered areas in Eq. 1 and Eq. 2 are set to $\{48, 32, 16\}$ for the three layers. Refer to the supplement for more details. + +Training Recipe. The model is trained by Adam optimizer on two NVIDIA A6000 GPUs with a batch size of 12 for 50 epochs. The initial learning rate is $1e^{-4}$ and decreases by $1e^{-5}$ for every ten epochs. The input image size is $900 \times 1600$ [21, 37] and $300 \times 1280$ [21] on nuScenes [4] and ZJU-4DRadarCam [21], respectively. The $\lambda$ in Eq. 5 is set to one. All other settings are consistent with prior works [21, 37]. + +# 4.3. Comparisons with State-of-the-art Methods + +Comparisons with Independent Models. Independent models [12, 20, 26, 28, 29, 31, 37, 38, 40] predict dense + +metric depth without using relative depth predictors [3, 34, 35, 51]. Most of them adopt two-stage methods with intermediate quasi-dense depth, which are inefficient and not robust. Our TacoDepth significantly surpasses these methods in efficiency and accuracy with one-stage fusion. For example, in Table 1, compared with the recent CaFNet [40], our model reduces RMSE, MAE, and inference time by over $10.9\%$ , $13.1\%$ , and $74.3\%$ on nuScenes dataset [4]. Additionally, Li et al. [20] aim to improve the training paradigm under sparse supervision by their Disruption-Compensation framework. Without using such advanced training and supervision strategy, TacoDepth still outperforms Li et al. [20] by over $60.5\%$ , $4.9\%$ , and $8.1\%$ for runtime, MAE, and RMSE, showing the efficacy of our proposed architectures. + +Similarly, in Table 2, compared with Singh et al. [37], + +
TypeMethod0 - 50m0 - 70m0 - 80m
MAE↓RMSE↓iMAE↓iRMSE↓MAE↓RMSE↓iMAE↓iRMSE↓MAE↓RMSE↓iMAE↓iRMSE↓
Indep-endentDORN [28] (ICIP'21)2210.24129.719.831.92402.24625.219.831.92447.64760.019.931.9
Singh et al. [37] (CVPR'23)1785.43704.618.135.31932.74137.118.035.21979.54309.317.935.1
TacoDepth (Ours)1120.12686.712.825.01181.82906.312.724.91201.12990.712.724.9
Plug-inRadarCam-Depth [21] (ICRA'24)1067.52817.410.522.91157.03117.710.422.91183.53229.010.422.8
TacoDepth (Ours)930.22477.39.320.8983.12779.69.320.91032.52850.39.420.9
+ +Table 2. Comparisons with state-of-the-art methods on ZJU-4DRadarCam [21] (in millimeters). Plug-in models are compared using the same relative depth predictor DPT-Hybrid [35]. Since many prior works do not include experiments on this new dataset, we adhere to the same compared methods [21, 28, 37] and metrics as RadarCam-Depth [21]. Best performance is in boldface. Second best is underlined. + +![](images/e46bdf0b07e2ce89f1b4d1888720194317b6101fe78e2c0d317541fad59cfe01.jpg) +Figure 5. Visual results of plug-in models [21] on the ZJU-4DRadarCam [21]. The same depth predictor DPT-Hybrid [35] is adopted for fair comparisons. Regions with obvious differences are highlighted in the rectangular boxes. Best view zoomed in on-screen for details. + +TacoDepth reduces depth errors by over $37.3\%$ , $27.5\%$ , $28.7\%$ , and $29.1\%$ for MAE, RMSE, iMAE, and iRMSE on ZJU-4DRadarCam [21]. The superior efficiency and accuracy prove the effectiveness of our one-stage model design. + +Comparisons with Plug-in Models. Plug-in models [21] convert relative depth to metric depth by fusing the RGB and Radar data. Our method supports both independent and plug-in processing by the auxiliary input branch. For the plug-in mode, in Table 1, TacoDepth decreases the MAE and RMSE by over $11.7\%$ and $13.9\%$ than RadarCam-Depth [21] on nuScenes [4]. Owing to their complex four-stage framework [21], we even improve the speed by $91.8\%$ . Similarly, in Table 2, TacoDepth reduces the MAE, RMSE, iMAE, and iRMSE by over $12.8\%$ , $10.8\%$ , $9.6\%$ , and $8.3\%$ on ZJU-4DaradarCam [21]. The results prove the generality and flexibility of TacoDepth, providing a better balance of model efficiency and accuracy for different users. + +To be mentioned, in the plug-in mode, TacoDepth can + +be compatible with various relative depth predictors [3, 35, 51]. It can benefit from stronger depth models [3, 51] to achieve higher accuracy, which will be presented in Table 4. + +Qualitative Comparisons. The visual results of independent models [31, 37] are showcased in Fig. 4. Both daytime and nighttime samples are compared. TacoDepth can predict accurate depth maps with more complete structures (e.g., the cars and buses) and more meticulous details (e.g., the pedestrian overpass and metal guardrail). The plug-in models [21] are compared in Fig. 5. Based on the same depth predictor DPT-Hybrid [35], our method produces noticeably better structures and details in the rectangular boxes. The improvements are attributed to our graph-based Radar structure extractor and pyramid-based Radar fusion module, which extract and fuse Radar structures effectively. + +Model Robustness. Multi-stage methods [26, 28, 31, 37, 40] predict quasi-dense depth as an intermediate step. However, as shown in Fig. 2, the intermediate results are still + +
TypeMethodParams (M)↓FLOPs (G)↓Time (ms)↓
Indep-endentDORN [28] (ICIP'21)183.38923.47392.5
Singh et al. [37] (CVPR'23)22.81502.0994.2
TacoDepth (Ours)13.47139.3026.7
Plug-inRadarCam-Depth [21] (ICRA'24)33.26619.02358.3
TacoDepth (Ours)14.25139.8729.3
+ +Table 3. Model efficiency. We report FLOPs, model parameters, and runtime of different methods. Since varied depth predictors can be applied, the results of plug-in models do not contain initial depth prediction. FLOPs and runtime are evaluated with one $900 \times 1600$ image and 30 Radar points on one NVIDIA A6000 GPU. + +
InitialRadarCam-Depth [21]TacoDepth (Ours)
δ1↑Rel↓MAE↓δ1↑Rel↓MAE↓δ1↑Rel↓MAE↓
DPT-Hybrid [35]0.9030.092-0.9210.0871157.00.9320.076983.1
MiDaS-v3.1 [3]0.9170.081-0.9260.0861177.20.9440.071852.4
Depth-Anything-v2 [51]0.9320.069-0.9400.073924.70.9610.059730.2
+ +sparse and noisy, especially on challenging nighttime and glaring scenes, leading to disrupted structures, blurred details, and obvious artifacts in their final depth predictions, as presented in Fig. 4. In contrast, we propose the one-stage TacoDepth without the quasi-dense depth. By effectively integrating the complementary image details and Radar structures, our model produces robust depth for both daytime and nighttime scenarios in Fig. 4. See supplement for more quantitative and visual results on nighttime scenes. + +# 4.4. Model Efficiency + +We evaluate the processing time of different methods in Table 1. Compared with the second-best independent model of Li et al. [20] and the plug-in model RadarCam-Depth [21], TacoDepth improves inference speed by $60.5\%$ and $91.8\%$ . + +To further showcase our efficiency, we report model parameters and FLOPs in Table 3. Compared with the two-stage Singh et al. [37], the one-stage TacoDepth reduces the parameters and FLOPs by $40.9\%$ and $72.2\%$ . Compared with the four-stage plug-in model RadarCam-Depth [21], we decrease parameters and FLOPs by $57.1\%$ and $77.4\%$ . + +Note that, the differences between the plug-in and independent modes of our method arise from the auxiliary branch to fuse initial relative depth. TacoDepth achieves real-time processing of 37.5 fps by independent inference. + +# 4.5. Ablation Studies + +Plug-in Inference. In Table 4, our TacoDepth outperforms the RadarCam-Depth [21] with three different relative depth predictors [3, 35, 51]. Using more advanced depth predictors, e.g., MiDaS-v3.1 [3] and Depth-Anything-v2 [51], can further boost our accuracy without extra effort. The results demonstrate the versatility and flexibility of our approach. Efficacy of TacoDepth. In Table 5, we ablate the graph + +Table 4. Plug-in models with varied depth predictors. Our method outperforms RadarCam-Depth [21] with different depth predictors [3, 35, 51] on ZJU-4DRadarCam [21] (0-70 meters). + +
MethodRGBRadarGEPFMAE↓RMSE↓
RGB Baseline✓✗✗✗2474.35402.1
Singh et al. [37] (CVPR'23)✓✓✗✗2073.24590.7
TacoDepth (w/o GE)✓✓✗✓1815.64189.8
TacoDepth (Ours)✓✓✓✓1712.63960.5
+ +Table 5. Efficacy of the TacoDepth. We ablate the effectiveness of graph-based Radar structure extractor (GE) and pyramid-based Radar fusion (PF) for Radar structure extraction and integration. We adopt the same baseline as Singh et al. [37]. TacoDepth $(w / o$ GE) uses the same MLP to extract Radar point features as prior art [37]. The results are evaluated on nuScenes [4] in 0-70 meters. + +
MetricL=1L=2L=3L=4
MAE↓1840.61787.81712.61729.3
RMSE↓4203.74046.93960.53984.2
+ +Table 6. Layer number L. Results are on nuScenes [4] in $0 - 70\mathrm{m}$ + +based Radar structure extractor (GE) and pyramid-based Radar fusion module (PF). With the same MLP to extract Radar point features [37], TacoDepth $(w / o$ GE) reduces MAE and RMSE by $12.4\%$ and $8.7\%$ than Singh et al. [37], which are attributed to the PF. Replacing the MLP [37] with GE further decreases MAE and RMSE by $5.7\%$ and $5.5\%$ . + +We adopt the RGB baseline of Singh et al. [37], a common single-image depth U-Net. Overall, TacoDepth improves accuracy beyond the baseline by $30.8\%$ and $26.7\%$ (MAE and RMSE), proving the efficacy of our approach. + +Layer Number $L$ . In Table 6, we ablate the layer number $L$ of GE and PF. Compared with shallow layers $(L = 1,2)$ , deeper layers $(L = 3)$ better capture structures of point clouds [41, 45], producing higher accuracy. Due to Radar sparsity, more layers $(L = 4)$ does not yield improvements but decreases efficiency. We adopt $L = 3$ in experiments. + +# 5. Conclusion + +In this paper, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation model with one-stage fusion. Different from the multi-stage methods with quasidense depth, TacoDepth utilizes the graph-based Radar structure extractor and pyramid-based Radar fusion module to capture and integrate Radar features, achieving superior efficiency and robustness. Moreover, TacoDepth is flexible for independent and plug-in processing, offering a better balance of speed and accuracy. Our work provides a new perspective on the efficient Radar-Camera depth estimation. Limitations. We currently provide one efficient implementation of our TacoDepth. In future work, more techniques can be explored for varied performances and applications. + +Acknowledgments. The study is supported under RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 1 grant (RG14/22). + +# References + +[1] C Bradford Barber, David P Dobkin, and Hannu Huhdanpaa. The quickhull algorithm for convex hulls. ACM Transactions on Mathematical Software (TOMS), 22(4):469-483, 1996. 5 +[2] Oded Bialer and Yuval Haitman. Radsimreal: Bridging the gap between synthetic and real data in radar object detection with simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15407-15416, 2024. 1, 2 +[3] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 1, 2, 4, 5, 6, 7, 8 +[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11618-11628, 2020. 1, 2, 5, 6, 7, 8 +[5] Yujeong Chae, Hyeonseong Kim, and Kuk-Jin Yoon. Towards robust 3d object detection with lidar and 4d radar fusion in various weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15162-15172, 2024. 1, 2 +[6] Feipeng Chen, Lihui Wang, Pinzhang Zhao, and Jian Wei. 3d grid-based vdbscan clustering and radar-monocular depth completion. IEEE Sensors Journal, 2024. 1 +[7] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 4 +[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. 35, 2022. 2, 4 +[9] Fangqiang Ding, Andras Palffy, Dariu M. Gavrila, and Chris Xiaoxuan Lu. Hidden gems: 4d radar scene flow learning using cross-modal supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9340-9349, 2023. 1, 2 +[10] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6824-6835, 2021. 2, 4 +[11] Kexue Fu, Shaolei Liu, Xiaoyuan Luo, and Manning Wang. Robust point cloud registration framework based on deep graph matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8893-8902, 2021. 2, 3, 4, 6 +[12] Stefano Gasperini, Patrick Koch, Vinzensz Dallabetta, Nassir Navab, Benjamin Busam, and Federico Tombari. R4dyn: Exploring radar for self-supervised monocular depth estimation of dynamic scenes. In International Conference on 3D Vision (3DV), pages 751-760, 2021. 1, 5, 6 +[13] Kyle Harlow, Hyesu Jang, Timothy D. Barfoot, Ayoung Kim, and Christoffer Heckman. A new wave in robotics: Survey on recent mmwave radar applications in robotics. IEEE Transactions on Robotics, 40:4544-4560, 2024. 1, 2 + +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 2, 4, 5 +[15] Daniel Casado Herreraz, Matthias Zeller, Le Chang, Ignacio Vizzo, Michael Heidingsfeld, and Cyril Stachniss. Radar-only odometry and mapping for autonomous vehicles. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 539-548. IEEE, 2019. 1, 2 +[16] Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Implicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24118-24129, 2024. 1, 2 +[17] Andrey Ignatov, Grigory Malivenko, David Plowman, Samarth Shukla, and Radu Timofte. Fast and accurate single-image depth estimation on mobile devices, mobile ai 2021 challenge: Report. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2545-2557, 2021. 2 +[18] Yinuo Jiang, Beitong Zhou, Xiaoyu Liu, Qingyi Li, and Cheng Cheng. Gtinet: Global topology-aware interactions for unsupervised point cloud registration. IEEE Transactions on Circuits and Systems for Video Technology, 34(7): 6363-6375, 2024. 2 +[19] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, 2012. 2, 4 +[20] Huadong Li, Minhao Jing, Jiajun Liang, Haoqiang Fan, and Renhe Ji. Sparse beats dense: Rethinking supervision in radar-camera depth completion. arXiv preprint arXiv:2312.00844, 2023. 1, 5, 6, 8 +[21] Han Li, Yukai Ma, Yaqing Gu, Kewei Hu, Yong Liu, and Xingxing Zuo. Radarcam-depth: Radar-camera fusion for depth estimation with learned metric scale. arXiv preprint arXiv: 2401.04325, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[22] Jiaqi Li, Yiran Wang, Zihao Huang, Jinghong Zheng, Ke Xian, Zhiguo Cao, and Jianming Zhang. Diffusion-augmented depth prediction with sparse annotations. In Proceedings of the 31st ACM International Conference on Multimedia, page 2865-2876, New York, NY, USA, 2023. Association for Computing Machinery. 1 +[23] Jiaqi Li, Yiran Wang, Jinghong Zheng, Zihao Huang, Ke Xian, Zhiguo Cao, and Jianming Zhang. Self-distilled depth refinement with noisy poisson fusion, 2024. 2 +[24] Xingyi Li, Chaoyi Hong, Yiran Wang, Zhiguo Cao, Ke Xian, and Guosheng Lin. Symmernf: Learning to explore symmetry prior for single-view view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1726-1742, 2022. 2 +[25] Zecheng Li, Yuying Song, Fuyuan Ai, Chunyi Song, and Zhiwei Xu. Semantic-guided depth completion from monocular images and 4d radar data. IEEE Transactions on Intelligent Vehicles, pages 1-13, 2024. 1 + +[26] Juan-Ting Lin, Dengxin Dai, and Luc Van Gool. Depth estimation from monocular images and sparse radar data. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10233-10240. IEEE, 2020. 1, 2, 3, 4, 5, 6, 7 +[27] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. pages 14928-14937, 2024. 2 +[28] Chen-Chou Lo and Patrick Vandewalle. Depth estimation from monocular images and sparse radar using deep ordinal regression network. In IEEE International Conference on Image Processing (ICIP), pages 3343-3347, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[29] Chen-Chou Lo and Patrick Vandewalle. Rcdpt: Radar-camera fusion dense prediction transformer. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 3, 4, 5, 6 +[30] Chen-Chou Lo and Patrick Vandewalle. Expanding sparse radar depth based on joint bilateral filter for radar-guided monocular depth estimation. Sensors, 24(6):1864, 2024. +[31] Yunfei Long, Daniel Morris, Xiaoming Liu, Marcos Castro, Punarjay Chakravarty, and Praveen Narayanan. Radar-camera pixel depth association for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12507-12516, 2021. 1, 2, 3, 4, 5, 6, 7 +[32] Guohao Peng, Heshan Li, Yangyang Zhao, Jun Zhang, Zhenyu Wu, Pengyu Zheng, and Danwei Wang. Transloc4d: Transformer-based 4d radar place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17595-17605, 2024. 1, 2 +[33] Pierluigi Zama Ramirez, Fabio Tosi, Luigi Di Stefano, Radu Timofte, Alex Costanzino, Matteo Poggi, Samuele Salti, Stefano Mattoccia, Yangyang Zhang, Cailin Wu, Zhuangda He, Shuangshuang Yin, Jiaxu Dong, Yangchenxu Liu, Hao Jiang, Jun Shi, Yong A, Yixiang Jin, Dingzhe Li, Bingxin Ke, Anton Obukhov, Tinafu Wang, Nando Metzger, Shengyu Huang, Konrad Schindler, Yachuan Huang, Jiaqi Li, Junrui Zhang, Yiran Wang, Zihao Huang, Tianqi Liu, Zhiguo Cao, Pengzhi Li, Jui-Lin Wang, Wenjie Zhu, Hui Geng, Yuxin Zhang, Long Lan, Kele Xu, Tao Sun, Qisheng Xu, Sourav Saini, Aashray Gupta, Sahaj K. Mistry, Aryan Shukla, Vinit Jakhetiya, Sunil Jaiswal, Yuejin Sun, Zhuofan Zheng, Yi Ning, Jen-Hao Cheng, Hou-I Liu, Hsiang-Wei Huang, Cheng-Yen Yang, Zhongyu Jiang, Yi-Hao Peng, Aishi Huang, and Jenq-Neng Hwang. Ntire 2024 challenge on hr depth from images of specular and transparent surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6499–6512, 2024. 2 +[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(03):1623-1637, 2020. 1, 2, 3, 4, 5, 6 +[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi + +sion transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12179-12188, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[36] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks and Learning Systems, 20(1):61-80, 2009. 2, 3, 4, 6 +[37] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9275-9285, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[38] Tieshuai Song, Bin Yang, Jun Wang, Guidong He, Zhao Dong, and Fengjun Zhong. mmwave radar and image fusion for depth completion: a two-stage fusion network. In The 27th International Conference on Information Fusion (FUSION), 2024. 1, 2, 3, 4, 5, 6 +[39] Christian Stetco, Barnaba Ubezio, Stephan Muhlbacher-Karrer, and Hubert Zangl. Radar sensors in collaborative robotics: Fast simulation and experimental validation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10452-10458. IEEE, 2020. 1, 2 +[40] Huawei Sun, Hao Feng, Julius Ott, Lorenzo Servadei, and Robert Wille. Cafnet: A confidence-driven framework for radar camera depth estimation. arXiv preprint arXiv:2407.00697, 2024. 1, 2, 3, 4, 5, 6, 7 +[41] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6411–6420, 2019. 2, 3, 4, 8 +[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, 2017. 4 +[43] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3056-3065, 2019. 2, 3, 4 +[44] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3056-3065, 2019. 3 +[45] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5), 2019. 2, 3, 4, 8 +[46] Yiran Wang, Xingyi Li, Min Shi, Ke Xian, and Zhiguo Cao. Knowledge distillation for fast and accurate monocular depth estimation on mobile devices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2457-2465, 2021. 2 +[47] Yiran Wang, Zhiyu Pan, Xingyi Li, Zhiguo Cao, Ke Xian, and Jianming Zhang. Less is more: Consistent video depth estimation with masked frames modeling. In Proceedings + +of the 30th ACM International Conference on Multimedia, page 6347-6358, New York, NY, USA, 2022. Association for Computing Machinery. 2 +[48] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9466-9476, 2023. 2 +[49] Yiran Wang, Min Shi, Jiaqi Li, Chaoyi Hong, Zihao Huang, Juewen Peng, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Nvds+: Towards efficient and versatile neural stabilizer for video depth estimation. IEEE transactions on pattern analysis and machine intelligence, pages 1-18, 2024. 1 +[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, + +Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. 34:12077-12090, 2021. 2, 4 +[51] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 4, 5, 6, 7, 8 +[52] Junrui Zhang, Jiaqi Li, Yachuan Huang, Yiran Wang, Jinghong Zheng, Liao Shen, and Zhiguo Cao. Towards robust monocular depth estimation in non-lambertian surfaces, 2024. 5 +[53] Yi Zhou, Lulu Liu, Haocheng Zhao, Miguel López-Benítez, Limin Yu, and Yutao Yue. Towards deep radar perception for autonomous driving: Datasets, methods, and challenges. Sensors, 22(11):4208, 2022. 1, 2 \ No newline at end of file diff --git a/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/images.zip b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e3f568698c58fc86b1a13e09bf88895557d91542 --- /dev/null +++ b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c28556a90d4fb846e1adab66a4b5c901daebaf16283227cec72558bc360fcaf +size 642026 diff --git a/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/layout.json b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..63f5516cf80918f39d8f3f33fad25c47ac0c43a0 --- /dev/null +++ b/2025/TacoDepth_ Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion/layout.json @@ -0,0 +1,9308 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "type": "text", + "content": "TacoDepth: Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "spans": [ + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": "Yiran Wang" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": " Jiaqi Li" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": " Chaoyi Hong" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": " Ruibo Li" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": " \nLiusheng Sun" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": " Xiao Song" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": " Zhe Wang" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": " Zhiguo Cao" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": " Guosheng Lin" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 160, + 492, + 203 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 203, + 526, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 203, + 526, + 217 + ], + "spans": [ + { + "bbox": [ + 85, + 203, + 526, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 203, + 526, + 217 + ], + "type": "text", + "content": "School of AIA, Huazhong University of Science and Technology " + }, + { + "bbox": [ + 85, + 203, + 526, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 203, + 526, + 217 + ], + "type": "text", + "content": "SenseTime Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 178, + 219, + 432, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 219, + 432, + 232 + ], + "spans": [ + { + "bbox": [ + 178, + 219, + 432, + 232 + ], + "type": "text", + "content": "{wangyiran,zgcao}@hust.edu.cn, gslin@ntu.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 186, + 233, + 421, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 233, + 421, + 245 + ], + "spans": [ + { + "bbox": [ + 186, + 233, + 421, + 245 + ], + "type": "text", + "content": "https://github.com/RaymondWang987/TacoDepth" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 266, + 200, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 266, + 200, + 280 + ], + "spans": [ + { + "bbox": [ + 152, + 266, + 200, + 280 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 292, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 557 + ], + "type": "text", + "content": "Radar-Camera depth estimation aims to predict dense and accurate metric depth by fusing input images and Radar data. Model efficiency is crucial for this task in pursuit of real-time processing on autonomous vehicles and robotic platforms. However, due to the sparsity of Radar returns, the prevailing methods adopt multi-stage frameworks with intermediate quasi-dense depth, which are time-consuming and not robust. To address these challenges, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation model with one-stage fusion. Specifically, the graph-based Radar structure extractor and the pyramid-based Radar fusion module are designed to capture and integrate the graph structures of Radar point clouds, delivering superior model efficiency and robustness without relying on the intermediate depth results. Moreover, TacoDepth can be flexible for different inference modes, providing a better balance of speed and accuracy. Extensive experiments are conducted to demonstrate the efficacy of our method. Compared with the previous state-of-the-art approach, TacoDepth improves depth accuracy and processing speed by " + }, + { + "bbox": [ + 55, + 292, + 296, + 557 + ], + "type": "inline_equation", + "content": "12.8\\%" + }, + { + "bbox": [ + 55, + 292, + 296, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 292, + 296, + 557 + ], + "type": "inline_equation", + "content": "91.8\\%" + }, + { + "bbox": [ + 55, + 292, + 296, + 557 + ], + "type": "text", + "content": ". Our work provides a new perspective on efficient Radar-Camera depth estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 598, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 296, + 696 + ], + "type": "text", + "content": "Radar-Camera depth estimation [6, 12, 20, 21, 25, 26, 28-31, 37] is a vital task for 3D perception and motion planning in autonomous driving [4, 15, 22, 49, 53] and robotics [13, 21, 39]. The task aims to predict dense metric depth by integrating images with mmWave Radar data, perceiving metric scale in different environments. As a typical range sensor on driving and robotic platforms, Radar is widely equipped and used for its high affordability and low power consump" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 266, + 561, + 415 + ], + "blocks": [ + { + "bbox": [ + 317, + 266, + 561, + 415 + ], + "lines": [ + { + "bbox": [ + 317, + 266, + 561, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 561, + 415 + ], + "type": "image", + "image_path": "5ea893d73a9f562e8ff7a9996d340f22ad41f96b1382fcc6f7a46389e4b8e5f5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 418, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 475 + ], + "type": "text", + "content": "Figure 1. Performance and efficiency. Circle area indicates inference time (ms). Smaller circles showcase faster speed. The X-axis and Y-axis represent MAE and RMSE metrics for depth errors on the nuScenes [4] dataset. Lower MAE and RMSE mean higher accuracy. Our TacoDepth outperforms prior arts by large margins." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 495, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 542 + ], + "type": "text", + "content": "tion. It is valuable to explore Radar-Camera fusion in various tasks [2, 5, 9, 16, 32, 37] for its superior all-weather reliability. Besides, for real-time processing on vehicles and robotics, improving efficiency is crucial for fusion models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "The obstacle for fusing Radar in depth estimation lies in its sparsity and noises. Due to the large beam width, Radar point clouds are " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "1000\\mathrm{x}" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " sparser with higher noise levels than LiDAR returns. To deal with the issue, the prevailing methods [21, 26, 28, 29, 31, 37, 38, 40] adopt complex multi-stage frameworks. Intermediate quasi-dense depth is predicted to bridge the gap between sparse Radar and final dense depth. For instance, two-stage methods [26, 28, 29, 31, 37, 38, 40] predict final depth based on intermediate results, as independent models without utilizing other depth predictors [3, 34, 35, 51]. RadarCam-Depth [21], as a plugin module, converts relative depth [34, 35] to metric scale in four stages by post-processing. These multi-stage methods are inefficient and time-consuming, as presented in Fig. 1." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 217, + 713 + ], + "type": "text", + "content": "* Equal contribution, † Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10523" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "Meanwhile, despite being described as quasi-dense, the intermediate depth [21, 26, 28, 29, 31, 37, 40] remains sparse and noisy. As shown in Fig. 2, only few pixels exhibit valid depth. For some adverse lighting or weather conditions, e.g., glaring and nighttime scenarios, even no pixels are predicted with depth values. These defective intermediate results lead to disrupted structures, blurred details, or noticeable artifacts in final predictions, as shown in Fig. 4. With these observations, a question naturally arises: can models directly predict accurate and dense depth without the intermediate results, achieving one-stage Radar-Camera depth estimation with higher efficiency and robustness?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 218, + 296, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 457 + ], + "type": "text", + "content": "To answer the above question, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation framework with one-stage fusion. To simplify multi-stage fusion [21, 26, 28, 29, 31, 37, 38, 40] into one stage, TacoDepth needs to extract and integrate Radar information more effectively. Prior arts simply extract features from Radar coordinates [21, 37] or projected depth [26, 31], sensitive to noises and outliers [11]. In contrast, we utilize a graph-based Radar structure extractor to capture geometric structures and topologies of point clouds, improving robustness to inaccurate points [11, 18, 36, 43]. After that, our pyramid-based Radar fusion module integrates image and Radar features hierarchically. Shallow layers capture image details and Radar coordinates, while deep layers fuse scene semantics and structures [10, 14, 19, 41, 45, 50]. In each layer, Radar-centered flash attention mechanism is applied to establish the cross-modal correlations. For one Radar point, lightweight and accelerated flash attention [8] is calculated within Radar-centered areas for efficiency. Finally, TacoDepth predicts accurate and dense depth by a decoder." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 460, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 592 + ], + "type": "text", + "content": "Moreover, we further analyze the characteristics of different inference strategies in prior arts [21, 26, 28, 29, 31, 37, 38, 40]. Independent models [26, 28, 29, 31, 37, 40] are more efficient without initial depth prediction and postprocessing, while the plug-in model [21] can be more accurate by leveraging cutting-edge relative depth predictors [3, 23, 34, 51]. Therefore, TacoDepth aims to be flexible for the two settings. An auxiliary and optional branch is added for initial depth, allowing TacoDepth to support both independent and plug-in processing. Our approach can better balance depth accuracy and model efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "We conduct evaluations on nuScenes [4] and ZJU-4DRadarCam [21] datasets with different depth ranges, following prior arts [21, 31, 37]. As shown in Fig. 1, whether compared with independent [26, 28, 31, 37, 40] or plugin in [21] models, TacoDepth significantly outperforms prior arts [21, 26, 28, 31, 37, 40] in accuracy and efficiency. Besides, our approach achieves real-time processing of over 37 fps as an independent model and strong compatibility with various depth predictors [3, 34, 35, 51] as a plug-in model, further proving the efficiency and flexibility of TacoDepth." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 553, + 134 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 134 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 134 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 134 + ], + "type": "image", + "image_path": "1e7180a4707dc7ac431aa404e436ae2c9b86da99bc267c0cfa0a28614cba35cc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 137, + 555, + 171 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 171 + ], + "type": "text", + "content": "Figure 2. Intermediate quasi-dense depth maps [21, 26, 28, 29, 31, 37, 40] remain sparse and noisy. Pixels with valid depth values are visualized by the gray areas in the red rectangular boxes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 186, + 535, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 535, + 197 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 535, + 197 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 201, + 553, + 273 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 201, + 553, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 201, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 314, + 201, + 553, + 224 + ], + "type": "text", + "content": "- We propose TacoDepth, an efficient and accurate Radar-Camera depth estimation method with one-stage fusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 225, + 553, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 225, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 314, + 225, + 553, + 248 + ], + "type": "text", + "content": "- We design the pyramid-based Radar fusion module, fusing Radar and image features effectively and efficiently." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 249, + 553, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 249, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 314, + 249, + 553, + 273 + ], + "type": "text", + "content": "- TacoDepth is flexible for both independent and plug-in processing, balancing model efficiency and performance." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 285, + 400, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 400, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 400, + 298 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 306, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 521 + ], + "type": "text", + "content": "MmWave Radar for Vehicles and Robotics. Autonomous driving [4, 15, 48, 53] and robotics [13, 21, 24, 39] rely on multiple sensors to perceive the environment, e.g., camera, LiDAR, and mmWave Radar. LiDAR and Radar provide the metric scale of different scenarios as a complement to cameras. Compared with LiDAR, Radar is cheaper and more widely-equipped with lower power consumption. The reflection properties of mmWave also provide Radar with farther sensing ranges and stronger robustness to all-weather conditions. The main drawback of Radar lies in its sparsity and noise. Due to the large beam width, Radar point clouds are " + }, + { + "bbox": [ + 313, + 306, + 555, + 521 + ], + "type": "inline_equation", + "content": "1000\\mathrm{x}" + }, + { + "bbox": [ + 313, + 306, + 555, + 521 + ], + "type": "text", + "content": " sparser [37] than LiDAR. Besides, Radar captures coordinates and depth values with lower accuracy. For instance, 3D Radar suffers from height ambiguity [31] due to insufficient antenna elements along the elevation axis, while this problem can be alleviated by the emerging 4D Radar. Thus, integrating Radar data is valuable for various tasks [2, 5, 9, 16, 27, 32, 37] in driving and robotic systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Radar-Camera Depth Estimation. The task aims to predict dense metric depth by merging images and Radar, facilitating 3D perception and motion planning of vehicles and robotics. Model efficiency is crucial for these real-time applications. However, to overcome the sparsity and noise of Radar, prior arts [21, 26, 28, 29, 31, 37, 38, 40] adopt inefficient multi-stage frameworks. The final depth is predicted based on intermediate quasi-dense depth. These methods can be divided into independent and plug-in models. Most independent models [26, 28, 29, 31, 37, 40] predict final depth in two stages without relying on other depth predictors [17, 33-35, 46, 47]. For example, Lin et al. [26] and DORN [28] use two CNNs [14] for quasi-dense and final depth. RC-PDA [31] and RC-DPT [29] convert intermediate results into multiple channels for varied confidence levels. Singh et al. [37] propose gated fusion to filter the inac" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10524" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 380, + 281 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 380, + 281 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 380, + 281 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 380, + 281 + ], + "type": "image", + "image_path": "7af03cc5a3f094eef113e03d9d84f96b8eeaef12d4f649907d8ef409a66c004b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": "Figure 3. Overview of the TacoDepth. The graph-based Radar structure extractor captures graph structures of Radar point clouds through the node feature " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": " and edge feature " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "E_{l}" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": " in a certain layer " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": ". The pyramid-based Radar fusion module integrates image and Radar features in a pyramidal hierarchical manner. In each layer, to efficiently build cross-modal correspondences, the Radar-centered flash attention is calculated within Radar-centered areas based on horizontal coordinates, e.g., pixels in the orange area as queries, while Radar points in the red area as keys and values. Our TacoDepth achieves efficient and accurate Radar-Camera depth estimation in one stage. During inference, the model is flexible and supports both independent and plug-in processing, facilitating a better balance of model efficiency and accuracy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 541, + 281 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 541, + 281 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 541, + 281 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 541, + 281 + ], + "type": "image", + "image_path": "8ece566f18f36fa5fc0ecaf50fc399d52b01522b368a268e9d4d6ec4fe6fae46.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 370, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 370, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 370, + 295, + 467 + ], + "type": "text", + "content": "curate quasi-dense depth. Besides, RadarCam-Depth [21] designs a plug-in module, aligning relative depth [34, 35] to metric scale in four stages, including initial depth, global scale alignment, quasi-dense depth, and scale learner. These methods [21, 26, 28, 29, 31, 37, 38, 40] are inefficient with complex stages. Defective intermediate depth also limits their robustness. In contrast, we propose TacoDepth model to achieve efficient, accurate, and flexible one-stage fusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 476, + 127, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 127, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 127, + 491 + ], + "type": "text", + "content": "3. TacoDepth" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 497, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 295, + 593 + ], + "type": "text", + "content": "We present a detailed illustration of TacoDepth, an efficient Radar-Camera depth estimation framework with one-stage fusion. In Sec. 3.1, we provide an overall description to outline our approach. In Sec. 3.2, we elaborate on the detailed architecture of the pyramid-based Radar fusion module. TacoDepth can be flexible for independent and plug-in processing, which will be discussed in Sec. 3.3. Besides, the training procedure for TacoDepth is illustrated in Sec. 3.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 600, + 123, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 123, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 123, + 611 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "Here we present an overview of our TacoDepth framework. Given a single image " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "I \\in \\mathcal{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": " and a Radar point cloud " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "P \\in \\mathcal{R}^{K \\times 3}" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": " as inputs, our goal is to predict dense metric depth " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "D \\in \\mathcal{R}^{H \\times W}" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": " represent image height and weight. The point cloud " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "p \\in \\mathcal{R}^3" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Radar-Camera depth estimation can be divided into two primary subtasks: the extraction and the integration of Radar information. Accordingly, as depicted in Fig. 3," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 370, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 407 + ], + "type": "text", + "content": "TacoDepth comprises a graph-based Radar structure extractor and a pyramid-based Radar fusion module to undertake these two important steps respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 408, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 493 + ], + "type": "text", + "content": "Firstly, for Radar feature extraction, previous methods [21, 37] simply extract features from the coordinates of individual Radar points, focusing solely on the point-level information. For example, Singh et al. [37] extract point features in 32,256 dimensions from the three-dimensional coordinates using a multi-layer perception (MLP), which inevitably introduces substantial redundancy and noise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 699 + ], + "type": "text", + "content": "To achieve efficient Radar-Camera fusion in a single stage, TacoDepth needs to extract more comprehensive and effective Radar information, prompting us to reconsider the complementarity between camera images and Radar data. Compared to the coordinates, the geometric structures of point clouds, e.g., the distances or relations between Radar points, are potentially more informative for Radar-Camera fusion. The overall topologies are also more robust to inaccurate outliers [11]. To this end, we employ the graph-based Radar structure extractor with a lightweight GNN [11, 36, 43] architecture. Specifically, Radar points are regarded as the graph nodes. The adjacency matrix of all points represents the graph edges. In each layer, the node features of Radar points can be updated and aggregated along edges by PCA-GM [44]. The graph-based Radar structure extractor can capture point coordinates and graph structures from shallow layers to deep layers [11, 41, 45]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "After extracting the Radar structures, the subsequent step" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10525" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "is to merge RGB and Radar features. Prior arts [21, 26, 28, 29, 31, 37, 38, 40] adopt multi-stage methods with intermediate quasi-dense depth, which are inefficient and not robust as illustrated in Sec. 1. In contrast, we propose our pyramid-based Radar fusion module for efficient one-stage fusion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 337 + ], + "type": "text", + "content": "To be specific, features of the two modalities are integrated in a pyramidal hierarchical manner. Shallow layers can merge image details with Radar coordinates, whereas deeper layers could fuse scene semantics and geometric structures [10, 14, 19, 41, 45, 50]. In each layer, we leverage Radar-centered flash attention for cross-modal correlations and feature fusion. The queries are derived from image pixels, while keys and values originate from Radar points. For each point, we delineate Radar-centered areas based on the horizontal coordinates, as exemplified by the orange and red areas in Fig. 3, given that the projected horizontal coordinates of Radar points are generally more precise than the height dimension [31, 37]. Lightweight flash attention [8] is calculated within these areas to maintain efficiency, since external pixels and points can not be correlated. Finally, with the fused features from all layers, TacoDepth predicts accurate and dense metric depth by a depth decoder [37]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 338, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 445 + ], + "type": "text", + "content": "Furthermore, TacoDepth is adaptable for both independent and plug-in inference. An optional input branch is introduced to handle the initial relative depth. For the independent inference, our model achieves real-time processing of over 37 fps. For the plug-in manner, TacoDepth demonstrates compatibility with various cutting-edge depth predictors [3, 34, 35, 51], as proved in Sec. 4.5. Consequently, our approach can deliver a balanced trade-off between performance and efficiency for different users and applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 217, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 217, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 217, + 467 + ], + "type": "text", + "content": "3.2. Pyramid-based Radar Fusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": "For the Radar point cloud " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": ", our graph-based Radar structure extractor derives node feature " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": " and edge feature " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "E_{l}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": " at a certain layer " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "l \\in \\{1,2,\\dots ,L\\}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": " represents the layer number of the GNN module [11, 36, 43]. For the input image " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": ", the image encoder [14, 21, 37] extracts multi-scale features " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\{F_1,F_2,\\dots ,F_{2L}\\}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": ". The pyramid-based Radar fusion module hierarchically integrates the multi-modal features. " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "E_{l}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": " will be fused with image features in adjacent layers. We merge the node feature " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": " with image feature " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "F_{2l - 1}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": ", and the edge feature " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "E_{l}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "inline_equation", + "content": "F_{2l}" + }, + { + "bbox": [ + 55, + 473, + 295, + 593 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "The prerequisite for feature fusion is cross-modal correspondences between Radar points and RGB pixels. Due to the noisy elevation coordinates of 3D Radar, prior arts [21, 29, 31, 37, 40] try to build one-to-many mappings by the intermediate quasi-dense depth. They adopt explicit classification supervision with the depth discrepancies between Li-DAR and Radar as pseudo-labels. However, the quasi-dense depth prediction leads to inefficiency and erroneous final outcomes, as discussed in Sec. 1. Given the absence of reliable explicit labels and supervision, we hope our TacoDepth" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 534, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 534, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 534, + 84 + ], + "type": "text", + "content": "can learn and establish the correspondences implicitly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": "Therefore, for efficient Radar-Camera correlations, we propose the Radar-centered flash attention in TacoDepth. We define Radar-centered areas based on the precise horizontal coordinates of Radar points and image pixels. For each point, the attention is computed only with image pixels in the area, since external pixels can not be correlated. Here we take " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "E_{l} \\in \\mathcal{R}^{K \\times K}" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "F_{2l} \\in \\mathcal{R}^{h \\times w \\times C_l}" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": " as examples, where " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": " are feature height and width. " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": " refer to the point number and layer feature channel. For a Radar point " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": " with the horizontal coordinate " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": ", as showcased by the exemplary orange area in Fig. 3, we only retain image pixels " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": " whose coordinates " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": " are close to " + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 313, + 85, + 553, + 228 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 232, + 553, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 232, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 319, + 232, + 553, + 248 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {2 l} = F _ {2 l} \\left[ x _ {p} - a _ {l} < x _ {m} < x _ {p} + a _ {l} \\right], \\hat {F} _ {2 l} \\in \\mathcal {R} ^ {\\hat {M} \\times C _ {l}}, \\tag {1}", + "image_path": "ff13b02320b1bd929f29666359e0914d1016bde8757f777dc66ea95b4a180917.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "inline_equation", + "content": "a_{l}" + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "text", + "content": " indicates the width of the Radar-centered area in layer " + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "text", + "content": ". The notation " + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "text", + "content": " depicts the fetching operation of pixels or points. " + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\hat{F}_{2l}" + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "text", + "content": " is the image feature of the remaining " + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "text", + "content": " pixels to generate queries. For a random pixel " + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "inline_equation", + "content": "\\hat{F}_{2l}" + }, + { + "bbox": [ + 313, + 252, + 553, + 324 + ], + "type": "text", + "content": ", as represented by the red area in Fig. 3, the attention calculation is also restricted to adjacent Radar points:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 328, + 553, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 328, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 322, + 328, + 553, + 344 + ], + "type": "interline_equation", + "content": "\\hat {E} _ {l} = E _ {l} \\left[ x _ {m} - a _ {l} < x _ {p} < x _ {m} + a _ {l} \\right], \\hat {E} _ {l} \\in \\mathcal {R} ^ {\\hat {K} \\times K}. \\tag {2}", + "image_path": "03313e8d3e04bc98044e3b8d4aaab6d0c0e4d94023c60722f29b7c8a636c4571.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "content": "The feature dimension maintains the original " + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "content": " to preserve integral graph topologies. " + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "inline_equation", + "content": "\\hat{E}_l" + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "content": " is the edge feature of the retained " + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "inline_equation", + "content": "\\hat{K}" + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "content": " points for keys and values. In this way, the computational costs can be significantly decreased with the small subset of " + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "content": " pixels and " + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "inline_equation", + "content": "\\hat{K}" + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "content": " points. For the pixel " + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "content": ", we generate queries from " + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "inline_equation", + "content": "\\hat{F}_{2l}[m] \\in \\mathcal{R}^{C_l}" + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "content": ", while keys and values are derived from " + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "inline_equation", + "content": "\\hat{E}_l" + }, + { + "bbox": [ + 313, + 348, + 553, + 431 + ], + "type": "text", + "content": ". Radar-centered attention is computed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 330, + 435, + 553, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 435, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 330, + 435, + 553, + 472 + ], + "type": "interline_equation", + "content": "F _ {2 l} ^ {\\prime} [ m ] = \\operatorname {s o f t m a x} \\frac {W _ {q} \\hat {F} _ {2 l} [ m ] \\left(W _ {k} \\hat {E} _ {l}\\right) ^ {T}}{\\sqrt {C _ {l}}} W _ {v} \\hat {E} _ {l}, \\tag {3}", + "image_path": "c346eb509a4cd787b95384d61477e065ebc04a3bab687af20fa58a92a6213880.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "text", + "content": " are learnable linear projections. The fused Radar-Camera feature " + }, + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "inline_equation", + "content": "F_{2l}^{\\prime} \\in \\mathcal{R}^{h \\times w \\times C_l}" + }, + { + "bbox": [ + 313, + 474, + 553, + 557 + ], + "type": "text", + "content": " can be acquired for the current layer. The above calculation is incorporated into attention blocks with MLPs and residual connections. To further improve model efficiency, we employ lightweight and accelerated flash attention modules [7, 8] to replace the original attention computation [42]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "text", + "content": "Similarly, the node feature " + }, + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "text", + "content": " can be merged with the image feature " + }, + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "inline_equation", + "content": "F_{2l-1}" + }, + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "text", + "content": " by slightly adjusting the dimensions of linear projections. Ultimately, the acquired Radar-Camera features " + }, + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "inline_equation", + "content": "\\{F_1', F_2', \\dots, F_{2L}'\\}" + }, + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "text", + "content": " are fed to a common depth decoder [21, 37] to predict accurate metric depth." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "As a consequence, the proposed pyramid-based Radar fusion module enables efficient and effective Radar-Camera fusion in a single stage. The Radar-centered flash attention establishes cross-modal correlations with minimal computational costs. Without relying on the previous multi-stage frameworks and intermediate quasi-dense depth [21, 26, 28, 29, 31, 37, 38, 40], TacoDepth achieves higher efficiency and stronger robustness for Radar-Camera depth estimation." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10526" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 163, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 163, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 163, + 83 + ], + "type": "text", + "content": "3.3. Flexible Inference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "text", + "content": "Considering the varied inference manners in prior arts [21, 26, 28, 29, 31, 37, 38, 40], the plug-in module [21] could achieve higher accuracy using state-of-the-art relative depth predictors [34, 35, 52], while independent models [26, 28, 29, 31, 37, 38, 40] are capable of attaining faster inference speed without initial depth prediction and post-processing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 162, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 162, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 296, + 233 + ], + "type": "text", + "content": "Therefore, TacoDepth is designed to be flexible, supporting both independent and plug-in predictions. For the plug-in inference, an initial relative depth map is converted to metric scale by fusing camera and Radar information. If we denote the TacoDepth model as " + }, + { + "bbox": [ + 55, + 162, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta}" + }, + { + "bbox": [ + 55, + 162, + 296, + 233 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 55, + 162, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 162, + 296, + 233 + ], + "type": "text", + "content": ", the depth prediction process can then be formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 244, + 294, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 244, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 244, + 294, + 258 + ], + "type": "interline_equation", + "content": "D = \\mathcal {T} _ {\\theta} (I, P \\mid D ^ {*}), \\tag {4}", + "image_path": "b38354d95b4966f52e2b6d6b9eba82dcce51a925e00aa2858d17757de669262f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "text", + "content": "where the vertical bar " + }, + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "inline_equation", + "content": "(|)" + }, + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "text", + "content": " separates the required and optional inputs. " + }, + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "text", + "content": " is the predicted depth of our model. " + }, + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "inline_equation", + "content": "D^{*}" + }, + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "text", + "content": " is the optional input of initial relative depth from depth predictors [3, 34, 35, 51]. We introduce an auxiliary branch to process and fuse " + }, + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "inline_equation", + "content": "D^{*}" + }, + { + "bbox": [ + 55, + 262, + 296, + 396 + ], + "type": "text", + "content": " with image features, while keeping all other settings identical to the independent mode. Our plugin in manner demonstrates compatibility with various depth predictors, e.g., MiDaS [3, 34], DPT [35], and Depth-Anything-v2 [51]. In this way, TacoDepth can achieve a better balance of model efficiency and accuracy, offering superior flexibility for different users and applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 190, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 190, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 190, + 415 + ], + "type": "text", + "content": "3.4. Training the TacoDepth" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "content": "In the training phase, the training data [4, 21] is randomly divided into two equal portions in each epoch. For the auxiliary branch, one half of the data is provided with initial relative depth, while the other half adopts zero as input. This training strategy facilitates the simultaneous optimization of the independent inference and plug-in prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": "As for the training loss, we employ the same depth ground truth and " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " loss as the prior art [37]. Specifically, LiDAR-captured depth " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "D_{gt}" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " and reprojection-accumulated depth " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "D_{acc}" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " are utilized for supervision [21, 31, 37]. The overall loss function " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\ell_{L_1}" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " can be expressed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 559, + 296, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 559, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 61, + 559, + 296, + 590 + ], + "type": "interline_equation", + "content": "\\ell_ {L _ {1}} = \\frac {1}{| \\Omega_ {g t} |} \\sum_ {\\Omega_ {g t}} | D - D _ {g t} | + \\frac {\\lambda}{| \\Omega_ {a c c} |} \\sum_ {\\Omega_ {a c c}} | D - D _ {a c c} |, (5)", + "image_path": "0cc0110d8192b22bfdb43e4923924dafff805a51ad493feaa87081633acfc980.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\Omega_{gt}" + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\Omega_{acc}" + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": " depict the masks of " + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "inline_equation", + "content": "D_{gt}" + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "inline_equation", + "content": "D_{acc}" + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": " with valid depth values. " + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 598, + 295, + 624 + ], + "type": "text", + "content": " is the coefficient of the second term." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 634, + 137, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 137, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 137, + 647 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "To prove the efficacy of our TacoDepth, we conduct experiments on the nuScenes [4] and ZJU-4DRadarCam [21] datasets following prior arts [20, 21, 26, 28, 29, 31, 37, 40]. In Sec. 4.1, we describe the experimental datasets and evaluation protocols. The implementation details of our model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "are illustrated in Sec. 4.2. In Sec. 4.3, we compare the TacoDepth with previous Radar-Camera depth estimation approaches to demonstrate our state-of-the-art performance. The inference speed and model efficiency of different methods are evaluated and discussed in Sec. 4.4. In Sec. 4.5, we perform ablation studies to expound on our specific designs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 153, + 497, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 497, + 165 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 497, + 165 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Protocols" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 171, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 338 + ], + "type": "text", + "content": "Due to the scarcity of public datasets for Radar-Camera depth estimation, prior works [20, 26, 28, 29, 31, 37, 38, 40] only conduct experiments on nuScenes dataset [4] with 3D Radar. Recently, RadarCam-Depth [21] releases their ZJU-4DRadarCam dataset [21] with 4D Radar to alleviate this problem. We adopt these two datasets [4, 21] for experiments, proving our effectiveness on both 3D and 4D Radar. NuScenes Dataset. As an outdoor driving dataset, the nuScenes [4] is collected by vehicles equipped with cameras, 3D Radar, LiDAR, and IMU. It contains 1,000 scenes and around 40,000 synchronized Radar-Camera keyframes captured in Boston and Singapore. We follow the same train-test split as prior arts [21, 26, 29, 31, 37, 40], with 700 scenes for training, 150 for validation, and 150 for testing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 339, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 422 + ], + "type": "text", + "content": "ZJU-4DRadarCam Dataset. The dataset [21] is collected by a ground robot fitted with more advanced camera, 4D Radar, and LiDAR sensors, offering denser Radar data and LiDAR depth than the nuScenes [4] dataset. It comprises 33,409 Radar-Camera keyframes in total. We follow the official data split [21] with 29,312 frames for training and validation, and 4,097 samples for testing in our experiments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 423, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 567 + ], + "type": "text", + "content": "Evaluation Protocols. We evaluate independent and plug-in models separately for fair comparisons on the two datasets [4, 21] with depth ranges of 50, 70, and 80 meters [21, 31, 37, 40]. On nuScenes [4], previous state-of-the-art methods [12, 20, 21, 26, 28, 31, 37, 40] are compared comprehensively to the best of our knowledge. For the recently published ZJU-4DRadarCam [21], since most prior works do not include experiments on the dataset, we can only follow the same comparative methods [21, 28, 37] and results as RadarCam-Depth [21]. Besides, we also evaluate plug-in models with different depth predictors [3, 35, 51] to showcase model flexibility on the ZJU-4DRadarCam [21]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt the commonly-applied depth metrics MAE, RMSE, iMAE, iRMSE, " + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 313, + 567, + 553, + 604 + ], + "type": "text", + "content": ", and Rel as prior arts [21, 26, 31, 37]. See supplement for more details." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 611, + 447, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 447, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 447, + 624 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "Data Processing. We adhere to the data processing procedures of prior arts [21, 37]. For the nuScenes [4], the " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "D_{acc}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " in Eq. 5 is generated by accumulating the " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "D_{gt}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " of 160 nearby frames [21, 37]. For the ZJU-4DRadarCam dataset [21], the " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "D_{acc}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " is obtained via the linear interpolation [1, 21] of " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "D_{gt}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": ". Model Architecture. In line with prior arts [21, 37], the ResNet-18 [14] is adopted as the image encoder. The graph" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10527" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 529, + 216 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 216 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 216 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 216 + ], + "type": "table", + "html": "
TypeMethodRadar FramesImagesTime(ms)↓0 - 50m0 - 70m0 - 80m
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
Indep-endentLin et al. [26] (IROS'20)31117.62034.94316.52294.75338.22371.05623.0
RC-PDA [31] (CVPR'21)5379.22225.04156.53326.16700.63713.67692.8
R4Dyn [12] (3DV'21)41------6434.0
DORN [28] (ICIP'21)5(×3)1392.51926.64124.82380.65252.72467.75554.3
Singh et al. [37] (CVPR'23)1194.21727.73746.82073.24590.72179.34898.7
CaFNet [40] (IROS'24)11103.91674.23674.52010.34493.12109.84765.6
Li et al. [20] (ECCV'24)1167.61524.53567.31822.94303.61927.04609.6
TacoDepth (Ours)1126.71423.63275.81712.63960.51833.44150.2
Plug-inRadarCam-Depth [21] (ICRA'24)11358.31286.12964.31587.93662.51689.73948.0
TacoDepth (Ours)1129.31046.82487.51347.13152.81492.43324.8
", + "image_path": "b0138951cc71c5ede6cdc87185b52c3d8c5b18f227274b09d184c56c0ff1ab18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 219, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 555, + 285 + ], + "type": "text", + "content": "Table 1. Comparisons with state-of-the-art methods on the nuScenes dataset [4] (in millimeters). Independent [12, 20, 26, 28, 31, 37, 40] and plug-in [21] models are evaluated separately, with maximum evaluation distances of 50, 70, and 80 meters as prior arts [21, 37]. Plug-in models are compared using the same depth predictor DPT-Hybrid [35] as RadarCam-Depth [21]. We report the average runtime of processing one " + }, + { + "bbox": [ + 55, + 219, + 555, + 285 + ], + "type": "inline_equation", + "content": "900 \\times 1600" + }, + { + "bbox": [ + 55, + 219, + 555, + 285 + ], + "type": "text", + "content": " frame and 30 Radar points by different methods on one NVIDIA RTX A6000 GPU. The inference time of plug-in models does not contain initial depth prediction. Some methods [12, 26, 28, 31] require multiple frames and Radar scans for inference, while others [20, 21, 37, 40] only need a single frame as input. The best performance is in boldface. Second best is underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 82, + 290, + 529, + 485 + ], + "blocks": [ + { + "bbox": [ + 82, + 290, + 529, + 485 + ], + "lines": [ + { + "bbox": [ + 82, + 290, + 529, + 485 + ], + "spans": [ + { + "bbox": [ + 82, + 290, + 529, + 485 + ], + "type": "image", + "image_path": "dbaae734e43abccef215ebd53b3aeb0d7c53b5feb682ca5416b2f07576172f44.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 490, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 55, + 490, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 555, + 514 + ], + "type": "text", + "content": "Figure 4. Visual results of independent models on nuScenes [4]. Both daytime and nighttime samples are presented. Prior arts [31, 37] exhibit disrupted structures and noticeable artifacts. TacoDepth produces accurate depth with finer details and more complete structures." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": "based Radar structure extractor contains " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": " layers with the GNN [11, 36] architecture. The widths " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "a_{l}" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": " of Radar-centered areas in Eq. 1 and Eq. 2 are set to " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\{48, 32, 16\\}" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": " for the three layers. Refer to the supplement for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": "Training Recipe. The model is trained by Adam optimizer on two NVIDIA A6000 GPUs with a batch size of 12 for 50 epochs. The initial learning rate is " + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": " and decreases by " + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": " for every ten epochs. The input image size is " + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "inline_equation", + "content": "900 \\times 1600" + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": " [21, 37] and " + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "inline_equation", + "content": "300 \\times 1280" + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": " [21] on nuScenes [4] and ZJU-4DRadarCam [21], respectively. The " + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 582, + 295, + 666 + ], + "type": "text", + "content": " in Eq. 5 is set to one. All other settings are consistent with prior works [21, 37]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 284, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 284, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 284, + 685 + ], + "type": "text", + "content": "4.3. Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Comparisons with Independent Models. Independent models [12, 20, 26, 28, 29, 31, 37, 38, 40] predict dense" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "content": "metric depth without using relative depth predictors [3, 34, 35, 51]. Most of them adopt two-stage methods with intermediate quasi-dense depth, which are inefficient and not robust. Our TacoDepth significantly surpasses these methods in efficiency and accuracy with one-stage fusion. For example, in Table 1, compared with the recent CaFNet [40], our model reduces RMSE, MAE, and inference time by over " + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "inline_equation", + "content": "10.9\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "inline_equation", + "content": "74.3\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "content": " on nuScenes dataset [4]. Additionally, Li et al. [20] aim to improve the training paradigm under sparse supervision by their Disruption-Compensation framework. Without using such advanced training and supervision strategy, TacoDepth still outperforms Li et al. [20] by over " + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "inline_equation", + "content": "60.5\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "content": " for runtime, MAE, and RMSE, showing the efficacy of our proposed architectures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Similarly, in Table 2, compared with Singh et al. [37]," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10528" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 529, + 148 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 148 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 148 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 148 + ], + "type": "table", + "html": "
TypeMethod0 - 50m0 - 70m0 - 80m
MAE↓RMSE↓iMAE↓iRMSE↓MAE↓RMSE↓iMAE↓iRMSE↓MAE↓RMSE↓iMAE↓iRMSE↓
Indep-endentDORN [28] (ICIP'21)2210.24129.719.831.92402.24625.219.831.92447.64760.019.931.9
Singh et al. [37] (CVPR'23)1785.43704.618.135.31932.74137.118.035.21979.54309.317.935.1
TacoDepth (Ours)1120.12686.712.825.01181.82906.312.724.91201.12990.712.724.9
Plug-inRadarCam-Depth [21] (ICRA'24)1067.52817.410.522.91157.03117.710.422.91183.53229.010.422.8
TacoDepth (Ours)930.22477.39.320.8983.12779.69.320.91032.52850.39.420.9
", + "image_path": "95ed752ab587a53076e237bf0903093410c0e458cf6d9f0f770bae89856c2ca9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 151, + 555, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 151, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 555, + 185 + ], + "type": "text", + "content": "Table 2. Comparisons with state-of-the-art methods on ZJU-4DRadarCam [21] (in millimeters). Plug-in models are compared using the same relative depth predictor DPT-Hybrid [35]. Since many prior works do not include experiments on this new dataset, we adhere to the same compared methods [21, 28, 37] and metrics as RadarCam-Depth [21]. Best performance is in boldface. Second best is underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 82, + 193, + 529, + 445 + ], + "blocks": [ + { + "bbox": [ + 82, + 193, + 529, + 445 + ], + "lines": [ + { + "bbox": [ + 82, + 193, + 529, + 445 + ], + "spans": [ + { + "bbox": [ + 82, + 193, + 529, + 445 + ], + "type": "image", + "image_path": "e46bdf0b07e2ce89f1b4d1888720194317b6101fe78e2c0d317541fad59cfe01.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 448, + 555, + 471 + ], + "lines": [ + { + "bbox": [ + 55, + 448, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 555, + 471 + ], + "type": "text", + "content": "Figure 5. Visual results of plug-in models [21] on the ZJU-4DRadarCam [21]. The same depth predictor DPT-Hybrid [35] is adopted for fair comparisons. Regions with obvious differences are highlighted in the rectangular boxes. Best view zoomed in on-screen for details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "text", + "content": "TacoDepth reduces depth errors by over " + }, + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "inline_equation", + "content": "37.3\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "inline_equation", + "content": "27.5\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "inline_equation", + "content": "28.7\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "inline_equation", + "content": "29.1\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 539 + ], + "type": "text", + "content": " for MAE, RMSE, iMAE, and iRMSE on ZJU-4DRadarCam [21]. The superior efficiency and accuracy prove the effectiveness of our one-stage model design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "text", + "content": "Comparisons with Plug-in Models. Plug-in models [21] convert relative depth to metric depth by fusing the RGB and Radar data. Our method supports both independent and plug-in processing by the auxiliary input branch. For the plug-in mode, in Table 1, TacoDepth decreases the MAE and RMSE by over " + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "inline_equation", + "content": "13.9\\%" + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "text", + "content": " than RadarCam-Depth [21] on nuScenes [4]. Owing to their complex four-stage framework [21], we even improve the speed by " + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "inline_equation", + "content": "91.8\\%" + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "text", + "content": ". Similarly, in Table 2, TacoDepth reduces the MAE, RMSE, iMAE, and iRMSE by over " + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "inline_equation", + "content": "12.8\\%" + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "inline_equation", + "content": "10.8\\%" + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 54, + 542, + 295, + 698 + ], + "type": "text", + "content": " on ZJU-4DaradarCam [21]. The results prove the generality and flexibility of TacoDepth, providing a better balance of model efficiency and accuracy for different users." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "To be mentioned, in the plug-in mode, TacoDepth can" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 491, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 528 + ], + "type": "text", + "content": "be compatible with various relative depth predictors [3, 35, 51]. It can benefit from stronger depth models [3, 51] to achieve higher accuracy, which will be presented in Table 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 530, + 556, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 556, + 674 + ], + "type": "text", + "content": "Qualitative Comparisons. The visual results of independent models [31, 37] are showcased in Fig. 4. Both daytime and nighttime samples are compared. TacoDepth can predict accurate depth maps with more complete structures (e.g., the cars and buses) and more meticulous details (e.g., the pedestrian overpass and metal guardrail). The plug-in models [21] are compared in Fig. 5. Based on the same depth predictor DPT-Hybrid [35], our method produces noticeably better structures and details in the rectangular boxes. The improvements are attributed to our graph-based Radar structure extractor and pyramid-based Radar fusion module, which extract and fuse Radar structures effectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Model Robustness. Multi-stage methods [26, 28, 31, 37, 40] predict quasi-dense depth as an intermediate step. However, as shown in Fig. 2, the intermediate results are still" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10529" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 294, + 133 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 133 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 133 + ], + "type": "table", + "html": "
TypeMethodParams (M)↓FLOPs (G)↓Time (ms)↓
Indep-endentDORN [28] (ICIP'21)183.38923.47392.5
Singh et al. [37] (CVPR'23)22.81502.0994.2
TacoDepth (Ours)13.47139.3026.7
Plug-inRadarCam-Depth [21] (ICRA'24)33.26619.02358.3
TacoDepth (Ours)14.25139.8729.3
", + "image_path": "1acc8f52cdcd4886da6f9bce4bd0ca90d719228d7f971d3877c62a175b4b6595.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 200, + 294, + 248 + ], + "blocks": [ + { + "bbox": [ + 55, + 140, + 295, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 196 + ], + "type": "text", + "content": "Table 3. Model efficiency. We report FLOPs, model parameters, and runtime of different methods. Since varied depth predictors can be applied, the results of plug-in models do not contain initial depth prediction. FLOPs and runtime are evaluated with one " + }, + { + "bbox": [ + 55, + 140, + 295, + 196 + ], + "type": "inline_equation", + "content": "900 \\times 1600" + }, + { + "bbox": [ + 55, + 140, + 295, + 196 + ], + "type": "text", + "content": " image and 30 Radar points on one NVIDIA A6000 GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 200, + 294, + 248 + ], + "lines": [ + { + "bbox": [ + 58, + 200, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 58, + 200, + 294, + 248 + ], + "type": "table", + "html": "
InitialRadarCam-Depth [21]TacoDepth (Ours)
δ1↑Rel↓MAE↓δ1↑Rel↓MAE↓δ1↑Rel↓MAE↓
DPT-Hybrid [35]0.9030.092-0.9210.0871157.00.9320.076983.1
MiDaS-v3.1 [3]0.9170.081-0.9260.0861177.20.9440.071852.4
Depth-Anything-v2 [51]0.9320.069-0.9400.073924.70.9610.059730.2
", + "image_path": "0137d81d96af557e94b1acac99b67e072ca08b94850bcf4a2d64924a9d1324a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 302, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 302, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 54, + 302, + 295, + 410 + ], + "type": "text", + "content": "sparse and noisy, especially on challenging nighttime and glaring scenes, leading to disrupted structures, blurred details, and obvious artifacts in their final depth predictions, as presented in Fig. 4. In contrast, we propose the one-stage TacoDepth without the quasi-dense depth. By effectively integrating the complementary image details and Radar structures, our model produces robust depth for both daytime and nighttime scenarios in Fig. 4. See supplement for more quantitative and visual results on nighttime scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 157, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 157, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 157, + 431 + ], + "type": "text", + "content": "4.4. Model Efficiency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "type": "text", + "content": "We evaluate the processing time of different methods in Table 1. Compared with the second-best independent model of Li et al. [20] and the plug-in model RadarCam-Depth [21], TacoDepth improves inference speed by " + }, + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "type": "inline_equation", + "content": "60.5\\%" + }, + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "type": "inline_equation", + "content": "91.8\\%" + }, + { + "bbox": [ + 55, + 435, + 295, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "text", + "content": "To further showcase our efficiency, we report model parameters and FLOPs in Table 3. Compared with the two-stage Singh et al. [37], the one-stage TacoDepth reduces the parameters and FLOPs by " + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "inline_equation", + "content": "40.9\\%" + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "inline_equation", + "content": "72.2\\%" + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "text", + "content": ". Compared with the four-stage plug-in model RadarCam-Depth [21], we decrease parameters and FLOPs by " + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "inline_equation", + "content": "77.4\\%" + }, + { + "bbox": [ + 55, + 484, + 295, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 604 + ], + "type": "text", + "content": "Note that, the differences between the plug-in and independent modes of our method arise from the auxiliary branch to fuse initial relative depth. TacoDepth achieves real-time processing of 37.5 fps by independent inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 156, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 156, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 156, + 624 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "Plug-in Inference. In Table 4, our TacoDepth outperforms the RadarCam-Depth [21] with three different relative depth predictors [3, 35, 51]. Using more advanced depth predictors, e.g., MiDaS-v3.1 [3] and Depth-Anything-v2 [51], can further boost our accuracy without extra effort. The results demonstrate the versatility and flexibility of our approach. Efficacy of TacoDepth. In Table 5, we ablate the graph" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 72, + 553, + 128 + ], + "blocks": [ + { + "bbox": [ + 55, + 255, + 295, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 289 + ], + "type": "text", + "content": "Table 4. Plug-in models with varied depth predictors. Our method outperforms RadarCam-Depth [21] with different depth predictors [3, 35, 51] on ZJU-4DRadarCam [21] (0-70 meters)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 72, + 553, + 128 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 128 + ], + "type": "table", + "html": "
MethodRGBRadarGEPFMAE↓RMSE↓
RGB Baseline✓✗✗✗2474.35402.1
Singh et al. [37] (CVPR'23)✓✓✗✗2073.24590.7
TacoDepth (w/o GE)✓✓✗✓1815.64189.8
TacoDepth (Ours)✓✓✓✓1712.63960.5
", + "image_path": "680ddb6326fdff00ac0a511809c0d8f784355409019629fa0bf2407b04b39111.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 348, + 202, + 520, + 243 + ], + "blocks": [ + { + "bbox": [ + 313, + 131, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 131, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 131, + 555, + 198 + ], + "type": "text", + "content": "Table 5. Efficacy of the TacoDepth. We ablate the effectiveness of graph-based Radar structure extractor (GE) and pyramid-based Radar fusion (PF) for Radar structure extraction and integration. We adopt the same baseline as Singh et al. [37]. TacoDepth " + }, + { + "bbox": [ + 313, + 131, + 555, + 198 + ], + "type": "inline_equation", + "content": "(w / o" + }, + { + "bbox": [ + 313, + 131, + 555, + 198 + ], + "type": "text", + "content": " GE) uses the same MLP to extract Radar point features as prior art [37]. The results are evaluated on nuScenes [4] in 0-70 meters." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 202, + 520, + 243 + ], + "lines": [ + { + "bbox": [ + 348, + 202, + 520, + 243 + ], + "spans": [ + { + "bbox": [ + 348, + 202, + 520, + 243 + ], + "type": "table", + "html": "
MetricL=1L=2L=3L=4
MAE↓1840.61787.81712.61729.3
RMSE↓4203.74046.93960.53984.2
", + "image_path": "c1a82b7127a5e76c32c8c230479175c3fcf1812f1243fb4a5ec971c5b1ea0bf4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 245, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 314, + 245, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 314, + 245, + 553, + 257 + ], + "type": "text", + "content": "Table 6. Layer number L. Results are on nuScenes [4] in " + }, + { + "bbox": [ + 314, + 245, + 553, + 257 + ], + "type": "inline_equation", + "content": "0 - 70\\mathrm{m}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "text", + "content": "based Radar structure extractor (GE) and pyramid-based Radar fusion module (PF). With the same MLP to extract Radar point features [37], TacoDepth " + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "inline_equation", + "content": "(w / o" + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "text", + "content": " GE) reduces MAE and RMSE by " + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "inline_equation", + "content": "12.4\\%" + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "text", + "content": " than Singh et al. [37], which are attributed to the PF. Replacing the MLP [37] with GE further decreases MAE and RMSE by " + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 313, + 272, + 553, + 343 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 344, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 553, + 392 + ], + "type": "text", + "content": "We adopt the RGB baseline of Singh et al. [37], a common single-image depth U-Net. Overall, TacoDepth improves accuracy beyond the baseline by " + }, + { + "bbox": [ + 313, + 344, + 553, + 392 + ], + "type": "inline_equation", + "content": "30.8\\%" + }, + { + "bbox": [ + 313, + 344, + 553, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 344, + 553, + 392 + ], + "type": "inline_equation", + "content": "26.7\\%" + }, + { + "bbox": [ + 313, + 344, + 553, + 392 + ], + "type": "text", + "content": " (MAE and RMSE), proving the efficacy of our approach." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "text", + "content": "Layer Number " + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "text", + "content": ". In Table 6, we ablate the layer number " + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "text", + "content": " of GE and PF. Compared with shallow layers " + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "inline_equation", + "content": "(L = 1,2)" + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "text", + "content": ", deeper layers " + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "inline_equation", + "content": "(L = 3)" + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "text", + "content": " better capture structures of point clouds [41, 45], producing higher accuracy. Due to Radar sparsity, more layers " + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "inline_equation", + "content": "(L = 4)" + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "text", + "content": " does not yield improvements but decreases efficiency. We adopt " + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 313, + 392, + 553, + 464 + ], + "type": "text", + "content": " in experiments." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 476, + 388, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 388, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 388, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 497, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 653 + ], + "type": "text", + "content": "In this paper, we propose TacoDepth, an efficient and accurate Radar-Camera depth estimation model with one-stage fusion. Different from the multi-stage methods with quasidense depth, TacoDepth utilizes the graph-based Radar structure extractor and pyramid-based Radar fusion module to capture and integrate Radar features, achieving superior efficiency and robustness. Moreover, TacoDepth is flexible for independent and plug-in processing, offering a better balance of speed and accuracy. Our work provides a new perspective on the efficient Radar-Camera depth estimation. Limitations. We currently provide one efficient implementation of our TacoDepth. In future work, more techniques can be explored for varied performances and applications." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgments. The study is supported under RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This research is also supported by the MoE AcRF Tier 1 grant (RG14/22)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10530" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] C Bradford Barber, David P Dobkin, and Hannu Huhdanpaa. The quickhull algorithm for convex hulls. ACM Transactions on Mathematical Software (TOMS), 22(4):469-483, 1996. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 296, + 178 + ], + "type": "text", + "content": "[2] Oded Bialer and Yuval Haitman. Radsimreal: Bridging the gap between synthetic and real data in radar object detection with simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15407-15416, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 178, + 295, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 178, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 178, + 295, + 220 + ], + "type": "text", + "content": "[3] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3. 1-a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 223, + 296, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 223, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 62, + 223, + 296, + 298 + ], + "type": "text", + "content": "[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11618-11628, 2020. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 300, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 296, + 354 + ], + "type": "text", + "content": "[5] Yujeong Chae, Hyeonseong Kim, and Kuk-Jin Yoon. Towards robust 3d object detection with lidar and 4d radar fusion in various weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15162-15172, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 387 + ], + "type": "text", + "content": "[6] Feipeng Chen, Lihui Wang, Pinzhang Zhao, and Jian Wei. 3d grid-based vdbscan clustering and radar-monocular depth completion. IEEE Sensors Journal, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 419 + ], + "type": "text", + "content": "[7] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 453 + ], + "type": "text", + "content": "[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. 35, 2022. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 453, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 296, + 508 + ], + "type": "text", + "content": "[9] Fangqiang Ding, Andras Palffy, Dariu M. Gavrila, and Chris Xiaoxuan Lu. Hidden gems: 4d radar scene flow learning using cross-modal supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9340-9349, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 508, + 296, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 508, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 57, + 508, + 296, + 562 + ], + "type": "text", + "content": "[10] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6824-6835, 2021. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 563, + 295, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 563, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 57, + 563, + 295, + 616 + ], + "type": "text", + "content": "[11] Kexue Fu, Shaolei Liu, Xiaoyuan Luo, and Manning Wang. Robust point cloud registration framework based on deep graph matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8893-8902, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 617, + 295, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 617, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 57, + 617, + 295, + 671 + ], + "type": "text", + "content": "[12] Stefano Gasperini, Patrick Koch, Vinzensz Dallabetta, Nassir Navab, Benjamin Busam, and Federico Tombari. R4dyn: Exploring radar for self-supervised monocular depth estimation of dynamic scenes. In International Conference on 3D Vision (3DV), pages 751-760, 2021. 1, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 672, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 672, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 57, + 672, + 295, + 715 + ], + "type": "text", + "content": "[13] Kyle Harlow, Hyesu Jang, Timothy D. Barfoot, Ayoung Kim, and Christoffer Heckman. A new wave in robotics: Survey on recent mmwave radar applications in robotics. IEEE Transactions on Robotics, 40:4544-4560, 2024. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 698 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 2, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "type": "text", + "content": "[15] Daniel Casado Herreraz, Matthias Zeller, Le Chang, Ignacio Vizzo, Michael Heidingsfeld, and Cyril Stachniss. Radar-only odometry and mapping for autonomous vehicles. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 539-548. IEEE, 2019. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "type": "text", + "content": "[16] Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Implicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24118-24129, 2024. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 306 + ], + "type": "text", + "content": "[17] Andrey Ignatov, Grigory Malivenko, David Plowman, Samarth Shukla, and Radu Timofte. Fast and accurate single-image depth estimation on mobile devices, mobile ai 2021 challenge: Report. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2545-2557, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "text", + "content": "[18] Yinuo Jiang, Beitong Zhou, Xiaoyu Liu, Qingyi Li, and Cheng Cheng. Gtinet: Global topology-aware interactions for unsupervised point cloud registration. IEEE Transactions on Circuits and Systems for Video Technology, 34(7): 6363-6375, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 363, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 406 + ], + "type": "text", + "content": "[19] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, 2012. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "type": "text", + "content": "[20] Huadong Li, Minhao Jing, Jiajun Liang, Haoqiang Fan, and Renhe Ji. Sparse beats dense: Rethinking supervision in radar-camera depth completion. arXiv preprint arXiv:2312.00844, 2023. 1, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "text", + "content": "[21] Han Li, Yukai Ma, Yaqing Gu, Kewei Hu, Yong Liu, and Xingxing Zuo. Radarcam-depth: Radar-camera fusion for depth estimation with learned metric scale. arXiv preprint arXiv: 2401.04325, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 563 + ], + "type": "text", + "content": "[22] Jiaqi Li, Yiran Wang, Zihao Huang, Jinghong Zheng, Ke Xian, Zhiguo Cao, and Jianming Zhang. Diffusion-augmented depth prediction with sparse annotations. In Proceedings of the 31st ACM International Conference on Multimedia, page 2865-2876, New York, NY, USA, 2023. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 564, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 564, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 564, + 553, + 597 + ], + "type": "text", + "content": "[23] Jiaqi Li, Yiran Wang, Jinghong Zheng, Zihao Huang, Ke Xian, Zhiguo Cao, and Jianming Zhang. Self-distilled depth refinement with noisy poisson fusion, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 598, + 553, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 598, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 316, + 598, + 553, + 652 + ], + "type": "text", + "content": "[24] Xingyi Li, Chaoyi Hong, Yiran Wang, Zhiguo Cao, Ke Xian, and Guosheng Lin. Symmernf: Learning to explore symmetry prior for single-view view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1726-1742, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 654, + 553, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 698 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 698 + ], + "type": "text", + "content": "[25] Zecheng Li, Yuying Song, Fuyuan Ai, Chunyi Song, and Zhiwei Xu. Semantic-guided depth completion from monocular images and 4d radar data. IEEE Transactions on Intelligent Vehicles, pages 1-13, 2024. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10531" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 715 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[26] Juan-Ting Lin, Dengxin Dai, and Luc Van Gool. Depth estimation from monocular images and sparse radar data. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10233-10240. IEEE, 2020. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[27] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. pages 14928-14937, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[28] Chen-Chou Lo and Patrick Vandewalle. Depth estimation from monocular images and sparse radar using deep ordinal regression network. In IEEE International Conference on Image Processing (ICIP), pages 3343-3347, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[29] Chen-Chou Lo and Patrick Vandewalle. Rcdpt: Radar-camera fusion dense prediction transformer. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "text", + "content": "[30] Chen-Chou Lo and Patrick Vandewalle. Expanding sparse radar depth based on joint bilateral filter for radar-guided monocular depth estimation. Sensors, 24(6):1864, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 373 + ], + "type": "text", + "content": "[31] Yunfei Long, Daniel Morris, Xiaoming Liu, Marcos Castro, Punarjay Chakravarty, and Praveen Narayanan. Radar-camera pixel depth association for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12507-12516, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 439 + ], + "type": "text", + "content": "[32] Guohao Peng, Heshan Li, Yangyang Zhao, Jun Zhang, Zhenyu Wu, Pengyu Zheng, and Danwei Wang. Transloc4d: Transformer-based 4d radar place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17595-17605, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 441, + 294, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 650 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 650 + ], + "type": "text", + "content": "[33] Pierluigi Zama Ramirez, Fabio Tosi, Luigi Di Stefano, Radu Timofte, Alex Costanzino, Matteo Poggi, Samuele Salti, Stefano Mattoccia, Yangyang Zhang, Cailin Wu, Zhuangda He, Shuangshuang Yin, Jiaxu Dong, Yangchenxu Liu, Hao Jiang, Jun Shi, Yong A, Yixiang Jin, Dingzhe Li, Bingxin Ke, Anton Obukhov, Tinafu Wang, Nando Metzger, Shengyu Huang, Konrad Schindler, Yachuan Huang, Jiaqi Li, Junrui Zhang, Yiran Wang, Zihao Huang, Tianqi Liu, Zhiguo Cao, Pengzhi Li, Jui-Lin Wang, Wenjie Zhu, Hui Geng, Yuxin Zhang, Long Lan, Kele Xu, Tao Sun, Qisheng Xu, Sourav Saini, Aashray Gupta, Sahaj K. Mistry, Aryan Shukla, Vinit Jakhetiya, Sunil Jaiswal, Yuejin Sun, Zhuofan Zheng, Yi Ning, Jen-Hao Cheng, Hou-I Liu, Hsiang-Wei Huang, Cheng-Yen Yang, Zhongyu Jiang, Yi-Hao Peng, Aishi Huang, and Jenq-Neng Hwang. Ntire 2024 challenge on hr depth from images of specular and transparent surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6499–6512, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 651, + 294, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 294, + 705 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 294, + 705 + ], + "type": "text", + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(03):1623-1637, 2020. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 706, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 706, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 706, + 294, + 715 + ], + "type": "text", + "content": "[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 710 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "sion transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12179-12188, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 108, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 555, + 150 + ], + "type": "text", + "content": "[36] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks and Learning Systems, 20(1):61-80, 2009. 2, 3, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "type": "text", + "content": "[37] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9275-9285, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 219, + 554, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 273 + ], + "type": "text", + "content": "[38] Tieshuai Song, Bin Yang, Jun Wang, Guidong He, Zhao Dong, and Fengjun Zhong. mmwave radar and image fusion for depth completion: a two-stage fusion network. In The 27th International Conference on Information Fusion (FUSION), 2024. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 275, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 554, + 329 + ], + "type": "text", + "content": "[39] Christian Stetco, Barnaba Ubezio, Stephan Muhlbacher-Karrer, and Hubert Zangl. Radar sensors in collaborative robotics: Fast simulation and experimental validation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10452-10458. IEEE, 2020. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 331, + 554, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 373 + ], + "type": "text", + "content": "[40] Huawei Sun, Hao Feng, Julius Ott, Lorenzo Servadei, and Robert Wille. Cafnet: A confidence-driven framework for radar camera depth estimation. arXiv preprint arXiv:2407.00697, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 375, + 554, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 554, + 440 + ], + "type": "text", + "content": "[41] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6411–6420, 2019. 2, 3, 4, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 442, + 554, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 554, + 485 + ], + "type": "text", + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, 2017. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 487, + 554, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 554, + 530 + ], + "type": "text", + "content": "[43] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3056-3065, 2019. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 532, + 554, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 554, + 575 + ], + "type": "text", + "content": "[44] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3056-3065, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 577, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 554, + 620 + ], + "type": "text", + "content": "[45] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5), 2019. 2, 3, 4, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 622, + 555, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 555, + 676 + ], + "type": "text", + "content": "[46] Yiran Wang, Xingyi Li, Min Shi, Ke Xian, and Zhiguo Cao. Knowledge distillation for fast and accurate monocular depth estimation on mobile devices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2457-2465, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 677, + 554, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 554, + 710 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 554, + 710 + ], + "type": "text", + "content": "[47] Yiran Wang, Zhiyu Pan, Xingyi Li, Zhiguo Cao, Ke Xian, and Jianming Zhang. Less is more: Consistent video depth estimation with masked frames modeling. In Proceedings" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10532" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 239 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "of the 30th ACM International Conference on Multimedia, page 6347-6358, New York, NY, USA, 2022. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[48] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9466-9476, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 226 + ], + "type": "text", + "content": "[49] Yiran Wang, Min Shi, Jiaqi Li, Chaoyi Hong, Zihao Huang, Juewen Peng, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Nvds+: Towards efficient and versatile neural stabilizer for video depth estimation. IEEE transactions on pattern analysis and machine intelligence, pages 1-18, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 228, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 228, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 228, + 294, + 239 + ], + "type": "text", + "content": "[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar," + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 240 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. 34:12077-12090, 2021. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[51] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "text", + "content": "[52] Junrui Zhang, Jiaqi Li, Yachuan Huang, Yiran Wang, Jinghong Zheng, Liao Shen, and Zhiguo Cao. Towards robust monocular depth estimation in non-lambertian surfaces, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[53] Yi Zhou, Lulu Liu, Haocheng Zhao, Miguel López-Benítez, Limin Yu, and Yutao Yue. Towards deep radar perception for autonomous driving: Datasets, methods, and challenges. Sensors, 22(11):4208, 2022. 1, 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10533" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_content_list.json b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..78cc1103c01083efbcdf611d56f3fff8135b4818 --- /dev/null +++ b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_content_list.json @@ -0,0 +1,1957 @@ +[ + { + "type": "text", + "text": "TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yoon Gyo Jung\\*1 Jaewoo Park\\*2 Jaeho Yoon\\*3 Kuan-Chuan Peng4 Wonchul $\\mathrm{Kim}^2$ Andrew Beng Jin Teoh Octavia Camps", + "bbox": [ + 225, + 179, + 772, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Northeastern University $^{2}$ AiV Co. $^{3}$ Yonsei University $^{4}$ Mitsubishi Electric Research Laboratories", + "bbox": [ + 254, + 224, + 740, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 250, + 330, + 323, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We aim to solve unsupervised anomaly detection in a practical challenging environment where the normal dataset is both contaminated with defective regions and its product class distribution is tailed but unknown. We observe that existing models suffer from tail-versus-noise trade-off where if a model is robust against pixel noise, then its performance deteriorates on tail class samples, and vice versa. To mitigate the issue, we handle the tail class and noise samples independently. To this end, we propose TailSampler, a novel class size predictor that estimates the class cardinality of samples based on a symmetric assumption on the class-wise distribution of embedding similarities. TailSampler can be utilized to sample the tail class samples exclusively, allowing to handle them separately. Based on these facets, we build a memory-based anomaly detection model TailedCore, whose memory both well captures tail class information and is noise-robust. We extensively validate the effectiveness of TailedCore on the unsupervised long-tail noisy anomaly detection setting, and show that TailedCore outperforms the state-of-the-art in most settings. Code is available in TailedCore", + "bbox": [ + 88, + 364, + 485, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 696, + 215, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the complex landscape of anomaly detection, the challenge often lies in navigating practical scenarios that feature a diverse and imbalanced data distribution [6, 8, 11, 21]. We delve into a challenging yet highly realistic scenario in anomaly detection, where the training dataset, comprising multiple product classes [39], is beset by two significant complications: contamination with noise [20] and presence of tailed class distributions [25]. In this context, 'tail' (few-shot) classes have limited few-shot samples, unlike 'head' (many-shot) classes which are data-rich but might contain samples with defects. Crucially, in this unsupervised set-", + "bbox": [ + 89, + 723, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08f2b2036c81a6d05697400ceebc490f94245dfd6bab865ac9ff521699b12b5f.jpg", + "image_caption": [ + "SimpleNet", + "WinCLIP", + "#", + "IVQ", + "#", + "bftPatch", + "", + "。", + "PaDiM", + "", + "Figure 1. Tail class (x-axis) and noisy head class (y-axis) performance comparison. The tail-versus-noise trade-off is shown across memory-based anomaly detection models (circles), and is more indicative in anomaly classification task evaluated by image-level AUROC (left)." + ], + "image_footnote": [], + "bbox": [ + 517, + 316, + 705, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ting, the class information of these products remains unknown to the model trainer. This setting mirrors a common predicament in large-scale industrial applications, where companies are required to manage vast production lines encompassing a wide array of product types, each with varying production rates.", + "bbox": [ + 509, + 535, + 906, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A critical issue arises in this scenario, which we term the \"tail-versus-noise dilemma.\" The dilemma underscores a fundamental trade-off where models robust to class imbalance overfit to defect noises, likely misclassifying many defects as normal. In contrast, existing models that are robust against defect noise tend to underfit, missing vital details in tail classes. Fig. 1 shows this tail-versus-noise trade-off which is prevalent through existing anomaly detection algorithms. This intrinsic characteristic comes from the fact that few-shot class features are equivalent to the noisy abnormal features in terms of their statistical occurrence. Consequently, existing models find it difficult to strike a balance between avoiding overfitting to noise and effectively capturing the nuances of less represented classes.", + "bbox": [ + 509, + 612, + 908, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To resolve this challenge, we propose to handle tail classes independently from head classes by exclusively sampling them. To this end, we devise a novel class size predictor TailSampler that estimates the cardinality of the class of any training sample. For accurate estimation, TailSampler estimates the class sizes based on a reflective symmetry between inter-class", + "bbox": [ + 509, + 810, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* equal contribution † corresponding author", + "bbox": [ + 89, + 886, + 312, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25539", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and intra-class similarity distributions in the embedding space. The estimated class sizes of the training samples then enable to sample tail class features exclusively from the training set.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By utilizing the aggregated few-shot class instances, we devise a memory-based anomaly detector TailedCore, which is built upon the noise-discriminated model by augmenting the tail class patch features sampled with TailSampler. As the memory of TailedCore is both clean from noisy defect patches and preserves the few-shot class information, it exhibits robustness against both noise contamination and class imbalance of the training set.", + "bbox": [ + 89, + 137, + 482, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 258, + 395, + 271 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore a practical and challenging anomaly detection scenario, where a single detection model is trained on multi-class product training samples that involve noise contamination and whose class distribution is long-tailed.", + "- We propose a memory-based anomaly detector TailedCore whose patch memory bank is both noise-free and augmented with the representative patches of few-shot class instances. To build TailedCore, we introduce TailSampler that can aggregate few-shot class instances exclusively. TailSampler is based on a novel class size predictor that estimates the class size of training samples based on the reflective symmetry between inter-class and intra-class class-wise embedding similarity distributions.", + "- We conduct extensive analyses and comparative studies on both TailedCore and few-shot sampler along with the proposed class size predictor on the proposed unsupervised long-tail noisy anomaly detection benchmark." + ], + "bbox": [ + 89, + 273, + 482, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The codes for TailedCore along with the data generation will be open sourced. We use the term 'few-shot' and 'tail' interchangeably to denote the class with a small number of samples.", + "bbox": [ + 89, + 531, + 483, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 89, + 590, + 228, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Anomaly detection in extreme scenarios", + "text_level": 1, + "bbox": [ + 89, + 617, + 431, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One vitally important aspect of anomaly detection [9, 34] is the environment where these models are tested. This exploration has led to the identification of several challenging scenarios, including multi-class settings with training data spanning multiple product types [26, 39], zero-shot and few-shot settings with minimal training images [15, 19, 23, 38, 43], and noisy settings featuring uncleaned training data with defective images [7, 20, 28].", + "bbox": [ + 89, + 638, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-class anomaly detection In contrast to single-product settings, anomaly detection models trained on multi-class datasets often encounter \"shortcut learning\" [39], leading to low anomaly scores for normal images. This issue is particularly pronounced in reconstruction-based models, which tend to reconstruct all samples, regardless of their normality or abnormality. To resolve it, UniAD [39] employs neighbor masked attention within a transformer model, focusing attention on local parts to prevent shortcut learning. HVQ [26] utilizes vector quantization to discretize the latent space, restricting the network's recon", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "striction fidelity on normal data distributions. UniFormaly [22] increases sensitivity to irrelevant patterns by masking less relevant patches and employing top-k matchings with a pretrained foundation model. These multi-class anomaly detection methods do not consider either long-tail classes or contaminated training data setups, and do not work well in the combined \"noisy long-tail\" setup, which is shown in our experimental results.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zero/few-shot anomaly detection The inception of few-shot learning benchmarks in anomaly detection marked a significant advancement, with methods leveraging template matching [18] and graph representations [38] for minimal training data scenarios. AnomalyCLIP [43], WinCLIP [19], and AnomalyGPT [17], on the other hand, detect anomalies in the few-shot classes based on the acquired knowledge of pretrained foundation models such as CLIP [33] and GPT [32]. However, these methods often require sophisticated prompt engineering or delicate prompt-tuning, and have yet to achieve optimal precision and recall in the single-class training setting with a sufficient amount of training samples. In contrast, our proposed TailedCore needs none of these process.", + "bbox": [ + 511, + 199, + 908, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Noise anomaly detection Dealing with training datasets contaminated with defective samples presents another extreme learning scenario. SoftPatch [20] and InReaCh [28] focus on purifying memory coresets through outlier detection algorithms and distance association filtering, respectively. IGD [7] adopts a robust Gaussian modeling approach, addressing both contaminated training data and small dataset challenges. Yet, these methods do not fully consider scenarios involving both data contamination and class imbalance, unlike our proposed TailedCore.", + "bbox": [ + 511, + 398, + 908, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Long-tail noisy learning", + "text_level": 1, + "bbox": [ + 511, + 544, + 733, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the realm of image classification, several studies have focused on learning scenarios involving training sets with label contamination and class imbalance. To address these challenges, the existing literature relies on the robustness and unsupervised aspects of self-supervised learning [14, 27, 37, 41]. The learned embeddings of self-supervised frameworks are less impacted by label contamination and class imbalances. However, these works differ from ours in that they deal with label-level contamination rather than pixel-level defects. Furthermore, these methods require explicit supervised labels to remediate the contamination. In contrast, our work targets noise in pixels without any class label information.", + "bbox": [ + 511, + 566, + 908, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Few-shot/outlier sampling", + "text_level": 1, + "bbox": [ + 511, + 757, + 750, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Both classical clustering methods and outlier detection algorithms can be applied to our task, as they facilitate the identification of few-shot tail class samples, either indirectly or directly. Unsupervised clustering algorithms [2, 13, 16] can estimate the cluster sizes but struggle to identify small sample clusters in the imbalanced data. DBSCAN [13], on the other hand, captures both clusters and outliers. This however can be problematic when attempting to identify clusters with only one", + "bbox": [ + 511, + 780, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25540", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sample (1-shot) as it may be confused between few-shot samples and outliers. Unlike clustering methods, outlier detection algorithms [1, 5, 24, 35] are adept at identifying 1-shot samples which inherently resemble outliers. However, their effectiveness diminishes when a few-shot class contains more than one sample, as these no longer fit the typical outlier profile, potentially leading to incorrect identification of few-shot class samples. Both algorithm types, due to their inherent design, struggle to accurately predict class/cluster sizes in datasets that are both imbalanced and contaminated. Our proposed class size predictor specifically addresses this issue, aiming to reliably and accurately identify few-shot class samples and enable their exclusive sampling. Ultimately, the identified few-shot class samples are augmented to the noise-discriminated patch set, constituting the memory-based anomaly detection model TailedCore, which is robust against both class imbalance and training set noise.", + "bbox": [ + 89, + 90, + 483, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 89, + 344, + 215, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Memory-based anomaly detection", + "text_level": 1, + "bbox": [ + 89, + 369, + 387, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train a memory-based anomaly detector, a feature extractor $f$ (e.g., WideResNet) extracts a feature map $\\phi_i \\in \\mathbb{R}^{C \\times H \\times W}$ from each training input image $x_i$ . Then, a coreset selection algorithm $S_{core}$ is applied to the set $P = \\{\\phi_i^{(h,w)}\\}$ of patch features $\\phi_i^{(h,w)}$ , resulting in a memory bank", + "bbox": [ + 89, + 391, + 483, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\text {P a t c h C o r e}} := \\mathcal {S} _ {\\text {c o r e}} (P) \\subset P \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 479, + 483, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that contains representative patch features of $P$ . For instance, PatchCore opts for greedy sampling strategy to obtain the coreset. In the inference stage, the anomaly score $s^{(h,w)}$ of the testing image $x$ is computed for each patch feature $\\phi^{(h,w)}$ by measuring its distance to the memory", + "bbox": [ + 89, + 503, + 483, + 579 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s c o r e} ^ {(h, w)} = \\operatorname {d i s t} \\left(\\phi^ {(h, w)}, M\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 584, + 483, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $dist$ is the nearest neighbor distance in [34]. The pixel-level anomaly score is computed by up-scaling the score map $score = [score^{(h,w)}] \\in \\mathbb{R}^{H,W}$ , and its image-level anomaly score is defined as its pixel maximum $\\max_{h,w} score^{(h,w)}$ .", + "bbox": [ + 89, + 609, + 483, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Noise discrimination of memory", + "text_level": 1, + "bbox": [ + 89, + 679, + 374, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The training set $X = \\{x_{i}\\}$ may be contaminated with defect samples. The naively attained memory bank then contains abnormal patch features, and assigns a low anomaly score on defect images, resulting in poor performance. To resolve this, [20] applies a noise discrimination algorithm $S_{clean}$ on the patch feature set $P$ prior to coreset sampling:", + "bbox": [ + 89, + 700, + 483, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {\\text {c l e a n}} (P) = \\left\\{\\phi_ {i} ^ {(h, w)}: \\text {o u t l i e r} \\left(\\phi_ {i} ^ {(h, w)}\\right) < \\tau_ {0. 1 5} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 797, + 483, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "aggregating only the clean normal patches with low outlier scores. SoftPatch [20] exploits local outlier factor (LOF) for outlier scoring, and selects the threshold $\\tau_{0.15}$ to remove only $15\\%$ of whole patches. In inference, SoftPatch utilizes the memory set $M_{SoftPatch} = S_{core}(S_{clean}(P))$", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/06b4eb98e12f75e6d68ca17b8960d9f9ea0cbb1c919b0e1cd543fb93e8ec03a3.jpg", + "image_caption": [ + "Figure 2. (left) The ratio of removed patches based on highest outlier scores by Eq. (3), which shows that most of few-shot class patches are lost. (right) The ratio of sampled patches by greedy coreset sampling from PatchCore which favors both few-shot and anomaly samples." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 709, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/87aadc70551d1e473d78aa2ab458fd971f200f34b2fb6e4e45987b03de6c0509.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 90, + 901, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Problem and motivation", + "text_level": 1, + "bbox": [ + 511, + 271, + 728, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Problem", + "text_level": 1, + "bbox": [ + 511, + 296, + 616, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a practical anomaly detection setting where the normal training data $X = \\{x_{i}\\}$ is both multi-class with tail distribution and noise-contaminated. Particularly, each $x_{i}$ is paired with the product class label $y_{i} \\in \\mathcal{Y} = \\{1,2,\\dots,|\\mathcal{Y}|\\}$ (e.g., hazelnut, screw, pill, etc.) and an anomaly label $y_{i}^{a} \\in \\mathcal{Y}^{a} := \\{0,1\\}$ with 1 indicating anomaly. However, this labeling information is unknown to the model trainer. The class distribution is long-tailed such that there are tail (i.e., few-shot) classes whose number of training samples is very low; i.e.,", + "bbox": [ + 511, + 320, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left| C _ {k} \\right| = \\left| \\left\\{x _ {i}: y _ {i} ^ {c} = k \\right\\} \\right| \\leq K \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 470, + 906, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with a very small $K$ (e.g., 1, 4, or 20), where $|\\cdot|$ indicates the number of elements in the set. Moreover, some of the training samples in $X$ are anomalies, containing defect blobs in their images. We assume the tail class samples are all normal since in the unsupervised setting its normality is determined by its majority. Overall, the training set $X$ with the above descriptions constitutes unsupervised tailed noisy anomaly detection setting.", + "bbox": [ + 511, + 500, + 906, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Motivation: Tail-versus-noise trade-off", + "text_level": 1, + "bbox": [ + 511, + 619, + 849, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An ideal model to handle a long-tail noisy dataset should selectively sample the few-shot samples while removing noisy samples. However, we observe that the existing models suffer from a trade-off between class-imbalance and noise robustness. For memory-based noise-discriminative models in particular (e.g., SoftPatch), the methods focus on sampling the majority patch features, which often ignores less dominating but important patches.", + "bbox": [ + 511, + 642, + 905, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2 left, the noise discrimination process removes both defect and few-shot class patches, losing most of the information in the tail distribution. On the other hand, PatchCore samples few-shot samples well due to greedy coreset sampling which captures maximally different features. This sampling principle, however, favors the noise patch features as well as shown in Fig. 2 (right) and the low performance on noisy head classes in Fig. 1. Previous methods, hence, face the tail-versus-noise dilemma described above.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25541", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To resolve this, we propose to handle the tail classes independently from the head classes that potentially contain defective local regions. This, however, needs exclusive sampling of tail classes. To achieve this, we use the globally average pooled embedding vector, which is invariant against local pixel variation and rather exclusively corresponds to the object-centric nature of image, i.e., class.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Method", + "text_level": 1, + "bbox": [ + 89, + 210, + 178, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to build a coreset-based anomaly detection model using patch features. We require its memory to aptly capture tail class information and also to be absent of noise patches as much as possible. Particularly, we augment the noise discriminated memory $M_{clean}$ with the tail class memory $M_{tail}$ , which contains the patches of few-shot class samples exclusively:", + "bbox": [ + 89, + 238, + 483, + 329 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\text {T a i l e d C o r e}} = M _ {\\text {c l e a n}} \\cup M _ {\\text {t a i l}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 342, + 483, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The main challenge lies on sampling the tail class patches exclusively. To this end, we devise a few-shot sampler by estimating the class size of samples based on its embedding (e.g., global average pooling layer of the encoder). Using the estimated class sizes of training samples, the few-shot sampler enables us to exclusively sample tail class samples and the patches thereof.", + "bbox": [ + 89, + 371, + 483, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Few-shot sampler: TailSampler", + "text_level": 1, + "bbox": [ + 89, + 472, + 372, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TailSampler first estimates the class size of every training sample. Then, by automatically determining the maximum number of tail classes, denoted by $y_{\\mathrm{max}}$ , TailSampler collects the tail class samples' patch features exclusively.", + "bbox": [ + 89, + 494, + 483, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1.1.Class size predictor", + "text_level": 1, + "bbox": [ + 89, + 565, + 272, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To sample patches exclusively from few-shot class, we identify samples from few-shot classes by estimating their \"class sizes\", namely the size of class each sample belongs to. To accurately measure the class size, we hypothesize as follows: In the embedding space, samples within an appropriate angle are majorly of the same class, and this angle is likely to be the half of maximum angle to another embedding. In addition, the neighborhoods of neighborhoods within the half angle give accurate class size estimate. Based on these hypotheses, we proceed as follows:", + "bbox": [ + 89, + 584, + 483, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we adaptively set angle $\\alpha_{i}$ for each embedding $e_i$ from training samples $x_{i}$ . The coverage of angle $\\alpha_{i}$ decides the neighbors where a large value would assign many neighbors. We set the angle to be half of its maximum within the train set, and hypothesize that most of neighbors within the region is likely of the same class. Particularly, let $H_{i} = \\{e\\in Z:\\angle (e_{i},e)\\leq m_{i} / 2\\}$ denote the half-angle region where the maximum angle is $m_{i}\\coloneqq \\max_{e\\in Z}\\angle (e_{i},e)$ and the set $Z$ contains all train embeddings. Then, the adaptive angle $\\alpha_{i}$ is defined to contain $p$ -th percentile of the half-angle region:", + "bbox": [ + 89, + 720, + 483, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\angle \\left(e _ {i}, e _ {\\left(p \\cdot | H _ {i} |\\right)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 885, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the index $(j)$ of $e_{(j)}$ is sorted in the increasing order of the angle $\\angle (e_i,e_{(j)})$ . We set $p = 0.85$ in all experiments unless specified otherwise.", + "bbox": [ + 511, + 90, + 905, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After setting the adaptive angle $\\alpha_{i}$ for each train embedding $e_i$ , we estimate its class size based on neighborhoods of neighborhoods. Let $N_{\\alpha}(e_i) = \\{e \\in Z : \\angle(e_i, e) \\leq \\alpha\\}$ denote the neighborhood of $e_i$ , which is the set of all train embedding $e$ within the angle $\\alpha$ of $e_i$ . Then, the class size is estimated by the mode of the sizes of neighborhoods of the neighborhood:", + "bbox": [ + 511, + 136, + 906, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa_ {i} = \\underset {e \\in N _ {\\alpha_ {i}} (e _ {i})} {\\operatorname {m o d e}} \\left(| N _ {\\alpha (e)} (e) |\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 239, + 906, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha(e)$ is the adaptive angle with respect to the embedding $e$ belonging to the neighborhood $N_{\\alpha_i}(e_i)$ of embedding $e_i$ . Using the neighborhoods of neighborhood and mode on it gives more robust estimation of the class size than mere count of the direct neighborhood of $e_i$ .", + "bbox": [ + 511, + 277, + 906, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposition in Supp. 9 shows that the angle of neighborhood given by Eq. (6) corresponds to the decision boundary that maximizes the inter-class separation under regularities.", + "bbox": [ + 511, + 354, + 906, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1.2. Estimation of maximum size of tail class", + "text_level": 1, + "bbox": [ + 511, + 407, + 839, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After estimating the class size of each training sample by $\\kappa_{i}$ , we determine the maximum size for the tail classes based on the elbow technique [36]. In particular, by utilizing the sample-wise class sizes $\\kappa_{i}$ , we first obtain rough estimation of the size of each class set $\\eta_{y} \\approx |C_{y}|$ . Then, we estimate the maximum size of tail classes $K_{max}$ by finding a point where the class size most abruptly changes. This optimal shift point is found by the elbow technique. Concretely, we first estimate the size of each class $y$ , $\\eta_{y} \\approx |C_{y}|$ , inductively by", + "bbox": [ + 511, + 426, + 906, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\eta_ {1} = \\operatorname {r o u n d} \\left(\\frac {1}{\\kappa_ {(1)}} \\sum_ {i = 1} ^ {\\kappa_ {(1)}} \\kappa_ {(i)}\\right), \\\\ \\eta_ {y + 1} = \\operatorname {r o u n d} \\left(\\frac {1}{\\kappa_ {\\left(\\eta_ {y} + 1\\right)}} \\sum_ {i = \\eta_ {y} + 1} ^ {\\min \\left(\\kappa_ {\\left(\\eta_ {y} + 1\\right)}, | X |\\right)} \\kappa_ {(i)}\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 575, + 906, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the index $(i)$ of $\\kappa_{(i)}$ is sorted in the increasing order $\\kappa_{(1)}\\leq \\dots \\leq \\kappa_{(|X|)}$ .", + "bbox": [ + 511, + 678, + 905, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After we acquire the set $\\{(y,\\eta_y):y\\in (1,\\ldots ,|\\mathcal{Y}|)\\}$ through the inductive process above, we define the maximum size of tail classes as the elbow point of the estimated sizes of classes $\\eta_{y}$ since at this point the class size increases abruptly. The elbow is computed by linearly connecting $(1,\\eta_{1})$ and $(|\\mathcal{V}|,\\eta_{|\\mathcal{V}|})$ , and finding the maximum orthogonal distance onto the line from the set of points $\\{(y,\\eta_y):y\\in (1,\\dots |\\mathcal{Y}|)\\}$ as shown below:", + "bbox": [ + 511, + 710, + 905, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\max } := \\eta_ {y _ {\\max }} =: e l b o w (\\eta_ {1}, \\dots , \\eta_ {| \\mathcal {Y} |}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 828, + 906, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 857, + 557, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {\\max } = \\underset {y \\in \\{1, \\dots , | \\mathcal {Y} | \\}} {\\operatorname {a r g m a x}} \\frac {| a y - \\eta_ {y} + b |}{\\sqrt {a ^ {2} + 1}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 869, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25542", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ba9e9fcdd99af9e088ed6ee4234468f17bb787d7ab9350bc27caf2a2cdf5cdbc.jpg", + "image_caption": [ + "Figure 3. (a) Sampling process description of TailedCore (ours) and (b) the illustration of how we use the elbow method [36]." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and where the line $\\eta_y = ay + b$ intersects $(1, \\eta_1)$ and $(|\\mathcal{Y}|, \\eta_{|\\mathcal{Y}|})$ with $a = (\\eta_{|\\mathcal{Y}|} - \\eta_1) / (|\\mathcal{Y}| - 1)$ and $b = \\eta_1 - m$ . Fig. 3 (b) illustrates Eq. (10). We choose the elbow point $(y_{max}, \\eta_{y_{max}})$ as the maximum size of tail classes $K_{max}$ .", + "bbox": [ + 89, + 285, + 482, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The maximum size of tail classes $K_{max}$ allows us to sample the patch features of tail classes exclusively:", + "bbox": [ + 89, + 345, + 483, + 376 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {\\text {t a i l}} (P) = \\left\\{\\phi_ {i} ^ {(h, w)}: \\kappa_ {i} \\leq K _ {\\max } \\right\\}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 383, + 483, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\kappa_{i}$ given by Eq. (7).", + "bbox": [ + 89, + 411, + 248, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Proposed method: TailedCore", + "text_level": 1, + "bbox": [ + 89, + 433, + 359, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our anomaly detection model is built upon the memory bank constructed as below. First, we sample the base memory bank as in [20] with noise-discrimination described in Eq. (3). Then, the base memory is augmented with TailSampler $S_{tail}$ (Eq. (11)). Overall, TailedCore's memory bank is defined by $M_{TailedCore} = S_{core}(S_{clean}(P)) \\cup S_{core}(S_{tail}(P))$ , where $S_{core}$ is the greedy coreset sampling in Eq. (1), $S_{clean}$ is the noise discrimination process that filters out defect patches based on their outlier scores (Eq. (3)), and the few-shot sampler $S_{tail}$ augments the tail class patches exclusively on the noise-discriminated base memory bank. The whole process is shown in Fig 3. As TailedCore's patch memory bank is both noise-clean and contains the representative features of few-shot classes, it is robust against both class imbalance and normal sample contamination. In inference, the anomaly score is obtained by Eq. (2) as in the conventional memory-based detector.", + "bbox": [ + 88, + 455, + 483, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 709, + 217, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Section 6 consists of dataset description, comparison to the SOTA models, and the detailed ablation study and analysis of TailedCore's components.", + "bbox": [ + 89, + 734, + 483, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Datasets for unsupervised long-tail noisy anomaly detection", + "text_level": 1, + "bbox": [ + 89, + 787, + 482, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We devise the long-tail noisy anomaly detection datasets by modifying the class-wise sample distributions of the widely used anomaly detection datasets MVTecAD [3] and VisA [44]. We consider three different types of long-tail distributions. Particularly, datasets with step-shaped tail distribution are devised", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3ef8b5b6e645be1cf73068d3c97c3e44f76327e51d299c8f692a1e3d7782a9c2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDim [9] ICPR2182.4580.9582.0677.4781.2879.1971.5481.7575.63
HVQ [26] NeurIPS2383.4680.2382.9982.0185.5083.5674.1590.1580.55
WinCLIP [19] CVPR2389.3590.1190.3791.6088.2190.3791.8088.2390.37
AnomalyCLIP [43] ICLR2490.9390.9891.4891.8290.8391.4891.2191.9091.48
PatchCore [34] CVPR2293.3387.5989.1892.1971.1883.8386.3670.4880.01
SoftPatch [20] NeurIPS2284.6886.9587.7167.6597.5479.6460.6697.4975.40
TailedCore (ours)96.5595.2496.1295.8295.3495.7193.5495.7794.43
", + "bbox": [ + 514, + 282, + 905, + 382 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/031aa3ae5552f9cf33f916efdbc963e8591a850759a7fdac9be9e1adc2d2d8dc.jpg", + "table_caption": [ + "Table 1. Anomaly classification on MVTecAD with image-level AUROC $(\\%)$ . We report the mean over 5 random seeds for each measurement. Notations: $C_h / C_t$ : head / tail classes." + ], + "table_footnote": [], + "table_body": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2170.7083.3578.6460.6588.9372.4355.9886.7568.80
HVQ [26]NeuRPS2373.4784.0368.2568.2589.3077.0261.5780.4069.42
WinCLIP [19]CVPR2373.2576.9275.4775.9874.7675.4778.8070.8075.47
AnomalyCLIP [43]ICLR2481.9682.4882.0582.2881.7482.0583.2680.3482.05
PatchCore [34]CVPR2286.1185.7385.5983.5367.5176.8579.3368.5674.84
SoftPatch [20]NeuRPS2278.0492.1686.5659.7095.9774.8152.6194.1769.92
TailedCore (ours)87.5593.0690.8585.1695.9189.6482.9794.1187.61
", + "bbox": [ + 516, + 433, + 903, + 532 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7821c6f9d7eec1b9fdaff970d5a291d4f8344ef529962f2d7a20df4b45535157.jpg", + "table_caption": [ + "Table 2. Anomaly classification on VisA with image-level AUROC $(\\%)$ . The format and evaluation protocol are the same as Tab. 1." + ], + "table_footnote": [], + "table_body": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2190.1192.6691.4382.5395.2987.6778.8095.5485.50
HVQ [26]NeuRPS2393.6386.8590.5590.7392.5891.5386.3695.2089.90
WinCLIP [19]CVPR2382.0384.0682.2980.6084.6382.2980.1685.4882.29
AnomalyCLIP [43]ICLR2491.2491.6991.0889.9692.6691.0889.3493.6891.08
PatchCore [34]CVPR2293.5687.9889.9393.5472.0985.1992.0271.3583.75
SoftPatch [20]NeuRPS2292.1993.8393.4180.9896.4987.2470.3496.8980.99
TailedCore (ours)96.0895.0195.2995.5693.2094.7494.1993.7093.99
", + "bbox": [ + 516, + 570, + 903, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Anomaly segmentation on MVTecAD with pixel-level AUROC (\\%). We report the mean over 5 random seeds for each measurement. Notations: $C_h / C_t$ : head / tail classes.", + "bbox": [ + 511, + 670, + 906, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by making the particular classes to have only a small number of samples; namely $K = 1$ , and $K = 4$ , imposing extreme class imbalance on the datasets. We name these two datasets by 'step $(K = 1)$ ' and 'step $(K = 4)$ '. On the other hand, another realistic type of tail-distribution is considered by making the class distribution to follow the Pareto distribution with the shape parameter 0.6. For the step-tailed distribution datasets, we impose $40\\%$ of classes to be head classes, containing the original number of training samples. Thus, in MVTecAD, 6 of 15 classes are head classes. For the Pareto tailed distribution, we regard as few-shot (i.e., tail classes) the classes whose number of samples", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25543", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fad6ae5166c7874704a10d07e64ce5af9dea7949adb2e5ed7989a4dee7c8083f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2189.0295.1082.8183.9097.3689.5182.5796.5788.40
HVQ [26]NeuIPS2395.2797.6096.7193.8898.3495.7490.5895.5192.63
WinCLIP [19]CVPR2371.9473.9773.1974.6071.2173.1973.8172.3273.19
AnomalyCLIP [43]ICLR2495.6095.4695.5195.5495.4895.5196.1694.6095.51
PatchCore [34]CVPR2296.8487.9991.1395.3962.9681.8894.1165.3082.10
SoftPatch [20]NeuIPS2293.2096.7495.2783.9597.1089.4380.7396.8287.43
TailedCore (ours)97.9897.2597.4896.8097.0296.8996.1297.3996.65
", + "bbox": [ + 94, + 88, + 480, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Anomaly segmentation on VisA with pixel-level AUROC $(\\%)$ . The format and evaluation protocol are the same as Tab. 3.", + "bbox": [ + 89, + 188, + 482, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is less than 20 [25]. To impose noisy condition on the datasets, the anomaly samples (i.e., the samples with defective regions) are added to the training samples such that the noisy samples constitute $10\\%$ of the training set following the overlap protocol in SoftPatch [20] for fair comparison. In all cases, the testing sets of datasets remain the same as the original MVTecAD and VisA. A more detailed configuration is given in Supp. 11.", + "bbox": [ + 89, + 228, + 483, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Comparison to the state-of-the-art (SOTA)", + "text_level": 1, + "bbox": [ + 89, + 344, + 452, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare TailedCore to the SOTA anomaly detection methods in the unsupervised long-tail noisy settings. We carefully select our baselines. HVQ excels in multi-class learning, PaDiM and SoftPatch are robust against noises as PaDiM is based on Gaussian modeling of the data, while SoftPatch explicitly removes the defect patches by outlier scoring. Using foundation models, WinClip and AnoamlyClip are strong for zero-shot learning. PatchCore are near-perfect in the single-class anomaly detection settings. We show the full experimental results in Supp. 12. All models are trained in the multi-class setting; namely, one model for one dataset with all classes rather than one model per class. The detection performance, however, is measured for each product class under image-level and pixel-level AUROC, and averaged for reporting and comparison. For comparison, we consider models that are specifically robust against noise, strong on multi-class training, and foundation models that can exhibit zero/few-shot learning capability.", + "bbox": [ + 88, + 366, + 482, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tabs. 1, 2, 3, and 4 show the superiority of TailedCore across all types of long-tail distributions. As anticipated, all baseline models encounter the tail-versus-noise trade-off, resulting in suboptimal performance either on tail classes with limited samples or on head classes with noisy training samples. These trends become more pronounced in highly unbalanced datasets, where the few-shot classes contain only 1 or 4 samples.", + "bbox": [ + 89, + 623, + 483, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Ablation study and analysis", + "text_level": 1, + "bbox": [ + 89, + 738, + 339, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3.1. Tail (few-shot) sampling", + "text_level": 1, + "bbox": [ + 89, + 761, + 305, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate how accurately our proposed algorithm $S_{tail}$ can sample tail (few-shot) class samples exclusively without including noise samples. First of all, we measure the ratio of mis-included tail class samples TailSampler which is the ratio of not sampled tail class samples to total tail class samples. A single mis-inclusion can vitally damage anomaly detection performance especially in case of the one-shot setting $K = 1$ as the model wouldn't be able to utilize the only sample of the class.", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, the exclusive aspect of few-shot sampling is crucial; the few-shot sampled set shouldn't include anomaly and head class samples as possible. Included anomaly samples would harm the head class performance and head class sampled by few-shot sampler might cause class imbalance issue as well as the possibility of sampling anomaly from head classes. Hence, we measure the ratios of included anomalies and head-class samples along with AUROC which is measured on the outlier scores derived by a given algorithm and ground truth binary labels that indicate whether a given sample is tail class or head class. For TailSampler, the outlier score of each train sample is given as the negative of predicted class size. The included anomaly ratio is defined as the ratio of number of anomaly samples to the number of whole anomaly samples. The ratio of included head-class sample is defined as the number of included head class to the number of included tail class samples.", + "bbox": [ + 511, + 90, + 906, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For extensive comparison, we compare TailSampler with widely used clustering algorithms [2, 13, 16], outlier detectors [1, 5, 24, 35], and density estimators [10, 30]. These methods provide clustering size and/or outlier score/prediction. In cases when only outlier scores are available, we apply the elbow method to induce deterministic prediction of whether a given sample is few-shot or not. We use default settings provided by Sklearn [31] and PyOD [42]. For all methods, we utilize the embedding vectors of WideResNet-50 [40]. Experiments with different architectures and details are provided in Supp. 10.", + "bbox": [ + 511, + 332, + 905, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The combined ratio, which is the sum of missing tail samples and included anomaly samples, indicates that TailSampler is most effective for exclusively sampling few-shot class samples. The most critical factor is to not miss few-shot class samples. On this ratio, TailSampler achieves the lowest rate except for the algorithms that trivially sample all few-shot class samples. We find that most of the classical algorithms for clustering and outlier detection are not usable as they either miss too many few-shot samples or includes a large number of anomaly samples. We note that the class size predictor in our few-shot samples resembles DBSCAN and Affinity Propagation, respectively, in terms of using neighborhoods and voting mechanism. None of these methods however have instance-specific parameters that are adaptively adjusted across different samples.", + "bbox": [ + 511, + 484, + 908, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3.2. Class size prediction", + "text_level": 1, + "bbox": [ + 511, + 700, + 702, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We measure how accurately our class size predictor estimates the class size of each sample by measuring the class size prediction error. It is measured for each train sample and weight-averaged through the whole train samples:", + "bbox": [ + 511, + 720, + 905, + 781 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ne r r o r = \\frac {1}{| \\mathcal {X} |} w _ {i} \\sum_ {i = 1} ^ {| \\mathcal {X} |} \\left| \\kappa_ {i} - \\left| C _ {k (i)} \\right| \\right| \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 787, + 906, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $k(i)$ indicates the index such that $x_{i} \\in C_{k(i)}$ (namely, $C_{k(i)}$ is the class of $x_{i}$ ). The weight is chosen by $w_{i} = 1 / |C_{k(i)}|$ , which is to give more penalty on the few-shot class samples more than the head class samples.", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25544", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/152311089b96de8555f960d6ca9fed416477e5d0357d3e8fc144abc2b7317b43.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 282, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e729b819c2118e56ba16c1a0a1940bd4dfc07c914c7583bb628dfb4aae8f6df4.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 93, + 214, + 281, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebb0db69dd702d850c868a7a4c91b7993907b663f3bdd5a94387380c61cb7050.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 93, + 339, + 281, + 445 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/411a6a35c177bd2d5d4c88cf36b7b0b5e5b712fa13ab9f0d988198bcbf586c88.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 290, + 89, + 478, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b774650a828046de9a5043b7a620221b1bae70aa92a70778ba9170a06713eb24.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 292, + 214, + 478, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48d33364309a3fa03f7362eb1046fb4206ff6652013af25084b37575b0dff952.jpg", + "image_caption": [ + "(f)", + "Figure 4. The plot of noise ratio of train data versus anomaly classification (image-level AUROC) and segmentation (pixel-level AUROC) performance. In all cases, TailedCore outperforms both PatchCore and SoftPatch. Dataset is MVTecAD with different tail distributions. The lines show the mean and shades show standard deviation of multiple runs of experiments with different seeds." + ], + "image_footnote": [], + "bbox": [ + 292, + 339, + 478, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 5 shows that most of the clustering based methods vastly fail on this task, and the estimation by our class size predictor outperforms them significantly. We believe this is partly due to suboptimal hyperparameters used for these algorithms. Due to their classical nature, however, it is unclear how to tune them without the availability of the validation set in our scenario.", + "bbox": [ + 89, + 564, + 483, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3.3. Abalation study on class size predictor", + "text_level": 1, + "bbox": [ + 89, + 667, + 405, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct detailed ablation study on the proposed class size prediction in Eq. (7). One hyperparameter used in TailSampler $S_{tail}$ of TailedCore is the percentiling proportion $p$ to consider only the majority of nearby samples within the neighborhood. We analyze how this parameter can affect the anomaly detection performance of TailedCore. Tab. 5 indicates that using the nearest $85\\%$ embeddings (i.e., $p = 85\\%$ ) is the best choice. Using all of the available neighborhoods (i.e., $p = 100\\%$ ) would include different class samples in the neighborhood, deteriorating class estimation. On the other hand, counting only with the half of samples (i.e., $p = 50\\%$ ) in the neighborhood may lose too much information.", + "bbox": [ + 89, + 686, + 482, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another component in the class size predictor that requires analysis is its voting mechanism used in Eq. (7). One may", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d6d38b54e36654772fdedf91eef296db15e3a89e07c007efadba19196acc91d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodcombined ratio of missing \\( {C}_{t} \\) & included AS (↓)ratio of missing \\( {C}_{t} \\) samples (↓)ratio of included AS (↓)ratio of included \\( {C}_{h} \\) samples (↓)class size prediction error (↓)AUROC (%) (↑)
LOF\\( {89.58} \\pm {44.42} \\)\\( {31.22} \\pm {34.88} \\)\\( {58.36} \\pm {9.54} \\)\\( {92.50} \\pm {7.87} \\)\\( {1.00} \\pm {0.00} \\)\\( {77.08} \\pm {25.86} \\)
IF\\( {18.15} \\pm {15.91} \\)\\( {7.25} \\pm {10.05} \\)\\( {10.90} \\pm {5.86} \\)\\( {69.02} \\pm {25.96} \\)\\( {1.00} \\pm {0.00} \\)\\( {98.96} \\pm {1.27} \\)
OCSVM\\( {49.95} \\pm {9.76} \\)\\( {0.78} \\pm {2.19} \\)\\( {49.17} \\pm {7.57} \\)\\( {95.06} \\pm {6.08} \\)\\( {1.00} \\pm {0.00} \\)\\( {96.72} \\pm {3.59} \\)
DBSCAN\\( {35.36} \\pm {49.55} \\)\\( {30.64} \\pm {44.38} \\)\\( {4.72} \\pm {5.17} \\)\\( {25.57} \\pm {27.24} \\)\\( {0.35} \\pm {0.44} \\)\\( {84.46} \\pm {22.18} \\)
KMeans\\( {94.67} \\pm {44.95} \\)\\( {47.22} \\pm {35.49} \\)\\( {47.45} \\pm {9.46} \\)\\( {97.40} \\pm {3.99} \\)\\( {4.27} \\pm {1.40} \\)\\( {52.58} \\pm {20.46} \\)
GMM\\( {100.00} \\pm {0.00} \\)\\( \\mathbf{{0.00}} \\pm \\mathbf{{0.00}} \\)\\( {100.00} \\pm {0.00} \\)\\( {97.58} \\pm {2.86} \\)\\( {12.50} \\pm {1.53} \\)\\( {50.00} \\pm {0.00} \\)
KDE\\( {29.81} \\pm {24.15} \\)\\( {2.14} \\pm {6.26} \\)\\( {27.67} \\pm {17.89} \\)\\( {87.47} \\pm {10.35} \\)\\( {1.00} \\pm {0.00} \\)\\( {97.98} \\pm {2.19} \\)
AP\\( {27.81} \\pm {32.59} \\)\\( {12.98} \\pm {24.01} \\)\\( {14.83} \\pm {8.58} \\)\\( {77.45} \\pm {28.21} \\)\\( {0.70} \\pm {0.22} \\)\\( {93.03} \\pm {12.8} \\)
KNN\\( {43.49} \\pm {31.25} \\)\\( {14.46} \\pm {24.07} \\)\\( {29.03} \\pm {7.18} \\)\\( {83.12} \\pm {18.28} \\)\\( {1.00} \\pm {0.00} \\)\\( {94.51} \\pm {10.34} \\)
TailSampler (ours)\\( {5.10} \\pm {5.07} \\)\\( {0.12} \\pm {0.65} \\)\\( {4.98} \\pm {4.42} \\)\\( {31.96} \\pm {26.05} \\)\\( \\mathbf{{0.19}} \\pm \\mathbf{{0.05}} \\)\\( {99.78} \\pm {0.30} \\)
Ablation on the class size predictor:
TailSampler (p=100)\\( {8.48} \\pm {10.88} \\)\\( {1.07} \\pm {5.23} \\)\\( {7.41} \\pm {5.65} \\)\\( {42.64} \\pm {31.86} \\)\\( {0.51} \\pm {0.04} \\)\\( {99.76} \\pm {0.29} \\)
TailSampler (p=50)\\( {7.37} \\pm {10.35} \\)\\( {3.46} \\pm {6.80} \\)\\( {3.91} \\pm {3.55} \\)\\( {27.90} \\pm {27.36} \\)\\( {0.12} \\pm {0.12} \\)\\( {97.36} \\pm {5.65} \\)
TailSampler (top-1)\\( {14.47} \\pm {7.99} \\)\\( {0.20} \\pm {0.78} \\)\\( {14.27} \\pm {7.21} \\)\\( {73.97} \\pm {27.95} \\)\\( {0.85} \\pm {0.04} \\)\\( {99.73} \\pm {0.34} \\)
TailSampler (average)\\( {6.48} \\pm {10.11} \\)\\( {3.65} \\pm {6.99} \\)\\( {2.83} \\pm {3.12} \\)\\( {22.16} \\pm {24.33} \\)\\( {0.21} \\pm {0.07} \\)\\( {99.35} \\pm {2.59} \\)
", + "bbox": [ + 516, + 89, + 903, + 232 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4de34cd9027a14185d4cc02a9c0faa460224929897c1654e7cdc6853f105d1e0.jpg", + "table_caption": [ + "Table 5. The analysis on the proposed TailSampler and its class size predictor. The results are obtained from both MVTecAD and VisA and averaged across all different types of tail distributions and random seeds (overall 60 different seeds) with the standard deviations indicated on the right. The best and second best performances are marked in bold and underline, respectively. Acronyms/notations: IF: Isolation Forest; AP: Affinity Propagation; AS: anomaly samples; $C_h / C_t$ : head / tail classes." + ], + "table_footnote": [], + "table_body": "
tasktail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
anomaly classificationw/o FSA82.2291.2288.1360.0197.6075.0566.2897.8578.94
w/ FSA96.5595.2496.1393.5495.7894.4495.8295.3495.72
anomaly segmentationw/o FSA91.6094.9493.5468.9996.8580.1479.3996.3586.23
w/ FSA96.0995.0195.3094.1993.7194.0095.5693.2094.75
", + "bbox": [ + 514, + 368, + 906, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Abaltion on the few-shot augmentation (FSA) of TailedCore. The image-level and pixel-level AUROC metrics $(\\%)$ are reported on the MVTecAD dataset with different types of tail distributions for anomaly classification and segmentation tasks, respectively. Notations: $C_h / C_t$ : head / tail classes.", + "bbox": [ + 511, + 448, + 906, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "simply use the nearest embedding's neighborhood size instead, or the class size of $x_{i}$ can be estimated by averaging instead of majority vote. As indicated in Tab. 5, both the nearest embedding (i.e., top-1) and averaging are inferior to majority voting. The former is inevitably less robust, while averaging can still be sensitive to the outlier values of neighborhood size.", + "bbox": [ + 511, + 547, + 906, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3.4. Ablation on TailedCore", + "text_level": 1, + "bbox": [ + 511, + 664, + 723, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the impact of augmentation of few-shot class patches by TailSampler in TailedCore. The TailedCore model without these augmented patches has the same memory bank as SoftPatch, but its inference scoring is simpler. Tab. 6 indicates that without the augmented patches from TailSampler, the memory-based model cannot detect anomalies in few-shot class samples. This shows the necessity of tail class augmentation by TailSampler. We note some performance decrease in the head class detection performance since few-shot augmentation may involve few anomaly samples from the head classes that behave like few-shot class instances. This issue intrinsically is due to the weakness of encoder embedding, and the proposed TailSampler based on such embeddings can induce incorrectly sampled few-shot class instances, which we discuss in Sec. 7.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25545", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9717bda343f63692eaf05562fe64c0ffcd246b969c9c5565d6bb41162b2cfb4e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 285, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce20f1f5e2917667e94b2ca569c2c2bd8cae69cde8903b1a9d60b5645b29507b.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 300, + 90, + 488, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/08a3066e9f32c6a13df8d9ae59142a883f26650743126ca83c2a02d0c0ba89a4.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 694, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2acdf6056a1babed87312a360c29e6b8814b976420236aa54344350c4396e70.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 712, + 90, + 898, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d27c9a2cc6be928f5abb9e05bb2f50a32ab6c24ff2e678ad5e3d94b012f00912.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 94, + 219, + 284, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4883daa739be54c80bc4e97a87f1d4b85029f91b9b8ba1be30c94f110de2885.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 300, + 219, + 488, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcb0e23af33ba75dc74d66826c9a39155bcd6a5b8a0faf7871478dd736a261bf.jpg", + "image_caption": [ + "(g)", + "Figure 5. Classification accuracy between tail classes and noisy samples versus metrics relevant to class size prediction and few-shot sampling on MVtecAD and VisA datasets with step-like tailed distribution and $K = 4$ . The correlation is strong for (a) mis-sampling ratio, (b) ratio of missing few-shot samples, (e) class size prediction error, and (f) AUROC for predicting whether a sample is few-shot or not. We show that improving the discriminative aspect of network's embeddings improves few-shot sampling and prediction error of class size predictor, which in turn improves (g) anomaly classification (image-level AUROC) and (h) anomaly segmentation (pixel-level AUROC) performance of TailedCore." + ], + "image_footnote": [], + "bbox": [ + 506, + 219, + 694, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ad25f98f81dfe8375ee1b7274baf1e43b4a0313173ca8c87b8ecc06af7045fc.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 710, + 219, + 898, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Noise ratio", + "text_level": 1, + "bbox": [ + 89, + 431, + 210, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the impact of noise ratio. Training data with a greater number of defect samples increase the chance of sampling noise instances for the few-shot sampler. We analyze this trend in Fig. 4, but the impact is not severe. In all cases, TailedCore either outperforms or is on par with both PatchCore and SoftPatch.", + "bbox": [ + 89, + 455, + 483, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Relation between the discriminative aspect of embedding and TailedCore", + "text_level": 1, + "bbox": [ + 89, + 542, + 483, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TailedCore utilizes the embedding features $e_i$ of training samples $x_i$ for few-shot sampling and augmentation thereof. Therefore, TailedCore's detection performance highly relies on the discriminative aspect of the embedding feature. Poorly designed embedding features will not discriminate different class samples in the embedding space. In this case, TailSampler cannot sample tail class samples and it may overly many head class samples. TailedCore's memory bank obtained so will neither be class-balanced nor noise-free, making the inference of TailedCore poor.", + "bbox": [ + 89, + 580, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, a highly discriminative embedding features will make TailedCore's class size prediction accurate, thereby enabling it to sample the few-shot class samples exclusively without aggregating any anomaly and head class samples. In which case, TailedCore is both noise-free and class-balanced.", + "bbox": [ + 89, + 733, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We test this behavior by Oracle labeling and training the embedding features with class labels. Fig. 5 shows that improving discriminative quality of embedding (measured by its classification accuracy on noise and tail class samples) accordingly improves the relevant metrics. Particularly, more discriminative embeddings make more accurate class size", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "prediction, improving TailSampler along with its anomaly classification and segmentation performances.", + "bbox": [ + 511, + 433, + 906, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitations", + "text_level": 1, + "bbox": [ + 511, + 479, + 629, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The class size prediction in TailSampler can fail if the reflective-symmetric assumption on the inter-class and intra-class similarities breaks down. Particularly, this case can occur if the utilized encoder embedding is either poor or not aligned with the label space of classes. In some cases, the geometric aspect of defect samples highly resembles that of few-shot class instances in the embedding space. We believe that this is due to imperfect nature of encoder embedding, which is often more sensitive to spurious features than the notion of class and object-centric aspect [29]. This issue can be improved by enhancing the encoder network.", + "bbox": [ + 511, + 505, + 908, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 672, + 627, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we introduced TailedCore, a novel memory-based anomaly detection model designed to effectively address the challenges of unsupervised anomaly detection in noisy and long-tail class distributions. By developing a unique class size predictor and a tailored memory bank, we successfully navigated the tail-versus-noise dilemma, enhancing the model's performance in identifying anomalies within a contaminated and imbalanced dataset. Our comprehensive evaluations demonstrated TailedCore's superior ability to mitigate noise contamination and class imbalance, underscoring its significance in advancing anomaly detection technologies. This work not only showcases a significant stride in handling real-world anomaly detection scenarios but also lays a robust foundation for future research in this critical field.", + "bbox": [ + 511, + 699, + 908, + 910 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25546", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments Kuan-Chuan Peng was exclusively supported by Mitsubishi Electric Research Laboratories. This work was also supported by the Technology Innovation Program (20023364) funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea), NSF grant 2038493, ONR grant N00014-21-1-2431, NIH grant R01CA240771 from NCI, and U.S. Department of Homeland Security grant 22STESE00001-03-02. The views and conclusions contained in this document are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the U.S. Department of Homeland Security.", + "bbox": [ + 89, + 90, + 485, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 270, + 183, + 286 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fabrizio Angiulli and Clara Pizzuti. Fast outlier detection in high dimensional spaces. In European conference on principles of data mining and knowledge discovery, pages 15-27. Springer, 2002. 3, 6", + "[2] David Arthur, Sergei Vassilvitskii, et al. k-means++: The advantages of careful seeding. In Soda, pages 1027-1035, 2007. 2, 6", + "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 5", + "[4] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4):947-969, 2022. 3", + "[5] Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and Jörg Sander. Lof: identifying density-based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pages 93-104, 2000. 3, 6", + "[6] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1", + "[7] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 2", + "[8] Zhuohang Chen, Jinglong Chen, Yong Feng, Shen Liu, Tianci Zhang, Kaiyu Zhang, and Wenrong Xiao. Imbalance fault diagnosis under long-tailed distribution: Challenges, solutions and prospects. Knowledge-Based Systems, 258:110008, 2022. 1", + "[9] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 2, 5, 6", + "[10] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 6", + "[11] Jan Diers and Christian Pigorsch. A survey of methods for automated quality control based on images. International Journal of Computer Vision, pages 1-29, 2023. 1" + ], + "bbox": [ + 93, + 295, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[13] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 2, 6", + "[14] Chaowei Fang, Lechao Cheng, Huiyan Qi, and Dingwen Zhang. Combating noisy labels in long-tailed image classification. arXiv preprint arXiv:2209.00273, 2022. 2", + "[15] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 2", + "[16] Brendan J Frey and Delbert Dueck. Clustering by passing messages between data points. science, 315(5814):972-976, 2007. 2, 6", + "[17] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. arXiv preprint arXiv:2308.15366, 2023. 2", + "[18] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 2", + "[19] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero-/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2, 5, 6", + "[20] Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai Wu, Yong Liu, Chengjie Wang, and Feng Zheng. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35:15433-15445, 2022. 1, 2, 3, 5, 6", + "[21] Justin M Johnson and Taghi M Khoshgoftaar. Survey on deep learning with class imbalance. Journal of Big Data, 6(1):1-54, 2019. 1", + "[22] Yujin Lee, Harin Lim, Seoyoon Jang, and Hyunsoo Yoon. Uniformality: Towards task-agnostic unified framework for visual anomaly detection. Available at SSRN 4586132. 2", + "[23] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. arXiv preprint arXiv:2401.16753, 2024. 2", + "[24] Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 eighth ieee international conference on data mining, pages 413-422. IEEE, 2008. 3, 6", + "[25] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 1, 6", + "[26] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems, 36, 2023. 2, 5, 6" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25547", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yang Lu, Yiliang Zhang, Bo Han, Yiu-ming Cheung, and Hanzi Wang. Label-noise learning with intrinsically long-tailed data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1369-1378, 2023. 2", + "[28] Declan McIntosh and Alexandra Branzan Albu. Inter-realization channels: Unsupervised anomaly detection beyond one-class classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6285-6295, 2023. 2", + "[29] Yifei Ming, Hang Yin, and Yixuan Li. On the impact of spurious correlation for out-of-distribution detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10051-10059, 2022. 8", + "[30] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33 (3):1065-1076, 1962. 6", + "[31] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011. 6", + "[32] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[34] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 2, 3, 5, 6", + "[35] Bernhard Schölkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. Advances in neural information processing systems, 12, 1999. 3, 6", + "[36] Robert L Thorndike. Who belongs in the family? Psychometrika, 18(4):267-276, 1953. 4, 5", + "[37] Tong Wei, Jiang-Xin Shi, Wei-Wei Tu, and Yu-Feng Li. Robust long-tailed learning under label noise. arXiv preprint arXiv:2108.11569, 2021. 2", + "[38] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Feng Zheng, and Yaochu Jin. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. arXiv preprint arXiv:2301.12082, 2023. 2", + "[39] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 1, 2", + "[40] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6", + "[41] Manyi Zhang, Xuyang Zhao, Jun Yao, Chun Yuan, and Weiran Huang. When noisy labels meet long tail dilemmas: A representation calibration method. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15890-15900, 2023. 2" + ], + "bbox": [ + 89, + 92, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Yue Zhao, Zain Nasrullah, and Zheng Li. Pyod: A python toolbox for scalable outlier detection. Journal of Machine Learning Research, 20(96):1-7, 2019. 6", + "[43] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In International Conference on Learning Representations, 2024. 2, 5, 6", + "[44] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. arXiv preprint arXiv:2207.14315, 2022. 5" + ], + "bbox": [ + 514, + 92, + 903, + 246 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25548", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_model.json b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1e56363711bbf0b3b344ba53c3d79c08df0e9708 --- /dev/null +++ b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_model.json @@ -0,0 +1,2640 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.18, + 0.773, + 0.218 + ], + "angle": 0, + "content": "Yoon Gyo Jung\\*1 Jaewoo Park\\*2 Jaeho Yoon\\*3 Kuan-Chuan Peng4 Wonchul \\(\\mathrm{Kim}^2\\) Andrew Beng Jin Teoh Octavia Camps" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.226, + 0.741, + 0.273 + ], + "angle": 0, + "content": "\\(^{1}\\)Northeastern University \\(^{2}\\)AiV Co. \\(^{3}\\)Yonsei University \\(^{4}\\)Mitsubishi Electric Research Laboratories" + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.332, + 0.325, + 0.348 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.365, + 0.486, + 0.665 + ], + "angle": 0, + "content": "We aim to solve unsupervised anomaly detection in a practical challenging environment where the normal dataset is both contaminated with defective regions and its product class distribution is tailed but unknown. We observe that existing models suffer from tail-versus-noise trade-off where if a model is robust against pixel noise, then its performance deteriorates on tail class samples, and vice versa. To mitigate the issue, we handle the tail class and noise samples independently. To this end, we propose TailSampler, a novel class size predictor that estimates the class cardinality of samples based on a symmetric assumption on the class-wise distribution of embedding similarities. TailSampler can be utilized to sample the tail class samples exclusively, allowing to handle them separately. Based on these facets, we build a memory-based anomaly detection model TailedCore, whose memory both well captures tail class information and is noise-robust. We extensively validate the effectiveness of TailedCore on the unsupervised long-tail noisy anomaly detection setting, and show that TailedCore outperforms the state-of-the-art in most settings. Code is available in TailedCore" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.216, + 0.713 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.486, + 0.875 + ], + "angle": 0, + "content": "In the complex landscape of anomaly detection, the challenge often lies in navigating practical scenarios that feature a diverse and imbalanced data distribution [6, 8, 11, 21]. We delve into a challenging yet highly realistic scenario in anomaly detection, where the training dataset, comprising multiple product classes [39], is beset by two significant complications: contamination with noise [20] and presence of tailed class distributions [25]. In this context, 'tail' (few-shot) classes have limited few-shot samples, unlike 'head' (many-shot) classes which are data-rich but might contain samples with defects. Crucially, in this unsupervised set-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.318, + 0.707, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.446, + 0.583, + 0.453 + ], + "angle": 0, + "content": "SimpleNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.446, + 0.638, + 0.453 + ], + "angle": 0, + "content": "WinCLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.446, + 0.678, + 0.453 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.446, + 0.71, + 0.453 + ], + "angle": 0, + "content": "IVQ" + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.446, + 0.742, + 0.453 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.446, + 0.772, + 0.453 + ], + "angle": 0, + "content": "bftPatch" + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.446, + 0.805, + 0.453 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.446, + 0.836, + 0.453 + ], + "angle": 0, + "content": "。" + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.446, + 0.864, + 0.453 + ], + "angle": 0, + "content": "PaDiM" + }, + { + "type": "image_caption", + "bbox": [ + 0.866, + 0.446, + 0.891, + 0.453 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.465, + 0.908, + 0.521 + ], + "angle": 0, + "content": "Figure 1. Tail class (x-axis) and noisy head class (y-axis) performance comparison. The tail-versus-noise trade-off is shown across memory-based anomaly detection models (circles), and is more indicative in anomaly classification task evaluated by image-level AUROC (left)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.536, + 0.907, + 0.612 + ], + "angle": 0, + "content": "ting, the class information of these products remains unknown to the model trainer. This setting mirrors a common predicament in large-scale industrial applications, where companies are required to manage vast production lines encompassing a wide array of product types, each with varying production rates." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.909, + 0.809 + ], + "angle": 0, + "content": "A critical issue arises in this scenario, which we term the \"tail-versus-noise dilemma.\" The dilemma underscores a fundamental trade-off where models robust to class imbalance overfit to defect noises, likely misclassifying many defects as normal. In contrast, existing models that are robust against defect noise tend to underfit, missing vital details in tail classes. Fig. 1 shows this tail-versus-noise trade-off which is prevalent through existing anomaly detection algorithms. This intrinsic characteristic comes from the fact that few-shot class features are equivalent to the noisy abnormal features in terms of their statistical occurrence. Consequently, existing models find it difficult to strike a balance between avoiding overfitting to noise and effectively capturing the nuances of less represented classes." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To resolve this challenge, we propose to handle tail classes independently from head classes by exclusively sampling them. To this end, we devise a novel class size predictor TailSampler that estimates the cardinality of the class of any training sample. For accurate estimation, TailSampler estimates the class sizes based on a reflective symmetry between inter-class" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.887, + 0.313, + 0.901 + ], + "angle": 0, + "content": "* equal contribution † corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25539" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "and intra-class similarity distributions in the embedding space. The estimated class sizes of the training samples then enable to sample tail class features exclusively from the training set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.483, + 0.258 + ], + "angle": 0, + "content": "By utilizing the aggregated few-shot class instances, we devise a memory-based anomaly detector TailedCore, which is built upon the noise-discriminated model by augmenting the tail class patch features sampled with TailSampler. As the memory of TailedCore is both clean from noisy defect patches and preserves the few-shot class information, it exhibits robustness against both noise contamination and class imbalance of the training set." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.259, + 0.397, + 0.272 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.275, + 0.483, + 0.334 + ], + "angle": 0, + "content": "- We explore a practical and challenging anomaly detection scenario, where a single detection model is trained on multi-class product training samples that involve noise contamination and whose class distribution is long-tailed." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.336, + 0.483, + 0.47 + ], + "angle": 0, + "content": "- We propose a memory-based anomaly detector TailedCore whose patch memory bank is both noise-free and augmented with the representative patches of few-shot class instances. To build TailedCore, we introduce TailSampler that can aggregate few-shot class instances exclusively. TailSampler is based on a novel class size predictor that estimates the class size of training samples based on the reflective symmetry between inter-class and intra-class class-wise embedding similarity distributions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.471, + 0.483, + 0.531 + ], + "angle": 0, + "content": "- We conduct extensive analyses and comparative studies on both TailedCore and few-shot sampler along with the proposed class size predictor on the proposed unsupervised long-tail noisy anomaly detection benchmark." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.275, + 0.483, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.577 + ], + "angle": 0, + "content": "The codes for TailedCore along with the data generation will be open sourced. We use the term 'few-shot' and 'tail' interchangeably to denote the class with a small number of samples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.229, + 0.608 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.433, + 0.634 + ], + "angle": 0, + "content": "2.1. Anomaly detection in extreme scenarios" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.746 + ], + "angle": 0, + "content": "One vitally important aspect of anomaly detection [9, 34] is the environment where these models are tested. This exploration has led to the identification of several challenging scenarios, including multi-class settings with training data spanning multiple product types [26, 39], zero-shot and few-shot settings with minimal training images [15, 19, 23, 38, 43], and noisy settings featuring uncleaned training data with defective images [7, 20, 28]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Multi-class anomaly detection In contrast to single-product settings, anomaly detection models trained on multi-class datasets often encounter \"shortcut learning\" [39], leading to low anomaly scores for normal images. This issue is particularly pronounced in reconstruction-based models, which tend to reconstruct all samples, regardless of their normality or abnormality. To resolve it, UniAD [39] employs neighbor masked attention within a transformer model, focusing attention on local parts to prevent shortcut learning. HVQ [26] utilizes vector quantization to discretize the latent space, restricting the network's recon" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.198 + ], + "angle": 0, + "content": "striction fidelity on normal data distributions. UniFormaly [22] increases sensitivity to irrelevant patterns by masking less relevant patches and employing top-k matchings with a pretrained foundation model. These multi-class anomaly detection methods do not consider either long-tail classes or contaminated training data setups, and do not work well in the combined \"noisy long-tail\" setup, which is shown in our experimental results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.909, + 0.396 + ], + "angle": 0, + "content": "Zero/few-shot anomaly detection The inception of few-shot learning benchmarks in anomaly detection marked a significant advancement, with methods leveraging template matching [18] and graph representations [38] for minimal training data scenarios. AnomalyCLIP [43], WinCLIP [19], and AnomalyGPT [17], on the other hand, detect anomalies in the few-shot classes based on the acquired knowledge of pretrained foundation models such as CLIP [33] and GPT [32]. However, these methods often require sophisticated prompt engineering or delicate prompt-tuning, and have yet to achieve optimal precision and recall in the single-class training setting with a sufficient amount of training samples. In contrast, our proposed TailedCore needs none of these process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.909, + 0.535 + ], + "angle": 0, + "content": "Noise anomaly detection Dealing with training datasets contaminated with defective samples presents another extreme learning scenario. SoftPatch [20] and InReaCh [28] focus on purifying memory coresets through outlier detection algorithms and distance association filtering, respectively. IGD [7] adopts a robust Gaussian modeling approach, addressing both contaminated training data and small dataset challenges. Yet, these methods do not fully consider scenarios involving both data contamination and class imbalance, unlike our proposed TailedCore." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.734, + 0.56 + ], + "angle": 0, + "content": "2.2. Long-tail noisy learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.748 + ], + "angle": 0, + "content": "In the realm of image classification, several studies have focused on learning scenarios involving training sets with label contamination and class imbalance. To address these challenges, the existing literature relies on the robustness and unsupervised aspects of self-supervised learning [14, 27, 37, 41]. The learned embeddings of self-supervised frameworks are less impacted by label contamination and class imbalances. However, these works differ from ours in that they deal with label-level contamination rather than pixel-level defects. Furthermore, these methods require explicit supervised labels to remediate the contamination. In contrast, our work targets noise in pixels without any class label information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.75, + 0.774 + ], + "angle": 0, + "content": "2.3. Few-shot/outlier sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Both classical clustering methods and outlier detection algorithms can be applied to our task, as they facilitate the identification of few-shot tail class samples, either indirectly or directly. Unsupervised clustering algorithms [2, 13, 16] can estimate the cluster sizes but struggle to identify small sample clusters in the imbalanced data. DBSCAN [13], on the other hand, captures both clusters and outliers. This however can be problematic when attempting to identify clusters with only one" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25540" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.334 + ], + "angle": 0, + "content": "sample (1-shot) as it may be confused between few-shot samples and outliers. Unlike clustering methods, outlier detection algorithms [1, 5, 24, 35] are adept at identifying 1-shot samples which inherently resemble outliers. However, their effectiveness diminishes when a few-shot class contains more than one sample, as these no longer fit the typical outlier profile, potentially leading to incorrect identification of few-shot class samples. Both algorithm types, due to their inherent design, struggle to accurately predict class/cluster sizes in datasets that are both imbalanced and contaminated. Our proposed class size predictor specifically addresses this issue, aiming to reliably and accurately identify few-shot class samples and enable their exclusive sampling. Ultimately, the identified few-shot class samples are augmented to the noise-discriminated patch set, constituting the memory-based anomaly detection model TailedCore, which is robust against both class imbalance and training set noise." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.345, + 0.216, + 0.362 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.37, + 0.388, + 0.387 + ], + "angle": 0, + "content": "3.1. Memory-based anomaly detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.473 + ], + "angle": 0, + "content": "To train a memory-based anomaly detector, a feature extractor \\(f\\) (e.g., WideResNet) extracts a feature map \\(\\phi_i \\in \\mathbb{R}^{C \\times H \\times W}\\) from each training input image \\(x_i\\). Then, a coreset selection algorithm \\(S_{core}\\) is applied to the set \\(P = \\{\\phi_i^{(h,w)}\\}\\) of patch features \\(\\phi_i^{(h,w)}\\), resulting in a memory bank" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.48, + 0.484, + 0.498 + ], + "angle": 0, + "content": "\\[\nM _ {\\text {P a t c h C o r e}} := \\mathcal {S} _ {\\text {c o r e}} (P) \\subset P \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.58 + ], + "angle": 0, + "content": "that contains representative patch features of \\( P \\). For instance, PatchCore opts for greedy sampling strategy to obtain the coreset. In the inference stage, the anomaly score \\( s^{(h,w)} \\) of the testing image \\( x \\) is computed for each patch feature \\( \\phi^{(h,w)} \\) by measuring its distance to the memory" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.585, + 0.484, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s c o r e} ^ {(h, w)} = \\operatorname {d i s t} \\left(\\phi^ {(h, w)}, M\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.673 + ], + "angle": 0, + "content": "where \\(dist\\) is the nearest neighbor distance in [34]. The pixel-level anomaly score is computed by up-scaling the score map \\(score = [score^{(h,w)}] \\in \\mathbb{R}^{H,W}\\), and its image-level anomaly score is defined as its pixel maximum \\(\\max_{h,w} score^{(h,w)}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.375, + 0.696 + ], + "angle": 0, + "content": "3.2. Noise discrimination of memory" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.484, + 0.793 + ], + "angle": 0, + "content": "The training set \\( X = \\{x_{i}\\} \\) may be contaminated with defect samples. The naively attained memory bank then contains abnormal patch features, and assigns a low anomaly score on defect images, resulting in poor performance. To resolve this, [20] applies a noise discrimination algorithm \\( S_{clean} \\) on the patch feature set \\( P \\) prior to coreset sampling:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.798, + 0.484, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {\\text {c l e a n}} (P) = \\left\\{\\phi_ {i} ^ {(h, w)}: \\text {o u t l i e r} \\left(\\phi_ {i} ^ {(h, w)}\\right) < \\tau_ {0. 1 5} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "aggregating only the clean normal patches with low outlier scores. SoftPatch [20] exploits local outlier factor (LOF) for outlier scoring, and selects the threshold \\(\\tau_{0.15}\\) to remove only \\(15\\%\\) of whole patches. In inference, SoftPatch utilizes the memory set \\(M_{SoftPatch} = S_{core}(S_{clean}(P))\\)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.71, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.091, + 0.903, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.203, + 0.906, + 0.26 + ], + "angle": 0, + "content": "Figure 2. (left) The ratio of removed patches based on highest outlier scores by Eq. (3), which shows that most of few-shot class patches are lost. (right) The ratio of sampled patches by greedy coreset sampling from PatchCore which favors both few-shot and anomaly samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.272, + 0.73, + 0.287 + ], + "angle": 0, + "content": "4. Problem and motivation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.297, + 0.617, + 0.312 + ], + "angle": 0, + "content": "4.1. Problem" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.907, + 0.457 + ], + "angle": 0, + "content": "We consider a practical anomaly detection setting where the normal training data \\( X = \\{x_{i}\\} \\) is both multi-class with tail distribution and noise-contaminated. Particularly, each \\( x_{i} \\) is paired with the product class label \\( y_{i} \\in \\mathcal{Y} = \\{1,2,\\dots,|\\mathcal{Y}|\\} \\) (e.g., hazelnut, screw, pill, etc.) and an anomaly label \\( y_{i}^{a} \\in \\mathcal{Y}^{a} := \\{0,1\\} \\) with 1 indicating anomaly. However, this labeling information is unknown to the model trainer. The class distribution is long-tailed such that there are tail (i.e., few-shot) classes whose number of training samples is very low; i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.471, + 0.907, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\left| C _ {k} \\right| = \\left| \\left\\{x _ {i}: y _ {i} ^ {c} = k \\right\\} \\right| \\leq K \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.907, + 0.608 + ], + "angle": 0, + "content": "with a very small \\( K \\) (e.g., 1, 4, or 20), where \\( |\\cdot| \\) indicates the number of elements in the set. Moreover, some of the training samples in \\( X \\) are anomalies, containing defect blobs in their images. We assume the tail class samples are all normal since in the unsupervised setting its normality is determined by its majority. Overall, the training set \\( X \\) with the above descriptions constitutes unsupervised tailed noisy anomaly detection setting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.85, + 0.635 + ], + "angle": 0, + "content": "4.2. Motivation: Tail-versus-noise trade-off" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.764 + ], + "angle": 0, + "content": "An ideal model to handle a long-tail noisy dataset should selectively sample the few-shot samples while removing noisy samples. However, we observe that the existing models suffer from a trade-off between class-imbalance and noise robustness. For memory-based noise-discriminative models in particular (e.g., SoftPatch), the methods focus on sampling the majority patch features, which often ignores less dominating but important patches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 2 left, the noise discrimination process removes both defect and few-shot class patches, losing most of the information in the tail distribution. On the other hand, PatchCore samples few-shot samples well due to greedy coreset sampling which captures maximally different features. This sampling principle, however, favors the noise patch features as well as shown in Fig. 2 (right) and the low performance on noisy head classes in Fig. 1. Previous methods, hence, face the tail-versus-noise dilemma described above." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25541" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "To resolve this, we propose to handle the tail classes independently from the head classes that potentially contain defective local regions. This, however, needs exclusive sampling of tail classes. To achieve this, we use the globally average pooled embedding vector, which is invariant against local pixel variation and rather exclusively corresponds to the object-centric nature of image, i.e., class." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.212, + 0.179, + 0.228 + ], + "angle": 0, + "content": "5. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.239, + 0.485, + 0.33 + ], + "angle": 0, + "content": "We aim to build a coreset-based anomaly detection model using patch features. We require its memory to aptly capture tail class information and also to be absent of noise patches as much as possible. Particularly, we augment the noise discriminated memory \\( M_{clean} \\) with the tail class memory \\( M_{tail} \\), which contains the patches of few-shot class samples exclusively:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.343, + 0.484, + 0.358 + ], + "angle": 0, + "content": "\\[\nM _ {\\text {T a i l e d C o r e}} = M _ {\\text {c l e a n}} \\cup M _ {\\text {t a i l}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.485, + 0.464 + ], + "angle": 0, + "content": "The main challenge lies on sampling the tail class patches exclusively. To this end, we devise a few-shot sampler by estimating the class size of samples based on its embedding (e.g., global average pooling layer of the encoder). Using the estimated class sizes of training samples, the few-shot sampler enables us to exclusively sample tail class samples and the patches thereof." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.473, + 0.373, + 0.49 + ], + "angle": 0, + "content": "5.1. Few-shot sampler: TailSampler" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.484, + 0.557 + ], + "angle": 0, + "content": "TailSampler first estimates the class size of every training sample. Then, by automatically determining the maximum number of tail classes, denoted by \\( y_{\\mathrm{max}} \\), TailSampler collects the tail class samples' patch features exclusively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.566, + 0.274, + 0.581 + ], + "angle": 0, + "content": "5.1.1.Class size predictor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.72 + ], + "angle": 0, + "content": "To sample patches exclusively from few-shot class, we identify samples from few-shot classes by estimating their \"class sizes\", namely the size of class each sample belongs to. To accurately measure the class size, we hypothesize as follows: In the embedding space, samples within an appropriate angle are majorly of the same class, and this angle is likely to be the half of maximum angle to another embedding. In addition, the neighborhoods of neighborhoods within the half angle give accurate class size estimate. Based on these hypotheses, we proceed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.484, + 0.873 + ], + "angle": 0, + "content": "First, we adaptively set angle \\(\\alpha_{i}\\) for each embedding \\(e_i\\) from training samples \\(x_{i}\\). The coverage of angle \\(\\alpha_{i}\\) decides the neighbors where a large value would assign many neighbors. We set the angle to be half of its maximum within the train set, and hypothesize that most of neighbors within the region is likely of the same class. Particularly, let \\(H_{i} = \\{e\\in Z:\\angle (e_{i},e)\\leq m_{i} / 2\\}\\) denote the half-angle region where the maximum angle is \\(m_{i}\\coloneqq \\max_{e\\in Z}\\angle (e_{i},e)\\) and the set \\(Z\\) contains all train embeddings. Then, the adaptive angle \\(\\alpha_{i}\\) is defined to contain \\(p\\)-th percentile of the half-angle region:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.886, + 0.484, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\angle \\left(e _ {i}, e _ {\\left(p \\cdot | H _ {i} |\\right)}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.136 + ], + "angle": 0, + "content": "where the index \\((j)\\) of \\(e_{(j)}\\) is sorted in the increasing order of the angle \\(\\angle (e_i,e_{(j)})\\). We set \\(p = 0.85\\) in all experiments unless specified otherwise." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.228 + ], + "angle": 0, + "content": "After setting the adaptive angle \\(\\alpha_{i}\\) for each train embedding \\(e_i\\), we estimate its class size based on neighborhoods of neighborhoods. Let \\(N_{\\alpha}(e_i) = \\{e \\in Z : \\angle(e_i, e) \\leq \\alpha\\}\\) denote the neighborhood of \\(e_i\\), which is the set of all train embedding \\(e\\) within the angle \\(\\alpha\\) of \\(e_i\\). Then, the class size is estimated by the mode of the sizes of neighborhoods of the neighborhood:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.24, + 0.907, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\kappa_ {i} = \\underset {e \\in N _ {\\alpha_ {i}} (e _ {i})} {\\operatorname {m o d e}} \\left(| N _ {\\alpha (e)} (e) |\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.278, + 0.907, + 0.354 + ], + "angle": 0, + "content": "where \\(\\alpha(e)\\) is the adaptive angle with respect to the embedding \\(e\\) belonging to the neighborhood \\(N_{\\alpha_i}(e_i)\\) of embedding \\(e_i\\). Using the neighborhoods of neighborhood and mode on it gives more robust estimation of the class size than mere count of the direct neighborhood of \\(e_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.908, + 0.4 + ], + "angle": 0, + "content": "The proposition in Supp. 9 shows that the angle of neighborhood given by Eq. (6) corresponds to the decision boundary that maximizes the inter-class separation under regularities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.84, + 0.422 + ], + "angle": 0, + "content": "5.1.2. Estimation of maximum size of tail class" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.565 + ], + "angle": 0, + "content": "After estimating the class size of each training sample by \\(\\kappa_{i}\\), we determine the maximum size for the tail classes based on the elbow technique [36]. In particular, by utilizing the sample-wise class sizes \\(\\kappa_{i}\\), we first obtain rough estimation of the size of each class set \\(\\eta_{y} \\approx |C_{y}|\\). Then, we estimate the maximum size of tail classes \\(K_{max}\\) by finding a point where the class size most abruptly changes. This optimal shift point is found by the elbow technique. Concretely, we first estimate the size of each class \\(y\\), \\(\\eta_{y} \\approx |C_{y}|\\), inductively by" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.576, + 0.907, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\eta_ {1} = \\operatorname {r o u n d} \\left(\\frac {1}{\\kappa_ {(1)}} \\sum_ {i = 1} ^ {\\kappa_ {(1)}} \\kappa_ {(i)}\\right), \\\\ \\eta_ {y + 1} = \\operatorname {r o u n d} \\left(\\frac {1}{\\kappa_ {\\left(\\eta_ {y} + 1\\right)}} \\sum_ {i = \\eta_ {y} + 1} ^ {\\min \\left(\\kappa_ {\\left(\\eta_ {y} + 1\\right)}, | X |\\right)} \\kappa_ {(i)}\\right), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.679, + 0.906, + 0.711 + ], + "angle": 0, + "content": "where the index \\((i)\\) of \\(\\kappa_{(i)}\\) is sorted in the increasing order \\(\\kappa_{(1)}\\leq \\dots \\leq \\kappa_{(|X|)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.906, + 0.817 + ], + "angle": 0, + "content": "After we acquire the set \\(\\{(y,\\eta_y):y\\in (1,\\ldots ,|\\mathcal{Y}|)\\}\\) through the inductive process above, we define the maximum size of tail classes as the elbow point of the estimated sizes of classes \\(\\eta_{y}\\) since at this point the class size increases abruptly. The elbow is computed by linearly connecting \\((1,\\eta_{1})\\) and \\((|\\mathcal{V}|,\\eta_{|\\mathcal{V}|})\\) , and finding the maximum orthogonal distance onto the line from the set of points \\(\\{(y,\\eta_y):y\\in (1,\\dots |\\mathcal{Y}|)\\}\\) as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.829, + 0.907, + 0.846 + ], + "angle": 0, + "content": "\\[\nK _ {\\max } := \\eta_ {y _ {\\max }} =: e l b o w (\\eta_ {1}, \\dots , \\eta_ {| \\mathcal {Y} |}) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.858, + 0.558, + 0.871 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.87, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\ny _ {\\max } = \\underset {y \\in \\{1, \\dots , | \\mathcal {Y} | \\}} {\\operatorname {a r g m a x}} \\frac {| a y - \\eta_ {y} + b |}{\\sqrt {a ^ {2} + 1}} \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25542" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.258, + 0.851, + 0.273 + ], + "angle": 0, + "content": "Figure 3. (a) Sampling process description of TailedCore (ours) and (b) the illustration of how we use the elbow method [36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.483, + 0.346 + ], + "angle": 0, + "content": "and where the line \\(\\eta_y = ay + b\\) intersects \\((1, \\eta_1)\\) and \\((|\\mathcal{Y}|, \\eta_{|\\mathcal{Y}|})\\) with \\(a = (\\eta_{|\\mathcal{Y}|} - \\eta_1) / (|\\mathcal{Y}| - 1)\\) and \\(b = \\eta_1 - m\\). Fig. 3 (b) illustrates Eq. (10). We choose the elbow point \\((y_{max}, \\eta_{y_{max}})\\) as the maximum size of tail classes \\(K_{max}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.484, + 0.377 + ], + "angle": 0, + "content": "The maximum size of tail classes \\( K_{max} \\) allows us to sample the patch features of tail classes exclusively:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.384, + 0.484, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {\\text {t a i l}} (P) = \\left\\{\\phi_ {i} ^ {(h, w)}: \\kappa_ {i} \\leq K _ {\\max } \\right\\}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.249, + 0.427 + ], + "angle": 0, + "content": "with \\(\\kappa_{i}\\) given by Eq. (7)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.434, + 0.36, + 0.451 + ], + "angle": 0, + "content": "5.2. Proposed method: TailedCore" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.456, + 0.485, + 0.699 + ], + "angle": 0, + "content": "Our anomaly detection model is built upon the memory bank constructed as below. First, we sample the base memory bank as in [20] with noise-discrimination described in Eq. (3). Then, the base memory is augmented with TailSampler \\( S_{tail} \\) (Eq. (11)). Overall, TailedCore's memory bank is defined by \\( M_{TailedCore} = S_{core}(S_{clean}(P)) \\cup S_{core}(S_{tail}(P)) \\), where \\( S_{core} \\) is the greedy coreset sampling in Eq. (1), \\( S_{clean} \\) is the noise discrimination process that filters out defect patches based on their outlier scores (Eq. (3)), and the few-shot sampler \\( S_{tail} \\) augments the tail class patches exclusively on the noise-discriminated base memory bank. The whole process is shown in Fig 3. As TailedCore's patch memory bank is both noise-clean and contains the representative features of few-shot classes, it is robust against both class imbalance and normal sample contamination. In inference, the anomaly score is obtained by Eq. (2) as in the conventional memory-based detector." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.218, + 0.727 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.781 + ], + "angle": 0, + "content": "Section 6 consists of dataset description, comparison to the SOTA models, and the detailed ablation study and analysis of TailedCore's components." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.788, + 0.483, + 0.82 + ], + "angle": 0, + "content": "6.1. Datasets for unsupervised long-tail noisy anomaly detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We devise the long-tail noisy anomaly detection datasets by modifying the class-wise sample distributions of the widely used anomaly detection datasets MVTecAD [3] and VisA [44]. We consider three different types of long-tail distributions. Particularly, datasets with step-shaped tail distribution are devised" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.283, + 0.906, + 0.383 + ], + "angle": 0, + "content": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDim [9] ICPR2182.4580.9582.0677.4781.2879.1971.5481.7575.63
HVQ [26] NeurIPS2383.4680.2382.9982.0185.5083.5674.1590.1580.55
WinCLIP [19] CVPR2389.3590.1190.3791.6088.2190.3791.8088.2390.37
AnomalyCLIP [43] ICLR2490.9390.9891.4891.8290.8391.4891.2191.9091.48
PatchCore [34] CVPR2293.3387.5989.1892.1971.1883.8386.3670.4880.01
SoftPatch [20] NeurIPS2284.6886.9587.7167.6597.5479.6460.6697.4975.40
TailedCore (ours)96.5595.2496.1295.8295.3495.7193.5495.7794.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.384, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Table 1. Anomaly classification on MVTecAD with image-level AUROC \\((\\%)\\). We report the mean over 5 random seeds for each measurement. Notations: \\(C_h / C_t\\): head / tail classes." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.534 + ], + "angle": 0, + "content": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2170.7083.3578.6460.6588.9372.4355.9886.7568.80
HVQ [26]NeuRPS2373.4784.0368.2568.2589.3077.0261.5780.4069.42
WinCLIP [19]CVPR2373.2576.9275.4775.9874.7675.4778.8070.8075.47
AnomalyCLIP [43]ICLR2481.9682.4882.0582.2881.7482.0583.2680.3482.05
PatchCore [34]CVPR2286.1185.7385.5983.5367.5176.8579.3368.5674.84
SoftPatch [20]NeuRPS2278.0492.1686.5659.7095.9774.8152.6194.1769.92
TailedCore (ours)87.5593.0690.8585.1695.9189.6482.9794.1187.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.534, + 0.905, + 0.562 + ], + "angle": 0, + "content": "Table 2. Anomaly classification on VisA with image-level AUROC \\((\\%)\\). The format and evaluation protocol are the same as Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.672 + ], + "angle": 0, + "content": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2190.1192.6691.4382.5395.2987.6778.8095.5485.50
HVQ [26]NeuRPS2393.6386.8590.5590.7392.5891.5386.3695.2089.90
WinCLIP [19]CVPR2382.0384.0682.2980.6084.6382.2980.1685.4882.29
AnomalyCLIP [43]ICLR2491.2491.6991.0889.9692.6691.0889.3493.6891.08
PatchCore [34]CVPR2293.5687.9889.9393.5472.0985.1992.0271.3583.75
SoftPatch [20]NeuRPS2292.1993.8393.4180.9896.4987.2470.3496.8980.99
TailedCore (ours)96.0895.0195.2995.5693.2094.7494.1993.7093.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.671, + 0.908, + 0.713 + ], + "angle": 0, + "content": "Table 3. Anomaly segmentation on MVTecAD with pixel-level AUROC (\\%). We report the mean over 5 random seeds for each measurement. Notations: \\(C_h / C_t\\): head / tail classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "by making the particular classes to have only a small number of samples; namely \\( K = 1 \\), and \\( K = 4 \\), imposing extreme class imbalance on the datasets. We name these two datasets by 'step \\( (K = 1) \\)' and 'step \\( (K = 4) \\)'. On the other hand, another realistic type of tail-distribution is considered by making the class distribution to follow the Pareto distribution with the shape parameter 0.6. For the step-tailed distribution datasets, we impose \\( 40\\% \\) of classes to be head classes, containing the original number of training samples. Thus, in MVTecAD, 6 of 15 classes are head classes. For the Pareto tailed distribution, we regard as few-shot (i.e., tail classes) the classes whose number of samples" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25543" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.188 + ], + "angle": 0, + "content": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2189.0295.1082.8183.9097.3689.5182.5796.5788.40
HVQ [26]NeuIPS2395.2797.6096.7193.8898.3495.7490.5895.5192.63
WinCLIP [19]CVPR2371.9473.9773.1974.6071.2173.1973.8172.3273.19
AnomalyCLIP [43]ICLR2495.6095.4695.5195.5495.4895.5196.1694.6095.51
PatchCore [34]CVPR2296.8487.9991.1395.3962.9681.8894.1165.3082.10
SoftPatch [20]NeuIPS2293.2096.7495.2783.9597.1089.4380.7396.8287.43
TailedCore (ours)97.9897.2597.4896.8097.0296.8996.1297.3996.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.189, + 0.483, + 0.216 + ], + "angle": 0, + "content": "Table 4. Anomaly segmentation on VisA with pixel-level AUROC \\((\\%)\\). The format and evaluation protocol are the same as Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.229, + 0.484, + 0.336 + ], + "angle": 0, + "content": "is less than 20 [25]. To impose noisy condition on the datasets, the anomaly samples (i.e., the samples with defective regions) are added to the training samples such that the noisy samples constitute \\(10\\%\\) of the training set following the overlap protocol in SoftPatch [20] for fair comparison. In all cases, the testing sets of datasets remain the same as the original MVTecAD and VisA. A more detailed configuration is given in Supp. 11." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.345, + 0.453, + 0.361 + ], + "angle": 0, + "content": "6.2. Comparison to the state-of-the-art (SOTA)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.367, + 0.483, + 0.623 + ], + "angle": 0, + "content": "We compare TailedCore to the SOTA anomaly detection methods in the unsupervised long-tail noisy settings. We carefully select our baselines. HVQ excels in multi-class learning, PaDiM and SoftPatch are robust against noises as PaDiM is based on Gaussian modeling of the data, while SoftPatch explicitly removes the defect patches by outlier scoring. Using foundation models, WinClip and AnoamlyClip are strong for zero-shot learning. PatchCore are near-perfect in the single-class anomaly detection settings. We show the full experimental results in Supp. 12. All models are trained in the multi-class setting; namely, one model for one dataset with all classes rather than one model per class. The detection performance, however, is measured for each product class under image-level and pixel-level AUROC, and averaged for reporting and comparison. For comparison, we consider models that are specifically robust against noise, strong on multi-class training, and foundation models that can exhibit zero/few-shot learning capability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.73 + ], + "angle": 0, + "content": "Tabs. 1, 2, 3, and 4 show the superiority of TailedCore across all types of long-tail distributions. As anticipated, all baseline models encounter the tail-versus-noise trade-off, resulting in suboptimal performance either on tail classes with limited samples or on head classes with noisy training samples. These trends become more pronounced in highly unbalanced datasets, where the few-shot classes contain only 1 or 4 samples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.739, + 0.34, + 0.755 + ], + "angle": 0, + "content": "6.3. Ablation study and analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.762, + 0.307, + 0.777 + ], + "angle": 0, + "content": "6.3.1. Tail (few-shot) sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate how accurately our proposed algorithm \\( S_{tail} \\) can sample tail (few-shot) class samples exclusively without including noise samples. First of all, we measure the ratio of mis-included tail class samples TailSampler which is the ratio of not sampled tail class samples to total tail class samples. A single mis-inclusion can vitally damage anomaly detection performance especially in case of the one-shot setting \\( K = 1 \\) as the model wouldn't be able to utilize the only sample of the class." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.332 + ], + "angle": 0, + "content": "Additionally, the exclusive aspect of few-shot sampling is crucial; the few-shot sampled set shouldn't include anomaly and head class samples as possible. Included anomaly samples would harm the head class performance and head class sampled by few-shot sampler might cause class imbalance issue as well as the possibility of sampling anomaly from head classes. Hence, we measure the ratios of included anomalies and head-class samples along with AUROC which is measured on the outlier scores derived by a given algorithm and ground truth binary labels that indicate whether a given sample is tail class or head class. For TailSampler, the outlier score of each train sample is given as the negative of predicted class size. The included anomaly ratio is defined as the ratio of number of anomaly samples to the number of whole anomaly samples. The ratio of included head-class sample is defined as the number of included head class to the number of included tail class samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.484 + ], + "angle": 0, + "content": "For extensive comparison, we compare TailSampler with widely used clustering algorithms [2, 13, 16], outlier detectors [1, 5, 24, 35], and density estimators [10, 30]. These methods provide clustering size and/or outlier score/prediction. In cases when only outlier scores are available, we apply the elbow method to induce deterministic prediction of whether a given sample is few-shot or not. We use default settings provided by Sklearn [31] and PyOD [42]. For all methods, we utilize the embedding vectors of WideResNet-50 [40]. Experiments with different architectures and details are provided in Supp. 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.909, + 0.697 + ], + "angle": 0, + "content": "The combined ratio, which is the sum of missing tail samples and included anomaly samples, indicates that TailSampler is most effective for exclusively sampling few-shot class samples. The most critical factor is to not miss few-shot class samples. On this ratio, TailSampler achieves the lowest rate except for the algorithms that trivially sample all few-shot class samples. We find that most of the classical algorithms for clustering and outlier detection are not usable as they either miss too many few-shot samples or includes a large number of anomaly samples. We note that the class size predictor in our few-shot samples resembles DBSCAN and Affinity Propagation, respectively, in terms of using neighborhoods and voting mechanism. None of these methods however have instance-specific parameters that are adaptively adjusted across different samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.702, + 0.704, + 0.717 + ], + "angle": 0, + "content": "6.3.2. Class size prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.906, + 0.782 + ], + "angle": 0, + "content": "We measure how accurately our class size predictor estimates the class size of each sample by measuring the class size prediction error. It is measured for each train sample and weight-averaged through the whole train samples:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.789, + 0.907, + 0.832 + ], + "angle": 0, + "content": "\\[\ne r r o r = \\frac {1}{| \\mathcal {X} |} w _ {i} \\sum_ {i = 1} ^ {| \\mathcal {X} |} \\left| \\kappa_ {i} - \\left| C _ {k (i)} \\right| \\right| \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\( k(i) \\) indicates the index such that \\( x_{i} \\in C_{k(i)} \\) (namely, \\( C_{k(i)} \\) is the class of \\( x_{i} \\)). The weight is chosen by \\( w_{i} = 1 / |C_{k(i)}| \\), which is to give more penalty on the few-shot class samples more than the head class samples." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25544" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.283, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.203, + 0.197, + 0.213 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.215, + 0.282, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.327, + 0.197, + 0.337 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.34, + 0.282, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.452, + 0.197, + 0.462 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.09, + 0.48, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.202, + 0.394, + 0.213 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.215, + 0.48, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.327, + 0.394, + 0.338 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.34, + 0.48, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.451, + 0.394, + 0.462 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.545 + ], + "angle": 0, + "content": "Figure 4. The plot of noise ratio of train data versus anomaly classification (image-level AUROC) and segmentation (pixel-level AUROC) performance. In all cases, TailedCore outperforms both PatchCore and SoftPatch. Dataset is MVTecAD with different tail distributions. The lines show the mean and shades show standard deviation of multiple runs of experiments with different seeds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Tab. 5 shows that most of the clustering based methods vastly fail on this task, and the estimation by our class size predictor outperforms them significantly. We believe this is partly due to suboptimal hyperparameters used for these algorithms. Due to their classical nature, however, it is unclear how to tune them without the availability of the validation set in our scenario." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.668, + 0.406, + 0.684 + ], + "angle": 0, + "content": "6.3.3. Abalation study on class size predictor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.868 + ], + "angle": 0, + "content": "We conduct detailed ablation study on the proposed class size prediction in Eq. (7). One hyperparameter used in TailSampler \\( S_{tail} \\) of TailedCore is the percentiling proportion \\( p \\) to consider only the majority of nearby samples within the neighborhood. We analyze how this parameter can affect the anomaly detection performance of TailedCore. Tab. 5 indicates that using the nearest \\( 85\\% \\) embeddings (i.e., \\( p = 85\\% \\)) is the best choice. Using all of the available neighborhoods (i.e., \\( p = 100\\% \\)) would include different class samples in the neighborhood, deteriorating class estimation. On the other hand, counting only with the half of samples (i.e., \\( p = 50\\% \\)) in the neighborhood may lose too much information." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Another component in the class size predictor that requires analysis is its voting mechanism used in Eq. (7). One may" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.233 + ], + "angle": 0, + "content": "
methodcombined ratio of missing \\( {C}_{t} \\) & included AS (↓)ratio of missing \\( {C}_{t} \\) samples (↓)ratio of included AS (↓)ratio of included \\( {C}_{h} \\) samples (↓)class size prediction error (↓)AUROC (%) (↑)
LOF\\( {89.58} \\pm {44.42} \\)\\( {31.22} \\pm {34.88} \\)\\( {58.36} \\pm {9.54} \\)\\( {92.50} \\pm {7.87} \\)\\( {1.00} \\pm {0.00} \\)\\( {77.08} \\pm {25.86} \\)
IF\\( {18.15} \\pm {15.91} \\)\\( {7.25} \\pm {10.05} \\)\\( {10.90} \\pm {5.86} \\)\\( {69.02} \\pm {25.96} \\)\\( {1.00} \\pm {0.00} \\)\\( {98.96} \\pm {1.27} \\)
OCSVM\\( {49.95} \\pm {9.76} \\)\\( {0.78} \\pm {2.19} \\)\\( {49.17} \\pm {7.57} \\)\\( {95.06} \\pm {6.08} \\)\\( {1.00} \\pm {0.00} \\)\\( {96.72} \\pm {3.59} \\)
DBSCAN\\( {35.36} \\pm {49.55} \\)\\( {30.64} \\pm {44.38} \\)\\( {4.72} \\pm {5.17} \\)\\( {25.57} \\pm {27.24} \\)\\( {0.35} \\pm {0.44} \\)\\( {84.46} \\pm {22.18} \\)
KMeans\\( {94.67} \\pm {44.95} \\)\\( {47.22} \\pm {35.49} \\)\\( {47.45} \\pm {9.46} \\)\\( {97.40} \\pm {3.99} \\)\\( {4.27} \\pm {1.40} \\)\\( {52.58} \\pm {20.46} \\)
GMM\\( {100.00} \\pm {0.00} \\)\\( \\mathbf{{0.00}} \\pm \\mathbf{{0.00}} \\)\\( {100.00} \\pm {0.00} \\)\\( {97.58} \\pm {2.86} \\)\\( {12.50} \\pm {1.53} \\)\\( {50.00} \\pm {0.00} \\)
KDE\\( {29.81} \\pm {24.15} \\)\\( {2.14} \\pm {6.26} \\)\\( {27.67} \\pm {17.89} \\)\\( {87.47} \\pm {10.35} \\)\\( {1.00} \\pm {0.00} \\)\\( {97.98} \\pm {2.19} \\)
AP\\( {27.81} \\pm {32.59} \\)\\( {12.98} \\pm {24.01} \\)\\( {14.83} \\pm {8.58} \\)\\( {77.45} \\pm {28.21} \\)\\( {0.70} \\pm {0.22} \\)\\( {93.03} \\pm {12.8} \\)
KNN\\( {43.49} \\pm {31.25} \\)\\( {14.46} \\pm {24.07} \\)\\( {29.03} \\pm {7.18} \\)\\( {83.12} \\pm {18.28} \\)\\( {1.00} \\pm {0.00} \\)\\( {94.51} \\pm {10.34} \\)
TailSampler (ours)\\( {5.10} \\pm {5.07} \\)\\( {0.12} \\pm {0.65} \\)\\( {4.98} \\pm {4.42} \\)\\( {31.96} \\pm {26.05} \\)\\( \\mathbf{{0.19}} \\pm \\mathbf{{0.05}} \\)\\( {99.78} \\pm {0.30} \\)
Ablation on the class size predictor:
TailSampler (p=100)\\( {8.48} \\pm {10.88} \\)\\( {1.07} \\pm {5.23} \\)\\( {7.41} \\pm {5.65} \\)\\( {42.64} \\pm {31.86} \\)\\( {0.51} \\pm {0.04} \\)\\( {99.76} \\pm {0.29} \\)
TailSampler (p=50)\\( {7.37} \\pm {10.35} \\)\\( {3.46} \\pm {6.80} \\)\\( {3.91} \\pm {3.55} \\)\\( {27.90} \\pm {27.36} \\)\\( {0.12} \\pm {0.12} \\)\\( {97.36} \\pm {5.65} \\)
TailSampler (top-1)\\( {14.47} \\pm {7.99} \\)\\( {0.20} \\pm {0.78} \\)\\( {14.27} \\pm {7.21} \\)\\( {73.97} \\pm {27.95} \\)\\( {0.85} \\pm {0.04} \\)\\( {99.73} \\pm {0.34} \\)
TailSampler (average)\\( {6.48} \\pm {10.11} \\)\\( {3.65} \\pm {6.99} \\)\\( {2.83} \\pm {3.12} \\)\\( {22.16} \\pm {24.33} \\)\\( {0.21} \\pm {0.07} \\)\\( {99.35} \\pm {2.59} \\)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.234, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Table 5. The analysis on the proposed TailSampler and its class size predictor. The results are obtained from both MVTecAD and VisA and averaged across all different types of tail distributions and random seeds (overall 60 different seeds) with the standard deviations indicated on the right. The best and second best performances are marked in bold and underline, respectively. Acronyms/notations: IF: Isolation Forest; AP: Affinity Propagation; AS: anomaly samples; \\( C_h / C_t \\): head / tail classes." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.369, + 0.907, + 0.447 + ], + "angle": 0, + "content": "
tasktail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
anomaly classificationw/o FSA82.2291.2288.1360.0197.6075.0566.2897.8578.94
w/ FSA96.5595.2496.1393.5495.7894.4495.8295.3495.72
anomaly segmentationw/o FSA91.6094.9493.5468.9996.8580.1479.3996.3586.23
w/ FSA96.0995.0195.3094.1993.7194.0095.5693.2094.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.449, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Table 6. Abaltion on the few-shot augmentation (FSA) of TailedCore. The image-level and pixel-level AUROC metrics \\((\\%)\\) are reported on the MVTecAD dataset with different types of tail distributions for anomaly classification and segmentation tasks, respectively. Notations: \\(C_h / C_t\\): head / tail classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.641 + ], + "angle": 0, + "content": "simply use the nearest embedding's neighborhood size instead, or the class size of \\( x_{i} \\) can be estimated by averaging instead of majority vote. As indicated in Tab. 5, both the nearest embedding (i.e., top-1) and averaging are inferior to majority voting. The former is inevitably less robust, while averaging can still be sensitive to the outlier values of neighborhood size." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.724, + 0.68 + ], + "angle": 0, + "content": "6.3.4. Ablation on TailedCore" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We analyze the impact of augmentation of few-shot class patches by TailSampler in TailedCore. The TailedCore model without these augmented patches has the same memory bank as SoftPatch, but its inference scoring is simpler. Tab. 6 indicates that without the augmented patches from TailSampler, the memory-based model cannot detect anomalies in few-shot class samples. This shows the necessity of tail class augmentation by TailSampler. We note some performance decrease in the head class detection performance since few-shot augmentation may involve few anomaly samples from the head classes that behave like few-shot class instances. This issue intrinsically is due to the weakness of encoder embedding, and the proposed TailSampler based on such embeddings can induce incorrectly sampled few-shot class instances, which we discuss in Sec. 7." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25545" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.287, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.206, + 0.2, + 0.217 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.092, + 0.49, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.206, + 0.405, + 0.217 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.695, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.206, + 0.61, + 0.217 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.092, + 0.9, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.205, + 0.816, + 0.217 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.22, + 0.285, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.335, + 0.2, + 0.344 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.22, + 0.49, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.335, + 0.405, + 0.344 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.22, + 0.695, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.335, + 0.61, + 0.346 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.22, + 0.9, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.335, + 0.815, + 0.345 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.346, + 0.907, + 0.415 + ], + "angle": 0, + "content": "Figure 5. Classification accuracy between tail classes and noisy samples versus metrics relevant to class size prediction and few-shot sampling on MVtecAD and VisA datasets with step-like tailed distribution and \\( K = 4 \\). The correlation is strong for (a) mis-sampling ratio, (b) ratio of missing few-shot samples, (e) class size prediction error, and (f) AUROC for predicting whether a sample is few-shot or not. We show that improving the discriminative aspect of network's embeddings improves few-shot sampling and prediction error of class size predictor, which in turn improves (g) anomaly classification (image-level AUROC) and (h) anomaly segmentation (pixel-level AUROC) performance of TailedCore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.433, + 0.212, + 0.449 + ], + "angle": 0, + "content": "6.4. Noise ratio" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.533 + ], + "angle": 0, + "content": "We analyze the impact of noise ratio. Training data with a greater number of defect samples increase the chance of sampling noise instances for the few-shot sampler. We analyze this trend in Fig. 4, but the impact is not severe. In all cases, TailedCore either outperforms or is on par with both PatchCore and SoftPatch." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.544, + 0.484, + 0.576 + ], + "angle": 0, + "content": "6.5. Relation between the discriminative aspect of embedding and TailedCore" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.734 + ], + "angle": 0, + "content": "TailedCore utilizes the embedding features \\( e_i \\) of training samples \\( x_i \\) for few-shot sampling and augmentation thereof. Therefore, TailedCore's detection performance highly relies on the discriminative aspect of the embedding feature. Poorly designed embedding features will not discriminate different class samples in the embedding space. In this case, TailSampler cannot sample tail class samples and it may overly many head class samples. TailedCore's memory bank obtained so will neither be class-balanced nor noise-free, making the inference of TailedCore poor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.809 + ], + "angle": 0, + "content": "In contrast, a highly discriminative embedding features will make TailedCore's class size prediction accurate, thereby enabling it to sample the few-shot class samples exclusively without aggregating any anomaly and head class samples. In which case, TailedCore is both noise-free and class-balanced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We test this behavior by Oracle labeling and training the embedding features with class labels. Fig. 5 shows that improving discriminative quality of embedding (measured by its classification accuracy on noise and tail class samples) accordingly improves the relevant metrics. Particularly, more discriminative embeddings make more accurate class size" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.465 + ], + "angle": 0, + "content": "prediction, improving TailSampler along with its anomaly classification and segmentation performances." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.48, + 0.63, + 0.496 + ], + "angle": 0, + "content": "7. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.909, + 0.658 + ], + "angle": 0, + "content": "The class size prediction in TailSampler can fail if the reflective-symmetric assumption on the inter-class and intra-class similarities breaks down. Particularly, this case can occur if the utilized encoder embedding is either poor or not aligned with the label space of classes. In some cases, the geometric aspect of defect samples highly resembles that of few-shot class instances in the embedding space. We believe that this is due to imperfect nature of encoder embedding, which is often more sensitive to spurious features than the notion of class and object-centric aspect [29]. This issue can be improved by enhancing the encoder network." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.673, + 0.628, + 0.689 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.911 + ], + "angle": 0, + "content": "In this study, we introduced TailedCore, a novel memory-based anomaly detection model designed to effectively address the challenges of unsupervised anomaly detection in noisy and long-tail class distributions. By developing a unique class size predictor and a tailored memory bank, we successfully navigated the tail-versus-noise dilemma, enhancing the model's performance in identifying anomalies within a contaminated and imbalanced dataset. Our comprehensive evaluations demonstrated TailedCore's superior ability to mitigate noise contamination and class imbalance, underscoring its significance in advancing anomaly detection technologies. This work not only showcases a significant stride in handling real-world anomaly detection scenarios but also lays a robust foundation for future research in this critical field." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25546" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.259 + ], + "angle": 0, + "content": "Acknowledgments Kuan-Chuan Peng was exclusively supported by Mitsubishi Electric Research Laboratories. This work was also supported by the Technology Innovation Program (20023364) funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea), NSF grant 2038493, ONR grant N00014-21-1-2431, NIH grant R01CA240771 from NCI, and U.S. Department of Homeland Security grant 22STESE00001-03-02. The views and conclusions contained in this document are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the U.S. Department of Homeland Security." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.271, + 0.184, + 0.287 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.296, + 0.486, + 0.351 + ], + "angle": 0, + "content": "[1] Fabrizio Angiulli and Clara Pizzuti. Fast outlier detection in high dimensional spaces. In European conference on principles of data mining and knowledge discovery, pages 15-27. Springer, 2002. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.352, + 0.485, + 0.381 + ], + "angle": 0, + "content": "[2] David Arthur, Sergei Vassilvitskii, et al. k-means++: The advantages of careful seeding. In Soda, pages 1027-1035, 2007. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.381, + 0.485, + 0.45 + ], + "angle": 0, + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.451, + 0.485, + 0.507 + ], + "angle": 0, + "content": "[4] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4):947-969, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.485, + 0.563 + ], + "angle": 0, + "content": "[5] Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and Jörg Sander. Lof: identifying density-based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pages 93-104, 2000. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.485, + 0.619 + ], + "angle": 0, + "content": "[6] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.485, + 0.675 + ], + "angle": 0, + "content": "[7] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.485, + 0.732 + ], + "angle": 0, + "content": "[8] Zhuohang Chen, Jinglong Chen, Yong Feng, Shen Liu, Tianci Zhang, Kaiyu Zhang, and Wenrong Xiao. Imbalance fault diagnosis under long-tailed distribution: Challenges, solutions and prospects. Knowledge-Based Systems, 258:110008, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.485, + 0.802 + ], + "angle": 0, + "content": "[9] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.486, + 0.858 + ], + "angle": 0, + "content": "[10] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[11] Jan Diers and Christian Pigorsch. A survey of methods for automated quality control based on images. International Journal of Computer Vision, pages 1-29, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.296, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[13] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[14] Chaowei Fang, Lechao Cheng, Huiyan Qi, and Dingwen Zhang. Combating noisy labels in long-tailed image classification. arXiv preprint arXiv:2209.00273, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[15] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[16] Brendan J Frey and Delbert Dueck. Clustering by passing messages between data points. science, 315(5814):972-976, 2007. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[17] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. arXiv preprint arXiv:2308.15366, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.475 + ], + "angle": 0, + "content": "[18] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.907, + 0.546 + ], + "angle": 0, + "content": "[19] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero-/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[20] Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai Wu, Yong Liu, Chengjie Wang, and Feng Zheng. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35:15433-15445, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.907, + 0.644 + ], + "angle": 0, + "content": "[21] Justin M Johnson and Taghi M Khoshgoftaar. Survey on deep learning with class imbalance. Journal of Big Data, 6(1):1-54, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.907, + 0.688 + ], + "angle": 0, + "content": "[22] Yujin Lee, Harin Lim, Seoyoon Jang, and Hyunsoo Yoon. Uniformality: Towards task-agnostic unified framework for visual anomaly detection. Available at SSRN 4586132. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.907, + 0.744 + ], + "angle": 0, + "content": "[23] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. arXiv preprint arXiv:2401.16753, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[24] Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 eighth ieee international conference on data mining, pages 413-422. IEEE, 2008. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.845 + ], + "angle": 0, + "content": "[25] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[26] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems, 36, 2023. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25547" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[27] Yang Lu, Yiliang Zhang, Bo Han, Yiu-ming Cheung, and Hanzi Wang. Label-noise learning with intrinsically long-tailed data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1369-1378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.151, + 0.486, + 0.207 + ], + "angle": 0, + "content": "[28] Declan McIntosh and Alexandra Branzan Albu. Inter-realization channels: Unsupervised anomaly detection beyond one-class classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6285-6295, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.209, + 0.485, + 0.263 + ], + "angle": 0, + "content": "[29] Yifei Ming, Hang Yin, and Yixuan Li. On the impact of spurious correlation for out-of-distribution detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10051-10059, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.266, + 0.483, + 0.307 + ], + "angle": 0, + "content": "[30] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33 (3):1065-1076, 1962. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.31, + 0.485, + 0.379 + ], + "angle": 0, + "content": "[31] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.381, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[32] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.424, + 0.485, + 0.495 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.497, + 0.485, + 0.566 + ], + "angle": 0, + "content": "[34] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.568, + 0.485, + 0.624 + ], + "angle": 0, + "content": "[35] Bernhard Schölkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. Advances in neural information processing systems, 12, 1999. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.626, + 0.485, + 0.654 + ], + "angle": 0, + "content": "[36] Robert L Thorndike. Who belongs in the family? Psychometrika, 18(4):267-276, 1953. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.655, + 0.485, + 0.697 + ], + "angle": 0, + "content": "[37] Tong Wei, Jiang-Xin Shi, Wei-Wei Tu, and Yu-Feng Li. Robust long-tailed learning under label noise. arXiv preprint arXiv:2108.11569, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.7, + 0.485, + 0.741 + ], + "angle": 0, + "content": "[38] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Feng Zheng, and Yaochu Jin. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. arXiv preprint arXiv:2301.12082, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.744, + 0.485, + 0.799 + ], + "angle": 0, + "content": "[39] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.802, + 0.485, + 0.83 + ], + "angle": 0, + "content": "[40] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.832, + 0.485, + 0.9 + ], + "angle": 0, + "content": "[41] Manyi Zhang, Xuyang Zhao, Jun Yao, Chun Yuan, and Weiran Huang. When noisy labels meet long tail dilemmas: A representation calibration method. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15890-15900, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[42] Yue Zhao, Zain Nasrullah, and Zheng Li. Pyod: A python toolbox for scalable outlier detection. Journal of Machine Learning Research, 20(96):1-7, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[43] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In International Conference on Learning Representations, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[44] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. arXiv preprint arXiv:2207.14315, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25548" + } + ] +] \ No newline at end of file diff --git a/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_origin.pdf b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..43027324e90f7736a4e0b393b0f5e61a3b2ebd50 --- /dev/null +++ b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/5a75e7a0-a353-42d1-89a1-a561d7342d9c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1a40c74fef884ecc96d7029a6bdb6c37a40e9b64a83b64fea6712b9fb4afd1a +size 1918588 diff --git a/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/full.md b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d404b24bfb0b50d9ed089e6e839e22ad9fac53ae --- /dev/null +++ b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/full.md @@ -0,0 +1,392 @@ +# TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection + +Yoon Gyo Jung\*1 Jaewoo Park\*2 Jaeho Yoon\*3 Kuan-Chuan Peng4 Wonchul $\mathrm{Kim}^2$ Andrew Beng Jin Teoh Octavia Camps + +$^{1}$ Northeastern University $^{2}$ AiV Co. $^{3}$ Yonsei University $^{4}$ Mitsubishi Electric Research Laboratories + +# Abstract + +We aim to solve unsupervised anomaly detection in a practical challenging environment where the normal dataset is both contaminated with defective regions and its product class distribution is tailed but unknown. We observe that existing models suffer from tail-versus-noise trade-off where if a model is robust against pixel noise, then its performance deteriorates on tail class samples, and vice versa. To mitigate the issue, we handle the tail class and noise samples independently. To this end, we propose TailSampler, a novel class size predictor that estimates the class cardinality of samples based on a symmetric assumption on the class-wise distribution of embedding similarities. TailSampler can be utilized to sample the tail class samples exclusively, allowing to handle them separately. Based on these facets, we build a memory-based anomaly detection model TailedCore, whose memory both well captures tail class information and is noise-robust. We extensively validate the effectiveness of TailedCore on the unsupervised long-tail noisy anomaly detection setting, and show that TailedCore outperforms the state-of-the-art in most settings. Code is available in TailedCore + +# 1. Introduction + +In the complex landscape of anomaly detection, the challenge often lies in navigating practical scenarios that feature a diverse and imbalanced data distribution [6, 8, 11, 21]. We delve into a challenging yet highly realistic scenario in anomaly detection, where the training dataset, comprising multiple product classes [39], is beset by two significant complications: contamination with noise [20] and presence of tailed class distributions [25]. In this context, 'tail' (few-shot) classes have limited few-shot samples, unlike 'head' (many-shot) classes which are data-rich but might contain samples with defects. Crucially, in this unsupervised set- + +![](images/08f2b2036c81a6d05697400ceebc490f94245dfd6bab865ac9ff521699b12b5f.jpg) +SimpleNet +WinCLIP +# +IVQ +# +bftPatch + +。 +PaDiM + +Figure 1. Tail class (x-axis) and noisy head class (y-axis) performance comparison. The tail-versus-noise trade-off is shown across memory-based anomaly detection models (circles), and is more indicative in anomaly classification task evaluated by image-level AUROC (left). + +ting, the class information of these products remains unknown to the model trainer. This setting mirrors a common predicament in large-scale industrial applications, where companies are required to manage vast production lines encompassing a wide array of product types, each with varying production rates. + +A critical issue arises in this scenario, which we term the "tail-versus-noise dilemma." The dilemma underscores a fundamental trade-off where models robust to class imbalance overfit to defect noises, likely misclassifying many defects as normal. In contrast, existing models that are robust against defect noise tend to underfit, missing vital details in tail classes. Fig. 1 shows this tail-versus-noise trade-off which is prevalent through existing anomaly detection algorithms. This intrinsic characteristic comes from the fact that few-shot class features are equivalent to the noisy abnormal features in terms of their statistical occurrence. Consequently, existing models find it difficult to strike a balance between avoiding overfitting to noise and effectively capturing the nuances of less represented classes. + +To resolve this challenge, we propose to handle tail classes independently from head classes by exclusively sampling them. To this end, we devise a novel class size predictor TailSampler that estimates the cardinality of the class of any training sample. For accurate estimation, TailSampler estimates the class sizes based on a reflective symmetry between inter-class + +and intra-class similarity distributions in the embedding space. The estimated class sizes of the training samples then enable to sample tail class features exclusively from the training set. + +By utilizing the aggregated few-shot class instances, we devise a memory-based anomaly detector TailedCore, which is built upon the noise-discriminated model by augmenting the tail class patch features sampled with TailSampler. As the memory of TailedCore is both clean from noisy defect patches and preserves the few-shot class information, it exhibits robustness against both noise contamination and class imbalance of the training set. + +Our contributions are summarized as follows: + +- We explore a practical and challenging anomaly detection scenario, where a single detection model is trained on multi-class product training samples that involve noise contamination and whose class distribution is long-tailed. +- We propose a memory-based anomaly detector TailedCore whose patch memory bank is both noise-free and augmented with the representative patches of few-shot class instances. To build TailedCore, we introduce TailSampler that can aggregate few-shot class instances exclusively. TailSampler is based on a novel class size predictor that estimates the class size of training samples based on the reflective symmetry between inter-class and intra-class class-wise embedding similarity distributions. +- We conduct extensive analyses and comparative studies on both TailedCore and few-shot sampler along with the proposed class size predictor on the proposed unsupervised long-tail noisy anomaly detection benchmark. + +The codes for TailedCore along with the data generation will be open sourced. We use the term 'few-shot' and 'tail' interchangeably to denote the class with a small number of samples. + +# 2. Related works + +# 2.1. Anomaly detection in extreme scenarios + +One vitally important aspect of anomaly detection [9, 34] is the environment where these models are tested. This exploration has led to the identification of several challenging scenarios, including multi-class settings with training data spanning multiple product types [26, 39], zero-shot and few-shot settings with minimal training images [15, 19, 23, 38, 43], and noisy settings featuring uncleaned training data with defective images [7, 20, 28]. + +Multi-class anomaly detection In contrast to single-product settings, anomaly detection models trained on multi-class datasets often encounter "shortcut learning" [39], leading to low anomaly scores for normal images. This issue is particularly pronounced in reconstruction-based models, which tend to reconstruct all samples, regardless of their normality or abnormality. To resolve it, UniAD [39] employs neighbor masked attention within a transformer model, focusing attention on local parts to prevent shortcut learning. HVQ [26] utilizes vector quantization to discretize the latent space, restricting the network's recon + +striction fidelity on normal data distributions. UniFormaly [22] increases sensitivity to irrelevant patterns by masking less relevant patches and employing top-k matchings with a pretrained foundation model. These multi-class anomaly detection methods do not consider either long-tail classes or contaminated training data setups, and do not work well in the combined "noisy long-tail" setup, which is shown in our experimental results. + +Zero/few-shot anomaly detection The inception of few-shot learning benchmarks in anomaly detection marked a significant advancement, with methods leveraging template matching [18] and graph representations [38] for minimal training data scenarios. AnomalyCLIP [43], WinCLIP [19], and AnomalyGPT [17], on the other hand, detect anomalies in the few-shot classes based on the acquired knowledge of pretrained foundation models such as CLIP [33] and GPT [32]. However, these methods often require sophisticated prompt engineering or delicate prompt-tuning, and have yet to achieve optimal precision and recall in the single-class training setting with a sufficient amount of training samples. In contrast, our proposed TailedCore needs none of these process. + +Noise anomaly detection Dealing with training datasets contaminated with defective samples presents another extreme learning scenario. SoftPatch [20] and InReaCh [28] focus on purifying memory coresets through outlier detection algorithms and distance association filtering, respectively. IGD [7] adopts a robust Gaussian modeling approach, addressing both contaminated training data and small dataset challenges. Yet, these methods do not fully consider scenarios involving both data contamination and class imbalance, unlike our proposed TailedCore. + +# 2.2. Long-tail noisy learning + +In the realm of image classification, several studies have focused on learning scenarios involving training sets with label contamination and class imbalance. To address these challenges, the existing literature relies on the robustness and unsupervised aspects of self-supervised learning [14, 27, 37, 41]. The learned embeddings of self-supervised frameworks are less impacted by label contamination and class imbalances. However, these works differ from ours in that they deal with label-level contamination rather than pixel-level defects. Furthermore, these methods require explicit supervised labels to remediate the contamination. In contrast, our work targets noise in pixels without any class label information. + +# 2.3. Few-shot/outlier sampling + +Both classical clustering methods and outlier detection algorithms can be applied to our task, as they facilitate the identification of few-shot tail class samples, either indirectly or directly. Unsupervised clustering algorithms [2, 13, 16] can estimate the cluster sizes but struggle to identify small sample clusters in the imbalanced data. DBSCAN [13], on the other hand, captures both clusters and outliers. This however can be problematic when attempting to identify clusters with only one + +sample (1-shot) as it may be confused between few-shot samples and outliers. Unlike clustering methods, outlier detection algorithms [1, 5, 24, 35] are adept at identifying 1-shot samples which inherently resemble outliers. However, their effectiveness diminishes when a few-shot class contains more than one sample, as these no longer fit the typical outlier profile, potentially leading to incorrect identification of few-shot class samples. Both algorithm types, due to their inherent design, struggle to accurately predict class/cluster sizes in datasets that are both imbalanced and contaminated. Our proposed class size predictor specifically addresses this issue, aiming to reliably and accurately identify few-shot class samples and enable their exclusive sampling. Ultimately, the identified few-shot class samples are augmented to the noise-discriminated patch set, constituting the memory-based anomaly detection model TailedCore, which is robust against both class imbalance and training set noise. + +# 3. Background + +# 3.1. Memory-based anomaly detection + +To train a memory-based anomaly detector, a feature extractor $f$ (e.g., WideResNet) extracts a feature map $\phi_i \in \mathbb{R}^{C \times H \times W}$ from each training input image $x_i$ . Then, a coreset selection algorithm $S_{core}$ is applied to the set $P = \{\phi_i^{(h,w)}\}$ of patch features $\phi_i^{(h,w)}$ , resulting in a memory bank + +$$ +M _ {\text {P a t c h C o r e}} := \mathcal {S} _ {\text {c o r e}} (P) \subset P \tag {1} +$$ + +that contains representative patch features of $P$ . For instance, PatchCore opts for greedy sampling strategy to obtain the coreset. In the inference stage, the anomaly score $s^{(h,w)}$ of the testing image $x$ is computed for each patch feature $\phi^{(h,w)}$ by measuring its distance to the memory + +$$ +\operatorname {s c o r e} ^ {(h, w)} = \operatorname {d i s t} \left(\phi^ {(h, w)}, M\right), \tag {2} +$$ + +where $dist$ is the nearest neighbor distance in [34]. The pixel-level anomaly score is computed by up-scaling the score map $score = [score^{(h,w)}] \in \mathbb{R}^{H,W}$ , and its image-level anomaly score is defined as its pixel maximum $\max_{h,w} score^{(h,w)}$ . + +# 3.2. Noise discrimination of memory + +The training set $X = \{x_{i}\}$ may be contaminated with defect samples. The naively attained memory bank then contains abnormal patch features, and assigns a low anomaly score on defect images, resulting in poor performance. To resolve this, [20] applies a noise discrimination algorithm $S_{clean}$ on the patch feature set $P$ prior to coreset sampling: + +$$ +\mathcal {S} _ {\text {c l e a n}} (P) = \left\{\phi_ {i} ^ {(h, w)}: \text {o u t l i e r} \left(\phi_ {i} ^ {(h, w)}\right) < \tau_ {0. 1 5} \right\}, \tag {3} +$$ + +aggregating only the clean normal patches with low outlier scores. SoftPatch [20] exploits local outlier factor (LOF) for outlier scoring, and selects the threshold $\tau_{0.15}$ to remove only $15\%$ of whole patches. In inference, SoftPatch utilizes the memory set $M_{SoftPatch} = S_{core}(S_{clean}(P))$ + +![](images/06b4eb98e12f75e6d68ca17b8960d9f9ea0cbb1c919b0e1cd543fb93e8ec03a3.jpg) +Figure 2. (left) The ratio of removed patches based on highest outlier scores by Eq. (3), which shows that most of few-shot class patches are lost. (right) The ratio of sampled patches by greedy coreset sampling from PatchCore which favors both few-shot and anomaly samples. + +![](images/87aadc70551d1e473d78aa2ab458fd971f200f34b2fb6e4e45987b03de6c0509.jpg) + +# 4. Problem and motivation + +# 4.1. Problem + +We consider a practical anomaly detection setting where the normal training data $X = \{x_{i}\}$ is both multi-class with tail distribution and noise-contaminated. Particularly, each $x_{i}$ is paired with the product class label $y_{i} \in \mathcal{Y} = \{1,2,\dots,|\mathcal{Y}|\}$ (e.g., hazelnut, screw, pill, etc.) and an anomaly label $y_{i}^{a} \in \mathcal{Y}^{a} := \{0,1\}$ with 1 indicating anomaly. However, this labeling information is unknown to the model trainer. The class distribution is long-tailed such that there are tail (i.e., few-shot) classes whose number of training samples is very low; i.e., + +$$ +\left| C _ {k} \right| = \left| \left\{x _ {i}: y _ {i} ^ {c} = k \right\} \right| \leq K \tag {4} +$$ + +with a very small $K$ (e.g., 1, 4, or 20), where $|\cdot|$ indicates the number of elements in the set. Moreover, some of the training samples in $X$ are anomalies, containing defect blobs in their images. We assume the tail class samples are all normal since in the unsupervised setting its normality is determined by its majority. Overall, the training set $X$ with the above descriptions constitutes unsupervised tailed noisy anomaly detection setting. + +# 4.2. Motivation: Tail-versus-noise trade-off + +An ideal model to handle a long-tail noisy dataset should selectively sample the few-shot samples while removing noisy samples. However, we observe that the existing models suffer from a trade-off between class-imbalance and noise robustness. For memory-based noise-discriminative models in particular (e.g., SoftPatch), the methods focus on sampling the majority patch features, which often ignores less dominating but important patches. + +As shown in Fig. 2 left, the noise discrimination process removes both defect and few-shot class patches, losing most of the information in the tail distribution. On the other hand, PatchCore samples few-shot samples well due to greedy coreset sampling which captures maximally different features. This sampling principle, however, favors the noise patch features as well as shown in Fig. 2 (right) and the low performance on noisy head classes in Fig. 1. Previous methods, hence, face the tail-versus-noise dilemma described above. + +To resolve this, we propose to handle the tail classes independently from the head classes that potentially contain defective local regions. This, however, needs exclusive sampling of tail classes. To achieve this, we use the globally average pooled embedding vector, which is invariant against local pixel variation and rather exclusively corresponds to the object-centric nature of image, i.e., class. + +# 5. Method + +We aim to build a coreset-based anomaly detection model using patch features. We require its memory to aptly capture tail class information and also to be absent of noise patches as much as possible. Particularly, we augment the noise discriminated memory $M_{clean}$ with the tail class memory $M_{tail}$ , which contains the patches of few-shot class samples exclusively: + +$$ +M _ {\text {T a i l e d C o r e}} = M _ {\text {c l e a n}} \cup M _ {\text {t a i l}}. \tag {5} +$$ + +The main challenge lies on sampling the tail class patches exclusively. To this end, we devise a few-shot sampler by estimating the class size of samples based on its embedding (e.g., global average pooling layer of the encoder). Using the estimated class sizes of training samples, the few-shot sampler enables us to exclusively sample tail class samples and the patches thereof. + +# 5.1. Few-shot sampler: TailSampler + +TailSampler first estimates the class size of every training sample. Then, by automatically determining the maximum number of tail classes, denoted by $y_{\mathrm{max}}$ , TailSampler collects the tail class samples' patch features exclusively. + +# 5.1.1.Class size predictor + +To sample patches exclusively from few-shot class, we identify samples from few-shot classes by estimating their "class sizes", namely the size of class each sample belongs to. To accurately measure the class size, we hypothesize as follows: In the embedding space, samples within an appropriate angle are majorly of the same class, and this angle is likely to be the half of maximum angle to another embedding. In addition, the neighborhoods of neighborhoods within the half angle give accurate class size estimate. Based on these hypotheses, we proceed as follows: + +First, we adaptively set angle $\alpha_{i}$ for each embedding $e_i$ from training samples $x_{i}$ . The coverage of angle $\alpha_{i}$ decides the neighbors where a large value would assign many neighbors. We set the angle to be half of its maximum within the train set, and hypothesize that most of neighbors within the region is likely of the same class. Particularly, let $H_{i} = \{e\in Z:\angle (e_{i},e)\leq m_{i} / 2\}$ denote the half-angle region where the maximum angle is $m_{i}\coloneqq \max_{e\in Z}\angle (e_{i},e)$ and the set $Z$ contains all train embeddings. Then, the adaptive angle $\alpha_{i}$ is defined to contain $p$ -th percentile of the half-angle region: + +$$ +\alpha_ {i} = \angle \left(e _ {i}, e _ {\left(p \cdot | H _ {i} |\right)}\right), \tag {6} +$$ + +where the index $(j)$ of $e_{(j)}$ is sorted in the increasing order of the angle $\angle (e_i,e_{(j)})$ . We set $p = 0.85$ in all experiments unless specified otherwise. + +After setting the adaptive angle $\alpha_{i}$ for each train embedding $e_i$ , we estimate its class size based on neighborhoods of neighborhoods. Let $N_{\alpha}(e_i) = \{e \in Z : \angle(e_i, e) \leq \alpha\}$ denote the neighborhood of $e_i$ , which is the set of all train embedding $e$ within the angle $\alpha$ of $e_i$ . Then, the class size is estimated by the mode of the sizes of neighborhoods of the neighborhood: + +$$ +\kappa_ {i} = \underset {e \in N _ {\alpha_ {i}} (e _ {i})} {\operatorname {m o d e}} \left(| N _ {\alpha (e)} (e) |\right), \tag {7} +$$ + +where $\alpha(e)$ is the adaptive angle with respect to the embedding $e$ belonging to the neighborhood $N_{\alpha_i}(e_i)$ of embedding $e_i$ . Using the neighborhoods of neighborhood and mode on it gives more robust estimation of the class size than mere count of the direct neighborhood of $e_i$ . + +The proposition in Supp. 9 shows that the angle of neighborhood given by Eq. (6) corresponds to the decision boundary that maximizes the inter-class separation under regularities. + +# 5.1.2. Estimation of maximum size of tail class + +After estimating the class size of each training sample by $\kappa_{i}$ , we determine the maximum size for the tail classes based on the elbow technique [36]. In particular, by utilizing the sample-wise class sizes $\kappa_{i}$ , we first obtain rough estimation of the size of each class set $\eta_{y} \approx |C_{y}|$ . Then, we estimate the maximum size of tail classes $K_{max}$ by finding a point where the class size most abruptly changes. This optimal shift point is found by the elbow technique. Concretely, we first estimate the size of each class $y$ , $\eta_{y} \approx |C_{y}|$ , inductively by + +$$ +\begin{array}{l} \eta_ {1} = \operatorname {r o u n d} \left(\frac {1}{\kappa_ {(1)}} \sum_ {i = 1} ^ {\kappa_ {(1)}} \kappa_ {(i)}\right), \\ \eta_ {y + 1} = \operatorname {r o u n d} \left(\frac {1}{\kappa_ {\left(\eta_ {y} + 1\right)}} \sum_ {i = \eta_ {y} + 1} ^ {\min \left(\kappa_ {\left(\eta_ {y} + 1\right)}, | X |\right)} \kappa_ {(i)}\right), \tag {8} \\ \end{array} +$$ + +where the index $(i)$ of $\kappa_{(i)}$ is sorted in the increasing order $\kappa_{(1)}\leq \dots \leq \kappa_{(|X|)}$ . + +After we acquire the set $\{(y,\eta_y):y\in (1,\ldots ,|\mathcal{Y}|)\}$ through the inductive process above, we define the maximum size of tail classes as the elbow point of the estimated sizes of classes $\eta_{y}$ since at this point the class size increases abruptly. The elbow is computed by linearly connecting $(1,\eta_{1})$ and $(|\mathcal{V}|,\eta_{|\mathcal{V}|})$ , and finding the maximum orthogonal distance onto the line from the set of points $\{(y,\eta_y):y\in (1,\dots |\mathcal{Y}|)\}$ as shown below: + +$$ +K _ {\max } := \eta_ {y _ {\max }} =: e l b o w (\eta_ {1}, \dots , \eta_ {| \mathcal {Y} |}) \tag {9} +$$ + +where + +$$ +y _ {\max } = \underset {y \in \{1, \dots , | \mathcal {Y} | \}} {\operatorname {a r g m a x}} \frac {| a y - \eta_ {y} + b |}{\sqrt {a ^ {2} + 1}} \tag {10} +$$ + +![](images/ba9e9fcdd99af9e088ed6ee4234468f17bb787d7ab9350bc27caf2a2cdf5cdbc.jpg) +Figure 3. (a) Sampling process description of TailedCore (ours) and (b) the illustration of how we use the elbow method [36]. + +and where the line $\eta_y = ay + b$ intersects $(1, \eta_1)$ and $(|\mathcal{Y}|, \eta_{|\mathcal{Y}|})$ with $a = (\eta_{|\mathcal{Y}|} - \eta_1) / (|\mathcal{Y}| - 1)$ and $b = \eta_1 - m$ . Fig. 3 (b) illustrates Eq. (10). We choose the elbow point $(y_{max}, \eta_{y_{max}})$ as the maximum size of tail classes $K_{max}$ . + +The maximum size of tail classes $K_{max}$ allows us to sample the patch features of tail classes exclusively: + +$$ +\mathcal {S} _ {\text {t a i l}} (P) = \left\{\phi_ {i} ^ {(h, w)}: \kappa_ {i} \leq K _ {\max } \right\}. \tag {11} +$$ + +with $\kappa_{i}$ given by Eq. (7). + +# 5.2. Proposed method: TailedCore + +Our anomaly detection model is built upon the memory bank constructed as below. First, we sample the base memory bank as in [20] with noise-discrimination described in Eq. (3). Then, the base memory is augmented with TailSampler $S_{tail}$ (Eq. (11)). Overall, TailedCore's memory bank is defined by $M_{TailedCore} = S_{core}(S_{clean}(P)) \cup S_{core}(S_{tail}(P))$ , where $S_{core}$ is the greedy coreset sampling in Eq. (1), $S_{clean}$ is the noise discrimination process that filters out defect patches based on their outlier scores (Eq. (3)), and the few-shot sampler $S_{tail}$ augments the tail class patches exclusively on the noise-discriminated base memory bank. The whole process is shown in Fig 3. As TailedCore's patch memory bank is both noise-clean and contains the representative features of few-shot classes, it is robust against both class imbalance and normal sample contamination. In inference, the anomaly score is obtained by Eq. (2) as in the conventional memory-based detector. + +# 6. Experiments + +Section 6 consists of dataset description, comparison to the SOTA models, and the detailed ablation study and analysis of TailedCore's components. + +# 6.1. Datasets for unsupervised long-tail noisy anomaly detection + +We devise the long-tail noisy anomaly detection datasets by modifying the class-wise sample distributions of the widely used anomaly detection datasets MVTecAD [3] and VisA [44]. We consider three different types of long-tail distributions. Particularly, datasets with step-shaped tail distribution are devised + +
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDim [9] ICPR2182.4580.9582.0677.4781.2879.1971.5481.7575.63
HVQ [26] NeurIPS2383.4680.2382.9982.0185.5083.5674.1590.1580.55
WinCLIP [19] CVPR2389.3590.1190.3791.6088.2190.3791.8088.2390.37
AnomalyCLIP [43] ICLR2490.9390.9891.4891.8290.8391.4891.2191.9091.48
PatchCore [34] CVPR2293.3387.5989.1892.1971.1883.8386.3670.4880.01
SoftPatch [20] NeurIPS2284.6886.9587.7167.6597.5479.6460.6697.4975.40
TailedCore (ours)96.5595.2496.1295.8295.3495.7193.5495.7794.43
+ +Table 1. Anomaly classification on MVTecAD with image-level AUROC $(\%)$ . We report the mean over 5 random seeds for each measurement. Notations: $C_h / C_t$ : head / tail classes. + +
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2170.7083.3578.6460.6588.9372.4355.9886.7568.80
HVQ [26]NeuRPS2373.4784.0368.2568.2589.3077.0261.5780.4069.42
WinCLIP [19]CVPR2373.2576.9275.4775.9874.7675.4778.8070.8075.47
AnomalyCLIP [43]ICLR2481.9682.4882.0582.2881.7482.0583.2680.3482.05
PatchCore [34]CVPR2286.1185.7385.5983.5367.5176.8579.3368.5674.84
SoftPatch [20]NeuRPS2278.0492.1686.5659.7095.9774.8152.6194.1769.92
TailedCore (ours)87.5593.0690.8585.1695.9189.6482.9794.1187.61
+ +Table 2. Anomaly classification on VisA with image-level AUROC $(\%)$ . The format and evaluation protocol are the same as Tab. 1. + +
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2190.1192.6691.4382.5395.2987.6778.8095.5485.50
HVQ [26]NeuRPS2393.6386.8590.5590.7392.5891.5386.3695.2089.90
WinCLIP [19]CVPR2382.0384.0682.2980.6084.6382.2980.1685.4882.29
AnomalyCLIP [43]ICLR2491.2491.6991.0889.9692.6691.0889.3493.6891.08
PatchCore [34]CVPR2293.5687.9889.9393.5472.0985.1992.0271.3583.75
SoftPatch [20]NeuRPS2292.1993.8393.4180.9896.4987.2470.3496.8980.99
TailedCore (ours)96.0895.0195.2995.5693.2094.7494.1993.7093.99
+ +Table 3. Anomaly segmentation on MVTecAD with pixel-level AUROC (\%). We report the mean over 5 random seeds for each measurement. Notations: $C_h / C_t$ : head / tail classes. + +by making the particular classes to have only a small number of samples; namely $K = 1$ , and $K = 4$ , imposing extreme class imbalance on the datasets. We name these two datasets by 'step $(K = 1)$ ' and 'step $(K = 4)$ '. On the other hand, another realistic type of tail-distribution is considered by making the class distribution to follow the Pareto distribution with the shape parameter 0.6. For the step-tailed distribution datasets, we impose $40\%$ of classes to be head classes, containing the original number of training samples. Thus, in MVTecAD, 6 of 15 classes are head classes. For the Pareto tailed distribution, we regard as few-shot (i.e., tail classes) the classes whose number of samples + +
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2189.0295.1082.8183.9097.3689.5182.5796.5788.40
HVQ [26]NeuIPS2395.2797.6096.7193.8898.3495.7490.5895.5192.63
WinCLIP [19]CVPR2371.9473.9773.1974.6071.2173.1973.8172.3273.19
AnomalyCLIP [43]ICLR2495.6095.4695.5195.5495.4895.5196.1694.6095.51
PatchCore [34]CVPR2296.8487.9991.1395.3962.9681.8894.1165.3082.10
SoftPatch [20]NeuIPS2293.2096.7495.2783.9597.1089.4380.7396.8287.43
TailedCore (ours)97.9897.2597.4896.8097.0296.8996.1297.3996.65
+ +Table 4. Anomaly segmentation on VisA with pixel-level AUROC $(\%)$ . The format and evaluation protocol are the same as Tab. 3. + +is less than 20 [25]. To impose noisy condition on the datasets, the anomaly samples (i.e., the samples with defective regions) are added to the training samples such that the noisy samples constitute $10\%$ of the training set following the overlap protocol in SoftPatch [20] for fair comparison. In all cases, the testing sets of datasets remain the same as the original MVTecAD and VisA. A more detailed configuration is given in Supp. 11. + +# 6.2. Comparison to the state-of-the-art (SOTA) + +We compare TailedCore to the SOTA anomaly detection methods in the unsupervised long-tail noisy settings. We carefully select our baselines. HVQ excels in multi-class learning, PaDiM and SoftPatch are robust against noises as PaDiM is based on Gaussian modeling of the data, while SoftPatch explicitly removes the defect patches by outlier scoring. Using foundation models, WinClip and AnoamlyClip are strong for zero-shot learning. PatchCore are near-perfect in the single-class anomaly detection settings. We show the full experimental results in Supp. 12. All models are trained in the multi-class setting; namely, one model for one dataset with all classes rather than one model per class. The detection performance, however, is measured for each product class under image-level and pixel-level AUROC, and averaged for reporting and comparison. For comparison, we consider models that are specifically robust against noise, strong on multi-class training, and foundation models that can exhibit zero/few-shot learning capability. + +Tabs. 1, 2, 3, and 4 show the superiority of TailedCore across all types of long-tail distributions. As anticipated, all baseline models encounter the tail-versus-noise trade-off, resulting in suboptimal performance either on tail classes with limited samples or on head classes with noisy training samples. These trends become more pronounced in highly unbalanced datasets, where the few-shot classes contain only 1 or 4 samples. + +# 6.3. Ablation study and analysis + +# 6.3.1. Tail (few-shot) sampling + +We evaluate how accurately our proposed algorithm $S_{tail}$ can sample tail (few-shot) class samples exclusively without including noise samples. First of all, we measure the ratio of mis-included tail class samples TailSampler which is the ratio of not sampled tail class samples to total tail class samples. A single mis-inclusion can vitally damage anomaly detection performance especially in case of the one-shot setting $K = 1$ as the model wouldn't be able to utilize the only sample of the class. + +Additionally, the exclusive aspect of few-shot sampling is crucial; the few-shot sampled set shouldn't include anomaly and head class samples as possible. Included anomaly samples would harm the head class performance and head class sampled by few-shot sampler might cause class imbalance issue as well as the possibility of sampling anomaly from head classes. Hence, we measure the ratios of included anomalies and head-class samples along with AUROC which is measured on the outlier scores derived by a given algorithm and ground truth binary labels that indicate whether a given sample is tail class or head class. For TailSampler, the outlier score of each train sample is given as the negative of predicted class size. The included anomaly ratio is defined as the ratio of number of anomaly samples to the number of whole anomaly samples. The ratio of included head-class sample is defined as the number of included head class to the number of included tail class samples. + +For extensive comparison, we compare TailSampler with widely used clustering algorithms [2, 13, 16], outlier detectors [1, 5, 24, 35], and density estimators [10, 30]. These methods provide clustering size and/or outlier score/prediction. In cases when only outlier scores are available, we apply the elbow method to induce deterministic prediction of whether a given sample is few-shot or not. We use default settings provided by Sklearn [31] and PyOD [42]. For all methods, we utilize the embedding vectors of WideResNet-50 [40]. Experiments with different architectures and details are provided in Supp. 10. + +The combined ratio, which is the sum of missing tail samples and included anomaly samples, indicates that TailSampler is most effective for exclusively sampling few-shot class samples. The most critical factor is to not miss few-shot class samples. On this ratio, TailSampler achieves the lowest rate except for the algorithms that trivially sample all few-shot class samples. We find that most of the classical algorithms for clustering and outlier detection are not usable as they either miss too many few-shot samples or includes a large number of anomaly samples. We note that the class size predictor in our few-shot samples resembles DBSCAN and Affinity Propagation, respectively, in terms of using neighborhoods and voting mechanism. None of these methods however have instance-specific parameters that are adaptively adjusted across different samples. + +# 6.3.2. Class size prediction + +We measure how accurately our class size predictor estimates the class size of each sample by measuring the class size prediction error. It is measured for each train sample and weight-averaged through the whole train samples: + +$$ +e r r o r = \frac {1}{| \mathcal {X} |} w _ {i} \sum_ {i = 1} ^ {| \mathcal {X} |} \left| \kappa_ {i} - \left| C _ {k (i)} \right| \right| \tag {12} +$$ + +where $k(i)$ indicates the index such that $x_{i} \in C_{k(i)}$ (namely, $C_{k(i)}$ is the class of $x_{i}$ ). The weight is chosen by $w_{i} = 1 / |C_{k(i)}|$ , which is to give more penalty on the few-shot class samples more than the head class samples. + +![](images/152311089b96de8555f960d6ca9fed416477e5d0357d3e8fc144abc2b7317b43.jpg) +(a) + +![](images/e729b819c2118e56ba16c1a0a1940bd4dfc07c914c7583bb628dfb4aae8f6df4.jpg) +(c) + +![](images/ebb0db69dd702d850c868a7a4c91b7993907b663f3bdd5a94387380c61cb7050.jpg) +(e) + +![](images/411a6a35c177bd2d5d4c88cf36b7b0b5e5b712fa13ab9f0d988198bcbf586c88.jpg) +(b) + +![](images/b774650a828046de9a5043b7a620221b1bae70aa92a70778ba9170a06713eb24.jpg) +(d) + +![](images/48d33364309a3fa03f7362eb1046fb4206ff6652013af25084b37575b0dff952.jpg) +(f) +Figure 4. The plot of noise ratio of train data versus anomaly classification (image-level AUROC) and segmentation (pixel-level AUROC) performance. In all cases, TailedCore outperforms both PatchCore and SoftPatch. Dataset is MVTecAD with different tail distributions. The lines show the mean and shades show standard deviation of multiple runs of experiments with different seeds. + +Tab. 5 shows that most of the clustering based methods vastly fail on this task, and the estimation by our class size predictor outperforms them significantly. We believe this is partly due to suboptimal hyperparameters used for these algorithms. Due to their classical nature, however, it is unclear how to tune them without the availability of the validation set in our scenario. + +# 6.3.3. Abalation study on class size predictor + +We conduct detailed ablation study on the proposed class size prediction in Eq. (7). One hyperparameter used in TailSampler $S_{tail}$ of TailedCore is the percentiling proportion $p$ to consider only the majority of nearby samples within the neighborhood. We analyze how this parameter can affect the anomaly detection performance of TailedCore. Tab. 5 indicates that using the nearest $85\%$ embeddings (i.e., $p = 85\%$ ) is the best choice. Using all of the available neighborhoods (i.e., $p = 100\%$ ) would include different class samples in the neighborhood, deteriorating class estimation. On the other hand, counting only with the half of samples (i.e., $p = 50\%$ ) in the neighborhood may lose too much information. + +Another component in the class size predictor that requires analysis is its voting mechanism used in Eq. (7). One may + +
methodcombined ratio of missing \( {C}_{t} \) & included AS (↓)ratio of missing \( {C}_{t} \) samples (↓)ratio of included AS (↓)ratio of included \( {C}_{h} \) samples (↓)class size prediction error (↓)AUROC (%) (↑)
LOF\( {89.58} \pm {44.42} \)\( {31.22} \pm {34.88} \)\( {58.36} \pm {9.54} \)\( {92.50} \pm {7.87} \)\( {1.00} \pm {0.00} \)\( {77.08} \pm {25.86} \)
IF\( {18.15} \pm {15.91} \)\( {7.25} \pm {10.05} \)\( {10.90} \pm {5.86} \)\( {69.02} \pm {25.96} \)\( {1.00} \pm {0.00} \)\( {98.96} \pm {1.27} \)
OCSVM\( {49.95} \pm {9.76} \)\( {0.78} \pm {2.19} \)\( {49.17} \pm {7.57} \)\( {95.06} \pm {6.08} \)\( {1.00} \pm {0.00} \)\( {96.72} \pm {3.59} \)
DBSCAN\( {35.36} \pm {49.55} \)\( {30.64} \pm {44.38} \)\( {4.72} \pm {5.17} \)\( {25.57} \pm {27.24} \)\( {0.35} \pm {0.44} \)\( {84.46} \pm {22.18} \)
KMeans\( {94.67} \pm {44.95} \)\( {47.22} \pm {35.49} \)\( {47.45} \pm {9.46} \)\( {97.40} \pm {3.99} \)\( {4.27} \pm {1.40} \)\( {52.58} \pm {20.46} \)
GMM\( {100.00} \pm {0.00} \)\( \mathbf{{0.00}} \pm \mathbf{{0.00}} \)\( {100.00} \pm {0.00} \)\( {97.58} \pm {2.86} \)\( {12.50} \pm {1.53} \)\( {50.00} \pm {0.00} \)
KDE\( {29.81} \pm {24.15} \)\( {2.14} \pm {6.26} \)\( {27.67} \pm {17.89} \)\( {87.47} \pm {10.35} \)\( {1.00} \pm {0.00} \)\( {97.98} \pm {2.19} \)
AP\( {27.81} \pm {32.59} \)\( {12.98} \pm {24.01} \)\( {14.83} \pm {8.58} \)\( {77.45} \pm {28.21} \)\( {0.70} \pm {0.22} \)\( {93.03} \pm {12.8} \)
KNN\( {43.49} \pm {31.25} \)\( {14.46} \pm {24.07} \)\( {29.03} \pm {7.18} \)\( {83.12} \pm {18.28} \)\( {1.00} \pm {0.00} \)\( {94.51} \pm {10.34} \)
TailSampler (ours)\( {5.10} \pm {5.07} \)\( {0.12} \pm {0.65} \)\( {4.98} \pm {4.42} \)\( {31.96} \pm {26.05} \)\( \mathbf{{0.19}} \pm \mathbf{{0.05}} \)\( {99.78} \pm {0.30} \)
Ablation on the class size predictor:
TailSampler (p=100)\( {8.48} \pm {10.88} \)\( {1.07} \pm {5.23} \)\( {7.41} \pm {5.65} \)\( {42.64} \pm {31.86} \)\( {0.51} \pm {0.04} \)\( {99.76} \pm {0.29} \)
TailSampler (p=50)\( {7.37} \pm {10.35} \)\( {3.46} \pm {6.80} \)\( {3.91} \pm {3.55} \)\( {27.90} \pm {27.36} \)\( {0.12} \pm {0.12} \)\( {97.36} \pm {5.65} \)
TailSampler (top-1)\( {14.47} \pm {7.99} \)\( {0.20} \pm {0.78} \)\( {14.27} \pm {7.21} \)\( {73.97} \pm {27.95} \)\( {0.85} \pm {0.04} \)\( {99.73} \pm {0.34} \)
TailSampler (average)\( {6.48} \pm {10.11} \)\( {3.65} \pm {6.99} \)\( {2.83} \pm {3.12} \)\( {22.16} \pm {24.33} \)\( {0.21} \pm {0.07} \)\( {99.35} \pm {2.59} \)
+ +Table 5. The analysis on the proposed TailSampler and its class size predictor. The results are obtained from both MVTecAD and VisA and averaged across all different types of tail distributions and random seeds (overall 60 different seeds) with the standard deviations indicated on the right. The best and second best performances are marked in bold and underline, respectively. Acronyms/notations: IF: Isolation Forest; AP: Affinity Propagation; AS: anomaly samples; $C_h / C_t$ : head / tail classes. + +
tasktail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
anomaly classificationw/o FSA82.2291.2288.1360.0197.6075.0566.2897.8578.94
w/ FSA96.5595.2496.1393.5495.7894.4495.8295.3495.72
anomaly segmentationw/o FSA91.6094.9493.5468.9996.8580.1479.3996.3586.23
w/ FSA96.0995.0195.3094.1993.7194.0095.5693.2094.75
+ +Table 6. Abaltion on the few-shot augmentation (FSA) of TailedCore. The image-level and pixel-level AUROC metrics $(\%)$ are reported on the MVTecAD dataset with different types of tail distributions for anomaly classification and segmentation tasks, respectively. Notations: $C_h / C_t$ : head / tail classes. + +simply use the nearest embedding's neighborhood size instead, or the class size of $x_{i}$ can be estimated by averaging instead of majority vote. As indicated in Tab. 5, both the nearest embedding (i.e., top-1) and averaging are inferior to majority voting. The former is inevitably less robust, while averaging can still be sensitive to the outlier values of neighborhood size. + +# 6.3.4. Ablation on TailedCore + +We analyze the impact of augmentation of few-shot class patches by TailSampler in TailedCore. The TailedCore model without these augmented patches has the same memory bank as SoftPatch, but its inference scoring is simpler. Tab. 6 indicates that without the augmented patches from TailSampler, the memory-based model cannot detect anomalies in few-shot class samples. This shows the necessity of tail class augmentation by TailSampler. We note some performance decrease in the head class detection performance since few-shot augmentation may involve few anomaly samples from the head classes that behave like few-shot class instances. This issue intrinsically is due to the weakness of encoder embedding, and the proposed TailSampler based on such embeddings can induce incorrectly sampled few-shot class instances, which we discuss in Sec. 7. + +![](images/9717bda343f63692eaf05562fe64c0ffcd246b969c9c5565d6bb41162b2cfb4e.jpg) +(a) + +![](images/ce20f1f5e2917667e94b2ca569c2c2bd8cae69cde8903b1a9d60b5645b29507b.jpg) +(b) + +![](images/08a3066e9f32c6a13df8d9ae59142a883f26650743126ca83c2a02d0c0ba89a4.jpg) +(c) + +![](images/d2acdf6056a1babed87312a360c29e6b8814b976420236aa54344350c4396e70.jpg) +(d) + +![](images/d27c9a2cc6be928f5abb9e05bb2f50a32ab6c24ff2e678ad5e3d94b012f00912.jpg) +(e) + +![](images/f4883daa739be54c80bc4e97a87f1d4b85029f91b9b8ba1be30c94f110de2885.jpg) +(f) + +![](images/bcb0e23af33ba75dc74d66826c9a39155bcd6a5b8a0faf7871478dd736a261bf.jpg) +(g) +Figure 5. Classification accuracy between tail classes and noisy samples versus metrics relevant to class size prediction and few-shot sampling on MVtecAD and VisA datasets with step-like tailed distribution and $K = 4$ . The correlation is strong for (a) mis-sampling ratio, (b) ratio of missing few-shot samples, (e) class size prediction error, and (f) AUROC for predicting whether a sample is few-shot or not. We show that improving the discriminative aspect of network's embeddings improves few-shot sampling and prediction error of class size predictor, which in turn improves (g) anomaly classification (image-level AUROC) and (h) anomaly segmentation (pixel-level AUROC) performance of TailedCore. + +![](images/9ad25f98f81dfe8375ee1b7274baf1e43b4a0313173ca8c87b8ecc06af7045fc.jpg) +(h) + +# 6.4. Noise ratio + +We analyze the impact of noise ratio. Training data with a greater number of defect samples increase the chance of sampling noise instances for the few-shot sampler. We analyze this trend in Fig. 4, but the impact is not severe. In all cases, TailedCore either outperforms or is on par with both PatchCore and SoftPatch. + +# 6.5. Relation between the discriminative aspect of embedding and TailedCore + +TailedCore utilizes the embedding features $e_i$ of training samples $x_i$ for few-shot sampling and augmentation thereof. Therefore, TailedCore's detection performance highly relies on the discriminative aspect of the embedding feature. Poorly designed embedding features will not discriminate different class samples in the embedding space. In this case, TailSampler cannot sample tail class samples and it may overly many head class samples. TailedCore's memory bank obtained so will neither be class-balanced nor noise-free, making the inference of TailedCore poor. + +In contrast, a highly discriminative embedding features will make TailedCore's class size prediction accurate, thereby enabling it to sample the few-shot class samples exclusively without aggregating any anomaly and head class samples. In which case, TailedCore is both noise-free and class-balanced. + +We test this behavior by Oracle labeling and training the embedding features with class labels. Fig. 5 shows that improving discriminative quality of embedding (measured by its classification accuracy on noise and tail class samples) accordingly improves the relevant metrics. Particularly, more discriminative embeddings make more accurate class size + +prediction, improving TailSampler along with its anomaly classification and segmentation performances. + +# 7. Limitations + +The class size prediction in TailSampler can fail if the reflective-symmetric assumption on the inter-class and intra-class similarities breaks down. Particularly, this case can occur if the utilized encoder embedding is either poor or not aligned with the label space of classes. In some cases, the geometric aspect of defect samples highly resembles that of few-shot class instances in the embedding space. We believe that this is due to imperfect nature of encoder embedding, which is often more sensitive to spurious features than the notion of class and object-centric aspect [29]. This issue can be improved by enhancing the encoder network. + +# 8. Conclusion + +In this study, we introduced TailedCore, a novel memory-based anomaly detection model designed to effectively address the challenges of unsupervised anomaly detection in noisy and long-tail class distributions. By developing a unique class size predictor and a tailored memory bank, we successfully navigated the tail-versus-noise dilemma, enhancing the model's performance in identifying anomalies within a contaminated and imbalanced dataset. Our comprehensive evaluations demonstrated TailedCore's superior ability to mitigate noise contamination and class imbalance, underscoring its significance in advancing anomaly detection technologies. This work not only showcases a significant stride in handling real-world anomaly detection scenarios but also lays a robust foundation for future research in this critical field. + +Acknowledgments Kuan-Chuan Peng was exclusively supported by Mitsubishi Electric Research Laboratories. This work was also supported by the Technology Innovation Program (20023364) funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea), NSF grant 2038493, ONR grant N00014-21-1-2431, NIH grant R01CA240771 from NCI, and U.S. Department of Homeland Security grant 22STESE00001-03-02. The views and conclusions contained in this document are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the U.S. Department of Homeland Security. + +# References + +[1] Fabrizio Angiulli and Clara Pizzuti. Fast outlier detection in high dimensional spaces. In European conference on principles of data mining and knowledge discovery, pages 15-27. Springer, 2002. 3, 6 +[2] David Arthur, Sergei Vassilvitskii, et al. k-means++: The advantages of careful seeding. In Soda, pages 1027-1035, 2007. 2, 6 +[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 5 +[4] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4):947-969, 2022. 3 +[5] Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and Jörg Sander. Lof: identifying density-based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pages 93-104, 2000. 3, 6 +[6] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1 +[7] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 2 +[8] Zhuohang Chen, Jinglong Chen, Yong Feng, Shen Liu, Tianci Zhang, Kaiyu Zhang, and Wenrong Xiao. Imbalance fault diagnosis under long-tailed distribution: Challenges, solutions and prospects. Knowledge-Based Systems, 258:110008, 2022. 1 +[9] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 2, 5, 6 +[10] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 6 +[11] Jan Diers and Christian Pigorsch. A survey of methods for automated quality control based on images. International Journal of Computer Vision, pages 1-29, 2023. 1 + +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[13] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 2, 6 +[14] Chaowei Fang, Lechao Cheng, Huiyan Qi, and Dingwen Zhang. Combating noisy labels in long-tailed image classification. arXiv preprint arXiv:2209.00273, 2022. 2 +[15] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 2 +[16] Brendan J Frey and Delbert Dueck. Clustering by passing messages between data points. science, 315(5814):972-976, 2007. 2, 6 +[17] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. arXiv preprint arXiv:2308.15366, 2023. 2 +[18] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 2 +[19] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero-/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2, 5, 6 +[20] Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai Wu, Yong Liu, Chengjie Wang, and Feng Zheng. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35:15433-15445, 2022. 1, 2, 3, 5, 6 +[21] Justin M Johnson and Taghi M Khoshgoftaar. Survey on deep learning with class imbalance. Journal of Big Data, 6(1):1-54, 2019. 1 +[22] Yujin Lee, Harin Lim, Seoyoon Jang, and Hyunsoo Yoon. Uniformality: Towards task-agnostic unified framework for visual anomaly detection. Available at SSRN 4586132. 2 +[23] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. arXiv preprint arXiv:2401.16753, 2024. 2 +[24] Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 eighth ieee international conference on data mining, pages 413-422. IEEE, 2008. 3, 6 +[25] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 1, 6 +[26] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems, 36, 2023. 2, 5, 6 + +[27] Yang Lu, Yiliang Zhang, Bo Han, Yiu-ming Cheung, and Hanzi Wang. Label-noise learning with intrinsically long-tailed data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1369-1378, 2023. 2 +[28] Declan McIntosh and Alexandra Branzan Albu. Inter-realization channels: Unsupervised anomaly detection beyond one-class classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6285-6295, 2023. 2 +[29] Yifei Ming, Hang Yin, and Yixuan Li. On the impact of spurious correlation for out-of-distribution detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10051-10059, 2022. 8 +[30] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33 (3):1065-1076, 1962. 6 +[31] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011. 6 +[32] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[34] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 2, 3, 5, 6 +[35] Bernhard Schölkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. Advances in neural information processing systems, 12, 1999. 3, 6 +[36] Robert L Thorndike. Who belongs in the family? Psychometrika, 18(4):267-276, 1953. 4, 5 +[37] Tong Wei, Jiang-Xin Shi, Wei-Wei Tu, and Yu-Feng Li. Robust long-tailed learning under label noise. arXiv preprint arXiv:2108.11569, 2021. 2 +[38] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Feng Zheng, and Yaochu Jin. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. arXiv preprint arXiv:2301.12082, 2023. 2 +[39] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 1, 2 +[40] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6 +[41] Manyi Zhang, Xuyang Zhao, Jun Yao, Chun Yuan, and Weiran Huang. When noisy labels meet long tail dilemmas: A representation calibration method. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15890-15900, 2023. 2 + +[42] Yue Zhao, Zain Nasrullah, and Zheng Li. Pyod: A python toolbox for scalable outlier detection. Journal of Machine Learning Research, 20(96):1-7, 2019. 6 +[43] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In International Conference on Learning Representations, 2024. 2, 5, 6 +[44] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. arXiv preprint arXiv:2207.14315, 2022. 5 \ No newline at end of file diff --git a/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/images.zip b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3e2e81e4359503d99d53f971e6d6b5585ec762e3 --- /dev/null +++ b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9faa1d26f273f8be2d43a1dcb38b88061358a29e0235a80e6d9b07a322e335e2 +size 547183 diff --git a/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/layout.json b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8c5e78f1fd950592edd586d2b154e317b87deade --- /dev/null +++ b/2025/TailedCore_ Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection/layout.json @@ -0,0 +1,10541 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "text", + "content": "TailedCore: Few-Shot Sampling for Unsupervised Long-Tail Noisy Anomaly Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 473, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 473, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 473, + 172 + ], + "type": "text", + "content": "Yoon Gyo Jung\\*1 Jaewoo Park\\*2 Jaeho Yoon\\*3 Kuan-Chuan Peng4 Wonchul " + }, + { + "bbox": [ + 138, + 142, + 473, + 172 + ], + "type": "inline_equation", + "content": "\\mathrm{Kim}^2" + }, + { + "bbox": [ + 138, + 142, + 473, + 172 + ], + "type": "text", + "content": " Andrew Beng Jin Teoh Octavia Camps" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "spans": [ + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "type": "text", + "content": "Northeastern University " + }, + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "type": "text", + "content": "AiV Co. " + }, + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "type": "text", + "content": "Yonsei University " + }, + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 156, + 178, + 453, + 216 + ], + "type": "text", + "content": "Mitsubishi Electric Research Laboratories" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 262, + 198, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 262, + 198, + 275 + ], + "spans": [ + { + "bbox": [ + 153, + 262, + 198, + 275 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 289, + 297, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 297, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 297, + 526 + ], + "type": "text", + "content": "We aim to solve unsupervised anomaly detection in a practical challenging environment where the normal dataset is both contaminated with defective regions and its product class distribution is tailed but unknown. We observe that existing models suffer from tail-versus-noise trade-off where if a model is robust against pixel noise, then its performance deteriorates on tail class samples, and vice versa. To mitigate the issue, we handle the tail class and noise samples independently. To this end, we propose TailSampler, a novel class size predictor that estimates the class cardinality of samples based on a symmetric assumption on the class-wise distribution of embedding similarities. TailSampler can be utilized to sample the tail class samples exclusively, allowing to handle them separately. Based on these facets, we build a memory-based anomaly detection model TailedCore, whose memory both well captures tail class information and is noise-robust. We extensively validate the effectiveness of TailedCore on the unsupervised long-tail noisy anomaly detection setting, and show that TailedCore outperforms the state-of-the-art in most settings. Code is available in TailedCore" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 132, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 132, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 132, + 564 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 573, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 297, + 693 + ], + "type": "text", + "content": "In the complex landscape of anomaly detection, the challenge often lies in navigating practical scenarios that feature a diverse and imbalanced data distribution [6, 8, 11, 21]. We delve into a challenging yet highly realistic scenario in anomaly detection, where the training dataset, comprising multiple product classes [39], is beset by two significant complications: contamination with noise [20] and presence of tailed class distributions [25]. In this context, 'tail' (few-shot) classes have limited few-shot samples, unlike 'head' (many-shot) classes which are data-rich but might contain samples with defects. Crucially, in this unsupervised set-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 251, + 432, + 352 + ], + "blocks": [ + { + "bbox": [ + 317, + 251, + 432, + 352 + ], + "lines": [ + { + "bbox": [ + 317, + 251, + 432, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 432, + 352 + ], + "type": "image", + "image_path": "08f2b2036c81a6d05697400ceebc490f94245dfd6bab865ac9ff521699b12b5f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 353, + 356, + 358 + ], + "lines": [ + { + "bbox": [ + 326, + 353, + 356, + 358 + ], + "spans": [ + { + "bbox": [ + 326, + 353, + 356, + 358 + ], + "type": "text", + "content": "SimpleNet" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 353, + 390, + 358 + ], + "lines": [ + { + "bbox": [ + 366, + 353, + 390, + 358 + ], + "spans": [ + { + "bbox": [ + 366, + 353, + 390, + 358 + ], + "type": "text", + "content": "WinCLIP" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 353, + 414, + 358 + ], + "lines": [ + { + "bbox": [ + 391, + 353, + 414, + 358 + ], + "spans": [ + { + "bbox": [ + 391, + 353, + 414, + 358 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 353, + 434, + 358 + ], + "lines": [ + { + "bbox": [ + 416, + 353, + 434, + 358 + ], + "spans": [ + { + "bbox": [ + 416, + 353, + 434, + 358 + ], + "type": "text", + "content": "IVQ" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 353, + 454, + 358 + ], + "lines": [ + { + "bbox": [ + 435, + 353, + 454, + 358 + ], + "spans": [ + { + "bbox": [ + 435, + 353, + 454, + 358 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 455, + 353, + 472, + 358 + ], + "lines": [ + { + "bbox": [ + 455, + 353, + 472, + 358 + ], + "spans": [ + { + "bbox": [ + 455, + 353, + 472, + 358 + ], + "type": "text", + "content": "bftPatch" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 353, + 492, + 358 + ], + "lines": [ + { + "bbox": [ + 473, + 353, + 492, + 358 + ], + "spans": [ + { + "bbox": [ + 473, + 353, + 492, + 358 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 353, + 511, + 358 + ], + "lines": [ + { + "bbox": [ + 493, + 353, + 511, + 358 + ], + "spans": [ + { + "bbox": [ + 493, + 353, + 511, + 358 + ], + "type": "text", + "content": "。" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 512, + 353, + 528, + 358 + ], + "lines": [ + { + "bbox": [ + 512, + 353, + 528, + 358 + ], + "spans": [ + { + "bbox": [ + 512, + 353, + 528, + 358 + ], + "type": "text", + "content": "PaDiM" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 529, + 353, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 529, + 353, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 529, + 353, + 545, + 358 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 368, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 412 + ], + "type": "text", + "content": "Figure 1. Tail class (x-axis) and noisy head class (y-axis) performance comparison. The tail-versus-noise trade-off is shown across memory-based anomaly detection models (circles), and is more indicative in anomaly classification task evaluated by image-level AUROC (left)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 424, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 424, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 312, + 424, + 555, + 484 + ], + "type": "text", + "content": "ting, the class information of these products remains unknown to the model trainer. This setting mirrors a common predicament in large-scale industrial applications, where companies are required to manage vast production lines encompassing a wide array of product types, each with varying production rates." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 485, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 556, + 640 + ], + "type": "text", + "content": "A critical issue arises in this scenario, which we term the \"tail-versus-noise dilemma.\" The dilemma underscores a fundamental trade-off where models robust to class imbalance overfit to defect noises, likely misclassifying many defects as normal. In contrast, existing models that are robust against defect noise tend to underfit, missing vital details in tail classes. Fig. 1 shows this tail-versus-noise trade-off which is prevalent through existing anomaly detection algorithms. This intrinsic characteristic comes from the fact that few-shot class features are equivalent to the noisy abnormal features in terms of their statistical occurrence. Consequently, existing models find it difficult to strike a balance between avoiding overfitting to noise and effectively capturing the nuances of less represented classes." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 642, + 556, + 713 + ], + "type": "text", + "content": "To resolve this challenge, we propose to handle tail classes independently from head classes by exclusively sampling them. To this end, we devise a novel class size predictor TailSampler that estimates the cardinality of the class of any training sample. For accurate estimation, TailSampler estimates the class sizes based on a reflective symmetry between inter-class" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 702, + 191, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 191, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 191, + 713 + ], + "type": "text", + "content": "* equal contribution † corresponding author" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25539" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "and intra-class similarity distributions in the embedding space. The estimated class sizes of the training samples then enable to sample tail class features exclusively from the training set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 295, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 295, + 204 + ], + "type": "text", + "content": "By utilizing the aggregated few-shot class instances, we devise a memory-based anomaly detector TailedCore, which is built upon the noise-discriminated model by augmenting the tail class patch features sampled with TailSampler. As the memory of TailedCore is both clean from noisy defect patches and preserves the few-shot class information, it exhibits robustness against both noise contamination and class imbalance of the training set." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 205, + 242, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 205, + 242, + 215 + ], + "spans": [ + { + "bbox": [ + 67, + 205, + 242, + 215 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 217, + 295, + 420 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 217, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 264 + ], + "type": "text", + "content": "- We explore a practical and challenging anomaly detection scenario, where a single detection model is trained on multi-class product training samples that involve noise contamination and whose class distribution is long-tailed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 372 + ], + "type": "text", + "content": "- We propose a memory-based anomaly detector TailedCore whose patch memory bank is both noise-free and augmented with the representative patches of few-shot class instances. To build TailedCore, we introduce TailSampler that can aggregate few-shot class instances exclusively. TailSampler is based on a novel class size predictor that estimates the class size of training samples based on the reflective symmetry between inter-class and intra-class class-wise embedding similarity distributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 373, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 420 + ], + "type": "text", + "content": "- We conduct extensive analyses and comparative studies on both TailedCore and few-shot sampler along with the proposed class size predictor on the proposed unsupervised long-tail noisy anomaly detection benchmark." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 421, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 456 + ], + "type": "text", + "content": "The codes for TailedCore along with the data generation will be open sourced. We use the term 'few-shot' and 'tail' interchangeably to denote the class with a small number of samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 468, + 140, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 140, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 140, + 481 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 489, + 264, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 264, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 264, + 502 + ], + "type": "text", + "content": "2.1. Anomaly detection in extreme scenarios" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 506, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 590 + ], + "type": "text", + "content": "One vitally important aspect of anomaly detection [9, 34] is the environment where these models are tested. This exploration has led to the identification of several challenging scenarios, including multi-class settings with training data spanning multiple product types [26, 39], zero-shot and few-shot settings with minimal training images [15, 19, 23, 38, 43], and noisy settings featuring uncleaned training data with defective images [7, 20, 28]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Multi-class anomaly detection In contrast to single-product settings, anomaly detection models trained on multi-class datasets often encounter \"shortcut learning\" [39], leading to low anomaly scores for normal images. This issue is particularly pronounced in reconstruction-based models, which tend to reconstruct all samples, regardless of their normality or abnormality. To resolve it, UniAD [39] employs neighbor masked attention within a transformer model, focusing attention on local parts to prevent shortcut learning. HVQ [26] utilizes vector quantization to discretize the latent space, restricting the network's recon" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "striction fidelity on normal data distributions. UniFormaly [22] increases sensitivity to irrelevant patterns by masking less relevant patches and employing top-k matchings with a pretrained foundation model. These multi-class anomaly detection methods do not consider either long-tail classes or contaminated training data setups, and do not work well in the combined \"noisy long-tail\" setup, which is shown in our experimental results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 158, + 556, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 556, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 556, + 313 + ], + "type": "text", + "content": "Zero/few-shot anomaly detection The inception of few-shot learning benchmarks in anomaly detection marked a significant advancement, with methods leveraging template matching [18] and graph representations [38] for minimal training data scenarios. AnomalyCLIP [43], WinCLIP [19], and AnomalyGPT [17], on the other hand, detect anomalies in the few-shot classes based on the acquired knowledge of pretrained foundation models such as CLIP [33] and GPT [32]. However, these methods often require sophisticated prompt engineering or delicate prompt-tuning, and have yet to achieve optimal precision and recall in the single-class training setting with a sufficient amount of training samples. In contrast, our proposed TailedCore needs none of these process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 316, + 556, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 423 + ], + "type": "text", + "content": "Noise anomaly detection Dealing with training datasets contaminated with defective samples presents another extreme learning scenario. SoftPatch [20] and InReaCh [28] focus on purifying memory coresets through outlier detection algorithms and distance association filtering, respectively. IGD [7] adopts a robust Gaussian modeling approach, addressing both contaminated training data and small dataset challenges. Yet, these methods do not fully consider scenarios involving both data contamination and class imbalance, unlike our proposed TailedCore." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 431, + 449, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 449, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 449, + 443 + ], + "type": "text", + "content": "2.2. Long-tail noisy learning" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "content": "In the realm of image classification, several studies have focused on learning scenarios involving training sets with label contamination and class imbalance. To address these challenges, the existing literature relies on the robustness and unsupervised aspects of self-supervised learning [14, 27, 37, 41]. The learned embeddings of self-supervised frameworks are less impacted by label contamination and class imbalances. However, these works differ from ours in that they deal with label-level contamination rather than pixel-level defects. Furthermore, these methods require explicit supervised labels to remediate the contamination. In contrast, our work targets noise in pixels without any class label information." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 600, + 459, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 459, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 459, + 613 + ], + "type": "text", + "content": "2.3. Few-shot/outlier sampling" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": "Both classical clustering methods and outlier detection algorithms can be applied to our task, as they facilitate the identification of few-shot tail class samples, either indirectly or directly. Unsupervised clustering algorithms [2, 13, 16] can estimate the cluster sizes but struggle to identify small sample clusters in the imbalanced data. DBSCAN [13], on the other hand, captures both clusters and outliers. This however can be problematic when attempting to identify clusters with only one" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25540" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": "sample (1-shot) as it may be confused between few-shot samples and outliers. Unlike clustering methods, outlier detection algorithms [1, 5, 24, 35] are adept at identifying 1-shot samples which inherently resemble outliers. However, their effectiveness diminishes when a few-shot class contains more than one sample, as these no longer fit the typical outlier profile, potentially leading to incorrect identification of few-shot class samples. Both algorithm types, due to their inherent design, struggle to accurately predict class/cluster sizes in datasets that are both imbalanced and contaminated. Our proposed class size predictor specifically addresses this issue, aiming to reliably and accurately identify few-shot class samples and enable their exclusive sampling. Ultimately, the identified few-shot class samples are augmented to the noise-discriminated patch set, constituting the memory-based anomaly detection model TailedCore, which is robust against both class imbalance and training set noise." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 273, + 132, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 132, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 132, + 286 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 237, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 237, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 237, + 306 + ], + "type": "text", + "content": "3.1. Memory-based anomaly detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "text", + "content": "To train a memory-based anomaly detector, a feature extractor " + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "text", + "content": " (e.g., WideResNet) extracts a feature map " + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\phi_i \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "text", + "content": " from each training input image " + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "text", + "content": ". Then, a coreset selection algorithm " + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "inline_equation", + "content": "S_{core}" + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "text", + "content": " is applied to the set " + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "inline_equation", + "content": "P = \\{\\phi_i^{(h,w)}\\}" + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "text", + "content": " of patch features " + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\phi_i^{(h,w)}" + }, + { + "bbox": [ + 55, + 310, + 296, + 374 + ], + "type": "text", + "content": ", resulting in a memory bank" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 380, + 296, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 380, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 117, + 380, + 296, + 394 + ], + "type": "interline_equation", + "content": "M _ {\\text {P a t c h C o r e}} := \\mathcal {S} _ {\\text {c o r e}} (P) \\subset P \\tag {1}", + "image_path": "4fd7875eafdaf7e451eb6fc63d5605960585b8e5a0d286c3a6297b3b6a91d553.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "text", + "content": "that contains representative patch features of " + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "text", + "content": ". For instance, PatchCore opts for greedy sampling strategy to obtain the coreset. In the inference stage, the anomaly score " + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "inline_equation", + "content": "s^{(h,w)}" + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "text", + "content": " of the testing image " + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "text", + "content": " is computed for each patch feature " + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\phi^{(h,w)}" + }, + { + "bbox": [ + 55, + 399, + 296, + 459 + ], + "type": "text", + "content": " by measuring its distance to the memory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 463, + 296, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 463, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 115, + 463, + 296, + 478 + ], + "type": "interline_equation", + "content": "\\operatorname {s c o r e} ^ {(h, w)} = \\operatorname {d i s t} \\left(\\phi^ {(h, w)}, M\\right), \\tag {2}", + "image_path": "386078b1b5d209dadc4033b24d01eabd5ced09b3350d73a3ecd322de0cd98f6e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "dist" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": " is the nearest neighbor distance in [34]. The pixel-level anomaly score is computed by up-scaling the score map " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "score = [score^{(h,w)}] \\in \\mathbb{R}^{H,W}" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": ", and its image-level anomaly score is defined as its pixel maximum " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\max_{h,w} score^{(h,w)}" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 538, + 229, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 229, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 229, + 551 + ], + "type": "text", + "content": "3.2. Noise discrimination of memory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "text", + "content": "The training set " + }, + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "inline_equation", + "content": "X = \\{x_{i}\\}" + }, + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "text", + "content": " may be contaminated with defect samples. The naively attained memory bank then contains abnormal patch features, and assigns a low anomaly score on defect images, resulting in poor performance. To resolve this, [20] applies a noise discrimination algorithm " + }, + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "inline_equation", + "content": "S_{clean}" + }, + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "text", + "content": " on the patch feature set " + }, + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "text", + "content": " prior to coreset sampling:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 84, + 632, + 296, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 632, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 84, + 632, + 296, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {\\text {c l e a n}} (P) = \\left\\{\\phi_ {i} ^ {(h, w)}: \\text {o u t l i e r} \\left(\\phi_ {i} ^ {(h, w)}\\right) < \\tau_ {0. 1 5} \\right\\}, \\tag {3}", + "image_path": "fa0b5cae5cd41ea8ba12e7e84e8f820aa866339d9bba9c671ea9934a370821d0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "aggregating only the clean normal patches with low outlier scores. SoftPatch [20] exploits local outlier factor (LOF) for outlier scoring, and selects the threshold " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{0.15}" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " to remove only " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " of whole patches. In inference, SoftPatch utilizes the memory set " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "M_{SoftPatch} = S_{core}(S_{clean}(P))" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 434, + 159 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 434, + 159 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 434, + 159 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 434, + 159 + ], + "type": "image", + "image_path": "06b4eb98e12f75e6d68ca17b8960d9f9ea0cbb1c919b0e1cd543fb93e8ec03a3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 160, + 554, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 554, + 205 + ], + "type": "text", + "content": "Figure 2. (left) The ratio of removed patches based on highest outlier scores by Eq. (3), which shows that most of few-shot class patches are lost. (right) The ratio of sampled patches by greedy coreset sampling from PatchCore which favors both few-shot and anomaly samples." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 552, + 159 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 552, + 159 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 552, + 159 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 552, + 159 + ], + "type": "image", + "image_path": "87aadc70551d1e473d78aa2ab458fd971f200f34b2fb6e4e45987b03de6c0509.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 215, + 446, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 446, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 446, + 227 + ], + "type": "text", + "content": "4. Problem and motivation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 235, + 377, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 377, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 377, + 247 + ], + "type": "text", + "content": "4.1. Problem" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "text", + "content": "We consider a practical anomaly detection setting where the normal training data " + }, + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "inline_equation", + "content": "X = \\{x_{i}\\}" + }, + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "text", + "content": " is both multi-class with tail distribution and noise-contaminated. Particularly, each " + }, + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "text", + "content": " is paired with the product class label " + }, + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "inline_equation", + "content": "y_{i} \\in \\mathcal{Y} = \\{1,2,\\dots,|\\mathcal{Y}|\\}" + }, + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "text", + "content": " (e.g., hazelnut, screw, pill, etc.) and an anomaly label " + }, + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "inline_equation", + "content": "y_{i}^{a} \\in \\mathcal{Y}^{a} := \\{0,1\\}" + }, + { + "bbox": [ + 313, + 254, + 555, + 361 + ], + "type": "text", + "content": " with 1 indicating anomaly. However, this labeling information is unknown to the model trainer. The class distribution is long-tailed such that there are tail (i.e., few-shot) classes whose number of training samples is very low; i.e.," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 384, + 373, + 555, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 373, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 384, + 373, + 555, + 387 + ], + "type": "interline_equation", + "content": "\\left| C _ {k} \\right| = \\left| \\left\\{x _ {i}: y _ {i} ^ {c} = k \\right\\} \\right| \\leq K \\tag {4}", + "image_path": "ce15a7376abf981686a190baae7bbcad9d4b695903d7438a0030ed43419135e5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "text", + "content": "with a very small " + }, + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "text", + "content": " (e.g., 1, 4, or 20), where " + }, + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "text", + "content": " indicates the number of elements in the set. Moreover, some of the training samples in " + }, + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "text", + "content": " are anomalies, containing defect blobs in their images. We assume the tail class samples are all normal since in the unsupervised setting its normality is determined by its majority. Overall, the training set " + }, + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 396, + 555, + 481 + ], + "type": "text", + "content": " with the above descriptions constitutes unsupervised tailed noisy anomaly detection setting." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 491, + 520, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 520, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 520, + 502 + ], + "type": "text", + "content": "4.2. Motivation: Tail-versus-noise trade-off" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 509, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 605 + ], + "type": "text", + "content": "An ideal model to handle a long-tail noisy dataset should selectively sample the few-shot samples while removing noisy samples. However, we observe that the existing models suffer from a trade-off between class-imbalance and noise robustness. For memory-based noise-discriminative models in particular (e.g., SoftPatch), the methods focus on sampling the majority patch features, which often ignores less dominating but important patches." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "As shown in Fig. 2 left, the noise discrimination process removes both defect and few-shot class patches, losing most of the information in the tail distribution. On the other hand, PatchCore samples few-shot samples well due to greedy coreset sampling which captures maximally different features. This sampling principle, however, favors the noise patch features as well as shown in Fig. 2 (right) and the low performance on noisy head classes in Fig. 1. Previous methods, hence, face the tail-versus-noise dilemma described above." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25541" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "To resolve this, we propose to handle the tail classes independently from the head classes that potentially contain defective local regions. This, however, needs exclusive sampling of tail classes. To achieve this, we use the globally average pooled embedding vector, which is invariant against local pixel variation and rather exclusively corresponds to the object-centric nature of image, i.e., class." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 109, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 109, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 109, + 180 + ], + "type": "text", + "content": "5. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 189, + 296, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 296, + 261 + ], + "type": "text", + "content": "We aim to build a coreset-based anomaly detection model using patch features. We require its memory to aptly capture tail class information and also to be absent of noise patches as much as possible. Particularly, we augment the noise discriminated memory " + }, + { + "bbox": [ + 55, + 189, + 296, + 261 + ], + "type": "inline_equation", + "content": "M_{clean}" + }, + { + "bbox": [ + 55, + 189, + 296, + 261 + ], + "type": "text", + "content": " with the tail class memory " + }, + { + "bbox": [ + 55, + 189, + 296, + 261 + ], + "type": "inline_equation", + "content": "M_{tail}" + }, + { + "bbox": [ + 55, + 189, + 296, + 261 + ], + "type": "text", + "content": ", which contains the patches of few-shot class samples exclusively:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 271, + 296, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 271, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 115, + 271, + 296, + 283 + ], + "type": "interline_equation", + "content": "M _ {\\text {T a i l e d C o r e}} = M _ {\\text {c l e a n}} \\cup M _ {\\text {t a i l}}. \\tag {5}", + "image_path": "788e9d61fcf77009d28d2a8469342c8be36710dbbb1c19a3f640a0e068e3e9df.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 294, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 367 + ], + "type": "text", + "content": "The main challenge lies on sampling the tail class patches exclusively. To this end, we devise a few-shot sampler by estimating the class size of samples based on its embedding (e.g., global average pooling layer of the encoder). Using the estimated class sizes of training samples, the few-shot sampler enables us to exclusively sample tail class samples and the patches thereof." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 374, + 228, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 228, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 228, + 388 + ], + "type": "text", + "content": "5.1. Few-shot sampler: TailSampler" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "content": "TailSampler first estimates the class size of every training sample. Then, by automatically determining the maximum number of tail classes, denoted by " + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 392, + 296, + 441 + ], + "type": "text", + "content": ", TailSampler collects the tail class samples' patch features exclusively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 448, + 167, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 167, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 167, + 460 + ], + "type": "text", + "content": "5.1.1.Class size predictor" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "text", + "content": "To sample patches exclusively from few-shot class, we identify samples from few-shot classes by estimating their \"class sizes\", namely the size of class each sample belongs to. To accurately measure the class size, we hypothesize as follows: In the embedding space, samples within an appropriate angle are majorly of the same class, and this angle is likely to be the half of maximum angle to another embedding. In addition, the neighborhoods of neighborhoods within the half angle give accurate class size estimate. Based on these hypotheses, we proceed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": "First, we adaptively set angle " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": " for each embedding " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": " from training samples " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": ". The coverage of angle " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": " decides the neighbors where a large value would assign many neighbors. We set the angle to be half of its maximum within the train set, and hypothesize that most of neighbors within the region is likely of the same class. Particularly, let " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "H_{i} = \\{e\\in Z:\\angle (e_{i},e)\\leq m_{i} / 2\\}" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": " denote the half-angle region where the maximum angle is " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "m_{i}\\coloneqq \\max_{e\\in Z}\\angle (e_{i},e)" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": " and the set " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": " contains all train embeddings. Then, the adaptive angle " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": " is defined to contain " + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": "-th percentile of the half-angle region:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 701, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 701, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 135, + 701, + 296, + 715 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\angle \\left(e _ {i}, e _ {\\left(p \\cdot | H _ {i} |\\right)}\\right), \\tag {6}", + "image_path": "a4e37fa760376bdaa647e7f361cd9c8d148aa32f3cbb3a56c9357c62625d52a4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "content": "where the index " + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "inline_equation", + "content": "(j)" + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "inline_equation", + "content": "e_{(j)}" + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "content": " is sorted in the increasing order of the angle " + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "inline_equation", + "content": "\\angle (e_i,e_{(j)})" + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "inline_equation", + "content": "p = 0.85" + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "content": " in all experiments unless specified otherwise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "text", + "content": "After setting the adaptive angle " + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "text", + "content": " for each train embedding " + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "text", + "content": ", we estimate its class size based on neighborhoods of neighborhoods. Let " + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "inline_equation", + "content": "N_{\\alpha}(e_i) = \\{e \\in Z : \\angle(e_i, e) \\leq \\alpha\\}" + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "text", + "content": " denote the neighborhood of " + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "text", + "content": ", which is the set of all train embedding " + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "text", + "content": " within the angle " + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 313, + 108, + 555, + 180 + ], + "type": "text", + "content": ". Then, the class size is estimated by the mode of the sizes of neighborhoods of the neighborhood:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 190, + 555, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 190, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 381, + 190, + 555, + 211 + ], + "type": "interline_equation", + "content": "\\kappa_ {i} = \\underset {e \\in N _ {\\alpha_ {i}} (e _ {i})} {\\operatorname {m o d e}} \\left(| N _ {\\alpha (e)} (e) |\\right), \\tag {7}", + "image_path": "87b77910c416ef17a2c2422bc424b228ced12e13c1c7d872f25a1857feb36be0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "inline_equation", + "content": "\\alpha(e)" + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "text", + "content": " is the adaptive angle with respect to the embedding " + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "text", + "content": " belonging to the neighborhood " + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "inline_equation", + "content": "N_{\\alpha_i}(e_i)" + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "text", + "content": " of embedding " + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "text", + "content": ". Using the neighborhoods of neighborhood and mode on it gives more robust estimation of the class size than mere count of the direct neighborhood of " + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 313, + 220, + 555, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 281, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 316 + ], + "type": "text", + "content": "The proposition in Supp. 9 shows that the angle of neighborhood given by Eq. (6) corresponds to the decision boundary that maximizes the inter-class separation under regularities." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 323, + 514, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 514, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 514, + 334 + ], + "type": "text", + "content": "5.1.2. Estimation of maximum size of tail class" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": "After estimating the class size of each training sample by " + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\kappa_{i}" + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": ", we determine the maximum size for the tail classes based on the elbow technique [36]. In particular, by utilizing the sample-wise class sizes " + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\kappa_{i}" + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": ", we first obtain rough estimation of the size of each class set " + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\eta_{y} \\approx |C_{y}|" + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": ". Then, we estimate the maximum size of tail classes " + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "inline_equation", + "content": "K_{max}" + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": " by finding a point where the class size most abruptly changes. This optimal shift point is found by the elbow technique. Concretely, we first estimate the size of each class " + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\eta_{y} \\approx |C_{y}|" + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": ", inductively by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 456, + 555, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 456, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 345, + 456, + 555, + 529 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\eta_ {1} = \\operatorname {r o u n d} \\left(\\frac {1}{\\kappa_ {(1)}} \\sum_ {i = 1} ^ {\\kappa_ {(1)}} \\kappa_ {(i)}\\right), \\\\ \\eta_ {y + 1} = \\operatorname {r o u n d} \\left(\\frac {1}{\\kappa_ {\\left(\\eta_ {y} + 1\\right)}} \\sum_ {i = \\eta_ {y} + 1} ^ {\\min \\left(\\kappa_ {\\left(\\eta_ {y} + 1\\right)}, | X |\\right)} \\kappa_ {(i)}\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "1404ebb1e121988b659efcd4caedd41d7a8664bc90edf035c2b06418b69579e8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 537, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 554, + 563 + ], + "type": "text", + "content": "where the index " + }, + { + "bbox": [ + 313, + 537, + 554, + 563 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 313, + 537, + 554, + 563 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 537, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\kappa_{(i)}" + }, + { + "bbox": [ + 313, + 537, + 554, + 563 + ], + "type": "text", + "content": " is sorted in the increasing order " + }, + { + "bbox": [ + 313, + 537, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\kappa_{(1)}\\leq \\dots \\leq \\kappa_{(|X|)}" + }, + { + "bbox": [ + 313, + 537, + 554, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": "After we acquire the set " + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\{(y,\\eta_y):y\\in (1,\\ldots ,|\\mathcal{Y}|)\\}" + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": " through the inductive process above, we define the maximum size of tail classes as the elbow point of the estimated sizes of classes " + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\eta_{y}" + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": " since at this point the class size increases abruptly. The elbow is computed by linearly connecting " + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "inline_equation", + "content": "(1,\\eta_{1})" + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "inline_equation", + "content": "(|\\mathcal{V}|,\\eta_{|\\mathcal{V}|})" + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": " , and finding the maximum orthogonal distance onto the line from the set of points " + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\{(y,\\eta_y):y\\in (1,\\dots |\\mathcal{Y}|)\\}" + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": " as shown below:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 365, + 656, + 555, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 656, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 365, + 656, + 555, + 670 + ], + "type": "interline_equation", + "content": "K _ {\\max } := \\eta_ {y _ {\\max }} =: e l b o w (\\eta_ {1}, \\dots , \\eta_ {| \\mathcal {Y} |}) \\tag {9}", + "image_path": "48582415cbced2110065144b18f1297c81fc58d5501d51c7bc1475746a5efeb2.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 679, + 341, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 679, + 341, + 689 + ], + "spans": [ + { + "bbox": [ + 314, + 679, + 341, + 689 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 373, + 689, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 373, + 689, + 555, + 715 + ], + "type": "interline_equation", + "content": "y _ {\\max } = \\underset {y \\in \\{1, \\dots , | \\mathcal {Y} | \\}} {\\operatorname {a r g m a x}} \\frac {| a y - \\eta_ {y} + b |}{\\sqrt {a ^ {2} + 1}} \\tag {10}", + "image_path": "6283176b9927cc9d37376f1e21781ba3fe8cf5782f479a4d308d3dcbcaa0849e.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25542" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 203 + ], + "type": "image", + "image_path": "ba9e9fcdd99af9e088ed6ee4234468f17bb787d7ab9350bc27caf2a2cdf5cdbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 204, + 520, + 216 + ], + "lines": [ + { + "bbox": [ + 88, + 204, + 520, + 216 + ], + "spans": [ + { + "bbox": [ + 88, + 204, + 520, + 216 + ], + "type": "text", + "content": "Figure 3. (a) Sampling process description of TailedCore (ours) and (b) the illustration of how we use the elbow method [36]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "text", + "content": "and where the line " + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "inline_equation", + "content": "\\eta_y = ay + b" + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "text", + "content": " intersects " + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "inline_equation", + "content": "(1, \\eta_1)" + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "inline_equation", + "content": "(|\\mathcal{Y}|, \\eta_{|\\mathcal{Y}|})" + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "inline_equation", + "content": "a = (\\eta_{|\\mathcal{Y}|} - \\eta_1) / (|\\mathcal{Y}| - 1)" + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "inline_equation", + "content": "b = \\eta_1 - m" + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "text", + "content": ". Fig. 3 (b) illustrates Eq. (10). We choose the elbow point " + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "inline_equation", + "content": "(y_{max}, \\eta_{y_{max}})" + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "text", + "content": " as the maximum size of tail classes " + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "inline_equation", + "content": "K_{max}" + }, + { + "bbox": [ + 55, + 226, + 295, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 296, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 298 + ], + "type": "text", + "content": "The maximum size of tail classes " + }, + { + "bbox": [ + 55, + 274, + 296, + 298 + ], + "type": "inline_equation", + "content": "K_{max}" + }, + { + "bbox": [ + 55, + 274, + 296, + 298 + ], + "type": "text", + "content": " allows us to sample the patch features of tail classes exclusively:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 304, + 296, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 304, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 110, + 304, + 296, + 320 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {\\text {t a i l}} (P) = \\left\\{\\phi_ {i} ^ {(h, w)}: \\kappa_ {i} \\leq K _ {\\max } \\right\\}. \\tag {11}", + "image_path": "4391a72f26bc3ff28fe4f504897e7c6029fc28d1b12cb6648f9630a8a12289c6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 326, + 152, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 152, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 152, + 338 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 326, + 152, + 338 + ], + "type": "inline_equation", + "content": "\\kappa_{i}" + }, + { + "bbox": [ + 55, + 326, + 152, + 338 + ], + "type": "text", + "content": " given by Eq. (7)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 343, + 220, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 220, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 220, + 357 + ], + "type": "text", + "content": "5.2. Proposed method: TailedCore" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "text", + "content": "Our anomaly detection model is built upon the memory bank constructed as below. First, we sample the base memory bank as in [20] with noise-discrimination described in Eq. (3). Then, the base memory is augmented with TailSampler " + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "inline_equation", + "content": "S_{tail}" + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "text", + "content": " (Eq. (11)). Overall, TailedCore's memory bank is defined by " + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "inline_equation", + "content": "M_{TailedCore} = S_{core}(S_{clean}(P)) \\cup S_{core}(S_{tail}(P))" + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "inline_equation", + "content": "S_{core}" + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "text", + "content": " is the greedy coreset sampling in Eq. (1), " + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "inline_equation", + "content": "S_{clean}" + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "text", + "content": " is the noise discrimination process that filters out defect patches based on their outlier scores (Eq. (3)), and the few-shot sampler " + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "inline_equation", + "content": "S_{tail}" + }, + { + "bbox": [ + 54, + 361, + 296, + 553 + ], + "type": "text", + "content": " augments the tail class patches exclusively on the noise-discriminated base memory bank. The whole process is shown in Fig 3. As TailedCore's patch memory bank is both noise-clean and contains the representative features of few-shot classes, it is robust against both class imbalance and normal sample contamination. In inference, the anomaly score is obtained by Eq. (2) as in the conventional memory-based detector." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 562, + 133, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 133, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 133, + 575 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 618 + ], + "type": "text", + "content": "Section 6 consists of dataset description, comparison to the SOTA models, and the detailed ablation study and analysis of TailedCore's components." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 624, + 295, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 649 + ], + "type": "text", + "content": "6.1. Datasets for unsupervised long-tail noisy anomaly detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "We devise the long-tail noisy anomaly detection datasets by modifying the class-wise sample distributions of the widely used anomaly detection datasets MVTecAD [3] and VisA [44]. We consider three different types of long-tail distributions. Particularly, datasets with step-shaped tail distribution are devised" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 224, + 554, + 303 + ], + "blocks": [ + { + "bbox": [ + 315, + 224, + 554, + 303 + ], + "lines": [ + { + "bbox": [ + 315, + 224, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 315, + 224, + 554, + 303 + ], + "type": "table", + "html": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDim [9] ICPR2182.4580.9582.0677.4781.2879.1971.5481.7575.63
HVQ [26] NeurIPS2383.4680.2382.9982.0185.5083.5674.1590.1580.55
WinCLIP [19] CVPR2389.3590.1190.3791.6088.2190.3791.8088.2390.37
AnomalyCLIP [43] ICLR2490.9390.9891.4891.8290.8391.4891.2191.9091.48
PatchCore [34] CVPR2293.3387.5989.1892.1971.1883.8386.3670.4880.01
SoftPatch [20] NeurIPS2284.6886.9587.7167.6597.5479.6460.6697.4975.40
TailedCore (ours)96.5595.2496.1295.8295.3495.7193.5495.7794.43
", + "image_path": "3ef8b5b6e645be1cf73068d3c97c3e44f76327e51d299c8f692a1e3d7782a9c2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 316, + 343, + 553, + 422 + ], + "blocks": [ + { + "bbox": [ + 314, + 304, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 314, + 304, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 555, + 335 + ], + "type": "text", + "content": "Table 1. Anomaly classification on MVTecAD with image-level AUROC " + }, + { + "bbox": [ + 314, + 304, + 555, + 335 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 314, + 304, + 555, + 335 + ], + "type": "text", + "content": ". We report the mean over 5 random seeds for each measurement. Notations: " + }, + { + "bbox": [ + 314, + 304, + 555, + 335 + ], + "type": "inline_equation", + "content": "C_h / C_t" + }, + { + "bbox": [ + 314, + 304, + 555, + 335 + ], + "type": "text", + "content": ": head / tail classes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 343, + 553, + 422 + ], + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 422 + ], + "type": "table", + "html": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2170.7083.3578.6460.6588.9372.4355.9886.7568.80
HVQ [26]NeuRPS2373.4784.0368.2568.2589.3077.0261.5780.4069.42
WinCLIP [19]CVPR2373.2576.9275.4775.9874.7675.4778.8070.8075.47
AnomalyCLIP [43]ICLR2481.9682.4882.0582.2881.7482.0583.2680.3482.05
PatchCore [34]CVPR2286.1185.7385.5983.5367.5176.8579.3368.5674.84
SoftPatch [20]NeuRPS2278.0492.1686.5659.7095.9774.8152.6194.1769.92
TailedCore (ours)87.5593.0690.8585.1695.9189.6482.9794.1187.61
", + "image_path": "031aa3ae5552f9cf33f916efdbc963e8591a850759a7fdac9be9e1adc2d2d8dc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 316, + 452, + 553, + 532 + ], + "blocks": [ + { + "bbox": [ + 314, + 422, + 553, + 445 + ], + "lines": [ + { + "bbox": [ + 314, + 422, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 553, + 445 + ], + "type": "text", + "content": "Table 2. Anomaly classification on VisA with image-level AUROC " + }, + { + "bbox": [ + 314, + 422, + 553, + 445 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 314, + 422, + 553, + 445 + ], + "type": "text", + "content": ". The format and evaluation protocol are the same as Tab. 1." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 452, + 553, + 532 + ], + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 532 + ], + "type": "table", + "html": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2190.1192.6691.4382.5395.2987.6778.8095.5485.50
HVQ [26]NeuRPS2393.6386.8590.5590.7392.5891.5386.3695.2089.90
WinCLIP [19]CVPR2382.0384.0682.2980.6084.6382.2980.1685.4882.29
AnomalyCLIP [43]ICLR2491.2491.6991.0889.9692.6691.0889.3493.6891.08
PatchCore [34]CVPR2293.5687.9889.9393.5472.0985.1992.0271.3583.75
SoftPatch [20]NeuRPS2292.1993.8393.4180.9896.4987.2470.3496.8980.99
TailedCore (ours)96.0895.0195.2995.5693.2094.7494.1993.7093.99
", + "image_path": "7821c6f9d7eec1b9fdaff970d5a291d4f8344ef529962f2d7a20df4b45535157.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 531, + 555, + 564 + ], + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 564 + ], + "type": "text", + "content": "Table 3. Anomaly segmentation on MVTecAD with pixel-level AUROC (\\%). We report the mean over 5 random seeds for each measurement. Notations: " + }, + { + "bbox": [ + 313, + 531, + 555, + 564 + ], + "type": "inline_equation", + "content": "C_h / C_t" + }, + { + "bbox": [ + 313, + 531, + 555, + 564 + ], + "type": "text", + "content": ": head / tail classes." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "by making the particular classes to have only a small number of samples; namely " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", imposing extreme class imbalance on the datasets. We name these two datasets by 'step " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "(K = 1)" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "' and 'step " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "(K = 4)" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "'. On the other hand, another realistic type of tail-distribution is considered by making the class distribution to follow the Pareto distribution with the shape parameter 0.6. For the step-tailed distribution datasets, we impose " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " of classes to be head classes, containing the original number of training samples. Thus, in MVTecAD, 6 of 15 classes are head classes. For the Pareto tailed distribution, we regard as few-shot (i.e., tail classes) the classes whose number of samples" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25543" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 148 + ], + "type": "table", + "html": "
tail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
PaDiM [9]ICPR2189.0295.1082.8183.9097.3689.5182.5796.5788.40
HVQ [26]NeuIPS2395.2797.6096.7193.8898.3495.7490.5895.5192.63
WinCLIP [19]CVPR2371.9473.9773.1974.6071.2173.1973.8172.3273.19
AnomalyCLIP [43]ICLR2495.6095.4695.5195.5495.4895.5196.1694.6095.51
PatchCore [34]CVPR2296.8487.9991.1395.3962.9681.8894.1165.3082.10
SoftPatch [20]NeuIPS2293.2096.7495.2783.9597.1089.4380.7396.8287.43
TailedCore (ours)97.9897.2597.4896.8097.0296.8996.1297.3996.65
", + "image_path": "fad6ae5166c7874704a10d07e64ce5af9dea7949adb2e5ed7989a4dee7c8083f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 149, + 295, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 149, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 295, + 171 + ], + "type": "text", + "content": "Table 4. Anomaly segmentation on VisA with pixel-level AUROC " + }, + { + "bbox": [ + 55, + 149, + 295, + 171 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 149, + 295, + 171 + ], + "type": "text", + "content": ". The format and evaluation protocol are the same as Tab. 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 181, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 296, + 266 + ], + "type": "text", + "content": "is less than 20 [25]. To impose noisy condition on the datasets, the anomaly samples (i.e., the samples with defective regions) are added to the training samples such that the noisy samples constitute " + }, + { + "bbox": [ + 55, + 181, + 296, + 266 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 181, + 296, + 266 + ], + "type": "text", + "content": " of the training set following the overlap protocol in SoftPatch [20] for fair comparison. In all cases, the testing sets of datasets remain the same as the original MVTecAD and VisA. A more detailed configuration is given in Supp. 11." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 273, + 277, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 277, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 277, + 285 + ], + "type": "text", + "content": "6.2. Comparison to the state-of-the-art (SOTA)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 290, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 290, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 54, + 290, + 295, + 493 + ], + "type": "text", + "content": "We compare TailedCore to the SOTA anomaly detection methods in the unsupervised long-tail noisy settings. We carefully select our baselines. HVQ excels in multi-class learning, PaDiM and SoftPatch are robust against noises as PaDiM is based on Gaussian modeling of the data, while SoftPatch explicitly removes the defect patches by outlier scoring. Using foundation models, WinClip and AnoamlyClip are strong for zero-shot learning. PatchCore are near-perfect in the single-class anomaly detection settings. We show the full experimental results in Supp. 12. All models are trained in the multi-class setting; namely, one model for one dataset with all classes rather than one model per class. The detection performance, however, is measured for each product class under image-level and pixel-level AUROC, and averaged for reporting and comparison. For comparison, we consider models that are specifically robust against noise, strong on multi-class training, and foundation models that can exhibit zero/few-shot learning capability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 494, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 578 + ], + "type": "text", + "content": "Tabs. 1, 2, 3, and 4 show the superiority of TailedCore across all types of long-tail distributions. As anticipated, all baseline models encounter the tail-versus-noise trade-off, resulting in suboptimal performance either on tail classes with limited samples or on head classes with noisy training samples. These trends become more pronounced in highly unbalanced datasets, where the few-shot classes contain only 1 or 4 samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 585, + 208, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 208, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 208, + 597 + ], + "type": "text", + "content": "6.3. Ablation study and analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 603, + 187, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 187, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 187, + 615 + ], + "type": "text", + "content": "6.3.1. Tail (few-shot) sampling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "We evaluate how accurately our proposed algorithm " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "S_{tail}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " can sample tail (few-shot) class samples exclusively without including noise samples. First of all, we measure the ratio of mis-included tail class samples TailSampler which is the ratio of not sampled tail class samples to total tail class samples. A single mis-inclusion can vitally damage anomaly detection performance especially in case of the one-shot setting " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " as the model wouldn't be able to utilize the only sample of the class." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "text", + "content": "Additionally, the exclusive aspect of few-shot sampling is crucial; the few-shot sampled set shouldn't include anomaly and head class samples as possible. Included anomaly samples would harm the head class performance and head class sampled by few-shot sampler might cause class imbalance issue as well as the possibility of sampling anomaly from head classes. Hence, we measure the ratios of included anomalies and head-class samples along with AUROC which is measured on the outlier scores derived by a given algorithm and ground truth binary labels that indicate whether a given sample is tail class or head class. For TailSampler, the outlier score of each train sample is given as the negative of predicted class size. The included anomaly ratio is defined as the ratio of number of anomaly samples to the number of whole anomaly samples. The ratio of included head-class sample is defined as the number of included head class to the number of included tail class samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 263, + 554, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 383 + ], + "type": "text", + "content": "For extensive comparison, we compare TailSampler with widely used clustering algorithms [2, 13, 16], outlier detectors [1, 5, 24, 35], and density estimators [10, 30]. These methods provide clustering size and/or outlier score/prediction. In cases when only outlier scores are available, we apply the elbow method to induce deterministic prediction of whether a given sample is few-shot or not. We use default settings provided by Sklearn [31] and PyOD [42]. For all methods, we utilize the embedding vectors of WideResNet-50 [40]. Experiments with different architectures and details are provided in Supp. 10." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 552 + ], + "type": "text", + "content": "The combined ratio, which is the sum of missing tail samples and included anomaly samples, indicates that TailSampler is most effective for exclusively sampling few-shot class samples. The most critical factor is to not miss few-shot class samples. On this ratio, TailSampler achieves the lowest rate except for the algorithms that trivially sample all few-shot class samples. We find that most of the classical algorithms for clustering and outlier detection are not usable as they either miss too many few-shot samples or includes a large number of anomaly samples. We note that the class size predictor in our few-shot samples resembles DBSCAN and Affinity Propagation, respectively, in terms of using neighborhoods and voting mechanism. None of these methods however have instance-specific parameters that are adaptively adjusted across different samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 555, + 430, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 430, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 430, + 567 + ], + "type": "text", + "content": "6.3.2. Class size prediction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 571, + 554, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 619 + ], + "type": "text", + "content": "We measure how accurately our class size predictor estimates the class size of each sample by measuring the class size prediction error. It is measured for each train sample and weight-averaged through the whole train samples:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 372, + 624, + 555, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 624, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 372, + 624, + 555, + 658 + ], + "type": "interline_equation", + "content": "e r r o r = \\frac {1}{| \\mathcal {X} |} w _ {i} \\sum_ {i = 1} ^ {| \\mathcal {X} |} \\left| \\kappa_ {i} - \\left| C _ {k (i)} \\right| \\right| \\tag {12}", + "image_path": "d417d8c1f165abcba119a47f45d731db02273f3ec5ffd2089b9abc0c125e4b4c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "k(i)" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": " indicates the index such that " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "x_{i} \\in C_{k(i)}" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": " (namely, " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "C_{k(i)}" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": " is the class of " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "). The weight is chosen by " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "w_{i} = 1 / |C_{k(i)}|" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": ", which is to give more penalty on the few-shot class samples more than the head class samples." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25544" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 173, + 155 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 173, + 155 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 173, + 155 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 173, + 155 + ], + "type": "image", + "image_path": "152311089b96de8555f960d6ca9fed416477e5d0357d3e8fc144abc2b7317b43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 160, + 120, + 168 + ], + "lines": [ + { + "bbox": [ + 111, + 160, + 120, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 160, + 120, + 168 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 170, + 172, + 254 + ], + "blocks": [ + { + "bbox": [ + 57, + 170, + 172, + 254 + ], + "lines": [ + { + "bbox": [ + 57, + 170, + 172, + 254 + ], + "spans": [ + { + "bbox": [ + 57, + 170, + 172, + 254 + ], + "type": "image", + "image_path": "e729b819c2118e56ba16c1a0a1940bd4dfc07c914c7583bb628dfb4aae8f6df4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 258, + 120, + 266 + ], + "lines": [ + { + "bbox": [ + 111, + 258, + 120, + 266 + ], + "spans": [ + { + "bbox": [ + 111, + 258, + 120, + 266 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 269, + 172, + 353 + ], + "blocks": [ + { + "bbox": [ + 57, + 269, + 172, + 353 + ], + "lines": [ + { + "bbox": [ + 57, + 269, + 172, + 353 + ], + "spans": [ + { + "bbox": [ + 57, + 269, + 172, + 353 + ], + "type": "image", + "image_path": "ebb0db69dd702d850c868a7a4c91b7993907b663f3bdd5a94387380c61cb7050.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 357, + 120, + 365 + ], + "lines": [ + { + "bbox": [ + 111, + 357, + 120, + 365 + ], + "spans": [ + { + "bbox": [ + 111, + 357, + 120, + 365 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 71, + 293, + 156 + ], + "blocks": [ + { + "bbox": [ + 178, + 71, + 293, + 156 + ], + "lines": [ + { + "bbox": [ + 178, + 71, + 293, + 156 + ], + "spans": [ + { + "bbox": [ + 178, + 71, + 293, + 156 + ], + "type": "image", + "image_path": "411a6a35c177bd2d5d4c88cf36b7b0b5e5b712fa13ab9f0d988198bcbf586c88.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 159, + 241, + 168 + ], + "lines": [ + { + "bbox": [ + 231, + 159, + 241, + 168 + ], + "spans": [ + { + "bbox": [ + 231, + 159, + 241, + 168 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 179, + 170, + 293, + 254 + ], + "blocks": [ + { + "bbox": [ + 179, + 170, + 293, + 254 + ], + "lines": [ + { + "bbox": [ + 179, + 170, + 293, + 254 + ], + "spans": [ + { + "bbox": [ + 179, + 170, + 293, + 254 + ], + "type": "image", + "image_path": "b774650a828046de9a5043b7a620221b1bae70aa92a70778ba9170a06713eb24.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 258, + 241, + 267 + ], + "lines": [ + { + "bbox": [ + 231, + 258, + 241, + 267 + ], + "spans": [ + { + "bbox": [ + 231, + 258, + 241, + 267 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 179, + 269, + 293, + 353 + ], + "blocks": [ + { + "bbox": [ + 179, + 269, + 293, + 353 + ], + "lines": [ + { + "bbox": [ + 179, + 269, + 293, + 353 + ], + "spans": [ + { + "bbox": [ + 179, + 269, + 293, + 353 + ], + "type": "image", + "image_path": "48d33364309a3fa03f7362eb1046fb4206ff6652013af25084b37575b0dff952.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 357, + 241, + 365 + ], + "lines": [ + { + "bbox": [ + 231, + 357, + 241, + 365 + ], + "spans": [ + { + "bbox": [ + 231, + 357, + 241, + 365 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 366, + 295, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 431 + ], + "type": "text", + "content": "Figure 4. The plot of noise ratio of train data versus anomaly classification (image-level AUROC) and segmentation (pixel-level AUROC) performance. In all cases, TailedCore outperforms both PatchCore and SoftPatch. Dataset is MVTecAD with different tail distributions. The lines show the mean and shades show standard deviation of multiple runs of experiments with different seeds." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 519 + ], + "type": "text", + "content": "Tab. 5 shows that most of the clustering based methods vastly fail on this task, and the estimation by our class size predictor outperforms them significantly. We believe this is partly due to suboptimal hyperparameters used for these algorithms. Due to their classical nature, however, it is unclear how to tune them without the availability of the validation set in our scenario." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 529, + 248, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 248, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 248, + 541 + ], + "type": "text", + "content": "6.3.3. Abalation study on class size predictor" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": "We conduct detailed ablation study on the proposed class size prediction in Eq. (7). One hyperparameter used in TailSampler " + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "inline_equation", + "content": "S_{tail}" + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": " of TailedCore is the percentiling proportion " + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": " to consider only the majority of nearby samples within the neighborhood. We analyze how this parameter can affect the anomaly detection performance of TailedCore. Tab. 5 indicates that using the nearest " + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": " embeddings (i.e., " + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "inline_equation", + "content": "p = 85\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": ") is the best choice. Using all of the available neighborhoods (i.e., " + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "inline_equation", + "content": "p = 100\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": ") would include different class samples in the neighborhood, deteriorating class estimation. On the other hand, counting only with the half of samples (i.e., " + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "inline_equation", + "content": "p = 50\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": ") in the neighborhood may lose too much information." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Another component in the class size predictor that requires analysis is its voting mechanism used in Eq. (7). One may" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 553, + 184 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 184 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 184 + ], + "type": "table", + "html": "
methodcombined ratio of missing \\( {C}_{t} \\) & included AS (↓)ratio of missing \\( {C}_{t} \\) samples (↓)ratio of included AS (↓)ratio of included \\( {C}_{h} \\) samples (↓)class size prediction error (↓)AUROC (%) (↑)
LOF\\( {89.58} \\pm {44.42} \\)\\( {31.22} \\pm {34.88} \\)\\( {58.36} \\pm {9.54} \\)\\( {92.50} \\pm {7.87} \\)\\( {1.00} \\pm {0.00} \\)\\( {77.08} \\pm {25.86} \\)
IF\\( {18.15} \\pm {15.91} \\)\\( {7.25} \\pm {10.05} \\)\\( {10.90} \\pm {5.86} \\)\\( {69.02} \\pm {25.96} \\)\\( {1.00} \\pm {0.00} \\)\\( {98.96} \\pm {1.27} \\)
OCSVM\\( {49.95} \\pm {9.76} \\)\\( {0.78} \\pm {2.19} \\)\\( {49.17} \\pm {7.57} \\)\\( {95.06} \\pm {6.08} \\)\\( {1.00} \\pm {0.00} \\)\\( {96.72} \\pm {3.59} \\)
DBSCAN\\( {35.36} \\pm {49.55} \\)\\( {30.64} \\pm {44.38} \\)\\( {4.72} \\pm {5.17} \\)\\( {25.57} \\pm {27.24} \\)\\( {0.35} \\pm {0.44} \\)\\( {84.46} \\pm {22.18} \\)
KMeans\\( {94.67} \\pm {44.95} \\)\\( {47.22} \\pm {35.49} \\)\\( {47.45} \\pm {9.46} \\)\\( {97.40} \\pm {3.99} \\)\\( {4.27} \\pm {1.40} \\)\\( {52.58} \\pm {20.46} \\)
GMM\\( {100.00} \\pm {0.00} \\)\\( \\mathbf{{0.00}} \\pm \\mathbf{{0.00}} \\)\\( {100.00} \\pm {0.00} \\)\\( {97.58} \\pm {2.86} \\)\\( {12.50} \\pm {1.53} \\)\\( {50.00} \\pm {0.00} \\)
KDE\\( {29.81} \\pm {24.15} \\)\\( {2.14} \\pm {6.26} \\)\\( {27.67} \\pm {17.89} \\)\\( {87.47} \\pm {10.35} \\)\\( {1.00} \\pm {0.00} \\)\\( {97.98} \\pm {2.19} \\)
AP\\( {27.81} \\pm {32.59} \\)\\( {12.98} \\pm {24.01} \\)\\( {14.83} \\pm {8.58} \\)\\( {77.45} \\pm {28.21} \\)\\( {0.70} \\pm {0.22} \\)\\( {93.03} \\pm {12.8} \\)
KNN\\( {43.49} \\pm {31.25} \\)\\( {14.46} \\pm {24.07} \\)\\( {29.03} \\pm {7.18} \\)\\( {83.12} \\pm {18.28} \\)\\( {1.00} \\pm {0.00} \\)\\( {94.51} \\pm {10.34} \\)
TailSampler (ours)\\( {5.10} \\pm {5.07} \\)\\( {0.12} \\pm {0.65} \\)\\( {4.98} \\pm {4.42} \\)\\( {31.96} \\pm {26.05} \\)\\( \\mathbf{{0.19}} \\pm \\mathbf{{0.05}} \\)\\( {99.78} \\pm {0.30} \\)
Ablation on the class size predictor:
TailSampler (p=100)\\( {8.48} \\pm {10.88} \\)\\( {1.07} \\pm {5.23} \\)\\( {7.41} \\pm {5.65} \\)\\( {42.64} \\pm {31.86} \\)\\( {0.51} \\pm {0.04} \\)\\( {99.76} \\pm {0.29} \\)
TailSampler (p=50)\\( {7.37} \\pm {10.35} \\)\\( {3.46} \\pm {6.80} \\)\\( {3.91} \\pm {3.55} \\)\\( {27.90} \\pm {27.36} \\)\\( {0.12} \\pm {0.12} \\)\\( {97.36} \\pm {5.65} \\)
TailSampler (top-1)\\( {14.47} \\pm {7.99} \\)\\( {0.20} \\pm {0.78} \\)\\( {14.27} \\pm {7.21} \\)\\( {73.97} \\pm {27.95} \\)\\( {0.85} \\pm {0.04} \\)\\( {99.73} \\pm {0.34} \\)
TailSampler (average)\\( {6.48} \\pm {10.11} \\)\\( {3.65} \\pm {6.99} \\)\\( {2.83} \\pm {3.12} \\)\\( {22.16} \\pm {24.33} \\)\\( {0.21} \\pm {0.07} \\)\\( {99.35} \\pm {2.59} \\)
", + "image_path": "d6d38b54e36654772fdedf91eef296db15e3a89e07c007efadba19196acc91d9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 315, + 292, + 555, + 354 + ], + "blocks": [ + { + "bbox": [ + 313, + 185, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 185, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 555, + 270 + ], + "type": "text", + "content": "Table 5. The analysis on the proposed TailSampler and its class size predictor. The results are obtained from both MVTecAD and VisA and averaged across all different types of tail distributions and random seeds (overall 60 different seeds) with the standard deviations indicated on the right. The best and second best performances are marked in bold and underline, respectively. Acronyms/notations: IF: Isolation Forest; AP: Affinity Propagation; AS: anomaly samples; " + }, + { + "bbox": [ + 313, + 185, + 555, + 270 + ], + "type": "inline_equation", + "content": "C_h / C_t" + }, + { + "bbox": [ + 313, + 185, + 555, + 270 + ], + "type": "text", + "content": ": head / tail classes." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 292, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 315, + 292, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 292, + 555, + 354 + ], + "type": "table", + "html": "
tasktail typeParetostep (K=4)step (K=1)
class typeCtChallCtChallCtChall
anomaly classificationw/o FSA82.2291.2288.1360.0197.6075.0566.2897.8578.94
w/ FSA96.5595.2496.1393.5495.7894.4495.8295.3495.72
anomaly segmentationw/o FSA91.6094.9493.5468.9996.8580.1479.3996.3586.23
w/ FSA96.0995.0195.3094.1993.7194.0095.5693.2094.75
", + "image_path": "4de34cd9027a14185d4cc02a9c0faa460224929897c1654e7cdc6853f105d1e0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 355, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 409 + ], + "type": "text", + "content": "Table 6. Abaltion on the few-shot augmentation (FSA) of TailedCore. The image-level and pixel-level AUROC metrics " + }, + { + "bbox": [ + 313, + 355, + 555, + 409 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 355, + 555, + 409 + ], + "type": "text", + "content": " are reported on the MVTecAD dataset with different types of tail distributions for anomaly classification and segmentation tasks, respectively. Notations: " + }, + { + "bbox": [ + 313, + 355, + 555, + 409 + ], + "type": "inline_equation", + "content": "C_h / C_t" + }, + { + "bbox": [ + 313, + 355, + 555, + 409 + ], + "type": "text", + "content": ": head / tail classes." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "type": "text", + "content": "simply use the nearest embedding's neighborhood size instead, or the class size of " + }, + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 434, + 555, + 507 + ], + "type": "text", + "content": " can be estimated by averaging instead of majority vote. As indicated in Tab. 5, both the nearest embedding (i.e., top-1) and averaging are inferior to majority voting. The former is inevitably less robust, while averaging can still be sensitive to the outlier values of neighborhood size." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 526, + 443, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 443, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 443, + 538 + ], + "type": "text", + "content": "6.3.4. Ablation on TailedCore" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "We analyze the impact of augmentation of few-shot class patches by TailSampler in TailedCore. The TailedCore model without these augmented patches has the same memory bank as SoftPatch, but its inference scoring is simpler. Tab. 6 indicates that without the augmented patches from TailSampler, the memory-based model cannot detect anomalies in few-shot class samples. This shows the necessity of tail class augmentation by TailSampler. We note some performance decrease in the head class detection performance since few-shot augmentation may involve few anomaly samples from the head classes that behave like few-shot class instances. This issue intrinsically is due to the weakness of encoder embedding, and the proposed TailSampler based on such embeddings can induce incorrectly sampled few-shot class instances, which we discuss in Sec. 7." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25545" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 175, + 158 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 175, + 158 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 175, + 158 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 175, + 158 + ], + "type": "image", + "image_path": "9717bda343f63692eaf05562fe64c0ffcd246b969c9c5565d6bb41162b2cfb4e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 163, + 122, + 171 + ], + "lines": [ + { + "bbox": [ + 112, + 163, + 122, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 163, + 122, + 171 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 72, + 299, + 158 + ], + "blocks": [ + { + "bbox": [ + 184, + 72, + 299, + 158 + ], + "lines": [ + { + "bbox": [ + 184, + 72, + 299, + 158 + ], + "spans": [ + { + "bbox": [ + 184, + 72, + 299, + 158 + ], + "type": "image", + "image_path": "ce20f1f5e2917667e94b2ca569c2c2bd8cae69cde8903b1a9d60b5645b29507b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 163, + 247, + 171 + ], + "lines": [ + { + "bbox": [ + 237, + 163, + 247, + 171 + ], + "spans": [ + { + "bbox": [ + 237, + 163, + 247, + 171 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 425, + 158 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 425, + 158 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 425, + 158 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 425, + 158 + ], + "type": "image", + "image_path": "08a3066e9f32c6a13df8d9ae59142a883f26650743126ca83c2a02d0c0ba89a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 163, + 373, + 171 + ], + "lines": [ + { + "bbox": [ + 362, + 163, + 373, + 171 + ], + "spans": [ + { + "bbox": [ + 362, + 163, + 373, + 171 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 550, + 158 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 550, + 158 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 550, + 158 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 550, + 158 + ], + "type": "image", + "image_path": "d2acdf6056a1babed87312a360c29e6b8814b976420236aa54344350c4396e70.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 162, + 499, + 171 + ], + "lines": [ + { + "bbox": [ + 488, + 162, + 499, + 171 + ], + "spans": [ + { + "bbox": [ + 488, + 162, + 499, + 171 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 174, + 174, + 259 + ], + "blocks": [ + { + "bbox": [ + 58, + 174, + 174, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 174, + 174, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 174, + 174, + 259 + ], + "type": "image", + "image_path": "d27c9a2cc6be928f5abb9e05bb2f50a32ab6c24ff2e678ad5e3d94b012f00912.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 265, + 122, + 272 + ], + "lines": [ + { + "bbox": [ + 112, + 265, + 122, + 272 + ], + "spans": [ + { + "bbox": [ + 112, + 265, + 122, + 272 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 184, + 174, + 299, + 259 + ], + "blocks": [ + { + "bbox": [ + 184, + 174, + 299, + 259 + ], + "lines": [ + { + "bbox": [ + 184, + 174, + 299, + 259 + ], + "spans": [ + { + "bbox": [ + 184, + 174, + 299, + 259 + ], + "type": "image", + "image_path": "f4883daa739be54c80bc4e97a87f1d4b85029f91b9b8ba1be30c94f110de2885.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 265, + 247, + 272 + ], + "lines": [ + { + "bbox": [ + 238, + 265, + 247, + 272 + ], + "spans": [ + { + "bbox": [ + 238, + 265, + 247, + 272 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 174, + 425, + 259 + ], + "blocks": [ + { + "bbox": [ + 310, + 174, + 425, + 259 + ], + "lines": [ + { + "bbox": [ + 310, + 174, + 425, + 259 + ], + "spans": [ + { + "bbox": [ + 310, + 174, + 425, + 259 + ], + "type": "image", + "image_path": "bcb0e23af33ba75dc74d66826c9a39155bcd6a5b8a0faf7871478dd736a261bf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 265, + 373, + 274 + ], + "lines": [ + { + "bbox": [ + 362, + 265, + 373, + 274 + ], + "spans": [ + { + "bbox": [ + 362, + 265, + 373, + 274 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 274, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 328 + ], + "type": "text", + "content": "Figure 5. Classification accuracy between tail classes and noisy samples versus metrics relevant to class size prediction and few-shot sampling on MVtecAD and VisA datasets with step-like tailed distribution and " + }, + { + "bbox": [ + 55, + 274, + 555, + 328 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 55, + 274, + 555, + 328 + ], + "type": "text", + "content": ". The correlation is strong for (a) mis-sampling ratio, (b) ratio of missing few-shot samples, (e) class size prediction error, and (f) AUROC for predicting whether a sample is few-shot or not. We show that improving the discriminative aspect of network's embeddings improves few-shot sampling and prediction error of class size predictor, which in turn improves (g) anomaly classification (image-level AUROC) and (h) anomaly segmentation (pixel-level AUROC) performance of TailedCore." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 174, + 550, + 259 + ], + "blocks": [ + { + "bbox": [ + 435, + 174, + 550, + 259 + ], + "lines": [ + { + "bbox": [ + 435, + 174, + 550, + 259 + ], + "spans": [ + { + "bbox": [ + 435, + 174, + 550, + 259 + ], + "type": "image", + "image_path": "9ad25f98f81dfe8375ee1b7274baf1e43b4a0313173ca8c87b8ecc06af7045fc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 265, + 498, + 273 + ], + "lines": [ + { + "bbox": [ + 488, + 265, + 498, + 273 + ], + "spans": [ + { + "bbox": [ + 488, + 265, + 498, + 273 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 342, + 129, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 129, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 129, + 355 + ], + "type": "text", + "content": "6.4. Noise ratio" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 422 + ], + "type": "text", + "content": "We analyze the impact of noise ratio. Training data with a greater number of defect samples increase the chance of sampling noise instances for the few-shot sampler. We analyze this trend in Fig. 4, but the impact is not severe. In all cases, TailedCore either outperforms or is on par with both PatchCore and SoftPatch." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 430, + 296, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 456 + ], + "type": "text", + "content": "6.5. Relation between the discriminative aspect of embedding and TailedCore" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "type": "text", + "content": "TailedCore utilizes the embedding features " + }, + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "type": "text", + "content": " of training samples " + }, + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 460, + 295, + 581 + ], + "type": "text", + "content": " for few-shot sampling and augmentation thereof. Therefore, TailedCore's detection performance highly relies on the discriminative aspect of the embedding feature. Poorly designed embedding features will not discriminate different class samples in the embedding space. In this case, TailSampler cannot sample tail class samples and it may overly many head class samples. TailedCore's memory bank obtained so will neither be class-balanced nor noise-free, making the inference of TailedCore poor." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 581, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 640 + ], + "type": "text", + "content": "In contrast, a highly discriminative embedding features will make TailedCore's class size prediction accurate, thereby enabling it to sample the few-shot class samples exclusively without aggregating any anomaly and head class samples. In which case, TailedCore is both noise-free and class-balanced." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "We test this behavior by Oracle labeling and training the embedding features with class labels. Fig. 5 shows that improving discriminative quality of embedding (measured by its classification accuracy on noise and tail class samples) accordingly improves the relevant metrics. Particularly, more discriminative embeddings make more accurate class size" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 343, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 368 + ], + "type": "text", + "content": "prediction, improving TailSampler along with its anomaly classification and segmentation performances." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 380, + 385, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 385, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 385, + 392 + ], + "type": "text", + "content": "7. Limitations" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 400, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 521 + ], + "type": "text", + "content": "The class size prediction in TailSampler can fail if the reflective-symmetric assumption on the inter-class and intra-class similarities breaks down. Particularly, this case can occur if the utilized encoder embedding is either poor or not aligned with the label space of classes. In some cases, the geometric aspect of defect samples highly resembles that of few-shot class instances in the embedding space. We believe that this is due to imperfect nature of encoder embedding, which is often more sensitive to spurious features than the notion of class and object-centric aspect [29]. This issue can be improved by enhancing the encoder network." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 533, + 384, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 384, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 384, + 545 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 554, + 556, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 721 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 721 + ], + "type": "text", + "content": "In this study, we introduced TailedCore, a novel memory-based anomaly detection model designed to effectively address the challenges of unsupervised anomaly detection in noisy and long-tail class distributions. By developing a unique class size predictor and a tailored memory bank, we successfully navigated the tail-versus-noise dilemma, enhancing the model's performance in identifying anomalies within a contaminated and imbalanced dataset. Our comprehensive evaluations demonstrated TailedCore's superior ability to mitigate noise contamination and class imbalance, underscoring its significance in advancing anomaly detection technologies. This work not only showcases a significant stride in handling real-world anomaly detection scenarios but also lays a robust foundation for future research in this critical field." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25546" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "type": "text", + "content": "Acknowledgments Kuan-Chuan Peng was exclusively supported by Mitsubishi Electric Research Laboratories. This work was also supported by the Technology Innovation Program (20023364) funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea), NSF grant 2038493, ONR grant N00014-21-1-2431, NIH grant R01CA240771 from NCI, and U.S. Department of Homeland Security grant 22STESE00001-03-02. The views and conclusions contained in this document are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of the U.S. Department of Homeland Security." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 214, + 112, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 214, + 112, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 214, + 112, + 227 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 234, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 234, + 297, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 234, + 297, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 234, + 297, + 277 + ], + "type": "text", + "content": "[1] Fabrizio Angiulli and Clara Pizzuti. Fast outlier detection in high dimensional spaces. In European conference on principles of data mining and knowledge discovery, pages 15-27. Springer, 2002. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 278, + 296, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 278, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 61, + 278, + 296, + 301 + ], + "type": "text", + "content": "[2] David Arthur, Sergei Vassilvitskii, et al. k-means++: The advantages of careful seeding. In Soda, pages 1027-1035, 2007. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 301, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 301, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 61, + 301, + 296, + 356 + ], + "type": "text", + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 357, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 357, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 61, + 357, + 296, + 401 + ], + "type": "text", + "content": "[4] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4):947-969, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 402, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 445 + ], + "type": "text", + "content": "[5] Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and Jörg Sander. Lof: identifying density-based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data, pages 93-104, 2000. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "text", + "content": "[6] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "type": "text", + "content": "[7] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 579 + ], + "type": "text", + "content": "[8] Zhuohang Chen, Jinglong Chen, Yong Feng, Shen Liu, Tianci Zhang, Kaiyu Zhang, and Wenrong Xiao. Imbalance fault diagnosis under long-tailed distribution: Challenges, solutions and prospects. Knowledge-Based Systems, 258:110008, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 580, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 635 + ], + "type": "text", + "content": "[9] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 297, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 297, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 297, + 679 + ], + "type": "text", + "content": "[10] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological), 39(1):1-22, 1977. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 714 + ], + "type": "text", + "content": "[11] Jan Diers and Christian Pigorsch. A survey of methods for automated quality control based on images. International Journal of Computer Vision, pages 1-29, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 162 + ], + "type": "text", + "content": "[13] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 196 + ], + "type": "text", + "content": "[14] Chaowei Fang, Lechao Cheng, Huiyan Qi, and Dingwen Zhang. Combating noisy labels in long-tailed image classification. arXiv preprint arXiv:2209.00273, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 251 + ], + "type": "text", + "content": "[15] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 285 + ], + "type": "text", + "content": "[16] Brendan J Frey and Delbert Dueck. Clustering by passing messages between data points. science, 315(5814):972-976, 2007. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "text", + "content": "[17] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. arXiv preprint arXiv:2308.15366, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "type": "text", + "content": "[18] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 432 + ], + "type": "text", + "content": "[19] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero-/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 555, + 476 + ], + "type": "text", + "content": "[20] Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai Wu, Yong Liu, Chengjie Wang, and Feng Zheng. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35:15433-15445, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 510 + ], + "type": "text", + "content": "[21] Justin M Johnson and Taghi M Khoshgoftaar. Survey on deep learning with class imbalance. Journal of Big Data, 6(1):1-54, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 512, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 555, + 544 + ], + "type": "text", + "content": "[22] Yujin Lee, Harin Lim, Seoyoon Jang, and Hyunsoo Yoon. Uniformality: Towards task-agnostic unified framework for visual anomaly detection. Available at SSRN 4586132. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 555, + 589 + ], + "type": "text", + "content": "[23] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. arXiv preprint arXiv:2401.16753, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 624 + ], + "type": "text", + "content": "[24] Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 eighth ieee international conference on data mining, pages 413-422. IEEE, 2008. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 669 + ], + "type": "text", + "content": "[25] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 1, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 714 + ], + "type": "text", + "content": "[26] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems, 36, 2023. 2, 5, 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25547" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 117 + ], + "type": "text", + "content": "[27] Yang Lu, Yiliang Zhang, Bo Han, Yiu-ming Cheung, and Hanzi Wang. Label-noise learning with intrinsically long-tailed data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1369-1378, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 119, + 297, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 297, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 297, + 163 + ], + "type": "text", + "content": "[28] Declan McIntosh and Alexandra Branzan Albu. Inter-realization channels: Unsupervised anomaly detection beyond one-class classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6285-6295, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 296, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 296, + 208 + ], + "type": "text", + "content": "[29] Yifei Ming, Hang Yin, and Yixuan Li. On the impact of spurious correlation for out-of-distribution detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10051-10059, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 210, + 295, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 295, + 243 + ], + "type": "text", + "content": "[30] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33 (3):1065-1076, 1962. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 245, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 296, + 300 + ], + "type": "text", + "content": "[31] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 301, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 295, + 334 + ], + "type": "text", + "content": "[32] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 335, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 296, + 392 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 296, + 448 + ], + "type": "text", + "content": "[34] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 296, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 494 + ], + "type": "text", + "content": "[35] Bernhard Schölkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. Advances in neural information processing systems, 12, 1999. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 495, + 296, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 495, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 495, + 296, + 517 + ], + "type": "text", + "content": "[36] Robert L Thorndike. Who belongs in the family? Psychometrika, 18(4):267-276, 1953. 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 518, + 296, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 518, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 518, + 296, + 552 + ], + "type": "text", + "content": "[37] Tong Wei, Jiang-Xin Shi, Wei-Wei Tu, and Yu-Feng Li. Robust long-tailed learning under label noise. arXiv preprint arXiv:2108.11569, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 554, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 296, + 586 + ], + "type": "text", + "content": "[38] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Feng Zheng, and Yaochu Jin. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. arXiv preprint arXiv:2301.12082, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 589, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 296, + 632 + ], + "type": "text", + "content": "[39] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 657 + ], + "type": "text", + "content": "[40] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[41] Manyi Zhang, Xuyang Zhao, Jun Yao, Chun Yuan, and Weiran Huang. When noisy labels meet long tail dilemmas: A representation calibration method. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15890-15900, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 195 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "text", + "content": "[42] Yue Zhao, Zain Nasrullah, and Zheng Li. Pyod: A python toolbox for scalable outlier detection. Journal of Machine Learning Research, 20(96):1-7, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "type": "text", + "content": "[43] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In International Conference on Learning Representations, 2024. 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "type": "text", + "content": "[44] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. arXiv preprint arXiv:2207.14315, 2022. 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25548" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/c1e057f3-058c-42b0-8d35-8e21af099503_content_list.json b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/c1e057f3-058c-42b0-8d35-8e21af099503_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d0881865fc90668c987a614f7e913e5f354abf68 --- /dev/null +++ b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/c1e057f3-058c-42b0-8d35-8e21af099503_content_list.json @@ -0,0 +1,1752 @@ +[ + { + "type": "text", + "text": "Take the Bull by the Horns: Learning to Segment Hard Samples", + "text_level": 1, + "bbox": [ + 174, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuan Guo, Jingyu Kong, Yu Wang, Yuping Duan* \nCenter for Applied Mathematics, Tianjin University, Tianjin 300072, P.R.China \nKey Laboratory of Digital Technology in Medical Diagnostics of Zhejiang Province, 310030, P.R.China \nSchool of Mathematical Sciences, Beijing Normal University, Beijing, 100875, P.R.China", + "bbox": [ + 101, + 180, + 919, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guoyuan6060@163.com;kjyanhui@163.com;flimanadam@gmail.com;doveduan@gmail.com", + "bbox": [ + 174, + 253, + 839, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Medical image segmentation is vital for clinical applications, with hard samples playing a key role in segmentation accuracy. We propose an effective image segmentation framework that includes mechanisms for identifying and segmenting hard samples. It derives a novel image segmentation paradigm: 1) Learning to identify hard samples: automatically selecting inherent hard samples from different datasets, and 2) Learning to segment hard samples: achieving the segmentation of hard samples through effective feature augmentation on dedicated networks. We name our method 'Learning to Segment hard samples' (L2S). The hard sample identification module comprises a backbone model and a classifier, which dynamically uncovers inherent dataset patterns. The hard sample segmentation module utilizes the diffusion process for feature augmentation and incorporates a more sophisticated segmentation network to achieve precise segmentation. We justify our motivation through solid theoretical analysis and extensive experiments. Evaluations across various modalities show that our L2S outperforms other SOTA methods, particularly by substantially improving the segmentation accuracy of hard samples. On ISIC dataset, our L2S improves the Dice score on hard samples and overall segmentation by $8.97\\%$ and $1.01\\%$ , respectively, compared to SOTA methods. The code is available at https://github.com/TqIYuanGie/L2S.", + "bbox": [ + 89, + 335, + 483, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 743, + 220, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image segmentation is vital in medicine for identifying regions of interest to aid diagnosis and treatment planning. While traditional methods struggle with image complexity, deep learning excels in automatically extracting high-level features and capturing intricate patterns.", + "bbox": [ + 89, + 768, + 482, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning-based image segmentation methods can be broadly categorized into end-to-end and coarse-to-fine", + "bbox": [ + 89, + 845, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6673986b253dade01d0b8403716e23931b25430fd9df92f304a093e0a69fbad1.jpg", + "image_caption": [ + "Figure 1. Illustration of Hard samples. Several types of challenges that lead to hard samples are provided, which exhibit a long-tail distribution." + ], + "image_footnote": [], + "bbox": [ + 514, + 301, + 906, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "approaches. End-to-end methods, such as U-Net [30], V-Net [27], and nnU-Net [18], map input images directly to segmentation outputs within a unified framework, often using CNNs as benchmarks. Recent transformer-based models, like TransUNet [5] and Swin-UNet [4], improve performance by capturing long-range dependencies. However, these models struggle with segmenting small anatomical structures or pathological regions where fine details and precise boundaries are crucial.", + "bbox": [ + 511, + 549, + 906, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Coarse-to-fine methods, on the other hand, begin with rough segmentation and progressively refine it, allowing the model to capture broader context before focusing on intricate details [6, 20, 43, 45]. This approach effectively addresses some limitations of end-to-end methods in segmenting small or complex structures. However, these models can still struggle with hard samples in medical image segmentation, as the quality of refinement depends on the coarse predictions, which may fail to capture subtle features essential for complex cases.", + "bbox": [ + 511, + 686, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hard samples, such as heterogeneous tumor regions, vessel bifurcations, stenosis, imaging artifacts (e.g., MRI motion artifacts), and rare data in training sets, present significant challenges for accurate segmentation and often", + "bbox": [ + 511, + 840, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15642", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/be9f2e141694943c93c0c5c3ca82975a3deee320a8dad3ec935914c58a56c786.jpg", + "image_caption": [ + "Figure 2. Compared with the previous two types of methods, our method is more specifically targeted at hard samples." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 473, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hold critical clinical significance; see Figure 1. This paper proposes an effective segmentation framework to improve medical image segmentation, particularly for hard samples. Our model, L2S (Learning to Segment Hard Samples), consists of two components: identifying hard samples and segmenting hard samples. First, we introduce an adaptive method to distinguish normal and hard samples, using a dynamic threshold to identify hard samples for different datasets. Second, we develop a specialized hard sample segmentation network based on a multiscale Transformer network, i.e., Mask2Former [7], where effective feature enhancement and selection methods are introduced to precisely segment hard samples. Additionally, as shown in Figure 2, we introduce an implicit regularization term between hard sample generation and segmentation, where random noise in the diffusion process introduces uncertainty to prevent overfitting and improve model generalization. To sum up, our main contributions can be concluded as follows:", + "bbox": [ + 91, + 295, + 483, + 568 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design an effective segmentation framework that introduces a novel paradigm to enable the model to identify and segment hard samples inherent to the dataset.", + "- We propose to use a diffusion process as feature augmentation, which serves as an implicit regularization to enhance the model's performance and generalization ability in segmenting hard samples.", + "- Our theoretical proof establishes the equivalence between the diffusion process and second-order regularization, offering deeper insights into the superior performance of our method on hard samples.", + "- Our L2S framework demonstrates substantial performance gains across a range of medical segmentation datasets through leveraging dataset-specific characteristics to guide the model in segmenting hard samples." + ], + "bbox": [ + 89, + 571, + 482, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 89, + 814, + 235, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hard sample mining. Hard sample mining has been a central focus in various machine learning tasks and has played a significant role in applications such as classification and detection [9, 22, 31, 33]. It has also made significant ad", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vancements in the field of image segmentation, including layer cascade techniques for segmenting regions of different hardness [23], loss-based identification and retraining of hard regions [44], edge refinement from feature-level image rendering [21], and pixel hardness learning using global and historical loss values [42]. These methods have proven effective in improving model robustness by focusing learning on challenging examples.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, research on hard sample mining in the field of medical image segmentation remains insufficient. Due to the complexity and variability of medical images, which often involve rare pathologies, irregular boundaries, and limited annotated data, the above methods face significant limitations. In addition, Hard samples in medical images often have significant clinical implications. Effectively handling these challenging cases is crucial for clinical applications, and more data-efficient methods are required to learn from rare and precious samples.", + "bbox": [ + 511, + 214, + 903, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data augmentation and model improvements. Data augmentation is an effective way to improve data utilization and model performance. Traditional methods like rotations, flips, and noise addition [32] are homogeneous, while heterogeneous strategies like random proportion mixing [14, 35, 48], random masking [10, 53] and image blending [3, 46] increase data diversity. However, designing proper augmentation for medical images requires considerable domain knowledge and is even impossible for some complicated features. To effectively extract delicate features, various sophisticated architectures are proposed for medical image segmentation. Multi-scale methods [24, 52, 54] capture broad and detailed information across spatial resolutions, improving adaptability to diverse samples. Multimodal designs [39, 50] integrate data from different imaging modalities, compensating for individual limitations. Attention mechanisms [13, 16, 29] focus on key regions, improving precision with hard samples. Uncertainty estimation identifies ambiguous regions, improving robustness. Although these methods effectively enhance overall segmentation performance, the improvements for hard samples are usually significantly less than for easy samples, which is undesirable in medical applications. More importantly, humans can adjust their learning and reasoning strategies when dealing with hard and easy samples, a characteristic not fully leveraged by current methods.", + "bbox": [ + 511, + 369, + 906, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning to learn. Learning to learn is a new paradigm focused on scarce and imbalanced data, including balanced replay [8, 17], dynamic gradient re-balancing [36], feature augmentation [37, 47], and contrastive learning [15, 38, 49, 51]. Among them, DINO [49] enhances hard object detection accuracy by comparing positive samples with challenging negative boxes. The idea of using large-scale negative sampling to improve performance is also a viable solution for hard sample segmentation. Motivated by", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15643", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03637708cea63fadbecf6aa38b007ca719202f4f00b325b151fb9a2387af9399.jpg", + "image_caption": [ + "(b) Learning to segment hard samples", + "(a) Learning to identify hard samples", + "Figure 3. Architecture of our L2S model. (a) Learning to identify hard samples, including a backbone model, a pre-trained classifier, and a post-former module. (b) Learning to segment hard samples, including a feature diffusion and hard sample segmentation module." + ], + "image_footnote": [], + "bbox": [ + 116, + 108, + 880, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "these methods, we aim to design a segmentation architecture that automatically identifies hard samples and adapts its learning and inferencing frameworks accordingly.", + "bbox": [ + 88, + 465, + 482, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our methodology", + "text_level": 1, + "bbox": [ + 89, + 526, + 264, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 3, the proposed L2S has two main components: identifying and segmenting hard samples. First, a backbone model and a pretrained classifier extracts features to identify hard samples. Hard samples are then enhanced through a diffusion process, followed by a specialized segmentation model. By using distinct models for normal and hard samples and focusing on hard sample segmentation, our method improves overall accuracy.", + "bbox": [ + 88, + 551, + 482, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Learning to identify hard samples", + "text_level": 1, + "bbox": [ + 89, + 681, + 387, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hard sample identification. To effectively learn from hard samples, the first step is to establish an approach that automatically assesses the hardness of each sample. In segmentation tasks, it is reasonable to hypothesize that images yielding unsatisfactory segmentation indicate 'hard samples', whereas high-quality segmentation belongs to the 'normal samples'. Accordingly, the most straightforward approach to measure the hardness is to evaluate the quality of segmentation. A simple yet efficient method is to train a neural network that categorizes segmentation outcomes as 'qualified' or 'unqualified'. To this purpose, we develop the network by integrating a segmentation backbone and a pre-trained classifier. During training, the seg", + "bbox": [ + 89, + 703, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mentation results are compared with the ground truth, and the pre-trained classifier is further trained using the obtained classification labels for distinction during testing. The segmentation backbone employs a simple UNet architecture, enhanced with random perturbations in the final decoder layer, allowing the model to simultaneously produce multiple segmentation outcomes for a single image.", + "bbox": [ + 511, + 465, + 906, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train the network, we label the segmentation results as 'qualified' or 'unqualified' based on a Dice score threshold derived from the ground truth. Since this threshold reflects instantaneous hardness, which may change during training, we use Bayesian optimization for dynamic adjustment instead of a fixed threshold. We initialize the threshold to $\\eta^0$ and use the model performance in the validation set as the objective function $h(x; \\theta, \\eta)$ . A Gaussian Process (GP) is used to approximate this objective function", + "bbox": [ + 511, + 573, + 908, + 709 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh (x; \\theta , \\eta) \\sim \\mathcal {G P} (m (\\eta); k (\\eta , \\eta^ {\\prime})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 715, + 903, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $m(\\eta)$ is the mean function, $k(\\eta ,\\eta^{\\prime})$ is the covariance function. The value of $\\eta$ is dynamically updated by maximizing the Expected Improvement (EI) function as follows", + "bbox": [ + 511, + 739, + 906, + 787 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\eta^ {k + 1} = \\arg \\max _ {\\eta} \\left(\\mu (\\eta) - h (\\eta^ {k})\\right) \\Phi (Z) + \\sigma (\\eta) \\phi (Z),\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 792, + 885, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu (\\eta)$ and $\\sigma (\\eta)$ are the predicted mean and standard variation of the objective function, and $\\Phi$ and $\\phi$ are the cumulative distribution function and probability density function with $Z = \\left(\\mu (\\eta) - h(\\eta^{k})\\right) / \\sigma (\\eta)$ . The gradient-descent optimization is used to solve the process. Then, we", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15644", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "reclassify normal and hard samples using the updated $\\eta^{k + 1}$ and repeat the above process in subsequent validation.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Post-Former. We further refine the qualified segmentations associated with normal samples by a Post-Former module, which is a two-layer Transformer. Specifically, the segmentations provide $q$ to access $k$ and $v$ from the shallow image features of the encoder of the segmentation backbone, and the refinement is performed through a cross-attention layer and a self-attention layer.", + "bbox": [ + 89, + 136, + 483, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning to segment hard samples", + "text_level": 1, + "bbox": [ + 89, + 250, + 392, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature augmentation via diffusion process. It is well known that feature augmentation can enhance the generalizability of image segmentation models and improve the performance of hard samples. Wei et al. [40] analyzed the generalization error bound for deep neural network (DNN) models, showing that it is influenced by the Jacobian matrix and the trace of the Hessian matrix of the loss function. Let $\\bar{\\ell} = \\min \\{B,\\ell \\}$ represent the truncation of the loss function to some fixed bound $B > 0$ . Given $P$ as the population data distribution and $P_{n}$ as the distribution over training samples, the following generalization bound holds:", + "bbox": [ + 89, + 272, + 483, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 (Theorem 4.1 in [40]). With probability $1 - \\delta$ over the training examples, for all weight matrices $W$ satisfying the norm bound $\\| W^T\\|_{2,1} \\leq A$ , the following bound holds", + "bbox": [ + 89, + 444, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {P} \\left[ \\bar {\\ell} \\right] - 1. 0 1 \\mathbb {E} _ {P _ {n}} \\left[ \\bar {\\ell} \\right] \\leq \\frac {\\left(A \\mu (W)\\right) ^ {\\frac {2}{3}} (\\theta B) ^ {\\frac {1}{3}}}{n ^ {\\frac {1}{3}}} \\\\ + \\frac {A \\sqrt {B \\nu (W)} \\theta}{\\sqrt {n}} + \\frac {B A ^ {2} \\theta}{n \\left(\\log^ {2} \\left(\\frac {B A ^ {2} \\theta}{\\nu (W) n}\\right) + 1\\right)} + \\zeta , \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 503, + 482, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu (W)$ and $\\nu (W)$ are the Jacobian and Hessian matrix over samples, respectively, $\\theta = \\ln^3 (nM)\\max_i\\| x_i\\| _2^2$ and $\\zeta$ is a low order term.", + "bbox": [ + 89, + 592, + 482, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, we can control both the Jacobian norm and the trace of the Hessian matrix to enhance the generalization of DNN models. While gradient descent primarily focuses on first-order derivatives, it lacks an effective mechanism to constrain the trace of the Hessian. Liu et al. [25] addressed this gap by introducing the trace of the Hessian as an additional penalty term in the loss function", + "bbox": [ + 89, + 646, + 483, + 751 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {a l l} (f (x), y) \\approx \\ell (f (x), y) + \\lambda \\cdot \\operatorname {T r} \\left(H _ {\\ell , W}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 760, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ controls the strength of the Hessian regularization.", + "bbox": [ + 89, + 785, + 482, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we implement feature augmentation using a diffusion model, which is equivalent to implicitly penalizing the trace of the Hessian (see Section 3.3 for details). Specifically, we use random noise to progressively corrupt clean labels $x_0$ into noisy labels $x_{t}$ as", + "bbox": [ + 89, + 801, + 483, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 883, + 372, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,I)$ , $\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_t$ , $\\alpha_t = 1 - \\beta_t$ and $\\beta_t$ is used to parameterize the forward process for $t = 1, \\dots, T$ . The forward process generates a series of noisy labels, enriching the feature space of hard samples. The reverse process iteratively removes noises to estimate clean labels. We use an encoder-decoder network as the denoiser, with $g_\\theta(\\cdot)$ encoding the noise variable and $d_\\theta(\\cdot)$ decoding to predict the $\\epsilon$ . We then have", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = d _ {\\theta} \\left(x _ {t}; m _ {\\theta} \\left(c _ {f}, g _ {\\theta} (x _ {t})\\right), t\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 223, + 828, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the conditional image embedding $c_{f}$ extracted from the backbone model is fused with the noise embedding $g_{\\theta}(x_{t})$ in mix-former module $m_{\\theta}(\\cdot)$ . After training, the diffusion model's decoder is discarded, and its gradients guide the encoder to better represent noisy labels as augmented features for hard sample segmentation. Notably, applying a similar diffusion procedure to normal samples yields minimal improvements, suggesting that the optimization strategy for samples of varying hardness should differ.", + "bbox": [ + 511, + 250, + 906, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mix-Former. For a segmentation task, the diffusion process must be conditional on the original image, which the proposed Mix-Former module implements. It integrates the noise embedding, hereafter referred to as $g(\\epsilon)$ for clarity, and the conditional image embedding $c_{f}$ obtained by the segmentation backbone model. In particular, Mix-Former consists of $N$ (with $N = 3$ in our implementation) transformer blocks, each containing two cross-attention modules. We define the basic form of the attention mechanism with domain transformation as", + "bbox": [ + 511, + 398, + 905, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {A t t n} (q, k) = F ^ {- 1} \\left[ \\operatorname {s o f t m a x} \\left(\\frac {\\left(F (q) \\mathcal {W} _ {q}\\right) \\cdot \\left(F (k) \\mathcal {W} _ {k}\\right) ^ {T}}{\\sqrt {d _ {k}}}\\right) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 542, + 895, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{W}$ is the projection matrix in Fourier space, $F$ and $F^{-1}$ are the Fast Fourier Transform (FFT) and inverse transform, respectively. Here, frequency domain learning can enhances the model's ability to capture complex dependencies between noise and semantic information, thereby improving performance.", + "bbox": [ + 511, + 580, + 905, + 672 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/28000a14b79c4654dac1cc502672780191a05c5f70749cbc2422549c8f0c6f51.jpg", + "image_caption": [ + "Figure 4. The structure of Mix-Former, where $g(\\epsilon)$ and $c_{f}$ are fused in the mix-former through $N$ layers of repeated attention." + ], + "image_footnote": [], + "bbox": [ + 524, + 685, + 895, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 4, the first module integrates $g(\\epsilon)$ into $c_{f}^{0}$ , while the second module encodes the embedding $c_{f}^{1}$ back into $g(\\epsilon)$ . The bidirectional encoding enables the", + "bbox": [ + 511, + 853, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$g(\\varepsilon)$", + "bbox": [ + 728, + 904, + 754, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15645", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model to learn the interaction between noise and semantic features to obtain a more robust representation [41].", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hard sample segmentation. Hard samples exhibit low-quality segmentation that deviates significantly from their labels. Thus, we concatenate the unqualified segmentation with its original image and use an encoder and pixel decoder to extract multi-scale features. A learnable object query accesses these features via cross-attention in the Transformer decoder. Augmented features from the diffusion model further enhance feature extraction. The learnable query dynamically selects relevant information for precise segmentation of hard samples. The Transformer decoder contains four cross-attention layers; see Figure 5. The first three layers fuse the learnable query with features from three scales of the pixel decoder, while the fourth layer integrates augmented features from the diffusion process to enhance segmentation. The details are as follows", + "bbox": [ + 89, + 136, + 483, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} = \\left\\{ \\begin{array}{l l} \\operatorname {s o f t m a x} \\left(\\frac {\\left(W _ {q} L _ {q}\\right) \\cdot \\left(W _ {k} M _ {f} ^ {i}\\right) ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot W _ {v} M _ {f} ^ {i}, & i = 1, 2, 3, \\\\ F _ {A t t n} (z _ {3}, g (\\epsilon)) \\cdot W _ {v} g _ {\\natural} (\\epsilon), & i = 4, \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 369, + 486, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{q}$ represents the learnable query, $M_{f}^{i}$ represents the multi-scale features by scaling with 32, 16 and $\\S$ respectively, and $z_{i}$ represents the attention-weighted feature. Such pattern is repeated in a round-robin fashion for all subsequent layers.", + "bbox": [ + 89, + 431, + 483, + 507 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e5254675850841c516dd3a5f6eed51bb9309efb7d5d718bd1fcb8fcaa0f0f201.jpg", + "image_caption": [ + "Figure 5. Hard samples segmentation decoder, where $L_{q}$ access both multi-scale features $M_{f}$ and augmented features $g(\\epsilon)$ ." + ], + "image_footnote": [], + "bbox": [ + 101, + 500, + 473, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Implicit regularization", + "text_level": 1, + "bbox": [ + 89, + 698, + 302, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we formally demonstrate that the diffusion process in our model works as an implicit regularization to improve the model's generalization ability to hard samples. We aim to find the hard sample segmentation network $f(x;\\theta)$ that minimizes the expected loss $\\mathcal{L} = \\mathbb{E}[\\ell (f(x;\\theta),y)]$ . Then we have the following result", + "bbox": [ + 89, + 720, + 483, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1. For $\\epsilon \\sim \\mathcal{N}(0,I)$ , let $g(\\epsilon)$ denote the augmented features obtained by the diffusion probabilistic model. Then the expected loss becomes", + "bbox": [ + 89, + 820, + 483, + 866 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ \\ell (f (x + g (\\epsilon)), y) ] \\approx \\ell (f (x), y) + \\frac {1}{2} \\operatorname {T r} (H \\Sigma_ {g}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 875, + 483, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. By incorporating the augmented features extracted from the noisy data, we have the expected loss function as", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} [ \\ell (f (x + g (\\epsilon)), y) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 131, + 797, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which can be approximated by Taylor's expansion", + "bbox": [ + 511, + 157, + 846, + 172 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\approx \\mathbb {E} \\big [ \\ell (f (x), y) + \\nabla_ {x} \\ell (f (x), y) \\cdot g (\\epsilon) + \\frac {1}{2} g (\\epsilon) ^ {T} H g (\\epsilon) \\big ],\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 181, + 906, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\nabla_{x}\\ell (f(x),y)$ and $H$ to be the gradient and Hessian of the loss function w.r.t. the input $x$ . Since $\\epsilon$ is of zero mean, by assuming that $g$ is center distributed, i.e., $\\mathbb{E}[g(\\epsilon)] = 0$ , we have", + "bbox": [ + 511, + 217, + 906, + 277 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ \\nabla_ {x} \\ell (f (x), y) \\cdot g (\\epsilon) ] = \\mathbb {E} [ g (\\epsilon) ] \\cdot \\nabla_ {x} \\ell (f (x), y) = 0.\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 287, + 887, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the properties of matrix trace, there is", + "bbox": [ + 511, + 313, + 828, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ g (\\epsilon) ^ {T} H g (\\epsilon) ] = \\mathbb {E} [ \\operatorname {T r} (H g (\\epsilon) g (\\epsilon) ^ {T}) ] = \\operatorname {T r} (H \\mathbb {E} [ g (\\epsilon) g (\\epsilon) ^ {T} ]),\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 334, + 895, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assuming the covariance matrix of $g(\\epsilon)$ is $\\Sigma_g$ , we have $\\Sigma_g = \\mathbb{E}[g(\\epsilon)g(\\epsilon)^T]$ . It follows that", + "bbox": [ + 511, + 359, + 905, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ g (\\epsilon) ^ {T} H g (\\epsilon) \\right] = \\operatorname {T r} \\left(H \\Sigma_ {g}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 398, + 803, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the expected loss function becomes", + "bbox": [ + 513, + 426, + 792, + 440 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ \\ell (f (x + g (\\epsilon)), y) ] \\approx \\ell (f (x), y) + \\frac {1}{2} \\operatorname {T r} (H \\Sigma_ {g}),\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 450, + 874, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which completes the proof.", + "bbox": [ + 513, + 486, + 696, + 501 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/51e90a82c225beb8a6662c886bcfddaa33d1d3a5e1b94861c003c3f343be5a7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 888, + 487, + 905, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By penalizing the trace of the Hessian matrix, the model is driven towards flatter minima, where the loss function is relatively stable, thereby improving generalization for hard sample segmentation [25].", + "bbox": [ + 511, + 511, + 906, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss function", + "text_level": 1, + "bbox": [ + 513, + 580, + 653, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion loss. The Diffusion Loss measures the difference between the ground truth noise added to the label and the noise predicted by the diffusion model", + "bbox": [ + 511, + 602, + 905, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} = \\left\\| \\epsilon - d _ {\\theta} \\left(t, x _ {t}, c _ {f}\\right) \\right\\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 655, + 903, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_{f}$ is the conditional image embedding, $\\epsilon$ is the noise, and $d_{\\theta}(t,x_t,c_f)$ is the noise predicted at time step $t$ .", + "bbox": [ + 511, + 683, + 903, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segmentation loss. For hard samples, the segmentation loss $\\mathcal{L}_{hard} = \\mathcal{L}_{mask} + \\lambda_{cls}\\mathcal{L}_{cls}$ combines mask loss and classification loss. The mask loss includes binary cross-entropy and dice loss: $\\mathcal{L}_{mask} = \\lambda_{bce}\\mathcal{L}_{bce} + \\lambda_{dice}\\mathcal{L}_{dice}$ , while the classification loss is cross-entropy. For normal samples, the segmentation loss $\\mathcal{L}_{normal}$ uses only the mask loss. The total segmentation loss is:", + "bbox": [ + 511, + 714, + 905, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e g}} = \\lambda_ {\\text {h a r d}} \\mathcal {L} _ {\\text {h a r d}} + \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 829, + 903, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with hyperparameters set as $\\lambda_{bce} = \\lambda_{dice} = \\lambda_{cls} = 5$ , $\\lambda_{hard} = 5$ , and $\\lambda_{normal} = 2$ . The dice loss for the backbone model is weighted at 5, while the binary cross-entropy", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15646", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/93778f2cf1288b5b97313419ed06beae3b41a2578756557c8582e2efd47e242c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCOSMOS 2022ISIC 2018Kvasir-SEGBUSIAverage
DSCIoUDSC†DSCIoUDSC†DSCIoUDSC†DSCIoUDSC†DSCIoU
UNet[30]82.4472.7641.7687.1479.7342.2985.6378.6748.3576.4167.6141.0882.9074.69
PolypPVT[11]82.3372.3434.4389.8482.6152.4691.4486.3960.2681.0572.5146.6286.1778.46
nnUNet[18]84.1275.5855.7988.9181.6546.3190.5585.1853.4780.9472.4342.8786.1378.71
MedT[34]82.8173.1539.2688.8481.4944.0589.6884.1251.2380.4471.2740.1985.4477.51
TransUNet[5]82.9273.1843.6889.4482.3751.7691.0485.7556.7380.3271.3845.3185.9378.17
TransFuses[50]83.1773.8945.3589.9683.2254.9391.2786.3158.6481.5073.2548.2586.4879.17
SwinUNet[4]83.6374.7748.1789.5682.4049.5790.2285.0951.1879.7670.1939.7385.7978.11
MedSegDiffV2[41]82.8972.9339.5489.7882.6551.1291.0686.2152.3080.5371.4442.7486.0778.31
Mask2former[7]82.7672.8241.3289.5982.4649.5190.8685.2853.7780.1970.6343.3685.8577.80
L2S84.1675.6566.6790.9784.3463.9091.8787.0264.3481.8273.2753.2787.2180.07
", + "bbox": [ + 91, + 89, + 916, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of L2S with SOTA segmentation methods over different image modalities, where the results of SOTA methods are reproduced using their publicly available implementation with five-fold cross validation. $\\mathrm{DSC}^{\\dagger}$ is computed on the hard samples with dice score below 0.7. We add the standard deviation and statistical significance in suppl. materials.", + "bbox": [ + 89, + 289, + 906, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "loss for the classifier is weighted at 10.", + "bbox": [ + 89, + 358, + 349, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistency loss. To ensure consistency between normal and hard samples, which share the same label, we introduce the following consistency loss:", + "bbox": [ + 89, + 373, + 482, + 419 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n s}} = \\left\\| p \\left(\\mathcal {D} _ {n}\\right) - f \\left(\\mathcal {D} _ {h}\\right) \\right\\| _ {2} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 431, + 482, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the post-former $p(\\cdot)$ fine-tunes the normal sample set, and $f(\\cdot)$ represents the segmentation network for the hard samples. To sum up, the total loss for our model is the weighted sum of all the aforementioned terms as", + "bbox": [ + 89, + 460, + 483, + 522 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {s e g}} + \\lambda_ {\\text {d i f f}} \\mathcal {L} _ {\\text {d i f f}} + \\lambda_ {\\text {c o n s}} \\mathcal {L} _ {\\text {c o n s}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 536, + 482, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the weights are fixed as $\\lambda_{diff} = 5$ and $\\lambda_{cons} = 5$ .", + "bbox": [ + 89, + 564, + 465, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental results", + "text_level": 1, + "bbox": [ + 89, + 595, + 290, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and implementation details", + "text_level": 1, + "bbox": [ + 89, + 619, + 405, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted experiments on six public medical segmentation datasets to evaluate the performance of our model across various imaging modalities. These include the COSMOS2022 dataset [55], ISIC2018 dataset [28], Kvasir-SEG dataset [19], and the BUSI dataset [1] for carotid artery, skin lesion, polyp and breast ultrasound segmentation, respectively. The BTCV dataset [12], containing 12 anatomical structures, was used for abdominal multi-organ segmentation, while the ACDC dataset [2] was utilized for segmenting cardiac structures in MRI images.", + "bbox": [ + 88, + 642, + 482, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the hard sample segmentation network, we used a pre-trained Swin Transformer [26] as the encoder and MSDeformAttn [56] as the pixel decoder, similar to Mask2Former [7]. The Transformer decoder was configured with $L = 3$ (12 layers in total) and 100 queries by default. We trained our model using the AdamW optimizer with a learning rate of 0.01 and a weight decay of $1e^{-4}$ .", + "bbox": [ + 88, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training images were resized to $224 \\times 224$ for the COSMOS2022, ACDC, and BTCV datasets. For other datasets, the images were resized to $448 \\times 448$ with color jittering and random horizontal flipping. The training schedule included 12,000 iterations for COSMOS2022, 30,000 for ACDC and BTCV, and 20,000 for other datasets. All experiments were conducted using PyTorch 1.10.0 on 4 NVIDIA RTX 3090 GPUs. The Dice score with an initial threshold of 0.7 was used to distinguish between normal and hard samples. During training, normal and hard samples generated by the backbone model were stored separately. For each input image, a pair of normal and hard samples was randomly selected from their respective sets for further segmentation. Each image's two sets contain at least one element to ensure the balance.", + "bbox": [ + 511, + 358, + 906, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compared our L2S with SOTA segmentation methods, including CNN-based methods like UNet [30], nnUNet [18], and transformer-based methods such as Polyp-PVT [11], MedT [34], TransUNet [5], TransFuse [50], SwinUNet [4], and the diffusion-based method MedSegDiffV2 [41]. We also compared L2S with the baseline model Mask2Former [7] effectiveness of the proposed framework.", + "bbox": [ + 511, + 587, + 908, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results evaluation", + "text_level": 1, + "bbox": [ + 511, + 709, + 689, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons on binary segmentation tasks. Our L2S model is compared with SOTA segmentation methods across four tasks; see Table 1, where the best results are shown in bold, and the same is applied in the following tables. The results demonstrate that our L2S model achieves the highest average Dice of $87.21\\%$ and IoU of $80.07\\%$ . Improving hard sample segmentation yields significant gains: L2S outperforms the SOTA methods by 10. $88\\%$ in COSMOS2022, $8.97\\%$ in ISIC2018, and $4.08\\%$ in the polyp-specific PolypPVT model in Kvasir-SEG. These results highlight that the existing methods perform poorly", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15647", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/58e3803bfcbfef9e5f0711f6533a9d382ba5dd9b6d26a6c45f6bf00bfb11e5e6.jpg", + "image_caption": [ + "Figure 6. Qualitative segmentation results for hard samples, including carotid artery, breast ultrasound, polyp, and skin lesions." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 898, + 330 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a8d351744400e1e8ce3520a963947af6e269fde0e1b47d1ce4f86c8fac866e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSpleenR.KidL.KidGall.Eso.LiverStom.AortaIVCVeinsPanc.AGAvg.
UNet[30]0.9330.8790.9130.7120.7080.9390.8200.8570.7850.7070.6810.6370.797
nnUNet[18]0.9420.8940.9100.7040.7230.9480.8240.8770.7820.7200.6800.6160.802
TransUNet[5]0.9520.9270.9290.6620.7570.9690.8890.9200.8330.7910.7750.6370.837
TransFuses[50]0.9690.9240.9310.7720.7880.9740.9030.9140.8360.8020.7850.7410.861
SwinUNet[4]0.9390.9160.9080.7270.7340.9410.8520.9030.7980.7550.7330.6460.821
MedSegDiffV2[41]0.9480.9250.9160.7940.8220.9530.8910.9180.8390.7980.7910.7270.860
Mask2former[7]0.9560.9130.9210.7340.7550.9510.8920.9110.8350.7950.7530.6860.842
L2S0.9730.9340.9450.8260.8050.9720.9180.9220.8460.8240.8130.8020.882
", + "bbox": [ + 96, + 368, + 901, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on hard samples. Our L2S achieves over a $10\\%$ improvement in accuracy compared to Mask2Former. A detailed analysis of the selected hard samples is provided in Supplementary Material B.4 to validate our assumption that hard samples correspond to poor segmentation. Figure 6 presents the segmentation results of representative hard samples. Our L2S model excels at segmenting complex structures. Qualitative results for carotid arteries, breast ultrasound, polyps, and skin lesions show that the predictions of our L2S better match the ground truth. For carotid artery segmentation, our L2S model precisely identifies stenosis, offering important clinical value. Overall, our designs significantly improve robustness and generalization. The visualized results are available in the supplementary materials.", + "bbox": [ + 88, + 566, + 482, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on multi-organ segmentation. Table 2 shows that our L2S outperforms the SOTA methods on BTCV multi-organ dataset, achieving the highest average Dice score of $88.2\\%$ . Table 3 lists the Dice scores of L2S and the comparison methods on ACDC dataset for cardiac organ segmentation. Our L2S gains an average Dice score of $91.36\\%$ , outperforming nnUNet by $0.33\\%$ and improving right ventricle segmentation by $0.48\\%$ . The comparison", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8dcf6587506b6848bcbbe5f176347d144a04b93bef6113b14dbff3c5f9ffaeb9.jpg", + "table_caption": [ + "Table 2. Comparison of L2S with SOTA segmentation methods over multi-organ segmentation BTCV dataset in term of Dice Score." + ], + "table_footnote": [], + "table_body": "
MethodsRVMyoLVAvg.
UNet86.7980.4593.6386.96
nnUNet89.5988.3195.2091.03
TransUNet87.0587.1794.5289.58
TransFuses87.9487.0395.0990.02
SwinUNet86.5785.3394.1088.67
Mask2former86.7486.4993.9289.05
L2S(UNet)90.0788.6595.3691.36
L2S(Mask2former)90.4289.0295.6191.68
", + "bbox": [ + 522, + 564, + 898, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results of cardiac organ segmentation on ACDC dataset, where dice scores are reported for individual organs.", + "bbox": [ + 511, + 717, + 905, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "experiments on different backbone models show that more powerful Mask2Former can further enhance segmentation accuracy. Thus, the L2S framework is highly compatible and can be integrated with different segmentation models.", + "bbox": [ + 511, + 775, + 906, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 511, + 847, + 676, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted a series of ablation studies on the ISIC2018 dataset unless otherwise noted. Table 4 shows an ablation", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15648", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "study on FFT, consistency loss, and adaptive threshold selection. As shown, FFT significantly impacts the Dice score with an improvement of $0.31\\%$ , highlighting the importance of feature alignment.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89e7104f9d85cef90f83db47a3320a9d17ce894ac35b97146a0da9940971dd26.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDSC (%)IoU (%)
L2S90.9784.34
- FFT90.66 (-0.3)83.95 (-0.4)
- consistency loss90.78 (-0.2)84.07 (-0.3)
- adaptive threshold90.86 (-0.1)84.24 (-0.1)
- all 3 components90.54 (-0.4)83.72 (-0.6)
", + "bbox": [ + 109, + 164, + 464, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hard sample identification and augmentation. Table 5 presents an ablation study evaluating the impact of incorporating normal and hard sample identification mechanisms and feature augmentation modules on model performance. The results show that removing the identification mechanism (i.e., no longer distinguishing the segmentation results of the backbone model, treating all samples in the dataset as hard samples for subsequent processing, and therefore using no classifier during testing) causes a decrease in the Dice score from $90.87\\%$ to $89.59\\%$ , which indicates that complex processing may adversely affect the original normal samples in the dataset. In addition, using diffusion models for feature augmentation significantly improves the hard sample segmentation, increasing the Dice score by $4.52\\%$ . Incorporating the Post-former module slightly enhances segmentation by refining normal samples.", + "bbox": [ + 88, + 338, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dcdca98fcef5f52924f2583565fecc56e22b06d7ba7645ac3f1970397b1ea56d.jpg", + "table_caption": [ + "Table 4. Three optimization improvements in network operation, where we remove one component at a time." + ], + "table_footnote": [], + "table_body": "
CLFDDPMP-FormerDSC (%)IoU (%)DSC† (%)
89.5282.35-
✓90.2883.7458.96
✓89.5982.48-
✓✓90.8784.11-
✓✓90.3683.7959.38
✓✓✓90.9784.3463.90
", + "bbox": [ + 91, + 593, + 488, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hard sample segmentation. Table 6 shows the impact of various design choices on the performance of the hard sample segmentation network, measured by the Dice score. Using query vectors alongside hard sample features boost the Dice score from $55.83\\%$ to $59.88\\%$ . Besides, multiscale features and query learnability can further enhance hard sample segmentation accuracy, by increasing Dice of more than $1.5\\%$ and $3\\%$ , respectively. We also compare the", + "bbox": [ + 89, + 779, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6d977e3d75d6af65d2aad1866dc07762f5b617243cdabb7eb530dfebf4146dee.jpg", + "table_caption": [ + "Table 5. Ablation study on our L2S method, where 'CLF', 'DDPM' and 'P-Former' denote the pretrained classifier, feature augmentation, and Post-Former, respectively." + ], + "table_footnote": [], + "table_body": "
MethodDSC† (%)FLOPs
single scale (w/o query)55.8315.2G
single scale62.2432.7G
- learnable query features59.8832.7G
L2S63.9024.3G
- learnable query features60.8224.3G
", + "bbox": [ + 524, + 88, + 893, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FLOPs of different designs for hard sample segmentation on the ISIC dataset. The multiscale approach achieved better performance with fewer FLOPs.", + "bbox": [ + 511, + 257, + 903, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Discussion", + "text_level": 1, + "bbox": [ + 511, + 320, + 632, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a specialized framework for hard sample segmentation in medical images, advancing current methods. Our approach includes a dedicated sample generator and a classifier to distinguish normal from hard samples. In addition, our results also indicate that classifier accuracy significantly influences overall model performance, and substituting direct labels during testing can yield additional gains; see Table 7. While not the primary focus, these enhancements suggest valuable directions for future research, particularly in improving model learning and segmentation on complex medical image datasets.", + "bbox": [ + 511, + 344, + 906, + 512 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69a2ffb1547af5a539a51addc41df3292e8b00725ec9a7d366b93bc55a0b4691.jpg", + "table_caption": [ + "Table 6. Ablation study on Learnable query and Mutil-scale features, where 'single-scale' is the single resolution of $1/8$ ." + ], + "table_footnote": [], + "table_body": "
DatasetAcc (%)Rec (%)DSC (%)DSC* (%)
COSMOS202292.782.384.1684.34
ISIC201883.372.690.9791.28
BUSI78.670.481.8290.04
Kvasir-SEG85.975.291.8792.10
ACDC87.571.791.3691.45
", + "bbox": [ + 514, + 527, + 910, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. The impact of the classifier to segmentation, where Acc and Rec denote the overall accuracy and recall of hard samples, respectively. $\\mathrm{DSC}^*$ is obtained by classification using true labels.", + "bbox": [ + 511, + 638, + 906, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Concluding remarks", + "text_level": 1, + "bbox": [ + 511, + 720, + 710, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel segmentation framework that can identify hard samples and enhance segmentation performance by employing stronger networks and more sophisticated optimization strategies on selected hard samples. The framework incorporates a diffusion process as an implicit regularization to handle hard samples, achieved by leveraging the inherent distribution bias of the dataset combined with random noise. Extensive experiments across various datasets demonstrated the superiority of our method over SOTA approaches, particularly for challenging samples.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15649", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 89, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 174, + 187, + 189 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6", + "[2] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514-2525, 2018. 6", + "[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 2", + "[4] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 1, 6, 7", + "[5] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 6, 7", + "[6] Mengzhao Chen, Mingbao Lin, Ke Li, Yunhang Shen, Yongjian Wu, Fei Chao, and Rongrong Ji. Cf-vit: A general coarse-to-fine method for vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7042-7052, 2023. 1", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 6, 7", + "[8] Jiequan Cui, Yuhui Yuan, Zhisheng Zhong, Zhuotao Tian, Han Hu, Stephen Lin, and Jiaya Jia. Region rebalance for long-tailed semantic segmentation. arXiv preprint arXiv:2204.01969, 2022. 2", + "[9] Yin Cui, Feng Zhou, Yuanqing Lin, and Serge Belongie. Fine-grained categorization and dataset bootstrapping using deep metric learning with humans in the loop. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1153-1162, 2016. 2", + "[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 2", + "[11] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 6" + ], + "bbox": [ + 93, + 199, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "The work is supported by the National Natural Science Foundation of China (NSFC 12071345).", + "[12] Xi Fang and Pingkun Yan. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging, 39(11):3619–3629, 2020. 6", + "[13] Jun Fu, Jing Liu, Hajjie Tian, et al. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3146-3154, 2019. 2", + "[14] Ethan Harris, Antonia Marcu, Matthew Painter, Mahesan Niranjan, Adam Prügel-Bennett, and Jonathon Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020. 2", + "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7132-7141, 2018. 2", + "[17] Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to segment the tail. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14045-14054, 2020. 2", + "[18] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 1, 6, 7", + "[19] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pál Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26, pages 451–462. Springer, 2020. 6", + "[20] Longlong Jing, Yucheng Chen, and Yingli Tian. Coarse-to-fine semantic segmentation from image-level labels. IEEE transactions on image processing, 29:225-236, 2019. 1", + "[21] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 2", + "[22] Aybora Koksal, Önder Tuzcuoglu, Kutalmış Gökalp İnce, Yoldaş Ataseven, and A Aydin Alatan. Improved hard example mining approach for single shot object detectors. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3536-3540. IEEE, 2022. 2", + "[23] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3202, 2017. 2", + "[24] Xuhang Lian, Yanwei Pang, Jungong Han, and Jing Pan. Cascaded hierarchical atrous spatial pyramid pooling module for semantic segmentation. Pattern Recognition, 110: 107622, 2021. 2" + ], + "bbox": [ + 93, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15650", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Yucong Liu, Shixing Yu, and Tong Lin. Regularizing deep neural networks with stochastic estimators of hessian trace. arXiv preprint arXiv:2208.05924, 2022. 4, 5", + "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6", + "[27] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 1", + "[28] Md Ashraful Alam Milton. Automated skin lesion classification using ensemble of deep neural networks in isic 2018: Skin lesion analysis towards melanoma detection challenge. arXiv preprint arXiv:1901.10802, 2019. 6", + "[29] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 2", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 6, 7", + "[31] Seong-Eun Ryu and Kyung-Yong Chung. Detection model of occluded object based on yolo using hard-example mining and augmentation policy optimization. Applied Sciences, 11 (15):7093, 2021. 2", + "[32] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 2", + "[33] Abhinav Shrivastava, Rahul Sukthankar, and Mohammad Hejrati. Training region-based object detectors with online hard example mining. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 761-769, 2016. 2", + "[34] Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Hacihaliloglu, and Vishal M Patel. Medical transformer: Gated axial-attention for medical image segmentation. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, part I 24, pages 36-46. Springer, 2021. 6", + "[35] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In International conference on machine learning, pages 6438-6447. PMLR, 2019. 2", + "[36] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9695-9704, 2021. 2" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model-agnostic segmentation refinement with discrete diffusion process. arXiv preprint arXiv:2312.12425, 2023. 2", + "[38] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7303-7313, 2021. 2", + "[39] Xue Wang, Zhanshan Li, Yongping Huang, and Yingying Jiao. Multimodal medical image segmentation using multiscale context-aware network. Neurocomputing, 486:135-146, 2022. 2", + "[40] Colin Wei, Sham Kakade, and Tengyu Ma. The implicit and explicit regularization effects of dropout. In International conference on machine learning, pages 10181-10192. PMLR, 2020. 4", + "[41] Junde Wu, Wei Ji, Huazhu Fu, Min Xu, Yueming Jin, and Yanwu Xu. Medsegdiff-v2: Diffusion-based medical image segmentation with transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6030-6038, 2024. 5, 6, 7", + "[42] Xin Xiao, Daiguo Zhou, Jiagao Hu, Yi Hu, and Yongchao Xu. Not all pixels are equal: Learning pixel hardness for semantic segmentation. arXiv preprint arXiv:2305.08462, 2023. 2", + "[43] Jiahao Xie, Sheng Zhang, Jianwei Lu, and Ye Luo. L-snet: From region localization to scale invariant medical image segmentation. In 2021 IEEE International Conference on Image Processing (ICIP), pages 41-45. IEEE, 2021. 1", + "[44] Jin Yin, Pengfei Xia, and Jingsong He. Online hard region mining for semantic segmentation. Neural Processing Letters, 50:2665-2679, 2019. 2", + "[45] Qihang Yu, Dong Yang, Holger Roth, Yutong Bai, Yixiao Zhang, Alan L Yuille, and Daguang Xu. C2fnas: Coarse-to-fine neural architecture search for 3d medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4126-4135, 2020. 1", + "[46] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2", + "[47] Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3457-3466, 2021. 2", + "[48] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2", + "[49] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15651", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, Part I 24, pages 14-24. Springer, 2021. 2, 6, 7", + "[51] Xiangyun Zhao, Raviteja Vemulapalli, Philip Andrew Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive learning for label efficient semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10623-10633, 2021. 2", + "[52] Xiaoqi Zhao, Hongpeng Jia, Youwei Pang, Long Lv, Feng Tian, Lihe Zhang, Weibing Sun, and Huchuan Lu. M $^{\\wedge}\\{2\\}$ snet: Multi-scale in multi-scale subtraction network for medical image segmentation. arXiv preprint arXiv:2303.10894, 2023. 2", + "[53] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13001-13008, 2020. 2", + "[54] Tongxue Zhou, Su Ruan, and Stephane Canu. A review: Deep learning for medical image segmentation using multimodality fusion. *Array*, 3:100004, 2019. 2", + "[55] Chenglu Zhu, Xiaoyan Wang, Shengyong Chen, Zhongzhao Teng, Cong Bai, Xiaojie Huang, Ming Xia, Zhanpeng Shao, Zheng Gu, and Peiliang Sun. Complex carotid artery segmentation in multi-contrast mr sequences by improved optimal surface graph cuts based on flow line learning. Medical & Biological Engineering & Computing, 60(9):2693-2706, 2022. 6", + "[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 6" + ], + "bbox": [ + 91, + 92, + 480, + 583 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15652", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/c1e057f3-058c-42b0-8d35-8e21af099503_model.json b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/c1e057f3-058c-42b0-8d35-8e21af099503_model.json new file mode 100644 index 0000000000000000000000000000000000000000..056a67a25daf21522d9af9739c29fffccb91b080 --- /dev/null +++ b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/c1e057f3-058c-42b0-8d35-8e21af099503_model.json @@ -0,0 +1,2411 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "Take the Bull by the Horns: Learning to Segment Hard Samples" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.181, + 0.92, + 0.252 + ], + "angle": 0, + "content": "Yuan Guo, Jingyu Kong, Yu Wang, Yuping Duan* \nCenter for Applied Mathematics, Tianjin University, Tianjin 300072, P.R.China \nKey Laboratory of Digital Technology in Medical Diagnostics of Zhejiang Province, 310030, P.R.China \nSchool of Mathematical Sciences, Beijing Normal University, Beijing, 100875, P.R.China" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.255, + 0.841, + 0.27 + ], + "angle": 0, + "content": "guoyuan6060@163.com;kjyanhui@163.com;flimanadam@gmail.com;doveduan@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.485, + 0.715 + ], + "angle": 0, + "content": "Medical image segmentation is vital for clinical applications, with hard samples playing a key role in segmentation accuracy. We propose an effective image segmentation framework that includes mechanisms for identifying and segmenting hard samples. It derives a novel image segmentation paradigm: 1) Learning to identify hard samples: automatically selecting inherent hard samples from different datasets, and 2) Learning to segment hard samples: achieving the segmentation of hard samples through effective feature augmentation on dedicated networks. We name our method 'Learning to Segment hard samples' (L2S). The hard sample identification module comprises a backbone model and a classifier, which dynamically uncovers inherent dataset patterns. The hard sample segmentation module utilizes the diffusion process for feature augmentation and incorporates a more sophisticated segmentation network to achieve precise segmentation. We justify our motivation through solid theoretical analysis and extensive experiments. Evaluations across various modalities show that our L2S outperforms other SOTA methods, particularly by substantially improving the segmentation accuracy of hard samples. On ISIC dataset, our L2S improves the Dice score on hard samples and overall segmentation by \\(8.97\\%\\) and \\(1.01\\%\\), respectively, compared to SOTA methods. The code is available at https://github.com/TqIYuanGie/L2S." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.744, + 0.222, + 0.76 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.846 + ], + "angle": 0, + "content": "Image segmentation is vital in medicine for identifying regions of interest to aid diagnosis and treatment planning. While traditional methods struggle with image complexity, deep learning excels in automatically extracting high-level features and capturing intricate patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.847, + 0.483, + 0.876 + ], + "angle": 0, + "content": "Deep learning-based image segmentation methods can be broadly categorized into end-to-end and coarse-to-fine" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.303, + 0.907, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.518 + ], + "angle": 0, + "content": "Figure 1. Illustration of Hard samples. Several types of challenges that lead to hard samples are provided, which exhibit a long-tail distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.686 + ], + "angle": 0, + "content": "approaches. End-to-end methods, such as U-Net [30], V-Net [27], and nnU-Net [18], map input images directly to segmentation outputs within a unified framework, often using CNNs as benchmarks. Recent transformer-based models, like TransUNet [5] and Swin-UNet [4], improve performance by capturing long-range dependencies. However, these models struggle with segmenting small anatomical structures or pathological regions where fine details and precise boundaries are crucial." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Coarse-to-fine methods, on the other hand, begin with rough segmentation and progressively refine it, allowing the model to capture broader context before focusing on intricate details [6, 20, 43, 45]. This approach effectively addresses some limitations of end-to-end methods in segmenting small or complex structures. However, these models can still struggle with hard samples in medical image segmentation, as the quality of refinement depends on the coarse predictions, which may fail to capture subtle features essential for complex cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Hard samples, such as heterogeneous tumor regions, vessel bifurcations, stenosis, imaging artifacts (e.g., MRI motion artifacts), and rare data in training sets, present significant challenges for accurate segmentation and often" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15642" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.475, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.241, + 0.483, + 0.269 + ], + "angle": 0, + "content": "Figure 2. Compared with the previous two types of methods, our method is more specifically targeted at hard samples." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.296, + 0.485, + 0.569 + ], + "angle": 0, + "content": "hold critical clinical significance; see Figure 1. This paper proposes an effective segmentation framework to improve medical image segmentation, particularly for hard samples. Our model, L2S (Learning to Segment Hard Samples), consists of two components: identifying hard samples and segmenting hard samples. First, we introduce an adaptive method to distinguish normal and hard samples, using a dynamic threshold to identify hard samples for different datasets. Second, we develop a specialized hard sample segmentation network based on a multiscale Transformer network, i.e., Mask2Former [7], where effective feature enhancement and selection methods are introduced to precisely segment hard samples. Additionally, as shown in Figure 2, we introduce an implicit regularization term between hard sample generation and segmentation, where random noise in the diffusion process introduces uncertainty to prevent overfitting and improve model generalization. To sum up, our main contributions can be concluded as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.573, + 0.483, + 0.617 + ], + "angle": 0, + "content": "- We design an effective segmentation framework that introduces a novel paradigm to enable the model to identify and segment hard samples inherent to the dataset." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.618, + 0.483, + 0.678 + ], + "angle": 0, + "content": "- We propose to use a diffusion process as feature augmentation, which serves as an implicit regularization to enhance the model's performance and generalization ability in segmenting hard samples." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.679, + 0.483, + 0.738 + ], + "angle": 0, + "content": "- Our theoretical proof establishes the equivalence between the diffusion process and second-order regularization, offering deeper insights into the superior performance of our method on hard samples." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.739, + 0.483, + 0.8 + ], + "angle": 0, + "content": "- Our L2S framework demonstrates substantial performance gains across a range of medical segmentation datasets through leveraging dataset-specific characteristics to guide the model in segmenting hard samples." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.573, + 0.483, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.236, + 0.83 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Hard sample mining. Hard sample mining has been a central focus in various machine learning tasks and has played a significant role in applications such as classification and detection [9, 22, 31, 33]. It has also made significant ad" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "vancements in the field of image segmentation, including layer cascade techniques for segmenting regions of different hardness [23], loss-based identification and retraining of hard regions [44], edge refinement from feature-level image rendering [21], and pixel hardness learning using global and historical loss values [42]. These methods have proven effective in improving model robustness by focusing learning on challenging examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.905, + 0.367 + ], + "angle": 0, + "content": "However, research on hard sample mining in the field of medical image segmentation remains insufficient. Due to the complexity and variability of medical images, which often involve rare pathologies, irregular boundaries, and limited annotated data, the above methods face significant limitations. In addition, Hard samples in medical images often have significant clinical implications. Effectively handling these challenging cases is crucial for clinical applications, and more data-efficient methods are required to learn from rare and precious samples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.37, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Data augmentation and model improvements. Data augmentation is an effective way to improve data utilization and model performance. Traditional methods like rotations, flips, and noise addition [32] are homogeneous, while heterogeneous strategies like random proportion mixing [14, 35, 48], random masking [10, 53] and image blending [3, 46] increase data diversity. However, designing proper augmentation for medical images requires considerable domain knowledge and is even impossible for some complicated features. To effectively extract delicate features, various sophisticated architectures are proposed for medical image segmentation. Multi-scale methods [24, 52, 54] capture broad and detailed information across spatial resolutions, improving adaptability to diverse samples. Multimodal designs [39, 50] integrate data from different imaging modalities, compensating for individual limitations. Attention mechanisms [13, 16, 29] focus on key regions, improving precision with hard samples. Uncertainty estimation identifies ambiguous regions, improving robustness. Although these methods effectively enhance overall segmentation performance, the improvements for hard samples are usually significantly less than for easy samples, which is undesirable in medical applications. More importantly, humans can adjust their learning and reasoning strategies when dealing with hard and easy samples, a characteristic not fully leveraged by current methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Learning to learn. Learning to learn is a new paradigm focused on scarce and imbalanced data, including balanced replay [8, 17], dynamic gradient re-balancing [36], feature augmentation [37, 47], and contrastive learning [15, 38, 49, 51]. Among them, DINO [49] enhances hard object detection accuracy by comparing positive samples with challenging negative boxes. The idea of using large-scale negative sampling to improve performance is also a viable solution for hard sample segmentation. Motivated by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15643" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.092, + 0.394, + 0.109 + ], + "angle": 0, + "content": "(b) Learning to segment hard samples" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.109, + 0.882, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.383, + 0.388, + 0.399 + ], + "angle": 0, + "content": "(a) Learning to identify hard samples" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.412, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 3. Architecture of our L2S model. (a) Learning to identify hard samples, including a backbone model, a pre-trained classifier, and a post-former module. (b) Learning to segment hard samples, including a feature diffusion and hard sample segmentation module." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.467, + 0.483, + 0.514 + ], + "angle": 0, + "content": "these methods, we aim to design a segmentation architecture that automatically identifies hard samples and adapts its learning and inferencing frameworks accordingly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.265, + 0.544 + ], + "angle": 0, + "content": "3. Our methodology" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.483, + 0.675 + ], + "angle": 0, + "content": "As shown in Figure 3, the proposed L2S has two main components: identifying and segmenting hard samples. First, a backbone model and a pretrained classifier extracts features to identify hard samples. Hard samples are then enhanced through a diffusion process, followed by a specialized segmentation model. By using distinct models for normal and hard samples and focusing on hard sample segmentation, our method improves overall accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.388, + 0.699 + ], + "angle": 0, + "content": "3.1. Learning to identify hard samples" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Hard sample identification. To effectively learn from hard samples, the first step is to establish an approach that automatically assesses the hardness of each sample. In segmentation tasks, it is reasonable to hypothesize that images yielding unsatisfactory segmentation indicate 'hard samples', whereas high-quality segmentation belongs to the 'normal samples'. Accordingly, the most straightforward approach to measure the hardness is to evaluate the quality of segmentation. A simple yet efficient method is to train a neural network that categorizes segmentation outcomes as 'qualified' or 'unqualified'. To this purpose, we develop the network by integrating a segmentation backbone and a pre-trained classifier. During training, the seg" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.573 + ], + "angle": 0, + "content": "mentation results are compared with the ground truth, and the pre-trained classifier is further trained using the obtained classification labels for distinction during testing. The segmentation backbone employs a simple UNet architecture, enhanced with random perturbations in the final decoder layer, allowing the model to simultaneously produce multiple segmentation outcomes for a single image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.909, + 0.71 + ], + "angle": 0, + "content": "To train the network, we label the segmentation results as 'qualified' or 'unqualified' based on a Dice score threshold derived from the ground truth. Since this threshold reflects instantaneous hardness, which may change during training, we use Bayesian optimization for dynamic adjustment instead of a fixed threshold. We initialize the threshold to \\(\\eta^0\\) and use the model performance in the validation set as the objective function \\(h(x; \\theta, \\eta)\\). A Gaussian Process (GP) is used to approximate this objective function" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.717, + 0.905, + 0.734 + ], + "angle": 0, + "content": "\\[\nh (x; \\theta , \\eta) \\sim \\mathcal {G P} (m (\\eta); k (\\eta , \\eta^ {\\prime})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.908, + 0.788 + ], + "angle": 0, + "content": "where \\(m(\\eta)\\) is the mean function, \\(k(\\eta ,\\eta^{\\prime})\\) is the covariance function. The value of \\(\\eta\\) is dynamically updated by maximizing the Expected Improvement (EI) function as follows" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.793, + 0.887, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\eta^ {k + 1} = \\arg \\max _ {\\eta} \\left(\\mu (\\eta) - h (\\eta^ {k})\\right) \\Phi (Z) + \\sigma (\\eta) \\phi (Z),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mu (\\eta)\\) and \\(\\sigma (\\eta)\\) are the predicted mean and standard variation of the objective function, and \\(\\Phi\\) and \\(\\phi\\) are the cumulative distribution function and probability density function with \\(Z = \\left(\\mu (\\eta) - h(\\eta^{k})\\right) / \\sigma (\\eta)\\). The gradient-descent optimization is used to solve the process. Then, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15644" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.123 + ], + "angle": 0, + "content": "reclassify normal and hard samples using the updated \\(\\eta^{k + 1}\\) and repeat the above process in subsequent validation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.484, + 0.244 + ], + "angle": 0, + "content": "Post-Former. We further refine the qualified segmentations associated with normal samples by a Post-Former module, which is a two-layer Transformer. Specifically, the segmentations provide \\( q \\) to access \\( k \\) and \\( v \\) from the shallow image features of the encoder of the segmentation backbone, and the refinement is performed through a cross-attention layer and a self-attention layer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.251, + 0.393, + 0.267 + ], + "angle": 0, + "content": "3.2. Learning to segment hard samples" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.439 + ], + "angle": 0, + "content": "Feature augmentation via diffusion process. It is well known that feature augmentation can enhance the generalizability of image segmentation models and improve the performance of hard samples. Wei et al. [40] analyzed the generalization error bound for deep neural network (DNN) models, showing that it is influenced by the Jacobian matrix and the trace of the Hessian matrix of the loss function. Let \\(\\bar{\\ell} = \\min \\{B,\\ell \\}\\) represent the truncation of the loss function to some fixed bound \\(B > 0\\). Given \\(P\\) as the population data distribution and \\(P_{n}\\) as the distribution over training samples, the following generalization bound holds:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.485, + 0.501 + ], + "angle": 0, + "content": "Theorem 1 (Theorem 4.1 in [40]). With probability \\(1 - \\delta\\) over the training examples, for all weight matrices \\(W\\) satisfying the norm bound \\(\\| W^T\\|_{2,1} \\leq A\\), the following bound holds" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.505, + 0.483, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {P} \\left[ \\bar {\\ell} \\right] - 1. 0 1 \\mathbb {E} _ {P _ {n}} \\left[ \\bar {\\ell} \\right] \\leq \\frac {\\left(A \\mu (W)\\right) ^ {\\frac {2}{3}} (\\theta B) ^ {\\frac {1}{3}}}{n ^ {\\frac {1}{3}}} \\\\ + \\frac {A \\sqrt {B \\nu (W)} \\theta}{\\sqrt {n}} + \\frac {B A ^ {2} \\theta}{n \\left(\\log^ {2} \\left(\\frac {B A ^ {2} \\theta}{\\nu (W) n}\\right) + 1\\right)} + \\zeta , \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.639 + ], + "angle": 0, + "content": "where \\(\\mu (W)\\) and \\(\\nu (W)\\) are the Jacobian and Hessian matrix over samples, respectively, \\(\\theta = \\ln^3 (nM)\\max_i\\| x_i\\| _2^2\\) and \\(\\zeta\\) is a low order term." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.752 + ], + "angle": 0, + "content": "Thus, we can control both the Jacobian norm and the trace of the Hessian matrix to enhance the generalization of DNN models. While gradient descent primarily focuses on first-order derivatives, it lacks an effective mechanism to constrain the trace of the Hessian. Liu et al. [25] addressed this gap by introducing the trace of the Hessian as an additional penalty term in the loss function" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.761, + 0.483, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\ell_ {a l l} (f (x), y) \\approx \\ell (f (x), y) + \\lambda \\cdot \\operatorname {T r} \\left(H _ {\\ell , W}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.483, + 0.801 + ], + "angle": 0, + "content": "where \\(\\lambda\\) controls the strength of the Hessian regularization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.484, + 0.877 + ], + "angle": 0, + "content": "In this work, we implement feature augmentation using a diffusion model, which is equivalent to implicitly penalizing the trace of the Hessian (see Section 3.3 for details). Specifically, we use random noise to progressively corrupt clean labels \\( x_0 \\) into noisy labels \\( x_{t} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.885, + 0.374, + 0.902 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.212 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,I)\\), \\(\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_t\\), \\(\\alpha_t = 1 - \\beta_t\\) and \\(\\beta_t\\) is used to parameterize the forward process for \\(t = 1, \\dots, T\\). The forward process generates a series of noisy labels, enriching the feature space of hard samples. The reverse process iteratively removes noises to estimate clean labels. We use an encoder-decoder network as the denoiser, with \\(g_\\theta(\\cdot)\\) encoding the noise variable and \\(d_\\theta(\\cdot)\\) decoding to predict the \\(\\epsilon\\). We then have" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.224, + 0.829, + 0.24 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = d _ {\\theta} \\left(x _ {t}; m _ {\\theta} \\left(c _ {f}, g _ {\\theta} (x _ {t})\\right), t\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.907, + 0.386 + ], + "angle": 0, + "content": "where the conditional image embedding \\( c_{f} \\) extracted from the backbone model is fused with the noise embedding \\( g_{\\theta}(x_{t}) \\) in mix-former module \\( m_{\\theta}(\\cdot) \\). After training, the diffusion model's decoder is discarded, and its gradients guide the encoder to better represent noisy labels as augmented features for hard sample segmentation. Notably, applying a similar diffusion procedure to normal samples yields minimal improvements, suggesting that the optimization strategy for samples of varying hardness should differ." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.906, + 0.538 + ], + "angle": 0, + "content": "Mix-Former. For a segmentation task, the diffusion process must be conditional on the original image, which the proposed Mix-Former module implements. It integrates the noise embedding, hereafter referred to as \\( g(\\epsilon) \\) for clarity, and the conditional image embedding \\( c_{f} \\) obtained by the segmentation backbone model. In particular, Mix-Former consists of \\( N \\) (with \\( N = 3 \\) in our implementation) transformer blocks, each containing two cross-attention modules. We define the basic form of the attention mechanism with domain transformation as" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.544, + 0.897, + 0.577 + ], + "angle": 0, + "content": "\\[\nF _ {A t t n} (q, k) = F ^ {- 1} \\left[ \\operatorname {s o f t m a x} \\left(\\frac {\\left(F (q) \\mathcal {W} _ {q}\\right) \\cdot \\left(F (k) \\mathcal {W} _ {k}\\right) ^ {T}}{\\sqrt {d _ {k}}}\\right) \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.582, + 0.906, + 0.673 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}\\) is the projection matrix in Fourier space, \\(F\\) and \\(F^{-1}\\) are the Fast Fourier Transform (FFT) and inverse transform, respectively. Here, frequency domain learning can enhances the model's ability to capture complex dependencies between noise and semantic information, thereby improving performance." + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.686, + 0.897, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.81, + 0.907, + 0.841 + ], + "angle": 0, + "content": "Figure 4. The structure of Mix-Former, where \\( g(\\epsilon) \\) and \\( c_{f} \\) are fused in the mix-former through \\( N \\) layers of repeated attention." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.854, + 0.906, + 0.904 + ], + "angle": 0, + "content": "As shown in Figure 4, the first module integrates \\( g(\\epsilon) \\) into \\( c_{f}^{0} \\), while the second module encodes the embedding \\( c_{f}^{1} \\) back into \\( g(\\epsilon) \\). The bidirectional encoding enables the" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.905, + 0.756, + 0.918 + ], + "angle": 0, + "content": "\\(g(\\varepsilon)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15645" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "model to learn the interaction between noise and semantic features to obtain a more robust representation [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.484, + 0.362 + ], + "angle": 0, + "content": "Hard sample segmentation. Hard samples exhibit low-quality segmentation that deviates significantly from their labels. Thus, we concatenate the unqualified segmentation with its original image and use an encoder and pixel decoder to extract multi-scale features. A learnable object query accesses these features via cross-attention in the Transformer decoder. Augmented features from the diffusion model further enhance feature extraction. The learnable query dynamically selects relevant information for precise segmentation of hard samples. The Transformer decoder contains four cross-attention layers; see Figure 5. The first three layers fuse the learnable query with features from three scales of the pixel decoder, while the fourth layer integrates augmented features from the diffusion process to enhance segmentation. The details are as follows" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.371, + 0.487, + 0.424 + ], + "angle": 0, + "content": "\\[\nz _ {i} = \\left\\{ \\begin{array}{l l} \\operatorname {s o f t m a x} \\left(\\frac {\\left(W _ {q} L _ {q}\\right) \\cdot \\left(W _ {k} M _ {f} ^ {i}\\right) ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot W _ {v} M _ {f} ^ {i}, & i = 1, 2, 3, \\\\ F _ {A t t n} (z _ {3}, g (\\epsilon)) \\cdot W _ {v} g _ {\\natural} (\\epsilon), & i = 4, \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.508 + ], + "angle": 0, + "content": "where \\(L_{q}\\) represents the learnable query, \\(M_{f}^{i}\\) represents the multi-scale features by scaling with 32, 16 and \\(\\S\\) respectively, and \\(z_{i}\\) represents the attention-weighted feature. Such pattern is repeated in a round-robin fashion for all subsequent layers." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.501, + 0.475, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Figure 5. Hard samples segmentation decoder, where \\( L_{q} \\) access both multi-scale features \\( M_{f} \\) and augmented features \\( g(\\epsilon) \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.699, + 0.303, + 0.715 + ], + "angle": 0, + "content": "3.3. Implicit regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.484, + 0.813 + ], + "angle": 0, + "content": "Subsequently, we formally demonstrate that the diffusion process in our model works as an implicit regularization to improve the model's generalization ability to hard samples. We aim to find the hard sample segmentation network \\( f(x;\\theta) \\) that minimizes the expected loss \\( \\mathcal{L} = \\mathbb{E}[\\ell (f(x;\\theta),y)] \\). Then we have the following result" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Proposition 1. For \\(\\epsilon \\sim \\mathcal{N}(0,I)\\), let \\(g(\\epsilon)\\) denote the augmented features obtained by the diffusion probabilistic model. Then the expected loss becomes" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.876, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ \\ell (f (x + g (\\epsilon)), y) ] \\approx \\ell (f (x), y) + \\frac {1}{2} \\operatorname {T r} (H \\Sigma_ {g}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "Proof. By incorporating the augmented features extracted from the noisy data, we have the expected loss function as" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.132, + 0.799, + 0.149 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} [ \\ell (f (x + g (\\epsilon)), y) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.158, + 0.847, + 0.173 + ], + "angle": 0, + "content": "which can be approximated by Taylor's expansion" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.182, + 0.907, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\approx \\mathbb {E} \\big [ \\ell (f (x), y) + \\nabla_ {x} \\ell (f (x), y) \\cdot g (\\epsilon) + \\frac {1}{2} g (\\epsilon) ^ {T} H g (\\epsilon) \\big ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.218, + 0.907, + 0.278 + ], + "angle": 0, + "content": "with \\(\\nabla_{x}\\ell (f(x),y)\\) and \\(H\\) to be the gradient and Hessian of the loss function w.r.t. the input \\(x\\). Since \\(\\epsilon\\) is of zero mean, by assuming that \\(g\\) is center distributed, i.e., \\(\\mathbb{E}[g(\\epsilon)] = 0\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.289, + 0.888, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ \\nabla_ {x} \\ell (f (x), y) \\cdot g (\\epsilon) ] = \\mathbb {E} [ g (\\epsilon) ] \\cdot \\nabla_ {x} \\ell (f (x), y) = 0.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.314, + 0.83, + 0.329 + ], + "angle": 0, + "content": "Based on the properties of matrix trace, there is" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.335, + 0.897, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ g (\\epsilon) ^ {T} H g (\\epsilon) ] = \\mathbb {E} [ \\operatorname {T r} (H g (\\epsilon) g (\\epsilon) ^ {T}) ] = \\operatorname {T r} (H \\mathbb {E} [ g (\\epsilon) g (\\epsilon) ^ {T} ]),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.361, + 0.906, + 0.392 + ], + "angle": 0, + "content": "Assuming the covariance matrix of \\( g(\\epsilon) \\) is \\( \\Sigma_g \\), we have \\( \\Sigma_g = \\mathbb{E}[g(\\epsilon)g(\\epsilon)^T] \\). It follows that" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.4, + 0.805, + 0.418 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ g (\\epsilon) ^ {T} H g (\\epsilon) \\right] = \\operatorname {T r} \\left(H \\Sigma_ {g}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.427, + 0.793, + 0.441 + ], + "angle": 0, + "content": "Then, the expected loss function becomes" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.451, + 0.875, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ \\ell (f (x + g (\\epsilon)), y) ] \\approx \\ell (f (x), y) + \\frac {1}{2} \\operatorname {T r} (H \\Sigma_ {g}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.487, + 0.697, + 0.502 + ], + "angle": 0, + "content": "which completes the proof." + }, + { + "type": "image", + "bbox": [ + 0.889, + 0.488, + 0.906, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.573 + ], + "angle": 0, + "content": "By penalizing the trace of the Hessian matrix, the model is driven towards flatter minima, where the loss function is relatively stable, thereby improving generalization for hard sample segmentation [25]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.581, + 0.655, + 0.595 + ], + "angle": 0, + "content": "3.4. Loss function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.648 + ], + "angle": 0, + "content": "Diffusion loss. The Diffusion Loss measures the difference between the ground truth noise added to the label and the noise predicted by the diffusion model" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.656, + 0.905, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} = \\left\\| \\epsilon - d _ {\\theta} \\left(t, x _ {t}, c _ {f}\\right) \\right\\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.904, + 0.714 + ], + "angle": 0, + "content": "where \\(c_{f}\\) is the conditional image embedding, \\(\\epsilon\\) is the noise, and \\(d_{\\theta}(t,x_t,c_f)\\) is the noise predicted at time step \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.82 + ], + "angle": 0, + "content": "Segmentation loss. For hard samples, the segmentation loss \\(\\mathcal{L}_{hard} = \\mathcal{L}_{mask} + \\lambda_{cls}\\mathcal{L}_{cls}\\) combines mask loss and classification loss. The mask loss includes binary cross-entropy and dice loss: \\(\\mathcal{L}_{mask} = \\lambda_{bce}\\mathcal{L}_{bce} + \\lambda_{dice}\\mathcal{L}_{dice}\\), while the classification loss is cross-entropy. For normal samples, the segmentation loss \\(\\mathcal{L}_{normal}\\) uses only the mask loss. The total segmentation loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.83, + 0.905, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e g}} = \\lambda_ {\\text {h a r d}} \\mathcal {L} _ {\\text {h a r d}} + \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "with hyperparameters set as \\(\\lambda_{bce} = \\lambda_{dice} = \\lambda_{cls} = 5\\), \\(\\lambda_{hard} = 5\\), and \\(\\lambda_{normal} = 2\\). The dice loss for the backbone model is weighted at 5, while the binary cross-entropy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15646" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.09, + 0.917, + 0.281 + ], + "angle": 0, + "content": "
MethodCOSMOS 2022ISIC 2018Kvasir-SEGBUSIAverage
DSCIoUDSC†DSCIoUDSC†DSCIoUDSC†DSCIoUDSC†DSCIoU
UNet[30]82.4472.7641.7687.1479.7342.2985.6378.6748.3576.4167.6141.0882.9074.69
PolypPVT[11]82.3372.3434.4389.8482.6152.4691.4486.3960.2681.0572.5146.6286.1778.46
nnUNet[18]84.1275.5855.7988.9181.6546.3190.5585.1853.4780.9472.4342.8786.1378.71
MedT[34]82.8173.1539.2688.8481.4944.0589.6884.1251.2380.4471.2740.1985.4477.51
TransUNet[5]82.9273.1843.6889.4482.3751.7691.0485.7556.7380.3271.3845.3185.9378.17
TransFuses[50]83.1773.8945.3589.9683.2254.9391.2786.3158.6481.5073.2548.2586.4879.17
SwinUNet[4]83.6374.7748.1789.5682.4049.5790.2285.0951.1879.7670.1939.7385.7978.11
MedSegDiffV2[41]82.8972.9339.5489.7882.6551.1291.0686.2152.3080.5371.4442.7486.0778.31
Mask2former[7]82.7672.8241.3289.5982.4649.5190.8685.2853.7780.1970.6343.3685.8577.80
L2S84.1675.6566.6790.9784.3463.9091.8787.0264.3481.8273.2753.2787.2180.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.29, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Table 1. Comparison of L2S with SOTA segmentation methods over different image modalities, where the results of SOTA methods are reproduced using their publicly available implementation with five-fold cross validation. \\( \\mathrm{DSC}^{\\dagger} \\) is computed on the hard samples with dice score below 0.7. We add the standard deviation and statistical significance in suppl. materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.351, + 0.374 + ], + "angle": 0, + "content": "loss for the classifier is weighted at 10." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.42 + ], + "angle": 0, + "content": "Consistency loss. To ensure consistency between normal and hard samples, which share the same label, we introduce the following consistency loss:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.432, + 0.483, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n s}} = \\left\\| p \\left(\\mathcal {D} _ {n}\\right) - f \\left(\\mathcal {D} _ {h}\\right) \\right\\| _ {2} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.523 + ], + "angle": 0, + "content": "where the post-former \\( p(\\cdot) \\) fine-tunes the normal sample set, and \\( f(\\cdot) \\) represents the segmentation network for the hard samples. To sum up, the total loss for our model is the weighted sum of all the aforementioned terms as" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.537, + 0.483, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {s e g}} + \\lambda_ {\\text {d i f f}} \\mathcal {L} _ {\\text {d i f f}} + \\lambda_ {\\text {c o n s}} \\mathcal {L} _ {\\text {c o n s}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.466, + 0.582 + ], + "angle": 0, + "content": "where the weights are fixed as \\(\\lambda_{diff} = 5\\) and \\(\\lambda_{cons} = 5\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.596, + 0.291, + 0.613 + ], + "angle": 0, + "content": "4. Experimental results" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.621, + 0.406, + 0.638 + ], + "angle": 0, + "content": "4.1. Datasets and implementation details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.795 + ], + "angle": 0, + "content": "We conducted experiments on six public medical segmentation datasets to evaluate the performance of our model across various imaging modalities. These include the COSMOS2022 dataset [55], ISIC2018 dataset [28], Kvasir-SEG dataset [19], and the BUSI dataset [1] for carotid artery, skin lesion, polyp and breast ultrasound segmentation, respectively. The BTCV dataset [12], containing 12 anatomical structures, was used for abdominal multi-organ segmentation, while the ACDC dataset [2] was utilized for segmenting cardiac structures in MRI images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For the hard sample segmentation network, we used a pre-trained Swin Transformer [26] as the encoder and MSDeformAttn [56] as the pixel decoder, similar to Mask2Former [7]. The Transformer decoder was configured with \\( L = 3 \\) (12 layers in total) and 100 queries by default. We trained our model using the AdamW optimizer with a learning rate of 0.01 and a weight decay of \\( 1e^{-4} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Training images were resized to \\(224 \\times 224\\) for the COSMOS2022, ACDC, and BTCV datasets. For other datasets, the images were resized to \\(448 \\times 448\\) with color jittering and random horizontal flipping. The training schedule included 12,000 iterations for COSMOS2022, 30,000 for ACDC and BTCV, and 20,000 for other datasets. All experiments were conducted using PyTorch 1.10.0 on 4 NVIDIA RTX 3090 GPUs. The Dice score with an initial threshold of 0.7 was used to distinguish between normal and hard samples. During training, normal and hard samples generated by the backbone model were stored separately. For each input image, a pair of normal and hard samples was randomly selected from their respective sets for further segmentation. Each image's two sets contain at least one element to ensure the balance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.909, + 0.695 + ], + "angle": 0, + "content": "We compared our L2S with SOTA segmentation methods, including CNN-based methods like UNet [30], nnUNet [18], and transformer-based methods such as Polyp-PVT [11], MedT [34], TransUNet [5], TransFuse [50], SwinUNet [4], and the diffusion-based method MedSegDiffV2 [41]. We also compared L2S with the baseline model Mask2Former [7] effectiveness of the proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.691, + 0.725 + ], + "angle": 0, + "content": "4.2. Results evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparisons on binary segmentation tasks. Our L2S model is compared with SOTA segmentation methods across four tasks; see Table 1, where the best results are shown in bold, and the same is applied in the following tables. The results demonstrate that our L2S model achieves the highest average Dice of \\(87.21\\%\\) and IoU of \\(80.07\\%\\). Improving hard sample segmentation yields significant gains: L2S outperforms the SOTA methods by 10. \\(88\\%\\) in COSMOS2022, \\(8.97\\%\\) in ISIC2018, and \\(4.08\\%\\) in the polyp-specific PolypPVT model in Kvasir-SEG. These results highlight that the existing methods perform poorly" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15647" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.899, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.34, + 0.873, + 0.356 + ], + "angle": 0, + "content": "Figure 6. Qualitative segmentation results for hard samples, including carotid artery, breast ultrasound, polyp, and skin lesions." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.369, + 0.903, + 0.516 + ], + "angle": 0, + "content": "
MethodSpleenR.KidL.KidGall.Eso.LiverStom.AortaIVCVeinsPanc.AGAvg.
UNet[30]0.9330.8790.9130.7120.7080.9390.8200.8570.7850.7070.6810.6370.797
nnUNet[18]0.9420.8940.9100.7040.7230.9480.8240.8770.7820.7200.6800.6160.802
TransUNet[5]0.9520.9270.9290.6620.7570.9690.8890.9200.8330.7910.7750.6370.837
TransFuses[50]0.9690.9240.9310.7720.7880.9740.9030.9140.8360.8020.7850.7410.861
SwinUNet[4]0.9390.9160.9080.7270.7340.9410.8520.9030.7980.7550.7330.6460.821
MedSegDiffV2[41]0.9480.9250.9160.7940.8220.9530.8910.9180.8390.7980.7910.7270.860
Mask2former[7]0.9560.9130.9210.7340.7550.9510.8920.9110.8350.7950.7530.6860.842
L2S0.9730.9340.9450.8260.8050.9720.9180.9220.8460.8240.8130.8020.882
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.526, + 0.888, + 0.541 + ], + "angle": 0, + "content": "Table 2. Comparison of L2S with SOTA segmentation methods over multi-organ segmentation BTCV dataset in term of Dice Score." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.568, + 0.483, + 0.779 + ], + "angle": 0, + "content": "on hard samples. Our L2S achieves over a \\(10\\%\\) improvement in accuracy compared to Mask2Former. A detailed analysis of the selected hard samples is provided in Supplementary Material B.4 to validate our assumption that hard samples correspond to poor segmentation. Figure 6 presents the segmentation results of representative hard samples. Our L2S model excels at segmenting complex structures. Qualitative results for carotid arteries, breast ultrasound, polyps, and skin lesions show that the predictions of our L2S better match the ground truth. For carotid artery segmentation, our L2S model precisely identifies stenosis, offering important clinical value. Overall, our designs significantly improve robustness and generalization. The visualized results are available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparisons on multi-organ segmentation. Table 2 shows that our L2S outperforms the SOTA methods on BTCV multi-organ dataset, achieving the highest average Dice score of \\(88.2\\%\\). Table 3 lists the Dice scores of L2S and the comparison methods on ACDC dataset for cardiac organ segmentation. Our L2S gains an average Dice score of \\(91.36\\%\\), outperforming nnUNet by \\(0.33\\%\\) and improving right ventricle segmentation by \\(0.48\\%\\). The comparison" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.565, + 0.899, + 0.709 + ], + "angle": 0, + "content": "
MethodsRVMyoLVAvg.
UNet86.7980.4593.6386.96
nnUNet89.5988.3195.2091.03
TransUNet87.0587.1794.5289.58
TransFuses87.9487.0395.0990.02
SwinUNet86.5785.3394.1088.67
Mask2former86.7486.4993.9289.05
L2S(UNet)90.0788.6595.3691.36
L2S(Mask2former)90.4289.0295.6191.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.718, + 0.906, + 0.748 + ], + "angle": 0, + "content": "Table 3. Results of cardiac organ segmentation on ACDC dataset, where dice scores are reported for individual organs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.776, + 0.907, + 0.837 + ], + "angle": 0, + "content": "experiments on different backbone models show that more powerful Mask2Former can further enhance segmentation accuracy. Thus, the L2S framework is highly compatible and can be integrated with different segmentation models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.677, + 0.863 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We conducted a series of ablation studies on the ISIC2018 dataset unless otherwise noted. Table 4 shows an ablation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15648" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "study on FFT, consistency loss, and adaptive threshold selection. As shown, FFT significantly impacts the Dice score with an improvement of \\(0.31\\%\\), highlighting the importance of feature alignment." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.165, + 0.465, + 0.284 + ], + "angle": 0, + "content": "
MethodDSC (%)IoU (%)
L2S90.9784.34
- FFT90.66 (-0.3)83.95 (-0.4)
- consistency loss90.78 (-0.2)84.07 (-0.3)
- adaptive threshold90.86 (-0.1)84.24 (-0.1)
- all 3 components90.54 (-0.4)83.72 (-0.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.322 + ], + "angle": 0, + "content": "Table 4. Three optimization improvements in network operation, where we remove one component at a time." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.483, + 0.581 + ], + "angle": 0, + "content": "Hard sample identification and augmentation. Table 5 presents an ablation study evaluating the impact of incorporating normal and hard sample identification mechanisms and feature augmentation modules on model performance. The results show that removing the identification mechanism (i.e., no longer distinguishing the segmentation results of the backbone model, treating all samples in the dataset as hard samples for subsequent processing, and therefore using no classifier during testing) causes a decrease in the Dice score from \\(90.87\\%\\) to \\(89.59\\%\\), which indicates that complex processing may adversely affect the original normal samples in the dataset. In addition, using diffusion models for feature augmentation significantly improves the hard sample segmentation, increasing the Dice score by \\(4.52\\%\\). Incorporating the Post-former module slightly enhances segmentation by refining normal samples." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.594, + 0.489, + 0.711 + ], + "angle": 0, + "content": "
CLFDDPMP-FormerDSC (%)IoU (%)DSC† (%)
89.5282.35-
✓90.2883.7458.96
✓89.5982.48-
✓✓90.8784.11-
✓✓90.3683.7959.38
✓✓✓90.9784.3463.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Table 5. Ablation study on our L2S method, where 'CLF', 'DDPM' and 'P-Former' denote the pretrained classifier, feature augmentation, and Post-Former, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Hard sample segmentation. Table 6 shows the impact of various design choices on the performance of the hard sample segmentation network, measured by the Dice score. Using query vectors alongside hard sample features boost the Dice score from \\(55.83\\%\\) to \\(59.88\\%\\). Besides, multiscale features and query learnability can further enhance hard sample segmentation accuracy, by increasing Dice of more than \\(1.5\\%\\) and \\(3\\%\\), respectively. We also compare the" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.089, + 0.895, + 0.188 + ], + "angle": 0, + "content": "
MethodDSC† (%)FLOPs
single scale (w/o query)55.8315.2G
single scale62.2432.7G
- learnable query features59.8832.7G
L2S63.9024.3G
- learnable query features60.8224.3G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Table 6. Ablation study on Learnable query and Mutil-scale features, where 'single-scale' is the single resolution of \\(1/8\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.304 + ], + "angle": 0, + "content": "FLOPs of different designs for hard sample segmentation on the ISIC dataset. The multiscale approach achieved better performance with fewer FLOPs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.321, + 0.633, + 0.336 + ], + "angle": 0, + "content": "4.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.513 + ], + "angle": 0, + "content": "We introduce a specialized framework for hard sample segmentation in medical images, advancing current methods. Our approach includes a dedicated sample generator and a classifier to distinguish normal from hard samples. In addition, our results also indicate that classifier accuracy significantly influences overall model performance, and substituting direct labels during testing can yield additional gains; see Table 7. While not the primary focus, these enhancements suggest valuable directions for future research, particularly in improving model learning and segmentation on complex medical image datasets." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.529, + 0.911, + 0.63 + ], + "angle": 0, + "content": "
DatasetAcc (%)Rec (%)DSC (%)DSC* (%)
COSMOS202292.782.384.1684.34
ISIC201883.372.690.9791.28
BUSI78.670.481.8290.04
Kvasir-SEG85.975.291.8792.10
ACDC87.571.791.3691.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.639, + 0.907, + 0.682 + ], + "angle": 0, + "content": "Table 7. The impact of the classifier to segmentation, where Acc and Rec denote the overall accuracy and recall of hard samples, respectively. \\( \\mathrm{DSC}^* \\) is obtained by classification using true labels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.722, + 0.712, + 0.74 + ], + "angle": 0, + "content": "5. Concluding remarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We propose a novel segmentation framework that can identify hard samples and enhance segmentation performance by employing stronger networks and more sophisticated optimization strategies on selected hard samples. The framework incorporates a diffusion process as an implicit regularization to handle hard samples, achieved by leveraging the inherent distribution bias of the dataset combined with random noise. Extensive experiments across various datasets demonstrated the superiority of our method over SOTA approaches, particularly for challenging samples." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15649" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.147 + ], + "angle": 0, + "content": "The work is supported by the National Natural Science Foundation of China (NSFC 12071345)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.175, + 0.188, + 0.19 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.244, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[2] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514-2525, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.343, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[4] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.526 + ], + "angle": 0, + "content": "[5] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.484, + 0.598 + ], + "angle": 0, + "content": "[6] Mengzhao Chen, Mingbao Lin, Ke Li, Yunhang Shen, Yongjian Wu, Fei Chao, and Rongrong Ji. Cf-vit: A general coarse-to-fine method for vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7042-7052, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[8] Jiequan Cui, Yuhui Yuan, Zhisheng Zhong, Zhuotao Tian, Han Hu, Stephen Lin, and Jiaya Jia. Region rebalance for long-tailed semantic segmentation. arXiv preprint arXiv:2204.01969, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[9] Yin Cui, Feng Zhou, Yuanqing Lin, and Serge Belongie. Fine-grained categorization and dataset bootstrapping using deep metric learning with humans in the loop. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1153-1162, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.2, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[12] Xi Fang and Pingkun Yan. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging, 39(11):3619–3629, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[13] Jun Fu, Jing Liu, Hajjie Tian, et al. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3146-3154, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[14] Ethan Harris, Antonia Marcu, Matthew Painter, Mahesan Niranjan, Adam Prügel-Bennett, and Jonathon Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7132-7141, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[17] Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to segment the tail. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14045-14054, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[18] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[19] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pál Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26, pages 451–462. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.906, + 0.646 + ], + "angle": 0, + "content": "[20] Longlong Jing, Yucheng Chen, and Yingli Tian. Coarse-to-fine semantic segmentation from image-level labels. IEEE transactions on image processing, 29:225-236, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[22] Aybora Koksal, Önder Tuzcuoglu, Kutalmış Gökalp İnce, Yoldaş Ataseven, and A Aydin Alatan. Improved hard example mining approach for single shot object detectors. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3536-3540. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[23] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3202, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[24] Xuhang Lian, Yanwei Pang, Jungong Han, and Jing Pan. Cascaded hierarchical atrous spatial pyramid pooling module for semantic segmentation. Pattern Recognition, 110: 107622, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15650" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Yucong Liu, Shixing Yu, and Tong Lin. Regularizing deep neural networks with stochastic estimators of hessian trace. arXiv preprint arXiv:2208.05924, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[27] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[28] Md Ashraful Alam Milton. Automated skin lesion classification using ensemble of deep neural networks in isic 2018: Skin lesion analysis towards melanoma detection challenge. arXiv preprint arXiv:1901.10802, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[29] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[31] Seong-Eun Ryu and Kyung-Yong Chung. Detection model of occluded object based on yolo using hard-example mining and augmentation policy optimization. Applied Sciences, 11 (15):7093, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[32] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[33] Abhinav Shrivastava, Rahul Sukthankar, and Mohammad Hejrati. Training region-based object detectors with online hard example mining. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 761-769, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[34] Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Hacihaliloglu, and Vishal M Patel. Medical transformer: Gated axial-attention for medical image segmentation. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, part I 24, pages 36-46. Springer, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[35] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In International conference on machine learning, pages 6438-6447. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9695-9704, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[37] Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model-agnostic segmentation refinement with discrete diffusion process. arXiv preprint arXiv:2312.12425, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.907, + 0.22 + ], + "angle": 0, + "content": "[38] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7303-7313, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[39] Xue Wang, Zhanshan Li, Yongping Huang, and Yingying Jiao. Multimodal medical image segmentation using multiscale context-aware network. Neurocomputing, 486:135-146, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.28, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[40] Colin Wei, Sham Kakade, and Tengyu Ma. The implicit and explicit regularization effects of dropout. In International conference on machine learning, pages 10181-10192. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.338, + 0.905, + 0.407 + ], + "angle": 0, + "content": "[41] Junde Wu, Wei Ji, Huazhu Fu, Min Xu, Yueming Jin, and Yanwu Xu. Medsegdiff-v2: Diffusion-based medical image segmentation with transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6030-6038, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.41, + 0.905, + 0.465 + ], + "angle": 0, + "content": "[42] Xin Xiao, Daiguo Zhou, Jiagao Hu, Yi Hu, and Yongchao Xu. Not all pixels are equal: Learning pixel hardness for semantic segmentation. arXiv preprint arXiv:2305.08462, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[43] Jiahao Xie, Sheng Zhang, Jianwei Lu, and Ye Luo. L-snet: From region localization to scale invariant medical image segmentation. In 2021 IEEE International Conference on Image Processing (ICIP), pages 41-45. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[44] Jin Yin, Pengfei Xia, and Jingsong He. Online hard region mining for semantic segmentation. Neural Processing Letters, 50:2665-2679, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.654 + ], + "angle": 0, + "content": "[45] Qihang Yu, Dong Yang, Holger Roth, Yutong Bai, Yixiao Zhang, Alan L Yuille, and Daguang Xu. C2fnas: Coarse-to-fine neural architecture search for 3d medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4126-4135, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[46] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[47] Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3457-3466, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[48] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15651" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[50] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, Part I 24, pages 14-24. Springer, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[51] Xiangyun Zhao, Raviteja Vemulapalli, Philip Andrew Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive learning for label efficient semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10623-10633, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[52] Xiaoqi Zhao, Hongpeng Jia, Youwei Pang, Long Lv, Feng Tian, Lihe Zhang, Weibing Sun, and Huchuan Lu. M\\(^{\\wedge}\\{2\\}\\) snet: Multi-scale in multi-scale subtraction network for medical image segmentation. arXiv preprint arXiv:2303.10894, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.332, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[53] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13001-13008, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[54] Tongxue Zhou, Su Ruan, and Stephane Canu. A review: Deep learning for medical image segmentation using multimodality fusion. *Array*, 3:100004, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.527 + ], + "angle": 0, + "content": "[55] Chenglu Zhu, Xiaoyan Wang, Shengyong Chen, Zhongzhao Teng, Cong Bai, Xiaojie Huang, Ming Xia, Zhanpeng Shao, Zheng Gu, and Peiliang Sun. Complex carotid artery segmentation in multi-contrast mr sequences by improved optimal surface graph cuts based on flow line learning. Medical & Biological Engineering & Computing, 60(9):2693-2706, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.53, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15652" + } + ] +] \ No newline at end of file diff --git a/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/c1e057f3-058c-42b0-8d35-8e21af099503_origin.pdf b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/c1e057f3-058c-42b0-8d35-8e21af099503_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..92c72a59f4446375d6eaf989dea5508c05239290 --- /dev/null +++ b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/c1e057f3-058c-42b0-8d35-8e21af099503_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04be9995cd5eff0eb7fec1ee5ceddd728322506f22d752a3827cb69b2dd1a7bb +size 1010898 diff --git a/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/full.md b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/full.md new file mode 100644 index 0000000000000000000000000000000000000000..32dd17dce16d8ddf1e0416dfa00b729c28562103 --- /dev/null +++ b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/full.md @@ -0,0 +1,371 @@ +# Take the Bull by the Horns: Learning to Segment Hard Samples + +Yuan Guo, Jingyu Kong, Yu Wang, Yuping Duan* +Center for Applied Mathematics, Tianjin University, Tianjin 300072, P.R.China +Key Laboratory of Digital Technology in Medical Diagnostics of Zhejiang Province, 310030, P.R.China +School of Mathematical Sciences, Beijing Normal University, Beijing, 100875, P.R.China + +guoyuan6060@163.com;kjyanhui@163.com;flimanadam@gmail.com;doveduan@gmail.com + +# Abstract + +Medical image segmentation is vital for clinical applications, with hard samples playing a key role in segmentation accuracy. We propose an effective image segmentation framework that includes mechanisms for identifying and segmenting hard samples. It derives a novel image segmentation paradigm: 1) Learning to identify hard samples: automatically selecting inherent hard samples from different datasets, and 2) Learning to segment hard samples: achieving the segmentation of hard samples through effective feature augmentation on dedicated networks. We name our method 'Learning to Segment hard samples' (L2S). The hard sample identification module comprises a backbone model and a classifier, which dynamically uncovers inherent dataset patterns. The hard sample segmentation module utilizes the diffusion process for feature augmentation and incorporates a more sophisticated segmentation network to achieve precise segmentation. We justify our motivation through solid theoretical analysis and extensive experiments. Evaluations across various modalities show that our L2S outperforms other SOTA methods, particularly by substantially improving the segmentation accuracy of hard samples. On ISIC dataset, our L2S improves the Dice score on hard samples and overall segmentation by $8.97\%$ and $1.01\%$ , respectively, compared to SOTA methods. The code is available at https://github.com/TqIYuanGie/L2S. + +# 1. Introduction + +Image segmentation is vital in medicine for identifying regions of interest to aid diagnosis and treatment planning. While traditional methods struggle with image complexity, deep learning excels in automatically extracting high-level features and capturing intricate patterns. + +Deep learning-based image segmentation methods can be broadly categorized into end-to-end and coarse-to-fine + +![](images/6673986b253dade01d0b8403716e23931b25430fd9df92f304a093e0a69fbad1.jpg) +Figure 1. Illustration of Hard samples. Several types of challenges that lead to hard samples are provided, which exhibit a long-tail distribution. + +approaches. End-to-end methods, such as U-Net [30], V-Net [27], and nnU-Net [18], map input images directly to segmentation outputs within a unified framework, often using CNNs as benchmarks. Recent transformer-based models, like TransUNet [5] and Swin-UNet [4], improve performance by capturing long-range dependencies. However, these models struggle with segmenting small anatomical structures or pathological regions where fine details and precise boundaries are crucial. + +Coarse-to-fine methods, on the other hand, begin with rough segmentation and progressively refine it, allowing the model to capture broader context before focusing on intricate details [6, 20, 43, 45]. This approach effectively addresses some limitations of end-to-end methods in segmenting small or complex structures. However, these models can still struggle with hard samples in medical image segmentation, as the quality of refinement depends on the coarse predictions, which may fail to capture subtle features essential for complex cases. + +Hard samples, such as heterogeneous tumor regions, vessel bifurcations, stenosis, imaging artifacts (e.g., MRI motion artifacts), and rare data in training sets, present significant challenges for accurate segmentation and often + +![](images/be9f2e141694943c93c0c5c3ca82975a3deee320a8dad3ec935914c58a56c786.jpg) +Figure 2. Compared with the previous two types of methods, our method is more specifically targeted at hard samples. + +hold critical clinical significance; see Figure 1. This paper proposes an effective segmentation framework to improve medical image segmentation, particularly for hard samples. Our model, L2S (Learning to Segment Hard Samples), consists of two components: identifying hard samples and segmenting hard samples. First, we introduce an adaptive method to distinguish normal and hard samples, using a dynamic threshold to identify hard samples for different datasets. Second, we develop a specialized hard sample segmentation network based on a multiscale Transformer network, i.e., Mask2Former [7], where effective feature enhancement and selection methods are introduced to precisely segment hard samples. Additionally, as shown in Figure 2, we introduce an implicit regularization term between hard sample generation and segmentation, where random noise in the diffusion process introduces uncertainty to prevent overfitting and improve model generalization. To sum up, our main contributions can be concluded as follows: + +- We design an effective segmentation framework that introduces a novel paradigm to enable the model to identify and segment hard samples inherent to the dataset. +- We propose to use a diffusion process as feature augmentation, which serves as an implicit regularization to enhance the model's performance and generalization ability in segmenting hard samples. +- Our theoretical proof establishes the equivalence between the diffusion process and second-order regularization, offering deeper insights into the superior performance of our method on hard samples. +- Our L2S framework demonstrates substantial performance gains across a range of medical segmentation datasets through leveraging dataset-specific characteristics to guide the model in segmenting hard samples. + +# 2. Related works + +Hard sample mining. Hard sample mining has been a central focus in various machine learning tasks and has played a significant role in applications such as classification and detection [9, 22, 31, 33]. It has also made significant ad + +vancements in the field of image segmentation, including layer cascade techniques for segmenting regions of different hardness [23], loss-based identification and retraining of hard regions [44], edge refinement from feature-level image rendering [21], and pixel hardness learning using global and historical loss values [42]. These methods have proven effective in improving model robustness by focusing learning on challenging examples. + +However, research on hard sample mining in the field of medical image segmentation remains insufficient. Due to the complexity and variability of medical images, which often involve rare pathologies, irregular boundaries, and limited annotated data, the above methods face significant limitations. In addition, Hard samples in medical images often have significant clinical implications. Effectively handling these challenging cases is crucial for clinical applications, and more data-efficient methods are required to learn from rare and precious samples. + +Data augmentation and model improvements. Data augmentation is an effective way to improve data utilization and model performance. Traditional methods like rotations, flips, and noise addition [32] are homogeneous, while heterogeneous strategies like random proportion mixing [14, 35, 48], random masking [10, 53] and image blending [3, 46] increase data diversity. However, designing proper augmentation for medical images requires considerable domain knowledge and is even impossible for some complicated features. To effectively extract delicate features, various sophisticated architectures are proposed for medical image segmentation. Multi-scale methods [24, 52, 54] capture broad and detailed information across spatial resolutions, improving adaptability to diverse samples. Multimodal designs [39, 50] integrate data from different imaging modalities, compensating for individual limitations. Attention mechanisms [13, 16, 29] focus on key regions, improving precision with hard samples. Uncertainty estimation identifies ambiguous regions, improving robustness. Although these methods effectively enhance overall segmentation performance, the improvements for hard samples are usually significantly less than for easy samples, which is undesirable in medical applications. More importantly, humans can adjust their learning and reasoning strategies when dealing with hard and easy samples, a characteristic not fully leveraged by current methods. + +Learning to learn. Learning to learn is a new paradigm focused on scarce and imbalanced data, including balanced replay [8, 17], dynamic gradient re-balancing [36], feature augmentation [37, 47], and contrastive learning [15, 38, 49, 51]. Among them, DINO [49] enhances hard object detection accuracy by comparing positive samples with challenging negative boxes. The idea of using large-scale negative sampling to improve performance is also a viable solution for hard sample segmentation. Motivated by + +![](images/03637708cea63fadbecf6aa38b007ca719202f4f00b325b151fb9a2387af9399.jpg) +(b) Learning to segment hard samples +(a) Learning to identify hard samples +Figure 3. Architecture of our L2S model. (a) Learning to identify hard samples, including a backbone model, a pre-trained classifier, and a post-former module. (b) Learning to segment hard samples, including a feature diffusion and hard sample segmentation module. + +these methods, we aim to design a segmentation architecture that automatically identifies hard samples and adapts its learning and inferencing frameworks accordingly. + +# 3. Our methodology + +As shown in Figure 3, the proposed L2S has two main components: identifying and segmenting hard samples. First, a backbone model and a pretrained classifier extracts features to identify hard samples. Hard samples are then enhanced through a diffusion process, followed by a specialized segmentation model. By using distinct models for normal and hard samples and focusing on hard sample segmentation, our method improves overall accuracy. + +# 3.1. Learning to identify hard samples + +Hard sample identification. To effectively learn from hard samples, the first step is to establish an approach that automatically assesses the hardness of each sample. In segmentation tasks, it is reasonable to hypothesize that images yielding unsatisfactory segmentation indicate 'hard samples', whereas high-quality segmentation belongs to the 'normal samples'. Accordingly, the most straightforward approach to measure the hardness is to evaluate the quality of segmentation. A simple yet efficient method is to train a neural network that categorizes segmentation outcomes as 'qualified' or 'unqualified'. To this purpose, we develop the network by integrating a segmentation backbone and a pre-trained classifier. During training, the seg + +mentation results are compared with the ground truth, and the pre-trained classifier is further trained using the obtained classification labels for distinction during testing. The segmentation backbone employs a simple UNet architecture, enhanced with random perturbations in the final decoder layer, allowing the model to simultaneously produce multiple segmentation outcomes for a single image. + +To train the network, we label the segmentation results as 'qualified' or 'unqualified' based on a Dice score threshold derived from the ground truth. Since this threshold reflects instantaneous hardness, which may change during training, we use Bayesian optimization for dynamic adjustment instead of a fixed threshold. We initialize the threshold to $\eta^0$ and use the model performance in the validation set as the objective function $h(x; \theta, \eta)$ . A Gaussian Process (GP) is used to approximate this objective function + +$$ +h (x; \theta , \eta) \sim \mathcal {G P} (m (\eta); k (\eta , \eta^ {\prime})), \tag {1} +$$ + +where $m(\eta)$ is the mean function, $k(\eta ,\eta^{\prime})$ is the covariance function. The value of $\eta$ is dynamically updated by maximizing the Expected Improvement (EI) function as follows + +$$ +\eta^ {k + 1} = \arg \max _ {\eta} \left(\mu (\eta) - h (\eta^ {k})\right) \Phi (Z) + \sigma (\eta) \phi (Z), +$$ + +where $\mu (\eta)$ and $\sigma (\eta)$ are the predicted mean and standard variation of the objective function, and $\Phi$ and $\phi$ are the cumulative distribution function and probability density function with $Z = \left(\mu (\eta) - h(\eta^{k})\right) / \sigma (\eta)$ . The gradient-descent optimization is used to solve the process. Then, we + +reclassify normal and hard samples using the updated $\eta^{k + 1}$ and repeat the above process in subsequent validation. + +Post-Former. We further refine the qualified segmentations associated with normal samples by a Post-Former module, which is a two-layer Transformer. Specifically, the segmentations provide $q$ to access $k$ and $v$ from the shallow image features of the encoder of the segmentation backbone, and the refinement is performed through a cross-attention layer and a self-attention layer. + +# 3.2. Learning to segment hard samples + +Feature augmentation via diffusion process. It is well known that feature augmentation can enhance the generalizability of image segmentation models and improve the performance of hard samples. Wei et al. [40] analyzed the generalization error bound for deep neural network (DNN) models, showing that it is influenced by the Jacobian matrix and the trace of the Hessian matrix of the loss function. Let $\bar{\ell} = \min \{B,\ell \}$ represent the truncation of the loss function to some fixed bound $B > 0$ . Given $P$ as the population data distribution and $P_{n}$ as the distribution over training samples, the following generalization bound holds: + +Theorem 1 (Theorem 4.1 in [40]). With probability $1 - \delta$ over the training examples, for all weight matrices $W$ satisfying the norm bound $\| W^T\|_{2,1} \leq A$ , the following bound holds + +$$ +\begin{array}{l} \mathbb {E} _ {P} \left[ \bar {\ell} \right] - 1. 0 1 \mathbb {E} _ {P _ {n}} \left[ \bar {\ell} \right] \leq \frac {\left(A \mu (W)\right) ^ {\frac {2}{3}} (\theta B) ^ {\frac {1}{3}}}{n ^ {\frac {1}{3}}} \\ + \frac {A \sqrt {B \nu (W)} \theta}{\sqrt {n}} + \frac {B A ^ {2} \theta}{n \left(\log^ {2} \left(\frac {B A ^ {2} \theta}{\nu (W) n}\right) + 1\right)} + \zeta , \tag {2} \\ \end{array} +$$ + +where $\mu (W)$ and $\nu (W)$ are the Jacobian and Hessian matrix over samples, respectively, $\theta = \ln^3 (nM)\max_i\| x_i\| _2^2$ and $\zeta$ is a low order term. + +Thus, we can control both the Jacobian norm and the trace of the Hessian matrix to enhance the generalization of DNN models. While gradient descent primarily focuses on first-order derivatives, it lacks an effective mechanism to constrain the trace of the Hessian. Liu et al. [25] addressed this gap by introducing the trace of the Hessian as an additional penalty term in the loss function + +$$ +\ell_ {a l l} (f (x), y) \approx \ell (f (x), y) + \lambda \cdot \operatorname {T r} \left(H _ {\ell , W}\right), \tag {3} +$$ + +where $\lambda$ controls the strength of the Hessian regularization. + +In this work, we implement feature augmentation using a diffusion model, which is equivalent to implicitly penalizing the trace of the Hessian (see Section 3.3 for details). Specifically, we use random noise to progressively corrupt clean labels $x_0$ into noisy labels $x_{t}$ as + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon +$$ + +where $\epsilon \sim \mathcal{N}(0,I)$ , $\bar{\alpha}_t = \prod_{i=1}^t \alpha_t$ , $\alpha_t = 1 - \beta_t$ and $\beta_t$ is used to parameterize the forward process for $t = 1, \dots, T$ . The forward process generates a series of noisy labels, enriching the feature space of hard samples. The reverse process iteratively removes noises to estimate clean labels. We use an encoder-decoder network as the denoiser, with $g_\theta(\cdot)$ encoding the noise variable and $d_\theta(\cdot)$ decoding to predict the $\epsilon$ . We then have + +$$ +x _ {t - 1} = d _ {\theta} \left(x _ {t}; m _ {\theta} \left(c _ {f}, g _ {\theta} (x _ {t})\right), t\right), +$$ + +where the conditional image embedding $c_{f}$ extracted from the backbone model is fused with the noise embedding $g_{\theta}(x_{t})$ in mix-former module $m_{\theta}(\cdot)$ . After training, the diffusion model's decoder is discarded, and its gradients guide the encoder to better represent noisy labels as augmented features for hard sample segmentation. Notably, applying a similar diffusion procedure to normal samples yields minimal improvements, suggesting that the optimization strategy for samples of varying hardness should differ. + +Mix-Former. For a segmentation task, the diffusion process must be conditional on the original image, which the proposed Mix-Former module implements. It integrates the noise embedding, hereafter referred to as $g(\epsilon)$ for clarity, and the conditional image embedding $c_{f}$ obtained by the segmentation backbone model. In particular, Mix-Former consists of $N$ (with $N = 3$ in our implementation) transformer blocks, each containing two cross-attention modules. We define the basic form of the attention mechanism with domain transformation as + +$$ +F _ {A t t n} (q, k) = F ^ {- 1} \left[ \operatorname {s o f t m a x} \left(\frac {\left(F (q) \mathcal {W} _ {q}\right) \cdot \left(F (k) \mathcal {W} _ {k}\right) ^ {T}}{\sqrt {d _ {k}}}\right) \right], +$$ + +where $\mathcal{W}$ is the projection matrix in Fourier space, $F$ and $F^{-1}$ are the Fast Fourier Transform (FFT) and inverse transform, respectively. Here, frequency domain learning can enhances the model's ability to capture complex dependencies between noise and semantic information, thereby improving performance. + +![](images/28000a14b79c4654dac1cc502672780191a05c5f70749cbc2422549c8f0c6f51.jpg) +Figure 4. The structure of Mix-Former, where $g(\epsilon)$ and $c_{f}$ are fused in the mix-former through $N$ layers of repeated attention. + +As shown in Figure 4, the first module integrates $g(\epsilon)$ into $c_{f}^{0}$ , while the second module encodes the embedding $c_{f}^{1}$ back into $g(\epsilon)$ . The bidirectional encoding enables the + +$g(\varepsilon)$ + +model to learn the interaction between noise and semantic features to obtain a more robust representation [41]. + +Hard sample segmentation. Hard samples exhibit low-quality segmentation that deviates significantly from their labels. Thus, we concatenate the unqualified segmentation with its original image and use an encoder and pixel decoder to extract multi-scale features. A learnable object query accesses these features via cross-attention in the Transformer decoder. Augmented features from the diffusion model further enhance feature extraction. The learnable query dynamically selects relevant information for precise segmentation of hard samples. The Transformer decoder contains four cross-attention layers; see Figure 5. The first three layers fuse the learnable query with features from three scales of the pixel decoder, while the fourth layer integrates augmented features from the diffusion process to enhance segmentation. The details are as follows + +$$ +z _ {i} = \left\{ \begin{array}{l l} \operatorname {s o f t m a x} \left(\frac {\left(W _ {q} L _ {q}\right) \cdot \left(W _ {k} M _ {f} ^ {i}\right) ^ {T}}{\sqrt {d _ {k}}}\right) \cdot W _ {v} M _ {f} ^ {i}, & i = 1, 2, 3, \\ F _ {A t t n} (z _ {3}, g (\epsilon)) \cdot W _ {v} g _ {\natural} (\epsilon), & i = 4, \end{array} \right. +$$ + +where $L_{q}$ represents the learnable query, $M_{f}^{i}$ represents the multi-scale features by scaling with 32, 16 and $\S$ respectively, and $z_{i}$ represents the attention-weighted feature. Such pattern is repeated in a round-robin fashion for all subsequent layers. + +![](images/e5254675850841c516dd3a5f6eed51bb9309efb7d5d718bd1fcb8fcaa0f0f201.jpg) +Figure 5. Hard samples segmentation decoder, where $L_{q}$ access both multi-scale features $M_{f}$ and augmented features $g(\epsilon)$ . + +# 3.3. Implicit regularization + +Subsequently, we formally demonstrate that the diffusion process in our model works as an implicit regularization to improve the model's generalization ability to hard samples. We aim to find the hard sample segmentation network $f(x;\theta)$ that minimizes the expected loss $\mathcal{L} = \mathbb{E}[\ell (f(x;\theta),y)]$ . Then we have the following result + +Proposition 1. For $\epsilon \sim \mathcal{N}(0,I)$ , let $g(\epsilon)$ denote the augmented features obtained by the diffusion probabilistic model. Then the expected loss becomes + +$$ +\mathbb {E} [ \ell (f (x + g (\epsilon)), y) ] \approx \ell (f (x), y) + \frac {1}{2} \operatorname {T r} (H \Sigma_ {g}). \tag {4} +$$ + +Proof. By incorporating the augmented features extracted from the noisy data, we have the expected loss function as + +$$ +\mathcal {L} = \mathbb {E} [ \ell (f (x + g (\epsilon)), y) ], +$$ + +which can be approximated by Taylor's expansion + +$$ +\mathcal {L} \approx \mathbb {E} \big [ \ell (f (x), y) + \nabla_ {x} \ell (f (x), y) \cdot g (\epsilon) + \frac {1}{2} g (\epsilon) ^ {T} H g (\epsilon) \big ], +$$ + +with $\nabla_{x}\ell (f(x),y)$ and $H$ to be the gradient and Hessian of the loss function w.r.t. the input $x$ . Since $\epsilon$ is of zero mean, by assuming that $g$ is center distributed, i.e., $\mathbb{E}[g(\epsilon)] = 0$ , we have + +$$ +\mathbb {E} [ \nabla_ {x} \ell (f (x), y) \cdot g (\epsilon) ] = \mathbb {E} [ g (\epsilon) ] \cdot \nabla_ {x} \ell (f (x), y) = 0. +$$ + +Based on the properties of matrix trace, there is + +$$ +\mathbb {E} [ g (\epsilon) ^ {T} H g (\epsilon) ] = \mathbb {E} [ \operatorname {T r} (H g (\epsilon) g (\epsilon) ^ {T}) ] = \operatorname {T r} (H \mathbb {E} [ g (\epsilon) g (\epsilon) ^ {T} ]), +$$ + +Assuming the covariance matrix of $g(\epsilon)$ is $\Sigma_g$ , we have $\Sigma_g = \mathbb{E}[g(\epsilon)g(\epsilon)^T]$ . It follows that + +$$ +\mathbb {E} \left[ g (\epsilon) ^ {T} H g (\epsilon) \right] = \operatorname {T r} \left(H \Sigma_ {g}\right). +$$ + +Then, the expected loss function becomes + +$$ +\mathbb {E} [ \ell (f (x + g (\epsilon)), y) ] \approx \ell (f (x), y) + \frac {1}{2} \operatorname {T r} (H \Sigma_ {g}), +$$ + +which completes the proof. + +![](images/51e90a82c225beb8a6662c886bcfddaa33d1d3a5e1b94861c003c3f343be5a7e.jpg) + +By penalizing the trace of the Hessian matrix, the model is driven towards flatter minima, where the loss function is relatively stable, thereby improving generalization for hard sample segmentation [25]. + +# 3.4. Loss function + +Diffusion loss. The Diffusion Loss measures the difference between the ground truth noise added to the label and the noise predicted by the diffusion model + +$$ +\mathcal {L} _ {\text {d i f f}} = \left\| \epsilon - d _ {\theta} \left(t, x _ {t}, c _ {f}\right) \right\| _ {2} ^ {2}, \tag {5} +$$ + +where $c_{f}$ is the conditional image embedding, $\epsilon$ is the noise, and $d_{\theta}(t,x_t,c_f)$ is the noise predicted at time step $t$ . + +Segmentation loss. For hard samples, the segmentation loss $\mathcal{L}_{hard} = \mathcal{L}_{mask} + \lambda_{cls}\mathcal{L}_{cls}$ combines mask loss and classification loss. The mask loss includes binary cross-entropy and dice loss: $\mathcal{L}_{mask} = \lambda_{bce}\mathcal{L}_{bce} + \lambda_{dice}\mathcal{L}_{dice}$ , while the classification loss is cross-entropy. For normal samples, the segmentation loss $\mathcal{L}_{normal}$ uses only the mask loss. The total segmentation loss is: + +$$ +\mathcal {L} _ {\text {s e g}} = \lambda_ {\text {h a r d}} \mathcal {L} _ {\text {h a r d}} + \lambda_ {\text {n o r m a l}} \mathcal {L} _ {\text {n o r m a l}}, \tag {6} +$$ + +with hyperparameters set as $\lambda_{bce} = \lambda_{dice} = \lambda_{cls} = 5$ , $\lambda_{hard} = 5$ , and $\lambda_{normal} = 2$ . The dice loss for the backbone model is weighted at 5, while the binary cross-entropy + +
MethodCOSMOS 2022ISIC 2018Kvasir-SEGBUSIAverage
DSCIoUDSC†DSCIoUDSC†DSCIoUDSC†DSCIoUDSC†DSCIoU
UNet[30]82.4472.7641.7687.1479.7342.2985.6378.6748.3576.4167.6141.0882.9074.69
PolypPVT[11]82.3372.3434.4389.8482.6152.4691.4486.3960.2681.0572.5146.6286.1778.46
nnUNet[18]84.1275.5855.7988.9181.6546.3190.5585.1853.4780.9472.4342.8786.1378.71
MedT[34]82.8173.1539.2688.8481.4944.0589.6884.1251.2380.4471.2740.1985.4477.51
TransUNet[5]82.9273.1843.6889.4482.3751.7691.0485.7556.7380.3271.3845.3185.9378.17
TransFuses[50]83.1773.8945.3589.9683.2254.9391.2786.3158.6481.5073.2548.2586.4879.17
SwinUNet[4]83.6374.7748.1789.5682.4049.5790.2285.0951.1879.7670.1939.7385.7978.11
MedSegDiffV2[41]82.8972.9339.5489.7882.6551.1291.0686.2152.3080.5371.4442.7486.0778.31
Mask2former[7]82.7672.8241.3289.5982.4649.5190.8685.2853.7780.1970.6343.3685.8577.80
L2S84.1675.6566.6790.9784.3463.9091.8787.0264.3481.8273.2753.2787.2180.07
+ +Table 1. Comparison of L2S with SOTA segmentation methods over different image modalities, where the results of SOTA methods are reproduced using their publicly available implementation with five-fold cross validation. $\mathrm{DSC}^{\dagger}$ is computed on the hard samples with dice score below 0.7. We add the standard deviation and statistical significance in suppl. materials. + +loss for the classifier is weighted at 10. + +Consistency loss. To ensure consistency between normal and hard samples, which share the same label, we introduce the following consistency loss: + +$$ +\mathcal {L} _ {\text {c o n s}} = \left\| p \left(\mathcal {D} _ {n}\right) - f \left(\mathcal {D} _ {h}\right) \right\| _ {2} ^ {2}, \tag {7} +$$ + +where the post-former $p(\cdot)$ fine-tunes the normal sample set, and $f(\cdot)$ represents the segmentation network for the hard samples. To sum up, the total loss for our model is the weighted sum of all the aforementioned terms as + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s e g}} + \lambda_ {\text {d i f f}} \mathcal {L} _ {\text {d i f f}} + \lambda_ {\text {c o n s}} \mathcal {L} _ {\text {c o n s}}, \tag {8} +$$ + +where the weights are fixed as $\lambda_{diff} = 5$ and $\lambda_{cons} = 5$ . + +# 4. Experimental results + +# 4.1. Datasets and implementation details + +We conducted experiments on six public medical segmentation datasets to evaluate the performance of our model across various imaging modalities. These include the COSMOS2022 dataset [55], ISIC2018 dataset [28], Kvasir-SEG dataset [19], and the BUSI dataset [1] for carotid artery, skin lesion, polyp and breast ultrasound segmentation, respectively. The BTCV dataset [12], containing 12 anatomical structures, was used for abdominal multi-organ segmentation, while the ACDC dataset [2] was utilized for segmenting cardiac structures in MRI images. + +For the hard sample segmentation network, we used a pre-trained Swin Transformer [26] as the encoder and MSDeformAttn [56] as the pixel decoder, similar to Mask2Former [7]. The Transformer decoder was configured with $L = 3$ (12 layers in total) and 100 queries by default. We trained our model using the AdamW optimizer with a learning rate of 0.01 and a weight decay of $1e^{-4}$ . + +Training images were resized to $224 \times 224$ for the COSMOS2022, ACDC, and BTCV datasets. For other datasets, the images were resized to $448 \times 448$ with color jittering and random horizontal flipping. The training schedule included 12,000 iterations for COSMOS2022, 30,000 for ACDC and BTCV, and 20,000 for other datasets. All experiments were conducted using PyTorch 1.10.0 on 4 NVIDIA RTX 3090 GPUs. The Dice score with an initial threshold of 0.7 was used to distinguish between normal and hard samples. During training, normal and hard samples generated by the backbone model were stored separately. For each input image, a pair of normal and hard samples was randomly selected from their respective sets for further segmentation. Each image's two sets contain at least one element to ensure the balance. + +We compared our L2S with SOTA segmentation methods, including CNN-based methods like UNet [30], nnUNet [18], and transformer-based methods such as Polyp-PVT [11], MedT [34], TransUNet [5], TransFuse [50], SwinUNet [4], and the diffusion-based method MedSegDiffV2 [41]. We also compared L2S with the baseline model Mask2Former [7] effectiveness of the proposed framework. + +# 4.2. Results evaluation + +Comparisons on binary segmentation tasks. Our L2S model is compared with SOTA segmentation methods across four tasks; see Table 1, where the best results are shown in bold, and the same is applied in the following tables. The results demonstrate that our L2S model achieves the highest average Dice of $87.21\%$ and IoU of $80.07\%$ . Improving hard sample segmentation yields significant gains: L2S outperforms the SOTA methods by 10. $88\%$ in COSMOS2022, $8.97\%$ in ISIC2018, and $4.08\%$ in the polyp-specific PolypPVT model in Kvasir-SEG. These results highlight that the existing methods perform poorly + +![](images/58e3803bfcbfef9e5f0711f6533a9d382ba5dd9b6d26a6c45f6bf00bfb11e5e6.jpg) +Figure 6. Qualitative segmentation results for hard samples, including carotid artery, breast ultrasound, polyp, and skin lesions. + +
MethodSpleenR.KidL.KidGall.Eso.LiverStom.AortaIVCVeinsPanc.AGAvg.
UNet[30]0.9330.8790.9130.7120.7080.9390.8200.8570.7850.7070.6810.6370.797
nnUNet[18]0.9420.8940.9100.7040.7230.9480.8240.8770.7820.7200.6800.6160.802
TransUNet[5]0.9520.9270.9290.6620.7570.9690.8890.9200.8330.7910.7750.6370.837
TransFuses[50]0.9690.9240.9310.7720.7880.9740.9030.9140.8360.8020.7850.7410.861
SwinUNet[4]0.9390.9160.9080.7270.7340.9410.8520.9030.7980.7550.7330.6460.821
MedSegDiffV2[41]0.9480.9250.9160.7940.8220.9530.8910.9180.8390.7980.7910.7270.860
Mask2former[7]0.9560.9130.9210.7340.7550.9510.8920.9110.8350.7950.7530.6860.842
L2S0.9730.9340.9450.8260.8050.9720.9180.9220.8460.8240.8130.8020.882
+ +on hard samples. Our L2S achieves over a $10\%$ improvement in accuracy compared to Mask2Former. A detailed analysis of the selected hard samples is provided in Supplementary Material B.4 to validate our assumption that hard samples correspond to poor segmentation. Figure 6 presents the segmentation results of representative hard samples. Our L2S model excels at segmenting complex structures. Qualitative results for carotid arteries, breast ultrasound, polyps, and skin lesions show that the predictions of our L2S better match the ground truth. For carotid artery segmentation, our L2S model precisely identifies stenosis, offering important clinical value. Overall, our designs significantly improve robustness and generalization. The visualized results are available in the supplementary materials. + +Comparisons on multi-organ segmentation. Table 2 shows that our L2S outperforms the SOTA methods on BTCV multi-organ dataset, achieving the highest average Dice score of $88.2\%$ . Table 3 lists the Dice scores of L2S and the comparison methods on ACDC dataset for cardiac organ segmentation. Our L2S gains an average Dice score of $91.36\%$ , outperforming nnUNet by $0.33\%$ and improving right ventricle segmentation by $0.48\%$ . The comparison + +Table 2. Comparison of L2S with SOTA segmentation methods over multi-organ segmentation BTCV dataset in term of Dice Score. + +
MethodsRVMyoLVAvg.
UNet86.7980.4593.6386.96
nnUNet89.5988.3195.2091.03
TransUNet87.0587.1794.5289.58
TransFuses87.9487.0395.0990.02
SwinUNet86.5785.3394.1088.67
Mask2former86.7486.4993.9289.05
L2S(UNet)90.0788.6595.3691.36
L2S(Mask2former)90.4289.0295.6191.68
+ +Table 3. Results of cardiac organ segmentation on ACDC dataset, where dice scores are reported for individual organs. + +experiments on different backbone models show that more powerful Mask2Former can further enhance segmentation accuracy. Thus, the L2S framework is highly compatible and can be integrated with different segmentation models. + +# 4.3. Ablation studies + +We conducted a series of ablation studies on the ISIC2018 dataset unless otherwise noted. Table 4 shows an ablation + +study on FFT, consistency loss, and adaptive threshold selection. As shown, FFT significantly impacts the Dice score with an improvement of $0.31\%$ , highlighting the importance of feature alignment. + +
MethodDSC (%)IoU (%)
L2S90.9784.34
- FFT90.66 (-0.3)83.95 (-0.4)
- consistency loss90.78 (-0.2)84.07 (-0.3)
- adaptive threshold90.86 (-0.1)84.24 (-0.1)
- all 3 components90.54 (-0.4)83.72 (-0.6)
+ +Hard sample identification and augmentation. Table 5 presents an ablation study evaluating the impact of incorporating normal and hard sample identification mechanisms and feature augmentation modules on model performance. The results show that removing the identification mechanism (i.e., no longer distinguishing the segmentation results of the backbone model, treating all samples in the dataset as hard samples for subsequent processing, and therefore using no classifier during testing) causes a decrease in the Dice score from $90.87\%$ to $89.59\%$ , which indicates that complex processing may adversely affect the original normal samples in the dataset. In addition, using diffusion models for feature augmentation significantly improves the hard sample segmentation, increasing the Dice score by $4.52\%$ . Incorporating the Post-former module slightly enhances segmentation by refining normal samples. + +Table 4. Three optimization improvements in network operation, where we remove one component at a time. + +
CLFDDPMP-FormerDSC (%)IoU (%)DSC† (%)
89.5282.35-
✓90.2883.7458.96
✓89.5982.48-
✓✓90.8784.11-
✓✓90.3683.7959.38
✓✓✓90.9784.3463.90
+ +Hard sample segmentation. Table 6 shows the impact of various design choices on the performance of the hard sample segmentation network, measured by the Dice score. Using query vectors alongside hard sample features boost the Dice score from $55.83\%$ to $59.88\%$ . Besides, multiscale features and query learnability can further enhance hard sample segmentation accuracy, by increasing Dice of more than $1.5\%$ and $3\%$ , respectively. We also compare the + +Table 5. Ablation study on our L2S method, where 'CLF', 'DDPM' and 'P-Former' denote the pretrained classifier, feature augmentation, and Post-Former, respectively. + +
MethodDSC† (%)FLOPs
single scale (w/o query)55.8315.2G
single scale62.2432.7G
- learnable query features59.8832.7G
L2S63.9024.3G
- learnable query features60.8224.3G
+ +FLOPs of different designs for hard sample segmentation on the ISIC dataset. The multiscale approach achieved better performance with fewer FLOPs. + +# 4.4. Discussion + +We introduce a specialized framework for hard sample segmentation in medical images, advancing current methods. Our approach includes a dedicated sample generator and a classifier to distinguish normal from hard samples. In addition, our results also indicate that classifier accuracy significantly influences overall model performance, and substituting direct labels during testing can yield additional gains; see Table 7. While not the primary focus, these enhancements suggest valuable directions for future research, particularly in improving model learning and segmentation on complex medical image datasets. + +Table 6. Ablation study on Learnable query and Mutil-scale features, where 'single-scale' is the single resolution of $1/8$ . + +
DatasetAcc (%)Rec (%)DSC (%)DSC* (%)
COSMOS202292.782.384.1684.34
ISIC201883.372.690.9791.28
BUSI78.670.481.8290.04
Kvasir-SEG85.975.291.8792.10
ACDC87.571.791.3691.45
+ +Table 7. The impact of the classifier to segmentation, where Acc and Rec denote the overall accuracy and recall of hard samples, respectively. $\mathrm{DSC}^*$ is obtained by classification using true labels. + +# 5. Concluding remarks + +We propose a novel segmentation framework that can identify hard samples and enhance segmentation performance by employing stronger networks and more sophisticated optimization strategies on selected hard samples. The framework incorporates a diffusion process as an implicit regularization to handle hard samples, achieved by leveraging the inherent distribution bias of the dataset combined with random noise. Extensive experiments across various datasets demonstrated the superiority of our method over SOTA approaches, particularly for challenging samples. + +# 6. Acknowledgment + +# References + +[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6 +[2] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514-2525, 2018. 6 +[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 2 +[4] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 1, 6, 7 +[5] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 6, 7 +[6] Mengzhao Chen, Mingbao Lin, Ke Li, Yunhang Shen, Yongjian Wu, Fei Chao, and Rongrong Ji. Cf-vit: A general coarse-to-fine method for vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7042-7052, 2023. 1 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 6, 7 +[8] Jiequan Cui, Yuhui Yuan, Zhisheng Zhong, Zhuotao Tian, Han Hu, Stephen Lin, and Jiaya Jia. Region rebalance for long-tailed semantic segmentation. arXiv preprint arXiv:2204.01969, 2022. 2 +[9] Yin Cui, Feng Zhou, Yuanqing Lin, and Serge Belongie. Fine-grained categorization and dataset bootstrapping using deep metric learning with humans in the loop. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1153-1162, 2016. 2 +[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 2 +[11] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 6 + +The work is supported by the National Natural Science Foundation of China (NSFC 12071345). +[12] Xi Fang and Pingkun Yan. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging, 39(11):3619–3629, 2020. 6 +[13] Jun Fu, Jing Liu, Hajjie Tian, et al. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3146-3154, 2019. 2 +[14] Ethan Harris, Antonia Marcu, Matthew Painter, Mahesan Niranjan, Adam Prügel-Bennett, and Jonathon Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020. 2 +[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7132-7141, 2018. 2 +[17] Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to segment the tail. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14045-14054, 2020. 2 +[18] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 1, 6, 7 +[19] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pál Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26, pages 451–462. Springer, 2020. 6 +[20] Longlong Jing, Yucheng Chen, and Yingli Tian. Coarse-to-fine semantic segmentation from image-level labels. IEEE transactions on image processing, 29:225-236, 2019. 1 +[21] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 2 +[22] Aybora Koksal, Önder Tuzcuoglu, Kutalmış Gökalp İnce, Yoldaş Ataseven, and A Aydin Alatan. Improved hard example mining approach for single shot object detectors. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3536-3540. IEEE, 2022. 2 +[23] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3202, 2017. 2 +[24] Xuhang Lian, Yanwei Pang, Jungong Han, and Jing Pan. Cascaded hierarchical atrous spatial pyramid pooling module for semantic segmentation. Pattern Recognition, 110: 107622, 2021. 2 + +[25] Yucong Liu, Shixing Yu, and Tong Lin. Regularizing deep neural networks with stochastic estimators of hessian trace. arXiv preprint arXiv:2208.05924, 2022. 4, 5 +[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6 +[27] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 1 +[28] Md Ashraful Alam Milton. Automated skin lesion classification using ensemble of deep neural networks in isic 2018: Skin lesion analysis towards melanoma detection challenge. arXiv preprint arXiv:1901.10802, 2019. 6 +[29] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 2 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 6, 7 +[31] Seong-Eun Ryu and Kyung-Yong Chung. Detection model of occluded object based on yolo using hard-example mining and augmentation policy optimization. Applied Sciences, 11 (15):7093, 2021. 2 +[32] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 2 +[33] Abhinav Shrivastava, Rahul Sukthankar, and Mohammad Hejrati. Training region-based object detectors with online hard example mining. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 761-769, 2016. 2 +[34] Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Hacihaliloglu, and Vishal M Patel. Medical transformer: Gated axial-attention for medical image segmentation. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, part I 24, pages 36-46. Springer, 2021. 6 +[35] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In International conference on machine learning, pages 6438-6447. PMLR, 2019. 2 +[36] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9695-9704, 2021. 2 + +[37] Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model-agnostic segmentation refinement with discrete diffusion process. arXiv preprint arXiv:2312.12425, 2023. 2 +[38] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7303-7313, 2021. 2 +[39] Xue Wang, Zhanshan Li, Yongping Huang, and Yingying Jiao. Multimodal medical image segmentation using multiscale context-aware network. Neurocomputing, 486:135-146, 2022. 2 +[40] Colin Wei, Sham Kakade, and Tengyu Ma. The implicit and explicit regularization effects of dropout. In International conference on machine learning, pages 10181-10192. PMLR, 2020. 4 +[41] Junde Wu, Wei Ji, Huazhu Fu, Min Xu, Yueming Jin, and Yanwu Xu. Medsegdiff-v2: Diffusion-based medical image segmentation with transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6030-6038, 2024. 5, 6, 7 +[42] Xin Xiao, Daiguo Zhou, Jiagao Hu, Yi Hu, and Yongchao Xu. Not all pixels are equal: Learning pixel hardness for semantic segmentation. arXiv preprint arXiv:2305.08462, 2023. 2 +[43] Jiahao Xie, Sheng Zhang, Jianwei Lu, and Ye Luo. L-snet: From region localization to scale invariant medical image segmentation. In 2021 IEEE International Conference on Image Processing (ICIP), pages 41-45. IEEE, 2021. 1 +[44] Jin Yin, Pengfei Xia, and Jingsong He. Online hard region mining for semantic segmentation. Neural Processing Letters, 50:2665-2679, 2019. 2 +[45] Qihang Yu, Dong Yang, Holger Roth, Yutong Bai, Yixiao Zhang, Alan L Yuille, and Daguang Xu. C2fnas: Coarse-to-fine neural architecture search for 3d medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4126-4135, 2020. 1 +[46] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2 +[47] Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3457-3466, 2021. 2 +[48] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2 +[49] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2 + +[50] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, Part I 24, pages 14-24. Springer, 2021. 2, 6, 7 +[51] Xiangyun Zhao, Raviteja Vemulapalli, Philip Andrew Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive learning for label efficient semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10623-10633, 2021. 2 +[52] Xiaoqi Zhao, Hongpeng Jia, Youwei Pang, Long Lv, Feng Tian, Lihe Zhang, Weibing Sun, and Huchuan Lu. M $^{\wedge}\{2\}$ snet: Multi-scale in multi-scale subtraction network for medical image segmentation. arXiv preprint arXiv:2303.10894, 2023. 2 +[53] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13001-13008, 2020. 2 +[54] Tongxue Zhou, Su Ruan, and Stephane Canu. A review: Deep learning for medical image segmentation using multimodality fusion. *Array*, 3:100004, 2019. 2 +[55] Chenglu Zhu, Xiaoyan Wang, Shengyong Chen, Zhongzhao Teng, Cong Bai, Xiaojie Huang, Ming Xia, Zhanpeng Shao, Zheng Gu, and Peiliang Sun. Complex carotid artery segmentation in multi-contrast mr sequences by improved optimal surface graph cuts based on flow line learning. Medical & Biological Engineering & Computing, 60(9):2693-2706, 2022. 6 +[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 6 \ No newline at end of file diff --git a/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/images.zip b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9b1ac28a7e55c5537fd571e548332acaa0416d77 --- /dev/null +++ b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d038efaeee9be171111f0f0773541ba23dda6192c991a9f0435f2179ffd9eea1 +size 740577 diff --git a/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/layout.json b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..372701866fa2e8c7ef32d0a8a3bc8cfa8bcfe90e --- /dev/null +++ b/2025/Take the Bull by the Horns_ Learning to Segment Hard Samples/layout.json @@ -0,0 +1,9668 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 503, + 121 + ], + "type": "text", + "content": "Take the Bull by the Horns: Learning to Segment Hard Samples" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 143, + 563, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 143, + 563, + 199 + ], + "spans": [ + { + "bbox": [ + 62, + 143, + 563, + 199 + ], + "type": "text", + "content": "Yuan Guo, Jingyu Kong, Yu Wang, Yuping Duan* \nCenter for Applied Mathematics, Tianjin University, Tianjin 300072, P.R.China \nKey Laboratory of Digital Technology in Medical Diagnostics of Zhejiang Province, 310030, P.R.China \nSchool of Mathematical Sciences, Beijing Normal University, Beijing, 100875, P.R.China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 201, + 514, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 201, + 514, + 213 + ], + "spans": [ + { + "bbox": [ + 107, + 201, + 514, + 213 + ], + "type": "text", + "content": "guoyuan6060@163.com;kjyanhui@163.com;flimanadam@gmail.com;doveduan@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "type": "text", + "content": "Medical image segmentation is vital for clinical applications, with hard samples playing a key role in segmentation accuracy. We propose an effective image segmentation framework that includes mechanisms for identifying and segmenting hard samples. It derives a novel image segmentation paradigm: 1) Learning to identify hard samples: automatically selecting inherent hard samples from different datasets, and 2) Learning to segment hard samples: achieving the segmentation of hard samples through effective feature augmentation on dedicated networks. We name our method 'Learning to Segment hard samples' (L2S). The hard sample identification module comprises a backbone model and a classifier, which dynamically uncovers inherent dataset patterns. The hard sample segmentation module utilizes the diffusion process for feature augmentation and incorporates a more sophisticated segmentation network to achieve precise segmentation. We justify our motivation through solid theoretical analysis and extensive experiments. Evaluations across various modalities show that our L2S outperforms other SOTA methods, particularly by substantially improving the segmentation accuracy of hard samples. On ISIC dataset, our L2S improves the Dice score on hard samples and overall segmentation by " + }, + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "type": "inline_equation", + "content": "8.97\\%" + }, + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "type": "inline_equation", + "content": "1.01\\%" + }, + { + "bbox": [ + 55, + 266, + 296, + 566 + ], + "type": "text", + "content": ", respectively, compared to SOTA methods. The code is available at https://github.com/TqIYuanGie/L2S." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 135, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 670 + ], + "type": "text", + "content": "Image segmentation is vital in medicine for identifying regions of interest to aid diagnosis and treatment planning. While traditional methods struggle with image complexity, deep learning excels in automatically extracting high-level features and capturing intricate patterns." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "type": "text", + "content": "Deep learning-based image segmentation methods can be broadly categorized into end-to-end and coarse-to-fine" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 239, + 555, + 370 + ], + "blocks": [ + { + "bbox": [ + 315, + 239, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 315, + 239, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 555, + 370 + ], + "type": "image", + "image_path": "6673986b253dade01d0b8403716e23931b25430fd9df92f304a093e0a69fbad1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 377, + 555, + 410 + ], + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 410 + ], + "type": "text", + "content": "Figure 1. Illustration of Hard samples. Several types of challenges that lead to hard samples are provided, which exhibit a long-tail distribution." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 435, + 555, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 543 + ], + "type": "text", + "content": "approaches. End-to-end methods, such as U-Net [30], V-Net [27], and nnU-Net [18], map input images directly to segmentation outputs within a unified framework, often using CNNs as benchmarks. Recent transformer-based models, like TransUNet [5] and Swin-UNet [4], improve performance by capturing long-range dependencies. However, these models struggle with segmenting small anatomical structures or pathological regions where fine details and precise boundaries are crucial." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 663 + ], + "type": "text", + "content": "Coarse-to-fine methods, on the other hand, begin with rough segmentation and progressively refine it, allowing the model to capture broader context before focusing on intricate details [6, 20, 43, 45]. This approach effectively addresses some limitations of end-to-end methods in segmenting small or complex structures. However, these models can still struggle with hard samples in medical image segmentation, as the quality of refinement depends on the coarse predictions, which may fail to capture subtle features essential for complex cases." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": "Hard samples, such as heterogeneous tumor regions, vessel bifurcations, stenosis, imaging artifacts (e.g., MRI motion artifacts), and rare data in training sets, present significant challenges for accurate segmentation and often" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15642" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 290, + 182 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 290, + 182 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 290, + 182 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 290, + 182 + ], + "type": "image", + "image_path": "be9f2e141694943c93c0c5c3ca82975a3deee320a8dad3ec935914c58a56c786.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "type": "text", + "content": "Figure 2. Compared with the previous two types of methods, our method is more specifically targeted at hard samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 234, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 296, + 450 + ], + "type": "text", + "content": "hold critical clinical significance; see Figure 1. This paper proposes an effective segmentation framework to improve medical image segmentation, particularly for hard samples. Our model, L2S (Learning to Segment Hard Samples), consists of two components: identifying hard samples and segmenting hard samples. First, we introduce an adaptive method to distinguish normal and hard samples, using a dynamic threshold to identify hard samples for different datasets. Second, we develop a specialized hard sample segmentation network based on a multiscale Transformer network, i.e., Mask2Former [7], where effective feature enhancement and selection methods are introduced to precisely segment hard samples. Additionally, as shown in Figure 2, we introduce an implicit regularization term between hard sample generation and segmentation, where random noise in the diffusion process introduces uncertainty to prevent overfitting and improve model generalization. To sum up, our main contributions can be concluded as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 295, + 633 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 453, + 295, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 488 + ], + "type": "text", + "content": "- We design an effective segmentation framework that introduces a novel paradigm to enable the model to identify and segment hard samples inherent to the dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 489, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 536 + ], + "type": "text", + "content": "- We propose to use a diffusion process as feature augmentation, which serves as an implicit regularization to enhance the model's performance and generalization ability in segmenting hard samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 537, + 295, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 584 + ], + "type": "text", + "content": "- Our theoretical proof establishes the equivalence between the diffusion process and second-order regularization, offering deeper insights into the superior performance of our method on hard samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 585, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 585, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 585, + 295, + 633 + ], + "type": "text", + "content": "- Our L2S framework demonstrates substantial performance gains across a range of medical segmentation datasets through leveraging dataset-specific characteristics to guide the model in segmenting hard samples." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 645, + 144, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 144, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 144, + 657 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Hard sample mining. Hard sample mining has been a central focus in various machine learning tasks and has played a significant role in applications such as classification and detection [9, 22, 31, 33]. It has also made significant ad" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "vancements in the field of image segmentation, including layer cascade techniques for segmenting regions of different hardness [23], loss-based identification and retraining of hard regions [44], edge refinement from feature-level image rendering [21], and pixel hardness learning using global and historical loss values [42]. These methods have proven effective in improving model robustness by focusing learning on challenging examples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 170, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 290 + ], + "type": "text", + "content": "However, research on hard sample mining in the field of medical image segmentation remains insufficient. Due to the complexity and variability of medical images, which often involve rare pathologies, irregular boundaries, and limited annotated data, the above methods face significant limitations. In addition, Hard samples in medical images often have significant clinical implications. Effectively handling these challenging cases is crucial for clinical applications, and more data-efficient methods are required to learn from rare and precious samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "type": "text", + "content": "Data augmentation and model improvements. Data augmentation is an effective way to improve data utilization and model performance. Traditional methods like rotations, flips, and noise addition [32] are homogeneous, while heterogeneous strategies like random proportion mixing [14, 35, 48], random masking [10, 53] and image blending [3, 46] increase data diversity. However, designing proper augmentation for medical images requires considerable domain knowledge and is even impossible for some complicated features. To effectively extract delicate features, various sophisticated architectures are proposed for medical image segmentation. Multi-scale methods [24, 52, 54] capture broad and detailed information across spatial resolutions, improving adaptability to diverse samples. Multimodal designs [39, 50] integrate data from different imaging modalities, compensating for individual limitations. Attention mechanisms [13, 16, 29] focus on key regions, improving precision with hard samples. Uncertainty estimation identifies ambiguous regions, improving robustness. Although these methods effectively enhance overall segmentation performance, the improvements for hard samples are usually significantly less than for easy samples, which is undesirable in medical applications. More importantly, humans can adjust their learning and reasoning strategies when dealing with hard and easy samples, a characteristic not fully leveraged by current methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "Learning to learn. Learning to learn is a new paradigm focused on scarce and imbalanced data, including balanced replay [8, 17], dynamic gradient re-balancing [36], feature augmentation [37, 47], and contrastive learning [15, 38, 49, 51]. Among them, DINO [49] enhances hard object detection accuracy by comparing positive samples with challenging negative boxes. The idea of using large-scale negative sampling to improve performance is also a viable solution for hard sample segmentation. Motivated by" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15643" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 86, + 539, + 302 + ], + "blocks": [ + { + "bbox": [ + 71, + 72, + 241, + 86 + ], + "lines": [ + { + "bbox": [ + 71, + 72, + 241, + 86 + ], + "spans": [ + { + "bbox": [ + 71, + 72, + 241, + 86 + ], + "type": "text", + "content": "(b) Learning to segment hard samples" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 86, + 539, + 302 + ], + "lines": [ + { + "bbox": [ + 71, + 86, + 539, + 302 + ], + "spans": [ + { + "bbox": [ + 71, + 86, + 539, + 302 + ], + "type": "image", + "image_path": "03637708cea63fadbecf6aa38b007ca719202f4f00b325b151fb9a2387af9399.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 303, + 237, + 316 + ], + "lines": [ + { + "bbox": [ + 72, + 303, + 237, + 316 + ], + "spans": [ + { + "bbox": [ + 72, + 303, + 237, + 316 + ], + "type": "text", + "content": "(a) Learning to identify hard samples" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 326, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 326, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 555, + 350 + ], + "type": "text", + "content": "Figure 3. Architecture of our L2S model. (a) Learning to identify hard samples, including a backbone model, a pre-trained classifier, and a post-former module. (b) Learning to segment hard samples, including a feature diffusion and hard sample segmentation module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 369, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 295, + 407 + ], + "type": "text", + "content": "these methods, we aim to design a segmentation architecture that automatically identifies hard samples and adapts its learning and inferencing frameworks accordingly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 417, + 162, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 162, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 162, + 430 + ], + "type": "text", + "content": "3. Our methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 437, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 534 + ], + "type": "text", + "content": "As shown in Figure 3, the proposed L2S has two main components: identifying and segmenting hard samples. First, a backbone model and a pretrained classifier extracts features to identify hard samples. Hard samples are then enhanced through a diffusion process, followed by a specialized segmentation model. By using distinct models for normal and hard samples and focusing on hard sample segmentation, our method improves overall accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 237, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 237, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 237, + 553 + ], + "type": "text", + "content": "3.1. Learning to identify hard samples" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": "Hard sample identification. To effectively learn from hard samples, the first step is to establish an approach that automatically assesses the hardness of each sample. In segmentation tasks, it is reasonable to hypothesize that images yielding unsatisfactory segmentation indicate 'hard samples', whereas high-quality segmentation belongs to the 'normal samples'. Accordingly, the most straightforward approach to measure the hardness is to evaluate the quality of segmentation. A simple yet efficient method is to train a neural network that categorizes segmentation outcomes as 'qualified' or 'unqualified'. To this purpose, we develop the network by integrating a segmentation backbone and a pre-trained classifier. During training, the seg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 453 + ], + "type": "text", + "content": "mentation results are compared with the ground truth, and the pre-trained classifier is further trained using the obtained classification labels for distinction during testing. The segmentation backbone employs a simple UNet architecture, enhanced with random perturbations in the final decoder layer, allowing the model to simultaneously produce multiple segmentation outcomes for a single image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 454, + 556, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 556, + 562 + ], + "type": "text", + "content": "To train the network, we label the segmentation results as 'qualified' or 'unqualified' based on a Dice score threshold derived from the ground truth. Since this threshold reflects instantaneous hardness, which may change during training, we use Bayesian optimization for dynamic adjustment instead of a fixed threshold. We initialize the threshold to " + }, + { + "bbox": [ + 313, + 454, + 556, + 562 + ], + "type": "inline_equation", + "content": "\\eta^0" + }, + { + "bbox": [ + 313, + 454, + 556, + 562 + ], + "type": "text", + "content": " and use the model performance in the validation set as the objective function " + }, + { + "bbox": [ + 313, + 454, + 556, + 562 + ], + "type": "inline_equation", + "content": "h(x; \\theta, \\eta)" + }, + { + "bbox": [ + 313, + 454, + 556, + 562 + ], + "type": "text", + "content": ". A Gaussian Process (GP) is used to approximate this objective function" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 567, + 553, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 567, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 365, + 567, + 553, + 581 + ], + "type": "interline_equation", + "content": "h (x; \\theta , \\eta) \\sim \\mathcal {G P} (m (\\eta); k (\\eta , \\eta^ {\\prime})), \\tag {1}", + "image_path": "bc0eff7996a9ed0169f15cafe79bd52b8116ee20aaf9ebe432cc8fdc33c0e1fa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 586, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 586, + 555, + 624 + ], + "type": "inline_equation", + "content": "m(\\eta)" + }, + { + "bbox": [ + 313, + 586, + 555, + 624 + ], + "type": "text", + "content": " is the mean function, " + }, + { + "bbox": [ + 313, + 586, + 555, + 624 + ], + "type": "inline_equation", + "content": "k(\\eta ,\\eta^{\\prime})" + }, + { + "bbox": [ + 313, + 586, + 555, + 624 + ], + "type": "text", + "content": " is the covariance function. The value of " + }, + { + "bbox": [ + 313, + 586, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 586, + 555, + 624 + ], + "type": "text", + "content": " is dynamically updated by maximizing the Expected Improvement (EI) function as follows" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 628, + 542, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 628, + 542, + 648 + ], + "spans": [ + { + "bbox": [ + 324, + 628, + 542, + 648 + ], + "type": "interline_equation", + "content": "\\eta^ {k + 1} = \\arg \\max _ {\\eta} \\left(\\mu (\\eta) - h (\\eta^ {k})\\right) \\Phi (Z) + \\sigma (\\eta) \\phi (Z),", + "image_path": "9c20df71cbf2d60f3e733634530769b1c77bf2ded069c0d4c3192d4f3796e3a9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mu (\\eta)" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\sigma (\\eta)" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " are the predicted mean and standard variation of the objective function, and " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " are the cumulative distribution function and probability density function with " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "Z = \\left(\\mu (\\eta) - h(\\eta^{k})\\right) / \\sigma (\\eta)" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": ". The gradient-descent optimization is used to solve the process. Then, we" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15644" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "reclassify normal and hard samples using the updated " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\eta^{k + 1}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " and repeat the above process in subsequent validation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": "Post-Former. We further refine the qualified segmentations associated with normal samples by a Post-Former module, which is a two-layer Transformer. Specifically, the segmentations provide " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": " to access " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 108, + 296, + 193 + ], + "type": "text", + "content": " from the shallow image features of the encoder of the segmentation backbone, and the refinement is performed through a cross-attention layer and a self-attention layer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 240, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 240, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 240, + 211 + ], + "type": "text", + "content": "3.2. Learning to segment hard samples" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "text", + "content": "Feature augmentation via diffusion process. It is well known that feature augmentation can enhance the generalizability of image segmentation models and improve the performance of hard samples. Wei et al. [40] analyzed the generalization error bound for deep neural network (DNN) models, showing that it is influenced by the Jacobian matrix and the trace of the Hessian matrix of the loss function. Let " + }, + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\bar{\\ell} = \\min \\{B,\\ell \\}" + }, + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "text", + "content": " represent the truncation of the loss function to some fixed bound " + }, + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "inline_equation", + "content": "B > 0" + }, + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "text", + "content": " as the population data distribution and " + }, + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "inline_equation", + "content": "P_{n}" + }, + { + "bbox": [ + 55, + 216, + 296, + 347 + ], + "type": "text", + "content": " as the distribution over training samples, the following generalization bound holds:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 352, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 396 + ], + "type": "text", + "content": "Theorem 1 (Theorem 4.1 in [40]). With probability " + }, + { + "bbox": [ + 55, + 352, + 296, + 396 + ], + "type": "inline_equation", + "content": "1 - \\delta" + }, + { + "bbox": [ + 55, + 352, + 296, + 396 + ], + "type": "text", + "content": " over the training examples, for all weight matrices " + }, + { + "bbox": [ + 55, + 352, + 296, + 396 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 352, + 296, + 396 + ], + "type": "text", + "content": " satisfying the norm bound " + }, + { + "bbox": [ + 55, + 352, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\| W^T\\|_{2,1} \\leq A" + }, + { + "bbox": [ + 55, + 352, + 296, + 396 + ], + "type": "text", + "content": ", the following bound holds" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 399, + 295, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 399, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 61, + 399, + 295, + 468 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {P} \\left[ \\bar {\\ell} \\right] - 1. 0 1 \\mathbb {E} _ {P _ {n}} \\left[ \\bar {\\ell} \\right] \\leq \\frac {\\left(A \\mu (W)\\right) ^ {\\frac {2}{3}} (\\theta B) ^ {\\frac {1}{3}}}{n ^ {\\frac {1}{3}}} \\\\ + \\frac {A \\sqrt {B \\nu (W)} \\theta}{\\sqrt {n}} + \\frac {B A ^ {2} \\theta}{n \\left(\\log^ {2} \\left(\\frac {B A ^ {2} \\theta}{\\nu (W) n}\\right) + 1\\right)} + \\zeta , \\tag {2} \\\\ \\end{array}", + "image_path": "ddfde572db09c7cedcb5ae19f8fda64f20ff36228044a5ab5f04856ff7df05c1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mu (W)" + }, + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\nu (W)" + }, + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "text", + "content": " are the Jacobian and Hessian matrix over samples, respectively, " + }, + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\theta = \\ln^3 (nM)\\max_i\\| x_i\\| _2^2" + }, + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 55, + 469, + 295, + 506 + ], + "type": "text", + "content": " is a low order term." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "text", + "content": "Thus, we can control both the Jacobian norm and the trace of the Hessian matrix to enhance the generalization of DNN models. While gradient descent primarily focuses on first-order derivatives, it lacks an effective mechanism to constrain the trace of the Hessian. Liu et al. [25] addressed this gap by introducing the trace of the Hessian as an additional penalty term in the loss function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 602, + 295, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 602, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 86, + 602, + 295, + 616 + ], + "type": "interline_equation", + "content": "\\ell_ {a l l} (f (x), y) \\approx \\ell (f (x), y) + \\lambda \\cdot \\operatorname {T r} \\left(H _ {\\ell , W}\\right), \\tag {3}", + "image_path": "d47e824ac8f28fec72118024880d29c8322a648ef0f858ea06d44f986b957811.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 622, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 295, + 634 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 622, + 295, + 634 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 622, + 295, + 634 + ], + "type": "text", + "content": " controls the strength of the Hessian regularization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": "In this work, we implement feature augmentation using a diffusion model, which is equivalent to implicitly penalizing the trace of the Hessian (see Section 3.3 for details). Specifically, we use random noise to progressively corrupt clean labels " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": " into noisy labels " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 122, + 700, + 228, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 700, + 228, + 714 + ], + "spans": [ + { + "bbox": [ + 122, + 700, + 228, + 714 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon", + "image_path": "e96ec1e2ccc410b1414685aef057f6de827810da8d97e1f09c9f7bb51eab885b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\alpha_t = 1 - \\beta_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " is used to parameterize the forward process for " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "t = 1, \\dots, T" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ". The forward process generates a series of noisy labels, enriching the feature space of hard samples. The reverse process iteratively removes noises to estimate clean labels. We use an encoder-decoder network as the denoiser, with " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "g_\\theta(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " encoding the noise variable and " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "d_\\theta(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " decoding to predict the " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ". We then have" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 177, + 507, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 177, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 361, + 177, + 507, + 190 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = d _ {\\theta} \\left(x _ {t}; m _ {\\theta} \\left(c _ {f}, g _ {\\theta} (x _ {t})\\right), t\\right),", + "image_path": "a407044bd63dc3e7928df6863f88d1b41375143ea9e2a96abf17e956f0ac030b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "text", + "content": "where the conditional image embedding " + }, + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "inline_equation", + "content": "c_{f}" + }, + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "text", + "content": " extracted from the backbone model is fused with the noise embedding " + }, + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "inline_equation", + "content": "g_{\\theta}(x_{t})" + }, + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "text", + "content": " in mix-former module " + }, + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "inline_equation", + "content": "m_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 198, + 555, + 305 + ], + "type": "text", + "content": ". After training, the diffusion model's decoder is discarded, and its gradients guide the encoder to better represent noisy labels as augmented features for hard sample segmentation. Notably, applying a similar diffusion procedure to normal samples yields minimal improvements, suggesting that the optimization strategy for samples of varying hardness should differ." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "text", + "content": "Mix-Former. For a segmentation task, the diffusion process must be conditional on the original image, which the proposed Mix-Former module implements. It integrates the noise embedding, hereafter referred to as " + }, + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "inline_equation", + "content": "g(\\epsilon)" + }, + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "text", + "content": " for clarity, and the conditional image embedding " + }, + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "inline_equation", + "content": "c_{f}" + }, + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "text", + "content": " obtained by the segmentation backbone model. In particular, Mix-Former consists of " + }, + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 313, + 316, + 554, + 426 + ], + "type": "text", + "content": " in our implementation) transformer blocks, each containing two cross-attention modules. We define the basic form of the attention mechanism with domain transformation as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 430, + 548, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 430, + 548, + 456 + ], + "spans": [ + { + "bbox": [ + 319, + 430, + 548, + 456 + ], + "type": "interline_equation", + "content": "F _ {A t t n} (q, k) = F ^ {- 1} \\left[ \\operatorname {s o f t m a x} \\left(\\frac {\\left(F (q) \\mathcal {W} _ {q}\\right) \\cdot \\left(F (k) \\mathcal {W} _ {k}\\right) ^ {T}}{\\sqrt {d _ {k}}}\\right) \\right],", + "image_path": "f19c91ede96c490a4ababb9aa39705746843340b81837120d96164695929219b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 460, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 460, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 460, + 554, + 533 + ], + "type": "text", + "content": " is the projection matrix in Fourier space, " + }, + { + "bbox": [ + 313, + 460, + 554, + 533 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 460, + 554, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 554, + 533 + ], + "type": "inline_equation", + "content": "F^{-1}" + }, + { + "bbox": [ + 313, + 460, + 554, + 533 + ], + "type": "text", + "content": " are the Fast Fourier Transform (FFT) and inverse transform, respectively. Here, frequency domain learning can enhances the model's ability to capture complex dependencies between noise and semantic information, thereby improving performance." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 321, + 543, + 548, + 633 + ], + "blocks": [ + { + "bbox": [ + 321, + 543, + 548, + 633 + ], + "lines": [ + { + "bbox": [ + 321, + 543, + 548, + 633 + ], + "spans": [ + { + "bbox": [ + 321, + 543, + 548, + 633 + ], + "type": "image", + "image_path": "28000a14b79c4654dac1cc502672780191a05c5f70749cbc2422549c8f0c6f51.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 641, + 555, + 666 + ], + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 666 + ], + "type": "text", + "content": "Figure 4. The structure of Mix-Former, where " + }, + { + "bbox": [ + 313, + 641, + 555, + 666 + ], + "type": "inline_equation", + "content": "g(\\epsilon)" + }, + { + "bbox": [ + 313, + 641, + 555, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 555, + 666 + ], + "type": "inline_equation", + "content": "c_{f}" + }, + { + "bbox": [ + 313, + 641, + 555, + 666 + ], + "type": "text", + "content": " are fused in the mix-former through " + }, + { + "bbox": [ + 313, + 641, + 555, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 641, + 555, + 666 + ], + "type": "text", + "content": " layers of repeated attention." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "text", + "content": "As shown in Figure 4, the first module integrates " + }, + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "inline_equation", + "content": "g(\\epsilon)" + }, + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "inline_equation", + "content": "c_{f}^{0}" + }, + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "text", + "content": ", while the second module encodes the embedding " + }, + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "inline_equation", + "content": "c_{f}^{1}" + }, + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "text", + "content": " back into " + }, + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "inline_equation", + "content": "g(\\epsilon)" + }, + { + "bbox": [ + 313, + 676, + 554, + 715 + ], + "type": "text", + "content": ". The bidirectional encoding enables the" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 446, + 716, + 462, + 727 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 716, + 462, + 727 + ], + "spans": [ + { + "bbox": [ + 446, + 716, + 462, + 727 + ], + "type": "inline_equation", + "content": "g(\\varepsilon)" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15645" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "model to learn the interaction between noise and semantic features to obtain a more robust representation [41]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 296, + 286 + ], + "type": "text", + "content": "Hard sample segmentation. Hard samples exhibit low-quality segmentation that deviates significantly from their labels. Thus, we concatenate the unqualified segmentation with its original image and use an encoder and pixel decoder to extract multi-scale features. A learnable object query accesses these features via cross-attention in the Transformer decoder. Augmented features from the diffusion model further enhance feature extraction. The learnable query dynamically selects relevant information for precise segmentation of hard samples. The Transformer decoder contains four cross-attention layers; see Figure 5. The first three layers fuse the learnable query with features from three scales of the pixel decoder, while the fourth layer integrates augmented features from the diffusion process to enhance segmentation. The details are as follows" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 298, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 298, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 298, + 335 + ], + "type": "interline_equation", + "content": "z _ {i} = \\left\\{ \\begin{array}{l l} \\operatorname {s o f t m a x} \\left(\\frac {\\left(W _ {q} L _ {q}\\right) \\cdot \\left(W _ {k} M _ {f} ^ {i}\\right) ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot W _ {v} M _ {f} ^ {i}, & i = 1, 2, 3, \\\\ F _ {A t t n} (z _ {3}, g (\\epsilon)) \\cdot W _ {v} g _ {\\natural} (\\epsilon), & i = 4, \\end{array} \\right.", + "image_path": "dcd1cd747fe80703211fa5a02455aba089c0b4ea9c239610f7a7d0955d3fe958.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "inline_equation", + "content": "L_{q}" + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "text", + "content": " represents the learnable query, " + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "inline_equation", + "content": "M_{f}^{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "text", + "content": " represents the multi-scale features by scaling with 32, 16 and " + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "text", + "content": " respectively, and " + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "text", + "content": " represents the attention-weighted feature. Such pattern is repeated in a round-robin fashion for all subsequent layers." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 396, + 290, + 503 + ], + "blocks": [ + { + "bbox": [ + 62, + 396, + 290, + 503 + ], + "lines": [ + { + "bbox": [ + 62, + 396, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 290, + 503 + ], + "type": "image", + "image_path": "e5254675850841c516dd3a5f6eed51bb9309efb7d5d718bd1fcb8fcaa0f0f201.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 511, + 295, + 534 + ], + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 534 + ], + "type": "text", + "content": "Figure 5. Hard samples segmentation decoder, where " + }, + { + "bbox": [ + 55, + 511, + 295, + 534 + ], + "type": "inline_equation", + "content": "L_{q}" + }, + { + "bbox": [ + 55, + 511, + 295, + 534 + ], + "type": "text", + "content": " access both multi-scale features " + }, + { + "bbox": [ + 55, + 511, + 295, + 534 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 55, + 511, + 295, + 534 + ], + "type": "text", + "content": " and augmented features " + }, + { + "bbox": [ + 55, + 511, + 295, + 534 + ], + "type": "inline_equation", + "content": "g(\\epsilon)" + }, + { + "bbox": [ + 55, + 511, + 295, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 553, + 185, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 185, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 185, + 566 + ], + "type": "text", + "content": "3.3. Implicit regularization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 571, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 643 + ], + "type": "text", + "content": "Subsequently, we formally demonstrate that the diffusion process in our model works as an implicit regularization to improve the model's generalization ability to hard samples. We aim to find the hard sample segmentation network " + }, + { + "bbox": [ + 55, + 571, + 296, + 643 + ], + "type": "inline_equation", + "content": "f(x;\\theta)" + }, + { + "bbox": [ + 55, + 571, + 296, + 643 + ], + "type": "text", + "content": " that minimizes the expected loss " + }, + { + "bbox": [ + 55, + 571, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathbb{E}[\\ell (f(x;\\theta),y)]" + }, + { + "bbox": [ + 55, + 571, + 296, + 643 + ], + "type": "text", + "content": ". Then we have the following result" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "content": "Proposition 1. For " + }, + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "inline_equation", + "content": "g(\\epsilon)" + }, + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "content": " denote the augmented features obtained by the diffusion probabilistic model. Then the expected loss becomes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 693, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 296, + 717 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ \\ell (f (x + g (\\epsilon)), y) ] \\approx \\ell (f (x), y) + \\frac {1}{2} \\operatorname {T r} (H \\Sigma_ {g}). \\tag {4}", + "image_path": "52160a18accfd6d59e2ff5b4e20e727a5a02c11f897936f4524ab14c3b8e13e2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "Proof. By incorporating the augmented features extracted from the noisy data, we have the expected loss function as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 380, + 104, + 488, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 104, + 488, + 118 + ], + "spans": [ + { + "bbox": [ + 380, + 104, + 488, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} [ \\ell (f (x + g (\\epsilon)), y) ],", + "image_path": "30348826e83172bd882608e0031ec97f87eea72a318bf3231dba7f513279ee84.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 125, + 518, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 518, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 518, + 137 + ], + "type": "text", + "content": "which can be approximated by Taylor's expansion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 144, + 555, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 167 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\approx \\mathbb {E} \\big [ \\ell (f (x), y) + \\nabla_ {x} \\ell (f (x), y) \\cdot g (\\epsilon) + \\frac {1}{2} g (\\epsilon) ^ {T} H g (\\epsilon) \\big ],", + "image_path": "060623585ff06bbc9f6edcbc42d92562bdaa1664e5a72b44efa347aa63a2ebbc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\ell (f(x),y)" + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "text", + "content": " to be the gradient and Hessian of the loss function w.r.t. the input " + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "text", + "content": " is of zero mean, by assuming that " + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "text", + "content": " is center distributed, i.e., " + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[g(\\epsilon)] = 0" + }, + { + "bbox": [ + 313, + 172, + 555, + 220 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 228, + 543, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 228, + 543, + 242 + ], + "spans": [ + { + "bbox": [ + 324, + 228, + 543, + 242 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ \\nabla_ {x} \\ell (f (x), y) \\cdot g (\\epsilon) ] = \\mathbb {E} [ g (\\epsilon) ] \\cdot \\nabla_ {x} \\ell (f (x), y) = 0.", + "image_path": "84d405eaddbd3d21b539cde049745ba4ea20b4a61680cf342ab5aceadffb29b8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 248, + 507, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 507, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 507, + 260 + ], + "type": "text", + "content": "Based on the properties of matrix trace, there is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 265, + 548, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 265, + 548, + 279 + ], + "spans": [ + { + "bbox": [ + 318, + 265, + 548, + 279 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ g (\\epsilon) ^ {T} H g (\\epsilon) ] = \\mathbb {E} [ \\operatorname {T r} (H g (\\epsilon) g (\\epsilon) ^ {T}) ] = \\operatorname {T r} (H \\mathbb {E} [ g (\\epsilon) g (\\epsilon) ^ {T} ]),", + "image_path": "e157b09a9b000e2ca4a5ed8bb69db985705909ee5ee309b16ac43cb264f73f0c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "text", + "content": "Assuming the covariance matrix of " + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "inline_equation", + "content": "g(\\epsilon)" + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "inline_equation", + "content": "\\Sigma_g" + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "inline_equation", + "content": "\\Sigma_g = \\mathbb{E}[g(\\epsilon)g(\\epsilon)^T]" + }, + { + "bbox": [ + 313, + 285, + 554, + 310 + ], + "type": "text", + "content": ". It follows that" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 375, + 316, + 492, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 316, + 492, + 331 + ], + "spans": [ + { + "bbox": [ + 375, + 316, + 492, + 331 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ g (\\epsilon) ^ {T} H g (\\epsilon) \\right] = \\operatorname {T r} \\left(H \\Sigma_ {g}\\right).", + "image_path": "019d4eb4f46de56f7d573b139643dc23c7a5c55db948979d1ddd62a8ab53797e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 338, + 485, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 485, + 349 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 485, + 349 + ], + "type": "text", + "content": "Then, the expected loss function becomes" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 332, + 357, + 535, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 357, + 535, + 379 + ], + "spans": [ + { + "bbox": [ + 332, + 357, + 535, + 379 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ \\ell (f (x + g (\\epsilon)), y) ] \\approx \\ell (f (x), y) + \\frac {1}{2} \\operatorname {T r} (H \\Sigma_ {g}),", + "image_path": "bde2e7a0398f57cb54c5091ae3471fd403631258bdc5cb7075ae55a6f360ce15.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 385, + 426, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 426, + 397 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 426, + 397 + ], + "type": "text", + "content": "which completes the proof." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 544, + 386, + 554, + 395 + ], + "blocks": [ + { + "bbox": [ + 544, + 386, + 554, + 395 + ], + "lines": [ + { + "bbox": [ + 544, + 386, + 554, + 395 + ], + "spans": [ + { + "bbox": [ + 544, + 386, + 554, + 395 + ], + "type": "image", + "image_path": "51e90a82c225beb8a6662c886bcfddaa33d1d3a5e1b94861c003c3f343be5a7e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "content": "By penalizing the trace of the Hessian matrix, the model is driven towards flatter minima, where the loss function is relatively stable, thereby improving generalization for hard sample segmentation [25]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 460, + 400, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 460, + 400, + 471 + ], + "spans": [ + { + "bbox": [ + 314, + 460, + 400, + 471 + ], + "type": "text", + "content": "3.4. Loss function" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 513 + ], + "type": "text", + "content": "Diffusion loss. The Diffusion Loss measures the difference between the ground truth noise added to the label and the noise predicted by the diffusion model" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 372, + 519, + 553, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 519, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 372, + 519, + 553, + 534 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} = \\left\\| \\epsilon - d _ {\\theta} \\left(t, x _ {t}, c _ {f}\\right) \\right\\| _ {2} ^ {2}, \\tag {5}", + "image_path": "8e7540d7f86d8e8537d1ec4d5b81d6067fb80e9ce59adb4679f99d0ddb7946d5.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "inline_equation", + "content": "c_{f}" + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "text", + "content": " is the conditional image embedding, " + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "text", + "content": " is the noise, and " + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "inline_equation", + "content": "d_{\\theta}(t,x_t,c_f)" + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "text", + "content": " is the noise predicted at time step " + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 566, + 554, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 649 + ], + "type": "text", + "content": "Segmentation loss. For hard samples, the segmentation loss " + }, + { + "bbox": [ + 313, + 566, + 554, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{hard} = \\mathcal{L}_{mask} + \\lambda_{cls}\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 566, + 554, + 649 + ], + "type": "text", + "content": " combines mask loss and classification loss. The mask loss includes binary cross-entropy and dice loss: " + }, + { + "bbox": [ + 313, + 566, + 554, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask} = \\lambda_{bce}\\mathcal{L}_{bce} + \\lambda_{dice}\\mathcal{L}_{dice}" + }, + { + "bbox": [ + 313, + 566, + 554, + 649 + ], + "type": "text", + "content": ", while the classification loss is cross-entropy. For normal samples, the segmentation loss " + }, + { + "bbox": [ + 313, + 566, + 554, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{normal}" + }, + { + "bbox": [ + 313, + 566, + 554, + 649 + ], + "type": "text", + "content": " uses only the mask loss. The total segmentation loss is:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 351, + 657, + 553, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 657, + 553, + 670 + ], + "spans": [ + { + "bbox": [ + 351, + 657, + 553, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e g}} = \\lambda_ {\\text {h a r d}} \\mathcal {L} _ {\\text {h a r d}} + \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}}, \\tag {6}", + "image_path": "85dfb9a0a16175aae3101b85ea0c959f80e31f560f0d692c5f4eb2ff6004dedd.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "with hyperparameters set as " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{bce} = \\lambda_{dice} = \\lambda_{cls} = 5" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{hard} = 5" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{normal} = 2" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ". The dice loss for the backbone model is weighted at 5, while the binary cross-entropy" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15646" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 561, + 222 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 561, + 222 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 561, + 222 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 561, + 222 + ], + "type": "table", + "html": "
MethodCOSMOS 2022ISIC 2018Kvasir-SEGBUSIAverage
DSCIoUDSC†DSCIoUDSC†DSCIoUDSC†DSCIoUDSC†DSCIoU
UNet[30]82.4472.7641.7687.1479.7342.2985.6378.6748.3576.4167.6141.0882.9074.69
PolypPVT[11]82.3372.3434.4389.8482.6152.4691.4486.3960.2681.0572.5146.6286.1778.46
nnUNet[18]84.1275.5855.7988.9181.6546.3190.5585.1853.4780.9472.4342.8786.1378.71
MedT[34]82.8173.1539.2688.8481.4944.0589.6884.1251.2380.4471.2740.1985.4477.51
TransUNet[5]82.9273.1843.6889.4482.3751.7691.0485.7556.7380.3271.3845.3185.9378.17
TransFuses[50]83.1773.8945.3589.9683.2254.9391.2786.3158.6481.5073.2548.2586.4879.17
SwinUNet[4]83.6374.7748.1789.5682.4049.5790.2285.0951.1879.7670.1939.7385.7978.11
MedSegDiffV2[41]82.8972.9339.5489.7882.6551.1291.0686.2152.3080.5371.4442.7486.0778.31
Mask2former[7]82.7672.8241.3289.5982.4649.5190.8685.2853.7780.1970.6343.3685.8577.80
L2S84.1675.6566.6790.9784.3463.9091.8787.0264.3481.8273.2753.2787.2180.07
", + "image_path": "93778f2cf1288b5b97313419ed06beae3b41a2578756557c8582e2efd47e242c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 555, + 264 + ], + "type": "text", + "content": "Table 1. Comparison of L2S with SOTA segmentation methods over different image modalities, where the results of SOTA methods are reproduced using their publicly available implementation with five-fold cross validation. " + }, + { + "bbox": [ + 55, + 229, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{DSC}^{\\dagger}" + }, + { + "bbox": [ + 55, + 229, + 555, + 264 + ], + "type": "text", + "content": " is computed on the hard samples with dice score below 0.7. We add the standard deviation and statistical significance in suppl. materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 284, + 214, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 214, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 214, + 296 + ], + "type": "text", + "content": "loss for the classifier is weighted at 10." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 296, + 295, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 332 + ], + "type": "text", + "content": "Consistency loss. To ensure consistency between normal and hard samples, which share the same label, we introduce the following consistency loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 342, + 295, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 342, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 114, + 342, + 295, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n s}} = \\left\\| p \\left(\\mathcal {D} _ {n}\\right) - f \\left(\\mathcal {D} _ {h}\\right) \\right\\| _ {2} ^ {2}, \\tag {7}", + "image_path": "ab62e5104546a6dd23fdea3c5b4483571e42cc7931b200d0aca08eee326ccfef.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "type": "text", + "content": "where the post-former " + }, + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "type": "text", + "content": " fine-tunes the normal sample set, and " + }, + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 365, + 296, + 414 + ], + "type": "text", + "content": " represents the segmentation network for the hard samples. To sum up, the total loss for our model is the weighted sum of all the aforementioned terms as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 425, + 295, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 425, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 94, + 425, + 295, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {s e g}} + \\lambda_ {\\text {d i f f}} \\mathcal {L} _ {\\text {d i f f}} + \\lambda_ {\\text {c o n s}} \\mathcal {L} _ {\\text {c o n s}}, \\tag {8}", + "image_path": "cd592c05ea454c4e134d958470ed0e9c8f037f4c380e2ea304e3d8ba8237b9c6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 447, + 285, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 285, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 285, + 460 + ], + "type": "text", + "content": "where the weights are fixed as " + }, + { + "bbox": [ + 55, + 447, + 285, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{diff} = 5" + }, + { + "bbox": [ + 55, + 447, + 285, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 447, + 285, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{cons} = 5" + }, + { + "bbox": [ + 55, + 447, + 285, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 472, + 178, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 178, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 178, + 485 + ], + "type": "text", + "content": "4. Experimental results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 248, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 248, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 248, + 505 + ], + "type": "text", + "content": "4.1. Datasets and implementation details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 509, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 629 + ], + "type": "text", + "content": "We conducted experiments on six public medical segmentation datasets to evaluate the performance of our model across various imaging modalities. These include the COSMOS2022 dataset [55], ISIC2018 dataset [28], Kvasir-SEG dataset [19], and the BUSI dataset [1] for carotid artery, skin lesion, polyp and breast ultrasound segmentation, respectively. The BTCV dataset [12], containing 12 anatomical structures, was used for abdominal multi-organ segmentation, while the ACDC dataset [2] was utilized for segmenting cardiac structures in MRI images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": "For the hard sample segmentation network, we used a pre-trained Swin Transformer [26] as the encoder and MSDeformAttn [56] as the pixel decoder, similar to Mask2Former [7]. The Transformer decoder was configured with " + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": " (12 layers in total) and 100 queries by default. We trained our model using the AdamW optimizer with a learning rate of 0.01 and a weight decay of " + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 284, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 463 + ], + "type": "text", + "content": "Training images were resized to " + }, + { + "bbox": [ + 313, + 284, + 555, + 463 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 284, + 555, + 463 + ], + "type": "text", + "content": " for the COSMOS2022, ACDC, and BTCV datasets. For other datasets, the images were resized to " + }, + { + "bbox": [ + 313, + 284, + 555, + 463 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 313, + 284, + 555, + 463 + ], + "type": "text", + "content": " with color jittering and random horizontal flipping. The training schedule included 12,000 iterations for COSMOS2022, 30,000 for ACDC and BTCV, and 20,000 for other datasets. All experiments were conducted using PyTorch 1.10.0 on 4 NVIDIA RTX 3090 GPUs. The Dice score with an initial threshold of 0.7 was used to distinguish between normal and hard samples. During training, normal and hard samples generated by the backbone model were stored separately. For each input image, a pair of normal and hard samples was randomly selected from their respective sets for further segmentation. Each image's two sets contain at least one element to ensure the balance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 465, + 556, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 556, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 556, + 550 + ], + "type": "text", + "content": "We compared our L2S with SOTA segmentation methods, including CNN-based methods like UNet [30], nnUNet [18], and transformer-based methods such as Polyp-PVT [11], MedT [34], TransUNet [5], TransFuse [50], SwinUNet [4], and the diffusion-based method MedSegDiffV2 [41]. We also compared L2S with the baseline model Mask2Former [7] effectiveness of the proposed framework." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 562, + 422, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 422, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 422, + 574 + ], + "type": "text", + "content": "4.2. Results evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Comparisons on binary segmentation tasks. Our L2S model is compared with SOTA segmentation methods across four tasks; see Table 1, where the best results are shown in bold, and the same is applied in the following tables. The results demonstrate that our L2S model achieves the highest average Dice of " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "87.21\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " and IoU of " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "80.07\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ". Improving hard sample segmentation yields significant gains: L2S outperforms the SOTA methods by 10. " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " in COSMOS2022, " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "8.97\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " in ISIC2018, and " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "4.08\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " in the polyp-specific PolypPVT model in Kvasir-SEG. These results highlight that the existing methods perform poorly" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15647" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 550, + 262 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 550, + 262 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 550, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 550, + 262 + ], + "type": "image", + "image_path": "58e3803bfcbfef9e5f0711f6533a9d382ba5dd9b6d26a6c45f6bf00bfb11e5e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 269, + 534, + 281 + ], + "lines": [ + { + "bbox": [ + 75, + 269, + 534, + 281 + ], + "spans": [ + { + "bbox": [ + 75, + 269, + 534, + 281 + ], + "type": "text", + "content": "Figure 6. Qualitative segmentation results for hard samples, including carotid artery, breast ultrasound, polyp, and skin lesions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 292, + 552, + 408 + ], + "blocks": [ + { + "bbox": [ + 59, + 292, + 552, + 408 + ], + "lines": [ + { + "bbox": [ + 59, + 292, + 552, + 408 + ], + "spans": [ + { + "bbox": [ + 59, + 292, + 552, + 408 + ], + "type": "table", + "html": "
MethodSpleenR.KidL.KidGall.Eso.LiverStom.AortaIVCVeinsPanc.AGAvg.
UNet[30]0.9330.8790.9130.7120.7080.9390.8200.8570.7850.7070.6810.6370.797
nnUNet[18]0.9420.8940.9100.7040.7230.9480.8240.8770.7820.7200.6800.6160.802
TransUNet[5]0.9520.9270.9290.6620.7570.9690.8890.9200.8330.7910.7750.6370.837
TransFuses[50]0.9690.9240.9310.7720.7880.9740.9030.9140.8360.8020.7850.7410.861
SwinUNet[4]0.9390.9160.9080.7270.7340.9410.8520.9030.7980.7550.7330.6460.821
MedSegDiffV2[41]0.9480.9250.9160.7940.8220.9530.8910.9180.8390.7980.7910.7270.860
Mask2former[7]0.9560.9130.9210.7340.7550.9510.8920.9110.8350.7950.7530.6860.842
L2S0.9730.9340.9450.8260.8050.9720.9180.9220.8460.8240.8130.8020.882
", + "image_path": "8a8d351744400e1e8ce3520a963947af6e269fde0e1b47d1ce4f86c8fac866e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 449, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 616 + ], + "type": "text", + "content": "on hard samples. Our L2S achieves over a " + }, + { + "bbox": [ + 54, + 449, + 295, + 616 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 449, + 295, + 616 + ], + "type": "text", + "content": " improvement in accuracy compared to Mask2Former. A detailed analysis of the selected hard samples is provided in Supplementary Material B.4 to validate our assumption that hard samples correspond to poor segmentation. Figure 6 presents the segmentation results of representative hard samples. Our L2S model excels at segmenting complex structures. Qualitative results for carotid arteries, breast ultrasound, polyps, and skin lesions show that the predictions of our L2S better match the ground truth. For carotid artery segmentation, our L2S model precisely identifies stenosis, offering important clinical value. Overall, our designs significantly improve robustness and generalization. The visualized results are available in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "Comparisons on multi-organ segmentation. Table 2 shows that our L2S outperforms the SOTA methods on BTCV multi-organ dataset, achieving the highest average Dice score of " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "88.2\\%" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Table 3 lists the Dice scores of L2S and the comparison methods on ACDC dataset for cardiac organ segmentation. Our L2S gains an average Dice score of " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "91.36\\%" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": ", outperforming nnUNet by " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.33\\%" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": " and improving right ventricle segmentation by " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.48\\%" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": ". The comparison" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 447, + 550, + 561 + ], + "blocks": [ + { + "bbox": [ + 65, + 416, + 543, + 428 + ], + "lines": [ + { + "bbox": [ + 65, + 416, + 543, + 428 + ], + "spans": [ + { + "bbox": [ + 65, + 416, + 543, + 428 + ], + "type": "text", + "content": "Table 2. Comparison of L2S with SOTA segmentation methods over multi-organ segmentation BTCV dataset in term of Dice Score." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 447, + 550, + 561 + ], + "lines": [ + { + "bbox": [ + 320, + 447, + 550, + 561 + ], + "spans": [ + { + "bbox": [ + 320, + 447, + 550, + 561 + ], + "type": "table", + "html": "
MethodsRVMyoLVAvg.
UNet86.7980.4593.6386.96
nnUNet89.5988.3195.2091.03
TransUNet87.0587.1794.5289.58
TransFuses87.9487.0395.0990.02
SwinUNet86.5785.3394.1088.67
Mask2former86.7486.4993.9289.05
L2S(UNet)90.0788.6595.3691.36
L2S(Mask2former)90.4289.0295.6191.68
", + "image_path": "8dcf6587506b6848bcbbe5f176347d144a04b93bef6113b14dbff3c5f9ffaeb9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 568, + 554, + 592 + ], + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 592 + ], + "type": "text", + "content": "Table 3. Results of cardiac organ segmentation on ACDC dataset, where dice scores are reported for individual organs." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 614, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 555, + 662 + ], + "type": "text", + "content": "experiments on different backbone models show that more powerful Mask2Former can further enhance segmentation accuracy. Thus, the L2S framework is highly compatible and can be integrated with different segmentation models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 671, + 414, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 414, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 414, + 683 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "We conducted a series of ablation studies on the ISIC2018 dataset unless otherwise noted. Table 4 shows an ablation" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15648" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "study on FFT, consistency loss, and adaptive threshold selection. As shown, FFT significantly impacts the Dice score with an improvement of " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "0.31\\%" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": ", highlighting the importance of feature alignment." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 130, + 284, + 224 + ], + "blocks": [ + { + "bbox": [ + 67, + 130, + 284, + 224 + ], + "lines": [ + { + "bbox": [ + 67, + 130, + 284, + 224 + ], + "spans": [ + { + "bbox": [ + 67, + 130, + 284, + 224 + ], + "type": "table", + "html": "
MethodDSC (%)IoU (%)
L2S90.9784.34
- FFT90.66 (-0.3)83.95 (-0.4)
- consistency loss90.78 (-0.2)84.07 (-0.3)
- adaptive threshold90.86 (-0.1)84.24 (-0.1)
- all 3 components90.54 (-0.4)83.72 (-0.6)
", + "image_path": "89e7104f9d85cef90f83db47a3320a9d17ce894ac35b97146a0da9940971dd26.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "text", + "content": "Hard sample identification and augmentation. Table 5 presents an ablation study evaluating the impact of incorporating normal and hard sample identification mechanisms and feature augmentation modules on model performance. The results show that removing the identification mechanism (i.e., no longer distinguishing the segmentation results of the backbone model, treating all samples in the dataset as hard samples for subsequent processing, and therefore using no classifier during testing) causes a decrease in the Dice score from " + }, + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "inline_equation", + "content": "90.87\\%" + }, + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "inline_equation", + "content": "89.59\\%" + }, + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "text", + "content": ", which indicates that complex processing may adversely affect the original normal samples in the dataset. In addition, using diffusion models for feature augmentation significantly improves the hard sample segmentation, increasing the Dice score by " + }, + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "inline_equation", + "content": "4.52\\%" + }, + { + "bbox": [ + 54, + 268, + 295, + 460 + ], + "type": "text", + "content": ". Incorporating the Post-former module slightly enhances segmentation by refining normal samples." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 470, + 299, + 563 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "type": "text", + "content": "Table 4. Three optimization improvements in network operation, where we remove one component at a time." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 470, + 299, + 563 + ], + "lines": [ + { + "bbox": [ + 56, + 470, + 299, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 470, + 299, + 563 + ], + "type": "table", + "html": "
CLFDDPMP-FormerDSC (%)IoU (%)DSC† (%)
89.5282.35-
✓90.2883.7458.96
✓89.5982.48-
✓✓90.8784.11-
✓✓90.3683.7959.38
✓✓✓90.9784.3463.90
", + "image_path": "dcdca98fcef5f52924f2583565fecc56e22b06d7ba7645ac3f1970397b1ea56d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": "Hard sample segmentation. Table 6 shows the impact of various design choices on the performance of the hard sample segmentation network, measured by the Dice score. Using query vectors alongside hard sample features boost the Dice score from " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "55.83\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "59.88\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": ". Besides, multiscale features and query learnability can further enhance hard sample segmentation accuracy, by increasing Dice of more than " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": ", respectively. We also compare the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 70, + 547, + 148 + ], + "blocks": [ + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "type": "text", + "content": "Table 5. Ablation study on our L2S method, where 'CLF', 'DDPM' and 'P-Former' denote the pretrained classifier, feature augmentation, and Post-Former, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 70, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 547, + 148 + ], + "type": "table", + "html": "
MethodDSC† (%)FLOPs
single scale (w/o query)55.8315.2G
single scale62.2432.7G
- learnable query features59.8832.7G
L2S63.9024.3G
- learnable query features60.8224.3G
", + "image_path": "6d977e3d75d6af65d2aad1866dc07762f5b617243cdabb7eb530dfebf4146dee.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 204, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 240 + ], + "type": "text", + "content": "FLOPs of different designs for hard sample segmentation on the ISIC dataset. The multiscale approach achieved better performance with fewer FLOPs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 254, + 387, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 387, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 387, + 266 + ], + "type": "text", + "content": "4.4. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 273, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 406 + ], + "type": "text", + "content": "We introduce a specialized framework for hard sample segmentation in medical images, advancing current methods. Our approach includes a dedicated sample generator and a classifier to distinguish normal from hard samples. In addition, our results also indicate that classifier accuracy significantly influences overall model performance, and substituting direct labels during testing can yield additional gains; see Table 7. While not the primary focus, these enhancements suggest valuable directions for future research, particularly in improving model learning and segmentation on complex medical image datasets." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 418, + 557, + 498 + ], + "blocks": [ + { + "bbox": [ + 313, + 156, + 553, + 179 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 179 + ], + "type": "text", + "content": "Table 6. Ablation study on Learnable query and Mutil-scale features, where 'single-scale' is the single resolution of " + }, + { + "bbox": [ + 313, + 156, + 553, + 179 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 313, + 156, + 553, + 179 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 418, + 557, + 498 + ], + "lines": [ + { + "bbox": [ + 315, + 418, + 557, + 498 + ], + "spans": [ + { + "bbox": [ + 315, + 418, + 557, + 498 + ], + "type": "table", + "html": "
DatasetAcc (%)Rec (%)DSC (%)DSC* (%)
COSMOS202292.782.384.1684.34
ISIC201883.372.690.9791.28
BUSI78.670.481.8290.04
Kvasir-SEG85.975.291.8792.10
ACDC87.571.791.3691.45
", + "image_path": "69a2ffb1547af5a539a51addc41df3292e8b00725ec9a7d366b93bc55a0b4691.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 506, + 555, + 540 + ], + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 540 + ], + "type": "text", + "content": "Table 7. The impact of the classifier to segmentation, where Acc and Rec denote the overall accuracy and recall of hard samples, respectively. " + }, + { + "bbox": [ + 313, + 506, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{DSC}^*" + }, + { + "bbox": [ + 313, + 506, + 555, + 540 + ], + "type": "text", + "content": " is obtained by classification using true labels." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 571, + 435, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 435, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 435, + 586 + ], + "type": "text", + "content": "5. Concluding remarks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "We propose a novel segmentation framework that can identify hard samples and enhance segmentation performance by employing stronger networks and more sophisticated optimization strategies on selected hard samples. The framework incorporates a diffusion process as an implicit regularization to handle hard samples, achieved by leveraging the inherent distribution bias of the dataset combined with random noise. Extensive experiments across various datasets demonstrated the superiority of our method over SOTA approaches, particularly for challenging samples." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15649" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 138, + 115, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 138, + 115, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 138, + 115, + 150 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 158, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 158, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 294, + 191 + ], + "type": "text", + "content": "[1] Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images. Data in brief, 28:104863, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 193, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 296, + 269 + ], + "type": "text", + "content": "[2] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514-2525, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 271, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 271, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 271, + 296, + 304 + ], + "type": "text", + "content": "[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 296, + 360 + ], + "type": "text", + "content": "[4] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 1, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 296, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 416 + ], + "type": "text", + "content": "[5] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 419, + 296, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 296, + 473 + ], + "type": "text", + "content": "[6] Mengzhao Chen, Mingbao Lin, Ke Li, Yunhang Shen, Yongjian Wu, Fei Chao, and Rongrong Ji. Cf-vit: A general coarse-to-fine method for vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7042-7052, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 531 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "text", + "content": "[8] Jiequan Cui, Yuhui Yuan, Zhisheng Zhong, Zhuotao Tian, Han Hu, Stephen Lin, and Jiaya Jia. Region rebalance for long-tailed semantic segmentation. arXiv preprint arXiv:2204.01969, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 633 + ], + "type": "text", + "content": "[9] Yin Cui, Feng Zhou, Yuanqing Lin, and Serge Belongie. Fine-grained categorization and dataset bootstrapping using deep metric learning with humans in the loop. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1153-1162, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "text", + "content": "[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[11] Bo Dong, Wenhai Wang, Deng-Ping Fan, Jinpeng Li, Huazhu Fu, and Ling Shao. Polyp-pvt: Polyp segmentation with pyramid vision transformers. arXiv preprint arXiv:2108.06932, 2021. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "type": "text", + "content": "The work is supported by the National Natural Science Foundation of China (NSFC 12071345)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[12] Xi Fang and Pingkun Yan. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging, 39(11):3619–3629, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 162 + ], + "type": "text", + "content": "[13] Jun Fu, Jing Liu, Hajjie Tian, et al. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3146-3154, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 554, + 206 + ], + "type": "text", + "content": "[14] Ethan Harris, Antonia Marcu, Matthew Painter, Mahesan Niranjan, Adam Prügel-Bennett, and Jonathon Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "type": "text", + "content": "[15] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7132-7141, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 310, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 554, + 363 + ], + "type": "text", + "content": "[17] Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to segment the tail. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14045-14054, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 366, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 554, + 409 + ], + "type": "text", + "content": "[18] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 1, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 477 + ], + "type": "text", + "content": "[19] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pál Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia modeling: 26th international conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, proceedings, part II 26, pages 451–462. Springer, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 479, + 554, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 554, + 511 + ], + "type": "text", + "content": "[20] Longlong Jing, Yucheng Chen, and Yingli Tian. Coarse-to-fine semantic segmentation from image-level labels. IEEE transactions on image processing, 29:225-236, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 513, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 554, + 555 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 557, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 554, + 612 + ], + "type": "text", + "content": "[22] Aybora Koksal, Önder Tuzcuoglu, Kutalmış Gökalp İnce, Yoldaş Ataseven, and A Aydin Alatan. Improved hard example mining approach for single shot object detectors. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3536-3540. IEEE, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "type": "text", + "content": "[23] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3202, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[24] Xuhang Lian, Yanwei Pang, Jungong Han, and Jing Pan. Cascaded hierarchical atrous spatial pyramid pooling module for semantic segmentation. Pattern Recognition, 110: 107622, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15650" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[25] Yucong Liu, Shixing Yu, and Tong Lin. Regularizing deep neural networks with stochastic estimators of hessian trace. arXiv preprint arXiv:2208.05924, 2022. 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 161 + ], + "type": "text", + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "text", + "content": "[27] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "type": "text", + "content": "[28] Md Ashraful Alam Milton. Automated skin lesion classification using ensemble of deep neural networks in isic 2018: Skin lesion analysis towards melanoma detection challenge. arXiv preprint arXiv:1901.10802, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[29] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 382 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[31] Seong-Eun Ryu and Kyung-Yong Chung. Detection model of occluded object based on yolo using hard-example mining and augmentation policy optimization. Applied Sciences, 11 (15):7093, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "text", + "content": "[32] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[33] Abhinav Shrivastava, Rahul Sukthankar, and Mohammad Hejrati. Training region-based object detectors with online hard example mining. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 761-769, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 592 + ], + "type": "text", + "content": "[34] Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Hacihaliloglu, and Vishal M Patel. Medical transformer: Gated axial-attention for medical image segmentation. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, part I 24, pages 36-46. Springer, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 647 + ], + "type": "text", + "content": "[35] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In International conference on machine learning, pages 6438-6447. PMLR, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "text", + "content": "[36] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9695-9704, 2021. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "type": "text", + "content": "[37] Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model-agnostic segmentation refinement with discrete diffusion process. arXiv preprint arXiv:2312.12425, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 119, + 555, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 174 + ], + "type": "text", + "content": "[38] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7303-7313, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[39] Xue Wang, Zhanshan Li, Yongping Huang, and Yingying Jiao. Multimodal medical image segmentation using multiscale context-aware network. Neurocomputing, 486:135-146, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 221, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 221, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 221, + 553, + 266 + ], + "type": "text", + "content": "[40] Colin Wei, Sham Kakade, and Tengyu Ma. The implicit and explicit regularization effects of dropout. In International conference on machine learning, pages 10181-10192. PMLR, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 267, + 553, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 267, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 267, + 553, + 322 + ], + "type": "text", + "content": "[41] Junde Wu, Wei Ji, Huazhu Fu, Min Xu, Yueming Jin, and Yanwu Xu. Medsegdiff-v2: Diffusion-based medical image segmentation with transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6030-6038, 2024. 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 324, + 553, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 324, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 316, + 324, + 553, + 368 + ], + "type": "text", + "content": "[42] Xin Xiao, Daiguo Zhou, Jiagao Hu, Yi Hu, and Yongchao Xu. Not all pixels are equal: Learning pixel hardness for semantic segmentation. arXiv preprint arXiv:2305.08462, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[43] Jiahao Xie, Sheng Zhang, Jianwei Lu, and Ye Luo. L-snet: From region localization to scale invariant medical image segmentation. In 2021 IEEE International Conference on Image Processing (ICIP), pages 41-45. IEEE, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 417, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 449 + ], + "type": "text", + "content": "[44] Jin Yin, Pengfei Xia, and Jingsong He. Online hard region mining for semantic segmentation. Neural Processing Letters, 50:2665-2679, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 452, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 517 + ], + "type": "text", + "content": "[45] Qihang Yu, Dong Yang, Holger Roth, Yutong Bai, Yixiao Zhang, Alan L Yuille, and Daguang Xu. C2fnas: Coarse-to-fine neural architecture search for 3d medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4126-4135, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "text", + "content": "[46] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 577, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 632 + ], + "type": "text", + "content": "[47] Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3457-3466, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 634, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 553, + 667 + ], + "type": "text", + "content": "[48] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[49] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15651" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 462 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[50] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, Part I 24, pages 14-24. Springer, 2021. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "text", + "content": "[51] Xiangyun Zhao, Raviteja Vemulapalli, Philip Andrew Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive learning for label efficient semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10623-10633, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 207, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 294, + 260 + ], + "type": "text", + "content": "[52] Xiaoqi Zhao, Hongpeng Jia, Youwei Pang, Long Lv, Feng Tian, Lihe Zhang, Weibing Sun, and Huchuan Lu. M" + }, + { + "bbox": [ + 56, + 207, + 294, + 260 + ], + "type": "inline_equation", + "content": "^{\\wedge}\\{2\\}" + }, + { + "bbox": [ + 56, + 207, + 294, + 260 + ], + "type": "text", + "content": " snet: Multi-scale in multi-scale subtraction network for medical image segmentation. arXiv preprint arXiv:2303.10894, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 262, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 305 + ], + "type": "text", + "content": "[53] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13001-13008, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 340 + ], + "type": "text", + "content": "[54] Tongxue Zhou, Su Ruan, and Stephane Canu. A review: Deep learning for medical image segmentation using multimodality fusion. *Array*, 3:100004, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 417 + ], + "type": "text", + "content": "[55] Chenglu Zhu, Xiaoyan Wang, Shengyong Chen, Zhongzhao Teng, Cong Bai, Xiaojie Huang, Ming Xia, Zhanpeng Shao, Zheng Gu, and Peiliang Sun. Complex carotid artery segmentation in multi-contrast mr sequences by improved optimal surface graph cuts based on flow line learning. Medical & Biological Engineering & Computing, 60(9):2693-2706, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 419, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 419, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 419, + 294, + 462 + ], + "type": "text", + "content": "[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 6" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15652" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/e2161c88-f390-4144-acf4-52130bcde3bd_content_list.json b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/e2161c88-f390-4144-acf4-52130bcde3bd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bacfe0922d4d9847b11e24acb261412dc5ee25ca --- /dev/null +++ b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/e2161c88-f390-4144-acf4-52130bcde3bd_content_list.json @@ -0,0 +1,1426 @@ +[ + { + "type": "text", + "text": "Taming Teacher Forcing for Masked Autoregressive Video Generation", + "text_level": 1, + "bbox": [ + 145, + 130, + 852, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deyu Zhou1 Quan Sun2 Yuang Peng2,4 Kun Yan2 Runpei Dong3 Duomin Wang2 Zheng Ge2 Nan Duan2 Xiangyu Zhang2", + "bbox": [ + 142, + 178, + 852, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ HKUST(GZ) $^{2}$ StepFun $^{3}$ UIUC $^{4}$ THU MAGI-VIDEO-GENERATION.GITHUB.IO", + "bbox": [ + 285, + 220, + 710, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce MAGI, a hybrid video generation framework that combines masked modeling for intra-frame generation with causal modeling for next-frame generation. Our key innovation, Complete Teacher Forcing (CTF), conditions masked frames on complete observation frames rather than masked ones (namely Masked Teacher Forcing, MTF), enabling a smooth transition from token-level (patch-level) to frame-level autoregressive generation. CTF significantly outperforms MTF, achieving a $+23\\%$ improvement in FVD scores on first-frame conditioned video prediction. To address issues like exposure bias, we employ targeted training strategies, setting a new benchmark in autoregressive video generation. Experiments show that MAGI can generate long, coherent video sequences exceeding 100 frames, even when trained on as few as 16 frames, highlighting its potential for scalable, high-quality video generation.", + "bbox": [ + 88, + 325, + 485, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 597, + 220, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating order matters in autoregressive image generation. While existing approaches mostly reply on simple raster-scan order, recent studies [8, 22] have demonstrated superior results through alternative generation strategies. These include randomized spatial order with bidirectional attention [4, 22, 51] and generation along increased scales [5, 8]. Despite its fundamental importance, in the realm of autoregressive video generation, however, the discussion of generation order has been largely overlooked.", + "bbox": [ + 89, + 622, + 482, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing autoregressive video generation methods can be categorized into two groups according to the prediction granularity, as shown in Tab. 1. The first group, exemplified by MAGViT [51], adopts a masked modeling approach* where visual tokens are generated in descending order according to their logit probabilities. While this approach uses bi-directional attention across frames, it presents two signif", + "bbox": [ + 89, + 758, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "icant limitations: substantial computational overhead during inference since it is unable to utilize KV Cache [9, 23], and neglect temporal causality between consecutive frames. These oversights are particularly noteworthy given the remarkable success of causal modeling in large language models. Some other methods in this group [2, 6] have explored frame-level prediction with casual temporal attention. However, these methods are limited due to incomplete history observation, which causes an intrinsic training-inference gap.", + "bbox": [ + 511, + 294, + 903, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The second group comprises fully autoregressive approaches operating on visual patches [20, 46, 48, 50], exemplified by recent works like VideoPoet [20] and Emu3 [46]. Despite their successful temporal modeling, these methods have not incorporated contemporary advances from image generation research in intra-frame generation. Instead, they continue to rely on raster-scan order - an approach demonstrated sub-optimal in image generation [22].", + "bbox": [ + 511, + 446, + 905, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce Masked Autoregressive video Generation (MAGI), a hybrid framework that synergizes the strengths of both video generation paradigms by combining masked modeling for intra-frame generation with causal modeling for inter-frame dependencies. We start with a naive implementation. This initial implementation explores a straightforward approach where tokens in each next frame are generated conditioned on previous masked frames (Figure 1(a)). While this can be viewed as a natural extension of MaskGIT [4] and Muse [5] to autoregressive video generation, our analysis reveals a fundamental limitation in its implementation of teacher forcing (TF) [47]. To be specific, TF involves feeding the model the groundtruth (correct) output from the previous time step as input during training, rather than using the model's own predicted output. However, our initial approach, which we term as Masked Teacher Forcing (MTF), deviates from this principle by substituting ground-truth tokens with mask tokens. Despite similar approaches being explored in interactive generative models (e.g., Genie [2]), we identify a critical training-inference discrepancy: during training, next-frame generation is conditioned on a mixture of visible", + "bbox": [ + 511, + 569, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*We follow the definition in MAR [22], which considers masked image generation as a form of auto-regressive generation with a certain order.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7374", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/938e2ae0f55b4ea66b5a1dd04a9ffc685e49dcc5d79e75eaf80ea1ddb204b944.jpg", + "table_caption": [ + "Table 1. Comparison of autoregressive paradigms across different methods. We compare MAGI with various methods based on temporal attention mechanisms (how each frame attends to others), support for KV Cache [23] and variable context lengths, observation completeness (the completeness of the context frames used for conditioning), and the prediction granularity (patch or frame). $\\dagger$ : MAGVIT predicts masked tokens of multiple frames simultaneously. Genie [2] and MAGI predict multiple tokens of single frame simultaneously." + ], + "table_footnote": [], + "table_body": "
Prediction GranularityMethodTemporal AttentionKV CacheVar. ContextObservation
PatchVideoGPT [50], Phenaki [43], Omni [45]Causal✓✓Complete
FrameMAGVIT† [51], GameNGen [40]Bidirectional✗✗Complete
Diffusion Forcing [6]Causal✓✓Noisy
Genie [2]Causal✓✓Masked
MAGI (with CTF)Causal✓✓Complete
", + "bbox": [ + 93, + 155, + 903, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and masked tokens, while during inference, it relies solely on complete visual tokens from previous frames. Our experiments demonstrate that this inconsistency compromises generation quality, especially motion coherency.", + "bbox": [ + 89, + 273, + 482, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose Complete Teacher Forcing (CTF), an advanced paradigm that maintains training-test consistency by conditioning next-frame generation on complete visual observations from history frames during training. This is achieved through a novel mechanism that prepends complete video tokens to the masked input sequence and employs a carefully designed attention mask. Empirical results show that CTF better captures motion, outperforming MTF by $23\\%$ in terms of FVD scores on first-frame conditioned video prediction.", + "bbox": [ + 88, + 335, + 482, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we tackle the persistent challenge of exposure bias [32] in autoregressive video generation through two complementary strategies: dynamic interval training and noise injection [40]. Dynamic interval training refers to randomly sampling frames with varying intervals during training, which introduces diversity into the data distribution and helps the model to better generalize to different temporal frequencies. Dynamic noise injection, on the other hand, involves adding random noise to observation frames during training, which helps improve the model's robustness by simulating the possible errors may occur in inference. Combined with these two techniques, MAGI establishes a robust baseline for autoregressive video generation. Through extensive evaluations, we demonstrate that MAGI achieves superior video generation quality and length scalability. For example, MAGI can generate coherent video sequences that exceed 100 frames, even when trained on sequences as short as 16 frames. This underscores the potential of our approach for scalable video generation.", + "bbox": [ + 89, + 487, + 482, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Problem Statement", + "text_level": 1, + "bbox": [ + 89, + 791, + 421, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Autoregressive Video Generation", + "text_level": 1, + "bbox": [ + 89, + 816, + 380, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a video $V \\in \\mathbb{R}^{N \\times H \\times W \\times 3}$ , the sequence of $N$ frames with height $H$ and width $W$ can be denoted as $V = \\{f_i \\in \\mathbb{R}^{H \\times W \\times 3} | i = 1, \\dots, N\\}$ . Each frame can be further divided into $m$ patches, represented as", + "bbox": [ + 89, + 838, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\{\\mathbf{x}_{11}, \\mathbf{x}_{12}, \\ldots, \\mathbf{x}_{1m}, \\ldots, \\mathbf{x}_{n1}, \\mathbf{x}_{n2}, \\ldots, \\mathbf{x}_{nm}\\}$ . Autoregressive video generation models primarily differ in the granularity of causality (patch-level vs. frame-level) and the training paradigm (teacher forcing), discussed below.", + "bbox": [ + 511, + 273, + 903, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Patch-level Methods Patch-level autoregressive models [46, 50] operate by modeling the video on a patch-by-patch basis. The likelihood of generating entire video with a model parameterized by $\\theta$ is decomposed into the product of posterior of each patch conditioned all preceding patches:", + "bbox": [ + 511, + 334, + 903, + 411 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np (V) = \\prod_ {t = 1} ^ {N \\times m} p \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {1}, \\mathbf {x} _ {2}, \\dots , \\mathbf {x} _ {t - 1}; \\theta\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 417, + 903, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $p_t$ represents a patch in a predefined ordering. The model generates each patch sequentially by conditioning on all previously generated patches, capturing both spatial and temporal dependencies.", + "bbox": [ + 511, + 467, + 906, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Frame-level Methods Frame-level autoregressive models [2, 40] model the video at the granularity of entire frames. Similarly, the likelihood of video modeling is:", + "bbox": [ + 511, + 527, + 903, + 574 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np (V) = \\prod_ {t = 1} ^ {n} p \\left(f _ {t} \\mid f _ {1}, f _ {2}, \\dots , f _ {t - 1}; \\theta\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 580, + 903, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $f_{t}$ is the entire frame generated conditioned on all history frames. In this fashion, the patches in each frame are generated in parallel.", + "bbox": [ + 511, + 628, + 903, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Teacher Forcing Teacher forcing [47] is a widely adopted training strategy in autoregressive models, where the model conditions on the true previous frame or patch rather than its own predictions [10, 47]. This method is commonly used in patch-level autoregressive video models [20, 45, 48, 50] as well as in language models [28, 29], where the model is supervised through next token or patch prediction, effectively \"shifting\" one token or patch at a time during training.", + "bbox": [ + 511, + 674, + 906, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Can we achieve frame-level teacher forcing?", + "text_level": 1, + "bbox": [ + 511, + 801, + 885, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While teacher forcing is effective in autoregressive language models [28, 29], its application to frame-level video generation is largely unexplored. Few works adopt a straightforward \"shift one frame\" paradigm for frame-level video models.", + "bbox": [ + 511, + 825, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7375", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa7d1b37decb3ef7fdf4fdaa6f64c900dc1de1efdd31294131cbe3c3f4e03ff5.jpg", + "image_caption": [ + "Figure 1. Conceptual comparison of masked teacher forcing and our proposed complete teacher forcing mechanisms in autoregressive video generation. The blocks here illustrate individual frames. These two mechanisms differ in how history observation frames are used to condition next frame generation during training. (a): The observation frames are masked frames [2, 13] that are partially observable. (b): Every frame is generated conditioned on fully observable history frames." + ], + "image_footnote": [], + "bbox": [ + 196, + 94, + 794, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead, existing methods often rely on masked frame prediction, where masked frames are conditioned on other frames. For example, Genie [2] extends MaskGIT [4] to video generation by conditioning masked frames on other masked teacher frames—a strategy we refer to as Masked Teacher Forcing (MTF). Similarly, Diffusion Forcing [6] introduces temporal causal attention with a novel noise strategy but conditions each frame on noisy teacher frames, further diverging from traditional teacher forcing. GameNGen [40] employs bidirectional diffusion models with binary mask tokens to represent masked and conditional frames; however, it relies on fixed-length conditional frames, limiting autoregressive flexibility and preventing the use of efficient mechanisms like KV Cache—one notable advantage of traditional autoregressive methods. Generating videos autoregressively at the patch level is also suboptimal because inter-frames lack clear raster-scan spatial causality. In contrast, videos exhibit strong temporal causality, suggesting that predicting videos frame by frame, while generating the tokens within each frame in parallel, is more effective.", + "bbox": [ + 91, + 395, + 483, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these issues, MAGI fully leverages teacher forcing for frame-level autoregressive video generation, enabling more flexible, efficient, and scalable video generation.", + "bbox": [ + 89, + 712, + 482, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 89, + 784, + 199, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing frame-level autoregressive video generation models, such as GameNGen [40], typically use binary masks to distinguish between frames used for prediction and those used for conditioning. These models are trained to predict future frames conditioned on a fixed-length sequence of preceding frames. This fixed conditioning contrasts with", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the greater flexibility of autoregressive language models, which condition on prefixes of variable length. In this work, we leverage MAR [22], a state-of-the-art masked image generation framework, and investigate crucial design choices for masked autoregressive video generation. Our focus includes exploring different autoregressive paradigms and methods for mitigating exposure bias and error accumulation during autoregressive inference.", + "bbox": [ + 511, + 395, + 906, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked Teacher Forcing vs. Complete Teacher Forcing. A natural extension of masked image generation [4, 22] to autoregressive video prediction is to modify the attention mechanism [2], incorporating causal temporal attention during the training phase for masked frames. We refer to this approach as Masked Teacher Forcing (MTF), as illustrated on the right of Fig. 1. However, this method introduces a significant training-inference gap: during training, the model attends to a high mask ratio, while at inference time, it must predict future frames conditioned on previously generated unmasked ones. This mismatch between training and inference undermines the model's ability to generate realistic videos, ultimately limiting its performance.", + "bbox": [ + 511, + 534, + 906, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The formal formulation of MTF is as follows:", + "bbox": [ + 532, + 729, + 836, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(f _ {j} ^ {m} \\mid f _ {1} ^ {m}, f _ {2} ^ {m}, \\dots , f _ {j - 1} ^ {m}; \\theta\\right), \\quad j \\in \\{1, 2, \\dots , n \\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 753, + 906, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{j}^{m}$ represents the $j$ -th masked frame, while the frames $f_{1}^{m}, f_{2}^{m}, \\ldots, f_{j-1}^{m}$ are the previously predicted masked frames. In this formulation, during training, each masked frame attends to previously masked frames and itself. However, the approach introduces a significant gap between the training and the inference phase, because the model must predict masked frames conditioned on previously unmasked generated ones at the inference stage.", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7376", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cf088ff7a49e0fa9532de1f576a7494a69c24f4feb6ee2df4e814647ca11c50f.jpg", + "image_caption": [ + "Figure 2. Overview of MAGI video generation framework. MAGI receives observation frames and corresponding masked frames as inputs, enabling autoregressive video generation with Complete Teacher Forcing (Sec. 3)." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 477, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue and more effectively bridge the gap between training and inference, we propose a novel autoregressive paradigm, termed Complete Teacher Forcing (CTF), as shown on the left of Fig. 1. Unlike MTF, CTF conditions on unmasked observed frames during training, predicting each masked frame. Formally, let the observation frames be denoted as $\\{f_1,f_2,\\dots ,f_n\\}$ , where $n$ is the number of frames. The autoregressive formulation for predicting the $j$ -th masked frame, $f_{j}^{m}$ , is conditioned on both the previous unmasked observation frames and itself, i.e.,", + "bbox": [ + 89, + 338, + 483, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(f _ {j} ^ {m} \\mid f _ {1}, f _ {2}, \\dots , f _ {j - 1}; \\theta\\right), \\quad j \\in \\{1, 2, \\dots , n \\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 502, + 482, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p(f_{j}^{m}\\mid \\cdot ;\\theta)$ is the autoregressive model. Note that $f_{1}^{m}$ is conditioned only on itself, which resembles image generation. For all $j > 1$ , $f_{j}^{m}$ attends to both the previous unmasked observation frames and itself, making this approach consistent between training and inference.", + "bbox": [ + 89, + 531, + 483, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Addressing Exposure Bias & Error Accumulation Although CTF reduces the gap between training and inference, it still faces challenges related to exposure bias [55] and error accumulation [21]. Specifically, the model may struggle to generate realistic video sequences during inference when relying on self-predicted frames.", + "bbox": [ + 89, + 628, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate these issues, we introduce two strategies:", + "bbox": [ + 107, + 719, + 460, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Dynamic interval training. Video clips are sampled with random frame intervals, which forces the model to learn longer temporal dependencies and larger motion ranges, thus improving prediction stability for long videos. However, we find that a vanilla usage of this strategy leads to videos generated with uncontrollable motion range during inference. For example, prediction with a random motion range leads to unsatisfactory video generation results on UCF-101 that requires generation in 25 FPS. To support controllable generation to handle varying user-specified frame intervals, we introduce learnable interval", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39999c985198ed1caf1198e62569c2922f431c4e3d924a5c3548ea56d8c69e6b.jpg", + "image_caption": [ + "Figure 3. Temporal attention mask in our CTF during training. The attention within observation frames causal, while the attention within masked frames are atrous. In this fashion, each masked frame attends to itself and unmasked history observation frames. During inference, a standard causal mask is employed, where each frame attends only to previously generated frames." + ], + "image_footnote": [], + "bbox": [ + 596, + 92, + 834, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "embeddings which encodes different intervals into specific embeddings like positional encoding [42]. Specifically, the $x$ -th interval embedding will be added to hidden states when sampling with an interval $x$ which enables the model's awareness of the desired generation interval.", + "bbox": [ + 524, + 383, + 906, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Dynamic noise injection. As pointed out by Valevski et al. [40], there is a domain shift caused by teacher forcing and auto-regressive modeling, and thus data corruption using an injected noise during training is useful. Inspired by this practice, we also adopt a dynamic noise injection strategy. In addition, similar to interval embeddings, a learnable noise level embedding is concatenated with hidden states as the model inputs. Both techniques are crucial for improving the robustness of CTF by ensuring better generalization and stability during inference.", + "bbox": [ + 513, + 460, + 908, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Architecture", + "text_level": 1, + "bbox": [ + 511, + 627, + 643, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce MAGI, as shown in Fig. 2, a novel family of masked autoregressive video generation models, which incorporates the techniques outlined above. We implement MAGI with a Transformer architecture, detailed as follows.", + "bbox": [ + 511, + 652, + 906, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer Decoder & Temporal Attention MAGI employs a stack of spatial-temporal Transformer blocks consisting of interleaved 2D spatial attention layers and 1D temporal attention layers. As shown in Fig. 2, weptides the complete observation frames with masked frames as the input of our Transformer. For CTF, we design a special temporal attention mask, as shown in Fig. 3. In the temporal layer, each frame only attends to itself and its preceding observation frames. The observation frames only attend to previous frames. Thus, the attention paradigm of CTF is consistent between training and inference.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7377", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Head [22] Atop the Transformer decoder, we stack multi-layer perceptron (MLP) layers as the diffusion head, following MAR [22]. This component predicts masked tokens through a denoising diffusion procedure [16], enhancing the model's capacity for autoregressive generation.", + "bbox": [ + 89, + 90, + 480, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learnable Positional Embeddings To distinguish between masked and unmasked frames, we introduce two distinct learnable positional embeddings: one for the masked frames and another for the observation frames. Additionally, we adopt a learnable positional embedding for the diffusion head, as proposed in MAR [22].", + "bbox": [ + 89, + 181, + 480, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial and Temporal Positional Embeddings We utilize sinusoidal positional embeddings [42] for both 2D spatial and 1D temporal encodings, ensuring that the model effectively captures spatiotemporal relationships across frames.", + "bbox": [ + 89, + 272, + 480, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frame Interval Embeddings As part of our dynamic interval training strategy, we introduce a learnable interval embedding, with a vocabulary length of 25, which covers frame intervals ranging from 1 to 25. This embedding allows the model to capture long-range temporal dependencies across varying frame intervals.", + "bbox": [ + 89, + 333, + 480, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Noise Level Embeddings To support dynamic noise injection, we incorporate random Gaussian noise into the observation frames during training, following GameNGen [40]. We use a noise level with a range of [1, 5], which is encoded through a learnable vocabulary embedding, with a dimension the same as the Transformer's hidden dimension. This enables the model to adapt to different noise levels during the denoising process.", + "bbox": [ + 89, + 422, + 480, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 556, + 223, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 582, + 279, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets For the five-frame conditional video prediction tasks, we use the Kinetics-600 dataset [3], which consists of 480,000 videos spanning 600 action categories. For unconditional and first-frame conditional video generation $\\ddagger$ , we utilize the UCF-101 dataset [35], containing over 13,000 clips across 101 human action classes and with totally 27 hours of recording length.", + "bbox": [ + 89, + 603, + 480, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We set the learning rate to $2 \\times 10^{-4}$ for Kinetics-600 and $1 \\times 10^{-4}$ for UCF-101. Our largest model is trained for 150 epochs on Kinetics-600 and 1,400 epochs on UCF-101, with batch sizes of 256 and 128, respectively. Other hyperparameters follow the settings in MAR [22]. For Kinetics-600, we use the 3D-VAE from OmniTokenizer [45]. For UCF-101, we use the 2D-VAE from Stable Diffusion 1.4 [31] and the 3D-VAE from Cosmos [26]. During training, we sample 17 frames for the 3D-VAE and 16 frames for the 2D-VAE, respectively. All models are trained at a resolution of $256 \\times 256$ .", + "bbox": [ + 89, + 713, + 480, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference and Evaluation For inference, we follow the strategy outlined in MAR [22], using 64 iterative steps for masked frame prediction per frame. The inference procedure is significantly accelerated using KV Cache. For evaluation on Kinetics-600, we generate 50,000 videos using context frames randomly sampled from the test set and compute the Fréchet Video Distance (FVD) [39] against the ground-truth videos, both resized to $64 \\times 64$ resolution. For UCF-101, we follow the evaluation protocol of previous works [24, 45]. We randomly sample 2,048 videos and compute the FVD against 2,048 randomly selected ground-truth videos from the dataset.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Teacher Forcing Matters in Autoregressive Video Generation", + "text_level": 1, + "bbox": [ + 511, + 280, + 903, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CTF Outperforms MTF Our CTF significantly outperforms MTF by nearly $23\\%$ on first-frame conditioned video prediction. As shown in Fig. 5b, although MTF achieves slightly better frame-wise FID, it results in substantially worse overall FVD compared to CTF. Per-frame analysis (Fig. 4c-d) reveals that CTF better models temporal motion, while MTF generates high-quality static images lacking temporal coherence.", + "bbox": [ + 511, + 316, + 903, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Why Does CTF Achieve Superior FVD? We argue that although MTF's training on highly masked observations (using an optimal mask ratio of $70\\% - 100\\%$ ) is beneficial for generating visually similar frames and achieving lower FID scores, it imposes a significant limitation. Because MTF is trained to predict frames with minimal past information, it cannot effectively utilize full observations during inference. In contrast, CTF, trained on full observations, captures motion more accurately, leading to superior FVD scores.", + "bbox": [ + 511, + 455, + 905, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Addressing Exposure Bias", + "text_level": 1, + "bbox": [ + 511, + 599, + 751, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Strategies Matters An ablation study (Fig. 5c) demonstrates the importance of our proposed training strategies: Removing either or both dynamic interval training $^{\\S}$ and dynamic noise injection significantly degrades CTF's FVD and FID, underscoring the exposure bias in autoregressive prediction. Combining both yields the best performance, demonstrating their synergistic effect.", + "bbox": [ + 511, + 622, + 903, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Does MTF Also Benefit from These Training Strategies? These training strategies prove beneficial not only for CTF but also for MTF. While we focus our quantitative analysis on CTF and omit the full results for MTF due to space limitations, similar trends of performance improvement were observed. Notably, even when MTF benefits from these strategies, CTF consistently achieves superior results (see Fig. 5b), demonstrating its inherent advantages.", + "bbox": [ + 511, + 746, + 903, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{\\S}$ Further analysis, detailed in Appendix B.2, demonstrates how dynamic interval training mitigates exposure bias.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Class-conditional results are included in Appendix B.1.", + "bbox": [ + 107, + 886, + 408, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7378", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/694a6964ee89c28ae2311e21086a9e1f7e32d2c32b0bd5465f6769371fb53d4c.jpg", + "image_caption": [ + "Figure 4. Case Study of Proposed Training Techniques: This figure evaluates the impact of dynamic interval training and dynamic noise injection on CTF and MTF by: 1) visualizing CTF with and without these strategies; and 2) comparing CTF and MTF when both use them. All methods perform first-frame conditional video prediction on UCF-101 [35]. The results demonstrate the efficacy of the proposed training strategies and the superior motion and temporal coherence of CTF." + ], + "image_footnote": [], + "bbox": [ + 114, + 103, + 630, + 614 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d2b827bba5b5bfe63087066207ce06be0c5fa7c39e96fbb7069efd1d8342463f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 103, + 880, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Benchmarking with Previous Methods", + "text_level": 1, + "bbox": [ + 89, + 703, + 421, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video Prediction We evaluate our model trained on Kinetics-600 [3] against existing non-autoregressive (NAR) and autoregressive (AR) methods (Tab. 2). Our method, MAGI, achieves an FVD score of 11.5, establishing a new state-of-the-art among AR models and significantly outperforming the patch-level AR method Omni [45]. Notably, our MAGI achieves -21.4 significantly lower FVD than Omni whose FVD is 32.9. This result demonstrates the effectiveness of frame-level autoregressive modeling with Complete Teacher Forcing. Examples of generated videos are provided in Appendix Fig. 7.", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unconditional Video Generation For unconditional video generation on UCF-101 [35] (Tab. 3), our method achieves state-of-the-art results among AR models, outperforming Latte [24], a DiT-based NAR model, by approximately 50 FVD points using the same VAE. Furthermore, using a stronger VAE (Cosmos [26]), our method becomes competitive with state-of-the-art NAR methods.", + "bbox": [ + 511, + 703, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 818, + 679, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 outlines the core differences between MAGI and other methods. MAGI achieves a smooth transition from patch-level to frame-level autoregressive generation while retaining key advantages, analyzed below.", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7379", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da383fab327517f737b56b64a51191efd3aedce132363d216ec2e716e2ff7a36.jpg", + "image_caption": [ + "(a) Speed comparison." + ], + "image_footnote": [], + "bbox": [ + 94, + 97, + 339, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06b4ae7a80482e3156ec42db492237ad0a84b1425acd998ddcbc16b3f457e251.jpg", + "image_caption": [ + "(b) MTF v.s. CTF.", + "Figure 5. a) Speed of generation process with varying numbers of frames. MAGI achieves a significant speedup by utilizing KV Cache. b) Masked Teacher Forcing (MTF) v.s. Complete Teacher Forcing (CTF). Both methods utilize the proposed training strategies — dynamic interval training and dynamic noise injection. We report FID scores for individual frames and FVD scores for all frames on UCF101 [35]. CTF achieves significantly better FVD scores but slightly worse FID scores compared to MTF. This result demonstrates that CTF better captures motion with temporal coherence, even when the quality of individual frames is lower. c) The results of ablation study on first-frame conditional video predcition of UCF-101 [35]. The \"star\" indicates the FVD scores of each method with the same color." + ], + "image_footnote": [], + "bbox": [ + 372, + 97, + 630, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee43c7245cf011fbcb9022b6664a083e5e2b45f8623ddb4ef3386ccf24d7b57f.jpg", + "image_caption": [ + "(c) Ablation on first-frame conditional generation." + ], + "image_footnote": [], + "bbox": [ + 635, + 98, + 903, + 227 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3afd00392388f4f9e4826b7fb84851b5ff32f26bb3995cf4b56fac28e89c6868.jpg", + "table_caption": [ + "Table 2. Video Prediction on Kinetics-600 [3]. The results are evaluated on the testset on Kinetics-600. We report the $\\mathrm{FVD}_{64,50\\mathrm{K}}$ obtained from $50\\mathrm{K}$ samples in the resolution of $64\\times 64$ . NAR: Non-autoregressive methods. AR: Autoregressive methods." + ], + "table_footnote": [], + "table_body": "
TypeMethodVAEFVD64,50K ↓
NARVideo Diffusion [17]-16.2
NARRIN [18]-10.8
NARMAGVIT [51]MAGVIT [51]9.9
NARMAGVIT-v2 [52]MAGVIT-v2 [52]4.3
ARViVQVAE [44]VQVAE [41]64.3
ARPhenaki [43]VQVAE [41]36.4
AROmni [45]Omni [45]32.9
ARMAGIOmni [45]11.5
", + "bbox": [ + 94, + 422, + 480, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generation Order Matters Our experiments (Fig. 5b and 5c) demonstrate that single-frame FID increases as the number of AR steps grows, regardless of the training paradigms and strategies employed. This finding underscores the importance of generation order in autoregressive video generation. Since frame-level generation requires fewer AR steps than patch-level generation, it can better preserve image quality over long sequences by mitigating error accumulation. Moreover, fewer autoregressive steps reduce the propagation of errors that typically accumulate in sequential generation processes, leading to more coherent and stable video outputs. This efficiency not only enhances visual fidelity but also contributes to faster inference times, which are critical for practical applications.", + "bbox": [ + 88, + 588, + 482, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KV Cache A key practical advantage of MAGI is its efficient inference. MAGI's frame-level KV Cache enables approximately linear scaling of inference time with the number of generated frames (Fig. 5a), a significant advantage over the parallel computation required by NAR methods.", + "bbox": [ + 88, + 824, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bc074493cdebe2122195518364999b7f859d9a48f1dbada68e932ca05b589f70.jpg", + "table_caption": [ + "Table 3. Unconditional video generation on UCF-101 [35]. We report the $\\mathrm{FVD}_{256,2048}$ which is the FVD obtained from 2048 samples in the resolution of $256\\times 256$" + ], + "table_footnote": [], + "table_body": "
TypeMethodVAEFVD256, 2048 ↓
NARMoCoGAN [38]-2886.9
NARMoCoGAN-HD [37]-1729.6
NARDIGAN [53]-1630.2
NARStyleGAN-V [34]-1431.0
NARPVDM [54]PVDM [54]1141.9
NARMoStGAN-V [33]-1380.3
NARLatte [24]SD1.4 [31]477.9
NARDiM [25]SD1.4 [31]358.8
NARMatten [36]SD1.4 [31]210.6
ARVideoGPT [50]3D VQVAE [41]2880.6
ARMAGISD1.4 [31]420.6
ARMAGICosmos [26]297.8
", + "bbox": [ + 517, + 410, + 903, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Practical Advantages of Autoregressive Generation To explore this, we train MAGI on UCF-101 [35] using randomly sampled 16-frame clips and evaluate its long-term prediction capability in a first-frame conditional generation setting. Fig. 6 demonstrates the promising predictive performance of MAGI, achieving reasonable results for sequences up to 100 frames in specific scenarios, particularly those with relatively static backgrounds and simple object motions. We observe performance degradation for videos with non-periodic motions (e.g., diving), which we attribute, in part, to the simplicity of the UCF-101 dataset and the challenging extrapolation required by the first-frame conditional setting. Specifically, for non-periodic actions like diving, the model lacks cues to predict subsequent actions once the primary action is complete, rendering metrics like FVD inappropriate. Therefore, we focus on demonstrating MAGI's long-range prediction capability in these specific scenarios to showcase its potential for capturing long-range temporal dependencies and motivate further investigation.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7380", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a054ebd7b67da3fe5235072f683b4bf369a0574a3c62e702d0b0f9d0c01259f1.jpg", + "image_caption": [ + "Figure 6. Long-term Video Prediction. MAGI predicts over 100 frames from a single input frame, maintaining reasonable motion even when trained on only 16 frames." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Related Work", + "text_level": 1, + "bbox": [ + 89, + 351, + 230, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Autoregressive Visual Generation Autoregressive language modeling [28, 29] has propelled the development of visual content generation using discrete visual tokens [41]. In this framework, pre-trained visual tokenizers like VQVAE [41] map visual patches into a discrete latent space, allowing visual generation to be approached similarly to language modeling. Early works such as DALL-E [30] focus on text-to-image generation by learning a joint distribution between text and discrete image representations using an autoregressive cross-entropy loss. Concurrently, VideoGPT [50] extends this idea to video generation, employing discrete tokens for autoregressive video prediction. VideoPoet [20] further advances this approach by integrating a causal video tokenizer, MAGVIT-v2 [52], for multimodal video generation. OmniTokenizer [45] proposes a unified tokenizer for both discrete and continuous representations, enabling patch-level autoregressive video generation. In contrast, our work focuses on frame-level causality rather than patch-level, avoiding the limitations of raster-scan order and outperforming patch-level baselines.", + "bbox": [ + 88, + 377, + 482, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Masked and Diffusion Video Generation Diffusion models have recently gained prominence in visual generation tasks [7, 11, 14, 16, 31], extending effectively to video generation. Video diffusion models [1, 17] employ bidirectional attention and binary mask embeddings to enable frame-level autoregressive prediction. Notable works such as GameNGen [40] use bidirectional diffusion models for real-time game generation. However, due to their bidirectional nature, these models cannot leverage KV Cache for extended video generation, limiting their scalability. Several masked video generators, such as Genie [2], extend MaskGIT [4] into a causal-attention-based architecture for video generation. Despite their advantages, these", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "methods suffer from the training-inference gap inherent in masked autoregressive modeling, which negatively impacts generation quality. In contrast, our approach fully leverages KV Cache during inference, facilitated by our training paradigm that bridges the training-inference gap through a novel Complete Teacher Forcing paradigm.", + "bbox": [ + 511, + 352, + 905, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Addressing Exposure Bias for Autoregressive Video Generation Autoregressive models often suffer from exposure bias [55] and error accumulation [21], which degrade performance over long sequences. To mitigate these issues, we adopt noise injection from GameN-Gen [40]—where Gaussian noise is added to observation frame latents during training. Additionally, we introduce dynamic interval training, exposing the model to frames with varying temporal intervals to enhance generalization. As shown in Fig. 4, both strategies individually improve robustness, and their combination yields even greater performance gains in autoregressive video generation.", + "bbox": [ + 511, + 455, + 906, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present MAGI, a hybrid framework that combines masked and causal modeling to achieve efficient and scalable video generation. We identify that the teacher forcing paradigm makes a significant difference—our Complete Teacher Forcing (CTF) approach effectively bridges the training-inference gap inherent in Masked Teacher Forcing (MTF). Additionally, we introduce essential training strategies to alleviate exposure bias. Through comprehensive experiments, we validate the effectiveness of each component. Our final model achieves state-of-the-art performance among autoregressive video generation methods. We also demonstrate the generation of long, coherent video sequences exceeding 100 frames from training sequences as short as 16 frames, highlighting MAGI's potential as a scalable autoregressive video generation model.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7381", + "bbox": [ + 482, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to thank Tianhong Li for his insightful suggestions, invaluable help, and exceptional support. We would also like to thank Chenfei Wu, Haoyang Huang, Guoqing Ma, Hongyu Zhou, Liangyu Chen, Chunrui Han, Yimin Jiang and Yu Deng for their constructive discussions and advice, which greatly improved this work.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 218, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 8", + "[2] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2, 3, 8", + "[3] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808.01340, 2018. 5, 6, 7", + "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022, 1, 3, 8", + "[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1", + "[6] Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. arXiv preprint arXiv:2407.01392, 2024. 1, 2, 3", + "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 8", + "[8] Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 1", + "[9] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive KV cache compression for llms. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1", + "[10] Anirudh Goyal, Alex Lamb, Ying Zhang, Saizheng Zhang, Aaron C. Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In" + ], + "bbox": [ + 93, + 243, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 4601-4609, 2016. 2", + "bbox": [ + 544, + 92, + 906, + 146 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 8", + "bbox": [ + 516, + 148, + 905, + 215 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[12] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 13", + "bbox": [ + 517, + 217, + 905, + 284 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 3", + "[14] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2(3):4, 2022. 8", + "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 12", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5, 8, 13", + "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 7, 8", + "[18] Allan Jabri, David Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. arXiv preprint arXiv:2212.11972, 2022. 7", + "[19] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 13", + "[20] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 1, 2, 8, 12", + "[21] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Conference on robot learning, pages 143-156. PMLR, 2017. 4, 8", + "[22] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 5, 12, 13", + "[23] Shi Luohe, Hongyi Zhang, Yao Yao, Zuchao Li, and hai zhao. Keep the cost down: A review on methods to optimize LLM's KV-cache consumption. In First Conference on Language Modeling, 2024. 1, 2" + ], + "bbox": [ + 93, + 243, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7382", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 6, 7", + "[25] Shentong Mo and Yapeng Tian. Scaling diffusion mamba with bidirectional ssms for efficient image and video generation. arXiv preprint arXiv:2405.15881, 2024. 7", + "[26] NVIDIA. Cosmos tokenizer. https://github.com/NVIDIA/Cosmos-Tokenizer, 2024. Accessed: 2024-11-15. 5, 6, 7, 12", + "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 13", + "[28] Alec Radford. Improving language understanding by generative pre-training. 2018. 2, 8", + "[29] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 8", + "[30] A. Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 8", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5, 7, 8", + "[32] Florian Schmidt. Generalization in generation: A closer look at exposure bias. In Proceedings of the 3rd Workshop on Neural Generation and Translation@EMNLP-IJCNLP 2019, Hong Kong, November 4, 2019, pages 157-167. Association for Computational Linguistics, 2019. 2", + "[33] Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Mostgan-v: Video generation with temporal motion styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5652-5661, 2023. 7", + "[34] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3626-3636, 2022. 7", + "[35] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5, 6, 7", + "[36] Yao Teng, Yue Wu, Han Shi, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024. 7", + "[37] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021. 7", + "[38] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 7", + "[39] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5", + "[40] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. arXiv preprint arXiv:2408.14837, 2024. 2, 3, 4, 5, 8", + "[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7, 8", + "[42] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4, 5", + "[43] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022. 2, 7", + "[44] Jacob Walker, Ali Razavi, and Aäron van den Oord. Predicting video with vqvae. arXiv preprint arXiv:2103.01950, 2021. 7", + "[45] Junke Wang, Yi Jiang, Zehuan Yuan, Binyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. arXiv preprint arXiv:2406.09399, 2024. 2, 5, 6, 7, 8, 12", + "[46] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2", + "[47] Ronald J. Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural Computation, 1(2):270-280, 1989. 1, 2", + "[48] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In European conference on computer vision, pages 720-736. Springer, 2022. 1, 2", + "[49] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution text-to-image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025. 13", + "[50] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 1, 2, 7, 8", + "[51] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 1, 2, 7, 12", + "[52] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7383", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 7, 8, 12", + "[53] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 7", + "[54] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 7", + "[55] Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. Bridging the gap between training and inference for neural machine translation. arXiv preprint arXiv:1906.02448, 2019. 4, 8" + ], + "bbox": [ + 91, + 90, + 483, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7384", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/e2161c88-f390-4144-acf4-52130bcde3bd_model.json b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/e2161c88-f390-4144-acf4-52130bcde3bd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..41494118ec14a40d2f6b8e50522bfc79b82e6da2 --- /dev/null +++ b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/e2161c88-f390-4144-acf4-52130bcde3bd_model.json @@ -0,0 +1,2048 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.853, + 0.153 + ], + "angle": 0, + "content": "Taming Teacher Forcing for Masked Autoregressive Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.179, + 0.854, + 0.219 + ], + "angle": 0, + "content": "Deyu Zhou1 Quan Sun2 Yuang Peng2,4 Kun Yan2 Runpei Dong3 Duomin Wang2 Zheng Ge2 Nan Duan2 Xiangyu Zhang2" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.222, + 0.712, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)HKUST(GZ) \\(^{2}\\)StepFun \\(^{3}\\)UIUC \\(^{4}\\)THU MAGI-VIDEO-GENERATION.GITHUB.IO" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.294, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.569 + ], + "angle": 0, + "content": "We introduce MAGI, a hybrid video generation framework that combines masked modeling for intra-frame generation with causal modeling for next-frame generation. Our key innovation, Complete Teacher Forcing (CTF), conditions masked frames on complete observation frames rather than masked ones (namely Masked Teacher Forcing, MTF), enabling a smooth transition from token-level (patch-level) to frame-level autoregressive generation. CTF significantly outperforms MTF, achieving a \\(+23\\%\\) improvement in FVD scores on first-frame conditioned video prediction. To address issues like exposure bias, we employ targeted training strategies, setting a new benchmark in autoregressive video generation. Experiments show that MAGI can generate long, coherent video sequences exceeding 100 frames, even when trained on as few as 16 frames, highlighting its potential for scalable, high-quality video generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.598, + 0.222, + 0.613 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Generating order matters in autoregressive image generation. While existing approaches mostly reply on simple raster-scan order, recent studies [8, 22] have demonstrated superior results through alternative generation strategies. These include randomized spatial order with bidirectional attention [4, 22, 51] and generation along increased scales [5, 8]. Despite its fundamental importance, in the realm of autoregressive video generation, however, the discussion of generation order has been largely overlooked." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Existing autoregressive video generation methods can be categorized into two groups according to the prediction granularity, as shown in Tab. 1. The first group, exemplified by MAGViT [51], adopts a masked modeling approach* where visual tokens are generated in descending order according to their logit probabilities. While this approach uses bi-directional attention across frames, it presents two signif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.905, + 0.446 + ], + "angle": 0, + "content": "icant limitations: substantial computational overhead during inference since it is unable to utilize KV Cache [9, 23], and neglect temporal causality between consecutive frames. These oversights are particularly noteworthy given the remarkable success of causal modeling in large language models. Some other methods in this group [2, 6] have explored frame-level prediction with casual temporal attention. However, these methods are limited due to incomplete history observation, which causes an intrinsic training-inference gap." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.569 + ], + "angle": 0, + "content": "The second group comprises fully autoregressive approaches operating on visual patches [20, 46, 48, 50], exemplified by recent works like VideoPoet [20] and Emu3 [46]. Despite their successful temporal modeling, these methods have not incorporated contemporary advances from image generation research in intra-frame generation. Instead, they continue to rely on raster-scan order - an approach demonstrated sub-optimal in image generation [22]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce Masked Autoregressive video Generation (MAGI), a hybrid framework that synergizes the strengths of both video generation paradigms by combining masked modeling for intra-frame generation with causal modeling for inter-frame dependencies. We start with a naive implementation. This initial implementation explores a straightforward approach where tokens in each next frame are generated conditioned on previous masked frames (Figure 1(a)). While this can be viewed as a natural extension of MaskGIT [4] and Muse [5] to autoregressive video generation, our analysis reveals a fundamental limitation in its implementation of teacher forcing (TF) [47]. To be specific, TF involves feeding the model the groundtruth (correct) output from the previous time step as input during training, rather than using the model's own predicted output. However, our initial approach, which we term as Masked Teacher Forcing (MTF), deviates from this principle by substituting ground-truth tokens with mask tokens. Despite similar approaches being explored in interactive generative models (e.g., Genie [2]), we identify a critical training-inference discrepancy: during training, next-frame generation is conditioned on a mixture of visible" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*We follow the definition in MAR [22], which considers masked image generation as a form of auto-regressive generation with a certain order." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7374" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Table 1. Comparison of autoregressive paradigms across different methods. We compare MAGI with various methods based on temporal attention mechanisms (how each frame attends to others), support for KV Cache [23] and variable context lengths, observation completeness (the completeness of the context frames used for conditioning), and the prediction granularity (patch or frame). \\(\\dagger\\): MAGVIT predicts masked tokens of multiple frames simultaneously. Genie [2] and MAGI predict multiple tokens of single frame simultaneously." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.156, + 0.905, + 0.259 + ], + "angle": 0, + "content": "
Prediction GranularityMethodTemporal AttentionKV CacheVar. ContextObservation
PatchVideoGPT [50], Phenaki [43], Omni [45]Causal✓✓Complete
FrameMAGVIT† [51], GameNGen [40]Bidirectional✗✗Complete
Diffusion Forcing [6]Causal✓✓Noisy
Genie [2]Causal✓✓Masked
MAGI (with CTF)Causal✓✓Complete
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.336 + ], + "angle": 0, + "content": "and masked tokens, while during inference, it relies solely on complete visual tokens from previous frames. Our experiments demonstrate that this inconsistency compromises generation quality, especially motion coherency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.336, + 0.483, + 0.488 + ], + "angle": 0, + "content": "To address these limitations, we propose Complete Teacher Forcing (CTF), an advanced paradigm that maintains training-test consistency by conditioning next-frame generation on complete visual observations from history frames during training. This is achieved through a novel mechanism that prepends complete video tokens to the masked input sequence and employs a carefully designed attention mask. Empirical results show that CTF better captures motion, outperforming MTF by \\(23\\%\\) in terms of FVD scores on first-frame conditioned video prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.483, + 0.777 + ], + "angle": 0, + "content": "In addition, we tackle the persistent challenge of exposure bias [32] in autoregressive video generation through two complementary strategies: dynamic interval training and noise injection [40]. Dynamic interval training refers to randomly sampling frames with varying intervals during training, which introduces diversity into the data distribution and helps the model to better generalize to different temporal frequencies. Dynamic noise injection, on the other hand, involves adding random noise to observation frames during training, which helps improve the model's robustness by simulating the possible errors may occur in inference. Combined with these two techniques, MAGI establishes a robust baseline for autoregressive video generation. Through extensive evaluations, we demonstrate that MAGI achieves superior video generation quality and length scalability. For example, MAGI can generate coherent video sequences that exceed 100 frames, even when trained on sequences as short as 16 frames. This underscores the potential of our approach for scalable video generation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.792, + 0.422, + 0.809 + ], + "angle": 0, + "content": "2. Background and Problem Statement" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.382, + 0.833 + ], + "angle": 0, + "content": "2.1. Autoregressive Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.839, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Given a video \\(V \\in \\mathbb{R}^{N \\times H \\times W \\times 3}\\), the sequence of \\(N\\) frames with height \\(H\\) and width \\(W\\) can be denoted as \\(V = \\{f_i \\in \\mathbb{R}^{H \\times W \\times 3} | i = 1, \\dots, N\\}\\). Each frame can be further divided into \\(m\\) patches, represented as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.905, + 0.335 + ], + "angle": 0, + "content": "\\(\\{\\mathbf{x}_{11}, \\mathbf{x}_{12}, \\ldots, \\mathbf{x}_{1m}, \\ldots, \\mathbf{x}_{n1}, \\mathbf{x}_{n2}, \\ldots, \\mathbf{x}_{nm}\\}\\). Autoregressive video generation models primarily differ in the granularity of causality (patch-level vs. frame-level) and the training paradigm (teacher forcing), discussed below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.905, + 0.412 + ], + "angle": 0, + "content": "Patch-level Methods Patch-level autoregressive models [46, 50] operate by modeling the video on a patch-by-patch basis. The likelihood of generating entire video with a model parameterized by \\(\\theta\\) is decomposed into the product of posterior of each patch conditioned all preceding patches:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.419, + 0.905, + 0.461 + ], + "angle": 0, + "content": "\\[\np (V) = \\prod_ {t = 1} ^ {N \\times m} p \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {1}, \\mathbf {x} _ {2}, \\dots , \\mathbf {x} _ {t - 1}; \\theta\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.528 + ], + "angle": 0, + "content": "where \\( p_t \\) represents a patch in a predefined ordering. The model generates each patch sequentially by conditioning on all previously generated patches, capturing both spatial and temporal dependencies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.905, + 0.575 + ], + "angle": 0, + "content": "Frame-level Methods Frame-level autoregressive models [2, 40] model the video at the granularity of entire frames. Similarly, the likelihood of video modeling is:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.582, + 0.905, + 0.621 + ], + "angle": 0, + "content": "\\[\np (V) = \\prod_ {t = 1} ^ {n} p \\left(f _ {t} \\mid f _ {1}, f _ {2}, \\dots , f _ {t - 1}; \\theta\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.674 + ], + "angle": 0, + "content": "where \\( f_{t} \\) is the entire frame generated conditioned on all history frames. In this fashion, the patches in each frame are generated in parallel." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.796 + ], + "angle": 0, + "content": "Teacher Forcing Teacher forcing [47] is a widely adopted training strategy in autoregressive models, where the model conditions on the true previous frame or patch rather than its own predictions [10, 47]. This method is commonly used in patch-level autoregressive video models [20, 45, 48, 50] as well as in language models [28, 29], where the model is supervised through next token or patch prediction, effectively \"shifting\" one token or patch at a time during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.887, + 0.82 + ], + "angle": 0, + "content": "2.2. Can we achieve frame-level teacher forcing?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "While teacher forcing is effective in autoregressive language models [28, 29], its application to frame-level video generation is largely unexplored. Few works adopt a straightforward \"shift one frame\" paradigm for frame-level video models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7375" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.197, + 0.095, + 0.795, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.313, + 0.907, + 0.37 + ], + "angle": 0, + "content": "Figure 1. Conceptual comparison of masked teacher forcing and our proposed complete teacher forcing mechanisms in autoregressive video generation. The blocks here illustrate individual frames. These two mechanisms differ in how history observation frames are used to condition next frame generation during training. (a): The observation frames are masked frames [2, 13] that are partially observable. (b): Every frame is generated conditioned on fully observable history frames." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.396, + 0.485, + 0.711 + ], + "angle": 0, + "content": "Instead, existing methods often rely on masked frame prediction, where masked frames are conditioned on other frames. For example, Genie [2] extends MaskGIT [4] to video generation by conditioning masked frames on other masked teacher frames—a strategy we refer to as Masked Teacher Forcing (MTF). Similarly, Diffusion Forcing [6] introduces temporal causal attention with a novel noise strategy but conditions each frame on noisy teacher frames, further diverging from traditional teacher forcing. GameNGen [40] employs bidirectional diffusion models with binary mask tokens to represent masked and conditional frames; however, it relies on fixed-length conditional frames, limiting autoregressive flexibility and preventing the use of efficient mechanisms like KV Cache—one notable advantage of traditional autoregressive methods. Generating videos autoregressively at the patch level is also suboptimal because inter-frames lack clear raster-scan spatial causality. In contrast, videos exhibit strong temporal causality, suggesting that predicting videos frame by frame, while generating the tokens within each frame in parallel, is more effective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.772 + ], + "angle": 0, + "content": "To address these issues, MAGI fully leverages teacher forcing for frame-level autoregressive video generation, enabling more flexible, efficient, and scalable video generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.2, + 0.803 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Existing frame-level autoregressive video generation models, such as GameNGen [40], typically use binary masks to distinguish between frames used for prediction and those used for conditioning. These models are trained to predict future frames conditioned on a fixed-length sequence of preceding frames. This fixed conditioning contrasts with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.517 + ], + "angle": 0, + "content": "the greater flexibility of autoregressive language models, which condition on prefixes of variable length. In this work, we leverage MAR [22], a state-of-the-art masked image generation framework, and investigate crucial design choices for masked autoregressive video generation. Our focus includes exploring different autoregressive paradigms and methods for mitigating exposure bias and error accumulation during autoregressive inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.731 + ], + "angle": 0, + "content": "Masked Teacher Forcing vs. Complete Teacher Forcing. A natural extension of masked image generation [4, 22] to autoregressive video prediction is to modify the attention mechanism [2], incorporating causal temporal attention during the training phase for masked frames. We refer to this approach as Masked Teacher Forcing (MTF), as illustrated on the right of Fig. 1. However, this method introduces a significant training-inference gap: during training, the model attends to a high mask ratio, while at inference time, it must predict future frames conditioned on previously generated unmasked ones. This mismatch between training and inference undermines the model's ability to generate realistic videos, ultimately limiting its performance." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.731, + 0.838, + 0.745 + ], + "angle": 0, + "content": "The formal formulation of MTF is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.755, + 0.907, + 0.773 + ], + "angle": 0, + "content": "\\[\np \\left(f _ {j} ^ {m} \\mid f _ {1} ^ {m}, f _ {2} ^ {m}, \\dots , f _ {j - 1} ^ {m}; \\theta\\right), \\quad j \\in \\{1, 2, \\dots , n \\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\( f_{j}^{m} \\) represents the \\( j \\)-th masked frame, while the frames \\( f_{1}^{m}, f_{2}^{m}, \\ldots, f_{j-1}^{m} \\) are the previously predicted masked frames. In this formulation, during training, each masked frame attends to previously masked frames and itself. However, the approach introduces a significant gap between the training and the inference phase, because the model must predict masked frames conditioned on previously unmasked generated ones at the inference stage." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7376" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.478, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.32 + ], + "angle": 0, + "content": "Figure 2. Overview of MAGI video generation framework. MAGI receives observation frames and corresponding masked frames as inputs, enabling autoregressive video generation with Complete Teacher Forcing (Sec. 3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.491 + ], + "angle": 0, + "content": "To address this issue and more effectively bridge the gap between training and inference, we propose a novel autoregressive paradigm, termed Complete Teacher Forcing (CTF), as shown on the left of Fig. 1. Unlike MTF, CTF conditions on unmasked observed frames during training, predicting each masked frame. Formally, let the observation frames be denoted as \\(\\{f_1,f_2,\\dots ,f_n\\}\\) , where \\(n\\) is the number of frames. The autoregressive formulation for predicting the \\(j\\) -th masked frame, \\(f_{j}^{m}\\) , is conditioned on both the previous unmasked observation frames and itself, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.503, + 0.483, + 0.521 + ], + "angle": 0, + "content": "\\[\np \\left(f _ {j} ^ {m} \\mid f _ {1}, f _ {2}, \\dots , f _ {j - 1}; \\theta\\right), \\quad j \\in \\{1, 2, \\dots , n \\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.609 + ], + "angle": 0, + "content": "where \\( p(f_{j}^{m}\\mid \\cdot ;\\theta) \\) is the autoregressive model. Note that \\( f_{1}^{m} \\) is conditioned only on itself, which resembles image generation. For all \\( j > 1 \\), \\( f_{j}^{m} \\) attends to both the previous unmasked observation frames and itself, making this approach consistent between training and inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Addressing Exposure Bias & Error Accumulation Although CTF reduces the gap between training and inference, it still faces challenges related to exposure bias [55] and error accumulation [21]. Specifically, the model may struggle to generate realistic video sequences during inference when relying on self-predicted frames." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.72, + 0.462, + 0.735 + ], + "angle": 0, + "content": "To mitigate these issues, we introduce two strategies:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- Dynamic interval training. Video clips are sampled with random frame intervals, which forces the model to learn longer temporal dependencies and larger motion ranges, thus improving prediction stability for long videos. However, we find that a vanilla usage of this strategy leads to videos generated with uncontrollable motion range during inference. For example, prediction with a random motion range leads to unsatisfactory video generation results on UCF-101 that requires generation in 25 FPS. To support controllable generation to handle varying user-specified frame intervals, we introduce learnable interval" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.093, + 0.835, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.287, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Figure 3. Temporal attention mask in our CTF during training. The attention within observation frames causal, while the attention within masked frames are atrous. In this fashion, each masked frame attends to itself and unmasked history observation frames. During inference, a standard causal mask is employed, where each frame attends only to previously generated frames." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.385, + 0.907, + 0.461 + ], + "angle": 0, + "content": "embeddings which encodes different intervals into specific embeddings like positional encoding [42]. Specifically, the \\( x \\)-th interval embedding will be added to hidden states when sampling with an interval \\( x \\) which enables the model's awareness of the desired generation interval." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.461, + 0.909, + 0.613 + ], + "angle": 0, + "content": "- Dynamic noise injection. As pointed out by Valevski et al. [40], there is a domain shift caused by teacher forcing and auto-regressive modeling, and thus data corruption using an injected noise during training is useful. Inspired by this practice, we also adopt a dynamic noise injection strategy. In addition, similar to interval embeddings, a learnable noise level embedding is concatenated with hidden states as the model inputs. Both techniques are crucial for improving the robustness of CTF by ensuring better generalization and stability during inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.645, + 0.643 + ], + "angle": 0, + "content": "4. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.907, + 0.713 + ], + "angle": 0, + "content": "We introduce MAGI, as shown in Fig. 2, a novel family of masked autoregressive video generation models, which incorporates the techniques outlined above. We implement MAGI with a Transformer architecture, detailed as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Transformer Decoder & Temporal Attention MAGI employs a stack of spatial-temporal Transformer blocks consisting of interleaved 2D spatial attention layers and 1D temporal attention layers. As shown in Fig. 2, weptides the complete observation frames with masked frames as the input of our Transformer. For CTF, we design a special temporal attention mask, as shown in Fig. 3. In the temporal layer, each frame only attends to itself and its preceding observation frames. The observation frames only attend to previous frames. Thus, the attention paradigm of CTF is consistent between training and inference." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7377" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "Diffusion Head [22] Atop the Transformer decoder, we stack multi-layer perceptron (MLP) layers as the diffusion head, following MAR [22]. This component predicts masked tokens through a denoising diffusion procedure [16], enhancing the model's capacity for autoregressive generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.182, + 0.482, + 0.272 + ], + "angle": 0, + "content": "Learnable Positional Embeddings To distinguish between masked and unmasked frames, we introduce two distinct learnable positional embeddings: one for the masked frames and another for the observation frames. Additionally, we adopt a learnable positional embedding for the diffusion head, as proposed in MAR [22]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.332 + ], + "angle": 0, + "content": "Spatial and Temporal Positional Embeddings We utilize sinusoidal positional embeddings [42] for both 2D spatial and 1D temporal encodings, ensuring that the model effectively captures spatiotemporal relationships across frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.482, + 0.422 + ], + "angle": 0, + "content": "Frame Interval Embeddings As part of our dynamic interval training strategy, we introduce a learnable interval embedding, with a vocabulary length of 25, which covers frame intervals ranging from 1 to 25. This embedding allows the model to capture long-range temporal dependencies across varying frame intervals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.482, + 0.545 + ], + "angle": 0, + "content": "Noise Level Embeddings To support dynamic noise injection, we incorporate random Gaussian noise into the observation frames during training, following GameNGen [40]. We use a noise level with a range of [1, 5], which is encoded through a learnable vocabulary embedding, with a dimension the same as the Transformer's hidden dimension. This enables the model to adapt to different noise levels during the denoising process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.224, + 0.575 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.281, + 0.599 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.482, + 0.71 + ], + "angle": 0, + "content": "Datasets For the five-frame conditional video prediction tasks, we use the Kinetics-600 dataset [3], which consists of 480,000 videos spanning 600 action categories. For unconditional and first-frame conditional video generation\\(\\ddagger\\), we utilize the UCF-101 dataset [35], containing over 13,000 clips across 101 human action classes and with totally 27 hours of recording length." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.482, + 0.879 + ], + "angle": 0, + "content": "Implementation Details We set the learning rate to \\(2 \\times 10^{-4}\\) for Kinetics-600 and \\(1 \\times 10^{-4}\\) for UCF-101. Our largest model is trained for 150 epochs on Kinetics-600 and 1,400 epochs on UCF-101, with batch sizes of 256 and 128, respectively. Other hyperparameters follow the settings in MAR [22]. For Kinetics-600, we use the 3D-VAE from OmniTokenizer [45]. For UCF-101, we use the 2D-VAE from Stable Diffusion 1.4 [31] and the 3D-VAE from Cosmos [26]. During training, we sample 17 frames for the 3D-VAE and 16 frames for the 2D-VAE, respectively. All models are trained at a resolution of \\(256 \\times 256\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Inference and Evaluation For inference, we follow the strategy outlined in MAR [22], using 64 iterative steps for masked frame prediction per frame. The inference procedure is significantly accelerated using KV Cache. For evaluation on Kinetics-600, we generate 50,000 videos using context frames randomly sampled from the test set and compute the Fréchet Video Distance (FVD) [39] against the ground-truth videos, both resized to \\(64 \\times 64\\) resolution. For UCF-101, we follow the evaluation protocol of previous works [24, 45]. We randomly sample 2,048 videos and compute the FVD against 2,048 randomly selected ground-truth videos from the dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.281, + 0.905, + 0.311 + ], + "angle": 0, + "content": "5.2. Teacher Forcing Matters in Autoregressive Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.439 + ], + "angle": 0, + "content": "CTF Outperforms MTF Our CTF significantly outperforms MTF by nearly \\(23\\%\\) on first-frame conditioned video prediction. As shown in Fig. 5b, although MTF achieves slightly better frame-wise FID, it results in substantially worse overall FVD compared to CTF. Per-frame analysis (Fig. 4c-d) reveals that CTF better models temporal motion, while MTF generates high-quality static images lacking temporal coherence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.593 + ], + "angle": 0, + "content": "Why Does CTF Achieve Superior FVD? We argue that although MTF's training on highly masked observations (using an optimal mask ratio of \\(70\\% - 100\\%\\)) is beneficial for generating visually similar frames and achieving lower FID scores, it imposes a significant limitation. Because MTF is trained to predict frames with minimal past information, it cannot effectively utilize full observations during inference. In contrast, CTF, trained on full observations, captures motion more accurately, leading to superior FVD scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.753, + 0.618 + ], + "angle": 0, + "content": "5.3. Addressing Exposure Bias" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.905, + 0.73 + ], + "angle": 0, + "content": "Training Strategies Matters An ablation study (Fig. 5c) demonstrates the importance of our proposed training strategies: Removing either or both dynamic interval training\\(^{\\S}\\) and dynamic noise injection significantly degrades CTF's FVD and FID, underscoring the exposure bias in autoregressive prediction. Combining both yields the best performance, demonstrating their synergistic effect." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.905, + 0.868 + ], + "angle": 0, + "content": "Does MTF Also Benefit from These Training Strategies? These training strategies prove beneficial not only for CTF but also for MTF. While we focus our quantitative analysis on CTF and omit the full results for MTF due to space limitations, similar trends of performance improvement were observed. Notably, even when MTF benefits from these strategies, CTF consistently achieves superior results (see Fig. 5b), demonstrating its inherent advantages." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\(^{\\S}\\)Further analysis, detailed in Appendix B.2, demonstrates how dynamic interval training mitigates exposure bias." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.409, + 0.901 + ], + "angle": 0, + "content": "Class-conditional results are included in Appendix B.1." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7378" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.104, + 0.631, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.104, + 0.882, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.624, + 0.907, + 0.681 + ], + "angle": 0, + "content": "Figure 4. Case Study of Proposed Training Techniques: This figure evaluates the impact of dynamic interval training and dynamic noise injection on CTF and MTF by: 1) visualizing CTF with and without these strategies; and 2) comparing CTF and MTF when both use them. All methods perform first-frame conditional video prediction on UCF-101 [35]. The results demonstrate the efficacy of the proposed training strategies and the superior motion and temporal coherence of CTF." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.704, + 0.423, + 0.72 + ], + "angle": 0, + "content": "5.4. Benchmarking with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Video Prediction We evaluate our model trained on Kinetics-600 [3] against existing non-autoregressive (NAR) and autoregressive (AR) methods (Tab. 2). Our method, MAGI, achieves an FVD score of 11.5, establishing a new state-of-the-art among AR models and significantly outperforming the patch-level AR method Omni [45]. Notably, our MAGI achieves -21.4 significantly lower FVD than Omni whose FVD is 32.9. This result demonstrates the effectiveness of frame-level autoregressive modeling with Complete Teacher Forcing. Examples of generated videos are provided in Appendix Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Unconditional Video Generation For unconditional video generation on UCF-101 [35] (Tab. 3), our method achieves state-of-the-art results among AR models, outperforming Latte [24], a DiT-based NAR model, by approximately 50 FVD points using the same VAE. Furthermore, using a stronger VAE (Cosmos [26]), our method becomes competitive with state-of-the-art NAR methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.681, + 0.835 + ], + "angle": 0, + "content": "5.5. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Tab. 1 outlines the core differences between MAGI and other methods. MAGI achieves a smooth transition from patch-level to frame-level autoregressive generation while retaining key advantages, analyzed below." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7379" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.098, + 0.341, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.231, + 0.288, + 0.244 + ], + "angle": 0, + "content": "(a) Speed comparison." + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.098, + 0.631, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.23, + 0.551, + 0.242 + ], + "angle": 0, + "content": "(b) MTF v.s. CTF." + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.099, + 0.904, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.231, + 0.901, + 0.243 + ], + "angle": 0, + "content": "(c) Ablation on first-frame conditional generation." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.255, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Figure 5. a) Speed of generation process with varying numbers of frames. MAGI achieves a significant speedup by utilizing KV Cache. b) Masked Teacher Forcing (MTF) v.s. Complete Teacher Forcing (CTF). Both methods utilize the proposed training strategies — dynamic interval training and dynamic noise injection. We report FID scores for individual frames and FVD scores for all frames on UCF101 [35]. CTF achieves significantly better FVD scores but slightly worse FID scores compared to MTF. This result demonstrates that CTF better captures motion with temporal coherence, even when the quality of individual frames is lower. c) The results of ablation study on first-frame conditional video predcition of UCF-101 [35]. The \"star\" indicates the FVD scores of each method with the same color." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.363, + 0.483, + 0.42 + ], + "angle": 0, + "content": "Table 2. Video Prediction on Kinetics-600 [3]. The results are evaluated on the testset on Kinetics-600. We report the \\(\\mathrm{FVD}_{64,50\\mathrm{K}}\\) obtained from \\(50\\mathrm{K}\\) samples in the resolution of \\(64\\times 64\\). NAR: Non-autoregressive methods. AR: Autoregressive methods." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.424, + 0.481, + 0.564 + ], + "angle": 0, + "content": "
TypeMethodVAEFVD64,50K ↓
NARVideo Diffusion [17]-16.2
NARRIN [18]-10.8
NARMAGVIT [51]MAGVIT [51]9.9
NARMAGVIT-v2 [52]MAGVIT-v2 [52]4.3
ARViVQVAE [44]VQVAE [41]64.3
ARPhenaki [43]VQVAE [41]36.4
AROmni [45]Omni [45]32.9
ARMAGIOmni [45]11.5
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.589, + 0.483, + 0.801 + ], + "angle": 0, + "content": "Generation Order Matters Our experiments (Fig. 5b and 5c) demonstrate that single-frame FID increases as the number of AR steps grows, regardless of the training paradigms and strategies employed. This finding underscores the importance of generation order in autoregressive video generation. Since frame-level generation requires fewer AR steps than patch-level generation, it can better preserve image quality over long sequences by mitigating error accumulation. Moreover, fewer autoregressive steps reduce the propagation of errors that typically accumulate in sequential generation processes, leading to more coherent and stable video outputs. This efficiency not only enhances visual fidelity but also contributes to faster inference times, which are critical for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "KV Cache A key practical advantage of MAGI is its efficient inference. MAGI's frame-level KV Cache enables approximately linear scaling of inference time with the number of generated frames (Fig. 5a), a significant advantage over the parallel computation required by NAR methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.363, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Table 3. Unconditional video generation on UCF-101 [35]. We report the \\(\\mathrm{FVD}_{256,2048}\\) which is the FVD obtained from 2048 samples in the resolution of \\(256\\times 256\\)" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.411, + 0.905, + 0.591 + ], + "angle": 0, + "content": "
TypeMethodVAEFVD256, 2048 ↓
NARMoCoGAN [38]-2886.9
NARMoCoGAN-HD [37]-1729.6
NARDIGAN [53]-1630.2
NARStyleGAN-V [34]-1431.0
NARPVDM [54]PVDM [54]1141.9
NARMoStGAN-V [33]-1380.3
NARLatte [24]SD1.4 [31]477.9
NARDiM [25]SD1.4 [31]358.8
NARMatten [36]SD1.4 [31]210.6
ARVideoGPT [50]3D VQVAE [41]2880.6
ARMAGISD1.4 [31]420.6
ARMAGICosmos [26]297.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Practical Advantages of Autoregressive Generation To explore this, we train MAGI on UCF-101 [35] using randomly sampled 16-frame clips and evaluate its long-term prediction capability in a first-frame conditional generation setting. Fig. 6 demonstrates the promising predictive performance of MAGI, achieving reasonable results for sequences up to 100 frames in specific scenarios, particularly those with relatively static backgrounds and simple object motions. We observe performance degradation for videos with non-periodic motions (e.g., diving), which we attribute, in part, to the simplicity of the UCF-101 dataset and the challenging extrapolation required by the first-frame conditional setting. Specifically, for non-periodic actions like diving, the model lacks cues to predict subsequent actions once the primary action is complete, rendering metrics like FVD inappropriate. Therefore, we focus on demonstrating MAGI's long-range prediction capability in these specific scenarios to showcase its potential for capturing long-range temporal dependencies and motivate further investigation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7380" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.308, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 6. Long-term Video Prediction. MAGI predicts over 100 frames from a single input frame, maintaining reasonable motion even when trained on only 16 frames." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.352, + 0.232, + 0.367 + ], + "angle": 0, + "content": "6. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.378, + 0.483, + 0.681 + ], + "angle": 0, + "content": "Autoregressive Visual Generation Autoregressive language modeling [28, 29] has propelled the development of visual content generation using discrete visual tokens [41]. In this framework, pre-trained visual tokenizers like VQVAE [41] map visual patches into a discrete latent space, allowing visual generation to be approached similarly to language modeling. Early works such as DALL-E [30] focus on text-to-image generation by learning a joint distribution between text and discrete image representations using an autoregressive cross-entropy loss. Concurrently, VideoGPT [50] extends this idea to video generation, employing discrete tokens for autoregressive video prediction. VideoPoet [20] further advances this approach by integrating a causal video tokenizer, MAGVIT-v2 [52], for multimodal video generation. OmniTokenizer [45] proposes a unified tokenizer for both discrete and continuous representations, enabling patch-level autoregressive video generation. In contrast, our work focuses on frame-level causality rather than patch-level, avoiding the limitations of raster-scan order and outperforming patch-level baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Masked and Diffusion Video Generation Diffusion models have recently gained prominence in visual generation tasks [7, 11, 14, 16, 31], extending effectively to video generation. Video diffusion models [1, 17] employ bidirectional attention and binary mask embeddings to enable frame-level autoregressive prediction. Notable works such as GameNGen [40] use bidirectional diffusion models for real-time game generation. However, due to their bidirectional nature, these models cannot leverage KV Cache for extended video generation, limiting their scalability. Several masked video generators, such as Genie [2], extend MaskGIT [4] into a causal-attention-based architecture for video generation. Despite their advantages, these" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.445 + ], + "angle": 0, + "content": "methods suffer from the training-inference gap inherent in masked autoregressive modeling, which negatively impacts generation quality. In contrast, our approach fully leverages KV Cache during inference, facilitated by our training paradigm that bridges the training-inference gap through a novel Complete Teacher Forcing paradigm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Addressing Exposure Bias for Autoregressive Video Generation Autoregressive models often suffer from exposure bias [55] and error accumulation [21], which degrade performance over long sequences. To mitigate these issues, we adopt noise injection from GameN-Gen [40]—where Gaussian noise is added to observation frame latents during training. Additionally, we introduce dynamic interval training, exposing the model to frames with varying temporal intervals to enhance generalization. As shown in Fig. 4, both strategies individually improve robustness, and their combination yields even greater performance gains in autoregressive video generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.65, + 0.634, + 0.665 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present MAGI, a hybrid framework that combines masked and causal modeling to achieve efficient and scalable video generation. We identify that the teacher forcing paradigm makes a significant difference—our Complete Teacher Forcing (CTF) approach effectively bridges the training-inference gap inherent in Masked Teacher Forcing (MTF). Additionally, we introduce essential training strategies to alleviate exposure bias. Through comprehensive experiments, we validate the effectiveness of each component. Our final model achieves state-of-the-art performance among autoregressive video generation methods. We also demonstrate the generation of long, coherent video sequences exceeding 100 frames from training sequences as short as 16 frames, highlighting MAGI's potential as a scalable autoregressive video generation model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "7381" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "We would like to thank Tianhong Li for his insightful suggestions, invaluable help, and exceptional support. We would also like to thank Chenfei Wu, Haoyang Huang, Guoqing Ma, Hongyu Zhou, Liangyu Chen, Chunrui Han, Yimin Jiang and Yu Deng for their constructive discussions and advice, which greatly improved this work." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.219, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.484, + 0.312 + ], + "angle": 0, + "content": "[1] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.314, + 0.484, + 0.383 + ], + "angle": 0, + "content": "[2] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.384, + 0.483, + 0.425 + ], + "angle": 0, + "content": "[3] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808.01340, 2018. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.426, + 0.484, + 0.493 + ], + "angle": 0, + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022, 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.484, + 0.578 + ], + "angle": 0, + "content": "[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[6] Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. arXiv preprint arXiv:2407.01392, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[8] Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive KV cache compression for llms. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Anirudh Goyal, Alex Lamb, Ying Zhang, Saizheng Zhang, Aaron C. Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 4601-4609, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.218, + 0.906, + 0.285 + ], + "angle": 0, + "content": "[12] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[14] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2(3):4, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.454 + ], + "angle": 0, + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[18] Allan Jabri, David Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. arXiv preprint arXiv:2212.11972, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[19] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[20] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 1, 2, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[21] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Conference on robot learning, pages 143-156. PMLR, 2017. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[22] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Shi Luohe, Hongyi Zhang, Yao Yao, Zuchao Li, and hai zhao. Keep the cost down: A review on methods to optimize LLM's KV-cache consumption. In First Conference on Language Modeling, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7382" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[24] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[25] Shentong Mo and Yapeng Tian. Scaling diffusion mamba with bidirectional ssms for efficient image and video generation. arXiv preprint arXiv:2405.15881, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[26] NVIDIA. Cosmos tokenizer. https://github.com/NVIDIA/Cosmos-Tokenizer, 2024. Accessed: 2024-11-15. 5, 6, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[28] Alec Radford. Improving language understanding by generative pre-training. 2018. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[29] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[30] A. Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[32] Florian Schmidt. Generalization in generation: A closer look at exposure bias. In Proceedings of the 3rd Workshop on Neural Generation and Translation@EMNLP-IJCNLP 2019, Hong Kong, November 4, 2019, pages 157-167. Association for Computational Linguistics, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[33] Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Mostgan-v: Video generation with temporal motion styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5652-5661, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[34] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3626-3636, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[35] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[36] Yao Teng, Yue Wu, Han Shi, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[37] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[39] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[40] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. arXiv preprint arXiv:2408.14837, 2024. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[42] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[43] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[44] Jacob Walker, Ali Razavi, and Aäron van den Oord. Predicting video with vqvae. arXiv preprint arXiv:2103.01950, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[45] Junke Wang, Yi Jiang, Zehuan Yuan, Binyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. arXiv preprint arXiv:2406.09399, 2024. 2, 5, 6, 7, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[46] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[47] Ronald J. Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural Computation, 1(2):270-280, 1989. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[48] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In European conference on computer vision, pages 720-736. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[49] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution text-to-image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[50] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[51] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 1, 2, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7383" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.135 + ], + "angle": 0, + "content": "Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 7, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[53] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[54] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[55] Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. Bridging the gap between training and inference for neural machine translation. arXiv preprint arXiv:1906.02448, 2019. 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7384" + } + ] +] \ No newline at end of file diff --git a/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/e2161c88-f390-4144-acf4-52130bcde3bd_origin.pdf b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/e2161c88-f390-4144-acf4-52130bcde3bd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5692a98174ec0c72f77fc3ab172b3d746ffb0916 --- /dev/null +++ b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/e2161c88-f390-4144-acf4-52130bcde3bd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a7fc22b65f125a1c380207270960877cdd2c5b7b865166c1cfbebbc6ed2dcae +size 1697552 diff --git a/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/full.md b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..abc8a772a1e5637aa5af0c936d2bc71428401882 --- /dev/null +++ b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/full.md @@ -0,0 +1,274 @@ +# Taming Teacher Forcing for Masked Autoregressive Video Generation + +Deyu Zhou1 Quan Sun2 Yuang Peng2,4 Kun Yan2 Runpei Dong3 Duomin Wang2 Zheng Ge2 Nan Duan2 Xiangyu Zhang2 + +$^{1}$ HKUST(GZ) $^{2}$ StepFun $^{3}$ UIUC $^{4}$ THU MAGI-VIDEO-GENERATION.GITHUB.IO + +# Abstract + +We introduce MAGI, a hybrid video generation framework that combines masked modeling for intra-frame generation with causal modeling for next-frame generation. Our key innovation, Complete Teacher Forcing (CTF), conditions masked frames on complete observation frames rather than masked ones (namely Masked Teacher Forcing, MTF), enabling a smooth transition from token-level (patch-level) to frame-level autoregressive generation. CTF significantly outperforms MTF, achieving a $+23\%$ improvement in FVD scores on first-frame conditioned video prediction. To address issues like exposure bias, we employ targeted training strategies, setting a new benchmark in autoregressive video generation. Experiments show that MAGI can generate long, coherent video sequences exceeding 100 frames, even when trained on as few as 16 frames, highlighting its potential for scalable, high-quality video generation. + +# 1. Introduction + +Generating order matters in autoregressive image generation. While existing approaches mostly reply on simple raster-scan order, recent studies [8, 22] have demonstrated superior results through alternative generation strategies. These include randomized spatial order with bidirectional attention [4, 22, 51] and generation along increased scales [5, 8]. Despite its fundamental importance, in the realm of autoregressive video generation, however, the discussion of generation order has been largely overlooked. + +Existing autoregressive video generation methods can be categorized into two groups according to the prediction granularity, as shown in Tab. 1. The first group, exemplified by MAGViT [51], adopts a masked modeling approach* where visual tokens are generated in descending order according to their logit probabilities. While this approach uses bi-directional attention across frames, it presents two signif + +icant limitations: substantial computational overhead during inference since it is unable to utilize KV Cache [9, 23], and neglect temporal causality between consecutive frames. These oversights are particularly noteworthy given the remarkable success of causal modeling in large language models. Some other methods in this group [2, 6] have explored frame-level prediction with casual temporal attention. However, these methods are limited due to incomplete history observation, which causes an intrinsic training-inference gap. + +The second group comprises fully autoregressive approaches operating on visual patches [20, 46, 48, 50], exemplified by recent works like VideoPoet [20] and Emu3 [46]. Despite their successful temporal modeling, these methods have not incorporated contemporary advances from image generation research in intra-frame generation. Instead, they continue to rely on raster-scan order - an approach demonstrated sub-optimal in image generation [22]. + +In this paper, we introduce Masked Autoregressive video Generation (MAGI), a hybrid framework that synergizes the strengths of both video generation paradigms by combining masked modeling for intra-frame generation with causal modeling for inter-frame dependencies. We start with a naive implementation. This initial implementation explores a straightforward approach where tokens in each next frame are generated conditioned on previous masked frames (Figure 1(a)). While this can be viewed as a natural extension of MaskGIT [4] and Muse [5] to autoregressive video generation, our analysis reveals a fundamental limitation in its implementation of teacher forcing (TF) [47]. To be specific, TF involves feeding the model the groundtruth (correct) output from the previous time step as input during training, rather than using the model's own predicted output. However, our initial approach, which we term as Masked Teacher Forcing (MTF), deviates from this principle by substituting ground-truth tokens with mask tokens. Despite similar approaches being explored in interactive generative models (e.g., Genie [2]), we identify a critical training-inference discrepancy: during training, next-frame generation is conditioned on a mixture of visible + +Table 1. Comparison of autoregressive paradigms across different methods. We compare MAGI with various methods based on temporal attention mechanisms (how each frame attends to others), support for KV Cache [23] and variable context lengths, observation completeness (the completeness of the context frames used for conditioning), and the prediction granularity (patch or frame). $\dagger$ : MAGVIT predicts masked tokens of multiple frames simultaneously. Genie [2] and MAGI predict multiple tokens of single frame simultaneously. + +
Prediction GranularityMethodTemporal AttentionKV CacheVar. ContextObservation
PatchVideoGPT [50], Phenaki [43], Omni [45]Causal✓✓Complete
FrameMAGVIT† [51], GameNGen [40]Bidirectional✗✗Complete
Diffusion Forcing [6]Causal✓✓Noisy
Genie [2]Causal✓✓Masked
MAGI (with CTF)Causal✓✓Complete
+ +and masked tokens, while during inference, it relies solely on complete visual tokens from previous frames. Our experiments demonstrate that this inconsistency compromises generation quality, especially motion coherency. + +To address these limitations, we propose Complete Teacher Forcing (CTF), an advanced paradigm that maintains training-test consistency by conditioning next-frame generation on complete visual observations from history frames during training. This is achieved through a novel mechanism that prepends complete video tokens to the masked input sequence and employs a carefully designed attention mask. Empirical results show that CTF better captures motion, outperforming MTF by $23\%$ in terms of FVD scores on first-frame conditioned video prediction. + +In addition, we tackle the persistent challenge of exposure bias [32] in autoregressive video generation through two complementary strategies: dynamic interval training and noise injection [40]. Dynamic interval training refers to randomly sampling frames with varying intervals during training, which introduces diversity into the data distribution and helps the model to better generalize to different temporal frequencies. Dynamic noise injection, on the other hand, involves adding random noise to observation frames during training, which helps improve the model's robustness by simulating the possible errors may occur in inference. Combined with these two techniques, MAGI establishes a robust baseline for autoregressive video generation. Through extensive evaluations, we demonstrate that MAGI achieves superior video generation quality and length scalability. For example, MAGI can generate coherent video sequences that exceed 100 frames, even when trained on sequences as short as 16 frames. This underscores the potential of our approach for scalable video generation. + +# 2. Background and Problem Statement + +# 2.1. Autoregressive Video Generation + +Given a video $V \in \mathbb{R}^{N \times H \times W \times 3}$ , the sequence of $N$ frames with height $H$ and width $W$ can be denoted as $V = \{f_i \in \mathbb{R}^{H \times W \times 3} | i = 1, \dots, N\}$ . Each frame can be further divided into $m$ patches, represented as + +$\{\mathbf{x}_{11}, \mathbf{x}_{12}, \ldots, \mathbf{x}_{1m}, \ldots, \mathbf{x}_{n1}, \mathbf{x}_{n2}, \ldots, \mathbf{x}_{nm}\}$ . Autoregressive video generation models primarily differ in the granularity of causality (patch-level vs. frame-level) and the training paradigm (teacher forcing), discussed below. + +Patch-level Methods Patch-level autoregressive models [46, 50] operate by modeling the video on a patch-by-patch basis. The likelihood of generating entire video with a model parameterized by $\theta$ is decomposed into the product of posterior of each patch conditioned all preceding patches: + +$$ +p (V) = \prod_ {t = 1} ^ {N \times m} p \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {1}, \mathbf {x} _ {2}, \dots , \mathbf {x} _ {t - 1}; \theta\right), \tag {1} +$$ + +where $p_t$ represents a patch in a predefined ordering. The model generates each patch sequentially by conditioning on all previously generated patches, capturing both spatial and temporal dependencies. + +Frame-level Methods Frame-level autoregressive models [2, 40] model the video at the granularity of entire frames. Similarly, the likelihood of video modeling is: + +$$ +p (V) = \prod_ {t = 1} ^ {n} p \left(f _ {t} \mid f _ {1}, f _ {2}, \dots , f _ {t - 1}; \theta\right), \tag {2} +$$ + +where $f_{t}$ is the entire frame generated conditioned on all history frames. In this fashion, the patches in each frame are generated in parallel. + +Teacher Forcing Teacher forcing [47] is a widely adopted training strategy in autoregressive models, where the model conditions on the true previous frame or patch rather than its own predictions [10, 47]. This method is commonly used in patch-level autoregressive video models [20, 45, 48, 50] as well as in language models [28, 29], where the model is supervised through next token or patch prediction, effectively "shifting" one token or patch at a time during training. + +# 2.2. Can we achieve frame-level teacher forcing? + +While teacher forcing is effective in autoregressive language models [28, 29], its application to frame-level video generation is largely unexplored. Few works adopt a straightforward "shift one frame" paradigm for frame-level video models. + +![](images/aa7d1b37decb3ef7fdf4fdaa6f64c900dc1de1efdd31294131cbe3c3f4e03ff5.jpg) +Figure 1. Conceptual comparison of masked teacher forcing and our proposed complete teacher forcing mechanisms in autoregressive video generation. The blocks here illustrate individual frames. These two mechanisms differ in how history observation frames are used to condition next frame generation during training. (a): The observation frames are masked frames [2, 13] that are partially observable. (b): Every frame is generated conditioned on fully observable history frames. + +Instead, existing methods often rely on masked frame prediction, where masked frames are conditioned on other frames. For example, Genie [2] extends MaskGIT [4] to video generation by conditioning masked frames on other masked teacher frames—a strategy we refer to as Masked Teacher Forcing (MTF). Similarly, Diffusion Forcing [6] introduces temporal causal attention with a novel noise strategy but conditions each frame on noisy teacher frames, further diverging from traditional teacher forcing. GameNGen [40] employs bidirectional diffusion models with binary mask tokens to represent masked and conditional frames; however, it relies on fixed-length conditional frames, limiting autoregressive flexibility and preventing the use of efficient mechanisms like KV Cache—one notable advantage of traditional autoregressive methods. Generating videos autoregressively at the patch level is also suboptimal because inter-frames lack clear raster-scan spatial causality. In contrast, videos exhibit strong temporal causality, suggesting that predicting videos frame by frame, while generating the tokens within each frame in parallel, is more effective. + +To address these issues, MAGI fully leverages teacher forcing for frame-level autoregressive video generation, enabling more flexible, efficient, and scalable video generation. + +# 3. Approach + +Existing frame-level autoregressive video generation models, such as GameNGen [40], typically use binary masks to distinguish between frames used for prediction and those used for conditioning. These models are trained to predict future frames conditioned on a fixed-length sequence of preceding frames. This fixed conditioning contrasts with + +the greater flexibility of autoregressive language models, which condition on prefixes of variable length. In this work, we leverage MAR [22], a state-of-the-art masked image generation framework, and investigate crucial design choices for masked autoregressive video generation. Our focus includes exploring different autoregressive paradigms and methods for mitigating exposure bias and error accumulation during autoregressive inference. + +Masked Teacher Forcing vs. Complete Teacher Forcing. A natural extension of masked image generation [4, 22] to autoregressive video prediction is to modify the attention mechanism [2], incorporating causal temporal attention during the training phase for masked frames. We refer to this approach as Masked Teacher Forcing (MTF), as illustrated on the right of Fig. 1. However, this method introduces a significant training-inference gap: during training, the model attends to a high mask ratio, while at inference time, it must predict future frames conditioned on previously generated unmasked ones. This mismatch between training and inference undermines the model's ability to generate realistic videos, ultimately limiting its performance. + +The formal formulation of MTF is as follows: + +$$ +p \left(f _ {j} ^ {m} \mid f _ {1} ^ {m}, f _ {2} ^ {m}, \dots , f _ {j - 1} ^ {m}; \theta\right), \quad j \in \{1, 2, \dots , n \}. \tag {3} +$$ + +where $f_{j}^{m}$ represents the $j$ -th masked frame, while the frames $f_{1}^{m}, f_{2}^{m}, \ldots, f_{j-1}^{m}$ are the previously predicted masked frames. In this formulation, during training, each masked frame attends to previously masked frames and itself. However, the approach introduces a significant gap between the training and the inference phase, because the model must predict masked frames conditioned on previously unmasked generated ones at the inference stage. + +![](images/cf088ff7a49e0fa9532de1f576a7494a69c24f4feb6ee2df4e814647ca11c50f.jpg) +Figure 2. Overview of MAGI video generation framework. MAGI receives observation frames and corresponding masked frames as inputs, enabling autoregressive video generation with Complete Teacher Forcing (Sec. 3). + +To address this issue and more effectively bridge the gap between training and inference, we propose a novel autoregressive paradigm, termed Complete Teacher Forcing (CTF), as shown on the left of Fig. 1. Unlike MTF, CTF conditions on unmasked observed frames during training, predicting each masked frame. Formally, let the observation frames be denoted as $\{f_1,f_2,\dots ,f_n\}$ , where $n$ is the number of frames. The autoregressive formulation for predicting the $j$ -th masked frame, $f_{j}^{m}$ , is conditioned on both the previous unmasked observation frames and itself, i.e., + +$$ +p \left(f _ {j} ^ {m} \mid f _ {1}, f _ {2}, \dots , f _ {j - 1}; \theta\right), \quad j \in \{1, 2, \dots , n \}, \tag {4} +$$ + +where $p(f_{j}^{m}\mid \cdot ;\theta)$ is the autoregressive model. Note that $f_{1}^{m}$ is conditioned only on itself, which resembles image generation. For all $j > 1$ , $f_{j}^{m}$ attends to both the previous unmasked observation frames and itself, making this approach consistent between training and inference. + +Addressing Exposure Bias & Error Accumulation Although CTF reduces the gap between training and inference, it still faces challenges related to exposure bias [55] and error accumulation [21]. Specifically, the model may struggle to generate realistic video sequences during inference when relying on self-predicted frames. + +To mitigate these issues, we introduce two strategies: + +- Dynamic interval training. Video clips are sampled with random frame intervals, which forces the model to learn longer temporal dependencies and larger motion ranges, thus improving prediction stability for long videos. However, we find that a vanilla usage of this strategy leads to videos generated with uncontrollable motion range during inference. For example, prediction with a random motion range leads to unsatisfactory video generation results on UCF-101 that requires generation in 25 FPS. To support controllable generation to handle varying user-specified frame intervals, we introduce learnable interval + +![](images/39999c985198ed1caf1198e62569c2922f431c4e3d924a5c3548ea56d8c69e6b.jpg) +Figure 3. Temporal attention mask in our CTF during training. The attention within observation frames causal, while the attention within masked frames are atrous. In this fashion, each masked frame attends to itself and unmasked history observation frames. During inference, a standard causal mask is employed, where each frame attends only to previously generated frames. + +embeddings which encodes different intervals into specific embeddings like positional encoding [42]. Specifically, the $x$ -th interval embedding will be added to hidden states when sampling with an interval $x$ which enables the model's awareness of the desired generation interval. + +- Dynamic noise injection. As pointed out by Valevski et al. [40], there is a domain shift caused by teacher forcing and auto-regressive modeling, and thus data corruption using an injected noise during training is useful. Inspired by this practice, we also adopt a dynamic noise injection strategy. In addition, similar to interval embeddings, a learnable noise level embedding is concatenated with hidden states as the model inputs. Both techniques are crucial for improving the robustness of CTF by ensuring better generalization and stability during inference. + +# 4. Architecture + +We introduce MAGI, as shown in Fig. 2, a novel family of masked autoregressive video generation models, which incorporates the techniques outlined above. We implement MAGI with a Transformer architecture, detailed as follows. + +Transformer Decoder & Temporal Attention MAGI employs a stack of spatial-temporal Transformer blocks consisting of interleaved 2D spatial attention layers and 1D temporal attention layers. As shown in Fig. 2, weptides the complete observation frames with masked frames as the input of our Transformer. For CTF, we design a special temporal attention mask, as shown in Fig. 3. In the temporal layer, each frame only attends to itself and its preceding observation frames. The observation frames only attend to previous frames. Thus, the attention paradigm of CTF is consistent between training and inference. + +Diffusion Head [22] Atop the Transformer decoder, we stack multi-layer perceptron (MLP) layers as the diffusion head, following MAR [22]. This component predicts masked tokens through a denoising diffusion procedure [16], enhancing the model's capacity for autoregressive generation. + +Learnable Positional Embeddings To distinguish between masked and unmasked frames, we introduce two distinct learnable positional embeddings: one for the masked frames and another for the observation frames. Additionally, we adopt a learnable positional embedding for the diffusion head, as proposed in MAR [22]. + +Spatial and Temporal Positional Embeddings We utilize sinusoidal positional embeddings [42] for both 2D spatial and 1D temporal encodings, ensuring that the model effectively captures spatiotemporal relationships across frames. + +Frame Interval Embeddings As part of our dynamic interval training strategy, we introduce a learnable interval embedding, with a vocabulary length of 25, which covers frame intervals ranging from 1 to 25. This embedding allows the model to capture long-range temporal dependencies across varying frame intervals. + +Noise Level Embeddings To support dynamic noise injection, we incorporate random Gaussian noise into the observation frames during training, following GameNGen [40]. We use a noise level with a range of [1, 5], which is encoded through a learnable vocabulary embedding, with a dimension the same as the Transformer's hidden dimension. This enables the model to adapt to different noise levels during the denoising process. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets For the five-frame conditional video prediction tasks, we use the Kinetics-600 dataset [3], which consists of 480,000 videos spanning 600 action categories. For unconditional and first-frame conditional video generation $\ddagger$ , we utilize the UCF-101 dataset [35], containing over 13,000 clips across 101 human action classes and with totally 27 hours of recording length. + +Implementation Details We set the learning rate to $2 \times 10^{-4}$ for Kinetics-600 and $1 \times 10^{-4}$ for UCF-101. Our largest model is trained for 150 epochs on Kinetics-600 and 1,400 epochs on UCF-101, with batch sizes of 256 and 128, respectively. Other hyperparameters follow the settings in MAR [22]. For Kinetics-600, we use the 3D-VAE from OmniTokenizer [45]. For UCF-101, we use the 2D-VAE from Stable Diffusion 1.4 [31] and the 3D-VAE from Cosmos [26]. During training, we sample 17 frames for the 3D-VAE and 16 frames for the 2D-VAE, respectively. All models are trained at a resolution of $256 \times 256$ . + +Inference and Evaluation For inference, we follow the strategy outlined in MAR [22], using 64 iterative steps for masked frame prediction per frame. The inference procedure is significantly accelerated using KV Cache. For evaluation on Kinetics-600, we generate 50,000 videos using context frames randomly sampled from the test set and compute the Fréchet Video Distance (FVD) [39] against the ground-truth videos, both resized to $64 \times 64$ resolution. For UCF-101, we follow the evaluation protocol of previous works [24, 45]. We randomly sample 2,048 videos and compute the FVD against 2,048 randomly selected ground-truth videos from the dataset. + +# 5.2. Teacher Forcing Matters in Autoregressive Video Generation + +CTF Outperforms MTF Our CTF significantly outperforms MTF by nearly $23\%$ on first-frame conditioned video prediction. As shown in Fig. 5b, although MTF achieves slightly better frame-wise FID, it results in substantially worse overall FVD compared to CTF. Per-frame analysis (Fig. 4c-d) reveals that CTF better models temporal motion, while MTF generates high-quality static images lacking temporal coherence. + +Why Does CTF Achieve Superior FVD? We argue that although MTF's training on highly masked observations (using an optimal mask ratio of $70\% - 100\%$ ) is beneficial for generating visually similar frames and achieving lower FID scores, it imposes a significant limitation. Because MTF is trained to predict frames with minimal past information, it cannot effectively utilize full observations during inference. In contrast, CTF, trained on full observations, captures motion more accurately, leading to superior FVD scores. + +# 5.3. Addressing Exposure Bias + +Training Strategies Matters An ablation study (Fig. 5c) demonstrates the importance of our proposed training strategies: Removing either or both dynamic interval training $^{\S}$ and dynamic noise injection significantly degrades CTF's FVD and FID, underscoring the exposure bias in autoregressive prediction. Combining both yields the best performance, demonstrating their synergistic effect. + +Does MTF Also Benefit from These Training Strategies? These training strategies prove beneficial not only for CTF but also for MTF. While we focus our quantitative analysis on CTF and omit the full results for MTF due to space limitations, similar trends of performance improvement were observed. Notably, even when MTF benefits from these strategies, CTF consistently achieves superior results (see Fig. 5b), demonstrating its inherent advantages. + +![](images/694a6964ee89c28ae2311e21086a9e1f7e32d2c32b0bd5465f6769371fb53d4c.jpg) +Figure 4. Case Study of Proposed Training Techniques: This figure evaluates the impact of dynamic interval training and dynamic noise injection on CTF and MTF by: 1) visualizing CTF with and without these strategies; and 2) comparing CTF and MTF when both use them. All methods perform first-frame conditional video prediction on UCF-101 [35]. The results demonstrate the efficacy of the proposed training strategies and the superior motion and temporal coherence of CTF. + +![](images/d2b827bba5b5bfe63087066207ce06be0c5fa7c39e96fbb7069efd1d8342463f.jpg) + +# 5.4. Benchmarking with Previous Methods + +Video Prediction We evaluate our model trained on Kinetics-600 [3] against existing non-autoregressive (NAR) and autoregressive (AR) methods (Tab. 2). Our method, MAGI, achieves an FVD score of 11.5, establishing a new state-of-the-art among AR models and significantly outperforming the patch-level AR method Omni [45]. Notably, our MAGI achieves -21.4 significantly lower FVD than Omni whose FVD is 32.9. This result demonstrates the effectiveness of frame-level autoregressive modeling with Complete Teacher Forcing. Examples of generated videos are provided in Appendix Fig. 7. + +Unconditional Video Generation For unconditional video generation on UCF-101 [35] (Tab. 3), our method achieves state-of-the-art results among AR models, outperforming Latte [24], a DiT-based NAR model, by approximately 50 FVD points using the same VAE. Furthermore, using a stronger VAE (Cosmos [26]), our method becomes competitive with state-of-the-art NAR methods. + +# 5.5. Further Analysis + +Tab. 1 outlines the core differences between MAGI and other methods. MAGI achieves a smooth transition from patch-level to frame-level autoregressive generation while retaining key advantages, analyzed below. + +![](images/da383fab327517f737b56b64a51191efd3aedce132363d216ec2e716e2ff7a36.jpg) +(a) Speed comparison. + +![](images/06b4ae7a80482e3156ec42db492237ad0a84b1425acd998ddcbc16b3f457e251.jpg) +(b) MTF v.s. CTF. +Figure 5. a) Speed of generation process with varying numbers of frames. MAGI achieves a significant speedup by utilizing KV Cache. b) Masked Teacher Forcing (MTF) v.s. Complete Teacher Forcing (CTF). Both methods utilize the proposed training strategies — dynamic interval training and dynamic noise injection. We report FID scores for individual frames and FVD scores for all frames on UCF101 [35]. CTF achieves significantly better FVD scores but slightly worse FID scores compared to MTF. This result demonstrates that CTF better captures motion with temporal coherence, even when the quality of individual frames is lower. c) The results of ablation study on first-frame conditional video predcition of UCF-101 [35]. The "star" indicates the FVD scores of each method with the same color. + +![](images/ee43c7245cf011fbcb9022b6664a083e5e2b45f8623ddb4ef3386ccf24d7b57f.jpg) +(c) Ablation on first-frame conditional generation. + +Table 2. Video Prediction on Kinetics-600 [3]. The results are evaluated on the testset on Kinetics-600. We report the $\mathrm{FVD}_{64,50\mathrm{K}}$ obtained from $50\mathrm{K}$ samples in the resolution of $64\times 64$ . NAR: Non-autoregressive methods. AR: Autoregressive methods. + +
TypeMethodVAEFVD64,50K ↓
NARVideo Diffusion [17]-16.2
NARRIN [18]-10.8
NARMAGVIT [51]MAGVIT [51]9.9
NARMAGVIT-v2 [52]MAGVIT-v2 [52]4.3
ARViVQVAE [44]VQVAE [41]64.3
ARPhenaki [43]VQVAE [41]36.4
AROmni [45]Omni [45]32.9
ARMAGIOmni [45]11.5
+ +Generation Order Matters Our experiments (Fig. 5b and 5c) demonstrate that single-frame FID increases as the number of AR steps grows, regardless of the training paradigms and strategies employed. This finding underscores the importance of generation order in autoregressive video generation. Since frame-level generation requires fewer AR steps than patch-level generation, it can better preserve image quality over long sequences by mitigating error accumulation. Moreover, fewer autoregressive steps reduce the propagation of errors that typically accumulate in sequential generation processes, leading to more coherent and stable video outputs. This efficiency not only enhances visual fidelity but also contributes to faster inference times, which are critical for practical applications. + +KV Cache A key practical advantage of MAGI is its efficient inference. MAGI's frame-level KV Cache enables approximately linear scaling of inference time with the number of generated frames (Fig. 5a), a significant advantage over the parallel computation required by NAR methods. + +Table 3. Unconditional video generation on UCF-101 [35]. We report the $\mathrm{FVD}_{256,2048}$ which is the FVD obtained from 2048 samples in the resolution of $256\times 256$ + +
TypeMethodVAEFVD256, 2048 ↓
NARMoCoGAN [38]-2886.9
NARMoCoGAN-HD [37]-1729.6
NARDIGAN [53]-1630.2
NARStyleGAN-V [34]-1431.0
NARPVDM [54]PVDM [54]1141.9
NARMoStGAN-V [33]-1380.3
NARLatte [24]SD1.4 [31]477.9
NARDiM [25]SD1.4 [31]358.8
NARMatten [36]SD1.4 [31]210.6
ARVideoGPT [50]3D VQVAE [41]2880.6
ARMAGISD1.4 [31]420.6
ARMAGICosmos [26]297.8
+ +Practical Advantages of Autoregressive Generation To explore this, we train MAGI on UCF-101 [35] using randomly sampled 16-frame clips and evaluate its long-term prediction capability in a first-frame conditional generation setting. Fig. 6 demonstrates the promising predictive performance of MAGI, achieving reasonable results for sequences up to 100 frames in specific scenarios, particularly those with relatively static backgrounds and simple object motions. We observe performance degradation for videos with non-periodic motions (e.g., diving), which we attribute, in part, to the simplicity of the UCF-101 dataset and the challenging extrapolation required by the first-frame conditional setting. Specifically, for non-periodic actions like diving, the model lacks cues to predict subsequent actions once the primary action is complete, rendering metrics like FVD inappropriate. Therefore, we focus on demonstrating MAGI's long-range prediction capability in these specific scenarios to showcase its potential for capturing long-range temporal dependencies and motivate further investigation. + +![](images/a054ebd7b67da3fe5235072f683b4bf369a0574a3c62e702d0b0f9d0c01259f1.jpg) +Figure 6. Long-term Video Prediction. MAGI predicts over 100 frames from a single input frame, maintaining reasonable motion even when trained on only 16 frames. + +# 6. Related Work + +Autoregressive Visual Generation Autoregressive language modeling [28, 29] has propelled the development of visual content generation using discrete visual tokens [41]. In this framework, pre-trained visual tokenizers like VQVAE [41] map visual patches into a discrete latent space, allowing visual generation to be approached similarly to language modeling. Early works such as DALL-E [30] focus on text-to-image generation by learning a joint distribution between text and discrete image representations using an autoregressive cross-entropy loss. Concurrently, VideoGPT [50] extends this idea to video generation, employing discrete tokens for autoregressive video prediction. VideoPoet [20] further advances this approach by integrating a causal video tokenizer, MAGVIT-v2 [52], for multimodal video generation. OmniTokenizer [45] proposes a unified tokenizer for both discrete and continuous representations, enabling patch-level autoregressive video generation. In contrast, our work focuses on frame-level causality rather than patch-level, avoiding the limitations of raster-scan order and outperforming patch-level baselines. + +Masked and Diffusion Video Generation Diffusion models have recently gained prominence in visual generation tasks [7, 11, 14, 16, 31], extending effectively to video generation. Video diffusion models [1, 17] employ bidirectional attention and binary mask embeddings to enable frame-level autoregressive prediction. Notable works such as GameNGen [40] use bidirectional diffusion models for real-time game generation. However, due to their bidirectional nature, these models cannot leverage KV Cache for extended video generation, limiting their scalability. Several masked video generators, such as Genie [2], extend MaskGIT [4] into a causal-attention-based architecture for video generation. Despite their advantages, these + +methods suffer from the training-inference gap inherent in masked autoregressive modeling, which negatively impacts generation quality. In contrast, our approach fully leverages KV Cache during inference, facilitated by our training paradigm that bridges the training-inference gap through a novel Complete Teacher Forcing paradigm. + +Addressing Exposure Bias for Autoregressive Video Generation Autoregressive models often suffer from exposure bias [55] and error accumulation [21], which degrade performance over long sequences. To mitigate these issues, we adopt noise injection from GameN-Gen [40]—where Gaussian noise is added to observation frame latents during training. Additionally, we introduce dynamic interval training, exposing the model to frames with varying temporal intervals to enhance generalization. As shown in Fig. 4, both strategies individually improve robustness, and their combination yields even greater performance gains in autoregressive video generation. + +# 7. Conclusion + +We present MAGI, a hybrid framework that combines masked and causal modeling to achieve efficient and scalable video generation. We identify that the teacher forcing paradigm makes a significant difference—our Complete Teacher Forcing (CTF) approach effectively bridges the training-inference gap inherent in Masked Teacher Forcing (MTF). Additionally, we introduce essential training strategies to alleviate exposure bias. Through comprehensive experiments, we validate the effectiveness of each component. Our final model achieves state-of-the-art performance among autoregressive video generation methods. We also demonstrate the generation of long, coherent video sequences exceeding 100 frames from training sequences as short as 16 frames, highlighting MAGI's potential as a scalable autoregressive video generation model. + +# Acknowledgements + +We would like to thank Tianhong Li for his insightful suggestions, invaluable help, and exceptional support. We would also like to thank Chenfei Wu, Haoyang Huang, Guoqing Ma, Hongyu Zhou, Liangyu Chen, Chunrui Han, Yimin Jiang and Yu Deng for their constructive discussions and advice, which greatly improved this work. + +# References + +[1] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 8 +[2] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2, 3, 8 +[3] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808.01340, 2018. 5, 6, 7 +[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022, 1, 3, 8 +[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1 +[6] Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. arXiv preprint arXiv:2407.01392, 2024. 1, 2, 3 +[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 8 +[8] Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 1 +[9] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive KV cache compression for llms. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1 +[10] Anirudh Goyal, Alex Lamb, Ying Zhang, Saizheng Zhang, Aaron C. Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In + +Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 4601-4609, 2016. 2 + +[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 8 + +[12] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 13 + +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 3 +[14] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2(3):4, 2022. 8 +[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 12 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5, 8, 13 +[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 7, 8 +[18] Allan Jabri, David Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. arXiv preprint arXiv:2212.11972, 2022. 7 +[19] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 13 +[20] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 1, 2, 8, 12 +[21] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Conference on robot learning, pages 143-156. PMLR, 2017. 4, 8 +[22] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 5, 12, 13 +[23] Shi Luohe, Hongyi Zhang, Yao Yao, Zuchao Li, and hai zhao. Keep the cost down: A review on methods to optimize LLM's KV-cache consumption. In First Conference on Language Modeling, 2024. 1, 2 + +[24] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 6, 7 +[25] Shentong Mo and Yapeng Tian. Scaling diffusion mamba with bidirectional ssms for efficient image and video generation. arXiv preprint arXiv:2405.15881, 2024. 7 +[26] NVIDIA. Cosmos tokenizer. https://github.com/NVIDIA/Cosmos-Tokenizer, 2024. Accessed: 2024-11-15. 5, 6, 7, 12 +[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 13 +[28] Alec Radford. Improving language understanding by generative pre-training. 2018. 2, 8 +[29] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 8 +[30] A. Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 8 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5, 7, 8 +[32] Florian Schmidt. Generalization in generation: A closer look at exposure bias. In Proceedings of the 3rd Workshop on Neural Generation and Translation@EMNLP-IJCNLP 2019, Hong Kong, November 4, 2019, pages 157-167. Association for Computational Linguistics, 2019. 2 +[33] Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Mostgan-v: Video generation with temporal motion styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5652-5661, 2023. 7 +[34] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3626-3636, 2022. 7 +[35] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5, 6, 7 +[36] Yao Teng, Yue Wu, Han Shi, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024. 7 +[37] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021. 7 +[38] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for + +video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 7 +[39] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5 +[40] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. arXiv preprint arXiv:2408.14837, 2024. 2, 3, 4, 5, 8 +[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7, 8 +[42] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4, 5 +[43] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022. 2, 7 +[44] Jacob Walker, Ali Razavi, and Aäron van den Oord. Predicting video with vqvae. arXiv preprint arXiv:2103.01950, 2021. 7 +[45] Junke Wang, Yi Jiang, Zehuan Yuan, Binyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. arXiv preprint arXiv:2406.09399, 2024. 2, 5, 6, 7, 8, 12 +[46] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2 +[47] Ronald J. Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural Computation, 1(2):270-280, 1989. 1, 2 +[48] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In European conference on computer vision, pages 720-736. Springer, 2022. 1, 2 +[49] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution text-to-image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025. 13 +[50] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 1, 2, 7, 8 +[51] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 1, 2, 7, 12 +[52] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh + +Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 7, 8, 12 +[53] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 7 +[54] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 7 +[55] Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. Bridging the gap between training and inference for neural machine translation. arXiv preprint arXiv:1906.02448, 2019. 4, 8 \ No newline at end of file diff --git a/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/images.zip b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8d1208ea13bf09fae6713f59083812b3df9c4ad6 --- /dev/null +++ b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:003f99617406b6c07386647c80d4b61273c43a59ce3f12626093a35e7b68fc14 +size 601884 diff --git a/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/layout.json b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..291569cadff9dda06f22321c2564a3d04aaf27bc --- /dev/null +++ b/2025/Taming Teacher Forcing for Masked Autoregressive Video Generation/layout.json @@ -0,0 +1,7095 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 522, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 522, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 522, + 121 + ], + "type": "text", + "content": "Taming Teacher Forcing for Masked Autoregressive Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 141, + 522, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 141, + 522, + 173 + ], + "spans": [ + { + "bbox": [ + 87, + 141, + 522, + 173 + ], + "type": "text", + "content": "Deyu Zhou1 Quan Sun2 Yuang Peng2,4 Kun Yan2 Runpei Dong3 Duomin Wang2 Zheng Ge2 Nan Duan2 Xiangyu Zhang2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "spans": [ + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "type": "text", + "content": "HKUST(GZ) " + }, + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "type": "text", + "content": "StepFun " + }, + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "type": "text", + "content": "UIUC " + }, + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 175, + 175, + 435, + 204 + ], + "type": "text", + "content": "THU MAGI-VIDEO-GENERATION.GITHUB.IO" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 258, + 297, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 450 + ], + "type": "text", + "content": "We introduce MAGI, a hybrid video generation framework that combines masked modeling for intra-frame generation with causal modeling for next-frame generation. Our key innovation, Complete Teacher Forcing (CTF), conditions masked frames on complete observation frames rather than masked ones (namely Masked Teacher Forcing, MTF), enabling a smooth transition from token-level (patch-level) to frame-level autoregressive generation. CTF significantly outperforms MTF, achieving a " + }, + { + "bbox": [ + 54, + 258, + 297, + 450 + ], + "type": "inline_equation", + "content": "+23\\%" + }, + { + "bbox": [ + 54, + 258, + 297, + 450 + ], + "type": "text", + "content": " improvement in FVD scores on first-frame conditioned video prediction. To address issues like exposure bias, we employ targeted training strategies, setting a new benchmark in autoregressive video generation. Experiments show that MAGI can generate long, coherent video sequences exceeding 100 frames, even when trained on as few as 16 frames, highlighting its potential for scalable, high-quality video generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 135, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 135, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 135, + 485 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 493, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 601 + ], + "type": "text", + "content": "Generating order matters in autoregressive image generation. While existing approaches mostly reply on simple raster-scan order, recent studies [8, 22] have demonstrated superior results through alternative generation strategies. These include randomized spatial order with bidirectional attention [4, 22, 51] and generation along increased scales [5, 8]. Despite its fundamental importance, in the realm of autoregressive video generation, however, the discussion of generation order has been largely overlooked." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": "Existing autoregressive video generation methods can be categorized into two groups according to the prediction granularity, as shown in Tab. 1. The first group, exemplified by MAGViT [51], adopts a masked modeling approach* where visual tokens are generated in descending order according to their logit probabilities. While this approach uses bi-directional attention across frames, it presents two signif" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 233, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 553, + 353 + ], + "type": "text", + "content": "icant limitations: substantial computational overhead during inference since it is unable to utilize KV Cache [9, 23], and neglect temporal causality between consecutive frames. These oversights are particularly noteworthy given the remarkable success of causal modeling in large language models. Some other methods in this group [2, 6] have explored frame-level prediction with casual temporal attention. However, these methods are limited due to incomplete history observation, which causes an intrinsic training-inference gap." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 450 + ], + "type": "text", + "content": "The second group comprises fully autoregressive approaches operating on visual patches [20, 46, 48, 50], exemplified by recent works like VideoPoet [20] and Emu3 [46]. Despite their successful temporal modeling, these methods have not incorporated contemporary advances from image generation research in intra-frame generation. Instead, they continue to rely on raster-scan order - an approach demonstrated sub-optimal in image generation [22]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 451, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce Masked Autoregressive video Generation (MAGI), a hybrid framework that synergizes the strengths of both video generation paradigms by combining masked modeling for intra-frame generation with causal modeling for inter-frame dependencies. We start with a naive implementation. This initial implementation explores a straightforward approach where tokens in each next frame are generated conditioned on previous masked frames (Figure 1(a)). While this can be viewed as a natural extension of MaskGIT [4] and Muse [5] to autoregressive video generation, our analysis reveals a fundamental limitation in its implementation of teacher forcing (TF) [47]. To be specific, TF involves feeding the model the groundtruth (correct) output from the previous time step as input during training, rather than using the model's own predicted output. However, our initial approach, which we term as Masked Teacher Forcing (MTF), deviates from this principle by substituting ground-truth tokens with mask tokens. Despite similar approaches being explored in interactive generative models (e.g., Genie [2]), we identify a critical training-inference discrepancy: during training, next-frame generation is conditioned on a mixture of visible" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*We follow the definition in MAR [22], which considers masked image generation as a form of auto-regressive generation with a certain order." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7374" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 123, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "type": "text", + "content": "Table 1. Comparison of autoregressive paradigms across different methods. We compare MAGI with various methods based on temporal attention mechanisms (how each frame attends to others), support for KV Cache [23] and variable context lengths, observation completeness (the completeness of the context frames used for conditioning), and the prediction granularity (patch or frame). " + }, + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "type": "text", + "content": ": MAGVIT predicts masked tokens of multiple frames simultaneously. Genie [2] and MAGI predict multiple tokens of single frame simultaneously." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 123, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 123, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 123, + 553, + 205 + ], + "type": "table", + "html": "
Prediction GranularityMethodTemporal AttentionKV CacheVar. ContextObservation
PatchVideoGPT [50], Phenaki [43], Omni [45]Causal✓✓Complete
FrameMAGVIT† [51], GameNGen [40]Bidirectional✗✗Complete
Diffusion Forcing [6]Causal✓✓Noisy
Genie [2]Causal✓✓Masked
MAGI (with CTF)Causal✓✓Complete
", + "image_path": "938e2ae0f55b4ea66b5a1dd04a9ffc685e49dcc5d79e75eaf80ea1ddb204b944.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 217, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 266 + ], + "type": "text", + "content": "and masked tokens, while during inference, it relies solely on complete visual tokens from previous frames. Our experiments demonstrate that this inconsistency compromises generation quality, especially motion coherency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 266, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 295, + 386 + ], + "type": "text", + "content": "To address these limitations, we propose Complete Teacher Forcing (CTF), an advanced paradigm that maintains training-test consistency by conditioning next-frame generation on complete visual observations from history frames during training. This is achieved through a novel mechanism that prepends complete video tokens to the masked input sequence and employs a carefully designed attention mask. Empirical results show that CTF better captures motion, outperforming MTF by " + }, + { + "bbox": [ + 54, + 266, + 295, + 386 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 54, + 266, + 295, + 386 + ], + "type": "text", + "content": " in terms of FVD scores on first-frame conditioned video prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 386, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 615 + ], + "type": "text", + "content": "In addition, we tackle the persistent challenge of exposure bias [32] in autoregressive video generation through two complementary strategies: dynamic interval training and noise injection [40]. Dynamic interval training refers to randomly sampling frames with varying intervals during training, which introduces diversity into the data distribution and helps the model to better generalize to different temporal frequencies. Dynamic noise injection, on the other hand, involves adding random noise to observation frames during training, which helps improve the model's robustness by simulating the possible errors may occur in inference. Combined with these two techniques, MAGI establishes a robust baseline for autoregressive video generation. Through extensive evaluations, we demonstrate that MAGI achieves superior video generation quality and length scalability. For example, MAGI can generate coherent video sequences that exceed 100 frames, even when trained on sequences as short as 16 frames. This underscores the potential of our approach for scalable video generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 627, + 258, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 258, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 258, + 640 + ], + "type": "text", + "content": "2. Background and Problem Statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 233, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 233, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 233, + 659 + ], + "type": "text", + "content": "2.1. Autoregressive Video Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": "Given a video " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{N \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": ", the sequence of " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": " frames with height " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": " can be denoted as " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "V = \\{f_i \\in \\mathbb{R}^{H \\times W \\times 3} | i = 1, \\dots, N\\}" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": ". Each frame can be further divided into " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": " patches, represented as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 217, + 553, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{11}, \\mathbf{x}_{12}, \\ldots, \\mathbf{x}_{1m}, \\ldots, \\mathbf{x}_{n1}, \\mathbf{x}_{n2}, \\ldots, \\mathbf{x}_{nm}\\}" + }, + { + "bbox": [ + 313, + 217, + 553, + 265 + ], + "type": "text", + "content": ". Autoregressive video generation models primarily differ in the granularity of causality (patch-level vs. frame-level) and the training paradigm (teacher forcing), discussed below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 265, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 326 + ], + "type": "text", + "content": "Patch-level Methods Patch-level autoregressive models [46, 50] operate by modeling the video on a patch-by-patch basis. The likelihood of generating entire video with a model parameterized by " + }, + { + "bbox": [ + 313, + 265, + 553, + 326 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 265, + 553, + 326 + ], + "type": "text", + "content": " is decomposed into the product of posterior of each patch conditioned all preceding patches:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 331, + 553, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 331, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 348, + 331, + 553, + 365 + ], + "type": "interline_equation", + "content": "p (V) = \\prod_ {t = 1} ^ {N \\times m} p \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {1}, \\mathbf {x} _ {2}, \\dots , \\mathbf {x} _ {t - 1}; \\theta\\right), \\tag {1}", + "image_path": "00de8b1e8b6bbf3dd015f141b2e7f07da09cac02fa1ad6b40d435cdf01bb7c27.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 370, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 370, + 555, + 418 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 313, + 370, + 555, + 418 + ], + "type": "text", + "content": " represents a patch in a predefined ordering. The model generates each patch sequentially by conditioning on all previously generated patches, capturing both spatial and temporal dependencies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 418, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 553, + 455 + ], + "type": "text", + "content": "Frame-level Methods Frame-level autoregressive models [2, 40] model the video at the granularity of entire frames. Similarly, the likelihood of video modeling is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 460, + 553, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 460, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 354, + 460, + 553, + 491 + ], + "type": "interline_equation", + "content": "p (V) = \\prod_ {t = 1} ^ {n} p \\left(f _ {t} \\mid f _ {1}, f _ {2}, \\dots , f _ {t - 1}; \\theta\\right), \\tag {2}", + "image_path": "a22f4e4d1d8f8dc31b7fa4582a4639b1695fdc5f7870b8daaafb5b7235a576e7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "text", + "content": " is the entire frame generated conditioned on all history frames. In this fashion, the patches in each frame are generated in parallel." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 630 + ], + "type": "text", + "content": "Teacher Forcing Teacher forcing [47] is a widely adopted training strategy in autoregressive models, where the model conditions on the true previous frame or patch rather than its own predictions [10, 47]. This method is commonly used in patch-level autoregressive video models [20, 45, 48, 50] as well as in language models [28, 29], where the model is supervised through next token or patch prediction, effectively \"shifting\" one token or patch at a time during training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 635, + 542, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 542, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 542, + 649 + ], + "type": "text", + "content": "2.2. Can we achieve frame-level teacher forcing?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": "While teacher forcing is effective in autoregressive language models [28, 29], its application to frame-level video generation is largely unexplored. Few works adopt a straightforward \"shift one frame\" paradigm for frame-level video models." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7375" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 75, + 486, + 238 + ], + "blocks": [ + { + "bbox": [ + 120, + 75, + 486, + 238 + ], + "lines": [ + { + "bbox": [ + 120, + 75, + 486, + 238 + ], + "spans": [ + { + "bbox": [ + 120, + 75, + 486, + 238 + ], + "type": "image", + "image_path": "aa7d1b37decb3ef7fdf4fdaa6f64c900dc1de1efdd31294131cbe3c3f4e03ff5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "text", + "content": "Figure 1. Conceptual comparison of masked teacher forcing and our proposed complete teacher forcing mechanisms in autoregressive video generation. The blocks here illustrate individual frames. These two mechanisms differ in how history observation frames are used to condition next frame generation during training. (a): The observation frames are masked frames [2, 13] that are partially observable. (b): Every frame is generated conditioned on fully observable history frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 313, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 296, + 563 + ], + "type": "text", + "content": "Instead, existing methods often rely on masked frame prediction, where masked frames are conditioned on other frames. For example, Genie [2] extends MaskGIT [4] to video generation by conditioning masked frames on other masked teacher frames—a strategy we refer to as Masked Teacher Forcing (MTF). Similarly, Diffusion Forcing [6] introduces temporal causal attention with a novel noise strategy but conditions each frame on noisy teacher frames, further diverging from traditional teacher forcing. GameNGen [40] employs bidirectional diffusion models with binary mask tokens to represent masked and conditional frames; however, it relies on fixed-length conditional frames, limiting autoregressive flexibility and preventing the use of efficient mechanisms like KV Cache—one notable advantage of traditional autoregressive methods. Generating videos autoregressively at the patch level is also suboptimal because inter-frames lack clear raster-scan spatial causality. In contrast, videos exhibit strong temporal causality, suggesting that predicting videos frame by frame, while generating the tokens within each frame in parallel, is more effective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 564, + 295, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 611 + ], + "type": "text", + "content": "To address these issues, MAGI fully leverages teacher forcing for frame-level autoregressive video generation, enabling more flexible, efficient, and scalable video generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 621, + 122, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 122, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 122, + 635 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Existing frame-level autoregressive video generation models, such as GameNGen [40], typically use binary masks to distinguish between frames used for prediction and those used for conditioning. These models are trained to predict future frames conditioned on a fixed-length sequence of preceding frames. This fixed conditioning contrasts with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 313, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 409 + ], + "type": "text", + "content": "the greater flexibility of autoregressive language models, which condition on prefixes of variable length. In this work, we leverage MAR [22], a state-of-the-art masked image generation framework, and investigate crucial design choices for masked autoregressive video generation. Our focus includes exploring different autoregressive paradigms and methods for mitigating exposure bias and error accumulation during autoregressive inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 423, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 578 + ], + "type": "text", + "content": "Masked Teacher Forcing vs. Complete Teacher Forcing. A natural extension of masked image generation [4, 22] to autoregressive video prediction is to modify the attention mechanism [2], incorporating causal temporal attention during the training phase for masked frames. We refer to this approach as Masked Teacher Forcing (MTF), as illustrated on the right of Fig. 1. However, this method introduces a significant training-inference gap: during training, the model attends to a high mask ratio, while at inference time, it must predict future frames conditioned on previously generated unmasked ones. This mismatch between training and inference undermines the model's ability to generate realistic videos, ultimately limiting its performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 578, + 512, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 578, + 512, + 590 + ], + "spans": [ + { + "bbox": [ + 326, + 578, + 512, + 590 + ], + "type": "text", + "content": "The formal formulation of MTF is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 597, + 555, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 597, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 325, + 597, + 555, + 612 + ], + "type": "interline_equation", + "content": "p \\left(f _ {j} ^ {m} \\mid f _ {1} ^ {m}, f _ {2} ^ {m}, \\dots , f _ {j - 1} ^ {m}; \\theta\\right), \\quad j \\in \\{1, 2, \\dots , n \\}. \\tag {3}", + "image_path": "55e3bce6c00de6f62b4c493a8b02e4449be02f68ed489aab8f4ddefcf6298dd9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "f_{j}^{m}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "-th masked frame, while the frames " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "f_{1}^{m}, f_{2}^{m}, \\ldots, f_{j-1}^{m}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " are the previously predicted masked frames. In this formulation, during training, each masked frame attends to previously masked frames and itself. However, the approach introduces a significant gap between the training and the inference phase, because the model must predict masked frames conditioned on previously unmasked generated ones at the inference stage." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7376" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 292, + 203 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 292, + 203 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 292, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 292, + 203 + ], + "type": "image", + "image_path": "cf088ff7a49e0fa9532de1f576a7494a69c24f4feb6ee2df4e814647ca11c50f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 207, + 295, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 253 + ], + "type": "text", + "content": "Figure 2. Overview of MAGI video generation framework. MAGI receives observation frames and corresponding masked frames as inputs, enabling autoregressive video generation with Complete Teacher Forcing (Sec. 3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "text", + "content": "To address this issue and more effectively bridge the gap between training and inference, we propose a novel autoregressive paradigm, termed Complete Teacher Forcing (CTF), as shown on the left of Fig. 1. Unlike MTF, CTF conditions on unmasked observed frames during training, predicting each masked frame. Formally, let the observation frames be denoted as " + }, + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\{f_1,f_2,\\dots ,f_n\\}" + }, + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "text", + "content": " is the number of frames. The autoregressive formulation for predicting the " + }, + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "text", + "content": " -th masked frame, " + }, + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "inline_equation", + "content": "f_{j}^{m}" + }, + { + "bbox": [ + 55, + 268, + 296, + 388 + ], + "type": "text", + "content": " , is conditioned on both the previous unmasked observation frames and itself, i.e.," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 398, + 295, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 398, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 70, + 398, + 295, + 412 + ], + "type": "interline_equation", + "content": "p \\left(f _ {j} ^ {m} \\mid f _ {1}, f _ {2}, \\dots , f _ {j - 1}; \\theta\\right), \\quad j \\in \\{1, 2, \\dots , n \\}, \\tag {4}", + "image_path": "e24734be252d5f3c9ddb57700effef9ca11e60da1a1cec5cc71545e2b2c1be65.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "inline_equation", + "content": "p(f_{j}^{m}\\mid \\cdot ;\\theta)" + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "text", + "content": " is the autoregressive model. Note that " + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "inline_equation", + "content": "f_{1}^{m}" + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "text", + "content": " is conditioned only on itself, which resembles image generation. For all " + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "inline_equation", + "content": "j > 1" + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "inline_equation", + "content": "f_{j}^{m}" + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "text", + "content": " attends to both the previous unmasked observation frames and itself, making this approach consistent between training and inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "type": "text", + "content": "Addressing Exposure Bias & Error Accumulation Although CTF reduces the gap between training and inference, it still faces challenges related to exposure bias [55] and error accumulation [21]. Specifically, the model may struggle to generate realistic video sequences during inference when relying on self-predicted frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 570, + 282, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 570, + 282, + 582 + ], + "spans": [ + { + "bbox": [ + 66, + 570, + 282, + 582 + ], + "type": "text", + "content": "To mitigate these issues, we introduce two strategies:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "- Dynamic interval training. Video clips are sampled with random frame intervals, which forces the model to learn longer temporal dependencies and larger motion ranges, thus improving prediction stability for long videos. However, we find that a vanilla usage of this strategy leads to videos generated with uncontrollable motion range during inference. For example, prediction with a random motion range leads to unsatisfactory video generation results on UCF-101 that requires generation in 25 FPS. To support controllable generation to handle varying user-specified frame intervals, we introduce learnable interval" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 365, + 73, + 511, + 219 + ], + "blocks": [ + { + "bbox": [ + 365, + 73, + 511, + 219 + ], + "lines": [ + { + "bbox": [ + 365, + 73, + 511, + 219 + ], + "spans": [ + { + "bbox": [ + 365, + 73, + 511, + 219 + ], + "type": "image", + "image_path": "39999c985198ed1caf1198e62569c2922f431c4e3d924a5c3548ea56d8c69e6b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 227, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 293 + ], + "type": "text", + "content": "Figure 3. Temporal attention mask in our CTF during training. The attention within observation frames causal, while the attention within masked frames are atrous. In this fashion, each masked frame attends to itself and unmasked history observation frames. During inference, a standard causal mask is employed, where each frame attends only to previously generated frames." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 304, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 304, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 321, + 304, + 555, + 365 + ], + "type": "text", + "content": "embeddings which encodes different intervals into specific embeddings like positional encoding [42]. Specifically, the " + }, + { + "bbox": [ + 321, + 304, + 555, + 365 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 321, + 304, + 555, + 365 + ], + "type": "text", + "content": "-th interval embedding will be added to hidden states when sampling with an interval " + }, + { + "bbox": [ + 321, + 304, + 555, + 365 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 321, + 304, + 555, + 365 + ], + "type": "text", + "content": " which enables the model's awareness of the desired generation interval." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 365, + 556, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 365, + 556, + 485 + ], + "spans": [ + { + "bbox": [ + 314, + 365, + 556, + 485 + ], + "type": "text", + "content": "- Dynamic noise injection. As pointed out by Valevski et al. [40], there is a domain shift caused by teacher forcing and auto-regressive modeling, and thus data corruption using an injected noise during training is useful. Inspired by this practice, we also adopt a dynamic noise injection strategy. In addition, similar to interval embeddings, a learnable noise level embedding is concatenated with hidden states as the model inputs. Both techniques are crucial for improving the robustness of CTF by ensuring better generalization and stability during inference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 497, + 394, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 394, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 394, + 509 + ], + "type": "text", + "content": "4. Architecture" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 517, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 564 + ], + "type": "text", + "content": "We introduce MAGI, as shown in Fig. 2, a novel family of masked autoregressive video generation models, which incorporates the techniques outlined above. We implement MAGI with a Transformer architecture, detailed as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Transformer Decoder & Temporal Attention MAGI employs a stack of spatial-temporal Transformer blocks consisting of interleaved 2D spatial attention layers and 1D temporal attention layers. As shown in Fig. 2, weptides the complete observation frames with masked frames as the input of our Transformer. For CTF, we design a special temporal attention mask, as shown in Fig. 3. In the temporal layer, each frame only attends to itself and its preceding observation frames. The observation frames only attend to previous frames. Thus, the attention paradigm of CTF is consistent between training and inference." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7377" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": "Diffusion Head [22] Atop the Transformer decoder, we stack multi-layer perceptron (MLP) layers as the diffusion head, following MAR [22]. This component predicts masked tokens through a denoising diffusion procedure [16], enhancing the model's capacity for autoregressive generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 215 + ], + "type": "text", + "content": "Learnable Positional Embeddings To distinguish between masked and unmasked frames, we introduce two distinct learnable positional embeddings: one for the masked frames and another for the observation frames. Additionally, we adopt a learnable positional embedding for the diffusion head, as proposed in MAR [22]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 262 + ], + "type": "text", + "content": "Spatial and Temporal Positional Embeddings We utilize sinusoidal positional embeddings [42] for both 2D spatial and 1D temporal encodings, ensuring that the model effectively captures spatiotemporal relationships across frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 334 + ], + "type": "text", + "content": "Frame Interval Embeddings As part of our dynamic interval training strategy, we introduce a learnable interval embedding, with a vocabulary length of 25, which covers frame intervals ranging from 1 to 25. This embedding allows the model to capture long-range temporal dependencies across varying frame intervals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 335, + 294, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 294, + 431 + ], + "type": "text", + "content": "Noise Level Embeddings To support dynamic noise injection, we incorporate random Gaussian noise into the observation frames during training, following GameNGen [40]. We use a noise level with a range of [1, 5], which is encoded through a learnable vocabulary embedding, with a dimension the same as the Transformer's hidden dimension. This enables the model to adapt to different noise levels during the denoising process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 137, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 137, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 137, + 455 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 461, + 171, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 171, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 171, + 474 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 478, + 294, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 294, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 294, + 562 + ], + "type": "text", + "content": "Datasets For the five-frame conditional video prediction tasks, we use the Kinetics-600 dataset [3], which consists of 480,000 videos spanning 600 action categories. For unconditional and first-frame conditional video generation" + }, + { + "bbox": [ + 55, + 478, + 294, + 562 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 478, + 294, + 562 + ], + "type": "text", + "content": ", we utilize the UCF-101 dataset [35], containing over 13,000 clips across 101 human action classes and with totally 27 hours of recording length." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 565, + 294, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 294, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 294, + 696 + ], + "type": "text", + "content": "Implementation Details We set the learning rate to " + }, + { + "bbox": [ + 55, + 565, + 294, + 696 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 565, + 294, + 696 + ], + "type": "text", + "content": " for Kinetics-600 and " + }, + { + "bbox": [ + 55, + 565, + 294, + 696 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 565, + 294, + 696 + ], + "type": "text", + "content": " for UCF-101. Our largest model is trained for 150 epochs on Kinetics-600 and 1,400 epochs on UCF-101, with batch sizes of 256 and 128, respectively. Other hyperparameters follow the settings in MAR [22]. For Kinetics-600, we use the 3D-VAE from OmniTokenizer [45]. For UCF-101, we use the 2D-VAE from Stable Diffusion 1.4 [31] and the 3D-VAE from Cosmos [26]. During training, we sample 17 frames for the 3D-VAE and 16 frames for the 2D-VAE, respectively. All models are trained at a resolution of " + }, + { + "bbox": [ + 55, + 565, + 294, + 696 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 565, + 294, + 696 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "Inference and Evaluation For inference, we follow the strategy outlined in MAR [22], using 64 iterative steps for masked frame prediction per frame. The inference procedure is significantly accelerated using KV Cache. For evaluation on Kinetics-600, we generate 50,000 videos using context frames randomly sampled from the test set and compute the Fréchet Video Distance (FVD) [39] against the ground-truth videos, both resized to " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " resolution. For UCF-101, we follow the evaluation protocol of previous works [24, 45]. We randomly sample 2,048 videos and compute the FVD against 2,048 randomly selected ground-truth videos from the dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 222, + 553, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 553, + 246 + ], + "type": "text", + "content": "5.2. Teacher Forcing Matters in Autoregressive Video Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 251, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 347 + ], + "type": "text", + "content": "CTF Outperforms MTF Our CTF significantly outperforms MTF by nearly " + }, + { + "bbox": [ + 313, + 251, + 553, + 347 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 313, + 251, + 553, + 347 + ], + "type": "text", + "content": " on first-frame conditioned video prediction. As shown in Fig. 5b, although MTF achieves slightly better frame-wise FID, it results in substantially worse overall FVD compared to CTF. Per-frame analysis (Fig. 4c-d) reveals that CTF better models temporal motion, while MTF generates high-quality static images lacking temporal coherence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 361, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 469 + ], + "type": "text", + "content": "Why Does CTF Achieve Superior FVD? We argue that although MTF's training on highly masked observations (using an optimal mask ratio of " + }, + { + "bbox": [ + 313, + 361, + 554, + 469 + ], + "type": "inline_equation", + "content": "70\\% - 100\\%" + }, + { + "bbox": [ + 313, + 361, + 554, + 469 + ], + "type": "text", + "content": ") is beneficial for generating visually similar frames and achieving lower FID scores, it imposes a significant limitation. Because MTF is trained to predict frames with minimal past information, it cannot effectively utilize full observations during inference. In contrast, CTF, trained on full observations, captures motion more accurately, leading to superior FVD scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 475, + 460, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 460, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 460, + 489 + ], + "type": "text", + "content": "5.3. Addressing Exposure Bias" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 493, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 553, + 578 + ], + "type": "text", + "content": "Training Strategies Matters An ablation study (Fig. 5c) demonstrates the importance of our proposed training strategies: Removing either or both dynamic interval training" + }, + { + "bbox": [ + 313, + 493, + 553, + 578 + ], + "type": "inline_equation", + "content": "^{\\S}" + }, + { + "bbox": [ + 313, + 493, + 553, + 578 + ], + "type": "text", + "content": " and dynamic noise injection significantly degrades CTF's FVD and FID, underscoring the exposure bias in autoregressive prediction. Combining both yields the best performance, demonstrating their synergistic effect." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 591, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 553, + 687 + ], + "type": "text", + "content": "Does MTF Also Benefit from These Training Strategies? These training strategies prove beneficial not only for CTF but also for MTF. While we focus our quantitative analysis on CTF and omit the full results for MTF due to space limitations, similar trends of performance improvement were observed. Notably, even when MTF benefits from these strategies, CTF consistently achieves superior results (see Fig. 5b), demonstrating its inherent advantages." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "inline_equation", + "content": "^{\\S}" + }, + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "Further analysis, detailed in Appendix B.2, demonstrates how dynamic interval training mitigates exposure bias." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "type": "text", + "content": "Class-conditional results are included in Appendix B.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7378" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 82, + 386, + 487 + ], + "blocks": [ + { + "bbox": [ + 70, + 82, + 386, + 487 + ], + "lines": [ + { + "bbox": [ + 70, + 82, + 386, + 487 + ], + "spans": [ + { + "bbox": [ + 70, + 82, + 386, + 487 + ], + "type": "image", + "image_path": "694a6964ee89c28ae2311e21086a9e1f7e32d2c32b0bd5465f6769371fb53d4c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 494, + 555, + 539 + ], + "lines": [ + { + "bbox": [ + 55, + 494, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 555, + 539 + ], + "type": "text", + "content": "Figure 4. Case Study of Proposed Training Techniques: This figure evaluates the impact of dynamic interval training and dynamic noise injection on CTF and MTF by: 1) visualizing CTF with and without these strategies; and 2) comparing CTF and MTF when both use them. All methods perform first-frame conditional video prediction on UCF-101 [35]. The results demonstrate the efficacy of the proposed training strategies and the superior motion and temporal coherence of CTF." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 389, + 82, + 539, + 487 + ], + "blocks": [ + { + "bbox": [ + 389, + 82, + 539, + 487 + ], + "lines": [ + { + "bbox": [ + 389, + 82, + 539, + 487 + ], + "spans": [ + { + "bbox": [ + 389, + 82, + 539, + 487 + ], + "type": "image", + "image_path": "d2b827bba5b5bfe63087066207ce06be0c5fa7c39e96fbb7069efd1d8342463f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 557, + 258, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 258, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 258, + 570 + ], + "type": "text", + "content": "5.4. Benchmarking with Previous Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Video Prediction We evaluate our model trained on Kinetics-600 [3] against existing non-autoregressive (NAR) and autoregressive (AR) methods (Tab. 2). Our method, MAGI, achieves an FVD score of 11.5, establishing a new state-of-the-art among AR models and significantly outperforming the patch-level AR method Omni [45]. Notably, our MAGI achieves -21.4 significantly lower FVD than Omni whose FVD is 32.9. This result demonstrates the effectiveness of frame-level autoregressive modeling with Complete Teacher Forcing. Examples of generated videos are provided in Appendix Fig. 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 557, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 641 + ], + "type": "text", + "content": "Unconditional Video Generation For unconditional video generation on UCF-101 [35] (Tab. 3), our method achieves state-of-the-art results among AR models, outperforming Latte [24], a DiT-based NAR model, by approximately 50 FVD points using the same VAE. Furthermore, using a stronger VAE (Cosmos [26]), our method becomes competitive with state-of-the-art NAR methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 648, + 416, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 416, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 416, + 661 + ], + "type": "text", + "content": "5.5. Further Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Tab. 1 outlines the core differences between MAGI and other methods. MAGI achieves a smooth transition from patch-level to frame-level autoregressive generation while retaining key advantages, analyzed below." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7379" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 77, + 208, + 179 + ], + "blocks": [ + { + "bbox": [ + 58, + 77, + 208, + 179 + ], + "lines": [ + { + "bbox": [ + 58, + 77, + 208, + 179 + ], + "spans": [ + { + "bbox": [ + 58, + 77, + 208, + 179 + ], + "type": "image", + "image_path": "da383fab327517f737b56b64a51191efd3aedce132363d216ec2e716e2ff7a36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 182, + 176, + 193 + ], + "lines": [ + { + "bbox": [ + 102, + 182, + 176, + 193 + ], + "spans": [ + { + "bbox": [ + 102, + 182, + 176, + 193 + ], + "type": "text", + "content": "(a) Speed comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 77, + 386, + 180 + ], + "blocks": [ + { + "bbox": [ + 228, + 77, + 386, + 180 + ], + "lines": [ + { + "bbox": [ + 228, + 77, + 386, + 180 + ], + "spans": [ + { + "bbox": [ + 228, + 77, + 386, + 180 + ], + "type": "image", + "image_path": "06b4ae7a80482e3156ec42db492237ad0a84b1425acd998ddcbc16b3f457e251.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 182, + 337, + 191 + ], + "lines": [ + { + "bbox": [ + 272, + 182, + 337, + 191 + ], + "spans": [ + { + "bbox": [ + 272, + 182, + 337, + 191 + ], + "type": "text", + "content": "(b) MTF v.s. CTF." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "type": "text", + "content": "Figure 5. a) Speed of generation process with varying numbers of frames. MAGI achieves a significant speedup by utilizing KV Cache. b) Masked Teacher Forcing (MTF) v.s. Complete Teacher Forcing (CTF). Both methods utilize the proposed training strategies — dynamic interval training and dynamic noise injection. We report FID scores for individual frames and FVD scores for all frames on UCF101 [35]. CTF achieves significantly better FVD scores but slightly worse FID scores compared to MTF. This result demonstrates that CTF better captures motion with temporal coherence, even when the quality of individual frames is lower. c) The results of ablation study on first-frame conditional video predcition of UCF-101 [35]. The \"star\" indicates the FVD scores of each method with the same color." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 78, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 389, + 78, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 389, + 78, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 389, + 78, + 553, + 180 + ], + "type": "image", + "image_path": "ee43c7245cf011fbcb9022b6664a083e5e2b45f8623ddb4ef3386ccf24d7b57f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 182, + 551, + 192 + ], + "lines": [ + { + "bbox": [ + 390, + 182, + 551, + 192 + ], + "spans": [ + { + "bbox": [ + 390, + 182, + 551, + 192 + ], + "type": "text", + "content": "(c) Ablation on first-frame conditional generation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 335, + 294, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 287, + 295, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 295, + 332 + ], + "type": "text", + "content": "Table 2. Video Prediction on Kinetics-600 [3]. The results are evaluated on the testset on Kinetics-600. We report the " + }, + { + "bbox": [ + 55, + 287, + 295, + 332 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{64,50\\mathrm{K}}" + }, + { + "bbox": [ + 55, + 287, + 295, + 332 + ], + "type": "text", + "content": " obtained from " + }, + { + "bbox": [ + 55, + 287, + 295, + 332 + ], + "type": "inline_equation", + "content": "50\\mathrm{K}" + }, + { + "bbox": [ + 55, + 287, + 295, + 332 + ], + "type": "text", + "content": " samples in the resolution of " + }, + { + "bbox": [ + 55, + 287, + 295, + 332 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 55, + 287, + 295, + 332 + ], + "type": "text", + "content": ". NAR: Non-autoregressive methods. AR: Autoregressive methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 335, + 294, + 446 + ], + "lines": [ + { + "bbox": [ + 58, + 335, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 58, + 335, + 294, + 446 + ], + "type": "table", + "html": "
TypeMethodVAEFVD64,50K ↓
NARVideo Diffusion [17]-16.2
NARRIN [18]-10.8
NARMAGVIT [51]MAGVIT [51]9.9
NARMAGVIT-v2 [52]MAGVIT-v2 [52]4.3
ARViVQVAE [44]VQVAE [41]64.3
ARPhenaki [43]VQVAE [41]36.4
AROmni [45]Omni [45]32.9
ARMAGIOmni [45]11.5
", + "image_path": "3afd00392388f4f9e4826b7fb84851b5ff32f26bb3995cf4b56fac28e89c6868.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 466, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 466, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 54, + 466, + 295, + 634 + ], + "type": "text", + "content": "Generation Order Matters Our experiments (Fig. 5b and 5c) demonstrate that single-frame FID increases as the number of AR steps grows, regardless of the training paradigms and strategies employed. This finding underscores the importance of generation order in autoregressive video generation. Since frame-level generation requires fewer AR steps than patch-level generation, it can better preserve image quality over long sequences by mitigating error accumulation. Moreover, fewer autoregressive steps reduce the propagation of errors that typically accumulate in sequential generation processes, leading to more coherent and stable video outputs. This efficiency not only enhances visual fidelity but also contributes to faster inference times, which are critical for practical applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": "KV Cache A key practical advantage of MAGI is its efficient inference. MAGI's frame-level KV Cache enables approximately linear scaling of inference time with the number of generated frames (Fig. 5a), a significant advantage over the parallel computation required by NAR methods." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 325, + 553, + 468 + ], + "blocks": [ + { + "bbox": [ + 313, + 287, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 321 + ], + "type": "text", + "content": "Table 3. Unconditional video generation on UCF-101 [35]. We report the " + }, + { + "bbox": [ + 313, + 287, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{256,2048}" + }, + { + "bbox": [ + 313, + 287, + 555, + 321 + ], + "type": "text", + "content": " which is the FVD obtained from 2048 samples in the resolution of " + }, + { + "bbox": [ + 313, + 287, + 555, + 321 + ], + "type": "inline_equation", + "content": "256\\times 256" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 325, + 553, + 468 + ], + "lines": [ + { + "bbox": [ + 317, + 325, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 317, + 325, + 553, + 468 + ], + "type": "table", + "html": "
TypeMethodVAEFVD256, 2048 ↓
NARMoCoGAN [38]-2886.9
NARMoCoGAN-HD [37]-1729.6
NARDIGAN [53]-1630.2
NARStyleGAN-V [34]-1431.0
NARPVDM [54]PVDM [54]1141.9
NARMoStGAN-V [33]-1380.3
NARLatte [24]SD1.4 [31]477.9
NARDiM [25]SD1.4 [31]358.8
NARMatten [36]SD1.4 [31]210.6
ARVideoGPT [50]3D VQVAE [41]2880.6
ARMAGISD1.4 [31]420.6
ARMAGICosmos [26]297.8
", + "image_path": "bc074493cdebe2122195518364999b7f859d9a48f1dbada68e932ca05b589f70.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "Practical Advantages of Autoregressive Generation To explore this, we train MAGI on UCF-101 [35] using randomly sampled 16-frame clips and evaluate its long-term prediction capability in a first-frame conditional generation setting. Fig. 6 demonstrates the promising predictive performance of MAGI, achieving reasonable results for sequences up to 100 frames in specific scenarios, particularly those with relatively static backgrounds and simple object motions. We observe performance degradation for videos with non-periodic motions (e.g., diving), which we attribute, in part, to the simplicity of the UCF-101 dataset and the challenging extrapolation required by the first-frame conditional setting. Specifically, for non-periodic actions like diving, the model lacks cues to predict subsequent actions once the primary action is complete, rendering metrics like FVD inappropriate. Therefore, we focus on demonstrating MAGI's long-range prediction capability in these specific scenarios to showcase its potential for capturing long-range temporal dependencies and motivate further investigation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7380" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 240 + ], + "type": "image", + "image_path": "a054ebd7b67da3fe5235072f683b4bf369a0574a3c62e702d0b0f9d0c01259f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "type": "text", + "content": "Figure 6. Long-term Video Prediction. MAGI predicts over 100 frames from a single input frame, maintaining reasonable motion even when trained on only 16 frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 141, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 141, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 141, + 290 + ], + "type": "text", + "content": "6. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 299, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 299, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 299, + 295, + 539 + ], + "type": "text", + "content": "Autoregressive Visual Generation Autoregressive language modeling [28, 29] has propelled the development of visual content generation using discrete visual tokens [41]. In this framework, pre-trained visual tokenizers like VQVAE [41] map visual patches into a discrete latent space, allowing visual generation to be approached similarly to language modeling. Early works such as DALL-E [30] focus on text-to-image generation by learning a joint distribution between text and discrete image representations using an autoregressive cross-entropy loss. Concurrently, VideoGPT [50] extends this idea to video generation, employing discrete tokens for autoregressive video prediction. VideoPoet [20] further advances this approach by integrating a causal video tokenizer, MAGVIT-v2 [52], for multimodal video generation. OmniTokenizer [45] proposes a unified tokenizer for both discrete and continuous representations, enabling patch-level autoregressive video generation. In contrast, our work focuses on frame-level causality rather than patch-level, avoiding the limitations of raster-scan order and outperforming patch-level baselines." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Masked and Diffusion Video Generation Diffusion models have recently gained prominence in visual generation tasks [7, 11, 14, 16, 31], extending effectively to video generation. Video diffusion models [1, 17] employ bidirectional attention and binary mask embeddings to enable frame-level autoregressive prediction. Notable works such as GameNGen [40] use bidirectional diffusion models for real-time game generation. However, due to their bidirectional nature, these models cannot leverage KV Cache for extended video generation, limiting their scalability. Several masked video generators, such as Genie [2], extend MaskGIT [4] into a causal-attention-based architecture for video generation. Despite their advantages, these" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 279, + 554, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 352 + ], + "type": "text", + "content": "methods suffer from the training-inference gap inherent in masked autoregressive modeling, which negatively impacts generation quality. In contrast, our approach fully leverages KV Cache during inference, facilitated by our training paradigm that bridges the training-inference gap through a novel Complete Teacher Forcing paradigm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 361, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 504 + ], + "type": "text", + "content": "Addressing Exposure Bias for Autoregressive Video Generation Autoregressive models often suffer from exposure bias [55] and error accumulation [21], which degrade performance over long sequences. To mitigate these issues, we adopt noise injection from GameN-Gen [40]—where Gaussian noise is added to observation frame latents during training. Additionally, we introduce dynamic interval training, exposing the model to frames with varying temporal intervals to enhance generalization. As shown in Fig. 4, both strategies individually improve robustness, and their combination yields even greater performance gains in autoregressive video generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "We present MAGI, a hybrid framework that combines masked and causal modeling to achieve efficient and scalable video generation. We identify that the teacher forcing paradigm makes a significant difference—our Complete Teacher Forcing (CTF) approach effectively bridges the training-inference gap inherent in Masked Teacher Forcing (MTF). Additionally, we introduce essential training strategies to alleviate exposure bias. Through comprehensive experiments, we validate the effectiveness of each component. Our final model achieves state-of-the-art performance among autoregressive video generation methods. We also demonstrate the generation of long, coherent video sequences exceeding 100 frames from training sequences as short as 16 frames, highlighting MAGI's potential as a scalable autoregressive video generation model." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "text", + "content": "7381" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "We would like to thank Tianhong Li for his insightful suggestions, invaluable help, and exceptional support. We would also like to thank Chenfei Wu, Haoyang Huang, Guoqing Ma, Hongyu Zhou, Liangyu Chen, Chunrui Han, Yimin Jiang and Yu Deng for their constructive discussions and advice, which greatly improved this work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 193, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 296, + 247 + ], + "type": "text", + "content": "[1] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 248, + 296, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 248, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 296, + 303 + ], + "type": "text", + "content": "[2] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 295, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 295, + 336 + ], + "type": "text", + "content": "[3] Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808.01340, 2018. 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 337, + 296, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 337, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 337, + 296, + 390 + ], + "type": "text", + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022, 1, 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 296, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 457 + ], + "type": "text", + "content": "[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T. Freeman, Michael Rubinstein, Yuanzhen Li, and Dilip Krishnan. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 502 + ], + "type": "text", + "content": "[6] Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. arXiv preprint arXiv:2407.01392, 2024. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 503, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 296, + 557 + ], + "type": "text", + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "type": "text", + "content": "[8] Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 614, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 678 + ], + "type": "text", + "content": "[9] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive KV cache compression for llms. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[10] Anirudh Goyal, Alex Lamb, Ying Zhang, Saizheng Zhang, Aaron C. Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 4601-4609, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 171 + ], + "type": "text", + "content": "[11] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 172, + 554, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 172, + 554, + 225 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 554, + 225 + ], + "type": "text", + "content": "[12] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 193, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15979-15988. IEEE, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 293, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 337 + ], + "type": "text", + "content": "[14] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2(3):4, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 359 + ], + "type": "text", + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5, 8, 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "type": "text", + "content": "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 469 + ], + "type": "text", + "content": "[18] Allan Jabri, David Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. arXiv preprint arXiv:2212.11972, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "text", + "content": "[19] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 13" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "text", + "content": "[20] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 1, 2, 8, 12" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "text", + "content": "[21] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Conference on robot learning, pages 143-156. PMLR, 2017. 4, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 614, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 667 + ], + "type": "text", + "content": "[22] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 5, 12, 13" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[23] Shi Luohe, Hongyi Zhang, Yao Yao, Zuchao Li, and hai zhao. Keep the cost down: A review on methods to optimize LLM's KV-cache consumption. In First Conference on Language Modeling, 2024. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7382" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[24] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[25] Shentong Mo and Yapeng Tian. Scaling diffusion mamba with bidirectional ssms for efficient image and video generation. arXiv preprint arXiv:2405.15881, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[26] NVIDIA. Cosmos tokenizer. https://github.com/NVIDIA/Cosmos-Tokenizer, 2024. Accessed: 2024-11-15. 5, 6, 7, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "text", + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 263 + ], + "type": "text", + "content": "[28] Alec Radford. Improving language understanding by generative pre-training. 2018. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[29] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[30] A. Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 409 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 410, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 465 + ], + "type": "text", + "content": "[32] Florian Schmidt. Generalization in generation: A closer look at exposure bias. In Proceedings of the 3rd Workshop on Neural Generation and Translation@EMNLP-IJCNLP 2019, Hong Kong, November 4, 2019, pages 157-167. Association for Computational Linguistics, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 510 + ], + "type": "text", + "content": "[33] Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Mostgan-v: Video generation with temporal motion styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5652-5661, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 566 + ], + "type": "text", + "content": "[34] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3626-3636, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "text", + "content": "[35] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[36] Yao Teng, Yue Wu, Han Shi, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "text", + "content": "[37] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "text", + "content": "[38] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[39] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[40] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines. arXiv preprint arXiv:2408.14837, 2024. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "text", + "content": "[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 242 + ], + "type": "text", + "content": "[42] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "text", + "content": "[43] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[44] Jacob Walker, Ali Razavi, and Aäron van den Oord. Predicting video with vqvae. arXiv preprint arXiv:2103.01950, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[45] Junke Wang, Yi Jiang, Zehuan Yuan, Binyue Peng, Zuxuan Wu, and Yu-Gang Jiang. Omnitokensizer: A joint image-video tokenizer for visual generation. arXiv preprint arXiv:2406.09399, 2024. 2, 5, 6, 7, 8, 12" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "type": "text", + "content": "[46] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "text", + "content": "[47] Ronald J. Williams and David Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural Computation, 1(2):270-280, 1989. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "text", + "content": "[48] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In European conference on computer vision, pages 720-736. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 523, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 589 + ], + "type": "text", + "content": "[49] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, et al. Sana: Efficient high-resolution text-to-image synthesis with linear diffusion transformers. In The Thirteenth International Conference on Learning Representations, 2025. 13" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "text", + "content": "[50] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 1, 2, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 689 + ], + "type": "text", + "content": "[51] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 1, 2, 7, 12" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "text", + "content": "[52] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7383" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 251 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 7, 8, 12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[53] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 205 + ], + "type": "text", + "content": "[54] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "text", + "content": "[55] Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. Bridging the gap between training and inference for neural machine translation. arXiv preprint arXiv:1906.02448, 2019. 4, 8" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "7384" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_content_list.json b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..68f3b0aa9033c14de5f0c7d438a05e759c8f9923 --- /dev/null +++ b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_content_list.json @@ -0,0 +1,1748 @@ +[ + { + "type": "text", + "text": "Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs", + "text_level": 1, + "bbox": [ + 181, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingji Zhong $^{1}$ Zhihao Li $^{2}$ Dave Zhenyu Chen $^{2}$ Lanqing Hong $^{2}$ Dan Xu $^{1}$", + "bbox": [ + 171, + 202, + 821, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Hong Kong University of Science and Technology 2Huawei Noah's Ark Lab", + "bbox": [ + 169, + 227, + 828, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yzhongbn,danxu}@cse.ust.hk, {zhihao.li,dave.zhenyuchen,honglanqing}@huawei.com", + "bbox": [ + 148, + 251, + 844, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9116717b497687e07d9578d6803f616b891debbc78b9a4b8dcf892ca67404706.jpg", + "image_caption": [ + "Figure 1. We tackle the critical issues of (a) extrapolation and (b) occlusion in sparse-input 3DGS by leveraging a video diffusion model. Vanilla generation often suffers from inconsistencies within the generated sequences (as highlighted by the yellow arrows), leading to black shadows in the rendered images. In contrast, our scene-grounding generation produces consistent sequences, effectively addressing these issues and enhancing overall quality (c), as indicated by the blue boxes. The numbers refer to PSNR values. Zoom in for better visualization." + ], + "image_footnote": [], + "bbox": [ + 99, + 285, + 890, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 638, + 326, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent successes in novel view synthesis using 3D Gaussian Splatting (3DGS), modeling scenes with sparse inputs remains a challenge. In this work, we address two critical yet overlooked issues in real-world sparse-input modeling: extrapolation and occlusion. To tackle these issues, we propose to use a reconstruction by generation pipeline that leverages learned priors from video diffusion models to provide plausible interpretations for regions outside the field of view or occluded. However, the generated sequences exhibit inconsistencies that do not fully benefit subsequent 3DGS modeling. To address the challenge of inconsistencies, we introduce a novel scene-grounding guidance based on rendered sequences from an optimized 3DGS, which tames the diffusion model to generate consistent sequences. This guidance is training-free and does not require any fine-tuning of the diffusion model. To facilitate", + "bbox": [ + 88, + 659, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "holistic scene modeling, we also propose a trajectory initialization method. It effectively identifies regions that are outside the field of view and occluded. We further design a scheme tailored for 3DGS optimization with generated sequences. Experiments demonstrate that our method significantly improves upon the baseline and achieves state-of-the-art performance on challenging benchmarks.", + "bbox": [ + 511, + 640, + 906, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 767, + 645, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in 3D scene representation such as Neural Radiance Fields (NeRF) [2-4, 29, 31, 43] have greatly boosted the performance of Novel View Synthesis (NVS). NeRF represents the scene with a Multi-Layer Perceptron (MLP) and renders high-fidelity images with volumetric rendering. More recently, 3D Gaussian Splatt-ing (3DGS) [16, 25, 26, 58] emerges as a powerful explicit", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6133", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representation that models the scene with a set of 3D gaussian primitives and renders images via differentiable splatting. 3DGS achieves comparable performance to NeRF while requiring significantly less training time and offering higher inference speeds.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite recent advances in scene representations based on 3DGS, modeling scenes with sparse inputs remains a significant challenge. The sparse supervision often leads radiance fields to learn degenerate representations due to shape-radiance ambiguity [59], regardless of whether the representation is NeRF or 3DGS. While there have been promising improvements [18, 32, 44, 63], the commonly used faceforwarding [14, 28] and object-oriented 'outside-in' viewing [29] settings oversimplify real-world sparse-input modeling, causing many methods to overlook two critical issues: (i) extrapolation - while the sparse inputs typically cover the scene as much as possible, there may still exist regions that are outside the field of view, as shown in Fig. 1 (a); (ii) occlusion - occlusion frequently occurs for novel views that deviate even slightly from the training input views, as illustrated in Fig. 1 (b). When rendering with an optimized 3DGS, these issues can cause severe artifacts, such as black holes, significantly degrading image quality. Therefore, handling these two issues, i.e., extrapolation and occlusion, is critical for real-world sparse-input modeling.", + "bbox": [ + 91, + 170, + 482, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above-discussed issues, we propose a novel reconstruction by generation pipeline based on 3DGS. Intuitively, we use video diffusion models [7, 12, 51, 57] to generate multi-view sequences, which provide plausible interpretations of the scene based on priors learned from large-scale datasets. These sequences significantly enlarge the viewing instances, offering a high potential to address the extrapolation and occlusion issues. With the sparse inputs and the generated sequences, we can optimize a 3DGS to model the scene. However, as shown in Fig. 1, this vanilla pipeline brings little improvement or may even degrade the performance. The main reason can be attributed to the multi-view inconsistency within the generated sequences. The inconsistency manifests in two aspects: (i) the appearance inconsistency between frames within a sequence; (ii) the generated sequence may contain hallucinated elements not present in the scene.", + "bbox": [ + 89, + 474, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fully leverage the learned prior from video diffusion models for sparse-input 3DGS, we further explore addressing the challenges of inconsistencies within the generated sequences. Unlike existing methods that resolve appearance inconsistencies by assigning per-frame learnable appearance embeddings [21, 27], we focus on taming video diffusion models to directly generate sequences with consistency. Inspired by training-free guidance methods for diffusion models [1, 38, 53, 56] that enable controllable generation through external guidance, we introduce a novel strategy called scene-grounding guidance to ensure consis", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tent generation without requiring further fine-tuning of the diffusion models. Specifically, the scene-grounding guidance is based on a rendered sequence from an optimized 3DGS. During each step of the denoising process, the noisy sequence receives gradients from the supervision of the rendered sequence. Although the rendered sequence does not provide perfect guidance, our key insight in employing it to address the inconsistency is twofold: (i) adjacent frames within the rendered sequence are highly consistent due to limited camera movement between them; (ii) the rendered sequence is scene-grounding, which can guide the diffusion model to avoid generating elements that do not exist in the scene. In addition to addressing the issues of extrapolation and occlusion, our method also enhances the overall quality of the rendered images, as demonstrated in Fig. 1 (c). To effectively identify regions that are outside the field of view or occluded, we propose a trajectory initialization strategy to determine the camera trajectory during sequence generation, which is also based on an optimized 3DGS. With the proposed method, we can perform a holistic modeling of the scene. Additionally, we introduce a scheme for optimizing 3DGS with generated sequences, focusing on loss and sampling designs, which further enhance overall performance. We evaluate our method on two challenging in- 108 door datasets, Replica [40] and ScanNet++ [54], where the issues of extrapolation and occlusion are pronounced. The experiments demonstrate that our method achieves notable improvements and establishes state-of-the-art performance. Our contributions are summarized as:", + "bbox": [ + 511, + 90, + 903, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- This paper is the first to explicitly address the challenges of extrapolation and occlusion in 3DGS modeling from sparse inputs.", + "- We propose a novel reconstruction by generation pipeline with a designed scene-grounding guidance, which tames the video diffusion models to generate consistent and plausible sequences to effectively tackle the issues of extrapolation and occlusion.", + "- We present a trajectory initialization strategy that effectively identifies regions that are outside the field of view and occluded, facilitating holistic scene modeling. We also introduce a scheme for optimizing 3DGS with generated sequences, further improving the performance.", + "- Our method demonstrates significant improvements over the baseline, achieving over $3.5\\mathrm{dB}$ and $2.5\\mathrm{dB}$ PSNR enhancements on the Replica [40] and ScanNet++ [54] datasets, respectively, thereby establishing state-of-the-art performance." + ], + "bbox": [ + 514, + 529, + 903, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 814, + 661, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Radiance Fields from Sparse Inputs. Although improvements have been made in scene representation [2, 3, 16, 25, 29, 31, 58], learning a robust radiance field typically requires dense inputs as supervision due to the radiance-shape", + "bbox": [ + 511, + 839, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6134", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1adff7a35be3b3e9be90abc25e4366e13fdfa15b87e931dbf42f2322f8656915.jpg", + "image_caption": [ + "Figure 2. Framework overview of our proposed method. It consists of three parts: scene-grounding guidance, trajectory initialization, and optimization scheme with generated sequences. Initially, a baseline 3DGS is trained using sparse inputs and initialized with the point cloud from DUSt3R [46]. Yellow regions denote uncovered areas, e.g., those outside the field of view or occluded. The trajectory initialization determines the paths for sequence generation based on renderings from the baseline 3DGS, facilitating holistic scene modeling. The video diffusion model receives an input image along with the trajectory for sequence generation, incorporating scene-grounding guidance during the denoising process to ensure consistent output. The guidance is based on the rendered sequences. Finally, the generated sequences are utilized to optimize the final 3DGS through a tailored optimization scheme." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 883, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ambiguity [59]. Current works can be roughly categorized into two lines of research. The first line of research focuses on pre-training generalizable radiance fields using multiview datasets. Generalizable NeRF [6, 24, 45, 55] typically projects the ray points to reference images and integrates 2D features as the auxiliary input of MLP. In contrast, generalizable 3DGS [5, 8, 22, 60] commonly perform dense depth prediction associated with gaussian properties. The second line of research leverages regularization techniques for optimization, which are applicable to both NeRF and 3DGS, including hand-crafted constraints [17, 32, 63], and those derived from pre-trained models [10, 18, 33, 35, 44, 50, 64]. In this paper, we explore the priors from video diffusion models [7, 12, 51, 57] for sparse-input 3DGS modeling. Unlike previous works, our method trains sparse-input 3DGS using augmented sequences from video diffusion models, which provide interpolation or extrapolation around the input views. We design a guidance strategy that tames the video diffusion model to generate more plausible and scene-grounded sequences, greatly enhancing the performance.", + "bbox": [ + 91, + 455, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Prior for Radiance Fields. Diffusion models [11, 36, 39] have shown remarkable generation capabilities. The strong prior knowledge embedded in diffusion models can facilitate the training of radiance fields. Specifically, several works leverage Score Distillation Sampling (SDS) [19, 20, 34, 48] using a frozen diffusion model to train a radiance field based on text prompts in a zero-shot manner. Other studies focus on training view-consistent [37, 42, 49] or quality-enhanced [23] diffusion", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "models, where the generated results can be directly applied to train a radiance field. Our method also uses the generated results to train a radiance field. However, our primary contribution lies in a novel guidance strategy that controls the generation to be consistent, which is crucial for sparse-input modeling augmented with generation.", + "bbox": [ + 511, + 455, + 906, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable Generation for Diffusion Models. Current works of controllable generation can be categorized into training-required and training-free methods. Training-required methods fine-tune the diffusion models with additional conditions [30, 61], or train an additional noise-conditioned external guidance function, e.g., classifier guidance [9] for denosing sampler. Training-free methods freeze the foundation diffusion model, and modify the denoising process with the control signal from the external guidance functions [1, 38, 53, 56]. These methods do not require training additional guidance functions or fine-tuning diffusion models; instead, they enable controllable generation in a plug-and-play manner. Our work is inspired by training-free methods. However, we concentrate on multi-view modeling, which necessitates high consistency control over the generated results. Besides, unlike these methods that typically rely on pre-trained models for guidance, we utilize rendered results to provide scene-grounding guidance.", + "bbox": [ + 511, + 546, + 908, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 829, + 723, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we utilize video diffusion models to tackle two critical issues in real-world sparse-input modeling: extrapolation and occlusion, as illustrated in Fig. 1. The overview", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6135", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of our method is illustrated in Fig. 2, which consists of three proposed components: a scene-grounding guidance (Sec. 3.2), a trajectory initialization strategy (Sec. 3.3), and a scheme for 3DGS optimization with generated sequences (Sec. 3.4). We will detail these components following a preliminary review of 3DGS and diffusion models.", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 194, + 220, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Gaussian Splating (3DGS) [16] represents a scene with a set of anisotropic 3D Gaussian primitives. Each Gaussian primitive is parametrized by a set of attributes: a center $\\mu$ , a scaling factor $s$ , a quaternion $q$ , an opacity value $\\alpha$ , and a feature vector $f$ . The basis function of each Gaussian primitive is formulated as $\\mathcal{G}(x) = \\exp(-\\frac{1}{2} (\\mathrm{x} - \\mu)^{\\mathrm{T}} \\Sigma^{-1} (\\mathrm{x} - \\mu))$ , where $\\Sigma$ is the covariance matrix derived from $s$ and $q$ . 3DGS renders the scene through a differentiable splatting, which firstly transforms the 3D Gaussian $\\mathcal{G}(x)$ into 2D Gaussian $\\mathcal{G}'(x)$ on the image plane via projection [65], and applies a tile-based rasterizer for rendering, which sorts the 2D Gaussians by depth and employs the $\\alpha$ -blending as follows:", + "bbox": [ + 89, + 217, + 483, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC \\left(x _ {p}\\right) = \\sum_ {i \\in K} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\quad \\sigma_ {i} = \\alpha_ {i} \\mathcal {G} _ {i} ^ {\\prime} \\left(x _ {p}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 414, + 482, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{p}$ is the pixel position, $K$ refers to the number of 2D Gaussians associated with the pixel, and $c$ represents the decoded color of feature $f$ .", + "bbox": [ + 89, + 455, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Models [11, 39] are a family of generative models that progressively perturb data with intensifying Gaussian noises (i.e., forward noising), and then learn to reverse this process for sample generation (i.e., reverse denoising). The key of the diffusion model is a U-Net $\\epsilon_{\\theta}$ which is trained to predict the noise that is injected in the current sample $\\mathbf{x}_t$ . The sampling is conducted by iterative denoising for $T$ steps [56] as follows:", + "bbox": [ + 89, + 501, + 483, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t - 1} = (1 + \\beta_ {t} / 2) \\mathbf {x} _ {t} + \\beta_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) + \\sqrt {\\beta_ {t}} \\mathbf {z} \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 628, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t)$ is the estimated score function which can be derived from $\\epsilon_{\\theta}(\\mathbf{x}_t,t);\\beta_t$ is pre-defined parameters; $\\mathbf{z}\\sim \\mathcal{N}(0,\\mathbf{I})$ . In this work, we leverage a camera-controlled image-to-video diffusion model [57], whose condition includes an image for the first frame, and the camera trajectory for the path of the generated sequence. The model is operated in a latent space of dimension $d$ , supporting the sequence length of $L$ , thus $\\mathbf{x}_t\\in \\mathbb{R}^{L\\times h\\times w\\times d}$ .", + "bbox": [ + 89, + 652, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Generation via Scene-Grounding Guidance", + "text_level": 1, + "bbox": [ + 89, + 787, + 457, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Applying the generated sequences from the video diffusion model can provide plausible interpretations of regions not covered by the sparse inputs. However, as illustrated in Fig. 1, the inconsistency within the generated sequences manifests as: (i) appearance inconsistencies across frames and (ii) the occurrence of non-existent elements, which can", + "bbox": [ + 89, + 810, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Generation with Scene-Grounding Guidance", + "bbox": [ + 514, + 93, + 898, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: Function GENERATOR $(\\mathcal{R},I,\\{\\phi_j\\}_{j = 1}^L)$", + "bbox": [ + 521, + 113, + 813, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2: Input: Optimized 3DGS model $\\mathcal{R}$ , input image $I$ , camera trajectory of a sequence $\\{\\phi_j\\}_{j=1}^L$ .", + "bbox": [ + 524, + 131, + 903, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3: Given: Latent image-to-video diffusion model $\\epsilon_{\\theta}$ , VAE decoder $\\mathcal{D}$ , pre-defined $\\beta_{t}$ , $\\bar{\\alpha}_{t}$ and guidance scale $\\gamma_{t}$ .", + "bbox": [ + 524, + 160, + 903, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4: Abbreviate $\\epsilon_{\\theta}(\\mathbf{x}_t,t,I,\\{\\phi_j\\}_{j = 1}^L)$ as $\\epsilon_{\\theta}(\\mathbf{x}_t,t)$", + "bbox": [ + 524, + 190, + 838, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5: $\\mathbf{S},\\mathbf{M} =$ rasterize $(\\{\\phi_j\\}_{j = 1}^L,\\mathcal{R})$ $\\triangleright$ Eq. (1)& (5)", + "bbox": [ + 524, + 207, + 900, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6: $\\mathbf{x}_T\\sim \\mathcal{N}(0,\\mathbf{I})$", + "bbox": [ + 524, + 223, + 637, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "7: for $t = T, \\dots, 1$ do", + "bbox": [ + 524, + 238, + 674, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "8: $\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})$ if $t > 1$ , else $\\mathbf{z} = \\mathbf{0}$", + "bbox": [ + 524, + 252, + 772, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "9: $\\hat{\\mathbf{x}}_{t - 1} = (1 + \\frac{1}{2}\\beta_t)\\mathbf{x}_t - \\frac{\\beta_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\epsilon_\\theta (\\mathbf{x}_t,t) + \\sqrt{\\beta_t}\\mathbf{z}$", + "bbox": [ + 524, + 267, + 875, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "10: $\\mathbf{x}_{0|t} = \\frac{1}{\\sqrt{\\bar{\\alpha}_t}} (\\mathbf{x}_t - \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon_\\theta (\\mathbf{x}_t,t))$", + "bbox": [ + 519, + 286, + 797, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "11: $\\mathbf{X}_{0|t} = \\mathcal{D}(\\mathbf{x}_{0|t})$", + "bbox": [ + 519, + 305, + 663, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "12: $\\mathbf{g}_t = \\nabla_{\\mathbf{x}_t}\\mathcal{L}(\\mathbf{S},\\mathbf{M},\\mathbf{X}_{0|t})$ $\\triangleright$ Eq. (6)", + "bbox": [ + 519, + 319, + 900, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "13: $\\mathbf{x}_{t - 1} = \\hat{\\mathbf{x}}_{t - 1} - \\gamma_t\\mathbf{g}_t$ $\\triangleright$ Eq. (2)& (4)", + "bbox": [ + 519, + 335, + 903, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "14: end for", + "bbox": [ + 519, + 349, + 596, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "15: return $\\mathcal{D}(\\mathbf{x}_0)$", + "bbox": [ + 519, + 364, + 640, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "negatively impact the 3DGS modeling. In this section, we propose an innovative scene-grounding guidance method that directs the video diffusion model to generate consistent sequences, significantly enhancing the performance of sparse-input 3DGS.", + "bbox": [ + 511, + 397, + 906, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by previous training-free guidance methods [1, 56] that achieve their objectives by modifying the sampler in Eq. (2), we adopt a similar approach to attain the goal of consistency. Specifically, we firstly replace $\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t)$ with a conditional score function $\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t|\\mathcal{Q})$ , where $\\mathcal{Q}$ refers to the target of consistency. The conditional score function can be expanded by the Bayesian rule as:", + "bbox": [ + 511, + 472, + 905, + 577 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t} | \\mathcal {Q}) = \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) + \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathcal {Q} | \\mathbf {x} _ {t}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 584, + 903, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\nabla_{\\mathbf{x}_t}\\log p(\\mathcal{Q}|\\mathbf{x}_t)$ can be considered as a guidance term that injects the consistency constraint into Eq. (2). We further formulate $p(\\mathcal{Q}|\\mathbf{x}_t)$ as: $p(\\mathcal{Q}|\\mathbf{x}_t) = \\exp (-\\lambda \\mathcal{L}(\\mathcal{Q},\\mathbf{x}_t)) / Z$ where $\\mathcal{L}(\\mathcal{Q},\\mathbf{x}_t)$ indicates how well the current sample $\\mathbf{x}_t$ is aligned with the target, and $Z$ is a normalization term. The guidance term can thus be implemented using the gradient of the following loss function:", + "bbox": [ + 511, + 606, + 905, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathcal {Q} | \\mathbf {x} _ {t}) \\propto - \\nabla_ {\\mathbf {x} _ {t}} \\mathcal {L} (\\mathcal {Q}, \\mathbf {x} _ {t}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 727, + 903, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is appended to Eq. (2) to achieve the target of consistency during the denoising sampling.", + "bbox": [ + 511, + 750, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The remaining problem lies in how to define the consistency target $\\mathcal{Q}$ . Unlike previous works [1, 38, 56] that define the target based on external pre-trained models, we establish the target using a rendered sequence from an optimized 3DGS model $\\mathcal{R}$ . Though the rendered sequence is not perfect, our key insights are as follows: (i) the rendered images of adjacent frames are highly consistent, as the camera movement between them is typically minor; (ii)", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6136", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the rendered frames provide scene grounding, clearly indicating which elements exist in the scene. Therefore, the rendered sequence can serve as an effective guidance for the generated sequence to achieve the target of consistency.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a camera trajectory $\\{\\phi_j\\}_{j = 1}^L$ for the sequence generation, we first utilize the optimized 3DGS to render a sequence $\\{S_j\\}_{j = 1}^L$ , along with a mask sequence $\\{M_j\\}_{j = 1}^L$ that indicates the regions not covered by the sparse inputs. To get the mask, we first render a transmittance map, which is obtained by $\\alpha$ -blending (as Eq. (1)) on the opacity. For each pixel $x_{p}$ , the $\\alpha$ -blending is formulated as:", + "bbox": [ + 89, + 152, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO \\left(x _ {p}\\right) = \\sum_ {i \\in K} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 262, + 482, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ and $O$ refer to the opacity of the gaussian and the transmittance map, respectively. The mask is then acquired by thresholding the transmittance map with a value $\\eta_{\\mathrm{mask}}$ : $M = (O < \\eta_{\\mathrm{mask}})$ . For convenience, we stack $\\{S_j\\}_{j=1}^L$ and $\\{M_j\\}_{j=1}^L$ to $\\mathbf{S}$ and $\\mathbf{M}$ , which are of shape $L \\times H \\times W \\times 3$ and $L \\times H \\times W \\times 1$ , respectively. Since the target of consistency is based on the rendered sequence in clean data space, to receive the guidance, we transform the noisy latent $\\mathbf{x}_t$ into a latent $\\mathbf{x}_{0|t}$ in the clean data space, based on prediction from the model $\\epsilon_\\theta$ : $\\mathbf{x}_{0|t} = (\\mathbf{x}_t - \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_\\theta(\\mathbf{x}_t, t)) / \\sqrt{\\bar{\\alpha}_t}$ . With the consistency target $\\mathcal{Q}$ that is based on the rendered sequence $\\mathbf{S}$ , we formulate the function $\\mathcal{L}$ in the guidance term (Eq. (4)) as:", + "bbox": [ + 89, + 304, + 483, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\mathbf {S}, \\mathbf {M}, \\mathbf {X} _ {0 | t}) = \\| \\bar {\\mathbf {M}} \\odot (\\mathbf {S} - \\mathbf {X} _ {0 | t}) \\| _ {1} \\tag {6} \\\\ + \\lambda_ {\\text {p e r c}} \\mathcal {L} _ {\\text {p e r c}} (\\bar {\\mathbf {M}} \\odot \\mathbf {S}, \\bar {\\mathbf {M}} \\odot \\mathbf {X} _ {0 | t}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 503, + 482, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_{0|t}$ is decoded from the latent $\\mathbf{x}_{0|t}$ by a VAE decoder, $\\bar{\\mathbf{M}}$ is the result of flipping the values in $\\mathbf{M}$ , $\\odot$ is the Hadamard product, and $\\mathcal{L}_{\\mathrm{perc}}$ is a perceptual loss [15] with its corresponding weight as $\\lambda_{\\mathrm{perc}}$ .", + "bbox": [ + 89, + 545, + 483, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the guidance from Eq. (6), the denoising process balances the consistency constraint and the prior from the diffusion model, integrating them into plausible generation results. This guidance does not involve any fine-tuning of the diffusion model, thereby preserving its generative capabilities. The detailed pipeline is outlined in Alg. 1.", + "bbox": [ + 89, + 607, + 483, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Trajectory Initialization Strategy", + "text_level": 1, + "bbox": [ + 89, + 710, + 380, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable holistic modeling of the scene, the camera trajectories for the video diffusion model should cover regions that are outside the field of view or occluded as much as possible. The generated sequences can thus provide plausible interpretations for these regions, which serve as the basis for optimizing the subsequent 3DGS model. Similar to the scene-grounding guidance discussed in Sec. 3.2, the proposed trajectory initialization method is also based on an optimized 3DGS model. For the $i$ -th sparse input view with camera pose $\\varphi_{i}$ , we first sample a set of candidate poses around it, as depicted in Fig. 3. Suppose", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1ef091fba6b0377cc0b70a326fbc2fcf34cfcd09333ecee5591c04b445a6485.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed trajectory initialization strategy. The yellow parts represent unobserved regions. For each input view, we sample a set of candidate poses around it, and render at these poses using an optimized 3DGS. We select candidate poses whose renderings exhibit significant holes (highlighted by red boxes), and interpolate trajectories between these candidate poses and the input view's pose." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 898, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "there are a total of $m$ candidate poses, we use the optimized 3DGS model $\\mathcal{R}$ to render images for these poses as $\\{\\hat{S}_c^{(i)},\\hat{M}_c^{(i)}\\}_{c = 1}^m =$ rasterize $(\\{\\hat{\\phi}_c^{(i)}\\}_{c = 1}^m,\\mathcal{R})$ . For poses where the rendered images exhibit significant black holes, as indicated by the mask $\\hat{M}$ calculated from Eq. (5), we interpolate a trajectory of length $L$ (matching the length of video diffusion model) between the input camera pose and these poses as follows: $\\{\\phi_j^{(i,c)}\\}_{j = 1}^L = \\mathrm{interp}(\\varphi_i,\\hat{\\phi}_c^{(i)})$ where $\\hat{\\phi}_c^{(i)}$ refers to one selected candidate pose from the $i$ -th input. In practice, we select the top- $k$ candidate poses based on the sizes of their corresponding masks. Then, we build a trajectory pool by traversing all input views and their respective selected candidate poses as:", + "bbox": [ + 511, + 407, + 906, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi = \\left\\{\\left\\{\\phi_ {j} ^ {(i, c)} \\right\\} _ {j = 1} ^ {L} \\mid i, c \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 617, + 903, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where each element in the pool is sampled for the sequence generation.", + "bbox": [ + 511, + 640, + 905, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. 3DGS Optimization with Generation", + "text_level": 1, + "bbox": [ + 511, + 679, + 831, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given sparse inputs of $N$ images along with their poses, i.e., $\\{C_i^{\\mathrm{gt}},\\varphi_i\\}_{i = 1}^N$ , we aim at optimizing a 3DGS model with the auxiliary generated sequences. For simplicity, we refer to the input images paired with their poses as 'input views', and we term the generated images with their associated poses as 'generated views'. During each iteration, we sample an input view and a generated view for supervision. Specifically, for the input view, we employ the default reconstruction loss [16] written as:", + "bbox": [ + 511, + 700, + 905, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {i n p u t}} = (1 - \\lambda) \\mathcal {L} _ {1} \\left(C _ {i}, C _ {i} ^ {\\mathrm {g t}}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(C _ {i}, C _ {i} ^ {\\mathrm {g t}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 844, + 903, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C_i$ refers to the rendered image and $\\lambda$ is a weighting factor. For supervision of generated views, we find that the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6137", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 2 3DGS Optimization with Generation", + "bbox": [ + 91, + 92, + 424, + 107 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: Sparse inputs of N images $\\{C_i^{\\mathrm{gt}},\\varphi_i\\}_{i = 1}^N$", + "2: Given: Number of iterations $N_{\\mathrm{iter}}$ , generation interval $N_{\\mathrm{gen}}$ , ratio of samples from other sequences $\\eta$ .", + "3: Variable: Global list of generated views $\\mathbf{G} = []$", + "4: Baseline 3DGS model optimization $\\Rightarrow \\mathcal{R}$", + "5: Trajectory initialization $\\Rightarrow \\Phi$ $\\triangleright$ Eq. (7)", + "6: for $t = 0, \\dots, N_{\\mathrm{iter}} - 1$ do", + "7: If $t\\% N_{\\mathrm{gen}} = 0$ then", + "8: Sample an input view $I$", + "9: Sample a trajectory around $I$ from $\\Phi \\Rightarrow \\{\\phi_j\\}_{j=1}^L$", + "10: $\\mathbf{S} = \\mathrm{GENERATOR}(\\mathcal{R},I,\\{\\phi_j\\}_{j = 1}^L)$", + "11: Append S to G", + "12: End If", + "13: Sample an input view to get $\\mathcal{L}^{\\mathrm{input}}$ $\\triangleright$ Eq. (8)", + "14: If rand() $\\geq \\eta$ then", + "15: Sample a generated view from S", + "16: Else Sample a generated view from G", + "17: End If", + "18: Use the generated view to get $\\mathcal{L}^{\\mathrm{gen}}$ $\\triangleright$ Eq. (9)", + "19: $(\\mathcal{L}^{\\mathrm{input}} + \\mathcal{L}^{\\mathrm{gen}})$ .backward()", + "20: #Densification and opacity reset", + "21: end for" + ], + "bbox": [ + 99, + 114, + 480, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "reconstruction loss does not effectively fill the hole regions, and increasing its weight leads to performance degradation due to flaws in the generated images. To address this issue, we propose using perceptual loss [15]. The perceptual loss is calculated over the entire image, allowing those hole regions to significantly influence the gradients, thereby effectively driving the model to fill those holes. Thus, the loss on the generated views is formulated as:", + "bbox": [ + 89, + 463, + 483, + 583 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {g e n}} = \\lambda_ {\\text {g e n 1}} \\mathcal {L} _ {1} \\left(C _ {j}, S _ {j}\\right) + \\lambda_ {\\text {g e n 2}} \\mathcal {L} _ {\\text {p e r c}} \\left(C _ {j}, S _ {j}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 594, + 480, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S_{j}$ refers to the generated image, $\\lambda_{\\mathrm{gen1}}$ and $\\lambda_{\\mathrm{gen2}}$ are two balancing factors, respectively.", + "bbox": [ + 89, + 621, + 482, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We empirically find that conducting local sampling within a specific optimization interval, where a substantial portion of generated views is sampled from the same sequence of local regions, enhances visual quality. However, sampling exclusively from a single sequence can lead to a forgetting issue, where optimized information about holes in other regions becomes diluted. Therefore, within each interval of local sampling, we also include generated views from other sequences with a ratio $\\eta$ . The optimization pipeline is presented in Alg. 2.", + "bbox": [ + 89, + 652, + 483, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 809, + 220, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 89, + 833, + 287, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Metrics. We target addressing the issues of extrapolation and occlusion for sparse-input 3DGS scene modeling, which are overlooked by current benchmarks. To", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "w/ Vanilla Generation", + "bbox": [ + 522, + 89, + 637, + 99 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1537861a4c58d25b5858af5487478160c838e5ec1cb31f2e7d8a653ff32f50c9.jpg", + "image_caption": [ + "Figure 4. Sequences from the vanilla generation suffer from inconsistencies. A 3DGS model optimized with these sequences renders images with black shadows, highlighted by red boxes, while our method solves this issue with the scene-grounding guidance." + ], + "image_footnote": [], + "bbox": [ + 519, + 99, + 642, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "w/ Guided Generation", + "bbox": [ + 658, + 89, + 769, + 99 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3984260e981f383a62bb0a63b6b2b8a8aeec30db5060c0f6c27955d2f2e114c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 101, + 771, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ground Truth", + "bbox": [ + 800, + 89, + 870, + 98 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9aa4ee5f3f265f18b84669cd92f1f5fe9e3984d8821aaa6ea64696a00ef24ed5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 101, + 898, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "evaluate the effectiveness of our method, we conduct experiments on a new benchmark created from two indoor datasets, i.e., the synthetic Replica [40] and the realistic ScanNet++ [54], which consists of 6 and 4 scenes, respectively. Although the selected six input views for each scene can cover most regions, there are still areas outside the field of view. Moreover, the 'inside-out' viewing directions make occlusion common in this benchmark. More details can be found in the supplement. For quantitative comparisons, we report PSNR, SSIM [47], and LPIPS [62] scores.", + "bbox": [ + 511, + 241, + 903, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline. We train a baseline 3DGS model initialized with the point cloud from DUSt3R [46], incorporating the gaussian unpooling in FSGS [64], which makes the optimized model a strong baseline. Based on this we conduct experiments to verify the effectiveness of our method. The model is denoted as 'Baseline 3DGS' in the following.", + "bbox": [ + 511, + 393, + 903, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The baseline model described above serves as the model $\\mathcal{R}$ for scene-grounding guidance and trajectory initialization. For sequence generation, we employ the camera-controlled image-to-video diffusion model [57] which supports the generation of $L = 25$ frames. The weighting factors, $\\lambda$ , $\\lambda_{\\mathrm{perc}}$ , $\\lambda_{\\mathrm{gen1}}$ , and $\\lambda_{\\mathrm{gen2}}$ are set to $0.2$ , $10^{-4}$ , $0.1$ , and $0.01$ , respectively. The threshold of $\\eta_{\\mathrm{mask}}$ is set to $0.9$ , while $\\eta$ is set to $0.5$ . The generation interval $N_{\\mathrm{gen}}$ is set to $260$ and $N_{\\mathrm{iter}}$ is set to $10,000$ .", + "bbox": [ + 511, + 484, + 905, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 633, + 651, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on Replica. As shown in Tab. 1, our method achieves the highest performance on the Replica dataset, outperforming DNGaussian [18] and FSGS [64] by a significant margin of over $3.0\\mathrm{dB}$ in PSNR. Fig. 5 illustrates that our method effectively addresses occlusion and extrapolation, while other 3DGS-based methods struggle with these challenges. Additionally, their depth regularization often compromise thin structures, such as the flower in the vase in the second row. FreeNeRF [52] exhibits severe artifacts because it cannot effectively utilize the strong prior from the DUSt3R point cloud. Although FreeNeRF can fill hole regions through neighboring interpolation (e.g., the wall behind the chair in the first row and the ceiling in the second row), the results frequently exhibit blurring or artifacts.", + "bbox": [ + 511, + 657, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on ScanNet++. ScanNet++ is a dataset captured in realistic scenes, so it is more complicated and chal", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6138", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f90b6a651d05826be96b312818235e9c13bb555a2d8087df277def0cd44845e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 88, + 880, + 428 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/646482c860e183294032331487e2c29a18b1a4fe28ef7576fb92bb1550dc1a01.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodReplica [40]ScanNet++ [54]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Mip-NeRF [2]18.120.7070.39119.580.7550.389
InfoNeRF [17]13.070.5980.55214.540.6460.495
DietNeRF [13]18.990.6760.44419.760.7190.431
FreeNeRF [52]20.990.7650.32420.170.7560.368
3DGS†[16]22.800.8180.17921.410.8170.211
DNGaussian [18]17.630.7180.43519.010.7540.367
DNGaussian†[18]22.710.8210.18920.680.7880.281
FSGS [64]20.220.7600.30417.950.7300.373
FSGS†[64]22.990.8330.20521.230.8130.257
Ours26.350.8720.12223.890.8500.182
", + "bbox": [ + 96, + 479, + 483, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons on the Replica and ScanNet++ datasets. Including our approach, 3DGS-based methods marked with $\\updownarrow$ are initialized with the point cloud from DUSt3R [46].", + "bbox": [ + 89, + 648, + 482, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lenging than the synthetic Replica dataset. The results in Tab. 1 demonstrate that our method has a clear advantage over current approaches, surpassing FSGS by more than $2.5\\mathrm{dB}$ in PSNR. As depicted in Fig. 5, our method effectively addresses the extrapolation issue (e.g., the ceiling in the fourth row) and mitigates needle-like artifacts observed in the rendered images of DNGaussian [18] and FSGS [64] (the third row). Furthermore, the comparisons in the third row highlight our method's superiority in preserving finer details compared to all other methods.", + "bbox": [ + 89, + 695, + 483, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 852, + 256, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our technical contributions consist of three key components. We analyze their effects on the Replica dataset.", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8938fcdfba6d9110e773d45996333126be74c9a43c506768960b25c6d0947339.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons on the Replica and ScanNet++ datasets. All 3DGS-based methods are optimized using the initialized point cloud from DUSt3R [46]. Our method effectively addresses the issues of extrapolation and occlusion while preserving finer details and reducing artifacts. For better visualization, please zoom in on the results.", + "Figure 6. Our method not only effectively addresses extrapolation and occlusion (red boxes), improving the overall quality (blue boxes), but also predicts more plausible geometry." + ], + "image_footnote": [], + "bbox": [ + 516, + 470, + 906, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generation with Scene-Grounding Guidance. Optimizing a 3DGS with sequences from vanilla generation results in quality degradation. In Tab. 2 (a), while the full image metrics are enhanced due to slightly improved modeling at occluded regions, the visual quality degrades, as indicated by PSNR of observable regions dropping from $25.45\\mathrm{dB}$ to $25.00\\mathrm{dB}$ . This degradation is attributed to inconsistencies within generated sequences, which can result in black shadows in rendered images as illustrated in Fig. 4. In contrast, our scene-grounding guidance ensures that the generated sequences remain consistent, significantly enhancing the modeling capability in regions outside the field of view and occluded, while also improving the overall quality, evidenced by the 'w/ Guided Generation' results in Tab. 2 (a).", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6139", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c1133b71fc460ada0e2a7bed5e42bc8af506832f33cec42d376229107ea0a765.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(a)Gen.Guide.Traj.Full ImageObservable Regions
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Baseline 3DGS22.800.8180.17925.450.8600.129
w/ Vanilla Generation✓23.690.8400.16025.000.8700.119
w/ Guided Generation✓✓25.030.8520.13926.520.8810.101
w/ Guided Generation&Traj.✓✓✓25.580.8590.13826.530.8830.100
", + "bbox": [ + 101, + 89, + 620, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1ccbb7f1f0417432ea9a8f7d4f0c28c5258e374b550890e591e3353ef7a3a182.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(b)PSNR↑SSIM↑LPIPS↓
Baseline 3DGS22.800.8180.179
w/ Guided Generation&Traj.25.580.8590.138
w/ perceptual loss26.350.8720.122
w/o local sampling26.280.8710.127
w/o global list26.010.8670.122
", + "bbox": [ + 627, + 89, + 906, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c28a4bff8d684a5dcb1ef8a36a8b73af00d4928123255edeb3373fc5082eb13b.jpg", + "image_caption": [ + "Baseline 3DGS", + "w/ Guided Genration&Traj.", + "Figure 7. The perceptual loss for generated views greatly increases the modeling capability at hole regions." + ], + "image_footnote": [], + "bbox": [ + 94, + 241, + 220, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/556a5d5d14b0e3d3d91a798cc8d95e9552d71d192dd8fe3a9141006973072802.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 241, + 346, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83f0dd9468285be22122a5df6b9c290869192250138ec402ad5363aa239f19fd.jpg", + "image_caption": [ + "w/ Perceptual Loss" + ], + "image_footnote": [], + "bbox": [ + 351, + 241, + 475, + 314 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b1ff34b35f867a7e31a993273004cc69fc69115cf001bf307fdfb98c56593fa5.jpg", + "table_caption": [ + "Table 2. Ablation experiments on the Replica dataset. (a) Effectiveness of the proposed scene-grounding guidance (Guide.) for generation, and the trajectory initialization strategy (Traj.). (Gen.) indicates utilizing generated sequences for modeling. Metrics of observable regions mask out regions outside the field of view or occluded. (b) Effectiveness of the proposed scheme for 3DGS optimization." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS↓
Baseline 3DGS+LaMa [41]24.560.8330.167
Baseline 3DGS+SDS [34] with SDInpaint24.330.8440.153
Baseline 3DGS+SDS [34] with SDInpaint*25.150.8530.141
Ours26.350.8720.122
", + "bbox": [ + 109, + 348, + 460, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparisons with inpainting methods on the Replica dataset. * indicates the usage of our trajectory initialization.", + "bbox": [ + 89, + 416, + 482, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Trajectory Initialization Strategy. Tab. 2 (a) further demonstrates that the proposed trajectory initialization strategy significantly boosts the performance, notated as 'w/ Guided Generation&Traj'. The improvement mainly arises from enhanced modeling of the regions outside the field of view or the occluded areas, as the metrics of visible regions plateau while the overall image metrics improve by over 0.5 dB in PSNR. This indicates that the initialization strategy effectively identifies hole regions for holistic modeling.", + "bbox": [ + 88, + 459, + 482, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scheme for 3DGS Optimization with Generation. We verify the effectiveness of the proposed scheme in Tab. 2 (b). Specifically, the perceptual loss of Eq. (9) increases PSNR by over $0.5\\mathrm{dB}$ , which is crucial for the model to fill the hole regions, as shown in Fig. 7. We empirically find that local sampling brings improvement in Sec. 3.4. This is evidenced by the performance decrease of 'w/o local sampling', which randomly samples generated views from all generated sequences. Alg. 2 shows that we use a global list to avoid the forgetting problem, and its necessity is verified by an over $0.3\\mathrm{dB}$ PSNR drop observed with 'w/o global list'. Combining these contributions, our full model effectively addresses extrapolation and occlusion while enhancing overall image quality, meanwhile exhibiting much better geometry, as shown in Fig. 6.", + "bbox": [ + 89, + 595, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Further Comparisons with Inpainting Methods", + "text_level": 1, + "bbox": [ + 89, + 835, + 482, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extrapolation and occlusion can also be addressed using inpainting methods. We thus compare our approach with two inpainting-based methods. One method applies LaMa [41]", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/765d64f45012320363a4026bdc7ec0b3d41053160dd277e3b3f23f00e4a0da0b.jpg", + "image_caption": [ + "LaMa", + "Figure 8. Qualitative comparisons with other inpainting methods. Our approach not only produces more plausible appearances around the inpainting regions but also predicts more consistent geometries in fine-grained local areas." + ], + "image_footnote": [], + "bbox": [ + 517, + 241, + 640, + 385 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be06c768b96b4c9495a667ba7c86f35f83435b2414f25ad3a9408f6338da935f.jpg", + "image_caption": [ + "SDS with SDInpaint*" + ], + "image_footnote": [], + "bbox": [ + 645, + 241, + 769, + 385 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bed2b29a0e7c0318fee352dbcbf14a6bbfb44561648ba35990edd819de56304b.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 774, + 241, + 898, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inpainting on hole regions, while the other optimizes a 3DGS by Score Distillation Sampling (SDS) [34] based on a SDInpaint model [36]. We also incorporate our proposed trajectory initialization into the SDS method to enhance the optimization of inpainting regions. Results in Tab. 3 show that our method outperforms these two methods by more than $1.0\\mathrm{dB}$ in PSNR. Qualitative results in Fig. 8 indicate that, under certain conditions, the SDS-based method produces inpainted areas with strange appearances, while LaMa tends to create blurring artifacts interpolated from neighboring regions. Besides, Fig. 8 also shows that our method predicts better geometry in regions with local details. The inpainting results from our approach are more plausible due to the well-designed guidance, which effectively exploits prior knowledge from the diffusion model.", + "bbox": [ + 511, + 453, + 906, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 685, + 633, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have explored to address the critical issues of extrapolation and occlusion in sparse-input 3DGS modeling. We propose using video diffusion models that provide plausible interpretations for regions that are outside the field of view and occluded. To resolve inconsistencies within generated sequences, we introduce a novel scene-grounding guidance that controls the diffusion model to generate consistent sequences without any fine-tuning. Additionally, we propose a trajectory initialization strategy to enhance holistic modeling and develop a scheme for optimizing 3DGS with generated sequences. Extensive experiments validate our approach, demonstrating that it outperforms current methods by a significant margin.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6140", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research is supported in part by the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, and Tencent Rhino-Bird Focused Research Program.", + "bbox": [ + 89, + 90, + 485, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 195, + 187, + 210 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In CVPR, 2023. 2, 3, 4", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 2, 7", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 1", + "[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, 2024. 3", + "[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 3", + "[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3", + "[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2024. 3", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3", + "[10] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 3", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4", + "[12] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3", + "[13] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In ICCV, 2021. 7" + ], + "bbox": [ + 93, + 220, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2", + "[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 5, 6", + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 4, 5, 7", + "[17] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 3, 7", + "[18] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 2, 3, 6, 7", + "[19] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In CVPR, 2024. 3", + "[20] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3", + "[21] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, et al. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In CVPR, 2024. 2", + "[22] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In ECCV, 2024. 3", + "[23] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs-enhancer: Enhancing unbounded 3d gaussian splattering with view-consistent 2d diffusion priors. In NeurIPS, 2024. 3", + "[24] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 3", + "[25] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In CVPR, 2024. 1, 2", + "[26] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Francisco Vicente Carrasco, Markus Steinberger, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. arXiv preprint arXiv:2406.15643, 2024. 1", + "[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2", + "[28] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM TOG, 38 (4):1-14, 2019. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6141", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 3", + "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2", + "[32] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2, 3", + "[33] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In ECCV, 2024. 3", + "[34] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3, 8", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 8", + "[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3", + "[38] Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. Loss-guided diffusion models for plug-and-play controllable generation. In ICML, 2023. 2, 3, 4", + "[39] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3, 4", + "[40] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7", + "[41] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 8", + "[42] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv preprint arXiv:2402.12712, 2024. 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1", + "[44] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In ICCV, 2023. 2, 3", + "[45] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 3", + "[46] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 3, 6, 7", + "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 6", + "[48] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 3", + "[49] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 3", + "[50] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In CVPR, 2023. 3", + "[51] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2, 3", + "[52] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In CVPR, 2023. 6, 7", + "[53] Haotian Ye, Haowei Lin, Jiaqi Han, Minkai Xu, Sheng Liu, Yitao Liang, Jianzhu Ma, James Zou, and Stefano Ermon. Tfg: Unified training-free guidance for diffusion models. arXiv preprint arXiv:2409.15761, 2024. 2, 3", + "[54] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 2, 6, 7", + "[55] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 3", + "[56] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In ICCV, 2023. 2, 3, 4", + "[57] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2, 3, 4, 6", + "[58] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, 2024. 1, 2" + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6142", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 3", + "[60] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In ECCV, 2024. 3", + "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3", + "[62] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[63] Yingji Zhong, Lanqing Hong, Zhenguo Li, and Dan Xu. Cvtxrf: Contrastive in-voxel transformer for 3d consistent radiance fields from sparse inputs. In CVPR, 2024. 2, 3", + "[64] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, 2024. 3, 6, 7", + "[65] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002. 4" + ], + "bbox": [ + 91, + 90, + 480, + 404 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6143", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_model.json b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d3ee81846565354c743cf2e27513ceba75bd942b --- /dev/null +++ b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_model.json @@ -0,0 +1,2730 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.819, + 0.177 + ], + "angle": 0, + "content": "Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.203, + 0.822, + 0.223 + ], + "angle": 0, + "content": "Yingji Zhong\\(^{1}\\) Zhihao Li\\(^{2}\\) Dave Zhenyu Chen\\(^{2}\\) Lanqing Hong\\(^{2}\\) Dan Xu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.228, + 0.83, + 0.248 + ], + "angle": 0, + "content": "1The Hong Kong University of Science and Technology 2Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.252, + 0.845, + 0.269 + ], + "angle": 0, + "content": "{yzhongbn,danxu}@cse.ust.hk, {zhihao.li,dave.zhenyuchen,honglanqing}@huawei.com" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.286, + 0.892, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.573, + 0.907, + 0.63 + ], + "angle": 0, + "content": "Figure 1. We tackle the critical issues of (a) extrapolation and (b) occlusion in sparse-input 3DGS by leveraging a video diffusion model. Vanilla generation often suffers from inconsistencies within the generated sequences (as highlighted by the yellow arrows), leading to black shadows in the rendered images. In contrast, our scene-grounding generation produces consistent sequences, effectively addressing these issues and enhancing overall quality (c), as indicated by the blue boxes. The numbers refer to PSNR values. Zoom in for better visualization." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.64, + 0.327, + 0.654 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Despite recent successes in novel view synthesis using 3D Gaussian Splatting (3DGS), modeling scenes with sparse inputs remains a challenge. In this work, we address two critical yet overlooked issues in real-world sparse-input modeling: extrapolation and occlusion. To tackle these issues, we propose to use a reconstruction by generation pipeline that leverages learned priors from video diffusion models to provide plausible interpretations for regions outside the field of view or occluded. However, the generated sequences exhibit inconsistencies that do not fully benefit subsequent 3DGS modeling. To address the challenge of inconsistencies, we introduce a novel scene-grounding guidance based on rendered sequences from an optimized 3DGS, which tames the diffusion model to generate consistent sequences. This guidance is training-free and does not require any fine-tuning of the diffusion model. To facilitate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.748 + ], + "angle": 0, + "content": "holistic scene modeling, we also propose a trajectory initialization method. It effectively identifies regions that are outside the field of view and occluded. We further design a scheme tailored for 3DGS optimization with generated sequences. Experiments demonstrate that our method significantly improves upon the baseline and achieves state-of-the-art performance on challenging benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.768, + 0.646, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent advances in 3D scene representation such as Neural Radiance Fields (NeRF) [2-4, 29, 31, 43] have greatly boosted the performance of Novel View Synthesis (NVS). NeRF represents the scene with a Multi-Layer Perceptron (MLP) and renders high-fidelity images with volumetric rendering. More recently, 3D Gaussian Splatt-ing (3DGS) [16, 25, 26, 58] emerges as a powerful explicit" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6133" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "representation that models the scene with a set of 3D gaussian primitives and renders images via differentiable splatting. 3DGS achieves comparable performance to NeRF while requiring significantly less training time and offering higher inference speeds." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.171, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Despite recent advances in scene representations based on 3DGS, modeling scenes with sparse inputs remains a significant challenge. The sparse supervision often leads radiance fields to learn degenerate representations due to shape-radiance ambiguity [59], regardless of whether the representation is NeRF or 3DGS. While there have been promising improvements [18, 32, 44, 63], the commonly used faceforwarding [14, 28] and object-oriented 'outside-in' viewing [29] settings oversimplify real-world sparse-input modeling, causing many methods to overlook two critical issues: (i) extrapolation - while the sparse inputs typically cover the scene as much as possible, there may still exist regions that are outside the field of view, as shown in Fig. 1 (a); (ii) occlusion - occlusion frequently occurs for novel views that deviate even slightly from the training input views, as illustrated in Fig. 1 (b). When rendering with an optimized 3DGS, these issues can cause severe artifacts, such as black holes, significantly degrading image quality. Therefore, handling these two issues, i.e., extrapolation and occlusion, is critical for real-world sparse-input modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.732 + ], + "angle": 0, + "content": "To address the above-discussed issues, we propose a novel reconstruction by generation pipeline based on 3DGS. Intuitively, we use video diffusion models [7, 12, 51, 57] to generate multi-view sequences, which provide plausible interpretations of the scene based on priors learned from large-scale datasets. These sequences significantly enlarge the viewing instances, offering a high potential to address the extrapolation and occlusion issues. With the sparse inputs and the generated sequences, we can optimize a 3DGS to model the scene. However, as shown in Fig. 1, this vanilla pipeline brings little improvement or may even degrade the performance. The main reason can be attributed to the multi-view inconsistency within the generated sequences. The inconsistency manifests in two aspects: (i) the appearance inconsistency between frames within a sequence; (ii) the generated sequence may contain hallucinated elements not present in the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To fully leverage the learned prior from video diffusion models for sparse-input 3DGS, we further explore addressing the challenges of inconsistencies within the generated sequences. Unlike existing methods that resolve appearance inconsistencies by assigning per-frame learnable appearance embeddings [21, 27], we focus on taming video diffusion models to directly generate sequences with consistency. Inspired by training-free guidance methods for diffusion models [1, 38, 53, 56] that enable controllable generation through external guidance, we introduce a novel strategy called scene-grounding guidance to ensure consis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.529 + ], + "angle": 0, + "content": "tent generation without requiring further fine-tuning of the diffusion models. Specifically, the scene-grounding guidance is based on a rendered sequence from an optimized 3DGS. During each step of the denoising process, the noisy sequence receives gradients from the supervision of the rendered sequence. Although the rendered sequence does not provide perfect guidance, our key insight in employing it to address the inconsistency is twofold: (i) adjacent frames within the rendered sequence are highly consistent due to limited camera movement between them; (ii) the rendered sequence is scene-grounding, which can guide the diffusion model to avoid generating elements that do not exist in the scene. In addition to addressing the issues of extrapolation and occlusion, our method also enhances the overall quality of the rendered images, as demonstrated in Fig. 1 (c). To effectively identify regions that are outside the field of view or occluded, we propose a trajectory initialization strategy to determine the camera trajectory during sequence generation, which is also based on an optimized 3DGS. With the proposed method, we can perform a holistic modeling of the scene. Additionally, we introduce a scheme for optimizing 3DGS with generated sequences, focusing on loss and sampling designs, which further enhance overall performance. We evaluate our method on two challenging in- 108 door datasets, Replica [40] and ScanNet++ [54], where the issues of extrapolation and occlusion are pronounced. The experiments demonstrate that our method achieves notable improvements and establishes state-of-the-art performance. Our contributions are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.53, + 0.905, + 0.575 + ], + "angle": 0, + "content": "- This paper is the first to explicitly address the challenges of extrapolation and occlusion in 3DGS modeling from sparse inputs." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.575, + 0.905, + 0.649 + ], + "angle": 0, + "content": "- We propose a novel reconstruction by generation pipeline with a designed scene-grounding guidance, which tames the video diffusion models to generate consistent and plausible sequences to effectively tackle the issues of extrapolation and occlusion." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.651, + 0.905, + 0.726 + ], + "angle": 0, + "content": "- We present a trajectory initialization strategy that effectively identifies regions that are outside the field of view and occluded, facilitating holistic scene modeling. We also introduce a scheme for optimizing 3DGS with generated sequences, further improving the performance." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.727, + 0.905, + 0.801 + ], + "angle": 0, + "content": "- Our method demonstrates significant improvements over the baseline, achieving over \\(3.5\\mathrm{dB}\\) and \\(2.5\\mathrm{dB}\\) PSNR enhancements on the Replica [40] and ScanNet++ [54] datasets, respectively, thereby establishing state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.53, + 0.905, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.815, + 0.663, + 0.831 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Radiance Fields from Sparse Inputs. Although improvements have been made in scene representation [2, 3, 16, 25, 29, 31, 58], learning a robust radiance field typically requires dense inputs as supervision due to the radiance-shape" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6134" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.885, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.345, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Framework overview of our proposed method. It consists of three parts: scene-grounding guidance, trajectory initialization, and optimization scheme with generated sequences. Initially, a baseline 3DGS is trained using sparse inputs and initialized with the point cloud from DUSt3R [46]. Yellow regions denote uncovered areas, e.g., those outside the field of view or occluded. The trajectory initialization determines the paths for sequence generation based on renderings from the baseline 3DGS, facilitating holistic scene modeling. The video diffusion model receives an input image along with the trajectory for sequence generation, incorporating scene-grounding guidance during the denoising process to ensure consistent output. The guidance is based on the rendered sequences. Finally, the generated sequences are utilized to optimize the final 3DGS through a tailored optimization scheme." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.456, + 0.485, + 0.76 + ], + "angle": 0, + "content": "ambiguity [59]. Current works can be roughly categorized into two lines of research. The first line of research focuses on pre-training generalizable radiance fields using multiview datasets. Generalizable NeRF [6, 24, 45, 55] typically projects the ray points to reference images and integrates 2D features as the auxiliary input of MLP. In contrast, generalizable 3DGS [5, 8, 22, 60] commonly perform dense depth prediction associated with gaussian properties. The second line of research leverages regularization techniques for optimization, which are applicable to both NeRF and 3DGS, including hand-crafted constraints [17, 32, 63], and those derived from pre-trained models [10, 18, 33, 35, 44, 50, 64]. In this paper, we explore the priors from video diffusion models [7, 12, 51, 57] for sparse-input 3DGS modeling. Unlike previous works, our method trains sparse-input 3DGS using augmented sequences from video diffusion models, which provide interpolation or extrapolation around the input views. We design a guidance strategy that tames the video diffusion model to generate more plausible and scene-grounded sequences, greatly enhancing the performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Diffusion Prior for Radiance Fields. Diffusion models [11, 36, 39] have shown remarkable generation capabilities. The strong prior knowledge embedded in diffusion models can facilitate the training of radiance fields. Specifically, several works leverage Score Distillation Sampling (SDS) [19, 20, 34, 48] using a frozen diffusion model to train a radiance field based on text prompts in a zero-shot manner. Other studies focus on training view-consistent [37, 42, 49] or quality-enhanced [23] diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.547 + ], + "angle": 0, + "content": "models, where the generated results can be directly applied to train a radiance field. Our method also uses the generated results to train a radiance field. However, our primary contribution lies in a novel guidance strategy that controls the generation to be consistent, which is crucial for sparse-input modeling augmented with generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.82 + ], + "angle": 0, + "content": "Controllable Generation for Diffusion Models. Current works of controllable generation can be categorized into training-required and training-free methods. Training-required methods fine-tune the diffusion models with additional conditions [30, 61], or train an additional noise-conditioned external guidance function, e.g., classifier guidance [9] for denosing sampler. Training-free methods freeze the foundation diffusion model, and modify the denoising process with the control signal from the external guidance functions [1, 38, 53, 56]. These methods do not require training additional guidance functions or fine-tuning diffusion models; instead, they enable controllable generation in a plug-and-play manner. Our work is inspired by training-free methods. However, we concentrate on multi-view modeling, which necessitates high consistency control over the generated results. Besides, unlike these methods that typically rely on pre-trained models for guidance, we utilize rendered results to provide scene-grounding guidance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.725, + 0.848 + ], + "angle": 0, + "content": "3. The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "In this paper, we utilize video diffusion models to tackle two critical issues in real-world sparse-input modeling: extrapolation and occlusion, as illustrated in Fig. 1. The overview" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6135" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "of our method is illustrated in Fig. 2, which consists of three proposed components: a scene-grounding guidance (Sec. 3.2), a trajectory initialization strategy (Sec. 3.3), and a scheme for 3DGS optimization with generated sequences (Sec. 3.4). We will detail these components following a preliminary review of 3DGS and diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.195, + 0.221, + 0.21 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.484, + 0.414 + ], + "angle": 0, + "content": "3D Gaussian Splating (3DGS) [16] represents a scene with a set of anisotropic 3D Gaussian primitives. Each Gaussian primitive is parametrized by a set of attributes: a center \\(\\mu\\), a scaling factor \\(s\\), a quaternion \\(q\\), an opacity value \\(\\alpha\\), and a feature vector \\(f\\). The basis function of each Gaussian primitive is formulated as \\(\\mathcal{G}(x) = \\exp(-\\frac{1}{2} (\\mathrm{x} - \\mu)^{\\mathrm{T}} \\Sigma^{-1} (\\mathrm{x} - \\mu))\\), where \\(\\Sigma\\) is the covariance matrix derived from \\(s\\) and \\(q\\). 3DGS renders the scene through a differentiable splatting, which firstly transforms the 3D Gaussian \\(\\mathcal{G}(x)\\) into 2D Gaussian \\(\\mathcal{G}'(x)\\) on the image plane via projection [65], and applies a tile-based rasterizer for rendering, which sorts the 2D Gaussians by depth and employs the \\(\\alpha\\)-blending as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.415, + 0.483, + 0.456 + ], + "angle": 0, + "content": "\\[\nC \\left(x _ {p}\\right) = \\sum_ {i \\in K} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\quad \\sigma_ {i} = \\alpha_ {i} \\mathcal {G} _ {i} ^ {\\prime} \\left(x _ {p}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.501 + ], + "angle": 0, + "content": "where \\( x_{p} \\) is the pixel position, \\( K \\) refers to the number of 2D Gaussians associated with the pixel, and \\( c \\) represents the decoded color of feature \\( f \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Diffusion Models [11, 39] are a family of generative models that progressively perturb data with intensifying Gaussian noises (i.e., forward noising), and then learn to reverse this process for sample generation (i.e., reverse denoising). The key of the diffusion model is a U-Net \\(\\epsilon_{\\theta}\\) which is trained to predict the noise that is injected in the current sample \\(\\mathbf{x}_t\\). The sampling is conducted by iterative denoising for \\(T\\) steps [56] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.63, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t - 1} = (1 + \\beta_ {t} / 2) \\mathbf {x} _ {t} + \\beta_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) + \\sqrt {\\beta_ {t}} \\mathbf {z} \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.775 + ], + "angle": 0, + "content": "where \\(\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t)\\) is the estimated score function which can be derived from \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t);\\beta_t\\) is pre-defined parameters; \\(\\mathbf{z}\\sim \\mathcal{N}(0,\\mathbf{I})\\). In this work, we leverage a camera-controlled image-to-video diffusion model [57], whose condition includes an image for the first frame, and the camera trajectory for the path of the generated sequence. The model is operated in a latent space of dimension \\(d\\), supporting the sequence length of \\(L\\), thus \\(\\mathbf{x}_t\\in \\mathbb{R}^{L\\times h\\times w\\times d}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.459, + 0.803 + ], + "angle": 0, + "content": "3.2. Generation via Scene-Grounding Guidance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Applying the generated sequences from the video diffusion model can provide plausible interpretations of regions not covered by the sparse inputs. However, as illustrated in Fig. 1, the inconsistency within the generated sequences manifests as: (i) appearance inconsistencies across frames and (ii) the occurrence of non-existent elements, which can" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.094, + 0.9, + 0.108 + ], + "angle": 0, + "content": "Algorithm 1 Generation with Scene-Grounding Guidance" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.114, + 0.815, + 0.132 + ], + "angle": 0, + "content": "1: Function GENERATOR \\((\\mathcal{R},I,\\{\\phi_j\\}_{j = 1}^L)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.132, + 0.905, + 0.161 + ], + "angle": 0, + "content": "2: Input: Optimized 3DGS model \\(\\mathcal{R}\\), input image \\(I\\), camera trajectory of a sequence \\(\\{\\phi_j\\}_{j=1}^L\\)." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.161, + 0.905, + 0.19 + ], + "angle": 0, + "content": "3: Given: Latent image-to-video diffusion model \\(\\epsilon_{\\theta}\\), VAE decoder \\(\\mathcal{D}\\), pre-defined \\(\\beta_{t}\\), \\(\\bar{\\alpha}_{t}\\) and guidance scale \\(\\gamma_{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.191, + 0.839, + 0.208 + ], + "angle": 0, + "content": "4: Abbreviate \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t,I,\\{\\phi_j\\}_{j = 1}^L)\\) as \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.208, + 0.901, + 0.224 + ], + "angle": 0, + "content": "5: \\(\\mathbf{S},\\mathbf{M} =\\) rasterize \\((\\{\\phi_j\\}_{j = 1}^L,\\mathcal{R})\\) \\(\\triangleright\\) Eq. (1)& (5)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.224, + 0.638, + 0.238 + ], + "angle": 0, + "content": "6: \\(\\mathbf{x}_T\\sim \\mathcal{N}(0,\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.239, + 0.676, + 0.251 + ], + "angle": 0, + "content": "7: for \\( t = T, \\dots, 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.253, + 0.774, + 0.267 + ], + "angle": 0, + "content": "8: \\(\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})\\) if \\(t > 1\\), else \\(\\mathbf{z} = \\mathbf{0}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.268, + 0.876, + 0.287 + ], + "angle": 0, + "content": "9: \\(\\hat{\\mathbf{x}}_{t - 1} = (1 + \\frac{1}{2}\\beta_t)\\mathbf{x}_t - \\frac{\\beta_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\epsilon_\\theta (\\mathbf{x}_t,t) + \\sqrt{\\beta_t}\\mathbf{z}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.287, + 0.799, + 0.305 + ], + "angle": 0, + "content": "10: \\(\\mathbf{x}_{0|t} = \\frac{1}{\\sqrt{\\bar{\\alpha}_t}} (\\mathbf{x}_t - \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon_\\theta (\\mathbf{x}_t,t))\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.306, + 0.665, + 0.32 + ], + "angle": 0, + "content": "11: \\(\\mathbf{X}_{0|t} = \\mathcal{D}(\\mathbf{x}_{0|t})\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.32, + 0.901, + 0.335 + ], + "angle": 0, + "content": "12: \\(\\mathbf{g}_t = \\nabla_{\\mathbf{x}_t}\\mathcal{L}(\\mathbf{S},\\mathbf{M},\\mathbf{X}_{0|t})\\) \\(\\triangleright\\) Eq. (6)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.336, + 0.905, + 0.35 + ], + "angle": 0, + "content": "13: \\(\\mathbf{x}_{t - 1} = \\hat{\\mathbf{x}}_{t - 1} - \\gamma_t\\mathbf{g}_t\\) \\(\\triangleright\\) Eq. (2)& (4)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.351, + 0.597, + 0.363 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.365, + 0.641, + 0.38 + ], + "angle": 0, + "content": "15: return \\(\\mathcal{D}(\\mathbf{x}_0)\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.472 + ], + "angle": 0, + "content": "negatively impact the 3DGS modeling. In this section, we propose an innovative scene-grounding guidance method that directs the video diffusion model to generate consistent sequences, significantly enhancing the performance of sparse-input 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.473, + 0.906, + 0.578 + ], + "angle": 0, + "content": "Inspired by previous training-free guidance methods [1, 56] that achieve their objectives by modifying the sampler in Eq. (2), we adopt a similar approach to attain the goal of consistency. Specifically, we firstly replace \\(\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t)\\) with a conditional score function \\(\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t|\\mathcal{Q})\\), where \\(\\mathcal{Q}\\) refers to the target of consistency. The conditional score function can be expanded by the Bayesian rule as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.585, + 0.905, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t} | \\mathcal {Q}) = \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) + \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathcal {Q} | \\mathbf {x} _ {t}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.713 + ], + "angle": 0, + "content": "where \\(\\nabla_{\\mathbf{x}_t}\\log p(\\mathcal{Q}|\\mathbf{x}_t)\\) can be considered as a guidance term that injects the consistency constraint into Eq. (2). We further formulate \\(p(\\mathcal{Q}|\\mathbf{x}_t)\\) as: \\(p(\\mathcal{Q}|\\mathbf{x}_t) = \\exp (-\\lambda \\mathcal{L}(\\mathcal{Q},\\mathbf{x}_t)) / Z\\) where \\(\\mathcal{L}(\\mathcal{Q},\\mathbf{x}_t)\\) indicates how well the current sample \\(\\mathbf{x}_t\\) is aligned with the target, and \\(Z\\) is a normalization term. The guidance term can thus be implemented using the gradient of the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.728, + 0.905, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathcal {Q} | \\mathbf {x} _ {t}) \\propto - \\nabla_ {\\mathbf {x} _ {t}} \\mathcal {L} (\\mathcal {Q}, \\mathbf {x} _ {t}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.904, + 0.78 + ], + "angle": 0, + "content": "which is appended to Eq. (2) to achieve the target of consistency during the denoising sampling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The remaining problem lies in how to define the consistency target \\(\\mathcal{Q}\\). Unlike previous works [1, 38, 56] that define the target based on external pre-trained models, we establish the target using a rendered sequence from an optimized 3DGS model \\(\\mathcal{R}\\). Though the rendered sequence is not perfect, our key insights are as follows: (i) the rendered images of adjacent frames are highly consistent, as the camera movement between them is typically minor; (ii)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6136" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "the rendered frames provide scene grounding, clearly indicating which elements exist in the scene. Therefore, the rendered sequence can serve as an effective guidance for the generated sequence to achieve the target of consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.484, + 0.26 + ], + "angle": 0, + "content": "Given a camera trajectory \\(\\{\\phi_j\\}_{j = 1}^L\\) for the sequence generation, we first utilize the optimized 3DGS to render a sequence \\(\\{S_j\\}_{j = 1}^L\\), along with a mask sequence \\(\\{M_j\\}_{j = 1}^L\\) that indicates the regions not covered by the sparse inputs. To get the mask, we first render a transmittance map, which is obtained by \\(\\alpha\\)-blending (as Eq. (1)) on the opacity. For each pixel \\(x_{p}\\), the \\(\\alpha\\)-blending is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.263, + 0.483, + 0.304 + ], + "angle": 0, + "content": "\\[\nO \\left(x _ {p}\\right) = \\sum_ {i \\in K} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.502 + ], + "angle": 0, + "content": "where \\(\\sigma\\) and \\(O\\) refer to the opacity of the gaussian and the transmittance map, respectively. The mask is then acquired by thresholding the transmittance map with a value \\(\\eta_{\\mathrm{mask}}\\): \\(M = (O < \\eta_{\\mathrm{mask}})\\). For convenience, we stack \\(\\{S_j\\}_{j=1}^L\\) and \\(\\{M_j\\}_{j=1}^L\\) to \\(\\mathbf{S}\\) and \\(\\mathbf{M}\\), which are of shape \\(L \\times H \\times W \\times 3\\) and \\(L \\times H \\times W \\times 1\\), respectively. Since the target of consistency is based on the rendered sequence in clean data space, to receive the guidance, we transform the noisy latent \\(\\mathbf{x}_t\\) into a latent \\(\\mathbf{x}_{0|t}\\) in the clean data space, based on prediction from the model \\(\\epsilon_\\theta\\): \\(\\mathbf{x}_{0|t} = (\\mathbf{x}_t - \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_\\theta(\\mathbf{x}_t, t)) / \\sqrt{\\bar{\\alpha}_t}\\). With the consistency target \\(\\mathcal{Q}\\) that is based on the rendered sequence \\(\\mathbf{S}\\), we formulate the function \\(\\mathcal{L}\\) in the guidance term (Eq. (4)) as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.505, + 0.483, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\mathbf {S}, \\mathbf {M}, \\mathbf {X} _ {0 | t}) = \\| \\bar {\\mathbf {M}} \\odot (\\mathbf {S} - \\mathbf {X} _ {0 | t}) \\| _ {1} \\tag {6} \\\\ + \\lambda_ {\\text {p e r c}} \\mathcal {L} _ {\\text {p e r c}} (\\bar {\\mathbf {M}} \\odot \\mathbf {S}, \\bar {\\mathbf {M}} \\odot \\mathbf {X} _ {0 | t}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.607 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_{0|t}\\) is decoded from the latent \\(\\mathbf{x}_{0|t}\\) by a VAE decoder, \\(\\bar{\\mathbf{M}}\\) is the result of flipping the values in \\(\\mathbf{M}\\), \\(\\odot\\) is the Hadamard product, and \\(\\mathcal{L}_{\\mathrm{perc}}\\) is a perceptual loss [15] with its corresponding weight as \\(\\lambda_{\\mathrm{perc}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.699 + ], + "angle": 0, + "content": "With the guidance from Eq. (6), the denoising process balances the consistency constraint and the prior from the diffusion model, integrating them into plausible generation results. This guidance does not involve any fine-tuning of the diffusion model, thereby preserving its generative capabilities. The detailed pipeline is outlined in Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.382, + 0.728 + ], + "angle": 0, + "content": "3.3. Trajectory Initialization Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To enable holistic modeling of the scene, the camera trajectories for the video diffusion model should cover regions that are outside the field of view or occluded as much as possible. The generated sequences can thus provide plausible interpretations for these regions, which serve as the basis for optimizing the subsequent 3DGS model. Similar to the scene-grounding guidance discussed in Sec. 3.2, the proposed trajectory initialization method is also based on an optimized 3DGS model. For the \\(i\\)-th sparse input view with camera pose \\(\\varphi_{i}\\), we first sample a set of candidate poses around it, as depicted in Fig. 3. Suppose" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.9, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.301, + 0.906, + 0.399 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed trajectory initialization strategy. The yellow parts represent unobserved regions. For each input view, we sample a set of candidate poses around it, and render at these poses using an optimized 3DGS. We select candidate poses whose renderings exhibit significant holes (highlighted by red boxes), and interpolate trajectories between these candidate poses and the input view's pose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.613 + ], + "angle": 0, + "content": "there are a total of \\(m\\) candidate poses, we use the optimized 3DGS model \\(\\mathcal{R}\\) to render images for these poses as \\(\\{\\hat{S}_c^{(i)},\\hat{M}_c^{(i)}\\}_{c = 1}^m =\\) rasterize \\((\\{\\hat{\\phi}_c^{(i)}\\}_{c = 1}^m,\\mathcal{R})\\) . For poses where the rendered images exhibit significant black holes, as indicated by the mask \\(\\hat{M}\\) calculated from Eq. (5), we interpolate a trajectory of length \\(L\\) (matching the length of video diffusion model) between the input camera pose and these poses as follows: \\(\\{\\phi_j^{(i,c)}\\}_{j = 1}^L = \\mathrm{interp}(\\varphi_i,\\hat{\\phi}_c^{(i)})\\) where \\(\\hat{\\phi}_c^{(i)}\\) refers to one selected candidate pose from the \\(i\\) -th input. In practice, we select the top- \\(k\\) candidate poses based on the sizes of their corresponding masks. Then, we build a trajectory pool by traversing all input views and their respective selected candidate poses as:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.618, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\Phi = \\left\\{\\left\\{\\phi_ {j} ^ {(i, c)} \\right\\} _ {j = 1} ^ {L} \\mid i, c \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.906, + 0.671 + ], + "angle": 0, + "content": "where each element in the pool is sampled for the sequence generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.833, + 0.695 + ], + "angle": 0, + "content": "3.4. 3DGS Optimization with Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.906, + 0.837 + ], + "angle": 0, + "content": "Given sparse inputs of \\(N\\) images along with their poses, i.e., \\(\\{C_i^{\\mathrm{gt}},\\varphi_i\\}_{i = 1}^N\\), we aim at optimizing a 3DGS model with the auxiliary generated sequences. For simplicity, we refer to the input images paired with their poses as 'input views', and we term the generated images with their associated poses as 'generated views'. During each iteration, we sample an input view and a generated view for supervision. Specifically, for the input view, we employ the default reconstruction loss [16] written as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.845, + 0.905, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {i n p u t}} = (1 - \\lambda) \\mathcal {L} _ {1} \\left(C _ {i}, C _ {i} ^ {\\mathrm {g t}}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(C _ {i}, C _ {i} ^ {\\mathrm {g t}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\( C_i \\) refers to the rendered image and \\( \\lambda \\) is a weighting factor. For supervision of generated views, we find that the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6137" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.093, + 0.426, + 0.108 + ], + "angle": 0, + "content": "Algorithm 2 3DGS Optimization with Generation" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.116, + 0.439, + 0.132 + ], + "angle": 0, + "content": "1: Input: Sparse inputs of N images \\(\\{C_i^{\\mathrm{gt}},\\varphi_i\\}_{i = 1}^N\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.132, + 0.482, + 0.162 + ], + "angle": 0, + "content": "2: Given: Number of iterations \\( N_{\\mathrm{iter}} \\), generation interval \\( N_{\\mathrm{gen}} \\), ratio of samples from other sequences \\( \\eta \\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.162, + 0.445, + 0.177 + ], + "angle": 0, + "content": "3: Variable: Global list of generated views \\(\\mathbf{G} = []\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.178, + 0.395, + 0.191 + ], + "angle": 0, + "content": "4: Baseline 3DGS model optimization \\(\\Rightarrow \\mathcal{R}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.192, + 0.478, + 0.207 + ], + "angle": 0, + "content": "5: Trajectory initialization \\(\\Rightarrow \\Phi\\) \\(\\triangleright\\) Eq. (7)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.208, + 0.303, + 0.222 + ], + "angle": 0, + "content": "6: for \\( t = 0, \\dots, N_{\\mathrm{iter}} - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.223, + 0.278, + 0.238 + ], + "angle": 0, + "content": "7: If \\(t\\% N_{\\mathrm{gen}} = 0\\) then" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.238, + 0.309, + 0.251 + ], + "angle": 0, + "content": "8: Sample an input view \\(I\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.252, + 0.476, + 0.269 + ], + "angle": 0, + "content": "9: Sample a trajectory around \\(I\\) from \\(\\Phi \\Rightarrow \\{\\phi_j\\}_{j=1}^L\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.27, + 0.395, + 0.285 + ], + "angle": 0, + "content": "10: \\(\\mathbf{S} = \\mathrm{GENERATOR}(\\mathcal{R},I,\\{\\phi_j\\}_{j = 1}^L)\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.286, + 0.257, + 0.299 + ], + "angle": 0, + "content": "11: Append S to G" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.301, + 0.186, + 0.312 + ], + "angle": 0, + "content": "12: End If" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.315, + 0.478, + 0.33 + ], + "angle": 0, + "content": "13: Sample an input view to get \\(\\mathcal{L}^{\\mathrm{input}}\\) \\(\\triangleright\\) Eq. (8)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.331, + 0.266, + 0.344 + ], + "angle": 0, + "content": "14: If rand() \\(\\geq \\eta\\) then" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.345, + 0.369, + 0.359 + ], + "angle": 0, + "content": "15: Sample a generated view from S" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.36, + 0.389, + 0.374 + ], + "angle": 0, + "content": "16: Else Sample a generated view from G" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.376, + 0.186, + 0.388 + ], + "angle": 0, + "content": "17: End If" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.391, + 0.478, + 0.405 + ], + "angle": 0, + "content": "18: Use the generated view to get \\(\\mathcal{L}^{\\mathrm{gen}}\\) \\(\\triangleright\\) Eq. (9)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.406, + 0.326, + 0.42 + ], + "angle": 0, + "content": "19: \\((\\mathcal{L}^{\\mathrm{input}} + \\mathcal{L}^{\\mathrm{gen}})\\) .backward()" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.421, + 0.353, + 0.434 + ], + "angle": 0, + "content": "20: #Densification and opacity reset" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.437, + 0.173, + 0.448 + ], + "angle": 0, + "content": "21: end for" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.116, + 0.482, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.484, + 0.584 + ], + "angle": 0, + "content": "reconstruction loss does not effectively fill the hole regions, and increasing its weight leads to performance degradation due to flaws in the generated images. To address this issue, we propose using perceptual loss [15]. The perceptual loss is calculated over the entire image, allowing those hole regions to significantly influence the gradients, thereby effectively driving the model to fill those holes. Thus, the loss on the generated views is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.595, + 0.482, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {g e n}} = \\lambda_ {\\text {g e n 1}} \\mathcal {L} _ {1} \\left(C _ {j}, S _ {j}\\right) + \\lambda_ {\\text {g e n 2}} \\mathcal {L} _ {\\text {p e r c}} \\left(C _ {j}, S _ {j}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.652 + ], + "angle": 0, + "content": "where \\(S_{j}\\) refers to the generated image, \\(\\lambda_{\\mathrm{gen1}}\\) and \\(\\lambda_{\\mathrm{gen2}}\\) are two balancing factors, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.484, + 0.804 + ], + "angle": 0, + "content": "We empirically find that conducting local sampling within a specific optimization interval, where a substantial portion of generated views is sampled from the same sequence of local regions, enhances visual quality. However, sampling exclusively from a single sequence can lead to a forgetting issue, where optimized information about holes in other regions becomes diluted. Therefore, within each interval of local sampling, we also include generated views from other sequences with a ratio \\(\\eta\\). The optimization pipeline is presented in Alg. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.81, + 0.222, + 0.826 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.288, + 0.85 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets and Metrics. We target addressing the issues of extrapolation and occlusion for sparse-input 3DGS scene modeling, which are overlooked by current benchmarks. To" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.09, + 0.638, + 0.1 + ], + "angle": 0, + "content": "w/ Vanilla Generation" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.101, + 0.643, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.659, + 0.09, + 0.771, + 0.1 + ], + "angle": 0, + "content": "w/ Guided Generation" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.102, + 0.772, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.802, + 0.09, + 0.872, + 0.099 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.102, + 0.899, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.176, + 0.905, + 0.231 + ], + "angle": 0, + "content": "Figure 4. Sequences from the vanilla generation suffer from inconsistencies. A 3DGS model optimized with these sequences renders images with black shadows, highlighted by red boxes, while our method solves this issue with the scene-grounding guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.905, + 0.393 + ], + "angle": 0, + "content": "evaluate the effectiveness of our method, we conduct experiments on a new benchmark created from two indoor datasets, i.e., the synthetic Replica [40] and the realistic ScanNet++ [54], which consists of 6 and 4 scenes, respectively. Although the selected six input views for each scene can cover most regions, there are still areas outside the field of view. Moreover, the 'inside-out' viewing directions make occlusion common in this benchmark. More details can be found in the supplement. For quantitative comparisons, we report PSNR, SSIM [47], and LPIPS [62] scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.485 + ], + "angle": 0, + "content": "Baseline. We train a baseline 3DGS model initialized with the point cloud from DUSt3R [46], incorporating the gaussian unpooling in FSGS [64], which makes the optimized model a strong baseline. Based on this we conduct experiments to verify the effectiveness of our method. The model is denoted as 'Baseline 3DGS' in the following." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.906, + 0.623 + ], + "angle": 0, + "content": "Implementation Details. The baseline model described above serves as the model \\(\\mathcal{R}\\) for scene-grounding guidance and trajectory initialization. For sequence generation, we employ the camera-controlled image-to-video diffusion model [57] which supports the generation of \\(L = 25\\) frames. The weighting factors, \\(\\lambda\\), \\(\\lambda_{\\mathrm{perc}}\\), \\(\\lambda_{\\mathrm{gen1}}\\), and \\(\\lambda_{\\mathrm{gen2}}\\) are set to \\(0.2\\), \\(10^{-4}\\), \\(0.1\\), and \\(0.01\\), respectively. The threshold of \\(\\eta_{\\mathrm{mask}}\\) is set to \\(0.9\\), while \\(\\eta\\) is set to \\(0.5\\). The generation interval \\(N_{\\mathrm{gen}}\\) is set to \\(260\\) and \\(N_{\\mathrm{iter}}\\) is set to \\(10,000\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.653, + 0.651 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Comparison on Replica. As shown in Tab. 1, our method achieves the highest performance on the Replica dataset, outperforming DNGaussian [18] and FSGS [64] by a significant margin of over \\(3.0\\mathrm{dB}\\) in PSNR. Fig. 5 illustrates that our method effectively addresses occlusion and extrapolation, while other 3DGS-based methods struggle with these challenges. Additionally, their depth regularization often compromise thin structures, such as the flower in the vase in the second row. FreeNeRF [52] exhibits severe artifacts because it cannot effectively utilize the strong prior from the DUSt3R point cloud. Although FreeNeRF can fill hole regions through neighboring interpolation (e.g., the wall behind the chair in the first row and the ceiling in the second row), the results frequently exhibit blurring or artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Comparison on ScanNet++. ScanNet++ is a dataset captured in realistic scenes, so it is more complicated and chal" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6138" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.882, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.43, + 0.907, + 0.471 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons on the Replica and ScanNet++ datasets. All 3DGS-based methods are optimized using the initialized point cloud from DUSt3R [46]. Our method effectively addresses the issues of extrapolation and occlusion while preserving finer details and reducing artifacts. For better visualization, please zoom in on the results." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.481, + 0.485, + 0.648 + ], + "angle": 0, + "content": "
MethodReplica [40]ScanNet++ [54]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Mip-NeRF [2]18.120.7070.39119.580.7550.389
InfoNeRF [17]13.070.5980.55214.540.6460.495
DietNeRF [13]18.990.6760.44419.760.7190.431
FreeNeRF [52]20.990.7650.32420.170.7560.368
3DGS†[16]22.800.8180.17921.410.8170.211
DNGaussian [18]17.630.7180.43519.010.7540.367
DNGaussian†[18]22.710.8210.18920.680.7880.281
FSGS [64]20.220.7600.30417.950.7300.373
FSGS†[64]22.990.8330.20521.230.8130.257
Ours26.350.8720.12223.890.8500.182
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.692 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on the Replica and ScanNet++ datasets. Including our approach, 3DGS-based methods marked with \\(\\updownarrow\\) are initialized with the point cloud from DUSt3R [46]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.484, + 0.848 + ], + "angle": 0, + "content": "lenging than the synthetic Replica dataset. The results in Tab. 1 demonstrate that our method has a clear advantage over current approaches, surpassing FSGS by more than \\(2.5\\mathrm{dB}\\) in PSNR. As depicted in Fig. 5, our method effectively addresses the extrapolation issue (e.g., the ceiling in the fourth row) and mitigates needle-like artifacts observed in the rendered images of DNGaussian [18] and FSGS [64] (the third row). Furthermore, the comparisons in the third row highlight our method's superiority in preserving finer details compared to all other methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.853, + 0.258, + 0.867 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our technical contributions consist of three key components. We analyze their effects on the Replica dataset." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.471, + 0.907, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.637, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Figure 6. Our method not only effectively addresses extrapolation and occlusion (red boxes), improving the overall quality (blue boxes), but also predicts more plausible geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Generation with Scene-Grounding Guidance. Optimizing a 3DGS with sequences from vanilla generation results in quality degradation. In Tab. 2 (a), while the full image metrics are enhanced due to slightly improved modeling at occluded regions, the visual quality degrades, as indicated by PSNR of observable regions dropping from \\(25.45\\mathrm{dB}\\) to \\(25.00\\mathrm{dB}\\). This degradation is attributed to inconsistencies within generated sequences, which can result in black shadows in rendered images as illustrated in Fig. 4. In contrast, our scene-grounding guidance ensures that the generated sequences remain consistent, significantly enhancing the modeling capability in regions outside the field of view and occluded, while also improving the overall quality, evidenced by the 'w/ Guided Generation' results in Tab. 2 (a)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6139" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.09, + 0.621, + 0.178 + ], + "angle": 0, + "content": "
(a)Gen.Guide.Traj.Full ImageObservable Regions
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Baseline 3DGS22.800.8180.17925.450.8600.129
w/ Vanilla Generation✓23.690.8400.16025.000.8700.119
w/ Guided Generation✓✓25.030.8520.13926.520.8810.101
w/ Guided Generation&Traj.✓✓✓25.580.8590.13826.530.8830.100
" + }, + { + "type": "table", + "bbox": [ + 0.629, + 0.09, + 0.908, + 0.178 + ], + "angle": 0, + "content": "
(b)PSNR↑SSIM↑LPIPS↓
Baseline 3DGS22.800.8180.179
w/ Guided Generation&Traj.25.580.8590.138
w/ perceptual loss26.350.8720.122
w/o local sampling26.280.8710.127
w/o global list26.010.8670.122
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.18, + 0.907, + 0.223 + ], + "angle": 0, + "content": "Table 2. Ablation experiments on the Replica dataset. (a) Effectiveness of the proposed scene-grounding guidance (Guide.) for generation, and the trajectory initialization strategy (Traj.). (Gen.) indicates utilizing generated sequences for modeling. Metrics of observable regions mask out regions outside the field of view or occluded. (b) Effectiveness of the proposed scheme for 3DGS optimization." + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.228, + 0.192, + 0.239 + ], + "angle": 0, + "content": "Baseline 3DGS" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.228, + 0.358, + 0.24 + ], + "angle": 0, + "content": "w/ Guided Genration&Traj." + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.228, + 0.468, + 0.24 + ], + "angle": 0, + "content": "w/ Perceptual Loss" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.242, + 0.222, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.242, + 0.348, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.242, + 0.476, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.317, + 0.483, + 0.344 + ], + "angle": 0, + "content": "Figure 7. The perceptual loss for generated views greatly increases the modeling capability at hole regions." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.349, + 0.462, + 0.416 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS↓
Baseline 3DGS+LaMa [41]24.560.8330.167
Baseline 3DGS+SDS [34] with SDInpaint24.330.8440.153
Baseline 3DGS+SDS [34] with SDInpaint*25.150.8530.141
Ours26.350.8720.122
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.417, + 0.483, + 0.446 + ], + "angle": 0, + "content": "Table 3. Comparisons with inpainting methods on the Replica dataset. * indicates the usage of our trajectory initialization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.46, + 0.483, + 0.595 + ], + "angle": 0, + "content": "Trajectory Initialization Strategy. Tab. 2 (a) further demonstrates that the proposed trajectory initialization strategy significantly boosts the performance, notated as 'w/ Guided Generation&Traj'. The improvement mainly arises from enhanced modeling of the regions outside the field of view or the occluded areas, as the metrics of visible regions plateau while the overall image metrics improve by over 0.5 dB in PSNR. This indicates that the initialization strategy effectively identifies hole regions for holistic modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Scheme for 3DGS Optimization with Generation. We verify the effectiveness of the proposed scheme in Tab. 2 (b). Specifically, the perceptual loss of Eq. (9) increases PSNR by over \\(0.5\\mathrm{dB}\\), which is crucial for the model to fill the hole regions, as shown in Fig. 7. We empirically find that local sampling brings improvement in Sec. 3.4. This is evidenced by the performance decrease of 'w/o local sampling', which randomly samples generated views from all generated sequences. Alg. 2 shows that we use a global list to avoid the forgetting problem, and its necessity is verified by an over \\(0.3\\mathrm{dB}\\) PSNR drop observed with 'w/o global list'. Combining these contributions, our full model effectively addresses extrapolation and occlusion while enhancing overall image quality, meanwhile exhibiting much better geometry, as shown in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.836, + 0.483, + 0.852 + ], + "angle": 0, + "content": "4.4. Further Comparisons with Inpainting Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Extrapolation and occlusion can also be addressed using inpainting methods. We thus compare our approach with two inpainting-based methods. One method applies LaMa [41]" + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.228, + 0.596, + 0.24 + ], + "angle": 0, + "content": "LaMa" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.242, + 0.642, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.228, + 0.767, + 0.24 + ], + "angle": 0, + "content": "SDS with SDInpaint*" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.242, + 0.771, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.228, + 0.856, + 0.24 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.242, + 0.899, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.389, + 0.907, + 0.444 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons with other inpainting methods. Our approach not only produces more plausible appearances around the inpainting regions but also predicts more consistent geometries in fine-grained local areas." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.681 + ], + "angle": 0, + "content": "inpainting on hole regions, while the other optimizes a 3DGS by Score Distillation Sampling (SDS) [34] based on a SDInpaint model [36]. We also incorporate our proposed trajectory initialization into the SDS method to enhance the optimization of inpainting regions. Results in Tab. 3 show that our method outperforms these two methods by more than \\(1.0\\mathrm{dB}\\) in PSNR. Qualitative results in Fig. 8 indicate that, under certain conditions, the SDS-based method produces inpainted areas with strange appearances, while LaMa tends to create blurring artifacts interpolated from neighboring regions. Besides, Fig. 8 also shows that our method predicts better geometry in regions with local details. The inpainting results from our approach are more plausible due to the well-designed guidance, which effectively exploits prior knowledge from the diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.686, + 0.634, + 0.701 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have explored to address the critical issues of extrapolation and occlusion in sparse-input 3DGS modeling. We propose using video diffusion models that provide plausible interpretations for regions that are outside the field of view and occluded. To resolve inconsistencies within generated sequences, we introduce a novel scene-grounding guidance that controls the diffusion model to generate consistent sequences without any fine-tuning. Additionally, we propose a trajectory initialization strategy to enhance holistic modeling and develop a scheme for optimizing 3DGS with generated sequences. Extensive experiments validate our approach, demonstrating that it outperforms current methods by a significant margin." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6140" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.184 + ], + "angle": 0, + "content": "Acknowledgement. This research is supported in part by the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, and Tencent Rhino-Bird Focused Research Program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.196, + 0.188, + 0.212 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.221, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In CVPR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.278, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.334, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.69, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[10] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In ICCV, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.221, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[14] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[17] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[18] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[19] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[20] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[21] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, et al. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[22] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[23] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs-enhancer: Enhancing unbounded 3d gaussian splattering with view-consistent 2d diffusion priors. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[24] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.905, + 0.703 + ], + "angle": 0, + "content": "[25] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[26] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Francisco Vicente Carrasco, Markus Steinberger, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. arXiv preprint arXiv:2406.15643, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM TOG, 38 (4):1-14, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6141" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[32] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[33] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[34] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.632 + ], + "angle": 0, + "content": "[38] Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. Loss-guided diffusion models for plug-and-play controllable generation. In ICML, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[39] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.677, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[40] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[41] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[42] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv preprint arXiv:2402.12712, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[43] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[44] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.191, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[45] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.261, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[46] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.304, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.345, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[48] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[49] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.457, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[50] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.498, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[51] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.568, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[52] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In CVPR, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[53] Haotian Ye, Haowei Lin, Jiaqi Han, Minkai Xu, Sheng Liu, Yitao Liang, Jianzhu Ma, James Zou, and Stefano Ermon. Tfg: Unified training-free guidance for diffusion models. arXiv preprint arXiv:2409.15761, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[54] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.707, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[55] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[56] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In ICCV, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.791, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[57] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[58] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6142" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[59] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[60] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[62] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[63] Yingji Zhong, Lanqing Hong, Zhenguo Li, and Dan Xu. Cvtxrf: Contrastive in-voxel transformer for 3d consistent radiance fields from sparse inputs. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[64] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[65] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6143" + } + ] +] \ No newline at end of file diff --git a/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_origin.pdf b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..962249f050031591a0d966e8d815c0630e97ed4e --- /dev/null +++ b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/b3776566-4465-43cf-8d8c-c85e0e038e48_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c41a544a61b4ed0f8aa69a06bdaf7347268a7cc998042ce553198acf4b31cb46 +size 5881181 diff --git a/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/full.md b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..30fb4a9faabe0f67c999cd48c629529a8501dbe2 --- /dev/null +++ b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/full.md @@ -0,0 +1,372 @@ +# Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs + +Yingji Zhong $^{1}$ Zhihao Li $^{2}$ Dave Zhenyu Chen $^{2}$ Lanqing Hong $^{2}$ Dan Xu $^{1}$ + +1The Hong Kong University of Science and Technology 2Huawei Noah's Ark Lab + +{yzhongbn,danxu}@cse.ust.hk, {zhihao.li,dave.zhenyuchen,honglanqing}@huawei.com + +![](images/9116717b497687e07d9578d6803f616b891debbc78b9a4b8dcf892ca67404706.jpg) +Figure 1. We tackle the critical issues of (a) extrapolation and (b) occlusion in sparse-input 3DGS by leveraging a video diffusion model. Vanilla generation often suffers from inconsistencies within the generated sequences (as highlighted by the yellow arrows), leading to black shadows in the rendered images. In contrast, our scene-grounding generation produces consistent sequences, effectively addressing these issues and enhancing overall quality (c), as indicated by the blue boxes. The numbers refer to PSNR values. Zoom in for better visualization. + +# Abstract + +Despite recent successes in novel view synthesis using 3D Gaussian Splatting (3DGS), modeling scenes with sparse inputs remains a challenge. In this work, we address two critical yet overlooked issues in real-world sparse-input modeling: extrapolation and occlusion. To tackle these issues, we propose to use a reconstruction by generation pipeline that leverages learned priors from video diffusion models to provide plausible interpretations for regions outside the field of view or occluded. However, the generated sequences exhibit inconsistencies that do not fully benefit subsequent 3DGS modeling. To address the challenge of inconsistencies, we introduce a novel scene-grounding guidance based on rendered sequences from an optimized 3DGS, which tames the diffusion model to generate consistent sequences. This guidance is training-free and does not require any fine-tuning of the diffusion model. To facilitate + +holistic scene modeling, we also propose a trajectory initialization method. It effectively identifies regions that are outside the field of view and occluded. We further design a scheme tailored for 3DGS optimization with generated sequences. Experiments demonstrate that our method significantly improves upon the baseline and achieves state-of-the-art performance on challenging benchmarks. + +# 1. Introduction + +Recent advances in 3D scene representation such as Neural Radiance Fields (NeRF) [2-4, 29, 31, 43] have greatly boosted the performance of Novel View Synthesis (NVS). NeRF represents the scene with a Multi-Layer Perceptron (MLP) and renders high-fidelity images with volumetric rendering. More recently, 3D Gaussian Splatt-ing (3DGS) [16, 25, 26, 58] emerges as a powerful explicit + +representation that models the scene with a set of 3D gaussian primitives and renders images via differentiable splatting. 3DGS achieves comparable performance to NeRF while requiring significantly less training time and offering higher inference speeds. + +Despite recent advances in scene representations based on 3DGS, modeling scenes with sparse inputs remains a significant challenge. The sparse supervision often leads radiance fields to learn degenerate representations due to shape-radiance ambiguity [59], regardless of whether the representation is NeRF or 3DGS. While there have been promising improvements [18, 32, 44, 63], the commonly used faceforwarding [14, 28] and object-oriented 'outside-in' viewing [29] settings oversimplify real-world sparse-input modeling, causing many methods to overlook two critical issues: (i) extrapolation - while the sparse inputs typically cover the scene as much as possible, there may still exist regions that are outside the field of view, as shown in Fig. 1 (a); (ii) occlusion - occlusion frequently occurs for novel views that deviate even slightly from the training input views, as illustrated in Fig. 1 (b). When rendering with an optimized 3DGS, these issues can cause severe artifacts, such as black holes, significantly degrading image quality. Therefore, handling these two issues, i.e., extrapolation and occlusion, is critical for real-world sparse-input modeling. + +To address the above-discussed issues, we propose a novel reconstruction by generation pipeline based on 3DGS. Intuitively, we use video diffusion models [7, 12, 51, 57] to generate multi-view sequences, which provide plausible interpretations of the scene based on priors learned from large-scale datasets. These sequences significantly enlarge the viewing instances, offering a high potential to address the extrapolation and occlusion issues. With the sparse inputs and the generated sequences, we can optimize a 3DGS to model the scene. However, as shown in Fig. 1, this vanilla pipeline brings little improvement or may even degrade the performance. The main reason can be attributed to the multi-view inconsistency within the generated sequences. The inconsistency manifests in two aspects: (i) the appearance inconsistency between frames within a sequence; (ii) the generated sequence may contain hallucinated elements not present in the scene. + +To fully leverage the learned prior from video diffusion models for sparse-input 3DGS, we further explore addressing the challenges of inconsistencies within the generated sequences. Unlike existing methods that resolve appearance inconsistencies by assigning per-frame learnable appearance embeddings [21, 27], we focus on taming video diffusion models to directly generate sequences with consistency. Inspired by training-free guidance methods for diffusion models [1, 38, 53, 56] that enable controllable generation through external guidance, we introduce a novel strategy called scene-grounding guidance to ensure consis + +tent generation without requiring further fine-tuning of the diffusion models. Specifically, the scene-grounding guidance is based on a rendered sequence from an optimized 3DGS. During each step of the denoising process, the noisy sequence receives gradients from the supervision of the rendered sequence. Although the rendered sequence does not provide perfect guidance, our key insight in employing it to address the inconsistency is twofold: (i) adjacent frames within the rendered sequence are highly consistent due to limited camera movement between them; (ii) the rendered sequence is scene-grounding, which can guide the diffusion model to avoid generating elements that do not exist in the scene. In addition to addressing the issues of extrapolation and occlusion, our method also enhances the overall quality of the rendered images, as demonstrated in Fig. 1 (c). To effectively identify regions that are outside the field of view or occluded, we propose a trajectory initialization strategy to determine the camera trajectory during sequence generation, which is also based on an optimized 3DGS. With the proposed method, we can perform a holistic modeling of the scene. Additionally, we introduce a scheme for optimizing 3DGS with generated sequences, focusing on loss and sampling designs, which further enhance overall performance. We evaluate our method on two challenging in- 108 door datasets, Replica [40] and ScanNet++ [54], where the issues of extrapolation and occlusion are pronounced. The experiments demonstrate that our method achieves notable improvements and establishes state-of-the-art performance. Our contributions are summarized as: + +- This paper is the first to explicitly address the challenges of extrapolation and occlusion in 3DGS modeling from sparse inputs. +- We propose a novel reconstruction by generation pipeline with a designed scene-grounding guidance, which tames the video diffusion models to generate consistent and plausible sequences to effectively tackle the issues of extrapolation and occlusion. +- We present a trajectory initialization strategy that effectively identifies regions that are outside the field of view and occluded, facilitating holistic scene modeling. We also introduce a scheme for optimizing 3DGS with generated sequences, further improving the performance. +- Our method demonstrates significant improvements over the baseline, achieving over $3.5\mathrm{dB}$ and $2.5\mathrm{dB}$ PSNR enhancements on the Replica [40] and ScanNet++ [54] datasets, respectively, thereby establishing state-of-the-art performance. + +# 2. Related Works + +Radiance Fields from Sparse Inputs. Although improvements have been made in scene representation [2, 3, 16, 25, 29, 31, 58], learning a robust radiance field typically requires dense inputs as supervision due to the radiance-shape + +![](images/1adff7a35be3b3e9be90abc25e4366e13fdfa15b87e931dbf42f2322f8656915.jpg) +Figure 2. Framework overview of our proposed method. It consists of three parts: scene-grounding guidance, trajectory initialization, and optimization scheme with generated sequences. Initially, a baseline 3DGS is trained using sparse inputs and initialized with the point cloud from DUSt3R [46]. Yellow regions denote uncovered areas, e.g., those outside the field of view or occluded. The trajectory initialization determines the paths for sequence generation based on renderings from the baseline 3DGS, facilitating holistic scene modeling. The video diffusion model receives an input image along with the trajectory for sequence generation, incorporating scene-grounding guidance during the denoising process to ensure consistent output. The guidance is based on the rendered sequences. Finally, the generated sequences are utilized to optimize the final 3DGS through a tailored optimization scheme. + +ambiguity [59]. Current works can be roughly categorized into two lines of research. The first line of research focuses on pre-training generalizable radiance fields using multiview datasets. Generalizable NeRF [6, 24, 45, 55] typically projects the ray points to reference images and integrates 2D features as the auxiliary input of MLP. In contrast, generalizable 3DGS [5, 8, 22, 60] commonly perform dense depth prediction associated with gaussian properties. The second line of research leverages regularization techniques for optimization, which are applicable to both NeRF and 3DGS, including hand-crafted constraints [17, 32, 63], and those derived from pre-trained models [10, 18, 33, 35, 44, 50, 64]. In this paper, we explore the priors from video diffusion models [7, 12, 51, 57] for sparse-input 3DGS modeling. Unlike previous works, our method trains sparse-input 3DGS using augmented sequences from video diffusion models, which provide interpolation or extrapolation around the input views. We design a guidance strategy that tames the video diffusion model to generate more plausible and scene-grounded sequences, greatly enhancing the performance. + +Diffusion Prior for Radiance Fields. Diffusion models [11, 36, 39] have shown remarkable generation capabilities. The strong prior knowledge embedded in diffusion models can facilitate the training of radiance fields. Specifically, several works leverage Score Distillation Sampling (SDS) [19, 20, 34, 48] using a frozen diffusion model to train a radiance field based on text prompts in a zero-shot manner. Other studies focus on training view-consistent [37, 42, 49] or quality-enhanced [23] diffusion + +models, where the generated results can be directly applied to train a radiance field. Our method also uses the generated results to train a radiance field. However, our primary contribution lies in a novel guidance strategy that controls the generation to be consistent, which is crucial for sparse-input modeling augmented with generation. + +Controllable Generation for Diffusion Models. Current works of controllable generation can be categorized into training-required and training-free methods. Training-required methods fine-tune the diffusion models with additional conditions [30, 61], or train an additional noise-conditioned external guidance function, e.g., classifier guidance [9] for denosing sampler. Training-free methods freeze the foundation diffusion model, and modify the denoising process with the control signal from the external guidance functions [1, 38, 53, 56]. These methods do not require training additional guidance functions or fine-tuning diffusion models; instead, they enable controllable generation in a plug-and-play manner. Our work is inspired by training-free methods. However, we concentrate on multi-view modeling, which necessitates high consistency control over the generated results. Besides, unlike these methods that typically rely on pre-trained models for guidance, we utilize rendered results to provide scene-grounding guidance. + +# 3. The Proposed Method + +In this paper, we utilize video diffusion models to tackle two critical issues in real-world sparse-input modeling: extrapolation and occlusion, as illustrated in Fig. 1. The overview + +of our method is illustrated in Fig. 2, which consists of three proposed components: a scene-grounding guidance (Sec. 3.2), a trajectory initialization strategy (Sec. 3.3), and a scheme for 3DGS optimization with generated sequences (Sec. 3.4). We will detail these components following a preliminary review of 3DGS and diffusion models. + +# 3.1. Preliminary + +3D Gaussian Splating (3DGS) [16] represents a scene with a set of anisotropic 3D Gaussian primitives. Each Gaussian primitive is parametrized by a set of attributes: a center $\mu$ , a scaling factor $s$ , a quaternion $q$ , an opacity value $\alpha$ , and a feature vector $f$ . The basis function of each Gaussian primitive is formulated as $\mathcal{G}(x) = \exp(-\frac{1}{2} (\mathrm{x} - \mu)^{\mathrm{T}} \Sigma^{-1} (\mathrm{x} - \mu))$ , where $\Sigma$ is the covariance matrix derived from $s$ and $q$ . 3DGS renders the scene through a differentiable splatting, which firstly transforms the 3D Gaussian $\mathcal{G}(x)$ into 2D Gaussian $\mathcal{G}'(x)$ on the image plane via projection [65], and applies a tile-based rasterizer for rendering, which sorts the 2D Gaussians by depth and employs the $\alpha$ -blending as follows: + +$$ +C \left(x _ {p}\right) = \sum_ {i \in K} c _ {i} \sigma_ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \sigma_ {j}\right), \quad \sigma_ {i} = \alpha_ {i} \mathcal {G} _ {i} ^ {\prime} \left(x _ {p}\right), \tag {1} +$$ + +where $x_{p}$ is the pixel position, $K$ refers to the number of 2D Gaussians associated with the pixel, and $c$ represents the decoded color of feature $f$ . + +Diffusion Models [11, 39] are a family of generative models that progressively perturb data with intensifying Gaussian noises (i.e., forward noising), and then learn to reverse this process for sample generation (i.e., reverse denoising). The key of the diffusion model is a U-Net $\epsilon_{\theta}$ which is trained to predict the noise that is injected in the current sample $\mathbf{x}_t$ . The sampling is conducted by iterative denoising for $T$ steps [56] as follows: + +$$ +\mathbf {x} _ {t - 1} = (1 + \beta_ {t} / 2) \mathbf {x} _ {t} + \beta_ {t} \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t}) + \sqrt {\beta_ {t}} \mathbf {z} \quad (2) +$$ + +where $\nabla_{\mathbf{x}_t}\log p(\mathbf{x}_t)$ is the estimated score function which can be derived from $\epsilon_{\theta}(\mathbf{x}_t,t);\beta_t$ is pre-defined parameters; $\mathbf{z}\sim \mathcal{N}(0,\mathbf{I})$ . In this work, we leverage a camera-controlled image-to-video diffusion model [57], whose condition includes an image for the first frame, and the camera trajectory for the path of the generated sequence. The model is operated in a latent space of dimension $d$ , supporting the sequence length of $L$ , thus $\mathbf{x}_t\in \mathbb{R}^{L\times h\times w\times d}$ . + +# 3.2. Generation via Scene-Grounding Guidance + +Applying the generated sequences from the video diffusion model can provide plausible interpretations of regions not covered by the sparse inputs. However, as illustrated in Fig. 1, the inconsistency within the generated sequences manifests as: (i) appearance inconsistencies across frames and (ii) the occurrence of non-existent elements, which can + +Algorithm 1 Generation with Scene-Grounding Guidance + +1: Function GENERATOR $(\mathcal{R},I,\{\phi_j\}_{j = 1}^L)$ + +2: Input: Optimized 3DGS model $\mathcal{R}$ , input image $I$ , camera trajectory of a sequence $\{\phi_j\}_{j=1}^L$ . + +3: Given: Latent image-to-video diffusion model $\epsilon_{\theta}$ , VAE decoder $\mathcal{D}$ , pre-defined $\beta_{t}$ , $\bar{\alpha}_{t}$ and guidance scale $\gamma_{t}$ . + +4: Abbreviate $\epsilon_{\theta}(\mathbf{x}_t,t,I,\{\phi_j\}_{j = 1}^L)$ as $\epsilon_{\theta}(\mathbf{x}_t,t)$ + +5: $\mathbf{S},\mathbf{M} =$ rasterize $(\{\phi_j\}_{j = 1}^L,\mathcal{R})$ $\triangleright$ Eq. (1)& (5) + +6: $\mathbf{x}_T\sim \mathcal{N}(0,\mathbf{I})$ + +7: for $t = T, \dots, 1$ do + +8: $\mathbf{z} \sim \mathcal{N}(0, \mathbf{I})$ if $t > 1$ , else $\mathbf{z} = \mathbf{0}$ + +9: $\hat{\mathbf{x}}_{t - 1} = (1 + \frac{1}{2}\beta_t)\mathbf{x}_t - \frac{\beta_t}{\sqrt{1 - \bar{\alpha}_t}}\epsilon_\theta (\mathbf{x}_t,t) + \sqrt{\beta_t}\mathbf{z}$ + +10: $\mathbf{x}_{0|t} = \frac{1}{\sqrt{\bar{\alpha}_t}} (\mathbf{x}_t - \sqrt{1 - \bar{\alpha}_t}\epsilon_\theta (\mathbf{x}_t,t))$ + +11: $\mathbf{X}_{0|t} = \mathcal{D}(\mathbf{x}_{0|t})$ + +12: $\mathbf{g}_t = \nabla_{\mathbf{x}_t}\mathcal{L}(\mathbf{S},\mathbf{M},\mathbf{X}_{0|t})$ $\triangleright$ Eq. (6) + +13: $\mathbf{x}_{t - 1} = \hat{\mathbf{x}}_{t - 1} - \gamma_t\mathbf{g}_t$ $\triangleright$ Eq. (2)& (4) + +14: end for + +15: return $\mathcal{D}(\mathbf{x}_0)$ + +negatively impact the 3DGS modeling. In this section, we propose an innovative scene-grounding guidance method that directs the video diffusion model to generate consistent sequences, significantly enhancing the performance of sparse-input 3DGS. + +Inspired by previous training-free guidance methods [1, 56] that achieve their objectives by modifying the sampler in Eq. (2), we adopt a similar approach to attain the goal of consistency. Specifically, we firstly replace $\nabla_{\mathbf{x}_t}\log p(\mathbf{x}_t)$ with a conditional score function $\nabla_{\mathbf{x}_t}\log p(\mathbf{x}_t|\mathcal{Q})$ , where $\mathcal{Q}$ refers to the target of consistency. The conditional score function can be expanded by the Bayesian rule as: + +$$ +\nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t} | \mathcal {Q}) = \nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t}) + \nabla_ {\mathbf {x} _ {t}} \log p (\mathcal {Q} | \mathbf {x} _ {t}), \tag {3} +$$ + +where $\nabla_{\mathbf{x}_t}\log p(\mathcal{Q}|\mathbf{x}_t)$ can be considered as a guidance term that injects the consistency constraint into Eq. (2). We further formulate $p(\mathcal{Q}|\mathbf{x}_t)$ as: $p(\mathcal{Q}|\mathbf{x}_t) = \exp (-\lambda \mathcal{L}(\mathcal{Q},\mathbf{x}_t)) / Z$ where $\mathcal{L}(\mathcal{Q},\mathbf{x}_t)$ indicates how well the current sample $\mathbf{x}_t$ is aligned with the target, and $Z$ is a normalization term. The guidance term can thus be implemented using the gradient of the following loss function: + +$$ +\nabla_ {\mathbf {x} _ {t}} \log p (\mathcal {Q} | \mathbf {x} _ {t}) \propto - \nabla_ {\mathbf {x} _ {t}} \mathcal {L} (\mathcal {Q}, \mathbf {x} _ {t}), \tag {4} +$$ + +which is appended to Eq. (2) to achieve the target of consistency during the denoising sampling. + +The remaining problem lies in how to define the consistency target $\mathcal{Q}$ . Unlike previous works [1, 38, 56] that define the target based on external pre-trained models, we establish the target using a rendered sequence from an optimized 3DGS model $\mathcal{R}$ . Though the rendered sequence is not perfect, our key insights are as follows: (i) the rendered images of adjacent frames are highly consistent, as the camera movement between them is typically minor; (ii) + +the rendered frames provide scene grounding, clearly indicating which elements exist in the scene. Therefore, the rendered sequence can serve as an effective guidance for the generated sequence to achieve the target of consistency. + +Given a camera trajectory $\{\phi_j\}_{j = 1}^L$ for the sequence generation, we first utilize the optimized 3DGS to render a sequence $\{S_j\}_{j = 1}^L$ , along with a mask sequence $\{M_j\}_{j = 1}^L$ that indicates the regions not covered by the sparse inputs. To get the mask, we first render a transmittance map, which is obtained by $\alpha$ -blending (as Eq. (1)) on the opacity. For each pixel $x_{p}$ , the $\alpha$ -blending is formulated as: + +$$ +O \left(x _ {p}\right) = \sum_ {i \in K} \sigma_ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \sigma_ {j}\right), \tag {5} +$$ + +where $\sigma$ and $O$ refer to the opacity of the gaussian and the transmittance map, respectively. The mask is then acquired by thresholding the transmittance map with a value $\eta_{\mathrm{mask}}$ : $M = (O < \eta_{\mathrm{mask}})$ . For convenience, we stack $\{S_j\}_{j=1}^L$ and $\{M_j\}_{j=1}^L$ to $\mathbf{S}$ and $\mathbf{M}$ , which are of shape $L \times H \times W \times 3$ and $L \times H \times W \times 1$ , respectively. Since the target of consistency is based on the rendered sequence in clean data space, to receive the guidance, we transform the noisy latent $\mathbf{x}_t$ into a latent $\mathbf{x}_{0|t}$ in the clean data space, based on prediction from the model $\epsilon_\theta$ : $\mathbf{x}_{0|t} = (\mathbf{x}_t - \sqrt{1 - \bar{\alpha}_t} \epsilon_\theta(\mathbf{x}_t, t)) / \sqrt{\bar{\alpha}_t}$ . With the consistency target $\mathcal{Q}$ that is based on the rendered sequence $\mathbf{S}$ , we formulate the function $\mathcal{L}$ in the guidance term (Eq. (4)) as: + +$$ +\begin{array}{l} \mathcal {L} (\mathbf {S}, \mathbf {M}, \mathbf {X} _ {0 | t}) = \| \bar {\mathbf {M}} \odot (\mathbf {S} - \mathbf {X} _ {0 | t}) \| _ {1} \tag {6} \\ + \lambda_ {\text {p e r c}} \mathcal {L} _ {\text {p e r c}} (\bar {\mathbf {M}} \odot \mathbf {S}, \bar {\mathbf {M}} \odot \mathbf {X} _ {0 | t}), \\ \end{array} +$$ + +where $\mathbf{X}_{0|t}$ is decoded from the latent $\mathbf{x}_{0|t}$ by a VAE decoder, $\bar{\mathbf{M}}$ is the result of flipping the values in $\mathbf{M}$ , $\odot$ is the Hadamard product, and $\mathcal{L}_{\mathrm{perc}}$ is a perceptual loss [15] with its corresponding weight as $\lambda_{\mathrm{perc}}$ . + +With the guidance from Eq. (6), the denoising process balances the consistency constraint and the prior from the diffusion model, integrating them into plausible generation results. This guidance does not involve any fine-tuning of the diffusion model, thereby preserving its generative capabilities. The detailed pipeline is outlined in Alg. 1. + +# 3.3. Trajectory Initialization Strategy + +To enable holistic modeling of the scene, the camera trajectories for the video diffusion model should cover regions that are outside the field of view or occluded as much as possible. The generated sequences can thus provide plausible interpretations for these regions, which serve as the basis for optimizing the subsequent 3DGS model. Similar to the scene-grounding guidance discussed in Sec. 3.2, the proposed trajectory initialization method is also based on an optimized 3DGS model. For the $i$ -th sparse input view with camera pose $\varphi_{i}$ , we first sample a set of candidate poses around it, as depicted in Fig. 3. Suppose + +![](images/d1ef091fba6b0377cc0b70a326fbc2fcf34cfcd09333ecee5591c04b445a6485.jpg) +Figure 3. Illustration of the proposed trajectory initialization strategy. The yellow parts represent unobserved regions. For each input view, we sample a set of candidate poses around it, and render at these poses using an optimized 3DGS. We select candidate poses whose renderings exhibit significant holes (highlighted by red boxes), and interpolate trajectories between these candidate poses and the input view's pose. + +there are a total of $m$ candidate poses, we use the optimized 3DGS model $\mathcal{R}$ to render images for these poses as $\{\hat{S}_c^{(i)},\hat{M}_c^{(i)}\}_{c = 1}^m =$ rasterize $(\{\hat{\phi}_c^{(i)}\}_{c = 1}^m,\mathcal{R})$ . For poses where the rendered images exhibit significant black holes, as indicated by the mask $\hat{M}$ calculated from Eq. (5), we interpolate a trajectory of length $L$ (matching the length of video diffusion model) between the input camera pose and these poses as follows: $\{\phi_j^{(i,c)}\}_{j = 1}^L = \mathrm{interp}(\varphi_i,\hat{\phi}_c^{(i)})$ where $\hat{\phi}_c^{(i)}$ refers to one selected candidate pose from the $i$ -th input. In practice, we select the top- $k$ candidate poses based on the sizes of their corresponding masks. Then, we build a trajectory pool by traversing all input views and their respective selected candidate poses as: + +$$ +\Phi = \left\{\left\{\phi_ {j} ^ {(i, c)} \right\} _ {j = 1} ^ {L} \mid i, c \right\}, \tag {7} +$$ + +where each element in the pool is sampled for the sequence generation. + +# 3.4. 3DGS Optimization with Generation + +Given sparse inputs of $N$ images along with their poses, i.e., $\{C_i^{\mathrm{gt}},\varphi_i\}_{i = 1}^N$ , we aim at optimizing a 3DGS model with the auxiliary generated sequences. For simplicity, we refer to the input images paired with their poses as 'input views', and we term the generated images with their associated poses as 'generated views'. During each iteration, we sample an input view and a generated view for supervision. Specifically, for the input view, we employ the default reconstruction loss [16] written as: + +$$ +\mathcal {L} ^ {\text {i n p u t}} = (1 - \lambda) \mathcal {L} _ {1} \left(C _ {i}, C _ {i} ^ {\mathrm {g t}}\right) + \lambda \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}} \left(C _ {i}, C _ {i} ^ {\mathrm {g t}}\right), \tag {8} +$$ + +where $C_i$ refers to the rendered image and $\lambda$ is a weighting factor. For supervision of generated views, we find that the + +Algorithm 2 3DGS Optimization with Generation + +1: Input: Sparse inputs of N images $\{C_i^{\mathrm{gt}},\varphi_i\}_{i = 1}^N$ +2: Given: Number of iterations $N_{\mathrm{iter}}$ , generation interval $N_{\mathrm{gen}}$ , ratio of samples from other sequences $\eta$ . +3: Variable: Global list of generated views $\mathbf{G} = []$ +4: Baseline 3DGS model optimization $\Rightarrow \mathcal{R}$ +5: Trajectory initialization $\Rightarrow \Phi$ $\triangleright$ Eq. (7) +6: for $t = 0, \dots, N_{\mathrm{iter}} - 1$ do +7: If $t\% N_{\mathrm{gen}} = 0$ then +8: Sample an input view $I$ +9: Sample a trajectory around $I$ from $\Phi \Rightarrow \{\phi_j\}_{j=1}^L$ +10: $\mathbf{S} = \mathrm{GENERATOR}(\mathcal{R},I,\{\phi_j\}_{j = 1}^L)$ +11: Append S to G +12: End If +13: Sample an input view to get $\mathcal{L}^{\mathrm{input}}$ $\triangleright$ Eq. (8) +14: If rand() $\geq \eta$ then +15: Sample a generated view from S +16: Else Sample a generated view from G +17: End If +18: Use the generated view to get $\mathcal{L}^{\mathrm{gen}}$ $\triangleright$ Eq. (9) +19: $(\mathcal{L}^{\mathrm{input}} + \mathcal{L}^{\mathrm{gen}})$ .backward() +20: #Densification and opacity reset +21: end for + +reconstruction loss does not effectively fill the hole regions, and increasing its weight leads to performance degradation due to flaws in the generated images. To address this issue, we propose using perceptual loss [15]. The perceptual loss is calculated over the entire image, allowing those hole regions to significantly influence the gradients, thereby effectively driving the model to fill those holes. Thus, the loss on the generated views is formulated as: + +$$ +\mathcal {L} ^ {\text {g e n}} = \lambda_ {\text {g e n 1}} \mathcal {L} _ {1} \left(C _ {j}, S _ {j}\right) + \lambda_ {\text {g e n 2}} \mathcal {L} _ {\text {p e r c}} \left(C _ {j}, S _ {j}\right), \tag {9} +$$ + +where $S_{j}$ refers to the generated image, $\lambda_{\mathrm{gen1}}$ and $\lambda_{\mathrm{gen2}}$ are two balancing factors, respectively. + +We empirically find that conducting local sampling within a specific optimization interval, where a substantial portion of generated views is sampled from the same sequence of local regions, enhances visual quality. However, sampling exclusively from a single sequence can lead to a forgetting issue, where optimized information about holes in other regions becomes diluted. Therefore, within each interval of local sampling, we also include generated views from other sequences with a ratio $\eta$ . The optimization pipeline is presented in Alg. 2. + +# 4. Experiments + +# 4.1. Experimental Setups + +Datasets and Metrics. We target addressing the issues of extrapolation and occlusion for sparse-input 3DGS scene modeling, which are overlooked by current benchmarks. To + +w/ Vanilla Generation + +![](images/1537861a4c58d25b5858af5487478160c838e5ec1cb31f2e7d8a653ff32f50c9.jpg) +Figure 4. Sequences from the vanilla generation suffer from inconsistencies. A 3DGS model optimized with these sequences renders images with black shadows, highlighted by red boxes, while our method solves this issue with the scene-grounding guidance. + +w/ Guided Generation + +![](images/3984260e981f383a62bb0a63b6b2b8a8aeec30db5060c0f6c27955d2f2e114c4.jpg) + +Ground Truth + +![](images/9aa4ee5f3f265f18b84669cd92f1f5fe9e3984d8821aaa6ea64696a00ef24ed5.jpg) + +evaluate the effectiveness of our method, we conduct experiments on a new benchmark created from two indoor datasets, i.e., the synthetic Replica [40] and the realistic ScanNet++ [54], which consists of 6 and 4 scenes, respectively. Although the selected six input views for each scene can cover most regions, there are still areas outside the field of view. Moreover, the 'inside-out' viewing directions make occlusion common in this benchmark. More details can be found in the supplement. For quantitative comparisons, we report PSNR, SSIM [47], and LPIPS [62] scores. + +Baseline. We train a baseline 3DGS model initialized with the point cloud from DUSt3R [46], incorporating the gaussian unpooling in FSGS [64], which makes the optimized model a strong baseline. Based on this we conduct experiments to verify the effectiveness of our method. The model is denoted as 'Baseline 3DGS' in the following. + +Implementation Details. The baseline model described above serves as the model $\mathcal{R}$ for scene-grounding guidance and trajectory initialization. For sequence generation, we employ the camera-controlled image-to-video diffusion model [57] which supports the generation of $L = 25$ frames. The weighting factors, $\lambda$ , $\lambda_{\mathrm{perc}}$ , $\lambda_{\mathrm{gen1}}$ , and $\lambda_{\mathrm{gen2}}$ are set to $0.2$ , $10^{-4}$ , $0.1$ , and $0.01$ , respectively. The threshold of $\eta_{\mathrm{mask}}$ is set to $0.9$ , while $\eta$ is set to $0.5$ . The generation interval $N_{\mathrm{gen}}$ is set to $260$ and $N_{\mathrm{iter}}$ is set to $10,000$ . + +# 4.2. Comparisons + +Comparison on Replica. As shown in Tab. 1, our method achieves the highest performance on the Replica dataset, outperforming DNGaussian [18] and FSGS [64] by a significant margin of over $3.0\mathrm{dB}$ in PSNR. Fig. 5 illustrates that our method effectively addresses occlusion and extrapolation, while other 3DGS-based methods struggle with these challenges. Additionally, their depth regularization often compromise thin structures, such as the flower in the vase in the second row. FreeNeRF [52] exhibits severe artifacts because it cannot effectively utilize the strong prior from the DUSt3R point cloud. Although FreeNeRF can fill hole regions through neighboring interpolation (e.g., the wall behind the chair in the first row and the ceiling in the second row), the results frequently exhibit blurring or artifacts. + +Comparison on ScanNet++. ScanNet++ is a dataset captured in realistic scenes, so it is more complicated and chal + +![](images/2f90b6a651d05826be96b312818235e9c13bb555a2d8087df277def0cd44845e.jpg) + +
MethodReplica [40]ScanNet++ [54]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Mip-NeRF [2]18.120.7070.39119.580.7550.389
InfoNeRF [17]13.070.5980.55214.540.6460.495
DietNeRF [13]18.990.6760.44419.760.7190.431
FreeNeRF [52]20.990.7650.32420.170.7560.368
3DGS†[16]22.800.8180.17921.410.8170.211
DNGaussian [18]17.630.7180.43519.010.7540.367
DNGaussian†[18]22.710.8210.18920.680.7880.281
FSGS [64]20.220.7600.30417.950.7300.373
FSGS†[64]22.990.8330.20521.230.8130.257
Ours26.350.8720.12223.890.8500.182
+ +Table 1. Quantitative comparisons on the Replica and ScanNet++ datasets. Including our approach, 3DGS-based methods marked with $\updownarrow$ are initialized with the point cloud from DUSt3R [46]. + +lenging than the synthetic Replica dataset. The results in Tab. 1 demonstrate that our method has a clear advantage over current approaches, surpassing FSGS by more than $2.5\mathrm{dB}$ in PSNR. As depicted in Fig. 5, our method effectively addresses the extrapolation issue (e.g., the ceiling in the fourth row) and mitigates needle-like artifacts observed in the rendered images of DNGaussian [18] and FSGS [64] (the third row). Furthermore, the comparisons in the third row highlight our method's superiority in preserving finer details compared to all other methods. + +# 4.3. Ablation Studies + +Our technical contributions consist of three key components. We analyze their effects on the Replica dataset. + +![](images/8938fcdfba6d9110e773d45996333126be74c9a43c506768960b25c6d0947339.jpg) +Figure 5. Qualitative comparisons on the Replica and ScanNet++ datasets. All 3DGS-based methods are optimized using the initialized point cloud from DUSt3R [46]. Our method effectively addresses the issues of extrapolation and occlusion while preserving finer details and reducing artifacts. For better visualization, please zoom in on the results. +Figure 6. Our method not only effectively addresses extrapolation and occlusion (red boxes), improving the overall quality (blue boxes), but also predicts more plausible geometry. + +Generation with Scene-Grounding Guidance. Optimizing a 3DGS with sequences from vanilla generation results in quality degradation. In Tab. 2 (a), while the full image metrics are enhanced due to slightly improved modeling at occluded regions, the visual quality degrades, as indicated by PSNR of observable regions dropping from $25.45\mathrm{dB}$ to $25.00\mathrm{dB}$ . This degradation is attributed to inconsistencies within generated sequences, which can result in black shadows in rendered images as illustrated in Fig. 4. In contrast, our scene-grounding guidance ensures that the generated sequences remain consistent, significantly enhancing the modeling capability in regions outside the field of view and occluded, while also improving the overall quality, evidenced by the 'w/ Guided Generation' results in Tab. 2 (a). + +
(a)Gen.Guide.Traj.Full ImageObservable Regions
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Baseline 3DGS22.800.8180.17925.450.8600.129
w/ Vanilla Generation✓23.690.8400.16025.000.8700.119
w/ Guided Generation✓✓25.030.8520.13926.520.8810.101
w/ Guided Generation&Traj.✓✓✓25.580.8590.13826.530.8830.100
+ +
(b)PSNR↑SSIM↑LPIPS↓
Baseline 3DGS22.800.8180.179
w/ Guided Generation&Traj.25.580.8590.138
w/ perceptual loss26.350.8720.122
w/o local sampling26.280.8710.127
w/o global list26.010.8670.122
+ +![](images/c28a4bff8d684a5dcb1ef8a36a8b73af00d4928123255edeb3373fc5082eb13b.jpg) +Baseline 3DGS +w/ Guided Genration&Traj. +Figure 7. The perceptual loss for generated views greatly increases the modeling capability at hole regions. + +![](images/556a5d5d14b0e3d3d91a798cc8d95e9552d71d192dd8fe3a9141006973072802.jpg) + +![](images/83f0dd9468285be22122a5df6b9c290869192250138ec402ad5363aa239f19fd.jpg) +w/ Perceptual Loss + +Table 2. Ablation experiments on the Replica dataset. (a) Effectiveness of the proposed scene-grounding guidance (Guide.) for generation, and the trajectory initialization strategy (Traj.). (Gen.) indicates utilizing generated sequences for modeling. Metrics of observable regions mask out regions outside the field of view or occluded. (b) Effectiveness of the proposed scheme for 3DGS optimization. + +
PSNR↑SSIM↑LPIPS↓
Baseline 3DGS+LaMa [41]24.560.8330.167
Baseline 3DGS+SDS [34] with SDInpaint24.330.8440.153
Baseline 3DGS+SDS [34] with SDInpaint*25.150.8530.141
Ours26.350.8720.122
+ +Table 3. Comparisons with inpainting methods on the Replica dataset. * indicates the usage of our trajectory initialization. + +Trajectory Initialization Strategy. Tab. 2 (a) further demonstrates that the proposed trajectory initialization strategy significantly boosts the performance, notated as 'w/ Guided Generation&Traj'. The improvement mainly arises from enhanced modeling of the regions outside the field of view or the occluded areas, as the metrics of visible regions plateau while the overall image metrics improve by over 0.5 dB in PSNR. This indicates that the initialization strategy effectively identifies hole regions for holistic modeling. + +Scheme for 3DGS Optimization with Generation. We verify the effectiveness of the proposed scheme in Tab. 2 (b). Specifically, the perceptual loss of Eq. (9) increases PSNR by over $0.5\mathrm{dB}$ , which is crucial for the model to fill the hole regions, as shown in Fig. 7. We empirically find that local sampling brings improvement in Sec. 3.4. This is evidenced by the performance decrease of 'w/o local sampling', which randomly samples generated views from all generated sequences. Alg. 2 shows that we use a global list to avoid the forgetting problem, and its necessity is verified by an over $0.3\mathrm{dB}$ PSNR drop observed with 'w/o global list'. Combining these contributions, our full model effectively addresses extrapolation and occlusion while enhancing overall image quality, meanwhile exhibiting much better geometry, as shown in Fig. 6. + +# 4.4. Further Comparisons with Inpainting Methods + +Extrapolation and occlusion can also be addressed using inpainting methods. We thus compare our approach with two inpainting-based methods. One method applies LaMa [41] + +![](images/765d64f45012320363a4026bdc7ec0b3d41053160dd277e3b3f23f00e4a0da0b.jpg) +LaMa +Figure 8. Qualitative comparisons with other inpainting methods. Our approach not only produces more plausible appearances around the inpainting regions but also predicts more consistent geometries in fine-grained local areas. + +![](images/be06c768b96b4c9495a667ba7c86f35f83435b2414f25ad3a9408f6338da935f.jpg) +SDS with SDInpaint* + +![](images/bed2b29a0e7c0318fee352dbcbf14a6bbfb44561648ba35990edd819de56304b.jpg) +Ours + +inpainting on hole regions, while the other optimizes a 3DGS by Score Distillation Sampling (SDS) [34] based on a SDInpaint model [36]. We also incorporate our proposed trajectory initialization into the SDS method to enhance the optimization of inpainting regions. Results in Tab. 3 show that our method outperforms these two methods by more than $1.0\mathrm{dB}$ in PSNR. Qualitative results in Fig. 8 indicate that, under certain conditions, the SDS-based method produces inpainted areas with strange appearances, while LaMa tends to create blurring artifacts interpolated from neighboring regions. Besides, Fig. 8 also shows that our method predicts better geometry in regions with local details. The inpainting results from our approach are more plausible due to the well-designed guidance, which effectively exploits prior knowledge from the diffusion model. + +# 5. Conclusion + +In this paper, we have explored to address the critical issues of extrapolation and occlusion in sparse-input 3DGS modeling. We propose using video diffusion models that provide plausible interpretations for regions that are outside the field of view and occluded. To resolve inconsistencies within generated sequences, we introduce a novel scene-grounding guidance that controls the diffusion model to generate consistent sequences without any fine-tuning. Additionally, we propose a trajectory initialization strategy to enhance holistic modeling and develop a scheme for optimizing 3DGS with generated sequences. Extensive experiments validate our approach, demonstrating that it outperforms current methods by a significant margin. + +Acknowledgement. This research is supported in part by the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, and Tencent Rhino-Bird Focused Research Program. + +# References + +[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In CVPR, 2023. 2, 3, 4 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 2, 7 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 1 +[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, 2024. 3 +[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 3 +[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3 +[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2024. 3 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3 +[10] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 3 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4 +[12] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3 +[13] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In ICCV, 2021. 7 + +[14] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2 +[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 5, 6 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 4, 5, 7 +[17] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 3, 7 +[18] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 2, 3, 6, 7 +[19] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In CVPR, 2024. 3 +[20] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3 +[21] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, et al. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In CVPR, 2024. 2 +[22] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In ECCV, 2024. 3 +[23] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs-enhancer: Enhancing unbounded 3d gaussian splattering with view-consistent 2d diffusion priors. In NeurIPS, 2024. 3 +[24] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 3 +[25] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In CVPR, 2024. 1, 2 +[26] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Francisco Vicente Carrasco, Markus Steinberger, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. arXiv preprint arXiv:2406.15643, 2024. 1 +[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2 +[28] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM TOG, 38 (4):1-14, 2019. 2 + +[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 3 +[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2 +[32] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2, 3 +[33] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In ECCV, 2024. 3 +[34] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3, 8 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 8 +[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3 +[38] Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. Loss-guided diffusion models for plug-and-play controllable generation. In ICML, 2023. 2, 3, 4 +[39] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3, 4 +[40] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7 +[41] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 8 +[42] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv preprint arXiv:2402.12712, 2024. 3 + +[43] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1 +[44] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In ICCV, 2023. 2, 3 +[45] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 3 +[46] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 3, 6, 7 +[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 6 +[48] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 3 +[49] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 3 +[50] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In CVPR, 2023. 3 +[51] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2, 3 +[52] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In CVPR, 2023. 6, 7 +[53] Haotian Ye, Haowei Lin, Jiaqi Han, Minkai Xu, Sheng Liu, Yitao Liang, Jianzhu Ma, James Zou, and Stefano Ermon. Tfg: Unified training-free guidance for diffusion models. arXiv preprint arXiv:2409.15761, 2024. 2, 3 +[54] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 2, 6, 7 +[55] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 3 +[56] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In ICCV, 2023. 2, 3, 4 +[57] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2, 3, 4, 6 +[58] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, 2024. 1, 2 + +[59] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 3 +[60] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In ECCV, 2024. 3 +[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3 +[62] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[63] Yingji Zhong, Lanqing Hong, Zhenguo Li, and Dan Xu. Cvtxrf: Contrastive in-voxel transformer for 3d consistent radiance fields from sparse inputs. In CVPR, 2024. 2, 3 +[64] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, 2024. 3, 6, 7 +[65] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002. 4 \ No newline at end of file diff --git a/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/images.zip b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9b427bda79939d592931f2190d23fac29ec48c86 --- /dev/null +++ b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6245c4eda1460d055131b97bb673e19789ec25196452ac5f0618e69b437de29 +size 778567 diff --git a/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/layout.json b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e6411680a8894cf657c0c8734ad9054c9cf554d8 --- /dev/null +++ b/2025/Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs/layout.json @@ -0,0 +1,11451 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "text", + "content": "Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "text", + "content": "Yingji Zhong" + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "text", + "content": " Zhihao Li" + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "text", + "content": " Dave Zhenyu Chen" + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "text", + "content": " Lanqing Hong" + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "text", + "content": " Dan Xu" + }, + { + "bbox": [ + 105, + 160, + 503, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 180, + 507, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 180, + 507, + 196 + ], + "spans": [ + { + "bbox": [ + 104, + 180, + 507, + 196 + ], + "type": "text", + "content": "1The Hong Kong University of Science and Technology 2Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 199, + 517, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 199, + 517, + 213 + ], + "spans": [ + { + "bbox": [ + 91, + 199, + 517, + 213 + ], + "type": "text", + "content": "{yzhongbn,danxu}@cse.ust.hk, {zhihao.li,dave.zhenyuchen,honglanqing}@huawei.com" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 61, + 226, + 545, + 450 + ], + "blocks": [ + { + "bbox": [ + 61, + 226, + 545, + 450 + ], + "lines": [ + { + "bbox": [ + 61, + 226, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 61, + 226, + 545, + 450 + ], + "type": "image", + "image_path": "9116717b497687e07d9578d6803f616b891debbc78b9a4b8dcf892ca67404706.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 453, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 453, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 555, + 498 + ], + "type": "text", + "content": "Figure 1. We tackle the critical issues of (a) extrapolation and (b) occlusion in sparse-input 3DGS by leveraging a video diffusion model. Vanilla generation often suffers from inconsistencies within the generated sequences (as highlighted by the yellow arrows), leading to black shadows in the rendered images. In contrast, our scene-grounding generation produces consistent sequences, effectively addressing these issues and enhancing overall quality (c), as indicated by the blue boxes. The numbers refer to PSNR values. Zoom in for better visualization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 506, + 200, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 506, + 200, + 517 + ], + "spans": [ + { + "bbox": [ + 151, + 506, + 200, + 517 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 522, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 297, + 714 + ], + "type": "text", + "content": "Despite recent successes in novel view synthesis using 3D Gaussian Splatting (3DGS), modeling scenes with sparse inputs remains a challenge. In this work, we address two critical yet overlooked issues in real-world sparse-input modeling: extrapolation and occlusion. To tackle these issues, we propose to use a reconstruction by generation pipeline that leverages learned priors from video diffusion models to provide plausible interpretations for regions outside the field of view or occluded. However, the generated sequences exhibit inconsistencies that do not fully benefit subsequent 3DGS modeling. To address the challenge of inconsistencies, we introduce a novel scene-grounding guidance based on rendered sequences from an optimized 3DGS, which tames the diffusion model to generate consistent sequences. This guidance is training-free and does not require any fine-tuning of the diffusion model. To facilitate" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "type": "text", + "content": "holistic scene modeling, we also propose a trajectory initialization method. It effectively identifies regions that are outside the field of view and occluded. We further design a scheme tailored for 3DGS optimization with generated sequences. Experiments demonstrate that our method significantly improves upon the baseline and achieves state-of-the-art performance on challenging benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 608, + 395, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 608, + 395, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 608, + 395, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Recent advances in 3D scene representation such as Neural Radiance Fields (NeRF) [2-4, 29, 31, 43] have greatly boosted the performance of Novel View Synthesis (NVS). NeRF represents the scene with a Multi-Layer Perceptron (MLP) and renders high-fidelity images with volumetric rendering. More recently, 3D Gaussian Splatt-ing (3DGS) [16, 25, 26, 58] emerges as a powerful explicit" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6133" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "representation that models the scene with a set of 3D gaussian primitives and renders images via differentiable splatting. 3DGS achieves comparable performance to NeRF while requiring significantly less training time and offering higher inference speeds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 135, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 135, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 135, + 295, + 374 + ], + "type": "text", + "content": "Despite recent advances in scene representations based on 3DGS, modeling scenes with sparse inputs remains a significant challenge. The sparse supervision often leads radiance fields to learn degenerate representations due to shape-radiance ambiguity [59], regardless of whether the representation is NeRF or 3DGS. While there have been promising improvements [18, 32, 44, 63], the commonly used faceforwarding [14, 28] and object-oriented 'outside-in' viewing [29] settings oversimplify real-world sparse-input modeling, causing many methods to overlook two critical issues: (i) extrapolation - while the sparse inputs typically cover the scene as much as possible, there may still exist regions that are outside the field of view, as shown in Fig. 1 (a); (ii) occlusion - occlusion frequently occurs for novel views that deviate even slightly from the training input views, as illustrated in Fig. 1 (b). When rendering with an optimized 3DGS, these issues can cause severe artifacts, such as black holes, significantly degrading image quality. Therefore, handling these two issues, i.e., extrapolation and occlusion, is critical for real-world sparse-input modeling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 376, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 579 + ], + "type": "text", + "content": "To address the above-discussed issues, we propose a novel reconstruction by generation pipeline based on 3DGS. Intuitively, we use video diffusion models [7, 12, 51, 57] to generate multi-view sequences, which provide plausible interpretations of the scene based on priors learned from large-scale datasets. These sequences significantly enlarge the viewing instances, offering a high potential to address the extrapolation and occlusion issues. With the sparse inputs and the generated sequences, we can optimize a 3DGS to model the scene. However, as shown in Fig. 1, this vanilla pipeline brings little improvement or may even degrade the performance. The main reason can be attributed to the multi-view inconsistency within the generated sequences. The inconsistency manifests in two aspects: (i) the appearance inconsistency between frames within a sequence; (ii) the generated sequence may contain hallucinated elements not present in the scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "To fully leverage the learned prior from video diffusion models for sparse-input 3DGS, we further explore addressing the challenges of inconsistencies within the generated sequences. Unlike existing methods that resolve appearance inconsistencies by assigning per-frame learnable appearance embeddings [21, 27], we focus on taming video diffusion models to directly generate sequences with consistency. Inspired by training-free guidance methods for diffusion models [1, 38, 53, 56] that enable controllable generation through external guidance, we introduce a novel strategy called scene-grounding guidance to ensure consis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 418 + ], + "type": "text", + "content": "tent generation without requiring further fine-tuning of the diffusion models. Specifically, the scene-grounding guidance is based on a rendered sequence from an optimized 3DGS. During each step of the denoising process, the noisy sequence receives gradients from the supervision of the rendered sequence. Although the rendered sequence does not provide perfect guidance, our key insight in employing it to address the inconsistency is twofold: (i) adjacent frames within the rendered sequence are highly consistent due to limited camera movement between them; (ii) the rendered sequence is scene-grounding, which can guide the diffusion model to avoid generating elements that do not exist in the scene. In addition to addressing the issues of extrapolation and occlusion, our method also enhances the overall quality of the rendered images, as demonstrated in Fig. 1 (c). To effectively identify regions that are outside the field of view or occluded, we propose a trajectory initialization strategy to determine the camera trajectory during sequence generation, which is also based on an optimized 3DGS. With the proposed method, we can perform a holistic modeling of the scene. Additionally, we introduce a scheme for optimizing 3DGS with generated sequences, focusing on loss and sampling designs, which further enhance overall performance. We evaluate our method on two challenging in- 108 door datasets, Replica [40] and ScanNet++ [54], where the issues of extrapolation and occlusion are pronounced. The experiments demonstrate that our method achieves notable improvements and establishes state-of-the-art performance. Our contributions are summarized as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 419, + 553, + 634 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 315, + 419, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 419, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 315, + 419, + 553, + 455 + ], + "type": "text", + "content": "- This paper is the first to explicitly address the challenges of extrapolation and occlusion in 3DGS modeling from sparse inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 455, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 455, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 455, + 553, + 514 + ], + "type": "text", + "content": "- We propose a novel reconstruction by generation pipeline with a designed scene-grounding guidance, which tames the video diffusion models to generate consistent and plausible sequences to effectively tackle the issues of extrapolation and occlusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 515, + 553, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 515, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 315, + 515, + 553, + 574 + ], + "type": "text", + "content": "- We present a trajectory initialization strategy that effectively identifies regions that are outside the field of view and occluded, facilitating holistic scene modeling. We also introduce a scheme for optimizing 3DGS with generated sequences, further improving the performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 575, + 553, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 575, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 315, + 575, + 553, + 634 + ], + "type": "text", + "content": "- Our method demonstrates significant improvements over the baseline, achieving over " + }, + { + "bbox": [ + 315, + 575, + 553, + 634 + ], + "type": "inline_equation", + "content": "3.5\\mathrm{dB}" + }, + { + "bbox": [ + 315, + 575, + 553, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 575, + 553, + 634 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{dB}" + }, + { + "bbox": [ + 315, + 575, + 553, + 634 + ], + "type": "text", + "content": " PSNR enhancements on the Replica [40] and ScanNet++ [54] datasets, respectively, thereby establishing state-of-the-art performance." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 645, + 405, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 645, + 405, + 658 + ], + "spans": [ + { + "bbox": [ + 314, + 645, + 405, + 658 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "type": "text", + "content": "Radiance Fields from Sparse Inputs. Although improvements have been made in scene representation [2, 3, 16, 25, 29, 31, 58], learning a robust radiance field typically requires dense inputs as supervision due to the radiance-shape" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6134" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 541, + 270 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 270 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 270 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 270 + ], + "type": "image", + "image_path": "1adff7a35be3b3e9be90abc25e4366e13fdfa15b87e931dbf42f2322f8656915.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 273, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 273, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. Framework overview of our proposed method. It consists of three parts: scene-grounding guidance, trajectory initialization, and optimization scheme with generated sequences. Initially, a baseline 3DGS is trained using sparse inputs and initialized with the point cloud from DUSt3R [46]. Yellow regions denote uncovered areas, e.g., those outside the field of view or occluded. The trajectory initialization determines the paths for sequence generation based on renderings from the baseline 3DGS, facilitating holistic scene modeling. The video diffusion model receives an input image along with the trajectory for sequence generation, incorporating scene-grounding guidance during the denoising process to ensure consistent output. The guidance is based on the rendered sequences. Finally, the generated sequences are utilized to optimize the final 3DGS through a tailored optimization scheme." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 361, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 601 + ], + "type": "text", + "content": "ambiguity [59]. Current works can be roughly categorized into two lines of research. The first line of research focuses on pre-training generalizable radiance fields using multiview datasets. Generalizable NeRF [6, 24, 45, 55] typically projects the ray points to reference images and integrates 2D features as the auxiliary input of MLP. In contrast, generalizable 3DGS [5, 8, 22, 60] commonly perform dense depth prediction associated with gaussian properties. The second line of research leverages regularization techniques for optimization, which are applicable to both NeRF and 3DGS, including hand-crafted constraints [17, 32, 63], and those derived from pre-trained models [10, 18, 33, 35, 44, 50, 64]. In this paper, we explore the priors from video diffusion models [7, 12, 51, 57] for sparse-input 3DGS modeling. Unlike previous works, our method trains sparse-input 3DGS using augmented sequences from video diffusion models, which provide interpolation or extrapolation around the input views. We design a guidance strategy that tames the video diffusion model to generate more plausible and scene-grounded sequences, greatly enhancing the performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "Diffusion Prior for Radiance Fields. Diffusion models [11, 36, 39] have shown remarkable generation capabilities. The strong prior knowledge embedded in diffusion models can facilitate the training of radiance fields. Specifically, several works leverage Score Distillation Sampling (SDS) [19, 20, 34, 48] using a frozen diffusion model to train a radiance field based on text prompts in a zero-shot manner. Other studies focus on training view-consistent [37, 42, 49] or quality-enhanced [23] diffusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 361, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 433 + ], + "type": "text", + "content": "models, where the generated results can be directly applied to train a radiance field. Our method also uses the generated results to train a radiance field. However, our primary contribution lies in a novel guidance strategy that controls the generation to be consistent, which is crucial for sparse-input modeling augmented with generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 433, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 649 + ], + "type": "text", + "content": "Controllable Generation for Diffusion Models. Current works of controllable generation can be categorized into training-required and training-free methods. Training-required methods fine-tune the diffusion models with additional conditions [30, 61], or train an additional noise-conditioned external guidance function, e.g., classifier guidance [9] for denosing sampler. Training-free methods freeze the foundation diffusion model, and modify the denoising process with the control signal from the external guidance functions [1, 38, 53, 56]. These methods do not require training additional guidance functions or fine-tuning diffusion models; instead, they enable controllable generation in a plug-and-play manner. Our work is inspired by training-free methods. However, we concentrate on multi-view modeling, which necessitates high consistency control over the generated results. Besides, unlike these methods that typically rely on pre-trained models for guidance, we utilize rendered results to provide scene-grounding guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 657, + 443, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 443, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 443, + 671 + ], + "type": "text", + "content": "3. The Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we utilize video diffusion models to tackle two critical issues in real-world sparse-input modeling: extrapolation and occlusion, as illustrated in Fig. 1. The overview" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6135" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "of our method is illustrated in Fig. 2, which consists of three proposed components: a scene-grounding guidance (Sec. 3.2), a trajectory initialization strategy (Sec. 3.3), and a scheme for 3DGS optimization with generated sequences (Sec. 3.4). We will detail these components following a preliminary review of 3DGS and diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 154, + 135, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 154, + 135, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 135, + 166 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": "3D Gaussian Splating (3DGS) [16] represents a scene with a set of anisotropic 3D Gaussian primitives. Each Gaussian primitive is parametrized by a set of attributes: a center " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": ", a scaling factor " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": ", a quaternion " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": ", an opacity value " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": ", and a feature vector " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": ". The basis function of each Gaussian primitive is formulated as " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(x) = \\exp(-\\frac{1}{2} (\\mathrm{x} - \\mu)^{\\mathrm{T}} \\Sigma^{-1} (\\mathrm{x} - \\mu))" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": " is the covariance matrix derived from " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": ". 3DGS renders the scene through a differentiable splatting, which firstly transforms the 3D Gaussian " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(x)" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": " into 2D Gaussian " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{G}'(x)" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": " on the image plane via projection [65], and applies a tile-based rasterizer for rendering, which sorts the 2D Gaussians by depth and employs the " + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 172, + 296, + 327 + ], + "type": "text", + "content": "-blending as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 328, + 295, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 328, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 67, + 328, + 295, + 361 + ], + "type": "interline_equation", + "content": "C \\left(x _ {p}\\right) = \\sum_ {i \\in K} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\quad \\sigma_ {i} = \\alpha_ {i} \\mathcal {G} _ {i} ^ {\\prime} \\left(x _ {p}\\right), \\tag {1}", + "image_path": "eb53a99828d9d936523b5b738041e0b58e867a9d274b0c75fe246c2d295322d1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "text", + "content": " is the pixel position, " + }, + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "text", + "content": " refers to the number of 2D Gaussians associated with the pixel, and " + }, + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "text", + "content": " represents the decoded color of feature " + }, + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 361, + 296, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "text", + "content": "Diffusion Models [11, 39] are a family of generative models that progressively perturb data with intensifying Gaussian noises (i.e., forward noising), and then learn to reverse this process for sample generation (i.e., reverse denoising). The key of the diffusion model is a U-Net " + }, + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "text", + "content": " which is trained to predict the noise that is injected in the current sample " + }, + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "text", + "content": ". The sampling is conducted by iterative denoising for " + }, + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 397, + 296, + 493 + ], + "type": "text", + "content": " steps [56] as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 498, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 498, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 67, + 498, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t - 1} = (1 + \\beta_ {t} / 2) \\mathbf {x} _ {t} + \\beta_ {t} \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) + \\sqrt {\\beta_ {t}} \\mathbf {z} \\quad (2)", + "image_path": "2a7e5ec2ec775b3c014f7d2cb7436de6e87402d9a3dde8081150315aca1b66ff.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "text", + "content": " is the estimated score function which can be derived from " + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t);\\beta_t" + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "text", + "content": " is pre-defined parameters; " + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{z}\\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "text", + "content": ". In this work, we leverage a camera-controlled image-to-video diffusion model [57], whose condition includes an image for the first frame, and the camera trajectory for the path of the generated sequence. The model is operated in a latent space of dimension " + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "text", + "content": ", supporting the sequence length of " + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "text", + "content": ", thus " + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t\\in \\mathbb{R}^{L\\times h\\times w\\times d}" + }, + { + "bbox": [ + 55, + 517, + 296, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 280, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 280, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 280, + 635 + ], + "type": "text", + "content": "3.2. Generation via Scene-Grounding Guidance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "content": "Applying the generated sequences from the video diffusion model can provide plausible interpretations of regions not covered by the sparse inputs. However, as illustrated in Fig. 1, the inconsistency within the generated sequences manifests as: (i) appearance inconsistencies across frames and (ii) the occurrence of non-existent elements, which can" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 74, + 550, + 85 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 74, + 550, + 85 + ], + "spans": [ + { + "bbox": [ + 315, + 74, + 550, + 85 + ], + "type": "text", + "content": "Algorithm 1 Generation with Scene-Grounding Guidance" + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 319, + 90, + 498, + 104 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 90, + 498, + 104 + ], + "spans": [ + { + "bbox": [ + 319, + 90, + 498, + 104 + ], + "type": "text", + "content": "1: Function GENERATOR " + }, + { + "bbox": [ + 319, + 90, + 498, + 104 + ], + "type": "inline_equation", + "content": "(\\mathcal{R},I,\\{\\phi_j\\}_{j = 1}^L)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 104, + 553, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 104, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 321, + 104, + 553, + 127 + ], + "type": "text", + "content": "2: Input: Optimized 3DGS model " + }, + { + "bbox": [ + 321, + 104, + 553, + 127 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 321, + 104, + 553, + 127 + ], + "type": "text", + "content": ", input image " + }, + { + "bbox": [ + 321, + 104, + 553, + 127 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 321, + 104, + 553, + 127 + ], + "type": "text", + "content": ", camera trajectory of a sequence " + }, + { + "bbox": [ + 321, + 104, + 553, + 127 + ], + "type": "inline_equation", + "content": "\\{\\phi_j\\}_{j=1}^L" + }, + { + "bbox": [ + 321, + 104, + 553, + 127 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "text", + "content": "3: Given: Latent image-to-video diffusion model " + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "text", + "content": ", VAE decoder " + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "text", + "content": ", pre-defined " + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}" + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "text", + "content": " and guidance scale " + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 321, + 127, + 553, + 150 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 151, + 513, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 151, + 513, + 164 + ], + "spans": [ + { + "bbox": [ + 321, + 151, + 513, + 164 + ], + "type": "text", + "content": "4: Abbreviate " + }, + { + "bbox": [ + 321, + 151, + 513, + 164 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t,I,\\{\\phi_j\\}_{j = 1}^L)" + }, + { + "bbox": [ + 321, + 151, + 513, + 164 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 321, + 151, + 513, + 164 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 164, + 551, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 164, + 551, + 177 + ], + "spans": [ + { + "bbox": [ + 321, + 164, + 551, + 177 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 321, + 164, + 551, + 177 + ], + "type": "inline_equation", + "content": "\\mathbf{S},\\mathbf{M} =" + }, + { + "bbox": [ + 321, + 164, + 551, + 177 + ], + "type": "text", + "content": " rasterize " + }, + { + "bbox": [ + 321, + 164, + 551, + 177 + ], + "type": "inline_equation", + "content": "(\\{\\phi_j\\}_{j = 1}^L,\\mathcal{R})" + }, + { + "bbox": [ + 321, + 164, + 551, + 177 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 321, + 164, + 551, + 177 + ], + "type": "text", + "content": " Eq. (1)& (5)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 177, + 390, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 177, + 390, + 188 + ], + "spans": [ + { + "bbox": [ + 321, + 177, + 390, + 188 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 321, + 177, + 390, + 188 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(0,\\mathbf{I})" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 189, + 413, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 189, + 413, + 198 + ], + "spans": [ + { + "bbox": [ + 321, + 189, + 413, + 198 + ], + "type": "text", + "content": "7: for " + }, + { + "bbox": [ + 321, + 189, + 413, + 198 + ], + "type": "inline_equation", + "content": "t = T, \\dots, 1" + }, + { + "bbox": [ + 321, + 189, + 413, + 198 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 200, + 473, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 200, + 473, + 211 + ], + "spans": [ + { + "bbox": [ + 321, + 200, + 473, + 211 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 321, + 200, + 473, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 321, + 200, + 473, + 211 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 321, + 200, + 473, + 211 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 321, + 200, + 473, + 211 + ], + "type": "text", + "content": ", else " + }, + { + "bbox": [ + 321, + 200, + 473, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathbf{0}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 212, + 536, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 212, + 536, + 227 + ], + "spans": [ + { + "bbox": [ + 321, + 212, + 536, + 227 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 321, + 212, + 536, + 227 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{t - 1} = (1 + \\frac{1}{2}\\beta_t)\\mathbf{x}_t - \\frac{\\beta_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\epsilon_\\theta (\\mathbf{x}_t,t) + \\sqrt{\\beta_t}\\mathbf{z}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 227, + 488, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 227, + 488, + 241 + ], + "spans": [ + { + "bbox": [ + 318, + 227, + 488, + 241 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 318, + 227, + 488, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|t} = \\frac{1}{\\sqrt{\\bar{\\alpha}_t}} (\\mathbf{x}_t - \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon_\\theta (\\mathbf{x}_t,t))" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 242, + 406, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 242, + 406, + 253 + ], + "spans": [ + { + "bbox": [ + 318, + 242, + 406, + 253 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 318, + 242, + 406, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{0|t} = \\mathcal{D}(\\mathbf{x}_{0|t})" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 253, + 551, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 253, + 551, + 265 + ], + "spans": [ + { + "bbox": [ + 318, + 253, + 551, + 265 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 318, + 253, + 551, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_t = \\nabla_{\\mathbf{x}_t}\\mathcal{L}(\\mathbf{S},\\mathbf{M},\\mathbf{X}_{0|t})" + }, + { + "bbox": [ + 318, + 253, + 551, + 265 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 253, + 551, + 265 + ], + "type": "text", + "content": " Eq. (6)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 266, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 266, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 318, + 266, + 553, + 277 + ], + "type": "text", + "content": "13: " + }, + { + "bbox": [ + 318, + 266, + 553, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1} = \\hat{\\mathbf{x}}_{t - 1} - \\gamma_t\\mathbf{g}_t" + }, + { + "bbox": [ + 318, + 266, + 553, + 277 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 266, + 553, + 277 + ], + "type": "text", + "content": " Eq. (2)& (4)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 277, + 365, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 277, + 365, + 287 + ], + "spans": [ + { + "bbox": [ + 318, + 277, + 365, + 287 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 289, + 392, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 289, + 392, + 300 + ], + "spans": [ + { + "bbox": [ + 318, + 289, + 392, + 300 + ], + "type": "text", + "content": "15: return " + }, + { + "bbox": [ + 318, + 289, + 392, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathbf{x}_0)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 315, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 373 + ], + "type": "text", + "content": "negatively impact the 3DGS modeling. In this section, we propose an innovative scene-grounding guidance method that directs the video diffusion model to generate consistent sequences, significantly enhancing the performance of sparse-input 3DGS." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 374, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 457 + ], + "type": "text", + "content": "Inspired by previous training-free guidance methods [1, 56] that achieve their objectives by modifying the sampler in Eq. (2), we adopt a similar approach to attain the goal of consistency. Specifically, we firstly replace " + }, + { + "bbox": [ + 313, + 374, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 374, + 554, + 457 + ], + "type": "text", + "content": " with a conditional score function " + }, + { + "bbox": [ + 313, + 374, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(\\mathbf{x}_t|\\mathcal{Q})" + }, + { + "bbox": [ + 313, + 374, + 554, + 457 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 374, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 374, + 554, + 457 + ], + "type": "text", + "content": " refers to the target of consistency. The conditional score function can be expanded by the Bayesian rule as:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 319, + 463, + 553, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 463, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 319, + 463, + 553, + 475 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t} | \\mathcal {Q}) = \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) + \\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathcal {Q} | \\mathbf {x} _ {t}), \\tag {3}", + "image_path": "748fd0b17f03dce232d7591392d1874cd4322c53f5841358e0acf7c2bcc16743.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(\\mathcal{Q}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "content": " can be considered as a guidance term that injects the consistency constraint into Eq. (2). We further formulate " + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "inline_equation", + "content": "p(\\mathcal{Q}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "inline_equation", + "content": "p(\\mathcal{Q}|\\mathbf{x}_t) = \\exp (-\\lambda \\mathcal{L}(\\mathcal{Q},\\mathbf{x}_t)) / Z" + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathcal{Q},\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "content": " indicates how well the current sample " + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "content": " is aligned with the target, and " + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "content": " is a normalization term. The guidance term can thus be implemented using the gradient of the following loss function:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 361, + 576, + 553, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 576, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 361, + 576, + 553, + 589 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathcal {Q} | \\mathbf {x} _ {t}) \\propto - \\nabla_ {\\mathbf {x} _ {t}} \\mathcal {L} (\\mathcal {Q}, \\mathbf {x} _ {t}), \\tag {4}", + "image_path": "1f38d90ca0e19b3d550dfab99d9a264aabfb4034a7c9571531e6523b1887cc81.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "type": "text", + "content": "which is appended to Eq. (2) to achieve the target of consistency during the denoising sampling." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "The remaining problem lies in how to define the consistency target " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": ". Unlike previous works [1, 38, 56] that define the target based on external pre-trained models, we establish the target using a rendered sequence from an optimized 3DGS model " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": ". Though the rendered sequence is not perfect, our key insights are as follows: (i) the rendered images of adjacent frames are highly consistent, as the camera movement between them is typically minor; (ii)" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6136" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "the rendered frames provide scene grounding, clearly indicating which elements exist in the scene. Therefore, the rendered sequence can serve as an effective guidance for the generated sequence to achieve the target of consistency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": "Given a camera trajectory " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\{\\phi_j\\}_{j = 1}^L" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": " for the sequence generation, we first utilize the optimized 3DGS to render a sequence " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\{S_j\\}_{j = 1}^L" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": ", along with a mask sequence " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\{M_j\\}_{j = 1}^L" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": " that indicates the regions not covered by the sparse inputs. To get the mask, we first render a transmittance map, which is obtained by " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": "-blending (as Eq. (1)) on the opacity. For each pixel " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": "-blending is formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 208, + 295, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 208, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 114, + 208, + 295, + 240 + ], + "type": "interline_equation", + "content": "O \\left(x _ {p}\\right) = \\sum_ {i \\in K} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\tag {5}", + "image_path": "dbc3db87158d63ca5069413525cd1f34dcb08e4ff8c963a94af48329f9400f8e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " refer to the opacity of the gaussian and the transmittance map, respectively. The mask is then acquired by thresholding the transmittance map with a value " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{mask}}" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "M = (O < \\eta_{\\mathrm{mask}})" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": ". For convenience, we stack " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\{S_j\\}_{j=1}^L" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\{M_j\\}_{j=1}^L" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": ", which are of shape " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "L \\times H \\times W \\times 3" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "L \\times H \\times W \\times 1" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": ", respectively. Since the target of consistency is based on the rendered sequence in clean data space, to receive the guidance, we transform the noisy latent " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " into a latent " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|t}" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " in the clean data space, based on prediction from the model " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|t} = (\\mathbf{x}_t - \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon_\\theta(\\mathbf{x}_t, t)) / \\sqrt{\\bar{\\alpha}_t}" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": ". With the consistency target " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " that is based on the rendered sequence " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": ", we formulate the function " + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 241, + 296, + 397 + ], + "type": "text", + "content": " in the guidance term (Eq. (4)) as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 399, + 295, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 399, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 67, + 399, + 295, + 429 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\mathbf {S}, \\mathbf {M}, \\mathbf {X} _ {0 | t}) = \\| \\bar {\\mathbf {M}} \\odot (\\mathbf {S} - \\mathbf {X} _ {0 | t}) \\| _ {1} \\tag {6} \\\\ + \\lambda_ {\\text {p e r c}} \\mathcal {L} _ {\\text {p e r c}} (\\bar {\\mathbf {M}} \\odot \\mathbf {S}, \\bar {\\mathbf {M}} \\odot \\mathbf {X} _ {0 | t}), \\\\ \\end{array}", + "image_path": "68f8a157b1f1f55b55de33102129063abe69b0f9ca3ae4e2bab3e2d5cef01600.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{0|t}" + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "text", + "content": " is decoded from the latent " + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0|t}" + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "text", + "content": " by a VAE decoder, " + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{M}}" + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "text", + "content": " is the result of flipping the values in " + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "text", + "content": " is the Hadamard product, and " + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{perc}}" + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "text", + "content": " is a perceptual loss [15] with its corresponding weight as " + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{perc}}" + }, + { + "bbox": [ + 55, + 432, + 296, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 553 + ], + "type": "text", + "content": "With the guidance from Eq. (6), the denoising process balances the consistency constraint and the prior from the diffusion model, integrating them into plausible generation results. This guidance does not involve any fine-tuning of the diffusion model, thereby preserving its generative capabilities. The detailed pipeline is outlined in Alg. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 233, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 233, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 233, + 576 + ], + "type": "text", + "content": "3.3. Trajectory Initialization Strategy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "To enable holistic modeling of the scene, the camera trajectories for the video diffusion model should cover regions that are outside the field of view or occluded as much as possible. The generated sequences can thus provide plausible interpretations for these regions, which serve as the basis for optimizing the subsequent 3DGS model. Similar to the scene-grounding guidance discussed in Sec. 3.2, the proposed trajectory initialization method is also based on an optimized 3DGS model. For the " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "-th sparse input view with camera pose " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\varphi_{i}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", we first sample a set of candidate poses around it, as depicted in Fig. 3. Suppose" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 550, + 230 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 550, + 230 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 550, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 550, + 230 + ], + "type": "image", + "image_path": "d1ef091fba6b0377cc0b70a326fbc2fcf34cfcd09333ecee5591c04b445a6485.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 238, + 554, + 316 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 554, + 316 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed trajectory initialization strategy. The yellow parts represent unobserved regions. For each input view, we sample a set of candidate poses around it, and render at these poses using an optimized 3DGS. We select candidate poses whose renderings exhibit significant holes (highlighted by red boxes), and interpolate trajectories between these candidate poses and the input view's pose." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": "there are a total of " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " candidate poses, we use the optimized 3DGS model " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " to render images for these poses as " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\{\\hat{S}_c^{(i)},\\hat{M}_c^{(i)}\\}_{c = 1}^m =" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " rasterize " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "(\\{\\hat{\\phi}_c^{(i)}\\}_{c = 1}^m,\\mathcal{R})" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " . For poses where the rendered images exhibit significant black holes, as indicated by the mask " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " calculated from Eq. (5), we interpolate a trajectory of length " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " (matching the length of video diffusion model) between the input camera pose and these poses as follows: " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\{\\phi_j^{(i,c)}\\}_{j = 1}^L = \\mathrm{interp}(\\varphi_i,\\hat{\\phi}_c^{(i)})" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_c^{(i)}" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " refers to one selected candidate pose from the " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " -th input. In practice, we select the top- " + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 323, + 555, + 485 + ], + "type": "text", + "content": " candidate poses based on the sizes of their corresponding masks. Then, we build a trajectory pool by traversing all input views and their respective selected candidate poses as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 384, + 489, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 489, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 384, + 489, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\Phi = \\left\\{\\left\\{\\phi_ {j} ^ {(i, c)} \\right\\} _ {j = 1} ^ {L} \\mid i, c \\right\\}, \\tag {7}", + "image_path": "84002709c9ef7930539781dc10a6b2ff043bd8ab0ca03df6e953d9de019f5b79.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 507, + 554, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 554, + 531 + ], + "type": "text", + "content": "where each element in the pool is sampled for the sequence generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 538, + 509, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 509, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 509, + 550 + ], + "type": "text", + "content": "3.4. 3DGS Optimization with Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "type": "text", + "content": "Given sparse inputs of " + }, + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "type": "text", + "content": " images along with their poses, i.e., " + }, + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\{C_i^{\\mathrm{gt}},\\varphi_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 555, + 554, + 662 + ], + "type": "text", + "content": ", we aim at optimizing a 3DGS model with the auxiliary generated sequences. For simplicity, we refer to the input images paired with their poses as 'input views', and we term the generated images with their associated poses as 'generated views'. During each iteration, we sample an input view and a generated view for supervision. Specifically, for the input view, we employ the default reconstruction loss [16] written as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 669, + 553, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 669, + 553, + 684 + ], + "spans": [ + { + "bbox": [ + 319, + 669, + 553, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {i n p u t}} = (1 - \\lambda) \\mathcal {L} _ {1} \\left(C _ {i}, C _ {i} ^ {\\mathrm {g t}}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(C _ {i}, C _ {i} ^ {\\mathrm {g t}}\\right), \\tag {8}", + "image_path": "7c55af3feb0cf2c384ad255a32f851062cf3ac66dc2a24f5d1f0e51952ceacb9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " refers to the rendered image and " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " is a weighting factor. For supervision of generated views, we find that the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6137" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 260, + 85 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 260, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 260, + 85 + ], + "type": "text", + "content": "Algorithm 2 3DGS Optimization with Generation" + } + ] + } + ], + "index": 0, + "type": "text" + }, + { + "bbox": [ + 61, + 91, + 294, + 354 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 62, + 91, + 268, + 104 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 91, + 268, + 104 + ], + "spans": [ + { + "bbox": [ + 62, + 91, + 268, + 104 + ], + "type": "text", + "content": "1: Input: Sparse inputs of N images " + }, + { + "bbox": [ + 62, + 91, + 268, + 104 + ], + "type": "inline_equation", + "content": "\\{C_i^{\\mathrm{gt}},\\varphi_i\\}_{i = 1}^N" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 104, + 294, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 104, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 62, + 104, + 294, + 128 + ], + "type": "text", + "content": "2: Given: Number of iterations " + }, + { + "bbox": [ + 62, + 104, + 294, + 128 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{iter}}" + }, + { + "bbox": [ + 62, + 104, + 294, + 128 + ], + "type": "text", + "content": ", generation interval " + }, + { + "bbox": [ + 62, + 104, + 294, + 128 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{gen}}" + }, + { + "bbox": [ + 62, + 104, + 294, + 128 + ], + "type": "text", + "content": ", ratio of samples from other sequences " + }, + { + "bbox": [ + 62, + 104, + 294, + 128 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 62, + 104, + 294, + 128 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 128, + 272, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 128, + 272, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 128, + 272, + 140 + ], + "type": "text", + "content": "3: Variable: Global list of generated views " + }, + { + "bbox": [ + 62, + 128, + 272, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{G} = []" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 140, + 241, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 140, + 241, + 151 + ], + "spans": [ + { + "bbox": [ + 62, + 140, + 241, + 151 + ], + "type": "text", + "content": "4: Baseline 3DGS model optimization " + }, + { + "bbox": [ + 62, + 140, + 241, + 151 + ], + "type": "inline_equation", + "content": "\\Rightarrow \\mathcal{R}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 152, + 292, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 152, + 292, + 163 + ], + "spans": [ + { + "bbox": [ + 62, + 152, + 292, + 163 + ], + "type": "text", + "content": "5: Trajectory initialization " + }, + { + "bbox": [ + 62, + 152, + 292, + 163 + ], + "type": "inline_equation", + "content": "\\Rightarrow \\Phi" + }, + { + "bbox": [ + 62, + 152, + 292, + 163 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 62, + 152, + 292, + 163 + ], + "type": "text", + "content": " Eq. (7)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 164, + 185, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 164, + 185, + 175 + ], + "spans": [ + { + "bbox": [ + 62, + 164, + 185, + 175 + ], + "type": "text", + "content": "6: for " + }, + { + "bbox": [ + 62, + 164, + 185, + 175 + ], + "type": "inline_equation", + "content": "t = 0, \\dots, N_{\\mathrm{iter}} - 1" + }, + { + "bbox": [ + 62, + 164, + 185, + 175 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 176, + 170, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 176, + 170, + 188 + ], + "spans": [ + { + "bbox": [ + 62, + 176, + 170, + 188 + ], + "type": "text", + "content": "7: If " + }, + { + "bbox": [ + 62, + 176, + 170, + 188 + ], + "type": "inline_equation", + "content": "t\\% N_{\\mathrm{gen}} = 0" + }, + { + "bbox": [ + 62, + 176, + 170, + 188 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 188, + 189, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 188, + 189, + 198 + ], + "spans": [ + { + "bbox": [ + 62, + 188, + 189, + 198 + ], + "type": "text", + "content": "8: Sample an input view " + }, + { + "bbox": [ + 62, + 188, + 189, + 198 + ], + "type": "inline_equation", + "content": "I" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 199, + 291, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 291, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 291, + 213 + ], + "type": "text", + "content": "9: Sample a trajectory around " + }, + { + "bbox": [ + 62, + 199, + 291, + 213 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 62, + 199, + 291, + 213 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 62, + 199, + 291, + 213 + ], + "type": "inline_equation", + "content": "\\Phi \\Rightarrow \\{\\phi_j\\}_{j=1}^L" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 213, + 241, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 241, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 241, + 225 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 61, + 213, + 241, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\mathrm{GENERATOR}(\\mathcal{R},I,\\{\\phi_j\\}_{j = 1}^L)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 226, + 157, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 226, + 157, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 226, + 157, + 236 + ], + "type": "text", + "content": "11: Append S to G" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 238, + 113, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 238, + 113, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 238, + 113, + 247 + ], + "type": "text", + "content": "12: End If" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 249, + 292, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 249, + 292, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 249, + 292, + 261 + ], + "type": "text", + "content": "13: Sample an input view to get " + }, + { + "bbox": [ + 61, + 249, + 292, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{input}}" + }, + { + "bbox": [ + 61, + 249, + 292, + 261 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 61, + 249, + 292, + 261 + ], + "type": "text", + "content": " Eq. (8)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 262, + 162, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 262, + 162, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 162, + 272 + ], + "type": "text", + "content": "14: If rand() " + }, + { + "bbox": [ + 61, + 262, + 162, + 272 + ], + "type": "inline_equation", + "content": "\\geq \\eta" + }, + { + "bbox": [ + 61, + 262, + 162, + 272 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 273, + 225, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 273, + 225, + 284 + ], + "spans": [ + { + "bbox": [ + 61, + 273, + 225, + 284 + ], + "type": "text", + "content": "15: Sample a generated view from S" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 61, + 285, + 238, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 238, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 238, + 296 + ], + "type": "text", + "content": "16: Else Sample a generated view from G" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 61, + 297, + 113, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 113, + 307 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 113, + 307 + ], + "type": "text", + "content": "17: End If" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 61, + 309, + 292, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 309, + 292, + 320 + ], + "spans": [ + { + "bbox": [ + 61, + 309, + 292, + 320 + ], + "type": "text", + "content": "18: Use the generated view to get " + }, + { + "bbox": [ + 61, + 309, + 292, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{gen}}" + }, + { + "bbox": [ + 61, + 309, + 292, + 320 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 61, + 309, + 292, + 320 + ], + "type": "text", + "content": " Eq. (9)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 61, + 321, + 199, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 199, + 332 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 199, + 332 + ], + "type": "text", + "content": "19: " + }, + { + "bbox": [ + 61, + 321, + 199, + 332 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}^{\\mathrm{input}} + \\mathcal{L}^{\\mathrm{gen}})" + }, + { + "bbox": [ + 61, + 321, + 199, + 332 + ], + "type": "text", + "content": " .backward()" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 61, + 333, + 216, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 333, + 216, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 333, + 216, + 343 + ], + "type": "text", + "content": "20: #Densification and opacity reset" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 61, + 346, + 105, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 346, + 105, + 354 + ], + "spans": [ + { + "bbox": [ + 61, + 346, + 105, + 354 + ], + "type": "text", + "content": "21: end for" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 367, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 462 + ], + "type": "text", + "content": "reconstruction loss does not effectively fill the hole regions, and increasing its weight leads to performance degradation due to flaws in the generated images. To address this issue, we propose using perceptual loss [15]. The perceptual loss is calculated over the entire image, allowing those hole regions to significantly influence the gradients, thereby effectively driving the model to fill those holes. Thus, the loss on the generated views is formulated as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 71, + 471, + 294, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 471, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 71, + 471, + 294, + 485 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {g e n}} = \\lambda_ {\\text {g e n 1}} \\mathcal {L} _ {1} \\left(C _ {j}, S _ {j}\\right) + \\lambda_ {\\text {g e n 2}} \\mathcal {L} _ {\\text {p e r c}} \\left(C _ {j}, S _ {j}\\right), \\tag {9}", + "image_path": "668a81ed17c949c29d771d12da7e1a762c39468e0bf8df1b297bef421c1e9e63.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "text", + "content": " refers to the generated image, " + }, + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{gen1}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{gen2}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "text", + "content": " are two balancing factors, respectively." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 517, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 636 + ], + "type": "text", + "content": "We empirically find that conducting local sampling within a specific optimization interval, where a substantial portion of generated views is sampled from the same sequence of local regions, enhances visual quality. However, sampling exclusively from a single sequence can lead to a forgetting issue, where optimized information about holes in other regions becomes diluted. Therefore, within each interval of local sampling, we also include generated views from other sequences with a ratio " + }, + { + "bbox": [ + 55, + 517, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 517, + 296, + 636 + ], + "type": "text", + "content": ". The optimization pipeline is presented in Alg. 2." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 641, + 135, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 135, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 135, + 654 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 660, + 176, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 176, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 176, + 673 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Datasets and Metrics. We target addressing the issues of extrapolation and occlusion for sparse-input 3DGS scene modeling, which are overlooked by current benchmarks. To" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 71, + 390, + 79 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 71, + 390, + 79 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 390, + 79 + ], + "type": "text", + "content": "w/ Vanilla Generation" + } + ] + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 318, + 79, + 393, + 137 + ], + "blocks": [ + { + "bbox": [ + 318, + 79, + 393, + 137 + ], + "lines": [ + { + "bbox": [ + 318, + 79, + 393, + 137 + ], + "spans": [ + { + "bbox": [ + 318, + 79, + 393, + 137 + ], + "type": "image", + "image_path": "1537861a4c58d25b5858af5487478160c838e5ec1cb31f2e7d8a653ff32f50c9.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 139, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 313, + 139, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 553, + 182 + ], + "type": "text", + "content": "Figure 4. Sequences from the vanilla generation suffer from inconsistencies. A 3DGS model optimized with these sequences renders images with black shadows, highlighted by red boxes, while our method solves this issue with the scene-grounding guidance." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 403, + 71, + 471, + 79 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 71, + 471, + 79 + ], + "spans": [ + { + "bbox": [ + 403, + 71, + 471, + 79 + ], + "type": "text", + "content": "w/ Guided Generation" + } + ] + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 395, + 80, + 472, + 137 + ], + "blocks": [ + { + "bbox": [ + 395, + 80, + 472, + 137 + ], + "lines": [ + { + "bbox": [ + 395, + 80, + 472, + 137 + ], + "spans": [ + { + "bbox": [ + 395, + 80, + 472, + 137 + ], + "type": "image", + "image_path": "3984260e981f383a62bb0a63b6b2b8a8aeec30db5060c0f6c27955d2f2e114c4.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 490, + 71, + 533, + 78 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 490, + 71, + 533, + 78 + ], + "spans": [ + { + "bbox": [ + 490, + 71, + 533, + 78 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 475, + 80, + 550, + 137 + ], + "blocks": [ + { + "bbox": [ + 475, + 80, + 550, + 137 + ], + "lines": [ + { + "bbox": [ + 475, + 80, + 550, + 137 + ], + "spans": [ + { + "bbox": [ + 475, + 80, + 550, + 137 + ], + "type": "image", + "image_path": "9aa4ee5f3f265f18b84669cd92f1f5fe9e3984d8821aaa6ea64696a00ef24ed5.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 191, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 553, + 311 + ], + "type": "text", + "content": "evaluate the effectiveness of our method, we conduct experiments on a new benchmark created from two indoor datasets, i.e., the synthetic Replica [40] and the realistic ScanNet++ [54], which consists of 6 and 4 scenes, respectively. Although the selected six input views for each scene can cover most regions, there are still areas outside the field of view. Moreover, the 'inside-out' viewing directions make occlusion common in this benchmark. More details can be found in the supplement. For quantitative comparisons, we report PSNR, SSIM [47], and LPIPS [62] scores." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 312, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 384 + ], + "type": "text", + "content": "Baseline. We train a baseline 3DGS model initialized with the point cloud from DUSt3R [46], incorporating the gaussian unpooling in FSGS [64], which makes the optimized model a strong baseline. Based on this we conduct experiments to verify the effectiveness of our method. The model is denoted as 'Baseline 3DGS' in the following." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": "Implementation Details. The baseline model described above serves as the model " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": " for scene-grounding guidance and trajectory initialization. For sequence generation, we employ the camera-controlled image-to-video diffusion model [57] which supports the generation of " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "L = 25" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": " frames. The weighting factors, " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{perc}}" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{gen1}}" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{gen2}}" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "0.2" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "0.1" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "0.01" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": ", respectively. The threshold of " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{mask}}" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "0.9" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "0.5" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": ". The generation interval " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{gen}}" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "260" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{iter}}" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "inline_equation", + "content": "10,000" + }, + { + "bbox": [ + 313, + 384, + 554, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 502, + 399, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 399, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 399, + 515 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 521, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 689 + ], + "type": "text", + "content": "Comparison on Replica. As shown in Tab. 1, our method achieves the highest performance on the Replica dataset, outperforming DNGaussian [18] and FSGS [64] by a significant margin of over " + }, + { + "bbox": [ + 313, + 521, + 554, + 689 + ], + "type": "inline_equation", + "content": "3.0\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 521, + 554, + 689 + ], + "type": "text", + "content": " in PSNR. Fig. 5 illustrates that our method effectively addresses occlusion and extrapolation, while other 3DGS-based methods struggle with these challenges. Additionally, their depth regularization often compromise thin structures, such as the flower in the vase in the second row. FreeNeRF [52] exhibits severe artifacts because it cannot effectively utilize the strong prior from the DUSt3R point cloud. Although FreeNeRF can fill hole regions through neighboring interpolation (e.g., the wall behind the chair in the first row and the ceiling in the second row), the results frequently exhibit blurring or artifacts." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Comparison on ScanNet++. ScanNet++ is a dataset captured in realistic scenes, so it is more complicated and chal" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6138" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 539, + 339 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 539, + 339 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 539, + 339 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 539, + 339 + ], + "type": "image", + "image_path": "2f90b6a651d05826be96b312818235e9c13bb555a2d8087df277def0cd44845e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 380, + 296, + 513 + ], + "blocks": [ + { + "bbox": [ + 59, + 380, + 296, + 513 + ], + "lines": [ + { + "bbox": [ + 59, + 380, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 59, + 380, + 296, + 513 + ], + "type": "table", + "html": "
MethodReplica [40]ScanNet++ [54]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Mip-NeRF [2]18.120.7070.39119.580.7550.389
InfoNeRF [17]13.070.5980.55214.540.6460.495
DietNeRF [13]18.990.6760.44419.760.7190.431
FreeNeRF [52]20.990.7650.32420.170.7560.368
3DGS†[16]22.800.8180.17921.410.8170.211
DNGaussian [18]17.630.7180.43519.010.7540.367
DNGaussian†[18]22.710.8210.18920.680.7880.281
FSGS [64]20.220.7600.30417.950.7300.373
FSGS†[64]22.990.8330.20521.230.8130.257
Ours26.350.8720.12223.890.8500.182
", + "image_path": "646482c860e183294032331487e2c29a18b1a4fe28ef7576fb92bb1550dc1a01.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on the Replica and ScanNet++ datasets. Including our approach, 3DGS-based methods marked with " + }, + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\updownarrow" + }, + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "type": "text", + "content": " are initialized with the point cloud from DUSt3R [46]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 551, + 296, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 296, + 671 + ], + "type": "text", + "content": "lenging than the synthetic Replica dataset. The results in Tab. 1 demonstrate that our method has a clear advantage over current approaches, surpassing FSGS by more than " + }, + { + "bbox": [ + 55, + 551, + 296, + 671 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 551, + 296, + 671 + ], + "type": "text", + "content": " in PSNR. As depicted in Fig. 5, our method effectively addresses the extrapolation issue (e.g., the ceiling in the fourth row) and mitigates needle-like artifacts observed in the rendered images of DNGaussian [18] and FSGS [64] (the third row). Furthermore, the comparisons in the third row highlight our method's superiority in preserving finer details compared to all other methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 675, + 157, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 157, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 157, + 686 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Our technical contributions consist of three key components. We analyze their effects on the Replica dataset." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 373, + 555, + 502 + ], + "blocks": [ + { + "bbox": [ + 55, + 340, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 555, + 373 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons on the Replica and ScanNet++ datasets. All 3DGS-based methods are optimized using the initialized point cloud from DUSt3R [46]. Our method effectively addresses the issues of extrapolation and occlusion while preserving finer details and reducing artifacts. For better visualization, please zoom in on the results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 373, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 316, + 373, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 555, + 502 + ], + "type": "image", + "image_path": "8938fcdfba6d9110e773d45996333126be74c9a43c506768960b25c6d0947339.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "type": "text", + "content": "Figure 6. Our method not only effectively addresses extrapolation and occlusion (red boxes), improving the overall quality (blue boxes), but also predicts more plausible geometry." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Generation with Scene-Grounding Guidance. Optimizing a 3DGS with sequences from vanilla generation results in quality degradation. In Tab. 2 (a), while the full image metrics are enhanced due to slightly improved modeling at occluded regions, the visual quality degrades, as indicated by PSNR of observable regions dropping from " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "25.45\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "25.00\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": ". This degradation is attributed to inconsistencies within generated sequences, which can result in black shadows in rendered images as illustrated in Fig. 4. In contrast, our scene-grounding guidance ensures that the generated sequences remain consistent, significantly enhancing the modeling capability in regions outside the field of view and occluded, while also improving the overall quality, evidenced by the 'w/ Guided Generation' results in Tab. 2 (a)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6139" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 71, + 380, + 140 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 380, + 140 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 380, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 380, + 140 + ], + "type": "table", + "html": "
(a)Gen.Guide.Traj.Full ImageObservable Regions
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Baseline 3DGS22.800.8180.17925.450.8600.129
w/ Vanilla Generation✓23.690.8400.16025.000.8700.119
w/ Guided Generation✓✓25.030.8520.13926.520.8810.101
w/ Guided Generation&Traj.✓✓✓25.580.8590.13826.530.8830.100
", + "image_path": "c1133b71fc460ada0e2a7bed5e42bc8af506832f33cec42d376229107ea0a765.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 384, + 71, + 555, + 140 + ], + "blocks": [ + { + "bbox": [ + 384, + 71, + 555, + 140 + ], + "lines": [ + { + "bbox": [ + 384, + 71, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 384, + 71, + 555, + 140 + ], + "type": "table", + "html": "
(b)PSNR↑SSIM↑LPIPS↓
Baseline 3DGS22.800.8180.179
w/ Guided Generation&Traj.25.580.8590.138
w/ perceptual loss26.350.8720.122
w/o local sampling26.280.8710.127
w/o global list26.010.8670.122
", + "image_path": "1ccbb7f1f0417432ea9a8f7d4f0c28c5258e374b550890e591e3353ef7a3a182.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 191, + 135, + 249 + ], + "blocks": [ + { + "bbox": [ + 67, + 180, + 117, + 189 + ], + "lines": [ + { + "bbox": [ + 67, + 180, + 117, + 189 + ], + "spans": [ + { + "bbox": [ + 67, + 180, + 117, + 189 + ], + "type": "text", + "content": "Baseline 3DGS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 180, + 219, + 190 + ], + "lines": [ + { + "bbox": [ + 132, + 180, + 219, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 219, + 190 + ], + "type": "text", + "content": "w/ Guided Genration&Traj." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 191, + 135, + 249 + ], + "lines": [ + { + "bbox": [ + 58, + 191, + 135, + 249 + ], + "spans": [ + { + "bbox": [ + 58, + 191, + 135, + 249 + ], + "type": "image", + "image_path": "c28a4bff8d684a5dcb1ef8a36a8b73af00d4928123255edeb3373fc5082eb13b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 295, + 272 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 272 + ], + "type": "text", + "content": "Figure 7. The perceptual loss for generated views greatly increases the modeling capability at hole regions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 136, + 191, + 212, + 249 + ], + "blocks": [ + { + "bbox": [ + 136, + 191, + 212, + 249 + ], + "lines": [ + { + "bbox": [ + 136, + 191, + 212, + 249 + ], + "spans": [ + { + "bbox": [ + 136, + 191, + 212, + 249 + ], + "type": "image", + "image_path": "556a5d5d14b0e3d3d91a798cc8d95e9552d71d192dd8fe3a9141006973072802.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 215, + 191, + 291, + 249 + ], + "blocks": [ + { + "bbox": [ + 225, + 180, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 225, + 180, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 225, + 180, + 286, + 190 + ], + "type": "text", + "content": "w/ Perceptual Loss" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 215, + 191, + 291, + 249 + ], + "lines": [ + { + "bbox": [ + 215, + 191, + 291, + 249 + ], + "spans": [ + { + "bbox": [ + 215, + 191, + 291, + 249 + ], + "type": "image", + "image_path": "83f0dd9468285be22122a5df6b9c290869192250138ec402ad5363aa239f19fd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 67, + 276, + 282, + 329 + ], + "blocks": [ + { + "bbox": [ + 55, + 142, + 555, + 176 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 555, + 176 + ], + "type": "text", + "content": "Table 2. Ablation experiments on the Replica dataset. (a) Effectiveness of the proposed scene-grounding guidance (Guide.) for generation, and the trajectory initialization strategy (Traj.). (Gen.) indicates utilizing generated sequences for modeling. Metrics of observable regions mask out regions outside the field of view or occluded. (b) Effectiveness of the proposed scheme for 3DGS optimization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 276, + 282, + 329 + ], + "lines": [ + { + "bbox": [ + 67, + 276, + 282, + 329 + ], + "spans": [ + { + "bbox": [ + 67, + 276, + 282, + 329 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS↓
Baseline 3DGS+LaMa [41]24.560.8330.167
Baseline 3DGS+SDS [34] with SDInpaint24.330.8440.153
Baseline 3DGS+SDS [34] with SDInpaint*25.150.8530.141
Ours26.350.8720.122
", + "image_path": "b1ff34b35f867a7e31a993273004cc69fc69115cf001bf307fdfb98c56593fa5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 330, + 295, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 330, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 295, + 353 + ], + "type": "text", + "content": "Table 3. Comparisons with inpainting methods on the Replica dataset. * indicates the usage of our trajectory initialization." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 364, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 364, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 54, + 364, + 295, + 471 + ], + "type": "text", + "content": "Trajectory Initialization Strategy. Tab. 2 (a) further demonstrates that the proposed trajectory initialization strategy significantly boosts the performance, notated as 'w/ Guided Generation&Traj'. The improvement mainly arises from enhanced modeling of the regions outside the field of view or the occluded areas, as the metrics of visible regions plateau while the overall image metrics improve by over 0.5 dB in PSNR. This indicates that the initialization strategy effectively identifies hole regions for holistic modeling." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 472, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 653 + ], + "type": "text", + "content": "Scheme for 3DGS Optimization with Generation. We verify the effectiveness of the proposed scheme in Tab. 2 (b). Specifically, the perceptual loss of Eq. (9) increases PSNR by over " + }, + { + "bbox": [ + 55, + 472, + 295, + 653 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 472, + 295, + 653 + ], + "type": "text", + "content": ", which is crucial for the model to fill the hole regions, as shown in Fig. 7. We empirically find that local sampling brings improvement in Sec. 3.4. This is evidenced by the performance decrease of 'w/o local sampling', which randomly samples generated views from all generated sequences. Alg. 2 shows that we use a global list to avoid the forgetting problem, and its necessity is verified by an over " + }, + { + "bbox": [ + 55, + 472, + 295, + 653 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 472, + 295, + 653 + ], + "type": "text", + "content": " PSNR drop observed with 'w/o global list'. Combining these contributions, our full model effectively addresses extrapolation and occlusion while enhancing overall image quality, meanwhile exhibiting much better geometry, as shown in Fig. 6." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 662, + 295, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 295, + 674 + ], + "type": "text", + "content": "4.4. Further Comparisons with Inpainting Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "Extrapolation and occlusion can also be addressed using inpainting methods. We thus compare our approach with two inpainting-based methods. One method applies LaMa [41]" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 191, + 392, + 305 + ], + "blocks": [ + { + "bbox": [ + 345, + 180, + 364, + 190 + ], + "lines": [ + { + "bbox": [ + 345, + 180, + 364, + 190 + ], + "spans": [ + { + "bbox": [ + 345, + 180, + 364, + 190 + ], + "type": "text", + "content": "LaMa" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 191, + 392, + 305 + ], + "lines": [ + { + "bbox": [ + 317, + 191, + 392, + 305 + ], + "spans": [ + { + "bbox": [ + 317, + 191, + 392, + 305 + ], + "type": "image", + "image_path": "765d64f45012320363a4026bdc7ec0b3d41053160dd277e3b3f23f00e4a0da0b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 308, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 351 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons with other inpainting methods. Our approach not only produces more plausible appearances around the inpainting regions but also predicts more consistent geometries in fine-grained local areas." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 395, + 191, + 471, + 305 + ], + "blocks": [ + { + "bbox": [ + 402, + 180, + 469, + 190 + ], + "lines": [ + { + "bbox": [ + 402, + 180, + 469, + 190 + ], + "spans": [ + { + "bbox": [ + 402, + 180, + 469, + 190 + ], + "type": "text", + "content": "SDS with SDInpaint*" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 191, + 471, + 305 + ], + "lines": [ + { + "bbox": [ + 395, + 191, + 471, + 305 + ], + "spans": [ + { + "bbox": [ + 395, + 191, + 471, + 305 + ], + "type": "image", + "image_path": "be06c768b96b4c9495a667ba7c86f35f83435b2414f25ad3a9408f6338da935f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 474, + 191, + 550, + 305 + ], + "blocks": [ + { + "bbox": [ + 506, + 180, + 523, + 190 + ], + "lines": [ + { + "bbox": [ + 506, + 180, + 523, + 190 + ], + "spans": [ + { + "bbox": [ + 506, + 180, + 523, + 190 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 191, + 550, + 305 + ], + "lines": [ + { + "bbox": [ + 474, + 191, + 550, + 305 + ], + "spans": [ + { + "bbox": [ + 474, + 191, + 550, + 305 + ], + "type": "image", + "image_path": "bed2b29a0e7c0318fee352dbcbf14a6bbfb44561648ba35990edd819de56304b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 359, + 555, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 539 + ], + "type": "text", + "content": "inpainting on hole regions, while the other optimizes a 3DGS by Score Distillation Sampling (SDS) [34] based on a SDInpaint model [36]. We also incorporate our proposed trajectory initialization into the SDS method to enhance the optimization of inpainting regions. Results in Tab. 3 show that our method outperforms these two methods by more than " + }, + { + "bbox": [ + 313, + 359, + 555, + 539 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 359, + 555, + 539 + ], + "type": "text", + "content": " in PSNR. Qualitative results in Fig. 8 indicate that, under certain conditions, the SDS-based method produces inpainted areas with strange appearances, while LaMa tends to create blurring artifacts interpolated from neighboring regions. Besides, Fig. 8 also shows that our method predicts better geometry in regions with local details. The inpainting results from our approach are more plausible due to the well-designed guidance, which effectively exploits prior knowledge from the diffusion model." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 543, + 388, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 388, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 388, + 555 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we have explored to address the critical issues of extrapolation and occlusion in sparse-input 3DGS modeling. We propose using video diffusion models that provide plausible interpretations for regions that are outside the field of view and occluded. To resolve inconsistencies within generated sequences, we introduce a novel scene-grounding guidance that controls the diffusion model to generate consistent sequences without any fine-tuning. Additionally, we propose a trajectory initialization strategy to enhance holistic modeling and develop a scheme for optimizing 3DGS with generated sequences. Extensive experiments validate our approach, demonstrating that it outperforms current methods by a significant margin." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6140" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "type": "text", + "content": "Acknowledgement. This research is supported in part by the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, SAIL Research Project, HKUST-Zeekr Collaborative Research Fund, HKUST-WeBank Joint Lab Project, and Tencent Rhino-Bird Focused Research Program." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 175, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 175, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 296, + 218 + ], + "type": "text", + "content": "[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In CVPR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 220, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 296, + 262 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 264, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 264, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 264, + 295, + 297 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 294, + 331 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. ICCV, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 333, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 295, + 374 + ], + "type": "text", + "content": "[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 377, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 294, + 420 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 477 + ], + "type": "text", + "content": "[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 479, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 294, + 521 + ], + "type": "text", + "content": "[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 523, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 294, + 544 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 546, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 546, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 546, + 295, + 600 + ], + "type": "text", + "content": "[10] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 624 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "text", + "content": "[12] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[13] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In ICCV, 2021. 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[14] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "text", + "content": "[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "text", + "content": "[17] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[18] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "text", + "content": "[19] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In CVPR, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "type": "text", + "content": "[20] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "text", + "content": "[21] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, et al. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In CVPR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 400, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 443 + ], + "type": "text", + "content": "[22] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splattering reconstruction from multi-view stereo. In ECCV, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "type": "text", + "content": "[23] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs-enhancer: Enhancing unbounded 3d gaussian splattering with view-consistent 2d diffusion priors. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 521 + ], + "type": "text", + "content": "[24] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 524, + 553, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 553, + 556 + ], + "type": "text", + "content": "[25] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 558, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 611 + ], + "type": "text", + "content": "[26] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Francisco Vicente Carrasco, Markus Steinberger, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. arXiv preprint arXiv:2406.15643, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 613, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 657 + ], + "type": "text", + "content": "[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM TOG, 38 (4):1-14, 2019. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6141" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[32] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[33] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In ECCV, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "type": "text", + "content": "[34] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "text", + "content": "[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 500 + ], + "type": "text", + "content": "[38] Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. Loss-guided diffusion models for plug-and-play controllable generation. In ICML, 2023. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 502, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 294, + 533 + ], + "type": "text", + "content": "[39] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "type": "text", + "content": "[40] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 645 + ], + "type": "text", + "content": "[41] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[42] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv preprint arXiv:2402.12712, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[43] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 118, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 149 + ], + "type": "text", + "content": "[44] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 151, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 151, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 315, + 151, + 553, + 204 + ], + "type": "text", + "content": "[45] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 206, + 553, + 239 + ], + "type": "text", + "content": "[46] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 240, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 240, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 315, + 240, + 553, + 271 + ], + "type": "text", + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 273, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 273, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 315, + 273, + 553, + 315 + ], + "type": "text", + "content": "[48] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 553, + 360 + ], + "type": "text", + "content": "[49] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 361, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 553, + 392 + ], + "type": "text", + "content": "[50] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 394, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 394, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 315, + 394, + 553, + 447 + ], + "type": "text", + "content": "[51] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 449, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 449, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 315, + 449, + 553, + 480 + ], + "type": "text", + "content": "[52] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In CVPR, 2023. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 315, + 482, + 553, + 525 + ], + "type": "text", + "content": "[53] Haotian Ye, Haowei Lin, Jiaqi Han, Minkai Xu, Sheng Liu, Yitao Liang, Jianzhu Ma, James Zou, and Stefano Ermon. Tfg: Unified training-free guidance for diffusion models. arXiv preprint arXiv:2409.15761, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 315, + 526, + 553, + 558 + ], + "type": "text", + "content": "[54] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 559, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 559, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 315, + 559, + 553, + 591 + ], + "type": "text", + "content": "[55] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 315, + 593, + 553, + 624 + ], + "type": "text", + "content": "[56] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In ICCV, 2023. 2, 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 626, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 626, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 315, + 626, + 553, + 679 + ], + "type": "text", + "content": "[57] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 681, + 553, + 712 + ], + "type": "text", + "content": "[58] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6142" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 320 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[59] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 149 + ], + "type": "text", + "content": "[60] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In ECCV, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "text", + "content": "[62] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "text", + "content": "[63] Yingji Zhong, Lanqing Hong, Zhenguo Li, and Dan Xu. Cvtxrf: Contrastive in-voxel transformer for 3d consistent radiance fields from sparse inputs. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "text", + "content": "[64] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, 2024. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 287, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 320 + ], + "type": "text", + "content": "[65] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002. 4" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6143" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_content_list.json b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d8c568d9ccff7488465ba437fd730d1e27f09fef --- /dev/null +++ b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_content_list.json @@ -0,0 +1,1401 @@ +[ + { + "type": "text", + "text": "TaoAvatar: Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 189, + 128, + 810, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianchuan Chen,\\* Jingchuan Hu,\\* Gaige Wang,\\* Zhonghua Jiang \nTiansong Zhou, Zhiwen Chen\\*, Chengfei Lv\\* \nAlibaba Group, Hangzhou, China", + "bbox": [ + 230, + 203, + 764, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6de75c108a6d16181aa0ae30ea0d797e628fcf6f4302f98a0b4aaec16eb8a37.jpg", + "image_caption": [ + "Figure 1. TaoAvatar generates photorealistic, topology-consistent 3D full-body avatars from multi-view sequences. It provides high-quality, real-time rendering with low storage requirements, compatible across various mobile and AR devices like the Apple Vision Pro." + ], + "image_footnote": [], + "bbox": [ + 91, + 256, + 679, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33d932f309550bd7b5d915d8ec306113b28eb2dd54dd2b3da55006a9f95d3678.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 257, + 901, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 477, + 326, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Realistic 3D full-body talking avatars hold great potential in AR, with applications ranging from e-commerce live streaming to holographic communication. Despite advances in 3D Gaussian Splatting (3DGS) for lifelike avatar creation, existing methods struggle with fine-grained control of facial expressions and body movements in full-body talking tasks. Additionally, they often lack sufficient details and cannot run in real-time on mobile devices. We present TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar driven by various signals. Our approach starts by creating a personalized clothed human parametric template that binds Gaussians to represent appearances. We then pre-train a StyleUnet-based network to handle complex pose-dependent non-rigid deformation, which can capture high-frequency appearance details but is too resource-intensive for mobile devices. To overcome this, we \"bake\" the non-rigid deformations into a lightweight MLP-based network using a distillation technique and develop blend shapes to compensate for details. Extensive experiments show that TaoAvatar achieves state-of-the-art rendering quality while running in real-time across various devices, maintaining 90 FPS on high-definition stereo devices such as the Apple Vision Pro.", + "bbox": [ + 86, + 510, + 485, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 477, + 643, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating lifelike 3D human avatars is a dynamic and rapidly advancing research area in computer vision and graphics, essential for applications in AR/VR, 3D entertainment, and virtual communication. Despite decades of research, achieving realistic and immersive avatars on mobile devices—particularly AR—remains challenging due to computational limitations. Current industry solutions, such as MetaHuman, often rely on high-precision scans and extensive manual effort by artists for 3D modeling, rigging, and motion capture.", + "bbox": [ + 511, + 505, + 906, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, academic researchers have adopted parametric models [37, 43] for human representation, which are constructed from massive body scans and can capture a performer's shape, expression, and pose from images or videos. These methods [4-6, 27] extend the naked model [37, 43] by adding offsets to represent tailored clothing, but are still struggling with complex geometries and high-frequency details, such as loose skirts and fine hair, due to limitations of topology and texture resolution. With the rise of neural radiance fields (NeRF) [40], many researchers tend to define implicit neural human representations in canonical space, animated by backward linear blend skinning [44, 65] of these parametric models [37, 43]. While these methods can handle arbitrary topology and deliver higher rendering quality, they are highly time-consuming due to the large amount of sampling required", + "bbox": [ + 511, + 659, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. ‡Project Leader and †Corresponding author. Project: https://PixelAI-Team.github.io/TaoAvatar", + "bbox": [ + 89, + 875, + 482, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10723", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "during volume rendering. Although many efforts [9, 24] have significantly improved rendering efficiency by introducing iNGP [42], the animated results still fail to escape the Uncanny Valley effect when driven. Recently, 3D Gaussian Splatting (3DGS) [29], an explicit and efficient point-based representation, has gained tremendous attention from researchers for its ability to deliver both high-quality rendering and real-time performance. Compared to implicit representations [40, 54], the explicit point-based representation is more compatible with being integrated with the mesh-based parametric models [37, 43], which can be directly driven by forward linear blend skinning. Many recent methods [20, 34, 46, 48], combining 3DGS and the parametric models [37, 43], have made significant strides toward realizing lifelike 3D human avatars. However, achieving real-time rendering on mobile devices remains challenging, especially for AR glasses, which require high resolution, stereo rendering, and seamless interaction.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar designed for running in real-time on Augmented Reality (AR) devices. Unlike previous approaches [8, 10, 34], which develop implicit parameterized models from scratch, we construct a personalized, clothed parameterized template that preserves the facial expressions and gesture controls inherent to the native SMPLX [43]. At the same time, we bind the Gaussians to the triangles as texture to create a hybrid avatar representation. Inspired by [34], we pre-train a large StyleUnet-based teacher network to learn pose-dependent dynamic deformation maps of Gaussians in 2D space by front and back orthogonal projection. Although the teacher network is powerful enough to capture high-frequency appearance details in various poses, its large number of parameters makes it difficult to run in real-time on AR devices. To achieve high-performance rendering, we bake the dynamic deformation of Gaussians into the non-rigid deformation field of the mesh using a distillation technique, employing a compact MLP-based student network. Meanwhile, we propose two lightweight, learnable blend shapes to compensate for the non-rigid deformation of the Gaussians. After fin-tuning, our student model achieves high-performance rendering without compromising quality.", + "bbox": [ + 91, + 367, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate our approach, we propose a multi-view dataset named TalkBody4D, focusing on prevalent full-body talking scenarios encountered in daily life, which includes rich facial expressions and gestures with synchronous audios different from existing full-body motion datasets [11, 22, 69]. Experimental results demonstrate that our method effectively captures high-frequency appearance details of human avatars while maintaining a lightweight architecture capable of rendering stereo images at 2K resolution on common mobile and AR devices, including the Apple Vision Pro. The resulting full-body avatar is highly", + "bbox": [ + 89, + 734, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "expressive, enabling users to animate it with facial expressions, hand gestures, and body poses, thereby highlighting the potential applications of our method. Our contributions can be summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce TaoAvatar, a novel teacher-student framework which yields topologically consistent, well-aligned geometry, further creating a high-fidelity, lightweight, 3DGS-based drivable full-body talking avatar.", + "- We propose two innovative strategies including non-rigid deformation baking and two lightweight blend shapes compensations to ensure high-quality rendering and efficient performance on mobile and AR devices, achieving 2K resolution rendering at 90 FPS on high-definition stereo devices like the Apple Vision Pro.", + "- We contribute TalkBody4D, a multi-view dataset to be released, designed for full-body talking scenarios, featuring diverse facial expressions and gestures with synchronous audios. Extensive experiments demonstrate that our approach outperforms other state-of-the-art methods in both quality and performance." + ], + "bbox": [ + 511, + 152, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 407, + 651, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Avatar Representation. Traditional computer graphics techniques [12, 19, 50, 51] use 3D scanning or multiview stereo (MVS) to create full-body meshes, which are then combined with a 3D skeleton to produce drivable templates. However, this approach is both expensive and inefficient. As parametric body models, SMPL [37] and SMPLX [43] are widely used for their versatility but lack detail in simulating loose clothing. Some works [4, 5, 27] have attempted to add clothing offsets to SMPLX, effectively handling tight clothing but struggling with loose garments. Due to the limited expressiveness of traditional meshes, researchers have combined SMPL with advanced rendering techniques. Some methods [7, 15, 23, 36, 57] use NeRF [40] to implicitly express human bodies, mapping points from the observation space backward to canonical space. Although high-quality, volume rendering is slow and poses challenges for real-time applications. In contrast, 3DGS [29] uses 3D Gaussian distributions for scenes, offering real-time rendering [38, 58], fast training, and the ability to handle complex materials such as hair and translucency. Some methods [34, 46, 48, 56, 63] combining SMPLX [43] and 3DGS [29], like GaussianAvatar [20], use explicit GS point clouds and forward skinning to efficiently simulate the deformations, offering good rendering quality and speed. This representation presents potential solutions for achieving both real-time performance and computational efficiency.", + "bbox": [ + 511, + 431, + 903, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Nonrigid Deformation Modeling. Various methods have been developed to simulate non-rigid human deformation realistically. The MLP-based methods [21, 30, 33, 41], like 3DGS-Avatar [46], are noted for lightweight", + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10724", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "flexibility but struggle with handling novel poses. Physical simulation methods [59, 68] produce physics-based animations but require unified topology, and complex modeling, and are computationally expensive, which negatively impacts real-time performance. Additionally, some methods [17, 18] use GNNs to learn non-rigid deformations in a data-driven manner. These methods are faster than traditional cloth simulations. Recently, dynamic texture methods are emerging, these methods [26, 32, 36] model complex details in the 2D texture space, while AnimatableGS [34] and MeshAvatar [10] enhance modeling with front and back projection strategies which avoid texture unwarping.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Drivable Talking Avatars. Research on drivable digital humans is rapidly advancing for various applications. In 3D Talking Heads, methods [13, 45, 47, 60, 61, 66] excel in capturing and driving facial expressions, enhancing the realism of virtual characters. For drivable bodies, these methods [8, 20, 34, 56, 70] achieve precise restoration of body movements by learning pose-dependent deformation fields. Furthermore, some methods [49, 59] not only support detailed body modeling and animation but also enable head expressions and lip-syncing. Despite enhancing realism, running these methods in real-time on mobile or AR devices is challenging due to their high computational demands, exceeding current edge platform capabilities. Optimizing computational efficiency without compromising high-quality output is a key research focus.", + "bbox": [ + 89, + 272, + 482, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 511, + 181, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Leveraging multi-view RGB videos of a performer alongside corresponding SMPLX registrations for each frame, we aim to develop a high-fidelity, lightweight full-body talking avatar capable of high-resolution and high-performance rendering on mobile devices. To achieve this, we first create a hybrid clothed parametric representation by integrating 3D Gaussian Splatting (3DGS) [29] and SMPLX [43], which is more expressive for simulating loose clothing and hair, while maintaining high-quality rendering (Sec. 3.1). Next, we propose a teacher-student framework to reconstruct high-frequency, pose-dependent dynamic details as illustrated in Fig. 2. By leveraging non-rigid deformation baking and incorporating two lightweight blend shape compensations, our method achieves a tradeoff between high quality and high performance (Sec. 3.2).", + "bbox": [ + 89, + 537, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Hybrid Clothed Parametric Representations", + "text_level": 1, + "bbox": [ + 89, + 772, + 467, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Clothed Extension of SMPLX. To extend SMPLX to complete clothed human geometry with clothes, hair, etc., we choose a frame close to T-pose as a reference, which provides more visible details and less sticky geometry. First, we utilize NeuS2 [54] to reconstruct the geometry of this reference frame and parse out non-body components (e.g., clothes) from it using [53]. However, the body's skeleton", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cannot directly drive these components. We estimate the shape and pose of the SMPLX [37] in the current frame and use robust skinning transfer [3] to propagate the body's skinning weights to these non-body components. Finally, we apply inverse skinning to transform these components back to the standard T-pose, resulting in a comprehensive personalized parametric model, referred to as SMPLX++.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Binding Gaussians on Mesh as Texture. After obtaining the complete geometric model, we bind 3D Gaussians to the mesh's triangles as textures, enabling them to move synchronously with the mesh. Inspired by these hybrid representations [45, 48], we define the attributes of the Gaussians in a local coordinate system relative to the triangles. Specifically, for each triangle, we randomly initialize $k$ Gaussians, each of which maintains the local attributes $\\{f,(u,v),\\gamma ,\\mathbf{r},\\mathbf{s},o,\\mathbf{sh}\\}$ , where $f$ is the index of the parent triangle, $(u,v)$ is the barycentric coordinates, $\\gamma$ represents the translation along the triangle's normal direction, $\\mathbf{r}$ and $\\mathbf{s}$ denote rotation and scaling in the local space, $\\mathbf{o}$ indicates opacity, and $\\mathbf{sh}$ consists of spherical harmonic coefficients. Additionally, we define the normal of each Gaussian as $\\mathbf{n}_g = [1,0,0]$ within the local space, ensuring consistency with the mesh's normal for subsequent transformations. To deform each Gaussian from the local to the world space, we construct a transformation based on its parent triangle $F_{f}(\\mathbf{v}_{1},\\mathbf{v}_{2},\\mathbf{v}_{3})$ , where $\\mathbf{v}_1$ , $\\mathbf{v}_2$ , $\\mathbf{v}_3$ are the vertices of the triangle $F_{f}$ in the world space:", + "bbox": [ + 511, + 196, + 906, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} = u \\cdot \\mathbf {v} _ {1} + v \\cdot \\mathbf {v} _ {2} + (1 - u - v) \\cdot \\mathbf {v} _ {3},\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 513, + 803, + 527 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} = [ \\mathbf {n}, \\mathbf {q}, \\mathbf {n} \\times \\mathbf {q} ], \\mathbf {q} = \\frac {\\left(\\mathbf {v} _ {2} + \\mathbf {v} _ {3}\\right) / 2 - \\mathbf {v} _ {1}}{\\left\\| \\left(\\mathbf {v} _ {2} + \\mathbf {v} _ {3}\\right) / 2 - \\mathbf {v} _ {1} \\right\\|}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 531, + 903, + 564 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ne = \\left(\\left\\| \\mathbf {v} _ {1} - \\mathbf {v} _ {2} \\right\\| + \\left\\| \\mathbf {v} _ {2} - \\mathbf {v} _ {3} \\right\\| + \\left\\| \\mathbf {v} _ {1} - \\mathbf {v} _ {3} \\right\\|\\right) / 3,\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 566, + 864, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{n} = \\frac{(\\mathbf{v}_1 - \\mathbf{v}_2)\\times(\\mathbf{v}_3 - \\mathbf{v}_1)}{\\|\\mathbf{v}_1 - \\mathbf{v}_2\\|\\cdot\\|\\mathbf{v}_3 - \\mathbf{v}_1\\|}$ is the normal of the triangle, $\\mathbf{p}$ denotes the surface point on the parent triangle, and $e$ denotes the average edge length.", + "bbox": [ + 511, + 595, + 903, + 646 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {w} = \\mathbf {p} + \\gamma \\mathbf {R n}, \\quad \\mathbf {r} _ {w} = \\mathbf {R r}, \\quad \\mathbf {s} _ {w} = e \\cdot \\mathbf {s}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 657, + 903, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "here, $\\mathbf{u}_w$ , $\\mathbf{r}_w$ , $\\mathbf{s}_w$ represent the position, rotation, and scaling in the world spaces after transformation, respectively. Furthermore, $\\mathbf{c}_w = SH(\\mathbf{sh},\\mathbf{R}^{-1}\\mathbf{d})$ denotes the color along the view direction $\\mathbf{d}$ in the world space.", + "bbox": [ + 511, + 686, + 903, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dynamic Mesh-based Gaussian Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 757, + 903, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning Dynamic Non-Rigid Gaussian Deformation Maps. Although SMPLX++ can be directly driven by the skeleton, linear blend skinning is insufficient for performing dynamic non-rigid deformation such as clothing folds and swaying motions. Like dynamic Gaussian avatars [26, 34], we utilize a large StyleUnet [52] as the teacher network to capture complex pose-dependent dynamic non-rigid deformation of Gaussians in 2D texture space. Following [34],", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10725", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ffc1c21440c2523504d70f66488fb12509676fa526cc770024f9f855c5f36072.jpg", + "image_caption": [ + "Figure 2. Illustration of our Method. Our pipeline begins by reconstructing (a) a cloth-extended SMPLX mesh with aligned Gaussian textures. To address complex dynamic non-rigid deformations, (b) we employ a teacher StyleUnet to learn pose-dependent non-rigid maps, which are then baked into a lightweight student MLP to infer non-rigid deformations of our template mesh. For high-fidelity rendering, (c) we introduce learnable Gaussian blend shapes to enhance appearance details." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 885, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we rasterize the T-posed mesh, which is colored using posed coordinates and segmentation colors, to obtain the front and back position maps $\\{\\mathcal{P}_f,\\mathcal{P}_b\\}$ . These position maps are then fed into the StyleUnet network along with view directions to generate the non-rigid deformation maps $\\{\\Delta \\mathcal{U}_f,\\Delta \\mathcal{U}_b\\}$ and other residual Gaussian attribute maps. We add these residuals to the Gaussian's local attributes and transform them into the world space according to Eq. (2). To train the StyleUnet-based teacher network $\\mathcal{T}$ , we adopt the losses including L1, D-SSIM [29], and perceptual loss [64]. Additionally, we introduce a normal loss $\\mathcal{L}_{nor}$ . Consequently, the total reconstruction loss $\\mathcal{L}_{rec}$ is defined as:", + "bbox": [ + 88, + 492, + 483, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\mathcal {L} _ {1} + \\lambda_ {\\text {s s i m}} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {\\text {l p i p s}} \\mathcal {L} _ {\\text {l p i p s}} + \\lambda_ {\\text {n o r}} \\mathcal {L} _ {\\text {n o r}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 688, + 482, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{*}$ is the loss weights. The normal loss is defined as $\\mathcal{L}_{nor} = \\| N_t - N_g\\|$ , where $N_{t}$ is the rendered normal image by the teacher network, and the ground truth normal image $N_{g}$ can be obtained from the method [54].", + "bbox": [ + 89, + 718, + 482, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baking Non-Rigid Gaussian Deformation into Mesh Deformation. Although the StyleUnet-based network $\\mathcal{T}$ can achieve good results, it struggles to run in real-time on mobile devices due to its large number of parameters, especially for AR devices, which require high resolution, stereo rendering, and high performance. Inspired by knowledge distillation, we bake the Gaussian non-rigid deformation of the teacher network into the mesh non-rigid deformation", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "field, which is a compact MLP-based student network $S$ .", + "bbox": [ + 511, + 492, + 888, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\bar {\\mathbf {v}} _ {i} = \\mathcal {S} \\left(\\bar {\\mathbf {v}} _ {i}, \\theta , \\mathbf {z} _ {t}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 516, + 903, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{v}}_i$ is the $i$ -th vertex coordinate in the canonical space, $\\theta$ is the pose parameter, and $\\mathbf{z}_t$ is a learnable embedding for each frame to compensate for inaccurate pose estimation. To train the student network $\\mathcal{S}$ with the capability of pose-dependent non-rigid deformation, we directly supervise its output using the Gaussian non-rigid deformation maps $\\{\\Delta \\mathcal{U}_f,\\Delta \\mathcal{U}_b\\}$ of the teacher network $\\mathcal{T}$ . Specifically, we render the mesh non-rigid deformations $\\{\\Delta \\mathbf{v}_i\\}_{i = 1}^N$ to front and back deformation maps $\\{\\Delta \\mathcal{V}_f,\\Delta \\mathcal{V}_b\\}$ in the canonical space, and then calculate the differences:", + "bbox": [ + 511, + 540, + 906, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n o n} = \\left\\| \\Delta \\mathcal {V} _ {f} - \\Delta \\mathcal {U} _ {f} \\right\\| + \\left\\| \\Delta \\mathcal {V} _ {b} - \\Delta \\mathcal {U} _ {b} \\right\\|. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 700, + 903, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, to mitigate the intersection between clothing and the body, we introduce a semantic loss $\\mathcal{L}_{sem}$ . For each vertex, we assign a semantic label $\\mathbf{e}_i = \\mathbf{c}_i + \\sin (\\tau \\bar{\\mathbf{v}}_i)$ that integrates the candidate coordinates $\\bar{\\mathbf{v}}_i$ and the artificial segmentation color $\\mathbf{c}_i$ , and $\\tau$ is a scale factor. The semantic label for each Gaussian can be obtained by interpolating the semantic label of the three vertices of its parent triangle. Subsequently, we render the Gaussian semantic map $\\mathcal{E}_t$ and the mesh semantic map $\\mathcal{E}_s$ in the world space separately and compute the semantic loss $\\mathcal{L}_{sem}$ :", + "bbox": [ + 511, + 724, + 906, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e m} = \\left\\| \\mathcal {E} _ {s} - \\mathcal {E} _ {t} \\right\\|. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 883, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10726", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, we refine the local attributes by computing the reconstruction loss $\\mathcal{L}_{rec}$ , which takes the rendered image of the teacher network as ground truth. The total loss $\\mathcal{L}_{bak}$ during baking is:", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b a k} = \\mathcal {L} _ {r e c} + \\lambda_ {n o n} \\mathcal {L} _ {n o n} + \\lambda_ {s e m} \\mathcal {L} _ {s e m}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 162, + 482, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compensating Gaussian Deformation with Lightweight Blend Shapes. Inspired by pose corrective blend shapes in SMPL [37], we build learnable blend shapes for position and color for each Gaussian. The position blend shapes, $\\mathbf{U} \\in \\mathbb{R}^{3 \\times n}$ , primarily compensate for Gaussian positional adjustments (e.g., hair fluttering, clothing folds) across different poses. Meanwhile, the color blend shapes, $\\mathbf{C} \\in \\mathbb{R}^{3 \\times n}$ , address appearance changes, such as shadow variations caused by self-obscuration. To obtain the corresponding driving coefficients, we employ a mapping network $\\mathcal{H}$ for the head using the expression parameters $\\epsilon$ as input and a mapping network $\\mathcal{B}$ for the body which inputs the body pose parameters $\\theta$ , to get the coefficients $\\mathbf{z}_h \\in \\mathbb{R}^{n_h}$ for the head and $\\mathbf{z}_b \\in \\mathbb{R}^{n_b}$ for the body, respectively, enabling independent and fine-grained control over head and body deformations. The compensation of position and color for each Gaussian can be expressed as:", + "bbox": [ + 89, + 188, + 483, + 445 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\mathbf {u} = B S \\left(\\mathbf {z} _ {h} \\oplus \\mathbf {z} _ {b}; \\mathbf {U}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 455, + 369, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\mathbf {c} = B S \\left(\\mathbf {z} _ {h} \\oplus \\mathbf {z} _ {b}; \\mathbf {C}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 473, + 480, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus$ is a concatenation operation, resulting in the size $n = n_h + n_b$ . The function $BS(\\cdot ,\\cdot)$ represents the blend shapes operation mentioned in SMPL [37]. We add these residuals to local attributes:", + "bbox": [ + 89, + 500, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {w} = \\mathbf {p} + \\mathbf {R} (\\gamma \\mathbf {n} + \\delta \\mathbf {u}),\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 568, + 372, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {w} = S H (\\mathbf {R} ^ {- 1} \\mathbf {d}) + \\delta \\mathbf {c}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 583, + 480, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the fine-tuning stage on training data, we freeze the non-rigid deformation field network $S$ , optimize the two mapping networks $\\mathcal{H}$ and $\\mathcal{B}$ , and two blend shapes $\\mathbf{U}$ and $\\mathbf{C}$ . The loss during fine-tuning is the same as that pre-training the teacher network.", + "bbox": [ + 89, + 614, + 482, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After baking mesh non-rigid deformation and applying Gaussian non-rigid compensation, the performance of our student model is significantly enhanced compared to the teacher model, while still maintaining high quality.", + "bbox": [ + 89, + 689, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 763, + 223, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 787, + 297, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To evaluate the performance on the full-body talking task, we introduce a new dataset, TalkBody4D, designed for common full-body talking scenarios in everyday life, primarily characterized by rich mouth movements and diverse gestures with synchronous audios. The dataset comprises 4 distinct identities, each wearing 2 different outfits.", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each talking sequence consists of rough 6k frames with 4K image resolution, 60 views with 48 full-body, and 12 face-region views. To evaluate performance on complex motions and expressions, we select 4 sequences (04, 05, 06, 08) from the ActorHQ [22] dataset, supplement with 2 dancing and 2 exaggerated expression sequences. The resolution of our training and testing images is $1500 \\times 2000$ for all experiments. SMPLX registrations for these sequences can be obtained by first initializing SMPLX parameters using existing tools [1, 2, 45], followed by fine-tuning the poses with a photometric loss to improve geometry alignment.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. The clothed parametric template has roughly $22k$ vertices and $45k$ faces in total, which contains additional $23k$ faces for clothes, hair, and shoes, compared to the naive SMPLX. We randomly initialize about $220k$ Gaussians in total to bind with the template as texture, for each triangle containing roughly 4 to 6, and optimize these Gaussians for $10k$ iterations using the multi-view images of the reference frame. We pre-train our StyleUnet-based[34] teacher network for $600k$ iterations, and set the loss weights $\\lambda_{ssim} = 0.2$ , $\\lambda_{lpips} = 0.01$ , $\\lambda_{nor} = 0.02$ and batch size is 1 during the training process. Then we bake the teacher network into a small 5-layer MLP-based student network, which set $\\lambda_{non} = 0.1$ , $\\lambda_{sem} = 1.0$ and optimize $30k$ iterations. Finally, we finetune the mapping networks and blend shapes for Gaussian non-rigid compensation on training data $100k$ iterations at the batch size of 4. The size of two blend shapes for position and color is $n = 28$ , in which $n_h = 8$ for the head and $n_b = 20$ for the body.", + "bbox": [ + 511, + 258, + 906, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison", + "text_level": 1, + "bbox": [ + 511, + 544, + 643, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct an extensive comparison between our method and recent state-of-the-art full-body avatar methods include GaussianAvatar [20], 3DGS-Avatar [46], MeshAvatar [10], and AnimatableGS [34]. All of these methods learn pose-dependent non-rigid deformations in the canonical space and then utilize the skeleton-based rigid transformations of SMPL [37] or SMPLX [43] into observation space to model the dynamic human. We uniformly use SMPLX [43] across all these methods to ensure a fair comparison.", + "bbox": [ + 511, + 566, + 905, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison on Full-body Talking. We conduct experiments on typical full-body talking scenarios encountered in daily life, which include a variety of mouth movements and gestures. We provide the quantitative comparison on full-body talking as shown in Tab. 1. We also show the qualitative comparison in Fig. 3, and our method can generate more realistic and detailed results, especially in the face region. 3DGS-Avatar [46] directly learns two MLP-based networks to handle the pose-dependent non-rigid deformation of Gaussians but produces very blurry results due to the low-frequency bias of MLPs. GaussianAvatar [20] uses a 2D CNN network by encoding the position map in SMPLX's UV space to generate the non-rigid deformation", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10727", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0bf0305810caf19bbfc67cf176a043b3929f8cbd206ce269bb72cd330093f867.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons on full-body talking tasks. Our method outperforms state-of-the-art methods by producing clearer clothing dynamics and enhanced facial details." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 903, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "maps of Gaussians. However, it is restricted by the mini-clothed topology of SMPLX, struggling to handle loose skirts. MeshAvatar [10] chooses mesh as a human representation, which is insufficient to express the geometry of details, such as hair, and glasses. AnimatableGS [34] can generate high-quality dynamic appearance and complex non-rigid deformation. However, it is restricted by the implicit template of learning from scratch, which makes it difficult to handle fine-grained expressions, like blinking and flashing teeth as shown in the face region in Fig. 3. Therefore we adopt the SMPLX++ model as the template whose face and hands retain the native SMPLX prior. However, due to the amount parameters of the teacher StyleUnet [52], it achieves poor performance and is hard to run on a mobile", + "bbox": [ + 89, + 681, + 485, + 893 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "device in real time. Through baking non-rigid deformation, we can achieve $150+$ FPS at the resolution of $1500 \\times 2000$ on Nvidia RTX4090 using only a small MLP-based mesh non-rigid deformation network and two lightweight learnable blend shapes as shown in Tab. 1.", + "bbox": [ + 511, + 681, + 906, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on Complex Motion and Expression. Besides full-body talking scenes, our method is also capable of reconstructing complex motions as well as exaggerated expressions as shown in Fig. 4. We provide quantitative comparisons in Tab. 2, our student network achieves competitive results compared to the teacher network, outperforming other methods. Especially for exaggerated expressions, our method significantly outperforms others in reconstructing the face region as shown in Tab. 2. Moreover, our method", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10728", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/751fc55e47ac640ac687ddffdbd70656dea9a876d1c330fd3069607caa0fcce4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Novel ViewNovel Gestures and ExpressionSpeed
ModelPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓FPS
GaussianAvatar[20]26.58 (23.57).9313 (.8159).10577 (.25242)25.99 (23.15).9232 (.8092).12265 (.26207)54
3DGS-Avatar[46]28.91 (23.95).9411 (.8303).07984 (.20450)26.46 (23.32).9157 (.8184).11804 (.21632)55
MeshAvatar[10]28.53 (24.55).9360 (.8083).09470 (.25572)27.08 (23.58).9229 (.7965).10783 (.24947)22
AnimatableGS[34]32.50 (26.42).9599 (.8587).06695 (.19535)28.05 (23.68).9328 (.8142).09191 (.22673)16
Ours (Teacher)33.45 (27.01).9649 (.8741).04986 (.15613)28.28 (24.28).9336 (.8291).07385 (.18176)16
Ours (Student)33.81 (27.80).9689 (.8975).06437 (.14218)28.38 (24.99).9389 (.8525).08874 (.13364)156
", + "bbox": [ + 99, + 88, + 903, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/306b4d44d5fb81c6683afe0b858d0b4ff8dd3d37c0a63d59c10d0321ecd66943.jpg", + "image_caption": [ + "Figure 4. Results in challenging scenarios. Our method can obtain high-quality reconstruction even for challenging pose and expressions." + ], + "image_footnote": [], + "bbox": [ + 96, + 262, + 859, + 526 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9d88b34f9692057bf8110f22bdce96cf644218a7ea7666b6535f62c2b582d977.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on full-body talking task. The results inside the parentheses are evaluated for the face area, and the inference speed is evaluated on Nvidia RTX4090 when rendering images at a resolution of ${1500} \\times {2000}$ ." + ], + "table_footnote": [], + "table_body": "
ModelPSNR↑SSIM↑LPIPS↓
GaussianAvatar[20]25.94 (24.33).9294 (.8251).10478 (.24179)
3DGS-Avatar[46]30.04 (25.08).9403 (.8458).08471 (.18044)
MeshAvatar[10]28.51 (24.94).9334 (.8100).08846 (.23517)
AnimatableGS[34]31.81 (26.79).9493 (.8608).07586 (.19521)
Ours (Teacher)32.80 (27.40).9533 (.8768).05581 (.14996)
Ours (Student)32.72 (27.35).9579 (.8836).07326 (.13914)
", + "bbox": [ + 94, + 566, + 490, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons about complex motions and expressions reconstruction. The results inside the parentheses are evaluated for face regions with exaggerated expressions.", + "bbox": [ + 89, + 676, + 482, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "can obtain high-quality normals as shown in Fig. 4, which can be used for image-based relighting.", + "bbox": [ + 89, + 733, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on Novel Expression and Novel Gesture. Benefiting from our parametric human model SMPLX++, our full-body avatars can be expression-driven and gesture-driven. We provide quantitative comparisons of self-driven on full-body talking scenes as shown in Tab. 1. As shown in Fig. 5, we provide visualizations of different characters driven by the same skeleton. In addition, our model can accept expression parameter inputs generated from audio by UniTalker [14] as shown in Fig. 5. More demos are pro", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vided in the supplementary materials.", + "bbox": [ + 513, + 569, + 761, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 599, + 666, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we discuss the key contributions of our method. Additional quantitative results and experiments are available in the supplementary materials.", + "bbox": [ + 511, + 625, + 906, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Impact of Template. The choice of template has a crucial impact on our approach. If we directly use SMPLX as our geometric template, as shown in Tab. 3 (w SMPLX), leading to poor results with numerous artifacts, especially for loose skirts as seen in Fig. 6. After extending the geometry of SMPLX to a complete template with clothes, hair, and shoes, this quality is significantly improved as shown in Tab. 3 (w/o Mesh Non.), which allows our method to reconstruct more challenging clothes.", + "bbox": [ + 511, + 672, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Impact of Non-rigid Deformation. Despite having a complete geometry, the rigid transformations of the skeleton are not able to properly represent non-rigid transformations such as skirt swinging and hair flying. We concurrently introduce mesh non-rigid deformation and Gaussian non-rigid deformation in our method. The geometry with-", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10729", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd81a408de9da3e103ed8a1462d3aaccdb6b9831218adfc16104d323aa67ca29.jpg", + "image_caption": [ + "Figure 5. Novel pose and expression animation. TaoAvatar can be driven by the same skeleton and expression parameters vividly." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "out mesh non-rigid deformation is unable to attach to the surface of the character as shown in Fig. 6 (w/o Mesh Non.). The Gaussian non-rigid compensation also plays a crucial role in quantitative metrics. Without the blend shapes for Gaussian non-rigid compensation, the results are prone to produce artifacts due to imprecise geometry and variable appearance as shown in Fig. 6 (w/o Gau Non.).", + "bbox": [ + 89, + 539, + 482, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Impact of Baking Teacher. Compared to directly training a small MLP-based network to handle non-rigid deformations as shown in Tab. 3 (w/o Teacher), our proposed multi-stage training strategy is more efficient. With direct supervision of the non-rigid deformation of the teacher network, it is easier for the student network to decouple the geometry deformation and appearance variations.", + "bbox": [ + 89, + 646, + 482, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Application", + "text_level": 1, + "bbox": [ + 89, + 768, + 212, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TaoAvatar offers a lightweight, comprehensive representation for 3D talking body scenarios, easily deployable on various mobile devices using existing tools [25] as demonstrated in Fig. 1. It seamlessly integrates with AI models to enable real-time dialogue interactions. we deployed a 3D digital human agent on the Apple Vision Pro, which interacts with users through an ASR-LLM-TTS pipeline [16, 31,", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59d40bf71ace3044765c29d31cff170e4af8e88e200a969b6f952fc8fa6453a7.jpg", + "image_caption": [ + "Figure 6. Ablation Study. Red and Green boxes show artifacts and their improved counterparts." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 292 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e45798e617a0d2e041901cb9d698ad4b8be61e79733661336c64e63f0bad4832.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPSNR↑SSIM↑LPIPS↓P2S↓Chamfer↓
w SMPLX28.47.9476.07899.7690.9995
w/o Mesh Non.32.10.9734.03814.4877.5007
w/o Gau Non.31.16.9686.03932.2968.3068
w/o Teacher32.67.9751.03769.5236.5359
Full33.29.9772.03464.2953.3052
", + "bbox": [ + 517, + 339, + 916, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation Study. We use the mesh reconstructed by NeuS2 [54] as a pseudo-truth for geometry evaluation.", + "bbox": [ + 511, + 436, + 906, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "39, 62]. Facial expressions and gestures are dynamically controlled by an Audio2BS model [14], allowing the agent to respond naturally with synchronized speech, expressions, and movements. A live demonstration is available in our supplementary materials.", + "bbox": [ + 511, + 479, + 906, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 511, + 568, + 627, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this work, we present TaoAvatar, a lightweight, lifelike full-body talking avatar solution. We demonstrate how the teacher-student framework captures high-definition facial and body details while ensuring real-time performance on AR devices. TaoAvatar can be driven by diverse signals, including facial expressions, hand gestures, and body poses. Through both quantitative and qualitative evaluations, we showcase the advantages of our approach. Additionally, we validate its practical potential with a real-world application on the Apple Vision Pro.", + "bbox": [ + 511, + 594, + 906, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Works. TaoAvatar encounters challenges in modeling flexible clothing deformation under exaggerated body poses, which are out-of-distribution of training data. A possible solution is to integrate GNN simulators [17, 18] handling larger hemlines, which are compatible with our approach.", + "bbox": [ + 511, + 744, + 905, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Potential Social Impact. TaoAvatar can synthesize life-like talking digital humans within an augmented reality environment, generating fabricated 3D content or 2D videos. Therefore, responsible use of this technology is essential.", + "bbox": [ + 511, + 835, + 905, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10730", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://github.com/zju3dv/EasyMocap.5, 13", + "[2] https://github.com/ShenhanQian/VHAP.5,13", + "[3] Rinet Abdrashitov, Kim Raichstat, Jared Monsen, and David Hill. Robust skin weights transfer via weight inpainting. In SIGGRAPH Asia 2023 Technical Communications, SA Technical Communications 2023, Sydney, NSW, Australia, December 12-15, 2023, pages 25:1-25:4. ACM, 2023. 3, 13", + "[4] Thiemo Alldieck, Marcus A. Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 8387-8397. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 2", + "[5] Thiemo Alldieck, Marcus A. Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1175-1186. Computer Vision Foundation / IEEE, 2019. 2", + "[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 5419-5429. IEEE, 2019. 1", + "[7] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos, 2021. 2", + "[8] Xu Chen, Yufeng Zheng, Michael J. Black, Otmar Hilliges, and Andreas Geiger. SNARF: differentiable forward skinning for animating non-rigid neural implicit shapes. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 11574-11584. IEEE, 2021. 2, 3", + "[9] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023. 2", + "[10] Yushuo Chen, Zerong Zheng, Zhe Li, Chao Xu, and Yebin Liu. Meshavatar: Learning high-quality triangular human avatars from multi-view videos. CoRR, abs/2407.08414, 2024. 2, 3, 5, 6, 7, 13, 14, 16", + "[11] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 19925-19936. IEEE, 2023. 2", + "[12] Edilson De Aguiar, Christian Theobalt, Carsten Stoll, and Hans-Peter Seidel. Marker-less deformable mesh tracking" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for human shape and motion capture. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2", + "[13] Helisa Dhamo, Yinyu Nie, Arthur Moreau, Jifei Song, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Headgas: Real-time animatable head avatars via 3d gaussian splatting. In ECCV (2), pages 459-476. Springer, 2024. 3", + "[14] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through A unified model. CoRR, abs/2408.00762, 2024. 7, 8, 14", + "[15] Yao Feng, Weiyang Liu, Timo Bolkart, Jinlong Yang, Marc Pollefeys, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. CoRR, abs/2309.06441, 2023. 2", + "[16] Zhifu Gao, Shiliang Zhang, Ian McLoughlin, and Zhijie Yan. Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to-end speech recognition. In 23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18-22, 2022, pages 2063-2067. ISCA, 2022. 8, 14", + "[17] Artur Grigorev, Michael J. Black, and Otmar Hilliges. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16965-16974. IEEE, 2023. 3, 8", + "[18] Artur Grigorev, Giorgio Becherini, Michael J. Black, Otmar Hilliges, and Bernhard Thomaszewski. Contourcraft: Learning to resolve intersections in neural multi-garment simulations. In ACM SIGGRAPH 2024 Conference Papers, SIGGRAPH 2024, Denver, CO, USA, 27 July 2024-1 August 2024, page 81. ACM, 2024. 3, 8", + "[19] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics, 40(4), 2021. 2", + "[20] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 634-644. IEEE, 2024. 2, 3, 5, 7, 14, 16", + "[21] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20418-20431. IEEE, 2024. 2", + "[22] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Trans. Graph., 42(4):160:1-160:12, 2023. 2, 5", + "[23] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In IEEE/CVF Conference on Computer Vision and" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10731", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5595-5605. IEEE, 2022. 2", + "[24] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16922-16932, 2023. 2", + "[25] Xiaotang Jiang, Huan Wang, Yiliu Chen, Ziqi Wu, Lichuan Wang, Bin Zou, Yafeng Yang, Zongyang Cui, Yu Cai, Tianhang Yu, Chengfei Lyu, and Zhihua Wu. MNN: A universal and efficient inference engine. In Proceedings of the Third Conference on Machine Learning and Systems, MLSys 2020, Austin, TX, USA, March 2-4, 2020. mlsys.org, 2020. 8", + "[26] Yujiang Jiang, Qingmin Liao, Xiaoyu Li, Li Ma, Qi Zhang, Chaopeng Zhang, Zongqing Lu, and Ying Shan. UV gaus-sians: Joint learning of mesh deformation and gaussian textures for human avatar modeling. CoRR, abs/2403.11589, 2024. 3", + "[27] Yujiang Jiang, Qingmin Liao, Zhaolong Wang, Xiangru Lin, Zongqing Lu, Yuxi Zhao, Hanqing Wei, Jingrui Ye, Yu Zhang, and Zhijing Shao. Smplx-lite: A realistic and drivable avatar benchmark with rich geometry and texture annotations. In IEEE International Conference on Multimedia and Expo, ICME 2024, Niagara Falls, ON, Canada, July 15-19, 2024, pages 1-6. IEEE, 2024. 1, 2", + "[28] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5322-5332. IEEE, 2024. 16", + "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 4", + "[30] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: human gaussian splats. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 505-515. IEEE, 2024. 2", + "[31] Jungil Kong, Jihoon Park, Beomjeong Kim, Jeongmin Kim, Dohee Kong, and Sangjin Kim. VITS2: improving quality and efficiency of single-stage text-to-speech with adversarial learning and architecture design. In 24th Annual Conference of the International Speech Communication Association, Interspeech 2023, Dublin, Ireland, August 20-24, 2023, pages 4374-4378. ISCA, 2023. 8, 14", + "[32] Youngjoong Kwon, Baole Fang, Yixing Lu, Haoye Dong, Cheng Zhang, Francisco Vicente Carrasco, Albert Mosella-Montoro, Jianjin Xu, Shingo Takagi, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Generalizable human gaussians for sparse view synthesis. In ECCV (78), pages 451-468. Springer, 2024. 3", + "[33] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. GART: gaussian articulated template models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19876-19887. IEEE, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19711-19722. IEEE, 2024. 2, 3, 5, 6, 7, 13, 14, 15, 16", + "[35] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3d gaussian splatting for inverse rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21644-21653. IEEE, 2024. 16", + "[36] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: neural free-view synthesis of human actors with pose control. ACM Trans. Graph., 40(6):219:1-219:16, 2021. 2, 3", + "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 1, 2, 3, 5", + "[38] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 2", + "[39] Chengfei Lv, Chaoyue Niu, Renjie Gu, Xiaotang Jiang, Zhaode Wang, Bin Liu, Ziqi Wu, Qiulin Yao, Congyu Huang, Panos Huang, Tao Huang, Hui Shu, Jinde Song, Bin Zou, Peng Lan, Guohuan Xu, Fei Wu, Shaojie Tang, Fan Wu, and Guihai Chen. Walle: An end-to-end, general-purpose, and large-scale production system for device-cloud collaborative machine learning. In 16th USENIX Symposium on Operating Systems Design and Implementation, OSDI 2022, Carlsbad, CA, USA, July 11-13, 2022, pages 249-265. USENIX Association, 2022. 8", + "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 13, 16", + "[41] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In CVPR, pages 788-798. IEEE, 2024. 2", + "[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2", + "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1, 2, 3, 5, 14", + "[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 1", + "[45] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gauss" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10732", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20299-20309. IEEE, 2024. 3, 5", + "[46] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5020-5030. IEEE, 2024. 2, 5, 7, 14, 16", + "[47] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 130-141. IEEE, 2024. 3", + "[48] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 1606-1616. IEEE, 2024. 2, 3", + "[49] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16911-16921. IEEE, 2023. 3", + "[50] Carsten Stoll, Juergen Gall, Edilson De Aguiar, Sebastian Thrun, and Christian Theobalt. Video-based reconstruction of animatable human characters. ACM Transactions on Graphics (TOG), 29(6):1-10, 2010. 2", + "[51] Daniel Vlasic, Pieter Peers, Ilya Baran, Paul E. Debevec, Jovan Popovic, Szymon Rusinkiewicz, and Wojciech Matusik. Dynamic shape capture using multi-view photometric stereo. ACM Trans. Graph., 28(5):174, 2009. 2", + "[52] Lizhen Wang, Xiaochen Zhao, Jingxiang Sun, Yuxiang Zhang, Hongwen Zhang, Tao Yu, and Yebin Liu. Stylea-vatar: Real-time photo-realistic portrait avatar from a single video. In ACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH 2023, Los Angeles, CA, USA, August 6-10, 2023, pages 67:1-67:10. ACM, 2023. 3, 6, 15", + "[53] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 550-560. IEEE, 2024. 3, 13", + "[54] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3272-3283. IEEE, 2023. 2, 3, 4, 8, 13, 15, 16", + "[55] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 15" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Jing Wen, Xiaoming Zhao, Zhongzheng Ren, Alexander G. Schwing, and Shenlong Wang. Gomavatar: Efficient animatable human modeling from monocular video using gaussians-on-mesh. In CVPR, pages 2059-2069. IEEE, 2024. 2, 3", + "[57] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16189-16199. IEEE, 2022. 2", + "[58] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 2", + "[59] Donglai Xiang, Timur M. Bagautdinov, Tuur Stuyck, Fabian Prada, Javier Romero, Weipeng Xu, Shunsuke Saito, Jingfan Guo, Breannan Smith, Takaaki Shiratori, Yaser Sheikh, Jessica K. Hodgins, and Chenglei Wu. Dressing avatars: Deep photorealistic appearance for physically simulated clothing. ACM Trans. Graph., 41(6):222:1-222:15, 2022. 3", + "[60] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In CVPR, pages 1802-1812. IEEE, 2024. 3", + "[61] Yuelang Xu, Bengwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 1931-1941. IEEE, 2024. 3", + "[62] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. Qwen2 technical report. CoRR, abs/2407.10671, 2024. 8, 14", + "[63] Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. In CVPR, pages 896-905. IEEE, 2024. 2", + "[64] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595. Computer Vision Foundation / IEEE Computer Society, 2018. 4, 15" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10733", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humanerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 1", + "[66] Zhongyuan Zhao, Zhenyu Bao, Qing Li, Guoping Qiu, and Kanglin Liu. Psavatar: A point-based shape model for real-time head avatar animation with 3d gaussian splatting, 2024. 3", + "[67] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CoRR, abs/2401.03407, 2024. 15", + "[68] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas J. Guibas, and Gordon Wetzstein. Physavatar: Learning the physics of dressed 3d avatars from visual observations. CoRR, abs/2404.04421, 2024. 3", + "[69] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM Trans. Graph., 42(4):158:1-158:19, 2023. 2", + "[70] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4574-4584. IEEE, 2023. 3" + ], + "bbox": [ + 91, + 92, + 480, + 483 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10734", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_model.json b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5ec4ff6f31d88cacc1b6c541a4a0b49e6ef88933 --- /dev/null +++ b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_model.json @@ -0,0 +1,2182 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.191, + 0.13, + 0.812, + 0.178 + ], + "angle": 0, + "content": "TaoAvatar: Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.204, + 0.766, + 0.257 + ], + "angle": 0, + "content": "Jianchuan Chen,\\* Jingchuan Hu,\\* Gaige Wang,\\* Zhonghua Jiang \nTiansong Zhou, Zhiwen Chen\\*, Chengfei Lv\\* \nAlibaba Group, Hangzhou, China" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.257, + 0.681, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.258, + 0.902, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.435, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Figure 1. TaoAvatar generates photorealistic, topology-consistent 3D full-body avatars from multi-view sequences. It provides high-quality, real-time rendering with low storage requirements, compatible across various mobile and AR devices like the Apple Vision Pro." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.478, + 0.327, + 0.495 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.511, + 0.486, + 0.859 + ], + "angle": 0, + "content": "Realistic 3D full-body talking avatars hold great potential in AR, with applications ranging from e-commerce live streaming to holographic communication. Despite advances in 3D Gaussian Splatting (3DGS) for lifelike avatar creation, existing methods struggle with fine-grained control of facial expressions and body movements in full-body talking tasks. Additionally, they often lack sufficient details and cannot run in real-time on mobile devices. We present TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar driven by various signals. Our approach starts by creating a personalized clothed human parametric template that binds Gaussians to represent appearances. We then pre-train a StyleUnet-based network to handle complex pose-dependent non-rigid deformation, which can capture high-frequency appearance details but is too resource-intensive for mobile devices. To overcome this, we \"bake\" the non-rigid deformations into a lightweight MLP-based network using a distillation technique and develop blend shapes to compensate for details. Extensive experiments show that TaoAvatar achieves state-of-the-art rendering quality while running in real-time across various devices, maintaining 90 FPS on high-definition stereo devices such as the Apple Vision Pro." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.478, + 0.645, + 0.493 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Creating lifelike 3D human avatars is a dynamic and rapidly advancing research area in computer vision and graphics, essential for applications in AR/VR, 3D entertainment, and virtual communication. Despite decades of research, achieving realistic and immersive avatars on mobile devices—particularly AR—remains challenging due to computational limitations. Current industry solutions, such as MetaHuman, often rely on high-precision scans and extensive manual effort by artists for 3D modeling, rigging, and motion capture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In recent years, academic researchers have adopted parametric models [37, 43] for human representation, which are constructed from massive body scans and can capture a performer's shape, expression, and pose from images or videos. These methods [4-6, 27] extend the naked model [37, 43] by adding offsets to represent tailored clothing, but are still struggling with complex geometries and high-frequency details, such as loose skirts and fine hair, due to limitations of topology and texture resolution. With the rise of neural radiance fields (NeRF) [40], many researchers tend to define implicit neural human representations in canonical space, animated by backward linear blend skinning [44, 65] of these parametric models [37, 43]. While these methods can handle arbitrary topology and deliver higher rendering quality, they are highly time-consuming due to the large amount of sampling required" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.903 + ], + "angle": 0, + "content": "*Equal contribution. ‡Project Leader and †Corresponding author. Project: https://PixelAI-Team.github.io/TaoAvatar" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "10723" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "during volume rendering. Although many efforts [9, 24] have significantly improved rendering efficiency by introducing iNGP [42], the animated results still fail to escape the Uncanny Valley effect when driven. Recently, 3D Gaussian Splatting (3DGS) [29], an explicit and efficient point-based representation, has gained tremendous attention from researchers for its ability to deliver both high-quality rendering and real-time performance. Compared to implicit representations [40, 54], the explicit point-based representation is more compatible with being integrated with the mesh-based parametric models [37, 43], which can be directly driven by forward linear blend skinning. Many recent methods [20, 34, 46, 48], combining 3DGS and the parametric models [37, 43], have made significant strides toward realizing lifelike 3D human avatars. However, achieving real-time rendering on mobile devices remains challenging, especially for AR glasses, which require high resolution, stereo rendering, and seamless interaction." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.368, + 0.483, + 0.732 + ], + "angle": 0, + "content": "We introduce TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar designed for running in real-time on Augmented Reality (AR) devices. Unlike previous approaches [8, 10, 34], which develop implicit parameterized models from scratch, we construct a personalized, clothed parameterized template that preserves the facial expressions and gesture controls inherent to the native SMPLX [43]. At the same time, we bind the Gaussians to the triangles as texture to create a hybrid avatar representation. Inspired by [34], we pre-train a large StyleUnet-based teacher network to learn pose-dependent dynamic deformation maps of Gaussians in 2D space by front and back orthogonal projection. Although the teacher network is powerful enough to capture high-frequency appearance details in various poses, its large number of parameters makes it difficult to run in real-time on AR devices. To achieve high-performance rendering, we bake the dynamic deformation of Gaussians into the non-rigid deformation field of the mesh using a distillation technique, employing a compact MLP-based student network. Meanwhile, we propose two lightweight, learnable blend shapes to compensate for the non-rigid deformation of the Gaussians. After fin-tuning, our student model achieves high-performance rendering without compromising quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To evaluate our approach, we propose a multi-view dataset named TalkBody4D, focusing on prevalent full-body talking scenarios encountered in daily life, which includes rich facial expressions and gestures with synchronous audios different from existing full-body motion datasets [11, 22, 69]. Experimental results demonstrate that our method effectively captures high-frequency appearance details of human avatars while maintaining a lightweight architecture capable of rendering stereo images at 2K resolution on common mobile and AR devices, including the Apple Vision Pro. The resulting full-body avatar is highly" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "expressive, enabling users to animate it with facial expressions, hand gestures, and body poses, thereby highlighting the potential applications of our method. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.213 + ], + "angle": 0, + "content": "- We introduce TaoAvatar, a novel teacher-student framework which yields topologically consistent, well-aligned geometry, further creating a high-fidelity, lightweight, 3DGS-based drivable full-body talking avatar." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- We propose two innovative strategies including non-rigid deformation baking and two lightweight blend shapes compensations to ensure high-quality rendering and efficient performance on mobile and AR devices, achieving 2K resolution rendering at 90 FPS on high-definition stereo devices like the Apple Vision Pro." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- We contribute TalkBody4D, a multi-view dataset to be released, designed for full-body talking scenarios, featuring diverse facial expressions and gestures with synchronous audios. Extensive experiments demonstrate that our approach outperforms other state-of-the-art methods in both quality and performance." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.652, + 0.423 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.84 + ], + "angle": 0, + "content": "3D Avatar Representation. Traditional computer graphics techniques [12, 19, 50, 51] use 3D scanning or multiview stereo (MVS) to create full-body meshes, which are then combined with a 3D skeleton to produce drivable templates. However, this approach is both expensive and inefficient. As parametric body models, SMPL [37] and SMPLX [43] are widely used for their versatility but lack detail in simulating loose clothing. Some works [4, 5, 27] have attempted to add clothing offsets to SMPLX, effectively handling tight clothing but struggling with loose garments. Due to the limited expressiveness of traditional meshes, researchers have combined SMPL with advanced rendering techniques. Some methods [7, 15, 23, 36, 57] use NeRF [40] to implicitly express human bodies, mapping points from the observation space backward to canonical space. Although high-quality, volume rendering is slow and poses challenges for real-time applications. In contrast, 3DGS [29] uses 3D Gaussian distributions for scenes, offering real-time rendering [38, 58], fast training, and the ability to handle complex materials such as hair and translucency. Some methods [34, 46, 48, 56, 63] combining SMPLX [43] and 3DGS [29], like GaussianAvatar [20], use explicit GS point clouds and forward skinning to efficiently simulate the deformations, offering good rendering quality and speed. This representation presents potential solutions for achieving both real-time performance and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Dynamic Nonrigid Deformation Modeling. Various methods have been developed to simulate non-rigid human deformation realistically. The MLP-based methods [21, 30, 33, 41], like 3DGS-Avatar [46], are noted for lightweight" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10724" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "flexibility but struggle with handling novel poses. Physical simulation methods [59, 68] produce physics-based animations but require unified topology, and complex modeling, and are computationally expensive, which negatively impacts real-time performance. Additionally, some methods [17, 18] use GNNs to learn non-rigid deformations in a data-driven manner. These methods are faster than traditional cloth simulations. Recently, dynamic texture methods are emerging, these methods [26, 32, 36] model complex details in the 2D texture space, while AnimatableGS [34] and MeshAvatar [10] enhance modeling with front and back projection strategies which avoid texture unwarping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.501 + ], + "angle": 0, + "content": "3D Drivable Talking Avatars. Research on drivable digital humans is rapidly advancing for various applications. In 3D Talking Heads, methods [13, 45, 47, 60, 61, 66] excel in capturing and driving facial expressions, enhancing the realism of virtual characters. For drivable bodies, these methods [8, 20, 34, 56, 70] achieve precise restoration of body movements by learning pose-dependent deformation fields. Furthermore, some methods [49, 59] not only support detailed body modeling and animation but also enable head expressions and lip-syncing. Despite enhancing realism, running these methods in real-time on mobile or AR devices is challenging due to their high computational demands, exceeding current edge platform capabilities. Optimizing computational efficiency without compromising high-quality output is a key research focus." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.182, + 0.527 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.765 + ], + "angle": 0, + "content": "Leveraging multi-view RGB videos of a performer alongside corresponding SMPLX registrations for each frame, we aim to develop a high-fidelity, lightweight full-body talking avatar capable of high-resolution and high-performance rendering on mobile devices. To achieve this, we first create a hybrid clothed parametric representation by integrating 3D Gaussian Splatting (3DGS) [29] and SMPLX [43], which is more expressive for simulating loose clothing and hair, while maintaining high-quality rendering (Sec. 3.1). Next, we propose a teacher-student framework to reconstruct high-frequency, pose-dependent dynamic details as illustrated in Fig. 2. By leveraging non-rigid deformation baking and incorporating two lightweight blend shape compensations, our method achieves a tradeoff between high quality and high performance (Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.468, + 0.79 + ], + "angle": 0, + "content": "3.1. Hybrid Clothed Parametric Representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The Clothed Extension of SMPLX. To extend SMPLX to complete clothed human geometry with clothes, hair, etc., we choose a frame close to T-pose as a reference, which provides more visible details and less sticky geometry. First, we utilize NeuS2 [54] to reconstruct the geometry of this reference frame and parse out non-body components (e.g., clothes) from it using [53]. However, the body's skeleton" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.197 + ], + "angle": 0, + "content": "cannot directly drive these components. We estimate the shape and pose of the SMPLX [37] in the current frame and use robust skinning transfer [3] to propagate the body's skinning weights to these non-body components. Finally, we apply inverse skinning to transform these components back to the standard T-pose, resulting in a comprehensive personalized parametric model, referred to as SMPLX++." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Binding Gaussians on Mesh as Texture. After obtaining the complete geometric model, we bind 3D Gaussians to the mesh's triangles as textures, enabling them to move synchronously with the mesh. Inspired by these hybrid representations [45, 48], we define the attributes of the Gaussians in a local coordinate system relative to the triangles. Specifically, for each triangle, we randomly initialize \\( k \\) Gaussians, each of which maintains the local attributes \\( \\{f,(u,v),\\gamma ,\\mathbf{r},\\mathbf{s},o,\\mathbf{sh}\\} \\), where \\( f \\) is the index of the parent triangle, \\( (u,v) \\) is the barycentric coordinates, \\( \\gamma \\) represents the translation along the triangle's normal direction, \\( \\mathbf{r} \\) and \\( \\mathbf{s} \\) denote rotation and scaling in the local space, \\( \\mathbf{o} \\) indicates opacity, and \\( \\mathbf{sh} \\) consists of spherical harmonic coefficients. Additionally, we define the normal of each Gaussian as \\( \\mathbf{n}_g = [1,0,0] \\) within the local space, ensuring consistency with the mesh's normal for subsequent transformations. To deform each Gaussian from the local to the world space, we construct a transformation based on its parent triangle \\( F_{f}(\\mathbf{v}_{1},\\mathbf{v}_{2},\\mathbf{v}_{3}) \\), where \\( \\mathbf{v}_1 \\), \\( \\mathbf{v}_2 \\), \\( \\mathbf{v}_3 \\) are the vertices of the triangle \\( F_{f} \\) in the world space:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.514, + 0.805, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} = u \\cdot \\mathbf {v} _ {1} + v \\cdot \\mathbf {v} _ {2} + (1 - u - v) \\cdot \\mathbf {v} _ {3},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.532, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} = [ \\mathbf {n}, \\mathbf {q}, \\mathbf {n} \\times \\mathbf {q} ], \\mathbf {q} = \\frac {\\left(\\mathbf {v} _ {2} + \\mathbf {v} _ {3}\\right) / 2 - \\mathbf {v} _ {1}}{\\left\\| \\left(\\mathbf {v} _ {2} + \\mathbf {v} _ {3}\\right) / 2 - \\mathbf {v} _ {1} \\right\\|}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.567, + 0.866, + 0.584 + ], + "angle": 0, + "content": "\\[\ne = \\left(\\left\\| \\mathbf {v} _ {1} - \\mathbf {v} _ {2} \\right\\| + \\left\\| \\mathbf {v} _ {2} - \\mathbf {v} _ {3} \\right\\| + \\left\\| \\mathbf {v} _ {1} - \\mathbf {v} _ {3} \\right\\|\\right) / 3,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.905, + 0.647 + ], + "angle": 0, + "content": "where \\(\\mathbf{n} = \\frac{(\\mathbf{v}_1 - \\mathbf{v}_2)\\times(\\mathbf{v}_3 - \\mathbf{v}_1)}{\\|\\mathbf{v}_1 - \\mathbf{v}_2\\|\\cdot\\|\\mathbf{v}_3 - \\mathbf{v}_1\\|}\\) is the normal of the triangle, \\(\\mathbf{p}\\) denotes the surface point on the parent triangle, and \\(e\\) denotes the average edge length." + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.659, + 0.905, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {w} = \\mathbf {p} + \\gamma \\mathbf {R n}, \\quad \\mathbf {r} _ {w} = \\mathbf {R r}, \\quad \\mathbf {s} _ {w} = e \\cdot \\mathbf {s}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.905, + 0.749 + ], + "angle": 0, + "content": "here, \\(\\mathbf{u}_w\\), \\(\\mathbf{r}_w\\), \\(\\mathbf{s}_w\\) represent the position, rotation, and scaling in the world spaces after transformation, respectively. Furthermore, \\(\\mathbf{c}_w = SH(\\mathbf{sh},\\mathbf{R}^{-1}\\mathbf{d})\\) denotes the color along the view direction \\(\\mathbf{d}\\) in the world space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.905, + 0.773 + ], + "angle": 0, + "content": "3.2. Dynamic Mesh-based Gaussian Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Learning Dynamic Non-Rigid Gaussian Deformation Maps. Although SMPLX++ can be directly driven by the skeleton, linear blend skinning is insufficient for performing dynamic non-rigid deformation such as clothing folds and swaying motions. Like dynamic Gaussian avatars [26, 34], we utilize a large StyleUnet [52] as the teacher network to capture complex pose-dependent dynamic non-rigid deformation of Gaussians in 2D texture space. Following [34]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10725" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.887, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.411, + 0.908, + 0.468 + ], + "angle": 0, + "content": "Figure 2. Illustration of our Method. Our pipeline begins by reconstructing (a) a cloth-extended SMPLX mesh with aligned Gaussian textures. To address complex dynamic non-rigid deformations, (b) we employ a teacher StyleUnet to learn pose-dependent non-rigid maps, which are then baked into a lightweight student MLP to infer non-rigid deformations of our template mesh. For high-fidelity rendering, (c) we introduce learnable Gaussian blend shapes to enhance appearance details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.484, + 0.676 + ], + "angle": 0, + "content": "we rasterize the T-posed mesh, which is colored using posed coordinates and segmentation colors, to obtain the front and back position maps \\(\\{\\mathcal{P}_f,\\mathcal{P}_b\\}\\). These position maps are then fed into the StyleUnet network along with view directions to generate the non-rigid deformation maps \\(\\{\\Delta \\mathcal{U}_f,\\Delta \\mathcal{U}_b\\}\\) and other residual Gaussian attribute maps. We add these residuals to the Gaussian's local attributes and transform them into the world space according to Eq. (2). To train the StyleUnet-based teacher network \\(\\mathcal{T}\\), we adopt the losses including L1, D-SSIM [29], and perceptual loss [64]. Additionally, we introduce a normal loss \\(\\mathcal{L}_{nor}\\). Consequently, the total reconstruction loss \\(\\mathcal{L}_{rec}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.689, + 0.483, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\mathcal {L} _ {1} + \\lambda_ {\\text {s s i m}} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {\\text {l p i p s}} \\mathcal {L} _ {\\text {l p i p s}} + \\lambda_ {\\text {n o r}} \\mathcal {L} _ {\\text {n o r}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.78 + ], + "angle": 0, + "content": "where \\(\\lambda_{*}\\) is the loss weights. The normal loss is defined as \\(\\mathcal{L}_{nor} = \\| N_t - N_g\\|\\), where \\(N_{t}\\) is the rendered normal image by the teacher network, and the ground truth normal image \\(N_{g}\\) can be obtained from the method [54]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Baking Non-Rigid Gaussian Deformation into Mesh Deformation. Although the StyleUnet-based network \\(\\mathcal{T}\\) can achieve good results, it struggles to run in real-time on mobile devices due to its large number of parameters, especially for AR devices, which require high resolution, stereo rendering, and high performance. Inspired by knowledge distillation, we bake the Gaussian non-rigid deformation of the teacher network into the mesh non-rigid deformation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.493, + 0.889, + 0.509 + ], + "angle": 0, + "content": "field, which is a compact MLP-based student network \\(S\\)." + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.517, + 0.905, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\Delta \\bar {\\mathbf {v}} _ {i} = \\mathcal {S} \\left(\\bar {\\mathbf {v}} _ {i}, \\theta , \\mathbf {z} _ {t}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.694 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{v}}_i\\) is the \\(i\\)-th vertex coordinate in the canonical space, \\(\\theta\\) is the pose parameter, and \\(\\mathbf{z}_t\\) is a learnable embedding for each frame to compensate for inaccurate pose estimation. To train the student network \\(\\mathcal{S}\\) with the capability of pose-dependent non-rigid deformation, we directly supervise its output using the Gaussian non-rigid deformation maps \\(\\{\\Delta \\mathcal{U}_f,\\Delta \\mathcal{U}_b\\}\\) of the teacher network \\(\\mathcal{T}\\). Specifically, we render the mesh non-rigid deformations \\(\\{\\Delta \\mathbf{v}_i\\}_{i = 1}^N\\) to front and back deformation maps \\(\\{\\Delta \\mathcal{V}_f,\\Delta \\mathcal{V}_b\\}\\) in the canonical space, and then calculate the differences:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.702, + 0.905, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n o n} = \\left\\| \\Delta \\mathcal {V} _ {f} - \\Delta \\mathcal {U} _ {f} \\right\\| + \\left\\| \\Delta \\mathcal {V} _ {b} - \\Delta \\mathcal {U} _ {b} \\right\\|. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.907, + 0.877 + ], + "angle": 0, + "content": "Furthermore, to mitigate the intersection between clothing and the body, we introduce a semantic loss \\(\\mathcal{L}_{sem}\\). For each vertex, we assign a semantic label \\(\\mathbf{e}_i = \\mathbf{c}_i + \\sin (\\tau \\bar{\\mathbf{v}}_i)\\) that integrates the candidate coordinates \\(\\bar{\\mathbf{v}}_i\\) and the artificial segmentation color \\(\\mathbf{c}_i\\), and \\(\\tau\\) is a scale factor. The semantic label for each Gaussian can be obtained by interpolating the semantic label of the three vertices of its parent triangle. Subsequently, we render the Gaussian semantic map \\(\\mathcal{E}_t\\) and the mesh semantic map \\(\\mathcal{E}_s\\) in the world space separately and compute the semantic loss \\(\\mathcal{L}_{sem}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e m} = \\left\\| \\mathcal {E} _ {s} - \\mathcal {E} _ {t} \\right\\|. \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10726" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "In addition, we refine the local attributes by computing the reconstruction loss \\(\\mathcal{L}_{rec}\\), which takes the rendered image of the teacher network as ground truth. The total loss \\(\\mathcal{L}_{bak}\\) during baking is:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.163, + 0.483, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b a k} = \\mathcal {L} _ {r e c} + \\lambda_ {n o n} \\mathcal {L} _ {n o n} + \\lambda_ {s e m} \\mathcal {L} _ {s e m}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.484, + 0.446 + ], + "angle": 0, + "content": "Compensating Gaussian Deformation with Lightweight Blend Shapes. Inspired by pose corrective blend shapes in SMPL [37], we build learnable blend shapes for position and color for each Gaussian. The position blend shapes, \\(\\mathbf{U} \\in \\mathbb{R}^{3 \\times n}\\), primarily compensate for Gaussian positional adjustments (e.g., hair fluttering, clothing folds) across different poses. Meanwhile, the color blend shapes, \\(\\mathbf{C} \\in \\mathbb{R}^{3 \\times n}\\), address appearance changes, such as shadow variations caused by self-obscuration. To obtain the corresponding driving coefficients, we employ a mapping network \\(\\mathcal{H}\\) for the head using the expression parameters \\(\\epsilon\\) as input and a mapping network \\(\\mathcal{B}\\) for the body which inputs the body pose parameters \\(\\theta\\), to get the coefficients \\(\\mathbf{z}_h \\in \\mathbb{R}^{n_h}\\) for the head and \\(\\mathbf{z}_b \\in \\mathbb{R}^{n_b}\\) for the body, respectively, enabling independent and fine-grained control over head and body deformations. The compensation of position and color for each Gaussian can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.456, + 0.37, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\delta \\mathbf {u} = B S \\left(\\mathbf {z} _ {h} \\oplus \\mathbf {z} _ {b}; \\mathbf {U}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.474, + 0.482, + 0.491 + ], + "angle": 0, + "content": "\\[\n\\delta \\mathbf {c} = B S \\left(\\mathbf {z} _ {h} \\oplus \\mathbf {z} _ {b}; \\mathbf {C}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.56 + ], + "angle": 0, + "content": "where \\(\\oplus\\) is a concatenation operation, resulting in the size \\(n = n_h + n_b\\). The function \\(BS(\\cdot ,\\cdot)\\) represents the blend shapes operation mentioned in SMPL [37]. We add these residuals to local attributes:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.569, + 0.373, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {w} = \\mathbf {p} + \\mathbf {R} (\\gamma \\mathbf {n} + \\delta \\mathbf {u}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.584, + 0.482, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {w} = S H (\\mathbf {R} ^ {- 1} \\mathbf {d}) + \\delta \\mathbf {c}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.689 + ], + "angle": 0, + "content": "For the fine-tuning stage on training data, we freeze the non-rigid deformation field network \\(S\\), optimize the two mapping networks \\(\\mathcal{H}\\) and \\(\\mathcal{B}\\), and two blend shapes \\(\\mathbf{U}\\) and \\(\\mathbf{C}\\). The loss during fine-tuning is the same as that pre-training the teacher network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.751 + ], + "angle": 0, + "content": "After baking mesh non-rigid deformation and applying Gaussian non-rigid compensation, the performance of our student model is significantly enhanced compared to the teacher model, while still maintaining high quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.764, + 0.224, + 0.78 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.299, + 0.805 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. To evaluate the performance on the full-body talking task, we introduce a new dataset, TalkBody4D, designed for common full-body talking scenarios in everyday life, primarily characterized by rich mouth movements and diverse gestures with synchronous audios. The dataset comprises 4 distinct identities, each wearing 2 different outfits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "Each talking sequence consists of rough 6k frames with 4K image resolution, 60 views with 48 full-body, and 12 face-region views. To evaluate performance on complex motions and expressions, we select 4 sequences (04, 05, 06, 08) from the ActorHQ [22] dataset, supplement with 2 dancing and 2 exaggerated expression sequences. The resolution of our training and testing images is \\(1500 \\times 2000\\) for all experiments. SMPLX registrations for these sequences can be obtained by first initializing SMPLX parameters using existing tools [1, 2, 45], followed by fine-tuning the poses with a photometric loss to improve geometry alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Implementation Details. The clothed parametric template has roughly \\(22k\\) vertices and \\(45k\\) faces in total, which contains additional \\(23k\\) faces for clothes, hair, and shoes, compared to the naive SMPLX. We randomly initialize about \\(220k\\) Gaussians in total to bind with the template as texture, for each triangle containing roughly 4 to 6, and optimize these Gaussians for \\(10k\\) iterations using the multi-view images of the reference frame. We pre-train our StyleUnet-based[34] teacher network for \\(600k\\) iterations, and set the loss weights \\(\\lambda_{ssim} = 0.2\\), \\(\\lambda_{lpips} = 0.01\\), \\(\\lambda_{nor} = 0.02\\) and batch size is 1 during the training process. Then we bake the teacher network into a small 5-layer MLP-based student network, which set \\(\\lambda_{non} = 0.1\\), \\(\\lambda_{sem} = 1.0\\) and optimize \\(30k\\) iterations. Finally, we finetune the mapping networks and blend shapes for Gaussian non-rigid compensation on training data \\(100k\\) iterations at the batch size of 4. The size of two blend shapes for position and color is \\(n = 28\\), in which \\(n_h = 8\\) for the head and \\(n_b = 20\\) for the body." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.645, + 0.56 + ], + "angle": 0, + "content": "4.2. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.703 + ], + "angle": 0, + "content": "We conduct an extensive comparison between our method and recent state-of-the-art full-body avatar methods include GaussianAvatar [20], 3DGS-Avatar [46], MeshAvatar [10], and AnimatableGS [34]. All of these methods learn pose-dependent non-rigid deformations in the canonical space and then utilize the skeleton-based rigid transformations of SMPL [37] or SMPLX [43] into observation space to model the dynamic human. We uniformly use SMPLX [43] across all these methods to ensure a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparison on Full-body Talking. We conduct experiments on typical full-body talking scenarios encountered in daily life, which include a variety of mouth movements and gestures. We provide the quantitative comparison on full-body talking as shown in Tab. 1. We also show the qualitative comparison in Fig. 3, and our method can generate more realistic and detailed results, especially in the face region. 3DGS-Avatar [46] directly learns two MLP-based networks to handle the pose-dependent non-rigid deformation of Gaussians but produces very blurry results due to the low-frequency bias of MLPs. GaussianAvatar [20] uses a 2D CNN network by encoding the position map in SMPLX's UV space to generate the non-rigid deformation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10727" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.905, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.64, + 0.908, + 0.669 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons on full-body talking tasks. Our method outperforms state-of-the-art methods by producing clearer clothing dynamics and enhanced facial details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.486, + 0.895 + ], + "angle": 0, + "content": "maps of Gaussians. However, it is restricted by the mini-clothed topology of SMPLX, struggling to handle loose skirts. MeshAvatar [10] chooses mesh as a human representation, which is insufficient to express the geometry of details, such as hair, and glasses. AnimatableGS [34] can generate high-quality dynamic appearance and complex non-rigid deformation. However, it is restricted by the implicit template of learning from scratch, which makes it difficult to handle fine-grained expressions, like blinking and flashing teeth as shown in the face region in Fig. 3. Therefore we adopt the SMPLX++ model as the template whose face and hands retain the native SMPLX prior. However, due to the amount parameters of the teacher StyleUnet [52], it achieves poor performance and is hard to run on a mobile" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.759 + ], + "angle": 0, + "content": "device in real time. Through baking non-rigid deformation, we can achieve \\(150+\\) FPS at the resolution of \\(1500 \\times 2000\\) on Nvidia RTX4090 using only a small MLP-based mesh non-rigid deformation network and two lightweight learnable blend shapes as shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparison on Complex Motion and Expression. Besides full-body talking scenes, our method is also capable of reconstructing complex motions as well as exaggerated expressions as shown in Fig. 4. We provide quantitative comparisons in Tab. 2, our student network achieves competitive results compared to the teacher network, outperforming other methods. Especially for exaggerated expressions, our method significantly outperforms others in reconstructing the face region as shown in Tab. 2. Moreover, our method" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10728" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.905, + 0.219 + ], + "angle": 0, + "content": "
Novel ViewNovel Gestures and ExpressionSpeed
ModelPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓FPS
GaussianAvatar[20]26.58 (23.57).9313 (.8159).10577 (.25242)25.99 (23.15).9232 (.8092).12265 (.26207)54
3DGS-Avatar[46]28.91 (23.95).9411 (.8303).07984 (.20450)26.46 (23.32).9157 (.8184).11804 (.21632)55
MeshAvatar[10]28.53 (24.55).9360 (.8083).09470 (.25572)27.08 (23.58).9229 (.7965).10783 (.24947)22
AnimatableGS[34]32.50 (26.42).9599 (.8587).06695 (.19535)28.05 (23.68).9328 (.8142).09191 (.22673)16
Ours (Teacher)33.45 (27.01).9649 (.8741).04986 (.15613)28.28 (24.28).9336 (.8291).07385 (.18176)16
Ours (Student)33.81 (27.80).9689 (.8975).06437 (.14218)28.38 (24.99).9389 (.8525).08874 (.13364)156
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.229, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on full-body talking task. The results inside the parentheses are evaluated for the face area, and the inference speed is evaluated on Nvidia RTX4090 when rendering images at a resolution of \\( {1500} \\times {2000} \\) ." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.263, + 0.861, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.53, + 0.907, + 0.546 + ], + "angle": 0, + "content": "Figure 4. Results in challenging scenarios. Our method can obtain high-quality reconstruction even for challenging pose and expressions." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.567, + 0.491, + 0.666 + ], + "angle": 0, + "content": "
ModelPSNR↑SSIM↑LPIPS↓
GaussianAvatar[20]25.94 (24.33).9294 (.8251).10478 (.24179)
3DGS-Avatar[46]30.04 (25.08).9403 (.8458).08471 (.18044)
MeshAvatar[10]28.51 (24.94).9334 (.8100).08846 (.23517)
AnimatableGS[34]31.81 (26.79).9493 (.8608).07586 (.19521)
Ours (Teacher)32.80 (27.40).9533 (.8768).05581 (.14996)
Ours (Student)32.72 (27.35).9579 (.8836).07326 (.13914)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.677, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons about complex motions and expressions reconstruction. The results inside the parentheses are evaluated for face regions with exaggerated expressions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.765 + ], + "angle": 0, + "content": "can obtain high-quality normals as shown in Fig. 4, which can be used for image-based relighting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparison on Novel Expression and Novel Gesture. Benefiting from our parametric human model SMPLX++, our full-body avatars can be expression-driven and gesture-driven. We provide quantitative comparisons of self-driven on full-body talking scenes as shown in Tab. 1. As shown in Fig. 5, we provide visualizations of different characters driven by the same skeleton. In addition, our model can accept expression parameter inputs generated from audio by UniTalker [14] as shown in Fig. 5. More demos are pro" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.57, + 0.763, + 0.587 + ], + "angle": 0, + "content": "vided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.667, + 0.618 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.671 + ], + "angle": 0, + "content": "In this subsection, we discuss the key contributions of our method. Additional quantitative results and experiments are available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.809 + ], + "angle": 0, + "content": "The Impact of Template. The choice of template has a crucial impact on our approach. If we directly use SMPLX as our geometric template, as shown in Tab. 3 (w SMPLX), leading to poor results with numerous artifacts, especially for loose skirts as seen in Fig. 6. After extending the geometry of SMPLX to a complete template with clothes, hair, and shoes, this quality is significantly improved as shown in Tab. 3 (w/o Mesh Non.), which allows our method to reconstruct more challenging clothes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The Impact of Non-rigid Deformation. Despite having a complete geometry, the rigid transformations of the skeleton are not able to properly represent non-rigid transformations such as skirt swinging and hair flying. We concurrently introduce mesh non-rigid deformation and Gaussian non-rigid deformation in our method. The geometry with-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10729" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.48, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.495, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Figure 5. Novel pose and expression animation. TaoAvatar can be driven by the same skeleton and expression parameters vividly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.483, + 0.646 + ], + "angle": 0, + "content": "out mesh non-rigid deformation is unable to attach to the surface of the character as shown in Fig. 6 (w/o Mesh Non.). The Gaussian non-rigid compensation also plays a crucial role in quantitative metrics. Without the blend shapes for Gaussian non-rigid compensation, the results are prone to produce artifacts due to imprecise geometry and variable appearance as shown in Fig. 6 (w/o Gau Non.)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.483, + 0.753 + ], + "angle": 0, + "content": "The Impact of Baking Teacher. Compared to directly training a small MLP-based network to handle non-rigid deformations as shown in Tab. 3 (w/o Teacher), our proposed multi-stage training strategy is more efficient. With direct supervision of the non-rigid deformation of the teacher network, it is easier for the student network to decouple the geometry deformation and appearance variations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.214, + 0.787 + ], + "angle": 0, + "content": "5. Application" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "TaoAvatar offers a lightweight, comprehensive representation for 3D talking body scenarios, easily deployable on various mobile devices using existing tools [25] as demonstrated in Fig. 1. It seamlessly integrates with AI models to enable real-time dialogue interactions. we deployed a 3D digital human agent on the Apple Vision Pro, which interacts with users through an ASR-LLM-TTS pipeline [16, 31," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.299, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 6. Ablation Study. Red and Green boxes show artifacts and their improved counterparts." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.34, + 0.918, + 0.432 + ], + "angle": 0, + "content": "
ModelPSNR↑SSIM↑LPIPS↓P2S↓Chamfer↓
w SMPLX28.47.9476.07899.7690.9995
w/o Mesh Non.32.10.9734.03814.4877.5007
w/o Gau Non.31.16.9686.03932.2968.3068
w/o Teacher32.67.9751.03769.5236.5359
Full33.29.9772.03464.2953.3052
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.438, + 0.907, + 0.467 + ], + "angle": 0, + "content": "Table 3. Ablation Study. We use the mesh reconstructed by NeuS2 [54] as a pseudo-truth for geometry evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.907, + 0.556 + ], + "angle": 0, + "content": "39, 62]. Facial expressions and gestures are dynamically controlled by an Audio2BS model [14], allowing the agent to respond naturally with synchronized speech, expressions, and movements. A live demonstration is available in our supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.569, + 0.628, + 0.585 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.745 + ], + "angle": 0, + "content": "Conclusion. In this work, we present TaoAvatar, a lightweight, lifelike full-body talking avatar solution. We demonstrate how the teacher-student framework captures high-definition facial and body details while ensuring real-time performance on AR devices. TaoAvatar can be driven by diverse signals, including facial expressions, hand gestures, and body poses. Through both quantitative and qualitative evaluations, we showcase the advantages of our approach. Additionally, we validate its practical potential with a real-world application on the Apple Vision Pro." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.836 + ], + "angle": 0, + "content": "Limitations and Future Works. TaoAvatar encounters challenges in modeling flexible clothing deformation under exaggerated body poses, which are out-of-distribution of training data. A possible solution is to integrate GNN simulators [17, 18] handling larger hemlines, which are compatible with our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.836, + 0.906, + 0.898 + ], + "angle": 0, + "content": "Potential Social Impact. TaoAvatar can synthesize life-like talking digital humans within an augmented reality environment, generating fabricated 3D content or 2D videos. Therefore, responsible use of this technology is essential." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10730" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.475, + 0.129 + ], + "angle": 0, + "content": "[1] https://github.com/zju3dv/EasyMocap.5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.131, + 0.475, + 0.144 + ], + "angle": 0, + "content": "[2] https://github.com/ShenhanQian/VHAP.5,13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.145, + 0.483, + 0.214 + ], + "angle": 0, + "content": "[3] Rinet Abdrashitov, Kim Raichstat, Jared Monsen, and David Hill. Robust skin weights transfer via weight inpainting. In SIGGRAPH Asia 2023 Technical Communications, SA Technical Communications 2023, Sydney, NSW, Australia, December 12-15, 2023, pages 25:1-25:4. ACM, 2023. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.312 + ], + "angle": 0, + "content": "[4] Thiemo Alldieck, Marcus A. Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 8387-8397. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[5] Thiemo Alldieck, Marcus A. Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1175-1186. Computer Vision Foundation / IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 5419-5429. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.536 + ], + "angle": 0, + "content": "[7] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[8] Xu Chen, Yufeng Zheng, Michael J. Black, Otmar Hilliges, and Andreas Geiger. SNARF: differentiable forward skinning for animating non-rigid neural implicit shapes. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 11574-11584. IEEE, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[9] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Yushuo Chen, Zerong Zheng, Zhe Li, Chao Xu, and Yebin Liu. Meshavatar: Learning high-quality triangular human avatars from multi-view videos. CoRR, abs/2407.08414, 2024. 2, 3, 5, 6, 7, 13, 14, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 19925-19936. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Edilson De Aguiar, Christian Theobalt, Carsten Stoll, and Hans-Peter Seidel. Marker-less deformable mesh tracking" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "for human shape and motion capture. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[13] Helisa Dhamo, Yinyu Nie, Arthur Moreau, Jifei Song, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Headgas: Real-time animatable head avatars via 3d gaussian splatting. In ECCV (2), pages 459-476. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through A unified model. CoRR, abs/2408.00762, 2024. 7, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[15] Yao Feng, Weiyang Liu, Timo Bolkart, Jinlong Yang, Marc Pollefeys, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. CoRR, abs/2309.06441, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.308, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[16] Zhifu Gao, Shiliang Zhang, Ian McLoughlin, and Zhijie Yan. Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to-end speech recognition. In 23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18-22, 2022, pages 2063-2067. ISCA, 2022. 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[17] Artur Grigorev, Michael J. Black, and Otmar Hilliges. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16965-16974. IEEE, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[18] Artur Grigorev, Giorgio Becherini, Michael J. Black, Otmar Hilliges, and Bernhard Thomaszewski. Contourcraft: Learning to resolve intersections in neural multi-garment simulations. In ACM SIGGRAPH 2024 Conference Papers, SIGGRAPH 2024, Denver, CO, USA, 27 July 2024-1 August 2024, page 81. ACM, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[19] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics, 40(4), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[20] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 634-644. IEEE, 2024. 2, 3, 5, 7, 14, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[21] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20418-20431. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[22] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Trans. Graph., 42(4):160:1-160:12, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In IEEE/CVF Conference on Computer Vision and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10731" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5595-5605. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[24] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16922-16932, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[25] Xiaotang Jiang, Huan Wang, Yiliu Chen, Ziqi Wu, Lichuan Wang, Bin Zou, Yafeng Yang, Zongyang Cui, Yu Cai, Tianhang Yu, Chengfei Lyu, and Zhihua Wu. MNN: A universal and efficient inference engine. In Proceedings of the Third Conference on Machine Learning and Systems, MLSys 2020, Austin, TX, USA, March 2-4, 2020. mlsys.org, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.275, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[26] Yujiang Jiang, Qingmin Liao, Xiaoyu Li, Li Ma, Qi Zhang, Chaopeng Zhang, Zongqing Lu, and Ying Shan. UV gaus-sians: Joint learning of mesh deformation and gaussian textures for human avatar modeling. CoRR, abs/2403.11589, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[27] Yujiang Jiang, Qingmin Liao, Zhaolong Wang, Xiangru Lin, Zongqing Lu, Yuxi Zhao, Hanqing Wei, Jingrui Ye, Yu Zhang, and Zhijing Shao. Smplx-lite: A realistic and drivable avatar benchmark with rich geometry and texture annotations. In IEEE International Conference on Multimedia and Expo, ICME 2024, Niagara Falls, ON, Canada, July 15-19, 2024, pages 1-6. IEEE, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[28] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5322-5332. IEEE, 2024. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[30] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: human gaussian splats. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 505-515. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[31] Jungil Kong, Jihoon Park, Beomjeong Kim, Jeongmin Kim, Dohee Kong, and Sangjin Kim. VITS2: improving quality and efficiency of single-stage text-to-speech with adversarial learning and architecture design. In 24th Annual Conference of the International Speech Communication Association, Interspeech 2023, Dublin, Ireland, August 20-24, 2023, pages 4374-4378. ISCA, 2023. 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[32] Youngjoong Kwon, Baole Fang, Yixing Lu, Haoye Dong, Cheng Zhang, Francisco Vicente Carrasco, Albert Mosella-Montoro, Jianjin Xu, Shingo Takagi, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Generalizable human gaussians for sparse view synthesis. In ECCV (78), pages 451-468. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[33] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. GART: gaussian articulated template models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19876-19887. IEEE, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[34] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19711-19722. IEEE, 2024. 2, 3, 5, 6, 7, 13, 14, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[35] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3d gaussian splatting for inverse rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21644-21653. IEEE, 2024. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[36] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: neural free-view synthesis of human actors with pose control. ACM Trans. Graph., 40(6):219:1-219:16, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[38] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[39] Chengfei Lv, Chaoyue Niu, Renjie Gu, Xiaotang Jiang, Zhaode Wang, Bin Liu, Ziqi Wu, Qiulin Yao, Congyu Huang, Panos Huang, Tao Huang, Hui Shu, Jinde Song, Bin Zou, Peng Lan, Guohuan Xu, Fei Wu, Shaojie Tang, Fan Wu, and Guihai Chen. Walle: An end-to-end, general-purpose, and large-scale production system for device-cloud collaborative machine learning. In 16th USENIX Symposium on Operating Systems Design and Implementation, OSDI 2022, Carlsbad, CA, USA, July 11-13, 2022, pages 249-265. USENIX Association, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.538, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 13, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[41] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In CVPR, pages 788-798. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1, 2, 3, 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[45] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gauss" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10732" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "sianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20299-20309. IEEE, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.231 + ], + "angle": 0, + "content": "[46] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5020-5030. IEEE, 2024. 2, 5, 7, 14, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[47] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 130-141. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[48] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 1606-1616. IEEE, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[49] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16911-16921. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[50] Carsten Stoll, Juergen Gall, Edilson De Aguiar, Sebastian Thrun, and Christian Theobalt. Video-based reconstruction of animatable human characters. ACM Transactions on Graphics (TOG), 29(6):1-10, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[51] Daniel Vlasic, Pieter Peers, Ilya Baran, Paul E. Debevec, Jovan Popovic, Szymon Rusinkiewicz, and Wojciech Matusik. Dynamic shape capture using multi-view photometric stereo. ACM Trans. Graph., 28(5):174, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[52] Lizhen Wang, Xiaochen Zhao, Jingxiang Sun, Yuxiang Zhang, Hongwen Zhang, Tao Yu, and Yebin Liu. Stylea-vatar: Real-time photo-realistic portrait avatar from a single video. In ACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH 2023, Los Angeles, CA, USA, August 6-10, 2023, pages 67:1-67:10. ACM, 2023. 3, 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[53] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 550-560. IEEE, 2024. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[54] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3272-3283. IEEE, 2023. 2, 3, 4, 8, 13, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[55] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 15" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[56] Jing Wen, Xiaoming Zhao, Zhongzheng Ren, Alexander G. Schwing, and Shenlong Wang. Gomavatar: Efficient animatable human modeling from monocular video using gaussians-on-mesh. In CVPR, pages 2059-2069. IEEE, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[57] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16189-16199. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[58] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[59] Donglai Xiang, Timur M. Bagautdinov, Tuur Stuyck, Fabian Prada, Javier Romero, Weipeng Xu, Shunsuke Saito, Jingfan Guo, Breannan Smith, Takaaki Shiratori, Yaser Sheikh, Jessica K. Hodgins, and Chenglei Wu. Dressing avatars: Deep photorealistic appearance for physically simulated clothing. ACM Trans. Graph., 41(6):222:1-222:15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.456 + ], + "angle": 0, + "content": "[60] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In CVPR, pages 1802-1812. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[61] Yuelang Xu, Bengwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 1931-1941. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[62] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. Qwen2 technical report. CoRR, abs/2407.10671, 2024. 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[63] Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. In CVPR, pages 896-905. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[64] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595. Computer Vision Foundation / IEEE Computer Society, 2018. 4, 15" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10733" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[65] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humanerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[66] Zhongyuan Zhao, Zhenyu Bao, Qing Li, Guoping Qiu, and Kanglin Liu. Psavatar: A point-based shape model for real-time head avatar animation with 3d gaussian splatting, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[67] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CoRR, abs/2401.03407, 2024. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[68] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas J. Guibas, and Gordon Wetzstein. Physavatar: Learning the physics of dressed 3d avatars from visual observations. CoRR, abs/2404.04421, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[69] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM Trans. Graph., 42(4):158:1-158:19, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.484 + ], + "angle": 0, + "content": "[70] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4574-4584. IEEE, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10734" + } + ] +] \ No newline at end of file diff --git a/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_origin.pdf b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4431b2931b03c3ef1e6804691192b11aa88ea6c7 --- /dev/null +++ b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/45c8e882-5f08-40bd-989e-2c96f33116f0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aeaca02703284e657c909f2cb9fe60c1222691035c6d0b9b346fadd5ba7c52c5 +size 9984094 diff --git a/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/full.md b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d27ba103d2d5ad593e76d5c39afbcb0ee31dd56c --- /dev/null +++ b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/full.md @@ -0,0 +1,299 @@ +# TaoAvatar: Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting + +Jianchuan Chen,\* Jingchuan Hu,\* Gaige Wang,\* Zhonghua Jiang +Tiansong Zhou, Zhiwen Chen\*, Chengfei Lv\* +Alibaba Group, Hangzhou, China + +![](images/e6de75c108a6d16181aa0ae30ea0d797e628fcf6f4302f98a0b4aaec16eb8a37.jpg) +Figure 1. TaoAvatar generates photorealistic, topology-consistent 3D full-body avatars from multi-view sequences. It provides high-quality, real-time rendering with low storage requirements, compatible across various mobile and AR devices like the Apple Vision Pro. + +![](images/33d932f309550bd7b5d915d8ec306113b28eb2dd54dd2b3da55006a9f95d3678.jpg) + +# Abstract + +Realistic 3D full-body talking avatars hold great potential in AR, with applications ranging from e-commerce live streaming to holographic communication. Despite advances in 3D Gaussian Splatting (3DGS) for lifelike avatar creation, existing methods struggle with fine-grained control of facial expressions and body movements in full-body talking tasks. Additionally, they often lack sufficient details and cannot run in real-time on mobile devices. We present TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar driven by various signals. Our approach starts by creating a personalized clothed human parametric template that binds Gaussians to represent appearances. We then pre-train a StyleUnet-based network to handle complex pose-dependent non-rigid deformation, which can capture high-frequency appearance details but is too resource-intensive for mobile devices. To overcome this, we "bake" the non-rigid deformations into a lightweight MLP-based network using a distillation technique and develop blend shapes to compensate for details. Extensive experiments show that TaoAvatar achieves state-of-the-art rendering quality while running in real-time across various devices, maintaining 90 FPS on high-definition stereo devices such as the Apple Vision Pro. + +# 1. Introduction + +Creating lifelike 3D human avatars is a dynamic and rapidly advancing research area in computer vision and graphics, essential for applications in AR/VR, 3D entertainment, and virtual communication. Despite decades of research, achieving realistic and immersive avatars on mobile devices—particularly AR—remains challenging due to computational limitations. Current industry solutions, such as MetaHuman, often rely on high-precision scans and extensive manual effort by artists for 3D modeling, rigging, and motion capture. + +In recent years, academic researchers have adopted parametric models [37, 43] for human representation, which are constructed from massive body scans and can capture a performer's shape, expression, and pose from images or videos. These methods [4-6, 27] extend the naked model [37, 43] by adding offsets to represent tailored clothing, but are still struggling with complex geometries and high-frequency details, such as loose skirts and fine hair, due to limitations of topology and texture resolution. With the rise of neural radiance fields (NeRF) [40], many researchers tend to define implicit neural human representations in canonical space, animated by backward linear blend skinning [44, 65] of these parametric models [37, 43]. While these methods can handle arbitrary topology and deliver higher rendering quality, they are highly time-consuming due to the large amount of sampling required + +during volume rendering. Although many efforts [9, 24] have significantly improved rendering efficiency by introducing iNGP [42], the animated results still fail to escape the Uncanny Valley effect when driven. Recently, 3D Gaussian Splatting (3DGS) [29], an explicit and efficient point-based representation, has gained tremendous attention from researchers for its ability to deliver both high-quality rendering and real-time performance. Compared to implicit representations [40, 54], the explicit point-based representation is more compatible with being integrated with the mesh-based parametric models [37, 43], which can be directly driven by forward linear blend skinning. Many recent methods [20, 34, 46, 48], combining 3DGS and the parametric models [37, 43], have made significant strides toward realizing lifelike 3D human avatars. However, achieving real-time rendering on mobile devices remains challenging, especially for AR glasses, which require high resolution, stereo rendering, and seamless interaction. + +We introduce TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar designed for running in real-time on Augmented Reality (AR) devices. Unlike previous approaches [8, 10, 34], which develop implicit parameterized models from scratch, we construct a personalized, clothed parameterized template that preserves the facial expressions and gesture controls inherent to the native SMPLX [43]. At the same time, we bind the Gaussians to the triangles as texture to create a hybrid avatar representation. Inspired by [34], we pre-train a large StyleUnet-based teacher network to learn pose-dependent dynamic deformation maps of Gaussians in 2D space by front and back orthogonal projection. Although the teacher network is powerful enough to capture high-frequency appearance details in various poses, its large number of parameters makes it difficult to run in real-time on AR devices. To achieve high-performance rendering, we bake the dynamic deformation of Gaussians into the non-rigid deformation field of the mesh using a distillation technique, employing a compact MLP-based student network. Meanwhile, we propose two lightweight, learnable blend shapes to compensate for the non-rigid deformation of the Gaussians. After fin-tuning, our student model achieves high-performance rendering without compromising quality. + +To evaluate our approach, we propose a multi-view dataset named TalkBody4D, focusing on prevalent full-body talking scenarios encountered in daily life, which includes rich facial expressions and gestures with synchronous audios different from existing full-body motion datasets [11, 22, 69]. Experimental results demonstrate that our method effectively captures high-frequency appearance details of human avatars while maintaining a lightweight architecture capable of rendering stereo images at 2K resolution on common mobile and AR devices, including the Apple Vision Pro. The resulting full-body avatar is highly + +expressive, enabling users to animate it with facial expressions, hand gestures, and body poses, thereby highlighting the potential applications of our method. Our contributions can be summarized as follows: + +- We introduce TaoAvatar, a novel teacher-student framework which yields topologically consistent, well-aligned geometry, further creating a high-fidelity, lightweight, 3DGS-based drivable full-body talking avatar. +- We propose two innovative strategies including non-rigid deformation baking and two lightweight blend shapes compensations to ensure high-quality rendering and efficient performance on mobile and AR devices, achieving 2K resolution rendering at 90 FPS on high-definition stereo devices like the Apple Vision Pro. +- We contribute TalkBody4D, a multi-view dataset to be released, designed for full-body talking scenarios, featuring diverse facial expressions and gestures with synchronous audios. Extensive experiments demonstrate that our approach outperforms other state-of-the-art methods in both quality and performance. + +# 2. Related work + +3D Avatar Representation. Traditional computer graphics techniques [12, 19, 50, 51] use 3D scanning or multiview stereo (MVS) to create full-body meshes, which are then combined with a 3D skeleton to produce drivable templates. However, this approach is both expensive and inefficient. As parametric body models, SMPL [37] and SMPLX [43] are widely used for their versatility but lack detail in simulating loose clothing. Some works [4, 5, 27] have attempted to add clothing offsets to SMPLX, effectively handling tight clothing but struggling with loose garments. Due to the limited expressiveness of traditional meshes, researchers have combined SMPL with advanced rendering techniques. Some methods [7, 15, 23, 36, 57] use NeRF [40] to implicitly express human bodies, mapping points from the observation space backward to canonical space. Although high-quality, volume rendering is slow and poses challenges for real-time applications. In contrast, 3DGS [29] uses 3D Gaussian distributions for scenes, offering real-time rendering [38, 58], fast training, and the ability to handle complex materials such as hair and translucency. Some methods [34, 46, 48, 56, 63] combining SMPLX [43] and 3DGS [29], like GaussianAvatar [20], use explicit GS point clouds and forward skinning to efficiently simulate the deformations, offering good rendering quality and speed. This representation presents potential solutions for achieving both real-time performance and computational efficiency. + +Dynamic Nonrigid Deformation Modeling. Various methods have been developed to simulate non-rigid human deformation realistically. The MLP-based methods [21, 30, 33, 41], like 3DGS-Avatar [46], are noted for lightweight + +flexibility but struggle with handling novel poses. Physical simulation methods [59, 68] produce physics-based animations but require unified topology, and complex modeling, and are computationally expensive, which negatively impacts real-time performance. Additionally, some methods [17, 18] use GNNs to learn non-rigid deformations in a data-driven manner. These methods are faster than traditional cloth simulations. Recently, dynamic texture methods are emerging, these methods [26, 32, 36] model complex details in the 2D texture space, while AnimatableGS [34] and MeshAvatar [10] enhance modeling with front and back projection strategies which avoid texture unwarping. + +3D Drivable Talking Avatars. Research on drivable digital humans is rapidly advancing for various applications. In 3D Talking Heads, methods [13, 45, 47, 60, 61, 66] excel in capturing and driving facial expressions, enhancing the realism of virtual characters. For drivable bodies, these methods [8, 20, 34, 56, 70] achieve precise restoration of body movements by learning pose-dependent deformation fields. Furthermore, some methods [49, 59] not only support detailed body modeling and animation but also enable head expressions and lip-syncing. Despite enhancing realism, running these methods in real-time on mobile or AR devices is challenging due to their high computational demands, exceeding current edge platform capabilities. Optimizing computational efficiency without compromising high-quality output is a key research focus. + +# 3. Method + +Leveraging multi-view RGB videos of a performer alongside corresponding SMPLX registrations for each frame, we aim to develop a high-fidelity, lightweight full-body talking avatar capable of high-resolution and high-performance rendering on mobile devices. To achieve this, we first create a hybrid clothed parametric representation by integrating 3D Gaussian Splatting (3DGS) [29] and SMPLX [43], which is more expressive for simulating loose clothing and hair, while maintaining high-quality rendering (Sec. 3.1). Next, we propose a teacher-student framework to reconstruct high-frequency, pose-dependent dynamic details as illustrated in Fig. 2. By leveraging non-rigid deformation baking and incorporating two lightweight blend shape compensations, our method achieves a tradeoff between high quality and high performance (Sec. 3.2). + +# 3.1. Hybrid Clothed Parametric Representations + +The Clothed Extension of SMPLX. To extend SMPLX to complete clothed human geometry with clothes, hair, etc., we choose a frame close to T-pose as a reference, which provides more visible details and less sticky geometry. First, we utilize NeuS2 [54] to reconstruct the geometry of this reference frame and parse out non-body components (e.g., clothes) from it using [53]. However, the body's skeleton + +cannot directly drive these components. We estimate the shape and pose of the SMPLX [37] in the current frame and use robust skinning transfer [3] to propagate the body's skinning weights to these non-body components. Finally, we apply inverse skinning to transform these components back to the standard T-pose, resulting in a comprehensive personalized parametric model, referred to as SMPLX++. + +Binding Gaussians on Mesh as Texture. After obtaining the complete geometric model, we bind 3D Gaussians to the mesh's triangles as textures, enabling them to move synchronously with the mesh. Inspired by these hybrid representations [45, 48], we define the attributes of the Gaussians in a local coordinate system relative to the triangles. Specifically, for each triangle, we randomly initialize $k$ Gaussians, each of which maintains the local attributes $\{f,(u,v),\gamma ,\mathbf{r},\mathbf{s},o,\mathbf{sh}\}$ , where $f$ is the index of the parent triangle, $(u,v)$ is the barycentric coordinates, $\gamma$ represents the translation along the triangle's normal direction, $\mathbf{r}$ and $\mathbf{s}$ denote rotation and scaling in the local space, $\mathbf{o}$ indicates opacity, and $\mathbf{sh}$ consists of spherical harmonic coefficients. Additionally, we define the normal of each Gaussian as $\mathbf{n}_g = [1,0,0]$ within the local space, ensuring consistency with the mesh's normal for subsequent transformations. To deform each Gaussian from the local to the world space, we construct a transformation based on its parent triangle $F_{f}(\mathbf{v}_{1},\mathbf{v}_{2},\mathbf{v}_{3})$ , where $\mathbf{v}_1$ , $\mathbf{v}_2$ , $\mathbf{v}_3$ are the vertices of the triangle $F_{f}$ in the world space: + +$$ +\mathbf {p} = u \cdot \mathbf {v} _ {1} + v \cdot \mathbf {v} _ {2} + (1 - u - v) \cdot \mathbf {v} _ {3}, +$$ + +$$ +\mathbf {R} = [ \mathbf {n}, \mathbf {q}, \mathbf {n} \times \mathbf {q} ], \mathbf {q} = \frac {\left(\mathbf {v} _ {2} + \mathbf {v} _ {3}\right) / 2 - \mathbf {v} _ {1}}{\left\| \left(\mathbf {v} _ {2} + \mathbf {v} _ {3}\right) / 2 - \mathbf {v} _ {1} \right\|}, \tag {1} +$$ + +$$ +e = \left(\left\| \mathbf {v} _ {1} - \mathbf {v} _ {2} \right\| + \left\| \mathbf {v} _ {2} - \mathbf {v} _ {3} \right\| + \left\| \mathbf {v} _ {1} - \mathbf {v} _ {3} \right\|\right) / 3, +$$ + +where $\mathbf{n} = \frac{(\mathbf{v}_1 - \mathbf{v}_2)\times(\mathbf{v}_3 - \mathbf{v}_1)}{\|\mathbf{v}_1 - \mathbf{v}_2\|\cdot\|\mathbf{v}_3 - \mathbf{v}_1\|}$ is the normal of the triangle, $\mathbf{p}$ denotes the surface point on the parent triangle, and $e$ denotes the average edge length. + +$$ +\mathbf {u} _ {w} = \mathbf {p} + \gamma \mathbf {R n}, \quad \mathbf {r} _ {w} = \mathbf {R r}, \quad \mathbf {s} _ {w} = e \cdot \mathbf {s}, \tag {2} +$$ + +here, $\mathbf{u}_w$ , $\mathbf{r}_w$ , $\mathbf{s}_w$ represent the position, rotation, and scaling in the world spaces after transformation, respectively. Furthermore, $\mathbf{c}_w = SH(\mathbf{sh},\mathbf{R}^{-1}\mathbf{d})$ denotes the color along the view direction $\mathbf{d}$ in the world space. + +# 3.2. Dynamic Mesh-based Gaussian Reconstruction + +Learning Dynamic Non-Rigid Gaussian Deformation Maps. Although SMPLX++ can be directly driven by the skeleton, linear blend skinning is insufficient for performing dynamic non-rigid deformation such as clothing folds and swaying motions. Like dynamic Gaussian avatars [26, 34], we utilize a large StyleUnet [52] as the teacher network to capture complex pose-dependent dynamic non-rigid deformation of Gaussians in 2D texture space. Following [34], + +![](images/ffc1c21440c2523504d70f66488fb12509676fa526cc770024f9f855c5f36072.jpg) +Figure 2. Illustration of our Method. Our pipeline begins by reconstructing (a) a cloth-extended SMPLX mesh with aligned Gaussian textures. To address complex dynamic non-rigid deformations, (b) we employ a teacher StyleUnet to learn pose-dependent non-rigid maps, which are then baked into a lightweight student MLP to infer non-rigid deformations of our template mesh. For high-fidelity rendering, (c) we introduce learnable Gaussian blend shapes to enhance appearance details. + +we rasterize the T-posed mesh, which is colored using posed coordinates and segmentation colors, to obtain the front and back position maps $\{\mathcal{P}_f,\mathcal{P}_b\}$ . These position maps are then fed into the StyleUnet network along with view directions to generate the non-rigid deformation maps $\{\Delta \mathcal{U}_f,\Delta \mathcal{U}_b\}$ and other residual Gaussian attribute maps. We add these residuals to the Gaussian's local attributes and transform them into the world space according to Eq. (2). To train the StyleUnet-based teacher network $\mathcal{T}$ , we adopt the losses including L1, D-SSIM [29], and perceptual loss [64]. Additionally, we introduce a normal loss $\mathcal{L}_{nor}$ . Consequently, the total reconstruction loss $\mathcal{L}_{rec}$ is defined as: + +$$ +\mathcal {L} _ {\text {r e c}} = \mathcal {L} _ {1} + \lambda_ {\text {s s i m}} \mathcal {L} _ {\text {s s i m}} + \lambda_ {\text {l p i p s}} \mathcal {L} _ {\text {l p i p s}} + \lambda_ {\text {n o r}} \mathcal {L} _ {\text {n o r}}, \tag {3} +$$ + +where $\lambda_{*}$ is the loss weights. The normal loss is defined as $\mathcal{L}_{nor} = \| N_t - N_g\|$ , where $N_{t}$ is the rendered normal image by the teacher network, and the ground truth normal image $N_{g}$ can be obtained from the method [54]. + +Baking Non-Rigid Gaussian Deformation into Mesh Deformation. Although the StyleUnet-based network $\mathcal{T}$ can achieve good results, it struggles to run in real-time on mobile devices due to its large number of parameters, especially for AR devices, which require high resolution, stereo rendering, and high performance. Inspired by knowledge distillation, we bake the Gaussian non-rigid deformation of the teacher network into the mesh non-rigid deformation + +field, which is a compact MLP-based student network $S$ . + +$$ +\Delta \bar {\mathbf {v}} _ {i} = \mathcal {S} \left(\bar {\mathbf {v}} _ {i}, \theta , \mathbf {z} _ {t}\right), \tag {4} +$$ + +where $\bar{\mathbf{v}}_i$ is the $i$ -th vertex coordinate in the canonical space, $\theta$ is the pose parameter, and $\mathbf{z}_t$ is a learnable embedding for each frame to compensate for inaccurate pose estimation. To train the student network $\mathcal{S}$ with the capability of pose-dependent non-rigid deformation, we directly supervise its output using the Gaussian non-rigid deformation maps $\{\Delta \mathcal{U}_f,\Delta \mathcal{U}_b\}$ of the teacher network $\mathcal{T}$ . Specifically, we render the mesh non-rigid deformations $\{\Delta \mathbf{v}_i\}_{i = 1}^N$ to front and back deformation maps $\{\Delta \mathcal{V}_f,\Delta \mathcal{V}_b\}$ in the canonical space, and then calculate the differences: + +$$ +\mathcal {L} _ {n o n} = \left\| \Delta \mathcal {V} _ {f} - \Delta \mathcal {U} _ {f} \right\| + \left\| \Delta \mathcal {V} _ {b} - \Delta \mathcal {U} _ {b} \right\|. \tag {5} +$$ + +Furthermore, to mitigate the intersection between clothing and the body, we introduce a semantic loss $\mathcal{L}_{sem}$ . For each vertex, we assign a semantic label $\mathbf{e}_i = \mathbf{c}_i + \sin (\tau \bar{\mathbf{v}}_i)$ that integrates the candidate coordinates $\bar{\mathbf{v}}_i$ and the artificial segmentation color $\mathbf{c}_i$ , and $\tau$ is a scale factor. The semantic label for each Gaussian can be obtained by interpolating the semantic label of the three vertices of its parent triangle. Subsequently, we render the Gaussian semantic map $\mathcal{E}_t$ and the mesh semantic map $\mathcal{E}_s$ in the world space separately and compute the semantic loss $\mathcal{L}_{sem}$ : + +$$ +\mathcal {L} _ {s e m} = \left\| \mathcal {E} _ {s} - \mathcal {E} _ {t} \right\|. \tag {6} +$$ + +In addition, we refine the local attributes by computing the reconstruction loss $\mathcal{L}_{rec}$ , which takes the rendered image of the teacher network as ground truth. The total loss $\mathcal{L}_{bak}$ during baking is: + +$$ +\mathcal {L} _ {b a k} = \mathcal {L} _ {r e c} + \lambda_ {n o n} \mathcal {L} _ {n o n} + \lambda_ {s e m} \mathcal {L} _ {s e m}. \tag {7} +$$ + +Compensating Gaussian Deformation with Lightweight Blend Shapes. Inspired by pose corrective blend shapes in SMPL [37], we build learnable blend shapes for position and color for each Gaussian. The position blend shapes, $\mathbf{U} \in \mathbb{R}^{3 \times n}$ , primarily compensate for Gaussian positional adjustments (e.g., hair fluttering, clothing folds) across different poses. Meanwhile, the color blend shapes, $\mathbf{C} \in \mathbb{R}^{3 \times n}$ , address appearance changes, such as shadow variations caused by self-obscuration. To obtain the corresponding driving coefficients, we employ a mapping network $\mathcal{H}$ for the head using the expression parameters $\epsilon$ as input and a mapping network $\mathcal{B}$ for the body which inputs the body pose parameters $\theta$ , to get the coefficients $\mathbf{z}_h \in \mathbb{R}^{n_h}$ for the head and $\mathbf{z}_b \in \mathbb{R}^{n_b}$ for the body, respectively, enabling independent and fine-grained control over head and body deformations. The compensation of position and color for each Gaussian can be expressed as: + +$$ +\delta \mathbf {u} = B S \left(\mathbf {z} _ {h} \oplus \mathbf {z} _ {b}; \mathbf {U}\right), +$$ + +$$ +\delta \mathbf {c} = B S \left(\mathbf {z} _ {h} \oplus \mathbf {z} _ {b}; \mathbf {C}\right), \tag {8} +$$ + +where $\oplus$ is a concatenation operation, resulting in the size $n = n_h + n_b$ . The function $BS(\cdot ,\cdot)$ represents the blend shapes operation mentioned in SMPL [37]. We add these residuals to local attributes: + +$$ +\mathbf {u} _ {w} = \mathbf {p} + \mathbf {R} (\gamma \mathbf {n} + \delta \mathbf {u}), +$$ + +$$ +\mathbf {c} _ {w} = S H (\mathbf {R} ^ {- 1} \mathbf {d}) + \delta \mathbf {c}. \tag {9} +$$ + +For the fine-tuning stage on training data, we freeze the non-rigid deformation field network $S$ , optimize the two mapping networks $\mathcal{H}$ and $\mathcal{B}$ , and two blend shapes $\mathbf{U}$ and $\mathbf{C}$ . The loss during fine-tuning is the same as that pre-training the teacher network. + +After baking mesh non-rigid deformation and applying Gaussian non-rigid compensation, the performance of our student model is significantly enhanced compared to the teacher model, while still maintaining high quality. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. To evaluate the performance on the full-body talking task, we introduce a new dataset, TalkBody4D, designed for common full-body talking scenarios in everyday life, primarily characterized by rich mouth movements and diverse gestures with synchronous audios. The dataset comprises 4 distinct identities, each wearing 2 different outfits. + +Each talking sequence consists of rough 6k frames with 4K image resolution, 60 views with 48 full-body, and 12 face-region views. To evaluate performance on complex motions and expressions, we select 4 sequences (04, 05, 06, 08) from the ActorHQ [22] dataset, supplement with 2 dancing and 2 exaggerated expression sequences. The resolution of our training and testing images is $1500 \times 2000$ for all experiments. SMPLX registrations for these sequences can be obtained by first initializing SMPLX parameters using existing tools [1, 2, 45], followed by fine-tuning the poses with a photometric loss to improve geometry alignment. + +Implementation Details. The clothed parametric template has roughly $22k$ vertices and $45k$ faces in total, which contains additional $23k$ faces for clothes, hair, and shoes, compared to the naive SMPLX. We randomly initialize about $220k$ Gaussians in total to bind with the template as texture, for each triangle containing roughly 4 to 6, and optimize these Gaussians for $10k$ iterations using the multi-view images of the reference frame. We pre-train our StyleUnet-based[34] teacher network for $600k$ iterations, and set the loss weights $\lambda_{ssim} = 0.2$ , $\lambda_{lpips} = 0.01$ , $\lambda_{nor} = 0.02$ and batch size is 1 during the training process. Then we bake the teacher network into a small 5-layer MLP-based student network, which set $\lambda_{non} = 0.1$ , $\lambda_{sem} = 1.0$ and optimize $30k$ iterations. Finally, we finetune the mapping networks and blend shapes for Gaussian non-rigid compensation on training data $100k$ iterations at the batch size of 4. The size of two blend shapes for position and color is $n = 28$ , in which $n_h = 8$ for the head and $n_b = 20$ for the body. + +# 4.2. Comparison + +We conduct an extensive comparison between our method and recent state-of-the-art full-body avatar methods include GaussianAvatar [20], 3DGS-Avatar [46], MeshAvatar [10], and AnimatableGS [34]. All of these methods learn pose-dependent non-rigid deformations in the canonical space and then utilize the skeleton-based rigid transformations of SMPL [37] or SMPLX [43] into observation space to model the dynamic human. We uniformly use SMPLX [43] across all these methods to ensure a fair comparison. + +Comparison on Full-body Talking. We conduct experiments on typical full-body talking scenarios encountered in daily life, which include a variety of mouth movements and gestures. We provide the quantitative comparison on full-body talking as shown in Tab. 1. We also show the qualitative comparison in Fig. 3, and our method can generate more realistic and detailed results, especially in the face region. 3DGS-Avatar [46] directly learns two MLP-based networks to handle the pose-dependent non-rigid deformation of Gaussians but produces very blurry results due to the low-frequency bias of MLPs. GaussianAvatar [20] uses a 2D CNN network by encoding the position map in SMPLX's UV space to generate the non-rigid deformation + +![](images/0bf0305810caf19bbfc67cf176a043b3929f8cbd206ce269bb72cd330093f867.jpg) +Figure 3. Qualitative comparisons on full-body talking tasks. Our method outperforms state-of-the-art methods by producing clearer clothing dynamics and enhanced facial details. + +maps of Gaussians. However, it is restricted by the mini-clothed topology of SMPLX, struggling to handle loose skirts. MeshAvatar [10] chooses mesh as a human representation, which is insufficient to express the geometry of details, such as hair, and glasses. AnimatableGS [34] can generate high-quality dynamic appearance and complex non-rigid deformation. However, it is restricted by the implicit template of learning from scratch, which makes it difficult to handle fine-grained expressions, like blinking and flashing teeth as shown in the face region in Fig. 3. Therefore we adopt the SMPLX++ model as the template whose face and hands retain the native SMPLX prior. However, due to the amount parameters of the teacher StyleUnet [52], it achieves poor performance and is hard to run on a mobile + +device in real time. Through baking non-rigid deformation, we can achieve $150+$ FPS at the resolution of $1500 \times 2000$ on Nvidia RTX4090 using only a small MLP-based mesh non-rigid deformation network and two lightweight learnable blend shapes as shown in Tab. 1. + +Comparison on Complex Motion and Expression. Besides full-body talking scenes, our method is also capable of reconstructing complex motions as well as exaggerated expressions as shown in Fig. 4. We provide quantitative comparisons in Tab. 2, our student network achieves competitive results compared to the teacher network, outperforming other methods. Especially for exaggerated expressions, our method significantly outperforms others in reconstructing the face region as shown in Tab. 2. Moreover, our method + +
Novel ViewNovel Gestures and ExpressionSpeed
ModelPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓FPS
GaussianAvatar[20]26.58 (23.57).9313 (.8159).10577 (.25242)25.99 (23.15).9232 (.8092).12265 (.26207)54
3DGS-Avatar[46]28.91 (23.95).9411 (.8303).07984 (.20450)26.46 (23.32).9157 (.8184).11804 (.21632)55
MeshAvatar[10]28.53 (24.55).9360 (.8083).09470 (.25572)27.08 (23.58).9229 (.7965).10783 (.24947)22
AnimatableGS[34]32.50 (26.42).9599 (.8587).06695 (.19535)28.05 (23.68).9328 (.8142).09191 (.22673)16
Ours (Teacher)33.45 (27.01).9649 (.8741).04986 (.15613)28.28 (24.28).9336 (.8291).07385 (.18176)16
Ours (Student)33.81 (27.80).9689 (.8975).06437 (.14218)28.38 (24.99).9389 (.8525).08874 (.13364)156
+ +![](images/306b4d44d5fb81c6683afe0b858d0b4ff8dd3d37c0a63d59c10d0321ecd66943.jpg) +Figure 4. Results in challenging scenarios. Our method can obtain high-quality reconstruction even for challenging pose and expressions. + +Table 1. Quantitative comparisons on full-body talking task. The results inside the parentheses are evaluated for the face area, and the inference speed is evaluated on Nvidia RTX4090 when rendering images at a resolution of ${1500} \times {2000}$ . + +
ModelPSNR↑SSIM↑LPIPS↓
GaussianAvatar[20]25.94 (24.33).9294 (.8251).10478 (.24179)
3DGS-Avatar[46]30.04 (25.08).9403 (.8458).08471 (.18044)
MeshAvatar[10]28.51 (24.94).9334 (.8100).08846 (.23517)
AnimatableGS[34]31.81 (26.79).9493 (.8608).07586 (.19521)
Ours (Teacher)32.80 (27.40).9533 (.8768).05581 (.14996)
Ours (Student)32.72 (27.35).9579 (.8836).07326 (.13914)
+ +Table 2. Quantitative comparisons about complex motions and expressions reconstruction. The results inside the parentheses are evaluated for face regions with exaggerated expressions. + +can obtain high-quality normals as shown in Fig. 4, which can be used for image-based relighting. + +Comparison on Novel Expression and Novel Gesture. Benefiting from our parametric human model SMPLX++, our full-body avatars can be expression-driven and gesture-driven. We provide quantitative comparisons of self-driven on full-body talking scenes as shown in Tab. 1. As shown in Fig. 5, we provide visualizations of different characters driven by the same skeleton. In addition, our model can accept expression parameter inputs generated from audio by UniTalker [14] as shown in Fig. 5. More demos are pro + +vided in the supplementary materials. + +# 4.3. Ablation Study + +In this subsection, we discuss the key contributions of our method. Additional quantitative results and experiments are available in the supplementary materials. + +The Impact of Template. The choice of template has a crucial impact on our approach. If we directly use SMPLX as our geometric template, as shown in Tab. 3 (w SMPLX), leading to poor results with numerous artifacts, especially for loose skirts as seen in Fig. 6. After extending the geometry of SMPLX to a complete template with clothes, hair, and shoes, this quality is significantly improved as shown in Tab. 3 (w/o Mesh Non.), which allows our method to reconstruct more challenging clothes. + +The Impact of Non-rigid Deformation. Despite having a complete geometry, the rigid transformations of the skeleton are not able to properly represent non-rigid transformations such as skirt swinging and hair flying. We concurrently introduce mesh non-rigid deformation and Gaussian non-rigid deformation in our method. The geometry with- + +![](images/bd81a408de9da3e103ed8a1462d3aaccdb6b9831218adfc16104d323aa67ca29.jpg) +Figure 5. Novel pose and expression animation. TaoAvatar can be driven by the same skeleton and expression parameters vividly. + +out mesh non-rigid deformation is unable to attach to the surface of the character as shown in Fig. 6 (w/o Mesh Non.). The Gaussian non-rigid compensation also plays a crucial role in quantitative metrics. Without the blend shapes for Gaussian non-rigid compensation, the results are prone to produce artifacts due to imprecise geometry and variable appearance as shown in Fig. 6 (w/o Gau Non.). + +The Impact of Baking Teacher. Compared to directly training a small MLP-based network to handle non-rigid deformations as shown in Tab. 3 (w/o Teacher), our proposed multi-stage training strategy is more efficient. With direct supervision of the non-rigid deformation of the teacher network, it is easier for the student network to decouple the geometry deformation and appearance variations. + +# 5. Application + +TaoAvatar offers a lightweight, comprehensive representation for 3D talking body scenarios, easily deployable on various mobile devices using existing tools [25] as demonstrated in Fig. 1. It seamlessly integrates with AI models to enable real-time dialogue interactions. we deployed a 3D digital human agent on the Apple Vision Pro, which interacts with users through an ASR-LLM-TTS pipeline [16, 31, + +![](images/59d40bf71ace3044765c29d31cff170e4af8e88e200a969b6f952fc8fa6453a7.jpg) +Figure 6. Ablation Study. Red and Green boxes show artifacts and their improved counterparts. + +
ModelPSNR↑SSIM↑LPIPS↓P2S↓Chamfer↓
w SMPLX28.47.9476.07899.7690.9995
w/o Mesh Non.32.10.9734.03814.4877.5007
w/o Gau Non.31.16.9686.03932.2968.3068
w/o Teacher32.67.9751.03769.5236.5359
Full33.29.9772.03464.2953.3052
+ +Table 3. Ablation Study. We use the mesh reconstructed by NeuS2 [54] as a pseudo-truth for geometry evaluation. + +39, 62]. Facial expressions and gestures are dynamically controlled by an Audio2BS model [14], allowing the agent to respond naturally with synchronized speech, expressions, and movements. A live demonstration is available in our supplementary materials. + +# 6. Discussion + +Conclusion. In this work, we present TaoAvatar, a lightweight, lifelike full-body talking avatar solution. We demonstrate how the teacher-student framework captures high-definition facial and body details while ensuring real-time performance on AR devices. TaoAvatar can be driven by diverse signals, including facial expressions, hand gestures, and body poses. Through both quantitative and qualitative evaluations, we showcase the advantages of our approach. Additionally, we validate its practical potential with a real-world application on the Apple Vision Pro. + +Limitations and Future Works. TaoAvatar encounters challenges in modeling flexible clothing deformation under exaggerated body poses, which are out-of-distribution of training data. A possible solution is to integrate GNN simulators [17, 18] handling larger hemlines, which are compatible with our approach. + +Potential Social Impact. TaoAvatar can synthesize life-like talking digital humans within an augmented reality environment, generating fabricated 3D content or 2D videos. Therefore, responsible use of this technology is essential. + +# References + +[1] https://github.com/zju3dv/EasyMocap.5, 13 +[2] https://github.com/ShenhanQian/VHAP.5,13 +[3] Rinet Abdrashitov, Kim Raichstat, Jared Monsen, and David Hill. Robust skin weights transfer via weight inpainting. In SIGGRAPH Asia 2023 Technical Communications, SA Technical Communications 2023, Sydney, NSW, Australia, December 12-15, 2023, pages 25:1-25:4. ACM, 2023. 3, 13 +[4] Thiemo Alldieck, Marcus A. Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 8387-8397. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 2 +[5] Thiemo Alldieck, Marcus A. Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1175-1186. Computer Vision Foundation / IEEE, 2019. 2 +[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 5419-5429. IEEE, 2019. 1 +[7] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos, 2021. 2 +[8] Xu Chen, Yufeng Zheng, Michael J. Black, Otmar Hilliges, and Andreas Geiger. SNARF: differentiable forward skinning for animating non-rigid neural implicit shapes. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 11574-11584. IEEE, 2021. 2, 3 +[9] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023. 2 +[10] Yushuo Chen, Zerong Zheng, Zhe Li, Chao Xu, and Yebin Liu. Meshavatar: Learning high-quality triangular human avatars from multi-view videos. CoRR, abs/2407.08414, 2024. 2, 3, 5, 6, 7, 13, 14, 16 +[11] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 19925-19936. IEEE, 2023. 2 +[12] Edilson De Aguiar, Christian Theobalt, Carsten Stoll, and Hans-Peter Seidel. Marker-less deformable mesh tracking + +for human shape and motion capture. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2 +[13] Helisa Dhamo, Yinyu Nie, Arthur Moreau, Jifei Song, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Headgas: Real-time animatable head avatars via 3d gaussian splatting. In ECCV (2), pages 459-476. Springer, 2024. 3 +[14] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through A unified model. CoRR, abs/2408.00762, 2024. 7, 8, 14 +[15] Yao Feng, Weiyang Liu, Timo Bolkart, Jinlong Yang, Marc Pollefeys, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. CoRR, abs/2309.06441, 2023. 2 +[16] Zhifu Gao, Shiliang Zhang, Ian McLoughlin, and Zhijie Yan. Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to-end speech recognition. In 23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18-22, 2022, pages 2063-2067. ISCA, 2022. 8, 14 +[17] Artur Grigorev, Michael J. Black, and Otmar Hilliges. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16965-16974. IEEE, 2023. 3, 8 +[18] Artur Grigorev, Giorgio Becherini, Michael J. Black, Otmar Hilliges, and Bernhard Thomaszewski. Contourcraft: Learning to resolve intersections in neural multi-garment simulations. In ACM SIGGRAPH 2024 Conference Papers, SIGGRAPH 2024, Denver, CO, USA, 27 July 2024-1 August 2024, page 81. ACM, 2024. 3, 8 +[19] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics, 40(4), 2021. 2 +[20] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 634-644. IEEE, 2024. 2, 3, 5, 7, 14, 16 +[21] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20418-20431. IEEE, 2024. 2 +[22] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Trans. Graph., 42(4):160:1-160:12, 2023. 2, 5 +[23] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In IEEE/CVF Conference on Computer Vision and + +Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5595-5605. IEEE, 2022. 2 +[24] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16922-16932, 2023. 2 +[25] Xiaotang Jiang, Huan Wang, Yiliu Chen, Ziqi Wu, Lichuan Wang, Bin Zou, Yafeng Yang, Zongyang Cui, Yu Cai, Tianhang Yu, Chengfei Lyu, and Zhihua Wu. MNN: A universal and efficient inference engine. In Proceedings of the Third Conference on Machine Learning and Systems, MLSys 2020, Austin, TX, USA, March 2-4, 2020. mlsys.org, 2020. 8 +[26] Yujiang Jiang, Qingmin Liao, Xiaoyu Li, Li Ma, Qi Zhang, Chaopeng Zhang, Zongqing Lu, and Ying Shan. UV gaus-sians: Joint learning of mesh deformation and gaussian textures for human avatar modeling. CoRR, abs/2403.11589, 2024. 3 +[27] Yujiang Jiang, Qingmin Liao, Zhaolong Wang, Xiangru Lin, Zongqing Lu, Yuxi Zhao, Hanqing Wei, Jingrui Ye, Yu Zhang, and Zhijing Shao. Smplx-lite: A realistic and drivable avatar benchmark with rich geometry and texture annotations. In IEEE International Conference on Multimedia and Expo, ICME 2024, Niagara Falls, ON, Canada, July 15-19, 2024, pages 1-6. IEEE, 2024. 1, 2 +[28] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5322-5332. IEEE, 2024. 16 +[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 4 +[30] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: human gaussian splats. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 505-515. IEEE, 2024. 2 +[31] Jungil Kong, Jihoon Park, Beomjeong Kim, Jeongmin Kim, Dohee Kong, and Sangjin Kim. VITS2: improving quality and efficiency of single-stage text-to-speech with adversarial learning and architecture design. In 24th Annual Conference of the International Speech Communication Association, Interspeech 2023, Dublin, Ireland, August 20-24, 2023, pages 4374-4378. ISCA, 2023. 8, 14 +[32] Youngjoong Kwon, Baole Fang, Yixing Lu, Haoye Dong, Cheng Zhang, Francisco Vicente Carrasco, Albert Mosella-Montoro, Jianjin Xu, Shingo Takagi, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Generalizable human gaussians for sparse view synthesis. In ECCV (78), pages 451-468. Springer, 2024. 3 +[33] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. GART: gaussian articulated template models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19876-19887. IEEE, 2024. 2 + +[34] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19711-19722. IEEE, 2024. 2, 3, 5, 6, 7, 13, 14, 15, 16 +[35] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3d gaussian splatting for inverse rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21644-21653. IEEE, 2024. 16 +[36] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: neural free-view synthesis of human actors with pose control. ACM Trans. Graph., 40(6):219:1-219:16, 2021. 2, 3 +[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 1, 2, 3, 5 +[38] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 2 +[39] Chengfei Lv, Chaoyue Niu, Renjie Gu, Xiaotang Jiang, Zhaode Wang, Bin Liu, Ziqi Wu, Qiulin Yao, Congyu Huang, Panos Huang, Tao Huang, Hui Shu, Jinde Song, Bin Zou, Peng Lan, Guohuan Xu, Fei Wu, Shaojie Tang, Fan Wu, and Guihai Chen. Walle: An end-to-end, general-purpose, and large-scale production system for device-cloud collaborative machine learning. In 16th USENIX Symposium on Operating Systems Design and Implementation, OSDI 2022, Carlsbad, CA, USA, July 11-13, 2022, pages 249-265. USENIX Association, 2022. 8 +[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 13, 16 +[41] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In CVPR, pages 788-798. IEEE, 2024. 2 +[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2 +[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1, 2, 3, 5, 14 +[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 1 +[45] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gauss + +sianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20299-20309. IEEE, 2024. 3, 5 +[46] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5020-5030. IEEE, 2024. 2, 5, 7, 14, 16 +[47] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 130-141. IEEE, 2024. 3 +[48] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 1606-1616. IEEE, 2024. 2, 3 +[49] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16911-16921. IEEE, 2023. 3 +[50] Carsten Stoll, Juergen Gall, Edilson De Aguiar, Sebastian Thrun, and Christian Theobalt. Video-based reconstruction of animatable human characters. ACM Transactions on Graphics (TOG), 29(6):1-10, 2010. 2 +[51] Daniel Vlasic, Pieter Peers, Ilya Baran, Paul E. Debevec, Jovan Popovic, Szymon Rusinkiewicz, and Wojciech Matusik. Dynamic shape capture using multi-view photometric stereo. ACM Trans. Graph., 28(5):174, 2009. 2 +[52] Lizhen Wang, Xiaochen Zhao, Jingxiang Sun, Yuxiang Zhang, Hongwen Zhang, Tao Yu, and Yebin Liu. Stylea-vatar: Real-time photo-realistic portrait avatar from a single video. In ACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH 2023, Los Angeles, CA, USA, August 6-10, 2023, pages 67:1-67:10. ACM, 2023. 3, 6, 15 +[53] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 550-560. IEEE, 2024. 3, 13 +[54] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3272-3283. IEEE, 2023. 2, 3, 4, 8, 13, 15, 16 +[55] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 15 + +[56] Jing Wen, Xiaoming Zhao, Zhongzheng Ren, Alexander G. Schwing, and Shenlong Wang. Gomavatar: Efficient animatable human modeling from monocular video using gaussians-on-mesh. In CVPR, pages 2059-2069. IEEE, 2024. 2, 3 +[57] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16189-16199. IEEE, 2022. 2 +[58] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 2 +[59] Donglai Xiang, Timur M. Bagautdinov, Tuur Stuyck, Fabian Prada, Javier Romero, Weipeng Xu, Shunsuke Saito, Jingfan Guo, Breannan Smith, Takaaki Shiratori, Yaser Sheikh, Jessica K. Hodgins, and Chenglei Wu. Dressing avatars: Deep photorealistic appearance for physically simulated clothing. ACM Trans. Graph., 41(6):222:1-222:15, 2022. 3 +[60] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In CVPR, pages 1802-1812. IEEE, 2024. 3 +[61] Yuelang Xu, Bengwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 1931-1941. IEEE, 2024. 3 +[62] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. Qwen2 technical report. CoRR, abs/2407.10671, 2024. 8, 14 +[63] Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. In CVPR, pages 896-905. IEEE, 2024. 2 +[64] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595. Computer Vision Foundation / IEEE Computer Society, 2018. 4, 15 + +[65] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humanerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 1 +[66] Zhongyuan Zhao, Zhenyu Bao, Qing Li, Guoping Qiu, and Kanglin Liu. Psavatar: A point-based shape model for real-time head avatar animation with 3d gaussian splatting, 2024. 3 +[67] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CoRR, abs/2401.03407, 2024. 15 +[68] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas J. Guibas, and Gordon Wetzstein. Physavatar: Learning the physics of dressed 3d avatars from visual observations. CoRR, abs/2404.04421, 2024. 3 +[69] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM Trans. Graph., 42(4):158:1-158:19, 2023. 2 +[70] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4574-4584. IEEE, 2023. 3 \ No newline at end of file diff --git a/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/images.zip b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44720d03dc1b446ea2dd42b22394c762f6f9a368 --- /dev/null +++ b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73717092509171469a2574cb730fd38ccc082bd4adfd558f77032bd148e39b23 +size 894068 diff --git a/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/layout.json b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6b038460b73bce30ebba1f518027d289b1a14a60 --- /dev/null +++ b/2025/TaoAvatar_ Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting/layout.json @@ -0,0 +1,8370 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 102, + 496, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 102, + 496, + 140 + ], + "spans": [ + { + "bbox": [ + 116, + 102, + 496, + 140 + ], + "type": "text", + "content": "TaoAvatar: Real-Time Lifelike Full-Body Talking Avatars for Augmented Reality via 3D Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 161, + 468, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 161, + 468, + 203 + ], + "spans": [ + { + "bbox": [ + 141, + 161, + 468, + 203 + ], + "type": "text", + "content": "Jianchuan Chen,\\* Jingchuan Hu,\\* Gaige Wang,\\* Zhonghua Jiang \nTiansong Zhou, Zhiwen Chen\\*, Chengfei Lv\\* \nAlibaba Group, Hangzhou, China" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 203, + 416, + 342 + ], + "blocks": [ + { + "bbox": [ + 56, + 203, + 416, + 342 + ], + "lines": [ + { + "bbox": [ + 56, + 203, + 416, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 203, + 416, + 342 + ], + "type": "image", + "image_path": "e6de75c108a6d16181aa0ae30ea0d797e628fcf6f4302f98a0b4aaec16eb8a37.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 344, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 344, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 555, + 368 + ], + "type": "text", + "content": "Figure 1. TaoAvatar generates photorealistic, topology-consistent 3D full-body avatars from multi-view sequences. It provides high-quality, real-time rendering with low storage requirements, compatible across various mobile and AR devices like the Apple Vision Pro." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 418, + 204, + 552, + 342 + ], + "blocks": [ + { + "bbox": [ + 418, + 204, + 552, + 342 + ], + "lines": [ + { + "bbox": [ + 418, + 204, + 552, + 342 + ], + "spans": [ + { + "bbox": [ + 418, + 204, + 552, + 342 + ], + "type": "image", + "image_path": "33d932f309550bd7b5d915d8ec306113b28eb2dd54dd2b3da55006a9f95d3678.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 378, + 200, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 378, + 200, + 392 + ], + "spans": [ + { + "bbox": [ + 151, + 378, + 200, + 392 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 404, + 297, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 297, + 680 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 297, + 680 + ], + "type": "text", + "content": "Realistic 3D full-body talking avatars hold great potential in AR, with applications ranging from e-commerce live streaming to holographic communication. Despite advances in 3D Gaussian Splatting (3DGS) for lifelike avatar creation, existing methods struggle with fine-grained control of facial expressions and body movements in full-body talking tasks. Additionally, they often lack sufficient details and cannot run in real-time on mobile devices. We present TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar driven by various signals. Our approach starts by creating a personalized clothed human parametric template that binds Gaussians to represent appearances. We then pre-train a StyleUnet-based network to handle complex pose-dependent non-rigid deformation, which can capture high-frequency appearance details but is too resource-intensive for mobile devices. To overcome this, we \"bake\" the non-rigid deformations into a lightweight MLP-based network using a distillation technique and develop blend shapes to compensate for details. Extensive experiments show that TaoAvatar achieves state-of-the-art rendering quality while running in real-time across various devices, maintaining 90 FPS on high-definition stereo devices such as the Apple Vision Pro." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 378, + 394, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 378, + 394, + 390 + ], + "spans": [ + { + "bbox": [ + 315, + 378, + 394, + 390 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 400, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 520 + ], + "type": "text", + "content": "Creating lifelike 3D human avatars is a dynamic and rapidly advancing research area in computer vision and graphics, essential for applications in AR/VR, 3D entertainment, and virtual communication. Despite decades of research, achieving realistic and immersive avatars on mobile devices—particularly AR—remains challenging due to computational limitations. Current industry solutions, such as MetaHuman, often rely on high-precision scans and extensive manual effort by artists for 3D modeling, rigging, and motion capture." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "content": "In recent years, academic researchers have adopted parametric models [37, 43] for human representation, which are constructed from massive body scans and can capture a performer's shape, expression, and pose from images or videos. These methods [4-6, 27] extend the naked model [37, 43] by adding offsets to represent tailored clothing, but are still struggling with complex geometries and high-frequency details, such as loose skirts and fine hair, due to limitations of topology and texture resolution. With the rise of neural radiance fields (NeRF) [40], many researchers tend to define implicit neural human representations in canonical space, animated by backward linear blend skinning [44, 65] of these parametric models [37, 43]. While these methods can handle arbitrary topology and deliver higher rendering quality, they are highly time-consuming due to the large amount of sampling required" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 715 + ], + "type": "text", + "content": "*Equal contribution. ‡Project Leader and †Corresponding author. Project: https://PixelAI-Team.github.io/TaoAvatar" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "10723" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "during volume rendering. Although many efforts [9, 24] have significantly improved rendering efficiency by introducing iNGP [42], the animated results still fail to escape the Uncanny Valley effect when driven. Recently, 3D Gaussian Splatting (3DGS) [29], an explicit and efficient point-based representation, has gained tremendous attention from researchers for its ability to deliver both high-quality rendering and real-time performance. Compared to implicit representations [40, 54], the explicit point-based representation is more compatible with being integrated with the mesh-based parametric models [37, 43], which can be directly driven by forward linear blend skinning. Many recent methods [20, 34, 46, 48], combining 3DGS and the parametric models [37, 43], have made significant strides toward realizing lifelike 3D human avatars. However, achieving real-time rendering on mobile devices remains challenging, especially for AR glasses, which require high resolution, stereo rendering, and seamless interaction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 291, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 295, + 579 + ], + "type": "text", + "content": "We introduce TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar designed for running in real-time on Augmented Reality (AR) devices. Unlike previous approaches [8, 10, 34], which develop implicit parameterized models from scratch, we construct a personalized, clothed parameterized template that preserves the facial expressions and gesture controls inherent to the native SMPLX [43]. At the same time, we bind the Gaussians to the triangles as texture to create a hybrid avatar representation. Inspired by [34], we pre-train a large StyleUnet-based teacher network to learn pose-dependent dynamic deformation maps of Gaussians in 2D space by front and back orthogonal projection. Although the teacher network is powerful enough to capture high-frequency appearance details in various poses, its large number of parameters makes it difficult to run in real-time on AR devices. To achieve high-performance rendering, we bake the dynamic deformation of Gaussians into the non-rigid deformation field of the mesh using a distillation technique, employing a compact MLP-based student network. Meanwhile, we propose two lightweight, learnable blend shapes to compensate for the non-rigid deformation of the Gaussians. After fin-tuning, our student model achieves high-performance rendering without compromising quality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": "To evaluate our approach, we propose a multi-view dataset named TalkBody4D, focusing on prevalent full-body talking scenarios encountered in daily life, which includes rich facial expressions and gestures with synchronous audios different from existing full-body motion datasets [11, 22, 69]. Experimental results demonstrate that our method effectively captures high-frequency appearance details of human avatars while maintaining a lightweight architecture capable of rendering stereo images at 2K resolution on common mobile and AR devices, including the Apple Vision Pro. The resulting full-body avatar is highly" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "expressive, enabling users to animate it with facial expressions, hand gestures, and body poses, thereby highlighting the potential applications of our method. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 121, + 553, + 312 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "content": "- We introduce TaoAvatar, a novel teacher-student framework which yields topologically consistent, well-aligned geometry, further creating a high-fidelity, lightweight, 3DGS-based drivable full-body talking avatar." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "content": "- We propose two innovative strategies including non-rigid deformation baking and two lightweight blend shapes compensations to ensure high-quality rendering and efficient performance on mobile and AR devices, achieving 2K resolution rendering at 90 FPS on high-definition stereo devices like the Apple Vision Pro." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 239, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 312 + ], + "type": "text", + "content": "- We contribute TalkBody4D, a multi-view dataset to be released, designed for full-body talking scenarios, featuring diverse facial expressions and gestures with synchronous audios. Extensive experiments demonstrate that our approach outperforms other state-of-the-art methods in both quality and performance." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 323, + 399, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 399, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 399, + 335 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 342, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 665 + ], + "type": "text", + "content": "3D Avatar Representation. Traditional computer graphics techniques [12, 19, 50, 51] use 3D scanning or multiview stereo (MVS) to create full-body meshes, which are then combined with a 3D skeleton to produce drivable templates. However, this approach is both expensive and inefficient. As parametric body models, SMPL [37] and SMPLX [43] are widely used for their versatility but lack detail in simulating loose clothing. Some works [4, 5, 27] have attempted to add clothing offsets to SMPLX, effectively handling tight clothing but struggling with loose garments. Due to the limited expressiveness of traditional meshes, researchers have combined SMPL with advanced rendering techniques. Some methods [7, 15, 23, 36, 57] use NeRF [40] to implicitly express human bodies, mapping points from the observation space backward to canonical space. Although high-quality, volume rendering is slow and poses challenges for real-time applications. In contrast, 3DGS [29] uses 3D Gaussian distributions for scenes, offering real-time rendering [38, 58], fast training, and the ability to handle complex materials such as hair and translucency. Some methods [34, 46, 48, 56, 63] combining SMPLX [43] and 3DGS [29], like GaussianAvatar [20], use explicit GS point clouds and forward skinning to efficiently simulate the deformations, offering good rendering quality and speed. This representation presents potential solutions for achieving both real-time performance and computational efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "Dynamic Nonrigid Deformation Modeling. Various methods have been developed to simulate non-rigid human deformation realistically. The MLP-based methods [21, 30, 33, 41], like 3DGS-Avatar [46], are noted for lightweight" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10724" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "flexibility but struggle with handling novel poses. Physical simulation methods [59, 68] produce physics-based animations but require unified topology, and complex modeling, and are computationally expensive, which negatively impacts real-time performance. Additionally, some methods [17, 18] use GNNs to learn non-rigid deformations in a data-driven manner. These methods are faster than traditional cloth simulations. Recently, dynamic texture methods are emerging, these methods [26, 32, 36] model complex details in the 2D texture space, while AnimatableGS [34] and MeshAvatar [10] enhance modeling with front and back projection strategies which avoid texture unwarping." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 396 + ], + "type": "text", + "content": "3D Drivable Talking Avatars. Research on drivable digital humans is rapidly advancing for various applications. In 3D Talking Heads, methods [13, 45, 47, 60, 61, 66] excel in capturing and driving facial expressions, enhancing the realism of virtual characters. For drivable bodies, these methods [8, 20, 34, 56, 70] achieve precise restoration of body movements by learning pose-dependent deformation fields. Furthermore, some methods [49, 59] not only support detailed body modeling and animation but also enable head expressions and lip-syncing. Despite enhancing realism, running these methods in real-time on mobile or AR devices is challenging due to their high computational demands, exceeding current edge platform capabilities. Optimizing computational efficiency without compromising high-quality output is a key research focus." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 405, + 111, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 111, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 111, + 417 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "type": "text", + "content": "Leveraging multi-view RGB videos of a performer alongside corresponding SMPLX registrations for each frame, we aim to develop a high-fidelity, lightweight full-body talking avatar capable of high-resolution and high-performance rendering on mobile devices. To achieve this, we first create a hybrid clothed parametric representation by integrating 3D Gaussian Splatting (3DGS) [29] and SMPLX [43], which is more expressive for simulating loose clothing and hair, while maintaining high-quality rendering (Sec. 3.1). Next, we propose a teacher-student framework to reconstruct high-frequency, pose-dependent dynamic details as illustrated in Fig. 2. By leveraging non-rigid deformation baking and incorporating two lightweight blend shape compensations, our method achieves a tradeoff between high quality and high performance (Sec. 3.2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 612, + 286, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 286, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 286, + 625 + ], + "type": "text", + "content": "3.1. Hybrid Clothed Parametric Representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "The Clothed Extension of SMPLX. To extend SMPLX to complete clothed human geometry with clothes, hair, etc., we choose a frame close to T-pose as a reference, which provides more visible details and less sticky geometry. First, we utilize NeuS2 [54] to reconstruct the geometry of this reference frame and parse out non-body components (e.g., clothes) from it using [53]. However, the body's skeleton" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "cannot directly drive these components. We estimate the shape and pose of the SMPLX [37] in the current frame and use robust skinning transfer [3] to propagate the body's skinning weights to these non-body components. Finally, we apply inverse skinning to transform these components back to the standard T-pose, resulting in a comprehensive personalized parametric model, referred to as SMPLX++." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": "Binding Gaussians on Mesh as Texture. After obtaining the complete geometric model, we bind 3D Gaussians to the mesh's triangles as textures, enabling them to move synchronously with the mesh. Inspired by these hybrid representations [45, 48], we define the attributes of the Gaussians in a local coordinate system relative to the triangles. Specifically, for each triangle, we randomly initialize " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " Gaussians, each of which maintains the local attributes " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\{f,(u,v),\\gamma ,\\mathbf{r},\\mathbf{s},o,\\mathbf{sh}\\}" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " is the index of the parent triangle, " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " is the barycentric coordinates, " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " represents the translation along the triangle's normal direction, " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " denote rotation and scaling in the local space, " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " indicates opacity, and " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{sh}" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " consists of spherical harmonic coefficients. Additionally, we define the normal of each Gaussian as " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_g = [1,0,0]" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " within the local space, ensuring consistency with the mesh's normal for subsequent transformations. To deform each Gaussian from the local to the world space, we construct a transformation based on its parent triangle " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "F_{f}(\\mathbf{v}_{1},\\mathbf{v}_{2},\\mathbf{v}_{3})" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_1" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_2" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_3" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " are the vertices of the triangle " + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "inline_equation", + "content": "F_{f}" + }, + { + "bbox": [ + 313, + 156, + 555, + 396 + ], + "type": "text", + "content": " in the world space:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 332, + 407, + 492, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 407, + 492, + 418 + ], + "spans": [ + { + "bbox": [ + 332, + 407, + 492, + 418 + ], + "type": "interline_equation", + "content": "\\mathbf {p} = u \\cdot \\mathbf {v} _ {1} + v \\cdot \\mathbf {v} _ {2} + (1 - u - v) \\cdot \\mathbf {v} _ {3},", + "image_path": "6835e2938e101c968eb46399e44098cb0eec8d57c03581b550b30f6731fd4cc8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 421, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 421, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 329, + 421, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\mathbf {R} = [ \\mathbf {n}, \\mathbf {q}, \\mathbf {n} \\times \\mathbf {q} ], \\mathbf {q} = \\frac {\\left(\\mathbf {v} _ {2} + \\mathbf {v} _ {3}\\right) / 2 - \\mathbf {v} _ {1}}{\\left\\| \\left(\\mathbf {v} _ {2} + \\mathbf {v} _ {3}\\right) / 2 - \\mathbf {v} _ {1} \\right\\|}, \\tag {1}", + "image_path": "94be4e42ed12138f6ca2772b3729910fb9da7e8acf5fb8bdff994778748f5afa.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 333, + 449, + 529, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 449, + 529, + 462 + ], + "spans": [ + { + "bbox": [ + 333, + 449, + 529, + 462 + ], + "type": "interline_equation", + "content": "e = \\left(\\left\\| \\mathbf {v} _ {1} - \\mathbf {v} _ {2} \\right\\| + \\left\\| \\mathbf {v} _ {2} - \\mathbf {v} _ {3} \\right\\| + \\left\\| \\mathbf {v} _ {1} - \\mathbf {v} _ {3} \\right\\|\\right) / 3,", + "image_path": "49d24a6a86812c59ce80b8fd8198fb2695c8797861cd0268e8b42d07b35ebc8b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 553, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 472, + 553, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{n} = \\frac{(\\mathbf{v}_1 - \\mathbf{v}_2)\\times(\\mathbf{v}_3 - \\mathbf{v}_1)}{\\|\\mathbf{v}_1 - \\mathbf{v}_2\\|\\cdot\\|\\mathbf{v}_3 - \\mathbf{v}_1\\|}" + }, + { + "bbox": [ + 313, + 472, + 553, + 512 + ], + "type": "text", + "content": " is the normal of the triangle, " + }, + { + "bbox": [ + 313, + 472, + 553, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 472, + 553, + 512 + ], + "type": "text", + "content": " denotes the surface point on the parent triangle, and " + }, + { + "bbox": [ + 313, + 472, + 553, + 512 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 472, + 553, + 512 + ], + "type": "text", + "content": " denotes the average edge length." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 344, + 521, + 553, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 521, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 344, + 521, + 553, + 534 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {w} = \\mathbf {p} + \\gamma \\mathbf {R n}, \\quad \\mathbf {r} _ {w} = \\mathbf {R r}, \\quad \\mathbf {s} _ {w} = e \\cdot \\mathbf {s}, \\tag {2}", + "image_path": "5ac2b513725ab94293a2687c9e6510075247d0b582e84479d6a34d64d4606f27.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_w" + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_w" + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_w" + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "text", + "content": " represent the position, rotation, and scaling in the world spaces after transformation, respectively. Furthermore, " + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_w = SH(\\mathbf{sh},\\mathbf{R}^{-1}\\mathbf{d})" + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "text", + "content": " denotes the color along the view direction " + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "text", + "content": " in the world space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 600, + 553, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 553, + 612 + ], + "type": "text", + "content": "3.2. Dynamic Mesh-based Gaussian Reconstruction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "Learning Dynamic Non-Rigid Gaussian Deformation Maps. Although SMPLX++ can be directly driven by the skeleton, linear blend skinning is insufficient for performing dynamic non-rigid deformation such as clothing folds and swaying motions. Like dynamic Gaussian avatars [26, 34], we utilize a large StyleUnet [52] as the teacher network to capture complex pose-dependent dynamic non-rigid deformation of Gaussians in 2D texture space. Following [34]," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10725" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 542, + 320 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 542, + 320 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 542, + 320 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 542, + 320 + ], + "type": "image", + "image_path": "ffc1c21440c2523504d70f66488fb12509676fa526cc770024f9f855c5f36072.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 325, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 54, + 325, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 325, + 555, + 370 + ], + "type": "text", + "content": "Figure 2. Illustration of our Method. Our pipeline begins by reconstructing (a) a cloth-extended SMPLX mesh with aligned Gaussian textures. To address complex dynamic non-rigid deformations, (b) we employ a teacher StyleUnet to learn pose-dependent non-rigid maps, which are then baked into a lightweight student MLP to infer non-rigid deformations of our template mesh. For high-fidelity rendering, (c) we introduce learnable Gaussian blend shapes to enhance appearance details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "text", + "content": "we rasterize the T-posed mesh, which is colored using posed coordinates and segmentation colors, to obtain the front and back position maps " + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_f,\\mathcal{P}_b\\}" + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "text", + "content": ". These position maps are then fed into the StyleUnet network along with view directions to generate the non-rigid deformation maps " + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\{\\Delta \\mathcal{U}_f,\\Delta \\mathcal{U}_b\\}" + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "text", + "content": " and other residual Gaussian attribute maps. We add these residuals to the Gaussian's local attributes and transform them into the world space according to Eq. (2). To train the StyleUnet-based teacher network " + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "text", + "content": ", we adopt the losses including L1, D-SSIM [29], and perceptual loss [64]. Additionally, we introduce a normal loss " + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nor}" + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "text", + "content": ". Consequently, the total reconstruction loss " + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 54, + 390, + 296, + 535 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 545, + 295, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 545, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 60, + 545, + 295, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} = \\mathcal {L} _ {1} + \\lambda_ {\\text {s s i m}} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {\\text {l p i p s}} \\mathcal {L} _ {\\text {l p i p s}} + \\lambda_ {\\text {n o r}} \\mathcal {L} _ {\\text {n o r}}, \\tag {3}", + "image_path": "a28fc9575418cb86b6657cd9a99c47f43afc9e23573c8b0fbe806b9c9fc4b0c8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{*}" + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "text", + "content": " is the loss weights. The normal loss is defined as " + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nor} = \\| N_t - N_g\\|" + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "text", + "content": " is the rendered normal image by the teacher network, and the ground truth normal image " + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "text", + "content": " can be obtained from the method [54]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Baking Non-Rigid Gaussian Deformation into Mesh Deformation. Although the StyleUnet-based network " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " can achieve good results, it struggles to run in real-time on mobile devices due to its large number of parameters, especially for AR devices, which require high resolution, stereo rendering, and high performance. Inspired by knowledge distillation, we bake the Gaussian non-rigid deformation of the teacher network into the mesh non-rigid deformation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 390, + 544, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 544, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 544, + 403 + ], + "type": "text", + "content": "field, which is a compact MLP-based student network " + }, + { + "bbox": [ + 313, + 390, + 544, + 403 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 390, + 544, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 391, + 409, + 553, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 409, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 391, + 409, + 553, + 422 + ], + "type": "interline_equation", + "content": "\\Delta \\bar {\\mathbf {v}} _ {i} = \\mathcal {S} \\left(\\bar {\\mathbf {v}} _ {i}, \\theta , \\mathbf {z} _ {t}\\right), \\tag {4}", + "image_path": "9f45c95286b0a4ccc2a9f0699cf2a44effe1457805919957229c888e09bfe29b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{v}}_i" + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": "-th vertex coordinate in the canonical space, " + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": " is the pose parameter, and " + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": " is a learnable embedding for each frame to compensate for inaccurate pose estimation. To train the student network " + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": " with the capability of pose-dependent non-rigid deformation, we directly supervise its output using the Gaussian non-rigid deformation maps " + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\{\\Delta \\mathcal{U}_f,\\Delta \\mathcal{U}_b\\}" + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": " of the teacher network " + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": ". Specifically, we render the mesh non-rigid deformations " + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\{\\Delta \\mathbf{v}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": " to front and back deformation maps " + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\{\\Delta \\mathcal{V}_f,\\Delta \\mathcal{V}_b\\}" + }, + { + "bbox": [ + 313, + 428, + 555, + 549 + ], + "type": "text", + "content": " in the canonical space, and then calculate the differences:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 555, + 553, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 555, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 347, + 555, + 553, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n o n} = \\left\\| \\Delta \\mathcal {V} _ {f} - \\Delta \\mathcal {U} _ {f} \\right\\| + \\left\\| \\Delta \\mathcal {V} _ {b} - \\Delta \\mathcal {U} _ {b} \\right\\|. \\tag {5}", + "image_path": "7c281d8047b206945c66a73fe72b229e63d9b0bc897a619ba6e75d831fcbf72d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "content": "Furthermore, to mitigate the intersection between clothing and the body, we introduce a semantic loss " + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sem}" + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "content": ". For each vertex, we assign a semantic label " + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i = \\mathbf{c}_i + \\sin (\\tau \\bar{\\mathbf{v}}_i)" + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "content": " that integrates the candidate coordinates " + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{v}}_i" + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "content": " and the artificial segmentation color " + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "content": " is a scale factor. The semantic label for each Gaussian can be obtained by interpolating the semantic label of the three vertices of its parent triangle. Subsequently, we render the Gaussian semantic map " + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "content": " and the mesh semantic map " + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_s" + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "content": " in the world space separately and compute the semantic loss " + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sem}" + }, + { + "bbox": [ + 313, + 574, + 555, + 694 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 392, + 700, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 392, + 700, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e m} = \\left\\| \\mathcal {E} _ {s} - \\mathcal {E} _ {t} \\right\\|. \\tag {6}", + "image_path": "e0f1cf6c57c812ea6a3f825e833372a075c26c663c0504186c11ab898cadd1ed.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10726" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "In addition, we refine the local attributes by computing the reconstruction loss " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": ", which takes the rendered image of the teacher network as ground truth. The total loss " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bak}" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": " during baking is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 129, + 295, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 129, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 93, + 129, + 295, + 142 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b a k} = \\mathcal {L} _ {r e c} + \\lambda_ {n o n} \\mathcal {L} _ {n o n} + \\lambda_ {s e m} \\mathcal {L} _ {s e m}. \\tag {7}", + "image_path": "7b60f7ecc030629289c2167953caf268fd51b4347cdca2cc339e1a645f864a95.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "content": "Compensating Gaussian Deformation with Lightweight Blend Shapes. Inspired by pose corrective blend shapes in SMPL [37], we build learnable blend shapes for position and color for each Gaussian. The position blend shapes, " + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{3 \\times n}" + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "content": ", primarily compensate for Gaussian positional adjustments (e.g., hair fluttering, clothing folds) across different poses. Meanwhile, the color blend shapes, " + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{3 \\times n}" + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "content": ", address appearance changes, such as shadow variations caused by self-obscuration. To obtain the corresponding driving coefficients, we employ a mapping network " + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "content": " for the head using the expression parameters " + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "content": " as input and a mapping network " + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "content": " for the body which inputs the body pose parameters " + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "content": ", to get the coefficients " + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_h \\in \\mathbb{R}^{n_h}" + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "content": " for the head and " + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_b \\in \\mathbb{R}^{n_b}" + }, + { + "bbox": [ + 55, + 149, + 296, + 353 + ], + "type": "text", + "content": " for the body, respectively, enabling independent and fine-grained control over head and body deformations. The compensation of position and color for each Gaussian can be expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 361, + 226, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 361, + 226, + 374 + ], + "spans": [ + { + "bbox": [ + 124, + 361, + 226, + 374 + ], + "type": "interline_equation", + "content": "\\delta \\mathbf {u} = B S \\left(\\mathbf {z} _ {h} \\oplus \\mathbf {z} _ {b}; \\mathbf {U}\\right),", + "image_path": "7403595f149d2f55bb7598abc5976fe250dda56c4999d8e0b6aa2eb5e71664bd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 375, + 294, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 375, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 127, + 375, + 294, + 388 + ], + "type": "interline_equation", + "content": "\\delta \\mathbf {c} = B S \\left(\\mathbf {z} _ {h} \\oplus \\mathbf {z} _ {b}; \\mathbf {C}\\right), \\tag {8}", + "image_path": "6a9ef37d83d1f12aa927f1793a6b504eb08f238369082e2b3d69fff86e4daef8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 396, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 396, + 296, + 443 + ], + "type": "text", + "content": " is a concatenation operation, resulting in the size " + }, + { + "bbox": [ + 55, + 396, + 296, + 443 + ], + "type": "inline_equation", + "content": "n = n_h + n_b" + }, + { + "bbox": [ + 55, + 396, + 296, + 443 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 55, + 396, + 296, + 443 + ], + "type": "inline_equation", + "content": "BS(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 396, + 296, + 443 + ], + "type": "text", + "content": " represents the blend shapes operation mentioned in SMPL [37]. We add these residuals to local attributes:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 450, + 228, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 450, + 228, + 464 + ], + "spans": [ + { + "bbox": [ + 122, + 450, + 228, + 464 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {w} = \\mathbf {p} + \\mathbf {R} (\\gamma \\mathbf {n} + \\delta \\mathbf {u}),", + "image_path": "e922d1c76c9f30ebfbdc4afd0b843586541e962762c4a267d216628bc8221934.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 124, + 462, + 294, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 462, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 124, + 462, + 294, + 479 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {w} = S H (\\mathbf {R} ^ {- 1} \\mathbf {d}) + \\delta \\mathbf {c}. \\tag {9}", + "image_path": "6ee3fc6fc2713c1da9415436f7ac5752ae64575a81ccce884450ddae8a97d088.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "text", + "content": "For the fine-tuning stage on training data, we freeze the non-rigid deformation field network " + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "text", + "content": ", optimize the two mapping networks " + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "text", + "content": ", and two blend shapes " + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 487, + 295, + 545 + ], + "type": "text", + "content": ". The loss during fine-tuning is the same as that pre-training the teacher network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "content": "After baking mesh non-rigid deformation and applying Gaussian non-rigid compensation, the performance of our student model is significantly enhanced compared to the teacher model, while still maintaining high quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 137, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 137, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 137, + 617 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 624, + 182, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 182, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 182, + 637 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Datasets. To evaluate the performance on the full-body talking task, we introduce a new dataset, TalkBody4D, designed for common full-body talking scenarios in everyday life, primarily characterized by rich mouth movements and diverse gestures with synchronous audios. The dataset comprises 4 distinct identities, each wearing 2 different outfits." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "Each talking sequence consists of rough 6k frames with 4K image resolution, 60 views with 48 full-body, and 12 face-region views. To evaluate performance on complex motions and expressions, we select 4 sequences (04, 05, 06, 08) from the ActorHQ [22] dataset, supplement with 2 dancing and 2 exaggerated expression sequences. The resolution of our training and testing images is " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "1500 \\times 2000" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " for all experiments. SMPLX registrations for these sequences can be obtained by first initializing SMPLX parameters using existing tools [1, 2, 45], followed by fine-tuning the poses with a photometric loss to improve geometry alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": "Implementation Details. The clothed parametric template has roughly " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "22k" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " vertices and " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "45k" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " faces in total, which contains additional " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "23k" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " faces for clothes, hair, and shoes, compared to the naive SMPLX. We randomly initialize about " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "220k" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " Gaussians in total to bind with the template as texture, for each triangle containing roughly 4 to 6, and optimize these Gaussians for " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "10k" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " iterations using the multi-view images of the reference frame. We pre-train our StyleUnet-based[34] teacher network for " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "600k" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " iterations, and set the loss weights " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\lambda_{ssim} = 0.2" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\lambda_{lpips} = 0.01" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\lambda_{nor} = 0.02" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " and batch size is 1 during the training process. Then we bake the teacher network into a small 5-layer MLP-based student network, which set " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\lambda_{non} = 0.1" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\lambda_{sem} = 1.0" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " and optimize " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "30k" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " iterations. Finally, we finetune the mapping networks and blend shapes for Gaussian non-rigid compensation on training data " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " iterations at the batch size of 4. The size of two blend shapes for position and color is " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "n = 28" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "n_h = 8" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " for the head and " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "n_b = 20" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " for the body." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 431, + 394, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 394, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 394, + 443 + ], + "type": "text", + "content": "4.2. Comparison" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 449, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 556 + ], + "type": "text", + "content": "We conduct an extensive comparison between our method and recent state-of-the-art full-body avatar methods include GaussianAvatar [20], 3DGS-Avatar [46], MeshAvatar [10], and AnimatableGS [34]. All of these methods learn pose-dependent non-rigid deformations in the canonical space and then utilize the skeleton-based rigid transformations of SMPL [37] or SMPLX [43] into observation space to model the dynamic human. We uniformly use SMPLX [43] across all these methods to ensure a fair comparison." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Comparison on Full-body Talking. We conduct experiments on typical full-body talking scenarios encountered in daily life, which include a variety of mouth movements and gestures. We provide the quantitative comparison on full-body talking as shown in Tab. 1. We also show the qualitative comparison in Fig. 3, and our method can generate more realistic and detailed results, especially in the face region. 3DGS-Avatar [46] directly learns two MLP-based networks to handle the pose-dependent non-rigid deformation of Gaussians but produces very blurry results due to the low-frequency bias of MLPs. GaussianAvatar [20] uses a 2D CNN network by encoding the position map in SMPLX's UV space to generate the non-rigid deformation" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10727" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 553, + 506 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 506 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 506 + ], + "type": "image", + "image_path": "0bf0305810caf19bbfc67cf176a043b3929f8cbd206ce269bb72cd330093f867.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 506, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 506, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 555, + 529 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons on full-body talking tasks. Our method outperforms state-of-the-art methods by producing clearer clothing dynamics and enhanced facial details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 540, + 297, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 297, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 297, + 708 + ], + "type": "text", + "content": "maps of Gaussians. However, it is restricted by the mini-clothed topology of SMPLX, struggling to handle loose skirts. MeshAvatar [10] chooses mesh as a human representation, which is insufficient to express the geometry of details, such as hair, and glasses. AnimatableGS [34] can generate high-quality dynamic appearance and complex non-rigid deformation. However, it is restricted by the implicit template of learning from scratch, which makes it difficult to handle fine-grained expressions, like blinking and flashing teeth as shown in the face region in Fig. 3. Therefore we adopt the SMPLX++ model as the template whose face and hands retain the native SMPLX prior. However, due to the amount parameters of the teacher StyleUnet [52], it achieves poor performance and is hard to run on a mobile" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": "device in real time. Through baking non-rigid deformation, we can achieve " + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "inline_equation", + "content": "150+" + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": " FPS at the resolution of " + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "inline_equation", + "content": "1500 \\times 2000" + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": " on Nvidia RTX4090 using only a small MLP-based mesh non-rigid deformation network and two lightweight learnable blend shapes as shown in Tab. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Comparison on Complex Motion and Expression. Besides full-body talking scenes, our method is also capable of reconstructing complex motions as well as exaggerated expressions as shown in Fig. 4. We provide quantitative comparisons in Tab. 2, our student network achieves competitive results compared to the teacher network, outperforming other methods. Especially for exaggerated expressions, our method significantly outperforms others in reconstructing the face region as shown in Tab. 2. Moreover, our method" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10728" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 553, + 173 + ], + "type": "table", + "html": "
Novel ViewNovel Gestures and ExpressionSpeed
ModelPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓FPS
GaussianAvatar[20]26.58 (23.57).9313 (.8159).10577 (.25242)25.99 (23.15).9232 (.8092).12265 (.26207)54
3DGS-Avatar[46]28.91 (23.95).9411 (.8303).07984 (.20450)26.46 (23.32).9157 (.8184).11804 (.21632)55
MeshAvatar[10]28.53 (24.55).9360 (.8083).09470 (.25572)27.08 (23.58).9229 (.7965).10783 (.24947)22
AnimatableGS[34]32.50 (26.42).9599 (.8587).06695 (.19535)28.05 (23.68).9328 (.8142).09191 (.22673)16
Ours (Teacher)33.45 (27.01).9649 (.8741).04986 (.15613)28.28 (24.28).9336 (.8291).07385 (.18176)16
Ours (Student)33.81 (27.80).9689 (.8975).06437 (.14218)28.38 (24.99).9389 (.8525).08874 (.13364)156
", + "image_path": "751fc55e47ac640ac687ddffdbd70656dea9a876d1c330fd3069607caa0fcce4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 208, + 526, + 417 + ], + "blocks": [ + { + "bbox": [ + 59, + 208, + 526, + 417 + ], + "lines": [ + { + "bbox": [ + 59, + 208, + 526, + 417 + ], + "spans": [ + { + "bbox": [ + 59, + 208, + 526, + 417 + ], + "type": "image", + "image_path": "306b4d44d5fb81c6683afe0b858d0b4ff8dd3d37c0a63d59c10d0321ecd66943.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 419, + 555, + 432 + ], + "lines": [ + { + "bbox": [ + 55, + 419, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 555, + 432 + ], + "type": "text", + "content": "Figure 4. Results in challenging scenarios. Our method can obtain high-quality reconstruction even for challenging pose and expressions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 449, + 300, + 527 + ], + "blocks": [ + { + "bbox": [ + 55, + 181, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 181, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 555, + 205 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on full-body talking task. The results inside the parentheses are evaluated for the face area, and the inference speed is evaluated on Nvidia RTX4090 when rendering images at a resolution of " + }, + { + "bbox": [ + 55, + 181, + 555, + 205 + ], + "type": "inline_equation", + "content": "{1500} \\times {2000}" + }, + { + "bbox": [ + 55, + 181, + 555, + 205 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 449, + 300, + 527 + ], + "lines": [ + { + "bbox": [ + 58, + 449, + 300, + 527 + ], + "spans": [ + { + "bbox": [ + 58, + 449, + 300, + 527 + ], + "type": "table", + "html": "
ModelPSNR↑SSIM↑LPIPS↓
GaussianAvatar[20]25.94 (24.33).9294 (.8251).10478 (.24179)
3DGS-Avatar[46]30.04 (25.08).9403 (.8458).08471 (.18044)
MeshAvatar[10]28.51 (24.94).9334 (.8100).08846 (.23517)
AnimatableGS[34]31.81 (26.79).9493 (.8608).07586 (.19521)
Ours (Teacher)32.80 (27.40).9533 (.8768).05581 (.14996)
Ours (Student)32.72 (27.35).9579 (.8836).07326 (.13914)
", + "image_path": "9d88b34f9692057bf8110f22bdce96cf644218a7ea7666b6535f62c2b582d977.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 536, + 295, + 570 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 570 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons about complex motions and expressions reconstruction. The results inside the parentheses are evaluated for face regions with exaggerated expressions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "text", + "content": "can obtain high-quality normals as shown in Fig. 4, which can be used for image-based relighting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "Comparison on Novel Expression and Novel Gesture. Benefiting from our parametric human model SMPLX++, our full-body avatars can be expression-driven and gesture-driven. We provide quantitative comparisons of self-driven on full-body talking scenes as shown in Tab. 1. As shown in Fig. 5, we provide visualizations of different characters driven by the same skeleton. In addition, our model can accept expression parameter inputs generated from audio by UniTalker [14] as shown in Fig. 5. More demos are pro" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 451, + 466, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 451, + 466, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 451, + 466, + 464 + ], + "type": "text", + "content": "vided in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 475, + 408, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 408, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 408, + 489 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 531 + ], + "type": "text", + "content": "In this subsection, we discuss the key contributions of our method. Additional quantitative results and experiments are available in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": "The Impact of Template. The choice of template has a crucial impact on our approach. If we directly use SMPLX as our geometric template, as shown in Tab. 3 (w SMPLX), leading to poor results with numerous artifacts, especially for loose skirts as seen in Fig. 6. After extending the geometry of SMPLX to a complete template with clothes, hair, and shoes, this quality is significantly improved as shown in Tab. 3 (w/o Mesh Non.), which allows our method to reconstruct more challenging clothes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "The Impact of Non-rigid Deformation. Despite having a complete geometry, the rigid transformations of the skeleton are not able to properly represent non-rigid transformations such as skirt swinging and hair flying. We concurrently introduce mesh non-rigid deformation and Gaussian non-rigid deformation in our method. The geometry with-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10729" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 388 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 388 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 388 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 388 + ], + "type": "image", + "image_path": "bd81a408de9da3e103ed8a1462d3aaccdb6b9831218adfc16104d323aa67ca29.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 392, + 295, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 415 + ], + "type": "text", + "content": "Figure 5. Novel pose and expression animation. TaoAvatar can be driven by the same skeleton and expression parameters vividly." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 427, + 295, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 511 + ], + "type": "text", + "content": "out mesh non-rigid deformation is unable to attach to the surface of the character as shown in Fig. 6 (w/o Mesh Non.). The Gaussian non-rigid compensation also plays a crucial role in quantitative metrics. Without the blend shapes for Gaussian non-rigid compensation, the results are prone to produce artifacts due to imprecise geometry and variable appearance as shown in Fig. 6 (w/o Gau Non.)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 512, + 295, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 295, + 596 + ], + "type": "text", + "content": "The Impact of Baking Teacher. Compared to directly training a small MLP-based network to handle non-rigid deformations as shown in Tab. 3 (w/o Teacher), our proposed multi-stage training strategy is more efficient. With direct supervision of the non-rigid deformation of the teacher network, it is easier for the student network to decouple the geometry deformation and appearance variations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 609, + 130, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 130, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 130, + 623 + ], + "type": "text", + "content": "5. Application" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "TaoAvatar offers a lightweight, comprehensive representation for 3D talking body scenarios, easily deployable on various mobile devices using existing tools [25] as demonstrated in Fig. 1. It seamlessly integrates with AI models to enable real-time dialogue interactions. we deployed a 3D digital human agent on the Apple Vision Pro, which interacts with users through an ASR-LLM-TTS pipeline [16, 31," + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 232 + ], + "type": "image", + "image_path": "59d40bf71ace3044765c29d31cff170e4af8e88e200a969b6f952fc8fa6453a7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 236, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 258 + ], + "type": "text", + "content": "Figure 6. Ablation Study. Red and Green boxes show artifacts and their improved counterparts." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 269, + 561, + 342 + ], + "blocks": [ + { + "bbox": [ + 317, + 269, + 561, + 342 + ], + "lines": [ + { + "bbox": [ + 317, + 269, + 561, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 269, + 561, + 342 + ], + "type": "table", + "html": "
ModelPSNR↑SSIM↑LPIPS↓P2S↓Chamfer↓
w SMPLX28.47.9476.07899.7690.9995
w/o Mesh Non.32.10.9734.03814.4877.5007
w/o Gau Non.31.16.9686.03932.2968.3068
w/o Teacher32.67.9751.03769.5236.5359
Full33.29.9772.03464.2953.3052
", + "image_path": "e45798e617a0d2e041901cb9d698ad4b8be61e79733661336c64e63f0bad4832.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 346, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 369 + ], + "type": "text", + "content": "Table 3. Ablation Study. We use the mesh reconstructed by NeuS2 [54] as a pseudo-truth for geometry evaluation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 380, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 440 + ], + "type": "text", + "content": "39, 62]. Facial expressions and gestures are dynamically controlled by an Audio2BS model [14], allowing the agent to respond naturally with synchronized speech, expressions, and movements. A live demonstration is available in our supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 384, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 384, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 384, + 463 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 590 + ], + "type": "text", + "content": "Conclusion. In this work, we present TaoAvatar, a lightweight, lifelike full-body talking avatar solution. We demonstrate how the teacher-student framework captures high-definition facial and body details while ensuring real-time performance on AR devices. TaoAvatar can be driven by diverse signals, including facial expressions, hand gestures, and body poses. Through both quantitative and qualitative evaluations, we showcase the advantages of our approach. Additionally, we validate its practical potential with a real-world application on the Apple Vision Pro." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "content": "Limitations and Future Works. TaoAvatar encounters challenges in modeling flexible clothing deformation under exaggerated body poses, which are out-of-distribution of training data. A possible solution is to integrate GNN simulators [17, 18] handling larger hemlines, which are compatible with our approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 662, + 554, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 554, + 711 + ], + "type": "text", + "content": "Potential Social Impact. TaoAvatar can synthesize life-like talking digital humans within an augmented reality environment, generating fabricated 3D content or 2D videos. Therefore, responsible use of this technology is essential." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10730" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 290, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 290, + 102 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 290, + 102 + ], + "type": "text", + "content": "[1] https://github.com/zju3dv/EasyMocap.5, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 103, + 290, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 290, + 114 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 290, + 114 + ], + "type": "text", + "content": "[2] https://github.com/ShenhanQian/VHAP.5,13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 114, + 295, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 114, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 62, + 114, + 295, + 169 + ], + "type": "text", + "content": "[3] Rinet Abdrashitov, Kim Raichstat, Jared Monsen, and David Hill. Robust skin weights transfer via weight inpainting. In SIGGRAPH Asia 2023 Technical Communications, SA Technical Communications 2023, Sydney, NSW, Australia, December 12-15, 2023, pages 25:1-25:4. ACM, 2023. 3, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 170, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 247 + ], + "type": "text", + "content": "[4] Thiemo Alldieck, Marcus A. Magnor, Weipeng Xu, Christian Theobalt, and Gerard Pons-Moll. Video based reconstruction of 3d people models. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 8387-8397. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 323 + ], + "type": "text", + "content": "[5] Thiemo Alldieck, Marcus A. Magnor, Bharat Lal Bhatnagar, Christian Theobalt, and Gerard Pons-Moll. Learning to reconstruct people in clothing from a single RGB camera. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1175-1186. Computer Vision Foundation / IEEE, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 390 + ], + "type": "text", + "content": "[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 5419-5429. IEEE, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 424 + ], + "type": "text", + "content": "[7] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Linchao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 491 + ], + "type": "text", + "content": "[8] Xu Chen, Yufeng Zheng, Michael J. Black, Otmar Hilliges, and Andreas Geiger. SNARF: differentiable forward skinning for animating non-rigid neural implicit shapes. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 11574-11584. IEEE, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 492, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 545 + ], + "type": "text", + "content": "[9] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 548, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 295, + 590 + ], + "type": "text", + "content": "[10] Yushuo Chen, Zerong Zheng, Zhe Li, Chao Xu, and Yebin Liu. Meshavatar: Learning high-quality triangular human avatars from multi-view videos. CoRR, abs/2407.08414, 2024. 2, 3, 5, 6, 7, 13, 14, 16" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 592, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 689 + ], + "type": "text", + "content": "[11] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 19925-19936. IEEE, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Edilson De Aguiar, Christian Theobalt, Carsten Stoll, and Hans-Peter Seidel. Marker-less deformable mesh tracking" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "for human shape and motion capture. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 152 + ], + "type": "text", + "content": "[13] Helisa Dhamo, Yinyu Nie, Arthur Moreau, Jifei Song, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Headgas: Real-time animatable head avatars via 3d gaussian splatting. In ECCV (2), pages 459-476. Springer, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 153, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 195 + ], + "type": "text", + "content": "[14] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through A unified model. CoRR, abs/2408.00762, 2024. 7, 8, 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[15] Yao Feng, Weiyang Liu, Timo Bolkart, Jinlong Yang, Marc Pollefeys, and Michael J. Black. Learning disentangled avatars with hybrid 3d representations. CoRR, abs/2309.06441, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "text", + "content": "[16] Zhifu Gao, Shiliang Zhang, Ian McLoughlin, and Zhijie Yan. Paraformer: Fast and accurate parallel transformer for non-autoregressive end-to-end speech recognition. In 23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18-22, 2022, pages 2063-2067. ISCA, 2022. 8, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 376 + ], + "type": "text", + "content": "[17] Artur Grigorev, Michael J. Black, and Otmar Hilliges. HOOD: hierarchical graphs for generalized modelling of clothing dynamics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16965-16974. IEEE, 2023. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 377, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 443 + ], + "type": "text", + "content": "[18] Artur Grigorev, Giorgio Becherini, Michael J. Black, Otmar Hilliges, and Bernhard Thomaszewski. Contourcraft: Learning to resolve intersections in neural multi-garment simulations. In ACM SIGGRAPH 2024 Conference Papers, SIGGRAPH 2024, Denver, CO, USA, 27 July 2024-1 August 2024, page 81. ACM, 2024. 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "text", + "content": "[19] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics, 40(4), 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 490, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 567 + ], + "type": "text", + "content": "[20] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 634-644. IEEE, 2024. 2, 3, 5, 7, 14, 16" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "text", + "content": "[21] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20418-20431. IEEE, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "text", + "content": "[22] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Trans. Graph., 42(4):160:1-160:12, 2023. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[23] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In IEEE/CVF Conference on Computer Vision and" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10731" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5595-5605. IEEE, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[24] Tianjian Jiang, Xu Chen, Jie Song, and Otmar Hilliges. Instantavatar: Learning avatars from monocular video in 60 seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16922-16932, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 217 + ], + "type": "text", + "content": "[25] Xiaotang Jiang, Huan Wang, Yiliu Chen, Ziqi Wu, Lichuan Wang, Bin Zou, Yafeng Yang, Zongyang Cui, Yu Cai, Tianhang Yu, Chengfei Lyu, and Zhihua Wu. MNN: A universal and efficient inference engine. In Proceedings of the Third Conference on Machine Learning and Systems, MLSys 2020, Austin, TX, USA, March 2-4, 2020. mlsys.org, 2020. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[26] Yujiang Jiang, Qingmin Liao, Xiaoyu Li, Li Ma, Qi Zhang, Chaopeng Zhang, Zongqing Lu, and Ying Shan. UV gaus-sians: Joint learning of mesh deformation and gaussian textures for human avatar modeling. CoRR, abs/2403.11589, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 348 + ], + "type": "text", + "content": "[27] Yujiang Jiang, Qingmin Liao, Zhaolong Wang, Xiangru Lin, Zongqing Lu, Yuxi Zhao, Hanqing Wei, Jingrui Ye, Yu Zhang, and Zhijing Shao. Smplx-lite: A realistic and drivable avatar benchmark with rich geometry and texture annotations. In IEEE International Conference on Multimedia and Expo, ICME 2024, Niagara Falls, ON, Canada, July 15-19, 2024, pages 1-6. IEEE, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 415 + ], + "type": "text", + "content": "[28] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5322-5332. IEEE, 2024. 16" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "text", + "content": "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[30] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: human gaussian splats. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 505-515. IEEE, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 515, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 591 + ], + "type": "text", + "content": "[31] Jungil Kong, Jihoon Park, Beomjeong Kim, Jeongmin Kim, Dohee Kong, and Sangjin Kim. VITS2: improving quality and efficiency of single-stage text-to-speech with adversarial learning and architecture design. In 24th Annual Conference of the International Speech Communication Association, Interspeech 2023, Dublin, Ireland, August 20-24, 2023, pages 4374-4378. ISCA, 2023. 8, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 592, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 658 + ], + "type": "text", + "content": "[32] Youngjoong Kwon, Baole Fang, Yixing Lu, Haoye Dong, Cheng Zhang, Francisco Vicente Carrasco, Albert Mosella-Montoro, Jianjin Xu, Shingo Takagi, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Generalizable human gaussians for sparse view synthesis. In ECCV (78), pages 451-468. Springer, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[33] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. GART: gaussian articulated template models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19876-19887. IEEE, 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 138 + ], + "type": "text", + "content": "[34] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19711-19722. IEEE, 2024. 2, 3, 5, 6, 7, 13, 14, 15, 16" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[35] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3d gaussian splatting for inverse rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21644-21653. IEEE, 2024. 16" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "type": "text", + "content": "[36] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: neural free-view synthesis of human actors with pose control. ACM Trans. Graph., 40(6):219:1-219:16, 2021. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 1, 2, 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 283, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 316 + ], + "type": "text", + "content": "[38] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 316, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 425 + ], + "type": "text", + "content": "[39] Chengfei Lv, Chaoyue Niu, Renjie Gu, Xiaotang Jiang, Zhaode Wang, Bin Liu, Ziqi Wu, Qiulin Yao, Congyu Huang, Panos Huang, Tao Huang, Hui Shu, Jinde Song, Bin Zou, Peng Lan, Guohuan Xu, Fei Wu, Shaojie Tang, Fan Wu, and Guihai Chen. Walle: An end-to-end, general-purpose, and large-scale production system for device-cloud collaborative machine learning. In 16th USENIX Symposium on Operating Systems Design and Implementation, OSDI 2022, Carlsbad, CA, USA, July 11-13, 2022, pages 249-265. USENIX Association, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 13, 16" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[41] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In CVPR, pages 788-798. IEEE, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "text", + "content": "[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 570, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 635 + ], + "type": "text", + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1, 2, 3, 5, 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 635, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 691 + ], + "type": "text", + "content": "[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[45] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gauss" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10732" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "sianavatars: Photorealistic head avatars with rigged 3d gaus-sians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20299-20309. IEEE, 2024. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 182 + ], + "type": "text", + "content": "[46] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5020-5030. IEEE, 2024. 2, 5, 7, 14, 16" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "text", + "content": "[47] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 130-141. IEEE, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 315 + ], + "type": "text", + "content": "[48] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. Splattingavatar: Realistic real-time human avatars with mesh-embedded gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 1606-1616. IEEE, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "type": "text", + "content": "[49] Kaiyue Shen, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Julien Valentin, Jie Song, and Otmar Hilliges. X-avatar: Expressive human avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 16911-16921. IEEE, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "text", + "content": "[50] Carsten Stoll, Juergen Gall, Edilson De Aguiar, Sebastian Thrun, and Christian Theobalt. Video-based reconstruction of animatable human characters. ACM Transactions on Graphics (TOG), 29(6):1-10, 2010. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[51] Daniel Vlasic, Pieter Peers, Ilya Baran, Paul E. Debevec, Jovan Popovic, Szymon Rusinkiewicz, and Wojciech Matusik. Dynamic shape capture using multi-view photometric stereo. ACM Trans. Graph., 28(5):174, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "text", + "content": "[52] Lizhen Wang, Xiaochen Zhao, Jingxiang Sun, Yuxiang Zhang, Hongwen Zhang, Tao Yu, and Yebin Liu. Stylea-vatar: Real-time photo-realistic portrait avatar from a single video. In ACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH 2023, Los Angeles, CA, USA, August 6-10, 2023, pages 67:1-67:10. ACM, 2023. 3, 6, 15" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "type": "text", + "content": "[53] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 550-560. IEEE, 2024. 3, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 603, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 669 + ], + "type": "text", + "content": "[54] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3272-3283. IEEE, 2023. 2, 3, 4, 8, 13, 15, 16" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[55] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4): 600-612, 2004. 15" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[56] Jing Wen, Xiaoming Zhao, Zhongzheng Ren, Alexander G. Schwing, and Shenlong Wang. Gomavatar: Efficient animatable human modeling from monocular video using gaussians-on-mesh. In CVPR, pages 2059-2069. IEEE, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 128, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 194 + ], + "type": "text", + "content": "[57] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 16189-16199. IEEE, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "text", + "content": "[58] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "type": "text", + "content": "[59] Donglai Xiang, Timur M. Bagautdinov, Tuur Stuyck, Fabian Prada, Javier Romero, Weipeng Xu, Shunsuke Saito, Jingfan Guo, Breannan Smith, Takaaki Shiratori, Yaser Sheikh, Jessica K. Hodgins, and Chenglei Wu. Dressing avatars: Deep photorealistic appearance for physically simulated clothing. ACM Trans. Graph., 41(6):222:1-222:15, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 327, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 361 + ], + "type": "text", + "content": "[60] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In CVPR, pages 1802-1812. IEEE, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 361, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 426 + ], + "type": "text", + "content": "[61] Yuelang Xu, Bengwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 1931-1941. IEEE, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 427, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 591 + ], + "type": "text", + "content": "[62] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. Qwen2 technical report. CoRR, abs/2407.10671, 2024. 8, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[63] Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. In CVPR, pages 896-905. IEEE, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "type": "text", + "content": "[64] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-595. Computer Vision Foundation / IEEE Computer Society, 2018. 4, 15" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10733" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 383 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[65] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humanerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 171 + ], + "type": "text", + "content": "[66] Zhongyuan Zhao, Zhenyu Bao, Qing Li, Guoping Qiu, and Kanglin Liu. Psavatar: A point-based shape model for real-time head avatar animation with 3d gaussian splatting, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[67] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CoRR, abs/2401.03407, 2024. 15" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "type": "text", + "content": "[68] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas J. Guibas, and Gordon Wetzstein. Physavatar: Learning the physics of dressed 3d avatars from visual observations. CoRR, abs/2404.04421, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 327 + ], + "type": "text", + "content": "[69] Zerong Zheng, Xiaochen Zhao, Hongwen Zhang, Boning Liu, and Yebin Liu. Avatarrex: Real-time expressive full-body avatars. ACM Trans. Graph., 42(4):158:1-158:19, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 383 + ], + "type": "text", + "content": "[70] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4574-4584. IEEE, 2023. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10734" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Targeted Forgetting of Image Subgroups in CLIP Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_content_list.json b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..70656db7070ebd45020d20d62a9e44f9414d4741 --- /dev/null +++ b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_content_list.json @@ -0,0 +1,1456 @@ +[ + { + "type": "text", + "text": "Targeted Forgetting of Image Subgroups in CLIP Models", + "text_level": 1, + "bbox": [ + 209, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeliang Zhang $^{1}$ Gaowen Liu $^{2}$ Charles Fleming $^{2}$ Ramana Rao Kompella $^{2}$ Chenliang Xu $^{1}$ $^{1}$ University of Rochester $^{2}$ Cisco Research", + "bbox": [ + 109, + 179, + 883, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zeliang.zhang, chenliang.xu}@rochester.edu, {gaoliu, chflemin, rkompell}@cisco.com", + "bbox": [ + 130, + 218, + 861, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation models (FMs) such as CLIP have demonstrated impressive zero-shot performance across various tasks by leveraging large-scale, unsupervised pre-training. However, they often inherit harmful or unwanted knowledge from noisy internet-sourced datasets, compromising their reliability in real-world applications. Existing model unlearning methods either rely on access to pre-trained datasets or focus on coarse-grained unlearning (e.g., entire classes), leaving a critical gap for fine-grained unlearning. In this paper, we address the challenging scenario of selectively forgetting specific portions of knowledge within a class—without access to pre-trained data—while preserving the model's overall performance. We propose a novel three-stage approach that progressively unlearns targeted knowledge while mitigating over-forgetting. It consists of (1) a forgetting stage to fine-tune the CLIP on samples to be forgotten, (2) a reminding stage to restore performance on retained samples, and (3) a restoring stage to recover zero-shot capabilities using model souping. Additionally, we introduce knowledge distillation to handle the distribution disparity between forgetting/retaining samples and unseen pre-trained data. Extensive experiments on CIFAR-10, ImageNet-1K, and style datasets demonstrate that our approach effectively unlearns specific subgroups while maintaining strong zero-shot performance on semantically similar subgroups and other categories, significantly outperforming baseline unlearning methods, which lose effectiveness under the CLIP unlearning setting.", + "bbox": [ + 89, + 300, + 485, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 753, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation models (FMs) are deep neural networks pretrained on large-scale datasets [50]. By employing unsupervised training, these models capture complex patterns and relationships within the data, enabling a deep understanding of various domains [4]. For example, CLIP [43], a vision-language model, learns correspondences between natural language and images. Trained on the large-scale LAION-5B dataset [47], CLIP demonstrates strong zero", + "bbox": [ + 89, + 779, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba6cd99b7c5621bd4b277f2eed7cbf59e04a64fe32deca2be9df55c53528bb89.jpg", + "image_caption": [ + "Figure 1. In the task of CLIP unlearning, classical machine unlearning methods often fail to generalize across multiple datasets due to limited access to the original pre-training data. In contrast, our proposed method does not require access to the pre-training dataset. After unlearning, CLIP not only forgets the specified subgroup of images but also maintains consistent performance on various unseen datasets in the zero-shot classification setting." + ], + "image_footnote": [], + "bbox": [ + 519, + 267, + 906, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shot performance [10, 22, 44]. However, its reliability in real-world applications is questionable [63, 64]. The training data, sourced from the internet, includes uncensored content such as discriminatory and violent imagery, copyright violations, and personal information [21, 42, 59]. This problematic content can be unintentionally learned during pre-training, raising concerns about the reliability and safety of downstream models [9, 31]. Therefore, it is critical to remove such harmful knowledge to ensure trustworthy FMs.", + "bbox": [ + 511, + 593, + 908, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numerous studies have explored model unlearning, which can be broadly divided into exact [46, 53, 58] and approximate strategies [11, 23, 38]. Exact unlearning aims to completely eliminate the influence of forgotten data from the model, often by retraining it [7, 58, 62], ensuring the model behaves as if the data had never been included in training. While thorough, this approach is computationally expensive and impractical for large-scale models [32, 48]. In contrast, approximate unlearning offers a more efficient alternative, using techniques like variational inference [20, 38] or adaptive regularization [27, 37, 61] to partially reduce the data's", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9870", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "impact. However, these methods face several challenges when applied to FMs.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, as shown in Fig. 1, the training data for FMs is often vast or inaccessible [5, 47], rendering exact unlearning impractical. Second, while FMs are trained on large-scale datasets using unsupervised learning, the fine-tuning typically used in approximate unlearning strategies [26, 51] can degrade generalization across other knowledge domains [1, 35, 41], leading to significant performance drops, particularly in zero-shot tasks [55]. Third, the knowledge to be forgotten is often deeply entangled with other learned information [17, 49]. This entanglement makes it difficult to isolate and remove specific knowledge from the model parameters, particularly when distributional disparities exist between the data we want to unlearn and the pre-trained data.", + "bbox": [ + 89, + 123, + 482, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few studies have specifically addressed the challenge of unlearning in foundational models, particularly in CLIP. Most existing approaches focus on unlearning entire classes or broad concepts, yet real-world applications often require a more targeted approach, where only certain portions of knowledge within a class are forgotten. For example, we may want the model to forget representations of Disney's Mickey Mouse while retaining versions created by the public, a task known as subgroup image forgetting. Moreover, many current unlearning methods rely on access to the original pre-trained dataset to restore model performance, which is often unavailable. This limitation underscores a critical question in current research: how can models like CLIP achieve selective, fine-grained unlearning without relying on pre-trained data, while still maintaining generalization across other tasks?", + "bbox": [ + 89, + 321, + 482, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we bridge this gap by tackling a more challenging and practical scenario: enabling CLIP to selectively forget specific portions of knowledge within a single class without access to pre-trained data, framing this as a fine-grained unlearning task. Rather than unlearning an entire class or concept (e.g., all images of airplanes), we focus on selectively forgetting images from a particular subset (e.g., Boeing airplanes) while preserving the model's knowledge of other subsets (e.g., different manufacturers). This selective forgetting is particularly challenging due to the feature entanglement across subsets within a class, and it is exceedingly difficult to define and collect all subsets within the same class in a foundational model setting, which is trained on vast, large-scale data.", + "bbox": [ + 89, + 566, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenge of partial knowledge removal, we propose a novel, three-stage approach: forgetting, reminding, and restoring. In the forgetting stage, we identify layers essential for representing the forgetting samples but less critical to other data, fine-tuning the model on these samples to selectively remove targeted knowledge. We observe that models can over- forget samples from similar distributions, so we introduce a retaining set with a similar distribution to the", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "targeted subgroup, aiding layer identification and supporting the reminding stage. During reminding, we fine-tune the model on the retaining samples to reinforce CLIP's performance on data prone to over-forgetting due to distributional similarity. In the final restore stage, we employ model souping to fully recover CLIP's zero-shot performance. Throughout this process, we address the distributional gap between forgetting/retaining samples and unseen pre-trained data by distilling knowledge from CLIP, achieving a more precise representation without relying on the original pre-trained dataset. Extensive experiments demonstrate our method's effectiveness in selectively unlearning subgroup images while preserving robust zero-shot performance across diverse distributions and datasets.", + "bbox": [ + 511, + 90, + 906, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 535, + 303, + 823, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We formulate the problem of subgroup image unlearning within the CLIP model.", + "2. We propose a three-stage method for unlearning in CLIP, consisting of forgetting, reminding, and restoring. Our method is data-efficient, which does not require access to CLIP's pre-training dataset, making it more practical for real-world applications.", + "3. We conduct extensive experiments to validate the effectiveness of our proposed method, including unlearning tasks on CIFAR-10, ImageNet-1K, and style forgetting. Additionally, we demonstrate the scalability of our approach across various applications through retrieval-based evaluations." + ], + "bbox": [ + 514, + 316, + 903, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 527, + 650, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Machine Unlearning. Various strategies have been developed to facilitate machine unlearning in models. Notably, Golatkar et al. [19], Warnecke et al. [54] employ a method of continuous fine-tuning (FT) on the retain dataset—those data elements the model is expected to remember. This approach not only approximates the forgetting of the target dataset but also mitigates the issue of overforgetting [32, 39], in contrast to exclusively fine-tuning on the forgetting dataset. Becker and Liebig [3], Liu et al. [32] leverage Fisher information, which assesses the sensitivity of parameters concerning the forgetting dataset, to achieve model unlearning. Meanwhile, Thudi et al. [52] unroll the knowledge embedded in the forgetting dataset from the model parameters using stochastic gradient ascent (GA). Further, [18] introduces the use of local Lipschitz (LIP) perturbations on the forgetting dataset as a technique for unlearning in image classification models, which Kravets and Namboodiri [29] subsequently adapts to the CLIP model. Additionally, Chandawat et al. [12] proposes an error minimization-maximization (EMMN) framework that aims to maximize errors on the forgetting dataset while minimizing them on the retraining dataset, thus achieving a balance between forgetting and retaining performance. Among these methods, EMMN [12] and LIP [29] stand out", + "bbox": [ + 511, + 553, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9871", + "bbox": [ + 482, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as they do not require access to the training datasets (both forgetting and retaining), unlike other approaches.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the Importance of Unlearning in CLIP. Trained on a billion-scale dataset of image-text pairs [47], the CLIP model [43] excels in zero-shot classification tasks. Despite its strengths, the presence of uncensored samples within the training data has raised practical concerns [24]. Given the impracticality of retraining the CLIP model from scratch due to significant resource demands and the inaccessibility of the large-scale dataset, it is imperative to explore machine unlearning methods to selectively erase problematic knowledge from the model. Poppi et al. [40] suggest an approach to first distinguish between safe and unsafe content, followed by fine-tuning CLIP to alter the embeddings of unsafe content. Meanwhile, Cai et al. [8] focuses on identifying and unlearning from the most influential layer in CLIP, using a subset of the training set to restore zero-shot capabilities. [29] expand zero-shot machine unlearning techniques to the CLIP model, enabling the forgetting of entire concepts as indicated by prompts. Concurrently, Kravets and Namboodiri [29] apply Layer-wise Relevance Propagation (LoRA) to the text encoder of CLIP, specifically fine-tuning this component to unlearn a given concept within the model.", + "bbox": [ + 91, + 122, + 486, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 450, + 297, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The CLIP model, denoted $g(\\cdot, \\cdot) = \\{g^{\\mathrm{img}}(\\cdot), g^{\\mathrm{txt}}(\\cdot)\\}$ , processes image and text inputs into embeddings, $e^{\\mathrm{img}}$ and $e^{\\mathrm{txt}}$ . It aims for high cosine similarity between these embeddings when they are correctly paired, and low similarity otherwise, using a threshold $\\delta$ .", + "bbox": [ + 89, + 476, + 483, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose a large dataset $D = \\{x_{n}^{\\mathrm{img}}, x_{n}^{\\mathrm{txt}}, c_{n}, g_{n}\\}_{n=1}^{N}$ , where we denote $x^{\\mathrm{img}}$ as the image, $x^{txt}$ as the paired text learned by the CLIP, $c$ as the coarse main category, and $g$ is the fine-grained subclass label. The forgetting dataset $D^{f} = \\{x_{n}^{\\mathrm{img}}, x_{n}^{\\mathrm{txt}}, c_{n}, g_{n}\\}_{n=1}^{N_{f}}$ is usually a small set compared with $D$ . The goal is to unlearn the associations between $x^{\\mathrm{img}}$ and $x^{txt}$ on $D^{f}$ without impacting the model's performance on the remaining dataset $D^{r} = D / D_{f}$ , namely the retain set. The task involves selectively erasing knowledge from $D^{f}$ while maintaining performance on $D^{r}$ , which can be formulated as follows:", + "bbox": [ + 89, + 551, + 483, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\forall \\left(x _ {i} ^ {\\mathrm {i m g}}, x ^ {\\mathrm {t x t}}\\right) \\in D ^ {f}, \\frac {\\hat {g} ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot \\hat {g} ^ {\\mathrm {t x t}} \\left(x ^ {\\mathrm {t x t}}\\right)}{\\| \\hat {g} ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot \\hat {g} ^ {\\mathrm {t x t}} \\left(x ^ {\\mathrm {t x t}}\\right) \\|} < \\delta , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 727, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\forall (x _ {i} ^ {\\mathrm {i m g}}, x _ {i} ^ {\\mathrm {t x t}}) \\in D ^ {r}, \\hat {g} (x _ {i} ^ {\\mathrm {i m g}}, x _ {i} ^ {\\mathrm {t x t}}) \\approx g (x _ {i} ^ {\\mathrm {i m g}}, x _ {i} ^ {\\mathrm {t x t}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 766, + 439, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{g}$ represents the unlearned CLIP model and $g$ the original model.", + "bbox": [ + 89, + 794, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implementing the approach outlined in Eq. (1) is challenging due to several factors:", + "bbox": [ + 89, + 825, + 483, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. The large dataset $D$ used for training $g$ is inaccessible, and similarly, $D_r$ is also unavailable. We can only access a small subset of $D_r$ to recover model performance", + "bbox": [ + 91, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "after forgetting, blocking the direct use of conventional machine unlearning methods.", + "bbox": [ + 532, + 90, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. In addition, we may only have access to a coarse label $c$ as the surrogate of $x^{txt}$ rather than a fine-grained subgroup label $g$ . This limitation increases the risk of inaccurately erasing targeted knowledge and over-forgetting unrelated knowledge that is similar to the target.", + "3. There may be a substantial distribution shift between $D_{f}$ and $D$ . With a large number of parameters, directly finetuning on the small dataset $D_{f}$ could cause the model to overfit and lose generalization ability." + ], + "bbox": [ + 511, + 121, + 906, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, the goal of this task is twofold: it involves selectively forgetting only specific image subgroups because of coarse labeling, while simultaneously preserving the model's robust zero-shot capabilities across various datasets, including those characterized by similarly coarse labels.", + "bbox": [ + 511, + 257, + 906, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 345, + 648, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the challenge of subgroup image unlearning, we propose a novel methodology that comprises three stages: forgetting, reminding, and restoring, as depicted in Fig. 2.", + "bbox": [ + 511, + 371, + 906, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Forgetting", + "text_level": 1, + "bbox": [ + 511, + 425, + 632, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous studies [14, 34] indicate that selectively fine-tuning critical layers for a target dataset can effectively mitigate over-forgetting [33], typically using Fisher information [36, 57] to assess the importance of each layer. However, while Fisher information measures parameter sensitivity to the target dataset, it overlooks potential impacts on other datasets, which are inaccessible in our task. Our experiments reveal that relying solely on the forgetting dataset to modify the CLIP model often leads to catastrophic over-forgetting on similar subgroups with distributions akin to the forgetting data, as shown in Tab. 1 for the CLIP model optimized by GA. To address this, we propose using relative Fisher information which combines both the forgetting and retaining samples as a more nuanced metric for identifying essential layers for fine-tuning, thereby enhancing the precision of the unlearning process.", + "bbox": [ + 511, + 446, + 908, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concretely, given the forgetting dataset $\\mathcal{D}^f = \\{(x_i^{\\mathrm{img}}, x^{\\mathrm{txt}})\\}$ , we manually construct a retain dataset $\\mathcal{D}^r = \\{(x_i^{\\mathrm{img}}, x_i^{\\mathrm{txt}})\\}$ with a distribution similar to $\\mathcal{D}^f$ , used for alleviating the aforementioned subgroup over-forgetting. Here, $\\mathcal{D}^r$ is comprised of similar subgroups of images that we aim to preserve in the model's memory. In practical scenarios, we often use coarse labels as surrogate image descriptions, so $x^{\\mathrm{txt}}$ from the forgetting dataset may also appear in $\\mathcal{D}^r$ . We calculate the relative Fisher information for the $l$ -th layer in $g^{\\mathrm{img}}(\\cdot)$ as follows:", + "bbox": [ + 511, + 688, + 908, + 843 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} ^ {l} = \\frac {\\mathbb {E} _ {\\left(x _ {i} ^ {\\mathrm {i m g}} , x ^ {\\mathrm {t x t}}\\right) \\sim \\mathcal {D} ^ {f}} \\left[ \\nabla_ {\\theta^ {l}} ^ {2} \\left(g ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot g ^ {\\mathrm {t x t}} \\left(x ^ {\\mathrm {t x t}}\\right)\\right) \\right]}{\\mathbb {E} _ {\\left(x _ {i} ^ {\\mathrm {i m g}} , x _ {i} ^ {\\mathrm {t x t}}\\right) \\sim \\mathcal {D} ^ {r}} \\left[ \\nabla_ {\\theta^ {l}} ^ {2} \\left(g ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot g ^ {\\mathrm {t x t}} \\left(x _ {i} ^ {\\mathrm {t x t}}\\right)\\right) \\right]}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 852, + 906, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9872", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/66fbaebd037d2b8e98bc47ce995193b54015a996c2a3098f5f397f2e95dd3b0a.jpg", + "image_caption": [ + "Figure 2. Overview of our method to unlearn the CLIP on a subgroup of images. We first compute the relative fisher information on the forgetting dataset and retain dataset to evaluate the layer importance. We select important layers for the following fine-tuning process. During the fine-tuning process for knowledge reminding, we align the distribution of our retain dataset to that of the pre-training dataset learned by the model. A model merging operation is performed at the last step of knowledge restoration." + ], + "image_footnote": [], + "bbox": [ + 178, + 88, + 823, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta^l$ represents the parameters of the $l$ -th layer.", + "bbox": [ + 89, + 395, + 429, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then rank the importance of different layers based on the computed relative Fisher information, as detailed in Eq. (2). Layers with higher relative importance are selected for targeted unlearning, with all other layers frozen to minimize the impact on non-target knowledge. In our experiments, applying the low-rank adaptation (LoRA) technique [25] to the selected layers significantly enhances the targeted unlearning performance while minimizing effects on other retained knowledge.", + "bbox": [ + 89, + 411, + 483, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we fine-tune our model on the forgetting dataset $\\mathcal{D}_f$ to minimize the following loss function:", + "bbox": [ + 89, + 547, + 483, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\mathcal {L} _ {f} = \\sum_ {i} \\frac {g ^ {i m g} \\left(x _ {i} ^ {i m g}\\right) \\cdot g ^ {t x t} \\left(x ^ {t x t}\\right)}{\\| g ^ {i m g} \\left(x _ {i} ^ {i m g}\\right) \\cdot g ^ {t x t} \\left(x ^ {t x t}\\right) \\|}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 590, + 483, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where images in the forgetting dataset are associated through the same coarse label $x^{txt}$ .", + "bbox": [ + 89, + 640, + 482, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By fine-tuning the CLIP model on images marked for unlearning with coarse labels, we aim to selectively erase specific knowledge. However, this adjustment often leads to catastrophic overforgetting, affecting the model's ability across both similar and different image distributions. For example, attempting to forget the \"airplane\" class in CIFAR-10 may also affect the recognition of other classes within CIFAR-10 and in ImageNet-1K. To tackle these challenges, particularly when full access to the pre-training dataset is unavailable, we implement reminding and restoration stages.", + "bbox": [ + 89, + 671, + 483, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Reminding", + "text_level": 1, + "bbox": [ + 89, + 832, + 212, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An intuitive strategy to retain non-targeted knowledge in CLIP is to fine-tune it on our retaining set, built during the forgetting stage for computing $\\mathcal{I}$ . However, two challenges", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hinder the direct application of this fine-tuning step. First, our retaining set may have a different distribution than the pre-training dataset, potentially causing inefficient learning. Second, our retaining set is typically limited, and fine-tuning on it could lead to further catastrophic over-forgetting in various unseen datasets and tasks.", + "bbox": [ + 511, + 395, + 906, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the first challenge, inspired by research showing that training data features are encoded in model parameters [60], we employ an optimization-based strategy to align the retaining dataset's distribution with the pre-training dataset. Specifically, for a given batch of image-text pairs in the retaining dataset $\\mathcal{D}_r = \\{(x_i^{\\mathrm{img}}, x_i^{\\mathrm{xt}})\\}_{i=1}^B$ , we introduce a batch of perturbations $\\delta_{i_{i=1}}^{B}$ on the images. These perturbations are optimized by minimizing the loss:", + "bbox": [ + 511, + 486, + 908, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {a} = \\sum_ {l} \\left\\| \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\mu_ {l} ^ {\\text {i m g}} \\left(x _ {i} ^ {\\text {i m g}} + \\delta_ {i}\\right) - \\mathrm {B N} _ {l} ^ {\\mu} \\right\\| \\tag {4} \\\\ + \\sum_ {l} \\left\\| \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sigma_ {l} ^ {i m g} \\left(x _ {i} ^ {i m g} + \\delta_ {i}\\right) - \\mathbf {B N} _ {l} ^ {\\sigma} \\right\\|, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 619, + 906, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_l^{img}(\\cdot)$ and $\\sigma_l^{img}$ are the mean and variance of the batch of intermediate features before the $l$ -th BatchNormalization (BN) layer, computed by the image encoder of original CLIP. $\\mathrm{BN}_l^\\mu$ and $\\mathrm{BN}_l^\\sigma$ are the mean and variance parameters of the $l$ -th BN layer, encoding the global distribution information of the pre-training dataset.", + "bbox": [ + 511, + 717, + 906, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then fine-tune the CLIP model on the retaining dataset, now aligned with the pre-training dataset's distribution. This serves as a reminder of the knowledge encoded in the original model, effectively achieved without direct access to the pre-training dataset. This strategy helps prevent further catastrophic forgetting that could arise from distributional", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9873", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "disparities between the retaining dataset and the original training dataset.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the second challenge of overfitting during fine-tuning due to the large amount of parameters, we propose minimizing the impact on the original model parameters by using an exponential moving average (EMA) [28] for the final model parameters, as follows:", + "bbox": [ + 89, + 122, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {\\mathrm {e m a}} = \\alpha \\cdot \\theta^ {\\mathrm {e m a}} + (1 - \\alpha) \\cdot \\theta , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 214, + 483, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta^{\\mathrm{ema}}$ is initialized as the original CLIP parameters $\\theta_0$ , $\\theta$ represents the continuously optimized parameters from $\\theta_0$ , and $\\alpha$ is the decay factor. This approach effectively balances parameter retention with necessary updates.", + "bbox": [ + 89, + 239, + 483, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Restoration", + "text_level": 1, + "bbox": [ + 89, + 311, + 218, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to fine-tuning on the retain dataset for reminding, we also employ a model merging strategy to restore the zero-shot performance of CLIP, a technique empirically shown to locate flat optima on the loss landscape, enhancing generalization [56].", + "bbox": [ + 89, + 334, + 483, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first compile a small calibration dataset $\\mathcal{D}_m = \\{(x_i,y_i)\\}_{i = 1}^N$ for zero-shot image classification, which may include portions of the retain dataset not used in fine-tuning. We then execute model merging with varying coefficients as follows:", + "bbox": [ + 89, + 411, + 483, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {\\theta} {\\arg \\max } \\operatorname {A c c} (g, \\mathcal {D} _ {m}), \\tag {6} \\\\ s. t. \\theta = \\alpha \\cdot \\theta^ {f} + (1 - \\alpha) \\cdot \\theta^ {o r i}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 497, + 483, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Acc(g, \\mathcal{D}_m)$ assesses the zero-shot classification performance of model $g$ on $\\mathcal{D}_m$ , and $\\theta^f$ and $\\theta^{ori}$ represent the parameters of the CLIP model post-unlearning and the original model, respectively.", + "bbox": [ + 89, + 551, + 483, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Evaluations", + "text_level": 1, + "bbox": [ + 89, + 628, + 214, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by focusing on the forgetting subgroups within the ImageNet dataset, where the images are semantically grouped by Breeds. We select two superclasses: 'monkey' and 'turtle.' Each superclass contains four distinct subgroups. We specifically target one subgroup in each superclass for unlearning: 'marmoset' in the monkey superclass and 'box turtle' in the turtle superclass.", + "bbox": [ + 89, + 654, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 771, + 277, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings. We conduct experiments in three distinct settings: in-domain forgetting, near-domain forgetting, and style forgetting, as detailed below:", + "bbox": [ + 89, + 794, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- In-domain forgetting: Experiments are conducted on the Breeds dataset [45], a subset of the ImageNet-1K dataset [16] consisting of 4 subgroups within 17 superclasses. The images, with dimensions of $224 \\times 224$ , closely", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "match the distribution of images in the pre-trained LAION-5B dataset used for CLIP. For each superclass, one subgroup is selected for unlearning, while the images from the remaining three subgroups form the retaining set.", + "bbox": [ + 524, + 90, + 906, + 151 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Out-of-domain forgetting: Experiments are performed on the CIFAR-10 dataset [30], which differs significantly from the image distribution in LAION-5B. To align with CLIP's input format, we resize the images from $32 \\times 32$ to $224 \\times 224$ . In this setting, we select a class shared between CIFAR-10 and ImageNet-1K. The images of this selected class form the forgetting dataset, while the remaining classes in CIFAR-10 constitute the retaining set.", + "- Style forgetting: Experiments are conducted on one class in the ImageNet-1K dataset [16] transformed by Stable Diffusion [13]. The images are randomly transformed into three styles: sketch style, poster art, and black-and-white drawing. For a given style to forget, images in the remaining styles are treated as the retaining set." + ], + "bbox": [ + 513, + 152, + 906, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines and implementations. We select six methods as our baselines: FT [54], GA [52], Fisher [32], LIP [18], and EMMN [12], and CLIP-LIP [29]. Except the CLIP-LIP, these methods are developed for conventional models that require a classification head. To address this difference, we use the probability from the zero-shot classification task as a surrogate for the classification logits. In these approaches, the text prompt is provided by the superclass names in both the forgetting and retain sets. For all learning-based methods, recognizing the sensitivity of CLIP parameters during the fine-tuning process, we use the Adam optimizer with a small learning rate of $1 \\times 10^{-6}$ for up to 2 epochs. For LIP, we set the number of noisy copies to 10. Additionally, the variance scaling factor $\\alpha$ in Fisher is set to 0.2. We provide the full implementation in Appendix A.1.", + "bbox": [ + 511, + 367, + 906, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Evaluations.** The performance of image subgroup unlearning in CLIP is evaluated along three dimensions. First, we must ensure that CLIP successfully forgets the images from the specified subgroup. Second, CLIP must continue to perform consistently on the retain set, which includes the other subgroups that are not intended to be forgotten. Third, CLIP must maintain strong zero-shot performance across various tasks and datasets. These considerations lead us to evaluate CLIP's performance not only on the targeted forgetting subgroups and unaffected subgroups, but also on various unseen datasets during the forgetting process, including Food [6], STL [15], and ObjectNet [2]. Additionally, we will use the post-unlearning CLIP to retrieve images in order to assess the impact of forgetting on its learned representations. Additionally, we use the classification accuracy ratio between the original and post-unlearned CLIP models to gauge the degree of knowledge forgetting and retention. More specifically, the ratio is computed as $\\min\\left(\\frac{\\text{Acc\\_learn}}{\\text{Acc}_{ori}}, 1\\right)$ , where we cap the score at 1 to represent full knowledge restoration. This ratio is shown in the subscript of each result. We further", + "bbox": [ + 511, + 598, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9874", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6790c4fa66c2a078aaa350c00e9ab582bbff7da4870e39d844eca33096d766e3.jpg", + "table_caption": [ + "Table 1. Evaluation results for forgetting the subclass 'marmoset' within the superclass 'monkey' (left) and 'box turtle' within the superclass 'turtle' (right) in the ImageNet-1K dataset are presented (Target). Each superclass contains four subgroups, with the remaining three designated as the retain set (Retain)." + ], + "table_footnote": [], + "table_body": "
BackboneMethodImageNetCIFARFoodSTLObjectNetScoreBackboneMethodImageNetAll (↑)CIFARFoodSTLObjectNetScore
Target (↓)Retain (↑)All (↑)Target (↓)Retain (↑)
RN50Original51.054.759.870.460.992.068.9-RN50Original64.570.359.870.460.992.068.9-
FT2.5.4.987.1.100.025.6.42.813.5.19.217.4.28.543.1.46.819.6.58.451.6FT0.0.0.072.1.100.026.0.43.512.6.17.928.6.47.070.3.76.420.2.93.359.2
GA0.0.0.00.9.1.632.2.53.816.4.23.322.9.37.663.3.68.822.1.32.145.3GA0.0.0.013.3.18.942.1.70.414.6.20.740.0.55.777.8.84.621.3.30.955.8
Fisher0.2.0.30.5.0.91.3.2.110.4.14.70.2.0.310.5.11.43.0.14.419.1Fisher1.2.1.92.0.1.83.5.5.99.7.13.80.4.0.710.5.14.11.6.23.319.3
LIP0.7.1.40.2.3.61.3.2.210.7.15.20.2.0.310.3.11.21.6.2.318.6LIP0.4.0.63.4.4.815.2.25.410.6.15.16.1.10.023.7.25.811.3.16.428.1
EMMN0.0.0.056.7.100.028.4.47.513.1.18.620.4.33.554.9.59.720.5.29.755.6EMMN0.1.0.257.2.81.428.1.47.012.5.17.822.1.36.364.7.70.317.0.25.554.0
CLIP-LIP2.65.10.4.0.717.7.29.618.8.26.74.2.6.950.8.55.212.5.18.133.2CLIP-LIP1.9.2.92.0.2.831.9.53.319.5.27.76.9.11.363.6.69.134.2.69.644.4
Ours0.0.0.050.2.91.754.5.91.185.7.100.062.9.100.081.5.88.645.2.56.691.0Ours0.0.0.069.4.98.750.6.84.654.5.77.450.5.82.987.6.95.243.1.62.685.9
RN101Original66.252.662.364.457.392.469.2-RN101Original66.277.162.364.457.392.469.2-
FT1.3.2.075.2.100.027.6.44.311.8.18.34.4.7.635.1.48.017.3.25.047.3FT0.0.0.068.0.88.232.4.52.011.4.17.727.8.48.566.0.71.423.5.34.058.8
GA0.7.1.10.3.0.536.2.58.119.9.9.929.3.51.158.6.63.419.4.88.047.3GA0.0.0.022.1.28.746.4.74.512.2.18.940.4.70.575.6.81.825.9.37.458.8
Fisher0.9.1.40.4.0.70.3.0.511.7.18.20.2.0.39.6.10.31.0.1.418.6Fisher1.9.2.92.4.3.10.9.1.411.0.17.10.3.0.59.5.10.30.6.0.918.6
LIP0.2.0.30.3.0.52.0.3.210.4.16.10.2.0.310.7.11.61.9.2.719.2LIP0.81.23.2.4.212.9.20.710.3.16.08.4.14.716.9.18.314.7.21.227.7
EMMN0.7.1.162.5.100.030.1.48.322.0.35.110.8.18.810.0.10.823.8.34.449.5EMMN0.0.0.048.6.83.032.5.52.29.2.14.322.4.39.160.8.65.822.9.33.152.5
CLIP-LIP2.7.4.11.5.2.932.0.61.428.6.44.410.9.19.049.6.53.714.9.21.541.2CLIP-LIP4.0.6.910.8.14.038.1.61.238.5.59.817.6.02.766.8.72.318.6.26.951.1
Ours0.0.0.047.7.90.758.5.93.969.7.100.056.9.99.393.5.100.045.9.66.392.9Ours0.0.0.080.1.100.059.3.95.270.9.100.055.8.97.491.5.99.046.2.66.894.1
", + "bbox": [ + 94, + 143, + 879, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/014c762956b287042f5aeddaa43f90b13f9a217e31b85321e4cf8bb6c7568a5a.jpg", + "image_caption": [ + "Figure 3. Retrieval results on the ImageNet-1K dataset before and after the unlearning process in the CLIP model. We highlight the \"marmoset\" subgroup in red, with other \"monkey\" subgroups shown in green." + ], + "image_footnote": [], + "bbox": [ + 155, + 297, + 844, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "compute the mean of these ratios across all classes, denoted as the Score, to provide a comprehensive evaluation metric.", + "bbox": [ + 89, + 554, + 482, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. In-domain subgroup forgetting in ImageNet", + "text_level": 1, + "bbox": [ + 89, + 590, + 465, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Results are shown in Tab. 1. We observe that both the Fisher and LIP methods fail to optimize the model, resulting in severe catastrophic forgetting across the tested datasets. In comparison, FT, GA, EMMN, and CLIP-LIP perform slightly better but still suffer from catastrophic forgetting to varying degrees. For instance, optimizing only the retaining set with the FT strategy enables the CLIP model to forget the targeted subgroup while retaining knowledge of the intended subgroups. However, this approach leads to noticeable forgetting across other datasets, including different superclasses in ImageNet and unrelated datasets. Additionally, when applying gradient ascent to the forgetting subgroup with the GA strategy, CLIP exhibits forgetting within other subgroups of the same superclasses, and its zero-shot performance deteriorates due to catastrophic forgetting. Even with the inclusion of both forgetting and retaining samples in the fine-tuning process under the EMMN strategy, these issues remain inadequately addressed. By incorporating layer importance relative to the forgetting dataset to minimize the adverse ef", + "bbox": [ + 89, + 613, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fects of fine-tuning and leveraging a model merging strategy, we can preserve most of the model's zero-shot performance while ensuring that the targeted subgroup is forgotten.", + "bbox": [ + 511, + 554, + 906, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also present the retrieval results using our fine-tuned CLIP model, which has forgotten the target subgroup, in Fig. 3. Before unlearning, when using the prompt 'monkey' to retrieve images, the model retrieves images from various subgroups, including 'marmoset' and 'howler monkey,' among others. Additionally, when prompted with specific subclass names, as shown in the second and third rows in figure, the model retrieves corresponding subgroup images. After unlearning, we observe that images of 'marmoset' are no longer retrieved when prompted with either the superclass name 'monkey' or the specific subgroup name 'marmoset,' indicating that the knowledge of 'marmoset' has been effectively erased from CLIP's parameter space. We also include retrieval results when using the prompt 'panda'; the consistent performance between the pre- and post-unlearning versions of CLIP shows that knowledge of other classes remains largely unaffected.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9875", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4adc4b5aed144b80cc69c9adcd900ac67bd55d893118a00cce39805015a50dba.jpg", + "table_caption": [ + "Table 2. Evaluation results for forgetting the \"airplane\" (left) and \"truck\" (right) class in CIFAR-10 are presented. To assess the unaffected subgroup performance, we also separately retrieve the \"airplane\" class from ImageNet-1K (see the column labeled Target in ImageNet). The CIFAR-10 dataset is used exclusively as the retain set (see the column labeled Retain in CIFAR)." + ], + "table_footnote": [], + "table_body": "
BackboneMethodCIFAR Target (↓)Retain (↑)FoodSTLObjectNetImageNet Target AllScoreBackboneMethodCIFAR Target (↓)Retain (↑)FoodSTLObjectNetImageNet Target AllScore
RN50Original81.370.460.992.068.953.659.8-RN50Original54.073.460.992.068.952.359.8-
FT0.00.092.6100.01.11.841.745.30.20.30.00.00.30.535.4FT0.00.056.376.71.32.129.311.80.40.60.00.00.20.330.2
GA0.00.014.210.21.01.614.115.30.10.10.20.40.10.219.7GA0.00.020.217.51.22.019.321.00.30.40.00.00.10.221.6
Fisher0.00.013.118.61.01.615.216.50.30.40.00.00.10.219.6Fisher0.00.012.416.91.22.013.921.00.20.40.00.00.20.219.2
LIP0.00.015.722.31.32.114.916.20.30.40.10.20.20.320.2LIP0.00.013.918.90.91.512.313.40.40.60.10.20.20.319.3
EMMN0.00.080.7100.01.62.641.645.20.20.30.00.00.50.835.6EMMN0.00.048.566.11.32.121.223.00.20.30.00.00.10.227.4
CLIP-LIP1.11.427.238.67.111.749.945.231.215.334.564.426.844.851.1CLIP-LIP3.87.045.361.716.627.374.280.733.648.828.953.339.566.161.8
Ours7.61365.102.551.985.285.793.257.984.049.892.950.183.888.9Ours0.91.763.266.150.582.984.491.756.491.951.585.852.187.189.5
RN101Original69.264.457.392.469.251.962.3-RN101Original91.479.657.392.469.250.662.3-
FT0.00.094.6100.01.11.958.463.20.40.60.00.00.20.338.0FT0.00.058.373.21.42.440.433.70.30.40.10.20.30.531.5
GA0.00.011.117.20.91.615.016.20.30.40.00.00.10.219.4GA0.00.020.225.41.22.119.320.90.20.30.00.00.10.221.3
Fisher0.00.012.920.01.01.716.918.30.30.40.10.20.20.320.1Fisher0.00.013.116.51.01.714.215.40.30.40.10.20.10.219.2
LIP0.00.013.821.41.62.815.717.00.20.30.00.00.30.520.3LIP0.00.015.719.71.62.815.717.00.30.40.00.00.20.320.0
EMMN0.00.081.2100.01.22.156.761.40.30.40.10.20.20.337.8EMMN0.00.048.360.71.42.435.848.70.10.10.10.20.20.328.9
CLIP-LIP7.410.735.585.110.418.263.168.330.432.920.699.719.831.849.5CLIP-LIP9.610.552.866.319.534.076.082.350.773.342.584.043.670.071.3
Ours8.812.763.208.148.885.287.794.958.184.045.788.141.266.186.2Ours3.23.570.388.446.280.686.693.758.384.243.986.846.674.886.4
", + "bbox": [ + 94, + 142, + 879, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1acade394e65f2e4e936d3a22d8743ba67c0d202a90a441d755b675ffedfaeeb.jpg", + "image_caption": [ + "Figure 4. Retrieval results of the CLIP pre- and post-unlearning process." + ], + "image_footnote": [], + "bbox": [ + 135, + 300, + 859, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Cross-domain forgetting in CIFAR-10", + "text_level": 1, + "bbox": [ + 89, + 550, + 416, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then validate the effectiveness of our method in forgetting images belonging to common classes in CIFAR-10, whose concepts also appear in other datasets, such as ImageNet. The main challenge lies in the significant distributional differences between CIFAR-10 and the pre-trained dataset, which can result in severe catastrophic forgetting during the finetuning process for unlearning. In this experiment, we select 'airplane' and 'truck' as the target groups to unlearn from CLIP. In addition to evaluating the zero-shot classification performance of these targeted classes in CIFAR-10, we also assess their performance in ImageNet to determine whether unlearning is accurately confined to the specified distribution, without affecting the entire concept.", + "bbox": [ + 88, + 577, + 485, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the results in Tab. 2. Compared to the results on ImageNet, models fine-tuned on the CIFAR-10 dataset exhibit more severe catastrophic forgetting, even with methods like FT, GA, and EMMN, which retain some zero-shot capability after forgetting specific classes in ImageNet. Notably, models fine-tuned on CIFAR-10 to retain performance demonstrate strong zero-shot performance on STL-10, a dataset with a distribution similar to CIFAR-10, highlighting", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the impact of the fine-tuning dataset on model performance. In contrast, our method shows remarkable robustness under such a significant distribution shift. Specifically, while the post-unlearning CLIP achieves random classification performance on the forgotten class, it retains at least $66.1\\%$ of its original zero-shot performance.", + "bbox": [ + 511, + 551, + 906, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also retrieve related classes in CIFAR-10 and ImageNet using the pre- and post-unlearning CLIP models for the 'airplane' class. As shown in Fig. 4, we observe that the post-unlearning CLIP is unable to retrieve airplanes from the CIFAR-10 dataset but can still retrieve airplanes from the ImageNet dataset, indicating precise forgetting limited to the target distribution. The consistency of retrieved results between the pre- and post-unlearning CLIP further demonstrates the effectiveness of our method in preserving unrelated knowledge within the model.", + "bbox": [ + 511, + 641, + 908, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Style forgetting", + "text_level": 1, + "bbox": [ + 511, + 803, + 668, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate performance on forgetting a specific style within a given class, representing a more fine-grained subgroup image unlearning challenge. Specifically, we use Stable Diffusion to generate or transform existing images into three distinct styles: poster art, black-and-white drawing,", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9876", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd23cc594ac3c10f42f5bf2654b2e4b8c22697962ccb9529354d45f365c060ed.jpg", + "image_caption": [ + "Figure 5. Examples of stylized images of the sports car in the ImageNet-1K dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 478, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/37ff6f81bc7855b0a19e3f60466760a6df7aa27f0cb729d63ba2188e818e065a.jpg", + "table_caption": [ + "Table 3. Evaluation results of CLIP after forgetting sketch-style \"auto-cars\" using various methods." + ], + "table_footnote": [], + "table_body": "
MethodCIFARFoodSTLObjectNetImageNetScore
Target style (↓)Retain (↑)All (↑)
Original60.852.288.465.468.182.558.6-
FT10.116.61.93.610.511.93.96.099.5100.098.3100.01.11.920.0
GA46.776.839.876.283.294.157.888.42.02.950.361.052.188.983.2
Fisher10.517.31.52.915.918.01.11.70.10.10.50.60.20.320.1
LIP11.218.40.91.713.615.44.26.40.00.00.00.00.50.920.4
EMMN16.326.826.50.839.244.320.130.714.521.370.285.14.88.246.4
CLIP-LIP15.425.36.512.518.721.28.412.81.62.32.42.91.32.224.9
Ours49.280.943.082.480.891.458.990.12.13.166.080.051.988.687.2
", + "bbox": [ + 93, + 297, + 483, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and sketch, as illustrated in Figure 5. We first fine-tune the CLIP on images with different styles, of which the zero-shot performance is reported at the first line in Tab. 3. Then, different unlearning methods are applied to erase the model's knowledge of cars in the sketch style, while retaining images of cars in the other styles. Results are shown in Table 3.", + "bbox": [ + 88, + 412, + 483, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Among the baselines, we find that GA performs the best, effectively forgetting the targeted style while retaining most unrelated knowledge, as indicated by its strong zero-shot performance across different datasets. In comparison, simply fine-tuning CLIP on the retaining set leads to severe catastrophic forgetting, as shown in the results of FT and EMMN. The EMMN results also suggest that incorporating the retaining set into the unlearning process is not always beneficial. Although using the retaining set could potentially reduce model over-forgetting, an imprecise selection of the retaining set and an imbalanced loss between it and the forgetting set can degrade the model's generalization ability on other datasets. By selecting the optimal layer for fine-tuning to minimize the impact on the model's original performance, our method achieves the best overall results, maintaining at least $80.4\\%$ of the model's zero-shot capacity.", + "bbox": [ + 89, + 505, + 483, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation study and discussion", + "text_level": 1, + "bbox": [ + 89, + 756, + 356, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the importance of reminding stage. The reminding stage follows the forgetting stage, using the retaining set to partially recover any unintended knowledge loss. Notably, our retaining set typically differs from the original pre-training set, potentially further compromising CLIP's zero-shot classification capability. To examine this, we conduct an ablation study on the impact of iteration count during the reminding stage. Using CIFAR-10 as the unlearning", + "bbox": [ + 88, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1afa3872f578b119b5078ca98051c0640763531a1d9d7cb511a313b40d9721df.jpg", + "table_caption": [ + "Table 4. Ablation study on the used number of steps at retaining stage on the zero-shot capacity of CLIP. We set the batch size as 64 for fine-tuning and $\\alpha$ as 0.65 for model merging." + ], + "table_footnote": [], + "table_body": "
# steps51020304050100
Accr59.863.265.641.435.434.932.5
Accin44.941.238.125.88.52.10.8
", + "bbox": [ + 545, + 142, + 877, + 189 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ab435860db9fd91f7f9db7fd7831296e54afb60d858ba3438774cc87714461c3.jpg", + "table_caption": [ + "Table 5. Ablation study on the $\\alpha$ for the model merging operation at the restoring stage to the zero-shot capacity of CLIP." + ], + "table_footnote": [], + "table_body": "
α0.10.30.50.650.70.9
Accf0.00.02.48.812.868.8
Accr22.727.542.763.266.778.7
Accin1.04.422.3941.243.560.5
", + "bbox": [ + 565, + 250, + 857, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dataset for CLIP, we evaluate performance on both the retain set $(\\mathrm{ACC}_r)$ and ImageNet $(\\mathrm{ACC}_{in})$ . As shown in Tab. 4, we observe that with an increasing number of steps, classification accuracy on the retain set initially improves before declining, while accuracy on the ImageNet-1K dataset consistently decreases, with a significant drop at higher iteration counts. This suggests that while additional steps in the reminding stage help the model relearn knowledge that was unintentionally forgotten, they also lead to a larger distribution shift from the original parameter space, ultimately impairing zero-shot performance.", + "bbox": [ + 511, + 340, + 908, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the model merging. In our work, we use model merging to fully restore zero-shot capacity. While our method leverages a calibration dataset to identify the optimal coefficient, $\\alpha$ , we also conduct an additional experiment by directly varying the coefficients to evaluate both forgetting performance $(\\mathrm{ACC}_f)$ and recovery performance $(\\mathrm{ACC}_r / \\mathrm{ACC}_{in})$ . As shown in Tab. 5, achieving an optimal $\\alpha$ is crucial to balancing forgetting and retaining performance. With increasing $\\alpha$ , the model's retaining performance and zero-shot classification capacity improve continuously. However, a large $\\alpha$ can also cause the model to relearn data that should be forgotten, as indicated by the $68.8\\%$ accuracy on the forgetting set when $\\alpha$ is set to 0.9.", + "bbox": [ + 511, + 512, + 910, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 734, + 633, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study the problem of forgetting a subgroup of images in the CLIP model. We propose a three-stage method consisting of the forgetting, retaining, and restoring steps. Our method does not require any pre-trained dataset, allowing it to achieve a balanced approach by selectively forgetting the targeted subgroup of images while preserving the knowledge that should not be forgotten. Experimental results validate the effectiveness of our method and highlight the challenges of this task, as most classical methods fail.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9877", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andrew Bai, Chih-Kuan Yeh, Cho-Jui Hsieh, and Ankur Taly. Which pretrain samples to rehearse when finetuning pretrained models? arXiv preprint arXiv:2402.08096, 2024. 2", + "[2] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 5", + "[3] Alexander Becker and Thomas Liebig. Evaluating machine unlearning via epistemic uncertainty. arXiv preprint arXiv:2208.10836, 2022. 2", + "[4] Yoshua Bengio, Aaron C Courville, and Pascal Vincent. Unsupervised feature learning and deep learning: A review and new perspectives. CoRR, abs/1206.5538, 1(2665):2012, 2012. 1", + "[5] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5", + "[7] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141-159. IEEE, 2021. 1", + "[8] Zikui Cai, Yaoteng Tan, and M Salman Asif. Single layer single gradient unlearning. arXiv preprint arXiv:2407.11867, 2024.3", + "[9] Hao Chen, Bhiksha Raj, Xing Xie, and Jindong Wang. On catastrophic inheritance of large foundation models. arXiv preprint arXiv:2402.01909, 2024. 1", + "[10] Ruizhe Cheng, Bichen Wu, Peizhao Zhang, Peter Vajda, and Joseph E Gonzalez. Data-efficient language-supervised zero-shot learning with self-distillation. In CVPR, pages 3119-3124, 2021. 1", + "[11] Eli Chien, Chao Pan, and Olgica Milenkovic. Efficient model updates for approximate unlearning of graph-structured data. In The Eleventh International Conference on Learning Representations, 2022. 1", + "[12] Vikram S Chandawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security, 18:2345-2354, 2023. 2, 5", + "[13] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 5", + "[14] Lluc Clavera Comas. Machine unlearning: fisher infomation matrix and selective forgetting in deep networks. B.S. thesis, Universitat Politècnica de Catalunya, 2024. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 5", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5", + "[17] Ronen Eldan and Mark Russinovich. Who's harry potter? approximate unlearning in llms. arXiv preprint arXiv:2310.02238, 2023. 2", + "[18] Jack Foster, Kyle Fogarty, Stefan Schoepf, Cengiz Öztireli, and Alexandra Brintrup. Zero-shot machine unlearning at scale via lipschitz regularization. arXiv preprint arXiv:2402.01401, 2024. 2, 5", + "[19] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9304–9312, 2020. 2", + "[20] Jinu Gong, Osvaldo Simeone, and Joonhyuk Kang. Bayesian variational federated learning and unlearning in decentralized networks. In 2021 IEEE 22nd International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pages 216-220. IEEE, 2021. 1", + "[21] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. NeurIPS, 36, 2024. 1", + "[22] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In AAAI, pages 746-754, 2023. 1", + "[23] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. *NeurIPS*, 34:16319-16330, 2021. 1", + "[24] Kris Holt. Researchers found child abuse material in the largest ai image generation dataset, 2024. Accessed: 2024-10-25. 3", + "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4", + "[26] Mark He Huang, Lin Geng Foo, and Jun Liu. Learning to unlearn for robust machine unlearning. arXiv preprint arXiv:2407.10494, 2024. 2", + "[27] Zhehao Huang, Xinwen Cheng, JingHao Zheng, Haoran Wang, Zhengbao He, Tao Li, and Xiaolin Huang. Unified gradient-based machine unlearning with remain geometry enhancement. arXiv preprint arXiv:2409.19732, 2024. 1", + "[28] Frank Klinker. Exponential moving average versus moving exponential average. Mathematische Semesterberichte, 58: 97-107, 2011. 5", + "[29] Alexey Kravets and Vinay P. Namboodiri. Zero-shot CLIP class forgetting via text-image space adaptation. Transactions on Machine Learning Research, 2025. 2, 3, 5" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9878", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[31] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In CVPR, pages 20071-20082, 2023. 1", + "[32] Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, PRANAY SHARMA, Sijia Liu, et al. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 5", + "[33] Ji Liu, Jiaxiang Ren, Ruoming Jin, Zijie Zhang, Yang Zhou, Patrick Valduriez, and Dejing Dou. Fisher information-based efficient curriculum federated learning with large language models. arXiv preprint arXiv:2410.00131, 2024. 3", + "[34] Yufang Liu, Changzhi Sun, Yuanbin Wu, and Aimin Zhou. Unlearning with fisher masking. arXiv preprint arXiv:2310.05331, 2023. 3", + "[35] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Xi-angyang Ji, Antoni Chan, and Rong Jin. Improved fine-tuning by better leveraging pre-training data. Advances in Neural Information Processing Systems, 35:32568-32581, 2022. 2", + "[36] Alexander Ly, Maarten Marsman, Josine Verhagen, Raoul PPP Grasman, and Eric-Jan Wagenmakers. A tutorial on fisher information. Journal of Mathematical Psychology, 80:40-55, 2017. 3", + "[37] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N Ravi. Deep unlearning via randomized conditionally independent hessian. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10422-10431, 2022. 1", + "[38] Quoc Phong Nguyen, Bryan Kian Hsiang Low, and Patrick Jaillet. Variational bayesian unlearning. Advances in Neural Information Processing Systems, 33:16025-16036, 2020. 1", + "[39] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. *Neural networks*, 113:54–71, 2019. 2", + "[40] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara, et al. Safe-clip: Removing nsfw concepts from vision-and-language models. In Proceedings of the European Conference on Computer Vision, 2024. 3", + "[41] Clifton Poth, Jonas Pfeiffer, Andreas Rückle, and Iryna Gurevych. What to pre-train on? efficient intermediate task selection. arXiv preprint arXiv:2104.08247, 2021. 2", + "[42] Jenny Quang. Does training ai violate copyright law? Berkeley Tech. LJ, 36:1407, 2021. 1", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3", + "[44] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, pages 2765-2775, 2023. 1" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Shibani Santurkar, Dimitris Tsipras, and Aleksander Madry. Breeds: Benchmarks for subpopulation shift. arXiv preprint arXiv:2008.04859, 2020. 5", + "[46] Sebastian Schelter, Mozhdeh Ariannezhad, and Maarten de Rijke. Forget me now: Fast and exact unlearning in neighborhood-based recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2011–2015, 2023. 1", + "[47] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 1, 2, 3", + "[48] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 34:18075-18086, 2021. 1", + "[49] Adam Shostack. The boy who survived: Removing harry potter from an llm is harder than reported. arXiv preprint arXiv:2403.12082, 2024. 2", + "[50] Shashank Subramanian, Peter Harrington, Kurt Keutzer, Wahid Bhimji, Dmitriy Morozov, Michael W Mahoney, and Amir Gholami. Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. NeurIPS, 36, 2024. 1", + "[51] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[52] Anvith Thudi, Gabriel Deza, Varun Chandrasekaran, and Nicolas Papernot. Unrolling sgd: Understanding factors influencing machine unlearning. In 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P), pages 303-319. IEEE, 2022. 2, 5", + "[53] Enayat Ullah, Tung Mai, Anup Rao, Ryan A Rossi, and Raman Arora. Machine unlearning via algorithmic stability. In Conference on Learning Theory, pages 4126-4142. PMLR, 2021. 1", + "[54] Alexander Warnecke, Lukas Pirch, Christian Wressnegger, and Konrad Rieck. Machine unlearning of features and labels. In 30th Annual Network and Distributed System Security Symposium, NDSS 2023, San Diego, California, USA, February 27 - March 3, 2023. The Internet Society, 2023. 2, 5", + "[55] Yixuan Wei, Han Hu, Zhenda Xie, Ze Liu, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Improving clip fine-tuning performance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5439-5449, 2023. 2", + "[56] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 5" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9879", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Tomoya Yamashita, Masanori Yamada, and Takashi Shibata. One-shot machine unlearning with mnemonic code. In The 16th Asian Conference on Machine Learning (Conference Track). 3", + "[58] Haonan Yan, Xiaoguang Li, Ziyao Guo, Hui Li, Fenghua Li, and Xiaodong Lin. Arcane: An efficient architecture for exact machine unlearning. In *IJCAI*, page 19, 2022. 1", + "[59] Kaiyu Yang, Clint Qinami, Li Fei-Fei, Jia Deng, and Olga Russakovsky. Towards fairer datasets: Filtering and balancing the distribution of the people subtree in the imagenet hierarchy. In Proceedings of the 2020 conference on fairness, accountability, and transparency, pages 547–558, 2020. 1", + "[60] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. NeurIPS, 36, 2024. 4", + "[61] Binchi Zhang, Yushun Dong, Tianhao Wang, and Jundong Li. Towards certified unlearning for deep neural networks. arXiv preprint arXiv:2408.00920, 2024. 1", + "[62] Jiahao Zhang. Graph unlearning with efficient partial retraining. In Companion Proceedings of the ACM on Web Conference 2024, pages 1218-1221, 2024. 1", + "[63] Zeliang Zhang, Mingqian Feng, Zhiheng Li, and Chenliang Xu. Discover and mitigate multiple biased subgroups in image classifiers. In CVPR, 2024. 1", + "[64] Zeliang Zhang, Zhuo Liu, Mingqian Feng, and Chenliang Xu. Can clip count stars? an empirical study on quantity bias in clip. In EMNLP (Findings), 2024. 1" + ], + "bbox": [ + 91, + 92, + 485, + 474 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9880", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Targeted Forgetting of Image Subgroups in CLIP Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_model.json b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2354a8aa022b8db6bad3af5d795a394a0c135a97 --- /dev/null +++ b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Targeted Forgetting of Image Subgroups in CLIP Models" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.18, + 0.884, + 0.217 + ], + "angle": 0, + "content": "Zeliang Zhang\\(^{1}\\) Gaowen Liu\\(^{2}\\) Charles Fleming\\(^{2}\\) Ramana Rao Kompella\\(^{2}\\) Chenliang Xu\\(^{1}\\) \n\\(^{1}\\)University of Rochester \\(^{2}\\)Cisco Research" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.219, + 0.862, + 0.235 + ], + "angle": 0, + "content": "{zeliang.zhang, chenliang.xu}@rochester.edu, {gaoliu, chflemin, rkompell}@cisco.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.725 + ], + "angle": 0, + "content": "Foundation models (FMs) such as CLIP have demonstrated impressive zero-shot performance across various tasks by leveraging large-scale, unsupervised pre-training. However, they often inherit harmful or unwanted knowledge from noisy internet-sourced datasets, compromising their reliability in real-world applications. Existing model unlearning methods either rely on access to pre-trained datasets or focus on coarse-grained unlearning (e.g., entire classes), leaving a critical gap for fine-grained unlearning. In this paper, we address the challenging scenario of selectively forgetting specific portions of knowledge within a class—without access to pre-trained data—while preserving the model's overall performance. We propose a novel three-stage approach that progressively unlearns targeted knowledge while mitigating over-forgetting. It consists of (1) a forgetting stage to fine-tune the CLIP on samples to be forgotten, (2) a reminding stage to restore performance on retained samples, and (3) a restoring stage to recover zero-shot capabilities using model souping. Additionally, we introduce knowledge distillation to handle the distribution disparity between forgetting/retaining samples and unseen pre-trained data. Extensive experiments on CIFAR-10, ImageNet-1K, and style datasets demonstrate that our approach effectively unlearns specific subgroups while maintaining strong zero-shot performance on semantically similar subgroups and other categories, significantly outperforming baseline unlearning methods, which lose effectiveness under the CLIP unlearning setting." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.755, + 0.222, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Foundation models (FMs) are deep neural networks pretrained on large-scale datasets [50]. By employing unsupervised training, these models capture complex patterns and relationships within the data, enabling a deep understanding of various domains [4]. For example, CLIP [43], a vision-language model, learns correspondences between natural language and images. Trained on the large-scale LAION-5B dataset [47], CLIP demonstrates strong zero" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.268, + 0.907, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.462, + 0.909, + 0.56 + ], + "angle": 0, + "content": "Figure 1. In the task of CLIP unlearning, classical machine unlearning methods often fail to generalize across multiple datasets due to limited access to the original pre-training data. In contrast, our proposed method does not require access to the pre-training dataset. After unlearning, CLIP not only forgets the specified subgroup of images but also maintains consistent performance on various unseen datasets in the zero-shot classification setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.909, + 0.732 + ], + "angle": 0, + "content": "shot performance [10, 22, 44]. However, its reliability in real-world applications is questionable [63, 64]. The training data, sourced from the internet, includes uncensored content such as discriminatory and violent imagery, copyright violations, and personal information [21, 42, 59]. This problematic content can be unintentionally learned during pre-training, raising concerns about the reliability and safety of downstream models [9, 31]. Therefore, it is critical to remove such harmful knowledge to ensure trustworthy FMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Numerous studies have explored model unlearning, which can be broadly divided into exact [46, 53, 58] and approximate strategies [11, 23, 38]. Exact unlearning aims to completely eliminate the influence of forgotten data from the model, often by retraining it [7, 58, 62], ensuring the model behaves as if the data had never been included in training. While thorough, this approach is computationally expensive and impractical for large-scale models [32, 48]. In contrast, approximate unlearning offers a more efficient alternative, using techniques like variational inference [20, 38] or adaptive regularization [27, 37, 61] to partially reduce the data's" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9870" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "impact. However, these methods face several challenges when applied to FMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.124, + 0.483, + 0.32 + ], + "angle": 0, + "content": "First, as shown in Fig. 1, the training data for FMs is often vast or inaccessible [5, 47], rendering exact unlearning impractical. Second, while FMs are trained on large-scale datasets using unsupervised learning, the fine-tuning typically used in approximate unlearning strategies [26, 51] can degrade generalization across other knowledge domains [1, 35, 41], leading to significant performance drops, particularly in zero-shot tasks [55]. Third, the knowledge to be forgotten is often deeply entangled with other learned information [17, 49]. This entanglement makes it difficult to isolate and remove specific knowledge from the model parameters, particularly when distributional disparities exist between the data we want to unlearn and the pre-trained data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.483, + 0.563 + ], + "angle": 0, + "content": "Few studies have specifically addressed the challenge of unlearning in foundational models, particularly in CLIP. Most existing approaches focus on unlearning entire classes or broad concepts, yet real-world applications often require a more targeted approach, where only certain portions of knowledge within a class are forgotten. For example, we may want the model to forget representations of Disney's Mickey Mouse while retaining versions created by the public, a task known as subgroup image forgetting. Moreover, many current unlearning methods rely on access to the original pre-trained dataset to restore model performance, which is often unavailable. This limitation underscores a critical question in current research: how can models like CLIP achieve selective, fine-grained unlearning without relying on pre-trained data, while still maintaining generalization across other tasks?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.778 + ], + "angle": 0, + "content": "In this work, we bridge this gap by tackling a more challenging and practical scenario: enabling CLIP to selectively forget specific portions of knowledge within a single class without access to pre-trained data, framing this as a fine-grained unlearning task. Rather than unlearning an entire class or concept (e.g., all images of airplanes), we focus on selectively forgetting images from a particular subset (e.g., Boeing airplanes) while preserving the model's knowledge of other subsets (e.g., different manufacturers). This selective forgetting is particularly challenging due to the feature entanglement across subsets within a class, and it is exceedingly difficult to define and collect all subsets within the same class in a foundational model setting, which is trained on vast, large-scale data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the challenge of partial knowledge removal, we propose a novel, three-stage approach: forgetting, reminding, and restoring. In the forgetting stage, we identify layers essential for representing the forgetting samples but less critical to other data, fine-tuning the model on these samples to selectively remove targeted knowledge. We observe that models can over- forget samples from similar distributions, so we introduce a retaining set with a similar distribution to the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.302 + ], + "angle": 0, + "content": "targeted subgroup, aiding layer identification and supporting the reminding stage. During reminding, we fine-tune the model on the retaining samples to reinforce CLIP's performance on data prone to over-forgetting due to distributional similarity. In the final restore stage, we employ model souping to fully recover CLIP's zero-shot performance. Throughout this process, we address the distributional gap between forgetting/retaining samples and unseen pre-trained data by distilling knowledge from CLIP, achieving a more precise representation without relying on the original pre-trained dataset. Extensive experiments demonstrate our method's effectiveness in selectively unlearning subgroup images while preserving robust zero-shot performance across diverse distributions and datasets." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.304, + 0.825, + 0.317 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.318, + 0.904, + 0.347 + ], + "angle": 0, + "content": "1. We formulate the problem of subgroup image unlearning within the CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.349, + 0.905, + 0.422 + ], + "angle": 0, + "content": "2. We propose a three-stage method for unlearning in CLIP, consisting of forgetting, reminding, and restoring. Our method is data-efficient, which does not require access to CLIP's pre-training dataset, making it more practical for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.424, + 0.905, + 0.512 + ], + "angle": 0, + "content": "3. We conduct extensive experiments to validate the effectiveness of our proposed method, including unlearning tasks on CIFAR-10, ImageNet-1K, and style forgetting. Additionally, we demonstrate the scalability of our approach across various applications through retrieval-based evaluations." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.318, + 0.905, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.529, + 0.651, + 0.544 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Machine Unlearning. Various strategies have been developed to facilitate machine unlearning in models. Notably, Golatkar et al. [19], Warnecke et al. [54] employ a method of continuous fine-tuning (FT) on the retain dataset—those data elements the model is expected to remember. This approach not only approximates the forgetting of the target dataset but also mitigates the issue of overforgetting [32, 39], in contrast to exclusively fine-tuning on the forgetting dataset. Becker and Liebig [3], Liu et al. [32] leverage Fisher information, which assesses the sensitivity of parameters concerning the forgetting dataset, to achieve model unlearning. Meanwhile, Thudi et al. [52] unroll the knowledge embedded in the forgetting dataset from the model parameters using stochastic gradient ascent (GA). Further, [18] introduces the use of local Lipschitz (LIP) perturbations on the forgetting dataset as a technique for unlearning in image classification models, which Kravets and Namboodiri [29] subsequently adapts to the CLIP model. Additionally, Chandawat et al. [12] proposes an error minimization-maximization (EMMN) framework that aims to maximize errors on the forgetting dataset while minimizing them on the retraining dataset, thus achieving a balance between forgetting and retaining performance. Among these methods, EMMN [12] and LIP [29] stand out" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.513, + 0.956 + ], + "angle": 0, + "content": "9871" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "as they do not require access to the training datasets (both forgetting and retaining), unlike other approaches." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.123, + 0.487, + 0.44 + ], + "angle": 0, + "content": "On the Importance of Unlearning in CLIP. Trained on a billion-scale dataset of image-text pairs [47], the CLIP model [43] excels in zero-shot classification tasks. Despite its strengths, the presence of uncensored samples within the training data has raised practical concerns [24]. Given the impracticality of retraining the CLIP model from scratch due to significant resource demands and the inaccessibility of the large-scale dataset, it is imperative to explore machine unlearning methods to selectively erase problematic knowledge from the model. Poppi et al. [40] suggest an approach to first distinguish between safe and unsafe content, followed by fine-tuning CLIP to alter the embeddings of unsafe content. Meanwhile, Cai et al. [8] focuses on identifying and unlearning from the most influential layer in CLIP, using a subset of the training set to restore zero-shot capabilities. [29] expand zero-shot machine unlearning techniques to the CLIP model, enabling the forgetting of entire concepts as indicated by prompts. Concurrently, Kravets and Namboodiri [29] apply Layer-wise Relevance Propagation (LoRA) to the text encoder of CLIP, specifically fine-tuning this component to unlearn a given concept within the model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.452, + 0.299, + 0.468 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.485, + 0.551 + ], + "angle": 0, + "content": "The CLIP model, denoted \\( g(\\cdot, \\cdot) = \\{g^{\\mathrm{img}}(\\cdot), g^{\\mathrm{txt}}(\\cdot)\\} \\), processes image and text inputs into embeddings, \\( e^{\\mathrm{img}} \\) and \\( e^{\\mathrm{txt}} \\). It aims for high cosine similarity between these embeddings when they are correctly paired, and low similarity otherwise, using a threshold \\( \\delta \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.72 + ], + "angle": 0, + "content": "Suppose a large dataset \\( D = \\{x_{n}^{\\mathrm{img}}, x_{n}^{\\mathrm{txt}}, c_{n}, g_{n}\\}_{n=1}^{N} \\), where we denote \\( x^{\\mathrm{img}} \\) as the image, \\( x^{txt} \\) as the paired text learned by the CLIP, \\( c \\) as the coarse main category, and \\( g \\) is the fine-grained subclass label. The forgetting dataset \\( D^{f} = \\{x_{n}^{\\mathrm{img}}, x_{n}^{\\mathrm{txt}}, c_{n}, g_{n}\\}_{n=1}^{N_{f}} \\) is usually a small set compared with \\( D \\). The goal is to unlearn the associations between \\( x^{\\mathrm{img}} \\) and \\( x^{txt} \\) on \\( D^{f} \\) without impacting the model's performance on the remaining dataset \\( D^{r} = D / D_{f} \\), namely the retain set. The task involves selectively erasing knowledge from \\( D^{f} \\) while maintaining performance on \\( D^{r} \\), which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.728, + 0.484, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\forall \\left(x _ {i} ^ {\\mathrm {i m g}}, x ^ {\\mathrm {t x t}}\\right) \\in D ^ {f}, \\frac {\\hat {g} ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot \\hat {g} ^ {\\mathrm {t x t}} \\left(x ^ {\\mathrm {t x t}}\\right)}{\\| \\hat {g} ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot \\hat {g} ^ {\\mathrm {t x t}} \\left(x ^ {\\mathrm {t x t}}\\right) \\|} < \\delta , \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.767, + 0.44, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\forall (x _ {i} ^ {\\mathrm {i m g}}, x _ {i} ^ {\\mathrm {t x t}}) \\in D ^ {r}, \\hat {g} (x _ {i} ^ {\\mathrm {i m g}}, x _ {i} ^ {\\mathrm {t x t}}) \\approx g (x _ {i} ^ {\\mathrm {i m g}}, x _ {i} ^ {\\mathrm {t x t}}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.825 + ], + "angle": 0, + "content": "where \\(\\hat{g}\\) represents the unlearned CLIP model and \\(g\\) the original model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.855 + ], + "angle": 0, + "content": "Implementing the approach outlined in Eq. (1) is challenging due to several factors:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "1. The large dataset \\( D \\) used for training \\( g \\) is inaccessible, and similarly, \\( D_r \\) is also unavailable. We can only access a small subset of \\( D_r \\) to recover model performance" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "after forgetting, blocking the direct use of conventional machine unlearning methods." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.196 + ], + "angle": 0, + "content": "2. In addition, we may only have access to a coarse label \\( c \\) as the surrogate of \\( x^{txt} \\) rather than a fine-grained subgroup label \\( g \\). This limitation increases the risk of inaccurately erasing targeted knowledge and over-forgetting unrelated knowledge that is similar to the target." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.907, + 0.257 + ], + "angle": 0, + "content": "3. There may be a substantial distribution shift between \\( D_{f} \\) and \\( D \\). With a large number of parameters, directly finetuning on the small dataset \\( D_{f} \\) could cause the model to overfit and lose generalization ability." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.122, + 0.907, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Thus, the goal of this task is twofold: it involves selectively forgetting only specific image subgroups because of coarse labeling, while simultaneously preserving the model's robust zero-shot capabilities across various datasets, including those characterized by similarly coarse labels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.346, + 0.649, + 0.364 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.418 + ], + "angle": 0, + "content": "To address the challenge of subgroup image unlearning, we propose a novel methodology that comprises three stages: forgetting, reminding, and restoring, as depicted in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.426, + 0.633, + 0.442 + ], + "angle": 0, + "content": "4.1. Forgetting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.689 + ], + "angle": 0, + "content": "Previous studies [14, 34] indicate that selectively fine-tuning critical layers for a target dataset can effectively mitigate over-forgetting [33], typically using Fisher information [36, 57] to assess the importance of each layer. However, while Fisher information measures parameter sensitivity to the target dataset, it overlooks potential impacts on other datasets, which are inaccessible in our task. Our experiments reveal that relying solely on the forgetting dataset to modify the CLIP model often leads to catastrophic over-forgetting on similar subgroups with distributions akin to the forgetting data, as shown in Tab. 1 for the CLIP model optimized by GA. To address this, we propose using relative Fisher information which combines both the forgetting and retaining samples as a more nuanced metric for identifying essential layers for fine-tuning, thereby enhancing the precision of the unlearning process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.909, + 0.844 + ], + "angle": 0, + "content": "Concretely, given the forgetting dataset \\(\\mathcal{D}^f = \\{(x_i^{\\mathrm{img}}, x^{\\mathrm{txt}})\\}\\), we manually construct a retain dataset \\(\\mathcal{D}^r = \\{(x_i^{\\mathrm{img}}, x_i^{\\mathrm{txt}})\\}\\) with a distribution similar to \\(\\mathcal{D}^f\\), used for alleviating the aforementioned subgroup over-forgetting. Here, \\(\\mathcal{D}^r\\) is comprised of similar subgroups of images that we aim to preserve in the model's memory. In practical scenarios, we often use coarse labels as surrogate image descriptions, so \\(x^{\\mathrm{txt}}\\) from the forgetting dataset may also appear in \\(\\mathcal{D}^r\\). We calculate the relative Fisher information for the \\(l\\)-th layer in \\(g^{\\mathrm{img}}(\\cdot)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.853, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} ^ {l} = \\frac {\\mathbb {E} _ {\\left(x _ {i} ^ {\\mathrm {i m g}} , x ^ {\\mathrm {t x t}}\\right) \\sim \\mathcal {D} ^ {f}} \\left[ \\nabla_ {\\theta^ {l}} ^ {2} \\left(g ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot g ^ {\\mathrm {t x t}} \\left(x ^ {\\mathrm {t x t}}\\right)\\right) \\right]}{\\mathbb {E} _ {\\left(x _ {i} ^ {\\mathrm {i m g}} , x _ {i} ^ {\\mathrm {t x t}}\\right) \\sim \\mathcal {D} ^ {r}} \\left[ \\nabla_ {\\theta^ {l}} ^ {2} \\left(g ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot g ^ {\\mathrm {t x t}} \\left(x _ {i} ^ {\\mathrm {t x t}}\\right)\\right) \\right]}, \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.089, + 0.825, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.313, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Overview of our method to unlearn the CLIP on a subgroup of images. We first compute the relative fisher information on the forgetting dataset and retain dataset to evaluate the layer importance. We select important layers for the following fine-tuning process. During the fine-tuning process for knowledge reminding, we align the distribution of our retain dataset to that of the pre-training dataset learned by the model. A model merging operation is performed at the last step of knowledge restoration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.431, + 0.411 + ], + "angle": 0, + "content": "where \\(\\theta^l\\) represents the parameters of the \\(l\\)-th layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.547 + ], + "angle": 0, + "content": "We then rank the importance of different layers based on the computed relative Fisher information, as detailed in Eq. (2). Layers with higher relative importance are selected for targeted unlearning, with all other layers frozen to minimize the impact on non-target knowledge. In our experiments, applying the low-rank adaptation (LoRA) technique [25] to the selected layers significantly enhances the targeted unlearning performance while minimizing effects on other retained knowledge." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.548, + 0.484, + 0.579 + ], + "angle": 0, + "content": "Finally, we fine-tune our model on the forgetting dataset \\(\\mathcal{D}_f\\) to minimize the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.591, + 0.484, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\min \\mathcal {L} _ {f} = \\sum_ {i} \\frac {g ^ {i m g} \\left(x _ {i} ^ {i m g}\\right) \\cdot g ^ {t x t} \\left(x ^ {t x t}\\right)}{\\| g ^ {i m g} \\left(x _ {i} ^ {i m g}\\right) \\cdot g ^ {t x t} \\left(x ^ {t x t}\\right) \\|}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.67 + ], + "angle": 0, + "content": "where images in the forgetting dataset are associated through the same coarse label \\( x^{txt} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.822 + ], + "angle": 0, + "content": "By fine-tuning the CLIP model on images marked for unlearning with coarse labels, we aim to selectively erase specific knowledge. However, this adjustment often leads to catastrophic overforgetting, affecting the model's ability across both similar and different image distributions. For example, attempting to forget the \"airplane\" class in CIFAR-10 may also affect the recognition of other classes within CIFAR-10 and in ImageNet-1K. To tackle these challenges, particularly when full access to the pre-training dataset is unavailable, we implement reminding and restoration stages." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.214, + 0.85 + ], + "angle": 0, + "content": "4.2. Reminding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "An intuitive strategy to retain non-targeted knowledge in CLIP is to fine-tune it on our retaining set, built during the forgetting stage for computing \\(\\mathcal{I}\\). However, two challenges" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.485 + ], + "angle": 0, + "content": "hinder the direct application of this fine-tuning step. First, our retaining set may have a different distribution than the pre-training dataset, potentially causing inefficient learning. Second, our retaining set is typically limited, and fine-tuning on it could lead to further catastrophic over-forgetting in various unseen datasets and tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.909, + 0.61 + ], + "angle": 0, + "content": "To address the first challenge, inspired by research showing that training data features are encoded in model parameters [60], we employ an optimization-based strategy to align the retaining dataset's distribution with the pre-training dataset. Specifically, for a given batch of image-text pairs in the retaining dataset \\(\\mathcal{D}_r = \\{(x_i^{\\mathrm{img}}, x_i^{\\mathrm{xt}})\\}_{i=1}^B\\), we introduce a batch of perturbations \\(\\delta_{i_{i=1}}^{B}\\) on the images. These perturbations are optimized by minimizing the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.62, + 0.907, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {a} = \\sum_ {l} \\left\\| \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\mu_ {l} ^ {\\text {i m g}} \\left(x _ {i} ^ {\\text {i m g}} + \\delta_ {i}\\right) - \\mathrm {B N} _ {l} ^ {\\mu} \\right\\| \\tag {4} \\\\ + \\sum_ {l} \\left\\| \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sigma_ {l} ^ {i m g} \\left(x _ {i} ^ {i m g} + \\delta_ {i}\\right) - \\mathbf {B N} _ {l} ^ {\\sigma} \\right\\|, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mu_l^{img}(\\cdot)\\) and \\(\\sigma_l^{img}\\) are the mean and variance of the batch of intermediate features before the \\(l\\)-th BatchNormalization (BN) layer, computed by the image encoder of original CLIP. \\(\\mathrm{BN}_l^\\mu\\) and \\(\\mathrm{BN}_l^\\sigma\\) are the mean and variance parameters of the \\(l\\)-th BN layer, encoding the global distribution information of the pre-training dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We then fine-tune the CLIP model on the retaining dataset, now aligned with the pre-training dataset's distribution. This serves as a reminder of the knowledge encoded in the original model, effectively achieved without direct access to the pre-training dataset. This strategy helps prevent further catastrophic forgetting that could arise from distributional" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9873" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "disparities between the retaining dataset and the original training dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.484, + 0.198 + ], + "angle": 0, + "content": "To address the second challenge of overfitting during fine-tuning due to the large amount of parameters, we propose minimizing the impact on the original model parameters by using an exponential moving average (EMA) [28] for the final model parameters, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.215, + 0.484, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\theta^ {\\mathrm {e m a}} = \\alpha \\cdot \\theta^ {\\mathrm {e m a}} + (1 - \\alpha) \\cdot \\theta , \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.485, + 0.302 + ], + "angle": 0, + "content": "where \\(\\theta^{\\mathrm{ema}}\\) is initialized as the original CLIP parameters \\(\\theta_0\\), \\(\\theta\\) represents the continuously optimized parameters from \\(\\theta_0\\), and \\(\\alpha\\) is the decay factor. This approach effectively balances parameter retention with necessary updates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.313, + 0.219, + 0.327 + ], + "angle": 0, + "content": "4.3. Restoration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.41 + ], + "angle": 0, + "content": "In addition to fine-tuning on the retain dataset for reminding, we also employ a model merging strategy to restore the zero-shot performance of CLIP, a technique empirically shown to locate flat optima on the loss landscape, enhancing generalization [56]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.486 + ], + "angle": 0, + "content": "Specifically, we first compile a small calibration dataset \\(\\mathcal{D}_m = \\{(x_i,y_i)\\}_{i = 1}^N\\) for zero-shot image classification, which may include portions of the retain dataset not used in fine-tuning. We then execute model merging with varying coefficients as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.498, + 0.484, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {\\theta} {\\arg \\max } \\operatorname {A c c} (g, \\mathcal {D} _ {m}), \\tag {6} \\\\ s. t. \\theta = \\alpha \\cdot \\theta^ {f} + (1 - \\alpha) \\cdot \\theta^ {o r i}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.613 + ], + "angle": 0, + "content": "where \\( Acc(g, \\mathcal{D}_m) \\) assesses the zero-shot classification performance of model \\( g \\) on \\( \\mathcal{D}_m \\), and \\( \\theta^f \\) and \\( \\theta^{ori} \\) represent the parameters of the CLIP model post-unlearning and the original model, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.629, + 0.215, + 0.644 + ], + "angle": 0, + "content": "5. Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.76 + ], + "angle": 0, + "content": "We begin by focusing on the forgetting subgroups within the ImageNet dataset, where the images are semantically grouped by Breeds. We select two superclasses: 'monkey' and 'turtle.' Each superclass contains four distinct subgroups. We specifically target one subgroup in each superclass for unlearning: 'marmoset' in the monkey superclass and 'box turtle' in the turtle superclass." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.279, + 0.789 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Settings. We conduct experiments in three distinct settings: in-domain forgetting, near-domain forgetting, and style forgetting, as detailed below:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- In-domain forgetting: Experiments are conducted on the Breeds dataset [45], a subset of the ImageNet-1K dataset [16] consisting of 4 subgroups within 17 superclasses. The images, with dimensions of \\(224 \\times 224\\), closely" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.908, + 0.152 + ], + "angle": 0, + "content": "match the distribution of images in the pre-trained LAION-5B dataset used for CLIP. For each superclass, one subgroup is selected for unlearning, while the images from the remaining three subgroups form the retaining set." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.907, + 0.272 + ], + "angle": 0, + "content": "- Out-of-domain forgetting: Experiments are performed on the CIFAR-10 dataset [30], which differs significantly from the image distribution in LAION-5B. To align with CLIP's input format, we resize the images from \\(32 \\times 32\\) to \\(224 \\times 224\\). In this setting, we select a class shared between CIFAR-10 and ImageNet-1K. The images of this selected class form the forgetting dataset, while the remaining classes in CIFAR-10 constitute the retaining set." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.273, + 0.906, + 0.363 + ], + "angle": 0, + "content": "- Style forgetting: Experiments are conducted on one class in the ImageNet-1K dataset [16] transformed by Stable Diffusion [13]. The images are randomly transformed into three styles: sketch style, poster art, and black-and-white drawing. For a given style to forget, images in the remaining styles are treated as the retaining set." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.153, + 0.907, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.908, + 0.595 + ], + "angle": 0, + "content": "Baselines and implementations. We select six methods as our baselines: FT [54], GA [52], Fisher [32], LIP [18], and EMMN [12], and CLIP-LIP [29]. Except the CLIP-LIP, these methods are developed for conventional models that require a classification head. To address this difference, we use the probability from the zero-shot classification task as a surrogate for the classification logits. In these approaches, the text prompt is provided by the superclass names in both the forgetting and retain sets. For all learning-based methods, recognizing the sensitivity of CLIP parameters during the fine-tuning process, we use the Adam optimizer with a small learning rate of \\( 1 \\times 10^{-6} \\) for up to 2 epochs. For LIP, we set the number of noisy copies to 10. Additionally, the variance scaling factor \\( \\alpha \\) in Fisher is set to 0.2. We provide the full implementation in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.901 + ], + "angle": 0, + "content": "**Evaluations.** The performance of image subgroup unlearning in CLIP is evaluated along three dimensions. First, we must ensure that CLIP successfully forgets the images from the specified subgroup. Second, CLIP must continue to perform consistently on the retain set, which includes the other subgroups that are not intended to be forgotten. Third, CLIP must maintain strong zero-shot performance across various tasks and datasets. These considerations lead us to evaluate CLIP's performance not only on the targeted forgetting subgroups and unaffected subgroups, but also on various unseen datasets during the forgetting process, including Food [6], STL [15], and ObjectNet [2]. Additionally, we will use the post-unlearning CLIP to retrieve images in order to assess the impact of forgetting on its learned representations. Additionally, we use the classification accuracy ratio between the original and post-unlearned CLIP models to gauge the degree of knowledge forgetting and retention. More specifically, the ratio is computed as \\(\\min\\left(\\frac{\\text{Acc\\_learn}}{\\text{Acc}_{ori}}, 1\\right)\\), where we cap the score at 1 to represent full knowledge restoration. This ratio is shown in the subscript of each result. We further" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9874" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 1. Evaluation results for forgetting the subclass 'marmoset' within the superclass 'monkey' (left) and 'box turtle' within the superclass 'turtle' (right) in the ImageNet-1K dataset are presented (Target). Each superclass contains four subgroups, with the remaining three designated as the retain set (Retain)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.144, + 0.88, + 0.287 + ], + "angle": 0, + "content": "
BackboneMethodImageNetCIFARFoodSTLObjectNetScoreBackboneMethodImageNetAll (↑)CIFARFoodSTLObjectNetScore
Target (↓)Retain (↑)All (↑)Target (↓)Retain (↑)
RN50Original51.054.759.870.460.992.068.9-RN50Original64.570.359.870.460.992.068.9-
FT2.5.4.987.1.100.025.6.42.813.5.19.217.4.28.543.1.46.819.6.58.451.6FT0.0.0.072.1.100.026.0.43.512.6.17.928.6.47.070.3.76.420.2.93.359.2
GA0.0.0.00.9.1.632.2.53.816.4.23.322.9.37.663.3.68.822.1.32.145.3GA0.0.0.013.3.18.942.1.70.414.6.20.740.0.55.777.8.84.621.3.30.955.8
Fisher0.2.0.30.5.0.91.3.2.110.4.14.70.2.0.310.5.11.43.0.14.419.1Fisher1.2.1.92.0.1.83.5.5.99.7.13.80.4.0.710.5.14.11.6.23.319.3
LIP0.7.1.40.2.3.61.3.2.210.7.15.20.2.0.310.3.11.21.6.2.318.6LIP0.4.0.63.4.4.815.2.25.410.6.15.16.1.10.023.7.25.811.3.16.428.1
EMMN0.0.0.056.7.100.028.4.47.513.1.18.620.4.33.554.9.59.720.5.29.755.6EMMN0.1.0.257.2.81.428.1.47.012.5.17.822.1.36.364.7.70.317.0.25.554.0
CLIP-LIP2.65.10.4.0.717.7.29.618.8.26.74.2.6.950.8.55.212.5.18.133.2CLIP-LIP1.9.2.92.0.2.831.9.53.319.5.27.76.9.11.363.6.69.134.2.69.644.4
Ours0.0.0.050.2.91.754.5.91.185.7.100.062.9.100.081.5.88.645.2.56.691.0Ours0.0.0.069.4.98.750.6.84.654.5.77.450.5.82.987.6.95.243.1.62.685.9
RN101Original66.252.662.364.457.392.469.2-RN101Original66.277.162.364.457.392.469.2-
FT1.3.2.075.2.100.027.6.44.311.8.18.34.4.7.635.1.48.017.3.25.047.3FT0.0.0.068.0.88.232.4.52.011.4.17.727.8.48.566.0.71.423.5.34.058.8
GA0.7.1.10.3.0.536.2.58.119.9.9.929.3.51.158.6.63.419.4.88.047.3GA0.0.0.022.1.28.746.4.74.512.2.18.940.4.70.575.6.81.825.9.37.458.8
Fisher0.9.1.40.4.0.70.3.0.511.7.18.20.2.0.39.6.10.31.0.1.418.6Fisher1.9.2.92.4.3.10.9.1.411.0.17.10.3.0.59.5.10.30.6.0.918.6
LIP0.2.0.30.3.0.52.0.3.210.4.16.10.2.0.310.7.11.61.9.2.719.2LIP0.81.23.2.4.212.9.20.710.3.16.08.4.14.716.9.18.314.7.21.227.7
EMMN0.7.1.162.5.100.030.1.48.322.0.35.110.8.18.810.0.10.823.8.34.449.5EMMN0.0.0.048.6.83.032.5.52.29.2.14.322.4.39.160.8.65.822.9.33.152.5
CLIP-LIP2.7.4.11.5.2.932.0.61.428.6.44.410.9.19.049.6.53.714.9.21.541.2CLIP-LIP4.0.6.910.8.14.038.1.61.238.5.59.817.6.02.766.8.72.318.6.26.951.1
Ours0.0.0.047.7.90.758.5.93.969.7.100.056.9.99.393.5.100.045.9.66.392.9Ours0.0.0.080.1.100.059.3.95.270.9.100.055.8.97.491.5.99.046.2.66.894.1
" + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.298, + 0.845, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.5, + 0.908, + 0.53 + ], + "angle": 0, + "content": "Figure 3. Retrieval results on the ImageNet-1K dataset before and after the unlearning process in the CLIP model. We highlight the \"marmoset\" subgroup in red, with other \"monkey\" subgroups shown in green." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.587 + ], + "angle": 0, + "content": "compute the mean of these ratios across all classes, denoted as the Score, to provide a comprehensive evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.592, + 0.466, + 0.609 + ], + "angle": 0, + "content": "5.2. In-domain subgroup forgetting in ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.486, + 0.901 + ], + "angle": 0, + "content": "The Results are shown in Tab. 1. We observe that both the Fisher and LIP methods fail to optimize the model, resulting in severe catastrophic forgetting across the tested datasets. In comparison, FT, GA, EMMN, and CLIP-LIP perform slightly better but still suffer from catastrophic forgetting to varying degrees. For instance, optimizing only the retaining set with the FT strategy enables the CLIP model to forget the targeted subgroup while retaining knowledge of the intended subgroups. However, this approach leads to noticeable forgetting across other datasets, including different superclasses in ImageNet and unrelated datasets. Additionally, when applying gradient ascent to the forgetting subgroup with the GA strategy, CLIP exhibits forgetting within other subgroups of the same superclasses, and its zero-shot performance deteriorates due to catastrophic forgetting. Even with the inclusion of both forgetting and retaining samples in the fine-tuning process under the EMMN strategy, these issues remain inadequately addressed. By incorporating layer importance relative to the forgetting dataset to minimize the adverse ef" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.908, + 0.601 + ], + "angle": 0, + "content": "fects of fine-tuning and leveraging a model merging strategy, we can preserve most of the model's zero-shot performance while ensuring that the targeted subgroup is forgotten." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We also present the retrieval results using our fine-tuned CLIP model, which has forgotten the target subgroup, in Fig. 3. Before unlearning, when using the prompt 'monkey' to retrieve images, the model retrieves images from various subgroups, including 'marmoset' and 'howler monkey,' among others. Additionally, when prompted with specific subclass names, as shown in the second and third rows in figure, the model retrieves corresponding subgroup images. After unlearning, we observe that images of 'marmoset' are no longer retrieved when prompted with either the superclass name 'monkey' or the specific subgroup name 'marmoset,' indicating that the knowledge of 'marmoset' has been effectively erased from CLIP's parameter space. We also include retrieval results when using the prompt 'panda'; the consistent performance between the pre- and post-unlearning versions of CLIP shows that knowledge of other classes remains largely unaffected." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9875" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.132 + ], + "angle": 0, + "content": "Table 2. Evaluation results for forgetting the \"airplane\" (left) and \"truck\" (right) class in CIFAR-10 are presented. To assess the unaffected subgroup performance, we also separately retrieve the \"airplane\" class from ImageNet-1K (see the column labeled Target in ImageNet). The CIFAR-10 dataset is used exclusively as the retain set (see the column labeled Retain in CIFAR)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.143, + 0.88, + 0.287 + ], + "angle": 0, + "content": "
BackboneMethodCIFAR Target (↓)Retain (↑)FoodSTLObjectNetImageNet Target AllScoreBackboneMethodCIFAR Target (↓)Retain (↑)FoodSTLObjectNetImageNet Target AllScore
RN50Original81.370.460.992.068.953.659.8-RN50Original54.073.460.992.068.952.359.8-
FT0.00.092.6100.01.11.841.745.30.20.30.00.00.30.535.4FT0.00.056.376.71.32.129.311.80.40.60.00.00.20.330.2
GA0.00.014.210.21.01.614.115.30.10.10.20.40.10.219.7GA0.00.020.217.51.22.019.321.00.30.40.00.00.10.221.6
Fisher0.00.013.118.61.01.615.216.50.30.40.00.00.10.219.6Fisher0.00.012.416.91.22.013.921.00.20.40.00.00.20.219.2
LIP0.00.015.722.31.32.114.916.20.30.40.10.20.20.320.2LIP0.00.013.918.90.91.512.313.40.40.60.10.20.20.319.3
EMMN0.00.080.7100.01.62.641.645.20.20.30.00.00.50.835.6EMMN0.00.048.566.11.32.121.223.00.20.30.00.00.10.227.4
CLIP-LIP1.11.427.238.67.111.749.945.231.215.334.564.426.844.851.1CLIP-LIP3.87.045.361.716.627.374.280.733.648.828.953.339.566.161.8
Ours7.61365.102.551.985.285.793.257.984.049.892.950.183.888.9Ours0.91.763.266.150.582.984.491.756.491.951.585.852.187.189.5
RN101Original69.264.457.392.469.251.962.3-RN101Original91.479.657.392.469.250.662.3-
FT0.00.094.6100.01.11.958.463.20.40.60.00.00.20.338.0FT0.00.058.373.21.42.440.433.70.30.40.10.20.30.531.5
GA0.00.011.117.20.91.615.016.20.30.40.00.00.10.219.4GA0.00.020.225.41.22.119.320.90.20.30.00.00.10.221.3
Fisher0.00.012.920.01.01.716.918.30.30.40.10.20.20.320.1Fisher0.00.013.116.51.01.714.215.40.30.40.10.20.10.219.2
LIP0.00.013.821.41.62.815.717.00.20.30.00.00.30.520.3LIP0.00.015.719.71.62.815.717.00.30.40.00.00.20.320.0
EMMN0.00.081.2100.01.22.156.761.40.30.40.10.20.20.337.8EMMN0.00.048.360.71.42.435.848.70.10.10.10.20.20.328.9
CLIP-LIP7.410.735.585.110.418.263.168.330.432.920.699.719.831.849.5CLIP-LIP9.610.552.866.319.534.076.082.350.773.342.584.043.670.071.3
Ours8.812.763.208.148.885.287.794.958.184.045.788.141.266.186.2Ours3.23.570.388.446.280.686.693.758.384.243.986.846.674.886.4
" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.301, + 0.86, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.51, + 0.713, + 0.526 + ], + "angle": 0, + "content": "Figure 4. Retrieval results of the CLIP pre- and post-unlearning process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.418, + 0.567 + ], + "angle": 0, + "content": "5.3. Cross-domain forgetting in CIFAR-10" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.578, + 0.486, + 0.776 + ], + "angle": 0, + "content": "We then validate the effectiveness of our method in forgetting images belonging to common classes in CIFAR-10, whose concepts also appear in other datasets, such as ImageNet. The main challenge lies in the significant distributional differences between CIFAR-10 and the pre-trained dataset, which can result in severe catastrophic forgetting during the finetuning process for unlearning. In this experiment, we select 'airplane' and 'truck' as the target groups to unlearn from CLIP. In addition to evaluating the zero-shot classification performance of these targeted classes in CIFAR-10, we also assess their performance in ImageNet to determine whether unlearning is accurately confined to the specified distribution, without affecting the entire concept." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We present the results in Tab. 2. Compared to the results on ImageNet, models fine-tuned on the CIFAR-10 dataset exhibit more severe catastrophic forgetting, even with methods like FT, GA, and EMMN, which retain some zero-shot capability after forgetting specific classes in ImageNet. Notably, models fine-tuned on CIFAR-10 to retain performance demonstrate strong zero-shot performance on STL-10, a dataset with a distribution similar to CIFAR-10, highlighting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.908, + 0.642 + ], + "angle": 0, + "content": "the impact of the fine-tuning dataset on model performance. In contrast, our method shows remarkable robustness under such a significant distribution shift. Specifically, while the post-unlearning CLIP achieves random classification performance on the forgotten class, it retains at least \\(66.1\\%\\) of its original zero-shot performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.793 + ], + "angle": 0, + "content": "We also retrieve related classes in CIFAR-10 and ImageNet using the pre- and post-unlearning CLIP models for the 'airplane' class. As shown in Fig. 4, we observe that the post-unlearning CLIP is unable to retrieve airplanes from the CIFAR-10 dataset but can still retrieve airplanes from the ImageNet dataset, indicating precise forgetting limited to the target distribution. The consistency of retrieved results between the pre- and post-unlearning CLIP further demonstrates the effectiveness of our method in preserving unrelated knowledge within the model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.669, + 0.82 + ], + "angle": 0, + "content": "5.4. Style forgetting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We further evaluate performance on forgetting a specific style within a given class, representing a more fine-grained subgroup image unlearning challenge. Specifically, we use Stable Diffusion to generate or transform existing images into three distinct styles: poster art, black-and-white drawing," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.48, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.216, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Figure 5. Examples of stylized images of the sports car in the ImageNet-1K dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.26, + 0.483, + 0.289 + ], + "angle": 0, + "content": "Table 3. Evaluation results of CLIP after forgetting sketch-style \"auto-cars\" using various methods." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.299, + 0.485, + 0.389 + ], + "angle": 0, + "content": "
MethodCIFARFoodSTLObjectNetImageNetScore
Target style (↓)Retain (↑)All (↑)
Original60.852.288.465.468.182.558.6-
FT10.116.61.93.610.511.93.96.099.5100.098.3100.01.11.920.0
GA46.776.839.876.283.294.157.888.42.02.950.361.052.188.983.2
Fisher10.517.31.52.915.918.01.11.70.10.10.50.60.20.320.1
LIP11.218.40.91.713.615.44.26.40.00.00.00.00.50.920.4
EMMN16.326.826.50.839.244.320.130.714.521.370.285.14.88.246.4
CLIP-LIP15.425.36.512.518.721.28.412.81.62.32.42.91.32.224.9
Ours49.280.943.082.480.891.458.990.12.13.166.080.051.988.687.2
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.413, + 0.484, + 0.504 + ], + "angle": 0, + "content": "and sketch, as illustrated in Figure 5. We first fine-tune the CLIP on images with different styles, of which the zero-shot performance is reported at the first line in Tab. 3. Then, different unlearning methods are applied to erase the model's knowledge of cars in the sketch style, while retaining images of cars in the other styles. Results are shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.747 + ], + "angle": 0, + "content": "Among the baselines, we find that GA performs the best, effectively forgetting the targeted style while retaining most unrelated knowledge, as indicated by its strong zero-shot performance across different datasets. In comparison, simply fine-tuning CLIP on the retaining set leads to severe catastrophic forgetting, as shown in the results of FT and EMMN. The EMMN results also suggest that incorporating the retaining set into the unlearning process is not always beneficial. Although using the retaining set could potentially reduce model over-forgetting, an imprecise selection of the retaining set and an imbalanced loss between it and the forgetting set can degrade the model's generalization ability on other datasets. By selecting the optimal layer for fine-tuning to minimize the impact on the model's original performance, our method achieves the best overall results, maintaining at least \\(80.4\\%\\) of the model's zero-shot capacity." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.757, + 0.357, + 0.773 + ], + "angle": 0, + "content": "5.5. Ablation study and discussion" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "On the importance of reminding stage. The reminding stage follows the forgetting stage, using the retaining set to partially recover any unintended knowledge loss. Notably, our retaining set typically differs from the original pre-training set, potentially further compromising CLIP's zero-shot classification capability. To examine this, we conduct an ablation study on the impact of iteration count during the reminding stage. Using CIFAR-10 as the unlearning" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 4. Ablation study on the used number of steps at retaining stage on the zero-shot capacity of CLIP. We set the batch size as 64 for fine-tuning and \\(\\alpha\\) as 0.65 for model merging." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.143, + 0.878, + 0.19 + ], + "angle": 0, + "content": "
# steps51020304050100
Accr59.863.265.641.435.434.932.5
Accin44.941.238.125.88.52.10.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.211, + 0.907, + 0.24 + ], + "angle": 0, + "content": "Table 5. Ablation study on the \\( \\alpha \\) for the model merging operation at the restoring stage to the zero-shot capacity of CLIP." + }, + { + "type": "table", + "bbox": [ + 0.566, + 0.25, + 0.859, + 0.308 + ], + "angle": 0, + "content": "
α0.10.30.50.650.70.9
Accf0.00.02.48.812.868.8
Accr22.727.542.763.266.778.7
Accin1.04.422.3941.243.560.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.909, + 0.509 + ], + "angle": 0, + "content": "dataset for CLIP, we evaluate performance on both the retain set \\((\\mathrm{ACC}_r)\\) and ImageNet \\((\\mathrm{ACC}_{in})\\). As shown in Tab. 4, we observe that with an increasing number of steps, classification accuracy on the retain set initially improves before declining, while accuracy on the ImageNet-1K dataset consistently decreases, with a significant drop at higher iteration counts. This suggests that while additional steps in the reminding stage help the model relearn knowledge that was unintentionally forgotten, they also lead to a larger distribution shift from the original parameter space, ultimately impairing zero-shot performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.911, + 0.709 + ], + "angle": 0, + "content": "On the model merging. In our work, we use model merging to fully restore zero-shot capacity. While our method leverages a calibration dataset to identify the optimal coefficient, \\(\\alpha\\), we also conduct an additional experiment by directly varying the coefficients to evaluate both forgetting performance \\((\\mathrm{ACC}_f)\\) and recovery performance \\((\\mathrm{ACC}_r / \\mathrm{ACC}_{in})\\). As shown in Tab. 5, achieving an optimal \\(\\alpha\\) is crucial to balancing forgetting and retaining performance. With increasing \\(\\alpha\\), the model's retaining performance and zero-shot classification capacity improve continuously. However, a large \\(\\alpha\\) can also cause the model to relearn data that should be forgotten, as indicated by the \\(68.8\\%\\) accuracy on the forgetting set when \\(\\alpha\\) is set to 0.9." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.735, + 0.634, + 0.751 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we study the problem of forgetting a subgroup of images in the CLIP model. We propose a three-stage method consisting of the forgetting, retaining, and restoring steps. Our method does not require any pre-trained dataset, allowing it to achieve a balanced approach by selectively forgetting the targeted subgroup of images while preserving the knowledge that should not be forgotten. Experimental results validate the effectiveness of our method and highlight the challenges of this task, as most classical methods fail." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9877" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Andrew Bai, Chih-Kuan Yeh, Cho-Jui Hsieh, and Ankur Taly. Which pretrain samples to rehearse when finetuning pretrained models? arXiv preprint arXiv:2402.08096, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[3] Alexander Becker and Thomas Liebig. Evaluating machine unlearning via epistemic uncertainty. arXiv preprint arXiv:2208.10836, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[4] Yoshua Bengio, Aaron C Courville, and Pascal Vincent. Unsupervised feature learning and deep learning: A review and new perspectives. CoRR, abs/1206.5538, 1(2665):2012, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[5] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.484, + 0.535 + ], + "angle": 0, + "content": "[7] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141-159. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.484, + 0.576 + ], + "angle": 0, + "content": "[8] Zikui Cai, Yaoteng Tan, and M Salman Asif. Single layer single gradient unlearning. arXiv preprint arXiv:2407.11867, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.62 + ], + "angle": 0, + "content": "[9] Hao Chen, Bhiksha Raj, Xing Xie, and Jindong Wang. On catastrophic inheritance of large foundation models. arXiv preprint arXiv:2402.01909, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.622, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[10] Ruizhe Cheng, Bichen Wu, Peizhao Zhang, Peter Vajda, and Joseph E Gonzalez. Data-efficient language-supervised zero-shot learning with self-distillation. In CVPR, pages 3119-3124, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[11] Eli Chien, Chao Pan, and Olgica Milenkovic. Efficient model updates for approximate unlearning of graph-structured data. In The Eleventh International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[12] Vikram S Chandawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security, 18:2345-2354, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[13] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] Lluc Clavera Comas. Machine unlearning: fisher infomation matrix and selective forgetting in deep networks. B.S. thesis, Universitat Politècnica de Catalunya, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[15] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[17] Ronen Eldan and Mark Russinovich. Who's harry potter? approximate unlearning in llms. arXiv preprint arXiv:2310.02238, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[18] Jack Foster, Kyle Fogarty, Stefan Schoepf, Cengiz Öztireli, and Alexandra Brintrup. Zero-shot machine unlearning at scale via lipschitz regularization. arXiv preprint arXiv:2402.01401, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[19] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9304–9312, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[20] Jinu Gong, Osvaldo Simeone, and Joonhyuk Kang. Bayesian variational federated learning and unlearning in decentralized networks. In 2021 IEEE 22nd International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pages 216-220. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.515 + ], + "angle": 0, + "content": "[21] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. NeurIPS, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[22] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In AAAI, pages 746-754, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[23] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. *NeurIPS*, 34:16319-16330, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[24] Kris Holt. Researchers found child abuse material in the largest ai image generation dataset, 2024. Accessed: 2024-10-25. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.906, + 0.715 + ], + "angle": 0, + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[26] Mark He Huang, Lin Geng Foo, and Jun Liu. Learning to unlearn for robust machine unlearning. arXiv preprint arXiv:2407.10494, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[27] Zhehao Huang, Xinwen Cheng, JingHao Zheng, Haoran Wang, Zhengbao He, Tao Li, and Xiaolin Huang. Unified gradient-based machine unlearning with remain geometry enhancement. arXiv preprint arXiv:2409.19732, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[28] Frank Klinker. Exponential moving average versus moving exponential average. Mathematische Semesterberichte, 58: 97-107, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[29] Alexey Kravets and Vinay P. Namboodiri. Zero-shot CLIP class forgetting via text-image space adaptation. Transactions on Machine Learning Research, 2025. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9878" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[31] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In CVPR, pages 20071-20082, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.483, + 0.245 + ], + "angle": 0, + "content": "[32] Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, PRANAY SHARMA, Sijia Liu, et al. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.483, + 0.301 + ], + "angle": 0, + "content": "[33] Ji Liu, Jiaxiang Ren, Ruoming Jin, Zijie Zhang, Yang Zhou, Patrick Valduriez, and Dejing Dou. Fisher information-based efficient curriculum federated learning with large language models. arXiv preprint arXiv:2410.00131, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[34] Yufang Liu, Changzhi Sun, Yuanbin Wu, and Aimin Zhou. Unlearning with fisher masking. arXiv preprint arXiv:2310.05331, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[35] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Xi-angyang Ji, Antoni Chan, and Rong Jin. Improved fine-tuning by better leveraging pre-training data. Advances in Neural Information Processing Systems, 35:32568-32581, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[36] Alexander Ly, Maarten Marsman, Josine Verhagen, Raoul PPP Grasman, and Eric-Jan Wagenmakers. A tutorial on fisher information. Journal of Mathematical Psychology, 80:40-55, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[37] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N Ravi. Deep unlearning via randomized conditionally independent hessian. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10422-10431, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[38] Quoc Phong Nguyen, Bryan Kian Hsiang Low, and Patrick Jaillet. Variational bayesian unlearning. Advances in Neural Information Processing Systems, 33:16025-16036, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[39] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. *Neural networks*, 113:54–71, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[40] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara, et al. Safe-clip: Removing nsfw concepts from vision-and-language models. In Proceedings of the European Conference on Computer Vision, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[41] Clifton Poth, Jonas Pfeiffer, Andreas Rückle, and Iryna Gurevych. What to pre-train on? efficient intermediate task selection. arXiv preprint arXiv:2104.08247, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[42] Jenny Quang. Does training ai violate copyright law? Berkeley Tech. LJ, 36:1407, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[44] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, pages 2765-2775, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.133 + ], + "angle": 0, + "content": "[45] Shibani Santurkar, Dimitris Tsipras, and Aleksander Madry. Breeds: Benchmarks for subpopulation shift. arXiv preprint arXiv:2008.04859, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[46] Sebastian Schelter, Mozhdeh Ariannezhad, and Maarten de Rijke. Forget me now: Fast and exact unlearning in neighborhood-based recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2011–2015, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[47] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[48] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 34:18075-18086, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[49] Adam Shostack. The boy who survived: Removing harry potter from an llm is harder than reported. arXiv preprint arXiv:2403.12082, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.907, + 0.474 + ], + "angle": 0, + "content": "[50] Shashank Subramanian, Peter Harrington, Kurt Keutzer, Wahid Bhimji, Dmitriy Morozov, Michael W Mahoney, and Amir Gholami. Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. NeurIPS, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.907, + 0.531 + ], + "angle": 0, + "content": "[51] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[52] Anvith Thudi, Gabriel Deza, Varun Chandrasekaran, and Nicolas Papernot. Unrolling sgd: Understanding factors influencing machine unlearning. In 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P), pages 303-319. IEEE, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.907, + 0.659 + ], + "angle": 0, + "content": "[53] Enayat Ullah, Tung Mai, Anup Rao, Ryan A Rossi, and Raman Arora. Machine unlearning via algorithmic stability. In Conference on Learning Theory, pages 4126-4142. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[54] Alexander Warnecke, Lukas Pirch, Christian Wressnegger, and Konrad Rieck. Machine unlearning of features and labels. In 30th Annual Network and Distributed System Security Symposium, NDSS 2023, San Diego, California, USA, February 27 - March 3, 2023. The Internet Society, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.907, + 0.801 + ], + "angle": 0, + "content": "[55] Yixuan Wei, Han Hu, Zhenda Xie, Ze Liu, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Improving clip fine-tuning performance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5439-5449, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[56] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9879" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[57] Tomoya Yamashita, Masanori Yamada, and Takashi Shibata. One-shot machine unlearning with mnemonic code. In The 16th Asian Conference on Machine Learning (Conference Track). 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[58] Haonan Yan, Xiaoguang Li, Ziyao Guo, Hui Li, Fenghua Li, and Xiaodong Lin. Arcane: An efficient architecture for exact machine unlearning. In *IJCAI*, page 19, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[59] Kaiyu Yang, Clint Qinami, Li Fei-Fei, Jia Deng, and Olga Russakovsky. Towards fairer datasets: Filtering and balancing the distribution of the people subtree in the imagenet hierarchy. In Proceedings of the 2020 conference on fairness, accountability, and transparency, pages 547–558, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[60] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. NeurIPS, 36, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[61] Binchi Zhang, Yushun Dong, Tianhao Wang, and Jundong Li. Towards certified unlearning for deep neural networks. arXiv preprint arXiv:2408.00920, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[62] Jiahao Zhang. Graph unlearning with efficient partial retraining. In Companion Proceedings of the ACM on Web Conference 2024, pages 1218-1221, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[63] Zeliang Zhang, Mingqian Feng, Zhiheng Li, and Chenliang Xu. Discover and mitigate multiple biased subgroups in image classifiers. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[64] Zeliang Zhang, Zhuo Liu, Mingqian Feng, and Chenliang Xu. Can clip count stars? an empirical study on quantity bias in clip. In EMNLP (Findings), 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9880" + } + ] +] \ No newline at end of file diff --git a/2025/Targeted Forgetting of Image Subgroups in CLIP Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_origin.pdf b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d94f6162b4b6f3f2edf9c03c2b364b4af7773b22 --- /dev/null +++ b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/d5d3c532-3ca2-4b7b-a059-b24be7107188_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad583c14428fd35daa7b386ab7269a9d3960d6063d4f42ade0069f6afec9c3be +size 6358613 diff --git a/2025/Targeted Forgetting of Image Subgroups in CLIP Models/full.md b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b25db54eaf628798b6e9c3c99b8abc3b2c491146 --- /dev/null +++ b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/full.md @@ -0,0 +1,301 @@ +# Targeted Forgetting of Image Subgroups in CLIP Models + +Zeliang Zhang $^{1}$ Gaowen Liu $^{2}$ Charles Fleming $^{2}$ Ramana Rao Kompella $^{2}$ Chenliang Xu $^{1}$ $^{1}$ University of Rochester $^{2}$ Cisco Research + +{zeliang.zhang, chenliang.xu}@rochester.edu, {gaoliu, chflemin, rkompell}@cisco.com + +# Abstract + +Foundation models (FMs) such as CLIP have demonstrated impressive zero-shot performance across various tasks by leveraging large-scale, unsupervised pre-training. However, they often inherit harmful or unwanted knowledge from noisy internet-sourced datasets, compromising their reliability in real-world applications. Existing model unlearning methods either rely on access to pre-trained datasets or focus on coarse-grained unlearning (e.g., entire classes), leaving a critical gap for fine-grained unlearning. In this paper, we address the challenging scenario of selectively forgetting specific portions of knowledge within a class—without access to pre-trained data—while preserving the model's overall performance. We propose a novel three-stage approach that progressively unlearns targeted knowledge while mitigating over-forgetting. It consists of (1) a forgetting stage to fine-tune the CLIP on samples to be forgotten, (2) a reminding stage to restore performance on retained samples, and (3) a restoring stage to recover zero-shot capabilities using model souping. Additionally, we introduce knowledge distillation to handle the distribution disparity between forgetting/retaining samples and unseen pre-trained data. Extensive experiments on CIFAR-10, ImageNet-1K, and style datasets demonstrate that our approach effectively unlearns specific subgroups while maintaining strong zero-shot performance on semantically similar subgroups and other categories, significantly outperforming baseline unlearning methods, which lose effectiveness under the CLIP unlearning setting. + +# 1. Introduction + +Foundation models (FMs) are deep neural networks pretrained on large-scale datasets [50]. By employing unsupervised training, these models capture complex patterns and relationships within the data, enabling a deep understanding of various domains [4]. For example, CLIP [43], a vision-language model, learns correspondences between natural language and images. Trained on the large-scale LAION-5B dataset [47], CLIP demonstrates strong zero + +![](images/ba6cd99b7c5621bd4b277f2eed7cbf59e04a64fe32deca2be9df55c53528bb89.jpg) +Figure 1. In the task of CLIP unlearning, classical machine unlearning methods often fail to generalize across multiple datasets due to limited access to the original pre-training data. In contrast, our proposed method does not require access to the pre-training dataset. After unlearning, CLIP not only forgets the specified subgroup of images but also maintains consistent performance on various unseen datasets in the zero-shot classification setting. + +shot performance [10, 22, 44]. However, its reliability in real-world applications is questionable [63, 64]. The training data, sourced from the internet, includes uncensored content such as discriminatory and violent imagery, copyright violations, and personal information [21, 42, 59]. This problematic content can be unintentionally learned during pre-training, raising concerns about the reliability and safety of downstream models [9, 31]. Therefore, it is critical to remove such harmful knowledge to ensure trustworthy FMs. + +Numerous studies have explored model unlearning, which can be broadly divided into exact [46, 53, 58] and approximate strategies [11, 23, 38]. Exact unlearning aims to completely eliminate the influence of forgotten data from the model, often by retraining it [7, 58, 62], ensuring the model behaves as if the data had never been included in training. While thorough, this approach is computationally expensive and impractical for large-scale models [32, 48]. In contrast, approximate unlearning offers a more efficient alternative, using techniques like variational inference [20, 38] or adaptive regularization [27, 37, 61] to partially reduce the data's + +impact. However, these methods face several challenges when applied to FMs. + +First, as shown in Fig. 1, the training data for FMs is often vast or inaccessible [5, 47], rendering exact unlearning impractical. Second, while FMs are trained on large-scale datasets using unsupervised learning, the fine-tuning typically used in approximate unlearning strategies [26, 51] can degrade generalization across other knowledge domains [1, 35, 41], leading to significant performance drops, particularly in zero-shot tasks [55]. Third, the knowledge to be forgotten is often deeply entangled with other learned information [17, 49]. This entanglement makes it difficult to isolate and remove specific knowledge from the model parameters, particularly when distributional disparities exist between the data we want to unlearn and the pre-trained data. + +Few studies have specifically addressed the challenge of unlearning in foundational models, particularly in CLIP. Most existing approaches focus on unlearning entire classes or broad concepts, yet real-world applications often require a more targeted approach, where only certain portions of knowledge within a class are forgotten. For example, we may want the model to forget representations of Disney's Mickey Mouse while retaining versions created by the public, a task known as subgroup image forgetting. Moreover, many current unlearning methods rely on access to the original pre-trained dataset to restore model performance, which is often unavailable. This limitation underscores a critical question in current research: how can models like CLIP achieve selective, fine-grained unlearning without relying on pre-trained data, while still maintaining generalization across other tasks? + +In this work, we bridge this gap by tackling a more challenging and practical scenario: enabling CLIP to selectively forget specific portions of knowledge within a single class without access to pre-trained data, framing this as a fine-grained unlearning task. Rather than unlearning an entire class or concept (e.g., all images of airplanes), we focus on selectively forgetting images from a particular subset (e.g., Boeing airplanes) while preserving the model's knowledge of other subsets (e.g., different manufacturers). This selective forgetting is particularly challenging due to the feature entanglement across subsets within a class, and it is exceedingly difficult to define and collect all subsets within the same class in a foundational model setting, which is trained on vast, large-scale data. + +To address the challenge of partial knowledge removal, we propose a novel, three-stage approach: forgetting, reminding, and restoring. In the forgetting stage, we identify layers essential for representing the forgetting samples but less critical to other data, fine-tuning the model on these samples to selectively remove targeted knowledge. We observe that models can over- forget samples from similar distributions, so we introduce a retaining set with a similar distribution to the + +targeted subgroup, aiding layer identification and supporting the reminding stage. During reminding, we fine-tune the model on the retaining samples to reinforce CLIP's performance on data prone to over-forgetting due to distributional similarity. In the final restore stage, we employ model souping to fully recover CLIP's zero-shot performance. Throughout this process, we address the distributional gap between forgetting/retaining samples and unseen pre-trained data by distilling knowledge from CLIP, achieving a more precise representation without relying on the original pre-trained dataset. Extensive experiments demonstrate our method's effectiveness in selectively unlearning subgroup images while preserving robust zero-shot performance across diverse distributions and datasets. + +We summarize our contributions as follows: + +1. We formulate the problem of subgroup image unlearning within the CLIP model. +2. We propose a three-stage method for unlearning in CLIP, consisting of forgetting, reminding, and restoring. Our method is data-efficient, which does not require access to CLIP's pre-training dataset, making it more practical for real-world applications. +3. We conduct extensive experiments to validate the effectiveness of our proposed method, including unlearning tasks on CIFAR-10, ImageNet-1K, and style forgetting. Additionally, we demonstrate the scalability of our approach across various applications through retrieval-based evaluations. + +# 2. Related work + +Machine Unlearning. Various strategies have been developed to facilitate machine unlearning in models. Notably, Golatkar et al. [19], Warnecke et al. [54] employ a method of continuous fine-tuning (FT) on the retain dataset—those data elements the model is expected to remember. This approach not only approximates the forgetting of the target dataset but also mitigates the issue of overforgetting [32, 39], in contrast to exclusively fine-tuning on the forgetting dataset. Becker and Liebig [3], Liu et al. [32] leverage Fisher information, which assesses the sensitivity of parameters concerning the forgetting dataset, to achieve model unlearning. Meanwhile, Thudi et al. [52] unroll the knowledge embedded in the forgetting dataset from the model parameters using stochastic gradient ascent (GA). Further, [18] introduces the use of local Lipschitz (LIP) perturbations on the forgetting dataset as a technique for unlearning in image classification models, which Kravets and Namboodiri [29] subsequently adapts to the CLIP model. Additionally, Chandawat et al. [12] proposes an error minimization-maximization (EMMN) framework that aims to maximize errors on the forgetting dataset while minimizing them on the retraining dataset, thus achieving a balance between forgetting and retaining performance. Among these methods, EMMN [12] and LIP [29] stand out + +as they do not require access to the training datasets (both forgetting and retaining), unlike other approaches. + +On the Importance of Unlearning in CLIP. Trained on a billion-scale dataset of image-text pairs [47], the CLIP model [43] excels in zero-shot classification tasks. Despite its strengths, the presence of uncensored samples within the training data has raised practical concerns [24]. Given the impracticality of retraining the CLIP model from scratch due to significant resource demands and the inaccessibility of the large-scale dataset, it is imperative to explore machine unlearning methods to selectively erase problematic knowledge from the model. Poppi et al. [40] suggest an approach to first distinguish between safe and unsafe content, followed by fine-tuning CLIP to alter the embeddings of unsafe content. Meanwhile, Cai et al. [8] focuses on identifying and unlearning from the most influential layer in CLIP, using a subset of the training set to restore zero-shot capabilities. [29] expand zero-shot machine unlearning techniques to the CLIP model, enabling the forgetting of entire concepts as indicated by prompts. Concurrently, Kravets and Namboodiri [29] apply Layer-wise Relevance Propagation (LoRA) to the text encoder of CLIP, specifically fine-tuning this component to unlearn a given concept within the model. + +# 3. Problem Formulation + +The CLIP model, denoted $g(\cdot, \cdot) = \{g^{\mathrm{img}}(\cdot), g^{\mathrm{txt}}(\cdot)\}$ , processes image and text inputs into embeddings, $e^{\mathrm{img}}$ and $e^{\mathrm{txt}}$ . It aims for high cosine similarity between these embeddings when they are correctly paired, and low similarity otherwise, using a threshold $\delta$ . + +Suppose a large dataset $D = \{x_{n}^{\mathrm{img}}, x_{n}^{\mathrm{txt}}, c_{n}, g_{n}\}_{n=1}^{N}$ , where we denote $x^{\mathrm{img}}$ as the image, $x^{txt}$ as the paired text learned by the CLIP, $c$ as the coarse main category, and $g$ is the fine-grained subclass label. The forgetting dataset $D^{f} = \{x_{n}^{\mathrm{img}}, x_{n}^{\mathrm{txt}}, c_{n}, g_{n}\}_{n=1}^{N_{f}}$ is usually a small set compared with $D$ . The goal is to unlearn the associations between $x^{\mathrm{img}}$ and $x^{txt}$ on $D^{f}$ without impacting the model's performance on the remaining dataset $D^{r} = D / D_{f}$ , namely the retain set. The task involves selectively erasing knowledge from $D^{f}$ while maintaining performance on $D^{r}$ , which can be formulated as follows: + +$$ +\forall \left(x _ {i} ^ {\mathrm {i m g}}, x ^ {\mathrm {t x t}}\right) \in D ^ {f}, \frac {\hat {g} ^ {\mathrm {i m g}} \left(x _ {i} ^ {\mathrm {i m g}}\right) \cdot \hat {g} ^ {\mathrm {t x t}} \left(x ^ {\mathrm {t x t}}\right)}{\| \hat {g} ^ {\mathrm {i m g}} \left(x _ {i} ^ {\mathrm {i m g}}\right) \cdot \hat {g} ^ {\mathrm {t x t}} \left(x ^ {\mathrm {t x t}}\right) \|} < \delta , \tag {1} +$$ + +$$ +\forall (x _ {i} ^ {\mathrm {i m g}}, x _ {i} ^ {\mathrm {t x t}}) \in D ^ {r}, \hat {g} (x _ {i} ^ {\mathrm {i m g}}, x _ {i} ^ {\mathrm {t x t}}) \approx g (x _ {i} ^ {\mathrm {i m g}}, x _ {i} ^ {\mathrm {t x t}}), +$$ + +where $\hat{g}$ represents the unlearned CLIP model and $g$ the original model. + +Implementing the approach outlined in Eq. (1) is challenging due to several factors: + +1. The large dataset $D$ used for training $g$ is inaccessible, and similarly, $D_r$ is also unavailable. We can only access a small subset of $D_r$ to recover model performance + +after forgetting, blocking the direct use of conventional machine unlearning methods. + +2. In addition, we may only have access to a coarse label $c$ as the surrogate of $x^{txt}$ rather than a fine-grained subgroup label $g$ . This limitation increases the risk of inaccurately erasing targeted knowledge and over-forgetting unrelated knowledge that is similar to the target. +3. There may be a substantial distribution shift between $D_{f}$ and $D$ . With a large number of parameters, directly finetuning on the small dataset $D_{f}$ could cause the model to overfit and lose generalization ability. + +Thus, the goal of this task is twofold: it involves selectively forgetting only specific image subgroups because of coarse labeling, while simultaneously preserving the model's robust zero-shot capabilities across various datasets, including those characterized by similarly coarse labels. + +# 4. Methodology + +To address the challenge of subgroup image unlearning, we propose a novel methodology that comprises three stages: forgetting, reminding, and restoring, as depicted in Fig. 2. + +# 4.1. Forgetting + +Previous studies [14, 34] indicate that selectively fine-tuning critical layers for a target dataset can effectively mitigate over-forgetting [33], typically using Fisher information [36, 57] to assess the importance of each layer. However, while Fisher information measures parameter sensitivity to the target dataset, it overlooks potential impacts on other datasets, which are inaccessible in our task. Our experiments reveal that relying solely on the forgetting dataset to modify the CLIP model often leads to catastrophic over-forgetting on similar subgroups with distributions akin to the forgetting data, as shown in Tab. 1 for the CLIP model optimized by GA. To address this, we propose using relative Fisher information which combines both the forgetting and retaining samples as a more nuanced metric for identifying essential layers for fine-tuning, thereby enhancing the precision of the unlearning process. + +Concretely, given the forgetting dataset $\mathcal{D}^f = \{(x_i^{\mathrm{img}}, x^{\mathrm{txt}})\}$ , we manually construct a retain dataset $\mathcal{D}^r = \{(x_i^{\mathrm{img}}, x_i^{\mathrm{txt}})\}$ with a distribution similar to $\mathcal{D}^f$ , used for alleviating the aforementioned subgroup over-forgetting. Here, $\mathcal{D}^r$ is comprised of similar subgroups of images that we aim to preserve in the model's memory. In practical scenarios, we often use coarse labels as surrogate image descriptions, so $x^{\mathrm{txt}}$ from the forgetting dataset may also appear in $\mathcal{D}^r$ . We calculate the relative Fisher information for the $l$ -th layer in $g^{\mathrm{img}}(\cdot)$ as follows: + +$$ +\mathcal {I} ^ {l} = \frac {\mathbb {E} _ {\left(x _ {i} ^ {\mathrm {i m g}} , x ^ {\mathrm {t x t}}\right) \sim \mathcal {D} ^ {f}} \left[ \nabla_ {\theta^ {l}} ^ {2} \left(g ^ {\mathrm {i m g}} \left(x _ {i} ^ {\mathrm {i m g}}\right) \cdot g ^ {\mathrm {t x t}} \left(x ^ {\mathrm {t x t}}\right)\right) \right]}{\mathbb {E} _ {\left(x _ {i} ^ {\mathrm {i m g}} , x _ {i} ^ {\mathrm {t x t}}\right) \sim \mathcal {D} ^ {r}} \left[ \nabla_ {\theta^ {l}} ^ {2} \left(g ^ {\mathrm {i m g}} \left(x _ {i} ^ {\mathrm {i m g}}\right) \cdot g ^ {\mathrm {t x t}} \left(x _ {i} ^ {\mathrm {t x t}}\right)\right) \right]}, \tag {2} +$$ + +![](images/66fbaebd037d2b8e98bc47ce995193b54015a996c2a3098f5f397f2e95dd3b0a.jpg) +Figure 2. Overview of our method to unlearn the CLIP on a subgroup of images. We first compute the relative fisher information on the forgetting dataset and retain dataset to evaluate the layer importance. We select important layers for the following fine-tuning process. During the fine-tuning process for knowledge reminding, we align the distribution of our retain dataset to that of the pre-training dataset learned by the model. A model merging operation is performed at the last step of knowledge restoration. + +where $\theta^l$ represents the parameters of the $l$ -th layer. + +We then rank the importance of different layers based on the computed relative Fisher information, as detailed in Eq. (2). Layers with higher relative importance are selected for targeted unlearning, with all other layers frozen to minimize the impact on non-target knowledge. In our experiments, applying the low-rank adaptation (LoRA) technique [25] to the selected layers significantly enhances the targeted unlearning performance while minimizing effects on other retained knowledge. + +Finally, we fine-tune our model on the forgetting dataset $\mathcal{D}_f$ to minimize the following loss function: + +$$ +\min \mathcal {L} _ {f} = \sum_ {i} \frac {g ^ {i m g} \left(x _ {i} ^ {i m g}\right) \cdot g ^ {t x t} \left(x ^ {t x t}\right)}{\| g ^ {i m g} \left(x _ {i} ^ {i m g}\right) \cdot g ^ {t x t} \left(x ^ {t x t}\right) \|}, \tag {3} +$$ + +where images in the forgetting dataset are associated through the same coarse label $x^{txt}$ . + +By fine-tuning the CLIP model on images marked for unlearning with coarse labels, we aim to selectively erase specific knowledge. However, this adjustment often leads to catastrophic overforgetting, affecting the model's ability across both similar and different image distributions. For example, attempting to forget the "airplane" class in CIFAR-10 may also affect the recognition of other classes within CIFAR-10 and in ImageNet-1K. To tackle these challenges, particularly when full access to the pre-training dataset is unavailable, we implement reminding and restoration stages. + +# 4.2. Reminding + +An intuitive strategy to retain non-targeted knowledge in CLIP is to fine-tune it on our retaining set, built during the forgetting stage for computing $\mathcal{I}$ . However, two challenges + +hinder the direct application of this fine-tuning step. First, our retaining set may have a different distribution than the pre-training dataset, potentially causing inefficient learning. Second, our retaining set is typically limited, and fine-tuning on it could lead to further catastrophic over-forgetting in various unseen datasets and tasks. + +To address the first challenge, inspired by research showing that training data features are encoded in model parameters [60], we employ an optimization-based strategy to align the retaining dataset's distribution with the pre-training dataset. Specifically, for a given batch of image-text pairs in the retaining dataset $\mathcal{D}_r = \{(x_i^{\mathrm{img}}, x_i^{\mathrm{xt}})\}_{i=1}^B$ , we introduce a batch of perturbations $\delta_{i_{i=1}}^{B}$ on the images. These perturbations are optimized by minimizing the loss: + +$$ +\begin{array}{l} \mathcal {L} _ {a} = \sum_ {l} \left\| \frac {1}{B} \sum_ {i = 1} ^ {B} \mu_ {l} ^ {\text {i m g}} \left(x _ {i} ^ {\text {i m g}} + \delta_ {i}\right) - \mathrm {B N} _ {l} ^ {\mu} \right\| \tag {4} \\ + \sum_ {l} \left\| \frac {1}{B} \sum_ {i = 1} ^ {B} \sigma_ {l} ^ {i m g} \left(x _ {i} ^ {i m g} + \delta_ {i}\right) - \mathbf {B N} _ {l} ^ {\sigma} \right\|, \\ \end{array} +$$ + +where $\mu_l^{img}(\cdot)$ and $\sigma_l^{img}$ are the mean and variance of the batch of intermediate features before the $l$ -th BatchNormalization (BN) layer, computed by the image encoder of original CLIP. $\mathrm{BN}_l^\mu$ and $\mathrm{BN}_l^\sigma$ are the mean and variance parameters of the $l$ -th BN layer, encoding the global distribution information of the pre-training dataset. + +We then fine-tune the CLIP model on the retaining dataset, now aligned with the pre-training dataset's distribution. This serves as a reminder of the knowledge encoded in the original model, effectively achieved without direct access to the pre-training dataset. This strategy helps prevent further catastrophic forgetting that could arise from distributional + +disparities between the retaining dataset and the original training dataset. + +To address the second challenge of overfitting during fine-tuning due to the large amount of parameters, we propose minimizing the impact on the original model parameters by using an exponential moving average (EMA) [28] for the final model parameters, as follows: + +$$ +\theta^ {\mathrm {e m a}} = \alpha \cdot \theta^ {\mathrm {e m a}} + (1 - \alpha) \cdot \theta , \tag {5} +$$ + +where $\theta^{\mathrm{ema}}$ is initialized as the original CLIP parameters $\theta_0$ , $\theta$ represents the continuously optimized parameters from $\theta_0$ , and $\alpha$ is the decay factor. This approach effectively balances parameter retention with necessary updates. + +# 4.3. Restoration + +In addition to fine-tuning on the retain dataset for reminding, we also employ a model merging strategy to restore the zero-shot performance of CLIP, a technique empirically shown to locate flat optima on the loss landscape, enhancing generalization [56]. + +Specifically, we first compile a small calibration dataset $\mathcal{D}_m = \{(x_i,y_i)\}_{i = 1}^N$ for zero-shot image classification, which may include portions of the retain dataset not used in fine-tuning. We then execute model merging with varying coefficients as follows: + +$$ +\begin{array}{l} \underset {\theta} {\arg \max } \operatorname {A c c} (g, \mathcal {D} _ {m}), \tag {6} \\ s. t. \theta = \alpha \cdot \theta^ {f} + (1 - \alpha) \cdot \theta^ {o r i}, \\ \end{array} +$$ + +where $Acc(g, \mathcal{D}_m)$ assesses the zero-shot classification performance of model $g$ on $\mathcal{D}_m$ , and $\theta^f$ and $\theta^{ori}$ represent the parameters of the CLIP model post-unlearning and the original model, respectively. + +# 5. Evaluations + +We begin by focusing on the forgetting subgroups within the ImageNet dataset, where the images are semantically grouped by Breeds. We select two superclasses: 'monkey' and 'turtle.' Each superclass contains four distinct subgroups. We specifically target one subgroup in each superclass for unlearning: 'marmoset' in the monkey superclass and 'box turtle' in the turtle superclass. + +# 5.1. Experimental setup + +Settings. We conduct experiments in three distinct settings: in-domain forgetting, near-domain forgetting, and style forgetting, as detailed below: + +- In-domain forgetting: Experiments are conducted on the Breeds dataset [45], a subset of the ImageNet-1K dataset [16] consisting of 4 subgroups within 17 superclasses. The images, with dimensions of $224 \times 224$ , closely + +match the distribution of images in the pre-trained LAION-5B dataset used for CLIP. For each superclass, one subgroup is selected for unlearning, while the images from the remaining three subgroups form the retaining set. + +- Out-of-domain forgetting: Experiments are performed on the CIFAR-10 dataset [30], which differs significantly from the image distribution in LAION-5B. To align with CLIP's input format, we resize the images from $32 \times 32$ to $224 \times 224$ . In this setting, we select a class shared between CIFAR-10 and ImageNet-1K. The images of this selected class form the forgetting dataset, while the remaining classes in CIFAR-10 constitute the retaining set. +- Style forgetting: Experiments are conducted on one class in the ImageNet-1K dataset [16] transformed by Stable Diffusion [13]. The images are randomly transformed into three styles: sketch style, poster art, and black-and-white drawing. For a given style to forget, images in the remaining styles are treated as the retaining set. + +Baselines and implementations. We select six methods as our baselines: FT [54], GA [52], Fisher [32], LIP [18], and EMMN [12], and CLIP-LIP [29]. Except the CLIP-LIP, these methods are developed for conventional models that require a classification head. To address this difference, we use the probability from the zero-shot classification task as a surrogate for the classification logits. In these approaches, the text prompt is provided by the superclass names in both the forgetting and retain sets. For all learning-based methods, recognizing the sensitivity of CLIP parameters during the fine-tuning process, we use the Adam optimizer with a small learning rate of $1 \times 10^{-6}$ for up to 2 epochs. For LIP, we set the number of noisy copies to 10. Additionally, the variance scaling factor $\alpha$ in Fisher is set to 0.2. We provide the full implementation in Appendix A.1. + +**Evaluations.** The performance of image subgroup unlearning in CLIP is evaluated along three dimensions. First, we must ensure that CLIP successfully forgets the images from the specified subgroup. Second, CLIP must continue to perform consistently on the retain set, which includes the other subgroups that are not intended to be forgotten. Third, CLIP must maintain strong zero-shot performance across various tasks and datasets. These considerations lead us to evaluate CLIP's performance not only on the targeted forgetting subgroups and unaffected subgroups, but also on various unseen datasets during the forgetting process, including Food [6], STL [15], and ObjectNet [2]. Additionally, we will use the post-unlearning CLIP to retrieve images in order to assess the impact of forgetting on its learned representations. Additionally, we use the classification accuracy ratio between the original and post-unlearned CLIP models to gauge the degree of knowledge forgetting and retention. More specifically, the ratio is computed as $\min\left(\frac{\text{Acc\_learn}}{\text{Acc}_{ori}}, 1\right)$ , where we cap the score at 1 to represent full knowledge restoration. This ratio is shown in the subscript of each result. We further + +Table 1. Evaluation results for forgetting the subclass 'marmoset' within the superclass 'monkey' (left) and 'box turtle' within the superclass 'turtle' (right) in the ImageNet-1K dataset are presented (Target). Each superclass contains four subgroups, with the remaining three designated as the retain set (Retain). + +
BackboneMethodImageNetCIFARFoodSTLObjectNetScoreBackboneMethodImageNetAll (↑)CIFARFoodSTLObjectNetScore
Target (↓)Retain (↑)All (↑)Target (↓)Retain (↑)
RN50Original51.054.759.870.460.992.068.9-RN50Original64.570.359.870.460.992.068.9-
FT2.5.4.987.1.100.025.6.42.813.5.19.217.4.28.543.1.46.819.6.58.451.6FT0.0.0.072.1.100.026.0.43.512.6.17.928.6.47.070.3.76.420.2.93.359.2
GA0.0.0.00.9.1.632.2.53.816.4.23.322.9.37.663.3.68.822.1.32.145.3GA0.0.0.013.3.18.942.1.70.414.6.20.740.0.55.777.8.84.621.3.30.955.8
Fisher0.2.0.30.5.0.91.3.2.110.4.14.70.2.0.310.5.11.43.0.14.419.1Fisher1.2.1.92.0.1.83.5.5.99.7.13.80.4.0.710.5.14.11.6.23.319.3
LIP0.7.1.40.2.3.61.3.2.210.7.15.20.2.0.310.3.11.21.6.2.318.6LIP0.4.0.63.4.4.815.2.25.410.6.15.16.1.10.023.7.25.811.3.16.428.1
EMMN0.0.0.056.7.100.028.4.47.513.1.18.620.4.33.554.9.59.720.5.29.755.6EMMN0.1.0.257.2.81.428.1.47.012.5.17.822.1.36.364.7.70.317.0.25.554.0
CLIP-LIP2.65.10.4.0.717.7.29.618.8.26.74.2.6.950.8.55.212.5.18.133.2CLIP-LIP1.9.2.92.0.2.831.9.53.319.5.27.76.9.11.363.6.69.134.2.69.644.4
Ours0.0.0.050.2.91.754.5.91.185.7.100.062.9.100.081.5.88.645.2.56.691.0Ours0.0.0.069.4.98.750.6.84.654.5.77.450.5.82.987.6.95.243.1.62.685.9
RN101Original66.252.662.364.457.392.469.2-RN101Original66.277.162.364.457.392.469.2-
FT1.3.2.075.2.100.027.6.44.311.8.18.34.4.7.635.1.48.017.3.25.047.3FT0.0.0.068.0.88.232.4.52.011.4.17.727.8.48.566.0.71.423.5.34.058.8
GA0.7.1.10.3.0.536.2.58.119.9.9.929.3.51.158.6.63.419.4.88.047.3GA0.0.0.022.1.28.746.4.74.512.2.18.940.4.70.575.6.81.825.9.37.458.8
Fisher0.9.1.40.4.0.70.3.0.511.7.18.20.2.0.39.6.10.31.0.1.418.6Fisher1.9.2.92.4.3.10.9.1.411.0.17.10.3.0.59.5.10.30.6.0.918.6
LIP0.2.0.30.3.0.52.0.3.210.4.16.10.2.0.310.7.11.61.9.2.719.2LIP0.81.23.2.4.212.9.20.710.3.16.08.4.14.716.9.18.314.7.21.227.7
EMMN0.7.1.162.5.100.030.1.48.322.0.35.110.8.18.810.0.10.823.8.34.449.5EMMN0.0.0.048.6.83.032.5.52.29.2.14.322.4.39.160.8.65.822.9.33.152.5
CLIP-LIP2.7.4.11.5.2.932.0.61.428.6.44.410.9.19.049.6.53.714.9.21.541.2CLIP-LIP4.0.6.910.8.14.038.1.61.238.5.59.817.6.02.766.8.72.318.6.26.951.1
Ours0.0.0.047.7.90.758.5.93.969.7.100.056.9.99.393.5.100.045.9.66.392.9Ours0.0.0.080.1.100.059.3.95.270.9.100.055.8.97.491.5.99.046.2.66.894.1
+ +![](images/014c762956b287042f5aeddaa43f90b13f9a217e31b85321e4cf8bb6c7568a5a.jpg) +Figure 3. Retrieval results on the ImageNet-1K dataset before and after the unlearning process in the CLIP model. We highlight the "marmoset" subgroup in red, with other "monkey" subgroups shown in green. + +compute the mean of these ratios across all classes, denoted as the Score, to provide a comprehensive evaluation metric. + +# 5.2. In-domain subgroup forgetting in ImageNet + +The Results are shown in Tab. 1. We observe that both the Fisher and LIP methods fail to optimize the model, resulting in severe catastrophic forgetting across the tested datasets. In comparison, FT, GA, EMMN, and CLIP-LIP perform slightly better but still suffer from catastrophic forgetting to varying degrees. For instance, optimizing only the retaining set with the FT strategy enables the CLIP model to forget the targeted subgroup while retaining knowledge of the intended subgroups. However, this approach leads to noticeable forgetting across other datasets, including different superclasses in ImageNet and unrelated datasets. Additionally, when applying gradient ascent to the forgetting subgroup with the GA strategy, CLIP exhibits forgetting within other subgroups of the same superclasses, and its zero-shot performance deteriorates due to catastrophic forgetting. Even with the inclusion of both forgetting and retaining samples in the fine-tuning process under the EMMN strategy, these issues remain inadequately addressed. By incorporating layer importance relative to the forgetting dataset to minimize the adverse ef + +fects of fine-tuning and leveraging a model merging strategy, we can preserve most of the model's zero-shot performance while ensuring that the targeted subgroup is forgotten. + +We also present the retrieval results using our fine-tuned CLIP model, which has forgotten the target subgroup, in Fig. 3. Before unlearning, when using the prompt 'monkey' to retrieve images, the model retrieves images from various subgroups, including 'marmoset' and 'howler monkey,' among others. Additionally, when prompted with specific subclass names, as shown in the second and third rows in figure, the model retrieves corresponding subgroup images. After unlearning, we observe that images of 'marmoset' are no longer retrieved when prompted with either the superclass name 'monkey' or the specific subgroup name 'marmoset,' indicating that the knowledge of 'marmoset' has been effectively erased from CLIP's parameter space. We also include retrieval results when using the prompt 'panda'; the consistent performance between the pre- and post-unlearning versions of CLIP shows that knowledge of other classes remains largely unaffected. + +Table 2. Evaluation results for forgetting the "airplane" (left) and "truck" (right) class in CIFAR-10 are presented. To assess the unaffected subgroup performance, we also separately retrieve the "airplane" class from ImageNet-1K (see the column labeled Target in ImageNet). The CIFAR-10 dataset is used exclusively as the retain set (see the column labeled Retain in CIFAR). + +
BackboneMethodCIFAR Target (↓)Retain (↑)FoodSTLObjectNetImageNet Target AllScoreBackboneMethodCIFAR Target (↓)Retain (↑)FoodSTLObjectNetImageNet Target AllScore
RN50Original81.370.460.992.068.953.659.8-RN50Original54.073.460.992.068.952.359.8-
FT0.00.092.6100.01.11.841.745.30.20.30.00.00.30.535.4FT0.00.056.376.71.32.129.311.80.40.60.00.00.20.330.2
GA0.00.014.210.21.01.614.115.30.10.10.20.40.10.219.7GA0.00.020.217.51.22.019.321.00.30.40.00.00.10.221.6
Fisher0.00.013.118.61.01.615.216.50.30.40.00.00.10.219.6Fisher0.00.012.416.91.22.013.921.00.20.40.00.00.20.219.2
LIP0.00.015.722.31.32.114.916.20.30.40.10.20.20.320.2LIP0.00.013.918.90.91.512.313.40.40.60.10.20.20.319.3
EMMN0.00.080.7100.01.62.641.645.20.20.30.00.00.50.835.6EMMN0.00.048.566.11.32.121.223.00.20.30.00.00.10.227.4
CLIP-LIP1.11.427.238.67.111.749.945.231.215.334.564.426.844.851.1CLIP-LIP3.87.045.361.716.627.374.280.733.648.828.953.339.566.161.8
Ours7.61365.102.551.985.285.793.257.984.049.892.950.183.888.9Ours0.91.763.266.150.582.984.491.756.491.951.585.852.187.189.5
RN101Original69.264.457.392.469.251.962.3-RN101Original91.479.657.392.469.250.662.3-
FT0.00.094.6100.01.11.958.463.20.40.60.00.00.20.338.0FT0.00.058.373.21.42.440.433.70.30.40.10.20.30.531.5
GA0.00.011.117.20.91.615.016.20.30.40.00.00.10.219.4GA0.00.020.225.41.22.119.320.90.20.30.00.00.10.221.3
Fisher0.00.012.920.01.01.716.918.30.30.40.10.20.20.320.1Fisher0.00.013.116.51.01.714.215.40.30.40.10.20.10.219.2
LIP0.00.013.821.41.62.815.717.00.20.30.00.00.30.520.3LIP0.00.015.719.71.62.815.717.00.30.40.00.00.20.320.0
EMMN0.00.081.2100.01.22.156.761.40.30.40.10.20.20.337.8EMMN0.00.048.360.71.42.435.848.70.10.10.10.20.20.328.9
CLIP-LIP7.410.735.585.110.418.263.168.330.432.920.699.719.831.849.5CLIP-LIP9.610.552.866.319.534.076.082.350.773.342.584.043.670.071.3
Ours8.812.763.208.148.885.287.794.958.184.045.788.141.266.186.2Ours3.23.570.388.446.280.686.693.758.384.243.986.846.674.886.4
+ +![](images/1acade394e65f2e4e936d3a22d8743ba67c0d202a90a441d755b675ffedfaeeb.jpg) +Figure 4. Retrieval results of the CLIP pre- and post-unlearning process. + +# 5.3. Cross-domain forgetting in CIFAR-10 + +We then validate the effectiveness of our method in forgetting images belonging to common classes in CIFAR-10, whose concepts also appear in other datasets, such as ImageNet. The main challenge lies in the significant distributional differences between CIFAR-10 and the pre-trained dataset, which can result in severe catastrophic forgetting during the finetuning process for unlearning. In this experiment, we select 'airplane' and 'truck' as the target groups to unlearn from CLIP. In addition to evaluating the zero-shot classification performance of these targeted classes in CIFAR-10, we also assess their performance in ImageNet to determine whether unlearning is accurately confined to the specified distribution, without affecting the entire concept. + +We present the results in Tab. 2. Compared to the results on ImageNet, models fine-tuned on the CIFAR-10 dataset exhibit more severe catastrophic forgetting, even with methods like FT, GA, and EMMN, which retain some zero-shot capability after forgetting specific classes in ImageNet. Notably, models fine-tuned on CIFAR-10 to retain performance demonstrate strong zero-shot performance on STL-10, a dataset with a distribution similar to CIFAR-10, highlighting + +the impact of the fine-tuning dataset on model performance. In contrast, our method shows remarkable robustness under such a significant distribution shift. Specifically, while the post-unlearning CLIP achieves random classification performance on the forgotten class, it retains at least $66.1\%$ of its original zero-shot performance. + +We also retrieve related classes in CIFAR-10 and ImageNet using the pre- and post-unlearning CLIP models for the 'airplane' class. As shown in Fig. 4, we observe that the post-unlearning CLIP is unable to retrieve airplanes from the CIFAR-10 dataset but can still retrieve airplanes from the ImageNet dataset, indicating precise forgetting limited to the target distribution. The consistency of retrieved results between the pre- and post-unlearning CLIP further demonstrates the effectiveness of our method in preserving unrelated knowledge within the model. + +# 5.4. Style forgetting + +We further evaluate performance on forgetting a specific style within a given class, representing a more fine-grained subgroup image unlearning challenge. Specifically, we use Stable Diffusion to generate or transform existing images into three distinct styles: poster art, black-and-white drawing, + +![](images/fd23cc594ac3c10f42f5bf2654b2e4b8c22697962ccb9529354d45f365c060ed.jpg) +Figure 5. Examples of stylized images of the sports car in the ImageNet-1K dataset. + +Table 3. Evaluation results of CLIP after forgetting sketch-style "auto-cars" using various methods. + +
MethodCIFARFoodSTLObjectNetImageNetScore
Target style (↓)Retain (↑)All (↑)
Original60.852.288.465.468.182.558.6-
FT10.116.61.93.610.511.93.96.099.5100.098.3100.01.11.920.0
GA46.776.839.876.283.294.157.888.42.02.950.361.052.188.983.2
Fisher10.517.31.52.915.918.01.11.70.10.10.50.60.20.320.1
LIP11.218.40.91.713.615.44.26.40.00.00.00.00.50.920.4
EMMN16.326.826.50.839.244.320.130.714.521.370.285.14.88.246.4
CLIP-LIP15.425.36.512.518.721.28.412.81.62.32.42.91.32.224.9
Ours49.280.943.082.480.891.458.990.12.13.166.080.051.988.687.2
+ +and sketch, as illustrated in Figure 5. We first fine-tune the CLIP on images with different styles, of which the zero-shot performance is reported at the first line in Tab. 3. Then, different unlearning methods are applied to erase the model's knowledge of cars in the sketch style, while retaining images of cars in the other styles. Results are shown in Table 3. + +Among the baselines, we find that GA performs the best, effectively forgetting the targeted style while retaining most unrelated knowledge, as indicated by its strong zero-shot performance across different datasets. In comparison, simply fine-tuning CLIP on the retaining set leads to severe catastrophic forgetting, as shown in the results of FT and EMMN. The EMMN results also suggest that incorporating the retaining set into the unlearning process is not always beneficial. Although using the retaining set could potentially reduce model over-forgetting, an imprecise selection of the retaining set and an imbalanced loss between it and the forgetting set can degrade the model's generalization ability on other datasets. By selecting the optimal layer for fine-tuning to minimize the impact on the model's original performance, our method achieves the best overall results, maintaining at least $80.4\%$ of the model's zero-shot capacity. + +# 5.5. Ablation study and discussion + +On the importance of reminding stage. The reminding stage follows the forgetting stage, using the retaining set to partially recover any unintended knowledge loss. Notably, our retaining set typically differs from the original pre-training set, potentially further compromising CLIP's zero-shot classification capability. To examine this, we conduct an ablation study on the impact of iteration count during the reminding stage. Using CIFAR-10 as the unlearning + +Table 4. Ablation study on the used number of steps at retaining stage on the zero-shot capacity of CLIP. We set the batch size as 64 for fine-tuning and $\alpha$ as 0.65 for model merging. + +
# steps51020304050100
Accr59.863.265.641.435.434.932.5
Accin44.941.238.125.88.52.10.8
+ +Table 5. Ablation study on the $\alpha$ for the model merging operation at the restoring stage to the zero-shot capacity of CLIP. + +
α0.10.30.50.650.70.9
Accf0.00.02.48.812.868.8
Accr22.727.542.763.266.778.7
Accin1.04.422.3941.243.560.5
+ +dataset for CLIP, we evaluate performance on both the retain set $(\mathrm{ACC}_r)$ and ImageNet $(\mathrm{ACC}_{in})$ . As shown in Tab. 4, we observe that with an increasing number of steps, classification accuracy on the retain set initially improves before declining, while accuracy on the ImageNet-1K dataset consistently decreases, with a significant drop at higher iteration counts. This suggests that while additional steps in the reminding stage help the model relearn knowledge that was unintentionally forgotten, they also lead to a larger distribution shift from the original parameter space, ultimately impairing zero-shot performance. + +On the model merging. In our work, we use model merging to fully restore zero-shot capacity. While our method leverages a calibration dataset to identify the optimal coefficient, $\alpha$ , we also conduct an additional experiment by directly varying the coefficients to evaluate both forgetting performance $(\mathrm{ACC}_f)$ and recovery performance $(\mathrm{ACC}_r / \mathrm{ACC}_{in})$ . As shown in Tab. 5, achieving an optimal $\alpha$ is crucial to balancing forgetting and retaining performance. With increasing $\alpha$ , the model's retaining performance and zero-shot classification capacity improve continuously. However, a large $\alpha$ can also cause the model to relearn data that should be forgotten, as indicated by the $68.8\%$ accuracy on the forgetting set when $\alpha$ is set to 0.9. + +# 6. Conclusion + +In this paper, we study the problem of forgetting a subgroup of images in the CLIP model. We propose a three-stage method consisting of the forgetting, retaining, and restoring steps. Our method does not require any pre-trained dataset, allowing it to achieve a balanced approach by selectively forgetting the targeted subgroup of images while preserving the knowledge that should not be forgotten. Experimental results validate the effectiveness of our method and highlight the challenges of this task, as most classical methods fail. + +# References + +[1] Andrew Bai, Chih-Kuan Yeh, Cho-Jui Hsieh, and Ankur Taly. Which pretrain samples to rehearse when finetuning pretrained models? arXiv preprint arXiv:2402.08096, 2024. 2 +[2] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 5 +[3] Alexander Becker and Thomas Liebig. Evaluating machine unlearning via epistemic uncertainty. arXiv preprint arXiv:2208.10836, 2022. 2 +[4] Yoshua Bengio, Aaron C Courville, and Pascal Vincent. Unsupervised feature learning and deep learning: A review and new perspectives. CoRR, abs/1206.5538, 1(2665):2012, 2012. 1 +[5] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. Advances in Neural Information Processing Systems, 36, 2024. 2 +[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5 +[7] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141-159. IEEE, 2021. 1 +[8] Zikui Cai, Yaoteng Tan, and M Salman Asif. Single layer single gradient unlearning. arXiv preprint arXiv:2407.11867, 2024.3 +[9] Hao Chen, Bhiksha Raj, Xing Xie, and Jindong Wang. On catastrophic inheritance of large foundation models. arXiv preprint arXiv:2402.01909, 2024. 1 +[10] Ruizhe Cheng, Bichen Wu, Peizhao Zhang, Peter Vajda, and Joseph E Gonzalez. Data-efficient language-supervised zero-shot learning with self-distillation. In CVPR, pages 3119-3124, 2021. 1 +[11] Eli Chien, Chao Pan, and Olgica Milenkovic. Efficient model updates for approximate unlearning of graph-structured data. In The Eleventh International Conference on Learning Representations, 2022. 1 +[12] Vikram S Chandawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security, 18:2345-2354, 2023. 2, 5 +[13] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 5 +[14] Lluc Clavera Comas. Machine unlearning: fisher infomation matrix and selective forgetting in deep networks. B.S. thesis, Universitat Politècnica de Catalunya, 2024. 3 + +[15] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 5 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5 +[17] Ronen Eldan and Mark Russinovich. Who's harry potter? approximate unlearning in llms. arXiv preprint arXiv:2310.02238, 2023. 2 +[18] Jack Foster, Kyle Fogarty, Stefan Schoepf, Cengiz Öztireli, and Alexandra Brintrup. Zero-shot machine unlearning at scale via lipschitz regularization. arXiv preprint arXiv:2402.01401, 2024. 2, 5 +[19] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9304–9312, 2020. 2 +[20] Jinu Gong, Osvaldo Simeone, and Joonhyuk Kang. Bayesian variational federated learning and unlearning in decentralized networks. In 2021 IEEE 22nd International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pages 216-220. IEEE, 2021. 1 +[21] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. NeurIPS, 36, 2024. 1 +[22] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In AAAI, pages 746-754, 2023. 1 +[23] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. *NeurIPS*, 34:16319-16330, 2021. 1 +[24] Kris Holt. Researchers found child abuse material in the largest ai image generation dataset, 2024. Accessed: 2024-10-25. 3 +[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4 +[26] Mark He Huang, Lin Geng Foo, and Jun Liu. Learning to unlearn for robust machine unlearning. arXiv preprint arXiv:2407.10494, 2024. 2 +[27] Zhehao Huang, Xinwen Cheng, JingHao Zheng, Haoran Wang, Zhengbao He, Tao Li, and Xiaolin Huang. Unified gradient-based machine unlearning with remain geometry enhancement. arXiv preprint arXiv:2409.19732, 2024. 1 +[28] Frank Klinker. Exponential moving average versus moving exponential average. Mathematische Semesterberichte, 58: 97-107, 2011. 5 +[29] Alexey Kravets and Vinay P. Namboodiri. Zero-shot CLIP class forgetting via text-image space adaptation. Transactions on Machine Learning Research, 2025. 2, 3, 5 + +[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[31] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In CVPR, pages 20071-20082, 2023. 1 +[32] Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, PRANAY SHARMA, Sijia Liu, et al. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 5 +[33] Ji Liu, Jiaxiang Ren, Ruoming Jin, Zijie Zhang, Yang Zhou, Patrick Valduriez, and Dejing Dou. Fisher information-based efficient curriculum federated learning with large language models. arXiv preprint arXiv:2410.00131, 2024. 3 +[34] Yufang Liu, Changzhi Sun, Yuanbin Wu, and Aimin Zhou. Unlearning with fisher masking. arXiv preprint arXiv:2310.05331, 2023. 3 +[35] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Xi-angyang Ji, Antoni Chan, and Rong Jin. Improved fine-tuning by better leveraging pre-training data. Advances in Neural Information Processing Systems, 35:32568-32581, 2022. 2 +[36] Alexander Ly, Maarten Marsman, Josine Verhagen, Raoul PPP Grasman, and Eric-Jan Wagenmakers. A tutorial on fisher information. Journal of Mathematical Psychology, 80:40-55, 2017. 3 +[37] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N Ravi. Deep unlearning via randomized conditionally independent hessian. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10422-10431, 2022. 1 +[38] Quoc Phong Nguyen, Bryan Kian Hsiang Low, and Patrick Jaillet. Variational bayesian unlearning. Advances in Neural Information Processing Systems, 33:16025-16036, 2020. 1 +[39] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. *Neural networks*, 113:54–71, 2019. 2 +[40] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara, et al. Safe-clip: Removing nsfw concepts from vision-and-language models. In Proceedings of the European Conference on Computer Vision, 2024. 3 +[41] Clifton Poth, Jonas Pfeiffer, Andreas Rückle, and Iryna Gurevych. What to pre-train on? efficient intermediate task selection. arXiv preprint arXiv:2104.08247, 2021. 2 +[42] Jenny Quang. Does training ai violate copyright law? Berkeley Tech. LJ, 36:1407, 2021. 1 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3 +[44] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, pages 2765-2775, 2023. 1 + +[45] Shibani Santurkar, Dimitris Tsipras, and Aleksander Madry. Breeds: Benchmarks for subpopulation shift. arXiv preprint arXiv:2008.04859, 2020. 5 +[46] Sebastian Schelter, Mozhdeh Ariannezhad, and Maarten de Rijke. Forget me now: Fast and exact unlearning in neighborhood-based recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2011–2015, 2023. 1 +[47] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 1, 2, 3 +[48] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 34:18075-18086, 2021. 1 +[49] Adam Shostack. The boy who survived: Removing harry potter from an llm is harder than reported. arXiv preprint arXiv:2403.12082, 2024. 2 +[50] Shashank Subramanian, Peter Harrington, Kurt Keutzer, Wahid Bhimji, Dmitriy Morozov, Michael W Mahoney, and Amir Gholami. Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. NeurIPS, 36, 2024. 1 +[51] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[52] Anvith Thudi, Gabriel Deza, Varun Chandrasekaran, and Nicolas Papernot. Unrolling sgd: Understanding factors influencing machine unlearning. In 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P), pages 303-319. IEEE, 2022. 2, 5 +[53] Enayat Ullah, Tung Mai, Anup Rao, Ryan A Rossi, and Raman Arora. Machine unlearning via algorithmic stability. In Conference on Learning Theory, pages 4126-4142. PMLR, 2021. 1 +[54] Alexander Warnecke, Lukas Pirch, Christian Wressnegger, and Konrad Rieck. Machine unlearning of features and labels. In 30th Annual Network and Distributed System Security Symposium, NDSS 2023, San Diego, California, USA, February 27 - March 3, 2023. The Internet Society, 2023. 2, 5 +[55] Yixuan Wei, Han Hu, Zhenda Xie, Ze Liu, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Improving clip fine-tuning performance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5439-5449, 2023. 2 +[56] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 5 + +[57] Tomoya Yamashita, Masanori Yamada, and Takashi Shibata. One-shot machine unlearning with mnemonic code. In The 16th Asian Conference on Machine Learning (Conference Track). 3 +[58] Haonan Yan, Xiaoguang Li, Ziyao Guo, Hui Li, Fenghua Li, and Xiaodong Lin. Arcane: An efficient architecture for exact machine unlearning. In *IJCAI*, page 19, 2022. 1 +[59] Kaiyu Yang, Clint Qinami, Li Fei-Fei, Jia Deng, and Olga Russakovsky. Towards fairer datasets: Filtering and balancing the distribution of the people subtree in the imagenet hierarchy. In Proceedings of the 2020 conference on fairness, accountability, and transparency, pages 547–558, 2020. 1 +[60] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. NeurIPS, 36, 2024. 4 +[61] Binchi Zhang, Yushun Dong, Tianhao Wang, and Jundong Li. Towards certified unlearning for deep neural networks. arXiv preprint arXiv:2408.00920, 2024. 1 +[62] Jiahao Zhang. Graph unlearning with efficient partial retraining. In Companion Proceedings of the ACM on Web Conference 2024, pages 1218-1221, 2024. 1 +[63] Zeliang Zhang, Mingqian Feng, Zhiheng Li, and Chenliang Xu. Discover and mitigate multiple biased subgroups in image classifiers. In CVPR, 2024. 1 +[64] Zeliang Zhang, Zhuo Liu, Mingqian Feng, and Chenliang Xu. Can clip count stars? an empirical study on quantity bias in clip. In EMNLP (Findings), 2024. 1 \ No newline at end of file diff --git a/2025/Targeted Forgetting of Image Subgroups in CLIP Models/images.zip b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e32c57c22cd7301fe4c07dcdaeb60bf89e972c2b --- /dev/null +++ b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2693e9d6000a7ea1a442102e29731fdc676fde5f7661d9839df12bcfbe67d2cc +size 598150 diff --git a/2025/Targeted Forgetting of Image Subgroups in CLIP Models/layout.json b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3151f822ff5987dd1d97ec68dafc0e93bed51add --- /dev/null +++ b/2025/Targeted Forgetting of Image Subgroups in CLIP Models/layout.json @@ -0,0 +1,8424 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 128, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 128, + 103, + 483, + 121 + ], + "type": "text", + "content": "Targeted Forgetting of Image Subgroups in CLIP Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "content": "Zeliang Zhang" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "content": " Gaowen Liu" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "content": " Charles Fleming" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "content": " Ramana Rao Kompella" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "content": " Chenliang Xu" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "content": "University of Rochester " + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 541, + 171 + ], + "type": "text", + "content": "Cisco Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 173, + 527, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 173, + 527, + 186 + ], + "spans": [ + { + "bbox": [ + 80, + 173, + 527, + 186 + ], + "type": "text", + "content": "{zeliang.zhang, chenliang.xu}@rochester.edu, {gaoliu, chflemin, rkompell}@cisco.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 297, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 574 + ], + "type": "text", + "content": "Foundation models (FMs) such as CLIP have demonstrated impressive zero-shot performance across various tasks by leveraging large-scale, unsupervised pre-training. However, they often inherit harmful or unwanted knowledge from noisy internet-sourced datasets, compromising their reliability in real-world applications. Existing model unlearning methods either rely on access to pre-trained datasets or focus on coarse-grained unlearning (e.g., entire classes), leaving a critical gap for fine-grained unlearning. In this paper, we address the challenging scenario of selectively forgetting specific portions of knowledge within a class—without access to pre-trained data—while preserving the model's overall performance. We propose a novel three-stage approach that progressively unlearns targeted knowledge while mitigating over-forgetting. It consists of (1) a forgetting stage to fine-tune the CLIP on samples to be forgotten, (2) a reminding stage to restore performance on retained samples, and (3) a restoring stage to recover zero-shot capabilities using model souping. Additionally, we introduce knowledge distillation to handle the distribution disparity between forgetting/retaining samples and unseen pre-trained data. Extensive experiments on CIFAR-10, ImageNet-1K, and style datasets demonstrate that our approach effectively unlearns specific subgroups while maintaining strong zero-shot performance on semantically similar subgroups and other categories, significantly outperforming baseline unlearning methods, which lose effectiveness under the CLIP unlearning setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 597, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 597, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 597, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "content": "Foundation models (FMs) are deep neural networks pretrained on large-scale datasets [50]. By employing unsupervised training, these models capture complex patterns and relationships within the data, enabling a deep understanding of various domains [4]. For example, CLIP [43], a vision-language model, learns correspondences between natural language and images. Trained on the large-scale LAION-5B dataset [47], CLIP demonstrates strong zero" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 212, + 555, + 355 + ], + "blocks": [ + { + "bbox": [ + 318, + 212, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 318, + 212, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 318, + 212, + 555, + 355 + ], + "type": "image", + "image_path": "ba6cd99b7c5621bd4b277f2eed7cbf59e04a64fe32deca2be9df55c53528bb89.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 365, + 556, + 443 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 443 + ], + "type": "text", + "content": "Figure 1. In the task of CLIP unlearning, classical machine unlearning methods often fail to generalize across multiple datasets due to limited access to the original pre-training data. In contrast, our proposed method does not require access to the pre-training dataset. After unlearning, CLIP not only forgets the specified subgroup of images but also maintains consistent performance on various unseen datasets in the zero-shot classification setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 470, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 556, + 579 + ], + "type": "text", + "content": "shot performance [10, 22, 44]. However, its reliability in real-world applications is questionable [63, 64]. The training data, sourced from the internet, includes uncensored content such as discriminatory and violent imagery, copyright violations, and personal information [21, 42, 59]. This problematic content can be unintentionally learned during pre-training, raising concerns about the reliability and safety of downstream models [9, 31]. Therefore, it is critical to remove such harmful knowledge to ensure trustworthy FMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Numerous studies have explored model unlearning, which can be broadly divided into exact [46, 53, 58] and approximate strategies [11, 23, 38]. Exact unlearning aims to completely eliminate the influence of forgotten data from the model, often by retraining it [7, 58, 62], ensuring the model behaves as if the data had never been included in training. While thorough, this approach is computationally expensive and impractical for large-scale models [32, 48]. In contrast, approximate unlearning offers a more efficient alternative, using techniques like variational inference [20, 38] or adaptive regularization [27, 37, 61] to partially reduce the data's" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9870" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "impact. However, these methods face several challenges when applied to FMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 295, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 295, + 253 + ], + "type": "text", + "content": "First, as shown in Fig. 1, the training data for FMs is often vast or inaccessible [5, 47], rendering exact unlearning impractical. Second, while FMs are trained on large-scale datasets using unsupervised learning, the fine-tuning typically used in approximate unlearning strategies [26, 51] can degrade generalization across other knowledge domains [1, 35, 41], leading to significant performance drops, particularly in zero-shot tasks [55]. Third, the knowledge to be forgotten is often deeply entangled with other learned information [17, 49]. This entanglement makes it difficult to isolate and remove specific knowledge from the model parameters, particularly when distributional disparities exist between the data we want to unlearn and the pre-trained data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 445 + ], + "type": "text", + "content": "Few studies have specifically addressed the challenge of unlearning in foundational models, particularly in CLIP. Most existing approaches focus on unlearning entire classes or broad concepts, yet real-world applications often require a more targeted approach, where only certain portions of knowledge within a class are forgotten. For example, we may want the model to forget representations of Disney's Mickey Mouse while retaining versions created by the public, a task known as subgroup image forgetting. Moreover, many current unlearning methods rely on access to the original pre-trained dataset to restore model performance, which is often unavailable. This limitation underscores a critical question in current research: how can models like CLIP achieve selective, fine-grained unlearning without relying on pre-trained data, while still maintaining generalization across other tasks?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 616 + ], + "type": "text", + "content": "In this work, we bridge this gap by tackling a more challenging and practical scenario: enabling CLIP to selectively forget specific portions of knowledge within a single class without access to pre-trained data, framing this as a fine-grained unlearning task. Rather than unlearning an entire class or concept (e.g., all images of airplanes), we focus on selectively forgetting images from a particular subset (e.g., Boeing airplanes) while preserving the model's knowledge of other subsets (e.g., different manufacturers). This selective forgetting is particularly challenging due to the feature entanglement across subsets within a class, and it is exceedingly difficult to define and collect all subsets within the same class in a foundational model setting, which is trained on vast, large-scale data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "To address the challenge of partial knowledge removal, we propose a novel, three-stage approach: forgetting, reminding, and restoring. In the forgetting stage, we identify layers essential for representing the forgetting samples but less critical to other data, fine-tuning the model on these samples to selectively remove targeted knowledge. We observe that models can over- forget samples from similar distributions, so we introduce a retaining set with a similar distribution to the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": "targeted subgroup, aiding layer identification and supporting the reminding stage. During reminding, we fine-tune the model on the retaining samples to reinforce CLIP's performance on data prone to over-forgetting due to distributional similarity. In the final restore stage, we employ model souping to fully recover CLIP's zero-shot performance. Throughout this process, we address the distributional gap between forgetting/retaining samples and unseen pre-trained data by distilling knowledge from CLIP, achieving a more precise representation without relying on the original pre-trained dataset. Extensive experiments demonstrate our method's effectiveness in selectively unlearning subgroup images while preserving robust zero-shot performance across diverse distributions and datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 328, + 240, + 504, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 240, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 328, + 240, + 504, + 251 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 251, + 553, + 405 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 315, + 251, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 251, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 251, + 553, + 274 + ], + "type": "text", + "content": "1. We formulate the problem of subgroup image unlearning within the CLIP model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 276, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 553, + 334 + ], + "type": "text", + "content": "2. We propose a three-stage method for unlearning in CLIP, consisting of forgetting, reminding, and restoring. Our method is data-efficient, which does not require access to CLIP's pre-training dataset, making it more practical for real-world applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 335, + 553, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 335, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 553, + 405 + ], + "type": "text", + "content": "3. We conduct extensive experiments to validate the effectiveness of our proposed method, including unlearning tasks on CIFAR-10, ImageNet-1K, and style forgetting. Additionally, we demonstrate the scalability of our approach across various applications through retrieval-based evaluations." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 418, + 398, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 398, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 398, + 430 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 713 + ], + "type": "text", + "content": "Machine Unlearning. Various strategies have been developed to facilitate machine unlearning in models. Notably, Golatkar et al. [19], Warnecke et al. [54] employ a method of continuous fine-tuning (FT) on the retain dataset—those data elements the model is expected to remember. This approach not only approximates the forgetting of the target dataset but also mitigates the issue of overforgetting [32, 39], in contrast to exclusively fine-tuning on the forgetting dataset. Becker and Liebig [3], Liu et al. [32] leverage Fisher information, which assesses the sensitivity of parameters concerning the forgetting dataset, to achieve model unlearning. Meanwhile, Thudi et al. [52] unroll the knowledge embedded in the forgetting dataset from the model parameters using stochastic gradient ascent (GA). Further, [18] introduces the use of local Lipschitz (LIP) perturbations on the forgetting dataset as a technique for unlearning in image classification models, which Kravets and Namboodiri [29] subsequently adapts to the CLIP model. Additionally, Chandawat et al. [12] proposes an error minimization-maximization (EMMN) framework that aims to maximize errors on the forgetting dataset while minimizing them on the retraining dataset, thus achieving a balance between forgetting and retaining performance. Among these methods, EMMN [12] and LIP [29] stand out" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "text", + "content": "9871" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "as they do not require access to the training datasets (both forgetting and retaining), unlike other approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 298, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 298, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 298, + 348 + ], + "type": "text", + "content": "On the Importance of Unlearning in CLIP. Trained on a billion-scale dataset of image-text pairs [47], the CLIP model [43] excels in zero-shot classification tasks. Despite its strengths, the presence of uncensored samples within the training data has raised practical concerns [24]. Given the impracticality of retraining the CLIP model from scratch due to significant resource demands and the inaccessibility of the large-scale dataset, it is imperative to explore machine unlearning methods to selectively erase problematic knowledge from the model. Poppi et al. [40] suggest an approach to first distinguish between safe and unsafe content, followed by fine-tuning CLIP to alter the embeddings of unsafe content. Meanwhile, Cai et al. [8] focuses on identifying and unlearning from the most influential layer in CLIP, using a subset of the training set to restore zero-shot capabilities. [29] expand zero-shot machine unlearning techniques to the CLIP model, enabling the forgetting of entire concepts as indicated by prompts. Concurrently, Kravets and Namboodiri [29] apply Layer-wise Relevance Propagation (LoRA) to the text encoder of CLIP, specifically fine-tuning this component to unlearn a given concept within the model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 357, + 182, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 182, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 182, + 370 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "text", + "content": "The CLIP model, denoted " + }, + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "inline_equation", + "content": "g(\\cdot, \\cdot) = \\{g^{\\mathrm{img}}(\\cdot), g^{\\mathrm{txt}}(\\cdot)\\}" + }, + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "text", + "content": ", processes image and text inputs into embeddings, " + }, + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "inline_equation", + "content": "e^{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "inline_equation", + "content": "e^{\\mathrm{txt}}" + }, + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "text", + "content": ". It aims for high cosine similarity between these embeddings when they are correctly paired, and low similarity otherwise, using a threshold " + }, + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 377, + 296, + 436 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": "Suppose a large dataset " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "D = \\{x_{n}^{\\mathrm{img}}, x_{n}^{\\mathrm{txt}}, c_{n}, g_{n}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": ", where we denote " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": " as the image, " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "x^{txt}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": " as the paired text learned by the CLIP, " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": " as the coarse main category, and " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": " is the fine-grained subclass label. The forgetting dataset " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "D^{f} = \\{x_{n}^{\\mathrm{img}}, x_{n}^{\\mathrm{txt}}, c_{n}, g_{n}\\}_{n=1}^{N_{f}}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": " is usually a small set compared with " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": ". The goal is to unlearn the associations between " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "x^{txt}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "D^{f}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": " without impacting the model's performance on the remaining dataset " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "D^{r} = D / D_{f}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": ", namely the retain set. The task involves selectively erasing knowledge from " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "D^{f}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": " while maintaining performance on " + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "inline_equation", + "content": "D^{r}" + }, + { + "bbox": [ + 55, + 437, + 296, + 570 + ], + "type": "text", + "content": ", which can be formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 576, + 296, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 576, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 70, + 576, + 296, + 605 + ], + "type": "interline_equation", + "content": "\\forall \\left(x _ {i} ^ {\\mathrm {i m g}}, x ^ {\\mathrm {t x t}}\\right) \\in D ^ {f}, \\frac {\\hat {g} ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot \\hat {g} ^ {\\mathrm {t x t}} \\left(x ^ {\\mathrm {t x t}}\\right)}{\\| \\hat {g} ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot \\hat {g} ^ {\\mathrm {t x t}} \\left(x ^ {\\mathrm {t x t}}\\right) \\|} < \\delta , \\tag {1}", + "image_path": "0b9acc4dc466a6dab909fa7efe4e6897e73b8bcf378f7bd71cec3ee23e0e64ba.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 607, + 269, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 607, + 269, + 623 + ], + "spans": [ + { + "bbox": [ + 76, + 607, + 269, + 623 + ], + "type": "interline_equation", + "content": "\\forall (x _ {i} ^ {\\mathrm {i m g}}, x _ {i} ^ {\\mathrm {t x t}}) \\in D ^ {r}, \\hat {g} (x _ {i} ^ {\\mathrm {i m g}}, x _ {i} ^ {\\mathrm {t x t}}) \\approx g (x _ {i} ^ {\\mathrm {i m g}}, x _ {i} ^ {\\mathrm {t x t}}),", + "image_path": "fd7bc3f94547459b3f58116b4ce466f452577acce9caedfad0cae3e6f5c5ba7d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "content": " represents the unlearned CLIP model and " + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "content": " the original model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 677 + ], + "type": "text", + "content": "Implementing the approach outlined in Eq. (1) is challenging due to several factors:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "text", + "content": "1. The large dataset " + }, + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "text", + "content": " used for training " + }, + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "text", + "content": " is inaccessible, and similarly, " + }, + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "text", + "content": " is also unavailable. We can only access a small subset of " + }, + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 56, + 677, + 296, + 713 + ], + "type": "text", + "content": " to recover model performance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 553, + 95 + ], + "type": "text", + "content": "after forgetting, blocking the direct use of conventional machine unlearning methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 96, + 555, + 203 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 96, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 155 + ], + "type": "text", + "content": "2. In addition, we may only have access to a coarse label " + }, + { + "bbox": [ + 314, + 96, + 553, + 155 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 314, + 96, + 553, + 155 + ], + "type": "text", + "content": " as the surrogate of " + }, + { + "bbox": [ + 314, + 96, + 553, + 155 + ], + "type": "inline_equation", + "content": "x^{txt}" + }, + { + "bbox": [ + 314, + 96, + 553, + 155 + ], + "type": "text", + "content": " rather than a fine-grained subgroup label " + }, + { + "bbox": [ + 314, + 96, + 553, + 155 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 314, + 96, + 553, + 155 + ], + "type": "text", + "content": ". This limitation increases the risk of inaccurately erasing targeted knowledge and over-forgetting unrelated knowledge that is similar to the target." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": "3. There may be a substantial distribution shift between " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": ". With a large number of parameters, directly finetuning on the small dataset " + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 313, + 156, + 555, + 203 + ], + "type": "text", + "content": " could cause the model to overfit and lose generalization ability." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 204, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 264 + ], + "type": "text", + "content": "Thus, the goal of this task is twofold: it involves selectively forgetting only specific image subgroups because of coarse labeling, while simultaneously preserving the model's robust zero-shot capabilities across various datasets, including those characterized by similarly coarse labels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 274, + 397, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 397, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 397, + 288 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 294, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 331 + ], + "type": "text", + "content": "To address the challenge of subgroup image unlearning, we propose a novel methodology that comprises three stages: forgetting, reminding, and restoring, as depicted in Fig. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 337, + 387, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 387, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 387, + 350 + ], + "type": "text", + "content": "4.1. Forgetting" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 354, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 545 + ], + "type": "text", + "content": "Previous studies [14, 34] indicate that selectively fine-tuning critical layers for a target dataset can effectively mitigate over-forgetting [33], typically using Fisher information [36, 57] to assess the importance of each layer. However, while Fisher information measures parameter sensitivity to the target dataset, it overlooks potential impacts on other datasets, which are inaccessible in our task. Our experiments reveal that relying solely on the forgetting dataset to modify the CLIP model often leads to catastrophic over-forgetting on similar subgroups with distributions akin to the forgetting data, as shown in Tab. 1 for the CLIP model optimized by GA. To address this, we propose using relative Fisher information which combines both the forgetting and retaining samples as a more nuanced metric for identifying essential layers for fine-tuning, thereby enhancing the precision of the unlearning process." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "content": "Concretely, given the forgetting dataset " + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^f = \\{(x_i^{\\mathrm{img}}, x^{\\mathrm{txt}})\\}" + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "content": ", we manually construct a retain dataset " + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^r = \\{(x_i^{\\mathrm{img}}, x_i^{\\mathrm{txt}})\\}" + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "content": " with a distribution similar to " + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^f" + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "content": ", used for alleviating the aforementioned subgroup over-forgetting. Here, " + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^r" + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "content": " is comprised of similar subgroups of images that we aim to preserve in the model's memory. In practical scenarios, we often use coarse labels as surrogate image descriptions, so " + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "content": " from the forgetting dataset may also appear in " + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^r" + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "content": ". We calculate the relative Fisher information for the " + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "content": "-th layer in " + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "inline_equation", + "content": "g^{\\mathrm{img}}(\\cdot)" + }, + { + "bbox": [ + 313, + 545, + 556, + 668 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 675, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 675, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 326, + 675, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {I} ^ {l} = \\frac {\\mathbb {E} _ {\\left(x _ {i} ^ {\\mathrm {i m g}} , x ^ {\\mathrm {t x t}}\\right) \\sim \\mathcal {D} ^ {f}} \\left[ \\nabla_ {\\theta^ {l}} ^ {2} \\left(g ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot g ^ {\\mathrm {t x t}} \\left(x ^ {\\mathrm {t x t}}\\right)\\right) \\right]}{\\mathbb {E} _ {\\left(x _ {i} ^ {\\mathrm {i m g}} , x _ {i} ^ {\\mathrm {t x t}}\\right) \\sim \\mathcal {D} ^ {r}} \\left[ \\nabla_ {\\theta^ {l}} ^ {2} \\left(g ^ {\\mathrm {i m g}} \\left(x _ {i} ^ {\\mathrm {i m g}}\\right) \\cdot g ^ {\\mathrm {t x t}} \\left(x _ {i} ^ {\\mathrm {t x t}}\\right)\\right) \\right]}, \\tag {2}", + "image_path": "a647de81ec64d1e9df003d757c13eba2f1a21727bbc98cec2a5b8308e643557a.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9872" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 70, + 504, + 240 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 504, + 240 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 504, + 240 + ], + "type": "image", + "image_path": "66fbaebd037d2b8e98bc47ce995193b54015a996c2a3098f5f397f2e95dd3b0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": "Figure 2. Overview of our method to unlearn the CLIP on a subgroup of images. We first compute the relative fisher information on the forgetting dataset and retain dataset to evaluate the layer importance. We select important layers for the following fine-tuning process. During the fine-tuning process for knowledge reminding, we align the distribution of our retain dataset to that of the pre-training dataset learned by the model. A model merging operation is performed at the last step of knowledge restoration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 313, + 263, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 263, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 263, + 325 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 313, + 263, + 325 + ], + "type": "inline_equation", + "content": "\\theta^l" + }, + { + "bbox": [ + 55, + 313, + 263, + 325 + ], + "type": "text", + "content": " represents the parameters of the " + }, + { + "bbox": [ + 55, + 313, + 263, + 325 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 313, + 263, + 325 + ], + "type": "text", + "content": "-th layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 326, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 433 + ], + "type": "text", + "content": "We then rank the importance of different layers based on the computed relative Fisher information, as detailed in Eq. (2). Layers with higher relative importance are selected for targeted unlearning, with all other layers frozen to minimize the impact on non-target knowledge. In our experiments, applying the low-rank adaptation (LoRA) technique [25] to the selected layers significantly enhances the targeted unlearning performance while minimizing effects on other retained knowledge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 458 + ], + "type": "text", + "content": "Finally, we fine-tune our model on the forgetting dataset " + }, + { + "bbox": [ + 55, + 434, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 55, + 434, + 296, + 458 + ], + "type": "text", + "content": " to minimize the following loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 468, + 296, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 468, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 88, + 468, + 296, + 498 + ], + "type": "interline_equation", + "content": "\\min \\mathcal {L} _ {f} = \\sum_ {i} \\frac {g ^ {i m g} \\left(x _ {i} ^ {i m g}\\right) \\cdot g ^ {t x t} \\left(x ^ {t x t}\\right)}{\\| g ^ {i m g} \\left(x _ {i} ^ {i m g}\\right) \\cdot g ^ {t x t} \\left(x ^ {t x t}\\right) \\|}, \\tag {3}", + "image_path": "37bf8b323f0d1aaeabcdde96fd99ff100a42d99822e3d42acd880c6b46a0cfc6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 507, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 530 + ], + "type": "text", + "content": "where images in the forgetting dataset are associated through the same coarse label " + }, + { + "bbox": [ + 55, + 507, + 295, + 530 + ], + "type": "inline_equation", + "content": "x^{txt}" + }, + { + "bbox": [ + 55, + 507, + 295, + 530 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 532, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 651 + ], + "type": "text", + "content": "By fine-tuning the CLIP model on images marked for unlearning with coarse labels, we aim to selectively erase specific knowledge. However, this adjustment often leads to catastrophic overforgetting, affecting the model's ability across both similar and different image distributions. For example, attempting to forget the \"airplane\" class in CIFAR-10 may also affect the recognition of other classes within CIFAR-10 and in ImageNet-1K. To tackle these challenges, particularly when full access to the pre-training dataset is unavailable, we implement reminding and restoration stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 130, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 130, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 130, + 673 + ], + "type": "text", + "content": "4.2. Reminding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "An intuitive strategy to retain non-targeted knowledge in CLIP is to fine-tune it on our retaining set, built during the forgetting stage for computing " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". However, two challenges" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 313, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 384 + ], + "type": "text", + "content": "hinder the direct application of this fine-tuning step. First, our retaining set may have a different distribution than the pre-training dataset, potentially causing inefficient learning. Second, our retaining set is typically limited, and fine-tuning on it could lead to further catastrophic over-forgetting in various unseen datasets and tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 385, + 556, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 556, + 483 + ], + "type": "text", + "content": "To address the first challenge, inspired by research showing that training data features are encoded in model parameters [60], we employ an optimization-based strategy to align the retaining dataset's distribution with the pre-training dataset. Specifically, for a given batch of image-text pairs in the retaining dataset " + }, + { + "bbox": [ + 313, + 385, + 556, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r = \\{(x_i^{\\mathrm{img}}, x_i^{\\mathrm{xt}})\\}_{i=1}^B" + }, + { + "bbox": [ + 313, + 385, + 556, + 483 + ], + "type": "text", + "content": ", we introduce a batch of perturbations " + }, + { + "bbox": [ + 313, + 385, + 556, + 483 + ], + "type": "inline_equation", + "content": "\\delta_{i_{i=1}}^{B}" + }, + { + "bbox": [ + 313, + 385, + 556, + 483 + ], + "type": "text", + "content": " on the images. These perturbations are optimized by minimizing the loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 340, + 491, + 555, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 491, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 340, + 491, + 555, + 559 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {a} = \\sum_ {l} \\left\\| \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\mu_ {l} ^ {\\text {i m g}} \\left(x _ {i} ^ {\\text {i m g}} + \\delta_ {i}\\right) - \\mathrm {B N} _ {l} ^ {\\mu} \\right\\| \\tag {4} \\\\ + \\sum_ {l} \\left\\| \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sigma_ {l} ^ {i m g} \\left(x _ {i} ^ {i m g} + \\delta_ {i}\\right) - \\mathbf {B N} _ {l} ^ {\\sigma} \\right\\|, \\\\ \\end{array}", + "image_path": "e3502c3bbc12a736219463c31166fc75c0f7786029f63802a089a209d2dc3413.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mu_l^{img}(\\cdot)" + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\sigma_l^{img}" + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "text", + "content": " are the mean and variance of the batch of intermediate features before the " + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "text", + "content": "-th BatchNormalization (BN) layer, computed by the image encoder of original CLIP. " + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{BN}_l^\\mu" + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{BN}_l^\\sigma" + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "text", + "content": " are the mean and variance parameters of the " + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 568, + 555, + 641 + ], + "type": "text", + "content": "-th BN layer, encoding the global distribution information of the pre-training dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "We then fine-tune the CLIP model on the retaining dataset, now aligned with the pre-training dataset's distribution. This serves as a reminder of the knowledge encoded in the original model, effectively achieved without direct access to the pre-training dataset. This strategy helps prevent further catastrophic forgetting that could arise from distributional" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9873" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "disparities between the retaining dataset and the original training dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 296, + 156 + ], + "type": "text", + "content": "To address the second challenge of overfitting during fine-tuning due to the large amount of parameters, we propose minimizing the impact on the original model parameters by using an exponential moving average (EMA) [28] for the final model parameters, as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 113, + 170, + 296, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 170, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 113, + 170, + 296, + 184 + ], + "type": "interline_equation", + "content": "\\theta^ {\\mathrm {e m a}} = \\alpha \\cdot \\theta^ {\\mathrm {e m a}} + (1 - \\alpha) \\cdot \\theta , \\tag {5}", + "image_path": "24da4a911a78b4ce43262b900e752325acfabfcfde6669053b76c7053bbf0e75.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{ema}}" + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "text", + "content": " is initialized as the original CLIP parameters " + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "text", + "content": " represents the continuously optimized parameters from " + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 190, + 296, + 239 + ], + "type": "text", + "content": " is the decay factor. This approach effectively balances parameter retention with necessary updates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 247, + 134, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 134, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 134, + 258 + ], + "type": "text", + "content": "4.3. Restoration" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 265, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 324 + ], + "type": "text", + "content": "In addition to fine-tuning on the retain dataset for reminding, we also employ a model merging strategy to restore the zero-shot performance of CLIP, a technique empirically shown to locate flat optima on the loss landscape, enhancing generalization [56]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 326, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 384 + ], + "type": "text", + "content": "Specifically, we first compile a small calibration dataset " + }, + { + "bbox": [ + 55, + 326, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m = \\{(x_i,y_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 326, + 296, + 384 + ], + "type": "text", + "content": " for zero-shot image classification, which may include portions of the retain dataset not used in fine-tuning. We then execute model merging with varying coefficients as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 394, + 296, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 394, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 108, + 394, + 296, + 426 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {\\theta} {\\arg \\max } \\operatorname {A c c} (g, \\mathcal {D} _ {m}), \\tag {6} \\\\ s. t. \\theta = \\alpha \\cdot \\theta^ {f} + (1 - \\alpha) \\cdot \\theta^ {o r i}, \\\\ \\end{array}", + "image_path": "294f8bdf44204e8098a4f2c0edb26f1d38ccd3cb19d25afa06f1093b68bde75b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "inline_equation", + "content": "Acc(g, \\mathcal{D}_m)" + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "content": " assesses the zero-shot classification performance of model " + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\theta^f" + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\theta^{ori}" + }, + { + "bbox": [ + 55, + 437, + 296, + 485 + ], + "type": "text", + "content": " represent the parameters of the CLIP model post-unlearning and the original model, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 131, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 131, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 131, + 510 + ], + "type": "text", + "content": "5. Evaluations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 518, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 601 + ], + "type": "text", + "content": "We begin by focusing on the forgetting subgroups within the ImageNet dataset, where the images are semantically grouped by Breeds. We select two superclasses: 'monkey' and 'turtle.' Each superclass contains four distinct subgroups. We specifically target one subgroup in each superclass for unlearning: 'marmoset' in the monkey superclass and 'box turtle' in the turtle superclass." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 611, + 170, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 170, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 170, + 624 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 665 + ], + "type": "text", + "content": "Settings. We conduct experiments in three distinct settings: in-domain forgetting, near-domain forgetting, and style forgetting, as detailed below:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "- In-domain forgetting: Experiments are conducted on the Breeds dataset [45], a subset of the ImageNet-1K dataset [16] consisting of 4 subgroups within 17 superclasses. The images, with dimensions of " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": ", closely" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 555, + 120 + ], + "type": "text", + "content": "match the distribution of images in the pre-trained LAION-5B dataset used for CLIP. For each superclass, one subgroup is selected for unlearning, while the images from the remaining three subgroups form the retaining set." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 121, + 555, + 287 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 121, + 555, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 555, + 215 + ], + "type": "text", + "content": "- Out-of-domain forgetting: Experiments are performed on the CIFAR-10 dataset [30], which differs significantly from the image distribution in LAION-5B. To align with CLIP's input format, we resize the images from " + }, + { + "bbox": [ + 314, + 121, + 555, + 215 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 314, + 121, + 555, + 215 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 314, + 121, + 555, + 215 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 314, + 121, + 555, + 215 + ], + "type": "text", + "content": ". In this setting, we select a class shared between CIFAR-10 and ImageNet-1K. The images of this selected class form the forgetting dataset, while the remaining classes in CIFAR-10 constitute the retaining set." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 216, + 554, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 216, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 216, + 554, + 287 + ], + "type": "text", + "content": "- Style forgetting: Experiments are conducted on one class in the ImageNet-1K dataset [16] transformed by Stable Diffusion [13]. The images are randomly transformed into three styles: sketch style, poster art, and black-and-white drawing. For a given style to forget, images in the remaining styles are treated as the retaining set." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 291, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 471 + ], + "type": "text", + "content": "Baselines and implementations. We select six methods as our baselines: FT [54], GA [52], Fisher [32], LIP [18], and EMMN [12], and CLIP-LIP [29]. Except the CLIP-LIP, these methods are developed for conventional models that require a classification head. To address this difference, we use the probability from the zero-shot classification task as a surrogate for the classification logits. In these approaches, the text prompt is provided by the superclass names in both the forgetting and retain sets. For all learning-based methods, recognizing the sensitivity of CLIP parameters during the fine-tuning process, we use the Adam optimizer with a small learning rate of " + }, + { + "bbox": [ + 313, + 291, + 555, + 471 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 291, + 555, + 471 + ], + "type": "text", + "content": " for up to 2 epochs. For LIP, we set the number of noisy copies to 10. Additionally, the variance scaling factor " + }, + { + "bbox": [ + 313, + 291, + 555, + 471 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 291, + 555, + 471 + ], + "type": "text", + "content": " in Fisher is set to 0.2. We provide the full implementation in Appendix A.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 474, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 713 + ], + "type": "text", + "content": "**Evaluations.** The performance of image subgroup unlearning in CLIP is evaluated along three dimensions. First, we must ensure that CLIP successfully forgets the images from the specified subgroup. Second, CLIP must continue to perform consistently on the retain set, which includes the other subgroups that are not intended to be forgotten. Third, CLIP must maintain strong zero-shot performance across various tasks and datasets. These considerations lead us to evaluate CLIP's performance not only on the targeted forgetting subgroups and unaffected subgroups, but also on various unseen datasets during the forgetting process, including Food [6], STL [15], and ObjectNet [2]. Additionally, we will use the post-unlearning CLIP to retrieve images in order to assess the impact of forgetting on its learned representations. Additionally, we use the classification accuracy ratio between the original and post-unlearned CLIP models to gauge the degree of knowledge forgetting and retention. More specifically, the ratio is computed as " + }, + { + "bbox": [ + 313, + 474, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\min\\left(\\frac{\\text{Acc\\_learn}}{\\text{Acc}_{ori}}, 1\\right)" + }, + { + "bbox": [ + 313, + 474, + 556, + 713 + ], + "type": "text", + "content": ", where we cap the score at 1 to represent full knowledge restoration. This ratio is shown in the subscript of each result. We further" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9874" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 114, + 538, + 227 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Evaluation results for forgetting the subclass 'marmoset' within the superclass 'monkey' (left) and 'box turtle' within the superclass 'turtle' (right) in the ImageNet-1K dataset are presented (Target). Each superclass contains four subgroups, with the remaining three designated as the retain set (Retain)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 114, + 538, + 227 + ], + "lines": [ + { + "bbox": [ + 58, + 114, + 538, + 227 + ], + "spans": [ + { + "bbox": [ + 58, + 114, + 538, + 227 + ], + "type": "table", + "html": "
BackboneMethodImageNetCIFARFoodSTLObjectNetScoreBackboneMethodImageNetAll (↑)CIFARFoodSTLObjectNetScore
Target (↓)Retain (↑)All (↑)Target (↓)Retain (↑)
RN50Original51.054.759.870.460.992.068.9-RN50Original64.570.359.870.460.992.068.9-
FT2.5.4.987.1.100.025.6.42.813.5.19.217.4.28.543.1.46.819.6.58.451.6FT0.0.0.072.1.100.026.0.43.512.6.17.928.6.47.070.3.76.420.2.93.359.2
GA0.0.0.00.9.1.632.2.53.816.4.23.322.9.37.663.3.68.822.1.32.145.3GA0.0.0.013.3.18.942.1.70.414.6.20.740.0.55.777.8.84.621.3.30.955.8
Fisher0.2.0.30.5.0.91.3.2.110.4.14.70.2.0.310.5.11.43.0.14.419.1Fisher1.2.1.92.0.1.83.5.5.99.7.13.80.4.0.710.5.14.11.6.23.319.3
LIP0.7.1.40.2.3.61.3.2.210.7.15.20.2.0.310.3.11.21.6.2.318.6LIP0.4.0.63.4.4.815.2.25.410.6.15.16.1.10.023.7.25.811.3.16.428.1
EMMN0.0.0.056.7.100.028.4.47.513.1.18.620.4.33.554.9.59.720.5.29.755.6EMMN0.1.0.257.2.81.428.1.47.012.5.17.822.1.36.364.7.70.317.0.25.554.0
CLIP-LIP2.65.10.4.0.717.7.29.618.8.26.74.2.6.950.8.55.212.5.18.133.2CLIP-LIP1.9.2.92.0.2.831.9.53.319.5.27.76.9.11.363.6.69.134.2.69.644.4
Ours0.0.0.050.2.91.754.5.91.185.7.100.062.9.100.081.5.88.645.2.56.691.0Ours0.0.0.069.4.98.750.6.84.654.5.77.450.5.82.987.6.95.243.1.62.685.9
RN101Original66.252.662.364.457.392.469.2-RN101Original66.277.162.364.457.392.469.2-
FT1.3.2.075.2.100.027.6.44.311.8.18.34.4.7.635.1.48.017.3.25.047.3FT0.0.0.068.0.88.232.4.52.011.4.17.727.8.48.566.0.71.423.5.34.058.8
GA0.7.1.10.3.0.536.2.58.119.9.9.929.3.51.158.6.63.419.4.88.047.3GA0.0.0.022.1.28.746.4.74.512.2.18.940.4.70.575.6.81.825.9.37.458.8
Fisher0.9.1.40.4.0.70.3.0.511.7.18.20.2.0.39.6.10.31.0.1.418.6Fisher1.9.2.92.4.3.10.9.1.411.0.17.10.3.0.59.5.10.30.6.0.918.6
LIP0.2.0.30.3.0.52.0.3.210.4.16.10.2.0.310.7.11.61.9.2.719.2LIP0.81.23.2.4.212.9.20.710.3.16.08.4.14.716.9.18.314.7.21.227.7
EMMN0.7.1.162.5.100.030.1.48.322.0.35.110.8.18.810.0.10.823.8.34.449.5EMMN0.0.0.048.6.83.032.5.52.29.2.14.322.4.39.160.8.65.822.9.33.152.5
CLIP-LIP2.7.4.11.5.2.932.0.61.428.6.44.410.9.19.049.6.53.714.9.21.541.2CLIP-LIP4.0.6.910.8.14.038.1.61.238.5.59.817.6.02.766.8.72.318.6.26.951.1
Ours0.0.0.047.7.90.758.5.93.969.7.100.056.9.99.393.5.100.045.9.66.392.9Ours0.0.0.080.1.100.059.3.95.270.9.100.055.8.97.491.5.99.046.2.66.894.1
", + "image_path": "6790c4fa66c2a078aaa350c00e9ab582bbff7da4870e39d844eca33096d766e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 95, + 236, + 517, + 387 + ], + "blocks": [ + { + "bbox": [ + 95, + 236, + 517, + 387 + ], + "lines": [ + { + "bbox": [ + 95, + 236, + 517, + 387 + ], + "spans": [ + { + "bbox": [ + 95, + 236, + 517, + 387 + ], + "type": "image", + "image_path": "014c762956b287042f5aeddaa43f90b13f9a217e31b85321e4cf8bb6c7568a5a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 396, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 54, + 396, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 555, + 419 + ], + "type": "text", + "content": "Figure 3. Retrieval results on the ImageNet-1K dataset before and after the unlearning process in the CLIP model. We highlight the \"marmoset\" subgroup in red, with other \"monkey\" subgroups shown in green." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 439, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 464 + ], + "type": "text", + "content": "compute the mean of these ratios across all classes, denoted as the Score, to provide a comprehensive evaluation metric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 285, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 285, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 285, + 482 + ], + "type": "text", + "content": "5.2. In-domain subgroup forgetting in ImageNet" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 297, + 713 + ], + "type": "text", + "content": "The Results are shown in Tab. 1. We observe that both the Fisher and LIP methods fail to optimize the model, resulting in severe catastrophic forgetting across the tested datasets. In comparison, FT, GA, EMMN, and CLIP-LIP perform slightly better but still suffer from catastrophic forgetting to varying degrees. For instance, optimizing only the retaining set with the FT strategy enables the CLIP model to forget the targeted subgroup while retaining knowledge of the intended subgroups. However, this approach leads to noticeable forgetting across other datasets, including different superclasses in ImageNet and unrelated datasets. Additionally, when applying gradient ascent to the forgetting subgroup with the GA strategy, CLIP exhibits forgetting within other subgroups of the same superclasses, and its zero-shot performance deteriorates due to catastrophic forgetting. Even with the inclusion of both forgetting and retaining samples in the fine-tuning process under the EMMN strategy, these issues remain inadequately addressed. By incorporating layer importance relative to the forgetting dataset to minimize the adverse ef" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 439, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 475 + ], + "type": "text", + "content": "fects of fine-tuning and leveraging a model merging strategy, we can preserve most of the model's zero-shot performance while ensuring that the targeted subgroup is forgotten." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "We also present the retrieval results using our fine-tuned CLIP model, which has forgotten the target subgroup, in Fig. 3. Before unlearning, when using the prompt 'monkey' to retrieve images, the model retrieves images from various subgroups, including 'marmoset' and 'howler monkey,' among others. Additionally, when prompted with specific subclass names, as shown in the second and third rows in figure, the model retrieves corresponding subgroup images. After unlearning, we observe that images of 'marmoset' are no longer retrieved when prompted with either the superclass name 'monkey' or the specific subgroup name 'marmoset,' indicating that the knowledge of 'marmoset' has been effectively erased from CLIP's parameter space. We also include retrieval results when using the prompt 'panda'; the consistent performance between the pre- and post-unlearning versions of CLIP shows that knowledge of other classes remains largely unaffected." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9875" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 113, + 538, + 227 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 2. Evaluation results for forgetting the \"airplane\" (left) and \"truck\" (right) class in CIFAR-10 are presented. To assess the unaffected subgroup performance, we also separately retrieve the \"airplane\" class from ImageNet-1K (see the column labeled Target in ImageNet). The CIFAR-10 dataset is used exclusively as the retain set (see the column labeled Retain in CIFAR)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 113, + 538, + 227 + ], + "lines": [ + { + "bbox": [ + 58, + 113, + 538, + 227 + ], + "spans": [ + { + "bbox": [ + 58, + 113, + 538, + 227 + ], + "type": "table", + "html": "
BackboneMethodCIFAR Target (↓)Retain (↑)FoodSTLObjectNetImageNet Target AllScoreBackboneMethodCIFAR Target (↓)Retain (↑)FoodSTLObjectNetImageNet Target AllScore
RN50Original81.370.460.992.068.953.659.8-RN50Original54.073.460.992.068.952.359.8-
FT0.00.092.6100.01.11.841.745.30.20.30.00.00.30.535.4FT0.00.056.376.71.32.129.311.80.40.60.00.00.20.330.2
GA0.00.014.210.21.01.614.115.30.10.10.20.40.10.219.7GA0.00.020.217.51.22.019.321.00.30.40.00.00.10.221.6
Fisher0.00.013.118.61.01.615.216.50.30.40.00.00.10.219.6Fisher0.00.012.416.91.22.013.921.00.20.40.00.00.20.219.2
LIP0.00.015.722.31.32.114.916.20.30.40.10.20.20.320.2LIP0.00.013.918.90.91.512.313.40.40.60.10.20.20.319.3
EMMN0.00.080.7100.01.62.641.645.20.20.30.00.00.50.835.6EMMN0.00.048.566.11.32.121.223.00.20.30.00.00.10.227.4
CLIP-LIP1.11.427.238.67.111.749.945.231.215.334.564.426.844.851.1CLIP-LIP3.87.045.361.716.627.374.280.733.648.828.953.339.566.161.8
Ours7.61365.102.551.985.285.793.257.984.049.892.950.183.888.9Ours0.91.763.266.150.582.984.491.756.491.951.585.852.187.189.5
RN101Original69.264.457.392.469.251.962.3-RN101Original91.479.657.392.469.250.662.3-
FT0.00.094.6100.01.11.958.463.20.40.60.00.00.20.338.0FT0.00.058.373.21.42.440.433.70.30.40.10.20.30.531.5
GA0.00.011.117.20.91.615.016.20.30.40.00.00.10.219.4GA0.00.020.225.41.22.119.320.90.20.30.00.00.10.221.3
Fisher0.00.012.920.01.01.716.918.30.30.40.10.20.20.320.1Fisher0.00.013.116.51.01.714.215.40.30.40.10.20.10.219.2
LIP0.00.013.821.41.62.815.717.00.20.30.00.00.30.520.3LIP0.00.015.719.71.62.815.717.00.30.40.00.00.20.320.0
EMMN0.00.081.2100.01.22.156.761.40.30.40.10.20.20.337.8EMMN0.00.048.360.71.42.435.848.70.10.10.10.20.20.328.9
CLIP-LIP7.410.735.585.110.418.263.168.330.432.920.699.719.831.849.5CLIP-LIP9.610.552.866.319.534.076.082.350.773.342.584.043.670.071.3
Ours8.812.763.208.148.885.287.794.958.184.045.788.141.266.186.2Ours3.23.570.388.446.280.686.693.758.384.243.986.846.674.886.4
", + "image_path": "4adc4b5aed144b80cc69c9adcd900ac67bd55d893118a00cce39805015a50dba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 83, + 238, + 526, + 392 + ], + "blocks": [ + { + "bbox": [ + 83, + 238, + 526, + 392 + ], + "lines": [ + { + "bbox": [ + 83, + 238, + 526, + 392 + ], + "spans": [ + { + "bbox": [ + 83, + 238, + 526, + 392 + ], + "type": "image", + "image_path": "1acade394e65f2e4e936d3a22d8743ba67c0d202a90a441d755b675ffedfaeeb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 403, + 436, + 416 + ], + "lines": [ + { + "bbox": [ + 173, + 403, + 436, + 416 + ], + "spans": [ + { + "bbox": [ + 173, + 403, + 436, + 416 + ], + "type": "text", + "content": "Figure 4. Retrieval results of the CLIP pre- and post-unlearning process." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 255, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 255, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 255, + 449 + ], + "type": "text", + "content": "5.3. Cross-domain forgetting in CIFAR-10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 457, + 297, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 457, + 297, + 614 + ], + "spans": [ + { + "bbox": [ + 54, + 457, + 297, + 614 + ], + "type": "text", + "content": "We then validate the effectiveness of our method in forgetting images belonging to common classes in CIFAR-10, whose concepts also appear in other datasets, such as ImageNet. The main challenge lies in the significant distributional differences between CIFAR-10 and the pre-trained dataset, which can result in severe catastrophic forgetting during the finetuning process for unlearning. In this experiment, we select 'airplane' and 'truck' as the target groups to unlearn from CLIP. In addition to evaluating the zero-shot classification performance of these targeted classes in CIFAR-10, we also assess their performance in ImageNet to determine whether unlearning is accurately confined to the specified distribution, without affecting the entire concept." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "We present the results in Tab. 2. Compared to the results on ImageNet, models fine-tuned on the CIFAR-10 dataset exhibit more severe catastrophic forgetting, even with methods like FT, GA, and EMMN, which retain some zero-shot capability after forgetting specific classes in ImageNet. Notably, models fine-tuned on CIFAR-10 to retain performance demonstrate strong zero-shot performance on STL-10, a dataset with a distribution similar to CIFAR-10, highlighting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 437, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 508 + ], + "type": "text", + "content": "the impact of the fine-tuning dataset on model performance. In contrast, our method shows remarkable robustness under such a significant distribution shift. Specifically, while the post-unlearning CLIP achieves random classification performance on the forgotten class, it retains at least " + }, + { + "bbox": [ + 313, + 437, + 555, + 508 + ], + "type": "inline_equation", + "content": "66.1\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 508 + ], + "type": "text", + "content": " of its original zero-shot performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 508, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 628 + ], + "type": "text", + "content": "We also retrieve related classes in CIFAR-10 and ImageNet using the pre- and post-unlearning CLIP models for the 'airplane' class. As shown in Fig. 4, we observe that the post-unlearning CLIP is unable to retrieve airplanes from the CIFAR-10 dataset but can still retrieve airplanes from the ImageNet dataset, indicating precise forgetting limited to the target distribution. The consistency of retrieved results between the pre- and post-unlearning CLIP further demonstrates the effectiveness of our method in preserving unrelated knowledge within the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 636, + 409, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 409, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 409, + 649 + ], + "type": "text", + "content": "5.4. Style forgetting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "We further evaluate performance on forgetting a specific style within a given class, representing a more fine-grained subgroup image unlearning challenge. Specifically, we use Stable Diffusion to generate or transform existing images into three distinct styles: poster art, black-and-white drawing," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9876" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 293, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 293, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 293, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 293, + 159 + ], + "type": "image", + "image_path": "fd23cc594ac3c10f42f5bf2654b2e4b8c22697962ccb9529354d45f365c060ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "type": "text", + "content": "Figure 5. Examples of stylized images of the sports car in the ImageNet-1K dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 236, + 296, + 308 + ], + "blocks": [ + { + "bbox": [ + 54, + 205, + 295, + 228 + ], + "lines": [ + { + "bbox": [ + 54, + 205, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 295, + 228 + ], + "type": "text", + "content": "Table 3. Evaluation results of CLIP after forgetting sketch-style \"auto-cars\" using various methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 236, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 57, + 236, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 236, + 296, + 308 + ], + "type": "table", + "html": "
MethodCIFARFoodSTLObjectNetImageNetScore
Target style (↓)Retain (↑)All (↑)
Original60.852.288.465.468.182.558.6-
FT10.116.61.93.610.511.93.96.099.5100.098.3100.01.11.920.0
GA46.776.839.876.283.294.157.888.42.02.950.361.052.188.983.2
Fisher10.517.31.52.915.918.01.11.70.10.10.50.60.20.320.1
LIP11.218.40.91.713.615.44.26.40.00.00.00.00.50.920.4
EMMN16.326.826.50.839.244.320.130.714.521.370.285.14.88.246.4
CLIP-LIP15.425.36.512.518.721.28.412.81.62.32.42.91.32.224.9
Ours49.280.943.082.480.891.458.990.12.13.166.080.051.988.687.2
", + "image_path": "37ff6f81bc7855b0a19e3f60466760a6df7aa27f0cb729d63ba2188e818e065a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 327, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 296, + 399 + ], + "type": "text", + "content": "and sketch, as illustrated in Figure 5. We first fine-tune the CLIP on images with different styles, of which the zero-shot performance is reported at the first line in Tab. 3. Then, different unlearning methods are applied to erase the model's knowledge of cars in the sketch style, while retaining images of cars in the other styles. Results are shown in Table 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 400, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 591 + ], + "type": "text", + "content": "Among the baselines, we find that GA performs the best, effectively forgetting the targeted style while retaining most unrelated knowledge, as indicated by its strong zero-shot performance across different datasets. In comparison, simply fine-tuning CLIP on the retaining set leads to severe catastrophic forgetting, as shown in the results of FT and EMMN. The EMMN results also suggest that incorporating the retaining set into the unlearning process is not always beneficial. Although using the retaining set could potentially reduce model over-forgetting, an imprecise selection of the retaining set and an imbalanced loss between it and the forgetting set can degrade the model's generalization ability on other datasets. By selecting the optimal layer for fine-tuning to minimize the impact on the model's original performance, our method achieves the best overall results, maintaining at least " + }, + { + "bbox": [ + 55, + 400, + 296, + 591 + ], + "type": "inline_equation", + "content": "80.4\\%" + }, + { + "bbox": [ + 55, + 400, + 296, + 591 + ], + "type": "text", + "content": " of the model's zero-shot capacity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 599, + 218, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 218, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 218, + 612 + ], + "type": "text", + "content": "5.5. Ablation study and discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": "On the importance of reminding stage. The reminding stage follows the forgetting stage, using the retaining set to partially recover any unintended knowledge loss. Notably, our retaining set typically differs from the original pre-training set, potentially further compromising CLIP's zero-shot classification capability. To examine this, we conduct an ablation study on the impact of iteration count during the reminding stage. Using CIFAR-10 as the unlearning" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 334, + 113, + 537, + 150 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "type": "text", + "content": "Table 4. Ablation study on the used number of steps at retaining stage on the zero-shot capacity of CLIP. We set the batch size as 64 for fine-tuning and " + }, + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "type": "text", + "content": " as 0.65 for model merging." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 113, + 537, + 150 + ], + "lines": [ + { + "bbox": [ + 334, + 113, + 537, + 150 + ], + "spans": [ + { + "bbox": [ + 334, + 113, + 537, + 150 + ], + "type": "table", + "html": "
# steps51020304050100
Accr59.863.265.641.435.434.932.5
Accin44.941.238.125.88.52.10.8
", + "image_path": "1afa3872f578b119b5078ca98051c0640763531a1d9d7cb511a313b40d9721df.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 346, + 198, + 525, + 243 + ], + "blocks": [ + { + "bbox": [ + 313, + 167, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 190 + ], + "type": "text", + "content": "Table 5. Ablation study on the " + }, + { + "bbox": [ + 313, + 167, + 555, + 190 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 167, + 555, + 190 + ], + "type": "text", + "content": " for the model merging operation at the restoring stage to the zero-shot capacity of CLIP." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 346, + 198, + 525, + 243 + ], + "lines": [ + { + "bbox": [ + 346, + 198, + 525, + 243 + ], + "spans": [ + { + "bbox": [ + 346, + 198, + 525, + 243 + ], + "type": "table", + "html": "
α0.10.30.50.650.70.9
Accf0.00.02.48.812.868.8
Accr22.727.542.763.266.778.7
Accin1.04.422.3941.243.560.5
", + "image_path": "ab435860db9fd91f7f9db7fd7831296e54afb60d858ba3438774cc87714461c3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 270, + 556, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 556, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 556, + 403 + ], + "type": "text", + "content": "dataset for CLIP, we evaluate performance on both the retain set " + }, + { + "bbox": [ + 313, + 270, + 556, + 403 + ], + "type": "inline_equation", + "content": "(\\mathrm{ACC}_r)" + }, + { + "bbox": [ + 313, + 270, + 556, + 403 + ], + "type": "text", + "content": " and ImageNet " + }, + { + "bbox": [ + 313, + 270, + 556, + 403 + ], + "type": "inline_equation", + "content": "(\\mathrm{ACC}_{in})" + }, + { + "bbox": [ + 313, + 270, + 556, + 403 + ], + "type": "text", + "content": ". As shown in Tab. 4, we observe that with an increasing number of steps, classification accuracy on the retain set initially improves before declining, while accuracy on the ImageNet-1K dataset consistently decreases, with a significant drop at higher iteration counts. This suggests that while additional steps in the reminding stage help the model relearn knowledge that was unintentionally forgotten, they also lead to a larger distribution shift from the original parameter space, ultimately impairing zero-shot performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "content": "On the model merging. In our work, we use model merging to fully restore zero-shot capacity. While our method leverages a calibration dataset to identify the optimal coefficient, " + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "content": ", we also conduct an additional experiment by directly varying the coefficients to evaluate both forgetting performance " + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "inline_equation", + "content": "(\\mathrm{ACC}_f)" + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "content": " and recovery performance " + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "inline_equation", + "content": "(\\mathrm{ACC}_r / \\mathrm{ACC}_{in})" + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "content": ". As shown in Tab. 5, achieving an optimal " + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "content": " is crucial to balancing forgetting and retaining performance. With increasing " + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "content": ", the model's retaining performance and zero-shot classification capacity improve continuously. However, a large " + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "content": " can also cause the model to relearn data that should be forgotten, as indicated by the " + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "inline_equation", + "content": "68.8\\%" + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "content": " accuracy on the forgetting set when " + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 406, + 557, + 561 + ], + "type": "text", + "content": " is set to 0.9." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 388, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 388, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 388, + 594 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we study the problem of forgetting a subgroup of images in the CLIP model. We propose a three-stage method consisting of the forgetting, retaining, and restoring steps. Our method does not require any pre-trained dataset, allowing it to achieve a balanced approach by selectively forgetting the targeted subgroup of images while preserving the knowledge that should not be forgotten. Experimental results validate the effectiveness of our method and highlight the challenges of this task, as most classical methods fail." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9877" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Andrew Bai, Chih-Kuan Yeh, Cho-Jui Hsieh, and Ankur Taly. Which pretrain samples to rehearse when finetuning pretrained models? arXiv preprint arXiv:2402.08096, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 296, + 223 + ], + "type": "text", + "content": "[3] Alexander Becker and Thomas Liebig. Evaluating machine unlearning via epistemic uncertainty. arXiv preprint arXiv:2208.10836, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 296, + 267 + ], + "type": "text", + "content": "[4] Yoshua Bengio, Aaron C Courville, and Pascal Vincent. Unsupervised feature learning and deep learning: A review and new perspectives. CoRR, abs/1206.5538, 1(2665):2012, 2012. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 312 + ], + "type": "text", + "content": "[5] Abeba Birhane, Sanghyun Han, Vishnu Boddeti, Sasha Luccioni, et al. Into the laion's den: Investigating hate in multimodal datasets. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 314, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 296, + 369 + ], + "type": "text", + "content": "[6] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 423 + ], + "type": "text", + "content": "[7] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE Symposium on Security and Privacy (SP), pages 141-159. IEEE, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 296, + 456 + ], + "type": "text", + "content": "[8] Zikui Cai, Yaoteng Tan, and M Salman Asif. Single layer single gradient unlearning. arXiv preprint arXiv:2407.11867, 2024.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 491 + ], + "type": "text", + "content": "[9] Hao Chen, Bhiksha Raj, Xing Xie, and Jindong Wang. On catastrophic inheritance of large foundation models. arXiv preprint arXiv:2402.01909, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 492, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 492, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 57, + 492, + 296, + 534 + ], + "type": "text", + "content": "[10] Ruizhe Cheng, Bichen Wu, Peizhao Zhang, Peter Vajda, and Joseph E Gonzalez. Data-efficient language-supervised zero-shot learning with self-distillation. In CVPR, pages 3119-3124, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 296, + 579 + ], + "type": "text", + "content": "[11] Eli Chien, Chao Pan, and Olgica Milenkovic. Efficient model updates for approximate unlearning of graph-structured data. In The Eleventh International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 581, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 296, + 624 + ], + "type": "text", + "content": "[12] Vikram S Chandawat, Ayush K Tarun, Murari Mandal, and Mohan Kankanhalli. Zero-shot machine unlearning. IEEE Transactions on Information Forensics and Security, 18:2345-2354, 2023. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 679 + ], + "type": "text", + "content": "[13] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[14] Lluc Clavera Comas. Machine unlearning: fisher infomation matrix and selective forgetting in deep networks. B.S. thesis, Universitat Politècnica de Catalunya, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[15] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 175, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 206 + ], + "type": "text", + "content": "[17] Ronen Eldan and Mark Russinovich. Who's harry potter? approximate unlearning in llms. arXiv preprint arXiv:2310.02238, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 208, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 554, + 251 + ], + "type": "text", + "content": "[18] Jack Foster, Kyle Fogarty, Stefan Schoepf, Cengiz Öztireli, and Alexandra Brintrup. Zero-shot machine unlearning at scale via lipschitz regularization. arXiv preprint arXiv:2402.01401, 2024. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 253, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 554, + 306 + ], + "type": "text", + "content": "[19] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9304–9312, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 363 + ], + "type": "text", + "content": "[20] Jinu Gong, Osvaldo Simeone, and Joonhyuk Kang. Bayesian variational federated learning and unlearning in decentralized networks. In 2021 IEEE 22nd International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pages 216-220. IEEE, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 554, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 407 + ], + "type": "text", + "content": "[21] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. NeurIPS, 36, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "text", + "content": "[22] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In AAAI, pages 746-754, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 455, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 487 + ], + "type": "text", + "content": "[23] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. *NeurIPS*, 34:16319-16330, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 489, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 554, + 520 + ], + "type": "text", + "content": "[24] Kris Holt. Researchers found child abuse material in the largest ai image generation dataset, 2024. Accessed: 2024-10-25. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 523, + 554, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 554, + 566 + ], + "type": "text", + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 568, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 554, + 599 + ], + "type": "text", + "content": "[26] Mark He Huang, Lin Geng Foo, and Jun Liu. Learning to unlearn for robust machine unlearning. arXiv preprint arXiv:2407.10494, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 601, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 554, + 645 + ], + "type": "text", + "content": "[27] Zhehao Huang, Xinwen Cheng, JingHao Zheng, Haoran Wang, Zhengbao He, Tao Li, and Xiaolin Huang. Unified gradient-based machine unlearning with remain geometry enhancement. arXiv preprint arXiv:2409.19732, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 678 + ], + "type": "text", + "content": "[28] Frank Klinker. Exponential moving average versus moving exponential average. Mathematische Semesterberichte, 58: 97-107, 2011. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[29] Alexey Kravets and Vinay P. Namboodiri. Zero-shot CLIP class forgetting via text-image space adaptation. Transactions on Machine Learning Research, 2025. 2, 3, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9878" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 148 + ], + "type": "text", + "content": "[31] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In CVPR, pages 20071-20082, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 295, + 194 + ], + "type": "text", + "content": "[32] Jiancheng Liu, Parikshit Ram, Yuguang Yao, Gaowen Liu, Yang Liu, PRANAY SHARMA, Sijia Liu, et al. Model sparsity can simplify machine unlearning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 295, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 295, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 295, + 238 + ], + "type": "text", + "content": "[33] Ji Liu, Jiaxiang Ren, Ruoming Jin, Zijie Zhang, Yang Zhou, Patrick Valduriez, and Dejing Dou. Fisher information-based efficient curriculum federated learning with large language models. arXiv preprint arXiv:2410.00131, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 270 + ], + "type": "text", + "content": "[34] Yufang Liu, Changzhi Sun, Yuanbin Wu, and Aimin Zhou. Unlearning with fisher masking. arXiv preprint arXiv:2310.05331, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 271, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 295, + 316 + ], + "type": "text", + "content": "[35] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Xi-angyang Ji, Antoni Chan, and Rong Jin. Improved fine-tuning by better leveraging pre-training data. Advances in Neural Information Processing Systems, 35:32568-32581, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 358 + ], + "type": "text", + "content": "[36] Alexander Ly, Maarten Marsman, Josine Verhagen, Raoul PPP Grasman, and Eric-Jan Wagenmakers. A tutorial on fisher information. Journal of Mathematical Psychology, 80:40-55, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 414 + ], + "type": "text", + "content": "[37] Ronak Mehta, Sourav Pal, Vikas Singh, and Sathya N Ravi. Deep unlearning via randomized conditionally independent hessian. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10422-10431, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 448 + ], + "type": "text", + "content": "[38] Quoc Phong Nguyen, Bryan Kian Hsiang Low, and Patrick Jaillet. Variational bayesian unlearning. Advances in Neural Information Processing Systems, 33:16025-16036, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 491 + ], + "type": "text", + "content": "[39] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. *Neural networks*, 113:54–71, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 295, + 546 + ], + "type": "text", + "content": "[40] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara, et al. Safe-clip: Removing nsfw concepts from vision-and-language models. In Proceedings of the European Conference on Computer Vision, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 581 + ], + "type": "text", + "content": "[41] Clifton Poth, Jonas Pfeiffer, Andreas Rückle, and Iryna Gurevych. What to pre-train on? efficient intermediate task selection. arXiv preprint arXiv:2104.08247, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 602 + ], + "type": "text", + "content": "[42] Jenny Quang. Does training ai violate copyright law? Berkeley Tech. LJ, 36:1407, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 603, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 668 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[44] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In CVPR, pages 2765-2775, 2023. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "text", + "content": "[45] Shibani Santurkar, Dimitris Tsipras, and Aleksander Madry. Breeds: Benchmarks for subpopulation shift. arXiv preprint arXiv:2008.04859, 2020. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 172 + ], + "type": "text", + "content": "[46] Sebastian Schelter, Mozhdeh Ariannezhad, and Maarten de Rijke. Forget me now: Fast and exact unlearning in neighborhood-based recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2011–2015, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 175, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 555, + 239 + ], + "type": "text", + "content": "[47] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "text", + "content": "[48] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 34:18075-18086, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 319 + ], + "type": "text", + "content": "[49] Adam Shostack. The boy who survived: Removing harry potter from an llm is harder than reported. arXiv preprint arXiv:2403.12082, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 321, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 555, + 375 + ], + "type": "text", + "content": "[50] Shashank Subramanian, Peter Harrington, Kurt Keutzer, Wahid Bhimji, Dmitriy Morozov, Michael W Mahoney, and Amir Gholami. Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. NeurIPS, 36, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 420 + ], + "type": "text", + "content": "[51] Ayush K Tarun, Vikram S Chandawat, Murari Mandal, and Mohan Kankanhalli. Fast yet effective machine unlearning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 422, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 555, + 476 + ], + "type": "text", + "content": "[52] Anvith Thudi, Gabriel Deza, Varun Chandrasekaran, and Nicolas Papernot. Unrolling sgd: Understanding factors influencing machine unlearning. In 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P), pages 303-319. IEEE, 2022. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 521 + ], + "type": "text", + "content": "[53] Enayat Ullah, Tung Mai, Anup Rao, Ryan A Rossi, and Raman Arora. Machine unlearning via algorithmic stability. In Conference on Learning Theory, pages 4126-4142. PMLR, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 555, + 578 + ], + "type": "text", + "content": "[54] Alexander Warnecke, Lukas Pirch, Christian Wressnegger, and Konrad Rieck. Machine unlearning of features and labels. In 30th Annual Network and Distributed System Security Symposium, NDSS 2023, San Diego, California, USA, February 27 - March 3, 2023. The Internet Society, 2023. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 580, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 634 + ], + "type": "text", + "content": "[55] Yixuan Wei, Han Hu, Zhenda Xie, Ze Liu, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Improving clip fine-tuning performance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5439-5449, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 713 + ], + "type": "text", + "content": "[56] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International conference on machine learning, pages 23965-23998. PMLR, 2022. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9879" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 376 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[57] Tomoya Yamashita, Masanori Yamada, and Takashi Shibata. One-shot machine unlearning with mnemonic code. In The 16th Asian Conference on Machine Learning (Conference Track). 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[58] Haonan Yan, Xiaoguang Li, Ziyao Guo, Hui Li, Fenghua Li, and Xiaodong Lin. Arcane: An efficient architecture for exact machine unlearning. In *IJCAI*, page 19, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 207 + ], + "type": "text", + "content": "[59] Kaiyu Yang, Clint Qinami, Li Fei-Fei, Jia Deng, and Olga Russakovsky. Towards fairer datasets: Filtering and balancing the distribution of the people subtree in the imagenet hierarchy. In Proceedings of the 2020 conference on fairness, accountability, and transparency, pages 547–558, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 295, + 240 + ], + "type": "text", + "content": "[60] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. NeurIPS, 36, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "text", + "content": "[61] Binchi Zhang, Yushun Dong, Tianhao Wang, and Jundong Li. Towards certified unlearning for deep neural networks. arXiv preprint arXiv:2408.00920, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 295, + 308 + ], + "type": "text", + "content": "[62] Jiahao Zhang. Graph unlearning with efficient partial retraining. In Companion Proceedings of the ACM on Web Conference 2024, pages 1218-1221, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 342 + ], + "type": "text", + "content": "[63] Zeliang Zhang, Mingqian Feng, Zhiheng Li, and Chenliang Xu. Discover and mitigate multiple biased subgroups in image classifiers. In CVPR, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 376 + ], + "type": "text", + "content": "[64] Zeliang Zhang, Zhuo Liu, Mingqian Feng, and Chenliang Xu. Can clip count stars? an empirical study on quantity bias in clip. In EMNLP (Findings), 2024. 1" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9880" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_content_list.json b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1fa94e4af86f6bf13911151d8d6b80f651006c5a --- /dev/null +++ b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_content_list.json @@ -0,0 +1,1491 @@ +[ + { + "type": "text", + "text": "Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics", + "text_level": 1, + "bbox": [ + 104, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shibo Zhao $^{1\\dagger*}$ , Sifan Zhou $^{1\\dagger}$ , Raphael Blanchard $^{1}$ , Yuheng Qiu $^{1}$ , Wenshan Wang $^{1}$ , Sebastian Scherer $^{1}$ , Carnegie Mellon University", + "bbox": [ + 111, + 179, + 849, + 223 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d7d98a2eaf60957bb593c4f2df6a5c979b3d7568fac2fbe056b222499820582.jpg", + "image_caption": [ + "Fig. 1. t-SNE Visualization of our Tartan IMU Foundation Model. Our foundation model is trained on over 100 hours of diverse IMU data collected from various platforms, including ground vehicles, drones, legged robots, and human motion. This model maps IMU data into a unified high-dimensional shared space, capturing broad, generalizable IMU knowledge." + ], + "image_footnote": [], + "bbox": [ + 98, + 255, + 872, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 593, + 313, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent advances in deep learning, most existing learning IMU odometry methods are trained on specific datasets, lack generalization, and are prone to overfitting, which limits their real-world application. To address these challenges, we present Tartan IMU, a foundation model designed for generalizable, IMU-based state estimation across diverse robotic platforms. Our approach consists of three-stage: First, a pre-trained foundation model leverages over 100 hours of multi-platform data to establish general motion knowledge, achieving $36\\%$ improvement in ATE over specialized models. Second, to adapt to previously unseen tasks, we use Low-Rank Adaptation (LoRA), allowing positive transfer with only $1.1M$ trainable parameters. Finally, to support robotics deployment, we introduce online test-time adaptation, which eliminates the boundary between training and testing, allowing the model to continuously \"learn as it operates\" at 200 FPS in real-time.", + "bbox": [ + 75, + 612, + 473, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: https://superodometry.com/tartanimu.", + "bbox": [ + 500, + 594, + 843, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 623, + 632, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Inertial Measurement Unit (IMU) is a critical component in robotic systems, widely used across platforms such as autonomous drones, vehicles, legged robots, and VR/AR devices for human interaction. By measuring linear acceleration and angular velocity, the IMU enables accurate motion tracking and reliable control across diverse environments. Compared to other positioning solutions such as visual odometry[47] and LiDAR odometry[48], Inertial odometry (IO) offers greater robustness in challenging conditions as it is unaffected by visual degradation such as low lighting and LiDAR degradation such as smoke [49].", + "bbox": [ + 496, + 642, + 890, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, current IO solutions rely on double integration to integrate the pose from acceleration and angular velocity, which is susceptible to drift due to factors such as sensor bias, temperature fluctuations, and sensor white noise [19]. Even minor errors can rapidly accumulate during integration, causing significant drift and unbounded er", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author. † Equal contribution.", + "bbox": [ + 94, + 886, + 331, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22520", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rors within seconds [10]. To solve this problem, previous research incorporates domain-specific knowledge such as dead reckoning[10] or other sensor constraints[19]. While these approaches show effectiveness in particular application areas, they are still difficult to address the fundamental drift problem of inertial odometry [10].", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, learning-based inertial odometry methods have demonstrated promising performance in addressing challenges of inertial positioning. For example, TLIO [25] has been effective in pedestrian motion estimation, IMO for drone motion estimation [14], AI-IMU [6] for vehicle motion estimation, and deep IMU bias estimation for legged robot motion estimation [7]. However, most of these approaches are still difficult when applied in real-world scenarios. The key challenges can be summarized as follows:", + "bbox": [ + 75, + 181, + 470, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalizability: Existing models are tailored to either locomotion-specific or device-specific, which often struggle to generalize in out-of-distribution data. To the best of our knowledge, no current method has been evaluated comprehensively during long-term operation and across various motion patterns and devices.", + "bbox": [ + 75, + 319, + 468, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adapability: Existing learning models struggle to adapt to new motion types or new platforms. Deploying learning-based IO across various platforms requires that models quickly adapt to new motion modalities, even with limited data. This adaptability is essential for the practical deployment of learning-based IO in real-world applications.", + "bbox": [ + 75, + 410, + 468, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Drawing from the success of large-scale, high-quality, and diverse data in developing general models that outperform specialized models in natural language processing [26] and visual perception [27], we explore this paradigm in the context of IMU odometry. While this approach has proven effective across many domains, no prior work has applied it to inertial positioning. In fact, IMU data differs fundamentally from language and vision data due to its continuous, dynamic, and noisy nature. Even small changes in motion patterns or devices can create significant domain gaps between the training and testing phases, making IMU generalization particularly challenging.", + "bbox": [ + 75, + 502, + 468, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we try to answer a key question: How to achieve the foundation model for IMU state estimation in robotics? We define an ideal IMU foundation model should be: i) generalizable positioning across robotic platforms, ii) positive knowledge transfer to new platforms, and iii) fast adaptation for new knowledge. We introduce TartanIMU, an IMU foundation model built on a three-stage pretrain-adapt-test framework: pre-training on a large dataset, finetuning on unseen data, and online adaptation for real-world testing (as shown in Figure 1). The main contributions are:", + "bbox": [ + 75, + 685, + 468, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- IMU Foundation Model: We introduce the first IMU foundation model, featuring a shared backbone that enables scalable, cross-platform motion estimation, as illustrated in Figure 1. TartanIMU is trained on a diverse", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dataset of over 100 hours of data collected from various robotic platforms, including ground vehicles, drones, legged robots, and human motion. As a result, it achieves a notable accuracy improvement of $36\\%$ across all platforms, consistently outperforming expert-tuned models.", + "bbox": [ + 511, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Efficient Fine-tuning for Unseen Motion: Leveraging our pre-trained IMU foundation model, it can quickly adapt to new setups with minimal data. To achieve this, we integrated LoRA [21] into our architecture, requiring only 1.1M parameters to adapt unseen scenes. This approach not only mitigates the problem of catastrophic forgetting but also allows flexible insertion of adapters tailored to different robotic platforms.", + "- Online Adaptation: We propose an online adaptation method that dynamically selects motion patterns for fast training. Unlike conventional learning approaches where the pre-trained model remains static during inference, our method eliminates the boundary between the training and testing phases—we learn as we operate. This approach progressively enhances performance and achieves rapid real-time adaptation." + ], + "bbox": [ + 500, + 167, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We hope that Tartan IMU represents a step towards such general-purpose IMU foundation models that can serve as a foundation for diverse mobile robotic applications.", + "bbox": [ + 500, + 412, + 890, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 469, + 640, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional model-based inertial odometry (IO) methods, such as those using kinematic motion models [19], estimate relative motion through double integration. However, these methods are prone to drift over time due to sensor noise, necessitating external corrections such as visual input [28]. Recent data-driven approaches [12, 29, 46] aim to address this issue by modeling and compensating for noise while quantifying uncertainty in IMU measurements. Nevertheless, error accumulation during double integration remains a significant challenge.", + "bbox": [ + 496, + 489, + 890, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning has emerged as a promising alternative, driving the development of learning-based IO methods. Early approaches utilized CNNs [11, 24, 50], LSTMs [12, 17, 18, 33, 40, 41], and Transformers [1, 31, 39, 42, 43] to estimate motion. Hybrid methods, such as TLIO [25] and IDOL [35], combined learning-based techniques with kinematic models to enhance robustness and accuracy. Furthermore, learning-based approaches have been applied to tasks like motion classification and human gait estimation [4, 38], with RIO [9] introducing rotation-equivalent augmentation to improve robustness.", + "bbox": [ + 496, + 643, + 890, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advancements, most learning-based IO methods are tailored for pedestrian navigation, limiting their ability to generalize to out-of-distribution scenarios [7]. Recent research has extended their application to diverse platforms, including legged robots [7, 8], wheeled robots [5, 6, 15], and racing drones [14]. For planar-motion", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22521", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1ed936a8e16468a4aabc76762439196e34f922388e28c8630545c9ff207fc3f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 88, + 344, + 239 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3b892534bdf82172d26958675636fcdb1c225303e611ee07b2d8da4ed6dd683e.jpg", + "image_caption": [ + "Stage 2: Fine-tune for Unseen Environments" + ], + "image_footnote": [], + "bbox": [ + 117, + 258, + 344, + 371 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/35123c12c4eb0d1dc6e7af19a636e546a8e4ccded9d869cdb44d1e5d282ca1c6.jpg", + "image_caption": [ + "Stage 3: Online Adaptation" + ], + "image_footnote": [], + "bbox": [ + 112, + 400, + 338, + 523 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c20f73908ab6e63e3addb29c0d29feaee5ca2ed17c05e706f0c7531593960d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 87, + 869, + 239 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/93b8a77a9ce4f2feaabd00b202cadfdbb495dbdf529a5216dcfb1d962d401f92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 241, + 861, + 375 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/14ade9df7871a143737211cc4a9128eb28c69cb8201d6a1b59441c9478c28a1d.jpg", + "image_caption": [ + "Fig. 2. Three Learning Stages of Tartan IMU a) Pretrained IMU Model: features a shared backbone to capture generalizable IMU knowledge. b) Efficient Fine-Tuning: utilizes an adapter to enable positive transfer for new tasks. c) Online Adaptation: employs an adaptive memory buffer to support on-the-fly model updates during deployment." + ], + "image_footnote": [], + "bbox": [ + 361, + 378, + 859, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "systems, such as ground robots and pedestrians, zero-velocity updates (ZUPTs) are often integrated with motion models [37], feature extractors [44], or uncertainty quantification frameworks [6] to mitigate drift. To improve the generalization of learning inertial odometry, EqNIO [22] introduces canonical displacement priors in the network design.", + "bbox": [ + 75, + 598, + 468, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these efforts have expanded the scope of learning-based IO, there is currently no IMU odometry model that achieves generalization across diverse platforms, adapts seamlessly to unseen environments, and serves as a foundation for diverse IMU-based motion estimation tasks. Challenges related to generalizability, and adaptability remain significant barriers to the broader adoption of learning-based IMU odometry across a wide range of applications. Further details are provided in the supplementary materials.", + "bbox": [ + 75, + 688, + 470, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 832, + 168, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pipeline Overview", + "text_level": 1, + "bbox": [ + 76, + 851, + 253, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The framework of the proposed Tartan IMU is illustrated in Figure 2. Tartan IMU comprises three key components:", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pre-trained IMU model trained from large-scale datasets, an adapter network fine-tuning for unseen motion, and online test-time adaptation to further boost the accuracy and robustness of odometry during deployment.", + "bbox": [ + 496, + 598, + 892, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Pretrained IMU Model", + "text_level": 1, + "bbox": [ + 500, + 667, + 714, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve broad generalization and few-shot learning capabilities, our Tartan IMU model is trained on a diverse array of datasets. The model architecture includes four main components: data alignment and augmentation, a heterogeneous shared IMU backbone, a multi-head for different robots, and a loss function shown as Figure 2 stage 1.", + "bbox": [ + 496, + 686, + 893, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Alignment The datasets we use involve various IMU sensors with differing axis orientations. To ensure consistency, we standardize the coordinate system to X-forward, Y-left, and Z-up in the IMU body frame [30]. We also normalize the sampling frequency to $200\\mathrm{Hz}$ across all datasets for consistency, balancing temporal resolution with computational efficiency. This setup offers two key benefits: (i) sufficient temporal resolution to accurately capture motion", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22522", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dynamics; (ii) a consistent input structure that reduces variability from different sensor sampling rates, supporting uniformity in training and inference.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Augmentation To enhance model generalization, we apply rotation-equivariance [9] in the network. This method assumes that if the IMU axis undergoes a specific rotation, the predicted and ground truth trajectories should transform correspondingly. This alignment and augmentation process is critical in improving the model's overall performance.", + "bbox": [ + 75, + 137, + 468, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Heterogeneous Shared Backbone After data alignment and augmentation, we introduce our foundation IMU encoder which employed a ResNet-style architecture [20] to extract latent features from diverse IMU inputs. Given the continuous and regular nature of robotic motion, which often follows specific patterns over time, we utilized an LSTM network to capture these temporal dynamics. The LSTM outputs are then passed to a decoder, which regresses velocity and covariance shown in Figure 2, stage1.", + "bbox": [ + 75, + 228, + 468, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our network's data flow, the earlier ResNet layers function as a heterogeneous* shared backbone, mapping diverse IMU data from different robots into a unified high-dimensional representation space. This shared representation allows IMU data from new platforms to require only minimal data and training to adapt to the shared \"knowledge.\" As a result, this representation space becomes generalizable across diverse robots, capturing broader, world-level knowledge. Our aim is to pre-train foundation models that can map raw IMU signals from various robots into a shared latent space. To showcase that our shared backbone effectively captures motion-specific representations across different robots, we use t-SNE to visualize the learned latent space (see Figure 2, Stage 1). The visualization reveals clear clusters corresponding to four robot types—vehicle (red), drones (green), legged robots (gray), and human (blue)—highlighting the model's ability to learn a unified representation while preserving the distinctions between different motion patterns.", + "bbox": [ + 75, + 364, + 468, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-head Following the heterogeneous shared backbone, the final component of our network is a robot-decoupled, multi-head design that maps the shared latent representation to the velocity estimation space. During training, different regression heads are assigned to different robot types, benefiting from joint training within the shared backbone, as illustrated in Figure 2, stage 1. This multi-head setup provides two key advantages: (i) it enables the model to learn diverse motion patterns in parallel, enhancing adaptability across different robots during deployment; and (ii) it stabilizes the training process by avoiding competing learning objectives (e.g., drone motion and car motion are significantly different). We adopt two fully connected layers as decoders to regress the velocity of the window and the cor", + "bbox": [ + 75, + 652, + 468, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "responding covariance. The mathematical formulation of the network is as follows:", + "bbox": [ + 500, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(\\hat {v}, \\hat {u}) = f \\left(\\left(^ {B} \\mathbf {a} _ {n - N}, ^ {B} \\boldsymbol {\\omega} _ {n - N}\\right), \\dots , \\left(^ {B} \\mathbf {a} _ {n}, ^ {B} \\boldsymbol {\\omega} _ {n}\\right), \\mathbf {h} _ {n - N}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 122, + 890, + 142 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { B } \\mathbf { a } _ { n } = { } _ { W } ^ { B } \\mathbf { R } _ { n } \\left( { } _ { B } ^ { W } \\mathbf { R } _ { n } \\left( \\mathbf { a } - \\mathbf { b } _ { a } \\right) - { } ^ { W } \\mathbf { g } \\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 143, + 759, + 162 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n^ {B} \\boldsymbol {\\omega} _ {n} = \\boldsymbol {\\omega} - \\mathbf {b} _ {g} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 165, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $f(\\cdot)$ represents the neural network function that processes inputs from the IMU sensor. a denotes acceleration, $\\omega$ is angular velocity from the IMU, and $\\mathbf{h}_{n - N}$ refers to the hidden state produced by the LSTM at the previous time step. For each IMU measurement, we remove the gravity vector $W\\mathbf{g} = [0,0,9.8]$ mapped into the IMU body frame. At each time step, the network predicts the current motion based on the hidden state $\\mathbf{h}_{n - N}$ and a local window of $N$ samples of acceleration and angular velocity in the body frame $\\mathbf{B}$ . The output consists of the estimated relative velocity $\\hat{\\mathbf{v}}$ and associated uncertainties $\\hat{\\mathbf{u}}$ . We also assume gyro bias $\\mathbf{b}_g$ and acceleration bias $\\mathbf{b}_a$ are known.", + "bbox": [ + 500, + 196, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss Function Our network employs relative motion loss functions for each robot head, ensuring high accuracy in local measurements, as shown in Figure 2 Stage 1. The relative loss function helps the network capture motion dynamics within a single local window, improving the smoothness of the predicted trajectory. To optimize this relative loss, we use the Mean Squared Error (MSE), defined as follows:", + "bbox": [ + 498, + 378, + 890, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R L} ^ {M S E} (\\mathbf {v}, \\hat {\\mathbf {v}}) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\sum_ {j = m} ^ {m + M} \\left(v _ {j \\rightarrow j + 1} - \\hat {v} _ {j \\rightarrow j + 1}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 492, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to existing works [25][13], we formulate the loss in the body frame, similar to [30] rather than in global coordinates, which allows the model to learn more generalized features instead of simply memorizing fixed global coordinates from training trajectories. Here, $\\hat{v}_{j\\rightarrow j + 1}$ denotes the 3D body velocity output of the network for the $j$ -th window in the body frame, while $v_{j\\rightarrow j + 1}$ represents the corresponding ground truth in the same frame. The variable $n$ indicates the batch size during training, $m$ the starting window of the sequence, and $M$ the total number of LSTM windows. In practice, we use a 10-window LSTM, where each window spans 1 second of IMU data at $200\\mathrm{Hz}$ , trained with a $1\\mathrm{Hz}$ supervision signal.", + "bbox": [ + 498, + 547, + 890, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Covariance: The covariance estimation follows the methodology outlined in [25]. The network outputs a three-dimensional vector, where each element represents the logarithm of a diagonal element of the covariance matrix $\\Sigma$ .", + "bbox": [ + 498, + 744, + 890, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {\\mathrm {N L L}} = \\frac {1}{2} (v - \\hat {v}) ^ {T} \\Sigma^ {- 1} (v - \\hat {v}) + \\frac {1}{2} \\ln | \\Sigma | \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 815, + 890, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Fine-tuning for Unseen Environments", + "text_level": 1, + "bbox": [ + 500, + 852, + 823, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While our pretrained IMU model is designed to be generalizable across various platforms, we cannot guarantee", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "* Heterogeneity refers to the varying motion patterns across different robot platforms and different IMU devices.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "22523", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d87835894a69f649de1ad71a354f2783b8143511dd9b78eae722246171774bc0.jpg", + "image_caption": [ + "Fig. 3. Distribution of motion patterns classified by GMMs. The histograms with overlaid Gaussian fits show clusters based on linear and angular velocities. The GMM identifies motion types: right turns (red) with negative linear velocity Y, left turns (blue) with positive linear velocity Y, stationary states (green) near zero velocities, and forward motion (purple) with positive linear velocities." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 346, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5166550e55e44de7d117752bb0d97e499a38c6f8f1100d3c8f60b75eca9a1ed9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 85, + 614, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f67119bb06d38602b71242a2168b5bf8371e3c677b35056390242cc4ca731298.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 85, + 888, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"zero-shot\" capability for every robot platform or motion pattern. Additionally, because training efficiency and real-time performance are critical for most robotic state estimation tasks, fine-tuning every parameter of the pretrained model becomes impractical. To address this, we propose adopting a Low-Rank Adapter (LoRA)[21], which freezes the entire pretrained IMU model and only updates a small, zero-initialized adapter network with 5 million parameters (see Figure 2 Stage 2). To update each pretrained weight matrix $W_0 \\in \\mathbb{R}^{d \\times k}$ in Tartan IMU, we apply a low-rank decomposition, representing $W_0 + \\Delta W = W_0 + BA$ , where $B \\in \\mathbb{R}^{d \\times r}$ and $A \\in \\mathbb{R}^{r \\times k}$ , with $r \\ll \\min(d, k)$ . During adaptation, $W_0$ remains fixed, while $A$ and $B$ are trainable. Both $W_0$ and $\\Delta W = BA$ process the input, and their outputs are summed. This yields:", + "bbox": [ + 75, + 277, + 472, + 505 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh = W _ {0} x + \\Delta W x = W _ {0} x + B A x \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 517, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following LoRA [21], $A$ is initialized by a random Gaussian distribution and $B$ at zero, making $\\Delta W$ start as zero. We then scale $\\Delta Wx$ by $\\frac{\\alpha}{r}$ for controlled adaptation, where $\\alpha$ is a constant in $r$ . This approach preserves the original knowledge from the pretrained model while progressively incorporating new information during adaption.", + "bbox": [ + 75, + 544, + 468, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Online Adaptation", + "text_level": 1, + "bbox": [ + 76, + 640, + 256, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While LoRA effectively addresses the domain gap between training and testing, it requires pre-collected datasets, which may not be feasible in real-world operations due to the need for rapid adaptation when domain shifts occur. To overcome this limitation, we propose a novel online learning pipeline that enables our model to adapt to new robotic platforms on-the-fly. We introduce a \"learning within SLAM\" approach, where SLAM acts as a teacher model, providing relative poses as supervision, while our IMU model functions as a student, continuously adapting to reduce the domain gap (see Figure 2 stage 3). Our approach has to overcome two major challenges associated with online adaptation: i) Limited computation time, ii) Restricted data availability", + "bbox": [ + 75, + 657, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Training Buffer Selection To address challenges $i$ and $ii$ , we found that the key lies in having a diverse and concise dataset training buffer. We introduce a", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dynamic training buffer selection mechanism that preserves diverse samples by focusing on sample similarity. This approach dynamically selects distinct training samples from the input distribution, minimizing overlap and ensuring a balanced dataset without overburdening computational resources. For motion sample similarity comparison, we employ Gaussian Mixture Models (GMMs) [32] to cluster IMU data by distinct motion patterns, as shown in Figure 3. By maintaining a uniform sample distribution across these clusters (see Figure 2, stage 3), the buffer reduces memory load, accelerates online adaptation, and enhances accuracy by learning from diverse motion representations.", + "bbox": [ + 496, + 277, + 893, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 465, + 633, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Data: We train Tartan IMU on a large-scale, heterogeneous dataset encompassing over 100 hours of real-world IMU data from eight distinct robotic platforms with diverse dynamics. Sourced from prior works such as SubT-MRS [49], IDOL [35], Blackbird [3], and the UZH dataset [16], this dataset includes wheeled robots, drones, legged robots, and human-held devices, providing raw IMU data and ground-truth trajectories. To enhance the model's generalization capability, we applied data alignment and data augmentation described in section 3.2.", + "bbox": [ + 496, + 486, + 890, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiment Goals: Our approach is designed to achieve strong generalization, efficient fine-tuning, rapid adaptation, and resistance to forgetting for IMU foundation model. To validate these, we examine four key questions:", + "bbox": [ + 496, + 637, + 890, + 698 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q1: Does our model generalize across platforms and outperform platform-specific models?", + "Q2: Can it adapt to new environments through finetuning with minimal data?", + "Q3: Is it capable of real-time adaptation to a new platform with high accuracy?", + "Q4: Can it mitigate catastrophic forgetting problems?" + ], + "bbox": [ + 498, + 698, + 890, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Generalization Evaluation", + "text_level": 1, + "bbox": [ + 500, + 806, + 738, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address Q1, we focus on two key aspects for training a more generalized model: (1) heterogeneous shared backbone; (2) large-scale, high-quality, diverse data.", + "bbox": [ + 496, + 825, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To illustrate the impact of model architecture, we evaluated various IMU models on data from four motion plat", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22524", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1550a90c0bee80e0ecc4da7a01af7d1e1347ce22dd1aec2a09d8eef7fd320f90.jpg", + "image_caption": [ + "Fig. 4. Qualitative analysis of our proposed TartanIMU compared to other specialized IMU models. Our TartanIMU outperforms different specialized IMU models across various robotic platforms, including wheeled (first row), human (second row), quadruped (third row), and drone (fourth row) systems. Our method provides the best pose estimation results shown in red trajectory." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 448 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/786a249e96eec1e59c47096ed42304e9b4e7dce92490eb691588e4b662d67f2c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Robot PlatformAI-IMU [6]RNIN-VIO [13]TLIO [25]IMO [14]TartanIMUImprovement
ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATET-RTE
Wheeled (Car)[49]7.683.337.825.068.123.738.123.736.172.52↑19.63 %↑24.36%
Handled (Human)[35]8.264.897.625.616.964.8210.195.674.321.95↑47.69%↑60.05%
Legged (Dog)[49]3.231.603.101.583.611.733.351.641.460.79↑54.83%↑50.63%
Aerial (Drone)[2]4.141.454.321.513.931.403.721.343.321.04↑19.81%↑28.97%
", + "bbox": [ + 86, + 503, + 879, + 584 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c4ec1d344c9c63594bab88e74ddad4df68607f1a9d3dcca5e48a6f4876b8e43c.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of pose estimation between the proposed TartanIMU and existing IMU methods [6, 13, 14, 25]. Our TartanIMU outperforms different specialized IMU models, achieving a $35.5\\%$ on ATE and $41.0\\%$ on T-RTE across various robotic platforms including wheeled, handled, legged and aerial systems." + ], + "table_footnote": [], + "table_body": "
Robot PlatformTartan IMU
ATE↓T-RTE↓
SingleAllImproveSingleAllImprove
Wheeled[49]7.516.17↑22%3.172.52↑26%
Handled[35]6.644.32↑54%4.241.95↑117%
Legged[49]3.641.46↑150%1.940.79↑146%
Aerial[2]3.843.32↑16%1.361.04↑30%
Average5.413.82↑42%2.681.57↑71%
", + "bbox": [ + 83, + 645, + 460, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Accuracy comparison of our Tartan IMU pre-trained model trained with data from single-robot vs. all-robots. These results confirm that training in a wide variety of robot data greatly enhances the model's performance: up to $42\\%$ average improvement on ATE and $71\\%$ average improvement on T-RTE", + "bbox": [ + 75, + 776, + 468, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "forms: aerial, wheeled, legged, and handheld. For a fair comparison, all models—AI-IMU [6], RNIN-VIO [13], TLIO [25], and IMO [14]—were trained on the same large", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset and tested on identical unseen data.", + "bbox": [ + 500, + 648, + 784, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4, our IMU pre-trained model, designed with a heterogeneous shared backbone network, demonstrated substantial few-shot learning capabilities, allowing it to generalize well across different robotic platforms. Without any additional fine-tuning, it was tested on four distinct systems—aerial, wheeled, legged, and handheld—and consistently outperformed models specifically optimized for each of these applications. This model provided more accurate and reliable odometry estimates across a variety of motion patterns. Notably, it achieved an average improvement of $35.5\\%$ in Absolute Trajectory Error (ATE) and $41.0\\%$ in Time-Relative Trajectory Error (T-RTE) over the second-best performing model. These results underscore the effectiveness of the heterogeneous shared backbone architecture in achieving robust, platform", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22525", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c9ebf24a8d5eaf117ecbdabd22f5d373b16a3af26353a05bee20578d08f7340.jpg", + "image_caption": [ + "Fine-tuning Our Model on Unseen Dataset", + "Fig. 5. Comparison of Broader Generalization Results of Our IMU Pre-trained Model with a SOTA Model on Unseen Dataset We fine-tuned both the IMU pre-trained model and the SOTA model [13] from the source domain (SubT dataset [49]) to the target domain (Tartandrive dataset [34]). The quantitative comparison show that our model adapts quicker, requiring $33\\%$ fewer iterations compared to the SOTA method." + ], + "image_footnote": [], + "bbox": [ + 98, + 107, + 867, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "agnostic generalization across diverse robotic motion types.", + "bbox": [ + 75, + 407, + 467, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluated the impact of data diversity on model performance. Specifically, we compared the results of our IMU-pretrained model when trained on data from a single platform versus data from various robotic types. As shown in Table.2, models trained on diverse datasets demonstrated substantial improvements. For instance, in the \"quadruped\" sequence, the model trained on data from aerial, wheeled, legged, and handheld platforms significantly outperformed the model trained solely on legged robot data, with Absolute Trajectory Error (ATE) reduced by up to $150\\%$ and Translational Relative Trajectory Error (T-RTE) improving by $146\\%$ . These results confirm that training on a broad range of robotic data significantly enhances model performance, even when there are notable differences between the training and testing domains. This highlights the critical importance of large-scale, high-quality, and diverse datasets in improving the generalization capabilities of the IMU model (see supplementary video).", + "bbox": [ + 75, + 422, + 468, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Fine-tuning Evaluation for Unseen Motion", + "text_level": 1, + "bbox": [ + 76, + 700, + 439, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although our model was trained on a large dataset (see Q1), it still encounters generalization challenges when deployed on new robotic platforms with differing motion patterns. To address Q2, we employed LORA with minimal data to evaluate our model's adaptability to an unseen platform. Specifically, our model was trained exclusively on the SubT dataset [49] and tested on the unseen TartanDrive dataset [34].", + "bbox": [ + 75, + 719, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results, displayed in Figure 5, demonstrate the model's performance on a test sequence excluded from the training dataset. We compared our model against a state-of-the-art (SOTA) baseline model [13], trained on the same", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SubT dataset under identical conditions. Over a 60-epoch fine-tuning period (Figure 5), our model achieved significantly more accurate odometry predictions, closely aligning with the ground truth (dashed line). In contrast, while the SOTA model showed some improvement, it retained considerable trajectory errors at the 60-epoch mark.", + "bbox": [ + 496, + 407, + 890, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These findings indicate that our IMU pre-trained model adapts to new domains faster and with $33\\%$ fewer iterations compared to the SOTA method, making it highly effective for real-world applications across diverse robotic platforms and unfamiliar environments (see supplementary video).", + "bbox": [ + 496, + 498, + 890, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Online Adaptation Evaluation", + "text_level": 1, + "bbox": [ + 498, + 579, + 767, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Q3 poses a difficult challenge, as it involves continual learning [23, 36], that requires updating the model effectively to balance prior knowledge with new experience.", + "bbox": [ + 496, + 598, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, our foundation model was initially trained on a UGV operating at a maximum speed of $5\\mathrm{m / s}$ while testing was conducted on an off-road vehicle reaching speeds up to $15\\mathrm{m / s}$ -introducing a substantial domain shift. To address this, we implemented an adaptive training buffer selection strategy, detailed in Sec. 3.4. As shown in Figure 6, the model undergoes real-time adaptation from the \"SubT UGV model\" to the \"Tartan drive car model.\" In the figure, the blue line represents the estimated trajectory from the Tartan IMU model, while the dashed line indicates the ground truth. Notably, the testing data is incrementally added to the training set, leading to an uneven data distribution. However, our approach dynamically selects meaningful training samples (stationary, forward, left turn, and right turn) to ensure a balanced and diverse memory buffer, which facilitates online adaptation in real-time. Figure 6 demonstrates that our pipeline can effectively adapt in real", + "bbox": [ + 496, + 645, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22526", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b63c9ccc0628c2355594851d97f4ab6b28cbb044e132d84ef97c4e5da7e5e57.jpg", + "image_caption": [ + "Fig. 6. Performance of online adaptation. The \"SubT UGV model\" adapts to the \"Tartan drive model\" within 105 seconds. Training data is incrementally added to the buffer, represented by the red trajectory at the bottom, and categorized into segments of stationary, forward motion, left turn, and right turn, each with varying time lengths. The height of each 3D bar indicates the duration of each segment in the buffer. After buffer selection, the data distribution becomes uniform, facilitating rapid adaptation." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b0b9fcfc6d22a05c7779779ab7e0f20088a853f16f91f332a3b5e4101dc53c3.jpg", + "image_caption": [ + "Fig. 7. Non-forgetting Characteristics of Our Model. We evaluate our Tartan IMU pretrained model's generalization before and after fine-tuning from the SubT dataset [49] (source domain) to the TartanDrive dataset [34] (target domain). This demonstrates the non-forgetting capability of our LoRA-based fine-tuning adaptation, ensuring effective generalization across different scenarios." + ], + "image_footnote": [], + "bbox": [ + 96, + 340, + 450, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "time, completing the adaptation process in 105 seconds. We also incorporated online adaptation into the SLAM pipeline. Please see supplementary videos for this section.", + "bbox": [ + 75, + 681, + 468, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Non-forgetting Characteristics Evaluation", + "text_level": 1, + "bbox": [ + 76, + 731, + 434, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address Q4, we evaluate 2 models across source and target domains: a fully fine-tuned IMU model and an IMU model fine-tuned with a low-rank adapter (LoRA) [21].", + "bbox": [ + 75, + 750, + 468, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 7(a), we present the odometry results from inference using our base Tartan IMU model without fin-tuning. Figure 7(b) shows model performance in an unseen environment (target domain) outside the pre-training dataset. Both models adapt to the new environment after fine-tuning. In Figure 7(c), we assess model performance on the source domain after domain adaptation. The", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fully fine-tuned expert model performs worse than its original performance, demonstrating catastrophic forgetting. In contrast, the LoRA-based model maintains accurate estimations on the source domain, even after adapting to new knowledge. This demonstrates that LoRA-based approach effectively mitigates the forgetting problem.", + "bbox": [ + 496, + 344, + 892, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 443, + 619, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces Tartan IMU, a foundational model designed for generalizable IMU-based pose estimation across diverse robotic platforms. Tartan IMU improves accuracy by $36\\%$ compared to specialized models, leveraging over 100 hours of multi-platform data within its three-stage architecture. It utilizes LoRA fine-tuning to enable efficient domain adaptation with minimal additional parameters. Additionally, Tartan IMU supports online test-time adaptation, allowing for continuous performance improvement during deployment. The combination of these three stages makes Tartan IMU a versatile and robust solution for a wide range of robotic platforms.", + "bbox": [ + 496, + 463, + 892, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 500, + 652, + 620, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While TartanIMU exhibits strong generalization across vehicles, drones, and legged robots, it still cannot support arbitrary robotic platforms. However, our experiments show that the car motion head generalizes well to TartanDrive and SubT vehicles. We believe our categories—car, humanoid, quadruped, and drone—encompass most robots. For unseen platforms, introducing a new motion head or leveraging a mixture of existing experts (MoE)[45] presents a promising future direction. At the same time, we plan to explore the integration of both real-world and simulated IMU data to develop a more generalizable model to further improve the generalization capability. Additionally, our GMM-based buffer selection may struggle with complex robot motion patterns. Developing a more robust, robot-agnostic adaptation strategy would be a next step.", + "bbox": [ + 496, + 672, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22527", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 76, + 90, + 235, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We sincerely thank the anonymous reviewers for their valuable feedback on this work. This research was supported by funding from the U.S. Army Research Lab under grants W911NF-23-S-0001, W911NF-21-20152, W911NF-24-20125, and W911NF-17-S-0003.", + "bbox": [ + 76, + 108, + 468, + 179 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 199, + 174, + 214 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yasin Almalioglu, Mehmet Turan, Muhamad Risqi U Sapatra, Pedro PB de Gusmao, Andrew Markham, and Niki Trigoni. Selfvio: Self-supervised deep monocular visual-inertial odometry and depth estimation. Neural Networks, 150:119-136, 2022. 2", + "[2] Amado Antonini, Winter Guerra, Varun Murali, Thomas Sayre-McCord, and Sertac Karaman. The blackbird uav dataset. The International Journal of Robotics Research, 0 (0):0278364920908331, 0. 6", + "[3] Amado Antonini, Winter Guerra, Varun Murali, Thomas Sayre-McCord, and Sertac Karaman. The blackbird dataset: A large-scale dataset for uav perception in aggressive flight. In International Symposium on Experimental Robotics, pages 130-139. Springer, 2018. 5", + "[4] Omri Asraf, Firas Shama, and Itzik Klein. Pdrnet: A deep-learning pedestrian dead reckoning framework. IEEE Sensors Journal, 22(6):4932-4939, 2021. 2", + "[5] Martin Brossard, Axel Barrau, and Silvere Bonnabel. Rinsw: Robust inertial navigation system on wheels. The IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2", + "[6] Martin Brossard, Axel Barrau, and Silvère Bonnabel. Ai-imu dead-reckoning. IEEE Transactions on Intelligent Vehicles, 5(4):585-595, 2020. 2, 3, 6", + "[7] Russell Buchanan, Varun Agrawal, Marco Camurri, Frank Dellaert, and Maurice Fallon. Deep imu bias inference for robust visual-inertial odometry with factor graphs. IEEE Robotics and Automation Letters, 8(1):41-48, 2022. 2", + "[8] Russell Buchanan, Marco Camurri, Frank Dellaert, and Maurice Fallon. Learning inertial odometry for dynamic legged robot state estimation. In Conference on Robot Learning, pages 1575-1584. PMLR, 2022. 2", + "[9] Xiya Cao, Caifa Zhou, Dandan Zeng, and Yongliang Wang. Rio: Rotation-equivariance supervised learning of robust inertial odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6614-6623, 2022. 2, 4", + "[10] Changhao Chen and Xianfei Pan. Deep learning for inertial positioning: A survey. IEEE Transactions on Intelligent Transportation Systems, 2024. 2", + "[11] Changhao Chen, Xiaoxuan Lu, Andrew Markham, and Niki Trigoni. Ionet: Learning to cure the curse of drift in inertial odometry. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2", + "[12] Changhao Chen, Yishu Miao, Chris Xiaoxuan Lu, Linhai Xie, Phil Blunsom, Andrew Markham, and Niki Trigoni. Motiontransformer: Transferring neural inertial tracking between domains. In The Conference on Artificial Intelligence (AAAI), 2019. 2" + ], + "bbox": [ + 78, + 219, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Danpeng Chen, Nan Wang, Runsen Xu, Weijian Xie, Hu-jun Bao, and Guofeng Zhang. Rnin-vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 275-283, 2021. 4, 6, 7", + "[14] Giovanni Cioffi, Leonard Bauersfeld, Elia Kaufmann, and Davide Scaramuzza. Learned inertial odometry for autonomous drone racing. 2023. 2, 6", + "[15] Santiago Cortés, Arno Solin, and Juho Kannala. Deep learning based speed estimation for constraining strapdown inertial navigation on smartphones. In 2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2018. 2", + "[16] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 5", + "[17] Mahdi Abolfazli Esfahani, Han Wang, Keyu Wu, and Shenghai Yuan. Aboldeepio: A novel deep inertial odometry network for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 21(5):1941-1950, 2019. 2", + "[18] Tobias Feigl, Sebastian Kram, Philipp Woller, Ramiz H Siddiqui, Michael Philippsen, and Christopher Mutschler. A bidirectional LSTM for estimating dynamic human velocities from a single imu. In 2019 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1-8. IEEE, 2019. 2", + "[19] Christian Forster, Luca Carlone, Frank Dellaert, and Davide Scaramuzza. Imu preintegration on manifold for efficient visual-inertial maximum-a-posteriori estimation. In Robotics: Science and Systems XI, 2015. 1, 2", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 5, 8", + "[22] Royina Karegoudra Jayanth, Yinshuang Xu, Ziyun Wang, Evangelos Chatzipantazis, Daniel Gehrig, and Kostas Dani-ilidis. Eqnio: Subequivariant neural inertial odometry, 2024. 3", + "[23] Timothee Lesort, Vincenzo Lomonaco, Andrei Stoian, Davide Maltoni, David Filliat, and Natalia Diaz-Rodriguez. Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges, 2019. 7", + "[24] Feng Liu, Hongyu Ge, Dan Tao, Ruipeng Gao, and Zhang Zhang. Smartphone-based pedestrian inertial tracking: Dataset, model, and deployment. IEEE Transactions on Instrumentation and Measurement, 2023. 2", + "[25] Wenxin Liu, David Caruso, Eddy Iig, Jing Dong, Anastasios I Mourikis, Kostas Daniilidis, Vijay Kumar, and Jakob Engel. Tlio: Tight learned inertial odometry. IEEE Robotics and Automation Letters, 5(4):5653-5660, 2020. 2, 4, 6", + "[26] R OpenAI et al. Gpt-4 technical report. ArXiv, 2303:08774, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22528", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2", + "[28] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE transactions on robotics, 34(4):1004-1020, 2018. 2", + "[29] Yuheng Qiu, Chen Wang, Can Xu, Yutian Chen, Xunfei Zhou, Youjie Xia, and Sebastian Scherer. Airimu: Learning uncertainty propagation for inertial odometry. 2023. 2", + "[30] Yuheng Qiu, Can Xu, Yutian Chen, Shibo Zhao, Junyi Geng, and Sebastian Scherer. Airio: Learning inertial odometry with enhanced imu feature observability, 2025. 3, 4", + "[31] Bingbing Rao, Ehsan Kazemi, Yifan Ding, Devu M Shila, Frank M Tucker, and Liqiang Wang. Ctin: Robust contextual transformer network for inertial navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5413-5421, 2022. 2", + "[32] Douglas A. Reynolds. Gaussian mixture models. In Encyclopedia of Biometrics, 2018. 5", + "[33] João Paulo Silva do Monte Lima, Hideaki Uchiyama, and Rin-ichiro Taniguchi. End-to-end learning framework for imu-based 6-dof odometry. Sensors, 19(17):3777, 2019. 2", + "[34] Matthew Sivaprakasam, Parv Maheshwari, Mateo Guaman Castro, Samuel Triest, Micah Nye, Steve Willits, Andrew Saba, Wenshan Wang, and Sebastian Scherer. Tartandrive 2.0: More modalities and better infrastructure to further self-supervised learning research in off-road driving tasks. arXiv preprint arXiv:2402.01913, 2024. 7, 8", + "[35] Scott Sun, Dennis Melamed, and Kris Kitani. *Idol: Inertial deep orientation-estimation and localization*. In *Proceedings of the AAAI Conference on Artificial Intelligence*, pages 6128–6137, 2021. 2, 5, 6", + "[36] Niko Sunderhauf, Oliver Brock, Walter Scheirer, Raia Hadsell, Dieter Fox, Jürgen Leitner, Ben Upcroft, Pieter Abbeel, Wolfram Burgard, Michael Milford, and Peter Corke. The limits and potentials of deep learning for robotics. The International Journal of Robotics Research, 37(4-5):405-420, 2018. 7", + "[37] Hailiang Tang, Xiaoji Niu, Tisheng Zhang, You Li, and Jingnan Liu. Odonet: Untethered speed aiding for vehicle navigation without hardware wheeled odometer. IEEE Sensors Journal, 2022. 3", + "[38] Xiaoqiang Teng, Pengfei Xu, Deke Guo, Yulan Guo, Runbo Hu, Hua Chai, and Didi Chuxing. Arpdr: An accurate and robust pedestrian dead reckoning system for indoor localization on handheld smartphones. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10888-10893. IEEE, 2020. 2", + "[39] Yiming Tu and Jin Xie. Undeeplio: Unsupervised deep lidar-inertial odometry. In Asian Conference on Pattern Recognition, pages 189-202. Springer, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Brandon Wagstaff and Jonathan Kelly. Lstm-based zero-velocity detection for robust inertial navigation. In 2018 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1–8. IEEE, 2018. 2", + "[41] Qu Wang, Haiyong Luo, Langlang Ye, Aidong Men, Fang Zhao, Yan Huang, and Shanghai Ou. Pedestrian heading estimation based on spatial transformer networks and hierarchical LSTM. IEEE Access, 7:162309-162322, 2019. 2", + "[42] Yingying Wang, Hu Cheng, and Max Q-H Meng. A2dio: Attention-driven deep inertial odometry for pedestrian localization based on 6d imu. In 2022 International Conference on Robotics and Automation (ICRA), pages 819-825. IEEE, 2022. 2", + "[43] Peng Wei, Guoliang Hua, Weibo Huang, Fanyang Meng, and Hong Liu. Unsupervised monocular visual-inertial odometry network. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 2347-2354, 2021. 2", + "[44] Xinguo Yu, Ben Liu, Xinyue Lan, Zhuoling Xiao, Shuisheng Lin, Bo Yan, and Liang Zhou. Azupt: Adaptive zero velocity update based on neural networks for pedestrian tracking. In 2019 IEEE Global Communications Conference (GLOBECOM), pages 1-6. IEEE, 2019. 3", + "[45] Seniha Esen Yuksel, Joseph N. Wilson, and Paul D. Gader. Twenty years of mixture of experts. IEEE Transactions on Neural Networks and Learning Systems, 23(8):1177-1193, 2012. 8", + "[46] Ming Zhang, Mingming Zhang, Yiming Chen, and Mingyang Li. Imu data processing for inertial aided navigation: A recurrent neural network based approach. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3992-3998. IEEE, 2021. 2", + "[47] Shibo Zhao, Peng Wang, Hengrui Zhang, Zheng Fang, and Sebastian Scherer. Tp-tio: A robust thermal-inertial odometry with deep thermalpoint. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4505-4512. IEEE, 2020. 1", + "[48] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729–8736. IEEE, 2021. 1", + "[49] Shibo Zhao, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, et al. Subt-mrs dataset: Pushing slam towards all-weather environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22647–22657, 2024. 1, 5, 6, 7, 8", + "[50] Baoding Zhou, Zhining Gu, Fuqiang Gu, Peng Wu, Chengjing Yang, Xu Liu, Linchao Li, Yan Li, and Qingquan Li. Deepvip: Deep learning-based vehicle indoor positioning using smartphones. IEEE Transactions on Vehicular Technology, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 851 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22529", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_model.json b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_model.json new file mode 100644 index 0000000000000000000000000000000000000000..00ddb73df5ff2b0fa26c3f932f5ed38d38d834dd --- /dev/null +++ b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.18, + 0.851, + 0.224 + ], + "angle": 0, + "content": "Shibo Zhao\\(^{1\\dagger*}\\), Sifan Zhou\\(^{1\\dagger}\\), Raphael Blanchard\\(^{1}\\), Yuheng Qiu\\(^{1}\\), Wenshan Wang\\(^{1}\\), Sebastian Scherer\\(^{1}\\), Carnegie Mellon University" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.256, + 0.874, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.539, + 0.893, + 0.582 + ], + "angle": 0, + "content": "Fig. 1. t-SNE Visualization of our Tartan IMU Foundation Model. Our foundation model is trained on over 100 hours of diverse IMU data collected from various platforms, including ground vehicles, drones, legged robots, and human motion. This model maps IMU data into a unified high-dimensional shared space, capturing broad, generalizable IMU knowledge." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.594, + 0.314, + 0.609 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.474, + 0.871 + ], + "angle": 0, + "content": "Despite recent advances in deep learning, most existing learning IMU odometry methods are trained on specific datasets, lack generalization, and are prone to overfitting, which limits their real-world application. To address these challenges, we present Tartan IMU, a foundation model designed for generalizable, IMU-based state estimation across diverse robotic platforms. Our approach consists of three-stage: First, a pre-trained foundation model leverages over 100 hours of multi-platform data to establish general motion knowledge, achieving \\(36\\%\\) improvement in ATE over specialized models. Second, to adapt to previously unseen tasks, we use Low-Rank Adaptation (LoRA), allowing positive transfer with only \\(1.1M\\) trainable parameters. Finally, to support robotics deployment, we introduce online test-time adaptation, which eliminates the boundary between training and testing, allowing the model to continuously \"learn as it operates\" at 200 FPS in real-time." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.595, + 0.844, + 0.611 + ], + "angle": 0, + "content": "Project page: https://superodometry.com/tartanimu." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.624, + 0.633, + 0.639 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.81 + ], + "angle": 0, + "content": "The Inertial Measurement Unit (IMU) is a critical component in robotic systems, widely used across platforms such as autonomous drones, vehicles, legged robots, and VR/AR devices for human interaction. By measuring linear acceleration and angular velocity, the IMU enables accurate motion tracking and reliable control across diverse environments. Compared to other positioning solutions such as visual odometry[47] and LiDAR odometry[48], Inertial odometry (IO) offers greater robustness in challenging conditions as it is unaffected by visual degradation such as low lighting and LiDAR degradation such as smoke [49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "However, current IO solutions rely on double integration to integrate the pose from acceleration and angular velocity, which is susceptible to drift due to factors such as sensor bias, temperature fluctuations, and sensor white noise [19]. Even minor errors can rapidly accumulate during integration, causing significant drift and unbounded er" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.333, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author. † Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22520" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "rors within seconds [10]. To solve this problem, previous research incorporates domain-specific knowledge such as dead reckoning[10] or other sensor constraints[19]. While these approaches show effectiveness in particular application areas, they are still difficult to address the fundamental drift problem of inertial odometry [10]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.319 + ], + "angle": 0, + "content": "Recently, learning-based inertial odometry methods have demonstrated promising performance in addressing challenges of inertial positioning. For example, TLIO [25] has been effective in pedestrian motion estimation, IMO for drone motion estimation [14], AI-IMU [6] for vehicle motion estimation, and deep IMU bias estimation for legged robot motion estimation [7]. However, most of these approaches are still difficult when applied in real-world scenarios. The key challenges can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.47, + 0.41 + ], + "angle": 0, + "content": "Generalizability: Existing models are tailored to either locomotion-specific or device-specific, which often struggle to generalize in out-of-distribution data. To the best of our knowledge, no current method has been evaluated comprehensively during long-term operation and across various motion patterns and devices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.469, + 0.502 + ], + "angle": 0, + "content": "Adapability: Existing learning models struggle to adapt to new motion types or new platforms. Deploying learning-based IO across various platforms requires that models quickly adapt to new motion modalities, even with limited data. This adaptability is essential for the practical deployment of learning-based IO in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.469, + 0.685 + ], + "angle": 0, + "content": "Drawing from the success of large-scale, high-quality, and diverse data in developing general models that outperform specialized models in natural language processing [26] and visual perception [27], we explore this paradigm in the context of IMU odometry. While this approach has proven effective across many domains, no prior work has applied it to inertial positioning. In fact, IMU data differs fundamentally from language and vision data due to its continuous, dynamic, and noisy nature. Even small changes in motion patterns or devices can create significant domain gaps between the training and testing phases, making IMU generalization particularly challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.837 + ], + "angle": 0, + "content": "In this paper, we try to answer a key question: How to achieve the foundation model for IMU state estimation in robotics? We define an ideal IMU foundation model should be: i) generalizable positioning across robotic platforms, ii) positive knowledge transfer to new platforms, and iii) fast adaptation for new knowledge. We introduce TartanIMU, an IMU foundation model built on a three-stage pretrain-adapt-test framework: pre-training on a large dataset, finetuning on unseen data, and online adaptation for real-world testing (as shown in Figure 1). The main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- IMU Foundation Model: We introduce the first IMU foundation model, featuring a shared backbone that enables scalable, cross-platform motion estimation, as illustrated in Figure 1. TartanIMU is trained on a diverse" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "dataset of over 100 hours of data collected from various robotic platforms, including ground vehicles, drones, legged robots, and human motion. As a result, it achieves a notable accuracy improvement of \\(36\\%\\) across all platforms, consistently outperforming expert-tuned models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.288 + ], + "angle": 0, + "content": "- Efficient Fine-tuning for Unseen Motion: Leveraging our pre-trained IMU foundation model, it can quickly adapt to new setups with minimal data. To achieve this, we integrated LoRA [21] into our architecture, requiring only 1.1M parameters to adapt unseen scenes. This approach not only mitigates the problem of catastrophic forgetting but also allows flexible insertion of adapters tailored to different robotic platforms." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.409 + ], + "angle": 0, + "content": "- Online Adaptation: We propose an online adaptation method that dynamically selects motion patterns for fast training. Unlike conventional learning approaches where the pre-trained model remains static during inference, our method eliminates the boundary between the training and testing phases—we learn as we operate. This approach progressively enhances performance and achieves rapid real-time adaptation." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.413, + 0.892, + 0.459 + ], + "angle": 0, + "content": "We hope that Tartan IMU represents a step towards such general-purpose IMU foundation models that can serve as a foundation for diverse mobile robotic applications." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.47, + 0.642, + 0.485 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Traditional model-based inertial odometry (IO) methods, such as those using kinematic motion models [19], estimate relative motion through double integration. However, these methods are prone to drift over time due to sensor noise, necessitating external corrections such as visual input [28]. Recent data-driven approaches [12, 29, 46] aim to address this issue by modeling and compensating for noise while quantifying uncertainty in IMU measurements. Nevertheless, error accumulation during double integration remains a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Deep learning has emerged as a promising alternative, driving the development of learning-based IO methods. Early approaches utilized CNNs [11, 24, 50], LSTMs [12, 17, 18, 33, 40, 41], and Transformers [1, 31, 39, 42, 43] to estimate motion. Hybrid methods, such as TLIO [25] and IDOL [35], combined learning-based techniques with kinematic models to enhance robustness and accuracy. Furthermore, learning-based approaches have been applied to tasks like motion classification and human gait estimation [4, 38], with RIO [9] introducing rotation-equivalent augmentation to improve robustness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Despite these advancements, most learning-based IO methods are tailored for pedestrian navigation, limiting their ability to generalize to out-of-distribution scenarios [7]. Recent research has extended their application to diverse platforms, including legged robots [7, 8], wheeled robots [5, 6, 15], and racing drones [14]. For planar-motion" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22521" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.345, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.244, + 0.344, + 0.258 + ], + "angle": 0, + "content": "Stage 2: Fine-tune for Unseen Environments" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.259, + 0.346, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.384, + 0.271, + 0.397 + ], + "angle": 0, + "content": "Stage 3: Online Adaptation" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.401, + 0.339, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.088, + 0.87, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.242, + 0.862, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.379, + 0.861, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.54, + 0.895, + 0.585 + ], + "angle": 0, + "content": "Fig. 2. Three Learning Stages of Tartan IMU a) Pretrained IMU Model: features a shared backbone to capture generalizable IMU knowledge. b) Efficient Fine-Tuning: utilizes an adapter to enable positive transfer for new tasks. c) Online Adaptation: employs an adaptive memory buffer to support on-the-fly model updates during deployment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.689 + ], + "angle": 0, + "content": "systems, such as ground robots and pedestrians, zero-velocity updates (ZUPTs) are often integrated with motion models [37], feature extractors [44], or uncertainty quantification frameworks [6] to mitigate drift. To improve the generalization of learning inertial odometry, EqNIO [22] introduces canonical displacement priors in the network design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.826 + ], + "angle": 0, + "content": "While these efforts have expanded the scope of learning-based IO, there is currently no IMU odometry model that achieves generalization across diverse platforms, adapts seamlessly to unseen environments, and serves as a foundation for diverse IMU-based motion estimation tasks. Challenges related to generalizability, and adaptability remain significant barriers to the broader adoption of learning-based IMU odometry across a wide range of applications. Further details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.169, + 0.848 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.852, + 0.254, + 0.868 + ], + "angle": 0, + "content": "3.1. Pipeline Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The framework of the proposed Tartan IMU is illustrated in Figure 2. Tartan IMU comprises three key components:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.66 + ], + "angle": 0, + "content": "pre-trained IMU model trained from large-scale datasets, an adapter network fine-tuning for unseen motion, and online test-time adaptation to further boost the accuracy and robustness of odometry during deployment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.668, + 0.715, + 0.683 + ], + "angle": 0, + "content": "3.2. Pretrained IMU Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.894, + 0.779 + ], + "angle": 0, + "content": "To achieve broad generalization and few-shot learning capabilities, our Tartan IMU model is trained on a diverse array of datasets. The model architecture includes four main components: data alignment and augmentation, a heterogeneous shared IMU backbone, a multi-head for different robots, and a loss function shown as Figure 2 stage 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Data Alignment The datasets we use involve various IMU sensors with differing axis orientations. To ensure consistency, we standardize the coordinate system to X-forward, Y-left, and Z-up in the IMU body frame [30]. We also normalize the sampling frequency to \\(200\\mathrm{Hz}\\) across all datasets for consistency, balancing temporal resolution with computational efficiency. This setup offers two key benefits: (i) sufficient temporal resolution to accurately capture motion" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22522" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "dynamics; (ii) a consistent input structure that reduces variability from different sensor sampling rates, supporting uniformity in training and inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.469, + 0.228 + ], + "angle": 0, + "content": "Data Augmentation To enhance model generalization, we apply rotation-equivariance [9] in the network. This method assumes that if the IMU axis undergoes a specific rotation, the predicted and ground truth trajectories should transform correspondingly. This alignment and augmentation process is critical in improving the model's overall performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.469, + 0.365 + ], + "angle": 0, + "content": "Heterogeneous Shared Backbone After data alignment and augmentation, we introduce our foundation IMU encoder which employed a ResNet-style architecture [20] to extract latent features from diverse IMU inputs. Given the continuous and regular nature of robotic motion, which often follows specific patterns over time, we utilized an LSTM network to capture these temporal dynamics. The LSTM outputs are then passed to a decoder, which regresses velocity and covariance shown in Figure 2, stage1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.469, + 0.652 + ], + "angle": 0, + "content": "In our network's data flow, the earlier ResNet layers function as a heterogeneous* shared backbone, mapping diverse IMU data from different robots into a unified high-dimensional representation space. This shared representation allows IMU data from new platforms to require only minimal data and training to adapt to the shared \"knowledge.\" As a result, this representation space becomes generalizable across diverse robots, capturing broader, world-level knowledge. Our aim is to pre-train foundation models that can map raw IMU signals from various robots into a shared latent space. To showcase that our shared backbone effectively captures motion-specific representations across different robots, we use t-SNE to visualize the learned latent space (see Figure 2, Stage 1). The visualization reveals clear clusters corresponding to four robot types—vehicle (red), drones (green), legged robots (gray), and human (blue)—highlighting the model's ability to learn a unified representation while preserving the distinctions between different motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.469, + 0.864 + ], + "angle": 0, + "content": "Multi-head Following the heterogeneous shared backbone, the final component of our network is a robot-decoupled, multi-head design that maps the shared latent representation to the velocity estimation space. During training, different regression heads are assigned to different robot types, benefiting from joint training within the shared backbone, as illustrated in Figure 2, stage 1. This multi-head setup provides two key advantages: (i) it enables the model to learn diverse motion patterns in parallel, enhancing adaptability across different robots during deployment; and (ii) it stabilizes the training process by avoiding competing learning objectives (e.g., drone motion and car motion are significantly different). We adopt two fully connected layers as decoders to regress the velocity of the window and the cor" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "responding covariance. The mathematical formulation of the network is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.123, + 0.892, + 0.143 + ], + "angle": 0, + "content": "\\[\n(\\hat {v}, \\hat {u}) = f \\left(\\left(^ {B} \\mathbf {a} _ {n - N}, ^ {B} \\boldsymbol {\\omega} _ {n - N}\\right), \\dots , \\left(^ {B} \\mathbf {a} _ {n}, ^ {B} \\boldsymbol {\\omega} _ {n}\\right), \\mathbf {h} _ {n - N}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.145, + 0.76, + 0.163 + ], + "angle": 0, + "content": "\\[\n{ } ^ { B } \\mathbf { a } _ { n } = { } _ { W } ^ { B } \\mathbf { R } _ { n } \\left( { } _ { B } ^ { W } \\mathbf { R } _ { n } \\left( \\mathbf { a } - \\mathbf { b } _ { a } \\right) - { } ^ { W } \\mathbf { g } \\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.166, + 0.892, + 0.197 + ], + "angle": 0, + "content": "\\[\n^ {B} \\boldsymbol {\\omega} _ {n} = \\boldsymbol {\\omega} - \\mathbf {b} _ {g} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.198, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Here, \\( f(\\cdot) \\) represents the neural network function that processes inputs from the IMU sensor. a denotes acceleration, \\( \\omega \\) is angular velocity from the IMU, and \\( \\mathbf{h}_{n - N} \\) refers to the hidden state produced by the LSTM at the previous time step. For each IMU measurement, we remove the gravity vector \\( W\\mathbf{g} = [0,0,9.8] \\) mapped into the IMU body frame. At each time step, the network predicts the current motion based on the hidden state \\( \\mathbf{h}_{n - N} \\) and a local window of \\( N \\) samples of acceleration and angular velocity in the body frame \\( \\mathbf{B} \\). The output consists of the estimated relative velocity \\( \\hat{\\mathbf{v}} \\) and associated uncertainties \\( \\hat{\\mathbf{u}} \\). We also assume gyro bias \\( \\mathbf{b}_g \\) and acceleration bias \\( \\mathbf{b}_a \\) are known." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.484 + ], + "angle": 0, + "content": "Loss Function Our network employs relative motion loss functions for each robot head, ensuring high accuracy in local measurements, as shown in Figure 2 Stage 1. The relative loss function helps the network capture motion dynamics within a single local window, improving the smoothness of the predicted trajectory. To optimize this relative loss, we use the Mean Squared Error (MSE), defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.493, + 0.892, + 0.542 + ], + "angle": 0, + "content": "\\[\nL _ {R L} ^ {M S E} (\\mathbf {v}, \\hat {\\mathbf {v}}) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\sum_ {j = m} ^ {m + M} \\left(v _ {j \\rightarrow j + 1} - \\hat {v} _ {j \\rightarrow j + 1}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.549, + 0.892, + 0.744 + ], + "angle": 0, + "content": "In contrast to existing works [25][13], we formulate the loss in the body frame, similar to [30] rather than in global coordinates, which allows the model to learn more generalized features instead of simply memorizing fixed global coordinates from training trajectories. Here, \\(\\hat{v}_{j\\rightarrow j + 1}\\) denotes the 3D body velocity output of the network for the \\(j\\)-th window in the body frame, while \\(v_{j\\rightarrow j + 1}\\) represents the corresponding ground truth in the same frame. The variable \\(n\\) indicates the batch size during training, \\(m\\) the starting window of the sequence, and \\(M\\) the total number of LSTM windows. In practice, we use a 10-window LSTM, where each window spans 1 second of IMU data at \\(200\\mathrm{Hz}\\), trained with a \\(1\\mathrm{Hz}\\) supervision signal." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.745, + 0.892, + 0.806 + ], + "angle": 0, + "content": "Covariance: The covariance estimation follows the methodology outlined in [25]. The network outputs a three-dimensional vector, where each element represents the logarithm of a diagonal element of the covariance matrix \\(\\Sigma\\)." + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.816, + 0.892, + 0.845 + ], + "angle": 0, + "content": "\\[\nL ^ {\\mathrm {N L L}} = \\frac {1}{2} (v - \\hat {v}) ^ {T} \\Sigma^ {- 1} (v - \\hat {v}) + \\frac {1}{2} \\ln | \\Sigma | \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.853, + 0.825, + 0.868 + ], + "angle": 0, + "content": "3.3. Fine-tuning for Unseen Environments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While our pretrained IMU model is designed to be generalizable across various platforms, we cannot guarantee" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "* Heterogeneity refers to the varying motion patterns across different robot platforms and different IMU devices." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22523" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.347, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.087, + 0.616, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.087, + 0.889, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.894, + 0.263 + ], + "angle": 0, + "content": "Fig. 3. Distribution of motion patterns classified by GMMs. The histograms with overlaid Gaussian fits show clusters based on linear and angular velocities. The GMM identifies motion types: right turns (red) with negative linear velocity Y, left turns (blue) with positive linear velocity Y, stationary states (green) near zero velocities, and forward motion (purple) with positive linear velocities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.473, + 0.506 + ], + "angle": 0, + "content": "\"zero-shot\" capability for every robot platform or motion pattern. Additionally, because training efficiency and real-time performance are critical for most robotic state estimation tasks, fine-tuning every parameter of the pretrained model becomes impractical. To address this, we propose adopting a Low-Rank Adapter (LoRA)[21], which freezes the entire pretrained IMU model and only updates a small, zero-initialized adapter network with 5 million parameters (see Figure 2 Stage 2). To update each pretrained weight matrix \\( W_0 \\in \\mathbb{R}^{d \\times k} \\) in Tartan IMU, we apply a low-rank decomposition, representing \\( W_0 + \\Delta W = W_0 + BA \\), where \\( B \\in \\mathbb{R}^{d \\times r} \\) and \\( A \\in \\mathbb{R}^{r \\times k} \\), with \\( r \\ll \\min(d, k) \\). During adaptation, \\( W_0 \\) remains fixed, while \\( A \\) and \\( B \\) are trainable. Both \\( W_0 \\) and \\( \\Delta W = BA \\) process the input, and their outputs are summed. This yields:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.518, + 0.47, + 0.535 + ], + "angle": 0, + "content": "\\[\nh = W _ {0} x + \\Delta W x = W _ {0} x + B A x \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Following LoRA [21], \\(A\\) is initialized by a random Gaussian distribution and \\(B\\) at zero, making \\(\\Delta W\\) start as zero. We then scale \\(\\Delta Wx\\) by \\(\\frac{\\alpha}{r}\\) for controlled adaptation, where \\(\\alpha\\) is a constant in \\(r\\). This approach preserves the original knowledge from the pretrained model while progressively incorporating new information during adaption." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.641, + 0.258, + 0.657 + ], + "angle": 0, + "content": "3.4. Online Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.856 + ], + "angle": 0, + "content": "While LoRA effectively addresses the domain gap between training and testing, it requires pre-collected datasets, which may not be feasible in real-world operations due to the need for rapid adaptation when domain shifts occur. To overcome this limitation, we propose a novel online learning pipeline that enables our model to adapt to new robotic platforms on-the-fly. We introduce a \"learning within SLAM\" approach, where SLAM acts as a teacher model, providing relative poses as supervision, while our IMU model functions as a student, continuously adapting to reduce the domain gap (see Figure 2 stage 3). Our approach has to overcome two major challenges associated with online adaptation: i) Limited computation time, ii) Restricted data availability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Adaptive Training Buffer Selection To address challenges \\(i\\) and \\(ii\\), we found that the key lies in having a diverse and concise dataset training buffer. We introduce a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.895, + 0.461 + ], + "angle": 0, + "content": "dynamic training buffer selection mechanism that preserves diverse samples by focusing on sample similarity. This approach dynamically selects distinct training samples from the input distribution, minimizing overlap and ensuring a balanced dataset without overburdening computational resources. For motion sample similarity comparison, we employ Gaussian Mixture Models (GMMs) [32] to cluster IMU data by distinct motion patterns, as shown in Figure 3. By maintaining a uniform sample distribution across these clusters (see Figure 2, stage 3), the buffer reduces memory load, accelerates online adaptation, and enhances accuracy by learning from diverse motion representations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.467, + 0.634, + 0.484 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Training Data: We train Tartan IMU on a large-scale, heterogeneous dataset encompassing over 100 hours of real-world IMU data from eight distinct robotic platforms with diverse dynamics. Sourced from prior works such as SubT-MRS [49], IDOL [35], Blackbird [3], and the UZH dataset [16], this dataset includes wheeled robots, drones, legged robots, and human-held devices, providing raw IMU data and ground-truth trajectories. To enhance the model's generalization capability, we applied data alignment and data augmentation described in section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Experiment Goals: Our approach is designed to achieve strong generalization, efficient fine-tuning, rapid adaptation, and resistance to forgetting for IMU foundation model. To validate these, we examine four key questions:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.699, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Q1: Does our model generalize across platforms and outperform platform-specific models?" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.729, + 0.892, + 0.758 + ], + "angle": 0, + "content": "Q2: Can it adapt to new environments through finetuning with minimal data?" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.759, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Q3: Is it capable of real-time adaptation to a new platform with high accuracy?" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.789, + 0.875, + 0.805 + ], + "angle": 0, + "content": "Q4: Can it mitigate catastrophic forgetting problems?" + }, + { + "type": "list", + "bbox": [ + 0.499, + 0.699, + 0.892, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.74, + 0.822 + ], + "angle": 0, + "content": "4.1. Generalization Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.871 + ], + "angle": 0, + "content": "To address Q1, we focus on two key aspects for training a more generalized model: (1) heterogeneous shared backbone; (2) large-scale, high-quality, diverse data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To illustrate the impact of model architecture, we evaluated various IMU models on data from four motion plat" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22524" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.895, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.456, + 0.893, + 0.499 + ], + "angle": 0, + "content": "Fig. 4. Qualitative analysis of our proposed TartanIMU compared to other specialized IMU models. Our TartanIMU outperforms different specialized IMU models across various robotic platforms, including wheeled (first row), human (second row), quadruped (third row), and drone (fourth row) systems. Our method provides the best pose estimation results shown in red trajectory." + }, + { + "type": "table", + "bbox": [ + 0.087, + 0.504, + 0.88, + 0.585 + ], + "angle": 0, + "content": "
Robot PlatformAI-IMU [6]RNIN-VIO [13]TLIO [25]IMO [14]TartanIMUImprovement
ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATET-RTE
Wheeled (Car)[49]7.683.337.825.068.123.738.123.736.172.52↑19.63 %↑24.36%
Handled (Human)[35]8.264.897.625.616.964.8210.195.674.321.95↑47.69%↑60.05%
Legged (Dog)[49]3.231.603.101.583.611.733.351.641.460.79↑54.83%↑50.63%
Aerial (Drone)[2]4.141.454.321.513.931.403.721.343.321.04↑19.81%↑28.97%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.591, + 0.891, + 0.634 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of pose estimation between the proposed TartanIMU and existing IMU methods [6, 13, 14, 25]. Our TartanIMU outperforms different specialized IMU models, achieving a \\(35.5\\%\\) on ATE and \\(41.0\\%\\) on T-RTE across various robotic platforms including wheeled, handled, legged and aerial systems." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.646, + 0.462, + 0.773 + ], + "angle": 0, + "content": "
Robot PlatformTartan IMU
ATE↓T-RTE↓
SingleAllImproveSingleAllImprove
Wheeled[49]7.516.17↑22%3.172.52↑26%
Handled[35]6.644.32↑54%4.241.95↑117%
Legged[49]3.641.46↑150%1.940.79↑146%
Aerial[2]3.843.32↑16%1.361.04↑30%
Average5.413.82↑42%2.681.57↑71%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.777, + 0.47, + 0.849 + ], + "angle": 0, + "content": "Table 2. Accuracy comparison of our Tartan IMU pre-trained model trained with data from single-robot vs. all-robots. These results confirm that training in a wide variety of robot data greatly enhances the model's performance: up to \\(42\\%\\) average improvement on ATE and \\(71\\%\\) average improvement on T-RTE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "forms: aerial, wheeled, legged, and handheld. For a fair comparison, all models—AI-IMU [6], RNIN-VIO [13], TLIO [25], and IMO [14]—were trained on the same large" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.649, + 0.785, + 0.663 + ], + "angle": 0, + "content": "dataset and tested on identical unseen data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Fig. 4, our IMU pre-trained model, designed with a heterogeneous shared backbone network, demonstrated substantial few-shot learning capabilities, allowing it to generalize well across different robotic platforms. Without any additional fine-tuning, it was tested on four distinct systems—aerial, wheeled, legged, and handheld—and consistently outperformed models specifically optimized for each of these applications. This model provided more accurate and reliable odometry estimates across a variety of motion patterns. Notably, it achieved an average improvement of \\(35.5\\%\\) in Absolute Trajectory Error (ATE) and \\(41.0\\%\\) in Time-Relative Trajectory Error (T-RTE) over the second-best performing model. These results underscore the effectiveness of the heterogeneous shared backbone architecture in achieving robust, platform" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22525" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.09, + 0.582, + 0.101 + ], + "angle": 0, + "content": "Fine-tuning Our Model on Unseen Dataset" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.108, + 0.868, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.894, + 0.398 + ], + "angle": 0, + "content": "Fig. 5. Comparison of Broader Generalization Results of Our IMU Pre-trained Model with a SOTA Model on Unseen Dataset We fine-tuned both the IMU pre-trained model and the SOTA model [13] from the source domain (SubT dataset [49]) to the target domain (Tartandrive dataset [34]). The quantitative comparison show that our model adapts quicker, requiring \\(33\\%\\) fewer iterations compared to the SOTA method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.468, + 0.422 + ], + "angle": 0, + "content": "agnostic generalization across diverse robotic motion types." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.469, + 0.695 + ], + "angle": 0, + "content": "We also evaluated the impact of data diversity on model performance. Specifically, we compared the results of our IMU-pretrained model when trained on data from a single platform versus data from various robotic types. As shown in Table.2, models trained on diverse datasets demonstrated substantial improvements. For instance, in the \"quadruped\" sequence, the model trained on data from aerial, wheeled, legged, and handheld platforms significantly outperformed the model trained solely on legged robot data, with Absolute Trajectory Error (ATE) reduced by up to \\(150\\%\\) and Translational Relative Trajectory Error (T-RTE) improving by \\(146\\%\\). These results confirm that training on a broad range of robotic data significantly enhances model performance, even when there are notable differences between the training and testing domains. This highlights the critical importance of large-scale, high-quality, and diverse datasets in improving the generalization capabilities of the IMU model (see supplementary video)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.701, + 0.44, + 0.717 + ], + "angle": 0, + "content": "4.2. Fine-tuning Evaluation for Unseen Motion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.839 + ], + "angle": 0, + "content": "Although our model was trained on a large dataset (see Q1), it still encounters generalization challenges when deployed on new robotic platforms with differing motion patterns. To address Q2, we employed LORA with minimal data to evaluate our model's adaptability to an unseen platform. Specifically, our model was trained exclusively on the SubT dataset [49] and tested on the unseen TartanDrive dataset [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The results, displayed in Figure 5, demonstrate the model's performance on a test sequence excluded from the training dataset. We compared our model against a state-of-the-art (SOTA) baseline model [13], trained on the same" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.498 + ], + "angle": 0, + "content": "SubT dataset under identical conditions. Over a 60-epoch fine-tuning period (Figure 5), our model achieved significantly more accurate odometry predictions, closely aligning with the ground truth (dashed line). In contrast, while the SOTA model showed some improvement, it retained considerable trajectory errors at the 60-epoch mark." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.575 + ], + "angle": 0, + "content": "These findings indicate that our IMU pre-trained model adapts to new domains faster and with \\(33\\%\\) fewer iterations compared to the SOTA method, making it highly effective for real-world applications across diverse robotic platforms and unfamiliar environments (see supplementary video)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.58, + 0.768, + 0.595 + ], + "angle": 0, + "content": "4.3. Online Adaptation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Q3 poses a difficult challenge, as it involves continual learning [23, 36], that requires updating the model effectively to balance prior knowledge with new experience." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this experiment, our foundation model was initially trained on a UGV operating at a maximum speed of \\(5\\mathrm{m / s}\\) while testing was conducted on an off-road vehicle reaching speeds up to \\(15\\mathrm{m / s}\\) -introducing a substantial domain shift. To address this, we implemented an adaptive training buffer selection strategy, detailed in Sec. 3.4. As shown in Figure 6, the model undergoes real-time adaptation from the \"SubT UGV model\" to the \"Tartan drive car model.\" In the figure, the blue line represents the estimated trajectory from the Tartan IMU model, while the dashed line indicates the ground truth. Notably, the testing data is incrementally added to the training set, leading to an uneven data distribution. However, our approach dynamically selects meaningful training samples (stationary, forward, left turn, and right turn) to ensure a balanced and diverse memory buffer, which facilitates online adaptation in real-time. Figure 6 demonstrates that our pipeline can effectively adapt in real" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22526" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.893, + 0.335 + ], + "angle": 0, + "content": "Fig. 6. Performance of online adaptation. The \"SubT UGV model\" adapts to the \"Tartan drive model\" within 105 seconds. Training data is incrementally added to the buffer, represented by the red trajectory at the bottom, and categorized into segments of stationary, forward motion, left turn, and right turn, each with varying time lengths. The height of each 3D bar indicates the duration of each segment in the buffer. After buffer selection, the data distribution becomes uniform, facilitating rapid adaptation." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.341, + 0.451, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Fig. 7. Non-forgetting Characteristics of Our Model. We evaluate our Tartan IMU pretrained model's generalization before and after fine-tuning from the SubT dataset [49] (source domain) to the TartanDrive dataset [34] (target domain). This demonstrates the non-forgetting capability of our LoRA-based fine-tuning adaptation, ensuring effective generalization across different scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.728 + ], + "angle": 0, + "content": "time, completing the adaptation process in 105 seconds. We also incorporated online adaptation into the SLAM pipeline. Please see supplementary videos for this section." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.436, + 0.748 + ], + "angle": 0, + "content": "4.4. Non-forgetting Characteristics Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.796 + ], + "angle": 0, + "content": "To address Q4, we evaluate 2 models across source and target domains: a fully fine-tuned IMU model and an IMU model fine-tuned with a low-rank adapter (LoRA) [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In Figure 7(a), we present the odometry results from inference using our base Tartan IMU model without fin-tuning. Figure 7(b) shows model performance in an unseen environment (target domain) outside the pre-training dataset. Both models adapt to the new environment after fine-tuning. In Figure 7(c), we assess model performance on the source domain after domain adaptation. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.893, + 0.437 + ], + "angle": 0, + "content": "fully fine-tuned expert model performs worse than its original performance, demonstrating catastrophic forgetting. In contrast, the LoRA-based model maintains accurate estimations on the source domain, even after adapting to new knowledge. This demonstrates that LoRA-based approach effectively mitigates the forgetting problem." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.62, + 0.459 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.893, + 0.646 + ], + "angle": 0, + "content": "This paper introduces Tartan IMU, a foundational model designed for generalizable IMU-based pose estimation across diverse robotic platforms. Tartan IMU improves accuracy by \\(36\\%\\) compared to specialized models, leveraging over 100 hours of multi-platform data within its three-stage architecture. It utilizes LoRA fine-tuning to enable efficient domain adaptation with minimal additional parameters. Additionally, Tartan IMU supports online test-time adaptation, allowing for continuous performance improvement during deployment. The combination of these three stages makes Tartan IMU a versatile and robust solution for a wide range of robotic platforms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.654, + 0.622, + 0.669 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.901 + ], + "angle": 0, + "content": "While TartanIMU exhibits strong generalization across vehicles, drones, and legged robots, it still cannot support arbitrary robotic platforms. However, our experiments show that the car motion head generalizes well to TartanDrive and SubT vehicles. We believe our categories—car, humanoid, quadruped, and drone—encompass most robots. For unseen platforms, introducing a new motion head or leveraging a mixture of existing experts (MoE)[45] presents a promising future direction. At the same time, we plan to explore the integration of both real-world and simulated IMU data to develop a more generalizable model to further improve the generalization capability. Additionally, our GMM-based buffer selection may struggle with complex robot motion patterns. Developing a more robust, robot-agnostic adaptation strategy would be a next step." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22527" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.236, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.109, + 0.47, + 0.18 + ], + "angle": 0, + "content": "We sincerely thank the anonymous reviewers for their valuable feedback on this work. This research was supported by funding from the U.S. Army Research Lab under grants W911NF-23-S-0001, W911NF-21-20152, W911NF-24-20125, and W911NF-17-S-0003." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.2, + 0.175, + 0.215 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.22, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[1] Yasin Almalioglu, Mehmet Turan, Muhamad Risqi U Sapatra, Pedro PB de Gusmao, Andrew Markham, and Niki Trigoni. Selfvio: Self-supervised deep monocular visual-inertial odometry and depth estimation. Neural Networks, 150:119-136, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.29, + 0.472, + 0.344 + ], + "angle": 0, + "content": "[2] Amado Antonini, Winter Guerra, Varun Murali, Thomas Sayre-McCord, and Sertac Karaman. The blackbird uav dataset. The International Journal of Robotics Research, 0 (0):0278364920908331, 0. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[3] Amado Antonini, Winter Guerra, Varun Murali, Thomas Sayre-McCord, and Sertac Karaman. The blackbird dataset: A large-scale dataset for uav perception in aggressive flight. In International Symposium on Experimental Robotics, pages 130-139. Springer, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[4] Omri Asraf, Firas Shama, and Itzik Klein. Pdrnet: A deep-learning pedestrian dead reckoning framework. IEEE Sensors Journal, 22(6):4932-4939, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.457, + 0.469, + 0.512 + ], + "angle": 0, + "content": "[5] Martin Brossard, Axel Barrau, and Silvere Bonnabel. Rinsw: Robust inertial navigation system on wheels. The IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.513, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[6] Martin Brossard, Axel Barrau, and Silvère Bonnabel. Ai-imu dead-reckoning. IEEE Transactions on Intelligent Vehicles, 5(4):585-595, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.554, + 0.469, + 0.609 + ], + "angle": 0, + "content": "[7] Russell Buchanan, Varun Agrawal, Marco Camurri, Frank Dellaert, and Maurice Fallon. Deep imu bias inference for robust visual-inertial odometry with factor graphs. IEEE Robotics and Automation Letters, 8(1):41-48, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.61, + 0.469, + 0.664 + ], + "angle": 0, + "content": "[8] Russell Buchanan, Marco Camurri, Frank Dellaert, and Maurice Fallon. Learning inertial odometry for dynamic legged robot state estimation. In Conference on Robot Learning, pages 1575-1584. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.665, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[9] Xiya Cao, Caifa Zhou, Dandan Zeng, and Yongliang Wang. Rio: Rotation-equivariance supervised learning of robust inertial odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6614-6623, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.735, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[10] Changhao Chen and Xianfei Pan. Deep learning for inertial positioning: A survey. IEEE Transactions on Intelligent Transportation Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.776, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[11] Changhao Chen, Xiaoxuan Lu, Andrew Markham, and Niki Trigoni. Ionet: Learning to cure the curse of drift in inertial odometry. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[12] Changhao Chen, Yishu Miao, Chris Xiaoxuan Lu, Linhai Xie, Phil Blunsom, Andrew Markham, and Niki Trigoni. Motiontransformer: Transferring neural inertial tracking between domains. In The Conference on Artificial Intelligence (AAAI), 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.22, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[13] Danpeng Chen, Nan Wang, Runsen Xu, Weijian Xie, Hu-jun Bao, and Guofeng Zhang. Rnin-vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 275-283, 2021. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[14] Giovanni Cioffi, Leonard Bauersfeld, Elia Kaufmann, and Davide Scaramuzza. Learned inertial odometry for autonomous drone racing. 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[15] Santiago Cortés, Arno Solin, and Juho Kannala. Deep learning based speed estimation for constraining strapdown inertial navigation on smartphones. In 2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[16] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[17] Mahdi Abolfazli Esfahani, Han Wang, Keyu Wu, and Shenghai Yuan. Aboldeepio: A novel deep inertial odometry network for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 21(5):1941-1950, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[18] Tobias Feigl, Sebastian Kram, Philipp Woller, Ramiz H Siddiqui, Michael Philippsen, and Christopher Mutschler. A bidirectional LSTM for estimating dynamic human velocities from a single imu. In 2019 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1-8. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[19] Christian Forster, Luca Carlone, Frank Dellaert, and Davide Scaramuzza. Imu preintegration on manifold for efficient visual-inertial maximum-a-posteriori estimation. In Robotics: Science and Systems XI, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.538, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[22] Royina Karegoudra Jayanth, Yinshuang Xu, Ziyun Wang, Evangelos Chatzipantazis, Daniel Gehrig, and Kostas Dani-ilidis. Eqnio: Subequivariant neural inertial odometry, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[23] Timothee Lesort, Vincenzo Lomonaco, Andrei Stoian, Davide Maltoni, David Filliat, and Natalia Diaz-Rodriguez. Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[24] Feng Liu, Hongyu Ge, Dan Tao, Ruipeng Gao, and Zhang Zhang. Smartphone-based pedestrian inertial tracking: Dataset, model, and deployment. IEEE Transactions on Instrumentation and Measurement, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[25] Wenxin Liu, David Caruso, Eddy Iig, Jing Dong, Anastasios I Mourikis, Kostas Daniilidis, Vijay Kumar, and Jakob Engel. Tlio: Tight learned inertial odometry. IEEE Robotics and Automation Letters, 5(4):5653-5660, 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] R OpenAI et al. Gpt-4 technical report. ArXiv, 2303:08774, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22528" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[27] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.219, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[28] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE transactions on robotics, 34(4):1004-1020, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.262, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[29] Yuheng Qiu, Chen Wang, Can Xu, Yutian Chen, Xunfei Zhou, Youjie Xia, and Sebastian Scherer. Airimu: Learning uncertainty propagation for inertial odometry. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[30] Yuheng Qiu, Can Xu, Yutian Chen, Shibo Zhao, Junyi Geng, and Sebastian Scherer. Airio: Learning inertial odometry with enhanced imu feature observability, 2025. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[31] Bingbing Rao, Ehsan Kazemi, Yifan Ding, Devu M Shila, Frank M Tucker, and Liqiang Wang. Ctin: Robust contextual transformer network for inertial navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5413-5421, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[32] Douglas A. Reynolds. Gaussian mixture models. In Encyclopedia of Biometrics, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.49 + ], + "angle": 0, + "content": "[33] João Paulo Silva do Monte Lima, Hideaki Uchiyama, and Rin-ichiro Taniguchi. End-to-end learning framework for imu-based 6-dof odometry. Sensors, 19(17):3777, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[34] Matthew Sivaprakasam, Parv Maheshwari, Mateo Guaman Castro, Samuel Triest, Micah Nye, Steve Willits, Andrew Saba, Wenshan Wang, and Sebastian Scherer. Tartandrive 2.0: More modalities and better infrastructure to further self-supervised learning research in off-road driving tasks. arXiv preprint arXiv:2402.01913, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[35] Scott Sun, Dennis Melamed, and Kris Kitani. *Idol: Inertial deep orientation-estimation and localization*. In *Proceedings of the AAAI Conference on Artificial Intelligence*, pages 6128–6137, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[36] Niko Sunderhauf, Oliver Brock, Walter Scheirer, Raia Hadsell, Dieter Fox, Jürgen Leitner, Ben Upcroft, Pieter Abbeel, Wolfram Burgard, Michael Milford, and Peter Corke. The limits and potentials of deep learning for robotics. The International Journal of Robotics Research, 37(4-5):405-420, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[37] Hailiang Tang, Xiaoji Niu, Tisheng Zhang, You Li, and Jingnan Liu. Odonet: Untethered speed aiding for vehicle navigation without hardware wheeled odometer. IEEE Sensors Journal, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[38] Xiaoqiang Teng, Pengfei Xu, Deke Guo, Yulan Guo, Runbo Hu, Hua Chai, and Didi Chuxing. Arpdr: An accurate and robust pedestrian dead reckoning system for indoor localization on handheld smartphones. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10888-10893. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Yiming Tu and Jin Xie. Undeeplio: Unsupervised deep lidar-inertial odometry. In Asian Conference on Pattern Recognition, pages 189-202. Springer, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[40] Brandon Wagstaff and Jonathan Kelly. Lstm-based zero-velocity detection for robust inertial navigation. In 2018 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1–8. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[41] Qu Wang, Haiyong Luo, Langlang Ye, Aidong Men, Fang Zhao, Yan Huang, and Shanghai Ou. Pedestrian heading estimation based on spatial transformer networks and hierarchical LSTM. IEEE Access, 7:162309-162322, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[42] Yingying Wang, Hu Cheng, and Max Q-H Meng. A2dio: Attention-driven deep inertial odometry for pedestrian localization based on 6d imu. In 2022 International Conference on Robotics and Automation (ICRA), pages 819-825. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[43] Peng Wei, Guoliang Hua, Weibo Huang, Fanyang Meng, and Hong Liu. Unsupervised monocular visual-inertial odometry network. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 2347-2354, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[44] Xinguo Yu, Ben Liu, Xinyue Lan, Zhuoling Xiao, Shuisheng Lin, Bo Yan, and Liang Zhou. Azupt: Adaptive zero velocity update based on neural networks for pedestrian tracking. In 2019 IEEE Global Communications Conference (GLOBECOM), pages 1-6. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.417, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[45] Seniha Esen Yuksel, Joseph N. Wilson, and Paul D. Gader. Twenty years of mixture of experts. IEEE Transactions on Neural Networks and Learning Systems, 23(8):1177-1193, 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[46] Ming Zhang, Mingming Zhang, Yiming Chen, and Mingyang Li. Imu data processing for inertial aided navigation: A recurrent neural network based approach. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3992-3998. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[47] Shibo Zhao, Peng Wang, Hengrui Zhang, Zheng Fang, and Sebastian Scherer. Tp-tio: A robust thermal-inertial odometry with deep thermalpoint. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4505-4512. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.615, + 0.892, + 0.695 + ], + "angle": 0, + "content": "[48] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729–8736. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.699, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[49] Shibo Zhao, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, et al. Subt-mrs dataset: Pushing slam towards all-weather environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22647–22657, 2024. 1, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.784, + 0.892, + 0.852 + ], + "angle": 0, + "content": "[50] Baoding Zhou, Zhining Gu, Fuqiang Gu, Peng Wu, Chengjing Yang, Xu Liu, Linchao Li, Yan Li, and Qingquan Li. Deepvip: Deep learning-based vehicle indoor positioning using smartphones. IEEE Transactions on Vehicular Technology, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22529" + } + ] +] \ No newline at end of file diff --git a/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_origin.pdf b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cc2d347e5cf09945cca492891162fda7069cc6ce --- /dev/null +++ b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/183cae4a-40b3-4a32-ae82-e543fd104c41_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bf43dce5ce9cee71f31bd69b59ae40f530903cc21489b5efb2105208c34ccc9 +size 3009675 diff --git a/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/full.md b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8f026eea2f9352872569eec561f470c96c05ad93 --- /dev/null +++ b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/full.md @@ -0,0 +1,289 @@ +# Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics + +Shibo Zhao $^{1\dagger*}$ , Sifan Zhou $^{1\dagger}$ , Raphael Blanchard $^{1}$ , Yuheng Qiu $^{1}$ , Wenshan Wang $^{1}$ , Sebastian Scherer $^{1}$ , Carnegie Mellon University + +![](images/6d7d98a2eaf60957bb593c4f2df6a5c979b3d7568fac2fbe056b222499820582.jpg) +Fig. 1. t-SNE Visualization of our Tartan IMU Foundation Model. Our foundation model is trained on over 100 hours of diverse IMU data collected from various platforms, including ground vehicles, drones, legged robots, and human motion. This model maps IMU data into a unified high-dimensional shared space, capturing broad, generalizable IMU knowledge. + +# Abstract + +Despite recent advances in deep learning, most existing learning IMU odometry methods are trained on specific datasets, lack generalization, and are prone to overfitting, which limits their real-world application. To address these challenges, we present Tartan IMU, a foundation model designed for generalizable, IMU-based state estimation across diverse robotic platforms. Our approach consists of three-stage: First, a pre-trained foundation model leverages over 100 hours of multi-platform data to establish general motion knowledge, achieving $36\%$ improvement in ATE over specialized models. Second, to adapt to previously unseen tasks, we use Low-Rank Adaptation (LoRA), allowing positive transfer with only $1.1M$ trainable parameters. Finally, to support robotics deployment, we introduce online test-time adaptation, which eliminates the boundary between training and testing, allowing the model to continuously "learn as it operates" at 200 FPS in real-time. + +Project page: https://superodometry.com/tartanimu. + +# 1. Introduction + +The Inertial Measurement Unit (IMU) is a critical component in robotic systems, widely used across platforms such as autonomous drones, vehicles, legged robots, and VR/AR devices for human interaction. By measuring linear acceleration and angular velocity, the IMU enables accurate motion tracking and reliable control across diverse environments. Compared to other positioning solutions such as visual odometry[47] and LiDAR odometry[48], Inertial odometry (IO) offers greater robustness in challenging conditions as it is unaffected by visual degradation such as low lighting and LiDAR degradation such as smoke [49]. + +However, current IO solutions rely on double integration to integrate the pose from acceleration and angular velocity, which is susceptible to drift due to factors such as sensor bias, temperature fluctuations, and sensor white noise [19]. Even minor errors can rapidly accumulate during integration, causing significant drift and unbounded er + +rors within seconds [10]. To solve this problem, previous research incorporates domain-specific knowledge such as dead reckoning[10] or other sensor constraints[19]. While these approaches show effectiveness in particular application areas, they are still difficult to address the fundamental drift problem of inertial odometry [10]. + +Recently, learning-based inertial odometry methods have demonstrated promising performance in addressing challenges of inertial positioning. For example, TLIO [25] has been effective in pedestrian motion estimation, IMO for drone motion estimation [14], AI-IMU [6] for vehicle motion estimation, and deep IMU bias estimation for legged robot motion estimation [7]. However, most of these approaches are still difficult when applied in real-world scenarios. The key challenges can be summarized as follows: + +Generalizability: Existing models are tailored to either locomotion-specific or device-specific, which often struggle to generalize in out-of-distribution data. To the best of our knowledge, no current method has been evaluated comprehensively during long-term operation and across various motion patterns and devices. + +Adapability: Existing learning models struggle to adapt to new motion types or new platforms. Deploying learning-based IO across various platforms requires that models quickly adapt to new motion modalities, even with limited data. This adaptability is essential for the practical deployment of learning-based IO in real-world applications. + +Drawing from the success of large-scale, high-quality, and diverse data in developing general models that outperform specialized models in natural language processing [26] and visual perception [27], we explore this paradigm in the context of IMU odometry. While this approach has proven effective across many domains, no prior work has applied it to inertial positioning. In fact, IMU data differs fundamentally from language and vision data due to its continuous, dynamic, and noisy nature. Even small changes in motion patterns or devices can create significant domain gaps between the training and testing phases, making IMU generalization particularly challenging. + +In this paper, we try to answer a key question: How to achieve the foundation model for IMU state estimation in robotics? We define an ideal IMU foundation model should be: i) generalizable positioning across robotic platforms, ii) positive knowledge transfer to new platforms, and iii) fast adaptation for new knowledge. We introduce TartanIMU, an IMU foundation model built on a three-stage pretrain-adapt-test framework: pre-training on a large dataset, finetuning on unseen data, and online adaptation for real-world testing (as shown in Figure 1). The main contributions are: + +- IMU Foundation Model: We introduce the first IMU foundation model, featuring a shared backbone that enables scalable, cross-platform motion estimation, as illustrated in Figure 1. TartanIMU is trained on a diverse + +dataset of over 100 hours of data collected from various robotic platforms, including ground vehicles, drones, legged robots, and human motion. As a result, it achieves a notable accuracy improvement of $36\%$ across all platforms, consistently outperforming expert-tuned models. + +- Efficient Fine-tuning for Unseen Motion: Leveraging our pre-trained IMU foundation model, it can quickly adapt to new setups with minimal data. To achieve this, we integrated LoRA [21] into our architecture, requiring only 1.1M parameters to adapt unseen scenes. This approach not only mitigates the problem of catastrophic forgetting but also allows flexible insertion of adapters tailored to different robotic platforms. +- Online Adaptation: We propose an online adaptation method that dynamically selects motion patterns for fast training. Unlike conventional learning approaches where the pre-trained model remains static during inference, our method eliminates the boundary between the training and testing phases—we learn as we operate. This approach progressively enhances performance and achieves rapid real-time adaptation. + +We hope that Tartan IMU represents a step towards such general-purpose IMU foundation models that can serve as a foundation for diverse mobile robotic applications. + +# 2. Related Work + +Traditional model-based inertial odometry (IO) methods, such as those using kinematic motion models [19], estimate relative motion through double integration. However, these methods are prone to drift over time due to sensor noise, necessitating external corrections such as visual input [28]. Recent data-driven approaches [12, 29, 46] aim to address this issue by modeling and compensating for noise while quantifying uncertainty in IMU measurements. Nevertheless, error accumulation during double integration remains a significant challenge. + +Deep learning has emerged as a promising alternative, driving the development of learning-based IO methods. Early approaches utilized CNNs [11, 24, 50], LSTMs [12, 17, 18, 33, 40, 41], and Transformers [1, 31, 39, 42, 43] to estimate motion. Hybrid methods, such as TLIO [25] and IDOL [35], combined learning-based techniques with kinematic models to enhance robustness and accuracy. Furthermore, learning-based approaches have been applied to tasks like motion classification and human gait estimation [4, 38], with RIO [9] introducing rotation-equivalent augmentation to improve robustness. + +Despite these advancements, most learning-based IO methods are tailored for pedestrian navigation, limiting their ability to generalize to out-of-distribution scenarios [7]. Recent research has extended their application to diverse platforms, including legged robots [7, 8], wheeled robots [5, 6, 15], and racing drones [14]. For planar-motion + +![](images/1ed936a8e16468a4aabc76762439196e34f922388e28c8630545c9ff207fc3f9.jpg) + +![](images/3b892534bdf82172d26958675636fcdb1c225303e611ee07b2d8da4ed6dd683e.jpg) +Stage 2: Fine-tune for Unseen Environments + +![](images/35123c12c4eb0d1dc6e7af19a636e546a8e4ccded9d869cdb44d1e5d282ca1c6.jpg) +Stage 3: Online Adaptation + +![](images/2c20f73908ab6e63e3addb29c0d29feaee5ca2ed17c05e706f0c7531593960d0.jpg) + +![](images/93b8a77a9ce4f2feaabd00b202cadfdbb495dbdf529a5216dcfb1d962d401f92.jpg) + +![](images/14ade9df7871a143737211cc4a9128eb28c69cb8201d6a1b59441c9478c28a1d.jpg) +Fig. 2. Three Learning Stages of Tartan IMU a) Pretrained IMU Model: features a shared backbone to capture generalizable IMU knowledge. b) Efficient Fine-Tuning: utilizes an adapter to enable positive transfer for new tasks. c) Online Adaptation: employs an adaptive memory buffer to support on-the-fly model updates during deployment. + +systems, such as ground robots and pedestrians, zero-velocity updates (ZUPTs) are often integrated with motion models [37], feature extractors [44], or uncertainty quantification frameworks [6] to mitigate drift. To improve the generalization of learning inertial odometry, EqNIO [22] introduces canonical displacement priors in the network design. + +While these efforts have expanded the scope of learning-based IO, there is currently no IMU odometry model that achieves generalization across diverse platforms, adapts seamlessly to unseen environments, and serves as a foundation for diverse IMU-based motion estimation tasks. Challenges related to generalizability, and adaptability remain significant barriers to the broader adoption of learning-based IMU odometry across a wide range of applications. Further details are provided in the supplementary materials. + +# 3. Method + +# 3.1. Pipeline Overview + +The framework of the proposed Tartan IMU is illustrated in Figure 2. Tartan IMU comprises three key components: + +pre-trained IMU model trained from large-scale datasets, an adapter network fine-tuning for unseen motion, and online test-time adaptation to further boost the accuracy and robustness of odometry during deployment. + +# 3.2. Pretrained IMU Model + +To achieve broad generalization and few-shot learning capabilities, our Tartan IMU model is trained on a diverse array of datasets. The model architecture includes four main components: data alignment and augmentation, a heterogeneous shared IMU backbone, a multi-head for different robots, and a loss function shown as Figure 2 stage 1. + +Data Alignment The datasets we use involve various IMU sensors with differing axis orientations. To ensure consistency, we standardize the coordinate system to X-forward, Y-left, and Z-up in the IMU body frame [30]. We also normalize the sampling frequency to $200\mathrm{Hz}$ across all datasets for consistency, balancing temporal resolution with computational efficiency. This setup offers two key benefits: (i) sufficient temporal resolution to accurately capture motion + +dynamics; (ii) a consistent input structure that reduces variability from different sensor sampling rates, supporting uniformity in training and inference. + +Data Augmentation To enhance model generalization, we apply rotation-equivariance [9] in the network. This method assumes that if the IMU axis undergoes a specific rotation, the predicted and ground truth trajectories should transform correspondingly. This alignment and augmentation process is critical in improving the model's overall performance. + +Heterogeneous Shared Backbone After data alignment and augmentation, we introduce our foundation IMU encoder which employed a ResNet-style architecture [20] to extract latent features from diverse IMU inputs. Given the continuous and regular nature of robotic motion, which often follows specific patterns over time, we utilized an LSTM network to capture these temporal dynamics. The LSTM outputs are then passed to a decoder, which regresses velocity and covariance shown in Figure 2, stage1. + +In our network's data flow, the earlier ResNet layers function as a heterogeneous* shared backbone, mapping diverse IMU data from different robots into a unified high-dimensional representation space. This shared representation allows IMU data from new platforms to require only minimal data and training to adapt to the shared "knowledge." As a result, this representation space becomes generalizable across diverse robots, capturing broader, world-level knowledge. Our aim is to pre-train foundation models that can map raw IMU signals from various robots into a shared latent space. To showcase that our shared backbone effectively captures motion-specific representations across different robots, we use t-SNE to visualize the learned latent space (see Figure 2, Stage 1). The visualization reveals clear clusters corresponding to four robot types—vehicle (red), drones (green), legged robots (gray), and human (blue)—highlighting the model's ability to learn a unified representation while preserving the distinctions between different motion patterns. + +Multi-head Following the heterogeneous shared backbone, the final component of our network is a robot-decoupled, multi-head design that maps the shared latent representation to the velocity estimation space. During training, different regression heads are assigned to different robot types, benefiting from joint training within the shared backbone, as illustrated in Figure 2, stage 1. This multi-head setup provides two key advantages: (i) it enables the model to learn diverse motion patterns in parallel, enhancing adaptability across different robots during deployment; and (ii) it stabilizes the training process by avoiding competing learning objectives (e.g., drone motion and car motion are significantly different). We adopt two fully connected layers as decoders to regress the velocity of the window and the cor + +responding covariance. The mathematical formulation of the network is as follows: + +$$ +(\hat {v}, \hat {u}) = f \left(\left(^ {B} \mathbf {a} _ {n - N}, ^ {B} \boldsymbol {\omega} _ {n - N}\right), \dots , \left(^ {B} \mathbf {a} _ {n}, ^ {B} \boldsymbol {\omega} _ {n}\right), \mathbf {h} _ {n - N}\right) +$$ + +$$ +{ } ^ { B } \mathbf { a } _ { n } = { } _ { W } ^ { B } \mathbf { R } _ { n } \left( { } _ { B } ^ { W } \mathbf { R } _ { n } \left( \mathbf { a } - \mathbf { b } _ { a } \right) - { } ^ { W } \mathbf { g } \right) +$$ + +$$ +^ {B} \boldsymbol {\omega} _ {n} = \boldsymbol {\omega} - \mathbf {b} _ {g} \tag {1} +$$ + +Here, $f(\cdot)$ represents the neural network function that processes inputs from the IMU sensor. a denotes acceleration, $\omega$ is angular velocity from the IMU, and $\mathbf{h}_{n - N}$ refers to the hidden state produced by the LSTM at the previous time step. For each IMU measurement, we remove the gravity vector $W\mathbf{g} = [0,0,9.8]$ mapped into the IMU body frame. At each time step, the network predicts the current motion based on the hidden state $\mathbf{h}_{n - N}$ and a local window of $N$ samples of acceleration and angular velocity in the body frame $\mathbf{B}$ . The output consists of the estimated relative velocity $\hat{\mathbf{v}}$ and associated uncertainties $\hat{\mathbf{u}}$ . We also assume gyro bias $\mathbf{b}_g$ and acceleration bias $\mathbf{b}_a$ are known. + +Loss Function Our network employs relative motion loss functions for each robot head, ensuring high accuracy in local measurements, as shown in Figure 2 Stage 1. The relative loss function helps the network capture motion dynamics within a single local window, improving the smoothness of the predicted trajectory. To optimize this relative loss, we use the Mean Squared Error (MSE), defined as follows: + +$$ +L _ {R L} ^ {M S E} (\mathbf {v}, \hat {\mathbf {v}}) = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\sum_ {j = m} ^ {m + M} \left(v _ {j \rightarrow j + 1} - \hat {v} _ {j \rightarrow j + 1}\right)\right) \tag {2} +$$ + +In contrast to existing works [25][13], we formulate the loss in the body frame, similar to [30] rather than in global coordinates, which allows the model to learn more generalized features instead of simply memorizing fixed global coordinates from training trajectories. Here, $\hat{v}_{j\rightarrow j + 1}$ denotes the 3D body velocity output of the network for the $j$ -th window in the body frame, while $v_{j\rightarrow j + 1}$ represents the corresponding ground truth in the same frame. The variable $n$ indicates the batch size during training, $m$ the starting window of the sequence, and $M$ the total number of LSTM windows. In practice, we use a 10-window LSTM, where each window spans 1 second of IMU data at $200\mathrm{Hz}$ , trained with a $1\mathrm{Hz}$ supervision signal. + +Covariance: The covariance estimation follows the methodology outlined in [25]. The network outputs a three-dimensional vector, where each element represents the logarithm of a diagonal element of the covariance matrix $\Sigma$ . + +$$ +L ^ {\mathrm {N L L}} = \frac {1}{2} (v - \hat {v}) ^ {T} \Sigma^ {- 1} (v - \hat {v}) + \frac {1}{2} \ln | \Sigma | \tag {3} +$$ + +# 3.3. Fine-tuning for Unseen Environments + +While our pretrained IMU model is designed to be generalizable across various platforms, we cannot guarantee + +![](images/d87835894a69f649de1ad71a354f2783b8143511dd9b78eae722246171774bc0.jpg) +Fig. 3. Distribution of motion patterns classified by GMMs. The histograms with overlaid Gaussian fits show clusters based on linear and angular velocities. The GMM identifies motion types: right turns (red) with negative linear velocity Y, left turns (blue) with positive linear velocity Y, stationary states (green) near zero velocities, and forward motion (purple) with positive linear velocities. + +![](images/5166550e55e44de7d117752bb0d97e499a38c6f8f1100d3c8f60b75eca9a1ed9.jpg) + +![](images/f67119bb06d38602b71242a2168b5bf8371e3c677b35056390242cc4ca731298.jpg) + +"zero-shot" capability for every robot platform or motion pattern. Additionally, because training efficiency and real-time performance are critical for most robotic state estimation tasks, fine-tuning every parameter of the pretrained model becomes impractical. To address this, we propose adopting a Low-Rank Adapter (LoRA)[21], which freezes the entire pretrained IMU model and only updates a small, zero-initialized adapter network with 5 million parameters (see Figure 2 Stage 2). To update each pretrained weight matrix $W_0 \in \mathbb{R}^{d \times k}$ in Tartan IMU, we apply a low-rank decomposition, representing $W_0 + \Delta W = W_0 + BA$ , where $B \in \mathbb{R}^{d \times r}$ and $A \in \mathbb{R}^{r \times k}$ , with $r \ll \min(d, k)$ . During adaptation, $W_0$ remains fixed, while $A$ and $B$ are trainable. Both $W_0$ and $\Delta W = BA$ process the input, and their outputs are summed. This yields: + +$$ +h = W _ {0} x + \Delta W x = W _ {0} x + B A x \tag {4} +$$ + +Following LoRA [21], $A$ is initialized by a random Gaussian distribution and $B$ at zero, making $\Delta W$ start as zero. We then scale $\Delta Wx$ by $\frac{\alpha}{r}$ for controlled adaptation, where $\alpha$ is a constant in $r$ . This approach preserves the original knowledge from the pretrained model while progressively incorporating new information during adaption. + +# 3.4. Online Adaptation + +While LoRA effectively addresses the domain gap between training and testing, it requires pre-collected datasets, which may not be feasible in real-world operations due to the need for rapid adaptation when domain shifts occur. To overcome this limitation, we propose a novel online learning pipeline that enables our model to adapt to new robotic platforms on-the-fly. We introduce a "learning within SLAM" approach, where SLAM acts as a teacher model, providing relative poses as supervision, while our IMU model functions as a student, continuously adapting to reduce the domain gap (see Figure 2 stage 3). Our approach has to overcome two major challenges associated with online adaptation: i) Limited computation time, ii) Restricted data availability + +Adaptive Training Buffer Selection To address challenges $i$ and $ii$ , we found that the key lies in having a diverse and concise dataset training buffer. We introduce a + +dynamic training buffer selection mechanism that preserves diverse samples by focusing on sample similarity. This approach dynamically selects distinct training samples from the input distribution, minimizing overlap and ensuring a balanced dataset without overburdening computational resources. For motion sample similarity comparison, we employ Gaussian Mixture Models (GMMs) [32] to cluster IMU data by distinct motion patterns, as shown in Figure 3. By maintaining a uniform sample distribution across these clusters (see Figure 2, stage 3), the buffer reduces memory load, accelerates online adaptation, and enhances accuracy by learning from diverse motion representations. + +# 4. Experiments + +Training Data: We train Tartan IMU on a large-scale, heterogeneous dataset encompassing over 100 hours of real-world IMU data from eight distinct robotic platforms with diverse dynamics. Sourced from prior works such as SubT-MRS [49], IDOL [35], Blackbird [3], and the UZH dataset [16], this dataset includes wheeled robots, drones, legged robots, and human-held devices, providing raw IMU data and ground-truth trajectories. To enhance the model's generalization capability, we applied data alignment and data augmentation described in section 3.2. + +Experiment Goals: Our approach is designed to achieve strong generalization, efficient fine-tuning, rapid adaptation, and resistance to forgetting for IMU foundation model. To validate these, we examine four key questions: + +Q1: Does our model generalize across platforms and outperform platform-specific models? +Q2: Can it adapt to new environments through finetuning with minimal data? +Q3: Is it capable of real-time adaptation to a new platform with high accuracy? +Q4: Can it mitigate catastrophic forgetting problems? + +# 4.1. Generalization Evaluation + +To address Q1, we focus on two key aspects for training a more generalized model: (1) heterogeneous shared backbone; (2) large-scale, high-quality, diverse data. + +To illustrate the impact of model architecture, we evaluated various IMU models on data from four motion plat + +![](images/1550a90c0bee80e0ecc4da7a01af7d1e1347ce22dd1aec2a09d8eef7fd320f90.jpg) +Fig. 4. Qualitative analysis of our proposed TartanIMU compared to other specialized IMU models. Our TartanIMU outperforms different specialized IMU models across various robotic platforms, including wheeled (first row), human (second row), quadruped (third row), and drone (fourth row) systems. Our method provides the best pose estimation results shown in red trajectory. + +
Robot PlatformAI-IMU [6]RNIN-VIO [13]TLIO [25]IMO [14]TartanIMUImprovement
ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATET-RTE
Wheeled (Car)[49]7.683.337.825.068.123.738.123.736.172.52↑19.63 %↑24.36%
Handled (Human)[35]8.264.897.625.616.964.8210.195.674.321.95↑47.69%↑60.05%
Legged (Dog)[49]3.231.603.101.583.611.733.351.641.460.79↑54.83%↑50.63%
Aerial (Drone)[2]4.141.454.321.513.931.403.721.343.321.04↑19.81%↑28.97%
+ +Table 1. Quantitative comparison of pose estimation between the proposed TartanIMU and existing IMU methods [6, 13, 14, 25]. Our TartanIMU outperforms different specialized IMU models, achieving a $35.5\%$ on ATE and $41.0\%$ on T-RTE across various robotic platforms including wheeled, handled, legged and aerial systems. + +
Robot PlatformTartan IMU
ATE↓T-RTE↓
SingleAllImproveSingleAllImprove
Wheeled[49]7.516.17↑22%3.172.52↑26%
Handled[35]6.644.32↑54%4.241.95↑117%
Legged[49]3.641.46↑150%1.940.79↑146%
Aerial[2]3.843.32↑16%1.361.04↑30%
Average5.413.82↑42%2.681.57↑71%
+ +Table 2. Accuracy comparison of our Tartan IMU pre-trained model trained with data from single-robot vs. all-robots. These results confirm that training in a wide variety of robot data greatly enhances the model's performance: up to $42\%$ average improvement on ATE and $71\%$ average improvement on T-RTE + +forms: aerial, wheeled, legged, and handheld. For a fair comparison, all models—AI-IMU [6], RNIN-VIO [13], TLIO [25], and IMO [14]—were trained on the same large + +dataset and tested on identical unseen data. + +As illustrated in Fig. 4, our IMU pre-trained model, designed with a heterogeneous shared backbone network, demonstrated substantial few-shot learning capabilities, allowing it to generalize well across different robotic platforms. Without any additional fine-tuning, it was tested on four distinct systems—aerial, wheeled, legged, and handheld—and consistently outperformed models specifically optimized for each of these applications. This model provided more accurate and reliable odometry estimates across a variety of motion patterns. Notably, it achieved an average improvement of $35.5\%$ in Absolute Trajectory Error (ATE) and $41.0\%$ in Time-Relative Trajectory Error (T-RTE) over the second-best performing model. These results underscore the effectiveness of the heterogeneous shared backbone architecture in achieving robust, platform + +![](images/1c9ebf24a8d5eaf117ecbdabd22f5d373b16a3af26353a05bee20578d08f7340.jpg) +Fine-tuning Our Model on Unseen Dataset +Fig. 5. Comparison of Broader Generalization Results of Our IMU Pre-trained Model with a SOTA Model on Unseen Dataset We fine-tuned both the IMU pre-trained model and the SOTA model [13] from the source domain (SubT dataset [49]) to the target domain (Tartandrive dataset [34]). The quantitative comparison show that our model adapts quicker, requiring $33\%$ fewer iterations compared to the SOTA method. + +agnostic generalization across diverse robotic motion types. + +We also evaluated the impact of data diversity on model performance. Specifically, we compared the results of our IMU-pretrained model when trained on data from a single platform versus data from various robotic types. As shown in Table.2, models trained on diverse datasets demonstrated substantial improvements. For instance, in the "quadruped" sequence, the model trained on data from aerial, wheeled, legged, and handheld platforms significantly outperformed the model trained solely on legged robot data, with Absolute Trajectory Error (ATE) reduced by up to $150\%$ and Translational Relative Trajectory Error (T-RTE) improving by $146\%$ . These results confirm that training on a broad range of robotic data significantly enhances model performance, even when there are notable differences between the training and testing domains. This highlights the critical importance of large-scale, high-quality, and diverse datasets in improving the generalization capabilities of the IMU model (see supplementary video). + +# 4.2. Fine-tuning Evaluation for Unseen Motion + +Although our model was trained on a large dataset (see Q1), it still encounters generalization challenges when deployed on new robotic platforms with differing motion patterns. To address Q2, we employed LORA with minimal data to evaluate our model's adaptability to an unseen platform. Specifically, our model was trained exclusively on the SubT dataset [49] and tested on the unseen TartanDrive dataset [34]. + +The results, displayed in Figure 5, demonstrate the model's performance on a test sequence excluded from the training dataset. We compared our model against a state-of-the-art (SOTA) baseline model [13], trained on the same + +SubT dataset under identical conditions. Over a 60-epoch fine-tuning period (Figure 5), our model achieved significantly more accurate odometry predictions, closely aligning with the ground truth (dashed line). In contrast, while the SOTA model showed some improvement, it retained considerable trajectory errors at the 60-epoch mark. + +These findings indicate that our IMU pre-trained model adapts to new domains faster and with $33\%$ fewer iterations compared to the SOTA method, making it highly effective for real-world applications across diverse robotic platforms and unfamiliar environments (see supplementary video). + +# 4.3. Online Adaptation Evaluation + +Q3 poses a difficult challenge, as it involves continual learning [23, 36], that requires updating the model effectively to balance prior knowledge with new experience. + +In this experiment, our foundation model was initially trained on a UGV operating at a maximum speed of $5\mathrm{m / s}$ while testing was conducted on an off-road vehicle reaching speeds up to $15\mathrm{m / s}$ -introducing a substantial domain shift. To address this, we implemented an adaptive training buffer selection strategy, detailed in Sec. 3.4. As shown in Figure 6, the model undergoes real-time adaptation from the "SubT UGV model" to the "Tartan drive car model." In the figure, the blue line represents the estimated trajectory from the Tartan IMU model, while the dashed line indicates the ground truth. Notably, the testing data is incrementally added to the training set, leading to an uneven data distribution. However, our approach dynamically selects meaningful training samples (stationary, forward, left turn, and right turn) to ensure a balanced and diverse memory buffer, which facilitates online adaptation in real-time. Figure 6 demonstrates that our pipeline can effectively adapt in real + +![](images/4b63c9ccc0628c2355594851d97f4ab6b28cbb044e132d84ef97c4e5da7e5e57.jpg) +Fig. 6. Performance of online adaptation. The "SubT UGV model" adapts to the "Tartan drive model" within 105 seconds. Training data is incrementally added to the buffer, represented by the red trajectory at the bottom, and categorized into segments of stationary, forward motion, left turn, and right turn, each with varying time lengths. The height of each 3D bar indicates the duration of each segment in the buffer. After buffer selection, the data distribution becomes uniform, facilitating rapid adaptation. + +![](images/2b0b9fcfc6d22a05c7779779ab7e0f20088a853f16f91f332a3b5e4101dc53c3.jpg) +Fig. 7. Non-forgetting Characteristics of Our Model. We evaluate our Tartan IMU pretrained model's generalization before and after fine-tuning from the SubT dataset [49] (source domain) to the TartanDrive dataset [34] (target domain). This demonstrates the non-forgetting capability of our LoRA-based fine-tuning adaptation, ensuring effective generalization across different scenarios. + +time, completing the adaptation process in 105 seconds. We also incorporated online adaptation into the SLAM pipeline. Please see supplementary videos for this section. + +# 4.4. Non-forgetting Characteristics Evaluation + +To address Q4, we evaluate 2 models across source and target domains: a fully fine-tuned IMU model and an IMU model fine-tuned with a low-rank adapter (LoRA) [21]. + +In Figure 7(a), we present the odometry results from inference using our base Tartan IMU model without fin-tuning. Figure 7(b) shows model performance in an unseen environment (target domain) outside the pre-training dataset. Both models adapt to the new environment after fine-tuning. In Figure 7(c), we assess model performance on the source domain after domain adaptation. The + +fully fine-tuned expert model performs worse than its original performance, demonstrating catastrophic forgetting. In contrast, the LoRA-based model maintains accurate estimations on the source domain, even after adapting to new knowledge. This demonstrates that LoRA-based approach effectively mitigates the forgetting problem. + +# 5. Conclusion + +This paper introduces Tartan IMU, a foundational model designed for generalizable IMU-based pose estimation across diverse robotic platforms. Tartan IMU improves accuracy by $36\%$ compared to specialized models, leveraging over 100 hours of multi-platform data within its three-stage architecture. It utilizes LoRA fine-tuning to enable efficient domain adaptation with minimal additional parameters. Additionally, Tartan IMU supports online test-time adaptation, allowing for continuous performance improvement during deployment. The combination of these three stages makes Tartan IMU a versatile and robust solution for a wide range of robotic platforms. + +# 6. Limitations + +While TartanIMU exhibits strong generalization across vehicles, drones, and legged robots, it still cannot support arbitrary robotic platforms. However, our experiments show that the car motion head generalizes well to TartanDrive and SubT vehicles. We believe our categories—car, humanoid, quadruped, and drone—encompass most robots. For unseen platforms, introducing a new motion head or leveraging a mixture of existing experts (MoE)[45] presents a promising future direction. At the same time, we plan to explore the integration of both real-world and simulated IMU data to develop a more generalizable model to further improve the generalization capability. Additionally, our GMM-based buffer selection may struggle with complex robot motion patterns. Developing a more robust, robot-agnostic adaptation strategy would be a next step. + +# Acknowledgments + +We sincerely thank the anonymous reviewers for their valuable feedback on this work. This research was supported by funding from the U.S. Army Research Lab under grants W911NF-23-S-0001, W911NF-21-20152, W911NF-24-20125, and W911NF-17-S-0003. + +# References + +[1] Yasin Almalioglu, Mehmet Turan, Muhamad Risqi U Sapatra, Pedro PB de Gusmao, Andrew Markham, and Niki Trigoni. Selfvio: Self-supervised deep monocular visual-inertial odometry and depth estimation. Neural Networks, 150:119-136, 2022. 2 +[2] Amado Antonini, Winter Guerra, Varun Murali, Thomas Sayre-McCord, and Sertac Karaman. The blackbird uav dataset. The International Journal of Robotics Research, 0 (0):0278364920908331, 0. 6 +[3] Amado Antonini, Winter Guerra, Varun Murali, Thomas Sayre-McCord, and Sertac Karaman. The blackbird dataset: A large-scale dataset for uav perception in aggressive flight. In International Symposium on Experimental Robotics, pages 130-139. Springer, 2018. 5 +[4] Omri Asraf, Firas Shama, and Itzik Klein. Pdrnet: A deep-learning pedestrian dead reckoning framework. IEEE Sensors Journal, 22(6):4932-4939, 2021. 2 +[5] Martin Brossard, Axel Barrau, and Silvere Bonnabel. Rinsw: Robust inertial navigation system on wheels. The IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2 +[6] Martin Brossard, Axel Barrau, and Silvère Bonnabel. Ai-imu dead-reckoning. IEEE Transactions on Intelligent Vehicles, 5(4):585-595, 2020. 2, 3, 6 +[7] Russell Buchanan, Varun Agrawal, Marco Camurri, Frank Dellaert, and Maurice Fallon. Deep imu bias inference for robust visual-inertial odometry with factor graphs. IEEE Robotics and Automation Letters, 8(1):41-48, 2022. 2 +[8] Russell Buchanan, Marco Camurri, Frank Dellaert, and Maurice Fallon. Learning inertial odometry for dynamic legged robot state estimation. In Conference on Robot Learning, pages 1575-1584. PMLR, 2022. 2 +[9] Xiya Cao, Caifa Zhou, Dandan Zeng, and Yongliang Wang. Rio: Rotation-equivariance supervised learning of robust inertial odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6614-6623, 2022. 2, 4 +[10] Changhao Chen and Xianfei Pan. Deep learning for inertial positioning: A survey. IEEE Transactions on Intelligent Transportation Systems, 2024. 2 +[11] Changhao Chen, Xiaoxuan Lu, Andrew Markham, and Niki Trigoni. Ionet: Learning to cure the curse of drift in inertial odometry. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2 +[12] Changhao Chen, Yishu Miao, Chris Xiaoxuan Lu, Linhai Xie, Phil Blunsom, Andrew Markham, and Niki Trigoni. Motiontransformer: Transferring neural inertial tracking between domains. In The Conference on Artificial Intelligence (AAAI), 2019. 2 + +[13] Danpeng Chen, Nan Wang, Runsen Xu, Weijian Xie, Hu-jun Bao, and Guofeng Zhang. Rnin-vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 275-283, 2021. 4, 6, 7 +[14] Giovanni Cioffi, Leonard Bauersfeld, Elia Kaufmann, and Davide Scaramuzza. Learned inertial odometry for autonomous drone racing. 2023. 2, 6 +[15] Santiago Cortés, Arno Solin, and Juho Kannala. Deep learning based speed estimation for constraining strapdown inertial navigation on smartphones. In 2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2018. 2 +[16] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 5 +[17] Mahdi Abolfazli Esfahani, Han Wang, Keyu Wu, and Shenghai Yuan. Aboldeepio: A novel deep inertial odometry network for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 21(5):1941-1950, 2019. 2 +[18] Tobias Feigl, Sebastian Kram, Philipp Woller, Ramiz H Siddiqui, Michael Philippsen, and Christopher Mutschler. A bidirectional LSTM for estimating dynamic human velocities from a single imu. In 2019 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1-8. IEEE, 2019. 2 +[19] Christian Forster, Luca Carlone, Frank Dellaert, and Davide Scaramuzza. Imu preintegration on manifold for efficient visual-inertial maximum-a-posteriori estimation. In Robotics: Science and Systems XI, 2015. 1, 2 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 5, 8 +[22] Royina Karegoudra Jayanth, Yinshuang Xu, Ziyun Wang, Evangelos Chatzipantazis, Daniel Gehrig, and Kostas Dani-ilidis. Eqnio: Subequivariant neural inertial odometry, 2024. 3 +[23] Timothee Lesort, Vincenzo Lomonaco, Andrei Stoian, Davide Maltoni, David Filliat, and Natalia Diaz-Rodriguez. Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges, 2019. 7 +[24] Feng Liu, Hongyu Ge, Dan Tao, Ruipeng Gao, and Zhang Zhang. Smartphone-based pedestrian inertial tracking: Dataset, model, and deployment. IEEE Transactions on Instrumentation and Measurement, 2023. 2 +[25] Wenxin Liu, David Caruso, Eddy Iig, Jing Dong, Anastasios I Mourikis, Kostas Daniilidis, Vijay Kumar, and Jakob Engel. Tlio: Tight learned inertial odometry. IEEE Robotics and Automation Letters, 5(4):5653-5660, 2020. 2, 4, 6 +[26] R OpenAI et al. Gpt-4 technical report. ArXiv, 2303:08774, 2023. 2 + +[27] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2 +[28] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE transactions on robotics, 34(4):1004-1020, 2018. 2 +[29] Yuheng Qiu, Chen Wang, Can Xu, Yutian Chen, Xunfei Zhou, Youjie Xia, and Sebastian Scherer. Airimu: Learning uncertainty propagation for inertial odometry. 2023. 2 +[30] Yuheng Qiu, Can Xu, Yutian Chen, Shibo Zhao, Junyi Geng, and Sebastian Scherer. Airio: Learning inertial odometry with enhanced imu feature observability, 2025. 3, 4 +[31] Bingbing Rao, Ehsan Kazemi, Yifan Ding, Devu M Shila, Frank M Tucker, and Liqiang Wang. Ctin: Robust contextual transformer network for inertial navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5413-5421, 2022. 2 +[32] Douglas A. Reynolds. Gaussian mixture models. In Encyclopedia of Biometrics, 2018. 5 +[33] João Paulo Silva do Monte Lima, Hideaki Uchiyama, and Rin-ichiro Taniguchi. End-to-end learning framework for imu-based 6-dof odometry. Sensors, 19(17):3777, 2019. 2 +[34] Matthew Sivaprakasam, Parv Maheshwari, Mateo Guaman Castro, Samuel Triest, Micah Nye, Steve Willits, Andrew Saba, Wenshan Wang, and Sebastian Scherer. Tartandrive 2.0: More modalities and better infrastructure to further self-supervised learning research in off-road driving tasks. arXiv preprint arXiv:2402.01913, 2024. 7, 8 +[35] Scott Sun, Dennis Melamed, and Kris Kitani. *Idol: Inertial deep orientation-estimation and localization*. In *Proceedings of the AAAI Conference on Artificial Intelligence*, pages 6128–6137, 2021. 2, 5, 6 +[36] Niko Sunderhauf, Oliver Brock, Walter Scheirer, Raia Hadsell, Dieter Fox, Jürgen Leitner, Ben Upcroft, Pieter Abbeel, Wolfram Burgard, Michael Milford, and Peter Corke. The limits and potentials of deep learning for robotics. The International Journal of Robotics Research, 37(4-5):405-420, 2018. 7 +[37] Hailiang Tang, Xiaoji Niu, Tisheng Zhang, You Li, and Jingnan Liu. Odonet: Untethered speed aiding for vehicle navigation without hardware wheeled odometer. IEEE Sensors Journal, 2022. 3 +[38] Xiaoqiang Teng, Pengfei Xu, Deke Guo, Yulan Guo, Runbo Hu, Hua Chai, and Didi Chuxing. Arpdr: An accurate and robust pedestrian dead reckoning system for indoor localization on handheld smartphones. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10888-10893. IEEE, 2020. 2 +[39] Yiming Tu and Jin Xie. Undeeplio: Unsupervised deep lidar-inertial odometry. In Asian Conference on Pattern Recognition, pages 189-202. Springer, 2022. 2 + +[40] Brandon Wagstaff and Jonathan Kelly. Lstm-based zero-velocity detection for robust inertial navigation. In 2018 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1–8. IEEE, 2018. 2 +[41] Qu Wang, Haiyong Luo, Langlang Ye, Aidong Men, Fang Zhao, Yan Huang, and Shanghai Ou. Pedestrian heading estimation based on spatial transformer networks and hierarchical LSTM. IEEE Access, 7:162309-162322, 2019. 2 +[42] Yingying Wang, Hu Cheng, and Max Q-H Meng. A2dio: Attention-driven deep inertial odometry for pedestrian localization based on 6d imu. In 2022 International Conference on Robotics and Automation (ICRA), pages 819-825. IEEE, 2022. 2 +[43] Peng Wei, Guoliang Hua, Weibo Huang, Fanyang Meng, and Hong Liu. Unsupervised monocular visual-inertial odometry network. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 2347-2354, 2021. 2 +[44] Xinguo Yu, Ben Liu, Xinyue Lan, Zhuoling Xiao, Shuisheng Lin, Bo Yan, and Liang Zhou. Azupt: Adaptive zero velocity update based on neural networks for pedestrian tracking. In 2019 IEEE Global Communications Conference (GLOBECOM), pages 1-6. IEEE, 2019. 3 +[45] Seniha Esen Yuksel, Joseph N. Wilson, and Paul D. Gader. Twenty years of mixture of experts. IEEE Transactions on Neural Networks and Learning Systems, 23(8):1177-1193, 2012. 8 +[46] Ming Zhang, Mingming Zhang, Yiming Chen, and Mingyang Li. Imu data processing for inertial aided navigation: A recurrent neural network based approach. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3992-3998. IEEE, 2021. 2 +[47] Shibo Zhao, Peng Wang, Hengrui Zhang, Zheng Fang, and Sebastian Scherer. Tp-tio: A robust thermal-inertial odometry with deep thermalpoint. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4505-4512. IEEE, 2020. 1 +[48] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729–8736. IEEE, 2021. 1 +[49] Shibo Zhao, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, et al. Subt-mrs dataset: Pushing slam towards all-weather environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22647–22657, 2024. 1, 5, 6, 7, 8 +[50] Baoding Zhou, Zhining Gu, Fuqiang Gu, Peng Wu, Chengjing Yang, Xu Liu, Linchao Li, Yan Li, and Qingquan Li. Deepvip: Deep learning-based vehicle indoor positioning using smartphones. IEEE Transactions on Vehicular Technology, 2022. 2 \ No newline at end of file diff --git a/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/images.zip b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ec4b78c2404d91ec584e9fe85cbd688696666309 --- /dev/null +++ b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bceb05ad9da46a5921c046608fba72db49d961b63c8b9630c07082ec04fa460c +size 726704 diff --git a/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/layout.json b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d25ef290dd4a22b5cb34ca71871fe5e7d20580fb --- /dev/null +++ b/2025/Tartan IMU_ A Light Foundation Model for Inertial Positioning in Robotics/layout.json @@ -0,0 +1,7679 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 531, + 121 + ], + "type": "text", + "content": "Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "spans": [ + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "text", + "content": "Shibo Zhao" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "inline_equation", + "content": "^{1\\dagger*}" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "text", + "content": ", Sifan Zhou" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "text", + "content": ", Raphael Blanchard" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "text", + "content": ", Yuheng Qiu" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "text", + "content": ", Wenshan Wang" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "text", + "content": ", Sebastian Scherer" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 142, + 520, + 177 + ], + "type": "text", + "content": ", Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 202, + 534, + 418 + ], + "blocks": [ + { + "bbox": [ + 60, + 202, + 534, + 418 + ], + "lines": [ + { + "bbox": [ + 60, + 202, + 534, + 418 + ], + "spans": [ + { + "bbox": [ + 60, + 202, + 534, + 418 + ], + "type": "image", + "image_path": "6d7d98a2eaf60957bb593c4f2df6a5c979b3d7568fac2fbe056b222499820582.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 426, + 546, + 460 + ], + "lines": [ + { + "bbox": [ + 45, + 426, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 45, + 426, + 546, + 460 + ], + "type": "text", + "content": "Fig. 1. t-SNE Visualization of our Tartan IMU Foundation Model. Our foundation model is trained on over 100 hours of diverse IMU data collected from various platforms, including ground vehicles, drones, legged robots, and human motion. This model maps IMU data into a unified high-dimensional shared space, capturing broad, generalizable IMU knowledge." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 470, + 192, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 470, + 192, + 482 + ], + "spans": [ + { + "bbox": [ + 143, + 470, + 192, + 482 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 485, + 290, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 290, + 689 + ], + "type": "text", + "content": "Despite recent advances in deep learning, most existing learning IMU odometry methods are trained on specific datasets, lack generalization, and are prone to overfitting, which limits their real-world application. To address these challenges, we present Tartan IMU, a foundation model designed for generalizable, IMU-based state estimation across diverse robotic platforms. Our approach consists of three-stage: First, a pre-trained foundation model leverages over 100 hours of multi-platform data to establish general motion knowledge, achieving " + }, + { + "bbox": [ + 46, + 485, + 290, + 689 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 46, + 485, + 290, + 689 + ], + "type": "text", + "content": " improvement in ATE over specialized models. Second, to adapt to previously unseen tasks, we use Low-Rank Adaptation (LoRA), allowing positive transfer with only " + }, + { + "bbox": [ + 46, + 485, + 290, + 689 + ], + "type": "inline_equation", + "content": "1.1M" + }, + { + "bbox": [ + 46, + 485, + 290, + 689 + ], + "type": "text", + "content": " trainable parameters. Finally, to support robotics deployment, we introduce online test-time adaptation, which eliminates the boundary between training and testing, allowing the model to continuously \"learn as it operates\" at 200 FPS in real-time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 471, + 516, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 516, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 516, + 483 + ], + "type": "text", + "content": "Project page: https://superodometry.com/tartanimu." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 494, + 387, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 494, + 387, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 494, + 387, + 506 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 509, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 641 + ], + "type": "text", + "content": "The Inertial Measurement Unit (IMU) is a critical component in robotic systems, widely used across platforms such as autonomous drones, vehicles, legged robots, and VR/AR devices for human interaction. By measuring linear acceleration and angular velocity, the IMU enables accurate motion tracking and reliable control across diverse environments. Compared to other positioning solutions such as visual odometry[47] and LiDAR odometry[48], Inertial odometry (IO) offers greater robustness in challenging conditions as it is unaffected by visual degradation such as low lighting and LiDAR degradation such as smoke [49]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "However, current IO solutions rely on double integration to integrate the pose from acceleration and angular velocity, which is susceptible to drift due to factors such as sensor bias, temperature fluctuations, and sensor white noise [19]. Even minor errors can rapidly accumulate during integration, causing significant drift and unbounded er" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "type": "text", + "content": "* Corresponding author. † Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22520" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "rors within seconds [10]. To solve this problem, previous research incorporates domain-specific knowledge such as dead reckoning[10] or other sensor constraints[19]. While these approaches show effectiveness in particular application areas, they are still difficult to address the fundamental drift problem of inertial odometry [10]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 252 + ], + "type": "text", + "content": "Recently, learning-based inertial odometry methods have demonstrated promising performance in addressing challenges of inertial positioning. For example, TLIO [25] has been effective in pedestrian motion estimation, IMO for drone motion estimation [14], AI-IMU [6] for vehicle motion estimation, and deep IMU bias estimation for legged robot motion estimation [7]. However, most of these approaches are still difficult when applied in real-world scenarios. The key challenges can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 253, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 287, + 324 + ], + "type": "text", + "content": "Generalizability: Existing models are tailored to either locomotion-specific or device-specific, which often struggle to generalize in out-of-distribution data. To the best of our knowledge, no current method has been evaluated comprehensively during long-term operation and across various motion patterns and devices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 397 + ], + "type": "text", + "content": "Adapability: Existing learning models struggle to adapt to new motion types or new platforms. Deploying learning-based IO across various platforms requires that models quickly adapt to new motion modalities, even with limited data. This adaptability is essential for the practical deployment of learning-based IO in real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 398, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 542 + ], + "type": "text", + "content": "Drawing from the success of large-scale, high-quality, and diverse data in developing general models that outperform specialized models in natural language processing [26] and visual perception [27], we explore this paradigm in the context of IMU odometry. While this approach has proven effective across many domains, no prior work has applied it to inertial positioning. In fact, IMU data differs fundamentally from language and vision data due to its continuous, dynamic, and noisy nature. Even small changes in motion patterns or devices can create significant domain gaps between the training and testing phases, making IMU generalization particularly challenging." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 543, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 662 + ], + "type": "text", + "content": "In this paper, we try to answer a key question: How to achieve the foundation model for IMU state estimation in robotics? We define an ideal IMU foundation model should be: i) generalizable positioning across robotic platforms, ii) positive knowledge transfer to new platforms, and iii) fast adaptation for new knowledge. We introduce TartanIMU, an IMU foundation model built on a three-stage pretrain-adapt-test framework: pre-training on a large dataset, finetuning on unseen data, and online adaptation for real-world testing (as shown in Figure 1). The main contributions are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "- IMU Foundation Model: We introduce the first IMU foundation model, featuring a shared backbone that enables scalable, cross-platform motion estimation, as illustrated in Figure 1. TartanIMU is trained on a diverse" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 132 + ], + "type": "text", + "content": "dataset of over 100 hours of data collected from various robotic platforms, including ground vehicles, drones, legged robots, and human motion. As a result, it achieves a notable accuracy improvement of " + }, + { + "bbox": [ + 313, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 313, + 72, + 545, + 132 + ], + "type": "text", + "content": " across all platforms, consistently outperforming expert-tuned models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 133, + 545, + 323 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 133, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 228 + ], + "type": "text", + "content": "- Efficient Fine-tuning for Unseen Motion: Leveraging our pre-trained IMU foundation model, it can quickly adapt to new setups with minimal data. To achieve this, we integrated LoRA [21] into our architecture, requiring only 1.1M parameters to adapt unseen scenes. This approach not only mitigates the problem of catastrophic forgetting but also allows flexible insertion of adapters tailored to different robotic platforms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 323 + ], + "type": "text", + "content": "- Online Adaptation: We propose an online adaptation method that dynamically selects motion patterns for fast training. Unlike conventional learning approaches where the pre-trained model remains static during inference, our method eliminates the boundary between the training and testing phases—we learn as we operate. This approach progressively enhances performance and achieves rapid real-time adaptation." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 327, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 545, + 363 + ], + "type": "text", + "content": "We hope that Tartan IMU represents a step towards such general-purpose IMU foundation models that can serve as a foundation for diverse mobile robotic applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 372, + 392, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 392, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 392, + 384 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 388, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 509 + ], + "type": "text", + "content": "Traditional model-based inertial odometry (IO) methods, such as those using kinematic motion models [19], estimate relative motion through double integration. However, these methods are prone to drift over time due to sensor noise, necessitating external corrections such as visual input [28]. Recent data-driven approaches [12, 29, 46] aim to address this issue by modeling and compensating for noise while quantifying uncertainty in IMU measurements. Nevertheless, error accumulation during double integration remains a significant challenge." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "content": "Deep learning has emerged as a promising alternative, driving the development of learning-based IO methods. Early approaches utilized CNNs [11, 24, 50], LSTMs [12, 17, 18, 33, 40, 41], and Transformers [1, 31, 39, 42, 43] to estimate motion. Hybrid methods, such as TLIO [25] and IDOL [35], combined learning-based techniques with kinematic models to enhance robustness and accuracy. Furthermore, learning-based approaches have been applied to tasks like motion classification and human gait estimation [4, 38], with RIO [9] introducing rotation-equivalent augmentation to improve robustness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Despite these advancements, most learning-based IO methods are tailored for pedestrian navigation, limiting their ability to generalize to out-of-distribution scenarios [7]. Recent research has extended their application to diverse platforms, including legged robots [7, 8], wheeled robots [5, 6, 15], and racing drones [14]. For planar-motion" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22521" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 211, + 190 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 211, + 190 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 211, + 190 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 211, + 190 + ], + "type": "image", + "image_path": "1ed936a8e16468a4aabc76762439196e34f922388e28c8630545c9ff207fc3f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 205, + 211, + 294 + ], + "blocks": [ + { + "bbox": [ + 61, + 193, + 210, + 204 + ], + "lines": [ + { + "bbox": [ + 61, + 193, + 210, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 210, + 204 + ], + "type": "text", + "content": "Stage 2: Fine-tune for Unseen Environments" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 205, + 211, + 294 + ], + "lines": [ + { + "bbox": [ + 72, + 205, + 211, + 294 + ], + "spans": [ + { + "bbox": [ + 72, + 205, + 211, + 294 + ], + "type": "image", + "image_path": "3b892534bdf82172d26958675636fcdb1c225303e611ee07b2d8da4ed6dd683e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 317, + 207, + 415 + ], + "blocks": [ + { + "bbox": [ + 70, + 304, + 165, + 314 + ], + "lines": [ + { + "bbox": [ + 70, + 304, + 165, + 314 + ], + "spans": [ + { + "bbox": [ + 70, + 304, + 165, + 314 + ], + "type": "text", + "content": "Stage 3: Online Adaptation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 317, + 207, + 415 + ], + "lines": [ + { + "bbox": [ + 69, + 317, + 207, + 415 + ], + "spans": [ + { + "bbox": [ + 69, + 317, + 207, + 415 + ], + "type": "image", + "image_path": "35123c12c4eb0d1dc6e7af19a636e546a8e4ccded9d869cdb44d1e5d282ca1c6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 221, + 69, + 532, + 190 + ], + "blocks": [ + { + "bbox": [ + 221, + 69, + 532, + 190 + ], + "lines": [ + { + "bbox": [ + 221, + 69, + 532, + 190 + ], + "spans": [ + { + "bbox": [ + 221, + 69, + 532, + 190 + ], + "type": "image", + "image_path": "2c20f73908ab6e63e3addb29c0d29feaee5ca2ed17c05e706f0c7531593960d0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 221, + 191, + 527, + 297 + ], + "blocks": [ + { + "bbox": [ + 221, + 191, + 527, + 297 + ], + "lines": [ + { + "bbox": [ + 221, + 191, + 527, + 297 + ], + "spans": [ + { + "bbox": [ + 221, + 191, + 527, + 297 + ], + "type": "image", + "image_path": "93b8a77a9ce4f2feaabd00b202cadfdbb495dbdf529a5216dcfb1d962d401f92.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 221, + 300, + 526, + 417 + ], + "blocks": [ + { + "bbox": [ + 221, + 300, + 526, + 417 + ], + "lines": [ + { + "bbox": [ + 221, + 300, + 526, + 417 + ], + "spans": [ + { + "bbox": [ + 221, + 300, + 526, + 417 + ], + "type": "image", + "image_path": "14ade9df7871a143737211cc4a9128eb28c69cb8201d6a1b59441c9478c28a1d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 427, + 547, + 463 + ], + "lines": [ + { + "bbox": [ + 45, + 427, + 547, + 463 + ], + "spans": [ + { + "bbox": [ + 45, + 427, + 547, + 463 + ], + "type": "text", + "content": "Fig. 2. Three Learning Stages of Tartan IMU a) Pretrained IMU Model: features a shared backbone to capture generalizable IMU knowledge. b) Efficient Fine-Tuning: utilizes an adapter to enable positive transfer for new tasks. c) Online Adaptation: employs an adaptive memory buffer to support on-the-fly model updates during deployment." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 474, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 545 + ], + "type": "text", + "content": "systems, such as ground robots and pedestrians, zero-velocity updates (ZUPTs) are often integrated with motion models [37], feature extractors [44], or uncertainty quantification frameworks [6] to mitigate drift. To improve the generalization of learning inertial odometry, EqNIO [22] introduces canonical displacement priors in the network design." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 545, + 288, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 654 + ], + "type": "text", + "content": "While these efforts have expanded the scope of learning-based IO, there is currently no IMU odometry model that achieves generalization across diverse platforms, adapts seamlessly to unseen environments, and serves as a foundation for diverse IMU-based motion estimation tasks. Challenges related to generalizability, and adaptability remain significant barriers to the broader adoption of learning-based IMU odometry across a wide range of applications. Further details are provided in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 659, + 103, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 103, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 103, + 671 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 674, + 155, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 155, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 155, + 687 + ], + "type": "text", + "content": "3.1. Pipeline Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "The framework of the proposed Tartan IMU is illustrated in Figure 2. Tartan IMU comprises three key components:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 546, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 522 + ], + "type": "text", + "content": "pre-trained IMU model trained from large-scale datasets, an adapter network fine-tuning for unseen motion, and online test-time adaptation to further boost the accuracy and robustness of odometry during deployment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 529, + 437, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 437, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 437, + 540 + ], + "type": "text", + "content": "3.2. Pretrained IMU Model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 544, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 616 + ], + "type": "text", + "content": "To achieve broad generalization and few-shot learning capabilities, our Tartan IMU model is trained on a diverse array of datasets. The model architecture includes four main components: data alignment and augmentation, a heterogeneous shared IMU backbone, a multi-head for different robots, and a loss function shown as Figure 2 stage 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Data Alignment The datasets we use involve various IMU sensors with differing axis orientations. To ensure consistency, we standardize the coordinate system to X-forward, Y-left, and Z-up in the IMU body frame [30]. We also normalize the sampling frequency to " + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "inline_equation", + "content": "200\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": " across all datasets for consistency, balancing temporal resolution with computational efficiency. This setup offers two key benefits: (i) sufficient temporal resolution to accurately capture motion" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22522" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "dynamics; (ii) a consistent input structure that reduces variability from different sensor sampling rates, supporting uniformity in training and inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 180 + ], + "type": "text", + "content": "Data Augmentation To enhance model generalization, we apply rotation-equivariance [9] in the network. This method assumes that if the IMU axis undergoes a specific rotation, the predicted and ground truth trajectories should transform correspondingly. This alignment and augmentation process is critical in improving the model's overall performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 181, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 289 + ], + "type": "text", + "content": "Heterogeneous Shared Backbone After data alignment and augmentation, we introduce our foundation IMU encoder which employed a ResNet-style architecture [20] to extract latent features from diverse IMU inputs. Given the continuous and regular nature of robotic motion, which often follows specific patterns over time, we utilized an LSTM network to capture these temporal dynamics. The LSTM outputs are then passed to a decoder, which regresses velocity and covariance shown in Figure 2, stage1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 516 + ], + "type": "text", + "content": "In our network's data flow, the earlier ResNet layers function as a heterogeneous* shared backbone, mapping diverse IMU data from different robots into a unified high-dimensional representation space. This shared representation allows IMU data from new platforms to require only minimal data and training to adapt to the shared \"knowledge.\" As a result, this representation space becomes generalizable across diverse robots, capturing broader, world-level knowledge. Our aim is to pre-train foundation models that can map raw IMU signals from various robots into a shared latent space. To showcase that our shared backbone effectively captures motion-specific representations across different robots, we use t-SNE to visualize the learned latent space (see Figure 2, Stage 1). The visualization reveals clear clusters corresponding to four robot types—vehicle (red), drones (green), legged robots (gray), and human (blue)—highlighting the model's ability to learn a unified representation while preserving the distinctions between different motion patterns." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 517, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 684 + ], + "type": "text", + "content": "Multi-head Following the heterogeneous shared backbone, the final component of our network is a robot-decoupled, multi-head design that maps the shared latent representation to the velocity estimation space. During training, different regression heads are assigned to different robot types, benefiting from joint training within the shared backbone, as illustrated in Figure 2, stage 1. This multi-head setup provides two key advantages: (i) it enables the model to learn diverse motion patterns in parallel, enhancing adaptability across different robots during deployment; and (ii) it stabilizes the training process by avoiding competing learning objectives (e.g., drone motion and car motion are significantly different). We adopt two fully connected layers as decoders to regress the velocity of the window and the cor" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "content": "responding covariance. The mathematical formulation of the network is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 97, + 545, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 97, + 545, + 113 + ], + "spans": [ + { + "bbox": [ + 306, + 97, + 545, + 113 + ], + "type": "interline_equation", + "content": "(\\hat {v}, \\hat {u}) = f \\left(\\left(^ {B} \\mathbf {a} _ {n - N}, ^ {B} \\boldsymbol {\\omega} _ {n - N}\\right), \\dots , \\left(^ {B} \\mathbf {a} _ {n}, ^ {B} \\boldsymbol {\\omega} _ {n}\\right), \\mathbf {h} _ {n - N}\\right)", + "image_path": "bf07094f3158e412a244d5d114b3288f1585436e4622086d35be19bdced62dc3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 114, + 465, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 114, + 465, + 129 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 465, + 129 + ], + "type": "interline_equation", + "content": "{ } ^ { B } \\mathbf { a } _ { n } = { } _ { W } ^ { B } \\mathbf { R } _ { n } \\left( { } _ { B } ^ { W } \\mathbf { R } _ { n } \\left( \\mathbf { a } - \\mathbf { b } _ { a } \\right) - { } ^ { W } \\mathbf { g } \\right)", + "image_path": "4861dd20097bc2213a634f526e8db06df0b9833a777d9e182a120879b71151cd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 131, + 545, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 131, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 310, + 131, + 545, + 156 + ], + "type": "interline_equation", + "content": "^ {B} \\boldsymbol {\\omega} _ {n} = \\boldsymbol {\\omega} - \\mathbf {b} _ {g} \\tag {1}", + "image_path": "611b82fa801231b38d65e8a8efc2e3afc27211b4f7febfdf9b7105d8a458f8b5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": " represents the neural network function that processes inputs from the IMU sensor. a denotes acceleration, " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": " is angular velocity from the IMU, and " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{n - N}" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": " refers to the hidden state produced by the LSTM at the previous time step. For each IMU measurement, we remove the gravity vector " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "W\\mathbf{g} = [0,0,9.8]" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": " mapped into the IMU body frame. At each time step, the network predicts the current motion based on the hidden state " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{n - N}" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": " and a local window of " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": " samples of acceleration and angular velocity in the body frame " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": ". The output consists of the estimated relative velocity " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": " and associated uncertainties " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": ". We also assume gyro bias " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_g" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": " and acceleration bias " + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_a" + }, + { + "bbox": [ + 306, + 156, + 545, + 300 + ], + "type": "text", + "content": " are known." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 300, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 383 + ], + "type": "text", + "content": "Loss Function Our network employs relative motion loss functions for each robot head, ensuring high accuracy in local measurements, as shown in Figure 2 Stage 1. The relative loss function helps the network capture motion dynamics within a single local window, improving the smoothness of the predicted trajectory. To optimize this relative loss, we use the Mean Squared Error (MSE), defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 390, + 545, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 390, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 311, + 390, + 545, + 429 + ], + "type": "interline_equation", + "content": "L _ {R L} ^ {M S E} (\\mathbf {v}, \\hat {\\mathbf {v}}) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\sum_ {j = m} ^ {m + M} \\left(v _ {j \\rightarrow j + 1} - \\hat {v} _ {j \\rightarrow j + 1}\\right)\\right) \\tag {2}", + "image_path": "e237e786a061fa7076fb84898229e6fbc01a76fdff1064fe9b37e77d719ea05a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "content": "In contrast to existing works [25][13], we formulate the loss in the body frame, similar to [30] rather than in global coordinates, which allows the model to learn more generalized features instead of simply memorizing fixed global coordinates from training trajectories. Here, " + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\hat{v}_{j\\rightarrow j + 1}" + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "content": " denotes the 3D body velocity output of the network for the " + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "content": "-th window in the body frame, while " + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "v_{j\\rightarrow j + 1}" + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "content": " represents the corresponding ground truth in the same frame. The variable " + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "content": " indicates the batch size during training, " + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "content": " the starting window of the sequence, and " + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "content": " the total number of LSTM windows. In practice, we use a 10-window LSTM, where each window spans 1 second of IMU data at " + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "200\\mathrm{Hz}" + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "content": ", trained with a " + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "1\\mathrm{Hz}" + }, + { + "bbox": [ + 305, + 434, + 545, + 589 + ], + "type": "text", + "content": " supervision signal." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 590, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 590, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 545, + 638 + ], + "type": "text", + "content": "Covariance: The covariance estimation follows the methodology outlined in [25]. The network outputs a three-dimensional vector, where each element represents the logarithm of a diagonal element of the covariance matrix " + }, + { + "bbox": [ + 305, + 590, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 305, + 590, + 545, + 638 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 646, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 646, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 339, + 646, + 545, + 669 + ], + "type": "interline_equation", + "content": "L ^ {\\mathrm {N L L}} = \\frac {1}{2} (v - \\hat {v}) ^ {T} \\Sigma^ {- 1} (v - \\hat {v}) + \\frac {1}{2} \\ln | \\Sigma | \\tag {3}", + "image_path": "201f8656d37147b7c219beb52e0472fbe2a710fc9f321c45051b77f322556d19.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 675, + 504, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 675, + 504, + 687 + ], + "spans": [ + { + "bbox": [ + 306, + 675, + 504, + 687 + ], + "type": "text", + "content": "3.3. Fine-tuning for Unseen Environments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "While our pretrained IMU model is designed to be generalizable across various platforms, we cannot guarantee" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "* Heterogeneity refers to the varying motion patterns across different robot platforms and different IMU devices." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22523" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 212, + 167 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 212, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 212, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 212, + 167 + ], + "type": "image", + "image_path": "d87835894a69f649de1ad71a354f2783b8143511dd9b78eae722246171774bc0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 174, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 547, + 208 + ], + "type": "text", + "content": "Fig. 3. Distribution of motion patterns classified by GMMs. The histograms with overlaid Gaussian fits show clusters based on linear and angular velocities. The GMM identifies motion types: right turns (red) with negative linear velocity Y, left turns (blue) with positive linear velocity Y, stationary states (green) near zero velocities, and forward motion (purple) with positive linear velocities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 214, + 68, + 376, + 167 + ], + "blocks": [ + { + "bbox": [ + 214, + 68, + 376, + 167 + ], + "lines": [ + { + "bbox": [ + 214, + 68, + 376, + 167 + ], + "spans": [ + { + "bbox": [ + 214, + 68, + 376, + 167 + ], + "type": "image", + "image_path": "5166550e55e44de7d117752bb0d97e499a38c6f8f1100d3c8f60b75eca9a1ed9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 68, + 544, + 167 + ], + "blocks": [ + { + "bbox": [ + 379, + 68, + 544, + 167 + ], + "lines": [ + { + "bbox": [ + 379, + 68, + 544, + 167 + ], + "spans": [ + { + "bbox": [ + 379, + 68, + 544, + 167 + ], + "type": "image", + "image_path": "f67119bb06d38602b71242a2168b5bf8371e3c677b35056390242cc4ca731298.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": "\"zero-shot\" capability for every robot platform or motion pattern. Additionally, because training efficiency and real-time performance are critical for most robotic state estimation tasks, fine-tuning every parameter of the pretrained model becomes impractical. To address this, we propose adopting a Low-Rank Adapter (LoRA)[21], which freezes the entire pretrained IMU model and only updates a small, zero-initialized adapter network with 5 million parameters (see Figure 2 Stage 2). To update each pretrained weight matrix " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": " in Tartan IMU, we apply a low-rank decomposition, representing " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "W_0 + \\Delta W = W_0 + BA" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d, k)" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": ". During adaptation, " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "W_0" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": " remains fixed, while " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": " are trainable. Both " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "W_0" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "inline_equation", + "content": "\\Delta W = BA" + }, + { + "bbox": [ + 46, + 220, + 289, + 400 + ], + "type": "text", + "content": " process the input, and their outputs are summed. This yields:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 410, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 410, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 94, + 410, + 287, + 423 + ], + "type": "interline_equation", + "content": "h = W _ {0} x + \\Delta W x = W _ {0} x + B A x \\tag {4}", + "image_path": "719aa9062d5c3a7889ce6e53c69dfd8d534af73eb659f07e1f53f2b356f7c2b6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "text", + "content": "Following LoRA [21], " + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "text", + "content": " is initialized by a random Gaussian distribution and " + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "text", + "content": " at zero, making " + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "text", + "content": " start as zero. We then scale " + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\Delta Wx" + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\frac{\\alpha}{r}" + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "text", + "content": " for controlled adaptation, where " + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "text", + "content": " is a constant in " + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 431, + 287, + 504 + ], + "type": "text", + "content": ". This approach preserves the original knowledge from the pretrained model while progressively incorporating new information during adaption." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 507, + 157, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 157, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 157, + 520 + ], + "type": "text", + "content": "3.4. Online Adaptation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "type": "text", + "content": "While LoRA effectively addresses the domain gap between training and testing, it requires pre-collected datasets, which may not be feasible in real-world operations due to the need for rapid adaptation when domain shifts occur. To overcome this limitation, we propose a novel online learning pipeline that enables our model to adapt to new robotic platforms on-the-fly. We introduce a \"learning within SLAM\" approach, where SLAM acts as a teacher model, providing relative poses as supervision, while our IMU model functions as a student, continuously adapting to reduce the domain gap (see Figure 2 stage 3). Our approach has to overcome two major challenges associated with online adaptation: i) Limited computation time, ii) Restricted data availability" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Adaptive Training Buffer Selection To address challenges " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": ", we found that the key lies in having a diverse and concise dataset training buffer. We introduce a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 220, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 365 + ], + "type": "text", + "content": "dynamic training buffer selection mechanism that preserves diverse samples by focusing on sample similarity. This approach dynamically selects distinct training samples from the input distribution, minimizing overlap and ensuring a balanced dataset without overburdening computational resources. For motion sample similarity comparison, we employ Gaussian Mixture Models (GMMs) [32] to cluster IMU data by distinct motion patterns, as shown in Figure 3. By maintaining a uniform sample distribution across these clusters (see Figure 2, stage 3), the buffer reduces memory load, accelerates online adaptation, and enhances accuracy by learning from diverse motion representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 369, + 388, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 388, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 388, + 383 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 385, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 504 + ], + "type": "text", + "content": "Training Data: We train Tartan IMU on a large-scale, heterogeneous dataset encompassing over 100 hours of real-world IMU data from eight distinct robotic platforms with diverse dynamics. Sourced from prior works such as SubT-MRS [49], IDOL [35], Blackbird [3], and the UZH dataset [16], this dataset includes wheeled robots, drones, legged robots, and human-held devices, providing raw IMU data and ground-truth trajectories. To enhance the model's generalization capability, we applied data alignment and data augmentation described in section 3.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 505, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 553 + ], + "type": "text", + "content": "Experiment Goals: Our approach is designed to achieve strong generalization, efficient fine-tuning, rapid adaptation, and resistance to forgetting for IMU foundation model. To validate these, we examine four key questions:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 553, + 545, + 637 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 305, + 553, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 553, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 553, + 545, + 576 + ], + "type": "text", + "content": "Q1: Does our model generalize across platforms and outperform platform-specific models?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 577, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 577, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 577, + 545, + 600 + ], + "type": "text", + "content": "Q2: Can it adapt to new environments through finetuning with minimal data?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 601, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 601, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 601, + 545, + 624 + ], + "type": "text", + "content": "Q3: Is it capable of real-time adaptation to a new platform with high accuracy?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 624, + 535, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 535, + 637 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 535, + 637 + ], + "type": "text", + "content": "Q4: Can it mitigate catastrophic forgetting problems?" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 639, + 452, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 452, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 452, + 651 + ], + "type": "text", + "content": "4.1. Generalization Evaluation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "To address Q1, we focus on two key aspects for training a more generalized model: (1) heterogeneous shared backbone; (2) large-scale, high-quality, diverse data." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "To illustrate the impact of model architecture, we evaluated various IMU models on data from four motion plat" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22524" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 355 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 355 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 355 + ], + "type": "image", + "image_path": "1550a90c0bee80e0ecc4da7a01af7d1e1347ce22dd1aec2a09d8eef7fd320f90.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 361, + 546, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 546, + 395 + ], + "type": "text", + "content": "Fig. 4. Qualitative analysis of our proposed TartanIMU compared to other specialized IMU models. Our TartanIMU outperforms different specialized IMU models across various robotic platforms, including wheeled (first row), human (second row), quadruped (third row), and drone (fourth row) systems. Our method provides the best pose estimation results shown in red trajectory." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 399, + 538, + 463 + ], + "blocks": [ + { + "bbox": [ + 53, + 399, + 538, + 463 + ], + "lines": [ + { + "bbox": [ + 53, + 399, + 538, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 538, + 463 + ], + "type": "table", + "html": "
Robot PlatformAI-IMU [6]RNIN-VIO [13]TLIO [25]IMO [14]TartanIMUImprovement
ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATE ↓T-RTE ↓ATET-RTE
Wheeled (Car)[49]7.683.337.825.068.123.738.123.736.172.52↑19.63 %↑24.36%
Handled (Human)[35]8.264.897.625.616.964.8210.195.674.321.95↑47.69%↑60.05%
Legged (Dog)[49]3.231.603.101.583.611.733.351.641.460.79↑54.83%↑50.63%
Aerial (Drone)[2]4.141.454.321.513.931.403.721.343.321.04↑19.81%↑28.97%
", + "image_path": "786a249e96eec1e59c47096ed42304e9b4e7dce92490eb691588e4b662d67f2c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 51, + 511, + 282, + 612 + ], + "blocks": [ + { + "bbox": [ + 46, + 468, + 545, + 502 + ], + "lines": [ + { + "bbox": [ + 46, + 468, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 545, + 502 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of pose estimation between the proposed TartanIMU and existing IMU methods [6, 13, 14, 25]. Our TartanIMU outperforms different specialized IMU models, achieving a " + }, + { + "bbox": [ + 46, + 468, + 545, + 502 + ], + "type": "inline_equation", + "content": "35.5\\%" + }, + { + "bbox": [ + 46, + 468, + 545, + 502 + ], + "type": "text", + "content": " on ATE and " + }, + { + "bbox": [ + 46, + 468, + 545, + 502 + ], + "type": "inline_equation", + "content": "41.0\\%" + }, + { + "bbox": [ + 46, + 468, + 545, + 502 + ], + "type": "text", + "content": " on T-RTE across various robotic platforms including wheeled, handled, legged and aerial systems." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 511, + 282, + 612 + ], + "lines": [ + { + "bbox": [ + 51, + 511, + 282, + 612 + ], + "spans": [ + { + "bbox": [ + 51, + 511, + 282, + 612 + ], + "type": "table", + "html": "
Robot PlatformTartan IMU
ATE↓T-RTE↓
SingleAllImproveSingleAllImprove
Wheeled[49]7.516.17↑22%3.172.52↑26%
Handled[35]6.644.32↑54%4.241.95↑117%
Legged[49]3.641.46↑150%1.940.79↑146%
Aerial[2]3.843.32↑16%1.361.04↑30%
Average5.413.82↑42%2.681.57↑71%
", + "image_path": "c4ec1d344c9c63594bab88e74ddad4df68607f1a9d3dcca5e48a6f4876b8e43c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 615, + 287, + 672 + ], + "lines": [ + { + "bbox": [ + 46, + 615, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 287, + 672 + ], + "type": "text", + "content": "Table 2. Accuracy comparison of our Tartan IMU pre-trained model trained with data from single-robot vs. all-robots. These results confirm that training in a wide variety of robot data greatly enhances the model's performance: up to " + }, + { + "bbox": [ + 46, + 615, + 287, + 672 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 46, + 615, + 287, + 672 + ], + "type": "text", + "content": " average improvement on ATE and " + }, + { + "bbox": [ + 46, + 615, + 287, + 672 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 46, + 615, + 287, + 672 + ], + "type": "text", + "content": " average improvement on T-RTE" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "forms: aerial, wheeled, legged, and handheld. For a fair comparison, all models—AI-IMU [6], RNIN-VIO [13], TLIO [25], and IMO [14]—were trained on the same large" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 514, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 480, + 525 + ], + "type": "text", + "content": "dataset and tested on identical unseen data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "As illustrated in Fig. 4, our IMU pre-trained model, designed with a heterogeneous shared backbone network, demonstrated substantial few-shot learning capabilities, allowing it to generalize well across different robotic platforms. Without any additional fine-tuning, it was tested on four distinct systems—aerial, wheeled, legged, and handheld—and consistently outperformed models specifically optimized for each of these applications. This model provided more accurate and reliable odometry estimates across a variety of motion patterns. Notably, it achieved an average improvement of " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "35.5\\%" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " in Absolute Trajectory Error (ATE) and " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "41.0\\%" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " in Time-Relative Trajectory Error (T-RTE) over the second-best performing model. These results underscore the effectiveness of the heterogeneous shared backbone architecture in achieving robust, platform" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22525" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 85, + 531, + 268 + ], + "blocks": [ + { + "bbox": [ + 219, + 71, + 356, + 79 + ], + "lines": [ + { + "bbox": [ + 219, + 71, + 356, + 79 + ], + "spans": [ + { + "bbox": [ + 219, + 71, + 356, + 79 + ], + "type": "text", + "content": "Fine-tuning Our Model on Unseen Dataset" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 85, + 531, + 268 + ], + "lines": [ + { + "bbox": [ + 60, + 85, + 531, + 268 + ], + "spans": [ + { + "bbox": [ + 60, + 85, + 531, + 268 + ], + "type": "image", + "image_path": "1c9ebf24a8d5eaf117ecbdabd22f5d373b16a3af26353a05bee20578d08f7340.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 547, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 547, + 315 + ], + "type": "text", + "content": "Fig. 5. Comparison of Broader Generalization Results of Our IMU Pre-trained Model with a SOTA Model on Unseen Dataset We fine-tuned both the IMU pre-trained model and the SOTA model [13] from the source domain (SubT dataset [49]) to the target domain (Tartandrive dataset [34]). The quantitative comparison show that our model adapts quicker, requiring " + }, + { + "bbox": [ + 46, + 271, + 547, + 315 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 46, + 271, + 547, + 315 + ], + "type": "text", + "content": " fewer iterations compared to the SOTA method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 286, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 334 + ], + "type": "text", + "content": "agnostic generalization across diverse robotic motion types." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 550 + ], + "type": "text", + "content": "We also evaluated the impact of data diversity on model performance. Specifically, we compared the results of our IMU-pretrained model when trained on data from a single platform versus data from various robotic types. As shown in Table.2, models trained on diverse datasets demonstrated substantial improvements. For instance, in the \"quadruped\" sequence, the model trained on data from aerial, wheeled, legged, and handheld platforms significantly outperformed the model trained solely on legged robot data, with Absolute Trajectory Error (ATE) reduced by up to " + }, + { + "bbox": [ + 46, + 335, + 287, + 550 + ], + "type": "inline_equation", + "content": "150\\%" + }, + { + "bbox": [ + 46, + 335, + 287, + 550 + ], + "type": "text", + "content": " and Translational Relative Trajectory Error (T-RTE) improving by " + }, + { + "bbox": [ + 46, + 335, + 287, + 550 + ], + "type": "inline_equation", + "content": "146\\%" + }, + { + "bbox": [ + 46, + 335, + 287, + 550 + ], + "type": "text", + "content": ". These results confirm that training on a broad range of robotic data significantly enhances model performance, even when there are notable differences between the training and testing domains. This highlights the critical importance of large-scale, high-quality, and diverse datasets in improving the generalization capabilities of the IMU model (see supplementary video)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 555, + 269, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 269, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 269, + 567 + ], + "type": "text", + "content": "4.2. Fine-tuning Evaluation for Unseen Motion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 664 + ], + "type": "text", + "content": "Although our model was trained on a large dataset (see Q1), it still encounters generalization challenges when deployed on new robotic platforms with differing motion patterns. To address Q2, we employed LORA with minimal data to evaluate our model's adaptability to an unseen platform. Specifically, our model was trained exclusively on the SubT dataset [49] and tested on the unseen TartanDrive dataset [34]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "The results, displayed in Figure 5, demonstrate the model's performance on a test sequence excluded from the training dataset. We compared our model against a state-of-the-art (SOTA) baseline model [13], trained on the same" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 323, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 394 + ], + "type": "text", + "content": "SubT dataset under identical conditions. Over a 60-epoch fine-tuning period (Figure 5), our model achieved significantly more accurate odometry predictions, closely aligning with the ground truth (dashed line). In contrast, while the SOTA model showed some improvement, it retained considerable trajectory errors at the 60-epoch mark." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "text", + "content": "These findings indicate that our IMU pre-trained model adapts to new domains faster and with " + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "text", + "content": " fewer iterations compared to the SOTA method, making it highly effective for real-world applications across diverse robotic platforms and unfamiliar environments (see supplementary video)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 459, + 470, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 459, + 470, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 459, + 470, + 471 + ], + "type": "text", + "content": "4.3. Online Adaptation Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 510 + ], + "type": "text", + "content": "Q3 poses a difficult challenge, as it involves continual learning [23, 36], that requires updating the model effectively to balance prior knowledge with new experience." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 511, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 546, + 713 + ], + "type": "text", + "content": "In this experiment, our foundation model was initially trained on a UGV operating at a maximum speed of " + }, + { + "bbox": [ + 304, + 511, + 546, + 713 + ], + "type": "inline_equation", + "content": "5\\mathrm{m / s}" + }, + { + "bbox": [ + 304, + 511, + 546, + 713 + ], + "type": "text", + "content": " while testing was conducted on an off-road vehicle reaching speeds up to " + }, + { + "bbox": [ + 304, + 511, + 546, + 713 + ], + "type": "inline_equation", + "content": "15\\mathrm{m / s}" + }, + { + "bbox": [ + 304, + 511, + 546, + 713 + ], + "type": "text", + "content": " -introducing a substantial domain shift. To address this, we implemented an adaptive training buffer selection strategy, detailed in Sec. 3.4. As shown in Figure 6, the model undergoes real-time adaptation from the \"SubT UGV model\" to the \"Tartan drive car model.\" In the figure, the blue line represents the estimated trajectory from the Tartan IMU model, while the dashed line indicates the ground truth. Notably, the testing data is incrementally added to the training set, leading to an uneven data distribution. However, our approach dynamically selects meaningful training samples (stationary, forward, left turn, and right turn) to ensure a balanced and diverse memory buffer, which facilitates online adaptation in real-time. Figure 6 demonstrates that our pipeline can effectively adapt in real" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22526" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 216 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 216 + ], + "type": "image", + "image_path": "4b63c9ccc0628c2355594851d97f4ab6b28cbb044e132d84ef97c4e5da7e5e57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 219, + 546, + 265 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 546, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 546, + 265 + ], + "type": "text", + "content": "Fig. 6. Performance of online adaptation. The \"SubT UGV model\" adapts to the \"Tartan drive model\" within 105 seconds. Training data is incrementally added to the buffer, represented by the red trajectory at the bottom, and categorized into segments of stationary, forward motion, left turn, and right turn, each with varying time lengths. The height of each 3D bar indicates the duration of each segment in the buffer. After buffer selection, the data distribution becomes uniform, facilitating rapid adaptation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 270, + 276, + 456 + ], + "blocks": [ + { + "bbox": [ + 59, + 270, + 276, + 456 + ], + "lines": [ + { + "bbox": [ + 59, + 270, + 276, + 456 + ], + "spans": [ + { + "bbox": [ + 59, + 270, + 276, + 456 + ], + "type": "image", + "image_path": "2b0b9fcfc6d22a05c7779779ab7e0f20088a853f16f91f332a3b5e4101dc53c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 463, + 287, + 531 + ], + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 531 + ], + "type": "text", + "content": "Fig. 7. Non-forgetting Characteristics of Our Model. We evaluate our Tartan IMU pretrained model's generalization before and after fine-tuning from the SubT dataset [49] (source domain) to the TartanDrive dataset [34] (target domain). This demonstrates the non-forgetting capability of our LoRA-based fine-tuning adaptation, ensuring effective generalization across different scenarios." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": "time, completing the adaptation process in 105 seconds. We also incorporated online adaptation into the SLAM pipeline. Please see supplementary videos for this section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 579, + 266, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 266, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 266, + 592 + ], + "type": "text", + "content": "4.4. Non-forgetting Characteristics Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 630 + ], + "type": "text", + "content": "To address Q4, we evaluate 2 models across source and target domains: a fully fine-tuned IMU model and an IMU model fine-tuned with a low-rank adapter (LoRA) [21]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "In Figure 7(a), we present the odometry results from inference using our base Tartan IMU model without fin-tuning. Figure 7(b) shows model performance in an unseen environment (target domain) outside the pre-training dataset. Both models adapt to the new environment after fine-tuning. In Figure 7(c), we assess model performance on the source domain after domain adaptation. The" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 273, + 546, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 546, + 346 + ], + "type": "text", + "content": "fully fine-tuned expert model performs worse than its original performance, demonstrating catastrophic forgetting. In contrast, the LoRA-based model maintains accurate estimations on the source domain, even after adapting to new knowledge. This demonstrates that LoRA-based approach effectively mitigates the forgetting problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 351, + 379, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 379, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 379, + 363 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 367, + 546, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 546, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 546, + 511 + ], + "type": "text", + "content": "This paper introduces Tartan IMU, a foundational model designed for generalizable IMU-based pose estimation across diverse robotic platforms. Tartan IMU improves accuracy by " + }, + { + "bbox": [ + 304, + 367, + 546, + 511 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 304, + 367, + 546, + 511 + ], + "type": "text", + "content": " compared to specialized models, leveraging over 100 hours of multi-platform data within its three-stage architecture. It utilizes LoRA fine-tuning to enable efficient domain adaptation with minimal additional parameters. Additionally, Tartan IMU supports online test-time adaptation, allowing for continuous performance improvement during deployment. The combination of these three stages makes Tartan IMU a versatile and robust solution for a wide range of robotic platforms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 517, + 380, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 380, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 380, + 529 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "content": "While TartanIMU exhibits strong generalization across vehicles, drones, and legged robots, it still cannot support arbitrary robotic platforms. However, our experiments show that the car motion head generalizes well to TartanDrive and SubT vehicles. We believe our categories—car, humanoid, quadruped, and drone—encompass most robots. For unseen platforms, introducing a new motion head or leveraging a mixture of existing experts (MoE)[45] presents a promising future direction. At the same time, we plan to explore the integration of both real-world and simulated IMU data to develop a more generalizable model to further improve the generalization capability. Additionally, our GMM-based buffer selection may struggle with complex robot motion patterns. Developing a more robust, robot-agnostic adaptation strategy would be a next step." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22527" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 144, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 144, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 144, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 86, + 287, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 86, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 47, + 86, + 287, + 142 + ], + "type": "text", + "content": "We sincerely thank the anonymous reviewers for their valuable feedback on this work. This research was supported by funding from the U.S. Army Research Lab under grants W911NF-23-S-0001, W911NF-21-20152, W911NF-24-20125, and W911NF-17-S-0003." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 158, + 107, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 158, + 107, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 158, + 107, + 170 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 174, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 174, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 174, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 174, + 287, + 228 + ], + "type": "text", + "content": "[1] Yasin Almalioglu, Mehmet Turan, Muhamad Risqi U Sapatra, Pedro PB de Gusmao, Andrew Markham, and Niki Trigoni. Selfvio: Self-supervised deep monocular visual-inertial odometry and depth estimation. Neural Networks, 150:119-136, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 229, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 229, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 229, + 288, + 272 + ], + "type": "text", + "content": "[2] Amado Antonini, Winter Guerra, Varun Murali, Thomas Sayre-McCord, and Sertac Karaman. The blackbird uav dataset. The International Journal of Robotics Research, 0 (0):0278364920908331, 0. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 287, + 328 + ], + "type": "text", + "content": "[3] Amado Antonini, Winter Guerra, Varun Murali, Thomas Sayre-McCord, and Sertac Karaman. The blackbird dataset: A large-scale dataset for uav perception in aggressive flight. In International Symposium on Experimental Robotics, pages 130-139. Springer, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 360 + ], + "type": "text", + "content": "[4] Omri Asraf, Firas Shama, and Itzik Klein. Pdrnet: A deep-learning pedestrian dead reckoning framework. IEEE Sensors Journal, 22(6):4932-4939, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 361, + 287, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 287, + 405 + ], + "type": "text", + "content": "[5] Martin Brossard, Axel Barrau, and Silvere Bonnabel. Rinsw: Robust inertial navigation system on wheels. The IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 406, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 406, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 406, + 287, + 437 + ], + "type": "text", + "content": "[6] Martin Brossard, Axel Barrau, and Silvère Bonnabel. Ai-imu dead-reckoning. IEEE Transactions on Intelligent Vehicles, 5(4):585-595, 2020. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 438, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 438, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 53, + 438, + 287, + 482 + ], + "type": "text", + "content": "[7] Russell Buchanan, Varun Agrawal, Marco Camurri, Frank Dellaert, and Maurice Fallon. Deep imu bias inference for robust visual-inertial odometry with factor graphs. IEEE Robotics and Automation Letters, 8(1):41-48, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 483, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 483, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 483, + 287, + 525 + ], + "type": "text", + "content": "[8] Russell Buchanan, Marco Camurri, Frank Dellaert, and Maurice Fallon. Learning inertial odometry for dynamic legged robot state estimation. In Conference on Robot Learning, pages 1575-1584. PMLR, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 526, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 287, + 581 + ], + "type": "text", + "content": "[9] Xiya Cao, Caifa Zhou, Dandan Zeng, and Yongliang Wang. Rio: Rotation-equivariance supervised learning of robust inertial odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6614-6623, 2022. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "type": "text", + "content": "[10] Changhao Chen and Xianfei Pan. Deep learning for inertial positioning: A survey. IEEE Transactions on Intelligent Transportation Systems, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[11] Changhao Chen, Xiaoxuan Lu, Andrew Markham, and Niki Trigoni. Ionet: Learning to cure the curse of drift in inertial odometry. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Changhao Chen, Yishu Miao, Chris Xiaoxuan Lu, Linhai Xie, Phil Blunsom, Andrew Markham, and Niki Trigoni. Motiontransformer: Transferring neural inertial tracking between domains. In The Conference on Artificial Intelligence (AAAI), 2019. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[13] Danpeng Chen, Nan Wang, Runsen Xu, Weijian Xie, Hu-jun Bao, and Guofeng Zhang. Rnin-vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 275-283, 2021. 4, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "text", + "content": "[14] Giovanni Cioffi, Leonard Bauersfeld, Elia Kaufmann, and Davide Scaramuzza. Learned inertial odometry for autonomous drone racing. 2023. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 172, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 227 + ], + "type": "text", + "content": "[15] Santiago Cortés, Arno Solin, and Juho Kannala. Deep learning based speed estimation for constraining strapdown inertial navigation on smartphones. In 2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[16] Jeffrey Delmerico, Titus Cieslewski, Henri Rebecq, Matthias Faessler, and Davide Scaramuzza. Are we ready for autonomous drone racing? the UZH-FPV drone racing dataset. In IEEE Int. Conf. Robot. Autom. (ICRA), 2019. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "text", + "content": "[17] Mahdi Abolfazli Esfahani, Han Wang, Keyu Wu, and Shenghai Yuan. Aboldeepio: A novel deep inertial odometry network for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 21(5):1941-1950, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 380 + ], + "type": "text", + "content": "[18] Tobias Feigl, Sebastian Kram, Philipp Woller, Ramiz H Siddiqui, Michael Philippsen, and Christopher Mutschler. A bidirectional LSTM for estimating dynamic human velocities from a single imu. In 2019 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1-8. IEEE, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 425 + ], + "type": "text", + "content": "[19] Christian Forster, Luca Carlone, Frank Dellaert, and Davide Scaramuzza. Imu preintegration on manifold for efficient visual-inertial maximum-a-posteriori estimation. In Robotics: Science and Systems XI, 2015. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 514, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 557 + ], + "type": "text", + "content": "[22] Royina Karegoudra Jayanth, Yinshuang Xu, Ziyun Wang, Evangelos Chatzipantazis, Daniel Gehrig, and Kostas Dani-ilidis. Eqnio: Subequivariant neural inertial odometry, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "text", + "content": "[23] Timothee Lesort, Vincenzo Lomonaco, Andrei Stoian, Davide Maltoni, David Filliat, and Natalia Diaz-Rodriguez. Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges, 2019. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "text", + "content": "[24] Feng Liu, Hongyu Ge, Dan Tao, Ruipeng Gao, and Zhang Zhang. Smartphone-based pedestrian inertial tracking: Dataset, model, and deployment. IEEE Transactions on Instrumentation and Measurement, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "text", + "content": "[25] Wenxin Liu, David Caruso, Eddy Iig, Jing Dong, Anastasios I Mourikis, Kostas Daniilidis, Vijay Kumar, and Jakob Engel. Tlio: Tight learned inertial odometry. IEEE Robotics and Automation Letters, 5(4):5653-5660, 2020. 2, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[26] R OpenAI et al. Gpt-4 technical report. ArXiv, 2303:08774, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22528" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 171 + ], + "type": "text", + "content": "[27] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 173, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 173, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 287, + 205 + ], + "type": "text", + "content": "[28] Tong Qin, Peiliang Li, and Shaojie Shen. Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE transactions on robotics, 34(4):1004-1020, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 207, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 207, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 207, + 287, + 240 + ], + "type": "text", + "content": "[29] Yuheng Qiu, Chen Wang, Can Xu, Yutian Chen, Xunfei Zhou, Youjie Xia, and Sebastian Scherer. Airimu: Learning uncertainty propagation for inertial odometry. 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 287, + 274 + ], + "type": "text", + "content": "[30] Yuheng Qiu, Can Xu, Yutian Chen, Shibo Zhao, Junyi Geng, and Sebastian Scherer. Airio: Learning inertial odometry with enhanced imu feature observability, 2025. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 287, + 330 + ], + "type": "text", + "content": "[31] Bingbing Rao, Ehsan Kazemi, Yifan Ding, Devu M Shila, Frank M Tucker, and Liqiang Wang. Ctin: Robust contextual transformer network for inertial navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5413-5421, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 354 + ], + "type": "text", + "content": "[32] Douglas A. Reynolds. Gaussian mixture models. In Encyclopedia of Biometrics, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "text", + "content": "[33] João Paulo Silva do Monte Lima, Hideaki Uchiyama, and Rin-ichiro Taniguchi. End-to-end learning framework for imu-based 6-dof odometry. Sensors, 19(17):3777, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 455 + ], + "type": "text", + "content": "[34] Matthew Sivaprakasam, Parv Maheshwari, Mateo Guaman Castro, Samuel Triest, Micah Nye, Steve Willits, Andrew Saba, Wenshan Wang, and Sebastian Scherer. Tartandrive 2.0: More modalities and better infrastructure to further self-supervised learning research in off-road driving tasks. arXiv preprint arXiv:2402.01913, 2024. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "text", + "content": "[35] Scott Sun, Dennis Melamed, and Kris Kitani. *Idol: Inertial deep orientation-estimation and localization*. In *Proceedings of the AAAI Conference on Artificial Intelligence*, pages 6128–6137, 2021. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 566 + ], + "type": "text", + "content": "[36] Niko Sunderhauf, Oliver Brock, Walter Scheirer, Raia Hadsell, Dieter Fox, Jürgen Leitner, Ben Upcroft, Pieter Abbeel, Wolfram Burgard, Michael Milford, and Peter Corke. The limits and potentials of deep learning for robotics. The International Journal of Robotics Research, 37(4-5):405-420, 2018. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "text", + "content": "[37] Hailiang Tang, Xiaoji Niu, Tisheng Zhang, You Li, and Jingnan Liu. Odonet: Untethered speed aiding for vehicle navigation without hardware wheeled odometer. IEEE Sensors Journal, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[38] Xiaoqiang Teng, Pengfei Xu, Deke Guo, Yulan Guo, Runbo Hu, Hua Chai, and Didi Chuxing. Arpdr: An accurate and robust pedestrian dead reckoning system for indoor localization on handheld smartphones. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10888-10893. IEEE, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[39] Yiming Tu and Jin Xie. Undeeplio: Unsupervised deep lidar-inertial odometry. In Asian Conference on Pattern Recognition, pages 189-202. Springer, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 674 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[40] Brandon Wagstaff and Jonathan Kelly. Lstm-based zero-velocity detection for robust inertial navigation. In 2018 International Conference on Indoor Positioning and Indoor Navigation (IPIN), pages 1–8. IEEE, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[41] Qu Wang, Haiyong Luo, Langlang Ye, Aidong Men, Fang Zhao, Yan Huang, and Shanghai Ou. Pedestrian heading estimation based on spatial transformer networks and hierarchical LSTM. IEEE Access, 7:162309-162322, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[42] Yingying Wang, Hu Cheng, and Max Q-H Meng. A2dio: Attention-driven deep inertial odometry for pedestrian localization based on 6d imu. In 2022 International Conference on Robotics and Automation (ICRA), pages 819-825. IEEE, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "type": "text", + "content": "[43] Peng Wei, Guoliang Hua, Weibo Huang, Fanyang Meng, and Hong Liu. Unsupervised monocular visual-inertial odometry network. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 2347-2354, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 274, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 328 + ], + "type": "text", + "content": "[44] Xinguo Yu, Ben Liu, Xinyue Lan, Zhuoling Xiao, Shuisheng Lin, Bo Yan, and Liang Zhou. Azupt: Adaptive zero velocity update based on neural networks for pedestrian tracking. In 2019 IEEE Global Communications Conference (GLOBECOM), pages 1-6. IEEE, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 330, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 330, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 330, + 545, + 373 + ], + "type": "text", + "content": "[45] Seniha Esen Yuksel, Joseph N. Wilson, and Paul D. Gader. Twenty years of mixture of experts. IEEE Transactions on Neural Networks and Learning Systems, 23(8):1177-1193, 2012. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "type": "text", + "content": "[46] Ming Zhang, Mingming Zhang, Yiming Chen, and Mingyang Li. Imu data processing for inertial aided navigation: A recurrent neural network based approach. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3992-3998. IEEE, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 431, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 545, + 485 + ], + "type": "text", + "content": "[47] Shibo Zhao, Peng Wang, Hengrui Zhang, Zheng Fang, and Sebastian Scherer. Tp-tio: A robust thermal-inertial odometry with deep thermalpoint. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4505-4512. IEEE, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 487, + 545, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 550 + ], + "type": "text", + "content": "[48] Shibo Zhao, Hengrui Zhang, Peng Wang, Lucas Nogueira, and Sebastian Scherer. Super odometry: Imu-centric lidar-visual-inertial estimator for challenging environments. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8729–8736. IEEE, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 553, + 545, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 553, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 308, + 553, + 545, + 619 + ], + "type": "text", + "content": "[49] Shibo Zhao, Yuanjun Gao, Tianhao Wu, Damanpreet Singh, Rushan Jiang, Haoxiang Sun, Mansi Sarawata, Yuheng Qiu, Warren Whittaker, Ian Higgins, et al. Subt-mrs dataset: Pushing slam towards all-weather environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22647–22657, 2024. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 620, + 545, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 620, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 308, + 620, + 545, + 674 + ], + "type": "text", + "content": "[50] Baoding Zhou, Zhining Gu, Fuqiang Gu, Peng Wu, Chengjing Yang, Xu Liu, Linchao Li, Yan Li, and Qingquan Li. Deepvip: Deep learning-based vehicle indoor positioning using smartphones. IEEE Transactions on Vehicular Technology, 2022. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22529" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_content_list.json b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7e0cc5111a76018d6847d797f145aad447ca1469 --- /dev/null +++ b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_content_list.json @@ -0,0 +1,1698 @@ +[ + { + "type": "text", + "text": "Task Preference Optimization: Improving Multimodal Large Language Models with Vision Task Alignment", + "text_level": 1, + "bbox": [ + 86, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziang Yan\\*2,1, Zhilin Li\\*3,1, Yinan He\\*1", + "bbox": [ + 331, + 202, + 645, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenting Wang $^{4,1}$ , Kunchang Li $^{5,1}$ , Xinhao Li $^{6,1}$ , Xiangyu Zeng $^{6,1}$", + "bbox": [ + 228, + 220, + 750, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zilei Wang $^{3}$ , Yali Wang $^{5,1}$ , Yu Qiao $^{1}$ , Limin Wang $^{6,1}$ , Yi Wang $^{\\dagger,7}$", + "bbox": [ + 232, + 239, + 746, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai AI Laboratory $^{2}$ Zhejiang University $^{3}$ University of Science and Technology of China $^{4}$ Shanghai Jiao Tong University", + "bbox": [ + 91, + 258, + 890, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences Nanjing University Shanghai Innovation Institute", + "bbox": [ + 91, + 275, + 887, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/OpenGVLab/TPO", + "bbox": [ + 346, + 295, + 632, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current multimodal large language models (MLLMs) struggle with fine-grained or precise understanding of visuals although they give comprehensive perception and reasoning in a spectrum of vision applications. Recent studies either develop tool-using or unify specific visual tasks into the autoregressive framework, often at the expense of overall multimodal performance. To address this issue and enhance MLLMs with visual tasks in a scalable fashion, we propose Task Preference Optimization (TPO), a novel method that utilizes differentiable task preferences derived from typical fine-grained visual tasks. TPO introduces learnable task tokens that establish connections between multiple task-specific heads and the MLLM. By leveraging rich visual labels during training, TPO significantly enhances the MLLM's multimodal capabilities and task-specific performance. Through multi-task co-training within TPO, we observe synergistic benefits that elevate individual task performance beyond what is achievable through single-task training methodologies. Our instantiation of this approach with VideoChat and LLaVA demonstrates an overall $14.6\\%$ improvement in multimodal performance compared to baseline models. Additionally, MLLM-TPO demonstrates robust zero-shot capabilities across various tasks, performing comparably to state-of-the-art supervised models.", + "bbox": [ + 75, + 377, + 473, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 786, + 209, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal large language models (MLLMs) show impressive visual perception and reasoning capabilities, with applications in personal assistants [68], embodied systems [17], and scientific discovery [6]. Considering the", + "bbox": [ + 75, + 811, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e5025be807bb324267d565245547b627928c16cd143e0e1fed961d01a1e5b01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 358, + 527, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Show the tracking trajectory of cup with candy in middle. Which cup is the candy under, left one, middle one or right one?", + "bbox": [ + 542, + 348, + 769, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c3801e4a772805f647d6a89a8afce4327ab7ebb890e61751a2d7808bcb98f56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 347, + 883, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The cups are shuffled in the video, and the candy is under the middle cup.", + "bbox": [ + 514, + 398, + 799, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/557ea63bd3e484136e7dfeda25df6c6a64e6adb3d465243e4158f0d6592931c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 397, + 844, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cff516a6b12e6815e2d186d4a735a81fd8693a1de6e3fd5c49c7acd465e582ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 854, + 395, + 885, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d11b7e46cf72f07c856552c1f6f0c2bb9392b4f45278ffa2e1ef7237042e5b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 422, + 844, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The candy is under the candy is under the the left cup. first cup on the left.", + "bbox": [ + 508, + 470, + 846, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ad6c3de098d56f274573d4269296257f9180c82982ae5ef2019dd74b6290f7a.jpg", + "image_caption": [ + "Figure 1. TPO uses differentiable task preferences from dense visual supervisions via task-specific heads to enhance MLLMs in fine-grained understanding." + ], + "image_footnote": [], + "bbox": [ + 852, + 469, + 883, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "growing expectations of users in more accurate and detailed perception, taking the shell game as an example, further improving the generality of MLLM requires fine-grained knowledge representation beyond words, e.g. MLLM implicitly embeds how to track keys indicated by users.", + "bbox": [ + 496, + 568, + 892, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing studies address the enhancement of perceptual granularity in MLLMs by focusing on specific visual tasks (e.g. temporal grounding, segmentation, tracking) via MLLMs. They usually fine-tune MLLMs on more task data in text format or enable MLLMs to activate the corresponding task heads. Shikra [7] applies MLLM to localization tasks, transforms object coordinates into dialogue formats, and learns them autoregressively. Meanwhile, TimeChat [63] and VideoChat-T [96] treat event timestamps as text for autoregressive prediction, thereby endowing MLLMs with temporal grounding. They do improve specific task performance significantly, while at the cost of multimodal performance more or less. This is counterintuitive as seminal research proves different visual tasks are correlated and training them together often yields mutual gains [29, 47, 66]. We conjecture that the presentation of different tasks influences this training and that the conflict", + "bbox": [ + 496, + 643, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author.", + "bbox": [ + 99, + 886, + 338, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29880", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "arises from the learning discrepancies between discrete textual tokens and visually dense predictions. Our experiments in Section 4 validate that a decoupled representation design can effectively address this issue.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the multimodal capabilities of MLLMs, we explore optimization methods to meet multiple visual task requirements in an end-to-end manner. We propose task preference optimization (TPO), incorporating visual task knowledge into the MLLMs by jointly maximizing the likelihoods of visual task estimations and multimodal dialogue. Inspired by direct preference optimization [60] (DPO) and related methods, we treat visual task annotations as human preferences in particular demands, as shown in Figure 2. DPO aids LLM (or MLLM) in aligning with human preferences through a binary classification that directs the model to generate responses that people prefer. Similarly, TPO enhances MLLMs' visual sensing capabilities through differentiable task optimizations that guide MLLMs to yield dense predictions closely resembling human perception. To achieve this, TPO concretizes MLLM-specific visual perceptions into corresponding task tokens, disentangled from MLLM representation. Then TPO fine-tunes these task tokens and updates the MLLM accordingly.", + "bbox": [ + 75, + 152, + 470, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, TPO appends visual task heads to the partial output of the MLLM, using several learnable task tokens as inputs for these corresponding heads. During training, TPO enables the MLLM to first distinguish and activate the appropriate task tokens based on user instructions. Subsequently, TPO jointly trains the task tokens and their corresponding heads to enhance the MLLM's understanding of visual tasks. Finally, TPO trains the entire model—including task tokens and heads—on both multimodal and visual-specific task data, promoting the perception and reasoning capabilities of the MLLM. Additionally, we note that multi-task co-training yields greater improvements than single-task training.", + "bbox": [ + 75, + 441, + 468, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TPO demonstrates scalability across various MLLM approaches, encompassing a wide range of visual task categories and data quantities. We validate the effectiveness of TPO within widely used MLLMs, such as LLaVA [48] and VideoChat [43], as detailed in Section 4.3. By fine-tuning these open-source MLLMs with TPO, we enhance their visual understanding capabilities and improve dialogue performance. Additionally, we explore several key spatiotemporal and spatial perception tasks, including spatial grounding, tracking, and temporal grounding. Our findings indicate that these tasks can mutually enhance each other's performance, particularly contributing to the improvement of multimodal dialogue capabilities.", + "bbox": [ + 75, + 640, + 470, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as:", + "bbox": [ + 96, + 838, + 369, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a new training method for multimodal large language models, referred to as Task Preference Optimization (TPO). This method leverages supervised in", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e3b5d96dc60e98838535b1935eaa8b58307386009df569516318a6558db9d88b.jpg", + "image_caption": [ + "Figure 2. Comparison of Learning Method. A solid line indicates data flow, and a dotted line represents feedback. $\\star$ and $\\text{心}$ denote modules that are frozen and unfrozen." + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 890, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "formation from visual task-specific data to optimize the MLLM through the corresponding heads, resulting in significantly enhanced multimodal perception and reasoning performance. Specifically, TPO achieves an average improvement of $14.6\\%$ across multiple image and video multimodal benchmarks [19, 38, 45, 54, 104].", + "bbox": [ + 509, + 290, + 890, + 380 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- TPO effectively equips MLLM with the capability to address several key visual perception tasks through the introduced task heads. MLLM-TPO achieves comparable performance in spatial grounding, moment retrieval, highlight detection, referring segmentation, and tracking comparable to expert models across various benchmarks.", + "- TPO demonstrates scalability across various employed MLLMs, task heads, and scales of task data. We validate the effectiveness of TPO in multiple mainstream MLLMs, such as VideoChat2 [45] and LLaVA [39, 48]. Notably, multi-task joint training based on TPO enhances both multimodal performance and individual visual task, with improvements becoming increasingly significant as additional appropriate heads are introduced. Furthermore, the performance of the MLLM and task heads improves when scaling task data." + ], + "bbox": [ + 500, + 380, + 890, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 633, + 640, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Foundation Model. Vision foundation models [44, 66, 69, 71, 72, 74, 78, 84, 87, 103] are designed to be adaptable for various downstream tasks through extensive pre-training on large-scale and diverse datasets. VideoPrism [103] achieves the leading results in various video tasks by combining video-text contrastive learning and spatiotemporal token reconstruction using a dataset of public and proprietary videos. InternVideo2 [78] utilizes masked reconstruction, cross-modal contrastive learning, and next-token prediction to enhance the model's perceptiveness, semantic understanding, and reasoning capabilities.", + "bbox": [ + 496, + 657, + 890, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on vision foundation models, some studies [41, 51, 77, 106, 107] intend to incorporate downstream task heads into this framework and expect end-to-end task training. Unified-IO [51] introduces a versatile model capable of handling a wide range of tasks across vision and lan", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29881", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5f3fae02b4b65871ccf8c68d6b7127b7fd4b1a99e935fd534ef46c873e6c026c.jpg", + "image_caption": [ + "Figure 3. Overall Pipeline of TPO. The architecture of Task Preference Optimization (TPO) consists of four main components: (1) a vision encoder, (2) a connector, (3) a large language model, and (4) a series of visual task heads. Differently colored flame symbols indicate which components are unfrozen at various stages of the training process." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 885, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "guage domains, requiring minimal or no task-specific modifications or additional parameters. Uni-Perceivers [41, 106, 107] formulate different tasks to find the maximum likelihood target for each input through the representation similarity regardless of their modality. Nevertheless, these generalist models are limited to predefined tasks and cannot support flexible open-ended task customization based on language instructions.", + "bbox": [ + 75, + 415, + 472, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Large Language Model. The effective understanding and reasoning of LMMs has attracted the attention of numerous researchers. Limited to its input modality, researchers expand the visual capabilities of LLM, leading to MLLMs. Seminal works, such as BLIP-2 [42], LLaVA [48], and mPLUG-Owl [89], introduce image captioning and visual question answering based on LLM using visual instruction-tuning data. Some video-based MLLMs have been proposed, such as VideoChat [43], VideoChatGPT [52], and Video-LLaMA [99], enabling LLM to gain video understanding capabilities by encoding multiple video frames and using video instruction data.", + "bbox": [ + 75, + 551, + 468, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vanilla MLLMs have achieved impressive results in visual-related dialogues, but barely address some fine-grained tasks, such as segmentation and temporal grounding. To address this challenge, MLLMs usually take one of the following pipelines: pixel-to-sequence (P2S) and pixel-to-embedding (P2E). For P2S methods [7, 63, 70, 79, 90], MLLM directly outputs the textual predictions. TimeChat [63] introduces a time-aware frame encoder and a sliding video Q-former to enhance temporal perception. For P2E methods [3, 35, 75, 82, 97], MLLM compresses and inputs visuals into the downstream decoder for estima", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions. LISA [35], NExT-Chat [97], and VideoLISA [3] introduce SAM [33] as a segmentation tool, using a special token as a prompt to connect MLLM and SAM. Vision-LLM v2 [82] designs routing tokens and super-link queries to bridge MLLM with multiple decoders.", + "bbox": [ + 496, + 415, + 893, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alignment in MLLM. Aligning MLLM with human preferences or values is crucial for MLLM's development. Recent works [67, 94, 105] introduce alignment approaches to MLLM, including proximal policy optimization (PPO) [64] and direct preference optimization (DPO) [60], as shown in Figure 2. They usually exploit proprietary models like GPT4-V to build visual preference datasets and then tune MLLMs using PPO or DPO. Llava-RLHF [67] incorporates PPO into a MLLM framework Llava, argued by image captions or question-answers (QA), while Zhou et al. [105] and Zhang et al. [101] give DPO implementations for MLLMs where visual preference data are created by GPT-4V and other open-sourced MLLMs. RLHF-V [94] collects a dense human preference in segmentation and enhances MLLMs with DPO.", + "bbox": [ + 496, + 500, + 892, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 739, + 591, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Preference Optimization (TPO) aims to enable MLLMs to master classical visual perceptions (such as tracking, temporal grounding, etc.) for better task generalization, regarding that many multimodal reasoning tasks require precise visual cues for accurate and reliable responses. As shown in Figure 3, MLLM-TPO has a typical multimodal model $M$ (consisting of a vision encoder $E$ , a vision-language connector $C$ , and a large language model $G$ ) and a task preference model $P$ with a series of visual", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29882", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "task heads $\\{H_{i}\\}_{i = 1\\dots n}$ ( $n$ denotes the task head number). These heads connect with MLLM using the embeddings $\\{\\mathbf{e}_i\\}_{i = 1\\dots n}$ ( $\\mathbf{e}_i = G(\\mathbf{v}_i)\\in \\mathbb{R}^{1\\times C}$ ) transformed from the learnable task-specific tokens $\\{\\mathbf{v}_i\\}_{i = 1\\dots n}$ ( $\\mathbf{v}_i\\in \\mathbb{R}^{1\\times C}$ ) via MLLM.", + "bbox": [ + 75, + 90, + 467, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TPO employs a local-to-global training scheme, first adapting task heads to the MLLM and then training them jointly. Specifically, the MLLM $M$ starts to recognize $\\{\\mathbf{v}_i\\}_{i=1\\dots n}$ by updating $G$ according to user instructions, then we tune $\\{\\mathbf{v}_i\\}_{i=1\\dots n}$ and $\\{H_i\\}_{i=1\\dots n}$ for adapting visual heads to $M$ . Finally, we train both $M$ and $\\{H_i\\}_{i=1\\dots n}$ together. During inference, MLLM-TPO can respond to users' queries in text, and produce structured visual outputs (like masks, timestamps, trajectories, etc.) when users ask (e.g. yielding time results for \"find when the birthday party starts\"). We detail how to form MLLM-TPO structurally and train it in the following.", + "bbox": [ + 75, + 167, + 467, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Task Preference Models", + "text_level": 1, + "bbox": [ + 76, + 359, + 295, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The task preference model (TPM) $P$ contains a series of task tokens $\\{\\mathbf{v}_i\\}_{i = 1\\dots n}$ and heads $\\{H_i\\}_{i = 1\\dots n}$ . Before TPM works, the attached MLLM generates special tokens indicating task types from the input queries. Then TPM dynamically calls the task token $\\mathbf{v}_j$ based on the special token, transforms it to the task embedding $\\mathbf{e}_j$ via the LLM $G$ (the last hidden embedding), and feeds $\\mathbf{e}_j$ to the corresponding visual task head $\\{H_j\\}$ for specific task predictions. Considering existing MLLMs demonstrate remarkable capabilities in common object and scene recognition, yet struggle to accurately locate things or actions, the employed task heads mainly focus on spatiotemporal localization and tracking. Specifically, we give three fundamental task types for compensating mainstream MLLMs' gaps from expert models in visual perceptions: 1) region head, 2) temporal head, and 3) mask head. Their architectures are given below.", + "bbox": [ + 75, + 383, + 467, + 625 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Region Head. A two-layer multilayer perceptron (MLP) with ReLU activations is employed for the region head. It takes embeddings from LLM and regresses them to the bounding box coordinates for spatial grounding.", + "- Temporal Head. This head is composed of a video encoder, text encoder, and temporal grounding model for moment retrieval and highlight detection. The video and text queries are input into the video and text encoders to obtain their features respectively. Then we append the temporal task embedding after the text features. With the new text and video features, the MLLM estimates the start/end time and highlight score of the corresponding moment from the query via the temporal grounding model.", + "- Mask Head. Pixel-level tasks pose significant challenges for MLLMs due to the lack of the corresponding learning data and optimization in MLLMs. To this end, we introduce a specialized mask head, utilizing the image en" + ], + "bbox": [ + 76, + 628, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coder, mask decoder, and memory bank components from SAM2 [61], replacing the prompt encoder with a single MLP layer called the mask adapter. Specifically, for the given visuals and user queries, the features extracted from the aforementioned image encoder and the text representation corresponding to the mask task embeddings from LLM are employed, fed into the mask decoder to produce the desired mask.", + "bbox": [ + 511, + 90, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most known discriminative vision tasks can be addressed by one or a combination of these three task heads. TPM builds the architectural foundations for leveraging existing vision data with annotations to enhance MLLMs.", + "bbox": [ + 496, + 217, + 890, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Task Preference Optimization", + "text_level": 1, + "bbox": [ + 500, + 292, + 764, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TPO improves MLLMs with extra supervisions from visual task heads by back-propagating gradients from heads to update MLLMs using visual task data. It enables the language model $G$ in MLLM to discriminate specific task types when users demand (task assignment). Then, TPO trains TPM $P$ via compact task representations $\\mathbf{e}_i$ (task optimization) from $\\mathbf{v}_i$ . Lastly, we train $M$ and $P$ together, tuning $M$ for the refined spatiotemporal perception according to task preferences from $P$ . Its optimization objective is given as:", + "bbox": [ + 496, + 316, + 890, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {m l l m}} + \\underbrace {\\mathcal {L} _ {\\text {a s s i g n}} \\left(G \\left(\\mathbf {T} _ {\\mathrm {q}}\\right) , \\mathbf {s}\\right)} _ {\\text {T a k A s s i g n m e n t}} + \\sum_ {i = 1} ^ {n} \\underbrace {\\mathcal {L} _ {\\text {t a s k}} \\left(\\mathbf {A} _ {i} , H _ {i} \\left(G (\\mathbf {v} _ {i})\\right)\\right)} _ {\\text {T a k O p t i m i z a t i o n}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 468, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_{\\mathbf{q}}$ and $\\mathbf{s}$ denote a user query (some contain specific task needs like tracking) in the form of text token sequence and the ground truth task indicating token, respectively. $\\mathbf{A}_i$ is the task annotation of the input visual $\\mathbf{X}$ , so it could be a number tuple for describing regional localization and area or mask for delineating object shape and position. Here $(\\mathbf{X},\\mathbf{T}_{\\mathbf{q}},\\mathbf{T}_{\\mathbf{a}},\\mathbf{A}_i)$ stands for an input data tuple to MLLM-TPO. Usually, we have $(\\mathbf{X},\\mathbf{T}_{\\mathbf{q}},\\mathbf{T}_{\\mathbf{a}})$ for typical MLLM training while $(\\mathbf{X},\\mathbf{T}_{\\mathbf{q}},\\mathbf{A}_i)$ for task token and head training. $\\mathcal{L}_{\\mathrm{assign}}$ is the cross-entropy loss. Meanwhile $\\mathcal{L}_{\\mathrm{task}_i}$ varies according to the task, and is usually regression- or classification-related loss.", + "bbox": [ + 496, + 527, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train MLLM-TPO, we propose a 3-stage local-to-global training scheme, given in Figure 3. Stage 1 learns to identify the task type based on user queries. In stage 2, we train task heads along with their corresponding task tokens, respectively. Lastly, we co-train task heads with MLLM by both task data and multimodal conversation data. Our 3-stage training strategy mitigates the risk of degrading the MLLM's general abilities. We describe them as:", + "bbox": [ + 496, + 710, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 1: Task Assignment. We design a variety of dialogue templates for different visual tasks to perform task recognition instruction tuning for LLM. The LLM is trained using LoRA [25] in this stage.", + "bbox": [ + 496, + 847, + 890, + 909 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29883", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2: Vision Task Training. We integrate task tokens and task-specific heads into the model. By training on task-specific data, the model equips the capacity with fine-grained perception and aligns between the task head and the MLLM. The region head, temporal head, mask adapter, and corresponding task tokens are trained respectively. Similarly, the LLM is updated with LoRA.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 3: Multi-task Training. We tune the entire model on the mixed corpus combining multimodal conversations and specific task data. The synergistic training allows gradients from the heads to MLLM, supervised by human annotations in vision tasks. The vision encoder, connector, task tokens, region head, temporal head, mask adapter, and LLM (with LoRA) are joint-trained in this stage.", + "bbox": [ + 75, + 204, + 468, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 324, + 200, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We give the implementation and training & testing specifics of our TPO, and then show its results and ablations.", + "bbox": [ + 75, + 349, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We employ VideoChat2 [45], a video-based MLLM, as the primary framework in experiments. For its configuration, we employ UMT-L [44] as the vision encoder, Q-former in $\\mathrm{BERT}_{\\mathrm{base}}$ [32] as the connector, and Mistral-7B [28] as the language model (LLM).", + "bbox": [ + 75, + 390, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regarding the three task heads, the Region Head is initialized randomly. The Temporal Head utilizes CGDETR [55], with parameters also initialized randomly. The video features input to the temporal head are extracted from the pre-trained InternVideo2 [78], while query features are extracted using the LLM [12]. The Mask Head employs SAM2 [61] and is initialized with its pre-trained weights. Since SAM2 is originally designed solely for tracking, we incorporate a two-layer MLP to encode positioning prompts for spatiotemporal grounding. Additionally, to enable the MLLM to handle the spatial locations provided by the user, we utilize another two-layer MLP to encode the spatial input into the MLLM.", + "bbox": [ + 75, + 465, + 468, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model is trained on a variety of visual task datasets and conversation datasets, as shown in Table 1. More training details are provided in the Appendix.", + "bbox": [ + 75, + 662, + 468, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarks. We evaluate TPO on both general image/video understanding benchmarks (mainly in multiple choice form) and specific visual tasks (e.g. grounding, tracking, and so on). Specifically, we run our model along with other approaches on MVBench, VideoMME, NExTGQA, MLVU, MMIU, and SEED-Bench2. MVBench [45] is designed to evaluate multi-modal fine-grained video understanding tasks for clips (lasting around 8 to 16 seconds), consisting of 20 video tasks relying heavily on temporal perception that can hardly be addressed by single-frame analysis. Video-MME [19] is for evaluating MLLMs in both perception and reasoning across varying lengths of", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bef60dc2a6fcb1b8808b8b3581a15bc324fb77c91ce9f8a61dc0e4cd405ac5af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskSamplesDatasets
Segmentation114.6KMeViS [15], SAMv2 [61]
Temporal Grounding116.5KActivityNet [5], TACoS [62], QVHighlight [36], DiDeMo [24], QuerYD [57], HiREST [95], NLQ [22]
Spatial Grounding540.0KAllseeing-V2 [76], Visual Genome [34], RefCOCO [91], RefCOCO+ [91], RefCOCOg [53]
Conversation3MYouCook2 [14], ActivityNet [5], VideoChat2-IT [45], ShareGPT-4o [11], LLaVA-Hound-DPO [100], ShareGPT4V [8]
", + "bbox": [ + 504, + 88, + 888, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Overview of Datasets Used in TPO for Various Tasks.", + "bbox": [ + 501, + 207, + 888, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "videos, annotated by humans. NExT-GQA [83] builds on NExT-QA by introducing timestamps that are crucial for understanding questions and determining answers. This requires MLLM to perform multi-step reasoning and emphasize a deeper understanding of both visual and textual content. MLVU [104] is constructed from a wide variety of long videos, with lengths ranging from 3 minutes to 2 hours, and includes nine distinct evaluation tasks. MMIU [54] is a comprehensive evaluation suite designed to assess MLLMs across a wide range of multi-image tasks. It encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions. SEED2-Bench2 [38] is a MLLM benchmark, featuring 24K multiple-choice questions with human annotations. It spans 27 evaluation dimensions, assessing both text and image generation.", + "bbox": [ + 496, + 247, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concerning visual tasks, we test spatial grounding, moment retrieval, highlight detection, tracking, and referring segmentation, including 7 related benchmarks [15, 18, 21, 27, 36, 65, 91], and the results from several corresponding state-of-the-art expert models.", + "bbox": [ + 496, + 489, + 890, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. General Understanding Evaluation", + "text_level": 1, + "bbox": [ + 498, + 575, + 805, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Video Understanding. TPO improves its baseline (VideoChat2) on several video understanding benchmarks with a notable margin. As shown in Table 2, VideoChat-TPO, using a 7B LLM and only 16 input frames, achieves a 66.8 average score on MVBench [45], increasing by $6.4\\%$ over VideoChat2 and exceeding the performance of ST-LLM [49] which uses 64 frames. Considering MVBench focuses on characterizing subtle temporal variations involving fine-grained action types, action order, moving direction, and so on, TPO's seamless integration of detailed video understanding (like spatiotemporal grounding and tracking) into its optimized MLLM makes it effectively handle these fine-grained tasks.", + "bbox": [ + 496, + 598, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On VideoMME [19], VideoChat-TPO outperforms the compared models, achieving a $9.3\\%$ improvement over VideoChat2 and demonstrating significant gains across short and medium-length videos, regardless of subtitle availability. On MLVU [104], VideoChat-TPO achieves a M-AVG score of 54.7, surpassing VideoChat2 by $10.2\\%$ . This confirms the effectiveness of TPO's enhanced percep", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29884", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/82b1c3d0d062e96100534244ae2736ebe180b7a8f66e83057cac96d518b184bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelLLM ParamsFramesMVBench [45]VideoMME [19]MLVU [104]
OverallShortMediumLong
AVGw/o s.w/ s.w/o s.w/ s.w/o s.w/ s.w/o s.w/ s.M-AVG
TimeChat [63]7B9638.534.336.939.143.131.833.932.133.630.9
Video-LLAVA [46]7B843.041.141.946.947.338.740.437.837.947.3
ShareGPT4Video [9]7B1651.239.943.648.353.636.339.335.037.946.4
LLaVA-Next-Video [102]7B1644.038.040.844.647.437.739.431.935.639.3
ST-LLM [49]7B6454.937.942.345.748.436.841.431.336.9-
PLLAVA-34B [85]34B1658.140.035.047.236.238.235.934.732.953.6
Chat-UniVi [30]7B6440.840.645.945.751.241.347.339.143.4-
VideoChat2(baseline) [45]7B1660.439.543.848.352.837.039.433.239.244.5
VideoChat-TPO7B1666.8 (+6.4)48.8 (+9.3)53.8 (+10.0)58.864.946.750.041.046.454.7 (+10.2)
", + "bbox": [ + 78, + 88, + 890, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance on Multimodal Video Understanding. We compare our model to others using LLMs of the same generation or 16-frame input. w/o s. indicates without subtitle, while w s. indicates with subtitle. M-AVG refers to the mean average of MLVU.", + "bbox": [ + 76, + 252, + 892, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tions of long-form video evaluations. These results across three benchmarks demonstrate the notable advancements in multimodal video understanding achieved through TPO.", + "bbox": [ + 75, + 306, + 470, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Grounded Video QA. Table 3 shows that VideoChat-TPO outperforms other models, achieving superior accuracy (Acc) and intersection over union (IoU) scores in NExT-GQA [83]. Its intersection over prediction (IoP) scores are comparable to those of LLoVi [98], which employs large, closed-source commercial models like GPT-4 [1]. The high Acc@IoP score not only reflects the enhanced capability of TPO-optimized VideoChat in effectively understanding and interpreting video content, but also demonstrates its robustness in handling complex reasoning scenarios. Furthermore, the higher Acc@GQA score indicates that the model successfully integrates fine-grained temporal grounding with complex reasoning tasks, enabling it to accurately provide temporal clues necessary for inferring answers. From this perspective, TPO training framework significantly allows MLLMs to overcome their limitations in identifying and localizing temporal elements within video data. This enhancement positions MLLM-TPO with a competitive edge in temporal-related reasoning.", + "bbox": [ + 73, + 377, + 472, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multimodal Multi-image Understanding. To explore TPO's effect on multi-image understanding, we test it on the MMIU and SEED-Bench2, as shown in Table 4. On MMIU, VideoChat-TPO achieved an overall score of 40.2, a $5.2\\%$ improvement over VideoChat2. Besides temporal changes, MMIU evaluates models' spatial sensing and semantic relations in scenes. On SEED-Bench2, across 27 performance indicators, VideoChat-TPO achieves $41.7\\%$ improvement on average performance compared to VideoChat2. VideoChat-TPO's superior performance on multi-image understanding compared to LLaVA-Interleave and InternLM-XComposer demonstrates that TPO's vision enhancements improve the MLLM's spatial perception and image understanding.", + "bbox": [ + 75, + 688, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Vision Task Evaluation", + "text_level": 1, + "bbox": [ + 498, + 306, + 714, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moment Retrieval. Moment Retrieval is to locate the target segments in a video based on queries. Table 5 and 6 compare the zero-shot and fine-tuned moment retrieval performance of VideoChat-TPO against other expert models and MLLMs. In zero-shot settings, VideoChat-TPO achieves a $R@1$ ( $IoU = 0.5$ ) score of 40.2 on the Charades-STA [21]. It surpasses the previous state-of-the-art (SOTA) MLLM specialized for temporal tasks, ChatVTG [59], and the expert model, UniVTG [47]. This demonstrates VideoChat-TPO's ability to accurately locate video moments corresponding to given text queries, thereby enhancing the practical applicability of MLLMs.", + "bbox": [ + 496, + 330, + 893, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Highlight Detection. Highlight detection generates salient scores for emphasizing frames based on the input language query. We compare the fine-tuning highlight detection performance of VideoChat-TPO with that of other expert models and MLLMs, as shown in Table 6. VideoChat-TPO notably outperforms the previous state-of-the-art MLLM (such as TimeChat [63]) on both Charades-STA and QVHighlight. For expert models like QD-DETR and UniVTG, VideoChat-TPO beats them on Charades-STA non-trivially and achieves a comparable performance on QVHighlight. This also demonstrates the progress in extending MLLMs to broad temporal tasks.", + "bbox": [ + 496, + 527, + 893, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial Grounding. To verify the fine-grained localization ability of the model, we run the spatial grounding task which inputs textual descriptions into the model and gets the corresponding bounding box on RefCOCO [91]. As shown in Table 7, we compare VideoChat-TPO with pixel-to-sequence models, i.e., VisionLLM-H [75], pixel-to-embedding methods, i.e., NExT-Chat [97], and expert models, i.e., G-DINO [50]. VideoChat-TPO performs better than the pixel-to-embedding methods by using only a simple task head and achieves comparable performance to pixel-to-sequence models fine-tuned on a large amount of spatial grounding data, as well as to specialized models.", + "bbox": [ + 496, + 723, + 895, + 905 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29885", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dffe59f164e7302a05a21b978b3137098655f9f5e87c8a166016e6c64aa51726.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAcc@IoPAcc@GQAmIoPIoP@0.3IoP@0.5mIoUIoU@0.3IoU@0.5
VIOLETv2 [20]54.912.823.625.123.33.14.31.3
SeViLA [93]72.516.629.534.722.921.729.213.8
LangRepo [31]59.617.131.3-28.718.5-12.2
FrozenBiLM NG+ [88]73.817.524.228.523.79.613.56.1
VideoStreaming [58]57.417.832.2-31.019.3-13.3
LLoVi [98]65.924.337.3-36.920.0-15.3
HawkEye [79]-----25.737.019.5
VideoChat-TPO77.725.535.647.532.827.741.223.4
", + "bbox": [ + 81, + 88, + 563, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e57fe6e6565d9a2d55e8ca11e5b090b23a36ace39419386959fef58c2de47753.jpg", + "table_caption": [ + "Table 3. Performance on Grounded QA." + ], + "table_footnote": [], + "table_body": "
ModelMMIU [54]SEED2I [38]SEED2M [38]
LLaVA-v1.5 [48]19.258.339.2
ShareGPT4V [8]18.5--
OpenFlamingo [2]22.336.643.5
LLaVA-Interleave [40]32.4--
VideoChat2 [45]35.026.527.6
VideoChatGPT [52]-38.349.8
InternLM-XComposer [16]21.965.449.8
VideoChat-TPO40.2 (+5.2)67.3 (+40.8)70.0 (+42.4)
", + "bbox": [ + 566, + 88, + 890, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f08653a17a3782dd31721fcf3ddbbcda8945fe6dea4bb53798be42a8ff77aa8.jpg", + "table_caption": [ + "Table 4. Performance on Image Understanding." + ], + "table_footnote": [], + "table_body": "
ModelCharades-STA [21]
R@0.3R@0.5R@0.7mIoU
UniVTG [47]44.125.210.027.1
VideoChat2 [45]38.014.33.824.6
VTimeLLM [26]51.027.511.431.2
TimeChat [63]-32.213.4-
HawkEye [79]50.631.414.533.7
ChatVTG [59]52.733.015.934.9
VideoChat-TPO58.340.218.438.1
", + "bbox": [ + 84, + 215, + 330, + 323 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a45b786cd89462262bcdb77ee4cdffc3b47947fdadeb30e36efc8616fbbdfb49.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCharades-STA [21]QVHighlight [36]
R@0.3R@0.5R@0.7mIoUmAPHIT@1
M-DETR [36]65.852.130.645.535.755.6
QD-DETR [56]-57.332.6-38.962.4
UniVTG [47]72.660.238.652.240.566.3
TimeChat [63]-46.723.7-21.737.9
HawkEye [79]72.558.328.8---
VideoChat-TPO77.065.040.755.038.665.4
", + "bbox": [ + 331, + 217, + 684, + 323 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5f8ad1cb55a8b866d964ae881f62850cf00a1366eafe793d75af0ba19035ab7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsRefCOCO [91]
valtestAtestB
MAntNet ★ [92]76.480.469.3
OFA-L [73]80.083.776.4
G-DINO-L ★ [50]90.693.288.2
VisionLLM-H [75]-86.7-
Shikra-7B [7]87.090.680.2
NExT-Chat-7B [97]85.590.077.9
VideoChat-TPO85.990.881.3
", + "bbox": [ + 687, + 215, + 885, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Zero-Shot Performance on Table 6. Fine-tuning Performance on Moment Re-Moment Retrieval. Gray means no LLM. Retrieval and Highlight Detection. Gray means no LLM.", + "bbox": [ + 81, + 325, + 679, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Spatial Grounding Task. $\\star$ with a refined decoder.", + "bbox": [ + 686, + 325, + 885, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tracking. In the tracking task, the model receives the object coordinates from the first frame and outputs the coordinates for the remaining frames in the video. We evaluate the VideoChat-TPO on the mainstream tracking benchmarks LAOST [18] and GOT-10k [27], as shown in Table 8. In zero-shot testing, VideoChat-TPO ranks best among all MLLMs, even surpassing some fine-tuned expert models notably with around $10\\%$ increase on success rate, such as SiamFC [4] and ATOM [13]. With the TPO training method, the model is optimized from mask sequences that hard to represent by words, allowing the MLLM to achieve strong motion characterization for multimodal generalization in highly dynamic and occluded scenes.", + "bbox": [ + 75, + 380, + 472, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Referring Segmentation. The referring segmentation task requires the model to output the mask sequence corresponding to the specified prompt. This capability is newly activated in VideoChat-TPO, as it is not supported or designed in the original VideoChat2 or SAM2. We compare the referring segmentation performance of VideoChat-TPO with other expert models and MLLMs in Table 9. The zero-shot capability of VideoChat-TPO is comparable to the fine-tuning capabilities of other expert models, i.e. ReferFormer [81], showing its notable open-world video segmentation performance. Through TPO training, MLLM effectively optimizes its understanding in both tracking and segmenting objects indicated by users. It further enables the model to excel in pixel-level tasks, offering perceptual advantages to it in practical applications like robotic control.", + "bbox": [ + 75, + 601, + 468, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 844, + 241, + 858 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze the effectiveness of the key components of TPO and evaluate its scalability.", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Extending to Other MLLMs. To evaluate the effectiveness of the TPO method on different MLLMs, we apply TPO to LLaVA (LLaVA-OneVision [39]) in addition to VideoChat. In LLaVA, fine-grained video understanding is a weak spot. Despite the conversation data used in LLaVA-OneVision being very similar to, or even more extensive than, what we employ, there remains significant room for improvement in its performance on short-term fine-grained understanding tasks. As shown in Table 10, with TPO, LLaVA not only gains capabilities that it originally lacked but also achieves an $8.1\\%$ improvement on MVBench, demonstrating the universality of the TPO method.", + "bbox": [ + 496, + 380, + 890, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task Preference Model vs. Textualized Task Data. We compare the TPO method with the approach of textualizing task data and training it in an autoregressive manner, which is a straightforward and efficient method for MLLMs to learn specific tasks. We use the same data as the TPO method and convert the data related to the task head into conversation format. Due to the limitations of textual representation, much data must be input in a more discrete manner. For instance, in tracking supervision data, we follow Merlin [90] by converting masks into a sequence of spatial coordinates to serve as both the model's input and output. As shown in Table 11, MLLM-TPO achieves a $2.1\\%$ performance improvement on MVBench [45] compared to MLLMs trained with textualized task data. This demonstrates that the task head in the TPO method enables more effective utilization of the original supervision signals compared to the next-token prediction approach, which inevitably incurs information loss when converting fine-grained tasks into textual annotations. The next-token prediction method cannot truly capture the nuanced information present in videos, while TPO allows the MLLM to", + "bbox": [ + 496, + 583, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29886", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1ddd0901546927d15e10310821ff869a618f56fd2e7896e2d0c221823f736589.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelLaSOT [18]GOT-10k [27]
SuccessPnormPOverlapSR0.5SR0.75
SiamFC [4]33.642.033.934.835.39.8
ATOM [13]51.5--55.663.440.2
SiamRPN++ [37]49.656.949.151.861.832.5
SiamFC++ [86]54.462.354.759.569.547.9
LLaVA-1.5 [48]19.416.512.823.520.29.7
Merlin [90]39.840.238.151.455.942.8
VideoChat-TPO69.480.176.970.679.866.0
", + "bbox": [ + 81, + 89, + 385, + 189 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cf6b2ab84f6c59b2891f49da02576b58175b55067db9404c2cf2bad48887f963.jpg", + "table_caption": [ + "Table 8. Performance on Tracking Benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodRef-YouTube-VOS [65]MeViS [15]
J&FJFJ&FJF
ReferFormer [81]62.961.364.631.029.832.2
OnlineRefer [80]62.961.064.7---
LISA [35]52.652.153.0---
VideoLISA [3]63.761.763.744.441.347.6
VideoChat-TPO63.952.375.447.042.651.3
", + "bbox": [ + 395, + 89, + 694, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f81727b7e4fc0a473de5db5e4faacd6339b3157aa9b41692f0898d58ab69588c.jpg", + "table_caption": [ + "Table 9. Performance on Referring Segmentation. TPO is evaluated in a zero-shot on Ref-YouTube." + ], + "table_footnote": [], + "table_body": "
ModelMVBench
LLaVA-OV [39]56.7
LLaVA-OV-TPO64.8 (+8.1)
", + "bbox": [ + 707, + 89, + 887, + 137 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/795164e10745c8629936ac36c5ff534ecefbb76493b2acd509e4a4c4e85e4d53.jpg", + "table_caption": [ + "Table 10. Applying TPO on LLaVA-OV. TPO significantly enhances the ability to understand fine-grained video details." + ], + "table_footnote": [], + "table_body": "
ModelCharades-STA [21]MVBench
R@0.3R@0.5R@0.7mIoUAVG
VideoChat-TPO58.340.218.438.166.8
w/o reasoning data56.438.317.135.666.1
replace by simple head31.517.86.115.465.8
textualized task data33.718.66.216.064.7
", + "bbox": [ + 81, + 231, + 465, + 311 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/707067be93d98c4638e5e8b0e154a301cf638a00823d5f4960acc0a3689db5fa.jpg", + "table_caption": [ + "Table 11. Ablation of Reasoning Data and Head Performance." + ], + "table_footnote": [], + "table_body": "
Exp.TRMCR@0.5Acc@0.5J&F
1✓30.2--
2✓-77.3-
3✓--55.1
4✓✓-80.258.3
5✓✓✓36.781.661.4
6✓✓✓✓40.282.063.9
", + "bbox": [ + 109, + 342, + 434, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 12. Impact of TPO Components and Data. T, R, M, and C denote temporal head, region head, mask head, and conversation data respectively. R1@0.5 means R1@0.5 in Charades-STA, Acc@0.5 represents the mean of Acc@0.5 in all COCO datasets, $\\mathcal{J}\\&\\mathcal{F}$ means $\\mathcal{J}\\&\\mathcal{F}$ in Ref-YouTube-VOS.", + "bbox": [ + 76, + 449, + 468, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learn perceptual information beyond conversational data.", + "bbox": [ + 76, + 535, + 450, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Task Head Performance on TPO. A stronger task head is expected to enhance the multimodal capabilities of TPO more effectively than a weaker one. To further explore the impact of the task head on model performance, we replaced the temporal head from CG-DETR [55] with a simple two-layer MLP. As shown in Table 11, the simple temporal head exhibits a significant performance decline compared to CG-DETR on the corresponding tasks. However, it still achieves a $1.1\\%$ improvement compared to the model without TPO. In contrast to the simpler head, the well-designed head can provide more accurate expert knowledge through its architecture and pre-trained task preference weights, enabling better utilization of the data.", + "bbox": [ + 75, + 559, + 468, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Data Scaling. Concerning foundation or large models, we expect to enhance their model performance by incorporating additional data at scale. In our approach, we expand the existing conversation data by integrating two reasoning datasets [5, 14] with temporal information. By incorporating these datasets, we observe a notable enhancement in the capabilities of the temporal head in Table 11. Furthermore, as shown in Table 12, adding conversation data further improves performance across various", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tasks. Overall, increasing the amount of conversation and fine-grained task data leads to noticeable improvements in model performance, demonstrating the effectiveness of data scaling in training multimodal models with TPO.", + "bbox": [ + 496, + 233, + 890, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Synergistic Gains from Co-training. Table 12 presents the impact of incorporating different visual tasks into TPO on performance. This indicates that the collaborative learning of visual tasks facilitates the transfer of knowledge, ultimately leading to better performance across the board. Overall, these results confirm that co-training not only benefits individual tasks but also creates a synergistic effect that enhances overall capabilities in visual understanding.", + "bbox": [ + 496, + 306, + 890, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 443, + 625, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces Task Preference Optimization (TPO). It enhances the overall multimodal performance of MLLMs by enhancing their precise visual understanding. TPO achieves this by integrating differentiable task preferences derived from fine-grained visual tasks. It introduces learnable task tokens that serve as bridges between multiple task-specific heads and the MLLM. Through the joint optimization of these task tokens, heads, and the MLLM, TPO leads to a substantial improvement in multimodal dialogue capabilities and performance across various visual tasks. Our results demonstrate the effectiveness of TPO in scaling MLLMs with task-specific data and seamlessly integrating them with existing expert vision models. We believe this study clarifies the prerequisites for fusing MLLMs with models and data from the pre-large model era, effectively bridging the gap between expert and generalist models, as well as between generation and understanding.", + "bbox": [ + 496, + 465, + 890, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Currently, TPO focuses exclusively on discriminative visual tasks, overlooking generative ones. Meanwhile, the framework is supervised by human annotations, which neglects potentially valuable unsupervised or self-supervised learning approaches such as contrastive learning [10, 23]. This limitation inherently restricts the scalability of TPO in terms of both task diversity and requirements. While we demonstrate TPO's potential for enhancing MLLMs through increased task-specific data, a comprehensive investigation into the broader impact of this scaling remains a crucial area for future research.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29887", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is partially supported by the National Key R&D Program of China (No. 2022ZD0160102), Jiangsu Frontier Technology R&D Program (No. BF2024076), and the Science and Technology Commission of Shanghai Municipality under Grant No. 23YF1461900.", + "bbox": [ + 76, + 113, + 468, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 210, + 174, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6", + "[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortman, and Ludwig Schmidt. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 7", + "[3] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 3, 8", + "[4] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In ECCV, pages 850–865, 2016. 7, 8", + "[5] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 5, 8", + "[6] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv preprint arXiv:2304.02948, 2023. 1", + "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1, 3, 7", + "[8] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In ECCV, 2024. 5, 7", + "[9] Lin Chen, Xin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, Li Yuan, Yu Qiao, Dahua Lin, Feng Zhao, and Jiaqi Wang. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 6", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learn" + ], + "bbox": [ + 86, + 237, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 8", + "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 5", + "[12] Yiming Cui, Ziqing Yang, and Xin Yao. Efficient and effective text encoding for chinese llama and alpaca. arXiv preprint arXiv:2304.08177, 2023. 5", + "[13] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In CVPR, pages 4660-4669, 2019. 7, 8", + "[14] Pradipto Das, Chenliang Xu, Richard F Doell, and Jason J Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In CVPR, pages 2634-2641, 2013. 5, 8", + "[15] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 5, 8", + "[16] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 7", + "[17] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 1", + "[18] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, pages 5374-5383, 2019. 5, 7, 8", + "[19] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 5, 6", + "[20] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. An empirical study of end-to-end video-language transformers with masked visual modeling. In CVPR, 2023. 7", + "[21] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 5, 6, 7, 8", + "[22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18995-19012, 2022. 5", + "[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 8" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29888", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 5", + "[25] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 4", + "[26] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024. 7", + "[27] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 43(5):1562-1577, 2019. 5, 7, 8", + "[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 5", + "[29] Fan Jiang and Zilei Wang. Sparse sharing relation network for panoptic driving perception. In ACMMM, pages 800-808, 2023. 1", + "[30] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In CVPR, pages 13700-13710, 2024. 6", + "[31] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv preprint arXiv:2403.14622, 2024.7", + "[32] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 5", + "[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3", + "[34] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 5", + "[35] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, pages 9579-9589, 2024. 3, 8", + "[36] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. NeurIPS, 34:11846-11858, 2021. 5, 7", + "[37] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. SiamRPN++: Evolution of siamese visual tracking with very deep networks. In CVPR, pages 4282-4291, 2019. 8", + "[38] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In CVPR, pages 13299-13308, 2024. 2, 5, 7" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 7, 8", + "[40] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7", + "[41] Hao Li, Jinguo Zhu, Xiaohu Jiang, Xizhou Zhu, Hongsheng Li, Chun Yuan, Xiaohua Wang, Yu Qiao, Xiaogang Wang, Wenhai Wang, et al. Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In CVPR, pages 2691-2700, 2023. 2, 3", + "[42] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 3", + "[43] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 3", + "[44] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2, 5", + "[45] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 5, 6, 7", + "[46] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6", + "[47] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 6, 7", + "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 2, 3, 7, 8", + "[49] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In ECCV, 2024. 5, 6", + "[50] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In ECCV, 2024. 6, 7", + "[51] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 2", + "[52] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 3, 7" + ], + "bbox": [ + 509, + 90, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29889", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. 5", + "[54] Fanqing Meng, Jin Wang, Chuanhao Li, Quanfeng Lu, Hao Tian, Jiaqi Liao, Xizhou Zhu, Jifeng Dai, Yu Qiao, Ping Luo, et al. Mmiu: Multimodal multi-image understanding for evaluating large vision-language models. arXiv preprint arXiv:2408.02718, 2024. 2, 5, 7", + "[55] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 5, 8", + "[56] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 7", + "[57] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP, pages 2265-2269. IEEE, 2021. 5", + "[58] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuangrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009, 2024.7", + "[59] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPR, pages 1847-1856, 2024. 6, 7", + "[60] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 36, 2024. 2, 3", + "[61] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4, 5", + "[62] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics, 1:25-36, 2013. 5", + "[63] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. CVPR, abs/2312.02051, 2024. 1, 3, 6, 7", + "[64] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3", + "[65] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223. Springer, 2020. 5, 8", + "[66] Jing Shao, Siyu Chen, Yangguang Li, Kun Wang, Zhenfei Yin, Yinan He, Jianing Teng, Qinghong Sun, Mengya Gao," + ], + "bbox": [ + 86, + 92, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jihao Liu, et al. Intern: A new learning paradigm towards general vision. arXiv preprint arXiv:2111.08687, 2021. 1, 2", + "[67] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 3", + "[68] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1", + "[69] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, pages 10078-10093, 2022. 2", + "[70] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3", + "[71] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, pages 20-36. Springer, 2016. 2", + "[72] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, 2023. 2", + "[73] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, pages 23318-23340. PMLR, 2022. 7", + "[74] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, pages 6312-6322, 2023. 2", + "[75] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2024. 3, 6, 7", + "[76] Weiyun Wang, Yiming Ren, Haowen Luo, Tiantong Li, Chenxiang Yan, Zhe Chen, Wenhai Wang, Qingyun Li, Lewei Lu, Xizhou Zhu, et al. The all-seeing project v2: Towards general relation comprehension of the open world. arXiv preprint arXiv:2402.19474, 2024. 5", + "[77] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In ICCV, pages 1130-1140, 2023. 2", + "[78] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. In ECCV, 2024. 2, 5" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29890", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3, 7", + "[80] Dongming Wu, Tiancai Wang, Yang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In ICCV, pages 2761-2770, 2023. 8", + "[81] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 7, 8", + "[82] Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Wenhai Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. arXiv preprint arXiv:2406.08394, 2024. 3", + "[83] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In CVPR, 2024. 5, 6", + "[84] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pretraining for zero-shot video-text understanding. In EMNLP, pages 6787-6800, 2021. 2", + "[85] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 6", + "[86] Yinda Xu et al. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, pages 140-148, 2020. 8", + "[87] Shen Yan, Tao Zhu, Zirui Wang, Yuan Cao, Mi Zhang, Soham Ghosh, Yonghui Wu, and Jiahui Yu. Video-text modeling with zero-shot transfer from contrastive captioners. ArXiv, abs/2212.04979, 2022. 2", + "[88] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. In NeurIPS, 2022. 7", + "[89] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 3", + "[90] En Yu, Liang Zhao, Yana Wei, Jinrong Yang, Dongming Wu, Lingyu Kong, Haoran Wei, Tiancai Wang, Zheng Ge, Xiangyu Zhang, et al. Merlin: Empowering multimodal llms with foresight minds. In ECCV, pages 425-443. Springer, 2025. 3, 7, 8", + "[91] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 5, 6, 7", + "[92] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In CVPR, pages 1307-1315, 2018. 7" + ], + "bbox": [ + 86, + 92, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[93] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In NeurIPS, 2023. 7", + "[94] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In CVPR, pages 13807-13816, 2024. 3", + "[95] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In CVPR, pages 23056-23065, 2023. 5", + "[96] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In ICLR. 1", + "[97] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An Imm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 3, 6, 7", + "[98] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lllm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 6, 7", + "[99] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, pages 543–553, 2023. 3", + "[100] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 5", + "[101] Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 3", + "[102] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model, 2024. 6", + "[103] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In ICML, 2024. 2", + "[104] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2, 5, 6", + "[105] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. In ICLR Workshop on Reliable and Responsible Foundation Models, 2024. 3", + "[106] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "29891", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conditional moes. In NeurIPS, pages 2664-2678, 2022. 2, 3", + "[107] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 2, 3" + ], + "bbox": [ + 78, + 90, + 467, + 189 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "29892", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_model.json b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9341761d4823c1e09e1dcb2060e2274433b3ce98 --- /dev/null +++ b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_model.json @@ -0,0 +1,2866 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.087, + 0.13, + 0.885, + 0.177 + ], + "angle": 0, + "content": "Task Preference Optimization: Improving Multimodal Large Language Models with Vision Task Alignment" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.203, + 0.646, + 0.221 + ], + "angle": 0, + "content": "Ziang Yan\\*2,1, Zhilin Li\\*3,1, Yinan He\\*1" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.221, + 0.751, + 0.24 + ], + "angle": 0, + "content": "Chenting Wang\\(^{4,1}\\), Kunchang Li\\(^{5,1}\\), Xinhao Li\\(^{6,1}\\), Xiangyu Zeng\\(^{6,1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.24, + 0.748, + 0.258 + ], + "angle": 0, + "content": "Zilei Wang\\(^{3}\\), Yali Wang\\(^{5,1}\\), Yu Qiao\\(^{1}\\), Limin Wang\\(^{6,1}\\), Yi Wang\\(^{\\dagger,7}\\)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.259, + 0.892, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai AI Laboratory \\(^{2}\\)Zhejiang University \\(^{3}\\)University of Science and Technology of China \\(^{4}\\)Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.888, + 0.292 + ], + "angle": 0, + "content": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences Nanjing University Shanghai Innovation Institute" + }, + { + "type": "text", + "bbox": [ + 0.347, + 0.296, + 0.633, + 0.309 + ], + "angle": 0, + "content": "https://github.com/OpenGVLab/TPO" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.474, + 0.755 + ], + "angle": 0, + "content": "Current multimodal large language models (MLLMs) struggle with fine-grained or precise understanding of visuals although they give comprehensive perception and reasoning in a spectrum of vision applications. Recent studies either develop tool-using or unify specific visual tasks into the autoregressive framework, often at the expense of overall multimodal performance. To address this issue and enhance MLLMs with visual tasks in a scalable fashion, we propose Task Preference Optimization (TPO), a novel method that utilizes differentiable task preferences derived from typical fine-grained visual tasks. TPO introduces learnable task tokens that establish connections between multiple task-specific heads and the MLLM. By leveraging rich visual labels during training, TPO significantly enhances the MLLM's multimodal capabilities and task-specific performance. Through multi-task co-training within TPO, we observe synergistic benefits that elevate individual task performance beyond what is achievable through single-task training methodologies. Our instantiation of this approach with VideoChat and LLaVA demonstrates an overall \\(14.6\\%\\) improvement in multimodal performance compared to baseline models. Additionally, MLLM-TPO demonstrates robust zero-shot capabilities across various tasks, performing comparably to state-of-the-art supervised models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.21, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Multimodal large language models (MLLMs) show impressive visual perception and reasoning capabilities, with applications in personal assistants [68], embodied systems [17], and scientific discovery [6]. Considering the" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.359, + 0.528, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.543, + 0.349, + 0.771, + 0.387 + ], + "angle": 0, + "content": "Show the tracking trajectory of cup with candy in middle. Which cup is the candy under, left one, middle one or right one?" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.348, + 0.884, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.399, + 0.8, + 0.423 + ], + "angle": 0, + "content": "The cups are shuffled in the video, and the candy is under the middle cup." + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.398, + 0.846, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.856, + 0.396, + 0.887, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.424, + 0.845, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.471, + 0.848, + 0.497 + ], + "angle": 0, + "content": "The candy is under the candy is under the the left cup. first cup on the left." + }, + { + "type": "image", + "bbox": [ + 0.853, + 0.47, + 0.885, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.511, + 0.893, + 0.553 + ], + "angle": 0, + "content": "Figure 1. TPO uses differentiable task preferences from dense visual supervisions via task-specific heads to enhance MLLMs in fine-grained understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.644 + ], + "angle": 0, + "content": "growing expectations of users in more accurate and detailed perception, taking the shell game as an example, further improving the generality of MLLM requires fine-grained knowledge representation beyond words, e.g. MLLM implicitly embeds how to track keys indicated by users." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Existing studies address the enhancement of perceptual granularity in MLLMs by focusing on specific visual tasks (e.g. temporal grounding, segmentation, tracking) via MLLMs. They usually fine-tune MLLMs on more task data in text format or enable MLLMs to activate the corresponding task heads. Shikra [7] applies MLLM to localization tasks, transforms object coordinates into dialogue formats, and learns them autoregressively. Meanwhile, TimeChat [63] and VideoChat-T [96] treat event timestamps as text for autoregressive prediction, thereby endowing MLLMs with temporal grounding. They do improve specific task performance significantly, while at the cost of multimodal performance more or less. This is counterintuitive as seminal research proves different visual tasks are correlated and training them together often yields mutual gains [29, 47, 66]. We conjecture that the presentation of different tasks influences this training and that the conflict" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.339, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "29880" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "arises from the learning discrepancies between discrete textual tokens and visually dense predictions. Our experiments in Section 4 validate that a decoupled representation design can effectively address this issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.471, + 0.442 + ], + "angle": 0, + "content": "To enhance the multimodal capabilities of MLLMs, we explore optimization methods to meet multiple visual task requirements in an end-to-end manner. We propose task preference optimization (TPO), incorporating visual task knowledge into the MLLMs by jointly maximizing the likelihoods of visual task estimations and multimodal dialogue. Inspired by direct preference optimization [60] (DPO) and related methods, we treat visual task annotations as human preferences in particular demands, as shown in Figure 2. DPO aids LLM (or MLLM) in aligning with human preferences through a binary classification that directs the model to generate responses that people prefer. Similarly, TPO enhances MLLMs' visual sensing capabilities through differentiable task optimizations that guide MLLMs to yield dense predictions closely resembling human perception. To achieve this, TPO concretizes MLLM-specific visual perceptions into corresponding task tokens, disentangled from MLLM representation. Then TPO fine-tunes these task tokens and updates the MLLM accordingly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Specifically, TPO appends visual task heads to the partial output of the MLLM, using several learnable task tokens as inputs for these corresponding heads. During training, TPO enables the MLLM to first distinguish and activate the appropriate task tokens based on user instructions. Subsequently, TPO jointly trains the task tokens and their corresponding heads to enhance the MLLM's understanding of visual tasks. Finally, TPO trains the entire model—including task tokens and heads—on both multimodal and visual-specific task data, promoting the perception and reasoning capabilities of the MLLM. Additionally, we note that multi-task co-training yields greater improvements than single-task training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.471, + 0.837 + ], + "angle": 0, + "content": "TPO demonstrates scalability across various MLLM approaches, encompassing a wide range of visual task categories and data quantities. We validate the effectiveness of TPO within widely used MLLMs, such as LLaVA [48] and VideoChat [43], as detailed in Section 4.3. By fine-tuning these open-source MLLMs with TPO, we enhance their visual understanding capabilities and improve dialogue performance. Additionally, we explore several key spatiotemporal and spatial perception tasks, including spatial grounding, tracking, and temporal grounding. Our findings indicate that these tasks can mutually enhance each other's performance, particularly contributing to the improvement of multimodal dialogue capabilities." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.839, + 0.37, + 0.853 + ], + "angle": 0, + "content": "Our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We propose a new training method for multimodal large language models, referred to as Task Preference Optimization (TPO). This method leverages supervised in" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.091, + 0.892, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.236, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Figure 2. Comparison of Learning Method. A solid line indicates data flow, and a dotted line represents feedback. \\(\\star\\) and \\(\\text{心}\\) denote modules that are frozen and unfrozen." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.291, + 0.892, + 0.381 + ], + "angle": 0, + "content": "formation from visual task-specific data to optimize the MLLM through the corresponding heads, resulting in significantly enhanced multimodal perception and reasoning performance. Specifically, TPO achieves an average improvement of \\(14.6\\%\\) across multiple image and video multimodal benchmarks [19, 38, 45, 54, 104]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.381, + 0.892, + 0.472 + ], + "angle": 0, + "content": "- TPO effectively equips MLLM with the capability to address several key visual perception tasks through the introduced task heads. MLLM-TPO achieves comparable performance in spatial grounding, moment retrieval, highlight detection, referring segmentation, and tracking comparable to expert models across various benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.472, + 0.892, + 0.622 + ], + "angle": 0, + "content": "- TPO demonstrates scalability across various employed MLLMs, task heads, and scales of task data. We validate the effectiveness of TPO in multiple mainstream MLLMs, such as VideoChat2 [45] and LLaVA [39, 48]. Notably, multi-task joint training based on TPO enhances both multimodal performance and individual visual task, with improvements becoming increasingly significant as additional appropriate heads are introduced. Furthermore, the performance of the MLLM and task heads improves when scaling task data." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.381, + 0.892, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.642, + 0.649 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.826 + ], + "angle": 0, + "content": "Vision Foundation Model. Vision foundation models [44, 66, 69, 71, 72, 74, 78, 84, 87, 103] are designed to be adaptable for various downstream tasks through extensive pre-training on large-scale and diverse datasets. VideoPrism [103] achieves the leading results in various video tasks by combining video-text contrastive learning and spatiotemporal token reconstruction using a dataset of public and proprietary videos. InternVideo2 [78] utilizes masked reconstruction, cross-modal contrastive learning, and next-token prediction to enhance the model's perceptiveness, semantic understanding, and reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Based on vision foundation models, some studies [41, 51, 77, 106, 107] intend to incorporate downstream task heads into this framework and expect end-to-end task training. Unified-IO [51] introduces a versatile model capable of handling a wide range of tasks across vision and lan" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29881" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.887, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.347, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Figure 3. Overall Pipeline of TPO. The architecture of Task Preference Optimization (TPO) consists of four main components: (1) a vision encoder, (2) a connector, (3) a large language model, and (4) a series of visual task heads. Differently colored flame symbols indicate which components are unfrozen at various stages of the training process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.473, + 0.538 + ], + "angle": 0, + "content": "guage domains, requiring minimal or no task-specific modifications or additional parameters. Uni-Perceivers [41, 106, 107] formulate different tasks to find the maximum likelihood target for each input through the representation similarity regardless of their modality. Nevertheless, these generalist models are limited to predefined tasks and cannot support flexible open-ended task customization based on language instructions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Multimodal Large Language Model. The effective understanding and reasoning of LMMs has attracted the attention of numerous researchers. Limited to its input modality, researchers expand the visual capabilities of LLM, leading to MLLMs. Seminal works, such as BLIP-2 [42], LLaVA [48], and mPLUG-Owl [89], introduce image captioning and visual question answering based on LLM using visual instruction-tuning data. Some video-based MLLMs have been proposed, such as VideoChat [43], VideoChatGPT [52], and Video-LLaMA [99], enabling LLM to gain video understanding capabilities by encoding multiple video frames and using video instruction data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Vanilla MLLMs have achieved impressive results in visual-related dialogues, but barely address some fine-grained tasks, such as segmentation and temporal grounding. To address this challenge, MLLMs usually take one of the following pipelines: pixel-to-sequence (P2S) and pixel-to-embedding (P2E). For P2S methods [7, 63, 70, 79, 90], MLLM directly outputs the textual predictions. TimeChat [63] introduces a time-aware frame encoder and a sliding video Q-former to enhance temporal perception. For P2E methods [3, 35, 75, 82, 97], MLLM compresses and inputs visuals into the downstream decoder for estima" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.493 + ], + "angle": 0, + "content": "tions. LISA [35], NExT-Chat [97], and VideoLISA [3] introduce SAM [33] as a segmentation tool, using a special token as a prompt to connect MLLM and SAM. Vision-LLM v2 [82] designs routing tokens and super-link queries to bridge MLLM with multiple decoders." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.893, + 0.728 + ], + "angle": 0, + "content": "Alignment in MLLM. Aligning MLLM with human preferences or values is crucial for MLLM's development. Recent works [67, 94, 105] introduce alignment approaches to MLLM, including proximal policy optimization (PPO) [64] and direct preference optimization (DPO) [60], as shown in Figure 2. They usually exploit proprietary models like GPT4-V to build visual preference datasets and then tune MLLMs using PPO or DPO. Llava-RLHF [67] incorporates PPO into a MLLM framework Llava, argued by image captions or question-answers (QA), while Zhou et al. [105] and Zhang et al. [101] give DPO implementations for MLLMs where visual preference data are created by GPT-4V and other open-sourced MLLMs. RLHF-V [94] collects a dense human preference in segmentation and enhances MLLMs with DPO." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.593, + 0.755 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Task Preference Optimization (TPO) aims to enable MLLMs to master classical visual perceptions (such as tracking, temporal grounding, etc.) for better task generalization, regarding that many multimodal reasoning tasks require precise visual cues for accurate and reliable responses. As shown in Figure 3, MLLM-TPO has a typical multimodal model \\( M \\) (consisting of a vision encoder \\( E \\), a vision-language connector \\( C \\), and a large language model \\( G \\)) and a task preference model \\( P \\) with a series of visual" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29882" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.468, + 0.166 + ], + "angle": 0, + "content": "task heads \\(\\{H_{i}\\}_{i = 1\\dots n}\\) (\\(n\\) denotes the task head number). These heads connect with MLLM using the embeddings \\(\\{\\mathbf{e}_i\\}_{i = 1\\dots n}\\) (\\(\\mathbf{e}_i = G(\\mathbf{v}_i)\\in \\mathbb{R}^{1\\times C}\\)) transformed from the learnable task-specific tokens \\(\\{\\mathbf{v}_i\\}_{i = 1\\dots n}\\) (\\(\\mathbf{v}_i\\in \\mathbb{R}^{1\\times C}\\)) via MLLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.349 + ], + "angle": 0, + "content": "TPO employs a local-to-global training scheme, first adapting task heads to the MLLM and then training them jointly. Specifically, the MLLM \\( M \\) starts to recognize \\( \\{\\mathbf{v}_i\\}_{i=1\\dots n} \\) by updating \\( G \\) according to user instructions, then we tune \\( \\{\\mathbf{v}_i\\}_{i=1\\dots n} \\) and \\( \\{H_i\\}_{i=1\\dots n} \\) for adapting visual heads to \\( M \\). Finally, we train both \\( M \\) and \\( \\{H_i\\}_{i=1\\dots n} \\) together. During inference, MLLM-TPO can respond to users' queries in text, and produce structured visual outputs (like masks, timestamps, trajectories, etc.) when users ask (e.g. yielding time results for \"find when the birthday party starts\"). We detail how to form MLLM-TPO structurally and train it in the following." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.36, + 0.296, + 0.374 + ], + "angle": 0, + "content": "3.1. Task Preference Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.468, + 0.626 + ], + "angle": 0, + "content": "The task preference model (TPM) \\( P \\) contains a series of task tokens \\( \\{\\mathbf{v}_i\\}_{i = 1\\dots n} \\) and heads \\( \\{H_i\\}_{i = 1\\dots n} \\). Before TPM works, the attached MLLM generates special tokens indicating task types from the input queries. Then TPM dynamically calls the task token \\( \\mathbf{v}_j \\) based on the special token, transforms it to the task embedding \\( \\mathbf{e}_j \\) via the LLM \\( G \\) (the last hidden embedding), and feeds \\( \\mathbf{e}_j \\) to the corresponding visual task head \\( \\{H_j\\} \\) for specific task predictions. Considering existing MLLMs demonstrate remarkable capabilities in common object and scene recognition, yet struggle to accurately locate things or actions, the employed task heads mainly focus on spatiotemporal localization and tracking. Specifically, we give three fundamental task types for compensating mainstream MLLMs' gaps from expert models in visual perceptions: 1) region head, 2) temporal head, and 3) mask head. Their architectures are given below." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.629, + 0.468, + 0.689 + ], + "angle": 0, + "content": "- Region Head. A two-layer multilayer perceptron (MLP) with ReLU activations is employed for the region head. It takes embeddings from LLM and regresses them to the bounding box coordinates for spatial grounding." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.69, + 0.468, + 0.839 + ], + "angle": 0, + "content": "- Temporal Head. This head is composed of a video encoder, text encoder, and temporal grounding model for moment retrieval and highlight detection. The video and text queries are input into the video and text encoders to obtain their features respectively. Then we append the temporal task embedding after the text features. With the new text and video features, the MLLM estimates the start/end time and highlight score of the corresponding moment from the query via the temporal grounding model." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- Mask Head. Pixel-level tasks pose significant challenges for MLLMs due to the lack of the corresponding learning data and optimization in MLLMs. To this end, we introduce a specialized mask head, utilizing the image en" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.629, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "coder, mask decoder, and memory bank components from SAM2 [61], replacing the prompt encoder with a single MLP layer called the mask adapter. Specifically, for the given visuals and user queries, the features extracted from the aforementioned image encoder and the text representation corresponding to the mask task embeddings from LLM are employed, fed into the mask decoder to produce the desired mask." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.218, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Most known discriminative vision tasks can be addressed by one or a combination of these three task heads. TPM builds the architectural foundations for leveraging existing vision data with annotations to enhance MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.293, + 0.766, + 0.309 + ], + "angle": 0, + "content": "3.2. Task Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.455 + ], + "angle": 0, + "content": "TPO improves MLLMs with extra supervisions from visual task heads by back-propagating gradients from heads to update MLLMs using visual task data. It enables the language model \\(G\\) in MLLM to discriminate specific task types when users demand (task assignment). Then, TPO trains TPM \\(P\\) via compact task representations \\(\\mathbf{e}_i\\) (task optimization) from \\(\\mathbf{v}_i\\). Lastly, we train \\(M\\) and \\(P\\) together, tuning \\(M\\) for the refined spatiotemporal perception according to task preferences from \\(P\\). Its optimization objective is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.469, + 0.891, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {m l l m}} + \\underbrace {\\mathcal {L} _ {\\text {a s s i g n}} \\left(G \\left(\\mathbf {T} _ {\\mathrm {q}}\\right) , \\mathbf {s}\\right)} _ {\\text {T a k A s s i g n m e n t}} + \\sum_ {i = 1} ^ {n} \\underbrace {\\mathcal {L} _ {\\text {t a s k}} \\left(\\mathbf {A} _ {i} , H _ {i} \\left(G (\\mathbf {v} _ {i})\\right)\\right)} _ {\\text {T a k O p t i m i z a t i o n}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.709 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_{\\mathbf{q}}\\) and \\(\\mathbf{s}\\) denote a user query (some contain specific task needs like tracking) in the form of text token sequence and the ground truth task indicating token, respectively. \\(\\mathbf{A}_i\\) is the task annotation of the input visual \\(\\mathbf{X}\\), so it could be a number tuple for describing regional localization and area or mask for delineating object shape and position. Here \\((\\mathbf{X},\\mathbf{T}_{\\mathbf{q}},\\mathbf{T}_{\\mathbf{a}},\\mathbf{A}_i)\\) stands for an input data tuple to MLLM-TPO. Usually, we have \\((\\mathbf{X},\\mathbf{T}_{\\mathbf{q}},\\mathbf{T}_{\\mathbf{a}})\\) for typical MLLM training while \\((\\mathbf{X},\\mathbf{T}_{\\mathbf{q}},\\mathbf{A}_i)\\) for task token and head training. \\(\\mathcal{L}_{\\mathrm{assign}}\\) is the cross-entropy loss. Meanwhile \\(\\mathcal{L}_{\\mathrm{task}_i}\\) varies according to the task, and is usually regression- or classification-related loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.832 + ], + "angle": 0, + "content": "To train MLLM-TPO, we propose a 3-stage local-to-global training scheme, given in Figure 3. Stage 1 learns to identify the task type based on user queries. In stage 2, we train task heads along with their corresponding task tokens, respectively. Lastly, we co-train task heads with MLLM by both task data and multimodal conversation data. Our 3-stage training strategy mitigates the risk of degrading the MLLM's general abilities. We describe them as:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.848, + 0.892, + 0.91 + ], + "angle": 0, + "content": "Stage 1: Task Assignment. We design a variety of dialogue templates for different visual tasks to perform task recognition instruction tuning for LLM. The LLM is trained using LoRA [25] in this stage." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29883" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "Stage 2: Vision Task Training. We integrate task tokens and task-specific heads into the model. By training on task-specific data, the model equips the capacity with fine-grained perception and aligns between the task head and the MLLM. The region head, temporal head, mask adapter, and corresponding task tokens are trained respectively. Similarly, the LLM is updated with LoRA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.205, + 0.47, + 0.311 + ], + "angle": 0, + "content": "Stage 3: Multi-task Training. We tune the entire model on the mixed corpus combining multimodal conversations and specific task data. The synergistic training allows gradients from the heads to MLLM, supervised by human annotations in vision tasks. The vision encoder, connector, task tokens, region head, temporal head, mask adapter, and LLM (with LoRA) are joint-trained in this stage." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.201, + 0.342 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.38 + ], + "angle": 0, + "content": "We give the implementation and training & testing specifics of our TPO, and then show its results and ablations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.469, + 0.466 + ], + "angle": 0, + "content": "Implementation Details. We employ VideoChat2 [45], a video-based MLLM, as the primary framework in experiments. For its configuration, we employ UMT-L [44] as the vision encoder, Q-former in \\(\\mathrm{BERT}_{\\mathrm{base}}\\) [32] as the connector, and Mistral-7B [28] as the language model (LLM)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.469, + 0.662 + ], + "angle": 0, + "content": "Regarding the three task heads, the Region Head is initialized randomly. The Temporal Head utilizes CGDETR [55], with parameters also initialized randomly. The video features input to the temporal head are extracted from the pre-trained InternVideo2 [78], while query features are extracted using the LLM [12]. The Mask Head employs SAM2 [61] and is initialized with its pre-trained weights. Since SAM2 is originally designed solely for tracking, we incorporate a two-layer MLP to encode positioning prompts for spatiotemporal grounding. Additionally, to enable the MLLM to handle the spatial locations provided by the user, we utilize another two-layer MLP to encode the spatial input into the MLLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.469, + 0.709 + ], + "angle": 0, + "content": "Our model is trained on a variety of visual task datasets and conversation datasets, as shown in Table 1. More training details are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Benchmarks. We evaluate TPO on both general image/video understanding benchmarks (mainly in multiple choice form) and specific visual tasks (e.g. grounding, tracking, and so on). Specifically, we run our model along with other approaches on MVBench, VideoMME, NExTGQA, MLVU, MMIU, and SEED-Bench2. MVBench [45] is designed to evaluate multi-modal fine-grained video understanding tasks for clips (lasting around 8 to 16 seconds), consisting of 20 video tasks relying heavily on temporal perception that can hardly be addressed by single-frame analysis. Video-MME [19] is for evaluating MLLMs in both perception and reasoning across varying lengths of" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.206 + ], + "angle": 0, + "content": "
TaskSamplesDatasets
Segmentation114.6KMeViS [15], SAMv2 [61]
Temporal Grounding116.5KActivityNet [5], TACoS [62], QVHighlight [36], DiDeMo [24], QuerYD [57], HiREST [95], NLQ [22]
Spatial Grounding540.0KAllseeing-V2 [76], Visual Genome [34], RefCOCO [91], RefCOCO+ [91], RefCOCOg [53]
Conversation3MYouCook2 [14], ActivityNet [5], VideoChat2-IT [45], ShareGPT-4o [11], LLaVA-Hound-DPO [100], ShareGPT4V [8]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.502, + 0.208, + 0.89, + 0.221 + ], + "angle": 0, + "content": "Table 1. Overview of Datasets Used in TPO for Various Tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.892, + 0.489 + ], + "angle": 0, + "content": "videos, annotated by humans. NExT-GQA [83] builds on NExT-QA by introducing timestamps that are crucial for understanding questions and determining answers. This requires MLLM to perform multi-step reasoning and emphasize a deeper understanding of both visual and textual content. MLVU [104] is constructed from a wide variety of long videos, with lengths ranging from 3 minutes to 2 hours, and includes nine distinct evaluation tasks. MMIU [54] is a comprehensive evaluation suite designed to assess MLLMs across a wide range of multi-image tasks. It encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions. SEED2-Bench2 [38] is a MLLM benchmark, featuring 24K multiple-choice questions with human annotations. It spans 27 evaluation dimensions, assessing both text and image generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Concerning visual tasks, we test spatial grounding, moment retrieval, highlight detection, tracking, and referring segmentation, including 7 related benchmarks [15, 18, 21, 27, 36, 65, 91], and the results from several corresponding state-of-the-art expert models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.576, + 0.807, + 0.592 + ], + "angle": 0, + "content": "4.1. General Understanding Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Multimodal Video Understanding. TPO improves its baseline (VideoChat2) on several video understanding benchmarks with a notable margin. As shown in Table 2, VideoChat-TPO, using a 7B LLM and only 16 input frames, achieves a 66.8 average score on MVBench [45], increasing by \\(6.4\\%\\) over VideoChat2 and exceeding the performance of ST-LLM [49] which uses 64 frames. Considering MVBench focuses on characterizing subtle temporal variations involving fine-grained action types, action order, moving direction, and so on, TPO's seamless integration of detailed video understanding (like spatiotemporal grounding and tracking) into its optimized MLLM makes it effectively handle these fine-grained tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "On VideoMME [19], VideoChat-TPO outperforms the compared models, achieving a \\(9.3\\%\\) improvement over VideoChat2 and demonstrating significant gains across short and medium-length videos, regardless of subtitle availability. On MLVU [104], VideoChat-TPO achieves a M-AVG score of 54.7, surpassing VideoChat2 by \\(10.2\\%\\). This confirms the effectiveness of TPO's enhanced percep" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29884" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.891, + 0.251 + ], + "angle": 0, + "content": "
ModelLLM ParamsFramesMVBench [45]VideoMME [19]MLVU [104]
OverallShortMediumLong
AVGw/o s.w/ s.w/o s.w/ s.w/o s.w/ s.w/o s.w/ s.M-AVG
TimeChat [63]7B9638.534.336.939.143.131.833.932.133.630.9
Video-LLAVA [46]7B843.041.141.946.947.338.740.437.837.947.3
ShareGPT4Video [9]7B1651.239.943.648.353.636.339.335.037.946.4
LLaVA-Next-Video [102]7B1644.038.040.844.647.437.739.431.935.639.3
ST-LLM [49]7B6454.937.942.345.748.436.841.431.336.9-
PLLAVA-34B [85]34B1658.140.035.047.236.238.235.934.732.953.6
Chat-UniVi [30]7B6440.840.645.945.751.241.347.339.143.4-
VideoChat2(baseline) [45]7B1660.439.543.848.352.837.039.433.239.244.5
VideoChat-TPO7B1666.8 (+6.4)48.8 (+9.3)53.8 (+10.0)58.864.946.750.041.046.454.7 (+10.2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.253, + 0.893, + 0.281 + ], + "angle": 0, + "content": "Table 2. Performance on Multimodal Video Understanding. We compare our model to others using LLMs of the same generation or 16-frame input. w/o s. indicates without subtitle, while w s. indicates with subtitle. M-AVG refers to the mean average of MLVU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.471, + 0.355 + ], + "angle": 0, + "content": "tions of long-form video evaluations. These results across three benchmarks demonstrate the notable advancements in multimodal video understanding achieved through TPO." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.378, + 0.473, + 0.665 + ], + "angle": 0, + "content": "Grounded Video QA. Table 3 shows that VideoChat-TPO outperforms other models, achieving superior accuracy (Acc) and intersection over union (IoU) scores in NExT-GQA [83]. Its intersection over prediction (IoP) scores are comparable to those of LLoVi [98], which employs large, closed-source commercial models like GPT-4 [1]. The high Acc@IoP score not only reflects the enhanced capability of TPO-optimized VideoChat in effectively understanding and interpreting video content, but also demonstrates its robustness in handling complex reasoning scenarios. Furthermore, the higher Acc@GQA score indicates that the model successfully integrates fine-grained temporal grounding with complex reasoning tasks, enabling it to accurately provide temporal clues necessary for inferring answers. From this perspective, TPO training framework significantly allows MLLMs to overcome their limitations in identifying and localizing temporal elements within video data. This enhancement positions MLLM-TPO with a competitive edge in temporal-related reasoning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Multimodal Multi-image Understanding. To explore TPO's effect on multi-image understanding, we test it on the MMIU and SEED-Bench2, as shown in Table 4. On MMIU, VideoChat-TPO achieved an overall score of 40.2, a \\(5.2\\%\\) improvement over VideoChat2. Besides temporal changes, MMIU evaluates models' spatial sensing and semantic relations in scenes. On SEED-Bench2, across 27 performance indicators, VideoChat-TPO achieves \\(41.7\\%\\) improvement on average performance compared to VideoChat2. VideoChat-TPO's superior performance on multi-image understanding compared to LLaVA-Interleave and InternLM-XComposer demonstrates that TPO's vision enhancements improve the MLLM's spatial perception and image understanding." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.307, + 0.715, + 0.323 + ], + "angle": 0, + "content": "4.2. Vision Task Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.895, + 0.513 + ], + "angle": 0, + "content": "Moment Retrieval. Moment Retrieval is to locate the target segments in a video based on queries. Table 5 and 6 compare the zero-shot and fine-tuned moment retrieval performance of VideoChat-TPO against other expert models and MLLMs. In zero-shot settings, VideoChat-TPO achieves a \\( R@1 \\) (\\( IoU = 0.5 \\)) score of 40.2 on the Charades-STA [21]. It surpasses the previous state-of-the-art (SOTA) MLLM specialized for temporal tasks, ChatVTG [59], and the expert model, UniVTG [47]. This demonstrates VideoChat-TPO's ability to accurately locate video moments corresponding to given text queries, thereby enhancing the practical applicability of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.895, + 0.71 + ], + "angle": 0, + "content": "Highlight Detection. Highlight detection generates salient scores for emphasizing frames based on the input language query. We compare the fine-tuning highlight detection performance of VideoChat-TPO with that of other expert models and MLLMs, as shown in Table 6. VideoChat-TPO notably outperforms the previous state-of-the-art MLLM (such as TimeChat [63]) on both Charades-STA and QVHighlight. For expert models like QD-DETR and UniVTG, VideoChat-TPO beats them on Charades-STA non-trivially and achieves a comparable performance on QVHighlight. This also demonstrates the progress in extending MLLMs to broad temporal tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.897, + 0.906 + ], + "angle": 0, + "content": "Spatial Grounding. To verify the fine-grained localization ability of the model, we run the spatial grounding task which inputs textual descriptions into the model and gets the corresponding bounding box on RefCOCO [91]. As shown in Table 7, we compare VideoChat-TPO with pixel-to-sequence models, i.e., VisionLLM-H [75], pixel-to-embedding methods, i.e., NExT-Chat [97], and expert models, i.e., G-DINO [50]. VideoChat-TPO performs better than the pixel-to-embedding methods by using only a simple task head and achieves comparable performance to pixel-to-sequence models fine-tuned on a large amount of spatial grounding data, as well as to specialized models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29885" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.565, + 0.187 + ], + "angle": 0, + "content": "
ModelAcc@IoPAcc@GQAmIoPIoP@0.3IoP@0.5mIoUIoU@0.3IoU@0.5
VIOLETv2 [20]54.912.823.625.123.33.14.31.3
SeViLA [93]72.516.629.534.722.921.729.213.8
LangRepo [31]59.617.131.3-28.718.5-12.2
FrozenBiLM NG+ [88]73.817.524.228.523.79.613.56.1
VideoStreaming [58]57.417.832.2-31.019.3-13.3
LLoVi [98]65.924.337.3-36.920.0-15.3
HawkEye [79]-----25.737.019.5
VideoChat-TPO77.725.535.647.532.827.741.223.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.197, + 0.189, + 0.45, + 0.203 + ], + "angle": 0, + "content": "Table 3. Performance on Grounded QA." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.089, + 0.891, + 0.187 + ], + "angle": 0, + "content": "
ModelMMIU [54]SEED2I [38]SEED2M [38]
LLaVA-v1.5 [48]19.258.339.2
ShareGPT4V [8]18.5--
OpenFlamingo [2]22.336.643.5
LLaVA-Interleave [40]32.4--
VideoChat2 [45]35.026.527.6
VideoChatGPT [52]-38.349.8
InternLM-XComposer [16]21.965.449.8
VideoChat-TPO40.2 (+5.2)67.3 (+40.8)70.0 (+42.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.577, + 0.189, + 0.877, + 0.203 + ], + "angle": 0, + "content": "Table 4. Performance on Image Understanding." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.217, + 0.331, + 0.324 + ], + "angle": 0, + "content": "
ModelCharades-STA [21]
R@0.3R@0.5R@0.7mIoU
UniVTG [47]44.125.210.027.1
VideoChat2 [45]38.014.33.824.6
VTimeLLM [26]51.027.511.431.2
TimeChat [63]-32.213.4-
HawkEye [79]50.631.414.533.7
ChatVTG [59]52.733.015.934.9
VideoChat-TPO58.340.218.438.1
" + }, + { + "type": "table", + "bbox": [ + 0.332, + 0.218, + 0.685, + 0.324 + ], + "angle": 0, + "content": "
ModelCharades-STA [21]QVHighlight [36]
R@0.3R@0.5R@0.7mIoUmAPHIT@1
M-DETR [36]65.852.130.645.535.755.6
QD-DETR [56]-57.332.6-38.962.4
UniVTG [47]72.660.238.652.240.566.3
TimeChat [63]-46.723.7-21.737.9
HawkEye [79]72.558.328.8---
VideoChat-TPO77.065.040.755.038.665.4
" + }, + { + "type": "table", + "bbox": [ + 0.689, + 0.217, + 0.887, + 0.324 + ], + "angle": 0, + "content": "
MethodsRefCOCO [91]
valtestAtestB
MAntNet ★ [92]76.480.469.3
OFA-L [73]80.083.776.4
G-DINO-L ★ [50]90.693.288.2
VisionLLM-H [75]-86.7-
Shikra-7B [7]87.090.680.2
NExT-Chat-7B [97]85.590.077.9
VideoChat-TPO85.990.881.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.326, + 0.68, + 0.354 + ], + "angle": 0, + "content": "Table 5. Zero-Shot Performance on Table 6. Fine-tuning Performance on Moment Re-Moment Retrieval. Gray means no LLM. Retrieval and Highlight Detection. Gray means no LLM." + }, + { + "type": "table_caption", + "bbox": [ + 0.687, + 0.327, + 0.886, + 0.354 + ], + "angle": 0, + "content": "Table 7. Spatial Grounding Task. \\(\\star\\) with a refined decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.473, + 0.579 + ], + "angle": 0, + "content": "Tracking. In the tracking task, the model receives the object coordinates from the first frame and outputs the coordinates for the remaining frames in the video. We evaluate the VideoChat-TPO on the mainstream tracking benchmarks LAOST [18] and GOT-10k [27], as shown in Table 8. In zero-shot testing, VideoChat-TPO ranks best among all MLLMs, even surpassing some fine-tuned expert models notably with around \\(10\\%\\) increase on success rate, such as SiamFC [4] and ATOM [13]. With the TPO training method, the model is optimized from mask sequences that hard to represent by words, allowing the MLLM to achieve strong motion characterization for multimodal generalization in highly dynamic and occluded scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.47, + 0.83 + ], + "angle": 0, + "content": "Referring Segmentation. The referring segmentation task requires the model to output the mask sequence corresponding to the specified prompt. This capability is newly activated in VideoChat-TPO, as it is not supported or designed in the original VideoChat2 or SAM2. We compare the referring segmentation performance of VideoChat-TPO with other expert models and MLLMs in Table 9. The zero-shot capability of VideoChat-TPO is comparable to the fine-tuning capabilities of other expert models, i.e. ReferFormer [81], showing its notable open-world video segmentation performance. Through TPO training, MLLM effectively optimizes its understanding in both tracking and segmenting objects indicated by users. It further enables the model to excel in pixel-level tasks, offering perceptual advantages to it in practical applications like robotic control." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.242, + 0.859 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "In this section, we analyze the effectiveness of the key components of TPO and evaluate its scalability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Extending to Other MLLMs. To evaluate the effectiveness of the TPO method on different MLLMs, we apply TPO to LLaVA (LLaVA-OneVision [39]) in addition to VideoChat. In LLaVA, fine-grained video understanding is a weak spot. Despite the conversation data used in LLaVA-OneVision being very similar to, or even more extensive than, what we employ, there remains significant room for improvement in its performance on short-term fine-grained understanding tasks. As shown in Table 10, with TPO, LLaVA not only gains capabilities that it originally lacked but also achieves an \\(8.1\\%\\) improvement on MVBench, demonstrating the universality of the TPO method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Task Preference Model vs. Textualized Task Data. We compare the TPO method with the approach of textualizing task data and training it in an autoregressive manner, which is a straightforward and efficient method for MLLMs to learn specific tasks. We use the same data as the TPO method and convert the data related to the task head into conversation format. Due to the limitations of textual representation, much data must be input in a more discrete manner. For instance, in tracking supervision data, we follow Merlin [90] by converting masks into a sequence of spatial coordinates to serve as both the model's input and output. As shown in Table 11, MLLM-TPO achieves a \\(2.1\\%\\) performance improvement on MVBench [45] compared to MLLMs trained with textualized task data. This demonstrates that the task head in the TPO method enables more effective utilization of the original supervision signals compared to the next-token prediction approach, which inevitably incurs information loss when converting fine-grained tasks into textual annotations. The next-token prediction method cannot truly capture the nuanced information present in videos, while TPO allows the MLLM to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29886" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.387, + 0.19 + ], + "angle": 0, + "content": "
ModelLaSOT [18]GOT-10k [27]
SuccessPnormPOverlapSR0.5SR0.75
SiamFC [4]33.642.033.934.835.39.8
ATOM [13]51.5--55.663.440.2
SiamRPN++ [37]49.656.949.151.861.832.5
SiamFC++ [86]54.462.354.759.569.547.9
LLaVA-1.5 [48]19.416.512.823.520.29.7
Merlin [90]39.840.238.151.455.942.8
VideoChat-TPO69.480.176.970.679.866.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.193, + 0.386, + 0.206 + ], + "angle": 0, + "content": "Table 8. Performance on Tracking Benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.396, + 0.09, + 0.696, + 0.173 + ], + "angle": 0, + "content": "
MethodRef-YouTube-VOS [65]MeViS [15]
J&FJFJ&FJF
ReferFormer [81]62.961.364.631.029.832.2
OnlineRefer [80]62.961.064.7---
LISA [35]52.652.153.0---
VideoLISA [3]63.761.763.744.441.347.6
VideoChat-TPO63.952.375.447.042.651.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.39, + 0.179, + 0.7, + 0.206 + ], + "angle": 0, + "content": "Table 9. Performance on Referring Segmentation. TPO is evaluated in a zero-shot on Ref-YouTube." + }, + { + "type": "table", + "bbox": [ + 0.708, + 0.09, + 0.888, + 0.138 + ], + "angle": 0, + "content": "
ModelMVBench
LLaVA-OV [39]56.7
LLaVA-OV-TPO64.8 (+8.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.701, + 0.151, + 0.892, + 0.207 + ], + "angle": 0, + "content": "Table 10. Applying TPO on LLaVA-OV. TPO significantly enhances the ability to understand fine-grained video details." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.232, + 0.466, + 0.312 + ], + "angle": 0, + "content": "
ModelCharades-STA [21]MVBench
R@0.3R@0.5R@0.7mIoUAVG
VideoChat-TPO58.340.218.438.166.8
w/o reasoning data56.438.317.135.666.1
replace by simple head31.517.86.115.465.8
textualized task data33.718.66.216.064.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.315, + 0.468, + 0.328 + ], + "angle": 0, + "content": "Table 11. Ablation of Reasoning Data and Head Performance." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.343, + 0.435, + 0.447 + ], + "angle": 0, + "content": "
Exp.TRMCR@0.5Acc@0.5J&F
1✓30.2--
2✓-77.3-
3✓--55.1
4✓✓-80.258.3
5✓✓✓36.781.661.4
6✓✓✓✓40.282.063.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.45, + 0.47, + 0.517 + ], + "angle": 0, + "content": "Table 12. Impact of TPO Components and Data. T, R, M, and C denote temporal head, region head, mask head, and conversation data respectively. R1@0.5 means R1@0.5 in Charades-STA, Acc@0.5 represents the mean of Acc@0.5 in all COCO datasets, \\(\\mathcal{J}\\&\\mathcal{F}\\) means \\(\\mathcal{J}\\&\\mathcal{F}\\) in Ref-YouTube-VOS." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.536, + 0.452, + 0.551 + ], + "angle": 0, + "content": "learn perceptual information beyond conversational data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.469, + 0.756 + ], + "angle": 0, + "content": "Impact of Task Head Performance on TPO. A stronger task head is expected to enhance the multimodal capabilities of TPO more effectively than a weaker one. To further explore the impact of the task head on model performance, we replaced the temporal head from CG-DETR [55] with a simple two-layer MLP. As shown in Table 11, the simple temporal head exhibits a significant performance decline compared to CG-DETR on the corresponding tasks. However, it still achieves a \\(1.1\\%\\) improvement compared to the model without TPO. In contrast to the simpler head, the well-designed head can provide more accurate expert knowledge through its architecture and pre-trained task preference weights, enabling better utilization of the data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Impact of Data Scaling. Concerning foundation or large models, we expect to enhance their model performance by incorporating additional data at scale. In our approach, we expand the existing conversation data by integrating two reasoning datasets [5, 14] with temporal information. By incorporating these datasets, we observe a notable enhancement in the capabilities of the temporal head in Table 11. Furthermore, as shown in Table 12, adding conversation data further improves performance across various" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.294 + ], + "angle": 0, + "content": "tasks. Overall, increasing the amount of conversation and fine-grained task data leads to noticeable improvements in model performance, demonstrating the effectiveness of data scaling in training multimodal models with TPO." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Synergistic Gains from Co-training. Table 12 presents the impact of incorporating different visual tasks into TPO on performance. This indicates that the collaborative learning of visual tasks facilitates the transfer of knowledge, ultimately leading to better performance across the board. Overall, these results confirm that co-training not only benefits individual tasks but also creates a synergistic effect that enhances overall capabilities in visual understanding." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.627, + 0.459 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.723 + ], + "angle": 0, + "content": "This paper introduces Task Preference Optimization (TPO). It enhances the overall multimodal performance of MLLMs by enhancing their precise visual understanding. TPO achieves this by integrating differentiable task preferences derived from fine-grained visual tasks. It introduces learnable task tokens that serve as bridges between multiple task-specific heads and the MLLM. Through the joint optimization of these task tokens, heads, and the MLLM, TPO leads to a substantial improvement in multimodal dialogue capabilities and performance across various visual tasks. Our results demonstrate the effectiveness of TPO in scaling MLLMs with task-specific data and seamlessly integrating them with existing expert vision models. We believe this study clarifies the prerequisites for fusing MLLMs with models and data from the pre-large model era, effectively bridging the gap between expert and generalist models, as well as between generation and understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limitations. Currently, TPO focuses exclusively on discriminative visual tasks, overlooking generative ones. Meanwhile, the framework is supervised by human annotations, which neglects potentially valuable unsupervised or self-supervised learning approaches such as contrastive learning [10, 23]. This limitation inherently restricts the scalability of TPO in terms of both task diversity and requirements. While we demonstrate TPO's potential for enhancing MLLMs through increased task-specific data, a comprehensive investigation into the broader impact of this scaling remains a crucial area for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29887" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.114, + 0.47, + 0.186 + ], + "angle": 0, + "content": "This work is partially supported by the National Key R&D Program of China (No. 2022ZD0160102), Jiangsu Frontier Technology R&D Program (No. BF2024076), and the Science and Technology Commission of Shanghai Municipality under Grant No. 23YF1461900." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.212, + 0.175, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.238, + 0.47, + 0.307 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.309, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortman, and Ludwig Schmidt. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.422, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[3] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.492, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[4] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In ECCV, pages 850–865, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.549, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[5] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.606, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[6] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv preprint arXiv:2304.02948, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.676, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.733, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[8] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In ECCV, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.789, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[9] Lin Chen, Xin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, Li Yuan, Yu Qiao, Dahua Lin, Feng Zhao, and Jiaqi Wang. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learn" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.238, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "ing of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.123, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[12] Yiming Cui, Ziqing Yang, and Xin Yao. Efficient and effective text encoding for chinese llama and alpaca. arXiv preprint arXiv:2304.08177, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.236, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[13] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In CVPR, pages 4660-4669, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.28, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[14] Pradipto Das, Chenliang Xu, Richard F Doell, and Jason J Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In CVPR, pages 2634-2641, 2013. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.336, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[15] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.393, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[16] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.477, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[17] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.535, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[18] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, pages 5374-5383, 2019. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.605, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[19] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[20] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. An empirical study of end-to-end video-language transformers with masked visual modeling. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.747, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[21] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.79, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18995-19012, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 8" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29888" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[24] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[25] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.192, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[26] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.235, + 0.471, + 0.277 + ], + "angle": 0, + "content": "[27] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 43(5):1562-1577, 2019. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.471, + 0.39 + ], + "angle": 0, + "content": "[29] Fan Jiang and Zilei Wang. Sparse sharing relation network for panoptic driving perception. In ACMMM, pages 800-808, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.392, + 0.471, + 0.447 + ], + "angle": 0, + "content": "[30] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In CVPR, pages 13700-13710, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.449, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[31] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv preprint arXiv:2403.14622, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.505, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[32] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.562, + 0.471, + 0.617 + ], + "angle": 0, + "content": "[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.619, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[34] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.689, + 0.471, + 0.744 + ], + "angle": 0, + "content": "[35] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, pages 9579-9589, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.746, + 0.471, + 0.787 + ], + "angle": 0, + "content": "[36] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. NeurIPS, 34:11846-11858, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.789, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[37] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. SiamRPN++: Evolution of siamese visual tracking with very deep networks. In CVPR, pages 4282-4291, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.845, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[38] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In CVPR, pages 13299-13308, 2024. 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[39] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[41] Hao Li, Jinguo Zhu, Xiaohu Jiang, Xizhou Zhu, Hongsheng Li, Chun Yuan, Xiaohua Wang, Yu Qiao, Xiaogang Wang, Wenhai Wang, et al. Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In CVPR, pages 2691-2700, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[42] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[43] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[44] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.447, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[45] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.518, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[46] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.575, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[47] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.646, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.676, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[49] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In ECCV, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.718, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[50] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In ECCV, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[51] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29889" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[53] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.15, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[54] Fanqing Meng, Jin Wang, Chuanhao Li, Quanfeng Lu, Hao Tian, Jiaqi Liao, Xizhou Zhu, Jifeng Dai, Yu Qiao, Ping Luo, et al. Mmiu: Multimodal multi-image understanding for evaluating large vision-language models. arXiv preprint arXiv:2408.02718, 2024. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[55] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[56] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[57] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP, pages 2265-2269. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.391, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[58] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuangrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.448, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[59] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPR, pages 1847-1856, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[60] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.562, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[61] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.66, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[62] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics, 1:25-36, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.717, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[63] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. CVPR, abs/2312.02051, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.774, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[64] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.817, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[65] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223. Springer, 2020. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[66] Jing Shao, Siyu Chen, Yangguang Li, Kun Wang, Zhenfei Yin, Yinan He, Jianing Teng, Qinghong Sun, Mengya Gao," + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Jihao Liu, et al. Intern: A new learning paradigm towards general vision. arXiv preprint arXiv:2111.08687, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.135, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[67] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.205, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[68] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.275, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[69] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, pages 10078-10093, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[70] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.401, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[71] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, pages 20-36. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.457, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[72] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[73] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, pages 23318-23340. PMLR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[74] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, pages 6312-6322, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[75] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[76] Weiyun Wang, Yiming Ren, Haowen Luo, Tiantong Li, Chenxiang Yan, Zhe Chen, Wenhai Wang, Qingyun Li, Lewei Lu, Xizhou Zhu, et al. The all-seeing project v2: Towards general relation comprehension of the open world. arXiv preprint arXiv:2402.19474, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.79, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[77] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In ICCV, pages 1130-1140, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[78] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. In ECCV, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29890" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[79] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.15, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[80] Dongming Wu, Tiancai Wang, Yang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In ICCV, pages 2761-2770, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.207, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[81] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.25, + 0.468, + 0.331 + ], + "angle": 0, + "content": "[82] Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Wenhai Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. arXiv preprint arXiv:2406.08394, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.468, + 0.375 + ], + "angle": 0, + "content": "[83] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In CVPR, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.377, + 0.468, + 0.446 + ], + "angle": 0, + "content": "[84] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pretraining for zero-shot video-text understanding. In EMNLP, pages 6787-6800, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.448, + 0.468, + 0.503 + ], + "angle": 0, + "content": "[85] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.505, + 0.468, + 0.546 + ], + "angle": 0, + "content": "[86] Yinda Xu et al. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, pages 140-148, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.548, + 0.468, + 0.602 + ], + "angle": 0, + "content": "[87] Shen Yan, Tao Zhu, Zirui Wang, Yuan Cao, Mi Zhang, Soham Ghosh, Yonghui Wu, and Jiahui Yu. Video-text modeling with zero-shot transfer from contrastive captioners. ArXiv, abs/2212.04979, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.468, + 0.646 + ], + "angle": 0, + "content": "[88] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. In NeurIPS, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.648, + 0.468, + 0.73 + ], + "angle": 0, + "content": "[89] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.733, + 0.468, + 0.801 + ], + "angle": 0, + "content": "[90] En Yu, Liang Zhao, Yana Wei, Jinrong Yang, Dongming Wu, Lingyu Kong, Haoran Wei, Tiancai Wang, Zheng Ge, Xiangyu Zhang, et al. Merlin: Empowering multimodal llms with foresight minds. In ECCV, pages 425-443. Springer, 2025. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.803, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[91] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[92] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In CVPR, pages 1307-1315, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[93] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In NeurIPS, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[94] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In CVPR, pages 13807-13816, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[95] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In CVPR, pages 23056-23065, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[96] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In ICLR. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[97] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An Imm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[98] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lllm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.429, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[99] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, pages 543–553, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[100] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[101] Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[102] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[103] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[104] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[105] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. In ICLR Workshop on Reliable and Responsible Foundation Models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[106] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29891" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "conditional moes. In NeurIPS, pages 2664-2678, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[107] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29892" + } + ] +] \ No newline at end of file diff --git a/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_origin.pdf b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..310c5a01c602558da883ac5253ff8a4b3e3a052d --- /dev/null +++ b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/775c99dd-79ec-4cb9-9585-2a300953790b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44d43f83e7bb3d0da818b85526f417f94d578369e88763bafe0464885ee1f714 +size 1579968 diff --git a/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/full.md b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..467ef4a382a6c1f2fd3ad576384322321caeaf6c --- /dev/null +++ b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/full.md @@ -0,0 +1,356 @@ +# Task Preference Optimization: Improving Multimodal Large Language Models with Vision Task Alignment + +Ziang Yan\*2,1, Zhilin Li\*3,1, Yinan He\*1 + +Chenting Wang $^{4,1}$ , Kunchang Li $^{5,1}$ , Xinhao Li $^{6,1}$ , Xiangyu Zeng $^{6,1}$ + +Zilei Wang $^{3}$ , Yali Wang $^{5,1}$ , Yu Qiao $^{1}$ , Limin Wang $^{6,1}$ , Yi Wang $^{\dagger,7}$ + +$^{1}$ Shanghai AI Laboratory $^{2}$ Zhejiang University $^{3}$ University of Science and Technology of China $^{4}$ Shanghai Jiao Tong University + +Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences Nanjing University Shanghai Innovation Institute + +https://github.com/OpenGVLab/TPO + +# Abstract + +Current multimodal large language models (MLLMs) struggle with fine-grained or precise understanding of visuals although they give comprehensive perception and reasoning in a spectrum of vision applications. Recent studies either develop tool-using or unify specific visual tasks into the autoregressive framework, often at the expense of overall multimodal performance. To address this issue and enhance MLLMs with visual tasks in a scalable fashion, we propose Task Preference Optimization (TPO), a novel method that utilizes differentiable task preferences derived from typical fine-grained visual tasks. TPO introduces learnable task tokens that establish connections between multiple task-specific heads and the MLLM. By leveraging rich visual labels during training, TPO significantly enhances the MLLM's multimodal capabilities and task-specific performance. Through multi-task co-training within TPO, we observe synergistic benefits that elevate individual task performance beyond what is achievable through single-task training methodologies. Our instantiation of this approach with VideoChat and LLaVA demonstrates an overall $14.6\%$ improvement in multimodal performance compared to baseline models. Additionally, MLLM-TPO demonstrates robust zero-shot capabilities across various tasks, performing comparably to state-of-the-art supervised models. + +# 1. Introduction + +Multimodal large language models (MLLMs) show impressive visual perception and reasoning capabilities, with applications in personal assistants [68], embodied systems [17], and scientific discovery [6]. Considering the + +![](images/9e5025be807bb324267d565245547b627928c16cd143e0e1fed961d01a1e5b01.jpg) + +Show the tracking trajectory of cup with candy in middle. Which cup is the candy under, left one, middle one or right one? + +![](images/3c3801e4a772805f647d6a89a8afce4327ab7ebb890e61751a2d7808bcb98f56.jpg) + +The cups are shuffled in the video, and the candy is under the middle cup. + +![](images/557ea63bd3e484136e7dfeda25df6c6a64e6adb3d465243e4158f0d6592931c1.jpg) + +![](images/cff516a6b12e6815e2d186d4a735a81fd8693a1de6e3fd5c49c7acd465e582ef.jpg) + +![](images/5d11b7e46cf72f07c856552c1f6f0c2bb9392b4f45278ffa2e1ef7237042e5b8.jpg) + +The candy is under the candy is under the the left cup. first cup on the left. + +![](images/9ad6c3de098d56f274573d4269296257f9180c82982ae5ef2019dd74b6290f7a.jpg) +Figure 1. TPO uses differentiable task preferences from dense visual supervisions via task-specific heads to enhance MLLMs in fine-grained understanding. + +growing expectations of users in more accurate and detailed perception, taking the shell game as an example, further improving the generality of MLLM requires fine-grained knowledge representation beyond words, e.g. MLLM implicitly embeds how to track keys indicated by users. + +Existing studies address the enhancement of perceptual granularity in MLLMs by focusing on specific visual tasks (e.g. temporal grounding, segmentation, tracking) via MLLMs. They usually fine-tune MLLMs on more task data in text format or enable MLLMs to activate the corresponding task heads. Shikra [7] applies MLLM to localization tasks, transforms object coordinates into dialogue formats, and learns them autoregressively. Meanwhile, TimeChat [63] and VideoChat-T [96] treat event timestamps as text for autoregressive prediction, thereby endowing MLLMs with temporal grounding. They do improve specific task performance significantly, while at the cost of multimodal performance more or less. This is counterintuitive as seminal research proves different visual tasks are correlated and training them together often yields mutual gains [29, 47, 66]. We conjecture that the presentation of different tasks influences this training and that the conflict + +arises from the learning discrepancies between discrete textual tokens and visually dense predictions. Our experiments in Section 4 validate that a decoupled representation design can effectively address this issue. + +To enhance the multimodal capabilities of MLLMs, we explore optimization methods to meet multiple visual task requirements in an end-to-end manner. We propose task preference optimization (TPO), incorporating visual task knowledge into the MLLMs by jointly maximizing the likelihoods of visual task estimations and multimodal dialogue. Inspired by direct preference optimization [60] (DPO) and related methods, we treat visual task annotations as human preferences in particular demands, as shown in Figure 2. DPO aids LLM (or MLLM) in aligning with human preferences through a binary classification that directs the model to generate responses that people prefer. Similarly, TPO enhances MLLMs' visual sensing capabilities through differentiable task optimizations that guide MLLMs to yield dense predictions closely resembling human perception. To achieve this, TPO concretizes MLLM-specific visual perceptions into corresponding task tokens, disentangled from MLLM representation. Then TPO fine-tunes these task tokens and updates the MLLM accordingly. + +Specifically, TPO appends visual task heads to the partial output of the MLLM, using several learnable task tokens as inputs for these corresponding heads. During training, TPO enables the MLLM to first distinguish and activate the appropriate task tokens based on user instructions. Subsequently, TPO jointly trains the task tokens and their corresponding heads to enhance the MLLM's understanding of visual tasks. Finally, TPO trains the entire model—including task tokens and heads—on both multimodal and visual-specific task data, promoting the perception and reasoning capabilities of the MLLM. Additionally, we note that multi-task co-training yields greater improvements than single-task training. + +TPO demonstrates scalability across various MLLM approaches, encompassing a wide range of visual task categories and data quantities. We validate the effectiveness of TPO within widely used MLLMs, such as LLaVA [48] and VideoChat [43], as detailed in Section 4.3. By fine-tuning these open-source MLLMs with TPO, we enhance their visual understanding capabilities and improve dialogue performance. Additionally, we explore several key spatiotemporal and spatial perception tasks, including spatial grounding, tracking, and temporal grounding. Our findings indicate that these tasks can mutually enhance each other's performance, particularly contributing to the improvement of multimodal dialogue capabilities. + +Our contributions can be summarized as: + +- We propose a new training method for multimodal large language models, referred to as Task Preference Optimization (TPO). This method leverages supervised in + +![](images/e3b5d96dc60e98838535b1935eaa8b58307386009df569516318a6558db9d88b.jpg) +Figure 2. Comparison of Learning Method. A solid line indicates data flow, and a dotted line represents feedback. $\star$ and $\text{心}$ denote modules that are frozen and unfrozen. + +formation from visual task-specific data to optimize the MLLM through the corresponding heads, resulting in significantly enhanced multimodal perception and reasoning performance. Specifically, TPO achieves an average improvement of $14.6\%$ across multiple image and video multimodal benchmarks [19, 38, 45, 54, 104]. + +- TPO effectively equips MLLM with the capability to address several key visual perception tasks through the introduced task heads. MLLM-TPO achieves comparable performance in spatial grounding, moment retrieval, highlight detection, referring segmentation, and tracking comparable to expert models across various benchmarks. +- TPO demonstrates scalability across various employed MLLMs, task heads, and scales of task data. We validate the effectiveness of TPO in multiple mainstream MLLMs, such as VideoChat2 [45] and LLaVA [39, 48]. Notably, multi-task joint training based on TPO enhances both multimodal performance and individual visual task, with improvements becoming increasingly significant as additional appropriate heads are introduced. Furthermore, the performance of the MLLM and task heads improves when scaling task data. + +# 2. Related Work + +Vision Foundation Model. Vision foundation models [44, 66, 69, 71, 72, 74, 78, 84, 87, 103] are designed to be adaptable for various downstream tasks through extensive pre-training on large-scale and diverse datasets. VideoPrism [103] achieves the leading results in various video tasks by combining video-text contrastive learning and spatiotemporal token reconstruction using a dataset of public and proprietary videos. InternVideo2 [78] utilizes masked reconstruction, cross-modal contrastive learning, and next-token prediction to enhance the model's perceptiveness, semantic understanding, and reasoning capabilities. + +Based on vision foundation models, some studies [41, 51, 77, 106, 107] intend to incorporate downstream task heads into this framework and expect end-to-end task training. Unified-IO [51] introduces a versatile model capable of handling a wide range of tasks across vision and lan + +![](images/5f3fae02b4b65871ccf8c68d6b7127b7fd4b1a99e935fd534ef46c873e6c026c.jpg) +Figure 3. Overall Pipeline of TPO. The architecture of Task Preference Optimization (TPO) consists of four main components: (1) a vision encoder, (2) a connector, (3) a large language model, and (4) a series of visual task heads. Differently colored flame symbols indicate which components are unfrozen at various stages of the training process. + +guage domains, requiring minimal or no task-specific modifications or additional parameters. Uni-Perceivers [41, 106, 107] formulate different tasks to find the maximum likelihood target for each input through the representation similarity regardless of their modality. Nevertheless, these generalist models are limited to predefined tasks and cannot support flexible open-ended task customization based on language instructions. + +Multimodal Large Language Model. The effective understanding and reasoning of LMMs has attracted the attention of numerous researchers. Limited to its input modality, researchers expand the visual capabilities of LLM, leading to MLLMs. Seminal works, such as BLIP-2 [42], LLaVA [48], and mPLUG-Owl [89], introduce image captioning and visual question answering based on LLM using visual instruction-tuning data. Some video-based MLLMs have been proposed, such as VideoChat [43], VideoChatGPT [52], and Video-LLaMA [99], enabling LLM to gain video understanding capabilities by encoding multiple video frames and using video instruction data. + +Vanilla MLLMs have achieved impressive results in visual-related dialogues, but barely address some fine-grained tasks, such as segmentation and temporal grounding. To address this challenge, MLLMs usually take one of the following pipelines: pixel-to-sequence (P2S) and pixel-to-embedding (P2E). For P2S methods [7, 63, 70, 79, 90], MLLM directly outputs the textual predictions. TimeChat [63] introduces a time-aware frame encoder and a sliding video Q-former to enhance temporal perception. For P2E methods [3, 35, 75, 82, 97], MLLM compresses and inputs visuals into the downstream decoder for estima + +tions. LISA [35], NExT-Chat [97], and VideoLISA [3] introduce SAM [33] as a segmentation tool, using a special token as a prompt to connect MLLM and SAM. Vision-LLM v2 [82] designs routing tokens and super-link queries to bridge MLLM with multiple decoders. + +Alignment in MLLM. Aligning MLLM with human preferences or values is crucial for MLLM's development. Recent works [67, 94, 105] introduce alignment approaches to MLLM, including proximal policy optimization (PPO) [64] and direct preference optimization (DPO) [60], as shown in Figure 2. They usually exploit proprietary models like GPT4-V to build visual preference datasets and then tune MLLMs using PPO or DPO. Llava-RLHF [67] incorporates PPO into a MLLM framework Llava, argued by image captions or question-answers (QA), while Zhou et al. [105] and Zhang et al. [101] give DPO implementations for MLLMs where visual preference data are created by GPT-4V and other open-sourced MLLMs. RLHF-V [94] collects a dense human preference in segmentation and enhances MLLMs with DPO. + +# 3. Method + +Task Preference Optimization (TPO) aims to enable MLLMs to master classical visual perceptions (such as tracking, temporal grounding, etc.) for better task generalization, regarding that many multimodal reasoning tasks require precise visual cues for accurate and reliable responses. As shown in Figure 3, MLLM-TPO has a typical multimodal model $M$ (consisting of a vision encoder $E$ , a vision-language connector $C$ , and a large language model $G$ ) and a task preference model $P$ with a series of visual + +task heads $\{H_{i}\}_{i = 1\dots n}$ ( $n$ denotes the task head number). These heads connect with MLLM using the embeddings $\{\mathbf{e}_i\}_{i = 1\dots n}$ ( $\mathbf{e}_i = G(\mathbf{v}_i)\in \mathbb{R}^{1\times C}$ ) transformed from the learnable task-specific tokens $\{\mathbf{v}_i\}_{i = 1\dots n}$ ( $\mathbf{v}_i\in \mathbb{R}^{1\times C}$ ) via MLLM. + +TPO employs a local-to-global training scheme, first adapting task heads to the MLLM and then training them jointly. Specifically, the MLLM $M$ starts to recognize $\{\mathbf{v}_i\}_{i=1\dots n}$ by updating $G$ according to user instructions, then we tune $\{\mathbf{v}_i\}_{i=1\dots n}$ and $\{H_i\}_{i=1\dots n}$ for adapting visual heads to $M$ . Finally, we train both $M$ and $\{H_i\}_{i=1\dots n}$ together. During inference, MLLM-TPO can respond to users' queries in text, and produce structured visual outputs (like masks, timestamps, trajectories, etc.) when users ask (e.g. yielding time results for "find when the birthday party starts"). We detail how to form MLLM-TPO structurally and train it in the following. + +# 3.1. Task Preference Models + +The task preference model (TPM) $P$ contains a series of task tokens $\{\mathbf{v}_i\}_{i = 1\dots n}$ and heads $\{H_i\}_{i = 1\dots n}$ . Before TPM works, the attached MLLM generates special tokens indicating task types from the input queries. Then TPM dynamically calls the task token $\mathbf{v}_j$ based on the special token, transforms it to the task embedding $\mathbf{e}_j$ via the LLM $G$ (the last hidden embedding), and feeds $\mathbf{e}_j$ to the corresponding visual task head $\{H_j\}$ for specific task predictions. Considering existing MLLMs demonstrate remarkable capabilities in common object and scene recognition, yet struggle to accurately locate things or actions, the employed task heads mainly focus on spatiotemporal localization and tracking. Specifically, we give three fundamental task types for compensating mainstream MLLMs' gaps from expert models in visual perceptions: 1) region head, 2) temporal head, and 3) mask head. Their architectures are given below. + +- Region Head. A two-layer multilayer perceptron (MLP) with ReLU activations is employed for the region head. It takes embeddings from LLM and regresses them to the bounding box coordinates for spatial grounding. +- Temporal Head. This head is composed of a video encoder, text encoder, and temporal grounding model for moment retrieval and highlight detection. The video and text queries are input into the video and text encoders to obtain their features respectively. Then we append the temporal task embedding after the text features. With the new text and video features, the MLLM estimates the start/end time and highlight score of the corresponding moment from the query via the temporal grounding model. +- Mask Head. Pixel-level tasks pose significant challenges for MLLMs due to the lack of the corresponding learning data and optimization in MLLMs. To this end, we introduce a specialized mask head, utilizing the image en + +coder, mask decoder, and memory bank components from SAM2 [61], replacing the prompt encoder with a single MLP layer called the mask adapter. Specifically, for the given visuals and user queries, the features extracted from the aforementioned image encoder and the text representation corresponding to the mask task embeddings from LLM are employed, fed into the mask decoder to produce the desired mask. + +Most known discriminative vision tasks can be addressed by one or a combination of these three task heads. TPM builds the architectural foundations for leveraging existing vision data with annotations to enhance MLLMs. + +# 3.2. Task Preference Optimization + +TPO improves MLLMs with extra supervisions from visual task heads by back-propagating gradients from heads to update MLLMs using visual task data. It enables the language model $G$ in MLLM to discriminate specific task types when users demand (task assignment). Then, TPO trains TPM $P$ via compact task representations $\mathbf{e}_i$ (task optimization) from $\mathbf{v}_i$ . Lastly, we train $M$ and $P$ together, tuning $M$ for the refined spatiotemporal perception according to task preferences from $P$ . Its optimization objective is given as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {m l l m}} + \underbrace {\mathcal {L} _ {\text {a s s i g n}} \left(G \left(\mathbf {T} _ {\mathrm {q}}\right) , \mathbf {s}\right)} _ {\text {T a k A s s i g n m e n t}} + \sum_ {i = 1} ^ {n} \underbrace {\mathcal {L} _ {\text {t a s k}} \left(\mathbf {A} _ {i} , H _ {i} \left(G (\mathbf {v} _ {i})\right)\right)} _ {\text {T a k O p t i m i z a t i o n}}, \tag {1} +$$ + +where $\mathbf{T}_{\mathbf{q}}$ and $\mathbf{s}$ denote a user query (some contain specific task needs like tracking) in the form of text token sequence and the ground truth task indicating token, respectively. $\mathbf{A}_i$ is the task annotation of the input visual $\mathbf{X}$ , so it could be a number tuple for describing regional localization and area or mask for delineating object shape and position. Here $(\mathbf{X},\mathbf{T}_{\mathbf{q}},\mathbf{T}_{\mathbf{a}},\mathbf{A}_i)$ stands for an input data tuple to MLLM-TPO. Usually, we have $(\mathbf{X},\mathbf{T}_{\mathbf{q}},\mathbf{T}_{\mathbf{a}})$ for typical MLLM training while $(\mathbf{X},\mathbf{T}_{\mathbf{q}},\mathbf{A}_i)$ for task token and head training. $\mathcal{L}_{\mathrm{assign}}$ is the cross-entropy loss. Meanwhile $\mathcal{L}_{\mathrm{task}_i}$ varies according to the task, and is usually regression- or classification-related loss. + +To train MLLM-TPO, we propose a 3-stage local-to-global training scheme, given in Figure 3. Stage 1 learns to identify the task type based on user queries. In stage 2, we train task heads along with their corresponding task tokens, respectively. Lastly, we co-train task heads with MLLM by both task data and multimodal conversation data. Our 3-stage training strategy mitigates the risk of degrading the MLLM's general abilities. We describe them as: + +Stage 1: Task Assignment. We design a variety of dialogue templates for different visual tasks to perform task recognition instruction tuning for LLM. The LLM is trained using LoRA [25] in this stage. + +Stage 2: Vision Task Training. We integrate task tokens and task-specific heads into the model. By training on task-specific data, the model equips the capacity with fine-grained perception and aligns between the task head and the MLLM. The region head, temporal head, mask adapter, and corresponding task tokens are trained respectively. Similarly, the LLM is updated with LoRA. + +Stage 3: Multi-task Training. We tune the entire model on the mixed corpus combining multimodal conversations and specific task data. The synergistic training allows gradients from the heads to MLLM, supervised by human annotations in vision tasks. The vision encoder, connector, task tokens, region head, temporal head, mask adapter, and LLM (with LoRA) are joint-trained in this stage. + +# 4. Experiment + +We give the implementation and training & testing specifics of our TPO, and then show its results and ablations. + +Implementation Details. We employ VideoChat2 [45], a video-based MLLM, as the primary framework in experiments. For its configuration, we employ UMT-L [44] as the vision encoder, Q-former in $\mathrm{BERT}_{\mathrm{base}}$ [32] as the connector, and Mistral-7B [28] as the language model (LLM). + +Regarding the three task heads, the Region Head is initialized randomly. The Temporal Head utilizes CGDETR [55], with parameters also initialized randomly. The video features input to the temporal head are extracted from the pre-trained InternVideo2 [78], while query features are extracted using the LLM [12]. The Mask Head employs SAM2 [61] and is initialized with its pre-trained weights. Since SAM2 is originally designed solely for tracking, we incorporate a two-layer MLP to encode positioning prompts for spatiotemporal grounding. Additionally, to enable the MLLM to handle the spatial locations provided by the user, we utilize another two-layer MLP to encode the spatial input into the MLLM. + +Our model is trained on a variety of visual task datasets and conversation datasets, as shown in Table 1. More training details are provided in the Appendix. + +Benchmarks. We evaluate TPO on both general image/video understanding benchmarks (mainly in multiple choice form) and specific visual tasks (e.g. grounding, tracking, and so on). Specifically, we run our model along with other approaches on MVBench, VideoMME, NExTGQA, MLVU, MMIU, and SEED-Bench2. MVBench [45] is designed to evaluate multi-modal fine-grained video understanding tasks for clips (lasting around 8 to 16 seconds), consisting of 20 video tasks relying heavily on temporal perception that can hardly be addressed by single-frame analysis. Video-MME [19] is for evaluating MLLMs in both perception and reasoning across varying lengths of + +
TaskSamplesDatasets
Segmentation114.6KMeViS [15], SAMv2 [61]
Temporal Grounding116.5KActivityNet [5], TACoS [62], QVHighlight [36], DiDeMo [24], QuerYD [57], HiREST [95], NLQ [22]
Spatial Grounding540.0KAllseeing-V2 [76], Visual Genome [34], RefCOCO [91], RefCOCO+ [91], RefCOCOg [53]
Conversation3MYouCook2 [14], ActivityNet [5], VideoChat2-IT [45], ShareGPT-4o [11], LLaVA-Hound-DPO [100], ShareGPT4V [8]
+ +Table 1. Overview of Datasets Used in TPO for Various Tasks. + +videos, annotated by humans. NExT-GQA [83] builds on NExT-QA by introducing timestamps that are crucial for understanding questions and determining answers. This requires MLLM to perform multi-step reasoning and emphasize a deeper understanding of both visual and textual content. MLVU [104] is constructed from a wide variety of long videos, with lengths ranging from 3 minutes to 2 hours, and includes nine distinct evaluation tasks. MMIU [54] is a comprehensive evaluation suite designed to assess MLLMs across a wide range of multi-image tasks. It encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions. SEED2-Bench2 [38] is a MLLM benchmark, featuring 24K multiple-choice questions with human annotations. It spans 27 evaluation dimensions, assessing both text and image generation. + +Concerning visual tasks, we test spatial grounding, moment retrieval, highlight detection, tracking, and referring segmentation, including 7 related benchmarks [15, 18, 21, 27, 36, 65, 91], and the results from several corresponding state-of-the-art expert models. + +# 4.1. General Understanding Evaluation + +Multimodal Video Understanding. TPO improves its baseline (VideoChat2) on several video understanding benchmarks with a notable margin. As shown in Table 2, VideoChat-TPO, using a 7B LLM and only 16 input frames, achieves a 66.8 average score on MVBench [45], increasing by $6.4\%$ over VideoChat2 and exceeding the performance of ST-LLM [49] which uses 64 frames. Considering MVBench focuses on characterizing subtle temporal variations involving fine-grained action types, action order, moving direction, and so on, TPO's seamless integration of detailed video understanding (like spatiotemporal grounding and tracking) into its optimized MLLM makes it effectively handle these fine-grained tasks. + +On VideoMME [19], VideoChat-TPO outperforms the compared models, achieving a $9.3\%$ improvement over VideoChat2 and demonstrating significant gains across short and medium-length videos, regardless of subtitle availability. On MLVU [104], VideoChat-TPO achieves a M-AVG score of 54.7, surpassing VideoChat2 by $10.2\%$ . This confirms the effectiveness of TPO's enhanced percep + +
ModelLLM ParamsFramesMVBench [45]VideoMME [19]MLVU [104]
OverallShortMediumLong
AVGw/o s.w/ s.w/o s.w/ s.w/o s.w/ s.w/o s.w/ s.M-AVG
TimeChat [63]7B9638.534.336.939.143.131.833.932.133.630.9
Video-LLAVA [46]7B843.041.141.946.947.338.740.437.837.947.3
ShareGPT4Video [9]7B1651.239.943.648.353.636.339.335.037.946.4
LLaVA-Next-Video [102]7B1644.038.040.844.647.437.739.431.935.639.3
ST-LLM [49]7B6454.937.942.345.748.436.841.431.336.9-
PLLAVA-34B [85]34B1658.140.035.047.236.238.235.934.732.953.6
Chat-UniVi [30]7B6440.840.645.945.751.241.347.339.143.4-
VideoChat2(baseline) [45]7B1660.439.543.848.352.837.039.433.239.244.5
VideoChat-TPO7B1666.8 (+6.4)48.8 (+9.3)53.8 (+10.0)58.864.946.750.041.046.454.7 (+10.2)
+ +Table 2. Performance on Multimodal Video Understanding. We compare our model to others using LLMs of the same generation or 16-frame input. w/o s. indicates without subtitle, while w s. indicates with subtitle. M-AVG refers to the mean average of MLVU. + +tions of long-form video evaluations. These results across three benchmarks demonstrate the notable advancements in multimodal video understanding achieved through TPO. + +Grounded Video QA. Table 3 shows that VideoChat-TPO outperforms other models, achieving superior accuracy (Acc) and intersection over union (IoU) scores in NExT-GQA [83]. Its intersection over prediction (IoP) scores are comparable to those of LLoVi [98], which employs large, closed-source commercial models like GPT-4 [1]. The high Acc@IoP score not only reflects the enhanced capability of TPO-optimized VideoChat in effectively understanding and interpreting video content, but also demonstrates its robustness in handling complex reasoning scenarios. Furthermore, the higher Acc@GQA score indicates that the model successfully integrates fine-grained temporal grounding with complex reasoning tasks, enabling it to accurately provide temporal clues necessary for inferring answers. From this perspective, TPO training framework significantly allows MLLMs to overcome their limitations in identifying and localizing temporal elements within video data. This enhancement positions MLLM-TPO with a competitive edge in temporal-related reasoning. + +Multimodal Multi-image Understanding. To explore TPO's effect on multi-image understanding, we test it on the MMIU and SEED-Bench2, as shown in Table 4. On MMIU, VideoChat-TPO achieved an overall score of 40.2, a $5.2\%$ improvement over VideoChat2. Besides temporal changes, MMIU evaluates models' spatial sensing and semantic relations in scenes. On SEED-Bench2, across 27 performance indicators, VideoChat-TPO achieves $41.7\%$ improvement on average performance compared to VideoChat2. VideoChat-TPO's superior performance on multi-image understanding compared to LLaVA-Interleave and InternLM-XComposer demonstrates that TPO's vision enhancements improve the MLLM's spatial perception and image understanding. + +# 4.2. Vision Task Evaluation + +Moment Retrieval. Moment Retrieval is to locate the target segments in a video based on queries. Table 5 and 6 compare the zero-shot and fine-tuned moment retrieval performance of VideoChat-TPO against other expert models and MLLMs. In zero-shot settings, VideoChat-TPO achieves a $R@1$ ( $IoU = 0.5$ ) score of 40.2 on the Charades-STA [21]. It surpasses the previous state-of-the-art (SOTA) MLLM specialized for temporal tasks, ChatVTG [59], and the expert model, UniVTG [47]. This demonstrates VideoChat-TPO's ability to accurately locate video moments corresponding to given text queries, thereby enhancing the practical applicability of MLLMs. + +Highlight Detection. Highlight detection generates salient scores for emphasizing frames based on the input language query. We compare the fine-tuning highlight detection performance of VideoChat-TPO with that of other expert models and MLLMs, as shown in Table 6. VideoChat-TPO notably outperforms the previous state-of-the-art MLLM (such as TimeChat [63]) on both Charades-STA and QVHighlight. For expert models like QD-DETR and UniVTG, VideoChat-TPO beats them on Charades-STA non-trivially and achieves a comparable performance on QVHighlight. This also demonstrates the progress in extending MLLMs to broad temporal tasks. + +Spatial Grounding. To verify the fine-grained localization ability of the model, we run the spatial grounding task which inputs textual descriptions into the model and gets the corresponding bounding box on RefCOCO [91]. As shown in Table 7, we compare VideoChat-TPO with pixel-to-sequence models, i.e., VisionLLM-H [75], pixel-to-embedding methods, i.e., NExT-Chat [97], and expert models, i.e., G-DINO [50]. VideoChat-TPO performs better than the pixel-to-embedding methods by using only a simple task head and achieves comparable performance to pixel-to-sequence models fine-tuned on a large amount of spatial grounding data, as well as to specialized models. + +
ModelAcc@IoPAcc@GQAmIoPIoP@0.3IoP@0.5mIoUIoU@0.3IoU@0.5
VIOLETv2 [20]54.912.823.625.123.33.14.31.3
SeViLA [93]72.516.629.534.722.921.729.213.8
LangRepo [31]59.617.131.3-28.718.5-12.2
FrozenBiLM NG+ [88]73.817.524.228.523.79.613.56.1
VideoStreaming [58]57.417.832.2-31.019.3-13.3
LLoVi [98]65.924.337.3-36.920.0-15.3
HawkEye [79]-----25.737.019.5
VideoChat-TPO77.725.535.647.532.827.741.223.4
+ +Table 3. Performance on Grounded QA. + +
ModelMMIU [54]SEED2I [38]SEED2M [38]
LLaVA-v1.5 [48]19.258.339.2
ShareGPT4V [8]18.5--
OpenFlamingo [2]22.336.643.5
LLaVA-Interleave [40]32.4--
VideoChat2 [45]35.026.527.6
VideoChatGPT [52]-38.349.8
InternLM-XComposer [16]21.965.449.8
VideoChat-TPO40.2 (+5.2)67.3 (+40.8)70.0 (+42.4)
+ +Table 4. Performance on Image Understanding. + +
ModelCharades-STA [21]
R@0.3R@0.5R@0.7mIoU
UniVTG [47]44.125.210.027.1
VideoChat2 [45]38.014.33.824.6
VTimeLLM [26]51.027.511.431.2
TimeChat [63]-32.213.4-
HawkEye [79]50.631.414.533.7
ChatVTG [59]52.733.015.934.9
VideoChat-TPO58.340.218.438.1
+ +
ModelCharades-STA [21]QVHighlight [36]
R@0.3R@0.5R@0.7mIoUmAPHIT@1
M-DETR [36]65.852.130.645.535.755.6
QD-DETR [56]-57.332.6-38.962.4
UniVTG [47]72.660.238.652.240.566.3
TimeChat [63]-46.723.7-21.737.9
HawkEye [79]72.558.328.8---
VideoChat-TPO77.065.040.755.038.665.4
+ +
MethodsRefCOCO [91]
valtestAtestB
MAntNet ★ [92]76.480.469.3
OFA-L [73]80.083.776.4
G-DINO-L ★ [50]90.693.288.2
VisionLLM-H [75]-86.7-
Shikra-7B [7]87.090.680.2
NExT-Chat-7B [97]85.590.077.9
VideoChat-TPO85.990.881.3
+ +Table 5. Zero-Shot Performance on Table 6. Fine-tuning Performance on Moment Re-Moment Retrieval. Gray means no LLM. Retrieval and Highlight Detection. Gray means no LLM. + +Table 7. Spatial Grounding Task. $\star$ with a refined decoder. + +Tracking. In the tracking task, the model receives the object coordinates from the first frame and outputs the coordinates for the remaining frames in the video. We evaluate the VideoChat-TPO on the mainstream tracking benchmarks LAOST [18] and GOT-10k [27], as shown in Table 8. In zero-shot testing, VideoChat-TPO ranks best among all MLLMs, even surpassing some fine-tuned expert models notably with around $10\%$ increase on success rate, such as SiamFC [4] and ATOM [13]. With the TPO training method, the model is optimized from mask sequences that hard to represent by words, allowing the MLLM to achieve strong motion characterization for multimodal generalization in highly dynamic and occluded scenes. + +Referring Segmentation. The referring segmentation task requires the model to output the mask sequence corresponding to the specified prompt. This capability is newly activated in VideoChat-TPO, as it is not supported or designed in the original VideoChat2 or SAM2. We compare the referring segmentation performance of VideoChat-TPO with other expert models and MLLMs in Table 9. The zero-shot capability of VideoChat-TPO is comparable to the fine-tuning capabilities of other expert models, i.e. ReferFormer [81], showing its notable open-world video segmentation performance. Through TPO training, MLLM effectively optimizes its understanding in both tracking and segmenting objects indicated by users. It further enables the model to excel in pixel-level tasks, offering perceptual advantages to it in practical applications like robotic control. + +# 4.3. Ablation Studies + +In this section, we analyze the effectiveness of the key components of TPO and evaluate its scalability. + +Extending to Other MLLMs. To evaluate the effectiveness of the TPO method on different MLLMs, we apply TPO to LLaVA (LLaVA-OneVision [39]) in addition to VideoChat. In LLaVA, fine-grained video understanding is a weak spot. Despite the conversation data used in LLaVA-OneVision being very similar to, or even more extensive than, what we employ, there remains significant room for improvement in its performance on short-term fine-grained understanding tasks. As shown in Table 10, with TPO, LLaVA not only gains capabilities that it originally lacked but also achieves an $8.1\%$ improvement on MVBench, demonstrating the universality of the TPO method. + +Task Preference Model vs. Textualized Task Data. We compare the TPO method with the approach of textualizing task data and training it in an autoregressive manner, which is a straightforward and efficient method for MLLMs to learn specific tasks. We use the same data as the TPO method and convert the data related to the task head into conversation format. Due to the limitations of textual representation, much data must be input in a more discrete manner. For instance, in tracking supervision data, we follow Merlin [90] by converting masks into a sequence of spatial coordinates to serve as both the model's input and output. As shown in Table 11, MLLM-TPO achieves a $2.1\%$ performance improvement on MVBench [45] compared to MLLMs trained with textualized task data. This demonstrates that the task head in the TPO method enables more effective utilization of the original supervision signals compared to the next-token prediction approach, which inevitably incurs information loss when converting fine-grained tasks into textual annotations. The next-token prediction method cannot truly capture the nuanced information present in videos, while TPO allows the MLLM to + +
ModelLaSOT [18]GOT-10k [27]
SuccessPnormPOverlapSR0.5SR0.75
SiamFC [4]33.642.033.934.835.39.8
ATOM [13]51.5--55.663.440.2
SiamRPN++ [37]49.656.949.151.861.832.5
SiamFC++ [86]54.462.354.759.569.547.9
LLaVA-1.5 [48]19.416.512.823.520.29.7
Merlin [90]39.840.238.151.455.942.8
VideoChat-TPO69.480.176.970.679.866.0
+ +Table 8. Performance on Tracking Benchmarks. + +
MethodRef-YouTube-VOS [65]MeViS [15]
J&FJFJ&FJF
ReferFormer [81]62.961.364.631.029.832.2
OnlineRefer [80]62.961.064.7---
LISA [35]52.652.153.0---
VideoLISA [3]63.761.763.744.441.347.6
VideoChat-TPO63.952.375.447.042.651.3
+ +Table 9. Performance on Referring Segmentation. TPO is evaluated in a zero-shot on Ref-YouTube. + +
ModelMVBench
LLaVA-OV [39]56.7
LLaVA-OV-TPO64.8 (+8.1)
+ +Table 10. Applying TPO on LLaVA-OV. TPO significantly enhances the ability to understand fine-grained video details. + +
ModelCharades-STA [21]MVBench
R@0.3R@0.5R@0.7mIoUAVG
VideoChat-TPO58.340.218.438.166.8
w/o reasoning data56.438.317.135.666.1
replace by simple head31.517.86.115.465.8
textualized task data33.718.66.216.064.7
+ +Table 11. Ablation of Reasoning Data and Head Performance. + +
Exp.TRMCR@0.5Acc@0.5J&F
1✓30.2--
2✓-77.3-
3✓--55.1
4✓✓-80.258.3
5✓✓✓36.781.661.4
6✓✓✓✓40.282.063.9
+ +Table 12. Impact of TPO Components and Data. T, R, M, and C denote temporal head, region head, mask head, and conversation data respectively. R1@0.5 means R1@0.5 in Charades-STA, Acc@0.5 represents the mean of Acc@0.5 in all COCO datasets, $\mathcal{J}\&\mathcal{F}$ means $\mathcal{J}\&\mathcal{F}$ in Ref-YouTube-VOS. + +learn perceptual information beyond conversational data. + +Impact of Task Head Performance on TPO. A stronger task head is expected to enhance the multimodal capabilities of TPO more effectively than a weaker one. To further explore the impact of the task head on model performance, we replaced the temporal head from CG-DETR [55] with a simple two-layer MLP. As shown in Table 11, the simple temporal head exhibits a significant performance decline compared to CG-DETR on the corresponding tasks. However, it still achieves a $1.1\%$ improvement compared to the model without TPO. In contrast to the simpler head, the well-designed head can provide more accurate expert knowledge through its architecture and pre-trained task preference weights, enabling better utilization of the data. + +Impact of Data Scaling. Concerning foundation or large models, we expect to enhance their model performance by incorporating additional data at scale. In our approach, we expand the existing conversation data by integrating two reasoning datasets [5, 14] with temporal information. By incorporating these datasets, we observe a notable enhancement in the capabilities of the temporal head in Table 11. Furthermore, as shown in Table 12, adding conversation data further improves performance across various + +tasks. Overall, increasing the amount of conversation and fine-grained task data leads to noticeable improvements in model performance, demonstrating the effectiveness of data scaling in training multimodal models with TPO. + +Synergistic Gains from Co-training. Table 12 presents the impact of incorporating different visual tasks into TPO on performance. This indicates that the collaborative learning of visual tasks facilitates the transfer of knowledge, ultimately leading to better performance across the board. Overall, these results confirm that co-training not only benefits individual tasks but also creates a synergistic effect that enhances overall capabilities in visual understanding. + +# 5. Conclusions + +This paper introduces Task Preference Optimization (TPO). It enhances the overall multimodal performance of MLLMs by enhancing their precise visual understanding. TPO achieves this by integrating differentiable task preferences derived from fine-grained visual tasks. It introduces learnable task tokens that serve as bridges between multiple task-specific heads and the MLLM. Through the joint optimization of these task tokens, heads, and the MLLM, TPO leads to a substantial improvement in multimodal dialogue capabilities and performance across various visual tasks. Our results demonstrate the effectiveness of TPO in scaling MLLMs with task-specific data and seamlessly integrating them with existing expert vision models. We believe this study clarifies the prerequisites for fusing MLLMs with models and data from the pre-large model era, effectively bridging the gap between expert and generalist models, as well as between generation and understanding. + +Limitations. Currently, TPO focuses exclusively on discriminative visual tasks, overlooking generative ones. Meanwhile, the framework is supervised by human annotations, which neglects potentially valuable unsupervised or self-supervised learning approaches such as contrastive learning [10, 23]. This limitation inherently restricts the scalability of TPO in terms of both task diversity and requirements. While we demonstrate TPO's potential for enhancing MLLMs through increased task-specific data, a comprehensive investigation into the broader impact of this scaling remains a crucial area for future research. + +# Acknowledgements + +This work is partially supported by the National Key R&D Program of China (No. 2022ZD0160102), Jiangsu Frontier Technology R&D Program (No. BF2024076), and the Science and Technology Commission of Shanghai Municipality under Grant No. 23YF1461900. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6 +[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortman, and Ludwig Schmidt. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 7 +[3] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 3, 8 +[4] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In ECCV, pages 850–865, 2016. 7, 8 +[5] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 5, 8 +[6] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv preprint arXiv:2304.02948, 2023. 1 +[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1, 3, 7 +[8] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In ECCV, 2024. 5, 7 +[9] Lin Chen, Xin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, Li Yuan, Yu Qiao, Dahua Lin, Feng Zhao, and Jiaqi Wang. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 6 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learn + +ing of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 8 +[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 5 +[12] Yiming Cui, Ziqing Yang, and Xin Yao. Efficient and effective text encoding for chinese llama and alpaca. arXiv preprint arXiv:2304.08177, 2023. 5 +[13] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In CVPR, pages 4660-4669, 2019. 7, 8 +[14] Pradipto Das, Chenliang Xu, Richard F Doell, and Jason J Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In CVPR, pages 2634-2641, 2013. 5, 8 +[15] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 5, 8 +[16] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 7 +[17] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 1 +[18] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, pages 5374-5383, 2019. 5, 7, 8 +[19] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 5, 6 +[20] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. An empirical study of end-to-end video-language transformers with masked visual modeling. In CVPR, 2023. 7 +[21] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 5, 6, 7, 8 +[22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18995-19012, 2022. 5 +[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 8 + +[24] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 5 +[25] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 4 +[26] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024. 7 +[27] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 43(5):1562-1577, 2019. 5, 7, 8 +[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 5 +[29] Fan Jiang and Zilei Wang. Sparse sharing relation network for panoptic driving perception. In ACMMM, pages 800-808, 2023. 1 +[30] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In CVPR, pages 13700-13710, 2024. 6 +[31] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv preprint arXiv:2403.14622, 2024.7 +[32] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 5 +[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3 +[34] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 5 +[35] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, pages 9579-9589, 2024. 3, 8 +[36] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. NeurIPS, 34:11846-11858, 2021. 5, 7 +[37] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. SiamRPN++: Evolution of siamese visual tracking with very deep networks. In CVPR, pages 4282-4291, 2019. 8 +[38] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In CVPR, pages 13299-13308, 2024. 2, 5, 7 + +[39] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 7, 8 +[40] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7 +[41] Hao Li, Jinguo Zhu, Xiaohu Jiang, Xizhou Zhu, Hongsheng Li, Chun Yuan, Xiaohua Wang, Yu Qiao, Xiaogang Wang, Wenhai Wang, et al. Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In CVPR, pages 2691-2700, 2023. 2, 3 +[42] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 3 +[43] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 3 +[44] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2, 5 +[45] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 5, 6, 7 +[46] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6 +[47] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 6, 7 +[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 2, 3, 7, 8 +[49] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In ECCV, 2024. 5, 6 +[50] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In ECCV, 2024. 6, 7 +[51] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 2 +[52] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 3, 7 + +[53] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. 5 +[54] Fanqing Meng, Jin Wang, Chuanhao Li, Quanfeng Lu, Hao Tian, Jiaqi Liao, Xizhou Zhu, Jifeng Dai, Yu Qiao, Ping Luo, et al. Mmiu: Multimodal multi-image understanding for evaluating large vision-language models. arXiv preprint arXiv:2408.02718, 2024. 2, 5, 7 +[55] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 5, 8 +[56] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 7 +[57] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP, pages 2265-2269. IEEE, 2021. 5 +[58] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuangrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009, 2024.7 +[59] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPR, pages 1847-1856, 2024. 6, 7 +[60] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 36, 2024. 2, 3 +[61] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4, 5 +[62] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics, 1:25-36, 2013. 5 +[63] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. CVPR, abs/2312.02051, 2024. 1, 3, 6, 7 +[64] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3 +[65] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223. Springer, 2020. 5, 8 +[66] Jing Shao, Siyu Chen, Yangguang Li, Kun Wang, Zhenfei Yin, Yinan He, Jianing Teng, Qinghong Sun, Mengya Gao, + +Jihao Liu, et al. Intern: A new learning paradigm towards general vision. arXiv preprint arXiv:2111.08687, 2021. 1, 2 +[67] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 3 +[68] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1 +[69] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, pages 10078-10093, 2022. 2 +[70] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3 +[71] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, pages 20-36. Springer, 2016. 2 +[72] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, 2023. 2 +[73] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, pages 23318-23340. PMLR, 2022. 7 +[74] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, pages 6312-6322, 2023. 2 +[75] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2024. 3, 6, 7 +[76] Weiyun Wang, Yiming Ren, Haowen Luo, Tiantong Li, Chenxiang Yan, Zhe Chen, Wenhai Wang, Qingyun Li, Lewei Lu, Xizhou Zhu, et al. The all-seeing project v2: Towards general relation comprehension of the open world. arXiv preprint arXiv:2402.19474, 2024. 5 +[77] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In ICCV, pages 1130-1140, 2023. 2 +[78] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. In ECCV, 2024. 2, 5 + +[79] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3, 7 +[80] Dongming Wu, Tiancai Wang, Yang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In ICCV, pages 2761-2770, 2023. 8 +[81] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 7, 8 +[82] Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Wenhai Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. arXiv preprint arXiv:2406.08394, 2024. 3 +[83] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In CVPR, 2024. 5, 6 +[84] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pretraining for zero-shot video-text understanding. In EMNLP, pages 6787-6800, 2021. 2 +[85] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 6 +[86] Yinda Xu et al. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, pages 140-148, 2020. 8 +[87] Shen Yan, Tao Zhu, Zirui Wang, Yuan Cao, Mi Zhang, Soham Ghosh, Yonghui Wu, and Jiahui Yu. Video-text modeling with zero-shot transfer from contrastive captioners. ArXiv, abs/2212.04979, 2022. 2 +[88] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. In NeurIPS, 2022. 7 +[89] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 3 +[90] En Yu, Liang Zhao, Yana Wei, Jinrong Yang, Dongming Wu, Lingyu Kong, Haoran Wei, Tiancai Wang, Zheng Ge, Xiangyu Zhang, et al. Merlin: Empowering multimodal llms with foresight minds. In ECCV, pages 425-443. Springer, 2025. 3, 7, 8 +[91] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 5, 6, 7 +[92] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In CVPR, pages 1307-1315, 2018. 7 + +[93] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In NeurIPS, 2023. 7 +[94] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In CVPR, pages 13807-13816, 2024. 3 +[95] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In CVPR, pages 23056-23065, 2023. 5 +[96] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In ICLR. 1 +[97] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An Imm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 3, 6, 7 +[98] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lllm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 6, 7 +[99] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, pages 543–553, 2023. 3 +[100] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 5 +[101] Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 3 +[102] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model, 2024. 6 +[103] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In ICML, 2024. 2 +[104] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2, 5, 6 +[105] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. In ICLR Workshop on Reliable and Responsible Foundation Models, 2024. 3 +[106] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with + +conditional moes. In NeurIPS, pages 2664-2678, 2022. 2, 3 +[107] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 2, 3 \ No newline at end of file diff --git a/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/images.zip b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bf07f7bdb826167e025380e162b94e18c2355b82 --- /dev/null +++ b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6baf0879869a18a43e88122342b0ef5bcb9a00bfe4a6cfe1838dbe92d32d261 +size 476369 diff --git a/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/layout.json b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f52b013ad2ec38c663edeca88e99228c6ec90809 --- /dev/null +++ b/2025/Task Preference Optimization_ Improving Multimodal Large Language Models with Vision Task Alignment/layout.json @@ -0,0 +1,10197 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 541, + 140 + ], + "type": "text", + "content": "Task Preference Optimization: Improving Multimodal Large Language Models with Vision Task Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 160, + 395, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 160, + 395, + 175 + ], + "spans": [ + { + "bbox": [ + 203, + 160, + 395, + 175 + ], + "type": "text", + "content": "Ziang Yan\\*2,1, Zhilin Li\\*3,1, Yinan He\\*1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "spans": [ + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "type": "text", + "content": "Chenting Wang" + }, + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "type": "inline_equation", + "content": "^{4,1}" + }, + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "type": "text", + "content": ", Kunchang Li" + }, + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "type": "inline_equation", + "content": "^{5,1}" + }, + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "type": "text", + "content": ", Xinhao Li" + }, + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "type": "inline_equation", + "content": "^{6,1}" + }, + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "type": "text", + "content": ", Xiangyu Zeng" + }, + { + "bbox": [ + 140, + 175, + 459, + 190 + ], + "type": "inline_equation", + "content": "^{6,1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "spans": [ + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "text", + "content": "Zilei Wang" + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "text", + "content": ", Yali Wang" + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "inline_equation", + "content": "^{5,1}" + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "text", + "content": ", Yu Qiao" + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "text", + "content": ", Limin Wang" + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "inline_equation", + "content": "^{6,1}" + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "text", + "content": ", Yi Wang" + }, + { + "bbox": [ + 142, + 190, + 457, + 204 + ], + "type": "inline_equation", + "content": "^{\\dagger,7}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "type": "text", + "content": "University of Science and Technology of China " + }, + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 56, + 205, + 545, + 217 + ], + "type": "text", + "content": "Shanghai Jiao Tong University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 218, + 543, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 543, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 543, + 231 + ], + "type": "text", + "content": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences Nanjing University Shanghai Innovation Institute" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 234, + 387, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 234, + 387, + 244 + ], + "spans": [ + { + "bbox": [ + 212, + 234, + 387, + 244 + ], + "type": "text", + "content": "https://github.com/OpenGVLab/TPO" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 299, + 290, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 290, + 597 + ], + "type": "text", + "content": "Current multimodal large language models (MLLMs) struggle with fine-grained or precise understanding of visuals although they give comprehensive perception and reasoning in a spectrum of vision applications. Recent studies either develop tool-using or unify specific visual tasks into the autoregressive framework, often at the expense of overall multimodal performance. To address this issue and enhance MLLMs with visual tasks in a scalable fashion, we propose Task Preference Optimization (TPO), a novel method that utilizes differentiable task preferences derived from typical fine-grained visual tasks. TPO introduces learnable task tokens that establish connections between multiple task-specific heads and the MLLM. By leveraging rich visual labels during training, TPO significantly enhances the MLLM's multimodal capabilities and task-specific performance. Through multi-task co-training within TPO, we observe synergistic benefits that elevate individual task performance beyond what is achievable through single-task training methodologies. Our instantiation of this approach with VideoChat and LLaVA demonstrates an overall " + }, + { + "bbox": [ + 46, + 299, + 290, + 597 + ], + "type": "inline_equation", + "content": "14.6\\%" + }, + { + "bbox": [ + 46, + 299, + 290, + 597 + ], + "type": "text", + "content": " improvement in multimodal performance compared to baseline models. Additionally, MLLM-TPO demonstrates robust zero-shot capabilities across various tasks, performing comparably to state-of-the-art supervised models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "content": "Multimodal large language models (MLLMs) show impressive visual perception and reasoning capabilities, with applications in personal assistants [68], embodied systems [17], and scientific discovery [6]. Considering the" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 284, + 323, + 298 + ], + "blocks": [ + { + "bbox": [ + 310, + 284, + 323, + 298 + ], + "lines": [ + { + "bbox": [ + 310, + 284, + 323, + 298 + ], + "spans": [ + { + "bbox": [ + 310, + 284, + 323, + 298 + ], + "type": "image", + "image_path": "9e5025be807bb324267d565245547b627928c16cd143e0e1fed961d01a1e5b01.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 332, + 276, + 471, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 276, + 471, + 306 + ], + "spans": [ + { + "bbox": [ + 332, + 276, + 471, + 306 + ], + "type": "text", + "content": "Show the tracking trajectory of cup with candy in middle. Which cup is the candy under, left one, middle one or right one?" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 486, + 275, + 541, + 308 + ], + "blocks": [ + { + "bbox": [ + 486, + 275, + 541, + 308 + ], + "lines": [ + { + "bbox": [ + 486, + 275, + 541, + 308 + ], + "spans": [ + { + "bbox": [ + 486, + 275, + 541, + 308 + ], + "type": "image", + "image_path": "3c3801e4a772805f647d6a89a8afce4327ab7ebb890e61751a2d7808bcb98f56.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 316, + 489, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 316, + 489, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 489, + 335 + ], + "type": "text", + "content": "The cups are shuffled in the video, and the candy is under the middle cup." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 504, + 315, + 517, + 330 + ], + "blocks": [ + { + "bbox": [ + 504, + 315, + 517, + 330 + ], + "lines": [ + { + "bbox": [ + 504, + 315, + 517, + 330 + ], + "spans": [ + { + "bbox": [ + 504, + 315, + 517, + 330 + ], + "type": "image", + "image_path": "557ea63bd3e484136e7dfeda25df6c6a64e6adb3d465243e4158f0d6592931c1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 523, + 313, + 542, + 332 + ], + "blocks": [ + { + "bbox": [ + 523, + 313, + 542, + 332 + ], + "lines": [ + { + "bbox": [ + 523, + 313, + 542, + 332 + ], + "spans": [ + { + "bbox": [ + 523, + 313, + 542, + 332 + ], + "type": "image", + "image_path": "cff516a6b12e6815e2d186d4a735a81fd8693a1de6e3fd5c49c7acd465e582ef.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 311, + 335, + 517, + 367 + ], + "blocks": [ + { + "bbox": [ + 311, + 335, + 517, + 367 + ], + "lines": [ + { + "bbox": [ + 311, + 335, + 517, + 367 + ], + "spans": [ + { + "bbox": [ + 311, + 335, + 517, + 367 + ], + "type": "image", + "image_path": "5d11b7e46cf72f07c856552c1f6f0c2bb9392b4f45278ffa2e1ef7237042e5b8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 373, + 518, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 373, + 518, + 393 + ], + "spans": [ + { + "bbox": [ + 311, + 373, + 518, + 393 + ], + "type": "text", + "content": "The candy is under the candy is under the the left cup. first cup on the left." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 522, + 372, + 541, + 392 + ], + "blocks": [ + { + "bbox": [ + 522, + 372, + 541, + 392 + ], + "lines": [ + { + "bbox": [ + 522, + 372, + 541, + 392 + ], + "spans": [ + { + "bbox": [ + 522, + 372, + 541, + 392 + ], + "type": "image", + "image_path": "9ad6c3de098d56f274573d4269296257f9180c82982ae5ef2019dd74b6290f7a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 404, + 546, + 437 + ], + "lines": [ + { + "bbox": [ + 305, + 404, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 404, + 546, + 437 + ], + "type": "text", + "content": "Figure 1. TPO uses differentiable task preferences from dense visual supervisions via task-specific heads to enhance MLLMs in fine-grained understanding." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 450, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 510 + ], + "type": "text", + "content": "growing expectations of users in more accurate and detailed perception, taking the shell game as an example, further improving the generality of MLLM requires fine-grained knowledge representation beyond words, e.g. MLLM implicitly embeds how to track keys indicated by users." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "content": "Existing studies address the enhancement of perceptual granularity in MLLMs by focusing on specific visual tasks (e.g. temporal grounding, segmentation, tracking) via MLLMs. They usually fine-tune MLLMs on more task data in text format or enable MLLMs to activate the corresponding task heads. Shikra [7] applies MLLM to localization tasks, transforms object coordinates into dialogue formats, and learns them autoregressively. Meanwhile, TimeChat [63] and VideoChat-T [96] treat event timestamps as text for autoregressive prediction, thereby endowing MLLMs with temporal grounding. They do improve specific task performance significantly, while at the cost of multimodal performance more or less. This is counterintuitive as seminal research proves different visual tasks are correlated and training them together often yields mutual gains [29, 47, 66]. We conjecture that the presentation of different tasks influences this training and that the conflict" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 207, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 207, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 207, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29880" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "arises from the learning discrepancies between discrete textual tokens and visually dense predictions. Our experiments in Section 4 validate that a decoupled representation design can effectively address this issue." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 350 + ], + "type": "text", + "content": "To enhance the multimodal capabilities of MLLMs, we explore optimization methods to meet multiple visual task requirements in an end-to-end manner. We propose task preference optimization (TPO), incorporating visual task knowledge into the MLLMs by jointly maximizing the likelihoods of visual task estimations and multimodal dialogue. Inspired by direct preference optimization [60] (DPO) and related methods, we treat visual task annotations as human preferences in particular demands, as shown in Figure 2. DPO aids LLM (or MLLM) in aligning with human preferences through a binary classification that directs the model to generate responses that people prefer. Similarly, TPO enhances MLLMs' visual sensing capabilities through differentiable task optimizations that guide MLLMs to yield dense predictions closely resembling human perception. To achieve this, TPO concretizes MLLM-specific visual perceptions into corresponding task tokens, disentangled from MLLM representation. Then TPO fine-tunes these task tokens and updates the MLLM accordingly." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 506 + ], + "type": "text", + "content": "Specifically, TPO appends visual task heads to the partial output of the MLLM, using several learnable task tokens as inputs for these corresponding heads. During training, TPO enables the MLLM to first distinguish and activate the appropriate task tokens based on user instructions. Subsequently, TPO jointly trains the task tokens and their corresponding heads to enhance the MLLM's understanding of visual tasks. Finally, TPO trains the entire model—including task tokens and heads—on both multimodal and visual-specific task data, promoting the perception and reasoning capabilities of the MLLM. Additionally, we note that multi-task co-training yields greater improvements than single-task training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 507, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 288, + 662 + ], + "type": "text", + "content": "TPO demonstrates scalability across various MLLM approaches, encompassing a wide range of visual task categories and data quantities. We validate the effectiveness of TPO within widely used MLLMs, such as LLaVA [48] and VideoChat [43], as detailed in Section 4.3. By fine-tuning these open-source MLLMs with TPO, we enhance their visual understanding capabilities and improve dialogue performance. Additionally, we explore several key spatiotemporal and spatial perception tasks, including spatial grounding, tracking, and temporal grounding. Our findings indicate that these tasks can mutually enhance each other's performance, particularly contributing to the improvement of multimodal dialogue capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 664, + 226, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 664, + 226, + 675 + ], + "spans": [ + { + "bbox": [ + 59, + 664, + 226, + 675 + ], + "type": "text", + "content": "Our contributions can be summarized as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "- We propose a new training method for multimodal large language models, referred to as Task Preference Optimization (TPO). This method leverages supervised in" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 545, + 177 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 177 + ], + "type": "image", + "image_path": "e3b5d96dc60e98838535b1935eaa8b58307386009df569516318a6558db9d88b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 186, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 186, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 545, + 220 + ], + "type": "text", + "content": "Figure 2. Comparison of Learning Method. A solid line indicates data flow, and a dotted line represents feedback. " + }, + { + "bbox": [ + 305, + 186, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 305, + 186, + 545, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 186, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\text{心}" + }, + { + "bbox": [ + 305, + 186, + 545, + 220 + ], + "type": "text", + "content": " denote modules that are frozen and unfrozen." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 230, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 230, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 312, + 230, + 545, + 301 + ], + "type": "text", + "content": "formation from visual task-specific data to optimize the MLLM through the corresponding heads, resulting in significantly enhanced multimodal perception and reasoning performance. Specifically, TPO achieves an average improvement of " + }, + { + "bbox": [ + 312, + 230, + 545, + 301 + ], + "type": "inline_equation", + "content": "14.6\\%" + }, + { + "bbox": [ + 312, + 230, + 545, + 301 + ], + "type": "text", + "content": " across multiple image and video multimodal benchmarks [19, 38, 45, 54, 104]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 301, + 545, + 492 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 301, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 301, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 545, + 373 + ], + "type": "text", + "content": "- TPO effectively equips MLLM with the capability to address several key visual perception tasks through the introduced task heads. MLLM-TPO achieves comparable performance in spatial grounding, moment retrieval, highlight detection, referring segmentation, and tracking comparable to expert models across various benchmarks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 373, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 545, + 492 + ], + "type": "text", + "content": "- TPO demonstrates scalability across various employed MLLMs, task heads, and scales of task data. We validate the effectiveness of TPO in multiple mainstream MLLMs, such as VideoChat2 [45] and LLaVA [39, 48]. Notably, multi-task joint training based on TPO enhances both multimodal performance and individual visual task, with improvements becoming increasingly significant as additional appropriate heads are introduced. Furthermore, the performance of the MLLM and task heads improves when scaling task data." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 654 + ], + "type": "text", + "content": "Vision Foundation Model. Vision foundation models [44, 66, 69, 71, 72, 74, 78, 84, 87, 103] are designed to be adaptable for various downstream tasks through extensive pre-training on large-scale and diverse datasets. VideoPrism [103] achieves the leading results in various video tasks by combining video-text contrastive learning and spatiotemporal token reconstruction using a dataset of public and proprietary videos. InternVideo2 [78] utilizes masked reconstruction, cross-modal contrastive learning, and next-token prediction to enhance the model's perceptiveness, semantic understanding, and reasoning capabilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Based on vision foundation models, some studies [41, 51, 77, 106, 107] intend to incorporate downstream task heads into this framework and expect end-to-end task training. Unified-IO [51] introduces a versatile model capable of handling a wide range of tasks across vision and lan" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29881" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 542, + 265 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 542, + 265 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 542, + 265 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 542, + 265 + ], + "type": "image", + "image_path": "5f3fae02b4b65871ccf8c68d6b7127b7fd4b1a99e935fd534ef46c873e6c026c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "type": "text", + "content": "Figure 3. Overall Pipeline of TPO. The architecture of Task Preference Optimization (TPO) consists of four main components: (1) a vision encoder, (2) a connector, (3) a large language model, and (4) a series of visual task heads. Differently colored flame symbols indicate which components are unfrozen at various stages of the training process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 329, + 289, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 289, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 289, + 426 + ], + "type": "text", + "content": "guage domains, requiring minimal or no task-specific modifications or additional parameters. Uni-Perceivers [41, 106, 107] formulate different tasks to find the maximum likelihood target for each input through the representation similarity regardless of their modality. Nevertheless, these generalist models are limited to predefined tasks and cannot support flexible open-ended task customization based on language instructions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": "Multimodal Large Language Model. The effective understanding and reasoning of LMMs has attracted the attention of numerous researchers. Limited to its input modality, researchers expand the visual capabilities of LLM, leading to MLLMs. Seminal works, such as BLIP-2 [42], LLaVA [48], and mPLUG-Owl [89], introduce image captioning and visual question answering based on LLM using visual instruction-tuning data. Some video-based MLLMs have been proposed, such as VideoChat [43], VideoChatGPT [52], and Video-LLaMA [99], enabling LLM to gain video understanding capabilities by encoding multiple video frames and using video instruction data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Vanilla MLLMs have achieved impressive results in visual-related dialogues, but barely address some fine-grained tasks, such as segmentation and temporal grounding. To address this challenge, MLLMs usually take one of the following pipelines: pixel-to-sequence (P2S) and pixel-to-embedding (P2E). For P2S methods [7, 63, 70, 79, 90], MLLM directly outputs the textual predictions. TimeChat [63] introduces a time-aware frame encoder and a sliding video Q-former to enhance temporal perception. For P2E methods [3, 35, 75, 82, 97], MLLM compresses and inputs visuals into the downstream decoder for estima" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 329, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 390 + ], + "type": "text", + "content": "tions. LISA [35], NExT-Chat [97], and VideoLISA [3] introduce SAM [33] as a segmentation tool, using a special token as a prompt to connect MLLM and SAM. Vision-LLM v2 [82] designs routing tokens and super-link queries to bridge MLLM with multiple decoders." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 396, + 546, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 576 + ], + "type": "text", + "content": "Alignment in MLLM. Aligning MLLM with human preferences or values is crucial for MLLM's development. Recent works [67, 94, 105] introduce alignment approaches to MLLM, including proximal policy optimization (PPO) [64] and direct preference optimization (DPO) [60], as shown in Figure 2. They usually exploit proprietary models like GPT4-V to build visual preference datasets and then tune MLLMs using PPO or DPO. Llava-RLHF [67] incorporates PPO into a MLLM framework Llava, argued by image captions or question-answers (QA), while Zhou et al. [105] and Zhang et al. [101] give DPO implementations for MLLMs where visual preference data are created by GPT-4V and other open-sourced MLLMs. RLHF-V [94] collects a dense human preference in segmentation and enhances MLLMs with DPO." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 586, + 362, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 362, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 362, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Task Preference Optimization (TPO) aims to enable MLLMs to master classical visual perceptions (such as tracking, temporal grounding, etc.) for better task generalization, regarding that many multimodal reasoning tasks require precise visual cues for accurate and reliable responses. As shown in Figure 3, MLLM-TPO has a typical multimodal model " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " (consisting of a vision encoder " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ", a vision-language connector " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ", and a large language model " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ") and a task preference model " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " with a series of visual" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29882" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "task heads " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "\\{H_{i}\\}_{i = 1\\dots n}" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": " denotes the task head number). These heads connect with MLLM using the embeddings " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{e}_i\\}_{i = 1\\dots n}" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i = G(\\mathbf{v}_i)\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": ") transformed from the learnable task-specific tokens " + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_i\\}_{i = 1\\dots n}" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": ") via MLLM." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "content": "TPO employs a local-to-global training scheme, first adapting task heads to the MLLM and then training them jointly. Specifically, the MLLM " + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "content": " starts to recognize " + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_i\\}_{i=1\\dots n}" + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "content": " by updating " + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "content": " according to user instructions, then we tune " + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_i\\}_{i=1\\dots n}" + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "inline_equation", + "content": "\\{H_i\\}_{i=1\\dots n}" + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "content": " for adapting visual heads to " + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "content": ". Finally, we train both " + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "inline_equation", + "content": "\\{H_i\\}_{i=1\\dots n}" + }, + { + "bbox": [ + 46, + 133, + 286, + 276 + ], + "type": "text", + "content": " together. During inference, MLLM-TPO can respond to users' queries in text, and produce structured visual outputs (like masks, timestamps, trajectories, etc.) when users ask (e.g. yielding time results for \"find when the birthday party starts\"). We detail how to form MLLM-TPO structurally and train it in the following." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 285, + 181, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 181, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 181, + 296 + ], + "type": "text", + "content": "3.1. Task Preference Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "content": "The task preference model (TPM) " + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "content": " contains a series of task tokens " + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_i\\}_{i = 1\\dots n}" + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "content": " and heads " + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\{H_i\\}_{i = 1\\dots n}" + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "content": ". Before TPM works, the attached MLLM generates special tokens indicating task types from the input queries. Then TPM dynamically calls the task token " + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j" + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "content": " based on the special token, transforms it to the task embedding " + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_j" + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "content": " via the LLM " + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "content": " (the last hidden embedding), and feeds " + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_j" + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "content": " to the corresponding visual task head " + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\{H_j\\}" + }, + { + "bbox": [ + 46, + 304, + 286, + 495 + ], + "type": "text", + "content": " for specific task predictions. Considering existing MLLMs demonstrate remarkable capabilities in common object and scene recognition, yet struggle to accurately locate things or actions, the employed task heads mainly focus on spatiotemporal localization and tracking. Specifically, we give three fundamental task types for compensating mainstream MLLMs' gaps from expert models in visual perceptions: 1) region head, 2) temporal head, and 3) mask head. Their architectures are given below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 498, + 286, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 286, + 545 + ], + "type": "text", + "content": "- Region Head. A two-layer multilayer perceptron (MLP) with ReLU activations is employed for the region head. It takes embeddings from LLM and regresses them to the bounding box coordinates for spatial grounding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 546, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 286, + 664 + ], + "type": "text", + "content": "- Temporal Head. This head is composed of a video encoder, text encoder, and temporal grounding model for moment retrieval and highlight detection. The video and text queries are input into the video and text encoders to obtain their features respectively. Then we append the temporal task embedding after the text features. With the new text and video features, the MLLM estimates the start/end time and highlight score of the corresponding moment from the query via the temporal grounding model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": "- Mask Head. Pixel-level tasks pose significant challenges for MLLMs due to the lack of the corresponding learning data and optimization in MLLMs. To this end, we introduce a specialized mask head, utilizing the image en" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 167 + ], + "type": "text", + "content": "coder, mask decoder, and memory bank components from SAM2 [61], replacing the prompt encoder with a single MLP layer called the mask adapter. Specifically, for the given visuals and user queries, the features extracted from the aforementioned image encoder and the text representation corresponding to the mask task embeddings from LLM are employed, fed into the mask decoder to produce the desired mask." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 172, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 545, + 220 + ], + "type": "text", + "content": "Most known discriminative vision tasks can be addressed by one or a combination of these three task heads. TPM builds the architectural foundations for leveraging existing vision data with annotations to enhance MLLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 232, + 468, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 468, + 244 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 468, + 244 + ], + "type": "text", + "content": "3.2. Task Preference Optimization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": "TPO improves MLLMs with extra supervisions from visual task heads by back-propagating gradients from heads to update MLLMs using visual task data. It enables the language model " + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": " in MLLM to discriminate specific task types when users demand (task assignment). Then, TPO trains TPM " + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": " via compact task representations " + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i" + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": " (task optimization) from " + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": ". Lastly, we train " + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": " together, tuning " + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": " for the refined spatiotemporal perception according to task preferences from " + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 251, + 545, + 360 + ], + "type": "text", + "content": ". Its optimization objective is given as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 371, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 545, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {m l l m}} + \\underbrace {\\mathcal {L} _ {\\text {a s s i g n}} \\left(G \\left(\\mathbf {T} _ {\\mathrm {q}}\\right) , \\mathbf {s}\\right)} _ {\\text {T a k A s s i g n m e n t}} + \\sum_ {i = 1} ^ {n} \\underbrace {\\mathcal {L} _ {\\text {t a s k}} \\left(\\mathbf {A} _ {i} , H _ {i} \\left(G (\\mathbf {v} _ {i})\\right)\\right)} _ {\\text {T a k O p t i m i z a t i o n}}, \\tag {1}", + "image_path": "15e2f15bcc01fbd5e2480d5318b5b0a498b2b81ea69bdf3d6c4b34a92cb4fca1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathbf{q}}" + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": " denote a user query (some contain specific task needs like tracking) in the form of text token sequence and the ground truth task indicating token, respectively. " + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": " is the task annotation of the input visual " + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": ", so it could be a number tuple for describing regional localization and area or mask for delineating object shape and position. Here " + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathbf{T}_{\\mathbf{q}},\\mathbf{T}_{\\mathbf{a}},\\mathbf{A}_i)" + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": " stands for an input data tuple to MLLM-TPO. Usually, we have " + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathbf{T}_{\\mathbf{q}},\\mathbf{T}_{\\mathbf{a}})" + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": " for typical MLLM training while " + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathbf{T}_{\\mathbf{q}},\\mathbf{A}_i)" + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": " for task token and head training. " + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{assign}}" + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": " is the cross-entropy loss. Meanwhile " + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}_i}" + }, + { + "bbox": [ + 304, + 418, + 545, + 561 + ], + "type": "text", + "content": " varies according to the task, and is usually regression- or classification-related loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 563, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 658 + ], + "type": "text", + "content": "To train MLLM-TPO, we propose a 3-stage local-to-global training scheme, given in Figure 3. Stage 1 learns to identify the task type based on user queries. In stage 2, we train task heads along with their corresponding task tokens, respectively. Lastly, we co-train task heads with MLLM by both task data and multimodal conversation data. Our 3-stage training strategy mitigates the risk of degrading the MLLM's general abilities. We describe them as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 671, + 545, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 671, + 545, + 720 + ], + "spans": [ + { + "bbox": [ + 304, + 671, + 545, + 720 + ], + "type": "text", + "content": "Stage 1: Task Assignment. We design a variety of dialogue templates for different visual tasks to perform task recognition instruction tuning for LLM. The LLM is trained using LoRA [25] in this stage." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29883" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "Stage 2: Vision Task Training. We integrate task tokens and task-specific heads into the model. By training on task-specific data, the model equips the capacity with fine-grained perception and aligns between the task head and the MLLM. The region head, temporal head, mask adapter, and corresponding task tokens are trained respectively. Similarly, the LLM is updated with LoRA." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 162, + 287, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 162, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 287, + 246 + ], + "type": "text", + "content": "Stage 3: Multi-task Training. We tune the entire model on the mixed corpus combining multimodal conversations and specific task data. The synergistic training allows gradients from the heads to MLLM, supervised by human annotations in vision tasks. The vision encoder, connector, task tokens, region head, temporal head, mask adapter, and LLM (with LoRA) are joint-trained in this stage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 257, + 123, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 123, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 123, + 270 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "type": "text", + "content": "We give the implementation and training & testing specifics of our TPO, and then show its results and ablations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 309, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 287, + 369 + ], + "type": "text", + "content": "Implementation Details. We employ VideoChat2 [45], a video-based MLLM, as the primary framework in experiments. For its configuration, we employ UMT-L [44] as the vision encoder, Q-former in " + }, + { + "bbox": [ + 46, + 309, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{BERT}_{\\mathrm{base}}" + }, + { + "bbox": [ + 46, + 309, + 287, + 369 + ], + "type": "text", + "content": " [32] as the connector, and Mistral-7B [28] as the language model (LLM)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 369, + 287, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 524 + ], + "type": "text", + "content": "Regarding the three task heads, the Region Head is initialized randomly. The Temporal Head utilizes CGDETR [55], with parameters also initialized randomly. The video features input to the temporal head are extracted from the pre-trained InternVideo2 [78], while query features are extracted using the LLM [12]. The Mask Head employs SAM2 [61] and is initialized with its pre-trained weights. Since SAM2 is originally designed solely for tracking, we incorporate a two-layer MLP to encode positioning prompts for spatiotemporal grounding. Additionally, to enable the MLLM to handle the spatial locations provided by the user, we utilize another two-layer MLP to encode the spatial input into the MLLM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": "Our model is trained on a variety of visual task datasets and conversation datasets, as shown in Table 1. More training details are provided in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Benchmarks. We evaluate TPO on both general image/video understanding benchmarks (mainly in multiple choice form) and specific visual tasks (e.g. grounding, tracking, and so on). Specifically, we run our model along with other approaches on MVBench, VideoMME, NExTGQA, MLVU, MMIU, and SEED-Bench2. MVBench [45] is designed to evaluate multi-modal fine-grained video understanding tasks for clips (lasting around 8 to 16 seconds), consisting of 20 video tasks relying heavily on temporal perception that can hardly be addressed by single-frame analysis. Video-MME [19] is for evaluating MLLMs in both perception and reasoning across varying lengths of" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 163 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 163 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 163 + ], + "type": "table", + "html": "
TaskSamplesDatasets
Segmentation114.6KMeViS [15], SAMv2 [61]
Temporal Grounding116.5KActivityNet [5], TACoS [62], QVHighlight [36], DiDeMo [24], QuerYD [57], HiREST [95], NLQ [22]
Spatial Grounding540.0KAllseeing-V2 [76], Visual Genome [34], RefCOCO [91], RefCOCO+ [91], RefCOCOg [53]
Conversation3MYouCook2 [14], ActivityNet [5], VideoChat2-IT [45], ShareGPT-4o [11], LLaVA-Hound-DPO [100], ShareGPT4V [8]
", + "image_path": "bef60dc2a6fcb1b8808b8b3581a15bc324fb77c91ce9f8a61dc0e4cd405ac5af.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 164, + 544, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 164, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 544, + 175 + ], + "type": "text", + "content": "Table 1. Overview of Datasets Used in TPO for Various Tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 196, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 545, + 387 + ], + "type": "text", + "content": "videos, annotated by humans. NExT-GQA [83] builds on NExT-QA by introducing timestamps that are crucial for understanding questions and determining answers. This requires MLLM to perform multi-step reasoning and emphasize a deeper understanding of both visual and textual content. MLVU [104] is constructed from a wide variety of long videos, with lengths ranging from 3 minutes to 2 hours, and includes nine distinct evaluation tasks. MMIU [54] is a comprehensive evaluation suite designed to assess MLLMs across a wide range of multi-image tasks. It encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions. SEED2-Bench2 [38] is a MLLM benchmark, featuring 24K multiple-choice questions with human annotations. It spans 27 evaluation dimensions, assessing both text and image generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": "Concerning visual tasks, we test spatial grounding, moment retrieval, highlight detection, tracking, and referring segmentation, including 7 related benchmarks [15, 18, 21, 27, 36, 65, 91], and the results from several corresponding state-of-the-art expert models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 456, + 493, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 456, + 493, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 493, + 468 + ], + "type": "text", + "content": "4.1. General Understanding Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 628 + ], + "type": "text", + "content": "Multimodal Video Understanding. TPO improves its baseline (VideoChat2) on several video understanding benchmarks with a notable margin. As shown in Table 2, VideoChat-TPO, using a 7B LLM and only 16 input frames, achieves a 66.8 average score on MVBench [45], increasing by " + }, + { + "bbox": [ + 304, + 474, + 545, + 628 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 628 + ], + "type": "text", + "content": " over VideoChat2 and exceeding the performance of ST-LLM [49] which uses 64 frames. Considering MVBench focuses on characterizing subtle temporal variations involving fine-grained action types, action order, moving direction, and so on, TPO's seamless integration of detailed video understanding (like spatiotemporal grounding and tracking) into its optimized MLLM makes it effectively handle these fine-grained tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "On VideoMME [19], VideoChat-TPO outperforms the compared models, achieving a " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "9.3\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " improvement over VideoChat2 and demonstrating significant gains across short and medium-length videos, regardless of subtitle availability. On MLVU [104], VideoChat-TPO achieves a M-AVG score of 54.7, surpassing VideoChat2 by " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "10.2\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". This confirms the effectiveness of TPO's enhanced percep" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29884" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 198 + ], + "type": "table", + "html": "
ModelLLM ParamsFramesMVBench [45]VideoMME [19]MLVU [104]
OverallShortMediumLong
AVGw/o s.w/ s.w/o s.w/ s.w/o s.w/ s.w/o s.w/ s.M-AVG
TimeChat [63]7B9638.534.336.939.143.131.833.932.133.630.9
Video-LLAVA [46]7B843.041.141.946.947.338.740.437.837.947.3
ShareGPT4Video [9]7B1651.239.943.648.353.636.339.335.037.946.4
LLaVA-Next-Video [102]7B1644.038.040.844.647.437.739.431.935.639.3
ST-LLM [49]7B6454.937.942.345.748.436.841.431.336.9-
PLLAVA-34B [85]34B1658.140.035.047.236.238.235.934.732.953.6
Chat-UniVi [30]7B6440.840.645.945.751.241.347.339.143.4-
VideoChat2(baseline) [45]7B1660.439.543.848.352.837.039.433.239.244.5
VideoChat-TPO7B1666.8 (+6.4)48.8 (+9.3)53.8 (+10.0)58.864.946.750.041.046.454.7 (+10.2)
", + "image_path": "82b1c3d0d062e96100534244ae2736ebe180b7a8f66e83057cac96d518b184bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 200, + 546, + 222 + ], + "lines": [ + { + "bbox": [ + 47, + 200, + 546, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 546, + 222 + ], + "type": "text", + "content": "Table 2. Performance on Multimodal Video Understanding. We compare our model to others using LLMs of the same generation or 16-frame input. w/o s. indicates without subtitle, while w s. indicates with subtitle. M-AVG refers to the mean average of MLVU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 243, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 281 + ], + "type": "text", + "content": "tions of long-form video evaluations. These results across three benchmarks demonstrate the notable advancements in multimodal video understanding achieved through TPO." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 299, + 289, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 299, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 45, + 299, + 289, + 526 + ], + "type": "text", + "content": "Grounded Video QA. Table 3 shows that VideoChat-TPO outperforms other models, achieving superior accuracy (Acc) and intersection over union (IoU) scores in NExT-GQA [83]. Its intersection over prediction (IoP) scores are comparable to those of LLoVi [98], which employs large, closed-source commercial models like GPT-4 [1]. The high Acc@IoP score not only reflects the enhanced capability of TPO-optimized VideoChat in effectively understanding and interpreting video content, but also demonstrates its robustness in handling complex reasoning scenarios. Furthermore, the higher Acc@GQA score indicates that the model successfully integrates fine-grained temporal grounding with complex reasoning tasks, enabling it to accurately provide temporal clues necessary for inferring answers. From this perspective, TPO training framework significantly allows MLLMs to overcome their limitations in identifying and localizing temporal elements within video data. This enhancement positions MLLM-TPO with a competitive edge in temporal-related reasoning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": "Multimodal Multi-image Understanding. To explore TPO's effect on multi-image understanding, we test it on the MMIU and SEED-Bench2, as shown in Table 4. On MMIU, VideoChat-TPO achieved an overall score of 40.2, a " + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": " improvement over VideoChat2. Besides temporal changes, MMIU evaluates models' spatial sensing and semantic relations in scenes. On SEED-Bench2, across 27 performance indicators, VideoChat-TPO achieves " + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "inline_equation", + "content": "41.7\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": " improvement on average performance compared to VideoChat2. VideoChat-TPO's superior performance on multi-image understanding compared to LLaVA-Interleave and InternLM-XComposer demonstrates that TPO's vision enhancements improve the MLLM's spatial perception and image understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 243, + 437, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 437, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 437, + 255 + ], + "type": "text", + "content": "4.2. Vision Task Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 262, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 406 + ], + "type": "text", + "content": "Moment Retrieval. Moment Retrieval is to locate the target segments in a video based on queries. Table 5 and 6 compare the zero-shot and fine-tuned moment retrieval performance of VideoChat-TPO against other expert models and MLLMs. In zero-shot settings, VideoChat-TPO achieves a " + }, + { + "bbox": [ + 304, + 262, + 547, + 406 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 304, + 262, + 547, + 406 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 262, + 547, + 406 + ], + "type": "inline_equation", + "content": "IoU = 0.5" + }, + { + "bbox": [ + 304, + 262, + 547, + 406 + ], + "type": "text", + "content": ") score of 40.2 on the Charades-STA [21]. It surpasses the previous state-of-the-art (SOTA) MLLM specialized for temporal tasks, ChatVTG [59], and the expert model, UniVTG [47]. This demonstrates VideoChat-TPO's ability to accurately locate video moments corresponding to given text queries, thereby enhancing the practical applicability of MLLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 418, + 547, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 562 + ], + "type": "text", + "content": "Highlight Detection. Highlight detection generates salient scores for emphasizing frames based on the input language query. We compare the fine-tuning highlight detection performance of VideoChat-TPO with that of other expert models and MLLMs, as shown in Table 6. VideoChat-TPO notably outperforms the previous state-of-the-art MLLM (such as TimeChat [63]) on both Charades-STA and QVHighlight. For expert models like QD-DETR and UniVTG, VideoChat-TPO beats them on Charades-STA non-trivially and achieves a comparable performance on QVHighlight. This also demonstrates the progress in extending MLLMs to broad temporal tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 573, + 548, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 548, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 548, + 717 + ], + "type": "text", + "content": "Spatial Grounding. To verify the fine-grained localization ability of the model, we run the spatial grounding task which inputs textual descriptions into the model and gets the corresponding bounding box on RefCOCO [91]. As shown in Table 7, we compare VideoChat-TPO with pixel-to-sequence models, i.e., VisionLLM-H [75], pixel-to-embedding methods, i.e., NExT-Chat [97], and expert models, i.e., G-DINO [50]. VideoChat-TPO performs better than the pixel-to-embedding methods by using only a simple task head and achieves comparable performance to pixel-to-sequence models fine-tuned on a large amount of spatial grounding data, as well as to specialized models." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29885" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 345, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 345, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 345, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 345, + 148 + ], + "type": "table", + "html": "
ModelAcc@IoPAcc@GQAmIoPIoP@0.3IoP@0.5mIoUIoU@0.3IoU@0.5
VIOLETv2 [20]54.912.823.625.123.33.14.31.3
SeViLA [93]72.516.629.534.722.921.729.213.8
LangRepo [31]59.617.131.3-28.718.5-12.2
FrozenBiLM NG+ [88]73.817.524.228.523.79.613.56.1
VideoStreaming [58]57.417.832.2-31.019.3-13.3
LLoVi [98]65.924.337.3-36.920.0-15.3
HawkEye [79]-----25.737.019.5
VideoChat-TPO77.725.535.647.532.827.741.223.4
", + "image_path": "dffe59f164e7302a05a21b978b3137098655f9f5e87c8a166016e6c64aa51726.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 347, + 70, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 120, + 149, + 275, + 160 + ], + "lines": [ + { + "bbox": [ + 120, + 149, + 275, + 160 + ], + "spans": [ + { + "bbox": [ + 120, + 149, + 275, + 160 + ], + "type": "text", + "content": "Table 3. Performance on Grounded QA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 70, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 347, + 70, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 347, + 70, + 545, + 148 + ], + "type": "table", + "html": "
ModelMMIU [54]SEED2I [38]SEED2M [38]
LLaVA-v1.5 [48]19.258.339.2
ShareGPT4V [8]18.5--
OpenFlamingo [2]22.336.643.5
LLaVA-Interleave [40]32.4--
VideoChat2 [45]35.026.527.6
VideoChatGPT [52]-38.349.8
InternLM-XComposer [16]21.965.449.8
VideoChat-TPO40.2 (+5.2)67.3 (+40.8)70.0 (+42.4)
", + "image_path": "e57fe6e6565d9a2d55e8ca11e5b090b23a36ace39419386959fef58c2de47753.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 171, + 202, + 256 + ], + "blocks": [ + { + "bbox": [ + 353, + 149, + 536, + 160 + ], + "lines": [ + { + "bbox": [ + 353, + 149, + 536, + 160 + ], + "spans": [ + { + "bbox": [ + 353, + 149, + 536, + 160 + ], + "type": "text", + "content": "Table 4. Performance on Image Understanding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 171, + 202, + 256 + ], + "lines": [ + { + "bbox": [ + 52, + 171, + 202, + 256 + ], + "spans": [ + { + "bbox": [ + 52, + 171, + 202, + 256 + ], + "type": "table", + "html": "
ModelCharades-STA [21]
R@0.3R@0.5R@0.7mIoU
UniVTG [47]44.125.210.027.1
VideoChat2 [45]38.014.33.824.6
VTimeLLM [26]51.027.511.431.2
TimeChat [63]-32.213.4-
HawkEye [79]50.631.414.533.7
ChatVTG [59]52.733.015.934.9
VideoChat-TPO58.340.218.438.1
", + "image_path": "9f08653a17a3782dd31721fcf3ddbbcda8945fe6dea4bb53798be42a8ff77aa8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 203, + 172, + 419, + 256 + ], + "blocks": [ + { + "bbox": [ + 203, + 172, + 419, + 256 + ], + "lines": [ + { + "bbox": [ + 203, + 172, + 419, + 256 + ], + "spans": [ + { + "bbox": [ + 203, + 172, + 419, + 256 + ], + "type": "table", + "html": "
ModelCharades-STA [21]QVHighlight [36]
R@0.3R@0.5R@0.7mIoUmAPHIT@1
M-DETR [36]65.852.130.645.535.755.6
QD-DETR [56]-57.332.6-38.962.4
UniVTG [47]72.660.238.652.240.566.3
TimeChat [63]-46.723.7-21.737.9
HawkEye [79]72.558.328.8---
VideoChat-TPO77.065.040.755.038.665.4
", + "image_path": "a45b786cd89462262bcdb77ee4cdffc3b47947fdadeb30e36efc8616fbbdfb49.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 421, + 171, + 542, + 256 + ], + "blocks": [ + { + "bbox": [ + 421, + 171, + 542, + 256 + ], + "lines": [ + { + "bbox": [ + 421, + 171, + 542, + 256 + ], + "spans": [ + { + "bbox": [ + 421, + 171, + 542, + 256 + ], + "type": "table", + "html": "
MethodsRefCOCO [91]
valtestAtestB
MAntNet ★ [92]76.480.469.3
OFA-L [73]80.083.776.4
G-DINO-L ★ [50]90.693.288.2
VisionLLM-H [75]-86.7-
Shikra-7B [7]87.090.680.2
NExT-Chat-7B [97]85.590.077.9
VideoChat-TPO85.990.881.3
", + "image_path": "f5f8ad1cb55a8b866d964ae881f62850cf00a1366eafe793d75af0ba19035ab7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 50, + 258, + 416, + 280 + ], + "lines": [ + { + "bbox": [ + 50, + 258, + 416, + 280 + ], + "spans": [ + { + "bbox": [ + 50, + 258, + 416, + 280 + ], + "type": "text", + "content": "Table 5. Zero-Shot Performance on Table 6. Fine-tuning Performance on Moment Re-Moment Retrieval. Gray means no LLM. Retrieval and Highlight Detection. Gray means no LLM." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 420, + 258, + 542, + 280 + ], + "lines": [ + { + "bbox": [ + 420, + 258, + 542, + 280 + ], + "spans": [ + { + "bbox": [ + 420, + 258, + 542, + 280 + ], + "type": "text", + "content": "Table 7. Spatial Grounding Task. " + }, + { + "bbox": [ + 420, + 258, + 542, + 280 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 420, + 258, + 542, + 280 + ], + "type": "text", + "content": " with a refined decoder." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 301, + 289, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 289, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 289, + 458 + ], + "type": "text", + "content": "Tracking. In the tracking task, the model receives the object coordinates from the first frame and outputs the coordinates for the remaining frames in the video. We evaluate the VideoChat-TPO on the mainstream tracking benchmarks LAOST [18] and GOT-10k [27], as shown in Table 8. In zero-shot testing, VideoChat-TPO ranks best among all MLLMs, even surpassing some fine-tuned expert models notably with around " + }, + { + "bbox": [ + 46, + 301, + 289, + 458 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 301, + 289, + 458 + ], + "type": "text", + "content": " increase on success rate, such as SiamFC [4] and ATOM [13]. With the TPO training method, the model is optimized from mask sequences that hard to represent by words, allowing the MLLM to achieve strong motion characterization for multimodal generalization in highly dynamic and occluded scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 476, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 657 + ], + "type": "text", + "content": "Referring Segmentation. The referring segmentation task requires the model to output the mask sequence corresponding to the specified prompt. This capability is newly activated in VideoChat-TPO, as it is not supported or designed in the original VideoChat2 or SAM2. We compare the referring segmentation performance of VideoChat-TPO with other expert models and MLLMs in Table 9. The zero-shot capability of VideoChat-TPO is comparable to the fine-tuning capabilities of other expert models, i.e. ReferFormer [81], showing its notable open-world video segmentation performance. Through TPO training, MLLM effectively optimizes its understanding in both tracking and segmenting objects indicated by users. It further enables the model to excel in pixel-level tasks, offering perceptual advantages to it in practical applications like robotic control." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 669, + 148, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 148, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 148, + 680 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "In this section, we analyze the effectiveness of the key components of TPO and evaluate its scalability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 301, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 445 + ], + "type": "text", + "content": "Extending to Other MLLMs. To evaluate the effectiveness of the TPO method on different MLLMs, we apply TPO to LLaVA (LLaVA-OneVision [39]) in addition to VideoChat. In LLaVA, fine-grained video understanding is a weak spot. Despite the conversation data used in LLaVA-OneVision being very similar to, or even more extensive than, what we employ, there remains significant room for improvement in its performance on short-term fine-grained understanding tasks. As shown in Table 10, with TPO, LLaVA not only gains capabilities that it originally lacked but also achieves an " + }, + { + "bbox": [ + 304, + 301, + 545, + 445 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 445 + ], + "type": "text", + "content": " improvement on MVBench, demonstrating the universality of the TPO method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": "Task Preference Model vs. Textualized Task Data. We compare the TPO method with the approach of textualizing task data and training it in an autoregressive manner, which is a straightforward and efficient method for MLLMs to learn specific tasks. We use the same data as the TPO method and convert the data related to the task head into conversation format. Due to the limitations of textual representation, much data must be input in a more discrete manner. For instance, in tracking supervision data, we follow Merlin [90] by converting masks into a sequence of spatial coordinates to serve as both the model's input and output. As shown in Table 11, MLLM-TPO achieves a " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " performance improvement on MVBench [45] compared to MLLMs trained with textualized task data. This demonstrates that the task head in the TPO method enables more effective utilization of the original supervision signals compared to the next-token prediction approach, which inevitably incurs information loss when converting fine-grained tasks into textual annotations. The next-token prediction method cannot truly capture the nuanced information present in videos, while TPO allows the MLLM to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29886" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 236, + 150 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 236, + 150 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 236, + 150 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 236, + 150 + ], + "type": "table", + "html": "
ModelLaSOT [18]GOT-10k [27]
SuccessPnormPOverlapSR0.5SR0.75
SiamFC [4]33.642.033.934.835.39.8
ATOM [13]51.5--55.663.440.2
SiamRPN++ [37]49.656.949.151.861.832.5
SiamFC++ [86]54.462.354.759.569.547.9
LLaVA-1.5 [48]19.416.512.823.520.29.7
Merlin [90]39.840.238.151.455.942.8
VideoChat-TPO69.480.176.970.679.866.0
", + "image_path": "1ddd0901546927d15e10310821ff869a618f56fd2e7896e2d0c221823f736589.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 242, + 71, + 425, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 152, + 236, + 163 + ], + "lines": [ + { + "bbox": [ + 50, + 152, + 236, + 163 + ], + "spans": [ + { + "bbox": [ + 50, + 152, + 236, + 163 + ], + "type": "text", + "content": "Table 8. Performance on Tracking Benchmarks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 242, + 71, + 425, + 137 + ], + "lines": [ + { + "bbox": [ + 242, + 71, + 425, + 137 + ], + "spans": [ + { + "bbox": [ + 242, + 71, + 425, + 137 + ], + "type": "table", + "html": "
MethodRef-YouTube-VOS [65]MeViS [15]
J&FJFJ&FJF
ReferFormer [81]62.961.364.631.029.832.2
OnlineRefer [80]62.961.064.7---
LISA [35]52.652.153.0---
VideoLISA [3]63.761.763.744.441.347.6
VideoChat-TPO63.952.375.447.042.651.3
", + "image_path": "cf6b2ab84f6c59b2891f49da02576b58175b55067db9404c2cf2bad48887f963.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 433, + 71, + 543, + 109 + ], + "blocks": [ + { + "bbox": [ + 238, + 141, + 428, + 163 + ], + "lines": [ + { + "bbox": [ + 238, + 141, + 428, + 163 + ], + "spans": [ + { + "bbox": [ + 238, + 141, + 428, + 163 + ], + "type": "text", + "content": "Table 9. Performance on Referring Segmentation. TPO is evaluated in a zero-shot on Ref-YouTube." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 433, + 71, + 543, + 109 + ], + "lines": [ + { + "bbox": [ + 433, + 71, + 543, + 109 + ], + "spans": [ + { + "bbox": [ + 433, + 71, + 543, + 109 + ], + "type": "table", + "html": "
ModelMVBench
LLaVA-OV [39]56.7
LLaVA-OV-TPO64.8 (+8.1)
", + "image_path": "f81727b7e4fc0a473de5db5e4faacd6339b3157aa9b41692f0898d58ab69588c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 183, + 285, + 247 + ], + "blocks": [ + { + "bbox": [ + 429, + 119, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 429, + 119, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 429, + 119, + 545, + 163 + ], + "type": "text", + "content": "Table 10. Applying TPO on LLaVA-OV. TPO significantly enhances the ability to understand fine-grained video details." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 183, + 285, + 247 + ], + "lines": [ + { + "bbox": [ + 50, + 183, + 285, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 183, + 285, + 247 + ], + "type": "table", + "html": "
ModelCharades-STA [21]MVBench
R@0.3R@0.5R@0.7mIoUAVG
VideoChat-TPO58.340.218.438.166.8
w/o reasoning data56.438.317.135.666.1
replace by simple head31.517.86.115.465.8
textualized task data33.718.66.216.064.7
", + "image_path": "795164e10745c8629936ac36c5ff534ecefbb76493b2acd509e4a4c4e85e4d53.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 67, + 271, + 266, + 354 + ], + "blocks": [ + { + "bbox": [ + 49, + 249, + 286, + 259 + ], + "lines": [ + { + "bbox": [ + 49, + 249, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 249, + 286, + 259 + ], + "type": "text", + "content": "Table 11. Ablation of Reasoning Data and Head Performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 271, + 266, + 354 + ], + "lines": [ + { + "bbox": [ + 67, + 271, + 266, + 354 + ], + "spans": [ + { + "bbox": [ + 67, + 271, + 266, + 354 + ], + "type": "table", + "html": "
Exp.TRMCR@0.5Acc@0.5J&F
1✓30.2--
2✓-77.3-
3✓--55.1
4✓✓-80.258.3
5✓✓✓36.781.661.4
6✓✓✓✓40.282.063.9
", + "image_path": "707067be93d98c4638e5e8b0e154a301cf638a00823d5f4960acc0a3689db5fa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 356, + 287, + 409 + ], + "lines": [ + { + "bbox": [ + 47, + 356, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 287, + 409 + ], + "type": "text", + "content": "Table 12. Impact of TPO Components and Data. T, R, M, and C denote temporal head, region head, mask head, and conversation data respectively. R1@0.5 means R1@0.5 in Charades-STA, Acc@0.5 represents the mean of Acc@0.5 in all COCO datasets, " + }, + { + "bbox": [ + 47, + 356, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 47, + 356, + 287, + 409 + ], + "type": "text", + "content": " means " + }, + { + "bbox": [ + 47, + 356, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 47, + 356, + 287, + 409 + ], + "type": "text", + "content": " in Ref-YouTube-VOS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 424, + 276, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 276, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 276, + 436 + ], + "type": "text", + "content": "learn perceptual information beyond conversational data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": "Impact of Task Head Performance on TPO. A stronger task head is expected to enhance the multimodal capabilities of TPO more effectively than a weaker one. To further explore the impact of the task head on model performance, we replaced the temporal head from CG-DETR [55] with a simple two-layer MLP. As shown in Table 11, the simple temporal head exhibits a significant performance decline compared to CG-DETR on the corresponding tasks. However, it still achieves a " + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": " improvement compared to the model without TPO. In contrast to the simpler head, the well-designed head can provide more accurate expert knowledge through its architecture and pre-trained task preference weights, enabling better utilization of the data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Impact of Data Scaling. Concerning foundation or large models, we expect to enhance their model performance by incorporating additional data at scale. In our approach, we expand the existing conversation data by integrating two reasoning datasets [5, 14] with temporal information. By incorporating these datasets, we observe a notable enhancement in the capabilities of the temporal head in Table 11. Furthermore, as shown in Table 12, adding conversation data further improves performance across various" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 185, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 232 + ], + "type": "text", + "content": "tasks. Overall, increasing the amount of conversation and fine-grained task data leads to noticeable improvements in model performance, demonstrating the effectiveness of data scaling in training multimodal models with TPO." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 243, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 338 + ], + "type": "text", + "content": "Synergistic Gains from Co-training. Table 12 presents the impact of incorporating different visual tasks into TPO on performance. This indicates that the collaborative learning of visual tasks facilitates the transfer of knowledge, ultimately leading to better performance across the board. Overall, these results confirm that co-training not only benefits individual tasks but also creates a synergistic effect that enhances overall capabilities in visual understanding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 351, + 383, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 383, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 383, + 363 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 369, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 572 + ], + "type": "text", + "content": "This paper introduces Task Preference Optimization (TPO). It enhances the overall multimodal performance of MLLMs by enhancing their precise visual understanding. TPO achieves this by integrating differentiable task preferences derived from fine-grained visual tasks. It introduces learnable task tokens that serve as bridges between multiple task-specific heads and the MLLM. Through the joint optimization of these task tokens, heads, and the MLLM, TPO leads to a substantial improvement in multimodal dialogue capabilities and performance across various visual tasks. Our results demonstrate the effectiveness of TPO in scaling MLLMs with task-specific data and seamlessly integrating them with existing expert vision models. We believe this study clarifies the prerequisites for fusing MLLMs with models and data from the pre-large model era, effectively bridging the gap between expert and generalist models, as well as between generation and understanding." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Limitations. Currently, TPO focuses exclusively on discriminative visual tasks, overlooking generative ones. Meanwhile, the framework is supervised by human annotations, which neglects potentially valuable unsupervised or self-supervised learning approaches such as contrastive learning [10, 23]. This limitation inherently restricts the scalability of TPO in terms of both task diversity and requirements. While we demonstrate TPO's potential for enhancing MLLMs through increased task-specific data, a comprehensive investigation into the broader impact of this scaling remains a crucial area for future research." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29887" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 90, + 287, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 90, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 47, + 90, + 287, + 147 + ], + "type": "text", + "content": "This work is partially supported by the National Key R&D Program of China (No. 2022ZD0160102), Jiangsu Frontier Technology R&D Program (No. BF2024076), and the Science and Technology Commission of Shanghai Municipality under Grant No. 23YF1461900." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 167, + 107, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 167, + 107, + 180 + ], + "spans": [ + { + "bbox": [ + 48, + 167, + 107, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 188, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 58, + 188, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 188, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 287, + 243 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 244, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 244, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 58, + 244, + 287, + 330 + ], + "type": "text", + "content": "[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortman, and Ludwig Schmidt. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 334, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 334, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 58, + 334, + 287, + 387 + ], + "type": "text", + "content": "[3] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 389, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 389, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 58, + 389, + 287, + 432 + ], + "type": "text", + "content": "[4] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In ECCV, pages 850–865, 2016. 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 434, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 434, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 58, + 434, + 287, + 477 + ], + "type": "text", + "content": "[5] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, pages 961-970, 2015. 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 479, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 479, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 58, + 479, + 287, + 533 + ], + "type": "text", + "content": "[6] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv preprint arXiv:2304.02948, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 58, + 535, + 287, + 578 + ], + "type": "text", + "content": "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 580, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 580, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 58, + 580, + 287, + 622 + ], + "type": "text", + "content": "[8] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In ECCV, 2024. 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 624, + 287, + 689 + ], + "type": "text", + "content": "[9] Lin Chen, Xin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, Li Yuan, Yu Qiao, Dahua Lin, Feng Zhao, and Jiaqi Wang. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learn" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 544, + 95 + ], + "type": "text", + "content": "ing of visual representations. In ICML, pages 1597-1607. PMLR, 2020. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 97, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 97, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 97, + 545, + 150 + ], + "type": "text", + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 185 + ], + "type": "text", + "content": "[12] Yiming Cui, Ziqing Yang, and Xin Yao. Efficient and effective text encoding for chinese llama and alpaca. arXiv preprint arXiv:2304.08177, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 545, + 219 + ], + "type": "text", + "content": "[13] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In CVPR, pages 4660-4669, 2019. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 221, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 221, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 312, + 221, + 545, + 263 + ], + "type": "text", + "content": "[14] Pradipto Das, Chenliang Xu, Richard F Doell, and Jason J Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In CVPR, pages 2634-2641, 2013. 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 266, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 266, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 266, + 545, + 308 + ], + "type": "text", + "content": "[15] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 311, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 311, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 312, + 311, + 545, + 376 + ], + "type": "text", + "content": "[16] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 377, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 377, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 312, + 377, + 545, + 421 + ], + "type": "text", + "content": "[17] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 423, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 423, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 312, + 423, + 545, + 475 + ], + "type": "text", + "content": "[18] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, pages 5374-5383, 2019. 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 479, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 479, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 312, + 479, + 545, + 543 + ], + "type": "text", + "content": "[19] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 545, + 589 + ], + "type": "text", + "content": "[20] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. An empirical study of end-to-end video-language transformers with masked visual modeling. In CVPR, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 591, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 591, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 312, + 591, + 545, + 623 + ], + "type": "text", + "content": "[21] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 625, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 625, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 312, + 625, + 545, + 678 + ], + "type": "text", + "content": "[22] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, pages 18995-19012, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29888" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 106 + ], + "type": "text", + "content": "[24] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "text", + "content": "[25] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 288, + 185 + ], + "type": "text", + "content": "[26] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 219 + ], + "type": "text", + "content": "[27] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 43(5):1562-1577, 2019. 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 274 + ], + "type": "text", + "content": "[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 308 + ], + "type": "text", + "content": "[29] Fan Jiang and Zilei Wang. Sparse sharing relation network for panoptic driving perception. In ACMMM, pages 800-808, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 310, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 288, + 354 + ], + "type": "text", + "content": "[30] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In CVPR, pages 13700-13710, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 355, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 355, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 53, + 355, + 288, + 398 + ], + "type": "text", + "content": "[31] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv preprint arXiv:2403.14622, 2024.7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 399, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 443 + ], + "type": "text", + "content": "[32] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 445, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 288, + 488 + ], + "type": "text", + "content": "[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 490, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 288, + 544 + ], + "type": "text", + "content": "[34] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 545, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 288, + 589 + ], + "type": "text", + "content": "[35] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, pages 9579-9589, 2024. 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 590, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 288, + 623 + ], + "type": "text", + "content": "[36] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. NeurIPS, 34:11846-11858, 2021. 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 624, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 288, + 667 + ], + "type": "text", + "content": "[37] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. SiamRPN++: Evolution of siamese visual tracking with very deep networks. In CVPR, pages 4282-4291, 2019. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 288, + 713 + ], + "type": "text", + "content": "[38] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In CVPR, pages 13299-13308, 2024. 2, 5, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 312, + 72, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 117 + ], + "type": "text", + "content": "[39] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 162 + ], + "type": "text", + "content": "[40] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 545, + 217 + ], + "type": "text", + "content": "[41] Hao Li, Jinguo Zhu, Xiaohu Jiang, Xizhou Zhu, Hongsheng Li, Chun Yuan, Xiaohua Wang, Yu Qiao, Xiaogang Wang, Wenhai Wang, et al. Uni-perceiver v2: A generalist model for large-scale vision and vision-language tasks. In CVPR, pages 2691-2700, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 219, + 545, + 262 + ], + "type": "text", + "content": "[42] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "type": "text", + "content": "[43] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 309, + 545, + 352 + ], + "type": "text", + "content": "[44] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 545, + 408 + ], + "type": "text", + "content": "[45] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 312, + 410, + 545, + 453 + ], + "type": "text", + "content": "[46] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 455, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 455, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 312, + 455, + 545, + 509 + ], + "type": "text", + "content": "[47] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 511, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 511, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 511, + 545, + 533 + ], + "type": "text", + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 312, + 535, + 545, + 567 + ], + "type": "text", + "content": "[49] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In ECCV, 2024. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 568, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 568, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 312, + 568, + 545, + 622 + ], + "type": "text", + "content": "[50] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In ECCV, 2024. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "type": "text", + "content": "[51] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 669, + 545, + 713 + ], + "type": "text", + "content": "[52] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 3, 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29889" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "text", + "content": "[53] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 173 + ], + "type": "text", + "content": "[54] Fanqing Meng, Jin Wang, Chuanhao Li, Quanfeng Lu, Hao Tian, Jiaqi Liao, Xizhou Zhu, Jifeng Dai, Yu Qiao, Ping Luo, et al. Mmiu: Multimodal multi-image understanding for evaluating large vision-language models. arXiv preprint arXiv:2408.02718, 2024. 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 287, + 218 + ], + "type": "text", + "content": "[55] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "type": "text", + "content": "[56] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "text", + "content": "[57] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP, pages 2265-2269. IEEE, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 309, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 309, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 53, + 309, + 287, + 353 + ], + "type": "text", + "content": "[58] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuangrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009, 2024.7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "text", + "content": "[59] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPR, pages 1847-1856, 2024. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 442 + ], + "type": "text", + "content": "[60] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 36, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 445, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 287, + 521 + ], + "type": "text", + "content": "[61] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 287, + 566 + ], + "type": "text", + "content": "[62] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics, 1:25-36, 2013. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 287, + 610 + ], + "type": "text", + "content": "[63] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. CVPR, abs/2312.02051, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 287, + 644 + ], + "type": "text", + "content": "[64] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "text", + "content": "[65] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223. Springer, 2020. 5, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "text", + "content": "[66] Jing Shao, Siyu Chen, Yangguang Li, Kun Wang, Zhenfei Yin, Yinan He, Jianing Teng, Qinghong Sun, Mengya Gao," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "text", + "content": "Jihao Liu, et al. Intern: A new learning paradigm towards general vision. arXiv preprint arXiv:2111.08687, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 106, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 106, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 106, + 545, + 160 + ], + "type": "text", + "content": "[67] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 162, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 216 + ], + "type": "text", + "content": "[68] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "type": "text", + "content": "[69] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, pages 10078-10093, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 545, + 316 + ], + "type": "text", + "content": "[70] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 317, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 317, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 312, + 317, + 545, + 360 + ], + "type": "text", + "content": "[71] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, pages 20-36. Springer, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 545, + 403 + ], + "type": "text", + "content": "[72] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 545, + 459 + ], + "type": "text", + "content": "[73] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, pages 23318-23340. PMLR, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "text", + "content": "[74] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, pages 6312-6322, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 545, + 568 + ], + "type": "text", + "content": "[75] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2024. 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 545, + 624 + ], + "type": "text", + "content": "[76] Weiyun Wang, Yiming Ren, Haowen Luo, Tiantong Li, Chenxiang Yan, Zhe Chen, Wenhai Wang, Qingyun Li, Lewei Lu, Xizhou Zhu, et al. The all-seeing project v2: Towards general relation comprehension of the open world. arXiv preprint arXiv:2402.19474, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 625, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 625, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 625, + 545, + 667 + ], + "type": "text", + "content": "[77] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In ICCV, pages 1130-1140, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "text", + "content": "[78] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. In ECCV, 2024. 2, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29890" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[79] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 162 + ], + "type": "text", + "content": "[80] Dongming Wu, Tiancai Wang, Yang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In ICCV, pages 2761-2770, 2023. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 163, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 287, + 196 + ], + "type": "text", + "content": "[81] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 286, + 262 + ], + "type": "text", + "content": "[82] Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Wenhai Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. arXiv preprint arXiv:2406.08394, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 286, + 297 + ], + "type": "text", + "content": "[83] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In CVPR, 2024. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 298, + 286, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 286, + 353 + ], + "type": "text", + "content": "[84] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pretraining for zero-shot video-text understanding. In EMNLP, pages 6787-6800, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 354, + 286, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 286, + 398 + ], + "type": "text", + "content": "[85] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 286, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 286, + 432 + ], + "type": "text", + "content": "[86] Yinda Xu et al. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, pages 140-148, 2020. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 434, + 286, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 286, + 476 + ], + "type": "text", + "content": "[87] Shen Yan, Tao Zhu, Zirui Wang, Yuan Cao, Mi Zhang, Soham Ghosh, Yonghui Wu, and Jiahui Yu. Video-text modeling with zero-shot transfer from contrastive captioners. ArXiv, abs/2212.04979, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 478, + 286, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 286, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 286, + 511 + ], + "type": "text", + "content": "[88] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. In NeurIPS, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 513, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 513, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 513, + 286, + 578 + ], + "type": "text", + "content": "[89] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chaoya Jiang, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 580, + 286, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 580, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 580, + 286, + 634 + ], + "type": "text", + "content": "[90] En Yu, Liang Zhao, Yana Wei, Jinrong Yang, Dongming Wu, Lingyu Kong, Haoran Wei, Tiancai Wang, Zheng Ge, Xiangyu Zhang, et al. Merlin: Empowering multimodal llms with foresight minds. In ECCV, pages 425-443. Springer, 2025. 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 635, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 286, + 668 + ], + "type": "text", + "content": "[91] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 286, + 713 + ], + "type": "text", + "content": "[92] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In CVPR, pages 1307-1315, 2018. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[93] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In NeurIPS, 2023. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "type": "text", + "content": "[94] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In CVPR, pages 13807-13816, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "text", + "content": "[95] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In CVPR, pages 23056-23065, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "type": "text", + "content": "[96] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In ICLR. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 545, + 293 + ], + "type": "text", + "content": "[97] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An Imm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 294, + 545, + 338 + ], + "type": "text", + "content": "[98] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lllm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 339, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 339, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 339, + 545, + 371 + ], + "type": "text", + "content": "[99] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, pages 543–553, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 372, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 426 + ], + "type": "text", + "content": "[100] Ruohong Zhang, Liangke Gui, Zhiqing Sun, Yihao Feng, Keyang Xu, Yuanhan Zhang, Di Fu, Chunyuan Li, Alexander Hauptmann, Yonatan Bisk, et al. Direct preference optimization of video large multimodal models from language model reward. arXiv preprint arXiv:2404.01258, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "text", + "content": "[101] Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "text", + "content": "[102] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "text", + "content": "[103] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In ICML, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "text", + "content": "[104] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 678 + ], + "type": "text", + "content": "[105] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. In ICLR Workshop on Reliable and Responsible Foundation Models, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[106] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29891" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 150 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "text", + "content": "conditional moes. In NeurIPS, pages 2664-2678, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "type": "text", + "content": "[107] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pretraining unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 2, 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29892" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_content_list.json b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..53c24d4309f5801aedba5e54fe164e8e719d914e --- /dev/null +++ b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_content_list.json @@ -0,0 +1,1738 @@ +[ + { + "type": "text", + "text": "Task Singular Vectors: Reducing Task Interference in Model Merging", + "text_level": 1, + "bbox": [ + 261, + 130, + 733, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Antonio Andrea Gargiulo\\* Simone Scardapane", + "bbox": [ + 187, + 198, + 398, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Donato Crisostomi* \nFabrizio Silvestri", + "bbox": [ + 433, + 198, + 591, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maria Sofia Bucarelli Emanuele Rodolà", + "bbox": [ + 632, + 198, + 805, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sapienza University of Rome", + "bbox": [ + 377, + 246, + 612, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gargiulo.1769185@studenti.uniroma1.it {crisostomi, rodola}@di.uniroma1.it simone.scardapane@uniroma1.it {bucarelli, fsilvestri}@diag.uniroma1.it", + "bbox": [ + 166, + 268, + 828, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 337, + 326, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task Arithmetic has emerged as a simple yet effective method to merge models without additional training. However, by treating entire networks as flat parameter vectors, it overlooks key structural information and is susceptible to task interference. In this paper, we study task vectors at the layer level, focusing on task layer matrices and their singular value decomposition. In particular, we concentrate on the resulting singular vectors, which we refer to as Task Singular Vectors (TSV). Recognizing that layer task matrices are often low-rank, we propose TSV-Compress (TSV-C), a simple procedure that compresses them to $10\\%$ of their original size while retaining $99\\%$ of accuracy. We further leverage this low-rank space to define a new measure of task interference based on the interaction of singular vectors from different tasks. Building on these findings, we introduce TSV-Merge (TSV-M), a novel model merging approach that combines compression with interference reduction, significantly outperforming existing methods.", + "bbox": [ + 88, + 369, + 485, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 670, + 220, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The widespread availability of pre-trained models and public repositories has driven the development of techniques for efficiently combining and reusing existing models. Among these techniques, model merging approaches enable the creation of multi-task models without additional training. One popular approach, Task Arithmetic (TA) [22], stands out for its simplicity and effectiveness. However, by treating entire networks as high-dimensional vectors, TA and subsequent works overlook crucial structural information [9, 10, 24, 32, 38, 51, 53]. This flattened view limits these approaches to coarse-grained measures like cosine similarity for assessing inter-task interactions.", + "bbox": [ + 89, + 695, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a112e20c2a5876d992350b71032a1805301920526a934c81d240c758be68c1b7.jpg", + "image_caption": [ + "Figure 1. Mean accuracy of a ViT-L-14 merged over 8, 14, and 20 tasks respectively. By significantly surpassing existing methods, TSV-M establishes the new state of the art in model merging." + ], + "image_footnote": [], + "bbox": [ + 517, + 338, + 903, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we instead focus on preserving and leveraging the natural structure of neural networks by examining weight differences at the layer level while retaining their matrix form wherever possible. Our approach begins by examining per-layer task matrices, which represent weight changes for each task, through singular value decomposition (SVD). This decomposition yields singular vectors and singular values that capture the most significant directions of variation within each layer. We term these singular vectors Task Singular Vectors (TSV), as they provide an interpretable basis for assessing task-specific contributions at each layer. Importantly, in accordance with recent literature on PEFT (cfr. Sec. 2), our analysis confirms that task matrices are inherently low-rank, meaning that only a small portion of TSVs is sufficient to represent the layer's function with high fidelity.", + "bbox": [ + 511, + 568, + 906, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building on this insight, we introduce TSV-Compress (TSV-C), a simple yet effective procedure to compress task vectors down to $10\\%$ of their original size while maintaining $99\\%$ of the original accuracy. Focusing on only a fraction of the most relevant TSVs for each task, TSV-Compress enables efficient storage and processing", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\star$ denotes equal contribution.", + "bbox": [ + 114, + 887, + 271, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18695", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of task-specific information without waiving performance.", + "bbox": [ + 89, + 90, + 477, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond compression, examining the interplay of TSVs across different tasks provides a geometrically informed framework for analyzing task interference at the individual layer level. By assessing how singular vectors from different tasks align or diverge within each layer, this approach offers a significantly more fine-grained understanding of inter-task interactions, going beyond global vector similarity metrics like cosine similarity. Building upon these principles, we introduce TSV-Merge (TSV-M), a novel model merging method that combines compression and task interference reduction. This is achieved by discarding the irrelevant singular vectors for each task and then reducing their inter-task interference with a whitening transformation over their similarity matrix. We empirically show our approach to effectively reduce task interference, and the reduction to be complementary to the compression step. We then evaluate our approach across several benchmarks and show it outperforms existing methods by an average of $\\sim 15\\%$ accuracy points, establishing the new state of the art by a large margin. We release all our code for research purposes1. In summary, our contribution is 4-fold:", + "bbox": [ + 91, + 107, + 483, + 424 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We study the singular value decomposition of per-layer task matrices, showing them to be low-rank and measuring the interplay of singular vectors from different tasks;", + "- We introduce TSV-C, a method that builds on this insight to compress task vectors by a factor of $10 \\times$ while preserving $99\\%$ of their original accuracy;", + "- We propose TSV-M, a model merging technique that complements compression with a task interference reduction step by applying a whitening transformation to decorrelate singular vectors across tasks;", + "- We conduct extensive experiments on multiple computer vision datasets, showing that TSV-M significantly outperforms existing model merging methods and provide in-depth analyses to uncover the underlying factors contributing to its success." + ], + "bbox": [ + 89, + 425, + 482, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 667, + 232, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Merging offers an efficient alternative to ensembling by combining existing models without the need for further training. Several approaches address this by determining the neuron permutations that align the models into a shared optimization basin, after which the models are combined using a straightforward averaging method. [1, 8, 25, 36, 43]. An alternative line of approaches focuses on the multi-task scenario, where a single pre-trained model is fine-tuned for different tasks [10, 22, 32, 40, 46, 48, 51, 53, 54]. The prerequisites for methods in this category are outlined in Table 1. Task Arithmetic [22] introduces the concept of task vectors, which are the weight differences", + "bbox": [ + 89, + 693, + 482, + 875 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/fbba5692ceb238ff5ab11ebd1035075fe388c609b6693548532c68ecbbb7219f.jpg", + "table_caption": [ + "Table 1. Model merging approaches and their requirements." + ], + "table_footnote": [], + "table_body": "
MethodAdditional trainingExtra storageValidat. data inputsdata labels
Weight Avg. [48]××××
Fisher-Merg. [32]××✓×
RegMean [24]××✓×
EMR-Merging [20]×✓××
TALL-Mask [46]×✓××
TSV-C (Ours)×✓××
Task Arith. [22]××✓✓
Ties-Merging [51]××✓✓
AdaMerging [52]✓×✓×
Consensus-TA [46]××✓✓
TSV-M (Ours)××××
", + "bbox": [ + 506, + 114, + 908, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "between fine-tuned models and the pre-trained base model. By averaging these vectors, a merged multi-task model can be created; conversely, negating a task vector allows for forgetting a specific task. TIES [51] addresses redundancies in model parameters by first selecting the top- $k$ most significant parameter changes and then constructing a sign vector based on the majority sign across all models. The latter is used to merge the task vectors disjointly, meaning the average is not computed when a parameter is zero or when parameters disagree in sign. Similarly, DARE [53] randomly resets redundant parameters to their pre-trained values and rescales the remaining parameters by a factor proportional to the dropped ones, aiming to reduce interference among tasks. Fisher Merging [32] and RegMean [24] merge models by performing weighted averaging, utilizing the Fisher information matrix and the inner product of input vectors, respectively. Model Broadcrumbs [10] focuses on merging only significant weights by discarding outliers and both minor and large perturbations in the fine-tuned parameters. More recently, Wang et al. [46] observed that tasks often utilize non-overlapping sets of weights. They propose two methods: the first, TALL-Mask, uses binary masks to activate important task-specific weights, requiring extra storage (as does our method TSV-C) for the masks. The second, Consensus Merging, leverages these masks to remove parameters that are important to less than two tasks. Like our approach, TwinMerging [31] and SMILE [44] apply SVD to layer task matrices. However, neglecting singular vector interference, they require a router to selectively activate a subset during inference.", + "bbox": [ + 511, + 356, + 906, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike existing approaches, our methods explicitly address task interference by leveraging the geometric structure of singular vectors to minimize interactions between task-specific parameters. Rather than averaging or selectively merging parameters as prior techniques do, we decorrelate singular vectors to directly reduce interference across tasks.", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/AntoAndGar/task_singular_vectors", + "bbox": [ + 107, + 886, + 400, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18696", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we achieve targeted compression by isolating each task's unique components, storing only essential parts, and ensuring minimal interference, a step beyond traditional SVD-based compression approaches.", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SVD for Model Compression A significant body of work explores low-rank decompositions for fully-connected and convolutional layers [11, 12, 16, 23, 27], as well as tensor decompositions [12, 16, 26]. These approaches typically achieve a layer-by-layer factorization of a trained network, focusing on minimizing the difference between the original and the low-rank approximated weight matrices. This is accomplished using techniques like SVD [12, 16, 23, 27] or iterative optimization [23, 26]. While other methods such as weight quantization, knowledge distillation, and pruning exist, we utilize SVD not only for low-rank approximation and compression but also because its matrix decomposition properties make it particularly effective in analyzing and mitigating interference.", + "bbox": [ + 89, + 178, + 483, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Task Singular Vectors", + "text_level": 1, + "bbox": [ + 89, + 407, + 302, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce key concepts for understanding our approach, including a new measure of task interference that is central to our method.", + "bbox": [ + 89, + 433, + 485, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 89, + 493, + 223, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We build on Task Arithmetic (TA), which defines task vectors capturing the differences in model weights for individual tasks. Formally, the weights $\\theta_{\\mathrm{MT}}$ of a multi-task model for $T$ tasks are computed by aggregating the task-specific weight differences, or task vectors, as follows:", + "bbox": [ + 89, + 516, + 483, + 594 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathrm {M T}} = \\theta_ {\\mathrm {p r e}} + \\alpha \\frac {\\sum_ {i = 1} ^ {T} \\tau_ {i}}{T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 604, + 483, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{pre}}$ is the set of pretrained model weights, $\\alpha$ is a scaling factor, and $\\tau_{i} = \\theta_{\\mathrm{ft}_{i}} - \\theta_{\\mathrm{pre}}$ is the task vector for task $i$ , with $\\theta_{\\mathrm{ft}_i}$ being the fine-tuned weights for the task. Differently from TA, however, we consider these operations at the layer level. From this perspective, Eq. (1) becomes", + "bbox": [ + 89, + 651, + 483, + 729 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathrm {M T}} ^ {(l)} = \\theta_ {\\text {p r e}} ^ {(l)} + \\alpha \\frac {\\sum_ {i = 1} ^ {T} \\Delta_ {i} ^ {(l)}}{T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 739, + 482, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{pre}}^{(l)}$ encodes the pretrained weights for layer $l$ , and $\\Delta_i^{(l)} = \\theta_{\\mathrm{ft}_i}^{(l)} - \\theta_{\\mathrm{pre}}^{(l)}$ is the task-specific weight difference for task $i$ at layer $l$ . When layer $l$ has a matrix structure, we call its corresponding $\\Delta_i^l$ the per-layer task matrix for task $i$ . When this is not the case, our framework defaults to standard TA. For brevity, we will generally omit the layer index and refer to the layer- $l$ task matrix $\\Delta_i^l$ as $\\Delta_i$ .", + "bbox": [ + 89, + 789, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decomposing layer task matrices", + "text_level": 1, + "bbox": [ + 513, + 90, + 750, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Treating layer-wise weights as structured entities rather than flattened vectors enables us to analyze their SVD, revealing low-rank properties and deeper insight into the inter-task interactions. Given two tasks $i,j$ , we consider the SVD of their task matrices $\\Delta_{i}$ and $\\Delta_{j}$ at a generic layer:", + "bbox": [ + 511, + 109, + 906, + 186 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {i} = U _ {i} \\Sigma_ {i} V _ {i} ^ {\\top}, \\qquad \\Delta_ {j} = U _ {j} \\Sigma_ {j} V _ {j} ^ {\\top}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 194, + 834, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $U_{i}, U_{j}$ and $V_{i}, V_{j}$ are the matrices of left and right singular vectors respectively, and $\\Sigma_{i}, \\Sigma_{j}$ are diagonal matrices of singular values. Due to their role in assessing task-specific contributions, we term the obtained singular vectors Task Singular Vectors (TSV).", + "bbox": [ + 511, + 223, + 906, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Importantly for our treatment, we can equivalently write the aggregated task matrices in Eq. (2) in terms of their singular vectors and values. By defining $U = [U_{1}\\dots U_{T}]$ as the column-wise concatenation of all the left TSVs, $V = [V_{1}\\dots V_{T}]$ as the row-wise concatenation of the right TSVs, and $\\Sigma$ as the block diagonal matrix with $\\{\\Sigma_i\\}_{i=1}^T$ along its diagonal, we can write:", + "bbox": [ + 511, + 299, + 906, + 405 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM = U \\Sigma V ^ {\\top} = \\sum_ {i = 1} ^ {T} U _ {i} \\Sigma_ {i} V _ {i} ^ {\\top} = \\sum_ {i = 1} ^ {T} \\Delta_ {i}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 415, + 906, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the simple case where $T = 2$ , $M$ would be given by:", + "bbox": [ + 511, + 465, + 880, + 483 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM = \\left[ \\begin{array}{c c} U _ {1} & U _ {2} \\end{array} \\right] \\left[ \\begin{array}{c c} \\Sigma_ {1} & 0 \\\\ 0 & \\Sigma_ {2} \\end{array} \\right] \\left[ \\begin{array}{c} V _ {1} ^ {\\top} \\\\ V _ {2} ^ {\\top} \\end{array} \\right] = \\Delta_ {1} + \\Delta_ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 489, + 893, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When concatenating singular components from different tasks, we violate certain properties of the SVD. Specifically, the matrices $U$ and $V$ become non-orthogonal because singular vectors from different tasks may overlap. Additionally, the singular values $\\Sigma_{i}$ from different tasks can vary significantly in magnitude, which may bias the merging process toward tasks with larger singular values.", + "bbox": [ + 511, + 534, + 906, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Low-rank nature of layer task matrices", + "text_level": 1, + "bbox": [ + 513, + 648, + 852, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start our analysis by studying the low-rank properties of per-layer task matrices. Interpreting SVD as a sum of rank-one matrices, for each task $i$ , by Eckart-Young's theorem [14] we get the best approximation (in Frobenius norm) of each task matrix $\\Delta_{i}$ by retaining only the top- $k$ singular values and their corresponding vectors:", + "bbox": [ + 511, + 671, + 908, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\Delta} _ {i} = \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \\top}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 772, + 906, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we show in Fig. 2, we find task matrices to be inherently low-rank: a small subset of TSVs is sufficient to represent the layer's function with high fidelity. In particular, even when preserving only $3\\%$ of the singular components per task, the mean accuracy drops by merely $1.5\\%$ . This is", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18697", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1842bab04fc51cc70444b101f9ef121909f45317df0acb0cf944e33b85c2ec50.jpg", + "image_caption": [ + "Figure 2. Mean absolute accuracy of the ViT-B-32 model across increasing fractions of retained singular components, averaged over 20 tasks. The red line represents the average accuracy of the original fine-tuned models with full-rank task matrices, while the green line shows the accuracies using low-rank approximations." + ], + "image_footnote": [], + "bbox": [ + 106, + 98, + 472, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noticeable considering that $97\\%$ of singular components in each layer matrix are discarded. Building on this insight, we propose in Sec. 4.1 a compression algorithm that maintains $99\\%$ of the accuracy while shrinking the task vectors to $10\\%$ of their original size.", + "bbox": [ + 89, + 344, + 483, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given this low-rank structure, it is natural to aggregate task matrices within their subspaces. We, therefore, obtain a reduced version of the aggregation matrix $M$ (Eq. (3)) as:", + "bbox": [ + 89, + 419, + 483, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} = \\sum_ {i = 1} ^ {T} \\hat {U} _ {i} \\hat {\\Sigma} _ {i} \\hat {V} _ {i} ^ {\\top} = \\sum_ {i = 1} ^ {T} \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \\top}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 476, + 482, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{U}_i$ and $\\hat{V}_i$ contain the top- $k$ left and right singular vectors for task $i$ , respectively, and $\\hat{\\Sigma}_i$ is the diagonal matrix of the top- $k$ singular values $\\sigma_j^i$ . This formulation highlights that the low-rank approximation $\\hat{M}$ utilizes the most relevant singular components from each task to effectively combine the task-specific weight differences.", + "bbox": [ + 89, + 523, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Singular Task interference", + "text_level": 1, + "bbox": [ + 89, + 626, + 331, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We hereby introduce a score of task interference based on the interplay of TSVs from different tasks, which we term Singular Task Interference (STI)", + "bbox": [ + 89, + 648, + 483, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S T I} \\left(\\left\\{\\Delta_ {i} \\right\\} _ {i = 1} ^ {T}\\right) = \\left\\| \\left(U ^ {\\top} U - I\\right) \\Sigma \\left(V ^ {\\top} V - I\\right) \\right\\| _ {1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 700, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $U, \\Sigma$ and $V$ are obtained by concatenating the singular value decompositions $\\left(\\{U_i\\}_{i=1}^T, \\{\\Sigma_i\\}_{i=1}^T, \\{V_i\\}_{i=1}^T\\right)$ of per-layer task matrices $\\{\\Delta_i\\}_{i=1}^T$ as detailed in Sec. 3.1. In the expression above, high inner product values for $U^\\top U$ and $V^\\top V$ imply a higher likelihood of interference, with minimal interference ideally yielding identity matrices.", + "bbox": [ + 89, + 731, + 483, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The underlying intuition is that overlapping singular vectors suggest shared features in the weight space across tasks. Such overlap can introduce interference when models are merged, ultimately degrading performance on individual tasks. We refer to Fig. 3 for a real example over eight tasks.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 513, + 90, + 622, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We demonstrate below how TSVs can be used for both compression and task interference reduction.", + "bbox": [ + 511, + 114, + 903, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. TSV for compression", + "text_level": 1, + "bbox": [ + 511, + 154, + 712, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the task is known or inferred (e.g. with a router, as in Mixture-of-Experts [15, 39] techniques), we can leverage the low-rank structure of per-layer task matrices (see Sec. 3.2) to effectively compress these ones while discarding task singular vectors from different tasks. In particular, given a known or inferred task index $h$ , we set other task components to zero, reducing Eq. (5) to:", + "bbox": [ + 511, + 176, + 905, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} = \\sum_ {i = 1} ^ {T} \\mathbb {1} _ {[ i = h ]} \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \\top} = \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {h} u _ {j} ^ {h} v _ {j} ^ {h \\top} = \\hat {\\Delta} _ {i}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 292, + 905, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This formula yields a low-rank approximation of $\\Delta_h$ , where only the top $k$ singular components of the task-specific matrix are retained. Here, the same number of components $k$ is taken from each task to ensure that each layer is compressed by a factor of $\\frac{1}{T}$ relative to the original dimension. Increasing $k$ improves approximation but reduces the compression.", + "bbox": [ + 511, + 345, + 905, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. TSV for model merging", + "text_level": 1, + "bbox": [ + 511, + 446, + 732, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In scenarios where the task identity is not known beforehand, we address the standard model merging problem by combining Task Singular Vectors (TSVs) to create a single model that performs well jointly across all tasks.", + "bbox": [ + 511, + 468, + 905, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reduce our measure of task interference, we decorrelate the TSVs of different tasks (encoded as columns in $\\hat{U}$ and $\\hat{V}$ ) by whitening these matrices to minimize their correlations. This can be achieved by applying the transformation $X \\mapsto X(X^{\\top}X)^{-\\frac{1}{2}}$ to both $\\hat{U}$ and $\\hat{V}$ . For improved numerical stability, we reformulate this whitening as an orthogonal Procrustes problem, seeking the orthogonal matrix $\\hat{U}_{\\perp}$ that minimizes the projection error:", + "bbox": [ + 511, + 529, + 905, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\hat {U} _ {\\perp}} \\| \\hat {U} _ {\\perp} - \\hat {U} \\| _ {F} \\quad \\text {s . t .} \\hat {U} _ {\\perp} ^ {\\top} \\hat {U} _ {\\perp} = I, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 657, + 903, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and similarly for $\\hat{V}$ . This problem admits a closed-form solution via the SVD $\\hat{U} = PDQ^{\\top}$ , yielding $\\hat{U}_{\\perp} = PQ^{\\top}$ (similarly for $\\hat{V}_{\\perp}$ ).", + "bbox": [ + 511, + 699, + 905, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 4.1. The transformations $X \\mapsto X(X^{\\top}X)^{-\\frac{1}{2}}$ (whitening) and $X \\mapsto PQ^{\\top}$ (Procrustes), where $X = PDQ^{\\top}$ is the SVD of $X$ , are equivalent.", + "bbox": [ + 511, + 752, + 905, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Given $X = PDQ^{\\top}$ and recalling that $D^{\\top}D = D^{2}$ for diagonal matrices, simple algebraic manipulation yields $X^{\\top}X = QD^{2}Q^{\\top}$ . It follows that $(X^{\\top}X)^{-1 / 2} = QD^{-1}Q^{\\top}$ , therefore the whitening transformation can be rewritten as $X(X^{\\top}X)^{-1 / 2} = (PDQ^{\\top})(QD^{-1}Q^{\\top}) = PDID^{-1}Q^{\\top} = PQ^{\\top}$ , completing the proof.", + "bbox": [ + 511, + 808, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18698", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b55846b13c15ccf51d7d4406a02a81453979021d81cb0dea9c2c5efdc8de5029.jpg", + "image_caption": [ + "Figure 3. Visualization of task interference among 8 tasks computed on the first attention layer of a ViT-B-32. The diagonal blocks display intra-task similarities, while the off-diagonal blocks illustrate inter-task similarities. The zoomed-in section highlights the interaction between the right singular vectors of the $3^{\\mathrm{rd}}$ and $4^{\\mathrm{th}}$ tasks." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 468, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2df729feab9265e61c1703f0e91034642d16d63b0c88e3aada49f8ef971ad59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 468, + 88, + 861, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f2ee7971f77a130fe96fe7cd91f91c9de2cd8b956b7cc7658366078b4c95042f.jpg", + "image_caption": [ + "Figure 4. Absolute accuracy of a ViT-B-32 merged over 8, 14, and 20 tasks, respectively." + ], + "image_footnote": [], + "bbox": [ + 96, + 405, + 356, + 623 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6c25b469be8f6f381fc43fbd4038701511e968c5dbc6cbde79245fcbec03f504.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 405, + 625, + 622 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a8d2c040cacf957448513937376889eb2841050a2f486a69a521a0d8a3ac02a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 405, + 901, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the (rank-reduced) TSV matrices $\\hat{U}$ and $\\hat{V}$ are decorrelated, we reconstruct the merged layer as in Eq. (5); see Algorithm 1 for the complete steps.", + "bbox": [ + 89, + 679, + 482, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 89, + 739, + 176, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our approaches over three different suites of tasks having cardinality 8, 14, and 20, respectively. The first one, introduced in [22], consists of datasets: Cars [28], DTD [4], EuroSAT [19], GTSRB [42], MNIST [30], RESISC45 [3], SUN397 [50], and SVHN [33]. The benchmark with 14 tasks builds on the preceding one, incorporating six additional datasets: CIFAR100 [29], STL10 [6], Flowers102 [34], OxfordIIITPet [35], PCAM [45], and FER2013 [18]. Finally, the 20-tasks benchmark in", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cludes the preceding 14 plus the following six: EMNIST [7], CIFAR10 [29], Food101 [2], FashionMNIST [49], RenderedSST2 [41], and KMNIST [5].", + "bbox": [ + 511, + 680, + 906, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Model merging results", + "text_level": 1, + "bbox": [ + 511, + 739, + 723, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on three variants of the CLIP [37] model, each employing a different size of ViT [13] visual encoder: ViT-B-32, ViT-B-16, and ViT-L-14. The main benchmarks involve merging 8, 14, and 20 tasks, mirroring the experimental setup described in Wang et al. [46]. We compare our approach against several training-free model merging methods, including weight averaging, Task Arithmetic [22], and Consensus Merging [46]. For reference, we include the performance of zero", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18699", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/03b97b0f268d348c59976af0030781f17df283fdb1389811fd8270b985d19647.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodViT-B-32ViT-B-16ViT-L-14
8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks
Zeroshot48.26(53.59)57.21(63.69)56.10(62.41)55.34(59.34)61.28(66.19)59.73(64.52)64.70(68.00)68.20(72.15)65.23(68.99)
Weight Averaging66.34(72.13)64.34(71.12)61.04(67.53)72.22(76.60)69.46(74.82)65.31(70.36)79.56(83.15)76.73(81.10)71.60(75.60)
Task Arithmetic70.79(76.55)65.32(72.09)60.52(66.79)75.41(79.58)70.52(75.89)65.78(70.76)84.93(88.65)79.41(83.95)74.01(78.07)
Consensus TA75.03(80.84)70.39(77.36)65.43(71.98)79.39(83.86)74.39(79.92)69.76(74.93)86.34(90.08)82.22(86.94)79.00(83.22)
TSV-M (Ours)85.86(92.31)80.06(87.88)77.07(84.29)89.01(93.94)84.58(91.01)80.57(86.45)92.98(96.98)89.17(94.43)87.72(92.50)
", + "bbox": [ + 94, + 88, + 903, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Average absolute accuracy results on model merging benchmarks; subscript (in parentheses) is the normalized average accuracy.", + "bbox": [ + 91, + 194, + 901, + 209 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 TSV-Merge." + ], + "code_body": "Require: Task matrices $\\Delta_1,\\ldots ,\\Delta_T$ , scaling factor $\\alpha$ \nEnsure: Merged model weights $\\theta_{\\mathrm{MT}}$ \n1: for $i = 1$ to $T$ do \n2: Compute SVD: $\\Delta_i = U_i\\Sigma_iV_i^\\top$ \n3: Retain first $\\frac{1}{T}$ singular components of $U_{i},\\Sigma_{i}$ , and $V_{i}$ \n4: end for \n5: Concatenate the matrices: \n6: $U\\gets [U_1|U_2|\\dots |U_T]$ \n7: $\\Sigma$ - block-diag( $\\Sigma_1,\\Sigma_2,\\dots ,\\Sigma_T)$ \n8: $V\\gets [V_{1}|V_{2}|\\dots |V_{T}]$ \n9: Compute the SVD of $U$ and $V$ .. \n10: $U = P_{U}D_{U}Q_{U}^{\\top}\\qquad V = P_{V}D_{V}Q_{V}^{\\top}$ \n11:Obtain the orthogonal matrices: \n12: $U_{\\perp} = P_{U}Q_{U}^{\\top}\\qquad V_{\\perp} = P_{V}Q_{V}^{\\top}$ \n13: Reconstruct the merged matrix: \n14: $\\hat{M}\\gets U_{\\perp}\\Sigma V_{\\perp}^{\\top}$ \n15: Construct merged model weights: \n16: $\\theta_{\\mathrm{MT}}\\gets \\theta_{\\mathrm{pre}} + \\alpha \\hat{M}$ \n17: return $\\theta_{\\mathrm{MT}}$", + "bbox": [ + 91, + 253, + 480, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shot models in Table 2 to represent the minimum achievable accuracy, and the average of individually fine-tuned models in Table 3 to indicate the maximum potential gains. Performance metrics were assessed using both the average absolute accuracy and the average normalized accuracy, calculated as detailed in Appendix B.1.", + "bbox": [ + 89, + 579, + 482, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As presented in Table 2, our method achieves state-of-the-art results across all benchmarks, regardless of the ViT size or the number of tasks involved. Notably, the most significant improvements were observed with the smaller ViT-B-32 encoder, where our approach outperforms Task Arithmetic and Consensus TA by an average of $+15.45\\%$ and $+10.71\\%$ absolute accuracy, respectively. Furthermore, when utilizing the ViT-L-14 model, our method attains an average normalized accuracy of $96.98\\%$ . This indicates that we can effectively replace eight individual task-specific models with a single merged model, incurring only a $-3.02\\%$ reduction in average accuracy. These results highlight the promise of our model merging technique as a cost-efficient alternative to mixtures of experts, multi-task learning, and assembling methods.", + "bbox": [ + 89, + 674, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report in Fig. 4 the per-dataset accuracies.", + "bbox": [ + 511, + 234, + 816, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Compression results", + "text_level": 1, + "bbox": [ + 511, + 260, + 705, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 3 we report the results for our compression algorithm TSV-C. We compare our method with TALL-masks [46], which stores very sparse binary masks for each task. As shown in Table 3, our method always retains more than $99\\%$ of the original accuracy for all considered benchmarks and models. The results of TSV-C are comparable to those of TALL-masks [46], with the former slightly underperforming TALL-masks for small ViTs and the latter gaining the upper hand in the larger-scale one. In general, however, the two approaches differ by less than $1\\%$ accuracy on average. Regarding the storage, our method stores only the top $\\frac{1}{T}$ singular vectors per task, which leads to a fixed storage requirement of approximately twice the size of the original model, regardless of the number of tasks. By storing binary masks, TALL-masks instead results in a storage size that varies with their compressibility, increasing with the number of tasks. This variability can lead to unpredictable storage requirements and may diminish compression benefits when masks are less compressible.", + "bbox": [ + 511, + 282, + 906, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our results show that TSV-C effectively balances compression and performance by leveraging the most significant TSVs. Maintaining near-original accuracy, our approach is particularly suitable for scenarios where storage constraints are critical but high model performance is still required.", + "bbox": [ + 511, + 570, + 906, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Analysis", + "text_level": 1, + "bbox": [ + 511, + 660, + 609, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we begin by conducting an ablation study to assess the contributions of interference reduction and low-rank compression to the overall performance. Next, we study how task interference varies across layers of different depths. Finally, we empirically show that our method does not require tuning a scaling coefficient.", + "bbox": [ + 511, + 686, + 905, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 787, + 663, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TSV-M combines low-rank approximation with task interference reduction. To evaluate the contribution of each component, we conducted an ablation study summarized in Table 4. Applying low-rank approximation alone to the layer task matrices results in worse performance than vanilla Task Arithmetic, shown in the first row where", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18700", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/20de4a1a895a1e27a49b7a87a05ac198e54636f24d3d3eb9b9768bc8fc103f50.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodViT-B-32ViT-B-16ViT-L-14
8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks
Finetuned92.83(100)90.88(100)91.37(100)94.64(100)92.76(100)93.17(100)95.81(100)94.29(100)94.73(100)
TALL-Mask+TIES93.13(100.37)90.92(100.04)91.11(99.70)94.68(100.04)92.69(99.90)93.05(99.87)95.96(100.15)93.40(99.09)93.91(99.16)
TSV-C (Ours)92.62(99.74)90.29(99.28)90.64(99.14)94.47(99.79)92.25(99.41)92.53(99.27)95.68(99.85)94.04(99.72)94.42(99.66)
", + "bbox": [ + 94, + 88, + 903, + 164 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b5d5906fe7d527fba09231a278f78b7d9b95f3e3d386c7cbdded913a55fbbf8.jpg", + "table_caption": [ + "Table 3. Average absolute accuracy results across all compression benchmarks for different models and varying number of tasks, subscript (in parentheses) the normalized average accuracies. Our TSV-C compression method consistently achieves over $99\\%$ of the original fine-tuned models' accuracy while significantly reducing storage requirements." + ], + "table_footnote": [], + "table_body": "
Low-rank approx.Interf. reductionViT-B-32
8 tasks14 tasks20 tasks
××76.5 (+0.0)72.1 (+0.0)66.8 (+0.0)
✓×75.2 (-1.3)71.0 (-1.1)66.3 (-0.5)
×✓82.6 (+7.4)75.7 (+4.7)69.9 (+3.6)
✓✓92.3 (+9.7)87.9 (+12.2)84.3 (+14.4)
", + "bbox": [ + 94, + 239, + 480, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison of different versions of Task Arithmetic, comprising either the low-rank approximation step, the interference reduction step, or both. The method performing both corresponds to the proposed TSV-Merge.", + "bbox": [ + 89, + 342, + 482, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf1bef04df4181e912a6d06a8efdbf544fc1f6b350bee695e433935858d44d98.jpg", + "image_caption": [ + "Figure 5. Approximation error from the orthogonalization of the TSVs through Procrustes for the ViT-B-32 model across 8 tasks. The violin plots represent layer-wise approximation error distributions for $U$ and $V$ in both full-rank and low-rank cases." + ], + "image_footnote": [], + "bbox": [ + 93, + 415, + 480, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "neither component is applied. In contrast, applying task interference reduction alone, while keeping full-rank matrices, significantly improves performance, with gains ranging from $+3.1\\%$ to $+6.1\\%$ . However, the best results are achieved only when both steps are combined, yielding substantial performance improvements of $+15.8\\%$ to $+17.5\\%$ .", + "bbox": [ + 89, + 656, + 482, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To explain why interference reduction applied to low-rank approximations in TSV-M outperforms its application to full-rank matrices, we analyze the errors introduced during the orthogonalization of the full-rank $U$ and $V$ matrices. As shown in Fig. 5, orthogonalizing full-rank matrices incurs significant approximation errors, measured by the sum of the Frobenius norms of reconstruction errors for the concatenated $U$ and $V$ matrices across all layers.", + "bbox": [ + 89, + 748, + 482, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For example, in the ViT-B-32 model with 8 tasks, the full-rank setting exhibits a wider error distribution with a", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "higher average, indicating greater variability and larger approximation discrepancies. In contrast, the low-rank setting produces a more compact and lower error distribution, suggesting better consistency in approximation across layers. In the following, we prove that this is not specific to the chosen model but a general property of our approach. The proof can be found in Appendix C.3.", + "bbox": [ + 511, + 243, + 903, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 6.1. Let $T \\in \\mathbb{N}$ such that $T > 4$ . Define $U = [U_1, \\ldots, U_T]$ as the matrix obtained by concatenating $T$ orthogonal matrices $U_i$ , each of shape $n \\times n$ . Let $\\widehat{U} = [\\widehat{U}_1, \\ldots, \\widehat{U}_T]$ be the matrix formed by truncating each $U_i$ to its first $k$ columns. Denote by $X$ and $\\widehat{X}$ the matrices resulting from Procrustes orthonormalization of $U$ and $\\widehat{U}$ , respectively. If $k \\leq n\\frac{T - 2\\sqrt{T}}{T}$ , then", + "bbox": [ + 511, + 353, + 906, + 465 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| U - X \\right\\| _ {F} \\geq \\left\\| \\widehat {U} - \\widehat {X} \\right\\| _ {F}.\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 469, + 800, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intuitively, the theorem asserts that the approximation error introduced by orthogonalization via Procrustes is smaller when applied to the concatenation of truncated matrices, compared to the concatenation of the original full-rank matrices, provided the rank of the truncated matrices satisfies certain conditions. To better understand the statement of the theorem, consider the case $T = 10$ , as in our scenario. The term $\\frac{T - 2\\sqrt{T}}{T} = \\frac{10 - 2\\sqrt{10}}{10} \\geq \\frac{1}{3}$ . In our case, we choose $k \\approx \\frac{n}{T}$ . For $T = 10$ , this gives $k \\approx \\frac{n}{10}$ , which is indeed smaller than $\\frac{n}{3}$ , satisfying the condition $k \\leq \\frac{1}{3} n$ .", + "bbox": [ + 511, + 494, + 905, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We emphasize that other orthogonalization methods, such as Gram-Schmidt, are ineffective in this context because they do not preserve the overall structure of the original data. Gram-Schmidt sequentially orthogonalizes vectors without minimizing deviation from the original set, potentially resulting in significant distortions.", + "bbox": [ + 511, + 646, + 905, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the substantial reduction in approximation error suggests that, while the advantages of interference reduction are considerable, they may be offset by the errors introduced when applied to full-rank matrices. In contrast, starting with low-rank approximations captures the essential information of each task, making the orthogonalization step less costly in terms of approximation error.", + "bbox": [ + 511, + 736, + 905, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Per-layer task interference", + "text_level": 1, + "bbox": [ + 511, + 848, + 754, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze task interference on a per-layer basis using Eq. (6). As shown in Fig. 6, interference is highest in the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18701", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7a7238853af6382173ed4d7135af578e9e7c67d2f2bb92821a7151a18d27d62.jpg", + "image_caption": [ + "Figure 6. Singular Task Interference (STI) across layers in a ViT-B-32 for 20 tasks. STI is high in early layers sharing common knowledge and lower in more specialized ones." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 475, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a5411dc1e6f1cfd5efac7fccfe7fe08c58c84950a6898ae253c3dcc440cdc9fc.jpg", + "image_caption": [ + "Figure 7. Best average normalized accuracy for different alpha values. The vertical labels indicate different sets of 8 tasks." + ], + "image_footnote": [], + "bbox": [ + 94, + 303, + 478, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "initial transformer layers and decreases significantly in the deeper layers. This aligns with the understanding that early, generalized layers capture common features across tasks, increasing the potential for conflict, while deeper layers are more specialized for specific tasks, reducing shared representation and interference. For brevity, we grouped layers within each transformer block in our analysis; each \"Layer $n$ \" in Fig. 6 includes two attention matrices and two MLP matrices. A layer-by-layer analysis is provided in Fig. 14.", + "bbox": [ + 89, + 551, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Choice of the interpolation coefficient", + "text_level": 1, + "bbox": [ + 89, + 696, + 415, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The aggregation in Eq. (2) involves a scaling coefficient $\\alpha$ , typically tuned using a validation set, that can have a significant impact on the accuracy of the final model. However, as shown in Fig. 7, our approach consistently achieves the best results with $\\alpha = 1.0$ , indicating that no additional scaling is necessary. Although purely empirical, this finding spares further tuning and validation data, thereby enhancing the practicality and efficiency of our method.", + "bbox": [ + 89, + 718, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Effect of task interference reduction", + "text_level": 1, + "bbox": [ + 89, + 848, + 403, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report in Fig. 8 the interference of a ViT-B-32 model both before and after the application of our TSV-M pipeline,", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7969a353a048b0e1838da3592cda9d730db2984a3796a0006e800d4e0e772a84.jpg", + "image_caption": [ + "Figure 8. Singular Task Interference (STI) and average normalized accuracy for Task Arithmetic and TSV-Merge on the ViT-B-32 model, evaluated across merges of 8, 14, and 20 tasks." + ], + "image_footnote": [], + "bbox": [ + 553, + 94, + 867, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "assessed across three distinct sets of tasks with different cardinalities. In all instances, we observed that a reduction in interference is closely associated with a significant gain in accuracy. This provides empirical evidence for the effectiveness of our approach in minimizing interference and, ultimately, enhancing the merging process.", + "bbox": [ + 511, + 319, + 903, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 429, + 638, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we tackled the problem of model merging by analyzing the SVD of per-layer task matrices, confirming their inherent low-rank structure and leveraging their singular vectors to define task interference.", + "bbox": [ + 511, + 455, + 903, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building on the former insight, we introduced TSV-Compress (TSV-C), a model compression algorithm that reduces task vectors to $10\\%$ of their original size while retaining $99\\%$ of the original accuracy. Unlike existing methods, our approach maintains a constant storage requirement, independent of the number of tasks. We then developed TSV-Merge (TSV-M), combining low-rank approximation with interference reduction to create a novel model merging technique. Achieving normalized accuracy up to $97\\%$ , our method offers a storage-efficient alternative to MoEs, ensembles, and multi-task learning.", + "bbox": [ + 511, + 518, + 905, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our extensive evaluation revealed that while interference reduction contributes most to performance improvements, the optimal results are achieved in combination with compression. We also observed that task interference decreases in deeper layers, and importantly, that our method does not require tuning a scaling coefficient.", + "bbox": [ + 511, + 686, + 903, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Exploring alternative methods for task importance and rank approximation could be a valuable direction for future work. Currently, we use a uniform rank across tasks, approximating each task vector to $\\frac{1}{T}$ of the layer's dimensionality. However, techniques like those from Idelbayev and Carreira-Perpinan [21] or optimal hard thresholds from Gavish and Donoho [17] could be applied to select the optimal rank for each task individually.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18702", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 225, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by PRIN 2022 project 20227YET9B \"AdVVent\" CUP code B53D23012830006 and by projects FAIR (PE0000013) and SERICS (PE00000014) under the MUR National Recovery and Resilience Plan funded by the European Union - NextGenerationEU. This work was moreover partially supported by the project NEREO (Neural Reasoning over Open Data) project funded by the Italian Ministry of Education and Research (PRIN) Grant no. 2022AEFHAZ.", + "bbox": [ + 89, + 109, + 485, + 247 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 262, + 187, + 277 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - Mining Discriminative Components with Random Forests. In Computer Vision - ECCV 2014, pages 446-461, Cham, 2014. Springer International Publishing. 5", + "[3] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 105(10):1865-1883, 2017. Conference Name: Proceedings of the IEEE. 5", + "[4] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing Textures in the Wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, Columbus, OH, USA, 2014. IEEE. 5", + "[5] Tarin Clanuwat, Mikel Bober-Irizar, Asanobu Kitamoto, Alex Lamb, Kazuaki Yamamoto, and David Ha. Deep learning for classical japanese literature. CoRR, abs/1812.01718, 2018. 5", + "[6] Adam Coates, Andrew Ng, and Honglak Lee. An Analysis of Single-Layer Networks in Unsupervised Feature Learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. ISSN: 1938-7228. 5", + "[7] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and André van Schaik. EMNIST: Extending MNIST to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 2921-2926, 2017. ISSN: 2161-4407. 5", + "[8] Donato Crisostomi, Marco Fumero, Daniele Baieri, Florian Bernard, and Emanuele Rodola. $c^2 m^3$ : Cycle-consistent multi-model merging. In Advances in Neural Information Processing Systems, 2025. 2", + "[9] Nico Daheim, Thomas Möllenhoff, Edoardo Ponti, Iryna Gurevych, and Mohammad Emtiyaz Khan. Model merging by uncertainty-based gradient matching. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[10] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse" + ], + "bbox": [ + 93, + 286, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "masks. In European Conference on Computer Vision, pages 270-287. Springer, 2025. 1, 2", + "[11] Misha Denil, Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato, and Nando De Freitas. Predicting parameters in deep learning. Advances in neural information processing systems, 26, 2013. 3", + "[12] Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. Advances in neural information processing systems, 27, 2014. 3, 18", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 5", + "[14] Carl Eckart and G. Marion Young. The approximation of one matrix by another of lower rank. Psychometrika, 1:211-218, 1936. 3", + "[15] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Workshop Track Proceedings, 2014. 4", + "[16] Timur Garipov, Dmitry Podoprikhin, Alexander Novikov, and Dmitry Vetrov. Ultimate tensorization: compressing convolutional and fc layers alike. arXiv preprint arXiv:1611.03214, 2016. 3", + "[17] Matan Gavish and David L. Donoho. The optimal hard threshold for singular values is $4 / \\sqrt{3}$ . IEEE Transactions on Information Theory, 60(8):5040-5053, 2014. 8", + "[18] Ian J. Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, Yingbo Zhou, Chetan Ramaiah, Fangxiang Feng, Ruifan Li, Xiaojie Wang, Dimitris Athanasakis, John Shawe-Taylor, Maxim Milakov, John Park, Radu Ionescu, Marius Popescu, Cristian Grozea, James Bergstra, Jingjing Xie, Lukasz Romaszko, Bing Xu, Zhang Chuang, and Yoshua Bengio. Challenges in Representation Learning: A Report on Three Machine Learning Contests. In Neural Information Processing, pages 117-124, Berlin, Heidelberg, 2013. Springer. 5", + "[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. Conference Name: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 5", + "[20] Chenyu Huang, Peng Ye, Tao Chen, Tong He, Xiangyu Yue, and Wanli Ouyang. EMR-merging: Tuning-free high-performance model merging. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2", + "[21] Yerlan Idelbayev and Miguel A Carreira-Perpinán. Low-rank compression of neural nets: Learning the rank of each layer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8049–8059, 2020. 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18703", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 13", + "[23] Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. In *British Machine Vision Conference*, BMVC 2014, Nottingham, UK, September 1-5, 2014. BMVA Press, 2014. 3", + "[24] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2023. 1, 2", + "[25] Keller Jordan, Hanie Sedghi, Olga Saukh, Rahim Entezari, and Behnam Neyshabur. REPAIR: REnormalizing permuted activations for interpolation repair. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[26] Yong-Deok Kim, Eunhyeok Park, Sungjoo Yoo, Taelim Choi, Lu Yang, and Dongjun Shin. Compression of deep convolutional neural networks for fast and low power mobile applications. In Proc. of the 4th Int. Conf. Learning Representations (ICLR 2016), San Juan, Puerto Rico, May 2–4, 2016. 3", + "[27] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 3", + "[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D Object Representations for Fine-Grained Categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, Sydney, Australia, 2013. IEEE. 5", + "[29] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario, 2009. 5", + "[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5", + "[31] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. Advances in Neural Information Processing Systems, 37:78905-78935, 2024. 2", + "[32] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 35:17703-17716, 2022. 1, 2", + "[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 5", + "[34] Maria-Elena Nilsback and Andrew Zisserman. Automated Flower Classification over a Large Number of Classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5", + "[35] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. ISSN: 1063-6919. 5" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Fidel A Guerrero Pña, Heitor Rapela Medeiros, Thomas Dubail, Masih Aminbeidokhti, Eric Granger, and Marco Pedersoli. Re-basin via implicit sinkhorn differentiation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[38] Alexandre Rame, Matthieu Kirchmeyer, Thibaud Rahier, Alain Rakotomamonjy, Patrick Gallinari, and Matthieu Cord. Diverse weight averaging for out-of-distribution generalization. Advances in Neural Information Processing Systems, 35:10821-10836, 2022. 1", + "[39] Noam Shazeer, *Azalia Mirhoseini*, *Krzysztof Maziarz*, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2017. 4", + "[40] Li Shen, Anke Tang, Enneng Yang, Guibing Guo, Yong Luo, Lefei Zhang, Xiaochun Cao, Bo Du, and Dacheng Tao. Efficient and effective weight-ensembling mixture of experts for multi-task model merging. arXiv preprint arXiv:2410.21804, 2024. 2", + "[41] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA, 2013. Association for Computational Linguistics. 5", + "[42] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The German Traffic Sign Recognition Benchmark: A multi-class classification competition. In *The 2011 International Joint Conference on Neural Networks*, pages 1453–1460, 2011. ISSN: 2161-4407. 5", + "[43] George Stoica, Daniel Bolya, Jakob Brandt Bjorner, Pratik Ramesh, Taylor Hearn, and Judy Hoffman. Zipit! merging models from different tasks without training. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[44] Anke Tang, Li Shen, Yong Luo, Shuai Xie, Han Hu, Lefei Zhang, Bo Du, and Dacheng Tao. Smile: Zero-shot sparse mixture of low-rank experts construction from pre-trained foundation models. arXiv preprint arXiv:2408.10174, 2024. 2", + "[45] Bastiaan S. Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation Equivariant CNNs for Digital Pathology. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2018, pages 210–218, Cham, 2018. Springer International Publishing. 5", + "[46] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. In *Forty-first International Conference on Machine Learning*, 2024. 2, 5, 6, 12, 13" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18704", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] B. P. Welford. Note on a method for calculating corrected sums of squares and products. Technometrics, 4(3):419-420, 1962. 12", + "[48] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 2", + "[49] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms, 2017. arXiv:1708.07747 [cs, stat]. 5", + "[50] Jianxiong Xiao, Krista A. Ehinger, James Hays, Antonio Torralba, and Aude Oliva. SUN Database: Exploring a Large Collection of Scene Categories. International Journal of Computer Vision, 119(1):3-22, 2016. 5", + "[51] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2", + "[52] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 18", + "[53] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2", + "[54] Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Fabrizio Silvestri, and Emanuele Rodola. Atm: Improving model merging by alternating tuning and merging, 2024. 2" + ], + "bbox": [ + 89, + 92, + 483, + 584 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18705", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_model.json b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..21d14c9258e43af9aaa10e83f2fb45a6241ec1a5 --- /dev/null +++ b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.262, + 0.131, + 0.734, + 0.177 + ], + "angle": 0, + "content": "Task Singular Vectors: Reducing Task Interference in Model Merging" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.199, + 0.4, + 0.24 + ], + "angle": 0, + "content": "Antonio Andrea Gargiulo\\* Simone Scardapane" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.199, + 0.593, + 0.238 + ], + "angle": 0, + "content": "Donato Crisostomi* \nFabrizio Silvestri" + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.199, + 0.807, + 0.238 + ], + "angle": 0, + "content": "Maria Sofia Bucarelli Emanuele Rodolà" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.247, + 0.613, + 0.264 + ], + "angle": 0, + "content": "Sapienza University of Rome" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.27, + 0.829, + 0.303 + ], + "angle": 0, + "content": "gargiulo.1769185@studenti.uniroma1.it {crisostomi, rodola}@di.uniroma1.it simone.scardapane@uniroma1.it {bucarelli, fsilvestri}@diag.uniroma1.it" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.338, + 0.327, + 0.354 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.486, + 0.643 + ], + "angle": 0, + "content": "Task Arithmetic has emerged as a simple yet effective method to merge models without additional training. However, by treating entire networks as flat parameter vectors, it overlooks key structural information and is susceptible to task interference. In this paper, we study task vectors at the layer level, focusing on task layer matrices and their singular value decomposition. In particular, we concentrate on the resulting singular vectors, which we refer to as Task Singular Vectors (TSV). Recognizing that layer task matrices are often low-rank, we propose TSV-Compress (TSV-C), a simple procedure that compresses them to \\(10\\%\\) of their original size while retaining \\(99\\%\\) of accuracy. We further leverage this low-rank space to define a new measure of task interference based on the interaction of singular vectors from different tasks. Building on these findings, we introduce TSV-Merge (TSV-M), a novel model merging approach that combines compression with interference reduction, significantly outperforming existing methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.222, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.485, + 0.879 + ], + "angle": 0, + "content": "The widespread availability of pre-trained models and public repositories has driven the development of techniques for efficiently combining and reusing existing models. Among these techniques, model merging approaches enable the creation of multi-task models without additional training. One popular approach, Task Arithmetic (TA) [22], stands out for its simplicity and effectiveness. However, by treating entire networks as high-dimensional vectors, TA and subsequent works overlook crucial structural information [9, 10, 24, 32, 38, 51, 53]. This flattened view limits these approaches to coarse-grained measures like cosine similarity for assessing inter-task interactions." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.339, + 0.905, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.5, + 0.908, + 0.543 + ], + "angle": 0, + "content": "Figure 1. Mean accuracy of a ViT-L-14 merged over 8, 14, and 20 tasks respectively. By significantly surpassing existing methods, TSV-M establishes the new state of the art in model merging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.811 + ], + "angle": 0, + "content": "In this work, we instead focus on preserving and leveraging the natural structure of neural networks by examining weight differences at the layer level while retaining their matrix form wherever possible. Our approach begins by examining per-layer task matrices, which represent weight changes for each task, through singular value decomposition (SVD). This decomposition yields singular vectors and singular values that capture the most significant directions of variation within each layer. We term these singular vectors Task Singular Vectors (TSV), as they provide an interpretable basis for assessing task-specific contributions at each layer. Importantly, in accordance with recent literature on PEFT (cfr. Sec. 2), our analysis confirms that task matrices are inherently low-rank, meaning that only a small portion of TSVs is sufficient to represent the layer's function with high fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Building on this insight, we introduce TSV-Compress (TSV-C), a simple yet effective procedure to compress task vectors down to \\(10\\%\\) of their original size while maintaining \\(99\\%\\) of the original accuracy. Focusing on only a fraction of the most relevant TSVs for each task, TSV-Compress enables efficient storage and processing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.272, + 0.9 + ], + "angle": 0, + "content": "\\(\\star\\) denotes equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18695" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.478, + 0.107 + ], + "angle": 0, + "content": "of task-specific information without waiving performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.108, + 0.485, + 0.425 + ], + "angle": 0, + "content": "Beyond compression, examining the interplay of TSVs across different tasks provides a geometrically informed framework for analyzing task interference at the individual layer level. By assessing how singular vectors from different tasks align or diverge within each layer, this approach offers a significantly more fine-grained understanding of inter-task interactions, going beyond global vector similarity metrics like cosine similarity. Building upon these principles, we introduce TSV-Merge (TSV-M), a novel model merging method that combines compression and task interference reduction. This is achieved by discarding the irrelevant singular vectors for each task and then reducing their inter-task interference with a whitening transformation over their similarity matrix. We empirically show our approach to effectively reduce task interference, and the reduction to be complementary to the compression step. We then evaluate our approach across several benchmarks and show it outperforms existing methods by an average of \\(\\sim 15\\%\\) accuracy points, establishing the new state of the art by a large margin. We release all our code for research purposes1. In summary, our contribution is 4-fold:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.426, + 0.483, + 0.47 + ], + "angle": 0, + "content": "- We study the singular value decomposition of per-layer task matrices, showing them to be low-rank and measuring the interplay of singular vectors from different tasks;" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.471, + 0.483, + 0.517 + ], + "angle": 0, + "content": "- We introduce TSV-C, a method that builds on this insight to compress task vectors by a factor of \\(10 \\times\\) while preserving \\(99\\%\\) of their original accuracy;" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.517, + 0.483, + 0.576 + ], + "angle": 0, + "content": "- We propose TSV-M, a model merging technique that complements compression with a task interference reduction step by applying a whitening transformation to decorrelate singular vectors across tasks;" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.577, + 0.483, + 0.652 + ], + "angle": 0, + "content": "- We conduct extensive experiments on multiple computer vision datasets, showing that TSV-M significantly outperforms existing model merging methods and provide in-depth analyses to uncover the underlying factors contributing to its success." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.426, + 0.483, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.233, + 0.684 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Model Merging offers an efficient alternative to ensembling by combining existing models without the need for further training. Several approaches address this by determining the neuron permutations that align the models into a shared optimization basin, after which the models are combined using a straightforward averaging method. [1, 8, 25, 36, 43]. An alternative line of approaches focuses on the multi-task scenario, where a single pre-trained model is fine-tuned for different tasks [10, 22, 32, 40, 46, 48, 51, 53, 54]. The prerequisites for methods in this category are outlined in Table 1. Task Arithmetic [22] introduces the concept of task vectors, which are the weight differences" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.09, + 0.887, + 0.104 + ], + "angle": 0, + "content": "Table 1. Model merging approaches and their requirements." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.115, + 0.909, + 0.332 + ], + "angle": 0, + "content": "
MethodAdditional trainingExtra storageValidat. data inputsdata labels
Weight Avg. [48]××××
Fisher-Merg. [32]××✓×
RegMean [24]××✓×
EMR-Merging [20]×✓××
TALL-Mask [46]×✓××
TSV-C (Ours)×✓××
Task Arith. [22]××✓✓
Ties-Merging [51]××✓✓
AdaMerging [52]✓×✓×
Consensus-TA [46]××✓✓
TSV-M (Ours)××××
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.81 + ], + "angle": 0, + "content": "between fine-tuned models and the pre-trained base model. By averaging these vectors, a merged multi-task model can be created; conversely, negating a task vector allows for forgetting a specific task. TIES [51] addresses redundancies in model parameters by first selecting the top-\\(k\\) most significant parameter changes and then constructing a sign vector based on the majority sign across all models. The latter is used to merge the task vectors disjointly, meaning the average is not computed when a parameter is zero or when parameters disagree in sign. Similarly, DARE [53] randomly resets redundant parameters to their pre-trained values and rescales the remaining parameters by a factor proportional to the dropped ones, aiming to reduce interference among tasks. Fisher Merging [32] and RegMean [24] merge models by performing weighted averaging, utilizing the Fisher information matrix and the inner product of input vectors, respectively. Model Broadcrumbs [10] focuses on merging only significant weights by discarding outliers and both minor and large perturbations in the fine-tuned parameters. More recently, Wang et al. [46] observed that tasks often utilize non-overlapping sets of weights. They propose two methods: the first, TALL-Mask, uses binary masks to activate important task-specific weights, requiring extra storage (as does our method TSV-C) for the masks. The second, Consensus Merging, leverages these masks to remove parameters that are important to less than two tasks. Like our approach, TwinMerging [31] and SMILE [44] apply SVD to layer task matrices. However, neglecting singular vector interference, they require a router to selectively activate a subset during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Unlike existing approaches, our methods explicitly address task interference by leveraging the geometric structure of singular vectors to minimize interactions between task-specific parameters. Rather than averaging or selectively merging parameters as prior techniques do, we decorrelate singular vectors to directly reduce interference across tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.401, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/AntoAndGar/task_singular_vectors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18696" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "Additionally, we achieve targeted compression by isolating each task's unique components, storing only essential parts, and ensuring minimal interference, a step beyond traditional SVD-based compression approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.179, + 0.485, + 0.391 + ], + "angle": 0, + "content": "SVD for Model Compression A significant body of work explores low-rank decompositions for fully-connected and convolutional layers [11, 12, 16, 23, 27], as well as tensor decompositions [12, 16, 26]. These approaches typically achieve a layer-by-layer factorization of a trained network, focusing on minimizing the difference between the original and the low-rank approximated weight matrices. This is accomplished using techniques like SVD [12, 16, 23, 27] or iterative optimization [23, 26]. While other methods such as weight quantization, knowledge distillation, and pruning exist, we utilize SVD not only for low-rank approximation and compression but also because its matrix decomposition properties make it particularly effective in analyzing and mitigating interference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.303, + 0.426 + ], + "angle": 0, + "content": "3. Task Singular Vectors" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.486, + 0.482 + ], + "angle": 0, + "content": "In this section, we introduce key concepts for understanding our approach, including a new measure of task interference that is central to our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.494, + 0.224, + 0.51 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.485, + 0.595 + ], + "angle": 0, + "content": "We build on Task Arithmetic (TA), which defines task vectors capturing the differences in model weights for individual tasks. Formally, the weights \\(\\theta_{\\mathrm{MT}}\\) of a multi-task model for \\(T\\) tasks are computed by aggregating the task-specific weight differences, or task vectors, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.606, + 0.484, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathrm {M T}} = \\theta_ {\\mathrm {p r e}} + \\alpha \\frac {\\sum_ {i = 1} ^ {T} \\tau_ {i}}{T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.485, + 0.73 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{pre}}\\) is the set of pretrained model weights, \\(\\alpha\\) is a scaling factor, and \\(\\tau_{i} = \\theta_{\\mathrm{ft}_{i}} - \\theta_{\\mathrm{pre}}\\) is the task vector for task \\(i\\), with \\(\\theta_{\\mathrm{ft}_i}\\) being the fine-tuned weights for the task. Differently from TA, however, we consider these operations at the layer level. From this perspective, Eq. (1) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.741, + 0.483, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathrm {M T}} ^ {(l)} = \\theta_ {\\text {p r e}} ^ {(l)} + \\alpha \\frac {\\sum_ {i = 1} ^ {T} \\Delta_ {i} ^ {(l)}}{T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.485, + 0.903 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{pre}}^{(l)}\\) encodes the pretrained weights for layer \\(l\\), and \\(\\Delta_i^{(l)} = \\theta_{\\mathrm{ft}_i}^{(l)} - \\theta_{\\mathrm{pre}}^{(l)}\\) is the task-specific weight difference for task \\(i\\) at layer \\(l\\). When layer \\(l\\) has a matrix structure, we call its corresponding \\(\\Delta_i^l\\) the per-layer task matrix for task \\(i\\). When this is not the case, our framework defaults to standard TA. For brevity, we will generally omit the layer index and refer to the layer-\\(l\\) task matrix \\(\\Delta_i^l\\) as \\(\\Delta_i\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.75, + 0.107 + ], + "angle": 0, + "content": "Decomposing layer task matrices" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.11, + 0.907, + 0.188 + ], + "angle": 0, + "content": "Treating layer-wise weights as structured entities rather than flattened vectors enables us to analyze their SVD, revealing low-rank properties and deeper insight into the inter-task interactions. Given two tasks \\(i,j\\), we consider the SVD of their task matrices \\(\\Delta_{i}\\) and \\(\\Delta_{j}\\) at a generic layer:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.195, + 0.836, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {i} = U _ {i} \\Sigma_ {i} V _ {i} ^ {\\top}, \\qquad \\Delta_ {j} = U _ {j} \\Sigma_ {j} V _ {j} ^ {\\top}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.224, + 0.907, + 0.3 + ], + "angle": 0, + "content": "where \\( U_{i}, U_{j} \\) and \\( V_{i}, V_{j} \\) are the matrices of left and right singular vectors respectively, and \\( \\Sigma_{i}, \\Sigma_{j} \\) are diagonal matrices of singular values. Due to their role in assessing task-specific contributions, we term the obtained singular vectors Task Singular Vectors (TSV)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.3, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Importantly for our treatment, we can equivalently write the aggregated task matrices in Eq. (2) in terms of their singular vectors and values. By defining \\( U = [U_{1}\\dots U_{T}] \\) as the column-wise concatenation of all the left TSVs, \\( V = [V_{1}\\dots V_{T}] \\) as the row-wise concatenation of the right TSVs, and \\( \\Sigma \\) as the block diagonal matrix with \\( \\{\\Sigma_i\\}_{i=1}^T \\) along its diagonal, we can write:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.416, + 0.907, + 0.458 + ], + "angle": 0, + "content": "\\[\nM = U \\Sigma V ^ {\\top} = \\sum_ {i = 1} ^ {T} U _ {i} \\Sigma_ {i} V _ {i} ^ {\\top} = \\sum_ {i = 1} ^ {T} \\Delta_ {i}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.467, + 0.881, + 0.484 + ], + "angle": 0, + "content": "In the simple case where \\( T = 2 \\), \\( M \\) would be given by:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.491, + 0.895, + 0.527 + ], + "angle": 0, + "content": "\\[\nM = \\left[ \\begin{array}{c c} U _ {1} & U _ {2} \\end{array} \\right] \\left[ \\begin{array}{c c} \\Sigma_ {1} & 0 \\\\ 0 & \\Sigma_ {2} \\end{array} \\right] \\left[ \\begin{array}{c} V _ {1} ^ {\\top} \\\\ V _ {2} ^ {\\top} \\end{array} \\right] = \\Delta_ {1} + \\Delta_ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.907, + 0.643 + ], + "angle": 0, + "content": "When concatenating singular components from different tasks, we violate certain properties of the SVD. Specifically, the matrices \\( U \\) and \\( V \\) become non-orthogonal because singular vectors from different tasks may overlap. Additionally, the singular values \\( \\Sigma_{i} \\) from different tasks can vary significantly in magnitude, which may bias the merging process toward tasks with larger singular values." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.65, + 0.853, + 0.667 + ], + "angle": 0, + "content": "3.2. Low-rank nature of layer task matrices" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.672, + 0.909, + 0.764 + ], + "angle": 0, + "content": "We start our analysis by studying the low-rank properties of per-layer task matrices. Interpreting SVD as a sum of rank-one matrices, for each task \\(i\\), by Eckart-Young's theorem [14] we get the best approximation (in Frobenius norm) of each task matrix \\(\\Delta_{i}\\) by retaining only the top-\\(k\\) singular values and their corresponding vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.773, + 0.907, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\hat {\\Delta} _ {i} = \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \\top}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "As we show in Fig. 2, we find task matrices to be inherently low-rank: a small subset of TSVs is sufficient to represent the layer's function with high fidelity. In particular, even when preserving only \\(3\\%\\) of the singular components per task, the mean accuracy drops by merely \\(1.5\\%\\). This is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18697" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.099, + 0.473, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.485, + 0.322 + ], + "angle": 0, + "content": "Figure 2. Mean absolute accuracy of the ViT-B-32 model across increasing fractions of retained singular components, averaged over 20 tasks. The red line represents the average accuracy of the original fine-tuned models with full-rank task matrices, while the green line shows the accuracies using low-rank approximations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.419 + ], + "angle": 0, + "content": "noticeable considering that \\(97\\%\\) of singular components in each layer matrix are discarded. Building on this insight, we propose in Sec. 4.1 a compression algorithm that maintains \\(99\\%\\) of the accuracy while shrinking the task vectors to \\(10\\%\\) of their original size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.484, + 0.466 + ], + "angle": 0, + "content": "Given this low-rank structure, it is natural to aggregate task matrices within their subspaces. We, therefore, obtain a reduced version of the aggregation matrix \\( M \\) (Eq. (3)) as:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.477, + 0.483, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\hat {M} = \\sum_ {i = 1} ^ {T} \\hat {U} _ {i} \\hat {\\Sigma} _ {i} \\hat {V} _ {i} ^ {\\top} = \\sum_ {i = 1} ^ {T} \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \\top}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.62 + ], + "angle": 0, + "content": "where \\(\\hat{U}_i\\) and \\(\\hat{V}_i\\) contain the top-\\(k\\) left and right singular vectors for task \\(i\\), respectively, and \\(\\hat{\\Sigma}_i\\) is the diagonal matrix of the top-\\(k\\) singular values \\(\\sigma_j^i\\). This formulation highlights that the low-rank approximation \\(\\hat{M}\\) utilizes the most relevant singular components from each task to effectively combine the task-specific weight differences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.627, + 0.332, + 0.643 + ], + "angle": 0, + "content": "3.3. Singular Task interference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.484, + 0.695 + ], + "angle": 0, + "content": "We hereby introduce a score of task interference based on the interplay of TSVs from different tasks, which we term Singular Task Interference (STI)" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.702, + 0.483, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S T I} \\left(\\left\\{\\Delta_ {i} \\right\\} _ {i = 1} ^ {T}\\right) = \\left\\| \\left(U ^ {\\top} U - I\\right) \\Sigma \\left(V ^ {\\top} V - I\\right) \\right\\| _ {1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.484, + 0.825 + ], + "angle": 0, + "content": "where \\(U, \\Sigma\\) and \\(V\\) are obtained by concatenating the singular value decompositions \\(\\left(\\{U_i\\}_{i=1}^T, \\{\\Sigma_i\\}_{i=1}^T, \\{V_i\\}_{i=1}^T\\right)\\) of per-layer task matrices \\(\\{\\Delta_i\\}_{i=1}^T\\) as detailed in Sec. 3.1. In the expression above, high inner product values for \\(U^\\top U\\) and \\(V^\\top V\\) imply a higher likelihood of interference, with minimal interference ideally yielding identity matrices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The underlying intuition is that overlapping singular vectors suggest shared features in the weight space across tasks. Such overlap can introduce interference when models are merged, ultimately degrading performance on individual tasks. We refer to Fig. 3 for a real example over eight tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.623, + 0.108 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.145 + ], + "angle": 0, + "content": "We demonstrate below how TSVs can be used for both compression and task interference reduction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.155, + 0.714, + 0.171 + ], + "angle": 0, + "content": "4.1. TSV for compression" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.177, + 0.906, + 0.283 + ], + "angle": 0, + "content": "When the task is known or inferred (e.g. with a router, as in Mixture-of-Experts [15, 39] techniques), we can leverage the low-rank structure of per-layer task matrices (see Sec. 3.2) to effectively compress these ones while discarding task singular vectors from different tasks. In particular, given a known or inferred task index \\( h \\), we set other task components to zero, reducing Eq. (5) to:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.294, + 0.906, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\hat {M} = \\sum_ {i = 1} ^ {T} \\mathbb {1} _ {[ i = h ]} \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \\top} = \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {h} u _ {j} ^ {h} v _ {j} ^ {h \\top} = \\hat {\\Delta} _ {i}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.906, + 0.438 + ], + "angle": 0, + "content": "This formula yields a low-rank approximation of \\(\\Delta_h\\), where only the top \\(k\\) singular components of the task-specific matrix are retained. Here, the same number of components \\(k\\) is taken from each task to ensure that each layer is compressed by a factor of \\(\\frac{1}{T}\\) relative to the original dimension. Increasing \\(k\\) improves approximation but reduces the compression." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.447, + 0.733, + 0.463 + ], + "angle": 0, + "content": "4.2. TSV for model merging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.906, + 0.529 + ], + "angle": 0, + "content": "In scenarios where the task identity is not known beforehand, we address the standard model merging problem by combining Task Singular Vectors (TSVs) to create a single model that performs well jointly across all tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.906, + 0.65 + ], + "angle": 0, + "content": "To reduce our measure of task interference, we decorrelate the TSVs of different tasks (encoded as columns in \\(\\hat{U}\\) and \\(\\hat{V}\\)) by whitening these matrices to minimize their correlations. This can be achieved by applying the transformation \\(X \\mapsto X(X^{\\top}X)^{-\\frac{1}{2}}\\) to both \\(\\hat{U}\\) and \\(\\hat{V}\\). For improved numerical stability, we reformulate this whitening as an orthogonal Procrustes problem, seeking the orthogonal matrix \\(\\hat{U}_{\\perp}\\) that minimizes the projection error:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.659, + 0.905, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\hat {U} _ {\\perp}} \\| \\hat {U} _ {\\perp} - \\hat {U} \\| _ {F} \\quad \\text {s . t .} \\hat {U} _ {\\perp} ^ {\\top} \\hat {U} _ {\\perp} = I, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.906, + 0.746 + ], + "angle": 0, + "content": "and similarly for \\(\\hat{V}\\). This problem admits a closed-form solution via the SVD \\(\\hat{U} = PDQ^{\\top}\\), yielding \\(\\hat{U}_{\\perp} = PQ^{\\top}\\) (similarly for \\(\\hat{V}_{\\perp}\\))." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.753, + 0.906, + 0.801 + ], + "angle": 0, + "content": "Proposition 4.1. The transformations \\( X \\mapsto X(X^{\\top}X)^{-\\frac{1}{2}} \\) (whitening) and \\( X \\mapsto PQ^{\\top} \\) (Procrustes), where \\( X = PDQ^{\\top} \\) is the SVD of \\( X \\), are equivalent." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.809, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Proof. Given \\( X = PDQ^{\\top} \\) and recalling that \\( D^{\\top}D = D^{2} \\) for diagonal matrices, simple algebraic manipulation yields \\( X^{\\top}X = QD^{2}Q^{\\top} \\). It follows that \\( (X^{\\top}X)^{-1 / 2} = QD^{-1}Q^{\\top} \\), therefore the whitening transformation can be rewritten as \\( X(X^{\\top}X)^{-1 / 2} = (PDQ^{\\top})(QD^{-1}Q^{\\top}) = PDID^{-1}Q^{\\top} = PQ^{\\top} \\), completing the proof." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18698" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.47, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.47, + 0.089, + 0.862, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.349, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 3. Visualization of task interference among 8 tasks computed on the first attention layer of a ViT-B-32. The diagonal blocks display intra-task similarities, while the off-diagonal blocks illustrate inter-task similarities. The zoomed-in section highlights the interaction between the right singular vectors of the \\(3^{\\mathrm{rd}}\\) and \\(4^{\\mathrm{th}}\\) tasks." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.406, + 0.357, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.406, + 0.627, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.406, + 0.902, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.64, + 0.772, + 0.656 + ], + "angle": 0, + "content": "Figure 4. Absolute accuracy of a ViT-B-32 merged over 8, 14, and 20 tasks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.727 + ], + "angle": 0, + "content": "Once the (rank-reduced) TSV matrices \\(\\hat{U}\\) and \\(\\hat{V}\\) are decorrelated, we reconstruct the merged layer as in Eq. (5); see Algorithm 1 for the complete steps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.178, + 0.755 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate our approaches over three different suites of tasks having cardinality 8, 14, and 20, respectively. The first one, introduced in [22], consists of datasets: Cars [28], DTD [4], EuroSAT [19], GTSRB [42], MNIST [30], RESISC45 [3], SUN397 [50], and SVHN [33]. The benchmark with 14 tasks builds on the preceding one, incorporating six additional datasets: CIFAR100 [29], STL10 [6], Flowers102 [34], OxfordIIITPet [35], PCAM [45], and FER2013 [18]. Finally, the 20-tasks benchmark in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.681, + 0.907, + 0.727 + ], + "angle": 0, + "content": "cludes the preceding 14 plus the following six: EMNIST [7], CIFAR10 [29], Food101 [2], FashionMNIST [49], RenderedSST2 [41], and KMNIST [5]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.724, + 0.758 + ], + "angle": 0, + "content": "5.1. Model merging results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We evaluate our method on three variants of the CLIP [37] model, each employing a different size of ViT [13] visual encoder: ViT-B-32, ViT-B-16, and ViT-L-14. The main benchmarks involve merging 8, 14, and 20 tasks, mirroring the experimental setup described in Wang et al. [46]. We compare our approach against several training-free model merging methods, including weight averaging, Task Arithmetic [22], and Consensus Merging [46]. For reference, we include the performance of zero" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18699" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
MethodViT-B-32ViT-B-16ViT-L-14
8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks
Zeroshot48.26(53.59)57.21(63.69)56.10(62.41)55.34(59.34)61.28(66.19)59.73(64.52)64.70(68.00)68.20(72.15)65.23(68.99)
Weight Averaging66.34(72.13)64.34(71.12)61.04(67.53)72.22(76.60)69.46(74.82)65.31(70.36)79.56(83.15)76.73(81.10)71.60(75.60)
Task Arithmetic70.79(76.55)65.32(72.09)60.52(66.79)75.41(79.58)70.52(75.89)65.78(70.76)84.93(88.65)79.41(83.95)74.01(78.07)
Consensus TA75.03(80.84)70.39(77.36)65.43(71.98)79.39(83.86)74.39(79.92)69.76(74.93)86.34(90.08)82.22(86.94)79.00(83.22)
TSV-M (Ours)85.86(92.31)80.06(87.88)77.07(84.29)89.01(93.94)84.58(91.01)80.57(86.45)92.98(96.98)89.17(94.43)87.72(92.50)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.195, + 0.903, + 0.21 + ], + "angle": 0, + "content": "Table 2. Average absolute accuracy results on model merging benchmarks; subscript (in parentheses) is the normalized average accuracy." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.235, + 0.276, + 0.251 + ], + "angle": 0, + "content": "Algorithm 1 TSV-Merge." + }, + { + "type": "algorithm", + "bbox": [ + 0.092, + 0.254, + 0.482, + 0.544 + ], + "angle": 0, + "content": "Require: Task matrices \\(\\Delta_1,\\ldots ,\\Delta_T\\) , scaling factor \\(\\alpha\\) \nEnsure: Merged model weights \\(\\theta_{\\mathrm{MT}}\\) \n1: for \\(i = 1\\) to \\(T\\) do \n2: Compute SVD: \\(\\Delta_i = U_i\\Sigma_iV_i^\\top\\) \n3: Retain first \\(\\frac{1}{T}\\) singular components of \\(U_{i},\\Sigma_{i}\\) , and \\(V_{i}\\) \n4: end for \n5: Concatenate the matrices: \n6: \\(U\\gets [U_1|U_2|\\dots |U_T]\\) \n7: \\(\\Sigma\\) - block-diag( \\(\\Sigma_1,\\Sigma_2,\\dots ,\\Sigma_T)\\) \n8: \\(V\\gets [V_{1}|V_{2}|\\dots |V_{T}]\\) \n9: Compute the SVD of \\(U\\) and \\(V\\) .. \n10: \\(U = P_{U}D_{U}Q_{U}^{\\top}\\qquad V = P_{V}D_{V}Q_{V}^{\\top}\\) \n11:Obtain the orthogonal matrices: \n12: \\(U_{\\perp} = P_{U}Q_{U}^{\\top}\\qquad V_{\\perp} = P_{V}Q_{V}^{\\top}\\) \n13: Reconstruct the merged matrix: \n14: \\(\\hat{M}\\gets U_{\\perp}\\Sigma V_{\\perp}^{\\top}\\) \n15: Construct merged model weights: \n16: \\(\\theta_{\\mathrm{MT}}\\gets \\theta_{\\mathrm{pre}} + \\alpha \\hat{M}\\) \n17: return \\(\\theta_{\\mathrm{MT}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.672 + ], + "angle": 0, + "content": "shot models in Table 2 to represent the minimum achievable accuracy, and the average of individually fine-tuned models in Table 3 to indicate the maximum potential gains. Performance metrics were assessed using both the average absolute accuracy and the average normalized accuracy, calculated as detailed in Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As presented in Table 2, our method achieves state-of-the-art results across all benchmarks, regardless of the ViT size or the number of tasks involved. Notably, the most significant improvements were observed with the smaller ViT-B-32 encoder, where our approach outperforms Task Arithmetic and Consensus TA by an average of \\(+15.45\\%\\) and \\(+10.71\\%\\) absolute accuracy, respectively. Furthermore, when utilizing the ViT-L-14 model, our method attains an average normalized accuracy of \\(96.98\\%\\). This indicates that we can effectively replace eight individual task-specific models with a single merged model, incurring only a \\(-3.02\\%\\) reduction in average accuracy. These results highlight the promise of our model merging technique as a cost-efficient alternative to mixtures of experts, multi-task learning, and assembling methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.236, + 0.818, + 0.252 + ], + "angle": 0, + "content": "We report in Fig. 4 the per-dataset accuracies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.261, + 0.707, + 0.277 + ], + "angle": 0, + "content": "5.2. Compression results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.907, + 0.57 + ], + "angle": 0, + "content": "In Table 3 we report the results for our compression algorithm TSV-C. We compare our method with TALL-masks [46], which stores very sparse binary masks for each task. As shown in Table 3, our method always retains more than \\(99\\%\\) of the original accuracy for all considered benchmarks and models. The results of TSV-C are comparable to those of TALL-masks [46], with the former slightly underperforming TALL-masks for small ViTs and the latter gaining the upper hand in the larger-scale one. In general, however, the two approaches differ by less than \\(1\\%\\) accuracy on average. Regarding the storage, our method stores only the top \\(\\frac{1}{T}\\) singular vectors per task, which leads to a fixed storage requirement of approximately twice the size of the original model, regardless of the number of tasks. By storing binary masks, TALL-masks instead results in a storage size that varies with their compressibility, increasing with the number of tasks. This variability can lead to unpredictable storage requirements and may diminish compression benefits when masks are less compressible." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.907, + 0.647 + ], + "angle": 0, + "content": "Our results show that TSV-C effectively balances compression and performance by leveraging the most significant TSVs. Maintaining near-original accuracy, our approach is particularly suitable for scenarios where storage constraints are critical but high model performance is still required." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.661, + 0.611, + 0.679 + ], + "angle": 0, + "content": "6. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.779 + ], + "angle": 0, + "content": "In this section, we begin by conducting an ablation study to assess the contributions of interference reduction and low-rank compression to the overall performance. Next, we study how task interference varies across layers of different depths. Finally, we empirically show that our method does not require tuning a scaling coefficient." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.664, + 0.804 + ], + "angle": 0, + "content": "6.1. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "TSV-M combines low-rank approximation with task interference reduction. To evaluate the contribution of each component, we conducted an ablation study summarized in Table 4. Applying low-rank approximation alone to the layer task matrices results in worse performance than vanilla Task Arithmetic, shown in the first row where" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18700" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.165 + ], + "angle": 0, + "content": "
MethodViT-B-32ViT-B-16ViT-L-14
8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks
Finetuned92.83(100)90.88(100)91.37(100)94.64(100)92.76(100)93.17(100)95.81(100)94.29(100)94.73(100)
TALL-Mask+TIES93.13(100.37)90.92(100.04)91.11(99.70)94.68(100.04)92.69(99.90)93.05(99.87)95.96(100.15)93.40(99.09)93.91(99.16)
TSV-C (Ours)92.62(99.74)90.29(99.28)90.64(99.14)94.47(99.79)92.25(99.41)92.53(99.27)95.68(99.85)94.04(99.72)94.42(99.66)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.175, + 0.908, + 0.218 + ], + "angle": 0, + "content": "Table 3. Average absolute accuracy results across all compression benchmarks for different models and varying number of tasks, subscript (in parentheses) the normalized average accuracies. Our TSV-C compression method consistently achieves over \\(99\\%\\) of the original fine-tuned models' accuracy while significantly reducing storage requirements." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.24, + 0.481, + 0.334 + ], + "angle": 0, + "content": "
Low-rank approx.Interf. reductionViT-B-32
8 tasks14 tasks20 tasks
××76.5 (+0.0)72.1 (+0.0)66.8 (+0.0)
✓×75.2 (-1.3)71.0 (-1.1)66.3 (-0.5)
×✓82.6 (+7.4)75.7 (+4.7)69.9 (+3.6)
✓✓92.3 (+9.7)87.9 (+12.2)84.3 (+14.4)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.343, + 0.483, + 0.4 + ], + "angle": 0, + "content": "Table 4. Comparison of different versions of Task Arithmetic, comprising either the low-rank approximation step, the interference reduction step, or both. The method performing both corresponds to the proposed TSV-Merge." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.416, + 0.482, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.571, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Figure 5. Approximation error from the orthogonalization of the TSVs through Procrustes for the ViT-B-32 model across 8 tasks. The violin plots represent layer-wise approximation error distributions for \\( U \\) and \\( V \\) in both full-rank and low-rank cases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.748 + ], + "angle": 0, + "content": "neither component is applied. In contrast, applying task interference reduction alone, while keeping full-rank matrices, significantly improves performance, with gains ranging from \\(+3.1\\%\\) to \\(+6.1\\%\\). However, the best results are achieved only when both steps are combined, yielding substantial performance improvements of \\(+15.8\\%\\) to \\(+17.5\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.869 + ], + "angle": 0, + "content": "To explain why interference reduction applied to low-rank approximations in TSV-M outperforms its application to full-rank matrices, we analyze the errors introduced during the orthogonalization of the full-rank \\(U\\) and \\(V\\) matrices. As shown in Fig. 5, orthogonalizing full-rank matrices incurs significant approximation errors, measured by the sum of the Frobenius norms of reconstruction errors for the concatenated \\(U\\) and \\(V\\) matrices across all layers." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For example, in the ViT-B-32 model with 8 tasks, the full-rank setting exhibits a wider error distribution with a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.905, + 0.35 + ], + "angle": 0, + "content": "higher average, indicating greater variability and larger approximation discrepancies. In contrast, the low-rank setting produces a more compact and lower error distribution, suggesting better consistency in approximation across layers. In the following, we prove that this is not specific to the chosen model but a general property of our approach. The proof can be found in Appendix C.3." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.354, + 0.907, + 0.466 + ], + "angle": 0, + "content": "Theorem 6.1. Let \\( T \\in \\mathbb{N} \\) such that \\( T > 4 \\). Define \\( U = [U_1, \\ldots, U_T] \\) as the matrix obtained by concatenating \\( T \\) orthogonal matrices \\( U_i \\), each of shape \\( n \\times n \\). Let \\( \\widehat{U} = [\\widehat{U}_1, \\ldots, \\widehat{U}_T] \\) be the matrix formed by truncating each \\( U_i \\) to its first \\( k \\) columns. Denote by \\( X \\) and \\( \\widehat{X} \\) the matrices resulting from Procrustes orthonormalization of \\( U \\) and \\( \\widehat{U} \\), respectively. If \\( k \\leq n\\frac{T - 2\\sqrt{T}}{T} \\), then" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.47, + 0.802, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\left\\| U - X \\right\\| _ {F} \\geq \\left\\| \\widehat {U} - \\widehat {X} \\right\\| _ {F}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.906, + 0.647 + ], + "angle": 0, + "content": "Intuitively, the theorem asserts that the approximation error introduced by orthogonalization via Procrustes is smaller when applied to the concatenation of truncated matrices, compared to the concatenation of the original full-rank matrices, provided the rank of the truncated matrices satisfies certain conditions. To better understand the statement of the theorem, consider the case \\( T = 10 \\), as in our scenario. The term \\( \\frac{T - 2\\sqrt{T}}{T} = \\frac{10 - 2\\sqrt{10}}{10} \\geq \\frac{1}{3} \\). In our case, we choose \\( k \\approx \\frac{n}{T} \\). For \\( T = 10 \\), this gives \\( k \\approx \\frac{n}{10} \\), which is indeed smaller than \\( \\frac{n}{3} \\), satisfying the condition \\( k \\leq \\frac{1}{3} n \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.906, + 0.736 + ], + "angle": 0, + "content": "We emphasize that other orthogonalization methods, such as Gram-Schmidt, are ineffective in this context because they do not preserve the overall structure of the original data. Gram-Schmidt sequentially orthogonalizes vectors without minimizing deviation from the original set, potentially resulting in significant distortions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.906, + 0.842 + ], + "angle": 0, + "content": "Finally, the substantial reduction in approximation error suggests that, while the advantages of interference reduction are considerable, they may be offset by the errors introduced when applied to full-rank matrices. In contrast, starting with low-rank approximations captures the essential information of each task, making the orthogonalization step less costly in terms of approximation error." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.755, + 0.865 + ], + "angle": 0, + "content": "6.2. Per-layer task interference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We analyze task interference on a per-layer basis using Eq. (6). As shown in Fig. 6, interference is highest in the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18701" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.476, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.246, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Figure 6. Singular Task Interference (STI) across layers in a ViT-B-32 for 20 tasks. STI is high in early layers sharing common knowledge and lower in more specialized ones." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.304, + 0.48, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.498, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Figure 7. Best average normalized accuracy for different alpha values. The vertical labels indicate different sets of 8 tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.688 + ], + "angle": 0, + "content": "initial transformer layers and decreases significantly in the deeper layers. This aligns with the understanding that early, generalized layers capture common features across tasks, increasing the potential for conflict, while deeper layers are more specialized for specific tasks, reducing shared representation and interference. For brevity, we grouped layers within each transformer block in our analysis; each \"Layer \\( n \\)\" in Fig. 6 includes two attention matrices and two MLP matrices. A layer-by-layer analysis is provided in Fig. 14." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.416, + 0.714 + ], + "angle": 0, + "content": "6.3. Choice of the interpolation coefficient" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.84 + ], + "angle": 0, + "content": "The aggregation in Eq. (2) involves a scaling coefficient \\(\\alpha\\), typically tuned using a validation set, that can have a significant impact on the accuracy of the final model. However, as shown in Fig. 7, our approach consistently achieves the best results with \\(\\alpha = 1.0\\), indicating that no additional scaling is necessary. Although purely empirical, this finding spares further tuning and validation data, thereby enhancing the practicality and efficiency of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.405, + 0.864 + ], + "angle": 0, + "content": "6.4. Effect of task interference reduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We report in Fig. 8 the interference of a ViT-B-32 model both before and after the application of our TSV-M pipeline," + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.095, + 0.868, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.247, + 0.905, + 0.289 + ], + "angle": 0, + "content": "Figure 8. Singular Task Interference (STI) and average normalized accuracy for Task Arithmetic and TSV-Merge on the ViT-B-32 model, evaluated across merges of 8, 14, and 20 tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.905, + 0.412 + ], + "angle": 0, + "content": "assessed across three distinct sets of tasks with different cardinalities. In all instances, we observed that a reduction in interference is closely associated with a significant gain in accuracy. This provides empirical evidence for the effectiveness of our approach in minimizing interference and, ultimately, enhancing the merging process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.43, + 0.64, + 0.446 + ], + "angle": 0, + "content": "7. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.905, + 0.517 + ], + "angle": 0, + "content": "In this paper, we tackled the problem of model merging by analyzing the SVD of per-layer task matrices, confirming their inherent low-rank structure and leveraging their singular vectors to define task interference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.686 + ], + "angle": 0, + "content": "Building on the former insight, we introduced TSV-Compress (TSV-C), a model compression algorithm that reduces task vectors to \\(10\\%\\) of their original size while retaining \\(99\\%\\) of the original accuracy. Unlike existing methods, our approach maintains a constant storage requirement, independent of the number of tasks. We then developed TSV-Merge (TSV-M), combining low-rank approximation with interference reduction to create a novel model merging technique. Achieving normalized accuracy up to \\(97\\%\\), our method offers a storage-efficient alternative to MoEs, ensembles, and multi-task learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.905, + 0.779 + ], + "angle": 0, + "content": "Our extensive evaluation revealed that while interference reduction contributes most to performance improvements, the optimal results are achieved in combination with compression. We also observed that task interference decreases in deeper layers, and importantly, that our method does not require tuning a scaling coefficient." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Exploring alternative methods for task importance and rank approximation could be a valuable direction for future work. Currently, we use a uniform rank across tasks, approximating each task vector to \\(\\frac{1}{T}\\) of the layer's dimensionality. However, techniques like those from Idelbayev and Carreira-Perpinan [21] or optimal hard thresholds from Gavish and Donoho [17] could be applied to select the optimal rank for each task individually." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18702" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.226, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.111, + 0.486, + 0.248 + ], + "angle": 0, + "content": "This work was partially supported by PRIN 2022 project 20227YET9B \"AdVVent\" CUP code B53D23012830006 and by projects FAIR (PE0000013) and SERICS (PE00000014) under the MUR National Recovery and Resilience Plan funded by the European Union - NextGenerationEU. This work was moreover partially supported by the project NEREO (Neural Reasoning over Open Data) project funded by the Italian Ministry of Education and Research (PRIN) Grant no. 2022AEFHAZ." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.263, + 0.188, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.287, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.346, + 0.484, + 0.401 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - Mining Discriminative Components with Random Forests. In Computer Vision - ECCV 2014, pages 446-461, Cham, 2014. Springer International Publishing. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[3] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 105(10):1865-1883, 2017. Conference Name: Proceedings of the IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[4] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing Textures in the Wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, Columbus, OH, USA, 2014. IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[5] Tarin Clanuwat, Mikel Bober-Irizar, Asanobu Kitamoto, Alex Lamb, Kazuaki Yamamoto, and David Ha. Deep learning for classical japanese literature. CoRR, abs/1812.01718, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[6] Adam Coates, Andrew Ng, and Honglak Lee. An Analysis of Single-Layer Networks in Unsupervised Feature Learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. ISSN: 1938-7228. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[7] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and André van Schaik. EMNIST: Extending MNIST to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 2921-2926, 2017. ISSN: 2161-4407. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.745, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[8] Donato Crisostomi, Marco Fumero, Daniele Baieri, Florian Bernard, and Emanuele Rodola. \\(c^2 m^3\\): Cycle-consistent multi-model merging. In Advances in Neural Information Processing Systems, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[9] Nico Daheim, Thomas Möllenhoff, Edoardo Ponti, Iryna Gurevych, and Mohammad Emtiyaz Khan. Model merging by uncertainty-based gradient matching. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.287, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "masks. In European Conference on Computer Vision, pages 270-287. Springer, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[11] Misha Denil, Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato, and Nando De Freitas. Predicting parameters in deep learning. Advances in neural information processing systems, 26, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[12] Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. Advances in neural information processing systems, 27, 2014. 3, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.904, + 0.357 + ], + "angle": 0, + "content": "[14] Carl Eckart and G. Marion Young. The approximation of one matrix by another of lower rank. Psychometrika, 1:211-218, 1936. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.36, + 0.904, + 0.427 + ], + "angle": 0, + "content": "[15] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Workshop Track Proceedings, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[16] Timur Garipov, Dmitry Podoprikhin, Alexander Novikov, and Dmitry Vetrov. Ultimate tensorization: compressing convolutional and fc layers alike. arXiv preprint arXiv:1611.03214, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[17] Matan Gavish and David L. Donoho. The optimal hard threshold for singular values is \\(4 / \\sqrt{3}\\). IEEE Transactions on Information Theory, 60(8):5040-5053, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.527, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[18] Ian J. Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, Yingbo Zhou, Chetan Ramaiah, Fangxiang Feng, Ruifan Li, Xiaojie Wang, Dimitris Athanasakis, John Shawe-Taylor, Maxim Milakov, John Park, Radu Ionescu, Marius Popescu, Cristian Grozea, James Bergstra, Jingjing Xie, Lukasz Romaszko, Bing Xu, Zhang Chuang, and Yoshua Bengio. Challenges in Representation Learning: A Report on Three Machine Learning Contests. In Neural Information Processing, pages 117-124, Berlin, Heidelberg, 2013. Springer. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.68, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. Conference Name: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[20] Chenyu Huang, Peng Ye, Tao Chen, Tong He, Xiangyu Yue, and Wanli Ouyang. EMR-merging: Tuning-free high-performance model merging. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Yerlan Idelbayev and Miguel A Carreira-Perpinán. Low-rank compression of neural nets: Learning the rank of each layer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8049–8059, 2020. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18703" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[22] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[23] Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. In *British Machine Vision Conference*, BMVC 2014, Nottingham, UK, September 1-5, 2014. BMVA Press, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[24] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[25] Keller Jordan, Hanie Sedghi, Olga Saukh, Rahim Entezari, and Behnam Neyshabur. REPAIR: REnormalizing permuted activations for interpolation repair. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[26] Yong-Deok Kim, Eunhyeok Park, Sungjoo Yoo, Taelim Choi, Lu Yang, and Dongjun Shin. Compression of deep convolutional neural networks for fast and low power mobile applications. In Proc. of the 4th Int. Conf. Learning Representations (ICLR 2016), San Juan, Puerto Rico, May 2–4, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[27] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D Object Representations for Fine-Grained Categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, Sydney, Australia, 2013. IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[29] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[31] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. Advances in Neural Information Processing Systems, 37:78905-78935, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[32] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 35:17703-17716, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[34] Maria-Elena Nilsback and Andrew Zisserman. Automated Flower Classification over a Large Number of Classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[35] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. ISSN: 1063-6919. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[36] Fidel A Guerrero Pña, Heitor Rapela Medeiros, Thomas Dubail, Masih Aminbeidokhti, Eric Granger, and Marco Pedersoli. Re-basin via implicit sinkhorn differentiation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[38] Alexandre Rame, Matthieu Kirchmeyer, Thibaud Rahier, Alain Rakotomamonjy, Patrick Gallinari, and Matthieu Cord. Diverse weight averaging for out-of-distribution generalization. Advances in Neural Information Processing Systems, 35:10821-10836, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[39] Noam Shazeer, *Azalia Mirhoseini*, *Krzysztof Maziarz*, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[40] Li Shen, Anke Tang, Enneng Yang, Guibing Guo, Yong Luo, Lefei Zhang, Xiaochun Cao, Bo Du, and Dacheng Tao. Efficient and effective weight-ensembling mixture of experts for multi-task model merging. arXiv preprint arXiv:2410.21804, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[41] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA, 2013. Association for Computational Linguistics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[42] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The German Traffic Sign Recognition Benchmark: A multi-class classification competition. In *The 2011 International Joint Conference on Neural Networks*, pages 1453–1460, 2011. ISSN: 2161-4407. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.69 + ], + "angle": 0, + "content": "[43] George Stoica, Daniel Bolya, Jakob Brandt Bjorner, Pratik Ramesh, Taylor Hearn, and Judy Hoffman. Zipit! merging models from different tasks without training. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[44] Anke Tang, Li Shen, Yong Luo, Shuai Xie, Han Hu, Lefei Zhang, Bo Du, and Dacheng Tao. Smile: Zero-shot sparse mixture of low-rank experts construction from pre-trained foundation models. arXiv preprint arXiv:2408.10174, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[45] Bastiaan S. Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation Equivariant CNNs for Digital Pathology. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2018, pages 210–218, Cham, 2018. Springer International Publishing. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[46] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. In *Forty-first International Conference on Machine Learning*, 2024. 2, 5, 6, 12, 13" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18704" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[47] B. P. Welford. Note on a method for calculating corrected sums of squares and products. Technometrics, 4(3):419-420, 1962. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.244 + ], + "angle": 0, + "content": "[48] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.249, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[49] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms, 2017. arXiv:1708.07747 [cs, stat]. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.291, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[50] Jianxiong Xiao, Krista A. Ehinger, James Hays, Antonio Torralba, and Aude Oliva. SUN Database: Exploring a Large Collection of Scene Categories. International Journal of Computer Vision, 119(1):3-22, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.348, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[51] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.404, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[52] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.475, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[53] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.532, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[54] Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Fabrizio Silvestri, and Emanuele Rodola. Atm: Improving model merging by alternating tuning and merging, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18705" + } + ] +] \ No newline at end of file diff --git a/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_origin.pdf b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..71f0becd1705b736063a0c660046c753253065c8 --- /dev/null +++ b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/a88c8ab6-3362-4dfb-a40f-5aa669be1c2b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb4ee846ccad43a0493063166a0469ccbf3f57dc05537907b293b8bd42b76f0f +size 1116992 diff --git a/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/full.md b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1d5a01166a651b448fb73dd74644a31b9f907735 --- /dev/null +++ b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/full.md @@ -0,0 +1,362 @@ +# Task Singular Vectors: Reducing Task Interference in Model Merging + +Antonio Andrea Gargiulo\* Simone Scardapane + +Donato Crisostomi* +Fabrizio Silvestri + +Maria Sofia Bucarelli Emanuele Rodolà + +Sapienza University of Rome + +gargiulo.1769185@studenti.uniroma1.it {crisostomi, rodola}@di.uniroma1.it simone.scardapane@uniroma1.it {bucarelli, fsilvestri}@diag.uniroma1.it + +# Abstract + +Task Arithmetic has emerged as a simple yet effective method to merge models without additional training. However, by treating entire networks as flat parameter vectors, it overlooks key structural information and is susceptible to task interference. In this paper, we study task vectors at the layer level, focusing on task layer matrices and their singular value decomposition. In particular, we concentrate on the resulting singular vectors, which we refer to as Task Singular Vectors (TSV). Recognizing that layer task matrices are often low-rank, we propose TSV-Compress (TSV-C), a simple procedure that compresses them to $10\%$ of their original size while retaining $99\%$ of accuracy. We further leverage this low-rank space to define a new measure of task interference based on the interaction of singular vectors from different tasks. Building on these findings, we introduce TSV-Merge (TSV-M), a novel model merging approach that combines compression with interference reduction, significantly outperforming existing methods. + +# 1. Introduction + +The widespread availability of pre-trained models and public repositories has driven the development of techniques for efficiently combining and reusing existing models. Among these techniques, model merging approaches enable the creation of multi-task models without additional training. One popular approach, Task Arithmetic (TA) [22], stands out for its simplicity and effectiveness. However, by treating entire networks as high-dimensional vectors, TA and subsequent works overlook crucial structural information [9, 10, 24, 32, 38, 51, 53]. This flattened view limits these approaches to coarse-grained measures like cosine similarity for assessing inter-task interactions. + +![](images/a112e20c2a5876d992350b71032a1805301920526a934c81d240c758be68c1b7.jpg) +Figure 1. Mean accuracy of a ViT-L-14 merged over 8, 14, and 20 tasks respectively. By significantly surpassing existing methods, TSV-M establishes the new state of the art in model merging. + +In this work, we instead focus on preserving and leveraging the natural structure of neural networks by examining weight differences at the layer level while retaining their matrix form wherever possible. Our approach begins by examining per-layer task matrices, which represent weight changes for each task, through singular value decomposition (SVD). This decomposition yields singular vectors and singular values that capture the most significant directions of variation within each layer. We term these singular vectors Task Singular Vectors (TSV), as they provide an interpretable basis for assessing task-specific contributions at each layer. Importantly, in accordance with recent literature on PEFT (cfr. Sec. 2), our analysis confirms that task matrices are inherently low-rank, meaning that only a small portion of TSVs is sufficient to represent the layer's function with high fidelity. + +Building on this insight, we introduce TSV-Compress (TSV-C), a simple yet effective procedure to compress task vectors down to $10\%$ of their original size while maintaining $99\%$ of the original accuracy. Focusing on only a fraction of the most relevant TSVs for each task, TSV-Compress enables efficient storage and processing + +of task-specific information without waiving performance. + +Beyond compression, examining the interplay of TSVs across different tasks provides a geometrically informed framework for analyzing task interference at the individual layer level. By assessing how singular vectors from different tasks align or diverge within each layer, this approach offers a significantly more fine-grained understanding of inter-task interactions, going beyond global vector similarity metrics like cosine similarity. Building upon these principles, we introduce TSV-Merge (TSV-M), a novel model merging method that combines compression and task interference reduction. This is achieved by discarding the irrelevant singular vectors for each task and then reducing their inter-task interference with a whitening transformation over their similarity matrix. We empirically show our approach to effectively reduce task interference, and the reduction to be complementary to the compression step. We then evaluate our approach across several benchmarks and show it outperforms existing methods by an average of $\sim 15\%$ accuracy points, establishing the new state of the art by a large margin. We release all our code for research purposes1. In summary, our contribution is 4-fold: + +- We study the singular value decomposition of per-layer task matrices, showing them to be low-rank and measuring the interplay of singular vectors from different tasks; +- We introduce TSV-C, a method that builds on this insight to compress task vectors by a factor of $10 \times$ while preserving $99\%$ of their original accuracy; +- We propose TSV-M, a model merging technique that complements compression with a task interference reduction step by applying a whitening transformation to decorrelate singular vectors across tasks; +- We conduct extensive experiments on multiple computer vision datasets, showing that TSV-M significantly outperforms existing model merging methods and provide in-depth analyses to uncover the underlying factors contributing to its success. + +# 2. Related Work + +Model Merging offers an efficient alternative to ensembling by combining existing models without the need for further training. Several approaches address this by determining the neuron permutations that align the models into a shared optimization basin, after which the models are combined using a straightforward averaging method. [1, 8, 25, 36, 43]. An alternative line of approaches focuses on the multi-task scenario, where a single pre-trained model is fine-tuned for different tasks [10, 22, 32, 40, 46, 48, 51, 53, 54]. The prerequisites for methods in this category are outlined in Table 1. Task Arithmetic [22] introduces the concept of task vectors, which are the weight differences + +Table 1. Model merging approaches and their requirements. + +
MethodAdditional trainingExtra storageValidat. data inputsdata labels
Weight Avg. [48]××××
Fisher-Merg. [32]××✓×
RegMean [24]××✓×
EMR-Merging [20]×✓××
TALL-Mask [46]×✓××
TSV-C (Ours)×✓××
Task Arith. [22]××✓✓
Ties-Merging [51]××✓✓
AdaMerging [52]✓×✓×
Consensus-TA [46]××✓✓
TSV-M (Ours)××××
+ +between fine-tuned models and the pre-trained base model. By averaging these vectors, a merged multi-task model can be created; conversely, negating a task vector allows for forgetting a specific task. TIES [51] addresses redundancies in model parameters by first selecting the top- $k$ most significant parameter changes and then constructing a sign vector based on the majority sign across all models. The latter is used to merge the task vectors disjointly, meaning the average is not computed when a parameter is zero or when parameters disagree in sign. Similarly, DARE [53] randomly resets redundant parameters to their pre-trained values and rescales the remaining parameters by a factor proportional to the dropped ones, aiming to reduce interference among tasks. Fisher Merging [32] and RegMean [24] merge models by performing weighted averaging, utilizing the Fisher information matrix and the inner product of input vectors, respectively. Model Broadcrumbs [10] focuses on merging only significant weights by discarding outliers and both minor and large perturbations in the fine-tuned parameters. More recently, Wang et al. [46] observed that tasks often utilize non-overlapping sets of weights. They propose two methods: the first, TALL-Mask, uses binary masks to activate important task-specific weights, requiring extra storage (as does our method TSV-C) for the masks. The second, Consensus Merging, leverages these masks to remove parameters that are important to less than two tasks. Like our approach, TwinMerging [31] and SMILE [44] apply SVD to layer task matrices. However, neglecting singular vector interference, they require a router to selectively activate a subset during inference. + +Unlike existing approaches, our methods explicitly address task interference by leveraging the geometric structure of singular vectors to minimize interactions between task-specific parameters. Rather than averaging or selectively merging parameters as prior techniques do, we decorrelate singular vectors to directly reduce interference across tasks. + +Additionally, we achieve targeted compression by isolating each task's unique components, storing only essential parts, and ensuring minimal interference, a step beyond traditional SVD-based compression approaches. + +SVD for Model Compression A significant body of work explores low-rank decompositions for fully-connected and convolutional layers [11, 12, 16, 23, 27], as well as tensor decompositions [12, 16, 26]. These approaches typically achieve a layer-by-layer factorization of a trained network, focusing on minimizing the difference between the original and the low-rank approximated weight matrices. This is accomplished using techniques like SVD [12, 16, 23, 27] or iterative optimization [23, 26]. While other methods such as weight quantization, knowledge distillation, and pruning exist, we utilize SVD not only for low-rank approximation and compression but also because its matrix decomposition properties make it particularly effective in analyzing and mitigating interference. + +# 3. Task Singular Vectors + +In this section, we introduce key concepts for understanding our approach, including a new measure of task interference that is central to our method. + +# 3.1. Background + +We build on Task Arithmetic (TA), which defines task vectors capturing the differences in model weights for individual tasks. Formally, the weights $\theta_{\mathrm{MT}}$ of a multi-task model for $T$ tasks are computed by aggregating the task-specific weight differences, or task vectors, as follows: + +$$ +\theta_ {\mathrm {M T}} = \theta_ {\mathrm {p r e}} + \alpha \frac {\sum_ {i = 1} ^ {T} \tau_ {i}}{T}, \tag {1} +$$ + +where $\theta_{\mathrm{pre}}$ is the set of pretrained model weights, $\alpha$ is a scaling factor, and $\tau_{i} = \theta_{\mathrm{ft}_{i}} - \theta_{\mathrm{pre}}$ is the task vector for task $i$ , with $\theta_{\mathrm{ft}_i}$ being the fine-tuned weights for the task. Differently from TA, however, we consider these operations at the layer level. From this perspective, Eq. (1) becomes + +$$ +\theta_ {\mathrm {M T}} ^ {(l)} = \theta_ {\text {p r e}} ^ {(l)} + \alpha \frac {\sum_ {i = 1} ^ {T} \Delta_ {i} ^ {(l)}}{T}, \tag {2} +$$ + +where $\theta_{\mathrm{pre}}^{(l)}$ encodes the pretrained weights for layer $l$ , and $\Delta_i^{(l)} = \theta_{\mathrm{ft}_i}^{(l)} - \theta_{\mathrm{pre}}^{(l)}$ is the task-specific weight difference for task $i$ at layer $l$ . When layer $l$ has a matrix structure, we call its corresponding $\Delta_i^l$ the per-layer task matrix for task $i$ . When this is not the case, our framework defaults to standard TA. For brevity, we will generally omit the layer index and refer to the layer- $l$ task matrix $\Delta_i^l$ as $\Delta_i$ . + +# Decomposing layer task matrices + +Treating layer-wise weights as structured entities rather than flattened vectors enables us to analyze their SVD, revealing low-rank properties and deeper insight into the inter-task interactions. Given two tasks $i,j$ , we consider the SVD of their task matrices $\Delta_{i}$ and $\Delta_{j}$ at a generic layer: + +$$ +\Delta_ {i} = U _ {i} \Sigma_ {i} V _ {i} ^ {\top}, \qquad \Delta_ {j} = U _ {j} \Sigma_ {j} V _ {j} ^ {\top} +$$ + +where $U_{i}, U_{j}$ and $V_{i}, V_{j}$ are the matrices of left and right singular vectors respectively, and $\Sigma_{i}, \Sigma_{j}$ are diagonal matrices of singular values. Due to their role in assessing task-specific contributions, we term the obtained singular vectors Task Singular Vectors (TSV). + +Importantly for our treatment, we can equivalently write the aggregated task matrices in Eq. (2) in terms of their singular vectors and values. By defining $U = [U_{1}\dots U_{T}]$ as the column-wise concatenation of all the left TSVs, $V = [V_{1}\dots V_{T}]$ as the row-wise concatenation of the right TSVs, and $\Sigma$ as the block diagonal matrix with $\{\Sigma_i\}_{i=1}^T$ along its diagonal, we can write: + +$$ +M = U \Sigma V ^ {\top} = \sum_ {i = 1} ^ {T} U _ {i} \Sigma_ {i} V _ {i} ^ {\top} = \sum_ {i = 1} ^ {T} \Delta_ {i}. \tag {3} +$$ + +In the simple case where $T = 2$ , $M$ would be given by: + +$$ +M = \left[ \begin{array}{c c} U _ {1} & U _ {2} \end{array} \right] \left[ \begin{array}{c c} \Sigma_ {1} & 0 \\ 0 & \Sigma_ {2} \end{array} \right] \left[ \begin{array}{c} V _ {1} ^ {\top} \\ V _ {2} ^ {\top} \end{array} \right] = \Delta_ {1} + \Delta_ {2}. +$$ + +When concatenating singular components from different tasks, we violate certain properties of the SVD. Specifically, the matrices $U$ and $V$ become non-orthogonal because singular vectors from different tasks may overlap. Additionally, the singular values $\Sigma_{i}$ from different tasks can vary significantly in magnitude, which may bias the merging process toward tasks with larger singular values. + +# 3.2. Low-rank nature of layer task matrices + +We start our analysis by studying the low-rank properties of per-layer task matrices. Interpreting SVD as a sum of rank-one matrices, for each task $i$ , by Eckart-Young's theorem [14] we get the best approximation (in Frobenius norm) of each task matrix $\Delta_{i}$ by retaining only the top- $k$ singular values and their corresponding vectors: + +$$ +\hat {\Delta} _ {i} = \sum_ {j = 1} ^ {k} \sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \top}. \tag {4} +$$ + +As we show in Fig. 2, we find task matrices to be inherently low-rank: a small subset of TSVs is sufficient to represent the layer's function with high fidelity. In particular, even when preserving only $3\%$ of the singular components per task, the mean accuracy drops by merely $1.5\%$ . This is + +![](images/1842bab04fc51cc70444b101f9ef121909f45317df0acb0cf944e33b85c2ec50.jpg) +Figure 2. Mean absolute accuracy of the ViT-B-32 model across increasing fractions of retained singular components, averaged over 20 tasks. The red line represents the average accuracy of the original fine-tuned models with full-rank task matrices, while the green line shows the accuracies using low-rank approximations. + +noticeable considering that $97\%$ of singular components in each layer matrix are discarded. Building on this insight, we propose in Sec. 4.1 a compression algorithm that maintains $99\%$ of the accuracy while shrinking the task vectors to $10\%$ of their original size. + +Given this low-rank structure, it is natural to aggregate task matrices within their subspaces. We, therefore, obtain a reduced version of the aggregation matrix $M$ (Eq. (3)) as: + +$$ +\hat {M} = \sum_ {i = 1} ^ {T} \hat {U} _ {i} \hat {\Sigma} _ {i} \hat {V} _ {i} ^ {\top} = \sum_ {i = 1} ^ {T} \sum_ {j = 1} ^ {k} \sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \top}, \tag {5} +$$ + +where $\hat{U}_i$ and $\hat{V}_i$ contain the top- $k$ left and right singular vectors for task $i$ , respectively, and $\hat{\Sigma}_i$ is the diagonal matrix of the top- $k$ singular values $\sigma_j^i$ . This formulation highlights that the low-rank approximation $\hat{M}$ utilizes the most relevant singular components from each task to effectively combine the task-specific weight differences. + +# 3.3. Singular Task interference + +We hereby introduce a score of task interference based on the interplay of TSVs from different tasks, which we term Singular Task Interference (STI) + +$$ +\operatorname {S T I} \left(\left\{\Delta_ {i} \right\} _ {i = 1} ^ {T}\right) = \left\| \left(U ^ {\top} U - I\right) \Sigma \left(V ^ {\top} V - I\right) \right\| _ {1}, \tag {6} +$$ + +where $U, \Sigma$ and $V$ are obtained by concatenating the singular value decompositions $\left(\{U_i\}_{i=1}^T, \{\Sigma_i\}_{i=1}^T, \{V_i\}_{i=1}^T\right)$ of per-layer task matrices $\{\Delta_i\}_{i=1}^T$ as detailed in Sec. 3.1. In the expression above, high inner product values for $U^\top U$ and $V^\top V$ imply a higher likelihood of interference, with minimal interference ideally yielding identity matrices. + +The underlying intuition is that overlapping singular vectors suggest shared features in the weight space across tasks. Such overlap can introduce interference when models are merged, ultimately degrading performance on individual tasks. We refer to Fig. 3 for a real example over eight tasks. + +# 4. Approach + +We demonstrate below how TSVs can be used for both compression and task interference reduction. + +# 4.1. TSV for compression + +When the task is known or inferred (e.g. with a router, as in Mixture-of-Experts [15, 39] techniques), we can leverage the low-rank structure of per-layer task matrices (see Sec. 3.2) to effectively compress these ones while discarding task singular vectors from different tasks. In particular, given a known or inferred task index $h$ , we set other task components to zero, reducing Eq. (5) to: + +$$ +\hat {M} = \sum_ {i = 1} ^ {T} \mathbb {1} _ {[ i = h ]} \sum_ {j = 1} ^ {k} \sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \top} = \sum_ {j = 1} ^ {k} \sigma_ {j} ^ {h} u _ {j} ^ {h} v _ {j} ^ {h \top} = \hat {\Delta} _ {i}. \tag {7} +$$ + +This formula yields a low-rank approximation of $\Delta_h$ , where only the top $k$ singular components of the task-specific matrix are retained. Here, the same number of components $k$ is taken from each task to ensure that each layer is compressed by a factor of $\frac{1}{T}$ relative to the original dimension. Increasing $k$ improves approximation but reduces the compression. + +# 4.2. TSV for model merging + +In scenarios where the task identity is not known beforehand, we address the standard model merging problem by combining Task Singular Vectors (TSVs) to create a single model that performs well jointly across all tasks. + +To reduce our measure of task interference, we decorrelate the TSVs of different tasks (encoded as columns in $\hat{U}$ and $\hat{V}$ ) by whitening these matrices to minimize their correlations. This can be achieved by applying the transformation $X \mapsto X(X^{\top}X)^{-\frac{1}{2}}$ to both $\hat{U}$ and $\hat{V}$ . For improved numerical stability, we reformulate this whitening as an orthogonal Procrustes problem, seeking the orthogonal matrix $\hat{U}_{\perp}$ that minimizes the projection error: + +$$ +\min _ {\hat {U} _ {\perp}} \| \hat {U} _ {\perp} - \hat {U} \| _ {F} \quad \text {s . t .} \hat {U} _ {\perp} ^ {\top} \hat {U} _ {\perp} = I, \tag {8} +$$ + +and similarly for $\hat{V}$ . This problem admits a closed-form solution via the SVD $\hat{U} = PDQ^{\top}$ , yielding $\hat{U}_{\perp} = PQ^{\top}$ (similarly for $\hat{V}_{\perp}$ ). + +Proposition 4.1. The transformations $X \mapsto X(X^{\top}X)^{-\frac{1}{2}}$ (whitening) and $X \mapsto PQ^{\top}$ (Procrustes), where $X = PDQ^{\top}$ is the SVD of $X$ , are equivalent. + +Proof. Given $X = PDQ^{\top}$ and recalling that $D^{\top}D = D^{2}$ for diagonal matrices, simple algebraic manipulation yields $X^{\top}X = QD^{2}Q^{\top}$ . It follows that $(X^{\top}X)^{-1 / 2} = QD^{-1}Q^{\top}$ , therefore the whitening transformation can be rewritten as $X(X^{\top}X)^{-1 / 2} = (PDQ^{\top})(QD^{-1}Q^{\top}) = PDID^{-1}Q^{\top} = PQ^{\top}$ , completing the proof. + +![](images/b55846b13c15ccf51d7d4406a02a81453979021d81cb0dea9c2c5efdc8de5029.jpg) +Figure 3. Visualization of task interference among 8 tasks computed on the first attention layer of a ViT-B-32. The diagonal blocks display intra-task similarities, while the off-diagonal blocks illustrate inter-task similarities. The zoomed-in section highlights the interaction between the right singular vectors of the $3^{\mathrm{rd}}$ and $4^{\mathrm{th}}$ tasks. + +![](images/a2df729feab9265e61c1703f0e91034642d16d63b0c88e3aada49f8ef971ad59.jpg) + +![](images/f2ee7971f77a130fe96fe7cd91f91c9de2cd8b956b7cc7658366078b4c95042f.jpg) +Figure 4. Absolute accuracy of a ViT-B-32 merged over 8, 14, and 20 tasks, respectively. + +![](images/6c25b469be8f6f381fc43fbd4038701511e968c5dbc6cbde79245fcbec03f504.jpg) + +![](images/a8d2c040cacf957448513937376889eb2841050a2f486a69a521a0d8a3ac02a7.jpg) + +Once the (rank-reduced) TSV matrices $\hat{U}$ and $\hat{V}$ are decorrelated, we reconstruct the merged layer as in Eq. (5); see Algorithm 1 for the complete steps. + +# 5. Results + +We evaluate our approaches over three different suites of tasks having cardinality 8, 14, and 20, respectively. The first one, introduced in [22], consists of datasets: Cars [28], DTD [4], EuroSAT [19], GTSRB [42], MNIST [30], RESISC45 [3], SUN397 [50], and SVHN [33]. The benchmark with 14 tasks builds on the preceding one, incorporating six additional datasets: CIFAR100 [29], STL10 [6], Flowers102 [34], OxfordIIITPet [35], PCAM [45], and FER2013 [18]. Finally, the 20-tasks benchmark in + +cludes the preceding 14 plus the following six: EMNIST [7], CIFAR10 [29], Food101 [2], FashionMNIST [49], RenderedSST2 [41], and KMNIST [5]. + +# 5.1. Model merging results + +We evaluate our method on three variants of the CLIP [37] model, each employing a different size of ViT [13] visual encoder: ViT-B-32, ViT-B-16, and ViT-L-14. The main benchmarks involve merging 8, 14, and 20 tasks, mirroring the experimental setup described in Wang et al. [46]. We compare our approach against several training-free model merging methods, including weight averaging, Task Arithmetic [22], and Consensus Merging [46]. For reference, we include the performance of zero + +
MethodViT-B-32ViT-B-16ViT-L-14
8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks
Zeroshot48.26(53.59)57.21(63.69)56.10(62.41)55.34(59.34)61.28(66.19)59.73(64.52)64.70(68.00)68.20(72.15)65.23(68.99)
Weight Averaging66.34(72.13)64.34(71.12)61.04(67.53)72.22(76.60)69.46(74.82)65.31(70.36)79.56(83.15)76.73(81.10)71.60(75.60)
Task Arithmetic70.79(76.55)65.32(72.09)60.52(66.79)75.41(79.58)70.52(75.89)65.78(70.76)84.93(88.65)79.41(83.95)74.01(78.07)
Consensus TA75.03(80.84)70.39(77.36)65.43(71.98)79.39(83.86)74.39(79.92)69.76(74.93)86.34(90.08)82.22(86.94)79.00(83.22)
TSV-M (Ours)85.86(92.31)80.06(87.88)77.07(84.29)89.01(93.94)84.58(91.01)80.57(86.45)92.98(96.98)89.17(94.43)87.72(92.50)
+ +Table 2. Average absolute accuracy results on model merging benchmarks; subscript (in parentheses) is the normalized average accuracy. + +Algorithm 1 TSV-Merge. +Require: Task matrices $\Delta_1,\ldots ,\Delta_T$ , scaling factor $\alpha$ +Ensure: Merged model weights $\theta_{\mathrm{MT}}$ +1: for $i = 1$ to $T$ do +2: Compute SVD: $\Delta_i = U_i\Sigma_iV_i^\top$ +3: Retain first $\frac{1}{T}$ singular components of $U_{i},\Sigma_{i}$ , and $V_{i}$ +4: end for +5: Concatenate the matrices: +6: $U\gets [U_1|U_2|\dots |U_T]$ +7: $\Sigma$ - block-diag( $\Sigma_1,\Sigma_2,\dots ,\Sigma_T)$ +8: $V\gets [V_{1}|V_{2}|\dots |V_{T}]$ +9: Compute the SVD of $U$ and $V$ .. +10: $U = P_{U}D_{U}Q_{U}^{\top}\qquad V = P_{V}D_{V}Q_{V}^{\top}$ +11:Obtain the orthogonal matrices: +12: $U_{\perp} = P_{U}Q_{U}^{\top}\qquad V_{\perp} = P_{V}Q_{V}^{\top}$ +13: Reconstruct the merged matrix: +14: $\hat{M}\gets U_{\perp}\Sigma V_{\perp}^{\top}$ +15: Construct merged model weights: +16: $\theta_{\mathrm{MT}}\gets \theta_{\mathrm{pre}} + \alpha \hat{M}$ +17: return $\theta_{\mathrm{MT}}$ + +shot models in Table 2 to represent the minimum achievable accuracy, and the average of individually fine-tuned models in Table 3 to indicate the maximum potential gains. Performance metrics were assessed using both the average absolute accuracy and the average normalized accuracy, calculated as detailed in Appendix B.1. + +As presented in Table 2, our method achieves state-of-the-art results across all benchmarks, regardless of the ViT size or the number of tasks involved. Notably, the most significant improvements were observed with the smaller ViT-B-32 encoder, where our approach outperforms Task Arithmetic and Consensus TA by an average of $+15.45\%$ and $+10.71\%$ absolute accuracy, respectively. Furthermore, when utilizing the ViT-L-14 model, our method attains an average normalized accuracy of $96.98\%$ . This indicates that we can effectively replace eight individual task-specific models with a single merged model, incurring only a $-3.02\%$ reduction in average accuracy. These results highlight the promise of our model merging technique as a cost-efficient alternative to mixtures of experts, multi-task learning, and assembling methods. + +We report in Fig. 4 the per-dataset accuracies. + +# 5.2. Compression results + +In Table 3 we report the results for our compression algorithm TSV-C. We compare our method with TALL-masks [46], which stores very sparse binary masks for each task. As shown in Table 3, our method always retains more than $99\%$ of the original accuracy for all considered benchmarks and models. The results of TSV-C are comparable to those of TALL-masks [46], with the former slightly underperforming TALL-masks for small ViTs and the latter gaining the upper hand in the larger-scale one. In general, however, the two approaches differ by less than $1\%$ accuracy on average. Regarding the storage, our method stores only the top $\frac{1}{T}$ singular vectors per task, which leads to a fixed storage requirement of approximately twice the size of the original model, regardless of the number of tasks. By storing binary masks, TALL-masks instead results in a storage size that varies with their compressibility, increasing with the number of tasks. This variability can lead to unpredictable storage requirements and may diminish compression benefits when masks are less compressible. + +Our results show that TSV-C effectively balances compression and performance by leveraging the most significant TSVs. Maintaining near-original accuracy, our approach is particularly suitable for scenarios where storage constraints are critical but high model performance is still required. + +# 6. Analysis + +In this section, we begin by conducting an ablation study to assess the contributions of interference reduction and low-rank compression to the overall performance. Next, we study how task interference varies across layers of different depths. Finally, we empirically show that our method does not require tuning a scaling coefficient. + +# 6.1. Ablation study + +TSV-M combines low-rank approximation with task interference reduction. To evaluate the contribution of each component, we conducted an ablation study summarized in Table 4. Applying low-rank approximation alone to the layer task matrices results in worse performance than vanilla Task Arithmetic, shown in the first row where + +
MethodViT-B-32ViT-B-16ViT-L-14
8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks
Finetuned92.83(100)90.88(100)91.37(100)94.64(100)92.76(100)93.17(100)95.81(100)94.29(100)94.73(100)
TALL-Mask+TIES93.13(100.37)90.92(100.04)91.11(99.70)94.68(100.04)92.69(99.90)93.05(99.87)95.96(100.15)93.40(99.09)93.91(99.16)
TSV-C (Ours)92.62(99.74)90.29(99.28)90.64(99.14)94.47(99.79)92.25(99.41)92.53(99.27)95.68(99.85)94.04(99.72)94.42(99.66)
+ +Table 3. Average absolute accuracy results across all compression benchmarks for different models and varying number of tasks, subscript (in parentheses) the normalized average accuracies. Our TSV-C compression method consistently achieves over $99\%$ of the original fine-tuned models' accuracy while significantly reducing storage requirements. + +
Low-rank approx.Interf. reductionViT-B-32
8 tasks14 tasks20 tasks
××76.5 (+0.0)72.1 (+0.0)66.8 (+0.0)
✓×75.2 (-1.3)71.0 (-1.1)66.3 (-0.5)
×✓82.6 (+7.4)75.7 (+4.7)69.9 (+3.6)
✓✓92.3 (+9.7)87.9 (+12.2)84.3 (+14.4)
+ +Table 4. Comparison of different versions of Task Arithmetic, comprising either the low-rank approximation step, the interference reduction step, or both. The method performing both corresponds to the proposed TSV-Merge. + +![](images/bf1bef04df4181e912a6d06a8efdbf544fc1f6b350bee695e433935858d44d98.jpg) +Figure 5. Approximation error from the orthogonalization of the TSVs through Procrustes for the ViT-B-32 model across 8 tasks. The violin plots represent layer-wise approximation error distributions for $U$ and $V$ in both full-rank and low-rank cases. + +neither component is applied. In contrast, applying task interference reduction alone, while keeping full-rank matrices, significantly improves performance, with gains ranging from $+3.1\%$ to $+6.1\%$ . However, the best results are achieved only when both steps are combined, yielding substantial performance improvements of $+15.8\%$ to $+17.5\%$ . + +To explain why interference reduction applied to low-rank approximations in TSV-M outperforms its application to full-rank matrices, we analyze the errors introduced during the orthogonalization of the full-rank $U$ and $V$ matrices. As shown in Fig. 5, orthogonalizing full-rank matrices incurs significant approximation errors, measured by the sum of the Frobenius norms of reconstruction errors for the concatenated $U$ and $V$ matrices across all layers. + +For example, in the ViT-B-32 model with 8 tasks, the full-rank setting exhibits a wider error distribution with a + +higher average, indicating greater variability and larger approximation discrepancies. In contrast, the low-rank setting produces a more compact and lower error distribution, suggesting better consistency in approximation across layers. In the following, we prove that this is not specific to the chosen model but a general property of our approach. The proof can be found in Appendix C.3. + +Theorem 6.1. Let $T \in \mathbb{N}$ such that $T > 4$ . Define $U = [U_1, \ldots, U_T]$ as the matrix obtained by concatenating $T$ orthogonal matrices $U_i$ , each of shape $n \times n$ . Let $\widehat{U} = [\widehat{U}_1, \ldots, \widehat{U}_T]$ be the matrix formed by truncating each $U_i$ to its first $k$ columns. Denote by $X$ and $\widehat{X}$ the matrices resulting from Procrustes orthonormalization of $U$ and $\widehat{U}$ , respectively. If $k \leq n\frac{T - 2\sqrt{T}}{T}$ , then + +$$ +\left\| U - X \right\| _ {F} \geq \left\| \widehat {U} - \widehat {X} \right\| _ {F}. +$$ + +Intuitively, the theorem asserts that the approximation error introduced by orthogonalization via Procrustes is smaller when applied to the concatenation of truncated matrices, compared to the concatenation of the original full-rank matrices, provided the rank of the truncated matrices satisfies certain conditions. To better understand the statement of the theorem, consider the case $T = 10$ , as in our scenario. The term $\frac{T - 2\sqrt{T}}{T} = \frac{10 - 2\sqrt{10}}{10} \geq \frac{1}{3}$ . In our case, we choose $k \approx \frac{n}{T}$ . For $T = 10$ , this gives $k \approx \frac{n}{10}$ , which is indeed smaller than $\frac{n}{3}$ , satisfying the condition $k \leq \frac{1}{3} n$ . + +We emphasize that other orthogonalization methods, such as Gram-Schmidt, are ineffective in this context because they do not preserve the overall structure of the original data. Gram-Schmidt sequentially orthogonalizes vectors without minimizing deviation from the original set, potentially resulting in significant distortions. + +Finally, the substantial reduction in approximation error suggests that, while the advantages of interference reduction are considerable, they may be offset by the errors introduced when applied to full-rank matrices. In contrast, starting with low-rank approximations captures the essential information of each task, making the orthogonalization step less costly in terms of approximation error. + +# 6.2. Per-layer task interference + +We analyze task interference on a per-layer basis using Eq. (6). As shown in Fig. 6, interference is highest in the + +![](images/d7a7238853af6382173ed4d7135af578e9e7c67d2f2bb92821a7151a18d27d62.jpg) +Figure 6. Singular Task Interference (STI) across layers in a ViT-B-32 for 20 tasks. STI is high in early layers sharing common knowledge and lower in more specialized ones. + +![](images/a5411dc1e6f1cfd5efac7fccfe7fe08c58c84950a6898ae253c3dcc440cdc9fc.jpg) +Figure 7. Best average normalized accuracy for different alpha values. The vertical labels indicate different sets of 8 tasks. + +initial transformer layers and decreases significantly in the deeper layers. This aligns with the understanding that early, generalized layers capture common features across tasks, increasing the potential for conflict, while deeper layers are more specialized for specific tasks, reducing shared representation and interference. For brevity, we grouped layers within each transformer block in our analysis; each "Layer $n$ " in Fig. 6 includes two attention matrices and two MLP matrices. A layer-by-layer analysis is provided in Fig. 14. + +# 6.3. Choice of the interpolation coefficient + +The aggregation in Eq. (2) involves a scaling coefficient $\alpha$ , typically tuned using a validation set, that can have a significant impact on the accuracy of the final model. However, as shown in Fig. 7, our approach consistently achieves the best results with $\alpha = 1.0$ , indicating that no additional scaling is necessary. Although purely empirical, this finding spares further tuning and validation data, thereby enhancing the practicality and efficiency of our method. + +# 6.4. Effect of task interference reduction + +We report in Fig. 8 the interference of a ViT-B-32 model both before and after the application of our TSV-M pipeline, + +![](images/7969a353a048b0e1838da3592cda9d730db2984a3796a0006e800d4e0e772a84.jpg) +Figure 8. Singular Task Interference (STI) and average normalized accuracy for Task Arithmetic and TSV-Merge on the ViT-B-32 model, evaluated across merges of 8, 14, and 20 tasks. + +assessed across three distinct sets of tasks with different cardinalities. In all instances, we observed that a reduction in interference is closely associated with a significant gain in accuracy. This provides empirical evidence for the effectiveness of our approach in minimizing interference and, ultimately, enhancing the merging process. + +# 7. Conclusions + +In this paper, we tackled the problem of model merging by analyzing the SVD of per-layer task matrices, confirming their inherent low-rank structure and leveraging their singular vectors to define task interference. + +Building on the former insight, we introduced TSV-Compress (TSV-C), a model compression algorithm that reduces task vectors to $10\%$ of their original size while retaining $99\%$ of the original accuracy. Unlike existing methods, our approach maintains a constant storage requirement, independent of the number of tasks. We then developed TSV-Merge (TSV-M), combining low-rank approximation with interference reduction to create a novel model merging technique. Achieving normalized accuracy up to $97\%$ , our method offers a storage-efficient alternative to MoEs, ensembles, and multi-task learning. + +Our extensive evaluation revealed that while interference reduction contributes most to performance improvements, the optimal results are achieved in combination with compression. We also observed that task interference decreases in deeper layers, and importantly, that our method does not require tuning a scaling coefficient. + +Exploring alternative methods for task importance and rank approximation could be a valuable direction for future work. Currently, we use a uniform rank across tasks, approximating each task vector to $\frac{1}{T}$ of the layer's dimensionality. However, techniques like those from Idelbayev and Carreira-Perpinan [21] or optimal hard thresholds from Gavish and Donoho [17] could be applied to select the optimal rank for each task individually. + +# Acknowledgments + +This work was partially supported by PRIN 2022 project 20227YET9B "AdVVent" CUP code B53D23012830006 and by projects FAIR (PE0000013) and SERICS (PE00000014) under the MUR National Recovery and Resilience Plan funded by the European Union - NextGenerationEU. This work was moreover partially supported by the project NEREO (Neural Reasoning over Open Data) project funded by the Italian Ministry of Education and Research (PRIN) Grant no. 2022AEFHAZ. + +# References + +[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - Mining Discriminative Components with Random Forests. In Computer Vision - ECCV 2014, pages 446-461, Cham, 2014. Springer International Publishing. 5 +[3] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 105(10):1865-1883, 2017. Conference Name: Proceedings of the IEEE. 5 +[4] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing Textures in the Wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, Columbus, OH, USA, 2014. IEEE. 5 +[5] Tarin Clanuwat, Mikel Bober-Irizar, Asanobu Kitamoto, Alex Lamb, Kazuaki Yamamoto, and David Ha. Deep learning for classical japanese literature. CoRR, abs/1812.01718, 2018. 5 +[6] Adam Coates, Andrew Ng, and Honglak Lee. An Analysis of Single-Layer Networks in Unsupervised Feature Learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. ISSN: 1938-7228. 5 +[7] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and André van Schaik. EMNIST: Extending MNIST to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 2921-2926, 2017. ISSN: 2161-4407. 5 +[8] Donato Crisostomi, Marco Fumero, Daniele Baieri, Florian Bernard, and Emanuele Rodola. $c^2 m^3$ : Cycle-consistent multi-model merging. In Advances in Neural Information Processing Systems, 2025. 2 +[9] Nico Daheim, Thomas Möllenhoff, Edoardo Ponti, Iryna Gurevych, and Mohammad Emtiyaz Khan. Model merging by uncertainty-based gradient matching. In The Twelfth International Conference on Learning Representations, 2024. 1 +[10] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse + +masks. In European Conference on Computer Vision, pages 270-287. Springer, 2025. 1, 2 +[11] Misha Denil, Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato, and Nando De Freitas. Predicting parameters in deep learning. Advances in neural information processing systems, 26, 2013. 3 +[12] Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. Advances in neural information processing systems, 27, 2014. 3, 18 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 5 +[14] Carl Eckart and G. Marion Young. The approximation of one matrix by another of lower rank. Psychometrika, 1:211-218, 1936. 3 +[15] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Workshop Track Proceedings, 2014. 4 +[16] Timur Garipov, Dmitry Podoprikhin, Alexander Novikov, and Dmitry Vetrov. Ultimate tensorization: compressing convolutional and fc layers alike. arXiv preprint arXiv:1611.03214, 2016. 3 +[17] Matan Gavish and David L. Donoho. The optimal hard threshold for singular values is $4 / \sqrt{3}$ . IEEE Transactions on Information Theory, 60(8):5040-5053, 2014. 8 +[18] Ian J. Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, Yingbo Zhou, Chetan Ramaiah, Fangxiang Feng, Ruifan Li, Xiaojie Wang, Dimitris Athanasakis, John Shawe-Taylor, Maxim Milakov, John Park, Radu Ionescu, Marius Popescu, Cristian Grozea, James Bergstra, Jingjing Xie, Lukasz Romaszko, Bing Xu, Zhang Chuang, and Yoshua Bengio. Challenges in Representation Learning: A Report on Three Machine Learning Contests. In Neural Information Processing, pages 117-124, Berlin, Heidelberg, 2013. Springer. 5 +[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. Conference Name: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 5 +[20] Chenyu Huang, Peng Ye, Tao Chen, Tong He, Xiangyu Yue, and Wanli Ouyang. EMR-merging: Tuning-free high-performance model merging. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[21] Yerlan Idelbayev and Miguel A Carreira-Perpinán. Low-rank compression of neural nets: Learning the rank of each layer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8049–8059, 2020. 8 + +[22] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 13 +[23] Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. In *British Machine Vision Conference*, BMVC 2014, Nottingham, UK, September 1-5, 2014. BMVA Press, 2014. 3 +[24] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2023. 1, 2 +[25] Keller Jordan, Hanie Sedghi, Olga Saukh, Rahim Entezari, and Behnam Neyshabur. REPAIR: REnormalizing permuted activations for interpolation repair. In The Eleventh International Conference on Learning Representations, 2023. 2 +[26] Yong-Deok Kim, Eunhyeok Park, Sungjoo Yoo, Taelim Choi, Lu Yang, and Dongjun Shin. Compression of deep convolutional neural networks for fast and low power mobile applications. In Proc. of the 4th Int. Conf. Learning Representations (ICLR 2016), San Juan, Puerto Rico, May 2–4, 2016. 3 +[27] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 3 +[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D Object Representations for Fine-Grained Categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, Sydney, Australia, 2013. IEEE. 5 +[29] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario, 2009. 5 +[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5 +[31] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. Advances in Neural Information Processing Systems, 37:78905-78935, 2024. 2 +[32] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 35:17703-17716, 2022. 1, 2 +[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 5 +[34] Maria-Elena Nilsback and Andrew Zisserman. Automated Flower Classification over a Large Number of Classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5 +[35] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. ISSN: 1063-6919. 5 + +[36] Fidel A Guerrero Pña, Heitor Rapela Medeiros, Thomas Dubail, Masih Aminbeidokhti, Eric Granger, and Marco Pedersoli. Re-basin via implicit sinkhorn differentiation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[38] Alexandre Rame, Matthieu Kirchmeyer, Thibaud Rahier, Alain Rakotomamonjy, Patrick Gallinari, and Matthieu Cord. Diverse weight averaging for out-of-distribution generalization. Advances in Neural Information Processing Systems, 35:10821-10836, 2022. 1 +[39] Noam Shazeer, *Azalia Mirhoseini*, *Krzysztof Maziarz*, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2017. 4 +[40] Li Shen, Anke Tang, Enneng Yang, Guibing Guo, Yong Luo, Lefei Zhang, Xiaochun Cao, Bo Du, and Dacheng Tao. Efficient and effective weight-ensembling mixture of experts for multi-task model merging. arXiv preprint arXiv:2410.21804, 2024. 2 +[41] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA, 2013. Association for Computational Linguistics. 5 +[42] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The German Traffic Sign Recognition Benchmark: A multi-class classification competition. In *The 2011 International Joint Conference on Neural Networks*, pages 1453–1460, 2011. ISSN: 2161-4407. 5 +[43] George Stoica, Daniel Bolya, Jakob Brandt Bjorner, Pratik Ramesh, Taylor Hearn, and Judy Hoffman. Zipit! merging models from different tasks without training. In The Twelfth International Conference on Learning Representations, 2024. 2 +[44] Anke Tang, Li Shen, Yong Luo, Shuai Xie, Han Hu, Lefei Zhang, Bo Du, and Dacheng Tao. Smile: Zero-shot sparse mixture of low-rank experts construction from pre-trained foundation models. arXiv preprint arXiv:2408.10174, 2024. 2 +[45] Bastiaan S. Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation Equivariant CNNs for Digital Pathology. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2018, pages 210–218, Cham, 2018. Springer International Publishing. 5 +[46] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. In *Forty-first International Conference on Machine Learning*, 2024. 2, 5, 6, 12, 13 + +[47] B. P. Welford. Note on a method for calculating corrected sums of squares and products. Technometrics, 4(3):419-420, 1962. 12 +[48] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 2 +[49] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms, 2017. arXiv:1708.07747 [cs, stat]. 5 +[50] Jianxiong Xiao, Krista A. Ehinger, James Hays, Antonio Torralba, and Aude Oliva. SUN Database: Exploring a Large Collection of Scene Categories. International Journal of Computer Vision, 119(1):3-22, 2016. 5 +[51] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[52] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 18 +[53] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2 +[54] Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Fabrizio Silvestri, and Emanuele Rodola. Atm: Improving model merging by alternating tuning and merging, 2024. 2 \ No newline at end of file diff --git a/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/images.zip b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..131e22d8713de537b8c0151807a3ea0a6e183cf1 --- /dev/null +++ b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaf46f6f37eb50c9b76bae7f98e380b3d516dd9a48ee6de5c47fa0403fda37b3 +size 605777 diff --git a/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/layout.json b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a580f02541553d049da6b8e9d9fc9c8033d8877 --- /dev/null +++ b/2025/Task Singular Vectors_ Reducing Task Interference in Model Merging/layout.json @@ -0,0 +1,10480 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 103, + 449, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 103, + 449, + 140 + ], + "spans": [ + { + "bbox": [ + 160, + 103, + 449, + 140 + ], + "type": "text", + "content": "Task Singular Vectors: Reducing Task Interference in Model Merging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 157, + 244, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 157, + 244, + 190 + ], + "spans": [ + { + "bbox": [ + 115, + 157, + 244, + 190 + ], + "type": "text", + "content": "Antonio Andrea Gargiulo\\* Simone Scardapane" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 265, + 157, + 362, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 157, + 362, + 188 + ], + "spans": [ + { + "bbox": [ + 265, + 157, + 362, + 188 + ], + "type": "text", + "content": "Donato Crisostomi* \nFabrizio Silvestri" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 387, + 157, + 493, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 157, + 493, + 188 + ], + "spans": [ + { + "bbox": [ + 387, + 157, + 493, + 188 + ], + "type": "text", + "content": "Maria Sofia Bucarelli Emanuele Rodolà" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 231, + 195, + 375, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 195, + 375, + 209 + ], + "spans": [ + { + "bbox": [ + 231, + 195, + 375, + 209 + ], + "type": "text", + "content": "Sapienza University of Rome" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 213, + 507, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 213, + 507, + 239 + ], + "spans": [ + { + "bbox": [ + 102, + 213, + 507, + 239 + ], + "type": "text", + "content": "gargiulo.1769185@studenti.uniroma1.it {crisostomi, rodola}@di.uniroma1.it simone.scardapane@uniroma1.it {bucarelli, fsilvestri}@diag.uniroma1.it" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 267, + 200, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 267, + 200, + 280 + ], + "spans": [ + { + "bbox": [ + 152, + 267, + 200, + 280 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 293, + 297, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 297, + 509 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 297, + 509 + ], + "type": "text", + "content": "Task Arithmetic has emerged as a simple yet effective method to merge models without additional training. However, by treating entire networks as flat parameter vectors, it overlooks key structural information and is susceptible to task interference. In this paper, we study task vectors at the layer level, focusing on task layer matrices and their singular value decomposition. In particular, we concentrate on the resulting singular vectors, which we refer to as Task Singular Vectors (TSV). Recognizing that layer task matrices are often low-rank, we propose TSV-Compress (TSV-C), a simple procedure that compresses them to " + }, + { + "bbox": [ + 54, + 293, + 297, + 509 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 293, + 297, + 509 + ], + "type": "text", + "content": " of their original size while retaining " + }, + { + "bbox": [ + 54, + 293, + 297, + 509 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 54, + 293, + 297, + 509 + ], + "type": "text", + "content": " of accuracy. We further leverage this low-rank space to define a new measure of task interference based on the interaction of singular vectors from different tasks. Building on these findings, we introduce TSV-Merge (TSV-M), a novel model merging approach that combines compression with interference reduction, significantly outperforming existing methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 531, + 135, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 135, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 551, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 296, + 696 + ], + "type": "text", + "content": "The widespread availability of pre-trained models and public repositories has driven the development of techniques for efficiently combining and reusing existing models. Among these techniques, model merging approaches enable the creation of multi-task models without additional training. One popular approach, Task Arithmetic (TA) [22], stands out for its simplicity and effectiveness. However, by treating entire networks as high-dimensional vectors, TA and subsequent works overlook crucial structural information [9, 10, 24, 32, 38, 51, 53]. This flattened view limits these approaches to coarse-grained measures like cosine similarity for assessing inter-task interactions." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 268, + 553, + 387 + ], + "blocks": [ + { + "bbox": [ + 317, + 268, + 553, + 387 + ], + "lines": [ + { + "bbox": [ + 317, + 268, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 553, + 387 + ], + "type": "image", + "image_path": "a112e20c2a5876d992350b71032a1805301920526a934c81d240c758be68c1b7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 396, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 430 + ], + "type": "text", + "content": "Figure 1. Mean accuracy of a ViT-L-14 merged over 8, 14, and 20 tasks respectively. By significantly surpassing existing methods, TSV-M establishes the new state of the art in model merging." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 642 + ], + "type": "text", + "content": "In this work, we instead focus on preserving and leveraging the natural structure of neural networks by examining weight differences at the layer level while retaining their matrix form wherever possible. Our approach begins by examining per-layer task matrices, which represent weight changes for each task, through singular value decomposition (SVD). This decomposition yields singular vectors and singular values that capture the most significant directions of variation within each layer. We term these singular vectors Task Singular Vectors (TSV), as they provide an interpretable basis for assessing task-specific contributions at each layer. Importantly, in accordance with recent literature on PEFT (cfr. Sec. 2), our analysis confirms that task matrices are inherently low-rank, meaning that only a small portion of TSVs is sufficient to represent the layer's function with high fidelity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Building on this insight, we introduce TSV-Compress (TSV-C), a simple yet effective procedure to compress task vectors down to " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " of their original size while maintaining " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " of the original accuracy. Focusing on only a fraction of the most relevant TSVs for each task, TSV-Compress enables efficient storage and processing" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 166, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 166, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 166, + 712 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 70, + 703, + 166, + 712 + ], + "type": "text", + "content": " denotes equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18695" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 292, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 292, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 292, + 84 + ], + "type": "text", + "content": "of task-specific information without waiving performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 85, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 85, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 85, + 296, + 336 + ], + "type": "text", + "content": "Beyond compression, examining the interplay of TSVs across different tasks provides a geometrically informed framework for analyzing task interference at the individual layer level. By assessing how singular vectors from different tasks align or diverge within each layer, this approach offers a significantly more fine-grained understanding of inter-task interactions, going beyond global vector similarity metrics like cosine similarity. Building upon these principles, we introduce TSV-Merge (TSV-M), a novel model merging method that combines compression and task interference reduction. This is achieved by discarding the irrelevant singular vectors for each task and then reducing their inter-task interference with a whitening transformation over their similarity matrix. We empirically show our approach to effectively reduce task interference, and the reduction to be complementary to the compression step. We then evaluate our approach across several benchmarks and show it outperforms existing methods by an average of " + }, + { + "bbox": [ + 56, + 85, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\sim 15\\%" + }, + { + "bbox": [ + 56, + 85, + 296, + 336 + ], + "type": "text", + "content": " accuracy points, establishing the new state of the art by a large margin. We release all our code for research purposes1. In summary, our contribution is 4-fold:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 337, + 295, + 516 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 295, + 372 + ], + "type": "text", + "content": "- We study the singular value decomposition of per-layer task matrices, showing them to be low-rank and measuring the interplay of singular vectors from different tasks;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 373, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 295, + 409 + ], + "type": "text", + "content": "- We introduce TSV-C, a method that builds on this insight to compress task vectors by a factor of " + }, + { + "bbox": [ + 56, + 373, + 295, + 409 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 56, + 373, + 295, + 409 + ], + "type": "text", + "content": " while preserving " + }, + { + "bbox": [ + 56, + 373, + 295, + 409 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 56, + 373, + 295, + 409 + ], + "type": "text", + "content": " of their original accuracy;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 409, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 295, + 456 + ], + "type": "text", + "content": "- We propose TSV-M, a model merging technique that complements compression with a task interference reduction step by applying a whitening transformation to decorrelate singular vectors across tasks;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 456, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 516 + ], + "type": "text", + "content": "- We conduct extensive experiments on multiple computer vision datasets, showing that TSV-M significantly outperforms existing model merging methods and provide in-depth analyses to uncover the underlying factors contributing to its success." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 529, + 142, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 142, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 142, + 541 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 549, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 295, + 693 + ], + "type": "text", + "content": "Model Merging offers an efficient alternative to ensembling by combining existing models without the need for further training. Several approaches address this by determining the neuron permutations that align the models into a shared optimization basin, after which the models are combined using a straightforward averaging method. [1, 8, 25, 36, 43]. An alternative line of approaches focuses on the multi-task scenario, where a single pre-trained model is fine-tuned for different tasks [10, 22, 32, 40, 46, 48, 51, 53, 54]. The prerequisites for methods in this category are outlined in Table 1. Task Arithmetic [22] introduces the concept of task vectors, which are the weight differences" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 91, + 556, + 262 + ], + "blocks": [ + { + "bbox": [ + 325, + 71, + 542, + 82 + ], + "lines": [ + { + "bbox": [ + 325, + 71, + 542, + 82 + ], + "spans": [ + { + "bbox": [ + 325, + 71, + 542, + 82 + ], + "type": "text", + "content": "Table 1. Model merging approaches and their requirements." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 91, + 556, + 262 + ], + "lines": [ + { + "bbox": [ + 310, + 91, + 556, + 262 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 556, + 262 + ], + "type": "table", + "html": "
MethodAdditional trainingExtra storageValidat. data inputsdata labels
Weight Avg. [48]××××
Fisher-Merg. [32]××✓×
RegMean [24]××✓×
EMR-Merging [20]×✓××
TALL-Mask [46]×✓××
TSV-C (Ours)×✓××
Task Arith. [22]××✓✓
Ties-Merging [51]××✓✓
AdaMerging [52]✓×✓×
Consensus-TA [46]××✓✓
TSV-M (Ours)××××
", + "image_path": "fbba5692ceb238ff5ab11ebd1035075fe388c609b6693548532c68ecbbb7219f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 282, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 641 + ], + "type": "text", + "content": "between fine-tuned models and the pre-trained base model. By averaging these vectors, a merged multi-task model can be created; conversely, negating a task vector allows for forgetting a specific task. TIES [51] addresses redundancies in model parameters by first selecting the top-" + }, + { + "bbox": [ + 313, + 282, + 555, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 282, + 555, + 641 + ], + "type": "text", + "content": " most significant parameter changes and then constructing a sign vector based on the majority sign across all models. The latter is used to merge the task vectors disjointly, meaning the average is not computed when a parameter is zero or when parameters disagree in sign. Similarly, DARE [53] randomly resets redundant parameters to their pre-trained values and rescales the remaining parameters by a factor proportional to the dropped ones, aiming to reduce interference among tasks. Fisher Merging [32] and RegMean [24] merge models by performing weighted averaging, utilizing the Fisher information matrix and the inner product of input vectors, respectively. Model Broadcrumbs [10] focuses on merging only significant weights by discarding outliers and both minor and large perturbations in the fine-tuned parameters. More recently, Wang et al. [46] observed that tasks often utilize non-overlapping sets of weights. They propose two methods: the first, TALL-Mask, uses binary masks to activate important task-specific weights, requiring extra storage (as does our method TSV-C) for the masks. The second, Consensus Merging, leverages these masks to remove parameters that are important to less than two tasks. Like our approach, TwinMerging [31] and SMILE [44] apply SVD to layer task matrices. However, neglecting singular vector interference, they require a router to selectively activate a subset during inference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Unlike existing approaches, our methods explicitly address task interference by leveraging the geometric structure of singular vectors to minimize interactions between task-specific parameters. Rather than averaging or selectively merging parameters as prior techniques do, we decorrelate singular vectors to directly reduce interference across tasks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 245, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 245, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 245, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 702, + 245, + 713 + ], + "type": "text", + "content": "https://github.com/AntoAndGar/task_singular_vectors" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18696" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "Additionally, we achieve targeted compression by isolating each task's unique components, storing only essential parts, and ensuring minimal interference, a step beyond traditional SVD-based compression approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 141, + 296, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 296, + 309 + ], + "type": "text", + "content": "SVD for Model Compression A significant body of work explores low-rank decompositions for fully-connected and convolutional layers [11, 12, 16, 23, 27], as well as tensor decompositions [12, 16, 26]. These approaches typically achieve a layer-by-layer factorization of a trained network, focusing on minimizing the difference between the original and the low-rank approximated weight matrices. This is accomplished using techniques like SVD [12, 16, 23, 27] or iterative optimization [23, 26]. While other methods such as weight quantization, knowledge distillation, and pruning exist, we utilize SVD not only for low-rank approximation and compression but also because its matrix decomposition properties make it particularly effective in analyzing and mitigating interference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 185, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 185, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 185, + 337 + ], + "type": "text", + "content": "3. Task Singular Vectors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 343, + 297, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 297, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 297, + 381 + ], + "type": "text", + "content": "In this section, we introduce key concepts for understanding our approach, including a new measure of task interference that is central to our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 391, + 137, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 137, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 137, + 403 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 409, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 471 + ], + "type": "text", + "content": "We build on Task Arithmetic (TA), which defines task vectors capturing the differences in model weights for individual tasks. Formally, the weights " + }, + { + "bbox": [ + 55, + 409, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{MT}}" + }, + { + "bbox": [ + 55, + 409, + 296, + 471 + ], + "type": "text", + "content": " of a multi-task model for " + }, + { + "bbox": [ + 55, + 409, + 296, + 471 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 409, + 296, + 471 + ], + "type": "text", + "content": " tasks are computed by aggregating the task-specific weight differences, or task vectors, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 123, + 479, + 296, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 479, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 123, + 479, + 296, + 508 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathrm {M T}} = \\theta_ {\\mathrm {p r e}} + \\alpha \\frac {\\sum_ {i = 1} ^ {T} \\tau_ {i}}{T}, \\tag {1}", + "image_path": "5479f6abf95c1cf0dc61f0048d77a6261927974c644b95b3a354ced6710a2815.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{pre}}" + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "text", + "content": " is the set of pretrained model weights, " + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "text", + "content": " is a scaling factor, and " + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\tau_{i} = \\theta_{\\mathrm{ft}_{i}} - \\theta_{\\mathrm{pre}}" + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "text", + "content": " is the task vector for task " + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{ft}_i}" + }, + { + "bbox": [ + 55, + 516, + 296, + 578 + ], + "type": "text", + "content": " being the fine-tuned weights for the task. Differently from TA, however, we consider these operations at the layer level. From this perspective, Eq. (1) becomes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 586, + 295, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 586, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 118, + 586, + 295, + 615 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathrm {M T}} ^ {(l)} = \\theta_ {\\text {p r e}} ^ {(l)} + \\alpha \\frac {\\sum_ {i = 1} ^ {T} \\Delta_ {i} ^ {(l)}}{T}, \\tag {2}", + "image_path": "c22fc2389da4dcb8fde17fb09ab1a446315c62cfd8a00cc9f00cab53d777eeec.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{pre}}^{(l)}" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": " encodes the pretrained weights for layer " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_i^{(l)} = \\theta_{\\mathrm{ft}_i}^{(l)} - \\theta_{\\mathrm{pre}}^{(l)}" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": " is the task-specific weight difference for task " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": ". When layer " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": " has a matrix structure, we call its corresponding " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_i^l" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": " the per-layer task matrix for task " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": ". When this is not the case, our framework defaults to standard TA. For brevity, we will generally omit the layer index and refer to the layer-" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": " task matrix " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_i^l" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_i" + }, + { + "bbox": [ + 55, + 625, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 459, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 459, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 459, + 84 + ], + "type": "text", + "content": "Decomposing layer task matrices" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "type": "text", + "content": "Treating layer-wise weights as structured entities rather than flattened vectors enables us to analyze their SVD, revealing low-rank properties and deeper insight into the inter-task interactions. Given two tasks " + }, + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "type": "text", + "content": ", we consider the SVD of their task matrices " + }, + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "type": "inline_equation", + "content": "\\Delta_{i}" + }, + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "type": "inline_equation", + "content": "\\Delta_{j}" + }, + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "type": "text", + "content": " at a generic layer:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 154, + 511, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 154, + 511, + 171 + ], + "spans": [ + { + "bbox": [ + 358, + 154, + 511, + 171 + ], + "type": "interline_equation", + "content": "\\Delta_ {i} = U _ {i} \\Sigma_ {i} V _ {i} ^ {\\top}, \\qquad \\Delta_ {j} = U _ {j} \\Sigma_ {j} V _ {j} ^ {\\top}", + "image_path": "443233c0f4dc895fde056b767b187c0b0c516211857b7629520e6c4106715406.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "inline_equation", + "content": "U_{i}, U_{j}" + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "inline_equation", + "content": "V_{i}, V_{j}" + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": " are the matrices of left and right singular vectors respectively, and " + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}, \\Sigma_{j}" + }, + { + "bbox": [ + 313, + 177, + 555, + 237 + ], + "type": "text", + "content": " are diagonal matrices of singular values. Due to their role in assessing task-specific contributions, we term the obtained singular vectors Task Singular Vectors (TSV)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "text", + "content": "Importantly for our treatment, we can equivalently write the aggregated task matrices in Eq. (2) in terms of their singular vectors and values. By defining " + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "inline_equation", + "content": "U = [U_{1}\\dots U_{T}]" + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "text", + "content": " as the column-wise concatenation of all the left TSVs, " + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "inline_equation", + "content": "V = [V_{1}\\dots V_{T}]" + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "text", + "content": " as the row-wise concatenation of the right TSVs, and " + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "text", + "content": " as the block diagonal matrix with " + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\{\\Sigma_i\\}_{i=1}^T" + }, + { + "bbox": [ + 313, + 237, + 555, + 321 + ], + "type": "text", + "content": " along its diagonal, we can write:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 350, + 329, + 555, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 329, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 350, + 329, + 555, + 362 + ], + "type": "interline_equation", + "content": "M = U \\Sigma V ^ {\\top} = \\sum_ {i = 1} ^ {T} U _ {i} \\Sigma_ {i} V _ {i} ^ {\\top} = \\sum_ {i = 1} ^ {T} \\Delta_ {i}. \\tag {3}", + "image_path": "5402b57585db74e681ea8599bad19bfa9ad6484b6ec3467db555cf1d34df6ef2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 369, + 539, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 539, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 539, + 383 + ], + "type": "text", + "content": "In the simple case where " + }, + { + "bbox": [ + 313, + 369, + 539, + 383 + ], + "type": "inline_equation", + "content": "T = 2" + }, + { + "bbox": [ + 313, + 369, + 539, + 383 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 369, + 539, + 383 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 369, + 539, + 383 + ], + "type": "text", + "content": " would be given by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 388, + 547, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 547, + 417 + ], + "type": "interline_equation", + "content": "M = \\left[ \\begin{array}{c c} U _ {1} & U _ {2} \\end{array} \\right] \\left[ \\begin{array}{c c} \\Sigma_ {1} & 0 \\\\ 0 & \\Sigma_ {2} \\end{array} \\right] \\left[ \\begin{array}{c} V _ {1} ^ {\\top} \\\\ V _ {2} ^ {\\top} \\end{array} \\right] = \\Delta_ {1} + \\Delta_ {2}.", + "image_path": "aa3ab867811cdc121726e3dd9c2da6c818122d832a3b44985dd16f233d62af60.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 423, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 509 + ], + "type": "text", + "content": "When concatenating singular components from different tasks, we violate certain properties of the SVD. Specifically, the matrices " + }, + { + "bbox": [ + 313, + 423, + 555, + 509 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 423, + 555, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 423, + 555, + 509 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 423, + 555, + 509 + ], + "type": "text", + "content": " become non-orthogonal because singular vectors from different tasks may overlap. Additionally, the singular values " + }, + { + "bbox": [ + 313, + 423, + 555, + 509 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 313, + 423, + 555, + 509 + ], + "type": "text", + "content": " from different tasks can vary significantly in magnitude, which may bias the merging process toward tasks with larger singular values." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 514, + 522, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 522, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 522, + 528 + ], + "type": "text", + "content": "3.2. Low-rank nature of layer task matrices" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 532, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 605 + ], + "type": "text", + "content": "We start our analysis by studying the low-rank properties of per-layer task matrices. Interpreting SVD as a sum of rank-one matrices, for each task " + }, + { + "bbox": [ + 313, + 532, + 556, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 532, + 556, + 605 + ], + "type": "text", + "content": ", by Eckart-Young's theorem [14] we get the best approximation (in Frobenius norm) of each task matrix " + }, + { + "bbox": [ + 313, + 532, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\Delta_{i}" + }, + { + "bbox": [ + 313, + 532, + 556, + 605 + ], + "type": "text", + "content": " by retaining only the top-" + }, + { + "bbox": [ + 313, + 532, + 556, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 532, + 556, + 605 + ], + "type": "text", + "content": " singular values and their corresponding vectors:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 393, + 612, + 555, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 612, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 393, + 612, + 555, + 647 + ], + "type": "interline_equation", + "content": "\\hat {\\Delta} _ {i} = \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \\top}. \\tag {4}", + "image_path": "fa9e4cfea615dcbf5bdef0ef0083091811b8fd312f45563f4c4fc05bcdca9328.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "As we show in Fig. 2, we find task matrices to be inherently low-rank: a small subset of TSVs is sufficient to represent the layer's function with high fidelity. In particular, even when preserving only " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": " of the singular components per task, the mean accuracy drops by merely " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ". This is" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18697" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 78, + 289, + 182 + ], + "blocks": [ + { + "bbox": [ + 65, + 78, + 289, + 182 + ], + "lines": [ + { + "bbox": [ + 65, + 78, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 65, + 78, + 289, + 182 + ], + "type": "image", + "image_path": "1842bab04fc51cc70444b101f9ef121909f45317df0acb0cf944e33b85c2ec50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 296, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 255 + ], + "type": "text", + "content": "Figure 2. Mean absolute accuracy of the ViT-B-32 model across increasing fractions of retained singular components, averaged over 20 tasks. The red line represents the average accuracy of the original fine-tuned models with full-rank task matrices, while the green line shows the accuracies using low-rank approximations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 273, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 331 + ], + "type": "text", + "content": "noticeable considering that " + }, + { + "bbox": [ + 55, + 273, + 296, + 331 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 55, + 273, + 296, + 331 + ], + "type": "text", + "content": " of singular components in each layer matrix are discarded. Building on this insight, we propose in Sec. 4.1 a compression algorithm that maintains " + }, + { + "bbox": [ + 55, + 273, + 296, + 331 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 273, + 296, + 331 + ], + "type": "text", + "content": " of the accuracy while shrinking the task vectors to " + }, + { + "bbox": [ + 55, + 273, + 296, + 331 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 273, + 296, + 331 + ], + "type": "text", + "content": " of their original size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 332, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 296, + 369 + ], + "type": "text", + "content": "Given this low-rank structure, it is natural to aggregate task matrices within their subspaces. We, therefore, obtain a reduced version of the aggregation matrix " + }, + { + "bbox": [ + 55, + 332, + 296, + 369 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 332, + 296, + 369 + ], + "type": "text", + "content": " (Eq. (3)) as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 377, + 295, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 377, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 94, + 377, + 295, + 411 + ], + "type": "interline_equation", + "content": "\\hat {M} = \\sum_ {i = 1} ^ {T} \\hat {U} _ {i} \\hat {\\Sigma} _ {i} \\hat {V} _ {i} ^ {\\top} = \\sum_ {i = 1} ^ {T} \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \\top}, \\tag {5}", + "image_path": "f88d854082a38356a3551e388c066910e5fe31eb4fe4ab077ea2c36f645a08b1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\hat{U}_i" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\hat{V}_i" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "content": " contain the top-" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "content": " left and right singular vectors for task " + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\Sigma}_i" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "content": " is the diagonal matrix of the top-" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "content": " singular values " + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\sigma_j^i" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "content": ". This formulation highlights that the low-rank approximation " + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 55, + 415, + 296, + 491 + ], + "type": "text", + "content": " utilizes the most relevant singular components from each task to effectively combine the task-specific weight differences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 496, + 203, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 203, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 203, + 509 + ], + "type": "text", + "content": "3.3. Singular Task interference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 550 + ], + "type": "text", + "content": "We hereby introduce a score of task interference based on the interplay of TSVs from different tasks, which we term Singular Task Interference (STI)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 555, + 295, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 555, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 67, + 555, + 295, + 574 + ], + "type": "interline_equation", + "content": "\\operatorname {S T I} \\left(\\left\\{\\Delta_ {i} \\right\\} _ {i = 1} ^ {T}\\right) = \\left\\| \\left(U ^ {\\top} U - I\\right) \\Sigma \\left(V ^ {\\top} V - I\\right) \\right\\| _ {1}, \\tag {6}", + "image_path": "37f5a0096265928f8cc2e2c0e60b15e3776d039d70d9617aa7d60fecf7c45278.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "inline_equation", + "content": "U, \\Sigma" + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "text", + "content": " are obtained by concatenating the singular value decompositions " + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\left(\\{U_i\\}_{i=1}^T, \\{\\Sigma_i\\}_{i=1}^T, \\{V_i\\}_{i=1}^T\\right)" + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "text", + "content": " of per-layer task matrices " + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\{\\Delta_i\\}_{i=1}^T" + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "text", + "content": " as detailed in Sec. 3.1. In the expression above, high inner product values for " + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "inline_equation", + "content": "U^\\top U" + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "inline_equation", + "content": "V^\\top V" + }, + { + "bbox": [ + 55, + 579, + 296, + 653 + ], + "type": "text", + "content": " imply a higher likelihood of interference, with minimal interference ideally yielding identity matrices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "The underlying intuition is that overlapping singular vectors suggest shared features in the weight space across tasks. Such overlap can introduce interference when models are merged, ultimately degrading performance on individual tasks. We refer to Fig. 3 for a real example over eight tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 381, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 381, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 381, + 85 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 91, + 553, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 114 + ], + "type": "text", + "content": "We demonstrate below how TSVs can be used for both compression and task interference reduction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 122, + 436, + 135 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 436, + 135 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 436, + 135 + ], + "type": "text", + "content": "4.1. TSV for compression" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 140, + 554, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 554, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 554, + 224 + ], + "type": "text", + "content": "When the task is known or inferred (e.g. with a router, as in Mixture-of-Experts [15, 39] techniques), we can leverage the low-rank structure of per-layer task matrices (see Sec. 3.2) to effectively compress these ones while discarding task singular vectors from different tasks. In particular, given a known or inferred task index " + }, + { + "bbox": [ + 313, + 140, + 554, + 224 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 140, + 554, + 224 + ], + "type": "text", + "content": ", we set other task components to zero, reducing Eq. (5) to:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 232, + 554, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 232, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 323, + 232, + 554, + 266 + ], + "type": "interline_equation", + "content": "\\hat {M} = \\sum_ {i = 1} ^ {T} \\mathbb {1} _ {[ i = h ]} \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {i} u _ {j} ^ {i} v _ {j} ^ {i \\top} = \\sum_ {j = 1} ^ {k} \\sigma_ {j} ^ {h} u _ {j} ^ {h} v _ {j} ^ {h \\top} = \\hat {\\Delta} _ {i}. \\tag {7}", + "image_path": "c3ad04461fa3fa289021fac71536bd10dc6b75f6b90cdf26b9aff52f6b3896f1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "text", + "content": "This formula yields a low-rank approximation of " + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\Delta_h" + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "text", + "content": ", where only the top " + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "text", + "content": " singular components of the task-specific matrix are retained. Here, the same number of components " + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "text", + "content": " is taken from each task to ensure that each layer is compressed by a factor of " + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\frac{1}{T}" + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "text", + "content": " relative to the original dimension. Increasing " + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "text", + "content": " improves approximation but reduces the compression." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 354, + 448, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 448, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 448, + 366 + ], + "type": "text", + "content": "4.2. TSV for model merging" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "text", + "content": "In scenarios where the task identity is not known beforehand, we address the standard model merging problem by combining Task Singular Vectors (TSVs) to create a single model that performs well jointly across all tasks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "content": "To reduce our measure of task interference, we decorrelate the TSVs of different tasks (encoded as columns in " + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\hat{U}" + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "content": ") by whitening these matrices to minimize their correlations. This can be achieved by applying the transformation " + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "inline_equation", + "content": "X \\mapsto X(X^{\\top}X)^{-\\frac{1}{2}}" + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "content": " to both " + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\hat{U}" + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "content": ". For improved numerical stability, we reformulate this whitening as an orthogonal Procrustes problem, seeking the orthogonal matrix " + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\hat{U}_{\\perp}" + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "content": " that minimizes the projection error:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 359, + 521, + 553, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 521, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 359, + 521, + 553, + 544 + ], + "type": "interline_equation", + "content": "\\min _ {\\hat {U} _ {\\perp}} \\| \\hat {U} _ {\\perp} - \\hat {U} \\| _ {F} \\quad \\text {s . t .} \\hat {U} _ {\\perp} ^ {\\top} \\hat {U} _ {\\perp} = I, \\tag {8}", + "image_path": "687a73369151516534b70fdbe8b11e5601132faa4c24171fa37c64214439c47f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": "and similarly for " + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": ". This problem admits a closed-form solution via the SVD " + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\hat{U} = PDQ^{\\top}" + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\hat{U}_{\\perp} = PQ^{\\top}" + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": " (similarly for " + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\hat{V}_{\\perp}" + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "text", + "content": "Proposition 4.1. The transformations " + }, + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "inline_equation", + "content": "X \\mapsto X(X^{\\top}X)^{-\\frac{1}{2}}" + }, + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "text", + "content": " (whitening) and " + }, + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "inline_equation", + "content": "X \\mapsto PQ^{\\top}" + }, + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "text", + "content": " (Procrustes), where " + }, + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "inline_equation", + "content": "X = PDQ^{\\top}" + }, + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "text", + "content": " is the SVD of " + }, + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 596, + 554, + 634 + ], + "type": "text", + "content": ", are equivalent." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "text", + "content": "Proof. Given " + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "inline_equation", + "content": "X = PDQ^{\\top}" + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "text", + "content": " and recalling that " + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "inline_equation", + "content": "D^{\\top}D = D^{2}" + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "text", + "content": " for diagonal matrices, simple algebraic manipulation yields " + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "inline_equation", + "content": "X^{\\top}X = QD^{2}Q^{\\top}" + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "text", + "content": ". It follows that " + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "inline_equation", + "content": "(X^{\\top}X)^{-1 / 2} = QD^{-1}Q^{\\top}" + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "text", + "content": ", therefore the whitening transformation can be rewritten as " + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "inline_equation", + "content": "X(X^{\\top}X)^{-1 / 2} = (PDQ^{\\top})(QD^{-1}Q^{\\top}) = PDID^{-1}Q^{\\top} = PQ^{\\top}" + }, + { + "bbox": [ + 313, + 640, + 554, + 714 + ], + "type": "text", + "content": ", completing the proof." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18698" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 287, + 266 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 287, + 266 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 287, + 266 + ], + "type": "image", + "image_path": "b55846b13c15ccf51d7d4406a02a81453979021d81cb0dea9c2c5efdc8de5029.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 276, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 54, + 276, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 555, + 309 + ], + "type": "text", + "content": "Figure 3. Visualization of task interference among 8 tasks computed on the first attention layer of a ViT-B-32. The diagonal blocks display intra-task similarities, while the off-diagonal blocks illustrate inter-task similarities. The zoomed-in section highlights the interaction between the right singular vectors of the " + }, + { + "bbox": [ + 54, + 276, + 555, + 309 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 309 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 276, + 555, + 309 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 309 + ], + "type": "text", + "content": " tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 287, + 70, + 527, + 266 + ], + "blocks": [ + { + "bbox": [ + 287, + 70, + 527, + 266 + ], + "lines": [ + { + "bbox": [ + 287, + 70, + 527, + 266 + ], + "spans": [ + { + "bbox": [ + 287, + 70, + 527, + 266 + ], + "type": "image", + "image_path": "a2df729feab9265e61c1703f0e91034642d16d63b0c88e3aada49f8ef971ad59.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 321, + 218, + 494 + ], + "blocks": [ + { + "bbox": [ + 59, + 321, + 218, + 494 + ], + "lines": [ + { + "bbox": [ + 59, + 321, + 218, + 494 + ], + "spans": [ + { + "bbox": [ + 59, + 321, + 218, + 494 + ], + "type": "image", + "image_path": "f2ee7971f77a130fe96fe7cd91f91c9de2cd8b956b7cc7658366078b4c95042f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 506, + 472, + 519 + ], + "lines": [ + { + "bbox": [ + 138, + 506, + 472, + 519 + ], + "spans": [ + { + "bbox": [ + 138, + 506, + 472, + 519 + ], + "type": "text", + "content": "Figure 4. Absolute accuracy of a ViT-B-32 merged over 8, 14, and 20 tasks, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 225, + 321, + 383, + 493 + ], + "blocks": [ + { + "bbox": [ + 225, + 321, + 383, + 493 + ], + "lines": [ + { + "bbox": [ + 225, + 321, + 383, + 493 + ], + "spans": [ + { + "bbox": [ + 225, + 321, + 383, + 493 + ], + "type": "image", + "image_path": "6c25b469be8f6f381fc43fbd4038701511e968c5dbc6cbde79245fcbec03f504.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 386, + 321, + 552, + 492 + ], + "blocks": [ + { + "bbox": [ + 386, + 321, + 552, + 492 + ], + "lines": [ + { + "bbox": [ + 386, + 321, + 552, + 492 + ], + "spans": [ + { + "bbox": [ + 386, + 321, + 552, + 492 + ], + "type": "image", + "image_path": "a8d2c040cacf957448513937376889eb2841050a2f486a69a521a0d8a3ac02a7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 538, + 295, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 575 + ], + "type": "text", + "content": "Once the (rank-reduced) TSV matrices " + }, + { + "bbox": [ + 55, + 538, + 295, + 575 + ], + "type": "inline_equation", + "content": "\\hat{U}" + }, + { + "bbox": [ + 55, + 538, + 295, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 538, + 295, + 575 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 55, + 538, + 295, + 575 + ], + "type": "text", + "content": " are decorrelated, we reconstruct the merged layer as in Eq. (5); see Algorithm 1 for the complete steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 586, + 108, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 108, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 108, + 597 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "We evaluate our approaches over three different suites of tasks having cardinality 8, 14, and 20, respectively. The first one, introduced in [22], consists of datasets: Cars [28], DTD [4], EuroSAT [19], GTSRB [42], MNIST [30], RESISC45 [3], SUN397 [50], and SVHN [33]. The benchmark with 14 tasks builds on the preceding one, incorporating six additional datasets: CIFAR100 [29], STL10 [6], Flowers102 [34], OxfordIIITPet [35], PCAM [45], and FER2013 [18]. Finally, the 20-tasks benchmark in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 539, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 575 + ], + "type": "text", + "content": "cludes the preceding 14 plus the following six: EMNIST [7], CIFAR10 [29], Food101 [2], FashionMNIST [49], RenderedSST2 [41], and KMNIST [5]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 586, + 443, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 443, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 443, + 600 + ], + "type": "text", + "content": "5.1. Model merging results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "We evaluate our method on three variants of the CLIP [37] model, each employing a different size of ViT [13] visual encoder: ViT-B-32, ViT-B-16, and ViT-L-14. The main benchmarks involve merging 8, 14, and 20 tasks, mirroring the experimental setup described in Wang et al. [46]. We compare our approach against several training-free model merging methods, including weight averaging, Task Arithmetic [22], and Consensus Merging [46]. For reference, we include the performance of zero" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18699" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 148 + ], + "type": "table", + "html": "
MethodViT-B-32ViT-B-16ViT-L-14
8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks
Zeroshot48.26(53.59)57.21(63.69)56.10(62.41)55.34(59.34)61.28(66.19)59.73(64.52)64.70(68.00)68.20(72.15)65.23(68.99)
Weight Averaging66.34(72.13)64.34(71.12)61.04(67.53)72.22(76.60)69.46(74.82)65.31(70.36)79.56(83.15)76.73(81.10)71.60(75.60)
Task Arithmetic70.79(76.55)65.32(72.09)60.52(66.79)75.41(79.58)70.52(75.89)65.78(70.76)84.93(88.65)79.41(83.95)74.01(78.07)
Consensus TA75.03(80.84)70.39(77.36)65.43(71.98)79.39(83.86)74.39(79.92)69.76(74.93)86.34(90.08)82.22(86.94)79.00(83.22)
TSV-M (Ours)85.86(92.31)80.06(87.88)77.07(84.29)89.01(93.94)84.58(91.01)80.57(86.45)92.98(96.98)89.17(94.43)87.72(92.50)
", + "image_path": "03b97b0f268d348c59976af0030781f17df283fdb1389811fd8270b985d19647.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 154, + 552, + 166 + ], + "lines": [ + { + "bbox": [ + 56, + 154, + 552, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 552, + 166 + ], + "type": "text", + "content": "Table 2. Average absolute accuracy results on model merging benchmarks; subscript (in parentheses) is the normalized average accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "code", + "bbox": [ + 56, + 201, + 294, + 430 + ], + "blocks": [ + { + "bbox": [ + 56, + 186, + 168, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 186, + 168, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 168, + 198 + ], + "type": "text", + "content": "Algorithm 1 TSV-Merge." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "lines": [ + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": "Require: Task matrices " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\Delta_1,\\ldots ,\\Delta_T" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " , scaling factor " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \nEnsure: Merged model weights " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{MT}}" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n1: for " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " do \n2: Compute SVD: " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\Delta_i = U_i\\Sigma_iV_i^\\top" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n3: Retain first " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\frac{1}{T}" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " singular components of " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "U_{i},\\Sigma_{i}" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n4: end for \n5: Concatenate the matrices: \n6: " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "U\\gets [U_1|U_2|\\dots |U_T]" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " - block-diag( " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\Sigma_1,\\Sigma_2,\\dots ,\\Sigma_T)" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "V\\gets [V_{1}|V_{2}|\\dots |V_{T}]" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n9: Compute the SVD of " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " .. \n10: " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "U = P_{U}D_{U}Q_{U}^{\\top}\\qquad V = P_{V}D_{V}Q_{V}^{\\top}" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n11:Obtain the orthogonal matrices: \n12: " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "U_{\\perp} = P_{U}Q_{U}^{\\top}\\qquad V_{\\perp} = P_{V}Q_{V}^{\\top}" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n13: Reconstruct the merged matrix: \n14: " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\hat{M}\\gets U_{\\perp}\\Sigma V_{\\perp}^{\\top}" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n15: Construct merged model weights: \n16: " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{MT}}\\gets \\theta_{\\mathrm{pre}} + \\alpha \\hat{M}" + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "text", + "content": " \n17: return " + }, + { + "bbox": [ + 56, + 201, + 294, + 430 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{MT}}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 459, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 532 + ], + "type": "text", + "content": "shot models in Table 2 to represent the minimum achievable accuracy, and the average of individually fine-tuned models in Table 3 to indicate the maximum potential gains. Performance metrics were assessed using both the average absolute accuracy and the average normalized accuracy, calculated as detailed in Appendix B.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": "As presented in Table 2, our method achieves state-of-the-art results across all benchmarks, regardless of the ViT size or the number of tasks involved. Notably, the most significant improvements were observed with the smaller ViT-B-32 encoder, where our approach outperforms Task Arithmetic and Consensus TA by an average of " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "+15.45\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "+10.71\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": " absolute accuracy, respectively. Furthermore, when utilizing the ViT-L-14 model, our method attains an average normalized accuracy of " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "96.98\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": ". This indicates that we can effectively replace eight individual task-specific models with a single merged model, incurring only a " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "-3.02\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": " reduction in average accuracy. These results highlight the promise of our model merging technique as a cost-efficient alternative to mixtures of experts, multi-task learning, and assembling methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 186, + 500, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 500, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 500, + 199 + ], + "type": "text", + "content": "We report in Fig. 4 the per-dataset accuracies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 206, + 432, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 432, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 432, + 219 + ], + "type": "text", + "content": "5.2. Compression results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 224, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 451 + ], + "type": "text", + "content": "In Table 3 we report the results for our compression algorithm TSV-C. We compare our method with TALL-masks [46], which stores very sparse binary masks for each task. As shown in Table 3, our method always retains more than " + }, + { + "bbox": [ + 313, + 224, + 555, + 451 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 224, + 555, + 451 + ], + "type": "text", + "content": " of the original accuracy for all considered benchmarks and models. The results of TSV-C are comparable to those of TALL-masks [46], with the former slightly underperforming TALL-masks for small ViTs and the latter gaining the upper hand in the larger-scale one. In general, however, the two approaches differ by less than " + }, + { + "bbox": [ + 313, + 224, + 555, + 451 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 224, + 555, + 451 + ], + "type": "text", + "content": " accuracy on average. Regarding the storage, our method stores only the top " + }, + { + "bbox": [ + 313, + 224, + 555, + 451 + ], + "type": "inline_equation", + "content": "\\frac{1}{T}" + }, + { + "bbox": [ + 313, + 224, + 555, + 451 + ], + "type": "text", + "content": " singular vectors per task, which leads to a fixed storage requirement of approximately twice the size of the original model, regardless of the number of tasks. By storing binary masks, TALL-masks instead results in a storage size that varies with their compressibility, increasing with the number of tasks. This variability can lead to unpredictable storage requirements and may diminish compression benefits when masks are less compressible." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 452, + 555, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 555, + 512 + ], + "type": "text", + "content": "Our results show that TSV-C effectively balances compression and performance by leveraging the most significant TSVs. Maintaining near-original accuracy, our approach is particularly suitable for scenarios where storage constraints are critical but high model performance is still required." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 523, + 373, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 373, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 373, + 537 + ], + "type": "text", + "content": "6. Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "content": "In this section, we begin by conducting an ablation study to assess the contributions of interference reduction and low-rank compression to the overall performance. Next, we study how task interference varies across layers of different depths. Finally, we empirically show that our method does not require tuning a scaling coefficient." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 624, + 406, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 406, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 406, + 636 + ], + "type": "text", + "content": "6.1. Ablation study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "TSV-M combines low-rank approximation with task interference reduction. To evaluate the contribution of each component, we conducted an ablation study summarized in Table 4. Applying low-rank approximation alone to the layer task matrices results in worse performance than vanilla Task Arithmetic, shown in the first row where" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18700" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 130 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 130 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 130 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 130 + ], + "type": "table", + "html": "
MethodViT-B-32ViT-B-16ViT-L-14
8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks8 tasks14 tasks20 tasks
Finetuned92.83(100)90.88(100)91.37(100)94.64(100)92.76(100)93.17(100)95.81(100)94.29(100)94.73(100)
TALL-Mask+TIES93.13(100.37)90.92(100.04)91.11(99.70)94.68(100.04)92.69(99.90)93.05(99.87)95.96(100.15)93.40(99.09)93.91(99.16)
TSV-C (Ours)92.62(99.74)90.29(99.28)90.64(99.14)94.47(99.79)92.25(99.41)92.53(99.27)95.68(99.85)94.04(99.72)94.42(99.66)
", + "image_path": "20de4a1a895a1e27a49b7a87a05ac198e54636f24d3d3eb9b9768bc8fc103f50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 190, + 294, + 264 + ], + "blocks": [ + { + "bbox": [ + 55, + 138, + 555, + 172 + ], + "lines": [ + { + "bbox": [ + 55, + 138, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 555, + 172 + ], + "type": "text", + "content": "Table 3. Average absolute accuracy results across all compression benchmarks for different models and varying number of tasks, subscript (in parentheses) the normalized average accuracies. Our TSV-C compression method consistently achieves over " + }, + { + "bbox": [ + 55, + 138, + 555, + 172 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 138, + 555, + 172 + ], + "type": "text", + "content": " of the original fine-tuned models' accuracy while significantly reducing storage requirements." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 190, + 294, + 264 + ], + "lines": [ + { + "bbox": [ + 58, + 190, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 58, + 190, + 294, + 264 + ], + "type": "table", + "html": "
Low-rank approx.Interf. reductionViT-B-32
8 tasks14 tasks20 tasks
××76.5 (+0.0)72.1 (+0.0)66.8 (+0.0)
✓×75.2 (-1.3)71.0 (-1.1)66.3 (-0.5)
×✓82.6 (+7.4)75.7 (+4.7)69.9 (+3.6)
✓✓92.3 (+9.7)87.9 (+12.2)84.3 (+14.4)
", + "image_path": "6b5d5906fe7d527fba09231a278f78b7d9b95f3e3d386c7cbdded913a55fbbf8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 295, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 295, + 316 + ], + "type": "text", + "content": "Table 4. Comparison of different versions of Task Arithmetic, comprising either the low-rank approximation step, the interference reduction step, or both. The method performing both corresponds to the proposed TSV-Merge." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 329, + 294, + 443 + ], + "blocks": [ + { + "bbox": [ + 57, + 329, + 294, + 443 + ], + "lines": [ + { + "bbox": [ + 57, + 329, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 57, + 329, + 294, + 443 + ], + "type": "image", + "image_path": "bf1bef04df4181e912a6d06a8efdbf544fc1f6b350bee695e433935858d44d98.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 452, + 295, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 452, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 295, + 495 + ], + "type": "text", + "content": "Figure 5. Approximation error from the orthogonalization of the TSVs through Procrustes for the ViT-B-32 model across 8 tasks. The violin plots represent layer-wise approximation error distributions for " + }, + { + "bbox": [ + 55, + 452, + 295, + 495 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 452, + 295, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 452, + 295, + 495 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 452, + 295, + 495 + ], + "type": "text", + "content": " in both full-rank and low-rank cases." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "text", + "content": "neither component is applied. In contrast, applying task interference reduction alone, while keeping full-rank matrices, significantly improves performance, with gains ranging from " + }, + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "inline_equation", + "content": "+3.1\\%" + }, + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "inline_equation", + "content": "+6.1\\%" + }, + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "text", + "content": ". However, the best results are achieved only when both steps are combined, yielding substantial performance improvements of " + }, + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "inline_equation", + "content": "+15.8\\%" + }, + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "inline_equation", + "content": "+17.5\\%" + }, + { + "bbox": [ + 55, + 520, + 295, + 592 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": "To explain why interference reduction applied to low-rank approximations in TSV-M outperforms its application to full-rank matrices, we analyze the errors introduced during the orthogonalization of the full-rank " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " matrices. As shown in Fig. 5, orthogonalizing full-rank matrices incurs significant approximation errors, measured by the sum of the Frobenius norms of reconstruction errors for the concatenated " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " matrices across all layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "For example, in the ViT-B-32 model with 8 tasks, the full-rank setting exhibits a wider error distribution with a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 193, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 277 + ], + "type": "text", + "content": "higher average, indicating greater variability and larger approximation discrepancies. In contrast, the low-rank setting produces a more compact and lower error distribution, suggesting better consistency in approximation across layers. In the following, we prove that this is not specific to the chosen model but a general property of our approach. The proof can be found in Appendix C.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": "Theorem 6.1. Let " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{N}" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "T > 4" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": ". Define " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "U = [U_1, \\ldots, U_T]" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": " as the matrix obtained by concatenating " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": " orthogonal matrices " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "U_i" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": ", each of shape " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\widehat{U} = [\\widehat{U}_1, \\ldots, \\widehat{U}_T]" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": " be the matrix formed by truncating each " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "U_i" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": " to its first " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": " columns. Denote by " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\widehat{X}" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": " the matrices resulting from Procrustes orthonormalization of " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\widehat{U}" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": ", respectively. If " + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "inline_equation", + "content": "k \\leq n\\frac{T - 2\\sqrt{T}}{T}" + }, + { + "bbox": [ + 313, + 280, + 555, + 369 + ], + "type": "text", + "content": ", then" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 372, + 490, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 372, + 490, + 388 + ], + "spans": [ + { + "bbox": [ + 378, + 372, + 490, + 388 + ], + "type": "interline_equation", + "content": "\\left\\| U - X \\right\\| _ {F} \\geq \\left\\| \\widehat {U} - \\widehat {X} \\right\\| _ {F}.", + "image_path": "cd614cee464944c0cfa62561c370e1b8cb85f4eff368374137ee11923a87746d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "text", + "content": "Intuitively, the theorem asserts that the approximation error introduced by orthogonalization via Procrustes is smaller when applied to the concatenation of truncated matrices, compared to the concatenation of the original full-rank matrices, provided the rank of the truncated matrices satisfies certain conditions. To better understand the statement of the theorem, consider the case " + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "inline_equation", + "content": "T = 10" + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "text", + "content": ", as in our scenario. The term " + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "inline_equation", + "content": "\\frac{T - 2\\sqrt{T}}{T} = \\frac{10 - 2\\sqrt{10}}{10} \\geq \\frac{1}{3}" + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "text", + "content": ". In our case, we choose " + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "inline_equation", + "content": "k \\approx \\frac{n}{T}" + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "inline_equation", + "content": "T = 10" + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "text", + "content": ", this gives " + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "inline_equation", + "content": "k \\approx \\frac{n}{10}" + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "text", + "content": ", which is indeed smaller than " + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "inline_equation", + "content": "\\frac{n}{3}" + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "text", + "content": ", satisfying the condition " + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "inline_equation", + "content": "k \\leq \\frac{1}{3} n" + }, + { + "bbox": [ + 313, + 392, + 554, + 512 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 512, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 554, + 582 + ], + "type": "text", + "content": "We emphasize that other orthogonalization methods, such as Gram-Schmidt, are ineffective in this context because they do not preserve the overall structure of the original data. Gram-Schmidt sequentially orthogonalizes vectors without minimizing deviation from the original set, potentially resulting in significant distortions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 583, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 554, + 666 + ], + "type": "text", + "content": "Finally, the substantial reduction in approximation error suggests that, while the advantages of interference reduction are considerable, they may be offset by the errors introduced when applied to full-rank matrices. In contrast, starting with low-rank approximations captures the essential information of each task, making the orthogonalization step less costly in terms of approximation error." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 672, + 462, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 462, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 462, + 685 + ], + "type": "text", + "content": "6.2. Per-layer task interference" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "We analyze task interference on a per-layer basis using Eq. (6). As shown in Fig. 6, interference is highest in the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18701" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 291, + 180 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 291, + 180 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 291, + 180 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 291, + 180 + ], + "type": "image", + "image_path": "d7a7238853af6382173ed4d7135af578e9e7c67d2f2bb92821a7151a18d27d62.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 227 + ], + "type": "text", + "content": "Figure 6. Singular Task Interference (STI) across layers in a ViT-B-32 for 20 tasks. STI is high in early layers sharing common knowledge and lower in more specialized ones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 240, + 293, + 384 + ], + "blocks": [ + { + "bbox": [ + 58, + 240, + 293, + 384 + ], + "lines": [ + { + "bbox": [ + 58, + 240, + 293, + 384 + ], + "spans": [ + { + "bbox": [ + 58, + 240, + 293, + 384 + ], + "type": "image", + "image_path": "a5411dc1e6f1cfd5efac7fccfe7fe08c58c84950a6898ae253c3dcc440cdc9fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 394, + 295, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 394, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 295, + 415 + ], + "type": "text", + "content": "Figure 7. Best average normalized accuracy for different alpha values. The vertical labels indicate different sets of 8 tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 544 + ], + "type": "text", + "content": "initial transformer layers and decreases significantly in the deeper layers. This aligns with the understanding that early, generalized layers capture common features across tasks, increasing the potential for conflict, while deeper layers are more specialized for specific tasks, reducing shared representation and interference. For brevity, we grouped layers within each transformer block in our analysis; each \"Layer " + }, + { + "bbox": [ + 55, + 437, + 295, + 544 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 437, + 295, + 544 + ], + "type": "text", + "content": "\" in Fig. 6 includes two attention matrices and two MLP matrices. A layer-by-layer analysis is provided in Fig. 14." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 254, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 254, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 254, + 565 + ], + "type": "text", + "content": "6.3. Choice of the interpolation coefficient" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "type": "text", + "content": "The aggregation in Eq. (2) involves a scaling coefficient " + }, + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "type": "text", + "content": ", typically tuned using a validation set, that can have a significant impact on the accuracy of the final model. However, as shown in Fig. 7, our approach consistently achieves the best results with " + }, + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\alpha = 1.0" + }, + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "type": "text", + "content": ", indicating that no additional scaling is necessary. Although purely empirical, this finding spares further tuning and validation data, thereby enhancing the practicality and efficiency of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 247, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 247, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 247, + 684 + ], + "type": "text", + "content": "6.4. Effect of task interference reduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "We report in Fig. 8 the interference of a ViT-B-32 model both before and after the application of our TSV-M pipeline," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 339, + 75, + 531, + 182 + ], + "blocks": [ + { + "bbox": [ + 339, + 75, + 531, + 182 + ], + "lines": [ + { + "bbox": [ + 339, + 75, + 531, + 182 + ], + "spans": [ + { + "bbox": [ + 339, + 75, + 531, + 182 + ], + "type": "image", + "image_path": "7969a353a048b0e1838da3592cda9d730db2984a3796a0006e800d4e0e772a84.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 195, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 195, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 553, + 228 + ], + "type": "text", + "content": "Figure 8. Singular Task Interference (STI) and average normalized accuracy for Task Arithmetic and TSV-Merge on the ViT-B-32 model, evaluated across merges of 8, 14, and 20 tasks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 253, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 326 + ], + "type": "text", + "content": "assessed across three distinct sets of tasks with different cardinalities. In all instances, we observed that a reduction in interference is closely associated with a significant gain in accuracy. This provides empirical evidence for the effectiveness of our approach in minimizing interference and, ultimately, enhancing the merging process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 340, + 391, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 340, + 391, + 353 + ], + "spans": [ + { + "bbox": [ + 314, + 340, + 391, + 353 + ], + "type": "text", + "content": "7. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 361, + 553, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 409 + ], + "type": "text", + "content": "In this paper, we tackled the problem of model merging by analyzing the SVD of per-layer task matrices, confirming their inherent low-rank structure and leveraging their singular vectors to define task interference." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 411, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 543 + ], + "type": "text", + "content": "Building on the former insight, we introduced TSV-Compress (TSV-C), a model compression algorithm that reduces task vectors to " + }, + { + "bbox": [ + 313, + 411, + 554, + 543 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 411, + 554, + 543 + ], + "type": "text", + "content": " of their original size while retaining " + }, + { + "bbox": [ + 313, + 411, + 554, + 543 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 411, + 554, + 543 + ], + "type": "text", + "content": " of the original accuracy. Unlike existing methods, our approach maintains a constant storage requirement, independent of the number of tasks. We then developed TSV-Merge (TSV-M), combining low-rank approximation with interference reduction to create a novel model merging technique. Achieving normalized accuracy up to " + }, + { + "bbox": [ + 313, + 411, + 554, + 543 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 313, + 411, + 554, + 543 + ], + "type": "text", + "content": ", our method offers a storage-efficient alternative to MoEs, ensembles, and multi-task learning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 616 + ], + "type": "text", + "content": "Our extensive evaluation revealed that while interference reduction contributes most to performance improvements, the optimal results are achieved in combination with compression. We also observed that task interference decreases in deeper layers, and importantly, that our method does not require tuning a scaling coefficient." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Exploring alternative methods for task importance and rank approximation could be a valuable direction for future work. Currently, we use a uniform rank across tasks, approximating each task vector to " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{T}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " of the layer's dimensionality. However, techniques like those from Idelbayev and Carreira-Perpinan [21] or optimal hard thresholds from Gavish and Donoho [17] could be applied to select the optimal rank for each task individually." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18702" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 138, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 138, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 297, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 297, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 297, + 196 + ], + "type": "text", + "content": "This work was partially supported by PRIN 2022 project 20227YET9B \"AdVVent\" CUP code B53D23012830006 and by projects FAIR (PE0000013) and SERICS (PE00000014) under the MUR National Recovery and Resilience Plan funded by the European Union - NextGenerationEU. This work was moreover partially supported by the project NEREO (Neural Reasoning over Open Data) project funded by the Italian Ministry of Education and Research (PRIN) Grant no. 2022AEFHAZ." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 208, + 115, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 115, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 115, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 227, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 227, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 227, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 227, + 294, + 272 + ], + "type": "text", + "content": "[1] Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 317 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - Mining Discriminative Components with Random Forests. In Computer Vision - ECCV 2014, pages 446-461, Cham, 2014. Springer International Publishing. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 319, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 361 + ], + "type": "text", + "content": "[3] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 105(10):1865-1883, 2017. Conference Name: Proceedings of the IEEE. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "type": "text", + "content": "[4] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing Textures in the Wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, Columbus, OH, USA, 2014. IEEE. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "text", + "content": "[5] Tarin Clanuwat, Mikel Bober-Irizar, Asanobu Kitamoto, Alex Lamb, Kazuaki Yamamoto, and David Ha. Deep learning for classical japanese literature. CoRR, abs/1812.01718, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 294, + 531 + ], + "type": "text", + "content": "[6] Adam Coates, Andrew Ng, and Honglak Lee. An Analysis of Single-Layer Networks in Unsupervised Feature Learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. ISSN: 1938-7228. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "text", + "content": "[7] Gregory Cohen, Saeed Afshar, Jonathan Tapson, and André van Schaik. EMNIST: Extending MNIST to handwritten letters. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 2921-2926, 2017. ISSN: 2161-4407. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 294, + 633 + ], + "type": "text", + "content": "[8] Donato Crisostomi, Marco Fumero, Daniele Baieri, Florian Bernard, and Emanuele Rodola. " + }, + { + "bbox": [ + 62, + 590, + 294, + 633 + ], + "type": "inline_equation", + "content": "c^2 m^3" + }, + { + "bbox": [ + 62, + 590, + 294, + 633 + ], + "type": "text", + "content": ": Cycle-consistent multi-model merging. In Advances in Neural Information Processing Systems, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 294, + 689 + ], + "type": "text", + "content": "[9] Nico Daheim, Thomas Möllenhoff, Edoardo Ponti, Iryna Gurevych, and Mohammad Emtiyaz Khan. Model merging by uncertainty-based gradient matching. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[10] MohammadReza Davari and Eugene Belilovsky. Model breadcrumbs: Scaling multi-task model merging with sparse" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "masks. In European Conference on Computer Vision, pages 270-287. Springer, 2025. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[11] Misha Denil, Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato, and Nando De Freitas. Predicting parameters in deep learning. Advances in neural information processing systems, 26, 2013. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[12] Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. Advances in neural information processing systems, 27, 2014. 3, 18" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 185, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 249 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "text", + "content": "[14] Carl Eckart and G. Marion Young. The approximation of one matrix by another of lower rank. Psychometrika, 1:211-218, 1936. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 338 + ], + "type": "text", + "content": "[15] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Workshop Track Proceedings, 2014. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 339, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 381 + ], + "type": "text", + "content": "[16] Timur Garipov, Dmitry Podoprikhin, Alexander Novikov, and Dmitry Vetrov. Ultimate tensorization: compressing convolutional and fc layers alike. arXiv preprint arXiv:1611.03214, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "text", + "content": "[17] Matan Gavish and David L. Donoho. The optimal hard threshold for singular values is " + }, + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "inline_equation", + "content": "4 / \\sqrt{3}" + }, + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "text", + "content": ". IEEE Transactions on Information Theory, 60(8):5040-5053, 2014. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 417, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 417, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 417, + 553, + 536 + ], + "type": "text", + "content": "[18] Ian J. Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, Yingbo Zhou, Chetan Ramaiah, Fangxiang Feng, Ruifan Li, Xiaojie Wang, Dimitris Athanasakis, John Shawe-Taylor, Maxim Milakov, John Park, Radu Ionescu, Marius Popescu, Cristian Grozea, James Bergstra, Jingjing Xie, Lukasz Romaszko, Bing Xu, Zhang Chuang, and Yoshua Bengio. Challenges in Representation Learning: A Report on Three Machine Learning Contests. In Neural Information Processing, pages 117-124, Berlin, Heidelberg, 2013. Springer. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 538, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 538, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 538, + 553, + 613 + ], + "type": "text", + "content": "[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. Conference Name: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 615, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 667 + ], + "type": "text", + "content": "[20] Chenyu Huang, Peng Ye, Tao Chen, Tong He, Xiangyu Yue, and Wanli Ouyang. EMR-merging: Tuning-free high-performance model merging. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[21] Yerlan Idelbayev and Miguel A Carreira-Perpinán. Low-rank compression of neural nets: Learning the rank of each layer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8049–8059, 2020. 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18703" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[22] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[23] Max Jaderberg, Andrea Vedaldi, and Andrew Zisserman. Speeding up convolutional neural networks with low rank expansions. In *British Machine Vision Conference*, BMVC 2014, Nottingham, UK, September 1-5, 2014. BMVA Press, 2014. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[24] Xisen Jin, Xiang Ren, Daniel Preotiac-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[25] Keller Jordan, Hanie Sedghi, Olga Saukh, Rahim Entezari, and Behnam Neyshabur. REPAIR: REnormalizing permuted activations for interpolation repair. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 340 + ], + "type": "text", + "content": "[26] Yong-Deok Kim, Eunhyeok Park, Sungjoo Yoo, Taelim Choi, Lu Yang, and Dongjun Shin. Compression of deep convolutional neural networks for fast and low power mobile applications. In Proc. of the 4th Int. Conf. Learning Representations (ICLR 2016), San Juan, Puerto Rico, May 2–4, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 363 + ], + "type": "text", + "content": "[27] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "text", + "content": "[28] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D Object Representations for Fine-Grained Categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, Sydney, Australia, 2013. IEEE. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "text", + "content": "[29] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario, 2009. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 487 + ], + "type": "text", + "content": "[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[31] Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. Twin-merging: Dynamic integration of modular expertise in model merging. Advances in Neural Information Processing Systems, 37:78905-78935, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "text", + "content": "[32] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 35:17703-17716, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "text", + "content": "[34] Maria-Elena Nilsback and Andrew Zisserman. Automated Flower Classification over a Large Number of Classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[35] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505, 2012. ISSN: 1063-6919. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[36] Fidel A Guerrero Pña, Heitor Rapela Medeiros, Thomas Dubail, Masih Aminbeidokhti, Eric Granger, and Marco Pedersoli. Re-basin via implicit sinkhorn differentiation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 193 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "text", + "content": "[38] Alexandre Rame, Matthieu Kirchmeyer, Thibaud Rahier, Alain Rakotomamonjy, Patrick Gallinari, and Matthieu Cord. Diverse weight averaging for out-of-distribution generalization. Advances in Neural Information Processing Systems, 35:10821-10836, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[39] Noam Shazeer, *Azalia Mirhoseini*, *Krzysztof Maziarz*, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2017. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "text", + "content": "[40] Li Shen, Anke Tang, Enneng Yang, Guibing Guo, Yong Luo, Lefei Zhang, Xiaochun Cao, Bo Du, and Dacheng Tao. Efficient and effective weight-ensembling mixture of experts for multi-task model merging. arXiv preprint arXiv:2410.21804, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 437 + ], + "type": "text", + "content": "[41] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA, 2013. Association for Computational Linguistics. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "text", + "content": "[42] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The German Traffic Sign Recognition Benchmark: A multi-class classification competition. In *The 2011 International Joint Conference on Neural Networks*, pages 1453–1460, 2011. ISSN: 2161-4407. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 553, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 546 + ], + "type": "text", + "content": "[43] George Stoica, Daniel Bolya, Jakob Brandt Bjorner, Pratik Ramesh, Taylor Hearn, and Judy Hoffman. Zipit! merging models from different tasks without training. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 601 + ], + "type": "text", + "content": "[44] Anke Tang, Li Shen, Yong Luo, Shuai Xie, Han Hu, Lefei Zhang, Bo Du, and Dacheng Tao. Smile: Zero-shot sparse mixture of low-rank experts construction from pre-trained foundation models. arXiv preprint arXiv:2408.10174, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "text", + "content": "[45] Bastiaan S. Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation Equivariant CNNs for Digital Pathology. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2018, pages 210–218, Cham, 2018. Springer International Publishing. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[46] Ke Wang, Nikolaos Dimitriadis, Guillermo Ortiz-Jimenez, François Fleuret, and Pascal Frossard. Localizing task information for improved model merging and compression. In *Forty-first International Conference on Machine Learning*, 2024. 2, 5, 6, 12, 13" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18704" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[47] B. P. Welford. Note on a method for calculating corrected sums of squares and products. Technometrics, 4(3):419-420, 1962. 12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 193 + ], + "type": "text", + "content": "[48] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Proceedings of the 39th International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 197, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 197, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 197, + 295, + 228 + ], + "type": "text", + "content": "[49] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms, 2017. arXiv:1708.07747 [cs, stat]. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 230, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 230, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 296, + 273 + ], + "type": "text", + "content": "[50] Jianxiong Xiao, Krista A. Ehinger, James Hays, Antonio Torralba, and Aude Oliva. SUN Database: Exploring a Large Collection of Scene Categories. International Journal of Computer Vision, 119(1):3-22, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 275, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 275, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 57, + 275, + 294, + 317 + ], + "type": "text", + "content": "[51] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 319, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 319, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 57, + 319, + 294, + 373 + ], + "type": "text", + "content": "[52] Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 18" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 376, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 376, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 294, + 418 + ], + "type": "text", + "content": "[53] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 421, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 421, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 57, + 421, + 294, + 463 + ], + "type": "text", + "content": "[54] Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Fabrizio Silvestri, and Emanuele Rodola. Atm: Improving model merging by alternating tuning and merging, 2024. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18705" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_content_list.json b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..60489c4660915e0c2082facf2ee4cd7760bc0c6d --- /dev/null +++ b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_content_list.json @@ -0,0 +1,1636 @@ +[ + { + "type": "text", + "text": "Task-Agnostic Guided Feature Expansion for Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 127, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bowen Zheng Da-Wei Zhou† Han-Jia Ye De-Chuan Zhan \nNational Key Laboratory for Novel Software Technology, Nanjing University, China \nSchool of Artificial Intelligence, Nanjing University, China", + "bbox": [ + 163, + 180, + 831, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhengbw, zhoudw, yehj, zhandc}@lamda.nju.edu.cn", + "bbox": [ + 295, + 236, + 694, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to learn new concepts while preserve the learned knowledge is desirable for learning systems in Class-Incremental Learning (CIL). Recently, feature expansion of the model become a prevalent solution for CIL, where the old features are fixed during the training of the new task while new features are expanded for the new tasks. However, such task-specific features learned from the new task may collide with the old features, leading to misclassification between tasks. Therefore, the expanded model is often encouraged to capture diverse features from the new task, aiming to avoid such collision. However, the existing solution is largely restricted to the samples from the current task, because of the poor accessibility to previous samples. To promote the learning and transferring of diverse features across tasks, we propose a framework called Task-Agnostic Guided Feature Expansion (TagFex). Firstly, it captures task-agnostic features continually with a separate model, providing extra task-agnostic features for subsequent tasks. Secondly, to obtain useful features from the task-agnostic model for the current task, it aggregates the task-agnostic features with the task-specific feature using a merge attention. Then the aggregated feature is transferred back into the task-specific feature for inference, helping the task-specific model capture diverse features. Extensive experiments show the effectiveness and superiority of TagFex on various CIL settings. Code is available at https://github.com/bwnzheng/TagFex_CVPR2025.", + "bbox": [ + 89, + 318, + 483, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 757, + 220, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For real-world applications of learning systems, new concepts and knowledge increase over time. Therefore, the ability to learn new concepts while preserve the learned knowledge is desirable for the learning system. This motivates the research in Continual Learning (or Incremental Learning). Class-Incremental Learning (CIL) [29, 46, 60]", + "bbox": [ + 89, + 782, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16c3665f3a4bfd9fcad34882baa8ac9429b3425b5421c646008c8355d04b4d75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 285, + 905, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37240b96634c3f68686f8ce40a047f877885765a253dacaa5fc9ac3147015077.jpg", + "image_caption": [ + "Figure 1. Average CKA feature similarities between different expanded models of DER and TagFex by task. It shows the expanded features learned by TagFex are less correlated.", + "Figure 2. GradCAM visualization comparison between different expanded models by DER and TagFex. It shows that the expanded models learned by the proposed framework focus on more diverse features." + ], + "image_footnote": [], + "bbox": [ + 516, + 503, + 906, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is one of the scenarios where new concepts incrementally emerge as new classes. The main challenge for Continual Learning is catastrophic forgetting [30], where the learning system usually forgets previously learned knowledge fast and catastrophically. Solving catastrophic forgetting alone is not enough for the model to learn new knowledge. Therefore, researchers seek for balance between stability (ability to resist changes) and plasticity (ability to adapt) [15].", + "bbox": [ + 511, + 688, + 905, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the promising strategies is to train a new feature extractor for each new task while keep the previously learned model intact during the training of the new task. Such methods are often referred to as expansion-based approaches [10, 20, 43, 51, 59]. Although expansion-based CIL achieve impressive performance on various scenarios", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 886, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10099", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2717197f2aecb83d630e3bf03780705b95000e7fccf46102be629461facf6f50.jpg", + "image_caption": [ + "(a) Feature Expansion Without Rehearsal Samples" + ], + "image_footnote": [], + "bbox": [ + 174, + 70, + 421, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6b00a305db853dc8e78cec9d796d9e702917790cc01fcf3b921a08515eb706b1.jpg", + "image_caption": [ + "(b) Feature Expansion with More Diverse Features", + "Figure 3. The illustration and mitigation of feature collision in expansion-based CIL. (a) Without the rehearsal samples, it is very difficult to distinguish two types of the fruits with the same color, because the features captured by just training on the newest task are insufficient to distinguish classes between two tasks. (b) With more diverse features, the model is able to further disengage from the spurious features, reducing the feature collision." + ], + "image_footnote": [], + "bbox": [ + 477, + 68, + 836, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in CIL, it relies on the feature diversity extracted by the new model. Generally, the models trained on different tasks capture different features. However, there could be new features in the new task colliding with the old ones, leading to classification bias towards the new task. To understand this, we could consider an extreme case where there are two incremental tasks with similar spurious features, e.g., task 1 with red apples and green pears, task 2 with red berries and green grapes. It is very difficult to distinguish two types of the fruits with the same color in such incremental scenarios. As shown in Figure 3a, even if we expand the feature, due to the lack of extra samples for feature capturing, the model would rely on the learned color feature to reach low loss on the new task, and finally make random guesses between the apple and the berry with a sample of red fruits, or simply leaning towards the newest task.", + "bbox": [ + 88, + 353, + 482, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, to mitigate feature collision, expansion-based methods often introduce an auxiliary classifier and save a small number of rehearsal samples from the previous tasks to distinguish current task samples with previous samples [10, 51]. With the auxiliary classifier, the new model is encouraged to capture diverse features other than the features extracted by the old models. However, the limited number of rehearsal samples makes the training of the auxiliary classifier imbalanced and suboptimal. To illustrate this, we perform experiments on the similarities between the learned expanded features. Specifically, we use CKA [23] as the similarity metric and calculate the similarities between each expanded feature on the initial task. The average results are shown in Figure 1, and the detailed numbers are shown in Figure 9 in the Appendix 11. From the figures, we can conclude that the features extracted by the models learned by existing expansion-based methods are relatively similar, with average CKA around 0.35, especially when comparing to TagFex proposed in this paper, with average CKA around 0.2. This can also be seen by Grad", + "bbox": [ + 91, + 598, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CAM [12, 38] visualizations in Figure 2. We can clearly see that the models learned by DER [51] focus on similar areas and those by TagFex focus more diverse areas.", + "bbox": [ + 511, + 353, + 903, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further encourage the diverse feature capture, a trivial solution would be simply increasing the number of rehearsal samples. But can we find more efficient approach to help the model acquire more discriminative features? To this end, we have to help the model to acquire such features that the model has not been able to extract in the classification task and take the advantage of such features for subsequent training. Figure 3 illustrates the general idea and motivation of the proposed framework.", + "bbox": [ + 511, + 400, + 905, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose to continually capture diverse features in a separate feature space, and then utilize it to help the incremental learning. Specifically, for the continual capture of diverse features, we have to find an approach to capture the features agnostic to the classification, which are referred to as task-agnostic features. Therefore, we apply continual self-supervised learning (CSSL) [4, 11, 16] to train a separate task-agnostic model to capture the task-agnostic features continually. In classification, the models are merely required to capture the minimal necessary features (task-specific features) for the classification within the task. Whereas, task-agnostic features contain more information that they are hard to be captured in the merely training of the classification task. Such information is vital for the new model to capture discriminative features from the new task. To make full utilization of the task-agnostic model, we further design a merge attention to adaptively unify task-agnostic features with task-specific features. Then, we also transfer the information in the unified representations back into the task-specific model. Therefore, we are able to mitigate the feature collision in CIL, improving the performance on various scenarios. The overall framework is called Task-Agnostic Guided Feature Expansion (TagFex). Moreover, extensive experiments are", + "bbox": [ + 511, + 539, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10100", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performed to show the effectiveness of the task-agnostic guided feature expansion and the superiority of TagFex against the state-of-the-art methods, even with memory-aligned protocol.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 165, + 240, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class-Incremental Learning (CIL) is one of the scenarios in continual learning where the model is learned task by task with a different set of classes. During inference, the task from which each sample is sampled is not available. Many techniques and frameworks are proposed to alleviate catastrophic forgetting and improve the performance in CIL. They can be roughly categorized into several forms, such as model expansion, rehearsal memory, model distillation and regularization.", + "bbox": [ + 89, + 190, + 482, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model expansion comes from the idea of parameter isolation for each task. It expands the feature space for each task. PNN [35] proposes learning a new backbone for each new task and fixing the former in incremental learning. It also adds layer-wise connections between old and new models to reuse former features. P&C [37] suggests a progression-compression protocol. It first expands the network to learn representative representations. After that, a compression process is conducted to control the total budget. DER [51] trains a separate backbone for each task, aggregating all of the features for classification. DyTox [10] learns a separate task token for each task. BEEF [44] incorporates energy-based theory to construct a bi-compatible framework. TCIL [20] aims to tackle the task confusion, establishing a multi-level knowledge distillation to propagate knowledge learned from the old tasks to the new one.", + "bbox": [ + 89, + 327, + 482, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are works introducing contrastive learning to continual learning, but from different perspectives. OCM [16] uses contrastive learning to get holistic representations in online incremental scenarios. C4IL [32] introduces contrastive learning on the learned features to alleviate interphase confusion in CIL. Most of them directly apply contrastive learning to the inference model, where the contrastive and classification loss are competing with each other over one model, differing from our proposed TagFex. We discuss other related works about CIL in Appendix 7.", + "bbox": [ + 89, + 568, + 482, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continual Self-Supervised Learning (CSSL) aims to continually learn good representations from non-stationary data. Rehearsal-based method LUMP [28] utilizes rehearsal samples to augment current task samples by mixup. Regularization-based method CaSSLe [11] encourages current model to maintain a consistency with previous state via a prediction head. $C^2$ ASR [4] selects the most representative and discriminative samples by estimating the augmentation stability for rehearsal.", + "bbox": [ + 89, + 719, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we treat the features in such representations as task-agnostic features, since they are agnostic to the classification task, which contain the information that", + "bbox": [ + 89, + 854, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "task-specific features do not contain. To the best of our knowledge, using task-agnostic features to mitigate feature collision in expansion-based CIL has rarely been explored. Our work proposes to continually capture task-agnostic features in each task, and utilizes such features to help the expanded feature acquire more diverse features, alleviating feature collision.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 213, + 651, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 238, + 718, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In CIL scenarios, we have multiple classification tasks to learn sequentially. Let $\\mathcal{D}_t$ be the training dataset of the $t$ th task. $(\\pmb{x}_i^{(t)},y_i^{(t)})\\in \\mathcal{D}_t$ is a sample. $\\pmb{x}_i^{(t)}$ is the input, $y_{i}^{(t)}$ is the label. Let $\\mathcal{C}_t = \\bigcup_i\\{y_i^{(t)}\\}$ be the class set of task $t$ . In CIL, $\\forall t_1\\neq t_2,\\mathcal{C}_{t_1}\\cap \\mathcal{C}_{t_2} = \\emptyset$ . In each task, we only train the model on $\\mathcal{D}_t$ , but test on all the tasks the model has trained on, i.e., the seen tasks. For example, when the model is training on task $t_i$ , the seen tasks are tasks $t_j(j\\leq i)$ . Therefore, the class set of the seen tasks is $\\mathcal{V}_t = \\bigcup_t\\mathcal{C}_t$ . In rehearsal-based CIL, a small number of rehearsal samples $\\mathcal{M}$ are allowed to be stored for later tasks. The number of rehearsal samples should be constant during the CIL training. The goal is to make the model get better performance on all the seen tasks.", + "bbox": [ + 511, + 261, + 906, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature Expansion of Task-Specific Models", + "text_level": 1, + "bbox": [ + 511, + 489, + 879, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section we introduce DER [51] as the basis of the feature expansion for our task-specific models. As shown in the middle part of Figure 4 (Task-specific Models), a new model $f_{\\mathrm{ts}}^{(t)}$ is created at the beginning of each task. Also, the previously learned models $f_{\\mathrm{ts}}^{(0)} \\sim f_{\\mathrm{ts}}^{(t-1)}$ are saved and frozen for feature extraction. The classifier is also expanded and the trained parameter weights from the old classifier is inherited. To predict the label of a sample, the concatenated feature from all of the models is used by the classifier. Formally, the classification loss $\\mathcal{L}_{\\mathrm{cls}}$ with cross-entropy (CE) for the sample $(\\boldsymbol{x}, \\boldsymbol{y})$ would be,", + "bbox": [ + 511, + 513, + 906, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l s}} (\\boldsymbol {x}, y) = \\ell_ {\\mathrm {C E}} ([ f _ {\\mathrm {t s}} ^ {(0)} (\\boldsymbol {x}), \\dots , f _ {\\mathrm {t s}} ^ {(t)} (\\boldsymbol {x}) ], y; W _ {\\mathrm {c l s}} ^ {(t)}), \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 696, + 906, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y \\in \\mathcal{Y}_t$ and $W_{\\mathrm{cls}}^{(t)}$ is the parameter weights of the classifier in task $t$ .", + "bbox": [ + 511, + 732, + 903, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, to help the new model to capture diverse features, an auxiliary classifier is trained based on the new feature. The training objective for the auxiliary classifier is to not only distinguish classes within the current task, but also distinguish the samples from the previous tasks. Consequently, a small number of rehearsal samples are saved from the previous tasks. Therefore, the auxiliary classifier outputs $|\\mathcal{C}_t| + 1$ logits, where the old classes are treated as a whole. Formally, the auxiliary loss $\\mathcal{L}_{\\mathrm{aux}}$ with cross-entropy", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3912d895934282c6c3fcb6f20daabb9484e6bea74cad9dc75b67c194d17a6169.jpg", + "image_caption": [ + "Figure 4. Overview of the proposed framework TagFex. The framework is composed of three parts: Task-agnostic Model (Section 4.1), Merge Attention (Section 4.2) and Task-specific Models (Section 3.2 and 4.3). Overall, our framework continually captures the task-agnostic feature in each task and transfers such feature back to the task-specific feature, helping the task-specific model obtain more diverse features to combat the feature collision." + ], + "image_footnote": [], + "bbox": [ + 174, + 66, + 823, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for the sample $(\\pmb {x},\\pmb {y})$ would be,", + "bbox": [ + 89, + 406, + 300, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a u x}} (\\boldsymbol {x}, y) = \\ell_ {\\mathrm {C E}} \\left(f _ {\\mathrm {t s}} ^ {(t)} (\\boldsymbol {x}), y; W _ {\\mathrm {a u x}} ^ {(t)}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 431, + 480, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $y \\in [|\\mathcal{C}_t| + 1]$ and $W_{\\mathrm{aux}}^{(t)}$ is the parameter weights of the auxiliary classifier in task $t$ .", + "bbox": [ + 89, + 460, + 483, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Discussions on the Feature Diversity in Task-Specific Models", + "text_level": 1, + "bbox": [ + 89, + 500, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In supervised classification within one task, the model is only required to capture the minimal necessary features for the classification within the task. For example, in binary classification of rabbit and whale, the model only needs to distinguish the colors of the background and the object to minimize the classification loss. Such features are useful within a specific task, we thus refer to such features as task-specific features. Therefore, the task-specific models trained by DER can only extract the task-specific features for each task. However, apart from the task-specific features, there are plenty of features not discovered by these models. Such features are not necessarily related to the classification of the task, so we refer them as task-agnostic features. If these features are available for subsequent tasks, the new task-specific model can refer to them for better feature learning.", + "bbox": [ + 88, + 537, + 482, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DER uses rehearsal samples and the auxiliary classifier to capture diverse features, which makes the situation better. However, since the number of rehearsal samples are limited, the imbalanced training makes it suboptimal to capture more diverse features. Therefore, to further improve the ability to capture more diverse features, we explore how to incorporate the task-agnostic feature into this task-specific features in the next section.", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Task-Agnostic Guided Feature Expansion", + "text_level": 1, + "bbox": [ + 511, + 404, + 885, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Continual Acquisition of Task-Agnostic Features", + "text_level": 1, + "bbox": [ + 511, + 434, + 905, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by recent studies from continual self-supervised learning [4, 11, 28], not only can we extract task-agnostic features with self-supervised learning, but also continually extract new task-agnostic features while keeping the old ones. For the continual acquisition of task-agnostic features, we follow CaSSLe [11] to continually capture more features in a predictive manner. Also, we take SimCLR [3] as the base self-supervised method for CaSSLe and discuss other self-supervised methods in section 5.5. As shown in the left part of Figure 4 (Task-agnostic Model), we train a separate model to maintain the task-agnostic feature for the learned tasks. So that, the task-agnostic feature will not be affected by the classification objectives. Also, we save the trained task-agnostic model for the next task, in order to support the predictive knowledge preservation in CaSSLe. The overall loss for the task-agnostic model in the initial task and incremental tasks are,", + "bbox": [ + 511, + 477, + 906, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t a}} ^ {(t = 0)} = - \\operatorname {I n f o N C E} \\left(\\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} _ {i = 1} ^ {B}, \\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i} ^ {\\prime}\\right) \\right\\} _ {i = 1} ^ {B}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 753, + 906, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {t a}} ^ {(t > 0)} = - \\operatorname {I n f o N C E} \\left(\\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} _ {i = 1} ^ {B}, \\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i} ^ {\\prime}\\right) \\right\\} _ {i = 1} ^ {B}\\right) (4) \\\\ - \\operatorname {I n f o N C E} \\left(\\left\\{g \\left(f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i}\\right)\\right) \\right\\} _ {i = 1} ^ {B}, \\left\\{f _ {\\mathrm {t a}} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} _ {i = 1} ^ {B}\\right), (5) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 776, + 903, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {w h e r e} \\quad \\operatorname {I n f o N C E} (\\{\\boldsymbol {z} _ {i} \\} _ {i = 1} ^ {B}, \\{\\boldsymbol {z} ^ {\\prime} \\} _ {i = 1} ^ {B}) \\\\ = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\sin \\left(z _ {i} , z _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\sum_ {j = 1 , j \\neq i} ^ {B} \\exp \\left(\\sin \\left(z _ {i} , z _ {j} ^ {\\prime}\\right) / \\tau\\right)}, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 834, + 903, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10102", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equation 3 and 4 are contrastive InfoNCE [33] loss to maximize the lower bound of the mutual information between the sample and its representation. The last term in equation 5 aims to train a predictor $g(\\cdot)$ to predict the last task-agnostic feature $(f_{\\mathrm{ta}}^{\\prime}(\\boldsymbol{x}_i))$ with the current one $(f_{\\mathrm{ta}}(\\boldsymbol{x}_i))$ , where $f_{\\mathrm{ta}}^{\\prime}(\\cdot)$ is the last task-agnostic model we save. $\\sin (\\cdot, \\cdot)$ in Equation 6 includes another projection head where we apply cosine similarity in the projected feature space. The core idea of the predictive loss is that if such loss is minimized, the current feature at least contains as much information as the last feature contains. Therefore, the expressiveness of the task-agnostic feature is increasing task by task as the incremental learning proceeds.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Adaptive Feature Aggregation by Merge Attention", + "text_level": 1, + "bbox": [ + 89, + 297, + 483, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now we have a powerful feature extractor to continually capture the task-agnostic features from all of the tasks we have seen. To put the task-agnostic model into good use, we design the merge attention to adaptively unify task-agnostic features with task-specific features in the incremental tasks.", + "bbox": [ + 89, + 335, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the right part of Figure 4 (Merge Attention), the merge attention takes the feature maps of both the task-agnostic and task-specific features of a sample as inputs, denoting them as $F_{\\mathrm{ta}}(\\pmb{x})$ and $F_{\\mathrm{ts}}(\\pmb{x})$ respectively. Like the patch embedding of ViT [8], we treat each spacial feature in the feature map $(H \\times W$ features) as a patch token in our attention. The feature maps are first normalized by two layer norms for training stability.", + "bbox": [ + 89, + 411, + 483, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {\\mathrm {t s}} = \\operatorname {L a y e r N o r m} _ {1} \\left(F _ {\\mathrm {t s}} (\\boldsymbol {x})\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 545, + 482, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {\\mathrm {t a}} = \\operatorname {L a y e r N o r m} _ {2} (\\operatorname {S t o p G r a d i e n t} (F _ {\\mathrm {t a}} (\\boldsymbol {x}))). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 564, + 480, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, in order to absorb useful features from task-agnostic features, we take $z_{\\mathrm{ts}}$ as the query, making them attend to both $z_{\\mathrm{ts}}$ and $z_{\\mathrm{ta}}$ . Furthermore, since $z_{\\mathrm{ts}}$ and $z_{\\mathrm{ta}}$ are in different feature spaces, we calculate their keys and values separately.", + "bbox": [ + 89, + 592, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = W _ {q} \\boldsymbol {z} _ {\\mathrm {t s}}, \\quad K _ {\\mathrm {t s}} = W _ {k} ^ {(\\mathrm {t s})} \\boldsymbol {z} _ {\\mathrm {t s}}, \\quad V _ {\\mathrm {t s}} = W _ {v} ^ {(\\mathrm {t s})} \\boldsymbol {z} _ {\\mathrm {t s}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 680, + 482, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\mathrm {t a}} = W _ {k} ^ {(\\mathrm {t a})} \\boldsymbol {z} _ {\\mathrm {t a}}, V _ {\\mathrm {t a}} = W _ {v} ^ {(\\mathrm {t a})} \\boldsymbol {z} _ {\\mathrm {t a}} (1 0)\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 702, + 480, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then concatenate their keys for the calculation of the attention map. Finally, we get the merged feature map by multi-head scaled dot product attention [1, 42] with concatenated keys and values. For each head of the embedding:", + "bbox": [ + 89, + 733, + 483, + 794 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO ^ {(h)} = \\operatorname {S o f t m a x} \\left(\\frac {Q ^ {(h)} \\left[ K _ {\\mathrm {t s}} ^ {(h)} , K _ {\\mathrm {t a}} ^ {(h)} \\right] ^ {T}}{\\sqrt {d / h}}\\right) \\left[ V _ {\\mathrm {t s}} ^ {(h)}, V _ {\\mathrm {t a}} ^ {(h)} \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 805, + 483, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d$ is the embedding dimension of each patch token, $h$ is the number of heads. Such extended self-attention makes the queries not only attend to task-specific features", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "themselves, but also attend to task-agnostic features. Therefore, the merged feature contains not only the information learned from the current task, but also relevant information from the task-agnostic feature.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To guide the merge process, we use global average pooling to get the feature for the merge classifier. The merge classifier is trained to predict the label of the current task samples.", + "bbox": [ + 511, + 151, + 905, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m c l s}} = \\ell_ {\\mathrm {C E}} (\\operatorname {G A P} (O), y). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 212, + 903, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that we do not propagate the gradients to the task-agnostic model. Therefore, no task-specific information about the classification would affect the task-agnostic model, keeping the task-agnostic feature uncorrelated with task-specific ones.", + "bbox": [ + 511, + 233, + 905, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Task-agnostic Guided Knowledge Transferring", + "text_level": 1, + "bbox": [ + 511, + 316, + 905, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the merge attention, we have obtained more diverse features and their corresponding logits. However, we do not directly use the logits of the merge classifier for the final inference, since the distributional drift happens when the task-agnostic model is continually updated as the incremental training goes on. We have to transfer the diverse features from merged features to task-specific features. In the end, we will only use the task-specific models with the classifier for inference.", + "bbox": [ + 511, + 339, + 905, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the knowledge transferring, we use KL divergence between the logits of the merge classifier $(\\log (p_{\\mathrm{m}}))$ and task-specific classifier $(\\log (p_{\\mathrm{ts}}))$ whose weights $W_{\\mathrm{cls}}^{(t)}$ are mentioned in section 3.2 within the current task, which is,", + "bbox": [ + 511, + 476, + 905, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a n s}} = D _ {\\mathrm {K L}} \\left(\\operatorname {S t o p G r a d i e n t} \\left(p _ {\\mathrm {m}} ^ {(i)}\\right) \\| p _ {\\mathrm {t s}} ^ {(i)}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 546, + 903, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, with the losses mentioned in section 3.2, the overall loss we apply on the task-specific model is,", + "bbox": [ + 511, + 575, + 905, + 606 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t s}} = \\mathcal {L} _ {\\mathrm {c l s}} + \\mathcal {L} _ {\\mathrm {a u x}} + \\mathcal {L} _ {\\mathrm {t r a n s}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 617, + 903, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, since the expansion of the task-specific models suffers from rapidly increasing number of parameters at both training and inference. To eliminate such drawback for practical use of our method, we adopt the FPGM [18] as the pruning strategy. It takes the geometric median within the kernel pairs as the similarity. Then the kernel pairs whose similarity above a threshold are treated as redundant and one of them is pruned. With this strategy, the model size could be largely reduced while not affecting the accuracy too much. Since TagFex only uses task-specific models for inference, we only prune the task-specific models. With such pruning strategy, the method is called TagFex-P.", + "bbox": [ + 511, + 643, + 906, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Summary", + "text_level": 1, + "bbox": [ + 511, + 833, + 625, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the proposed framework TagFex is composed of three parts: 1) the continual acquisition of the task-agnostic features, 2) the adaptive feature aggregation with", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10103", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cdf20d1e4de0f4461af905fc3f87f876de0e89a429c5a2fe87f56990dd76b58a.jpg", + "table_caption": [ + "Table 1. Performance results. Bold font represents the best results in the scenario. TagFex-P is the method with pruning strategy." + ], + "table_footnote": [], + "table_body": "
MethodsCIFAR100ImageNet100ImageNet1000
10-1050-1010-1050-10100-100
LastAvgLastAvgLastAvgLastAvgLastAvg
iCaRL [34]49.5264.6450.5660.0850.9867.1153.6962.5640.4757.55
BiC [50]50.7965.3843.8257.0442.4065.1349.9066.36--
DyTox [10]60.4373.50--67.6176.51--59.7568.14
BEEF [44]60.9871.9463.5170.7168.7877.6270.9877.2758.6767.09
DER [51]64.3575.3665.2772.6066.7177.1871.0877.7158.8366.87
TagFex68.2378.4570.3375.8770.8479.2775.5480.6461.4568.32
TagFex-P67.3478.0269.2674.2469.2178.5674.1379.8560.1467.65
", + "bbox": [ + 130, + 90, + 874, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/58871b260329848775ab5c02dc98c6585b1658589a936b1979b3169bd8fd64d0.jpg", + "image_caption": [ + "(a) CIFAR100 10-10" + ], + "image_footnote": [], + "bbox": [ + 93, + 297, + 294, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/32ef14dd48f004008bb202fb7bfdc870bad425b76ae763e1a745b2523b23f0a4.jpg", + "image_caption": [ + "(b) CIFAR100 50-10" + ], + "image_footnote": [], + "bbox": [ + 295, + 297, + 496, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/678b2748fdebb18395d61d6ef6516bcae33144b717281dc081c886e7c2912a81.jpg", + "image_caption": [ + "(c) ImageNet100 10-10", + "Figure 5. Accuracy curves for baselines and TagFex on various scenarios." + ], + "image_footnote": [], + "bbox": [ + 496, + 297, + 699, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/920ae2b3eafead75aa601e4816a935823168827bf080659250da194e78929289.jpg", + "image_caption": [ + "(d) ImageNet100 50-10" + ], + "image_footnote": [], + "bbox": [ + 700, + 297, + 901, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "merge attention, 3) the task-agnostic guided knowledge transferring. The task-agnostic features are acquired continually by a separate model, and such task-agnostic information is first aggregated with the task-specific features and then transferred back into the task-specific features. The overall loss of our framework during the training of each task is the weighted sum of the three parts,", + "bbox": [ + 88, + 464, + 482, + 570 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\mathrm {t a}} \\mathcal {L} _ {\\mathrm {t a}} + \\lambda_ {\\mathrm {m c l s}} \\mathcal {L} _ {\\mathrm {m c l s}} + \\mathcal {L} _ {\\mathrm {t s}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 578, + 480, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For practical use and fair comparison, we further adopt pruning strategy to compress the task-specific models without affecting the accuracy too much.", + "bbox": [ + 89, + 599, + 483, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 657, + 223, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 89, + 681, + 287, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Rehearsal Memory. Following most of the image classification benchmarks in CIL [34, 50], we use CIFAR100 and ImageNet to train the model incrementally. CIFAR100 [24] has 50,000 training and 10,000 testing samples with 100 classes in total. Each sample is a tiny image in $32 \\times 32$ pixels. ImageNet [6] has 1,300 training samples and 50 test samples for each class. The maximum number of rehearsal samples is 2,000 as default for CIFAR100 and ImageNet100, and 20,000 for ImageNet1000. We use herding [34] for the selection and eviction of the rehearsal samples between each task.", + "bbox": [ + 89, + 703, + 482, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Split. There are two common types of splits in CIL. The small base one equally divides all of the classes in", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a dataset [34]. The large base one uses half of the classes in a dataset as the base task (task 0), and equally divides the remaining classes [19, 52]. For a dataset with 100 classes, 10-10 means 10 classes in the base task and incremental tasks are also with 10 classes each, 50-10 means 50 classes in the base task and 10 classes in the incremental tasks.", + "bbox": [ + 511, + 464, + 905, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Backbones and Implementation. We follow DER [51] to use ResNet18 [17] as the backbone model on both CIFAR and ImageNet for both task-agnostic and task-specific models. The experiments are implemented based on the open-source code PyCIL [58]. We provide more implementation details in Appendix 8.", + "bbox": [ + 511, + 556, + 906, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Performance Results", + "text_level": 1, + "bbox": [ + 511, + 664, + 709, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test our proposed TagFex and TagFex-P in various class incremental scenarios. And we report the accuracy of the test dataset at the end of the entire incremental training (Last), and the average test accuracy at the end of each task training across all of the incremental stages (Avg). We run the DER and TagFex experiments for 5 times and report their mean values. The numerical results are shown in Table 1. As we can see from the table, TagFex achieves the best results on all of the scenarios, and generally achieves $3\\sim 4\\%$ percent performance gap compared to DER. The pruned TagFex-P shows minor negative impact on the performance, which effectively reduces the size of the model. As for the number of parameters, since TagFex does not include extra modules for inference, it holds the same num", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10104", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dff91846edeed4788ad71052cb87b2eec84ca3fca48599bc0d80885646e8704d.jpg", + "image_caption": [ + "(d) Task 4 Epoch 0" + ], + "image_footnote": [], + "bbox": [ + 91, + 65, + 385, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/055721836421157a33380a50ec8954dff5b6f71b64e7c25ee706fb461435c36d.jpg", + "image_caption": [ + "(e) Task 4 Epoch 80", + "Figure 6. Evolution of the attention map averaged across the samples in the merge attention. In each sub-figure, each row represents the attention weights for a query patch attending to key patches (red pixels means high attention weights). The left part of each sub-figure is the attention weights for task-agnostic key patches $(K_{\\mathrm{ta}})$ , and the right part is the attention weights for task-specific key patches $(K_{\\mathrm{ts}})$ which are separated by a dashed line in the middle. From the evolution of the attention maps, we can conclude that for each task, the high attention values are first on the task-agnostic side, and then transfer to the task-specific side, which represents task-agnostic features are useful for the task at early epochs, and such information is absorbed by the task-specific model at the end of the task. Note that the attention map is averaged across the samples, so the attention map appears more uniform." + ], + "image_footnote": [], + "bbox": [ + 400, + 66, + 645, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8498751de8ae6f0262854e76887551eef22624b5a9113268dc487052644ba222.jpg", + "image_caption": [ + "(f) Task 4 Epoch 160" + ], + "image_footnote": [], + "bbox": [ + 661, + 68, + 905, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ber of parameters as DER. TagFex-P effectively reduces the averaged incremental number of parameters from 61.6M to 11.6M on CIFAR100 10-10, and from 61.6M to 14.4M on ImageNet100 10-10. Additionally, the accuracy curves for baselines and TagFex on various scenarios are shown in Figure 5, indicating that TagFex achieve a large performance gap compared to baseline methods. More comparisons for the number of parameters are shown in Table 5 in Appendix 10.", + "bbox": [ + 88, + 420, + 483, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Memory Efficiency", + "text_level": 1, + "bbox": [ + 89, + 569, + 272, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we investigate the memory efficiency of our framework. TagFex saves an extra model for the next task to continually capture the task-agnostic features compared to DER [51]. For ResNet18 on ImageNet, a model occupies the same memory as around 300 rehearsal samples [59]. It is unfair for methods with less memory saving for the next task. To eliminate such unfairness, we perform experiments in the memory-aligned protocol, where we compare TagFex to DER with more rehearsal samples. The results are shown in Table 4. As we can see from the tables, TagFex is still competitive against DER with comparable number of extra rehearsal samples. This is more obvious when the memory size is larger (Memory Size 3000, 3300), since TagFex uses the rehearsal samples more effectively than DER.", + "bbox": [ + 88, + 592, + 482, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Visualizations on Merge Attention Maps", + "text_level": 1, + "bbox": [ + 89, + 816, + 437, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the efficacy of the merge attention, we investigate the evolution of the attention map in the merge attention during the training on ImageNet100. We record the attention map of some samples at the end of some check-", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "point epochs within each task, and see how the attention map evolves during the training of the task. The results are shown in Figure 6. From the evolution of the attention maps, we can conclude that for the incremental tasks the high attention values are first on the task-agnostic side, and then transfer to the task-specific side, which represents task-agnostic features are useful for the task at early epochs, and such information is absorbed by the task-specific model at the end of the task. The visualizations of more tasks are shown in Appendix 12.", + "bbox": [ + 511, + 420, + 906, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 587, + 681, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TagFex with different strategies for continual learning of the task-agnostic model. In section 4.1, we use SimCLR [3] as the default configuration in the learning of the task-agnostic model. However, we can use different self-supervised losses on the task-agnostic model. We test TagFex with three different self-supervised methods on CIFAR100, which is shown in Table 2. The three methods are VICReg [2] which uses a mix of variance, invariance and covariance regularizations, BYOL [14] which uses a momentum encoder, and Barlow Twins [53] which considers an objective function measuring the cross-correlation matrix between the features. The results show that TagFex is compatible to different self-supervised methods and performs well with them. It is worth to note that BYOL and Barlow Twins get better results than the default configuration of TagFex. However, they are more complicated which require another momentum updated model for training.", + "bbox": [ + 511, + 611, + 906, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameter analysis. TagFex has two hyperparameters $\\lambda_{\\mathrm{ta}}$ and $\\lambda_{\\mathrm{mcls}}$ controlling the loss from the task", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10105", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/73d0f0a176e5517616e6a545274dc0ac130a6be5122c4278a0c138498d449c4f.jpg", + "table_caption": [ + "Table 2. Results for TagFex with other self-supervised methods." + ], + "table_footnote": [], + "table_body": "
methods10-1050-10
LastAvgLastAvg
TagFex68.2378.4570.3375.87
TagFex w/ VICReg67.8578.2169.6875.44
TagFex w/ Barlow Twins68.6278.5570.2475.82
TagFex w/ BYOL69.0478.7671.1076.31
", + "bbox": [ + 116, + 112, + 558, + 241 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b6afd622da1bf0de65721334743ab2160ae4c752f950f6cffc7d7c4e2407b9bc.jpg", + "table_caption": [ + "Table 3. Ablation study on CIFAR100." + ], + "table_footnote": [], + "table_body": "
Task-agnostic Model\nw/ Merge AttentionContinual Task-agnostic ModelKnowledge Transferring10-1050-10
LastAvgLastAvg
✓✓64.4575.3465.4572.58
✓✓65.8676.3267.5173.29
✓✓✓68.2378.4570.3375.87
", + "bbox": [ + 99, + 303, + 629, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "agnostic model and the merge classifier respectively. We find the best combination of them through grid search on the average incremental accuracy (Avg). We perform grid search on CIFAR100 10-10 scenario. The results are shown in Figure 7. As we can see from the figure, TagFex is quite robust to these two hyperparameters. The performance is generally better when $\\lambda_{\\mathrm{ta}} \\geq 1$ The best performance is achieved when their values are balanced and their values are both 1.", + "bbox": [ + 89, + 406, + 482, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. To verify the effectiveness of each part of TagFex, we perform ablation study on three components of TagFex. Since training the task-agnostic model alone would not influence the task-specific models which are used for inference, and the merge attention cannot work alone, we have to bind them at first (task-agnostic model w/ merge attention). Then, we add continual learning to the task-agnostic model (continual task-agnostic model), and do not transfer the knowledge from the task-agnostic model to the task-specific model. Finally, we compare knowledge transferring with and without the continual learning of the task-agnostic model. We evaluate the configurations on CIFAR100, and compare their average incremental accuracies (Avg). The results are shown in Table 3. As we can see, without the final knowledge transferring, we cannot effectively transfer the information into the task-specific model for inference. Also, the best performance is achieved with continual learning of task-agnostic model.", + "bbox": [ + 89, + 542, + 482, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 829, + 209, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we explored the solutions for the feature collision in expansion-based class-incremental learning. We illustrates the feature collision and propose a framework", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ced331c96f8add2807ffeab411590e3644dc9ff58669d24c1c703832c2668b18.jpg", + "image_caption": [ + "Figure 7. Hyperparameter sensitivity analysis on CIFAR100 10-10." + ], + "image_footnote": [], + "bbox": [ + 578, + 71, + 911, + 223 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/73f6dea5bc9573bf2b9d509489bb5e757bb33232eeae4c615bc8d6dc822a745c.jpg", + "table_caption": [ + "Table 4. Memory-aligned average accuracy comparison on ImageNet100." + ], + "table_footnote": [], + "table_body": "
MethodsMemory SizeScenarios
10-1050-10
DER230077.5678.21
TagFex200079.2780.64
DER330078.4379.12
TagFex300080.8781.96
", + "bbox": [ + 647, + 303, + 898, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "called TagFex to transfer diverse features into new task-specific model. In order to acquire such diverse features, we propose to continually train a separate model to capture task-agnostic features in advance for later tasks. Then, we design a merge attention to adaptively select and aggregate the features to transfer. Finally, we transfer such features back into the task-specific model. Furthermore, we not only apply pruning techniques to reduce the number of parameters, but also perform extensive experiments to show the superiority of our proposed framework. Also, TagFex has competitive memory efficiency against simply adding more rehearsal samples. From the visualization on merge attention maps, we conclude that the high attention values are transferred from the task-agnostic part to the task-specific part, indicating that the knowledge is absorbed by the task-specific model. We also perform further analysis to thoroughly evaluate our framework.", + "bbox": [ + 511, + 406, + 906, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: Creating a new model for a new task would dramatically increase the number of parameters, even though we could address it by pruning techniques. In the future work, we would like to seek for more memory efficient methods and techniques.", + "bbox": [ + 511, + 662, + 906, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 513, + 765, + 681, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially supported by National Science and Technology Major Project (2022ZD0114805), NSFC (62476123, 62376118, 62006112, 62250069), Fundamental Research Funds for the Central Universities (2024300373,14380021), CCF-Tencent Rhino-Bird Open Research Fund RAGR20240101, Collaborative Innovation Center of Novel Software Technology and Industrialization.", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10106", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 5", + "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In International Conference on Learning Representations, 2022. 7", + "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4, 7", + "[4] Haoyang Cheng, Haitao Wen, Xiaoliang Zhang, Heqian Qiu, Lanxiao Wang, and Hongliang Li. Contrastive continuity on augmentation stability rehearsal for continual self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5707-5717, 2023. 2, 3, 4", + "[5] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 1", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255. IEEE, 2009. 6", + "[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 1", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5", + "[9] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 86-102. Springer, 2020. 1", + "[10] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9285-9295, 2022. 1, 2, 3, 6", + "[11] Enrico Fini, Victor G Turrisi Da Costa, Xavier Alameda-Pineda, Elisa Ricci, Karteek Alahari, and Julien Mairal. Self-supervised models are continual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9621-9630, 2022. 2, 3, 4", + "[12] Jacob Gildenblat and contributors. Pytorch library for" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.2", + "[13] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost Van De Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 1", + "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 7", + "[15] Stephen Grossberg. Adaptive resonance theory: How a brain learns to consciously attend, learn, and recognize a changing world. Neural networks, 37:1-47, 2013. 1", + "[16] Yiduo Guo, Bing Liu, and Dongyan Zhao. Online continual learning through mutual information maximization. In International Conference on Machine Learning, pages 8109-8126. PMLR, 2022. 2, 3", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6", + "[18] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4340-4349, 2019. 5", + "[19] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 6", + "[20] Bingchen Huang, Zhineng Chen, Peng Zhou, Jiayin Chen, and Zuxuan Wu. Resolving task confusion in dynamic expansion architectures for class incremental learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 908-916, 2023. 1, 3", + "[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European conference on computer vision (ECCV), pages 709-727. Springer, 2022. 1", + "[22] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 1", + "[23] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 2", + "[24] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6" + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10107", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017. 1", + "[26] Guoliang Lin, Hanlu Chu, and Hanjiang Lai. Towards better plasticity-stability trade-off in incremental learning: A simple linear connector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 89-98, 2022. 1", + "[27] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12245-12254, 2020. 1", + "[28] Divyam Madaan, Jaehong Yoon, Yuanchun Li, Yunxin Liu, and Sung Ju Hwang. Representational continuity for unsupervised continual learning. In International Conference on Learning Representations, 2022. 3, 4", + "[29] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1", + "[30] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychol. Learn. Motiv., pages 109-165. Elsevier, 1989. 1", + "[31] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Dilan Gorur, Razvan Pascanu, and Hassan Ghasemzadeh. Linear mode connectivity in multitask and continual learning. In International Conference on Learning Representations, 2020. 1", + "[32] Zixuan Ni, Haizhou Shi, Siliang Tang, Longhui Wei, Qi Tian, and Yueting Zhuang. Revisiting catastrophic forgetting in class incremental learning. arXiv preprint arXiv:2107.12308, 2021. 3", + "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5", + "[34] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2001-2010, 2017. 6, 1", + "[35] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3", + "[36] Dawid Rymarczyk, Joost Van De Weijer, Bartosz Zielinski, and Bartlomiej Twardowski. Icicle: Interpretable class incremental continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1887-1898, 2023. 1", + "[37] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International conference on machine learning, pages 4528-4537. PMLR, 2018. 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2", + "[39] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16722-16731, 2022. 1", + "[40] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in Neural Information Processing Systems, 30, 2017. 1", + "[41] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 99-108, 2022. 1", + "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[43] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 398–414. Springer, 2022. 1", + "[44] Fu-Yun Wang, Da-Wei Zhou, Liu Liu, Han-Jia Ye, Yatao Bian, De-Chuan Zhan, and Peilin Zhao. Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion. In The Eleventh International Conference on Learning Representations, 2023. 3, 6", + "[45] Jingya Wang, Yu Zhang, Bin Zhang, Jinxiang Xia, and Weidong Wang. Ipfa-net: Important points feature aggregating net for point cloud classification and segmentation. Chinese Journal of Electronics, 34(1):322-337, 2025.", + "[46] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[47] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In Advances in Neural Information Processing Systems, 2022. 1", + "[48] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In Computer Vision-ECCV 2022: 17th European Conference, Part XXVI, pages 631-648. Springer, 2022. 1", + "[49] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10108", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 374-382, 2019. 6, 1", + "[51] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3014-3023, 2021. 1, 2, 3, 6, 7", + "[52] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6982-6991, 2020. 6", + "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 7", + "[54] Ruru Zhang, E Haihong, and Meina Song. Fscil-eaca: Few-shot class-incremental learning network based on embedding augmentation and classifier adaptation for image classification. Chinese Journal of Electronics, 33(1):139-152, 2024.", + "[55] Tao Zhang, Ying Fu, and Jun Zhang. Deep guided attention network for joint denoising and demosaicing in real image. Chinese Journal of Electronics, 33(1):303-312, 2024.", + "[56] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Preserving locality in vision transformers for class incremental learning. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1157–1162, 2023. 1", + "[57] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Multi-layer rehearsal feature augmentation for class incremental learning. In Proceedings of the 41st International Conference on Machine Learning, pages 61649-61663. PMLR, 2024. 1", + "[58] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, and De-Chuan Zhan. Pycil: a python toolbox for class incremental learning. SCIENCE CHINA Information Sciences, 66(9):197101–, 2023. 6", + "[59] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning. In The Eleventh International Conference on Learning Representations, 2023. 1, 7", + "[60] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9851-9873, 2024. 1", + "[61] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 1" + ], + "bbox": [ + 91, + 92, + 480, + 824 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10109", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_model.json b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1fe2a24b4803e781e61ad14cf1f1d44be5402449 --- /dev/null +++ b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "Task-Agnostic Guided Feature Expansion for Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.181, + 0.832, + 0.234 + ], + "angle": 0, + "content": "Bowen Zheng Da-Wei Zhou† Han-Jia Ye De-Chuan Zhan \nNational Key Laboratory for Novel Software Technology, Nanjing University, China \nSchool of Artificial Intelligence, Nanjing University, China" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.237, + 0.696, + 0.253 + ], + "angle": 0, + "content": "{zhengbw, zhoudw, yehj, zhandc}@lamda.nju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.485, + 0.729 + ], + "angle": 0, + "content": "The ability to learn new concepts while preserve the learned knowledge is desirable for learning systems in Class-Incremental Learning (CIL). Recently, feature expansion of the model become a prevalent solution for CIL, where the old features are fixed during the training of the new task while new features are expanded for the new tasks. However, such task-specific features learned from the new task may collide with the old features, leading to misclassification between tasks. Therefore, the expanded model is often encouraged to capture diverse features from the new task, aiming to avoid such collision. However, the existing solution is largely restricted to the samples from the current task, because of the poor accessibility to previous samples. To promote the learning and transferring of diverse features across tasks, we propose a framework called Task-Agnostic Guided Feature Expansion (TagFex). Firstly, it captures task-agnostic features continually with a separate model, providing extra task-agnostic features for subsequent tasks. Secondly, to obtain useful features from the task-agnostic model for the current task, it aggregates the task-agnostic features with the task-specific feature using a merge attention. Then the aggregated feature is transferred back into the task-specific feature for inference, helping the task-specific model capture diverse features. Extensive experiments show the effectiveness and superiority of TagFex on various CIL settings. Code is available at https://github.com/bwnzheng/TagFex_CVPR2025." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.758, + 0.222, + 0.773 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.876 + ], + "angle": 0, + "content": "For real-world applications of learning systems, new concepts and knowledge increase over time. Therefore, the ability to learn new concepts while preserve the learned knowledge is desirable for the learning system. This motivates the research in Continual Learning (or Incremental Learning). Class-Incremental Learning (CIL) [29, 46, 60]" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.286, + 0.906, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.46, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Figure 1. Average CKA feature similarities between different expanded models of DER and TagFex by task. It shows the expanded features learned by TagFex are less correlated." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.684 + ], + "angle": 0, + "content": "Figure 2. GradCAM visualization comparison between different expanded models by DER and TagFex. It shows that the expanded models learned by the proposed framework focus on more diverse features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.81 + ], + "angle": 0, + "content": "is one of the scenarios where new concepts incrementally emerge as new classes. The main challenge for Continual Learning is catastrophic forgetting [30], where the learning system usually forgets previously learned knowledge fast and catastrophically. Solving catastrophic forgetting alone is not enough for the model to learn new knowledge. Therefore, researchers seek for balance between stability (ability to resist changes) and plasticity (ability to adapt) [15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "One of the promising strategies is to train a new feature extractor for each new task while keep the previously learned model intact during the training of the new task. Such methods are often referred to as expansion-based approaches [10, 20, 43, 51, 59]. Although expansion-based CIL achieve impressive performance on various scenarios" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.071, + 0.422, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.263, + 0.43, + 0.275 + ], + "angle": 0, + "content": "(a) Feature Expansion Without Rehearsal Samples" + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.069, + 0.838, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.262, + 0.794, + 0.273 + ], + "angle": 0, + "content": "(b) Feature Expansion with More Diverse Features" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.282, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 3. The illustration and mitigation of feature collision in expansion-based CIL. (a) Without the rehearsal samples, it is very difficult to distinguish two types of the fruits with the same color, because the features captured by just training on the newest task are insufficient to distinguish classes between two tasks. (b) With more diverse features, the model is able to further disengage from the spurious features, reducing the feature collision." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.483, + 0.595 + ], + "angle": 0, + "content": "in CIL, it relies on the feature diversity extracted by the new model. Generally, the models trained on different tasks capture different features. However, there could be new features in the new task colliding with the old ones, leading to classification bias towards the new task. To understand this, we could consider an extreme case where there are two incremental tasks with similar spurious features, e.g., task 1 with red apples and green pears, task 2 with red berries and green grapes. It is very difficult to distinguish two types of the fruits with the same color in such incremental scenarios. As shown in Figure 3a, even if we expand the feature, due to the lack of extra samples for feature capturing, the model would rely on the learned color feature to reach low loss on the new task, and finally make random guesses between the apple and the berry with a sample of red fruits, or simply leaning towards the newest task." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Therefore, to mitigate feature collision, expansion-based methods often introduce an auxiliary classifier and save a small number of rehearsal samples from the previous tasks to distinguish current task samples with previous samples [10, 51]. With the auxiliary classifier, the new model is encouraged to capture diverse features other than the features extracted by the old models. However, the limited number of rehearsal samples makes the training of the auxiliary classifier imbalanced and suboptimal. To illustrate this, we perform experiments on the similarities between the learned expanded features. Specifically, we use CKA [23] as the similarity metric and calculate the similarities between each expanded feature on the initial task. The average results are shown in Figure 1, and the detailed numbers are shown in Figure 9 in the Appendix 11. From the figures, we can conclude that the features extracted by the models learned by existing expansion-based methods are relatively similar, with average CKA around 0.35, especially when comparing to TagFex proposed in this paper, with average CKA around 0.2. This can also be seen by Grad" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.905, + 0.399 + ], + "angle": 0, + "content": "CAM [12, 38] visualizations in Figure 2. We can clearly see that the models learned by DER [51] focus on similar areas and those by TagFex focus more diverse areas." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.906, + 0.537 + ], + "angle": 0, + "content": "To further encourage the diverse feature capture, a trivial solution would be simply increasing the number of rehearsal samples. But can we find more efficient approach to help the model acquire more discriminative features? To this end, we have to help the model to acquire such features that the model has not been able to extract in the classification task and take the advantage of such features for subsequent training. Figure 3 illustrates the general idea and motivation of the proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose to continually capture diverse features in a separate feature space, and then utilize it to help the incremental learning. Specifically, for the continual capture of diverse features, we have to find an approach to capture the features agnostic to the classification, which are referred to as task-agnostic features. Therefore, we apply continual self-supervised learning (CSSL) [4, 11, 16] to train a separate task-agnostic model to capture the task-agnostic features continually. In classification, the models are merely required to capture the minimal necessary features (task-specific features) for the classification within the task. Whereas, task-agnostic features contain more information that they are hard to be captured in the merely training of the classification task. Such information is vital for the new model to capture discriminative features from the new task. To make full utilization of the task-agnostic model, we further design a merge attention to adaptively unify task-agnostic features with task-specific features. Then, we also transfer the information in the unified representations back into the task-specific model. Therefore, we are able to mitigate the feature collision in CIL, improving the performance on various scenarios. The overall framework is called Task-Agnostic Guided Feature Expansion (TagFex). Moreover, extensive experiments are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10100" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "performed to show the effectiveness of the task-agnostic guided feature expansion and the superiority of TagFex against the state-of-the-art methods, even with memory-aligned protocol." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.241, + 0.182 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Class-Incremental Learning (CIL) is one of the scenarios in continual learning where the model is learned task by task with a different set of classes. During inference, the task from which each sample is sampled is not available. Many techniques and frameworks are proposed to alleviate catastrophic forgetting and improve the performance in CIL. They can be roughly categorized into several forms, such as model expansion, rehearsal memory, model distillation and regularization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.483, + 0.568 + ], + "angle": 0, + "content": "Model expansion comes from the idea of parameter isolation for each task. It expands the feature space for each task. PNN [35] proposes learning a new backbone for each new task and fixing the former in incremental learning. It also adds layer-wise connections between old and new models to reuse former features. P&C [37] suggests a progression-compression protocol. It first expands the network to learn representative representations. After that, a compression process is conducted to control the total budget. DER [51] trains a separate backbone for each task, aggregating all of the features for classification. DyTox [10] learns a separate task token for each task. BEEF [44] incorporates energy-based theory to construct a bi-compatible framework. TCIL [20] aims to tackle the task confusion, establishing a multi-level knowledge distillation to propagate knowledge learned from the old tasks to the new one." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.719 + ], + "angle": 0, + "content": "There are works introducing contrastive learning to continual learning, but from different perspectives. OCM [16] uses contrastive learning to get holistic representations in online incremental scenarios. C4IL [32] introduces contrastive learning on the learned features to alleviate interphase confusion in CIL. Most of them directly apply contrastive learning to the inference model, where the contrastive and classification loss are competing with each other over one model, differing from our proposed TagFex. We discuss other related works about CIL in Appendix 7." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Continual Self-Supervised Learning (CSSL) aims to continually learn good representations from non-stationary data. Rehearsal-based method LUMP [28] utilizes rehearsal samples to augment current task samples by mixup. Regularization-based method CaSSLe [11] encourages current model to maintain a consistency with previous state via a prediction head. \\(C^2\\) ASR [4] selects the most representative and discriminative samples by estimating the augmentation stability for rehearsal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.9 + ], + "angle": 0, + "content": "In this paper, we treat the features in such representations as task-agnostic features, since they are agnostic to the classification task, which contain the information that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "task-specific features do not contain. To the best of our knowledge, using task-agnostic features to mitigate feature collision in expansion-based CIL has rarely been explored. Our work proposes to continually capture task-agnostic features in each task, and utilizes such features to help the expanded feature acquire more diverse features, alleviating feature collision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.214, + 0.652, + 0.23 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.239, + 0.719, + 0.254 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.907, + 0.478 + ], + "angle": 0, + "content": "In CIL scenarios, we have multiple classification tasks to learn sequentially. Let \\(\\mathcal{D}_t\\) be the training dataset of the \\(t\\)th task. \\((\\pmb{x}_i^{(t)},y_i^{(t)})\\in \\mathcal{D}_t\\) is a sample. \\(\\pmb{x}_i^{(t)}\\) is the input, \\(y_{i}^{(t)}\\) is the label. Let \\(\\mathcal{C}_t = \\bigcup_i\\{y_i^{(t)}\\}\\) be the class set of task \\(t\\). In CIL, \\(\\forall t_1\\neq t_2,\\mathcal{C}_{t_1}\\cap \\mathcal{C}_{t_2} = \\emptyset\\). In each task, we only train the model on \\(\\mathcal{D}_t\\), but test on all the tasks the model has trained on, i.e., the seen tasks. For example, when the model is training on task \\(t_i\\), the seen tasks are tasks \\(t_j(j\\leq i)\\). Therefore, the class set of the seen tasks is \\(\\mathcal{V}_t = \\bigcup_t\\mathcal{C}_t\\). In rehearsal-based CIL, a small number of rehearsal samples \\(\\mathcal{M}\\) are allowed to be stored for later tasks. The number of rehearsal samples should be constant during the CIL training. The goal is to make the model get better performance on all the seen tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.491, + 0.88, + 0.507 + ], + "angle": 0, + "content": "3.2. Feature Expansion of Task-Specific Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.685 + ], + "angle": 0, + "content": "In this section we introduce DER [51] as the basis of the feature expansion for our task-specific models. As shown in the middle part of Figure 4 (Task-specific Models), a new model \\( f_{\\mathrm{ts}}^{(t)} \\) is created at the beginning of each task. Also, the previously learned models \\( f_{\\mathrm{ts}}^{(0)} \\sim f_{\\mathrm{ts}}^{(t-1)} \\) are saved and frozen for feature extraction. The classifier is also expanded and the trained parameter weights from the old classifier is inherited. To predict the label of a sample, the concatenated feature from all of the models is used by the classifier. Formally, the classification loss \\( \\mathcal{L}_{\\mathrm{cls}} \\) with cross-entropy (CE) for the sample \\( (\\boldsymbol{x}, \\boldsymbol{y}) \\) would be," + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.698, + 0.907, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l s}} (\\boldsymbol {x}, y) = \\ell_ {\\mathrm {C E}} ([ f _ {\\mathrm {t s}} ^ {(0)} (\\boldsymbol {x}), \\dots , f _ {\\mathrm {t s}} ^ {(t)} (\\boldsymbol {x}) ], y; W _ {\\mathrm {c l s}} ^ {(t)}), \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.905, + 0.763 + ], + "angle": 0, + "content": "where \\(y \\in \\mathcal{Y}_t\\) and \\(W_{\\mathrm{cls}}^{(t)}\\) is the parameter weights of the classifier in task \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Additionally, to help the new model to capture diverse features, an auxiliary classifier is trained based on the new feature. The training objective for the auxiliary classifier is to not only distinguish classes within the current task, but also distinguish the samples from the previous tasks. Consequently, a small number of rehearsal samples are saved from the previous tasks. Therefore, the auxiliary classifier outputs \\( |\\mathcal{C}_t| + 1 \\) logits, where the old classes are treated as a whole. Formally, the auxiliary loss \\( \\mathcal{L}_{\\mathrm{aux}} \\) with cross-entropy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10101" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.068, + 0.825, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.341, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 4. Overview of the proposed framework TagFex. The framework is composed of three parts: Task-agnostic Model (Section 4.1), Merge Attention (Section 4.2) and Task-specific Models (Section 3.2 and 4.3). Overall, our framework continually captures the task-agnostic feature in each task and transfers such feature back to the task-specific feature, helping the task-specific model obtain more diverse features to combat the feature collision." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.407, + 0.302, + 0.423 + ], + "angle": 0, + "content": "for the sample \\((\\pmb {x},\\pmb {y})\\) would be," + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.432, + 0.482, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a u x}} (\\boldsymbol {x}, y) = \\ell_ {\\mathrm {C E}} \\left(f _ {\\mathrm {t s}} ^ {(t)} (\\boldsymbol {x}), y; W _ {\\mathrm {a u x}} ^ {(t)}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.494 + ], + "angle": 0, + "content": "where \\(y \\in [|\\mathcal{C}_t| + 1]\\) and \\(W_{\\mathrm{aux}}^{(t)}\\) is the parameter weights of the auxiliary classifier in task \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.533 + ], + "angle": 0, + "content": "3.3. Discussions on the Feature Diversity in Task-Specific Models" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.539, + 0.483, + 0.78 + ], + "angle": 0, + "content": "In supervised classification within one task, the model is only required to capture the minimal necessary features for the classification within the task. For example, in binary classification of rabbit and whale, the model only needs to distinguish the colors of the background and the object to minimize the classification loss. Such features are useful within a specific task, we thus refer to such features as task-specific features. Therefore, the task-specific models trained by DER can only extract the task-specific features for each task. However, apart from the task-specific features, there are plenty of features not discovered by these models. Such features are not necessarily related to the classification of the task, so we refer them as task-agnostic features. If these features are available for subsequent tasks, the new task-specific model can refer to them for better feature learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "DER uses rehearsal samples and the auxiliary classifier to capture diverse features, which makes the situation better. However, since the number of rehearsal samples are limited, the imbalanced training makes it suboptimal to capture more diverse features. Therefore, to further improve the ability to capture more diverse features, we explore how to incorporate the task-agnostic feature into this task-specific features in the next section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.886, + 0.423 + ], + "angle": 0, + "content": "4. Task-Agnostic Guided Feature Expansion" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.435, + 0.906, + 0.465 + ], + "angle": 0, + "content": "4.1. Continual Acquisition of Task-Agnostic Features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Inspired by recent studies from continual self-supervised learning [4, 11, 28], not only can we extract task-agnostic features with self-supervised learning, but also continually extract new task-agnostic features while keeping the old ones. For the continual acquisition of task-agnostic features, we follow CaSSLe [11] to continually capture more features in a predictive manner. Also, we take SimCLR [3] as the base self-supervised method for CaSSLe and discuss other self-supervised methods in section 5.5. As shown in the left part of Figure 4 (Task-agnostic Model), we train a separate model to maintain the task-agnostic feature for the learned tasks. So that, the task-agnostic feature will not be affected by the classification objectives. Also, we save the trained task-agnostic model for the next task, in order to support the predictive knowledge preservation in CaSSLe. The overall loss for the task-agnostic model in the initial task and incremental tasks are," + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.754, + 0.907, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t a}} ^ {(t = 0)} = - \\operatorname {I n f o N C E} \\left(\\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} _ {i = 1} ^ {B}, \\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i} ^ {\\prime}\\right) \\right\\} _ {i = 1} ^ {B}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.777, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {t a}} ^ {(t > 0)} = - \\operatorname {I n f o N C E} \\left(\\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} _ {i = 1} ^ {B}, \\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i} ^ {\\prime}\\right) \\right\\} _ {i = 1} ^ {B}\\right) (4) \\\\ - \\operatorname {I n f o N C E} \\left(\\left\\{g \\left(f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i}\\right)\\right) \\right\\} _ {i = 1} ^ {B}, \\left\\{f _ {\\mathrm {t a}} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} _ {i = 1} ^ {B}\\right), (5) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.835, + 0.905, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {w h e r e} \\quad \\operatorname {I n f o N C E} (\\{\\boldsymbol {z} _ {i} \\} _ {i = 1} ^ {B}, \\{\\boldsymbol {z} ^ {\\prime} \\} _ {i = 1} ^ {B}) \\\\ = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\sin \\left(z _ {i} , z _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\sum_ {j = 1 , j \\neq i} ^ {B} \\exp \\left(\\sin \\left(z _ {i} , z _ {j} ^ {\\prime}\\right) / \\tau\\right)}, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10102" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "Equation 3 and 4 are contrastive InfoNCE [33] loss to maximize the lower bound of the mutual information between the sample and its representation. The last term in equation 5 aims to train a predictor \\( g(\\cdot) \\) to predict the last task-agnostic feature \\( (f_{\\mathrm{ta}}^{\\prime}(\\boldsymbol{x}_i)) \\) with the current one \\( (f_{\\mathrm{ta}}(\\boldsymbol{x}_i)) \\), where \\( f_{\\mathrm{ta}}^{\\prime}(\\cdot) \\) is the last task-agnostic model we save. \\( \\sin (\\cdot, \\cdot) \\) in Equation 6 includes another projection head where we apply cosine similarity in the projected feature space. The core idea of the predictive loss is that if such loss is minimized, the current feature at least contains as much information as the last feature contains. Therefore, the expressiveness of the task-agnostic feature is increasing task by task as the incremental learning proceeds." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.298, + 0.484, + 0.328 + ], + "angle": 0, + "content": "4.2. Adaptive Feature Aggregation by Merge Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.412 + ], + "angle": 0, + "content": "Now we have a powerful feature extractor to continually capture the task-agnostic features from all of the tasks we have seen. To put the task-agnostic model into good use, we design the merge attention to adaptively unify task-agnostic features with task-specific features in the incremental tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.534 + ], + "angle": 0, + "content": "As shown in the right part of Figure 4 (Merge Attention), the merge attention takes the feature maps of both the task-agnostic and task-specific features of a sample as inputs, denoting them as \\( F_{\\mathrm{ta}}(\\pmb{x}) \\) and \\( F_{\\mathrm{ts}}(\\pmb{x}) \\) respectively. Like the patch embedding of ViT [8], we treat each spacial feature in the feature map \\( (H \\times W \\) features) as a patch token in our attention. The feature maps are first normalized by two layer norms for training stability." + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.546, + 0.483, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {\\mathrm {t s}} = \\operatorname {L a y e r N o r m} _ {1} \\left(F _ {\\mathrm {t s}} (\\boldsymbol {x})\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.565, + 0.482, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {\\mathrm {t a}} = \\operatorname {L a y e r N o r m} _ {2} (\\operatorname {S t o p G r a d i e n t} (F _ {\\mathrm {t a}} (\\boldsymbol {x}))). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Then, in order to absorb useful features from task-agnostic features, we take \\( z_{\\mathrm{ts}} \\) as the query, making them attend to both \\( z_{\\mathrm{ts}} \\) and \\( z_{\\mathrm{ta}} \\). Furthermore, since \\( z_{\\mathrm{ts}} \\) and \\( z_{\\mathrm{ta}} \\) are in different feature spaces, we calculate their keys and values separately." + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.681, + 0.483, + 0.7 + ], + "angle": 0, + "content": "\\[\nQ = W _ {q} \\boldsymbol {z} _ {\\mathrm {t s}}, \\quad K _ {\\mathrm {t s}} = W _ {k} ^ {(\\mathrm {t s})} \\boldsymbol {z} _ {\\mathrm {t s}}, \\quad V _ {\\mathrm {t s}} = W _ {v} ^ {(\\mathrm {t s})} \\boldsymbol {z} _ {\\mathrm {t s}}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.703, + 0.482, + 0.723 + ], + "angle": 0, + "content": "\\[\nK _ {\\mathrm {t a}} = W _ {k} ^ {(\\mathrm {t a})} \\boldsymbol {z} _ {\\mathrm {t a}}, V _ {\\mathrm {t a}} = W _ {v} ^ {(\\mathrm {t a})} \\boldsymbol {z} _ {\\mathrm {t a}} (1 0)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.484, + 0.795 + ], + "angle": 0, + "content": "We then concatenate their keys for the calculation of the attention map. Finally, we get the merged feature map by multi-head scaled dot product attention [1, 42] with concatenated keys and values. For each head of the embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.806, + 0.484, + 0.845 + ], + "angle": 0, + "content": "\\[\nO ^ {(h)} = \\operatorname {S o f t m a x} \\left(\\frac {Q ^ {(h)} \\left[ K _ {\\mathrm {t s}} ^ {(h)} , K _ {\\mathrm {t a}} ^ {(h)} \\right] ^ {T}}{\\sqrt {d / h}}\\right) \\left[ V _ {\\mathrm {t s}} ^ {(h)}, V _ {\\mathrm {t a}} ^ {(h)} \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(d\\) is the embedding dimension of each patch token, \\(h\\) is the number of heads. Such extended self-attention makes the queries not only attend to task-specific features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "themselves, but also attend to task-agnostic features. Therefore, the merged feature contains not only the information learned from the current task, but also relevant information from the task-agnostic feature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.906, + 0.212 + ], + "angle": 0, + "content": "To guide the merge process, we use global average pooling to get the feature for the merge classifier. The merge classifier is trained to predict the label of the current task samples." + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.213, + 0.905, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m c l s}} = \\ell_ {\\mathrm {C E}} (\\operatorname {G A P} (O), y). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.906, + 0.31 + ], + "angle": 0, + "content": "Note that we do not propagate the gradients to the task-agnostic model. Therefore, no task-specific information about the classification would affect the task-agnostic model, keeping the task-agnostic feature uncorrelated with task-specific ones." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.318, + 0.906, + 0.335 + ], + "angle": 0, + "content": "4.3. Task-agnostic Guided Knowledge Transferring" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.906, + 0.475 + ], + "angle": 0, + "content": "With the merge attention, we have obtained more diverse features and their corresponding logits. However, we do not directly use the logits of the merge classifier for the final inference, since the distributional drift happens when the task-agnostic model is continually updated as the incremental training goes on. We have to transfer the diverse features from merged features to task-specific features. In the end, we will only use the task-specific models with the classifier for inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.54 + ], + "angle": 0, + "content": "For the knowledge transferring, we use KL divergence between the logits of the merge classifier \\((\\log (p_{\\mathrm{m}}))\\) and task-specific classifier \\((\\log (p_{\\mathrm{ts}}))\\) whose weights \\(W_{\\mathrm{cls}}^{(t)}\\) are mentioned in section 3.2 within the current task, which is," + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.547, + 0.905, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a n s}} = D _ {\\mathrm {K L}} \\left(\\operatorname {S t o p G r a d i e n t} \\left(p _ {\\mathrm {m}} ^ {(i)}\\right) \\| p _ {\\mathrm {t s}} ^ {(i)}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.576, + 0.906, + 0.607 + ], + "angle": 0, + "content": "Therefore, with the losses mentioned in section 3.2, the overall loss we apply on the task-specific model is," + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.618, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t s}} = \\mathcal {L} _ {\\mathrm {c l s}} + \\mathcal {L} _ {\\mathrm {a u x}} + \\mathcal {L} _ {\\mathrm {t r a n s}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.827 + ], + "angle": 0, + "content": "In addition, since the expansion of the task-specific models suffers from rapidly increasing number of parameters at both training and inference. To eliminate such drawback for practical use of our method, we adopt the FPGM [18] as the pruning strategy. It takes the geometric median within the kernel pairs as the similarity. Then the kernel pairs whose similarity above a threshold are treated as redundant and one of them is pruned. With this strategy, the model size could be largely reduced while not affecting the accuracy too much. Since TagFex only uses task-specific models for inference, we only prune the task-specific models. With such pruning strategy, the method is called TagFex-P." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.627, + 0.85 + ], + "angle": 0, + "content": "4.4. Summary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In summary, the proposed framework TagFex is composed of three parts: 1) the continual acquisition of the task-agnostic features, 2) the adaptive feature aggregation with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10103" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.119, + 0.067, + 0.878, + 0.082 + ], + "angle": 0, + "content": "Table 1. Performance results. Bold font represents the best results in the scenario. TagFex-P is the method with pruning strategy." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.092, + 0.875, + 0.289 + ], + "angle": 0, + "content": "
MethodsCIFAR100ImageNet100ImageNet1000
10-1050-1010-1050-10100-100
LastAvgLastAvgLastAvgLastAvgLastAvg
iCaRL [34]49.5264.6450.5660.0850.9867.1153.6962.5640.4757.55
BiC [50]50.7965.3843.8257.0442.4065.1349.9066.36--
DyTox [10]60.4373.50--67.6176.51--59.7568.14
BEEF [44]60.9871.9463.5170.7168.7877.6270.9877.2758.6767.09
DER [51]64.3575.3665.2772.6066.7177.1871.0877.7158.8366.87
TagFex68.2378.4570.3375.8770.8479.2775.5480.6461.4568.32
TagFex-P67.3478.0269.2674.2469.2178.5674.1379.8560.1467.65
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.299, + 0.295, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.417, + 0.25, + 0.429 + ], + "angle": 0, + "content": "(a) CIFAR100 10-10" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.299, + 0.497, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.417, + 0.453, + 0.429 + ], + "angle": 0, + "content": "(b) CIFAR100 50-10" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.299, + 0.7, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.417, + 0.662, + 0.429 + ], + "angle": 0, + "content": "(c) ImageNet100 10-10" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.299, + 0.903, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.417, + 0.864, + 0.429 + ], + "angle": 0, + "content": "(d) ImageNet100 50-10" + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.436, + 0.716, + 0.45 + ], + "angle": 0, + "content": "Figure 5. Accuracy curves for baselines and TagFex on various scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.465, + 0.483, + 0.571 + ], + "angle": 0, + "content": "merge attention, 3) the task-agnostic guided knowledge transferring. The task-agnostic features are acquired continually by a separate model, and such task-agnostic information is first aggregated with the task-specific features and then transferred back into the task-specific features. The overall loss of our framework during the training of each task is the weighted sum of the three parts," + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.579, + 0.482, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\mathrm {t a}} \\mathcal {L} _ {\\mathrm {t a}} + \\lambda_ {\\mathrm {m c l s}} \\mathcal {L} _ {\\mathrm {m c l s}} + \\mathcal {L} _ {\\mathrm {t s}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.647 + ], + "angle": 0, + "content": "For practical use and fair comparison, we further adopt pruning strategy to compress the task-specific models without affecting the accuracy too much." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.224, + 0.675 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.288, + 0.699 + ], + "angle": 0, + "content": "5.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Datasets and Rehearsal Memory. Following most of the image classification benchmarks in CIL [34, 50], we use CIFAR100 and ImageNet to train the model incrementally. CIFAR100 [24] has 50,000 training and 10,000 testing samples with 100 classes in total. Each sample is a tiny image in \\(32 \\times 32\\) pixels. ImageNet [6] has 1,300 training samples and 50 test samples for each class. The maximum number of rehearsal samples is 2,000 as default for CIFAR100 and ImageNet100, and 20,000 for ImageNet1000. We use herding [34] for the selection and eviction of the rehearsal samples between each task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data Split. There are two common types of splits in CIL. The small base one equally divides all of the classes in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.906, + 0.555 + ], + "angle": 0, + "content": "a dataset [34]. The large base one uses half of the classes in a dataset as the base task (task 0), and equally divides the remaining classes [19, 52]. For a dataset with 100 classes, 10-10 means 10 classes in the base task and incremental tasks are also with 10 classes each, 50-10 means 50 classes in the base task and 10 classes in the incremental tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.649 + ], + "angle": 0, + "content": "Backbones and Implementation. We follow DER [51] to use ResNet18 [17] as the backbone model on both CIFAR and ImageNet for both task-agnostic and task-specific models. The experiments are implemented based on the open-source code PyCIL [58]. We provide more implementation details in Appendix 8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.71, + 0.68 + ], + "angle": 0, + "content": "5.2. Performance Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We test our proposed TagFex and TagFex-P in various class incremental scenarios. And we report the accuracy of the test dataset at the end of the entire incremental training (Last), and the average test accuracy at the end of each task training across all of the incremental stages (Avg). We run the DER and TagFex experiments for 5 times and report their mean values. The numerical results are shown in Table 1. As we can see from the table, TagFex achieves the best results on all of the scenarios, and generally achieves \\(3\\sim 4\\%\\) percent performance gap compared to DER. The pruned TagFex-P shows minor negative impact on the performance, which effectively reduces the size of the model. As for the number of parameters, since TagFex does not include extra modules for inference, it holds the same num" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10104" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.066, + 0.387, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.283, + 0.314, + 0.297 + ], + "angle": 0, + "content": "(d) Task 4 Epoch 0" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.068, + 0.646, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.283, + 0.577, + 0.296 + ], + "angle": 0, + "content": "(e) Task 4 Epoch 80" + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.069, + 0.906, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.283, + 0.841, + 0.296 + ], + "angle": 0, + "content": "(f) Task 4 Epoch 160" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.308, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Figure 6. Evolution of the attention map averaged across the samples in the merge attention. In each sub-figure, each row represents the attention weights for a query patch attending to key patches (red pixels means high attention weights). The left part of each sub-figure is the attention weights for task-agnostic key patches \\((K_{\\mathrm{ta}})\\), and the right part is the attention weights for task-specific key patches \\((K_{\\mathrm{ts}})\\) which are separated by a dashed line in the middle. From the evolution of the attention maps, we can conclude that for each task, the high attention values are first on the task-agnostic side, and then transfer to the task-specific side, which represents task-agnostic features are useful for the task at early epochs, and such information is absorbed by the task-specific model at the end of the task. Note that the attention map is averaged across the samples, so the attention map appears more uniform." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.421, + 0.484, + 0.556 + ], + "angle": 0, + "content": "ber of parameters as DER. TagFex-P effectively reduces the averaged incremental number of parameters from 61.6M to 11.6M on CIFAR100 10-10, and from 61.6M to 14.4M on ImageNet100 10-10. Additionally, the accuracy curves for baselines and TagFex on various scenarios are shown in Figure 5, indicating that TagFex achieve a large performance gap compared to baseline methods. More comparisons for the number of parameters are shown in Table 5 in Appendix 10." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.57, + 0.274, + 0.587 + ], + "angle": 0, + "content": "5.3. Memory Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.593, + 0.483, + 0.804 + ], + "angle": 0, + "content": "In this section, we investigate the memory efficiency of our framework. TagFex saves an extra model for the next task to continually capture the task-agnostic features compared to DER [51]. For ResNet18 on ImageNet, a model occupies the same memory as around 300 rehearsal samples [59]. It is unfair for methods with less memory saving for the next task. To eliminate such unfairness, we perform experiments in the memory-aligned protocol, where we compare TagFex to DER with more rehearsal samples. The results are shown in Table 4. As we can see from the tables, TagFex is still competitive against DER with comparable number of extra rehearsal samples. This is more obvious when the memory size is larger (Memory Size 3000, 3300), since TagFex uses the rehearsal samples more effectively than DER." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.818, + 0.438, + 0.835 + ], + "angle": 0, + "content": "5.4. Visualizations on Merge Attention Maps" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To verify the efficacy of the merge attention, we investigate the evolution of the attention map in the merge attention during the training on ImageNet100. We record the attention map of some samples at the end of some check-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.572 + ], + "angle": 0, + "content": "point epochs within each task, and see how the attention map evolves during the training of the task. The results are shown in Figure 6. From the evolution of the attention maps, we can conclude that for the incremental tasks the high attention values are first on the task-agnostic side, and then transfer to the task-specific side, which represents task-agnostic features are useful for the task at early epochs, and such information is absorbed by the task-specific model at the end of the task. The visualizations of more tasks are shown in Appendix 12." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.682, + 0.604 + ], + "angle": 0, + "content": "5.5. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.869 + ], + "angle": 0, + "content": "TagFex with different strategies for continual learning of the task-agnostic model. In section 4.1, we use SimCLR [3] as the default configuration in the learning of the task-agnostic model. However, we can use different self-supervised losses on the task-agnostic model. We test TagFex with three different self-supervised methods on CIFAR100, which is shown in Table 2. The three methods are VICReg [2] which uses a mix of variance, invariance and covariance regularizations, BYOL [14] which uses a momentum encoder, and Barlow Twins [53] which considers an objective function measuring the cross-correlation matrix between the features. The results show that TagFex is compatible to different self-supervised methods and performs well with them. It is worth to note that BYOL and Barlow Twins get better results than the default configuration of TagFex. However, they are more complicated which require another momentum updated model for training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Hyperparameter analysis. TagFex has two hyperparameters \\(\\lambda_{\\mathrm{ta}}\\) and \\(\\lambda_{\\mathrm{mcls}}\\) controlling the loss from the task" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10105" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.146, + 0.089, + 0.527, + 0.103 + ], + "angle": 0, + "content": "Table 2. Results for TagFex with other self-supervised methods." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.113, + 0.56, + 0.242 + ], + "angle": 0, + "content": "
methods10-1050-10
LastAvgLastAvg
TagFex68.2378.4570.3375.87
TagFex w/ VICReg67.8578.2169.6875.44
TagFex w/ Barlow Twins68.6278.5570.2475.82
TagFex w/ BYOL69.0478.7671.1076.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.249, + 0.279, + 0.482, + 0.293 + ], + "angle": 0, + "content": "Table 3. Ablation study on CIFAR100." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.304, + 0.63, + 0.398 + ], + "angle": 0, + "content": "
Task-agnostic Model\nw/ Merge AttentionContinual Task-agnostic ModelKnowledge Transferring10-1050-10
LastAvgLastAvg
✓✓64.4575.3465.4572.58
✓✓65.8676.3267.5173.29
✓✓✓68.2378.4570.3375.87
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.483, + 0.542 + ], + "angle": 0, + "content": "agnostic model and the merge classifier respectively. We find the best combination of them through grid search on the average incremental accuracy (Avg). We perform grid search on CIFAR100 10-10 scenario. The results are shown in Figure 7. As we can see from the figure, TagFex is quite robust to these two hyperparameters. The performance is generally better when \\(\\lambda_{\\mathrm{ta}} \\geq 1\\) The best performance is achieved when their values are balanced and their values are both 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.483, + 0.816 + ], + "angle": 0, + "content": "Ablation study. To verify the effectiveness of each part of TagFex, we perform ablation study on three components of TagFex. Since training the task-agnostic model alone would not influence the task-specific models which are used for inference, and the merge attention cannot work alone, we have to bind them at first (task-agnostic model w/ merge attention). Then, we add continual learning to the task-agnostic model (continual task-agnostic model), and do not transfer the knowledge from the task-agnostic model to the task-specific model. Finally, we compare knowledge transferring with and without the continual learning of the task-agnostic model. We evaluate the configurations on CIFAR100, and compare their average incremental accuracies (Avg). The results are shown in Table 3. As we can see, without the final knowledge transferring, we cannot effectively transfer the information into the task-specific model for inference. Also, the best performance is achieved with continual learning of task-agnostic model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.21, + 0.845 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we explored the solutions for the feature collision in expansion-based class-incremental learning. We illustrates the feature collision and propose a framework" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.072, + 0.912, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.226, + 0.91, + 0.253 + ], + "angle": 0, + "content": "Figure 7. Hyperparameter sensitivity analysis on CIFAR100 10-10." + }, + { + "type": "table_caption", + "bbox": [ + 0.636, + 0.265, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Table 4. Memory-aligned average accuracy comparison on ImageNet100." + }, + { + "type": "table", + "bbox": [ + 0.648, + 0.304, + 0.899, + 0.398 + ], + "angle": 0, + "content": "
MethodsMemory SizeScenarios
10-1050-10
DER230077.5678.21
TagFex200079.2780.64
DER330078.4379.12
TagFex300080.8781.96
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.663 + ], + "angle": 0, + "content": "called TagFex to transfer diverse features into new task-specific model. In order to acquire such diverse features, we propose to continually train a separate model to capture task-agnostic features in advance for later tasks. Then, we design a merge attention to adaptively select and aggregate the features to transfer. Finally, we transfer such features back into the task-specific model. Furthermore, we not only apply pruning techniques to reduce the number of parameters, but also perform extensive experiments to show the superiority of our proposed framework. Also, TagFex has competitive memory efficiency against simply adding more rehearsal samples. From the visualization on merge attention maps, we conclude that the high attention values are transferred from the task-agnostic part to the task-specific part, indicating that the knowledge is absorbed by the task-specific model. We also perform further analysis to thoroughly evaluate our framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.907, + 0.739 + ], + "angle": 0, + "content": "Limitations: Creating a new model for a new task would dramatically increase the number of parameters, even though we could address it by pruning techniques. In the future work, we would like to seek for more memory efficient methods and techniques." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.766, + 0.682, + 0.784 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work is partially supported by National Science and Technology Major Project (2022ZD0114805), NSFC (62476123, 62376118, 62006112, 62250069), Fundamental Research Funds for the Central Universities (2024300373,14380021), CCF-Tencent Rhino-Bird Open Research Fund RAGR20240101, Collaborative Innovation Center of Novel Software Technology and Industrialization." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10106" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.16, + 0.483, + 0.214 + ], + "angle": 0, + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In International Conference on Learning Representations, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.275, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[4] Haoyang Cheng, Haitao Wen, Xiaoliang Zhang, Heqian Qiu, Lanxiao Wang, and Hongliang Li. Contrastive continuity on augmentation stability rehearsal for continual self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5707-5717, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[5] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[9] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 86-102. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[10] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9285-9295, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.885 + ], + "angle": 0, + "content": "[11] Enrico Fini, Victor G Turrisi Da Costa, Xavier Alameda-Pineda, Elisa Ricci, Karteek Alahari, and Julien Mairal. Self-supervised models are continual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9621-9630, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.888, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Jacob Gildenblat and contributors. Pytorch library for" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[13] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost Van De Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[15] Stephen Grossberg. Adaptive resonance theory: How a brain learns to consciously attend, learn, and recognize a changing world. Neural networks, 37:1-47, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.335, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[16] Yiduo Guo, Bing Liu, and Dongyan Zhao. Online continual learning through mutual information maximization. In International Conference on Machine Learning, pages 8109-8126. PMLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[18] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4340-4349, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[19] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[20] Bingchen Huang, Zhineng Chen, Peng Zhou, Jiayin Chen, and Zuxuan Wu. Resolving task confusion in dynamic expansion architectures for class incremental learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 908-916, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European conference on computer vision (ECCV), pages 709-727. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[22] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10107" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[25] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[26] Guoliang Lin, Hanlu Chu, and Hanjiang Lai. Towards better plasticity-stability trade-off in incremental learning: A simple linear connector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 89-98, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[27] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12245-12254, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[28] Divyam Madaan, Jaehong Yoon, Yuanchun Li, Yunxin Liu, and Sung Ju Hwang. Representational continuity for unsupervised continual learning. In International Conference on Learning Representations, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[29] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[30] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychol. Learn. Motiv., pages 109-165. Elsevier, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[31] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Dilan Gorur, Razvan Pascanu, and Hassan Ghasemzadeh. Linear mode connectivity in multitask and continual learning. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[32] Zixuan Ni, Haizhou Shi, Siliang Tang, Longhui Wei, Qi Tian, and Yueting Zhuang. Revisiting catastrophic forgetting in class incremental learning. arXiv preprint arXiv:2107.12308, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[34] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2001-2010, 2017. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[35] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[36] Dawid Rymarczyk, Joost Van De Weijer, Bartosz Zielinski, and Bartlomiej Twardowski. Icicle: Interpretable class incremental continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1887-1898, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[37] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International conference on machine learning, pages 4528-4537. PMLR, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[38] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.176, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[39] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16722-16731, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[40] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in Neural Information Processing Systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[41] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 99-108, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[43] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 398–414. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[44] Fu-Yun Wang, Da-Wei Zhou, Liu Liu, Han-Jia Ye, Yatao Bian, De-Chuan Zhan, and Peilin Zhao. Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion. In The Eleventh International Conference on Learning Representations, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[45] Jingya Wang, Yu Zhang, Bin Zhang, Jinxiang Xia, and Weidong Wang. Ipfa-net: Important points feature aggregating net for point cloud classification and segmentation. Chinese Journal of Electronics, 34(1):322-337, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[46] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[47] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[48] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In Computer Vision-ECCV 2022: 17th European Conference, Part XXVI, pages 631-648. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[49] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10108" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[50] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 374-382, 2019. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[51] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3014-3023, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[52] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6982-6991, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[54] Ruru Zhang, E Haihong, and Meina Song. Fscil-eaca: Few-shot class-incremental learning network based on embedding augmentation and classifier adaptation for image classification. Chinese Journal of Electronics, 33(1):139-152, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[55] Tao Zhang, Ying Fu, and Jun Zhang. Deep guided attention network for joint denoising and demosaicing in real image. Chinese Journal of Electronics, 33(1):303-312, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[56] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Preserving locality in vision transformers for class incremental learning. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1157–1162, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[57] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Multi-layer rehearsal feature augmentation for class incremental learning. In Proceedings of the 41st International Conference on Machine Learning, pages 61649-61663. PMLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[58] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, and De-Chuan Zhan. Pycil: a python toolbox for class incremental learning. SCIENCE CHINA Information Sciences, 66(9):197101–, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[59] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning. In The Eleventh International Conference on Learning Representations, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.769 + ], + "angle": 0, + "content": "[60] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9851-9873, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.77, + 0.482, + 0.825 + ], + "angle": 0, + "content": "[61] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10109" + } + ] +] \ No newline at end of file diff --git a/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_origin.pdf b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c1be65f798e40095122978ec8fd976951fcbbfcf --- /dev/null +++ b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/3797d0e9-aa1c-4a42-ad74-2a4a0b89a0f9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e3ee913057e7d7864bd3b0ce48b58e84dc391155a4671cfe335f46e6ef12861 +size 950733 diff --git a/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/full.md b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..72dace56072e0077d37ba21d1f8a1c0f8203edd2 --- /dev/null +++ b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/full.md @@ -0,0 +1,347 @@ +# Task-Agnostic Guided Feature Expansion for Class-Incremental Learning + +Bowen Zheng Da-Wei Zhou† Han-Jia Ye De-Chuan Zhan +National Key Laboratory for Novel Software Technology, Nanjing University, China +School of Artificial Intelligence, Nanjing University, China + +{zhengbw, zhoudw, yehj, zhandc}@lamda.nju.edu.cn + +# Abstract + +The ability to learn new concepts while preserve the learned knowledge is desirable for learning systems in Class-Incremental Learning (CIL). Recently, feature expansion of the model become a prevalent solution for CIL, where the old features are fixed during the training of the new task while new features are expanded for the new tasks. However, such task-specific features learned from the new task may collide with the old features, leading to misclassification between tasks. Therefore, the expanded model is often encouraged to capture diverse features from the new task, aiming to avoid such collision. However, the existing solution is largely restricted to the samples from the current task, because of the poor accessibility to previous samples. To promote the learning and transferring of diverse features across tasks, we propose a framework called Task-Agnostic Guided Feature Expansion (TagFex). Firstly, it captures task-agnostic features continually with a separate model, providing extra task-agnostic features for subsequent tasks. Secondly, to obtain useful features from the task-agnostic model for the current task, it aggregates the task-agnostic features with the task-specific feature using a merge attention. Then the aggregated feature is transferred back into the task-specific feature for inference, helping the task-specific model capture diverse features. Extensive experiments show the effectiveness and superiority of TagFex on various CIL settings. Code is available at https://github.com/bwnzheng/TagFex_CVPR2025. + +# 1. Introduction + +For real-world applications of learning systems, new concepts and knowledge increase over time. Therefore, the ability to learn new concepts while preserve the learned knowledge is desirable for the learning system. This motivates the research in Continual Learning (or Incremental Learning). Class-Incremental Learning (CIL) [29, 46, 60] + +![](images/16c3665f3a4bfd9fcad34882baa8ac9429b3425b5421c646008c8355d04b4d75.jpg) + +![](images/37240b96634c3f68686f8ce40a047f877885765a253dacaa5fc9ac3147015077.jpg) +Figure 1. Average CKA feature similarities between different expanded models of DER and TagFex by task. It shows the expanded features learned by TagFex are less correlated. +Figure 2. GradCAM visualization comparison between different expanded models by DER and TagFex. It shows that the expanded models learned by the proposed framework focus on more diverse features. + +is one of the scenarios where new concepts incrementally emerge as new classes. The main challenge for Continual Learning is catastrophic forgetting [30], where the learning system usually forgets previously learned knowledge fast and catastrophically. Solving catastrophic forgetting alone is not enough for the model to learn new knowledge. Therefore, researchers seek for balance between stability (ability to resist changes) and plasticity (ability to adapt) [15]. + +One of the promising strategies is to train a new feature extractor for each new task while keep the previously learned model intact during the training of the new task. Such methods are often referred to as expansion-based approaches [10, 20, 43, 51, 59]. Although expansion-based CIL achieve impressive performance on various scenarios + +![](images/2717197f2aecb83d630e3bf03780705b95000e7fccf46102be629461facf6f50.jpg) +(a) Feature Expansion Without Rehearsal Samples + +![](images/6b00a305db853dc8e78cec9d796d9e702917790cc01fcf3b921a08515eb706b1.jpg) +(b) Feature Expansion with More Diverse Features +Figure 3. The illustration and mitigation of feature collision in expansion-based CIL. (a) Without the rehearsal samples, it is very difficult to distinguish two types of the fruits with the same color, because the features captured by just training on the newest task are insufficient to distinguish classes between two tasks. (b) With more diverse features, the model is able to further disengage from the spurious features, reducing the feature collision. + +in CIL, it relies on the feature diversity extracted by the new model. Generally, the models trained on different tasks capture different features. However, there could be new features in the new task colliding with the old ones, leading to classification bias towards the new task. To understand this, we could consider an extreme case where there are two incremental tasks with similar spurious features, e.g., task 1 with red apples and green pears, task 2 with red berries and green grapes. It is very difficult to distinguish two types of the fruits with the same color in such incremental scenarios. As shown in Figure 3a, even if we expand the feature, due to the lack of extra samples for feature capturing, the model would rely on the learned color feature to reach low loss on the new task, and finally make random guesses between the apple and the berry with a sample of red fruits, or simply leaning towards the newest task. + +Therefore, to mitigate feature collision, expansion-based methods often introduce an auxiliary classifier and save a small number of rehearsal samples from the previous tasks to distinguish current task samples with previous samples [10, 51]. With the auxiliary classifier, the new model is encouraged to capture diverse features other than the features extracted by the old models. However, the limited number of rehearsal samples makes the training of the auxiliary classifier imbalanced and suboptimal. To illustrate this, we perform experiments on the similarities between the learned expanded features. Specifically, we use CKA [23] as the similarity metric and calculate the similarities between each expanded feature on the initial task. The average results are shown in Figure 1, and the detailed numbers are shown in Figure 9 in the Appendix 11. From the figures, we can conclude that the features extracted by the models learned by existing expansion-based methods are relatively similar, with average CKA around 0.35, especially when comparing to TagFex proposed in this paper, with average CKA around 0.2. This can also be seen by Grad + +CAM [12, 38] visualizations in Figure 2. We can clearly see that the models learned by DER [51] focus on similar areas and those by TagFex focus more diverse areas. + +To further encourage the diverse feature capture, a trivial solution would be simply increasing the number of rehearsal samples. But can we find more efficient approach to help the model acquire more discriminative features? To this end, we have to help the model to acquire such features that the model has not been able to extract in the classification task and take the advantage of such features for subsequent training. Figure 3 illustrates the general idea and motivation of the proposed framework. + +In this paper, we propose to continually capture diverse features in a separate feature space, and then utilize it to help the incremental learning. Specifically, for the continual capture of diverse features, we have to find an approach to capture the features agnostic to the classification, which are referred to as task-agnostic features. Therefore, we apply continual self-supervised learning (CSSL) [4, 11, 16] to train a separate task-agnostic model to capture the task-agnostic features continually. In classification, the models are merely required to capture the minimal necessary features (task-specific features) for the classification within the task. Whereas, task-agnostic features contain more information that they are hard to be captured in the merely training of the classification task. Such information is vital for the new model to capture discriminative features from the new task. To make full utilization of the task-agnostic model, we further design a merge attention to adaptively unify task-agnostic features with task-specific features. Then, we also transfer the information in the unified representations back into the task-specific model. Therefore, we are able to mitigate the feature collision in CIL, improving the performance on various scenarios. The overall framework is called Task-Agnostic Guided Feature Expansion (TagFex). Moreover, extensive experiments are + +performed to show the effectiveness of the task-agnostic guided feature expansion and the superiority of TagFex against the state-of-the-art methods, even with memory-aligned protocol. + +# 2. Related Works + +Class-Incremental Learning (CIL) is one of the scenarios in continual learning where the model is learned task by task with a different set of classes. During inference, the task from which each sample is sampled is not available. Many techniques and frameworks are proposed to alleviate catastrophic forgetting and improve the performance in CIL. They can be roughly categorized into several forms, such as model expansion, rehearsal memory, model distillation and regularization. + +Model expansion comes from the idea of parameter isolation for each task. It expands the feature space for each task. PNN [35] proposes learning a new backbone for each new task and fixing the former in incremental learning. It also adds layer-wise connections between old and new models to reuse former features. P&C [37] suggests a progression-compression protocol. It first expands the network to learn representative representations. After that, a compression process is conducted to control the total budget. DER [51] trains a separate backbone for each task, aggregating all of the features for classification. DyTox [10] learns a separate task token for each task. BEEF [44] incorporates energy-based theory to construct a bi-compatible framework. TCIL [20] aims to tackle the task confusion, establishing a multi-level knowledge distillation to propagate knowledge learned from the old tasks to the new one. + +There are works introducing contrastive learning to continual learning, but from different perspectives. OCM [16] uses contrastive learning to get holistic representations in online incremental scenarios. C4IL [32] introduces contrastive learning on the learned features to alleviate interphase confusion in CIL. Most of them directly apply contrastive learning to the inference model, where the contrastive and classification loss are competing with each other over one model, differing from our proposed TagFex. We discuss other related works about CIL in Appendix 7. + +Continual Self-Supervised Learning (CSSL) aims to continually learn good representations from non-stationary data. Rehearsal-based method LUMP [28] utilizes rehearsal samples to augment current task samples by mixup. Regularization-based method CaSSLe [11] encourages current model to maintain a consistency with previous state via a prediction head. $C^2$ ASR [4] selects the most representative and discriminative samples by estimating the augmentation stability for rehearsal. + +In this paper, we treat the features in such representations as task-agnostic features, since they are agnostic to the classification task, which contain the information that + +task-specific features do not contain. To the best of our knowledge, using task-agnostic features to mitigate feature collision in expansion-based CIL has rarely been explored. Our work proposes to continually capture task-agnostic features in each task, and utilizes such features to help the expanded feature acquire more diverse features, alleviating feature collision. + +# 3. Preliminaries + +# 3.1. Problem Formulation + +In CIL scenarios, we have multiple classification tasks to learn sequentially. Let $\mathcal{D}_t$ be the training dataset of the $t$ th task. $(\pmb{x}_i^{(t)},y_i^{(t)})\in \mathcal{D}_t$ is a sample. $\pmb{x}_i^{(t)}$ is the input, $y_{i}^{(t)}$ is the label. Let $\mathcal{C}_t = \bigcup_i\{y_i^{(t)}\}$ be the class set of task $t$ . In CIL, $\forall t_1\neq t_2,\mathcal{C}_{t_1}\cap \mathcal{C}_{t_2} = \emptyset$ . In each task, we only train the model on $\mathcal{D}_t$ , but test on all the tasks the model has trained on, i.e., the seen tasks. For example, when the model is training on task $t_i$ , the seen tasks are tasks $t_j(j\leq i)$ . Therefore, the class set of the seen tasks is $\mathcal{V}_t = \bigcup_t\mathcal{C}_t$ . In rehearsal-based CIL, a small number of rehearsal samples $\mathcal{M}$ are allowed to be stored for later tasks. The number of rehearsal samples should be constant during the CIL training. The goal is to make the model get better performance on all the seen tasks. + +# 3.2. Feature Expansion of Task-Specific Models + +In this section we introduce DER [51] as the basis of the feature expansion for our task-specific models. As shown in the middle part of Figure 4 (Task-specific Models), a new model $f_{\mathrm{ts}}^{(t)}$ is created at the beginning of each task. Also, the previously learned models $f_{\mathrm{ts}}^{(0)} \sim f_{\mathrm{ts}}^{(t-1)}$ are saved and frozen for feature extraction. The classifier is also expanded and the trained parameter weights from the old classifier is inherited. To predict the label of a sample, the concatenated feature from all of the models is used by the classifier. Formally, the classification loss $\mathcal{L}_{\mathrm{cls}}$ with cross-entropy (CE) for the sample $(\boldsymbol{x}, \boldsymbol{y})$ would be, + +$$ +\mathcal {L} _ {\mathrm {c l s}} (\boldsymbol {x}, y) = \ell_ {\mathrm {C E}} ([ f _ {\mathrm {t s}} ^ {(0)} (\boldsymbol {x}), \dots , f _ {\mathrm {t s}} ^ {(t)} (\boldsymbol {x}) ], y; W _ {\mathrm {c l s}} ^ {(t)}), \quad (1) +$$ + +where $y \in \mathcal{Y}_t$ and $W_{\mathrm{cls}}^{(t)}$ is the parameter weights of the classifier in task $t$ . + +Additionally, to help the new model to capture diverse features, an auxiliary classifier is trained based on the new feature. The training objective for the auxiliary classifier is to not only distinguish classes within the current task, but also distinguish the samples from the previous tasks. Consequently, a small number of rehearsal samples are saved from the previous tasks. Therefore, the auxiliary classifier outputs $|\mathcal{C}_t| + 1$ logits, where the old classes are treated as a whole. Formally, the auxiliary loss $\mathcal{L}_{\mathrm{aux}}$ with cross-entropy + +![](images/3912d895934282c6c3fcb6f20daabb9484e6bea74cad9dc75b67c194d17a6169.jpg) +Figure 4. Overview of the proposed framework TagFex. The framework is composed of three parts: Task-agnostic Model (Section 4.1), Merge Attention (Section 4.2) and Task-specific Models (Section 3.2 and 4.3). Overall, our framework continually captures the task-agnostic feature in each task and transfers such feature back to the task-specific feature, helping the task-specific model obtain more diverse features to combat the feature collision. + +for the sample $(\pmb {x},\pmb {y})$ would be, + +$$ +\mathcal {L} _ {\mathrm {a u x}} (\boldsymbol {x}, y) = \ell_ {\mathrm {C E}} \left(f _ {\mathrm {t s}} ^ {(t)} (\boldsymbol {x}), y; W _ {\mathrm {a u x}} ^ {(t)}\right), \tag {2} +$$ + +where $y \in [|\mathcal{C}_t| + 1]$ and $W_{\mathrm{aux}}^{(t)}$ is the parameter weights of the auxiliary classifier in task $t$ . + +# 3.3. Discussions on the Feature Diversity in Task-Specific Models + +In supervised classification within one task, the model is only required to capture the minimal necessary features for the classification within the task. For example, in binary classification of rabbit and whale, the model only needs to distinguish the colors of the background and the object to minimize the classification loss. Such features are useful within a specific task, we thus refer to such features as task-specific features. Therefore, the task-specific models trained by DER can only extract the task-specific features for each task. However, apart from the task-specific features, there are plenty of features not discovered by these models. Such features are not necessarily related to the classification of the task, so we refer them as task-agnostic features. If these features are available for subsequent tasks, the new task-specific model can refer to them for better feature learning. + +DER uses rehearsal samples and the auxiliary classifier to capture diverse features, which makes the situation better. However, since the number of rehearsal samples are limited, the imbalanced training makes it suboptimal to capture more diverse features. Therefore, to further improve the ability to capture more diverse features, we explore how to incorporate the task-agnostic feature into this task-specific features in the next section. + +# 4. Task-Agnostic Guided Feature Expansion + +# 4.1. Continual Acquisition of Task-Agnostic Features + +Inspired by recent studies from continual self-supervised learning [4, 11, 28], not only can we extract task-agnostic features with self-supervised learning, but also continually extract new task-agnostic features while keeping the old ones. For the continual acquisition of task-agnostic features, we follow CaSSLe [11] to continually capture more features in a predictive manner. Also, we take SimCLR [3] as the base self-supervised method for CaSSLe and discuss other self-supervised methods in section 5.5. As shown in the left part of Figure 4 (Task-agnostic Model), we train a separate model to maintain the task-agnostic feature for the learned tasks. So that, the task-agnostic feature will not be affected by the classification objectives. Also, we save the trained task-agnostic model for the next task, in order to support the predictive knowledge preservation in CaSSLe. The overall loss for the task-agnostic model in the initial task and incremental tasks are, + +$$ +\mathcal {L} _ {\mathrm {t a}} ^ {(t = 0)} = - \operatorname {I n f o N C E} \left(\left\{f _ {\mathrm {t a}} \left(\boldsymbol {x} _ {i}\right) \right\} _ {i = 1} ^ {B}, \left\{f _ {\mathrm {t a}} \left(\boldsymbol {x} _ {i} ^ {\prime}\right) \right\} _ {i = 1} ^ {B}\right), \tag {3} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {t a}} ^ {(t > 0)} = - \operatorname {I n f o N C E} \left(\left\{f _ {\mathrm {t a}} \left(\boldsymbol {x} _ {i}\right) \right\} _ {i = 1} ^ {B}, \left\{f _ {\mathrm {t a}} \left(\boldsymbol {x} _ {i} ^ {\prime}\right) \right\} _ {i = 1} ^ {B}\right) (4) \\ - \operatorname {I n f o N C E} \left(\left\{g \left(f _ {\mathrm {t a}} \left(\boldsymbol {x} _ {i}\right)\right) \right\} _ {i = 1} ^ {B}, \left\{f _ {\mathrm {t a}} ^ {\prime} \left(\boldsymbol {x} _ {i}\right) \right\} _ {i = 1} ^ {B}\right), (5) \\ \end{array} +$$ + +$$ +\begin{array}{l} \mathrm {w h e r e} \quad \operatorname {I n f o N C E} (\{\boldsymbol {z} _ {i} \} _ {i = 1} ^ {B}, \{\boldsymbol {z} ^ {\prime} \} _ {i = 1} ^ {B}) \\ = \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp \left(\sin \left(z _ {i} , z _ {i} ^ {\prime}\right) / \tau\right)}{\sum_ {j = 1 , j \neq i} ^ {B} \exp \left(\sin \left(z _ {i} , z _ {j} ^ {\prime}\right) / \tau\right)}, \tag {6} \\ \end{array} +$$ + +Equation 3 and 4 are contrastive InfoNCE [33] loss to maximize the lower bound of the mutual information between the sample and its representation. The last term in equation 5 aims to train a predictor $g(\cdot)$ to predict the last task-agnostic feature $(f_{\mathrm{ta}}^{\prime}(\boldsymbol{x}_i))$ with the current one $(f_{\mathrm{ta}}(\boldsymbol{x}_i))$ , where $f_{\mathrm{ta}}^{\prime}(\cdot)$ is the last task-agnostic model we save. $\sin (\cdot, \cdot)$ in Equation 6 includes another projection head where we apply cosine similarity in the projected feature space. The core idea of the predictive loss is that if such loss is minimized, the current feature at least contains as much information as the last feature contains. Therefore, the expressiveness of the task-agnostic feature is increasing task by task as the incremental learning proceeds. + +# 4.2. Adaptive Feature Aggregation by Merge Attention + +Now we have a powerful feature extractor to continually capture the task-agnostic features from all of the tasks we have seen. To put the task-agnostic model into good use, we design the merge attention to adaptively unify task-agnostic features with task-specific features in the incremental tasks. + +As shown in the right part of Figure 4 (Merge Attention), the merge attention takes the feature maps of both the task-agnostic and task-specific features of a sample as inputs, denoting them as $F_{\mathrm{ta}}(\pmb{x})$ and $F_{\mathrm{ts}}(\pmb{x})$ respectively. Like the patch embedding of ViT [8], we treat each spacial feature in the feature map $(H \times W$ features) as a patch token in our attention. The feature maps are first normalized by two layer norms for training stability. + +$$ +\mathbf {z} _ {\mathrm {t s}} = \operatorname {L a y e r N o r m} _ {1} \left(F _ {\mathrm {t s}} (\boldsymbol {x})\right), \tag {7} +$$ + +$$ +\boldsymbol {z} _ {\mathrm {t a}} = \operatorname {L a y e r N o r m} _ {2} (\operatorname {S t o p G r a d i e n t} (F _ {\mathrm {t a}} (\boldsymbol {x}))). \tag {8} +$$ + +Then, in order to absorb useful features from task-agnostic features, we take $z_{\mathrm{ts}}$ as the query, making them attend to both $z_{\mathrm{ts}}$ and $z_{\mathrm{ta}}$ . Furthermore, since $z_{\mathrm{ts}}$ and $z_{\mathrm{ta}}$ are in different feature spaces, we calculate their keys and values separately. + +$$ +Q = W _ {q} \boldsymbol {z} _ {\mathrm {t s}}, \quad K _ {\mathrm {t s}} = W _ {k} ^ {(\mathrm {t s})} \boldsymbol {z} _ {\mathrm {t s}}, \quad V _ {\mathrm {t s}} = W _ {v} ^ {(\mathrm {t s})} \boldsymbol {z} _ {\mathrm {t s}}, \tag {9} +$$ + +$$ +K _ {\mathrm {t a}} = W _ {k} ^ {(\mathrm {t a})} \boldsymbol {z} _ {\mathrm {t a}}, V _ {\mathrm {t a}} = W _ {v} ^ {(\mathrm {t a})} \boldsymbol {z} _ {\mathrm {t a}} (1 0) +$$ + +We then concatenate their keys for the calculation of the attention map. Finally, we get the merged feature map by multi-head scaled dot product attention [1, 42] with concatenated keys and values. For each head of the embedding: + +$$ +O ^ {(h)} = \operatorname {S o f t m a x} \left(\frac {Q ^ {(h)} \left[ K _ {\mathrm {t s}} ^ {(h)} , K _ {\mathrm {t a}} ^ {(h)} \right] ^ {T}}{\sqrt {d / h}}\right) \left[ V _ {\mathrm {t s}} ^ {(h)}, V _ {\mathrm {t a}} ^ {(h)} \right], \tag {11} +$$ + +where $d$ is the embedding dimension of each patch token, $h$ is the number of heads. Such extended self-attention makes the queries not only attend to task-specific features + +themselves, but also attend to task-agnostic features. Therefore, the merged feature contains not only the information learned from the current task, but also relevant information from the task-agnostic feature. + +To guide the merge process, we use global average pooling to get the feature for the merge classifier. The merge classifier is trained to predict the label of the current task samples. + +$$ +\mathcal {L} _ {\mathrm {m c l s}} = \ell_ {\mathrm {C E}} (\operatorname {G A P} (O), y). \tag {12} +$$ + +Note that we do not propagate the gradients to the task-agnostic model. Therefore, no task-specific information about the classification would affect the task-agnostic model, keeping the task-agnostic feature uncorrelated with task-specific ones. + +# 4.3. Task-agnostic Guided Knowledge Transferring + +With the merge attention, we have obtained more diverse features and their corresponding logits. However, we do not directly use the logits of the merge classifier for the final inference, since the distributional drift happens when the task-agnostic model is continually updated as the incremental training goes on. We have to transfer the diverse features from merged features to task-specific features. In the end, we will only use the task-specific models with the classifier for inference. + +For the knowledge transferring, we use KL divergence between the logits of the merge classifier $(\log (p_{\mathrm{m}}))$ and task-specific classifier $(\log (p_{\mathrm{ts}}))$ whose weights $W_{\mathrm{cls}}^{(t)}$ are mentioned in section 3.2 within the current task, which is, + +$$ +\mathcal {L} _ {\text {t r a n s}} = D _ {\mathrm {K L}} \left(\operatorname {S t o p G r a d i e n t} \left(p _ {\mathrm {m}} ^ {(i)}\right) \| p _ {\mathrm {t s}} ^ {(i)}\right). \tag {13} +$$ + +Therefore, with the losses mentioned in section 3.2, the overall loss we apply on the task-specific model is, + +$$ +\mathcal {L} _ {\mathrm {t s}} = \mathcal {L} _ {\mathrm {c l s}} + \mathcal {L} _ {\mathrm {a u x}} + \mathcal {L} _ {\mathrm {t r a n s}}. \tag {14} +$$ + +In addition, since the expansion of the task-specific models suffers from rapidly increasing number of parameters at both training and inference. To eliminate such drawback for practical use of our method, we adopt the FPGM [18] as the pruning strategy. It takes the geometric median within the kernel pairs as the similarity. Then the kernel pairs whose similarity above a threshold are treated as redundant and one of them is pruned. With this strategy, the model size could be largely reduced while not affecting the accuracy too much. Since TagFex only uses task-specific models for inference, we only prune the task-specific models. With such pruning strategy, the method is called TagFex-P. + +# 4.4. Summary + +In summary, the proposed framework TagFex is composed of three parts: 1) the continual acquisition of the task-agnostic features, 2) the adaptive feature aggregation with + +Table 1. Performance results. Bold font represents the best results in the scenario. TagFex-P is the method with pruning strategy. + +
MethodsCIFAR100ImageNet100ImageNet1000
10-1050-1010-1050-10100-100
LastAvgLastAvgLastAvgLastAvgLastAvg
iCaRL [34]49.5264.6450.5660.0850.9867.1153.6962.5640.4757.55
BiC [50]50.7965.3843.8257.0442.4065.1349.9066.36--
DyTox [10]60.4373.50--67.6176.51--59.7568.14
BEEF [44]60.9871.9463.5170.7168.7877.6270.9877.2758.6767.09
DER [51]64.3575.3665.2772.6066.7177.1871.0877.7158.8366.87
TagFex68.2378.4570.3375.8770.8479.2775.5480.6461.4568.32
TagFex-P67.3478.0269.2674.2469.2178.5674.1379.8560.1467.65
+ +![](images/58871b260329848775ab5c02dc98c6585b1658589a936b1979b3169bd8fd64d0.jpg) +(a) CIFAR100 10-10 + +![](images/32ef14dd48f004008bb202fb7bfdc870bad425b76ae763e1a745b2523b23f0a4.jpg) +(b) CIFAR100 50-10 + +![](images/678b2748fdebb18395d61d6ef6516bcae33144b717281dc081c886e7c2912a81.jpg) +(c) ImageNet100 10-10 +Figure 5. Accuracy curves for baselines and TagFex on various scenarios. + +![](images/920ae2b3eafead75aa601e4816a935823168827bf080659250da194e78929289.jpg) +(d) ImageNet100 50-10 + +merge attention, 3) the task-agnostic guided knowledge transferring. The task-agnostic features are acquired continually by a separate model, and such task-agnostic information is first aggregated with the task-specific features and then transferred back into the task-specific features. The overall loss of our framework during the training of each task is the weighted sum of the three parts, + +$$ +\mathcal {L} = \lambda_ {\mathrm {t a}} \mathcal {L} _ {\mathrm {t a}} + \lambda_ {\mathrm {m c l s}} \mathcal {L} _ {\mathrm {m c l s}} + \mathcal {L} _ {\mathrm {t s}}. \tag {15} +$$ + +For practical use and fair comparison, we further adopt pruning strategy to compress the task-specific models without affecting the accuracy too much. + +# 5. Experiments + +# 5.1. Experimental Setups + +Datasets and Rehearsal Memory. Following most of the image classification benchmarks in CIL [34, 50], we use CIFAR100 and ImageNet to train the model incrementally. CIFAR100 [24] has 50,000 training and 10,000 testing samples with 100 classes in total. Each sample is a tiny image in $32 \times 32$ pixels. ImageNet [6] has 1,300 training samples and 50 test samples for each class. The maximum number of rehearsal samples is 2,000 as default for CIFAR100 and ImageNet100, and 20,000 for ImageNet1000. We use herding [34] for the selection and eviction of the rehearsal samples between each task. + +Data Split. There are two common types of splits in CIL. The small base one equally divides all of the classes in + +a dataset [34]. The large base one uses half of the classes in a dataset as the base task (task 0), and equally divides the remaining classes [19, 52]. For a dataset with 100 classes, 10-10 means 10 classes in the base task and incremental tasks are also with 10 classes each, 50-10 means 50 classes in the base task and 10 classes in the incremental tasks. + +Backbones and Implementation. We follow DER [51] to use ResNet18 [17] as the backbone model on both CIFAR and ImageNet for both task-agnostic and task-specific models. The experiments are implemented based on the open-source code PyCIL [58]. We provide more implementation details in Appendix 8. + +# 5.2. Performance Results + +We test our proposed TagFex and TagFex-P in various class incremental scenarios. And we report the accuracy of the test dataset at the end of the entire incremental training (Last), and the average test accuracy at the end of each task training across all of the incremental stages (Avg). We run the DER and TagFex experiments for 5 times and report their mean values. The numerical results are shown in Table 1. As we can see from the table, TagFex achieves the best results on all of the scenarios, and generally achieves $3\sim 4\%$ percent performance gap compared to DER. The pruned TagFex-P shows minor negative impact on the performance, which effectively reduces the size of the model. As for the number of parameters, since TagFex does not include extra modules for inference, it holds the same num + +![](images/dff91846edeed4788ad71052cb87b2eec84ca3fca48599bc0d80885646e8704d.jpg) +(d) Task 4 Epoch 0 + +![](images/055721836421157a33380a50ec8954dff5b6f71b64e7c25ee706fb461435c36d.jpg) +(e) Task 4 Epoch 80 +Figure 6. Evolution of the attention map averaged across the samples in the merge attention. In each sub-figure, each row represents the attention weights for a query patch attending to key patches (red pixels means high attention weights). The left part of each sub-figure is the attention weights for task-agnostic key patches $(K_{\mathrm{ta}})$ , and the right part is the attention weights for task-specific key patches $(K_{\mathrm{ts}})$ which are separated by a dashed line in the middle. From the evolution of the attention maps, we can conclude that for each task, the high attention values are first on the task-agnostic side, and then transfer to the task-specific side, which represents task-agnostic features are useful for the task at early epochs, and such information is absorbed by the task-specific model at the end of the task. Note that the attention map is averaged across the samples, so the attention map appears more uniform. + +![](images/8498751de8ae6f0262854e76887551eef22624b5a9113268dc487052644ba222.jpg) +(f) Task 4 Epoch 160 + +ber of parameters as DER. TagFex-P effectively reduces the averaged incremental number of parameters from 61.6M to 11.6M on CIFAR100 10-10, and from 61.6M to 14.4M on ImageNet100 10-10. Additionally, the accuracy curves for baselines and TagFex on various scenarios are shown in Figure 5, indicating that TagFex achieve a large performance gap compared to baseline methods. More comparisons for the number of parameters are shown in Table 5 in Appendix 10. + +# 5.3. Memory Efficiency + +In this section, we investigate the memory efficiency of our framework. TagFex saves an extra model for the next task to continually capture the task-agnostic features compared to DER [51]. For ResNet18 on ImageNet, a model occupies the same memory as around 300 rehearsal samples [59]. It is unfair for methods with less memory saving for the next task. To eliminate such unfairness, we perform experiments in the memory-aligned protocol, where we compare TagFex to DER with more rehearsal samples. The results are shown in Table 4. As we can see from the tables, TagFex is still competitive against DER with comparable number of extra rehearsal samples. This is more obvious when the memory size is larger (Memory Size 3000, 3300), since TagFex uses the rehearsal samples more effectively than DER. + +# 5.4. Visualizations on Merge Attention Maps + +To verify the efficacy of the merge attention, we investigate the evolution of the attention map in the merge attention during the training on ImageNet100. We record the attention map of some samples at the end of some check- + +point epochs within each task, and see how the attention map evolves during the training of the task. The results are shown in Figure 6. From the evolution of the attention maps, we can conclude that for the incremental tasks the high attention values are first on the task-agnostic side, and then transfer to the task-specific side, which represents task-agnostic features are useful for the task at early epochs, and such information is absorbed by the task-specific model at the end of the task. The visualizations of more tasks are shown in Appendix 12. + +# 5.5. Further Analysis + +TagFex with different strategies for continual learning of the task-agnostic model. In section 4.1, we use SimCLR [3] as the default configuration in the learning of the task-agnostic model. However, we can use different self-supervised losses on the task-agnostic model. We test TagFex with three different self-supervised methods on CIFAR100, which is shown in Table 2. The three methods are VICReg [2] which uses a mix of variance, invariance and covariance regularizations, BYOL [14] which uses a momentum encoder, and Barlow Twins [53] which considers an objective function measuring the cross-correlation matrix between the features. The results show that TagFex is compatible to different self-supervised methods and performs well with them. It is worth to note that BYOL and Barlow Twins get better results than the default configuration of TagFex. However, they are more complicated which require another momentum updated model for training. + +Hyperparameter analysis. TagFex has two hyperparameters $\lambda_{\mathrm{ta}}$ and $\lambda_{\mathrm{mcls}}$ controlling the loss from the task + +Table 2. Results for TagFex with other self-supervised methods. + +
methods10-1050-10
LastAvgLastAvg
TagFex68.2378.4570.3375.87
TagFex w/ VICReg67.8578.2169.6875.44
TagFex w/ Barlow Twins68.6278.5570.2475.82
TagFex w/ BYOL69.0478.7671.1076.31
+ +Table 3. Ablation study on CIFAR100. + +
Task-agnostic Model +w/ Merge AttentionContinual Task-agnostic ModelKnowledge Transferring10-1050-10
LastAvgLastAvg
✓✓64.4575.3465.4572.58
✓✓65.8676.3267.5173.29
✓✓✓68.2378.4570.3375.87
+ +agnostic model and the merge classifier respectively. We find the best combination of them through grid search on the average incremental accuracy (Avg). We perform grid search on CIFAR100 10-10 scenario. The results are shown in Figure 7. As we can see from the figure, TagFex is quite robust to these two hyperparameters. The performance is generally better when $\lambda_{\mathrm{ta}} \geq 1$ The best performance is achieved when their values are balanced and their values are both 1. + +Ablation study. To verify the effectiveness of each part of TagFex, we perform ablation study on three components of TagFex. Since training the task-agnostic model alone would not influence the task-specific models which are used for inference, and the merge attention cannot work alone, we have to bind them at first (task-agnostic model w/ merge attention). Then, we add continual learning to the task-agnostic model (continual task-agnostic model), and do not transfer the knowledge from the task-agnostic model to the task-specific model. Finally, we compare knowledge transferring with and without the continual learning of the task-agnostic model. We evaluate the configurations on CIFAR100, and compare their average incremental accuracies (Avg). The results are shown in Table 3. As we can see, without the final knowledge transferring, we cannot effectively transfer the information into the task-specific model for inference. Also, the best performance is achieved with continual learning of task-agnostic model. + +# 6. Conclusion + +In this paper, we explored the solutions for the feature collision in expansion-based class-incremental learning. We illustrates the feature collision and propose a framework + +![](images/ced331c96f8add2807ffeab411590e3644dc9ff58669d24c1c703832c2668b18.jpg) +Figure 7. Hyperparameter sensitivity analysis on CIFAR100 10-10. + +Table 4. Memory-aligned average accuracy comparison on ImageNet100. + +
MethodsMemory SizeScenarios
10-1050-10
DER230077.5678.21
TagFex200079.2780.64
DER330078.4379.12
TagFex300080.8781.96
+ +called TagFex to transfer diverse features into new task-specific model. In order to acquire such diverse features, we propose to continually train a separate model to capture task-agnostic features in advance for later tasks. Then, we design a merge attention to adaptively select and aggregate the features to transfer. Finally, we transfer such features back into the task-specific model. Furthermore, we not only apply pruning techniques to reduce the number of parameters, but also perform extensive experiments to show the superiority of our proposed framework. Also, TagFex has competitive memory efficiency against simply adding more rehearsal samples. From the visualization on merge attention maps, we conclude that the high attention values are transferred from the task-agnostic part to the task-specific part, indicating that the knowledge is absorbed by the task-specific model. We also perform further analysis to thoroughly evaluate our framework. + +Limitations: Creating a new model for a new task would dramatically increase the number of parameters, even though we could address it by pruning techniques. In the future work, we would like to seek for more memory efficient methods and techniques. + +# Acknowledgements + +This work is partially supported by National Science and Technology Major Project (2022ZD0114805), NSFC (62476123, 62376118, 62006112, 62250069), Fundamental Research Funds for the Central Universities (2024300373,14380021), CCF-Tencent Rhino-Bird Open Research Fund RAGR20240101, Collaborative Innovation Center of Novel Software Technology and Industrialization. + +# References + +[1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 5 +[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In International Conference on Learning Representations, 2022. 7 +[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4, 7 +[4] Haoyang Cheng, Haitao Wen, Xiaoliang Zhang, Heqian Qiu, Lanxiao Wang, and Hongliang Li. Contrastive continuity on augmentation stability rehearsal for continual self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5707-5717, 2023. 2, 3, 4 +[5] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 1 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255. IEEE, 2009. 6 +[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 1 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5 +[9] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 86-102. Springer, 2020. 1 +[10] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9285-9295, 2022. 1, 2, 3, 6 +[11] Enrico Fini, Victor G Turrisi Da Costa, Xavier Alameda-Pineda, Elisa Ricci, Karteek Alahari, and Julien Mairal. Self-supervised models are continual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9621-9630, 2022. 2, 3, 4 +[12] Jacob Gildenblat and contributors. Pytorch library for + +cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.2 +[13] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost Van De Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 1 +[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 7 +[15] Stephen Grossberg. Adaptive resonance theory: How a brain learns to consciously attend, learn, and recognize a changing world. Neural networks, 37:1-47, 2013. 1 +[16] Yiduo Guo, Bing Liu, and Dongyan Zhao. Online continual learning through mutual information maximization. In International Conference on Machine Learning, pages 8109-8126. PMLR, 2022. 2, 3 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6 +[18] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4340-4349, 2019. 5 +[19] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 6 +[20] Bingchen Huang, Zhineng Chen, Peng Zhou, Jiayin Chen, and Zuxuan Wu. Resolving task confusion in dynamic expansion architectures for class incremental learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 908-916, 2023. 1, 3 +[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European conference on computer vision (ECCV), pages 709-727. Springer, 2022. 1 +[22] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 1 +[23] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 2 +[24] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6 + +[25] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017. 1 +[26] Guoliang Lin, Hanlu Chu, and Hanjiang Lai. Towards better plasticity-stability trade-off in incremental learning: A simple linear connector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 89-98, 2022. 1 +[27] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12245-12254, 2020. 1 +[28] Divyam Madaan, Jaehong Yoon, Yuanchun Li, Yunxin Liu, and Sung Ju Hwang. Representational continuity for unsupervised continual learning. In International Conference on Learning Representations, 2022. 3, 4 +[29] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1 +[30] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychol. Learn. Motiv., pages 109-165. Elsevier, 1989. 1 +[31] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Dilan Gorur, Razvan Pascanu, and Hassan Ghasemzadeh. Linear mode connectivity in multitask and continual learning. In International Conference on Learning Representations, 2020. 1 +[32] Zixuan Ni, Haizhou Shi, Siliang Tang, Longhui Wei, Qi Tian, and Yueting Zhuang. Revisiting catastrophic forgetting in class incremental learning. arXiv preprint arXiv:2107.12308, 2021. 3 +[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5 +[34] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2001-2010, 2017. 6, 1 +[35] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3 +[36] Dawid Rymarczyk, Joost Van De Weijer, Bartosz Zielinski, and Bartlomiej Twardowski. Icicle: Interpretable class incremental continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1887-1898, 2023. 1 +[37] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International conference on machine learning, pages 4528-4537. PMLR, 2018. 3 + +[38] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2 +[39] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16722-16731, 2022. 1 +[40] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in Neural Information Processing Systems, 30, 2017. 1 +[41] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 99-108, 2022. 1 +[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[43] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 398–414. Springer, 2022. 1 +[44] Fu-Yun Wang, Da-Wei Zhou, Liu Liu, Han-Jia Ye, Yatao Bian, De-Chuan Zhan, and Peilin Zhao. Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion. In The Eleventh International Conference on Learning Representations, 2023. 3, 6 +[45] Jingya Wang, Yu Zhang, Bin Zhang, Jinxiang Xia, and Weidong Wang. Ipfa-net: Important points feature aggregating net for point cloud classification and segmentation. Chinese Journal of Electronics, 34(1):322-337, 2025. +[46] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[47] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In Advances in Neural Information Processing Systems, 2022. 1 +[48] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In Computer Vision-ECCV 2022: 17th European Conference, Part XXVI, pages 631-648. Springer, 2022. 1 +[49] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 1 + +[50] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 374-382, 2019. 6, 1 +[51] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3014-3023, 2021. 1, 2, 3, 6, 7 +[52] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6982-6991, 2020. 6 +[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 7 +[54] Ruru Zhang, E Haihong, and Meina Song. Fscil-eaca: Few-shot class-incremental learning network based on embedding augmentation and classifier adaptation for image classification. Chinese Journal of Electronics, 33(1):139-152, 2024. +[55] Tao Zhang, Ying Fu, and Jun Zhang. Deep guided attention network for joint denoising and demosaicing in real image. Chinese Journal of Electronics, 33(1):303-312, 2024. +[56] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Preserving locality in vision transformers for class incremental learning. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1157–1162, 2023. 1 +[57] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Multi-layer rehearsal feature augmentation for class incremental learning. In Proceedings of the 41st International Conference on Machine Learning, pages 61649-61663. PMLR, 2024. 1 +[58] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, and De-Chuan Zhan. Pycil: a python toolbox for class incremental learning. SCIENCE CHINA Information Sciences, 66(9):197101–, 2023. 6 +[59] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning. In The Eleventh International Conference on Learning Representations, 2023. 1, 7 +[60] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9851-9873, 2024. 1 +[61] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 1 \ No newline at end of file diff --git a/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/images.zip b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2c9348614af19b8ca6c6f3fa8b69c0675419acfe --- /dev/null +++ b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6663a48db7e30df40d705c6c6f4d18b6030575e544783916efab0f3987b91b4 +size 594152 diff --git a/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/layout.json b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..02eb4c9b2f1a8fcc0696a313c151684899553f9a --- /dev/null +++ b/2025/Task-Agnostic Guided Feature Expansion for Class-Incremental Learning/layout.json @@ -0,0 +1,8356 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "text", + "content": "Task-Agnostic Guided Feature Expansion for Class-Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 143, + 509, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 143, + 509, + 185 + ], + "spans": [ + { + "bbox": [ + 100, + 143, + 509, + 185 + ], + "type": "text", + "content": "Bowen Zheng Da-Wei Zhou† Han-Jia Ye De-Chuan Zhan \nNational Key Laboratory for Novel Software Technology, Nanjing University, China \nSchool of Artificial Intelligence, Nanjing University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 187, + 425, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 187, + 425, + 200 + ], + "spans": [ + { + "bbox": [ + 181, + 187, + 425, + 200 + ], + "type": "text", + "content": "{zhengbw, zhoudw, yehj, zhandc}@lamda.nju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 252, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 577 + ], + "type": "text", + "content": "The ability to learn new concepts while preserve the learned knowledge is desirable for learning systems in Class-Incremental Learning (CIL). Recently, feature expansion of the model become a prevalent solution for CIL, where the old features are fixed during the training of the new task while new features are expanded for the new tasks. However, such task-specific features learned from the new task may collide with the old features, leading to misclassification between tasks. Therefore, the expanded model is often encouraged to capture diverse features from the new task, aiming to avoid such collision. However, the existing solution is largely restricted to the samples from the current task, because of the poor accessibility to previous samples. To promote the learning and transferring of diverse features across tasks, we propose a framework called Task-Agnostic Guided Feature Expansion (TagFex). Firstly, it captures task-agnostic features continually with a separate model, providing extra task-agnostic features for subsequent tasks. Secondly, to obtain useful features from the task-agnostic model for the current task, it aggregates the task-agnostic features with the task-specific feature using a merge attention. Then the aggregated feature is transferred back into the task-specific feature for inference, helping the task-specific model capture diverse features. Extensive experiments show the effectiveness and superiority of TagFex on various CIL settings. Code is available at https://github.com/bwnzheng/TagFex_CVPR2025." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 600, + 135, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 135, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 135, + 612 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": "For real-world applications of learning systems, new concepts and knowledge increase over time. Therefore, the ability to learn new concepts while preserve the learned knowledge is desirable for the learning system. This motivates the research in Continual Learning (or Incremental Learning). Class-Incremental Learning (CIL) [29, 46, 60]" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 226, + 554, + 357 + ], + "blocks": [ + { + "bbox": [ + 315, + 226, + 554, + 357 + ], + "lines": [ + { + "bbox": [ + 315, + 226, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 315, + 226, + 554, + 357 + ], + "type": "image", + "image_path": "16c3665f3a4bfd9fcad34882baa8ac9429b3425b5421c646008c8355d04b4d75.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 399, + 555, + 491 + ], + "blocks": [ + { + "bbox": [ + 313, + 364, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 397 + ], + "type": "text", + "content": "Figure 1. Average CKA feature similarities between different expanded models of DER and TagFex by task. It shows the expanded features learned by TagFex are less correlated." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 399, + 555, + 491 + ], + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 491 + ], + "type": "image", + "image_path": "37240b96634c3f68686f8ce40a047f877885765a253dacaa5fc9ac3147015077.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 498, + 555, + 541 + ], + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 541 + ], + "type": "text", + "content": "Figure 2. GradCAM visualization comparison between different expanded models by DER and TagFex. It shows that the expanded models learned by the proposed framework focus on more diverse features." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 641 + ], + "type": "text", + "content": "is one of the scenarios where new concepts incrementally emerge as new classes. The main challenge for Continual Learning is catastrophic forgetting [30], where the learning system usually forgets previously learned knowledge fast and catastrophically. Solving catastrophic forgetting alone is not enough for the model to learn new knowledge. Therefore, researchers seek for balance between stability (ability to resist changes) and plasticity (ability to adapt) [15]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "One of the promising strategies is to train a new feature extractor for each new task while keep the previously learned model intact during the training of the new task. Such methods are often referred to as expansion-based approaches [10, 20, 43, 51, 59]. Although expansion-based CIL achieve impressive performance on various scenarios" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10099" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 56, + 258, + 205 + ], + "blocks": [ + { + "bbox": [ + 107, + 56, + 258, + 205 + ], + "lines": [ + { + "bbox": [ + 107, + 56, + 258, + 205 + ], + "spans": [ + { + "bbox": [ + 107, + 56, + 258, + 205 + ], + "type": "image", + "image_path": "2717197f2aecb83d630e3bf03780705b95000e7fccf46102be629461facf6f50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 208, + 263, + 217 + ], + "lines": [ + { + "bbox": [ + 100, + 208, + 263, + 217 + ], + "spans": [ + { + "bbox": [ + 100, + 208, + 263, + 217 + ], + "type": "text", + "content": "(a) Feature Expansion Without Rehearsal Samples" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 292, + 54, + 512, + 205 + ], + "blocks": [ + { + "bbox": [ + 292, + 54, + 512, + 205 + ], + "lines": [ + { + "bbox": [ + 292, + 54, + 512, + 205 + ], + "spans": [ + { + "bbox": [ + 292, + 54, + 512, + 205 + ], + "type": "image", + "image_path": "6b00a305db853dc8e78cec9d796d9e702917790cc01fcf3b921a08515eb706b1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 207, + 485, + 216 + ], + "lines": [ + { + "bbox": [ + 321, + 207, + 485, + 216 + ], + "spans": [ + { + "bbox": [ + 321, + 207, + 485, + 216 + ], + "type": "text", + "content": "(b) Feature Expansion with More Diverse Features" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 223, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 54, + 223, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 54, + 223, + 555, + 268 + ], + "type": "text", + "content": "Figure 3. The illustration and mitigation of feature collision in expansion-based CIL. (a) Without the rehearsal samples, it is very difficult to distinguish two types of the fruits with the same color, because the features captured by just training on the newest task are insufficient to distinguish classes between two tasks. (b) With more diverse features, the model is able to further disengage from the spurious features, reducing the feature collision." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 280, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 295, + 471 + ], + "type": "text", + "content": "in CIL, it relies on the feature diversity extracted by the new model. Generally, the models trained on different tasks capture different features. However, there could be new features in the new task colliding with the old ones, leading to classification bias towards the new task. To understand this, we could consider an extreme case where there are two incremental tasks with similar spurious features, e.g., task 1 with red apples and green pears, task 2 with red berries and green grapes. It is very difficult to distinguish two types of the fruits with the same color in such incremental scenarios. As shown in Figure 3a, even if we expand the feature, due to the lack of extra samples for feature capturing, the model would rely on the learned color feature to reach low loss on the new task, and finally make random guesses between the apple and the berry with a sample of red fruits, or simply leaning towards the newest task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 295, + 713 + ], + "type": "text", + "content": "Therefore, to mitigate feature collision, expansion-based methods often introduce an auxiliary classifier and save a small number of rehearsal samples from the previous tasks to distinguish current task samples with previous samples [10, 51]. With the auxiliary classifier, the new model is encouraged to capture diverse features other than the features extracted by the old models. However, the limited number of rehearsal samples makes the training of the auxiliary classifier imbalanced and suboptimal. To illustrate this, we perform experiments on the similarities between the learned expanded features. Specifically, we use CKA [23] as the similarity metric and calculate the similarities between each expanded feature on the initial task. The average results are shown in Figure 1, and the detailed numbers are shown in Figure 9 in the Appendix 11. From the figures, we can conclude that the features extracted by the models learned by existing expansion-based methods are relatively similar, with average CKA around 0.35, especially when comparing to TagFex proposed in this paper, with average CKA around 0.2. This can also be seen by Grad" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 280, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 553, + 316 + ], + "type": "text", + "content": "CAM [12, 38] visualizations in Figure 2. We can clearly see that the models learned by DER [51] focus on similar areas and those by TagFex focus more diverse areas." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 317, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 554, + 425 + ], + "type": "text", + "content": "To further encourage the diverse feature capture, a trivial solution would be simply increasing the number of rehearsal samples. But can we find more efficient approach to help the model acquire more discriminative features? To this end, we have to help the model to acquire such features that the model has not been able to extract in the classification task and take the advantage of such features for subsequent training. Figure 3 illustrates the general idea and motivation of the proposed framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 427, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose to continually capture diverse features in a separate feature space, and then utilize it to help the incremental learning. Specifically, for the continual capture of diverse features, we have to find an approach to capture the features agnostic to the classification, which are referred to as task-agnostic features. Therefore, we apply continual self-supervised learning (CSSL) [4, 11, 16] to train a separate task-agnostic model to capture the task-agnostic features continually. In classification, the models are merely required to capture the minimal necessary features (task-specific features) for the classification within the task. Whereas, task-agnostic features contain more information that they are hard to be captured in the merely training of the classification task. Such information is vital for the new model to capture discriminative features from the new task. To make full utilization of the task-agnostic model, we further design a merge attention to adaptively unify task-agnostic features with task-specific features. Then, we also transfer the information in the unified representations back into the task-specific model. Therefore, we are able to mitigate the feature collision in CIL, improving the performance on various scenarios. The overall framework is called Task-Agnostic Guided Feature Expansion (TagFex). Moreover, extensive experiments are" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10100" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "performed to show the effectiveness of the task-agnostic guided feature expansion and the superiority of TagFex against the state-of-the-art methods, even with memory-aligned protocol." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 147, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 147, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 147, + 144 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 151, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 295, + 258 + ], + "type": "text", + "content": "Class-Incremental Learning (CIL) is one of the scenarios in continual learning where the model is learned task by task with a different set of classes. During inference, the task from which each sample is sampled is not available. Many techniques and frameworks are proposed to alleviate catastrophic forgetting and improve the performance in CIL. They can be roughly categorized into several forms, such as model expansion, rehearsal memory, model distillation and regularization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 259, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 295, + 449 + ], + "type": "text", + "content": "Model expansion comes from the idea of parameter isolation for each task. It expands the feature space for each task. PNN [35] proposes learning a new backbone for each new task and fixing the former in incremental learning. It also adds layer-wise connections between old and new models to reuse former features. P&C [37] suggests a progression-compression protocol. It first expands the network to learn representative representations. After that, a compression process is conducted to control the total budget. DER [51] trains a separate backbone for each task, aggregating all of the features for classification. DyTox [10] learns a separate task token for each task. BEEF [44] incorporates energy-based theory to construct a bi-compatible framework. TCIL [20] aims to tackle the task confusion, establishing a multi-level knowledge distillation to propagate knowledge learned from the old tasks to the new one." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "type": "text", + "content": "There are works introducing contrastive learning to continual learning, but from different perspectives. OCM [16] uses contrastive learning to get holistic representations in online incremental scenarios. C4IL [32] introduces contrastive learning on the learned features to alleviate interphase confusion in CIL. Most of them directly apply contrastive learning to the inference model, where the contrastive and classification loss are competing with each other over one model, differing from our proposed TagFex. We discuss other related works about CIL in Appendix 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": "Continual Self-Supervised Learning (CSSL) aims to continually learn good representations from non-stationary data. Rehearsal-based method LUMP [28] utilizes rehearsal samples to augment current task samples by mixup. Regularization-based method CaSSLe [11] encourages current model to maintain a consistency with previous state via a prediction head. " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "C^2" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": " ASR [4] selects the most representative and discriminative samples by estimating the augmentation stability for rehearsal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "content": "In this paper, we treat the features in such representations as task-agnostic features, since they are agnostic to the classification task, which contain the information that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "task-specific features do not contain. To the best of our knowledge, using task-agnostic features to mitigate feature collision in expansion-based CIL has rarely been explored. Our work proposes to continually capture task-agnostic features in each task, and utilizes such features to help the expanded feature acquire more diverse features, alleviating feature collision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 169, + 399, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 399, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 399, + 182 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 189, + 440, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 440, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 440, + 201 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": "In CIL scenarios, we have multiple classification tasks to learn sequentially. Let " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": " be the training dataset of the " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": "th task. " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_i^{(t)},y_i^{(t)})\\in \\mathcal{D}_t" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": " is a sample. " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{(t)}" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": " is the input, " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "y_{i}^{(t)}" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": " is the label. Let " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t = \\bigcup_i\\{y_i^{(t)}\\}" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": " be the class set of task " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": ". In CIL, " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\forall t_1\\neq t_2,\\mathcal{C}_{t_1}\\cap \\mathcal{C}_{t_2} = \\emptyset" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": ". In each task, we only train the model on " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": ", but test on all the tasks the model has trained on, i.e., the seen tasks. For example, when the model is training on task " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": ", the seen tasks are tasks " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "t_j(j\\leq i)" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": ". Therefore, the class set of the seen tasks is " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_t = \\bigcup_t\\mathcal{C}_t" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": ". In rehearsal-based CIL, a small number of rehearsal samples " + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 207, + 555, + 378 + ], + "type": "text", + "content": " are allowed to be stored for later tasks. The number of rehearsal samples should be constant during the CIL training. The goal is to make the model get better performance on all the seen tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 388, + 538, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 538, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 538, + 401 + ], + "type": "text", + "content": "3.2. Feature Expansion of Task-Specific Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "text", + "content": "In this section we introduce DER [51] as the basis of the feature expansion for our task-specific models. As shown in the middle part of Figure 4 (Task-specific Models), a new model " + }, + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ts}}^{(t)}" + }, + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "text", + "content": " is created at the beginning of each task. Also, the previously learned models " + }, + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ts}}^{(0)} \\sim f_{\\mathrm{ts}}^{(t-1)}" + }, + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "text", + "content": " are saved and frozen for feature extraction. The classifier is also expanded and the trained parameter weights from the old classifier is inherited. To predict the label of a sample, the concatenated feature from all of the models is used by the classifier. Formally, the classification loss " + }, + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "text", + "content": " with cross-entropy (CE) for the sample " + }, + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{x}, \\boldsymbol{y})" + }, + { + "bbox": [ + 313, + 407, + 555, + 542 + ], + "type": "text", + "content": " would be," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 552, + 555, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 552, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 325, + 552, + 555, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l s}} (\\boldsymbol {x}, y) = \\ell_ {\\mathrm {C E}} ([ f _ {\\mathrm {t s}} ^ {(0)} (\\boldsymbol {x}), \\dots , f _ {\\mathrm {t s}} ^ {(t)} (\\boldsymbol {x}) ], y; W _ {\\mathrm {c l s}} ^ {(t)}), \\quad (1)", + "image_path": "fc7bbbf8e8afe15c256bcccc1a9275e9dd9ad8d92c01b287b06cd5e27a0200e1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}_t" + }, + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{cls}}^{(t)}" + }, + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "text", + "content": " is the parameter weights of the classifier in task " + }, + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 580, + 553, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Additionally, to help the new model to capture diverse features, an auxiliary classifier is trained based on the new feature. The training objective for the auxiliary classifier is to not only distinguish classes within the current task, but also distinguish the samples from the previous tasks. Consequently, a small number of rehearsal samples are saved from the previous tasks. Therefore, the auxiliary classifier outputs " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}_t| + 1" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " logits, where the old classes are treated as a whole. Formally, the auxiliary loss " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " with cross-entropy" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10101" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 53, + 504, + 262 + ], + "blocks": [ + { + "bbox": [ + 107, + 53, + 504, + 262 + ], + "lines": [ + { + "bbox": [ + 107, + 53, + 504, + 262 + ], + "spans": [ + { + "bbox": [ + 107, + 53, + 504, + 262 + ], + "type": "image", + "image_path": "3912d895934282c6c3fcb6f20daabb9484e6bea74cad9dc75b67c194d17a6169.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 270, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 54, + 270, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 555, + 314 + ], + "type": "text", + "content": "Figure 4. Overview of the proposed framework TagFex. The framework is composed of three parts: Task-agnostic Model (Section 4.1), Merge Attention (Section 4.2) and Task-specific Models (Section 3.2 and 4.3). Overall, our framework continually captures the task-agnostic feature in each task and transfers such feature back to the task-specific feature, helping the task-specific model obtain more diverse features to combat the feature collision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 322, + 184, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 184, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 184, + 335 + ], + "type": "text", + "content": "for the sample " + }, + { + "bbox": [ + 55, + 322, + 184, + 335 + ], + "type": "inline_equation", + "content": "(\\pmb {x},\\pmb {y})" + }, + { + "bbox": [ + 55, + 322, + 184, + 335 + ], + "type": "text", + "content": " would be," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 342, + 294, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 342, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 99, + 342, + 294, + 358 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a u x}} (\\boldsymbol {x}, y) = \\ell_ {\\mathrm {C E}} \\left(f _ {\\mathrm {t s}} ^ {(t)} (\\boldsymbol {x}), y; W _ {\\mathrm {a u x}} ^ {(t)}\\right), \\tag {2}", + "image_path": "eeed9505f938d146b56d7f46853611816bd970c675ccd4a3a24536d5e678231c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 296, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 391 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 365, + 296, + 391 + ], + "type": "inline_equation", + "content": "y \\in [|\\mathcal{C}_t| + 1]" + }, + { + "bbox": [ + 55, + 365, + 296, + 391 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 391 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{aux}}^{(t)}" + }, + { + "bbox": [ + 55, + 365, + 296, + 391 + ], + "type": "text", + "content": " is the parameter weights of the auxiliary classifier in task " + }, + { + "bbox": [ + 55, + 365, + 296, + 391 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 365, + 296, + 391 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 296, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 422 + ], + "type": "text", + "content": "3.3. Discussions on the Feature Diversity in Task-Specific Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 617 + ], + "type": "text", + "content": "In supervised classification within one task, the model is only required to capture the minimal necessary features for the classification within the task. For example, in binary classification of rabbit and whale, the model only needs to distinguish the colors of the background and the object to minimize the classification loss. Such features are useful within a specific task, we thus refer to such features as task-specific features. Therefore, the task-specific models trained by DER can only extract the task-specific features for each task. However, apart from the task-specific features, there are plenty of features not discovered by these models. Such features are not necessarily related to the classification of the task, so we refer them as task-agnostic features. If these features are available for subsequent tasks, the new task-specific model can refer to them for better feature learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "DER uses rehearsal samples and the auxiliary classifier to capture diverse features, which makes the situation better. However, since the number of rehearsal samples are limited, the imbalanced training makes it suboptimal to capture more diverse features. Therefore, to further improve the ability to capture more diverse features, we explore how to incorporate the task-agnostic feature into this task-specific features in the next section." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 320, + 542, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 542, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 542, + 335 + ], + "type": "text", + "content": "4. Task-Agnostic Guided Feature Expansion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 344, + 554, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 554, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 554, + 368 + ], + "type": "text", + "content": "4.1. Continual Acquisition of Task-Agnostic Features" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 378, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 582 + ], + "type": "text", + "content": "Inspired by recent studies from continual self-supervised learning [4, 11, 28], not only can we extract task-agnostic features with self-supervised learning, but also continually extract new task-agnostic features while keeping the old ones. For the continual acquisition of task-agnostic features, we follow CaSSLe [11] to continually capture more features in a predictive manner. Also, we take SimCLR [3] as the base self-supervised method for CaSSLe and discuss other self-supervised methods in section 5.5. As shown in the left part of Figure 4 (Task-agnostic Model), we train a separate model to maintain the task-agnostic feature for the learned tasks. So that, the task-agnostic feature will not be affected by the classification objectives. Also, we save the trained task-agnostic model for the next task, in order to support the predictive knowledge preservation in CaSSLe. The overall loss for the task-agnostic model in the initial task and incremental tasks are," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 597, + 555, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 597, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 321, + 597, + 555, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t a}} ^ {(t = 0)} = - \\operatorname {I n f o N C E} \\left(\\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} _ {i = 1} ^ {B}, \\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i} ^ {\\prime}\\right) \\right\\} _ {i = 1} ^ {B}\\right), \\tag {3}", + "image_path": "36600010f0a0332b84e11940a432a399b975cabdbb0e9dee41da8fd002ae3d0e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 615, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 322, + 615, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {t a}} ^ {(t > 0)} = - \\operatorname {I n f o N C E} \\left(\\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} _ {i = 1} ^ {B}, \\left\\{f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i} ^ {\\prime}\\right) \\right\\} _ {i = 1} ^ {B}\\right) (4) \\\\ - \\operatorname {I n f o N C E} \\left(\\left\\{g \\left(f _ {\\mathrm {t a}} \\left(\\boldsymbol {x} _ {i}\\right)\\right) \\right\\} _ {i = 1} ^ {B}, \\left\\{f _ {\\mathrm {t a}} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right) \\right\\} _ {i = 1} ^ {B}\\right), (5) \\\\ \\end{array}", + "image_path": "58e2a977bab93c267cedbf9d6a45fcdf5c7751710b4947a2d4472cbd6a277e03.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 661, + 553, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 661, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 661, + 553, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {w h e r e} \\quad \\operatorname {I n f o N C E} (\\{\\boldsymbol {z} _ {i} \\} _ {i = 1} ^ {B}, \\{\\boldsymbol {z} ^ {\\prime} \\} _ {i = 1} ^ {B}) \\\\ = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\sin \\left(z _ {i} , z _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\sum_ {j = 1 , j \\neq i} ^ {B} \\exp \\left(\\sin \\left(z _ {i} , z _ {j} ^ {\\prime}\\right) / \\tau\\right)}, \\tag {6} \\\\ \\end{array}", + "image_path": "0f07de1fb980de445120504fd07886d08c79ac37b669579da2453900ff0050e7.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10102" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "Equation 3 and 4 are contrastive InfoNCE [33] loss to maximize the lower bound of the mutual information between the sample and its representation. The last term in equation 5 aims to train a predictor " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " to predict the last task-agnostic feature " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "(f_{\\mathrm{ta}}^{\\prime}(\\boldsymbol{x}_i))" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " with the current one " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "(f_{\\mathrm{ta}}(\\boldsymbol{x}_i))" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ta}}^{\\prime}(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " is the last task-agnostic model we save. " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\sin (\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " in Equation 6 includes another projection head where we apply cosine similarity in the projected feature space. The core idea of the predictive loss is that if such loss is minimized, the current feature at least contains as much information as the last feature contains. Therefore, the expressiveness of the task-agnostic feature is increasing task by task as the incremental learning proceeds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 296, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 259 + ], + "type": "text", + "content": "4.2. Adaptive Feature Aggregation by Merge Attention" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 326 + ], + "type": "text", + "content": "Now we have a powerful feature extractor to continually capture the task-agnostic features from all of the tasks we have seen. To put the task-agnostic model into good use, we design the merge attention to adaptively unify task-agnostic features with task-specific features in the incremental tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 326, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 422 + ], + "type": "text", + "content": "As shown in the right part of Figure 4 (Merge Attention), the merge attention takes the feature maps of both the task-agnostic and task-specific features of a sample as inputs, denoting them as " + }, + { + "bbox": [ + 55, + 326, + 296, + 422 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ta}}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 326, + 296, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 326, + 296, + 422 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ts}}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 326, + 296, + 422 + ], + "type": "text", + "content": " respectively. Like the patch embedding of ViT [8], we treat each spacial feature in the feature map " + }, + { + "bbox": [ + 55, + 326, + 296, + 422 + ], + "type": "inline_equation", + "content": "(H \\times W" + }, + { + "bbox": [ + 55, + 326, + 296, + 422 + ], + "type": "text", + "content": " features) as a patch token in our attention. The feature maps are first normalized by two layer norms for training stability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 432, + 295, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 432, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 82, + 432, + 295, + 445 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {\\mathrm {t s}} = \\operatorname {L a y e r N o r m} _ {1} \\left(F _ {\\mathrm {t s}} (\\boldsymbol {x})\\right), \\tag {7}", + "image_path": "fcac31cd937e496f88d6f39bf64feafe6817a26baa471dd03ff69fcb658594de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 447, + 294, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 447, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 82, + 447, + 294, + 460 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {\\mathrm {t a}} = \\operatorname {L a y e r N o r m} _ {2} (\\operatorname {S t o p G r a d i e n t} (F _ {\\mathrm {t a}} (\\boldsymbol {x}))). \\tag {8}", + "image_path": "1f7d8502c15f6c3ac1ffd6324a83eab2d15ce18535d3d61bf86e3bd2e699319b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": "Then, in order to absorb useful features from task-agnostic features, we take " + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{ts}}" + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": " as the query, making them attend to both " + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{ts}}" + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{ta}}" + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": ". Furthermore, since " + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{ts}}" + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{ta}}" + }, + { + "bbox": [ + 55, + 469, + 296, + 529 + ], + "type": "text", + "content": " are in different feature spaces, we calculate their keys and values separately." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 539, + 295, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 539, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 69, + 539, + 295, + 554 + ], + "type": "interline_equation", + "content": "Q = W _ {q} \\boldsymbol {z} _ {\\mathrm {t s}}, \\quad K _ {\\mathrm {t s}} = W _ {k} ^ {(\\mathrm {t s})} \\boldsymbol {z} _ {\\mathrm {t s}}, \\quad V _ {\\mathrm {t s}} = W _ {v} ^ {(\\mathrm {t s})} \\boldsymbol {z} _ {\\mathrm {t s}}, \\tag {9}", + "image_path": "73c5171b87772a78050cc692d18e316366d70c8781ae375175cda1264be8ee8c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 556, + 294, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 294, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 294, + 572 + ], + "type": "interline_equation", + "content": "K _ {\\mathrm {t a}} = W _ {k} ^ {(\\mathrm {t a})} \\boldsymbol {z} _ {\\mathrm {t a}}, V _ {\\mathrm {t a}} = W _ {v} ^ {(\\mathrm {t a})} \\boldsymbol {z} _ {\\mathrm {t a}} (1 0)", + "image_path": "358f733d2929110702272df3ed0fe37053c0d295b6b7759b7ea5baf2ef7be49d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 581, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 296, + 629 + ], + "type": "text", + "content": "We then concatenate their keys for the calculation of the attention map. Finally, we get the merged feature map by multi-head scaled dot product attention [1, 42] with concatenated keys and values. For each head of the embedding:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 60, + 638, + 296, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 638, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 60, + 638, + 296, + 669 + ], + "type": "interline_equation", + "content": "O ^ {(h)} = \\operatorname {S o f t m a x} \\left(\\frac {Q ^ {(h)} \\left[ K _ {\\mathrm {t s}} ^ {(h)} , K _ {\\mathrm {t a}} ^ {(h)} \\right] ^ {T}}{\\sqrt {d / h}}\\right) \\left[ V _ {\\mathrm {t s}} ^ {(h)}, V _ {\\mathrm {t a}} ^ {(h)} \\right], \\tag {11}", + "image_path": "966e632b396c910b2c4f94f49eff8d45077d49ba058c3b8b46a20eaa1c4ba94a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is the embedding dimension of each patch token, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is the number of heads. Such extended self-attention makes the queries not only attend to task-specific features" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "themselves, but also attend to task-agnostic features. Therefore, the merged feature contains not only the information learned from the current task, but also relevant information from the task-agnostic feature." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 554, + 167 + ], + "type": "text", + "content": "To guide the merge process, we use global average pooling to get the feature for the merge classifier. The merge classifier is trained to predict the label of the current task samples." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 377, + 168, + 553, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 168, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 377, + 168, + 553, + 180 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m c l s}} = \\ell_ {\\mathrm {C E}} (\\operatorname {G A P} (O), y). \\tag {12}", + "image_path": "3b76aecf8f9fbe1a2084d0b87f0a72a579f89ed414cdaa5a6670d7dc779e6756.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 185, + 554, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 554, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 554, + 245 + ], + "type": "text", + "content": "Note that we do not propagate the gradients to the task-agnostic model. Therefore, no task-specific information about the classification would affect the task-agnostic model, keeping the task-agnostic feature uncorrelated with task-specific ones." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 251, + 554, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 265 + ], + "type": "text", + "content": "4.3. Task-agnostic Guided Knowledge Transferring" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 269, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 554, + 376 + ], + "type": "text", + "content": "With the merge attention, we have obtained more diverse features and their corresponding logits. However, we do not directly use the logits of the merge classifier for the final inference, since the distributional drift happens when the task-agnostic model is continually updated as the incremental training goes on. We have to transfer the diverse features from merged features to task-specific features. In the end, we will only use the task-specific models with the classifier for inference." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 377, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 427 + ], + "type": "text", + "content": "For the knowledge transferring, we use KL divergence between the logits of the merge classifier " + }, + { + "bbox": [ + 313, + 377, + 554, + 427 + ], + "type": "inline_equation", + "content": "(\\log (p_{\\mathrm{m}}))" + }, + { + "bbox": [ + 313, + 377, + 554, + 427 + ], + "type": "text", + "content": " and task-specific classifier " + }, + { + "bbox": [ + 313, + 377, + 554, + 427 + ], + "type": "inline_equation", + "content": "(\\log (p_{\\mathrm{ts}}))" + }, + { + "bbox": [ + 313, + 377, + 554, + 427 + ], + "type": "text", + "content": " whose weights " + }, + { + "bbox": [ + 313, + 377, + 554, + 427 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{cls}}^{(t)}" + }, + { + "bbox": [ + 313, + 377, + 554, + 427 + ], + "type": "text", + "content": " are mentioned in section 3.2 within the current task, which is," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 433, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 433, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 347, + 433, + 553, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a n s}} = D _ {\\mathrm {K L}} \\left(\\operatorname {S t o p G r a d i e n t} \\left(p _ {\\mathrm {m}} ^ {(i)}\\right) \\| p _ {\\mathrm {t s}} ^ {(i)}\\right). \\tag {13}", + "image_path": "004dadfb4d1af0eccf2bf9c3b07eadd4768acf382871e2a8fa67fcbaae18cbda.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 456, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 480 + ], + "type": "text", + "content": "Therefore, with the losses mentioned in section 3.2, the overall loss we apply on the task-specific model is," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 375, + 489, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 489, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 375, + 489, + 553, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t s}} = \\mathcal {L} _ {\\mathrm {c l s}} + \\mathcal {L} _ {\\mathrm {a u x}} + \\mathcal {L} _ {\\mathrm {t r a n s}}. \\tag {14}", + "image_path": "c27f512d2f833e3ad74613fa2e9e486670c519fb25b2698ef3db477eb7bebbd5.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 510, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 654 + ], + "type": "text", + "content": "In addition, since the expansion of the task-specific models suffers from rapidly increasing number of parameters at both training and inference. To eliminate such drawback for practical use of our method, we adopt the FPGM [18] as the pruning strategy. It takes the geometric median within the kernel pairs as the similarity. Then the kernel pairs whose similarity above a threshold are treated as redundant and one of them is pruned. With this strategy, the model size could be largely reduced while not affecting the accuracy too much. Since TagFex only uses task-specific models for inference, we only prune the task-specific models. With such pruning strategy, the method is called TagFex-P." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 660, + 383, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 383, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 383, + 673 + ], + "type": "text", + "content": "4.4. Summary" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "In summary, the proposed framework TagFex is composed of three parts: 1) the continual acquisition of the task-agnostic features, 2) the adaptive feature aggregation with" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10103" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 72, + 535, + 228 + ], + "blocks": [ + { + "bbox": [ + 72, + 53, + 537, + 64 + ], + "lines": [ + { + "bbox": [ + 72, + 53, + 537, + 64 + ], + "spans": [ + { + "bbox": [ + 72, + 53, + 537, + 64 + ], + "type": "text", + "content": "Table 1. Performance results. Bold font represents the best results in the scenario. TagFex-P is the method with pruning strategy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 72, + 535, + 228 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 535, + 228 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 535, + 228 + ], + "type": "table", + "html": "
MethodsCIFAR100ImageNet100ImageNet1000
10-1050-1010-1050-10100-100
LastAvgLastAvgLastAvgLastAvgLastAvg
iCaRL [34]49.5264.6450.5660.0850.9867.1153.6962.5640.4757.55
BiC [50]50.7965.3843.8257.0442.4065.1349.9066.36--
DyTox [10]60.4373.50--67.6176.51--59.7568.14
BEEF [44]60.9871.9463.5170.7168.7877.6270.9877.2758.6767.09
DER [51]64.3575.3665.2772.6066.7177.1871.0877.7158.8366.87
TagFex68.2378.4570.3375.8770.8479.2775.5480.6461.4568.32
TagFex-P67.3478.0269.2674.2469.2178.5674.1379.8560.1467.65
", + "image_path": "cdf20d1e4de0f4461af905fc3f87f876de0e89a429c5a2fe87f56990dd76b58a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 236, + 180, + 328 + ], + "blocks": [ + { + "bbox": [ + 57, + 236, + 180, + 328 + ], + "lines": [ + { + "bbox": [ + 57, + 236, + 180, + 328 + ], + "spans": [ + { + "bbox": [ + 57, + 236, + 180, + 328 + ], + "type": "image", + "image_path": "58871b260329848775ab5c02dc98c6585b1658589a936b1979b3169bd8fd64d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 330, + 153, + 339 + ], + "lines": [ + { + "bbox": [ + 84, + 330, + 153, + 339 + ], + "spans": [ + { + "bbox": [ + 84, + 330, + 153, + 339 + ], + "type": "text", + "content": "(a) CIFAR100 10-10" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 236, + 304, + 328 + ], + "blocks": [ + { + "bbox": [ + 181, + 236, + 304, + 328 + ], + "lines": [ + { + "bbox": [ + 181, + 236, + 304, + 328 + ], + "spans": [ + { + "bbox": [ + 181, + 236, + 304, + 328 + ], + "type": "image", + "image_path": "32ef14dd48f004008bb202fb7bfdc870bad425b76ae763e1a745b2523b23f0a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 330, + 277, + 339 + ], + "lines": [ + { + "bbox": [ + 208, + 330, + 277, + 339 + ], + "spans": [ + { + "bbox": [ + 208, + 330, + 277, + 339 + ], + "type": "text", + "content": "(b) CIFAR100 50-10" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 236, + 428, + 328 + ], + "blocks": [ + { + "bbox": [ + 304, + 236, + 428, + 328 + ], + "lines": [ + { + "bbox": [ + 304, + 236, + 428, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 428, + 328 + ], + "type": "image", + "image_path": "678b2748fdebb18395d61d6ef6516bcae33144b717281dc081c886e7c2912a81.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 330, + 405, + 339 + ], + "lines": [ + { + "bbox": [ + 328, + 330, + 405, + 339 + ], + "spans": [ + { + "bbox": [ + 328, + 330, + 405, + 339 + ], + "type": "text", + "content": "(c) ImageNet100 10-10" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 345, + 438, + 356 + ], + "lines": [ + { + "bbox": [ + 170, + 345, + 438, + 356 + ], + "spans": [ + { + "bbox": [ + 170, + 345, + 438, + 356 + ], + "type": "text", + "content": "Figure 5. Accuracy curves for baselines and TagFex on various scenarios." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 429, + 236, + 552, + 328 + ], + "blocks": [ + { + "bbox": [ + 429, + 236, + 552, + 328 + ], + "lines": [ + { + "bbox": [ + 429, + 236, + 552, + 328 + ], + "spans": [ + { + "bbox": [ + 429, + 236, + 552, + 328 + ], + "type": "image", + "image_path": "920ae2b3eafead75aa601e4816a935823168827bf080659250da194e78929289.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 330, + 528, + 339 + ], + "lines": [ + { + "bbox": [ + 451, + 330, + 528, + 339 + ], + "spans": [ + { + "bbox": [ + 451, + 330, + 528, + 339 + ], + "type": "text", + "content": "(d) ImageNet100 50-10" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 368, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 368, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 368, + 295, + 452 + ], + "type": "text", + "content": "merge attention, 3) the task-agnostic guided knowledge transferring. The task-agnostic features are acquired continually by a separate model, and such task-agnostic information is first aggregated with the task-specific features and then transferred back into the task-specific features. The overall loss of our framework during the training of each task is the weighted sum of the three parts," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 458, + 294, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 458, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 107, + 458, + 294, + 471 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\mathrm {t a}} \\mathcal {L} _ {\\mathrm {t a}} + \\lambda_ {\\mathrm {m c l s}} \\mathcal {L} _ {\\mathrm {m c l s}} + \\mathcal {L} _ {\\mathrm {t s}}. \\tag {15}", + "image_path": "d53b6043be3f54e6f9dae5e6e2687bf370fbbef9d88e0592421c48012b936c95.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 475, + 296, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 512 + ], + "type": "text", + "content": "For practical use and fair comparison, we further adopt pruning strategy to compress the task-specific models without affecting the accuracy too much." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 540, + 176, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 176, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 176, + 553 + ], + "type": "text", + "content": "5.1. Experimental Setups" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": "Datasets and Rehearsal Memory. Following most of the image classification benchmarks in CIL [34, 50], we use CIFAR100 and ImageNet to train the model incrementally. CIFAR100 [24] has 50,000 training and 10,000 testing samples with 100 classes in total. Each sample is a tiny image in " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " pixels. ImageNet [6] has 1,300 training samples and 50 test samples for each class. The maximum number of rehearsal samples is 2,000 as default for CIFAR100 and ImageNet100, and 20,000 for ImageNet1000. We use herding [34] for the selection and eviction of the rehearsal samples between each task." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Data Split. There are two common types of splits in CIL. The small base one equally divides all of the classes in" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 368, + 554, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 439 + ], + "type": "text", + "content": "a dataset [34]. The large base one uses half of the classes in a dataset as the base task (task 0), and equally divides the remaining classes [19, 52]. For a dataset with 100 classes, 10-10 means 10 classes in the base task and incremental tasks are also with 10 classes each, 50-10 means 50 classes in the base task and 10 classes in the incremental tasks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 441, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 514 + ], + "type": "text", + "content": "Backbones and Implementation. We follow DER [51] to use ResNet18 [17] as the backbone model on both CIFAR and ImageNet for both task-agnostic and task-specific models. The experiments are implemented based on the open-source code PyCIL [58]. We provide more implementation details in Appendix 8." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 526, + 434, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 434, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 434, + 538 + ], + "type": "text", + "content": "5.2. Performance Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "We test our proposed TagFex and TagFex-P in various class incremental scenarios. And we report the accuracy of the test dataset at the end of the entire incremental training (Last), and the average test accuracy at the end of each task training across all of the incremental stages (Avg). We run the DER and TagFex experiments for 5 times and report their mean values. The numerical results are shown in Table 1. As we can see from the table, TagFex achieves the best results on all of the scenarios, and generally achieves " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "3\\sim 4\\%" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " percent performance gap compared to DER. The pruned TagFex-P shows minor negative impact on the performance, which effectively reduces the size of the model. As for the number of parameters, since TagFex does not include extra modules for inference, it holds the same num" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10104" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 52, + 236, + 221 + ], + "blocks": [ + { + "bbox": [ + 56, + 52, + 236, + 221 + ], + "lines": [ + { + "bbox": [ + 56, + 52, + 236, + 221 + ], + "spans": [ + { + "bbox": [ + 56, + 52, + 236, + 221 + ], + "type": "image", + "image_path": "dff91846edeed4788ad71052cb87b2eec84ca3fca48599bc0d80885646e8704d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 224, + 192, + 235 + ], + "lines": [ + { + "bbox": [ + 128, + 224, + 192, + 235 + ], + "spans": [ + { + "bbox": [ + 128, + 224, + 192, + 235 + ], + "type": "text", + "content": "(d) Task 4 Epoch 0" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 245, + 53, + 395, + 220 + ], + "blocks": [ + { + "bbox": [ + 245, + 53, + 395, + 220 + ], + "lines": [ + { + "bbox": [ + 245, + 53, + 395, + 220 + ], + "spans": [ + { + "bbox": [ + 245, + 53, + 395, + 220 + ], + "type": "image", + "image_path": "055721836421157a33380a50ec8954dff5b6f71b64e7c25ee706fb461435c36d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 224, + 353, + 234 + ], + "lines": [ + { + "bbox": [ + 286, + 224, + 353, + 234 + ], + "spans": [ + { + "bbox": [ + 286, + 224, + 353, + 234 + ], + "type": "text", + "content": "(e) Task 4 Epoch 80" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 243, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 54, + 243, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 555, + 321 + ], + "type": "text", + "content": "Figure 6. Evolution of the attention map averaged across the samples in the merge attention. In each sub-figure, each row represents the attention weights for a query patch attending to key patches (red pixels means high attention weights). The left part of each sub-figure is the attention weights for task-agnostic key patches " + }, + { + "bbox": [ + 54, + 243, + 555, + 321 + ], + "type": "inline_equation", + "content": "(K_{\\mathrm{ta}})" + }, + { + "bbox": [ + 54, + 243, + 555, + 321 + ], + "type": "text", + "content": ", and the right part is the attention weights for task-specific key patches " + }, + { + "bbox": [ + 54, + 243, + 555, + 321 + ], + "type": "inline_equation", + "content": "(K_{\\mathrm{ts}})" + }, + { + "bbox": [ + 54, + 243, + 555, + 321 + ], + "type": "text", + "content": " which are separated by a dashed line in the middle. From the evolution of the attention maps, we can conclude that for each task, the high attention values are first on the task-agnostic side, and then transfer to the task-specific side, which represents task-agnostic features are useful for the task at early epochs, and such information is absorbed by the task-specific model at the end of the task. Note that the attention map is averaged across the samples, so the attention map appears more uniform." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 405, + 54, + 554, + 220 + ], + "blocks": [ + { + "bbox": [ + 405, + 54, + 554, + 220 + ], + "lines": [ + { + "bbox": [ + 405, + 54, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 405, + 54, + 554, + 220 + ], + "type": "image", + "image_path": "8498751de8ae6f0262854e76887551eef22624b5a9113268dc487052644ba222.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 224, + 514, + 234 + ], + "lines": [ + { + "bbox": [ + 445, + 224, + 514, + 234 + ], + "spans": [ + { + "bbox": [ + 445, + 224, + 514, + 234 + ], + "type": "text", + "content": "(f) Task 4 Epoch 160" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 333, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 333, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 296, + 440 + ], + "type": "text", + "content": "ber of parameters as DER. TagFex-P effectively reduces the averaged incremental number of parameters from 61.6M to 11.6M on CIFAR100 10-10, and from 61.6M to 14.4M on ImageNet100 10-10. Additionally, the accuracy curves for baselines and TagFex on various scenarios are shown in Figure 5, indicating that TagFex achieve a large performance gap compared to baseline methods. More comparisons for the number of parameters are shown in Table 5 in Appendix 10." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 451, + 167, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 167, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 167, + 464 + ], + "type": "text", + "content": "5.3. Memory Efficiency" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 469, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 469, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 54, + 469, + 295, + 636 + ], + "type": "text", + "content": "In this section, we investigate the memory efficiency of our framework. TagFex saves an extra model for the next task to continually capture the task-agnostic features compared to DER [51]. For ResNet18 on ImageNet, a model occupies the same memory as around 300 rehearsal samples [59]. It is unfair for methods with less memory saving for the next task. To eliminate such unfairness, we perform experiments in the memory-aligned protocol, where we compare TagFex to DER with more rehearsal samples. The results are shown in Table 4. As we can see from the tables, TagFex is still competitive against DER with comparable number of extra rehearsal samples. This is more obvious when the memory size is larger (Memory Size 3000, 3300), since TagFex uses the rehearsal samples more effectively than DER." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 268, + 661 + ], + "type": "text", + "content": "5.4. Visualizations on Merge Attention Maps" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "To verify the efficacy of the merge attention, we investigate the evolution of the attention map in the merge attention during the training on ImageNet100. We record the attention map of some samples at the end of some check-" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 333, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 453 + ], + "type": "text", + "content": "point epochs within each task, and see how the attention map evolves during the training of the task. The results are shown in Figure 6. From the evolution of the attention maps, we can conclude that for the incremental tasks the high attention values are first on the task-agnostic side, and then transfer to the task-specific side, which represents task-agnostic features are useful for the task at early epochs, and such information is absorbed by the task-specific model at the end of the task. The visualizations of more tasks are shown in Appendix 12." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 465, + 417, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 417, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 417, + 478 + ], + "type": "text", + "content": "5.5. Further Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 484, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 688 + ], + "type": "text", + "content": "TagFex with different strategies for continual learning of the task-agnostic model. In section 4.1, we use SimCLR [3] as the default configuration in the learning of the task-agnostic model. However, we can use different self-supervised losses on the task-agnostic model. We test TagFex with three different self-supervised methods on CIFAR100, which is shown in Table 2. The three methods are VICReg [2] which uses a mix of variance, invariance and covariance regularizations, BYOL [14] which uses a momentum encoder, and Barlow Twins [53] which considers an objective function measuring the cross-correlation matrix between the features. The results show that TagFex is compatible to different self-supervised methods and performs well with them. It is worth to note that BYOL and Barlow Twins get better results than the default configuration of TagFex. However, they are more complicated which require another momentum updated model for training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Hyperparameter analysis. TagFex has two hyperparameters " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ta}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{mcls}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " controlling the loss from the task" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10105" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 89, + 342, + 191 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 322, + 81 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 322, + 81 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 322, + 81 + ], + "type": "text", + "content": "Table 2. Results for TagFex with other self-supervised methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 89, + 342, + 191 + ], + "lines": [ + { + "bbox": [ + 71, + 89, + 342, + 191 + ], + "spans": [ + { + "bbox": [ + 71, + 89, + 342, + 191 + ], + "type": "table", + "html": "
methods10-1050-10
LastAvgLastAvg
TagFex68.2378.4570.3375.87
TagFex w/ VICReg67.8578.2169.6875.44
TagFex w/ Barlow Twins68.6278.5570.2475.82
TagFex w/ BYOL69.0478.7671.1076.31
", + "image_path": "73d0f0a176e5517616e6a545274dc0ac130a6be5122c4278a0c138498d449c4f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 240, + 385, + 315 + ], + "blocks": [ + { + "bbox": [ + 152, + 220, + 294, + 232 + ], + "lines": [ + { + "bbox": [ + 152, + 220, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 152, + 220, + 294, + 232 + ], + "type": "text", + "content": "Table 3. Ablation study on CIFAR100." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 240, + 385, + 315 + ], + "lines": [ + { + "bbox": [ + 61, + 240, + 385, + 315 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 385, + 315 + ], + "type": "table", + "html": "
Task-agnostic Model\nw/ Merge AttentionContinual Task-agnostic ModelKnowledge Transferring10-1050-10
LastAvgLastAvg
✓✓64.4575.3465.4572.58
✓✓65.8676.3267.5173.29
✓✓✓68.2378.4570.3375.87
", + "image_path": "b6afd622da1bf0de65721334743ab2160ae4c752f950f6cffc7d7c4e2407b9bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 322, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 295, + 429 + ], + "type": "text", + "content": "agnostic model and the merge classifier respectively. We find the best combination of them through grid search on the average incremental accuracy (Avg). We perform grid search on CIFAR100 10-10 scenario. The results are shown in Figure 7. As we can see from the figure, TagFex is quite robust to these two hyperparameters. The performance is generally better when " + }, + { + "bbox": [ + 55, + 322, + 295, + 429 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ta}} \\geq 1" + }, + { + "bbox": [ + 55, + 322, + 295, + 429 + ], + "type": "text", + "content": " The best performance is achieved when their values are balanced and their values are both 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 430, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 646 + ], + "type": "text", + "content": "Ablation study. To verify the effectiveness of each part of TagFex, we perform ablation study on three components of TagFex. Since training the task-agnostic model alone would not influence the task-specific models which are used for inference, and the merge attention cannot work alone, we have to bind them at first (task-agnostic model w/ merge attention). Then, we add continual learning to the task-agnostic model (continual task-agnostic model), and do not transfer the knowledge from the task-agnostic model to the task-specific model. Finally, we compare knowledge transferring with and without the continual learning of the task-agnostic model. We evaluate the configurations on CIFAR100, and compare their average incremental accuracies (Avg). The results are shown in Table 3. As we can see, without the final knowledge transferring, we cannot effectively transfer the information into the task-specific model for inference. Also, the best performance is achieved with continual learning of task-agnostic model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 657, + 128, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 128, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 128, + 669 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we explored the solutions for the feature collision in expansion-based class-incremental learning. We illustrates the feature collision and propose a framework" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 354, + 57, + 558, + 177 + ], + "blocks": [ + { + "bbox": [ + 354, + 57, + 558, + 177 + ], + "lines": [ + { + "bbox": [ + 354, + 57, + 558, + 177 + ], + "spans": [ + { + "bbox": [ + 354, + 57, + 558, + 177 + ], + "type": "image", + "image_path": "ced331c96f8add2807ffeab411590e3644dc9ff58669d24c1c703832c2668b18.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 178, + 556, + 200 + ], + "lines": [ + { + "bbox": [ + 354, + 178, + 556, + 200 + ], + "spans": [ + { + "bbox": [ + 354, + 178, + 556, + 200 + ], + "type": "text", + "content": "Figure 7. Hyperparameter sensitivity analysis on CIFAR100 10-10." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 396, + 240, + 550, + 315 + ], + "blocks": [ + { + "bbox": [ + 389, + 209, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 389, + 209, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 389, + 209, + 555, + 232 + ], + "type": "text", + "content": "Table 4. Memory-aligned average accuracy comparison on ImageNet100." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 396, + 240, + 550, + 315 + ], + "lines": [ + { + "bbox": [ + 396, + 240, + 550, + 315 + ], + "spans": [ + { + "bbox": [ + 396, + 240, + 550, + 315 + ], + "type": "table", + "html": "
MethodsMemory SizeScenarios
10-1050-10
DER230077.5678.21
TagFex200079.2780.64
DER330078.4379.12
TagFex300080.8781.96
", + "image_path": "73f6dea5bc9573bf2b9d509489bb5e757bb33232eeae4c615bc8d6dc822a745c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 322, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 525 + ], + "type": "text", + "content": "called TagFex to transfer diverse features into new task-specific model. In order to acquire such diverse features, we propose to continually train a separate model to capture task-agnostic features in advance for later tasks. Then, we design a merge attention to adaptively select and aggregate the features to transfer. Finally, we transfer such features back into the task-specific model. Furthermore, we not only apply pruning techniques to reduce the number of parameters, but also perform extensive experiments to show the superiority of our proposed framework. Also, TagFex has competitive memory efficiency against simply adding more rehearsal samples. From the visualization on merge attention maps, we conclude that the high attention values are transferred from the task-agnostic part to the task-specific part, indicating that the knowledge is absorbed by the task-specific model. We also perform further analysis to thoroughly evaluate our framework." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 525, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 585 + ], + "type": "text", + "content": "Limitations: Creating a new model for a new task would dramatically increase the number of parameters, even though we could address it by pruning techniques. In the future work, we would like to seek for more memory efficient methods and techniques." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 606, + 417, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 606, + 417, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 606, + 417, + 620 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "This work is partially supported by National Science and Technology Major Project (2022ZD0114805), NSFC (62476123, 62376118, 62006112, 62250069), Fundamental Research Funds for the Central Universities (2024300373,14380021), CCF-Tencent Rhino-Bird Open Research Fund RAGR20240101, Collaborative Innovation Center of Novel Software Technology and Industrialization." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10106" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 126, + 295, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 126, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 126, + 295, + 169 + ], + "type": "text", + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In International Conference on Learning Representations, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 215 + ], + "type": "text", + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 281 + ], + "type": "text", + "content": "[4] Haoyang Cheng, Haitao Wen, Xiaoliang Zhang, Heqian Qiu, Lanxiao Wang, and Hongliang Li. Contrastive continuity on augmentation stability rehearsal for continual self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5707-5717, 2023. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 338 + ], + "type": "text", + "content": "[5] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 294, + 396 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 398, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 294, + 452 + ], + "type": "text", + "content": "[7] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without memorizing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138-5146, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 454, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 294, + 529 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 295, + 586 + ], + "type": "text", + "content": "[9] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 86-102. Springer, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 590, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 644 + ], + "type": "text", + "content": "[10] Arthur Douillard, Alexandre Rame, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9285-9295, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 700 + ], + "type": "text", + "content": "[11] Enrico Fini, Victor G Turrisi Da Costa, Xavier Alameda-Pineda, Elisa Ricci, Karteek Alahari, and Julien Mairal. Self-supervised models are continual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9621-9630, 2022. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 703, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 703, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 703, + 295, + 713 + ], + "type": "text", + "content": "[12] Jacob Gildenblat and contributors. Pytorch library for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021.2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 162 + ], + "type": "text", + "content": "[13] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost Van De Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "type": "text", + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 232, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 262 + ], + "type": "text", + "content": "[15] Stephen Grossberg. Adaptive resonance theory: How a brain learns to consciously attend, learn, and recognize a changing world. Neural networks, 37:1-47, 2013. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 553, + 308 + ], + "type": "text", + "content": "[16] Yiduo Guo, Bing Liu, and Dongyan Zhao. Online continual learning through mutual information maximization. In International Conference on Machine Learning, pages 8109-8126. PMLR, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 555, + 409 + ], + "type": "text", + "content": "[18] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4340-4349, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "type": "text", + "content": "[19] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[20] Bingchen Huang, Zhineng Chen, Peng Zhou, Jiayin Chen, and Zuxuan Wu. Resolving task confusion in dynamic expansion architectures for class incremental learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 908-916, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 523, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 577 + ], + "type": "text", + "content": "[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European conference on computer vision (ECCV), pages 709-727. Springer, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 579, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 644 + ], + "type": "text", + "content": "[22] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[23] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519-3529. PMLR, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10107" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[25] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[26] Guoliang Lin, Hanlu Chu, and Hanjiang Lai. Towards better plasticity-stability trade-off in incremental learning: A simple linear connector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 89-98, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 295, + 216 + ], + "type": "text", + "content": "[27] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 12245-12254, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[28] Divyam Madaan, Jaehong Yoon, Yuanchun Li, Yunxin Liu, and Sung Ju Hwang. Representational continuity for unsupervised continual learning. In International Conference on Learning Representations, 2022. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 325 + ], + "type": "text", + "content": "[29] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[30] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychol. Learn. Motiv., pages 109-165. Elsevier, 1989. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[31] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Dilan Gorur, Razvan Pascanu, and Hassan Ghasemzadeh. Linear mode connectivity in multitask and continual learning. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[32] Zixuan Ni, Haizhou Shi, Siliang Tang, Longhui Wei, Qi Tian, and Yueting Zhuang. Revisiting catastrophic forgetting in class incremental learning. arXiv preprint arXiv:2107.12308, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "text", + "content": "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "type": "text", + "content": "[34] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2001-2010, 2017. 6, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[35] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "text", + "content": "[36] Dawid Rymarczyk, Joost Van De Weijer, Bartosz Zielinski, and Bartlomiej Twardowski. Icicle: Interpretable class incremental continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1887-1898, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[37] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International conference on machine learning, pages 4528-4537. PMLR, 2018. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[38] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 139, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 139, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 139, + 553, + 205 + ], + "type": "text", + "content": "[39] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16722-16731, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 205, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 238 + ], + "type": "text", + "content": "[40] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in Neural Information Processing Systems, 30, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[41] Rishabh Tiwari, Krishnateja Killamsetty, Rishabh Iyer, and Pradeep Shenoy. Gcr: Gradient coreset based replay buffer selection for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 99-108, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 293, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 337 + ], + "type": "text", + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "text", + "content": "[43] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 398–414. Springer, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "text", + "content": "[44] Fu-Yun Wang, Da-Wei Zhou, Liu Liu, Han-Jia Ye, Yatao Bian, De-Chuan Zhan, and Peilin Zhao. Beef: Bi-compatible class-incremental learning via energy-based expansion and fusion. In The Eleventh International Conference on Learning Representations, 2023. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[45] Jingya Wang, Yu Zhang, Bin Zhang, Jinxiang Xia, and Weidong Wang. Ipfa-net: Important points feature aggregating net for point cloud classification and segmentation. Chinese Journal of Electronics, 34(1):322-337, 2025." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 492, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 536 + ], + "type": "text", + "content": "[46] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[47] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. In Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 646 + ], + "type": "text", + "content": "[48] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In Computer Vision-ECCV 2022: 17th European Conference, Part XXVI, pages 631-648. Springer, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[49] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 139-149, 2022. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10108" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 653 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[50] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 374-382, 2019. 6, 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[51] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3014-3023, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 249 + ], + "type": "text", + "content": "[52] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6982-6991, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 296 + ], + "type": "text", + "content": "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 340 + ], + "type": "text", + "content": "[54] Ruru Zhang, E Haihong, and Meina Song. Fscil-eaca: Few-shot class-incremental learning network based on embedding augmentation and classifier adaptation for image classification. Chinese Journal of Electronics, 33(1):139-152, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "type": "text", + "content": "[55] Tao Zhang, Ying Fu, and Jun Zhang. Deep guided attention network for joint denoising and demosaicing in real image. Chinese Journal of Electronics, 33(1):303-312, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 419 + ], + "type": "text", + "content": "[56] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Preserving locality in vision transformers for class incremental learning. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1157–1162, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 473 + ], + "type": "text", + "content": "[57] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Multi-layer rehearsal feature augmentation for class incremental learning. In Proceedings of the 41st International Conference on Machine Learning, pages 61649-61663. PMLR, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "text", + "content": "[58] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, and De-Chuan Zhan. Pycil: a python toolbox for class incremental learning. SCIENCE CHINA Information Sciences, 66(9):197101–, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 564 + ], + "type": "text", + "content": "[59] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning. In The Eleventh International Conference on Learning Representations, 2023. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "text", + "content": "[60] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9851-9873, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 609, + 294, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 294, + 653 + ], + "type": "text", + "content": "[61] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10109" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Task-Aware Clustering for Prompting Vision-Language Models/aabdfb34-0949-4921-9337-b3ad9044174e_content_list.json b/2025/Task-Aware Clustering for Prompting Vision-Language Models/aabdfb34-0949-4921-9337-b3ad9044174e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c512af0a5b3ef06f75b34fab8dd656a58f69a76d --- /dev/null +++ b/2025/Task-Aware Clustering for Prompting Vision-Language Models/aabdfb34-0949-4921-9337-b3ad9044174e_content_list.json @@ -0,0 +1,1458 @@ +[ + { + "type": "text", + "text": "Task-Aware Clustering for Prompting Vision-Language Models", + "text_level": 1, + "bbox": [ + 179, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fusheng Hao $^{1,2}$ Fengxiang He $^{3}$ Fuxiang Wu $^{1,2}$ Tichao Wang $^{1,2}$ Chengqun Song $^{1,2}$ Jun Cheng $^{1,2*}$ $^{1}$ Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n $^{2}$ The Chinese University of Hong Kong", + "bbox": [ + 102, + 178, + 898, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt learning has attracted widespread attention in adapting vision-language models to downstream tasks. Existing methods largely rely on optimization strategies to ensure the task-awareness of learnable prompts. Due to the scarcity of task-specific data, overfitting is prone to occur. The resulting prompts often do not generalize well or exhibit limited task-awareness. To address this issue, we propose a novel Task-Aware Clustering (TAC) framework for prompting vision-language models, which increases the task-awareness of learnable prompts by introducing task-aware pre-context. The key ingredients are as follows: (a) generating task-aware pre-context based on task-aware clustering that can preserve the backbone structure of a downstream task with only a few clustering centers, (b) enhancing the task-awareness of learnable prompts by enabling them to interact with task-aware pre-context via the well-pretrained encoders, and (c) preventing the visual task-aware pre-context from interfering the interaction between patch embeddings by masked attention mechanism. Extensive experiments are conducted on benchmark datasets, covering the base-to-novel, domain generalization, and cross-dataset transfer settings. Ablation studies validate the effectiveness of key ingredients. Comparative results show the superiority of our TAC over competitive counterparts. The code is available at https://github.com/FushengHao/TAC.", + "bbox": [ + 88, + 301, + 485, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 726, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale visual and textual data pairs have been collected and employed to pretrain high-capacity Vision-Language Models (VLMs) [1, 20, 37]. Due to the alignment of the visual and textual modalities, the resulting models have emerged with remarkable generalization ability on various downstream tasks such as zero-shot image classification [37] and open-vocabulary object detection [11]. Despite achieving decent performance, the performance is still", + "bbox": [ + 89, + 752, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/269335cecc0f062bd1d8e48c4a5f795fa484d9cc991b18b39b8d350631129afe.jpg", + "image_caption": [ + "Figure 1. Motivation of our TAC. We improve the task-awareness of learnable prompts from a new perspective, i.e., introducing task-aware pre-context." + ], + "image_footnote": [], + "bbox": [ + 535, + 271, + 880, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lower than that of the customized models [61]. Considering the data-hungry nature of VLMs and the scarcity of task-specific data, it is difficult to improve performance through finetuning such models. As an alternative to finetuning, prompting has become as a powerful paradigm for adapting VLMs to downstream tasks [4, 7, 21, 22, 29, 51].", + "bbox": [ + 511, + 536, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompting is a technique that adapts VLMs by increasing their understanding of downstream tasks. The current methodology is to provide VLMs with task-relevant context while keeping the model weights frozen. The context is also known as prompts. Many effective prompt templates have been engineered. For example, textual templates like \"a photo of a [class]\" [37] have been hand-crafted to provide task-relevant context for the textual encoder, and visual templates like \"a red circle around an [object]\" [2] have been discovered to direct the visual encoder's attention to the [object]. To avoid the need for large-scale trial and error in prompt engineering, context optimization [61] has been proposed, which treats learnable vectors as prompts and ensure their task-awareness by optimization. This strategy has attracted much attention in recent years. For example, maximizing the ground-truth class scores [61] is proposed to ensure the task-awareness of learnable prompts, and self-regulating constraints [23] are further introduced to guide", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: jun.cheng@siat.ac.cn", + "bbox": [ + 107, + 886, + 354, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14745", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the optimization of learnable prompts for better generalization. It is worth noting that the number of labeled images per class used for prompt learning is not sufficient in the setting of context optimization1. Due to the scarcity of task-specific data, the resulting prompts are often overfitted [23] and do not generalize well, meaning that the task-awareness they exhibit is limited.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of imposing stronger constraints [39, 45, 46], we propose to increase the task-awareness by directly introducing task-aware pre-context; see Figure 1. Considering that the class descriptions of a downstream task are given when encountering the task, the class embeddings extracted by using the well-pretrained textual encoder are good candidates for task-aware pre-context. However, one big challenge arises in this practice, i.e., the number of extra embeddings it incurs is large and equal to that of classes involved. For example, the number is 1000 for the popular dataset ImageNet [9], which is significantly larger than the typical values of the number of embeddings input to the textual and visual encoders, i.e., 77 and 197, respectively. This not only incurs large additional computational overhead, but also exceeds the maximum input limit for the textual encoder. These difficulties inspire us to keep only the backbone structure of class embeddings, which is helpful for reducing the number of embeddings incurred significantly. Motivated by these observations, we explore the prompting strategy based on clustering.", + "bbox": [ + 91, + 198, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel Task-Aware Clustering (TAC) framework for prompting VLMs. The central idea of TAC is to exploit task-aware pre-context to improve the task-awareness of learnable prompts. To generate task-aware pre-context, we conduct clustering on the class embeddings of a downstream task and perform linear transformation on the clustering centers to yield task-aware pre-context, which leads to a task-aware pre-context generalization module that can keep the backbone structure of the downstream task while being lightweight. Then, the task-aware pre-context is combined with the learnable prompts and their interaction through the well-pretrained encoders is exploited to reinforce the task-awareness of learnable prompts, which leads to a task-awareness reinforcement module. Further, a mask matrix is developed and injected into the attention mechanism to eliminate the undesired interference of visual task-aware pre-context to patch embeddings induced by the gap between textual and visual modalities, which leads to undesired interaction masking module. To suppress the overfitting induced by the scarcity of task-specific data, the prompted class and visual embeddings are encouraged to be consistent with their CLIP [37] peers, which leads to a self-regulation module. Our TAC maintains simplicity in design and improves the task-awareness of the resulting prompts, thus yielding competitive results.", + "bbox": [ + 91, + 500, + 482, + 876 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our task-aware clustering framework for prompting has the following main contributions:", + "bbox": [ + 513, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose to increase the task-awareness of learnable prompts from a new perspective, i.e., introducing task-aware pre-context.", + "- We propose to generate task-aware pre-context based on task-aware clustering that can preserve the backbone structure of a downstream task.", + "- We exploit the well-pretrained encoders to reinforce the task-awareness of learnable prompts while masking out the undesired interaction.", + "- We demonstrate the superiority of our TAC in the baseto-novel, domain generalization, and cross-dataset transfer settings." + ], + "bbox": [ + 514, + 123, + 903, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 321, + 661, + 337 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) integrate the textual and visual modalities, bridging the gap between them through alignment. Such models are usually pretrained on large-scale visual and textual data pairs [6, 20, 37]. They can be broadly classified into four categories according to their training paradigms [12]. The first category employs contrastive-based training [19, 40], which emphasizes reducing the distance between positive pairs and increasing the distance between negative pairs. The second category utilizes masking-based training [27, 49], which emphasizes the reconstruction of missing patches/words by exploiting the unmasked caption/image. The third category leverages generative-based training [54, 55], which emphasizes the generation of entire images or captions based on conditional captions or images. The fourth category employs LLM-based training [5, 63], which emphasizes the usage of large language models to extract supervisory information. Since the textual and visual modalities are aligned in a shared embedding space, VLMs exhibits a strong ability to understand open-vocabulary concepts, which facilitates various downstream tasks. Although VLMs have achieved decent performance, their accuracies are still far inferior to those of customized models [61]. Considering the data-hungry nature of VLMs and the scarcity of task-specific data, finetuning such models to improve their performance faces significant challenges. As an alternative to finetuning, prompting [37, 61] has attracted much attention in adapting VLMs. Orthogonal directions include the utilization of unlabeled images [43] and historical test data [57].", + "bbox": [ + 511, + 348, + 906, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompting adapts VLMs by providing them with contextual information or task-aware prompts while freezing their model weights, which has achieved good performance in many downstream tasks [13, 17, 24, 31, 48]. Existing methods can be broadly classified into two categories according to whether domain expertise is required. The first category relies on hand-crafted prompt templates, which emphasizes", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "The typical value of this number is 16", + "bbox": [ + 109, + 886, + 320, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14746", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the reliance on domain expertise. CLIP [37] crafts the textual template, i.e., \"a photo of a [class]\". CirclePrompt [2] engineers the visual template, i.e., \"a red circle around an [object]\". LaViP [26] grounds visual prompt with language. ArGue [45] crafts attribute-guided prompts with the help of large language models. The second category relies on prompt learning [61], which emphasizes the introduction of learnable prompts and the optimization of these prompts to ensure their task-awareness. PromptSRC [23] guides the optimization of learnable prompts with self-regulating constraints. GalLoP [28] learns global-local prompts to capture multi-scale visual information. CoPrompt [39] improves the generation ability of learnable prompts by encouraging the prediction consistency between the pretrained and trainable models. TCP [52] introduces class-aware prompts that are produced by performing textual knowledge embedding on class embeddings. In this work, we increase the task-awareness by introducing task-aware pre-context, which is different from the current practice that imposes extra constraints.", + "bbox": [ + 89, + 90, + 483, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Clustering is a technique that can group similar data points into clusters or groups, in which K-Means is a classic example. Integrating clustering may improve the capabilities of deep models in capturing backbone structures, and their combination has attracted a lot of attention. DivClust [34] introduces a diversity controlling loss to ensure the desired degree of diversity for the produced multiple clusters. LCPGAN [56] associates intra-class clusters with learnable visual concepts to capture the class-wise and intra-class information for image synthesis. ZeroShotULD [41] clusters random pixel locations with nearest neighbor searching for unsupervised landmarks discovery. SPC [18] replaces the deterministic prediction with semantic clustering for domain generalized semantic segmentation. MoDE [33] decides clusters for data experts by a two-step clustering on captions and uses fine-grained clusters to coarsely represent data experts. LTE [53] clusters multiple observed motions by establishing the mapping from trajectories to trajectory embeddings. In this work, we perform clustering to preserve the backbone structure of a downstream task represented by class embeddings.", + "bbox": [ + 91, + 400, + 483, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 733, + 183, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin by reviewing the preliminaries and then detail our Task-Aware Clustering (TAC) framework for prompting vision-language models.", + "bbox": [ + 89, + 758, + 485, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 816, + 232, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP. CLIP [37] includes a textual encoder and a visual encoder, and is pretrained on around 400 million visual and textual data pairs with the contrastive-based training strategy. With the hand-crafted prompt templates like \"a photo", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of a [class]\", it emerges strong zero-shot prediction capabilities. Although CLIP has achieved good performance in the zero-shot image recognition task, its accuracies are still lower than those of the customized models [61]. Due to the data-hungry nature of CLIP and the scarcity of task-specific data, finetuning CLIP to improve its performance on downstream tasks is particularly challenging. Considering that prompt engineering relies on domain expertise and needs large-scale trial and error, prompt learning [61] has been proposed, in which learnable vectors are treated as prompts and their task-awareness is ensured through optimization. For ease of comparison with existing prompt learning approaches, we follow the current practice and choose CLIP as the foundation model.", + "bbox": [ + 511, + 90, + 906, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt learning. CoOp [61] is a typical approach for adapting CLIP in the context of prompt learning. We take a $C$ -class image classification task as an example to show how learnable prompts are optimized in CoOp. First, a set of learnable vectors with a quantity of $M^t$ is introduced and treated as textual prompts, i.e., $\\{\\pmb{p}_m^t\\}_{m=1}^{M^t}$ . Second, these prompts are combined with the word embeddings that describe the task, which leads to a learnable version of task descriptions, i.e., $\\{\\pmb{p}_1^t,\\dots,\\pmb{p}_{M^t}^t,\\pmb{c}_i\\}_{i=1}^C$ , where $\\pmb{c}_i$ is the word embedding(s) of the $i$ -th class. Third, the learnable task descriptions are fed into the textual encoder to yield their embeddings $\\{\\pmb{w}_c^{\\mathrm{CoOp}}\\}_{c=1}^C$ , while an image $\\pmb{x}$ is fed into the visual encoder to generate its embedding $\\pmb{f}$ . Fourth, the probability of $\\pmb{x}$ that belongs to the $c$ -th class can be calculated as:", + "bbox": [ + 511, + 301, + 908, + 527 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np ^ {\\mathrm {C o O p}} (c | \\boldsymbol {x}) = \\frac {\\exp \\left(\\cos \\left(\\boldsymbol {w} _ {c} ^ {\\mathrm {C o O p}} , \\boldsymbol {f}\\right) / \\tau\\right)}{\\sum_ {i = 1} ^ {C} \\exp \\left(\\cos \\left(\\boldsymbol {w} _ {i} ^ {\\mathrm {C o O p}}, \\boldsymbol {f}\\right) / \\tau\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 532, + 906, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature and $\\cos (\\cdot ,\\cdot)$ denotes the cosine similarity. Finally, the task-awareness of learnable textual prompts is ensured by minimizing the cross-entropy loss:", + "bbox": [ + 511, + 575, + 906, + 623 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} ^ {\\mathrm {C o O p}} = - \\sum_ {c = 1} ^ {C} \\mathbb {I} (c = y) \\log p ^ {\\mathrm {C o O p}} (c | \\boldsymbol {x}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 627, + 906, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y$ is the ground-truth label of $\\pmb{x}$ and $\\mathbb{I}(\\cdot)$ denotes the indicator function. It is worth noting that (a) only the learnable prompts need to be optimized and (b) the quantity of labeled images that can be used in each class is insufficient in the setting of prompt learning, where the typical value is 16.", + "bbox": [ + 511, + 674, + 908, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Afterwards, additional constraints [23, 32, 39, 59] have been imposed to improve the task-awareness of learnable prompts. Different from the current practice, we increase the task-awareness from a new perspective, i.e., introducing task-aware pre-context.", + "bbox": [ + 511, + 765, + 906, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task-aware Clustering for Prompting", + "text_level": 1, + "bbox": [ + 511, + 848, + 839, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline of our TAC is shown in Figure 2 and the key ingredients are detailed below.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14747", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9fccb3d167d80bb3593a66d3b54553a5721cf609eb020d9dd6c18d3f5edae6e8.jpg", + "image_caption": [ + "Figure 2. Pipeline of our TAC. First, the textual and visual task-aware pre-contexts are generated based on the backbone structure of a downstream task. Second, the task-aware pre-context is combined with the learnable prompts and their interaction through the well-pretrained encoders is exploited to reinforce the task-awareness of learnable prompts. Third, the undesired interaction is masked out with a customized mask matrix. Further, the textual and visual consistencies are imposed to alleviate the overfitting issue." + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 890, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-aware pre-context generation. The class embeddings extracted by using the textual encoder of CLIP are task-aware and good candidates for the desired pre-context. Considering that the number of such embeddings is equal to that of classes involved and this number is large, applying this strategy to practice is challenging. The first challenge is that the number of extra input embeddings incurred exceeds the maximum input limit for the textual encoder, which would make the textual encoder unusable. The second challenge is that the extra computational overhead incurred is very high due to the large number of additional embeddings input into the visual encoder.", + "bbox": [ + 88, + 583, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We address these challenges by yielding textual and visual task-aware pre-contexts based on clustering that can preserve the backbone structure of class embeddings. A significant additional benefit of this practice is that the number of extra embeddings incurred can be significantly reduced. Specifically, let $\\{\\boldsymbol{w}_c^{\\mathrm{CLIP}}\\}_{c=1}^C$ denote the class embeddings extracted with the textual encoder of CLIP. Then, K-Means is conducted on $\\{\\boldsymbol{w}_c^{\\mathrm{CLIP}}\\}_{c=1}^C$ , which yields a set of clustering centers with a quantity of $K$ , i.e., $\\{\\boldsymbol{\\mu}_i\\}_{i=1}^K$ . Further, the", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "textual and visual task-aware pre-contexts are generated as:", + "bbox": [ + 511, + 583, + 906, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {i} ^ {t} = \\boldsymbol {W} ^ {t} \\boldsymbol {\\mu} _ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 609, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {i} ^ {v} = \\boldsymbol {W} ^ {v} \\boldsymbol {\\mu} _ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 638, + 906, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\boldsymbol{W}^t$ and $\\boldsymbol{W}^v$ are the textual and visual linear transformations. It is worth noting that (a) the dimensions of $\\boldsymbol{\\mu}_i^t$ and $\\boldsymbol{\\mu}_i^v$ are set to be the internal dimensions of the textual and visual encoders, correspondingly, and (b) since only the weights of linear transformations needs to be optimized, our task-aware pre-context generation module is lightweight.", + "bbox": [ + 511, + 660, + 906, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-awareness reinforcement. After obtaining the textual and visual task-aware pre-contexts, the task-awareness of learnable prompts is reinforced by enabling them to interact with the corresponding pre-contexts via the well-pretrained textual and visual encoders. Specifically, let $L$ denote the quantity of transformer blocks contained in the visual/textual encoder. We construct the input embeddings to the $l$ -th transformer block of the textual encoder as:", + "bbox": [ + 511, + 752, + 908, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {t} _ {l} ^ {s o s}, \\boldsymbol {\\mu} _ {1} ^ {t}, \\dots , \\boldsymbol {\\mu} _ {K} ^ {t}, \\boldsymbol {p} _ {l 1} ^ {t}, \\dots , \\boldsymbol {p} _ {l M ^ {t}} ^ {t}, \\boldsymbol {c} _ {l i}, \\boldsymbol {t} _ {l} ^ {e o s} \\right\\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 883, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14748", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $l = 1,\\dots ,L$ , $\\pmb{t}_l^{sos}$ is the start token input to the $l$ -th transformer block, $\\pmb{t}_l^{eos}$ is the end token input to the $l$ -th transformer block, $\\{\\pmb{p}_{li}^t\\}_{i=1}^{M^t}$ is the set of learnable textual prompts input to the $l$ -th transformer block with a quantity of $M^t$ , and $\\pmb{c}_{li}$ is the word embedding(s) input to the $l$ -th transformer block. It is worth noting that (a) $\\pmb{t}_l^{sos}$ , $\\pmb{t}_l^{eos}$ , and $\\pmb{c}_{li}$ are the outputs of the previous transformer block and (b) $\\{\\pmb{p}_{li}^t\\}_{i=1}^{M^t}$ and $\\{\\pmb{\\mu}_i^t\\}_{i=1}^K$ are not the outputs of the previous transformer block; see Figure 2.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, we construct the input embeddings to the $l$ -th transformer block of the visual encoder as:", + "bbox": [ + 89, + 228, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {e} _ {l} ^ {c l s}, \\boldsymbol {x} _ {l 1} ^ {p}, \\dots , \\boldsymbol {x} _ {l M} ^ {p}, \\boldsymbol {\\mu} _ {1} ^ {v}, \\dots , \\boldsymbol {\\mu} _ {K} ^ {v}, \\boldsymbol {p} _ {l 1} ^ {v}, \\dots , \\boldsymbol {p} _ {l M ^ {v}} ^ {v} \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 270, + 483, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $l = 1,\\dots ,L$ $e_l^{cls}$ is the class token input to the $l$ -th transformer block, $\\{\\pmb{x}_{lm}^p\\}_{m = 1}^M$ is the set of patch embeddings input to the $l$ -th transformer block with a quantity of $M$ , and $\\{\\pmb{p}_{li}^v\\}_{i = 1}^{M^v}$ is the set of learnable visual prompts input to the $l$ -th transformer block with a quantity of $M^v$ . It is worth noting that (a) $e_l^{cls}$ and $\\{\\pmb{x}_{lm}^p\\}_{m = 1}^M$ are the outputs of the previous transformer block and (b) $\\{\\pmb{p}_{li}^v\\}_{i = 1}^{M^v}$ and $\\{\\pmb{\\mu}_i^v\\}_{i = 1}^K$ are not the outputs of the previous transformer block; see Figure 2.", + "bbox": [ + 89, + 299, + 483, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Undesired interaction masking. There is a modal gap between patch embeddings and visual task-aware pre-context. Allowing patch embeddings to interact with visual task-aware pre-context may weaken the discriminative ability of patch embeddings. To avoid this phenomenon, the information flow from visual task-aware pre-context to patch embeddings is prohibited, which is achieved by exploiting the masked attention mechanism: $\\mathbf{A} = \\mathrm{softmax}(Q\\mathbf{K}^T + \\mathcal{M})$ . Here, $\\mathbf{Q}$ and $\\mathbf{K}$ denote the query and key matrices, with the first dimension specifying the quantity of input embeddings, i.e., $N_v = 1 + M + K + M^v$ . They are the linear transformations of input embeddings. $\\mathbf{A} \\in \\mathbb{R}^{N_v \\times N_v}$ denotes the attention matrix. $\\mathcal{M} \\in \\mathbb{R}^{N_v \\times N_v}$ is the customized mask matrix, which is defined as:", + "bbox": [ + 89, + 436, + 483, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {i j} = \\left\\{ \\begin{array}{l l} - \\infty & \\text {i f} 1 < i \\leq 1 + M \\text {a n d} 1 + M < j \\leq 1 + M + K \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 657, + 482, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathcal{M}_{ij} = -\\infty$ is equal to $A_{ij} = 0$ , meaning that the $i$ -th input embedding does not interact with the $j$ -th peer. In effect, with $\\mathcal{M}$ , patch embeddings are made to not interact with visual task-aware pre-context. It is worth noting that (a) $\\mathcal{M}$ is injected into the visual encoder in a layer-wise manner; see Figure 2, and (b) the learnable prompts act as a glue between visual task-aware pre-context and patch embeddings.", + "bbox": [ + 89, + 704, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-regulation. Due to the scarcity of task-specific data, overfitting can easily occur. We address this issue by exploiting the strong generation ability exhibited by the well-pretrained encoders, i.e., encouraging the prompted class and visual embeddings to be consistent with their CLIP", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "peers. Specifically, let $\\{\\boldsymbol{w}_c^{\\mathrm{TAC}}\\}_{c=1}^C$ denote the class embeddings extracted by using our TAC. The textual consistency is imposed by:", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {t} = \\sum_ {i = 1} ^ {C} | \\boldsymbol {w} _ {c} ^ {\\mathrm {T A C}} - \\boldsymbol {w} _ {c} ^ {\\mathrm {C L I P}} |. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 146, + 906, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data augmentations are applied $N^d$ times on $x$ . Then, the augmented images are fed into our visual encoder to extract their embeddings, i.e., $\\{f_i^{\\mathrm{TAC}}\\}_{i=1}^{N^d}$ . Correspondingly, the visual consistency is imposed by:", + "bbox": [ + 511, + 198, + 905, + 261 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {v} = \\sum_ {i = 1} ^ {N ^ {d}} | \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}} - \\boldsymbol {f} ^ {\\mathrm {C L I P}} |. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 270, + 905, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $f^{\\mathrm{CLIP}}$ denotes the embedding extracted by using the visual encoder of CLIP on $x$ .", + "bbox": [ + 511, + 324, + 903, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall loss. The probability of $\\pmb{f}_i^{\\mathrm{TAC}}$ that belongs to the $c$ -th class can be calculated as:", + "bbox": [ + 511, + 354, + 903, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np ^ {\\mathrm {T A C}} \\left(c \\mid \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right) = \\frac {\\exp \\left(\\cos \\left(\\boldsymbol {w} _ {c} ^ {\\mathrm {T A C}} , \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(\\boldsymbol {w} _ {j} ^ {\\mathrm {T A C}}, \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right) / \\tau\\right)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 391, + 905, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the following cross-entropy loss can be constructed:", + "bbox": [ + 513, + 439, + 898, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} ^ {\\mathrm {T A C}} = - \\sum_ {i = 1} ^ {N ^ {d}} \\sum_ {c = 1} ^ {C} \\mathbb {I} (c = y) \\log p ^ {\\mathrm {T A C}} \\left(c \\mid \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 464, + 905, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After combining self-regulation, the overall loss becomes:", + "bbox": [ + 513, + 516, + 897, + 531 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {O v e r a l l}} = \\mathcal {L} _ {\\mathrm {C E}} ^ {\\mathrm {T A C}} + \\lambda \\mathcal {L} ^ {t} + \\beta \\mathcal {L} ^ {v}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 540, + 903, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\lambda$ and $\\beta$ are hyper-parameters that play a balancing role.", + "bbox": [ + 511, + 569, + 903, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Complexity. The discussion is three-fold. First, the clustering on $\\{\\boldsymbol{w}_c^{\\mathrm{CLIP}}\\}_{c=1}^C$ and the extraction of original class and visual embeddings are pre-computed and cached. Thus, the extra computational overhead incurred by them is marginal. Second, the textual task-aware pre-context does not incur extra computational overhead since the textual encoder of CLIP only accepts a sufficiently large fixed-number of input embeddings. Third, with the visual task-aware pre-context, the quantity of embeddings input to the visual encoder only increases by a very small amount, i.e., $2.5\\%$ . Thus, the extra computational overhead incurred is very limited. In summary, the extra computational overhead incurred by our TAC is marginal.", + "bbox": [ + 511, + 599, + 906, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 808, + 645, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment setup", + "text_level": 1, + "bbox": [ + 511, + 833, + 687, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation and datasets. The evaluation is conducted in the base-to-novel, domain generalization, and cross-dataset transfer settings, and the average accuracy of three", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14749", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1c0d17228419df4758dacac99a534b0a356e6a1b68b61b3cb650132908cbd93b.jpg", + "table_caption": [ + "Table 1. Comparative results in the base-to-novel setting. Bold accuracies are the highest. HM denotes the harmonic mean of the base and novel accuracies." + ], + "table_footnote": [], + "table_body": "
(a) Average(b) ImageNet(c) Caltech101(d) OxfordPets
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]69.3474.2271.70CLIP [37]72.4368.1470.22CLIP [37]96.8494.0095.40CLIP [37]91.1797.2694.12
CoOp [61]82.6963.2271.66CoOp [61]76.4767.8871.92CoOp [61]98.0089.8193.73CoOp [61]93.6795.2994.47
Co-CoOp [60]80.4771.6975.83Co-CoOp [60]75.9870.4373.10Co-CoOp [60]97.9693.8195.84Co-CoOp [60]95.2097.6996.43
ProGrad [62]82.4870.7576.16ProGrad [62]77.0266.6671.46ProGrad [62]98.0293.8995.91ProGrad [62]95.0797.6396.33
RPO [29]81.1375.0077.78RPO [29]76.6071.5774.00RPO [29]97.9794.3796.03RPO [29]94.6397.5096.05
MetaPrompt [58]83.3876.0979.57MetaPrompt [58]77.3971.0674.09MetaPrompt [58]98.2894.5896.39MetaPrompt [58]95.7196.9896.34
KAPT [21]78.4170.5274.26KAPT [21]71.1065.2068.02KAPT [21]97.1093.5395.28KAPT [21]93.1396.5394.80
KgCoOp [51]80.7373.6077.00KgCoOp [51]75.8369.9672.78KgCoOp [51]97.7294.3996.03KgCoOp [51]94.6597.7696.18
MaPLe [22]82.2875.1478.55MaPLe [22]76.6670.5473.47MaPLe [22]97.7494.3696.02MaPLe [22]95.4397.7696.58
PromptSRC [23]84.2676.1079.97PromptSRC [23]77.6070.7374.01PromptSRC [23]98.1094.0396.02PromptSRC [23]95.3397.3096.30
TCP [52]84.1375.3679.51TCP [52]77.2769.8773.38TCP [52]98.2394.6796.42TCP [52]94.6797.2095.92
CoPrompt [39]84.0077.2380.48CoPrompt [39]77.6771.2774.33CoPrompt [39]98.2794.9096.55CoPrompt [39]95.6798.1096.87
ArGue [45]83.6978.0780.78ArGue [45]76.9272.0674.41ArGue [45]98.4395.2096.79ArGue [45]95.3697.9596.64
TAC (Ours)85.2477.6081.24TAC (Ours)78.5771.0374.61TAC (Ours)98.5795.2796.89TAC (Ours)95.9398.1797.04
(e) StanfordCars(f) Flowers102(g) Food101(h) FGVCAircraft
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]63.3774.8968.65CLIP [37]72.0877.8074.83CLIP [37]90.1091.2290.66CLIP [37]27.1936.2931.09
CoOp [61]78.1260.4068.13CoOp [61]97.6059.6774.06CoOp [61]88.3382.2685.19CoOp [61]40.4422.3028.75
Co-CoOp [60]70.4973.5972.01Co-CoOp [60]94.8771.7581.71Co-CoOp [60]90.7091.2990.99Co-CoOp [60]33.4123.7127.74
ProGrad [62]77.6868.6372.88ProGrad [62]95.5471.8782.03ProGrad [62]90.3789.5989.98ProGrad [62]40.5427.5732.82
RPO [29]73.8775.5374.69RPO [29]94.1376.6784.50RPO [29]90.3390.8390.58RPO [29]37.3334.2035.70
MetaPrompt [58]75.4374.4374.93MetaPrompt [58]97.5374.5484.50MetaPrompt [58]90.7691.7791.26MetaPrompt [58]39.3837.5938.46
KAPT [21]69.4766.2067.79KAPT [21]95.0071.2081.40KAPT [21]86.1387.0686.59KAPT [21]29.6728.7329.19
KgCoOp [51]71.7675.0473.36KgCoOp [51]95.0074.7383.65KgCoOp [51]90.5091.7091.09KgCoOp [51]36.2133.5534.83
MaPLe [22]72.9474.0073.47MaPLe [22]95.9272.4682.56MaPLe [22]90.7192.0591.38MaPLe [22]37.4435.6136.50
PromptSRC [23]78.2774.9776.58PromptSRC [23]98.0776.5085.95PromptSRC [23]90.6791.5391.10PromptSRC [23]42.7337.8740.15
TCP [52]80.8074.1377.32TCP [52]97.7375.5785.23TCP [52]90.5791.3790.97TCP [52]41.9734.4337.83
CoPrompt [39]76.9774.4075.66CoPrompt [39]97.2776.6085.71CoPrompt [39]90.7392.0791.40CoPrompt [39]40.2039.3339.76
ArGue [45]75.6473.3874.49ArGue [45]98.3475.4185.36ArGue [45]92.3391.9692.14ArGue [45]40.4638.0339.21
TAC (Ours)81.6374.1777.72TAC (Ours)97.9776.8786.15TAC (Ours)90.8791.8791.37TAC (Ours)44.6037.7040.86
(i) SUN397(j) DTD(k) EuroSAT(l) UCF101
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]69.3675.3572.23CLIP [37]53.2459.9056.37CLIP [37]56.4864.0560.03CLIP [37]70.5377.5073.85
CoOp [61]80.6065.8972.51CoOp [61]79.4441.1854.24CoOp [61]92.1954.7468.69CoOp [61]84.6956.0567.46
Co-CoOp [60]79.7476.8678.27Co-CoOp [60]77.0156.0064.85Co-CoOp [60]87.4960.0471.21Co-CoOp [60]82.3373.4577.64
ProGrad [62]81.2674.1777.55ProGrad [62]77.3552.3562.45ProGrad [62]90.1160.8972.67ProGrad [62]84.3374.9479.35
RPO [29]80.6077.8079.18RPO [29]76.7062.1368.61RPO [29]86.6368.9776.79RPO [29]83.6775.4379.34
MetaPrompt [58]82.1079.0180.53MetaPrompt [58]82.5260.1069.55MetaPrompt [58]93.3778.3485.20MetaPrompt [58]84.7078.5681.51
KAPT [21]79.4074.3376.78KAPT [21]75.9758.3065.97KAPT [21]84.8067.5775.21KAPT [21]80.8367.1073.33
KgCoOp [51]80.2976.5378.36KgCoOp [51]77.5554.9964.35KgCoOp [51]85.6464.3473.48KgCoOp [51]82.8976.6779.65
MaPLe [22]80.8278.7079.75MaPLe [22]80.3659.1868.16MaPLe [22]94.0773.2382.35MaPLe [22]83.0078.6680.77
PromptSRC [23]82.6778.4780.52PromptSRC [23]83.3762.9771.75PromptSRC [23]92.9073.9082.32PromptSRC [23]81.7078.8082.74
TCP [52]82.6378.2080.35TCP [52]82.7758.0768.23TCP [52]91.6374.7382.32TCP [52]81.1380.7783.83
CoPrompt [39]82.6380.0381.31CoPrompt [39]83.1364.7372.79CoPrompt [39]94.6078.5785.84CoPrompt [39]86.9079.5783.07
ArGue [45]81.5280.7481.13ArGue [45]81.6066.5573.31ArGue [45]94.4388.2491.23ArGue [45]85.5679.2982.31
TAC (Ours)83.7080.0381.82TAC (Ours)83.3764.2772.58TAC (Ours)94.3782.6088.10TAC (Ours)88.0781.6784.75
", + "bbox": [ + 102, + 116, + 898, + 614 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/02340d6b08ae4eeca10f23fe85ddd21e1f520a2e09c7aa9e6f1b5cc4709a2b8d.jpg", + "table_caption": [ + "Table 2. Comparative results in the cross-dataset transfer setting. Bold accuracies are the highest. Optimization is conducted on ImageNet and evaluation is performed on other datasets." + ], + "table_footnote": [], + "table_body": "
SourceTarget
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101Average
CoOp [61]71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
Co-CoOp [60]71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
MaPLe [22]70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
Bayesian Prompt [10]70.9393.6790.6365.0070.9086.3024.9367.4746.1045.8768.6765.95
PromptSRC [23]71.2793.6090.2565.7070.2586.1523.9067.1046.8745.5068.7565.81
TCP [52]71.4093.9791.2564.6971.2186.6923.4567.1544.3551.4568.7366.29
CoPrompt [39]70.8094.5090.7365.6772.3086.4324.0067.5747.0751.9069.7367.00
TAC (Ours)72.7794.5390.6765.3072.2085.8323.5367.6347.5748.0770.0066.53
", + "bbox": [ + 99, + 656, + 897, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "runs is reported as the final performance. The datasets involved are as follows: ImageNet [9], Caltech101 [30], OxfordPets [36], StanfordCars [25], Flowers102 [35], Food101 [3], FGVCAircraft [44], SUN397 [50], DTD [8], EuroSAT [14], UCF101 [42], ImageNet-A [16], ImageNet-R [15], ImageNet-Sketch [47], and ImageNetV2 [38].", + "bbox": [ + 89, + 794, + 480, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Our TAC is built on top of", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CoOp [61] and PromptSRC [23]. The ViT-B/16 based CLIP [37] is used as the foundation model. The optimizer is SGD, with a learning rate of $2.5 \\times 10^{-3}$ , a weight decay of $5 \\times 10^{-4}$ , and a momentum of 0.9. The cosine annealing is used for learning rate scheduling. The random flip and resized crop are used for data augmentation. The batch size is 8. $K = 5$ , $M^t = 4$ , $M^v = 4$ , $N^d = 4$ , $\\lambda = 20$", + "bbox": [ + 511, + 792, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14750", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b1ecdd5413a2be6fdf980651bd3747fd2c1add895e84bc8fb2c0be117f3045f1.jpg", + "table_caption": [ + "Table 3. Comparative results in the domain generalization setting. Bold accuracies are the highest. Optimization is conducted on ImageNet and evaluation is performed on datasets with domain shifts." + ], + "table_footnote": [], + "table_body": "
SourceTarget
ImageNetImageNetV2ImageNet-SketchImageNet-AImageNet-RAverage
CLIP [37]66.7360.8346.1547.7773.9657.17
RPO [29]71.6765.1349.2750.1376.5760.27
LaViP [26]65.9561.6047.2348.9183.9360.41
TCP [52]71.2064.6049.5051.2076.7360.51
CoOp [61]71.5164.2047.9949.7175.2159.28
Co-CoOp [60]71.0264.0748.7550.6376.1859.90
ProGrad [62]72.2464.7347.6149.3974.5859.07
KgCoOp [51]71.2064.1048.9750.6976.7060.11
MaPLe [22]70.7264.0749.1550.9076.9860.26
Bayesian Prompt [10]70.9364.2349.2051.3377.0060.44
PromptSRC [23]71.2764.3549.5550.9077.8060.65
CoPrompt [39]70.8064.2549.4350.5077.5160.42
ArGue [45]71.8465.0249.2551.4776.9660.67
TAC (Ours)72.7765.9750.3051.7378.5061.63
", + "bbox": [ + 200, + 117, + 795, + 327 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fed9496978f3b3600b3d5e499564ac88297df6ac6ebd85d8b7ad80cf5b9cee0a.jpg", + "table_caption": [ + "Table 4. Ablation on the textual and visual task-aware precontexts. HM denotes the harmonic mean of the base and novel accuracies." + ], + "table_footnote": [], + "table_body": "
Pre-contextBaseNovelHM
None84.7075.5579.86
Textual84.5776.4880.32
Visual84.6776.0680.13
Textual & visual85.2477.6081.24
", + "bbox": [ + 112, + 385, + 459, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\beta = 25$ , and $\\tau = 0.07$ . In the base-to-novel setting, the number of epochs is 20 and the scale of random resized crop is (0.4, 1.0). In the domain generalization and cross-dataset transfer settings, the number of epochs is 5 and the scale of random resized crop is (0.8, 1.0). The learning rate is warmed up for one epoch with a rate of $1.0 \\times 10^{-5}$ .", + "bbox": [ + 89, + 478, + 483, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparative results", + "text_level": 1, + "bbox": [ + 89, + 582, + 284, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Base-to-novel. In this setting, optimization is conducted on the base set and evaluation is performed on the base and novel sets. Table 1 shows the comparative results. Our TAC achieves the best average base accuracy, the second-best average novel accuracy, and the best harmonic mean. These results demonstrate the generalizability of our task-aware clustering prompting framework.", + "bbox": [ + 89, + 604, + 483, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-dataset transfer. In this setting, optimization is conducted on ImageNet and evaluation is performed on other datasets. Table 2 shows the comparative results. Our TAC achieves the best accuracy on ImageNet. Also, our TAC yields the second-best average accuracy on other datasets, which is slightly worse than CoPrompt [39]. It is worth noting that our TAC beats CoPrompt in the baseto-novel and domain generalization settings. These results demonstrate the transferability of our task-aware clustering prompting framework.", + "bbox": [ + 89, + 718, + 482, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Domain generalization. In this setting, optimization is conducted on ImageNet and evaluation is performed on", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c9fd9d3d730c9b7c575320b4f7a52ae5452d70b110ec665a099424f31a387da3.jpg", + "table_caption": [ + "Table 5. Ablation on the number of clustering centers. HM denotes the harmonic mean of the base and novel accuracies." + ], + "table_footnote": [], + "table_body": "
Number of clustering centersBaseNovelHM
K=084.7075.5579.86
K=184.9577.3980.99
K=284.9477.4181.00
K=385.3077.3281.14
K=485.4077.3481.17
K=585.2477.6081.24
", + "bbox": [ + 532, + 371, + 883, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "datasets with domain shifts. Table 3 shows the comparative results. Our TAC achieves the best accuracy on ImageNet and the best average accuracy on the datasets with domain shifts. These results demonstrate the domain generalizability of our task-aware clustering prompting framework.", + "bbox": [ + 511, + 486, + 906, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 511, + 574, + 674, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation studies are conducted in the base-to-novel setting. The average base accuracy, the average novel accuracy, and their harmonic mean are reported. More is shown in the appendix.", + "bbox": [ + 511, + 597, + 905, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task-aware pre-context. We ablate the effect of the task-aware pre-context by removing the textual pre-context, the visual pre-context, or both. It is worth noting that the other ingredients are kept in this experiment. Table 4 shows the results. Our observations are as follows: (a) both the textual and visual task-aware pre-contexts can improve the harmonic mean and (b) their combination achieves the greatest performance gains. These observations demonstrate the effectiveness of introducing task-aware pre-context.", + "bbox": [ + 511, + 657, + 905, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of clustering centers. We ablate the effect of the number of clustering centers by setting $K$ to different values. It is worth noting that (a) the maximum value of $K$ is 5 since the dataset EuroSAT [14] only has 5 classes and (b) the other ingredients are kept in this experiment. Table 5 shows the results. Our observations are as follows: (a) $K = 0$ means that the task-aware pre-context is", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14751", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2743fb8577246e3f74fe479c4abff6478bbf19c47c90c05ce18f8acff3a0bfa3.jpg", + "table_caption": [ + "Table 6. Ablation on the self-regulation. HM denotes the harmonic mean of the base and novel accuracies." + ], + "table_footnote": [], + "table_body": "
Self-regulationBaseNovelHM
LCE TAC85.2974.1879.34
LCE TAC + λLt83.5273.7378.32
LCE TAC + βLv84.4873.8978.83
LCE TAC + λLt + βLv85.2477.6081.24
", + "bbox": [ + 109, + 116, + 459, + 199 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/beb33bad77cf866b5225641b273eabaa9d1c24c13095ca489e82aedb21ba13fe.jpg", + "table_caption": [ + "Table 7. Ablation on the undesired interaction masking. HM denotes the harmonic mean of the base and novel accuracies." + ], + "table_footnote": [], + "table_body": "
Undesired interaction maskingBaseNovelHM
×85.0077.0580.83
✓85.2477.6081.24
", + "bbox": [ + 109, + 237, + 459, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "not exploited, (b) the introduction of task-aware pre-context can improve performance by a noticeable margin, and (c) the biggest performance gains are achieved when $K = 5$ . These observations demonstrate the effectiveness of introducing task-aware pre-context and lead us to set the default value of $K$ to 5.", + "bbox": [ + 89, + 305, + 482, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-regulation. We ablate the effect of the self-regulation by removing the textual consistency, the visual consistency, or both. It is worth noting that the other ingredients are kept in this experiment. Table 6 shows the results. Our observations are as follows: (a) without imposing self-regulation, the accuracy on the novel set is very poor, which also leads to a low harmonic mean, (b) imposing the textual consistency or visual consistency alone brings a noticeable performance drop, and (c) imposing the two consistencies simultaneously leads to a significant performance gain. The reason for these observations might be that the alignment between textual and visual modalities is well preserved by imposing the two consistencies, while it is difficult to achieve this with only one consistency. These observations demonstrate the necessity of self-regulation.", + "bbox": [ + 89, + 397, + 482, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Undesired interaction masking. We ablate the effect of the undesired interaction masking by removing it. It is worth noting that (a) the mask matrix is inserted into the visual encoder in a layer-wise manner and (b) the other ingredients are kept in this experiment. Table 7 shows the results. With the undesired interaction masking, the harmonic mean is improved by a margin of $0.41\\%$ . Considering the performance gain is achieved in an experiment setting involving 11 datasets, this observation demonstrates the effectiveness of undesired interaction masking.", + "bbox": [ + 89, + 626, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complexity. We evaluate the complexity by comparing our TAC with the competitive counterparts. It is worth noting that (a) PromptSRC [23] inserts learnable prompts into the textual and visual encoders in a layer-wise manner and (b) our TAC inserts the learnable prompts, along with the task-aware pre-contexts, into the textual and visual encoders in the same way. Table 8 shows the results. Since the task-aware pre-context generation is lightweight and the quan-", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6aafa2449730a9dd3960d0044f74cc7561d588171b5e7bdcf3142ab96ae7dddb.jpg", + "table_caption": [ + "Table 8. Ablation on the complexity. Evaluation is performed on a RTX 3090 GPU." + ], + "table_footnote": [], + "table_body": "
MethodOptimization (ms/image)Inference (ms/image)
CLIP [37]-1.7
CoOp [61]68.13.0
CoCoOp [60]269.0126.6
MaPLe [22]70.21.7
PromptSRC [23]71.53.8
TAC (Ours)73.24.1
", + "bbox": [ + 532, + 116, + 880, + 213 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a001f0a12f8bcf713fa29926b0290eaf88988508bacb4426cf068530e9b32825.jpg", + "table_caption": [ + "Table 9. Ablation on the number of learnable prompts. HM denotes the harmonic mean of the base and novel accuracies." + ], + "table_footnote": [], + "table_body": "
Number of learnable promptsBaseNovelHM
184.1575.6379.66
284.6776.6180.44
485.2477.6081.24
685.0276.8580.73
", + "bbox": [ + 532, + 246, + 880, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tity of additional input embeddings incurred by the task-aware pre-context is small, the extra computational overhead incurred by our TAC is marginal compared to PromptSRC [23]. Moreover, it is worth noting our TAC outperforms PromptSRC with noticeable performance gains in all experiment settings.", + "bbox": [ + 511, + 332, + 906, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of learnable prompts. We ablate the effect of the number of learnable prompts by setting $M^t$ and $M^v$ to different values. It is worth noting that (a) $M^t$ and $M^v$ are set to be equal by default and (b) the other ingredients are kept in this experiment. Table 9 shows the results. Our observations are as follows: (a) increasing the number of learnable prompts can improve the harmonic mean and (b) the greatest performance gain is achieved when $M^t = 4$ and $M^v = 4$ . These observations lead us to set the default value of $M^t$ and $M^v$ to 4.", + "bbox": [ + 511, + 422, + 908, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 587, + 633, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt learning has emerged as a powerful tool for adapting vision-language models. However, due to the scarcity of task-specific data, the resulting prompts often exhibit limited task-awareness. In this work, we propose to increase the task-awareness by introducing task-aware pre-context, which results in a novel task-aware clustering framework for prompting. The key ingredients are three-fold. First, the backbone structure of a downstream task is preserved by task-aware clustering, based on which the task-aware pre-context is generated. Second, the task-awareness of learnable prompts is enhanced by allowing them to interact with task-aware pre-context via the well-pretrained encoders. Finally, the undesired interference of visual task-aware pre-context to patch embeddings is eliminated by masked attention mechanism. Extensive ablation studies confirm the effectiveness of key ingredients. Comparative results in the base-to-novel, domain generalization, and cross-dataset transfer settings show the advantage of our task-aware clustering for prompt learning.", + "bbox": [ + 511, + 612, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14752", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Guangdong Major Project of Basic and Applied Basic Research (2023B0303000016), National Natural Science Foundation of China (62206268, U21A20487, 62372440), Shenzhen Technology Project (JCYJ20240813154723031), Shenzhen High-tech Zone Development Special Plan Innovation Platform Construction Project, the proof of concept center for high precision and high resolution 4D imaging, Guangdong Technology Project (2023TX07Z126), and Yunnan Science & Technology Project (202302AD080008, 202305AF150152).", + "bbox": [ + 89, + 114, + 485, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 277, + 187, + 292 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1", + "[2] Shtedritski Aleksandar, Rupprecht Christian, and Vedaldi Andrea. What does clip know about a red circle? visual prompt engineering for vlms. In ICCV, pages 11987-11997, 2023. 1, 3", + "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446-461, 2014. 6", + "[4] Adrian Bulat and Georgios Tzimiropoulos. Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In CVPR, pages 23232-23241, 2023. 1", + "[5] Yun-Hao Cao, Kaixiang Ji, Ziyuan Huang, Chuanyang Zheng, Jiajia Liu, Jian Wang, Jingdong Chen, and Ming Yang. Towards better vision-inspired vision-language models. In CVPR, pages 13537-13547, 2024. 2", + "[6] Delong Chen, Zhao Wu, Fan Liu, Zaiquan Yang, Shaoqiu Zheng, Ying Tan, and Erjin Zhou. Protoclip: Prototypical contrastive language image pretraining. IEEE TNNLS (Early Access), pages 1-15, 2023. 2", + "[7] Eulrang Cho, Jooyeon Kim, and Hyunwoo J. Kim. Distribution-aware prompt tuning for vision-language models. In ICCV, pages 22004-22013, 2023. 1", + "[8] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 6", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 2, 6", + "[10] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G. Turrisi da Costa, Cees G.M. Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learning for image-language model generalization. In ICCV, pages 15237-15246, 2023. 6, 7" + ], + "bbox": [ + 93, + 301, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 1", + "[12] Bordes Florian, Pang Richard Yuanzhe, Ajay Anurag, Li Alexander C, Bardes Adrien, Petryk Suzanne, Manas Oscar, Lin Zhiqiu, Mahmoud Anas, Jayaraman Bargav, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2", + "[13] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE TNNLS (Early Access), pages 1-11, 2023. 2", + "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. J-STARS, 12(7):2217-2226, 2019. 6, 7", + "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6", + "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 6", + "[17] Nisha Huang, Yuxin Zhang, Fan Tang, Chongyang Ma, Haibin Huang, Weiming Dong, and Changsheng Xu. Diff-styler: Controllable dual diffusion for text-driven image stylization. IEEE TNNLS (Early Access), pages 1-14, 2024. 2", + "[18] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, pages 3061-3071, 2023. 3", + "[19] Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Retrieval-enhanced contrastive vision-text models. In ICLR, 2024. 2", + "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1, 2", + "[21] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 1, 6", + "[22] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 6, 7, 8", + "[23] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 6, 7, 8", + "[24] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14753", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Jonathan Krause, Michael Stark, Jia Deng, and Fei-Fei Li. 3d object representations for fine-grained categorization. In ICCV, pages 554–561, 2013. 6", + "[26] Nilakshan Kunananthaseelan, Jing Zhang, and Mehrtash Harandi. Lavip: Language-grounded visual prompting. In AAAI, pages 2840-2848, 2024. 3, 7", + "[27] Gukyeong Kwon, Zhaowei Cai, Avinash Ravichandran, Erhan Bas, Rahul Bhotika, and Stefano Soatto. Masked vision and language modeling for multi-modal representation learning. In ICLR, 2023. 2", + "[28] Marc Lafon, Elias Ramzi, Clément Rambour, Nicolas Audebert, and Nicolas Thome. Gallop: Learning global and local prompts for vision-language models. In ECCV, 2024. 3", + "[29] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J. Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 1, 6, 7", + "[30] Fei-Fei Li, Fergus Rob, and Perona Pietro. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshop, pages 178–178, 2004. 6", + "[31] Yanxin Long, Jianhua Han, Runhui Huang, Hang Xu, Yi Zhu, Chunjing Xu, and Xiaodan Liang. Fine-grained visual-text prompt-driven self-training for open-vocabulary object detection. IEEE TNNLS (Early Access), pages 1-11, 2023. 2", + "[32] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In CVPR, pages 5206-5215, 2022. 3", + "[33] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In CVPR, pages 26354-26363, 2024. 3", + "[34] Metaxas Ioannis Maniadis, Tzimiropoulos Georgios, and Paras Ioannis. Divclust: Controlling diversity in deep clustering. In CVPR, pages 3418-3428, 2023. 3", + "[35] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 6", + "[36] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505, 2012. 6", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 6, 7, 8", + "[38] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imagenet? In ICML, pages 5389-5400, 2019. 6", + "[39] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. In ICLR, 2024. 2, 3, 6, 7", + "[40] Lavoie Samuel, Kirichenko Polina, Ibrahim Mark, Assran Mido, Wilson Andrew Gordon, Courville Aaron, and Ballas Nicolas. Modeling caption diversity in contrastive vision-language pretraining. In ICML, pages 26070-26084, 2024. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Tourani Siddharth, Alwheibi Ahmed, Mahmood Arif, and Khan Muhammad Haris. Pose-guided self-training with two-stage clustering for unsupervised landmark discovery. In CVPR, pages 23041-23051, 2024. 3", + "[42] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6", + "[43] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In CVPR, pages 23209-23218, 2024. 2", + "[44] Maji Subhransu, Rahtu Esa, Kannala Juho, Blaschko Matthew, and Vedaldi Andrea. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6", + "[45] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argument: Attribute-guided prompt tuning for vision-language models. In CVPR, pages 28578-28587, 2024. 2, 3, 6, 7", + "[46] Yujun Tong, Da Li, Dongliang Chang, Tianwei Cao, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Vision-language subspace prompting, 2024. 2", + "[47] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 6", + "[48] Mengmeng Wang, Jiazheng Xing, Jianbiao Mei, Yong Liu, and Yunliang Jiang. Actionclip: Adapting language-image pretrained models for video action recognition. IEEE TNNLS (Early Access), pages 1-13, 2023. 2", + "[49] Zihao Wei, Zixuan Pan, and Andrew Owens. Efficient vision-language pre-training by cluster masking. In CVPR, pages 26815-26825, 2024. 2", + "[50] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 6", + "[51] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, pages 6757-6767, 2023. 1, 6, 7", + "[52] Hantao Yao, Rui Zhang, and Changsheng Xu. Tcp: Textual-based class-aware prompt tuning for visual-language model. In CVPR, pages 23438-23448, 2024. 3, 6, 7", + "[53] Lochman Yaroslava, Olsson Carl, and Zach Christopher. Learned trajectory embedding for subspace clustering. In CVPR, pages 19092-19102, 2024. 3", + "[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 2", + "[55] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 2", + "[56] Yunfei Zhang, Xiaoyang Huo, Tianyi Chen, Si Wu, and Hau San Wong. Exploring intra-class variation factors with learnable cluster prompts for semi-supervised image synthesis. In CVPR, pages 7392-7401, 2023. 3" + ], + "bbox": [ + 514, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14754", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In CVPR, pages 28718-28728, 2024. 2", + "[58] Cairong Zhao, Yubin Wang, Xinyang Jiang, Yifei Shen, Kaitao Song, Dongsheng Li, and Duoqian Miao. Learning domain invariant prompt for vision-language models. IEEE TIP, 33:1348-1360, 2024. 6", + "[59] Zhaoheng Zheng, Jingmin Wei, Xuefeng Hu, Haidong Zhu, and Ram Nevatia. Large language models are good prompt learners for low-shot image classification. In CVPR, pages 28453-28462, 2024. 3", + "[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 6, 7, 8", + "[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 6, 7, 8", + "[62] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In CVPR, pages 15659-15669, 2023. 6, 7", + "[63] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + ], + "bbox": [ + 89, + 92, + 482, + 445 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14755", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Task-Aware Clustering for Prompting Vision-Language Models/aabdfb34-0949-4921-9337-b3ad9044174e_model.json b/2025/Task-Aware Clustering for Prompting Vision-Language Models/aabdfb34-0949-4921-9337-b3ad9044174e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..881831fe629a289c9bd4250866eb831ab767da1a --- /dev/null +++ b/2025/Task-Aware Clustering for Prompting Vision-Language Models/aabdfb34-0949-4921-9337-b3ad9044174e_model.json @@ -0,0 +1,2180 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.82, + 0.154 + ], + "angle": 0, + "content": "Task-Aware Clustering for Prompting Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.179, + 0.899, + 0.237 + ], + "angle": 0, + "content": "Fusheng Hao\\(^{1,2}\\) Fengxiang He\\(^{3}\\) Fuxiang Wu\\(^{1,2}\\) Tichao Wang\\(^{1,2}\\) Chengqun Song\\(^{1,2}\\) Jun Cheng\\(^{1,2*}\\) \n\\(^{1}\\) Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n\\(^{2}\\) The Chinese University of Hong Kong" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.695 + ], + "angle": 0, + "content": "Prompt learning has attracted widespread attention in adapting vision-language models to downstream tasks. Existing methods largely rely on optimization strategies to ensure the task-awareness of learnable prompts. Due to the scarcity of task-specific data, overfitting is prone to occur. The resulting prompts often do not generalize well or exhibit limited task-awareness. To address this issue, we propose a novel Task-Aware Clustering (TAC) framework for prompting vision-language models, which increases the task-awareness of learnable prompts by introducing task-aware pre-context. The key ingredients are as follows: (a) generating task-aware pre-context based on task-aware clustering that can preserve the backbone structure of a downstream task with only a few clustering centers, (b) enhancing the task-awareness of learnable prompts by enabling them to interact with task-aware pre-context via the well-pretrained encoders, and (c) preventing the visual task-aware pre-context from interfering the interaction between patch embeddings by masked attention mechanism. Extensive experiments are conducted on benchmark datasets, covering the base-to-novel, domain generalization, and cross-dataset transfer settings. Ablation studies validate the effectiveness of key ingredients. Comparative results show the superiority of our TAC over competitive counterparts. The code is available at https://github.com/FushengHao/TAC." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Large-scale visual and textual data pairs have been collected and employed to pretrain high-capacity Vision-Language Models (VLMs) [1, 20, 37]. Due to the alignment of the visual and textual modalities, the resulting models have emerged with remarkable generalization ability on various downstream tasks such as zero-shot image classification [37] and open-vocabulary object detection [11]. Despite achieving decent performance, the performance is still" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.272, + 0.882, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.484, + 0.907, + 0.527 + ], + "angle": 0, + "content": "Figure 1. Motivation of our TAC. We improve the task-awareness of learnable prompts from a new perspective, i.e., introducing task-aware pre-context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.629 + ], + "angle": 0, + "content": "lower than that of the customized models [61]. Considering the data-hungry nature of VLMs and the scarcity of task-specific data, it is difficult to improve performance through finetuning such models. As an alternative to finetuning, prompting has become as a powerful paradigm for adapting VLMs to downstream tasks [4, 7, 21, 22, 29, 51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Prompting is a technique that adapts VLMs by increasing their understanding of downstream tasks. The current methodology is to provide VLMs with task-relevant context while keeping the model weights frozen. The context is also known as prompts. Many effective prompt templates have been engineered. For example, textual templates like \"a photo of a [class]\" [37] have been hand-crafted to provide task-relevant context for the textual encoder, and visual templates like \"a red circle around an [object]\" [2] have been discovered to direct the visual encoder's attention to the [object]. To avoid the need for large-scale trial and error in prompt engineering, context optimization [61] has been proposed, which treats learnable vectors as prompts and ensure their task-awareness by optimization. This strategy has attracted much attention in recent years. For example, maximizing the ground-truth class scores [61] is proposed to ensure the task-awareness of learnable prompts, and self-regulating constraints [23] are further introduced to guide" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.355, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: jun.cheng@siat.ac.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14745" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "the optimization of learnable prompts for better generalization. It is worth noting that the number of labeled images per class used for prompt learning is not sufficient in the setting of context optimization1. Due to the scarcity of task-specific data, the resulting prompts are often overfitted [23] and do not generalize well, meaning that the task-awareness they exhibit is limited." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.199, + 0.483, + 0.5 + ], + "angle": 0, + "content": "Instead of imposing stronger constraints [39, 45, 46], we propose to increase the task-awareness by directly introducing task-aware pre-context; see Figure 1. Considering that the class descriptions of a downstream task are given when encountering the task, the class embeddings extracted by using the well-pretrained textual encoder are good candidates for task-aware pre-context. However, one big challenge arises in this practice, i.e., the number of extra embeddings it incurs is large and equal to that of classes involved. For example, the number is 1000 for the popular dataset ImageNet [9], which is significantly larger than the typical values of the number of embeddings input to the textual and visual encoders, i.e., 77 and 197, respectively. This not only incurs large additional computational overhead, but also exceeds the maximum input limit for the textual encoder. These difficulties inspire us to keep only the backbone structure of class embeddings, which is helpful for reducing the number of embeddings incurred significantly. Motivated by these observations, we explore the prompting strategy based on clustering." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.501, + 0.483, + 0.877 + ], + "angle": 0, + "content": "In this work, we propose a novel Task-Aware Clustering (TAC) framework for prompting VLMs. The central idea of TAC is to exploit task-aware pre-context to improve the task-awareness of learnable prompts. To generate task-aware pre-context, we conduct clustering on the class embeddings of a downstream task and perform linear transformation on the clustering centers to yield task-aware pre-context, which leads to a task-aware pre-context generalization module that can keep the backbone structure of the downstream task while being lightweight. Then, the task-aware pre-context is combined with the learnable prompts and their interaction through the well-pretrained encoders is exploited to reinforce the task-awareness of learnable prompts, which leads to a task-awareness reinforcement module. Further, a mask matrix is developed and injected into the attention mechanism to eliminate the undesired interference of visual task-aware pre-context to patch embeddings induced by the gap between textual and visual modalities, which leads to undesired interaction masking module. To suppress the overfitting induced by the scarcity of task-specific data, the prompted class and visual embeddings are encouraged to be consistent with their CLIP [37] peers, which leads to a self-regulation module. Our TAC maintains simplicity in design and improves the task-awareness of the resulting prompts, thus yielding competitive results." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "In summary, our task-aware clustering framework for prompting has the following main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.124, + 0.905, + 0.168 + ], + "angle": 0, + "content": "- We propose to increase the task-awareness of learnable prompts from a new perspective, i.e., introducing task-aware pre-context." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.17, + 0.905, + 0.213 + ], + "angle": 0, + "content": "- We propose to generate task-aware pre-context based on task-aware clustering that can preserve the backbone structure of a downstream task." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.215, + 0.905, + 0.258 + ], + "angle": 0, + "content": "- We exploit the well-pretrained encoders to reinforce the task-awareness of learnable prompts while masking out the undesired interaction." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.259, + 0.905, + 0.305 + ], + "angle": 0, + "content": "- We demonstrate the superiority of our TAC in the baseto-novel, domain generalization, and cross-dataset transfer settings." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.124, + 0.905, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.323, + 0.663, + 0.338 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) integrate the textual and visual modalities, bridging the gap between them through alignment. Such models are usually pretrained on large-scale visual and textual data pairs [6, 20, 37]. They can be broadly classified into four categories according to their training paradigms [12]. The first category employs contrastive-based training [19, 40], which emphasizes reducing the distance between positive pairs and increasing the distance between negative pairs. The second category utilizes masking-based training [27, 49], which emphasizes the reconstruction of missing patches/words by exploiting the unmasked caption/image. The third category leverages generative-based training [54, 55], which emphasizes the generation of entire images or captions based on conditional captions or images. The fourth category employs LLM-based training [5, 63], which emphasizes the usage of large language models to extract supervisory information. Since the textual and visual modalities are aligned in a shared embedding space, VLMs exhibits a strong ability to understand open-vocabulary concepts, which facilitates various downstream tasks. Although VLMs have achieved decent performance, their accuracies are still far inferior to those of customized models [61]. Considering the data-hungry nature of VLMs and the scarcity of task-specific data, finetuning such models to improve their performance faces significant challenges. As an alternative to finetuning, prompting [37, 61] has attracted much attention in adapting VLMs. Orthogonal directions include the utilization of unlabeled images [43] and historical test data [57]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Prompting adapts VLMs by providing them with contextual information or task-aware prompts while freezing their model weights, which has achieved good performance in many downstream tasks [13, 17, 24, 31, 48]. Existing methods can be broadly classified into two categories according to whether domain expertise is required. The first category relies on hand-crafted prompt templates, which emphasizes" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.321, + 0.9 + ], + "angle": 0, + "content": "The typical value of this number is 16" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14746" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.394 + ], + "angle": 0, + "content": "the reliance on domain expertise. CLIP [37] crafts the textual template, i.e., \"a photo of a [class]\". CirclePrompt [2] engineers the visual template, i.e., \"a red circle around an [object]\". LaViP [26] grounds visual prompt with language. ArGue [45] crafts attribute-guided prompts with the help of large language models. The second category relies on prompt learning [61], which emphasizes the introduction of learnable prompts and the optimization of these prompts to ensure their task-awareness. PromptSRC [23] guides the optimization of learnable prompts with self-regulating constraints. GalLoP [28] learns global-local prompts to capture multi-scale visual information. CoPrompt [39] improves the generation ability of learnable prompts by encouraging the prediction consistency between the pretrained and trainable models. TCP [52] introduces class-aware prompts that are produced by performing textual knowledge embedding on class embeddings. In this work, we increase the task-awareness by introducing task-aware pre-context, which is different from the current practice that imposes extra constraints." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.401, + 0.485, + 0.719 + ], + "angle": 0, + "content": "Clustering is a technique that can group similar data points into clusters or groups, in which K-Means is a classic example. Integrating clustering may improve the capabilities of deep models in capturing backbone structures, and their combination has attracted a lot of attention. DivClust [34] introduces a diversity controlling loss to ensure the desired degree of diversity for the produced multiple clusters. LCPGAN [56] associates intra-class clusters with learnable visual concepts to capture the class-wise and intra-class information for image synthesis. ZeroShotULD [41] clusters random pixel locations with nearest neighbor searching for unsupervised landmarks discovery. SPC [18] replaces the deterministic prediction with semantic clustering for domain generalized semantic segmentation. MoDE [33] decides clusters for data experts by a two-step clustering on captions and uses fine-grained clusters to coarsely represent data experts. LTE [53] clusters multiple observed motions by establishing the mapping from trajectories to trajectory embeddings. In this work, we perform clustering to preserve the backbone structure of a downstream task represented by class embeddings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.184, + 0.75 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.486, + 0.807 + ], + "angle": 0, + "content": "In this section, we begin by reviewing the preliminaries and then detail our Task-Aware Clustering (TAC) framework for prompting vision-language models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.233, + 0.833 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "CLIP. CLIP [37] includes a textual encoder and a visual encoder, and is pretrained on around 400 million visual and textual data pairs with the contrastive-based training strategy. With the hand-crafted prompt templates like \"a photo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.302 + ], + "angle": 0, + "content": "of a [class]\", it emerges strong zero-shot prediction capabilities. Although CLIP has achieved good performance in the zero-shot image recognition task, its accuracies are still lower than those of the customized models [61]. Due to the data-hungry nature of CLIP and the scarcity of task-specific data, finetuning CLIP to improve its performance on downstream tasks is particularly challenging. Considering that prompt engineering relies on domain expertise and needs large-scale trial and error, prompt learning [61] has been proposed, in which learnable vectors are treated as prompts and their task-awareness is ensured through optimization. For ease of comparison with existing prompt learning approaches, we follow the current practice and choose CLIP as the foundation model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.909, + 0.529 + ], + "angle": 0, + "content": "Prompt learning. CoOp [61] is a typical approach for adapting CLIP in the context of prompt learning. We take a \\(C\\)-class image classification task as an example to show how learnable prompts are optimized in CoOp. First, a set of learnable vectors with a quantity of \\(M^t\\) is introduced and treated as textual prompts, i.e., \\(\\{\\pmb{p}_m^t\\}_{m=1}^{M^t}\\). Second, these prompts are combined with the word embeddings that describe the task, which leads to a learnable version of task descriptions, i.e., \\(\\{\\pmb{p}_1^t,\\dots,\\pmb{p}_{M^t}^t,\\pmb{c}_i\\}_{i=1}^C\\), where \\(\\pmb{c}_i\\) is the word embedding(s) of the \\(i\\)-th class. Third, the learnable task descriptions are fed into the textual encoder to yield their embeddings \\(\\{\\pmb{w}_c^{\\mathrm{CoOp}}\\}_{c=1}^C\\), while an image \\(\\pmb{x}\\) is fed into the visual encoder to generate its embedding \\(\\pmb{f}\\). Fourth, the probability of \\(\\pmb{x}\\) that belongs to the \\(c\\)-th class can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.533, + 0.907, + 0.573 + ], + "angle": 0, + "content": "\\[\np ^ {\\mathrm {C o O p}} (c | \\boldsymbol {x}) = \\frac {\\exp \\left(\\cos \\left(\\boldsymbol {w} _ {c} ^ {\\mathrm {C o O p}} , \\boldsymbol {f}\\right) / \\tau\\right)}{\\sum_ {i = 1} ^ {C} \\exp \\left(\\cos \\left(\\boldsymbol {w} _ {i} ^ {\\mathrm {C o O p}}, \\boldsymbol {f}\\right) / \\tau\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.577, + 0.908, + 0.624 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature and \\(\\cos (\\cdot ,\\cdot)\\) denotes the cosine similarity. Finally, the task-awareness of learnable textual prompts is ensured by minimizing the cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.628, + 0.907, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} ^ {\\mathrm {C o O p}} = - \\sum_ {c = 1} ^ {C} \\mathbb {I} (c = y) \\log p ^ {\\mathrm {C o O p}} (c | \\boldsymbol {x}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.909, + 0.764 + ], + "angle": 0, + "content": "where \\(y\\) is the ground-truth label of \\(\\pmb{x}\\) and \\(\\mathbb{I}(\\cdot)\\) denotes the indicator function. It is worth noting that (a) only the learnable prompts need to be optimized and (b) the quantity of labeled images that can be used in each class is insufficient in the setting of prompt learning, where the typical value is 16." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.908, + 0.843 + ], + "angle": 0, + "content": "Afterwards, additional constraints [23, 32, 39, 59] have been imposed to improve the task-awareness of learnable prompts. Different from the current practice, we increase the task-awareness from a new perspective, i.e., introducing task-aware pre-context." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.84, + 0.866 + ], + "angle": 0, + "content": "3.2. Task-aware Clustering for Prompting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "The pipeline of our TAC is shown in Figure 2 and the key ingredients are detailed below." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14747" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.092, + 0.891, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.516, + 0.908, + 0.575 + ], + "angle": 0, + "content": "Figure 2. Pipeline of our TAC. First, the textual and visual task-aware pre-contexts are generated based on the backbone structure of a downstream task. Second, the task-aware pre-context is combined with the learnable prompts and their interaction through the well-pretrained encoders is exploited to reinforce the task-awareness of learnable prompts. Third, the undesired interaction is masked out with a customized mask matrix. Further, the textual and visual consistencies are imposed to alleviate the overfitting issue." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Task-aware pre-context generation. The class embeddings extracted by using the textual encoder of CLIP are task-aware and good candidates for the desired pre-context. Considering that the number of such embeddings is equal to that of classes involved and this number is large, applying this strategy to practice is challenging. The first challenge is that the number of extra input embeddings incurred exceeds the maximum input limit for the textual encoder, which would make the textual encoder unusable. The second challenge is that the extra computational overhead incurred is very high due to the large number of additional embeddings input into the visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.904 + ], + "angle": 0, + "content": "We address these challenges by yielding textual and visual task-aware pre-contexts based on clustering that can preserve the backbone structure of class embeddings. A significant additional benefit of this practice is that the number of extra embeddings incurred can be significantly reduced. Specifically, let \\(\\{\\boldsymbol{w}_c^{\\mathrm{CLIP}}\\}_{c=1}^C\\) denote the class embeddings extracted with the textual encoder of CLIP. Then, K-Means is conducted on \\(\\{\\boldsymbol{w}_c^{\\mathrm{CLIP}}\\}_{c=1}^C\\), which yields a set of clustering centers with a quantity of \\(K\\), i.e., \\(\\{\\boldsymbol{\\mu}_i\\}_{i=1}^K\\). Further, the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.584, + 0.907, + 0.6 + ], + "angle": 0, + "content": "textual and visual task-aware pre-contexts are generated as:" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.61, + 0.907, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {i} ^ {t} = \\boldsymbol {W} ^ {t} \\boldsymbol {\\mu} _ {i}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.662, + 0.639, + 0.907, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {i} ^ {v} = \\boldsymbol {W} ^ {v} \\boldsymbol {\\mu} _ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.907, + 0.752 + ], + "angle": 0, + "content": "where \\( \\boldsymbol{W}^t \\) and \\( \\boldsymbol{W}^v \\) are the textual and visual linear transformations. It is worth noting that (a) the dimensions of \\( \\boldsymbol{\\mu}_i^t \\) and \\( \\boldsymbol{\\mu}_i^v \\) are set to be the internal dimensions of the textual and visual encoders, correspondingly, and (b) since only the weights of linear transformations needs to be optimized, our task-aware pre-context generation module is lightweight." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.909, + 0.873 + ], + "angle": 0, + "content": "Task-awareness reinforcement. After obtaining the textual and visual task-aware pre-contexts, the task-awareness of learnable prompts is reinforced by enabling them to interact with the corresponding pre-contexts via the well-pretrained textual and visual encoders. Specifically, let \\( L \\) denote the quantity of transformer blocks contained in the visual/textual encoder. We construct the input embeddings to the \\( l \\)-th transformer block of the textual encoder as:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.884, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {t} _ {l} ^ {s o s}, \\boldsymbol {\\mu} _ {1} ^ {t}, \\dots , \\boldsymbol {\\mu} _ {K} ^ {t}, \\boldsymbol {p} _ {l 1} ^ {t}, \\dots , \\boldsymbol {p} _ {l M ^ {t}} ^ {t}, \\boldsymbol {c} _ {l i}, \\boldsymbol {t} _ {l} ^ {e o s} \\right\\}. \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14748" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.484, + 0.228 + ], + "angle": 0, + "content": "Here, \\(l = 1,\\dots ,L\\), \\(\\pmb{t}_l^{sos}\\) is the start token input to the \\(l\\)-th transformer block, \\(\\pmb{t}_l^{eos}\\) is the end token input to the \\(l\\)-th transformer block, \\(\\{\\pmb{p}_{li}^t\\}_{i=1}^{M^t}\\) is the set of learnable textual prompts input to the \\(l\\)-th transformer block with a quantity of \\(M^t\\), and \\(\\pmb{c}_{li}\\) is the word embedding(s) input to the \\(l\\)-th transformer block. It is worth noting that (a) \\(\\pmb{t}_l^{sos}\\), \\(\\pmb{t}_l^{eos}\\), and \\(\\pmb{c}_{li}\\) are the outputs of the previous transformer block and (b) \\(\\{\\pmb{p}_{li}^t\\}_{i=1}^{M^t}\\) and \\(\\{\\pmb{\\mu}_i^t\\}_{i=1}^K\\) are not the outputs of the previous transformer block; see Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.229, + 0.484, + 0.26 + ], + "angle": 0, + "content": "Similarly, we construct the input embeddings to the \\(l\\)-th transformer block of the visual encoder as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.271, + 0.484, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {e} _ {l} ^ {c l s}, \\boldsymbol {x} _ {l 1} ^ {p}, \\dots , \\boldsymbol {x} _ {l M} ^ {p}, \\boldsymbol {\\mu} _ {1} ^ {v}, \\dots , \\boldsymbol {\\mu} _ {K} ^ {v}, \\boldsymbol {p} _ {l 1} ^ {v}, \\dots , \\boldsymbol {p} _ {l M ^ {v}} ^ {v} \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.436 + ], + "angle": 0, + "content": "Here, \\(l = 1,\\dots ,L\\) \\(e_l^{cls}\\) is the class token input to the \\(l\\)-th transformer block, \\(\\{\\pmb{x}_{lm}^p\\}_{m = 1}^M\\) is the set of patch embeddings input to the \\(l\\)-th transformer block with a quantity of \\(M\\), and \\(\\{\\pmb{p}_{li}^v\\}_{i = 1}^{M^v}\\) is the set of learnable visual prompts input to the \\(l\\)-th transformer block with a quantity of \\(M^v\\). It is worth noting that (a) \\(e_l^{cls}\\) and \\(\\{\\pmb{x}_{lm}^p\\}_{m = 1}^M\\) are the outputs of the previous transformer block and (b) \\(\\{\\pmb{p}_{li}^v\\}_{i = 1}^{M^v}\\) and \\(\\{\\pmb{\\mu}_i^v\\}_{i = 1}^K\\) are not the outputs of the previous transformer block; see Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.648 + ], + "angle": 0, + "content": "Undesired interaction masking. There is a modal gap between patch embeddings and visual task-aware pre-context. Allowing patch embeddings to interact with visual task-aware pre-context may weaken the discriminative ability of patch embeddings. To avoid this phenomenon, the information flow from visual task-aware pre-context to patch embeddings is prohibited, which is achieved by exploiting the masked attention mechanism: \\( \\mathbf{A} = \\mathrm{softmax}(Q\\mathbf{K}^T + \\mathcal{M}) \\). Here, \\( \\mathbf{Q} \\) and \\( \\mathbf{K} \\) denote the query and key matrices, with the first dimension specifying the quantity of input embeddings, i.e., \\( N_v = 1 + M + K + M^v \\). They are the linear transformations of input embeddings. \\( \\mathbf{A} \\in \\mathbb{R}^{N_v \\times N_v} \\) denotes the attention matrix. \\( \\mathcal{M} \\in \\mathbb{R}^{N_v \\times N_v} \\) is the customized mask matrix, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {i j} = \\left\\{ \\begin{array}{l l} - \\infty & \\text {i f} 1 < i \\leq 1 + M \\text {a n d} 1 + M < j \\leq 1 + M + K \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{M}_{ij} = -\\infty\\) is equal to \\(A_{ij} = 0\\), meaning that the \\(i\\)-th input embedding does not interact with the \\(j\\)-th peer. In effect, with \\(\\mathcal{M}\\), patch embeddings are made to not interact with visual task-aware pre-context. It is worth noting that (a) \\(\\mathcal{M}\\) is injected into the visual encoder in a layer-wise manner; see Figure 2, and (b) the learnable prompts act as a glue between visual task-aware pre-context and patch embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Self-regulation. Due to the scarcity of task-specific data, overfitting can easily occur. We address this issue by exploiting the strong generation ability exhibited by the well-pretrained encoders, i.e., encouraging the prompted class and visual embeddings to be consistent with their CLIP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.906, + 0.137 + ], + "angle": 0, + "content": "peers. Specifically, let \\(\\{\\boldsymbol{w}_c^{\\mathrm{TAC}}\\}_{c=1}^C\\) denote the class embeddings extracted by using our TAC. The textual consistency is imposed by:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.147, + 0.907, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {t} = \\sum_ {i = 1} ^ {C} | \\boldsymbol {w} _ {c} ^ {\\mathrm {T A C}} - \\boldsymbol {w} _ {c} ^ {\\mathrm {C L I P}} |. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.906, + 0.262 + ], + "angle": 0, + "content": "Data augmentations are applied \\( N^d \\) times on \\( x \\). Then, the augmented images are fed into our visual encoder to extract their embeddings, i.e., \\( \\{f_i^{\\mathrm{TAC}}\\}_{i=1}^{N^d} \\). Correspondingly, the visual consistency is imposed by:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.271, + 0.906, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {v} = \\sum_ {i = 1} ^ {N ^ {d}} | \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}} - \\boldsymbol {f} ^ {\\mathrm {C L I P}} |. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.904, + 0.354 + ], + "angle": 0, + "content": "Here, \\( f^{\\mathrm{CLIP}} \\) denotes the embedding extracted by using the visual encoder of CLIP on \\( x \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.904, + 0.385 + ], + "angle": 0, + "content": "Overall loss. The probability of \\( \\pmb{f}_i^{\\mathrm{TAC}} \\) that belongs to the \\( c \\)-th class can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.392, + 0.906, + 0.432 + ], + "angle": 0, + "content": "\\[\np ^ {\\mathrm {T A C}} \\left(c \\mid \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right) = \\frac {\\exp \\left(\\cos \\left(\\boldsymbol {w} _ {c} ^ {\\mathrm {T A C}} , \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(\\boldsymbol {w} _ {j} ^ {\\mathrm {T A C}}, \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right) / \\tau\\right)}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.899, + 0.455 + ], + "angle": 0, + "content": "Then, the following cross-entropy loss can be constructed:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.465, + 0.906, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} ^ {\\mathrm {T A C}} = - \\sum_ {i = 1} ^ {N ^ {d}} \\sum_ {c = 1} ^ {C} \\mathbb {I} (c = y) \\log p ^ {\\mathrm {T A C}} \\left(c \\mid \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.517, + 0.898, + 0.532 + ], + "angle": 0, + "content": "After combining self-regulation, the overall loss becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.541, + 0.905, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {O v e r a l l}} = \\mathcal {L} _ {\\mathrm {C E}} ^ {\\mathrm {T A C}} + \\lambda \\mathcal {L} ^ {t} + \\beta \\mathcal {L} ^ {v}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.904, + 0.598 + ], + "angle": 0, + "content": "Here, \\(\\lambda\\) and \\(\\beta\\) are hyper-parameters that play a balancing role." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.797 + ], + "angle": 0, + "content": "Complexity. The discussion is three-fold. First, the clustering on \\(\\{\\boldsymbol{w}_c^{\\mathrm{CLIP}}\\}_{c=1}^C\\) and the extraction of original class and visual embeddings are pre-computed and cached. Thus, the extra computational overhead incurred by them is marginal. Second, the textual task-aware pre-context does not incur extra computational overhead since the textual encoder of CLIP only accepts a sufficiently large fixed-number of input embeddings. Third, with the visual task-aware pre-context, the quantity of embeddings input to the visual encoder only increases by a very small amount, i.e., \\(2.5\\%\\). Thus, the extra computational overhead incurred is very limited. In summary, the extra computational overhead incurred by our TAC is marginal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.809, + 0.646, + 0.826 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.688, + 0.85 + ], + "angle": 0, + "content": "4.1. Experiment setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Evaluation and datasets. The evaluation is conducted in the base-to-novel, domain generalization, and cross-dataset transfer settings, and the average accuracy of three" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14749" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.116 + ], + "angle": 0, + "content": "Table 1. Comparative results in the base-to-novel setting. Bold accuracies are the highest. HM denotes the harmonic mean of the base and novel accuracies." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.117, + 0.899, + 0.616 + ], + "angle": 0, + "content": "
(a) Average(b) ImageNet(c) Caltech101(d) OxfordPets
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]69.3474.2271.70CLIP [37]72.4368.1470.22CLIP [37]96.8494.0095.40CLIP [37]91.1797.2694.12
CoOp [61]82.6963.2271.66CoOp [61]76.4767.8871.92CoOp [61]98.0089.8193.73CoOp [61]93.6795.2994.47
Co-CoOp [60]80.4771.6975.83Co-CoOp [60]75.9870.4373.10Co-CoOp [60]97.9693.8195.84Co-CoOp [60]95.2097.6996.43
ProGrad [62]82.4870.7576.16ProGrad [62]77.0266.6671.46ProGrad [62]98.0293.8995.91ProGrad [62]95.0797.6396.33
RPO [29]81.1375.0077.78RPO [29]76.6071.5774.00RPO [29]97.9794.3796.03RPO [29]94.6397.5096.05
MetaPrompt [58]83.3876.0979.57MetaPrompt [58]77.3971.0674.09MetaPrompt [58]98.2894.5896.39MetaPrompt [58]95.7196.9896.34
KAPT [21]78.4170.5274.26KAPT [21]71.1065.2068.02KAPT [21]97.1093.5395.28KAPT [21]93.1396.5394.80
KgCoOp [51]80.7373.6077.00KgCoOp [51]75.8369.9672.78KgCoOp [51]97.7294.3996.03KgCoOp [51]94.6597.7696.18
MaPLe [22]82.2875.1478.55MaPLe [22]76.6670.5473.47MaPLe [22]97.7494.3696.02MaPLe [22]95.4397.7696.58
PromptSRC [23]84.2676.1079.97PromptSRC [23]77.6070.7374.01PromptSRC [23]98.1094.0396.02PromptSRC [23]95.3397.3096.30
TCP [52]84.1375.3679.51TCP [52]77.2769.8773.38TCP [52]98.2394.6796.42TCP [52]94.6797.2095.92
CoPrompt [39]84.0077.2380.48CoPrompt [39]77.6771.2774.33CoPrompt [39]98.2794.9096.55CoPrompt [39]95.6798.1096.87
ArGue [45]83.6978.0780.78ArGue [45]76.9272.0674.41ArGue [45]98.4395.2096.79ArGue [45]95.3697.9596.64
TAC (Ours)85.2477.6081.24TAC (Ours)78.5771.0374.61TAC (Ours)98.5795.2796.89TAC (Ours)95.9398.1797.04
(e) StanfordCars(f) Flowers102(g) Food101(h) FGVCAircraft
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]63.3774.8968.65CLIP [37]72.0877.8074.83CLIP [37]90.1091.2290.66CLIP [37]27.1936.2931.09
CoOp [61]78.1260.4068.13CoOp [61]97.6059.6774.06CoOp [61]88.3382.2685.19CoOp [61]40.4422.3028.75
Co-CoOp [60]70.4973.5972.01Co-CoOp [60]94.8771.7581.71Co-CoOp [60]90.7091.2990.99Co-CoOp [60]33.4123.7127.74
ProGrad [62]77.6868.6372.88ProGrad [62]95.5471.8782.03ProGrad [62]90.3789.5989.98ProGrad [62]40.5427.5732.82
RPO [29]73.8775.5374.69RPO [29]94.1376.6784.50RPO [29]90.3390.8390.58RPO [29]37.3334.2035.70
MetaPrompt [58]75.4374.4374.93MetaPrompt [58]97.5374.5484.50MetaPrompt [58]90.7691.7791.26MetaPrompt [58]39.3837.5938.46
KAPT [21]69.4766.2067.79KAPT [21]95.0071.2081.40KAPT [21]86.1387.0686.59KAPT [21]29.6728.7329.19
KgCoOp [51]71.7675.0473.36KgCoOp [51]95.0074.7383.65KgCoOp [51]90.5091.7091.09KgCoOp [51]36.2133.5534.83
MaPLe [22]72.9474.0073.47MaPLe [22]95.9272.4682.56MaPLe [22]90.7192.0591.38MaPLe [22]37.4435.6136.50
PromptSRC [23]78.2774.9776.58PromptSRC [23]98.0776.5085.95PromptSRC [23]90.6791.5391.10PromptSRC [23]42.7337.8740.15
TCP [52]80.8074.1377.32TCP [52]97.7375.5785.23TCP [52]90.5791.3790.97TCP [52]41.9734.4337.83
CoPrompt [39]76.9774.4075.66CoPrompt [39]97.2776.6085.71CoPrompt [39]90.7392.0791.40CoPrompt [39]40.2039.3339.76
ArGue [45]75.6473.3874.49ArGue [45]98.3475.4185.36ArGue [45]92.3391.9692.14ArGue [45]40.4638.0339.21
TAC (Ours)81.6374.1777.72TAC (Ours)97.9776.8786.15TAC (Ours)90.8791.8791.37TAC (Ours)44.6037.7040.86
(i) SUN397(j) DTD(k) EuroSAT(l) UCF101
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]69.3675.3572.23CLIP [37]53.2459.9056.37CLIP [37]56.4864.0560.03CLIP [37]70.5377.5073.85
CoOp [61]80.6065.8972.51CoOp [61]79.4441.1854.24CoOp [61]92.1954.7468.69CoOp [61]84.6956.0567.46
Co-CoOp [60]79.7476.8678.27Co-CoOp [60]77.0156.0064.85Co-CoOp [60]87.4960.0471.21Co-CoOp [60]82.3373.4577.64
ProGrad [62]81.2674.1777.55ProGrad [62]77.3552.3562.45ProGrad [62]90.1160.8972.67ProGrad [62]84.3374.9479.35
RPO [29]80.6077.8079.18RPO [29]76.7062.1368.61RPO [29]86.6368.9776.79RPO [29]83.6775.4379.34
MetaPrompt [58]82.1079.0180.53MetaPrompt [58]82.5260.1069.55MetaPrompt [58]93.3778.3485.20MetaPrompt [58]84.7078.5681.51
KAPT [21]79.4074.3376.78KAPT [21]75.9758.3065.97KAPT [21]84.8067.5775.21KAPT [21]80.8367.1073.33
KgCoOp [51]80.2976.5378.36KgCoOp [51]77.5554.9964.35KgCoOp [51]85.6464.3473.48KgCoOp [51]82.8976.6779.65
MaPLe [22]80.8278.7079.75MaPLe [22]80.3659.1868.16MaPLe [22]94.0773.2382.35MaPLe [22]83.0078.6680.77
PromptSRC [23]82.6778.4780.52PromptSRC [23]83.3762.9771.75PromptSRC [23]92.9073.9082.32PromptSRC [23]81.7078.8082.74
TCP [52]82.6378.2080.35TCP [52]82.7758.0768.23TCP [52]91.6374.7382.32TCP [52]81.1380.7783.83
CoPrompt [39]82.6380.0381.31CoPrompt [39]83.1364.7372.79CoPrompt [39]94.6078.5785.84CoPrompt [39]86.9079.5783.07
ArGue [45]81.5280.7481.13ArGue [45]81.6066.5573.31ArGue [45]94.4388.2491.23ArGue [45]85.5679.2982.31
TAC (Ours)83.7080.0381.82TAC (Ours)83.3764.2772.58TAC (Ours)94.3782.6088.10TAC (Ours)88.0781.6784.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.629, + 0.908, + 0.655 + ], + "angle": 0, + "content": "Table 2. Comparative results in the cross-dataset transfer setting. Bold accuracies are the highest. Optimization is conducted on ImageNet and evaluation is performed on other datasets." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.657, + 0.898, + 0.776 + ], + "angle": 0, + "content": "
SourceTarget
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101Average
CoOp [61]71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
Co-CoOp [60]71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
MaPLe [22]70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
Bayesian Prompt [10]70.9393.6790.6365.0070.9086.3024.9367.4746.1045.8768.6765.95
PromptSRC [23]71.2793.6090.2565.7070.2586.1523.9067.1046.8745.5068.7565.81
TCP [52]71.4093.9791.2564.6971.2186.6923.4567.1544.3551.4568.7366.29
CoPrompt [39]70.8094.5090.7365.6772.3086.4324.0067.5747.0751.9069.7367.00
TAC (Ours)72.7794.5390.6765.3072.2085.8323.5367.6347.5748.0770.0066.53
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.884 + ], + "angle": 0, + "content": "runs is reported as the final performance. The datasets involved are as follows: ImageNet [9], Caltech101 [30], OxfordPets [36], StanfordCars [25], Flowers102 [35], Food101 [3], FGVCAircraft [44], SUN397 [50], DTD [8], EuroSAT [14], UCF101 [42], ImageNet-A [16], ImageNet-R [15], ImageNet-Sketch [47], and ImageNetV2 [38]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation details. Our TAC is built on top of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.907, + 0.901 + ], + "angle": 0, + "content": "CoOp [61] and PromptSRC [23]. The ViT-B/16 based CLIP [37] is used as the foundation model. The optimizer is SGD, with a learning rate of \\( 2.5 \\times 10^{-3} \\), a weight decay of \\( 5 \\times 10^{-4} \\), and a momentum of 0.9. The cosine annealing is used for learning rate scheduling. The random flip and resized crop are used for data augmentation. The batch size is 8. \\( K = 5 \\), \\( M^t = 4 \\), \\( M^v = 4 \\), \\( N^d = 4 \\), \\( \\lambda = 20 \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14750" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 3. Comparative results in the domain generalization setting. Bold accuracies are the highest. Optimization is conducted on ImageNet and evaluation is performed on datasets with domain shifts." + }, + { + "type": "table", + "bbox": [ + 0.202, + 0.118, + 0.797, + 0.328 + ], + "angle": 0, + "content": "
SourceTarget
ImageNetImageNetV2ImageNet-SketchImageNet-AImageNet-RAverage
CLIP [37]66.7360.8346.1547.7773.9657.17
RPO [29]71.6765.1349.2750.1376.5760.27
LaViP [26]65.9561.6047.2348.9183.9360.41
TCP [52]71.2064.6049.5051.2076.7360.51
CoOp [61]71.5164.2047.9949.7175.2159.28
Co-CoOp [60]71.0264.0748.7550.6376.1859.90
ProGrad [62]72.2464.7347.6149.3974.5859.07
KgCoOp [51]71.2064.1048.9750.6976.7060.11
MaPLe [22]70.7264.0749.1550.9076.9860.26
Bayesian Prompt [10]70.9364.2349.2051.3377.0060.44
PromptSRC [23]71.2764.3549.5550.9077.8060.65
CoPrompt [39]70.8064.2549.4350.5077.5160.42
ArGue [45]71.8465.0249.2551.4776.9660.67
TAC (Ours)72.7765.9750.3051.7378.5061.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.344, + 0.486, + 0.385 + ], + "angle": 0, + "content": "Table 4. Ablation on the textual and visual task-aware precontexts. HM denotes the harmonic mean of the base and novel accuracies." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.386, + 0.46, + 0.464 + ], + "angle": 0, + "content": "
Pre-contextBaseNovelHM
None84.7075.5579.86
Textual84.5776.4880.32
Visual84.6776.0680.13
Textual & visual85.2477.6081.24
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.571 + ], + "angle": 0, + "content": "\\(\\beta = 25\\), and \\(\\tau = 0.07\\). In the base-to-novel setting, the number of epochs is 20 and the scale of random resized crop is (0.4, 1.0). In the domain generalization and cross-dataset transfer settings, the number of epochs is 5 and the scale of random resized crop is (0.8, 1.0). The learning rate is warmed up for one epoch with a rate of \\(1.0 \\times 10^{-5}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.285, + 0.599 + ], + "angle": 0, + "content": "4.2. Comparative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Base-to-novel. In this setting, optimization is conducted on the base set and evaluation is performed on the base and novel sets. Table 1 shows the comparative results. Our TAC achieves the best average base accuracy, the second-best average novel accuracy, and the best harmonic mean. These results demonstrate the generalizability of our task-aware clustering prompting framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Cross-dataset transfer. In this setting, optimization is conducted on ImageNet and evaluation is performed on other datasets. Table 2 shows the comparative results. Our TAC achieves the best accuracy on ImageNet. Also, our TAC yields the second-best average accuracy on other datasets, which is slightly worse than CoPrompt [39]. It is worth noting that our TAC beats CoPrompt in the baseto-novel and domain generalization settings. These results demonstrate the transferability of our task-aware clustering prompting framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Domain generalization. In this setting, optimization is conducted on ImageNet and evaluation is performed on" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.344, + 0.906, + 0.371 + ], + "angle": 0, + "content": "Table 5. Ablation on the number of clustering centers. HM denotes the harmonic mean of the base and novel accuracies." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.372, + 0.884, + 0.472 + ], + "angle": 0, + "content": "
Number of clustering centersBaseNovelHM
K=084.7075.5579.86
K=184.9577.3980.99
K=284.9477.4181.00
K=385.3077.3281.14
K=485.4077.3481.17
K=585.2477.6081.24
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.563 + ], + "angle": 0, + "content": "datasets with domain shifts. Table 3 shows the comparative results. Our TAC achieves the best accuracy on ImageNet and the best average accuracy on the datasets with domain shifts. These results demonstrate the domain generalizability of our task-aware clustering prompting framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.676, + 0.59 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Ablation studies are conducted in the base-to-novel setting. The average base accuracy, the average novel accuracy, and their harmonic mean are reported. More is shown in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Task-aware pre-context. We ablate the effect of the task-aware pre-context by removing the textual pre-context, the visual pre-context, or both. It is worth noting that the other ingredients are kept in this experiment. Table 4 shows the results. Our observations are as follows: (a) both the textual and visual task-aware pre-contexts can improve the harmonic mean and (b) their combination achieves the greatest performance gains. These observations demonstrate the effectiveness of introducing task-aware pre-context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Number of clustering centers. We ablate the effect of the number of clustering centers by setting \\( K \\) to different values. It is worth noting that (a) the maximum value of \\( K \\) is 5 since the dataset EuroSAT [14] only has 5 classes and (b) the other ingredients are kept in this experiment. Table 5 shows the results. Our observations are as follows: (a) \\( K = 0 \\) means that the task-aware pre-context is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14751" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 6. Ablation on the self-regulation. HM denotes the harmonic mean of the base and novel accuracies." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.117, + 0.46, + 0.2 + ], + "angle": 0, + "content": "
Self-regulationBaseNovelHM
LCE TAC85.2974.1879.34
LCE TAC + λLt83.5273.7378.32
LCE TAC + βLv84.4873.8978.83
LCE TAC + λLt + βLv85.2477.6081.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.21, + 0.483, + 0.238 + ], + "angle": 0, + "content": "Table 7. Ablation on the undesired interaction masking. HM denotes the harmonic mean of the base and novel accuracies." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.238, + 0.46, + 0.29 + ], + "angle": 0, + "content": "
Undesired interaction maskingBaseNovelHM
×85.0077.0580.83
✓85.2477.6081.24
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.483, + 0.395 + ], + "angle": 0, + "content": "not exploited, (b) the introduction of task-aware pre-context can improve performance by a noticeable margin, and (c) the biggest performance gains are achieved when \\( K = 5 \\). These observations demonstrate the effectiveness of introducing task-aware pre-context and lead us to set the default value of \\( K \\) to 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Self-regulation. We ablate the effect of the self-regulation by removing the textual consistency, the visual consistency, or both. It is worth noting that the other ingredients are kept in this experiment. Table 6 shows the results. Our observations are as follows: (a) without imposing self-regulation, the accuracy on the novel set is very poor, which also leads to a low harmonic mean, (b) imposing the textual consistency or visual consistency alone brings a noticeable performance drop, and (c) imposing the two consistencies simultaneously leads to a significant performance gain. The reason for these observations might be that the alignment between textual and visual modalities is well preserved by imposing the two consistencies, while it is difficult to achieve this with only one consistency. These observations demonstrate the necessity of self-regulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Undesired interaction masking. We ablate the effect of the undesired interaction masking by removing it. It is worth noting that (a) the mask matrix is inserted into the visual encoder in a layer-wise manner and (b) the other ingredients are kept in this experiment. Table 7 shows the results. With the undesired interaction masking, the harmonic mean is improved by a margin of \\(0.41\\%\\). Considering the performance gain is achieved in an experiment setting involving 11 datasets, this observation demonstrates the effectiveness of undesired interaction masking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Complexity. We evaluate the complexity by comparing our TAC with the competitive counterparts. It is worth noting that (a) PromptSRC [23] inserts learnable prompts into the textual and visual encoders in a layer-wise manner and (b) our TAC inserts the learnable prompts, along with the task-aware pre-contexts, into the textual and visual encoders in the same way. Table 8 shows the results. Since the task-aware pre-context generation is lightweight and the quan-" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.116 + ], + "angle": 0, + "content": "Table 8. Ablation on the complexity. Evaluation is performed on a RTX 3090 GPU." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.117, + 0.882, + 0.214 + ], + "angle": 0, + "content": "
MethodOptimization (ms/image)Inference (ms/image)
CLIP [37]-1.7
CoOp [61]68.13.0
CoCoOp [60]269.0126.6
MaPLe [22]70.21.7
PromptSRC [23]71.53.8
TAC (Ours)73.24.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.22, + 0.905, + 0.247 + ], + "angle": 0, + "content": "Table 9. Ablation on the number of learnable prompts. HM denotes the harmonic mean of the base and novel accuracies." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.247, + 0.882, + 0.321 + ], + "angle": 0, + "content": "
Number of learnable promptsBaseNovelHM
184.1575.6379.66
284.6776.6180.44
485.2477.6081.24
685.0276.8580.73
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.424 + ], + "angle": 0, + "content": "tity of additional input embeddings incurred by the task-aware pre-context is small, the extra computational overhead incurred by our TAC is marginal compared to PromptSRC [23]. Moreover, it is worth noting our TAC outperforms PromptSRC with noticeable performance gains in all experiment settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.909, + 0.575 + ], + "angle": 0, + "content": "Number of learnable prompts. We ablate the effect of the number of learnable prompts by setting \\( M^t \\) and \\( M^v \\) to different values. It is worth noting that (a) \\( M^t \\) and \\( M^v \\) are set to be equal by default and (b) the other ingredients are kept in this experiment. Table 9 shows the results. Our observations are as follows: (a) increasing the number of learnable prompts can improve the harmonic mean and (b) the greatest performance gain is achieved when \\( M^t = 4 \\) and \\( M^v = 4 \\). These observations lead us to set the default value of \\( M^t \\) and \\( M^v \\) to 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.634, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Prompt learning has emerged as a powerful tool for adapting vision-language models. However, due to the scarcity of task-specific data, the resulting prompts often exhibit limited task-awareness. In this work, we propose to increase the task-awareness by introducing task-aware pre-context, which results in a novel task-aware clustering framework for prompting. The key ingredients are three-fold. First, the backbone structure of a downstream task is preserved by task-aware clustering, based on which the task-aware pre-context is generated. Second, the task-awareness of learnable prompts is enhanced by allowing them to interact with task-aware pre-context via the well-pretrained encoders. Finally, the undesired interference of visual task-aware pre-context to patch embeddings is eliminated by masked attention mechanism. Extensive ablation studies confirm the effectiveness of key ingredients. Comparative results in the base-to-novel, domain generalization, and cross-dataset transfer settings show the advantage of our task-aware clustering for prompt learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14752" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.268 + ], + "angle": 0, + "content": "This work was supported by Guangdong Major Project of Basic and Applied Basic Research (2023B0303000016), National Natural Science Foundation of China (62206268, U21A20487, 62372440), Shenzhen Technology Project (JCYJ20240813154723031), Shenzhen High-tech Zone Development Special Plan Innovation Platform Construction Project, the proof of concept center for high precision and high resolution 4D imaging, Guangdong Technology Project (2023TX07Z126), and Yunnan Science & Technology Project (202302AD080008, 202305AF150152)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.278, + 0.188, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.303, + 0.485, + 0.44 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.442, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[2] Shtedritski Aleksandar, Rupprecht Christian, and Vedaldi Andrea. What does clip know about a red circle? visual prompt engineering for vlms. In ICCV, pages 11987-11997, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.483, + 0.539 + ], + "angle": 0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446-461, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[4] Adrian Bulat and Georgios Tzimiropoulos. Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In CVPR, pages 23232-23241, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.596, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[5] Yun-Hao Cao, Kaixiang Ji, Ziyuan Huang, Chuanyang Zheng, Jiajia Liu, Jian Wang, Jingdong Chen, and Ming Yang. Towards better vision-inspired vision-language models. In CVPR, pages 13537-13547, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[6] Delong Chen, Zhao Wu, Fan Liu, Zaiquan Yang, Shaoqiu Zheng, Ying Tan, and Erjin Zhou. Protoclip: Prototypical contrastive language image pretraining. IEEE TNNLS (Early Access), pages 1-15, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[7] Eulrang Cho, Jooyeon Kim, and Hyunwoo J. Kim. Distribution-aware prompt tuning for vision-language models. In ICCV, pages 22004-22013, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[8] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[10] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G. Turrisi da Costa, Cees G.M. Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learning for image-language model generalization. In ICCV, pages 15237-15246, 2023. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.303, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[11] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[12] Bordes Florian, Pang Richard Yuanzhe, Ajay Anurag, Li Alexander C, Bardes Adrien, Petryk Suzanne, Manas Oscar, Lin Zhiqiu, Mahmoud Anas, Jayaraman Bargav, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[13] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE TNNLS (Early Access), pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. J-STARS, 12(7):2217-2226, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[17] Nisha Huang, Yuxin Zhang, Fan Tang, Chongyang Ma, Haibin Huang, Weiming Dong, and Changsheng Xu. Diff-styler: Controllable dual diffusion for text-driven image stylization. IEEE TNNLS (Early Access), pages 1-14, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[18] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, pages 3061-3071, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[19] Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Retrieval-enhanced contrastive vision-text models. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[21] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[22] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[23] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14753" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Jonathan Krause, Michael Stark, Jia Deng, and Fei-Fei Li. 3d object representations for fine-grained categorization. In ICCV, pages 554–561, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[26] Nilakshan Kunananthaseelan, Jing Zhang, and Mehrtash Harandi. Lavip: Language-grounded visual prompting. In AAAI, pages 2840-2848, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Gukyeong Kwon, Zhaowei Cai, Avinash Ravichandran, Erhan Bas, Rahul Bhotika, and Stefano Soatto. Masked vision and language modeling for multi-modal representation learning. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[28] Marc Lafon, Elias Ramzi, Clément Rambour, Nicolas Audebert, and Nicolas Thome. Gallop: Learning global and local prompts for vision-language models. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.275, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[29] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J. Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[30] Fei-Fei Li, Fergus Rob, and Perona Pietro. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshop, pages 178–178, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[31] Yanxin Long, Jianhua Han, Runhui Huang, Hang Xu, Yi Zhu, Chunjing Xu, and Xiaodan Liang. Fine-grained visual-text prompt-driven self-training for open-vocabulary object detection. IEEE TNNLS (Early Access), pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[32] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In CVPR, pages 5206-5215, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[33] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In CVPR, pages 26354-26363, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[34] Metaxas Ioannis Maniadis, Tzimiropoulos Georgios, and Paras Ioannis. Divclust: Controlling diversity in deep clustering. In CVPR, pages 3418-3428, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.637 + ], + "angle": 0, + "content": "[35] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[36] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In \\(CVPR\\), pages 3498-3505, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.79 + ], + "angle": 0, + "content": "[38] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imagenet? In ICML, pages 5389-5400, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[39] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. In ICLR, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Lavoie Samuel, Kirichenko Polina, Ibrahim Mark, Assran Mido, Wilson Andrew Gordon, Courville Aaron, and Ballas Nicolas. Modeling caption diversity in contrastive vision-language pretraining. In ICML, pages 26070-26084, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.149 + ], + "angle": 0, + "content": "[41] Tourani Siddharth, Alwheibi Ahmed, Mahmood Arif, and Khan Muhammad Haris. Pose-guided self-training with two-stage clustering for unsupervised landmark discovery. In CVPR, pages 23041-23051, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.15, + 0.906, + 0.192 + ], + "angle": 0, + "content": "[42] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.194, + 0.906, + 0.236 + ], + "angle": 0, + "content": "[43] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In CVPR, pages 23209-23218, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.238, + 0.906, + 0.292 + ], + "angle": 0, + "content": "[44] Maji Subhransu, Rahtu Esa, Kannala Juho, Blaschko Matthew, and Vedaldi Andrea. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.295, + 0.906, + 0.337 + ], + "angle": 0, + "content": "[45] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argument: Attribute-guided prompt tuning for vision-language models. In CVPR, pages 28578-28587, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.339, + 0.906, + 0.381 + ], + "angle": 0, + "content": "[46] Yujun Tong, Da Li, Dongliang Chang, Tianwei Cao, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Vision-language subspace prompting, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.382, + 0.906, + 0.424 + ], + "angle": 0, + "content": "[47] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.426, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[48] Mengmeng Wang, Jiazheng Xing, Jianbiao Mei, Yong Liu, and Yunliang Jiang. Actionclip: Adapting language-image pretrained models for video action recognition. IEEE TNNLS (Early Access), pages 1-13, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.484, + 0.906, + 0.526 + ], + "angle": 0, + "content": "[49] Zihao Wei, Zixuan Pan, and Andrew Owens. Efficient vision-language pre-training by cluster masking. In CVPR, pages 26815-26825, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.528, + 0.906, + 0.582 + ], + "angle": 0, + "content": "[50] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.584, + 0.906, + 0.627 + ], + "angle": 0, + "content": "[51] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, pages 6757-6767, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.629, + 0.906, + 0.671 + ], + "angle": 0, + "content": "[52] Hantao Yao, Rui Zhang, and Changsheng Xu. Tcp: Textual-based class-aware prompt tuning for visual-language model. In CVPR, pages 23438-23448, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.672, + 0.906, + 0.715 + ], + "angle": 0, + "content": "[53] Lochman Yaroslava, Olsson Carl, and Zach Christopher. Learned trajectory embedding for subspace clustering. In CVPR, pages 19092-19102, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.716, + 0.906, + 0.77 + ], + "angle": 0, + "content": "[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.773, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[55] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.845, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[56] Yunfei Zhang, Xiaoyang Huo, Tianyi Chen, Si Wu, and Hau San Wong. Exploring intra-class variation factors with learnable cluster prompts for semi-supervised image synthesis. In CVPR, pages 7392-7401, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14754" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[57] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In CVPR, pages 28718-28728, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[58] Cairong Zhao, Yubin Wang, Xinyang Jiang, Yifei Shen, Kaitao Song, Dongsheng Li, and Duoqian Miao. Learning domain invariant prompt for vision-language models. IEEE TIP, 33:1348-1360, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[59] Zhaoheng Zheng, Jingmin Wei, Xuefeng Hu, Haidong Zhu, and Ram Nevatia. Large language models are good prompt learners for low-shot image classification. In CVPR, pages 28453-28462, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.305, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.348, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[62] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In CVPR, pages 15659-15669, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[63] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14755" + } + ] +] \ No newline at end of file diff --git a/2025/Task-Aware Clustering for Prompting Vision-Language Models/aabdfb34-0949-4921-9337-b3ad9044174e_origin.pdf b/2025/Task-Aware Clustering for Prompting Vision-Language Models/aabdfb34-0949-4921-9337-b3ad9044174e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9d5d6d057f3b1e207ffc9f00fe6c4ad84ee7a8f0 --- /dev/null +++ b/2025/Task-Aware Clustering for Prompting Vision-Language Models/aabdfb34-0949-4921-9337-b3ad9044174e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9a2a63466f95ed3b24022e1c55a0714558a2971f642eb48dc016cb21d71b1fe +size 2476992 diff --git a/2025/Task-Aware Clustering for Prompting Vision-Language Models/full.md b/2025/Task-Aware Clustering for Prompting Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cd598418a9fc7e6bd04887a785daef93a5f32828 --- /dev/null +++ b/2025/Task-Aware Clustering for Prompting Vision-Language Models/full.md @@ -0,0 +1,311 @@ +# Task-Aware Clustering for Prompting Vision-Language Models + +Fusheng Hao $^{1,2}$ Fengxiang He $^{3}$ Fuxiang Wu $^{1,2}$ Tichao Wang $^{1,2}$ Chengqun Song $^{1,2}$ Jun Cheng $^{1,2*}$ $^{1}$ Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences + $^{2}$ The Chinese University of Hong Kong + +# Abstract + +Prompt learning has attracted widespread attention in adapting vision-language models to downstream tasks. Existing methods largely rely on optimization strategies to ensure the task-awareness of learnable prompts. Due to the scarcity of task-specific data, overfitting is prone to occur. The resulting prompts often do not generalize well or exhibit limited task-awareness. To address this issue, we propose a novel Task-Aware Clustering (TAC) framework for prompting vision-language models, which increases the task-awareness of learnable prompts by introducing task-aware pre-context. The key ingredients are as follows: (a) generating task-aware pre-context based on task-aware clustering that can preserve the backbone structure of a downstream task with only a few clustering centers, (b) enhancing the task-awareness of learnable prompts by enabling them to interact with task-aware pre-context via the well-pretrained encoders, and (c) preventing the visual task-aware pre-context from interfering the interaction between patch embeddings by masked attention mechanism. Extensive experiments are conducted on benchmark datasets, covering the base-to-novel, domain generalization, and cross-dataset transfer settings. Ablation studies validate the effectiveness of key ingredients. Comparative results show the superiority of our TAC over competitive counterparts. The code is available at https://github.com/FushengHao/TAC. + +# 1. Introduction + +Large-scale visual and textual data pairs have been collected and employed to pretrain high-capacity Vision-Language Models (VLMs) [1, 20, 37]. Due to the alignment of the visual and textual modalities, the resulting models have emerged with remarkable generalization ability on various downstream tasks such as zero-shot image classification [37] and open-vocabulary object detection [11]. Despite achieving decent performance, the performance is still + +![](images/269335cecc0f062bd1d8e48c4a5f795fa484d9cc991b18b39b8d350631129afe.jpg) +Figure 1. Motivation of our TAC. We improve the task-awareness of learnable prompts from a new perspective, i.e., introducing task-aware pre-context. + +lower than that of the customized models [61]. Considering the data-hungry nature of VLMs and the scarcity of task-specific data, it is difficult to improve performance through finetuning such models. As an alternative to finetuning, prompting has become as a powerful paradigm for adapting VLMs to downstream tasks [4, 7, 21, 22, 29, 51]. + +Prompting is a technique that adapts VLMs by increasing their understanding of downstream tasks. The current methodology is to provide VLMs with task-relevant context while keeping the model weights frozen. The context is also known as prompts. Many effective prompt templates have been engineered. For example, textual templates like "a photo of a [class]" [37] have been hand-crafted to provide task-relevant context for the textual encoder, and visual templates like "a red circle around an [object]" [2] have been discovered to direct the visual encoder's attention to the [object]. To avoid the need for large-scale trial and error in prompt engineering, context optimization [61] has been proposed, which treats learnable vectors as prompts and ensure their task-awareness by optimization. This strategy has attracted much attention in recent years. For example, maximizing the ground-truth class scores [61] is proposed to ensure the task-awareness of learnable prompts, and self-regulating constraints [23] are further introduced to guide + +the optimization of learnable prompts for better generalization. It is worth noting that the number of labeled images per class used for prompt learning is not sufficient in the setting of context optimization1. Due to the scarcity of task-specific data, the resulting prompts are often overfitted [23] and do not generalize well, meaning that the task-awareness they exhibit is limited. + +Instead of imposing stronger constraints [39, 45, 46], we propose to increase the task-awareness by directly introducing task-aware pre-context; see Figure 1. Considering that the class descriptions of a downstream task are given when encountering the task, the class embeddings extracted by using the well-pretrained textual encoder are good candidates for task-aware pre-context. However, one big challenge arises in this practice, i.e., the number of extra embeddings it incurs is large and equal to that of classes involved. For example, the number is 1000 for the popular dataset ImageNet [9], which is significantly larger than the typical values of the number of embeddings input to the textual and visual encoders, i.e., 77 and 197, respectively. This not only incurs large additional computational overhead, but also exceeds the maximum input limit for the textual encoder. These difficulties inspire us to keep only the backbone structure of class embeddings, which is helpful for reducing the number of embeddings incurred significantly. Motivated by these observations, we explore the prompting strategy based on clustering. + +In this work, we propose a novel Task-Aware Clustering (TAC) framework for prompting VLMs. The central idea of TAC is to exploit task-aware pre-context to improve the task-awareness of learnable prompts. To generate task-aware pre-context, we conduct clustering on the class embeddings of a downstream task and perform linear transformation on the clustering centers to yield task-aware pre-context, which leads to a task-aware pre-context generalization module that can keep the backbone structure of the downstream task while being lightweight. Then, the task-aware pre-context is combined with the learnable prompts and their interaction through the well-pretrained encoders is exploited to reinforce the task-awareness of learnable prompts, which leads to a task-awareness reinforcement module. Further, a mask matrix is developed and injected into the attention mechanism to eliminate the undesired interference of visual task-aware pre-context to patch embeddings induced by the gap between textual and visual modalities, which leads to undesired interaction masking module. To suppress the overfitting induced by the scarcity of task-specific data, the prompted class and visual embeddings are encouraged to be consistent with their CLIP [37] peers, which leads to a self-regulation module. Our TAC maintains simplicity in design and improves the task-awareness of the resulting prompts, thus yielding competitive results. + +In summary, our task-aware clustering framework for prompting has the following main contributions: + +- We propose to increase the task-awareness of learnable prompts from a new perspective, i.e., introducing task-aware pre-context. +- We propose to generate task-aware pre-context based on task-aware clustering that can preserve the backbone structure of a downstream task. +- We exploit the well-pretrained encoders to reinforce the task-awareness of learnable prompts while masking out the undesired interaction. +- We demonstrate the superiority of our TAC in the baseto-novel, domain generalization, and cross-dataset transfer settings. + +# 2. Related Works + +Vision-Language Models (VLMs) integrate the textual and visual modalities, bridging the gap between them through alignment. Such models are usually pretrained on large-scale visual and textual data pairs [6, 20, 37]. They can be broadly classified into four categories according to their training paradigms [12]. The first category employs contrastive-based training [19, 40], which emphasizes reducing the distance between positive pairs and increasing the distance between negative pairs. The second category utilizes masking-based training [27, 49], which emphasizes the reconstruction of missing patches/words by exploiting the unmasked caption/image. The third category leverages generative-based training [54, 55], which emphasizes the generation of entire images or captions based on conditional captions or images. The fourth category employs LLM-based training [5, 63], which emphasizes the usage of large language models to extract supervisory information. Since the textual and visual modalities are aligned in a shared embedding space, VLMs exhibits a strong ability to understand open-vocabulary concepts, which facilitates various downstream tasks. Although VLMs have achieved decent performance, their accuracies are still far inferior to those of customized models [61]. Considering the data-hungry nature of VLMs and the scarcity of task-specific data, finetuning such models to improve their performance faces significant challenges. As an alternative to finetuning, prompting [37, 61] has attracted much attention in adapting VLMs. Orthogonal directions include the utilization of unlabeled images [43] and historical test data [57]. + +Prompting adapts VLMs by providing them with contextual information or task-aware prompts while freezing their model weights, which has achieved good performance in many downstream tasks [13, 17, 24, 31, 48]. Existing methods can be broadly classified into two categories according to whether domain expertise is required. The first category relies on hand-crafted prompt templates, which emphasizes + +the reliance on domain expertise. CLIP [37] crafts the textual template, i.e., "a photo of a [class]". CirclePrompt [2] engineers the visual template, i.e., "a red circle around an [object]". LaViP [26] grounds visual prompt with language. ArGue [45] crafts attribute-guided prompts with the help of large language models. The second category relies on prompt learning [61], which emphasizes the introduction of learnable prompts and the optimization of these prompts to ensure their task-awareness. PromptSRC [23] guides the optimization of learnable prompts with self-regulating constraints. GalLoP [28] learns global-local prompts to capture multi-scale visual information. CoPrompt [39] improves the generation ability of learnable prompts by encouraging the prediction consistency between the pretrained and trainable models. TCP [52] introduces class-aware prompts that are produced by performing textual knowledge embedding on class embeddings. In this work, we increase the task-awareness by introducing task-aware pre-context, which is different from the current practice that imposes extra constraints. + +Clustering is a technique that can group similar data points into clusters or groups, in which K-Means is a classic example. Integrating clustering may improve the capabilities of deep models in capturing backbone structures, and their combination has attracted a lot of attention. DivClust [34] introduces a diversity controlling loss to ensure the desired degree of diversity for the produced multiple clusters. LCPGAN [56] associates intra-class clusters with learnable visual concepts to capture the class-wise and intra-class information for image synthesis. ZeroShotULD [41] clusters random pixel locations with nearest neighbor searching for unsupervised landmarks discovery. SPC [18] replaces the deterministic prediction with semantic clustering for domain generalized semantic segmentation. MoDE [33] decides clusters for data experts by a two-step clustering on captions and uses fine-grained clusters to coarsely represent data experts. LTE [53] clusters multiple observed motions by establishing the mapping from trajectories to trajectory embeddings. In this work, we perform clustering to preserve the backbone structure of a downstream task represented by class embeddings. + +# 3. Method + +In this section, we begin by reviewing the preliminaries and then detail our Task-Aware Clustering (TAC) framework for prompting vision-language models. + +# 3.1. Preliminaries + +CLIP. CLIP [37] includes a textual encoder and a visual encoder, and is pretrained on around 400 million visual and textual data pairs with the contrastive-based training strategy. With the hand-crafted prompt templates like "a photo + +of a [class]", it emerges strong zero-shot prediction capabilities. Although CLIP has achieved good performance in the zero-shot image recognition task, its accuracies are still lower than those of the customized models [61]. Due to the data-hungry nature of CLIP and the scarcity of task-specific data, finetuning CLIP to improve its performance on downstream tasks is particularly challenging. Considering that prompt engineering relies on domain expertise and needs large-scale trial and error, prompt learning [61] has been proposed, in which learnable vectors are treated as prompts and their task-awareness is ensured through optimization. For ease of comparison with existing prompt learning approaches, we follow the current practice and choose CLIP as the foundation model. + +Prompt learning. CoOp [61] is a typical approach for adapting CLIP in the context of prompt learning. We take a $C$ -class image classification task as an example to show how learnable prompts are optimized in CoOp. First, a set of learnable vectors with a quantity of $M^t$ is introduced and treated as textual prompts, i.e., $\{\pmb{p}_m^t\}_{m=1}^{M^t}$ . Second, these prompts are combined with the word embeddings that describe the task, which leads to a learnable version of task descriptions, i.e., $\{\pmb{p}_1^t,\dots,\pmb{p}_{M^t}^t,\pmb{c}_i\}_{i=1}^C$ , where $\pmb{c}_i$ is the word embedding(s) of the $i$ -th class. Third, the learnable task descriptions are fed into the textual encoder to yield their embeddings $\{\pmb{w}_c^{\mathrm{CoOp}}\}_{c=1}^C$ , while an image $\pmb{x}$ is fed into the visual encoder to generate its embedding $\pmb{f}$ . Fourth, the probability of $\pmb{x}$ that belongs to the $c$ -th class can be calculated as: + +$$ +p ^ {\mathrm {C o O p}} (c | \boldsymbol {x}) = \frac {\exp \left(\cos \left(\boldsymbol {w} _ {c} ^ {\mathrm {C o O p}} , \boldsymbol {f}\right) / \tau\right)}{\sum_ {i = 1} ^ {C} \exp \left(\cos \left(\boldsymbol {w} _ {i} ^ {\mathrm {C o O p}}, \boldsymbol {f}\right) / \tau\right)}, \tag {1} +$$ + +where $\tau$ is the temperature and $\cos (\cdot ,\cdot)$ denotes the cosine similarity. Finally, the task-awareness of learnable textual prompts is ensured by minimizing the cross-entropy loss: + +$$ +\mathcal {L} _ {\mathrm {C E}} ^ {\mathrm {C o O p}} = - \sum_ {c = 1} ^ {C} \mathbb {I} (c = y) \log p ^ {\mathrm {C o O p}} (c | \boldsymbol {x}), \tag {2} +$$ + +where $y$ is the ground-truth label of $\pmb{x}$ and $\mathbb{I}(\cdot)$ denotes the indicator function. It is worth noting that (a) only the learnable prompts need to be optimized and (b) the quantity of labeled images that can be used in each class is insufficient in the setting of prompt learning, where the typical value is 16. + +Afterwards, additional constraints [23, 32, 39, 59] have been imposed to improve the task-awareness of learnable prompts. Different from the current practice, we increase the task-awareness from a new perspective, i.e., introducing task-aware pre-context. + +# 3.2. Task-aware Clustering for Prompting + +The pipeline of our TAC is shown in Figure 2 and the key ingredients are detailed below. + +![](images/9fccb3d167d80bb3593a66d3b54553a5721cf609eb020d9dd6c18d3f5edae6e8.jpg) +Figure 2. Pipeline of our TAC. First, the textual and visual task-aware pre-contexts are generated based on the backbone structure of a downstream task. Second, the task-aware pre-context is combined with the learnable prompts and their interaction through the well-pretrained encoders is exploited to reinforce the task-awareness of learnable prompts. Third, the undesired interaction is masked out with a customized mask matrix. Further, the textual and visual consistencies are imposed to alleviate the overfitting issue. + +Task-aware pre-context generation. The class embeddings extracted by using the textual encoder of CLIP are task-aware and good candidates for the desired pre-context. Considering that the number of such embeddings is equal to that of classes involved and this number is large, applying this strategy to practice is challenging. The first challenge is that the number of extra input embeddings incurred exceeds the maximum input limit for the textual encoder, which would make the textual encoder unusable. The second challenge is that the extra computational overhead incurred is very high due to the large number of additional embeddings input into the visual encoder. + +We address these challenges by yielding textual and visual task-aware pre-contexts based on clustering that can preserve the backbone structure of class embeddings. A significant additional benefit of this practice is that the number of extra embeddings incurred can be significantly reduced. Specifically, let $\{\boldsymbol{w}_c^{\mathrm{CLIP}}\}_{c=1}^C$ denote the class embeddings extracted with the textual encoder of CLIP. Then, K-Means is conducted on $\{\boldsymbol{w}_c^{\mathrm{CLIP}}\}_{c=1}^C$ , which yields a set of clustering centers with a quantity of $K$ , i.e., $\{\boldsymbol{\mu}_i\}_{i=1}^K$ . Further, the + +textual and visual task-aware pre-contexts are generated as: + +$$ +\boldsymbol {\mu} _ {i} ^ {t} = \boldsymbol {W} ^ {t} \boldsymbol {\mu} _ {i}, \tag {3} +$$ + +$$ +\boldsymbol {\mu} _ {i} ^ {v} = \boldsymbol {W} ^ {v} \boldsymbol {\mu} _ {i}, \tag {4} +$$ + +where $\boldsymbol{W}^t$ and $\boldsymbol{W}^v$ are the textual and visual linear transformations. It is worth noting that (a) the dimensions of $\boldsymbol{\mu}_i^t$ and $\boldsymbol{\mu}_i^v$ are set to be the internal dimensions of the textual and visual encoders, correspondingly, and (b) since only the weights of linear transformations needs to be optimized, our task-aware pre-context generation module is lightweight. + +Task-awareness reinforcement. After obtaining the textual and visual task-aware pre-contexts, the task-awareness of learnable prompts is reinforced by enabling them to interact with the corresponding pre-contexts via the well-pretrained textual and visual encoders. Specifically, let $L$ denote the quantity of transformer blocks contained in the visual/textual encoder. We construct the input embeddings to the $l$ -th transformer block of the textual encoder as: + +$$ +\left\{\boldsymbol {t} _ {l} ^ {s o s}, \boldsymbol {\mu} _ {1} ^ {t}, \dots , \boldsymbol {\mu} _ {K} ^ {t}, \boldsymbol {p} _ {l 1} ^ {t}, \dots , \boldsymbol {p} _ {l M ^ {t}} ^ {t}, \boldsymbol {c} _ {l i}, \boldsymbol {t} _ {l} ^ {e o s} \right\}. \tag {5} +$$ + +Here, $l = 1,\dots ,L$ , $\pmb{t}_l^{sos}$ is the start token input to the $l$ -th transformer block, $\pmb{t}_l^{eos}$ is the end token input to the $l$ -th transformer block, $\{\pmb{p}_{li}^t\}_{i=1}^{M^t}$ is the set of learnable textual prompts input to the $l$ -th transformer block with a quantity of $M^t$ , and $\pmb{c}_{li}$ is the word embedding(s) input to the $l$ -th transformer block. It is worth noting that (a) $\pmb{t}_l^{sos}$ , $\pmb{t}_l^{eos}$ , and $\pmb{c}_{li}$ are the outputs of the previous transformer block and (b) $\{\pmb{p}_{li}^t\}_{i=1}^{M^t}$ and $\{\pmb{\mu}_i^t\}_{i=1}^K$ are not the outputs of the previous transformer block; see Figure 2. + +Similarly, we construct the input embeddings to the $l$ -th transformer block of the visual encoder as: + +$$ +\left\{\boldsymbol {e} _ {l} ^ {c l s}, \boldsymbol {x} _ {l 1} ^ {p}, \dots , \boldsymbol {x} _ {l M} ^ {p}, \boldsymbol {\mu} _ {1} ^ {v}, \dots , \boldsymbol {\mu} _ {K} ^ {v}, \boldsymbol {p} _ {l 1} ^ {v}, \dots , \boldsymbol {p} _ {l M ^ {v}} ^ {v} \right\}. \tag {6} +$$ + +Here, $l = 1,\dots ,L$ $e_l^{cls}$ is the class token input to the $l$ -th transformer block, $\{\pmb{x}_{lm}^p\}_{m = 1}^M$ is the set of patch embeddings input to the $l$ -th transformer block with a quantity of $M$ , and $\{\pmb{p}_{li}^v\}_{i = 1}^{M^v}$ is the set of learnable visual prompts input to the $l$ -th transformer block with a quantity of $M^v$ . It is worth noting that (a) $e_l^{cls}$ and $\{\pmb{x}_{lm}^p\}_{m = 1}^M$ are the outputs of the previous transformer block and (b) $\{\pmb{p}_{li}^v\}_{i = 1}^{M^v}$ and $\{\pmb{\mu}_i^v\}_{i = 1}^K$ are not the outputs of the previous transformer block; see Figure 2. + +Undesired interaction masking. There is a modal gap between patch embeddings and visual task-aware pre-context. Allowing patch embeddings to interact with visual task-aware pre-context may weaken the discriminative ability of patch embeddings. To avoid this phenomenon, the information flow from visual task-aware pre-context to patch embeddings is prohibited, which is achieved by exploiting the masked attention mechanism: $\mathbf{A} = \mathrm{softmax}(Q\mathbf{K}^T + \mathcal{M})$ . Here, $\mathbf{Q}$ and $\mathbf{K}$ denote the query and key matrices, with the first dimension specifying the quantity of input embeddings, i.e., $N_v = 1 + M + K + M^v$ . They are the linear transformations of input embeddings. $\mathbf{A} \in \mathbb{R}^{N_v \times N_v}$ denotes the attention matrix. $\mathcal{M} \in \mathbb{R}^{N_v \times N_v}$ is the customized mask matrix, which is defined as: + +$$ +\mathcal {M} _ {i j} = \left\{ \begin{array}{l l} - \infty & \text {i f} 1 < i \leq 1 + M \text {a n d} 1 + M < j \leq 1 + M + K \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {7} +$$ + +Here, $\mathcal{M}_{ij} = -\infty$ is equal to $A_{ij} = 0$ , meaning that the $i$ -th input embedding does not interact with the $j$ -th peer. In effect, with $\mathcal{M}$ , patch embeddings are made to not interact with visual task-aware pre-context. It is worth noting that (a) $\mathcal{M}$ is injected into the visual encoder in a layer-wise manner; see Figure 2, and (b) the learnable prompts act as a glue between visual task-aware pre-context and patch embeddings. + +Self-regulation. Due to the scarcity of task-specific data, overfitting can easily occur. We address this issue by exploiting the strong generation ability exhibited by the well-pretrained encoders, i.e., encouraging the prompted class and visual embeddings to be consistent with their CLIP + +peers. Specifically, let $\{\boldsymbol{w}_c^{\mathrm{TAC}}\}_{c=1}^C$ denote the class embeddings extracted by using our TAC. The textual consistency is imposed by: + +$$ +\mathcal {L} ^ {t} = \sum_ {i = 1} ^ {C} | \boldsymbol {w} _ {c} ^ {\mathrm {T A C}} - \boldsymbol {w} _ {c} ^ {\mathrm {C L I P}} |. \tag {8} +$$ + +Data augmentations are applied $N^d$ times on $x$ . Then, the augmented images are fed into our visual encoder to extract their embeddings, i.e., $\{f_i^{\mathrm{TAC}}\}_{i=1}^{N^d}$ . Correspondingly, the visual consistency is imposed by: + +$$ +\mathcal {L} ^ {v} = \sum_ {i = 1} ^ {N ^ {d}} | \boldsymbol {f} _ {i} ^ {\mathrm {T A C}} - \boldsymbol {f} ^ {\mathrm {C L I P}} |. \tag {9} +$$ + +Here, $f^{\mathrm{CLIP}}$ denotes the embedding extracted by using the visual encoder of CLIP on $x$ . + +Overall loss. The probability of $\pmb{f}_i^{\mathrm{TAC}}$ that belongs to the $c$ -th class can be calculated as: + +$$ +p ^ {\mathrm {T A C}} \left(c \mid \boldsymbol {f} _ {i} ^ {\mathrm {T A C}}\right) = \frac {\exp \left(\cos \left(\boldsymbol {w} _ {c} ^ {\mathrm {T A C}} , \boldsymbol {f} _ {i} ^ {\mathrm {T A C}}\right) / \tau\right)}{\sum_ {j = 1} ^ {C} \exp \left(\cos \left(\boldsymbol {w} _ {j} ^ {\mathrm {T A C}}, \boldsymbol {f} _ {i} ^ {\mathrm {T A C}}\right) / \tau\right)}. \tag {10} +$$ + +Then, the following cross-entropy loss can be constructed: + +$$ +\mathcal {L} _ {\mathrm {C E}} ^ {\mathrm {T A C}} = - \sum_ {i = 1} ^ {N ^ {d}} \sum_ {c = 1} ^ {C} \mathbb {I} (c = y) \log p ^ {\mathrm {T A C}} \left(c \mid \boldsymbol {f} _ {i} ^ {\mathrm {T A C}}\right). \tag {11} +$$ + +After combining self-regulation, the overall loss becomes: + +$$ +\mathcal {L} ^ {\text {O v e r a l l}} = \mathcal {L} _ {\mathrm {C E}} ^ {\mathrm {T A C}} + \lambda \mathcal {L} ^ {t} + \beta \mathcal {L} ^ {v}. \tag {12} +$$ + +Here, $\lambda$ and $\beta$ are hyper-parameters that play a balancing role. + +Complexity. The discussion is three-fold. First, the clustering on $\{\boldsymbol{w}_c^{\mathrm{CLIP}}\}_{c=1}^C$ and the extraction of original class and visual embeddings are pre-computed and cached. Thus, the extra computational overhead incurred by them is marginal. Second, the textual task-aware pre-context does not incur extra computational overhead since the textual encoder of CLIP only accepts a sufficiently large fixed-number of input embeddings. Third, with the visual task-aware pre-context, the quantity of embeddings input to the visual encoder only increases by a very small amount, i.e., $2.5\%$ . Thus, the extra computational overhead incurred is very limited. In summary, the extra computational overhead incurred by our TAC is marginal. + +# 4. Experiments + +# 4.1. Experiment setup + +Evaluation and datasets. The evaluation is conducted in the base-to-novel, domain generalization, and cross-dataset transfer settings, and the average accuracy of three + +Table 1. Comparative results in the base-to-novel setting. Bold accuracies are the highest. HM denotes the harmonic mean of the base and novel accuracies. + +
(a) Average(b) ImageNet(c) Caltech101(d) OxfordPets
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]69.3474.2271.70CLIP [37]72.4368.1470.22CLIP [37]96.8494.0095.40CLIP [37]91.1797.2694.12
CoOp [61]82.6963.2271.66CoOp [61]76.4767.8871.92CoOp [61]98.0089.8193.73CoOp [61]93.6795.2994.47
Co-CoOp [60]80.4771.6975.83Co-CoOp [60]75.9870.4373.10Co-CoOp [60]97.9693.8195.84Co-CoOp [60]95.2097.6996.43
ProGrad [62]82.4870.7576.16ProGrad [62]77.0266.6671.46ProGrad [62]98.0293.8995.91ProGrad [62]95.0797.6396.33
RPO [29]81.1375.0077.78RPO [29]76.6071.5774.00RPO [29]97.9794.3796.03RPO [29]94.6397.5096.05
MetaPrompt [58]83.3876.0979.57MetaPrompt [58]77.3971.0674.09MetaPrompt [58]98.2894.5896.39MetaPrompt [58]95.7196.9896.34
KAPT [21]78.4170.5274.26KAPT [21]71.1065.2068.02KAPT [21]97.1093.5395.28KAPT [21]93.1396.5394.80
KgCoOp [51]80.7373.6077.00KgCoOp [51]75.8369.9672.78KgCoOp [51]97.7294.3996.03KgCoOp [51]94.6597.7696.18
MaPLe [22]82.2875.1478.55MaPLe [22]76.6670.5473.47MaPLe [22]97.7494.3696.02MaPLe [22]95.4397.7696.58
PromptSRC [23]84.2676.1079.97PromptSRC [23]77.6070.7374.01PromptSRC [23]98.1094.0396.02PromptSRC [23]95.3397.3096.30
TCP [52]84.1375.3679.51TCP [52]77.2769.8773.38TCP [52]98.2394.6796.42TCP [52]94.6797.2095.92
CoPrompt [39]84.0077.2380.48CoPrompt [39]77.6771.2774.33CoPrompt [39]98.2794.9096.55CoPrompt [39]95.6798.1096.87
ArGue [45]83.6978.0780.78ArGue [45]76.9272.0674.41ArGue [45]98.4395.2096.79ArGue [45]95.3697.9596.64
TAC (Ours)85.2477.6081.24TAC (Ours)78.5771.0374.61TAC (Ours)98.5795.2796.89TAC (Ours)95.9398.1797.04
(e) StanfordCars(f) Flowers102(g) Food101(h) FGVCAircraft
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]63.3774.8968.65CLIP [37]72.0877.8074.83CLIP [37]90.1091.2290.66CLIP [37]27.1936.2931.09
CoOp [61]78.1260.4068.13CoOp [61]97.6059.6774.06CoOp [61]88.3382.2685.19CoOp [61]40.4422.3028.75
Co-CoOp [60]70.4973.5972.01Co-CoOp [60]94.8771.7581.71Co-CoOp [60]90.7091.2990.99Co-CoOp [60]33.4123.7127.74
ProGrad [62]77.6868.6372.88ProGrad [62]95.5471.8782.03ProGrad [62]90.3789.5989.98ProGrad [62]40.5427.5732.82
RPO [29]73.8775.5374.69RPO [29]94.1376.6784.50RPO [29]90.3390.8390.58RPO [29]37.3334.2035.70
MetaPrompt [58]75.4374.4374.93MetaPrompt [58]97.5374.5484.50MetaPrompt [58]90.7691.7791.26MetaPrompt [58]39.3837.5938.46
KAPT [21]69.4766.2067.79KAPT [21]95.0071.2081.40KAPT [21]86.1387.0686.59KAPT [21]29.6728.7329.19
KgCoOp [51]71.7675.0473.36KgCoOp [51]95.0074.7383.65KgCoOp [51]90.5091.7091.09KgCoOp [51]36.2133.5534.83
MaPLe [22]72.9474.0073.47MaPLe [22]95.9272.4682.56MaPLe [22]90.7192.0591.38MaPLe [22]37.4435.6136.50
PromptSRC [23]78.2774.9776.58PromptSRC [23]98.0776.5085.95PromptSRC [23]90.6791.5391.10PromptSRC [23]42.7337.8740.15
TCP [52]80.8074.1377.32TCP [52]97.7375.5785.23TCP [52]90.5791.3790.97TCP [52]41.9734.4337.83
CoPrompt [39]76.9774.4075.66CoPrompt [39]97.2776.6085.71CoPrompt [39]90.7392.0791.40CoPrompt [39]40.2039.3339.76
ArGue [45]75.6473.3874.49ArGue [45]98.3475.4185.36ArGue [45]92.3391.9692.14ArGue [45]40.4638.0339.21
TAC (Ours)81.6374.1777.72TAC (Ours)97.9776.8786.15TAC (Ours)90.8791.8791.37TAC (Ours)44.6037.7040.86
(i) SUN397(j) DTD(k) EuroSAT(l) UCF101
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]69.3675.3572.23CLIP [37]53.2459.9056.37CLIP [37]56.4864.0560.03CLIP [37]70.5377.5073.85
CoOp [61]80.6065.8972.51CoOp [61]79.4441.1854.24CoOp [61]92.1954.7468.69CoOp [61]84.6956.0567.46
Co-CoOp [60]79.7476.8678.27Co-CoOp [60]77.0156.0064.85Co-CoOp [60]87.4960.0471.21Co-CoOp [60]82.3373.4577.64
ProGrad [62]81.2674.1777.55ProGrad [62]77.3552.3562.45ProGrad [62]90.1160.8972.67ProGrad [62]84.3374.9479.35
RPO [29]80.6077.8079.18RPO [29]76.7062.1368.61RPO [29]86.6368.9776.79RPO [29]83.6775.4379.34
MetaPrompt [58]82.1079.0180.53MetaPrompt [58]82.5260.1069.55MetaPrompt [58]93.3778.3485.20MetaPrompt [58]84.7078.5681.51
KAPT [21]79.4074.3376.78KAPT [21]75.9758.3065.97KAPT [21]84.8067.5775.21KAPT [21]80.8367.1073.33
KgCoOp [51]80.2976.5378.36KgCoOp [51]77.5554.9964.35KgCoOp [51]85.6464.3473.48KgCoOp [51]82.8976.6779.65
MaPLe [22]80.8278.7079.75MaPLe [22]80.3659.1868.16MaPLe [22]94.0773.2382.35MaPLe [22]83.0078.6680.77
PromptSRC [23]82.6778.4780.52PromptSRC [23]83.3762.9771.75PromptSRC [23]92.9073.9082.32PromptSRC [23]81.7078.8082.74
TCP [52]82.6378.2080.35TCP [52]82.7758.0768.23TCP [52]91.6374.7382.32TCP [52]81.1380.7783.83
CoPrompt [39]82.6380.0381.31CoPrompt [39]83.1364.7372.79CoPrompt [39]94.6078.5785.84CoPrompt [39]86.9079.5783.07
ArGue [45]81.5280.7481.13ArGue [45]81.6066.5573.31ArGue [45]94.4388.2491.23ArGue [45]85.5679.2982.31
TAC (Ours)83.7080.0381.82TAC (Ours)83.3764.2772.58TAC (Ours)94.3782.6088.10TAC (Ours)88.0781.6784.75
+ +Table 2. Comparative results in the cross-dataset transfer setting. Bold accuracies are the highest. Optimization is conducted on ImageNet and evaluation is performed on other datasets. + +
SourceTarget
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101Average
CoOp [61]71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
Co-CoOp [60]71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
MaPLe [22]70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
Bayesian Prompt [10]70.9393.6790.6365.0070.9086.3024.9367.4746.1045.8768.6765.95
PromptSRC [23]71.2793.6090.2565.7070.2586.1523.9067.1046.8745.5068.7565.81
TCP [52]71.4093.9791.2564.6971.2186.6923.4567.1544.3551.4568.7366.29
CoPrompt [39]70.8094.5090.7365.6772.3086.4324.0067.5747.0751.9069.7367.00
TAC (Ours)72.7794.5390.6765.3072.2085.8323.5367.6347.5748.0770.0066.53
+ +runs is reported as the final performance. The datasets involved are as follows: ImageNet [9], Caltech101 [30], OxfordPets [36], StanfordCars [25], Flowers102 [35], Food101 [3], FGVCAircraft [44], SUN397 [50], DTD [8], EuroSAT [14], UCF101 [42], ImageNet-A [16], ImageNet-R [15], ImageNet-Sketch [47], and ImageNetV2 [38]. + +Implementation details. Our TAC is built on top of + +CoOp [61] and PromptSRC [23]. The ViT-B/16 based CLIP [37] is used as the foundation model. The optimizer is SGD, with a learning rate of $2.5 \times 10^{-3}$ , a weight decay of $5 \times 10^{-4}$ , and a momentum of 0.9. The cosine annealing is used for learning rate scheduling. The random flip and resized crop are used for data augmentation. The batch size is 8. $K = 5$ , $M^t = 4$ , $M^v = 4$ , $N^d = 4$ , $\lambda = 20$ + +Table 3. Comparative results in the domain generalization setting. Bold accuracies are the highest. Optimization is conducted on ImageNet and evaluation is performed on datasets with domain shifts. + +
SourceTarget
ImageNetImageNetV2ImageNet-SketchImageNet-AImageNet-RAverage
CLIP [37]66.7360.8346.1547.7773.9657.17
RPO [29]71.6765.1349.2750.1376.5760.27
LaViP [26]65.9561.6047.2348.9183.9360.41
TCP [52]71.2064.6049.5051.2076.7360.51
CoOp [61]71.5164.2047.9949.7175.2159.28
Co-CoOp [60]71.0264.0748.7550.6376.1859.90
ProGrad [62]72.2464.7347.6149.3974.5859.07
KgCoOp [51]71.2064.1048.9750.6976.7060.11
MaPLe [22]70.7264.0749.1550.9076.9860.26
Bayesian Prompt [10]70.9364.2349.2051.3377.0060.44
PromptSRC [23]71.2764.3549.5550.9077.8060.65
CoPrompt [39]70.8064.2549.4350.5077.5160.42
ArGue [45]71.8465.0249.2551.4776.9660.67
TAC (Ours)72.7765.9750.3051.7378.5061.63
+ +Table 4. Ablation on the textual and visual task-aware precontexts. HM denotes the harmonic mean of the base and novel accuracies. + +
Pre-contextBaseNovelHM
None84.7075.5579.86
Textual84.5776.4880.32
Visual84.6776.0680.13
Textual & visual85.2477.6081.24
+ +$\beta = 25$ , and $\tau = 0.07$ . In the base-to-novel setting, the number of epochs is 20 and the scale of random resized crop is (0.4, 1.0). In the domain generalization and cross-dataset transfer settings, the number of epochs is 5 and the scale of random resized crop is (0.8, 1.0). The learning rate is warmed up for one epoch with a rate of $1.0 \times 10^{-5}$ . + +# 4.2. Comparative results + +Base-to-novel. In this setting, optimization is conducted on the base set and evaluation is performed on the base and novel sets. Table 1 shows the comparative results. Our TAC achieves the best average base accuracy, the second-best average novel accuracy, and the best harmonic mean. These results demonstrate the generalizability of our task-aware clustering prompting framework. + +Cross-dataset transfer. In this setting, optimization is conducted on ImageNet and evaluation is performed on other datasets. Table 2 shows the comparative results. Our TAC achieves the best accuracy on ImageNet. Also, our TAC yields the second-best average accuracy on other datasets, which is slightly worse than CoPrompt [39]. It is worth noting that our TAC beats CoPrompt in the baseto-novel and domain generalization settings. These results demonstrate the transferability of our task-aware clustering prompting framework. + +Domain generalization. In this setting, optimization is conducted on ImageNet and evaluation is performed on + +Table 5. Ablation on the number of clustering centers. HM denotes the harmonic mean of the base and novel accuracies. + +
Number of clustering centersBaseNovelHM
K=084.7075.5579.86
K=184.9577.3980.99
K=284.9477.4181.00
K=385.3077.3281.14
K=485.4077.3481.17
K=585.2477.6081.24
+ +datasets with domain shifts. Table 3 shows the comparative results. Our TAC achieves the best accuracy on ImageNet and the best average accuracy on the datasets with domain shifts. These results demonstrate the domain generalizability of our task-aware clustering prompting framework. + +# 4.3. Ablation studies + +Ablation studies are conducted in the base-to-novel setting. The average base accuracy, the average novel accuracy, and their harmonic mean are reported. More is shown in the appendix. + +Task-aware pre-context. We ablate the effect of the task-aware pre-context by removing the textual pre-context, the visual pre-context, or both. It is worth noting that the other ingredients are kept in this experiment. Table 4 shows the results. Our observations are as follows: (a) both the textual and visual task-aware pre-contexts can improve the harmonic mean and (b) their combination achieves the greatest performance gains. These observations demonstrate the effectiveness of introducing task-aware pre-context. + +Number of clustering centers. We ablate the effect of the number of clustering centers by setting $K$ to different values. It is worth noting that (a) the maximum value of $K$ is 5 since the dataset EuroSAT [14] only has 5 classes and (b) the other ingredients are kept in this experiment. Table 5 shows the results. Our observations are as follows: (a) $K = 0$ means that the task-aware pre-context is + +Table 6. Ablation on the self-regulation. HM denotes the harmonic mean of the base and novel accuracies. + +
Self-regulationBaseNovelHM
LCE TAC85.2974.1879.34
LCE TAC + λLt83.5273.7378.32
LCE TAC + βLv84.4873.8978.83
LCE TAC + λLt + βLv85.2477.6081.24
+ +Table 7. Ablation on the undesired interaction masking. HM denotes the harmonic mean of the base and novel accuracies. + +
Undesired interaction maskingBaseNovelHM
×85.0077.0580.83
✓85.2477.6081.24
+ +not exploited, (b) the introduction of task-aware pre-context can improve performance by a noticeable margin, and (c) the biggest performance gains are achieved when $K = 5$ . These observations demonstrate the effectiveness of introducing task-aware pre-context and lead us to set the default value of $K$ to 5. + +Self-regulation. We ablate the effect of the self-regulation by removing the textual consistency, the visual consistency, or both. It is worth noting that the other ingredients are kept in this experiment. Table 6 shows the results. Our observations are as follows: (a) without imposing self-regulation, the accuracy on the novel set is very poor, which also leads to a low harmonic mean, (b) imposing the textual consistency or visual consistency alone brings a noticeable performance drop, and (c) imposing the two consistencies simultaneously leads to a significant performance gain. The reason for these observations might be that the alignment between textual and visual modalities is well preserved by imposing the two consistencies, while it is difficult to achieve this with only one consistency. These observations demonstrate the necessity of self-regulation. + +Undesired interaction masking. We ablate the effect of the undesired interaction masking by removing it. It is worth noting that (a) the mask matrix is inserted into the visual encoder in a layer-wise manner and (b) the other ingredients are kept in this experiment. Table 7 shows the results. With the undesired interaction masking, the harmonic mean is improved by a margin of $0.41\%$ . Considering the performance gain is achieved in an experiment setting involving 11 datasets, this observation demonstrates the effectiveness of undesired interaction masking. + +Complexity. We evaluate the complexity by comparing our TAC with the competitive counterparts. It is worth noting that (a) PromptSRC [23] inserts learnable prompts into the textual and visual encoders in a layer-wise manner and (b) our TAC inserts the learnable prompts, along with the task-aware pre-contexts, into the textual and visual encoders in the same way. Table 8 shows the results. Since the task-aware pre-context generation is lightweight and the quan- + +Table 8. Ablation on the complexity. Evaluation is performed on a RTX 3090 GPU. + +
MethodOptimization (ms/image)Inference (ms/image)
CLIP [37]-1.7
CoOp [61]68.13.0
CoCoOp [60]269.0126.6
MaPLe [22]70.21.7
PromptSRC [23]71.53.8
TAC (Ours)73.24.1
+ +Table 9. Ablation on the number of learnable prompts. HM denotes the harmonic mean of the base and novel accuracies. + +
Number of learnable promptsBaseNovelHM
184.1575.6379.66
284.6776.6180.44
485.2477.6081.24
685.0276.8580.73
+ +tity of additional input embeddings incurred by the task-aware pre-context is small, the extra computational overhead incurred by our TAC is marginal compared to PromptSRC [23]. Moreover, it is worth noting our TAC outperforms PromptSRC with noticeable performance gains in all experiment settings. + +Number of learnable prompts. We ablate the effect of the number of learnable prompts by setting $M^t$ and $M^v$ to different values. It is worth noting that (a) $M^t$ and $M^v$ are set to be equal by default and (b) the other ingredients are kept in this experiment. Table 9 shows the results. Our observations are as follows: (a) increasing the number of learnable prompts can improve the harmonic mean and (b) the greatest performance gain is achieved when $M^t = 4$ and $M^v = 4$ . These observations lead us to set the default value of $M^t$ and $M^v$ to 4. + +# 5. Conclusion + +Prompt learning has emerged as a powerful tool for adapting vision-language models. However, due to the scarcity of task-specific data, the resulting prompts often exhibit limited task-awareness. In this work, we propose to increase the task-awareness by introducing task-aware pre-context, which results in a novel task-aware clustering framework for prompting. The key ingredients are three-fold. First, the backbone structure of a downstream task is preserved by task-aware clustering, based on which the task-aware pre-context is generated. Second, the task-awareness of learnable prompts is enhanced by allowing them to interact with task-aware pre-context via the well-pretrained encoders. Finally, the undesired interference of visual task-aware pre-context to patch embeddings is eliminated by masked attention mechanism. Extensive ablation studies confirm the effectiveness of key ingredients. Comparative results in the base-to-novel, domain generalization, and cross-dataset transfer settings show the advantage of our task-aware clustering for prompt learning. + +# Acknowledgements + +This work was supported by Guangdong Major Project of Basic and Applied Basic Research (2023B0303000016), National Natural Science Foundation of China (62206268, U21A20487, 62372440), Shenzhen Technology Project (JCYJ20240813154723031), Shenzhen High-tech Zone Development Special Plan Innovation Platform Construction Project, the proof of concept center for high precision and high resolution 4D imaging, Guangdong Technology Project (2023TX07Z126), and Yunnan Science & Technology Project (202302AD080008, 202305AF150152). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1 +[2] Shtedritski Aleksandar, Rupprecht Christian, and Vedaldi Andrea. What does clip know about a red circle? visual prompt engineering for vlms. In ICCV, pages 11987-11997, 2023. 1, 3 +[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446-461, 2014. 6 +[4] Adrian Bulat and Georgios Tzimiropoulos. Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In CVPR, pages 23232-23241, 2023. 1 +[5] Yun-Hao Cao, Kaixiang Ji, Ziyuan Huang, Chuanyang Zheng, Jiajia Liu, Jian Wang, Jingdong Chen, and Ming Yang. Towards better vision-inspired vision-language models. In CVPR, pages 13537-13547, 2024. 2 +[6] Delong Chen, Zhao Wu, Fan Liu, Zaiquan Yang, Shaoqiu Zheng, Ying Tan, and Erjin Zhou. Protoclip: Prototypical contrastive language image pretraining. IEEE TNNLS (Early Access), pages 1-15, 2023. 2 +[7] Eulrang Cho, Jooyeon Kim, and Hyunwoo J. Kim. Distribution-aware prompt tuning for vision-language models. In ICCV, pages 22004-22013, 2023. 1 +[8] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 6 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 2, 6 +[10] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G. Turrisi da Costa, Cees G.M. Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learning for image-language model generalization. In ICCV, pages 15237-15246, 2023. 6, 7 + +[11] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 1 +[12] Bordes Florian, Pang Richard Yuanzhe, Ajay Anurag, Li Alexander C, Bardes Adrien, Petryk Suzanne, Manas Oscar, Lin Zhiqiu, Mahmoud Anas, Jayaraman Bargav, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2 +[13] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE TNNLS (Early Access), pages 1-11, 2023. 2 +[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. J-STARS, 12(7):2217-2226, 2019. 6, 7 +[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6 +[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 6 +[17] Nisha Huang, Yuxin Zhang, Fan Tang, Chongyang Ma, Haibin Huang, Weiming Dong, and Changsheng Xu. Diff-styler: Controllable dual diffusion for text-driven image stylization. IEEE TNNLS (Early Access), pages 1-14, 2024. 2 +[18] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, pages 3061-3071, 2023. 3 +[19] Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Retrieval-enhanced contrastive vision-text models. In ICLR, 2024. 2 +[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1, 2 +[21] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 1, 6 +[22] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 6, 7, 8 +[23] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 6, 7, 8 +[24] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model, 2022. 2 + +[25] Jonathan Krause, Michael Stark, Jia Deng, and Fei-Fei Li. 3d object representations for fine-grained categorization. In ICCV, pages 554–561, 2013. 6 +[26] Nilakshan Kunananthaseelan, Jing Zhang, and Mehrtash Harandi. Lavip: Language-grounded visual prompting. In AAAI, pages 2840-2848, 2024. 3, 7 +[27] Gukyeong Kwon, Zhaowei Cai, Avinash Ravichandran, Erhan Bas, Rahul Bhotika, and Stefano Soatto. Masked vision and language modeling for multi-modal representation learning. In ICLR, 2023. 2 +[28] Marc Lafon, Elias Ramzi, Clément Rambour, Nicolas Audebert, and Nicolas Thome. Gallop: Learning global and local prompts for vision-language models. In ECCV, 2024. 3 +[29] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J. Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 1, 6, 7 +[30] Fei-Fei Li, Fergus Rob, and Perona Pietro. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshop, pages 178–178, 2004. 6 +[31] Yanxin Long, Jianhua Han, Runhui Huang, Hang Xu, Yi Zhu, Chunjing Xu, and Xiaodan Liang. Fine-grained visual-text prompt-driven self-training for open-vocabulary object detection. IEEE TNNLS (Early Access), pages 1-11, 2023. 2 +[32] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In CVPR, pages 5206-5215, 2022. 3 +[33] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In CVPR, pages 26354-26363, 2024. 3 +[34] Metaxas Ioannis Maniadis, Tzimiropoulos Georgios, and Paras Ioannis. Divclust: Controlling diversity in deep clustering. In CVPR, pages 3418-3428, 2023. 3 +[35] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 6 +[36] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505, 2012. 6 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 6, 7, 8 +[38] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imagenet? In ICML, pages 5389-5400, 2019. 6 +[39] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. In ICLR, 2024. 2, 3, 6, 7 +[40] Lavoie Samuel, Kirichenko Polina, Ibrahim Mark, Assran Mido, Wilson Andrew Gordon, Courville Aaron, and Ballas Nicolas. Modeling caption diversity in contrastive vision-language pretraining. In ICML, pages 26070-26084, 2024. 2 + +[41] Tourani Siddharth, Alwheibi Ahmed, Mahmood Arif, and Khan Muhammad Haris. Pose-guided self-training with two-stage clustering for unsupervised landmark discovery. In CVPR, pages 23041-23051, 2024. 3 +[42] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6 +[43] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In CVPR, pages 23209-23218, 2024. 2 +[44] Maji Subhransu, Rahtu Esa, Kannala Juho, Blaschko Matthew, and Vedaldi Andrea. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6 +[45] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argument: Attribute-guided prompt tuning for vision-language models. In CVPR, pages 28578-28587, 2024. 2, 3, 6, 7 +[46] Yujun Tong, Da Li, Dongliang Chang, Tianwei Cao, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Vision-language subspace prompting, 2024. 2 +[47] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 6 +[48] Mengmeng Wang, Jiazheng Xing, Jianbiao Mei, Yong Liu, and Yunliang Jiang. Actionclip: Adapting language-image pretrained models for video action recognition. IEEE TNNLS (Early Access), pages 1-13, 2023. 2 +[49] Zihao Wei, Zixuan Pan, and Andrew Owens. Efficient vision-language pre-training by cluster masking. In CVPR, pages 26815-26825, 2024. 2 +[50] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 6 +[51] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, pages 6757-6767, 2023. 1, 6, 7 +[52] Hantao Yao, Rui Zhang, and Changsheng Xu. Tcp: Textual-based class-aware prompt tuning for visual-language model. In CVPR, pages 23438-23448, 2024. 3, 6, 7 +[53] Lochman Yaroslava, Olsson Carl, and Zach Christopher. Learned trajectory embedding for subspace clustering. In CVPR, pages 19092-19102, 2024. 3 +[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 2 +[55] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 2 +[56] Yunfei Zhang, Xiaoyang Huo, Tianyi Chen, Si Wu, and Hau San Wong. Exploring intra-class variation factors with learnable cluster prompts for semi-supervised image synthesis. In CVPR, pages 7392-7401, 2023. 3 + +[57] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In CVPR, pages 28718-28728, 2024. 2 +[58] Cairong Zhao, Yubin Wang, Xinyang Jiang, Yifei Shen, Kaitao Song, Dongsheng Li, and Duoqian Miao. Learning domain invariant prompt for vision-language models. IEEE TIP, 33:1348-1360, 2024. 6 +[59] Zhaoheng Zheng, Jingmin Wei, Xuefeng Hu, Haidong Zhu, and Ram Nevatia. Large language models are good prompt learners for low-shot image classification. In CVPR, pages 28453-28462, 2024. 3 +[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 6, 7, 8 +[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 6, 7, 8 +[62] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In CVPR, pages 15659-15669, 2023. 6, 7 +[63] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 \ No newline at end of file diff --git a/2025/Task-Aware Clustering for Prompting Vision-Language Models/images.zip b/2025/Task-Aware Clustering for Prompting Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cd19033116a8faf1b6e2da050e19baefeceec9eb --- /dev/null +++ b/2025/Task-Aware Clustering for Prompting Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d81094b88ace054d5eeee6c09c89f9299d47a5c35c2c95ae72ba450f957f65bd +size 803394 diff --git a/2025/Task-Aware Clustering for Prompting Vision-Language Models/layout.json b/2025/Task-Aware Clustering for Prompting Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..10f22117e24a7db1cb1a364eac311f2736b641b5 --- /dev/null +++ b/2025/Task-Aware Clustering for Prompting Vision-Language Models/layout.json @@ -0,0 +1,8834 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "text", + "content": "Task-Aware Clustering for Prompting Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "spans": [ + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "text", + "content": "Fusheng Hao" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "text", + "content": " Fengxiang He" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "text", + "content": " Fuxiang Wu" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "text", + "content": " Tichao Wang" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "text", + "content": " Chengqun Song" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "text", + "content": " Jun Cheng" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "text", + "content": " Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 141, + 550, + 187 + ], + "type": "text", + "content": " The Chinese University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 239, + 297, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 550 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 550 + ], + "type": "text", + "content": "Prompt learning has attracted widespread attention in adapting vision-language models to downstream tasks. Existing methods largely rely on optimization strategies to ensure the task-awareness of learnable prompts. Due to the scarcity of task-specific data, overfitting is prone to occur. The resulting prompts often do not generalize well or exhibit limited task-awareness. To address this issue, we propose a novel Task-Aware Clustering (TAC) framework for prompting vision-language models, which increases the task-awareness of learnable prompts by introducing task-aware pre-context. The key ingredients are as follows: (a) generating task-aware pre-context based on task-aware clustering that can preserve the backbone structure of a downstream task with only a few clustering centers, (b) enhancing the task-awareness of learnable prompts by enabling them to interact with task-aware pre-context via the well-pretrained encoders, and (c) preventing the visual task-aware pre-context from interfering the interaction between patch embeddings by masked attention mechanism. Extensive experiments are conducted on benchmark datasets, covering the base-to-novel, domain generalization, and cross-dataset transfer settings. Ablation studies validate the effectiveness of key ingredients. Comparative results show the superiority of our TAC over competitive counterparts. The code is available at https://github.com/FushengHao/TAC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "content": "Large-scale visual and textual data pairs have been collected and employed to pretrain high-capacity Vision-Language Models (VLMs) [1, 20, 37]. Due to the alignment of the visual and textual modalities, the resulting models have emerged with remarkable generalization ability on various downstream tasks such as zero-shot image classification [37] and open-vocabulary object detection [11]. Despite achieving decent performance, the performance is still" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 215, + 539, + 381 + ], + "blocks": [ + { + "bbox": [ + 328, + 215, + 539, + 381 + ], + "lines": [ + { + "bbox": [ + 328, + 215, + 539, + 381 + ], + "spans": [ + { + "bbox": [ + 328, + 215, + 539, + 381 + ], + "type": "image", + "image_path": "269335cecc0f062bd1d8e48c4a5f795fa484d9cc991b18b39b8d350631129afe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 383, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 417 + ], + "type": "text", + "content": "Figure 1. Motivation of our TAC. We improve the task-awareness of learnable prompts from a new perspective, i.e., introducing task-aware pre-context." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 425, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 498 + ], + "type": "text", + "content": "lower than that of the customized models [61]. Considering the data-hungry nature of VLMs and the scarcity of task-specific data, it is difficult to improve performance through finetuning such models. As an alternative to finetuning, prompting has become as a powerful paradigm for adapting VLMs to downstream tasks [4, 7, 21, 22, 29, 51]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "Prompting is a technique that adapts VLMs by increasing their understanding of downstream tasks. The current methodology is to provide VLMs with task-relevant context while keeping the model weights frozen. The context is also known as prompts. Many effective prompt templates have been engineered. For example, textual templates like \"a photo of a [class]\" [37] have been hand-crafted to provide task-relevant context for the textual encoder, and visual templates like \"a red circle around an [object]\" [2] have been discovered to direct the visual encoder's attention to the [object]. To avoid the need for large-scale trial and error in prompt engineering, context optimization [61] has been proposed, which treats learnable vectors as prompts and ensure their task-awareness by optimization. This strategy has attracted much attention in recent years. For example, maximizing the ground-truth class scores [61] is proposed to ensure the task-awareness of learnable prompts, and self-regulating constraints [23] are further introduced to guide" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "text", + "content": "*Corresponding author: jun.cheng@siat.ac.cn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14745" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "the optimization of learnable prompts for better generalization. It is worth noting that the number of labeled images per class used for prompt learning is not sufficient in the setting of context optimization1. Due to the scarcity of task-specific data, the resulting prompts are often overfitted [23] and do not generalize well, meaning that the task-awareness they exhibit is limited." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 157, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 295, + 396 + ], + "type": "text", + "content": "Instead of imposing stronger constraints [39, 45, 46], we propose to increase the task-awareness by directly introducing task-aware pre-context; see Figure 1. Considering that the class descriptions of a downstream task are given when encountering the task, the class embeddings extracted by using the well-pretrained textual encoder are good candidates for task-aware pre-context. However, one big challenge arises in this practice, i.e., the number of extra embeddings it incurs is large and equal to that of classes involved. For example, the number is 1000 for the popular dataset ImageNet [9], which is significantly larger than the typical values of the number of embeddings input to the textual and visual encoders, i.e., 77 and 197, respectively. This not only incurs large additional computational overhead, but also exceeds the maximum input limit for the textual encoder. These difficulties inspire us to keep only the backbone structure of class embeddings, which is helpful for reducing the number of embeddings incurred significantly. Motivated by these observations, we explore the prompting strategy based on clustering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 396, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 295, + 694 + ], + "type": "text", + "content": "In this work, we propose a novel Task-Aware Clustering (TAC) framework for prompting VLMs. The central idea of TAC is to exploit task-aware pre-context to improve the task-awareness of learnable prompts. To generate task-aware pre-context, we conduct clustering on the class embeddings of a downstream task and perform linear transformation on the clustering centers to yield task-aware pre-context, which leads to a task-aware pre-context generalization module that can keep the backbone structure of the downstream task while being lightweight. Then, the task-aware pre-context is combined with the learnable prompts and their interaction through the well-pretrained encoders is exploited to reinforce the task-awareness of learnable prompts, which leads to a task-awareness reinforcement module. Further, a mask matrix is developed and injected into the attention mechanism to eliminate the undesired interference of visual task-aware pre-context to patch embeddings induced by the gap between textual and visual modalities, which leads to undesired interaction masking module. To suppress the overfitting induced by the scarcity of task-specific data, the prompted class and visual embeddings are encouraged to be consistent with their CLIP [37] peers, which leads to a self-regulation module. Our TAC maintains simplicity in design and improves the task-awareness of the resulting prompts, thus yielding competitive results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 96 + ], + "type": "text", + "content": "In summary, our task-aware clustering framework for prompting has the following main contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 315, + 98, + 553, + 241 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 315, + 98, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 98, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 98, + 553, + 133 + ], + "type": "text", + "content": "- We propose to increase the task-awareness of learnable prompts from a new perspective, i.e., introducing task-aware pre-context." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 134, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 134, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 315, + 134, + 553, + 168 + ], + "type": "text", + "content": "- We propose to generate task-aware pre-context based on task-aware clustering that can preserve the backbone structure of a downstream task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 170, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 170, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 315, + 170, + 553, + 204 + ], + "type": "text", + "content": "- We exploit the well-pretrained encoders to reinforce the task-awareness of learnable prompts while masking out the undesired interaction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 205, + 553, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 205, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 315, + 205, + 553, + 241 + ], + "type": "text", + "content": "- We demonstrate the superiority of our TAC in the baseto-novel, domain generalization, and cross-dataset transfer settings." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 255, + 405, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 405, + 267 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 405, + 267 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 276, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 624 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) integrate the textual and visual modalities, bridging the gap between them through alignment. Such models are usually pretrained on large-scale visual and textual data pairs [6, 20, 37]. They can be broadly classified into four categories according to their training paradigms [12]. The first category employs contrastive-based training [19, 40], which emphasizes reducing the distance between positive pairs and increasing the distance between negative pairs. The second category utilizes masking-based training [27, 49], which emphasizes the reconstruction of missing patches/words by exploiting the unmasked caption/image. The third category leverages generative-based training [54, 55], which emphasizes the generation of entire images or captions based on conditional captions or images. The fourth category employs LLM-based training [5, 63], which emphasizes the usage of large language models to extract supervisory information. Since the textual and visual modalities are aligned in a shared embedding space, VLMs exhibits a strong ability to understand open-vocabulary concepts, which facilitates various downstream tasks. Although VLMs have achieved decent performance, their accuracies are still far inferior to those of customized models [61]. Considering the data-hungry nature of VLMs and the scarcity of task-specific data, finetuning such models to improve their performance faces significant challenges. As an alternative to finetuning, prompting [37, 61] has attracted much attention in adapting VLMs. Orthogonal directions include the utilization of unlabeled images [43] and historical test data [57]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Prompting adapts VLMs by providing them with contextual information or task-aware prompts while freezing their model weights, which has achieved good performance in many downstream tasks [13, 17, 24, 31, 48]. Existing methods can be broadly classified into two categories according to whether domain expertise is required. The first category relies on hand-crafted prompt templates, which emphasizes" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 702, + 196, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 196, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 196, + 712 + ], + "type": "text", + "content": "The typical value of this number is 16" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14746" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": "the reliance on domain expertise. CLIP [37] crafts the textual template, i.e., \"a photo of a [class]\". CirclePrompt [2] engineers the visual template, i.e., \"a red circle around an [object]\". LaViP [26] grounds visual prompt with language. ArGue [45] crafts attribute-guided prompts with the help of large language models. The second category relies on prompt learning [61], which emphasizes the introduction of learnable prompts and the optimization of these prompts to ensure their task-awareness. PromptSRC [23] guides the optimization of learnable prompts with self-regulating constraints. GalLoP [28] learns global-local prompts to capture multi-scale visual information. CoPrompt [39] improves the generation ability of learnable prompts by encouraging the prediction consistency between the pretrained and trainable models. TCP [52] introduces class-aware prompts that are produced by performing textual knowledge embedding on class embeddings. In this work, we increase the task-awareness by introducing task-aware pre-context, which is different from the current practice that imposes extra constraints." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 317, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 296, + 569 + ], + "type": "text", + "content": "Clustering is a technique that can group similar data points into clusters or groups, in which K-Means is a classic example. Integrating clustering may improve the capabilities of deep models in capturing backbone structures, and their combination has attracted a lot of attention. DivClust [34] introduces a diversity controlling loss to ensure the desired degree of diversity for the produced multiple clusters. LCPGAN [56] associates intra-class clusters with learnable visual concepts to capture the class-wise and intra-class information for image synthesis. ZeroShotULD [41] clusters random pixel locations with nearest neighbor searching for unsupervised landmarks discovery. SPC [18] replaces the deterministic prediction with semantic clustering for domain generalized semantic segmentation. MoDE [33] decides clusters for data experts by a two-step clustering on captions and uses fine-grained clusters to coarsely represent data experts. LTE [53] clusters multiple observed motions by establishing the mapping from trajectories to trajectory embeddings. In this work, we perform clustering to preserve the backbone structure of a downstream task represented by class embeddings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 581, + 112, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 112, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 112, + 594 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 601, + 297, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 297, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 297, + 639 + ], + "type": "text", + "content": "In this section, we begin by reviewing the preliminaries and then detail our Task-Aware Clustering (TAC) framework for prompting vision-language models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 647, + 142, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 142, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 142, + 659 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "CLIP. CLIP [37] includes a textual encoder and a visual encoder, and is pretrained on around 400 million visual and textual data pairs with the contrastive-based training strategy. With the hand-crafted prompt templates like \"a photo" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": "of a [class]\", it emerges strong zero-shot prediction capabilities. Although CLIP has achieved good performance in the zero-shot image recognition task, its accuracies are still lower than those of the customized models [61]. Due to the data-hungry nature of CLIP and the scarcity of task-specific data, finetuning CLIP to improve its performance on downstream tasks is particularly challenging. Considering that prompt engineering relies on domain expertise and needs large-scale trial and error, prompt learning [61] has been proposed, in which learnable vectors are treated as prompts and their task-awareness is ensured through optimization. For ease of comparison with existing prompt learning approaches, we follow the current practice and choose CLIP as the foundation model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": "Prompt learning. CoOp [61] is a typical approach for adapting CLIP in the context of prompt learning. We take a " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": "-class image classification task as an example to show how learnable prompts are optimized in CoOp. First, a set of learnable vectors with a quantity of " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "M^t" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": " is introduced and treated as textual prompts, i.e., " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_m^t\\}_{m=1}^{M^t}" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": ". Second, these prompts are combined with the word embeddings that describe the task, which leads to a learnable version of task descriptions, i.e., " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_1^t,\\dots,\\pmb{p}_{M^t}^t,\\pmb{c}_i\\}_{i=1}^C" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": " is the word embedding(s) of the " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": "-th class. Third, the learnable task descriptions are fed into the textual encoder to yield their embeddings " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\{\\pmb{w}_c^{\\mathrm{CoOp}}\\}_{c=1}^C" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": ", while an image " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": " is fed into the visual encoder to generate its embedding " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": ". Fourth, the probability of " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": " that belongs to the " + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 239, + 556, + 418 + ], + "type": "text", + "content": "-th class can be calculated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 342, + 422, + 555, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 422, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 342, + 422, + 555, + 453 + ], + "type": "interline_equation", + "content": "p ^ {\\mathrm {C o O p}} (c | \\boldsymbol {x}) = \\frac {\\exp \\left(\\cos \\left(\\boldsymbol {w} _ {c} ^ {\\mathrm {C o O p}} , \\boldsymbol {f}\\right) / \\tau\\right)}{\\sum_ {i = 1} ^ {C} \\exp \\left(\\cos \\left(\\boldsymbol {w} _ {i} ^ {\\mathrm {C o O p}}, \\boldsymbol {f}\\right) / \\tau\\right)}, \\tag {1}", + "image_path": "d6237165ec6ca5e3a1254362421de8aa5eff734e8ed7b5dd9e4f2d000585f887.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 456, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 494 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 456, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 456, + 555, + 494 + ], + "type": "text", + "content": " is the temperature and " + }, + { + "bbox": [ + 313, + 456, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 456, + 555, + 494 + ], + "type": "text", + "content": " denotes the cosine similarity. Finally, the task-awareness of learnable textual prompts is ensured by minimizing the cross-entropy loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 347, + 497, + 555, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 497, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 347, + 497, + 555, + 530 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} ^ {\\mathrm {C o O p}} = - \\sum_ {c = 1} ^ {C} \\mathbb {I} (c = y) \\log p ^ {\\mathrm {C o O p}} (c | \\boldsymbol {x}), \\tag {2}", + "image_path": "70d03fadfc84c56dcb55611235347ec2b083a7cf567adc51ef775e032aa87e71.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 534, + 556, + 605 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 534, + 556, + 605 + ], + "type": "text", + "content": " is the ground-truth label of " + }, + { + "bbox": [ + 313, + 534, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 534, + 556, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 313, + 534, + 556, + 605 + ], + "type": "text", + "content": " denotes the indicator function. It is worth noting that (a) only the learnable prompts need to be optimized and (b) the quantity of labeled images that can be used in each class is insufficient in the setting of prompt learning, where the typical value is 16." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 606, + 555, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 667 + ], + "type": "text", + "content": "Afterwards, additional constraints [23, 32, 39, 59] have been imposed to improve the task-awareness of learnable prompts. Different from the current practice, we increase the task-awareness from a new perspective, i.e., introducing task-aware pre-context." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 672, + 514, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 514, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 514, + 685 + ], + "type": "text", + "content": "3.2. Task-aware Clustering for Prompting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "The pipeline of our TAC is shown in Figure 2 and the key ingredients are detailed below." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14747" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 545, + 404 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 545, + 404 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 545, + 404 + ], + "type": "image", + "image_path": "9fccb3d167d80bb3593a66d3b54553a5721cf609eb020d9dd6c18d3f5edae6e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 408, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 54, + 408, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 555, + 455 + ], + "type": "text", + "content": "Figure 2. Pipeline of our TAC. First, the textual and visual task-aware pre-contexts are generated based on the backbone structure of a downstream task. Second, the task-aware pre-context is combined with the learnable prompts and their interaction through the well-pretrained encoders is exploited to reinforce the task-awareness of learnable prompts. Third, the undesired interaction is masked out with a customized mask matrix. Further, the textual and visual consistencies are imposed to alleviate the overfitting issue." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 462, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 295, + 605 + ], + "type": "text", + "content": "Task-aware pre-context generation. The class embeddings extracted by using the textual encoder of CLIP are task-aware and good candidates for the desired pre-context. Considering that the number of such embeddings is equal to that of classes involved and this number is large, applying this strategy to practice is challenging. The first challenge is that the number of extra input embeddings incurred exceeds the maximum input limit for the textual encoder, which would make the textual encoder unusable. The second challenge is that the extra computational overhead incurred is very high due to the large number of additional embeddings input into the visual encoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "We address these challenges by yielding textual and visual task-aware pre-contexts based on clustering that can preserve the backbone structure of class embeddings. A significant additional benefit of this practice is that the number of extra embeddings incurred can be significantly reduced. Specifically, let " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{w}_c^{\\mathrm{CLIP}}\\}_{c=1}^C" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " denote the class embeddings extracted with the textual encoder of CLIP. Then, K-Means is conducted on " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{w}_c^{\\mathrm{CLIP}}\\}_{c=1}^C" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": ", which yields a set of clustering centers with a quantity of " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{\\mu}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": ". Further, the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 462, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 475 + ], + "type": "text", + "content": "textual and visual task-aware pre-contexts are generated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 405, + 483, + 555, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 483, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 405, + 483, + 555, + 498 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {i} ^ {t} = \\boldsymbol {W} ^ {t} \\boldsymbol {\\mu} _ {i}, \\tag {3}", + "image_path": "01337b30a5cd90c97eab4584dd4ad3991212f8a3fdd270f3591273ed935c7e57.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 405, + 506, + 555, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 506, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 405, + 506, + 555, + 520 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {i} ^ {v} = \\boldsymbol {W} ^ {v} \\boldsymbol {\\mu} _ {i}, \\tag {4}", + "image_path": "c4e9e89e0c3a39ae67d19d2e6eecad84ca4f68e7dea9761d5471b39b286b2939.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W}^t" + }, + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W}^v" + }, + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "text", + "content": " are the textual and visual linear transformations. It is worth noting that (a) the dimensions of " + }, + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu}_i^t" + }, + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu}_i^v" + }, + { + "bbox": [ + 313, + 523, + 555, + 595 + ], + "type": "text", + "content": " are set to be the internal dimensions of the textual and visual encoders, correspondingly, and (b) since only the weights of linear transformations needs to be optimized, our task-aware pre-context generation module is lightweight." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 596, + 556, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 556, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 556, + 691 + ], + "type": "text", + "content": "Task-awareness reinforcement. After obtaining the textual and visual task-aware pre-contexts, the task-awareness of learnable prompts is reinforced by enabling them to interact with the corresponding pre-contexts via the well-pretrained textual and visual encoders. Specifically, let " + }, + { + "bbox": [ + 313, + 596, + 556, + 691 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 596, + 556, + 691 + ], + "type": "text", + "content": " denote the quantity of transformer blocks contained in the visual/textual encoder. We construct the input embeddings to the " + }, + { + "bbox": [ + 313, + 596, + 556, + 691 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 596, + 556, + 691 + ], + "type": "text", + "content": "-th transformer block of the textual encoder as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 342, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 342, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {t} _ {l} ^ {s o s}, \\boldsymbol {\\mu} _ {1} ^ {t}, \\dots , \\boldsymbol {\\mu} _ {K} ^ {t}, \\boldsymbol {p} _ {l 1} ^ {t}, \\dots , \\boldsymbol {p} _ {l M ^ {t}} ^ {t}, \\boldsymbol {c} _ {l i}, \\boldsymbol {t} _ {l} ^ {e o s} \\right\\}. \\tag {5}", + "image_path": "9c3655b9ecd5f8797269e52593d21cce7d5e3959acaedee46b25741f1e33b518.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14748" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "l = 1,\\dots ,L" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{t}_l^{sos}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " is the start token input to the " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "-th transformer block, " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{t}_l^{eos}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " is the end token input to the " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "-th transformer block, " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_{li}^t\\}_{i=1}^{M^t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " is the set of learnable textual prompts input to the " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "-th transformer block with a quantity of " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "M^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{li}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " is the word embedding(s) input to the " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "-th transformer block. It is worth noting that (a) " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{t}_l^{sos}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{t}_l^{eos}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{li}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " are the outputs of the previous transformer block and (b) " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_{li}^t\\}_{i=1}^{M^t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\mu}_i^t\\}_{i=1}^K" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " are not the outputs of the previous transformer block; see Figure 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "type": "text", + "content": "Similarly, we construct the input embeddings to the " + }, + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "type": "text", + "content": "-th transformer block of the visual encoder as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 214, + 296, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 214, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 66, + 214, + 296, + 230 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {e} _ {l} ^ {c l s}, \\boldsymbol {x} _ {l 1} ^ {p}, \\dots , \\boldsymbol {x} _ {l M} ^ {p}, \\boldsymbol {\\mu} _ {1} ^ {v}, \\dots , \\boldsymbol {\\mu} _ {K} ^ {v}, \\boldsymbol {p} _ {l 1} ^ {v}, \\dots , \\boldsymbol {p} _ {l M ^ {v}} ^ {v} \\right\\}. \\tag {6}", + "image_path": "f1eb61479ba1b5f264d2deaeb55910f77afe54ef8eb85a885620e76fc8e75eed.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "l = 1,\\dots ,L" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "e_l^{cls}" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": " is the class token input to the " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": "-th transformer block, " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_{lm}^p\\}_{m = 1}^M" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": " is the set of patch embeddings input to the " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": "-th transformer block with a quantity of " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_{li}^v\\}_{i = 1}^{M^v}" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": " is the set of learnable visual prompts input to the " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": "-th transformer block with a quantity of " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "M^v" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": ". It is worth noting that (a) " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "e_l^{cls}" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_{lm}^p\\}_{m = 1}^M" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": " are the outputs of the previous transformer block and (b) " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_{li}^v\\}_{i = 1}^{M^v}" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\mu}_i^v\\}_{i = 1}^K" + }, + { + "bbox": [ + 55, + 237, + 296, + 345 + ], + "type": "text", + "content": " are not the outputs of the previous transformer block; see Figure 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "text", + "content": "Undesired interaction masking. There is a modal gap between patch embeddings and visual task-aware pre-context. Allowing patch embeddings to interact with visual task-aware pre-context may weaken the discriminative ability of patch embeddings. To avoid this phenomenon, the information flow from visual task-aware pre-context to patch embeddings is prohibited, which is achieved by exploiting the masked attention mechanism: " + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = \\mathrm{softmax}(Q\\mathbf{K}^T + \\mathcal{M})" + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "text", + "content": " denote the query and key matrices, with the first dimension specifying the quantity of input embeddings, i.e., " + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "inline_equation", + "content": "N_v = 1 + M + K + M^v" + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "text", + "content": ". They are the linear transformations of input embeddings. " + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N_v \\times N_v}" + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "text", + "content": " denotes the attention matrix. " + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{N_v \\times N_v}" + }, + { + "bbox": [ + 55, + 346, + 296, + 513 + ], + "type": "text", + "content": " is the customized mask matrix, which is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 557 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {i j} = \\left\\{ \\begin{array}{l l} - \\infty & \\text {i f} 1 < i \\leq 1 + M \\text {a n d} 1 + M < j \\leq 1 + M + K \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}", + "image_path": "1a8f35eaad0a3155042e1e5a47424d030f2cbbda14710206ea37e05d011fea80.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{ij} = -\\infty" + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": " is equal to " + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "inline_equation", + "content": "A_{ij} = 0" + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": ", meaning that the " + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": "-th input embedding does not interact with the " + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": "-th peer. In effect, with " + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": ", patch embeddings are made to not interact with visual task-aware pre-context. It is worth noting that (a) " + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 558, + 296, + 652 + ], + "type": "text", + "content": " is injected into the visual encoder in a layer-wise manner; see Figure 2, and (b) the learnable prompts act as a glue between visual task-aware pre-context and patch embeddings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Self-regulation. Due to the scarcity of task-specific data, overfitting can easily occur. We address this issue by exploiting the strong generation ability exhibited by the well-pretrained encoders, i.e., encouraging the prompted class and visual embeddings to be consistent with their CLIP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "peers. Specifically, let " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{w}_c^{\\mathrm{TAC}}\\}_{c=1}^C" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " denote the class embeddings extracted by using our TAC. The textual consistency is imposed by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 379, + 116, + 555, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 116, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 379, + 116, + 555, + 149 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {t} = \\sum_ {i = 1} ^ {C} | \\boldsymbol {w} _ {c} ^ {\\mathrm {T A C}} - \\boldsymbol {w} _ {c} ^ {\\mathrm {C L I P}} |. \\tag {8}", + "image_path": "80597202bc9d26ed82aca62dcc691925525c9869b077cc6f1b2a7e6cd6656d04.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 157, + 554, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 554, + 207 + ], + "type": "text", + "content": "Data augmentations are applied " + }, + { + "bbox": [ + 313, + 157, + 554, + 207 + ], + "type": "inline_equation", + "content": "N^d" + }, + { + "bbox": [ + 313, + 157, + 554, + 207 + ], + "type": "text", + "content": " times on " + }, + { + "bbox": [ + 313, + 157, + 554, + 207 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 157, + 554, + 207 + ], + "type": "text", + "content": ". Then, the augmented images are fed into our visual encoder to extract their embeddings, i.e., " + }, + { + "bbox": [ + 313, + 157, + 554, + 207 + ], + "type": "inline_equation", + "content": "\\{f_i^{\\mathrm{TAC}}\\}_{i=1}^{N^d}" + }, + { + "bbox": [ + 313, + 157, + 554, + 207 + ], + "type": "text", + "content": ". Correspondingly, the visual consistency is imposed by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 380, + 214, + 554, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 214, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 380, + 214, + 554, + 249 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {v} = \\sum_ {i = 1} ^ {N ^ {d}} | \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}} - \\boldsymbol {f} ^ {\\mathrm {C L I P}} |. \\tag {9}", + "image_path": "08e3d74d792a99fa40fde2a40598bea8be11ab4cd33c9c4da2c79f9481b89bc0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "type": "text", + "content": " denotes the embedding extracted by using the visual encoder of CLIP on " + }, + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 257, + 553, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "text", + "content": "Overall loss. The probability of " + }, + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{TAC}}" + }, + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "text", + "content": " that belongs to the " + }, + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "text", + "content": "-th class can be calculated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 310, + 554, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 310, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 323, + 310, + 554, + 342 + ], + "type": "interline_equation", + "content": "p ^ {\\mathrm {T A C}} \\left(c \\mid \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right) = \\frac {\\exp \\left(\\cos \\left(\\boldsymbol {w} _ {c} ^ {\\mathrm {T A C}} , \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(\\boldsymbol {w} _ {j} ^ {\\mathrm {T A C}}, \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right) / \\tau\\right)}. \\tag {10}", + "image_path": "033a166b48d6ac19c6cc30ef2f6b1ab0a886bd127420540838bdfb98fde9623c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 348, + 550, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 550, + 360 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 550, + 360 + ], + "type": "text", + "content": "Then, the following cross-entropy loss can be constructed:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 327, + 368, + 554, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 368, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 327, + 368, + 554, + 402 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} ^ {\\mathrm {T A C}} = - \\sum_ {i = 1} ^ {N ^ {d}} \\sum_ {c = 1} ^ {C} \\mathbb {I} (c = y) \\log p ^ {\\mathrm {T A C}} \\left(c \\mid \\boldsymbol {f} _ {i} ^ {\\mathrm {T A C}}\\right). \\tag {11}", + "image_path": "cbf41866e15df525e7e3b1c08c0685a8f7f87d408771e3bdd5fadf3a2f18b196.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 409, + 549, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 409, + 549, + 421 + ], + "spans": [ + { + "bbox": [ + 314, + 409, + 549, + 421 + ], + "type": "text", + "content": "After combining self-regulation, the overall loss becomes:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 366, + 428, + 553, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 428, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 366, + 428, + 553, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {O v e r a l l}} = \\mathcal {L} _ {\\mathrm {C E}} ^ {\\mathrm {T A C}} + \\lambda \\mathcal {L} ^ {t} + \\beta \\mathcal {L} ^ {v}. \\tag {12}", + "image_path": "30f1994eeb9246ca072533637179602607ac6023f0c3e22c2980a42e7db75b30.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 451, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 553, + 473 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 451, + 553, + 473 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 451, + 553, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 451, + 553, + 473 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 451, + 553, + 473 + ], + "type": "text", + "content": " are hyper-parameters that play a balancing role." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": "Complexity. The discussion is three-fold. First, the clustering on " + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{w}_c^{\\mathrm{CLIP}}\\}_{c=1}^C" + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": " and the extraction of original class and visual embeddings are pre-computed and cached. Thus, the extra computational overhead incurred by them is marginal. Second, the textual task-aware pre-context does not incur extra computational overhead since the textual encoder of CLIP only accepts a sufficiently large fixed-number of input embeddings. Third, with the visual task-aware pre-context, the quantity of embeddings input to the visual encoder only increases by a very small amount, i.e., " + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": ". Thus, the extra computational overhead incurred is very limited. In summary, the extra computational overhead incurred by our TAC is marginal." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 640, + 395, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 395, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 395, + 654 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 660, + 421, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 421, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 421, + 673 + ], + "type": "text", + "content": "4.1. Experiment setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Evaluation and datasets. The evaluation is conducted in the base-to-novel, domain generalization, and cross-dataset transfer settings, and the average accuracy of three" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14749" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 92, + 550, + 487 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 91 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 91 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 91 + ], + "type": "text", + "content": "Table 1. Comparative results in the base-to-novel setting. Bold accuracies are the highest. HM denotes the harmonic mean of the base and novel accuracies." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 92, + 550, + 487 + ], + "lines": [ + { + "bbox": [ + 63, + 92, + 550, + 487 + ], + "spans": [ + { + "bbox": [ + 63, + 92, + 550, + 487 + ], + "type": "table", + "html": "
(a) Average(b) ImageNet(c) Caltech101(d) OxfordPets
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]69.3474.2271.70CLIP [37]72.4368.1470.22CLIP [37]96.8494.0095.40CLIP [37]91.1797.2694.12
CoOp [61]82.6963.2271.66CoOp [61]76.4767.8871.92CoOp [61]98.0089.8193.73CoOp [61]93.6795.2994.47
Co-CoOp [60]80.4771.6975.83Co-CoOp [60]75.9870.4373.10Co-CoOp [60]97.9693.8195.84Co-CoOp [60]95.2097.6996.43
ProGrad [62]82.4870.7576.16ProGrad [62]77.0266.6671.46ProGrad [62]98.0293.8995.91ProGrad [62]95.0797.6396.33
RPO [29]81.1375.0077.78RPO [29]76.6071.5774.00RPO [29]97.9794.3796.03RPO [29]94.6397.5096.05
MetaPrompt [58]83.3876.0979.57MetaPrompt [58]77.3971.0674.09MetaPrompt [58]98.2894.5896.39MetaPrompt [58]95.7196.9896.34
KAPT [21]78.4170.5274.26KAPT [21]71.1065.2068.02KAPT [21]97.1093.5395.28KAPT [21]93.1396.5394.80
KgCoOp [51]80.7373.6077.00KgCoOp [51]75.8369.9672.78KgCoOp [51]97.7294.3996.03KgCoOp [51]94.6597.7696.18
MaPLe [22]82.2875.1478.55MaPLe [22]76.6670.5473.47MaPLe [22]97.7494.3696.02MaPLe [22]95.4397.7696.58
PromptSRC [23]84.2676.1079.97PromptSRC [23]77.6070.7374.01PromptSRC [23]98.1094.0396.02PromptSRC [23]95.3397.3096.30
TCP [52]84.1375.3679.51TCP [52]77.2769.8773.38TCP [52]98.2394.6796.42TCP [52]94.6797.2095.92
CoPrompt [39]84.0077.2380.48CoPrompt [39]77.6771.2774.33CoPrompt [39]98.2794.9096.55CoPrompt [39]95.6798.1096.87
ArGue [45]83.6978.0780.78ArGue [45]76.9272.0674.41ArGue [45]98.4395.2096.79ArGue [45]95.3697.9596.64
TAC (Ours)85.2477.6081.24TAC (Ours)78.5771.0374.61TAC (Ours)98.5795.2796.89TAC (Ours)95.9398.1797.04
(e) StanfordCars(f) Flowers102(g) Food101(h) FGVCAircraft
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]63.3774.8968.65CLIP [37]72.0877.8074.83CLIP [37]90.1091.2290.66CLIP [37]27.1936.2931.09
CoOp [61]78.1260.4068.13CoOp [61]97.6059.6774.06CoOp [61]88.3382.2685.19CoOp [61]40.4422.3028.75
Co-CoOp [60]70.4973.5972.01Co-CoOp [60]94.8771.7581.71Co-CoOp [60]90.7091.2990.99Co-CoOp [60]33.4123.7127.74
ProGrad [62]77.6868.6372.88ProGrad [62]95.5471.8782.03ProGrad [62]90.3789.5989.98ProGrad [62]40.5427.5732.82
RPO [29]73.8775.5374.69RPO [29]94.1376.6784.50RPO [29]90.3390.8390.58RPO [29]37.3334.2035.70
MetaPrompt [58]75.4374.4374.93MetaPrompt [58]97.5374.5484.50MetaPrompt [58]90.7691.7791.26MetaPrompt [58]39.3837.5938.46
KAPT [21]69.4766.2067.79KAPT [21]95.0071.2081.40KAPT [21]86.1387.0686.59KAPT [21]29.6728.7329.19
KgCoOp [51]71.7675.0473.36KgCoOp [51]95.0074.7383.65KgCoOp [51]90.5091.7091.09KgCoOp [51]36.2133.5534.83
MaPLe [22]72.9474.0073.47MaPLe [22]95.9272.4682.56MaPLe [22]90.7192.0591.38MaPLe [22]37.4435.6136.50
PromptSRC [23]78.2774.9776.58PromptSRC [23]98.0776.5085.95PromptSRC [23]90.6791.5391.10PromptSRC [23]42.7337.8740.15
TCP [52]80.8074.1377.32TCP [52]97.7375.5785.23TCP [52]90.5791.3790.97TCP [52]41.9734.4337.83
CoPrompt [39]76.9774.4075.66CoPrompt [39]97.2776.6085.71CoPrompt [39]90.7392.0791.40CoPrompt [39]40.2039.3339.76
ArGue [45]75.6473.3874.49ArGue [45]98.3475.4185.36ArGue [45]92.3391.9692.14ArGue [45]40.4638.0339.21
TAC (Ours)81.6374.1777.72TAC (Ours)97.9776.8786.15TAC (Ours)90.8791.8791.37TAC (Ours)44.6037.7040.86
(i) SUN397(j) DTD(k) EuroSAT(l) UCF101
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [37]69.3675.3572.23CLIP [37]53.2459.9056.37CLIP [37]56.4864.0560.03CLIP [37]70.5377.5073.85
CoOp [61]80.6065.8972.51CoOp [61]79.4441.1854.24CoOp [61]92.1954.7468.69CoOp [61]84.6956.0567.46
Co-CoOp [60]79.7476.8678.27Co-CoOp [60]77.0156.0064.85Co-CoOp [60]87.4960.0471.21Co-CoOp [60]82.3373.4577.64
ProGrad [62]81.2674.1777.55ProGrad [62]77.3552.3562.45ProGrad [62]90.1160.8972.67ProGrad [62]84.3374.9479.35
RPO [29]80.6077.8079.18RPO [29]76.7062.1368.61RPO [29]86.6368.9776.79RPO [29]83.6775.4379.34
MetaPrompt [58]82.1079.0180.53MetaPrompt [58]82.5260.1069.55MetaPrompt [58]93.3778.3485.20MetaPrompt [58]84.7078.5681.51
KAPT [21]79.4074.3376.78KAPT [21]75.9758.3065.97KAPT [21]84.8067.5775.21KAPT [21]80.8367.1073.33
KgCoOp [51]80.2976.5378.36KgCoOp [51]77.5554.9964.35KgCoOp [51]85.6464.3473.48KgCoOp [51]82.8976.6779.65
MaPLe [22]80.8278.7079.75MaPLe [22]80.3659.1868.16MaPLe [22]94.0773.2382.35MaPLe [22]83.0078.6680.77
PromptSRC [23]82.6778.4780.52PromptSRC [23]83.3762.9771.75PromptSRC [23]92.9073.9082.32PromptSRC [23]81.7078.8082.74
TCP [52]82.6378.2080.35TCP [52]82.7758.0768.23TCP [52]91.6374.7382.32TCP [52]81.1380.7783.83
CoPrompt [39]82.6380.0381.31CoPrompt [39]83.1364.7372.79CoPrompt [39]94.6078.5785.84CoPrompt [39]86.9079.5783.07
ArGue [45]81.5280.7481.13ArGue [45]81.6066.5573.31ArGue [45]94.4388.2491.23ArGue [45]85.5679.2982.31
TAC (Ours)83.7080.0381.82TAC (Ours)83.3764.2772.58TAC (Ours)94.3782.6088.10TAC (Ours)88.0781.6784.75
", + "image_path": "1c0d17228419df4758dacac99a534b0a356e6a1b68b61b3cb650132908cbd93b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 520, + 549, + 614 + ], + "blocks": [ + { + "bbox": [ + 55, + 498, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 55, + 498, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 555, + 518 + ], + "type": "text", + "content": "Table 2. Comparative results in the cross-dataset transfer setting. Bold accuracies are the highest. Optimization is conducted on ImageNet and evaluation is performed on other datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 520, + 549, + 614 + ], + "lines": [ + { + "bbox": [ + 61, + 520, + 549, + 614 + ], + "spans": [ + { + "bbox": [ + 61, + 520, + 549, + 614 + ], + "type": "table", + "html": "
SourceTarget
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101Average
CoOp [61]71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
Co-CoOp [60]71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
MaPLe [22]70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
Bayesian Prompt [10]70.9393.6790.6365.0070.9086.3024.9367.4746.1045.8768.6765.95
PromptSRC [23]71.2793.6090.2565.7070.2586.1523.9067.1046.8745.5068.7565.81
TCP [52]71.4093.9791.2564.6971.2186.6923.4567.1544.3551.4568.7366.29
CoPrompt [39]70.8094.5090.7365.6772.3086.4324.0067.5747.0751.9069.7367.00
TAC (Ours)72.7794.5390.6765.3072.2085.8323.5367.6347.5748.0770.0066.53
", + "image_path": "02340d6b08ae4eeca10f23fe85ddd21e1f520a2e09c7aa9e6f1b5cc4709a2b8d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 294, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 700 + ], + "type": "text", + "content": "runs is reported as the final performance. The datasets involved are as follows: ImageNet [9], Caltech101 [30], OxfordPets [36], StanfordCars [25], Flowers102 [35], Food101 [3], FGVCAircraft [44], SUN397 [50], DTD [8], EuroSAT [14], UCF101 [42], ImageNet-A [16], ImageNet-R [15], ImageNet-Sketch [47], and ImageNetV2 [38]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Implementation details. Our TAC is built on top of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": "CoOp [61] and PromptSRC [23]. The ViT-B/16 based CLIP [37] is used as the foundation model. The optimizer is SGD, with a learning rate of " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-3}" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": ", a weight decay of " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": ", and a momentum of 0.9. The cosine annealing is used for learning rate scheduling. The random flip and resized crop are used for data augmentation. The batch size is 8. " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "M^t = 4" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "M^v = 4" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "N^d = 4" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 20" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14750" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 93, + 487, + 259 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "type": "text", + "content": "Table 3. Comparative results in the domain generalization setting. Bold accuracies are the highest. Optimization is conducted on ImageNet and evaluation is performed on datasets with domain shifts." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 123, + 93, + 487, + 259 + ], + "lines": [ + { + "bbox": [ + 123, + 93, + 487, + 259 + ], + "spans": [ + { + "bbox": [ + 123, + 93, + 487, + 259 + ], + "type": "table", + "html": "
SourceTarget
ImageNetImageNetV2ImageNet-SketchImageNet-AImageNet-RAverage
CLIP [37]66.7360.8346.1547.7773.9657.17
RPO [29]71.6765.1349.2750.1376.5760.27
LaViP [26]65.9561.6047.2348.9183.9360.41
TCP [52]71.2064.6049.5051.2076.7360.51
CoOp [61]71.5164.2047.9949.7175.2159.28
Co-CoOp [60]71.0264.0748.7550.6376.1859.90
ProGrad [62]72.2464.7347.6149.3974.5859.07
KgCoOp [51]71.2064.1048.9750.6976.7060.11
MaPLe [22]70.7264.0749.1550.9076.9860.26
Bayesian Prompt [10]70.9364.2349.2051.3377.0060.44
PromptSRC [23]71.2764.3549.5550.9077.8060.65
CoPrompt [39]70.8064.2549.4350.5077.5160.42
ArGue [45]71.8465.0249.2551.4776.9660.67
TAC (Ours)72.7765.9750.3051.7378.5061.63
", + "image_path": "b1ecdd5413a2be6fdf980651bd3747fd2c1add895e84bc8fb2c0be117f3045f1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 305, + 281, + 367 + ], + "blocks": [ + { + "bbox": [ + 55, + 272, + 297, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 297, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 297, + 304 + ], + "type": "text", + "content": "Table 4. Ablation on the textual and visual task-aware precontexts. HM denotes the harmonic mean of the base and novel accuracies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 305, + 281, + 367 + ], + "lines": [ + { + "bbox": [ + 69, + 305, + 281, + 367 + ], + "spans": [ + { + "bbox": [ + 69, + 305, + 281, + 367 + ], + "type": "table", + "html": "
Pre-contextBaseNovelHM
None84.7075.5579.86
Textual84.5776.4880.32
Visual84.6776.0680.13
Textual & visual85.2477.6081.24
", + "image_path": "fed9496978f3b3600b3d5e499564ac88297df6ac6ebd85d8b7ad80cf5b9cee0a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\beta = 25" + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\tau = 0.07" + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "text", + "content": ". In the base-to-novel setting, the number of epochs is 20 and the scale of random resized crop is (0.4, 1.0). In the domain generalization and cross-dataset transfer settings, the number of epochs is 5 and the scale of random resized crop is (0.8, 1.0). The learning rate is warmed up for one epoch with a rate of " + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "inline_equation", + "content": "1.0 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 174, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 174, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 174, + 474 + ], + "type": "text", + "content": "4.2. Comparative results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 479, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 564 + ], + "type": "text", + "content": "Base-to-novel. In this setting, optimization is conducted on the base set and evaluation is performed on the base and novel sets. Table 1 shows the comparative results. Our TAC achieves the best average base accuracy, the second-best average novel accuracy, and the best harmonic mean. These results demonstrate the generalizability of our task-aware clustering prompting framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "Cross-dataset transfer. In this setting, optimization is conducted on ImageNet and evaluation is performed on other datasets. Table 2 shows the comparative results. Our TAC achieves the best accuracy on ImageNet. Also, our TAC yields the second-best average accuracy on other datasets, which is slightly worse than CoPrompt [39]. It is worth noting that our TAC beats CoPrompt in the baseto-novel and domain generalization settings. These results demonstrate the transferability of our task-aware clustering prompting framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Domain generalization. In this setting, optimization is conducted on ImageNet and evaluation is performed on" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 326, + 294, + 541, + 373 + ], + "blocks": [ + { + "bbox": [ + 313, + 272, + 554, + 293 + ], + "lines": [ + { + "bbox": [ + 313, + 272, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 554, + 293 + ], + "type": "text", + "content": "Table 5. Ablation on the number of clustering centers. HM denotes the harmonic mean of the base and novel accuracies." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 294, + 541, + 373 + ], + "lines": [ + { + "bbox": [ + 326, + 294, + 541, + 373 + ], + "spans": [ + { + "bbox": [ + 326, + 294, + 541, + 373 + ], + "type": "table", + "html": "
Number of clustering centersBaseNovelHM
K=084.7075.5579.86
K=184.9577.3980.99
K=284.9477.4181.00
K=385.3077.3281.14
K=485.4077.3481.17
K=585.2477.6081.24
", + "image_path": "c9fd9d3d730c9b7c575320b4f7a52ae5452d70b110ec665a099424f31a387da3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 445 + ], + "type": "text", + "content": "datasets with domain shifts. Table 3 shows the comparative results. Our TAC achieves the best accuracy on ImageNet and the best average accuracy on the datasets with domain shifts. These results demonstrate the domain generalizability of our task-aware clustering prompting framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 455, + 413, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 413, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 413, + 467 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 521 + ], + "type": "text", + "content": "Ablation studies are conducted in the base-to-novel setting. The average base accuracy, the average novel accuracy, and their harmonic mean are reported. More is shown in the appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "content": "Task-aware pre-context. We ablate the effect of the task-aware pre-context by removing the textual pre-context, the visual pre-context, or both. It is worth noting that the other ingredients are kept in this experiment. Table 4 shows the results. Our observations are as follows: (a) both the textual and visual task-aware pre-contexts can improve the harmonic mean and (b) their combination achieves the greatest performance gains. These observations demonstrate the effectiveness of introducing task-aware pre-context." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "Number of clustering centers. We ablate the effect of the number of clustering centers by setting " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " to different values. It is worth noting that (a) the maximum value of " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " is 5 since the dataset EuroSAT [14] only has 5 classes and (b) the other ingredients are kept in this experiment. Table 5 shows the results. Our observations are as follows: (a) " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "K = 0" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " means that the task-aware pre-context is" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14751" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 92, + 281, + 158 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 6. Ablation on the self-regulation. HM denotes the harmonic mean of the base and novel accuracies." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 92, + 281, + 158 + ], + "lines": [ + { + "bbox": [ + 67, + 92, + 281, + 158 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 281, + 158 + ], + "type": "table", + "html": "
Self-regulationBaseNovelHM
LCE TAC85.2974.1879.34
LCE TAC + λLt83.5273.7378.32
LCE TAC + βLv84.4873.8978.83
LCE TAC + λLt + βLv85.2477.6081.24
", + "image_path": "2743fb8577246e3f74fe479c4abff6478bbf19c47c90c05ce18f8acff3a0bfa3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 188, + 281, + 229 + ], + "blocks": [ + { + "bbox": [ + 55, + 166, + 295, + 188 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 295, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 295, + 188 + ], + "type": "text", + "content": "Table 7. Ablation on the undesired interaction masking. HM denotes the harmonic mean of the base and novel accuracies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 188, + 281, + 229 + ], + "lines": [ + { + "bbox": [ + 67, + 188, + 281, + 229 + ], + "spans": [ + { + "bbox": [ + 67, + 188, + 281, + 229 + ], + "type": "table", + "html": "
Undesired interaction maskingBaseNovelHM
×85.0077.0580.83
✓85.2477.6081.24
", + "image_path": "beb33bad77cf866b5225641b273eabaa9d1c24c13095ca489e82aedb21ba13fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 242, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 295, + 312 + ], + "type": "text", + "content": "not exploited, (b) the introduction of task-aware pre-context can improve performance by a noticeable margin, and (c) the biggest performance gains are achieved when " + }, + { + "bbox": [ + 55, + 242, + 295, + 312 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 55, + 242, + 295, + 312 + ], + "type": "text", + "content": ". These observations demonstrate the effectiveness of introducing task-aware pre-context and lead us to set the default value of " + }, + { + "bbox": [ + 55, + 242, + 295, + 312 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 242, + 295, + 312 + ], + "type": "text", + "content": " to 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 315, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 495 + ], + "type": "text", + "content": "Self-regulation. We ablate the effect of the self-regulation by removing the textual consistency, the visual consistency, or both. It is worth noting that the other ingredients are kept in this experiment. Table 6 shows the results. Our observations are as follows: (a) without imposing self-regulation, the accuracy on the novel set is very poor, which also leads to a low harmonic mean, (b) imposing the textual consistency or visual consistency alone brings a noticeable performance drop, and (c) imposing the two consistencies simultaneously leads to a significant performance gain. The reason for these observations might be that the alignment between textual and visual modalities is well preserved by imposing the two consistencies, while it is difficult to achieve this with only one consistency. These observations demonstrate the necessity of self-regulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "content": "Undesired interaction masking. We ablate the effect of the undesired interaction masking by removing it. It is worth noting that (a) the mask matrix is inserted into the visual encoder in a layer-wise manner and (b) the other ingredients are kept in this experiment. Table 7 shows the results. With the undesired interaction masking, the harmonic mean is improved by a margin of " + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "inline_equation", + "content": "0.41\\%" + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "content": ". Considering the performance gain is achieved in an experiment setting involving 11 datasets, this observation demonstrates the effectiveness of undesired interaction masking." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "Complexity. We evaluate the complexity by comparing our TAC with the competitive counterparts. It is worth noting that (a) PromptSRC [23] inserts learnable prompts into the textual and visual encoders in a layer-wise manner and (b) our TAC inserts the learnable prompts, along with the task-aware pre-contexts, into the textual and visual encoders in the same way. Table 8 shows the results. Since the task-aware pre-context generation is lightweight and the quan-" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 92, + 539, + 169 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 91 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 91 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 91 + ], + "type": "text", + "content": "Table 8. Ablation on the complexity. Evaluation is performed on a RTX 3090 GPU." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 92, + 539, + 169 + ], + "lines": [ + { + "bbox": [ + 326, + 92, + 539, + 169 + ], + "spans": [ + { + "bbox": [ + 326, + 92, + 539, + 169 + ], + "type": "table", + "html": "
MethodOptimization (ms/image)Inference (ms/image)
CLIP [37]-1.7
CoOp [61]68.13.0
CoCoOp [60]269.0126.6
MaPLe [22]70.21.7
PromptSRC [23]71.53.8
TAC (Ours)73.24.1
", + "image_path": "6aafa2449730a9dd3960d0044f74cc7561d588171b5e7bdcf3142ab96ae7dddb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 326, + 195, + 539, + 254 + ], + "blocks": [ + { + "bbox": [ + 313, + 174, + 553, + 195 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 553, + 195 + ], + "type": "text", + "content": "Table 9. Ablation on the number of learnable prompts. HM denotes the harmonic mean of the base and novel accuracies." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 195, + 539, + 254 + ], + "lines": [ + { + "bbox": [ + 326, + 195, + 539, + 254 + ], + "spans": [ + { + "bbox": [ + 326, + 195, + 539, + 254 + ], + "type": "table", + "html": "
Number of learnable promptsBaseNovelHM
184.1575.6379.66
284.6776.6180.44
485.2477.6081.24
685.0276.8580.73
", + "image_path": "a001f0a12f8bcf713fa29926b0290eaf88988508bacb4426cf068530e9b32825.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 263, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 335 + ], + "type": "text", + "content": "tity of additional input embeddings incurred by the task-aware pre-context is small, the extra computational overhead incurred by our TAC is marginal compared to PromptSRC [23]. Moreover, it is worth noting our TAC outperforms PromptSRC with noticeable performance gains in all experiment settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": "Number of learnable prompts. We ablate the effect of the number of learnable prompts by setting " + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "inline_equation", + "content": "M^t" + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "inline_equation", + "content": "M^v" + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": " to different values. It is worth noting that (a) " + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "inline_equation", + "content": "M^t" + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "inline_equation", + "content": "M^v" + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": " are set to be equal by default and (b) the other ingredients are kept in this experiment. Table 9 shows the results. Our observations are as follows: (a) increasing the number of learnable prompts can improve the harmonic mean and (b) the greatest performance gain is achieved when " + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "inline_equation", + "content": "M^t = 4" + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "inline_equation", + "content": "M^v = 4" + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": ". These observations lead us to set the default value of " + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "inline_equation", + "content": "M^t" + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "inline_equation", + "content": "M^v" + }, + { + "bbox": [ + 313, + 335, + 556, + 455 + ], + "type": "text", + "content": " to 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 465, + 388, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 388, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 388, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 485, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 715 + ], + "type": "text", + "content": "Prompt learning has emerged as a powerful tool for adapting vision-language models. However, due to the scarcity of task-specific data, the resulting prompts often exhibit limited task-awareness. In this work, we propose to increase the task-awareness by introducing task-aware pre-context, which results in a novel task-aware clustering framework for prompting. The key ingredients are three-fold. First, the backbone structure of a downstream task is preserved by task-aware clustering, based on which the task-aware pre-context is generated. Second, the task-awareness of learnable prompts is enhanced by allowing them to interact with task-aware pre-context via the well-pretrained encoders. Finally, the undesired interference of visual task-aware pre-context to patch embeddings is eliminated by masked attention mechanism. Extensive ablation studies confirm the effectiveness of key ingredients. Comparative results in the base-to-novel, domain generalization, and cross-dataset transfer settings show the advantage of our task-aware clustering for prompt learning." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14752" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "content": "This work was supported by Guangdong Major Project of Basic and Applied Basic Research (2023B0303000016), National Natural Science Foundation of China (62206268, U21A20487, 62372440), Shenzhen Technology Project (JCYJ20240813154723031), Shenzhen High-tech Zone Development Special Plan Innovation Platform Construction Project, the proof of concept center for high precision and high resolution 4D imaging, Guangdong Technology Project (2023TX07Z126), and Yunnan Science & Technology Project (202302AD080008, 202305AF150152)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 239, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 239, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 296, + 348 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 350, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 296, + 392 + ], + "type": "text", + "content": "[2] Shtedritski Aleksandar, Rupprecht Christian, and Vedaldi Andrea. What does clip know about a red circle? visual prompt engineering for vlms. In ICCV, pages 11987-11997, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 394, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 295, + 426 + ], + "type": "text", + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446-461, 2014. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 427, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 295, + 470 + ], + "type": "text", + "content": "[4] Adrian Bulat and Georgios Tzimiropoulos. Lasp: Text-to-text optimization for language-aware soft prompting of vision & language models. In CVPR, pages 23232-23241, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 472, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 294, + 514 + ], + "type": "text", + "content": "[5] Yun-Hao Cao, Kaixiang Ji, Ziyuan Huang, Chuanyang Zheng, Jiajia Liu, Jian Wang, Jingdong Chen, and Ming Yang. Towards better vision-inspired vision-language models. In CVPR, pages 13537-13547, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 515, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 294, + 559 + ], + "type": "text", + "content": "[6] Delong Chen, Zhao Wu, Fan Liu, Zaiquan Yang, Shaoqiu Zheng, Ying Tan, and Erjin Zhou. Protoclip: Prototypical contrastive language image pretraining. IEEE TNNLS (Early Access), pages 1-15, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 559, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 592 + ], + "type": "text", + "content": "[7] Eulrang Cho, Jooyeon Kim, and Hyunwoo J. Kim. Distribution-aware prompt tuning for vision-language models. In ICCV, pages 22004-22013, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 592, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 294, + 624 + ], + "type": "text", + "content": "[8] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 294, + 658 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "text", + "content": "[10] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G. Turrisi da Costa, Cees G.M. Snoek, Georgios Tzimiropoulos, and Brais Martinez. Bayesian prompt learning for image-language model generalization. In ICCV, pages 15237-15246, 2023. 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[11] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "type": "text", + "content": "[12] Bordes Florian, Pang Richard Yuanzhe, Ajay Anurag, Li Alexander C, Bardes Adrien, Petryk Suzanne, Manas Oscar, Lin Zhiqiu, Mahmoud Anas, Jayaraman Bargav, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[13] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. IEEE TNNLS (Early Access), pages 1-11, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. J-STARS, 12(7):2217-2226, 2019. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[17] Nisha Huang, Yuxin Zhang, Fan Tang, Chongyang Ma, Haibin Huang, Weiming Dong, and Changsheng Xu. Diff-styler: Controllable dual diffusion for text-driven image stylization. IEEE TNNLS (Early Access), pages 1-14, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[18] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, pages 3061-3071, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "type": "text", + "content": "[19] Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Retrieval-enhanced contrastive vision-text models. In ICLR, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 533 + ], + "type": "text", + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "text", + "content": "[21] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 1, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[22] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[23] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[24] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14753" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[25] Jonathan Krause, Michael Stark, Jia Deng, and Fei-Fei Li. 3d object representations for fine-grained categorization. In ICCV, pages 554–561, 2013. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "text", + "content": "[26] Nilakshan Kunananthaseelan, Jing Zhang, and Mehrtash Harandi. Lavip: Language-grounded visual prompting. In AAAI, pages 2840-2848, 2024. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[27] Gukyeong Kwon, Zhaowei Cai, Avinash Ravichandran, Erhan Bas, Rahul Bhotika, and Stefano Soatto. Masked vision and language modeling for multi-modal representation learning. In ICLR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 217 + ], + "type": "text", + "content": "[28] Marc Lafon, Elias Ramzi, Clément Rambour, Nicolas Audebert, and Nicolas Thome. Gallop: Learning global and local prompts for vision-language models. In ECCV, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "type": "text", + "content": "[29] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J. Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 1, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[30] Fei-Fei Li, Fergus Rob, and Perona Pietro. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR Workshop, pages 178–178, 2004. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 358 + ], + "type": "text", + "content": "[31] Yanxin Long, Jianhua Han, Runhui Huang, Hang Xu, Yi Zhu, Chunjing Xu, and Xiaodan Liang. Fine-grained visual-text prompt-driven self-training for open-vocabulary object detection. IEEE TNNLS (Early Access), pages 1-11, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 393 + ], + "type": "text", + "content": "[32] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In CVPR, pages 5206-5215, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[33] Jiawei Ma, Po-Yao Huang, Saining Xie, Shang-Wen Li, Luke Zettlemoyer, Shih-Fu Chang, Wen-Tau Yih, and Hu Xu. Mode: Clip data experts via clustering. In CVPR, pages 26354-26363, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "text", + "content": "[34] Metaxas Ioannis Maniadis, Tzimiropoulos Georgios, and Paras Ioannis. Divclust: Controlling diversity in deep clustering. In CVPR, pages 3418-3428, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "type": "text", + "content": "[35] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729. IEEE, 2008. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 536 + ], + "type": "text", + "content": "[36] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In " + }, + { + "bbox": [ + 56, + 504, + 294, + 536 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 56, + 504, + 294, + 536 + ], + "type": "text", + "content": ", pages 3498-3505, 2012. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 592 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 593, + 294, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 625 + ], + "type": "text", + "content": "[38] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imagenet? In ICML, pages 5389-5400, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 656 + ], + "type": "text", + "content": "[39] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. In ICLR, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[40] Lavoie Samuel, Kirichenko Polina, Ibrahim Mark, Assran Mido, Wilson Andrew Gordon, Courville Aaron, and Ballas Nicolas. Modeling caption diversity in contrastive vision-language pretraining. In ICML, pages 26070-26084, 2024. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 118 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 118 + ], + "type": "text", + "content": "[41] Tourani Siddharth, Alwheibi Ahmed, Mahmood Arif, and Khan Muhammad Haris. Pose-guided self-training with two-stage clustering for unsupervised landmark discovery. In CVPR, pages 23041-23051, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 118, + 554, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 554, + 152 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 554, + 152 + ], + "type": "text", + "content": "[42] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 153, + 554, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 153, + 554, + 186 + ], + "spans": [ + { + "bbox": [ + 315, + 153, + 554, + 186 + ], + "type": "text", + "content": "[43] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In CVPR, pages 23209-23218, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 188, + 554, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 188, + 554, + 231 + ], + "spans": [ + { + "bbox": [ + 315, + 188, + 554, + 231 + ], + "type": "text", + "content": "[44] Maji Subhransu, Rahtu Esa, Kannala Juho, Blaschko Matthew, and Vedaldi Andrea. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 233, + 554, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 233, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 233, + 554, + 266 + ], + "type": "text", + "content": "[45] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argument: Attribute-guided prompt tuning for vision-language models. In CVPR, pages 28578-28587, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 268, + 554, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 268, + 554, + 301 + ], + "spans": [ + { + "bbox": [ + 315, + 268, + 554, + 301 + ], + "type": "text", + "content": "[46] Yujun Tong, Da Li, Dongliang Chang, Tianwei Cao, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Vision-language subspace prompting, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 302, + 554, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 302, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 302, + 554, + 335 + ], + "type": "text", + "content": "[47] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 337, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 337, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 315, + 337, + 554, + 381 + ], + "type": "text", + "content": "[48] Mengmeng Wang, Jiazheng Xing, Jianbiao Mei, Yong Liu, and Yunliang Jiang. Actionclip: Adapting language-image pretrained models for video action recognition. IEEE TNNLS (Early Access), pages 1-13, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 383, + 554, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 383, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 315, + 383, + 554, + 416 + ], + "type": "text", + "content": "[49] Zihao Wei, Zixuan Pan, and Andrew Owens. Efficient vision-language pre-training by cluster masking. In CVPR, pages 26815-26825, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 418, + 554, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 418, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 315, + 418, + 554, + 460 + ], + "type": "text", + "content": "[50] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 462, + 554, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 462, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 315, + 462, + 554, + 496 + ], + "type": "text", + "content": "[51] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, pages 6757-6767, 2023. 1, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 498, + 554, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 498, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 315, + 498, + 554, + 531 + ], + "type": "text", + "content": "[52] Hantao Yao, Rui Zhang, and Changsheng Xu. Tcp: Textual-based class-aware prompt tuning for visual-language model. In CVPR, pages 23438-23448, 2024. 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 532, + 554, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 532, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 315, + 532, + 554, + 566 + ], + "type": "text", + "content": "[53] Lochman Yaroslava, Olsson Carl, and Zach Christopher. Learned trajectory embedding for subspace clustering. In CVPR, pages 19092-19102, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 567, + 554, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 567, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 315, + 567, + 554, + 609 + ], + "type": "text", + "content": "[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 612, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 612, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 315, + 612, + 554, + 667 + ], + "type": "text", + "content": "[55] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 669, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 669, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 669, + 554, + 713 + ], + "type": "text", + "content": "[56] Yunfei Zhang, Xiaoyang Huo, Tianyi Chen, Si Wu, and Hau San Wong. Exploring intra-class variation factors with learnable cluster prompts for semi-supervised image synthesis. In CVPR, pages 7392-7401, 2023. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14754" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 353 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[57] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In CVPR, pages 28718-28728, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 161 + ], + "type": "text", + "content": "[58] Cairong Zhao, Yubin Wang, Xinyang Jiang, Yifei Shen, Kaitao Song, Dongsheng Li, and Duoqian Miao. Learning domain invariant prompt for vision-language models. IEEE TIP, 33:1348-1360, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 205 + ], + "type": "text", + "content": "[59] Zhaoheng Zheng, Jingmin Wei, Xuefeng Hu, Haidong Zhu, and Ram Nevatia. Large language models are good prompt learners for low-shot image classification. In CVPR, pages 28453-28462, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 294, + 240 + ], + "type": "text", + "content": "[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 241, + 294, + 274 + ], + "type": "text", + "content": "[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 275, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 275, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 275, + 294, + 308 + ], + "type": "text", + "content": "[62] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In CVPR, pages 15659-15669, 2023. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 57, + 309, + 294, + 353 + ], + "type": "text", + "content": "[63] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14755" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_content_list.json b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02d0eaef0b831724af7f36a1f6db15d8c28a3aa6 --- /dev/null +++ b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_content_list.json @@ -0,0 +1,1809 @@ +[ + { + "type": "text", + "text": "Task-Specific Gradient Adaptation for Few-Shot One-Class Classification", + "text_level": 1, + "bbox": [ + 129, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunlong Li Xiabi Liu* Liyuan Pan Yuchen Ren Beijing Institute of Technology", + "bbox": [ + 276, + 180, + 720, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yunlong.li, liuxiabi, liyuan.pan, yuchen. ren}@bit.edu.cn", + "bbox": [ + 243, + 219, + 746, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Optimization-based meta-learning methods for few-shot one-class classification (FS-OCC) aim to fine-tune a metatrained model to classify the positive and negative samples using only a few positive samples by adaptation. However, recent approaches primarily focus on adjusting existing meta-learning algorithms for FS-OCC, while overlooking issues stemming from the misalignment between the cross-entropy loss and OCC tasks during adaptation. This misalignment, combined with the limited availability of one-class samples and the restricted diversity of task-specific adaptation, can significantly exacerbate the adverse effects of gradient instability and generalization. To address these challenges, we propose a novel Task-Specific Gradient Adaptation (TSGA) for FS-OCC. Without extra supervision, TSGA learns to generate appropriate, stable gradients by leveraging label prediction and feature representation details of one-class samples and refines the adaptation process by recalibrating task-specific gradients and regularization terms. We evaluate TSGA on three challenging datasets and a real-world CNC Milling Machine application and demonstrate consistent improvements over baseline methods. Furthermore, we illustrate the critical impact of gradient instability and task-agnostic adaptation. Notably, TSGA achieves state-of-the-art results by effectively addressing these issues.", + "bbox": [ + 89, + 301, + 483, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 710, + 220, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One-class classification (OCC) is an essential task in various fields, e.g., novelty detection and anomaly detection in cybersecurity and medical diagnosis applications [9, 37, 39, 54]. The objective of OCC tasks is to classify new samples as either positive or negative when only positive samples are available during training. While previous OCC methods [1, 7, 16, 17, 31, 32, 36, 40, 41, 45, 46, 48] often require extensive amounts of positive data to train a classifier, few-shot settings frequently involve scenarios where only lim", + "bbox": [ + 89, + 737, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4149e3f1141f615d38beb24c253da9df611c00577843255c648e6f434f3b709d.jpg", + "image_caption": [ + "(a) OC-MAML" + ], + "image_footnote": [], + "bbox": [ + 517, + 266, + 705, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92d5cf94ce507afdbdacbd8b1a1a461b3418c1b5c5e7b2f8c09b9c3b88cbfa42.jpg", + "image_caption": [ + "(b) TSGA", + "Figure 1. An overview of OC-MAML and the proposed TSGA. Both (a) and (b) allow rapid adaptation of the model parameterized by $\\pmb{\\theta}$ , comprising a feature extractor $f$ and a decision maker $g$ . Here, $\\pmb{x}$ and $y$ denote the input data and corresponding label (only positive), respectively. Different from OC-MAML which uses the cross-entropy loss $\\ell$ in adaptation, we employ Loss Network $\\mathcal{N}$ to generate appropriate losses $\\mathcal{L}$ by leveraging the representations $\\boldsymbol{r} = f(\\boldsymbol{x})$ and predictions $\\hat{y} = g(\\boldsymbol{r})$ . Then, we use Adapter Network $\\mathcal{A}$ to obtain task-specific gradients $\\alpha \\nabla_{\\theta} \\mathcal{L}$ and regularization terms $\\beta \\theta$ , which refines the adaptation process." + ], + "image_footnote": [], + "bbox": [ + 712, + 267, + 901, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ited positive examples (i.e., “shots”) are available, as seen in many real-world applications, e.g., only small datasets are available in industrial manufacturing [52]. Humans intuitively perform the few-shot OCC (FS-OCC) tasks, yet machines still struggle to accomplish them effectively.", + "bbox": [ + 511, + 537, + 906, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mainstream approaches to address the FS-OCC rely on meta-learning [20], which generally fall into two categories: 1) metric-based [12, 25, 42] and 2) optimization-based [15] methods. The former seeks to learn discriminative representations that separate positive from negative query samples, while the latter fine-tunes a meta-trained classifier by adaptation (i.e., a few gradient steps) on one-class support samples. Due to more flexible adaptation and better task generalization, optimization-based meta-learning methods demonstrate significant potential for FS-OCC, e.g., One-Class Model-Agnostic Meta-Learning (OC-MAML [15]) is developed based on the popular meta-learning approach MAML[14] by simply adapting the episode generation strategy and aims to find a model initialization that can classify novel query samples after adaptation on the one-class support set (as shown in Figure 1a).", + "bbox": [ + 511, + 613, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following this paradigm, optimization-based approaches primarily focus on adapting existing algorithms (e.g., MetaSGD [28] and MetaOptNet [27]) to FS-OCC tasks, while", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30556", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/106684eeb0e477769b7e698f8b244173a0931fc0a75c987ca16cac73f31a3073.jpg", + "image_caption": [ + "Figure 2. Loss gradient dynamics visualization of OC-MAML and TSGA for meta-validation set at the first adaptation step for 10-shot FS-OCC tasks on miniImageNet. (a) The loss gradient is computed at the adaptation step, and loss variation is measured along the gradient direction. (b) $\\ell_2$ distance between the loss gradient at a given point and the gradients corresponding to different points along the gradient direction. (c) The Lipschitz gradient continuity of loss gradients." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 359, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/09bac1af8ecd89752b3e7762bdee509c02ae54574ecbf7113cfc12b71f25d64b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 85, + 632, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/09a4f4002a77eb80e1aa1fac3b95987d76a5f2cd6d1ffb0ca04a207c54d170ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 85, + 906, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "overlooking the inherent misalignment between the evaluation metrics and OCC objectives in adaptation and lacking regularization terms in the absence of negative support samples. As illustrated in Figure 2c, high values in the effective $\\beta$ smoothness of OC-MAML indicate cross-entropy loss results in inappropriate gradients and unstable model parameter updates for one-class samples. Moreover, the generalization is further limited when novel tasks diverge significantly from the base tasks encountered during training, as OC-MAML applies a task-agnostic adaptation strategy that disregards task-specific meta-knowledge, thereby constraining its full potential [4, 20, 43].", + "bbox": [ + 88, + 348, + 480, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we address the questions raised above by proposing Task-Specific Gradient Adaptation (TSGA), an optimization-based meta-learning framework for dealing with FS-OCC. Figure 1b provides an overview of our TSGA framework. The central idea behind our TSGA is to automatically generate suitable gradients for the one-class support samples and task-specific hyperparameters (i.e., the learning rate $\\alpha$ and weight decay $\\beta$ ) in adaptation. To this end, we design a learnable loss network $\\mathcal{N}$ to generate the losses by leveraging label prediction and feature representation details of one-class samples, based on which the gradient is calculated and an adapter network $\\mathcal{A}$ is proposed to adaptively recalibrate task-specific gradients and regularization terms during adaptation.", + "bbox": [ + 89, + 534, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of TSGA with three public datasets and a real-world CNC Milling Machines application. On average, TSGA demonstrates consistent improvements and achieves state-of-the-art results. In particular, by tackling the gradient instability and task-agnostic adaptation, it improves the classification accuracy of the OC-MAML [15] baseline by $1.53\\%$ and $2.60\\%$ on miniImageNet and CIFAR-FS in the 10-shot setting, respectively. Specifically, TSGA outperforms the state-of-the-art metric-based AHDD [42] by $1.48\\%$ and $1.50\\%$ on miniImageNet", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and CIFAR-FS in the 2-shot setting, respectively (Table 1). Contributions. In summary, our contributions are three-fold. 1) We design a learnable loss network to avoid gradient instability and overfitting caused by the cross-entropy loss for FS-OCC. 2) We adaptively refine the adaptation process, which captures task-specific meta-knowledge for generalization in novel tasks. 3) Our method achieves state-of-the-art performance. Extensive experiments on three challenging datasets and a real-world CNC Milling Machines application demonstrate its effectiveness.", + "bbox": [ + 511, + 348, + 906, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 516, + 653, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. One-class classification", + "text_level": 1, + "bbox": [ + 511, + 541, + 725, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One-class classification (OCC) methods can be divided into discriminative, generative, and self-supervised approaches [9, 37, 39]. Discriminative methods [17, 32, 41, 45] learn feature representations of positive samples and classify novel samples deviating significantly from the extracted positive feature representations as negative, e.g., Deep-SVDD [45] learns a hypersphere to enclose positive sample embeddings within a minimal volume. Generative methods [1, 31, 36, 46, 48] model the underlying distribution of positive data using models, e.g., Generative Adversarial Networks [36, 48] and Auto-Encoders [46]. Negative samples exhibit high reconstruction errors, as models trained exclusively on positive data fail to reconstruct negatives. Self-supervised methods [7, 16, 40] leverage various transforms (e.g., rotation, flipping) to generate pseudolabels and train a classifier to predict the transform applied. A query sample is classified as positive if predictions match the transform and negative otherwise.", + "bbox": [ + 511, + 564, + 906, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Few-shot one-class classification", + "text_level": 1, + "bbox": [ + 511, + 847, + 795, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to high data scarcity, most existing OCC methods are ill-suited for the few-shot setting, e.g., HTDG [49],", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30557", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which combines a generative model and self-supervision using limited positive samples, suffers from complex training pipeline and marginal performance gain. To address FS-OCC tasks more effectively, recent discriminative approaches often leverage pre-trained weight initialization [6] on large-scale datasets [11, 41] (e.g., ImageNet [13]) or adopt the paradigm of meta-learning [12, 25, 42]. Specifically, FewSOME [6] employs Siamese Networks initialized with pre-trained weights and fine-tunes using stop gradient [10] to prevent representational collapse. OWProtoNets [25] is the first method to integrate meta-learning baseline Prototypical Networks [50] with OCC, which computes a positive prototype by averaging support embeddings and designates the origin as the negative prototype. MetaSVDD [12] constructs a minimum-volume hypersphere enclosing support data without the pseudo-negative prototype. Its variant, OC-ProtoNet, avoids complex computations of the hypersphere center by simply averaging the support features. AHDD [42], a hypersphere-based data description with a learnable radius to define the optimal feature space region for positive samples. Due to the flexibility and task adaptation capabilities, optimization-based meta-learning methods have emerged in FS-OCC. OC-MAML [15] introduces a modified episodic sampling strategy of MAML [14] to learn model initializations suited for FS-OCC. Building on this framework, some other methods (e.g., Reptile[33], Meta-SGD[28], MetaOptNet [27]) also adapt to FS-OCC.", + "bbox": [ + 89, + 90, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Loss function learning", + "text_level": 1, + "bbox": [ + 89, + 510, + 302, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The optimization of loss functions constitutes a foundational component of machine learning [18, 35]. Manually designed losses, e.g., cross-entropy loss and focal loss [44], necessitate white-box formulation while remaining tailored to specific objectives by tedious manual effort. In contrast, automated loss function design has gradually attracted attention. Recent approaches [23, 29, 35] optimize neural networks by minimizing surrogate losses that approximate the non-differentiable evaluation metric (e.g., accuracy, F1, AUC, AP). In addition, recent research has focused on meta-learning for loss functions. Under the \"Learning to Teach\" paradigm [18, 21, 30, 55], a teacher dynamically adjusts the parametric loss functions to align with the performance metric, while the student model is updated to minimize the evolving loss during training. Complementary approaches in meta-learning [2, 5, 22] develop metalearners that learn gradient update values or a task-adaptive loss function for rapid adaptation to new tasks.", + "bbox": [ + 89, + 532, + 483, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Task-specific adaptation", + "text_level": 1, + "bbox": [ + 89, + 816, + 313, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conventional few-shot learning (FSL) assumes training and test samples originate from the same domain, a premise of ten unrealistic in real-world applications. Recent advancements [4, 34, 43, 56] focus on enhancing the generaliza", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion of pre-trained models under significant cross-category or cross-domain shifts through test-time adaptation, e.g., CNAPS [43] integrates task-specific adapters via FiLM layers [38] to adapt feature extractor to target tasks. Similarly, ALFA [4] estimates task-adaptive hyperparameters via an auxiliary network to enhance the fast adaptation.", + "bbox": [ + 511, + 90, + 905, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 194, + 604, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elaborate on our proposed TSGA. Before that, we introduce some preliminary concepts about the FS-OCC task and the mechanism of OC-MAML.", + "bbox": [ + 511, + 219, + 905, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 273, + 643, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assume we have an OCC model parameterized by $\\theta$ , which consists of a feature extractor $f$ that outputs a representation for each input image, and a decision maker $g$ that makes predictions for the positive or negative class. We consider a distribution over tasks $p(\\mathcal{T})$ that we want our model to be able to adapt to. Each FS-OCC task $\\mathcal{T}$ is randomly sampled from $p(\\mathcal{T})$ . The details of the FS-OCC episode sampling technique are provided in Supplementary Materials. Our goal is to apply the model to the test task $\\mathcal{T}_i = \\{\\mathcal{S}, \\mathcal{Q}\\}$ , by deriving task-specific knowledge on the $K$ -shot one-class support set $\\mathcal{S} = \\{(x_i^s, y_i^s)\\}_{s=1}^K$ , consisting of $K$ image-label pairs from the positive class, i.e., $y_i^s = 0$ . It is expected that the model can correctly partition the $M$ class-balanced query samples $\\mathcal{Q} = \\{(x_i^q)\\}_{q=1}^M$ to the positive if the prediction $\\hat{y}_i^q = 0$ , or $\\hat{y}_i^q = 1$ otherwise.", + "bbox": [ + 511, + 295, + 906, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAML and OC-MAML. MAML [14] learns a model initialization for few-shot learning. However, MAML requires access to data examples from each class to yield such generalization. OC-MAML [15] adapts MAML for the FS-OCC problem. It first trains the model on the one-class support set from the sampled task $\\mathcal{T}_i$ by taking a few gradient descent steps in the inner loop (as shown in Figure 1a), then updates the model initialization using the loss computed on a class-balanced query set in the outer loop,", + "bbox": [ + 511, + 521, + 905, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {\\theta} _ {i} ^ {\\prime} = \\boldsymbol {\\theta} - \\alpha \\nabla_ {\\boldsymbol {\\theta}} \\ell (\\mathcal {S}; \\boldsymbol {\\theta}), \\\\ \\boldsymbol {\\theta} \\leftarrow \\boldsymbol {\\theta} - \\eta \\nabla_ {\\boldsymbol {\\theta}} \\sum_ {\\mathcal {T} _ {i}} \\ell \\left(\\mathcal {Q}; \\boldsymbol {\\theta} _ {i} ^ {\\prime}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 665, + 906, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\eta$ are learning rates. The cross-entropy loss is used for $\\ell$ . Theoretically, the gradient in the update can be approximated by using Taylor series expansion [15, 33],", + "bbox": [ + 511, + 727, + 905, + 772 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {g} = \\bar {\\boldsymbol {g}} _ {2} - \\alpha \\bar {\\boldsymbol {H}} _ {2} \\bar {\\boldsymbol {g}} _ {1} - \\alpha \\bar {\\boldsymbol {H}} _ {1} \\bar {\\boldsymbol {g}} _ {2} + O \\left(\\alpha^ {2}\\right) \\\\ = \\bar {\\boldsymbol {g}} _ {2} - \\alpha \\frac {\\partial \\left(\\bar {\\boldsymbol {g}} _ {1} \\cdot \\bar {\\boldsymbol {g}} _ {2}\\right)}{\\partial \\boldsymbol {\\theta}} + O \\left(\\alpha^ {2}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 779, + 903, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\overline{\\pmb{g}}_i$ and $\\overline{H}_i$ denote the gradient and Hessian computed on the $i^{th}$ mini-batch (i.e., one-class support samples and class-balanced query samples, respectively) at the initial parameter point $\\theta$ . The key to the effectiveness of OC-MAML", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30558", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lies in maximizing the inner product of the gradients computed on the one-class support set and class-balanced query set, thus taking a gradient step using one yields a performance increase on the other. However, we argue that the gradients on the support set $\\overline{g}_1$ computed by cross-entropy loss are unstable due to the one-class task property. Next, we solve this problem by proposing our TSGA.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 89, + 207, + 202, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our TSGA is illustrated in Figure 1b, which mainly consists of the following steps. For each iteration:", + "bbox": [ + 89, + 229, + 482, + 258 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Step-1. A feature extractor $f$ takes the images $\\boldsymbol{x}_i$ from $S$ as input to obtain image representations $\\boldsymbol{r}_i = f(\\boldsymbol{x}_i)$ and a decision maker $g$ transforms the representations $\\boldsymbol{r}_i$ to predictions $\\hat{y}_i = g(\\boldsymbol{r}_i)$ .", + "Step-2. A loss network $\\mathcal{N}$ parameterized by $\\phi$ takes image representations and predictions as input to generate the loss values, based on which we can obtain the gradients with respect to the model by backpropagation.", + "Step-3. An adapter network $\\mathcal{A}$ parameterized by $\\varphi$ recalibrates the gradients and parameters of the model to task-specific updated gradients and regularization terms, which remodel the update rules in the adaptation.", + "Step-4. In the outer loop, a weighted sum of cross-entropy loss concerning the class-balanced query set is minimized after every adaptation step to update each module mentioned above." + ], + "bbox": [ + 89, + 260, + 483, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Instance-wise Loss Network", + "text_level": 1, + "bbox": [ + 89, + 513, + 341, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The motivation of the loss network $\\mathcal{N}$ is that the cross-entropy loss, which is applied for rapid adaptation in OC-MAML, is not aligned with evaluation metrics in the OCC tasks, leading to unsuitable gradients and unstable model parameter updates. Designing loss functions manually on existing problems is resource intensive, so we introduce a learnable loss network $\\mathcal{N}$ to make the gradient more stable.", + "bbox": [ + 89, + 536, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the support set $\\mathcal{S} = \\{(x_i^s, y_i^s)\\}_{s=1}^K$ of a mettraining task $\\mathcal{T}_i$ , each image $\\pmb{x}$ is fed into feature extractor $f$ to generate image representation $\\pmb{r}$ , and then is transformed to prediction $\\hat{y}$ by decision maker $g$ . Next, the prediction $\\hat{y}$ together with the corresponding representation $\\pmb{r}$ is input into $\\mathcal{N}$ to generate the loss value, formulated as", + "bbox": [ + 89, + 642, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta , \\mathcal {S}; \\phi) = \\frac {1}{K} \\sum_ {(\\boldsymbol {x}, y \\in \\mathcal {S})} \\mathcal {N} (\\hat {y}, \\boldsymbol {r}, y; \\phi). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 744, + 482, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the support data in FS-OCC tasks exclusively belonging to the positive class, the label $y = 0$ can be ignored. We concatenate representation and prediction as input. For simplicity, we employ a 4-layer Multi-Layer Perceptron (MLP) with ELU activation between the layers for stable gradients since it is $C^\\infty$ continuous and the hidden size of $\\mathcal{N}$ is set to 4.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Task-Specific Adapter Network", + "text_level": 1, + "bbox": [ + 513, + 90, + 790, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aside from generating learnable supervision for stable gradients, we customize the adaptation process for generalization in different test tasks $\\mathcal{T}_i$ . Intuitively, the direct way to determine the adaptation is to learn the coefficient (i.e., the learning rate $\\alpha$ and weight decay $\\beta$ [4, 28]) for the adaptation,", + "bbox": [ + 511, + 112, + 906, + 200 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {\\theta} _ {i} ^ {\\prime} = \\beta \\boldsymbol {\\theta} - \\alpha \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, S; \\phi) \\\\ = \\left[ \\boldsymbol {\\beta}, \\boldsymbol {\\alpha} \\right] \\left[ \\begin{array}{c} \\boldsymbol {\\theta} \\\\ - \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, \\mathcal {S}; \\phi) \\end{array} \\right], \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 202, + 903, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ and $\\alpha$ are layer-wise weight decay and learning rate, respectively, for the computational efficiency.", + "bbox": [ + 511, + 258, + 903, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The adapter network is thus developed to achieve task-specific adaptation by gating mechanism, which can be regarded as a self-attention function to capture task-specific contextual information and adaptively recalibrate the parameters and gradients,", + "bbox": [ + 511, + 287, + 905, + 363 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\boldsymbol {\\beta}, \\boldsymbol {\\alpha} ] = \\mathcal {A} \\left(\\left[ \\begin{array}{c} \\boldsymbol {\\theta} \\\\ - \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, \\mathcal {S}; \\boldsymbol {\\phi}) \\end{array} \\right]; \\varphi\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 372, + 903, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ average pooling to the parameters and gradients per layer and concatenate these two $L$ -dimension vectors (i.e., $2L$ -dimension) to feed into $\\mathcal{A}$ , where $L$ is the number of the model's layer. We employ a Transformer Encoder model as the adapter network and remove all normalization layers. The learned relative positional encoding is used to utilize the position information. The details of $\\mathcal{A}$ are discussed in Section 4.3.", + "bbox": [ + 511, + 415, + 905, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Training process", + "text_level": 1, + "bbox": [ + 513, + 544, + 678, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the outer loop, we use a weighted sum of cross-entropy loss on the class-balanced query samples after every adaptation step $t$ to improve gradient propagation and reduce gradient instability [3]. More formally:", + "bbox": [ + 511, + 566, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell \\left(\\mathcal {Q}; \\boldsymbol {\\theta} _ {i} ^ {\\prime}\\right) = \\mathbb {E} _ {\\ell} \\left[ \\sum_ {t = 1} ^ {T} w _ {t} \\ell \\left(\\boldsymbol {\\theta} _ {i} ^ {t}\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 637, + 903, + 684 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell \\left(\\boldsymbol {\\theta} _ {i} ^ {t}\\right) = \\frac {1}{M} \\sum_ {\\left(\\boldsymbol {x}, y \\in \\mathcal {Q}\\right)} \\ell (\\hat {y}, y; \\boldsymbol {\\theta} _ {i} ^ {t}), t \\in [ 1, T ],\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 680, + 857, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{t}$ denotes the importance weight of the query set loss at step $t$ and $\\pmb{\\theta}_{i}^{t}$ the parameters after every step $t$ ,", + "bbox": [ + 511, + 728, + 906, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {i} ^ {0} = \\boldsymbol {\\theta},\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 768, + 591, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {i} ^ {t} = \\beta \\boldsymbol {\\theta} _ {i} ^ {t - 1} - \\alpha \\nabla_ {\\boldsymbol {\\theta} _ {i} ^ {t - 1}} \\mathcal {L} \\left(\\boldsymbol {\\theta} _ {i} ^ {t - 1}, S; \\phi\\right), t \\in [ 1, T ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 787, + 903, + 810 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {i} ^ {\\prime} = \\boldsymbol {\\theta} _ {i} ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 811, + 604, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ is the number of adaptation steps.", + "bbox": [ + 511, + 840, + 795, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initially, each loss contributes equally to the overall loss. However, as iterations progress, the contribution from earlier steps is gradually reduced, while the contribution from", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30559", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Meta-training of TSGA" + ], + "code_body": "Require: Task distribution $p(\\mathcal{T})$ \nRequire: Learning rates $\\alpha ,\\eta$ \nRequire: Batch size for the inner and outer updates $K,M$ 1: Randomly initialize $\\theta ,\\phi$ \n2: while not done do \n3: Sample batch of tasks $\\mathcal{T}_i\\sim p(\\mathcal{T})$ \n4: for each sampled task $\\mathcal{T}_i$ do \n5: Sample one-class support set $\\mathcal{S} = \\{(x_i^s,y_i^s)\\}_{s = 1}^K$ from $\\mathcal{T}_i$ \n6: Sample class-balanced query set $\\mathcal{Q} = \\{(x_i^q,y_i^q)\\}_{q = 1}^M$ from $\\mathcal{T}_i$ \n7: Initialize $\\theta_{i}^{\\prime} = \\theta$ \n8: for number of adaptation steps do \n9: Compute loss $\\mathcal{L}(\\theta ,\\mathcal{S};\\phi)$ via Eq. (3). \n10: Compute hyperparameters $\\beta ,\\alpha$ via Eq. (5). \n11: Perform gradient descent to adapt $\\theta_{i}^{\\prime}$ via Eq. (7). \n12: end for \n13: Compute cross-entropy loss $\\ell (\\mathcal{Q};\\theta_i^{\\prime})$ via Eq. (6) \n14: end for \n15: Perform gradient descent to update $(\\theta ,\\phi ,\\varphi)\\gets (\\theta ,\\phi ,\\varphi) - \\eta \\nabla_{(\\theta ,\\phi ,\\varphi)}\\sum_{\\mathcal{T}_i}\\ell (\\mathcal{Q};\\theta_i^{\\prime}).$ \n16: end while", + "bbox": [ + 91, + 107, + 483, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "later steps is increased to update $\\theta$ , $\\phi$ and $\\varphi$ as in,", + "bbox": [ + 89, + 470, + 421, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {\\theta}, \\phi , \\varphi\\right) \\leftarrow \\left(\\boldsymbol {\\theta}, \\phi , \\varphi\\right) - \\eta \\nabla_ {\\left(\\boldsymbol {\\theta}, \\phi , \\varphi\\right)} \\sum_ {\\mathcal {T} _ {i}} \\ell \\left(\\mathcal {Q}; \\boldsymbol {\\theta} _ {i} ^ {\\prime}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 497, + 482, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall training procedure of our method is summarized in Algorithm 1.", + "bbox": [ + 89, + 537, + 482, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As an optimization-based meta-learning method inherited the advantages of model-agnostic from OC-MAML, our TSGA does not impose any constraint on the model and its target applications.", + "bbox": [ + 89, + 569, + 482, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computation cost. The overall number of learnable parameters increased from OC-MAML, together with the loss network $\\mathcal{N}$ and adapter network $\\mathcal{A}$ , is negligible. Besides, in the bi-level optimization scheme, the training time and computation complexity for TSGA and OC-MAML are roughly the same.", + "bbox": [ + 89, + 628, + 482, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 733, + 220, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 757, + 277, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our approach on three standard benchmarks for FS-OCC: Omniglot [26], miniImageNet [53], and CIFAR-FS [8], with additional experiments on the real-world CNC Milling Machine dataset (The details are in Section 4.2). Omniglot [26] comprises 1,623 handwritten characters from 50 alphabets. All images are resized to $28 \\times 28$ pixels and partitioned into 25 alphabets for meta-training, 5 for meta-validation, and 20 for", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "meta-testing. miniImageNet [53] includes 50,000 training and 10,000 test images $(84 \\times 84$ pixels), evenly distributed across 100 classes. We partition the dataset into 64 classes for meta-training, 16 for meta-validation, and 20 for metatesting. CIFAR-FS [8] is derived from CIFAR-100 [24] via random sampling, with 600 images per class $(32 \\times 32$ pixels). We partition the dataset into 64 classes for metatraining, 16 for meta-validation, and 20 for meta-testing.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We followed the experimental setups used by OC-MAML [15], including hyperparameter setting, data preprocessing, and network backbones. We employ two widely adopted architectures for few-shot learning [34, 50, 51, 53]: Convolutional Neural Network (CNN) with 4 convolutional blocks (Conv-4) and ResNet-12 [19], as established in prior work. Architectural details for Conv-4 and ResNet-12 are provided in the Supplementary Materials. Unless stated otherwise, Conv-4 serves as the default backbone. In the inner loop, we optimize the classifier parameters using the SGD optimizer for 5 adaptation steps. In the outer loop, we use the Adam optimizer with a learning rate of 0.001, $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ . During meta-training, each iteration contains a batch of FS-OCC tasks (we set batch size as 4) with training conducted over 20,000 iterations. Validation is performed every 100 iterations by evaluating performance on 160 randomly sampled FS-OCC tasks from the meta-validation set. The best model on the meta-validation set is selected for final evaluation during meta-testing.", + "bbox": [ + 511, + 212, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We verify the effectiveness of TSGA against a comprehensive range of FS-OCC methods, including 1) optimization-based methods: OC-MAML [15], OC-MetaOptNet, OC-MetaSGD and 2) metric-based methods: OC-ProtoNet [12], OW-ProtoNet [25], AHDD [42]. We directly cited the results of OW-ProtoNet and OC-ProtoNet from OC-MAML [15] and AHDD [42], which reproduced these results under a consistent experimental setup.", + "bbox": [ + 511, + 515, + 905, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metric. We evaluate our method on 600 unique FS-OCC tasks randomly sampled from the meta-testing set for each dataset and report average accuracy (in %) in $95\\%$ confidence intervals. We ensure reliable evaluation by conducting 5 random restarts, using average results for comprehensive comparisons.", + "bbox": [ + 511, + 637, + 905, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 513, + 739, + 607, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we perform extensive experiments to demonstrate the effectiveness of our TSGA. The analysis of variance (ANOVA) results are provided in Supplementary Materials, with all pairwise comparisons achieving statistical significance $(p < 0.05)$ .", + "bbox": [ + 511, + 763, + 903, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline comparison. To validate the effectiveness of TSGA for FS-OCC, we conduct experiments on three datasets with baseline approaches, as shown before. The quantitative results of the baseline methods and ours with", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30560", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4ea4b39d3becff03ac1c78eb874a0c20e5bd9931c414d0b3f95b551068164fd4.jpg", + "table_caption": [], + "table_footnote": [ + "† Our reproduction." + ], + "table_body": "
ModelBackboneK=2K=10
MINOMNCIFAR-FSMINOMNCIFAR-FS
OC-ProtoNet [12]Conv-464.488.864.566.491.667.6
OW-ProtoNet [25]Conv-467.0-70.974.4-76.7
AHDD [42]Conv-471.9396.8175.3477.0097.8780.95
OC-MetaOptNet [15]ResNet-1251.8-56.367.4-75.5
OC-MetaSGD [15]Conv-469.6-71.475.8-77.8
OC-MAML [15]Conv-469.196.670.076.297.679.1
TSGA (Ours)Conv-473.41 ± 0.498.11 ± 0.176.84 ± 0.477.73 ± 0.498.54 ± 0.181.70 ± 0.4
OC-MAML†[15]ResNet-1276.60 ± 0.498.61 ± 0.174.26 ± 0.477.27 ± 0.499.18 ± 0.182.05 ± 0.4
TSGA (Ours)ResNet-1278.04 ± 0.498.97 ± 0.179.53 ± 0.479.24 ± 0.499.29 ± 0.184.05 ± 0.4
", + "bbox": [ + 93, + 88, + 906, + 290 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e0adfbfd7d048564cdbec9cd2e2a7522c17193e8274caa5f7fe541d3490490b3.jpg", + "table_caption": [ + "Table 1. Comparison with the metric-based and optimization-based meta-learning methods for 2-shot and 10-shot OCC tasks on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS. We highlight the best number in **bold**." + ], + "table_footnote": [ + "† Our reproduction." + ], + "table_body": "
ModelMINOMNCIFAR-FS
OC-ProtoNet [12]61.8986.0262.15
OW-ProtoNet [25]62.0094.90-
AHDD [42]68.8995.1671.89
OC-MAML†[15]69.6297.2173.76
TSGA (Ours)71.6797.3274.77
", + "bbox": [ + 93, + 369, + 480, + 484 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/033abea0e75b2d30e224525b0241bbe67fa715fd94544f09c9e78a69415511dc.jpg", + "table_caption": [ + "Table 2. One-shot OCC tasks comparison results of ours and baseline methods on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS. We highlight the best number in bold." + ], + "table_footnote": [], + "table_body": "
ModelOP00OP01OP02Avg
OC-MAML75.5665.6395.1578.78
TSGA76.0566.6597.4580.05
", + "bbox": [ + 93, + 571, + 480, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "various training sample numbers ( $K$ -shot) are reported in Tables 1 and 2. We can observe from the results: 1) TSGA consistently improves the baseline OC-MAML by a large margin in all settings, demonstrating its effectiveness in improving gradient stability and model generalization to FS-OCC. 2) TSGA achieves significant performance gains on both the Conv-4 and ResNet-12 backbones, suggesting TSGA's flexibility. 3) TSGA can tackle both the one-shot and few-shot FS-OCC tasks, whereas several FS-OCC baselines (e.g., Meta-SVDD [12]) are unsuitable for one-shot OCC, qualitative results are shown in Section 4.4. State-of-the-art comparison. Here we compare our TSGA", + "bbox": [ + 89, + 717, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e7cc0348cc0324e080a61323d8bcc1bb3fc5fa534eeda516e83104c944acd99f.jpg", + "table_caption": [ + "Table 3. Comparison of OC-MAML and TSGA on several operations (i.e., OP) of the real-world CNC Milling Machine dataset, with only $K = 1$ normal examples." + ], + "table_footnote": [], + "table_body": "
ModelMINOMN
MAML [14]0.280.16
OC-MAML [15]0.420.23
TSGA (Ours)0.470.39
", + "bbox": [ + 568, + 369, + 849, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Comparison of cosine similarity between the gradients of one-class support set and class-balanced query set averaged over test tasks of miniImageNet (MIN) and Omniglot (OMN).", + "bbox": [ + 511, + 458, + 906, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with the state-of-the-art metric-based method AHDD [42] and conduct experiments on three datasets. AHDD learns radius $r$ , which is involved in the model optimization. This can be regarded as a type of loss function learning. During the meta-testing phase, the radius remains unchanged while our TSGA further captures the task-specific knowledge to refine the adaptation process. In Table 1, we can observe that TSGA considerably outperforms the strong baseline AHDD and establishes new state-of-the-art results in all settings, which further confirms the effectiveness of our TSGA and the mechanism of adaptation to the unseen test tasks of optimization-based methods.", + "bbox": [ + 511, + 534, + 906, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results of real-world CNC Milling Machine dataset. We conduct experiments on the industrial process monitoring of brownfield milling machines dataset [52] and report the results in Table 3. We compute accuracy for evaluation. The dataset consists of 1702 acceleration data built with 15 different tool operations that run on 3 machines at different time frames. The industrial dataset has an unbalanced rate of 23.3:1 between normal (i.e., positive) and non-normal data. Each data can be seen as a multivariate time series (three signals, i.e., $x, y, z$ -axis). We created samples by randomly cropping windows of length 2048 from the corresponding time series of each operation and partition oper-", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30561", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3cd251a9873e5c497015fcf71cf325a1376e053104617b15325780239c541b6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDSettingK=2K=10
MINOMNCIFAR-FSMINOMNCIFAR-FS
ABaseline69.196.670.076.297.679.1
B+A72.1397.8975.9176.4698.3181.11
C+N72.0997.9975.5976.6698.4581.13
D+A+N(TSGA)73.41 (+4.31)98.11 (+1.51)76.84 (+6.84)77.73 (+1.53)98.54 (+0.94)81.70 (+2.60)
", + "bbox": [ + 91, + 88, + 903, + 208 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3b2217fef662c7563f87943b12bfb30e7e8dd8f4258675619bbdd4b7a736af4b.jpg", + "table_caption": [ + "Table 5. Ablation studies for the designed components of TSGA (including the loss network $\\mathcal{N}$ and adapter network $\\mathcal{A}$ ) on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS for 2-shot and 10-shot FS-OCC tasks." + ], + "table_footnote": [], + "table_body": "
Cross-EntropyLoss NetworkK=2K=10
FeatureLabelMINOMNCIFAR-FSMINOMNCIFAR-FS
✓XX69.196.670.076.297.679.1
✓X✓71.13 ± 0.497.95 ± 0.176.66 ± 0.477.60 ± 0.498.27 ± 0.181.11 ± 0.4
✓✓✓72.15 ± 0.497.91 ± 0.177.11 ± 0.477.19 ± 0.498.60 ± 0.180.66 ± 0.4
XX✓72.61 ± 0.498.01 ± 0.176.18 ± 0.477.31 ± 0.498.45 ± 0.181.46 ± 0.4
X✓✓73.41 ± 0.498.11 ± 0.176.84 ± 0.477.73 ± 0.498.54 ± 0.181.70 ± 0.4
", + "bbox": [ + 91, + 260, + 906, + 412 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1b21abd0079701b4b5a6e893c18b691b9162244f5159b8c21932c2c0df8ca523.jpg", + "table_caption": [ + "Table 6. The impact of loss network components (Feature and Label represent feature representations and label predictions, respectively) and cross-entropy loss on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS for 2-shot and 10-shot OCC tasks." + ], + "table_footnote": [], + "table_body": "
# Layer# DimAccuracy# Layer# DimAccuracy
1481.604481.28
881.84880.44
2481.398481.47
881.99880.65
", + "bbox": [ + 91, + 474, + 480, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Different adapter networks results on CIFAR-FS in 10-shot OCC tasks. # Layer and # Dim represent the number of layers and the hidden dimension, respectively.", + "bbox": [ + 89, + 582, + 482, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ations into 10 types for meta-training, 2 for meta-validation, and 3 for meta-testing. TSGA consistently outperforms OC-MAML across the 3 types of operations and achieves $80\\%$ accuracy on average by using only $K = 1$ normal examples, demonstrating its competence in the time-series data domain and for industrial real-world applications.", + "bbox": [ + 89, + 655, + 482, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cosine Similarity Analysis. The success of OC-MAML lies in maximizing the inner product, and therefore the cosine similarity, of the gradients computed on the one-class support set and class-balanced query set $\\frac{\\partial(\\overline{g}_1\\cdot\\overline{g}_2)}{\\partial\\theta}$ . Following OC-MAML, for each task, we use the initialization meta-learned by each algorithm to compute the loss gradients of support and query examples, and then compute the average cosine similarity across tasks. The cosine similarity of MAML, OC-MAML, and TSGA on test tasks from two datasets is reported in Table 4, showing that TSGA achieves", + "bbox": [ + 88, + 747, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "higher similarity values, indicating its ability to effectively align support and query gradients. This validates the improved task-specific adaptation capabilities of TSGA.", + "bbox": [ + 511, + 476, + 906, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 531, + 663, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we conduct ablative analyses to investigate the designed components of TSGA in Table 5. We also study the impact of different supervision caused by loss network and adapter network architectures on model performance in Tables 6 and 7. Please refer to Supplementary Materials for more other details.", + "bbox": [ + 511, + 551, + 906, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the designed components. TSGA contains two components, including a loss network $\\mathcal{L}$ and an adapter network $\\mathcal{A}$ . We conduct component-wise analysis by alternatively adding one of them to understand the influence of each component in Table 5. We take the OC-MAML as baseline (A). $\\mathbb{B}$ or $\\mathbb{C}$ : only leverage loss network or adapter network. $\\mathbb{D}$ : our TSGA. As can be seen, each component in TSGA contributes to the performance. In particular, the results in $\\mathbb{D}$ suggest that the loss network and the adapter network complement each other to improve the FS-OCC performance. The results in $\\mathbb{B}$ and $\\mathbb{C}$ verify the effectiveness of learnable supervision for OCC tasks, and task-specific adaptation for various tasks, respectively.", + "bbox": [ + 511, + 643, + 908, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of the loss network. TSGA leverages label predictions and feature representations to learn loss values. It is important to answer the question: Does the feature part matter? To this end, we remove the feature representations.", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30562", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecd27b087e872ef4dea8303d476947b3170d542e0ae17bf4a699530805997fd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 89, + 254, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3eb8004aa2a5611b6d826fa8b2d4005351bdb76714226f1247683ee4d1b7d20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 90, + 416, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8012c05a1f691ed37b64a74a01c46fd0dda752e1667d463e6ef132fa0ec3c56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 90, + 578, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1ed004893d2036f21f8662bea6fed907aa7744b8412c76002cec877c30d47648.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 90, + 741, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e97212de60e6520feb265e84d5955ea4c25ee714318d0aef562572335cfb0e53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 90, + 903, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8e28e474bb0a02b4bdf0f9b604f712ac5f4c1e92be6e0eace92733ee95c138da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 189, + 254, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/648a4aa6feacef866d7a478db3e3c41609bc2b7eaa4b4cdcc9a7c9494eda481f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 189, + 416, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b561d026264bbbffc42b4286688c7c43a336561faab3943b2b800267807650b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 189, + 578, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ced786a4f3f6ef9b6b4ae55883294d29199db71d4cc7030261c4ccb819f60608.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 189, + 741, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/430d165be57993f87364065ed39d27f353bcede59fc5ae011ef510bc91664ead.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 189, + 903, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e85460941e7c1b0aa7b99b0c76b84e3c9594ca273b9a13cc188ab3fb64ba462b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 285, + 254, + 381 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7fb17025fc6feef622e851cf50470e22508651ca6fea99cba86199c793f7b684.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 285, + 416, + 381 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/acd905ba7a3ae03bb6eff2bedb42fb0d990e5b1d467757fb65dbc2a35e1fb2f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 285, + 578, + 381 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a6825eb0c9354d0e4d3bb55f6d513295bdc1756ac003190a08edab76df3107a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 285, + 740, + 381 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b69de7149a731f120dcdf80e743626d6a3dcb1baaf092a0b90ec26f9fd28a41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 285, + 903, + 381 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/039ee51b641a54a7b0207600b179790b3c35976624f9128bc60981811032b4a4.jpg", + "image_caption": [ + "Figure 3. Visualization of the generated values $\\alpha$ and $\\beta$ of adapter network, across adaptation steps and model layers (i.e., weights and biases of 4 convolutional layers and 1 fully connected layer in CONV-4) for 10-shot OCC tasks on miniImageNet." + ], + "image_footnote": [], + "bbox": [ + 94, + 381, + 254, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32618484f2e8033e40559bc1d3bcf28471ebd6cc0b9016692293476d1db05e0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 381, + 416, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b88c95f34af3c38c11131af22f33af961c398276715c90e0f8c1b6e5f766f16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 381, + 578, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b824d6165523811fba2f2ad7e883681e6ddcedcbc150ccec9009b4658847ea72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 381, + 740, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8614f4ba23e82d2599332cb85a33c8bb9ad5c4abe82d8ebdca70715cb9365cb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 381, + 903, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are reported in Table 6. Not surprisingly, without feature representations, performance gains decline. Further, when the cross-entropy loss and loss network are used together, there is no significant difference from when only the latter is used, implying that our loss network can comprehend the cross-entropy loss information.", + "bbox": [ + 88, + 545, + 482, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of the adapter network. We study the impacts of the number of layers and the hidden size of the adapter network on performance. Table 7 shows that a larger number of layers and a bigger hidden size do not result in significant performance gains. TSGA is in general not sensitive to the choice of the number of layers and the hidden size within a certain range.", + "bbox": [ + 88, + 637, + 482, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualization", + "text_level": 1, + "bbox": [ + 89, + 756, + 227, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, we provide qualitative visualizations to see how our method works. In Figure 2, we present the visualization of the loss gradient dynamics [47], evaluating optimization stability and smoothness by measuring loss variations, changes in gradients, and the maximum difference in gradients. As observed, TSGA exhibits lighter shades in Figures 2a and 2b and lower values in Figure 2c, which indicate a smoother loss landscape. In Figure 3, we show the", + "bbox": [ + 88, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "task-specific hyperparameters generated by the adapter network averaged across tasks, which adaptively vary across the inner-loop steps and model layers. The variability of the generated values highlights the effectiveness of our TSGA in refining optimization processes for the given task. More visualization of training curves is in Supplementary Materials.", + "bbox": [ + 511, + 545, + 906, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose TSGA for few-shot one-class classification (FS-OCC). TSGA learns to generate stable and smooth gradients by leveraging label prediction and feature representation details and enhances the adaptation process by recalibrating task-specific gradients and regularization terms. We evaluate TSGA on three challenging datasets and a real-world CNC Milling Machine dataset, demonstrating that TSGA consistently improves performance. We also reveal the severity of unstable gradients and ignoring of task-specific adaptation, by tackling these issues TSGA establishes new state-of-the-art results.", + "bbox": [ + 511, + 689, + 906, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the National Natural Science Foundation of China (82171965, 62302045).", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30563", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Latent space autoregression for novelty detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 481-490, 2019. 1, 2", + "[2] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. Advances in neural information processing systems, 29, 2016. 3", + "[3] Antreas Antoniou, Harri Edwards, and Amos Storkey. How to train your maml. In Seventh International Conference on Learning Representations, 2019. 4", + "[4] Sungyong Baik, Myungsub Choi, Janghoon Choi, Heewon Kim, and Kyoung Mu Lee. Meta-learning with adaptive hyperparameters. Advances in neural information processing systems, 33:20755-20765, 2020. 2, 3, 4", + "[5] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9465-9474, 2021. 3", + "[6] Niamh Belton, Misgina Tsighe Hagos, Aonghus Lawlor, and Kathleen M Curran. Fewsome: One-class few shot anomaly detection with siamese networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2978-2987, 2023. 3", + "[7] Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In International Conference on Learning Representations. 1, 2", + "[8] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 5", + "[9] Raghavendra Chalopathy and Sanjay Chawla. Deep learning for anomaly detection: A survey. arXiv preprint arXiv:1901.03407, 2019. 1, 2", + "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3", + "[11] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 3", + "[12] Gabriel Dahia and Maurício Pamplona Segundo. Meta learning for few-shot one-class classification. AI, 2(2):195-208, 2021. 1, 3, 5, 6", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3", + "[14] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 1, 3, 6" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Ahmed Frikha, Denis Krompaß, Hans-Georg Köpken, and Volker Tresp. Few-shot one-class classification via meta-learning. In Proceedings of the AAAI conference on artificial intelligence, pages 7448–7456, 2021. 1, 2, 3, 5, 6", + "[16] Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations. Advances in neural information processing systems, 31, 2018. 1, 2", + "[17] Sachin Goyal, Aditi Raghunathan, Moksh Jain, Harsha Vardhan Simhadri, and Prateek Jain. Drocc: Deep robust one-class classification. In International conference on machine learning, pages 3711-3721. PMLR, 2020. 1, 2", + "[18] Zhaoyang Hai, Liyuan Pan, Xiabi Liu, Zhengzheng Liu, and Mirna Yunita. L2t-dln: Learning to teach with dynamic loss network. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[20] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5149-5169, 2021. 1, 2", + "[21] Chen Huang, Shuangfei Zhai, Walter Talbott, Miguel Bautista Martin, Shih-Yu Sun, Carlos Guestrin, and Josh Susskind. Addressing the loss-metric mismatch with adaptive loss alignment. In International conference on machine learning, pages 2891-2900. PMLR, 2019. 3", + "[22] Chen Huang, Shuangfei Zhai, Pengsheng Guo, and Josh Susskind. Metricopt: Learning to optimize black-box evaluation metrics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 174-183, 2021. 3", + "[23] Tao Huang, Zekang Li, Hua Lu, Yong Shan, Shusheng Yang, Yang Feng, Fei Wang, Shan You, and Chang Xu. Relational surrogate loss learning. In International Conference on Learning Representations. 3", + "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[25] Anna Kruspe. One-way prototypical networks. arXiv preprint arXiv:1906.00820, 2019. 1, 3, 5, 6", + "[26] Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332-1338, 2015. 5", + "[27] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 1, 3", + "[28] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1, 3, 4", + "[29] Lanlan Liu, Mingzhe Wang, and Jia Deng. A unified framework of surrogate loss by refactoring and interpolation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 278-293. Springer, 2020. 3" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30564", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Qingliang Liu and Jinmei Lai. Stochastic loss function. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4884-4891, 2020. 3", + "[31] Wenrui Liu, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Diversity-measurable anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12147-12156, 2023. 1, 2", + "[32] Philipp Liznerski, Lukas Ruff, Robert A Vandermeulen, Billy Joe Franks, Marius Kloft, and Klaus Robert Muller. Explainable deep one-class classification. In International Conference on Learning Representations. 1, 2", + "[33] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 3", + "[34] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 3, 5", + "[35] Yash Patel, Tomáš Hodaž, and Jiří Matas. Learning surrogates via deep embedding. In European Conference on Computer Vision, pages 205-221. Springer, 2020. 3", + "[36] Pramuditha Perera, Ramesh Nallapati, and Bing Xiang. Ocean: One-class novelty detection using gans with constrained latent representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2898-2906, 2019. 1, 2", + "[37] Pramuditha Perera, Poojan Oza, and Vishal M Patel. One-class classification: A survey. arXiv preprint arXiv:2101.03064, 2021. 1, 2", + "[38] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3", + "[39] Marco AF Pimentel, David A Clifton, Lei Clifton, and Lionel Tarassenko. A review of novelty detection. Signal processing, 99:215-249, 2014. 1, 2", + "[40] Chen Qiu, Timo Pfrommer, Marius Kloft, Stephan Mandt, and Maja Rudolph. Neural transformation learning for deep anomaly detection beyond images. In International conference on machine learning, pages 8703-8714. PMLR, 2021. 1, 2", + "[41] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 1, 2, 3", + "[42] Yuchen Ren, Xiabi Liu, Liyuan Pan, and Lijuan Niu. Adaptive hypersphere data description for few-shot one-class classification. Applied Intelligence, pages 1-13, 2024. 1, 2, 3, 5, 6", + "[43] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E Turner. Fast and flexible multi-task classification using conditional neural adaptive processes. Advances in neural information processing systems, 32, 2019. 2, 3", + "[44] T-YLPG Ross and GKHP Dollár. Focal loss for dense object detection. In proceedings of the IEEE conference on" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition, pages 2980-2988, 2017.3", + "[45] Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. In International conference on machine learning, pages 4393-4402. PMLR, 2018. 1, 2", + "[46] Mohammad Sabokrou, Mohammad Khalooei, Mahmood Fathy, and Ehsan Adeli. Adversarily learned one-class classifier for novelty detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3379-3388, 2018. 1, 2", + "[47] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? Advances in neural information processing systems, 31, 2018. 8", + "[48] Thomas Schlegl, Philipp Seebock, Sebastian M Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In International conference on information processing in medical imaging, pages 146-157. Springer, 2017. 1, 2", + "[49] Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8495-8504, 2021. 2", + "[50] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3, 5", + "[51] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 5", + "[52] Mohamed-Ali Tnani, Michael Feil, and Klaus Diepold. Smart data collection system for brownfield cnc milling machines: A new benchmark dataset for data-driven machine monitoring. Procedia CIRP, 107:131-136, 2022. 1, 6", + "[53] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 5", + "[54] Yizhe Wang, Congqi Cao, and Yanning Zhang. Visualsemantic network: a visual and semantic enhanced model for gesture recognition. Visual Intelligence, 1(1):25, 2023. 1", + "[55] Lijun Wu, Fei Tian, Yingce Xia, Yang Fan, Tao Qin, Lai Jian-Huang, and Tie-Yan Liu. Learning to teach with dynamic loss functions. Advances in neural information processing systems, 31, 2018. 3", + "[56] Ji Zhang, Lianli Gao, Xu Luo, Hengtao Shen, and Jingkuan Song. Deta: Denoised task adaptation for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11541-11551, 2023. 3" + ], + "bbox": [ + 516, + 90, + 903, + 854 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30565", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_model.json b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..28d1b0e2f8ccc7ba093642aa0488395bac8ec737 --- /dev/null +++ b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.131, + 0.87, + 0.153 + ], + "angle": 0, + "content": "Task-Specific Gradient Adaptation for Few-Shot One-Class Classification" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.181, + 0.721, + 0.218 + ], + "angle": 0, + "content": "Yunlong Li Xiabi Liu* Liyuan Pan Yuchen Ren Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.22, + 0.748, + 0.235 + ], + "angle": 0, + "content": "{yunlong.li, liuxiabi, liyuan.pan, yuchen. ren}@bit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.485, + 0.681 + ], + "angle": 0, + "content": "Optimization-based meta-learning methods for few-shot one-class classification (FS-OCC) aim to fine-tune a metatrained model to classify the positive and negative samples using only a few positive samples by adaptation. However, recent approaches primarily focus on adjusting existing meta-learning algorithms for FS-OCC, while overlooking issues stemming from the misalignment between the cross-entropy loss and OCC tasks during adaptation. This misalignment, combined with the limited availability of one-class samples and the restricted diversity of task-specific adaptation, can significantly exacerbate the adverse effects of gradient instability and generalization. To address these challenges, we propose a novel Task-Specific Gradient Adaptation (TSGA) for FS-OCC. Without extra supervision, TSGA learns to generate appropriate, stable gradients by leveraging label prediction and feature representation details of one-class samples and refines the adaptation process by recalibrating task-specific gradients and regularization terms. We evaluate TSGA on three challenging datasets and a real-world CNC Milling Machine application and demonstrate consistent improvements over baseline methods. Furthermore, we illustrate the critical impact of gradient instability and task-agnostic adaptation. Notably, TSGA achieves state-of-the-art results by effectively addressing these issues." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.712, + 0.222, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.484, + 0.875 + ], + "angle": 0, + "content": "One-class classification (OCC) is an essential task in various fields, e.g., novelty detection and anomaly detection in cybersecurity and medical diagnosis applications [9, 37, 39, 54]. The objective of OCC tasks is to classify new samples as either positive or negative when only positive samples are available during training. While previous OCC methods [1, 7, 16, 17, 31, 32, 36, 40, 41, 45, 46, 48] often require extensive amounts of positive data to train a classifier, few-shot settings frequently involve scenarios where only lim" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.267, + 0.706, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.346, + 0.655, + 0.358 + ], + "angle": 0, + "content": "(a) OC-MAML" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.268, + 0.902, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.346, + 0.836, + 0.358 + ], + "angle": 0, + "content": "(b) TSGA" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Figure 1. An overview of OC-MAML and the proposed TSGA. Both (a) and (b) allow rapid adaptation of the model parameterized by \\(\\pmb{\\theta}\\), comprising a feature extractor \\(f\\) and a decision maker \\(g\\). Here, \\(\\pmb{x}\\) and \\(y\\) denote the input data and corresponding label (only positive), respectively. Different from OC-MAML which uses the cross-entropy loss \\(\\ell\\) in adaptation, we employ Loss Network \\(\\mathcal{N}\\) to generate appropriate losses \\(\\mathcal{L}\\) by leveraging the representations \\(\\boldsymbol{r} = f(\\boldsymbol{x})\\) and predictions \\(\\hat{y} = g(\\boldsymbol{r})\\). Then, we use Adapter Network \\(\\mathcal{A}\\) to obtain task-specific gradients \\(\\alpha \\nabla_{\\theta} \\mathcal{L}\\) and regularization terms \\(\\beta \\theta\\), which refines the adaptation process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.613 + ], + "angle": 0, + "content": "ited positive examples (i.e., “shots”) are available, as seen in many real-world applications, e.g., only small datasets are available in industrial manufacturing [52]. Humans intuitively perform the few-shot OCC (FS-OCC) tasks, yet machines still struggle to accomplish them effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Mainstream approaches to address the FS-OCC rely on meta-learning [20], which generally fall into two categories: 1) metric-based [12, 25, 42] and 2) optimization-based [15] methods. The former seeks to learn discriminative representations that separate positive from negative query samples, while the latter fine-tunes a meta-trained classifier by adaptation (i.e., a few gradient steps) on one-class support samples. Due to more flexible adaptation and better task generalization, optimization-based meta-learning methods demonstrate significant potential for FS-OCC, e.g., One-Class Model-Agnostic Meta-Learning (OC-MAML [15]) is developed based on the popular meta-learning approach MAML[14] by simply adapting the episode generation strategy and aims to find a model initialization that can classify novel query samples after adaptation on the one-class support set (as shown in Figure 1a)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Following this paradigm, optimization-based approaches primarily focus on adapting existing algorithms (e.g., MetaSGD [28] and MetaOptNet [27]) to FS-OCC tasks, while" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "30556" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.36, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.087, + 0.633, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.087, + 0.907, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.266, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Figure 2. Loss gradient dynamics visualization of OC-MAML and TSGA for meta-validation set at the first adaptation step for 10-shot FS-OCC tasks on miniImageNet. (a) The loss gradient is computed at the adaptation step, and loss variation is measured along the gradient direction. (b) \\(\\ell_2\\) distance between the loss gradient at a given point and the gradients corresponding to different points along the gradient direction. (c) The Lipschitz gradient continuity of loss gradients." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.349, + 0.482, + 0.531 + ], + "angle": 0, + "content": "overlooking the inherent misalignment between the evaluation metrics and OCC objectives in adaptation and lacking regularization terms in the absence of negative support samples. As illustrated in Figure 2c, high values in the effective \\(\\beta\\) smoothness of OC-MAML indicate cross-entropy loss results in inappropriate gradients and unstable model parameter updates for one-class samples. Moreover, the generalization is further limited when novel tasks diverge significantly from the base tasks encountered during training, as OC-MAML applies a task-agnostic adaptation strategy that disregards task-specific meta-knowledge, thereby constraining its full potential [4, 20, 43]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.746 + ], + "angle": 0, + "content": "In this work, we address the questions raised above by proposing Task-Specific Gradient Adaptation (TSGA), an optimization-based meta-learning framework for dealing with FS-OCC. Figure 1b provides an overview of our TSGA framework. The central idea behind our TSGA is to automatically generate suitable gradients for the one-class support samples and task-specific hyperparameters (i.e., the learning rate \\(\\alpha\\) and weight decay \\(\\beta\\)) in adaptation. To this end, we design a learnable loss network \\(\\mathcal{N}\\) to generate the losses by leveraging label prediction and feature representation details of one-class samples, based on which the gradient is calculated and an adapter network \\(\\mathcal{A}\\) is proposed to adaptively recalibrate task-specific gradients and regularization terms during adaptation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of TSGA with three public datasets and a real-world CNC Milling Machines application. On average, TSGA demonstrates consistent improvements and achieves state-of-the-art results. In particular, by tackling the gradient instability and task-agnostic adaptation, it improves the classification accuracy of the OC-MAML [15] baseline by \\(1.53\\%\\) and \\(2.60\\%\\) on miniImageNet and CIFAR-FS in the 10-shot setting, respectively. Specifically, TSGA outperforms the state-of-the-art metric-based AHDD [42] by \\(1.48\\%\\) and \\(1.50\\%\\) on miniImageNet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.502 + ], + "angle": 0, + "content": "and CIFAR-FS in the 2-shot setting, respectively (Table 1). Contributions. In summary, our contributions are three-fold. 1) We design a learnable loss network to avoid gradient instability and overfitting caused by the cross-entropy loss for FS-OCC. 2) We adaptively refine the adaptation process, which captures task-specific meta-knowledge for generalization in novel tasks. 3) Our method achieves state-of-the-art performance. Extensive experiments on three challenging datasets and a real-world CNC Milling Machines application demonstrate its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.517, + 0.655, + 0.532 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.727, + 0.557 + ], + "angle": 0, + "content": "2.1. One-class classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.837 + ], + "angle": 0, + "content": "One-class classification (OCC) methods can be divided into discriminative, generative, and self-supervised approaches [9, 37, 39]. Discriminative methods [17, 32, 41, 45] learn feature representations of positive samples and classify novel samples deviating significantly from the extracted positive feature representations as negative, e.g., Deep-SVDD [45] learns a hypersphere to enclose positive sample embeddings within a minimal volume. Generative methods [1, 31, 36, 46, 48] model the underlying distribution of positive data using models, e.g., Generative Adversarial Networks [36, 48] and Auto-Encoders [46]. Negative samples exhibit high reconstruction errors, as models trained exclusively on positive data fail to reconstruct negatives. Self-supervised methods [7, 16, 40] leverage various transforms (e.g., rotation, flipping) to generate pseudolabels and train a classifier to predict the transform applied. A query sample is classified as positive if predictions match the transform and negative otherwise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.796, + 0.863 + ], + "angle": 0, + "content": "2.2. Few-shot one-class classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Due to high data scarcity, most existing OCC methods are ill-suited for the few-shot setting, e.g., HTDG [49]," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30557" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.501 + ], + "angle": 0, + "content": "which combines a generative model and self-supervision using limited positive samples, suffers from complex training pipeline and marginal performance gain. To address FS-OCC tasks more effectively, recent discriminative approaches often leverage pre-trained weight initialization [6] on large-scale datasets [11, 41] (e.g., ImageNet [13]) or adopt the paradigm of meta-learning [12, 25, 42]. Specifically, FewSOME [6] employs Siamese Networks initialized with pre-trained weights and fine-tunes using stop gradient [10] to prevent representational collapse. OWProtoNets [25] is the first method to integrate meta-learning baseline Prototypical Networks [50] with OCC, which computes a positive prototype by averaging support embeddings and designates the origin as the negative prototype. MetaSVDD [12] constructs a minimum-volume hypersphere enclosing support data without the pseudo-negative prototype. Its variant, OC-ProtoNet, avoids complex computations of the hypersphere center by simply averaging the support features. AHDD [42], a hypersphere-based data description with a learnable radius to define the optimal feature space region for positive samples. Due to the flexibility and task adaptation capabilities, optimization-based meta-learning methods have emerged in FS-OCC. OC-MAML [15] introduces a modified episodic sampling strategy of MAML [14] to learn model initializations suited for FS-OCC. Building on this framework, some other methods (e.g., Reptile[33], Meta-SGD[28], MetaOptNet [27]) also adapt to FS-OCC." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.303, + 0.528 + ], + "angle": 0, + "content": "2.3. Loss function learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.807 + ], + "angle": 0, + "content": "The optimization of loss functions constitutes a foundational component of machine learning [18, 35]. Manually designed losses, e.g., cross-entropy loss and focal loss [44], necessitate white-box formulation while remaining tailored to specific objectives by tedious manual effort. In contrast, automated loss function design has gradually attracted attention. Recent approaches [23, 29, 35] optimize neural networks by minimizing surrogate losses that approximate the non-differentiable evaluation metric (e.g., accuracy, F1, AUC, AP). In addition, recent research has focused on meta-learning for loss functions. Under the \"Learning to Teach\" paradigm [18, 21, 30, 55], a teacher dynamically adjusts the parametric loss functions to align with the performance metric, while the student model is updated to minimize the evolving loss during training. Complementary approaches in meta-learning [2, 5, 22] develop metalearners that learn gradient update values or a task-adaptive loss function for rapid adaptation to new tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.315, + 0.835 + ], + "angle": 0, + "content": "2.4. Task-specific adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Conventional few-shot learning (FSL) assumes training and test samples originate from the same domain, a premise of ten unrealistic in real-world applications. Recent advancements [4, 34, 43, 56] focus on enhancing the generaliza" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.184 + ], + "angle": 0, + "content": "tion of pre-trained models under significant cross-category or cross-domain shifts through test-time adaptation, e.g., CNAPS [43] integrates task-specific adapters via FiLM layers [38] to adapt feature extractor to target tasks. Similarly, ALFA [4] estimates task-adaptive hyperparameters via an auxiliary network to enhance the fast adaptation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.606, + 0.21 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.906, + 0.266 + ], + "angle": 0, + "content": "In this section, we elaborate on our proposed TSGA. Before that, we introduce some preliminary concepts about the FS-OCC task and the mechanism of OC-MAML." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.274, + 0.644, + 0.29 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Assume we have an OCC model parameterized by \\(\\theta\\), which consists of a feature extractor \\(f\\) that outputs a representation for each input image, and a decision maker \\(g\\) that makes predictions for the positive or negative class. We consider a distribution over tasks \\(p(\\mathcal{T})\\) that we want our model to be able to adapt to. Each FS-OCC task \\(\\mathcal{T}\\) is randomly sampled from \\(p(\\mathcal{T})\\). The details of the FS-OCC episode sampling technique are provided in Supplementary Materials. Our goal is to apply the model to the test task \\(\\mathcal{T}_i = \\{\\mathcal{S}, \\mathcal{Q}\\}\\), by deriving task-specific knowledge on the \\(K\\)-shot one-class support set \\(\\mathcal{S} = \\{(x_i^s, y_i^s)\\}_{s=1}^K\\), consisting of \\(K\\) image-label pairs from the positive class, i.e., \\(y_i^s = 0\\). It is expected that the model can correctly partition the \\(M\\) class-balanced query samples \\(\\mathcal{Q} = \\{(x_i^q)\\}_{q=1}^M\\) to the positive if the prediction \\(\\hat{y}_i^q = 0\\), or \\(\\hat{y}_i^q = 1\\) otherwise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.658 + ], + "angle": 0, + "content": "MAML and OC-MAML. MAML [14] learns a model initialization for few-shot learning. However, MAML requires access to data examples from each class to yield such generalization. OC-MAML [15] adapts MAML for the FS-OCC problem. It first trains the model on the one-class support set from the sampled task \\(\\mathcal{T}_i\\) by taking a few gradient descent steps in the inner loop (as shown in Figure 1a), then updates the model initialization using the loss computed on a class-balanced query set in the outer loop," + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.666, + 0.907, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {\\theta} _ {i} ^ {\\prime} = \\boldsymbol {\\theta} - \\alpha \\nabla_ {\\boldsymbol {\\theta}} \\ell (\\mathcal {S}; \\boldsymbol {\\theta}), \\\\ \\boldsymbol {\\theta} \\leftarrow \\boldsymbol {\\theta} - \\eta \\nabla_ {\\boldsymbol {\\theta}} \\sum_ {\\mathcal {T} _ {i}} \\ell \\left(\\mathcal {Q}; \\boldsymbol {\\theta} _ {i} ^ {\\prime}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.906, + 0.773 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\eta\\) are learning rates. The cross-entropy loss is used for \\(\\ell\\). Theoretically, the gradient in the update can be approximated by using Taylor series expansion [15, 33]," + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.78, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {g} = \\bar {\\boldsymbol {g}} _ {2} - \\alpha \\bar {\\boldsymbol {H}} _ {2} \\bar {\\boldsymbol {g}} _ {1} - \\alpha \\bar {\\boldsymbol {H}} _ {1} \\bar {\\boldsymbol {g}} _ {2} + O \\left(\\alpha^ {2}\\right) \\\\ = \\bar {\\boldsymbol {g}} _ {2} - \\alpha \\frac {\\partial \\left(\\bar {\\boldsymbol {g}} _ {1} \\cdot \\bar {\\boldsymbol {g}} _ {2}\\right)}{\\partial \\boldsymbol {\\theta}} + O \\left(\\alpha^ {2}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\overline{\\pmb{g}}_i\\) and \\(\\overline{H}_i\\) denote the gradient and Hessian computed on the \\(i^{th}\\) mini-batch (i.e., one-class support samples and class-balanced query samples, respectively) at the initial parameter point \\(\\theta\\). The key to the effectiveness of OC-MAML" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30558" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "lies in maximizing the inner product of the gradients computed on the one-class support set and class-balanced query set, thus taking a gradient step using one yields a performance increase on the other. However, we argue that the gradients on the support set \\(\\overline{g}_1\\) computed by cross-entropy loss are unstable due to the one-class task property. Next, we solve this problem by proposing our TSGA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.208, + 0.203, + 0.223 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.483, + 0.26 + ], + "angle": 0, + "content": "Our TSGA is illustrated in Figure 1b, which mainly consists of the following steps. For each iteration:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.484, + 0.322 + ], + "angle": 0, + "content": "Step-1. A feature extractor \\( f \\) takes the images \\( \\boldsymbol{x}_i \\) from \\( S \\) as input to obtain image representations \\( \\boldsymbol{r}_i = f(\\boldsymbol{x}_i) \\) and a decision maker \\( g \\) transforms the representations \\( \\boldsymbol{r}_i \\) to predictions \\( \\hat{y}_i = g(\\boldsymbol{r}_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.383 + ], + "angle": 0, + "content": "Step-2. A loss network \\(\\mathcal{N}\\) parameterized by \\(\\phi\\) takes image representations and predictions as input to generate the loss values, based on which we can obtain the gradients with respect to the model by backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.444 + ], + "angle": 0, + "content": "Step-3. An adapter network \\(\\mathcal{A}\\) parameterized by \\(\\varphi\\) recalibrates the gradients and parameters of the model to task-specific updated gradients and regularization terms, which remodel the update rules in the adaptation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.503 + ], + "angle": 0, + "content": "Step-4. In the outer loop, a weighted sum of cross-entropy loss concerning the class-balanced query set is minimized after every adaptation step to update each module mentioned above." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.261, + 0.484, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.515, + 0.342, + 0.529 + ], + "angle": 0, + "content": "3.3. Instance-wise Loss Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.643 + ], + "angle": 0, + "content": "The motivation of the loss network \\(\\mathcal{N}\\) is that the cross-entropy loss, which is applied for rapid adaptation in OC-MAML, is not aligned with evaluation metrics in the OCC tasks, leading to unsuitable gradients and unstable model parameter updates. Designing loss functions manually on existing problems is resource intensive, so we introduce a learnable loss network \\(\\mathcal{N}\\) to make the gradient more stable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.735 + ], + "angle": 0, + "content": "Given the support set \\(\\mathcal{S} = \\{(x_i^s, y_i^s)\\}_{s=1}^K\\) of a mettraining task \\(\\mathcal{T}_i\\), each image \\(\\pmb{x}\\) is fed into feature extractor \\(f\\) to generate image representation \\(\\pmb{r}\\), and then is transformed to prediction \\(\\hat{y}\\) by decision maker \\(g\\). Next, the prediction \\(\\hat{y}\\) together with the corresponding representation \\(\\pmb{r}\\) is input into \\(\\mathcal{N}\\) to generate the loss value, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.745, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta , \\mathcal {S}; \\phi) = \\frac {1}{K} \\sum_ {(\\boldsymbol {x}, y \\in \\mathcal {S})} \\mathcal {N} (\\hat {y}, \\boldsymbol {r}, y; \\phi). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Due to the support data in FS-OCC tasks exclusively belonging to the positive class, the label \\( y = 0 \\) can be ignored. We concatenate representation and prediction as input. For simplicity, we employ a 4-layer Multi-Layer Perceptron (MLP) with ELU activation between the layers for stable gradients since it is \\( C^\\infty \\) continuous and the hidden size of \\( \\mathcal{N} \\) is set to 4." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.791, + 0.107 + ], + "angle": 0, + "content": "3.4. Task-Specific Adapter Network" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.202 + ], + "angle": 0, + "content": "Aside from generating learnable supervision for stable gradients, we customize the adaptation process for generalization in different test tasks \\(\\mathcal{T}_i\\). Intuitively, the direct way to determine the adaptation is to learn the coefficient (i.e., the learning rate \\(\\alpha\\) and weight decay \\(\\beta\\) [4, 28]) for the adaptation," + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.203, + 0.905, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {\\theta} _ {i} ^ {\\prime} = \\beta \\boldsymbol {\\theta} - \\alpha \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, S; \\phi) \\\\ = \\left[ \\boldsymbol {\\beta}, \\boldsymbol {\\alpha} \\right] \\left[ \\begin{array}{c} \\boldsymbol {\\theta} \\\\ - \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, \\mathcal {S}; \\phi) \\end{array} \\right], \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.905, + 0.289 + ], + "angle": 0, + "content": "where \\(\\beta\\) and \\(\\alpha\\) are layer-wise weight decay and learning rate, respectively, for the computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.906, + 0.364 + ], + "angle": 0, + "content": "The adapter network is thus developed to achieve task-specific adaptation by gating mechanism, which can be regarded as a self-attention function to capture task-specific contextual information and adaptively recalibrate the parameters and gradients," + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.373, + 0.905, + 0.408 + ], + "angle": 0, + "content": "\\[\n[ \\boldsymbol {\\beta}, \\boldsymbol {\\alpha} ] = \\mathcal {A} \\left(\\left[ \\begin{array}{c} \\boldsymbol {\\theta} \\\\ - \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, \\mathcal {S}; \\boldsymbol {\\phi}) \\end{array} \\right]; \\varphi\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.906, + 0.536 + ], + "angle": 0, + "content": "We employ average pooling to the parameters and gradients per layer and concatenate these two \\(L\\)-dimension vectors (i.e., \\(2L\\)-dimension) to feed into \\(\\mathcal{A}\\), where \\(L\\) is the number of the model's layer. We employ a Transformer Encoder model as the adapter network and remove all normalization layers. The learned relative positional encoding is used to utilize the position information. The details of \\(\\mathcal{A}\\) are discussed in Section 4.3." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.545, + 0.679, + 0.561 + ], + "angle": 0, + "content": "3.5. Training process" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.568, + 0.906, + 0.629 + ], + "angle": 0, + "content": "In the outer loop, we use a weighted sum of cross-entropy loss on the class-balanced query samples after every adaptation step \\( t \\) to improve gradient propagation and reduce gradient instability [3]. More formally:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.638, + 0.905, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\ell \\left(\\mathcal {Q}; \\boldsymbol {\\theta} _ {i} ^ {\\prime}\\right) = \\mathbb {E} _ {\\ell} \\left[ \\sum_ {t = 1} ^ {T} w _ {t} \\ell \\left(\\boldsymbol {\\theta} _ {i} ^ {t}\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.681, + 0.858, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\ell \\left(\\boldsymbol {\\theta} _ {i} ^ {t}\\right) = \\frac {1}{M} \\sum_ {\\left(\\boldsymbol {x}, y \\in \\mathcal {Q}\\right)} \\ell (\\hat {y}, y; \\boldsymbol {\\theta} _ {i} ^ {t}), t \\in [ 1, T ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.729, + 0.907, + 0.76 + ], + "angle": 0, + "content": "where \\(w_{t}\\) denotes the importance weight of the query set loss at step \\(t\\) and \\(\\pmb{\\theta}_{i}^{t}\\) the parameters after every step \\(t\\)," + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.769, + 0.593, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {i} ^ {0} = \\boldsymbol {\\theta},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.789, + 0.905, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {i} ^ {t} = \\beta \\boldsymbol {\\theta} _ {i} ^ {t - 1} - \\alpha \\nabla_ {\\boldsymbol {\\theta} _ {i} ^ {t - 1}} \\mathcal {L} \\left(\\boldsymbol {\\theta} _ {i} ^ {t - 1}, S; \\phi\\right), t \\in [ 1, T ], \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.813, + 0.605, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {i} ^ {\\prime} = \\boldsymbol {\\theta} _ {i} ^ {T},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.797, + 0.856 + ], + "angle": 0, + "content": "where \\(T\\) is the number of adaptation steps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Initially, each loss contributes equally to the overall loss. However, as iterations progress, the contribution from earlier steps is gradually reduced, while the contribution from" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30559" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.094, + 0.092, + 0.338, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Meta-training of TSGA" + }, + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.108, + 0.484, + 0.445 + ], + "angle": 0, + "content": "Require: Task distribution \\(p(\\mathcal{T})\\) \nRequire: Learning rates \\(\\alpha ,\\eta\\) \nRequire: Batch size for the inner and outer updates \\(K,M\\) 1: Randomly initialize \\(\\theta ,\\phi\\) \n2: while not done do \n3: Sample batch of tasks \\(\\mathcal{T}_i\\sim p(\\mathcal{T})\\) \n4: for each sampled task \\(\\mathcal{T}_i\\) do \n5: Sample one-class support set \\(\\mathcal{S} = \\{(x_i^s,y_i^s)\\}_{s = 1}^K\\) from \\(\\mathcal{T}_i\\) \n6: Sample class-balanced query set \\(\\mathcal{Q} = \\{(x_i^q,y_i^q)\\}_{q = 1}^M\\) from \\(\\mathcal{T}_i\\) \n7: Initialize \\(\\theta_{i}^{\\prime} = \\theta\\) \n8: for number of adaptation steps do \n9: Compute loss \\(\\mathcal{L}(\\theta ,\\mathcal{S};\\phi)\\) via Eq. (3). \n10: Compute hyperparameters \\(\\beta ,\\alpha\\) via Eq. (5). \n11: Perform gradient descent to adapt \\(\\theta_{i}^{\\prime}\\) via Eq. (7). \n12: end for \n13: Compute cross-entropy loss \\(\\ell (\\mathcal{Q};\\theta_i^{\\prime})\\) via Eq. (6) \n14: end for \n15: Perform gradient descent to update \\((\\theta ,\\phi ,\\varphi)\\gets (\\theta ,\\phi ,\\varphi) - \\eta \\nabla_{(\\theta ,\\phi ,\\varphi)}\\sum_{\\mathcal{T}_i}\\ell (\\mathcal{Q};\\theta_i^{\\prime}).\\) \n16: end while" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.423, + 0.489 + ], + "angle": 0, + "content": "later steps is increased to update \\(\\theta\\), \\(\\phi\\) and \\(\\varphi\\) as in," + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.498, + 0.483, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {\\theta}, \\phi , \\varphi\\right) \\leftarrow \\left(\\boldsymbol {\\theta}, \\phi , \\varphi\\right) - \\eta \\nabla_ {\\left(\\boldsymbol {\\theta}, \\phi , \\varphi\\right)} \\sum_ {\\mathcal {T} _ {i}} \\ell \\left(\\mathcal {Q}; \\boldsymbol {\\theta} _ {i} ^ {\\prime}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.569 + ], + "angle": 0, + "content": "The overall training procedure of our method is summarized in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.629 + ], + "angle": 0, + "content": "As an optimization-based meta-learning method inherited the advantages of model-agnostic from OC-MAML, our TSGA does not impose any constraint on the model and its target applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Computation cost. The overall number of learnable parameters increased from OC-MAML, together with the loss network \\(\\mathcal{N}\\) and adapter network \\(\\mathcal{A}\\), is negligible. Besides, in the bi-level optimization scheme, the training time and computation complexity for TSGA and OC-MAML are roughly the same." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.222, + 0.751 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.279, + 0.775 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate our approach on three standard benchmarks for FS-OCC: Omniglot [26], miniImageNet [53], and CIFAR-FS [8], with additional experiments on the real-world CNC Milling Machine dataset (The details are in Section 4.2). Omniglot [26] comprises 1,623 handwritten characters from 50 alphabets. All images are resized to \\(28 \\times 28\\) pixels and partitioned into 25 alphabets for meta-training, 5 for meta-validation, and 20 for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "meta-testing. miniImageNet [53] includes 50,000 training and 10,000 test images \\((84 \\times 84\\) pixels), evenly distributed across 100 classes. We partition the dataset into 64 classes for meta-training, 16 for meta-validation, and 20 for metatesting. CIFAR-FS [8] is derived from CIFAR-100 [24] via random sampling, with 600 images per class \\((32 \\times 32\\) pixels). We partition the dataset into 64 classes for metatraining, 16 for meta-validation, and 20 for meta-testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Implementation details. We followed the experimental setups used by OC-MAML [15], including hyperparameter setting, data preprocessing, and network backbones. We employ two widely adopted architectures for few-shot learning [34, 50, 51, 53]: Convolutional Neural Network (CNN) with 4 convolutional blocks (Conv-4) and ResNet-12 [19], as established in prior work. Architectural details for Conv-4 and ResNet-12 are provided in the Supplementary Materials. Unless stated otherwise, Conv-4 serves as the default backbone. In the inner loop, we optimize the classifier parameters using the SGD optimizer for 5 adaptation steps. In the outer loop, we use the Adam optimizer with a learning rate of 0.001, \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\). During meta-training, each iteration contains a batch of FS-OCC tasks (we set batch size as 4) with training conducted over 20,000 iterations. Validation is performed every 100 iterations by evaluating performance on 160 randomly sampled FS-OCC tasks from the meta-validation set. The best model on the meta-validation set is selected for final evaluation during meta-testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.637 + ], + "angle": 0, + "content": "Baselines. We verify the effectiveness of TSGA against a comprehensive range of FS-OCC methods, including 1) optimization-based methods: OC-MAML [15], OC-MetaOptNet, OC-MetaSGD and 2) metric-based methods: OC-ProtoNet [12], OW-ProtoNet [25], AHDD [42]. We directly cited the results of OW-ProtoNet and OC-ProtoNet from OC-MAML [15] and AHDD [42], which reproduced these results under a consistent experimental setup." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.906, + 0.729 + ], + "angle": 0, + "content": "Evaluation metric. We evaluate our method on 600 unique FS-OCC tasks randomly sampled from the meta-testing set for each dataset and report average accuracy (in %) in \\(95\\%\\) confidence intervals. We ensure reliable evaluation by conducting 5 random restarts, using average results for comprehensive comparisons." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.741, + 0.608, + 0.755 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.905, + 0.84 + ], + "angle": 0, + "content": "In this section, we perform extensive experiments to demonstrate the effectiveness of our TSGA. The analysis of variance (ANOVA) results are provided in Supplementary Materials, with all pairwise comparisons achieving statistical significance \\((p < 0.05)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Baseline comparison. To validate the effectiveness of TSGA for FS-OCC, we conduct experiments on three datasets with baseline approaches, as shown before. The quantitative results of the baseline methods and ours with" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30560" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.291 + ], + "angle": 0, + "content": "
ModelBackboneK=2K=10
MINOMNCIFAR-FSMINOMNCIFAR-FS
OC-ProtoNet [12]Conv-464.488.864.566.491.667.6
OW-ProtoNet [25]Conv-467.0-70.974.4-76.7
AHDD [42]Conv-471.9396.8175.3477.0097.8780.95
OC-MetaOptNet [15]ResNet-1251.8-56.367.4-75.5
OC-MetaSGD [15]Conv-469.6-71.475.8-77.8
OC-MAML [15]Conv-469.196.670.076.297.679.1
TSGA (Ours)Conv-473.41 ± 0.498.11 ± 0.176.84 ± 0.477.73 ± 0.498.54 ± 0.181.70 ± 0.4
OC-MAML†[15]ResNet-1276.60 ± 0.498.61 ± 0.174.26 ± 0.477.27 ± 0.499.18 ± 0.182.05 ± 0.4
TSGA (Ours)ResNet-1278.04 ± 0.498.97 ± 0.179.53 ± 0.479.24 ± 0.499.29 ± 0.184.05 ± 0.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.106, + 0.292, + 0.238, + 0.308 + ], + "angle": 0, + "content": "† Our reproduction." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.318, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Table 1. Comparison with the metric-based and optimization-based meta-learning methods for 2-shot and 10-shot OCC tasks on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS. We highlight the best number in **bold**." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.371, + 0.482, + 0.485 + ], + "angle": 0, + "content": "
ModelMINOMNCIFAR-FS
OC-ProtoNet [12]61.8986.0262.15
OW-ProtoNet [25]62.0094.90-
AHDD [42]68.8995.1671.89
OC-MAML†[15]69.6297.2173.76
TSGA (Ours)71.6797.3274.77
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.106, + 0.487, + 0.238, + 0.502 + ], + "angle": 0, + "content": "† Our reproduction." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.513, + 0.483, + 0.556 + ], + "angle": 0, + "content": "Table 2. One-shot OCC tasks comparison results of ours and baseline methods on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS. We highlight the best number in bold." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.572, + 0.482, + 0.635 + ], + "angle": 0, + "content": "
ModelOP00OP01OP02Avg
OC-MAML75.5665.6395.1578.78
TSGA76.0566.6597.4580.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Table 3. Comparison of OC-MAML and TSGA on several operations (i.e., OP) of the real-world CNC Milling Machine dataset, with only \\( K = 1 \\) normal examples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.901 + ], + "angle": 0, + "content": "various training sample numbers (\\(K\\)-shot) are reported in Tables 1 and 2. We can observe from the results: 1) TSGA consistently improves the baseline OC-MAML by a large margin in all settings, demonstrating its effectiveness in improving gradient stability and model generalization to FS-OCC. 2) TSGA achieves significant performance gains on both the Conv-4 and ResNet-12 backbones, suggesting TSGA's flexibility. 3) TSGA can tackle both the one-shot and few-shot FS-OCC tasks, whereas several FS-OCC baselines (e.g., Meta-SVDD [12]) are unsuitable for one-shot OCC, qualitative results are shown in Section 4.4. State-of-the-art comparison. Here we compare our TSGA" + }, + { + "type": "table", + "bbox": [ + 0.57, + 0.371, + 0.851, + 0.449 + ], + "angle": 0, + "content": "
ModelMINOMN
MAML [14]0.280.16
OC-MAML [15]0.420.23
TSGA (Ours)0.470.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.459, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Table 4. Comparison of cosine similarity between the gradients of one-class support set and class-balanced query set averaged over test tasks of miniImageNet (MIN) and Omniglot (OMN)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.716 + ], + "angle": 0, + "content": "with the state-of-the-art metric-based method AHDD [42] and conduct experiments on three datasets. AHDD learns radius \\( r \\), which is involved in the model optimization. This can be regarded as a type of loss function learning. During the meta-testing phase, the radius remains unchanged while our TSGA further captures the task-specific knowledge to refine the adaptation process. In Table 1, we can observe that TSGA considerably outperforms the strong baseline AHDD and establishes new state-of-the-art results in all settings, which further confirms the effectiveness of our TSGA and the mechanism of adaptation to the unseen test tasks of optimization-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Results of real-world CNC Milling Machine dataset. We conduct experiments on the industrial process monitoring of brownfield milling machines dataset [52] and report the results in Table 3. We compute accuracy for evaluation. The dataset consists of 1702 acceleration data built with 15 different tool operations that run on 3 machines at different time frames. The industrial dataset has an unbalanced rate of 23.3:1 between normal (i.e., positive) and non-normal data. Each data can be seen as a multivariate time series (three signals, i.e., \\( x, y, z \\)-axis). We created samples by randomly cropping windows of length 2048 from the corresponding time series of each operation and partition oper-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30561" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.209 + ], + "angle": 0, + "content": "
IDSettingK=2K=10
MINOMNCIFAR-FSMINOMNCIFAR-FS
ABaseline69.196.670.076.297.679.1
B+A72.1397.8975.9176.4698.3181.11
C+N72.0997.9975.5976.6698.4581.13
D+A+N(TSGA)73.41 (+4.31)98.11 (+1.51)76.84 (+6.84)77.73 (+1.53)98.54 (+0.94)81.70 (+2.60)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.219, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Table 5. Ablation studies for the designed components of TSGA (including the loss network \\(\\mathcal{N}\\) and adapter network \\(\\mathcal{A}\\)) on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS for 2-shot and 10-shot FS-OCC tasks." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.261, + 0.907, + 0.413 + ], + "angle": 0, + "content": "
Cross-EntropyLoss NetworkK=2K=10
FeatureLabelMINOMNCIFAR-FSMINOMNCIFAR-FS
✓XX69.196.670.076.297.679.1
✓X✓71.13 ± 0.497.95 ± 0.176.66 ± 0.477.60 ± 0.498.27 ± 0.181.11 ± 0.4
✓✓✓72.15 ± 0.497.91 ± 0.177.11 ± 0.477.19 ± 0.498.60 ± 0.180.66 ± 0.4
XX✓72.61 ± 0.498.01 ± 0.176.18 ± 0.477.31 ± 0.498.45 ± 0.181.46 ± 0.4
X✓✓73.41 ± 0.498.11 ± 0.176.84 ± 0.477.73 ± 0.498.54 ± 0.181.70 ± 0.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.423, + 0.907, + 0.452 + ], + "angle": 0, + "content": "Table 6. The impact of loss network components (Feature and Label represent feature representations and label predictions, respectively) and cross-entropy loss on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS for 2-shot and 10-shot OCC tasks." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.574 + ], + "angle": 0, + "content": "
# Layer# DimAccuracy# Layer# DimAccuracy
1481.604481.28
881.84880.44
2481.398481.47
881.99880.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Table 7. Different adapter networks results on CIFAR-FS in 10-shot OCC tasks. # Layer and # Dim represent the number of layers and the hidden dimension, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.746 + ], + "angle": 0, + "content": "ations into 10 types for meta-training, 2 for meta-validation, and 3 for meta-testing. TSGA consistently outperforms OC-MAML across the 3 types of operations and achieves \\(80\\%\\) accuracy on average by using only \\(K = 1\\) normal examples, demonstrating its competence in the time-series data domain and for industrial real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.748, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Cosine Similarity Analysis. The success of OC-MAML lies in maximizing the inner product, and therefore the cosine similarity, of the gradients computed on the one-class support set and class-balanced query set \\(\\frac{\\partial(\\overline{g}_1\\cdot\\overline{g}_2)}{\\partial\\theta}\\). Following OC-MAML, for each task, we use the initialization meta-learned by each algorithm to compute the loss gradients of support and query examples, and then compute the average cosine similarity across tasks. The cosine similarity of MAML, OC-MAML, and TSGA on test tasks from two datasets is reported in Table 4, showing that TSGA achieves" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.525 + ], + "angle": 0, + "content": "higher similarity values, indicating its ability to effectively align support and query gradients. This validates the improved task-specific adaptation capabilities of TSGA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.665, + 0.548 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Here, we conduct ablative analyses to investigate the designed components of TSGA in Table 5. We also study the impact of different supervision caused by loss network and adapter network architectures on model performance in Tables 6 and 7. Please refer to Supplementary Materials for more other details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of the designed components. TSGA contains two components, including a loss network \\(\\mathcal{L}\\) and an adapter network \\(\\mathcal{A}\\). We conduct component-wise analysis by alternatively adding one of them to understand the influence of each component in Table 5. We take the OC-MAML as baseline (A). \\(\\mathbb{B}\\) or \\(\\mathbb{C}\\): only leverage loss network or adapter network. \\(\\mathbb{D}\\): our TSGA. As can be seen, each component in TSGA contributes to the performance. In particular, the results in \\(\\mathbb{D}\\) suggest that the loss network and the adapter network complement each other to improve the FS-OCC performance. The results in \\(\\mathbb{B}\\) and \\(\\mathbb{C}\\) verify the effectiveness of learnable supervision for OCC tasks, and task-specific adaptation for various tasks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Analysis of the loss network. TSGA leverages label predictions and feature representations to learn loss values. It is important to answer the question: Does the feature part matter? To this end, we remove the feature representations." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30562" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.256, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.091, + 0.418, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.092, + 0.58, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.092, + 0.742, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.091, + 0.904, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.19, + 0.255, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.19, + 0.417, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.19, + 0.58, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.19, + 0.742, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.19, + 0.904, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.286, + 0.255, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.286, + 0.417, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.286, + 0.58, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.286, + 0.741, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.286, + 0.904, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.382, + 0.255, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.382, + 0.417, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.382, + 0.58, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.382, + 0.741, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.382, + 0.904, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.491, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Figure 3. Visualization of the generated values \\(\\alpha\\) and \\(\\beta\\) of adapter network, across adaptation steps and model layers (i.e., weights and biases of 4 convolutional layers and 1 fully connected layer in CONV-4) for 10-shot OCC tasks on miniImageNet." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.546, + 0.483, + 0.636 + ], + "angle": 0, + "content": "The results are reported in Table 6. Not surprisingly, without feature representations, performance gains decline. Further, when the cross-entropy loss and loss network are used together, there is no significant difference from when only the latter is used, implying that our loss network can comprehend the cross-entropy loss information." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.638, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Analysis of the adapter network. We study the impacts of the number of layers and the hidden size of the adapter network on performance. Table 7 shows that a larger number of layers and a bigger hidden size do not result in significant performance gains. TSGA is in general not sensitive to the choice of the number of layers and the hidden size within a certain range." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.228, + 0.772 + ], + "angle": 0, + "content": "4.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Here, we provide qualitative visualizations to see how our method works. In Figure 2, we present the visualization of the loss gradient dynamics [47], evaluating optimization stability and smoothness by measuring loss variations, changes in gradients, and the maximum difference in gradients. As observed, TSGA exhibits lighter shades in Figures 2a and 2b and lower values in Figure 2c, which indicate a smoother loss landscape. In Figure 3, we show the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.651 + ], + "angle": 0, + "content": "task-specific hyperparameters generated by the adapter network averaged across tasks, which adaptively vary across the inner-loop steps and model layers. The variability of the generated values highlights the effectiveness of our TSGA in refining optimization processes for the given task. More visualization of training curves is in Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.854 + ], + "angle": 0, + "content": "In this work, we propose TSGA for few-shot one-class classification (FS-OCC). TSGA learns to generate stable and smooth gradients by leveraging label prediction and feature representation details and enhances the adaptation process by recalibrating task-specific gradients and regularization terms. We evaluate TSGA on three challenging datasets and a real-world CNC Milling Machine dataset, demonstrating that TSGA consistently improves performance. We also reveal the severity of unstable gradients and ignoring of task-specific adaptation, by tackling these issues TSGA establishes new state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the National Natural Science Foundation of China (82171965, 62302045)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30563" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Latent space autoregression for novelty detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 481-490, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[3] Antreas Antoniou, Harri Edwards, and Amos Storkey. How to train your maml. In Seventh International Conference on Learning Representations, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] Sungyong Baik, Myungsub Choi, Janghoon Choi, Heewon Kim, and Kyoung Mu Lee. Meta-learning with adaptive hyperparameters. Advances in neural information processing systems, 33:20755-20765, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.428 + ], + "angle": 0, + "content": "[5] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9465-9474, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.483, + 0.499 + ], + "angle": 0, + "content": "[6] Niamh Belton, Misgina Tsighe Hagos, Aonghus Lawlor, and Kathleen M Curran. Fewsome: One-class few shot anomaly detection with siamese networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2978-2987, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[7] Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In International Conference on Learning Representations. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[8] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[9] Raghavendra Chalopathy and Sanjay Chawla. Deep learning for anomaly detection: A survey. arXiv preprint arXiv:1901.03407, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.631, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[11] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[12] Gabriel Dahia and Maurício Pamplona Segundo. Meta learning for few-shot one-class classification. AI, 2(2):195-208, 2021. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[15] Ahmed Frikha, Denis Krompaß, Hans-Georg Köpken, and Volker Tresp. Few-shot one-class classification via meta-learning. In Proceedings of the AAAI conference on artificial intelligence, pages 7448–7456, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[16] Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations. Advances in neural information processing systems, 31, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[17] Sachin Goyal, Aditi Raghunathan, Moksh Jain, Harsha Vardhan Simhadri, and Prateek Jain. Drocc: Deep robust one-class classification. In International conference on machine learning, pages 3711-3721. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[18] Zhaoyang Hai, Liyuan Pan, Xiabi Liu, Zhengzheng Liu, and Mirna Yunita. L2t-dln: Learning to teach with dynamic loss network. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[20] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5149-5169, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.483 + ], + "angle": 0, + "content": "[21] Chen Huang, Shuangfei Zhai, Walter Talbott, Miguel Bautista Martin, Shih-Yu Sun, Carlos Guestrin, and Josh Susskind. Addressing the loss-metric mismatch with adaptive loss alignment. In International conference on machine learning, pages 2891-2900. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.551 + ], + "angle": 0, + "content": "[22] Chen Huang, Shuangfei Zhai, Pengsheng Guo, and Josh Susskind. Metricopt: Learning to optimize black-box evaluation metrics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 174-183, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.608 + ], + "angle": 0, + "content": "[23] Tao Huang, Zekang Li, Hua Lu, Yong Shan, Shusheng Yang, Yang Feng, Fei Wang, Shan You, and Chang Xu. Relational surrogate loss learning. In International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.906, + 0.637 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.906, + 0.665 + ], + "angle": 0, + "content": "[25] Anna Kruspe. One-way prototypical networks. arXiv preprint arXiv:1906.00820, 2019. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.719 + ], + "angle": 0, + "content": "[26] Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332-1338, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[27] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[28] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[29] Lanlan Liu, Mingzhe Wang, and Jia Deng. A unified framework of surrogate loss by refactoring and interpolation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 278-293. Springer, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30564" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[30] Qingliang Liu and Jinmei Lai. Stochastic loss function. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4884-4891, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[31] Wenrui Liu, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Diversity-measurable anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12147-12156, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.203, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[32] Philipp Liznerski, Lukas Ruff, Robert A Vandermeulen, Billy Joe Franks, Marius Kloft, and Klaus Robert Muller. Explainable deep one-class classification. In International Conference on Learning Representations. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[33] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[34] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[35] Yash Patel, Tomáš Hodaž, and Jiří Matas. Learning surrogates via deep embedding. In European Conference on Computer Vision, pages 205-221. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[36] Pramuditha Perera, Ramesh Nallapati, and Bing Xiang. Ocean: One-class novelty detection using gans with constrained latent representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2898-2906, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[37] Pramuditha Perera, Poojan Oza, and Vishal M Patel. One-class classification: A survey. arXiv preprint arXiv:2101.03064, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[38] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[39] Marco AF Pimentel, David A Clifton, Lei Clifton, and Lionel Tarassenko. A review of novelty detection. Signal processing, 99:215-249, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[40] Chen Qiu, Timo Pfrommer, Marius Kloft, Stephan Mandt, and Maja Rudolph. Neural transformation learning for deep anomaly detection beyond images. In International conference on machine learning, pages 8703-8714. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[41] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[42] Yuchen Ren, Xiabi Liu, Liyuan Pan, and Lijuan Niu. Adaptive hypersphere data description for few-shot one-class classification. Applied Intelligence, pages 1-13, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[43] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E Turner. Fast and flexible multi-task classification using conditional neural adaptive processes. Advances in neural information processing systems, 32, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[44] T-YLPG Ross and GKHP Dollár. Focal loss for dense object detection. In proceedings of the IEEE conference on" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.119 + ], + "angle": 0, + "content": "computer vision and pattern recognition, pages 2980-2988, 2017.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[45] Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. In International conference on machine learning, pages 4393-4402. PMLR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[46] Mohammad Sabokrou, Mohammad Khalooei, Mahmood Fathy, and Ehsan Adeli. Adversarily learned one-class classifier for novelty detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3379-3388, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[47] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? Advances in neural information processing systems, 31, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[48] Thomas Schlegl, Philipp Seebock, Sebastian M Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In International conference on information processing in medical imaging, pages 146-157. Springer, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[49] Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8495-8504, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.472, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[50] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[51] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[52] Mohamed-Ali Tnani, Michael Feil, and Klaus Diepold. Smart data collection system for brownfield cnc milling machines: A new benchmark dataset for data-driven machine monitoring. Procedia CIRP, 107:131-136, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[53] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.699, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[54] Yizhe Wang, Congqi Cao, and Yanning Zhang. Visualsemantic network: a visual and semantic enhanced model for gesture recognition. Visual Intelligence, 1(1):25, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.742, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[55] Lijun Wu, Fei Tian, Yingce Xia, Yang Fan, Tao Qin, Lai Jian-Huang, and Tie-Yan Liu. Learning to teach with dynamic loss functions. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.799, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[56] Ji Zhang, Lianli Gao, Xu Luo, Hengtao Shen, and Jingkuan Song. Deta: Denoised task adaptation for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11541-11551, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30565" + } + ] +] \ No newline at end of file diff --git a/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_origin.pdf b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6d7283aa2e42d967cea857da39cae6391923c364 --- /dev/null +++ b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/9921c614-541a-4ded-93b9-8ac26e16d0a2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3003bf71c14dc01385ab63be7e158bd11094046d856ddad2949a522d15c92c57 +size 640442 diff --git a/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/full.md b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aad45addfe8bef5d876d12e984a8ed576acc8ec4 --- /dev/null +++ b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/full.md @@ -0,0 +1,380 @@ +# Task-Specific Gradient Adaptation for Few-Shot One-Class Classification + +Yunlong Li Xiabi Liu* Liyuan Pan Yuchen Ren Beijing Institute of Technology + +{yunlong.li, liuxiabi, liyuan.pan, yuchen. ren}@bit.edu.cn + +# Abstract + +Optimization-based meta-learning methods for few-shot one-class classification (FS-OCC) aim to fine-tune a metatrained model to classify the positive and negative samples using only a few positive samples by adaptation. However, recent approaches primarily focus on adjusting existing meta-learning algorithms for FS-OCC, while overlooking issues stemming from the misalignment between the cross-entropy loss and OCC tasks during adaptation. This misalignment, combined with the limited availability of one-class samples and the restricted diversity of task-specific adaptation, can significantly exacerbate the adverse effects of gradient instability and generalization. To address these challenges, we propose a novel Task-Specific Gradient Adaptation (TSGA) for FS-OCC. Without extra supervision, TSGA learns to generate appropriate, stable gradients by leveraging label prediction and feature representation details of one-class samples and refines the adaptation process by recalibrating task-specific gradients and regularization terms. We evaluate TSGA on three challenging datasets and a real-world CNC Milling Machine application and demonstrate consistent improvements over baseline methods. Furthermore, we illustrate the critical impact of gradient instability and task-agnostic adaptation. Notably, TSGA achieves state-of-the-art results by effectively addressing these issues. + +# 1. Introduction + +One-class classification (OCC) is an essential task in various fields, e.g., novelty detection and anomaly detection in cybersecurity and medical diagnosis applications [9, 37, 39, 54]. The objective of OCC tasks is to classify new samples as either positive or negative when only positive samples are available during training. While previous OCC methods [1, 7, 16, 17, 31, 32, 36, 40, 41, 45, 46, 48] often require extensive amounts of positive data to train a classifier, few-shot settings frequently involve scenarios where only lim + +![](images/4149e3f1141f615d38beb24c253da9df611c00577843255c648e6f434f3b709d.jpg) +(a) OC-MAML + +![](images/92d5cf94ce507afdbdacbd8b1a1a461b3418c1b5c5e7b2f8c09b9c3b88cbfa42.jpg) +(b) TSGA +Figure 1. An overview of OC-MAML and the proposed TSGA. Both (a) and (b) allow rapid adaptation of the model parameterized by $\pmb{\theta}$ , comprising a feature extractor $f$ and a decision maker $g$ . Here, $\pmb{x}$ and $y$ denote the input data and corresponding label (only positive), respectively. Different from OC-MAML which uses the cross-entropy loss $\ell$ in adaptation, we employ Loss Network $\mathcal{N}$ to generate appropriate losses $\mathcal{L}$ by leveraging the representations $\boldsymbol{r} = f(\boldsymbol{x})$ and predictions $\hat{y} = g(\boldsymbol{r})$ . Then, we use Adapter Network $\mathcal{A}$ to obtain task-specific gradients $\alpha \nabla_{\theta} \mathcal{L}$ and regularization terms $\beta \theta$ , which refines the adaptation process. + +ited positive examples (i.e., “shots”) are available, as seen in many real-world applications, e.g., only small datasets are available in industrial manufacturing [52]. Humans intuitively perform the few-shot OCC (FS-OCC) tasks, yet machines still struggle to accomplish them effectively. + +Mainstream approaches to address the FS-OCC rely on meta-learning [20], which generally fall into two categories: 1) metric-based [12, 25, 42] and 2) optimization-based [15] methods. The former seeks to learn discriminative representations that separate positive from negative query samples, while the latter fine-tunes a meta-trained classifier by adaptation (i.e., a few gradient steps) on one-class support samples. Due to more flexible adaptation and better task generalization, optimization-based meta-learning methods demonstrate significant potential for FS-OCC, e.g., One-Class Model-Agnostic Meta-Learning (OC-MAML [15]) is developed based on the popular meta-learning approach MAML[14] by simply adapting the episode generation strategy and aims to find a model initialization that can classify novel query samples after adaptation on the one-class support set (as shown in Figure 1a). + +Following this paradigm, optimization-based approaches primarily focus on adapting existing algorithms (e.g., MetaSGD [28] and MetaOptNet [27]) to FS-OCC tasks, while + +![](images/106684eeb0e477769b7e698f8b244173a0931fc0a75c987ca16cac73f31a3073.jpg) +Figure 2. Loss gradient dynamics visualization of OC-MAML and TSGA for meta-validation set at the first adaptation step for 10-shot FS-OCC tasks on miniImageNet. (a) The loss gradient is computed at the adaptation step, and loss variation is measured along the gradient direction. (b) $\ell_2$ distance between the loss gradient at a given point and the gradients corresponding to different points along the gradient direction. (c) The Lipschitz gradient continuity of loss gradients. + +![](images/09bac1af8ecd89752b3e7762bdee509c02ae54574ecbf7113cfc12b71f25d64b.jpg) + +![](images/09a4f4002a77eb80e1aa1fac3b95987d76a5f2cd6d1ffb0ca04a207c54d170ef.jpg) + +overlooking the inherent misalignment between the evaluation metrics and OCC objectives in adaptation and lacking regularization terms in the absence of negative support samples. As illustrated in Figure 2c, high values in the effective $\beta$ smoothness of OC-MAML indicate cross-entropy loss results in inappropriate gradients and unstable model parameter updates for one-class samples. Moreover, the generalization is further limited when novel tasks diverge significantly from the base tasks encountered during training, as OC-MAML applies a task-agnostic adaptation strategy that disregards task-specific meta-knowledge, thereby constraining its full potential [4, 20, 43]. + +In this work, we address the questions raised above by proposing Task-Specific Gradient Adaptation (TSGA), an optimization-based meta-learning framework for dealing with FS-OCC. Figure 1b provides an overview of our TSGA framework. The central idea behind our TSGA is to automatically generate suitable gradients for the one-class support samples and task-specific hyperparameters (i.e., the learning rate $\alpha$ and weight decay $\beta$ ) in adaptation. To this end, we design a learnable loss network $\mathcal{N}$ to generate the losses by leveraging label prediction and feature representation details of one-class samples, based on which the gradient is calculated and an adapter network $\mathcal{A}$ is proposed to adaptively recalibrate task-specific gradients and regularization terms during adaptation. + +We demonstrate the effectiveness of TSGA with three public datasets and a real-world CNC Milling Machines application. On average, TSGA demonstrates consistent improvements and achieves state-of-the-art results. In particular, by tackling the gradient instability and task-agnostic adaptation, it improves the classification accuracy of the OC-MAML [15] baseline by $1.53\%$ and $2.60\%$ on miniImageNet and CIFAR-FS in the 10-shot setting, respectively. Specifically, TSGA outperforms the state-of-the-art metric-based AHDD [42] by $1.48\%$ and $1.50\%$ on miniImageNet + +and CIFAR-FS in the 2-shot setting, respectively (Table 1). Contributions. In summary, our contributions are three-fold. 1) We design a learnable loss network to avoid gradient instability and overfitting caused by the cross-entropy loss for FS-OCC. 2) We adaptively refine the adaptation process, which captures task-specific meta-knowledge for generalization in novel tasks. 3) Our method achieves state-of-the-art performance. Extensive experiments on three challenging datasets and a real-world CNC Milling Machines application demonstrate its effectiveness. + +# 2. Related Work + +# 2.1. One-class classification + +One-class classification (OCC) methods can be divided into discriminative, generative, and self-supervised approaches [9, 37, 39]. Discriminative methods [17, 32, 41, 45] learn feature representations of positive samples and classify novel samples deviating significantly from the extracted positive feature representations as negative, e.g., Deep-SVDD [45] learns a hypersphere to enclose positive sample embeddings within a minimal volume. Generative methods [1, 31, 36, 46, 48] model the underlying distribution of positive data using models, e.g., Generative Adversarial Networks [36, 48] and Auto-Encoders [46]. Negative samples exhibit high reconstruction errors, as models trained exclusively on positive data fail to reconstruct negatives. Self-supervised methods [7, 16, 40] leverage various transforms (e.g., rotation, flipping) to generate pseudolabels and train a classifier to predict the transform applied. A query sample is classified as positive if predictions match the transform and negative otherwise. + +# 2.2. Few-shot one-class classification + +Due to high data scarcity, most existing OCC methods are ill-suited for the few-shot setting, e.g., HTDG [49], + +which combines a generative model and self-supervision using limited positive samples, suffers from complex training pipeline and marginal performance gain. To address FS-OCC tasks more effectively, recent discriminative approaches often leverage pre-trained weight initialization [6] on large-scale datasets [11, 41] (e.g., ImageNet [13]) or adopt the paradigm of meta-learning [12, 25, 42]. Specifically, FewSOME [6] employs Siamese Networks initialized with pre-trained weights and fine-tunes using stop gradient [10] to prevent representational collapse. OWProtoNets [25] is the first method to integrate meta-learning baseline Prototypical Networks [50] with OCC, which computes a positive prototype by averaging support embeddings and designates the origin as the negative prototype. MetaSVDD [12] constructs a minimum-volume hypersphere enclosing support data without the pseudo-negative prototype. Its variant, OC-ProtoNet, avoids complex computations of the hypersphere center by simply averaging the support features. AHDD [42], a hypersphere-based data description with a learnable radius to define the optimal feature space region for positive samples. Due to the flexibility and task adaptation capabilities, optimization-based meta-learning methods have emerged in FS-OCC. OC-MAML [15] introduces a modified episodic sampling strategy of MAML [14] to learn model initializations suited for FS-OCC. Building on this framework, some other methods (e.g., Reptile[33], Meta-SGD[28], MetaOptNet [27]) also adapt to FS-OCC. + +# 2.3. Loss function learning + +The optimization of loss functions constitutes a foundational component of machine learning [18, 35]. Manually designed losses, e.g., cross-entropy loss and focal loss [44], necessitate white-box formulation while remaining tailored to specific objectives by tedious manual effort. In contrast, automated loss function design has gradually attracted attention. Recent approaches [23, 29, 35] optimize neural networks by minimizing surrogate losses that approximate the non-differentiable evaluation metric (e.g., accuracy, F1, AUC, AP). In addition, recent research has focused on meta-learning for loss functions. Under the "Learning to Teach" paradigm [18, 21, 30, 55], a teacher dynamically adjusts the parametric loss functions to align with the performance metric, while the student model is updated to minimize the evolving loss during training. Complementary approaches in meta-learning [2, 5, 22] develop metalearners that learn gradient update values or a task-adaptive loss function for rapid adaptation to new tasks. + +# 2.4. Task-specific adaptation + +Conventional few-shot learning (FSL) assumes training and test samples originate from the same domain, a premise of ten unrealistic in real-world applications. Recent advancements [4, 34, 43, 56] focus on enhancing the generaliza + +tion of pre-trained models under significant cross-category or cross-domain shifts through test-time adaptation, e.g., CNAPS [43] integrates task-specific adapters via FiLM layers [38] to adapt feature extractor to target tasks. Similarly, ALFA [4] estimates task-adaptive hyperparameters via an auxiliary network to enhance the fast adaptation. + +# 3. Method + +In this section, we elaborate on our proposed TSGA. Before that, we introduce some preliminary concepts about the FS-OCC task and the mechanism of OC-MAML. + +# 3.1. Preliminary + +Assume we have an OCC model parameterized by $\theta$ , which consists of a feature extractor $f$ that outputs a representation for each input image, and a decision maker $g$ that makes predictions for the positive or negative class. We consider a distribution over tasks $p(\mathcal{T})$ that we want our model to be able to adapt to. Each FS-OCC task $\mathcal{T}$ is randomly sampled from $p(\mathcal{T})$ . The details of the FS-OCC episode sampling technique are provided in Supplementary Materials. Our goal is to apply the model to the test task $\mathcal{T}_i = \{\mathcal{S}, \mathcal{Q}\}$ , by deriving task-specific knowledge on the $K$ -shot one-class support set $\mathcal{S} = \{(x_i^s, y_i^s)\}_{s=1}^K$ , consisting of $K$ image-label pairs from the positive class, i.e., $y_i^s = 0$ . It is expected that the model can correctly partition the $M$ class-balanced query samples $\mathcal{Q} = \{(x_i^q)\}_{q=1}^M$ to the positive if the prediction $\hat{y}_i^q = 0$ , or $\hat{y}_i^q = 1$ otherwise. + +MAML and OC-MAML. MAML [14] learns a model initialization for few-shot learning. However, MAML requires access to data examples from each class to yield such generalization. OC-MAML [15] adapts MAML for the FS-OCC problem. It first trains the model on the one-class support set from the sampled task $\mathcal{T}_i$ by taking a few gradient descent steps in the inner loop (as shown in Figure 1a), then updates the model initialization using the loss computed on a class-balanced query set in the outer loop, + +$$ +\begin{array}{l} \boldsymbol {\theta} _ {i} ^ {\prime} = \boldsymbol {\theta} - \alpha \nabla_ {\boldsymbol {\theta}} \ell (\mathcal {S}; \boldsymbol {\theta}), \\ \boldsymbol {\theta} \leftarrow \boldsymbol {\theta} - \eta \nabla_ {\boldsymbol {\theta}} \sum_ {\mathcal {T} _ {i}} \ell \left(\mathcal {Q}; \boldsymbol {\theta} _ {i} ^ {\prime}\right), \tag {1} \\ \end{array} +$$ + +where $\alpha$ and $\eta$ are learning rates. The cross-entropy loss is used for $\ell$ . Theoretically, the gradient in the update can be approximated by using Taylor series expansion [15, 33], + +$$ +\begin{array}{l} \boldsymbol {g} = \bar {\boldsymbol {g}} _ {2} - \alpha \bar {\boldsymbol {H}} _ {2} \bar {\boldsymbol {g}} _ {1} - \alpha \bar {\boldsymbol {H}} _ {1} \bar {\boldsymbol {g}} _ {2} + O \left(\alpha^ {2}\right) \\ = \bar {\boldsymbol {g}} _ {2} - \alpha \frac {\partial \left(\bar {\boldsymbol {g}} _ {1} \cdot \bar {\boldsymbol {g}} _ {2}\right)}{\partial \boldsymbol {\theta}} + O \left(\alpha^ {2}\right), \tag {2} \\ \end{array} +$$ + +where $\overline{\pmb{g}}_i$ and $\overline{H}_i$ denote the gradient and Hessian computed on the $i^{th}$ mini-batch (i.e., one-class support samples and class-balanced query samples, respectively) at the initial parameter point $\theta$ . The key to the effectiveness of OC-MAML + +lies in maximizing the inner product of the gradients computed on the one-class support set and class-balanced query set, thus taking a gradient step using one yields a performance increase on the other. However, we argue that the gradients on the support set $\overline{g}_1$ computed by cross-entropy loss are unstable due to the one-class task property. Next, we solve this problem by proposing our TSGA. + +# 3.2. Overview + +Our TSGA is illustrated in Figure 1b, which mainly consists of the following steps. For each iteration: + +Step-1. A feature extractor $f$ takes the images $\boldsymbol{x}_i$ from $S$ as input to obtain image representations $\boldsymbol{r}_i = f(\boldsymbol{x}_i)$ and a decision maker $g$ transforms the representations $\boldsymbol{r}_i$ to predictions $\hat{y}_i = g(\boldsymbol{r}_i)$ . +Step-2. A loss network $\mathcal{N}$ parameterized by $\phi$ takes image representations and predictions as input to generate the loss values, based on which we can obtain the gradients with respect to the model by backpropagation. +Step-3. An adapter network $\mathcal{A}$ parameterized by $\varphi$ recalibrates the gradients and parameters of the model to task-specific updated gradients and regularization terms, which remodel the update rules in the adaptation. +Step-4. In the outer loop, a weighted sum of cross-entropy loss concerning the class-balanced query set is minimized after every adaptation step to update each module mentioned above. + +# 3.3. Instance-wise Loss Network + +The motivation of the loss network $\mathcal{N}$ is that the cross-entropy loss, which is applied for rapid adaptation in OC-MAML, is not aligned with evaluation metrics in the OCC tasks, leading to unsuitable gradients and unstable model parameter updates. Designing loss functions manually on existing problems is resource intensive, so we introduce a learnable loss network $\mathcal{N}$ to make the gradient more stable. + +Given the support set $\mathcal{S} = \{(x_i^s, y_i^s)\}_{s=1}^K$ of a mettraining task $\mathcal{T}_i$ , each image $\pmb{x}$ is fed into feature extractor $f$ to generate image representation $\pmb{r}$ , and then is transformed to prediction $\hat{y}$ by decision maker $g$ . Next, the prediction $\hat{y}$ together with the corresponding representation $\pmb{r}$ is input into $\mathcal{N}$ to generate the loss value, formulated as + +$$ +\mathcal {L} (\theta , \mathcal {S}; \phi) = \frac {1}{K} \sum_ {(\boldsymbol {x}, y \in \mathcal {S})} \mathcal {N} (\hat {y}, \boldsymbol {r}, y; \phi). \tag {3} +$$ + +Due to the support data in FS-OCC tasks exclusively belonging to the positive class, the label $y = 0$ can be ignored. We concatenate representation and prediction as input. For simplicity, we employ a 4-layer Multi-Layer Perceptron (MLP) with ELU activation between the layers for stable gradients since it is $C^\infty$ continuous and the hidden size of $\mathcal{N}$ is set to 4. + +# 3.4. Task-Specific Adapter Network + +Aside from generating learnable supervision for stable gradients, we customize the adaptation process for generalization in different test tasks $\mathcal{T}_i$ . Intuitively, the direct way to determine the adaptation is to learn the coefficient (i.e., the learning rate $\alpha$ and weight decay $\beta$ [4, 28]) for the adaptation, + +$$ +\begin{array}{l} \boldsymbol {\theta} _ {i} ^ {\prime} = \beta \boldsymbol {\theta} - \alpha \nabla_ {\boldsymbol {\theta}} \mathcal {L} (\boldsymbol {\theta}, S; \phi) \\ = \left[ \boldsymbol {\beta}, \boldsymbol {\alpha} \right] \left[ \begin{array}{c} \boldsymbol {\theta} \\ - \nabla_ {\boldsymbol {\theta}} \mathcal {L} (\boldsymbol {\theta}, \mathcal {S}; \phi) \end{array} \right], \tag {4} \\ \end{array} +$$ + +where $\beta$ and $\alpha$ are layer-wise weight decay and learning rate, respectively, for the computational efficiency. + +The adapter network is thus developed to achieve task-specific adaptation by gating mechanism, which can be regarded as a self-attention function to capture task-specific contextual information and adaptively recalibrate the parameters and gradients, + +$$ +[ \boldsymbol {\beta}, \boldsymbol {\alpha} ] = \mathcal {A} \left(\left[ \begin{array}{c} \boldsymbol {\theta} \\ - \nabla_ {\boldsymbol {\theta}} \mathcal {L} (\boldsymbol {\theta}, \mathcal {S}; \boldsymbol {\phi}) \end{array} \right]; \varphi\right). \tag {5} +$$ + +We employ average pooling to the parameters and gradients per layer and concatenate these two $L$ -dimension vectors (i.e., $2L$ -dimension) to feed into $\mathcal{A}$ , where $L$ is the number of the model's layer. We employ a Transformer Encoder model as the adapter network and remove all normalization layers. The learned relative positional encoding is used to utilize the position information. The details of $\mathcal{A}$ are discussed in Section 4.3. + +# 3.5. Training process + +In the outer loop, we use a weighted sum of cross-entropy loss on the class-balanced query samples after every adaptation step $t$ to improve gradient propagation and reduce gradient instability [3]. More formally: + +$$ +\ell \left(\mathcal {Q}; \boldsymbol {\theta} _ {i} ^ {\prime}\right) = \mathbb {E} _ {\ell} \left[ \sum_ {t = 1} ^ {T} w _ {t} \ell \left(\boldsymbol {\theta} _ {i} ^ {t}\right) \right], \tag {6} +$$ + +$$ +\ell \left(\boldsymbol {\theta} _ {i} ^ {t}\right) = \frac {1}{M} \sum_ {\left(\boldsymbol {x}, y \in \mathcal {Q}\right)} \ell (\hat {y}, y; \boldsymbol {\theta} _ {i} ^ {t}), t \in [ 1, T ], +$$ + +where $w_{t}$ denotes the importance weight of the query set loss at step $t$ and $\pmb{\theta}_{i}^{t}$ the parameters after every step $t$ , + +$$ +\boldsymbol {\theta} _ {i} ^ {0} = \boldsymbol {\theta}, +$$ + +$$ +\boldsymbol {\theta} _ {i} ^ {t} = \beta \boldsymbol {\theta} _ {i} ^ {t - 1} - \alpha \nabla_ {\boldsymbol {\theta} _ {i} ^ {t - 1}} \mathcal {L} \left(\boldsymbol {\theta} _ {i} ^ {t - 1}, S; \phi\right), t \in [ 1, T ], \tag {7} +$$ + +$$ +\boldsymbol {\theta} _ {i} ^ {\prime} = \boldsymbol {\theta} _ {i} ^ {T}, +$$ + +where $T$ is the number of adaptation steps. + +Initially, each loss contributes equally to the overall loss. However, as iterations progress, the contribution from earlier steps is gradually reduced, while the contribution from + +Algorithm 1 Meta-training of TSGA +Require: Task distribution $p(\mathcal{T})$ +Require: Learning rates $\alpha ,\eta$ +Require: Batch size for the inner and outer updates $K,M$ 1: Randomly initialize $\theta ,\phi$ +2: while not done do +3: Sample batch of tasks $\mathcal{T}_i\sim p(\mathcal{T})$ +4: for each sampled task $\mathcal{T}_i$ do +5: Sample one-class support set $\mathcal{S} = \{(x_i^s,y_i^s)\}_{s = 1}^K$ from $\mathcal{T}_i$ +6: Sample class-balanced query set $\mathcal{Q} = \{(x_i^q,y_i^q)\}_{q = 1}^M$ from $\mathcal{T}_i$ +7: Initialize $\theta_{i}^{\prime} = \theta$ +8: for number of adaptation steps do +9: Compute loss $\mathcal{L}(\theta ,\mathcal{S};\phi)$ via Eq. (3). +10: Compute hyperparameters $\beta ,\alpha$ via Eq. (5). +11: Perform gradient descent to adapt $\theta_{i}^{\prime}$ via Eq. (7). +12: end for +13: Compute cross-entropy loss $\ell (\mathcal{Q};\theta_i^{\prime})$ via Eq. (6) +14: end for +15: Perform gradient descent to update $(\theta ,\phi ,\varphi)\gets (\theta ,\phi ,\varphi) - \eta \nabla_{(\theta ,\phi ,\varphi)}\sum_{\mathcal{T}_i}\ell (\mathcal{Q};\theta_i^{\prime}).$ +16: end while + +later steps is increased to update $\theta$ , $\phi$ and $\varphi$ as in, + +$$ +\left(\boldsymbol {\theta}, \phi , \varphi\right) \leftarrow \left(\boldsymbol {\theta}, \phi , \varphi\right) - \eta \nabla_ {\left(\boldsymbol {\theta}, \phi , \varphi\right)} \sum_ {\mathcal {T} _ {i}} \ell \left(\mathcal {Q}; \boldsymbol {\theta} _ {i} ^ {\prime}\right). \tag {8} +$$ + +The overall training procedure of our method is summarized in Algorithm 1. + +As an optimization-based meta-learning method inherited the advantages of model-agnostic from OC-MAML, our TSGA does not impose any constraint on the model and its target applications. + +Computation cost. The overall number of learnable parameters increased from OC-MAML, together with the loss network $\mathcal{N}$ and adapter network $\mathcal{A}$ , is negligible. Besides, in the bi-level optimization scheme, the training time and computation complexity for TSGA and OC-MAML are roughly the same. + +# 4. Experiments + +# 4.1. Experimental setup + +Datasets. We evaluate our approach on three standard benchmarks for FS-OCC: Omniglot [26], miniImageNet [53], and CIFAR-FS [8], with additional experiments on the real-world CNC Milling Machine dataset (The details are in Section 4.2). Omniglot [26] comprises 1,623 handwritten characters from 50 alphabets. All images are resized to $28 \times 28$ pixels and partitioned into 25 alphabets for meta-training, 5 for meta-validation, and 20 for + +meta-testing. miniImageNet [53] includes 50,000 training and 10,000 test images $(84 \times 84$ pixels), evenly distributed across 100 classes. We partition the dataset into 64 classes for meta-training, 16 for meta-validation, and 20 for metatesting. CIFAR-FS [8] is derived from CIFAR-100 [24] via random sampling, with 600 images per class $(32 \times 32$ pixels). We partition the dataset into 64 classes for metatraining, 16 for meta-validation, and 20 for meta-testing. + +Implementation details. We followed the experimental setups used by OC-MAML [15], including hyperparameter setting, data preprocessing, and network backbones. We employ two widely adopted architectures for few-shot learning [34, 50, 51, 53]: Convolutional Neural Network (CNN) with 4 convolutional blocks (Conv-4) and ResNet-12 [19], as established in prior work. Architectural details for Conv-4 and ResNet-12 are provided in the Supplementary Materials. Unless stated otherwise, Conv-4 serves as the default backbone. In the inner loop, we optimize the classifier parameters using the SGD optimizer for 5 adaptation steps. In the outer loop, we use the Adam optimizer with a learning rate of 0.001, $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ . During meta-training, each iteration contains a batch of FS-OCC tasks (we set batch size as 4) with training conducted over 20,000 iterations. Validation is performed every 100 iterations by evaluating performance on 160 randomly sampled FS-OCC tasks from the meta-validation set. The best model on the meta-validation set is selected for final evaluation during meta-testing. + +Baselines. We verify the effectiveness of TSGA against a comprehensive range of FS-OCC methods, including 1) optimization-based methods: OC-MAML [15], OC-MetaOptNet, OC-MetaSGD and 2) metric-based methods: OC-ProtoNet [12], OW-ProtoNet [25], AHDD [42]. We directly cited the results of OW-ProtoNet and OC-ProtoNet from OC-MAML [15] and AHDD [42], which reproduced these results under a consistent experimental setup. + +Evaluation metric. We evaluate our method on 600 unique FS-OCC tasks randomly sampled from the meta-testing set for each dataset and report average accuracy (in %) in $95\%$ confidence intervals. We ensure reliable evaluation by conducting 5 random restarts, using average results for comprehensive comparisons. + +# 4.2. Results + +In this section, we perform extensive experiments to demonstrate the effectiveness of our TSGA. The analysis of variance (ANOVA) results are provided in Supplementary Materials, with all pairwise comparisons achieving statistical significance $(p < 0.05)$ . + +Baseline comparison. To validate the effectiveness of TSGA for FS-OCC, we conduct experiments on three datasets with baseline approaches, as shown before. The quantitative results of the baseline methods and ours with + +
ModelBackboneK=2K=10
MINOMNCIFAR-FSMINOMNCIFAR-FS
OC-ProtoNet [12]Conv-464.488.864.566.491.667.6
OW-ProtoNet [25]Conv-467.0-70.974.4-76.7
AHDD [42]Conv-471.9396.8175.3477.0097.8780.95
OC-MetaOptNet [15]ResNet-1251.8-56.367.4-75.5
OC-MetaSGD [15]Conv-469.6-71.475.8-77.8
OC-MAML [15]Conv-469.196.670.076.297.679.1
TSGA (Ours)Conv-473.41 ± 0.498.11 ± 0.176.84 ± 0.477.73 ± 0.498.54 ± 0.181.70 ± 0.4
OC-MAML†[15]ResNet-1276.60 ± 0.498.61 ± 0.174.26 ± 0.477.27 ± 0.499.18 ± 0.182.05 ± 0.4
TSGA (Ours)ResNet-1278.04 ± 0.498.97 ± 0.179.53 ± 0.479.24 ± 0.499.29 ± 0.184.05 ± 0.4
+ +† Our reproduction. + +Table 1. Comparison with the metric-based and optimization-based meta-learning methods for 2-shot and 10-shot OCC tasks on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS. We highlight the best number in **bold**. + +
ModelMINOMNCIFAR-FS
OC-ProtoNet [12]61.8986.0262.15
OW-ProtoNet [25]62.0094.90-
AHDD [42]68.8995.1671.89
OC-MAML†[15]69.6297.2173.76
TSGA (Ours)71.6797.3274.77
+ +† Our reproduction. + +Table 2. One-shot OCC tasks comparison results of ours and baseline methods on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS. We highlight the best number in bold. + +
ModelOP00OP01OP02Avg
OC-MAML75.5665.6395.1578.78
TSGA76.0566.6597.4580.05
+ +various training sample numbers ( $K$ -shot) are reported in Tables 1 and 2. We can observe from the results: 1) TSGA consistently improves the baseline OC-MAML by a large margin in all settings, demonstrating its effectiveness in improving gradient stability and model generalization to FS-OCC. 2) TSGA achieves significant performance gains on both the Conv-4 and ResNet-12 backbones, suggesting TSGA's flexibility. 3) TSGA can tackle both the one-shot and few-shot FS-OCC tasks, whereas several FS-OCC baselines (e.g., Meta-SVDD [12]) are unsuitable for one-shot OCC, qualitative results are shown in Section 4.4. State-of-the-art comparison. Here we compare our TSGA + +Table 3. Comparison of OC-MAML and TSGA on several operations (i.e., OP) of the real-world CNC Milling Machine dataset, with only $K = 1$ normal examples. + +
ModelMINOMN
MAML [14]0.280.16
OC-MAML [15]0.420.23
TSGA (Ours)0.470.39
+ +Table 4. Comparison of cosine similarity between the gradients of one-class support set and class-balanced query set averaged over test tasks of miniImageNet (MIN) and Omniglot (OMN). + +with the state-of-the-art metric-based method AHDD [42] and conduct experiments on three datasets. AHDD learns radius $r$ , which is involved in the model optimization. This can be regarded as a type of loss function learning. During the meta-testing phase, the radius remains unchanged while our TSGA further captures the task-specific knowledge to refine the adaptation process. In Table 1, we can observe that TSGA considerably outperforms the strong baseline AHDD and establishes new state-of-the-art results in all settings, which further confirms the effectiveness of our TSGA and the mechanism of adaptation to the unseen test tasks of optimization-based methods. + +Results of real-world CNC Milling Machine dataset. We conduct experiments on the industrial process monitoring of brownfield milling machines dataset [52] and report the results in Table 3. We compute accuracy for evaluation. The dataset consists of 1702 acceleration data built with 15 different tool operations that run on 3 machines at different time frames. The industrial dataset has an unbalanced rate of 23.3:1 between normal (i.e., positive) and non-normal data. Each data can be seen as a multivariate time series (three signals, i.e., $x, y, z$ -axis). We created samples by randomly cropping windows of length 2048 from the corresponding time series of each operation and partition oper- + +
IDSettingK=2K=10
MINOMNCIFAR-FSMINOMNCIFAR-FS
ABaseline69.196.670.076.297.679.1
B+A72.1397.8975.9176.4698.3181.11
C+N72.0997.9975.5976.6698.4581.13
D+A+N(TSGA)73.41 (+4.31)98.11 (+1.51)76.84 (+6.84)77.73 (+1.53)98.54 (+0.94)81.70 (+2.60)
+ +Table 5. Ablation studies for the designed components of TSGA (including the loss network $\mathcal{N}$ and adapter network $\mathcal{A}$ ) on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS for 2-shot and 10-shot FS-OCC tasks. + +
Cross-EntropyLoss NetworkK=2K=10
FeatureLabelMINOMNCIFAR-FSMINOMNCIFAR-FS
✓XX69.196.670.076.297.679.1
✓X✓71.13 ± 0.497.95 ± 0.176.66 ± 0.477.60 ± 0.498.27 ± 0.181.11 ± 0.4
✓✓✓72.15 ± 0.497.91 ± 0.177.11 ± 0.477.19 ± 0.498.60 ± 0.180.66 ± 0.4
XX✓72.61 ± 0.498.01 ± 0.176.18 ± 0.477.31 ± 0.498.45 ± 0.181.46 ± 0.4
X✓✓73.41 ± 0.498.11 ± 0.176.84 ± 0.477.73 ± 0.498.54 ± 0.181.70 ± 0.4
+ +Table 6. The impact of loss network components (Feature and Label represent feature representations and label predictions, respectively) and cross-entropy loss on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS for 2-shot and 10-shot OCC tasks. + +
# Layer# DimAccuracy# Layer# DimAccuracy
1481.604481.28
881.84880.44
2481.398481.47
881.99880.65
+ +Table 7. Different adapter networks results on CIFAR-FS in 10-shot OCC tasks. # Layer and # Dim represent the number of layers and the hidden dimension, respectively. + +ations into 10 types for meta-training, 2 for meta-validation, and 3 for meta-testing. TSGA consistently outperforms OC-MAML across the 3 types of operations and achieves $80\%$ accuracy on average by using only $K = 1$ normal examples, demonstrating its competence in the time-series data domain and for industrial real-world applications. + +Cosine Similarity Analysis. The success of OC-MAML lies in maximizing the inner product, and therefore the cosine similarity, of the gradients computed on the one-class support set and class-balanced query set $\frac{\partial(\overline{g}_1\cdot\overline{g}_2)}{\partial\theta}$ . Following OC-MAML, for each task, we use the initialization meta-learned by each algorithm to compute the loss gradients of support and query examples, and then compute the average cosine similarity across tasks. The cosine similarity of MAML, OC-MAML, and TSGA on test tasks from two datasets is reported in Table 4, showing that TSGA achieves + +higher similarity values, indicating its ability to effectively align support and query gradients. This validates the improved task-specific adaptation capabilities of TSGA. + +# 4.3. Ablation study + +Here, we conduct ablative analyses to investigate the designed components of TSGA in Table 5. We also study the impact of different supervision caused by loss network and adapter network architectures on model performance in Tables 6 and 7. Please refer to Supplementary Materials for more other details. + +Effectiveness of the designed components. TSGA contains two components, including a loss network $\mathcal{L}$ and an adapter network $\mathcal{A}$ . We conduct component-wise analysis by alternatively adding one of them to understand the influence of each component in Table 5. We take the OC-MAML as baseline (A). $\mathbb{B}$ or $\mathbb{C}$ : only leverage loss network or adapter network. $\mathbb{D}$ : our TSGA. As can be seen, each component in TSGA contributes to the performance. In particular, the results in $\mathbb{D}$ suggest that the loss network and the adapter network complement each other to improve the FS-OCC performance. The results in $\mathbb{B}$ and $\mathbb{C}$ verify the effectiveness of learnable supervision for OCC tasks, and task-specific adaptation for various tasks, respectively. + +Analysis of the loss network. TSGA leverages label predictions and feature representations to learn loss values. It is important to answer the question: Does the feature part matter? To this end, we remove the feature representations. + +![](images/ecd27b087e872ef4dea8303d476947b3170d542e0ae17bf4a699530805997fd9.jpg) + +![](images/f3eb8004aa2a5611b6d826fa8b2d4005351bdb76714226f1247683ee4d1b7d20.jpg) + +![](images/d8012c05a1f691ed37b64a74a01c46fd0dda752e1667d463e6ef132fa0ec3c56.jpg) + +![](images/1ed004893d2036f21f8662bea6fed907aa7744b8412c76002cec877c30d47648.jpg) + +![](images/e97212de60e6520feb265e84d5955ea4c25ee714318d0aef562572335cfb0e53.jpg) + +![](images/8e28e474bb0a02b4bdf0f9b604f712ac5f4c1e92be6e0eace92733ee95c138da.jpg) + +![](images/648a4aa6feacef866d7a478db3e3c41609bc2b7eaa4b4cdcc9a7c9494eda481f.jpg) + +![](images/1b561d026264bbbffc42b4286688c7c43a336561faab3943b2b800267807650b.jpg) + +![](images/ced786a4f3f6ef9b6b4ae55883294d29199db71d4cc7030261c4ccb819f60608.jpg) + +![](images/430d165be57993f87364065ed39d27f353bcede59fc5ae011ef510bc91664ead.jpg) + +![](images/e85460941e7c1b0aa7b99b0c76b84e3c9594ca273b9a13cc188ab3fb64ba462b.jpg) + +![](images/7fb17025fc6feef622e851cf50470e22508651ca6fea99cba86199c793f7b684.jpg) + +![](images/acd905ba7a3ae03bb6eff2bedb42fb0d990e5b1d467757fb65dbc2a35e1fb2f3.jpg) + +![](images/0a6825eb0c9354d0e4d3bb55f6d513295bdc1756ac003190a08edab76df3107a.jpg) + +![](images/2b69de7149a731f120dcdf80e743626d6a3dcb1baaf092a0b90ec26f9fd28a41.jpg) + +![](images/039ee51b641a54a7b0207600b179790b3c35976624f9128bc60981811032b4a4.jpg) +Figure 3. Visualization of the generated values $\alpha$ and $\beta$ of adapter network, across adaptation steps and model layers (i.e., weights and biases of 4 convolutional layers and 1 fully connected layer in CONV-4) for 10-shot OCC tasks on miniImageNet. + +![](images/32618484f2e8033e40559bc1d3bcf28471ebd6cc0b9016692293476d1db05e0d.jpg) + +![](images/0b88c95f34af3c38c11131af22f33af961c398276715c90e0f8c1b6e5f766f16.jpg) + +![](images/b824d6165523811fba2f2ad7e883681e6ddcedcbc150ccec9009b4658847ea72.jpg) + +![](images/8614f4ba23e82d2599332cb85a33c8bb9ad5c4abe82d8ebdca70715cb9365cb6.jpg) + +The results are reported in Table 6. Not surprisingly, without feature representations, performance gains decline. Further, when the cross-entropy loss and loss network are used together, there is no significant difference from when only the latter is used, implying that our loss network can comprehend the cross-entropy loss information. + +Analysis of the adapter network. We study the impacts of the number of layers and the hidden size of the adapter network on performance. Table 7 shows that a larger number of layers and a bigger hidden size do not result in significant performance gains. TSGA is in general not sensitive to the choice of the number of layers and the hidden size within a certain range. + +# 4.4. Visualization + +Here, we provide qualitative visualizations to see how our method works. In Figure 2, we present the visualization of the loss gradient dynamics [47], evaluating optimization stability and smoothness by measuring loss variations, changes in gradients, and the maximum difference in gradients. As observed, TSGA exhibits lighter shades in Figures 2a and 2b and lower values in Figure 2c, which indicate a smoother loss landscape. In Figure 3, we show the + +task-specific hyperparameters generated by the adapter network averaged across tasks, which adaptively vary across the inner-loop steps and model layers. The variability of the generated values highlights the effectiveness of our TSGA in refining optimization processes for the given task. More visualization of training curves is in Supplementary Materials. + +# 5. Conclusion + +In this work, we propose TSGA for few-shot one-class classification (FS-OCC). TSGA learns to generate stable and smooth gradients by leveraging label prediction and feature representation details and enhances the adaptation process by recalibrating task-specific gradients and regularization terms. We evaluate TSGA on three challenging datasets and a real-world CNC Milling Machine dataset, demonstrating that TSGA consistently improves performance. We also reveal the severity of unstable gradients and ignoring of task-specific adaptation, by tackling these issues TSGA establishes new state-of-the-art results. + +Acknowledgments. This work was supported by the National Natural Science Foundation of China (82171965, 62302045). + +# References + +[1] Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Latent space autoregression for novelty detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 481-490, 2019. 1, 2 +[2] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. Advances in neural information processing systems, 29, 2016. 3 +[3] Antreas Antoniou, Harri Edwards, and Amos Storkey. How to train your maml. In Seventh International Conference on Learning Representations, 2019. 4 +[4] Sungyong Baik, Myungsub Choi, Janghoon Choi, Heewon Kim, and Kyoung Mu Lee. Meta-learning with adaptive hyperparameters. Advances in neural information processing systems, 33:20755-20765, 2020. 2, 3, 4 +[5] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9465-9474, 2021. 3 +[6] Niamh Belton, Misgina Tsighe Hagos, Aonghus Lawlor, and Kathleen M Curran. Fewsome: One-class few shot anomaly detection with siamese networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2978-2987, 2023. 3 +[7] Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In International Conference on Learning Representations. 1, 2 +[8] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 5 +[9] Raghavendra Chalopathy and Sanjay Chawla. Deep learning for anomaly detection: A survey. arXiv preprint arXiv:1901.03407, 2019. 1, 2 +[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3 +[11] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 3 +[12] Gabriel Dahia and Maurício Pamplona Segundo. Meta learning for few-shot one-class classification. AI, 2(2):195-208, 2021. 1, 3, 5, 6 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3 +[14] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 1, 3, 6 + +[15] Ahmed Frikha, Denis Krompaß, Hans-Georg Köpken, and Volker Tresp. Few-shot one-class classification via meta-learning. In Proceedings of the AAAI conference on artificial intelligence, pages 7448–7456, 2021. 1, 2, 3, 5, 6 +[16] Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations. Advances in neural information processing systems, 31, 2018. 1, 2 +[17] Sachin Goyal, Aditi Raghunathan, Moksh Jain, Harsha Vardhan Simhadri, and Prateek Jain. Drocc: Deep robust one-class classification. In International conference on machine learning, pages 3711-3721. PMLR, 2020. 1, 2 +[18] Zhaoyang Hai, Liyuan Pan, Xiabi Liu, Zhengzheng Liu, and Mirna Yunita. L2t-dln: Learning to teach with dynamic loss network. Advances in Neural Information Processing Systems, 36, 2024. 3 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[20] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5149-5169, 2021. 1, 2 +[21] Chen Huang, Shuangfei Zhai, Walter Talbott, Miguel Bautista Martin, Shih-Yu Sun, Carlos Guestrin, and Josh Susskind. Addressing the loss-metric mismatch with adaptive loss alignment. In International conference on machine learning, pages 2891-2900. PMLR, 2019. 3 +[22] Chen Huang, Shuangfei Zhai, Pengsheng Guo, and Josh Susskind. Metricopt: Learning to optimize black-box evaluation metrics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 174-183, 2021. 3 +[23] Tao Huang, Zekang Li, Hua Lu, Yong Shan, Shusheng Yang, Yang Feng, Fei Wang, Shan You, and Chang Xu. Relational surrogate loss learning. In International Conference on Learning Representations. 3 +[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[25] Anna Kruspe. One-way prototypical networks. arXiv preprint arXiv:1906.00820, 2019. 1, 3, 5, 6 +[26] Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332-1338, 2015. 5 +[27] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 1, 3 +[28] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1, 3, 4 +[29] Lanlan Liu, Mingzhe Wang, and Jia Deng. A unified framework of surrogate loss by refactoring and interpolation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 278-293. Springer, 2020. 3 + +[30] Qingliang Liu and Jinmei Lai. Stochastic loss function. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4884-4891, 2020. 3 +[31] Wenrui Liu, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Diversity-measurable anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12147-12156, 2023. 1, 2 +[32] Philipp Liznerski, Lukas Ruff, Robert A Vandermeulen, Billy Joe Franks, Marius Kloft, and Klaus Robert Muller. Explainable deep one-class classification. In International Conference on Learning Representations. 1, 2 +[33] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 3 +[34] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 3, 5 +[35] Yash Patel, Tomáš Hodaž, and Jiří Matas. Learning surrogates via deep embedding. In European Conference on Computer Vision, pages 205-221. Springer, 2020. 3 +[36] Pramuditha Perera, Ramesh Nallapati, and Bing Xiang. Ocean: One-class novelty detection using gans with constrained latent representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2898-2906, 2019. 1, 2 +[37] Pramuditha Perera, Poojan Oza, and Vishal M Patel. One-class classification: A survey. arXiv preprint arXiv:2101.03064, 2021. 1, 2 +[38] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3 +[39] Marco AF Pimentel, David A Clifton, Lei Clifton, and Lionel Tarassenko. A review of novelty detection. Signal processing, 99:215-249, 2014. 1, 2 +[40] Chen Qiu, Timo Pfrommer, Marius Kloft, Stephan Mandt, and Maja Rudolph. Neural transformation learning for deep anomaly detection beyond images. In International conference on machine learning, pages 8703-8714. PMLR, 2021. 1, 2 +[41] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 1, 2, 3 +[42] Yuchen Ren, Xiabi Liu, Liyuan Pan, and Lijuan Niu. Adaptive hypersphere data description for few-shot one-class classification. Applied Intelligence, pages 1-13, 2024. 1, 2, 3, 5, 6 +[43] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E Turner. Fast and flexible multi-task classification using conditional neural adaptive processes. Advances in neural information processing systems, 32, 2019. 2, 3 +[44] T-YLPG Ross and GKHP Dollár. Focal loss for dense object detection. In proceedings of the IEEE conference on + +computer vision and pattern recognition, pages 2980-2988, 2017.3 +[45] Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. In International conference on machine learning, pages 4393-4402. PMLR, 2018. 1, 2 +[46] Mohammad Sabokrou, Mohammad Khalooei, Mahmood Fathy, and Ehsan Adeli. Adversarily learned one-class classifier for novelty detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3379-3388, 2018. 1, 2 +[47] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? Advances in neural information processing systems, 31, 2018. 8 +[48] Thomas Schlegl, Philipp Seebock, Sebastian M Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In International conference on information processing in medical imaging, pages 146-157. Springer, 2017. 1, 2 +[49] Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8495-8504, 2021. 2 +[50] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3, 5 +[51] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 5 +[52] Mohamed-Ali Tnani, Michael Feil, and Klaus Diepold. Smart data collection system for brownfield cnc milling machines: A new benchmark dataset for data-driven machine monitoring. Procedia CIRP, 107:131-136, 2022. 1, 6 +[53] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 5 +[54] Yizhe Wang, Congqi Cao, and Yanning Zhang. Visualsemantic network: a visual and semantic enhanced model for gesture recognition. Visual Intelligence, 1(1):25, 2023. 1 +[55] Lijun Wu, Fei Tian, Yingce Xia, Yang Fan, Tao Qin, Lai Jian-Huang, and Tie-Yan Liu. Learning to teach with dynamic loss functions. Advances in neural information processing systems, 31, 2018. 3 +[56] Ji Zhang, Lianli Gao, Xu Luo, Hengtao Shen, and Jingkuan Song. Deta: Denoised task adaptation for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11541-11551, 2023. 3 \ No newline at end of file diff --git a/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/images.zip b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..162a3cabc27d11ef800f3968ee6be966b3a38341 --- /dev/null +++ b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5035bf528554c3d54ea35cc3bf741b7c7a7ca8660357d56112bdf3c4e6bcc483 +size 554799 diff --git a/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/layout.json b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3874902e1ddba856884eccd6a4eff398bcf59712 --- /dev/null +++ b/2025/Task-Specific Gradient Adaptation for Few-Shot One-Class Classification/layout.json @@ -0,0 +1,10416 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 79, + 103, + 532, + 121 + ], + "type": "text", + "content": "Task-Specific Gradient Adaptation for Few-Shot One-Class Classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 169, + 143, + 441, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 143, + 441, + 172 + ], + "spans": [ + { + "bbox": [ + 169, + 143, + 441, + 172 + ], + "type": "text", + "content": "Yunlong Li Xiabi Liu* Liyuan Pan Yuchen Ren Beijing Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 174, + 457, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 174, + 457, + 186 + ], + "spans": [ + { + "bbox": [ + 149, + 174, + 457, + 186 + ], + "type": "text", + "content": "{yunlong.li, liuxiabi, liyuan.pan, yuchen. ren}@bit.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 539 + ], + "type": "text", + "content": "Optimization-based meta-learning methods for few-shot one-class classification (FS-OCC) aim to fine-tune a metatrained model to classify the positive and negative samples using only a few positive samples by adaptation. However, recent approaches primarily focus on adjusting existing meta-learning algorithms for FS-OCC, while overlooking issues stemming from the misalignment between the cross-entropy loss and OCC tasks during adaptation. This misalignment, combined with the limited availability of one-class samples and the restricted diversity of task-specific adaptation, can significantly exacerbate the adverse effects of gradient instability and generalization. To address these challenges, we propose a novel Task-Specific Gradient Adaptation (TSGA) for FS-OCC. Without extra supervision, TSGA learns to generate appropriate, stable gradients by leveraging label prediction and feature representation details of one-class samples and refines the adaptation process by recalibrating task-specific gradients and regularization terms. We evaluate TSGA on three challenging datasets and a real-world CNC Milling Machine application and demonstrate consistent improvements over baseline methods. Furthermore, we illustrate the critical impact of gradient instability and task-agnostic adaptation. Notably, TSGA achieves state-of-the-art results by effectively addressing these issues." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 563, + 135, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 135, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 135, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 584, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 296, + 693 + ], + "type": "text", + "content": "One-class classification (OCC) is an essential task in various fields, e.g., novelty detection and anomaly detection in cybersecurity and medical diagnosis applications [9, 37, 39, 54]. The objective of OCC tasks is to classify new samples as either positive or negative when only positive samples are available during training. While previous OCC methods [1, 7, 16, 17, 31, 32, 36, 40, 41, 45, 46, 48] often require extensive amounts of positive data to train a classifier, few-shot settings frequently involve scenarios where only lim" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 211, + 432, + 271 + ], + "blocks": [ + { + "bbox": [ + 317, + 211, + 432, + 271 + ], + "lines": [ + { + "bbox": [ + 317, + 211, + 432, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 211, + 432, + 271 + ], + "type": "image", + "image_path": "4149e3f1141f615d38beb24c253da9df611c00577843255c648e6f434f3b709d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 274, + 400, + 283 + ], + "lines": [ + { + "bbox": [ + 349, + 274, + 400, + 283 + ], + "spans": [ + { + "bbox": [ + 349, + 274, + 400, + 283 + ], + "type": "text", + "content": "(a) OC-MAML" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 436, + 212, + 552, + 272 + ], + "blocks": [ + { + "bbox": [ + 436, + 212, + 552, + 272 + ], + "lines": [ + { + "bbox": [ + 436, + 212, + 552, + 272 + ], + "spans": [ + { + "bbox": [ + 436, + 212, + 552, + 272 + ], + "type": "image", + "image_path": "92d5cf94ce507afdbdacbd8b1a1a461b3418c1b5c5e7b2f8c09b9c3b88cbfa42.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 274, + 511, + 283 + ], + "lines": [ + { + "bbox": [ + 477, + 274, + 511, + 283 + ], + "spans": [ + { + "bbox": [ + 477, + 274, + 511, + 283 + ], + "type": "text", + "content": "(b) TSGA" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. An overview of OC-MAML and the proposed TSGA. Both (a) and (b) allow rapid adaptation of the model parameterized by " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": ", comprising a feature extractor " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": " and a decision maker " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": " denote the input data and corresponding label (only positive), respectively. Different from OC-MAML which uses the cross-entropy loss " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": " in adaptation, we employ Loss Network " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": " to generate appropriate losses " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": " by leveraging the representations " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r} = f(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": " and predictions " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\hat{y} = g(\\boldsymbol{r})" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": ". Then, we use Adapter Network " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": " to obtain task-specific gradients " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\alpha \\nabla_{\\theta} \\mathcal{L}" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": " and regularization terms " + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\beta \\theta" + }, + { + "bbox": [ + 313, + 293, + 555, + 403 + ], + "type": "text", + "content": ", which refines the adaptation process." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 485 + ], + "type": "text", + "content": "ited positive examples (i.e., “shots”) are available, as seen in many real-world applications, e.g., only small datasets are available in industrial manufacturing [52]. Humans intuitively perform the few-shot OCC (FS-OCC) tasks, yet machines still struggle to accomplish them effectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 486, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 676 + ], + "type": "text", + "content": "Mainstream approaches to address the FS-OCC rely on meta-learning [20], which generally fall into two categories: 1) metric-based [12, 25, 42] and 2) optimization-based [15] methods. The former seeks to learn discriminative representations that separate positive from negative query samples, while the latter fine-tunes a meta-trained classifier by adaptation (i.e., a few gradient steps) on one-class support samples. Due to more flexible adaptation and better task generalization, optimization-based meta-learning methods demonstrate significant potential for FS-OCC, e.g., One-Class Model-Agnostic Meta-Learning (OC-MAML [15]) is developed based on the popular meta-learning approach MAML[14] by simply adapting the episode generation strategy and aims to find a model initialization that can classify novel query samples after adaptation on the one-class support set (as shown in Figure 1a)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Following this paradigm, optimization-based approaches primarily focus on adapting existing algorithms (e.g., MetaSGD [28] and MetaOptNet [27]) to FS-OCC tasks, while" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30556" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 220, + 202 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 220, + 202 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 220, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 220, + 202 + ], + "type": "image", + "image_path": "106684eeb0e477769b7e698f8b244173a0931fc0a75c987ca16cac73f31a3073.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 210, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 210, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 210, + 555, + 257 + ], + "type": "text", + "content": "Figure 2. Loss gradient dynamics visualization of OC-MAML and TSGA for meta-validation set at the first adaptation step for 10-shot FS-OCC tasks on miniImageNet. (a) The loss gradient is computed at the adaptation step, and loss variation is measured along the gradient direction. (b) " + }, + { + "bbox": [ + 54, + 210, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 54, + 210, + 555, + 257 + ], + "type": "text", + "content": " distance between the loss gradient at a given point and the gradients corresponding to different points along the gradient direction. (c) The Lipschitz gradient continuity of loss gradients." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 68, + 387, + 203 + ], + "blocks": [ + { + "bbox": [ + 223, + 68, + 387, + 203 + ], + "lines": [ + { + "bbox": [ + 223, + 68, + 387, + 203 + ], + "spans": [ + { + "bbox": [ + 223, + 68, + 387, + 203 + ], + "type": "image", + "image_path": "09bac1af8ecd89752b3e7762bdee509c02ae54574ecbf7113cfc12b71f25d64b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 68, + 555, + 204 + ], + "blocks": [ + { + "bbox": [ + 391, + 68, + 555, + 204 + ], + "lines": [ + { + "bbox": [ + 391, + 68, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 391, + 68, + 555, + 204 + ], + "type": "image", + "image_path": "09a4f4002a77eb80e1aa1fac3b95987d76a5f2cd6d1ffb0ca04a207c54d170ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 276, + 294, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 276, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 294, + 420 + ], + "type": "text", + "content": "overlooking the inherent misalignment between the evaluation metrics and OCC objectives in adaptation and lacking regularization terms in the absence of negative support samples. As illustrated in Figure 2c, high values in the effective " + }, + { + "bbox": [ + 54, + 276, + 294, + 420 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 276, + 294, + 420 + ], + "type": "text", + "content": " smoothness of OC-MAML indicate cross-entropy loss results in inappropriate gradients and unstable model parameter updates for one-class samples. Moreover, the generalization is further limited when novel tasks diverge significantly from the base tasks encountered during training, as OC-MAML applies a task-agnostic adaptation strategy that disregards task-specific meta-knowledge, thereby constraining its full potential [4, 20, 43]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "text", + "content": "In this work, we address the questions raised above by proposing Task-Specific Gradient Adaptation (TSGA), an optimization-based meta-learning framework for dealing with FS-OCC. Figure 1b provides an overview of our TSGA framework. The central idea behind our TSGA is to automatically generate suitable gradients for the one-class support samples and task-specific hyperparameters (i.e., the learning rate " + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "text", + "content": " and weight decay " + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "text", + "content": ") in adaptation. To this end, we design a learnable loss network " + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "text", + "content": " to generate the losses by leveraging label prediction and feature representation details of one-class samples, based on which the gradient is calculated and an adapter network " + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "text", + "content": " is proposed to adaptively recalibrate task-specific gradients and regularization terms during adaptation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "We demonstrate the effectiveness of TSGA with three public datasets and a real-world CNC Milling Machines application. On average, TSGA demonstrates consistent improvements and achieves state-of-the-art results. In particular, by tackling the gradient instability and task-agnostic adaptation, it improves the classification accuracy of the OC-MAML [15] baseline by " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.53\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "2.60\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " on miniImageNet and CIFAR-FS in the 10-shot setting, respectively. Specifically, TSGA outperforms the state-of-the-art metric-based AHDD [42] by " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.48\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.50\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " on miniImageNet" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 276, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 397 + ], + "type": "text", + "content": "and CIFAR-FS in the 2-shot setting, respectively (Table 1). Contributions. In summary, our contributions are three-fold. 1) We design a learnable loss network to avoid gradient instability and overfitting caused by the cross-entropy loss for FS-OCC. 2) We adaptively refine the adaptation process, which captures task-specific meta-knowledge for generalization in novel tasks. 3) Our method achieves state-of-the-art performance. Extensive experiments on three challenging datasets and a real-world CNC Milling Machines application demonstrate its effectiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 409, + 400, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 409, + 400, + 421 + ], + "spans": [ + { + "bbox": [ + 314, + 409, + 400, + 421 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 429, + 444, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 444, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 444, + 441 + ], + "type": "text", + "content": "2.1. One-class classification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 447, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 662 + ], + "type": "text", + "content": "One-class classification (OCC) methods can be divided into discriminative, generative, and self-supervised approaches [9, 37, 39]. Discriminative methods [17, 32, 41, 45] learn feature representations of positive samples and classify novel samples deviating significantly from the extracted positive feature representations as negative, e.g., Deep-SVDD [45] learns a hypersphere to enclose positive sample embeddings within a minimal volume. Generative methods [1, 31, 36, 46, 48] model the underlying distribution of positive data using models, e.g., Generative Adversarial Networks [36, 48] and Auto-Encoders [46]. Negative samples exhibit high reconstruction errors, as models trained exclusively on positive data fail to reconstruct negatives. Self-supervised methods [7, 16, 40] leverage various transforms (e.g., rotation, flipping) to generate pseudolabels and train a classifier to predict the transform applied. A query sample is classified as positive if predictions match the transform and negative otherwise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 671, + 487, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 487, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 487, + 683 + ], + "type": "text", + "content": "2.2. Few-shot one-class classification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Due to high data scarcity, most existing OCC methods are ill-suited for the few-shot setting, e.g., HTDG [49]," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30557" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "content": "which combines a generative model and self-supervision using limited positive samples, suffers from complex training pipeline and marginal performance gain. To address FS-OCC tasks more effectively, recent discriminative approaches often leverage pre-trained weight initialization [6] on large-scale datasets [11, 41] (e.g., ImageNet [13]) or adopt the paradigm of meta-learning [12, 25, 42]. Specifically, FewSOME [6] employs Siamese Networks initialized with pre-trained weights and fine-tunes using stop gradient [10] to prevent representational collapse. OWProtoNets [25] is the first method to integrate meta-learning baseline Prototypical Networks [50] with OCC, which computes a positive prototype by averaging support embeddings and designates the origin as the negative prototype. MetaSVDD [12] constructs a minimum-volume hypersphere enclosing support data without the pseudo-negative prototype. Its variant, OC-ProtoNet, avoids complex computations of the hypersphere center by simply averaging the support features. AHDD [42], a hypersphere-based data description with a learnable radius to define the optimal feature space region for positive samples. Due to the flexibility and task adaptation capabilities, optimization-based meta-learning methods have emerged in FS-OCC. OC-MAML [15] introduces a modified episodic sampling strategy of MAML [14] to learn model initializations suited for FS-OCC. Building on this framework, some other methods (e.g., Reptile[33], Meta-SGD[28], MetaOptNet [27]) also adapt to FS-OCC." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 404, + 185, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 185, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 185, + 418 + ], + "type": "text", + "content": "2.3. Loss function learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 422, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 639 + ], + "type": "text", + "content": "The optimization of loss functions constitutes a foundational component of machine learning [18, 35]. Manually designed losses, e.g., cross-entropy loss and focal loss [44], necessitate white-box formulation while remaining tailored to specific objectives by tedious manual effort. In contrast, automated loss function design has gradually attracted attention. Recent approaches [23, 29, 35] optimize neural networks by minimizing surrogate losses that approximate the non-differentiable evaluation metric (e.g., accuracy, F1, AUC, AP). In addition, recent research has focused on meta-learning for loss functions. Under the \"Learning to Teach\" paradigm [18, 21, 30, 55], a teacher dynamically adjusts the parametric loss functions to align with the performance metric, while the student model is updated to minimize the evolving loss during training. Complementary approaches in meta-learning [2, 5, 22] develop metalearners that learn gradient update values or a task-adaptive loss function for rapid adaptation to new tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 647, + 192, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 192, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 192, + 661 + ], + "type": "text", + "content": "2.4. Task-specific adaptation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Conventional few-shot learning (FSL) assumes training and test samples originate from the same domain, a premise of ten unrealistic in real-world applications. Recent advancements [4, 34, 43, 56] focus on enhancing the generaliza" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "tion of pre-trained models under significant cross-category or cross-domain shifts through test-time adaptation, e.g., CNAPS [43] integrates task-specific adapters via FiLM layers [38] to adapt feature extractor to target tasks. Similarly, ALFA [4] estimates task-adaptive hyperparameters via an auxiliary network to enhance the fast adaptation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 154, + 370, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 370, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 370, + 166 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 174, + 554, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 554, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 554, + 210 + ], + "type": "text", + "content": "In this section, we elaborate on our proposed TSGA. Before that, we introduce some preliminary concepts about the FS-OCC task and the mechanism of OC-MAML." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 217, + 394, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 394, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 394, + 229 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": "Assume we have an OCC model parameterized by " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": ", which consists of a feature extractor " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": " that outputs a representation for each input image, and a decision maker " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": " that makes predictions for the positive or negative class. We consider a distribution over tasks " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "p(\\mathcal{T})" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": " that we want our model to be able to adapt to. Each FS-OCC task " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": " is randomly sampled from " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "p(\\mathcal{T})" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": ". The details of the FS-OCC episode sampling technique are provided in Supplementary Materials. Our goal is to apply the model to the test task " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i = \\{\\mathcal{S}, \\mathcal{Q}\\}" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": ", by deriving task-specific knowledge on the " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": "-shot one-class support set " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{(x_i^s, y_i^s)\\}_{s=1}^K" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": ", consisting of " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": " image-label pairs from the positive class, i.e., " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "y_i^s = 0" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": ". It is expected that the model can correctly partition the " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": " class-balanced query samples " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{(x_i^q)\\}_{q=1}^M" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": " to the positive if the prediction " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^q = 0" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^q = 1" + }, + { + "bbox": [ + 313, + 234, + 555, + 413 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "text", + "content": "MAML and OC-MAML. MAML [14] learns a model initialization for few-shot learning. However, MAML requires access to data examples from each class to yield such generalization. OC-MAML [15] adapts MAML for the FS-OCC problem. It first trains the model on the one-class support set from the sampled task " + }, + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "text", + "content": " by taking a few gradient descent steps in the inner loop (as shown in Figure 1a), then updates the model initialization using the loss computed on a class-balanced query set in the outer loop," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 527, + 555, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 527, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 373, + 527, + 555, + 570 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {\\theta} _ {i} ^ {\\prime} = \\boldsymbol {\\theta} - \\alpha \\nabla_ {\\boldsymbol {\\theta}} \\ell (\\mathcal {S}; \\boldsymbol {\\theta}), \\\\ \\boldsymbol {\\theta} \\leftarrow \\boldsymbol {\\theta} - \\eta \\nabla_ {\\boldsymbol {\\theta}} \\sum_ {\\mathcal {T} _ {i}} \\ell \\left(\\mathcal {Q}; \\boldsymbol {\\theta} _ {i} ^ {\\prime}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "d5412d45bae0b6155f2390285003bd0fc8210c8c1e98557d77f3e8ad0193adae.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 576, + 554, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 554, + 612 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 576, + 554, + 612 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 576, + 554, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 576, + 554, + 612 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 576, + 554, + 612 + ], + "type": "text", + "content": " are learning rates. The cross-entropy loss is used for " + }, + { + "bbox": [ + 313, + 576, + 554, + 612 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 576, + 554, + 612 + ], + "type": "text", + "content": ". Theoretically, the gradient in the update can be approximated by using Taylor series expansion [15, 33]," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 617, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 617, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 353, + 617, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {g} = \\bar {\\boldsymbol {g}} _ {2} - \\alpha \\bar {\\boldsymbol {H}} _ {2} \\bar {\\boldsymbol {g}} _ {1} - \\alpha \\bar {\\boldsymbol {H}} _ {1} \\bar {\\boldsymbol {g}} _ {2} + O \\left(\\alpha^ {2}\\right) \\\\ = \\bar {\\boldsymbol {g}} _ {2} - \\alpha \\frac {\\partial \\left(\\bar {\\boldsymbol {g}} _ {1} \\cdot \\bar {\\boldsymbol {g}} _ {2}\\right)}{\\partial \\boldsymbol {\\theta}} + O \\left(\\alpha^ {2}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "121add3ae5358ae82a745a6336fb674d31f24b9a840b995949c76fa8605c550a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\overline{\\pmb{g}}_i" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\overline{H}_i" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " denote the gradient and Hessian computed on the " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " mini-batch (i.e., one-class support samples and class-balanced query samples, respectively) at the initial parameter point " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": ". The key to the effectiveness of OC-MAML" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30558" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "lies in maximizing the inner product of the gradients computed on the one-class support set and class-balanced query set, thus taking a gradient step using one yields a performance increase on the other. However, we argue that the gradients on the support set " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\overline{g}_1" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " computed by cross-entropy loss are unstable due to the one-class task property. Next, we solve this problem by proposing our TSGA." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 164, + 124, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 124, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 124, + 176 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 295, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 205 + ], + "type": "text", + "content": "Our TSGA is illustrated in Figure 1b, which mainly consists of the following steps. For each iteration:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 206, + 296, + 398 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "text", + "content": "Step-1. A feature extractor " + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "text", + "content": " takes the images " + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i" + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "text", + "content": " as input to obtain image representations " + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_i = f(\\boldsymbol{x}_i)" + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "text", + "content": " and a decision maker " + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "text", + "content": " transforms the representations " + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_i" + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "text", + "content": " to predictions " + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\hat{y}_i = g(\\boldsymbol{r}_i)" + }, + { + "bbox": [ + 55, + 206, + 296, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 255, + 296, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 303 + ], + "type": "text", + "content": "Step-2. A loss network " + }, + { + "bbox": [ + 55, + 255, + 296, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 255, + 296, + 303 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 55, + 255, + 296, + 303 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 255, + 296, + 303 + ], + "type": "text", + "content": " takes image representations and predictions as input to generate the loss values, based on which we can obtain the gradients with respect to the model by backpropagation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": "Step-3. An adapter network " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": " recalibrates the gradients and parameters of the model to task-specific updated gradients and regularization terms, which remodel the update rules in the adaptation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 351, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 398 + ], + "type": "text", + "content": "Step-4. In the outer loop, a weighted sum of cross-entropy loss concerning the class-balanced query set is minimized after every adaptation step to update each module mentioned above." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 407, + 209, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 209, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 209, + 418 + ], + "type": "text", + "content": "3.3. Instance-wise Loss Network" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "text", + "content": "The motivation of the loss network " + }, + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "text", + "content": " is that the cross-entropy loss, which is applied for rapid adaptation in OC-MAML, is not aligned with evaluation metrics in the OCC tasks, leading to unsuitable gradients and unstable model parameter updates. Designing loss functions manually on existing problems is resource intensive, so we introduce a learnable loss network " + }, + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "text", + "content": " to make the gradient more stable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": "Given the support set " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{(x_i^s, y_i^s)\\}_{s=1}^K" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " of a mettraining task " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": ", each image " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " is fed into feature extractor " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " to generate image representation " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": ", and then is transformed to prediction " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " by decision maker " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": ". Next, the prediction " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " together with the corresponding representation " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " is input into " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " to generate the loss value, formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 93, + 590, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 590, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 93, + 590, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta , \\mathcal {S}; \\phi) = \\frac {1}{K} \\sum_ {(\\boldsymbol {x}, y \\in \\mathcal {S})} \\mathcal {N} (\\hat {y}, \\boldsymbol {r}, y; \\phi). \\tag {3}", + "image_path": "17194ffc5af5880fc4e4ead2dd7393c1d83d0e43900bfad05e2ce9a630e7d34d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Due to the support data in FS-OCC tasks exclusively belonging to the positive class, the label " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " can be ignored. We concatenate representation and prediction as input. For simplicity, we employ a 4-layer Multi-Layer Perceptron (MLP) with ELU activation between the layers for stable gradients since it is " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "C^\\infty" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " continuous and the hidden size of " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " is set to 4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 484, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 484, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 484, + 84 + ], + "type": "text", + "content": "3.4. Task-Specific Adapter Network" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 89, + 555, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 159 + ], + "type": "text", + "content": "Aside from generating learnable supervision for stable gradients, we customize the adaptation process for generalization in different test tasks " + }, + { + "bbox": [ + 313, + 89, + 555, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 313, + 89, + 555, + 159 + ], + "type": "text", + "content": ". Intuitively, the direct way to determine the adaptation is to learn the coefficient (i.e., the learning rate " + }, + { + "bbox": [ + 313, + 89, + 555, + 159 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 89, + 555, + 159 + ], + "type": "text", + "content": " and weight decay " + }, + { + "bbox": [ + 313, + 89, + 555, + 159 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 89, + 555, + 159 + ], + "type": "text", + "content": " [4, 28]) for the adaptation," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 160, + 553, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 160, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 365, + 160, + 553, + 201 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {\\theta} _ {i} ^ {\\prime} = \\beta \\boldsymbol {\\theta} - \\alpha \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, S; \\phi) \\\\ = \\left[ \\boldsymbol {\\beta}, \\boldsymbol {\\alpha} \\right] \\left[ \\begin{array}{c} \\boldsymbol {\\theta} \\\\ - \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, \\mathcal {S}; \\phi) \\end{array} \\right], \\tag {4} \\\\ \\end{array}", + "image_path": "acb8d37e9d9774596934aa16524ea1d4792f65d68db9cc1d34d7b5cabd4e820e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "content": " are layer-wise weight decay and learning rate, respectively, for the computational efficiency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 228, + 554, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 288 + ], + "type": "text", + "content": "The adapter network is thus developed to achieve task-specific adaptation by gating mechanism, which can be regarded as a self-attention function to capture task-specific contextual information and adaptively recalibrate the parameters and gradients," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 350, + 295, + 553, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 295, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 350, + 295, + 553, + 323 + ], + "type": "interline_equation", + "content": "[ \\boldsymbol {\\beta}, \\boldsymbol {\\alpha} ] = \\mathcal {A} \\left(\\left[ \\begin{array}{c} \\boldsymbol {\\theta} \\\\ - \\nabla_ {\\boldsymbol {\\theta}} \\mathcal {L} (\\boldsymbol {\\theta}, \\mathcal {S}; \\boldsymbol {\\phi}) \\end{array} \\right]; \\varphi\\right). \\tag {5}", + "image_path": "eae26932f6f69fe6f6aba15f3bbb82758332d9a499e0823b7049f2c3cb63e74f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "text", + "content": "We employ average pooling to the parameters and gradients per layer and concatenate these two " + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "text", + "content": "-dimension vectors (i.e., " + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "inline_equation", + "content": "2L" + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "text", + "content": "-dimension) to feed into " + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "text", + "content": " is the number of the model's layer. We employ a Transformer Encoder model as the adapter network and remove all normalization layers. The learned relative positional encoding is used to utilize the position information. The details of " + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 329, + 554, + 424 + ], + "type": "text", + "content": " are discussed in Section 4.3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 431, + 415, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 415, + 444 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 415, + 444 + ], + "type": "text", + "content": "3.5. Training process" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": "In the outer loop, we use a weighted sum of cross-entropy loss on the class-balanced query samples after every adaptation step " + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": " to improve gradient propagation and reduce gradient instability [3]. More formally:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 343, + 505, + 553, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 505, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 343, + 505, + 553, + 542 + ], + "type": "interline_equation", + "content": "\\ell \\left(\\mathcal {Q}; \\boldsymbol {\\theta} _ {i} ^ {\\prime}\\right) = \\mathbb {E} _ {\\ell} \\left[ \\sum_ {t = 1} ^ {T} w _ {t} \\ell \\left(\\boldsymbol {\\theta} _ {i} ^ {t}\\right) \\right], \\tag {6}", + "image_path": "d49ccd156058de21eeabd2e879377047aef6e65fa7beaa590672b7c16d96c239.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 356, + 539, + 525, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 539, + 525, + 570 + ], + "spans": [ + { + "bbox": [ + 356, + 539, + 525, + 570 + ], + "type": "interline_equation", + "content": "\\ell \\left(\\boldsymbol {\\theta} _ {i} ^ {t}\\right) = \\frac {1}{M} \\sum_ {\\left(\\boldsymbol {x}, y \\in \\mathcal {Q}\\right)} \\ell (\\hat {y}, y; \\boldsymbol {\\theta} _ {i} ^ {t}), t \\in [ 1, T ],", + "image_path": "28314069dd82d9a23548376caf4a86d84a8029d70f0ea8f62ecdce9636c369be.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "text", + "content": " denotes the importance weight of the query set loss at step " + }, + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{i}^{t}" + }, + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "text", + "content": " the parameters after every step " + }, + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 577, + 555, + 601 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 328, + 609, + 362, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 609, + 362, + 622 + ], + "spans": [ + { + "bbox": [ + 328, + 609, + 362, + 622 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {i} ^ {0} = \\boldsymbol {\\theta},", + "image_path": "a2f0929081c49c1b2d93b6965da06bbbd31f9d121886b1de931d56a3d17656d0.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 329, + 624, + 553, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 624, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 329, + 624, + 553, + 642 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {i} ^ {t} = \\beta \\boldsymbol {\\theta} _ {i} ^ {t - 1} - \\alpha \\nabla_ {\\boldsymbol {\\theta} _ {i} ^ {t - 1}} \\mathcal {L} \\left(\\boldsymbol {\\theta} _ {i} ^ {t - 1}, S; \\phi\\right), t \\in [ 1, T ], \\tag {7}", + "image_path": "265e3377aa6c26f5f32495bf19a8c37f8a35a60acc7762a4966ee71dfa4790b6.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 329, + 643, + 370, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 643, + 370, + 658 + ], + "spans": [ + { + "bbox": [ + 329, + 643, + 370, + 658 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {i} ^ {\\prime} = \\boldsymbol {\\theta} _ {i} ^ {T},", + "image_path": "db7593bbff9be0c9f8978457c678bd9a58805f396f9466b6e94dadd8f60db4a3.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 666, + 487, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 487, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 487, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 487, + 677 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 666, + 487, + 677 + ], + "type": "text", + "content": " is the number of adaptation steps." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Initially, each loss contributes equally to the overall loss. However, as iterations progress, the contribution from earlier steps is gradually reduced, while the contribution from" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30559" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 85, + 296, + 352 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 206, + 83 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 206, + 83 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 206, + 83 + ], + "type": "text", + "content": "Algorithm 1 Meta-training of TSGA" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "lines": [ + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": "Require: Task distribution " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "p(\\mathcal{T})" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " \nRequire: Learning rates " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\alpha ,\\eta" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " \nRequire: Batch size for the inner and outer updates " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "K,M" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " 1: Randomly initialize " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\theta ,\\phi" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " \n2: while not done do \n3: Sample batch of tasks " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i\\sim p(\\mathcal{T})" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " \n4: for each sampled task " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " do \n5: Sample one-class support set " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{(x_i^s,y_i^s)\\}_{s = 1}^K" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " \n6: Sample class-balanced query set " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{(x_i^q,y_i^q)\\}_{q = 1}^M" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " \n7: Initialize " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{\\prime} = \\theta" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " \n8: for number of adaptation steps do \n9: Compute loss " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta ,\\mathcal{S};\\phi)" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " via Eq. (3). \n10: Compute hyperparameters " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\beta ,\\alpha" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " via Eq. (5). \n11: Perform gradient descent to adapt " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{\\prime}" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " via Eq. (7). \n12: end for \n13: Compute cross-entropy loss " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\ell (\\mathcal{Q};\\theta_i^{\\prime})" + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " via Eq. (6) \n14: end for \n15: Perform gradient descent to update " + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi ,\\varphi)\\gets (\\theta ,\\phi ,\\varphi) - \\eta \\nabla_{(\\theta ,\\phi ,\\varphi)}\\sum_{\\mathcal{T}_i}\\ell (\\mathcal{Q};\\theta_i^{\\prime})." + }, + { + "bbox": [ + 56, + 85, + 296, + 352 + ], + "type": "text", + "content": " \n16: end while" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 373, + 258, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 258, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 258, + 387 + ], + "type": "text", + "content": "later steps is increased to update " + }, + { + "bbox": [ + 55, + 373, + 258, + 387 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 373, + 258, + 387 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 373, + 258, + 387 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 373, + 258, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 258, + 387 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 373, + 258, + 387 + ], + "type": "text", + "content": " as in," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 394, + 295, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 394, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 295, + 420 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {\\theta}, \\phi , \\varphi\\right) \\leftarrow \\left(\\boldsymbol {\\theta}, \\phi , \\varphi\\right) - \\eta \\nabla_ {\\left(\\boldsymbol {\\theta}, \\phi , \\varphi\\right)} \\sum_ {\\mathcal {T} _ {i}} \\ell \\left(\\mathcal {Q}; \\boldsymbol {\\theta} _ {i} ^ {\\prime}\\right). \\tag {8}", + "image_path": "c76f10322bc955bdd1aea6c91864fe00f811cb73b393c1d8afa113ecd8685dff.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 450 + ], + "type": "text", + "content": "The overall training procedure of our method is summarized in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 451, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 498 + ], + "type": "text", + "content": "As an optimization-based meta-learning method inherited the advantages of model-agnostic from OC-MAML, our TSGA does not impose any constraint on the model and its target applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "text", + "content": "Computation cost. The overall number of learnable parameters increased from OC-MAML, together with the loss network " + }, + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "text", + "content": " and adapter network " + }, + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "text", + "content": ", is negligible. Besides, in the bi-level optimization scheme, the training time and computation complexity for TSGA and OC-MAML are roughly the same." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 170, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 170, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 170, + 613 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate our approach on three standard benchmarks for FS-OCC: Omniglot [26], miniImageNet [53], and CIFAR-FS [8], with additional experiments on the real-world CNC Milling Machine dataset (The details are in Section 4.2). Omniglot [26] comprises 1,623 handwritten characters from 50 alphabets. All images are resized to " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": " pixels and partitioned into 25 alphabets for meta-training, 5 for meta-validation, and 20 for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "meta-testing. miniImageNet [53] includes 50,000 training and 10,000 test images " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "(84 \\times 84" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " pixels), evenly distributed across 100 classes. We partition the dataset into 64 classes for meta-training, 16 for meta-validation, and 20 for metatesting. CIFAR-FS [8] is derived from CIFAR-100 [24] via random sampling, with 600 images per class " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "(32 \\times 32" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " pixels). We partition the dataset into 64 classes for metatraining, 16 for meta-validation, and 20 for meta-testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 168, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 407 + ], + "type": "text", + "content": "Implementation details. We followed the experimental setups used by OC-MAML [15], including hyperparameter setting, data preprocessing, and network backbones. We employ two widely adopted architectures for few-shot learning [34, 50, 51, 53]: Convolutional Neural Network (CNN) with 4 convolutional blocks (Conv-4) and ResNet-12 [19], as established in prior work. Architectural details for Conv-4 and ResNet-12 are provided in the Supplementary Materials. Unless stated otherwise, Conv-4 serves as the default backbone. In the inner loop, we optimize the classifier parameters using the SGD optimizer for 5 adaptation steps. In the outer loop, we use the Adam optimizer with a learning rate of 0.001, " + }, + { + "bbox": [ + 313, + 168, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 168, + 555, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 168, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 313, + 168, + 555, + 407 + ], + "type": "text", + "content": ". During meta-training, each iteration contains a batch of FS-OCC tasks (we set batch size as 4) with training conducted over 20,000 iterations. Validation is performed every 100 iterations by evaluating performance on 160 randomly sampled FS-OCC tasks from the meta-validation set. The best model on the meta-validation set is selected for final evaluation during meta-testing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 408, + 554, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 504 + ], + "type": "text", + "content": "Baselines. We verify the effectiveness of TSGA against a comprehensive range of FS-OCC methods, including 1) optimization-based methods: OC-MAML [15], OC-MetaOptNet, OC-MetaSGD and 2) metric-based methods: OC-ProtoNet [12], OW-ProtoNet [25], AHDD [42]. We directly cited the results of OW-ProtoNet and OC-ProtoNet from OC-MAML [15] and AHDD [42], which reproduced these results under a consistent experimental setup." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "text", + "content": "Evaluation metric. We evaluate our method on 600 unique FS-OCC tasks randomly sampled from the meta-testing set for each dataset and report average accuracy (in %) in " + }, + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "text", + "content": " confidence intervals. We ensure reliable evaluation by conducting 5 random restarts, using average results for comprehensive comparisons." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 586, + 372, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 372, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 372, + 597 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 665 + ], + "type": "text", + "content": "In this section, we perform extensive experiments to demonstrate the effectiveness of our TSGA. The analysis of variance (ANOVA) results are provided in Supplementary Materials, with all pairwise comparisons achieving statistical significance " + }, + { + "bbox": [ + 313, + 605, + 553, + 665 + ], + "type": "inline_equation", + "content": "(p < 0.05)" + }, + { + "bbox": [ + 313, + 605, + 553, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Baseline comparison. To validate the effectiveness of TSGA for FS-OCC, we conduct experiments on three datasets with baseline approaches, as shown before. The quantitative results of the baseline methods and ours with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30560" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 555, + 230 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 230 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 230 + ], + "type": "table", + "html": "
ModelBackboneK=2K=10
MINOMNCIFAR-FSMINOMNCIFAR-FS
OC-ProtoNet [12]Conv-464.488.864.566.491.667.6
OW-ProtoNet [25]Conv-467.0-70.974.4-76.7
AHDD [42]Conv-471.9396.8175.3477.0097.8780.95
OC-MetaOptNet [15]ResNet-1251.8-56.367.4-75.5
OC-MetaSGD [15]Conv-469.6-71.475.8-77.8
OC-MAML [15]Conv-469.196.670.076.297.679.1
TSGA (Ours)Conv-473.41 ± 0.498.11 ± 0.176.84 ± 0.477.73 ± 0.498.54 ± 0.181.70 ± 0.4
OC-MAML†[15]ResNet-1276.60 ± 0.498.61 ± 0.174.26 ± 0.477.27 ± 0.499.18 ± 0.182.05 ± 0.4
TSGA (Ours)ResNet-1278.04 ± 0.498.97 ± 0.179.53 ± 0.479.24 ± 0.499.29 ± 0.184.05 ± 0.4
", + "image_path": "4ea4b39d3becff03ac1c78eb874a0c20e5bd9931c414d0b3f95b551068164fd4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 64, + 231, + 145, + 243 + ], + "lines": [ + { + "bbox": [ + 64, + 231, + 145, + 243 + ], + "spans": [ + { + "bbox": [ + 64, + 231, + 145, + 243 + ], + "type": "text", + "content": "† Our reproduction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 293, + 294, + 384 + ], + "blocks": [ + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "text", + "content": "Table 1. Comparison with the metric-based and optimization-based meta-learning methods for 2-shot and 10-shot OCC tasks on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS. We highlight the best number in **bold**." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 293, + 294, + 384 + ], + "lines": [ + { + "bbox": [ + 57, + 293, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 57, + 293, + 294, + 384 + ], + "type": "table", + "html": "
ModelMINOMNCIFAR-FS
OC-ProtoNet [12]61.8986.0262.15
OW-ProtoNet [25]62.0094.90-
AHDD [42]68.8995.1671.89
OC-MAML†[15]69.6297.2173.76
TSGA (Ours)71.6797.3274.77
", + "image_path": "e0adfbfd7d048564cdbec9cd2e2a7522c17193e8274caa5f7fe541d3490490b3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 64, + 385, + 145, + 397 + ], + "lines": [ + { + "bbox": [ + 64, + 385, + 145, + 397 + ], + "spans": [ + { + "bbox": [ + 64, + 385, + 145, + 397 + ], + "type": "text", + "content": "† Our reproduction." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 453, + 294, + 502 + ], + "blocks": [ + { + "bbox": [ + 55, + 406, + 295, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 406, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 295, + 440 + ], + "type": "text", + "content": "Table 2. One-shot OCC tasks comparison results of ours and baseline methods on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS. We highlight the best number in bold." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 453, + 294, + 502 + ], + "lines": [ + { + "bbox": [ + 57, + 453, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 453, + 294, + 502 + ], + "type": "table", + "html": "
ModelOP00OP01OP02Avg
OC-MAML75.5665.6395.1578.78
TSGA76.0566.6597.4580.05
", + "image_path": "033abea0e75b2d30e224525b0241bbe67fa715fd94544f09c9e78a69415511dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 568, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 713 + ], + "type": "text", + "content": "various training sample numbers (" + }, + { + "bbox": [ + 55, + 568, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 568, + 295, + 713 + ], + "type": "text", + "content": "-shot) are reported in Tables 1 and 2. We can observe from the results: 1) TSGA consistently improves the baseline OC-MAML by a large margin in all settings, demonstrating its effectiveness in improving gradient stability and model generalization to FS-OCC. 2) TSGA achieves significant performance gains on both the Conv-4 and ResNet-12 backbones, suggesting TSGA's flexibility. 3) TSGA can tackle both the one-shot and few-shot FS-OCC tasks, whereas several FS-OCC baselines (e.g., Meta-SVDD [12]) are unsuitable for one-shot OCC, qualitative results are shown in Section 4.4. State-of-the-art comparison. Here we compare our TSGA" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 348, + 293, + 520, + 355 + ], + "blocks": [ + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "text", + "content": "Table 3. Comparison of OC-MAML and TSGA on several operations (i.e., OP) of the real-world CNC Milling Machine dataset, with only " + }, + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "text", + "content": " normal examples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 293, + 520, + 355 + ], + "lines": [ + { + "bbox": [ + 348, + 293, + 520, + 355 + ], + "spans": [ + { + "bbox": [ + 348, + 293, + 520, + 355 + ], + "type": "table", + "html": "
ModelMINOMN
MAML [14]0.280.16
OC-MAML [15]0.420.23
TSGA (Ours)0.470.39
", + "image_path": "e7cc0348cc0324e080a61323d8bcc1bb3fc5fa534eeda516e83104c944acd99f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 363, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 397 + ], + "type": "text", + "content": "Table 4. Comparison of cosine similarity between the gradients of one-class support set and class-balanced query set averaged over test tasks of miniImageNet (MIN) and Omniglot (OMN)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 423, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 567 + ], + "type": "text", + "content": "with the state-of-the-art metric-based method AHDD [42] and conduct experiments on three datasets. AHDD learns radius " + }, + { + "bbox": [ + 313, + 423, + 555, + 567 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 423, + 555, + 567 + ], + "type": "text", + "content": ", which is involved in the model optimization. This can be regarded as a type of loss function learning. During the meta-testing phase, the radius remains unchanged while our TSGA further captures the task-specific knowledge to refine the adaptation process. In Table 1, we can observe that TSGA considerably outperforms the strong baseline AHDD and establishes new state-of-the-art results in all settings, which further confirms the effectiveness of our TSGA and the mechanism of adaptation to the unseen test tasks of optimization-based methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Results of real-world CNC Milling Machine dataset. We conduct experiments on the industrial process monitoring of brownfield milling machines dataset [52] and report the results in Table 3. We compute accuracy for evaluation. The dataset consists of 1702 acceleration data built with 15 different tool operations that run on 3 machines at different time frames. The industrial dataset has an unbalanced rate of 23.3:1 between normal (i.e., positive) and non-normal data. Each data can be seen as a multivariate time series (three signals, i.e., " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "-axis). We created samples by randomly cropping windows of length 2048 from the corresponding time series of each operation and partition oper-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30561" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 165 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 165 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 165 + ], + "type": "table", + "html": "
IDSettingK=2K=10
MINOMNCIFAR-FSMINOMNCIFAR-FS
ABaseline69.196.670.076.297.679.1
B+A72.1397.8975.9176.4698.3181.11
C+N72.0997.9975.5976.6698.4581.13
D+A+N(TSGA)73.41 (+4.31)98.11 (+1.51)76.84 (+6.84)77.73 (+1.53)98.54 (+0.94)81.70 (+2.60)
", + "image_path": "3cd251a9873e5c497015fcf71cf325a1376e053104617b15325780239c541b6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 206, + 555, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "type": "text", + "content": "Table 5. Ablation studies for the designed components of TSGA (including the loss network " + }, + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "type": "text", + "content": " and adapter network " + }, + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "type": "text", + "content": ") on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS for 2-shot and 10-shot FS-OCC tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 206, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 56, + 206, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 555, + 327 + ], + "type": "table", + "html": "
Cross-EntropyLoss NetworkK=2K=10
FeatureLabelMINOMNCIFAR-FSMINOMNCIFAR-FS
✓XX69.196.670.076.297.679.1
✓X✓71.13 ± 0.497.95 ± 0.176.66 ± 0.477.60 ± 0.498.27 ± 0.181.11 ± 0.4
✓✓✓72.15 ± 0.497.91 ± 0.177.11 ± 0.477.19 ± 0.498.60 ± 0.180.66 ± 0.4
XX✓72.61 ± 0.498.01 ± 0.176.18 ± 0.477.31 ± 0.498.45 ± 0.181.46 ± 0.4
X✓✓73.41 ± 0.498.11 ± 0.176.84 ± 0.477.73 ± 0.498.54 ± 0.181.70 ± 0.4
", + "image_path": "3b2217fef662c7563f87943b12bfb30e7e8dd8f4258675619bbdd4b7a736af4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 376, + 294, + 454 + ], + "blocks": [ + { + "bbox": [ + 55, + 335, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 555, + 357 + ], + "type": "text", + "content": "Table 6. The impact of loss network components (Feature and Label represent feature representations and label predictions, respectively) and cross-entropy loss on miniImageNet (MIN), Omniglot (OMN), and CIFAR-FS for 2-shot and 10-shot OCC tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 376, + 294, + 454 + ], + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 454 + ], + "type": "table", + "html": "
# Layer# DimAccuracy# Layer# DimAccuracy
1481.604481.28
881.84880.44
2481.398481.47
881.99880.65
", + "image_path": "1b21abd0079701b4b5a6e893c18b691b9162244f5159b8c21932c2c0df8ca523.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 295, + 496 + ], + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 496 + ], + "type": "text", + "content": "Table 7. Different adapter networks results on CIFAR-FS in 10-shot OCC tasks. # Layer and # Dim represent the number of layers and the hidden dimension, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "content": "ations into 10 types for meta-training, 2 for meta-validation, and 3 for meta-testing. TSGA consistently outperforms OC-MAML across the 3 types of operations and achieves " + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "content": " accuracy on average by using only " + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "content": " normal examples, demonstrating its competence in the time-series data domain and for industrial real-world applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 592, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 592, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 592, + 295, + 714 + ], + "type": "text", + "content": "Cosine Similarity Analysis. The success of OC-MAML lies in maximizing the inner product, and therefore the cosine similarity, of the gradients computed on the one-class support set and class-balanced query set " + }, + { + "bbox": [ + 54, + 592, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\partial(\\overline{g}_1\\cdot\\overline{g}_2)}{\\partial\\theta}" + }, + { + "bbox": [ + 54, + 592, + 295, + 714 + ], + "type": "text", + "content": ". Following OC-MAML, for each task, we use the initialization meta-learned by each algorithm to compute the loss gradients of support and query examples, and then compute the average cosine similarity across tasks. The cosine similarity of MAML, OC-MAML, and TSGA on test tasks from two datasets is reported in Table 4, showing that TSGA achieves" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 377, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 415 + ], + "type": "text", + "content": "higher similarity values, indicating its ability to effectively align support and query gradients. This validates the improved task-specific adaptation capabilities of TSGA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 421, + 406, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 406, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 406, + 434 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "content": "Here, we conduct ablative analyses to investigate the designed components of TSGA in Table 5. We also study the impact of different supervision caused by loss network and adapter network architectures on model performance in Tables 6 and 7. Please refer to Supplementary Materials for more other details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": "Effectiveness of the designed components. TSGA contains two components, including a loss network " + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": " and an adapter network " + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": ". We conduct component-wise analysis by alternatively adding one of them to understand the influence of each component in Table 5. We take the OC-MAML as baseline (A). " + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": ": only leverage loss network or adapter network. " + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": ": our TSGA. As can be seen, each component in TSGA contributes to the performance. In particular, the results in " + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": " suggest that the loss network and the adapter network complement each other to improve the FS-OCC performance. The results in " + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 313, + 510, + 556, + 666 + ], + "type": "text", + "content": " verify the effectiveness of learnable supervision for OCC tasks, and task-specific adaptation for various tasks, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Analysis of the loss network. TSGA leverages label predictions and feature representations to learn loss values. It is important to answer the question: Does the feature part matter? To this end, we remove the feature representations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30562" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 156, + 149 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 156, + 149 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 156, + 149 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 156, + 149 + ], + "type": "image", + "image_path": "ecd27b087e872ef4dea8303d476947b3170d542e0ae17bf4a699530805997fd9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 158, + 72, + 255, + 149 + ], + "blocks": [ + { + "bbox": [ + 158, + 72, + 255, + 149 + ], + "lines": [ + { + "bbox": [ + 158, + 72, + 255, + 149 + ], + "spans": [ + { + "bbox": [ + 158, + 72, + 255, + 149 + ], + "type": "image", + "image_path": "f3eb8004aa2a5611b6d826fa8b2d4005351bdb76714226f1247683ee4d1b7d20.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 256, + 72, + 354, + 149 + ], + "blocks": [ + { + "bbox": [ + 256, + 72, + 354, + 149 + ], + "lines": [ + { + "bbox": [ + 256, + 72, + 354, + 149 + ], + "spans": [ + { + "bbox": [ + 256, + 72, + 354, + 149 + ], + "type": "image", + "image_path": "d8012c05a1f691ed37b64a74a01c46fd0dda752e1667d463e6ef132fa0ec3c56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 360, + 72, + 454, + 149 + ], + "blocks": [ + { + "bbox": [ + 360, + 72, + 454, + 149 + ], + "lines": [ + { + "bbox": [ + 360, + 72, + 454, + 149 + ], + "spans": [ + { + "bbox": [ + 360, + 72, + 454, + 149 + ], + "type": "image", + "image_path": "1ed004893d2036f21f8662bea6fed907aa7744b8412c76002cec877c30d47648.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 457, + 72, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 457, + 72, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 457, + 72, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 457, + 72, + 553, + 149 + ], + "type": "image", + "image_path": "e97212de60e6520feb265e84d5955ea4c25ee714318d0aef562572335cfb0e53.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 150, + 156, + 225 + ], + "blocks": [ + { + "bbox": [ + 62, + 150, + 156, + 225 + ], + "lines": [ + { + "bbox": [ + 62, + 150, + 156, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 150, + 156, + 225 + ], + "type": "image", + "image_path": "8e28e474bb0a02b4bdf0f9b604f712ac5f4c1e92be6e0eace92733ee95c138da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 157, + 150, + 255, + 225 + ], + "blocks": [ + { + "bbox": [ + 157, + 150, + 255, + 225 + ], + "lines": [ + { + "bbox": [ + 157, + 150, + 255, + 225 + ], + "spans": [ + { + "bbox": [ + 157, + 150, + 255, + 225 + ], + "type": "image", + "image_path": "648a4aa6feacef866d7a478db3e3c41609bc2b7eaa4b4cdcc9a7c9494eda481f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 257, + 150, + 354, + 225 + ], + "blocks": [ + { + "bbox": [ + 257, + 150, + 354, + 225 + ], + "lines": [ + { + "bbox": [ + 257, + 150, + 354, + 225 + ], + "spans": [ + { + "bbox": [ + 257, + 150, + 354, + 225 + ], + "type": "image", + "image_path": "1b561d026264bbbffc42b4286688c7c43a336561faab3943b2b800267807650b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 357, + 150, + 454, + 225 + ], + "blocks": [ + { + "bbox": [ + 357, + 150, + 454, + 225 + ], + "lines": [ + { + "bbox": [ + 357, + 150, + 454, + 225 + ], + "spans": [ + { + "bbox": [ + 357, + 150, + 454, + 225 + ], + "type": "image", + "image_path": "ced786a4f3f6ef9b6b4ae55883294d29199db71d4cc7030261c4ccb819f60608.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 457, + 150, + 553, + 225 + ], + "blocks": [ + { + "bbox": [ + 457, + 150, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 457, + 150, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 457, + 150, + 553, + 225 + ], + "type": "image", + "image_path": "430d165be57993f87364065ed39d27f353bcede59fc5ae011ef510bc91664ead.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 63, + 226, + 156, + 302 + ], + "blocks": [ + { + "bbox": [ + 63, + 226, + 156, + 302 + ], + "lines": [ + { + "bbox": [ + 63, + 226, + 156, + 302 + ], + "spans": [ + { + "bbox": [ + 63, + 226, + 156, + 302 + ], + "type": "image", + "image_path": "e85460941e7c1b0aa7b99b0c76b84e3c9594ca273b9a13cc188ab3fb64ba462b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 157, + 226, + 255, + 302 + ], + "blocks": [ + { + "bbox": [ + 157, + 226, + 255, + 302 + ], + "lines": [ + { + "bbox": [ + 157, + 226, + 255, + 302 + ], + "spans": [ + { + "bbox": [ + 157, + 226, + 255, + 302 + ], + "type": "image", + "image_path": "7fb17025fc6feef622e851cf50470e22508651ca6fea99cba86199c793f7b684.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 258, + 226, + 354, + 302 + ], + "blocks": [ + { + "bbox": [ + 258, + 226, + 354, + 302 + ], + "lines": [ + { + "bbox": [ + 258, + 226, + 354, + 302 + ], + "spans": [ + { + "bbox": [ + 258, + 226, + 354, + 302 + ], + "type": "image", + "image_path": "acd905ba7a3ae03bb6eff2bedb42fb0d990e5b1d467757fb65dbc2a35e1fb2f3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 356, + 226, + 453, + 302 + ], + "blocks": [ + { + "bbox": [ + 356, + 226, + 453, + 302 + ], + "lines": [ + { + "bbox": [ + 356, + 226, + 453, + 302 + ], + "spans": [ + { + "bbox": [ + 356, + 226, + 453, + 302 + ], + "type": "image", + "image_path": "0a6825eb0c9354d0e4d3bb55f6d513295bdc1756ac003190a08edab76df3107a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 459, + 226, + 553, + 302 + ], + "blocks": [ + { + "bbox": [ + 459, + 226, + 553, + 302 + ], + "lines": [ + { + "bbox": [ + 459, + 226, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 459, + 226, + 553, + 302 + ], + "type": "image", + "image_path": "2b69de7149a731f120dcdf80e743626d6a3dcb1baaf092a0b90ec26f9fd28a41.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 58, + 302, + 156, + 378 + ], + "blocks": [ + { + "bbox": [ + 58, + 302, + 156, + 378 + ], + "lines": [ + { + "bbox": [ + 58, + 302, + 156, + 378 + ], + "spans": [ + { + "bbox": [ + 58, + 302, + 156, + 378 + ], + "type": "image", + "image_path": "039ee51b641a54a7b0207600b179790b3c35976624f9128bc60981811032b4a4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 388, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 555, + 411 + ], + "type": "text", + "content": "Figure 3. Visualization of the generated values " + }, + { + "bbox": [ + 55, + 388, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 388, + 555, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 388, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 388, + 555, + 411 + ], + "type": "text", + "content": " of adapter network, across adaptation steps and model layers (i.e., weights and biases of 4 convolutional layers and 1 fully connected layer in CONV-4) for 10-shot OCC tasks on miniImageNet." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 157, + 302, + 255, + 378 + ], + "blocks": [ + { + "bbox": [ + 157, + 302, + 255, + 378 + ], + "lines": [ + { + "bbox": [ + 157, + 302, + 255, + 378 + ], + "spans": [ + { + "bbox": [ + 157, + 302, + 255, + 378 + ], + "type": "image", + "image_path": "32618484f2e8033e40559bc1d3bcf28471ebd6cc0b9016692293476d1db05e0d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 257, + 302, + 354, + 378 + ], + "blocks": [ + { + "bbox": [ + 257, + 302, + 354, + 378 + ], + "lines": [ + { + "bbox": [ + 257, + 302, + 354, + 378 + ], + "spans": [ + { + "bbox": [ + 257, + 302, + 354, + 378 + ], + "type": "image", + "image_path": "0b88c95f34af3c38c11131af22f33af961c398276715c90e0f8c1b6e5f766f16.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 356, + 302, + 453, + 378 + ], + "blocks": [ + { + "bbox": [ + 356, + 302, + 453, + 378 + ], + "lines": [ + { + "bbox": [ + 356, + 302, + 453, + 378 + ], + "spans": [ + { + "bbox": [ + 356, + 302, + 453, + 378 + ], + "type": "image", + "image_path": "b824d6165523811fba2f2ad7e883681e6ddcedcbc150ccec9009b4658847ea72.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 459, + 302, + 553, + 378 + ], + "blocks": [ + { + "bbox": [ + 459, + 302, + 553, + 378 + ], + "lines": [ + { + "bbox": [ + 459, + 302, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 459, + 302, + 553, + 378 + ], + "type": "image", + "image_path": "8614f4ba23e82d2599332cb85a33c8bb9ad5c4abe82d8ebdca70715cb9365cb6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 432, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 432, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 54, + 432, + 295, + 503 + ], + "type": "text", + "content": "The results are reported in Table 6. Not surprisingly, without feature representations, performance gains decline. Further, when the cross-entropy loss and loss network are used together, there is no significant difference from when only the latter is used, implying that our loss network can comprehend the cross-entropy loss information." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 54, + 505, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 505, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 505, + 295, + 589 + ], + "type": "text", + "content": "Analysis of the adapter network. We study the impacts of the number of layers and the hidden size of the adapter network on performance. Table 7 shows that a larger number of layers and a bigger hidden size do not result in significant performance gains. TSGA is in general not sensitive to the choice of the number of layers and the hidden size within a certain range." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 599, + 139, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 139, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 139, + 611 + ], + "type": "text", + "content": "4.4. Visualization" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": "Here, we provide qualitative visualizations to see how our method works. In Figure 2, we present the visualization of the loss gradient dynamics [47], evaluating optimization stability and smoothness by measuring loss variations, changes in gradients, and the maximum difference in gradients. As observed, TSGA exhibits lighter shades in Figures 2a and 2b and lower values in Figure 2c, which indicate a smoother loss landscape. In Figure 3, we show the" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 432, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 515 + ], + "type": "text", + "content": "task-specific hyperparameters generated by the adapter network averaged across tasks, which adaptively vary across the inner-loop steps and model layers. The variability of the generated values highlights the effectiveness of our TSGA in refining optimization processes for the given task. More visualization of training curves is in Supplementary Materials." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 546, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 676 + ], + "type": "text", + "content": "In this work, we propose TSGA for few-shot one-class classification (FS-OCC). TSGA learns to generate stable and smooth gradients by leveraging label prediction and feature representation details and enhances the adaptation process by recalibrating task-specific gradients and regularization terms. We evaluate TSGA on three challenging datasets and a real-world CNC Milling Machine dataset, demonstrating that TSGA consistently improves performance. We also reveal the severity of unstable gradients and ignoring of task-specific adaptation, by tackling these issues TSGA establishes new state-of-the-art results." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the National Natural Science Foundation of China (82171965, 62302045)." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30563" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Latent space autoregression for novelty detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 481-490, 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 204, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 236 + ], + "type": "text", + "content": "[3] Antreas Antoniou, Harri Edwards, and Amos Storkey. How to train your maml. In Seventh International Conference on Learning Representations, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 295, + 282 + ], + "type": "text", + "content": "[4] Sungyong Baik, Myungsub Choi, Janghoon Choi, Heewon Kim, and Kyoung Mu Lee. Meta-learning with adaptive hyperparameters. Advances in neural information processing systems, 33:20755-20765, 2020. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 338 + ], + "type": "text", + "content": "[5] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9465-9474, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 340, + 295, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 295, + 395 + ], + "type": "text", + "content": "[6] Niamh Belton, Misgina Tsighe Hagos, Aonghus Lawlor, and Kathleen M Curran. Fewsome: One-class few shot anomaly detection with siamese networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2978-2987, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 430 + ], + "type": "text", + "content": "[7] Liron Bergman and Yedid Hoshen. Classification-based anomaly detection for general data. In International Conference on Learning Representations. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "type": "text", + "content": "[8] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 497 + ], + "type": "text", + "content": "[9] Raghavendra Chalopathy and Sanjay Chawla. Deep learning for anomaly detection: A survey. arXiv preprint arXiv:1901.03407, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 499, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 499, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 499, + 295, + 542 + ], + "type": "text", + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 544, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 588 + ], + "type": "text", + "content": "[11] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 622 + ], + "type": "text", + "content": "[12] Gabriel Dahia and Maurício Pamplona Segundo. Meta learning for few-shot one-class classification. AI, 2(2):195-208, 2021. 1, 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[14] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 1, 3, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[15] Ahmed Frikha, Denis Krompaß, Hans-Georg Köpken, and Volker Tresp. Few-shot one-class classification via meta-learning. In Proceedings of the AAAI conference on artificial intelligence, pages 7448–7456, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 150 + ], + "type": "text", + "content": "[16] Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric transformations. Advances in neural information processing systems, 31, 2018. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 151, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 554, + 194 + ], + "type": "text", + "content": "[17] Sachin Goyal, Aditi Raghunathan, Moksh Jain, Harsha Vardhan Simhadri, and Prateek Jain. Drocc: Deep robust one-class classification. In International conference on machine learning, pages 3711-3721. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "text", + "content": "[18] Zhaoyang Hai, Liyuan Pan, Xiabi Liu, Zhengzheng Liu, and Mirna Yunita. L2t-dln: Learning to teach with dynamic loss network. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "type": "text", + "content": "[20] Timothy Hospedales, Antreas Antoniou, Paul Micaelli, and Amos Storkey. Meta-learning in neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(9):5149-5169, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 554, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 382 + ], + "type": "text", + "content": "[21] Chen Huang, Shuangfei Zhai, Walter Talbott, Miguel Bautista Martin, Shih-Yu Sun, Carlos Guestrin, and Josh Susskind. Addressing the loss-metric mismatch with adaptive loss alignment. In International conference on machine learning, pages 2891-2900. PMLR, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 554, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 436 + ], + "type": "text", + "content": "[22] Chen Huang, Shuangfei Zhai, Pengsheng Guo, and Josh Susskind. Metricopt: Learning to optimize black-box evaluation metrics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 174-183, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 437, + 554, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 481 + ], + "type": "text", + "content": "[23] Tao Huang, Zekang Li, Hua Lu, Yong Shan, Shusheng Yang, Yang Feng, Fei Wang, Shan You, and Chang Xu. Relational surrogate loss learning. In International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 482, + 554, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 554, + 504 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 505, + 554, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 554, + 526 + ], + "type": "text", + "content": "[25] Anna Kruspe. One-way prototypical networks. arXiv preprint arXiv:1906.00820, 2019. 1, 3, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 526, + 554, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 569 + ], + "type": "text", + "content": "[26] Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332-1338, 2015. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "type": "text", + "content": "[27] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "text", + "content": "[28] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1, 3, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[29] Lanlan Liu, Mingzhe Wang, and Jia Deng. A unified framework of surrogate loss by refactoring and interpolation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 278-293. Springer, 2020. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30564" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[30] Qingliang Liu and Jinmei Lai. Stochastic loss function. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4884-4891, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "type": "text", + "content": "[31] Wenrui Liu, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Diversity-measurable anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12147-12156, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 160, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 160, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 294, + 205 + ], + "type": "text", + "content": "[32] Philipp Liznerski, Lukas Ruff, Robert A Vandermeulen, Billy Joe Franks, Marius Kloft, and Klaus Robert Muller. Explainable deep one-class classification. In International Conference on Learning Representations. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 237 + ], + "type": "text", + "content": "[33] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "text", + "content": "[34] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 282, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 316 + ], + "type": "text", + "content": "[35] Yash Patel, Tomáš Hodaž, and Jiří Matas. Learning surrogates via deep embedding. In European Conference on Computer Vision, pages 205-221. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[36] Pramuditha Perera, Ramesh Nallapati, and Bing Xiang. Ocean: One-class novelty detection using gans with constrained latent representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2898-2906, 2019. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "text", + "content": "[37] Pramuditha Perera, Poojan Oza, and Vishal M Patel. One-class classification: A survey. arXiv preprint arXiv:2101.03064, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "text", + "content": "[38] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 481 + ], + "type": "text", + "content": "[39] Marco AF Pimentel, David A Clifton, Lei Clifton, and Lionel Tarassenko. A review of novelty detection. Signal processing, 99:215-249, 2014. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "text", + "content": "[40] Chen Qiu, Timo Pfrommer, Marius Kloft, Stephan Mandt, and Maja Rudolph. Neural transformation learning for deep anomaly detection beyond images. In International conference on machine learning, pages 8703-8714. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "type": "text", + "content": "[41] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[42] Yuchen Ren, Xiabi Liu, Liyuan Pan, and Lijuan Niu. Adaptive hypersphere data description for few-shot one-class classification. Applied Intelligence, pages 1-13, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "type": "text", + "content": "[43] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E Turner. Fast and flexible multi-task classification using conditional neural adaptive processes. Advances in neural information processing systems, 32, 2019. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "text", + "content": "[44] T-YLPG Ross and GKHP Dollár. Focal loss for dense object detection. In proceedings of the IEEE conference on" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 677 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "text", + "content": "computer vision and pattern recognition, pages 2980-2988, 2017.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "text", + "content": "[45] Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. In International conference on machine learning, pages 4393-4402. PMLR, 2018. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 151, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 207 + ], + "type": "text", + "content": "[46] Mohammad Sabokrou, Mohammad Khalooei, Mahmood Fathy, and Ehsan Adeli. Adversarily learned one-class classifier for novelty detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3379-3388, 2018. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[47] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? Advances in neural information processing systems, 31, 2018. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 252, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 319 + ], + "type": "text", + "content": "[48] Thomas Schlegl, Philipp Seebock, Sebastian M Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In International conference on information processing in medical imaging, pages 146-157. Springer, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[49] Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8495-8504, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 373, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 408 + ], + "type": "text", + "content": "[50] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 408, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 464 + ], + "type": "text", + "content": "[51] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 464, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 509 + ], + "type": "text", + "content": "[52] Mohamed-Ali Tnani, Michael Feil, and Klaus Diepold. Smart data collection system for brownfield cnc milling machines: A new benchmark dataset for data-driven machine monitoring. Procedia CIRP, 107:131-136, 2022. 1, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[53] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 553, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 553, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 553, + 553, + 587 + ], + "type": "text", + "content": "[54] Yizhe Wang, Congqi Cao, and Yanning Zhang. Visualsemantic network: a visual and semantic enhanced model for gesture recognition. Visual Intelligence, 1(1):25, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 587, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 587, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 587, + 553, + 632 + ], + "type": "text", + "content": "[55] Lijun Wu, Fei Tian, Yingce Xia, Yang Fan, Tao Qin, Lai Jian-Huang, and Tie-Yan Liu. Learning to teach with dynamic loss functions. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 632, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 632, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 632, + 553, + 677 + ], + "type": "text", + "content": "[56] Ji Zhang, Lianli Gao, Xu Luo, Hengtao Shen, and Jingkuan Song. Deta: Denoised task adaptation for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11541-11551, 2023. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30565" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_content_list.json b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..be23a7ca153ced812520c3bf44de5518e8bfae3c --- /dev/null +++ b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_content_list.json @@ -0,0 +1,2008 @@ +[ + { + "type": "text", + "text": "Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding", + "text_level": 1, + "bbox": [ + 94, + 128, + 903, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenbo Chen1*; Zhen Xu1,2*; Ruotao Xu2, Si Wu1,2†; Hau-San Wong3", + "bbox": [ + 233, + 202, + 761, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, South China University of Technology $^{2}$ Institute for Super Robotics (Huangpu)", + "bbox": [ + 158, + 220, + 836, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Computer Science, City University of Hong Kong", + "bbox": [ + 236, + 256, + 759, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cscwb106, csxuzhen}@mail.scut.edu.cn, {xrt, cswusi}@scut.edu.cn, cshswong@cityu.edu.hk", + "bbox": [ + 112, + 276, + 877, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of visual grounding is to establish connections between target objects and textual descriptions. Large Language Models (LLMs) have demonstrated strong comprehension abilities across a variety of visual tasks. To establish precise associations between the text and the corresponding visual region, we propose a Task-aware Cross-modal feature Refinement Transformer with LLMs for visual grounding, and our model is referred to as TCRT. To enable the LLM trained solely on text to understand images, we introduce an LLM adaptation module that extracts text-related visual features to bridge the domain discrepancy between the textual and visual modalities. We feed the text and visual features into the LLM to obtain task-aware priors. To enable the priors to guide the feature fusion process, we further incorporate a cross-modal feature fusion module, which allows task-aware embeddings to refine visual features and facilitate information interaction between the Referring Expression Comprehension (REC) and Referring Expression Segmentation (RES) tasks. We have performed extensive experiments to verify the effectiveness of the main components and demonstrate the superior performance of the proposed TCRT over state-of-the-art end-to-end visual grounding methods on RefCOCO, RefCOCOg, RefCOCO+ and ReferItGame.", + "bbox": [ + 88, + 358, + 485, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 750, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual grounding aims to predict the locations of the objects specified by textual prompts in images, and is widely applied in the field of human-computer interaction. This task typically involves referring expression comprehension [10, 17, 49, 59] and referring expression segmentation [22, 38]. The main challenge is to comprehensively un", + "bbox": [ + 89, + 773, + 483, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f4c1f36b10a23daa8ce5df93c45225be6adc8c0b185ac6ecb1488ce8e50cb48f.jpg", + "image_caption": [ + "TCRT w/o LLM", + "\"bride throwing bouquet" + ], + "image_footnote": [], + "bbox": [ + 522, + 335, + 616, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/838a2900a098502cd3251275b99d6c3552bebf15354bec8bda1f28a4a0faae6e.jpg", + "image_caption": [ + "\"Guy front in gray suit left\"" + ], + "image_footnote": [], + "bbox": [ + 522, + 393, + 616, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/44aa74d849d420267f520734e534fa16bcd1bc4138d5ced316d320dc07448cca.jpg", + "image_caption": [ + "\"man in right front with blue tie next to older woman", + "Figure 1. The visualization of the refined visual features from the cross-modal feature fusion module. In challenging environments, the introduction of LLMs can facilitate a better association between visual features and text." + ], + "image_footnote": [], + "bbox": [ + 522, + 455, + 616, + 510 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e38bddff4d03ab149e64c193388b3c560b3621c1a333227c8d25cf4b29b648ef.jpg", + "image_caption": [ + "TCRT" + ], + "image_footnote": [], + "bbox": [ + 712, + 335, + 803, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0962ec6767962463598628d3cc48dd6dd4c8707d784c10251d85cf4c2043000.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 396, + 803, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83e68a042ef610ba44bcc52c4e667e60436031a0f2272c2ac6a441ab209a27ff.jpg", + "image_caption": [ + "predicted by" + ], + "image_footnote": [], + "bbox": [ + 712, + 455, + 803, + 510 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e1319efd6665c724e37d17b0c6b12437b265ab2ad1953f9b183f6914b1df72cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 335, + 897, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/974bc6bdd894f345e634f5c8d15cbe248c63ba37df684f639d90a33d281e066e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 396, + 897, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc54ee27f0edb5b8c64c46f960933757efa8845894bb5673567acd8d0319ae5e.jpg", + "image_caption": [ + "ground truth" + ], + "image_footnote": [], + "bbox": [ + 805, + 455, + 897, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "derstand intra-modality information and accurately establish inter-modality correspondences.", + "bbox": [ + 511, + 625, + 906, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing visual grounding methods can be categorized into two main types: two-stage and end-to-end methods. The two-stage methods [45] first generate region proposals [36] that may match the text, and then rank the candidate regions according to their relevance to the text. When the proposals generated in the first stage do not contain the target object, the second stage will be unable to correct these proposals. On the other hand, end-to-end grounding methods perform the integration of visual and textual features within an object detector, producing the box coordinates of the target object. The end-to-end methods are typically faster than the two-stage ones. With the advancement of transformer in the vision domain, TransVG [9, 10] and SeqTR [59] employ self-attention or cross-attention mechanisms to fuse textual and visual features. Recently, LLaVA [25] has integrated visual understanding capabilities into large language mod", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Joint first authors.", + "bbox": [ + 107, + 875, + 212, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els (LLMs). MiniGPT-2 [4] and OMG-LLaVA [58] apply LLMs to visual grounding tasks by integrating textual and visual features. In this work, we make an attempt to adapt LLMs to capture the information of target objects and produce task-aware embeddings for precise grounding results.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, we propose a task-aware cross-modal feature refinement transformer with large language models for visual grounding, which is referred to as TCRT, leveraging priors from LLMs to guide the multimodal information fusion process and achieve precise alignment between textual and local visual features (see Figure 1). To enable the LLM to more effectively process image data and adapt to visual grounding tasks, we propose an LLM adaptation module that utilizes conditional normalization to adjust the statistical characteristics of visual features based on textual information, while introducing object-text semantic matching regularization to associate local visual features with textual features. Additionally, to utilize the priors provided by the adaptation module, we incorporate a task-aware cross-modal feature fusion module, which continuously refines visual features during multimodal fusion and enables the model to exploit the relationship between the REC and RES tasks. We perform extensive qualitative and quantitative experiments to validate the effectiveness of our model, and demonstrate its advantages on benchmark datasets. The main contributions of this work are summarized as follows:", + "bbox": [ + 91, + 169, + 482, + 484 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Different from the state-of-the-art methods that focus on the application of LLMs in integrating textual and visual features, we further explore the LLM's capability of learning task-aware embeddings to facilitate visual grounding.", + "- We incorporate an LLM adaptation module for visual grounding, which learns to capture the information of target objects by modulating the statistical characteristics of visual features conditioned on textual features.", + "- We further perform task-aware cross-modal feature fusion, allowing the task-aware embeddings derived from LLMs to guide the association of textual and visual information and lead to precise grounding results." + ], + "bbox": [ + 109, + 488, + 478, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 700, + 232, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Two-stage Visual Grounding", + "text_level": 1, + "bbox": [ + 89, + 726, + 346, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In two-stage methods, a region proposal network [15, 16, 36] is first used to produce a set of candidate regions that may correspond to the target object mentioned in the textual description. These candidate regions are then matched to the textual description, and the predicted result is determined by identifying the region with the highest score. The maximum-margin ranking loss [30] is widely adopted for the second stage. Several researches further enhance the two-stage methods by incorporating context modeling [18, 32, 44, 55] and making use of phrase cooccur", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ences [2, 5, 11]. MAttNet [52] decomposed the expression into three modules: relationship, location, and subject to improve the grounding accuracy. To enhance the model's capacity for learning detailed alignments between text and visuals, CM-A-E [27] proposed a method capable of generating challenging samples for training the model. Graph learning was incorporated into visual grounding by SGMN [46] to enable the model to better capture the relationships between different objects within the same scene. To increase the recall of target objects, Ref-NMS [6] utilized textual features to direct the non-maximum suppression. To enhance the model's cross-modality fusion capabilities, GroundingDINO [26] integrated grounded pretraining techniques into detection transformers.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. End-to-end Visual Grounding", + "text_level": 1, + "bbox": [ + 511, + 314, + 779, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end methods in visual grounding aim to directly predict the bounding boxes in an image from a text in a single pass, resulting in faster inference speeds compared to two-stage methods. As a pioneering work, FAOA [47] concatenated image and textual features, and fed the concatenated features into a prediction layer to achieve target object detection. To reduce the referring ambiguity, ReSC [48] used a sub-query construction to address limitations in processing complex textual descriptions. TransVG [9, 10] is one of the pioneering transformer-based visual grounding models that directly regresses coordinates. To achieve multi-task learning, RefTR [22] used the transformer architecture to integrate REC and RES tasks into a single framework. To address the gap between the visual and textual backbones, QRNet [49] incorporated textual information into the visual backbone. Furthermore, VG-LAW [38] actively extracted text-relevant visual features based on text-adaptive weights and performed multi-task visual grounding. Additionally, PVD [7] introduced a diffusion model into the visual grounding task.", + "bbox": [ + 511, + 335, + 903, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the development of LLMs, some Multimodal Large Language Models [25] (MLLMs) are gradually being applied to visual grounding tasks. VistaLLM [34] used an instruction-guided image tokenizer to filter and compress image embeddings based on task, extracting visual information relevant to the current task. To efficiently locate and understand multiple target regions in complex scenes, Groma [29] introduced a localized visual tokenization mechanism that decomposes an image into multiple regions of interest, encoding each region as an independent region token. To unleash the potential of large language models, Griffon [53] introduces a foundational language-prompted localization dataset with nearly 6 million pretraining samples.", + "bbox": [ + 511, + 638, + 903, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Large Language Models", + "text_level": 1, + "bbox": [ + 511, + 847, + 736, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the expansion of training data and a significant increase in model parameters, large language models (LLMs)", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3932", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/976b3e3cd8eb1c19697c61cabd6b556ac4756f0ce8920a55e449140240bc4f18.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed method. We first extract textual features $F_{txt}$ using the textual backbone $E_{txt}$ . These features are then injected into the visual backbone $E_{vis}$ , which consists of transformer blocks and conditional normalization, to extract text-related visual features $F_{vis}$ . The object-text semantic matching regularization is employed to associate visual region tokens with textual tokens. We input the visual features processed by the visual adapter $V_{proj}$ , along with the text, into LLaMa. We take the last token $f_{llm}$ output from LLaMa. $f_{llm}$ is then fed into a query projection layer to generate task-relevant tokens $f_{box}$ , $f_{seg}$ , and $f_{phr}$ , which are used to predict the bounding box, segmentation, and phrase, respectively. A task-aware cross-modal feature fusion module is then employed to integrate textual, visual, and task features. Finally, task-relevant tokens are each fed to the corresponding prediction heads to produce the final predictions." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 903, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "based on the Transformer [42] architecture have developed rapidly. GPT-3 [3], with its 175 billion parameters, demonstrates strong performance in few-shot contextual learning. Following the success of GPT-3, several other LLMs, including PaLM [8], OPT [57], BLOOM [21], and LLaMA [40, 41], have emerged in succession. Mistral [19] introduced a specialized window attention mechanism to better handle extended context modeling. In parallel, Mixtral $8 \\times 7$ B utilized sparse MoE layers [14] to scale up parameters more effectively, enabling it to achieve higher performance with a reduced number of active parameters.", + "bbox": [ + 88, + 415, + 480, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main differences between the proposed TCRT and existing visual grounding methods are summarized as follows: First, we perform statistical modulation on visual features conditioned on textural features to reduce domain discrepancy before integrating them, while the existing methods [4, 53, 54] only perform dimension adaption. Second, TCRT leverage the task-aware embeddings derived from LLMs to guide cross-modal feature fusion and association analysis, rather than directly predicting the bounding boxes of target objects as in previous methods [4, 53, 54].", + "bbox": [ + 88, + 580, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 747, + 225, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setting", + "text_level": 1, + "bbox": [ + 89, + 772, + 250, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin by defining multi-task visual grounding (MTVG), and then proceed to describe the architecture of the proposed framework. The goal of MTVG is to detect or segment a target object based on a textual description. Formally, given an image $I \\in \\mathbb{R}^{3 \\times H \\times W}$ and a corresponding text $q$ , this task is to predict a bounding box $b = \\{x, y, w, h\\}$ that encompasses the object men", + "bbox": [ + 88, + 794, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tioned in the text, as well as a pixel-wise segmentation mask $s\\in \\mathbb{R}^{1\\times H\\times W}$ that provides a more detailed and precise localization. A textual backbone $E_{txt}$ and a visual backbone $E_{vis}$ are used to extract textual and visual features, respectively.", + "bbox": [ + 511, + 415, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. LLM Adaptation", + "text_level": 1, + "bbox": [ + 511, + 501, + 681, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-trained LLMs possess language understanding capabilities and can handle many complex tasks in natural language processing. However, due to the domain discrepancy between textual data and visual data, directly applying LLMs to grounding tasks may yield suboptimal performance. Therefore, we propose an LLM adaptation module to perform cross-modal adaptation of image information, enabling LLMs to better handle information within visual data. As shown in Figure 2, we first utilize the text encoding model $E_{txt}$ to extract the textual features of $q$ : $F_{txt} = E_{txt}(q)$ , where $F_{txt}$ represents the textual features, which include three types of textual features: sentence feature $f_{txt}^{sen} \\in \\mathbb{R}^{1 \\times d}$ , individual word features $f_{txt} \\in \\mathbb{R}^{L \\times d}$ , and average pooled word feature $f_{txt}^{mean} \\in \\mathbb{R}^{1 \\times d}$ , where $d$ and $L$ represent the dimension of textual features and the number of tokens, respectively.", + "bbox": [ + 511, + 523, + 906, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable the visual encoder to comprehend textual information, we designed a text-aware visual encoder $E_{vis}$ . Specifically, a conditioned normalization is inserted after each transformer block in the visual encoder. In the conditioned normalization, we fuse the sentence features $f_{txt}^{sen}$ and the average-pooled word features $f_{txt}^{mean}$ into a textual conditional feature $f_{txt}^{cond}$ via a weighted combination:", + "bbox": [ + 511, + 766, + 905, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t x t} ^ {\\text {c o n d}} = \\alpha^ {\\text {s e n}} f _ {t x t} ^ {\\text {s e n}} + \\alpha^ {\\text {m e a n}} \\phi \\left(f _ {t x t} ^ {\\text {m e a n}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 883, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3933", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/77472a9cf2b8779aff75b70d7bf29a59104fe024f03c5632df6d65d920c0bea4.jpg", + "image_caption": [ + "Figure 3. Object-text semantic matching regularization. We constrain the image features using a mask to isolate the object-specific features. For matching text-object pairs, we consider them as positive samples, whereas non-matching pairs are regarded as negative samples." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 480, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha^{sen}$ and $\\alpha^{mean}$ are learnable parameters that control the contribution of $f_{txt}^{sen}$ and $f_{txt}^{mean}$ , respectively. $\\phi(\\cdot)$ represents the feed-forward network. The scaling $\\gamma$ and shifting $\\beta$ parameters are determined using $f_{txt}^{cond}:\\gamma = \\phi(f_{txt}^{cond}) + \\gamma^{learn}$ , $\\beta = \\phi(f_{txt}^{cond}) + \\beta^{learn}$ , where $\\gamma^{learn}$ and $\\beta^{learn}$ are learnable parameters. Given the visual features $\\bar{F}_i^*$ , which is the output of the $i$ -th layer of transformer block, we calculate the mean $\\mu(\\bar{F}_i^*)$ and standard deviation $\\sigma(\\bar{F}_i^*)$ . The visual features $\\bar{F}_i^*$ are then modulated using the resulting scaling $\\gamma$ and shifting $\\beta$ parameters, as follows:", + "bbox": [ + 89, + 327, + 483, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {F} _ {i} ^ {*} = \\gamma \\left(\\frac {\\bar {F} _ {i} ^ {*} - \\mu (\\bar {F} _ {i} ^ {*})}{\\sigma (\\bar {F} _ {i} ^ {*}) + \\epsilon}\\right) + \\beta , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 488, + 482, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a small constant introduced to avoid division by zero. We obtain the text-aware visual features $F_{vis}$ by flattening the feature map $\\overline{F}_4^*$ extracted from the visual backbone. The image features and the input embeddings of LLMs are typically in different dimensional spaces. We use a visual adapter $V_{proj}$ to map the image features to a dimension that matches the LLM input, allowing the LLM to directly process visual information. $V_{proj}$ is a three-layer MLP. We feed the text and the mapped image features into the LLM, and take the last token of the LLM output as the prior embedding $f_{llm}$ :", + "bbox": [ + 89, + 539, + 483, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {l l m} = \\Phi (q, V _ {p r o j} \\underbrace {\\left(E _ {v i s} (I , f _ {t x t} ^ {c o n d})\\right)} _ {F _ {v i s}}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 715, + 482, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Phi$ represents an LLM model, $q$ denotes the original query text. We pass $f_{llm}$ through three task-aware linear layers ( $W_{box}$ , $W_{seg}$ , and $W_{phr}$ ) to obtain the task-specific embeddings $f_{box}$ , $f_{seg}$ , and $f_{phr}$ , which are responsible for predicting the bounding box, segmentation, and phrase, respectively.", + "bbox": [ + 89, + 763, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The text usually describes a specific object within an image, which occupies only a portion of the image. Most visual features do not have a direct relationship with the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "text. This makes it challenging to achieve precise semantic alignment between the target object and textual tokens. Thus, we propose the object-text semantic matching regularization to associate the visual region tokens with the textual ones. Object-text matching is a binary classification task, and the model evaluates the multimodal features of an object-text pair to classify it as either positive (matched) or negative (unmatched). As shown in Figure 3, if the textual features and visual features represent the same object, they are considered a positive sample pair; otherwise, they are considered a negative sample pair. The ground truth of object segmentation is represented by $s^+$ . We apply an affine transformation to $s^+$ to obtain $s^-$ . $s^-$ represents the inaccurate segmentation result. We leverage the segmentation mask of the target object to extract its features. A learnable token $f_{match}$ is then introduced to determine whether they are matched:", + "bbox": [ + 511, + 90, + 906, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\{\\bar {f} _ {\\text {m a t c h}}, \\bar {F} _ {l}, \\bar {F} _ {v} \\right\\} = \\varphi \\left(\\left[ f _ {\\text {m a t c h}}, F _ {\\text {t x t}}, F _ {\\text {v i s}} \\right]\\right), \\\\ \\left\\{\\hat {f} _ {\\text {m a t c h}} ^ {j}, \\hat {F} _ {l} \\right\\} = \\theta \\left(\\left[ \\bar {f} _ {\\text {m a t c h}}, \\bar {F} _ {l} \\right], F _ {\\text {v i s}} \\odot s ^ {j}\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 354, + 903, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j \\in \\{+, -\\}$ . $\\varphi(q)$ and $\\theta(q, kv)$ represent multi-head self-attention and cross-attention, respectively. $\\odot$ denotes element-wise multiplication. $[\\cdot]$ denotes the concatenation operation. We use the $\\hat{f}_{match}^{j}$ as the joint representation of the object-text pair. A feed-forward network is employed to predict the two-class probability $p_{match}^{j}$ that indicates whether a textual feature matches a local visual feature. The object-text matching loss is:", + "bbox": [ + 511, + 412, + 906, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a t c h}} = \\mathbb {E} _ {(I, q)} [ \\sum_ {j} H \\left(y _ {\\text {m a t c h}} ^ {j}, p _ {\\text {m a t c h}} ^ {j}\\right) ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 547, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H$ represents cross-entropy and $y_{match}^{j}$ is a 2-dimensional one-hot vector that represents the ground truth label. We use the visual features of the target object extracted from the ground truth mask as positive sample, indicating that these features match the textual description. When performing affine transformation (partial overlap may occur), the resulting mask deviates from the target object, and the extracted features are considered as negative sample. Object-text matching loss further associates visual tokens with textual ones by assessing whether they match.", + "bbox": [ + 511, + 595, + 906, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Task-aware Cross-modal Feature Fusion", + "text_level": 1, + "bbox": [ + 511, + 757, + 861, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the LLM receives textual and visual information, it can generate more meaningful semantic representations. These representations contain the LLM's understanding of objects and context, forming prior information that aids downstream tasks. The task-aware cross-modal feature fusion module, denoted as $V_{F}$ , is incorporated to guide feature the fusion process. Specifically, we first employ a bidirectional multi-head attention module to fuse textual and", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3934", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "visual features. Each modality serves as the query while the other modality acts as the key and value in the attention computation. We concatenate the three tokens used for prediction, $f_{box}$ , $f_{seg}$ , and $f_{phr}$ , with the textual features $F_{txt}$ to obtain a task-relevant textual feature $F_{task}$ . Then, $F_{task}$ and $F_{vis}$ are fed into the bi-directional multi-head attention module to fuse the two types of features. This process can be represented as follows: $F_{task}^{v2t} = \\theta (F_{task}, F_{vis})$ , $F_{vis}^{t2v} = \\theta (F_{vis}, F_{task})$ , where $F_{task}^{v2t}$ represents the textual feature enriched with image features, comprising $F_{txt}^{v2t}$ , $f_{box}^{v2t}$ , $f_{seg}^{v2t}$ , and $f_{phr}^{v2t}$ . Meanwhile, $F_{vis}^{t2v}$ denotes the visual feature enriched with textual information.", + "bbox": [ + 89, + 90, + 483, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target objects are typically located in specific regions. To enable the model to focus on local visual features, we input $f_{llm}$ into an MLP to obtain the refinement prompt $f_{prmt}$ , which guides the feature refinement process. We then assign different weights to image features at different locations by computing the cosine similarity between the image feature $F_{vis}^{t2v}$ and $f_{prmt}$ . This process can be formulated as follows:", + "bbox": [ + 89, + 273, + 483, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {v} ^ {t 2 v} = F _ {v i s} ^ {t 2 v} + F _ {v i s} ^ {t 2 v} \\odot \\underbrace {\\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {v i s} ^ {t 2 v}\\right) \\otimes \\mathrm {M L P} \\left(f _ {l l m}\\right)\\right)} _ {\\mathcal {M}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 404, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes matrix multiplication and $\\odot$ represents element-wise multiplication. The refinement prompt $f_{prmt}$ is dot-multiplied with the image features at each position in $F_{vis}^{t2v}$ to obtain a weight matrix $\\mathcal{M}$ . The final refined visual features $\\hat{F}_v^{t2v}$ are obtained under the guidance of the LLM. The model can control the strength of the feature representations at different positions in the image by adjusting the weight matrix $\\mathcal{M}$ . By assigning lower weights to irrelevant regions, the model can better focus on the features of the target regions.", + "bbox": [ + 89, + 459, + 483, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There is a close relationship between the REC task and the RES task. Both require an understanding and localization of the target object or region. REC tasks generally offer locational information of the target objects, whereas RES tasks offer precise boundary details. To strengthen the relationship between the REC and RES tasks, we perform cross-task interaction to effectively exchange information between the box token $f_{\\mathrm{box}}^{v2t}$ and the segmentation token $f_{\\mathrm{seg}}^{v2t}$ . The process is as follows:", + "bbox": [ + 89, + 613, + 483, + 751 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{f _ {b o x} ^ {\\prime}, f _ {s e g} ^ {\\prime} \\right\\} = \\theta ([ f _ {b o x} ^ {v 2 t}, f _ {s e g} ^ {v 2 t} ], F _ {t x t} ^ {v 2 t}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 763, + 482, + 795 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {b o x} ^ {\\prime \\prime} = f _ {b o x} ^ {\\prime} + \\phi (f _ {s e g} ^ {\\prime}), f _ {s e g} ^ {\\prime \\prime} = f _ {s e g} ^ {\\prime} + \\phi (f _ {b o x} ^ {\\prime}),\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 792, + 439, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This design allows the model to exploit the correlation between the REC and RES tasks. The updated $f_{box}$ , $f_{seg}$ , and $f_{phrase}$ are each fed into their respective prediction heads to obtain the final bounding box, segmentation, and phrase predictions.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Training", + "text_level": 1, + "bbox": [ + 513, + 90, + 671, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detection loss. For the REC task, the loss for bounding box regression is calculated using a combination of GIoU loss and L1 loss. We use $b$ to denote the ground truth bounding box and $\\tilde{b}$ to represent the predicted bounding box. The following defines the detection loss function:", + "bbox": [ + 511, + 112, + 906, + 188 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\det } = \\mathbb {E} _ {(I, q)} \\left[ \\Phi_ {\\text {s m o o t h} - l 1} (b, \\tilde {b}) + \\Phi_ {\\text {g i o u}} (b, \\tilde {b}) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 194, + 906, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_{\\mathrm{smooth} - l1}$ and $\\Phi_{\\mathrm{giou}}$ denote the smooth L1 loss [15] and Generalized IoU loss [37], respectively.", + "bbox": [ + 511, + 220, + 905, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segmentation loss. For the RES task, the segmentation loss is computed by combining the focal loss with dice loss. The ground truth mask is denoted as $s$ , while the predicted mask is $\\hat{s}$ . The following defines the segmentation loss function:", + "bbox": [ + 511, + 250, + 905, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s e g}} = \\mathbb {E} _ {(I, q)} \\left[ \\Phi_ {\\text {f o c a l}} (s, \\hat {s}) + \\Phi_ {\\text {d i c e}} (s, \\hat {s}) \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 318, + 906, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_{\\mathrm{focal}}$ and $\\Phi_{\\mathrm{dice}}$ denote the smooth focal loss [24] and DICE/F-1 loss [31], respectively.", + "bbox": [ + 511, + 342, + 905, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phrase loss. To enable the model to more accurately capture the entities within the text, we utilize $f_{phr}$ to predict the words in the text related to the target object. We fed $f_{phr}$ into an MLP to predict a word probability $p_{phr}$ . Then, we compute the phrase loss as follows:", + "bbox": [ + 511, + 373, + 905, + 448 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p h r}} = \\mathbb {E} _ {(I, q)} [ H (y _ {\\text {p h r}}, p _ {\\text {p h r}}) ], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 455, + 906, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{phr}$ denotes the ground truth label.", + "bbox": [ + 513, + 479, + 795, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By incorporating these training loss functions, the optimization problem of our model can be formulated as follows:", + "bbox": [ + 511, + 496, + 905, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {E _ {t x t}, E _ {v i s}} \\lambda_ {\\text {m a t c h}} \\mathcal {L} _ {\\text {m a t c h}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 549, + 905, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {E _ {t x t}, E _ {v i s}, V _ {p r o j}, W _ {t a s k}, V _ {F}} \\mathcal {L} _ {\\mathrm {d e t}} + \\mathcal {L} _ {\\mathrm {s e g}} + \\lambda_ {\\mathrm {p h r}} \\mathcal {L} _ {\\mathrm {p h r}},\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 575, + 867, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\text{task} \\in \\{box,seg,phr\\}$ , and $\\lambda_{\\mathrm{match}}$ and $\\lambda_{\\mathrm{phr}}$ are used to achieve a balance among the training objectives. $\\mathcal{L}_{\\mathrm{det}}$ and $\\mathcal{L}_{\\mathrm{seg}}$ enable the model to perform basic detection and segmentation tasks. $\\mathcal{L}_{\\mathrm{phr}}$ enables the model to achieve a more detailed understanding of the target objects referred to in the text. $\\mathcal{L}_{\\mathrm{match}}$ is used to associate visual tokens with textual ones and only optimizes the textual backbone and visual backbone. Our LLM adaptation module and task-aware cross-modal feature fusion module are jointly optimized during the training process.", + "bbox": [ + 511, + 606, + 906, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 768, + 645, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe datasets, evaluation metrics, and implementation details. Next, we qualitatively and quantitatively compare our method with state-of-the-art methods. We further perform comprehensive experiments to analyze the main components of TCRT, including conditional normalization, object-text semantic matching regularization, and LLM-guided visual feature refinement.", + "bbox": [ + 511, + 794, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3935", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9d1d53b43963d8fcc810fd254d7ef6d37af83eb285af8f21280a2b7039363548.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsVisual BackboneMulti-taskVenueRefCOCORefCOCOgRefCOCO+ReferItGame test
valtestAtestBvaltestvaltestAtestB
Two-stage Methods
MAttNet [52]RN101×CVPR1876.6581.1469.9966.5867.2765.3371.6256.0229.04
CM-A-E [27]RN101×CVPR1978.3583.1471.3267.9968.6768.0973.6558.03-
Ref-NMS [6]RN101×AAAI2180.7084.0076.0470.5570.6268.2573.6859.42-
End-to-end Methods
MCN [28]DN53✓CVPR2080.0882.2974.9866.4666.0167.1672.8657.31-
RefTR [22]RN101✓NeurIPS2182.2385.5976.5769.4169.4071.5875.9662.1671.42
SeqTR [59]DN53✓ECCV2281.2385.0076.0871.3571.5868.8275.3758.7869.66
VG-LAW [38]ViT-B✓CVPR2386.6289.3283.1676.9076.9676.3781.0467.5077.22
PVD [7]Swin-B✓AAAI2484.9988.0280.0374.3474.6474.2779.0665.11-
Ferret-7B [50]ViT-L×ICLR2487.4991.3582.4583.9384.7680.7887.3873.14-
Vista-7B [34]ViT-L✓CVPR2488.191.583.086.086.484.190.375.8-
TCRTSwin-B✓-91.0792.8586.2486.8187.1385.7590.7677.5979.09
", + "bbox": [ + 161, + 88, + 836, + 281 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9007d0c37ac9c2bf075e0b9812a841e59dd7d35979026b5e9c9335e343be1f33.jpg", + "table_caption": [ + "Table 1. Comparison with existing visual grounding methods on RefCOCO, RefCOCOg, RefCOCO+, and ReferItGame for the REC task." + ], + "table_footnote": [], + "table_body": "
MethodsVisual BackboneMulti-taskVenuevalRefCOCO testAtestBRefCOCOgvalRefCOCO+ testAtestB
MCN [28]DN53✓CVPR2062.4464.2059.7149.2249.4050.6254.9944.69
RefTR [22]RN101✓NeurIPS2170.5673.4966.5758.7358.5161.0864.6552.73
SeqTR [59]DN53✓ECCV2267.2669.7964.1255.6755.6454.1458.9348.19
VG-LAW [38]ViT-B✓CVPR2375.6277.5172.8965.6366.0866.6370.3858.89
PVD [7]Swin-B✓AAAI2475.0777.2970.1363.2263.8964.3969.1557.19
Vista-7B [34]ViT-L✓CVPR2474.576.073.969.870.971.874.465.6
TCRTSwin-B✓-77.1278.1275.6770.8472.0872.7974.9666.53
", + "bbox": [ + 147, + 320, + 849, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with existing visual grounding methods on RefCOCO, RefCOCOg, and RefCOCO+ for the RES task.", + "bbox": [ + 130, + 441, + 862, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 482, + 295, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 506, + 194, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method, we carry out experiments on the RefCOCO [51], RefCOCOg [33], RefCOCO+ [51], and ReferItGame [20] datasets. RefCOCO, RefCOCOg, and RefCOCO+ datasets are derived from MS-COCO [23]. RefCOCO and RefCOCO+ datasets were partitioned into train, validation (val), testA, and testB subsets. RefCOCOg was partitioned into train, validation (val), and test subsets. In addition, the ReferItGame dataset, which was collected from SAIAPR-12 [13], comprises both train and test subsets.", + "bbox": [ + 89, + 525, + 483, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 686, + 269, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the RSC task, we employ Prec@0.5 as the metric for evaluation, which is consistent with prior works. If the IoU of the ground truth box with the predicted box surpasses 0.5, then the predicted box is deemed accurate. For the RES task, we evaluate performance by calculating MIoU.", + "bbox": [ + 89, + 705, + 483, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 790, + 302, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our visual backbone utilizes a Swin Transformer architecture pre-trained with Mask-RCNN on MSCOCO [23]. The input image size is set to $448 \\times 448$ . For the textual backbone, we adopt the T5 [35] model and limit the maximum length of text to 40. For the large language model, we use Llama3.1-8B [12], which is frozen during the training", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6188e1706f311249b9db962a43220c5e9af8df4e1af5512001fc699679ecd39b.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with representative methods on REC task. We also visualize the attention scores of each method. Blue and green boxes are predicted regions and the ground truth, respectively." + ], + "image_footnote": [], + "bbox": [ + 527, + 479, + 893, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "process. Our model undergoes end-to-end training for 70 epochs. The learning rate begins at $10^{-4}$ , and then we drop it down to $10^{-5}$ after 50 epochs. The optimizer used is AdamW. A dropout rate of 0.1 is applied. The batch size is configured to 16. We adopt the same data augmentation strategies as previous works [47, 48]. We implement our TCRT using PyTorch and perform the experiments on NVIDIA A800 GPUs.", + "bbox": [ + 511, + 780, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3936", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6358e372ced6bd6f61e4b7bd2eb9d4cac3a4a8cfa7a4e4c7f409f48bfc5bff7d.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 114, + 102, + 228, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a440a266aefd8f3e14a63d30003e4eb93bfe9860595e43f42ea83530fc5b7b31.jpg", + "image_caption": [ + "SeqTR" + ], + "image_footnote": [], + "bbox": [ + 230, + 102, + 344, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b69c4f3701904f548eab03417498a7420944447432e9ee5b00853df532275bb.jpg", + "image_caption": [ + "TCRT" + ], + "image_footnote": [], + "bbox": [ + 346, + 102, + 460, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c09b938cf065752ebab34e9edbba87df7fd16f1b1a83de0241066896a0c769e8.jpg", + "image_caption": [ + "\"fourth from bottom ri" + ], + "image_footnote": [], + "bbox": [ + 114, + 172, + 228, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3247815e9a75f295c49ff4d3a9f86ba2f67e515f178f0ae1fd26b4ea3bc7b87f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 172, + 344, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b75658858a563569c1ada2bf6d255c86c11d8ca9d1bf0d2007f3cce8112a4748.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 172, + 460, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46f741158de625298d7ea4f7002cc856c029f632428413bfc36f2435cce44bfd.jpg", + "image_caption": [ + "\"big couch at bottom of screen\"", + "\"giraffe appendage on right\"", + "Figure 5. Qualitative comparison with SeqTR on RES task." + ], + "image_footnote": [], + "bbox": [ + 114, + 244, + 228, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc203ab519cc9858ec1a133de9d9bd4d6f1339ddbf1253ebdfa08642abba7461.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 244, + 344, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab4df296ca3d6f0d26f985e3ca70a7babb978adfdfccc8ac0c3cb0db402369d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 244, + 460, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb4c6554ad13a1f7f18bb111117e4cfd1bacc8bca639b0232610f7c7a960aeee.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 125, + 369, + 233, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/537e606bc7a6af9eca68fb32645269722e0bab1228f18f9a5d0597fdf6175c2d.jpg", + "image_caption": [ + "TCRT w/o CN" + ], + "image_footnote": [], + "bbox": [ + 233, + 369, + 339, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8500d96e1912d6f5b999616ba8fbfbab8da4635b66ed0ad58c23835d56f2d249.jpg", + "image_caption": [ + "TCRT" + ], + "image_footnote": [], + "bbox": [ + 343, + 369, + 449, + 425 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba8020222ad4d5beb663cc58af574dae7d2b975f18f660d8c0e21de70accb3ed.jpg", + "image_caption": [ + "\"white laptop keyboard\"" + ], + "image_footnote": [], + "bbox": [ + 125, + 436, + 230, + 491 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47ac7ca98d42a6427ecf55ddb5143914c4ff5c50eb31d89023dee28c834bd1b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 436, + 339, + 491 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a50145551878df435be89db6073004bdf97a985133516c6339d31c453bf96b7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 436, + 447, + 491 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d99568173df8748042329cb6644c793bd27e2f82d9c738a370f0295c1b74339.jpg", + "image_caption": [ + "\"person half shown on far left\"", + "\"Guy lower right side facing away\"", + "Figure 6. Visualization of features extracted by the visual backbone. The second column shows the results of 'TCRT w/o CN', while the third column shows the results of TCRT. Red and white boxes are the predicted regions and the ground truth, respectively." + ], + "image_footnote": [], + "bbox": [ + 125, + 505, + 230, + 560 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14c1e7263b24a01e6c0bad7c318df54a295376faa762cb92ef0921029d9be390.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 505, + 339, + 560 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/430e80c12c761bb581589d32754d356fcd6060ba2418418c756570868c86fc58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 505, + 447, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-arts", + "text_level": 1, + "bbox": [ + 89, + 667, + 398, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of the proposed TCRT, we perform both qualitative and quantitative experiments across four datasets: RefCOCO, RefCOCOg, RefCOCO+, and ReferItGame.", + "bbox": [ + 89, + 689, + 483, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. REC task", + "text_level": 1, + "bbox": [ + 89, + 758, + 205, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our TCRT against existing state-of-the-art models for the REC task. As shown in Table 1, TCRT achieves the best performance across these datasets. Specifically, compared to Vista-7B, TCRT shows absolute improvements of $+2.97$ , $+1.35$ , and $+3.24$ percentage points on RefCOCO; $+1.65$ , $+0.46$ , and $+1.79$ percentage points on RefCOCO+; and $+0.81$ and $+0.73$ percentage points on RefCOCOg. We also present qualitative results from three", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bfd3dd46061ca25f22683c2161b057e7714c41c8b698cd261614785d77d42d12.jpg", + "image_caption": [ + "Figure 7. Comparison between the TCRT without object-text matching loss $(\\mathcal{L}_{\\mathrm{match}})$ and TCRT in terms of $\\mathcal{L}_{\\mathrm{det}}$" + ], + "image_footnote": [], + "bbox": [ + 547, + 90, + 875, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "examples of the RefCOCO testB set for the REC task in Figure 4. We can see that TCRT performs well even when an image contains multiple similar objects, such as \"second vase from front\". Furthermore, we visualized the attention scores within the transformer. We observe that TCRT focuses better on the target object compared to other methods.", + "bbox": [ + 511, + 315, + 906, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. RES task", + "text_level": 1, + "bbox": [ + 511, + 412, + 624, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of our TCRT for the RES task by comparing it with existing state-of-the-art methods. The comparative results are provided in Table 2. Compared with Vista-7B, TCRT achieves higher mIoU scores with absolute improvements of $+2.62$ , $+2.12$ , and $+1.77$ percentage points on RefCOCO, $+0.99$ , $+0.56$ , and $+0.93$ percentage points on RefCOCO+, $+1.04$ and $+1.18$ percentage points on RefCOCOg. We also present a comparison for the RES task in Figure 5.", + "bbox": [ + 511, + 431, + 906, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis of Main Components", + "text_level": 1, + "bbox": [ + 511, + 577, + 781, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1. Conditional Normalization", + "text_level": 1, + "bbox": [ + 511, + 599, + 746, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first conduct an experiment to demonstrate the effectiveness of conditional normalization (CN). We visualize the features extracted by the visual backbone on RefCOCOg test set. As shown in Figure 6, 'TCRT w/o CN' has more dispersed attention, which may lead to inaccurate predictions. Due to the smaller domain discrepancy of TCRT, this allows the LLM to focus more attention on the target objects.", + "bbox": [ + 511, + 617, + 905, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Object-text Semantic Matching Regularization", + "text_level": 1, + "bbox": [ + 511, + 746, + 883, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the object-text matching loss $(\\mathcal{L}_{\\mathrm{match}})$ to associate the visual tokens with the textual ones. To validate the effectiveness of $\\mathcal{L}_{\\mathrm{match}}$ we compare 'TCRT w/o $\\mathcal{L}_{\\mathrm{match}}$ and TCRT in terms of the detection loss $(\\mathcal{L}_{\\mathrm{det}})$ on RefCOCO val set. In Figure 7, by plotting the loss curves of the models during training, we observe that TCRT achieves lower loss values, which is attributed to its capability to learn semantic-discriminative visual features. This enables the LLM to more easily recognize task-relevant visual features.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3937", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd67a5d4df4a948bc5176529532f5e9d0e03741c989aa5466dcbd92db0569c33.jpg", + "image_caption": [ + "Figure 8. visualization of LLM-guided visual feature refinement. We visualize the weight matrices $\\mathcal{M}$ of each layer in the fusion module to validate the role of the LLM in refining visual features. Red and white boxes are the predicted regions and the ground truth, respectively." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 472, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/30fdeff73b8690ec6b7c831502cacf7057cc0e7a354a3efc9fd3a424e9f1311f.jpg", + "image_caption": [ + "Figure 9. Analysis of Visual Feature Distribution." + ], + "image_footnote": [], + "bbox": [ + 102, + 377, + 472, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3. LLM-guided Visual Feature Refinement", + "text_level": 1, + "bbox": [ + 89, + 627, + 416, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the weight matrix $\\mathcal{M}$ in Figure 8 to validate the effectiveness of visual feature refinement on RefCOCOg test set. We observe that the attention of the weight matrix is relatively evenly distributed in the earlier layers. In the later layers, the attention of the weight matrix is primarily focused on the target object. This indicates that LLM's prior knowledge enhances our model's ability to locate target objects by further refining the visual features.", + "bbox": [ + 89, + 646, + 483, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.4. Analysis of Visual Feature Distribution", + "text_level": 1, + "bbox": [ + 89, + 775, + 408, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness of our model in aligning visual and textual features, we extract the visual features after fusion with textual information and study their distribution in the spatial domain. Specifically, we label four target objects in an image and use three different textual descriptions for each object, resulting in twelve data samples. These samples are fed into the grounding model, and the visual", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/48e08a8a3f33d5fb4cce0c7701013b750afaad4ad66078c6e25adaeba5e88a7d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodvalRefCOCO
testAtestB
TCRT w/o VF & VA87.6288.6084.04
TCRT w/o VF88.3890.2485.57
TCRT91.0792.8586.24
", + "bbox": [ + 566, + 88, + 857, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation experiments on RefCOCO to evaluate the proposed LLM adaptation module $\\left( {V}_{A}\\right)$ and task-aware cross-modal feature fusion module $\\left( {V}_{F}\\right)$ .", + "bbox": [ + 511, + 181, + 906, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "features from the fusion module's output are subjected to dimensionality reduction and visualization. As shown in Figure 9, we observe that our method causes the visual features describing the same object to cluster together.", + "bbox": [ + 511, + 253, + 906, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 325, + 674, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We validate the design of our proposed TCRT by conducting ablative experiments on major components in the REC task. The results are presented in Table 3. We observe that when the LLM adaptation module and task-aware cross-modal feature fusion module are not used, 'TCRT w/o $V_{F}$ & $V_{A}$ ' achieves accuracy rates of $87.62\\%$ , $88.60\\%$ , and $84.04\\%$ on the RefCOCO dataset. After incorporating the LLM adaptation module, the accuracy of 'TCRT w/o $V_{F}$ ' increases by 0.76, 1.64, and 1.53 absolute percentage points. When the $V_{F}$ module is further introduced, TCRT achieves accuracy rates of $91.07\\%$ , $92.85\\%$ , and $86.24\\%$ .", + "bbox": [ + 511, + 344, + 906, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 527, + 633, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a task-aware cross-modal feature refinement transformer with large language models to facilitate visual grounding. To adapt LLMs to the grounding task, we design an adaptation module that modulates the statistical characteristics of visual features conditioned on textual features along with object-text semantic matching regularization. An LLM is employed to integrate the aligned information from textual and visual features and produce task-aware embeddings. We further incorporate a task-aware cross-modal feature fusion module, allowing the resulting embeddings to refine visual features and enable information interaction between the REC and RES tasks. Comprehensive experimental results verify the effectiveness of our proposed approach both quantitatively and qualitatively.", + "bbox": [ + 511, + 555, + 906, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 782, + 691, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by TCL Science and Technology Innovation Fund (Project No. 20231752), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), and in part by the GuangDong Basic and Applied Basic Research Foundation (Project No. 2024A1515011437).", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3938", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023.", + "[2] Mohit Bajaj, Lanjun Wang, and Leonid Sigal. G3raphground: Graph-based language grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4281-4290, 2019. 2", + "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3", + "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2, 3", + "[5] Kan Chen, Rama Kovvuri, and Ram Nevatia. Query-guided regression network with context policy for phrase grounding. In Proceedings of the IEEE International Conference on Computer Vision, pages 824-832, 2017. 2", + "[6] Long Chen, Wenbo Ma, Jun Xiao, Hanwang Zhang, and Shih-Fu Chang. Ref-nms: Breaking proposal bottlenecks in two-stage referring expression grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1036-1044, 2021. 2, 6", + "[7] Zesen Cheng, Kehan Li, Peng Jin, Siheng Li, Xiangyang Ji, Li Yuan, Chang Liu, and Jie Chen. Parallel vertex diffusion for unified visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1326-1334, 2024. 2, 6", + "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 3", + "[9] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1769-1779, 2021. 1, 2", + "[10] Jiajun Deng, Zhengyuan Yang, Daqing Liu, Tianlang Chen, Wengang Zhou, Yanyong Zhang, Houqiang Li, and Wanli Ouyang. Transvg++: End-to-end visual grounding with language conditioned vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2", + "[11] Pelin Dogan, Leonid Sigal, and Markus Gross. Neural sequential phrase grounding (seqground). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4175-4184, 2019. 2", + "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 6" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Hugo Jair Escalante, Carlos A Hernández, Jesus A Gonzalez, Aurelio López-López, Manuel Montes, Eduardo F Morales, L Enrique Sucar, Luis Villasenor, and Michael Grubinger. The segmented and annotated iapr tc-12 benchmark. Computer Vision and Image Understanding, 114(4): 419-428, 2010. 6", + "[14] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3", + "[15] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 1440-1448, 2015. 2, 5", + "[16] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 580-587, 2014. 2", + "[17] Chih-Hui Ho, Srikar Appalaraju, Bhavan Jasani, R Manmatha, and Nuno Vasconcelos. Yoro-lightweight end to end visual grounding. In European Conference on Computer Vision, pages 3-23. Springer, 2022. 1", + "[18] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1115-1124, 2017. 2", + "[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 3", + "[20] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 787-798, 2014. 6", + "[21] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv e-prints, pages arXiv-2211, 2022. 3", + "[22] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in Neural Information Processing Systems, 34:19652-19664, 2021. 1, 2, 6", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 5", + "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee." + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3939", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1, 2", + "[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2", + "[27] Xihui Liu, Zihao Wang, Jing Shao, Xiaogang Wang, and Hongsheng Li. Improving referring expression grounding with cross-modal attention-guided erasing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1950-1959, 2019. 2, 6", + "[28] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10034-10043, 2020. 6", + "[29] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. In European Conference on Computer Vision, pages 417-435. Springer, 2025. 2", + "[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11-20, 2016. 2", + "[31] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 Fourth International Conference on 3D Vision (3DV), pages 565-571. IEEE, 2016. 5", + "[32] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 2", + "[33] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6", + "[34] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. 2, 6", + "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 6", + "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 1, 2", + "[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 5", + "[38] Wei Su, Peihan Miao, Huanzhang Dou, Gaoang Wang, Liang Qiao, Zheyang Li, and Xi Li. Language adaptive weight generation for multi-task visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10857-10866, 2023. 1, 2, 6", + "[39] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities, 2023.", + "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3", + "[43] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023.", + "[44] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4145-4154, 2019. 2", + "[45] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 1", + "[46] Sibei Yang, Guanbin Li, and Yizhou Yu. Graph-structured referring expression reasoning in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9952–9961, 2020. 2", + "[47] Zhengyuan Yang, Boqing Gong, Liwei Wang, Wenbing Huang, Dong Yu, and Jiebo Luo. A fast and accurate one-stage approach to visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4683-4693, 2019. 2, 6", + "[48] Zhengyuan Yang, Tianlang Chen, Liwei Wang, and Jiebo Luo. Improving one-stage visual grounding by recursive subquery construction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 387-404. Springer, 2020. 2, 6", + "[49] Jiabo Ye, Junfeng Tian, Ming Yan, Xiaoshan Yang, Xuwu Wang, Ji Zhang, Liang He, and Xin Lin. Shifting more" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3940", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15502-15512, 2022. 1, 2", + "[50] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 6", + "[51] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6", + "[52] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 2, 6", + "[53] Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. In European Conference on Computer Vision, pages 405-422. Springer, 2025. 2, 3", + "[54] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An lmm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 3", + "[55] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4158-4166, 2018. 2", + "[56] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024.", + "[57] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3", + "[58] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. Advances in Neural Information Processing Systems, 37:71737-71767, 2024. 2", + "[59] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision, pages 598-615. Springer, 2022. 1, 6" + ], + "bbox": [ + 91, + 90, + 485, + 810 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3941", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_model.json b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..581014e4cd5d8f393d11f4caf50be01e2c9fafcf --- /dev/null +++ b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_model.json @@ -0,0 +1,2829 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.905, + 0.177 + ], + "angle": 0, + "content": "Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.203, + 0.762, + 0.221 + ], + "angle": 0, + "content": "Wenbo Chen1*; Zhen Xu1,2*; Ruotao Xu2, Si Wu1,2†; Hau-San Wong3" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.222, + 0.838, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, South China University of Technology \\(^{2}\\)Institute for Super Robotics (Huangpu)" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.257, + 0.76, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Computer Science, City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.277, + 0.879, + 0.293 + ], + "angle": 0, + "content": "{cscwb106, csxuzhen}@mail.scut.edu.cn, {xrt, cswusi}@scut.edu.cn, cshswong@cityu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.721 + ], + "angle": 0, + "content": "The goal of visual grounding is to establish connections between target objects and textual descriptions. Large Language Models (LLMs) have demonstrated strong comprehension abilities across a variety of visual tasks. To establish precise associations between the text and the corresponding visual region, we propose a Task-aware Cross-modal feature Refinement Transformer with LLMs for visual grounding, and our model is referred to as TCRT. To enable the LLM trained solely on text to understand images, we introduce an LLM adaptation module that extracts text-related visual features to bridge the domain discrepancy between the textual and visual modalities. We feed the text and visual features into the LLM to obtain task-aware priors. To enable the priors to guide the feature fusion process, we further incorporate a cross-modal feature fusion module, which allows task-aware embeddings to refine visual features and facilitate information interaction between the Referring Expression Comprehension (REC) and Referring Expression Segmentation (RES) tasks. We have performed extensive experiments to verify the effectiveness of the main components and demonstrate the superior performance of the proposed TCRT over state-of-the-art end-to-end visual grounding methods on RefCOCO, RefCOCOg, RefCOCO+ and ReferItGame." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Visual grounding aims to predict the locations of the objects specified by textual prompts in images, and is widely applied in the field of human-computer interaction. This task typically involves referring expression comprehension [10, 17, 49, 59] and referring expression segmentation [22, 38]. The main challenge is to comprehensively un" + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.327, + 0.648, + 0.335 + ], + "angle": 0, + "content": "TCRT w/o LLM" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.336, + 0.617, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.391, + 0.617, + 0.395 + ], + "angle": 0, + "content": "\"bride throwing bouquet" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.395, + 0.617, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.449, + 0.625, + 0.457 + ], + "angle": 0, + "content": "\"Guy front in gray suit left\"" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.457, + 0.617, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.512, + 0.722, + 0.52 + ], + "angle": 0, + "content": "\"man in right front with blue tie next to older woman" + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.327, + 0.819, + 0.334 + ], + "angle": 0, + "content": "TCRT" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.336, + 0.805, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.397, + 0.805, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.457, + 0.805, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.513, + 0.805, + 0.521 + ], + "angle": 0, + "content": "predicted by" + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.336, + 0.898, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.397, + 0.898, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.457, + 0.898, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.82, + 0.513, + 0.898, + 0.521 + ], + "angle": 0, + "content": "ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.533, + 0.907, + 0.588 + ], + "angle": 0, + "content": "Figure 1. The visualization of the refined visual features from the cross-modal feature fusion module. In challenging environments, the introduction of LLMs can facilitate a better association between visual features and text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.656 + ], + "angle": 0, + "content": "derstand intra-modality information and accurately establish inter-modality correspondences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Existing visual grounding methods can be categorized into two main types: two-stage and end-to-end methods. The two-stage methods [45] first generate region proposals [36] that may match the text, and then rank the candidate regions according to their relevance to the text. When the proposals generated in the first stage do not contain the target object, the second stage will be unable to correct these proposals. On the other hand, end-to-end grounding methods perform the integration of visual and textual features within an object detector, producing the box coordinates of the target object. The end-to-end methods are typically faster than the two-stage ones. With the advancement of transformer in the vision domain, TransVG [9, 10] and SeqTR [59] employ self-attention or cross-attention mechanisms to fuse textual and visual features. Recently, LLaVA [25] has integrated visual understanding capabilities into large language mod" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.214, + 0.887 + ], + "angle": 0, + "content": "*Joint first authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3931" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "els (LLMs). MiniGPT-2 [4] and OMG-LLaVA [58] apply LLMs to visual grounding tasks by integrating textual and visual features. In this work, we make an attempt to adapt LLMs to capture the information of target objects and produce task-aware embeddings for precise grounding results." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.17, + 0.483, + 0.485 + ], + "angle": 0, + "content": "More specifically, we propose a task-aware cross-modal feature refinement transformer with large language models for visual grounding, which is referred to as TCRT, leveraging priors from LLMs to guide the multimodal information fusion process and achieve precise alignment between textual and local visual features (see Figure 1). To enable the LLM to more effectively process image data and adapt to visual grounding tasks, we propose an LLM adaptation module that utilizes conditional normalization to adjust the statistical characteristics of visual features based on textual information, while introducing object-text semantic matching regularization to associate local visual features with textual features. Additionally, to utilize the priors provided by the adaptation module, we incorporate a task-aware cross-modal feature fusion module, which continuously refines visual features during multimodal fusion and enables the model to exploit the relationship between the REC and RES tasks. We perform extensive qualitative and quantitative experiments to validate the effectiveness of our model, and demonstrate its advantages on benchmark datasets. The main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.489, + 0.48, + 0.564 + ], + "angle": 0, + "content": "- Different from the state-of-the-art methods that focus on the application of LLMs in integrating textual and visual features, we further explore the LLM's capability of learning task-aware embeddings to facilitate visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.565, + 0.48, + 0.623 + ], + "angle": 0, + "content": "- We incorporate an LLM adaptation module for visual grounding, which learns to capture the information of target objects by modulating the statistical characteristics of visual features conditioned on textual features." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.625, + 0.48, + 0.685 + ], + "angle": 0, + "content": "- We further perform task-aware cross-modal feature fusion, allowing the task-aware embeddings derived from LLMs to guide the association of textual and visual information and lead to precise grounding results." + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.489, + 0.48, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.702, + 0.233, + 0.717 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.347, + 0.743 + ], + "angle": 0, + "content": "2.1. Two-stage Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In two-stage methods, a region proposal network [15, 16, 36] is first used to produce a set of candidate regions that may correspond to the target object mentioned in the textual description. These candidate regions are then matched to the textual description, and the predicted result is determined by identifying the region with the highest score. The maximum-margin ranking loss [30] is widely adopted for the second stage. Several researches further enhance the two-stage methods by incorporating context modeling [18, 32, 44, 55] and making use of phrase cooccur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "ences [2, 5, 11]. MAttNet [52] decomposed the expression into three modules: relationship, location, and subject to improve the grounding accuracy. To enhance the model's capacity for learning detailed alignments between text and visuals, CM-A-E [27] proposed a method capable of generating challenging samples for training the model. Graph learning was incorporated into visual grounding by SGMN [46] to enable the model to better capture the relationships between different objects within the same scene. To increase the recall of target objects, Ref-NMS [6] utilized textual features to direct the non-maximum suppression. To enhance the model's cross-modality fusion capabilities, GroundingDINO [26] integrated grounded pretraining techniques into detection transformers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.315, + 0.78, + 0.331 + ], + "angle": 0, + "content": "2.2. End-to-end Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.905, + 0.639 + ], + "angle": 0, + "content": "End-to-end methods in visual grounding aim to directly predict the bounding boxes in an image from a text in a single pass, resulting in faster inference speeds compared to two-stage methods. As a pioneering work, FAOA [47] concatenated image and textual features, and fed the concatenated features into a prediction layer to achieve target object detection. To reduce the referring ambiguity, ReSC [48] used a sub-query construction to address limitations in processing complex textual descriptions. TransVG [9, 10] is one of the pioneering transformer-based visual grounding models that directly regresses coordinates. To achieve multi-task learning, RefTR [22] used the transformer architecture to integrate REC and RES tasks into a single framework. To address the gap between the visual and textual backbones, QRNet [49] incorporated textual information into the visual backbone. Furthermore, VG-LAW [38] actively extracted text-relevant visual features based on text-adaptive weights and performed multi-task visual grounding. Additionally, PVD [7] introduced a diffusion model into the visual grounding task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.905, + 0.837 + ], + "angle": 0, + "content": "With the development of LLMs, some Multimodal Large Language Models [25] (MLLMs) are gradually being applied to visual grounding tasks. VistaLLM [34] used an instruction-guided image tokenizer to filter and compress image embeddings based on task, extracting visual information relevant to the current task. To efficiently locate and understand multiple target regions in complex scenes, Groma [29] introduced a localized visual tokenization mechanism that decomposes an image into multiple regions of interest, encoding each region as an independent region token. To unleash the potential of large language models, Griffon [53] introduces a foundational language-prompted localization dataset with nearly 6 million pretraining samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.738, + 0.864 + ], + "angle": 0, + "content": "2.3. Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "With the expansion of training data and a significant increase in model parameters, large language models (LLMs)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3932" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.904, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.278, + 0.907, + 0.39 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed method. We first extract textual features \\( F_{txt} \\) using the textual backbone \\( E_{txt} \\). These features are then injected into the visual backbone \\( E_{vis} \\), which consists of transformer blocks and conditional normalization, to extract text-related visual features \\( F_{vis} \\). The object-text semantic matching regularization is employed to associate visual region tokens with textual tokens. We input the visual features processed by the visual adapter \\( V_{proj} \\), along with the text, into LLaMa. We take the last token \\( f_{llm} \\) output from LLaMa. \\( f_{llm} \\) is then fed into a query projection layer to generate task-relevant tokens \\( f_{box} \\), \\( f_{seg} \\), and \\( f_{phr} \\), which are used to predict the bounding box, segmentation, and phrase, respectively. A task-aware cross-modal feature fusion module is then employed to integrate textual, visual, and task features. Finally, task-relevant tokens are each fed to the corresponding prediction heads to produce the final predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.482, + 0.581 + ], + "angle": 0, + "content": "based on the Transformer [42] architecture have developed rapidly. GPT-3 [3], with its 175 billion parameters, demonstrates strong performance in few-shot contextual learning. Following the success of GPT-3, several other LLMs, including PaLM [8], OPT [57], BLOOM [21], and LLaMA [40, 41], have emerged in succession. Mistral [19] introduced a specialized window attention mechanism to better handle extended context modeling. In parallel, Mixtral \\(8 \\times 7\\)B utilized sparse MoE layers [14] to scale up parameters more effectively, enabling it to achieve higher performance with a reduced number of active parameters." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.734 + ], + "angle": 0, + "content": "The main differences between the proposed TCRT and existing visual grounding methods are summarized as follows: First, we perform statistical modulation on visual features conditioned on textural features to reduce domain discrepancy before integrating them, while the existing methods [4, 53, 54] only perform dimension adaption. Second, TCRT leverage the task-aware embeddings derived from LLMs to guide cross-modal feature fusion and association analysis, rather than directly predicting the bounding boxes of target objects as in previous methods [4, 53, 54]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.226, + 0.765 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.251, + 0.789 + ], + "angle": 0, + "content": "3.1. Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we begin by defining multi-task visual grounding (MTVG), and then proceed to describe the architecture of the proposed framework. The goal of MTVG is to detect or segment a target object based on a textual description. Formally, given an image \\( I \\in \\mathbb{R}^{3 \\times H \\times W} \\) and a corresponding text \\( q \\), this task is to predict a bounding box \\( b = \\{x, y, w, h\\} \\) that encompasses the object men" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.907, + 0.492 + ], + "angle": 0, + "content": "tioned in the text, as well as a pixel-wise segmentation mask \\(s\\in \\mathbb{R}^{1\\times H\\times W}\\) that provides a more detailed and precise localization. A textual backbone \\(E_{txt}\\) and a visual backbone \\(E_{vis}\\) are used to extract textual and visual features, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.683, + 0.518 + ], + "angle": 0, + "content": "3.2. LLM Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.766 + ], + "angle": 0, + "content": "Pre-trained LLMs possess language understanding capabilities and can handle many complex tasks in natural language processing. However, due to the domain discrepancy between textual data and visual data, directly applying LLMs to grounding tasks may yield suboptimal performance. Therefore, we propose an LLM adaptation module to perform cross-modal adaptation of image information, enabling LLMs to better handle information within visual data. As shown in Figure 2, we first utilize the text encoding model \\( E_{txt} \\) to extract the textual features of \\( q \\): \\( F_{txt} = E_{txt}(q) \\), where \\( F_{txt} \\) represents the textual features, which include three types of textual features: sentence feature \\( f_{txt}^{sen} \\in \\mathbb{R}^{1 \\times d} \\), individual word features \\( f_{txt} \\in \\mathbb{R}^{L \\times d} \\), and average pooled word feature \\( f_{txt}^{mean} \\in \\mathbb{R}^{1 \\times d} \\), where \\( d \\) and \\( L \\) represent the dimension of textual features and the number of tokens, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.906, + 0.873 + ], + "angle": 0, + "content": "To enable the visual encoder to comprehend textual information, we designed a text-aware visual encoder \\( E_{vis} \\). Specifically, a conditioned normalization is inserted after each transformer block in the visual encoder. In the conditioned normalization, we fuse the sentence features \\( f_{txt}^{sen} \\) and the average-pooled word features \\( f_{txt}^{mean} \\) into a textual conditional feature \\( f_{txt}^{cond} \\) via a weighted combination:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.884, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\nf _ {t x t} ^ {\\text {c o n d}} = \\alpha^ {\\text {s e n}} f _ {t x t} ^ {\\text {s e n}} + \\alpha^ {\\text {m e a n}} \\phi \\left(f _ {t x t} ^ {\\text {m e a n}}\\right), \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3933" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.087, + 0.482, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.231, + 0.485, + 0.304 + ], + "angle": 0, + "content": "Figure 3. Object-text semantic matching regularization. We constrain the image features using a mask to isolate the object-specific features. For matching text-object pairs, we consider them as positive samples, whereas non-matching pairs are regarded as negative samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.481 + ], + "angle": 0, + "content": "where \\(\\alpha^{sen}\\) and \\(\\alpha^{mean}\\) are learnable parameters that control the contribution of \\(f_{txt}^{sen}\\) and \\(f_{txt}^{mean}\\), respectively. \\(\\phi(\\cdot)\\) represents the feed-forward network. The scaling \\(\\gamma\\) and shifting \\(\\beta\\) parameters are determined using \\(f_{txt}^{cond}:\\gamma = \\phi(f_{txt}^{cond}) + \\gamma^{learn}\\), \\(\\beta = \\phi(f_{txt}^{cond}) + \\beta^{learn}\\), where \\(\\gamma^{learn}\\) and \\(\\beta^{learn}\\) are learnable parameters. Given the visual features \\(\\bar{F}_i^*\\), which is the output of the \\(i\\)-th layer of transformer block, we calculate the mean \\(\\mu(\\bar{F}_i^*)\\) and standard deviation \\(\\sigma(\\bar{F}_i^*)\\). The visual features \\(\\bar{F}_i^*\\) are then modulated using the resulting scaling \\(\\gamma\\) and shifting \\(\\beta\\) parameters, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.489, + 0.483, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\bar {F} _ {i} ^ {*} = \\gamma \\left(\\frac {\\bar {F} _ {i} ^ {*} - \\mu (\\bar {F} _ {i} ^ {*})}{\\sigma (\\bar {F} _ {i} ^ {*}) + \\epsilon}\\right) + \\beta , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.707 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a small constant introduced to avoid division by zero. We obtain the text-aware visual features \\(F_{vis}\\) by flattening the feature map \\(\\overline{F}_4^*\\) extracted from the visual backbone. The image features and the input embeddings of LLMs are typically in different dimensional spaces. We use a visual adapter \\(V_{proj}\\) to map the image features to a dimension that matches the LLM input, allowing the LLM to directly process visual information. \\(V_{proj}\\) is a three-layer MLP. We feed the text and the mapped image features into the LLM, and take the last token of the LLM output as the prior embedding \\(f_{llm}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.716, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\nf _ {l l m} = \\Phi (q, V _ {p r o j} \\underbrace {\\left(E _ {v i s} (I , f _ {t x t} ^ {c o n d})\\right)} _ {F _ {v i s}}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\(\\Phi\\) represents an LLM model, \\(q\\) denotes the original query text. We pass \\(f_{llm}\\) through three task-aware linear layers (\\(W_{box}\\), \\(W_{seg}\\), and \\(W_{phr}\\)) to obtain the task-specific embeddings \\(f_{box}\\), \\(f_{seg}\\), and \\(f_{phr}\\), which are responsible for predicting the bounding box, segmentation, and phrase, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The text usually describes a specific object within an image, which occupies only a portion of the image. Most visual features do not have a direct relationship with the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.348 + ], + "angle": 0, + "content": "text. This makes it challenging to achieve precise semantic alignment between the target object and textual tokens. Thus, we propose the object-text semantic matching regularization to associate the visual region tokens with the textual ones. Object-text matching is a binary classification task, and the model evaluates the multimodal features of an object-text pair to classify it as either positive (matched) or negative (unmatched). As shown in Figure 3, if the textual features and visual features represent the same object, they are considered a positive sample pair; otherwise, they are considered a negative sample pair. The ground truth of object segmentation is represented by \\( s^+ \\). We apply an affine transformation to \\( s^+ \\) to obtain \\( s^- \\). \\( s^- \\) represents the inaccurate segmentation result. We leverage the segmentation mask of the target object to extract its features. A learnable token \\( f_{match} \\) is then introduced to determine whether they are matched:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.356, + 0.905, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\{\\bar {f} _ {\\text {m a t c h}}, \\bar {F} _ {l}, \\bar {F} _ {v} \\right\\} = \\varphi \\left(\\left[ f _ {\\text {m a t c h}}, F _ {\\text {t x t}}, F _ {\\text {v i s}} \\right]\\right), \\\\ \\left\\{\\hat {f} _ {\\text {m a t c h}} ^ {j}, \\hat {F} _ {l} \\right\\} = \\theta \\left(\\left[ \\bar {f} _ {\\text {m a t c h}}, \\bar {F} _ {l} \\right], F _ {\\text {v i s}} \\odot s ^ {j}\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.537 + ], + "angle": 0, + "content": "where \\( j \\in \\{+, -\\} \\). \\( \\varphi(q) \\) and \\( \\theta(q, kv) \\) represent multi-head self-attention and cross-attention, respectively. \\( \\odot \\) denotes element-wise multiplication. \\( [\\cdot] \\) denotes the concatenation operation. We use the \\( \\hat{f}_{match}^{j} \\) as the joint representation of the object-text pair. A feed-forward network is employed to predict the two-class probability \\( p_{match}^{j} \\) that indicates whether a textual feature matches a local visual feature. The object-text matching loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.549, + 0.907, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a t c h}} = \\mathbb {E} _ {(I, q)} [ \\sum_ {j} H \\left(y _ {\\text {m a t c h}} ^ {j}, p _ {\\text {m a t c h}} ^ {j}\\right) ], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.749 + ], + "angle": 0, + "content": "where \\( H \\) represents cross-entropy and \\( y_{match}^{j} \\) is a 2-dimensional one-hot vector that represents the ground truth label. We use the visual features of the target object extracted from the ground truth mask as positive sample, indicating that these features match the textual description. When performing affine transformation (partial overlap may occur), the resulting mask deviates from the target object, and the extracted features are considered as negative sample. Object-text matching loss further associates visual tokens with textual ones by assessing whether they match." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.862, + 0.773 + ], + "angle": 0, + "content": "3.3. Task-aware Cross-modal Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "When the LLM receives textual and visual information, it can generate more meaningful semantic representations. These representations contain the LLM's understanding of objects and context, forming prior information that aids downstream tasks. The task-aware cross-modal feature fusion module, denoted as \\( V_{F} \\), is incorporated to guide feature the fusion process. Specifically, we first employ a bidirectional multi-head attention module to fuse textual and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3934" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.272 + ], + "angle": 0, + "content": "visual features. Each modality serves as the query while the other modality acts as the key and value in the attention computation. We concatenate the three tokens used for prediction, \\( f_{box} \\), \\( f_{seg} \\), and \\( f_{phr} \\), with the textual features \\( F_{txt} \\) to obtain a task-relevant textual feature \\( F_{task} \\). Then, \\( F_{task} \\) and \\( F_{vis} \\) are fed into the bi-directional multi-head attention module to fuse the two types of features. This process can be represented as follows: \\( F_{task}^{v2t} = \\theta (F_{task}, F_{vis}) \\), \\( F_{vis}^{t2v} = \\theta (F_{vis}, F_{task}) \\), where \\( F_{task}^{v2t} \\) represents the textual feature enriched with image features, comprising \\( F_{txt}^{v2t} \\), \\( f_{box}^{v2t} \\), \\( f_{seg}^{v2t} \\), and \\( f_{phr}^{v2t} \\). Meanwhile, \\( F_{vis}^{t2v} \\) denotes the visual feature enriched with textual information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.484, + 0.395 + ], + "angle": 0, + "content": "Target objects are typically located in specific regions. To enable the model to focus on local visual features, we input \\( f_{llm} \\) into an MLP to obtain the refinement prompt \\( f_{prmt} \\), which guides the feature refinement process. We then assign different weights to image features at different locations by computing the cosine similarity between the image feature \\( F_{vis}^{t2v} \\) and \\( f_{prmt} \\). This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.405, + 0.484, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {v} ^ {t 2 v} = F _ {v i s} ^ {t 2 v} + F _ {v i s} ^ {t 2 v} \\odot \\underbrace {\\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {v i s} ^ {t 2 v}\\right) \\otimes \\mathrm {M L P} \\left(f _ {l l m}\\right)\\right)} _ {\\mathcal {M}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.612 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes matrix multiplication and \\(\\odot\\) represents element-wise multiplication. The refinement prompt \\(f_{prmt}\\) is dot-multiplied with the image features at each position in \\(F_{vis}^{t2v}\\) to obtain a weight matrix \\(\\mathcal{M}\\). The final refined visual features \\(\\hat{F}_v^{t2v}\\) are obtained under the guidance of the LLM. The model can control the strength of the feature representations at different positions in the image by adjusting the weight matrix \\(\\mathcal{M}\\). By assigning lower weights to irrelevant regions, the model can better focus on the features of the target regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.752 + ], + "angle": 0, + "content": "There is a close relationship between the REC task and the RES task. Both require an understanding and localization of the target object or region. REC tasks generally offer locational information of the target objects, whereas RES tasks offer precise boundary details. To strengthen the relationship between the REC and RES tasks, we perform cross-task interaction to effectively exchange information between the box token \\( f_{\\mathrm{box}}^{v2t} \\) and the segmentation token \\( f_{\\mathrm{seg}}^{v2t} \\). The process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.765, + 0.483, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\left\\{f _ {b o x} ^ {\\prime}, f _ {s e g} ^ {\\prime} \\right\\} = \\theta ([ f _ {b o x} ^ {v 2 t}, f _ {s e g} ^ {v 2 t} ], F _ {t x t} ^ {v 2 t}), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.794, + 0.441, + 0.814 + ], + "angle": 0, + "content": "\\[\nf _ {b o x} ^ {\\prime \\prime} = f _ {b o x} ^ {\\prime} + \\phi (f _ {s e g} ^ {\\prime}), f _ {s e g} ^ {\\prime \\prime} = f _ {s e g} ^ {\\prime} + \\phi (f _ {b o x} ^ {\\prime}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This design allows the model to exploit the correlation between the REC and RES tasks. The updated \\( f_{box} \\), \\( f_{seg} \\), and \\( f_{phrase} \\) are each fed into their respective prediction heads to obtain the final bounding box, segmentation, and phrase predictions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.673, + 0.108 + ], + "angle": 0, + "content": "3.4. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.189 + ], + "angle": 0, + "content": "Detection loss. For the REC task, the loss for bounding box regression is calculated using a combination of GIoU loss and L1 loss. We use \\( b \\) to denote the ground truth bounding box and \\( \\tilde{b} \\) to represent the predicted bounding box. The following defines the detection loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.195, + 0.907, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\det } = \\mathbb {E} _ {(I, q)} \\left[ \\Phi_ {\\text {s m o o t h} - l 1} (b, \\tilde {b}) + \\Phi_ {\\text {g i o u}} (b, \\tilde {b}) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.221, + 0.906, + 0.251 + ], + "angle": 0, + "content": "where \\(\\Phi_{\\mathrm{smooth} - l1}\\) and \\(\\Phi_{\\mathrm{giou}}\\) denote the smooth L1 loss [15] and Generalized IoU loss [37], respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.251, + 0.906, + 0.312 + ], + "angle": 0, + "content": "Segmentation loss. For the RES task, the segmentation loss is computed by combining the focal loss with dice loss. The ground truth mask is denoted as \\( s \\), while the predicted mask is \\( \\hat{s} \\). The following defines the segmentation loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.319, + 0.907, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s e g}} = \\mathbb {E} _ {(I, q)} \\left[ \\Phi_ {\\text {f o c a l}} (s, \\hat {s}) + \\Phi_ {\\text {d i c e}} (s, \\hat {s}) \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.343, + 0.906, + 0.373 + ], + "angle": 0, + "content": "where \\(\\Phi_{\\mathrm{focal}}\\) and \\(\\Phi_{\\mathrm{dice}}\\) denote the smooth focal loss [24] and DICE/F-1 loss [31], respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.374, + 0.906, + 0.449 + ], + "angle": 0, + "content": "Phrase loss. To enable the model to more accurately capture the entities within the text, we utilize \\( f_{phr} \\) to predict the words in the text related to the target object. We fed \\( f_{phr} \\) into an MLP to predict a word probability \\( p_{phr} \\). Then, we compute the phrase loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.457, + 0.907, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p h r}} = \\mathbb {E} _ {(I, q)} [ H (y _ {\\text {p h r}}, p _ {\\text {p h r}}) ], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.481, + 0.796, + 0.496 + ], + "angle": 0, + "content": "where \\(y_{phr}\\) denotes the ground truth label." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.497, + 0.906, + 0.54 + ], + "angle": 0, + "content": "By incorporating these training loss functions, the optimization problem of our model can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.55, + 0.906, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\min _ {E _ {t x t}, E _ {v i s}} \\lambda_ {\\text {m a t c h}} \\mathcal {L} _ {\\text {m a t c h}}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.576, + 0.868, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\min _ {E _ {t x t}, E _ {v i s}, V _ {p r o j}, W _ {t a s k}, V _ {F}} \\mathcal {L} _ {\\mathrm {d e t}} + \\mathcal {L} _ {\\mathrm {s e g}} + \\lambda_ {\\mathrm {p h r}} \\mathcal {L} _ {\\mathrm {p h r}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.607, + 0.907, + 0.76 + ], + "angle": 0, + "content": "where \\( \\text{task} \\in \\{box,seg,phr\\} \\), and \\( \\lambda_{\\mathrm{match}} \\) and \\( \\lambda_{\\mathrm{phr}} \\) are used to achieve a balance among the training objectives. \\( \\mathcal{L}_{\\mathrm{det}} \\) and \\( \\mathcal{L}_{\\mathrm{seg}} \\) enable the model to perform basic detection and segmentation tasks. \\( \\mathcal{L}_{\\mathrm{phr}} \\) enables the model to achieve a more detailed understanding of the target objects referred to in the text. \\( \\mathcal{L}_{\\mathrm{match}} \\) is used to associate visual tokens with textual ones and only optimizes the textual backbone and visual backbone. Our LLM adaptation module and task-aware cross-modal feature fusion module are jointly optimized during the training process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.646, + 0.788 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In this section, we first describe datasets, evaluation metrics, and implementation details. Next, we qualitatively and quantitatively compare our method with state-of-the-art methods. We further perform comprehensive experiments to analyze the main components of TCRT, including conditional normalization, object-text semantic matching regularization, and LLM-guided visual feature refinement." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3935" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.837, + 0.282 + ], + "angle": 0, + "content": "
MethodsVisual BackboneMulti-taskVenueRefCOCORefCOCOgRefCOCO+ReferItGame test
valtestAtestBvaltestvaltestAtestB
Two-stage Methods
MAttNet [52]RN101×CVPR1876.6581.1469.9966.5867.2765.3371.6256.0229.04
CM-A-E [27]RN101×CVPR1978.3583.1471.3267.9968.6768.0973.6558.03-
Ref-NMS [6]RN101×AAAI2180.7084.0076.0470.5570.6268.2573.6859.42-
End-to-end Methods
MCN [28]DN53✓CVPR2080.0882.2974.9866.4666.0167.1672.8657.31-
RefTR [22]RN101✓NeurIPS2182.2385.5976.5769.4169.4071.5875.9662.1671.42
SeqTR [59]DN53✓ECCV2281.2385.0076.0871.3571.5868.8275.3758.7869.66
VG-LAW [38]ViT-B✓CVPR2386.6289.3283.1676.9076.9676.3781.0467.5077.22
PVD [7]Swin-B✓AAAI2484.9988.0280.0374.3474.6474.2779.0665.11-
Ferret-7B [50]ViT-L×ICLR2487.4991.3582.4583.9384.7680.7887.3873.14-
Vista-7B [34]ViT-L✓CVPR2488.191.583.086.086.484.190.375.8-
TCRTSwin-B✓-91.0792.8586.2486.8187.1385.7590.7677.5979.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.292, + 0.905, + 0.308 + ], + "angle": 0, + "content": "Table 1. Comparison with existing visual grounding methods on RefCOCO, RefCOCOg, RefCOCO+, and ReferItGame for the REC task." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.321, + 0.851, + 0.433 + ], + "angle": 0, + "content": "
MethodsVisual BackboneMulti-taskVenuevalRefCOCO testAtestBRefCOCOgvalRefCOCO+ testAtestB
MCN [28]DN53✓CVPR2062.4464.2059.7149.2249.4050.6254.9944.69
RefTR [22]RN101✓NeurIPS2170.5673.4966.5758.7358.5161.0864.6552.73
SeqTR [59]DN53✓ECCV2267.2669.7964.1255.6755.6454.1458.9348.19
VG-LAW [38]ViT-B✓CVPR2375.6277.5172.8965.6366.0866.6370.3858.89
PVD [7]Swin-B✓AAAI2475.0777.2970.1363.2263.8964.3969.1557.19
Vista-7B [34]ViT-L✓CVPR2474.576.073.969.870.971.874.465.6
TCRTSwin-B✓-77.1278.1275.6770.8472.0872.7974.9666.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.132, + 0.443, + 0.863, + 0.457 + ], + "angle": 0, + "content": "Table 2. Comparison with existing visual grounding methods on RefCOCO, RefCOCOg, and RefCOCO+ for the RES task." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.483, + 0.296, + 0.5 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.507, + 0.196, + 0.52 + ], + "angle": 0, + "content": "4.1.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.676 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method, we carry out experiments on the RefCOCO [51], RefCOCOg [33], RefCOCO+ [51], and ReferItGame [20] datasets. RefCOCO, RefCOCOg, and RefCOCO+ datasets are derived from MS-COCO [23]. RefCOCO and RefCOCO+ datasets were partitioned into train, validation (val), testA, and testB subsets. RefCOCOg was partitioned into train, validation (val), and test subsets. In addition, the ReferItGame dataset, which was collected from SAIAPR-12 [13], comprises both train and test subsets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.271, + 0.7 + ], + "angle": 0, + "content": "4.1.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.782 + ], + "angle": 0, + "content": "For the RSC task, we employ Prec@0.5 as the metric for evaluation, which is consistent with prior works. If the IoU of the ground truth box with the predicted box surpasses 0.5, then the predicted box is deemed accurate. For the RES task, we evaluate performance by calculating MIoU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.303, + 0.805 + ], + "angle": 0, + "content": "4.1.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our visual backbone utilizes a Swin Transformer architecture pre-trained with Mask-RCNN on MSCOCO [23]. The input image size is set to \\(448 \\times 448\\). For the textual backbone, we adopt the T5 [35] model and limit the maximum length of text to 40. For the large language model, we use Llama3.1-8B [12], which is frozen during the training" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.48, + 0.895, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.693, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with representative methods on REC task. We also visualize the attention scores of each method. Blue and green boxes are predicted regions and the ground truth, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.9 + ], + "angle": 0, + "content": "process. Our model undergoes end-to-end training for 70 epochs. The learning rate begins at \\(10^{-4}\\), and then we drop it down to \\(10^{-5}\\) after 50 epochs. The optimizer used is AdamW. A dropout rate of 0.1 is applied. The batch size is configured to 16. We adopt the same data augmentation strategies as previous works [47, 48]. We implement our TCRT using PyTorch and perform the experiments on NVIDIA A800 GPUs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3936" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.089, + 0.218, + 0.102 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.089, + 0.312, + 0.102 + ], + "angle": 0, + "content": "SeqTR" + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.089, + 0.425, + 0.102 + ], + "angle": 0, + "content": "TCRT" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.103, + 0.229, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.103, + 0.345, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.103, + 0.461, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.163, + 0.228, + 0.174 + ], + "angle": 0, + "content": "\"fourth from bottom ri" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.174, + 0.229, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.174, + 0.345, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.174, + 0.461, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.234, + 0.277, + 0.245 + ], + "angle": 0, + "content": "\"big couch at bottom of screen\"" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.245, + 0.229, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.245, + 0.345, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.245, + 0.461, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.306, + 0.26, + 0.317 + ], + "angle": 0, + "content": "\"giraffe appendage on right\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.327, + 0.463, + 0.342 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison with SeqTR on RES task." + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.359, + 0.196, + 0.369 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.358, + 0.327, + 0.368 + ], + "angle": 0, + "content": "TCRT w/o CN" + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.358, + 0.413, + 0.368 + ], + "angle": 0, + "content": "TCRT" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.37, + 0.234, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.37, + 0.341, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.37, + 0.45, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.426, + 0.239, + 0.436 + ], + "angle": 0, + "content": "\"white laptop keyboard\"" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.437, + 0.232, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.438, + 0.34, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.438, + 0.449, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.494, + 0.268, + 0.504 + ], + "angle": 0, + "content": "\"person half shown on far left\"" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.506, + 0.232, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.506, + 0.34, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.506, + 0.449, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.563, + 0.283, + 0.573 + ], + "angle": 0, + "content": "\"Guy lower right side facing away\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.641 + ], + "angle": 0, + "content": "Figure 6. Visualization of features extracted by the visual backbone. The second column shows the results of 'TCRT w/o CN', while the third column shows the results of TCRT. Red and white boxes are the predicted regions and the ground truth, respectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.668, + 0.399, + 0.684 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.751 + ], + "angle": 0, + "content": "To evaluate the effectiveness of the proposed TCRT, we perform both qualitative and quantitative experiments across four datasets: RefCOCO, RefCOCOg, RefCOCO+, and ReferItGame." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.76, + 0.206, + 0.775 + ], + "angle": 0, + "content": "4.2.1. REC task" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate our TCRT against existing state-of-the-art models for the REC task. As shown in Table 1, TCRT achieves the best performance across these datasets. Specifically, compared to Vista-7B, TCRT shows absolute improvements of \\(+2.97\\), \\(+1.35\\), and \\(+3.24\\) percentage points on RefCOCO; \\(+1.65\\), \\(+0.46\\), and \\(+1.79\\) percentage points on RefCOCO+; and \\(+0.81\\) and \\(+0.73\\) percentage points on RefCOCOg. We also present qualitative results from three" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.091, + 0.877, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.261, + 0.906, + 0.29 + ], + "angle": 0, + "content": "Figure 7. Comparison between the TCRT without object-text matching loss \\((\\mathcal{L}_{\\mathrm{match}})\\) and TCRT in terms of \\(\\mathcal{L}_{\\mathrm{det}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.907, + 0.407 + ], + "angle": 0, + "content": "examples of the RefCOCO testB set for the REC task in Figure 4. We can see that TCRT performs well even when an image contains multiple similar objects, such as \"second vase from front\". Furthermore, we visualized the attention scores within the transformer. We observe that TCRT focuses better on the target object compared to other methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.414, + 0.625, + 0.427 + ], + "angle": 0, + "content": "4.2.2. RES task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.568 + ], + "angle": 0, + "content": "We evaluate the effectiveness of our TCRT for the RES task by comparing it with existing state-of-the-art methods. The comparative results are provided in Table 2. Compared with Vista-7B, TCRT achieves higher mIoU scores with absolute improvements of \\(+2.62\\), \\(+2.12\\), and \\(+1.77\\) percentage points on RefCOCO, \\(+0.99\\), \\(+0.56\\), and \\(+0.93\\) percentage points on RefCOCO+, \\(+1.04\\) and \\(+1.18\\) percentage points on RefCOCOg. We also present a comparison for the RES task in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.578, + 0.782, + 0.594 + ], + "angle": 0, + "content": "4.3. Analysis of Main Components" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.747, + 0.613 + ], + "angle": 0, + "content": "4.3.1. Conditional Normalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.906, + 0.739 + ], + "angle": 0, + "content": "We first conduct an experiment to demonstrate the effectiveness of conditional normalization (CN). We visualize the features extracted by the visual backbone on RefCOCOg test set. As shown in Figure 6, 'TCRT w/o CN' has more dispersed attention, which may lead to inaccurate predictions. Due to the smaller domain discrepancy of TCRT, this allows the LLM to focus more attention on the target objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.885, + 0.762 + ], + "angle": 0, + "content": "4.3.2. Object-text Semantic Matching Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We use the object-text matching loss \\((\\mathcal{L}_{\\mathrm{match}})\\) to associate the visual tokens with the textual ones. To validate the effectiveness of \\(\\mathcal{L}_{\\mathrm{match}}\\) we compare 'TCRT w/o \\(\\mathcal{L}_{\\mathrm{match}}\\) and TCRT in terms of the detection loss \\((\\mathcal{L}_{\\mathrm{det}})\\) on RefCOCO val set. In Figure 7, by plotting the loss curves of the models during training, we observe that TCRT achieves lower loss values, which is attributed to its capability to learn semantic-discriminative visual features. This enables the LLM to more easily recognize task-relevant visual features." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3937" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.473, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.294, + 0.483, + 0.365 + ], + "angle": 0, + "content": "Figure 8. visualization of LLM-guided visual feature refinement. We visualize the weight matrices \\(\\mathcal{M}\\) of each layer in the fusion module to validate the role of the LLM in refining visual features. Red and white boxes are the predicted regions and the ground truth, respectively." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.378, + 0.473, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.588, + 0.435, + 0.602 + ], + "angle": 0, + "content": "Figure 9. Analysis of Visual Feature Distribution." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.628, + 0.418, + 0.643 + ], + "angle": 0, + "content": "4.3.3. LLM-guided Visual Feature Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.768 + ], + "angle": 0, + "content": "We visualize the weight matrix \\(\\mathcal{M}\\) in Figure 8 to validate the effectiveness of visual feature refinement on RefCOCOg test set. We observe that the attention of the weight matrix is relatively evenly distributed in the earlier layers. In the later layers, the attention of the weight matrix is primarily focused on the target object. This indicates that LLM's prior knowledge enhances our model's ability to locate target objects by further refining the visual features." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.776, + 0.409, + 0.791 + ], + "angle": 0, + "content": "4.3.4. Analysis of Visual Feature Distribution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To validate the effectiveness of our model in aligning visual and textual features, we extract the visual features after fusion with textual information and study their distribution in the spatial domain. Specifically, we label four target objects in an image and use three different textual descriptions for each object, resulting in twelve data samples. These samples are fed into the grounding model, and the visual" + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.089, + 0.858, + 0.172 + ], + "angle": 0, + "content": "
MethodvalRefCOCO
testAtestB
TCRT w/o VF & VA87.6288.6084.04
TCRT w/o VF88.3890.2485.57
TCRT91.0792.8586.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.183, + 0.907, + 0.224 + ], + "angle": 0, + "content": "Table 3. Ablation experiments on RefCOCO to evaluate the proposed LLM adaptation module \\( \\left( {V}_{A}\\right) \\) and task-aware cross-modal feature fusion module \\( \\left( {V}_{F}\\right) \\) ." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.907, + 0.315 + ], + "angle": 0, + "content": "features from the fusion module's output are subjected to dimensionality reduction and visualization. As shown in Figure 9, we observe that our method causes the visual features describing the same object to cluster together." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.326, + 0.676, + 0.339 + ], + "angle": 0, + "content": "4.3.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.512 + ], + "angle": 0, + "content": "We validate the design of our proposed TCRT by conducting ablative experiments on major components in the REC task. The results are presented in Table 3. We observe that when the LLM adaptation module and task-aware cross-modal feature fusion module are not used, 'TCRT w/o \\( V_{F} \\) & \\( V_{A} \\)' achieves accuracy rates of \\( 87.62\\% \\), \\( 88.60\\% \\), and \\( 84.04\\% \\) on the RefCOCO dataset. After incorporating the LLM adaptation module, the accuracy of 'TCRT w/o \\( V_{F} \\)' increases by 0.76, 1.64, and 1.53 absolute percentage points. When the \\( V_{F} \\) module is further introduced, TCRT achieves accuracy rates of \\( 91.07\\% \\), \\( 92.85\\% \\), and \\( 86.24\\% \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.634, + 0.544 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.768 + ], + "angle": 0, + "content": "In this paper, we propose a task-aware cross-modal feature refinement transformer with large language models to facilitate visual grounding. To adapt LLMs to the grounding task, we design an adaptation module that modulates the statistical characteristics of visual features conditioned on textual features along with object-text semantic matching regularization. An LLM is employed to integrate the aligned information from textual and visual features and produce task-aware embeddings. We further incorporate a task-aware cross-modal feature fusion module, allowing the resulting embeddings to refine visual features and enable information interaction between the REC and RES tasks. Comprehensive experimental results verify the effectiveness of our proposed approach both quantitatively and qualitatively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.784, + 0.692, + 0.801 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by TCL Science and Technology Innovation Fund (Project No. 20231752), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), and in part by the GuangDong Basic and Applied Basic Research Foundation (Project No. 2024A1515011437)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3938" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Mohit Bajaj, Lanjun Wang, and Leonid Sigal. G3raphground: Graph-based language grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4281-4290, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.482, + 0.269 + ], + "angle": 0, + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.41 + ], + "angle": 0, + "content": "[5] Kan Chen, Rama Kovvuri, and Ram Nevatia. Query-guided regression network with context policy for phrase grounding. In Proceedings of the IEEE International Conference on Computer Vision, pages 824-832, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.484, + 0.479 + ], + "angle": 0, + "content": "[6] Long Chen, Wenbo Ma, Jun Xiao, Hanwang Zhang, and Shih-Fu Chang. Ref-nms: Breaking proposal bottlenecks in two-stage referring expression grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1036-1044, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[7] Zesen Cheng, Kehan Li, Peng Jin, Siheng Li, Xiangyang Ji, Li Yuan, Chang Liu, and Jie Chen. Parallel vertex diffusion for unified visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1326-1334, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[9] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1769-1779, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[10] Jiajun Deng, Zhengyuan Yang, Daqing Liu, Tianlang Chen, Wengang Zhou, Yanyong Zhang, Houqiang Li, and Wanli Ouyang. Transvg++: End-to-end visual grounding with language conditioned vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[11] Pelin Dogan, Leonid Sigal, and Markus Gross. Neural sequential phrase grounding (seqground). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4175-4184, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[13] Hugo Jair Escalante, Carlos A Hernández, Jesus A Gonzalez, Aurelio López-López, Manuel Montes, Eduardo F Morales, L Enrique Sucar, Luis Villasenor, and Michael Grubinger. The segmented and annotated iapr tc-12 benchmark. Computer Vision and Image Understanding, 114(4): 419-428, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[14] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.904, + 0.275 + ], + "angle": 0, + "content": "[15] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 1440-1448, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[16] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 580-587, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[17] Chih-Hui Ho, Srikar Appalaraju, Bhavan Jasani, R Manmatha, and Nuno Vasconcelos. Yoro-lightweight end to end visual grounding. In European Conference on Computer Vision, pages 3-23. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[18] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1115-1124, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.547, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[20] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 787-798, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[21] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv e-prints, pages arXiv-2211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[22] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in Neural Information Processing Systems, 34:19652-19664, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.745, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.905, + 0.885 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.887, + 0.904, + 0.902 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3939" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[27] Xihui Liu, Zihao Wang, Jing Shao, Xiaogang Wang, and Hongsheng Li. Improving referring expression grounding with cross-modal attention-guided erasing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1950-1959, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[28] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10034-10043, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[29] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. In European Conference on Computer Vision, pages 417-435. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11-20, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[31] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 Fourth International Conference on 3D Vision (3DV), pages 565-571. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[32] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[33] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[34] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.907, + 0.273 + ], + "angle": 0, + "content": "[38] Wei Su, Peihan Miao, Huanzhang Dou, Gaoang Wang, Liang Qiao, Zheyang Li, and Xi Li. Language adaptive weight generation for multi-task visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10857-10866, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[39] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[43] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[44] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4145-4154, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[45] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.907, + 0.719 + ], + "angle": 0, + "content": "[46] Sibei Yang, Guanbin Li, and Yizhou Yu. Graph-structured referring expression reasoning in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9952–9961, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[47] Zhengyuan Yang, Boqing Gong, Liwei Wang, Wenbing Huang, Dong Yu, and Jiebo Luo. A fast and accurate one-stage approach to visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4683-4693, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[48] Zhengyuan Yang, Tianlang Chen, Liwei Wang, and Jiebo Luo. Improving one-stage visual grounding by recursive subquery construction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 387-404. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Jiabo Ye, Junfeng Tian, Ming Yan, Xiaoshan Yang, Xuwu Wang, Ji Zhang, Liang He, and Xin Lin. Shifting more" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3940" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15502-15512, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[50] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.301 + ], + "angle": 0, + "content": "[51] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[52] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[53] Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. In European Conference on Computer Vision, pages 405-422. Springer, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.445, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[54] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An lmm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.501, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[55] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4158-4166, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[56] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[57] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.482, + 0.739 + ], + "angle": 0, + "content": "[58] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. Advances in Neural Information Processing Systems, 37:71737-71767, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[59] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision, pages 598-615. Springer, 2022. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3941" + } + ] +] \ No newline at end of file diff --git a/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_origin.pdf b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0fbf751cdc9cc79036988aaf7a7ca1030a6d47b3 --- /dev/null +++ b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/3f001702-7969-4b52-b9c0-e5b93c68bbfc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17d3c867cad72c23dd8b60e142c8d0c580c1e4e065369217965031a6eb6e5162 +size 2192684 diff --git a/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/full.md b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c4d299310a3b210e75e51a3762811cdc30596d85 --- /dev/null +++ b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/full.md @@ -0,0 +1,404 @@ +# Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding + +Wenbo Chen1*; Zhen Xu1,2*; Ruotao Xu2, Si Wu1,2†; Hau-San Wong3 + +$^{1}$ School of Computer Science and Engineering, South China University of Technology $^{2}$ Institute for Super Robotics (Huangpu) + +$^{3}$ Department of Computer Science, City University of Hong Kong + +{cscwb106, csxuzhen}@mail.scut.edu.cn, {xrt, cswusi}@scut.edu.cn, cshswong@cityu.edu.hk + +# Abstract + +The goal of visual grounding is to establish connections between target objects and textual descriptions. Large Language Models (LLMs) have demonstrated strong comprehension abilities across a variety of visual tasks. To establish precise associations between the text and the corresponding visual region, we propose a Task-aware Cross-modal feature Refinement Transformer with LLMs for visual grounding, and our model is referred to as TCRT. To enable the LLM trained solely on text to understand images, we introduce an LLM adaptation module that extracts text-related visual features to bridge the domain discrepancy between the textual and visual modalities. We feed the text and visual features into the LLM to obtain task-aware priors. To enable the priors to guide the feature fusion process, we further incorporate a cross-modal feature fusion module, which allows task-aware embeddings to refine visual features and facilitate information interaction between the Referring Expression Comprehension (REC) and Referring Expression Segmentation (RES) tasks. We have performed extensive experiments to verify the effectiveness of the main components and demonstrate the superior performance of the proposed TCRT over state-of-the-art end-to-end visual grounding methods on RefCOCO, RefCOCOg, RefCOCO+ and ReferItGame. + +# 1. Introduction + +Visual grounding aims to predict the locations of the objects specified by textual prompts in images, and is widely applied in the field of human-computer interaction. This task typically involves referring expression comprehension [10, 17, 49, 59] and referring expression segmentation [22, 38]. The main challenge is to comprehensively un + +![](images/f4c1f36b10a23daa8ce5df93c45225be6adc8c0b185ac6ecb1488ce8e50cb48f.jpg) +TCRT w/o LLM +"bride throwing bouquet + +![](images/838a2900a098502cd3251275b99d6c3552bebf15354bec8bda1f28a4a0faae6e.jpg) +"Guy front in gray suit left" + +![](images/44aa74d849d420267f520734e534fa16bcd1bc4138d5ced316d320dc07448cca.jpg) +"man in right front with blue tie next to older woman +Figure 1. The visualization of the refined visual features from the cross-modal feature fusion module. In challenging environments, the introduction of LLMs can facilitate a better association between visual features and text. + +![](images/e38bddff4d03ab149e64c193388b3c560b3621c1a333227c8d25cf4b29b648ef.jpg) +TCRT + +![](images/d0962ec6767962463598628d3cc48dd6dd4c8707d784c10251d85cf4c2043000.jpg) + +![](images/83e68a042ef610ba44bcc52c4e667e60436031a0f2272c2ac6a441ab209a27ff.jpg) +predicted by + +![](images/e1319efd6665c724e37d17b0c6b12437b265ab2ad1953f9b183f6914b1df72cd.jpg) + +![](images/974bc6bdd894f345e634f5c8d15cbe248c63ba37df684f639d90a33d281e066e.jpg) + +![](images/cc54ee27f0edb5b8c64c46f960933757efa8845894bb5673567acd8d0319ae5e.jpg) +ground truth + +derstand intra-modality information and accurately establish inter-modality correspondences. + +Existing visual grounding methods can be categorized into two main types: two-stage and end-to-end methods. The two-stage methods [45] first generate region proposals [36] that may match the text, and then rank the candidate regions according to their relevance to the text. When the proposals generated in the first stage do not contain the target object, the second stage will be unable to correct these proposals. On the other hand, end-to-end grounding methods perform the integration of visual and textual features within an object detector, producing the box coordinates of the target object. The end-to-end methods are typically faster than the two-stage ones. With the advancement of transformer in the vision domain, TransVG [9, 10] and SeqTR [59] employ self-attention or cross-attention mechanisms to fuse textual and visual features. Recently, LLaVA [25] has integrated visual understanding capabilities into large language mod + +els (LLMs). MiniGPT-2 [4] and OMG-LLaVA [58] apply LLMs to visual grounding tasks by integrating textual and visual features. In this work, we make an attempt to adapt LLMs to capture the information of target objects and produce task-aware embeddings for precise grounding results. + +More specifically, we propose a task-aware cross-modal feature refinement transformer with large language models for visual grounding, which is referred to as TCRT, leveraging priors from LLMs to guide the multimodal information fusion process and achieve precise alignment between textual and local visual features (see Figure 1). To enable the LLM to more effectively process image data and adapt to visual grounding tasks, we propose an LLM adaptation module that utilizes conditional normalization to adjust the statistical characteristics of visual features based on textual information, while introducing object-text semantic matching regularization to associate local visual features with textual features. Additionally, to utilize the priors provided by the adaptation module, we incorporate a task-aware cross-modal feature fusion module, which continuously refines visual features during multimodal fusion and enables the model to exploit the relationship between the REC and RES tasks. We perform extensive qualitative and quantitative experiments to validate the effectiveness of our model, and demonstrate its advantages on benchmark datasets. The main contributions of this work are summarized as follows: + +- Different from the state-of-the-art methods that focus on the application of LLMs in integrating textual and visual features, we further explore the LLM's capability of learning task-aware embeddings to facilitate visual grounding. +- We incorporate an LLM adaptation module for visual grounding, which learns to capture the information of target objects by modulating the statistical characteristics of visual features conditioned on textual features. +- We further perform task-aware cross-modal feature fusion, allowing the task-aware embeddings derived from LLMs to guide the association of textual and visual information and lead to precise grounding results. + +# 2. Related Work + +# 2.1. Two-stage Visual Grounding + +In two-stage methods, a region proposal network [15, 16, 36] is first used to produce a set of candidate regions that may correspond to the target object mentioned in the textual description. These candidate regions are then matched to the textual description, and the predicted result is determined by identifying the region with the highest score. The maximum-margin ranking loss [30] is widely adopted for the second stage. Several researches further enhance the two-stage methods by incorporating context modeling [18, 32, 44, 55] and making use of phrase cooccur + +ences [2, 5, 11]. MAttNet [52] decomposed the expression into three modules: relationship, location, and subject to improve the grounding accuracy. To enhance the model's capacity for learning detailed alignments between text and visuals, CM-A-E [27] proposed a method capable of generating challenging samples for training the model. Graph learning was incorporated into visual grounding by SGMN [46] to enable the model to better capture the relationships between different objects within the same scene. To increase the recall of target objects, Ref-NMS [6] utilized textual features to direct the non-maximum suppression. To enhance the model's cross-modality fusion capabilities, GroundingDINO [26] integrated grounded pretraining techniques into detection transformers. + +# 2.2. End-to-end Visual Grounding + +End-to-end methods in visual grounding aim to directly predict the bounding boxes in an image from a text in a single pass, resulting in faster inference speeds compared to two-stage methods. As a pioneering work, FAOA [47] concatenated image and textual features, and fed the concatenated features into a prediction layer to achieve target object detection. To reduce the referring ambiguity, ReSC [48] used a sub-query construction to address limitations in processing complex textual descriptions. TransVG [9, 10] is one of the pioneering transformer-based visual grounding models that directly regresses coordinates. To achieve multi-task learning, RefTR [22] used the transformer architecture to integrate REC and RES tasks into a single framework. To address the gap between the visual and textual backbones, QRNet [49] incorporated textual information into the visual backbone. Furthermore, VG-LAW [38] actively extracted text-relevant visual features based on text-adaptive weights and performed multi-task visual grounding. Additionally, PVD [7] introduced a diffusion model into the visual grounding task. + +With the development of LLMs, some Multimodal Large Language Models [25] (MLLMs) are gradually being applied to visual grounding tasks. VistaLLM [34] used an instruction-guided image tokenizer to filter and compress image embeddings based on task, extracting visual information relevant to the current task. To efficiently locate and understand multiple target regions in complex scenes, Groma [29] introduced a localized visual tokenization mechanism that decomposes an image into multiple regions of interest, encoding each region as an independent region token. To unleash the potential of large language models, Griffon [53] introduces a foundational language-prompted localization dataset with nearly 6 million pretraining samples. + +# 2.3. Large Language Models + +With the expansion of training data and a significant increase in model parameters, large language models (LLMs) + +![](images/976b3e3cd8eb1c19697c61cabd6b556ac4756f0ce8920a55e449140240bc4f18.jpg) +Figure 2. An overview of the proposed method. We first extract textual features $F_{txt}$ using the textual backbone $E_{txt}$ . These features are then injected into the visual backbone $E_{vis}$ , which consists of transformer blocks and conditional normalization, to extract text-related visual features $F_{vis}$ . The object-text semantic matching regularization is employed to associate visual region tokens with textual tokens. We input the visual features processed by the visual adapter $V_{proj}$ , along with the text, into LLaMa. We take the last token $f_{llm}$ output from LLaMa. $f_{llm}$ is then fed into a query projection layer to generate task-relevant tokens $f_{box}$ , $f_{seg}$ , and $f_{phr}$ , which are used to predict the bounding box, segmentation, and phrase, respectively. A task-aware cross-modal feature fusion module is then employed to integrate textual, visual, and task features. Finally, task-relevant tokens are each fed to the corresponding prediction heads to produce the final predictions. + +based on the Transformer [42] architecture have developed rapidly. GPT-3 [3], with its 175 billion parameters, demonstrates strong performance in few-shot contextual learning. Following the success of GPT-3, several other LLMs, including PaLM [8], OPT [57], BLOOM [21], and LLaMA [40, 41], have emerged in succession. Mistral [19] introduced a specialized window attention mechanism to better handle extended context modeling. In parallel, Mixtral $8 \times 7$ B utilized sparse MoE layers [14] to scale up parameters more effectively, enabling it to achieve higher performance with a reduced number of active parameters. + +The main differences between the proposed TCRT and existing visual grounding methods are summarized as follows: First, we perform statistical modulation on visual features conditioned on textural features to reduce domain discrepancy before integrating them, while the existing methods [4, 53, 54] only perform dimension adaption. Second, TCRT leverage the task-aware embeddings derived from LLMs to guide cross-modal feature fusion and association analysis, rather than directly predicting the bounding boxes of target objects as in previous methods [4, 53, 54]. + +# 3. Methodology + +# 3.1. Problem Setting + +In this section, we begin by defining multi-task visual grounding (MTVG), and then proceed to describe the architecture of the proposed framework. The goal of MTVG is to detect or segment a target object based on a textual description. Formally, given an image $I \in \mathbb{R}^{3 \times H \times W}$ and a corresponding text $q$ , this task is to predict a bounding box $b = \{x, y, w, h\}$ that encompasses the object men + +tioned in the text, as well as a pixel-wise segmentation mask $s\in \mathbb{R}^{1\times H\times W}$ that provides a more detailed and precise localization. A textual backbone $E_{txt}$ and a visual backbone $E_{vis}$ are used to extract textual and visual features, respectively. + +# 3.2. LLM Adaptation + +Pre-trained LLMs possess language understanding capabilities and can handle many complex tasks in natural language processing. However, due to the domain discrepancy between textual data and visual data, directly applying LLMs to grounding tasks may yield suboptimal performance. Therefore, we propose an LLM adaptation module to perform cross-modal adaptation of image information, enabling LLMs to better handle information within visual data. As shown in Figure 2, we first utilize the text encoding model $E_{txt}$ to extract the textual features of $q$ : $F_{txt} = E_{txt}(q)$ , where $F_{txt}$ represents the textual features, which include three types of textual features: sentence feature $f_{txt}^{sen} \in \mathbb{R}^{1 \times d}$ , individual word features $f_{txt} \in \mathbb{R}^{L \times d}$ , and average pooled word feature $f_{txt}^{mean} \in \mathbb{R}^{1 \times d}$ , where $d$ and $L$ represent the dimension of textual features and the number of tokens, respectively. + +To enable the visual encoder to comprehend textual information, we designed a text-aware visual encoder $E_{vis}$ . Specifically, a conditioned normalization is inserted after each transformer block in the visual encoder. In the conditioned normalization, we fuse the sentence features $f_{txt}^{sen}$ and the average-pooled word features $f_{txt}^{mean}$ into a textual conditional feature $f_{txt}^{cond}$ via a weighted combination: + +$$ +f _ {t x t} ^ {\text {c o n d}} = \alpha^ {\text {s e n}} f _ {t x t} ^ {\text {s e n}} + \alpha^ {\text {m e a n}} \phi \left(f _ {t x t} ^ {\text {m e a n}}\right), \tag {1} +$$ + +![](images/77472a9cf2b8779aff75b70d7bf29a59104fe024f03c5632df6d65d920c0bea4.jpg) +Figure 3. Object-text semantic matching regularization. We constrain the image features using a mask to isolate the object-specific features. For matching text-object pairs, we consider them as positive samples, whereas non-matching pairs are regarded as negative samples. + +where $\alpha^{sen}$ and $\alpha^{mean}$ are learnable parameters that control the contribution of $f_{txt}^{sen}$ and $f_{txt}^{mean}$ , respectively. $\phi(\cdot)$ represents the feed-forward network. The scaling $\gamma$ and shifting $\beta$ parameters are determined using $f_{txt}^{cond}:\gamma = \phi(f_{txt}^{cond}) + \gamma^{learn}$ , $\beta = \phi(f_{txt}^{cond}) + \beta^{learn}$ , where $\gamma^{learn}$ and $\beta^{learn}$ are learnable parameters. Given the visual features $\bar{F}_i^*$ , which is the output of the $i$ -th layer of transformer block, we calculate the mean $\mu(\bar{F}_i^*)$ and standard deviation $\sigma(\bar{F}_i^*)$ . The visual features $\bar{F}_i^*$ are then modulated using the resulting scaling $\gamma$ and shifting $\beta$ parameters, as follows: + +$$ +\bar {F} _ {i} ^ {*} = \gamma \left(\frac {\bar {F} _ {i} ^ {*} - \mu (\bar {F} _ {i} ^ {*})}{\sigma (\bar {F} _ {i} ^ {*}) + \epsilon}\right) + \beta , \tag {2} +$$ + +where $\epsilon$ is a small constant introduced to avoid division by zero. We obtain the text-aware visual features $F_{vis}$ by flattening the feature map $\overline{F}_4^*$ extracted from the visual backbone. The image features and the input embeddings of LLMs are typically in different dimensional spaces. We use a visual adapter $V_{proj}$ to map the image features to a dimension that matches the LLM input, allowing the LLM to directly process visual information. $V_{proj}$ is a three-layer MLP. We feed the text and the mapped image features into the LLM, and take the last token of the LLM output as the prior embedding $f_{llm}$ : + +$$ +f _ {l l m} = \Phi (q, V _ {p r o j} \underbrace {\left(E _ {v i s} (I , f _ {t x t} ^ {c o n d})\right)} _ {F _ {v i s}}), \tag {3} +$$ + +where $\Phi$ represents an LLM model, $q$ denotes the original query text. We pass $f_{llm}$ through three task-aware linear layers ( $W_{box}$ , $W_{seg}$ , and $W_{phr}$ ) to obtain the task-specific embeddings $f_{box}$ , $f_{seg}$ , and $f_{phr}$ , which are responsible for predicting the bounding box, segmentation, and phrase, respectively. + +The text usually describes a specific object within an image, which occupies only a portion of the image. Most visual features do not have a direct relationship with the + +text. This makes it challenging to achieve precise semantic alignment between the target object and textual tokens. Thus, we propose the object-text semantic matching regularization to associate the visual region tokens with the textual ones. Object-text matching is a binary classification task, and the model evaluates the multimodal features of an object-text pair to classify it as either positive (matched) or negative (unmatched). As shown in Figure 3, if the textual features and visual features represent the same object, they are considered a positive sample pair; otherwise, they are considered a negative sample pair. The ground truth of object segmentation is represented by $s^+$ . We apply an affine transformation to $s^+$ to obtain $s^-$ . $s^-$ represents the inaccurate segmentation result. We leverage the segmentation mask of the target object to extract its features. A learnable token $f_{match}$ is then introduced to determine whether they are matched: + +$$ +\begin{array}{l} \left\{\bar {f} _ {\text {m a t c h}}, \bar {F} _ {l}, \bar {F} _ {v} \right\} = \varphi \left(\left[ f _ {\text {m a t c h}}, F _ {\text {t x t}}, F _ {\text {v i s}} \right]\right), \\ \left\{\hat {f} _ {\text {m a t c h}} ^ {j}, \hat {F} _ {l} \right\} = \theta \left(\left[ \bar {f} _ {\text {m a t c h}}, \bar {F} _ {l} \right], F _ {\text {v i s}} \odot s ^ {j}\right), \tag {4} \\ \end{array} +$$ + +where $j \in \{+, -\}$ . $\varphi(q)$ and $\theta(q, kv)$ represent multi-head self-attention and cross-attention, respectively. $\odot$ denotes element-wise multiplication. $[\cdot]$ denotes the concatenation operation. We use the $\hat{f}_{match}^{j}$ as the joint representation of the object-text pair. A feed-forward network is employed to predict the two-class probability $p_{match}^{j}$ that indicates whether a textual feature matches a local visual feature. The object-text matching loss is: + +$$ +\mathcal {L} _ {\text {m a t c h}} = \mathbb {E} _ {(I, q)} [ \sum_ {j} H \left(y _ {\text {m a t c h}} ^ {j}, p _ {\text {m a t c h}} ^ {j}\right) ], \tag {5} +$$ + +where $H$ represents cross-entropy and $y_{match}^{j}$ is a 2-dimensional one-hot vector that represents the ground truth label. We use the visual features of the target object extracted from the ground truth mask as positive sample, indicating that these features match the textual description. When performing affine transformation (partial overlap may occur), the resulting mask deviates from the target object, and the extracted features are considered as negative sample. Object-text matching loss further associates visual tokens with textual ones by assessing whether they match. + +# 3.3. Task-aware Cross-modal Feature Fusion + +When the LLM receives textual and visual information, it can generate more meaningful semantic representations. These representations contain the LLM's understanding of objects and context, forming prior information that aids downstream tasks. The task-aware cross-modal feature fusion module, denoted as $V_{F}$ , is incorporated to guide feature the fusion process. Specifically, we first employ a bidirectional multi-head attention module to fuse textual and + +visual features. Each modality serves as the query while the other modality acts as the key and value in the attention computation. We concatenate the three tokens used for prediction, $f_{box}$ , $f_{seg}$ , and $f_{phr}$ , with the textual features $F_{txt}$ to obtain a task-relevant textual feature $F_{task}$ . Then, $F_{task}$ and $F_{vis}$ are fed into the bi-directional multi-head attention module to fuse the two types of features. This process can be represented as follows: $F_{task}^{v2t} = \theta (F_{task}, F_{vis})$ , $F_{vis}^{t2v} = \theta (F_{vis}, F_{task})$ , where $F_{task}^{v2t}$ represents the textual feature enriched with image features, comprising $F_{txt}^{v2t}$ , $f_{box}^{v2t}$ , $f_{seg}^{v2t}$ , and $f_{phr}^{v2t}$ . Meanwhile, $F_{vis}^{t2v}$ denotes the visual feature enriched with textual information. + +Target objects are typically located in specific regions. To enable the model to focus on local visual features, we input $f_{llm}$ into an MLP to obtain the refinement prompt $f_{prmt}$ , which guides the feature refinement process. We then assign different weights to image features at different locations by computing the cosine similarity between the image feature $F_{vis}^{t2v}$ and $f_{prmt}$ . This process can be formulated as follows: + +$$ +\hat {F} _ {v} ^ {t 2 v} = F _ {v i s} ^ {t 2 v} + F _ {v i s} ^ {t 2 v} \odot \underbrace {\left(\operatorname {C o n v} _ {1 \times 1} \left(F _ {v i s} ^ {t 2 v}\right) \otimes \mathrm {M L P} \left(f _ {l l m}\right)\right)} _ {\mathcal {M}} \tag {6} +$$ + +where $\otimes$ denotes matrix multiplication and $\odot$ represents element-wise multiplication. The refinement prompt $f_{prmt}$ is dot-multiplied with the image features at each position in $F_{vis}^{t2v}$ to obtain a weight matrix $\mathcal{M}$ . The final refined visual features $\hat{F}_v^{t2v}$ are obtained under the guidance of the LLM. The model can control the strength of the feature representations at different positions in the image by adjusting the weight matrix $\mathcal{M}$ . By assigning lower weights to irrelevant regions, the model can better focus on the features of the target regions. + +There is a close relationship between the REC task and the RES task. Both require an understanding and localization of the target object or region. REC tasks generally offer locational information of the target objects, whereas RES tasks offer precise boundary details. To strengthen the relationship between the REC and RES tasks, we perform cross-task interaction to effectively exchange information between the box token $f_{\mathrm{box}}^{v2t}$ and the segmentation token $f_{\mathrm{seg}}^{v2t}$ . The process is as follows: + +$$ +\left\{f _ {b o x} ^ {\prime}, f _ {s e g} ^ {\prime} \right\} = \theta ([ f _ {b o x} ^ {v 2 t}, f _ {s e g} ^ {v 2 t} ], F _ {t x t} ^ {v 2 t}), \tag {7} +$$ + +$$ +f _ {b o x} ^ {\prime \prime} = f _ {b o x} ^ {\prime} + \phi (f _ {s e g} ^ {\prime}), f _ {s e g} ^ {\prime \prime} = f _ {s e g} ^ {\prime} + \phi (f _ {b o x} ^ {\prime}), +$$ + +This design allows the model to exploit the correlation between the REC and RES tasks. The updated $f_{box}$ , $f_{seg}$ , and $f_{phrase}$ are each fed into their respective prediction heads to obtain the final bounding box, segmentation, and phrase predictions. + +# 3.4. Model Training + +Detection loss. For the REC task, the loss for bounding box regression is calculated using a combination of GIoU loss and L1 loss. We use $b$ to denote the ground truth bounding box and $\tilde{b}$ to represent the predicted bounding box. The following defines the detection loss function: + +$$ +\mathcal {L} _ {\det } = \mathbb {E} _ {(I, q)} \left[ \Phi_ {\text {s m o o t h} - l 1} (b, \tilde {b}) + \Phi_ {\text {g i o u}} (b, \tilde {b}) \right], \tag {8} +$$ + +where $\Phi_{\mathrm{smooth} - l1}$ and $\Phi_{\mathrm{giou}}$ denote the smooth L1 loss [15] and Generalized IoU loss [37], respectively. + +Segmentation loss. For the RES task, the segmentation loss is computed by combining the focal loss with dice loss. The ground truth mask is denoted as $s$ , while the predicted mask is $\hat{s}$ . The following defines the segmentation loss function: + +$$ +\mathcal {L} _ {\mathrm {s e g}} = \mathbb {E} _ {(I, q)} \left[ \Phi_ {\text {f o c a l}} (s, \hat {s}) + \Phi_ {\text {d i c e}} (s, \hat {s}) \right], \tag {9} +$$ + +where $\Phi_{\mathrm{focal}}$ and $\Phi_{\mathrm{dice}}$ denote the smooth focal loss [24] and DICE/F-1 loss [31], respectively. + +Phrase loss. To enable the model to more accurately capture the entities within the text, we utilize $f_{phr}$ to predict the words in the text related to the target object. We fed $f_{phr}$ into an MLP to predict a word probability $p_{phr}$ . Then, we compute the phrase loss as follows: + +$$ +\mathcal {L} _ {\mathrm {p h r}} = \mathbb {E} _ {(I, q)} [ H (y _ {\text {p h r}}, p _ {\text {p h r}}) ], \tag {10} +$$ + +where $y_{phr}$ denotes the ground truth label. + +By incorporating these training loss functions, the optimization problem of our model can be formulated as follows: + +$$ +\min _ {E _ {t x t}, E _ {v i s}} \lambda_ {\text {m a t c h}} \mathcal {L} _ {\text {m a t c h}}, \tag {11} +$$ + +$$ +\min _ {E _ {t x t}, E _ {v i s}, V _ {p r o j}, W _ {t a s k}, V _ {F}} \mathcal {L} _ {\mathrm {d e t}} + \mathcal {L} _ {\mathrm {s e g}} + \lambda_ {\mathrm {p h r}} \mathcal {L} _ {\mathrm {p h r}}, +$$ + +where $\text{task} \in \{box,seg,phr\}$ , and $\lambda_{\mathrm{match}}$ and $\lambda_{\mathrm{phr}}$ are used to achieve a balance among the training objectives. $\mathcal{L}_{\mathrm{det}}$ and $\mathcal{L}_{\mathrm{seg}}$ enable the model to perform basic detection and segmentation tasks. $\mathcal{L}_{\mathrm{phr}}$ enables the model to achieve a more detailed understanding of the target objects referred to in the text. $\mathcal{L}_{\mathrm{match}}$ is used to associate visual tokens with textual ones and only optimizes the textual backbone and visual backbone. Our LLM adaptation module and task-aware cross-modal feature fusion module are jointly optimized during the training process. + +# 4. Experiments + +In this section, we first describe datasets, evaluation metrics, and implementation details. Next, we qualitatively and quantitatively compare our method with state-of-the-art methods. We further perform comprehensive experiments to analyze the main components of TCRT, including conditional normalization, object-text semantic matching regularization, and LLM-guided visual feature refinement. + +
MethodsVisual BackboneMulti-taskVenueRefCOCORefCOCOgRefCOCO+ReferItGame test
valtestAtestBvaltestvaltestAtestB
Two-stage Methods
MAttNet [52]RN101×CVPR1876.6581.1469.9966.5867.2765.3371.6256.0229.04
CM-A-E [27]RN101×CVPR1978.3583.1471.3267.9968.6768.0973.6558.03-
Ref-NMS [6]RN101×AAAI2180.7084.0076.0470.5570.6268.2573.6859.42-
End-to-end Methods
MCN [28]DN53✓CVPR2080.0882.2974.9866.4666.0167.1672.8657.31-
RefTR [22]RN101✓NeurIPS2182.2385.5976.5769.4169.4071.5875.9662.1671.42
SeqTR [59]DN53✓ECCV2281.2385.0076.0871.3571.5868.8275.3758.7869.66
VG-LAW [38]ViT-B✓CVPR2386.6289.3283.1676.9076.9676.3781.0467.5077.22
PVD [7]Swin-B✓AAAI2484.9988.0280.0374.3474.6474.2779.0665.11-
Ferret-7B [50]ViT-L×ICLR2487.4991.3582.4583.9384.7680.7887.3873.14-
Vista-7B [34]ViT-L✓CVPR2488.191.583.086.086.484.190.375.8-
TCRTSwin-B✓-91.0792.8586.2486.8187.1385.7590.7677.5979.09
+ +Table 1. Comparison with existing visual grounding methods on RefCOCO, RefCOCOg, RefCOCO+, and ReferItGame for the REC task. + +
MethodsVisual BackboneMulti-taskVenuevalRefCOCO testAtestBRefCOCOgvalRefCOCO+ testAtestB
MCN [28]DN53✓CVPR2062.4464.2059.7149.2249.4050.6254.9944.69
RefTR [22]RN101✓NeurIPS2170.5673.4966.5758.7358.5161.0864.6552.73
SeqTR [59]DN53✓ECCV2267.2669.7964.1255.6755.6454.1458.9348.19
VG-LAW [38]ViT-B✓CVPR2375.6277.5172.8965.6366.0866.6370.3858.89
PVD [7]Swin-B✓AAAI2475.0777.2970.1363.2263.8964.3969.1557.19
Vista-7B [34]ViT-L✓CVPR2474.576.073.969.870.971.874.465.6
TCRTSwin-B✓-77.1278.1275.6770.8472.0872.7974.9666.53
+ +Table 2. Comparison with existing visual grounding methods on RefCOCO, RefCOCOg, and RefCOCO+ for the RES task. + +# 4.1. Experimental Settings + +# 4.1.1. Datasets + +To evaluate the effectiveness of our method, we carry out experiments on the RefCOCO [51], RefCOCOg [33], RefCOCO+ [51], and ReferItGame [20] datasets. RefCOCO, RefCOCOg, and RefCOCO+ datasets are derived from MS-COCO [23]. RefCOCO and RefCOCO+ datasets were partitioned into train, validation (val), testA, and testB subsets. RefCOCOg was partitioned into train, validation (val), and test subsets. In addition, the ReferItGame dataset, which was collected from SAIAPR-12 [13], comprises both train and test subsets. + +# 4.1.2. Evaluation Metrics + +For the RSC task, we employ Prec@0.5 as the metric for evaluation, which is consistent with prior works. If the IoU of the ground truth box with the predicted box surpasses 0.5, then the predicted box is deemed accurate. For the RES task, we evaluate performance by calculating MIoU. + +# 4.1.3. Implementation Details + +Our visual backbone utilizes a Swin Transformer architecture pre-trained with Mask-RCNN on MSCOCO [23]. The input image size is set to $448 \times 448$ . For the textual backbone, we adopt the T5 [35] model and limit the maximum length of text to 40. For the large language model, we use Llama3.1-8B [12], which is frozen during the training + +![](images/6188e1706f311249b9db962a43220c5e9af8df4e1af5512001fc699679ecd39b.jpg) +Figure 4. Qualitative comparison with representative methods on REC task. We also visualize the attention scores of each method. Blue and green boxes are predicted regions and the ground truth, respectively. + +process. Our model undergoes end-to-end training for 70 epochs. The learning rate begins at $10^{-4}$ , and then we drop it down to $10^{-5}$ after 50 epochs. The optimizer used is AdamW. A dropout rate of 0.1 is applied. The batch size is configured to 16. We adopt the same data augmentation strategies as previous works [47, 48]. We implement our TCRT using PyTorch and perform the experiments on NVIDIA A800 GPUs. + +![](images/6358e372ced6bd6f61e4b7bd2eb9d4cac3a4a8cfa7a4e4c7f409f48bfc5bff7d.jpg) +Ground Truth + +![](images/a440a266aefd8f3e14a63d30003e4eb93bfe9860595e43f42ea83530fc5b7b31.jpg) +SeqTR + +![](images/4b69c4f3701904f548eab03417498a7420944447432e9ee5b00853df532275bb.jpg) +TCRT + +![](images/c09b938cf065752ebab34e9edbba87df7fd16f1b1a83de0241066896a0c769e8.jpg) +"fourth from bottom ri + +![](images/3247815e9a75f295c49ff4d3a9f86ba2f67e515f178f0ae1fd26b4ea3bc7b87f.jpg) + +![](images/b75658858a563569c1ada2bf6d255c86c11d8ca9d1bf0d2007f3cce8112a4748.jpg) + +![](images/46f741158de625298d7ea4f7002cc856c029f632428413bfc36f2435cce44bfd.jpg) +"big couch at bottom of screen" +"giraffe appendage on right" +Figure 5. Qualitative comparison with SeqTR on RES task. + +![](images/fc203ab519cc9858ec1a133de9d9bd4d6f1339ddbf1253ebdfa08642abba7461.jpg) + +![](images/ab4df296ca3d6f0d26f985e3ca70a7babb978adfdfccc8ac0c3cb0db402369d0.jpg) + +![](images/eb4c6554ad13a1f7f18bb111117e4cfd1bacc8bca639b0232610f7c7a960aeee.jpg) +Image + +![](images/537e606bc7a6af9eca68fb32645269722e0bab1228f18f9a5d0597fdf6175c2d.jpg) +TCRT w/o CN + +![](images/8500d96e1912d6f5b999616ba8fbfbab8da4635b66ed0ad58c23835d56f2d249.jpg) +TCRT + +![](images/ba8020222ad4d5beb663cc58af574dae7d2b975f18f660d8c0e21de70accb3ed.jpg) +"white laptop keyboard" + +![](images/47ac7ca98d42a6427ecf55ddb5143914c4ff5c50eb31d89023dee28c834bd1b5.jpg) + +![](images/a50145551878df435be89db6073004bdf97a985133516c6339d31c453bf96b7e.jpg) + +![](images/3d99568173df8748042329cb6644c793bd27e2f82d9c738a370f0295c1b74339.jpg) +"person half shown on far left" +"Guy lower right side facing away" +Figure 6. Visualization of features extracted by the visual backbone. The second column shows the results of 'TCRT w/o CN', while the third column shows the results of TCRT. Red and white boxes are the predicted regions and the ground truth, respectively. + +![](images/14c1e7263b24a01e6c0bad7c318df54a295376faa762cb92ef0921029d9be390.jpg) + +![](images/430e80c12c761bb581589d32754d356fcd6060ba2418418c756570868c86fc58.jpg) + +# 4.2. Comparisons with State-of-the-arts + +To evaluate the effectiveness of the proposed TCRT, we perform both qualitative and quantitative experiments across four datasets: RefCOCO, RefCOCOg, RefCOCO+, and ReferItGame. + +# 4.2.1. REC task + +We evaluate our TCRT against existing state-of-the-art models for the REC task. As shown in Table 1, TCRT achieves the best performance across these datasets. Specifically, compared to Vista-7B, TCRT shows absolute improvements of $+2.97$ , $+1.35$ , and $+3.24$ percentage points on RefCOCO; $+1.65$ , $+0.46$ , and $+1.79$ percentage points on RefCOCO+; and $+0.81$ and $+0.73$ percentage points on RefCOCOg. We also present qualitative results from three + +![](images/bfd3dd46061ca25f22683c2161b057e7714c41c8b698cd261614785d77d42d12.jpg) +Figure 7. Comparison between the TCRT without object-text matching loss $(\mathcal{L}_{\mathrm{match}})$ and TCRT in terms of $\mathcal{L}_{\mathrm{det}}$ + +examples of the RefCOCO testB set for the REC task in Figure 4. We can see that TCRT performs well even when an image contains multiple similar objects, such as "second vase from front". Furthermore, we visualized the attention scores within the transformer. We observe that TCRT focuses better on the target object compared to other methods. + +# 4.2.2. RES task + +We evaluate the effectiveness of our TCRT for the RES task by comparing it with existing state-of-the-art methods. The comparative results are provided in Table 2. Compared with Vista-7B, TCRT achieves higher mIoU scores with absolute improvements of $+2.62$ , $+2.12$ , and $+1.77$ percentage points on RefCOCO, $+0.99$ , $+0.56$ , and $+0.93$ percentage points on RefCOCO+, $+1.04$ and $+1.18$ percentage points on RefCOCOg. We also present a comparison for the RES task in Figure 5. + +# 4.3. Analysis of Main Components + +# 4.3.1. Conditional Normalization + +We first conduct an experiment to demonstrate the effectiveness of conditional normalization (CN). We visualize the features extracted by the visual backbone on RefCOCOg test set. As shown in Figure 6, 'TCRT w/o CN' has more dispersed attention, which may lead to inaccurate predictions. Due to the smaller domain discrepancy of TCRT, this allows the LLM to focus more attention on the target objects. + +# 4.3.2. Object-text Semantic Matching Regularization + +We use the object-text matching loss $(\mathcal{L}_{\mathrm{match}})$ to associate the visual tokens with the textual ones. To validate the effectiveness of $\mathcal{L}_{\mathrm{match}}$ we compare 'TCRT w/o $\mathcal{L}_{\mathrm{match}}$ and TCRT in terms of the detection loss $(\mathcal{L}_{\mathrm{det}})$ on RefCOCO val set. In Figure 7, by plotting the loss curves of the models during training, we observe that TCRT achieves lower loss values, which is attributed to its capability to learn semantic-discriminative visual features. This enables the LLM to more easily recognize task-relevant visual features. + +![](images/bd67a5d4df4a948bc5176529532f5e9d0e03741c989aa5466dcbd92db0569c33.jpg) +Figure 8. visualization of LLM-guided visual feature refinement. We visualize the weight matrices $\mathcal{M}$ of each layer in the fusion module to validate the role of the LLM in refining visual features. Red and white boxes are the predicted regions and the ground truth, respectively. + +![](images/30fdeff73b8690ec6b7c831502cacf7057cc0e7a354a3efc9fd3a424e9f1311f.jpg) +Figure 9. Analysis of Visual Feature Distribution. + +# 4.3.3. LLM-guided Visual Feature Refinement + +We visualize the weight matrix $\mathcal{M}$ in Figure 8 to validate the effectiveness of visual feature refinement on RefCOCOg test set. We observe that the attention of the weight matrix is relatively evenly distributed in the earlier layers. In the later layers, the attention of the weight matrix is primarily focused on the target object. This indicates that LLM's prior knowledge enhances our model's ability to locate target objects by further refining the visual features. + +# 4.3.4. Analysis of Visual Feature Distribution + +To validate the effectiveness of our model in aligning visual and textual features, we extract the visual features after fusion with textual information and study their distribution in the spatial domain. Specifically, we label four target objects in an image and use three different textual descriptions for each object, resulting in twelve data samples. These samples are fed into the grounding model, and the visual + +
MethodvalRefCOCO
testAtestB
TCRT w/o VF & VA87.6288.6084.04
TCRT w/o VF88.3890.2485.57
TCRT91.0792.8586.24
+ +Table 3. Ablation experiments on RefCOCO to evaluate the proposed LLM adaptation module $\left( {V}_{A}\right)$ and task-aware cross-modal feature fusion module $\left( {V}_{F}\right)$ . + +features from the fusion module's output are subjected to dimensionality reduction and visualization. As shown in Figure 9, we observe that our method causes the visual features describing the same object to cluster together. + +# 4.3.5. Ablation Studies + +We validate the design of our proposed TCRT by conducting ablative experiments on major components in the REC task. The results are presented in Table 3. We observe that when the LLM adaptation module and task-aware cross-modal feature fusion module are not used, 'TCRT w/o $V_{F}$ & $V_{A}$ ' achieves accuracy rates of $87.62\%$ , $88.60\%$ , and $84.04\%$ on the RefCOCO dataset. After incorporating the LLM adaptation module, the accuracy of 'TCRT w/o $V_{F}$ ' increases by 0.76, 1.64, and 1.53 absolute percentage points. When the $V_{F}$ module is further introduced, TCRT achieves accuracy rates of $91.07\%$ , $92.85\%$ , and $86.24\%$ . + +# 5. Conclusion + +In this paper, we propose a task-aware cross-modal feature refinement transformer with large language models to facilitate visual grounding. To adapt LLMs to the grounding task, we design an adaptation module that modulates the statistical characteristics of visual features conditioned on textual features along with object-text semantic matching regularization. An LLM is employed to integrate the aligned information from textual and visual features and produce task-aware embeddings. We further incorporate a task-aware cross-modal feature fusion module, allowing the resulting embeddings to refine visual features and enable information interaction between the REC and RES tasks. Comprehensive experimental results verify the effectiveness of our proposed approach both quantitatively and qualitatively. + +# 6. Acknowledgments + +This work was supported in part by TCL Science and Technology Innovation Fund (Project No. 20231752), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), and in part by the GuangDong Basic and Applied Basic Research Foundation (Project No. 2024A1515011437). + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. +[2] Mohit Bajaj, Lanjun Wang, and Leonid Sigal. G3raphground: Graph-based language grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4281-4290, 2019. 2 +[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3 +[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2, 3 +[5] Kan Chen, Rama Kovvuri, and Ram Nevatia. Query-guided regression network with context policy for phrase grounding. In Proceedings of the IEEE International Conference on Computer Vision, pages 824-832, 2017. 2 +[6] Long Chen, Wenbo Ma, Jun Xiao, Hanwang Zhang, and Shih-Fu Chang. Ref-nms: Breaking proposal bottlenecks in two-stage referring expression grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1036-1044, 2021. 2, 6 +[7] Zesen Cheng, Kehan Li, Peng Jin, Siheng Li, Xiangyang Ji, Li Yuan, Chang Liu, and Jie Chen. Parallel vertex diffusion for unified visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1326-1334, 2024. 2, 6 +[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 3 +[9] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1769-1779, 2021. 1, 2 +[10] Jiajun Deng, Zhengyuan Yang, Daqing Liu, Tianlang Chen, Wengang Zhou, Yanyong Zhang, Houqiang Li, and Wanli Ouyang. Transvg++: End-to-end visual grounding with language conditioned vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2 +[11] Pelin Dogan, Leonid Sigal, and Markus Gross. Neural sequential phrase grounding (seqground). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4175-4184, 2019. 2 +[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 6 + +[13] Hugo Jair Escalante, Carlos A Hernández, Jesus A Gonzalez, Aurelio López-López, Manuel Montes, Eduardo F Morales, L Enrique Sucar, Luis Villasenor, and Michael Grubinger. The segmented and annotated iapr tc-12 benchmark. Computer Vision and Image Understanding, 114(4): 419-428, 2010. 6 +[14] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3 +[15] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 1440-1448, 2015. 2, 5 +[16] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 580-587, 2014. 2 +[17] Chih-Hui Ho, Srikar Appalaraju, Bhavan Jasani, R Manmatha, and Nuno Vasconcelos. Yoro-lightweight end to end visual grounding. In European Conference on Computer Vision, pages 3-23. Springer, 2022. 1 +[18] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1115-1124, 2017. 2 +[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 3 +[20] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 787-798, 2014. 6 +[21] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv e-prints, pages arXiv-2211, 2022. 3 +[22] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in Neural Information Processing Systems, 34:19652-19664, 2021. 1, 2, 6 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 5 +[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. + +Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1, 2 +[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2 +[27] Xihui Liu, Zihao Wang, Jing Shao, Xiaogang Wang, and Hongsheng Li. Improving referring expression grounding with cross-modal attention-guided erasing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1950-1959, 2019. 2, 6 +[28] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10034-10043, 2020. 6 +[29] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. In European Conference on Computer Vision, pages 417-435. Springer, 2025. 2 +[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11-20, 2016. 2 +[31] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 Fourth International Conference on 3D Vision (3DV), pages 565-571. IEEE, 2016. 5 +[32] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 2 +[33] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6 +[34] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. 2, 6 +[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 6 +[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region + +proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 1, 2 +[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 5 +[38] Wei Su, Peihan Miao, Huanzhang Dou, Gaoang Wang, Liang Qiao, Zheyang Li, and Xi Li. Language adaptive weight generation for multi-task visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10857-10866, 2023. 1, 2, 6 +[39] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities, 2023. +[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3 +[43] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. +[44] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4145-4154, 2019. 2 +[45] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 1 +[46] Sibei Yang, Guanbin Li, and Yizhou Yu. Graph-structured referring expression reasoning in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9952–9961, 2020. 2 +[47] Zhengyuan Yang, Boqing Gong, Liwei Wang, Wenbing Huang, Dong Yu, and Jiebo Luo. A fast and accurate one-stage approach to visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4683-4693, 2019. 2, 6 +[48] Zhengyuan Yang, Tianlang Chen, Liwei Wang, and Jiebo Luo. Improving one-stage visual grounding by recursive subquery construction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 387-404. Springer, 2020. 2, 6 +[49] Jiabo Ye, Junfeng Tian, Ming Yan, Xiaoshan Yang, Xuwu Wang, Ji Zhang, Liang He, and Xin Lin. Shifting more + +attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15502-15512, 2022. 1, 2 +[50] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 6 +[51] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6 +[52] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 2, 6 +[53] Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. In European Conference on Computer Vision, pages 405-422. Springer, 2025. 2, 3 +[54] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An lmm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 3 +[55] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4158-4166, 2018. 2 +[56] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024. +[57] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3 +[58] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. Advances in Neural Information Processing Systems, 37:71737-71767, 2024. 2 +[59] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision, pages 598-615. Springer, 2022. 1, 6 \ No newline at end of file diff --git a/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/images.zip b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4c062e98ad6f5133730178b9b28d5cfaa091ae69 --- /dev/null +++ b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eff142db81c8b97508fe1678d500831ed6316870226d4bed2a07d0c25719d56a +size 642569 diff --git a/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/layout.json b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..121052ba89ad5a52681ad9d045ab1ad298dbf7ac --- /dev/null +++ b/2025/Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding/layout.json @@ -0,0 +1,12213 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 140 + ], + "type": "text", + "content": "Task-aware Cross-modal Feature Refinement Transformer with Large Language Models for Visual Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 160, + 466, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 160, + 466, + 175 + ], + "spans": [ + { + "bbox": [ + 143, + 160, + 466, + 175 + ], + "type": "text", + "content": "Wenbo Chen1*; Zhen Xu1,2*; Ruotao Xu2, Si Wu1,2†; Hau-San Wong3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 175, + 512, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 175, + 512, + 203 + ], + "spans": [ + { + "bbox": [ + 97, + 175, + 512, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 175, + 512, + 203 + ], + "type": "text", + "content": "School of Computer Science and Engineering, South China University of Technology " + }, + { + "bbox": [ + 97, + 175, + 512, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 175, + 512, + 203 + ], + "type": "text", + "content": "Institute for Super Robotics (Huangpu)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 203, + 465, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 203, + 465, + 217 + ], + "spans": [ + { + "bbox": [ + 145, + 203, + 465, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 203, + 465, + 217 + ], + "type": "text", + "content": "Department of Computer Science, City University of Hong Kong" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 219, + 537, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 219, + 537, + 232 + ], + "spans": [ + { + "bbox": [ + 69, + 219, + 537, + 232 + ], + "type": "text", + "content": "{cscwb106, csxuzhen}@mail.scut.edu.cn, {xrt, cswusi}@scut.edu.cn, cshswong@cityu.edu.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 284, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 571 + ], + "type": "text", + "content": "The goal of visual grounding is to establish connections between target objects and textual descriptions. Large Language Models (LLMs) have demonstrated strong comprehension abilities across a variety of visual tasks. To establish precise associations between the text and the corresponding visual region, we propose a Task-aware Cross-modal feature Refinement Transformer with LLMs for visual grounding, and our model is referred to as TCRT. To enable the LLM trained solely on text to understand images, we introduce an LLM adaptation module that extracts text-related visual features to bridge the domain discrepancy between the textual and visual modalities. We feed the text and visual features into the LLM to obtain task-aware priors. To enable the priors to guide the feature fusion process, we further incorporate a cross-modal feature fusion module, which allows task-aware embeddings to refine visual features and facilitate information interaction between the Referring Expression Comprehension (REC) and Referring Expression Segmentation (RES) tasks. We have performed extensive experiments to verify the effectiveness of the main components and demonstrate the superior performance of the proposed TCRT over state-of-the-art end-to-end visual grounding methods on RefCOCO, RefCOCOg, RefCOCO+ and ReferItGame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 613, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 686 + ], + "type": "text", + "content": "Visual grounding aims to predict the locations of the objects specified by textual prompts in images, and is widely applied in the field of human-computer interaction. This task typically involves referring expression comprehension [10, 17, 49, 59] and referring expression segmentation [22, 38]. The main challenge is to comprehensively un" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 266, + 377, + 309 + ], + "blocks": [ + { + "bbox": [ + 358, + 258, + 396, + 265 + ], + "lines": [ + { + "bbox": [ + 358, + 258, + 396, + 265 + ], + "spans": [ + { + "bbox": [ + 358, + 258, + 396, + 265 + ], + "type": "text", + "content": "TCRT w/o LLM" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 266, + 377, + 309 + ], + "lines": [ + { + "bbox": [ + 320, + 266, + 377, + 309 + ], + "spans": [ + { + "bbox": [ + 320, + 266, + 377, + 309 + ], + "type": "image", + "image_path": "f4c1f36b10a23daa8ce5df93c45225be6adc8c0b185ac6ecb1488ce8e50cb48f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 309, + 377, + 312 + ], + "lines": [ + { + "bbox": [ + 320, + 309, + 377, + 312 + ], + "spans": [ + { + "bbox": [ + 320, + 309, + 377, + 312 + ], + "type": "text", + "content": "\"bride throwing bouquet" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 312, + 377, + 355 + ], + "blocks": [ + { + "bbox": [ + 320, + 312, + 377, + 355 + ], + "lines": [ + { + "bbox": [ + 320, + 312, + 377, + 355 + ], + "spans": [ + { + "bbox": [ + 320, + 312, + 377, + 355 + ], + "type": "image", + "image_path": "838a2900a098502cd3251275b99d6c3552bebf15354bec8bda1f28a4a0faae6e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 355, + 382, + 361 + ], + "lines": [ + { + "bbox": [ + 320, + 355, + 382, + 361 + ], + "spans": [ + { + "bbox": [ + 320, + 355, + 382, + 361 + ], + "type": "text", + "content": "\"Guy front in gray suit left\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 320, + 361, + 377, + 404 + ], + "blocks": [ + { + "bbox": [ + 320, + 361, + 377, + 404 + ], + "lines": [ + { + "bbox": [ + 320, + 361, + 377, + 404 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 377, + 404 + ], + "type": "image", + "image_path": "44aa74d849d420267f520734e534fa16bcd1bc4138d5ced316d320dc07448cca.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 405, + 441, + 411 + ], + "lines": [ + { + "bbox": [ + 323, + 405, + 441, + 411 + ], + "spans": [ + { + "bbox": [ + 323, + 405, + 441, + 411 + ], + "type": "text", + "content": "\"man in right front with blue tie next to older woman" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 422, + 555, + 465 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 465 + ], + "type": "text", + "content": "Figure 1. The visualization of the refined visual features from the cross-modal feature fusion module. In challenging environments, the introduction of LLMs can facilitate a better association between visual features and text." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 436, + 266, + 492, + 308 + ], + "blocks": [ + { + "bbox": [ + 486, + 258, + 501, + 264 + ], + "lines": [ + { + "bbox": [ + 486, + 258, + 501, + 264 + ], + "spans": [ + { + "bbox": [ + 486, + 258, + 501, + 264 + ], + "type": "text", + "content": "TCRT" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 266, + 492, + 308 + ], + "lines": [ + { + "bbox": [ + 436, + 266, + 492, + 308 + ], + "spans": [ + { + "bbox": [ + 436, + 266, + 492, + 308 + ], + "type": "image", + "image_path": "e38bddff4d03ab149e64c193388b3c560b3621c1a333227c8d25cf4b29b648ef.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 436, + 314, + 492, + 356 + ], + "blocks": [ + { + "bbox": [ + 436, + 314, + 492, + 356 + ], + "lines": [ + { + "bbox": [ + 436, + 314, + 492, + 356 + ], + "spans": [ + { + "bbox": [ + 436, + 314, + 492, + 356 + ], + "type": "image", + "image_path": "d0962ec6767962463598628d3cc48dd6dd4c8707d784c10251d85cf4c2043000.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 436, + 361, + 492, + 404 + ], + "blocks": [ + { + "bbox": [ + 436, + 361, + 492, + 404 + ], + "lines": [ + { + "bbox": [ + 436, + 361, + 492, + 404 + ], + "spans": [ + { + "bbox": [ + 436, + 361, + 492, + 404 + ], + "type": "image", + "image_path": "83e68a042ef610ba44bcc52c4e667e60436031a0f2272c2ac6a441ab209a27ff.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 406, + 492, + 412 + ], + "lines": [ + { + "bbox": [ + 447, + 406, + 492, + 412 + ], + "spans": [ + { + "bbox": [ + 447, + 406, + 492, + 412 + ], + "type": "text", + "content": "predicted by" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 493, + 266, + 549, + 308 + ], + "blocks": [ + { + "bbox": [ + 493, + 266, + 549, + 308 + ], + "lines": [ + { + "bbox": [ + 493, + 266, + 549, + 308 + ], + "spans": [ + { + "bbox": [ + 493, + 266, + 549, + 308 + ], + "type": "image", + "image_path": "e1319efd6665c724e37d17b0c6b12437b265ab2ad1953f9b183f6914b1df72cd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 493, + 314, + 549, + 356 + ], + "blocks": [ + { + "bbox": [ + 493, + 314, + 549, + 356 + ], + "lines": [ + { + "bbox": [ + 493, + 314, + 549, + 356 + ], + "spans": [ + { + "bbox": [ + 493, + 314, + 549, + 356 + ], + "type": "image", + "image_path": "974bc6bdd894f345e634f5c8d15cbe248c63ba37df684f639d90a33d281e066e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 493, + 361, + 549, + 404 + ], + "blocks": [ + { + "bbox": [ + 493, + 361, + 549, + 404 + ], + "lines": [ + { + "bbox": [ + 493, + 361, + 549, + 404 + ], + "spans": [ + { + "bbox": [ + 493, + 361, + 549, + 404 + ], + "type": "image", + "image_path": "cc54ee27f0edb5b8c64c46f960933757efa8845894bb5673567acd8d0319ae5e.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 406, + 549, + 412 + ], + "lines": [ + { + "bbox": [ + 501, + 406, + 549, + 412 + ], + "spans": [ + { + "bbox": [ + 501, + 406, + 549, + 412 + ], + "type": "text", + "content": "ground truth" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 495, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 519 + ], + "type": "text", + "content": "derstand intra-modality information and accurately establish inter-modality correspondences." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "Existing visual grounding methods can be categorized into two main types: two-stage and end-to-end methods. The two-stage methods [45] first generate region proposals [36] that may match the text, and then rank the candidate regions according to their relevance to the text. When the proposals generated in the first stage do not contain the target object, the second stage will be unable to correct these proposals. On the other hand, end-to-end grounding methods perform the integration of visual and textual features within an object detector, producing the box coordinates of the target object. The end-to-end methods are typically faster than the two-stage ones. With the advancement of transformer in the vision domain, TransVG [9, 10] and SeqTR [59] employ self-attention or cross-attention mechanisms to fuse textual and visual features. Recently, LLaVA [25] has integrated visual understanding capabilities into large language mod" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 130, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 130, + 702 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 130, + 702 + ], + "type": "text", + "content": "*Joint first authors." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3931" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "els (LLMs). MiniGPT-2 [4] and OMG-LLaVA [58] apply LLMs to visual grounding tasks by integrating textual and visual features. In this work, we make an attempt to adapt LLMs to capture the information of target objects and produce task-aware embeddings for precise grounding results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 134, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 134, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 295, + 384 + ], + "type": "text", + "content": "More specifically, we propose a task-aware cross-modal feature refinement transformer with large language models for visual grounding, which is referred to as TCRT, leveraging priors from LLMs to guide the multimodal information fusion process and achieve precise alignment between textual and local visual features (see Figure 1). To enable the LLM to more effectively process image data and adapt to visual grounding tasks, we propose an LLM adaptation module that utilizes conditional normalization to adjust the statistical characteristics of visual features based on textual information, while introducing object-text semantic matching regularization to associate local visual features with textual features. Additionally, to utilize the priors provided by the adaptation module, we incorporate a task-aware cross-modal feature fusion module, which continuously refines visual features during multimodal fusion and enables the model to exploit the relationship between the REC and RES tasks. We perform extensive qualitative and quantitative experiments to validate the effectiveness of our model, and demonstrate its advantages on benchmark datasets. The main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 387, + 293, + 542 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 67, + 387, + 293, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 387, + 293, + 446 + ], + "spans": [ + { + "bbox": [ + 67, + 387, + 293, + 446 + ], + "type": "text", + "content": "- Different from the state-of-the-art methods that focus on the application of LLMs in integrating textual and visual features, we further explore the LLM's capability of learning task-aware embeddings to facilitate visual grounding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 447, + 293, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 447, + 293, + 493 + ], + "spans": [ + { + "bbox": [ + 67, + 447, + 293, + 493 + ], + "type": "text", + "content": "- We incorporate an LLM adaptation module for visual grounding, which learns to capture the information of target objects by modulating the statistical characteristics of visual features conditioned on textual features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 495, + 293, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 495, + 293, + 542 + ], + "spans": [ + { + "bbox": [ + 67, + 495, + 293, + 542 + ], + "type": "text", + "content": "- We further perform task-aware cross-modal feature fusion, allowing the task-aware embeddings derived from LLMs to guide the association of textual and visual information and lead to precise grounding results." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 555, + 142, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 142, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 142, + 567 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 212, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 212, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 212, + 588 + ], + "type": "text", + "content": "2.1. Two-stage Visual Grounding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "In two-stage methods, a region proposal network [15, 16, 36] is first used to produce a set of candidate regions that may correspond to the target object mentioned in the textual description. These candidate regions are then matched to the textual description, and the predicted result is determined by identifying the region with the highest score. The maximum-margin ranking loss [30] is widely adopted for the second stage. Several researches further enhance the two-stage methods by incorporating context modeling [18, 32, 44, 55] and making use of phrase cooccur" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "ences [2, 5, 11]. MAttNet [52] decomposed the expression into three modules: relationship, location, and subject to improve the grounding accuracy. To enhance the model's capacity for learning detailed alignments between text and visuals, CM-A-E [27] proposed a method capable of generating challenging samples for training the model. Graph learning was incorporated into visual grounding by SGMN [46] to enable the model to better capture the relationships between different objects within the same scene. To increase the recall of target objects, Ref-NMS [6] utilized textual features to direct the non-maximum suppression. To enhance the model's cross-modality fusion capabilities, GroundingDINO [26] integrated grounded pretraining techniques into detection transformers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 249, + 477, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 477, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 477, + 262 + ], + "type": "text", + "content": "2.2. End-to-end Visual Grounding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 266, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 506 + ], + "type": "text", + "content": "End-to-end methods in visual grounding aim to directly predict the bounding boxes in an image from a text in a single pass, resulting in faster inference speeds compared to two-stage methods. As a pioneering work, FAOA [47] concatenated image and textual features, and fed the concatenated features into a prediction layer to achieve target object detection. To reduce the referring ambiguity, ReSC [48] used a sub-query construction to address limitations in processing complex textual descriptions. TransVG [9, 10] is one of the pioneering transformer-based visual grounding models that directly regresses coordinates. To achieve multi-task learning, RefTR [22] used the transformer architecture to integrate REC and RES tasks into a single framework. To address the gap between the visual and textual backbones, QRNet [49] incorporated textual information into the visual backbone. Furthermore, VG-LAW [38] actively extracted text-relevant visual features based on text-adaptive weights and performed multi-task visual grounding. Additionally, PVD [7] introduced a diffusion model into the visual grounding task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 662 + ], + "type": "text", + "content": "With the development of LLMs, some Multimodal Large Language Models [25] (MLLMs) are gradually being applied to visual grounding tasks. VistaLLM [34] used an instruction-guided image tokenizer to filter and compress image embeddings based on task, extracting visual information relevant to the current task. To efficiently locate and understand multiple target regions in complex scenes, Groma [29] introduced a localized visual tokenization mechanism that decomposes an image into multiple regions of interest, encoding each region as an independent region token. To unleash the potential of large language models, Griffon [53] introduces a foundational language-prompted localization dataset with nearly 6 million pretraining samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 671, + 451, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 451, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 451, + 684 + ], + "type": "text", + "content": "2.3. Large Language Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "With the expansion of training data and a significant increase in model parameters, large language models (LLMs)" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3932" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 553, + 208 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 553, + 208 + ], + "type": "image", + "image_path": "976b3e3cd8eb1c19697c61cabd6b556ac4756f0ce8920a55e449140240bc4f18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed method. We first extract textual features " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "F_{txt}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": " using the textual backbone " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": ". These features are then injected into the visual backbone " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "E_{vis}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": ", which consists of transformer blocks and conditional normalization, to extract text-related visual features " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "F_{vis}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": ". The object-text semantic matching regularization is employed to associate visual region tokens with textual tokens. We input the visual features processed by the visual adapter " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "V_{proj}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": ", along with the text, into LLaMa. We take the last token " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "f_{llm}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": " output from LLaMa. " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "f_{llm}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": " is then fed into a query projection layer to generate task-relevant tokens " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "f_{box}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "inline_equation", + "content": "f_{phr}" + }, + { + "bbox": [ + 55, + 220, + 555, + 308 + ], + "type": "text", + "content": ", which are used to predict the bounding box, segmentation, and phrase, respectively. A task-aware cross-modal feature fusion module is then employed to integrate textual, visual, and task features. Finally, task-relevant tokens are each fed to the corresponding prediction heads to produce the final predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 329, + 294, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 294, + 460 + ], + "type": "text", + "content": "based on the Transformer [42] architecture have developed rapidly. GPT-3 [3], with its 175 billion parameters, demonstrates strong performance in few-shot contextual learning. Following the success of GPT-3, several other LLMs, including PaLM [8], OPT [57], BLOOM [21], and LLaMA [40, 41], have emerged in succession. Mistral [19] introduced a specialized window attention mechanism to better handle extended context modeling. In parallel, Mixtral " + }, + { + "bbox": [ + 54, + 329, + 294, + 460 + ], + "type": "inline_equation", + "content": "8 \\times 7" + }, + { + "bbox": [ + 54, + 329, + 294, + 460 + ], + "type": "text", + "content": "B utilized sparse MoE layers [14] to scale up parameters more effectively, enabling it to achieve higher performance with a reduced number of active parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "content": "The main differences between the proposed TCRT and existing visual grounding methods are summarized as follows: First, we perform statistical modulation on visual features conditioned on textural features to reduce domain discrepancy before integrating them, while the existing methods [4, 53, 54] only perform dimension adaption. Second, TCRT leverage the task-aware embeddings derived from LLMs to guide cross-modal feature fusion and association analysis, rather than directly predicting the bounding boxes of target objects as in previous methods [4, 53, 54]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 592, + 138, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 138, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 138, + 605 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 153, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 153, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 153, + 624 + ], + "type": "text", + "content": "3.1. Problem Setting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": "In this section, we begin by defining multi-task visual grounding (MTVG), and then proceed to describe the architecture of the proposed framework. The goal of MTVG is to detect or segment a target object based on a textual description. Formally, given an image " + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": " and a corresponding text " + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": ", this task is to predict a bounding box " + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "b = \\{x, y, w, h\\}" + }, + { + "bbox": [ + 54, + 629, + 295, + 713 + ], + "type": "text", + "content": " that encompasses the object men" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 329, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 389 + ], + "type": "text", + "content": "tioned in the text, as well as a pixel-wise segmentation mask " + }, + { + "bbox": [ + 313, + 329, + 555, + 389 + ], + "type": "inline_equation", + "content": "s\\in \\mathbb{R}^{1\\times H\\times W}" + }, + { + "bbox": [ + 313, + 329, + 555, + 389 + ], + "type": "text", + "content": " that provides a more detailed and precise localization. A textual backbone " + }, + { + "bbox": [ + 313, + 329, + 555, + 389 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 313, + 329, + 555, + 389 + ], + "type": "text", + "content": " and a visual backbone " + }, + { + "bbox": [ + 313, + 329, + 555, + 389 + ], + "type": "inline_equation", + "content": "E_{vis}" + }, + { + "bbox": [ + 313, + 329, + 555, + 389 + ], + "type": "text", + "content": " are used to extract textual and visual features, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 397, + 417, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 417, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 417, + 410 + ], + "type": "text", + "content": "3.2. LLM Adaptation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": "Pre-trained LLMs possess language understanding capabilities and can handle many complex tasks in natural language processing. However, due to the domain discrepancy between textual data and visual data, directly applying LLMs to grounding tasks may yield suboptimal performance. Therefore, we propose an LLM adaptation module to perform cross-modal adaptation of image information, enabling LLMs to better handle information within visual data. As shown in Figure 2, we first utilize the text encoding model " + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "inline_equation", + "content": "E_{txt}" + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": " to extract the textual features of " + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "inline_equation", + "content": "F_{txt} = E_{txt}(q)" + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "inline_equation", + "content": "F_{txt}" + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": " represents the textual features, which include three types of textual features: sentence feature " + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "inline_equation", + "content": "f_{txt}^{sen} \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": ", individual word features " + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "inline_equation", + "content": "f_{txt} \\in \\mathbb{R}^{L \\times d}" + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": ", and average pooled word feature " + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "inline_equation", + "content": "f_{txt}^{mean} \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": " represent the dimension of textual features and the number of tokens, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "text", + "content": "To enable the visual encoder to comprehend textual information, we designed a text-aware visual encoder " + }, + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "inline_equation", + "content": "E_{vis}" + }, + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "text", + "content": ". Specifically, a conditioned normalization is inserted after each transformer block in the visual encoder. In the conditioned normalization, we fuse the sentence features " + }, + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "inline_equation", + "content": "f_{txt}^{sen}" + }, + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "text", + "content": " and the average-pooled word features " + }, + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "inline_equation", + "content": "f_{txt}^{mean}" + }, + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "text", + "content": " into a textual conditional feature " + }, + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "inline_equation", + "content": "f_{txt}^{cond}" + }, + { + "bbox": [ + 313, + 607, + 554, + 691 + ], + "type": "text", + "content": " via a weighted combination:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 352, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 352, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "f _ {t x t} ^ {\\text {c o n d}} = \\alpha^ {\\text {s e n}} f _ {t x t} ^ {\\text {s e n}} + \\alpha^ {\\text {m e a n}} \\phi \\left(f _ {t x t} ^ {\\text {m e a n}}\\right), \\tag {1}", + "image_path": "698058e6d9681d473b48db69dc96c7edd049626210384b0c15321bf824ffec5c.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3933" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 294, + 175 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 294, + 175 + ], + "type": "image", + "image_path": "77472a9cf2b8779aff75b70d7bf29a59104fe024f03c5632df6d65d920c0bea4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 296, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 240 + ], + "type": "text", + "content": "Figure 3. Object-text semantic matching regularization. We constrain the image features using a mask to isolate the object-specific features. For matching text-object pairs, we consider them as positive samples, whereas non-matching pairs are regarded as negative samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\alpha^{sen}" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\alpha^{mean}" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " are learnable parameters that control the contribution of " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "f_{txt}^{sen}" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "f_{txt}^{mean}" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " represents the feed-forward network. The scaling " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " and shifting " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " parameters are determined using " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "f_{txt}^{cond}:\\gamma = \\phi(f_{txt}^{cond}) + \\gamma^{learn}" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\beta = \\phi(f_{txt}^{cond}) + \\beta^{learn}" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\gamma^{learn}" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\beta^{learn}" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " are learnable parameters. Given the visual features " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\bar{F}_i^*" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": ", which is the output of the " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": "-th layer of transformer block, we calculate the mean " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mu(\\bar{F}_i^*)" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " and standard deviation " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\sigma(\\bar{F}_i^*)" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": ". The visual features " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\bar{F}_i^*" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " are then modulated using the resulting scaling " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " and shifting " + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 259, + 296, + 380 + ], + "type": "text", + "content": " parameters, as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 387, + 295, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 387, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 110, + 387, + 295, + 420 + ], + "type": "interline_equation", + "content": "\\bar {F} _ {i} ^ {*} = \\gamma \\left(\\frac {\\bar {F} _ {i} ^ {*} - \\mu (\\bar {F} _ {i} ^ {*})}{\\sigma (\\bar {F} _ {i} ^ {*}) + \\epsilon}\\right) + \\beta , \\tag {2}", + "image_path": "3d743253f17176a2085801887766c3d7604c6b5e6976644518d7c72cd704f29c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " is a small constant introduced to avoid division by zero. We obtain the text-aware visual features " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "F_{vis}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " by flattening the feature map " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\overline{F}_4^*" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " extracted from the visual backbone. The image features and the input embeddings of LLMs are typically in different dimensional spaces. We use a visual adapter " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "V_{proj}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " to map the image features to a dimension that matches the LLM input, allowing the LLM to directly process visual information. " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "V_{proj}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " is a three-layer MLP. We feed the text and the mapped image features into the LLM, and take the last token of the LLM output as the prior embedding " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "f_{llm}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 567, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 567, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 99, + 567, + 295, + 597 + ], + "type": "interline_equation", + "content": "f _ {l l m} = \\Phi (q, V _ {p r o j} \\underbrace {\\left(E _ {v i s} (I , f _ {t x t} ^ {c o n d})\\right)} _ {F _ {v i s}}), \\tag {3}", + "image_path": "54a725ad14ef2e0923582f43fa72f185eb8c8957e0ba424cb9574c2df7914867.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " represents an LLM model, " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " denotes the original query text. We pass " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "f_{llm}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " through three task-aware linear layers (" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "W_{box}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "W_{seg}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "W_{phr}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ") to obtain the task-specific embeddings " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "f_{box}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "f_{phr}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ", which are responsible for predicting the bounding box, segmentation, and phrase, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The text usually describes a specific object within an image, which occupies only a portion of the image. Most visual features do not have a direct relationship with the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "text", + "content": "text. This makes it challenging to achieve precise semantic alignment between the target object and textual tokens. Thus, we propose the object-text semantic matching regularization to associate the visual region tokens with the textual ones. Object-text matching is a binary classification task, and the model evaluates the multimodal features of an object-text pair to classify it as either positive (matched) or negative (unmatched). As shown in Figure 3, if the textual features and visual features represent the same object, they are considered a positive sample pair; otherwise, they are considered a negative sample pair. The ground truth of object segmentation is represented by " + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "inline_equation", + "content": "s^+" + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "text", + "content": ". We apply an affine transformation to " + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "inline_equation", + "content": "s^+" + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "inline_equation", + "content": "s^-" + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "inline_equation", + "content": "s^-" + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "text", + "content": " represents the inaccurate segmentation result. We leverage the segmentation mask of the target object to extract its features. A learnable token " + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "inline_equation", + "content": "f_{match}" + }, + { + "bbox": [ + 313, + 72, + 555, + 275 + ], + "type": "text", + "content": " is then introduced to determine whether they are matched:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 340, + 281, + 553, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 281, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 340, + 281, + 553, + 319 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\{\\bar {f} _ {\\text {m a t c h}}, \\bar {F} _ {l}, \\bar {F} _ {v} \\right\\} = \\varphi \\left(\\left[ f _ {\\text {m a t c h}}, F _ {\\text {t x t}}, F _ {\\text {v i s}} \\right]\\right), \\\\ \\left\\{\\hat {f} _ {\\text {m a t c h}} ^ {j}, \\hat {F} _ {l} \\right\\} = \\theta \\left(\\left[ \\bar {f} _ {\\text {m a t c h}}, \\bar {F} _ {l} \\right], F _ {\\text {v i s}} \\odot s ^ {j}\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "c5fd249fa71cd957413f468880e02c9b2f0f5426efa8fce3be5e6a068526842c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "inline_equation", + "content": "j \\in \\{+, -\\}" + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\varphi(q)" + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\theta(q, kv)" + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "text", + "content": " represent multi-head self-attention and cross-attention, respectively. " + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "text", + "content": " denotes element-wise multiplication. " + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "text", + "content": " denotes the concatenation operation. We use the " + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\hat{f}_{match}^{j}" + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "text", + "content": " as the joint representation of the object-text pair. A feed-forward network is employed to predict the two-class probability " + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "inline_equation", + "content": "p_{match}^{j}" + }, + { + "bbox": [ + 313, + 327, + 555, + 425 + ], + "type": "text", + "content": " that indicates whether a textual feature matches a local visual feature. The object-text matching loss is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 350, + 434, + 555, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 434, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 350, + 434, + 555, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a t c h}} = \\mathbb {E} _ {(I, q)} [ \\sum_ {j} H \\left(y _ {\\text {m a t c h}} ^ {j}, p _ {\\text {m a t c h}} ^ {j}\\right) ], \\tag {5}", + "image_path": "710ce5a3910063c2e3c14ad1af57646a65550e7e988f8ed45d0d13836d6298bd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 472, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 472, + 555, + 593 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 472, + 555, + 593 + ], + "type": "text", + "content": " represents cross-entropy and " + }, + { + "bbox": [ + 313, + 472, + 555, + 593 + ], + "type": "inline_equation", + "content": "y_{match}^{j}" + }, + { + "bbox": [ + 313, + 472, + 555, + 593 + ], + "type": "text", + "content": " is a 2-dimensional one-hot vector that represents the ground truth label. We use the visual features of the target object extracted from the ground truth mask as positive sample, indicating that these features match the textual description. When performing affine transformation (partial overlap may occur), the resulting mask deviates from the target object, and the extracted features are considered as negative sample. Object-text matching loss further associates visual tokens with textual ones by assessing whether they match." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 600, + 527, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 527, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 527, + 612 + ], + "type": "text", + "content": "3.3. Task-aware Cross-modal Feature Fusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "When the LLM receives textual and visual information, it can generate more meaningful semantic representations. These representations contain the LLM's understanding of objects and context, forming prior information that aids downstream tasks. The task-aware cross-modal feature fusion module, denoted as " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "V_{F}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ", is incorporated to guide feature the fusion process. Specifically, we first employ a bidirectional multi-head attention module to fuse textual and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3934" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "visual features. Each modality serves as the query while the other modality acts as the key and value in the attention computation. We concatenate the three tokens used for prediction, " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "f_{box}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "f_{phr}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", with the textual features " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "F_{txt}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " to obtain a task-relevant textual feature " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "F_{task}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "F_{task}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "F_{vis}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " are fed into the bi-directional multi-head attention module to fuse the two types of features. This process can be represented as follows: " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "F_{task}^{v2t} = \\theta (F_{task}, F_{vis})" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "F_{vis}^{t2v} = \\theta (F_{vis}, F_{task})" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "F_{task}^{v2t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " represents the textual feature enriched with image features, comprising " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "F_{txt}^{v2t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "f_{box}^{v2t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "f_{seg}^{v2t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "f_{phr}^{v2t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". Meanwhile, " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "F_{vis}^{t2v}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " denotes the visual feature enriched with textual information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "text", + "content": "Target objects are typically located in specific regions. To enable the model to focus on local visual features, we input " + }, + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "inline_equation", + "content": "f_{llm}" + }, + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "text", + "content": " into an MLP to obtain the refinement prompt " + }, + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "inline_equation", + "content": "f_{prmt}" + }, + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "text", + "content": ", which guides the feature refinement process. We then assign different weights to image features at different locations by computing the cosine similarity between the image feature " + }, + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "inline_equation", + "content": "F_{vis}^{t2v}" + }, + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "inline_equation", + "content": "f_{prmt}" + }, + { + "bbox": [ + 55, + 217, + 296, + 312 + ], + "type": "text", + "content": ". This process can be formulated as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 320, + 296, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 320, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 59, + 320, + 296, + 354 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {v} ^ {t 2 v} = F _ {v i s} ^ {t 2 v} + F _ {v i s} ^ {t 2 v} \\odot \\underbrace {\\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {v i s} ^ {t 2 v}\\right) \\otimes \\mathrm {M L P} \\left(f _ {l l m}\\right)\\right)} _ {\\mathcal {M}} \\tag {6}", + "image_path": "3fe01b28479b06afad51ac68c7dc31efbc50636e7876e33b8e71f5c7c565f402.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "text", + "content": " denotes matrix multiplication and " + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "text", + "content": " represents element-wise multiplication. The refinement prompt " + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "inline_equation", + "content": "f_{prmt}" + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "text", + "content": " is dot-multiplied with the image features at each position in " + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "inline_equation", + "content": "F_{vis}^{t2v}" + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "text", + "content": " to obtain a weight matrix " + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "text", + "content": ". The final refined visual features " + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\hat{F}_v^{t2v}" + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "text", + "content": " are obtained under the guidance of the LLM. The model can control the strength of the feature representations at different positions in the image by adjusting the weight matrix " + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 364, + 296, + 484 + ], + "type": "text", + "content": ". By assigning lower weights to irrelevant regions, the model can better focus on the features of the target regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 595 + ], + "type": "text", + "content": "There is a close relationship between the REC task and the RES task. Both require an understanding and localization of the target object or region. REC tasks generally offer locational information of the target objects, whereas RES tasks offer precise boundary details. To strengthen the relationship between the REC and RES tasks, we perform cross-task interaction to effectively exchange information between the box token " + }, + { + "bbox": [ + 55, + 486, + 296, + 595 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{box}}^{v2t}" + }, + { + "bbox": [ + 55, + 486, + 296, + 595 + ], + "type": "text", + "content": " and the segmentation token " + }, + { + "bbox": [ + 55, + 486, + 296, + 595 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{seg}}^{v2t}" + }, + { + "bbox": [ + 55, + 486, + 296, + 595 + ], + "type": "text", + "content": ". The process is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 605, + 295, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 605, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 84, + 605, + 295, + 630 + ], + "type": "interline_equation", + "content": "\\left\\{f _ {b o x} ^ {\\prime}, f _ {s e g} ^ {\\prime} \\right\\} = \\theta ([ f _ {b o x} ^ {v 2 t}, f _ {s e g} ^ {v 2 t} ], F _ {t x t} ^ {v 2 t}), \\tag {7}", + "image_path": "7fb50302675ecea33b0f1a4621bf214b6e04b1d2e1c621ce75e35d1e474b2d3e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 628, + 269, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 628, + 269, + 644 + ], + "spans": [ + { + "bbox": [ + 83, + 628, + 269, + 644 + ], + "type": "interline_equation", + "content": "f _ {b o x} ^ {\\prime \\prime} = f _ {b o x} ^ {\\prime} + \\phi (f _ {s e g} ^ {\\prime}), f _ {s e g} ^ {\\prime \\prime} = f _ {s e g} ^ {\\prime} + \\phi (f _ {b o x} ^ {\\prime}),", + "image_path": "edb377fd0e500be2a1af1c198f399f6f1b3f1479f24d635dde3536836c416d95.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "This design allows the model to exploit the correlation between the REC and RES tasks. The updated " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{box}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{seg}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{phrase}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " are each fed into their respective prediction heads to obtain the final bounding box, segmentation, and phrase predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 411, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 411, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 411, + 85 + ], + "type": "text", + "content": "3.4. Model Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "text", + "content": "Detection loss. For the REC task, the loss for bounding box regression is calculated using a combination of GIoU loss and L1 loss. We use " + }, + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "text", + "content": " to denote the ground truth bounding box and " + }, + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "inline_equation", + "content": "\\tilde{b}" + }, + { + "bbox": [ + 313, + 89, + 555, + 149 + ], + "type": "text", + "content": " to represent the predicted bounding box. The following defines the detection loss function:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 154, + 555, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 154, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 341, + 154, + 555, + 170 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\det } = \\mathbb {E} _ {(I, q)} \\left[ \\Phi_ {\\text {s m o o t h} - l 1} (b, \\tilde {b}) + \\Phi_ {\\text {g i o u}} (b, \\tilde {b}) \\right], \\tag {8}", + "image_path": "432fb020fa19cb7521062290c4e680440707e75f7233da9dc49e3870a1e54e73.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{smooth} - l1}" + }, + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{giou}}" + }, + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "text", + "content": " denote the smooth L1 loss [15] and Generalized IoU loss [37], respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 198, + 554, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 554, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 554, + 247 + ], + "type": "text", + "content": "Segmentation loss. For the RES task, the segmentation loss is computed by combining the focal loss with dice loss. The ground truth mask is denoted as " + }, + { + "bbox": [ + 313, + 198, + 554, + 247 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 198, + 554, + 247 + ], + "type": "text", + "content": ", while the predicted mask is " + }, + { + "bbox": [ + 313, + 198, + 554, + 247 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 313, + 198, + 554, + 247 + ], + "type": "text", + "content": ". The following defines the segmentation loss function:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 252, + 555, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 252, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 350, + 252, + 555, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s e g}} = \\mathbb {E} _ {(I, q)} \\left[ \\Phi_ {\\text {f o c a l}} (s, \\hat {s}) + \\Phi_ {\\text {d i c e}} (s, \\hat {s}) \\right], \\tag {9}", + "image_path": "f13899a1b92d1872e9c5a59940aacc2406cdd9d43d7948abab9d0501a1a6797f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 271, + 554, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 554, + 295 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 271, + 554, + 295 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{focal}}" + }, + { + "bbox": [ + 313, + 271, + 554, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 271, + 554, + 295 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{dice}}" + }, + { + "bbox": [ + 313, + 271, + 554, + 295 + ], + "type": "text", + "content": " denote the smooth focal loss [24] and DICE/F-1 loss [31], respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 296, + 554, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 355 + ], + "type": "text", + "content": "Phrase loss. To enable the model to more accurately capture the entities within the text, we utilize " + }, + { + "bbox": [ + 313, + 296, + 554, + 355 + ], + "type": "inline_equation", + "content": "f_{phr}" + }, + { + "bbox": [ + 313, + 296, + 554, + 355 + ], + "type": "text", + "content": " to predict the words in the text related to the target object. We fed " + }, + { + "bbox": [ + 313, + 296, + 554, + 355 + ], + "type": "inline_equation", + "content": "f_{phr}" + }, + { + "bbox": [ + 313, + 296, + 554, + 355 + ], + "type": "text", + "content": " into an MLP to predict a word probability " + }, + { + "bbox": [ + 313, + 296, + 554, + 355 + ], + "type": "inline_equation", + "content": "p_{phr}" + }, + { + "bbox": [ + 313, + 296, + 554, + 355 + ], + "type": "text", + "content": ". Then, we compute the phrase loss as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 361, + 555, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 361, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 372, + 361, + 555, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p h r}} = \\mathbb {E} _ {(I, q)} [ H (y _ {\\text {p h r}}, p _ {\\text {p h r}}) ], \\tag {10}", + "image_path": "4fa32ddf86e014e9870e7db37981d938a21ab75c74b479f17c4eafca7cd7cba1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 380, + 487, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 487, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 487, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 380, + 487, + 392 + ], + "type": "inline_equation", + "content": "y_{phr}" + }, + { + "bbox": [ + 314, + 380, + 487, + 392 + ], + "type": "text", + "content": " denotes the ground truth label." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 393, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 427 + ], + "type": "text", + "content": "By incorporating these training loss functions, the optimization problem of our model can be formulated as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 435, + 554, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 435, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 367, + 435, + 554, + 453 + ], + "type": "interline_equation", + "content": "\\min _ {E _ {t x t}, E _ {v i s}} \\lambda_ {\\text {m a t c h}} \\mathcal {L} _ {\\text {m a t c h}}, \\tag {11}", + "image_path": "3186e77aec53d9a0002dde3cfa3a27d24d6154bdda6e0fcf88d7c44af25281e5.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 338, + 456, + 531, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 456, + 531, + 475 + ], + "spans": [ + { + "bbox": [ + 338, + 456, + 531, + 475 + ], + "type": "interline_equation", + "content": "\\min _ {E _ {t x t}, E _ {v i s}, V _ {p r o j}, W _ {t a s k}, V _ {F}} \\mathcal {L} _ {\\mathrm {d e t}} + \\mathcal {L} _ {\\mathrm {s e g}} + \\lambda_ {\\mathrm {p h r}} \\mathcal {L} _ {\\mathrm {p h r}},", + "image_path": "eac4e20405d5ad00358218d4e7be113bb52892e711105f5a0a67efd07bc0829a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\text{task} \\in \\{box,seg,phr\\}" + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{phr}}" + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "content": " are used to achieve a balance among the training objectives. " + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}" + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "content": " enable the model to perform basic detection and segmentation tasks. " + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{phr}}" + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "content": " enables the model to achieve a more detailed understanding of the target objects referred to in the text. " + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 480, + 555, + 601 + ], + "type": "text", + "content": " is used to associate visual tokens with textual ones and only optimizes the textual backbone and visual backbone. Our LLM adaptation module and task-aware cross-modal feature fusion module are jointly optimized during the training process." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 609, + 395, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 395, + 624 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 395, + 624 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "content": "In this section, we first describe datasets, evaluation metrics, and implementation details. Next, we qualitatively and quantitatively compare our method with state-of-the-art methods. We further perform comprehensive experiments to analyze the main components of TCRT, including conditional normalization, object-text semantic matching regularization, and LLM-guided visual feature refinement." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3935" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 512, + 223 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 512, + 223 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 512, + 223 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 512, + 223 + ], + "type": "table", + "html": "
MethodsVisual BackboneMulti-taskVenueRefCOCORefCOCOgRefCOCO+ReferItGame test
valtestAtestBvaltestvaltestAtestB
Two-stage Methods
MAttNet [52]RN101×CVPR1876.6581.1469.9966.5867.2765.3371.6256.0229.04
CM-A-E [27]RN101×CVPR1978.3583.1471.3267.9968.6768.0973.6558.03-
Ref-NMS [6]RN101×AAAI2180.7084.0076.0470.5570.6268.2573.6859.42-
End-to-end Methods
MCN [28]DN53✓CVPR2080.0882.2974.9866.4666.0167.1672.8657.31-
RefTR [22]RN101✓NeurIPS2182.2385.5976.5769.4169.4071.5875.9662.1671.42
SeqTR [59]DN53✓ECCV2281.2385.0076.0871.3571.5868.8275.3758.7869.66
VG-LAW [38]ViT-B✓CVPR2386.6289.3283.1676.9076.9676.3781.0467.5077.22
PVD [7]Swin-B✓AAAI2484.9988.0280.0374.3474.6474.2779.0665.11-
Ferret-7B [50]ViT-L×ICLR2487.4991.3582.4583.9384.7680.7887.3873.14-
Vista-7B [34]ViT-L✓CVPR2488.191.583.086.086.484.190.375.8-
TCRTSwin-B✓-91.0792.8586.2486.8187.1385.7590.7677.5979.09
", + "image_path": "9d1d53b43963d8fcc810fd254d7ef6d37af83eb285af8f21280a2b7039363548.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 90, + 254, + 520, + 342 + ], + "blocks": [ + { + "bbox": [ + 55, + 231, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 553, + 243 + ], + "type": "text", + "content": "Table 1. Comparison with existing visual grounding methods on RefCOCO, RefCOCOg, RefCOCO+, and ReferItGame for the REC task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 90, + 254, + 520, + 342 + ], + "lines": [ + { + "bbox": [ + 90, + 254, + 520, + 342 + ], + "spans": [ + { + "bbox": [ + 90, + 254, + 520, + 342 + ], + "type": "table", + "html": "
MethodsVisual BackboneMulti-taskVenuevalRefCOCO testAtestBRefCOCOgvalRefCOCO+ testAtestB
MCN [28]DN53✓CVPR2062.4464.2059.7149.2249.4050.6254.9944.69
RefTR [22]RN101✓NeurIPS2170.5673.4966.5758.7358.5161.0864.6552.73
SeqTR [59]DN53✓ECCV2267.2669.7964.1255.6755.6454.1458.9348.19
VG-LAW [38]ViT-B✓CVPR2375.6277.5172.8965.6366.0866.6370.3858.89
PVD [7]Swin-B✓AAAI2475.0777.2970.1363.2263.8964.3969.1557.19
Vista-7B [34]ViT-L✓CVPR2474.576.073.969.870.971.874.465.6
TCRTSwin-B✓-77.1278.1275.6770.8472.0872.7974.9666.53
", + "image_path": "9007d0c37ac9c2bf075e0b9812a841e59dd7d35979026b5e9c9335e343be1f33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 350, + 528, + 361 + ], + "lines": [ + { + "bbox": [ + 80, + 350, + 528, + 361 + ], + "spans": [ + { + "bbox": [ + 80, + 350, + 528, + 361 + ], + "type": "text", + "content": "Table 2. Comparison with existing visual grounding methods on RefCOCO, RefCOCOg, and RefCOCO+ for the RES task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 382, + 181, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 181, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 181, + 396 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 119, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 119, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 119, + 411 + ], + "type": "text", + "content": "4.1.1. Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 416, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 535 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method, we carry out experiments on the RefCOCO [51], RefCOCOg [33], RefCOCO+ [51], and ReferItGame [20] datasets. RefCOCO, RefCOCOg, and RefCOCO+ datasets are derived from MS-COCO [23]. RefCOCO and RefCOCO+ datasets were partitioned into train, validation (val), testA, and testB subsets. RefCOCOg was partitioned into train, validation (val), and test subsets. In addition, the ReferItGame dataset, which was collected from SAIAPR-12 [13], comprises both train and test subsets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 165, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 165, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 165, + 554 + ], + "type": "text", + "content": "4.1.2. Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 559, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 619 + ], + "type": "text", + "content": "For the RSC task, we employ Prec@0.5 as the metric for evaluation, which is consistent with prior works. If the IoU of the ground truth box with the predicted box surpasses 0.5, then the predicted box is deemed accurate. For the RES task, we evaluate performance by calculating MIoU." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 626, + 185, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 185, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 185, + 637 + ], + "type": "text", + "content": "4.1.3. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Our visual backbone utilizes a Swin Transformer architecture pre-trained with Mask-RCNN on MSCOCO [23]. The input image size is set to " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". For the textual backbone, we adopt the T5 [35] model and limit the maximum length of text to 40. For the large language model, we use Llama3.1-8B [12], which is frozen during the training" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 380, + 547, + 540 + ], + "blocks": [ + { + "bbox": [ + 323, + 380, + 547, + 540 + ], + "lines": [ + { + "bbox": [ + 323, + 380, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 323, + 380, + 547, + 540 + ], + "type": "image", + "image_path": "6188e1706f311249b9db962a43220c5e9af8df4e1af5512001fc699679ecd39b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 548, + 555, + 593 + ], + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 593 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with representative methods on REC task. We also visualize the attention scores of each method. Blue and green boxes are predicted regions and the ground truth, respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "content": "process. Our model undergoes end-to-end training for 70 epochs. The learning rate begins at " + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "content": ", and then we drop it down to " + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "content": " after 50 epochs. The optimizer used is AdamW. A dropout rate of 0.1 is applied. The batch size is configured to 16. We adopt the same data augmentation strategies as previous works [47, 48]. We implement our TCRT using PyTorch and perform the experiments on NVIDIA A800 GPUs." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3936" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 81, + 140, + 128 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 133, + 80 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 133, + 80 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 133, + 80 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 81, + 140, + 128 + ], + "lines": [ + { + "bbox": [ + 70, + 81, + 140, + 128 + ], + "spans": [ + { + "bbox": [ + 70, + 81, + 140, + 128 + ], + "type": "image", + "image_path": "6358e372ced6bd6f61e4b7bd2eb9d4cac3a4a8cfa7a4e4c7f409f48bfc5bff7d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 141, + 81, + 211, + 129 + ], + "blocks": [ + { + "bbox": [ + 162, + 70, + 190, + 80 + ], + "lines": [ + { + "bbox": [ + 162, + 70, + 190, + 80 + ], + "spans": [ + { + "bbox": [ + 162, + 70, + 190, + 80 + ], + "type": "text", + "content": "SeqTR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 81, + 211, + 129 + ], + "lines": [ + { + "bbox": [ + 141, + 81, + 211, + 129 + ], + "spans": [ + { + "bbox": [ + 141, + 81, + 211, + 129 + ], + "type": "image", + "image_path": "a440a266aefd8f3e14a63d30003e4eb93bfe9860595e43f42ea83530fc5b7b31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 212, + 81, + 282, + 129 + ], + "blocks": [ + { + "bbox": [ + 234, + 70, + 260, + 80 + ], + "lines": [ + { + "bbox": [ + 234, + 70, + 260, + 80 + ], + "spans": [ + { + "bbox": [ + 234, + 70, + 260, + 80 + ], + "type": "text", + "content": "TCRT" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 212, + 81, + 282, + 129 + ], + "lines": [ + { + "bbox": [ + 212, + 81, + 282, + 129 + ], + "spans": [ + { + "bbox": [ + 212, + 81, + 282, + 129 + ], + "type": "image", + "image_path": "4b69c4f3701904f548eab03417498a7420944447432e9ee5b00853df532275bb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 137, + 140, + 184 + ], + "blocks": [ + { + "bbox": [ + 72, + 129, + 139, + 137 + ], + "lines": [ + { + "bbox": [ + 72, + 129, + 139, + 137 + ], + "spans": [ + { + "bbox": [ + 72, + 129, + 139, + 137 + ], + "type": "text", + "content": "\"fourth from bottom ri" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 137, + 140, + 184 + ], + "lines": [ + { + "bbox": [ + 70, + 137, + 140, + 184 + ], + "spans": [ + { + "bbox": [ + 70, + 137, + 140, + 184 + ], + "type": "image", + "image_path": "c09b938cf065752ebab34e9edbba87df7fd16f1b1a83de0241066896a0c769e8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 141, + 137, + 211, + 184 + ], + "blocks": [ + { + "bbox": [ + 141, + 137, + 211, + 184 + ], + "lines": [ + { + "bbox": [ + 141, + 137, + 211, + 184 + ], + "spans": [ + { + "bbox": [ + 141, + 137, + 211, + 184 + ], + "type": "image", + "image_path": "3247815e9a75f295c49ff4d3a9f86ba2f67e515f178f0ae1fd26b4ea3bc7b87f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 212, + 137, + 282, + 184 + ], + "blocks": [ + { + "bbox": [ + 212, + 137, + 282, + 184 + ], + "lines": [ + { + "bbox": [ + 212, + 137, + 282, + 184 + ], + "spans": [ + { + "bbox": [ + 212, + 137, + 282, + 184 + ], + "type": "image", + "image_path": "b75658858a563569c1ada2bf6d255c86c11d8ca9d1bf0d2007f3cce8112a4748.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 70, + 194, + 140, + 241 + ], + "blocks": [ + { + "bbox": [ + 72, + 185, + 169, + 194 + ], + "lines": [ + { + "bbox": [ + 72, + 185, + 169, + 194 + ], + "spans": [ + { + "bbox": [ + 72, + 185, + 169, + 194 + ], + "type": "text", + "content": "\"big couch at bottom of screen\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 194, + 140, + 241 + ], + "lines": [ + { + "bbox": [ + 70, + 194, + 140, + 241 + ], + "spans": [ + { + "bbox": [ + 70, + 194, + 140, + 241 + ], + "type": "image", + "image_path": "46f741158de625298d7ea4f7002cc856c029f632428413bfc36f2435cce44bfd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 242, + 159, + 251 + ], + "lines": [ + { + "bbox": [ + 74, + 242, + 159, + 251 + ], + "spans": [ + { + "bbox": [ + 74, + 242, + 159, + 251 + ], + "type": "text", + "content": "\"giraffe appendage on right\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 258, + 283, + 270 + ], + "lines": [ + { + "bbox": [ + 67, + 258, + 283, + 270 + ], + "spans": [ + { + "bbox": [ + 67, + 258, + 283, + 270 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison with SeqTR on RES task." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 141, + 194, + 211, + 241 + ], + "blocks": [ + { + "bbox": [ + 141, + 194, + 211, + 241 + ], + "lines": [ + { + "bbox": [ + 141, + 194, + 211, + 241 + ], + "spans": [ + { + "bbox": [ + 141, + 194, + 211, + 241 + ], + "type": "image", + "image_path": "fc203ab519cc9858ec1a133de9d9bd4d6f1339ddbf1253ebdfa08642abba7461.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 212, + 194, + 282, + 241 + ], + "blocks": [ + { + "bbox": [ + 212, + 194, + 282, + 241 + ], + "lines": [ + { + "bbox": [ + 212, + 194, + 282, + 241 + ], + "spans": [ + { + "bbox": [ + 212, + 194, + 282, + 241 + ], + "type": "image", + "image_path": "ab4df296ca3d6f0d26f985e3ca70a7babb978adfdfccc8ac0c3cb0db402369d0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 77, + 293, + 143, + 335 + ], + "blocks": [ + { + "bbox": [ + 99, + 284, + 119, + 292 + ], + "lines": [ + { + "bbox": [ + 99, + 284, + 119, + 292 + ], + "spans": [ + { + "bbox": [ + 99, + 284, + 119, + 292 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 293, + 143, + 335 + ], + "lines": [ + { + "bbox": [ + 77, + 293, + 143, + 335 + ], + "spans": [ + { + "bbox": [ + 77, + 293, + 143, + 335 + ], + "type": "image", + "image_path": "eb4c6554ad13a1f7f18bb111117e4cfd1bacc8bca639b0232610f7c7a960aeee.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 143, + 293, + 208, + 335 + ], + "blocks": [ + { + "bbox": [ + 153, + 283, + 200, + 291 + ], + "lines": [ + { + "bbox": [ + 153, + 283, + 200, + 291 + ], + "spans": [ + { + "bbox": [ + 153, + 283, + 200, + 291 + ], + "type": "text", + "content": "TCRT w/o CN" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 293, + 208, + 335 + ], + "lines": [ + { + "bbox": [ + 143, + 293, + 208, + 335 + ], + "spans": [ + { + "bbox": [ + 143, + 293, + 208, + 335 + ], + "type": "image", + "image_path": "537e606bc7a6af9eca68fb32645269722e0bab1228f18f9a5d0597fdf6175c2d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 210, + 293, + 275, + 337 + ], + "blocks": [ + { + "bbox": [ + 231, + 283, + 252, + 291 + ], + "lines": [ + { + "bbox": [ + 231, + 283, + 252, + 291 + ], + "spans": [ + { + "bbox": [ + 231, + 283, + 252, + 291 + ], + "type": "text", + "content": "TCRT" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 210, + 293, + 275, + 337 + ], + "lines": [ + { + "bbox": [ + 210, + 293, + 275, + 337 + ], + "spans": [ + { + "bbox": [ + 210, + 293, + 275, + 337 + ], + "type": "image", + "image_path": "8500d96e1912d6f5b999616ba8fbfbab8da4635b66ed0ad58c23835d56f2d249.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 77, + 346, + 141, + 389 + ], + "blocks": [ + { + "bbox": [ + 80, + 337, + 146, + 345 + ], + "lines": [ + { + "bbox": [ + 80, + 337, + 146, + 345 + ], + "spans": [ + { + "bbox": [ + 80, + 337, + 146, + 345 + ], + "type": "text", + "content": "\"white laptop keyboard\"" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 346, + 141, + 389 + ], + "lines": [ + { + "bbox": [ + 77, + 346, + 141, + 389 + ], + "spans": [ + { + "bbox": [ + 77, + 346, + 141, + 389 + ], + "type": "image", + "image_path": "ba8020222ad4d5beb663cc58af574dae7d2b975f18f660d8c0e21de70accb3ed.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 143, + 346, + 208, + 389 + ], + "blocks": [ + { + "bbox": [ + 143, + 346, + 208, + 389 + ], + "lines": [ + { + "bbox": [ + 143, + 346, + 208, + 389 + ], + "spans": [ + { + "bbox": [ + 143, + 346, + 208, + 389 + ], + "type": "image", + "image_path": "47ac7ca98d42a6427ecf55ddb5143914c4ff5c50eb31d89023dee28c834bd1b5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 210, + 346, + 274, + 389 + ], + "blocks": [ + { + "bbox": [ + 210, + 346, + 274, + 389 + ], + "lines": [ + { + "bbox": [ + 210, + 346, + 274, + 389 + ], + "spans": [ + { + "bbox": [ + 210, + 346, + 274, + 389 + ], + "type": "image", + "image_path": "a50145551878df435be89db6073004bdf97a985133516c6339d31c453bf96b7e.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 77, + 400, + 141, + 444 + ], + "blocks": [ + { + "bbox": [ + 80, + 391, + 164, + 399 + ], + "lines": [ + { + "bbox": [ + 80, + 391, + 164, + 399 + ], + "spans": [ + { + "bbox": [ + 80, + 391, + 164, + 399 + ], + "type": "text", + "content": "\"person half shown on far left\"" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 400, + 141, + 444 + ], + "lines": [ + { + "bbox": [ + 77, + 400, + 141, + 444 + ], + "spans": [ + { + "bbox": [ + 77, + 400, + 141, + 444 + ], + "type": "image", + "image_path": "3d99568173df8748042329cb6644c793bd27e2f82d9c738a370f0295c1b74339.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 445, + 173, + 453 + ], + "lines": [ + { + "bbox": [ + 81, + 445, + 173, + 453 + ], + "spans": [ + { + "bbox": [ + 81, + 445, + 173, + 453 + ], + "type": "text", + "content": "\"Guy lower right side facing away\"" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 463, + 295, + 507 + ], + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 507 + ], + "type": "text", + "content": "Figure 6. Visualization of features extracted by the visual backbone. The second column shows the results of 'TCRT w/o CN', while the third column shows the results of TCRT. Red and white boxes are the predicted regions and the ground truth, respectively." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 143, + 400, + 208, + 444 + ], + "blocks": [ + { + "bbox": [ + 143, + 400, + 208, + 444 + ], + "lines": [ + { + "bbox": [ + 143, + 400, + 208, + 444 + ], + "spans": [ + { + "bbox": [ + 143, + 400, + 208, + 444 + ], + "type": "image", + "image_path": "14c1e7263b24a01e6c0bad7c318df54a295376faa762cb92ef0921029d9be390.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 210, + 400, + 274, + 444 + ], + "blocks": [ + { + "bbox": [ + 210, + 400, + 274, + 444 + ], + "lines": [ + { + "bbox": [ + 210, + 400, + 274, + 444 + ], + "spans": [ + { + "bbox": [ + 210, + 400, + 274, + 444 + ], + "type": "image", + "image_path": "430e80c12c761bb581589d32754d356fcd6060ba2418418c756570868c86fc58.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 529, + 244, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 244, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 244, + 541 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-arts" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "content": "To evaluate the effectiveness of the proposed TCRT, we perform both qualitative and quantitative experiments across four datasets: RefCOCO, RefCOCOg, RefCOCO+, and ReferItGame." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 601, + 126, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 126, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 126, + 613 + ], + "type": "text", + "content": "4.2.1. REC task" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "We evaluate our TCRT against existing state-of-the-art models for the REC task. As shown in Table 1, TCRT achieves the best performance across these datasets. Specifically, compared to Vista-7B, TCRT shows absolute improvements of " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "+2.97" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "+1.35" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "+3.24" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " percentage points on RefCOCO; " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "+1.65" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "+0.46" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "+1.79" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " percentage points on RefCOCO+; and " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "+0.81" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "+0.73" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " percentage points on RefCOCOg. We also present qualitative results from three" + } + ] + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 335, + 72, + 536, + 198 + ], + "blocks": [ + { + "bbox": [ + 335, + 72, + 536, + 198 + ], + "lines": [ + { + "bbox": [ + 335, + 72, + 536, + 198 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 536, + 198 + ], + "type": "image", + "image_path": "bfd3dd46061ca25f22683c2161b057e7714c41c8b698cd261614785d77d42d12.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 206, + 554, + 229 + ], + "lines": [ + { + "bbox": [ + 313, + 206, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 554, + 229 + ], + "type": "text", + "content": "Figure 7. Comparison between the TCRT without object-text matching loss " + }, + { + "bbox": [ + 313, + 206, + 554, + 229 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{match}})" + }, + { + "bbox": [ + 313, + 206, + 554, + 229 + ], + "type": "text", + "content": " and TCRT in terms of " + }, + { + "bbox": [ + 313, + 206, + 554, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 250, + 555, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 322 + ], + "type": "text", + "content": "examples of the RefCOCO testB set for the REC task in Figure 4. We can see that TCRT performs well even when an image contains multiple similar objects, such as \"second vase from front\". Furthermore, we visualized the attention scores within the transformer. We observe that TCRT focuses better on the target object compared to other methods." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 327, + 382, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 382, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 382, + 338 + ], + "type": "text", + "content": "4.2.2. RES task" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "content": "We evaluate the effectiveness of our TCRT for the RES task by comparing it with existing state-of-the-art methods. The comparative results are provided in Table 2. Compared with Vista-7B, TCRT achieves higher mIoU scores with absolute improvements of " + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "inline_equation", + "content": "+2.62" + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "inline_equation", + "content": "+2.12" + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "inline_equation", + "content": "+1.77" + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "content": " percentage points on RefCOCO, " + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "inline_equation", + "content": "+0.99" + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "inline_equation", + "content": "+0.56" + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "inline_equation", + "content": "+0.93" + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "content": " percentage points on RefCOCO+, " + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "inline_equation", + "content": "+1.04" + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "inline_equation", + "content": "+1.18" + }, + { + "bbox": [ + 313, + 342, + 555, + 449 + ], + "type": "text", + "content": " percentage points on RefCOCOg. We also present a comparison for the RES task in Figure 5." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 457, + 478, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 478, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 478, + 470 + ], + "type": "text", + "content": "4.3. Analysis of Main Components" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 475, + 457, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 457, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 457, + 485 + ], + "type": "text", + "content": "4.3.1. Conditional Normalization" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 489, + 554, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 554, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 554, + 585 + ], + "type": "text", + "content": "We first conduct an experiment to demonstrate the effectiveness of conditional normalization (CN). We visualize the features extracted by the visual backbone on RefCOCOg test set. As shown in Figure 6, 'TCRT w/o CN' has more dispersed attention, which may lead to inaccurate predictions. Due to the smaller domain discrepancy of TCRT, this allows the LLM to focus more attention on the target objects." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 591, + 541, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 541, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 541, + 603 + ], + "type": "text", + "content": "4.3.2. Object-text Semantic Matching Regularization" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "We use the object-text matching loss " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{match}})" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " to associate the visual tokens with the textual ones. To validate the effectiveness of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " we compare 'TCRT w/o " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " and TCRT in terms of the detection loss " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{det}})" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " on RefCOCO val set. In Figure 7, by plotting the loss curves of the models during training, we observe that TCRT achieves lower loss values, which is attributed to its capability to learn semantic-discriminative visual features. This enables the LLM to more easily recognize task-relevant visual features." + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3937" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 289, + 223 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 289, + 223 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 289, + 223 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 289, + 223 + ], + "type": "image", + "image_path": "bd67a5d4df4a948bc5176529532f5e9d0e03741c989aa5466dcbd92db0569c33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 295, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 289 + ], + "type": "text", + "content": "Figure 8. visualization of LLM-guided visual feature refinement. We visualize the weight matrices " + }, + { + "bbox": [ + 55, + 232, + 295, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 232, + 295, + 289 + ], + "type": "text", + "content": " of each layer in the fusion module to validate the role of the LLM in refining visual features. Red and white boxes are the predicted regions and the ground truth, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 299, + 289, + 457 + ], + "blocks": [ + { + "bbox": [ + 63, + 299, + 289, + 457 + ], + "lines": [ + { + "bbox": [ + 63, + 299, + 289, + 457 + ], + "spans": [ + { + "bbox": [ + 63, + 299, + 289, + 457 + ], + "type": "image", + "image_path": "30fdeff73b8690ec6b7c831502cacf7057cc0e7a354a3efc9fd3a424e9f1311f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 465, + 266, + 476 + ], + "lines": [ + { + "bbox": [ + 85, + 465, + 266, + 476 + ], + "spans": [ + { + "bbox": [ + 85, + 465, + 266, + 476 + ], + "type": "text", + "content": "Figure 9. Analysis of Visual Feature Distribution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 497, + 255, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 255, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 255, + 509 + ], + "type": "text", + "content": "4.3.3. LLM-guided Visual Feature Refinement" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 512, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 608 + ], + "type": "text", + "content": "We visualize the weight matrix " + }, + { + "bbox": [ + 55, + 512, + 296, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 512, + 296, + 608 + ], + "type": "text", + "content": " in Figure 8 to validate the effectiveness of visual feature refinement on RefCOCOg test set. We observe that the attention of the weight matrix is relatively evenly distributed in the earlier layers. In the later layers, the attention of the weight matrix is primarily focused on the target object. This indicates that LLM's prior knowledge enhances our model's ability to locate target objects by further refining the visual features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 614, + 250, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 250, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 250, + 626 + ], + "type": "text", + "content": "4.3.4. Analysis of Visual Feature Distribution" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "To validate the effectiveness of our model in aligning visual and textual features, we extract the visual features after fusion with textual information and study their distribution in the spatial domain. Specifically, we label four target objects in an image and use three different textual descriptions for each object, resulting in twelve data samples. These samples are fed into the grounding model, and the visual" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 347, + 70, + 525, + 136 + ], + "blocks": [ + { + "bbox": [ + 347, + 70, + 525, + 136 + ], + "lines": [ + { + "bbox": [ + 347, + 70, + 525, + 136 + ], + "spans": [ + { + "bbox": [ + 347, + 70, + 525, + 136 + ], + "type": "table", + "html": "
MethodvalRefCOCO
testAtestB
TCRT w/o VF & VA87.6288.6084.04
TCRT w/o VF88.3890.2485.57
TCRT91.0792.8586.24
", + "image_path": "48e08a8a3f33d5fb4cce0c7701013b750afaad4ad66078c6e25adaeba5e88a7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 144, + 555, + 177 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 177 + ], + "type": "text", + "content": "Table 3. Ablation experiments on RefCOCO to evaluate the proposed LLM adaptation module " + }, + { + "bbox": [ + 313, + 144, + 555, + 177 + ], + "type": "inline_equation", + "content": "\\left( {V}_{A}\\right)" + }, + { + "bbox": [ + 313, + 144, + 555, + 177 + ], + "type": "text", + "content": " and task-aware cross-modal feature fusion module " + }, + { + "bbox": [ + 313, + 144, + 555, + 177 + ], + "type": "inline_equation", + "content": "\\left( {V}_{F}\\right)" + }, + { + "bbox": [ + 313, + 144, + 555, + 177 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 201, + 555, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 249 + ], + "type": "text", + "content": "features from the fusion module's output are subjected to dimensionality reduction and visualization. As shown in Figure 9, we observe that our method causes the visual features describing the same object to cluster together." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 258, + 413, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 413, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 413, + 268 + ], + "type": "text", + "content": "4.3.5. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": "We validate the design of our proposed TCRT by conducting ablative experiments on major components in the REC task. The results are presented in Table 3. We observe that when the LLM adaptation module and task-aware cross-modal feature fusion module are not used, 'TCRT w/o " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "V_{F}" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "V_{A}" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": "' achieves accuracy rates of " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "87.62\\%" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "88.60\\%" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "84.04\\%" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": " on the RefCOCO dataset. After incorporating the LLM adaptation module, the accuracy of 'TCRT w/o " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "V_{F}" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": "' increases by 0.76, 1.64, and 1.53 absolute percentage points. When the " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "V_{F}" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": " module is further introduced, TCRT achieves accuracy rates of " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "91.07\\%" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "92.85\\%" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "inline_equation", + "content": "86.24\\%" + }, + { + "bbox": [ + 313, + 273, + 555, + 405 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 418, + 388, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 388, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 388, + 430 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 440, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 608 + ], + "type": "text", + "content": "In this paper, we propose a task-aware cross-modal feature refinement transformer with large language models to facilitate visual grounding. To adapt LLMs to the grounding task, we design an adaptation module that modulates the statistical characteristics of visual features conditioned on textual features along with object-text semantic matching regularization. An LLM is employed to integrate the aligned information from textual and visual features and produce task-aware embeddings. We further incorporate a task-aware cross-modal feature fusion module, allowing the resulting embeddings to refine visual features and enable information interaction between the REC and RES tasks. Comprehensive experimental results verify the effectiveness of our proposed approach both quantitatively and qualitatively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 620, + 423, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 423, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 423, + 634 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "This work was supported in part by TCL Science and Technology Innovation Fund (Project No. 20231752), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), and in part by the GuangDong Basic and Applied Basic Research Foundation (Project No. 2024A1515011437)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3938" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "text", + "content": "[2] Mohit Bajaj, Lanjun Wang, and Leonid Sigal. G3raphground: Graph-based language grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4281-4290, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 294, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 294, + 213 + ], + "type": "text", + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 278 + ], + "type": "text", + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 324 + ], + "type": "text", + "content": "[5] Kan Chen, Rama Kovvuri, and Ram Nevatia. Query-guided regression network with context policy for phrase grounding. In Proceedings of the IEEE International Conference on Computer Vision, pages 824-832, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 296, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 296, + 379 + ], + "type": "text", + "content": "[6] Long Chen, Wenbo Ma, Jun Xiao, Hanwang Zhang, and Shih-Fu Chang. Ref-nms: Breaking proposal bottlenecks in two-stage referring expression grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1036-1044, 2021. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "text", + "content": "[7] Zesen Cheng, Kehan Li, Peng Jin, Siheng Li, Xiangyang Ji, Li Yuan, Chang Liu, and Jie Chen. Parallel vertex diffusion for unified visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1326-1334, 2024. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 501 + ], + "type": "text", + "content": "[8] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 556 + ], + "type": "text", + "content": "[9] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1769-1779, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 613 + ], + "type": "text", + "content": "[10] Jiajun Deng, Zhengyuan Yang, Daqing Liu, Tianlang Chen, Wengang Zhou, Yanyong Zhang, Houqiang Li, and Wanli Ouyang. Transvg++: End-to-end visual grounding with language conditioned vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 658 + ], + "type": "text", + "content": "[11] Pelin Dogan, Leonid Sigal, and Markus Gross. Neural sequential phrase grounding (seqground). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4175-4184, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "text", + "content": "[13] Hugo Jair Escalante, Carlos A Hernández, Jesus A Gonzalez, Aurelio López-López, Manuel Montes, Eduardo F Morales, L Enrique Sucar, Luis Villasenor, and Michael Grubinger. The segmented and annotated iapr tc-12 benchmark. Computer Vision and Image Understanding, 114(4): 419-428, 2010. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 140, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 184 + ], + "type": "text", + "content": "[14] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 185, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 217 + ], + "type": "text", + "content": "[15] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 1440-1448, 2015. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "text", + "content": "[16] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 580-587, 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[17] Chih-Hui Ho, Srikar Appalaraju, Bhavan Jasani, R Manmatha, and Nuno Vasconcelos. Yoro-lightweight end to end visual grounding. In European Conference on Computer Vision, pages 3-23. Springer, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 321, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 376 + ], + "type": "text", + "content": "[18] Ronghang Hu, Marcus Rohrbach, Jacob Andreas, Trevor Darrell, and Kate Saenko. Modeling relationships in referential expressions with compositional modular networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1115-1124, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "type": "text", + "content": "[19] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 433, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 433, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 433, + 553, + 487 + ], + "type": "text", + "content": "[20] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 787-798, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 488, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 544 + ], + "type": "text", + "content": "[21] Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv e-prints, pages arXiv-2211, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 588 + ], + "type": "text", + "content": "[22] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in Neural Information Processing Systems, 34:19652-19664, 2021. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 656 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 700 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 702, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 553, + 714 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3939" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[26] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[27] Xihui Liu, Zihao Wang, Jing Shao, Xiaogang Wang, and Hongsheng Li. Improving referring expression grounding with cross-modal attention-guided erasing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1950-1959, 2019. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 270 + ], + "type": "text", + "content": "[28] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10034-10043, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[29] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. In European Conference on Computer Vision, pages 417-435. Springer, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "type": "text", + "content": "[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11-20, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 425 + ], + "type": "text", + "content": "[31] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 Fourth International Conference on 3D Vision (3DV), pages 565-571. IEEE, 2016. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "text", + "content": "[32] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 557 + ], + "type": "text", + "content": "[33] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 635 + ], + "type": "text", + "content": "[34] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "text", + "content": "[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 162, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 555, + 216 + ], + "type": "text", + "content": "[38] Wei Su, Peihan Miao, Huanzhang Dou, Gaoang Wang, Liang Qiao, Zheyang Li, and Xi Li. Language adaptive weight generation for multi-task visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10857-10866, 2023. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "text", + "content": "[39] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[42] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[43] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "text", + "content": "[44] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4145-4154, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[45] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 526, + 555, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 555, + 569 + ], + "type": "text", + "content": "[46] Sibei Yang, Guanbin Li, and Yizhou Yu. Graph-structured referring expression reasoning in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9952–9961, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "text", + "content": "[47] Zhengyuan Yang, Boqing Gong, Liwei Wang, Wenbing Huang, Dong Yu, and Jiebo Luo. A fast and accurate one-stage approach to visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4683-4693, 2019. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "text", + "content": "[48] Zhengyuan Yang, Tianlang Chen, Liwei Wang, and Jiebo Luo. Improving one-stage visual grounding by recursive subquery construction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 387-404. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[49] Jiabo Ye, Junfeng Tian, Ming Yan, Xiaoshan Yang, Xuwu Wang, Ji Zhang, Liang He, and Xin Lin. Shifting more" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3940" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 642 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 75, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 297, + 116 + ], + "type": "text", + "content": "attention to visual backbone: Query-modulated refinement networks for end-to-end visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15502-15512, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[50] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 238 + ], + "type": "text", + "content": "[51] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 295 + ], + "type": "text", + "content": "[52] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315, 2018. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 350 + ], + "type": "text", + "content": "[53] Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. In European Conference on Computer Vision, pages 405-422. Springer, 2025. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 352, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 352, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 352, + 294, + 393 + ], + "type": "text", + "content": "[54] Ao Zhang, Liming Zhao, Chen-Wei Xie, Yun Zheng, Wei Ji, and Tat-Seng Chua. Next-chat: An lmm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "type": "text", + "content": "[55] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4158-4166, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 474 + ], + "type": "text", + "content": "[56] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "text", + "content": "[57] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 531, + 294, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 294, + 585 + ], + "type": "text", + "content": "[58] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. Advances in Neural Information Processing Systems, 37:71737-71767, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 587, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 642 + ], + "type": "text", + "content": "[59] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision, pages 598-615. Springer, 2022. 1, 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3941" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Task-driven Image Fusion with Learnable Fusion Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_content_list.json b/2025/Task-driven Image Fusion with Learnable Fusion Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..81b6fea9e8d615d0c0e94ac1a076a032be52b84d --- /dev/null +++ b/2025/Task-driven Image Fusion with Learnable Fusion Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_content_list.json @@ -0,0 +1,1537 @@ +[ + { + "type": "text", + "text": "Task-driven Image Fusion with Learnable Fusion Loss", + "text_level": 1, + "bbox": [ + 223, + 130, + 774, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haowen Bai $^{1}$ Jiangshe Zhang $^{1*}$ Zixiang Zhao $^{2*}$ Yichen Wu $^{3}$ Lilun Deng $^{1}$ Yukun Cui $^{1}$ Tao Feng $^{4}$ Shuang Xu $^{5}$", + "bbox": [ + 243, + 179, + 750, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xi'an Jiaotong University $^{2}$ ETH Zürich $^{3}$ City University of Hong Kong $^{4}$ Tsinghua University $^{5}$ Northwestern Polytechnical University", + "bbox": [ + 197, + 218, + 799, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hwbai@stu.xjtu.edu.cn", + "bbox": [ + 395, + 258, + 594, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 306, + 326, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal image fusion aggregates information from multiple sensor sources, achieving superior visual quality and perceptual features compared to single-source images, often improving downstream tasks. However, current fusion methods for downstream tasks still use predefined fusion objectives that potentially mismatch the downstream tasks, limiting adaptive guidance and reducing model flexibility. To address this, we propose Task-driven Image Fusion (TD-Fusion), a fusion framework incorporating a learnable fusion loss guided by task loss. Specifically, our fusion loss includes learnable parameters modeled by a neural network called the loss generation module. This module is supervised by the downstream task loss in a meta-learning manner. The learning objective is to minimize the task loss of fused images after optimizing the fusion module with the fusion loss. Iterative updates between the fusion module and the loss module ensure that the fusion network evolves toward minimizing task loss, guiding the fusion process toward the task objectives. TDFusion's training relies entirely on the downstream task loss, making it adaptable to any specific task. It can be applied to any architecture of fusion and task networks. Experiments demonstrate TDFusion's performance through fusion experiments conducted on four different datasets, in addition to evaluations on semantic segmentation and object detection tasks. The code is available at https://github.com/HaowenBai/TDFusion.", + "bbox": [ + 88, + 339, + 485, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 776, + 220, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal image fusion [25, 31, 51, 59, 72, 79, 82] combines information from multiple sensors to produce a more holistic and detailed representation. Infrared images capture thermal radiation regardless of lighting conditions, while visible images provide richer texture details. Fused im", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ages enhance downstream tasks through improved information density and robustness [18, 41, 45, 62, 63, 70, 71], outperforming single-modal inputs in semantic segmentation [13, 28, 34, 56], object detection [8, 32], and other related applications [1, 3, 20, 27]. Conventional methods typically treat fusion as image restoration using unsupervised loss [76, 83, 87-89] or perceptual loss [21, 23, 77]. These approaches prioritize visual-level fusion through predefined aggregation objectives, often neglecting semantic feature extraction. This limitation hinders scene interpretation and task performance [14, 24, 52, 64]. Recent advances explore the mutual enhancement between fusion and downstream tasks [37]. By cascading the fusion network with downstream task network [39, 49], the task loss constrains the fusion learning, ensuring the fused images meet the task requirements [32, 56]. Alternatively, some methods incorporate high-level visual task features [34, 76, 78] or focus on learning optimal initializations [40] to enhance fusion.", + "bbox": [ + 511, + 309, + 906, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While integrating downstream tasks, existing frameworks still rely on predefined fusion loss terms lacking dynamic adaptation. The impact of downstream tasks remains limited due to specific combinations. Manually defined losses preserve predefined guidance, frequently overlooking task-specific requirements. This guidance imposes manually designed prior constraints on the fusion process, limiting the dynamic and adaptive influence of downstream tasks on specific image pairs. These approaches, whether incorporating task features [34, 78] or employing task losses [32, 56], still face the limitations of fixed fusion loss terms. Task-specialized networks [76] create fusion-task dependencies, restricting flexibility and limiting their applicability to various high-level vision tasks. We address these limitations through a task-driven framework with learnable loss. The fusion loss contains learnable parameters, generated by a loss generation module, and is designed to retain the intensity information of the source images for specific downstream tasks. The purpose of updating the fusion loss is to guide the fusion network in generating fused images that minimize downstream task loss,", + "bbox": [ + 511, + 583, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7457", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thereby enhancing adaptability. Moreover, the fusion loss update relies on the downstream task loss, making it independent of any specific task or network architecture.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The loss generation module produces fusion losses for subsequent fusion module updates. This complexity presents challenges to standard end-to-end training. Fortunately, meta-learning techniques, which are strategies for learning how to learn, can effectively achieve the learning objectives of the loss generation module. This involves minimizing task loss for fused images through an optimized fusion loss. Meta-learning tackles common deep-learning challenges like limited data, high computational costs, and the need for better generalization. Core optimization areas include parameter tuning [11], optimization strategies [26], and network architectures search [30]. In this paper, we draw inspiration from the Model-Agnostic Meta-Learning (MAML) approach [11] to train our loss generation module. Specifically, training the loss generation module involves two stages: inner updates and outer updates. During inner updates, the output of the loss generation module updates a surrogate fusion module without altering the original parameters. During outer updates, the fused image from the surrogate module is fed into the task network. The resulting task loss then updates the loss generation module through backpropagation. This alternating training ensures that the loss generation module consistently produces fusion losses that minimize the downstream task loss of the fused images.", + "bbox": [ + 91, + 140, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper introduces TDFusion, a task-oriented fusion framework driven by downstream tasks. It consists of a fusion module, a task module, and a loss generation module that learns to optimize the fusion loss. The fusion loss incorporates intensity preferences from source images and gradient preservation, guided by downstream task loss to refine intensity preferences. This model follows the general form of fusion loss used in advanced methods [76, 83, 88, 89], ensuring adaptability to various tasks. The updates of the loss generation module are performed using a meta-learning approach, optimizing the loss generation module parameters based on task loss from the fused images after each update of the fusion module. This process ensures that the loss function guides the fusion module continuously, optimizing feature aggregation and minimizing downstream task losses. The loss generation module dynamically adjusts through alternating updates with the fusion and task modules, generating optimal fusion losses at each model state. Our contributions can be summarized as follows:", + "bbox": [ + 89, + 505, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose TDFusion, a meta-learning-based fusion framework that leverages the loss functions of downstream tasks for training. This method promotes task-driven fusion and alleviates challenges caused by the absence of ground truth. Moreover, this framework is agnostic of specific downstream tasks or network architectures, which enhances its adaptability and flexibility.", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our framework includes a dynamically updated, learnable fusion loss generation module. It selectively extracts source image information, minimizing the loss in downstream tasks. This ensures optimal fusion performance while maximizing adaptability to downstream tasks.", + "- We analyze the information preferences of downstream tasks such as semantic segmentation and object detection, providing deeper insights into multi-modal high-level vision tasks.", + "- TDFusion achieves outstanding performance in both fusion and high-level vision tasks, validated on four fusion datasets with semantic segmentation and object detection." + ], + "bbox": [ + 513, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 287, + 653, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Deep learning-based Image Fusion", + "text_level": 1, + "bbox": [ + 513, + 314, + 815, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning-based image fusion methods have revolutionized the field by exploiting the powerful feature extraction capabilities of neural networks [4, 6, 33, 35, 36, 42-44, 65, 77]. These methods are broadly categorized into discriminative and generative approaches. Discriminative methods [10, 81, 85, 88] leverage the strong reconstruction ability of neural networks to directly learn the mapping between source and fused images [21, 23, 80, 82]. Generative methods, on the other hand, model the image generation process using generative approaches, integrating source images from a distributional perspective. These include methods based on Generative Adversarial Networks [46-48, 87] and diffusion models [73, 84]. Unified fusion methods [66, 75, 89] bridge the gap between different fusion sub-tasks, incorporating strategies such as continual learning [67] and self-supervised decomposition techniques [29]. The introduction of registration modules helps to mitigate misalignment issues [16, 60, 68] in source images. Recent studies have further explored the synergy between fusion and high-level vision tasks. These include leveraging downstream task losses to optimize fusion networks [32, 56], embedding high-level task features [34, 76, 78], and employing initialization techniques [40].", + "bbox": [ + 511, + 337, + 903, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Meta-Learning in Vision", + "text_level": 1, + "bbox": [ + 513, + 696, + 740, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meta-learning develops algorithms to automatically fine-tune hyperparameters for specific tasks, showing its versatility and effectiveness across various domains. MAML [11] and its variations [12, 50, 53] focus on learning efficient initialization parameters to quickly adapt to new tasks using minimal data. Meta-SGD [26] extends MAML by learning optimal update directions and rates, beneficial in few-shot learning scenarios. Other approaches like MW-Net [55] and L2RW [54] emphasize selecting relevant sample weights to tackle noisy data using a compact validation set. Additionally, some studies focus on improving model adaptability through learning loss functions [2, 7, 15].", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7458", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ff889df2ea5748d77aca49d1752dbca2103ade44bfc297f2bb885d9f67cfe3cb.jpg", + "image_caption": [ + "Learning of Loss Generate Module", + "Learning of Fusion", + "Figure 1. The TDFusion workflow alternates between training the loss generation module and the fusion module. Training of the loss generation module involves both inner and outer updates, learned through meta-learning." + ], + "image_footnote": [], + "bbox": [ + 98, + 107, + 903, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In image fusion, learnable filters [22] enable the fusion of images at arbitrary resolutions. MetaFusion [78] introduces a mechanism that improves image fusion and object detection by aligning semantics with fusion-specific features. ReFusion [5] guides the learnable fusion loss for various fusion tasks by reconstructing the source images. Metalearning also supports neural architecture search [38, 40] to identify optimal network architectures for image fusion and customizes network initialization for various tasks [40].", + "bbox": [ + 88, + 425, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Comparison with Existing Approaches", + "text_level": 1, + "bbox": [ + 89, + 573, + 421, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel image fusion method tailored for downstream tasks, leveraging a learnable fusion loss driven by task-specific objectives. Our method employs a meta-learning algorithm that alternates between inner and outer updates, allowing the downstream task loss to guide the optimization of learnable fusion parameters. This results in fused images that minimize the downstream task loss, enhancing their adaptability across various tasks. Unlike previous methods, our approach develops a task-specific fusion loss, shifting focus from traditional factors such as resolution and network structure, and avoiding reliance on predefined fusion loss terms. This renders our fusion framework more flexible and applicable to various scenarios.", + "bbox": [ + 89, + 595, + 483, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 806, + 181, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 832, + 200, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our TDFusion framework, as shown in Fig. 1, consists of a fusion network $\\mathcal{F}(\\cdot)$ , a downstream task network $\\mathcal{T}(\\cdot)$ and a fusion loss generation module $\\mathcal{G}(\\cdot)$ , which produces", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "parameters for a learnable loss function. The parameters of these modules are denoted as $\\theta_{\\mathcal{F}}$ , $\\theta_{\\mathcal{T}}$ , and $\\theta_{\\mathcal{G}}$ , respectively. The one-step updated clones of $\\mathcal{F}$ and $\\mathcal{T}$ are denoted as $\\mathcal{F}'$ and $\\mathcal{T}'$ , with parameters $\\theta_{\\mathcal{F}'}$ and $\\theta_{\\mathcal{T}'}$ . During the updates, the fusion network and the loss generation module alternate in learning, as depicted by blue and purple in Fig. 1. The update of the loss generation module consists of inner and outer updates, detailed in the following subsections. $\\mathcal{L}_f$ and $\\mathcal{L}_t$ represent the learnable fusion loss and task-specific loss, with their formulations provided in the next section.", + "bbox": [ + 511, + 425, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Loss Function", + "text_level": 1, + "bbox": [ + 511, + 587, + 660, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The learnable fusion loss $\\mathcal{L}_f$ consists of the intensity term and the gradient term. The intensity term is defined by the output of the loss generation module $\\{w_a, w_b\\} = \\mathcal{G}(I_a, I_b)$ , where $w_a$ and $w_b$ control the intensity preference in the fusion loss. The Softmax function in the loss generation module ensures $w_a^{ij} + w_b^{ij} = 1$ for each pixel, thereby selectively retaining the intensity information from the source images. The gradient term emphasizes higher gradient values from the input images [76, 83, 89], aiming to preserve maximal information from the source. The detailed formulation of the learnable fusion loss is as follows:", + "bbox": [ + 511, + 608, + 906, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = \\mathcal {L} _ {f} ^ {\\text {i n t}} + \\alpha \\mathcal {L} _ {f} ^ {\\text {g r a d}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 785, + 906, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} ^ {i n t} = \\frac {1}{H W} \\sum_ {i j} \\left[ \\sum_ {k \\in \\{a, b \\}} w _ {k} ^ {i j} \\left(I _ {f} ^ {i j} - I _ {k} ^ {i j}\\right) ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 808, + 903, + 857 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} ^ {\\text {g r a d}} = \\frac {1}{H W} \\sum_ {i j} \\left| \\nabla I _ {f} ^ {i j} - \\max _ {k \\in \\{a, b \\}} \\left(\\nabla I _ {k} ^ {i j}\\right) \\right|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 859, + 903, + 896 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7459", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\nabla$ denotes the Sobel operator, commonly employed for gradient extraction in image fusion [32, 57, 83]. The parameter $\\alpha$ serves as a scaling factor, while $\\mathcal{L}_f^{int}$ and $\\mathcal{L}_f^{grad}$ represent the intensity loss and gradient loss, respectively. The weights $\\{w_a, w_b\\}$ control the emphasis of the loss function on the intensity information from each source image. These parameters enable the fusion process to selectively aggregate and incorporate relevant information from the source images. Variations in $\\theta_{\\mathcal{G}}$ lead to different configurations of $w_a$ and $w_b$ , thereby influencing the characteristics of the fusion loss. $\\theta_{\\mathcal{G}}$ undergoes updates driven by the high-level task loss during the in-step update process, as detailed in Sec. 3.4.", + "bbox": [ + 89, + 90, + 483, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The loss function $\\mathcal{L}_t$ depends on the specific task. In this study, we adopt SegFormer [9] and YOLOv8 [19] for the downstream tasks, employing cross-entropy loss and YOLO loss [19] for each task, respectively.", + "bbox": [ + 89, + 287, + 483, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset Partitioning", + "text_level": 1, + "bbox": [ + 89, + 359, + 284, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to enhance the effectiveness of the loss generation module $\\mathcal{G}$ in guiding the fusion tasks, we create non-overlapping subsets of size $M$ at each training epoch. The meta-training set $\\{I_a^{mtr}, I_b^{mtr}\\}$ and the meta-test set $\\{I_a^{mts}, I_b^{mts}\\}$ are randomly drawn from the fusion training set $\\{I_a^{ftr}, I_b^{ftr}\\}$ . These subsets are fed into the model sequentially during the training process of the loss generation module, covering both the inner and outer updates. The entire fusion training set $\\{I_a^{ftr}, I_b^{ftr}\\}$ is fully utilized during the training of the fusion network.", + "bbox": [ + 89, + 382, + 483, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Learning of loss generation module", + "text_level": 1, + "bbox": [ + 89, + 547, + 398, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 1 illustrates that the loss generation and fusion modules are trained alternately, ensuring the fusion loss is optimized at various stages of training and under different states of the fusion network. The process of training to optimize the fusion loss involves two key steps: the inner update and the outer update. In the inner update, clones of both the fusion network and the task network are generated, with each network undergoing a single training iteration using the fusion loss and task loss, respectively. This process is designed to obtain the state of the network guided by the fusion loss. During the outer update, the task loss of the fusion image produced by the updated clone is calculated. This loss is then backpropagated to the loss generation network. The goal of this step is to direct the fusion network to generate fusion images that result in lower downstream task loss, once guided by the fusion loss. The alternating updates between the inner and outer steps constitute the learning procedure for the loss generation module.", + "bbox": [ + 89, + 569, + 483, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1. Inner Update", + "text_level": 1, + "bbox": [ + 89, + 849, + 230, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the inner update phase, the fusion network $\\mathcal{F}$ undergoes a single update guided by the fusion loss, which", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "depends on the current state of network $\\mathcal{G}$ . This update primarily aims to compute the intermediate parameters $\\theta_{\\mathcal{F}'}$ and $\\theta_{\\mathcal{R}'}$ , which are crucial for updating $\\theta_{\\mathcal{G}}$ in the subsequent phase. The upper section of the purple region in Fig. 1 illustrates this process. During this phase, the images from the meta-training set $\\{I_a^{mtr}, I_b^{mtr}\\}$ are fed into the model:", + "bbox": [ + 511, + 90, + 905, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathcal {F} ^ {\\prime}} = \\theta_ {\\mathcal {F}} - \\eta_ {\\mathcal {F} ^ {\\prime}} \\frac {\\partial \\mathcal {L} _ {f} \\left(I _ {a} ^ {m t r} , I _ {b} ^ {m t r} , I _ {f} ^ {m t r} ; \\theta_ {\\mathcal {G}}\\right)}{\\partial \\theta_ {\\mathcal {F}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 186, + 906, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ undergoes a single gradient descent update. $\\theta_{\\mathcal{G}}$ represent the parameters of the loss generation module $\\mathcal{G}$ , which determine the parameters for the learnable fusion loss. And the notation $\\eta_{\\mathcal{F}'}$ refers to the step size. The module $\\mathcal{F}'$ temporarily substitutes $\\mathcal{F}$ , adjusting its parameters in one update step. Meanwhile, the parameters of $\\mathcal{F}$ , denoted as $\\theta_{\\mathcal{F}}$ , remain unchanged. Similarly, $\\mathcal{T}'$ is updated in a single step using the parameters $\\theta_{\\mathcal{T}}$ from the current task network $\\mathcal{T}$ :", + "bbox": [ + 511, + 222, + 906, + 357 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathcal {T} ^ {\\prime}} = \\theta_ {\\mathcal {T}} - \\eta_ {\\mathcal {T} ^ {\\prime}} \\frac {\\partial \\mathcal {L} _ {t} \\left(I _ {f} ^ {m t r}\\right)}{\\partial \\theta_ {\\mathcal {T}}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 361, + 906, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The parameters $\\theta_{\\mathcal{F}'}$ and $\\theta_{\\mathcal{T}'}$ are both updated during the inner update, which also ensures that the computation graph of $\\theta_{\\mathcal{F}'}$ with respect to $\\theta_{\\mathcal{G}}$ is preserved. This preserved graph is essential for optimizing $\\theta_{\\mathcal{G}}$ during the outer update.", + "bbox": [ + 511, + 396, + 905, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.2. Outer Update", + "text_level": 1, + "bbox": [ + 511, + 463, + 656, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The primary objective of the outer update is to evaluate and refine the fusion guidance capability of $\\mathcal{G}$ , specifically by strengthening the influence of the loss function $\\mathcal{L}_f$ in steering the fusion module $\\mathcal{F}$ . In the framework diagram, this stage is represented in the lower part of the purple region. The modules $\\mathcal{F}'$ and $\\mathcal{T}'$ , derived from the inner update, represent the current guidance capacity of $\\mathcal{G}$ . In an ideal scenario, the optimal fusion loss should enhance the performance of the downstream task on the fused image. In this stage, the meta-test set $\\{I_a^{mts}, I_b^{mts}\\}$ is employed. The parameters $\\theta_{\\mathcal{G}}$ are subsequently updated using the task loss $\\mathcal{L}_t$ , which is computed by $\\mathcal{F}'$ and $\\mathcal{T}'$ :", + "bbox": [ + 511, + 481, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathcal {G}} = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\frac {\\partial \\mathcal {L} _ {t} \\left(I _ {f} ^ {m t s}\\right)}{\\partial \\theta_ {\\mathcal {G}}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 667, + 905, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{f}^{mts} = \\mathcal{F}'(I_{a}^{mts}, I_{b}^{mts})$ , and the gradient $\\partial \\mathcal{L}_t / \\partial \\theta_G$ can be calculated as:", + "bbox": [ + 511, + 704, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L} _ {t}}{\\partial \\theta_ {\\mathcal {G}}} = \\frac {\\partial \\mathcal {L} _ {t}}{\\partial \\theta_ {\\mathcal {F} ^ {\\prime}}} * \\left(- \\eta_ {\\mathcal {F} ^ {\\prime}} \\frac {\\partial^ {2} \\mathcal {L} _ {f} \\left(I _ {a} ^ {m t r} , I _ {b} ^ {m t r} , I _ {f} ^ {m t r} ; \\theta_ {\\mathcal {G}}\\right)}{\\partial \\theta_ {\\mathcal {F}} \\partial \\theta_ {\\mathcal {G}}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 738, + 906, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. (6) holds because the task loss $\\mathcal{L}_t$ is determined by $I_{f}^{mts}$ , which in turn relies on $\\theta_{\\mathcal{F}'}$ . The optimization of $\\theta_{\\mathcal{G}}$ via $\\mathcal{L}_t$ is realized by preserving the computational relationship between $\\theta_{\\mathcal{F}'}$ and $\\theta_{\\mathcal{G}}$ throughout the inner update. The updated $\\mathcal{G}$ module gains the ability to generate enhanced fusion loss functions, enabling the fusion module to integrate relevant information from the source images more efficiently into the fused output for downstream tasks.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7460", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 TDFusion Training Algorithm" + ], + "code_body": "Require: Training set $\\{I_a^{ftr}, I_b^{ftr}\\}$ with size $N$ . Output: Thoroughly trained $\\theta_{\\mathcal{F}}, \\theta_{\\mathcal{T}}, \\theta_{\\mathcal{G}}$ . \n1: Initialize $\\theta_{\\mathcal{F}}, \\theta_{\\mathcal{T}}, \\theta_{\\mathcal{G}}$ . \n2: for epoch = 1 to L do \n3: Sample $\\{I_a^{mtr}, I_b^{mtr}\\}$ and $\\{I_a^{mts}, I_b^{mts}\\}$ . \n4: for step = 1 to M do \n % Inner update: apply G. \n5: Sample $(I_a^{mtr}, I_b^{mtr})$ and get $(I_f^{mtr}, y^{mtr})$ . \n6: Compute $\\theta_{\\mathcal{F}'}$ and $\\theta_{\\mathcal{T}'}$ by Eq. (4) and Eq. (5). \n% Outer update: optimize G. \n7: Sample $(I_a^{mts}, I_b^{mts})$ and get $(I_f^{mts}, y^{mts})$ . \n8: Update $\\theta_{\\mathcal{G}}$ by Eq. (6). \n9: end for \n10: for step = 1 to N do \n % Fusion update: optimize F and T. \n11: Sample $(I_a^{ftr}, I_b^{ftr})$ and get $(I_f^{ftr}, y^{ftr})$ . \n12: Update $\\theta_{\\mathcal{F}}$ and $\\theta_{\\mathcal{T}}$ by Eq. (8) and Eq. (9). \n13: end for \n14: end for", + "bbox": [ + 94, + 107, + 486, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Learning of Fusion Network", + "text_level": 1, + "bbox": [ + 89, + 424, + 344, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The alternating inner and outer update iterations form a flexible and effective mechanism to refine $\\mathcal{G}$ in response to the evolving state of $\\mathcal{F}$ . After refining $\\mathcal{G}$ , it is utilized to further improve the training of $\\mathcal{F}$ . This stage, denoted in blue in the diagrams, involves processing the fusion training set images $\\{I_a^{ftr}, I_b^{ftr}\\}$ . Both $\\mathcal{F}$ and $\\mathcal{T}$ are updated through the application of the fusion loss $\\mathcal{L}_f$ and task loss $\\mathcal{L}_t$ :", + "bbox": [ + 89, + 446, + 483, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathcal {F}} = \\theta_ {\\mathcal {F}} - \\eta_ {\\mathcal {F}} \\frac {\\partial \\mathcal {L} _ {f} \\left(I _ {a} ^ {f t r} , I _ {b} ^ {f t r} , I _ {f} ^ {f t r} ; \\theta_ {\\mathcal {G}}\\right)}{\\partial \\theta_ {\\mathcal {F}}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 558, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathcal {T}} = \\theta_ {\\mathcal {T}} - \\eta_ {\\mathcal {T}} \\frac {\\partial \\mathcal {L} _ {t} \\left(I _ {f} ^ {f t r}\\right)}{\\partial \\theta_ {\\mathcal {T}}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 607, + 480, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After completing several training sessions on the fusion network, the focus then shifts back to the learning phase of the fusion generation module. The fusion framework evolves through a series of alternating phases, with each phase finetuning the fusion loss based on the fusion network's current state. Such alternating phases ensure that the fusion network consistently applies the most suitable fusion loss during its progression. Ultimately, this results in the development of a highly efficient fusion network, optimized for peak performance. The complete training procedure is detailed in Algorithm 1.", + "bbox": [ + 89, + 645, + 483, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Network Architecture", + "text_level": 1, + "bbox": [ + 89, + 818, + 295, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TDFusion is composed of three modules: the fusion network, the downstream task network, and the loss generation network. The fusion network shares the same architecture as [5], a lightweight model built upon the Restormer", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Block (RTB) [74]. An adaptive fusion module is incorporated into this network to facilitate feature integration. Fig. 1 illustrates the structure of the loss generation module. It also employs the Restormer Block (RTB) [74] as its primary component, receiving inputs $\\{I_a, I_b\\}$ . After applying Softmax( $\\cdot$ ), the final output guarantees $w_a^{ij} + w_b^{ij} = 1$ for each pixel. This design ensures that the fused image meets the similarity constraints and removes the reliance on initialization within the loss generation module. The architecture of the downstream task network $\\mathcal{T}(\\cdot)$ depends on the specific task. For learning of loss generation module, we chose the most lightweight models of SegFormer [9] and YOLOv8 [19] for semantic segmentation and object detection, respectively.", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.7. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 511, + 311, + 707, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better understand the weighting mechanism of the loss generation module $\\mathcal{G}$ , we investigate the optimization procedure of $\\mathcal{G}$ , which generates the weights $\\{w_{a}, w_{b}\\}$ , denoted as $\\theta_{\\mathcal{G}}$ . For clarity, we rewrite Eq. (2) as follows:", + "bbox": [ + 511, + 333, + 906, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {f} ^ {i n t} = \\left[ w _ {a} \\odot \\left(I _ {a} - I _ {f}\\right) \\odot \\left(I _ {a} - I _ {f}\\right) \\right. \\\\ \\left. + w _ {b} \\odot \\left(I _ {b} - I _ {f}\\right) \\odot \\left(I _ {b} - I _ {f}\\right) \\right] \\times \\frac {1}{H W} \\\\ = \\left[ \\mathcal {G} \\left(I _ {a}, I _ {b}; \\theta_ {\\mathcal {G}}\\right) \\odot \\left(I _ {a} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\right. \\tag {10} \\\\ \\odot \\left(I _ {a} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\\\ + \\left(1 - \\mathcal {G} \\left(I _ {a}, I _ {b}; \\theta_ {\\mathcal {G}}\\right)\\right) \\odot \\left(I _ {b} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\\\ \\left. \\odot \\left(I _ {b} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\right] \\times \\frac {1}{H W}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 401, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $w_{a}, w_{b} \\in \\mathbb{R}^{H \\times W}$ , $I_{a}, I_{b} \\in \\mathbb{R}^{H \\times W}$ , and $\\odot$ denotes the element-wise multiplication operation. Let $\\Omega'$ be the set $\\{\\theta_{\\mathcal{F}'}', \\theta_{\\mathcal{T}''}\\}$ , this leads to the following expression:", + "bbox": [ + 511, + 534, + 906, + 582 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta_ {\\mathcal {G}} = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\frac {\\partial \\mathcal {L} _ {t} ^ {m t s} (\\Omega^ {'} (\\theta_ {\\mathcal {G}}))}{\\partial \\theta_ {\\mathcal {G}}} \\\\ = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\frac {\\partial \\mathcal {L} _ {t} ^ {m t s} (\\Omega^ {'} (\\theta_ {\\mathcal {G}}))}{\\partial \\Omega^ {'}} \\frac {\\partial \\Omega^ {'} (\\theta_ {\\mathcal {G}})}{\\partial \\theta_ {\\mathcal {G}}} \\\\ = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\eta_ {\\mathcal {F} ^ {\\prime}} \\underbrace {\\frac {\\partial \\mathcal {L} _ {t} ^ {m t s} \\left(\\Omega^ {'} \\left(\\theta_ {\\mathcal {G}}\\right)\\right)}{\\partial \\Omega^ {\\prime}}} _ {\\text {(a)}} \\times \\frac {\\partial \\mathcal {G} \\left(I _ {a} , I _ {b} ; \\theta_ {\\mathcal {G}}\\right)}{\\theta_ {\\mathcal {G}}} \\\\ \\times \\underbrace {\\left[ \\left(I _ {a} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) \\odot \\left(I _ {a} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) \\right.} _ {(b)} \\tag {11} \\\\ \\underbrace {- \\left(I _ {b} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) \\odot \\left(I _ {b} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) ]} _ {(b)} \\\\ = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\eta_ {\\mathcal {F} ^ {\\prime}} \\mathbf {G} \\times \\frac {\\partial \\mathcal {G} \\left(I _ {a} , I _ {b} ; \\theta_ {\\mathcal {G}}\\right)}{\\theta_ {\\mathcal {G}}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 589, + 903, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{G}$ denotes the inner product between two gradients: (a) the first one is derived from the task loss using a meta-testing set, and (b) the second is calculated from the fusion loss based on a meta-training set. Consequently, the", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7461", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65a1c92046391bc694d653a9a75e71f8a8ba94d6cfdba5037aa2a9bc3eb03716.jpg", + "image_caption": [ + "Figure 2. Visual comparison of fusion results. The cases are \"01258N\" in MSRS dataset, \"00122\" in FMB dataset, \"00449\" in M3FD dataset and \"200304\" in LLVIP dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimization of the module $\\theta_{\\mathcal{G}}$ is driven by the task loss, with the objective of preserving task-specific information throughout the fusion process.", + "bbox": [ + 89, + 454, + 482, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 516, + 215, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 542, + 171, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Setup. In our experiments, the epoch number $L$ and the training iterations of the loss generation module $M$ are set to 50 and 200. The learning iterations for the fusion network $N$ depend on the size of the dataset. We use the Adam optimizer with a learning rate of 1e-4, a batch size of 2, and a hyperparameter $\\alpha$ set to 1. All experiments are conducted on a PC with a single NVIDIA RTX 3090 GPU.", + "bbox": [ + 89, + 566, + 482, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics and Comparison Methods. The advanced fusion methods compared in our study include TarDAL [32], SegMIF [34], MURF [69], EMMA [86], DCINN [61], MRFS [76], and TIMFusion [40]. The fusion performance is evaluated using metrics including entropy (EN), spatial frequency (SF), sum of correlation differences (SCD), visual information fidelity (VIF), $Q^{AB / F}$ and structural similarity index metric (SSIM).", + "bbox": [ + 89, + 672, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Split. We use four datasets annotated for downstream tasks, including MSRS [58], FMB [34], M3FD [32], and LLVIP [17]. MSRS contains 1083/361 image pairs for training/test, and FMB contains 1220/280 pairs for training/test. We follow the splits of the original papers for both datasets. M3FD dataset consists of 4200 pairs of images for detection, with 300 pairs designated for fusion evalu-", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ation. The 4200 detection images are split into 3150 for training and 1050 for testing, ensuring that the 300 pairs for fusion evaluation are included within the detection test set. These 300 fusion images are then employed to assess the fusion performance. The original LLVIP dataset contains 12025/3463 image pairs as training/test set. Due to its large size, we select every 10th image to form our training and test sets, resulting in 1203/347 image pairs for training and testing. Our splits for M3FD and LLVIP will be available.", + "bbox": [ + 511, + 454, + 906, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Fusion Experiments", + "text_level": 1, + "bbox": [ + 511, + 604, + 705, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 2 presents a visual comparison of different methods. The fused images generated by TDFusion excel in preserving discriminative details, achieving balanced brightness, and maintaining clear object contours. It effectively preserves the target features from the infrared images and the background details from the visible images, resulting in fused images that are more natural and exhibit greater clarity across different environments. These results highlight the advantages of TDFusion in detail preservation and visual performance. More results are available in the supplementary material. Tab. 1 presents the quantitative comparison of fusion over four datasets. TDFusion outperforms other methods across most metrics. This suggests that TDFusion not only enhances image details but also provides consistent fusion results across various scenarios. Compared to other methods, TDFusion shows superior adaptability and robustness, particularly in handling diverse image characteristics and challenging fusion tasks.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7462", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2962883789f32476e8578fb3a0036fd2d65c0d14d779d891fbce45181a537779.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of Infrared-visible image fusion. The red and blue markers represent the best and second-best values." + ], + "table_footnote": [], + "table_body": "
Infrared-visible Image Fusion on MSRS [58] DatasetInfrared-visible Image Fusion on FMB [34] Dataset
EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑
TarDAL [32]5.285.980.710.210.180.47TarDAL [32]6.636.941.030.280.290.74
SegMIF [34]5.9511.101.570.440.630.55SegMIF [34]6.8313.691.720.390.650.60
MURF [69]5.0410.491.020.220.370.60MURF [69]6.3713.881.340.220.370.68
EMMA [86]6.7311.561.620.490.640.70EMMA [86]6.7715.001.500.420.650.72
DCINN [61]6.0010.511.490.410.570.52DCINN [61]6.4711.471.390.380.590.74
MRFS [76]7.008.861.420.370.490.55MRFS [76]6.7812.421.240.380.620.73
TIMFusion [40]6.279.671.340.320.480.68TIMFusion [40]6.5112.231.240.350.590.73
TDFusion (Ours)6.7411.301.860.500.670.70TDFusion (Ours)6.8614.161.760.430.680.75
Infrared-visible Image Fusion on M3FD [32] DatasetInfrared-visible Image Fusion on LLVIP [17] Dataset
EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑
TarDAL [32]6.877.631.290.270.300.71TarDAL [32]6.327.421.040.270.220.58
SegMIF [34]6.8514.141.720.370.600.59SegMIF [34]6.6815.461.380.400.660.57
MURF [69]6.5012.551.460.210.320.64MURF [69]6.1315.080.960.210.310.57
EMMA [86]6.9215.231.490.380.590.69EMMA [86]7.3515.371.570.410.640.66
DCINN [61]6.5911.211.460.340.510.72DCINN [61]6.9813.341.430.380.520.64
MRFS [76]6.9412.071.260.340.550.70MRFS [76]6.8311.041.230.310.420.64
TIMFusion [40]6.7512.311.370.350.530.70TIMFusion [40]6.5813.521.140.330.460.64
TDFusion (Ours)6.9914.491.830.410.650.72TDFusion (Ours)7.3616.381.750.460.700.67
", + "bbox": [ + 94, + 109, + 906, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5baf043834cf4bf3ae738fc93d4d04004447215b3caaee616a34c56078459c4.jpg", + "image_caption": [ + "Figure 3. Visual comparison for Semantic Segmentation. The cases are \"00726N\" in MSRS dataset and \"01438\" in FMB dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 387, + 903, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Downstream Applications", + "text_level": 1, + "bbox": [ + 89, + 561, + 326, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section validates the adaptability of fusion methods to downstream tasks. For a fair comparison, we adopt SegFormer [9] and YOLOv8 [19] as backbones and retrain the task networks for each fusion method over 300 epochs to evaluate their adaptability to semantic segmentation and object detection. Fig. 3 and Fig. 4 present the visual comparisons of semantic segmentation and object detection, respectively. TDFusion outperforms in image detail retention, edge clarity, and object recognition, effectively identifying and segmenting objects. For semantic segmentation, the generated maps clearly distinguish different class regions, closely matching the ground truth. In object detection, the fused images exhibit more precise boundary localization for salient objects. This indicates that TDFusion maintains a better balance between fine details and overall context. More results can be found in the supplementary material.", + "bbox": [ + 88, + 590, + 482, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows the performance comparison across different methods in semantic segmentation and object detection. TDFusion outperforms other methods on most metrics, particularly in mIoU and mAP. This demonstrates that TDFu", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2125701b7d516e5d4c63c9b132bb8e32698e56df81bc8751e98f5cf37e4999ac.jpg", + "table_caption": [ + "Table 2. Performance comparison of downstream applications. The red and blue markers represent the best and second-best." + ], + "table_footnote": [], + "table_body": "
MethodsSemantic SegmentationObject Detection
MSRSFMBM3FDLLVIP
mAccmIoUmAccmIoUmAP50mAP75AP50AP75
Infrared83.2369.4958.8551.9879.1253.0596.0372.07
Visible83.4473.7665.1257.9682.2154.8291.7848.66
TarDAL81.9371.3562.8655.3383.1656.3993.7962.71
SegMIF85.7374.2565.9758.4183.6158.2393.9566.45
MURF85.0374.0864.1056.9680.5854.2294.2468.04
EMMA85.9974.4862.4556.2883.7156.9194.0066.21
DCINN84.1174.3561.0954.8182.6957.3794.9268.34
MRFS84.7674.5061.9355.7183.2857.7493.0367.21
TIMFusion83.6773.5863.7057.2483.2256.0893.7661.33
TDFusion86.0475.0967.1760.5086.2759.7195.0069.18
", + "bbox": [ + 514, + 592, + 906, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sion effectively enhances the quality of fused images. It also improves the downstream task performance, especially in terms of accuracy and robustness in complex scenarios. Class-wise results are provided in the supplementary material.", + "bbox": [ + 511, + 748, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Task-driven Learnable Loss", + "text_level": 1, + "bbox": [ + 511, + 832, + 764, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our framework incorporates a learnable fusion loss that models the preferences of downstream tasks for information from source images. The models trained on FMB dataset", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7463", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5dd3227b95ffe4277b7733b909d3963d469aa7be65f04d769127fe7da6e3408e.jpg", + "image_caption": [ + "Figure 4. Visual comparison for Object Detection. The cases are \"02236\" in M3FD dataset and \"210145\" in LLVIP dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc24fe26ef625184c016f17bfbe192a79954665885b86321e4b8a355c403d80c.jpg", + "image_caption": [ + "Figure 5. Visualisation of learnable loss for downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 93, + 263, + 285, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d562ac6c344a11d3f1bdb8b4f9901cc2cfb66ba113513b9c6660c1871c99d2bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 263, + 480, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and LLVIP dataset labeled as $SS$ and $OD$ , are evaluated on MSRS dataset to simulate performance in unknown scenes, as shown in Fig. 5. The results demonstrate that the fusion model adaptively selects information from infrared and visible images to satisfy task requirements. In semantic segmentation, the model combines scene structure and texture, prioritizing boundaries. This improves segmentation under varying lighting conditions. Fusion weights $\\{w_{a}^{SS}, w_{b}^{SS}\\}$ indicate a preference for visible details and infrared advantages in low-light conditions. In object detection, the model focuses on edge and contrast information, especially for instances like pedestrians and vehicles. Higher fusion weights are assigned to bright regions in infrared images, enhancing target detection in low-light conditions. Fusion weights $\\{w_{a}^{OD}, w_{b}^{OD}\\}$ reflect this preference. Comparison of fusion losses across tasks reveals distinct differences, especially in highlighted regions, confirming that the model adapts to task-specific requirements by selecting the most relevant information from multimodal images. More results are provided in the supplementary material.", + "bbox": [ + 89, + 443, + 483, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 756, + 253, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To thoroughly evaluate the performance of our proposed algorithm, we conduct a series of ablation experiments on FMB dataset, and the detailed results are shown in Tab. 3. In Exp. I, we exclude the learnable fusion loss by fixing $w_{a}$ and $w_{b}$ to $1/2$ . In Exp. II, we omit the gradient loss from the loss function. In Exp. III, we also allow the fusion module parameters to be jointly optimized by both the task loss $\\mathcal{L}_{t}$ and fusion loss $\\mathcal{L}_{f}$ . In Exp. IV, we exclude the fusion", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65243a8d45bb361ca787183f5d3a3f8c462e1971257f7373dbf0b1989fe7c8a7.jpg", + "table_caption": [ + "Table 3. Ablation experiment of fusion. The red denotes the best." + ], + "table_footnote": [], + "table_body": "
Ablation Studies of fusion on FMB Dataset
ConfigurationsENSFSCDVIF\\(Q^{AB/F}\\)SSIM
Ifix \\(w_{a}\\) and \\(w_{b}\\) as 1/26.6013.731.580.390.600.72
IIw/o \\(\\mathcal{L}_{f}^{grad}\\)6.7711.651.630.370.640.73
III\\(\\theta_{\\mathcal{F}}\\) influenced by \\(\\mathcal{L}_{t}\\)6.8013.851.700.410.660.73
IVw/o Fusion learning6.8214.071.720.410.670.72
V\\(I_{f}=w_{a}*I_{a}+w_{b}*I_{b}\\)6.7511.491.650.380.620.73
Ours6.8614.161.760.430.680.75
", + "bbox": [ + 517, + 282, + 906, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "module's dedicated learning phase and update it during the outer update of the loss module, which tests the impact of the fusion module's training schedule. In Exp. V, we replace our fusion method with a discriminative approach. The performance decline observed across different configurations confirms the rationality and effectiveness of our proposed method. Visualization and more analysis are provided in the supplementary material.", + "bbox": [ + 511, + 397, + 906, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 532, + 633, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To overcome the limitations of predefined fusion losses, which often fail to effectively guide the fusion process for downstream tasks, we propose a meta-learning-based framework for task-guided fusion. This framework includes a loss generation module that outputs the parameters of the learnable fusion loss. The module is updated using a meta-learning approach, which alternates between inner and outer loop steps to enhance its ability to guide the fusion network. Under varying fusion conditions, this module generates the optimal fusion loss for the downstream task. This enables the fusion network to produce fused images that minimize the task-specific loss. The theoretical analysis explains how the downstream task loss guides the fusion loss in our framework. Experiments using four publicly available fusion datasets and downstream tasks including semantic segmentation and object detection, demonstrate the effectiveness of our approach.", + "bbox": [ + 511, + 558, + 906, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 829, + 671, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work has been supported by the National Natural Science Foundation of China under Grant 12201497 and 12371512.", + "bbox": [ + 511, + 854, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7464", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hessah Albanwan, Rongjun Qin, and Yang Tang. Image fusion in remote sensing: An overview and meta-analysis. Photogrammetric Engineering & Remote Sensing, 90(12): 755-775, 2024. 1", + "[2] Antreas Antoniou and Amos J. Storkey. Learning to learn by self-critique. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), pages 9936-9946, 2019. 2", + "[3] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yukun Cui, Chunxia Zhang, Zhenbo Guo, and Yongjun Wang. Simultaneous automatic picking and manual picking refinement for first-break. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1", + "[4] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Baisong Jiang, Lilun Deng, Yukun Cui, Shuang Xu, and Chunxia Zhang. Deep unfolding multi-modal image fusion network via attribution analysis. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2", + "[5] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yichen Wu, Lilun Deng, Yukun Cui, Baisong Jiang, and Shuang Xu. Refusion: Learning image fusion from reconstruction with learnable loss via meta-learning. International Journal of Computer Vision, pages 1-21, 2024. 3, 5", + "[6] Haowen Bai, Jiangshe Zhang, Zixiang Zhao, Lilun Deng, Yukun Cui, and Shuang Xu. Retinex-mef: Retinex-based glare effects aware unsupervised multi-exposure image fusion. arXiv preprint arXiv:2503.07235, 2025. 2", + "[7] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 9445-9454. IEEE, 2021. 2", + "[8] Yanpeng Cao, Dayan Guan, Weilin Huang, Jiangxin Yang, Yanlong Cao, and Yu Qiao. Pedestrian detection with unsupervised multispectral feature learning using deep neural networks. Information Fusion, 46:206-217, 2019. 1", + "[9] Bo Cheng, Xiang Li, Yujie Wei, Cheng Huang, Xiaoyong Zhang, Yandong Jiang, Tianyu Zhang, Na Xu, Shuai Yu, Xinxin Zhan, et al. Segformer: Simple and efficient design for semantic segmentation with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12022-12032, 2021. 4, 5, 7", + "[10] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3333-3348, 2021. 2", + "[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the International conference on machine learning (ICML), pages 1126–1135, 2017. 2", + "[12] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In Proceedings of the International conference on machine learning (ICML), pages 1920-1930, 2019. 2", + "[13] Qishen Ha, Kohei Watanabe, Takumi Karasawa, Yoshitaka Ushiku, and Tatsuya Harada. Mfnet: Towards real-time se" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mantic segmentation for autonomous vehicles with multispectral scenes. In IROS, pages 5108-5115. IEEE, 2017. 1", + "[14] Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Task-driven super resolution: Object detection in low-resolution images. In Neural Information Processing: 28th International Conference (ICONIP), pages 387-395. Springer, 2021. 1", + "[15] Rein Houthooft, Yuhua Chen, Phillip Isola, Bradley C. Stadie, Filip Wolski, Jonathan Ho, and Pieter Abbeel. Evolved policy gradients. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), pages 5405-5414, 2018. 2", + "[16] Zhanbo Huang, Jinyuan Liu, Xin Fan, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Reconet: Recurrent correction network for fast and efficient multi-modality image fusion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 539-555. Springer, 2022. 2", + "[17] Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible-infrared paired dataset for low-light vision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3496-3504, 2021. 6, 7", + "[18] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 4369-4376, 2020. 1", + "[19] Glenn Jocher. Ultralytics YOLOv8. https://github.com/ultralytics, 2023.4,5,7", + "[20] Chenglong Li, Chengli Zhu, Yan Huang, Jin Tang, and Liang Wang. Cross-modal ranking with soft consistency and noisy labels for robust RGB-T tracking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 831-847. Springer, 2018. 1", + "[21] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 1, 2", + "[22] Huafeng Li, Yueliang Cen, Yu Liu, Xun Chen, and Zhengtao Yu. Different input resolutions and arbitrary output resolution: A meta learning-based deep framework for infrared and visible image fusion. IEEE Transactions on Image Processing, 30:4070-4083, 2021. 3", + "[23] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 1, 2", + "[24] Siyuan Li, Iago Bruno Araujo, Wenqi Ren, Zhangyang Wang, Eric K. Tokuda, Roberto Hirata Junior, Roberto Marcondes Cesar Junior, Jiawan Zhang, Xiaojie Guo, and Xiaochun Cao. Single image deraining: A comprehensive benchmark analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3838-3847. Computer Vision Foundation / IEEE, 2019. 1", + "[25] Xiaoling Li, Yanfeng Li, Houjin Chen, Yahui Peng, and Pan Pan. Ccafusion: cross-modal coordinate attention network for infrared and visible image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7465", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 2", + "[27] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 1", + "[28] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 1", + "[29] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 719-735. Springer, 2022. 2", + "[30] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 2", + "[31] Jinyuan Liu, Xin Fan, Ji Jiang, Risheng Liu, and Zhongxuan Luo. Learning a deep multi-scale feature ensemble and an edge-attention guidance for image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 32(1): 105–119, 2021. 1", + "[32] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5792-5801, 2022. 1, 2, 4, 6, 7", + "[33] Jinyuan Liu, Jingjie Shang, Risheng Liu, and Xin Fan. Attention-guided global-local adversarial learning for detail-preserving multi-exposure image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5026-5040, 2022. 2", + "[34] Jinyuan Liu, Zhu Liu, Guanyao Wu, Long Ma, Risheng Liu, Wei Zhong, Zhongxuan Luo, and Xin Fan. Multi-interactive feature learning and a full-time multi-modality benchmark for image fusion and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8115–8124, 2023. 1, 2, 6, 7", + "[35] Jinyuan Liu, Guanyao Wu, Junsheng Luan, Zhiying Jiang, Risheng Liu, and Xin Fan. Holoco: Holistic and local contrastive learning network for multi-exposure image fusion. Information Fusion, 95:237-249, 2023. 2", + "[36] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, 132(5):1748-1775, 2024. 2", + "[37] Jinyuan Liu, Guanyao Wu, Zhu Liu, Di Wang, Zhiying Jiang, Long Ma, Wei Zhong, and Xin Fan. Infrared and visible image fusion: From data compatibility to task adaption. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Risheng Liu, Zhu Liu, Jinyuan Liu, and Xin Fan. Searching a hierarchically aggregated fusion architecture for fast multi-modality image fusion. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 1600-1608. ACM, 2021. 3", + "[39] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5953-5969, 2022. 1", + "[40] Risheng Liu, Zhu Liu, Jinyuan Liu, Xin Fan, and Zhongxuan Luo. A task-guided, implicitly-searched and metainitialized deep model for image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2, 3, 6, 7", + "[41] ShuMin Liu, Jiajia Chen, and Susanto Rahardja. A new multi-focus image fusion algorithm and its efficient implementation. IEEE Transactions on Circuits and Systems for Video Technology, 30(5):1374-1384, 2020. 1", + "[42] Zhu Liu, Jinyuan Liu, Guanyao Wu, Long Ma, Xin Fan, and Risheng Liu. Bi-level dynamic learning for jointly multimodality image fusion and beyond. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1240-1248, 2023. 2", + "[43] Zhu Liu, Jinyuan Liu, Benzhuang Zhang, Long Ma, Xin Fan, and Risheng Liu. Paif: Perception-aware infrared-visible image fusion for attack-tolerant semantic segmentation. In Proceedings of the 31st ACM international conference on multimedia, pages 3706-3714, 2023.", + "[44] Zhu Liu, Jinyuan Liu, Guanyao Wu, Zihang Chen, Xin Fan, and Risheng Liu. Searching a compact architecture for robust multi-exposure image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6224-6237, 2024. 2", + "[45] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information Fusion, 45:153-178, 2019. 1", + "[46] Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusion: A generative adversarial network for infrared and visible image fusion. Information Fusion, 48:11-26, 2019. 2", + "[47] Jiayi Ma, Pengwei Liang, Wei Yu, Chen Chen, Xiaojie Guo, Jia Wu, and Junjun Jiang. Infrared and visible image fusion via detail preserving adversarial learning. Information Fusion, 54:85-98, 2020.", + "[48] Jiayi Ma, Han Xu, Junjun Jiang, Xiaoguang Mei, and Xiao-Ping (Steven) Zhang. Ddcgan: A dual-discriminator conditional generative adversarial network for multi-resolution image fusion. IEEE Transactions on Image Processing, 29: 4980-4995, 2020. 2", + "[49] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5637-5646, 2022. 1", + "[50] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7466", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Seonghyun Park, An Gia Vien, and Chul Lee. Cross-modal transformers for infrared and visible image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 34(2):770-785, 2023. 1", + "[52] Yanting Pei, Yaping Huang, Qi Zou, Yuhang Lu, and Song Wang. Does haze removal help cnn-based image classification? In Proceedings of the European conference on computer vision (ECCV), pages 682-697, 2018. 1", + "[53] Xinran Qin, Yuhui Quan, Tongyao Pang, and Hui Ji. Ground-truth free meta-learning for deep compressive sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9947-9956, 2023. 2", + "[54] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In Proceedings of the International conference on machine learning (ICML), pages 4334-4343, 2018. 2", + "[55] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2019. 2", + "[56] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1, 2", + "[57] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 4", + "[58] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Infromation Fusion, 83-84:79-92, 2022. 6, 7", + "[59] Wei Tang, Fazhi He, Yu Liu, Yansong Duan, and Tongzhen Si. Datfuse: Infrared and visible image fusion via dual attention transformer. IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159-3172, 2023. 1", + "[60] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 3508-3515. ijcai.org, 2022. 2", + "[61] Wu Wang, Liang-Jian Deng, Ran Ran, and Gemine Vivone. A general paradigm with detail-preserving conditional invertible network for image fusion. International Journal of Computer Vision, 132(4):1029-1054, 2024. 6, 7", + "[62] Zeyu Xiao and Xinchao Wang. Event-based video superresolution via state space models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1", + "[63] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 2409-2418, 2024. 1" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 217-235. Springer, 2024. 1", + "[65] Han Xu, Jiayi Ma, Zhuliang Le, Junjun Jiang, and Xiaojie Guo. Fusiondn: A unified densely connected network for image fusion. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 12484-12491, 2020. 2", + "[66] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 2", + "[67] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 2", + "[68] Han Xu, Jiayi Ma, Jiteng Yuan, Zhuliang Le, and Wei Liu. Rfnet: Unsupervised network for mutually reinforcing multimodal image registration and fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19647-19656, 2022. 2", + "[69] Han Xu, Jiteng Yuan, and Jiayi Ma. Murf: Mutually reinforcing multi-modal image registration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 6, 7", + "[70] Shuang Xu, Ouafa Amira, Junmin Liu, Chun-Xia Zhang, Jiangshe Zhang, and Guanghai Li. HAM-MFN: hyperspectral and multispectral image multiscale fusion network with RAP loss. IEEE Transactions on Geoscience and Remote Sensing, 58(7):4618-4628, 2020. 1", + "[71] Shuang Xu, Lizhen Ji, Zhe Wang, Pengfei Li, Kai Sun, Chunxia Zhang, and Jiangshe Zhang. Towards reducing severe defocus spread effects for multi-focus image fusion via an optimization based strategy. IEEE Transactions Computational Imaging, 6:1561-1570, 2020. 1", + "[72] Yong Yang, Jiaxiang Liu, Shuying Huang, Weiguo Wan, Wenying Wen, and Juwei Guan. Infrared and visible image fusion via texture conditional generative adversarial network. IEEE Transactions on Circuits and Systems for Video Technology, 31(12):4771-4783, 2021. 1", + "[73] Xunpeng Yi, Linfeng Tang, Hao Zhang, Han Xu, and Jiayi Ma. Diff-if: Multi-modality image fusion via diffusion model with fusion knowledge prior. Information Fusion, 110:102450, 2024. 2", + "[74] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728-5739, 2022. 5", + "[75] Hao Zhang and Jiayi Ma. Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion. International Journal of Computer Vision, 129(10):2761-2785, 2021. 2", + "[76] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Com" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7467", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "puter Vision and Pattern Recognition (CVPR), pages 26974-26983, 2024. 1, 2, 3, 6, 7", + "[77] Yu Zhang, Yu Liu, Peng Sun, Han Yan, Xiaolin Zhao, and Li Zhang. IFCNN: A general image fusion framework based on convolutional neural network. Information Fusion, 54: 99-118, 2020. 1, 2", + "[78] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13955-13965, 2023. 1, 2, 3", + "[79] Yangyang Zhao, Qingchun Zheng, Peihao Zhu, Xu Zhang, and Wenpeng Ma. Tufusion: A transformer-based universal fusion algorithm for multimodal images. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1", + "[80] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Jiangshe Zhang, and Pengfei Li. Didfuse: Deep image decomposition for infrared and visible image fusion. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 970-976, 2020. 2", + "[81] Zixiang Zhao, Shuang Xu, Jiangshe Zhang, Chengyang Liang, Chunxia Zhang, and Junmin Liu. Efficient and model-based infrared and visible image fusion via algorithm unrolling. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1186-1196, 2022. 2", + "[82] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5906–5916, 2023. 1, 2", + "[83] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5906–5916, 2023. 1, 2, 3, 4", + "[84] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. DDFM: denoising diffusion model for multi-modality image fusion. CoRR, abs/2303.06840, 2023. 2", + "[85] Zixiang Zhao, Jiang-She Zhang, Haowen Bai, Yicheng Wang, Yukun Cui, Lilun Deng, Kai Sun, Chunxia Zhang, Junmin Liu, and Shuang Xu. Deep convolutional sparse coding networks for interpretable image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2369-2377, 2023. 2", + "[86] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25912-25921, 2024. 6, 7", + "[87] Huabing Zhou, Wei Wu, Yanduo Zhang, Jiayi Ma, and Haibin Ling. Semantic-supervised infrared and visible image" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fusion via a dual-discriminator generative adversarial network. IEEE Transactions on Multimedia, 25:635-648, 2023. 1, 2", + "[88] Man Zhou, Naishan Zheng, Xuanhua He, Danfeng Hong, and Jocelyn Chanussot. Probing synergistic high-order interaction for multi-modal image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[89] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 1, 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 258 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7468", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Task-driven Image Fusion with Learnable Fusion Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_model.json b/2025/Task-driven Image Fusion with Learnable Fusion Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9688eed883bad439f47a05c78a96f267038c8be4 --- /dev/null +++ b/2025/Task-driven Image Fusion with Learnable Fusion Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_model.json @@ -0,0 +1,2534 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.224, + 0.131, + 0.776, + 0.152 + ], + "angle": 0, + "content": "Task-driven Image Fusion with Learnable Fusion Loss" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.18, + 0.751, + 0.218 + ], + "angle": 0, + "content": "Haowen Bai\\(^{1}\\) Jiangshe Zhang\\(^{1*}\\) Zixiang Zhao\\(^{2*}\\) Yichen Wu\\(^{3}\\) Lilun Deng\\(^{1}\\) Yukun Cui\\(^{1}\\) Tao Feng\\(^{4}\\) Shuang Xu\\(^{5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.219, + 0.8, + 0.256 + ], + "angle": 0, + "content": "\\(^{1}\\)Xi'an Jiaotong University \\(^{2}\\)ETH Zürich \\(^{3}\\)City University of Hong Kong \\(^{4}\\)Tsinghua University \\(^{5}\\)Northwestern Polytechnical University" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.259, + 0.595, + 0.273 + ], + "angle": 0, + "content": "hwbai@stu.xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.308, + 0.327, + 0.324 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.486, + 0.747 + ], + "angle": 0, + "content": "Multi-modal image fusion aggregates information from multiple sensor sources, achieving superior visual quality and perceptual features compared to single-source images, often improving downstream tasks. However, current fusion methods for downstream tasks still use predefined fusion objectives that potentially mismatch the downstream tasks, limiting adaptive guidance and reducing model flexibility. To address this, we propose Task-driven Image Fusion (TD-Fusion), a fusion framework incorporating a learnable fusion loss guided by task loss. Specifically, our fusion loss includes learnable parameters modeled by a neural network called the loss generation module. This module is supervised by the downstream task loss in a meta-learning manner. The learning objective is to minimize the task loss of fused images after optimizing the fusion module with the fusion loss. Iterative updates between the fusion module and the loss module ensure that the fusion network evolves toward minimizing task loss, guiding the fusion process toward the task objectives. TDFusion's training relies entirely on the downstream task loss, making it adaptable to any specific task. It can be applied to any architecture of fusion and task networks. Experiments demonstrate TDFusion's performance through fusion experiments conducted on four different datasets, in addition to evaluations on semantic segmentation and object detection tasks. The code is available at https://github.com/HaowenBai/TDFusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.222, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Multi-modal image fusion [25, 31, 51, 59, 72, 79, 82] combines information from multiple sensors to produce a more holistic and detailed representation. Infrared images capture thermal radiation regardless of lighting conditions, while visible images provide richer texture details. Fused im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.908, + 0.583 + ], + "angle": 0, + "content": "ages enhance downstream tasks through improved information density and robustness [18, 41, 45, 62, 63, 70, 71], outperforming single-modal inputs in semantic segmentation [13, 28, 34, 56], object detection [8, 32], and other related applications [1, 3, 20, 27]. Conventional methods typically treat fusion as image restoration using unsupervised loss [76, 83, 87-89] or perceptual loss [21, 23, 77]. These approaches prioritize visual-level fusion through predefined aggregation objectives, often neglecting semantic feature extraction. This limitation hinders scene interpretation and task performance [14, 24, 52, 64]. Recent advances explore the mutual enhancement between fusion and downstream tasks [37]. By cascading the fusion network with downstream task network [39, 49], the task loss constrains the fusion learning, ensuring the fused images meet the task requirements [32, 56]. Alternatively, some methods incorporate high-level visual task features [34, 76, 78] or focus on learning optimal initializations [40] to enhance fusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.903 + ], + "angle": 0, + "content": "While integrating downstream tasks, existing frameworks still rely on predefined fusion loss terms lacking dynamic adaptation. The impact of downstream tasks remains limited due to specific combinations. Manually defined losses preserve predefined guidance, frequently overlooking task-specific requirements. This guidance imposes manually designed prior constraints on the fusion process, limiting the dynamic and adaptive influence of downstream tasks on specific image pairs. These approaches, whether incorporating task features [34, 78] or employing task losses [32, 56], still face the limitations of fixed fusion loss terms. Task-specialized networks [76] create fusion-task dependencies, restricting flexibility and limiting their applicability to various high-level vision tasks. We address these limitations through a task-driven framework with learnable loss. The fusion loss contains learnable parameters, generated by a loss generation module, and is designed to retain the intensity information of the source images for specific downstream tasks. The purpose of updating the fusion loss is to guide the fusion network in generating fused images that minimize downstream task loss," + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.242, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7457" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "thereby enhancing adaptability. Moreover, the fusion loss update relies on the downstream task loss, making it independent of any specific task or network architecture." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.141, + 0.483, + 0.502 + ], + "angle": 0, + "content": "The loss generation module produces fusion losses for subsequent fusion module updates. This complexity presents challenges to standard end-to-end training. Fortunately, meta-learning techniques, which are strategies for learning how to learn, can effectively achieve the learning objectives of the loss generation module. This involves minimizing task loss for fused images through an optimized fusion loss. Meta-learning tackles common deep-learning challenges like limited data, high computational costs, and the need for better generalization. Core optimization areas include parameter tuning [11], optimization strategies [26], and network architectures search [30]. In this paper, we draw inspiration from the Model-Agnostic Meta-Learning (MAML) approach [11] to train our loss generation module. Specifically, training the loss generation module involves two stages: inner updates and outer updates. During inner updates, the output of the loss generation module updates a surrogate fusion module without altering the original parameters. During outer updates, the fused image from the surrogate module is fed into the task network. The resulting task loss then updates the loss generation module through backpropagation. This alternating training ensures that the loss generation module consistently produces fusion losses that minimize the downstream task loss of the fused images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.793 + ], + "angle": 0, + "content": "This paper introduces TDFusion, a task-oriented fusion framework driven by downstream tasks. It consists of a fusion module, a task module, and a loss generation module that learns to optimize the fusion loss. The fusion loss incorporates intensity preferences from source images and gradient preservation, guided by downstream task loss to refine intensity preferences. This model follows the general form of fusion loss used in advanced methods [76, 83, 88, 89], ensuring adaptability to various tasks. The updates of the loss generation module are performed using a meta-learning approach, optimizing the loss generation module parameters based on task loss from the fused images after each update of the fusion module. This process ensures that the loss function guides the fusion module continuously, optimizing feature aggregation and minimizing downstream task losses. The loss generation module dynamically adjusts through alternating updates with the fusion and task modules, generating optimal fusion losses at each model state. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose TDFusion, a meta-learning-based fusion framework that leverages the loss functions of downstream tasks for training. This method promotes task-driven fusion and alleviates challenges caused by the absence of ground truth. Moreover, this framework is agnostic of specific downstream tasks or network architectures, which enhances its adaptability and flexibility." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "- Our framework includes a dynamically updated, learnable fusion loss generation module. It selectively extracts source image information, minimizing the loss in downstream tasks. This ensures optimal fusion performance while maximizing adaptability to downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- We analyze the information preferences of downstream tasks such as semantic segmentation and object detection, providing deeper insights into multi-modal high-level vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.228, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- TDFusion achieves outstanding performance in both fusion and high-level vision tasks, validated on four fusion datasets with semantic segmentation and object detection." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.289, + 0.655, + 0.305 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.315, + 0.816, + 0.331 + ], + "angle": 0, + "content": "2.1. Deep learning-based Image Fusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.338, + 0.905, + 0.685 + ], + "angle": 0, + "content": "Deep learning-based image fusion methods have revolutionized the field by exploiting the powerful feature extraction capabilities of neural networks [4, 6, 33, 35, 36, 42-44, 65, 77]. These methods are broadly categorized into discriminative and generative approaches. Discriminative methods [10, 81, 85, 88] leverage the strong reconstruction ability of neural networks to directly learn the mapping between source and fused images [21, 23, 80, 82]. Generative methods, on the other hand, model the image generation process using generative approaches, integrating source images from a distributional perspective. These include methods based on Generative Adversarial Networks [46-48, 87] and diffusion models [73, 84]. Unified fusion methods [66, 75, 89] bridge the gap between different fusion sub-tasks, incorporating strategies such as continual learning [67] and self-supervised decomposition techniques [29]. The introduction of registration modules helps to mitigate misalignment issues [16, 60, 68] in source images. Recent studies have further explored the synergy between fusion and high-level vision tasks. These include leveraging downstream task losses to optimize fusion networks [32, 56], embedding high-level task features [34, 76, 78], and employing initialization techniques [40]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.697, + 0.741, + 0.714 + ], + "angle": 0, + "content": "2.2. Meta-Learning in Vision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Meta-learning develops algorithms to automatically fine-tune hyperparameters for specific tasks, showing its versatility and effectiveness across various domains. MAML [11] and its variations [12, 50, 53] focus on learning efficient initialization parameters to quickly adapt to new tasks using minimal data. Meta-SGD [26] extends MAML by learning optimal update directions and rates, beneficial in few-shot learning scenarios. Other approaches like MW-Net [55] and L2RW [54] emphasize selecting relevant sample weights to tackle noisy data using a compact validation set. Additionally, some studies focus on improving model adaptability through learning loss functions [2, 7, 15]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7458" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.09, + 0.394, + 0.105 + ], + "angle": 0, + "content": "Learning of Loss Generate Module" + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.09, + 0.78, + 0.105 + ], + "angle": 0, + "content": "Learning of Fusion" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.108, + 0.904, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.39, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Figure 1. The TDFusion workflow alternates between training the loss generation module and the fusion module. Training of the loss generation module involves both inner and outer updates, learned through meta-learning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.426, + 0.484, + 0.562 + ], + "angle": 0, + "content": "In image fusion, learnable filters [22] enable the fusion of images at arbitrary resolutions. MetaFusion [78] introduces a mechanism that improves image fusion and object detection by aligning semantics with fusion-specific features. ReFusion [5] guides the learnable fusion loss for various fusion tasks by reconstructing the source images. Metalearning also supports neural architecture search [38, 40] to identify optimal network architectures for image fusion and customizes network initialization for various tasks [40]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.423, + 0.59 + ], + "angle": 0, + "content": "2.3. Comparison with Existing Approaches" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.792 + ], + "angle": 0, + "content": "We propose a novel image fusion method tailored for downstream tasks, leveraging a learnable fusion loss driven by task-specific objectives. Our method employs a meta-learning algorithm that alternates between inner and outer updates, allowing the downstream task loss to guide the optimization of learnable fusion parameters. This results in fused images that minimize the downstream task loss, enhancing their adaptability across various tasks. Unlike previous methods, our approach develops a task-specific fusion loss, shifting focus from traditional factors such as resolution and network structure, and avoiding reliance on predefined fusion loss terms. This renders our fusion framework more flexible and applicable to various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.808, + 0.182, + 0.823 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.201, + 0.848 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our TDFusion framework, as shown in Fig. 1, consists of a fusion network \\(\\mathcal{F}(\\cdot)\\), a downstream task network \\(\\mathcal{T}(\\cdot)\\) and a fusion loss generation module \\(\\mathcal{G}(\\cdot)\\), which produces" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.577 + ], + "angle": 0, + "content": "parameters for a learnable loss function. The parameters of these modules are denoted as \\(\\theta_{\\mathcal{F}}\\), \\(\\theta_{\\mathcal{T}}\\), and \\(\\theta_{\\mathcal{G}}\\), respectively. The one-step updated clones of \\(\\mathcal{F}\\) and \\(\\mathcal{T}\\) are denoted as \\(\\mathcal{F}'\\) and \\(\\mathcal{T}'\\), with parameters \\(\\theta_{\\mathcal{F}'}\\) and \\(\\theta_{\\mathcal{T}'}\\). During the updates, the fusion network and the loss generation module alternate in learning, as depicted by blue and purple in Fig. 1. The update of the loss generation module consists of inner and outer updates, detailed in the following subsections. \\(\\mathcal{L}_f\\) and \\(\\mathcal{L}_t\\) represent the learnable fusion loss and task-specific loss, with their formulations provided in the next section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.661, + 0.603 + ], + "angle": 0, + "content": "3.2. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.775 + ], + "angle": 0, + "content": "The learnable fusion loss \\(\\mathcal{L}_f\\) consists of the intensity term and the gradient term. The intensity term is defined by the output of the loss generation module \\(\\{w_a, w_b\\} = \\mathcal{G}(I_a, I_b)\\), where \\(w_a\\) and \\(w_b\\) control the intensity preference in the fusion loss. The Softmax function in the loss generation module ensures \\(w_a^{ij} + w_b^{ij} = 1\\) for each pixel, thereby selectively retaining the intensity information from the source images. The gradient term emphasizes higher gradient values from the input images [76, 83, 89], aiming to preserve maximal information from the source. The detailed formulation of the learnable fusion loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.786, + 0.907, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = \\mathcal {L} _ {f} ^ {\\text {i n t}} + \\alpha \\mathcal {L} _ {f} ^ {\\text {g r a d}}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.809, + 0.905, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} ^ {i n t} = \\frac {1}{H W} \\sum_ {i j} \\left[ \\sum_ {k \\in \\{a, b \\}} w _ {k} ^ {i j} \\left(I _ {f} ^ {i j} - I _ {k} ^ {i j}\\right) ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.86, + 0.905, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} ^ {\\text {g r a d}} = \\frac {1}{H W} \\sum_ {i j} \\left| \\nabla I _ {f} ^ {i j} - \\max _ {k \\in \\{a, b \\}} \\left(\\nabla I _ {k} ^ {i j}\\right) \\right|, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7459" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.287 + ], + "angle": 0, + "content": "where \\(\\nabla\\) denotes the Sobel operator, commonly employed for gradient extraction in image fusion [32, 57, 83]. The parameter \\(\\alpha\\) serves as a scaling factor, while \\(\\mathcal{L}_f^{int}\\) and \\(\\mathcal{L}_f^{grad}\\) represent the intensity loss and gradient loss, respectively. The weights \\(\\{w_a, w_b\\}\\) control the emphasis of the loss function on the intensity information from each source image. These parameters enable the fusion process to selectively aggregate and incorporate relevant information from the source images. Variations in \\(\\theta_{\\mathcal{G}}\\) lead to different configurations of \\(w_a\\) and \\(w_b\\), thereby influencing the characteristics of the fusion loss. \\(\\theta_{\\mathcal{G}}\\) undergoes updates driven by the high-level task loss during the in-step update process, as detailed in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.351 + ], + "angle": 0, + "content": "The loss function \\(\\mathcal{L}_t\\) depends on the specific task. In this study, we adopt SegFormer [9] and YOLOv8 [19] for the downstream tasks, employing cross-entropy loss and YOLO loss [19] for each task, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.285, + 0.377 + ], + "angle": 0, + "content": "3.3. Dataset Partitioning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.537 + ], + "angle": 0, + "content": "In order to enhance the effectiveness of the loss generation module \\(\\mathcal{G}\\) in guiding the fusion tasks, we create non-overlapping subsets of size \\(M\\) at each training epoch. The meta-training set \\(\\{I_a^{mtr}, I_b^{mtr}\\}\\) and the meta-test set \\(\\{I_a^{mts}, I_b^{mts}\\}\\) are randomly drawn from the fusion training set \\(\\{I_a^{ftr}, I_b^{ftr}\\}\\). These subsets are fed into the model sequentially during the training process of the loss generation module, covering both the inner and outer updates. The entire fusion training set \\(\\{I_a^{ftr}, I_b^{ftr}\\}\\) is fully utilized during the training of the fusion network." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.548, + 0.399, + 0.563 + ], + "angle": 0, + "content": "3.4. Learning of loss generation module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.844 + ], + "angle": 0, + "content": "Fig. 1 illustrates that the loss generation and fusion modules are trained alternately, ensuring the fusion loss is optimized at various stages of training and under different states of the fusion network. The process of training to optimize the fusion loss involves two key steps: the inner update and the outer update. In the inner update, clones of both the fusion network and the task network are generated, with each network undergoing a single training iteration using the fusion loss and task loss, respectively. This process is designed to obtain the state of the network guided by the fusion loss. During the outer update, the task loss of the fusion image produced by the updated clone is calculated. This loss is then backpropagated to the loss generation network. The goal of this step is to direct the fusion network to generate fusion images that result in lower downstream task loss, once guided by the fusion loss. The alternating updates between the inner and outer steps constitute the learning procedure for the loss generation module." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.851, + 0.232, + 0.866 + ], + "angle": 0, + "content": "3.4.1. Inner Update" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "During the inner update phase, the fusion network \\(\\mathcal{F}\\) undergoes a single update guided by the fusion loss, which" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.184 + ], + "angle": 0, + "content": "depends on the current state of network \\(\\mathcal{G}\\). This update primarily aims to compute the intermediate parameters \\(\\theta_{\\mathcal{F}'}\\) and \\(\\theta_{\\mathcal{R}'}\\), which are crucial for updating \\(\\theta_{\\mathcal{G}}\\) in the subsequent phase. The upper section of the purple region in Fig. 1 illustrates this process. During this phase, the images from the meta-training set \\(\\{I_a^{mtr}, I_b^{mtr}\\}\\) are fed into the model:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.188, + 0.907, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathcal {F} ^ {\\prime}} = \\theta_ {\\mathcal {F}} - \\eta_ {\\mathcal {F} ^ {\\prime}} \\frac {\\partial \\mathcal {L} _ {f} \\left(I _ {a} ^ {m t r} , I _ {b} ^ {m t r} , I _ {f} ^ {m t r} ; \\theta_ {\\mathcal {G}}\\right)}{\\partial \\theta_ {\\mathcal {F}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.907, + 0.358 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) undergoes a single gradient descent update. \\(\\theta_{\\mathcal{G}}\\) represent the parameters of the loss generation module \\(\\mathcal{G}\\), which determine the parameters for the learnable fusion loss. And the notation \\(\\eta_{\\mathcal{F}'}\\) refers to the step size. The module \\(\\mathcal{F}'\\) temporarily substitutes \\(\\mathcal{F}\\), adjusting its parameters in one update step. Meanwhile, the parameters of \\(\\mathcal{F}\\), denoted as \\(\\theta_{\\mathcal{F}}\\), remain unchanged. Similarly, \\(\\mathcal{T}'\\) is updated in a single step using the parameters \\(\\theta_{\\mathcal{T}}\\) from the current task network \\(\\mathcal{T}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.362, + 0.907, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathcal {T} ^ {\\prime}} = \\theta_ {\\mathcal {T}} - \\eta_ {\\mathcal {T} ^ {\\prime}} \\frac {\\partial \\mathcal {L} _ {t} \\left(I _ {f} ^ {m t r}\\right)}{\\partial \\theta_ {\\mathcal {T}}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.459 + ], + "angle": 0, + "content": "The parameters \\(\\theta_{\\mathcal{F}'}\\) and \\(\\theta_{\\mathcal{T}'}\\) are both updated during the inner update, which also ensures that the computation graph of \\(\\theta_{\\mathcal{F}'}\\) with respect to \\(\\theta_{\\mathcal{G}}\\) is preserved. This preserved graph is essential for optimizing \\(\\theta_{\\mathcal{G}}\\) during the outer update." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.464, + 0.657, + 0.479 + ], + "angle": 0, + "content": "3.4.2. Outer Update" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.664 + ], + "angle": 0, + "content": "The primary objective of the outer update is to evaluate and refine the fusion guidance capability of \\(\\mathcal{G}\\), specifically by strengthening the influence of the loss function \\(\\mathcal{L}_f\\) in steering the fusion module \\(\\mathcal{F}\\). In the framework diagram, this stage is represented in the lower part of the purple region. The modules \\(\\mathcal{F}'\\) and \\(\\mathcal{T}'\\), derived from the inner update, represent the current guidance capacity of \\(\\mathcal{G}\\). In an ideal scenario, the optimal fusion loss should enhance the performance of the downstream task on the fused image. In this stage, the meta-test set \\(\\{I_a^{mts}, I_b^{mts}\\}\\) is employed. The parameters \\(\\theta_{\\mathcal{G}}\\) are subsequently updated using the task loss \\(\\mathcal{L}_t\\), which is computed by \\(\\mathcal{F}'\\) and \\(\\mathcal{T}'\\):" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.669, + 0.906, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathcal {G}} = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\frac {\\partial \\mathcal {L} _ {t} \\left(I _ {f} ^ {m t s}\\right)}{\\partial \\theta_ {\\mathcal {G}}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.734 + ], + "angle": 0, + "content": "where \\(I_{f}^{mts} = \\mathcal{F}'(I_{a}^{mts}, I_{b}^{mts})\\), and the gradient \\(\\partial \\mathcal{L}_t / \\partial \\theta_G\\) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.739, + 0.907, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L} _ {t}}{\\partial \\theta_ {\\mathcal {G}}} = \\frac {\\partial \\mathcal {L} _ {t}}{\\partial \\theta_ {\\mathcal {F} ^ {\\prime}}} * \\left(- \\eta_ {\\mathcal {F} ^ {\\prime}} \\frac {\\partial^ {2} \\mathcal {L} _ {f} \\left(I _ {a} ^ {m t r} , I _ {b} ^ {m t r} , I _ {f} ^ {m t r} ; \\theta_ {\\mathcal {G}}\\right)}{\\partial \\theta_ {\\mathcal {F}} \\partial \\theta_ {\\mathcal {G}}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Eq. (6) holds because the task loss \\(\\mathcal{L}_t\\) is determined by \\(I_{f}^{mts}\\), which in turn relies on \\(\\theta_{\\mathcal{F}'}\\). The optimization of \\(\\theta_{\\mathcal{G}}\\) via \\(\\mathcal{L}_t\\) is realized by preserving the computational relationship between \\(\\theta_{\\mathcal{F}'}\\) and \\(\\theta_{\\mathcal{G}}\\) throughout the inner update. The updated \\(\\mathcal{G}\\) module gains the ability to generate enhanced fusion loss functions, enabling the fusion module to integrate relevant information from the source images more efficiently into the fused output for downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7460" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.092, + 0.383, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 TDFusion Training Algorithm" + }, + { + "type": "algorithm", + "bbox": [ + 0.095, + 0.108, + 0.487, + 0.405 + ], + "angle": 0, + "content": "Require: Training set \\(\\{I_a^{ftr}, I_b^{ftr}\\}\\) with size \\(N\\). Output: Thoroughly trained \\(\\theta_{\\mathcal{F}}, \\theta_{\\mathcal{T}}, \\theta_{\\mathcal{G}}\\). \n1: Initialize \\(\\theta_{\\mathcal{F}}, \\theta_{\\mathcal{T}}, \\theta_{\\mathcal{G}}\\). \n2: for epoch = 1 to L do \n3: Sample \\(\\{I_a^{mtr}, I_b^{mtr}\\}\\) and \\(\\{I_a^{mts}, I_b^{mts}\\}\\). \n4: for step = 1 to M do \n % Inner update: apply G. \n5: Sample \\((I_a^{mtr}, I_b^{mtr})\\) and get \\((I_f^{mtr}, y^{mtr})\\). \n6: Compute \\(\\theta_{\\mathcal{F}'}\\) and \\(\\theta_{\\mathcal{T}'}\\) by Eq. (4) and Eq. (5). \n% Outer update: optimize G. \n7: Sample \\((I_a^{mts}, I_b^{mts})\\) and get \\((I_f^{mts}, y^{mts})\\). \n8: Update \\(\\theta_{\\mathcal{G}}\\) by Eq. (6). \n9: end for \n10: for step = 1 to N do \n % Fusion update: optimize F and T. \n11: Sample \\((I_a^{ftr}, I_b^{ftr})\\) and get \\((I_f^{ftr}, y^{ftr})\\). \n12: Update \\(\\theta_{\\mathcal{F}}\\) and \\(\\theta_{\\mathcal{T}}\\) by Eq. (8) and Eq. (9). \n13: end for \n14: end for" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.425, + 0.345, + 0.442 + ], + "angle": 0, + "content": "3.5. Learning of Fusion Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.553 + ], + "angle": 0, + "content": "The alternating inner and outer update iterations form a flexible and effective mechanism to refine \\(\\mathcal{G}\\) in response to the evolving state of \\(\\mathcal{F}\\). After refining \\(\\mathcal{G}\\), it is utilized to further improve the training of \\(\\mathcal{F}\\). This stage, denoted in blue in the diagrams, involves processing the fusion training set images \\(\\{I_a^{ftr}, I_b^{ftr}\\}\\). Both \\(\\mathcal{F}\\) and \\(\\mathcal{T}\\) are updated through the application of the fusion loss \\(\\mathcal{L}_f\\) and task loss \\(\\mathcal{L}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.559, + 0.483, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathcal {F}} = \\theta_ {\\mathcal {F}} - \\eta_ {\\mathcal {F}} \\frac {\\partial \\mathcal {L} _ {f} \\left(I _ {a} ^ {f t r} , I _ {b} ^ {f t r} , I _ {f} ^ {f t r} ; \\theta_ {\\mathcal {G}}\\right)}{\\partial \\theta_ {\\mathcal {F}}}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.608, + 0.482, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathcal {T}} = \\theta_ {\\mathcal {T}} - \\eta_ {\\mathcal {T}} \\frac {\\partial \\mathcal {L} _ {t} \\left(I _ {f} ^ {f t r}\\right)}{\\partial \\theta_ {\\mathcal {T}}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.811 + ], + "angle": 0, + "content": "After completing several training sessions on the fusion network, the focus then shifts back to the learning phase of the fusion generation module. The fusion framework evolves through a series of alternating phases, with each phase finetuning the fusion loss based on the fusion network's current state. Such alternating phases ensure that the fusion network consistently applies the most suitable fusion loss during its progression. Ultimately, this results in the development of a highly efficient fusion network, optimized for peak performance. The complete training procedure is detailed in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.296, + 0.833 + ], + "angle": 0, + "content": "3.6. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "TDFusion is composed of three modules: the fusion network, the downstream task network, and the loss generation network. The fusion network shares the same architecture as [5], a lightweight model built upon the Restormer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.304 + ], + "angle": 0, + "content": "Block (RTB) [74]. An adaptive fusion module is incorporated into this network to facilitate feature integration. Fig. 1 illustrates the structure of the loss generation module. It also employs the Restormer Block (RTB) [74] as its primary component, receiving inputs \\(\\{I_a, I_b\\}\\). After applying Softmax(\\(\\cdot\\)), the final output guarantees \\(w_a^{ij} + w_b^{ij} = 1\\) for each pixel. This design ensures that the fused image meets the similarity constraints and removes the reliance on initialization within the loss generation module. The architecture of the downstream task network \\(\\mathcal{T}(\\cdot)\\) depends on the specific task. For learning of loss generation module, we chose the most lightweight models of SegFormer [9] and YOLOv8 [19] for semantic segmentation and object detection, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.312, + 0.708, + 0.327 + ], + "angle": 0, + "content": "3.7. Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.395 + ], + "angle": 0, + "content": "To better understand the weighting mechanism of the loss generation module \\(\\mathcal{G}\\), we investigate the optimization procedure of \\(\\mathcal{G}\\), which generates the weights \\(\\{w_{a}, w_{b}\\}\\), denoted as \\(\\theta_{\\mathcal{G}}\\). For clarity, we rewrite Eq. (2) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.402, + 0.905, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {f} ^ {i n t} = \\left[ w _ {a} \\odot \\left(I _ {a} - I _ {f}\\right) \\odot \\left(I _ {a} - I _ {f}\\right) \\right. \\\\ \\left. + w _ {b} \\odot \\left(I _ {b} - I _ {f}\\right) \\odot \\left(I _ {b} - I _ {f}\\right) \\right] \\times \\frac {1}{H W} \\\\ = \\left[ \\mathcal {G} \\left(I _ {a}, I _ {b}; \\theta_ {\\mathcal {G}}\\right) \\odot \\left(I _ {a} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\right. \\tag {10} \\\\ \\odot \\left(I _ {a} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\\\ + \\left(1 - \\mathcal {G} \\left(I _ {a}, I _ {b}; \\theta_ {\\mathcal {G}}\\right)\\right) \\odot \\left(I _ {b} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\\\ \\left. \\odot \\left(I _ {b} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\right] \\times \\frac {1}{H W}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.907, + 0.583 + ], + "angle": 0, + "content": "Here, \\(w_{a}, w_{b} \\in \\mathbb{R}^{H \\times W}\\), \\(I_{a}, I_{b} \\in \\mathbb{R}^{H \\times W}\\), and \\(\\odot\\) denotes the element-wise multiplication operation. Let \\(\\Omega'\\) be the set \\(\\{\\theta_{\\mathcal{F}'}', \\theta_{\\mathcal{T}''}\\}\\), this leads to the following expression:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.59, + 0.905, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta_ {\\mathcal {G}} = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\frac {\\partial \\mathcal {L} _ {t} ^ {m t s} (\\Omega^ {'} (\\theta_ {\\mathcal {G}}))}{\\partial \\theta_ {\\mathcal {G}}} \\\\ = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\frac {\\partial \\mathcal {L} _ {t} ^ {m t s} (\\Omega^ {'} (\\theta_ {\\mathcal {G}}))}{\\partial \\Omega^ {'}} \\frac {\\partial \\Omega^ {'} (\\theta_ {\\mathcal {G}})}{\\partial \\theta_ {\\mathcal {G}}} \\\\ = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\eta_ {\\mathcal {F} ^ {\\prime}} \\underbrace {\\frac {\\partial \\mathcal {L} _ {t} ^ {m t s} \\left(\\Omega^ {'} \\left(\\theta_ {\\mathcal {G}}\\right)\\right)}{\\partial \\Omega^ {\\prime}}} _ {\\text {(a)}} \\times \\frac {\\partial \\mathcal {G} \\left(I _ {a} , I _ {b} ; \\theta_ {\\mathcal {G}}\\right)}{\\theta_ {\\mathcal {G}}} \\\\ \\times \\underbrace {\\left[ \\left(I _ {a} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) \\odot \\left(I _ {a} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) \\right.} _ {(b)} \\tag {11} \\\\ \\underbrace {- \\left(I _ {b} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) \\odot \\left(I _ {b} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) ]} _ {(b)} \\\\ = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\eta_ {\\mathcal {F} ^ {\\prime}} \\mathbf {G} \\times \\frac {\\partial \\mathcal {G} \\left(I _ {a} , I _ {b} ; \\theta_ {\\mathcal {G}}\\right)}{\\theta_ {\\mathcal {G}}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{G}\\) denotes the inner product between two gradients: (a) the first one is derived from the task loss using a meta-testing set, and (b) the second is calculated from the fusion loss based on a meta-training set. Consequently, the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7461" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.907, + 0.446 + ], + "angle": 0, + "content": "Figure 2. Visual comparison of fusion results. The cases are \"01258N\" in MSRS dataset, \"00122\" in FMB dataset, \"00449\" in M3FD dataset and \"200304\" in LLVIP dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.501 + ], + "angle": 0, + "content": "optimization of the module \\(\\theta_{\\mathcal{G}}\\) is driven by the task loss, with the objective of preserving task-specific information throughout the fusion process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.517, + 0.216, + 0.535 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.544, + 0.173, + 0.56 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Experimental Setup. In our experiments, the epoch number \\( L \\) and the training iterations of the loss generation module \\( M \\) are set to 50 and 200. The learning iterations for the fusion network \\( N \\) depend on the size of the dataset. We use the Adam optimizer with a learning rate of 1e-4, a batch size of 2, and a hyperparameter \\( \\alpha \\) set to 1. All experiments are conducted on a PC with a single NVIDIA RTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Evaluation Metrics and Comparison Methods. The advanced fusion methods compared in our study include TarDAL [32], SegMIF [34], MURF [69], EMMA [86], DCINN [61], MRFS [76], and TIMFusion [40]. The fusion performance is evaluated using metrics including entropy (EN), spatial frequency (SF), sum of correlation differences (SCD), visual information fidelity (VIF), \\( Q^{AB / F} \\) and structural similarity index metric (SSIM)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Dataset Split. We use four datasets annotated for downstream tasks, including MSRS [58], FMB [34], M3FD [32], and LLVIP [17]. MSRS contains 1083/361 image pairs for training/test, and FMB contains 1220/280 pairs for training/test. We follow the splits of the original papers for both datasets. M3FD dataset consists of 4200 pairs of images for detection, with 300 pairs designated for fusion evalu-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.591 + ], + "angle": 0, + "content": "ation. The 4200 detection images are split into 3150 for training and 1050 for testing, ensuring that the 300 pairs for fusion evaluation are included within the detection test set. These 300 fusion images are then employed to assess the fusion performance. The original LLVIP dataset contains 12025/3463 image pairs as training/test set. Due to its large size, we select every 10th image to form our training and test sets, resulting in 1203/347 image pairs for training and testing. Our splits for M3FD and LLVIP will be available." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.706, + 0.621 + ], + "angle": 0, + "content": "4.2. Fusion Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fig. 2 presents a visual comparison of different methods. The fused images generated by TDFusion excel in preserving discriminative details, achieving balanced brightness, and maintaining clear object contours. It effectively preserves the target features from the infrared images and the background details from the visible images, resulting in fused images that are more natural and exhibit greater clarity across different environments. These results highlight the advantages of TDFusion in detail preservation and visual performance. More results are available in the supplementary material. Tab. 1 presents the quantitative comparison of fusion over four datasets. TDFusion outperforms other methods across most metrics. This suggests that TDFusion not only enhances image details but also provides consistent fusion results across various scenarios. Compared to other methods, TDFusion shows superior adaptability and robustness, particularly in handling diverse image characteristics and challenging fusion tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7462" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.905, + 0.107 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of Infrared-visible image fusion. The red and blue markers represent the best and second-best values." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.111, + 0.907, + 0.374 + ], + "angle": 0, + "content": "
Infrared-visible Image Fusion on MSRS [58] DatasetInfrared-visible Image Fusion on FMB [34] Dataset
EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑
TarDAL [32]5.285.980.710.210.180.47TarDAL [32]6.636.941.030.280.290.74
SegMIF [34]5.9511.101.570.440.630.55SegMIF [34]6.8313.691.720.390.650.60
MURF [69]5.0410.491.020.220.370.60MURF [69]6.3713.881.340.220.370.68
EMMA [86]6.7311.561.620.490.640.70EMMA [86]6.7715.001.500.420.650.72
DCINN [61]6.0010.511.490.410.570.52DCINN [61]6.4711.471.390.380.590.74
MRFS [76]7.008.861.420.370.490.55MRFS [76]6.7812.421.240.380.620.73
TIMFusion [40]6.279.671.340.320.480.68TIMFusion [40]6.5112.231.240.350.590.73
TDFusion (Ours)6.7411.301.860.500.670.70TDFusion (Ours)6.8614.161.760.430.680.75
Infrared-visible Image Fusion on M3FD [32] DatasetInfrared-visible Image Fusion on LLVIP [17] Dataset
EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑
TarDAL [32]6.877.631.290.270.300.71TarDAL [32]6.327.421.040.270.220.58
SegMIF [34]6.8514.141.720.370.600.59SegMIF [34]6.6815.461.380.400.660.57
MURF [69]6.5012.551.460.210.320.64MURF [69]6.1315.080.960.210.310.57
EMMA [86]6.9215.231.490.380.590.69EMMA [86]7.3515.371.570.410.640.66
DCINN [61]6.5911.211.460.340.510.72DCINN [61]6.9813.341.430.380.520.64
MRFS [76]6.9412.071.260.340.550.70MRFS [76]6.8311.041.230.310.420.64
TIMFusion [40]6.7512.311.370.350.530.70TIMFusion [40]6.5813.521.140.330.460.64
TDFusion (Ours)6.9914.491.830.410.650.72TDFusion (Ours)7.3616.381.750.460.700.67
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.388, + 0.905, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.542, + 0.88, + 0.555 + ], + "angle": 0, + "content": "Figure 3. Visual comparison for Semantic Segmentation. The cases are \"00726N\" in MSRS dataset and \"01438\" in FMB dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.327, + 0.58 + ], + "angle": 0, + "content": "4.3. Downstream Applications" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.592, + 0.483, + 0.834 + ], + "angle": 0, + "content": "This section validates the adaptability of fusion methods to downstream tasks. For a fair comparison, we adopt SegFormer [9] and YOLOv8 [19] as backbones and retrain the task networks for each fusion method over 300 epochs to evaluate their adaptability to semantic segmentation and object detection. Fig. 3 and Fig. 4 present the visual comparisons of semantic segmentation and object detection, respectively. TDFusion outperforms in image detail retention, edge clarity, and object recognition, effectively identifying and segmenting objects. For semantic segmentation, the generated maps clearly distinguish different class regions, closely matching the ground truth. In object detection, the fused images exhibit more precise boundary localization for salient objects. This indicates that TDFusion maintains a better balance between fine details and overall context. More results can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Tab. 2 shows the performance comparison across different methods in semantic segmentation and object detection. TDFusion outperforms other methods on most metrics, particularly in mIoU and mAP. This demonstrates that TDFu" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.562, + 0.905, + 0.592 + ], + "angle": 0, + "content": "Table 2. Performance comparison of downstream applications. The red and blue markers represent the best and second-best." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.593, + 0.907, + 0.746 + ], + "angle": 0, + "content": "
MethodsSemantic SegmentationObject Detection
MSRSFMBM3FDLLVIP
mAccmIoUmAccmIoUmAP50mAP75AP50AP75
Infrared83.2369.4958.8551.9879.1253.0596.0372.07
Visible83.4473.7665.1257.9682.2154.8291.7848.66
TarDAL81.9371.3562.8655.3383.1656.3993.7962.71
SegMIF85.7374.2565.9758.4183.6158.2393.9566.45
MURF85.0374.0864.1056.9680.5854.2294.2468.04
EMMA85.9974.4862.4556.2883.7156.9194.0066.21
DCINN84.1174.3561.0954.8182.6957.3794.9268.34
MRFS84.7674.5061.9355.7183.2857.7493.0367.21
TIMFusion83.6773.5863.7057.2483.2256.0893.7661.33
TDFusion86.0475.0967.1760.5086.2759.7195.0069.18
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.822 + ], + "angle": 0, + "content": "sion effectively enhances the quality of fused images. It also improves the downstream task performance, especially in terms of accuracy and robustness in complex scenarios. Class-wise results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.766, + 0.849 + ], + "angle": 0, + "content": "4.4. Task-driven Learnable Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our framework incorporates a learnable fusion loss that models the preferences of downstream tasks for information from source images. The models trained on FMB dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7463" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.245, + 0.862, + 0.258 + ], + "angle": 0, + "content": "Figure 4. Visual comparison for Object Detection. The cases are \"02236\" in M3FD dataset and \"210145\" in LLVIP dataset." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.264, + 0.287, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.264, + 0.482, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.421, + 0.472, + 0.434 + ], + "angle": 0, + "content": "Figure 5. Visualisation of learnable loss for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.747 + ], + "angle": 0, + "content": "and LLVIP dataset labeled as \\( SS \\) and \\( OD \\), are evaluated on MSRS dataset to simulate performance in unknown scenes, as shown in Fig. 5. The results demonstrate that the fusion model adaptively selects information from infrared and visible images to satisfy task requirements. In semantic segmentation, the model combines scene structure and texture, prioritizing boundaries. This improves segmentation under varying lighting conditions. Fusion weights \\( \\{w_{a}^{SS}, w_{b}^{SS}\\} \\) indicate a preference for visible details and infrared advantages in low-light conditions. In object detection, the model focuses on edge and contrast information, especially for instances like pedestrians and vehicles. Higher fusion weights are assigned to bright regions in infrared images, enhancing target detection in low-light conditions. Fusion weights \\( \\{w_{a}^{OD}, w_{b}^{OD}\\} \\) reflect this preference. Comparison of fusion losses across tasks reveals distinct differences, especially in highlighted regions, confirming that the model adapts to task-specific requirements by selecting the most relevant information from multimodal images. More results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.254, + 0.772 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To thoroughly evaluate the performance of our proposed algorithm, we conduct a series of ablation experiments on FMB dataset, and the detailed results are shown in Tab. 3. In Exp. I, we exclude the learnable fusion loss by fixing \\( w_{a} \\) and \\( w_{b} \\) to \\( 1/2 \\). In Exp. II, we omit the gradient loss from the loss function. In Exp. III, we also allow the fusion module parameters to be jointly optimized by both the task loss \\( \\mathcal{L}_{t} \\) and fusion loss \\( \\mathcal{L}_{f} \\). In Exp. IV, we exclude the fusion" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.265, + 0.904, + 0.28 + ], + "angle": 0, + "content": "Table 3. Ablation experiment of fusion. The red denotes the best." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.283, + 0.907, + 0.389 + ], + "angle": 0, + "content": "
Ablation Studies of fusion on FMB Dataset
ConfigurationsENSFSCDVIF\\(Q^{AB/F}\\)SSIM
Ifix \\(w_{a}\\) and \\(w_{b}\\) as 1/26.6013.731.580.390.600.72
IIw/o \\(\\mathcal{L}_{f}^{grad}\\)6.7711.651.630.370.640.73
III\\(\\theta_{\\mathcal{F}}\\) influenced by \\(\\mathcal{L}_{t}\\)6.8013.851.700.410.660.73
IVw/o Fusion learning6.8214.071.720.410.670.72
V\\(I_{f}=w_{a}*I_{a}+w_{b}*I_{b}\\)6.7511.491.650.380.620.73
Ours6.8614.161.760.430.680.75
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.519 + ], + "angle": 0, + "content": "module's dedicated learning phase and update it during the outer update of the loss module, which tests the impact of the fusion module's training schedule. In Exp. V, we replace our fusion method with a discriminative approach. The performance decline observed across different configurations confirms the rationality and effectiveness of our proposed method. Visualization and more analysis are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.634, + 0.549 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.907, + 0.817 + ], + "angle": 0, + "content": "To overcome the limitations of predefined fusion losses, which often fail to effectively guide the fusion process for downstream tasks, we propose a meta-learning-based framework for task-guided fusion. This framework includes a loss generation module that outputs the parameters of the learnable fusion loss. The module is updated using a meta-learning approach, which alternates between inner and outer loop steps to enhance its ability to guide the fusion network. Under varying fusion conditions, this module generates the optimal fusion loss for the downstream task. This enables the fusion network to produce fused images that minimize the task-specific loss. The theoretical analysis explains how the downstream task loss guides the fusion loss in our framework. Experiments using four publicly available fusion datasets and downstream tasks including semantic segmentation and object detection, demonstrate the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.673, + 0.847 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.9 + ], + "angle": 0, + "content": "This work has been supported by the National Natural Science Foundation of China under Grant 12201497 and 12371512." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7464" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Hessah Albanwan, Rongjun Qin, and Yang Tang. Image fusion in remote sensing: An overview and meta-analysis. Photogrammetric Engineering & Remote Sensing, 90(12): 755-775, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.225 + ], + "angle": 0, + "content": "[2] Antreas Antoniou and Amos J. Storkey. Learning to learn by self-critique. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), pages 9936-9946, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[3] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yukun Cui, Chunxia Zhang, Zhenbo Guo, and Yongjun Wang. Simultaneous automatic picking and manual picking refinement for first-break. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Baisong Jiang, Lilun Deng, Yukun Cui, Shuang Xu, and Chunxia Zhang. Deep unfolding multi-modal image fusion network via attribution analysis. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[5] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yichen Wu, Lilun Deng, Yukun Cui, Baisong Jiang, and Shuang Xu. Refusion: Learning image fusion from reconstruction with learnable loss via meta-learning. International Journal of Computer Vision, pages 1-21, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Haowen Bai, Jiangshe Zhang, Zixiang Zhao, Lilun Deng, Yukun Cui, and Shuang Xu. Retinex-mef: Retinex-based glare effects aware unsupervised multi-exposure image fusion. arXiv preprint arXiv:2503.07235, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 9445-9454. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[8] Yanpeng Cao, Dayan Guan, Weilin Huang, Jiangxin Yang, Yanlong Cao, and Yu Qiao. Pedestrian detection with unsupervised multispectral feature learning using deep neural networks. Information Fusion, 46:206-217, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[9] Bo Cheng, Xiang Li, Yujie Wei, Cheng Huang, Xiaoyong Zhang, Yandong Jiang, Tianyu Zhang, Na Xu, Shuai Yu, Xinxin Zhan, et al. Segformer: Simple and efficient design for semantic segmentation with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12022-12032, 2021. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[10] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3333-3348, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the International conference on machine learning (ICML), pages 1126–1135, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[12] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In Proceedings of the International conference on machine learning (ICML), pages 1920-1930, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Qishen Ha, Kohei Watanabe, Takumi Karasawa, Yoshitaka Ushiku, and Tatsuya Harada. Mfnet: Towards real-time se" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "mantic segmentation for autonomous vehicles with multispectral scenes. In IROS, pages 5108-5115. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.193 + ], + "angle": 0, + "content": "[14] Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Task-driven super resolution: Object detection in low-resolution images. In Neural Information Processing: 28th International Conference (ICONIP), pages 387-395. Springer, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[15] Rein Houthooft, Yuhua Chen, Phillip Isola, Bradley C. Stadie, Filip Wolski, Jonathan Ho, and Pieter Abbeel. Evolved policy gradients. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), pages 5405-5414, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.267, + 0.905, + 0.337 + ], + "angle": 0, + "content": "[16] Zhanbo Huang, Jinyuan Liu, Xin Fan, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Reconet: Recurrent correction network for fast and efficient multi-modality image fusion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 539-555. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.395 + ], + "angle": 0, + "content": "[17] Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible-infrared paired dataset for low-light vision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3496-3504, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.907, + 0.466 + ], + "angle": 0, + "content": "[18] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 4369-4376, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.904, + 0.497 + ], + "angle": 0, + "content": "[19] Glenn Jocher. Ultralytics YOLOv8. https://github.com/ultralytics, 2023.4,5,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.5, + 0.905, + 0.569 + ], + "angle": 0, + "content": "[20] Chenglong Li, Chengli Zhu, Yan Huang, Jin Tang, and Liang Wang. Cross-modal ranking with soft consistency and noisy labels for robust RGB-T tracking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 831-847. Springer, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[21] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[22] Huafeng Li, Yueliang Cen, Yu Liu, Xun Chen, and Zhengtao Yu. Different input resolutions and arbitrary output resolution: A meta learning-based deep framework for infrared and visible image fusion. IEEE Transactions on Image Processing, 30:4070-4083, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[23] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[24] Siyuan Li, Iago Bruno Araujo, Wenqi Ren, Zhangyang Wang, Eric K. Tokuda, Roberto Hirata Junior, Roberto Marcondes Cesar Junior, Jiawan Zhang, Xiaojie Guo, and Xiaochun Cao. Single image deraining: A comprehensive benchmark analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3838-3847. Computer Vision Foundation / IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Xiaoling Li, Yanfeng Li, Houjin Chen, Yahui Peng, and Pan Pan. Ccafusion: cross-modal coordinate attention network for infrared and visible image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7465" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[26] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[27] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[28] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[29] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 719-735. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[30] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[31] Jinyuan Liu, Xin Fan, Ji Jiang, Risheng Liu, and Zhongxuan Luo. Learning a deep multi-scale feature ensemble and an edge-attention guidance for image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 32(1): 105–119, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[32] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5792-5801, 2022. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[33] Jinyuan Liu, Jingjie Shang, Risheng Liu, and Xin Fan. Attention-guided global-local adversarial learning for detail-preserving multi-exposure image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5026-5040, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[34] Jinyuan Liu, Zhu Liu, Guanyao Wu, Long Ma, Risheng Liu, Wei Zhong, Zhongxuan Luo, and Xin Fan. Multi-interactive feature learning and a full-time multi-modality benchmark for image fusion and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8115–8124, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[35] Jinyuan Liu, Guanyao Wu, Junsheng Luan, Zhiying Jiang, Risheng Liu, and Xin Fan. Holoco: Holistic and local contrastive learning network for multi-exposure image fusion. Information Fusion, 95:237-249, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[36] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, 132(5):1748-1775, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[37] Jinyuan Liu, Guanyao Wu, Zhu Liu, Di Wang, Zhiying Jiang, Long Ma, Wei Zhong, and Xin Fan. Infrared and visible image fusion: From data compatibility to task adaption. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[38] Risheng Liu, Zhu Liu, Jinyuan Liu, and Xin Fan. Searching a hierarchically aggregated fusion architecture for fast multi-modality image fusion. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 1600-1608. ACM, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[39] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5953-5969, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[40] Risheng Liu, Zhu Liu, Jinyuan Liu, Xin Fan, and Zhongxuan Luo. A task-guided, implicitly-searched and metainitialized deep model for image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[41] ShuMin Liu, Jiajia Chen, and Susanto Rahardja. A new multi-focus image fusion algorithm and its efficient implementation. IEEE Transactions on Circuits and Systems for Video Technology, 30(5):1374-1384, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[42] Zhu Liu, Jinyuan Liu, Guanyao Wu, Long Ma, Xin Fan, and Risheng Liu. Bi-level dynamic learning for jointly multimodality image fusion and beyond. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1240-1248, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[43] Zhu Liu, Jinyuan Liu, Benzhuang Zhang, Long Ma, Xin Fan, and Risheng Liu. Paif: Perception-aware infrared-visible image fusion for attack-tolerant semantic segmentation. In Proceedings of the 31st ACM international conference on multimedia, pages 3706-3714, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[44] Zhu Liu, Jinyuan Liu, Guanyao Wu, Zihang Chen, Xin Fan, and Risheng Liu. Searching a compact architecture for robust multi-exposure image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6224-6237, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[45] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information Fusion, 45:153-178, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[46] Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusion: A generative adversarial network for infrared and visible image fusion. Information Fusion, 48:11-26, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[47] Jiayi Ma, Pengwei Liang, Wei Yu, Chen Chen, Xiaojie Guo, Jia Wu, and Junjun Jiang. Infrared and visible image fusion via detail preserving adversarial learning. Information Fusion, 54:85-98, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[48] Jiayi Ma, Han Xu, Junjun Jiang, Xiaoguang Mei, and Xiao-Ping (Steven) Zhang. Ddcgan: A dual-discriminator conditional generative adversarial network for multi-resolution image fusion. IEEE Transactions on Image Processing, 29: 4980-4995, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[49] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5637-5646, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7466" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[51] Seonghyun Park, An Gia Vien, and Chul Lee. Cross-modal transformers for infrared and visible image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 34(2):770-785, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[52] Yanting Pei, Yaping Huang, Qi Zou, Yuhang Lu, and Song Wang. Does haze removal help cnn-based image classification? In Proceedings of the European conference on computer vision (ECCV), pages 682-697, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[53] Xinran Qin, Yuhui Quan, Tongyao Pang, and Hui Ji. Ground-truth free meta-learning for deep compressive sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9947-9956, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[54] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In Proceedings of the International conference on machine learning (ICML), pages 4334-4343, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.407 + ], + "angle": 0, + "content": "[55] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[56] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[57] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[58] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Infromation Fusion, 83-84:79-92, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[59] Wei Tang, Fazhi He, Yu Liu, Yansong Duan, and Tongzhen Si. Datfuse: Infrared and visible image fusion via dual attention transformer. IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159-3172, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[60] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 3508-3515. ijcai.org, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.485, + 0.77 + ], + "angle": 0, + "content": "[61] Wu Wang, Liang-Jian Deng, Ran Ran, and Gemine Vivone. A general paradigm with detail-preserving conditional invertible network for image fusion. International Journal of Computer Vision, 132(4):1029-1054, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[62] Zeyu Xiao and Xinchao Wang. Event-based video superresolution via state space models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[63] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 2409-2418, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[64] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 217-235. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[65] Han Xu, Jiayi Ma, Zhuliang Le, Junjun Jiang, and Xiaojie Guo. Fusiondn: A unified densely connected network for image fusion. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 12484-12491, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[66] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[67] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.907, + 0.39 + ], + "angle": 0, + "content": "[68] Han Xu, Jiayi Ma, Jiteng Yuan, Zhuliang Le, and Wei Liu. Rfnet: Unsupervised network for mutually reinforcing multimodal image registration and fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19647-19656, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[69] Han Xu, Jiteng Yuan, and Jiayi Ma. Murf: Mutually reinforcing multi-modal image registration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[70] Shuang Xu, Ouafa Amira, Junmin Liu, Chun-Xia Zhang, Jiangshe Zhang, and Guanghai Li. HAM-MFN: hyperspectral and multispectral image multiscale fusion network with RAP loss. IEEE Transactions on Geoscience and Remote Sensing, 58(7):4618-4628, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[71] Shuang Xu, Lizhen Ji, Zhe Wang, Pengfei Li, Kai Sun, Chunxia Zhang, and Jiangshe Zhang. Towards reducing severe defocus spread effects for multi-focus image fusion via an optimization based strategy. IEEE Transactions Computational Imaging, 6:1561-1570, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[72] Yong Yang, Jiaxiang Liu, Shuying Huang, Weiguo Wan, Wenying Wen, and Juwei Guan. Infrared and visible image fusion via texture conditional generative adversarial network. IEEE Transactions on Circuits and Systems for Video Technology, 31(12):4771-4783, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[73] Xunpeng Yi, Linfeng Tang, Hao Zhang, Han Xu, and Jiayi Ma. Diff-if: Multi-modality image fusion via diffusion model with fusion knowledge prior. Information Fusion, 110:102450, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[74] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728-5739, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[75] Hao Zhang and Jiayi Ma. Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion. International Journal of Computer Vision, 129(10):2761-2785, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[76] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Com" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7467" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "puter Vision and Pattern Recognition (CVPR), pages 26974-26983, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[77] Yu Zhang, Yu Liu, Peng Sun, Han Yan, Xiaolin Zhao, and Li Zhang. IFCNN: A general image fusion framework based on convolutional neural network. Information Fusion, 54: 99-118, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[78] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13955-13965, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.483, + 0.306 + ], + "angle": 0, + "content": "[79] Yangyang Zhao, Qingchun Zheng, Peihao Zhu, Xu Zhang, and Wenpeng Ma. Tufusion: A transformer-based universal fusion algorithm for multimodal images. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.308, + 0.483, + 0.377 + ], + "angle": 0, + "content": "[80] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Jiangshe Zhang, and Pengfei Li. Didfuse: Deep image decomposition for infrared and visible image fusion. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 970-976, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[81] Zixiang Zhao, Shuang Xu, Jiangshe Zhang, Chengyang Liang, Chunxia Zhang, and Junmin Liu. Efficient and model-based infrared and visible image fusion via algorithm unrolling. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1186-1196, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.45, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[82] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5906–5916, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.484, + 0.618 + ], + "angle": 0, + "content": "[83] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5906–5916, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[84] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. DDFM: denoising diffusion model for multi-modality image fusion. CoRR, abs/2303.06840, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[85] Zixiang Zhao, Jiang-She Zhang, Haowen Bai, Yicheng Wang, Yukun Cui, Lilun Deng, Kai Sun, Chunxia Zhang, Junmin Liu, and Shuang Xu. Deep convolutional sparse coding networks for interpretable image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2369-2377, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[86] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25912-25921, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[87] Huabing Zhou, Wei Wu, Yanduo Zhang, Jiayi Ma, and Haibin Ling. Semantic-supervised infrared and visible image" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "fusion via a dual-discriminator generative adversarial network. IEEE Transactions on Multimedia, 25:635-648, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[88] Man Zhou, Naishan Zheng, Xuanhua He, Danfeng Hong, and Jocelyn Chanussot. Probing synergistic high-order interaction for multi-modal image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[89] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7468" + } + ] +] \ No newline at end of file diff --git a/2025/Task-driven Image Fusion with Learnable Fusion Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_origin.pdf b/2025/Task-driven Image Fusion with Learnable Fusion Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..21d4ca6ae9116362c6e5d4ccc3e8683d6df48e55 --- /dev/null +++ b/2025/Task-driven Image Fusion with Learnable Fusion Loss/37d9af9d-2102-464a-b77a-9ecf64006f8e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a8ec7287323fb2bd4157845c68d070b9076961d2c8200168d8341f68830ccfc +size 9092267 diff --git a/2025/Task-driven Image Fusion with Learnable Fusion Loss/full.md b/2025/Task-driven Image Fusion with Learnable Fusion Loss/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a99437cd60acb1d04df755d850e01e0d57c093a3 --- /dev/null +++ b/2025/Task-driven Image Fusion with Learnable Fusion Loss/full.md @@ -0,0 +1,352 @@ +# Task-driven Image Fusion with Learnable Fusion Loss + +Haowen Bai $^{1}$ Jiangshe Zhang $^{1*}$ Zixiang Zhao $^{2*}$ Yichen Wu $^{3}$ Lilun Deng $^{1}$ Yukun Cui $^{1}$ Tao Feng $^{4}$ Shuang Xu $^{5}$ + +$^{1}$ Xi'an Jiaotong University $^{2}$ ETH Zürich $^{3}$ City University of Hong Kong $^{4}$ Tsinghua University $^{5}$ Northwestern Polytechnical University + +hwbai@stu.xjtu.edu.cn + +# Abstract + +Multi-modal image fusion aggregates information from multiple sensor sources, achieving superior visual quality and perceptual features compared to single-source images, often improving downstream tasks. However, current fusion methods for downstream tasks still use predefined fusion objectives that potentially mismatch the downstream tasks, limiting adaptive guidance and reducing model flexibility. To address this, we propose Task-driven Image Fusion (TD-Fusion), a fusion framework incorporating a learnable fusion loss guided by task loss. Specifically, our fusion loss includes learnable parameters modeled by a neural network called the loss generation module. This module is supervised by the downstream task loss in a meta-learning manner. The learning objective is to minimize the task loss of fused images after optimizing the fusion module with the fusion loss. Iterative updates between the fusion module and the loss module ensure that the fusion network evolves toward minimizing task loss, guiding the fusion process toward the task objectives. TDFusion's training relies entirely on the downstream task loss, making it adaptable to any specific task. It can be applied to any architecture of fusion and task networks. Experiments demonstrate TDFusion's performance through fusion experiments conducted on four different datasets, in addition to evaluations on semantic segmentation and object detection tasks. The code is available at https://github.com/HaowenBai/TDFusion. + +# 1. Introduction + +Multi-modal image fusion [25, 31, 51, 59, 72, 79, 82] combines information from multiple sensors to produce a more holistic and detailed representation. Infrared images capture thermal radiation regardless of lighting conditions, while visible images provide richer texture details. Fused im + +ages enhance downstream tasks through improved information density and robustness [18, 41, 45, 62, 63, 70, 71], outperforming single-modal inputs in semantic segmentation [13, 28, 34, 56], object detection [8, 32], and other related applications [1, 3, 20, 27]. Conventional methods typically treat fusion as image restoration using unsupervised loss [76, 83, 87-89] or perceptual loss [21, 23, 77]. These approaches prioritize visual-level fusion through predefined aggregation objectives, often neglecting semantic feature extraction. This limitation hinders scene interpretation and task performance [14, 24, 52, 64]. Recent advances explore the mutual enhancement between fusion and downstream tasks [37]. By cascading the fusion network with downstream task network [39, 49], the task loss constrains the fusion learning, ensuring the fused images meet the task requirements [32, 56]. Alternatively, some methods incorporate high-level visual task features [34, 76, 78] or focus on learning optimal initializations [40] to enhance fusion. + +While integrating downstream tasks, existing frameworks still rely on predefined fusion loss terms lacking dynamic adaptation. The impact of downstream tasks remains limited due to specific combinations. Manually defined losses preserve predefined guidance, frequently overlooking task-specific requirements. This guidance imposes manually designed prior constraints on the fusion process, limiting the dynamic and adaptive influence of downstream tasks on specific image pairs. These approaches, whether incorporating task features [34, 78] or employing task losses [32, 56], still face the limitations of fixed fusion loss terms. Task-specialized networks [76] create fusion-task dependencies, restricting flexibility and limiting their applicability to various high-level vision tasks. We address these limitations through a task-driven framework with learnable loss. The fusion loss contains learnable parameters, generated by a loss generation module, and is designed to retain the intensity information of the source images for specific downstream tasks. The purpose of updating the fusion loss is to guide the fusion network in generating fused images that minimize downstream task loss, + +thereby enhancing adaptability. Moreover, the fusion loss update relies on the downstream task loss, making it independent of any specific task or network architecture. + +The loss generation module produces fusion losses for subsequent fusion module updates. This complexity presents challenges to standard end-to-end training. Fortunately, meta-learning techniques, which are strategies for learning how to learn, can effectively achieve the learning objectives of the loss generation module. This involves minimizing task loss for fused images through an optimized fusion loss. Meta-learning tackles common deep-learning challenges like limited data, high computational costs, and the need for better generalization. Core optimization areas include parameter tuning [11], optimization strategies [26], and network architectures search [30]. In this paper, we draw inspiration from the Model-Agnostic Meta-Learning (MAML) approach [11] to train our loss generation module. Specifically, training the loss generation module involves two stages: inner updates and outer updates. During inner updates, the output of the loss generation module updates a surrogate fusion module without altering the original parameters. During outer updates, the fused image from the surrogate module is fed into the task network. The resulting task loss then updates the loss generation module through backpropagation. This alternating training ensures that the loss generation module consistently produces fusion losses that minimize the downstream task loss of the fused images. + +This paper introduces TDFusion, a task-oriented fusion framework driven by downstream tasks. It consists of a fusion module, a task module, and a loss generation module that learns to optimize the fusion loss. The fusion loss incorporates intensity preferences from source images and gradient preservation, guided by downstream task loss to refine intensity preferences. This model follows the general form of fusion loss used in advanced methods [76, 83, 88, 89], ensuring adaptability to various tasks. The updates of the loss generation module are performed using a meta-learning approach, optimizing the loss generation module parameters based on task loss from the fused images after each update of the fusion module. This process ensures that the loss function guides the fusion module continuously, optimizing feature aggregation and minimizing downstream task losses. The loss generation module dynamically adjusts through alternating updates with the fusion and task modules, generating optimal fusion losses at each model state. Our contributions can be summarized as follows: + +- We propose TDFusion, a meta-learning-based fusion framework that leverages the loss functions of downstream tasks for training. This method promotes task-driven fusion and alleviates challenges caused by the absence of ground truth. Moreover, this framework is agnostic of specific downstream tasks or network architectures, which enhances its adaptability and flexibility. + +- Our framework includes a dynamically updated, learnable fusion loss generation module. It selectively extracts source image information, minimizing the loss in downstream tasks. This ensures optimal fusion performance while maximizing adaptability to downstream tasks. +- We analyze the information preferences of downstream tasks such as semantic segmentation and object detection, providing deeper insights into multi-modal high-level vision tasks. +- TDFusion achieves outstanding performance in both fusion and high-level vision tasks, validated on four fusion datasets with semantic segmentation and object detection. + +# 2. Related Work + +# 2.1. Deep learning-based Image Fusion + +Deep learning-based image fusion methods have revolutionized the field by exploiting the powerful feature extraction capabilities of neural networks [4, 6, 33, 35, 36, 42-44, 65, 77]. These methods are broadly categorized into discriminative and generative approaches. Discriminative methods [10, 81, 85, 88] leverage the strong reconstruction ability of neural networks to directly learn the mapping between source and fused images [21, 23, 80, 82]. Generative methods, on the other hand, model the image generation process using generative approaches, integrating source images from a distributional perspective. These include methods based on Generative Adversarial Networks [46-48, 87] and diffusion models [73, 84]. Unified fusion methods [66, 75, 89] bridge the gap between different fusion sub-tasks, incorporating strategies such as continual learning [67] and self-supervised decomposition techniques [29]. The introduction of registration modules helps to mitigate misalignment issues [16, 60, 68] in source images. Recent studies have further explored the synergy between fusion and high-level vision tasks. These include leveraging downstream task losses to optimize fusion networks [32, 56], embedding high-level task features [34, 76, 78], and employing initialization techniques [40]. + +# 2.2. Meta-Learning in Vision + +Meta-learning develops algorithms to automatically fine-tune hyperparameters for specific tasks, showing its versatility and effectiveness across various domains. MAML [11] and its variations [12, 50, 53] focus on learning efficient initialization parameters to quickly adapt to new tasks using minimal data. Meta-SGD [26] extends MAML by learning optimal update directions and rates, beneficial in few-shot learning scenarios. Other approaches like MW-Net [55] and L2RW [54] emphasize selecting relevant sample weights to tackle noisy data using a compact validation set. Additionally, some studies focus on improving model adaptability through learning loss functions [2, 7, 15]. + +![](images/ff889df2ea5748d77aca49d1752dbca2103ade44bfc297f2bb885d9f67cfe3cb.jpg) +Learning of Loss Generate Module +Learning of Fusion +Figure 1. The TDFusion workflow alternates between training the loss generation module and the fusion module. Training of the loss generation module involves both inner and outer updates, learned through meta-learning. + +In image fusion, learnable filters [22] enable the fusion of images at arbitrary resolutions. MetaFusion [78] introduces a mechanism that improves image fusion and object detection by aligning semantics with fusion-specific features. ReFusion [5] guides the learnable fusion loss for various fusion tasks by reconstructing the source images. Metalearning also supports neural architecture search [38, 40] to identify optimal network architectures for image fusion and customizes network initialization for various tasks [40]. + +# 2.3. Comparison with Existing Approaches + +We propose a novel image fusion method tailored for downstream tasks, leveraging a learnable fusion loss driven by task-specific objectives. Our method employs a meta-learning algorithm that alternates between inner and outer updates, allowing the downstream task loss to guide the optimization of learnable fusion parameters. This results in fused images that minimize the downstream task loss, enhancing their adaptability across various tasks. Unlike previous methods, our approach develops a task-specific fusion loss, shifting focus from traditional factors such as resolution and network structure, and avoiding reliance on predefined fusion loss terms. This renders our fusion framework more flexible and applicable to various scenarios. + +# 3. Method + +# 3.1. Overview + +Our TDFusion framework, as shown in Fig. 1, consists of a fusion network $\mathcal{F}(\cdot)$ , a downstream task network $\mathcal{T}(\cdot)$ and a fusion loss generation module $\mathcal{G}(\cdot)$ , which produces + +parameters for a learnable loss function. The parameters of these modules are denoted as $\theta_{\mathcal{F}}$ , $\theta_{\mathcal{T}}$ , and $\theta_{\mathcal{G}}$ , respectively. The one-step updated clones of $\mathcal{F}$ and $\mathcal{T}$ are denoted as $\mathcal{F}'$ and $\mathcal{T}'$ , with parameters $\theta_{\mathcal{F}'}$ and $\theta_{\mathcal{T}'}$ . During the updates, the fusion network and the loss generation module alternate in learning, as depicted by blue and purple in Fig. 1. The update of the loss generation module consists of inner and outer updates, detailed in the following subsections. $\mathcal{L}_f$ and $\mathcal{L}_t$ represent the learnable fusion loss and task-specific loss, with their formulations provided in the next section. + +# 3.2. Loss Function + +The learnable fusion loss $\mathcal{L}_f$ consists of the intensity term and the gradient term. The intensity term is defined by the output of the loss generation module $\{w_a, w_b\} = \mathcal{G}(I_a, I_b)$ , where $w_a$ and $w_b$ control the intensity preference in the fusion loss. The Softmax function in the loss generation module ensures $w_a^{ij} + w_b^{ij} = 1$ for each pixel, thereby selectively retaining the intensity information from the source images. The gradient term emphasizes higher gradient values from the input images [76, 83, 89], aiming to preserve maximal information from the source. The detailed formulation of the learnable fusion loss is as follows: + +$$ +\mathcal {L} _ {f} = \mathcal {L} _ {f} ^ {\text {i n t}} + \alpha \mathcal {L} _ {f} ^ {\text {g r a d}}, \tag {1} +$$ + +$$ +\mathcal {L} _ {f} ^ {i n t} = \frac {1}{H W} \sum_ {i j} \left[ \sum_ {k \in \{a, b \}} w _ {k} ^ {i j} \left(I _ {f} ^ {i j} - I _ {k} ^ {i j}\right) ^ {2} \right], \tag {2} +$$ + +$$ +\mathcal {L} _ {f} ^ {\text {g r a d}} = \frac {1}{H W} \sum_ {i j} \left| \nabla I _ {f} ^ {i j} - \max _ {k \in \{a, b \}} \left(\nabla I _ {k} ^ {i j}\right) \right|, \tag {3} +$$ + +where $\nabla$ denotes the Sobel operator, commonly employed for gradient extraction in image fusion [32, 57, 83]. The parameter $\alpha$ serves as a scaling factor, while $\mathcal{L}_f^{int}$ and $\mathcal{L}_f^{grad}$ represent the intensity loss and gradient loss, respectively. The weights $\{w_a, w_b\}$ control the emphasis of the loss function on the intensity information from each source image. These parameters enable the fusion process to selectively aggregate and incorporate relevant information from the source images. Variations in $\theta_{\mathcal{G}}$ lead to different configurations of $w_a$ and $w_b$ , thereby influencing the characteristics of the fusion loss. $\theta_{\mathcal{G}}$ undergoes updates driven by the high-level task loss during the in-step update process, as detailed in Sec. 3.4. + +The loss function $\mathcal{L}_t$ depends on the specific task. In this study, we adopt SegFormer [9] and YOLOv8 [19] for the downstream tasks, employing cross-entropy loss and YOLO loss [19] for each task, respectively. + +# 3.3. Dataset Partitioning + +In order to enhance the effectiveness of the loss generation module $\mathcal{G}$ in guiding the fusion tasks, we create non-overlapping subsets of size $M$ at each training epoch. The meta-training set $\{I_a^{mtr}, I_b^{mtr}\}$ and the meta-test set $\{I_a^{mts}, I_b^{mts}\}$ are randomly drawn from the fusion training set $\{I_a^{ftr}, I_b^{ftr}\}$ . These subsets are fed into the model sequentially during the training process of the loss generation module, covering both the inner and outer updates. The entire fusion training set $\{I_a^{ftr}, I_b^{ftr}\}$ is fully utilized during the training of the fusion network. + +# 3.4. Learning of loss generation module + +Fig. 1 illustrates that the loss generation and fusion modules are trained alternately, ensuring the fusion loss is optimized at various stages of training and under different states of the fusion network. The process of training to optimize the fusion loss involves two key steps: the inner update and the outer update. In the inner update, clones of both the fusion network and the task network are generated, with each network undergoing a single training iteration using the fusion loss and task loss, respectively. This process is designed to obtain the state of the network guided by the fusion loss. During the outer update, the task loss of the fusion image produced by the updated clone is calculated. This loss is then backpropagated to the loss generation network. The goal of this step is to direct the fusion network to generate fusion images that result in lower downstream task loss, once guided by the fusion loss. The alternating updates between the inner and outer steps constitute the learning procedure for the loss generation module. + +# 3.4.1. Inner Update + +During the inner update phase, the fusion network $\mathcal{F}$ undergoes a single update guided by the fusion loss, which + +depends on the current state of network $\mathcal{G}$ . This update primarily aims to compute the intermediate parameters $\theta_{\mathcal{F}'}$ and $\theta_{\mathcal{R}'}$ , which are crucial for updating $\theta_{\mathcal{G}}$ in the subsequent phase. The upper section of the purple region in Fig. 1 illustrates this process. During this phase, the images from the meta-training set $\{I_a^{mtr}, I_b^{mtr}\}$ are fed into the model: + +$$ +\theta_ {\mathcal {F} ^ {\prime}} = \theta_ {\mathcal {F}} - \eta_ {\mathcal {F} ^ {\prime}} \frac {\partial \mathcal {L} _ {f} \left(I _ {a} ^ {m t r} , I _ {b} ^ {m t r} , I _ {f} ^ {m t r} ; \theta_ {\mathcal {G}}\right)}{\partial \theta_ {\mathcal {F}}}, \tag {4} +$$ + +where $\mathcal{F}$ undergoes a single gradient descent update. $\theta_{\mathcal{G}}$ represent the parameters of the loss generation module $\mathcal{G}$ , which determine the parameters for the learnable fusion loss. And the notation $\eta_{\mathcal{F}'}$ refers to the step size. The module $\mathcal{F}'$ temporarily substitutes $\mathcal{F}$ , adjusting its parameters in one update step. Meanwhile, the parameters of $\mathcal{F}$ , denoted as $\theta_{\mathcal{F}}$ , remain unchanged. Similarly, $\mathcal{T}'$ is updated in a single step using the parameters $\theta_{\mathcal{T}}$ from the current task network $\mathcal{T}$ : + +$$ +\theta_ {\mathcal {T} ^ {\prime}} = \theta_ {\mathcal {T}} - \eta_ {\mathcal {T} ^ {\prime}} \frac {\partial \mathcal {L} _ {t} \left(I _ {f} ^ {m t r}\right)}{\partial \theta_ {\mathcal {T}}}. \tag {5} +$$ + +The parameters $\theta_{\mathcal{F}'}$ and $\theta_{\mathcal{T}'}$ are both updated during the inner update, which also ensures that the computation graph of $\theta_{\mathcal{F}'}$ with respect to $\theta_{\mathcal{G}}$ is preserved. This preserved graph is essential for optimizing $\theta_{\mathcal{G}}$ during the outer update. + +# 3.4.2. Outer Update + +The primary objective of the outer update is to evaluate and refine the fusion guidance capability of $\mathcal{G}$ , specifically by strengthening the influence of the loss function $\mathcal{L}_f$ in steering the fusion module $\mathcal{F}$ . In the framework diagram, this stage is represented in the lower part of the purple region. The modules $\mathcal{F}'$ and $\mathcal{T}'$ , derived from the inner update, represent the current guidance capacity of $\mathcal{G}$ . In an ideal scenario, the optimal fusion loss should enhance the performance of the downstream task on the fused image. In this stage, the meta-test set $\{I_a^{mts}, I_b^{mts}\}$ is employed. The parameters $\theta_{\mathcal{G}}$ are subsequently updated using the task loss $\mathcal{L}_t$ , which is computed by $\mathcal{F}'$ and $\mathcal{T}'$ : + +$$ +\theta_ {\mathcal {G}} = \theta_ {\mathcal {G}} - \eta_ {\mathcal {G}} \frac {\partial \mathcal {L} _ {t} \left(I _ {f} ^ {m t s}\right)}{\partial \theta_ {\mathcal {G}}}, \tag {6} +$$ + +where $I_{f}^{mts} = \mathcal{F}'(I_{a}^{mts}, I_{b}^{mts})$ , and the gradient $\partial \mathcal{L}_t / \partial \theta_G$ can be calculated as: + +$$ +\frac {\partial \mathcal {L} _ {t}}{\partial \theta_ {\mathcal {G}}} = \frac {\partial \mathcal {L} _ {t}}{\partial \theta_ {\mathcal {F} ^ {\prime}}} * \left(- \eta_ {\mathcal {F} ^ {\prime}} \frac {\partial^ {2} \mathcal {L} _ {f} \left(I _ {a} ^ {m t r} , I _ {b} ^ {m t r} , I _ {f} ^ {m t r} ; \theta_ {\mathcal {G}}\right)}{\partial \theta_ {\mathcal {F}} \partial \theta_ {\mathcal {G}}}\right). \tag {7} +$$ + +Eq. (6) holds because the task loss $\mathcal{L}_t$ is determined by $I_{f}^{mts}$ , which in turn relies on $\theta_{\mathcal{F}'}$ . The optimization of $\theta_{\mathcal{G}}$ via $\mathcal{L}_t$ is realized by preserving the computational relationship between $\theta_{\mathcal{F}'}$ and $\theta_{\mathcal{G}}$ throughout the inner update. The updated $\mathcal{G}$ module gains the ability to generate enhanced fusion loss functions, enabling the fusion module to integrate relevant information from the source images more efficiently into the fused output for downstream tasks. + +Algorithm 1 TDFusion Training Algorithm +Require: Training set $\{I_a^{ftr}, I_b^{ftr}\}$ with size $N$ . Output: Thoroughly trained $\theta_{\mathcal{F}}, \theta_{\mathcal{T}}, \theta_{\mathcal{G}}$ . +1: Initialize $\theta_{\mathcal{F}}, \theta_{\mathcal{T}}, \theta_{\mathcal{G}}$ . +2: for epoch = 1 to L do +3: Sample $\{I_a^{mtr}, I_b^{mtr}\}$ and $\{I_a^{mts}, I_b^{mts}\}$ . +4: for step = 1 to M do + % Inner update: apply G. +5: Sample $(I_a^{mtr}, I_b^{mtr})$ and get $(I_f^{mtr}, y^{mtr})$ . +6: Compute $\theta_{\mathcal{F}'}$ and $\theta_{\mathcal{T}'}$ by Eq. (4) and Eq. (5). +% Outer update: optimize G. +7: Sample $(I_a^{mts}, I_b^{mts})$ and get $(I_f^{mts}, y^{mts})$ . +8: Update $\theta_{\mathcal{G}}$ by Eq. (6). +9: end for +10: for step = 1 to N do + % Fusion update: optimize F and T. +11: Sample $(I_a^{ftr}, I_b^{ftr})$ and get $(I_f^{ftr}, y^{ftr})$ . +12: Update $\theta_{\mathcal{F}}$ and $\theta_{\mathcal{T}}$ by Eq. (8) and Eq. (9). +13: end for +14: end for + +# 3.5. Learning of Fusion Network + +The alternating inner and outer update iterations form a flexible and effective mechanism to refine $\mathcal{G}$ in response to the evolving state of $\mathcal{F}$ . After refining $\mathcal{G}$ , it is utilized to further improve the training of $\mathcal{F}$ . This stage, denoted in blue in the diagrams, involves processing the fusion training set images $\{I_a^{ftr}, I_b^{ftr}\}$ . Both $\mathcal{F}$ and $\mathcal{T}$ are updated through the application of the fusion loss $\mathcal{L}_f$ and task loss $\mathcal{L}_t$ : + +$$ +\theta_ {\mathcal {F}} = \theta_ {\mathcal {F}} - \eta_ {\mathcal {F}} \frac {\partial \mathcal {L} _ {f} \left(I _ {a} ^ {f t r} , I _ {b} ^ {f t r} , I _ {f} ^ {f t r} ; \theta_ {\mathcal {G}}\right)}{\partial \theta_ {\mathcal {F}}}, \tag {8} +$$ + +$$ +\theta_ {\mathcal {T}} = \theta_ {\mathcal {T}} - \eta_ {\mathcal {T}} \frac {\partial \mathcal {L} _ {t} \left(I _ {f} ^ {f t r}\right)}{\partial \theta_ {\mathcal {T}}}. \tag {9} +$$ + +After completing several training sessions on the fusion network, the focus then shifts back to the learning phase of the fusion generation module. The fusion framework evolves through a series of alternating phases, with each phase finetuning the fusion loss based on the fusion network's current state. Such alternating phases ensure that the fusion network consistently applies the most suitable fusion loss during its progression. Ultimately, this results in the development of a highly efficient fusion network, optimized for peak performance. The complete training procedure is detailed in Algorithm 1. + +# 3.6. Network Architecture + +TDFusion is composed of three modules: the fusion network, the downstream task network, and the loss generation network. The fusion network shares the same architecture as [5], a lightweight model built upon the Restormer + +Block (RTB) [74]. An adaptive fusion module is incorporated into this network to facilitate feature integration. Fig. 1 illustrates the structure of the loss generation module. It also employs the Restormer Block (RTB) [74] as its primary component, receiving inputs $\{I_a, I_b\}$ . After applying Softmax( $\cdot$ ), the final output guarantees $w_a^{ij} + w_b^{ij} = 1$ for each pixel. This design ensures that the fused image meets the similarity constraints and removes the reliance on initialization within the loss generation module. The architecture of the downstream task network $\mathcal{T}(\cdot)$ depends on the specific task. For learning of loss generation module, we chose the most lightweight models of SegFormer [9] and YOLOv8 [19] for semantic segmentation and object detection, respectively. + +# 3.7. Theoretical Analysis + +To better understand the weighting mechanism of the loss generation module $\mathcal{G}$ , we investigate the optimization procedure of $\mathcal{G}$ , which generates the weights $\{w_{a}, w_{b}\}$ , denoted as $\theta_{\mathcal{G}}$ . For clarity, we rewrite Eq. (2) as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {f} ^ {i n t} = \left[ w _ {a} \odot \left(I _ {a} - I _ {f}\right) \odot \left(I _ {a} - I _ {f}\right) \right. \\ \left. + w _ {b} \odot \left(I _ {b} - I _ {f}\right) \odot \left(I _ {b} - I _ {f}\right) \right] \times \frac {1}{H W} \\ = \left[ \mathcal {G} \left(I _ {a}, I _ {b}; \theta_ {\mathcal {G}}\right) \odot \left(I _ {a} - \mathcal {F} _ {\theta_ {\mathcal {F}}} \left(I _ {a}, I _ {b}\right)\right) \right. \tag {10} \\ \odot \left(I _ {a} - \mathcal {F} _ {\theta_ {\mathcal {F}}} \left(I _ {a}, I _ {b}\right)\right) \\ + \left(1 - \mathcal {G} \left(I _ {a}, I _ {b}; \theta_ {\mathcal {G}}\right)\right) \odot \left(I _ {b} - \mathcal {F} _ {\theta_ {\mathcal {F}}} \left(I _ {a}, I _ {b}\right)\right) \\ \left. \odot \left(I _ {b} - \mathcal {F} _ {\theta_ {\mathcal {F}}} \left(I _ {a}, I _ {b}\right)\right) \right] \times \frac {1}{H W}. \\ \end{array} +$$ + +Here, $w_{a}, w_{b} \in \mathbb{R}^{H \times W}$ , $I_{a}, I_{b} \in \mathbb{R}^{H \times W}$ , and $\odot$ denotes the element-wise multiplication operation. Let $\Omega'$ be the set $\{\theta_{\mathcal{F}'}', \theta_{\mathcal{T}''}\}$ , this leads to the following expression: + +$$ +\begin{array}{l} \theta_ {\mathcal {G}} = \theta_ {\mathcal {G}} - \eta_ {\mathcal {G}} \frac {\partial \mathcal {L} _ {t} ^ {m t s} (\Omega^ {'} (\theta_ {\mathcal {G}}))}{\partial \theta_ {\mathcal {G}}} \\ = \theta_ {\mathcal {G}} - \eta_ {\mathcal {G}} \frac {\partial \mathcal {L} _ {t} ^ {m t s} (\Omega^ {'} (\theta_ {\mathcal {G}}))}{\partial \Omega^ {'}} \frac {\partial \Omega^ {'} (\theta_ {\mathcal {G}})}{\partial \theta_ {\mathcal {G}}} \\ = \theta_ {\mathcal {G}} - \eta_ {\mathcal {G}} \eta_ {\mathcal {F} ^ {\prime}} \underbrace {\frac {\partial \mathcal {L} _ {t} ^ {m t s} \left(\Omega^ {'} \left(\theta_ {\mathcal {G}}\right)\right)}{\partial \Omega^ {\prime}}} _ {\text {(a)}} \times \frac {\partial \mathcal {G} \left(I _ {a} , I _ {b} ; \theta_ {\mathcal {G}}\right)}{\theta_ {\mathcal {G}}} \\ \times \underbrace {\left[ \left(I _ {a} - \frac {\partial \mathcal {F} _ {\theta_ {\mathcal {F}}}}{\partial \theta_ {\mathcal {F}}}\right) \odot \left(I _ {a} - \frac {\partial \mathcal {F} _ {\theta_ {\mathcal {F}}}}{\partial \theta_ {\mathcal {F}}}\right) \right.} _ {(b)} \tag {11} \\ \underbrace {- \left(I _ {b} - \frac {\partial \mathcal {F} _ {\theta_ {\mathcal {F}}}}{\partial \theta_ {\mathcal {F}}}\right) \odot \left(I _ {b} - \frac {\partial \mathcal {F} _ {\theta_ {\mathcal {F}}}}{\partial \theta_ {\mathcal {F}}}\right) ]} _ {(b)} \\ = \theta_ {\mathcal {G}} - \eta_ {\mathcal {G}} \eta_ {\mathcal {F} ^ {\prime}} \mathbf {G} \times \frac {\partial \mathcal {G} \left(I _ {a} , I _ {b} ; \theta_ {\mathcal {G}}\right)}{\theta_ {\mathcal {G}}}. \\ \end{array} +$$ + +Here, $\mathbf{G}$ denotes the inner product between two gradients: (a) the first one is derived from the task loss using a meta-testing set, and (b) the second is calculated from the fusion loss based on a meta-training set. Consequently, the + +![](images/65a1c92046391bc694d653a9a75e71f8a8ba94d6cfdba5037aa2a9bc3eb03716.jpg) +Figure 2. Visual comparison of fusion results. The cases are "01258N" in MSRS dataset, "00122" in FMB dataset, "00449" in M3FD dataset and "200304" in LLVIP dataset. + +optimization of the module $\theta_{\mathcal{G}}$ is driven by the task loss, with the objective of preserving task-specific information throughout the fusion process. + +# 4. Experiment + +# 4.1. Setup + +Experimental Setup. In our experiments, the epoch number $L$ and the training iterations of the loss generation module $M$ are set to 50 and 200. The learning iterations for the fusion network $N$ depend on the size of the dataset. We use the Adam optimizer with a learning rate of 1e-4, a batch size of 2, and a hyperparameter $\alpha$ set to 1. All experiments are conducted on a PC with a single NVIDIA RTX 3090 GPU. + +Evaluation Metrics and Comparison Methods. The advanced fusion methods compared in our study include TarDAL [32], SegMIF [34], MURF [69], EMMA [86], DCINN [61], MRFS [76], and TIMFusion [40]. The fusion performance is evaluated using metrics including entropy (EN), spatial frequency (SF), sum of correlation differences (SCD), visual information fidelity (VIF), $Q^{AB / F}$ and structural similarity index metric (SSIM). + +Dataset Split. We use four datasets annotated for downstream tasks, including MSRS [58], FMB [34], M3FD [32], and LLVIP [17]. MSRS contains 1083/361 image pairs for training/test, and FMB contains 1220/280 pairs for training/test. We follow the splits of the original papers for both datasets. M3FD dataset consists of 4200 pairs of images for detection, with 300 pairs designated for fusion evalu- + +ation. The 4200 detection images are split into 3150 for training and 1050 for testing, ensuring that the 300 pairs for fusion evaluation are included within the detection test set. These 300 fusion images are then employed to assess the fusion performance. The original LLVIP dataset contains 12025/3463 image pairs as training/test set. Due to its large size, we select every 10th image to form our training and test sets, resulting in 1203/347 image pairs for training and testing. Our splits for M3FD and LLVIP will be available. + +# 4.2. Fusion Experiments + +Fig. 2 presents a visual comparison of different methods. The fused images generated by TDFusion excel in preserving discriminative details, achieving balanced brightness, and maintaining clear object contours. It effectively preserves the target features from the infrared images and the background details from the visible images, resulting in fused images that are more natural and exhibit greater clarity across different environments. These results highlight the advantages of TDFusion in detail preservation and visual performance. More results are available in the supplementary material. Tab. 1 presents the quantitative comparison of fusion over four datasets. TDFusion outperforms other methods across most metrics. This suggests that TDFusion not only enhances image details but also provides consistent fusion results across various scenarios. Compared to other methods, TDFusion shows superior adaptability and robustness, particularly in handling diverse image characteristics and challenging fusion tasks. + +Table 1. Quantitative comparison of Infrared-visible image fusion. The red and blue markers represent the best and second-best values. + +
Infrared-visible Image Fusion on MSRS [58] DatasetInfrared-visible Image Fusion on FMB [34] Dataset
EN ↑SF ↑SCD ↑VIF ↑\(Q^{AB/F}\) ↑SSIM ↑EN ↑SF ↑SCD ↑VIF ↑\(Q^{AB/F}\) ↑SSIM ↑
TarDAL [32]5.285.980.710.210.180.47TarDAL [32]6.636.941.030.280.290.74
SegMIF [34]5.9511.101.570.440.630.55SegMIF [34]6.8313.691.720.390.650.60
MURF [69]5.0410.491.020.220.370.60MURF [69]6.3713.881.340.220.370.68
EMMA [86]6.7311.561.620.490.640.70EMMA [86]6.7715.001.500.420.650.72
DCINN [61]6.0010.511.490.410.570.52DCINN [61]6.4711.471.390.380.590.74
MRFS [76]7.008.861.420.370.490.55MRFS [76]6.7812.421.240.380.620.73
TIMFusion [40]6.279.671.340.320.480.68TIMFusion [40]6.5112.231.240.350.590.73
TDFusion (Ours)6.7411.301.860.500.670.70TDFusion (Ours)6.8614.161.760.430.680.75
Infrared-visible Image Fusion on M3FD [32] DatasetInfrared-visible Image Fusion on LLVIP [17] Dataset
EN ↑SF ↑SCD ↑VIF ↑\(Q^{AB/F}\) ↑SSIM ↑EN ↑SF ↑SCD ↑VIF ↑\(Q^{AB/F}\) ↑SSIM ↑
TarDAL [32]6.877.631.290.270.300.71TarDAL [32]6.327.421.040.270.220.58
SegMIF [34]6.8514.141.720.370.600.59SegMIF [34]6.6815.461.380.400.660.57
MURF [69]6.5012.551.460.210.320.64MURF [69]6.1315.080.960.210.310.57
EMMA [86]6.9215.231.490.380.590.69EMMA [86]7.3515.371.570.410.640.66
DCINN [61]6.5911.211.460.340.510.72DCINN [61]6.9813.341.430.380.520.64
MRFS [76]6.9412.071.260.340.550.70MRFS [76]6.8311.041.230.310.420.64
TIMFusion [40]6.7512.311.370.350.530.70TIMFusion [40]6.5813.521.140.330.460.64
TDFusion (Ours)6.9914.491.830.410.650.72TDFusion (Ours)7.3616.381.750.460.700.67
+ +![](images/c5baf043834cf4bf3ae738fc93d4d04004447215b3caaee616a34c56078459c4.jpg) +Figure 3. Visual comparison for Semantic Segmentation. The cases are "00726N" in MSRS dataset and "01438" in FMB dataset. + +# 4.3. Downstream Applications + +This section validates the adaptability of fusion methods to downstream tasks. For a fair comparison, we adopt SegFormer [9] and YOLOv8 [19] as backbones and retrain the task networks for each fusion method over 300 epochs to evaluate their adaptability to semantic segmentation and object detection. Fig. 3 and Fig. 4 present the visual comparisons of semantic segmentation and object detection, respectively. TDFusion outperforms in image detail retention, edge clarity, and object recognition, effectively identifying and segmenting objects. For semantic segmentation, the generated maps clearly distinguish different class regions, closely matching the ground truth. In object detection, the fused images exhibit more precise boundary localization for salient objects. This indicates that TDFusion maintains a better balance between fine details and overall context. More results can be found in the supplementary material. + +Tab. 2 shows the performance comparison across different methods in semantic segmentation and object detection. TDFusion outperforms other methods on most metrics, particularly in mIoU and mAP. This demonstrates that TDFu + +Table 2. Performance comparison of downstream applications. The red and blue markers represent the best and second-best. + +
MethodsSemantic SegmentationObject Detection
MSRSFMBM3FDLLVIP
mAccmIoUmAccmIoUmAP50mAP75AP50AP75
Infrared83.2369.4958.8551.9879.1253.0596.0372.07
Visible83.4473.7665.1257.9682.2154.8291.7848.66
TarDAL81.9371.3562.8655.3383.1656.3993.7962.71
SegMIF85.7374.2565.9758.4183.6158.2393.9566.45
MURF85.0374.0864.1056.9680.5854.2294.2468.04
EMMA85.9974.4862.4556.2883.7156.9194.0066.21
DCINN84.1174.3561.0954.8182.6957.3794.9268.34
MRFS84.7674.5061.9355.7183.2857.7493.0367.21
TIMFusion83.6773.5863.7057.2483.2256.0893.7661.33
TDFusion86.0475.0967.1760.5086.2759.7195.0069.18
+ +sion effectively enhances the quality of fused images. It also improves the downstream task performance, especially in terms of accuracy and robustness in complex scenarios. Class-wise results are provided in the supplementary material. + +# 4.4. Task-driven Learnable Loss + +Our framework incorporates a learnable fusion loss that models the preferences of downstream tasks for information from source images. The models trained on FMB dataset + +![](images/5dd3227b95ffe4277b7733b909d3963d469aa7be65f04d769127fe7da6e3408e.jpg) +Figure 4. Visual comparison for Object Detection. The cases are "02236" in M3FD dataset and "210145" in LLVIP dataset. + +![](images/cc24fe26ef625184c016f17bfbe192a79954665885b86321e4b8a355c403d80c.jpg) +Figure 5. Visualisation of learnable loss for downstream tasks. + +![](images/d562ac6c344a11d3f1bdb8b4f9901cc2cfb66ba113513b9c6660c1871c99d2bd.jpg) + +and LLVIP dataset labeled as $SS$ and $OD$ , are evaluated on MSRS dataset to simulate performance in unknown scenes, as shown in Fig. 5. The results demonstrate that the fusion model adaptively selects information from infrared and visible images to satisfy task requirements. In semantic segmentation, the model combines scene structure and texture, prioritizing boundaries. This improves segmentation under varying lighting conditions. Fusion weights $\{w_{a}^{SS}, w_{b}^{SS}\}$ indicate a preference for visible details and infrared advantages in low-light conditions. In object detection, the model focuses on edge and contrast information, especially for instances like pedestrians and vehicles. Higher fusion weights are assigned to bright regions in infrared images, enhancing target detection in low-light conditions. Fusion weights $\{w_{a}^{OD}, w_{b}^{OD}\}$ reflect this preference. Comparison of fusion losses across tasks reveals distinct differences, especially in highlighted regions, confirming that the model adapts to task-specific requirements by selecting the most relevant information from multimodal images. More results are provided in the supplementary material. + +# 4.5. Ablation Studies + +To thoroughly evaluate the performance of our proposed algorithm, we conduct a series of ablation experiments on FMB dataset, and the detailed results are shown in Tab. 3. In Exp. I, we exclude the learnable fusion loss by fixing $w_{a}$ and $w_{b}$ to $1/2$ . In Exp. II, we omit the gradient loss from the loss function. In Exp. III, we also allow the fusion module parameters to be jointly optimized by both the task loss $\mathcal{L}_{t}$ and fusion loss $\mathcal{L}_{f}$ . In Exp. IV, we exclude the fusion + +Table 3. Ablation experiment of fusion. The red denotes the best. + +
Ablation Studies of fusion on FMB Dataset
ConfigurationsENSFSCDVIF\(Q^{AB/F}\)SSIM
Ifix \(w_{a}\) and \(w_{b}\) as 1/26.6013.731.580.390.600.72
IIw/o \(\mathcal{L}_{f}^{grad}\)6.7711.651.630.370.640.73
III\(\theta_{\mathcal{F}}\) influenced by \(\mathcal{L}_{t}\)6.8013.851.700.410.660.73
IVw/o Fusion learning6.8214.071.720.410.670.72
V\(I_{f}=w_{a}*I_{a}+w_{b}*I_{b}\)6.7511.491.650.380.620.73
Ours6.8614.161.760.430.680.75
+ +module's dedicated learning phase and update it during the outer update of the loss module, which tests the impact of the fusion module's training schedule. In Exp. V, we replace our fusion method with a discriminative approach. The performance decline observed across different configurations confirms the rationality and effectiveness of our proposed method. Visualization and more analysis are provided in the supplementary material. + +# 5. Conclusion + +To overcome the limitations of predefined fusion losses, which often fail to effectively guide the fusion process for downstream tasks, we propose a meta-learning-based framework for task-guided fusion. This framework includes a loss generation module that outputs the parameters of the learnable fusion loss. The module is updated using a meta-learning approach, which alternates between inner and outer loop steps to enhance its ability to guide the fusion network. Under varying fusion conditions, this module generates the optimal fusion loss for the downstream task. This enables the fusion network to produce fused images that minimize the task-specific loss. The theoretical analysis explains how the downstream task loss guides the fusion loss in our framework. Experiments using four publicly available fusion datasets and downstream tasks including semantic segmentation and object detection, demonstrate the effectiveness of our approach. + +# Acknowledgement + +This work has been supported by the National Natural Science Foundation of China under Grant 12201497 and 12371512. + +# References + +[1] Hessah Albanwan, Rongjun Qin, and Yang Tang. Image fusion in remote sensing: An overview and meta-analysis. Photogrammetric Engineering & Remote Sensing, 90(12): 755-775, 2024. 1 +[2] Antreas Antoniou and Amos J. Storkey. Learning to learn by self-critique. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), pages 9936-9946, 2019. 2 +[3] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yukun Cui, Chunxia Zhang, Zhenbo Guo, and Yongjun Wang. Simultaneous automatic picking and manual picking refinement for first-break. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[4] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Baisong Jiang, Lilun Deng, Yukun Cui, Shuang Xu, and Chunxia Zhang. Deep unfolding multi-modal image fusion network via attribution analysis. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2 +[5] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yichen Wu, Lilun Deng, Yukun Cui, Baisong Jiang, and Shuang Xu. Refusion: Learning image fusion from reconstruction with learnable loss via meta-learning. International Journal of Computer Vision, pages 1-21, 2024. 3, 5 +[6] Haowen Bai, Jiangshe Zhang, Zixiang Zhao, Lilun Deng, Yukun Cui, and Shuang Xu. Retinex-mef: Retinex-based glare effects aware unsupervised multi-exposure image fusion. arXiv preprint arXiv:2503.07235, 2025. 2 +[7] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 9445-9454. IEEE, 2021. 2 +[8] Yanpeng Cao, Dayan Guan, Weilin Huang, Jiangxin Yang, Yanlong Cao, and Yu Qiao. Pedestrian detection with unsupervised multispectral feature learning using deep neural networks. Information Fusion, 46:206-217, 2019. 1 +[9] Bo Cheng, Xiang Li, Yujie Wei, Cheng Huang, Xiaoyong Zhang, Yandong Jiang, Tianyu Zhang, Na Xu, Shuai Yu, Xinxin Zhan, et al. Segformer: Simple and efficient design for semantic segmentation with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12022-12032, 2021. 4, 5, 7 +[10] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3333-3348, 2021. 2 +[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the International conference on machine learning (ICML), pages 1126–1135, 2017. 2 +[12] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In Proceedings of the International conference on machine learning (ICML), pages 1920-1930, 2019. 2 +[13] Qishen Ha, Kohei Watanabe, Takumi Karasawa, Yoshitaka Ushiku, and Tatsuya Harada. Mfnet: Towards real-time se + +mantic segmentation for autonomous vehicles with multispectral scenes. In IROS, pages 5108-5115. IEEE, 2017. 1 +[14] Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Task-driven super resolution: Object detection in low-resolution images. In Neural Information Processing: 28th International Conference (ICONIP), pages 387-395. Springer, 2021. 1 +[15] Rein Houthooft, Yuhua Chen, Phillip Isola, Bradley C. Stadie, Filip Wolski, Jonathan Ho, and Pieter Abbeel. Evolved policy gradients. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), pages 5405-5414, 2018. 2 +[16] Zhanbo Huang, Jinyuan Liu, Xin Fan, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Reconet: Recurrent correction network for fast and efficient multi-modality image fusion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 539-555. Springer, 2022. 2 +[17] Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible-infrared paired dataset for low-light vision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3496-3504, 2021. 6, 7 +[18] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 4369-4376, 2020. 1 +[19] Glenn Jocher. Ultralytics YOLOv8. https://github.com/ultralytics, 2023.4,5,7 +[20] Chenglong Li, Chengli Zhu, Yan Huang, Jin Tang, and Liang Wang. Cross-modal ranking with soft consistency and noisy labels for robust RGB-T tracking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 831-847. Springer, 2018. 1 +[21] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 1, 2 +[22] Huafeng Li, Yueliang Cen, Yu Liu, Xun Chen, and Zhengtao Yu. Different input resolutions and arbitrary output resolution: A meta learning-based deep framework for infrared and visible image fusion. IEEE Transactions on Image Processing, 30:4070-4083, 2021. 3 +[23] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 1, 2 +[24] Siyuan Li, Iago Bruno Araujo, Wenqi Ren, Zhangyang Wang, Eric K. Tokuda, Roberto Hirata Junior, Roberto Marcondes Cesar Junior, Jiawan Zhang, Xiaojie Guo, and Xiaochun Cao. Single image deraining: A comprehensive benchmark analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3838-3847. Computer Vision Foundation / IEEE, 2019. 1 +[25] Xiaoling Li, Yanfeng Li, Houjin Chen, Yahui Peng, and Pan Pan. Ccafusion: cross-modal coordinate attention network for infrared and visible image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1 + +[26] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 2 +[27] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 1 +[28] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 1 +[29] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 719-735. Springer, 2022. 2 +[30] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 2 +[31] Jinyuan Liu, Xin Fan, Ji Jiang, Risheng Liu, and Zhongxuan Luo. Learning a deep multi-scale feature ensemble and an edge-attention guidance for image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 32(1): 105–119, 2021. 1 +[32] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5792-5801, 2022. 1, 2, 4, 6, 7 +[33] Jinyuan Liu, Jingjie Shang, Risheng Liu, and Xin Fan. Attention-guided global-local adversarial learning for detail-preserving multi-exposure image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5026-5040, 2022. 2 +[34] Jinyuan Liu, Zhu Liu, Guanyao Wu, Long Ma, Risheng Liu, Wei Zhong, Zhongxuan Luo, and Xin Fan. Multi-interactive feature learning and a full-time multi-modality benchmark for image fusion and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8115–8124, 2023. 1, 2, 6, 7 +[35] Jinyuan Liu, Guanyao Wu, Junsheng Luan, Zhiying Jiang, Risheng Liu, and Xin Fan. Holoco: Holistic and local contrastive learning network for multi-exposure image fusion. Information Fusion, 95:237-249, 2023. 2 +[36] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, 132(5):1748-1775, 2024. 2 +[37] Jinyuan Liu, Guanyao Wu, Zhu Liu, Di Wang, Zhiying Jiang, Long Ma, Wei Zhong, and Xin Fan. Infrared and visible image fusion: From data compatibility to task adaption. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 + +[38] Risheng Liu, Zhu Liu, Jinyuan Liu, and Xin Fan. Searching a hierarchically aggregated fusion architecture for fast multi-modality image fusion. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 1600-1608. ACM, 2021. 3 +[39] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5953-5969, 2022. 1 +[40] Risheng Liu, Zhu Liu, Jinyuan Liu, Xin Fan, and Zhongxuan Luo. A task-guided, implicitly-searched and metainitialized deep model for image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2, 3, 6, 7 +[41] ShuMin Liu, Jiajia Chen, and Susanto Rahardja. A new multi-focus image fusion algorithm and its efficient implementation. IEEE Transactions on Circuits and Systems for Video Technology, 30(5):1374-1384, 2020. 1 +[42] Zhu Liu, Jinyuan Liu, Guanyao Wu, Long Ma, Xin Fan, and Risheng Liu. Bi-level dynamic learning for jointly multimodality image fusion and beyond. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1240-1248, 2023. 2 +[43] Zhu Liu, Jinyuan Liu, Benzhuang Zhang, Long Ma, Xin Fan, and Risheng Liu. Paif: Perception-aware infrared-visible image fusion for attack-tolerant semantic segmentation. In Proceedings of the 31st ACM international conference on multimedia, pages 3706-3714, 2023. +[44] Zhu Liu, Jinyuan Liu, Guanyao Wu, Zihang Chen, Xin Fan, and Risheng Liu. Searching a compact architecture for robust multi-exposure image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6224-6237, 2024. 2 +[45] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information Fusion, 45:153-178, 2019. 1 +[46] Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusion: A generative adversarial network for infrared and visible image fusion. Information Fusion, 48:11-26, 2019. 2 +[47] Jiayi Ma, Pengwei Liang, Wei Yu, Chen Chen, Xiaojie Guo, Jia Wu, and Junjun Jiang. Infrared and visible image fusion via detail preserving adversarial learning. Information Fusion, 54:85-98, 2020. +[48] Jiayi Ma, Han Xu, Junjun Jiang, Xiaoguang Mei, and Xiao-Ping (Steven) Zhang. Ddcgan: A dual-discriminator conditional generative adversarial network for multi-resolution image fusion. IEEE Transactions on Image Processing, 29: 4980-4995, 2020. 2 +[49] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5637-5646, 2022. 1 +[50] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2 + +[51] Seonghyun Park, An Gia Vien, and Chul Lee. Cross-modal transformers for infrared and visible image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 34(2):770-785, 2023. 1 +[52] Yanting Pei, Yaping Huang, Qi Zou, Yuhang Lu, and Song Wang. Does haze removal help cnn-based image classification? In Proceedings of the European conference on computer vision (ECCV), pages 682-697, 2018. 1 +[53] Xinran Qin, Yuhui Quan, Tongyao Pang, and Hui Ji. Ground-truth free meta-learning for deep compressive sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9947-9956, 2023. 2 +[54] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In Proceedings of the International conference on machine learning (ICML), pages 4334-4343, 2018. 2 +[55] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2019. 2 +[56] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1, 2 +[57] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 4 +[58] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Infromation Fusion, 83-84:79-92, 2022. 6, 7 +[59] Wei Tang, Fazhi He, Yu Liu, Yansong Duan, and Tongzhen Si. Datfuse: Infrared and visible image fusion via dual attention transformer. IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159-3172, 2023. 1 +[60] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 3508-3515. ijcai.org, 2022. 2 +[61] Wu Wang, Liang-Jian Deng, Ran Ran, and Gemine Vivone. A general paradigm with detail-preserving conditional invertible network for image fusion. International Journal of Computer Vision, 132(4):1029-1054, 2024. 6, 7 +[62] Zeyu Xiao and Xinchao Wang. Event-based video superresolution via state space models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1 +[63] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 2409-2418, 2024. 1 + +[64] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 217-235. Springer, 2024. 1 +[65] Han Xu, Jiayi Ma, Zhuliang Le, Junjun Jiang, and Xiaojie Guo. Fusiondn: A unified densely connected network for image fusion. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 12484-12491, 2020. 2 +[66] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 2 +[67] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 2 +[68] Han Xu, Jiayi Ma, Jiteng Yuan, Zhuliang Le, and Wei Liu. Rfnet: Unsupervised network for mutually reinforcing multimodal image registration and fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19647-19656, 2022. 2 +[69] Han Xu, Jiteng Yuan, and Jiayi Ma. Murf: Mutually reinforcing multi-modal image registration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 6, 7 +[70] Shuang Xu, Ouafa Amira, Junmin Liu, Chun-Xia Zhang, Jiangshe Zhang, and Guanghai Li. HAM-MFN: hyperspectral and multispectral image multiscale fusion network with RAP loss. IEEE Transactions on Geoscience and Remote Sensing, 58(7):4618-4628, 2020. 1 +[71] Shuang Xu, Lizhen Ji, Zhe Wang, Pengfei Li, Kai Sun, Chunxia Zhang, and Jiangshe Zhang. Towards reducing severe defocus spread effects for multi-focus image fusion via an optimization based strategy. IEEE Transactions Computational Imaging, 6:1561-1570, 2020. 1 +[72] Yong Yang, Jiaxiang Liu, Shuying Huang, Weiguo Wan, Wenying Wen, and Juwei Guan. Infrared and visible image fusion via texture conditional generative adversarial network. IEEE Transactions on Circuits and Systems for Video Technology, 31(12):4771-4783, 2021. 1 +[73] Xunpeng Yi, Linfeng Tang, Hao Zhang, Han Xu, and Jiayi Ma. Diff-if: Multi-modality image fusion via diffusion model with fusion knowledge prior. Information Fusion, 110:102450, 2024. 2 +[74] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728-5739, 2022. 5 +[75] Hao Zhang and Jiayi Ma. Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion. International Journal of Computer Vision, 129(10):2761-2785, 2021. 2 +[76] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Com + +puter Vision and Pattern Recognition (CVPR), pages 26974-26983, 2024. 1, 2, 3, 6, 7 +[77] Yu Zhang, Yu Liu, Peng Sun, Han Yan, Xiaolin Zhao, and Li Zhang. IFCNN: A general image fusion framework based on convolutional neural network. Information Fusion, 54: 99-118, 2020. 1, 2 +[78] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13955-13965, 2023. 1, 2, 3 +[79] Yangyang Zhao, Qingchun Zheng, Peihao Zhu, Xu Zhang, and Wenpeng Ma. Tufusion: A transformer-based universal fusion algorithm for multimodal images. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1 +[80] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Jiangshe Zhang, and Pengfei Li. Didfuse: Deep image decomposition for infrared and visible image fusion. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 970-976, 2020. 2 +[81] Zixiang Zhao, Shuang Xu, Jiangshe Zhang, Chengyang Liang, Chunxia Zhang, and Junmin Liu. Efficient and model-based infrared and visible image fusion via algorithm unrolling. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1186-1196, 2022. 2 +[82] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5906–5916, 2023. 1, 2 +[83] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5906–5916, 2023. 1, 2, 3, 4 +[84] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. DDFM: denoising diffusion model for multi-modality image fusion. CoRR, abs/2303.06840, 2023. 2 +[85] Zixiang Zhao, Jiang-She Zhang, Haowen Bai, Yicheng Wang, Yukun Cui, Lilun Deng, Kai Sun, Chunxia Zhang, Junmin Liu, and Shuang Xu. Deep convolutional sparse coding networks for interpretable image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2369-2377, 2023. 2 +[86] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25912-25921, 2024. 6, 7 +[87] Huabing Zhou, Wei Wu, Yanduo Zhang, Jiayi Ma, and Haibin Ling. Semantic-supervised infrared and visible image + +fusion via a dual-discriminator generative adversarial network. IEEE Transactions on Multimedia, 25:635-648, 2023. 1, 2 +[88] Man Zhou, Naishan Zheng, Xuanhua He, Danfeng Hong, and Jocelyn Chanussot. Probing synergistic high-order interaction for multi-modal image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[89] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 1, 2, 3 \ No newline at end of file diff --git a/2025/Task-driven Image Fusion with Learnable Fusion Loss/images.zip b/2025/Task-driven Image Fusion with Learnable Fusion Loss/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5982451c0d82b03950a426e857ea84054912daa --- /dev/null +++ b/2025/Task-driven Image Fusion with Learnable Fusion Loss/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b40471ef2fb9156f412db055209870f1c6ca46ba7c3de4105468721c143af90 +size 852119 diff --git a/2025/Task-driven Image Fusion with Learnable Fusion Loss/layout.json b/2025/Task-driven Image Fusion with Learnable Fusion Loss/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fd22441507de32fe2d863ac8128c49029d3aa5e1 --- /dev/null +++ b/2025/Task-driven Image Fusion with Learnable Fusion Loss/layout.json @@ -0,0 +1,10526 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 103, + 474, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 103, + 474, + 120 + ], + "spans": [ + { + "bbox": [ + 137, + 103, + 474, + 120 + ], + "type": "text", + "content": "Task-driven Image Fusion with Learnable Fusion Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "spans": [ + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "text", + "content": "Haowen Bai" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "text", + "content": " Jiangshe Zhang" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "text", + "content": " Zixiang Zhao" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "text", + "content": " Yichen Wu" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "text", + "content": " Lilun Deng" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "text", + "content": " Yukun Cui" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "text", + "content": " Tao Feng" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "text", + "content": " Shuang Xu" + }, + { + "bbox": [ + 149, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{5}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "spans": [ + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "text", + "content": "Xi'an Jiaotong University " + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "text", + "content": "ETH Zürich " + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "text", + "content": "City University of Hong Kong " + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 121, + 173, + 489, + 202 + ], + "type": "text", + "content": "Northwestern Polytechnical University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 205, + 364, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 205, + 364, + 216 + ], + "spans": [ + { + "bbox": [ + 242, + 205, + 364, + 216 + ], + "type": "text", + "content": "hwbai@stu.xjtu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 243, + 200, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 243, + 200, + 256 + ], + "spans": [ + { + "bbox": [ + 151, + 243, + 200, + 256 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 269, + 297, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 297, + 591 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 297, + 591 + ], + "type": "text", + "content": "Multi-modal image fusion aggregates information from multiple sensor sources, achieving superior visual quality and perceptual features compared to single-source images, often improving downstream tasks. However, current fusion methods for downstream tasks still use predefined fusion objectives that potentially mismatch the downstream tasks, limiting adaptive guidance and reducing model flexibility. To address this, we propose Task-driven Image Fusion (TD-Fusion), a fusion framework incorporating a learnable fusion loss guided by task loss. Specifically, our fusion loss includes learnable parameters modeled by a neural network called the loss generation module. This module is supervised by the downstream task loss in a meta-learning manner. The learning objective is to minimize the task loss of fused images after optimizing the fusion module with the fusion loss. Iterative updates between the fusion module and the loss module ensure that the fusion network evolves toward minimizing task loss, guiding the fusion process toward the task objectives. TDFusion's training relies entirely on the downstream task loss, making it adaptable to any specific task. It can be applied to any architecture of fusion and task networks. Experiments demonstrate TDFusion's performance through fusion experiments conducted on four different datasets, in addition to evaluations on semantic segmentation and object detection tasks. The code is available at https://github.com/HaowenBai/TDFusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 615, + 135, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 135, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 135, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "Multi-modal image fusion [25, 31, 51, 59, 72, 79, 82] combines information from multiple sensors to produce a more holistic and detailed representation. Infrared images capture thermal radiation regardless of lighting conditions, while visible images provide richer texture details. Fused im" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 245, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 461 + ], + "type": "text", + "content": "ages enhance downstream tasks through improved information density and robustness [18, 41, 45, 62, 63, 70, 71], outperforming single-modal inputs in semantic segmentation [13, 28, 34, 56], object detection [8, 32], and other related applications [1, 3, 20, 27]. Conventional methods typically treat fusion as image restoration using unsupervised loss [76, 83, 87-89] or perceptual loss [21, 23, 77]. These approaches prioritize visual-level fusion through predefined aggregation objectives, often neglecting semantic feature extraction. This limitation hinders scene interpretation and task performance [14, 24, 52, 64]. Recent advances explore the mutual enhancement between fusion and downstream tasks [37]. By cascading the fusion network with downstream task network [39, 49], the task loss constrains the fusion learning, ensuring the fused images meet the task requirements [32, 56]. Alternatively, some methods incorporate high-level visual task features [34, 76, 78] or focus on learning optimal initializations [40] to enhance fusion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": "While integrating downstream tasks, existing frameworks still rely on predefined fusion loss terms lacking dynamic adaptation. The impact of downstream tasks remains limited due to specific combinations. Manually defined losses preserve predefined guidance, frequently overlooking task-specific requirements. This guidance imposes manually designed prior constraints on the fusion process, limiting the dynamic and adaptive influence of downstream tasks on specific image pairs. These approaches, whether incorporating task features [34, 78] or employing task losses [32, 56], still face the limitations of fixed fusion loss terms. Task-specialized networks [76] create fusion-task dependencies, restricting flexibility and limiting their applicability to various high-level vision tasks. We address these limitations through a task-driven framework with learnable loss. The fusion loss contains learnable parameters, generated by a loss generation module, and is designed to retain the intensity information of the source images for specific downstream tasks. The purpose of updating the fusion loss is to guide the fusion network in generating fused images that minimize downstream task loss," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 713 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7457" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "thereby enhancing adaptability. Moreover, the fusion loss update relies on the downstream task loss, making it independent of any specific task or network architecture." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 111, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 111, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 111, + 295, + 397 + ], + "type": "text", + "content": "The loss generation module produces fusion losses for subsequent fusion module updates. This complexity presents challenges to standard end-to-end training. Fortunately, meta-learning techniques, which are strategies for learning how to learn, can effectively achieve the learning objectives of the loss generation module. This involves minimizing task loss for fused images through an optimized fusion loss. Meta-learning tackles common deep-learning challenges like limited data, high computational costs, and the need for better generalization. Core optimization areas include parameter tuning [11], optimization strategies [26], and network architectures search [30]. In this paper, we draw inspiration from the Model-Agnostic Meta-Learning (MAML) approach [11] to train our loss generation module. Specifically, training the loss generation module involves two stages: inner updates and outer updates. During inner updates, the output of the loss generation module updates a surrogate fusion module without altering the original parameters. During outer updates, the fused image from the surrogate module is fed into the task network. The resulting task loss then updates the loss generation module through backpropagation. This alternating training ensures that the loss generation module consistently produces fusion losses that minimize the downstream task loss of the fused images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 400, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 628 + ], + "type": "text", + "content": "This paper introduces TDFusion, a task-oriented fusion framework driven by downstream tasks. It consists of a fusion module, a task module, and a loss generation module that learns to optimize the fusion loss. The fusion loss incorporates intensity preferences from source images and gradient preservation, guided by downstream task loss to refine intensity preferences. This model follows the general form of fusion loss used in advanced methods [76, 83, 88, 89], ensuring adaptability to various tasks. The updates of the loss generation module are performed using a meta-learning approach, optimizing the loss generation module parameters based on task loss from the fused images after each update of the fusion module. This process ensures that the loss function guides the fusion module continuously, optimizing feature aggregation and minimizing downstream task losses. The loss generation module dynamically adjusts through alternating updates with the fusion and task modules, generating optimal fusion losses at each model state. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "- We propose TDFusion, a meta-learning-based fusion framework that leverages the loss functions of downstream tasks for training. This method promotes task-driven fusion and alleviates challenges caused by the absence of ground truth. Moreover, this framework is agnostic of specific downstream tasks or network architectures, which enhances its adaptability and flexibility." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 553, + 216 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 132 + ], + "type": "text", + "content": "- Our framework includes a dynamically updated, learnable fusion loss generation module. It selectively extracts source image information, minimizing the loss in downstream tasks. This ensures optimal fusion performance while maximizing adaptability to downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "content": "- We analyze the information preferences of downstream tasks such as semantic segmentation and object detection, providing deeper insights into multi-modal high-level vision tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 180, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 180, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 553, + 216 + ], + "type": "text", + "content": "- TDFusion achieves outstanding performance in both fusion and high-level vision tasks, validated on four fusion datasets with semantic segmentation and object detection." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 228, + 400, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 400, + 241 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 400, + 241 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 249, + 499, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 249, + 499, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 249, + 499, + 262 + ], + "type": "text", + "content": "2.1. Deep learning-based Image Fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 267, + 553, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 553, + 542 + ], + "type": "text", + "content": "Deep learning-based image fusion methods have revolutionized the field by exploiting the powerful feature extraction capabilities of neural networks [4, 6, 33, 35, 36, 42-44, 65, 77]. These methods are broadly categorized into discriminative and generative approaches. Discriminative methods [10, 81, 85, 88] leverage the strong reconstruction ability of neural networks to directly learn the mapping between source and fused images [21, 23, 80, 82]. Generative methods, on the other hand, model the image generation process using generative approaches, integrating source images from a distributional perspective. These include methods based on Generative Adversarial Networks [46-48, 87] and diffusion models [73, 84]. Unified fusion methods [66, 75, 89] bridge the gap between different fusion sub-tasks, incorporating strategies such as continual learning [67] and self-supervised decomposition techniques [29]. The introduction of registration modules helps to mitigate misalignment issues [16, 60, 68] in source images. Recent studies have further explored the synergy between fusion and high-level vision tasks. These include leveraging downstream task losses to optimize fusion networks [32, 56], embedding high-level task features [34, 76, 78], and employing initialization techniques [40]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 552, + 453, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 453, + 565 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 453, + 565 + ], + "type": "text", + "content": "2.2. Meta-Learning in Vision" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "Meta-learning develops algorithms to automatically fine-tune hyperparameters for specific tasks, showing its versatility and effectiveness across various domains. MAML [11] and its variations [12, 50, 53] focus on learning efficient initialization parameters to quickly adapt to new tasks using minimal data. Meta-SGD [26] extends MAML by learning optimal update directions and rates, beneficial in few-shot learning scenarios. Other approaches like MW-Net [55] and L2RW [54] emphasize selecting relevant sample weights to tackle noisy data using a compact validation set. Additionally, some studies focus on improving model adaptability through learning loss functions [2, 7, 15]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7458" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 85, + 553, + 305 + ], + "blocks": [ + { + "bbox": [ + 99, + 71, + 241, + 83 + ], + "lines": [ + { + "bbox": [ + 99, + 71, + 241, + 83 + ], + "spans": [ + { + "bbox": [ + 99, + 71, + 241, + 83 + ], + "type": "text", + "content": "Learning of Loss Generate Module" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 397, + 71, + 477, + 83 + ], + "lines": [ + { + "bbox": [ + 397, + 71, + 477, + 83 + ], + "spans": [ + { + "bbox": [ + 397, + 71, + 477, + 83 + ], + "type": "text", + "content": "Learning of Fusion" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 85, + 553, + 305 + ], + "lines": [ + { + "bbox": [ + 60, + 85, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 60, + 85, + 553, + 305 + ], + "type": "image", + "image_path": "ff889df2ea5748d77aca49d1752dbca2103ade44bfc297f2bb885d9f67cfe3cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 332 + ], + "type": "text", + "content": "Figure 1. The TDFusion workflow alternates between training the loss generation module and the fusion module. Training of the loss generation module involves both inner and outer updates, learned through meta-learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 337, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 337, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 296, + 445 + ], + "type": "text", + "content": "In image fusion, learnable filters [22] enable the fusion of images at arbitrary resolutions. MetaFusion [78] introduces a mechanism that improves image fusion and object detection by aligning semantics with fusion-specific features. ReFusion [5] guides the learnable fusion loss for various fusion tasks by reconstructing the source images. Metalearning also supports neural architecture search [38, 40] to identify optimal network architectures for image fusion and customizes network initialization for various tasks [40]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 454, + 258, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 258, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 258, + 467 + ], + "type": "text", + "content": "2.3. Comparison with Existing Approaches" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 627 + ], + "type": "text", + "content": "We propose a novel image fusion method tailored for downstream tasks, leveraging a learnable fusion loss driven by task-specific objectives. Our method employs a meta-learning algorithm that alternates between inner and outer updates, allowing the downstream task loss to guide the optimization of learnable fusion parameters. This results in fused images that minimize the downstream task loss, enhancing their adaptability across various tasks. Unlike previous methods, our approach develops a task-specific fusion loss, shifting focus from traditional factors such as resolution and network structure, and avoiding reliance on predefined fusion loss terms. This renders our fusion framework more flexible and applicable to various scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 639, + 111, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 111, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 111, + 651 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 123, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 123, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 123, + 671 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Our TDFusion framework, as shown in Fig. 1, consists of a fusion network " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", a downstream task network " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " and a fusion loss generation module " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", which produces" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": "parameters for a learnable loss function. The parameters of these modules are denoted as " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": ", respectively. The one-step updated clones of " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": " are denoted as " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": ", with parameters " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}'}" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{T}'}" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": ". During the updates, the fusion network and the loss generation module alternate in learning, as depicted by blue and purple in Fig. 1. The update of the loss generation module consists of inner and outer updates, detailed in the following subsections. " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 313, + 337, + 555, + 456 + ], + "type": "text", + "content": " represent the learnable fusion loss and task-specific loss, with their formulations provided in the next section." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 465, + 404, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 404, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 404, + 477 + ], + "type": "text", + "content": "3.2. Loss Function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "text", + "content": "The learnable fusion loss " + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "text", + "content": " consists of the intensity term and the gradient term. The intensity term is defined by the output of the loss generation module " + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\{w_a, w_b\\} = \\mathcal{G}(I_a, I_b)" + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "inline_equation", + "content": "w_a" + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "inline_equation", + "content": "w_b" + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "text", + "content": " control the intensity preference in the fusion loss. The Softmax function in the loss generation module ensures " + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "inline_equation", + "content": "w_a^{ij} + w_b^{ij} = 1" + }, + { + "bbox": [ + 313, + 482, + 555, + 613 + ], + "type": "text", + "content": " for each pixel, thereby selectively retaining the intensity information from the source images. The gradient term emphasizes higher gradient values from the input images [76, 83, 89], aiming to preserve maximal information from the source. The detailed formulation of the learnable fusion loss is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 622, + 555, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 622, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 341, + 622, + 555, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = \\mathcal {L} _ {f} ^ {\\text {i n t}} + \\alpha \\mathcal {L} _ {f} ^ {\\text {g r a d}}, \\tag {1}", + "image_path": "5cad2e7d3c7771d339e48328bbe877c64640f42b1ecfcbf232a40ea841b2d6b1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 640, + 553, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 640, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 335, + 640, + 553, + 679 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} ^ {i n t} = \\frac {1}{H W} \\sum_ {i j} \\left[ \\sum_ {k \\in \\{a, b \\}} w _ {k} ^ {i j} \\left(I _ {f} ^ {i j} - I _ {k} ^ {i j}\\right) ^ {2} \\right], \\tag {2}", + "image_path": "c5c17bed475b5098f57d2b4633fe2aab9b69a183a147cf356079fdd28c44e867.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 330, + 681, + 553, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 681, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 330, + 681, + 553, + 710 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} ^ {\\text {g r a d}} = \\frac {1}{H W} \\sum_ {i j} \\left| \\nabla I _ {f} ^ {i j} - \\max _ {k \\in \\{a, b \\}} \\left(\\nabla I _ {k} ^ {i j}\\right) \\right|, \\tag {3}", + "image_path": "d06c59270f6609f6f303fe1891f680b2a66b8aaf6d3f0db4ca0efce311d4a4c5.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7459" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": " denotes the Sobel operator, commonly employed for gradient extraction in image fusion [32, 57, 83]. The parameter " + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": " serves as a scaling factor, while " + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f^{int}" + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f^{grad}" + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": " represent the intensity loss and gradient loss, respectively. The weights " + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\{w_a, w_b\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": " control the emphasis of the loss function on the intensity information from each source image. These parameters enable the fusion process to selectively aggregate and incorporate relevant information from the source images. Variations in " + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": " lead to different configurations of " + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "inline_equation", + "content": "w_a" + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "inline_equation", + "content": "w_b" + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": ", thereby influencing the characteristics of the fusion loss. " + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": " undergoes updates driven by the high-level task loss during the in-step update process, as detailed in Sec. 3.4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 277 + ], + "type": "text", + "content": "The loss function " + }, + { + "bbox": [ + 55, + 228, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 55, + 228, + 296, + 277 + ], + "type": "text", + "content": " depends on the specific task. In this study, we adopt SegFormer [9] and YOLOv8 [19] for the downstream tasks, employing cross-entropy loss and YOLO loss [19] for each task, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 285, + 174, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 174, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 174, + 298 + ], + "type": "text", + "content": "3.3. Dataset Partitioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "text", + "content": "In order to enhance the effectiveness of the loss generation module " + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "text", + "content": " in guiding the fusion tasks, we create non-overlapping subsets of size " + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "text", + "content": " at each training epoch. The meta-training set " + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\{I_a^{mtr}, I_b^{mtr}\\}" + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "text", + "content": " and the meta-test set " + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\{I_a^{mts}, I_b^{mts}\\}" + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "text", + "content": " are randomly drawn from the fusion training set " + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\{I_a^{ftr}, I_b^{ftr}\\}" + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "text", + "content": ". These subsets are fed into the model sequentially during the training process of the loss generation module, covering both the inner and outer updates. The entire fusion training set " + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\{I_a^{ftr}, I_b^{ftr}\\}" + }, + { + "bbox": [ + 55, + 303, + 296, + 425 + ], + "type": "text", + "content": " is fully utilized during the training of the fusion network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 244, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 244, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 244, + 445 + ], + "type": "text", + "content": "3.4. Learning of loss generation module" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 451, + 296, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 668 + ], + "type": "text", + "content": "Fig. 1 illustrates that the loss generation and fusion modules are trained alternately, ensuring the fusion loss is optimized at various stages of training and under different states of the fusion network. The process of training to optimize the fusion loss involves two key steps: the inner update and the outer update. In the inner update, clones of both the fusion network and the task network are generated, with each network undergoing a single training iteration using the fusion loss and task loss, respectively. This process is designed to obtain the state of the network guided by the fusion loss. During the outer update, the task loss of the fusion image produced by the updated clone is calculated. This loss is then backpropagated to the loss generation network. The goal of this step is to direct the fusion network to generate fusion images that result in lower downstream task loss, once guided by the fusion loss. The alternating updates between the inner and outer steps constitute the learning procedure for the loss generation module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 673, + 141, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 141, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 141, + 685 + ], + "type": "text", + "content": "3.4.1. Inner Update" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "During the inner update phase, the fusion network " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " undergoes a single update guided by the fusion loss, which" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "depends on the current state of network " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": ". This update primarily aims to compute the intermediate parameters " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}'}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{R}'}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": ", which are crucial for updating " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": " in the subsequent phase. The upper section of the purple region in Fig. 1 illustrates this process. During this phase, the images from the meta-training set " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\{I_a^{mtr}, I_b^{mtr}\\}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": " are fed into the model:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 148, + 555, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 148, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 347, + 148, + 555, + 174 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathcal {F} ^ {\\prime}} = \\theta_ {\\mathcal {F}} - \\eta_ {\\mathcal {F} ^ {\\prime}} \\frac {\\partial \\mathcal {L} _ {f} \\left(I _ {a} ^ {m t r} , I _ {b} ^ {m t r} , I _ {f} ^ {m t r} ; \\theta_ {\\mathcal {G}}\\right)}{\\partial \\theta_ {\\mathcal {F}}}, \\tag {4}", + "image_path": "7f7a8fd8b1f49ea8fc4ea4d06182e7d19dd36f88557fa0effefc2ebb4803f9a3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": " undergoes a single gradient descent update. " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": " represent the parameters of the loss generation module " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": ", which determine the parameters for the learnable fusion loss. And the notation " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathcal{F}'}" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": " refers to the step size. The module " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": " temporarily substitutes " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": ", adjusting its parameters in one update step. Meanwhile, the parameters of " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": ", remain unchanged. Similarly, " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": " is updated in a single step using the parameters " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": " from the current task network " + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 176, + 555, + 283 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 286, + 555, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 286, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 376, + 286, + 555, + 311 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathcal {T} ^ {\\prime}} = \\theta_ {\\mathcal {T}} - \\eta_ {\\mathcal {T} ^ {\\prime}} \\frac {\\partial \\mathcal {L} _ {t} \\left(I _ {f} ^ {m t r}\\right)}{\\partial \\theta_ {\\mathcal {T}}}. \\tag {5}", + "image_path": "89b5d176d8476bc217fc6361978405983d8a9e0fb832b9b4b72e5d313f17b2cd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "text", + "content": "The parameters " + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}'}" + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{T}'}" + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "text", + "content": " are both updated during the inner update, which also ensures that the computation graph of " + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}'}" + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "text", + "content": " is preserved. This preserved graph is essential for optimizing " + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 314, + 554, + 363 + ], + "type": "text", + "content": " during the outer update." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 367, + 402, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 402, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 402, + 379 + ], + "type": "text", + "content": "3.4.2. Outer Update" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": "The primary objective of the outer update is to evaluate and refine the fusion guidance capability of " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": ", specifically by strengthening the influence of the loss function " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " in steering the fusion module " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": ". In the framework diagram, this stage is represented in the lower part of the purple region. The modules " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": ", derived from the inner update, represent the current guidance capacity of " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": ". In an ideal scenario, the optimal fusion loss should enhance the performance of the downstream task on the fused image. In this stage, the meta-test set " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\{I_a^{mts}, I_b^{mts}\\}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " is employed. The parameters " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " are subsequently updated using the task loss " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": ", which is computed by " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 381, + 529, + 554, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 529, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 381, + 529, + 554, + 554 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathcal {G}} = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\frac {\\partial \\mathcal {L} _ {t} \\left(I _ {f} ^ {m t s}\\right)}{\\partial \\theta_ {\\mathcal {G}}}, \\tag {6}", + "image_path": "0678d0e580f60f7a6051d3824d8d4aa65bebb9f761ef6ac48e15c26a556abbd3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "type": "inline_equation", + "content": "I_{f}^{mts} = \\mathcal{F}'(I_{a}^{mts}, I_{b}^{mts})" + }, + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "type": "text", + "content": ", and the gradient " + }, + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{L}_t / \\partial \\theta_G" + }, + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "type": "text", + "content": " can be calculated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 585, + 555, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 585, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 324, + 585, + 555, + 613 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L} _ {t}}{\\partial \\theta_ {\\mathcal {G}}} = \\frac {\\partial \\mathcal {L} _ {t}}{\\partial \\theta_ {\\mathcal {F} ^ {\\prime}}} * \\left(- \\eta_ {\\mathcal {F} ^ {\\prime}} \\frac {\\partial^ {2} \\mathcal {L} _ {f} \\left(I _ {a} ^ {m t r} , I _ {b} ^ {m t r} , I _ {f} ^ {m t r} ; \\theta_ {\\mathcal {G}}\\right)}{\\partial \\theta_ {\\mathcal {F}} \\partial \\theta_ {\\mathcal {G}}}\\right). \\tag {7}", + "image_path": "06bcfca395457befbd3df11d3ee31c32e5dc291145ff3779be7068ab368ebe27.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Eq. (6) holds because the task loss " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " is determined by " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "I_{f}^{mts}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": ", which in turn relies on " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}'}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": ". The optimization of " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " is realized by preserving the computational relationship between " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}'}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " throughout the inner update. The updated " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " module gains the ability to generate enhanced fusion loss functions, enabling the fusion module to integrate relevant information from the source images more efficiently into the fused output for downstream tasks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7460" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 58, + 85, + 298, + 320 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 234, + 83 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 234, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 234, + 83 + ], + "type": "text", + "content": "Algorithm 1 TDFusion Training Algorithm" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": "Require: Training set " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\{I_a^{ftr}, I_b^{ftr}\\}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": ". Output: Thoroughly trained " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}}, \\theta_{\\mathcal{T}}, \\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": ". \n1: Initialize " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}}, \\theta_{\\mathcal{T}}, \\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": ". \n2: for epoch = 1 to L do \n3: Sample " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\{I_a^{mtr}, I_b^{mtr}\\}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\{I_a^{mts}, I_b^{mts}\\}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": ". \n4: for step = 1 to M do \n % Inner update: apply G. \n5: Sample " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "(I_a^{mtr}, I_b^{mtr})" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " and get " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "(I_f^{mtr}, y^{mtr})" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": ". \n6: Compute " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}'}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{T}'}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " by Eq. (4) and Eq. (5). \n% Outer update: optimize G. \n7: Sample " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "(I_a^{mts}, I_b^{mts})" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " and get " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "(I_f^{mts}, y^{mts})" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": ". \n8: Update " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " by Eq. (6). \n9: end for \n10: for step = 1 to N do \n % Fusion update: optimize F and T. \n11: Sample " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "(I_a^{ftr}, I_b^{ftr})" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " and get " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "(I_f^{ftr}, y^{ftr})" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": ". \n12: Update " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{T}}" + }, + { + "bbox": [ + 58, + 85, + 298, + 320 + ], + "type": "text", + "content": " by Eq. (8) and Eq. (9). \n13: end for \n14: end for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 336, + 211, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 211, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 211, + 350 + ], + "type": "text", + "content": "3.5. Learning of Fusion Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": "The alternating inner and outer update iterations form a flexible and effective mechanism to refine " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " in response to the evolving state of " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": ". After refining " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": ", it is utilized to further improve the training of " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": ". This stage, denoted in blue in the diagrams, involves processing the fusion training set images " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\{I_a^{ftr}, I_b^{ftr}\\}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " are updated through the application of the fusion loss " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " and task loss " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 442, + 295, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 442, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 93, + 442, + 295, + 472 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathcal {F}} = \\theta_ {\\mathcal {F}} - \\eta_ {\\mathcal {F}} \\frac {\\partial \\mathcal {L} _ {f} \\left(I _ {a} ^ {f t r} , I _ {b} ^ {f t r} , I _ {f} ^ {f t r} ; \\theta_ {\\mathcal {G}}\\right)}{\\partial \\theta_ {\\mathcal {F}}}, \\tag {8}", + "image_path": "a95369b04e7a4f6fd737e26faef85ca6471104f1c2c4582b7ec25a1ae89b2b77.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 481, + 294, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 481, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 121, + 481, + 294, + 510 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathcal {T}} = \\theta_ {\\mathcal {T}} - \\eta_ {\\mathcal {T}} \\frac {\\partial \\mathcal {L} _ {t} \\left(I _ {f} ^ {f t r}\\right)}{\\partial \\theta_ {\\mathcal {T}}}. \\tag {9}", + "image_path": "62be7dda20b9d427f5685afb8d996b087c3837708dd887a6f91f36d31b16b68a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 511, + 296, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 642 + ], + "type": "text", + "content": "After completing several training sessions on the fusion network, the focus then shifts back to the learning phase of the fusion generation module. The fusion framework evolves through a series of alternating phases, with each phase finetuning the fusion loss based on the fusion network's current state. Such alternating phases ensure that the fusion network consistently applies the most suitable fusion loss during its progression. Ultimately, this results in the development of a highly efficient fusion network, optimized for peak performance. The complete training procedure is detailed in Algorithm 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 648, + 181, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 181, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 181, + 659 + ], + "type": "text", + "content": "3.6. Network Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "TDFusion is composed of three modules: the fusion network, the downstream task network, and the loss generation network. The fusion network shares the same architecture as [5], a lightweight model built upon the Restormer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "Block (RTB) [74]. An adaptive fusion module is incorporated into this network to facilitate feature integration. Fig. 1 illustrates the structure of the loss generation module. It also employs the Restormer Block (RTB) [74] as its primary component, receiving inputs " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\{I_a, I_b\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ". After applying Softmax(" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "), the final output guarantees " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "w_a^{ij} + w_b^{ij} = 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " for each pixel. This design ensures that the fused image meets the similarity constraints and removes the reliance on initialization within the loss generation module. The architecture of the downstream task network " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " depends on the specific task. For learning of loss generation module, we chose the most lightweight models of SegFormer [9] and YOLOv8 [19] for semantic segmentation and object detection, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 433, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 433, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 433, + 258 + ], + "type": "text", + "content": "3.7. Theoretical Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": "To better understand the weighting mechanism of the loss generation module " + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": ", we investigate the optimization procedure of " + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": ", which generates the weights " + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\{w_{a}, w_{b}\\}" + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": ". For clarity, we rewrite Eq. (2) as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 332, + 318, + 553, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 318, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 332, + 318, + 553, + 418 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {f} ^ {i n t} = \\left[ w _ {a} \\odot \\left(I _ {a} - I _ {f}\\right) \\odot \\left(I _ {a} - I _ {f}\\right) \\right. \\\\ \\left. + w _ {b} \\odot \\left(I _ {b} - I _ {f}\\right) \\odot \\left(I _ {b} - I _ {f}\\right) \\right] \\times \\frac {1}{H W} \\\\ = \\left[ \\mathcal {G} \\left(I _ {a}, I _ {b}; \\theta_ {\\mathcal {G}}\\right) \\odot \\left(I _ {a} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\right. \\tag {10} \\\\ \\odot \\left(I _ {a} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\\\ + \\left(1 - \\mathcal {G} \\left(I _ {a}, I _ {b}; \\theta_ {\\mathcal {G}}\\right)\\right) \\odot \\left(I _ {b} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\\\ \\left. \\odot \\left(I _ {b} - \\mathcal {F} _ {\\theta_ {\\mathcal {F}}} \\left(I _ {a}, I _ {b}\\right)\\right) \\right] \\times \\frac {1}{H W}. \\\\ \\end{array}", + "image_path": "f346a49fc07600a227c89e8284247099b1af7f5c999fa01c9009cc8c68c4b8d6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "inline_equation", + "content": "w_{a}, w_{b} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "inline_equation", + "content": "I_{a}, I_{b} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "text", + "content": " denotes the element-wise multiplication operation. Let " + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\Omega'" + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "text", + "content": " be the set " + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\{\\theta_{\\mathcal{F}'}', \\theta_{\\mathcal{T}''}\\}" + }, + { + "bbox": [ + 313, + 423, + 555, + 461 + ], + "type": "text", + "content": ", this leads to the following expression:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 467, + 553, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 467, + 553, + 660 + ], + "spans": [ + { + "bbox": [ + 331, + 467, + 553, + 660 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta_ {\\mathcal {G}} = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\frac {\\partial \\mathcal {L} _ {t} ^ {m t s} (\\Omega^ {'} (\\theta_ {\\mathcal {G}}))}{\\partial \\theta_ {\\mathcal {G}}} \\\\ = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\frac {\\partial \\mathcal {L} _ {t} ^ {m t s} (\\Omega^ {'} (\\theta_ {\\mathcal {G}}))}{\\partial \\Omega^ {'}} \\frac {\\partial \\Omega^ {'} (\\theta_ {\\mathcal {G}})}{\\partial \\theta_ {\\mathcal {G}}} \\\\ = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\eta_ {\\mathcal {F} ^ {\\prime}} \\underbrace {\\frac {\\partial \\mathcal {L} _ {t} ^ {m t s} \\left(\\Omega^ {'} \\left(\\theta_ {\\mathcal {G}}\\right)\\right)}{\\partial \\Omega^ {\\prime}}} _ {\\text {(a)}} \\times \\frac {\\partial \\mathcal {G} \\left(I _ {a} , I _ {b} ; \\theta_ {\\mathcal {G}}\\right)}{\\theta_ {\\mathcal {G}}} \\\\ \\times \\underbrace {\\left[ \\left(I _ {a} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) \\odot \\left(I _ {a} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) \\right.} _ {(b)} \\tag {11} \\\\ \\underbrace {- \\left(I _ {b} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) \\odot \\left(I _ {b} - \\frac {\\partial \\mathcal {F} _ {\\theta_ {\\mathcal {F}}}}{\\partial \\theta_ {\\mathcal {F}}}\\right) ]} _ {(b)} \\\\ = \\theta_ {\\mathcal {G}} - \\eta_ {\\mathcal {G}} \\eta_ {\\mathcal {F} ^ {\\prime}} \\mathbf {G} \\times \\frac {\\partial \\mathcal {G} \\left(I _ {a} , I _ {b} ; \\theta_ {\\mathcal {G}}\\right)}{\\theta_ {\\mathcal {G}}}. \\\\ \\end{array}", + "image_path": "9a635aff0e5c3c6dbbfc5cc9a44b2d7bf1cb9371ee67f62439b3017dca6ecbe0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " denotes the inner product between two gradients: (a) the first one is derived from the task loss using a meta-testing set, and (b) the second is calculated from the fusion loss based on a meta-training set. Consequently, the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7461" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 330 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 330 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 330 + ], + "type": "image", + "image_path": "65a1c92046391bc694d653a9a75e71f8a8ba94d6cfdba5037aa2a9bc3eb03716.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 353 + ], + "type": "text", + "content": "Figure 2. Visual comparison of fusion results. The cases are \"01258N\" in MSRS dataset, \"00122\" in FMB dataset, \"00449\" in M3FD dataset and \"200304\" in LLVIP dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 360, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 396 + ], + "type": "text", + "content": "optimization of the module " + }, + { + "bbox": [ + 55, + 360, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 55, + 360, + 295, + 396 + ], + "type": "text", + "content": " is driven by the task loss, with the objective of preserving task-specific information throughout the fusion process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 409, + 132, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 132, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 132, + 423 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 430, + 105, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 105, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 105, + 443 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": "Experimental Setup. In our experiments, the epoch number " + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": " and the training iterations of the loss generation module " + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": " are set to 50 and 200. The learning iterations for the fusion network " + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": " depend on the size of the dataset. We use the Adam optimizer with a learning rate of 1e-4, a batch size of 2, and a hyperparameter " + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": " set to 1. All experiments are conducted on a PC with a single NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": "Evaluation Metrics and Comparison Methods. The advanced fusion methods compared in our study include TarDAL [32], SegMIF [34], MURF [69], EMMA [86], DCINN [61], MRFS [76], and TIMFusion [40]. The fusion performance is evaluated using metrics including entropy (EN), spatial frequency (SF), sum of correlation differences (SCD), visual information fidelity (VIF), " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "Q^{AB / F}" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": " and structural similarity index metric (SSIM)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Dataset Split. We use four datasets annotated for downstream tasks, including MSRS [58], FMB [34], M3FD [32], and LLVIP [17]. MSRS contains 1083/361 image pairs for training/test, and FMB contains 1220/280 pairs for training/test. We follow the splits of the original papers for both datasets. M3FD dataset consists of 4200 pairs of images for detection, with 300 pairs designated for fusion evalu-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "text", + "content": "ation. The 4200 detection images are split into 3150 for training and 1050 for testing, ensuring that the 300 pairs for fusion evaluation are included within the detection test set. These 300 fusion images are then employed to assess the fusion performance. The original LLVIP dataset contains 12025/3463 image pairs as training/test set. Due to its large size, we select every 10th image to form our training and test sets, resulting in 1203/347 image pairs for training and testing. Our splits for M3FD and LLVIP will be available." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 479, + 432, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 432, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 432, + 491 + ], + "type": "text", + "content": "4.2. Fusion Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "Fig. 2 presents a visual comparison of different methods. The fused images generated by TDFusion excel in preserving discriminative details, achieving balanced brightness, and maintaining clear object contours. It effectively preserves the target features from the infrared images and the background details from the visible images, resulting in fused images that are more natural and exhibit greater clarity across different environments. These results highlight the advantages of TDFusion in detail preservation and visual performance. More results are available in the supplementary material. Tab. 1 presents the quantitative comparison of fusion over four datasets. TDFusion outperforms other methods across most metrics. This suggests that TDFusion not only enhances image details but also provides consistent fusion results across various scenarios. Compared to other methods, TDFusion shows superior adaptability and robustness, particularly in handling diverse image characteristics and challenging fusion tasks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7462" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 87, + 555, + 296 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 553, + 84 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 553, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 553, + 84 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of Infrared-visible image fusion. The red and blue markers represent the best and second-best values." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 87, + 555, + 296 + ], + "lines": [ + { + "bbox": [ + 58, + 87, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 58, + 87, + 555, + 296 + ], + "type": "table", + "html": "
Infrared-visible Image Fusion on MSRS [58] DatasetInfrared-visible Image Fusion on FMB [34] Dataset
EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑
TarDAL [32]5.285.980.710.210.180.47TarDAL [32]6.636.941.030.280.290.74
SegMIF [34]5.9511.101.570.440.630.55SegMIF [34]6.8313.691.720.390.650.60
MURF [69]5.0410.491.020.220.370.60MURF [69]6.3713.881.340.220.370.68
EMMA [86]6.7311.561.620.490.640.70EMMA [86]6.7715.001.500.420.650.72
DCINN [61]6.0010.511.490.410.570.52DCINN [61]6.4711.471.390.380.590.74
MRFS [76]7.008.861.420.370.490.55MRFS [76]6.7812.421.240.380.620.73
TIMFusion [40]6.279.671.340.320.480.68TIMFusion [40]6.5112.231.240.350.590.73
TDFusion (Ours)6.7411.301.860.500.670.70TDFusion (Ours)6.8614.161.760.430.680.75
Infrared-visible Image Fusion on M3FD [32] DatasetInfrared-visible Image Fusion on LLVIP [17] Dataset
EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑EN ↑SF ↑SCD ↑VIF ↑\\(Q^{AB/F}\\) ↑SSIM ↑
TarDAL [32]6.877.631.290.270.300.71TarDAL [32]6.327.421.040.270.220.58
SegMIF [34]6.8514.141.720.370.600.59SegMIF [34]6.6815.461.380.400.660.57
MURF [69]6.5012.551.460.210.320.64MURF [69]6.1315.080.960.210.310.57
EMMA [86]6.9215.231.490.380.590.69EMMA [86]7.3515.371.570.410.640.66
DCINN [61]6.5911.211.460.340.510.72DCINN [61]6.9813.341.430.380.520.64
MRFS [76]6.9412.071.260.340.550.70MRFS [76]6.8311.041.230.310.420.64
TIMFusion [40]6.7512.311.370.350.530.70TIMFusion [40]6.5813.521.140.330.460.64
TDFusion (Ours)6.9914.491.830.410.650.72TDFusion (Ours)7.3616.381.750.460.700.67
", + "image_path": "2962883789f32476e8578fb3a0036fd2d65c0d14d779d891fbce45181a537779.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 307, + 553, + 427 + ], + "blocks": [ + { + "bbox": [ + 58, + 307, + 553, + 427 + ], + "lines": [ + { + "bbox": [ + 58, + 307, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 58, + 307, + 553, + 427 + ], + "type": "image", + "image_path": "c5baf043834cf4bf3ae738fc93d4d04004447215b3caaee616a34c56078459c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 429, + 538, + 439 + ], + "lines": [ + { + "bbox": [ + 70, + 429, + 538, + 439 + ], + "spans": [ + { + "bbox": [ + 70, + 429, + 538, + 439 + ], + "type": "text", + "content": "Figure 3. Visual comparison for Semantic Segmentation. The cases are \"00726N\" in MSRS dataset and \"01438\" in FMB dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 200, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 200, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 200, + 459 + ], + "type": "text", + "content": "4.3. Downstream Applications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 468, + 295, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 468, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 54, + 468, + 295, + 660 + ], + "type": "text", + "content": "This section validates the adaptability of fusion methods to downstream tasks. For a fair comparison, we adopt SegFormer [9] and YOLOv8 [19] as backbones and retrain the task networks for each fusion method over 300 epochs to evaluate their adaptability to semantic segmentation and object detection. Fig. 3 and Fig. 4 present the visual comparisons of semantic segmentation and object detection, respectively. TDFusion outperforms in image detail retention, edge clarity, and object recognition, effectively identifying and segmenting objects. For semantic segmentation, the generated maps clearly distinguish different class regions, closely matching the ground truth. In object detection, the fused images exhibit more precise boundary localization for salient objects. This indicates that TDFusion maintains a better balance between fine details and overall context. More results can be found in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Tab. 2 shows the performance comparison across different methods in semantic segmentation and object detection. TDFusion outperforms other methods on most metrics, particularly in mIoU and mAP. This demonstrates that TDFu" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 469, + 555, + 590 + ], + "blocks": [ + { + "bbox": [ + 313, + 445, + 553, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 445, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 553, + 468 + ], + "type": "text", + "content": "Table 2. Performance comparison of downstream applications. The red and blue markers represent the best and second-best." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 469, + 555, + 590 + ], + "lines": [ + { + "bbox": [ + 315, + 469, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 315, + 469, + 555, + 590 + ], + "type": "table", + "html": "
MethodsSemantic SegmentationObject Detection
MSRSFMBM3FDLLVIP
mAccmIoUmAccmIoUmAP50mAP75AP50AP75
Infrared83.2369.4958.8551.9879.1253.0596.0372.07
Visible83.4473.7665.1257.9682.2154.8291.7848.66
TarDAL81.9371.3562.8655.3383.1656.3993.7962.71
SegMIF85.7374.2565.9758.4183.6158.2393.9566.45
MURF85.0374.0864.1056.9680.5854.2294.2468.04
EMMA85.9974.4862.4556.2883.7156.9194.0066.21
DCINN84.1174.3561.0954.8182.6957.3794.9268.34
MRFS84.7674.5061.9355.7183.2857.7493.0367.21
TIMFusion83.6773.5863.7057.2483.2256.0893.7661.33
TDFusion86.0475.0967.1760.5086.2759.7195.0069.18
", + "image_path": "2125701b7d516e5d4c63c9b132bb8e32698e56df81bc8751e98f5cf37e4999ac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 593, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 651 + ], + "type": "text", + "content": "sion effectively enhances the quality of fused images. It also improves the downstream task performance, especially in terms of accuracy and robustness in complex scenarios. Class-wise results are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 468, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 468, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 468, + 672 + ], + "type": "text", + "content": "4.4. Task-driven Learnable Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Our framework incorporates a learnable fusion loss that models the preferences of downstream tasks for information from source images. The models trained on FMB dataset" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7463" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 191 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 191 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 191 + ], + "type": "image", + "image_path": "5dd3227b95ffe4277b7733b909d3963d469aa7be65f04d769127fe7da6e3408e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 194, + 527, + 204 + ], + "lines": [ + { + "bbox": [ + 81, + 194, + 527, + 204 + ], + "spans": [ + { + "bbox": [ + 81, + 194, + 527, + 204 + ], + "type": "text", + "content": "Figure 4. Visual comparison for Object Detection. The cases are \"02236\" in M3FD dataset and \"210145\" in LLVIP dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 209, + 175, + 331 + ], + "blocks": [ + { + "bbox": [ + 57, + 209, + 175, + 331 + ], + "lines": [ + { + "bbox": [ + 57, + 209, + 175, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 175, + 331 + ], + "type": "image", + "image_path": "cc24fe26ef625184c016f17bfbe192a79954665885b86321e4b8a355c403d80c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 333, + 288, + 343 + ], + "lines": [ + { + "bbox": [ + 61, + 333, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 333, + 288, + 343 + ], + "type": "text", + "content": "Figure 5. Visualisation of learnable loss for downstream tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 209, + 294, + 331 + ], + "blocks": [ + { + "bbox": [ + 178, + 209, + 294, + 331 + ], + "lines": [ + { + "bbox": [ + 178, + 209, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 178, + 209, + 294, + 331 + ], + "type": "image", + "image_path": "d562ac6c344a11d3f1bdb8b4f9901cc2cfb66ba113513b9c6660c1871c99d2bd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "text", + "content": "and LLVIP dataset labeled as " + }, + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "inline_equation", + "content": "SS" + }, + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "inline_equation", + "content": "OD" + }, + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "text", + "content": ", are evaluated on MSRS dataset to simulate performance in unknown scenes, as shown in Fig. 5. The results demonstrate that the fusion model adaptively selects information from infrared and visible images to satisfy task requirements. In semantic segmentation, the model combines scene structure and texture, prioritizing boundaries. This improves segmentation under varying lighting conditions. Fusion weights " + }, + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "inline_equation", + "content": "\\{w_{a}^{SS}, w_{b}^{SS}\\}" + }, + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "text", + "content": " indicate a preference for visible details and infrared advantages in low-light conditions. In object detection, the model focuses on edge and contrast information, especially for instances like pedestrians and vehicles. Higher fusion weights are assigned to bright regions in infrared images, enhancing target detection in low-light conditions. Fusion weights " + }, + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "inline_equation", + "content": "\\{w_{a}^{OD}, w_{b}^{OD}\\}" + }, + { + "bbox": [ + 55, + 351, + 296, + 591 + ], + "type": "text", + "content": " reflect this preference. Comparison of fusion losses across tasks reveals distinct differences, especially in highlighted regions, confirming that the model adapts to task-specific requirements by selecting the most relevant information from multimodal images. More results are provided in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 599, + 155, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 155, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 155, + 611 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "To thoroughly evaluate the performance of our proposed algorithm, we conduct a series of ablation experiments on FMB dataset, and the detailed results are shown in Tab. 3. In Exp. I, we exclude the learnable fusion loss by fixing " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "w_{a}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "w_{b}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ". In Exp. II, we omit the gradient loss from the loss function. In Exp. III, we also allow the fusion module parameters to be jointly optimized by both the task loss " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{t}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " and fusion loss " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{f}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ". In Exp. IV, we exclude the fusion" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 224, + 555, + 308 + ], + "blocks": [ + { + "bbox": [ + 315, + 209, + 553, + 221 + ], + "lines": [ + { + "bbox": [ + 315, + 209, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 553, + 221 + ], + "type": "text", + "content": "Table 3. Ablation experiment of fusion. The red denotes the best." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 224, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 317, + 224, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 224, + 555, + 308 + ], + "type": "table", + "html": "
Ablation Studies of fusion on FMB Dataset
ConfigurationsENSFSCDVIF\\(Q^{AB/F}\\)SSIM
Ifix \\(w_{a}\\) and \\(w_{b}\\) as 1/26.6013.731.580.390.600.72
IIw/o \\(\\mathcal{L}_{f}^{grad}\\)6.7711.651.630.370.640.73
III\\(\\theta_{\\mathcal{F}}\\) influenced by \\(\\mathcal{L}_{t}\\)6.8013.851.700.410.660.73
IVw/o Fusion learning6.8214.071.720.410.670.72
V\\(I_{f}=w_{a}*I_{a}+w_{b}*I_{b}\\)6.7511.491.650.380.620.73
Ours6.8614.161.760.430.680.75
", + "image_path": "65243a8d45bb361ca787183f5d3a3f8c462e1971257f7373dbf0b1989fe7c8a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "text", + "content": "module's dedicated learning phase and update it during the outer update of the loss module, which tests the impact of the fusion module's training schedule. In Exp. V, we replace our fusion method with a discriminative approach. The performance decline observed across different configurations confirms the rationality and effectiveness of our proposed method. Visualization and more analysis are provided in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 422, + 388, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 388, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 388, + 434 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 442, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 647 + ], + "type": "text", + "content": "To overcome the limitations of predefined fusion losses, which often fail to effectively guide the fusion process for downstream tasks, we propose a meta-learning-based framework for task-guided fusion. This framework includes a loss generation module that outputs the parameters of the learnable fusion loss. The module is updated using a meta-learning approach, which alternates between inner and outer loop steps to enhance its ability to guide the fusion network. Under varying fusion conditions, this module generates the optimal fusion loss for the downstream task. This enables the fusion network to produce fused images that minimize the task-specific loss. The theoretical analysis explains how the downstream task loss guides the fusion loss in our framework. Experiments using four publicly available fusion datasets and downstream tasks including semantic segmentation and object detection, demonstrate the effectiveness of our approach." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 657, + 411, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 411, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 411, + 670 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 712 + ], + "type": "text", + "content": "This work has been supported by the National Natural Science Foundation of China under Grant 12201497 and 12371512." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7464" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Hessah Albanwan, Rongjun Qin, and Yang Tang. Image fusion in remote sensing: An overview and meta-analysis. Photogrammetric Engineering & Remote Sensing, 90(12): 755-775, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Antreas Antoniou and Amos J. Storkey. Learning to learn by self-critique. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), pages 9936-9946, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 296, + 235 + ], + "type": "text", + "content": "[3] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yukun Cui, Chunxia Zhang, Zhenbo Guo, and Yongjun Wang. Simultaneous automatic picking and manual picking refinement for first-break. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "text", + "content": "[4] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Baisong Jiang, Lilun Deng, Yukun Cui, Shuang Xu, and Chunxia Zhang. Deep unfolding multi-modal image fusion network via attribution analysis. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 346 + ], + "type": "text", + "content": "[5] Haowen Bai, Zixiang Zhao, Jiangshe Zhang, Yichen Wu, Lilun Deng, Yukun Cui, Baisong Jiang, and Shuang Xu. Refusion: Learning image fusion from reconstruction with learnable loss via meta-learning. International Journal of Computer Vision, pages 1-21, 2024. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "text", + "content": "[6] Haowen Bai, Jiangshe Zhang, Zixiang Zhao, Lilun Deng, Yukun Cui, and Shuang Xu. Retinex-mef: Retinex-based glare effects aware unsupervised multi-exposure image fusion. arXiv preprint arXiv:2503.07235, 2025. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 446 + ], + "type": "text", + "content": "[7] Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, and Kyoung Mu Lee. Meta-learning with task-adaptive loss function for few-shot learning. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 9445-9454. IEEE, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "text", + "content": "[8] Yanpeng Cao, Dayan Guan, Weilin Huang, Jiangxin Yang, Yanlong Cao, and Yu Qiao. Pedestrian detection with unsupervised multispectral feature learning using deep neural networks. Information Fusion, 46:206-217, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 556 + ], + "type": "text", + "content": "[9] Bo Cheng, Xiang Li, Yujie Wei, Cheng Huang, Xiaoyong Zhang, Yandong Jiang, Tianyu Zhang, Na Xu, Shuai Yu, Xinxin Zhan, et al. Segformer: Simple and efficient design for semantic segmentation with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12022-12032, 2021. 4, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 557, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 601 + ], + "type": "text", + "content": "[10] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3333-3348, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "text", + "content": "[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the International conference on machine learning (ICML), pages 1126–1135, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 690 + ], + "type": "text", + "content": "[12] Chelsea Finn, Aravind Rajeswaran, Sham Kakade, and Sergey Levine. Online meta-learning. In Proceedings of the International conference on machine learning (ICML), pages 1920-1930, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Qishen Ha, Kohei Watanabe, Takumi Karasawa, Yoshitaka Ushiku, and Tatsuya Harada. Mfnet: Towards real-time se" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "mantic segmentation for autonomous vehicles with multispectral scenes. In IROS, pages 5108-5115. IEEE, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "type": "text", + "content": "[14] Muhammad Haris, Greg Shakhnarovich, and Norimichi Ukita. Task-driven super resolution: Object detection in low-resolution images. In Neural Information Processing: 28th International Conference (ICONIP), pages 387-395. Springer, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 154, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 208 + ], + "type": "text", + "content": "[15] Rein Houthooft, Yuhua Chen, Phillip Isola, Bradley C. Stadie, Filip Wolski, Jonathan Ho, and Pieter Abbeel. Evolved policy gradients. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), pages 5405-5414, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 211, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 211, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 553, + 266 + ], + "type": "text", + "content": "[16] Zhanbo Huang, Jinyuan Liu, Xin Fan, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Reconet: Recurrent correction network for fast and efficient multi-modality image fusion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 539-555. Springer, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 268, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 312 + ], + "type": "text", + "content": "[17] Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible-infrared paired dataset for low-light vision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3496-3504, 2021. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 315, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 555, + 369 + ], + "type": "text", + "content": "[18] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 4369-4376, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 393 + ], + "type": "text", + "content": "[19] Glenn Jocher. Ultralytics YOLOv8. https://github.com/ultralytics, 2023.4,5,7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 396, + 553, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 553, + 450 + ], + "type": "text", + "content": "[20] Chenglong Li, Chengli Zhu, Yan Huang, Jin Tang, and Liang Wang. Cross-modal ranking with soft consistency and noisy labels for robust RGB-T tracking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 831-847. Springer, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 452, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 485 + ], + "type": "text", + "content": "[21] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 487, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 542 + ], + "type": "text", + "content": "[22] Huafeng Li, Yueliang Cen, Yu Liu, Xun Chen, and Zhengtao Yu. Different input resolutions and arbitrary output resolution: A meta learning-based deep framework for infrared and visible image fusion. IEEE Transactions on Image Processing, 30:4070-4083, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 578 + ], + "type": "text", + "content": "[23] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72-86, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 579, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 666 + ], + "type": "text", + "content": "[24] Siyuan Li, Iago Bruno Araujo, Wenqi Ren, Zhangyang Wang, Eric K. Tokuda, Roberto Hirata Junior, Roberto Marcondes Cesar Junior, Jiawan Zhang, Xiaojie Guo, and Xiaochun Cao. Single image deraining: A comprehensive benchmark analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3838-3847. Computer Vision Foundation / IEEE, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[25] Xiaoling Li, Yanfeng Li, Houjin Chen, Yahui Peng, and Pan Pan. Ccafusion: cross-modal coordinate attention network for infrared and visible image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7465" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[26] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Metasgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[27] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "text", + "content": "[28] Zhuoyuan Li, Zikun Yuan, Li Li, Dong Liu, Xiaohu Tang, and Feng Wu. Object segmentation-assisted inter prediction for versatile video coding. IEEE Transactions on Broadcasting, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 263 + ], + "type": "text", + "content": "[29] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In Proceedings of the European Conference on Computer Vision (ECCV), pages 719-735. Springer, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[30] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 354 + ], + "type": "text", + "content": "[31] Jinyuan Liu, Xin Fan, Ji Jiang, Risheng Liu, and Zhongxuan Luo. Learning a deep multi-scale feature ensemble and an edge-attention guidance for image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 32(1): 105–119, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 432 + ], + "type": "text", + "content": "[32] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5792-5801, 2022. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[33] Jinyuan Liu, Jingjie Shang, Risheng Liu, and Xin Fan. Attention-guided global-local adversarial learning for detail-preserving multi-exposure image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5026-5040, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 555 + ], + "type": "text", + "content": "[34] Jinyuan Liu, Zhu Liu, Guanyao Wu, Long Ma, Risheng Liu, Wei Zhong, Zhongxuan Luo, and Xin Fan. Multi-interactive feature learning and a full-time multi-modality benchmark for image fusion and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8115–8124, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "text", + "content": "[35] Jinyuan Liu, Guanyao Wu, Junsheng Luan, Zhiying Jiang, Risheng Liu, and Xin Fan. Holoco: Holistic and local contrastive learning network for multi-exposure image fusion. Information Fusion, 95:237-249, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "text", + "content": "[36] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, 132(5):1748-1775, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[37] Jinyuan Liu, Guanyao Wu, Zhu Liu, Di Wang, Zhiying Jiang, Long Ma, Wei Zhong, and Xin Fan. Infrared and visible image fusion: From data compatibility to task adaption. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[38] Risheng Liu, Zhu Liu, Jinyuan Liu, and Xin Fan. Searching a hierarchically aggregated fusion architecture for fast multi-modality image fusion. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 1600-1608. ACM, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[39] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5953-5969, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 229 + ], + "type": "text", + "content": "[40] Risheng Liu, Zhu Liu, Jinyuan Liu, Xin Fan, and Zhongxuan Luo. A task-guided, implicitly-searched and metainitialized deep model for image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[41] ShuMin Liu, Jiajia Chen, and Susanto Rahardja. A new multi-focus image fusion algorithm and its efficient implementation. IEEE Transactions on Circuits and Systems for Video Technology, 30(5):1374-1384, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "text", + "content": "[42] Zhu Liu, Jinyuan Liu, Guanyao Wu, Long Ma, Xin Fan, and Risheng Liu. Bi-level dynamic learning for jointly multimodality image fusion and beyond. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 1240-1248, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "type": "text", + "content": "[43] Zhu Liu, Jinyuan Liu, Benzhuang Zhang, Long Ma, Xin Fan, and Risheng Liu. Paif: Perception-aware infrared-visible image fusion for attack-tolerant semantic segmentation. In Proceedings of the 31st ACM international conference on multimedia, pages 3706-3714, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "text", + "content": "[44] Zhu Liu, Jinyuan Liu, Guanyao Wu, Zihang Chen, Xin Fan, and Risheng Liu. Searching a compact architecture for robust multi-exposure image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6224-6237, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 443, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 476 + ], + "type": "text", + "content": "[45] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information Fusion, 45:153-178, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[46] Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusion: A generative adversarial network for infrared and visible image fusion. Information Fusion, 48:11-26, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "text", + "content": "[47] Jiayi Ma, Pengwei Liang, Wei Yu, Chen Chen, Xiaojie Guo, Jia Wu, and Junjun Jiang. Infrared and visible image fusion via detail preserving adversarial learning. Information Fusion, 54:85-98, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "text", + "content": "[48] Jiayi Ma, Han Xu, Junjun Jiang, Xiaoguang Mei, and Xiao-Ping (Steven) Zhang. Ddcgan: A dual-discriminator conditional generative adversarial network for multi-resolution image fusion. IEEE Transactions on Image Processing, 29: 4980-4995, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "text", + "content": "[49] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5637-5646, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[50] Alex Nichol, Joshua Achiam, and John Schulman. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7466" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[51] Seonghyun Park, An Gia Vien, and Chul Lee. Cross-modal transformers for infrared and visible image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 34(2):770-785, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 163 + ], + "type": "text", + "content": "[52] Yanting Pei, Yaping Huang, Qi Zou, Yuhang Lu, and Song Wang. Does haze removal help cnn-based image classification? In Proceedings of the European conference on computer vision (ECCV), pages 682-697, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "type": "text", + "content": "[53] Xinran Qin, Yuhui Quan, Tongyao Pang, and Hui Ji. Ground-truth free meta-learning for deep compressive sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9947-9956, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "type": "text", + "content": "[54] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In Proceedings of the International conference on machine learning (ICML), pages 4334-4343, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 268, + 294, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 322 + ], + "type": "text", + "content": "[55] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 325, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 294, + 369 + ], + "type": "text", + "content": "[56] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "text", + "content": "[57] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82:28-42, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "type": "text", + "content": "[58] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Infromation Fusion, 83-84:79-92, 2022. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 463, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 507 + ], + "type": "text", + "content": "[59] Wei Tang, Fazhi He, Yu Liu, Yansong Duan, and Tongzhen Si. Datfuse: Infrared and visible image fusion via dual attention transformer. IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159-3172, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "text", + "content": "[60] Di Wang, Jinyuan Liu, Xin Fan, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 3508-3515. ijcai.org, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 609 + ], + "type": "text", + "content": "[61] Wu Wang, Liang-Jian Deng, Ran Ran, and Gemine Vivone. A general paradigm with detail-preserving conditional invertible network for image fusion. International Journal of Computer Vision, 132(4):1029-1054, 2024. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 612, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 656 + ], + "type": "text", + "content": "[62] Zeyu Xiao and Xinchao Wang. Event-based video superresolution via state space models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[63] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 2409-2418, 2024. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "text", + "content": "[64] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 217-235. Springer, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 162 + ], + "type": "text", + "content": "[65] Han Xu, Jiayi Ma, Zhuliang Le, Junjun Jiang, and Xiaojie Guo. Fusiondn: A unified densely connected network for image fusion. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 12484-12491, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[66] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[67] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "type": "text", + "content": "[68] Han Xu, Jiayi Ma, Jiteng Yuan, Zhuliang Le, and Wei Liu. Rfnet: Unsupervised network for mutually reinforcing multimodal image registration and fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19647-19656, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "type": "text", + "content": "[69] Han Xu, Jiteng Yuan, and Jiayi Ma. Murf: Mutually reinforcing multi-modal image registration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "type": "text", + "content": "[70] Shuang Xu, Ouafa Amira, Junmin Liu, Chun-Xia Zhang, Jiangshe Zhang, and Guanghai Li. HAM-MFN: hyperspectral and multispectral image multiscale fusion network with RAP loss. IEEE Transactions on Geoscience and Remote Sensing, 58(7):4618-4628, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "text", + "content": "[71] Shuang Xu, Lizhen Ji, Zhe Wang, Pengfei Li, Kai Sun, Chunxia Zhang, and Jiangshe Zhang. Towards reducing severe defocus spread effects for multi-focus image fusion via an optimization based strategy. IEEE Transactions Computational Imaging, 6:1561-1570, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "text", + "content": "[72] Yong Yang, Jiaxiang Liu, Shuying Huang, Weiguo Wan, Wenying Wen, and Juwei Guan. Infrared and visible image fusion via texture conditional generative adversarial network. IEEE Transactions on Circuits and Systems for Video Technology, 31(12):4771-4783, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[73] Xunpeng Yi, Linfeng Tang, Hao Zhang, Han Xu, and Jiayi Ma. Diff-if: Multi-modality image fusion via diffusion model with fusion knowledge prior. Information Fusion, 110:102450, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "type": "text", + "content": "[74] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728-5739, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[75] Hao Zhang and Jiayi Ma. Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion. International Journal of Computer Vision, 129(10):2761-2785, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[76] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Com" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7467" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "puter Vision and Pattern Recognition (CVPR), pages 26974-26983, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[77] Yu Zhang, Yu Liu, Peng Sun, Han Yan, Xiaolin Zhao, and Li Zhang. IFCNN: A general image fusion framework based on convolutional neural network. Information Fusion, 54: 99-118, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 296, + 196 + ], + "type": "text", + "content": "[78] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13955-13965, 2023. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 242 + ], + "type": "text", + "content": "[79] Yangyang Zhao, Qingchun Zheng, Peihao Zhu, Xu Zhang, and Wenpeng Ma. Tufusion: A transformer-based universal fusion algorithm for multimodal images. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 298 + ], + "type": "text", + "content": "[80] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Jiangshe Zhang, and Pengfei Li. Didfuse: Deep image decomposition for infrared and visible image fusion. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 970-976, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 300, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 295, + 354 + ], + "type": "text", + "content": "[81] Zixiang Zhao, Shuang Xu, Jiangshe Zhang, Chengyang Liang, Chunxia Zhang, and Junmin Liu. Efficient and model-based infrared and visible image fusion via algorithm unrolling. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1186-1196, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 356, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 356, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 296, + 421 + ], + "type": "text", + "content": "[82] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5906–5916, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 423, + 296, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 296, + 489 + ], + "type": "text", + "content": "[83] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5906–5916, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 491, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 295, + 544 + ], + "type": "text", + "content": "[84] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. DDFM: denoising diffusion model for multi-modality image fusion. CoRR, abs/2303.06840, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 621 + ], + "type": "text", + "content": "[85] Zixiang Zhao, Jiang-She Zhang, Haowen Bai, Yicheng Wang, Yukun Cui, Lilun Deng, Kai Sun, Chunxia Zhang, Junmin Liu, and Shuang Xu. Deep convolutional sparse coding networks for interpretable image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2369-2377, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 689 + ], + "type": "text", + "content": "[86] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25912-25921, 2024. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "text", + "content": "[87] Huabing Zhou, Wei Wu, Yanduo Zhang, Jiayi Ma, and Haibin Ling. Semantic-supervised infrared and visible image" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 205 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "fusion via a dual-discriminator generative adversarial network. IEEE Transactions on Multimedia, 25:635-648, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 151 + ], + "type": "text", + "content": "[88] Man Zhou, Naishan Zheng, Xuanhua He, Danfeng Hong, and Jocelyn Chanussot. Probing synergistic high-order interaction for multi-modal image fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[89] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 1, 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7468" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_content_list.json b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ce4f3b37ebd864902f8dfaee78360f8625d188b6 --- /dev/null +++ b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_content_list.json @@ -0,0 +1,1900 @@ +[ + { + "type": "text", + "text": "Taste More, Taste Better: Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maochen Yang $^{1}$ Zekun Li $^{1}$ Jian Zhang $^{1}$ Lei Qi $^{3}$ Yinghuan Shi $^{1,2,*}$ $^{1}$ Nanjing University, China $^{2}$ Suzhou Laboratory, China $^{3}$ Southeast University, China \n{yangmaochen, lizekun, zhangjian7369}@smail.nju.edu.cn qilei@seu.edu.cn, syh@nju.edu.cn", + "bbox": [ + 104, + 202, + 888, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised crowd counting is crucial for addressing the high annotation costs of densely populated scenes. Although several methods based on pseudo-labeling have been proposed, it remains challenging to effectively and accurately utilize unlabeled data. In this paper, we propose a novel framework called Taste More Taste Better (TMTB), which emphasizes both data and model aspects. Firstly, we explore a data augmentation technique well-suited for the crowd counting task. By inpainting the background regions, this technique can effectively enhance data diversity while preserving the fidelity of the entire scenes. Secondly, we introduce the Visual State Space Model as backbone to capture the global context information from crowd scenes, which is crucial for extremely crowded, low-light, and adverse weather scenarios. In addition to the traditional regression head for exact prediction, we employ an Anti-Noise classification head to provide less exact but more accurate supervision, since the regression head is sensitive to noise in manual annotations. We conduct extensive experiments on four benchmark datasets and show that our method outperforms state-of-the-art methods by a large margin. Code is publicly available on https://github.com/syhien/taste_more_tastebetter.", + "bbox": [ + 89, + 323, + 483, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 698, + 220, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Crowd counting is a fundamental task in computer vision that involves estimating the number of people in densely populated scenes. This task is of great importance in various public safety scenarios, including transport manage-", + "bbox": [ + 89, + 722, + 483, + 785 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47c9357176488bf309c4f0f5a4e889f133bf456093d71c136db3a4a851f5444f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 289, + 648, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abbcdea02b1500d23c7a11a6b90130f2af84030408220d84ba885b4b2ae827b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 290, + 759, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a14293dff0a56dad803657968e30e8bf7dd2ea8ae16df59c0177b190f60db381.jpg", + "image_caption": [ + "Original A density map" + ], + "image_footnote": [], + "bbox": [ + 766, + 290, + 870, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97aec8fc0222aadb987475cab9c9228e19efc5734f45015ee7009ce2cd478033.jpg", + "image_caption": [ + "Mixup (A + background of B)", + "Mixup (A + B)" + ], + "image_footnote": [], + "bbox": [ + 544, + 358, + 648, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/812e4c71821035264a2aefe801a99994482187486a7b629922447ceecb94e854.jpg", + "image_caption": [ + "Mixup (background only) density map", + "Mixup density map" + ], + "image_footnote": [], + "bbox": [ + 653, + 358, + 759, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c18cc0f918477e80c423786c991d9d4ac67788115c4182d50c6ec08d337f33a7.jpg", + "image_caption": [ + "Original B density map" + ], + "image_footnote": [], + "bbox": [ + 763, + 358, + 870, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5291994ce4a7d114b49708f949aa45d1a945e5c67d051fd6aeef41d6c8b0fb33.jpg", + "image_caption": [ + "Inpainted A" + ], + "image_footnote": [], + "bbox": [ + 545, + 428, + 648, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c03ec3d1cf06d34965e3e49dd036e8d69846e6c2c667ef7e61ed5969e698a1f3.jpg", + "image_caption": [ + "Inpainted A density map" + ], + "image_footnote": [], + "bbox": [ + 655, + 428, + 759, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/379c121efee75d12876182ab18f8e8a76abeab0a0ac954f2784f4c477f588be2.jpg", + "image_caption": [ + "Inpainted B" + ], + "image_footnote": [], + "bbox": [ + 544, + 497, + 648, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfc17588551bad9a13f1450df2644d242306fcc3e81795ebf82cf6c0a203a3bb.jpg", + "image_caption": [ + "Inpainted B density map" + ], + "image_footnote": [], + "bbox": [ + 653, + 497, + 759, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c37fd19f3b54f97b7ca17371b95fdf005f1b35f812261f75a306bef9a1ea9bb.jpg", + "image_caption": [ + "Figure 1. Comparisons of augmentations. As shown in the red bounding boxes, both types of Mixup generate disappointing density maps. Mixup is supposed to retain crowd information from two images, but actually it destroys the spatial structure. As shown in green bounding boxes, our inpainting augmentation is better suited for the crowd counting task, showing an impressive reduction in Mean Absolute Error (MAE)." + ], + "image_footnote": [], + "bbox": [ + 772, + 430, + 870, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ment [14, 50, 61] and disaster prevention [51]. The substantial demand has driven the advancement of counting methods [3, 25, 43, 45, 65]. However, a significant cost of manual annotation is unavoidable for these fully-supervised deep methods to achieve satisfactory performance. For example, the UCF-QNRF dataset with 1,535 images and 1,251,642 annotations amounts to approximately 2,000 human-hours [17]. Compared to the extremely expensive labeled data with annotations, acquiring unlabeled crowd images is much more affordable. Consequently, there is a growing interest in leveraging unlabeled data through semi-supervised crowd counting (SSCC) methods. The recent mainstream works on SSCC [21, 24, 40, 68] mostly employ the Mean Teacher [52] framework, where the teacher", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: Yinghuan Shi. This work was supported by National Science and Technology Major Project (2023ZD0120700), NSFC Project (62222604, 62206052, 624B2063), China Postdoctoral Science Foundation (2024M750424), Fundamental Research Funds for the Central Universities (020214380120, 020214380128), State Key Laboratory Fund (ZZKT2024A14), Postdoctoral Fellowship Program of CPSF (GZC20240252), Jiangsu Funding Program for Excellent Postdoctoral Talent (2024ZB242), and Jiangsu Science and Technology Major Project (BG2024031).", + "bbox": [ + 89, + 791, + 483, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24440", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model generates pseudo-labels for unlabeled data to guide the student model. The key to achieving superior performance with SSCC methods on a given set of images, where only a small portion is annotated, lies in how we handle the unlabeled samples and how we design an appropriate model to fully exploit the information within them.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It has been observed that introducing similar but different samples to the training data can enlarge the support of the training distribution and thus enhance the robustness and generalizability of the model [4, 46]. In the literature, numerous data augmentation techniques have demonstrated their effectiveness in generating additional virtual samples to support the training process in both fully-supervised and semi-supervised settings [2, 63, 66]. However, we find that these off-the-shelf techniques cannot suit the crowd counting task, as they may destroy the spatial structure of crowd scenes, leading to distorted density maps, as shown in Figure 1. Therefore, a well-suited data augmentation technique specifically designed for the crowd counting task is crucial for the model to taste more from the unlabeled data.", + "bbox": [ + 89, + 184, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the help of appropriate data augmentation, the diversity of unlabeled data during training can be significantly improved. However, this simultaneously brings new challenges for the teacher model to produce correct supervision signals. From the perspective of the model itself, we find that the Convolutional Neural Network (CNN) architecture commonly used in previous studies can be suboptimal for the crowd counting task, as it tends to overfit to local details and fails to capture global context information, which is crucial for extremely crowded, low-light, and adverse weather scenarios. From another perspective, it is also important to generate more accurate pseudo-labels while filtering out low-quality parts. Considering these two points, we explore a modern backbone network to replace CNNs and enhance it with a holistic strategy for producing and leveraging pseudo-labels, which helps to taste data better.", + "bbox": [ + 89, + 398, + 482, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel semi-supervised crowd counting method, named Taste More Taste Better (TMTB). We design a novel data augmentation technique, Inpainting Augmentation, which enhances the diversity of the dataset while keeping the integrity of the foreground regions. Observing that some regions in inpainted images are noisy and harmful, to better utilize the inpainted images, we employ a Visual State Space Model (VSSM) with an Anti-Noise classification branch for assistance, which provides inexact but accurate supervision. The proposed model generates more accurate density maps than state-of-the-art methods. Our method TMTB achieves significant improvements on several benchmark datasets, including ShanghaiTech Part A & B [69], UCF-QNRF [17], and JHU-Crowd++ [47]. We reduce the Mean Absolute Error of JHU-Crowd with $5\\%$ data labeled to 67.0, marking the first time it has been lower than 70.0. The method also presents strong generalization ability", + "bbox": [ + 89, + 643, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in cross-dataset evaluations.", + "bbox": [ + 513, + 92, + 699, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are as follows:", + "bbox": [ + 532, + 106, + 872, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We observed that the lack of suitable data augmentation methods and the weakness of capturing global context information are bottlenecks for the crowd counting task. The off-the-shelf data methods may destroy the spatial structure of crowd scenes, which is crucial for the regression of density maps. CNNs tend to overfit local details and struggle to capture global context information.", + "- We propose a novel semi-supervised crowd counting method. A data augmentation technique is designed to enhance diversity while keeping the integrity of the foreground regions. A VSSM with an Anti-Noise classification branch for assistance is employed, which provides inexact but accurate supervision, to better utilize the data.", + "- The proposed method remarkably outperforms state-of-the-art methods across four benchmark datasets and three label settings. It achieves a groundbreaking reduction in the Mean Absolute Error on the JHU-Crowd dataset with only $5\\%$ of the data labeled, lowering it to 67.0, a decrease of $12.4\\%$ . Furthermore, the method demonstrates exceptional generalization capabilities in cross-dataset Domain Generalization settings, even surpassing the performance of fully-supervised approaches, setting a new standard for excellence." + ], + "bbox": [ + 513, + 121, + 903, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Releted Works", + "text_level": 1, + "bbox": [ + 513, + 482, + 661, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-supervised Crowd Counting. Fully-supervised crowd counting methods[5, 13, 17, 31, 38, 42, 49, 69] maintain the best performance in crowd counting tasks. However, fully-supervised methods require a large quantity of labeled data, which is expensive to collect and annotate.", + "bbox": [ + 511, + 508, + 906, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, semi-supervised crowd counting methods have been proposed to address the scarcity of labeled data. The Mean Teacher [52], a successful semi-supervised learning framework that has been widely used in classification [10, 48, 64], segmentation [2, 56] and regression [8, 19, 62], is the most popular method in semi-supervised crowd counting. L2R [26] leveraged unlabeled data through a learning-to-rank framework. IRAST [27] introduced surrogate tasks, self-learning inter-related binary segmentation tasks, to improve performance. Moreover, MTCP [73] applied multiple auxiliary tasks into representation learning, which also benefit performance. DA-Count [23] employed a Transformer [53] to refine foreground features. CU [21] investigated a supervised uncertainty estimation strategy to better utilize unlabeled data. SALCrowd [68] proposed a method to select unlabeled data based on intra-scale uncertainty and inter-scale inconsistency metrics. MRC-Count [40] incorporated a fine-grained density classification task, proving the effectiveness of the classification task in crowd counting. Most previous approaches [21, 26, 27, 40, 68] are based on CNNs or par", + "bbox": [ + 511, + 584, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24441", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/76ef64fe59dac56b1b3f9d223f42aacb5703dbe5a17cc6197ed4e9dfa5228d8f.jpg", + "image_caption": [ + "Figure 2. The overall framework of our method TMTB. TMTB contains the Mean Teacher framework for semi-supervised learning, VSSM as the backbone, a classification branch for predicting masks, and an inpainter for inpainting augmentation. The inpainting process is conducted periodically, which generates or updates the inpainted images. For filtering out unreliable regions, the teacher model generates the weighted mask based on the inconsistency level of the inpainted image, which is applied in $\\mathcal{L}^{\\mathrm{inp}}$ . The teacher model generates pseudolabels for unlabeled data (inpainted images included), and the student model predicts on labeled data and strongly augmented unlabeled data (inpainted images included)." + ], + "image_footnote": [], + "bbox": [ + 114, + 74, + 883, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tially [23] based on Transformers, which are weak in modeling long-range dependencies. When facing extremely dense crowd scenes, the performance of these methods is not satisfactory.", + "bbox": [ + 88, + 433, + 480, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Data Augmentation. Deep learning models tend to overfit when the quantity of data is limited, which harms the generalization ability of models. Therefore, data augmentation was proposed to address this issue. For image-based tasks, basic data augmentation methods include random cropping, flipping, rotation, scaling, color jittering, and noise adding. Some advanced data augmentation methods, such as CutMix [63], Mixup [66], and AutoAugment [7], have been proposed. However, these methods are not suitable for crowd counting tasks. Both the density-based and point-based crowd counting methods are sensitive to the spatial structure of the crowd, and the above methods may destroy the spatial structure of the crowd. The damage to the integrity of the crowd will lead to a decrease in model performance. We study this in Table 2.", + "bbox": [ + 88, + 500, + 482, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Data Augmentation [11, 15, 16, 20, 41] is the process of creating new synthetic data points that can be added to existing datasets. This unique approach is widely used to improve the performance of models by increasing the quantity and diversity of training data while maintaining the quality of the data. In [67], queries-only data was used to improve the multilingual task, showing the effectiveness of generative data augmentation. FreeMask [59] resorted to synthetic images from generative models to ease the burden of both data collection and annotation procedures, and achieves comparable performance to real-data-based meth", + "bbox": [ + 88, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ods. G-DAUG $^{\\text{C}}$ [60] generated synthetic examples using pretrained language models, and selects the most informative and diverse set of examples for data augmentation, to achieve more accurate and robust learning in low-resource setting. DiffusionMix [18] employed an inpaint-and-mix strategy to generate label-preserved data for classification task. However, it requires pre-defined binary masks to concatenate the original and generated images, which is not available in crowd counting since the position and shape of the crowd are unpredictable. Compared to existing works, our contribution is to fill the gap in effective augmentation methods within the field of pixel-level predition tasks.", + "bbox": [ + 511, + 431, + 906, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 631, + 604, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a semi-supervised crowd counting method that enhances the dataset using a diffusion-based model's inpainting technique and utilizes a VSSM (Visual State Space Model) with Anti-Noise classification branch to improve data diversity and model performance.", + "bbox": [ + 511, + 657, + 905, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of our method, TMTB, is shown in Figure 2. Our method is built on the Mean Teacher framework, which is widely employed in semi-supervised learning. The teacher model shares the same architecture as the student model, and its weights are updated using the student model's weights through the Exponential Moving Average (EMA) strategy. VMamba [29], a novel VSSM (Visual State Space Model), serves as the backbone, replacing the CNN and Transformer. The image features extracted from VMamba are then fed into two independent branches: the counting branch and the classification branch, for pre", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24442", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dictions. The counting branch predicts a density map of the image, while the classification branch predicts the indices of pre-defined count interval bins. We build a module, Inpainter, designed to periodically applying background inpainting augmentation to the training set.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first describe our novel inpainting augmentation in Section 3.1, then we discuss our network architecture in Section 3.2, and finally, we present our loss functions in Section 3.3.", + "bbox": [ + 89, + 167, + 482, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Inpainting Augmentation", + "text_level": 1, + "bbox": [ + 89, + 239, + 320, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike classification or regression tasks, the ground truth density maps are generated by point-level annotations. The state-of-the-art data augmentation techniques, Mixup [66], CutMix [63], BCP [2], etc., are not applicable to crowd counting, since the augmentation may cause damage to ground truth density maps. We introduce the inpainting technique of diffusion-based models [15, 44] as a novel data augmentation method with the filtering capability of unreliable regions, which generates new samples, thereby expanding the diversity of dataset.", + "bbox": [ + 89, + 263, + 482, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the inpainting mask $M^{\\mathrm{inp}}$ , the count interval bins predicted by the classification branch are used to separate the foreground and the background of image $x$ . We generate the positive text prompt randomly to achieve more diverse unlabeled dataset, but the negative text prompt is kept the same for all samples. The inpainting process is conducted periodically during the training process using a preset time interval $T^{\\mathrm{inp}}$ :", + "bbox": [ + 89, + 415, + 483, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} M ^ {\\text {i n p}} = \\mathbf {0} (\\arg \\max (\\hat {p})) , \\\\ x ^ {\\text {i n p}} = \\mathcal {S D} (x, M ^ {\\text {i n p}}, r ^ {\\text {p o s}}, r ^ {\\text {n e g}}), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 546, + 480, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{p}_i$ is the predicted probability distribution from the classification branch, $\\mathbf{0}(\\cdot)$ is the indicator function, $x^{\\mathrm{inp}}$ is the new image with the background inpainted, $\\mathcal{SD}(\\cdot ,\\cdot ,\\cdot ,\\cdot)$ represents Stable Diffusion [44], which is a mature, excellent, and cost-effective diffusion model, and $r^{\\mathrm{pos}}$ and $r^{\\mathrm{neg}}$ are the positive and negative text prompts, respectively.", + "bbox": [ + 89, + 597, + 482, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 shows examples of both good and poor inpainting samples. All new samples are considered unlabeled, as the inpainting process may add additional humans for better coherence. For well-inpainted samples, the foreground regions are well preserved, while the background regions are inpainted with different patterns, thus introducing diversity to the dataset. Even though the inpainting process is not perfect, the model can still benefit from the augmented dataset. However, failures may sometimes occur, potentially leading to the model learning from noisy inputs. The heads of humans and other features in the unreliable regions of poorly-inpainted images are fragmented and indistinct, rendering them difficult to recognize. It is necessary to filter out these unreliable regions. The EMA model calculates", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bdc36da9be0e50dd7b749392b7f13ceb2eff1446f28a4a80bfc4e5660b7804cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 90, + 612, + 135 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0f92ba5a7ab830b4627ddfea077397e786625cfe7c2d654052dab268597d3b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 93, + 697, + 133 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4f27d485cf7d13bde71c285d6b41f5f8b072130bfefa7e2f8393bce1b8a1a7be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 138, + 612, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/025f93ebc72dbdf9dc285e09df03e7ae55ee027e6c7d1d7fa46f1061f8db5f22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 138, + 697, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f850b867f7f3619df9b21ebb8afa6ea90b60fb34b6f235ed8ad620cf14d070e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 184, + 612, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/30ee914bbc6d7ca52c8d2e31ed79801e03264d58d6fa788b41ad04760c9c2783.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 184, + 697, + 227 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0aa76cc00917bd5871fea08311e2040a992a332e5789600cf1280fc610a8d3e6.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 532, + 229, + 612, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50c7922964f420d591552c3474d0e7f1cc268e3df5b08a9c8e5a1c3e37c2cd1c.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 619, + 229, + 697, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3bb6e907c05a346057b0d1c786758b2af710582b9971316459c62977b44bf6f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 92, + 797, + 135 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fa554f5d095db2c28ba03de2179678f6459ad75050bc7b8948a88823f6adc0df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 138, + 797, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/df8f947ff5583cc6dd5c85e84d78c94ccdb86d94aa45626b261c2f8c2132999e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 184, + 797, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/217ecad13b4843318302778885535058acb664ae03829c254a1dfd1c72fce3bd.jpg", + "image_caption": [ + "(c)", + "Figure 3. Inpainting samples. The left part shows well-inpainted couples, while the right part shows poorly-inpainted couples. Column (a) and (c) are the original images, while column (b) and (d) are the inpainted images." + ], + "image_footnote": [], + "bbox": [ + 718, + 229, + 797, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e58cf3d47797b103aa8baba045040c275ac652bee2954f9d6a9a53553e164f83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 92, + 883, + 135 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/00a245736f630b4e92effb4fea8f8cec05bde2bebee3bbba959ac4a137f15246.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 138, + 883, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a8d75ad999d2c8b7fda930ac51b7eb891eb9b607341b6fa914bf706905835be2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 183, + 883, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a98e93f86a4b5f8b6e3649c7a3a223aa60f53430fe51d6b2206027484d83291.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 803, + 229, + 883, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the inconsistency level of the inpainted image by the difference in the classification branch between the weakly augmented version and the strongly augmented version. The inconsistency level and the weighted mask are defined as follows:", + "bbox": [ + 511, + 356, + 905, + 430 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} M _ {l} ^ {\\text {i n c o n}} = \\mathbf {0} \\left(\\left| \\left| p ^ {\\hat {s}} - p ^ {\\hat {w}} \\right| - l \\right|\\right), \\\\ \\omega_ {l} ^ {t} = \\operatorname {s o f t m a x} \\left(e ^ {- l \\cdot t / T ^ {\\mathrm {i n p w}}}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 441, + 903, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {\\text {i n c o n}} = \\sum_ {l = 0} ^ {L} M _ {l} ^ {\\text {i n c o n}} \\cdot \\omega_ {l} ^ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 491, + 771, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{p^s}$ and $\\hat{p^w}$ are the predicted probability distributions of the classification branch of EMA model in the strong and weak augmentation version, respectively. $M_{l}^{\\mathrm{incon}}$ is the mask of inconsistency level $l$ , $M^{\\mathrm{incon}}$ is the final weighted mask, $\\omega_l^t$ is the weight of inconsistency level $l$ at time $t$ , $T^{\\mathrm{inpw}}$ is the parameter controlling the weight decay speed, and $L$ is the number of inconsistency levels. The $L$ depends on the predefined count interval bins, and since we followed the interval bins settings in [55], the $L$ is set to 2 for the balance of diversity and reliability. The weighted mask is then used to adjust the loss associated with the inpainted image.", + "bbox": [ + 511, + 546, + 906, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. VSSM with Classification for Anti-Noise", + "text_level": 1, + "bbox": [ + 511, + 726, + 857, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual State Space Models (VSSMs). Existing crowd counting methods often encounter challenges such as congestion, blur, and occlusion in intricate scenes. CNN-based approaches tend to overfit to local details, while Transformer-based methods excel in global modeling but face heavy computational burdens due to their quadratic complexity in terms of image sizes. These limitations motivate us to introduce the VSSM, which captures comprehensive long-range information with linear time complexity, for the crowd counting task.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24443", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/12cca87cdaa083d9ec4bc9bf936bc52d361120c616e02f6ccf8c91be6d70fa72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 153, + 77, + 323, + 159 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aa5de0346394fe5fd1ccf8a8471f1593847b7ffbf033415979e0b8c5f51afe79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 77, + 468, + 159 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/840b9ba167a45bf2e54efcb55b10604a76aabfa4b3f198cef7076f55bc1a0c71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 78, + 651, + 159 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6083d173668cbaa579be6f9dae596ba72d1373e6bbff80d87604c00ceccde312.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 78, + 844, + 159 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec707a3dec81ae4e3e4ed7bac659fd77064f1c70d8a08addd03beac6bfcb14f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 155, + 161, + 325, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/649f28792ff5649ac80710d1d70c17e8b82d9400e6c42df4b2323d80ff00a56e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 161, + 467, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/21cb4283be0fbca87edb8969f7363f4710ecd00c3c354274ca3e5c1c9baeced6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 161, + 650, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ce423d4f10383661bbb537b6e2c9cde346c6dd30fc12be5b602876c958b15f63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 161, + 843, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/976da9fe03f040cc5011e9a8e5b74926a1ad48e4358b08f1d9cd6446a3ffa3f9.jpg", + "image_caption": [ + "Figure 4. Visualizations of the predictions on the test set of the JHU-Crowd++ dataset. The first row shows the input images. The second row displays the predictions of the SOTA method MRC-Crowd [40]. The third row presents the predictions of our method TMTB. All models are trained with $5\\%$ labeled data." + ], + "image_footnote": [], + "bbox": [ + 155, + 246, + 323, + 328 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/23d7638c046dab4b4d52fa2660fd5c1c5139fa6a738f34c446bb517d956ed437.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 246, + 467, + 328 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a3c99317cf5a12f8c7c359e3564d014847408255575c8ee6adf2545634db76f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 246, + 650, + 328 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39b616760553d3f6e9c0b669d9da97af9cefd58d77c32b5bd260aa75da1a44c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 246, + 843, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "State Space Models (SSMs) map the input stimulation $u(t)$ to the output response $y(t)$ by means of a hidden state $h(t)$ . These processes are formulated with the following form:", + "bbox": [ + 89, + 393, + 482, + 453 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} h ^ {\\prime} (t) = A h (t) + B u (t), \\\\ y (t) = C h (t) + D u (t), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 460, + 480, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A, B, C,$ and $D$ are the weighting parameters.", + "bbox": [ + 89, + 508, + 439, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Linear time-invariant SSMs (Equation (3)) lack the ability to capture the contextual information, [12] proposed a novel parameterization method for SSMs that integrates an input-dependent selection mechanism (referred to as S6). VMamba [29] replaces the S6 module, with the newly proposed 2D-Selective-Scan (SS2D) module which concurrently achieving global receptive fields, dynamic weights, and linear complexity. By adopting complementary 1D traversal paths, SS2D enables each pixel in the image to effectively integrate information from all other pixels in different directions, thereby facilitating the establishment of global receptive fields in the 2D space and generating more accurate density map. Details about the SS2D module could be found in [29]. The features extracted by the final SS2D block are then fed into two branches: one for density map regression and the other for count interval classification, which is designed for less inexact and more accurate supervision.", + "bbox": [ + 88, + 523, + 482, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exact: Regression Head. The regression head is designed to predict the density map for the input image. The density map is then summed to obtain the exact final count value. The regression head consists of up-sampling layers, convolutional layers, batch normalization layers, and ReLU activation layers. We followed MRC-Crowd [40] employing the loss function proposed in CUT [39], which significantly", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "boosts the performance in the foreground area. It could be expressed as:", + "bbox": [ + 511, + 393, + 906, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e g}} = \\frac {1}{J} \\sum_ {j = 1} ^ {J} \\left(1 - \\operatorname {S S I M} \\left(D _ {j} (\\hat {y} \\odot M ^ {g t}\\right), \\right. \\tag {4} \\\\ \\left. D _ {j} \\left(y ^ {g t} \\odot M ^ {g t}\\right)\\right) + \\alpha \\cdot \\mathcal {L} _ {T V} \\left(\\hat {y}, y ^ {g t}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 434, + 903, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{SSIM}(\\cdot ,\\cdot)$ is the structural similarity index measure, and $D_{j}$ denotes the average pooling operation that downsamples an image to $1 / 2^{j}$ of its original size. The operation $\\odot$ denotes the Hadamard product, and $M^{gt}$ is the binary mask that distinguishes the dense and sparse regions based on a predefined threshold $\\tau$ . It can be obtained using the indicator function $\\mathbf{1}$ as $M^{gt} = \\mathbf{1}(y^{gt} > \\tau)$ . $\\alpha$ is a hyperparameter that balances the two terms. $J$ and $\\alpha$ are set to 3 and 0.01, respectively. $\\mathcal{L}_{TV}$ is the total variation loss.", + "bbox": [ + 511, + 508, + 905, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accurate: Anti-Noise Classification Head. Many previous works [1, 31, 34, 54] have identified the problem of inaccurate ground truth annotations in crowd counting datasets, which render the ground truth density maps noisy and inaccurate. In this way, the model may learn from the noisy ground truth and generate inaccurate density maps. Inspired by previous works [23, 40, 55], we employ the Anti-Noise classification task as an auxiliary task, which is more robust in predicting inexactly and accurately. The Anti-Noise classification branch is supervised by the inexact but accurate count interval bins rather than the exact but maybe inaccurate density maps. Furthermore, we use the prediction of the this branch to generate reliable masks for inpainting samples, and we discuss this in Section 3.1. The goal of the Anti-Noise classification branch is to predict the accurate count interval bins, which are predefined by [55].", + "bbox": [ + 511, + 645, + 906, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Anti-Noise classification head consists of an up", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24444", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sampling layer, a convolutional layer, a ReLU activation layer and a final convolutional layer. The standard cross-entropy loss is applied. In the case of labeled data, the loss is defined as:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l s}} = \\frac {1}{N} \\sum_ {i} ^ {N} \\mathcal {H} \\left(p _ {i} ^ {g t}, \\hat {p} _ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 160, + 482, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{H}$ is the cross-entropy loss function, $p_i^{gt}$ is the ground truth probability distribution, and $\\hat{p}_i$ is the predicted probability distribution.", + "bbox": [ + 89, + 214, + 483, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Losses", + "text_level": 1, + "bbox": [ + 89, + 270, + 179, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to make full use of the labeled data $L = \\{(x_{i}^{l},DM_{i}^{gt})\\}_{i = 1}^{N_{l}}$ and unlabeled data $U = \\{(x_{i}^{u})\\}_{i = 1}^{N_{u}}$ , we employ a common semi-supervised learning strategy widely used in crowd counting. The labeled samples are used to directly train the student model, while the unlabeled samples are used for consistency regularization. For labeled data, the supervised loss $\\mathcal{L}^{\\mathrm{s}} = \\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{s}} + \\mathcal{L}_{\\mathrm{cls}}^{\\mathrm{s}}$ is a combination of the regression loss (Equation (4)) and the classification loss (Equation (5)). Given the absence of a confidence score for the predicted regression map, we choose to apply consistency regularization to all unlabeled samples, rather than developing an algorithm to assess prediction uncertainty. The EMA teacher model generates pseudo-labels based on the prediction of unlabeled data with weak augmentation. The student model makes predictions based on the strong augmentation version. The error loss between the two predictions is calculated as the consistency loss.", + "bbox": [ + 89, + 292, + 483, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite the common strong augmentation, we also introduce a patch-aligned random masking strategy [58] as a kind of strong augmentation. The patch-aligned random masking strategy randomly masks out some patches of various sizes in the image. The student model is required to predict based on masked images, while the teacher model is able to access the full image. The semi-supervised loss is defined as:", + "bbox": [ + 89, + 549, + 483, + 669 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {u}} = \\operatorname {M A E} \\left(\\hat {y} ^ {s t}, \\hat {y} ^ {e m a}\\right) + \\operatorname {M A E} \\left(\\hat {p} ^ {s t}, \\hat {p} ^ {e m a}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 680, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{y}^{st}$ and $\\hat{y}^{ema}$ are the predicted density maps of the student model and the teacher model, respectively. $\\hat{p}^{st}$ and $\\hat{p}^{ema}$ are the predicted count interval bins of the student model and the teacher model, respectively.", + "bbox": [ + 89, + 705, + 482, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistency loss is also applied to the inpainted images. However, to avoid the model learning from noisy regions with low quality, the loss is adjusted by the weighted mask of inconsistency level. The inpainting loss is defined as:", + "bbox": [ + 89, + 767, + 482, + 828 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {i n p}} = M ^ {\\text {i n c o n}} \\times \\operatorname {M A E} \\left(\\hat {y} ^ {s t}, \\hat {y} ^ {e m a}\\right) + M ^ {\\text {i n c o n}} \\times \\operatorname {M A E} \\left(\\hat {p} ^ {s t}, \\hat {p} ^ {e m a}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 838, + 506, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M^{\\mathrm{incon}}$ is the weighted mask of inconsistency level mentioned in Equation (2).", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final loss is a combination of the supervised, semi-supervised, and inpainting losses:", + "bbox": [ + 513, + 90, + 903, + 122 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} ^ {\\mathrm {s}} + \\lambda_ {w} \\cdot \\mathcal {L} ^ {\\mathrm {u}} + \\lambda_ {w} \\cdot \\mathcal {L} ^ {\\text {i n p}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 132, + 903, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_w$ is the warm-up weight for the semi-supervised loss and the inpainting loss.", + "bbox": [ + 511, + 152, + 903, + 184 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 195, + 645, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on four public datasets:", + "bbox": [ + 511, + 220, + 833, + 234 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- JHU-Crowd++ [47] is a comprehensive dataset with 4,372 images and 1.51 million annotations, collected under a variety of scenarios and environmental conditions.", + "- UCF-QNRF [17] contains 1535 images. The dataset has a greater number of high-count crowd images and annotations, and a wider variety of scenes containing the most diverse set of viewpoints, densities and lighting variations.", + "- ShanghaiTech A & B [69] are two datasets containing nearly 1,200 images and 330,000 labeled heads." + ], + "bbox": [ + 513, + 236, + 903, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To achieve fair comparisons, we follow the same semi-supervised experimental protocol as previous works [23, 24, 40], including the $5\\%$ , $10\\%$ , and $40\\%$ labeled data settings and the split of labeled and unlabeled data. The evaluation metrics are Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The detailed definitions of them are provided in the supplementary material.", + "bbox": [ + 511, + 388, + 905, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For some excessively high-resolution images, we scale them down, ensuring the longest side does not exceed 1,920 pixels. For the ground truth density map, we use geometry-adapted Gaussian kernels described in [22] for all datasets except the ShanghaiTech B dataset, and we fix the size of the kernel to 4. When training, the input images are randomly cropped to $512 \\times 512$ pixels for all datasets except for ShanghaiTech A. The inputs from ShanghaiTech A are cropped to $256 \\times 256$ pixels instead. The count interval bins, which should be predicted by the classification branch, are set to the same as those in [55]. The AdamW optimizer [30] with a learning rate of $10^{-5}$ is used to optimize the model. The weight decay is set to $10^{-4}$ . The downsample factor is set to 8. The batch size is set to 8. In $5\\%$ and $10\\%$ labeled data settings, the ratio of labeled data to unlabeled data is set to $2:6$ . In the $40\\%$ labeled data setting, the ratio is set to $4:4$ . As we mentioned in Section 3.3, we use the patch-aligned random masking strategy [58] as a form of strong augmentation. The masked patch size and the masked region ratio are set to $32 \\times 32$ and 0.3, respectively. The ablation details can be found in [40]. The teacher model is updated by the exponential moving average with a decay rate of 0.97. The weights of all losses are set to 1.0. For $\\mathcal{L}^{\\mathrm{u}}$ and $\\mathcal{L}^{\\mathrm{inp}}$ , we ramp the weights from 0 to 1.0 in the first 20 epochs. The max accepted inconsistency level $L$ is set to 2, and the hyperparameter $T^{\\mathrm{inpw}}$ is set to 100. The ablation study of $T^{\\mathrm{inpw}}$ is conducted in Section 4.2. Each training", + "bbox": [ + 511, + 493, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24445", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d57826a0937f83c71df6819d2c233a5ebbf3e5cd96eb5bdf80fb2c965739f71c.jpg", + "table_caption": [ + "Table 1. Comparisons with the SOTA methods on main public datasets. The best results are highlighted in bold, and the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodsVenueLabeled PercentageJHU-Crowd++UCF-QNRFShanghaiTech AShanghaiTech B
MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓
MT [52]NIPS'175%101.5363.5172.4284.9104.7156.919.333.2
L2R [26]CVPR'185%101.4338.8160.1272.3103.0155.420.327.6
DACount [23]ACM MM'225%82.2294.9120.2209.385.4134.512.622.8
OT-M [24]CVPR'235%82.7304.5118.4195.483.7133.312.621.5
MRC-Crowd [40]T-CSVT'245%76.5282.7101.4171.374.8117.311.717.8
TMTBOurs5%67.0252.296.3163.472.4126.410.619.6
MT [52]NIPS'1710%90.2319.3145.5250.394.5156.115.624.5
L2R [26]CVPR'1810%87.5315.3148.9249.890.3153.515.624.4
DACount [23]ACM MM'2210%75.9282.3109.0187.274.9115.511.119.1
MTCP [73]T-NNLS'2310%88.1318.7124.7206.381.3130.514.522.3
OT-M [24]CVPR'2310%73.0280.6113.1186.780.1118.510.818.2
CU [21]ICCV'2310%74.9281.7104.0164.370.8116.69.717.7
SALCrowd [68]ACM MM'2410%69.7263.5106.7171.369.7114.59.717.5
MRC-Crowd [40]T-CSVT'2410%70.7261.393.4153.267.3106.810.318.2
TMTBOurs10%66.3265.191.7156.365.7110.49.817.2
MT [52]NIPS'1740%121.5388.9147.2247.688.2151.115.925.7
L2R [26]CVPR'1840%123.6376.1145.1256.186.5148.216.825.1
SUA [33]ICCV'2140%80.7290.8130.3226.368.5121.914.120.6
DACount [23]ACM MM'2240%65.1260.091.1153.467.5110.79.614.6
OT-M [24]CVPR'2340%72.1272.0100.6167.670.7114.58.113.1
SALCrowd [68]ACM MM'2440%62.0255.195.0167.060.797.27.912.7
MRC-Crowd [40]T-CSVT'2440%60.0227.381.1131.562.195.57.813.3
TMTBOurs40%60.0240.381.4140.660.899.07.512.9
", + "bbox": [ + 133, + 127, + 864, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "process coordinates with two inpainting background processes. The inpainting period $T^{\\mathrm{inp}}$ is set to 80 epochs.", + "bbox": [ + 89, + 473, + 482, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 89, + 513, + 230, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our main experiments, the proposed TMTB demonstrates superior performance compared to state-of-the-art methods across four public datasets with three labeled data percentages. The results are shown in Table 1.", + "bbox": [ + 89, + 537, + 482, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, when utilizing only $5\\%$ labeled data, TMTB achieves the lowest MAEs of 67.0 on JHU-Crowd++, 96.3 on UCF-QNRF, 72.4 on ShanghaiTech A, and 10.6 on ShanghaiTech B, demonstrating the best performance. The result on JHU-Crowd++ not only surpasses the performance of SOTA methods with the same amount of labeled data but also demonstrates that TMTB remains superior even at the $10\\%$ label setting, highlighting its efficient label utilization and exceptional accuracy. With a $10\\%$ labeled data setting, TMTB continues to outstand, obtaining a leading MAE of 66.3 on JHU-Crowd++, 91.7 on UCF-QNRF, and 65.7 on ShanghaiTech A, further demonstrating its generalization ability and effectiveness. At $40\\%$ labeled data, TMTB achieves competitive results, matching the lowest MAE of 60.0 on JHU-Crowd++, and decreasing the MAE to a lowest of 7.5 on ShanghaiTech B, maintaining strong performance across the other datasets. These results underscore the capability of TMTB to achieve low error with minimal labeled data, offering substantial improvements over SOTA methods.", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0c86f0d597292c4619c9564ac79e547a1b15f18d63e6628eb109ef2b10d14f18.jpg", + "table_caption": [ + "Table 2. An study of common various augmentation strategies on ShanghaiTech-A dataset using $10\\%$ labeled data." + ], + "table_footnote": [], + "table_body": "
AugmentationMAE ↓RMSE ↓
Mixup [66]285.16431.83
Mixup [66] (background only)72.00116.85
CutMix [63]312.60462.33
BCP [2]70.99121.83
TMTB (Ours)65.74110.42
", + "bbox": [ + 566, + 508, + 857, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Augmentation Strategies. In order to investigate the effectiveness of the proposed inpainting augmentation strategy, we compare it with some off-the-shelf augmentation strategies, on the ShanghaiTech A dataset with $10\\%$ labeled data, including Mixup [66], CutMix [63], and BCP [2]. The results are shown in Table 2.", + "bbox": [ + 511, + 626, + 906, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mixup and CutMix both degrade the performance when applied to the whole image. The main reason is that the integrity of the foreground and background regions is destroyed, making the augmented images non-informative and hard to learn. These results demonstrate the importance of the integrity of the foreground and background regions in the crowd counting task. We then apply Mixup to the background region only, and the performance improves. This approach is somewhat similar to the inpainting augmentation, which enhances the background region while keeping the foreground region intact. However, the background regions are sampled from the domain of the original dataset, which", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24446", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c801b6246fbf41b34d67c3fbbe0f5670acd545769488abafe7e38a4c296730ec.jpg", + "table_caption": [ + "Table 3. Comparisons with SOTA single domain generalization methods on Q (UCF-QNRF) and A (ShanghaiTech-A) datasets. The best results are highlighted in bold, and the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodsVenueSource Labeled PercentageA → QQ → A
MAE ↓RMSE ↓MAE ↓RMSE ↓
IBN [36]ECCV'18100%280.2561.0105.9174.6
SW [37]ICCV'19100%285.5431.0102.4168.8
ISW [6]CVPR'21100%215.6399.783.4136.0
DG-MAN [32]ICCV'21100%129..1238.2--
DCCUS [9]AAAI'23100%119.4216.667.4112.8
MPCount [38]CVPR'24100%115.7199.865.5110.1
TMTBOurs40%112.5190.662.4103.5
", + "bbox": [ + 91, + 155, + 483, + 268 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d3c04cbac2af279a542d609f1152b050d4c8bb57c87f9be412aaff64780e7116.jpg", + "table_caption": [ + "Table 4. Comparisons with SOTA domain adaptation methods on Q (UCF-QNRF) and A (ShanghaiTech-A) datasets. The best results are highlighted in bold, and the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodsVenueSource Labeled PercentageTarget Domain VisibleA → QQ → A
MAE ↓RMSE ↓MAE ↓RMSE ↓
RBT [28]ACM MM'20100%✓175.0294.8--
C2MoT [57]ACM MM'21100%✓125.7218.3--
FGFD [70]ACM MM'22100%✓124.1242.070.2118.4
DAOT [71]ACM MM'23100%✓113.9215.667.0128.4
FSIM [72]T-MM'23100%✓105.3191.166.8111.5
TMTBOurs40%✗112.5190.662.4103.5
", + "bbox": [ + 91, + 358, + 483, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "may not be sufficiently diverse. BCP is a recent birectical augmentation strategy designed for medical image segmentation. We employ it in the crowd counting task and find it can also improve the performance. However, the performance is not competitive enough when compared to SOTA methods.", + "bbox": [ + 89, + 493, + 483, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Domain Generalization & Domain Adaptation. We compare our method with state-of-the-art single domain generalization methods and obtain remarkable results in Table 3. The results show that our method outperforms the SOTA methods on both UCF-QNRF $\\rightarrow$ ShanghaiTech-A and ShanghaiTech-A $\\rightarrow$ UCF-QNRF tasks, utilizing only $40\\%$ labeled data on the source domain instead of $100\\%$ . Note that IBN, SW, and ISW are originally designed for image classification or segmentation, and are adapted to crowd counting and reported by MPCount [38]. The results demonstrate the effectiveness of VSSM and inpainting augmentation in our method.", + "bbox": [ + 89, + 590, + 482, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The study on domain adaptation is also conducted, and the results are shown in Table 4. Domain adaptation setting is easier than domain generalization, as the target domain is visible. It is surprising that our method achieves competitive results compared to SOTA methods, even though we only have $40\\%$ labeled data available in the source domain and no access to the target domain. The comparisons again provide evidence of the generalization ability of TMTB.", + "bbox": [ + 89, + 779, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5415e1d88a501949bed81408c9b315bc635450522a384c925de36d3783c5a286.jpg", + "table_caption": [ + "Table 5. An ablation study on the effectiveness of various components on UCF-QNRF dataset using $5\\%$ labeled data." + ], + "table_footnote": [], + "table_body": "
VSSMLsLuLinpMAE ↓RMSE ↓
XLsereg115.35188.19
✓Lsereg110.41185.52
✓Lsereg + Lcsls105.88183.87
✓Lsereg + Lcsls✓103.58178.97
✓Lsereg + Lcsls✓ωl97.21166.06
✓Lsereg + Lcsls✓ωlt96.33163.42
", + "bbox": [ + 575, + 126, + 849, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 246, + 666, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Components. We conduct an ablation study to analyze the effectiveness of various components in our method on the UCF-QNRF dataset with $5\\%$ labeled data. The results are shown in Table 5. The VSSM is the core component. Compared to OT-M [24], the VSSM with supervised loss $\\mathcal{L}^{\\mathrm{s}}$ improves the performance by decreasing the MAE by $10.57\\%$ . After adding the semi-supervised loss $\\mathcal{L}^{\\mathrm{u}}$ , the performance is further improved, showing the effectiveness of our semi-supervised learning strategy. Next, we introduce our novel inpainting augmentation strategy and its mask of inconsistency level $M_l^{\\mathrm{uncon}}$ but with fixed weights $\\omega_{l}$ . The performance is further improved, and the MAE of 97.21 surpasses the SOTA method MRC-Crowd [40] at 101.4. Finally, we employ time-based weights $\\omega_l^t$ for the inpainting loss, and the performance is significantly improved, achieving the lowest MAE of 96.33 and RMSE of 163.42.", + "bbox": [ + 511, + 268, + 906, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Weight Decay Speed. We study the impact of the weight decay speed $T^{\\mathrm{inpw}}$ on the JHU-Crowd++ dataset with 5% labeled data. The detailed influence of $T^{\\mathrm{inpw}}$ on the MAE and RMSE is shown in supplementary material. We train and test TMTB with different $T^{\\mathrm{inpw}}$ values ranging from 80 to 120. The results show that both MAE and RMSE decrease as $T^{\\mathrm{inpw}}$ increases from 80 to 100. However, when $T^{\\mathrm{inpw}}$ is greater than 100, the performance begins to degrade. Therefore, we set $T^{\\mathrm{inpw}}$ to 100 in experiments.", + "bbox": [ + 511, + 510, + 908, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 659, + 633, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the literature of semi-supervised crowd counting, a number of methods based on pseudo-labeling have been explored. Despite their success, these methods still struggle to achieve satisfactory performance in extremely crowded, low-light, and adverse weather scenarios. In this work, we reveal that the lack of suitable data augmentation methods and the weakness of capturing global context information are bottlenecks. Therefore, we design a novel inpainting augmentation method, Inpainting Augmentation that works better for crowd counting, enabling tasting more. Moreover, we introduce the Visual State Space Model to capture long-range dependencies in crowd counting, enabling tasting better. Our proposed framework, TMTB, achieves superior performance compared to existing methods.", + "bbox": [ + 511, + 684, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24447", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shuai Bai, Zhiqun He, Yu Qiao, Hanzhe Hu, Wei Wu, and Junjie Yan. Adaptive dilated network with self-correction supervision for counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5", + "[2] Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11514-11524, 2023. 2, 4, 7", + "[3] Lokesh Boominathan, Srinivas S S Kruthiventi, and R. Venkatesh Babu. Crowdnet: A deep convolutional network for dense crowd counting. In Proceedings of the 24th ACM International Conference on Multimedia, page 640-644, New York, NY, USA, 2016. Association for Computing Machinery. 1", + "[4] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. In Advances in Neural Information Processing Systems. MIT Press, 2000. 2", + "[5] Lei Chen, Xinghang Gao, Fei Chao, Xiang Chang, Chih Min Lin, Xingen Gao, Shaopeng Lin, Hongyi Zhang, and Juqiang Lin. The effectiveness of a simplified model structure for crowd counting, 2024. 2", + "[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11580-11590, 2021. 8", + "[7] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[8] Weihang Dai, Xiaomeng Li, and Kwang-Ting Cheng. Semi-supervised deep regression with uncertainty consistency and variational model ensembling via bayesian neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 37(6):7304-7313, 2023. 2", + "[9] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):561-570, 2023. 8", + "[10] Yue Duan, Zhen Zhao, Lei Qi, Lei Wang, Luping Zhou, Yinghuan Shi, and Yang Gao. Mutexmatch: Semi-supervised learning with mutex-based consistency regularization. IEEE Trans. Neural Networks Learn. Syst., 35(6): 8441-8455, 2024. 2", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3", + "[12] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling, 2024. 5, 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Tao Han, Lei Bai, Lingbo Liu, and Wanli Ouyang. Steerer: Resolving scale variations for counting and localization via selective inheritance learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21848-21859, 2023. 2", + "[14] Marcus Handte, Muhammad Umer Iqbal, Stephan Wagner, Wolfgang Apolinarski, Pedro José Marrón, Eva Maria Muñoz Navarro, Santiago Martínez, Sara Izquierdo Barthelemy, and Mario González Fernández. Crowd density estimation for public transport vehicles. In EDBT/ICDT Workshops, pages 315-322. CiteSeer, 2014. 1", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 3, 4", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[17] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 1, 2, 6", + "[18] Khawar Islam, Muhammad Zaigham Zaheer, Arif Mahmood, and Karthik Nandakumar. Diffusemix: Label-preserving data augmentation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27621-27630, 2024. 3", + "[19] Neal Jean, Sang Michael Xie, and Stefano Ermon. Semisupervised deep kernel learning: Regression with unlabeled data by minimizing predictive variance. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 2", + "[20] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[21] Chen Li, Xiaoling Hu, Shahira Abousamra, and Chao Chen. Calibrating uncertainty for semi-supervised crowd counting. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16685-16695, 2023. 1, 2, 7", + "[22] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6", + "[23] Hui Lin, Zhiheng Ma, Xiaopeng Hong, Yaowei Wang, and Zhou Su. Semi-supervised crowd counting via density agency. In Proceedings of the 30th ACM International Conference on Multimedia, page 1416-1426, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 5, 6, 7", + "[24] Wei Lin and Antoni B. Chan. Optimal transport minimization: Crowd localization on density maps for semisupervised counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21663-21673, 2023. 1, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24448", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Weizhe Liu, Mathieu Salzmann, and Pascal Fua. Context-aware crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[26] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Leveraging unlabeled data for crowd counting by learning to rank. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7", + "[27] Yan Liu, Lingqiao Liu, Peng Wang, Pingping Zhang, and Yinjie Lei. Semi-supervised crowd counting via self-training on surrogate tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 242-259. Springer, 2020. 2", + "[28] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, page 129-137, New York, NY, USA, 2020. Association for Computing Machinery. 8", + "[29] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. VMamba: Visual state space model. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 5, 1", + "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6", + "[31] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 5", + "[32] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6630-6638, 2021. 8", + "[33] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15549-15559, 2021. 7", + "[34] Greg Olmschenk., Hao Tang., and Zhigang Zhu. Improving dense crowd counting convolutional neural networks using inverse k-nearest neighbor maps and multiscale upsampling. In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 5: VISAPP, pages 185-195. INSTICC, SciTePress, 2020. 5", + "[35] OpenAI. Gpt-4o system card, 2024. 1", + "[36] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 8", + "[37] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 8", + "[38] Zhuoxuan Peng and S.-H. Gary Chan. Single domain generalization for crowd counting. In Proceedings of the" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28025-28034, 2024. 2, 8", + "[39] Yifei Qian, Liangfei Zhang, Xiaopeng Hong, Carl Donovan, and Ognjen Arandjelovic. Segmentation assisted u-shaped multi-scale transformer for crowd counting. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 5", + "[40] Yifei Qian, Xiaopeng Hong, Zhongliang Guo, Ognjen Arandjelovic, and Carl R. Donovan. Semi-supervised crowd counting with contextual modeling: Facilitating holistic understanding of crowd scenes. IEEE Transactions on Circuits and Systems for Video Technology, 34(9):8230-8241, 2024. 1, 2, 5, 6, 7, 8", + "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3", + "[42] Yasiru Ranasinghe, Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M. Patel. Crowd-diff: Multi-hypothesis crowd density estimation using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12809-12819, 2024. 2", + "[43] Viresh Ranjan, Hieu Le, and Minh Hoai. Iterative crowd counting. In Proceedings of the European conference on computer vision (ECCV), pages 270-285, 2018. 1", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 4", + "[45] Miaojing Shi, Zhaohui Yang, Chao Xu, and Qijun Chen. Revisiting perspective information for efficient crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[46] Patrice Y. Simard, Yann A. LeCun, John S. Denker, and Bernard Victorri. Transformation Invariance in Pattern Recognition - Tangent Distance and Tangent Propagation, pages 235-269. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 2", + "[47] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2594-2609, 2022. 2, 6", + "[48] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In Advances in Neural Information Processing Systems, pages 596-608. Curran Associates, Inc., 2020. 2", + "[49] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3365-3374, 2021. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24449", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Hakki Soy. Edge ai-based crowd counting application for public transport stops. In Convergence of Deep Learning and Internet of Things: Computing and Technology, pages 182-205. IGI Global, 2023. 1", + "[51] Aijun Sun, Mao Liu, and Jianfeng Li. Real-time crowd massing risk supervision system based on massing crowd counting in public venue. In 2009 International Symposium on Computer Network and Multimedia Technology, pages 1-7, 2009. 1", + "[52] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 2, 7", + "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2", + "[54] Jia Wan and Antoni Chan. Adaptive density map generation for crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5", + "[55] Changan Wang, Qingyu Song, Boshen Zhang, Yabiao Wang, Ying Tai, Xuyi Hu, Chengjie Wang, Jilin Li, Jiayi Ma, and Yang Wu. Uniformity in heterogeneity: Diving deep into count interval partition for crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3234-3242, 2021. 4, 5, 6", + "[56] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4248-4257, 2022. 2", + "[57] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, page 658-666, New York, NY, USA, 2021. Association for Computing Machinery. 8", + "[58] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9653-9663, 2022. 6", + "[59] Lihe Yang, Xiaogang Xu, Bingyi Kang, Yinghuan Shi, and Hengshuang Zhao. Freemask: Synthetic images with dense annotations make stronger segmentation models. In Advances in Neural Information Processing Systems, pages 18659-18675. Curran Associates, Inc., 2023. 3", + "[60] Yiben Yang, Chaitanya Malaviya, Jared Fernandez, Swabha Swayamdipta, Ronan Le Bras, Ji-Ping Wang, Chandra Bhagavatula, Yejin Choi, and Doug Downey. Generative data augmentation for commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1008-1025, Online, 2020. Association for Computational Linguistics. 3", + "[61] Jun Yi, Yiran Pang, Wei Zhou, Meng Zhao, and Fujian Zheng. A perspective-embedded scale-selection network" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for crowd counting in public transportation. IEEE Transactions on Intelligent Transportation Systems, 25(5):3420-3432, 2024. 1", + "[62] Yingda Yin, Yingcheng Cai, He Wang, and Baoquan Chen. Fishermatch: Semi-supervised rotation regression via entropy-based filtering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11164-11173, 2022. 2", + "[63] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features, 2019. 2, 3, 4, 7", + "[64] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In Advances in Neural Information Processing Systems, pages 18408-18419. Curran Associates, Inc., 2021. 2", + "[65] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1", + "[66] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 2, 3, 4, 7", + "[67] Shimao Zhang, Changjiang Gao, Wenhao Zhu, Jiajun Chen, Xin Huang, Xue Han, Junlan Feng, Chao Deng, and Shujian Huang. Getting more from less: Large language models are good spontaneous multilingual learners. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8037-8051, Miami, Florida, USA, 2024. Association for Computational Linguistics. 3", + "[68] Shiwei Zhang, Wei Ke, Shuai Liu, Xiaopeng Hong, and Tong Zhang. Boosting semi-supervised crowd counting with scale-based active learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8681-8690, New York, NY, USA, 2024. Association for Computing Machinery. 1, 2, 7", + "[69] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6", + "[70] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, page 5659-5668, New York, NY, USA, 2022. Association for Computing Machinery. 8", + "[71] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. In Proceedings of the 31st ACM International Conference on Multimedia, page 4319–4329, New York, NY, USA, 2023. Association for Computing Machinery. 8" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24450", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 26:3842-3855, 2024. 8", + "[73] Pengfei Zhu, Jingqing Li, Bing Cao, and Qinghua Hu. Multitask credible pseudo-label learning for semi-supervised crowd counting. IEEE Transactions on Neural Networks and Learning Systems, 35(8):10394-10406, 2024. 2, 7" + ], + "bbox": [ + 91, + 90, + 480, + 204 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "24451", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_model.json b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..79269cdeeb101a35f6e4391125595282bca8385d --- /dev/null +++ b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_model.json @@ -0,0 +1,2809 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.79, + 0.177 + ], + "angle": 0, + "content": "Taste More, Taste Better: Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.203, + 0.889, + 0.258 + ], + "angle": 0, + "content": "Maochen Yang\\(^{1}\\) Zekun Li\\(^{1}\\) Jian Zhang\\(^{1}\\) Lei Qi\\(^{3}\\) Yinghuan Shi\\(^{1,2,*}\\) \n\\(^{1}\\)Nanjing University, China \\(^{2}\\)Suzhou Laboratory, China \\(^{3}\\)Southeast University, China \n{yangmaochen, lizekun, zhangjian7369}@smail.nju.edu.cn qilei@seu.edu.cn, syh@nju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.485, + 0.671 + ], + "angle": 0, + "content": "Semi-supervised crowd counting is crucial for addressing the high annotation costs of densely populated scenes. Although several methods based on pseudo-labeling have been proposed, it remains challenging to effectively and accurately utilize unlabeled data. In this paper, we propose a novel framework called Taste More Taste Better (TMTB), which emphasizes both data and model aspects. Firstly, we explore a data augmentation technique well-suited for the crowd counting task. By inpainting the background regions, this technique can effectively enhance data diversity while preserving the fidelity of the entire scenes. Secondly, we introduce the Visual State Space Model as backbone to capture the global context information from crowd scenes, which is crucial for extremely crowded, low-light, and adverse weather scenarios. In addition to the traditional regression head for exact prediction, we employ an Anti-Noise classification head to provide less exact but more accurate supervision, since the regression head is sensitive to noise in manual annotations. We conduct extensive experiments on four benchmark datasets and show that our method outperforms state-of-the-art methods by a large margin. Code is publicly available on https://github.com/syhien/taste_more_tastebetter." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.699, + 0.222, + 0.714 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.484, + 0.786 + ], + "angle": 0, + "content": "Crowd counting is a fundamental task in computer vision that involves estimating the number of people in densely populated scenes. This task is of great importance in various public safety scenarios, including transport manage-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.792, + 0.484, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author: Yinghuan Shi. This work was supported by National Science and Technology Major Project (2023ZD0120700), NSFC Project (62222604, 62206052, 624B2063), China Postdoctoral Science Foundation (2024M750424), Fundamental Research Funds for the Central Universities (020214380120, 020214380128), State Key Laboratory Fund (ZZKT2024A14), Postdoctoral Fellowship Program of CPSF (GZC20240252), Jiangsu Funding Program for Excellent Postdoctoral Talent (2024ZB242), and Jiangsu Science and Technology Major Project (BG2024031)." + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.29, + 0.65, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.345, + 0.641, + 0.359 + ], + "angle": 0, + "content": "Mixup (A + background of B)" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.291, + 0.761, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.344, + 0.756, + 0.359 + ], + "angle": 0, + "content": "Mixup (background only) density map" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.291, + 0.871, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.344, + 0.866, + 0.353 + ], + "angle": 0, + "content": "Original A density map" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.359, + 0.65, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.414, + 0.625, + 0.422 + ], + "angle": 0, + "content": "Mixup (A + B)" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.359, + 0.761, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.414, + 0.745, + 0.422 + ], + "angle": 0, + "content": "Mixup density map" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.359, + 0.871, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.414, + 0.867, + 0.422 + ], + "angle": 0, + "content": "Original B density map" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.429, + 0.65, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.483, + 0.622, + 0.491 + ], + "angle": 0, + "content": "Inpainted A" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.429, + 0.761, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.483, + 0.757, + 0.491 + ], + "angle": 0, + "content": "Inpainted A density map" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.498, + 0.65, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.552, + 0.622, + 0.56 + ], + "angle": 0, + "content": "Inpainted B" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.498, + 0.761, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.552, + 0.757, + 0.56 + ], + "angle": 0, + "content": "Inpainted B density map" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.431, + 0.871, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.664 + ], + "angle": 0, + "content": "Figure 1. Comparisons of augmentations. As shown in the red bounding boxes, both types of Mixup generate disappointing density maps. Mixup is supposed to retain crowd information from two images, but actually it destroys the spatial structure. As shown in green bounding boxes, our inpainting augmentation is better suited for the crowd counting task, showing an impressive reduction in Mean Absolute Error (MAE)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "ment [14, 50, 61] and disaster prevention [51]. The substantial demand has driven the advancement of counting methods [3, 25, 43, 45, 65]. However, a significant cost of manual annotation is unavoidable for these fully-supervised deep methods to achieve satisfactory performance. For example, the UCF-QNRF dataset with 1,535 images and 1,251,642 annotations amounts to approximately 2,000 human-hours [17]. Compared to the extremely expensive labeled data with annotations, acquiring unlabeled crowd images is much more affordable. Consequently, there is a growing interest in leveraging unlabeled data through semi-supervised crowd counting (SSCC) methods. The recent mainstream works on SSCC [21, 24, 40, 68] mostly employ the Mean Teacher [52] framework, where the teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24440" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "model generates pseudo-labels for unlabeled data to guide the student model. The key to achieving superior performance with SSCC methods on a given set of images, where only a small portion is annotated, lies in how we handle the unlabeled samples and how we design an appropriate model to fully exploit the information within them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.397 + ], + "angle": 0, + "content": "It has been observed that introducing similar but different samples to the training data can enlarge the support of the training distribution and thus enhance the robustness and generalizability of the model [4, 46]. In the literature, numerous data augmentation techniques have demonstrated their effectiveness in generating additional virtual samples to support the training process in both fully-supervised and semi-supervised settings [2, 63, 66]. However, we find that these off-the-shelf techniques cannot suit the crowd counting task, as they may destroy the spatial structure of crowd scenes, leading to distorted density maps, as shown in Figure 1. Therefore, a well-suited data augmentation technique specifically designed for the crowd counting task is crucial for the model to taste more from the unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.642 + ], + "angle": 0, + "content": "With the help of appropriate data augmentation, the diversity of unlabeled data during training can be significantly improved. However, this simultaneously brings new challenges for the teacher model to produce correct supervision signals. From the perspective of the model itself, we find that the Convolutional Neural Network (CNN) architecture commonly used in previous studies can be suboptimal for the crowd counting task, as it tends to overfit to local details and fails to capture global context information, which is crucial for extremely crowded, low-light, and adverse weather scenarios. From another perspective, it is also important to generate more accurate pseudo-labels while filtering out low-quality parts. Considering these two points, we explore a modern backbone network to replace CNNs and enhance it with a holistic strategy for producing and leveraging pseudo-labels, which helps to taste data better." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a novel semi-supervised crowd counting method, named Taste More Taste Better (TMTB). We design a novel data augmentation technique, Inpainting Augmentation, which enhances the diversity of the dataset while keeping the integrity of the foreground regions. Observing that some regions in inpainted images are noisy and harmful, to better utilize the inpainted images, we employ a Visual State Space Model (VSSM) with an Anti-Noise classification branch for assistance, which provides inexact but accurate supervision. The proposed model generates more accurate density maps than state-of-the-art methods. Our method TMTB achieves significant improvements on several benchmark datasets, including ShanghaiTech Part A & B [69], UCF-QNRF [17], and JHU-Crowd++ [47]. We reduce the Mean Absolute Error of JHU-Crowd with \\(5\\%\\) data labeled to 67.0, marking the first time it has been lower than 70.0. The method also presents strong generalization ability" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.7, + 0.106 + ], + "angle": 0, + "content": "in cross-dataset evaluations." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.107, + 0.874, + 0.121 + ], + "angle": 0, + "content": "The main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.228 + ], + "angle": 0, + "content": "- We observed that the lack of suitable data augmentation methods and the weakness of capturing global context information are bottlenecks for the crowd counting task. The off-the-shelf data methods may destroy the spatial structure of crowd scenes, which is crucial for the regression of density maps. CNNs tend to overfit local details and struggle to capture global context information." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.229, + 0.905, + 0.318 + ], + "angle": 0, + "content": "- We propose a novel semi-supervised crowd counting method. A data augmentation technique is designed to enhance diversity while keeping the integrity of the foreground regions. A VSSM with an Anti-Noise classification branch for assistance is employed, which provides inexact but accurate supervision, to better utilize the data." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.319, + 0.905, + 0.468 + ], + "angle": 0, + "content": "- The proposed method remarkably outperforms state-of-the-art methods across four benchmark datasets and three label settings. It achieves a groundbreaking reduction in the Mean Absolute Error on the JHU-Crowd dataset with only \\(5\\%\\) of the data labeled, lowering it to 67.0, a decrease of \\(12.4\\%\\). Furthermore, the method demonstrates exceptional generalization capabilities in cross-dataset Domain Generalization settings, even surpassing the performance of fully-supervised approaches, setting a new standard for excellence." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.662, + 0.499 + ], + "angle": 0, + "content": "2. Releted Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.583 + ], + "angle": 0, + "content": "Semi-supervised Crowd Counting. Fully-supervised crowd counting methods[5, 13, 17, 31, 38, 42, 49, 69] maintain the best performance in crowd counting tasks. However, fully-supervised methods require a large quantity of labeled data, which is expensive to collect and annotate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Therefore, semi-supervised crowd counting methods have been proposed to address the scarcity of labeled data. The Mean Teacher [52], a successful semi-supervised learning framework that has been widely used in classification [10, 48, 64], segmentation [2, 56] and regression [8, 19, 62], is the most popular method in semi-supervised crowd counting. L2R [26] leveraged unlabeled data through a learning-to-rank framework. IRAST [27] introduced surrogate tasks, self-learning inter-related binary segmentation tasks, to improve performance. Moreover, MTCP [73] applied multiple auxiliary tasks into representation learning, which also benefit performance. DA-Count [23] employed a Transformer [53] to refine foreground features. CU [21] investigated a supervised uncertainty estimation strategy to better utilize unlabeled data. SALCrowd [68] proposed a method to select unlabeled data based on intra-scale uncertainty and inter-scale inconsistency metrics. MRC-Count [40] incorporated a fine-grained density classification task, proving the effectiveness of the classification task in crowd counting. Most previous approaches [21, 26, 27, 40, 68] are based on CNNs or par" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24441" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.075, + 0.885, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.331, + 0.908, + 0.415 + ], + "angle": 0, + "content": "Figure 2. The overall framework of our method TMTB. TMTB contains the Mean Teacher framework for semi-supervised learning, VSSM as the backbone, a classification branch for predicting masks, and an inpainter for inpainting augmentation. The inpainting process is conducted periodically, which generates or updates the inpainted images. For filtering out unreliable regions, the teacher model generates the weighted mask based on the inconsistency level of the inpainted image, which is applied in \\(\\mathcal{L}^{\\mathrm{inp}}\\). The teacher model generates pseudolabels for unlabeled data (inpainted images included), and the student model predicts on labeled data and strongly augmented unlabeled data (inpainted images included)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.434, + 0.482, + 0.496 + ], + "angle": 0, + "content": "tially [23] based on Transformers, which are weak in modeling long-range dependencies. When facing extremely dense crowd scenes, the performance of these methods is not satisfactory." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.501, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Generative Data Augmentation. Deep learning models tend to overfit when the quantity of data is limited, which harms the generalization ability of models. Therefore, data augmentation was proposed to address this issue. For image-based tasks, basic data augmentation methods include random cropping, flipping, rotation, scaling, color jittering, and noise adding. Some advanced data augmentation methods, such as CutMix [63], Mixup [66], and AutoAugment [7], have been proposed. However, these methods are not suitable for crowd counting tasks. Both the density-based and point-based crowd counting methods are sensitive to the spatial structure of the crowd, and the above methods may destroy the spatial structure of the crowd. The damage to the integrity of the crowd will lead to a decrease in model performance. We study this in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Generative Data Augmentation [11, 15, 16, 20, 41] is the process of creating new synthetic data points that can be added to existing datasets. This unique approach is widely used to improve the performance of models by increasing the quantity and diversity of training data while maintaining the quality of the data. In [67], queries-only data was used to improve the multilingual task, showing the effectiveness of generative data augmentation. FreeMask [59] resorted to synthetic images from generative models to ease the burden of both data collection and annotation procedures, and achieves comparable performance to real-data-based meth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.616 + ], + "angle": 0, + "content": "ods. G-DAUG\\(^{\\text{C}}\\) [60] generated synthetic examples using pretrained language models, and selects the most informative and diverse set of examples for data augmentation, to achieve more accurate and robust learning in low-resource setting. DiffusionMix [18] employed an inpaint-and-mix strategy to generate label-preserved data for classification task. However, it requires pre-defined binary masks to concatenate the original and generated images, which is not available in crowd counting since the position and shape of the crowd are unpredictable. Compared to existing works, our contribution is to fill the gap in effective augmentation methods within the field of pixel-level predition tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.632, + 0.605, + 0.646 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.734 + ], + "angle": 0, + "content": "We propose a semi-supervised crowd counting method that enhances the dataset using a diffusion-based model's inpainting technique and utilizes a VSSM (Visual State Space Model) with Anti-Noise classification branch to improve data diversity and model performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The overall framework of our method, TMTB, is shown in Figure 2. Our method is built on the Mean Teacher framework, which is widely employed in semi-supervised learning. The teacher model shares the same architecture as the student model, and its weights are updated using the student model's weights through the Exponential Moving Average (EMA) strategy. VMamba [29], a novel VSSM (Visual State Space Model), serves as the backbone, replacing the CNN and Transformer. The image features extracted from VMamba are then fed into two independent branches: the counting branch and the classification branch, for pre" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24442" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "dictions. The counting branch predicts a density map of the image, while the classification branch predicts the indices of pre-defined count interval bins. We build a module, Inpainter, designed to periodically applying background inpainting augmentation to the training set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.228 + ], + "angle": 0, + "content": "We first describe our novel inpainting augmentation in Section 3.1, then we discuss our network architecture in Section 3.2, and finally, we present our loss functions in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.241, + 0.321, + 0.258 + ], + "angle": 0, + "content": "3.1. Inpainting Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Unlike classification or regression tasks, the ground truth density maps are generated by point-level annotations. The state-of-the-art data augmentation techniques, Mixup [66], CutMix [63], BCP [2], etc., are not applicable to crowd counting, since the augmentation may cause damage to ground truth density maps. We introduce the inpainting technique of diffusion-based models [15, 44] as a novel data augmentation method with the filtering capability of unreliable regions, which generates new samples, thereby expanding the diversity of dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.535 + ], + "angle": 0, + "content": "For the inpainting mask \\( M^{\\mathrm{inp}} \\), the count interval bins predicted by the classification branch are used to separate the foreground and the background of image \\( x \\). We generate the positive text prompt randomly to achieve more diverse unlabeled dataset, but the negative text prompt is kept the same for all samples. The inpainting process is conducted periodically during the training process using a preset time interval \\( T^{\\mathrm{inp}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.547, + 0.482, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} M ^ {\\text {i n p}} = \\mathbf {0} (\\arg \\max (\\hat {p})) , \\\\ x ^ {\\text {i n p}} = \\mathcal {S D} (x, M ^ {\\text {i n p}}, r ^ {\\text {p o s}}, r ^ {\\text {n e g}}), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.689 + ], + "angle": 0, + "content": "where \\(\\hat{p}_i\\) is the predicted probability distribution from the classification branch, \\(\\mathbf{0}(\\cdot)\\) is the indicator function, \\(x^{\\mathrm{inp}}\\) is the new image with the background inpainted, \\(\\mathcal{SD}(\\cdot ,\\cdot ,\\cdot ,\\cdot)\\) represents Stable Diffusion [44], which is a mature, excellent, and cost-effective diffusion model, and \\(r^{\\mathrm{pos}}\\) and \\(r^{\\mathrm{neg}}\\) are the positive and negative text prompts, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Figure 3 shows examples of both good and poor inpainting samples. All new samples are considered unlabeled, as the inpainting process may add additional humans for better coherence. For well-inpainted samples, the foreground regions are well preserved, while the background regions are inpainted with different patterns, thus introducing diversity to the dataset. Even though the inpainting process is not perfect, the model can still benefit from the augmented dataset. However, failures may sometimes occur, potentially leading to the model learning from noisy inputs. The heads of humans and other features in the unreliable regions of poorly-inpainted images are fragmented and indistinct, rendering them difficult to recognize. It is necessary to filter out these unreliable regions. The EMA model calculates" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.092, + 0.614, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.094, + 0.699, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.139, + 0.614, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.139, + 0.699, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.185, + 0.614, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.185, + 0.699, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.23, + 0.614, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.274, + 0.58, + 0.282 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.23, + 0.699, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.274, + 0.666, + 0.282 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.093, + 0.798, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.139, + 0.798, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.185, + 0.798, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.23, + 0.798, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.274, + 0.765, + 0.282 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.093, + 0.884, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.139, + 0.884, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.184, + 0.884, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.23, + 0.884, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.274, + 0.851, + 0.281 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.287, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 3. Inpainting samples. The left part shows well-inpainted couples, while the right part shows poorly-inpainted couples. Column (a) and (c) are the original images, while column (b) and (d) are the inpainted images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.906, + 0.431 + ], + "angle": 0, + "content": "the inconsistency level of the inpainted image by the difference in the classification branch between the weakly augmented version and the strongly augmented version. The inconsistency level and the weighted mask are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.443, + 0.905, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} M _ {l} ^ {\\text {i n c o n}} = \\mathbf {0} \\left(\\left| \\left| p ^ {\\hat {s}} - p ^ {\\hat {w}} \\right| - l \\right|\\right), \\\\ \\omega_ {l} ^ {t} = \\operatorname {s o f t m a x} \\left(e ^ {- l \\cdot t / T ^ {\\mathrm {i n p w}}}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.492, + 0.772, + 0.533 + ], + "angle": 0, + "content": "\\[\nM ^ {\\text {i n c o n}} = \\sum_ {l = 0} ^ {L} M _ {l} ^ {\\text {i n c o n}} \\cdot \\omega_ {l} ^ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.716 + ], + "angle": 0, + "content": "where \\(\\hat{p^s}\\) and \\(\\hat{p^w}\\) are the predicted probability distributions of the classification branch of EMA model in the strong and weak augmentation version, respectively. \\(M_{l}^{\\mathrm{incon}}\\) is the mask of inconsistency level \\(l\\), \\(M^{\\mathrm{incon}}\\) is the final weighted mask, \\(\\omega_l^t\\) is the weight of inconsistency level \\(l\\) at time \\(t\\), \\(T^{\\mathrm{inpw}}\\) is the parameter controlling the weight decay speed, and \\(L\\) is the number of inconsistency levels. The \\(L\\) depends on the predefined count interval bins, and since we followed the interval bins settings in [55], the \\(L\\) is set to 2 for the balance of diversity and reliability. The weighted mask is then used to adjust the loss associated with the inpainted image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.859, + 0.742 + ], + "angle": 0, + "content": "3.2. VSSM with Classification for Anti-Noise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Visual State Space Models (VSSMs). Existing crowd counting methods often encounter challenges such as congestion, blur, and occlusion in intricate scenes. CNN-based approaches tend to overfit to local details, while Transformer-based methods excel in global modeling but face heavy computational burdens due to their quadratic complexity in terms of image sizes. These limitations motivate us to introduce the VSSM, which captures comprehensive long-range information with linear time complexity, for the crowd counting task." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24443" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.155, + 0.078, + 0.325, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.078, + 0.469, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.079, + 0.653, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.079, + 0.845, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.162, + 0.326, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.162, + 0.468, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.162, + 0.651, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.162, + 0.844, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.247, + 0.325, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.247, + 0.468, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.247, + 0.651, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.247, + 0.844, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.339, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Visualizations of the predictions on the test set of the JHU-Crowd++ dataset. The first row shows the input images. The second row displays the predictions of the SOTA method MRC-Crowd [40]. The third row presents the predictions of our method TMTB. All models are trained with \\(5\\%\\) labeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.483, + 0.454 + ], + "angle": 0, + "content": "State Space Models (SSMs) map the input stimulation \\( u(t) \\) to the output response \\( y(t) \\) by means of a hidden state \\( h(t) \\). These processes are formulated with the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.462, + 0.481, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} h ^ {\\prime} (t) = A h (t) + B u (t), \\\\ y (t) = C h (t) + D u (t), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.441, + 0.523 + ], + "angle": 0, + "content": "where \\(A, B, C,\\) and \\(D\\) are the weighting parameters." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Linear time-invariant SSMs (Equation (3)) lack the ability to capture the contextual information, [12] proposed a novel parameterization method for SSMs that integrates an input-dependent selection mechanism (referred to as S6). VMamba [29] replaces the S6 module, with the newly proposed 2D-Selective-Scan (SS2D) module which concurrently achieving global receptive fields, dynamic weights, and linear complexity. By adopting complementary 1D traversal paths, SS2D enables each pixel in the image to effectively integrate information from all other pixels in different directions, thereby facilitating the establishment of global receptive fields in the 2D space and generating more accurate density map. Details about the SS2D module could be found in [29]. The features extracted by the final SS2D block are then fed into two branches: one for density map regression and the other for count interval classification, which is designed for less inexact and more accurate supervision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Exact: Regression Head. The regression head is designed to predict the density map for the input image. The density map is then summed to obtain the exact final count value. The regression head consists of up-sampling layers, convolutional layers, batch normalization layers, and ReLU activation layers. We followed MRC-Crowd [40] employing the loss function proposed in CUT [39], which significantly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.424 + ], + "angle": 0, + "content": "boosts the performance in the foreground area. It could be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.435, + 0.905, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e g}} = \\frac {1}{J} \\sum_ {j = 1} ^ {J} \\left(1 - \\operatorname {S S I M} \\left(D _ {j} (\\hat {y} \\odot M ^ {g t}\\right), \\right. \\tag {4} \\\\ \\left. D _ {j} \\left(y ^ {g t} \\odot M ^ {g t}\\right)\\right) + \\alpha \\cdot \\mathcal {L} _ {T V} \\left(\\hat {y}, y ^ {g t}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.644 + ], + "angle": 0, + "content": "where \\(\\mathrm{SSIM}(\\cdot ,\\cdot)\\) is the structural similarity index measure, and \\(D_{j}\\) denotes the average pooling operation that downsamples an image to \\(1 / 2^{j}\\) of its original size. The operation \\(\\odot\\) denotes the Hadamard product, and \\(M^{gt}\\) is the binary mask that distinguishes the dense and sparse regions based on a predefined threshold \\(\\tau\\). It can be obtained using the indicator function \\(\\mathbf{1}\\) as \\(M^{gt} = \\mathbf{1}(y^{gt} > \\tau)\\). \\(\\alpha\\) is a hyperparameter that balances the two terms. \\(J\\) and \\(\\alpha\\) are set to 3 and 0.01, respectively. \\(\\mathcal{L}_{TV}\\) is the total variation loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.907, + 0.886 + ], + "angle": 0, + "content": "Accurate: Anti-Noise Classification Head. Many previous works [1, 31, 34, 54] have identified the problem of inaccurate ground truth annotations in crowd counting datasets, which render the ground truth density maps noisy and inaccurate. In this way, the model may learn from the noisy ground truth and generate inaccurate density maps. Inspired by previous works [23, 40, 55], we employ the Anti-Noise classification task as an auxiliary task, which is more robust in predicting inexactly and accurately. The Anti-Noise classification branch is supervised by the inexact but accurate count interval bins rather than the exact but maybe inaccurate density maps. Furthermore, we use the prediction of the this branch to generate reliable masks for inpainting samples, and we discuss this in Section 3.1. The goal of the Anti-Noise classification branch is to predict the accurate count interval bins, which are predefined by [55]." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "The Anti-Noise classification head consists of an up" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24444" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "sampling layer, a convolutional layer, a ReLU activation layer and a final convolutional layer. The standard cross-entropy loss is applied. In the case of labeled data, the loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.161, + 0.483, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l s}} = \\frac {1}{N} \\sum_ {i} ^ {N} \\mathcal {H} \\left(p _ {i} ^ {g t}, \\hat {p} _ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.484, + 0.262 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}\\) is the cross-entropy loss function, \\(p_i^{gt}\\) is the ground truth probability distribution, and \\(\\hat{p}_i\\) is the predicted probability distribution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.271, + 0.18, + 0.285 + ], + "angle": 0, + "content": "3.3. Losses" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.549 + ], + "angle": 0, + "content": "In order to make full use of the labeled data \\( L = \\{(x_{i}^{l},DM_{i}^{gt})\\}_{i = 1}^{N_{l}} \\) and unlabeled data \\( U = \\{(x_{i}^{u})\\}_{i = 1}^{N_{u}} \\), we employ a common semi-supervised learning strategy widely used in crowd counting. The labeled samples are used to directly train the student model, while the unlabeled samples are used for consistency regularization. For labeled data, the supervised loss \\( \\mathcal{L}^{\\mathrm{s}} = \\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{s}} + \\mathcal{L}_{\\mathrm{cls}}^{\\mathrm{s}} \\) is a combination of the regression loss (Equation (4)) and the classification loss (Equation (5)). Given the absence of a confidence score for the predicted regression map, we choose to apply consistency regularization to all unlabeled samples, rather than developing an algorithm to assess prediction uncertainty. The EMA teacher model generates pseudo-labels based on the prediction of unlabeled data with weak augmentation. The student model makes predictions based on the strong augmentation version. The error loss between the two predictions is calculated as the consistency loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.67 + ], + "angle": 0, + "content": "Despite the common strong augmentation, we also introduce a patch-aligned random masking strategy [58] as a kind of strong augmentation. The patch-aligned random masking strategy randomly masks out some patches of various sizes in the image. The student model is required to predict based on masked images, while the teacher model is able to access the full image. The semi-supervised loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.681, + 0.483, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {u}} = \\operatorname {M A E} \\left(\\hat {y} ^ {s t}, \\hat {y} ^ {e m a}\\right) + \\operatorname {M A E} \\left(\\hat {p} ^ {s t}, \\hat {p} ^ {e m a}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.483, + 0.768 + ], + "angle": 0, + "content": "where \\(\\hat{y}^{st}\\) and \\(\\hat{y}^{ema}\\) are the predicted density maps of the student model and the teacher model, respectively. \\(\\hat{p}^{st}\\) and \\(\\hat{p}^{ema}\\) are the predicted count interval bins of the student model and the teacher model, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.829 + ], + "angle": 0, + "content": "Consistency loss is also applied to the inpainted images. However, to avoid the model learning from noisy regions with low quality, the loss is adjusted by the weighted mask of inconsistency level. The inpainting loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.839, + 0.508, + 0.871 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {i n p}} = M ^ {\\text {i n c o n}} \\times \\operatorname {M A E} \\left(\\hat {y} ^ {s t}, \\hat {y} ^ {e m a}\\right) + M ^ {\\text {i n c o n}} \\times \\operatorname {M A E} \\left(\\hat {p} ^ {s t}, \\hat {p} ^ {e m a}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( M^{\\mathrm{incon}} \\) is the weighted mask of inconsistency level mentioned in Equation (2)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.123 + ], + "angle": 0, + "content": "The final loss is a combination of the supervised, semi-supervised, and inpainting losses:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.133, + 0.905, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} ^ {\\mathrm {s}} + \\lambda_ {w} \\cdot \\mathcal {L} ^ {\\mathrm {u}} + \\lambda_ {w} \\cdot \\mathcal {L} ^ {\\text {i n p}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.185 + ], + "angle": 0, + "content": "where \\(\\lambda_w\\) is the warm-up weight for the semi-supervised loss and the inpainting loss." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.196, + 0.646, + 0.213 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.222, + 0.834, + 0.236 + ], + "angle": 0, + "content": "We evaluate our method on four public datasets:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.237, + 0.905, + 0.282 + ], + "angle": 0, + "content": "- JHU-Crowd++ [47] is a comprehensive dataset with 4,372 images and 1.51 million annotations, collected under a variety of scenarios and environmental conditions." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.283, + 0.905, + 0.356 + ], + "angle": 0, + "content": "- UCF-QNRF [17] contains 1535 images. The dataset has a greater number of high-count crowd images and annotations, and a wider variety of scenes containing the most diverse set of viewpoints, densities and lighting variations." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.358, + 0.905, + 0.388 + ], + "angle": 0, + "content": "- ShanghaiTech A & B [69] are two datasets containing nearly 1,200 images and 330,000 labeled heads." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.237, + 0.905, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.389, + 0.906, + 0.493 + ], + "angle": 0, + "content": "To achieve fair comparisons, we follow the same semi-supervised experimental protocol as previous works [23, 24, 40], including the \\(5\\%\\), \\(10\\%\\), and \\(40\\%\\) labeled data settings and the split of labeled and unlabeled data. The evaluation metrics are Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The detailed definitions of them are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.902 + ], + "angle": 0, + "content": "For some excessively high-resolution images, we scale them down, ensuring the longest side does not exceed 1,920 pixels. For the ground truth density map, we use geometry-adapted Gaussian kernels described in [22] for all datasets except the ShanghaiTech B dataset, and we fix the size of the kernel to 4. When training, the input images are randomly cropped to \\(512 \\times 512\\) pixels for all datasets except for ShanghaiTech A. The inputs from ShanghaiTech A are cropped to \\(256 \\times 256\\) pixels instead. The count interval bins, which should be predicted by the classification branch, are set to the same as those in [55]. The AdamW optimizer [30] with a learning rate of \\(10^{-5}\\) is used to optimize the model. The weight decay is set to \\(10^{-4}\\). The downsample factor is set to 8. The batch size is set to 8. In \\(5\\%\\) and \\(10\\%\\) labeled data settings, the ratio of labeled data to unlabeled data is set to \\(2:6\\). In the \\(40\\%\\) labeled data setting, the ratio is set to \\(4:4\\). As we mentioned in Section 3.3, we use the patch-aligned random masking strategy [58] as a form of strong augmentation. The masked patch size and the masked region ratio are set to \\(32 \\times 32\\) and 0.3, respectively. The ablation details can be found in [40]. The teacher model is updated by the exponential moving average with a decay rate of 0.97. The weights of all losses are set to 1.0. For \\(\\mathcal{L}^{\\mathrm{u}}\\) and \\(\\mathcal{L}^{\\mathrm{inp}}\\), we ramp the weights from 0 to 1.0 in the first 20 epochs. The max accepted inconsistency level \\(L\\) is set to 2, and the hyperparameter \\(T^{\\mathrm{inpw}}\\) is set to 100. The ablation study of \\(T^{\\mathrm{inpw}}\\) is conducted in Section 4.2. Each training" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24445" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparisons with the SOTA methods on main public datasets. The best results are highlighted in bold, and the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.128, + 0.865, + 0.451 + ], + "angle": 0, + "content": "
MethodsVenueLabeled PercentageJHU-Crowd++UCF-QNRFShanghaiTech AShanghaiTech B
MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓
MT [52]NIPS'175%101.5363.5172.4284.9104.7156.919.333.2
L2R [26]CVPR'185%101.4338.8160.1272.3103.0155.420.327.6
DACount [23]ACM MM'225%82.2294.9120.2209.385.4134.512.622.8
OT-M [24]CVPR'235%82.7304.5118.4195.483.7133.312.621.5
MRC-Crowd [40]T-CSVT'245%76.5282.7101.4171.374.8117.311.717.8
TMTBOurs5%67.0252.296.3163.472.4126.410.619.6
MT [52]NIPS'1710%90.2319.3145.5250.394.5156.115.624.5
L2R [26]CVPR'1810%87.5315.3148.9249.890.3153.515.624.4
DACount [23]ACM MM'2210%75.9282.3109.0187.274.9115.511.119.1
MTCP [73]T-NNLS'2310%88.1318.7124.7206.381.3130.514.522.3
OT-M [24]CVPR'2310%73.0280.6113.1186.780.1118.510.818.2
CU [21]ICCV'2310%74.9281.7104.0164.370.8116.69.717.7
SALCrowd [68]ACM MM'2410%69.7263.5106.7171.369.7114.59.717.5
MRC-Crowd [40]T-CSVT'2410%70.7261.393.4153.267.3106.810.318.2
TMTBOurs10%66.3265.191.7156.365.7110.49.817.2
MT [52]NIPS'1740%121.5388.9147.2247.688.2151.115.925.7
L2R [26]CVPR'1840%123.6376.1145.1256.186.5148.216.825.1
SUA [33]ICCV'2140%80.7290.8130.3226.368.5121.914.120.6
DACount [23]ACM MM'2240%65.1260.091.1153.467.5110.79.614.6
OT-M [24]CVPR'2340%72.1272.0100.6167.670.7114.58.113.1
SALCrowd [68]ACM MM'2440%62.0255.195.0167.060.797.27.912.7
MRC-Crowd [40]T-CSVT'2440%60.0227.381.1131.562.195.57.813.3
TMTBOurs40%60.0240.381.4140.660.899.07.512.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.506 + ], + "angle": 0, + "content": "process coordinates with two inpainting background processes. The inpainting period \\( T^{\\mathrm{inp}} \\) is set to 80 epochs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.515, + 0.231, + 0.53 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.598 + ], + "angle": 0, + "content": "In our main experiments, the proposed TMTB demonstrates superior performance compared to state-of-the-art methods across four public datasets with three labeled data percentages. The results are shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Notably, when utilizing only \\(5\\%\\) labeled data, TMTB achieves the lowest MAEs of 67.0 on JHU-Crowd++, 96.3 on UCF-QNRF, 72.4 on ShanghaiTech A, and 10.6 on ShanghaiTech B, demonstrating the best performance. The result on JHU-Crowd++ not only surpasses the performance of SOTA methods with the same amount of labeled data but also demonstrates that TMTB remains superior even at the \\(10\\%\\) label setting, highlighting its efficient label utilization and exceptional accuracy. With a \\(10\\%\\) labeled data setting, TMTB continues to outstand, obtaining a leading MAE of 66.3 on JHU-Crowd++, 91.7 on UCF-QNRF, and 65.7 on ShanghaiTech A, further demonstrating its generalization ability and effectiveness. At \\(40\\%\\) labeled data, TMTB achieves competitive results, matching the lowest MAE of 60.0 on JHU-Crowd++, and decreasing the MAE to a lowest of 7.5 on ShanghaiTech B, maintaining strong performance across the other datasets. These results underscore the capability of TMTB to achieve low error with minimal labeled data, offering substantial improvements over SOTA methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.472, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Table 2. An study of common various augmentation strategies on ShanghaiTech-A dataset using \\(10\\%\\) labeled data." + }, + { + "type": "table", + "bbox": [ + 0.567, + 0.51, + 0.858, + 0.6 + ], + "angle": 0, + "content": "
AugmentationMAE ↓RMSE ↓
Mixup [66]285.16431.83
Mixup [66] (background only)72.00116.85
CutMix [63]312.60462.33
BCP [2]70.99121.83
TMTB (Ours)65.74110.42
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Augmentation Strategies. In order to investigate the effectiveness of the proposed inpainting augmentation strategy, we compare it with some off-the-shelf augmentation strategies, on the ShanghaiTech A dataset with \\(10\\%\\) labeled data, including Mixup [66], CutMix [63], and BCP [2]. The results are shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Mixup and CutMix both degrade the performance when applied to the whole image. The main reason is that the integrity of the foreground and background regions is destroyed, making the augmented images non-informative and hard to learn. These results demonstrate the importance of the integrity of the foreground and background regions in the crowd counting task. We then apply Mixup to the background region only, and the performance improves. This approach is somewhat similar to the inpainting augmentation, which enhances the background region while keeping the foreground region intact. However, the background regions are sampled from the domain of the original dataset, which" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24446" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.147 + ], + "angle": 0, + "content": "Table 3. Comparisons with SOTA single domain generalization methods on Q (UCF-QNRF) and A (ShanghaiTech-A) datasets. The best results are highlighted in bold, and the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.156, + 0.485, + 0.269 + ], + "angle": 0, + "content": "
MethodsVenueSource Labeled PercentageA → QQ → A
MAE ↓RMSE ↓MAE ↓RMSE ↓
IBN [36]ECCV'18100%280.2561.0105.9174.6
SW [37]ICCV'19100%285.5431.0102.4168.8
ISW [6]CVPR'21100%215.6399.783.4136.0
DG-MAN [32]ICCV'21100%129..1238.2--
DCCUS [9]AAAI'23100%119.4216.667.4112.8
MPCount [38]CVPR'24100%115.7199.865.5110.1
TMTBOurs40%112.5190.662.4103.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.293, + 0.485, + 0.351 + ], + "angle": 0, + "content": "Table 4. Comparisons with SOTA domain adaptation methods on Q (UCF-QNRF) and A (ShanghaiTech-A) datasets. The best results are highlighted in bold, and the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.359, + 0.485, + 0.458 + ], + "angle": 0, + "content": "
MethodsVenueSource Labeled PercentageTarget Domain VisibleA → QQ → A
MAE ↓RMSE ↓MAE ↓RMSE ↓
RBT [28]ACM MM'20100%✓175.0294.8--
C2MoT [57]ACM MM'21100%✓125.7218.3--
FGFD [70]ACM MM'22100%✓124.1242.070.2118.4
DAOT [71]ACM MM'23100%✓113.9215.667.0128.4
FSIM [72]T-MM'23100%✓105.3191.166.8111.5
TMTBOurs40%✗112.5190.662.4103.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.485, + 0.586 + ], + "angle": 0, + "content": "may not be sufficiently diverse. BCP is a recent birectical augmentation strategy designed for medical image segmentation. We employ it in the crowd counting task and find it can also improve the performance. However, the performance is not competitive enough when compared to SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.483, + 0.775 + ], + "angle": 0, + "content": "Domain Generalization & Domain Adaptation. We compare our method with state-of-the-art single domain generalization methods and obtain remarkable results in Table 3. The results show that our method outperforms the SOTA methods on both UCF-QNRF \\(\\rightarrow\\) ShanghaiTech-A and ShanghaiTech-A \\(\\rightarrow\\) UCF-QNRF tasks, utilizing only \\(40\\%\\) labeled data on the source domain instead of \\(100\\%\\). Note that IBN, SW, and ISW are originally designed for image classification or segmentation, and are adapted to crowd counting and reported by MPCount [38]. The results demonstrate the effectiveness of VSSM and inpainting augmentation in our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The study on domain adaptation is also conducted, and the results are shown in Table 4. Domain adaptation setting is easier than domain generalization, as the target domain is visible. It is surprising that our method achieves competitive results compared to SOTA methods, even though we only have \\(40\\%\\) labeled data available in the source domain and no access to the target domain. The comparisons again provide evidence of the generalization ability of TMTB." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Table 5. An ablation study on the effectiveness of various components on UCF-QNRF dataset using \\(5\\%\\) labeled data." + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.127, + 0.851, + 0.226 + ], + "angle": 0, + "content": "
VSSMLsLuLinpMAE ↓RMSE ↓
XLsereg115.35188.19
✓Lsereg110.41185.52
✓Lsereg + Lcsls105.88183.87
✓Lsereg + Lcsls✓103.58178.97
✓Lsereg + Lcsls✓ωl97.21166.06
✓Lsereg + Lcsls✓ωlt96.33163.42
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.247, + 0.668, + 0.264 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Components. We conduct an ablation study to analyze the effectiveness of various components in our method on the UCF-QNRF dataset with \\(5\\%\\) labeled data. The results are shown in Table 5. The VSSM is the core component. Compared to OT-M [24], the VSSM with supervised loss \\(\\mathcal{L}^{\\mathrm{s}}\\) improves the performance by decreasing the MAE by \\(10.57\\%\\). After adding the semi-supervised loss \\(\\mathcal{L}^{\\mathrm{u}}\\), the performance is further improved, showing the effectiveness of our semi-supervised learning strategy. Next, we introduce our novel inpainting augmentation strategy and its mask of inconsistency level \\(M_l^{\\mathrm{uncon}}\\) but with fixed weights \\(\\omega_{l}\\). The performance is further improved, and the MAE of 97.21 surpasses the SOTA method MRC-Crowd [40] at 101.4. Finally, we employ time-based weights \\(\\omega_l^t\\) for the inpainting loss, and the performance is significantly improved, achieving the lowest MAE of 96.33 and RMSE of 163.42." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.511, + 0.909, + 0.648 + ], + "angle": 0, + "content": "Weight Decay Speed. We study the impact of the weight decay speed \\( T^{\\mathrm{inpw}} \\) on the JHU-Crowd++ dataset with 5% labeled data. The detailed influence of \\( T^{\\mathrm{inpw}} \\) on the MAE and RMSE is shown in supplementary material. We train and test TMTB with different \\( T^{\\mathrm{inpw}} \\) values ranging from 80 to 120. The results show that both MAE and RMSE decrease as \\( T^{\\mathrm{inpw}} \\) increases from 80 to 100. However, when \\( T^{\\mathrm{inpw}} \\) is greater than 100, the performance begins to degrade. Therefore, we set \\( T^{\\mathrm{inpw}} \\) to 100 in experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.66, + 0.634, + 0.677 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.899 + ], + "angle": 0, + "content": "In the literature of semi-supervised crowd counting, a number of methods based on pseudo-labeling have been explored. Despite their success, these methods still struggle to achieve satisfactory performance in extremely crowded, low-light, and adverse weather scenarios. In this work, we reveal that the lack of suitable data augmentation methods and the weakness of capturing global context information are bottlenecks. Therefore, we design a novel inpainting augmentation method, Inpainting Augmentation that works better for crowd counting, enabling tasting more. Moreover, we introduce the Visual State Space Model to capture long-range dependencies in crowd counting, enabling tasting better. Our proposed framework, TMTB, achieves superior performance compared to existing methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24447" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Shuai Bai, Zhiqun He, Yu Qiao, Hanzhe Hu, Wei Wu, and Junjie Yan. Adaptive dilated network with self-correction supervision for counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11514-11524, 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[3] Lokesh Boominathan, Srinivas S S Kruthiventi, and R. Venkatesh Babu. Crowdnet: A deep convolutional network for dense crowd counting. In Proceedings of the 24th ACM International Conference on Multimedia, page 640-644, New York, NY, USA, 2016. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[4] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. In Advances in Neural Information Processing Systems. MIT Press, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[5] Lei Chen, Xinghang Gao, Fei Chao, Xiang Chang, Chih Min Lin, Xingen Gao, Shaopeng Lin, Hongyi Zhang, and Juqiang Lin. The effectiveness of a simplified model structure for crowd counting, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11580-11590, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[7] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[8] Weihang Dai, Xiaomeng Li, and Kwang-Ting Cheng. Semi-supervised deep regression with uncertainty consistency and variational model ensembling via bayesian neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 37(6):7304-7313, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[9] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):561-570, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[10] Yue Duan, Zhen Zhao, Lei Qi, Lei Wang, Luping Zhou, Yinghuan Shi, and Yang Gao. Mutexmatch: Semi-supervised learning with mutex-based consistency regularization. IEEE Trans. Neural Networks Learn. Syst., 35(6): 8441-8455, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling, 2024. 5, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Tao Han, Lei Bai, Lingbo Liu, and Wanli Ouyang. Steerer: Resolving scale variations for counting and localization via selective inheritance learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21848-21859, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] Marcus Handte, Muhammad Umer Iqbal, Stephan Wagner, Wolfgang Apolinarski, Pedro José Marrón, Eva Maria Muñoz Navarro, Santiago Martínez, Sara Izquierdo Barthelemy, and Mario González Fernández. Crowd density estimation for public transport vehicles. In EDBT/ICDT Workshops, pages 315-322. CiteSeer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[17] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[18] Khawar Islam, Muhammad Zaigham Zaheer, Arif Mahmood, and Karthik Nandakumar. Diffusemix: Label-preserving data augmentation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27621-27630, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[19] Neal Jean, Sang Michael Xie, and Stefano Ermon. Semisupervised deep kernel learning: Regression with unlabeled data by minimizing predictive variance. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[20] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[21] Chen Li, Xiaoling Hu, Shahira Abousamra, and Chao Chen. Calibrating uncertainty for semi-supervised crowd counting. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16685-16695, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.677, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[22] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[23] Hui Lin, Zhiheng Ma, Xiaopeng Hong, Yaowei Wang, and Zhou Su. Semi-supervised crowd counting via density agency. In Proceedings of the 30th ACM International Conference on Multimedia, page 1416-1426, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Wei Lin and Antoni B. Chan. Optimal transport minimization: Crowd localization on density maps for semisupervised counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21663-21673, 2023. 1, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24448" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[25] Weizhe Liu, Mathieu Salzmann, and Pascal Fua. Context-aware crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[26] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Leveraging unlabeled data for crowd counting by learning to rank. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[27] Yan Liu, Lingqiao Liu, Peng Wang, Pingping Zhang, and Yinjie Lei. Semi-supervised crowd counting via self-training on surrogate tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 242-259. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[28] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, page 129-137, New York, NY, USA, 2020. Association for Computing Machinery. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[29] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. VMamba: Visual state space model. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[31] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[32] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6630-6638, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[33] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15549-15559, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.742 + ], + "angle": 0, + "content": "[34] Greg Olmschenk., Hao Tang., and Zhigang Zhu. Improving dense crowd counting convolutional neural networks using inverse k-nearest neighbor maps and multiscale upsampling. In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 5: VISAPP, pages 185-195. INSTICC, SciTePress, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.351, + 0.758 + ], + "angle": 0, + "content": "[35] OpenAI. Gpt-4o system card, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[36] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[37] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Zhuoxuan Peng and S.-H. Gary Chan. Single domain generalization for crowd counting. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28025-28034, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[39] Yifei Qian, Liangfei Zhang, Xiaopeng Hong, Carl Donovan, and Ognjen Arandjelovic. Segmentation assisted u-shaped multi-scale transformer for crowd counting. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[40] Yifei Qian, Xiaopeng Hong, Zhongliang Guo, Ognjen Arandjelovic, and Carl R. Donovan. Semi-supervised crowd counting with contextual modeling: Facilitating holistic understanding of crowd scenes. IEEE Transactions on Circuits and Systems for Video Technology, 34(9):8230-8241, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[42] Yasiru Ranasinghe, Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M. Patel. Crowd-diff: Multi-hypothesis crowd density estimation using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12809-12819, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[43] Viresh Ranjan, Hieu Le, and Minh Hoai. Iterative crowd counting. In Proceedings of the European conference on computer vision (ECCV), pages 270-285, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.907, + 0.546 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[45] Miaojing Shi, Zhaohui Yang, Chao Xu, and Qijun Chen. Revisiting perspective information for efficient crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[46] Patrice Y. Simard, Yann A. LeCun, John S. Denker, and Bernard Victorri. Transformation Invariance in Pattern Recognition - Tangent Distance and Tangent Propagation, pages 235-269. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.677, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[47] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2594-2609, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[48] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In Advances in Neural Information Processing Systems, pages 596-608. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[49] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3365-3374, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24449" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[50] Hakki Soy. Edge ai-based crowd counting application for public transport stops. In Convergence of Deep Learning and Internet of Things: Computing and Technology, pages 182-205. IGI Global, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[51] Aijun Sun, Mao Liu, and Jianfeng Li. Real-time crowd massing risk supervision system based on massing crowd counting in public venue. In 2009 International Symposium on Computer Network and Multimedia Technology, pages 1-7, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[52] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[54] Jia Wan and Antoni Chan. Adaptive density map generation for crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[55] Changan Wang, Qingyu Song, Boshen Zhang, Yabiao Wang, Ying Tai, Xuyi Hu, Chengjie Wang, Jilin Li, Jiayi Ma, and Yang Wu. Uniformity in heterogeneity: Diving deep into count interval partition for crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3234-3242, 2021. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[56] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4248-4257, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[57] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, page 658-666, New York, NY, USA, 2021. Association for Computing Machinery. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[58] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9653-9663, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[59] Lihe Yang, Xiaogang Xu, Bingyi Kang, Yinghuan Shi, and Hengshuang Zhao. Freemask: Synthetic images with dense annotations make stronger segmentation models. In Advances in Neural Information Processing Systems, pages 18659-18675. Curran Associates, Inc., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[60] Yiben Yang, Chaitanya Malaviya, Jared Fernandez, Swabha Swayamdipta, Ronan Le Bras, Ji-Ping Wang, Chandra Bhagavatula, Yejin Choi, and Doug Downey. Generative data augmentation for commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1008-1025, Online, 2020. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[61] Jun Yi, Yiran Pang, Wei Zhou, Meng Zhao, and Fujian Zheng. A perspective-embedded scale-selection network" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "for crowd counting in public transportation. IEEE Transactions on Intelligent Transportation Systems, 25(5):3420-3432, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[62] Yingda Yin, Yingcheng Cai, He Wang, and Baoquan Chen. Fishermatch: Semi-supervised rotation regression via entropy-based filtering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11164-11173, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[63] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features, 2019. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[64] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In Advances in Neural Information Processing Systems, pages 18408-18419. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[65] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[66] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.464, + 0.907, + 0.561 + ], + "angle": 0, + "content": "[67] Shimao Zhang, Changjiang Gao, Wenhao Zhu, Jiajun Chen, Xin Huang, Xue Han, Junlan Feng, Chao Deng, and Shujian Huang. Getting more from less: Large language models are good spontaneous multilingual learners. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8037-8051, Miami, Florida, USA, 2024. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[68] Shiwei Zhang, Wei Ke, Shuai Liu, Xiaopeng Hong, and Tong Zhang. Boosting semi-supervised crowd counting with scale-based active learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8681-8690, New York, NY, USA, 2024. Association for Computing Machinery. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[69] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[70] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, page 5659-5668, New York, NY, USA, 2022. Association for Computing Machinery. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[71] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. In Proceedings of the 31st ACM International Conference on Multimedia, page 4319–4329, New York, NY, USA, 2023. Association for Computing Machinery. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24450" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[72] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 26:3842-3855, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[73] Pengfei Zhu, Jingqing Li, Bing Cao, and Qinghua Hu. Multitask credible pseudo-label learning for semi-supervised crowd counting. IEEE Transactions on Neural Networks and Learning Systems, 35(8):10394-10406, 2024. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24451" + } + ] +] \ No newline at end of file diff --git a/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_origin.pdf b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..95755d5e5384788ab259ab291c57b52c0c16ae11 --- /dev/null +++ b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/97690d50-f3d9-442e-9ba4-c2196e1e5a9a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd3abb88e8c7e8d0723e8db00febf969d1d38b8901f6ecec6e5c1e7853c89235 +size 9921176 diff --git a/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/full.md b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ab9980b22d4afbf68d57443c9f2369ba606deddd --- /dev/null +++ b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/full.md @@ -0,0 +1,386 @@ +# Taste More, Taste Better: Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting + +Maochen Yang $^{1}$ Zekun Li $^{1}$ Jian Zhang $^{1}$ Lei Qi $^{3}$ Yinghuan Shi $^{1,2,*}$ $^{1}$ Nanjing University, China $^{2}$ Suzhou Laboratory, China $^{3}$ Southeast University, China +{yangmaochen, lizekun, zhangjian7369}@smail.nju.edu.cn qilei@seu.edu.cn, syh@nju.edu.cn + +# Abstract + +Semi-supervised crowd counting is crucial for addressing the high annotation costs of densely populated scenes. Although several methods based on pseudo-labeling have been proposed, it remains challenging to effectively and accurately utilize unlabeled data. In this paper, we propose a novel framework called Taste More Taste Better (TMTB), which emphasizes both data and model aspects. Firstly, we explore a data augmentation technique well-suited for the crowd counting task. By inpainting the background regions, this technique can effectively enhance data diversity while preserving the fidelity of the entire scenes. Secondly, we introduce the Visual State Space Model as backbone to capture the global context information from crowd scenes, which is crucial for extremely crowded, low-light, and adverse weather scenarios. In addition to the traditional regression head for exact prediction, we employ an Anti-Noise classification head to provide less exact but more accurate supervision, since the regression head is sensitive to noise in manual annotations. We conduct extensive experiments on four benchmark datasets and show that our method outperforms state-of-the-art methods by a large margin. Code is publicly available on https://github.com/syhien/taste_more_tastebetter. + +# 1. Introduction + +Crowd counting is a fundamental task in computer vision that involves estimating the number of people in densely populated scenes. This task is of great importance in various public safety scenarios, including transport manage- + +![](images/47c9357176488bf309c4f0f5a4e889f133bf456093d71c136db3a4a851f5444f.jpg) + +![](images/abbcdea02b1500d23c7a11a6b90130f2af84030408220d84ba885b4b2ae827b3.jpg) + +![](images/a14293dff0a56dad803657968e30e8bf7dd2ea8ae16df59c0177b190f60db381.jpg) +Original A density map + +![](images/97aec8fc0222aadb987475cab9c9228e19efc5734f45015ee7009ce2cd478033.jpg) +Mixup (A + background of B) +Mixup (A + B) + +![](images/812e4c71821035264a2aefe801a99994482187486a7b629922447ceecb94e854.jpg) +Mixup (background only) density map +Mixup density map + +![](images/c18cc0f918477e80c423786c991d9d4ac67788115c4182d50c6ec08d337f33a7.jpg) +Original B density map + +![](images/5291994ce4a7d114b49708f949aa45d1a945e5c67d051fd6aeef41d6c8b0fb33.jpg) +Inpainted A + +![](images/c03ec3d1cf06d34965e3e49dd036e8d69846e6c2c667ef7e61ed5969e698a1f3.jpg) +Inpainted A density map + +![](images/379c121efee75d12876182ab18f8e8a76abeab0a0ac954f2784f4c477f588be2.jpg) +Inpainted B + +![](images/cfc17588551bad9a13f1450df2644d242306fcc3e81795ebf82cf6c0a203a3bb.jpg) +Inpainted B density map + +![](images/1c37fd19f3b54f97b7ca17371b95fdf005f1b35f812261f75a306bef9a1ea9bb.jpg) +Figure 1. Comparisons of augmentations. As shown in the red bounding boxes, both types of Mixup generate disappointing density maps. Mixup is supposed to retain crowd information from two images, but actually it destroys the spatial structure. As shown in green bounding boxes, our inpainting augmentation is better suited for the crowd counting task, showing an impressive reduction in Mean Absolute Error (MAE). + +ment [14, 50, 61] and disaster prevention [51]. The substantial demand has driven the advancement of counting methods [3, 25, 43, 45, 65]. However, a significant cost of manual annotation is unavoidable for these fully-supervised deep methods to achieve satisfactory performance. For example, the UCF-QNRF dataset with 1,535 images and 1,251,642 annotations amounts to approximately 2,000 human-hours [17]. Compared to the extremely expensive labeled data with annotations, acquiring unlabeled crowd images is much more affordable. Consequently, there is a growing interest in leveraging unlabeled data through semi-supervised crowd counting (SSCC) methods. The recent mainstream works on SSCC [21, 24, 40, 68] mostly employ the Mean Teacher [52] framework, where the teacher + +model generates pseudo-labels for unlabeled data to guide the student model. The key to achieving superior performance with SSCC methods on a given set of images, where only a small portion is annotated, lies in how we handle the unlabeled samples and how we design an appropriate model to fully exploit the information within them. + +It has been observed that introducing similar but different samples to the training data can enlarge the support of the training distribution and thus enhance the robustness and generalizability of the model [4, 46]. In the literature, numerous data augmentation techniques have demonstrated their effectiveness in generating additional virtual samples to support the training process in both fully-supervised and semi-supervised settings [2, 63, 66]. However, we find that these off-the-shelf techniques cannot suit the crowd counting task, as they may destroy the spatial structure of crowd scenes, leading to distorted density maps, as shown in Figure 1. Therefore, a well-suited data augmentation technique specifically designed for the crowd counting task is crucial for the model to taste more from the unlabeled data. + +With the help of appropriate data augmentation, the diversity of unlabeled data during training can be significantly improved. However, this simultaneously brings new challenges for the teacher model to produce correct supervision signals. From the perspective of the model itself, we find that the Convolutional Neural Network (CNN) architecture commonly used in previous studies can be suboptimal for the crowd counting task, as it tends to overfit to local details and fails to capture global context information, which is crucial for extremely crowded, low-light, and adverse weather scenarios. From another perspective, it is also important to generate more accurate pseudo-labels while filtering out low-quality parts. Considering these two points, we explore a modern backbone network to replace CNNs and enhance it with a holistic strategy for producing and leveraging pseudo-labels, which helps to taste data better. + +In this paper, we propose a novel semi-supervised crowd counting method, named Taste More Taste Better (TMTB). We design a novel data augmentation technique, Inpainting Augmentation, which enhances the diversity of the dataset while keeping the integrity of the foreground regions. Observing that some regions in inpainted images are noisy and harmful, to better utilize the inpainted images, we employ a Visual State Space Model (VSSM) with an Anti-Noise classification branch for assistance, which provides inexact but accurate supervision. The proposed model generates more accurate density maps than state-of-the-art methods. Our method TMTB achieves significant improvements on several benchmark datasets, including ShanghaiTech Part A & B [69], UCF-QNRF [17], and JHU-Crowd++ [47]. We reduce the Mean Absolute Error of JHU-Crowd with $5\%$ data labeled to 67.0, marking the first time it has been lower than 70.0. The method also presents strong generalization ability + +in cross-dataset evaluations. + +The main contributions of this paper are as follows: + +- We observed that the lack of suitable data augmentation methods and the weakness of capturing global context information are bottlenecks for the crowd counting task. The off-the-shelf data methods may destroy the spatial structure of crowd scenes, which is crucial for the regression of density maps. CNNs tend to overfit local details and struggle to capture global context information. +- We propose a novel semi-supervised crowd counting method. A data augmentation technique is designed to enhance diversity while keeping the integrity of the foreground regions. A VSSM with an Anti-Noise classification branch for assistance is employed, which provides inexact but accurate supervision, to better utilize the data. +- The proposed method remarkably outperforms state-of-the-art methods across four benchmark datasets and three label settings. It achieves a groundbreaking reduction in the Mean Absolute Error on the JHU-Crowd dataset with only $5\%$ of the data labeled, lowering it to 67.0, a decrease of $12.4\%$ . Furthermore, the method demonstrates exceptional generalization capabilities in cross-dataset Domain Generalization settings, even surpassing the performance of fully-supervised approaches, setting a new standard for excellence. + +# 2. Releted Works + +Semi-supervised Crowd Counting. Fully-supervised crowd counting methods[5, 13, 17, 31, 38, 42, 49, 69] maintain the best performance in crowd counting tasks. However, fully-supervised methods require a large quantity of labeled data, which is expensive to collect and annotate. + +Therefore, semi-supervised crowd counting methods have been proposed to address the scarcity of labeled data. The Mean Teacher [52], a successful semi-supervised learning framework that has been widely used in classification [10, 48, 64], segmentation [2, 56] and regression [8, 19, 62], is the most popular method in semi-supervised crowd counting. L2R [26] leveraged unlabeled data through a learning-to-rank framework. IRAST [27] introduced surrogate tasks, self-learning inter-related binary segmentation tasks, to improve performance. Moreover, MTCP [73] applied multiple auxiliary tasks into representation learning, which also benefit performance. DA-Count [23] employed a Transformer [53] to refine foreground features. CU [21] investigated a supervised uncertainty estimation strategy to better utilize unlabeled data. SALCrowd [68] proposed a method to select unlabeled data based on intra-scale uncertainty and inter-scale inconsistency metrics. MRC-Count [40] incorporated a fine-grained density classification task, proving the effectiveness of the classification task in crowd counting. Most previous approaches [21, 26, 27, 40, 68] are based on CNNs or par + +![](images/76ef64fe59dac56b1b3f9d223f42aacb5703dbe5a17cc6197ed4e9dfa5228d8f.jpg) +Figure 2. The overall framework of our method TMTB. TMTB contains the Mean Teacher framework for semi-supervised learning, VSSM as the backbone, a classification branch for predicting masks, and an inpainter for inpainting augmentation. The inpainting process is conducted periodically, which generates or updates the inpainted images. For filtering out unreliable regions, the teacher model generates the weighted mask based on the inconsistency level of the inpainted image, which is applied in $\mathcal{L}^{\mathrm{inp}}$ . The teacher model generates pseudolabels for unlabeled data (inpainted images included), and the student model predicts on labeled data and strongly augmented unlabeled data (inpainted images included). + +tially [23] based on Transformers, which are weak in modeling long-range dependencies. When facing extremely dense crowd scenes, the performance of these methods is not satisfactory. + +Generative Data Augmentation. Deep learning models tend to overfit when the quantity of data is limited, which harms the generalization ability of models. Therefore, data augmentation was proposed to address this issue. For image-based tasks, basic data augmentation methods include random cropping, flipping, rotation, scaling, color jittering, and noise adding. Some advanced data augmentation methods, such as CutMix [63], Mixup [66], and AutoAugment [7], have been proposed. However, these methods are not suitable for crowd counting tasks. Both the density-based and point-based crowd counting methods are sensitive to the spatial structure of the crowd, and the above methods may destroy the spatial structure of the crowd. The damage to the integrity of the crowd will lead to a decrease in model performance. We study this in Table 2. + +Generative Data Augmentation [11, 15, 16, 20, 41] is the process of creating new synthetic data points that can be added to existing datasets. This unique approach is widely used to improve the performance of models by increasing the quantity and diversity of training data while maintaining the quality of the data. In [67], queries-only data was used to improve the multilingual task, showing the effectiveness of generative data augmentation. FreeMask [59] resorted to synthetic images from generative models to ease the burden of both data collection and annotation procedures, and achieves comparable performance to real-data-based meth + +ods. G-DAUG $^{\text{C}}$ [60] generated synthetic examples using pretrained language models, and selects the most informative and diverse set of examples for data augmentation, to achieve more accurate and robust learning in low-resource setting. DiffusionMix [18] employed an inpaint-and-mix strategy to generate label-preserved data for classification task. However, it requires pre-defined binary masks to concatenate the original and generated images, which is not available in crowd counting since the position and shape of the crowd are unpredictable. Compared to existing works, our contribution is to fill the gap in effective augmentation methods within the field of pixel-level predition tasks. + +# 3. Method + +We propose a semi-supervised crowd counting method that enhances the dataset using a diffusion-based model's inpainting technique and utilizes a VSSM (Visual State Space Model) with Anti-Noise classification branch to improve data diversity and model performance. + +The overall framework of our method, TMTB, is shown in Figure 2. Our method is built on the Mean Teacher framework, which is widely employed in semi-supervised learning. The teacher model shares the same architecture as the student model, and its weights are updated using the student model's weights through the Exponential Moving Average (EMA) strategy. VMamba [29], a novel VSSM (Visual State Space Model), serves as the backbone, replacing the CNN and Transformer. The image features extracted from VMamba are then fed into two independent branches: the counting branch and the classification branch, for pre + +dictions. The counting branch predicts a density map of the image, while the classification branch predicts the indices of pre-defined count interval bins. We build a module, Inpainter, designed to periodically applying background inpainting augmentation to the training set. + +We first describe our novel inpainting augmentation in Section 3.1, then we discuss our network architecture in Section 3.2, and finally, we present our loss functions in Section 3.3. + +# 3.1. Inpainting Augmentation + +Unlike classification or regression tasks, the ground truth density maps are generated by point-level annotations. The state-of-the-art data augmentation techniques, Mixup [66], CutMix [63], BCP [2], etc., are not applicable to crowd counting, since the augmentation may cause damage to ground truth density maps. We introduce the inpainting technique of diffusion-based models [15, 44] as a novel data augmentation method with the filtering capability of unreliable regions, which generates new samples, thereby expanding the diversity of dataset. + +For the inpainting mask $M^{\mathrm{inp}}$ , the count interval bins predicted by the classification branch are used to separate the foreground and the background of image $x$ . We generate the positive text prompt randomly to achieve more diverse unlabeled dataset, but the negative text prompt is kept the same for all samples. The inpainting process is conducted periodically during the training process using a preset time interval $T^{\mathrm{inp}}$ : + +$$ +\begin{array}{l} M ^ {\text {i n p}} = \mathbf {0} (\arg \max (\hat {p})) , \\ x ^ {\text {i n p}} = \mathcal {S D} (x, M ^ {\text {i n p}}, r ^ {\text {p o s}}, r ^ {\text {n e g}}), \tag {1} \\ \end{array} +$$ + +where $\hat{p}_i$ is the predicted probability distribution from the classification branch, $\mathbf{0}(\cdot)$ is the indicator function, $x^{\mathrm{inp}}$ is the new image with the background inpainted, $\mathcal{SD}(\cdot ,\cdot ,\cdot ,\cdot)$ represents Stable Diffusion [44], which is a mature, excellent, and cost-effective diffusion model, and $r^{\mathrm{pos}}$ and $r^{\mathrm{neg}}$ are the positive and negative text prompts, respectively. + +Figure 3 shows examples of both good and poor inpainting samples. All new samples are considered unlabeled, as the inpainting process may add additional humans for better coherence. For well-inpainted samples, the foreground regions are well preserved, while the background regions are inpainted with different patterns, thus introducing diversity to the dataset. Even though the inpainting process is not perfect, the model can still benefit from the augmented dataset. However, failures may sometimes occur, potentially leading to the model learning from noisy inputs. The heads of humans and other features in the unreliable regions of poorly-inpainted images are fragmented and indistinct, rendering them difficult to recognize. It is necessary to filter out these unreliable regions. The EMA model calculates + +![](images/bdc36da9be0e50dd7b749392b7f13ceb2eff1446f28a4a80bfc4e5660b7804cd.jpg) + +![](images/f0f92ba5a7ab830b4627ddfea077397e786625cfe7c2d654052dab268597d3b5.jpg) + +![](images/4f27d485cf7d13bde71c285d6b41f5f8b072130bfefa7e2f8393bce1b8a1a7be.jpg) + +![](images/025f93ebc72dbdf9dc285e09df03e7ae55ee027e6c7d1d7fa46f1061f8db5f22.jpg) + +![](images/2f850b867f7f3619df9b21ebb8afa6ea90b60fb34b6f235ed8ad620cf14d070e.jpg) + +![](images/30ee914bbc6d7ca52c8d2e31ed79801e03264d58d6fa788b41ad04760c9c2783.jpg) + +![](images/0aa76cc00917bd5871fea08311e2040a992a332e5789600cf1280fc610a8d3e6.jpg) +(a) + +![](images/50c7922964f420d591552c3474d0e7f1cc268e3df5b08a9c8e5a1c3e37c2cd1c.jpg) +(b) + +![](images/3bb6e907c05a346057b0d1c786758b2af710582b9971316459c62977b44bf6f6.jpg) + +![](images/fa554f5d095db2c28ba03de2179678f6459ad75050bc7b8948a88823f6adc0df.jpg) + +![](images/df8f947ff5583cc6dd5c85e84d78c94ccdb86d94aa45626b261c2f8c2132999e.jpg) + +![](images/217ecad13b4843318302778885535058acb664ae03829c254a1dfd1c72fce3bd.jpg) +(c) +Figure 3. Inpainting samples. The left part shows well-inpainted couples, while the right part shows poorly-inpainted couples. Column (a) and (c) are the original images, while column (b) and (d) are the inpainted images. + +![](images/e58cf3d47797b103aa8baba045040c275ac652bee2954f9d6a9a53553e164f83.jpg) + +![](images/00a245736f630b4e92effb4fea8f8cec05bde2bebee3bbba959ac4a137f15246.jpg) + +![](images/a8d75ad999d2c8b7fda930ac51b7eb891eb9b607341b6fa914bf706905835be2.jpg) + +![](images/7a98e93f86a4b5f8b6e3649c7a3a223aa60f53430fe51d6b2206027484d83291.jpg) +(d) + +the inconsistency level of the inpainted image by the difference in the classification branch between the weakly augmented version and the strongly augmented version. The inconsistency level and the weighted mask are defined as follows: + +$$ +\begin{array}{l} M _ {l} ^ {\text {i n c o n}} = \mathbf {0} \left(\left| \left| p ^ {\hat {s}} - p ^ {\hat {w}} \right| - l \right|\right), \\ \omega_ {l} ^ {t} = \operatorname {s o f t m a x} \left(e ^ {- l \cdot t / T ^ {\mathrm {i n p w}}}\right), \tag {2} \\ \end{array} +$$ + +$$ +M ^ {\text {i n c o n}} = \sum_ {l = 0} ^ {L} M _ {l} ^ {\text {i n c o n}} \cdot \omega_ {l} ^ {t}, +$$ + +where $\hat{p^s}$ and $\hat{p^w}$ are the predicted probability distributions of the classification branch of EMA model in the strong and weak augmentation version, respectively. $M_{l}^{\mathrm{incon}}$ is the mask of inconsistency level $l$ , $M^{\mathrm{incon}}$ is the final weighted mask, $\omega_l^t$ is the weight of inconsistency level $l$ at time $t$ , $T^{\mathrm{inpw}}$ is the parameter controlling the weight decay speed, and $L$ is the number of inconsistency levels. The $L$ depends on the predefined count interval bins, and since we followed the interval bins settings in [55], the $L$ is set to 2 for the balance of diversity and reliability. The weighted mask is then used to adjust the loss associated with the inpainted image. + +# 3.2. VSSM with Classification for Anti-Noise + +Visual State Space Models (VSSMs). Existing crowd counting methods often encounter challenges such as congestion, blur, and occlusion in intricate scenes. CNN-based approaches tend to overfit to local details, while Transformer-based methods excel in global modeling but face heavy computational burdens due to their quadratic complexity in terms of image sizes. These limitations motivate us to introduce the VSSM, which captures comprehensive long-range information with linear time complexity, for the crowd counting task. + +![](images/12cca87cdaa083d9ec4bc9bf936bc52d361120c616e02f6ccf8c91be6d70fa72.jpg) + +![](images/aa5de0346394fe5fd1ccf8a8471f1593847b7ffbf033415979e0b8c5f51afe79.jpg) + +![](images/840b9ba167a45bf2e54efcb55b10604a76aabfa4b3f198cef7076f55bc1a0c71.jpg) + +![](images/6083d173668cbaa579be6f9dae596ba72d1373e6bbff80d87604c00ceccde312.jpg) + +![](images/ec707a3dec81ae4e3e4ed7bac659fd77064f1c70d8a08addd03beac6bfcb14f5.jpg) + +![](images/649f28792ff5649ac80710d1d70c17e8b82d9400e6c42df4b2323d80ff00a56e.jpg) + +![](images/21cb4283be0fbca87edb8969f7363f4710ecd00c3c354274ca3e5c1c9baeced6.jpg) + +![](images/ce423d4f10383661bbb537b6e2c9cde346c6dd30fc12be5b602876c958b15f63.jpg) + +![](images/976da9fe03f040cc5011e9a8e5b74926a1ad48e4358b08f1d9cd6446a3ffa3f9.jpg) +Figure 4. Visualizations of the predictions on the test set of the JHU-Crowd++ dataset. The first row shows the input images. The second row displays the predictions of the SOTA method MRC-Crowd [40]. The third row presents the predictions of our method TMTB. All models are trained with $5\%$ labeled data. + +![](images/23d7638c046dab4b4d52fa2660fd5c1c5139fa6a738f34c446bb517d956ed437.jpg) + +![](images/a3c99317cf5a12f8c7c359e3564d014847408255575c8ee6adf2545634db76f4.jpg) + +![](images/39b616760553d3f6e9c0b669d9da97af9cefd58d77c32b5bd260aa75da1a44c8.jpg) + +State Space Models (SSMs) map the input stimulation $u(t)$ to the output response $y(t)$ by means of a hidden state $h(t)$ . These processes are formulated with the following form: + +$$ +\begin{array}{l} h ^ {\prime} (t) = A h (t) + B u (t), \\ y (t) = C h (t) + D u (t), \tag {3} \\ \end{array} +$$ + +where $A, B, C,$ and $D$ are the weighting parameters. + +Linear time-invariant SSMs (Equation (3)) lack the ability to capture the contextual information, [12] proposed a novel parameterization method for SSMs that integrates an input-dependent selection mechanism (referred to as S6). VMamba [29] replaces the S6 module, with the newly proposed 2D-Selective-Scan (SS2D) module which concurrently achieving global receptive fields, dynamic weights, and linear complexity. By adopting complementary 1D traversal paths, SS2D enables each pixel in the image to effectively integrate information from all other pixels in different directions, thereby facilitating the establishment of global receptive fields in the 2D space and generating more accurate density map. Details about the SS2D module could be found in [29]. The features extracted by the final SS2D block are then fed into two branches: one for density map regression and the other for count interval classification, which is designed for less inexact and more accurate supervision. + +Exact: Regression Head. The regression head is designed to predict the density map for the input image. The density map is then summed to obtain the exact final count value. The regression head consists of up-sampling layers, convolutional layers, batch normalization layers, and ReLU activation layers. We followed MRC-Crowd [40] employing the loss function proposed in CUT [39], which significantly + +boosts the performance in the foreground area. It could be expressed as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {r e g}} = \frac {1}{J} \sum_ {j = 1} ^ {J} \left(1 - \operatorname {S S I M} \left(D _ {j} (\hat {y} \odot M ^ {g t}\right), \right. \tag {4} \\ \left. D _ {j} \left(y ^ {g t} \odot M ^ {g t}\right)\right) + \alpha \cdot \mathcal {L} _ {T V} \left(\hat {y}, y ^ {g t}\right), \\ \end{array} +$$ + +where $\mathrm{SSIM}(\cdot ,\cdot)$ is the structural similarity index measure, and $D_{j}$ denotes the average pooling operation that downsamples an image to $1 / 2^{j}$ of its original size. The operation $\odot$ denotes the Hadamard product, and $M^{gt}$ is the binary mask that distinguishes the dense and sparse regions based on a predefined threshold $\tau$ . It can be obtained using the indicator function $\mathbf{1}$ as $M^{gt} = \mathbf{1}(y^{gt} > \tau)$ . $\alpha$ is a hyperparameter that balances the two terms. $J$ and $\alpha$ are set to 3 and 0.01, respectively. $\mathcal{L}_{TV}$ is the total variation loss. + +Accurate: Anti-Noise Classification Head. Many previous works [1, 31, 34, 54] have identified the problem of inaccurate ground truth annotations in crowd counting datasets, which render the ground truth density maps noisy and inaccurate. In this way, the model may learn from the noisy ground truth and generate inaccurate density maps. Inspired by previous works [23, 40, 55], we employ the Anti-Noise classification task as an auxiliary task, which is more robust in predicting inexactly and accurately. The Anti-Noise classification branch is supervised by the inexact but accurate count interval bins rather than the exact but maybe inaccurate density maps. Furthermore, we use the prediction of the this branch to generate reliable masks for inpainting samples, and we discuss this in Section 3.1. The goal of the Anti-Noise classification branch is to predict the accurate count interval bins, which are predefined by [55]. + +The Anti-Noise classification head consists of an up + +sampling layer, a convolutional layer, a ReLU activation layer and a final convolutional layer. The standard cross-entropy loss is applied. In the case of labeled data, the loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {c l s}} = \frac {1}{N} \sum_ {i} ^ {N} \mathcal {H} \left(p _ {i} ^ {g t}, \hat {p} _ {i}\right), \tag {5} +$$ + +where $\mathcal{H}$ is the cross-entropy loss function, $p_i^{gt}$ is the ground truth probability distribution, and $\hat{p}_i$ is the predicted probability distribution. + +# 3.3. Losses + +In order to make full use of the labeled data $L = \{(x_{i}^{l},DM_{i}^{gt})\}_{i = 1}^{N_{l}}$ and unlabeled data $U = \{(x_{i}^{u})\}_{i = 1}^{N_{u}}$ , we employ a common semi-supervised learning strategy widely used in crowd counting. The labeled samples are used to directly train the student model, while the unlabeled samples are used for consistency regularization. For labeled data, the supervised loss $\mathcal{L}^{\mathrm{s}} = \mathcal{L}_{\mathrm{reg}}^{\mathrm{s}} + \mathcal{L}_{\mathrm{cls}}^{\mathrm{s}}$ is a combination of the regression loss (Equation (4)) and the classification loss (Equation (5)). Given the absence of a confidence score for the predicted regression map, we choose to apply consistency regularization to all unlabeled samples, rather than developing an algorithm to assess prediction uncertainty. The EMA teacher model generates pseudo-labels based on the prediction of unlabeled data with weak augmentation. The student model makes predictions based on the strong augmentation version. The error loss between the two predictions is calculated as the consistency loss. + +Despite the common strong augmentation, we also introduce a patch-aligned random masking strategy [58] as a kind of strong augmentation. The patch-aligned random masking strategy randomly masks out some patches of various sizes in the image. The student model is required to predict based on masked images, while the teacher model is able to access the full image. The semi-supervised loss is defined as: + +$$ +\mathcal {L} ^ {\mathrm {u}} = \operatorname {M A E} \left(\hat {y} ^ {s t}, \hat {y} ^ {e m a}\right) + \operatorname {M A E} \left(\hat {p} ^ {s t}, \hat {p} ^ {e m a}\right), \tag {6} +$$ + +where $\hat{y}^{st}$ and $\hat{y}^{ema}$ are the predicted density maps of the student model and the teacher model, respectively. $\hat{p}^{st}$ and $\hat{p}^{ema}$ are the predicted count interval bins of the student model and the teacher model, respectively. + +Consistency loss is also applied to the inpainted images. However, to avoid the model learning from noisy regions with low quality, the loss is adjusted by the weighted mask of inconsistency level. The inpainting loss is defined as: + +$$ +\mathcal {L} ^ {\text {i n p}} = M ^ {\text {i n c o n}} \times \operatorname {M A E} \left(\hat {y} ^ {s t}, \hat {y} ^ {e m a}\right) + M ^ {\text {i n c o n}} \times \operatorname {M A E} \left(\hat {p} ^ {s t}, \hat {p} ^ {e m a}\right), \tag {7} +$$ + +where $M^{\mathrm{incon}}$ is the weighted mask of inconsistency level mentioned in Equation (2). + +The final loss is a combination of the supervised, semi-supervised, and inpainting losses: + +$$ +\mathcal {L} = \mathcal {L} ^ {\mathrm {s}} + \lambda_ {w} \cdot \mathcal {L} ^ {\mathrm {u}} + \lambda_ {w} \cdot \mathcal {L} ^ {\text {i n p}}, \tag {8} +$$ + +where $\lambda_w$ is the warm-up weight for the semi-supervised loss and the inpainting loss. + +# 4. Experiments + +We evaluate our method on four public datasets: + +- JHU-Crowd++ [47] is a comprehensive dataset with 4,372 images and 1.51 million annotations, collected under a variety of scenarios and environmental conditions. +- UCF-QNRF [17] contains 1535 images. The dataset has a greater number of high-count crowd images and annotations, and a wider variety of scenes containing the most diverse set of viewpoints, densities and lighting variations. +- ShanghaiTech A & B [69] are two datasets containing nearly 1,200 images and 330,000 labeled heads. + +To achieve fair comparisons, we follow the same semi-supervised experimental protocol as previous works [23, 24, 40], including the $5\%$ , $10\%$ , and $40\%$ labeled data settings and the split of labeled and unlabeled data. The evaluation metrics are Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The detailed definitions of them are provided in the supplementary material. + +For some excessively high-resolution images, we scale them down, ensuring the longest side does not exceed 1,920 pixels. For the ground truth density map, we use geometry-adapted Gaussian kernels described in [22] for all datasets except the ShanghaiTech B dataset, and we fix the size of the kernel to 4. When training, the input images are randomly cropped to $512 \times 512$ pixels for all datasets except for ShanghaiTech A. The inputs from ShanghaiTech A are cropped to $256 \times 256$ pixels instead. The count interval bins, which should be predicted by the classification branch, are set to the same as those in [55]. The AdamW optimizer [30] with a learning rate of $10^{-5}$ is used to optimize the model. The weight decay is set to $10^{-4}$ . The downsample factor is set to 8. The batch size is set to 8. In $5\%$ and $10\%$ labeled data settings, the ratio of labeled data to unlabeled data is set to $2:6$ . In the $40\%$ labeled data setting, the ratio is set to $4:4$ . As we mentioned in Section 3.3, we use the patch-aligned random masking strategy [58] as a form of strong augmentation. The masked patch size and the masked region ratio are set to $32 \times 32$ and 0.3, respectively. The ablation details can be found in [40]. The teacher model is updated by the exponential moving average with a decay rate of 0.97. The weights of all losses are set to 1.0. For $\mathcal{L}^{\mathrm{u}}$ and $\mathcal{L}^{\mathrm{inp}}$ , we ramp the weights from 0 to 1.0 in the first 20 epochs. The max accepted inconsistency level $L$ is set to 2, and the hyperparameter $T^{\mathrm{inpw}}$ is set to 100. The ablation study of $T^{\mathrm{inpw}}$ is conducted in Section 4.2. Each training + +Table 1. Comparisons with the SOTA methods on main public datasets. The best results are highlighted in bold, and the second-best results are underlined. + +
MethodsVenueLabeled PercentageJHU-Crowd++UCF-QNRFShanghaiTech AShanghaiTech B
MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓
MT [52]NIPS'175%101.5363.5172.4284.9104.7156.919.333.2
L2R [26]CVPR'185%101.4338.8160.1272.3103.0155.420.327.6
DACount [23]ACM MM'225%82.2294.9120.2209.385.4134.512.622.8
OT-M [24]CVPR'235%82.7304.5118.4195.483.7133.312.621.5
MRC-Crowd [40]T-CSVT'245%76.5282.7101.4171.374.8117.311.717.8
TMTBOurs5%67.0252.296.3163.472.4126.410.619.6
MT [52]NIPS'1710%90.2319.3145.5250.394.5156.115.624.5
L2R [26]CVPR'1810%87.5315.3148.9249.890.3153.515.624.4
DACount [23]ACM MM'2210%75.9282.3109.0187.274.9115.511.119.1
MTCP [73]T-NNLS'2310%88.1318.7124.7206.381.3130.514.522.3
OT-M [24]CVPR'2310%73.0280.6113.1186.780.1118.510.818.2
CU [21]ICCV'2310%74.9281.7104.0164.370.8116.69.717.7
SALCrowd [68]ACM MM'2410%69.7263.5106.7171.369.7114.59.717.5
MRC-Crowd [40]T-CSVT'2410%70.7261.393.4153.267.3106.810.318.2
TMTBOurs10%66.3265.191.7156.365.7110.49.817.2
MT [52]NIPS'1740%121.5388.9147.2247.688.2151.115.925.7
L2R [26]CVPR'1840%123.6376.1145.1256.186.5148.216.825.1
SUA [33]ICCV'2140%80.7290.8130.3226.368.5121.914.120.6
DACount [23]ACM MM'2240%65.1260.091.1153.467.5110.79.614.6
OT-M [24]CVPR'2340%72.1272.0100.6167.670.7114.58.113.1
SALCrowd [68]ACM MM'2440%62.0255.195.0167.060.797.27.912.7
MRC-Crowd [40]T-CSVT'2440%60.0227.381.1131.562.195.57.813.3
TMTBOurs40%60.0240.381.4140.660.899.07.512.9
+ +process coordinates with two inpainting background processes. The inpainting period $T^{\mathrm{inp}}$ is set to 80 epochs. + +# 4.1. Main Results + +In our main experiments, the proposed TMTB demonstrates superior performance compared to state-of-the-art methods across four public datasets with three labeled data percentages. The results are shown in Table 1. + +Notably, when utilizing only $5\%$ labeled data, TMTB achieves the lowest MAEs of 67.0 on JHU-Crowd++, 96.3 on UCF-QNRF, 72.4 on ShanghaiTech A, and 10.6 on ShanghaiTech B, demonstrating the best performance. The result on JHU-Crowd++ not only surpasses the performance of SOTA methods with the same amount of labeled data but also demonstrates that TMTB remains superior even at the $10\%$ label setting, highlighting its efficient label utilization and exceptional accuracy. With a $10\%$ labeled data setting, TMTB continues to outstand, obtaining a leading MAE of 66.3 on JHU-Crowd++, 91.7 on UCF-QNRF, and 65.7 on ShanghaiTech A, further demonstrating its generalization ability and effectiveness. At $40\%$ labeled data, TMTB achieves competitive results, matching the lowest MAE of 60.0 on JHU-Crowd++, and decreasing the MAE to a lowest of 7.5 on ShanghaiTech B, maintaining strong performance across the other datasets. These results underscore the capability of TMTB to achieve low error with minimal labeled data, offering substantial improvements over SOTA methods. + +Table 2. An study of common various augmentation strategies on ShanghaiTech-A dataset using $10\%$ labeled data. + +
AugmentationMAE ↓RMSE ↓
Mixup [66]285.16431.83
Mixup [66] (background only)72.00116.85
CutMix [63]312.60462.33
BCP [2]70.99121.83
TMTB (Ours)65.74110.42
+ +Augmentation Strategies. In order to investigate the effectiveness of the proposed inpainting augmentation strategy, we compare it with some off-the-shelf augmentation strategies, on the ShanghaiTech A dataset with $10\%$ labeled data, including Mixup [66], CutMix [63], and BCP [2]. The results are shown in Table 2. + +Mixup and CutMix both degrade the performance when applied to the whole image. The main reason is that the integrity of the foreground and background regions is destroyed, making the augmented images non-informative and hard to learn. These results demonstrate the importance of the integrity of the foreground and background regions in the crowd counting task. We then apply Mixup to the background region only, and the performance improves. This approach is somewhat similar to the inpainting augmentation, which enhances the background region while keeping the foreground region intact. However, the background regions are sampled from the domain of the original dataset, which + +Table 3. Comparisons with SOTA single domain generalization methods on Q (UCF-QNRF) and A (ShanghaiTech-A) datasets. The best results are highlighted in bold, and the second-best results are underlined. + +
MethodsVenueSource Labeled PercentageA → QQ → A
MAE ↓RMSE ↓MAE ↓RMSE ↓
IBN [36]ECCV'18100%280.2561.0105.9174.6
SW [37]ICCV'19100%285.5431.0102.4168.8
ISW [6]CVPR'21100%215.6399.783.4136.0
DG-MAN [32]ICCV'21100%129..1238.2--
DCCUS [9]AAAI'23100%119.4216.667.4112.8
MPCount [38]CVPR'24100%115.7199.865.5110.1
TMTBOurs40%112.5190.662.4103.5
+ +Table 4. Comparisons with SOTA domain adaptation methods on Q (UCF-QNRF) and A (ShanghaiTech-A) datasets. The best results are highlighted in bold, and the second-best results are underlined. + +
MethodsVenueSource Labeled PercentageTarget Domain VisibleA → QQ → A
MAE ↓RMSE ↓MAE ↓RMSE ↓
RBT [28]ACM MM'20100%✓175.0294.8--
C2MoT [57]ACM MM'21100%✓125.7218.3--
FGFD [70]ACM MM'22100%✓124.1242.070.2118.4
DAOT [71]ACM MM'23100%✓113.9215.667.0128.4
FSIM [72]T-MM'23100%✓105.3191.166.8111.5
TMTBOurs40%✗112.5190.662.4103.5
+ +may not be sufficiently diverse. BCP is a recent birectical augmentation strategy designed for medical image segmentation. We employ it in the crowd counting task and find it can also improve the performance. However, the performance is not competitive enough when compared to SOTA methods. + +Domain Generalization & Domain Adaptation. We compare our method with state-of-the-art single domain generalization methods and obtain remarkable results in Table 3. The results show that our method outperforms the SOTA methods on both UCF-QNRF $\rightarrow$ ShanghaiTech-A and ShanghaiTech-A $\rightarrow$ UCF-QNRF tasks, utilizing only $40\%$ labeled data on the source domain instead of $100\%$ . Note that IBN, SW, and ISW are originally designed for image classification or segmentation, and are adapted to crowd counting and reported by MPCount [38]. The results demonstrate the effectiveness of VSSM and inpainting augmentation in our method. + +The study on domain adaptation is also conducted, and the results are shown in Table 4. Domain adaptation setting is easier than domain generalization, as the target domain is visible. It is surprising that our method achieves competitive results compared to SOTA methods, even though we only have $40\%$ labeled data available in the source domain and no access to the target domain. The comparisons again provide evidence of the generalization ability of TMTB. + +Table 5. An ablation study on the effectiveness of various components on UCF-QNRF dataset using $5\%$ labeled data. + +
VSSMLsLuLinpMAE ↓RMSE ↓
XLsereg115.35188.19
✓Lsereg110.41185.52
✓Lsereg + Lcsls105.88183.87
✓Lsereg + Lcsls✓103.58178.97
✓Lsereg + Lcsls✓ωl97.21166.06
✓Lsereg + Lcsls✓ωlt96.33163.42
+ +# 4.2. Ablation Study + +Components. We conduct an ablation study to analyze the effectiveness of various components in our method on the UCF-QNRF dataset with $5\%$ labeled data. The results are shown in Table 5. The VSSM is the core component. Compared to OT-M [24], the VSSM with supervised loss $\mathcal{L}^{\mathrm{s}}$ improves the performance by decreasing the MAE by $10.57\%$ . After adding the semi-supervised loss $\mathcal{L}^{\mathrm{u}}$ , the performance is further improved, showing the effectiveness of our semi-supervised learning strategy. Next, we introduce our novel inpainting augmentation strategy and its mask of inconsistency level $M_l^{\mathrm{uncon}}$ but with fixed weights $\omega_{l}$ . The performance is further improved, and the MAE of 97.21 surpasses the SOTA method MRC-Crowd [40] at 101.4. Finally, we employ time-based weights $\omega_l^t$ for the inpainting loss, and the performance is significantly improved, achieving the lowest MAE of 96.33 and RMSE of 163.42. + +Weight Decay Speed. We study the impact of the weight decay speed $T^{\mathrm{inpw}}$ on the JHU-Crowd++ dataset with 5% labeled data. The detailed influence of $T^{\mathrm{inpw}}$ on the MAE and RMSE is shown in supplementary material. We train and test TMTB with different $T^{\mathrm{inpw}}$ values ranging from 80 to 120. The results show that both MAE and RMSE decrease as $T^{\mathrm{inpw}}$ increases from 80 to 100. However, when $T^{\mathrm{inpw}}$ is greater than 100, the performance begins to degrade. Therefore, we set $T^{\mathrm{inpw}}$ to 100 in experiments. + +# 5. Conclusion + +In the literature of semi-supervised crowd counting, a number of methods based on pseudo-labeling have been explored. Despite their success, these methods still struggle to achieve satisfactory performance in extremely crowded, low-light, and adverse weather scenarios. In this work, we reveal that the lack of suitable data augmentation methods and the weakness of capturing global context information are bottlenecks. Therefore, we design a novel inpainting augmentation method, Inpainting Augmentation that works better for crowd counting, enabling tasting more. Moreover, we introduce the Visual State Space Model to capture long-range dependencies in crowd counting, enabling tasting better. Our proposed framework, TMTB, achieves superior performance compared to existing methods. + +# References + +[1] Shuai Bai, Zhiqun He, Yu Qiao, Hanzhe Hu, Wei Wu, and Junjie Yan. Adaptive dilated network with self-correction supervision for counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5 +[2] Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11514-11524, 2023. 2, 4, 7 +[3] Lokesh Boominathan, Srinivas S S Kruthiventi, and R. Venkatesh Babu. Crowdnet: A deep convolutional network for dense crowd counting. In Proceedings of the 24th ACM International Conference on Multimedia, page 640-644, New York, NY, USA, 2016. Association for Computing Machinery. 1 +[4] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. In Advances in Neural Information Processing Systems. MIT Press, 2000. 2 +[5] Lei Chen, Xinghang Gao, Fei Chao, Xiang Chang, Chih Min Lin, Xingen Gao, Shaopeng Lin, Hongyi Zhang, and Juqiang Lin. The effectiveness of a simplified model structure for crowd counting, 2024. 2 +[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11580-11590, 2021. 8 +[7] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[8] Weihang Dai, Xiaomeng Li, and Kwang-Ting Cheng. Semi-supervised deep regression with uncertainty consistency and variational model ensembling via bayesian neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 37(6):7304-7313, 2023. 2 +[9] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):561-570, 2023. 8 +[10] Yue Duan, Zhen Zhao, Lei Qi, Lei Wang, Luping Zhou, Yinghuan Shi, and Yang Gao. Mutexmatch: Semi-supervised learning with mutex-based consistency regularization. IEEE Trans. Neural Networks Learn. Syst., 35(6): 8441-8455, 2024. 2 +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3 +[12] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling, 2024. 5, 2 + +[13] Tao Han, Lei Bai, Lingbo Liu, and Wanli Ouyang. Steerer: Resolving scale variations for counting and localization via selective inheritance learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21848-21859, 2023. 2 +[14] Marcus Handte, Muhammad Umer Iqbal, Stephan Wagner, Wolfgang Apolinarski, Pedro José Marrón, Eva Maria Muñoz Navarro, Santiago Martínez, Sara Izquierdo Barthelemy, and Mario González Fernández. Crowd density estimation for public transport vehicles. In EDBT/ICDT Workshops, pages 315-322. CiteSeer, 2014. 1 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 3, 4 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[17] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 1, 2, 6 +[18] Khawar Islam, Muhammad Zaigham Zaheer, Arif Mahmood, and Karthik Nandakumar. Diffusemix: Label-preserving data augmentation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27621-27630, 2024. 3 +[19] Neal Jean, Sang Michael Xie, and Stefano Ermon. Semisupervised deep kernel learning: Regression with unlabeled data by minimizing predictive variance. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 2 +[20] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[21] Chen Li, Xiaoling Hu, Shahira Abousamra, and Chao Chen. Calibrating uncertainty for semi-supervised crowd counting. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16685-16695, 2023. 1, 2, 7 +[22] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6 +[23] Hui Lin, Zhiheng Ma, Xiaopeng Hong, Yaowei Wang, and Zhou Su. Semi-supervised crowd counting via density agency. In Proceedings of the 30th ACM International Conference on Multimedia, page 1416-1426, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 5, 6, 7 +[24] Wei Lin and Antoni B. Chan. Optimal transport minimization: Crowd localization on density maps for semisupervised counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21663-21673, 2023. 1, 6, 7, 8 + +[25] Weizhe Liu, Mathieu Salzmann, and Pascal Fua. Context-aware crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[26] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Leveraging unlabeled data for crowd counting by learning to rank. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7 +[27] Yan Liu, Lingqiao Liu, Peng Wang, Pingping Zhang, and Yinjie Lei. Semi-supervised crowd counting via self-training on surrogate tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 242-259. Springer, 2020. 2 +[28] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, page 129-137, New York, NY, USA, 2020. Association for Computing Machinery. 8 +[29] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. VMamba: Visual state space model. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 5, 1 +[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6 +[31] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 5 +[32] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6630-6638, 2021. 8 +[33] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15549-15559, 2021. 7 +[34] Greg Olmschenk., Hao Tang., and Zhigang Zhu. Improving dense crowd counting convolutional neural networks using inverse k-nearest neighbor maps and multiscale upsampling. In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 5: VISAPP, pages 185-195. INSTICC, SciTePress, 2020. 5 +[35] OpenAI. Gpt-4o system card, 2024. 1 +[36] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 8 +[37] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 8 +[38] Zhuoxuan Peng and S.-H. Gary Chan. Single domain generalization for crowd counting. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28025-28034, 2024. 2, 8 +[39] Yifei Qian, Liangfei Zhang, Xiaopeng Hong, Carl Donovan, and Ognjen Arandjelovic. Segmentation assisted u-shaped multi-scale transformer for crowd counting. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 5 +[40] Yifei Qian, Xiaopeng Hong, Zhongliang Guo, Ognjen Arandjelovic, and Carl R. Donovan. Semi-supervised crowd counting with contextual modeling: Facilitating holistic understanding of crowd scenes. IEEE Transactions on Circuits and Systems for Video Technology, 34(9):8230-8241, 2024. 1, 2, 5, 6, 7, 8 +[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3 +[42] Yasiru Ranasinghe, Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M. Patel. Crowd-diff: Multi-hypothesis crowd density estimation using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12809-12819, 2024. 2 +[43] Viresh Ranjan, Hieu Le, and Minh Hoai. Iterative crowd counting. In Proceedings of the European conference on computer vision (ECCV), pages 270-285, 2018. 1 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 4 +[45] Miaojing Shi, Zhaohui Yang, Chao Xu, and Qijun Chen. Revisiting perspective information for efficient crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[46] Patrice Y. Simard, Yann A. LeCun, John S. Denker, and Bernard Victorri. Transformation Invariance in Pattern Recognition - Tangent Distance and Tangent Propagation, pages 235-269. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 2 +[47] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2594-2609, 2022. 2, 6 +[48] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In Advances in Neural Information Processing Systems, pages 596-608. Curran Associates, Inc., 2020. 2 +[49] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3365-3374, 2021. 2 + +[50] Hakki Soy. Edge ai-based crowd counting application for public transport stops. In Convergence of Deep Learning and Internet of Things: Computing and Technology, pages 182-205. IGI Global, 2023. 1 +[51] Aijun Sun, Mao Liu, and Jianfeng Li. Real-time crowd massing risk supervision system based on massing crowd counting in public venue. In 2009 International Symposium on Computer Network and Multimedia Technology, pages 1-7, 2009. 1 +[52] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 2, 7 +[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2 +[54] Jia Wan and Antoni Chan. Adaptive density map generation for crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5 +[55] Changan Wang, Qingyu Song, Boshen Zhang, Yabiao Wang, Ying Tai, Xuyi Hu, Chengjie Wang, Jilin Li, Jiayi Ma, and Yang Wu. Uniformity in heterogeneity: Diving deep into count interval partition for crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3234-3242, 2021. 4, 5, 6 +[56] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4248-4257, 2022. 2 +[57] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, page 658-666, New York, NY, USA, 2021. Association for Computing Machinery. 8 +[58] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9653-9663, 2022. 6 +[59] Lihe Yang, Xiaogang Xu, Bingyi Kang, Yinghuan Shi, and Hengshuang Zhao. Freemask: Synthetic images with dense annotations make stronger segmentation models. In Advances in Neural Information Processing Systems, pages 18659-18675. Curran Associates, Inc., 2023. 3 +[60] Yiben Yang, Chaitanya Malaviya, Jared Fernandez, Swabha Swayamdipta, Ronan Le Bras, Ji-Ping Wang, Chandra Bhagavatula, Yejin Choi, and Doug Downey. Generative data augmentation for commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1008-1025, Online, 2020. Association for Computational Linguistics. 3 +[61] Jun Yi, Yiran Pang, Wei Zhou, Meng Zhao, and Fujian Zheng. A perspective-embedded scale-selection network + +for crowd counting in public transportation. IEEE Transactions on Intelligent Transportation Systems, 25(5):3420-3432, 2024. 1 +[62] Yingda Yin, Yingcheng Cai, He Wang, and Baoquan Chen. Fishermatch: Semi-supervised rotation regression via entropy-based filtering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11164-11173, 2022. 2 +[63] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features, 2019. 2, 3, 4, 7 +[64] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In Advances in Neural Information Processing Systems, pages 18408-18419. Curran Associates, Inc., 2021. 2 +[65] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1 +[66] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 2, 3, 4, 7 +[67] Shimao Zhang, Changjiang Gao, Wenhao Zhu, Jiajun Chen, Xin Huang, Xue Han, Junlan Feng, Chao Deng, and Shujian Huang. Getting more from less: Large language models are good spontaneous multilingual learners. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8037-8051, Miami, Florida, USA, 2024. Association for Computational Linguistics. 3 +[68] Shiwei Zhang, Wei Ke, Shuai Liu, Xiaopeng Hong, and Tong Zhang. Boosting semi-supervised crowd counting with scale-based active learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8681-8690, New York, NY, USA, 2024. Association for Computing Machinery. 1, 2, 7 +[69] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6 +[70] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, page 5659-5668, New York, NY, USA, 2022. Association for Computing Machinery. 8 +[71] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. In Proceedings of the 31st ACM International Conference on Multimedia, page 4319–4329, New York, NY, USA, 2023. Association for Computing Machinery. 8 + +[72] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 26:3842-3855, 2024. 8 +[73] Pengfei Zhu, Jingqing Li, Bing Cao, and Qinghua Hu. Multitask credible pseudo-label learning for semi-supervised crowd counting. IEEE Transactions on Neural Networks and Learning Systems, 35(8):10394-10406, 2024. 2, 7 \ No newline at end of file diff --git a/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/images.zip b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..946385b9b4cb4475b18497e29c266a4ad8008a8c --- /dev/null +++ b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c89f8ee86b2b215ec703d7892c6cc1ed69caec1d798eab2aa69a19b7f091594e +size 693039 diff --git a/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/layout.json b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..52873cfc92a2cae83a4926a60b576eafec24c190 --- /dev/null +++ b/2025/Taste More, Taste Better_ Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting/layout.json @@ -0,0 +1,11020 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 140 + ], + "type": "text", + "content": "Taste More, Taste Better: Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "text", + "content": "Maochen Yang" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "text", + "content": " Zekun Li" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "text", + "content": " Jian Zhang" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "text", + "content": " Lei Qi" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "text", + "content": " Yinghuan Shi" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "text", + "content": "Nanjing University, China " + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "text", + "content": "Suzhou Laboratory, China " + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 160, + 544, + 204 + ], + "type": "text", + "content": "Southeast University, China \n{yangmaochen, lizekun, zhangjian7369}@smail.nju.edu.cn qilei@seu.edu.cn, syh@nju.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 531 + ], + "type": "text", + "content": "Semi-supervised crowd counting is crucial for addressing the high annotation costs of densely populated scenes. Although several methods based on pseudo-labeling have been proposed, it remains challenging to effectively and accurately utilize unlabeled data. In this paper, we propose a novel framework called Taste More Taste Better (TMTB), which emphasizes both data and model aspects. Firstly, we explore a data augmentation technique well-suited for the crowd counting task. By inpainting the background regions, this technique can effectively enhance data diversity while preserving the fidelity of the entire scenes. Secondly, we introduce the Visual State Space Model as backbone to capture the global context information from crowd scenes, which is crucial for extremely crowded, low-light, and adverse weather scenarios. In addition to the traditional regression head for exact prediction, we employ an Anti-Noise classification head to provide less exact but more accurate supervision, since the regression head is sensitive to noise in manual annotations. We conduct extensive experiments on four benchmark datasets and show that our method outperforms state-of-the-art methods by a large margin. Code is publicly available on https://github.com/syhien/taste_more_tastebetter." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 553, + 135, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 135, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 135, + 565 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 572, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 622 + ], + "type": "text", + "content": "Crowd counting is a fundamental task in computer vision that involves estimating the number of people in densely populated scenes. This task is of great importance in various public safety scenarios, including transport manage-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 333, + 229, + 397, + 272 + ], + "blocks": [ + { + "bbox": [ + 333, + 229, + 397, + 272 + ], + "lines": [ + { + "bbox": [ + 333, + 229, + 397, + 272 + ], + "spans": [ + { + "bbox": [ + 333, + 229, + 397, + 272 + ], + "type": "image", + "image_path": "47c9357176488bf309c4f0f5a4e889f133bf456093d71c136db3a4a851f5444f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 400, + 230, + 465, + 271 + ], + "blocks": [ + { + "bbox": [ + 400, + 230, + 465, + 271 + ], + "lines": [ + { + "bbox": [ + 400, + 230, + 465, + 271 + ], + "spans": [ + { + "bbox": [ + 400, + 230, + 465, + 271 + ], + "type": "image", + "image_path": "abbcdea02b1500d23c7a11a6b90130f2af84030408220d84ba885b4b2ae827b3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 469, + 230, + 533, + 271 + ], + "blocks": [ + { + "bbox": [ + 469, + 230, + 533, + 271 + ], + "lines": [ + { + "bbox": [ + 469, + 230, + 533, + 271 + ], + "spans": [ + { + "bbox": [ + 469, + 230, + 533, + 271 + ], + "type": "image", + "image_path": "a14293dff0a56dad803657968e30e8bf7dd2ea8ae16df59c0177b190f60db381.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 272, + 529, + 279 + ], + "lines": [ + { + "bbox": [ + 471, + 272, + 529, + 279 + ], + "spans": [ + { + "bbox": [ + 471, + 272, + 529, + 279 + ], + "type": "text", + "content": "Original A density map" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 333, + 284, + 397, + 326 + ], + "blocks": [ + { + "bbox": [ + 338, + 273, + 392, + 284 + ], + "lines": [ + { + "bbox": [ + 338, + 273, + 392, + 284 + ], + "spans": [ + { + "bbox": [ + 338, + 273, + 392, + 284 + ], + "type": "text", + "content": "Mixup (A + background of B)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 284, + 397, + 326 + ], + "lines": [ + { + "bbox": [ + 333, + 284, + 397, + 326 + ], + "spans": [ + { + "bbox": [ + 333, + 284, + 397, + 326 + ], + "type": "image", + "image_path": "97aec8fc0222aadb987475cab9c9228e19efc5734f45015ee7009ce2cd478033.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 327, + 382, + 334 + ], + "lines": [ + { + "bbox": [ + 347, + 327, + 382, + 334 + ], + "spans": [ + { + "bbox": [ + 347, + 327, + 382, + 334 + ], + "type": "text", + "content": "Mixup (A + B)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 400, + 284, + 465, + 327 + ], + "blocks": [ + { + "bbox": [ + 402, + 272, + 462, + 284 + ], + "lines": [ + { + "bbox": [ + 402, + 272, + 462, + 284 + ], + "spans": [ + { + "bbox": [ + 402, + 272, + 462, + 284 + ], + "type": "text", + "content": "Mixup (background only) density map" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 284, + 465, + 327 + ], + "lines": [ + { + "bbox": [ + 400, + 284, + 465, + 327 + ], + "spans": [ + { + "bbox": [ + 400, + 284, + 465, + 327 + ], + "type": "image", + "image_path": "812e4c71821035264a2aefe801a99994482187486a7b629922447ceecb94e854.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 327, + 455, + 334 + ], + "lines": [ + { + "bbox": [ + 410, + 327, + 455, + 334 + ], + "spans": [ + { + "bbox": [ + 410, + 327, + 455, + 334 + ], + "type": "text", + "content": "Mixup density map" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 467, + 284, + 533, + 327 + ], + "blocks": [ + { + "bbox": [ + 467, + 284, + 533, + 327 + ], + "lines": [ + { + "bbox": [ + 467, + 284, + 533, + 327 + ], + "spans": [ + { + "bbox": [ + 467, + 284, + 533, + 327 + ], + "type": "image", + "image_path": "c18cc0f918477e80c423786c991d9d4ac67788115c4182d50c6ec08d337f33a7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 327, + 530, + 334 + ], + "lines": [ + { + "bbox": [ + 473, + 327, + 530, + 334 + ], + "spans": [ + { + "bbox": [ + 473, + 327, + 530, + 334 + ], + "type": "text", + "content": "Original B density map" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 334, + 339, + 397, + 381 + ], + "blocks": [ + { + "bbox": [ + 334, + 339, + 397, + 381 + ], + "lines": [ + { + "bbox": [ + 334, + 339, + 397, + 381 + ], + "spans": [ + { + "bbox": [ + 334, + 339, + 397, + 381 + ], + "type": "image", + "image_path": "5291994ce4a7d114b49708f949aa45d1a945e5c67d051fd6aeef41d6c8b0fb33.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 382, + 380, + 388 + ], + "lines": [ + { + "bbox": [ + 351, + 382, + 380, + 388 + ], + "spans": [ + { + "bbox": [ + 351, + 382, + 380, + 388 + ], + "type": "text", + "content": "Inpainted A" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 401, + 339, + 465, + 380 + ], + "blocks": [ + { + "bbox": [ + 401, + 339, + 465, + 380 + ], + "lines": [ + { + "bbox": [ + 401, + 339, + 465, + 380 + ], + "spans": [ + { + "bbox": [ + 401, + 339, + 465, + 380 + ], + "type": "image", + "image_path": "c03ec3d1cf06d34965e3e49dd036e8d69846e6c2c667ef7e61ed5969e698a1f3.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 382, + 463, + 388 + ], + "lines": [ + { + "bbox": [ + 403, + 382, + 463, + 388 + ], + "spans": [ + { + "bbox": [ + 403, + 382, + 463, + 388 + ], + "type": "text", + "content": "Inpainted A density map" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 333, + 394, + 397, + 436 + ], + "blocks": [ + { + "bbox": [ + 333, + 394, + 397, + 436 + ], + "lines": [ + { + "bbox": [ + 333, + 394, + 397, + 436 + ], + "spans": [ + { + "bbox": [ + 333, + 394, + 397, + 436 + ], + "type": "image", + "image_path": "379c121efee75d12876182ab18f8e8a76abeab0a0ac954f2784f4c477f588be2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 437, + 380, + 443 + ], + "lines": [ + { + "bbox": [ + 351, + 437, + 380, + 443 + ], + "spans": [ + { + "bbox": [ + 351, + 437, + 380, + 443 + ], + "type": "text", + "content": "Inpainted B" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 400, + 394, + 465, + 436 + ], + "blocks": [ + { + "bbox": [ + 400, + 394, + 465, + 436 + ], + "lines": [ + { + "bbox": [ + 400, + 394, + 465, + 436 + ], + "spans": [ + { + "bbox": [ + 400, + 394, + 465, + 436 + ], + "type": "image", + "image_path": "cfc17588551bad9a13f1450df2644d242306fcc3e81795ebf82cf6c0a203a3bb.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 437, + 463, + 443 + ], + "lines": [ + { + "bbox": [ + 403, + 437, + 463, + 443 + ], + "spans": [ + { + "bbox": [ + 403, + 437, + 463, + 443 + ], + "type": "text", + "content": "Inpainted B density map" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 473, + 341, + 533, + 437 + ], + "blocks": [ + { + "bbox": [ + 473, + 341, + 533, + 437 + ], + "lines": [ + { + "bbox": [ + 473, + 341, + 533, + 437 + ], + "spans": [ + { + "bbox": [ + 473, + 341, + 533, + 437 + ], + "type": "image", + "image_path": "1c37fd19f3b54f97b7ca17371b95fdf005f1b35f812261f75a306bef9a1ea9bb.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 449, + 555, + 525 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 525 + ], + "type": "text", + "content": "Figure 1. Comparisons of augmentations. As shown in the red bounding boxes, both types of Mixup generate disappointing density maps. Mixup is supposed to retain crowd information from two images, but actually it destroys the spatial structure. As shown in green bounding boxes, our inpainting augmentation is better suited for the crowd counting task, showing an impressive reduction in Mean Absolute Error (MAE)." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "ment [14, 50, 61] and disaster prevention [51]. The substantial demand has driven the advancement of counting methods [3, 25, 43, 45, 65]. However, a significant cost of manual annotation is unavoidable for these fully-supervised deep methods to achieve satisfactory performance. For example, the UCF-QNRF dataset with 1,535 images and 1,251,642 annotations amounts to approximately 2,000 human-hours [17]. Compared to the extremely expensive labeled data with annotations, acquiring unlabeled crowd images is much more affordable. Consequently, there is a growing interest in leveraging unlabeled data through semi-supervised crowd counting (SSCC) methods. The recent mainstream works on SSCC [21, 24, 40, 68] mostly employ the Mean Teacher [52] framework, where the teacher" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 627, + 296, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 296, + 712 + ], + "type": "text", + "content": "* Corresponding author: Yinghuan Shi. This work was supported by National Science and Technology Major Project (2023ZD0120700), NSFC Project (62222604, 62206052, 624B2063), China Postdoctoral Science Foundation (2024M750424), Fundamental Research Funds for the Central Universities (020214380120, 020214380128), State Key Laboratory Fund (ZZKT2024A14), Postdoctoral Fellowship Program of CPSF (GZC20240252), Jiangsu Funding Program for Excellent Postdoctoral Talent (2024ZB242), and Jiangsu Science and Technology Major Project (BG2024031)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24440" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "model generates pseudo-labels for unlabeled data to guide the student model. The key to achieving superior performance with SSCC methods on a given set of images, where only a small portion is annotated, lies in how we handle the unlabeled samples and how we design an appropriate model to fully exploit the information within them." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 314 + ], + "type": "text", + "content": "It has been observed that introducing similar but different samples to the training data can enlarge the support of the training distribution and thus enhance the robustness and generalizability of the model [4, 46]. In the literature, numerous data augmentation techniques have demonstrated their effectiveness in generating additional virtual samples to support the training process in both fully-supervised and semi-supervised settings [2, 63, 66]. However, we find that these off-the-shelf techniques cannot suit the crowd counting task, as they may destroy the spatial structure of crowd scenes, leading to distorted density maps, as shown in Figure 1. Therefore, a well-suited data augmentation technique specifically designed for the crowd counting task is crucial for the model to taste more from the unlabeled data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 508 + ], + "type": "text", + "content": "With the help of appropriate data augmentation, the diversity of unlabeled data during training can be significantly improved. However, this simultaneously brings new challenges for the teacher model to produce correct supervision signals. From the perspective of the model itself, we find that the Convolutional Neural Network (CNN) architecture commonly used in previous studies can be suboptimal for the crowd counting task, as it tends to overfit to local details and fails to capture global context information, which is crucial for extremely crowded, low-light, and adverse weather scenarios. From another perspective, it is also important to generate more accurate pseudo-labels while filtering out low-quality parts. Considering these two points, we explore a modern backbone network to replace CNNs and enhance it with a holistic strategy for producing and leveraging pseudo-labels, which helps to taste data better." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we propose a novel semi-supervised crowd counting method, named Taste More Taste Better (TMTB). We design a novel data augmentation technique, Inpainting Augmentation, which enhances the diversity of the dataset while keeping the integrity of the foreground regions. Observing that some regions in inpainted images are noisy and harmful, to better utilize the inpainted images, we employ a Visual State Space Model (VSSM) with an Anti-Noise classification branch for assistance, which provides inexact but accurate supervision. The proposed model generates more accurate density maps than state-of-the-art methods. Our method TMTB achieves significant improvements on several benchmark datasets, including ShanghaiTech Part A & B [69], UCF-QNRF [17], and JHU-Crowd++ [47]. We reduce the Mean Absolute Error of JHU-Crowd with " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " data labeled to 67.0, marking the first time it has been lower than 70.0. The method also presents strong generalization ability" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 73, + 428, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 428, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 428, + 83 + ], + "type": "text", + "content": "in cross-dataset evaluations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 84, + 534, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 84, + 534, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 84, + 534, + 95 + ], + "type": "text", + "content": "The main contributions of this paper are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 96, + 553, + 370 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 96, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 180 + ], + "type": "text", + "content": "- We observed that the lack of suitable data augmentation methods and the weakness of capturing global context information are bottlenecks for the crowd counting task. The off-the-shelf data methods may destroy the spatial structure of crowd scenes, which is crucial for the regression of density maps. CNNs tend to overfit local details and struggle to capture global context information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 181, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 181, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 315, + 181, + 553, + 251 + ], + "type": "text", + "content": "- We propose a novel semi-supervised crowd counting method. A data augmentation technique is designed to enhance diversity while keeping the integrity of the foreground regions. A VSSM with an Anti-Noise classification branch for assistance is employed, which provides inexact but accurate supervision, to better utilize the data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 252, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 252, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 315, + 252, + 553, + 370 + ], + "type": "text", + "content": "- The proposed method remarkably outperforms state-of-the-art methods across four benchmark datasets and three label settings. It achieves a groundbreaking reduction in the Mean Absolute Error on the JHU-Crowd dataset with only " + }, + { + "bbox": [ + 315, + 252, + 553, + 370 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 315, + 252, + 553, + 370 + ], + "type": "text", + "content": " of the data labeled, lowering it to 67.0, a decrease of " + }, + { + "bbox": [ + 315, + 252, + 553, + 370 + ], + "type": "inline_equation", + "content": "12.4\\%" + }, + { + "bbox": [ + 315, + 252, + 553, + 370 + ], + "type": "text", + "content": ". Furthermore, the method demonstrates exceptional generalization capabilities in cross-dataset Domain Generalization settings, even surpassing the performance of fully-supervised approaches, setting a new standard for excellence." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 382, + 405, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 405, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 405, + 395 + ], + "type": "text", + "content": "2. Releted Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 461 + ], + "type": "text", + "content": "Semi-supervised Crowd Counting. Fully-supervised crowd counting methods[5, 13, 17, 31, 38, 42, 49, 69] maintain the best performance in crowd counting tasks. However, fully-supervised methods require a large quantity of labeled data, which is expensive to collect and annotate." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 463, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 715 + ], + "type": "text", + "content": "Therefore, semi-supervised crowd counting methods have been proposed to address the scarcity of labeled data. The Mean Teacher [52], a successful semi-supervised learning framework that has been widely used in classification [10, 48, 64], segmentation [2, 56] and regression [8, 19, 62], is the most popular method in semi-supervised crowd counting. L2R [26] leveraged unlabeled data through a learning-to-rank framework. IRAST [27] introduced surrogate tasks, self-learning inter-related binary segmentation tasks, to improve performance. Moreover, MTCP [73] applied multiple auxiliary tasks into representation learning, which also benefit performance. DA-Count [23] employed a Transformer [53] to refine foreground features. CU [21] investigated a supervised uncertainty estimation strategy to better utilize unlabeled data. SALCrowd [68] proposed a method to select unlabeled data based on intra-scale uncertainty and inter-scale inconsistency metrics. MRC-Count [40] incorporated a fine-grained density classification task, proving the effectiveness of the classification task in crowd counting. Most previous approaches [21, 26, 27, 40, 68] are based on CNNs or par" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24441" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 59, + 541, + 252 + ], + "blocks": [ + { + "bbox": [ + 70, + 59, + 541, + 252 + ], + "lines": [ + { + "bbox": [ + 70, + 59, + 541, + 252 + ], + "spans": [ + { + "bbox": [ + 70, + 59, + 541, + 252 + ], + "type": "image", + "image_path": "76ef64fe59dac56b1b3f9d223f42aacb5703dbe5a17cc6197ed4e9dfa5228d8f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 262, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 54, + 262, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 555, + 328 + ], + "type": "text", + "content": "Figure 2. The overall framework of our method TMTB. TMTB contains the Mean Teacher framework for semi-supervised learning, VSSM as the backbone, a classification branch for predicting masks, and an inpainter for inpainting augmentation. The inpainting process is conducted periodically, which generates or updates the inpainted images. For filtering out unreliable regions, the teacher model generates the weighted mask based on the inconsistency level of the inpainted image, which is applied in " + }, + { + "bbox": [ + 54, + 262, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{inp}}" + }, + { + "bbox": [ + 54, + 262, + 555, + 328 + ], + "type": "text", + "content": ". The teacher model generates pseudolabels for unlabeled data (inpainted images included), and the student model predicts on labeled data and strongly augmented unlabeled data (inpainted images included)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 343, + 294, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 343, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 343, + 294, + 392 + ], + "type": "text", + "content": "tially [23] based on Transformers, which are weak in modeling long-range dependencies. When facing extremely dense crowd scenes, the performance of these methods is not satisfactory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 396, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 295, + 577 + ], + "type": "text", + "content": "Generative Data Augmentation. Deep learning models tend to overfit when the quantity of data is limited, which harms the generalization ability of models. Therefore, data augmentation was proposed to address this issue. For image-based tasks, basic data augmentation methods include random cropping, flipping, rotation, scaling, color jittering, and noise adding. Some advanced data augmentation methods, such as CutMix [63], Mixup [66], and AutoAugment [7], have been proposed. However, these methods are not suitable for crowd counting tasks. Both the density-based and point-based crowd counting methods are sensitive to the spatial structure of the crowd, and the above methods may destroy the spatial structure of the crowd. The damage to the integrity of the crowd will lead to a decrease in model performance. We study this in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": "Generative Data Augmentation [11, 15, 16, 20, 41] is the process of creating new synthetic data points that can be added to existing datasets. This unique approach is widely used to improve the performance of models by increasing the quantity and diversity of training data while maintaining the quality of the data. In [67], queries-only data was used to improve the multilingual task, showing the effectiveness of generative data augmentation. FreeMask [59] resorted to synthetic images from generative models to ease the burden of both data collection and annotation procedures, and achieves comparable performance to real-data-based meth" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 342, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 487 + ], + "type": "text", + "content": "ods. G-DAUG" + }, + { + "bbox": [ + 313, + 342, + 555, + 487 + ], + "type": "inline_equation", + "content": "^{\\text{C}}" + }, + { + "bbox": [ + 313, + 342, + 555, + 487 + ], + "type": "text", + "content": " [60] generated synthetic examples using pretrained language models, and selects the most informative and diverse set of examples for data augmentation, to achieve more accurate and robust learning in low-resource setting. DiffusionMix [18] employed an inpaint-and-mix strategy to generate label-preserved data for classification task. However, it requires pre-defined binary masks to concatenate the original and generated images, which is not available in crowd counting since the position and shape of the crowd are unpredictable. Compared to existing works, our contribution is to fill the gap in effective augmentation methods within the field of pixel-level predition tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 500, + 370, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 500, + 370, + 511 + ], + "spans": [ + { + "bbox": [ + 314, + 500, + 370, + 511 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 521, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 581 + ], + "type": "text", + "content": "We propose a semi-supervised crowd counting method that enhances the dataset using a diffusion-based model's inpainting technique and utilizes a VSSM (Visual State Space Model) with Anti-Noise classification branch to improve data diversity and model performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "The overall framework of our method, TMTB, is shown in Figure 2. Our method is built on the Mean Teacher framework, which is widely employed in semi-supervised learning. The teacher model shares the same architecture as the student model, and its weights are updated using the student model's weights through the Exponential Moving Average (EMA) strategy. VMamba [29], a novel VSSM (Visual State Space Model), serves as the backbone, replacing the CNN and Transformer. The image features extracted from VMamba are then fed into two independent branches: the counting branch and the classification branch, for pre" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24442" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "dictions. The counting branch predicts a density map of the image, while the classification branch predicts the indices of pre-defined count interval bins. We build a module, Inpainter, designed to periodically applying background inpainting augmentation to the training set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "text", + "content": "We first describe our novel inpainting augmentation in Section 3.1, then we discuss our network architecture in Section 3.2, and finally, we present our loss functions in Section 3.3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 196, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 196, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 196, + 204 + ], + "type": "text", + "content": "3.1. Inpainting Augmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 209, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 327 + ], + "type": "text", + "content": "Unlike classification or regression tasks, the ground truth density maps are generated by point-level annotations. The state-of-the-art data augmentation techniques, Mixup [66], CutMix [63], BCP [2], etc., are not applicable to crowd counting, since the augmentation may cause damage to ground truth density maps. We introduce the inpainting technique of diffusion-based models [15, 44] as a novel data augmentation method with the filtering capability of unreliable regions, which generates new samples, thereby expanding the diversity of dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 329, + 296, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 423 + ], + "type": "text", + "content": "For the inpainting mask " + }, + { + "bbox": [ + 55, + 329, + 296, + 423 + ], + "type": "inline_equation", + "content": "M^{\\mathrm{inp}}" + }, + { + "bbox": [ + 55, + 329, + 296, + 423 + ], + "type": "text", + "content": ", the count interval bins predicted by the classification branch are used to separate the foreground and the background of image " + }, + { + "bbox": [ + 55, + 329, + 296, + 423 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 329, + 296, + 423 + ], + "type": "text", + "content": ". We generate the positive text prompt randomly to achieve more diverse unlabeled dataset, but the negative text prompt is kept the same for all samples. The inpainting process is conducted periodically during the training process using a preset time interval " + }, + { + "bbox": [ + 55, + 329, + 296, + 423 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inp}}" + }, + { + "bbox": [ + 55, + 329, + 296, + 423 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 433, + 294, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 433, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 294, + 464 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} M ^ {\\text {i n p}} = \\mathbf {0} (\\arg \\max (\\hat {p})) , \\\\ x ^ {\\text {i n p}} = \\mathcal {S D} (x, M ^ {\\text {i n p}}, r ^ {\\text {p o s}}, r ^ {\\text {n e g}}), \\tag {1} \\\\ \\end{array}", + "image_path": "b03e06ac825dd3d8738e8bbcf694574d712c2cff4ffbbb12b6a0bb8d044aa9d1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "text", + "content": " is the predicted probability distribution from the classification branch, " + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{0}(\\cdot)" + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "text", + "content": " is the indicator function, " + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{inp}}" + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "text", + "content": " is the new image with the background inpainted, " + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{SD}(\\cdot ,\\cdot ,\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "text", + "content": " represents Stable Diffusion [44], which is a mature, excellent, and cost-effective diffusion model, and " + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{pos}}" + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "inline_equation", + "content": "r^{\\mathrm{neg}}" + }, + { + "bbox": [ + 55, + 473, + 295, + 545 + ], + "type": "text", + "content": " are the positive and negative text prompts, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Figure 3 shows examples of both good and poor inpainting samples. All new samples are considered unlabeled, as the inpainting process may add additional humans for better coherence. For well-inpainted samples, the foreground regions are well preserved, while the background regions are inpainted with different patterns, thus introducing diversity to the dataset. Even though the inpainting process is not perfect, the model can still benefit from the augmented dataset. However, failures may sometimes occur, potentially leading to the model learning from noisy inputs. The heads of humans and other features in the unreliable regions of poorly-inpainted images are fragmented and indistinct, rendering them difficult to recognize. It is necessary to filter out these unreliable regions. The EMA model calculates" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 72, + 375, + 107 + ], + "blocks": [ + { + "bbox": [ + 326, + 72, + 375, + 107 + ], + "lines": [ + { + "bbox": [ + 326, + 72, + 375, + 107 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 375, + 107 + ], + "type": "image", + "image_path": "bdc36da9be0e50dd7b749392b7f13ceb2eff1446f28a4a80bfc4e5660b7804cd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 379, + 74, + 427, + 106 + ], + "blocks": [ + { + "bbox": [ + 379, + 74, + 427, + 106 + ], + "lines": [ + { + "bbox": [ + 379, + 74, + 427, + 106 + ], + "spans": [ + { + "bbox": [ + 379, + 74, + 427, + 106 + ], + "type": "image", + "image_path": "f0f92ba5a7ab830b4627ddfea077397e786625cfe7c2d654052dab268597d3b5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 110, + 375, + 144 + ], + "blocks": [ + { + "bbox": [ + 326, + 110, + 375, + 144 + ], + "lines": [ + { + "bbox": [ + 326, + 110, + 375, + 144 + ], + "spans": [ + { + "bbox": [ + 326, + 110, + 375, + 144 + ], + "type": "image", + "image_path": "4f27d485cf7d13bde71c285d6b41f5f8b072130bfefa7e2f8393bce1b8a1a7be.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 379, + 110, + 427, + 144 + ], + "blocks": [ + { + "bbox": [ + 379, + 110, + 427, + 144 + ], + "lines": [ + { + "bbox": [ + 379, + 110, + 427, + 144 + ], + "spans": [ + { + "bbox": [ + 379, + 110, + 427, + 144 + ], + "type": "image", + "image_path": "025f93ebc72dbdf9dc285e09df03e7ae55ee027e6c7d1d7fa46f1061f8db5f22.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 326, + 146, + 375, + 179 + ], + "blocks": [ + { + "bbox": [ + 326, + 146, + 375, + 179 + ], + "lines": [ + { + "bbox": [ + 326, + 146, + 375, + 179 + ], + "spans": [ + { + "bbox": [ + 326, + 146, + 375, + 179 + ], + "type": "image", + "image_path": "2f850b867f7f3619df9b21ebb8afa6ea90b60fb34b6f235ed8ad620cf14d070e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 379, + 146, + 427, + 180 + ], + "blocks": [ + { + "bbox": [ + 379, + 146, + 427, + 180 + ], + "lines": [ + { + "bbox": [ + 379, + 146, + 427, + 180 + ], + "spans": [ + { + "bbox": [ + 379, + 146, + 427, + 180 + ], + "type": "image", + "image_path": "30ee914bbc6d7ca52c8d2e31ed79801e03264d58d6fa788b41ad04760c9c2783.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 326, + 182, + 375, + 216 + ], + "blocks": [ + { + "bbox": [ + 326, + 182, + 375, + 216 + ], + "lines": [ + { + "bbox": [ + 326, + 182, + 375, + 216 + ], + "spans": [ + { + "bbox": [ + 326, + 182, + 375, + 216 + ], + "type": "image", + "image_path": "0aa76cc00917bd5871fea08311e2040a992a332e5789600cf1280fc610a8d3e6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 217, + 354, + 223 + ], + "lines": [ + { + "bbox": [ + 347, + 217, + 354, + 223 + ], + "spans": [ + { + "bbox": [ + 347, + 217, + 354, + 223 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 379, + 182, + 427, + 216 + ], + "blocks": [ + { + "bbox": [ + 379, + 182, + 427, + 216 + ], + "lines": [ + { + "bbox": [ + 379, + 182, + 427, + 216 + ], + "spans": [ + { + "bbox": [ + 379, + 182, + 427, + 216 + ], + "type": "image", + "image_path": "50c7922964f420d591552c3474d0e7f1cc268e3df5b08a9c8e5a1c3e37c2cd1c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 217, + 407, + 223 + ], + "lines": [ + { + "bbox": [ + 399, + 217, + 407, + 223 + ], + "spans": [ + { + "bbox": [ + 399, + 217, + 407, + 223 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 439, + 73, + 488, + 107 + ], + "blocks": [ + { + "bbox": [ + 439, + 73, + 488, + 107 + ], + "lines": [ + { + "bbox": [ + 439, + 73, + 488, + 107 + ], + "spans": [ + { + "bbox": [ + 439, + 73, + 488, + 107 + ], + "type": "image", + "image_path": "3bb6e907c05a346057b0d1c786758b2af710582b9971316459c62977b44bf6f6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 440, + 110, + 488, + 144 + ], + "blocks": [ + { + "bbox": [ + 440, + 110, + 488, + 144 + ], + "lines": [ + { + "bbox": [ + 440, + 110, + 488, + 144 + ], + "spans": [ + { + "bbox": [ + 440, + 110, + 488, + 144 + ], + "type": "image", + "image_path": "fa554f5d095db2c28ba03de2179678f6459ad75050bc7b8948a88823f6adc0df.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 440, + 146, + 488, + 179 + ], + "blocks": [ + { + "bbox": [ + 440, + 146, + 488, + 179 + ], + "lines": [ + { + "bbox": [ + 440, + 146, + 488, + 179 + ], + "spans": [ + { + "bbox": [ + 440, + 146, + 488, + 179 + ], + "type": "image", + "image_path": "df8f947ff5583cc6dd5c85e84d78c94ccdb86d94aa45626b261c2f8c2132999e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 440, + 182, + 488, + 216 + ], + "blocks": [ + { + "bbox": [ + 440, + 182, + 488, + 216 + ], + "lines": [ + { + "bbox": [ + 440, + 182, + 488, + 216 + ], + "spans": [ + { + "bbox": [ + 440, + 182, + 488, + 216 + ], + "type": "image", + "image_path": "217ecad13b4843318302778885535058acb664ae03829c254a1dfd1c72fce3bd.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 217, + 468, + 223 + ], + "lines": [ + { + "bbox": [ + 460, + 217, + 468, + 223 + ], + "spans": [ + { + "bbox": [ + 460, + 217, + 468, + 223 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 227, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 271 + ], + "type": "text", + "content": "Figure 3. Inpainting samples. The left part shows well-inpainted couples, while the right part shows poorly-inpainted couples. Column (a) and (c) are the original images, while column (b) and (d) are the inpainted images." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 492, + 73, + 541, + 107 + ], + "blocks": [ + { + "bbox": [ + 492, + 73, + 541, + 107 + ], + "lines": [ + { + "bbox": [ + 492, + 73, + 541, + 107 + ], + "spans": [ + { + "bbox": [ + 492, + 73, + 541, + 107 + ], + "type": "image", + "image_path": "e58cf3d47797b103aa8baba045040c275ac652bee2954f9d6a9a53553e164f83.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 492, + 110, + 541, + 144 + ], + "blocks": [ + { + "bbox": [ + 492, + 110, + 541, + 144 + ], + "lines": [ + { + "bbox": [ + 492, + 110, + 541, + 144 + ], + "spans": [ + { + "bbox": [ + 492, + 110, + 541, + 144 + ], + "type": "image", + "image_path": "00a245736f630b4e92effb4fea8f8cec05bde2bebee3bbba959ac4a137f15246.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 492, + 145, + 541, + 179 + ], + "blocks": [ + { + "bbox": [ + 492, + 145, + 541, + 179 + ], + "lines": [ + { + "bbox": [ + 492, + 145, + 541, + 179 + ], + "spans": [ + { + "bbox": [ + 492, + 145, + 541, + 179 + ], + "type": "image", + "image_path": "a8d75ad999d2c8b7fda930ac51b7eb891eb9b607341b6fa914bf706905835be2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 492, + 182, + 541, + 216 + ], + "blocks": [ + { + "bbox": [ + 492, + 182, + 541, + 216 + ], + "lines": [ + { + "bbox": [ + 492, + 182, + 541, + 216 + ], + "spans": [ + { + "bbox": [ + 492, + 182, + 541, + 216 + ], + "type": "image", + "image_path": "7a98e93f86a4b5f8b6e3649c7a3a223aa60f53430fe51d6b2206027484d83291.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 217, + 520, + 222 + ], + "lines": [ + { + "bbox": [ + 512, + 217, + 520, + 222 + ], + "spans": [ + { + "bbox": [ + 512, + 217, + 520, + 222 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 282, + 554, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 554, + 341 + ], + "type": "text", + "content": "the inconsistency level of the inpainted image by the difference in the classification branch between the weakly augmented version and the strongly augmented version. The inconsistency level and the weighted mask are defined as follows:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 367, + 350, + 553, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 367, + 350, + 553, + 392 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} M _ {l} ^ {\\text {i n c o n}} = \\mathbf {0} \\left(\\left| \\left| p ^ {\\hat {s}} - p ^ {\\hat {w}} \\right| - l \\right|\\right), \\\\ \\omega_ {l} ^ {t} = \\operatorname {s o f t m a x} \\left(e ^ {- l \\cdot t / T ^ {\\mathrm {i n p w}}}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "c70542fd1196fcd967c11247a1c4028acfcc80c949e5649b8f1d7d95e8b42b49.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 367, + 389, + 472, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 389, + 472, + 422 + ], + "spans": [ + { + "bbox": [ + 367, + 389, + 472, + 422 + ], + "type": "interline_equation", + "content": "M ^ {\\text {i n c o n}} = \\sum_ {l = 0} ^ {L} M _ {l} ^ {\\text {i n c o n}} \\cdot \\omega_ {l} ^ {t},", + "image_path": "a9eec7bfdc0389902cf5306fbec19c0d8936b588e7c5d7d9969d2d169e5795ff.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\hat{p^s}" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\hat{p^w}" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " are the predicted probability distributions of the classification branch of EMA model in the strong and weak augmentation version, respectively. " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "M_{l}^{\\mathrm{incon}}" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " is the mask of inconsistency level " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "M^{\\mathrm{incon}}" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " is the final weighted mask, " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\omega_l^t" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " is the weight of inconsistency level " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inpw}}" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " is the parameter controlling the weight decay speed, and " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " is the number of inconsistency levels. The " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " depends on the predefined count interval bins, and since we followed the interval bins settings in [55], the " + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 433, + 555, + 567 + ], + "type": "text", + "content": " is set to 2 for the balance of diversity and reliability. The weighted mask is then used to adjust the loss associated with the inpainted image." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 575, + 525, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 525, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 525, + 587 + ], + "type": "text", + "content": "3.2. VSSM with Classification for Anti-Noise" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Visual State Space Models (VSSMs). Existing crowd counting methods often encounter challenges such as congestion, blur, and occlusion in intricate scenes. CNN-based approaches tend to overfit to local details, while Transformer-based methods excel in global modeling but face heavy computational burdens due to their quadratic complexity in terms of image sizes. These limitations motivate us to introduce the VSSM, which captures comprehensive long-range information with linear time complexity, for the crowd counting task." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24443" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 61, + 198, + 126 + ], + "blocks": [ + { + "bbox": [ + 94, + 61, + 198, + 126 + ], + "lines": [ + { + "bbox": [ + 94, + 61, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 94, + 61, + 198, + 126 + ], + "type": "image", + "image_path": "12cca87cdaa083d9ec4bc9bf936bc52d361120c616e02f6ccf8c91be6d70fa72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 201, + 61, + 287, + 126 + ], + "blocks": [ + { + "bbox": [ + 201, + 61, + 287, + 126 + ], + "lines": [ + { + "bbox": [ + 201, + 61, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 201, + 61, + 287, + 126 + ], + "type": "image", + "image_path": "aa5de0346394fe5fd1ccf8a8471f1593847b7ffbf033415979e0b8c5f51afe79.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 288, + 62, + 399, + 126 + ], + "blocks": [ + { + "bbox": [ + 288, + 62, + 399, + 126 + ], + "lines": [ + { + "bbox": [ + 288, + 62, + 399, + 126 + ], + "spans": [ + { + "bbox": [ + 288, + 62, + 399, + 126 + ], + "type": "image", + "image_path": "840b9ba167a45bf2e54efcb55b10604a76aabfa4b3f198cef7076f55bc1a0c71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 403, + 62, + 517, + 126 + ], + "blocks": [ + { + "bbox": [ + 403, + 62, + 517, + 126 + ], + "lines": [ + { + "bbox": [ + 403, + 62, + 517, + 126 + ], + "spans": [ + { + "bbox": [ + 403, + 62, + 517, + 126 + ], + "type": "image", + "image_path": "6083d173668cbaa579be6f9dae596ba72d1373e6bbff80d87604c00ceccde312.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 95, + 128, + 199, + 193 + ], + "blocks": [ + { + "bbox": [ + 95, + 128, + 199, + 193 + ], + "lines": [ + { + "bbox": [ + 95, + 128, + 199, + 193 + ], + "spans": [ + { + "bbox": [ + 95, + 128, + 199, + 193 + ], + "type": "image", + "image_path": "ec707a3dec81ae4e3e4ed7bac659fd77064f1c70d8a08addd03beac6bfcb14f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 201, + 128, + 286, + 193 + ], + "blocks": [ + { + "bbox": [ + 201, + 128, + 286, + 193 + ], + "lines": [ + { + "bbox": [ + 201, + 128, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 201, + 128, + 286, + 193 + ], + "type": "image", + "image_path": "649f28792ff5649ac80710d1d70c17e8b82d9400e6c42df4b2323d80ff00a56e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 288, + 128, + 398, + 193 + ], + "blocks": [ + { + "bbox": [ + 288, + 128, + 398, + 193 + ], + "lines": [ + { + "bbox": [ + 288, + 128, + 398, + 193 + ], + "spans": [ + { + "bbox": [ + 288, + 128, + 398, + 193 + ], + "type": "image", + "image_path": "21cb4283be0fbca87edb8969f7363f4710ecd00c3c354274ca3e5c1c9baeced6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 403, + 128, + 516, + 193 + ], + "blocks": [ + { + "bbox": [ + 403, + 128, + 516, + 193 + ], + "lines": [ + { + "bbox": [ + 403, + 128, + 516, + 193 + ], + "spans": [ + { + "bbox": [ + 403, + 128, + 516, + 193 + ], + "type": "image", + "image_path": "ce423d4f10383661bbb537b6e2c9cde346c6dd30fc12be5b602876c958b15f63.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 95, + 195, + 198, + 260 + ], + "blocks": [ + { + "bbox": [ + 95, + 195, + 198, + 260 + ], + "lines": [ + { + "bbox": [ + 95, + 195, + 198, + 260 + ], + "spans": [ + { + "bbox": [ + 95, + 195, + 198, + 260 + ], + "type": "image", + "image_path": "976da9fe03f040cc5011e9a8e5b74926a1ad48e4358b08f1d9cd6446a3ffa3f9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "type": "text", + "content": "Figure 4. Visualizations of the predictions on the test set of the JHU-Crowd++ dataset. The first row shows the input images. The second row displays the predictions of the SOTA method MRC-Crowd [40]. The third row presents the predictions of our method TMTB. All models are trained with " + }, + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "type": "text", + "content": " labeled data." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 201, + 195, + 286, + 260 + ], + "blocks": [ + { + "bbox": [ + 201, + 195, + 286, + 260 + ], + "lines": [ + { + "bbox": [ + 201, + 195, + 286, + 260 + ], + "spans": [ + { + "bbox": [ + 201, + 195, + 286, + 260 + ], + "type": "image", + "image_path": "23d7638c046dab4b4d52fa2660fd5c1c5139fa6a738f34c446bb517d956ed437.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 288, + 195, + 398, + 260 + ], + "blocks": [ + { + "bbox": [ + 288, + 195, + 398, + 260 + ], + "lines": [ + { + "bbox": [ + 288, + 195, + 398, + 260 + ], + "spans": [ + { + "bbox": [ + 288, + 195, + 398, + 260 + ], + "type": "image", + "image_path": "a3c99317cf5a12f8c7c359e3564d014847408255575c8ee6adf2545634db76f4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 403, + 195, + 516, + 260 + ], + "blocks": [ + { + "bbox": [ + 403, + 195, + 516, + 260 + ], + "lines": [ + { + "bbox": [ + 403, + 195, + 516, + 260 + ], + "spans": [ + { + "bbox": [ + 403, + 195, + 516, + 260 + ], + "type": "image", + "image_path": "39b616760553d3f6e9c0b669d9da97af9cefd58d77c32b5bd260aa75da1a44c8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 312, + 295, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 359 + ], + "type": "text", + "content": "State Space Models (SSMs) map the input stimulation " + }, + { + "bbox": [ + 55, + 312, + 295, + 359 + ], + "type": "inline_equation", + "content": "u(t)" + }, + { + "bbox": [ + 55, + 312, + 295, + 359 + ], + "type": "text", + "content": " to the output response " + }, + { + "bbox": [ + 55, + 312, + 295, + 359 + ], + "type": "inline_equation", + "content": "y(t)" + }, + { + "bbox": [ + 55, + 312, + 295, + 359 + ], + "type": "text", + "content": " by means of a hidden state " + }, + { + "bbox": [ + 55, + 312, + 295, + 359 + ], + "type": "inline_equation", + "content": "h(t)" + }, + { + "bbox": [ + 55, + 312, + 295, + 359 + ], + "type": "text", + "content": ". These processes are formulated with the following form:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 124, + 365, + 294, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 365, + 294, + 394 + ], + "spans": [ + { + "bbox": [ + 124, + 365, + 294, + 394 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} h ^ {\\prime} (t) = A h (t) + B u (t), \\\\ y (t) = C h (t) + D u (t), \\tag {3} \\\\ \\end{array}", + "image_path": "904ee855d687c30a158878f0b67eced82815283fc407405549098774b591860a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 403, + 269, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 269, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 269, + 414 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 403, + 269, + 414 + ], + "type": "inline_equation", + "content": "A, B, C," + }, + { + "bbox": [ + 55, + 403, + 269, + 414 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 403, + 269, + 414 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 403, + 269, + 414 + ], + "type": "text", + "content": " are the weighting parameters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 415, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 295, + 629 + ], + "type": "text", + "content": "Linear time-invariant SSMs (Equation (3)) lack the ability to capture the contextual information, [12] proposed a novel parameterization method for SSMs that integrates an input-dependent selection mechanism (referred to as S6). VMamba [29] replaces the S6 module, with the newly proposed 2D-Selective-Scan (SS2D) module which concurrently achieving global receptive fields, dynamic weights, and linear complexity. By adopting complementary 1D traversal paths, SS2D enables each pixel in the image to effectively integrate information from all other pixels in different directions, thereby facilitating the establishment of global receptive fields in the 2D space and generating more accurate density map. Details about the SS2D module could be found in [29]. The features extracted by the final SS2D block are then fed into two branches: one for density map regression and the other for count interval classification, which is designed for less inexact and more accurate supervision." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Exact: Regression Head. The regression head is designed to predict the density map for the input image. The density map is then summed to obtain the exact final count value. The regression head consists of up-sampling layers, convolutional layers, batch normalization layers, and ReLU activation layers. We followed MRC-Crowd [40] employing the loss function proposed in CUT [39], which significantly" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 312, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 335 + ], + "type": "text", + "content": "boosts the performance in the foreground area. It could be expressed as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 341, + 344, + 553, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 344, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 341, + 344, + 553, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {r e g}} = \\frac {1}{J} \\sum_ {j = 1} ^ {J} \\left(1 - \\operatorname {S S I M} \\left(D _ {j} (\\hat {y} \\odot M ^ {g t}\\right), \\right. \\tag {4} \\\\ \\left. D _ {j} \\left(y ^ {g t} \\odot M ^ {g t}\\right)\\right) + \\alpha \\cdot \\mathcal {L} _ {T V} \\left(\\hat {y}, y ^ {g t}\\right), \\\\ \\end{array}", + "image_path": "daf1ed249149602980ac79816eaae7ff737e1d963307c262b3194dbfe9a229d9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{SSIM}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " is the structural similarity index measure, and " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "D_{j}" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " denotes the average pooling operation that downsamples an image to " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "1 / 2^{j}" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " of its original size. The operation " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " denotes the Hadamard product, and " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "M^{gt}" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " is the binary mask that distinguishes the dense and sparse regions based on a predefined threshold " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": ". It can be obtained using the indicator function " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "M^{gt} = \\mathbf{1}(y^{gt} > \\tau)" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " is a hyperparameter that balances the two terms. " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " are set to 3 and 0.01, respectively. " + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TV}" + }, + { + "bbox": [ + 313, + 403, + 554, + 510 + ], + "type": "text", + "content": " is the total variation loss." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 511, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 701 + ], + "type": "text", + "content": "Accurate: Anti-Noise Classification Head. Many previous works [1, 31, 34, 54] have identified the problem of inaccurate ground truth annotations in crowd counting datasets, which render the ground truth density maps noisy and inaccurate. In this way, the model may learn from the noisy ground truth and generate inaccurate density maps. Inspired by previous works [23, 40, 55], we employ the Anti-Noise classification task as an auxiliary task, which is more robust in predicting inexactly and accurately. The Anti-Noise classification branch is supervised by the inexact but accurate count interval bins rather than the exact but maybe inaccurate density maps. Furthermore, we use the prediction of the this branch to generate reliable masks for inpainting samples, and we discuss this in Section 3.1. The goal of the Anti-Noise classification branch is to predict the accurate count interval bins, which are predefined by [55]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "The Anti-Noise classification head consists of an up" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24444" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "sampling layer, a convolutional layer, a ReLU activation layer and a final convolutional layer. The standard cross-entropy loss is applied. In the case of labeled data, the loss is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 127, + 295, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 127, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 119, + 127, + 295, + 159 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l s}} = \\frac {1}{N} \\sum_ {i} ^ {N} \\mathcal {H} \\left(p _ {i} ^ {g t}, \\hat {p} _ {i}\\right), \\tag {5}", + "image_path": "ce9375f3cbbc694c5561b8a568b60122bc11c75b1f9bf90bed81a72c01b8fe8c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 170, + 296, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 296, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 170, + 296, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 170, + 296, + 207 + ], + "type": "text", + "content": " is the cross-entropy loss function, " + }, + { + "bbox": [ + 55, + 170, + 296, + 207 + ], + "type": "inline_equation", + "content": "p_i^{gt}" + }, + { + "bbox": [ + 55, + 170, + 296, + 207 + ], + "type": "text", + "content": " is the ground truth probability distribution, and " + }, + { + "bbox": [ + 55, + 170, + 296, + 207 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 55, + 170, + 296, + 207 + ], + "type": "text", + "content": " is the predicted probability distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 214, + 110, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 110, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 110, + 225 + ], + "type": "text", + "content": "3.3. Losses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 232, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 434 + ], + "type": "text", + "content": "In order to make full use of the labeled data " + }, + { + "bbox": [ + 55, + 232, + 296, + 434 + ], + "type": "inline_equation", + "content": "L = \\{(x_{i}^{l},DM_{i}^{gt})\\}_{i = 1}^{N_{l}}" + }, + { + "bbox": [ + 55, + 232, + 296, + 434 + ], + "type": "text", + "content": " and unlabeled data " + }, + { + "bbox": [ + 55, + 232, + 296, + 434 + ], + "type": "inline_equation", + "content": "U = \\{(x_{i}^{u})\\}_{i = 1}^{N_{u}}" + }, + { + "bbox": [ + 55, + 232, + 296, + 434 + ], + "type": "text", + "content": ", we employ a common semi-supervised learning strategy widely used in crowd counting. The labeled samples are used to directly train the student model, while the unlabeled samples are used for consistency regularization. For labeled data, the supervised loss " + }, + { + "bbox": [ + 55, + 232, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{s}} = \\mathcal{L}_{\\mathrm{reg}}^{\\mathrm{s}} + \\mathcal{L}_{\\mathrm{cls}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 232, + 296, + 434 + ], + "type": "text", + "content": " is a combination of the regression loss (Equation (4)) and the classification loss (Equation (5)). Given the absence of a confidence score for the predicted regression map, we choose to apply consistency regularization to all unlabeled samples, rather than developing an algorithm to assess prediction uncertainty. The EMA teacher model generates pseudo-labels based on the prediction of unlabeled data with weak augmentation. The student model makes predictions based on the strong augmentation version. The error loss between the two predictions is calculated as the consistency loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 435, + 296, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 530 + ], + "type": "text", + "content": "Despite the common strong augmentation, we also introduce a patch-aligned random masking strategy [58] as a kind of strong augmentation. The patch-aligned random masking strategy randomly masks out some patches of various sizes in the image. The student model is required to predict based on masked images, while the teacher model is able to access the full image. The semi-supervised loss is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 539, + 295, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 539, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 82, + 539, + 295, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {u}} = \\operatorname {M A E} \\left(\\hat {y} ^ {s t}, \\hat {y} ^ {e m a}\\right) + \\operatorname {M A E} \\left(\\hat {p} ^ {s t}, \\hat {p} ^ {e m a}\\right), \\tag {6}", + "image_path": "532b5ab70d2b7f0ae0b34f0c92ab7df7a8a9267cd711174f13217636f00120d0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\hat{y}^{st}" + }, + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\hat{y}^{ema}" + }, + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "text", + "content": " are the predicted density maps of the student model and the teacher model, respectively. " + }, + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\hat{p}^{st}" + }, + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "inline_equation", + "content": "\\hat{p}^{ema}" + }, + { + "bbox": [ + 55, + 559, + 295, + 608 + ], + "type": "text", + "content": " are the predicted count interval bins of the student model and the teacher model, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 608, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 656 + ], + "type": "text", + "content": "Consistency loss is also applied to the inpainted images. However, to avoid the model learning from noisy regions with low quality, the loss is adjusted by the weighted mask of inconsistency level. The inpainting loss is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 664, + 310, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 310, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 310, + 689 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {i n p}} = M ^ {\\text {i n c o n}} \\times \\operatorname {M A E} \\left(\\hat {y} ^ {s t}, \\hat {y} ^ {e m a}\\right) + M ^ {\\text {i n c o n}} \\times \\operatorname {M A E} \\left(\\hat {p} ^ {s t}, \\hat {p} ^ {e m a}\\right), \\tag {7}", + "image_path": "ae3c36bc6c805f88212bf7b2591df4c3bfa7ca5c390c007ec5f05f094de4407e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "M^{\\mathrm{incon}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the weighted mask of inconsistency level mentioned in Equation (2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 97 + ], + "type": "text", + "content": "The final loss is a combination of the supervised, semi-supervised, and inpainting losses:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 105, + 553, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 105, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 370, + 105, + 553, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} ^ {\\mathrm {s}} + \\lambda_ {w} \\cdot \\mathcal {L} ^ {\\mathrm {u}} + \\lambda_ {w} \\cdot \\mathcal {L} ^ {\\text {i n p}}, \\tag {8}", + "image_path": "073ed2ade4af9b6e44f9d7a892c4208d11727d9a9cf0ae025603daf0e27da3f0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 121, + 553, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 146 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 146 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 121, + 553, + 146 + ], + "type": "inline_equation", + "content": "\\lambda_w" + }, + { + "bbox": [ + 313, + 121, + 553, + 146 + ], + "type": "text", + "content": " is the warm-up weight for the semi-supervised loss and the inpainting loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 155, + 395, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 155, + 395, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 155, + 395, + 168 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 175, + 510, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 510, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 510, + 186 + ], + "type": "text", + "content": "We evaluate our method on four public datasets:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 187, + 553, + 307 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 314, + 187, + 553, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 187, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 314, + 187, + 553, + 223 + ], + "type": "text", + "content": "- JHU-Crowd++ [47] is a comprehensive dataset with 4,372 images and 1.51 million annotations, collected under a variety of scenarios and environmental conditions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 224, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 224, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 315, + 224, + 553, + 281 + ], + "type": "text", + "content": "- UCF-QNRF [17] contains 1535 images. The dataset has a greater number of high-count crowd images and annotations, and a wider variety of scenes containing the most diverse set of viewpoints, densities and lighting variations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 283, + 553, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 283, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 314, + 283, + 553, + 307 + ], + "type": "text", + "content": "- ShanghaiTech A & B [69] are two datasets containing nearly 1,200 images and 330,000 labeled heads." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 308, + 554, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 390 + ], + "type": "text", + "content": "To achieve fair comparisons, we follow the same semi-supervised experimental protocol as previous works [23, 24, 40], including the " + }, + { + "bbox": [ + 313, + 308, + 554, + 390 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 308, + 554, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 308, + 554, + 390 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 308, + 554, + 390 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 308, + 554, + 390 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 308, + 554, + 390 + ], + "type": "text", + "content": " labeled data settings and the split of labeled and unlabeled data. The evaluation metrics are Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The detailed definitions of them are provided in the supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": "For some excessively high-resolution images, we scale them down, ensuring the longest side does not exceed 1,920 pixels. For the ground truth density map, we use geometry-adapted Gaussian kernels described in [22] for all datasets except the ShanghaiTech B dataset, and we fix the size of the kernel to 4. When training, the input images are randomly cropped to " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " pixels for all datasets except for ShanghaiTech A. The inputs from ShanghaiTech A are cropped to " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " pixels instead. The count interval bins, which should be predicted by the classification branch, are set to the same as those in [55]. The AdamW optimizer [30] with a learning rate of " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " is used to optimize the model. The weight decay is set to " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": ". The downsample factor is set to 8. The batch size is set to 8. In " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " labeled data settings, the ratio of labeled data to unlabeled data is set to " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "2:6" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": ". In the " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " labeled data setting, the ratio is set to " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "4:4" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": ". As we mentioned in Section 3.3, we use the patch-aligned random masking strategy [58] as a form of strong augmentation. The masked patch size and the masked region ratio are set to " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " and 0.3, respectively. The ablation details can be found in [40]. The teacher model is updated by the exponential moving average with a decay rate of 0.97. The weights of all losses are set to 1.0. For " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{u}}" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{inp}}" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": ", we ramp the weights from 0 to 1.0 in the first 20 epochs. The max accepted inconsistency level " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " is set to 2, and the hyperparameter " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inpw}}" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " is set to 100. The ablation study of " + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inpw}}" + }, + { + "bbox": [ + 313, + 391, + 555, + 714 + ], + "type": "text", + "content": " is conducted in Section 4.2. Each training" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24445" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 101, + 529, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparisons with the SOTA methods on main public datasets. The best results are highlighted in bold, and the second-best results are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 101, + 529, + 357 + ], + "lines": [ + { + "bbox": [ + 82, + 101, + 529, + 357 + ], + "spans": [ + { + "bbox": [ + 82, + 101, + 529, + 357 + ], + "type": "table", + "html": "
MethodsVenueLabeled PercentageJHU-Crowd++UCF-QNRFShanghaiTech AShanghaiTech B
MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓MAE ↓RMSE ↓
MT [52]NIPS'175%101.5363.5172.4284.9104.7156.919.333.2
L2R [26]CVPR'185%101.4338.8160.1272.3103.0155.420.327.6
DACount [23]ACM MM'225%82.2294.9120.2209.385.4134.512.622.8
OT-M [24]CVPR'235%82.7304.5118.4195.483.7133.312.621.5
MRC-Crowd [40]T-CSVT'245%76.5282.7101.4171.374.8117.311.717.8
TMTBOurs5%67.0252.296.3163.472.4126.410.619.6
MT [52]NIPS'1710%90.2319.3145.5250.394.5156.115.624.5
L2R [26]CVPR'1810%87.5315.3148.9249.890.3153.515.624.4
DACount [23]ACM MM'2210%75.9282.3109.0187.274.9115.511.119.1
MTCP [73]T-NNLS'2310%88.1318.7124.7206.381.3130.514.522.3
OT-M [24]CVPR'2310%73.0280.6113.1186.780.1118.510.818.2
CU [21]ICCV'2310%74.9281.7104.0164.370.8116.69.717.7
SALCrowd [68]ACM MM'2410%69.7263.5106.7171.369.7114.59.717.5
MRC-Crowd [40]T-CSVT'2410%70.7261.393.4153.267.3106.810.318.2
TMTBOurs10%66.3265.191.7156.365.7110.49.817.2
MT [52]NIPS'1740%121.5388.9147.2247.688.2151.115.925.7
L2R [26]CVPR'1840%123.6376.1145.1256.186.5148.216.825.1
SUA [33]ICCV'2140%80.7290.8130.3226.368.5121.914.120.6
DACount [23]ACM MM'2240%65.1260.091.1153.467.5110.79.614.6
OT-M [24]CVPR'2340%72.1272.0100.6167.670.7114.58.113.1
SALCrowd [68]ACM MM'2440%62.0255.195.0167.060.797.27.912.7
MRC-Crowd [40]T-CSVT'2440%60.0227.381.1131.562.195.57.813.3
TMTBOurs40%60.0240.381.4140.660.899.07.512.9
", + "image_path": "d57826a0937f83c71df6819d2c233a5ebbf3e5cd96eb5bdf80fb2c965739f71c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 375, + 295, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 400 + ], + "type": "text", + "content": "process coordinates with two inpainting background processes. The inpainting period " + }, + { + "bbox": [ + 55, + 375, + 295, + 400 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inp}}" + }, + { + "bbox": [ + 55, + 375, + 295, + 400 + ], + "type": "text", + "content": " is set to 80 epochs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 141, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 141, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 141, + 419 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 473 + ], + "type": "text", + "content": "In our main experiments, the proposed TMTB demonstrates superior performance compared to state-of-the-art methods across four public datasets with three labeled data percentages. The results are shown in Table 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "Notably, when utilizing only " + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": " labeled data, TMTB achieves the lowest MAEs of 67.0 on JHU-Crowd++, 96.3 on UCF-QNRF, 72.4 on ShanghaiTech A, and 10.6 on ShanghaiTech B, demonstrating the best performance. The result on JHU-Crowd++ not only surpasses the performance of SOTA methods with the same amount of labeled data but also demonstrates that TMTB remains superior even at the " + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": " label setting, highlighting its efficient label utilization and exceptional accuracy. With a " + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": " labeled data setting, TMTB continues to outstand, obtaining a leading MAE of 66.3 on JHU-Crowd++, 91.7 on UCF-QNRF, and 65.7 on ShanghaiTech A, further demonstrating its generalization ability and effectiveness. At " + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": " labeled data, TMTB achieves competitive results, matching the lowest MAE of 60.0 on JHU-Crowd++, and decreasing the MAE to a lowest of 7.5 on ShanghaiTech B, maintaining strong performance across the other datasets. These results underscore the capability of TMTB to achieve low error with minimal labeled data, offering substantial improvements over SOTA methods." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 347, + 403, + 525, + 475 + ], + "blocks": [ + { + "bbox": [ + 313, + 373, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 396 + ], + "type": "text", + "content": "Table 2. An study of common various augmentation strategies on ShanghaiTech-A dataset using " + }, + { + "bbox": [ + 313, + 373, + 555, + 396 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 373, + 555, + 396 + ], + "type": "text", + "content": " labeled data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 403, + 525, + 475 + ], + "lines": [ + { + "bbox": [ + 347, + 403, + 525, + 475 + ], + "spans": [ + { + "bbox": [ + 347, + 403, + 525, + 475 + ], + "type": "table", + "html": "
AugmentationMAE ↓RMSE ↓
Mixup [66]285.16431.83
Mixup [66] (background only)72.00116.85
CutMix [63]312.60462.33
BCP [2]70.99121.83
TMTB (Ours)65.74110.42
", + "image_path": "0c86f0d597292c4619c9564ac79e547a1b15f18d63e6628eb109ef2b10d14f18.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "text", + "content": "Augmentation Strategies. In order to investigate the effectiveness of the proposed inpainting augmentation strategy, we compare it with some off-the-shelf augmentation strategies, on the ShanghaiTech A dataset with " + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 496, + 555, + 568 + ], + "type": "text", + "content": " labeled data, including Mixup [66], CutMix [63], and BCP [2]. The results are shown in Table 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Mixup and CutMix both degrade the performance when applied to the whole image. The main reason is that the integrity of the foreground and background regions is destroyed, making the augmented images non-informative and hard to learn. These results demonstrate the importance of the integrity of the foreground and background regions in the crowd counting task. We then apply Mixup to the background region only, and the performance improves. This approach is somewhat similar to the inpainting augmentation, which enhances the background region while keeping the foreground region intact. However, the background regions are sampled from the domain of the original dataset, which" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24446" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 123, + 296, + 213 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "text", + "content": "Table 3. Comparisons with SOTA single domain generalization methods on Q (UCF-QNRF) and A (ShanghaiTech-A) datasets. The best results are highlighted in bold, and the second-best results are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 123, + 296, + 213 + ], + "lines": [ + { + "bbox": [ + 56, + 123, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 296, + 213 + ], + "type": "table", + "html": "
MethodsVenueSource Labeled PercentageA → QQ → A
MAE ↓RMSE ↓MAE ↓RMSE ↓
IBN [36]ECCV'18100%280.2561.0105.9174.6
SW [37]ICCV'19100%285.5431.0102.4168.8
ISW [6]CVPR'21100%215.6399.783.4136.0
DG-MAN [32]ICCV'21100%129..1238.2--
DCCUS [9]AAAI'23100%119.4216.667.4112.8
MPCount [38]CVPR'24100%115.7199.865.5110.1
TMTBOurs40%112.5190.662.4103.5
", + "image_path": "c801b6246fbf41b34d67c3fbbe0f5670acd545769488abafe7e38a4c296730ec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 284, + 296, + 362 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 296, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 277 + ], + "type": "text", + "content": "Table 4. Comparisons with SOTA domain adaptation methods on Q (UCF-QNRF) and A (ShanghaiTech-A) datasets. The best results are highlighted in bold, and the second-best results are underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 284, + 296, + 362 + ], + "lines": [ + { + "bbox": [ + 56, + 284, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 296, + 362 + ], + "type": "table", + "html": "
MethodsVenueSource Labeled PercentageTarget Domain VisibleA → QQ → A
MAE ↓RMSE ↓MAE ↓RMSE ↓
RBT [28]ACM MM'20100%✓175.0294.8--
C2MoT [57]ACM MM'21100%✓125.7218.3--
FGFD [70]ACM MM'22100%✓124.1242.070.2118.4
DAOT [71]ACM MM'23100%✓113.9215.667.0128.4
FSIM [72]T-MM'23100%✓105.3191.166.8111.5
TMTBOurs40%✗112.5190.662.4103.5
", + "image_path": "d3c04cbac2af279a542d609f1152b050d4c8bb57c87f9be412aaff64780e7116.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "text", + "content": "may not be sufficiently diverse. BCP is a recent birectical augmentation strategy designed for medical image segmentation. We employ it in the crowd counting task and find it can also improve the performance. However, the performance is not competitive enough when compared to SOTA methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": "Domain Generalization & Domain Adaptation. We compare our method with state-of-the-art single domain generalization methods and obtain remarkable results in Table 3. The results show that our method outperforms the SOTA methods on both UCF-QNRF " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": " ShanghaiTech-A and ShanghaiTech-A " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": " UCF-QNRF tasks, utilizing only " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": " labeled data on the source domain instead of " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": ". Note that IBN, SW, and ISW are originally designed for image classification or segmentation, and are adapted to crowd counting and reported by MPCount [38]. The results demonstrate the effectiveness of VSSM and inpainting augmentation in our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": "The study on domain adaptation is also conducted, and the results are shown in Table 4. Domain adaptation setting is easier than domain generalization, as the target domain is visible. It is surprising that our method achieves competitive results compared to SOTA methods, even though we only have " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": " labeled data available in the source domain and no access to the target domain. The comparisons again provide evidence of the generalization ability of TMTB." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 352, + 100, + 520, + 178 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 95 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 95 + ], + "type": "text", + "content": "Table 5. An ablation study on the effectiveness of various components on UCF-QNRF dataset using " + }, + { + "bbox": [ + 313, + 70, + 553, + 95 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 70, + 553, + 95 + ], + "type": "text", + "content": " labeled data." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 100, + 520, + 178 + ], + "lines": [ + { + "bbox": [ + 352, + 100, + 520, + 178 + ], + "spans": [ + { + "bbox": [ + 352, + 100, + 520, + 178 + ], + "type": "table", + "html": "
VSSMLsLuLinpMAE ↓RMSE ↓
XLsereg115.35188.19
✓Lsereg110.41185.52
✓Lsereg + Lcsls105.88183.87
✓Lsereg + Lcsls✓103.58178.97
✓Lsereg + Lcsls✓ωl97.21166.06
✓Lsereg + Lcsls✓ωlt96.33163.42
", + "image_path": "5415e1d88a501949bed81408c9b315bc635450522a384c925de36d3783c5a286.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 195, + 408, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 408, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 408, + 209 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "text", + "content": "Components. We conduct an ablation study to analyze the effectiveness of various components in our method on the UCF-QNRF dataset with " + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "text", + "content": " labeled data. The results are shown in Table 5. The VSSM is the core component. Compared to OT-M [24], the VSSM with supervised loss " + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "text", + "content": " improves the performance by decreasing the MAE by " + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "inline_equation", + "content": "10.57\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "text", + "content": ". After adding the semi-supervised loss " + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{u}}" + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "text", + "content": ", the performance is further improved, showing the effectiveness of our semi-supervised learning strategy. Next, we introduce our novel inpainting augmentation strategy and its mask of inconsistency level " + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "inline_equation", + "content": "M_l^{\\mathrm{uncon}}" + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "text", + "content": " but with fixed weights " + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\omega_{l}" + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "text", + "content": ". The performance is further improved, and the MAE of 97.21 surpasses the SOTA method MRC-Crowd [40] at 101.4. Finally, we employ time-based weights " + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\omega_l^t" + }, + { + "bbox": [ + 313, + 213, + 555, + 403 + ], + "type": "text", + "content": " for the inpainting loss, and the performance is significantly improved, achieving the lowest MAE of 96.33 and RMSE of 163.42." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "text", + "content": "Weight Decay Speed. We study the impact of the weight decay speed " + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inpw}}" + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "text", + "content": " on the JHU-Crowd++ dataset with 5% labeled data. The detailed influence of " + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inpw}}" + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "text", + "content": " on the MAE and RMSE is shown in supplementary material. We train and test TMTB with different " + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inpw}}" + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "text", + "content": " values ranging from 80 to 120. The results show that both MAE and RMSE decrease as " + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inpw}}" + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "text", + "content": " increases from 80 to 100. However, when " + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inpw}}" + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "text", + "content": " is greater than 100, the performance begins to degrade. Therefore, we set " + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{inpw}}" + }, + { + "bbox": [ + 313, + 404, + 556, + 513 + ], + "type": "text", + "content": " to 100 in experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 388, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 388, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 388, + 536 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 542, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 712 + ], + "type": "text", + "content": "In the literature of semi-supervised crowd counting, a number of methods based on pseudo-labeling have been explored. Despite their success, these methods still struggle to achieve satisfactory performance in extremely crowded, low-light, and adverse weather scenarios. In this work, we reveal that the lack of suitable data augmentation methods and the weakness of capturing global context information are bottlenecks. Therefore, we design a novel inpainting augmentation method, Inpainting Augmentation that works better for crowd counting, enabling tasting more. Moreover, we introduce the Visual State Space Model to capture long-range dependencies in crowd counting, enabling tasting better. Our proposed framework, TMTB, achieves superior performance compared to existing methods." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24447" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Shuai Bai, Zhiqun He, Yu Qiao, Hanzhe Hu, Wei Wu, and Junjie Yan. Adaptive dilated network with self-correction supervision for counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11514-11524, 2023. 2, 4, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "text", + "content": "[3] Lokesh Boominathan, Srinivas S S Kruthiventi, and R. Venkatesh Babu. Crowdnet: A deep convolutional network for dense crowd counting. In Proceedings of the 24th ACM International Conference on Multimedia, page 640-644, New York, NY, USA, 2016. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "text", + "content": "[4] Olivier Chapelle, Jason Weston, Léon Bottou, and Vladimir Vapnik. Vicinal risk minimization. In Advances in Neural Information Processing Systems. MIT Press, 2000. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 350 + ], + "type": "text", + "content": "[5] Lei Chen, Xinghang Gao, Fei Chao, Xiang Chang, Chih Min Lin, Xingen Gao, Shaopeng Lin, Hongyi Zhang, and Juqiang Lin. The effectiveness of a simplified model structure for crowd counting, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 418 + ], + "type": "text", + "content": "[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11580-11590, 2021. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 474 + ], + "type": "text", + "content": "[7] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "text", + "content": "[8] Weihang Dai, Xiaomeng Li, and Kwang-Ting Cheng. Semi-supervised deep regression with uncertainty consistency and variational model ensembling via bayesian neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 37(6):7304-7313, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "text", + "content": "[9] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):561-570, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "text", + "content": "[10] Yue Duan, Zhen Zhao, Lei Qi, Lei Wang, Luping Zhou, Yinghuan Shi, and Yang Gao. Mutexmatch: Semi-supervised learning with mutex-based consistency regularization. IEEE Trans. Neural Networks Learn. Syst., 35(6): 8441-8455, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[12] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling, 2024. 5, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Tao Han, Lei Bai, Lingbo Liu, and Wanli Ouyang. Steerer: Resolving scale variations for counting and localization via selective inheritance learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21848-21859, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 195 + ], + "type": "text", + "content": "[14] Marcus Handte, Muhammad Umer Iqbal, Stephan Wagner, Wolfgang Apolinarski, Pedro José Marrón, Eva Maria Muñoz Navarro, Santiago Martínez, Sara Izquierdo Barthelemy, and Mario González Fernández. Crowd density estimation for public transport vehicles. In EDBT/ICDT Workshops, pages 315-322. CiteSeer, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 555, + 341 + ], + "type": "text", + "content": "[17] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "text", + "content": "[18] Khawar Islam, Muhammad Zaigham Zaheer, Arif Mahmood, and Karthik Nandakumar. Diffusemix: Label-preserving data augmentation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27621-27630, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[19] Neal Jean, Sang Michael Xie, and Stefano Ermon. Semisupervised deep kernel learning: Regression with unlabeled data by minimizing predictive variance. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 488 + ], + "type": "text", + "content": "[20] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 491, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 533 + ], + "type": "text", + "content": "[21] Chen Li, Xiaoling Hu, Shahira Abousamra, and Chao Chen. Calibrating uncertainty for semi-supervised crowd counting. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16685-16695, 2023. 1, 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 536, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 536, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 536, + 553, + 589 + ], + "type": "text", + "content": "[22] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 591, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 656 + ], + "type": "text", + "content": "[23] Hui Lin, Zhiheng Ma, Xiaopeng Hong, Yaowei Wang, and Zhou Su. Semi-supervised crowd counting via density agency. In Proceedings of the 30th ACM International Conference on Multimedia, page 1416-1426, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[24] Wei Lin and Antoni B. Chan. Optimal transport minimization: Crowd localization on density maps for semisupervised counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21663-21673, 2023. 1, 6, 7, 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24448" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "text", + "content": "[25] Weizhe Liu, Mathieu Salzmann, and Pascal Fua. Context-aware crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[26] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Leveraging unlabeled data for crowd counting by learning to rank. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "type": "text", + "content": "[27] Yan Liu, Lingqiao Liu, Peng Wang, Pingping Zhang, and Yinjie Lei. Semi-supervised crowd counting via self-training on surrogate tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 242-259. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "type": "text", + "content": "[28] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, page 129-137, New York, NY, USA, 2020. Association for Computing Machinery. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "text", + "content": "[29] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. VMamba: Visual state space model. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 5, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 363 + ], + "type": "text", + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[31] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[32] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6630-6638, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[33] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15549-15559, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 587 + ], + "type": "text", + "content": "[34] Greg Olmschenk., Hao Tang., and Zhigang Zhu. Improving dense crowd counting convolutional neural networks using inverse k-nearest neighbor maps and multiscale upsampling. In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 5: VISAPP, pages 185-195. INSTICC, SciTePress, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 214, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 214, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 214, + 600 + ], + "type": "text", + "content": "[35] OpenAI. Gpt-4o system card, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[36] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[37] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[38] Zhuoxuan Peng and S.-H. Gary Chan. Single domain generalization for crowd counting. In Proceedings of the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28025-28034, 2024. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "text", + "content": "[39] Yifei Qian, Liangfei Zhang, Xiaopeng Hong, Carl Donovan, and Ognjen Arandjelovic. Segmentation assisted u-shaped multi-scale transformer for crowd counting. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 217 + ], + "type": "text", + "content": "[40] Yifei Qian, Xiaopeng Hong, Zhongliang Guo, Ognjen Arandjelovic, and Carl R. Donovan. Semi-supervised crowd counting with contextual modeling: Facilitating holistic understanding of crowd scenes. IEEE Transactions on Circuits and Systems for Video Technology, 34(9):8230-8241, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "type": "text", + "content": "[42] Yasiru Ranasinghe, Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M. Patel. Crowd-diff: Multi-hypothesis crowd density estimation using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12809-12819, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "text", + "content": "[43] Viresh Ranjan, Hieu Le, and Minh Hoai. Iterative crowd counting. In Proceedings of the European conference on computer vision (ECCV), pages 270-285, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 378, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 555, + 432 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "text", + "content": "[45] Miaojing Shi, Zhaohui Yang, Chao Xu, and Qijun Chen. Revisiting perspective information for efficient crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "text", + "content": "[46] Patrice Y. Simard, Yann A. LeCun, John S. Denker, and Bernard Victorri. Transformation Invariance in Pattern Recognition - Tangent Distance and Tangent Propagation, pages 235-269. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 536, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 578 + ], + "type": "text", + "content": "[47] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2594-2609, 2022. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 580, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 645 + ], + "type": "text", + "content": "[48] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In Advances in Neural Information Processing Systems, pages 596-608. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[49] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3365-3374, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24449" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[50] Hakki Soy. Edge ai-based crowd counting application for public transport stops. In Convergence of Deep Learning and Internet of Things: Computing and Technology, pages 182-205. IGI Global, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[51] Aijun Sun, Mao Liu, and Jianfeng Li. Real-time crowd massing risk supervision system based on massing crowd counting in public venue. In 2009 International Symposium on Computer Network and Multimedia Technology, pages 1-7, 2009. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[52] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[54] Jia Wan and Antoni Chan. Adaptive density map generation for crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 376 + ], + "type": "text", + "content": "[55] Changan Wang, Qingyu Song, Boshen Zhang, Yabiao Wang, Ying Tai, Xuyi Hu, Chengjie Wang, Jilin Li, Jiayi Ma, and Yang Wu. Uniformity in heterogeneity: Diving deep into count interval partition for crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3234-3242, 2021. 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 442 + ], + "type": "text", + "content": "[56] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4248-4257, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[57] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, page 658-666, New York, NY, USA, 2021. Association for Computing Machinery. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[58] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9653-9663, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "text", + "content": "[59] Lihe Yang, Xiaogang Xu, Bingyi Kang, Yinghuan Shi, and Hengshuang Zhao. Freemask: Synthetic images with dense annotations make stronger segmentation models. In Advances in Neural Information Processing Systems, pages 18659-18675. Curran Associates, Inc., 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 689 + ], + "type": "text", + "content": "[60] Yiben Yang, Chaitanya Malaviya, Jared Fernandez, Swabha Swayamdipta, Ronan Le Bras, Ji-Ping Wang, Chandra Bhagavatula, Yejin Choi, and Doug Downey. Generative data augmentation for commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1008-1025, Online, 2020. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[61] Jun Yi, Yiran Pang, Wei Zhou, Meng Zhao, and Fujian Zheng. A perspective-embedded scale-selection network" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "for crowd counting in public transportation. IEEE Transactions on Intelligent Transportation Systems, 25(5):3420-3432, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[62] Yingda Yin, Yingcheng Cai, He Wang, and Baoquan Chen. Fishermatch: Semi-supervised rotation regression via entropy-based filtering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11164-11173, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "type": "text", + "content": "[63] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features, 2019. 2, 3, 4, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 209, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 274 + ], + "type": "text", + "content": "[64] Bowen Zhang, Yidong Wang, Wenxin Hou, HAO WU, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In Advances in Neural Information Processing Systems, pages 18408-18419. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 320 + ], + "type": "text", + "content": "[65] Cong Zhang, Hongsheng Li, Xiaogang Wang, and Xiaokang Yang. Cross-scene crowd counting via deep convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "text", + "content": "[66] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 2, 3, 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 367, + 555, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 367, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 367, + 555, + 444 + ], + "type": "text", + "content": "[67] Shimao Zhang, Changjiang Gao, Wenhao Zhu, Jiajun Chen, Xin Huang, Xue Han, Junlan Feng, Chao Deng, and Shujian Huang. Getting more from less: Large language models are good spontaneous multilingual learners. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8037-8051, Miami, Florida, USA, 2024. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 445, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 511 + ], + "type": "text", + "content": "[68] Shiwei Zhang, Wei Ke, Shuai Liu, Xiaopeng Hong, and Tong Zhang. Boosting semi-supervised crowd counting with scale-based active learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8681-8690, New York, NY, USA, 2024. Association for Computing Machinery. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "text", + "content": "[69] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 569, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 635 + ], + "type": "text", + "content": "[70] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, page 5659-5668, New York, NY, USA, 2022. Association for Computing Machinery. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 636, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 713 + ], + "type": "text", + "content": "[71] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. In Proceedings of the 31st ACM International Conference on Multimedia, page 4319–4329, New York, NY, USA, 2023. Association for Computing Machinery. 8" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24450" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[72] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 26:3842-3855, 2024. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[73] Pengfei Zhu, Jingqing Li, Bing Cao, and Qinghua Hu. Multitask credible pseudo-label learning for semi-supervised crowd counting. IEEE Transactions on Neural Networks and Learning Systems, 35(8):10394-10406, 2024. 2, 7" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24451" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Taxonomy-Aware Evaluation of Vision-Language Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_content_list.json b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bb1ce7a91d4f815843793cf11380a33e0e13dab5 --- /dev/null +++ b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_content_list.json @@ -0,0 +1,1513 @@ +[ + { + "type": "text", + "text": "Taxonomy-Aware Evaluation of Vision-Language Models", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vésteinn Snæbjarnarson $^{1,2*}$ Kevin Du $^{2}$ Niklas Stoehr $^{2}$ Serge Belongie $^{1}$ Ryan Cotterell $^{2}$ Nico Lang $^{1}$ Stella Frank $^{1}$ $^{1}$ University of Copenhagen $^{2}$ ETH Zürich", + "bbox": [ + 161, + 179, + 833, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When a vision-language model (VLM) is prompted to identify an entity depicted in an image, it may answer \"I see a conifer,\" rather than the specific label NORWAY SPRUCE. This raises two issues for evaluation: Firstly, the unconstrained generated text needs to be mapped to the evaluation label space (i.e., CONIFER). Secondly, a useful classification measure should give partial credit to less-specific, but not incorrect, answers (NORWAY SPRUCE being a type of CONIFER). To meet these requirements, we propose a framework for evaluating unconstrained text predictions such as those generated from a vision-language model against a taxonomy. Specifically, we propose the use of hierarchical precision and recall measures to assess the level of correctness and specificity of predictions with regard to a taxonomy. Experimentally, we first show that existing text similarity measures do not capture taxonomic similarity well. We then develop and compare different methods to map textual VLM predictions onto a taxonomy. This allows us to compute hierarchical similarity measures between the generated text and the ground truth labels. Finally, we analyze modern VLMs on fine-grained visual classification tasks based on our proposed taxonomic evaluation scheme. Data and code are made available at https://github.com/vesteinn/vlm-eval.", + "bbox": [ + 88, + 325, + 485, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 715, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern vision-language models (VLMs) [1, 6, 12, 41, 46, 50] serve as viable one-stop shops for a wide variety of tasks at the intersection of vision and language. Given an image as input, VLMs generate unconstrained text, a key enabler of their flexibility across tasks. This flexibility, however, introduces additional challenges in their evaluation. In this paper, we study the problem of evaluating the unconstrained text generated by VLMs in the context of fine-grained visual categorization (FGVC), and put forward", + "bbox": [ + 89, + 741, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c50034492d3d7059650001866d0235dcd12abce808f5630bc933c180cd02bebe.jpg", + "image_caption": [ + "Figure 1. Vision-language-models (VLMs) as fine-grained classifiers. VLMs generate text with varying degrees of specificity and similarity to gold-standard label classes. We tackle the problem of aligning these outputs to taxonomic classes." + ], + "image_footnote": [], + "bbox": [ + 537, + 294, + 885, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "two desiderata when evaluating VLMs on FGVC (Fig. 1):", + "bbox": [ + 511, + 474, + 895, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Evaluation with Textual Awareness. When tasked with classifying entities within an image, VLMs often generate text that does not inherently fit within the predefined set of labels used in conventional evaluations. For example, a model might describe an image with the phrase pines in the snow, while the ground-truth label is simply PINE TREE. We suggest that a good evaluation should rely on fuzzy or representation-based matching techniques, which assess the similarity between the generated text and the predefined class labels, rather than whether the VLM generated the predefined class's label exactly. Such approaches have a long history in natural language processing (see, e.g., [42, 58]). We call an evaluation measure for FGVC that is compatible with unconstrained text textually aware.", + "bbox": [ + 514, + 498, + 906, + 741 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Evaluation with Hierarchical Awareness. Standard evaluation methods for VLM-based entity classification typically treat text generated by a VLM in a binary manner, i.e., it is either completely correct or incorrect [18, 31, 74]. However, many classification tasks involve label sets that are organized hierarchically, such as taxonomic trees or knowledge graphs. When such a structure is available, we suggest that evaluation measures take advantage of the structure, i.e., predictions should be judged based on their semantic", + "bbox": [ + 514, + 750, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Correspondence: vesteinnsnaebjarnarson@gmail.com.", + "bbox": [ + 107, + 887, + 467, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9109", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1cc79cace0e64468913364f8204cfa9af9d5b7bd60a46e40621d5616d3f7672.jpg", + "image_caption": [ + "Figure 2. Framework for taxonomy-aware VLM evaluation. We propose a method for granular evaluation of the open-ended image classification capabilities of VLMs. To this end, we map model predictions onto a taxonomy, e.g., sourced from a large knowledge graph. We then use hierarchical precision (hP) and recall (hR) to evaluate how taxonomically accurate and specific a model prediction is. This, in turn, lets us evaluate and compare different VLMs and prompts." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 506, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f066256a4675897a4adec1e660daa0590876e082b82e3d77a23023de4f2e7188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proximity to the correct label within the hierarchy. Fortunately, these structured label spaces are common in FGVC and other real-world classification settings. We call a VLM evaluation measure that is sensitive to taxonomic relationships hierarchically aware.", + "bbox": [ + 120, + 332, + 483, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the technical portion of the paper, we propose a novel evaluation measure for assessing VLMs' performance on FGVC tasks that incorporates both textual and hierarchical awareness. We term such an evaluation strategy taxonomy-aware. Fig. 2 illustrates the approach for taxonomy-aware VLM evaluation that we present in this paper. Our measure is based on hierarchical precision and recall [33] after placing strings generated by the VLM on a given taxonomy. To contextualize our taxonomy-aware evaluation scheme within the literature, we first evaluate existing text similarity measures from NLP to assess whether they implicitly capture taxonomic relationships. Our analysis reveals that these measures do not align well with the underlying taxonomic structure, motivating our novel taxonomy-aware evaluation framework that explicitly maps unconstrained text generated from a VLM onto a taxonomy.", + "bbox": [ + 88, + 417, + 483, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the empirical part of the paper, we explore our taxonomy-aware evaluation scheme on a range of VLMs, varying in size, training strategies, and performance, under our proposed taxonomy-aware evaluation scheme. To do so, we apply diverse prompt templates to control for accuracy, i.e., whether the model generates correct answers, and specificity, i.e., how detailed the response should be and whether to be more general when the model is not confident. In support of our evaluation, we leverage large, carefully constructed taxonomies derived from knowledge graphs extracted from Wikidata [71] and the Linnean Catalogue of Life [30]. These taxonomies span diverse domains, e.g., food, sports, animals, plants, cars, and landmarks, providing a holistic characterization of VLM behavior. Our results demonstrate that our proposed taxonomy-aware evaluation measures effectively quantify varying degrees of", + "bbox": [ + 89, + 659, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "specificity that more rigid evaluation measures like accuracy fail to capture. We find that larger commercial models respond more successfully to prompts requesting varying levels of specificity, revealing substantial differences compared to the prompt sensitivity exhibited by smaller, publicly available models. These findings illustrate how taxonomy-aware evaluation can uncover nuanced model behaviors when conventional evaluation measures cannot.", + "bbox": [ + 511, + 332, + 906, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 511, + 465, + 640, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We briefly review prior work on evaluating VLMs and textual similarity measures, with emphasis on approaches to visual categorization and hierarchical evaluation. A more general overview of the use of hierarchical labels for visual categorization is given in §17 in the supplement.", + "bbox": [ + 511, + 491, + 906, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Evaluating VLMs for Classification", + "text_level": 1, + "bbox": [ + 511, + 575, + 821, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent work has explored VLM evaluation in open-domain visual recognition and question answering, often as a byproduct of general model development. A dominant approach involves retrieving the most similar target class label to an unconstrained text using maximum inner-product search [15, 31]. The retrieved label can then be directly used for classification or further processed to extract candidate labels. In some cases [26], these extracted labels are compared to the original image using image-to-text similarity methods, such as CLIP [55]. In-context learning for classification [27] leverages large language models and a few example labels to generate or extract an image label from a longer caption. Other approaches rely on models like GPT-4 [50] to score generated answers based on a given question and reference answer [4, 34, 46, 87]. A recent method [26] expands classification label spaces by prompting models to produce increasingly specific labels using taxonomically structured templates extracted from ImageNet [18]. However, final model predictions are still evaluated using string-matching measures. Some multi-task VLM bench", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9110", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "marks include image classification tasks [9, 78], typically using exact-match evaluation or text-based measures borrowed from image captioning; see §17 in supplementary material. The Open-domain Visual Entity (OVEN) benchmark [31] relies on representation-based methods while other work avoids direct classification tasks altogether [84].", + "bbox": [ + 89, + 90, + 480, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Annotation and Prediction Similarity", + "text_level": 1, + "bbox": [ + 89, + 190, + 413, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language generation tasks such as image captioning have traditionally been evaluated using measure based on string similarity, e.g., ROUGE [42], BLEU [53],METEOR [5] and CIDER [69]. They primarily rely on surface form (string) overlap in prediction and ground truth labels, and as such, do not capture semantic similarities. This led to a family of trained measures that rely on a language model that, given a pair of inputs, predicts a similarity score, typically based on the semantic match. Examples of these are BERTScore [82], and SentenceBERT [58]. Causal and bi-directional language models have been shown to encode varied hierarchical semantic knowledge in human-interpretable metric spaces [28, 52, 54]. Natural language inference (NLI), also known as textual entailment [11, 73], can also capture hierarchical specificity judgments, since sub-categories entail their supercategories, i.e., This is a field sparrow entails This is a bird. In computer vision, this idea has been adopted as visual entailment [77].", + "bbox": [ + 89, + 213, + 483, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive vision and language models [55, 81] can be used to compare image and text pairs using top- $k$ retrieval by inner product similarity. This method relies on the intrinsic quality of the representations. While impressive, these have been shown to fail in cases requiring combining complex compositional information that was not seen together at training time [80]. More recently, prompt-based methods for scoring textual similarity have been proposed: For instance, the LLaVA models [46] are evaluated on the COCO[44] image captioning datasets by asking GPT-4 [50] to provide numerical values (1-10) for a range of measures such as relevance and accuracy. Another approach is fine-tuning large language models to act as evaluation judges [87]. This approach has also been claimed to outperform other evaluation methods for machine translation [34]. A more detailed description of the text similarity measures we compare against in this paper is given in the supplement in section $\\S 10$ .", + "bbox": [ + 89, + 484, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Taxonomy-aware Evaluation", + "text_level": 1, + "bbox": [ + 89, + 768, + 356, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Taxonomy-aware classification evaluation is a subtle and understudied problem [33, 35, 62]. It is necessary to assign partial credit to answers that share some taxonomic similarity, e.g., to answers that are within the correct sub-tree, even if they do not match the correct node exactly. Hierarchical precision (hP) and hierarchical recall (hR) [33] are two existing measures that suitably assign partial credit. Both", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "take the length of the longest common path from the root node of the taxonomy—equivalent to the size of the ancestor set—and normalize by either the length of the predicted path (hP) or the target path (hR).", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now introduce $\\mathbf{hP}$ and $\\mathbf{hR}$ formally. Let $\\Sigma$ be an alphabet of symbols, and $\\Sigma^{*}$ be the Kleene closure, i.e., the set of all strings whose symbols are drawn from $\\Sigma$ . We assume the VLM under consideration constitutes a distribution over $\\Sigma^{*}$ when it is conditioned on an image. We then define a taxonomy $\\mathcal{T}$ as a triplet $(V,\\mathcal{L},E)$ , where $V$ represents a finite set of nodes, $\\mathcal{L}\\colon V\\to \\Sigma^{*}$ is a function that maps each node to a string, called a label, and $E\\subset V\\times V$ is a set of directed edges between pairs of nodes. Additionally, we require that the directed graph $(V,E)$ forms a directed tree with a unique root $\\rho \\in V$ . The set of nodes encountered along the unique path from the taxonomy tree's root node $\\rho$ to a given node $v$ constitutes that node's ancestor set, which we denote $\\operatorname{anc}(v)$ . Formally, for any node $v\\in V$ , let $\\operatorname{par}(v)$ denote $v$ 's unique parent node in the tree; $\\operatorname{par}(\\rho)$ , i.e., with the root note $\\rho$ as an argument, is undefined. We then recursively define $\\operatorname{anc}(v)$ as:", + "bbox": [ + 511, + 152, + 906, + 409 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {a n c} (v) \\stackrel {\\text {d e f}} {=} \\left\\{\\rho \\right\\} \\quad (1 \\mathrm {a}, \\text {b a s e c a s e}; v = \\rho)\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 420, + 905, + 438 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {a n c} (v) \\stackrel {\\text {d e f}} {=} \\{v \\} \\cup \\operatorname {a n c} (\\operatorname {p a r} (v)). \\quad (1 b, \\text {i n d u c t i v e c a s e}; v \\neq \\rho)\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 440, + 905, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now discuss how to compare ground-truth and predicted nodes on a taxonomy. Let $\\{(v_n^{\\mathrm{gt}}, v_n^{\\mathrm{pr}})\\}_{n=1}^N$ be a set of pairs of nodes where $v_n^{\\mathrm{gt}}$ is the ground-truth node for the $n^{\\mathrm{th}}$ image and $v_n^{\\mathrm{pr}}$ is the node mapped from the string generated from the VLM conditioned on the $n^{\\mathrm{th}}$ image. (We describe a procedure to create this string-node mapping in §6.) We then define hP and hR over $\\{(v_n^{\\mathrm{gt}}, v_n^{\\mathrm{pr}})\\}_{n=1}^N$ as follows", + "bbox": [ + 511, + 470, + 905, + 577 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} \\mathbf {P} \\stackrel {\\text {d e f}} {=} \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\frac {\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {p r}}\\right) \\cap \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {g t}}\\right) \\right|}{\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {p r}}\\right) \\right|}, \\tag {2a}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 588, + 905, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} \\mathbf {R} \\stackrel {\\text {d e f}} {=} \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\frac {\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {p r}}\\right) \\cap \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {g t}}\\right) \\right|}{\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {g t}}\\right) \\right|}. \\tag {2b}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 633, + 903, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The hierarchical F1 (hF) is then defined as the harmonic mean of hP and hR,", + "bbox": [ + 511, + 684, + 905, + 714 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} \\mathbf {F} \\stackrel {\\text {d e f}} {=} \\frac {2 \\cdot \\mathbf {h} \\mathbf {P} \\cdot \\mathbf {h} \\mathbf {R}}{(\\mathbf {h} \\mathbf {P} + \\mathbf {h} \\mathbf {R})}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 724, + 903, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both hP and hR fall in the interval [0, 1]. The maximum is achieved when the target and predicted paths coincide. Less-than-perfect scores capture the extent to which the path deviates from the gold answer.", + "bbox": [ + 511, + 768, + 905, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hP Hierarchical precision captures the amount of incorrect information in the prediction, i.e., deviations from the correct path, relative to the number of nodes shared with the correct path.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9111", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/52e6d36176c7d6fc7bc8ffb25896dfbe5d9a2f9e0d38dfa3dd12aef346a52a42.jpg", + "image_caption": [ + "TRAIN", + "TELEPHOTO LENS", + "Prediction: A lens cover" + ], + "image_footnote": [ + "hP:1.00 hR:0.75" + ], + "bbox": [ + 143, + 88, + 285, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3179c14f5f01c22d2cb6cd3f8f0a06dd058f53363d25289a47cb8f959aa3c3d8.jpg", + "image_caption": [ + "Prediction: A mode of transport" + ], + "image_footnote": [ + "hP:0.80 hR:0.67" + ], + "bbox": [ + 290, + 108, + 419, + 160 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/87203c2258a6970c3db400abcd329e1130a5a2b1cf5e0917d91afc98e9d4924c.jpg", + "image_caption": [ + "PITH HELMET", + "Prediction: Neckwear", + "Figure 3. Illustrative examples. Hierarchical precision (hP) and recall (hR) calculations on random label pairs ( $\\S 5$ ). While hP penalizes incorrect labels, i.e., labels that are further away from the target's ancestor set, hR penalizes mistakes made higher up on the taxonomy." + ], + "image_footnote": [ + "hP:0.75 hR:0.50" + ], + "bbox": [ + 426, + 88, + 568, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9daca7581acc0e5907b52be97efed0edd0f2d74d2ea5fe30f7928cbc811cb51c.jpg", + "image_caption": [ + "STIRLING CASTLE", + "Prediction: Château de Ranrouët", + "POOL", + "Prediction: High jump" + ], + "image_footnote": [ + "hP:0.83 hR:0.83" + ], + "bbox": [ + 570, + 88, + 710, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cfa2ff0cc15dcd127bc7dbe7a3537c647d6f7b54d4d328256474468e228850d5.jpg", + "image_caption": [], + "image_footnote": [ + "hP:0.67 hR:0.67" + ], + "bbox": [ + 710, + 88, + 854, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hR Hierarchical recall measures the amount of correct information in the prediction, i.e., how much the path from the root to the predicted node overlaps with the path from the root to the ground truth node. Hierarchical recall thus penalizes missing coverage of the target path, especially early deviations. This can be seen as measuring the specificity of the prediction, in particular when the hP is 1.", + "bbox": [ + 89, + 295, + 483, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We give an example of the computation of $\\mathbf{hP}$ and $\\mathbf{hR}$ in Fig. 3, sampled from the OVEN datasets. In the case of the picture of TRAIN, we can see how only predicting MODE OF TRANSPORT leads to lower $\\mathbf{hR}$ since it lacks information, while $\\mathbf{hP}$ remains 1 since there is no incorrect information. For the sport POOL and STIRLING CASTLE, we notice how both scores drop, as incorrect nodes are predicted—these responses contain both incorrect information and some partially correct information. In the case of PITH HELMET, we see that the prediction is incorrect yet not very specific, resulting in a lower $\\mathbf{hR}$ than $\\mathbf{hP}$ . We also note that it is not possible to have a lower $\\mathbf{hP}$ than $\\mathbf{hR}$ if the ground truth nodes are leaf nodes.", + "bbox": [ + 89, + 422, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Taxonomically Linked Datasets", + "text_level": 1, + "bbox": [ + 89, + 632, + 377, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To investigate taxonomy-aware evaluation measures for text generated from a VLM, we consider two taxonomies linked to FGVC tasks. First, the Catalogue of Life taxonomy as contained in the iNaturalist21 dataset [30] is a clean, expert-curated taxonomy that maps evolutionary relationships between species and contains exactly 10,000 leaf nodes. The second is a more general taxonomy constructed from the Wikidata [71] knowledge graph, which links entities from FGVC datasets in the Open-domain Visual Entity (OVEN) benchmark [31].", + "bbox": [ + 89, + 657, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Extracting Taxonomies", + "text_level": 1, + "bbox": [ + 89, + 816, + 305, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the iNaturalist21 dataset includes a built-in taxonomy, this is not the case for other FGVC datasets. For those included in the OVEN collection, we utilize Wikidata—a large-scale knowledge graph derived from Wikipedia [71].", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike taxonomies, knowledge graphs are not constrained to tree or even directed acyclic graph structures, which necessitates the extraction of a taxonomy that fits the formal definition given in §3. To construct such a tree from Wikidata, we rely on the subclass of relation. When multiple paths exist from the root to a given node in the taxonomy, we retain the longest path. In the case of ties, we sample one of the tied paths uniformly at random. To ensure a rooted directed tree, we exclude higher-level abstract classes that introduce cycles. Because OVEN already links its datasets (§4.2) to Wikipedia entries, we can directly map them to Wikidata entities and incorporate them into our taxonomy.", + "bbox": [ + 511, + 295, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Entity Linked Datasets of Images", + "text_level": 1, + "bbox": [ + 511, + 502, + 805, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The iNaturalist21 dataset [30] was compiled and curated from the citizen science platform iNaturalist. It contains 2.7 million images spanning 10,000 species. The OVEN dataset [31] aggregates several FGVC datasets, linking them to Wikipedia identifiers and natural language questions to support the evaluation of open-domain visual entity recognition. It includes ImageNet21k-P [61, 63], iNaturalist2017 [29], Cars196 [36], SUN397 [76], Food101 [10], Sports100 [25], Aircraft [47], Oxford Flowers [49], and Google Landmarks v2 [56, 75]. We exclude iNaturalist2017 from our OVEN-based evaluations, as we separately analyze the newer iNaturalist21 dataset, which is aligned with the higher-quality Catalogue of Life taxonomy.", + "bbox": [ + 511, + 526, + 906, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. The Woes of Existing Measures", + "text_level": 1, + "bbox": [ + 511, + 738, + 799, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we evaluate whether non-taxonomic text similarity measures already encode sufficient taxonomic information, e.g., through label overlap or representational similarity. To investigate this, we conduct a targeted synthetic experiment comparing the taxonomic similarity measures (hP and hR), with textual and multimodal similarity measures from §2.2 commonly used to assess the quality of language model outputs. These measures are both string-based and representation-based.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9112", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a74c0ca1bb56a75b1527d0de4fe30db8f24b7286697d26ca744ba34f2df4b52c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MeasureiNat21OVEN
τ-hPτ-hRτ-hPτ-hR
EM0.010.070.010.01
Contained0.020.090.020.03
stringROUGE0.180.410.170.21
METEOR0.180.400.180.22
rep.BERTscore0.010.310.270.18
SentBERT0.020.250.340.31
NLI0.08-0.20-0.170.19
CLIP-t2t0.140.460.250.15
CLIP-i2t0.350.490.350.34
", + "bbox": [ + 91, + 88, + 485, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Correlation of text similarity measures with hierarchical metrics. Kendall's $\\tau$ correlations between taxonomic metrics hP, hR, and the different similarity measures on the synthetic evaluation sets. All correlations have a $p$ -value $< 0.0001$ .", + "bbox": [ + 89, + 286, + 483, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Setup. We construct two synthetic evaluation sets to determine to what extent text similarity measures reflect taxonomic structure. We consider the taxonomies from iNaturalist21 and the Wikidata (see §4) to generate controlled pairs of labels corresponding to pairs of nodes in each taxonomy. A reference node $v$ and a candidate node $v'$ are chosen such that $v$ is always a leaf node and $v'$ is at some given distance from $v$ . The distance between two nodes in the taxonomy is the length of the shortest undirected path between the two nodes in the taxonomy. To ensure a balanced dataset, we sample the node pairs such that their pairwise distance is uniformly distributed up to a maximum distance of 7, depending on the depth of the node. To evaluate hP, we sample 100,000 reference-candidate node pairs from the taxonomies. To evaluate hR in a controlled fashion, however, we additionally require that $v' \\in \\mathrm{anc}(v)$ , i.e., that $\\mathbf{hP} = 1$ . This allows us to isolate the sensitivity of the similarity measures at decreasing levels of specificity. $^{1}$ For each pair, we compute both the taxonomy-aware score and text-based measures between the corresponding labels $\\mathcal{L}(v)$ and $\\mathcal{L}(v')$ . We then measure the rank correlation between the taxonomy-aware score and various text-based measures using Kendall's $\\tau$ , which evaluates whether the similarity measure reflects taxonomic proximity under controlled conditions. By construction, we use the node label $\\mathcal{L}(v')$ as the prediction to avoid the problem of mapping unconstrained output onto the taxonomy. All of the measures we consider are text-based (see §10 in the supplement for details), except the CLIP image-to-text score, where we sample an image for the reference node from the corresponding dataset and measure CLIP model similarity between the image and the candidate label. We", + "bbox": [ + 91, + 369, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd61692ce31479b2c85a0d2d1b72c9b0912976249f281b87872c18d53a203420.jpg", + "image_caption": [ + "Figure 4. Agreement in node placement using different similarity measures. We observe variation in node placement for all measures. Only METEOR and ROUGE, and the CLIP variations when combined with Alg. 1 share predictions frequently." + ], + "image_footnote": [], + "bbox": [ + 531, + 92, + 880, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "choose high-performing models that are not prohibitively large for the model-based comparisons.2", + "bbox": [ + 511, + 378, + 906, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. Table 1 presents correlations between hP and hR on one hand and the text similarity measures from §2.2 and CLIP-i2t on the other, for the iNaturalist21 and Wiki-data labels. We observe that surface-level measures (Exact Match (EM), Contained) exhibit a low correlation with the taxonomy-aware measures hP and hR. ROUGE and METEOR, which are based on $n$ -gram overlap, capture taxonomic similarity more effectively. This is presumably due to label overlap, e.g., GRAY SEAL is a type of SEAL. The highest correlation is achieved by CLIP image-to-text (CLIP-i2t) for both datasets, followed by CLIP text-to-text (CLIP-t2t), except for hP on the iNaturalist21 dataset. In that case, CLIP-t2t fails to capture deviations from the correct taxonomic path. This suggests that CLIP's image features may encode more taxonomic information than its text representations, which can struggle due to the low-frequency vocabulary inherent to this task. NLI is an outlier, exhibiting a negative correlation in most cases. While NLI captures taxonomic similarity for closely related labels, it performs poorly on more distantly related pairs. This is visualized in Fig. 8 in the supplement.", + "bbox": [ + 511, + 428, + 908, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Summary. Overall, we observe that representation-based measures capture some notions of taxonomic distance along the path from the leaf node to the root (hR), but are, for the most part, poorly equipped to measure taxonomic distances across different subtrees (hP).", + "bbox": [ + 511, + 765, + 908, + 842 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Evaluating hP while controlling for hR is only possible for non-leaf reference nodes. Since our datasets always have leaf reference nodes, we did not explore this setting.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2All models are downloaded from. huggingface.co, i.e., BERTScore (microsoft/deberta-large-mnli), SentenceBERT (sentence-transformers/all-mpnet-base-v2), NLI (MoritzLaurer/mDeBERTa-v3-base-mnli-xnli) and CLIP (apple/DFN5B-CLIP-ViT-H-14).", + "bbox": [ + 511, + 849, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9113", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/44bff6acf4e192703565db272605c592c87eb25d4b8fe655717b9372d38d119e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MeasurehPhRhFAccuracy
Exact Match0.370.420.3917.5%
Contained0.490.500.4924.3%
BLEU0.710.530.5424.3%
METEOR0.620.650.6329.1%
ROUGE0.620.660.6330.0%
BERTScore0.450.520.4811.3%
SentBERT0.580.670.6220.0%
CLIP-t2t0.720.790.7530.3%
CLIP-i2t0.580.710.235.3%
CLIP-i2t+Alg. 10.800.810.8044.0%
CLIP-t2t+Alg. 10.790.820.8047.1%
", + "bbox": [ + 120, + 88, + 452, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Evaluating taxonomy mapping with human annotation. We report the performance of text similarity measures (directly) and CLIP image-to-text similarity for mapping VLM text predictions to the taxonomy using a subset of iNaturalist21 samples that were manually mapped to the correct nodes.", + "bbox": [ + 89, + 281, + 482, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Mapping VLM Predictions to Taxonomies", + "text_level": 1, + "bbox": [ + 89, + 359, + 467, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The key challenge for taxonomy-aware evaluation of unconstrained text generated from VLMs is mapping the text generated by a VLM onto nodes in a given taxonomy. A baseline approach to this mapping is to use text similarity measures (e.g., those in §2.2) to calculate the best match between the generated text and the labeled nodes in a taxonomy. We also develop a heuristic approach that performs this mapping using a general similarity measure as a subroutine, which makes using CLIP text-to-image similarity feasible. Specifically, this algorithm combines string-based matching with scores from a given similarity measure to position predictions onto the taxonomy. We describe this procedure in words below and give pseudocode in §11 (Alg. 1 supplement).", + "bbox": [ + 89, + 386, + 482, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our goal is to map each VLM-generated prediction string to a node in the taxonomy using a series of increasingly flexible matching strategies. Generally, we attempt exact matches first, then allow for partial overlaps or ambiguous cases, always preferring more specific nodes when multiple candidates are viable. Before matching, we normalize all predictions and node labels by lowercasing, replacing dashes with spaces, and stripping other punctuation. We then compute similarity scores between the prediction and all node labels using the CLIP similarity measures. More specifically, the mapping proceeds in stages. We first check whether any of the top- $k$ $(k = 10)$ highest-scoring nodes, according to the CLIP measure applied to their labels, appear verbatim in the prediction string; if so, we return the corresponding node. If this fails, we search for verbatim matches against all other labels in the taxonomy. If no full-string matches are found, we repeat these checks using $n$ -gram overlap for $n \\in \\{4,3,2\\}$ . If multiple candidates remain with similar scores, we apply a fallback based on score ambiguity: if the difference between the top two scores, and", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "between the top and $k^{\\text{th}}$ score, are both below threshold values, we treat the case as ambiguous. We then search for a common ancestor among the top candidates. If any node appears sufficiently often (at least four times) in the ancestor sets of multiple nodes in the top- $k$ highest-scoring nodes we return the most specific such node, defined as the one deepest in the taxonomy. If none of the prior criteria are met, we fall back to returning the single highest-scoring candidate. Thus, the algorithm always returns a node in the taxonomy. If multiple nodes match the above criteria, we return the most specific one, i.e., the node deepest in the taxonomy. The \"shared ancestor\" step is particularly beneficial for the CLIP image-to-text similarities since we do not have images corresponding to higher-order taxonomic concepts.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Mapping Quality", + "text_level": 1, + "bbox": [ + 511, + 314, + 683, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the quality of the taxonomic node proposals we measure their agreement against manually annotated node positions. We compare the similarity measures described in §2.2 (and Supplement §10) directly against all candidate node labels. We then do the same using Alg. 1, combined with the CLIP measures. To construct the test set, we gather 416 unique VLM outputs for the iNaturalist21 dataset, randomly sampled from the generations described in §7. For each output, we manually find the best matching taxonomic node given the string, e.g., matching This is a bird to AVES/BIRDS.", + "bbox": [ + 511, + 338, + 906, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. Results are shown in Tab. 2. We find string-matching measures (Exact Match and Contained) do fairly poorly, while other surface-level measures (BLEU, METEOR, and ROUGE) perform slightly better. The representation-based approaches do not perform better in general, except for CLIP text-to-text. Fig. 4 shows that the string-level measures have considerable diversity in their predictions, while the representation-based measures show more diversity. The best performance is found with CLIP text-to-text and CLIP image-to-text using the algorithm described above (Alg. 1), which results in the same hF score but slightly improved exact node match accuracy score for CLIP text-to-text + Alg. 1.", + "bbox": [ + 511, + 526, + 905, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7. Evaluating and Comparing VLMs", + "text_level": 1, + "bbox": [ + 511, + 738, + 823, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare one closed model and seven publicly available VLMs on fine-grained visual classification using our taxonomy-aware evaluation scheme. To map (unconstrained) strings generated from a VLM to nodes in the taxonomy, we use the best-performing similarity method (CLIP-t2t + Alg. 1), enabling the computation of hP, hR, and hF. These scores offer insight into the level of accuracy (hP) and specificity (hR) of VLM predictions for a given taxonomy. The resulting rankings differ significantly from", + "bbox": [ + 511, + 763, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9114", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/17295b3c219f7ae91a6947600b3ee958be156fe8fc4f47470097c65d7642f149.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model Name# ParametersTraining PipelineOpenVLM Avg.
GPT-4 [50]UnknownUnknown63.1
ILmXC2 [23]7BPretr., instr.-tuning62.2
OLM12B [79]12BPretr., instr.-tuning, RLHF54.6
QVLChat [3]9.6BPretr., instr.-tuning51.3
OLM3B [32]3BPretr., instr.-tuning47.8
LLaVA [45]7.2BPretr., instr.-tuning45.9
QwenVL [3]9.6BPretr.21.7
Fuyu [6]8BPretr.N/A
", + "bbox": [ + 89, + 88, + 480, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Overview of evaluated VLMs. Details on the models we evaluate, including number of parameters, training method, and average performance on a large vision-language benchmark.", + "bbox": [ + 89, + 236, + 482, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "those based on standard text similarity measures, highlighting the taxonomy-aware measure's ability to capture previously unmeasured aspects of model performance.", + "bbox": [ + 89, + 305, + 483, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental Setup. We compare eight VLMs that differ in model size, training configuration (pretrained-only, instruction-tuned, and preference-aligned [51]), and performance on the Open VLM Leaderboard [16], a collection of 13 VL benchmarks with 38 models. We summarize model sizes, training configuration, and average performance on the Open VLM leaderboard in Table 3. For both the OVEN-sourced FGVC datasets and iNaturalist21, we randomly sample 6000 images from the validation set. For iNaturalist21, we use the prompt What species is this?, while for OVEN the question is included in the dataset and varies by object type. To understand how well we can control hR and hP for instruction-tuned models, we construct two different sets of instructions: (1) one containing a minimal instruction for answering in a specified format, and (2) one containing an instruction to be specific. We use slight variations of the prompt formatting to fit each model (as specified by the model developers); the exact prompts are provided in the supplement (Tab. 7 and Tab. 8). Because non-instruction-tuned models do not respond meaningfully to instructions, we prompt them with a straightforward question-answer template $Q$ : {}A:. We use an orange-colored {} to indicate a variable in the template.", + "bbox": [ + 91, + 369, + 483, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 726, + 294, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How does the ranking of VLM performance change when considering taxonomy-aware measures? In", + "text_level": 1, + "bbox": [ + 89, + 750, + 483, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5, we rank eight VLMs based on their performance on FGVC according to the different similarity measures corresponding to the values given in Tab. 4 and Tab. 5 ( $\\S 14$ in supplement). We find that several text similarity measures yield similar rankings when used as direct evaluation criteria. In contrast, the taxonomy-aware measures lead to notable changes in model rankings. For instance, LLaVA ranks lowest under Exact Match for", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03ccd528a3a877782dcf4dc0c722d3d7cbbe9e755d2eaa5b0a769b31ce593594.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 95, + 895, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c137a57b5900c84316c3f9ff7471a49277b0d973d16ec6059948d53561d23e01.jpg", + "image_caption": [ + "Figure 5. Ranking of VLMs for the iNaturalist21 (top) and OVEN (bottom) datasets. We evaluate the ranking of VLMs (vertical axis) based on different evaluation measures (horizontal axis). The best model is shown in the top row. On the left, we see the model names ranked by exact match and on the right ranked by the hierarchical F1 hF. See §14, Tab. 4 in supplement for exact numbers." + ], + "image_footnote": [], + "bbox": [ + 524, + 208, + 895, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "iNaturalist21, yet achieves high hierarchical precision ( $\\mathbf{hP}$ ) and, consequently, a strong hierarchical F1 ( $\\mathbf{hF}$ ). This indicates that LLaVA deviates from the correct taxonomy path less frequently than other models. However, its low hierarchical recall ( $\\mathbf{hR}$ ) suggests that its predictions are often not specific. Another example is GPT-4, which ranks highest on all measures except $\\mathbf{hP}$ and $\\mathbf{hF}$ for iNaturalist21. Compared to QVLChat, which is ranked first in both $\\mathbf{hP}$ and $\\mathbf{hF}$ , we observe that GPT-4 makes more specific but erroneous predictions. Our taxonomy-aware measure offers a complementary perspective to traditional metrics by revealing performance differences that standard similarity-based measures overlook.", + "bbox": [ + 511, + 452, + 906, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Do taxonomy-aware measures capture prompt sensitivity? We prompt the VLMs using two distinct prompts, as described earlier, to examine how their predictions shift in terms of hierarchical precision (hP) and recall (hR), shown jointly in Fig. 7. Fuyu and QwenVL are not instruction-tuned and therefore do not respond differently, while LLaVA shows only minimal changes. In contrast, the other models clearly adjust their predictions based on the different instructions in the prompts. On the iNaturalist21 dataset, all models except GPT-4 exhibit a trade-off between hP and hR: higher precision tends to come at the cost of recall. GPT-4, however, offers both more accurate and more fine-grained classifications when prompted appropriately. On the OVEN dataset, results are more mixed: roughly half of the models improve on both axes with prompt changes.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9115", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c35983c1850958d662d8c80a44da3944dc31a7a67a0cb64e2fa9bb8889123cae.jpg", + "image_caption": [ + "Figure 6. Prompt tuning results for bird classifier. For High hP (●), we target prompts that prioritize hP. For Acc (▲) we optimize for prompts that give higher binary accuracy. This demonstrates how hP and hR can help tune a VLM application." + ], + "image_footnote": [], + "bbox": [ + 120, + 90, + 452, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These findings suggest that each application may prioritize different target values for $\\mathbf{hP}$ and $\\mathbf{hR}$ , and that models can adapt substantially to meet those needs. In such cases, prompt design becomes a powerful paradigm for tuning model behavior as further explored in the next section.", + "bbox": [ + 89, + 315, + 483, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2. An Application: High-Precision Classification", + "text_level": 1, + "bbox": [ + 89, + 400, + 478, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the usefulness of hR and hP in tuning prompts for a specific application, we prompt Llama 3.2 Vision (11B) [65] to classify birds. The application goal is to minimize the amount of wrong information in the prediction, which is equivalent to maximizing hierarchical precision but hard to formulate in terms of exact match binary classification accuracy. We use prompt-tuning with hP and hR as feedback signals, and compare to the same procedure but with accuracy as the feedback signal. We start with a basic prompt What is this an image of? Answer in the format 'A: .' and task ChatGPT [50] to improve on the prompt for 30 iterations. A system prompt describes the scores and the goal of developing a conservative classifier that gives a less specific prediction if uncertain (see §15 in the supplement for more details). In each iteration, we include past prompts along with their resulting hR and hP scores, or the accuracy score, in a few-shot learning setup. We tune the prompts with 20 images and evaluate on 200 images of birds from iNaturalist21. The results are shown in Fig. 6. While this is a toy example, we see how the taxonomy-aware measures not only enable a higher granularity of model performance analysis, but can also facilitate targeted application development.", + "bbox": [ + 91, + 422, + 483, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Discussion and Limitations", + "text_level": 1, + "bbox": [ + 89, + 784, + 344, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluating unconstrained text generated from a VLM is inherently challenging: there are countless entities to describe and even more ways to describe them. As the computer vision community shifts from closed-set recognition to open-ended recognition, evaluation increasingly resembles an NLP task. In this paper, we introduce structure into", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23719d906c712a5377c992d8275f07344af203736b7742367de560782056ad2a.jpg", + "image_caption": [ + "Figure 7. Prompt sensitivity captured by hR and hP. We compare two prompt versions, a minimal barebones version and one that asks for a taxonomically specific prediction." + ], + "image_footnote": [], + "bbox": [ + 531, + 88, + 897, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VLM evaluation by leveraging curated taxonomies and established hierarchical evaluation measures. Despite this, several limitations remain. First, placing unconstrained text on a taxonomy is a low-resource problem; there are no large-scale datasets available to evaluate our mapping approach, nor sufficient data to train a dedicated classifier that links VLM outputs to taxonomy nodes. To address this, we repurpose existing text similarity measures, along with CLIP, to perform the mapping. While our evaluation shows that this approach is non-trivial and introduces errors that affect downstream VLM evaluation, we are optimistic about future improvements, such as training on synthetic data generated by language models. Second, taxonomies, especially those derived from Wikidata, are inherently noisy. Even expert-curated taxonomies like the one used for iNaturalist21 can have uneven subtree granularity, complicating interpretations based on globally averaged measures. One potential solution would be to employ expert input to assign weighted edges to reflect semantic distances between nodes.", + "bbox": [ + 511, + 308, + 908, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 619, + 633, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text similarity measures from NLP are a natural place to look for evaluation of unconstrained VLM output. Additionally, many FGVC tasks come with side information about the class structure, namely taxonomies, which text similarity measures do not fully capture, as we have shown in this paper. To solve this gap, we have proposed a framework for using hierarchical evaluation measures to evaluate VLMs for FGVC, enabling a more fine-grained view of the specificity and accuracy of these models.", + "bbox": [ + 511, + 648, + 908, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 811, + 681, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vésteinn Snæbjarnarson, Serge Belongie, Nico Lang, and Stella Frank are supported by the Pioneer Centre for AI, DNRF grant number P1. Niklas Stoehr is supported by the Swiss Data Science Center (SDSC) PhD Fellowship.", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9116", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In Neural Information Processing Systems, pages 23716-23736, 2022. 1", + "[2] Majed A Alkhamees, Mohammed A Alnuem, Saleh M Al-Saleem, and Abdulrakeeb M Al-Ssulami. A semantic metric for concepts similarity in knowledge graphs. Journal of Information Science, 49:778-791, 2023. 4", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv, 2023. 7", + "[4] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. TouchStone: Evaluating vision-language models by language models. arXiv, 2308.16890, 2023. 2", + "[5] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 3, 2", + "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for AI agents, 2023. 1, 7", + "[7] Walter Bennette, Nathaniel Hofmann, Nathaniel Wilson, and Tyler Witter. Hierarchical open-set recognition for automatic target recognition. In 2021 IEEE Symposium Series on Computational Intelligence (SSCI). IEEE, 2021. 4", + "[8] Luca Bertinetto, Romain Mueller, Konstantinos Tertikas, Sina Samangooei, and Nicholas A Lord. Making better mistakes: Leveraging class hierarchies with deep networks. In Conference on Computer Vision and Pattern Recognition, pages 12506-12515, 2020. 4", + "[9] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. VisIT-Bench: A benchmark for vision-language instruction following inspired by real-world use. In Neural Information Processing Systems, 2023. 3", + "[10] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – Mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 4", + "[11] Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In Conference on Empirical Methods in Natural Language Processing, 2015. 3, 2" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. MiniGPT-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv, 2023. 1", + "[13] Elijah Cole, Kimberly Wilber, Grant Van Horn, Xuan Yang, Marco Fornoni, Pietro Perona, Serge Belongie, Andrew Howard, and Oisin Mac Aodha. On label granularity and object localization. In European Conference on Computer Vision, pages 604-620, 2022. 4", + "[14] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 4", + "[15] Alessandro Conti, Enrico Fini, Massimiliano Mancini, Paolo Rota, Yiming Wang, and Elisa Ricci. Vocabulary-free image classification. In Neural Information Processing Systems, 2023. 2", + "[16] OpenCompass Contributors. OpenCompass: A universal evaluation platform for foundation models, 2023. 7", + "[17] Riccardo De Lutio, Yihang She, Stefano D'Aronco, Stefania Russo, Philipp Brun, Jan D Wegner, and Konrad Schindler. Digital taxonomist: Identifying plant species in community scientists' photographs. ISPRS journal of photogrammetry and remote sensing, 182:112-121, 2021. 4", + "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 1, 2", + "[19] Jia Deng, Jonathan Krause, Alexander C. Berg, and Li Fei-Fei. Hedging your bets: Optimizing accuracy-specificity trade-offs in large scale visual recognition. In Conference on Computer Vision and Pattern Recognition, pages 3450–3457, 2012. 4", + "[20] Xiwen Dengxiong and Yu Kong. Ancestor search: Generalized open set recognition via hyperbolic side information learning. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4003-4012, 2023. 4", + "[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics, pages 4171–4186, 2019. 2", + "[22] Ankit Dhall, Anastasia Makarova, Octavian Ganea, Dario Pavllo, Michael Greeff, and Andreas Krause. Hierarchical image classification using entailment cone embeddings. In Conference on Computer Vision and Pattern Recognition, pages 836-837, 2020. 4", + "[23] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2: Mastering freeform text-image composition and comprehension in vision-language large model. arXiv, 2401.16420, 2024. 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9117", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] S. Doveh, A. Arbelle, S. Harary, E. Schwartz, R. Herzig, R. Giryes, R. Feris, R. Panda, S. Ullman, and L. Karlinsky. Teaching structured Vision & Language concepts to Vision & Language models. In Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 4", + "[25] Kaggle (Gerry). 100 sports image classification, 2022. 4", + "[26] Simon Ging, Maria Alejandra Bravo, and Thomas Brox. Open-ended VQA benchmarking of vision-language models by exploiting classification datasets and their semantic hierarchy. In International Conference on Learning Representations, 2024. 2", + "[27] Sherzod Hakimov and David Schlangen. Images in language space: Exploring the suitability of large language models for vision & language tasks. In Findings of the Association for Computational Linguistics, pages 14196-14210, 2023. 2", + "[28] Yuan He, Zhangdie Yuan, Jiaoyan Chen, and Ian Horrocks. Language models as hierarchy encoders. In Neural Information Processing Systems, 2024. 3", + "[29] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexx Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The iNaturalist species classification and detection dataset. In Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 4", + "[30] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin MacAodha. Benchmarking representation learning for natural world image collections. In Conference on Computer Vision and Pattern Recognition, pages 12879-12888, 2021. 2, 4", + "[31] Hexiang Hu, Yi Luan, Yang Chen, Urvashi Khandelwal, Mandar Joshi, Kenton Lee, Kristina Toutanova, and Ming-Wei Chang. Open-domain visual entity recognition: Towards recognizing millions of wikipedia entities. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12065–12075, 2023. 1, 2, 3, 4", + "[32] Shengding Hu, Yuge Tu, Xu Han, Ganqu Cui, Chaoqun He, Weilin Zhao, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Xinrong Zhang, Zhen Leng Thai, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, dahai li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the potential of small language models with scalable training strategies. In Conference on Language Modeling, 2024. 7", + "[33] Svetlana Kiritchenko, Stan Matwin, Richard Nock, and Fazel Famili. Learning and evaluation in the presence of class hierarchies: Application to text categorization. In Canadian AI, 2006. 2, 3", + "[34] Tom Kocmi and Christian Federmann. Large language models are state-of-the-art evaluators of translation quality. In Conference of the European Association for Machine Translation, pages 193–203, 2023. 2, 3", + "[35] Aris Kosmopoulos, Ioannis Partalas, Eric Gaussier, Georgios Paliouras, and Ion Androutsopoulos. Evaluation measures for hierarchical classification: A unified view and novel approaches. Data Mining and Knowledge Discovery, 29(3): 820-865, 2015. 3", + "[36] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D object representations for fine-grained categorization. In" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision Workshops, pages 554-561, 2013. 4", + "[37] Suren Kumar and Rui Zheng. Hierarchical category detector for clothing recognition from visual data. In International Conference on Computer Vision Workshops, pages 2306-2312, 2017. 4", + "[38] Nico Lang, Vésteinn Snæbjarnarson, Elijah Cole, Oisin Mac Aodha, Christian Igel, and Serge Belongie. From coarse to fine-grained open-set recognition. In Conference on Computer Vision and Pattern Recognition, pages 17804-17814, 2024. 4", + "[39] Kibok Lee, Kimin Lee, Kyle Min, Yuting Zhang, Jinwoo Shin, and Honglak Lee. Hierarchical novelty detection for visual object recognition. In Conference on Computer Vision and Pattern Recognition, pages 1034-1042, 2018. 4", + "[40] Clément Lefebvre and Niklas Stoehr. Rethinking the event coding pipeline with prompt entailment. In EACL Workshop Fact Extraction and VERIFICATION, 2023. 2", + "[41] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, 2023. 1", + "[42] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81. Association for Computational Linguistics, 2004. 1, 3, 2", + "[43] Dekang Lin. An information-theoretic definition of similarity. In International Conference on Machine Learning, page 296-304, 1998. 4", + "[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, pages 740-755, 2014. 3", + "[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Conference on Computer Vision and Pattern Recognition, 2023. 7", + "[46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Neural Information Processing Systems, 2023. 1, 2, 3", + "[47] Subhransu Maji, Esa Rahtu, Juho Kannala, Mathew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv, 1306.5151, 2013. 4", + "[48] Kenneth Marino, Ruslan Salakhutdinov, and Abhinav Gupta. The more you know: Using knowledge graphs for image classification. In Computer Vision and Pattern Recognition, 2016. 4", + "[49] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 4", + "[50] OpenAI. GPT-4 technical report, 2023. 1, 2, 3, 7, 8", + "[51] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9118", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Neural Information Processing Systems, 2022. 7", + "[52] Jeff Z. Pan, Simon Razniewski, Jan-Christoph Kalo, Sneha Singhania, Jiaoyan Chen, Stefan Dietze, Hajira Jabeen, Janna Omeliyanenko, Wen Zhang, Matteo Lissandrini, Russia Biswas, Gerard de Melo, Angela Bonifati, Edlira Vakaj, Mauro Dragoni, and Damien Graux. Large language models and knowledge graphs: Opportunities and challenges. Transactions on Graph Data and Knowledge, 2023. 3", + "[53] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. In Annual Meeting on Association for Computational Linguistics, page 311-318, 2002. 3, 2", + "[54] Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Conference on Empirical Methods in Natural Language Processing, pages 2463-2473, 2019. 3", + "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3", + "[56] Elias Ramzi, Nicolas Audebert, Clement Rambour, Andre Araujo, Xavier Bitot, and Nicolas Thome. Optimization of rank losses for image retrieval. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4", + "[57] Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees GM Snoek. SelEx: Self-expertise in fine-grained generalized category discovery. In European Conference on Computer Vision, 2024. 4", + "[58] Nils Reimers and Iryna Gurevych. Sentence-BERT: sentence embeddings using siamese BERT-networks. In Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3980-3990, 2019. 1, 3, 2", + "[59] Philip Resnik. A class-based approach to lexical discovery. In Annual Meeting of the Association for Computational Linguistics, pages 327-329, 1992. 4", + "[60] Philip Resnik. Using information content to evaluate semantic similarity in a taxonomy. In International Joint Conference on Artificial Intelligence, page 448-453, 1995. 4", + "[61] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. In Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 4", + "[62] Kevin Riehl, Michael Neunteufel, and Martin Hemberg. Hierarchical confusion matrix for classification performance evaluation. Journal of the Royal Statistical Society Series C: Applied Statistics, 72(5):1394-1412, 2023. 3", + "[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + ], + "bbox": [ + 91, + 90, + 483, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 4", + "[64] Paolo Simeone, Raul Santos-Rodriguez, Matt McVicar, Jeffrey Lijffijt, and Tijl De Bie. Hierarchical novelty detection. In Advances in Intelligent Data Analysis, pages 310-321, 2017. 4", + "[65] Hugo Touvron, Aaron Grattafori, Michael Auli, et al. The Llama 3 herd of models. arXiv, 2407.21783, 2024. 8", + "[66] Mehmet Ozgur Turkoglu, Stefano D'Aronco, Gregor Perich, Frank Liebisch, Constantin Streit, Konrad Schindler, and Jan Dirk Wegner. Crop mapping from image time series: Deep learning with multi-scale label hierarchies. Remote Sensing of Environment, 264:112603, 2021. 4", + "[67] Grant Van Horn, Steve Branson, Scott Loarie, Serge Belongie, and Pietro Perona. Lean multiclass crowdsourcing. In Conference on Computer Vision and Pattern Recognition, 2018. 4", + "[68] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In International Conference on Learning Representations, 2022. 4", + "[69] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Conference on Computer Vision and Pattern Recognition, 2015. 3", + "[70] Nakul Verma, Dhruv Mahajan, Sundararajan Sellamanickam, and Vinod Nair. Learning hierarchical similarity metrics. In Conference on Computer Vision and Pattern Recognition, pages 2280-2287, 2012. 4", + "[71] Denny Vrandecic and Markus Krötzsch. Wikidata: A free collaborative knowledgebase. Communications of the ACM, 57(10):78-85, 2014. 2, 4", + "[72] Cunxiang Wang, Sirui Cheng, Qipeng Guo, Yuanhao Yue, Bowen Ding, Zhikun Xu, Yidong Wang, Xiangkun Hu, Zheng Zhang, and Yue Zhang. Evaluating open-QA evaluation. In Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2", + "[73] Sinong Wang, Han Fang, Madian Khabsa, Hanzi Mao, and Hao Ma. Entailment as few-shot learner. arXiv, 2104.14690, 2021. 3, 2", + "[74] Xiu-Shen Wei, Yi-Zhen Song, Oisin Aodha, Jianxin Wu, Yuxin Peng, Jinhui Tang, Jian Yang, and Serge Belongie. Fine-grained image analysis with deep learning: A survey. IEEE Transactions on Pattern Analysis & Machine Intelligence, 44(12):8927-8948, 2022. 1", + "[75] Tobias Weyand, Andre Araujo, Bingyi Cao, and Jacjk Sim. Google Landmarks Dataset v2: A Large-Scale Benchmark for Instance-Level Recognition and Retrieval. In Conference on Computer Vision and Pattern Recognition, 2020. 4", + "[76] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 4", + "[77] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment: A novel task for fine-grained image understanding, 2019. 3, 2" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9119", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Fanqing Meng, Siyuan Huang, Meng Lei, Ping Luo, and Yu Qiao. LVLM-eHub: A comprehensive evaluation benchmark for large vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3", + "[79] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. RLHF-V: Towards trustworthy MLLMs via behavior alignment from fine-grained correctional human feedback. In Conference on Computer Vision and Pattern Recognition, 2024. 7", + "[80] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In International Conference on Learning Representations, 2023. 3", + "[81] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In International Conference on Computer Vision, pages 11941-11952, 2023. 3", + "[82] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. BERTScore: Evaluating text generation with BERT. In International Conference on Learning Representations, 2020. 3, 2", + "[83] Bingchen Zhao, Nico Lang, Serge Belongie, and Oisin Mac Aodha. Labeled data selection for category discovery. In European Conference on Computer Vision, 2024. 4", + "[84] Wangchunshu Zhou, Yan Zeng, Shizhe Diao, and Xinsong Zhang. VLUE: A multi-task multi-dimension benchmark for evaluating vision-language pre-training. In International Conference on Machine Learning, pages 27395-27411, 2022. 3", + "[85] Ganggao Zhu and Carlos A. Iglesias. Computing semantic similarity of concepts in knowledge graphs. IEEE Transactions on Knowledge and Data Engineering, 29(1):72-85, 2017. 4", + "[86] Ganggao Zhu and Carlos A. Iglesias. Exploiting semantic similarity for named entity disambiguation in knowledge graphs. Expert Systems with Applications, 101:8-24, 2018. 4", + "[87] Lianghui Zhu, Xinggang Wang, and Xinlong Wang. JudgeLM: Fine-tuned large language models are scalable judges. In International Conference on Learning Representations, 2025. 2, 3" + ], + "bbox": [ + 91, + 92, + 480, + 710 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9120", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Taxonomy-Aware Evaluation of Vision-Language Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_model.json b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d415a59914607c9b8249060305c72615b289aefd --- /dev/null +++ b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_model.json @@ -0,0 +1,2567 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Taxonomy-Aware Evaluation of Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.18, + 0.834, + 0.244 + ], + "angle": 0, + "content": "Vésteinn Snæbjarnarson\\(^{1,2*}\\) Kevin Du\\(^{2}\\) Niklas Stoehr\\(^{2}\\) Serge Belongie\\(^{1}\\) Ryan Cotterell\\(^{2}\\) Nico Lang\\(^{1}\\) Stella Frank\\(^{1}\\) \n\\(^{1}\\)University of Copenhagen \\(^{2}\\)ETH Zürich" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.294, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.691 + ], + "angle": 0, + "content": "When a vision-language model (VLM) is prompted to identify an entity depicted in an image, it may answer \"I see a conifer,\" rather than the specific label NORWAY SPRUCE. This raises two issues for evaluation: Firstly, the unconstrained generated text needs to be mapped to the evaluation label space (i.e., CONIFER). Secondly, a useful classification measure should give partial credit to less-specific, but not incorrect, answers (NORWAY SPRUCE being a type of CONIFER). To meet these requirements, we propose a framework for evaluating unconstrained text predictions such as those generated from a vision-language model against a taxonomy. Specifically, we propose the use of hierarchical precision and recall measures to assess the level of correctness and specificity of predictions with regard to a taxonomy. Experimentally, we first show that existing text similarity measures do not capture taxonomic similarity well. We then develop and compare different methods to map textual VLM predictions onto a taxonomy. This allows us to compute hierarchical similarity measures between the generated text and the ground truth labels. Finally, we analyze modern VLMs on fine-grained visual classification tasks based on our proposed taxonomic evaluation scheme. Data and code are made available at https://github.com/vesteinn/vlm-eval." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.717, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Modern vision-language models (VLMs) [1, 6, 12, 41, 46, 50] serve as viable one-stop shops for a wide variety of tasks at the intersection of vision and language. Given an image as input, VLMs generate unconstrained text, a key enabler of their flexibility across tasks. This flexibility, however, introduces additional challenges in their evaluation. In this paper, we study the problem of evaluating the unconstrained text generated by VLMs in the context of fine-grained visual categorization (FGVC), and put forward" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.295, + 0.886, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.413, + 0.907, + 0.469 + ], + "angle": 0, + "content": "Figure 1. Vision-language-models (VLMs) as fine-grained classifiers. VLMs generate text with varying degrees of specificity and similarity to gold-standard label classes. We tackle the problem of aligning these outputs to taxonomic classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.476, + 0.896, + 0.492 + ], + "angle": 0, + "content": "two desiderata when evaluating VLMs on FGVC (Fig. 1):" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.499, + 0.907, + 0.742 + ], + "angle": 0, + "content": "Evaluation with Textual Awareness. When tasked with classifying entities within an image, VLMs often generate text that does not inherently fit within the predefined set of labels used in conventional evaluations. For example, a model might describe an image with the phrase pines in the snow, while the ground-truth label is simply PINE TREE. We suggest that a good evaluation should rely on fuzzy or representation-based matching techniques, which assess the similarity between the generated text and the predefined class labels, rather than whether the VLM generated the predefined class's label exactly. Such approaches have a long history in natural language processing (see, e.g., [42, 58]). We call an evaluation measure for FGVC that is compatible with unconstrained text textually aware." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Evaluation with Hierarchical Awareness. Standard evaluation methods for VLM-based entity classification typically treat text generated by a VLM in a binary manner, i.e., it is either completely correct or incorrect [18, 31, 74]. However, many classification tasks involve label sets that are organized hierarchically, such as taxonomic trees or knowledge graphs. When such a structure is available, we suggest that evaluation measures take advantage of the structure, i.e., predictions should be judged based on their semantic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.468, + 0.901 + ], + "angle": 0, + "content": "*Correspondence: vesteinnsnaebjarnarson@gmail.com." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9109" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.508, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.908, + 0.307 + ], + "angle": 0, + "content": "Figure 2. Framework for taxonomy-aware VLM evaluation. We propose a method for granular evaluation of the open-ended image classification capabilities of VLMs. To this end, we map model predictions onto a taxonomy, e.g., sourced from a large knowledge graph. We then use hierarchical precision (hP) and recall (hR) to evaluate how taxonomically accurate and specific a model prediction is. This, in turn, lets us evaluate and compare different VLMs and prompts." + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.333, + 0.484, + 0.41 + ], + "angle": 0, + "content": "proximity to the correct label within the hierarchy. Fortunately, these structured label spaces are common in FGVC and other real-world classification settings. We call a VLM evaluation measure that is sensitive to taxonomic relationships hierarchically aware." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.418, + 0.484, + 0.659 + ], + "angle": 0, + "content": "In the technical portion of the paper, we propose a novel evaluation measure for assessing VLMs' performance on FGVC tasks that incorporates both textual and hierarchical awareness. We term such an evaluation strategy taxonomy-aware. Fig. 2 illustrates the approach for taxonomy-aware VLM evaluation that we present in this paper. Our measure is based on hierarchical precision and recall [33] after placing strings generated by the VLM on a given taxonomy. To contextualize our taxonomy-aware evaluation scheme within the literature, we first evaluate existing text similarity measures from NLP to assess whether they implicitly capture taxonomic relationships. Our analysis reveals that these measures do not align well with the underlying taxonomic structure, motivating our novel taxonomy-aware evaluation framework that explicitly maps unconstrained text generated from a VLM onto a taxonomy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the empirical part of the paper, we explore our taxonomy-aware evaluation scheme on a range of VLMs, varying in size, training strategies, and performance, under our proposed taxonomy-aware evaluation scheme. To do so, we apply diverse prompt templates to control for accuracy, i.e., whether the model generates correct answers, and specificity, i.e., how detailed the response should be and whether to be more general when the model is not confident. In support of our evaluation, we leverage large, carefully constructed taxonomies derived from knowledge graphs extracted from Wikidata [71] and the Linnean Catalogue of Life [30]. These taxonomies span diverse domains, e.g., food, sports, animals, plants, cars, and landmarks, providing a holistic characterization of VLM behavior. Our results demonstrate that our proposed taxonomy-aware evaluation measures effectively quantify varying degrees of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.453 + ], + "angle": 0, + "content": "specificity that more rigid evaluation measures like accuracy fail to capture. We find that larger commercial models respond more successfully to prompts requesting varying levels of specificity, revealing substantial differences compared to the prompt sensitivity exhibited by smaller, publicly available models. These findings illustrate how taxonomy-aware evaluation can uncover nuanced model behaviors when conventional evaluation measures cannot." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.642, + 0.484 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.568 + ], + "angle": 0, + "content": "We briefly review prior work on evaluating VLMs and textual similarity measures, with emphasis on approaches to visual categorization and hierarchical evaluation. A more general overview of the use of hierarchical labels for visual categorization is given in §17 in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.822, + 0.592 + ], + "angle": 0, + "content": "2.1. Evaluating VLMs for Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent work has explored VLM evaluation in open-domain visual recognition and question answering, often as a byproduct of general model development. A dominant approach involves retrieving the most similar target class label to an unconstrained text using maximum inner-product search [15, 31]. The retrieved label can then be directly used for classification or further processed to extract candidate labels. In some cases [26], these extracted labels are compared to the original image using image-to-text similarity methods, such as CLIP [55]. In-context learning for classification [27] leverages large language models and a few example labels to generate or extract an image label from a longer caption. Other approaches rely on models like GPT-4 [50] to score generated answers based on a given question and reference answer [4, 34, 46, 87]. A recent method [26] expands classification label spaces by prompting models to produce increasingly specific labels using taxonomically structured templates extracted from ImageNet [18]. However, final model predictions are still evaluated using string-matching measures. Some multi-task VLM bench" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9110" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.184 + ], + "angle": 0, + "content": "marks include image classification tasks [9, 78], typically using exact-match evaluation or text-based measures borrowed from image captioning; see §17 in supplementary material. The Open-domain Visual Entity (OVEN) benchmark [31] relies on representation-based methods while other work avoids direct classification tasks altogether [84]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.191, + 0.414, + 0.208 + ], + "angle": 0, + "content": "2.2. Annotation and Prediction Similarity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.485, + 0.485 + ], + "angle": 0, + "content": "Language generation tasks such as image captioning have traditionally been evaluated using measure based on string similarity, e.g., ROUGE [42], BLEU [53],METEOR [5] and CIDER [69]. They primarily rely on surface form (string) overlap in prediction and ground truth labels, and as such, do not capture semantic similarities. This led to a family of trained measures that rely on a language model that, given a pair of inputs, predicts a similarity score, typically based on the semantic match. Examples of these are BERTScore [82], and SentenceBERT [58]. Causal and bi-directional language models have been shown to encode varied hierarchical semantic knowledge in human-interpretable metric spaces [28, 52, 54]. Natural language inference (NLI), also known as textual entailment [11, 73], can also capture hierarchical specificity judgments, since sub-categories entail their supercategories, i.e., This is a field sparrow entails This is a bird. In computer vision, this idea has been adopted as visual entailment [77]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Contrastive vision and language models [55, 81] can be used to compare image and text pairs using top-\\(k\\) retrieval by inner product similarity. This method relies on the intrinsic quality of the representations. While impressive, these have been shown to fail in cases requiring combining complex compositional information that was not seen together at training time [80]. More recently, prompt-based methods for scoring textual similarity have been proposed: For instance, the LLaVA models [46] are evaluated on the COCO[44] image captioning datasets by asking GPT-4 [50] to provide numerical values (1-10) for a range of measures such as relevance and accuracy. Another approach is fine-tuning large language models to act as evaluation judges [87]. This approach has also been claimed to outperform other evaluation methods for machine translation [34]. A more detailed description of the text similarity measures we compare against in this paper is given in the supplement in section \\(\\S 10\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.357, + 0.787 + ], + "angle": 0, + "content": "3. Taxonomy-aware Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Taxonomy-aware classification evaluation is a subtle and understudied problem [33, 35, 62]. It is necessary to assign partial credit to answers that share some taxonomic similarity, e.g., to answers that are within the correct sub-tree, even if they do not match the correct node exactly. Hierarchical precision (hP) and hierarchical recall (hR) [33] are two existing measures that suitably assign partial credit. Both" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "take the length of the longest common path from the root node of the taxonomy—equivalent to the size of the ancestor set—and normalize by either the length of the predicted path (hP) or the target path (hR)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.41 + ], + "angle": 0, + "content": "We now introduce \\(\\mathbf{hP}\\) and \\(\\mathbf{hR}\\) formally. Let \\(\\Sigma\\) be an alphabet of symbols, and \\(\\Sigma^{*}\\) be the Kleene closure, i.e., the set of all strings whose symbols are drawn from \\(\\Sigma\\). We assume the VLM under consideration constitutes a distribution over \\(\\Sigma^{*}\\) when it is conditioned on an image. We then define a taxonomy \\(\\mathcal{T}\\) as a triplet \\((V,\\mathcal{L},E)\\), where \\(V\\) represents a finite set of nodes, \\(\\mathcal{L}\\colon V\\to \\Sigma^{*}\\) is a function that maps each node to a string, called a label, and \\(E\\subset V\\times V\\) is a set of directed edges between pairs of nodes. Additionally, we require that the directed graph \\((V,E)\\) forms a directed tree with a unique root \\(\\rho \\in V\\). The set of nodes encountered along the unique path from the taxonomy tree's root node \\(\\rho\\) to a given node \\(v\\) constitutes that node's ancestor set, which we denote \\(\\operatorname{anc}(v)\\). Formally, for any node \\(v\\in V\\), let \\(\\operatorname{par}(v)\\) denote \\(v\\)'s unique parent node in the tree; \\(\\operatorname{par}(\\rho)\\), i.e., with the root note \\(\\rho\\) as an argument, is undefined. We then recursively define \\(\\operatorname{anc}(v)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.421, + 0.906, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\operatorname {a n c} (v) \\stackrel {\\text {d e f}} {=} \\left\\{\\rho \\right\\} \\quad (1 \\mathrm {a}, \\text {b a s e c a s e}; v = \\rho)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.441, + 0.906, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\operatorname {a n c} (v) \\stackrel {\\text {d e f}} {=} \\{v \\} \\cup \\operatorname {a n c} (\\operatorname {p a r} (v)). \\quad (1 b, \\text {i n d u c t i v e c a s e}; v \\neq \\rho)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.906, + 0.578 + ], + "angle": 0, + "content": "We now discuss how to compare ground-truth and predicted nodes on a taxonomy. Let \\(\\{(v_n^{\\mathrm{gt}}, v_n^{\\mathrm{pr}})\\}_{n=1}^N\\) be a set of pairs of nodes where \\(v_n^{\\mathrm{gt}}\\) is the ground-truth node for the \\(n^{\\mathrm{th}}\\) image and \\(v_n^{\\mathrm{pr}}\\) is the node mapped from the string generated from the VLM conditioned on the \\(n^{\\mathrm{th}}\\) image. (We describe a procedure to create this string-node mapping in §6.) We then define hP and hR over \\(\\{(v_n^{\\mathrm{gt}}, v_n^{\\mathrm{pr}})\\}_{n=1}^N\\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.589, + 0.906, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} \\mathbf {P} \\stackrel {\\text {d e f}} {=} \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\frac {\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {p r}}\\right) \\cap \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {g t}}\\right) \\right|}{\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {p r}}\\right) \\right|}, \\tag {2a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.634, + 0.905, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} \\mathbf {R} \\stackrel {\\text {d e f}} {=} \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\frac {\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {p r}}\\right) \\cap \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {g t}}\\right) \\right|}{\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {g t}}\\right) \\right|}. \\tag {2b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.685, + 0.906, + 0.715 + ], + "angle": 0, + "content": "The hierarchical F1 (hF) is then defined as the harmonic mean of hP and hR," + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.725, + 0.905, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} \\mathbf {F} \\stackrel {\\text {d e f}} {=} \\frac {2 \\cdot \\mathbf {h} \\mathbf {P} \\cdot \\mathbf {h} \\mathbf {R}}{(\\mathbf {h} \\mathbf {P} + \\mathbf {h} \\mathbf {R})}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.769, + 0.906, + 0.83 + ], + "angle": 0, + "content": "Both hP and hR fall in the interval [0, 1]. The maximum is achieved when the target and predicted paths coincide. Less-than-perfect scores capture the extent to which the path deviates from the gold answer." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "hP Hierarchical precision captures the amount of incorrect information in the prediction, i.e., deviations from the correct path, relative to the number of nodes shared with the correct path." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9111" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.089, + 0.287, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.184, + 0.194, + 0.195 + ], + "angle": 0, + "content": "TRAIN" + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.203, + 0.262, + 0.227 + ], + "angle": 0, + "content": "Prediction: A mode of transport" + }, + { + "type": "image_footnote", + "bbox": [ + 0.145, + 0.234, + 0.253, + 0.246 + ], + "angle": 0, + "content": "hP:1.00 hR:0.75" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.109, + 0.421, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.183, + 0.414, + 0.195 + ], + "angle": 0, + "content": "TELEPHOTO LENS" + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.203, + 0.364, + 0.227 + ], + "angle": 0, + "content": "Prediction: A lens cover" + }, + { + "type": "image_footnote", + "bbox": [ + 0.287, + 0.234, + 0.395, + 0.246 + ], + "angle": 0, + "content": "hP:0.80 hR:0.67" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.089, + 0.57, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.184, + 0.528, + 0.195 + ], + "angle": 0, + "content": "PITH HELMET" + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.203, + 0.495, + 0.227 + ], + "angle": 0, + "content": "Prediction: Neckwear" + }, + { + "type": "image_footnote", + "bbox": [ + 0.429, + 0.234, + 0.537, + 0.246 + ], + "angle": 0, + "content": "hP:0.75 hR:0.50" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.089, + 0.712, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.184, + 0.702, + 0.195 + ], + "angle": 0, + "content": "STIRLING CASTLE" + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.203, + 0.692, + 0.227 + ], + "angle": 0, + "content": "Prediction: Château de Ranrouët" + }, + { + "type": "image_footnote", + "bbox": [ + 0.57, + 0.234, + 0.678, + 0.246 + ], + "angle": 0, + "content": "hP:0.83 hR:0.83" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.855, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.184, + 0.753, + 0.195 + ], + "angle": 0, + "content": "POOL" + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.203, + 0.779, + 0.228 + ], + "angle": 0, + "content": "Prediction: High jump" + }, + { + "type": "image_footnote", + "bbox": [ + 0.712, + 0.234, + 0.82, + 0.246 + ], + "angle": 0, + "content": "hP:0.67 hR:0.67" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.908, + 0.289 + ], + "angle": 0, + "content": "Figure 3. Illustrative examples. Hierarchical precision (hP) and recall (hR) calculations on random label pairs (\\(\\S 5\\)). While hP penalizes incorrect labels, i.e., labels that are further away from the target's ancestor set, hR penalizes mistakes made higher up on the taxonomy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.296, + 0.485, + 0.416 + ], + "angle": 0, + "content": "hR Hierarchical recall measures the amount of correct information in the prediction, i.e., how much the path from the root to the predicted node overlaps with the path from the root to the ground truth node. Hierarchical recall thus penalizes missing coverage of the target path, especially early deviations. This can be seen as measuring the specificity of the prediction, in particular when the hP is 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.621 + ], + "angle": 0, + "content": "We give an example of the computation of \\(\\mathbf{hP}\\) and \\(\\mathbf{hR}\\) in Fig. 3, sampled from the OVEN datasets. In the case of the picture of TRAIN, we can see how only predicting MODE OF TRANSPORT leads to lower \\(\\mathbf{hR}\\) since it lacks information, while \\(\\mathbf{hP}\\) remains 1 since there is no incorrect information. For the sport POOL and STIRLING CASTLE, we notice how both scores drop, as incorrect nodes are predicted—these responses contain both incorrect information and some partially correct information. In the case of PITH HELMET, we see that the prediction is incorrect yet not very specific, resulting in a lower \\(\\mathbf{hR}\\) than \\(\\mathbf{hP}\\). We also note that it is not possible to have a lower \\(\\mathbf{hP}\\) than \\(\\mathbf{hR}\\) if the ground truth nodes are leaf nodes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.379, + 0.649 + ], + "angle": 0, + "content": "4. Taxonomically Linked Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.809 + ], + "angle": 0, + "content": "To investigate taxonomy-aware evaluation measures for text generated from a VLM, we consider two taxonomies linked to FGVC tasks. First, the Catalogue of Life taxonomy as contained in the iNaturalist21 dataset [30] is a clean, expert-curated taxonomy that maps evolutionary relationships between species and contains exactly 10,000 leaf nodes. The second is a more general taxonomy constructed from the Wikidata [71] knowledge graph, which links entities from FGVC datasets in the Open-domain Visual Entity (OVEN) benchmark [31]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.306, + 0.833 + ], + "angle": 0, + "content": "4.1. Extracting Taxonomies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While the iNaturalist21 dataset includes a built-in taxonomy, this is not the case for other FGVC datasets. For those included in the OVEN collection, we utilize Wikidata—a large-scale knowledge graph derived from Wikipedia [71]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.908, + 0.492 + ], + "angle": 0, + "content": "Unlike taxonomies, knowledge graphs are not constrained to tree or even directed acyclic graph structures, which necessitates the extraction of a taxonomy that fits the formal definition given in §3. To construct such a tree from Wikidata, we rely on the subclass of relation. When multiple paths exist from the root to a given node in the taxonomy, we retain the longest path. In the case of ties, we sample one of the tied paths uniformly at random. To ensure a rooted directed tree, we exclude higher-level abstract classes that introduce cycles. Because OVEN already links its datasets (§4.2) to Wikipedia entries, we can directly map them to Wikidata entities and incorporate them into our taxonomy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.503, + 0.806, + 0.52 + ], + "angle": 0, + "content": "4.2. Entity Linked Datasets of Images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.724 + ], + "angle": 0, + "content": "The iNaturalist21 dataset [30] was compiled and curated from the citizen science platform iNaturalist. It contains 2.7 million images spanning 10,000 species. The OVEN dataset [31] aggregates several FGVC datasets, linking them to Wikipedia identifiers and natural language questions to support the evaluation of open-domain visual entity recognition. It includes ImageNet21k-P [61, 63], iNaturalist2017 [29], Cars196 [36], SUN397 [76], Food101 [10], Sports100 [25], Aircraft [47], Oxford Flowers [49], and Google Landmarks v2 [56, 75]. We exclude iNaturalist2017 from our OVEN-based evaluations, as we separately analyze the newer iNaturalist21 dataset, which is aligned with the higher-quality Catalogue of Life taxonomy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.8, + 0.756 + ], + "angle": 0, + "content": "5. The Woes of Existing Measures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we evaluate whether non-taxonomic text similarity measures already encode sufficient taxonomic information, e.g., through label overlap or representational similarity. To investigate this, we conduct a targeted synthetic experiment comparing the taxonomic similarity measures (hP and hR), with textual and multimodal similarity measures from §2.2 commonly used to assess the quality of language model outputs. These measures are both string-based and representation-based." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9112" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.486, + 0.278 + ], + "angle": 0, + "content": "
MeasureiNat21OVEN
τ-hPτ-hRτ-hPτ-hR
EM0.010.070.010.01
Contained0.020.090.020.03
stringROUGE0.180.410.170.21
METEOR0.180.400.180.22
rep.BERTscore0.010.310.270.18
SentBERT0.020.250.340.31
NLI0.08-0.20-0.170.19
CLIP-t2t0.140.460.250.15
CLIP-i2t0.350.490.350.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.287, + 0.485, + 0.345 + ], + "angle": 0, + "content": "Table 1. Correlation of text similarity measures with hierarchical metrics. Kendall's \\(\\tau\\) correlations between taxonomic metrics hP, hR, and the different similarity measures on the synthetic evaluation sets. All correlations have a \\(p\\)-value \\(< 0.0001\\)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.37, + 0.485, + 0.855 + ], + "angle": 0, + "content": "Experimental Setup. We construct two synthetic evaluation sets to determine to what extent text similarity measures reflect taxonomic structure. We consider the taxonomies from iNaturalist21 and the Wikidata (see §4) to generate controlled pairs of labels corresponding to pairs of nodes in each taxonomy. A reference node \\( v \\) and a candidate node \\( v' \\) are chosen such that \\( v \\) is always a leaf node and \\( v' \\) is at some given distance from \\( v \\). The distance between two nodes in the taxonomy is the length of the shortest undirected path between the two nodes in the taxonomy. To ensure a balanced dataset, we sample the node pairs such that their pairwise distance is uniformly distributed up to a maximum distance of 7, depending on the depth of the node. To evaluate hP, we sample 100,000 reference-candidate node pairs from the taxonomies. To evaluate hR in a controlled fashion, however, we additionally require that \\( v' \\in \\mathrm{anc}(v) \\), i.e., that \\( \\mathbf{hP} = 1 \\). This allows us to isolate the sensitivity of the similarity measures at decreasing levels of specificity.\\(^{1}\\) For each pair, we compute both the taxonomy-aware score and text-based measures between the corresponding labels \\( \\mathcal{L}(v) \\) and \\( \\mathcal{L}(v') \\). We then measure the rank correlation between the taxonomy-aware score and various text-based measures using Kendall's \\( \\tau \\), which evaluates whether the similarity measure reflects taxonomic proximity under controlled conditions. By construction, we use the node label \\( \\mathcal{L}(v') \\) as the prediction to avoid the problem of mapping unconstrained output onto the taxonomy. All of the measures we consider are text-based (see §10 in the supplement for details), except the CLIP image-to-text score, where we sample an image for the reference node from the corresponding dataset and measure CLIP model similarity between the image and the candidate label. We" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.093, + 0.882, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.315, + 0.907, + 0.372 + ], + "angle": 0, + "content": "Figure 4. Agreement in node placement using different similarity measures. We observe variation in node placement for all measures. Only METEOR and ROUGE, and the CLIP variations when combined with Alg. 1 share predictions frequently." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.41 + ], + "angle": 0, + "content": "choose high-performing models that are not prohibitively large for the model-based comparisons.2" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.909, + 0.747 + ], + "angle": 0, + "content": "Results. Table 1 presents correlations between hP and hR on one hand and the text similarity measures from §2.2 and CLIP-i2t on the other, for the iNaturalist21 and Wiki-data labels. We observe that surface-level measures (Exact Match (EM), Contained) exhibit a low correlation with the taxonomy-aware measures hP and hR. ROUGE and METEOR, which are based on \\( n \\)-gram overlap, capture taxonomic similarity more effectively. This is presumably due to label overlap, e.g., GRAY SEAL is a type of SEAL. The highest correlation is achieved by CLIP image-to-text (CLIP-i2t) for both datasets, followed by CLIP text-to-text (CLIP-t2t), except for hP on the iNaturalist21 dataset. In that case, CLIP-t2t fails to capture deviations from the correct taxonomic path. This suggests that CLIP's image features may encode more taxonomic information than its text representations, which can struggle due to the low-frequency vocabulary inherent to this task. NLI is an outlier, exhibiting a negative correlation in most cases. While NLI captures taxonomic similarity for closely related labels, it performs poorly on more distantly related pairs. This is visualized in Fig. 8 in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.843 + ], + "angle": 0, + "content": "Summary. Overall, we observe that representation-based measures capture some notions of taxonomic distance along the path from the leaf node to the root (hR), but are, for the most part, poorly equipped to measure taxonomic distances across different subtrees (hP)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Evaluating hP while controlling for hR is only possible for non-leaf reference nodes. Since our datasets always have leaf reference nodes, we did not explore this setting." + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.851, + 0.908, + 0.903 + ], + "angle": 0, + "content": "2All models are downloaded from. huggingface.co, i.e., BERTScore (microsoft/deberta-large-mnli), SentenceBERT (sentence-transformers/all-mpnet-base-v2), NLI (MoritzLaurer/mDeBERTa-v3-base-mnli-xnli) and CLIP (apple/DFN5B-CLIP-ViT-H-14)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9113" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.089, + 0.454, + 0.272 + ], + "angle": 0, + "content": "
MeasurehPhRhFAccuracy
Exact Match0.370.420.3917.5%
Contained0.490.500.4924.3%
BLEU0.710.530.5424.3%
METEOR0.620.650.6329.1%
ROUGE0.620.660.6330.0%
BERTScore0.450.520.4811.3%
SentBERT0.580.670.6220.0%
CLIP-t2t0.720.790.7530.3%
CLIP-i2t0.580.710.235.3%
CLIP-i2t+Alg. 10.800.810.8044.0%
CLIP-t2t+Alg. 10.790.820.8047.1%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.352 + ], + "angle": 0, + "content": "Table 2. Evaluating taxonomy mapping with human annotation. We report the performance of text similarity measures (directly) and CLIP image-to-text similarity for mapping VLM text predictions to the taxonomy using a subset of iNaturalist21 samples that were manually mapped to the correct nodes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.468, + 0.379 + ], + "angle": 0, + "content": "6. Mapping VLM Predictions to Taxonomies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.599 + ], + "angle": 0, + "content": "The key challenge for taxonomy-aware evaluation of unconstrained text generated from VLMs is mapping the text generated by a VLM onto nodes in a given taxonomy. A baseline approach to this mapping is to use text similarity measures (e.g., those in §2.2) to calculate the best match between the generated text and the labeled nodes in a taxonomy. We also develop a heuristic approach that performs this mapping using a general similarity measure as a subroutine, which makes using CLIP text-to-image similarity feasible. Specifically, this algorithm combines string-based matching with scores from a given similarity measure to position predictions onto the taxonomy. We describe this procedure in words below and give pseudocode in §11 (Alg. 1 supplement)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our goal is to map each VLM-generated prediction string to a node in the taxonomy using a series of increasingly flexible matching strategies. Generally, we attempt exact matches first, then allow for partial overlaps or ambiguous cases, always preferring more specific nodes when multiple candidates are viable. Before matching, we normalize all predictions and node labels by lowercasing, replacing dashes with spaces, and stripping other punctuation. We then compute similarity scores between the prediction and all node labels using the CLIP similarity measures. More specifically, the mapping proceeds in stages. We first check whether any of the top- \\(k\\) \\((k = 10)\\) highest-scoring nodes, according to the CLIP measure applied to their labels, appear verbatim in the prediction string; if so, we return the corresponding node. If this fails, we search for verbatim matches against all other labels in the taxonomy. If no full-string matches are found, we repeat these checks using \\(n\\)-gram overlap for \\(n \\in \\{4,3,2\\}\\). If multiple candidates remain with similar scores, we apply a fallback based on score ambiguity: if the difference between the top two scores, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "between the top and \\( k^{\\text{th}} \\) score, are both below threshold values, we treat the case as ambiguous. We then search for a common ancestor among the top candidates. If any node appears sufficiently often (at least four times) in the ancestor sets of multiple nodes in the top-\\( k \\) highest-scoring nodes we return the most specific such node, defined as the one deepest in the taxonomy. If none of the prior criteria are met, we fall back to returning the single highest-scoring candidate. Thus, the algorithm always returns a node in the taxonomy. If multiple nodes match the above criteria, we return the most specific one, i.e., the node deepest in the taxonomy. The \"shared ancestor\" step is particularly beneficial for the CLIP image-to-text similarities since we do not have images corresponding to higher-order taxonomic concepts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.315, + 0.684, + 0.332 + ], + "angle": 0, + "content": "6.1. Mapping Quality" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.504 + ], + "angle": 0, + "content": "To evaluate the quality of the taxonomic node proposals we measure their agreement against manually annotated node positions. We compare the similarity measures described in §2.2 (and Supplement §10) directly against all candidate node labels. We then do the same using Alg. 1, combined with the CLIP measures. To construct the test set, we gather 416 unique VLM outputs for the iNaturalist21 dataset, randomly sampled from the generations described in §7. For each output, we manually find the best matching taxonomic node given the string, e.g., matching This is a bird to AVES/BIRDS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.906, + 0.724 + ], + "angle": 0, + "content": "Results. Results are shown in Tab. 2. We find string-matching measures (Exact Match and Contained) do fairly poorly, while other surface-level measures (BLEU, METEOR, and ROUGE) perform slightly better. The representation-based approaches do not perform better in general, except for CLIP text-to-text. Fig. 4 shows that the string-level measures have considerable diversity in their predictions, while the representation-based measures show more diversity. The best performance is found with CLIP text-to-text and CLIP image-to-text using the algorithm described above (Alg. 1), which results in the same hF score but slightly improved exact node match accuracy score for CLIP text-to-text + Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.825, + 0.757 + ], + "angle": 0, + "content": "7. Evaluating and Comparing VLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We compare one closed model and seven publicly available VLMs on fine-grained visual classification using our taxonomy-aware evaluation scheme. To map (unconstrained) strings generated from a VLM to nodes in the taxonomy, we use the best-performing similarity method (CLIP-t2t + Alg. 1), enabling the computation of hP, hR, and hF. These scores offer insight into the level of accuracy (hP) and specificity (hR) of VLM predictions for a given taxonomy. The resulting rankings differ significantly from" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9114" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.482, + 0.227 + ], + "angle": 0, + "content": "
Model Name# ParametersTraining PipelineOpenVLM Avg.
GPT-4 [50]UnknownUnknown63.1
ILmXC2 [23]7BPretr., instr.-tuning62.2
OLM12B [79]12BPretr., instr.-tuning, RLHF54.6
QVLChat [3]9.6BPretr., instr.-tuning51.3
OLM3B [32]3BPretr., instr.-tuning47.8
LLaVA [45]7.2BPretr., instr.-tuning45.9
QwenVL [3]9.6BPretr.21.7
Fuyu [6]8BPretr.N/A
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.237, + 0.483, + 0.28 + ], + "angle": 0, + "content": "Table 3. Overview of evaluated VLMs. Details on the models we evaluate, including number of parameters, training method, and average performance on a large vision-language benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.484, + 0.352 + ], + "angle": 0, + "content": "those based on standard text similarity measures, highlighting the taxonomy-aware measure's ability to capture previously unmeasured aspects of model performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.37, + 0.485, + 0.718 + ], + "angle": 0, + "content": "Experimental Setup. We compare eight VLMs that differ in model size, training configuration (pretrained-only, instruction-tuned, and preference-aligned [51]), and performance on the Open VLM Leaderboard [16], a collection of 13 VL benchmarks with 38 models. We summarize model sizes, training configuration, and average performance on the Open VLM leaderboard in Table 3. For both the OVEN-sourced FGVC datasets and iNaturalist21, we randomly sample 6000 images from the validation set. For iNaturalist21, we use the prompt What species is this?, while for OVEN the question is included in the dataset and varies by object type. To understand how well we can control hR and hP for instruction-tuned models, we construct two different sets of instructions: (1) one containing a minimal instruction for answering in a specified format, and (2) one containing an instruction to be specific. We use slight variations of the prompt formatting to fit each model (as specified by the model developers); the exact prompts are provided in the supplement (Tab. 7 and Tab. 8). Because non-instruction-tuned models do not respond meaningfully to instructions, we prompt them with a straightforward question-answer template \\( Q \\): {}A:. We use an orange-colored {} to indicate a variable in the template." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.295, + 0.743 + ], + "angle": 0, + "content": "7.1. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.78 + ], + "angle": 0, + "content": "How does the ranking of VLM performance change when considering taxonomy-aware measures? In" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Fig. 5, we rank eight VLMs based on their performance on FGVC according to the different similarity measures corresponding to the values given in Tab. 4 and Tab. 5 (\\(\\S 14\\) in supplement). We find that several text similarity measures yield similar rankings when used as direct evaluation criteria. In contrast, the taxonomy-aware measures lead to notable changes in model rankings. For instance, LLaVA ranks lowest under Exact Match for" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.096, + 0.897, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.209, + 0.896, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.327, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Figure 5. Ranking of VLMs for the iNaturalist21 (top) and OVEN (bottom) datasets. We evaluate the ranking of VLMs (vertical axis) based on different evaluation measures (horizontal axis). The best model is shown in the top row. On the left, we see the model names ranked by exact match and on the right ranked by the hierarchical F1 hF. See §14, Tab. 4 in supplement for exact numbers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.651 + ], + "angle": 0, + "content": "iNaturalist21, yet achieves high hierarchical precision (\\(\\mathbf{hP}\\)) and, consequently, a strong hierarchical F1 (\\(\\mathbf{hF}\\)). This indicates that LLaVA deviates from the correct taxonomy path less frequently than other models. However, its low hierarchical recall (\\(\\mathbf{hR}\\)) suggests that its predictions are often not specific. Another example is GPT-4, which ranks highest on all measures except \\(\\mathbf{hP}\\) and \\(\\mathbf{hF}\\) for iNaturalist21. Compared to QVLChat, which is ranked first in both \\(\\mathbf{hP}\\) and \\(\\mathbf{hF}\\), we observe that GPT-4 makes more specific but erroneous predictions. Our taxonomy-aware measure offers a complementary perspective to traditional metrics by revealing performance differences that standard similarity-based measures overlook." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Do taxonomy-aware measures capture prompt sensitivity? We prompt the VLMs using two distinct prompts, as described earlier, to examine how their predictions shift in terms of hierarchical precision (hP) and recall (hR), shown jointly in Fig. 7. Fuyu and QwenVL are not instruction-tuned and therefore do not respond differently, while LLaVA shows only minimal changes. In contrast, the other models clearly adjust their predictions based on the different instructions in the prompts. On the iNaturalist21 dataset, all models except GPT-4 exhibit a trade-off between hP and hR: higher precision tends to come at the cost of recall. GPT-4, however, offers both more accurate and more fine-grained classifications when prompted appropriately. On the OVEN dataset, results are more mixed: roughly half of the models improve on both axes with prompt changes." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9115" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.092, + 0.454, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.233, + 0.486, + 0.292 + ], + "angle": 0, + "content": "Figure 6. Prompt tuning results for bird classifier. For High hP (●), we target prompts that prioritize hP. For Acc (▲) we optimize for prompts that give higher binary accuracy. This demonstrates how hP and hR can help tune a VLM application." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.392 + ], + "angle": 0, + "content": "These findings suggest that each application may prioritize different target values for \\(\\mathbf{hP}\\) and \\(\\mathbf{hR}\\), and that models can adapt substantially to meet those needs. In such cases, prompt design becomes a powerful paradigm for tuning model behavior as further explored in the next section." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.401, + 0.479, + 0.417 + ], + "angle": 0, + "content": "7.2. An Application: High-Precision Classification" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.424, + 0.485, + 0.773 + ], + "angle": 0, + "content": "To demonstrate the usefulness of hR and hP in tuning prompts for a specific application, we prompt Llama 3.2 Vision (11B) [65] to classify birds. The application goal is to minimize the amount of wrong information in the prediction, which is equivalent to maximizing hierarchical precision but hard to formulate in terms of exact match binary classification accuracy. We use prompt-tuning with hP and hR as feedback signals, and compare to the same procedure but with accuracy as the feedback signal. We start with a basic prompt What is this an image of? Answer in the format 'A: .' and task ChatGPT [50] to improve on the prompt for 30 iterations. A system prompt describes the scores and the goal of developing a conservative classifier that gives a less specific prediction if uncertain (see §15 in the supplement for more details). In each iteration, we include past prompts along with their resulting hR and hP scores, or the accuracy score, in a few-shot learning setup. We tune the prompts with 20 images and evaluate on 200 images of birds from iNaturalist21. The results are shown in Fig. 6. While this is a toy example, we see how the taxonomy-aware measures not only enable a higher granularity of model performance analysis, but can also facilitate targeted application development." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.785, + 0.345, + 0.801 + ], + "angle": 0, + "content": "8. Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Evaluating unconstrained text generated from a VLM is inherently challenging: there are countless entities to describe and even more ways to describe them. As the computer vision community shifts from closed-set recognition to open-ended recognition, evaluation increasingly resembles an NLP task. In this paper, we introduce structure into" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.089, + 0.898, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.232, + 0.908, + 0.275 + ], + "angle": 0, + "content": "Figure 7. Prompt sensitivity captured by hR and hP. We compare two prompt versions, a minimal barebones version and one that asks for a taxonomically specific prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.909, + 0.596 + ], + "angle": 0, + "content": "VLM evaluation by leveraging curated taxonomies and established hierarchical evaluation measures. Despite this, several limitations remain. First, placing unconstrained text on a taxonomy is a low-resource problem; there are no large-scale datasets available to evaluate our mapping approach, nor sufficient data to train a dedicated classifier that links VLM outputs to taxonomy nodes. To address this, we repurpose existing text similarity measures, along with CLIP, to perform the mapping. While our evaluation shows that this approach is non-trivial and introduces errors that affect downstream VLM evaluation, we are optimistic about future improvements, such as training on synthetic data generated by language models. Second, taxonomies, especially those derived from Wikidata, are inherently noisy. Even expert-curated taxonomies like the one used for iNaturalist21 can have uneven subtree granularity, complicating interpretations based on globally averaged measures. One potential solution would be to employ expert input to assign weighted edges to reflect semantic distances between nodes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.634, + 0.637 + ], + "angle": 0, + "content": "9. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.909, + 0.788 + ], + "angle": 0, + "content": "Text similarity measures from NLP are a natural place to look for evaluation of unconstrained VLM output. Additionally, many FGVC tasks come with side information about the class structure, namely taxonomies, which text similarity measures do not fully capture, as we have shown in this paper. To solve this gap, we have proposed a framework for using hierarchical evaluation measures to evaluate VLMs for FGVC, enabling a more fine-grained view of the specificity and accuracy of these models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.812, + 0.682, + 0.829 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Vésteinn Snæbjarnarson, Serge Belongie, Nico Lang, and Stella Frank are supported by the Pioneer Centre for AI, DNRF grant number P1. Niklas Stoehr is supported by the Swiss Data Science Center (SDSC) PhD Fellowship." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9116" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.265 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In Neural Information Processing Systems, pages 23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.271, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[2] Majed A Alkhamees, Mohammed A Alnuem, Saleh M Al-Saleem, and Abdulrakeeb M Al-Ssulami. A semantic metric for concepts similarity in knowledge graphs. Journal of Information Science, 49:778-791, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.328, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[4] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. TouchStone: Evaluating vision-language models by language models. arXiv, 2308.16890, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[5] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.482, + 0.583 + ], + "angle": 0, + "content": "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for AI agents, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[7] Walter Bennette, Nathaniel Hofmann, Nathaniel Wilson, and Tyler Witter. Hierarchical open-set recognition for automatic target recognition. In 2021 IEEE Symposium Series on Computational Intelligence (SSCI). IEEE, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[8] Luca Bertinetto, Romain Mueller, Konstantinos Tertikas, Sina Samangooei, and Nicholas A Lord. Making better mistakes: Leveraging class hierarchies with deep networks. In Conference on Computer Vision and Pattern Recognition, pages 12506-12515, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[9] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. VisIT-Bench: A benchmark for vision-language instruction following inspired by real-world use. In Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[10] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – Mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In Conference on Empirical Methods in Natural Language Processing, 2015. 3, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[12] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. MiniGPT-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[13] Elijah Cole, Kimberly Wilber, Grant Van Horn, Xuan Yang, Marco Fornoni, Pietro Perona, Serge Belongie, Andrew Howard, and Oisin Mac Aodha. On label granularity and object localization. In European Conference on Computer Vision, pages 604-620, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[14] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.306, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[15] Alessandro Conti, Enrico Fini, Massimiliano Mancini, Paolo Rota, Yiming Wang, and Elisa Ricci. Vocabulary-free image classification. In Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.904, + 0.39 + ], + "angle": 0, + "content": "[16] OpenCompass Contributors. OpenCompass: A universal evaluation platform for foundation models, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.393, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[17] Riccardo De Lutio, Yihang She, Stefano D'Aronco, Stefania Russo, Philipp Brun, Jan D Wegner, and Konrad Schindler. Digital taxonomist: Identifying plant species in community scientists' photographs. ISPRS journal of photogrammetry and remote sensing, 182:112-121, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.521, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[19] Jia Deng, Jonathan Krause, Alexander C. Berg, and Li Fei-Fei. Hedging your bets: Optimizing accuracy-specificity trade-offs in large scale visual recognition. In Conference on Computer Vision and Pattern Recognition, pages 3450–3457, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.592, + 0.905, + 0.647 + ], + "angle": 0, + "content": "[20] Xiwen Dengxiong and Yu Kong. Ancestor search: Generalized open set recognition via hyperbolic side information learning. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4003-4012, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.649, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics, pages 4171–4186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[22] Ankit Dhall, Anastasia Makarova, Octavian Ganea, Dario Pavllo, Michael Greeff, and Andreas Krause. Hierarchical image classification using entailment cone embeddings. In Conference on Computer Vision and Pattern Recognition, pages 836-837, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2: Mastering freeform text-image composition and comprehension in vision-language large model. arXiv, 2401.16420, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9117" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.162 + ], + "angle": 0, + "content": "[24] S. Doveh, A. Arbelle, S. Harary, E. Schwartz, R. Herzig, R. Giryes, R. Feris, R. Panda, S. Ullman, and L. Karlinsky. Teaching structured Vision & Language concepts to Vision & Language models. In Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.462, + 0.177 + ], + "angle": 0, + "content": "[25] Kaggle (Gerry). 100 sports image classification, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[26] Simon Ging, Maria Alejandra Bravo, and Thomas Brox. Open-ended VQA benchmarking of vision-language models by exploiting classification datasets and their semantic hierarchy. In International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[27] Sherzod Hakimov and David Schlangen. Images in language space: Exploring the suitability of large language models for vision & language tasks. In Findings of the Association for Computational Linguistics, pages 14196-14210, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[28] Yuan He, Zhangdie Yuan, Jiaoyan Chen, and Ian Horrocks. Language models as hierarchy encoders. In Neural Information Processing Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[29] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexx Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The iNaturalist species classification and detection dataset. In Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.49 + ], + "angle": 0, + "content": "[30] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin MacAodha. Benchmarking representation learning for natural world image collections. In Conference on Computer Vision and Pattern Recognition, pages 12879-12888, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[31] Hexiang Hu, Yi Luan, Yang Chen, Urvashi Khandelwal, Mandar Joshi, Kenton Lee, Kristina Toutanova, and Ming-Wei Chang. Open-domain visual entity recognition: Towards recognizing millions of wikipedia entities. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12065–12075, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[32] Shengding Hu, Yuge Tu, Xu Han, Ganqu Cui, Chaoqun He, Weilin Zhao, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Xinrong Zhang, Zhen Leng Thai, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, dahai li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the potential of small language models with scalable training strategies. In Conference on Language Modeling, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[33] Svetlana Kiritchenko, Stan Matwin, Richard Nock, and Fazel Famili. Learning and evaluation in the presence of class hierarchies: Application to text categorization. In Canadian AI, 2006. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[34] Tom Kocmi and Christian Federmann. Large language models are state-of-the-art evaluators of translation quality. In Conference of the European Association for Machine Translation, pages 193–203, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Aris Kosmopoulos, Ioannis Partalas, Eric Gaussier, Georgios Paliouras, and Ion Androutsopoulos. Evaluation measures for hierarchical classification: A unified view and novel approaches. Data Mining and Knowledge Discovery, 29(3): 820-865, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D object representations for fine-grained categorization. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision Workshops, pages 554-561, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[37] Suren Kumar and Rui Zheng. Hierarchical category detector for clothing recognition from visual data. In International Conference on Computer Vision Workshops, pages 2306-2312, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[38] Nico Lang, Vésteinn Snæbjarnarson, Elijah Cole, Oisin Mac Aodha, Christian Igel, and Serge Belongie. From coarse to fine-grained open-set recognition. In Conference on Computer Vision and Pattern Recognition, pages 17804-17814, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[39] Kibok Lee, Kimin Lee, Kyle Min, Yuting Zhang, Jinwoo Shin, and Honglak Lee. Hierarchical novelty detection for visual object recognition. In Conference on Computer Vision and Pattern Recognition, pages 1034-1042, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[40] Clément Lefebvre and Niklas Stoehr. Rethinking the event coding pipeline with prompt entailment. In EACL Workshop Fact Extraction and VERIFICATION, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[41] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[42] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81. Association for Computational Linguistics, 2004. 1, 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[43] Dekang Lin. An information-theoretic definition of similarity. In International Conference on Machine Learning, page 296-304, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, pages 740-755, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Conference on Computer Vision and Pattern Recognition, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Neural Information Processing Systems, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[47] Subhransu Maji, Esa Rahtu, Juho Kannala, Mathew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv, 1306.5151, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[48] Kenneth Marino, Ruslan Salakhutdinov, and Abhinav Gupta. The more you know: Using knowledge graphs for image classification. In Computer Vision and Pattern Recognition, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[49] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.854, + 0.844 + ], + "angle": 0, + "content": "[50] OpenAI. GPT-4 technical report, 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9118" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Neural Information Processing Systems, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[52] Jeff Z. Pan, Simon Razniewski, Jan-Christoph Kalo, Sneha Singhania, Jiaoyan Chen, Stefan Dietze, Hajira Jabeen, Janna Omeliyanenko, Wen Zhang, Matteo Lissandrini, Russia Biswas, Gerard de Melo, Angela Bonifati, Edlira Vakaj, Mauro Dragoni, and Damien Graux. Large language models and knowledge graphs: Opportunities and challenges. Transactions on Graph Data and Knowledge, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[53] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. In Annual Meeting on Association for Computational Linguistics, page 311-318, 2002. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[54] Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Conference on Empirical Methods in Natural Language Processing, pages 2463-2473, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[56] Elias Ramzi, Nicolas Audebert, Clement Rambour, Andre Araujo, Xavier Bitot, and Nicolas Thome. Optimization of rank losses for image retrieval. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.483, + 0.6 + ], + "angle": 0, + "content": "[57] Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees GM Snoek. SelEx: Self-expertise in fine-grained generalized category discovery. In European Conference on Computer Vision, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[58] Nils Reimers and Iryna Gurevych. Sentence-BERT: sentence embeddings using siamese BERT-networks. In Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3980-3990, 2019. 1, 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.483, + 0.714 + ], + "angle": 0, + "content": "[59] Philip Resnik. A class-based approach to lexical discovery. In Annual Meeting of the Association for Computational Linguistics, pages 327-329, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[60] Philip Resnik. Using information content to evaluate semantic similarity in a taxonomy. In International Joint Conference on Artificial Intelligence, page 448-453, 1995. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[61] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. In Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[62] Kevin Riehl, Michael Neunteufel, and Martin Hemberg. Hierarchical confusion matrix for classification performance evaluation. Journal of the Royal Statistical Society Series C: Applied Statistics, 72(5):1394-1412, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[64] Paolo Simeone, Raul Santos-Rodriguez, Matt McVicar, Jeffrey Lijffijt, and Tijl De Bie. Hierarchical novelty detection. In Advances in Intelligent Data Analysis, pages 310-321, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[65] Hugo Touvron, Aaron Grattafori, Michael Auli, et al. The Llama 3 herd of models. arXiv, 2407.21783, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[66] Mehmet Ozgur Turkoglu, Stefano D'Aronco, Gregor Perich, Frank Liebisch, Constantin Streit, Konrad Schindler, and Jan Dirk Wegner. Crop mapping from image time series: Deep learning with multi-scale label hierarchies. Remote Sensing of Environment, 264:112603, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[67] Grant Van Horn, Steve Branson, Scott Loarie, Serge Belongie, and Pietro Perona. Lean multiclass crowdsourcing. In Conference on Computer Vision and Pattern Recognition, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[68] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In International Conference on Learning Representations, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[69] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Conference on Computer Vision and Pattern Recognition, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[70] Nakul Verma, Dhruv Mahajan, Sundararajan Sellamanickam, and Vinod Nair. Learning hierarchical similarity metrics. In Conference on Computer Vision and Pattern Recognition, pages 2280-2287, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[71] Denny Vrandecic and Markus Krötzsch. Wikidata: A free collaborative knowledgebase. Communications of the ACM, 57(10):78-85, 2014. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[72] Cunxiang Wang, Sirui Cheng, Qipeng Guo, Yuanhao Yue, Bowen Ding, Zhikun Xu, Yidong Wang, Xiangkun Hu, Zheng Zhang, and Yue Zhang. Evaluating open-QA evaluation. In Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[73] Sinong Wang, Han Fang, Madian Khabsa, Hanzi Mao, and Hao Ma. Entailment as few-shot learner. arXiv, 2104.14690, 2021. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[74] Xiu-Shen Wei, Yi-Zhen Song, Oisin Aodha, Jianxin Wu, Yuxin Peng, Jinhui Tang, Jian Yang, and Serge Belongie. Fine-grained image analysis with deep learning: A survey. IEEE Transactions on Pattern Analysis & Machine Intelligence, 44(12):8927-8948, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[75] Tobias Weyand, Andre Araujo, Bingyi Cao, and Jacjk Sim. Google Landmarks Dataset v2: A Large-Scale Benchmark for Instance-Level Recognition and Retrieval. In Conference on Computer Vision and Pattern Recognition, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[76] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[77] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment: A novel task for fine-grained image understanding, 2019. 3, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9119" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[78] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Fanqing Meng, Siyuan Huang, Meng Lei, Ping Luo, and Yu Qiao. LVLM-eHub: A comprehensive evaluation benchmark for large vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[79] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. RLHF-V: Towards trustworthy MLLMs via behavior alignment from fine-grained correctional human feedback. In Conference on Computer Vision and Pattern Recognition, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[80] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[81] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In International Conference on Computer Vision, pages 11941-11952, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[82] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. BERTScore: Evaluating text generation with BERT. In International Conference on Learning Representations, 2020. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[83] Bingchen Zhao, Nico Lang, Serge Belongie, and Oisin Mac Aodha. Labeled data selection for category discovery. In European Conference on Computer Vision, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.541 + ], + "angle": 0, + "content": "[84] Wangchunshu Zhou, Yan Zeng, Shizhe Diao, and Xinsong Zhang. VLUE: A multi-task multi-dimension benchmark for evaluating vision-language pre-training. In International Conference on Machine Learning, pages 27395-27411, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.597 + ], + "angle": 0, + "content": "[85] Ganggao Zhu and Carlos A. Iglesias. Computing semantic similarity of concepts in knowledge graphs. IEEE Transactions on Knowledge and Data Engineering, 29(1):72-85, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[86] Ganggao Zhu and Carlos A. Iglesias. Exploiting semantic similarity for named entity disambiguation in knowledge graphs. Expert Systems with Applications, 101:8-24, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[87] Lianghui Zhu, Xinggang Wang, and Xinlong Wang. JudgeLM: Fine-tuned large language models are scalable judges. In International Conference on Learning Representations, 2025. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9120" + } + ] +] \ No newline at end of file diff --git a/2025/Taxonomy-Aware Evaluation of Vision-Language Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_origin.pdf b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..597d01d6d9f1941bc436f6e743d8b6ac43438eeb --- /dev/null +++ b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/777684c8-8659-4ecd-b9f3-bb5e792a1348_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96d1366369ab538439dcae99153f14713b3c62a2fdd7e73f737c9ba43cea6ee7 +size 1444869 diff --git a/2025/Taxonomy-Aware Evaluation of Vision-Language Models/full.md b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..154a92c26729056cd4bb93993787836d95029b43 --- /dev/null +++ b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/full.md @@ -0,0 +1,318 @@ +# Taxonomy-Aware Evaluation of Vision-Language Models + +Vésteinn Snæbjarnarson $^{1,2*}$ Kevin Du $^{2}$ Niklas Stoehr $^{2}$ Serge Belongie $^{1}$ Ryan Cotterell $^{2}$ Nico Lang $^{1}$ Stella Frank $^{1}$ $^{1}$ University of Copenhagen $^{2}$ ETH Zürich + +# Abstract + +When a vision-language model (VLM) is prompted to identify an entity depicted in an image, it may answer "I see a conifer," rather than the specific label NORWAY SPRUCE. This raises two issues for evaluation: Firstly, the unconstrained generated text needs to be mapped to the evaluation label space (i.e., CONIFER). Secondly, a useful classification measure should give partial credit to less-specific, but not incorrect, answers (NORWAY SPRUCE being a type of CONIFER). To meet these requirements, we propose a framework for evaluating unconstrained text predictions such as those generated from a vision-language model against a taxonomy. Specifically, we propose the use of hierarchical precision and recall measures to assess the level of correctness and specificity of predictions with regard to a taxonomy. Experimentally, we first show that existing text similarity measures do not capture taxonomic similarity well. We then develop and compare different methods to map textual VLM predictions onto a taxonomy. This allows us to compute hierarchical similarity measures between the generated text and the ground truth labels. Finally, we analyze modern VLMs on fine-grained visual classification tasks based on our proposed taxonomic evaluation scheme. Data and code are made available at https://github.com/vesteinn/vlm-eval. + +# 1. Introduction + +Modern vision-language models (VLMs) [1, 6, 12, 41, 46, 50] serve as viable one-stop shops for a wide variety of tasks at the intersection of vision and language. Given an image as input, VLMs generate unconstrained text, a key enabler of their flexibility across tasks. This flexibility, however, introduces additional challenges in their evaluation. In this paper, we study the problem of evaluating the unconstrained text generated by VLMs in the context of fine-grained visual categorization (FGVC), and put forward + +![](images/c50034492d3d7059650001866d0235dcd12abce808f5630bc933c180cd02bebe.jpg) +Figure 1. Vision-language-models (VLMs) as fine-grained classifiers. VLMs generate text with varying degrees of specificity and similarity to gold-standard label classes. We tackle the problem of aligning these outputs to taxonomic classes. + +two desiderata when evaluating VLMs on FGVC (Fig. 1): + +Evaluation with Textual Awareness. When tasked with classifying entities within an image, VLMs often generate text that does not inherently fit within the predefined set of labels used in conventional evaluations. For example, a model might describe an image with the phrase pines in the snow, while the ground-truth label is simply PINE TREE. We suggest that a good evaluation should rely on fuzzy or representation-based matching techniques, which assess the similarity between the generated text and the predefined class labels, rather than whether the VLM generated the predefined class's label exactly. Such approaches have a long history in natural language processing (see, e.g., [42, 58]). We call an evaluation measure for FGVC that is compatible with unconstrained text textually aware. + +Evaluation with Hierarchical Awareness. Standard evaluation methods for VLM-based entity classification typically treat text generated by a VLM in a binary manner, i.e., it is either completely correct or incorrect [18, 31, 74]. However, many classification tasks involve label sets that are organized hierarchically, such as taxonomic trees or knowledge graphs. When such a structure is available, we suggest that evaluation measures take advantage of the structure, i.e., predictions should be judged based on their semantic + +![](images/f1cc79cace0e64468913364f8204cfa9af9d5b7bd60a46e40621d5616d3f7672.jpg) +Figure 2. Framework for taxonomy-aware VLM evaluation. We propose a method for granular evaluation of the open-ended image classification capabilities of VLMs. To this end, we map model predictions onto a taxonomy, e.g., sourced from a large knowledge graph. We then use hierarchical precision (hP) and recall (hR) to evaluate how taxonomically accurate and specific a model prediction is. This, in turn, lets us evaluate and compare different VLMs and prompts. + +![](images/f066256a4675897a4adec1e660daa0590876e082b82e3d77a23023de4f2e7188.jpg) + +proximity to the correct label within the hierarchy. Fortunately, these structured label spaces are common in FGVC and other real-world classification settings. We call a VLM evaluation measure that is sensitive to taxonomic relationships hierarchically aware. + +In the technical portion of the paper, we propose a novel evaluation measure for assessing VLMs' performance on FGVC tasks that incorporates both textual and hierarchical awareness. We term such an evaluation strategy taxonomy-aware. Fig. 2 illustrates the approach for taxonomy-aware VLM evaluation that we present in this paper. Our measure is based on hierarchical precision and recall [33] after placing strings generated by the VLM on a given taxonomy. To contextualize our taxonomy-aware evaluation scheme within the literature, we first evaluate existing text similarity measures from NLP to assess whether they implicitly capture taxonomic relationships. Our analysis reveals that these measures do not align well with the underlying taxonomic structure, motivating our novel taxonomy-aware evaluation framework that explicitly maps unconstrained text generated from a VLM onto a taxonomy. + +In the empirical part of the paper, we explore our taxonomy-aware evaluation scheme on a range of VLMs, varying in size, training strategies, and performance, under our proposed taxonomy-aware evaluation scheme. To do so, we apply diverse prompt templates to control for accuracy, i.e., whether the model generates correct answers, and specificity, i.e., how detailed the response should be and whether to be more general when the model is not confident. In support of our evaluation, we leverage large, carefully constructed taxonomies derived from knowledge graphs extracted from Wikidata [71] and the Linnean Catalogue of Life [30]. These taxonomies span diverse domains, e.g., food, sports, animals, plants, cars, and landmarks, providing a holistic characterization of VLM behavior. Our results demonstrate that our proposed taxonomy-aware evaluation measures effectively quantify varying degrees of + +specificity that more rigid evaluation measures like accuracy fail to capture. We find that larger commercial models respond more successfully to prompts requesting varying levels of specificity, revealing substantial differences compared to the prompt sensitivity exhibited by smaller, publicly available models. These findings illustrate how taxonomy-aware evaluation can uncover nuanced model behaviors when conventional evaluation measures cannot. + +# 2. Background + +We briefly review prior work on evaluating VLMs and textual similarity measures, with emphasis on approaches to visual categorization and hierarchical evaluation. A more general overview of the use of hierarchical labels for visual categorization is given in §17 in the supplement. + +# 2.1. Evaluating VLMs for Classification + +Recent work has explored VLM evaluation in open-domain visual recognition and question answering, often as a byproduct of general model development. A dominant approach involves retrieving the most similar target class label to an unconstrained text using maximum inner-product search [15, 31]. The retrieved label can then be directly used for classification or further processed to extract candidate labels. In some cases [26], these extracted labels are compared to the original image using image-to-text similarity methods, such as CLIP [55]. In-context learning for classification [27] leverages large language models and a few example labels to generate or extract an image label from a longer caption. Other approaches rely on models like GPT-4 [50] to score generated answers based on a given question and reference answer [4, 34, 46, 87]. A recent method [26] expands classification label spaces by prompting models to produce increasingly specific labels using taxonomically structured templates extracted from ImageNet [18]. However, final model predictions are still evaluated using string-matching measures. Some multi-task VLM bench + +marks include image classification tasks [9, 78], typically using exact-match evaluation or text-based measures borrowed from image captioning; see §17 in supplementary material. The Open-domain Visual Entity (OVEN) benchmark [31] relies on representation-based methods while other work avoids direct classification tasks altogether [84]. + +# 2.2. Annotation and Prediction Similarity + +Language generation tasks such as image captioning have traditionally been evaluated using measure based on string similarity, e.g., ROUGE [42], BLEU [53],METEOR [5] and CIDER [69]. They primarily rely on surface form (string) overlap in prediction and ground truth labels, and as such, do not capture semantic similarities. This led to a family of trained measures that rely on a language model that, given a pair of inputs, predicts a similarity score, typically based on the semantic match. Examples of these are BERTScore [82], and SentenceBERT [58]. Causal and bi-directional language models have been shown to encode varied hierarchical semantic knowledge in human-interpretable metric spaces [28, 52, 54]. Natural language inference (NLI), also known as textual entailment [11, 73], can also capture hierarchical specificity judgments, since sub-categories entail their supercategories, i.e., This is a field sparrow entails This is a bird. In computer vision, this idea has been adopted as visual entailment [77]. + +Contrastive vision and language models [55, 81] can be used to compare image and text pairs using top- $k$ retrieval by inner product similarity. This method relies on the intrinsic quality of the representations. While impressive, these have been shown to fail in cases requiring combining complex compositional information that was not seen together at training time [80]. More recently, prompt-based methods for scoring textual similarity have been proposed: For instance, the LLaVA models [46] are evaluated on the COCO[44] image captioning datasets by asking GPT-4 [50] to provide numerical values (1-10) for a range of measures such as relevance and accuracy. Another approach is fine-tuning large language models to act as evaluation judges [87]. This approach has also been claimed to outperform other evaluation methods for machine translation [34]. A more detailed description of the text similarity measures we compare against in this paper is given in the supplement in section $\S 10$ . + +# 3. Taxonomy-aware Evaluation + +Taxonomy-aware classification evaluation is a subtle and understudied problem [33, 35, 62]. It is necessary to assign partial credit to answers that share some taxonomic similarity, e.g., to answers that are within the correct sub-tree, even if they do not match the correct node exactly. Hierarchical precision (hP) and hierarchical recall (hR) [33] are two existing measures that suitably assign partial credit. Both + +take the length of the longest common path from the root node of the taxonomy—equivalent to the size of the ancestor set—and normalize by either the length of the predicted path (hP) or the target path (hR). + +We now introduce $\mathbf{hP}$ and $\mathbf{hR}$ formally. Let $\Sigma$ be an alphabet of symbols, and $\Sigma^{*}$ be the Kleene closure, i.e., the set of all strings whose symbols are drawn from $\Sigma$ . We assume the VLM under consideration constitutes a distribution over $\Sigma^{*}$ when it is conditioned on an image. We then define a taxonomy $\mathcal{T}$ as a triplet $(V,\mathcal{L},E)$ , where $V$ represents a finite set of nodes, $\mathcal{L}\colon V\to \Sigma^{*}$ is a function that maps each node to a string, called a label, and $E\subset V\times V$ is a set of directed edges between pairs of nodes. Additionally, we require that the directed graph $(V,E)$ forms a directed tree with a unique root $\rho \in V$ . The set of nodes encountered along the unique path from the taxonomy tree's root node $\rho$ to a given node $v$ constitutes that node's ancestor set, which we denote $\operatorname{anc}(v)$ . Formally, for any node $v\in V$ , let $\operatorname{par}(v)$ denote $v$ 's unique parent node in the tree; $\operatorname{par}(\rho)$ , i.e., with the root note $\rho$ as an argument, is undefined. We then recursively define $\operatorname{anc}(v)$ as: + +$$ +\operatorname {a n c} (v) \stackrel {\text {d e f}} {=} \left\{\rho \right\} \quad (1 \mathrm {a}, \text {b a s e c a s e}; v = \rho) +$$ + +$$ +\operatorname {a n c} (v) \stackrel {\text {d e f}} {=} \{v \} \cup \operatorname {a n c} (\operatorname {p a r} (v)). \quad (1 b, \text {i n d u c t i v e c a s e}; v \neq \rho) +$$ + +We now discuss how to compare ground-truth and predicted nodes on a taxonomy. Let $\{(v_n^{\mathrm{gt}}, v_n^{\mathrm{pr}})\}_{n=1}^N$ be a set of pairs of nodes where $v_n^{\mathrm{gt}}$ is the ground-truth node for the $n^{\mathrm{th}}$ image and $v_n^{\mathrm{pr}}$ is the node mapped from the string generated from the VLM conditioned on the $n^{\mathrm{th}}$ image. (We describe a procedure to create this string-node mapping in §6.) We then define hP and hR over $\{(v_n^{\mathrm{gt}}, v_n^{\mathrm{pr}})\}_{n=1}^N$ as follows + +$$ +\mathbf {h} \mathbf {P} \stackrel {\text {d e f}} {=} \frac {1}{N} \sum_ {n = 1} ^ {N} \frac {\left| \operatorname {a n c} \left(v _ {n} ^ {\mathrm {p r}}\right) \cap \operatorname {a n c} \left(v _ {n} ^ {\mathrm {g t}}\right) \right|}{\left| \operatorname {a n c} \left(v _ {n} ^ {\mathrm {p r}}\right) \right|}, \tag {2a} +$$ + +$$ +\mathbf {h} \mathbf {R} \stackrel {\text {d e f}} {=} \frac {1}{N} \sum_ {n = 1} ^ {N} \frac {\left| \operatorname {a n c} \left(v _ {n} ^ {\mathrm {p r}}\right) \cap \operatorname {a n c} \left(v _ {n} ^ {\mathrm {g t}}\right) \right|}{\left| \operatorname {a n c} \left(v _ {n} ^ {\mathrm {g t}}\right) \right|}. \tag {2b} +$$ + +The hierarchical F1 (hF) is then defined as the harmonic mean of hP and hR, + +$$ +\mathbf {h} \mathbf {F} \stackrel {\text {d e f}} {=} \frac {2 \cdot \mathbf {h} \mathbf {P} \cdot \mathbf {h} \mathbf {R}}{(\mathbf {h} \mathbf {P} + \mathbf {h} \mathbf {R})}. \tag {3} +$$ + +Both hP and hR fall in the interval [0, 1]. The maximum is achieved when the target and predicted paths coincide. Less-than-perfect scores capture the extent to which the path deviates from the gold answer. + +hP Hierarchical precision captures the amount of incorrect information in the prediction, i.e., deviations from the correct path, relative to the number of nodes shared with the correct path. + +TRAIN +TELEPHOTO LENS +Prediction: A lens cover +![](images/52e6d36176c7d6fc7bc8ffb25896dfbe5d9a2f9e0d38dfa3dd12aef346a52a42.jpg) +hP:1.00 hR:0.75 + +Prediction: A mode of transport +![](images/3179c14f5f01c22d2cb6cd3f8f0a06dd058f53363d25289a47cb8f959aa3c3d8.jpg) +hP:0.80 hR:0.67 + +PITH HELMET +Prediction: Neckwear +Figure 3. Illustrative examples. Hierarchical precision (hP) and recall (hR) calculations on random label pairs ( $\S 5$ ). While hP penalizes incorrect labels, i.e., labels that are further away from the target's ancestor set, hR penalizes mistakes made higher up on the taxonomy. +![](images/87203c2258a6970c3db400abcd329e1130a5a2b1cf5e0917d91afc98e9d4924c.jpg) +hP:0.75 hR:0.50 + +STIRLING CASTLE +Prediction: Château de Ranrouët +POOL +Prediction: High jump +![](images/9daca7581acc0e5907b52be97efed0edd0f2d74d2ea5fe30f7928cbc811cb51c.jpg) +hP:0.83 hR:0.83 + +![](images/cfa2ff0cc15dcd127bc7dbe7a3537c647d6f7b54d4d328256474468e228850d5.jpg) +hP:0.67 hR:0.67 + +hR Hierarchical recall measures the amount of correct information in the prediction, i.e., how much the path from the root to the predicted node overlaps with the path from the root to the ground truth node. Hierarchical recall thus penalizes missing coverage of the target path, especially early deviations. This can be seen as measuring the specificity of the prediction, in particular when the hP is 1. + +We give an example of the computation of $\mathbf{hP}$ and $\mathbf{hR}$ in Fig. 3, sampled from the OVEN datasets. In the case of the picture of TRAIN, we can see how only predicting MODE OF TRANSPORT leads to lower $\mathbf{hR}$ since it lacks information, while $\mathbf{hP}$ remains 1 since there is no incorrect information. For the sport POOL and STIRLING CASTLE, we notice how both scores drop, as incorrect nodes are predicted—these responses contain both incorrect information and some partially correct information. In the case of PITH HELMET, we see that the prediction is incorrect yet not very specific, resulting in a lower $\mathbf{hR}$ than $\mathbf{hP}$ . We also note that it is not possible to have a lower $\mathbf{hP}$ than $\mathbf{hR}$ if the ground truth nodes are leaf nodes. + +# 4. Taxonomically Linked Datasets + +To investigate taxonomy-aware evaluation measures for text generated from a VLM, we consider two taxonomies linked to FGVC tasks. First, the Catalogue of Life taxonomy as contained in the iNaturalist21 dataset [30] is a clean, expert-curated taxonomy that maps evolutionary relationships between species and contains exactly 10,000 leaf nodes. The second is a more general taxonomy constructed from the Wikidata [71] knowledge graph, which links entities from FGVC datasets in the Open-domain Visual Entity (OVEN) benchmark [31]. + +# 4.1. Extracting Taxonomies + +While the iNaturalist21 dataset includes a built-in taxonomy, this is not the case for other FGVC datasets. For those included in the OVEN collection, we utilize Wikidata—a large-scale knowledge graph derived from Wikipedia [71]. + +Unlike taxonomies, knowledge graphs are not constrained to tree or even directed acyclic graph structures, which necessitates the extraction of a taxonomy that fits the formal definition given in §3. To construct such a tree from Wikidata, we rely on the subclass of relation. When multiple paths exist from the root to a given node in the taxonomy, we retain the longest path. In the case of ties, we sample one of the tied paths uniformly at random. To ensure a rooted directed tree, we exclude higher-level abstract classes that introduce cycles. Because OVEN already links its datasets (§4.2) to Wikipedia entries, we can directly map them to Wikidata entities and incorporate them into our taxonomy. + +# 4.2. Entity Linked Datasets of Images + +The iNaturalist21 dataset [30] was compiled and curated from the citizen science platform iNaturalist. It contains 2.7 million images spanning 10,000 species. The OVEN dataset [31] aggregates several FGVC datasets, linking them to Wikipedia identifiers and natural language questions to support the evaluation of open-domain visual entity recognition. It includes ImageNet21k-P [61, 63], iNaturalist2017 [29], Cars196 [36], SUN397 [76], Food101 [10], Sports100 [25], Aircraft [47], Oxford Flowers [49], and Google Landmarks v2 [56, 75]. We exclude iNaturalist2017 from our OVEN-based evaluations, as we separately analyze the newer iNaturalist21 dataset, which is aligned with the higher-quality Catalogue of Life taxonomy. + +# 5. The Woes of Existing Measures + +In this section, we evaluate whether non-taxonomic text similarity measures already encode sufficient taxonomic information, e.g., through label overlap or representational similarity. To investigate this, we conduct a targeted synthetic experiment comparing the taxonomic similarity measures (hP and hR), with textual and multimodal similarity measures from §2.2 commonly used to assess the quality of language model outputs. These measures are both string-based and representation-based. + +
MeasureiNat21OVEN
τ-hPτ-hRτ-hPτ-hR
EM0.010.070.010.01
Contained0.020.090.020.03
stringROUGE0.180.410.170.21
METEOR0.180.400.180.22
rep.BERTscore0.010.310.270.18
SentBERT0.020.250.340.31
NLI0.08-0.20-0.170.19
CLIP-t2t0.140.460.250.15
CLIP-i2t0.350.490.350.34
+ +Table 1. Correlation of text similarity measures with hierarchical metrics. Kendall's $\tau$ correlations between taxonomic metrics hP, hR, and the different similarity measures on the synthetic evaluation sets. All correlations have a $p$ -value $< 0.0001$ . + +Experimental Setup. We construct two synthetic evaluation sets to determine to what extent text similarity measures reflect taxonomic structure. We consider the taxonomies from iNaturalist21 and the Wikidata (see §4) to generate controlled pairs of labels corresponding to pairs of nodes in each taxonomy. A reference node $v$ and a candidate node $v'$ are chosen such that $v$ is always a leaf node and $v'$ is at some given distance from $v$ . The distance between two nodes in the taxonomy is the length of the shortest undirected path between the two nodes in the taxonomy. To ensure a balanced dataset, we sample the node pairs such that their pairwise distance is uniformly distributed up to a maximum distance of 7, depending on the depth of the node. To evaluate hP, we sample 100,000 reference-candidate node pairs from the taxonomies. To evaluate hR in a controlled fashion, however, we additionally require that $v' \in \mathrm{anc}(v)$ , i.e., that $\mathbf{hP} = 1$ . This allows us to isolate the sensitivity of the similarity measures at decreasing levels of specificity. $^{1}$ For each pair, we compute both the taxonomy-aware score and text-based measures between the corresponding labels $\mathcal{L}(v)$ and $\mathcal{L}(v')$ . We then measure the rank correlation between the taxonomy-aware score and various text-based measures using Kendall's $\tau$ , which evaluates whether the similarity measure reflects taxonomic proximity under controlled conditions. By construction, we use the node label $\mathcal{L}(v')$ as the prediction to avoid the problem of mapping unconstrained output onto the taxonomy. All of the measures we consider are text-based (see §10 in the supplement for details), except the CLIP image-to-text score, where we sample an image for the reference node from the corresponding dataset and measure CLIP model similarity between the image and the candidate label. We + +![](images/fd61692ce31479b2c85a0d2d1b72c9b0912976249f281b87872c18d53a203420.jpg) +Figure 4. Agreement in node placement using different similarity measures. We observe variation in node placement for all measures. Only METEOR and ROUGE, and the CLIP variations when combined with Alg. 1 share predictions frequently. + +choose high-performing models that are not prohibitively large for the model-based comparisons.2 + +Results. Table 1 presents correlations between hP and hR on one hand and the text similarity measures from §2.2 and CLIP-i2t on the other, for the iNaturalist21 and Wiki-data labels. We observe that surface-level measures (Exact Match (EM), Contained) exhibit a low correlation with the taxonomy-aware measures hP and hR. ROUGE and METEOR, which are based on $n$ -gram overlap, capture taxonomic similarity more effectively. This is presumably due to label overlap, e.g., GRAY SEAL is a type of SEAL. The highest correlation is achieved by CLIP image-to-text (CLIP-i2t) for both datasets, followed by CLIP text-to-text (CLIP-t2t), except for hP on the iNaturalist21 dataset. In that case, CLIP-t2t fails to capture deviations from the correct taxonomic path. This suggests that CLIP's image features may encode more taxonomic information than its text representations, which can struggle due to the low-frequency vocabulary inherent to this task. NLI is an outlier, exhibiting a negative correlation in most cases. While NLI captures taxonomic similarity for closely related labels, it performs poorly on more distantly related pairs. This is visualized in Fig. 8 in the supplement. + +Summary. Overall, we observe that representation-based measures capture some notions of taxonomic distance along the path from the leaf node to the root (hR), but are, for the most part, poorly equipped to measure taxonomic distances across different subtrees (hP). + +
MeasurehPhRhFAccuracy
Exact Match0.370.420.3917.5%
Contained0.490.500.4924.3%
BLEU0.710.530.5424.3%
METEOR0.620.650.6329.1%
ROUGE0.620.660.6330.0%
BERTScore0.450.520.4811.3%
SentBERT0.580.670.6220.0%
CLIP-t2t0.720.790.7530.3%
CLIP-i2t0.580.710.235.3%
CLIP-i2t+Alg. 10.800.810.8044.0%
CLIP-t2t+Alg. 10.790.820.8047.1%
+ +Table 2. Evaluating taxonomy mapping with human annotation. We report the performance of text similarity measures (directly) and CLIP image-to-text similarity for mapping VLM text predictions to the taxonomy using a subset of iNaturalist21 samples that were manually mapped to the correct nodes. + +# 6. Mapping VLM Predictions to Taxonomies + +The key challenge for taxonomy-aware evaluation of unconstrained text generated from VLMs is mapping the text generated by a VLM onto nodes in a given taxonomy. A baseline approach to this mapping is to use text similarity measures (e.g., those in §2.2) to calculate the best match between the generated text and the labeled nodes in a taxonomy. We also develop a heuristic approach that performs this mapping using a general similarity measure as a subroutine, which makes using CLIP text-to-image similarity feasible. Specifically, this algorithm combines string-based matching with scores from a given similarity measure to position predictions onto the taxonomy. We describe this procedure in words below and give pseudocode in §11 (Alg. 1 supplement). + +Our goal is to map each VLM-generated prediction string to a node in the taxonomy using a series of increasingly flexible matching strategies. Generally, we attempt exact matches first, then allow for partial overlaps or ambiguous cases, always preferring more specific nodes when multiple candidates are viable. Before matching, we normalize all predictions and node labels by lowercasing, replacing dashes with spaces, and stripping other punctuation. We then compute similarity scores between the prediction and all node labels using the CLIP similarity measures. More specifically, the mapping proceeds in stages. We first check whether any of the top- $k$ $(k = 10)$ highest-scoring nodes, according to the CLIP measure applied to their labels, appear verbatim in the prediction string; if so, we return the corresponding node. If this fails, we search for verbatim matches against all other labels in the taxonomy. If no full-string matches are found, we repeat these checks using $n$ -gram overlap for $n \in \{4,3,2\}$ . If multiple candidates remain with similar scores, we apply a fallback based on score ambiguity: if the difference between the top two scores, and + +between the top and $k^{\text{th}}$ score, are both below threshold values, we treat the case as ambiguous. We then search for a common ancestor among the top candidates. If any node appears sufficiently often (at least four times) in the ancestor sets of multiple nodes in the top- $k$ highest-scoring nodes we return the most specific such node, defined as the one deepest in the taxonomy. If none of the prior criteria are met, we fall back to returning the single highest-scoring candidate. Thus, the algorithm always returns a node in the taxonomy. If multiple nodes match the above criteria, we return the most specific one, i.e., the node deepest in the taxonomy. The "shared ancestor" step is particularly beneficial for the CLIP image-to-text similarities since we do not have images corresponding to higher-order taxonomic concepts. + +# 6.1. Mapping Quality + +To evaluate the quality of the taxonomic node proposals we measure their agreement against manually annotated node positions. We compare the similarity measures described in §2.2 (and Supplement §10) directly against all candidate node labels. We then do the same using Alg. 1, combined with the CLIP measures. To construct the test set, we gather 416 unique VLM outputs for the iNaturalist21 dataset, randomly sampled from the generations described in §7. For each output, we manually find the best matching taxonomic node given the string, e.g., matching This is a bird to AVES/BIRDS. + +Results. Results are shown in Tab. 2. We find string-matching measures (Exact Match and Contained) do fairly poorly, while other surface-level measures (BLEU, METEOR, and ROUGE) perform slightly better. The representation-based approaches do not perform better in general, except for CLIP text-to-text. Fig. 4 shows that the string-level measures have considerable diversity in their predictions, while the representation-based measures show more diversity. The best performance is found with CLIP text-to-text and CLIP image-to-text using the algorithm described above (Alg. 1), which results in the same hF score but slightly improved exact node match accuracy score for CLIP text-to-text + Alg. 1. + +# 7. Evaluating and Comparing VLMs + +We compare one closed model and seven publicly available VLMs on fine-grained visual classification using our taxonomy-aware evaluation scheme. To map (unconstrained) strings generated from a VLM to nodes in the taxonomy, we use the best-performing similarity method (CLIP-t2t + Alg. 1), enabling the computation of hP, hR, and hF. These scores offer insight into the level of accuracy (hP) and specificity (hR) of VLM predictions for a given taxonomy. The resulting rankings differ significantly from + +
Model Name# ParametersTraining PipelineOpenVLM Avg.
GPT-4 [50]UnknownUnknown63.1
ILmXC2 [23]7BPretr., instr.-tuning62.2
OLM12B [79]12BPretr., instr.-tuning, RLHF54.6
QVLChat [3]9.6BPretr., instr.-tuning51.3
OLM3B [32]3BPretr., instr.-tuning47.8
LLaVA [45]7.2BPretr., instr.-tuning45.9
QwenVL [3]9.6BPretr.21.7
Fuyu [6]8BPretr.N/A
+ +Table 3. Overview of evaluated VLMs. Details on the models we evaluate, including number of parameters, training method, and average performance on a large vision-language benchmark. + +those based on standard text similarity measures, highlighting the taxonomy-aware measure's ability to capture previously unmeasured aspects of model performance. + +Experimental Setup. We compare eight VLMs that differ in model size, training configuration (pretrained-only, instruction-tuned, and preference-aligned [51]), and performance on the Open VLM Leaderboard [16], a collection of 13 VL benchmarks with 38 models. We summarize model sizes, training configuration, and average performance on the Open VLM leaderboard in Table 3. For both the OVEN-sourced FGVC datasets and iNaturalist21, we randomly sample 6000 images from the validation set. For iNaturalist21, we use the prompt What species is this?, while for OVEN the question is included in the dataset and varies by object type. To understand how well we can control hR and hP for instruction-tuned models, we construct two different sets of instructions: (1) one containing a minimal instruction for answering in a specified format, and (2) one containing an instruction to be specific. We use slight variations of the prompt formatting to fit each model (as specified by the model developers); the exact prompts are provided in the supplement (Tab. 7 and Tab. 8). Because non-instruction-tuned models do not respond meaningfully to instructions, we prompt them with a straightforward question-answer template $Q$ : {}A:. We use an orange-colored {} to indicate a variable in the template. + +# 7.1. Experimental Results + +# How does the ranking of VLM performance change when considering taxonomy-aware measures? In + +Fig. 5, we rank eight VLMs based on their performance on FGVC according to the different similarity measures corresponding to the values given in Tab. 4 and Tab. 5 ( $\S 14$ in supplement). We find that several text similarity measures yield similar rankings when used as direct evaluation criteria. In contrast, the taxonomy-aware measures lead to notable changes in model rankings. For instance, LLaVA ranks lowest under Exact Match for + +![](images/03ccd528a3a877782dcf4dc0c722d3d7cbbe9e755d2eaa5b0a769b31ce593594.jpg) + +![](images/c137a57b5900c84316c3f9ff7471a49277b0d973d16ec6059948d53561d23e01.jpg) +Figure 5. Ranking of VLMs for the iNaturalist21 (top) and OVEN (bottom) datasets. We evaluate the ranking of VLMs (vertical axis) based on different evaluation measures (horizontal axis). The best model is shown in the top row. On the left, we see the model names ranked by exact match and on the right ranked by the hierarchical F1 hF. See §14, Tab. 4 in supplement for exact numbers. + +iNaturalist21, yet achieves high hierarchical precision ( $\mathbf{hP}$ ) and, consequently, a strong hierarchical F1 ( $\mathbf{hF}$ ). This indicates that LLaVA deviates from the correct taxonomy path less frequently than other models. However, its low hierarchical recall ( $\mathbf{hR}$ ) suggests that its predictions are often not specific. Another example is GPT-4, which ranks highest on all measures except $\mathbf{hP}$ and $\mathbf{hF}$ for iNaturalist21. Compared to QVLChat, which is ranked first in both $\mathbf{hP}$ and $\mathbf{hF}$ , we observe that GPT-4 makes more specific but erroneous predictions. Our taxonomy-aware measure offers a complementary perspective to traditional metrics by revealing performance differences that standard similarity-based measures overlook. + +Do taxonomy-aware measures capture prompt sensitivity? We prompt the VLMs using two distinct prompts, as described earlier, to examine how their predictions shift in terms of hierarchical precision (hP) and recall (hR), shown jointly in Fig. 7. Fuyu and QwenVL are not instruction-tuned and therefore do not respond differently, while LLaVA shows only minimal changes. In contrast, the other models clearly adjust their predictions based on the different instructions in the prompts. On the iNaturalist21 dataset, all models except GPT-4 exhibit a trade-off between hP and hR: higher precision tends to come at the cost of recall. GPT-4, however, offers both more accurate and more fine-grained classifications when prompted appropriately. On the OVEN dataset, results are more mixed: roughly half of the models improve on both axes with prompt changes. + +![](images/c35983c1850958d662d8c80a44da3944dc31a7a67a0cb64e2fa9bb8889123cae.jpg) +Figure 6. Prompt tuning results for bird classifier. For High hP (●), we target prompts that prioritize hP. For Acc (▲) we optimize for prompts that give higher binary accuracy. This demonstrates how hP and hR can help tune a VLM application. + +These findings suggest that each application may prioritize different target values for $\mathbf{hP}$ and $\mathbf{hR}$ , and that models can adapt substantially to meet those needs. In such cases, prompt design becomes a powerful paradigm for tuning model behavior as further explored in the next section. + +# 7.2. An Application: High-Precision Classification + +To demonstrate the usefulness of hR and hP in tuning prompts for a specific application, we prompt Llama 3.2 Vision (11B) [65] to classify birds. The application goal is to minimize the amount of wrong information in the prediction, which is equivalent to maximizing hierarchical precision but hard to formulate in terms of exact match binary classification accuracy. We use prompt-tuning with hP and hR as feedback signals, and compare to the same procedure but with accuracy as the feedback signal. We start with a basic prompt What is this an image of? Answer in the format 'A: .' and task ChatGPT [50] to improve on the prompt for 30 iterations. A system prompt describes the scores and the goal of developing a conservative classifier that gives a less specific prediction if uncertain (see §15 in the supplement for more details). In each iteration, we include past prompts along with their resulting hR and hP scores, or the accuracy score, in a few-shot learning setup. We tune the prompts with 20 images and evaluate on 200 images of birds from iNaturalist21. The results are shown in Fig. 6. While this is a toy example, we see how the taxonomy-aware measures not only enable a higher granularity of model performance analysis, but can also facilitate targeted application development. + +# 8. Discussion and Limitations + +Evaluating unconstrained text generated from a VLM is inherently challenging: there are countless entities to describe and even more ways to describe them. As the computer vision community shifts from closed-set recognition to open-ended recognition, evaluation increasingly resembles an NLP task. In this paper, we introduce structure into + +![](images/23719d906c712a5377c992d8275f07344af203736b7742367de560782056ad2a.jpg) +Figure 7. Prompt sensitivity captured by hR and hP. We compare two prompt versions, a minimal barebones version and one that asks for a taxonomically specific prediction. + +VLM evaluation by leveraging curated taxonomies and established hierarchical evaluation measures. Despite this, several limitations remain. First, placing unconstrained text on a taxonomy is a low-resource problem; there are no large-scale datasets available to evaluate our mapping approach, nor sufficient data to train a dedicated classifier that links VLM outputs to taxonomy nodes. To address this, we repurpose existing text similarity measures, along with CLIP, to perform the mapping. While our evaluation shows that this approach is non-trivial and introduces errors that affect downstream VLM evaluation, we are optimistic about future improvements, such as training on synthetic data generated by language models. Second, taxonomies, especially those derived from Wikidata, are inherently noisy. Even expert-curated taxonomies like the one used for iNaturalist21 can have uneven subtree granularity, complicating interpretations based on globally averaged measures. One potential solution would be to employ expert input to assign weighted edges to reflect semantic distances between nodes. + +# 9. Conclusion + +Text similarity measures from NLP are a natural place to look for evaluation of unconstrained VLM output. Additionally, many FGVC tasks come with side information about the class structure, namely taxonomies, which text similarity measures do not fully capture, as we have shown in this paper. To solve this gap, we have proposed a framework for using hierarchical evaluation measures to evaluate VLMs for FGVC, enabling a more fine-grained view of the specificity and accuracy of these models. + +# Acknowledgements + +Vésteinn Snæbjarnarson, Serge Belongie, Nico Lang, and Stella Frank are supported by the Pioneer Centre for AI, DNRF grant number P1. Niklas Stoehr is supported by the Swiss Data Science Center (SDSC) PhD Fellowship. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In Neural Information Processing Systems, pages 23716-23736, 2022. 1 +[2] Majed A Alkhamees, Mohammed A Alnuem, Saleh M Al-Saleem, and Abdulrakeeb M Al-Ssulami. A semantic metric for concepts similarity in knowledge graphs. Journal of Information Science, 49:778-791, 2023. 4 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv, 2023. 7 +[4] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. TouchStone: Evaluating vision-language models by language models. arXiv, 2308.16890, 2023. 2 +[5] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 3, 2 +[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for AI agents, 2023. 1, 7 +[7] Walter Bennette, Nathaniel Hofmann, Nathaniel Wilson, and Tyler Witter. Hierarchical open-set recognition for automatic target recognition. In 2021 IEEE Symposium Series on Computational Intelligence (SSCI). IEEE, 2021. 4 +[8] Luca Bertinetto, Romain Mueller, Konstantinos Tertikas, Sina Samangooei, and Nicholas A Lord. Making better mistakes: Leveraging class hierarchies with deep networks. In Conference on Computer Vision and Pattern Recognition, pages 12506-12515, 2020. 4 +[9] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. VisIT-Bench: A benchmark for vision-language instruction following inspired by real-world use. In Neural Information Processing Systems, 2023. 3 +[10] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – Mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 4 +[11] Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In Conference on Empirical Methods in Natural Language Processing, 2015. 3, 2 + +[12] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. MiniGPT-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv, 2023. 1 +[13] Elijah Cole, Kimberly Wilber, Grant Van Horn, Xuan Yang, Marco Fornoni, Pietro Perona, Serge Belongie, Andrew Howard, and Oisin Mac Aodha. On label granularity and object localization. In European Conference on Computer Vision, pages 604-620, 2022. 4 +[14] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 4 +[15] Alessandro Conti, Enrico Fini, Massimiliano Mancini, Paolo Rota, Yiming Wang, and Elisa Ricci. Vocabulary-free image classification. In Neural Information Processing Systems, 2023. 2 +[16] OpenCompass Contributors. OpenCompass: A universal evaluation platform for foundation models, 2023. 7 +[17] Riccardo De Lutio, Yihang She, Stefano D'Aronco, Stefania Russo, Philipp Brun, Jan D Wegner, and Konrad Schindler. Digital taxonomist: Identifying plant species in community scientists' photographs. ISPRS journal of photogrammetry and remote sensing, 182:112-121, 2021. 4 +[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 1, 2 +[19] Jia Deng, Jonathan Krause, Alexander C. Berg, and Li Fei-Fei. Hedging your bets: Optimizing accuracy-specificity trade-offs in large scale visual recognition. In Conference on Computer Vision and Pattern Recognition, pages 3450–3457, 2012. 4 +[20] Xiwen Dengxiong and Yu Kong. Ancestor search: Generalized open set recognition via hyperbolic side information learning. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4003-4012, 2023. 4 +[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics, pages 4171–4186, 2019. 2 +[22] Ankit Dhall, Anastasia Makarova, Octavian Ganea, Dario Pavllo, Michael Greeff, and Andreas Krause. Hierarchical image classification using entailment cone embeddings. In Conference on Computer Vision and Pattern Recognition, pages 836-837, 2020. 4 +[23] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2: Mastering freeform text-image composition and comprehension in vision-language large model. arXiv, 2401.16420, 2024. 7 + +[24] S. Doveh, A. Arbelle, S. Harary, E. Schwartz, R. Herzig, R. Giryes, R. Feris, R. Panda, S. Ullman, and L. Karlinsky. Teaching structured Vision & Language concepts to Vision & Language models. In Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 4 +[25] Kaggle (Gerry). 100 sports image classification, 2022. 4 +[26] Simon Ging, Maria Alejandra Bravo, and Thomas Brox. Open-ended VQA benchmarking of vision-language models by exploiting classification datasets and their semantic hierarchy. In International Conference on Learning Representations, 2024. 2 +[27] Sherzod Hakimov and David Schlangen. Images in language space: Exploring the suitability of large language models for vision & language tasks. In Findings of the Association for Computational Linguistics, pages 14196-14210, 2023. 2 +[28] Yuan He, Zhangdie Yuan, Jiaoyan Chen, and Ian Horrocks. Language models as hierarchy encoders. In Neural Information Processing Systems, 2024. 3 +[29] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexx Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The iNaturalist species classification and detection dataset. In Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 4 +[30] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin MacAodha. Benchmarking representation learning for natural world image collections. In Conference on Computer Vision and Pattern Recognition, pages 12879-12888, 2021. 2, 4 +[31] Hexiang Hu, Yi Luan, Yang Chen, Urvashi Khandelwal, Mandar Joshi, Kenton Lee, Kristina Toutanova, and Ming-Wei Chang. Open-domain visual entity recognition: Towards recognizing millions of wikipedia entities. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12065–12075, 2023. 1, 2, 3, 4 +[32] Shengding Hu, Yuge Tu, Xu Han, Ganqu Cui, Chaoqun He, Weilin Zhao, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Xinrong Zhang, Zhen Leng Thai, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, dahai li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the potential of small language models with scalable training strategies. In Conference on Language Modeling, 2024. 7 +[33] Svetlana Kiritchenko, Stan Matwin, Richard Nock, and Fazel Famili. Learning and evaluation in the presence of class hierarchies: Application to text categorization. In Canadian AI, 2006. 2, 3 +[34] Tom Kocmi and Christian Federmann. Large language models are state-of-the-art evaluators of translation quality. In Conference of the European Association for Machine Translation, pages 193–203, 2023. 2, 3 +[35] Aris Kosmopoulos, Ioannis Partalas, Eric Gaussier, Georgios Paliouras, and Ion Androutsopoulos. Evaluation measures for hierarchical classification: A unified view and novel approaches. Data Mining and Knowledge Discovery, 29(3): 820-865, 2015. 3 +[36] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D object representations for fine-grained categorization. In + +International Conference on Computer Vision Workshops, pages 554-561, 2013. 4 +[37] Suren Kumar and Rui Zheng. Hierarchical category detector for clothing recognition from visual data. In International Conference on Computer Vision Workshops, pages 2306-2312, 2017. 4 +[38] Nico Lang, Vésteinn Snæbjarnarson, Elijah Cole, Oisin Mac Aodha, Christian Igel, and Serge Belongie. From coarse to fine-grained open-set recognition. In Conference on Computer Vision and Pattern Recognition, pages 17804-17814, 2024. 4 +[39] Kibok Lee, Kimin Lee, Kyle Min, Yuting Zhang, Jinwoo Shin, and Honglak Lee. Hierarchical novelty detection for visual object recognition. In Conference on Computer Vision and Pattern Recognition, pages 1034-1042, 2018. 4 +[40] Clément Lefebvre and Niklas Stoehr. Rethinking the event coding pipeline with prompt entailment. In EACL Workshop Fact Extraction and VERIFICATION, 2023. 2 +[41] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, 2023. 1 +[42] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81. Association for Computational Linguistics, 2004. 1, 3, 2 +[43] Dekang Lin. An information-theoretic definition of similarity. In International Conference on Machine Learning, page 296-304, 1998. 4 +[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, pages 740-755, 2014. 3 +[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Conference on Computer Vision and Pattern Recognition, 2023. 7 +[46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Neural Information Processing Systems, 2023. 1, 2, 3 +[47] Subhransu Maji, Esa Rahtu, Juho Kannala, Mathew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv, 1306.5151, 2013. 4 +[48] Kenneth Marino, Ruslan Salakhutdinov, and Abhinav Gupta. The more you know: Using knowledge graphs for image classification. In Computer Vision and Pattern Recognition, 2016. 4 +[49] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 4 +[50] OpenAI. GPT-4 technical report, 2023. 1, 2, 3, 7, 8 +[51] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda + +Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Neural Information Processing Systems, 2022. 7 +[52] Jeff Z. Pan, Simon Razniewski, Jan-Christoph Kalo, Sneha Singhania, Jiaoyan Chen, Stefan Dietze, Hajira Jabeen, Janna Omeliyanenko, Wen Zhang, Matteo Lissandrini, Russia Biswas, Gerard de Melo, Angela Bonifati, Edlira Vakaj, Mauro Dragoni, and Damien Graux. Large language models and knowledge graphs: Opportunities and challenges. Transactions on Graph Data and Knowledge, 2023. 3 +[53] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. In Annual Meeting on Association for Computational Linguistics, page 311-318, 2002. 3, 2 +[54] Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Conference on Empirical Methods in Natural Language Processing, pages 2463-2473, 2019. 3 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3 +[56] Elias Ramzi, Nicolas Audebert, Clement Rambour, Andre Araujo, Xavier Bitot, and Nicolas Thome. Optimization of rank losses for image retrieval. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4 +[57] Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees GM Snoek. SelEx: Self-expertise in fine-grained generalized category discovery. In European Conference on Computer Vision, 2024. 4 +[58] Nils Reimers and Iryna Gurevych. Sentence-BERT: sentence embeddings using siamese BERT-networks. In Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3980-3990, 2019. 1, 3, 2 +[59] Philip Resnik. A class-based approach to lexical discovery. In Annual Meeting of the Association for Computational Linguistics, pages 327-329, 1992. 4 +[60] Philip Resnik. Using information content to evaluate semantic similarity in a taxonomy. In International Joint Conference on Artificial Intelligence, page 448-453, 1995. 4 +[61] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. In Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 4 +[62] Kevin Riehl, Michael Neunteufel, and Martin Hemberg. Hierarchical confusion matrix for classification performance evaluation. Journal of the Royal Statistical Society Series C: Applied Statistics, 72(5):1394-1412, 2023. 3 +[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, + +Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 4 +[64] Paolo Simeone, Raul Santos-Rodriguez, Matt McVicar, Jeffrey Lijffijt, and Tijl De Bie. Hierarchical novelty detection. In Advances in Intelligent Data Analysis, pages 310-321, 2017. 4 +[65] Hugo Touvron, Aaron Grattafori, Michael Auli, et al. The Llama 3 herd of models. arXiv, 2407.21783, 2024. 8 +[66] Mehmet Ozgur Turkoglu, Stefano D'Aronco, Gregor Perich, Frank Liebisch, Constantin Streit, Konrad Schindler, and Jan Dirk Wegner. Crop mapping from image time series: Deep learning with multi-scale label hierarchies. Remote Sensing of Environment, 264:112603, 2021. 4 +[67] Grant Van Horn, Steve Branson, Scott Loarie, Serge Belongie, and Pietro Perona. Lean multiclass crowdsourcing. In Conference on Computer Vision and Pattern Recognition, 2018. 4 +[68] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In International Conference on Learning Representations, 2022. 4 +[69] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Conference on Computer Vision and Pattern Recognition, 2015. 3 +[70] Nakul Verma, Dhruv Mahajan, Sundararajan Sellamanickam, and Vinod Nair. Learning hierarchical similarity metrics. In Conference on Computer Vision and Pattern Recognition, pages 2280-2287, 2012. 4 +[71] Denny Vrandecic and Markus Krötzsch. Wikidata: A free collaborative knowledgebase. Communications of the ACM, 57(10):78-85, 2014. 2, 4 +[72] Cunxiang Wang, Sirui Cheng, Qipeng Guo, Yuanhao Yue, Bowen Ding, Zhikun Xu, Yidong Wang, Xiangkun Hu, Zheng Zhang, and Yue Zhang. Evaluating open-QA evaluation. In Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2 +[73] Sinong Wang, Han Fang, Madian Khabsa, Hanzi Mao, and Hao Ma. Entailment as few-shot learner. arXiv, 2104.14690, 2021. 3, 2 +[74] Xiu-Shen Wei, Yi-Zhen Song, Oisin Aodha, Jianxin Wu, Yuxin Peng, Jinhui Tang, Jian Yang, and Serge Belongie. Fine-grained image analysis with deep learning: A survey. IEEE Transactions on Pattern Analysis & Machine Intelligence, 44(12):8927-8948, 2022. 1 +[75] Tobias Weyand, Andre Araujo, Bingyi Cao, and Jacjk Sim. Google Landmarks Dataset v2: A Large-Scale Benchmark for Instance-Level Recognition and Retrieval. In Conference on Computer Vision and Pattern Recognition, 2020. 4 +[76] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 4 +[77] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment: A novel task for fine-grained image understanding, 2019. 3, 2 + +[78] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Fanqing Meng, Siyuan Huang, Meng Lei, Ping Luo, and Yu Qiao. LVLM-eHub: A comprehensive evaluation benchmark for large vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3 +[79] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. RLHF-V: Towards trustworthy MLLMs via behavior alignment from fine-grained correctional human feedback. In Conference on Computer Vision and Pattern Recognition, 2024. 7 +[80] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In International Conference on Learning Representations, 2023. 3 +[81] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In International Conference on Computer Vision, pages 11941-11952, 2023. 3 +[82] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. BERTScore: Evaluating text generation with BERT. In International Conference on Learning Representations, 2020. 3, 2 +[83] Bingchen Zhao, Nico Lang, Serge Belongie, and Oisin Mac Aodha. Labeled data selection for category discovery. In European Conference on Computer Vision, 2024. 4 +[84] Wangchunshu Zhou, Yan Zeng, Shizhe Diao, and Xinsong Zhang. VLUE: A multi-task multi-dimension benchmark for evaluating vision-language pre-training. In International Conference on Machine Learning, pages 27395-27411, 2022. 3 +[85] Ganggao Zhu and Carlos A. Iglesias. Computing semantic similarity of concepts in knowledge graphs. IEEE Transactions on Knowledge and Data Engineering, 29(1):72-85, 2017. 4 +[86] Ganggao Zhu and Carlos A. Iglesias. Exploiting semantic similarity for named entity disambiguation in knowledge graphs. Expert Systems with Applications, 101:8-24, 2018. 4 +[87] Lianghui Zhu, Xinggang Wang, and Xinlong Wang. JudgeLM: Fine-tuned large language models are scalable judges. In International Conference on Learning Representations, 2025. 2, 3 \ No newline at end of file diff --git a/2025/Taxonomy-Aware Evaluation of Vision-Language Models/images.zip b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..52c6720b1bf87e42827f40d891650f0ba5c73ef9 --- /dev/null +++ b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44e081e6cbbe0050f52a1944c8833963113f7e5d2d98142f9c8565a4beea089f +size 404378 diff --git a/2025/Taxonomy-Aware Evaluation of Vision-Language Models/layout.json b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..03ca3b7648c16364f3fa118665818215a0d534f3 --- /dev/null +++ b/2025/Taxonomy-Aware Evaluation of Vision-Language Models/layout.json @@ -0,0 +1,9438 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "text", + "content": "Taxonomy-Aware Evaluation of Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "content": "Vésteinn Snæbjarnarson" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "content": " Kevin Du" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "content": " Niklas Stoehr" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "content": " Serge Belongie" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "content": " Ryan Cotterell" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "content": " Nico Lang" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "content": " Stella Frank" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "content": "University of Copenhagen " + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 510, + 193 + ], + "type": "text", + "content": "ETH Zürich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 258, + 297, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 547 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 547 + ], + "type": "text", + "content": "When a vision-language model (VLM) is prompted to identify an entity depicted in an image, it may answer \"I see a conifer,\" rather than the specific label NORWAY SPRUCE. This raises two issues for evaluation: Firstly, the unconstrained generated text needs to be mapped to the evaluation label space (i.e., CONIFER). Secondly, a useful classification measure should give partial credit to less-specific, but not incorrect, answers (NORWAY SPRUCE being a type of CONIFER). To meet these requirements, we propose a framework for evaluating unconstrained text predictions such as those generated from a vision-language model against a taxonomy. Specifically, we propose the use of hierarchical precision and recall measures to assess the level of correctness and specificity of predictions with regard to a taxonomy. Experimentally, we first show that existing text similarity measures do not capture taxonomic similarity well. We then develop and compare different methods to map textual VLM predictions onto a taxonomy. This allows us to compute hierarchical similarity measures between the generated text and the ground truth labels. Finally, we analyze modern VLMs on fine-grained visual classification tasks based on our proposed taxonomic evaluation scheme. Data and code are made available at https://github.com/vesteinn/vlm-eval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 587, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 696 + ], + "type": "text", + "content": "Modern vision-language models (VLMs) [1, 6, 12, 41, 46, 50] serve as viable one-stop shops for a wide variety of tasks at the intersection of vision and language. Given an image as input, VLMs generate unconstrained text, a key enabler of their flexibility across tasks. This flexibility, however, introduces additional challenges in their evaluation. In this paper, we study the problem of evaluating the unconstrained text generated by VLMs in the context of fine-grained visual categorization (FGVC), and put forward" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 329, + 233, + 542, + 316 + ], + "blocks": [ + { + "bbox": [ + 329, + 233, + 542, + 316 + ], + "lines": [ + { + "bbox": [ + 329, + 233, + 542, + 316 + ], + "spans": [ + { + "bbox": [ + 329, + 233, + 542, + 316 + ], + "type": "image", + "image_path": "c50034492d3d7059650001866d0235dcd12abce808f5630bc933c180cd02bebe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 327, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 371 + ], + "type": "text", + "content": "Figure 1. Vision-language-models (VLMs) as fine-grained classifiers. VLMs generate text with varying degrees of specificity and similarity to gold-standard label classes. We tackle the problem of aligning these outputs to taxonomic classes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 548, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 548, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 548, + 389 + ], + "type": "text", + "content": "two desiderata when evaluating VLMs on FGVC (Fig. 1):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 395, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 395, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 315, + 395, + 555, + 587 + ], + "type": "text", + "content": "Evaluation with Textual Awareness. When tasked with classifying entities within an image, VLMs often generate text that does not inherently fit within the predefined set of labels used in conventional evaluations. For example, a model might describe an image with the phrase pines in the snow, while the ground-truth label is simply PINE TREE. We suggest that a good evaluation should rely on fuzzy or representation-based matching techniques, which assess the similarity between the generated text and the predefined class labels, rather than whether the VLM generated the predefined class's label exactly. Such approaches have a long history in natural language processing (see, e.g., [42, 58]). We call an evaluation measure for FGVC that is compatible with unconstrained text textually aware." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 594, + 556, + 714 + ], + "type": "text", + "content": "Evaluation with Hierarchical Awareness. Standard evaluation methods for VLM-based entity classification typically treat text generated by a VLM in a binary manner, i.e., it is either completely correct or incorrect [18, 31, 74]. However, many classification tasks involve label sets that are organized hierarchically, such as taxonomic trees or knowledge graphs. When such a structure is available, we suggest that evaluation measures take advantage of the structure, i.e., predictions should be judged based on their semantic" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 286, + 713 + ], + "type": "text", + "content": "*Correspondence: vesteinnsnaebjarnarson@gmail.com." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9109" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 310, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 310, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 310, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 310, + 190 + ], + "type": "image", + "image_path": "f1cc79cace0e64468913364f8204cfa9af9d5b7bd60a46e40621d5616d3f7672.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "type": "text", + "content": "Figure 2. Framework for taxonomy-aware VLM evaluation. We propose a method for granular evaluation of the open-ended image classification capabilities of VLMs. To this end, we map model predictions onto a taxonomy, e.g., sourced from a large knowledge graph. We then use hierarchical precision (hP) and recall (hR) to evaluate how taxonomically accurate and specific a model prediction is. This, in turn, lets us evaluate and compare different VLMs and prompts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 191 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 191 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 191 + ], + "type": "image", + "image_path": "f066256a4675897a4adec1e660daa0590876e082b82e3d77a23023de4f2e7188.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 263, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 263, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 74, + 263, + 296, + 324 + ], + "type": "text", + "content": "proximity to the correct label within the hierarchy. Fortunately, these structured label spaces are common in FGVC and other real-world classification settings. We call a VLM evaluation measure that is sensitive to taxonomic relationships hierarchically aware." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 331, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 296, + 521 + ], + "type": "text", + "content": "In the technical portion of the paper, we propose a novel evaluation measure for assessing VLMs' performance on FGVC tasks that incorporates both textual and hierarchical awareness. We term such an evaluation strategy taxonomy-aware. Fig. 2 illustrates the approach for taxonomy-aware VLM evaluation that we present in this paper. Our measure is based on hierarchical precision and recall [33] after placing strings generated by the VLM on a given taxonomy. To contextualize our taxonomy-aware evaluation scheme within the literature, we first evaluate existing text similarity measures from NLP to assess whether they implicitly capture taxonomic relationships. Our analysis reveals that these measures do not align well with the underlying taxonomic structure, motivating our novel taxonomy-aware evaluation framework that explicitly maps unconstrained text generated from a VLM onto a taxonomy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "In the empirical part of the paper, we explore our taxonomy-aware evaluation scheme on a range of VLMs, varying in size, training strategies, and performance, under our proposed taxonomy-aware evaluation scheme. To do so, we apply diverse prompt templates to control for accuracy, i.e., whether the model generates correct answers, and specificity, i.e., how detailed the response should be and whether to be more general when the model is not confident. In support of our evaluation, we leverage large, carefully constructed taxonomies derived from knowledge graphs extracted from Wikidata [71] and the Linnean Catalogue of Life [30]. These taxonomies span diverse domains, e.g., food, sports, animals, plants, cars, and landmarks, providing a holistic characterization of VLM behavior. Our results demonstrate that our proposed taxonomy-aware evaluation measures effectively quantify varying degrees of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "text", + "content": "specificity that more rigid evaluation measures like accuracy fail to capture. We find that larger commercial models respond more successfully to prompts requesting varying levels of specificity, revealing substantial differences compared to the prompt sensitivity exhibited by smaller, publicly available models. These findings illustrate how taxonomy-aware evaluation can uncover nuanced model behaviors when conventional evaluation measures cannot." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 369, + 392, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 392, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 392, + 383 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 389, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 449 + ], + "type": "text", + "content": "We briefly review prior work on evaluating VLMs and textual similarity measures, with emphasis on approaches to visual categorization and hierarchical evaluation. A more general overview of the use of hierarchical labels for visual categorization is given in §17 in the supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 456, + 503, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 503, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 503, + 468 + ], + "type": "text", + "content": "2.1. Evaluating VLMs for Classification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "Recent work has explored VLM evaluation in open-domain visual recognition and question answering, often as a byproduct of general model development. A dominant approach involves retrieving the most similar target class label to an unconstrained text using maximum inner-product search [15, 31]. The retrieved label can then be directly used for classification or further processed to extract candidate labels. In some cases [26], these extracted labels are compared to the original image using image-to-text similarity methods, such as CLIP [55]. In-context learning for classification [27] leverages large language models and a few example labels to generate or extract an image label from a longer caption. Other approaches rely on models like GPT-4 [50] to score generated answers based on a given question and reference answer [4, 34, 46, 87]. A recent method [26] expands classification label spaces by prompting models to produce increasingly specific labels using taxonomically structured templates extracted from ImageNet [18]. However, final model predictions are still evaluated using string-matching measures. Some multi-task VLM bench" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9110" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "content": "marks include image classification tasks [9, 78], typically using exact-match evaluation or text-based measures borrowed from image captioning; see §17 in supplementary material. The Open-domain Visual Entity (OVEN) benchmark [31] relies on representation-based methods while other work avoids direct classification tasks altogether [84]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 151, + 253, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 253, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 253, + 164 + ], + "type": "text", + "content": "2.2. Annotation and Prediction Similarity" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 169, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 384 + ], + "type": "text", + "content": "Language generation tasks such as image captioning have traditionally been evaluated using measure based on string similarity, e.g., ROUGE [42], BLEU [53],METEOR [5] and CIDER [69]. They primarily rely on surface form (string) overlap in prediction and ground truth labels, and as such, do not capture semantic similarities. This led to a family of trained measures that rely on a language model that, given a pair of inputs, predicts a similarity score, typically based on the semantic match. Examples of these are BERTScore [82], and SentenceBERT [58]. Causal and bi-directional language models have been shown to encode varied hierarchical semantic knowledge in human-interpretable metric spaces [28, 52, 54]. Natural language inference (NLI), also known as textual entailment [11, 73], can also capture hierarchical specificity judgments, since sub-categories entail their supercategories, i.e., This is a field sparrow entails This is a bird. In computer vision, this idea has been adopted as visual entailment [77]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 601 + ], + "type": "text", + "content": "Contrastive vision and language models [55, 81] can be used to compare image and text pairs using top-" + }, + { + "bbox": [ + 55, + 384, + 296, + 601 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 384, + 296, + 601 + ], + "type": "text", + "content": " retrieval by inner product similarity. This method relies on the intrinsic quality of the representations. While impressive, these have been shown to fail in cases requiring combining complex compositional information that was not seen together at training time [80]. More recently, prompt-based methods for scoring textual similarity have been proposed: For instance, the LLaVA models [46] are evaluated on the COCO[44] image captioning datasets by asking GPT-4 [50] to provide numerical values (1-10) for a range of measures such as relevance and accuracy. Another approach is fine-tuning large language models to act as evaluation judges [87]. This approach has also been claimed to outperform other evaluation methods for machine translation [34]. A more detailed description of the text similarity measures we compare against in this paper is given in the supplement in section " + }, + { + "bbox": [ + 55, + 384, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\S 10" + }, + { + "bbox": [ + 55, + 384, + 296, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 609, + 218, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 218, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 218, + 623 + ], + "type": "text", + "content": "3. Taxonomy-aware Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Taxonomy-aware classification evaluation is a subtle and understudied problem [33, 35, 62]. It is necessary to assign partial credit to answers that share some taxonomic similarity, e.g., to answers that are within the correct sub-tree, even if they do not match the correct node exactly. Hierarchical precision (hP) and hierarchical recall (hR) [33] are two existing measures that suitably assign partial credit. Both" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "take the length of the longest common path from the root node of the taxonomy—equivalent to the size of the ancestor set—and normalize by either the length of the predicted path (hP) or the target path (hR)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": "We now introduce " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{hP}" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{hR}" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " formally. Let " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " be an alphabet of symbols, and " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\Sigma^{*}" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " be the Kleene closure, i.e., the set of all strings whose symbols are drawn from " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": ". We assume the VLM under consideration constitutes a distribution over " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\Sigma^{*}" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " when it is conditioned on an image. We then define a taxonomy " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " as a triplet " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "(V,\\mathcal{L},E)" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " represents a finite set of nodes, " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\colon V\\to \\Sigma^{*}" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " is a function that maps each node to a string, called a label, and " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "E\\subset V\\times V" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " is a set of directed edges between pairs of nodes. Additionally, we require that the directed graph " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "(V,E)" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " forms a directed tree with a unique root " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\rho \\in V" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": ". The set of nodes encountered along the unique path from the taxonomy tree's root node " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " to a given node " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " constitutes that node's ancestor set, which we denote " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\operatorname{anc}(v)" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": ". Formally, for any node " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "v\\in V" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\operatorname{par}(v)" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": "'s unique parent node in the tree; " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\operatorname{par}(\\rho)" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": ", i.e., with the root note " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " as an argument, is undefined. We then recursively define " + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\operatorname{anc}(v)" + }, + { + "bbox": [ + 313, + 121, + 555, + 324 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 333, + 554, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 333, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 326, + 333, + 554, + 347 + ], + "type": "interline_equation", + "content": "\\operatorname {a n c} (v) \\stackrel {\\text {d e f}} {=} \\left\\{\\rho \\right\\} \\quad (1 \\mathrm {a}, \\text {b a s e c a s e}; v = \\rho)", + "image_path": "96229b6bf266fd0e56fb264c04c578b6305ed80ea4033b16dccda2aebf84bc71.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 349, + 554, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 349, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 328, + 349, + 554, + 363 + ], + "type": "interline_equation", + "content": "\\operatorname {a n c} (v) \\stackrel {\\text {d e f}} {=} \\{v \\} \\cup \\operatorname {a n c} (\\operatorname {p a r} (v)). \\quad (1 b, \\text {i n d u c t i v e c a s e}; v \\neq \\rho)", + "image_path": "af9034cb5a2daa548cc2a441b18d770b1b62e1a71226b827e533462ccb20b932.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "text", + "content": "We now discuss how to compare ground-truth and predicted nodes on a taxonomy. Let " + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\{(v_n^{\\mathrm{gt}}, v_n^{\\mathrm{pr}})\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "text", + "content": " be a set of pairs of nodes where " + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "inline_equation", + "content": "v_n^{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "text", + "content": " is the ground-truth node for the " + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "inline_equation", + "content": "n^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "text", + "content": " image and " + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "inline_equation", + "content": "v_n^{\\mathrm{pr}}" + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "text", + "content": " is the node mapped from the string generated from the VLM conditioned on the " + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "inline_equation", + "content": "n^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "text", + "content": " image. (We describe a procedure to create this string-node mapping in §6.) We then define hP and hR over " + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\{(v_n^{\\mathrm{gt}}, v_n^{\\mathrm{pr}})\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 373, + 554, + 457 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 466, + 554, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 466, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 358, + 466, + 554, + 498 + ], + "type": "interline_equation", + "content": "\\mathbf {h} \\mathbf {P} \\stackrel {\\text {d e f}} {=} \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\frac {\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {p r}}\\right) \\cap \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {g t}}\\right) \\right|}{\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {p r}}\\right) \\right|}, \\tag {2a}", + "image_path": "00b477c76c8545951102b535dd77ff8a6dab708af5e1401833905cb8ab40e4df.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 502, + 553, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 502, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 358, + 502, + 553, + 533 + ], + "type": "interline_equation", + "content": "\\mathbf {h} \\mathbf {R} \\stackrel {\\text {d e f}} {=} \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\frac {\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {p r}}\\right) \\cap \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {g t}}\\right) \\right|}{\\left| \\operatorname {a n c} \\left(v _ {n} ^ {\\mathrm {g t}}\\right) \\right|}. \\tag {2b}", + "image_path": "ac31af212e748493f66a96ee210347fb7ac9a34924dbfc30fbed719753059724.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 542, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 554, + 566 + ], + "type": "text", + "content": "The hierarchical F1 (hF) is then defined as the harmonic mean of hP and hR," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 394, + 574, + 553, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 574, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 394, + 574, + 553, + 600 + ], + "type": "interline_equation", + "content": "\\mathbf {h} \\mathbf {F} \\stackrel {\\text {d e f}} {=} \\frac {2 \\cdot \\mathbf {h} \\mathbf {P} \\cdot \\mathbf {h} \\mathbf {R}}{(\\mathbf {h} \\mathbf {P} + \\mathbf {h} \\mathbf {R})}. \\tag {3}", + "image_path": "728d07734374f1774d64aa2865123b35a807cf63c0a5462fdcbf780e0bd669fc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 609, + 554, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 554, + 657 + ], + "type": "text", + "content": "Both hP and hR fall in the interval [0, 1]. The maximum is achieved when the target and predicted paths coincide. Less-than-perfect scores capture the extent to which the path deviates from the gold answer." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "hP Hierarchical precision captures the amount of incorrect information in the prediction, i.e., deviations from the correct path, relative to the number of nodes shared with the correct path." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9111" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 70, + 175, + 144 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 175, + 144 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 175, + 144 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 175, + 144 + ], + "type": "image", + "image_path": "52e6d36176c7d6fc7bc8ffb25896dfbe5d9a2f9e0d38dfa3dd12aef346a52a42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 145, + 118, + 154 + ], + "lines": [ + { + "bbox": [ + 89, + 145, + 118, + 154 + ], + "spans": [ + { + "bbox": [ + 89, + 145, + 118, + 154 + ], + "type": "text", + "content": "TRAIN" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 185, + 154, + 194 + ], + "lines": [ + { + "bbox": [ + 88, + 185, + 154, + 194 + ], + "spans": [ + { + "bbox": [ + 88, + 185, + 154, + 194 + ], + "type": "text", + "content": "hP:1.00 hR:0.75" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 176, + 144, + 253, + 154 + ], + "lines": [ + { + "bbox": [ + 176, + 144, + 253, + 154 + ], + "spans": [ + { + "bbox": [ + 176, + 144, + 253, + 154 + ], + "type": "text", + "content": "TELEPHOTO LENS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 160, + 222, + 179 + ], + "lines": [ + { + "bbox": [ + 175, + 160, + 222, + 179 + ], + "spans": [ + { + "bbox": [ + 175, + 160, + 222, + 179 + ], + "type": "text", + "content": "Prediction: A lens cover" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 86, + 257, + 127 + ], + "blocks": [ + { + "bbox": [ + 88, + 160, + 160, + 179 + ], + "lines": [ + { + "bbox": [ + 88, + 160, + 160, + 179 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 160, + 179 + ], + "type": "text", + "content": "Prediction: A mode of transport" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 86, + 257, + 127 + ], + "lines": [ + { + "bbox": [ + 178, + 86, + 257, + 127 + ], + "spans": [ + { + "bbox": [ + 178, + 86, + 257, + 127 + ], + "type": "image", + "image_path": "3179c14f5f01c22d2cb6cd3f8f0a06dd058f53363d25289a47cb8f959aa3c3d8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 185, + 241, + 194 + ], + "lines": [ + { + "bbox": [ + 175, + 185, + 241, + 194 + ], + "spans": [ + { + "bbox": [ + 175, + 185, + 241, + 194 + ], + "type": "text", + "content": "hP:0.80 hR:0.67" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 261, + 70, + 348, + 144 + ], + "blocks": [ + { + "bbox": [ + 261, + 70, + 348, + 144 + ], + "lines": [ + { + "bbox": [ + 261, + 70, + 348, + 144 + ], + "spans": [ + { + "bbox": [ + 261, + 70, + 348, + 144 + ], + "type": "image", + "image_path": "87203c2258a6970c3db400abcd329e1130a5a2b1cf5e0917d91afc98e9d4924c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 145, + 323, + 154 + ], + "lines": [ + { + "bbox": [ + 263, + 145, + 323, + 154 + ], + "spans": [ + { + "bbox": [ + 263, + 145, + 323, + 154 + ], + "type": "text", + "content": "PITH HELMET" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 262, + 160, + 302, + 179 + ], + "lines": [ + { + "bbox": [ + 262, + 160, + 302, + 179 + ], + "spans": [ + { + "bbox": [ + 262, + 160, + 302, + 179 + ], + "type": "text", + "content": "Prediction: Neckwear" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 262, + 185, + 328, + 194 + ], + "lines": [ + { + "bbox": [ + 262, + 185, + 328, + 194 + ], + "spans": [ + { + "bbox": [ + 262, + 185, + 328, + 194 + ], + "type": "text", + "content": "hP:0.75 hR:0.50" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "type": "text", + "content": "Figure 3. Illustrative examples. Hierarchical precision (hP) and recall (hR) calculations on random label pairs (" + }, + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\S 5" + }, + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "type": "text", + "content": "). While hP penalizes incorrect labels, i.e., labels that are further away from the target's ancestor set, hR penalizes mistakes made higher up on the taxonomy." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 349, + 70, + 435, + 144 + ], + "blocks": [ + { + "bbox": [ + 349, + 70, + 435, + 144 + ], + "lines": [ + { + "bbox": [ + 349, + 70, + 435, + 144 + ], + "spans": [ + { + "bbox": [ + 349, + 70, + 435, + 144 + ], + "type": "image", + "image_path": "9daca7581acc0e5907b52be97efed0edd0f2d74d2ea5fe30f7928cbc811cb51c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 145, + 429, + 154 + ], + "lines": [ + { + "bbox": [ + 350, + 145, + 429, + 154 + ], + "spans": [ + { + "bbox": [ + 350, + 145, + 429, + 154 + ], + "type": "text", + "content": "STIRLING CASTLE" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 160, + 423, + 179 + ], + "lines": [ + { + "bbox": [ + 349, + 160, + 423, + 179 + ], + "spans": [ + { + "bbox": [ + 349, + 160, + 423, + 179 + ], + "type": "text", + "content": "Prediction: Château de Ranrouët" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 348, + 185, + 414, + 194 + ], + "lines": [ + { + "bbox": [ + 348, + 185, + 414, + 194 + ], + "spans": [ + { + "bbox": [ + 348, + 185, + 414, + 194 + ], + "type": "text", + "content": "hP:0.83 hR:0.83" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 435, + 145, + 460, + 154 + ], + "lines": [ + { + "bbox": [ + 435, + 145, + 460, + 154 + ], + "spans": [ + { + "bbox": [ + 435, + 145, + 460, + 154 + ], + "type": "text", + "content": "POOL" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 160, + 476, + 180 + ], + "lines": [ + { + "bbox": [ + 435, + 160, + 476, + 180 + ], + "spans": [ + { + "bbox": [ + 435, + 160, + 476, + 180 + ], + "type": "text", + "content": "Prediction: High jump" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 523, + 144 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 523, + 144 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 523, + 144 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 523, + 144 + ], + "type": "image", + "image_path": "cfa2ff0cc15dcd127bc7dbe7a3537c647d6f7b54d4d328256474468e228850d5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 185, + 501, + 194 + ], + "lines": [ + { + "bbox": [ + 435, + 185, + 501, + 194 + ], + "spans": [ + { + "bbox": [ + 435, + 185, + 501, + 194 + ], + "type": "text", + "content": "hP:0.67 hR:0.67" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 234, + 296, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 329 + ], + "type": "text", + "content": "hR Hierarchical recall measures the amount of correct information in the prediction, i.e., how much the path from the root to the predicted node overlaps with the path from the root to the ground truth node. Hierarchical recall thus penalizes missing coverage of the target path, especially early deviations. This can be seen as measuring the specificity of the prediction, in particular when the hP is 1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": "We give an example of the computation of " + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{hP}" + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{hR}" + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": " in Fig. 3, sampled from the OVEN datasets. In the case of the picture of TRAIN, we can see how only predicting MODE OF TRANSPORT leads to lower " + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{hR}" + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": " since it lacks information, while " + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{hP}" + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": " remains 1 since there is no incorrect information. For the sport POOL and STIRLING CASTLE, we notice how both scores drop, as incorrect nodes are predicted—these responses contain both incorrect information and some partially correct information. In the case of PITH HELMET, we see that the prediction is incorrect yet not very specific, resulting in a lower " + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{hR}" + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{hP}" + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": ". We also note that it is not possible to have a lower " + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{hP}" + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{hR}" + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": " if the ground truth nodes are leaf nodes." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 501, + 231, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 231, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 231, + 514 + ], + "type": "text", + "content": "4. Taxonomically Linked Datasets" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": "To investigate taxonomy-aware evaluation measures for text generated from a VLM, we consider two taxonomies linked to FGVC tasks. First, the Catalogue of Life taxonomy as contained in the iNaturalist21 dataset [30] is a clean, expert-curated taxonomy that maps evolutionary relationships between species and contains exactly 10,000 leaf nodes. The second is a more general taxonomy constructed from the Wikidata [71] knowledge graph, which links entities from FGVC datasets in the Open-domain Visual Entity (OVEN) benchmark [31]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 647, + 187, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 187, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 187, + 659 + ], + "type": "text", + "content": "4.1. Extracting Taxonomies" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "While the iNaturalist21 dataset includes a built-in taxonomy, this is not the case for other FGVC datasets. For those included in the OVEN collection, we utilize Wikidata—a large-scale knowledge graph derived from Wikipedia [71]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 234, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 389 + ], + "type": "text", + "content": "Unlike taxonomies, knowledge graphs are not constrained to tree or even directed acyclic graph structures, which necessitates the extraction of a taxonomy that fits the formal definition given in §3. To construct such a tree from Wikidata, we rely on the subclass of relation. When multiple paths exist from the root to a given node in the taxonomy, we retain the longest path. In the case of ties, we sample one of the tied paths uniformly at random. To ensure a rooted directed tree, we exclude higher-level abstract classes that introduce cycles. Because OVEN already links its datasets (§4.2) to Wikipedia entries, we can directly map them to Wikidata entities and incorporate them into our taxonomy." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 398, + 493, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 493, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 493, + 411 + ], + "type": "text", + "content": "4.2. Entity Linked Datasets of Images" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 417, + 555, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 573 + ], + "type": "text", + "content": "The iNaturalist21 dataset [30] was compiled and curated from the citizen science platform iNaturalist. It contains 2.7 million images spanning 10,000 species. The OVEN dataset [31] aggregates several FGVC datasets, linking them to Wikipedia identifiers and natural language questions to support the evaluation of open-domain visual entity recognition. It includes ImageNet21k-P [61, 63], iNaturalist2017 [29], Cars196 [36], SUN397 [76], Food101 [10], Sports100 [25], Aircraft [47], Oxford Flowers [49], and Google Landmarks v2 [56, 75]. We exclude iNaturalist2017 from our OVEN-based evaluations, as we separately analyze the newer iNaturalist21 dataset, which is aligned with the higher-quality Catalogue of Life taxonomy." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 585, + 489, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 489, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 489, + 598 + ], + "type": "text", + "content": "5. The Woes of Existing Measures" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "In this section, we evaluate whether non-taxonomic text similarity measures already encode sufficient taxonomic information, e.g., through label overlap or representational similarity. To investigate this, we conduct a targeted synthetic experiment comparing the taxonomic similarity measures (hP and hR), with textual and multimodal similarity measures from §2.2 commonly used to assess the quality of language model outputs. These measures are both string-based and representation-based." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9112" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 297, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 297, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 297, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 297, + 220 + ], + "type": "table", + "html": "
MeasureiNat21OVEN
τ-hPτ-hRτ-hPτ-hR
EM0.010.070.010.01
Contained0.020.090.020.03
stringROUGE0.180.410.170.21
METEOR0.180.400.180.22
rep.BERTscore0.010.310.270.18
SentBERT0.020.250.340.31
NLI0.08-0.20-0.170.19
CLIP-t2t0.140.460.250.15
CLIP-i2t0.350.490.350.34
", + "image_path": "a74c0ca1bb56a75b1527d0de4fe30db8f24b7286697d26ca744ba34f2df4b52c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 227, + 296, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 296, + 273 + ], + "type": "text", + "content": "Table 1. Correlation of text similarity measures with hierarchical metrics. Kendall's " + }, + { + "bbox": [ + 55, + 227, + 296, + 273 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 227, + 296, + 273 + ], + "type": "text", + "content": " correlations between taxonomic metrics hP, hR, and the different similarity measures on the synthetic evaluation sets. All correlations have a " + }, + { + "bbox": [ + 55, + 227, + 296, + 273 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 227, + 296, + 273 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 55, + 227, + 296, + 273 + ], + "type": "inline_equation", + "content": "< 0.0001" + }, + { + "bbox": [ + 55, + 227, + 296, + 273 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": "Experimental Setup. We construct two synthetic evaluation sets to determine to what extent text similarity measures reflect taxonomic structure. We consider the taxonomies from iNaturalist21 and the Wikidata (see §4) to generate controlled pairs of labels corresponding to pairs of nodes in each taxonomy. A reference node " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": " and a candidate node " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": " are chosen such that " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": " is always a leaf node and " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": " is at some given distance from " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": ". The distance between two nodes in the taxonomy is the length of the shortest undirected path between the two nodes in the taxonomy. To ensure a balanced dataset, we sample the node pairs such that their pairwise distance is uniformly distributed up to a maximum distance of 7, depending on the depth of the node. To evaluate hP, we sample 100,000 reference-candidate node pairs from the taxonomies. To evaluate hR in a controlled fashion, however, we additionally require that " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "v' \\in \\mathrm{anc}(v)" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": ", i.e., that " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{hP} = 1" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": ". This allows us to isolate the sensitivity of the similarity measures at decreasing levels of specificity." + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": " For each pair, we compute both the taxonomy-aware score and text-based measures between the corresponding labels " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(v)" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(v')" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": ". We then measure the rank correlation between the taxonomy-aware score and various text-based measures using Kendall's " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": ", which evaluates whether the similarity measure reflects taxonomic proximity under controlled conditions. By construction, we use the node label " + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(v')" + }, + { + "bbox": [ + 56, + 293, + 296, + 677 + ], + "type": "text", + "content": " as the prediction to avoid the problem of mapping unconstrained output onto the taxonomy. All of the measures we consider are text-based (see §10 in the supplement for details), except the CLIP image-to-text score, where we sample an image for the reference node from the corresponding dataset and measure CLIP model similarity between the image and the candidate label. We" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 325, + 73, + 539, + 236 + ], + "blocks": [ + { + "bbox": [ + 325, + 73, + 539, + 236 + ], + "lines": [ + { + "bbox": [ + 325, + 73, + 539, + 236 + ], + "spans": [ + { + "bbox": [ + 325, + 73, + 539, + 236 + ], + "type": "image", + "image_path": "fd61692ce31479b2c85a0d2d1b72c9b0912976249f281b87872c18d53a203420.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 249, + 555, + 294 + ], + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 294 + ], + "type": "text", + "content": "Figure 4. Agreement in node placement using different similarity measures. We observe variation in node placement for all measures. Only METEOR and ROUGE, and the CLIP variations when combined with Alg. 1 share predictions frequently." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 300, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 324 + ], + "type": "text", + "content": "choose high-performing models that are not prohibitively large for the model-based comparisons.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 339, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 556, + 591 + ], + "type": "text", + "content": "Results. Table 1 presents correlations between hP and hR on one hand and the text similarity measures from §2.2 and CLIP-i2t on the other, for the iNaturalist21 and Wiki-data labels. We observe that surface-level measures (Exact Match (EM), Contained) exhibit a low correlation with the taxonomy-aware measures hP and hR. ROUGE and METEOR, which are based on " + }, + { + "bbox": [ + 313, + 339, + 556, + 591 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 339, + 556, + 591 + ], + "type": "text", + "content": "-gram overlap, capture taxonomic similarity more effectively. This is presumably due to label overlap, e.g., GRAY SEAL is a type of SEAL. The highest correlation is achieved by CLIP image-to-text (CLIP-i2t) for both datasets, followed by CLIP text-to-text (CLIP-t2t), except for hP on the iNaturalist21 dataset. In that case, CLIP-t2t fails to capture deviations from the correct taxonomic path. This suggests that CLIP's image features may encode more taxonomic information than its text representations, which can struggle due to the low-frequency vocabulary inherent to this task. NLI is an outlier, exhibiting a negative correlation in most cases. While NLI captures taxonomic similarity for closely related labels, it performs poorly on more distantly related pairs. This is visualized in Fig. 8 in the supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 606, + 556, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 667 + ], + "type": "text", + "content": "Summary. Overall, we observe that representation-based measures capture some notions of taxonomic distance along the path from the leaf node to the root (hR), but are, for the most part, poorly equipped to measure taxonomic distances across different subtrees (hP)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "Evaluating hP while controlling for hR is only possible for non-leaf reference nodes. Since our datasets always have leaf reference nodes, we did not explore this setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 673, + 555, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 555, + 715 + ], + "type": "text", + "content": "2All models are downloaded from. huggingface.co, i.e., BERTScore (microsoft/deberta-large-mnli), SentenceBERT (sentence-transformers/all-mpnet-base-v2), NLI (MoritzLaurer/mDeBERTa-v3-base-mnli-xnli) and CLIP (apple/DFN5B-CLIP-ViT-H-14)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9113" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 277, + 215 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 277, + 215 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 277, + 215 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 277, + 215 + ], + "type": "table", + "html": "
MeasurehPhRhFAccuracy
Exact Match0.370.420.3917.5%
Contained0.490.500.4924.3%
BLEU0.710.530.5424.3%
METEOR0.620.650.6329.1%
ROUGE0.620.660.6330.0%
BERTScore0.450.520.4811.3%
SentBERT0.580.670.6220.0%
CLIP-t2t0.720.790.7530.3%
CLIP-i2t0.580.710.235.3%
CLIP-i2t+Alg. 10.800.810.8044.0%
CLIP-t2t+Alg. 10.790.820.8047.1%
", + "image_path": "44bff6acf4e192703565db272605c592c87eb25d4b8fe655717b9372d38d119e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 295, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 278 + ], + "type": "text", + "content": "Table 2. Evaluating taxonomy mapping with human annotation. We report the performance of text similarity measures (directly) and CLIP image-to-text similarity for mapping VLM text predictions to the taxonomy using a subset of iNaturalist21 samples that were manually mapped to the correct nodes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 285, + 286, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 286, + 300 + ], + "type": "text", + "content": "6. Mapping VLM Predictions to Taxonomies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 306, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 474 + ], + "type": "text", + "content": "The key challenge for taxonomy-aware evaluation of unconstrained text generated from VLMs is mapping the text generated by a VLM onto nodes in a given taxonomy. A baseline approach to this mapping is to use text similarity measures (e.g., those in §2.2) to calculate the best match between the generated text and the labeled nodes in a taxonomy. We also develop a heuristic approach that performs this mapping using a general similarity measure as a subroutine, which makes using CLIP text-to-image similarity feasible. Specifically, this algorithm combines string-based matching with scores from a given similarity measure to position predictions onto the taxonomy. We describe this procedure in words below and give pseudocode in §11 (Alg. 1 supplement)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "Our goal is to map each VLM-generated prediction string to a node in the taxonomy using a series of increasingly flexible matching strategies. Generally, we attempt exact matches first, then allow for partial overlaps or ambiguous cases, always preferring more specific nodes when multiple candidates are viable. Before matching, we normalize all predictions and node labels by lowercasing, replacing dashes with spaces, and stripping other punctuation. We then compute similarity scores between the prediction and all node labels using the CLIP similarity measures. More specifically, the mapping proceeds in stages. We first check whether any of the top- " + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "inline_equation", + "content": "(k = 10)" + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": " highest-scoring nodes, according to the CLIP measure applied to their labels, appear verbatim in the prediction string; if so, we return the corresponding node. If this fails, we search for verbatim matches against all other labels in the taxonomy. If no full-string matches are found, we repeat these checks using " + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "-gram overlap for " + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "inline_equation", + "content": "n \\in \\{4,3,2\\}" + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": ". If multiple candidates remain with similar scores, we apply a fallback based on score ambiguity: if the difference between the top two scores, and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "between the top and " + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "k^{\\text{th}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": " score, are both below threshold values, we treat the case as ambiguous. We then search for a common ancestor among the top candidates. If any node appears sufficiently often (at least four times) in the ancestor sets of multiple nodes in the top-" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": " highest-scoring nodes we return the most specific such node, defined as the one deepest in the taxonomy. If none of the prior criteria are met, we fall back to returning the single highest-scoring candidate. Thus, the algorithm always returns a node in the taxonomy. If multiple nodes match the above criteria, we return the most specific one, i.e., the node deepest in the taxonomy. The \"shared ancestor\" step is particularly beneficial for the CLIP image-to-text similarities since we do not have images corresponding to higher-order taxonomic concepts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 249, + 418, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 418, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 418, + 262 + ], + "type": "text", + "content": "6.1. Mapping Quality" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 268, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 399 + ], + "type": "text", + "content": "To evaluate the quality of the taxonomic node proposals we measure their agreement against manually annotated node positions. We compare the similarity measures described in §2.2 (and Supplement §10) directly against all candidate node labels. We then do the same using Alg. 1, combined with the CLIP measures. To construct the test set, we gather 416 unique VLM outputs for the iNaturalist21 dataset, randomly sampled from the generations described in §7. For each output, we manually find the best matching taxonomic node given the string, e.g., matching This is a bird to AVES/BIRDS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 417, + 554, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 573 + ], + "type": "text", + "content": "Results. Results are shown in Tab. 2. We find string-matching measures (Exact Match and Contained) do fairly poorly, while other surface-level measures (BLEU, METEOR, and ROUGE) perform slightly better. The representation-based approaches do not perform better in general, except for CLIP text-to-text. Fig. 4 shows that the string-level measures have considerable diversity in their predictions, while the representation-based measures show more diversity. The best performance is found with CLIP text-to-text and CLIP image-to-text using the algorithm described above (Alg. 1), which results in the same hF score but slightly improved exact node match accuracy score for CLIP text-to-text + Alg. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 585, + 504, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 504, + 599 + ], + "type": "text", + "content": "7. Evaluating and Comparing VLMs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "content": "We compare one closed model and seven publicly available VLMs on fine-grained visual classification using our taxonomy-aware evaluation scheme. To map (unconstrained) strings generated from a VLM to nodes in the taxonomy, we use the best-performing similarity method (CLIP-t2t + Alg. 1), enabling the computation of hP, hR, and hF. These scores offer insight into the level of accuracy (hP) and specificity (hR) of VLM predictions for a given taxonomy. The resulting rankings differ significantly from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9114" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 294, + 179 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 179 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 179 + ], + "type": "table", + "html": "
Model Name# ParametersTraining PipelineOpenVLM Avg.
GPT-4 [50]UnknownUnknown63.1
ILmXC2 [23]7BPretr., instr.-tuning62.2
OLM12B [79]12BPretr., instr.-tuning, RLHF54.6
QVLChat [3]9.6BPretr., instr.-tuning51.3
OLM3B [32]3BPretr., instr.-tuning47.8
LLaVA [45]7.2BPretr., instr.-tuning45.9
QwenVL [3]9.6BPretr.21.7
Fuyu [6]8BPretr.N/A
", + "image_path": "17295b3c219f7ae91a6947600b3ee958be156fe8fc4f47470097c65d7642f149.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 187, + 295, + 221 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 295, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 295, + 221 + ], + "type": "text", + "content": "Table 3. Overview of evaluated VLMs. Details on the models we evaluate, including number of parameters, training method, and average performance on a large vision-language benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 242, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 278 + ], + "type": "text", + "content": "those based on standard text similarity measures, highlighting the taxonomy-aware measure's ability to capture previously unmeasured aspects of model performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 293, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 568 + ], + "type": "text", + "content": "Experimental Setup. We compare eight VLMs that differ in model size, training configuration (pretrained-only, instruction-tuned, and preference-aligned [51]), and performance on the Open VLM Leaderboard [16], a collection of 13 VL benchmarks with 38 models. We summarize model sizes, training configuration, and average performance on the Open VLM leaderboard in Table 3. For both the OVEN-sourced FGVC datasets and iNaturalist21, we randomly sample 6000 images from the validation set. For iNaturalist21, we use the prompt What species is this?, while for OVEN the question is included in the dataset and varies by object type. To understand how well we can control hR and hP for instruction-tuned models, we construct two different sets of instructions: (1) one containing a minimal instruction for answering in a specified format, and (2) one containing an instruction to be specific. We use slight variations of the prompt formatting to fit each model (as specified by the model developers); the exact prompts are provided in the supplement (Tab. 7 and Tab. 8). Because non-instruction-tuned models do not respond meaningfully to instructions, we prompt them with a straightforward question-answer template " + }, + { + "bbox": [ + 56, + 293, + 296, + 568 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 56, + 293, + 296, + 568 + ], + "type": "text", + "content": ": {}A:. We use an orange-colored {} to indicate a variable in the template." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 575, + 180, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 180, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 180, + 588 + ], + "type": "text", + "content": "7.1. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 617 + ], + "type": "text", + "content": "How does the ranking of VLM performance change when considering taxonomy-aware measures? In" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Fig. 5, we rank eight VLMs based on their performance on FGVC according to the different similarity measures corresponding to the values given in Tab. 4 and Tab. 5 (" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\S 14" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " in supplement). We find that several text similarity measures yield similar rankings when used as direct evaluation criteria. In contrast, the taxonomy-aware measures lead to notable changes in model rankings. For instance, LLaVA ranks lowest under Exact Match for" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 76, + 548, + 155 + ], + "blocks": [ + { + "bbox": [ + 321, + 76, + 548, + 155 + ], + "lines": [ + { + "bbox": [ + 321, + 76, + 548, + 155 + ], + "spans": [ + { + "bbox": [ + 321, + 76, + 548, + 155 + ], + "type": "image", + "image_path": "03ccd528a3a877782dcf4dc0c722d3d7cbbe9e755d2eaa5b0a769b31ce593594.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 165, + 548, + 244 + ], + "blocks": [ + { + "bbox": [ + 321, + 165, + 548, + 244 + ], + "lines": [ + { + "bbox": [ + 321, + 165, + 548, + 244 + ], + "spans": [ + { + "bbox": [ + 321, + 165, + 548, + 244 + ], + "type": "image", + "image_path": "c137a57b5900c84316c3f9ff7471a49277b0d973d16ec6059948d53561d23e01.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 258, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 335 + ], + "type": "text", + "content": "Figure 5. Ranking of VLMs for the iNaturalist21 (top) and OVEN (bottom) datasets. We evaluate the ranking of VLMs (vertical axis) based on different evaluation measures (horizontal axis). The best model is shown in the top row. On the left, we see the model names ranked by exact match and on the right ranked by the hierarchical F1 hF. See §14, Tab. 4 in supplement for exact numbers." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "text", + "content": "iNaturalist21, yet achieves high hierarchical precision (" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{hP}" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "text", + "content": ") and, consequently, a strong hierarchical F1 (" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{hF}" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "text", + "content": "). This indicates that LLaVA deviates from the correct taxonomy path less frequently than other models. However, its low hierarchical recall (" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{hR}" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "text", + "content": ") suggests that its predictions are often not specific. Another example is GPT-4, which ranks highest on all measures except " + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{hP}" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{hF}" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "text", + "content": " for iNaturalist21. Compared to QVLChat, which is ranked first in both " + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{hP}" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{hF}" + }, + { + "bbox": [ + 313, + 358, + 555, + 515 + ], + "type": "text", + "content": ", we observe that GPT-4 makes more specific but erroneous predictions. Our taxonomy-aware measure offers a complementary perspective to traditional metrics by revealing performance differences that standard similarity-based measures overlook." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Do taxonomy-aware measures capture prompt sensitivity? We prompt the VLMs using two distinct prompts, as described earlier, to examine how their predictions shift in terms of hierarchical precision (hP) and recall (hR), shown jointly in Fig. 7. Fuyu and QwenVL are not instruction-tuned and therefore do not respond differently, while LLaVA shows only minimal changes. In contrast, the other models clearly adjust their predictions based on the different instructions in the prompts. On the iNaturalist21 dataset, all models except GPT-4 exhibit a trade-off between hP and hR: higher precision tends to come at the cost of recall. GPT-4, however, offers both more accurate and more fine-grained classifications when prompted appropriately. On the OVEN dataset, results are more mixed: roughly half of the models improve on both axes with prompt changes." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9115" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 72, + 277, + 170 + ], + "blocks": [ + { + "bbox": [ + 74, + 72, + 277, + 170 + ], + "lines": [ + { + "bbox": [ + 74, + 72, + 277, + 170 + ], + "spans": [ + { + "bbox": [ + 74, + 72, + 277, + 170 + ], + "type": "image", + "image_path": "c35983c1850958d662d8c80a44da3944dc31a7a67a0cb64e2fa9bb8889123cae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 184, + 297, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 297, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 297, + 231 + ], + "type": "text", + "content": "Figure 6. Prompt tuning results for bird classifier. For High hP (●), we target prompts that prioritize hP. For Acc (▲) we optimize for prompts that give higher binary accuracy. This demonstrates how hP and hR can help tune a VLM application." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "content": "These findings suggest that each application may prioritize different target values for " + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{hP}" + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{hR}" + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "content": ", and that models can adapt substantially to meet those needs. In such cases, prompt design becomes a powerful paradigm for tuning model behavior as further explored in the next section." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 317, + 293, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 293, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 293, + 330 + ], + "type": "text", + "content": "7.2. An Application: High-Precision Classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 335, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 296, + 612 + ], + "type": "text", + "content": "To demonstrate the usefulness of hR and hP in tuning prompts for a specific application, we prompt Llama 3.2 Vision (11B) [65] to classify birds. The application goal is to minimize the amount of wrong information in the prediction, which is equivalent to maximizing hierarchical precision but hard to formulate in terms of exact match binary classification accuracy. We use prompt-tuning with hP and hR as feedback signals, and compare to the same procedure but with accuracy as the feedback signal. We start with a basic prompt What is this an image of? Answer in the format 'A: .' and task ChatGPT [50] to improve on the prompt for 30 iterations. A system prompt describes the scores and the goal of developing a conservative classifier that gives a less specific prediction if uncertain (see §15 in the supplement for more details). In each iteration, we include past prompts along with their resulting hR and hP scores, or the accuracy score, in a few-shot learning setup. We tune the prompts with 20 images and evaluate on 200 images of birds from iNaturalist21. The results are shown in Fig. 6. While this is a toy example, we see how the taxonomy-aware measures not only enable a higher granularity of model performance analysis, but can also facilitate targeted application development." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 621, + 211, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 211, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 211, + 634 + ], + "type": "text", + "content": "8. Discussion and Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Evaluating unconstrained text generated from a VLM is inherently challenging: there are countless entities to describe and even more ways to describe them. As the computer vision community shifts from closed-set recognition to open-ended recognition, evaluation increasingly resembles an NLP task. In this paper, we introduce structure into" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 325, + 70, + 549, + 171 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 549, + 171 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 549, + 171 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 549, + 171 + ], + "type": "image", + "image_path": "23719d906c712a5377c992d8275f07344af203736b7742367de560782056ad2a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 183, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 217 + ], + "type": "text", + "content": "Figure 7. Prompt sensitivity captured by hR and hP. We compare two prompt versions, a minimal barebones version and one that asks for a taxonomically specific prediction." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 244, + 556, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 556, + 472 + ], + "type": "text", + "content": "VLM evaluation by leveraging curated taxonomies and established hierarchical evaluation measures. Despite this, several limitations remain. First, placing unconstrained text on a taxonomy is a low-resource problem; there are no large-scale datasets available to evaluate our mapping approach, nor sufficient data to train a dedicated classifier that links VLM outputs to taxonomy nodes. To address this, we repurpose existing text similarity measures, along with CLIP, to perform the mapping. While our evaluation shows that this approach is non-trivial and introduces errors that affect downstream VLM evaluation, we are optimistic about future improvements, such as training on synthetic data generated by language models. Second, taxonomies, especially those derived from Wikidata, are inherently noisy. Even expert-curated taxonomies like the one used for iNaturalist21 can have uneven subtree granularity, complicating interpretations based on globally averaged measures. One potential solution would be to employ expert input to assign weighted edges to reflect semantic distances between nodes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 491, + 388, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 388, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 388, + 504 + ], + "type": "text", + "content": "9. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 514, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 624 + ], + "type": "text", + "content": "Text similarity measures from NLP are a natural place to look for evaluation of unconstrained VLM output. Additionally, many FGVC tasks come with side information about the class structure, namely taxonomies, which text similarity measures do not fully capture, as we have shown in this paper. To solve this gap, we have proposed a framework for using hierarchical evaluation measures to evaluate VLMs for FGVC, enabling a more fine-grained view of the specificity and accuracy of these models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 643, + 417, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 417, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 417, + 656 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Vésteinn Snæbjarnarson, Serge Belongie, Nico Lang, and Stella Frank are supported by the Pioneer Centre for AI, DNRF grant number P1. Niklas Stoehr is supported by the Swiss Data Science Center (SDSC) PhD Fellowship." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9116" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 209 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: A visual language model for few-shot learning. In Neural Information Processing Systems, pages 23716-23736, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 214, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 294, + 257 + ], + "type": "text", + "content": "[2] Majed A Alkhamees, Mohammed A Alnuem, Saleh M Al-Saleem, and Abdulrakeeb M Al-Ssulami. A semantic metric for concepts similarity in knowledge graphs. Journal of Information Science, 49:778-791, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 259, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 259, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 294, + 313 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 294, + 360 + ], + "type": "text", + "content": "[4] Shuai Bai, Shusheng Yang, Jinze Bai, Peng Wang, Xingxuan Zhang, Junyang Lin, Xinggang Wang, Chang Zhou, and Jingren Zhou. TouchStone: Evaluating vision-language models by language models. arXiv, 2308.16890, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 362, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 416 + ], + "type": "text", + "content": "[5] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 3, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 418, + 294, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 294, + 461 + ], + "type": "text", + "content": "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for AI agents, 2023. 1, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 464, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 508 + ], + "type": "text", + "content": "[7] Walter Bennette, Nathaniel Hofmann, Nathaniel Wilson, and Tyler Witter. Hierarchical open-set recognition for automatic target recognition. In 2021 IEEE Symposium Series on Computational Intelligence (SSCI). IEEE, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 510, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 294, + 564 + ], + "type": "text", + "content": "[8] Luca Bertinetto, Romain Mueller, Konstantinos Tertikas, Sina Samangooei, and Nicholas A Lord. Making better mistakes: Leveraging class hierarchies with deep networks. In Conference on Computer Vision and Pattern Recognition, pages 12506-12515, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 621 + ], + "type": "text", + "content": "[9] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. VisIT-Bench: A benchmark for vision-language instruction following inspired by real-world use. In Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "type": "text", + "content": "[10] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – Mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[11] Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In Conference on Empirical Methods in Natural Language Processing, 2015. 3, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[12] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechu Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. MiniGPT-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 553, + 184 + ], + "type": "text", + "content": "[13] Elijah Cole, Kimberly Wilber, Grant Van Horn, Xuan Yang, Marco Fornoni, Pietro Perona, Serge Belongie, Andrew Howard, and Oisin Mac Aodha. On label granularity and object localization. In European Conference on Computer Vision, pages 604-620, 2022. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[14] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 242, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 242, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 317, + 242, + 553, + 284 + ], + "type": "text", + "content": "[15] Alessandro Conti, Enrico Fini, Massimiliano Mancini, Paolo Rota, Yiming Wang, and Elisa Ricci. Vocabulary-free image classification. In Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 287, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 308 + ], + "type": "text", + "content": "[16] OpenCompass Contributors. OpenCompass: A universal evaluation platform for foundation models, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 311, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 311, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 311, + 553, + 365 + ], + "type": "text", + "content": "[17] Riccardo De Lutio, Yihang She, Stefano D'Aronco, Stefania Russo, Philipp Brun, Jan D Wegner, and Konrad Schindler. Digital taxonomist: Identifying plant species in community scientists' photographs. ISPRS journal of photogrammetry and remote sensing, 182:112-121, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 366, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 553, + 410 + ], + "type": "text", + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 412, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 412, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 412, + 553, + 465 + ], + "type": "text", + "content": "[19] Jia Deng, Jonathan Krause, Alexander C. Berg, and Li Fei-Fei. Hedging your bets: Optimizing accuracy-specificity trade-offs in large scale visual recognition. In Conference on Computer Vision and Pattern Recognition, pages 3450–3457, 2012. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 468, + 553, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 512 + ], + "type": "text", + "content": "[20] Xiwen Dengxiong and Yu Kong. Ancestor search: Generalized open set recognition via hyperbolic side information learning. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4003-4012, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 514, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 553, + 567 + ], + "type": "text", + "content": "[21] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics, pages 4171–4186, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 624 + ], + "type": "text", + "content": "[22] Ankit Dhall, Anastasia Makarova, Octavian Ganea, Dario Pavllo, Michael Greeff, and Andreas Krause. Hierarchical image classification using entailment cone embeddings. In Conference on Computer Vision and Pattern Recognition, pages 836-837, 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 625, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 713 + ], + "type": "text", + "content": "[23] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2: Mastering freeform text-image composition and comprehension in vision-language large model. arXiv, 2401.16420, 2024. 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9117" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 128 + ], + "type": "text", + "content": "[24] S. Doveh, A. Arbelle, S. Harary, E. Schwartz, R. Herzig, R. Giryes, R. Feris, R. Panda, S. Ullman, and L. Karlinsky. Teaching structured Vision & Language concepts to Vision & Language models. In Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 282, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 282, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 282, + 140 + ], + "type": "text", + "content": "[25] Kaggle (Gerry). 100 sports image classification, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "text", + "content": "[26] Simon Ging, Maria Alejandra Bravo, and Thomas Brox. Open-ended VQA benchmarking of vision-language models by exploiting classification datasets and their semantic hierarchy. In International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[27] Sherzod Hakimov and David Schlangen. Images in language space: Exploring the suitability of large language models for vision & language tasks. In Findings of the Association for Computational Linguistics, pages 14196-14210, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "text", + "content": "[28] Yuan He, Zhangdie Yuan, Jiaoyan Chen, and Ian Horrocks. Language models as hierarchy encoders. In Neural Information Processing Systems, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 331 + ], + "type": "text", + "content": "[29] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alexx Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The iNaturalist species classification and detection dataset. In Conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 388 + ], + "type": "text", + "content": "[30] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin MacAodha. Benchmarking representation learning for natural world image collections. In Conference on Computer Vision and Pattern Recognition, pages 12879-12888, 2021. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 294, + 455 + ], + "type": "text", + "content": "[31] Hexiang Hu, Yi Luan, Yang Chen, Urvashi Khandelwal, Mandar Joshi, Kenton Lee, Kristina Toutanova, and Ming-Wei Chang. Open-domain visual entity recognition: Towards recognizing millions of wikipedia entities. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12065–12075, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 544 + ], + "type": "text", + "content": "[32] Shengding Hu, Yuge Tu, Xu Han, Ganqu Cui, Chaoqun He, Weilin Zhao, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Xinrong Zhang, Zhen Leng Thai, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, dahai li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the potential of small language models with scalable training strategies. In Conference on Language Modeling, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[33] Svetlana Kiritchenko, Stan Matwin, Richard Nock, and Fazel Famili. Learning and evaluation in the presence of class hierarchies: Application to text categorization. In Canadian AI, 2006. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[34] Tom Kocmi and Christian Federmann. Large language models are state-of-the-art evaluators of translation quality. In Conference of the European Association for Machine Translation, pages 193–203, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[35] Aris Kosmopoulos, Ioannis Partalas, Eric Gaussier, Georgios Paliouras, and Ion Androutsopoulos. Evaluation measures for hierarchical classification: A unified view and novel approaches. Data Mining and Knowledge Discovery, 29(3): 820-865, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D object representations for fine-grained categorization. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision Workshops, pages 554-561, 2013. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[37] Suren Kumar and Rui Zheng. Hierarchical category detector for clothing recognition from visual data. In International Conference on Computer Vision Workshops, pages 2306-2312, 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[38] Nico Lang, Vésteinn Snæbjarnarson, Elijah Cole, Oisin Mac Aodha, Christian Igel, and Serge Belongie. From coarse to fine-grained open-set recognition. In Conference on Computer Vision and Pattern Recognition, pages 17804-17814, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[39] Kibok Lee, Kimin Lee, Kyle Min, Yuting Zhang, Jinwoo Shin, and Honglak Lee. Hierarchical novelty detection for visual object recognition. In Conference on Computer Vision and Pattern Recognition, pages 1034-1042, 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "text", + "content": "[40] Clément Lefebvre and Niklas Stoehr. Rethinking the event coding pipeline with prompt entailment. In EACL Workshop Fact Extraction and VERIFICATION, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[41] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[42] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81. Association for Computational Linguistics, 2004. 1, 3, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "text", + "content": "[43] Dekang Lin. An information-theoretic definition of similarity. In International Conference on Machine Learning, page 296-304, 1998. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[44] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, pages 740-755, 2014. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 455, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 497 + ], + "type": "text", + "content": "[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Conference on Computer Vision and Pattern Recognition, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 500, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 532 + ], + "type": "text", + "content": "[46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Neural Information Processing Systems, 2023. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "text", + "content": "[47] Subhransu Maji, Esa Rahtu, Juho Kannala, Mathew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv, 1306.5151, 2013. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 610 + ], + "type": "text", + "content": "[48] Kenneth Marino, Ruslan Salakhutdinov, and Abhinav Gupta. The more you know: Using knowledge graphs for image classification. In Computer Vision and Pattern Recognition, 2016. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[49] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 522, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 522, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 522, + 668 + ], + "type": "text", + "content": "[50] OpenAI. GPT-4 technical report, 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[51] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9118" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In Neural Information Processing Systems, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 205 + ], + "type": "text", + "content": "[52] Jeff Z. Pan, Simon Razniewski, Jan-Christoph Kalo, Sneha Singhania, Jiaoyan Chen, Stefan Dietze, Hajira Jabeen, Janna Omeliyanenko, Wen Zhang, Matteo Lissandrini, Russia Biswas, Gerard de Melo, Angela Bonifati, Edlira Vakaj, Mauro Dragoni, and Damien Graux. Large language models and knowledge graphs: Opportunities and challenges. Transactions on Graph Data and Knowledge, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 206, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 251 + ], + "type": "text", + "content": "[53] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. In Annual Meeting on Association for Computational Linguistics, page 311-318, 2002. 3, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 306 + ], + "type": "text", + "content": "[54] Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Conference on Empirical Methods in Natural Language Processing, pages 2463-2473, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 384 + ], + "type": "text", + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 386, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 295, + 430 + ], + "type": "text", + "content": "[56] Elias Ramzi, Nicolas Audebert, Clement Rambour, Andre Araujo, Xavier Bitot, and Nicolas Thome. Optimization of rank losses for image retrieval. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 431, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 295, + 475 + ], + "type": "text", + "content": "[57] Sarah Rastegar, Mohammadreza Salehi, Yuki M Asano, Hazel Doughty, and Cees GM Snoek. SelEx: Self-expertise in fine-grained generalized category discovery. In European Conference on Computer Vision, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 476, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 295, + 531 + ], + "type": "text", + "content": "[58] Nils Reimers and Iryna Gurevych. Sentence-BERT: sentence embeddings using siamese BERT-networks. In Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3980-3990, 2019. 1, 3, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 533, + 295, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 295, + 565 + ], + "type": "text", + "content": "[59] Philip Resnik. A class-based approach to lexical discovery. In Annual Meeting of the Association for Computational Linguistics, pages 327-329, 1992. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 600 + ], + "type": "text", + "content": "[60] Philip Resnik. Using information content to evaluate semantic similarity in a taxonomy. In International Joint Conference on Artificial Intelligence, page 448-453, 1995. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 644 + ], + "type": "text", + "content": "[61] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. In Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 295, + 689 + ], + "type": "text", + "content": "[62] Kevin Riehl, Michael Neunteufel, and Martin Hemberg. Hierarchical confusion matrix for classification performance evaluation. Journal of the Royal Statistical Society Series C: Applied Statistics, 72(5):1394-1412, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "text", + "content": "[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 116 + ], + "type": "text", + "content": "Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[64] Paolo Simeone, Raul Santos-Rodriguez, Matt McVicar, Jeffrey Lijffijt, and Tijl De Bie. Hierarchical novelty detection. In Advances in Intelligent Data Analysis, pages 310-321, 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 182 + ], + "type": "text", + "content": "[65] Hugo Touvron, Aaron Grattafori, Michael Auli, et al. The Llama 3 herd of models. arXiv, 2407.21783, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "text", + "content": "[66] Mehmet Ozgur Turkoglu, Stefano D'Aronco, Gregor Perich, Frank Liebisch, Constantin Streit, Konrad Schindler, and Jan Dirk Wegner. Crop mapping from image time series: Deep learning with multi-scale label hierarchies. Remote Sensing of Environment, 264:112603, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "text", + "content": "[67] Grant Van Horn, Steve Branson, Scott Loarie, Serge Belongie, and Pietro Perona. Lean multiclass crowdsourcing. In Conference on Computer Vision and Pattern Recognition, 2018. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 282, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 326 + ], + "type": "text", + "content": "[68] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In International Conference on Learning Representations, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "text", + "content": "[69] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Conference on Computer Vision and Pattern Recognition, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[70] Nakul Verma, Dhruv Mahajan, Sundararajan Sellamanickam, and Vinod Nair. Learning hierarchical similarity metrics. In Conference on Computer Vision and Pattern Recognition, pages 2280-2287, 2012. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "text", + "content": "[71] Denny Vrandecic and Markus Krötzsch. Wikidata: A free collaborative knowledgebase. Communications of the ACM, 57(10):78-85, 2014. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[72] Cunxiang Wang, Sirui Cheng, Qipeng Guo, Yuanhao Yue, Bowen Ding, Zhikun Xu, Yidong Wang, Xiangkun Hu, Zheng Zhang, and Yue Zhang. Evaluating open-QA evaluation. In Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 535 + ], + "type": "text", + "content": "[73] Sinong Wang, Han Fang, Madian Khabsa, Hanzi Mao, and Hao Ma. Entailment as few-shot learner. arXiv, 2104.14690, 2021. 3, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 536, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 591 + ], + "type": "text", + "content": "[74] Xiu-Shen Wei, Yi-Zhen Song, Oisin Aodha, Jianxin Wu, Yuxin Peng, Jinhui Tang, Jian Yang, and Serge Belongie. Fine-grained image analysis with deep learning: A survey. IEEE Transactions on Pattern Analysis & Machine Intelligence, 44(12):8927-8948, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[75] Tobias Weyand, Andre Araujo, Bingyi Cao, and Jacjk Sim. Google Landmarks Dataset v2: A Large-Scale Benchmark for Instance-Level Recognition and Retrieval. In Conference on Computer Vision and Pattern Recognition, 2020. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 680 + ], + "type": "text", + "content": "[76] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[77] Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. Visual entailment: A novel task for fine-grained image understanding, 2019. 3, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9119" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 563 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[78] Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Fanqing Meng, Siyuan Huang, Meng Lei, Ping Luo, and Yu Qiao. LVLM-eHub: A comprehensive evaluation benchmark for large vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "text", + "content": "[79] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. RLHF-V: Towards trustworthy MLLMs via behavior alignment from fine-grained correctional human feedback. In Conference on Computer Vision and Pattern Recognition, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 249 + ], + "type": "text", + "content": "[80] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 294 + ], + "type": "text", + "content": "[81] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In International Conference on Computer Vision, pages 11941-11952, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 339 + ], + "type": "text", + "content": "[82] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. BERTScore: Evaluating text generation with BERT. In International Conference on Learning Representations, 2020. 3, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "type": "text", + "content": "[83] Bingchen Zhao, Nico Lang, Serge Belongie, and Oisin Mac Aodha. Labeled data selection for category discovery. In European Conference on Computer Vision, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 428 + ], + "type": "text", + "content": "[84] Wangchunshu Zhou, Yan Zeng, Shizhe Diao, and Xinsong Zhang. VLUE: A multi-task multi-dimension benchmark for evaluating vision-language pre-training. In International Conference on Machine Learning, pages 27395-27411, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 472 + ], + "type": "text", + "content": "[85] Ganggao Zhu and Carlos A. Iglesias. Computing semantic similarity of concepts in knowledge graphs. IEEE Transactions on Knowledge and Data Engineering, 29(1):72-85, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 476, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 294, + 517 + ], + "type": "text", + "content": "[86] Ganggao Zhu and Carlos A. Iglesias. Exploiting semantic similarity for named entity disambiguation in knowledge graphs. Expert Systems with Applications, 101:8-24, 2018. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[87] Lianghui Zhu, Xinggang Wang, and Xinlong Wang. JudgeLM: Fine-tuned large language models are scalable judges. In International Conference on Learning Representations, 2025. 2, 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9120" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_content_list.json b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6001ee0d5af7328462cf69f0108165ba318ff227 --- /dev/null +++ b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_content_list.json @@ -0,0 +1,2053 @@ +[ + { + "type": "text", + "text": "Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution", + "text_level": 1, + "bbox": [ + 114, + 130, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyuan You $^{12}$ , Xin Cai $^{2}$ , Jinjin Gu $^{4}$ , Tianfan Xue $^{23\\dagger}$ , Chao Dong $^{134\\dagger}$", + "bbox": [ + 220, + 202, + 767, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences", + "bbox": [ + 192, + 223, + 803, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2The Chinese University of Hong Kong, 3Shanghai AI Laboratory", + "bbox": [ + 236, + 241, + 759, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Shenzhen University of Advanced Technology", + "bbox": [ + 310, + 260, + 686, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhiyuanyou@link.cuhk.edu.hk, caixin@link.cuhk.edu.hk, jinjin.gu@sydney.edu.au", + "bbox": [ + 156, + 279, + 839, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tfxue@ie.cuhk.edu.hk, chao.dong@siat.ac.cn † Corresponding Author", + "bbox": [ + 227, + 296, + 769, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 345, + 326, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid advancement of Multi-modal Large Language Models (MLLMs), MLLM-based Image Quality Assessment (IQA) methods have shown promising performance in linguistic quality description. However, current methods still fall short in accurately scoring image quality. In this work, we aim to leverage MLLMs to regress accurate quality scores. A key challenge is that the quality score is inherently continuous, typically modeled as a Gaussian distribution, whereas MLLMs generate discrete token outputs. This mismatch necessitates score discretization. Previous approaches discretize the mean score into a one-hot label, resulting in information loss and failing to capture inter-image relationships. We propose a distribution-based approach that discretizes the score distribution into a soft label. This method preserves the characteristics of the score distribution, achieving high accuracy and maintaining inter-image relationships. Moreover, to address dataset variation, where different IQA datasets exhibit various distributions, we introduce a fidelity loss based on Thurstone's model. This loss captures intra-dataset relationships, facilitating co-training across multiple IQA datasets. With these designs, we develop the distribution-based Depicted image Quality Assessment model for Score regression (DeQA-Score). Experiments across multiple benchmarks show that DeQA-Score stably outperforms baselines in score regression. Also, DeQA-Score can predict the score distribution that closely aligns with human annotations. Codes and model weights have been released in https://depictqa.github.io/deqa-score/.", + "bbox": [ + 89, + 378, + 483, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 830, + 220, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Quality Assessment (IQA) focuses on perceiving the visual quality of images [26, 41, 66, 84], and is widely adopted in various application scenarios, e.g., image com", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c3d5035d233ce48d76be96cbb7e2a3f6b830219e529a2f8c65952b102abf3e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 344, + 658, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc8d52c5ed98902340f73dffdd37719e2c097a2b605a4a9483a026b1587f3ad4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 345, + 772, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/115dd61a17c167a6b0fa757ebf7934db7d109da1216511689698f7c84fd62d72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 345, + 890, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7c00aeb46368cd6361ea1728f51370187a326ab8d512615c08d2b3733d753bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 425, + 656, + 510 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/68cb15a2ddd2e1f44ca67c4db6f133fe9877e832352fd6ea2957f4780c657fbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 425, + 772, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d80734b4ead7143813d3aa8d0a8a2ec9d4eacd184ffb73d35b65b5fdd5e9ca17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 425, + 890, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/759e462d9e30c7c8dd92e2e2369bbee0dc69ade76285eb83e82af4777da3ed09.jpg", + "image_caption": [ + "Figure 1. Illustration of our distribution-based soft label. Image A is overexposed, while Image B and C are well-exposed with better and similar quality. \"exce.\" means \"excellent\". To train an MLLM as a quality scorer, continuous scores need to be discretized to discrete level tokens as the training label. Q-Align [70] discretizes human-labeled mean score to a one-hot label, causing information loss. We discretize the score distribution, which is approximated as a Gaussian distribution, to obtain a more accurate soft label. Our soft label maintains better inter-image relationships that the Image B is better than A (A < B), and Image B and C share similar quality (B ≈ C), while the one-hot label can not." + ], + "image_footnote": [], + "bbox": [ + 526, + 510, + 656, + 595 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23c670d66a7d30f9dc0cadb426387dd71bda4645f463c8d68283befc68fea9dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 510, + 772, + 595 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3350e31f5688ab3c66084a6c2ca00e59dd95afa04748f6715069c00ad613a883.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 510, + 890, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pression and transmission [5, 52], processing smartphone-captured photos [17], and evaluating generation or restoration models [10, 12, 30, 99], etc. With the emergence of Multi-modal Large Language Models (MLLMs) [32, 44, 75, 95], MLLM-based IQA begins to attract more research interests [7, 69, 71, 79, 80]. These methods mainly leverage MLLMs to describe image quality using language, recognizing that language better mirrors human expression, and captures the multifaceted nature of IQA tasks [80].", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14483", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, for quality scoring, which needs an accurate numerical score to quantify an image's quality, existing description-centered MLLMs [69, 71, 79, 80] perform less accurately than traditional IQA methods. This hinders their applications on real-world IQA scenarios. Considering MLLMs' good generalization ability on various downstream tasks and promising performance on quality description, it is possible for MLLMs to score the image quality. Therefore, in this work, we aim to explore how MLLMs can effectively regress accurate quality scores by leveraging their broad training and inherent adaptability.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We find that the principal challenge in training MLLMs to predict accurate quality scores is the distribution gap between MLLMs' outputs and quality scores. MLLMs generate outputs as discrete tokens, while the quality score is inherently continuous, usually approximated to a Gaussian distribution. Previous approaches, like Q-Align [70], attempt to address the distribution gap by directly discretizing the mean score (i.e., the expectation of estimated score distribution) into a one-hot label, as depicted in Fig. 1. However, such a one-hot discretization method introduces information loss, causing mismatches between the discretized labels and the original continuous scores. For instance, in Fig. 1, Image A is overexposed with a lower mean score than Image B, but their one-hot labels appear identical. Conversely, Image B and C are well-exposed with similar quality scores, yet their discretized labels are different.", + "bbox": [ + 89, + 260, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the distribution gap, we propose a novel soft label to bridge continuous and discrete distributions. As illustrated in Fig. 1, instead of discretizing the mean of score distribution, we discretize the continuous score distribution itself into a discrete distribution of different levels. During training, this discrete distribution is used as the soft training label. This enables effective training with minimal information loss. Also, our distribution-based soft label better preserves inter-image relationship. For example, in Fig. 1, Image B is better than Image A, and Image B and C share similar quality, while the simple one-hot label fails to capture such relationships. At inference, MLLMs output discrete token results, which can be recovered to continuous score distribution. This method ensures that the discrete outputs align well with the underlying continuous distribution, allowing to accurately predict the desired score.", + "bbox": [ + 89, + 505, + 482, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another key challenge is the dataset variation, where different IQA datasets exhibit distinct distributions, making it hard to simply combine them for co-training. Due to this variation, images from different datasets may have nearly the same quality scores (i.e., linearly re-scaled scores) but drastically different perceptual quality [87, 98]. A previous method, UNIQUE [87], has demonstrated that the fidelity loss [63] under the Thurstone's model [60] can effectively capture the intra-dataset relationships, thus facilitating the joint training of multiple IQA datasets. However,", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the fidelity loss requires the model to predict an accurate score distribution, which MLLMs trained with the one-hot label cannot provide. Fortunately, MLLMs trained with our distribution-based soft label can predict such an accurate score distribution, allowing the adaptation of the fidelity loss. We apply this fidelity loss to our MLLM-based quality scorer and demonstrate its effectiveness in this new context.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Equipped with these designs, we successfully develop a distribution-based multi-modal image quality assessment model, termed DeQA-Score. DeQA-Score achieves better score regression results than baseline methods on both in-distribution (e.g., $1.3\\%$ in PLCC on KonIQ [21]) and out-of-distribution datasets (e.g., $4.6\\%$ in PLCC on LIVEWILD [18]). The advantage of our DeQA-Score becomes more pronounced with the co-training across multiple IQA datasets. Moreover, our method can further predict the distribution of quality scores, which is closely aligned with the distribution derived from extensive human annotations. We hope that our DeQA-Score will inspire more research to explore the potential of quality score distribution.", + "bbox": [ + 511, + 198, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 404, + 661, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Quality Assessment (IQA) methods mainly rely on quality scores to assess image quality. They can be categorized into full-reference and non-reference methods through whether there is a high-quality reference image.", + "bbox": [ + 511, + 430, + 903, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Full-reference IQA methods compute a similarity score as the quality score between a distorted image and a high-quality reference. Classical works rely on human-designed metrics such as structural similarity [66], phase congruency with gradient magnitude [82], image information [53], etc. The rapid advancement of deep learning has also inspired learning-based IQA methods that measure image quality through data-driven training. Pioneered by LPIPS [84] and PieAPP [48], driven by large-scale IQA datasets [24, 31, 52, 84], deep-learning methods [4, 6, 8, 15, 16, 19, 76, 94] have exhibited quite high accuracy in quality score regression.", + "bbox": [ + 511, + 491, + 903, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-reference IQA methods directly regress a quality score without a reference image. Initially, hand-crafted natural image statistics are adopted [37, 40-43, 50, 58]. Subsequently, deep-learning-based methods [25, 34, 46, 55, 56, 78, 88, 93, 96] replace hand-crafted statistics by learning quality priors from human-annotated IQA datasets [17, 21, 30, 31]. Recent works further improve the regression accuracy by introducing multi-dataset co-training strategy [87], multi-scale features [26], CLIP pre-training [64], multidimension attention [73], multitask learning [89], and so on.", + "bbox": [ + 511, + 657, + 903, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal Large Language Models (MLLMs) incorporate visual modality into Large Language Models (LLMs) [1, 13, 59, 61] to achieve general visual ability by leveraging the broad training and powerful generalization ability of LLMs. Existing MLLMs [3, 14, 32, 44, 67, 74, 77, 83, 85, 95] have demonstrated a general visual ability", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14484", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and can tackle a variety of multi-modality tasks, like visual question answering [20, 35, 36], document understanding [38, 39, 54], image captioning [2, 9, 81], etc. Given MLLMs' advanced performance in these high-level perception tasks, it is natural to explore how MLLMs perform in low-level perception tasks that are highly related to IQA.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MLLM-based IQA methods leverages MLLMs' foundational knowledge to achieve better IQA performance or more detailed assessment results [72, 92]. Q-Bench [68, 91] shows that general MLLMs have some low-level perception ability. [97] evaluates various MLLMs on the widely-adopted Two-Alternative Forced Choice (2AFC) task. Q-Instruct [69] enhances the low-level perception ability of MLLMs by introducing a large-scale dataset. CoInstruct [71] concentrates on the quality comparison among multiple images. DepictQA [79, 80] performs single-image assessment and paired-image comparison in both full-reference and non-reference settings. Q-Ground [7] focuses on the visual quality grounding task, and performs fine-scale detailed visual quality analysis. Some of these works have proposed MLLM-based quality scoring methods, e.g., binary softmax strategy in [68, 91] and pair-wise voting in [79, 80]. Nonetheless, these works principally focus on the quality description ability using languages.", + "bbox": [ + 89, + 183, + 482, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MLLM-based quality scorer. Initially, Q-Bench [68, 91] proposes a binary softmax strategy, enabling MLLMs to generate quality scores by predicting two discrete quality levels (i.e., good or poor). This strategy is adopted by Q-Instruct [69] and Co-Instruct [71]. Compare2Score [98] achieves quality scoring by teaching MLLMs to compare pair-wise images. Inspired by humans' annotation process, Q-Align [70] discretizes scores to five discrete levels using one-hot label to train MLLMs, resulting in more accurate score regression. Dog-IQA [33] adopts the one-hot label for training-free IQA by incorporating specific standards and local semantic objects. However, as discussed in Sec. 1, one-hot label causes some information loss and can not fully maintain the inter-image relationship.", + "bbox": [ + 89, + 458, + 482, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Soft Label Construction", + "text_level": 1, + "bbox": [ + 89, + 683, + 316, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting One-hot Label", + "text_level": 1, + "bbox": [ + 89, + 707, + 318, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training with one-hot label. MLLMs only accept discrete tokens as inputs and outputs. Therefore, as shown in Fig. 2a, continuous scores1 must be discretized into discrete tokens to train the MLLMs. Q-Align [70] adopts a one-hot discretization method. Specifically, as depicted in Fig. 2b, the score range [1, 5] is uniformly divided into five intervals. These intervals represent five rating levels, {\"bad\", \"poor\", \"fair\", \"good\", \"excellent\"}, which is the rating standard defined by [51]. In Fig. 2a, the Mean Opinion Score (MOS)", + "bbox": [ + 89, + 729, + 482, + 866 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca9f5b35f46c8c79885aa840602e14e4f79c4e744d4f73ecb592170353afc3da.jpg", + "image_caption": [ + "Rate the quality of this image.", + "(a)", + "5 candidate level tokens of :" + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 676, + 165 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4229c4b759b0484b598fc30218bb6045d56e7434fa040201503b2aadcc619013.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 109, + 887, + 127 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/85607b301653b1169f3e87de3fa19289e5db6691da3367bbc06d6b436ce25a8a.jpg", + "image_caption": [ + "[ bad, poor, fair, good, excellent ]" + ], + "image_footnote": [], + "bbox": [ + 527, + 170, + 715, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6282334889ff01d98345e4281bb475fb11e76726f56abadb26e95fe41a8faeef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 170, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6ce449d04ec5b3a3e190fa4076ec1dba39143bb254b5a5bb4d57c0f12fcd496.jpg", + "image_caption": [ + "(b)", + "Continuous score" + ], + "image_footnote": [], + "bbox": [ + 526, + 229, + 715, + 286 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/40446f0e90d988740585592af2819b30f8e65aea7a21c3f2cc2bf85f3227c8f0.jpg", + "image_caption": [ + "Discretized label", + "Figure 2. (a) MLLMs need discrete level tokens for training. Thus, continuous scores must be converted into discrete level tokens as training labels. (b) Comparison between one-hot label [70] and our soft label. One-hot label uniformly divides the score range into 5 intervals to discretize the mean score, i.e., the expectation of estimated score distribution. One-hot label assumes that different levels are identically independent, which is inaccurate. In contrast, we discretize the estimated Gaussian distribution of the score. Our soft label better preserves the relationships between levels, e.g., \"fair\" is closer to \"good\" than \"excellent\"." + ], + "image_footnote": [], + "bbox": [ + 723, + 229, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the image, 4.30, fails into the fifth intervals, resulting in the discretized label \"excellent\". Accordingly, the response is formulated as \"The quality of this image is excellent\". Using the response as the training labels, following [13, 61], MLLMs can be trained with next token prediction loss, i.e., cross-entropy loss between predicted logits and labels.", + "bbox": [ + 511, + 455, + 905, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Limitations of one-hot label. First, ideal training with cross-entropy loss using one-hot labels assumes that the five levels are independent (i.e., orthogonal in a latent space). However, intuitively, these levels are not truly independent, e.g., \"fair\" is closer to \"good\" than \"excellent\". Second, one-hot discretization only considers the mean score, which is the expectation of the score distribution, ignoring the distribution's uncertainty (i.e., variance). Third, near the border of two levels, images with similar quality can be divided into different levels. For example, in Fig. 1, Image B (MOS 3.38) and Image C (MOS 3.49) share similar mean scores, but are differently classified as \"fair\" and \"good\", due to their proximity to the border of these two levels, i.e., 3.4 in Fig. 2b. Finally, one-hot label causes information loss. In Q-Align, the five levels are respectively assigned scores $\\{1, 2, 3, 4, 5\\}$ , with which we calculate the discretization precision in Tab. 1. The discretization error is around $10 \\times$ to $35 \\times$ larger than our soft label that we will soon introduce, thus the one-hot label keeps lower correlation with MOS.", + "bbox": [ + 511, + 549, + 906, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Soft Label", + "text_level": 1, + "bbox": [ + 511, + 847, + 629, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discretization using score distribution. As demonstrated in [87], the quality score of one image, $x$ , is typically mod-", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 Unless otherwise specified, the mean quality scores in all datasets will be normalized to [1, 5], and the variances will be normalized accordingly.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14485", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/77b37992814759cf5537e64fac46787714d569c8b9ac43ff268ec4204d7416f8.jpg", + "table_caption": [ + "Table 1. Comparison of discretization precision between Q-Align's one-hot label [70] and our soft label. The precision of MOS discretization is measured by L1 Error / RMSE and PLCC / SRCC. The distance of original distribution and recovered distribution is measured by JS divergence / Wasserstein distance." + ], + "table_footnote": [], + "table_body": "
MetricsDiscretizationKonIQ [21]SPAQ [17]KADID [31]
L1 / RSMEOne-hot label0.302 / 0.3740.299 / 0.3660.327 / 0.389
Soft label0.008 / 0.0180.011 / 0.0250.024 / 0.038
PLCC / SRCCOne-hot label0.961 / 0.9520.969 / 0.9680.979 / 0.982
Soft label1.000 / 1.0001.000 / 1.0001.000 / 1.000
JS / W-Dist.One-hot label---
Soft label0.001 / 0.0380.022 / 0.0840.010 / 0.140
", + "bbox": [ + 91, + 161, + 483, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "eled as a Gaussian distribution, $x \\sim \\mathcal{N}(\\mu, \\sigma^2)$ , where $\\mu$ and $\\sigma^2$ are the Mean Opinion Score (MOS) and variance of multiple human annotations, respectively. The distribution is represented by its probability density function, $f(x)$ . As shown in the second row of Fig. 2b, we select five central points, $c_i \\in \\{1, 2, 3, 4, 5\\}$ , $i \\in \\{0, 1, 2, 3, 4\\}$ , as the centers of the five discrete levels. The width of each level region, $d$ , is set as 1. The possibility of the score falling into the $i^{\\mathrm{th}}$ level is denoted as $p_i^{raw}$ and is calculated as:", + "bbox": [ + 89, + 279, + 483, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} ^ {r a w} = \\int_ {c _ {i} - d / 2} ^ {c _ {i} + d / 2} f (x) \\mathrm {d x}, \\quad i \\in \\{0, 1, 2, 3, 4 \\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 422, + 482, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assign the five levels with the same textual tokens as Q-Align, i.e., {\"bad\", \"poor\", \"fair\", \"good\", \"excellent)}. Correspondingly, the score of $i^{\\text{th}}$ level is $c_i$ . This defines a rough discrete distribution as $P^{raw}(c_i) = p_i^{raw}$ .", + "bbox": [ + 89, + 465, + 483, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Post-adjustment. The discrete distribution is rough because of truncation errors. As shown in the second row of Fig. 2b, there are two truncation regions (i.e., $x < c_0 - d/2 = 0.5$ and $x > c_4 + d/2 = 5.5$ ), which are not considered in Eq. (1). This can cause two problems. First, the sum of $\\{p_i^{raw}\\}$ does not equal 1, meaning $p_i^{raw}$ is not strictly a discrete distribution. Second, the expectation of this discrete distribution does not equal the expectation of score distribution, $\\mu$ , leading to discretization errors.", + "bbox": [ + 89, + 527, + 483, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To correct these truncation errors, we propose a postadjustment method. Specifically, we apply a linear transformation to adjust $p_i^{raw}$ to more accurate $p_i$ as:", + "bbox": [ + 89, + 664, + 483, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\alpha p _ {i} ^ {r a w} + \\beta\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 715, + 328, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\quad \\sum_ {i} p _ {i} = 1, \\quad \\mu^ {r e c} = \\sum_ {i} p _ {i} c _ {i} = \\mu , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 736, + 482, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_{rec}$ means recovered expectation from discrete distribution. The parameters of the linear transformation, $\\alpha$ and $\\beta$ , can be determined by solving these two constraint conditions. One example of post-adjustment is given in Fig. 3. Moreover, the statistics of these two parameters in Tab. 2 show that the average values of $\\alpha$ and $\\beta$ are quite close to 1 and 0, respectively, implying that the post-adjustment is minimal. This effectively defines our soft label, which follows the discrete distribution as $P(c_i) = p_i$ .", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c9d680b802574353b486b41d6d0bdd4c49ba5f3ed9748fce5bcef0953a96916.jpg", + "image_caption": [ + "Figure 3. Illustration of the soft label adjustment." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 671, + 176 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1b5ed3fbbe8b423313f8b6d1a089ad6b6fcd371f258caa6a1df2ce7c2294a11f.jpg", + "table_caption": [ + "Table 2. Statistics of the label adjustment parameters. Averaged $\\alpha$ and $\\beta$ are very approaching 1 and 0, respectively, meaning that the postadjustment is quite small." + ], + "table_footnote": [], + "table_body": "
KonIQ [21]SPAQ [17]KADID [31]
α1.036±0.0411.033±0.0851.118±0.091
β-0.005±0.005-0.005±0.014-0.015±0.010
", + "bbox": [ + 676, + 161, + 903, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ideally, the discretization error of the MOS will be 0. However, the possibility of some levels may slightly fall below zero after the linear transformation. We address this by simply clipping these values to zero. The impact of this clipping on accuracy is negligible, as shown in Tab. 1. After clipping, our discretization error for MOS remains around $10 \\times$ to $35 \\times$ smaller than Q-Align's one-hot label. Also, our soft label maintains a perfect (i.e., 1.000) Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank-order Correlation Coefficient (SRCC) with the original MOS.", + "bbox": [ + 511, + 217, + 906, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recovery from discrete to continuous score distribution.", + "text_level": 1, + "bbox": [ + 511, + 371, + 903, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in the second condition of Eq. (2), we can compute the expectation of discrete distribution, $P(c_{i})$ , as the recovered expectation of original distribution. Besides the expectation, we can also recover the original variance using the variance of the discrete distribution, $P(c_{i})$ , as:", + "bbox": [ + 511, + 387, + 905, + 463 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu^ {r e c} = \\sum_ {i} p _ {i} c _ {i}, \\quad (\\sigma^ {r e c}) ^ {2} = \\sum_ {i} p _ {i} (c _ {i} - \\mu^ {r e c}) ^ {2}. \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 472, + 906, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we calculate the distance between recovered distribution, $\\mathcal{N}(\\mu^{rec}, (\\sigma^{rec})^2)$ , and original distribution, $\\mathcal{N}(\\mu, \\sigma^2)$ as provided in Tab. 1. Our recovered distribution is closely aligned with human-annotated score distribution. In contrast, one-hot label can not recover such a distribution.", + "bbox": [ + 511, + 503, + 905, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. DeQA-Score Model", + "text_level": 1, + "bbox": [ + 511, + 590, + 702, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Training with Our Soft Label", + "text_level": 1, + "bbox": [ + 511, + 616, + 774, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model architecture. As illustrated in Fig. 4, we follow the widely adopted MLLM architecture [32, 75] to construct our model. First, a vision encoder is adopted to represent the input image to visual tokens. Then, a vision-to-text connector projects the visual tokens into the textual space. Following [75], we use a vision abstractor as the connector. The abstractor can further reduce the number of visual tokens to 64. Finally, the visual and textual tokens are fused together and fed into an LLM for response prediction.", + "bbox": [ + 511, + 638, + 906, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next token prediction loss for common response tokens. As described in Sec. 3.1 and illustrated in Fig. 4, the constructed response follows the template, \"The quality of this image is $<$ level $>$ \". For common response tokens other than the level token, we compute the next token prediction loss, i.e., the cross-entropy loss $\\mathcal{L}_{ce}$ , following [45, 61].", + "bbox": [ + 511, + 777, + 905, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "KL divergence loss for the level token. There are five candidate values for the level token: {\"bad\", \"poor\", \"fair\",", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14486", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/034b24dc80d8eddb0c048dd00428bc8c05d4ad8297bdf2b6cfd8595889e0d82c.jpg", + "image_caption": [ + "Figure 4. Framework of our DeQA-Score trained with soft label. For the “” token, the KL divergence loss is calculated between predicted probabilities and our soft label. For other tokens, common cross-entropy loss for next token prediction is calculated." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 880, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"good\", \"excellent\". Our soft label, $p_i$ , assigns a ground-truth probability to each value. Let the predicted probabilities be $p_i^{pred}$ , the KL divergence loss is calculated through:", + "bbox": [ + 89, + 263, + 482, + 310 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k l} = \\sum_ {i} p _ {i} \\log \\left(p _ {i} ^ {\\text {p r e d}} / p _ {i}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 321, + 482, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The predicted probabilities, $p_i^{pred}$ , are obtained by applying a softmax activation function on model's output logits. Note that during training, this softmax function is applied across all textual tokens, not just across the five predefined levels. This approach encourages the model to reduce the possibilities of textual tokens outside the five levels.", + "bbox": [ + 89, + 353, + 483, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The two terms of losses, $\\mathcal{L}_{ce}$ and $\\mathcal{L}_{kl}$ , are both autoregressive losses, and thus can be easily combined. During training, the cross-entropy loss, $\\mathcal{L}_{ce}$ , will decrease near to 0 fast, since it is quite easy to optimize by simply remembering the response template \"The quality of this image is\".", + "bbox": [ + 89, + 445, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Score Estimation during Inference", + "text_level": 1, + "bbox": [ + 89, + 531, + 392, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, using the model-predicted probabilities of the five levels, $p_i^{pred}$ , we can estimate the expectation and variance of the score distribution the same as Eq. (3):", + "bbox": [ + 89, + 550, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu^ {p r e d} = \\sum_ {i} p _ {i} ^ {p r e d} c _ {i}, (\\sigma^ {p r e d}) ^ {2} = \\sum_ {i} p _ {i} ^ {p r e d} (c _ {i} - \\mu^ {p r e d}) ^ {2}. (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 601, + 482, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here we follow Q-Align [70] to apply a close-set softmax on the five levels to obtain the probabilities, $p_i^{pred}$ , to prevent the influence of other textual tokens. Our statistics in Tab. A2 reveal that the trained MLLM will always predict the five levels, thus applying softmax on either the five levels or all textual tokens yields nearly identical results. As given in Tab. 5, our predicted score distribution, $\\mathcal{N}(\\mu^{pred}, (\\sigma^{pred})^2)$ , aligns well with human annotations.", + "bbox": [ + 89, + 626, + 482, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Joint Training on Multiple IQA Datasets", + "text_level": 1, + "bbox": [ + 89, + 757, + 439, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As demonstrated in [87, 98], different IQA datasets exhibit distinct distributions. Due to this dataset variation, images from different datasets with nearly identical scores (i.e., linearly re-scaled scores) can have drastically different perceptual quality (see examples in Fig. A1). The reason is that different datasets have different reflections between quality scores and visual quality. This disparity makes it infeasible to merely regress absolute quality scores for multi-dataset", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/861682710f7e97a603ccd26e55fbaae04cbfe1c32a455b1e1a02c1c4897b9183.jpg", + "image_caption": [ + "Figure 5. Fidelity loss to maintain intra-dataset relationships. Consider two images, Image A and B, sampled from the same dataset. Using our distribution-based soft label, DeQA-Score predicts the expectation, $\\mu^{pred}$ , and variance, $(\\sigma^{pred})^2$ , of each image's score distribution, with which we can predict the probability that Image A is better than Image B, $P(A > B)$ . This predicted probability is then used to calculate the fidelity loss with the ground-truth probability. Fidelity loss preserves intra-dataset relationships, facilitating the co-training across multiple IQA datasets." + ], + "image_footnote": [], + "bbox": [ + 524, + 262, + 893, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "co-training. To address this challenge, UNIQUE [87] applies fidelity loss [63] based on Thurstone's model [60]. The fidelity loss is based on the principal that IQA emphasizes the ranking relationships within the same dataset, rather than the absolute values of quality scores. Specifically, two images are sampled from the same dataset, whose quality scores can be comparable. These two images are then compared using human-annotated quality scores. The fidelity loss encourages the model to learn such comparison results, which facilitates the model to focus more on the ranking relationships in each dataset, instead of the absolute values of quality scores. Thus, the fidelity loss helps multi-dataset co-training. Our soft label is naturally compatible with the fidelity loss, since our method can predict the score distribution, which is what the fidelity loss needs.", + "bbox": [ + 511, + 549, + 906, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ground-truth construction. As shown in Fig. 5, a pair of images, Image A and B, are sampled from the same dataset. We assume that the quality scores of the two images, $x_A, x_B$ , follow the Gaussian distribution with means and variances collected by subjective testing, i.e., $x_A \\sim \\mathcal{N}(\\mu_A, \\sigma_A^2)$ , $x_B \\sim \\mathcal{N}(\\mu_B, \\sigma_B^2)$ . Assuming that their quality scores are uncorrelated, the score difference, $x_A - x_B$ , also follows a Gaussian distribution, $\\mathcal{N}(\\mu_A - \\mu_B, \\sigma_A^2 + \\sigma_B^2)$ .", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14487", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the possibility that Image A has a better perceptual quality than Image B can be calculated as:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np (A > B) = p \\left(x _ {A} > x _ {B}\\right) = \\Phi \\left(\\frac {\\mu_ {A} - \\mu_ {B}}{\\sqrt {\\sigma_ {A} ^ {2} + \\sigma_ {B} ^ {2}}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 128, + 482, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Phi (\\cdot)$ is the Gaussian cumulative distribution function.", + "bbox": [ + 89, + 179, + 482, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model prediction. As described in Sec. 4.2, our model can predict the expectation and variance of the score distribution. Denoting the predicted expectation and variance of two images as $\\mu_A^{pred}$ , $(\\sigma_A^{pred})^2$ , $\\mu_B^{pred}$ , $(\\sigma_B^{pred})^2$ , same to Eq. (6), we can predict the possibility of Image A surpassing Image B in perceptual quality as:", + "bbox": [ + 89, + 196, + 483, + 287 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np ^ {p r e d} (A > B) = \\Phi \\left(\\frac {\\mu_ {A} ^ {p r e d} - \\mu_ {B} ^ {p r e d}}{\\sqrt {(\\sigma_ {A} ^ {p r e d}) ^ {2} + (\\sigma_ {B} ^ {p r e d}) ^ {2}}}\\right). \\quad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 296, + 482, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training with fidelity loss. Following [87], we choose fidelity loss as the similarity measure to train the model as:", + "bbox": [ + 89, + 356, + 483, + 387 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {f d} = 1 - \\sqrt {p (A > B) p ^ {p r e d} (A > B)} \\\\ - \\sqrt {\\left(1 - p (A > B)\\right) \\left(1 - p ^ {p r e d} (A > B)\\right)}. \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 409, + 480, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adding two auto-regressive losses, the final training loss is:", + "bbox": [ + 89, + 470, + 482, + 484 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {f d} + \\gamma \\left(\\mathcal {L} _ {c e} + \\mathcal {L} _ {k l}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 500, + 480, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma$ is a weighting term to adjust the loss scale.", + "bbox": [ + 89, + 521, + 431, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 549, + 223, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets, Baselines, Metrics, and Details", + "text_level": 1, + "bbox": [ + 89, + 573, + 434, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For a comprehensive evaluation, we include a wide range of IQA datasets: (a) in-the-wild images: KonIQ [21], SPAQ [17], LIVE-Wild [18], and FLIVE [78]; (b) synthetic distortions: KADID [31], TID2013 [47], and CSIQ [29]; (c) model-processed distortions: PIPAL [24]; (d) AI-generated images: AGIQA-3K [30]. Among these datasets, KonIQ, SPAQ, KADID, and PIPAL are split into training and test sets. Other datasets are used to assess the out-of-distribution generalization ability. The Mean Opinion Scores (MOS) of these datasets are normalized to the range of [1,5], with the variances normalized accordingly.", + "bbox": [ + 89, + 595, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline methods. To demonstrate the effectiveness of our distribution-based soft label, we primarily compare with Q-Align [70], which utilizes the one-hot label. The base models for DeQA-Score and Q-Align are kept identical for fair comparison. We also include previous IQA methods as baselines, including handcrafted methods: NIQE [41] and BRISQUE [40], and deep-learning-based methods: NIMA [57], HyperIQA [55], DBCNN [86], MUSIQ [26], CLIP-IQA+ [64], ManIQA [73], and Compare2Score [98].", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following [26, 70, 73], we use the Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank-order Correlation Coefficient (SRCC) as metrics to evaluate the score regression performance. Additionally, we employ KL divergence, JS divergence, and Wasserstein distance to assess the results of score distribution prediction.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We use mPLUG-Owl2 [75] as our base model, with a CLIP-pretrained ViT-L [49] as the image encoder, a six-layer Q-Former as the visual abstractor, and LLaMA-2-7B [62] as the LLM. The number of visual tokens output by the vision abstractor is set to 64, and the hidden dimension of LLM is 4096. The pretrained weights of mPLUG-Owl2 are used for model initialization. The weighting term $\\gamma$ is empirically set to 0.05, adjusting the two loss terms roughly to the same scale. We adopt AdamW [23] as the optimizer. The initial learning rate is set to 2e-5 and decays gradually using the cosine decay strategy. Our model is trained with a batch size of 64 for 3 epochs. Using 8 NVIDIA RTX A6000 GPUs, the training process on KonIQ dataset is completed in around 1.5 hours.", + "bbox": [ + 511, + 184, + 906, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results of Score Regression", + "text_level": 1, + "bbox": [ + 511, + 406, + 761, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Score regression results of single-dataset training. We first train our model on a single dataset, i.e., KonIQ, and then evaluate the model on both in-distribution and out-of-distribution datasets, following [28, 70]. The score regression results are shown in Tab. 3. On the in-distribution KonIQ test set, our model achieves superior score regression results compared to all baseline methods, e.g., a $1.3\\%$ improvement in PLCC over Q-Align. Furthermore, on out-of-distribution datasets, our DeQA-Score maintains its advantage with the best overall results across nearly all benchmarks, e.g., a $4.6\\%$ improvement in PLCC on LIVE-Wild. These findings confirm the effectiveness of our method.", + "bbox": [ + 511, + 429, + 905, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, we observe that all models perform poorly on the PIPAL dataset. For instance, our best results reach only 0.462 on PLCC. The main reason is the distortion gap between the training dataset, KonIQ (in-the-wild distortions), and PIPAL (model-processed distortions). This highlights the necessity of co-training with multiple IQA datasets containing various distortion types. This observation motivates our subsequent experiments on multi-dataset co-training.", + "bbox": [ + 511, + 611, + 905, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Score regression results of multi-dataset co-training. As discussed above, co-training on multiple IQA datasets is necessary to obtain a unified model for various categories of distortions. Therefore, we conduct multi-dataset cotraining experiments in Tab. 4. First, our DeQA-Score consistently outperforms Q-Align across different training combinations. The performance gain attributes to both the soft label and fidelity loss, as confirmed in Tab. 6. Second, comparing #0 to #1, when KADID (synthetic distortions) is added to the training set, the performance on TID2013 and CSIQ (both synthetic distortions) improves", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14488", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/26d388c9aababa3bae1d671d379fd4bfbcc109fb7d2f794c91b807a4ad58f84f.jpg", + "table_caption": [ + "Table 3. Score regression results of single-dataset training with the PLCC / SRCC metrics. All methods except handcrafted ones are trained on the KonIQ dataset. DeQA-Score outperforms all baseline methods across nearly all benchmarks." + ], + "table_footnote": [], + "table_body": "
CategoryMethodsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KCSIQFLIVE
HandcraftedNIQE [41]0.533 / 0.5300.679 / 0.6640.468 / 0.4050.195 / 0.1610.493 / 0.4490.560 / 0.5330.718 / 0.6280.147 / 0.100
BRISQUE [40]0.225 / 0.2260.490 / 0.4060.429 / 0.3560.267 / 0.2320.361 / 0.3130.541 / 0.4970.740 / 0.5560.108 / 0.054
Non-MLLMNIMA [57]0.896 / 0.8590.838 / 0.8560.532 / 0.5350.390 / 0.3990.814 / 0.7710.715 / 0.6540.695 / 0.6490.561 / 0.467
HyperIQA [55]0.917 / 0.9060.791 / 0.7880.506 / 0.4680.410 / 0.4030.772 / 0.7490.702 / 0.6400.752 / 0.7170.485 / 0.383
DBCNN [86]0.884 / 0.8750.812 / 0.8060.497 / 0.4840.384 / 0.3810.773 / 0.7550.730 / 0.6410.586 / 0.5720.485 / 0.385
MUSIQ [26]0.924 / 0.9290.868 / 0.8630.575 / 0.5560.431 / 0.4310.789 / 0.8300.722 / 0.6300.771 / 0.7100.565 / 0.467
CLIP-IQA+ [64]0.909 / 0.8950.866 / 0.8640.653 / 0.6540.427 / 0.4190.832 / 0.8050.736 / 0.6850.772 / 0.7190.427 / 0.316
ManIQA [73]0.849 / 0.8340.768 / 0.7580.499 / 0.4650.457 / 0.4520.849 / 0.8320.723 / 0.6360.623 / 0.6270.512 / 0.401
MLLM-basedCompare2Score [98]0.923 / 0.9100.867 / 0.8600.500 / 0.4530.354 / 0.3420.786 / 0.7720.777 / 0.6710.735 / 0.7050.474 / 0.413
Q-Align [70]0.941 / 0.9400.886 / 0.8870.674 / 0.6840.403 / 0.4190.853 / 0.8600.772 / 0.7350.785 / 0.7370.554 / 0.483
DeQA-Score0.953 / 0.9410.895 / 0.8960.694 / 0.6870.472 / 0.4780.892 / 0.8790.809 / 0.7290.787 / 0.7440.589 / 0.501
", + "bbox": [ + 91, + 121, + 906, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7a2af31349da31ccf2e629c4996caddcbaa45a0dd7aefbd60d4390d7ac45a3f8.jpg", + "table_caption": [ + "Table 4. Score regression results of co-training on multiple IQA datasets with the PLCC / SRCC metrics. Our DeQA-Score consistently surpasses the baseline method, Q-Align, across all benchmarks with different training combinations." + ], + "table_footnote": [], + "table_body": "
#Trained onMethodsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
0KonIQ, SPAQQ-Align [70]0.943 / 0.9400.933 / 0.9310.692 / 0.7080.401 / 0.4110.883 / 0.8790.795 / 0.7270.671 / 0.5680.795 / 0.767
DeQA-Score0.953 / 0.9430.936 / 0.9330.724 / 0.7190.468 / 0.4740.902 / 0.8880.810 / 0.7380.706 / 0.6160.836 / 0.785
1KonIQ, SPAQ, KADIDQ-Align [70]0.945 / 0.9380.933 / 0.9310.935 / 0.9340.409 / 0.4200.887 / 0.8830.788 / 0.7330.829 / 0.8080.876 / 0.845
DeQA-Score0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
2KonIQ, SPAQ, KADID, PIPALQ-Align [70]0.926 / 0.9320.917 / 0.9200.950 / 0.9540.702 / 0.6710.853 / 0.8450.765 / 0.7220.811 / 0.7950.838 / 0.789
DeQA-Score0.958 / 0.9460.932 / 0.9290.963 / 0.9610.724 / 0.6900.877 / 0.8570.770 / 0.7350.828 / 0.7960.863 / 0.807
", + "bbox": [ + 91, + 329, + 903, + 441 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f6d2c44707fda39404a9cd80c837066aa970ac4700cde32b035cf6720a609083.jpg", + "table_caption": [ + "Table 5. Score distribution prediction results with JS divergence / Wasserstein distance between the predicted Gaussian distribution and the human labeled Gaussian distribution as metrics. Models are trained on the KonIQ, SPAQ, and KADID datasets. DeQA-Score achieves a much closer alignment with human annotations." + ], + "table_footnote": [], + "table_body": "
DatasetsQ-Align [70]DeQA-Score (Ours)
KonIQ0.415 / 0.6110.014 / 0.186
SPAQ0.192 / 0.4550.070 / 0.332
KADID0.519 / 0.7230.029 / 0.304
LIVE-Wild0.254 / 0.6420.052 / 0.417
AGIQA-3K0.558 / 0.7300.177 / 0.592
", + "bbox": [ + 93, + 518, + 480, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "significantly (e.g., $33.1\\%$ in SRCC on TID2013), underscoring that distortion type is a major gap between training and testing. Third, from #1 to #2, adding PIPAL for training greatly enhances the results on PIPAL (from 0.495 to 0.724 in PLCC) but reduces the performance on several other datasets, e.g., from 0.900 to 0.863 in PLCC on CSIQ. This performance decline is likely due to the unique nature of model-processed distortions in PIPAL, which differ greatly from other distortions. Finally, despite never trained on AI-generated images, DeQA-Score performs well on AGIQA-3K with around 0.8 PLCC, reflecting the generalization ability enabled by MLLMs' broad pre-training.", + "bbox": [ + 88, + 625, + 482, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More results on AIGC images are provided in Tab. A4.", + "bbox": [ + 89, + 809, + 465, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Results of Score Distribution Prediction", + "text_level": 1, + "bbox": [ + 89, + 833, + 431, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As stated in Sec. 4.2, the MLLM-output probabilities allow us to predict a score distribution, which can be evaluated against the human labeled ground-truth distribution. We", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "calculate the distance between these two distributions and present the results in Tab. 5. The results illustrate that our predicted score distribution aligns much better with humans than Q-Align. As shown in Fig. 6, Q-Align tends to predict a single level token, bringing a quite small predicted variance, which compromises the distribution metrics.", + "bbox": [ + 511, + 450, + 906, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results are shown in Fig. 6. Across different image types: in-the-wild images (Fig. 6ab), AI-generated images (Fig. 6c), or synthetic distortions (Fig. 6d), and for varying quality levels: high-quality (Fig. 6ac), fair-quality (Fig. 6b), or poor-quality images (Fig. 6d), our method consistently predicts accurate quality scores. Additionally, our DeQA-Score provides a score distribution that aligns more closely with humans with smaller KL divergence. In contrast, Q-Align, which is trained with one-hot labels, tends to predict a single label, resulting in higher KL divergence.", + "bbox": [ + 511, + 545, + 908, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 710, + 679, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of soft label and fidelity loss is studied and shown in Tab. 6. In #1, \"Uncertainty\" is utilized to verify the effects of adding annotation variances without considering the score distortion. We directly append uncertainty within the responses, i.e., \"The quality of this image is with uncertainty\". The has five possible values: {\"minimal\", \"low\", \"moderate\", \"high\", \"severe\"}, derived by the one-hot discretization. In #2, model trained on one-hot label can not predict a reliable variance, thus we employ binary fidelity loss [89], which does not require predicted variance.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14489", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c699409b820ce91ede379bc8da6cf3fe9c1ca7a3022565144543be88ed462c3f.jpg", + "image_caption": [ + "(a) Score: $3.93 \\pm 0.79$" + ], + "image_footnote": [], + "bbox": [ + 112, + 90, + 243, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc877a31a5fed4204432786271d4cc0da1438ba39b46f9e32abe8d535e5a5e50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 90, + 374, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8714ef4df6cf8bcee3b733e0b6d67a887732ed3f0245c7b066eae90c6237a818.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 90, + 504, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b35d2693e00ddaf1a5869e88e5dfd1c1fb60de964b0bb631c1275f10ec648a01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 90, + 620, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3319e264f5260424b2edf37eda1075b0c5e27834d81eb45fbbadad8f360aa35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 90, + 751, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/252eee6bc93d1a830eb98985d222d5d34942159a6771d6521a610b8270e511ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 90, + 880, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/178a4aabf6be7abc4c222416926a22c9cde2ebab09fcca2919558c926b5dd88e.jpg", + "image_caption": [ + "(b) Score: $3.11 \\pm 0.66$" + ], + "image_footnote": [], + "bbox": [ + 109, + 191, + 243, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45657c6b07acc75babb77f2c0dfa1d13471e286a8556b66fe5798581cf8366e6.jpg", + "image_caption": [ + "Pred: 3.95, KL Div: 9.01" + ], + "image_footnote": [], + "bbox": [ + 248, + 193, + 374, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c570ba6f583d1d6ddf5d38abdd12d11914512d281ea3b6eb689e76ec73fb45bd.jpg", + "image_caption": [ + "Pred: 3.34, KL Div: 0.06", + "Figure 6. Qualitative results on KonIQ, SPAQ, AGIQA-3K, and KADID datasets. DeQA-Score predicts more accurate quality scores with more reasonable logits. Also, our method generates a score distribution that aligns better with humans, i.e., smaller KL divergence. However, Q-Align, which is trained on one-hot label, is more likely to predict a single label, resulting in higher KL divergence." + ], + "image_footnote": [], + "bbox": [ + 377, + 193, + 504, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c05a66f95588d56945a0f9153eb720fccf1f484dd78ced507edc05d977bf07f.jpg", + "image_caption": [ + "(d) Score: $1.92 \\pm 0.82$" + ], + "image_footnote": [], + "bbox": [ + 516, + 191, + 620, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c37c48761abc7d2a2469ed5c8e0c36f0474999395086f7f474996972224c0aa.jpg", + "image_caption": [ + "Pred: 1.04, KL Div: 16.69" + ], + "image_footnote": [], + "bbox": [ + 625, + 193, + 751, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d025297513e19736e64d8cc420319a8ae35d85b43196ac11890c821b49cd47d.jpg", + "image_caption": [ + "Pred: 1.76, KL Div: 0.14" + ], + "image_footnote": [], + "bbox": [ + 754, + 191, + 880, + 262 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/be24d649743e261fb51adf957c4281e93fbfa8828ae3710ca57a94ac30e2d9b1.jpg", + "table_caption": [ + "Table 6. Ablation studies on soft label and fidelity loss with PLCC / SRCC metrics. Models are trained on KonIQ, SPAQ, and KADID." + ], + "table_footnote": [], + "table_body": "
#One-hotSoftUncertaintyFidelityKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
0✓0.945 / 0.9380.933 / 0.9310.935 / 0.9340.409 / 0.4200.887 / 0.8830.788 / 0.7330.829 / 0.8080.876 / 0.845
1✓✓0.935 / 0.9360.933 / 0.9310.954 / 0.9510.439 / 0.4530.896 / 0.8870.764 / 0.7150.826 / 0.8030.884 / 0.849
2✓✓0.944 / 0.9360.935 / 0.9320.951 / 0.9500.473 / 0.4780.898 / 0.8860.788 / 0.7260.837 / 0.8100.876 / 0.839
3✓0.954 / 0.9430.937 / 0.9340.951 / 0.9470.472 / 0.4720.896 / 0.8790.791 / 0.7290.852 / 0.8210.894 / 0.855
4✓✓0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
", + "bbox": [ + 94, + 344, + 900, + 431 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d5908637415d6f56817dc07ad7c4d5ab24e970a365d0f64c955e1a545077165d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryLevel TextsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
Commonexce. / good / far / poor / bad0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
Randomexce. / poor / bad / good / fair0.942 / 0.9280.933 / 0.9290.950 / 0.9470.394 / 0.3960.863 / 0.8510.795 / 0.7270.814 / 0.7920.887 / 0.836
Reversebad / poor / fair / good / exce.0.943 / 0.9340.934 / 0.9310.947 / 0.9450.395 / 0.3860.836 / 0.8000.782 / 0.7140.811 / 0.7910.882 / 0.834
Numberone / two / three / four / five0.945 / 0.9320.934 / 0.9300.955 / 0.9520.421 / 0.4240.873 / 0.8570.766 / 0.7120.820 / 0.7940.882 / 0.835
Vehiclecar / bus / train / plane / ship0.953 / 0.9400.937 / 0.9330.954 / 0.9510.458 / 0.4560.889 / 0.8780.778 / 0.7020.854 / 0.8280.897 / 0.856
Mixtureapple / car / four / fair / dog0.947 / 0.9380.936 / 0.9320.954 / 0.9510.419 / 0.4150.865 / 0.8510.808 / 0.7370.827 / 0.8000.896 / 0.854
", + "bbox": [ + 93, + 455, + 900, + 555 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b05c9986ec578fc831ea43045f61215f8fb537d64696b22a740144c27204d08e.jpg", + "table_caption": [ + "Table 7. Effects of different level texts with PLCC / SRCC metrics. Models are trained on the KonIQ, SPAQ, and KADID datasets.", + "Table 8. Inference latency & throughput on RTX A6000 GPU." + ], + "table_footnote": [], + "table_body": "
Batch Size12481664256
Latency (s)0.0890.1210.1820.3150.5882.1147.929
Throughput (image/s)11.2416.5321.9825.4027.2130.2732.29
", + "bbox": [ + 93, + 579, + 480, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Several conclusions can be drawn from Tab. 6. First, comparing #0 with #3, our soft label significantly outperforms the one-hot label across all benchmarks, validating its effectiveness. Second, from #3 to #4, the performance of our soft label further improves when combined with the fidelity loss. Third, comparing #0 with #2, the one-hot label does not incorporate well with the fidelity loss, as it lacks score distribution information. Finally, from #0 to #1, simply adding uncertainty without considering score distribution does not yield stable or significant improvements.", + "bbox": [ + 89, + 641, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of different level texts are studied in Tab. 7. Beside the commonly used level texts, we consider five alternatives: (a) random sorted common texts; (b) reversed common texts; (c) simple numbers like one, two, etc.; (d) irrelevant words about vehicle; (e) a mixture of various types. The results indicate that common level texts achieve the best performance, as they align closely with the MLLMs' pre", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training. Interestingly, with random sorted / reversed texts, even a mixture of various level texts, which are inconsistent with MLLMs' prior knowledge, our model still learns these newly defined relationships and performs well. This shows the robust adaptability of MLLM-based IQA models.", + "bbox": [ + 511, + 568, + 903, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies on level number and training / fixing components are provided in Tabs. A6 and A7, respectively.", + "bbox": [ + 511, + 646, + 903, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Inference Latency", + "text_level": 1, + "bbox": [ + 511, + 686, + 687, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We test the inference latency on one single RTX A6000 GPU. As shown in Tab. 8, with the optimal GPU utilization (i.e., batch size 256), our DeQA-Score can process as many as 32.29 images per second in average. This high throughput emphasizes its suitability for real-time deployment.", + "bbox": [ + 511, + 709, + 903, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 799, + 638, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We develop DeQA-Score as a pioneering MLLM-based image quality scorer, empowered by distribution-based soft labels and fidelity loss, capable of both score regression and score distribution prediction, highlighting the importance of quality score distribution in MLLM-based IQA methods.", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14490", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was sponsored by RGC Early Career Scheme (ECS) No. 24209224, National Natural Science Foundation of China (Grant No.62276251), and the Joint Lab of CAS-HK.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 165, + 187, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019. 3", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2", + "[4] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE TIP, 2018. 2", + "[5] Xin Cai, Zhiyuan You, Hailong Zhang, Wentao Liu, Jinwei Gu, and Tianfan Xue. PhoCoLens: Photorealistic and consistent reconstruction in lensless imaging. In NeurIPS, 2024. 1", + "[6] Yue Cao, Zhaolin Wan, Dongwei Ren, Zifei Yan, and Wangmeng Zuo. Incorporating semi-supervised and positive unlabeled learning for boosting full reference image quality assessment. In CVPR, 2022. 2", + "[7] Chaofeng Chen, Yang Sensen, Haoning Wu, Liang Liao, Zicheng Zhang, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-Ground: Image quality grounding with large multi-modality models. In ACM MM, 2024. 1, 3", + "[8] Du Chen, Tianhe Wu, Kede Ma, and Lei Zhang. Toward generalized image quality assessment: Relaxing the perfect reference quality assumption. In CVPR, 2025. 2", + "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 3", + "[10] Zixuan Chen, Zewei He, Ziqian Lu, Xuecheng Sun, and Zhe-Ming Lu. Prompt-based test-time real image dehazing: a novel pipeline. In ECCV, 2024. 1", + "[11] Zijian Chen, Wei Sun, Haoning Wu, Zicheng Zhang, Jun Jia, Ru Huang, Xiongkuo Min, Guangtao Zhai, and Wenjun Zhang. Study of subjective and objective naturalness assessment of ai-generated images. IEEE TCSVT, 2024. 14", + "[12] Zixuan Chen, Yujin Wang, Xin Cai, Zhiyuan You, Zheming Lu, Fan Zhang, Shi Guo, and Tianfan Xue. UltraFusion: Ultra high dynamic imaging using exposure fusion. In CVPR, 2025. 1", + "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-" + ], + "bbox": [ + 93, + 191, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. https://vicuna.lmsys.org, 2023. 2, 3", + "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2", + "[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE TPAMI, 2020. 2", + "[16] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In ACM MM, 2021. 2", + "[17] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, 2020. 1, 2, 4, 6", + "[18] Deepti Ghadiyaram and C. Bovik Alan. Live in the wild image quality challenge database. http://live.ece.utexas.edu/research/ChallengeDB/index.html, 2015.2,6", + "[19] Abhijay Ghildyal and Feng Liu. Shift-tolerant perceptual similarity metric. In ECCV, 2022. 2", + "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3", + "[21] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. KonIQ-10K: An ecologically valid database for deep learning of blind image quality assessment. IEEE TIP, 2020. 2, 4, 6", + "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2021. 13, 14, 15", + "[23] Loshchilov Ilya and Hutter Frank. Decoupled weight decay regularization. In ICLR, 2019. 6, 13", + "[24] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. PIPAL: a large-scale image quality assessment dataset for perceptual image restoration. In ECCV, 2020. 2, 6", + "[25] Le Kang, Peng Ye, Yi Li, and David Doermann. Convolutional neural networks for no-reference image quality assessment. In CVPR, 2014. 2", + "[26] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality transformer. In CVPR, 2021. 1, 2, 6, 7", + "[27] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 2018. 13", + "[28] Daekyu Kwon, Dongyoung Kim, Sehwan Ki, Younghyun Jo, Hyong-Euk Lee, and Seon Joo Kim. ATTIQA: Generalizable image quality feature extractor using attribute-aware pretraining. In ACCV, 2024. 6" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14491", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Eric C Larson and Damon M Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 2010. 6", + "[30] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. AGIQA-3K: An open database for ai-generated image quality assessment. IEEE TCSVT, 2023. 1, 2, 6, 14", + "[31] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. KADID-10K: A large-scale artificially distorted iqa database. In International Conference on Quality of Multimedia Experience (QoMEX), 2019. 2, 4, 6", + "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1, 2, 4", + "[33] Kai Liu, Ziqing Zhang, Wenbo Li, Renjing Pei, Fenglong Song, Xiaohong Liu, Linghe Kong, and Yulun Zhang. Dog-IQA: Standard-guided zero-shot mllm for mix-grained image quality assessment. arXiv preprint arXiv:2410.02505, 2024.3", + "[34] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. RankIQA: Learning from rankings for no-reference image quality assessment. In ICCV, 2017. 2", + "[35] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 3", + "[36] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, 2022. 3", + "[37] Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding, 2017. 2", + "[38] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In ACL Findings, 2022. 3", + "[39] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. DocVQA: A dataset for vqa on document images. In WACV, 2021. 3", + "[40] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE TIP, 2012. 2, 6, 7", + "[41] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Sign. Process. Letters, 2012. 1, 6, 7", + "[42] Anush Krishna Moorthy and Alan Conrad Bovik. A two-step framework for constructing blind image quality indices. IEEE Sign. Process. Letters, 2010.", + "[43] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE TIP, 2011. 2", + "[44] OpenAI. GPT-4V(ision) system card, 2023. 1, 2", + "[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 4", + "[46] Da Pan, Ping Shi, Ming Hou, Zefeng Ying, Sizhe Fu, and Yuan Zhang. Blind predicting similar quality map for image quality assessment. In CVPR, 2018. 2", + "[47] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database TID2013: Peculiarities, results and perspectives. Signal processing: Image communication, 2015. 6", + "[48] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. PieAPP: Perceptual image-error assessment through pairwise preference. In CVPR, 2018. 2", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 13", + "[50] Michele A Saad, Alan C Bovik, and Christophe Charrier. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE TIP, 2012. 2", + "[51] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 2012. 3", + "[52] H Sheikh. LIVE image quality assessment database release 2. http://live.ece.utexas.edu/research/quality, 2005. 1, 2", + "[53] Hamid R Sheikh and Alan C Bovik. Image information and visual quality. IEEE TIP, 2006. 2", + "[54] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA models that can read. In CVPR, 2019. 3", + "[55] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In CVPR, 2020. 2, 6, 7", + "[56] Simen Sun, Tao Yu, Jiahua Xu, Wei Zhou, and Zhibo Chen. GraphIQA: Learning distortion graph representations for blind image quality assessment. IEEE TMM, 2022. 2", + "[57] Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment. IEEE TIP, 2018. 6, 7", + "[58] Huixuan Tang, Neel Joshi, and Ashish Kapoor. Learning a blind measure of perceptual image quality. In CVPR, 2011. 2", + "[59] InternLM Team. InternLM: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM, 2023. 2", + "[60] Louis L Thurstone. A law of comparative judgment. Psychological review, 1994. 2, 5", + "[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 3, 4", + "[62] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14492", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. \nLLaMA 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6, 13", + "[63] Ming-Feng Tsai, Tie-Yan Liu, Tao Qin, Hsin-Hsi Chen, and Wei-Ying Ma. FRank: A ranking method with fidelity loss. In International ACM SIGIR Conference on Research and Development in Information Retrieval, 2007. 2, 5", + "[64] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 2, 6, 7", + "[65] Jiarui Wang, Huiyu Duan, Jing Liu, Shi Chen, Xiongkuo Min, and Guangtao Zhai. AIGCIQA2023: A large-scale image quality assessment database for ai generated images: from the perspectives of quality, authenticity and correspondence. In CAAI International Conference on Artificial Intelligence, 2023. 14", + "[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 1, 2", + "[67] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv preprint arXiv:2312.06109, 2023. 2", + "[68] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-Bench: A benchmark for general-purpose foundation models on low-level vision. In ICLR, 2024. 3, 15, 16", + "[69] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-Instruct: Improving low-level visual abilities for multi-modality foundation models. In CVPR, 2024. 1, 2, 3, 15, 16", + "[70] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-Align: Teaching LMMs for visual scoring via discrete text-defined levels. In ICML, 2024. 1, 2, 3, 4, 5, 6, 7, 14, 15", + "[71] Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, et al. Towards open-ended visual quality comparison. In ECCV, 2024. 1, 2, 3", + "[72] Tianhe Wu, Kede Ma, Jie Liang, Yujiu Yang, and Lei Zhang. A comprehensive study of multimodal large language models for image quality assessment. In ECCV, 2024. 3", + "[73] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. MANIQA: Multi-dimension attention network for no-reference image quality assessment. In CVPRW, 2022. 2, 6, 7", + "[74] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mPLUG-Owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mPLUG-Owl2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 1, 4, 6, 13", + "[76] Guanghao Yin, Wei Wang, Zehuan Yuan, Chuchu Han, Wei Ji, Shouqian Sun, and Changhu Wang. Content-variant reference image quality assessment via knowledge distillation. In AAAI, 2022. 2", + "[77] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Lu Sheng, Lei Bai, Xiaoshui Huang, Zhiyong Wang, et al. Lamm: Language-assisted multimodal instruction-tuning dataset, framework, and benchmark. In NeurIPS, 2023. 2", + "[78] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (PaQ-2-PiQ): Mapping the perceptual space of picture quality. In CVPR, 2020. 2, 6", + "[79] Zhiyuan You, Jinjin Gu, Zheyuan Li, Xin Cai, Kaiwen Zhu, Chao Dong, and Tianfan Xue. Descriptive image quality assessment in the wild. arXiv preprint arXiv:2405.18842, 2024. 1, 2, 3", + "[80] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. In ECCV, 2024. 1, 2, 3", + "[81] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2014. 3", + "[82] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. FSIM: A feature similarity index for image quality assessment. IEEE TIP, 2011. 2", + "[83] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2", + "[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 1, 2", + "[85] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. LLaMA-Adapter: Efficient fine-tuning of large language models with zero-initialized attention. In ICLR, 2024. 2", + "[86] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE TCSVT, 2020. 6, 7", + "[87] Weixia Zhang, Kede Ma, Guangtao Zhai, and Xiaokang Yang. Uncertainty-aware blind image quality assessment in the laboratory and wild. IEEE TIP, 2021. 2, 3, 5, 6", + "[88] Weixia Zhang, Dingquan Li, Chao Ma, Guangtao Zhai, Xiaokang Yang, and Kede Ma. Continual learning for blind image quality assessment. IEEE TPAMI, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14493", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[89] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In CVPR, 2023. 2, 7, 14, 15", + "[90] Zicheng Zhang, Chunyi Li, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. A perceptual quality assessment exploration for aigc images. In ICME Workshops, 2023. 14", + "[91] Zicheng Zhang, Haoning Wu, Erli Zhang, Guangtao Zhai, and Weisi Lin. A benchmark for multi-modal foundation models on low-level vision: from single images to pairs. IEEE TPAMI, 2024. 3", + "[92] Zicheng Zhang, Yingjie Zhou, Chunyi Li, Baixuan Zhao, Xiaohong Liu, and Guangtao Zhai. Quality assessment in the era of large models: A survey. arXiv preprint arXiv:2409.00031, 2024. 3", + "[93] Heliang Zheng, Jianlong Fu, Yanhong Zeng, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. ICCV, 2021. 2", + "[94] Wei Zhou and Zhou Wang. Quality assessment of image super-resolution: Balancing deterministic and statistical fidelity. In ACM MM, 2022. 2", + "[95] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 1, 2", + "[96] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. MetaQA: deep meta-learning for no-reference image quality assessment. In CVPR, 2020. 2", + "[97] Hanwei Zhu, Xiangjie Sui, Baoliang Chen, Xuelin Liu, Peilin Chen, Yuming Fang, and Shiqi Wang. 2afc prompting of large multimodal models for image quality assessment. arXiv preprint arXiv:2402.01162, 2024. 3", + "[98] Hanwei Zhu, Haoning Wu, Yixuan Li, Zicheng Zhang, Bao-liang Chen, Lingyu Zhu, Yuming Fang, Guangtao Zhai, Weisi Lin, and Shiqi Wang. Adaptive image quality assessment via teaching large multimodal model to compare. arXiv preprint arXiv:2405.19298, 2024. 2, 3, 5, 6, 7, 14, 15", + "[99] Kaiwen Zhu, Jinjin Gu, Zhiyuan You, Yu Qiao, and Chao Dong. An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809, 2024. 1" + ], + "bbox": [ + 91, + 92, + 480, + 698 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14494", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_model.json b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a0357fe77d0518173b71b05f2508ba1594b57a09 --- /dev/null +++ b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_model.json @@ -0,0 +1,3227 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.885, + 0.177 + ], + "angle": 0, + "content": "Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.203, + 0.769, + 0.223 + ], + "angle": 0, + "content": "Zhiyuan You\\(^{12}\\), Xin Cai\\(^{2}\\), Jinjin Gu\\(^{4}\\), Tianfan Xue\\(^{23\\dagger}\\), Chao Dong\\(^{134\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.224, + 0.805, + 0.241 + ], + "angle": 0, + "content": "1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.242, + 0.76, + 0.26 + ], + "angle": 0, + "content": "2The Chinese University of Hong Kong, 3Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.261, + 0.687, + 0.278 + ], + "angle": 0, + "content": "4Shenzhen University of Advanced Technology" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.28, + 0.84, + 0.295 + ], + "angle": 0, + "content": "zhiyuanyou@link.cuhk.edu.hk, caixin@link.cuhk.edu.hk, jinjin.gu@sydney.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.297, + 0.77, + 0.312 + ], + "angle": 0, + "content": "tfxue@ie.cuhk.edu.hk, chao.dong@siat.ac.cn † Corresponding Author" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.347, + 0.327, + 0.364 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.485, + 0.818 + ], + "angle": 0, + "content": "With the rapid advancement of Multi-modal Large Language Models (MLLMs), MLLM-based Image Quality Assessment (IQA) methods have shown promising performance in linguistic quality description. However, current methods still fall short in accurately scoring image quality. In this work, we aim to leverage MLLMs to regress accurate quality scores. A key challenge is that the quality score is inherently continuous, typically modeled as a Gaussian distribution, whereas MLLMs generate discrete token outputs. This mismatch necessitates score discretization. Previous approaches discretize the mean score into a one-hot label, resulting in information loss and failing to capture inter-image relationships. We propose a distribution-based approach that discretizes the score distribution into a soft label. This method preserves the characteristics of the score distribution, achieving high accuracy and maintaining inter-image relationships. Moreover, to address dataset variation, where different IQA datasets exhibit various distributions, we introduce a fidelity loss based on Thurstone's model. This loss captures intra-dataset relationships, facilitating co-training across multiple IQA datasets. With these designs, we develop the distribution-based Depicted image Quality Assessment model for Score regression (DeQA-Score). Experiments across multiple benchmarks show that DeQA-Score stably outperforms baselines in score regression. Also, DeQA-Score can predict the score distribution that closely aligns with human annotations. Codes and model weights have been released in https://depictqa.github.io/deqa-score/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.831, + 0.222, + 0.847 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Image Quality Assessment (IQA) focuses on perceiving the visual quality of images [26, 41, 66, 84], and is widely adopted in various application scenarios, e.g., image com" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.345, + 0.659, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.347, + 0.774, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.346, + 0.892, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.426, + 0.658, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.426, + 0.774, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.426, + 0.892, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.511, + 0.658, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.511, + 0.774, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.511, + 0.892, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.603, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Figure 1. Illustration of our distribution-based soft label. Image A is overexposed, while Image B and C are well-exposed with better and similar quality. \"exce.\" means \"excellent\". To train an MLLM as a quality scorer, continuous scores need to be discretized to discrete level tokens as the training label. Q-Align [70] discretizes human-labeled mean score to a one-hot label, causing information loss. We discretize the score distribution, which is approximated as a Gaussian distribution, to obtain a more accurate soft label. Our soft label maintains better inter-image relationships that the Image B is better than A (A < B), and Image B and C share similar quality (B ≈ C), while the one-hot label can not." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "pression and transmission [5, 52], processing smartphone-captured photos [17], and evaluating generation or restoration models [10, 12, 30, 99], etc. With the emergence of Multi-modal Large Language Models (MLLMs) [32, 44, 75, 95], MLLM-based IQA begins to attract more research interests [7, 69, 71, 79, 80]. These methods mainly leverage MLLMs to describe image quality using language, recognizing that language better mirrors human expression, and captures the multifaceted nature of IQA tasks [80]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14483" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "However, for quality scoring, which needs an accurate numerical score to quantify an image's quality, existing description-centered MLLMs [69, 71, 79, 80] perform less accurately than traditional IQA methods. This hinders their applications on real-world IQA scenarios. Considering MLLMs' good generalization ability on various downstream tasks and promising performance on quality description, it is possible for MLLMs to score the image quality. Therefore, in this work, we aim to explore how MLLMs can effectively regress accurate quality scores by leveraging their broad training and inherent adaptability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.502 + ], + "angle": 0, + "content": "We find that the principal challenge in training MLLMs to predict accurate quality scores is the distribution gap between MLLMs' outputs and quality scores. MLLMs generate outputs as discrete tokens, while the quality score is inherently continuous, usually approximated to a Gaussian distribution. Previous approaches, like Q-Align [70], attempt to address the distribution gap by directly discretizing the mean score (i.e., the expectation of estimated score distribution) into a one-hot label, as depicted in Fig. 1. However, such a one-hot discretization method introduces information loss, causing mismatches between the discretized labels and the original continuous scores. For instance, in Fig. 1, Image A is overexposed with a lower mean score than Image B, but their one-hot labels appear identical. Conversely, Image B and C are well-exposed with similar quality scores, yet their discretized labels are different." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.747 + ], + "angle": 0, + "content": "To overcome the distribution gap, we propose a novel soft label to bridge continuous and discrete distributions. As illustrated in Fig. 1, instead of discretizing the mean of score distribution, we discretize the continuous score distribution itself into a discrete distribution of different levels. During training, this discrete distribution is used as the soft training label. This enables effective training with minimal information loss. Also, our distribution-based soft label better preserves inter-image relationship. For example, in Fig. 1, Image B is better than Image A, and Image B and C share similar quality, while the simple one-hot label fails to capture such relationships. At inference, MLLMs output discrete token results, which can be recovered to continuous score distribution. This method ensures that the discrete outputs align well with the underlying continuous distribution, allowing to accurately predict the desired score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Another key challenge is the dataset variation, where different IQA datasets exhibit distinct distributions, making it hard to simply combine them for co-training. Due to this variation, images from different datasets may have nearly the same quality scores (i.e., linearly re-scaled scores) but drastically different perceptual quality [87, 98]. A previous method, UNIQUE [87], has demonstrated that the fidelity loss [63] under the Thurstone's model [60] can effectively capture the intra-dataset relationships, thus facilitating the joint training of multiple IQA datasets. However," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.197 + ], + "angle": 0, + "content": "the fidelity loss requires the model to predict an accurate score distribution, which MLLMs trained with the one-hot label cannot provide. Fortunately, MLLMs trained with our distribution-based soft label can predict such an accurate score distribution, allowing the adaptation of the fidelity loss. We apply this fidelity loss to our MLLM-based quality scorer and demonstrate its effectiveness in this new context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.394 + ], + "angle": 0, + "content": "Equipped with these designs, we successfully develop a distribution-based multi-modal image quality assessment model, termed DeQA-Score. DeQA-Score achieves better score regression results than baseline methods on both in-distribution (e.g., \\(1.3\\%\\) in PLCC on KonIQ [21]) and out-of-distribution datasets (e.g., \\(4.6\\%\\) in PLCC on LIVEWILD [18]). The advantage of our DeQA-Score becomes more pronounced with the co-training across multiple IQA datasets. Moreover, our method can further predict the distribution of quality scores, which is closely aligned with the distribution derived from extensive human annotations. We hope that our DeQA-Score will inspire more research to explore the potential of quality score distribution." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.405, + 0.663, + 0.421 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.905, + 0.492 + ], + "angle": 0, + "content": "Image Quality Assessment (IQA) methods mainly rely on quality scores to assess image quality. They can be categorized into full-reference and non-reference methods through whether there is a high-quality reference image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.905, + 0.656 + ], + "angle": 0, + "content": "Full-reference IQA methods compute a similarity score as the quality score between a distorted image and a high-quality reference. Classical works rely on human-designed metrics such as structural similarity [66], phase congruency with gradient magnitude [82], image information [53], etc. The rapid advancement of deep learning has also inspired learning-based IQA methods that measure image quality through data-driven training. Pioneered by LPIPS [84] and PieAPP [48], driven by large-scale IQA datasets [24, 31, 52, 84], deep-learning methods [4, 6, 8, 15, 16, 19, 76, 94] have exhibited quite high accuracy in quality score regression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.905, + 0.808 + ], + "angle": 0, + "content": "Non-reference IQA methods directly regress a quality score without a reference image. Initially, hand-crafted natural image statistics are adopted [37, 40-43, 50, 58]. Subsequently, deep-learning-based methods [25, 34, 46, 55, 56, 78, 88, 93, 96] replace hand-crafted statistics by learning quality priors from human-annotated IQA datasets [17, 21, 30, 31]. Recent works further improve the regression accuracy by introducing multi-dataset co-training strategy [87], multi-scale features [26], CLIP pre-training [64], multidimension attention [73], multitask learning [89], and so on." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Multi-modal Large Language Models (MLLMs) incorporate visual modality into Large Language Models (LLMs) [1, 13, 59, 61] to achieve general visual ability by leveraging the broad training and powerful generalization ability of LLMs. Existing MLLMs [3, 14, 32, 44, 67, 74, 77, 83, 85, 95] have demonstrated a general visual ability" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14484" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "and can tackle a variety of multi-modality tasks, like visual question answering [20, 35, 36], document understanding [38, 39, 54], image captioning [2, 9, 81], etc. Given MLLMs' advanced performance in these high-level perception tasks, it is natural to explore how MLLMs perform in low-level perception tasks that are highly related to IQA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.483, + 0.457 + ], + "angle": 0, + "content": "MLLM-based IQA methods leverages MLLMs' foundational knowledge to achieve better IQA performance or more detailed assessment results [72, 92]. Q-Bench [68, 91] shows that general MLLMs have some low-level perception ability. [97] evaluates various MLLMs on the widely-adopted Two-Alternative Forced Choice (2AFC) task. Q-Instruct [69] enhances the low-level perception ability of MLLMs by introducing a large-scale dataset. CoInstruct [71] concentrates on the quality comparison among multiple images. DepictQA [79, 80] performs single-image assessment and paired-image comparison in both full-reference and non-reference settings. Q-Ground [7] focuses on the visual quality grounding task, and performs fine-scale detailed visual quality analysis. Some of these works have proposed MLLM-based quality scoring methods, e.g., binary softmax strategy in [68, 91] and pair-wise voting in [79, 80]. Nonetheless, these works principally focus on the quality description ability using languages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.671 + ], + "angle": 0, + "content": "MLLM-based quality scorer. Initially, Q-Bench [68, 91] proposes a binary softmax strategy, enabling MLLMs to generate quality scores by predicting two discrete quality levels (i.e., good or poor). This strategy is adopted by Q-Instruct [69] and Co-Instruct [71]. Compare2Score [98] achieves quality scoring by teaching MLLMs to compare pair-wise images. Inspired by humans' annotation process, Q-Align [70] discretizes scores to five discrete levels using one-hot label to train MLLMs, resulting in more accurate score regression. Dog-IQA [33] adopts the one-hot label for training-free IQA by incorporating specific standards and local semantic objects. However, as discussed in Sec. 1, one-hot label causes some information loss and can not fully maintain the inter-image relationship." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.684, + 0.317, + 0.699 + ], + "angle": 0, + "content": "3. Soft Label Construction" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.32, + 0.724 + ], + "angle": 0, + "content": "3.1. Revisiting One-hot Label" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.483, + 0.867 + ], + "angle": 0, + "content": "Training with one-hot label. MLLMs only accept discrete tokens as inputs and outputs. Therefore, as shown in Fig. 2a, continuous scores1 must be discretized into discrete tokens to train the MLLMs. Q-Align [70] adopts a one-hot discretization method. Specifically, as depicted in Fig. 2b, the score range [1, 5] is uniformly divided into five intervals. These intervals represent five rating levels, {\"bad\", \"poor\", \"fair\", \"good\", \"excellent\"}, which is the rating standard defined by [51]. In Fig. 2a, the Mean Opinion Score (MOS)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1 Unless otherwise specified, the mean quality scores in all datasets will be normalized to [1, 5], and the variances will be normalized accordingly." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.092, + 0.677, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.096, + 0.824, + 0.107 + ], + "angle": 0, + "content": "Rate the quality of this image." + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.11, + 0.888, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.155, + 0.542, + 0.166 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.139, + 0.856, + 0.149 + ], + "angle": 0, + "content": "5 candidate level tokens of :" + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.15, + 0.882, + 0.162 + ], + "angle": 0, + "content": "[ bad, poor, fair, good, excellent ]" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.171, + 0.716, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.171, + 0.891, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.231, + 0.716, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.231, + 0.891, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.291, + 0.542, + 0.302 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.291, + 0.685, + 0.301 + ], + "angle": 0, + "content": "Continuous score" + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.291, + 0.859, + 0.301 + ], + "angle": 0, + "content": "Discretized label" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.31, + 0.905, + 0.448 + ], + "angle": 0, + "content": "Figure 2. (a) MLLMs need discrete level tokens for training. Thus, continuous scores must be converted into discrete level tokens as training labels. (b) Comparison between one-hot label [70] and our soft label. One-hot label uniformly divides the score range into 5 intervals to discretize the mean score, i.e., the expectation of estimated score distribution. One-hot label assumes that different levels are identically independent, which is inaccurate. In contrast, we discretize the estimated Gaussian distribution of the score. Our soft label better preserves the relationships between levels, e.g., \"fair\" is closer to \"good\" than \"excellent\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.548 + ], + "angle": 0, + "content": "of the image, 4.30, fails into the fifth intervals, resulting in the discretized label \"excellent\". Accordingly, the response is formulated as \"The quality of this image is excellent\". Using the response as the training labels, following [13, 61], MLLMs can be trained with next token prediction loss, i.e., cross-entropy loss between predicted logits and labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Limitations of one-hot label. First, ideal training with cross-entropy loss using one-hot labels assumes that the five levels are independent (i.e., orthogonal in a latent space). However, intuitively, these levels are not truly independent, e.g., \"fair\" is closer to \"good\" than \"excellent\". Second, one-hot discretization only considers the mean score, which is the expectation of the score distribution, ignoring the distribution's uncertainty (i.e., variance). Third, near the border of two levels, images with similar quality can be divided into different levels. For example, in Fig. 1, Image B (MOS 3.38) and Image C (MOS 3.49) share similar mean scores, but are differently classified as \"fair\" and \"good\", due to their proximity to the border of these two levels, i.e., 3.4 in Fig. 2b. Finally, one-hot label causes information loss. In Q-Align, the five levels are respectively assigned scores \\(\\{1, 2, 3, 4, 5\\}\\), with which we calculate the discretization precision in Tab. 1. The discretization error is around \\(10 \\times\\) to \\(35 \\times\\) larger than our soft label that we will soon introduce, thus the one-hot label keeps lower correlation with MOS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.63, + 0.862 + ], + "angle": 0, + "content": "3.2. Soft Label" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Discretization using score distribution. As demonstrated in [87], the quality score of one image, \\( x \\), is typically mod-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14485" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.16 + ], + "angle": 0, + "content": "Table 1. Comparison of discretization precision between Q-Align's one-hot label [70] and our soft label. The precision of MOS discretization is measured by L1 Error / RMSE and PLCC / SRCC. The distance of original distribution and recovered distribution is measured by JS divergence / Wasserstein distance." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.276 + ], + "angle": 0, + "content": "
MetricsDiscretizationKonIQ [21]SPAQ [17]KADID [31]
L1 / RSMEOne-hot label0.302 / 0.3740.299 / 0.3660.327 / 0.389
Soft label0.008 / 0.0180.011 / 0.0250.024 / 0.038
PLCC / SRCCOne-hot label0.961 / 0.9520.969 / 0.9680.979 / 0.982
Soft label1.000 / 1.0001.000 / 1.0001.000 / 1.000
JS / W-Dist.One-hot label---
Soft label0.001 / 0.0380.022 / 0.0840.010 / 0.140
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.485, + 0.419 + ], + "angle": 0, + "content": "eled as a Gaussian distribution, \\(x \\sim \\mathcal{N}(\\mu, \\sigma^2)\\), where \\(\\mu\\) and \\(\\sigma^2\\) are the Mean Opinion Score (MOS) and variance of multiple human annotations, respectively. The distribution is represented by its probability density function, \\(f(x)\\). As shown in the second row of Fig. 2b, we select five central points, \\(c_i \\in \\{1, 2, 3, 4, 5\\}\\), \\(i \\in \\{0, 1, 2, 3, 4\\}\\), as the centers of the five discrete levels. The width of each level region, \\(d\\), is set as 1. The possibility of the score falling into the \\(i^{\\mathrm{th}}\\) level is denoted as \\(p_i^{raw}\\) and is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.424, + 0.483, + 0.462 + ], + "angle": 0, + "content": "\\[\np _ {i} ^ {r a w} = \\int_ {c _ {i} - d / 2} ^ {c _ {i} + d / 2} f (x) \\mathrm {d x}, \\quad i \\in \\{0, 1, 2, 3, 4 \\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.484, + 0.528 + ], + "angle": 0, + "content": "We assign the five levels with the same textual tokens as Q-Align, i.e., {\"bad\", \"poor\", \"fair\", \"good\", \"excellent)}. Correspondingly, the score of \\(i^{\\text{th}}\\) level is \\(c_i\\). This defines a rough discrete distribution as \\(P^{raw}(c_i) = p_i^{raw}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.484, + 0.665 + ], + "angle": 0, + "content": "Post-adjustment. The discrete distribution is rough because of truncation errors. As shown in the second row of Fig. 2b, there are two truncation regions (i.e., \\( x < c_0 - d/2 = 0.5 \\) and \\( x > c_4 + d/2 = 5.5 \\)), which are not considered in Eq. (1). This can cause two problems. First, the sum of \\( \\{p_i^{raw}\\} \\) does not equal 1, meaning \\( p_i^{raw} \\) is not strictly a discrete distribution. Second, the expectation of this discrete distribution does not equal the expectation of score distribution, \\( \\mu \\), leading to discretization errors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.484, + 0.711 + ], + "angle": 0, + "content": "To correct these truncation errors, we propose a postadjustment method. Specifically, we apply a linear transformation to adjust \\( p_i^{raw} \\) to more accurate \\( p_i \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.717, + 0.33, + 0.734 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\alpha p _ {i} ^ {r a w} + \\beta\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.737, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\quad \\sum_ {i} p _ {i} = 1, \\quad \\mu^ {r e c} = \\sum_ {i} p _ {i} c _ {i} = \\mu , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mu_{rec}\\) means recovered expectation from discrete distribution. The parameters of the linear transformation, \\(\\alpha\\) and \\(\\beta\\), can be determined by solving these two constraint conditions. One example of post-adjustment is given in Fig. 3. Moreover, the statistics of these two parameters in Tab. 2 show that the average values of \\(\\alpha\\) and \\(\\beta\\) are quite close to 1 and 0, respectively, implying that the post-adjustment is minimal. This effectively defines our soft label, which follows the discrete distribution as \\(P(c_i) = p_i\\)." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.089, + 0.673, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.181, + 0.677, + 0.21 + ], + "angle": 0, + "content": "Figure 3. Illustration of the soft label adjustment." + }, + { + "type": "table_caption", + "bbox": [ + 0.677, + 0.088, + 0.906, + 0.158 + ], + "angle": 0, + "content": "Table 2. Statistics of the label adjustment parameters. Averaged \\(\\alpha\\) and \\(\\beta\\) are very approaching 1 and 0, respectively, meaning that the postadjustment is quite small." + }, + { + "type": "table", + "bbox": [ + 0.678, + 0.162, + 0.905, + 0.208 + ], + "angle": 0, + "content": "
KonIQ [21]SPAQ [17]KADID [31]
α1.036±0.0411.033±0.0851.118±0.091
β-0.005±0.005-0.005±0.014-0.015±0.010
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.218, + 0.907, + 0.37 + ], + "angle": 0, + "content": "Ideally, the discretization error of the MOS will be 0. However, the possibility of some levels may slightly fall below zero after the linear transformation. We address this by simply clipping these values to zero. The impact of this clipping on accuracy is negligible, as shown in Tab. 1. After clipping, our discretization error for MOS remains around \\(10 \\times\\) to \\(35 \\times\\) smaller than Q-Align's one-hot label. Also, our soft label maintains a perfect (i.e., 1.000) Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank-order Correlation Coefficient (SRCC) with the original MOS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.372, + 0.905, + 0.387 + ], + "angle": 0, + "content": "Recovery from discrete to continuous score distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.464 + ], + "angle": 0, + "content": "As shown in the second condition of Eq. (2), we can compute the expectation of discrete distribution, \\( P(c_{i}) \\), as the recovered expectation of original distribution. Besides the expectation, we can also recover the original variance using the variance of the discrete distribution, \\( P(c_{i}) \\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.473, + 0.907, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mu^ {r e c} = \\sum_ {i} p _ {i} c _ {i}, \\quad (\\sigma^ {r e c}) ^ {2} = \\sum_ {i} p _ {i} (c _ {i} - \\mu^ {r e c}) ^ {2}. \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.906, + 0.58 + ], + "angle": 0, + "content": "Then we calculate the distance between recovered distribution, \\(\\mathcal{N}(\\mu^{rec}, (\\sigma^{rec})^2)\\), and original distribution, \\(\\mathcal{N}(\\mu, \\sigma^2)\\) as provided in Tab. 1. Our recovered distribution is closely aligned with human-annotated score distribution. In contrast, one-hot label can not recover such a distribution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.704, + 0.61 + ], + "angle": 0, + "content": "4. DeQA-Score Model" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.776, + 0.634 + ], + "angle": 0, + "content": "4.1. Training with Our Soft Label" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.776 + ], + "angle": 0, + "content": "Model architecture. As illustrated in Fig. 4, we follow the widely adopted MLLM architecture [32, 75] to construct our model. First, a vision encoder is adopted to represent the input image to visual tokens. Then, a vision-to-text connector projects the visual tokens into the textual space. Following [75], we use a vision abstractor as the connector. The abstractor can further reduce the number of visual tokens to 64. Finally, the visual and textual tokens are fused together and fed into an LLM for response prediction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.778, + 0.906, + 0.869 + ], + "angle": 0, + "content": "Next token prediction loss for common response tokens. As described in Sec. 3.1 and illustrated in Fig. 4, the constructed response follows the template, \"The quality of this image is \\(< \\) level \\(>\\)\". For common response tokens other than the level token, we compute the next token prediction loss, i.e., the cross-entropy loss \\(\\mathcal{L}_{ce}\\), following [45, 61]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "KL divergence loss for the level token. There are five candidate values for the level token: {\"bad\", \"poor\", \"fair\"," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14486" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.881, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.908, + 0.259 + ], + "angle": 0, + "content": "Figure 4. Framework of our DeQA-Score trained with soft label. For the “” token, the KL divergence loss is calculated between predicted probabilities and our soft label. For other tokens, common cross-entropy loss for next token prediction is calculated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.483, + 0.311 + ], + "angle": 0, + "content": "\"good\", \"excellent\". Our soft label, \\( p_i \\), assigns a ground-truth probability to each value. Let the predicted probabilities be \\( p_i^{pred} \\), the KL divergence loss is calculated through:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.322, + 0.483, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k l} = \\sum_ {i} p _ {i} \\log \\left(p _ {i} ^ {\\text {p r e d}} / p _ {i}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.446 + ], + "angle": 0, + "content": "The predicted probabilities, \\( p_i^{pred} \\), are obtained by applying a softmax activation function on model's output logits. Note that during training, this softmax function is applied across all textual tokens, not just across the five predefined levels. This approach encourages the model to reduce the possibilities of textual tokens outside the five levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.522 + ], + "angle": 0, + "content": "The two terms of losses, \\(\\mathcal{L}_{ce}\\) and \\(\\mathcal{L}_{kl}\\), are both autoregressive losses, and thus can be easily combined. During training, the cross-entropy loss, \\(\\mathcal{L}_{ce}\\), will decrease near to 0 fast, since it is quite easy to optimize by simply remembering the response template \"The quality of this image is\"." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.532, + 0.393, + 0.548 + ], + "angle": 0, + "content": "4.2. Score Estimation during Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.597 + ], + "angle": 0, + "content": "During inference, using the model-predicted probabilities of the five levels, \\( p_i^{pred} \\), we can estimate the expectation and variance of the score distribution the same as Eq. (3):" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.602, + 0.483, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mu^ {p r e d} = \\sum_ {i} p _ {i} ^ {p r e d} c _ {i}, (\\sigma^ {p r e d}) ^ {2} = \\sum_ {i} p _ {i} ^ {p r e d} (c _ {i} - \\mu^ {p r e d}) ^ {2}. (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Here we follow Q-Align [70] to apply a close-set softmax on the five levels to obtain the probabilities, \\( p_i^{pred} \\), to prevent the influence of other textual tokens. Our statistics in Tab. A2 reveal that the trained MLLM will always predict the five levels, thus applying softmax on either the five levels or all textual tokens yields nearly identical results. As given in Tab. 5, our predicted score distribution, \\( \\mathcal{N}(\\mu^{pred}, (\\sigma^{pred})^2) \\), aligns well with human annotations." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.44, + 0.775 + ], + "angle": 0, + "content": "4.3. Joint Training on Multiple IQA Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As demonstrated in [87, 98], different IQA datasets exhibit distinct distributions. Due to this dataset variation, images from different datasets with nearly identical scores (i.e., linearly re-scaled scores) can have drastically different perceptual quality (see examples in Fig. A1). The reason is that different datasets have different reflections between quality scores and visual quality. This disparity makes it infeasible to merely regress absolute quality scores for multi-dataset" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.263, + 0.895, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Figure 5. Fidelity loss to maintain intra-dataset relationships. Consider two images, Image A and B, sampled from the same dataset. Using our distribution-based soft label, DeQA-Score predicts the expectation, \\(\\mu^{pred}\\), and variance, \\((\\sigma^{pred})^2\\), of each image's score distribution, with which we can predict the probability that Image A is better than Image B, \\(P(A > B)\\). This predicted probability is then used to calculate the fidelity loss with the ground-truth probability. Fidelity loss preserves intra-dataset relationships, facilitating the co-training across multiple IQA datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.777 + ], + "angle": 0, + "content": "co-training. To address this challenge, UNIQUE [87] applies fidelity loss [63] based on Thurstone's model [60]. The fidelity loss is based on the principal that IQA emphasizes the ranking relationships within the same dataset, rather than the absolute values of quality scores. Specifically, two images are sampled from the same dataset, whose quality scores can be comparable. These two images are then compared using human-annotated quality scores. The fidelity loss encourages the model to learn such comparison results, which facilitates the model to focus more on the ranking relationships in each dataset, instead of the absolute values of quality scores. Thus, the fidelity loss helps multi-dataset co-training. Our soft label is naturally compatible with the fidelity loss, since our method can predict the score distribution, which is what the fidelity loss needs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.904 + ], + "angle": 0, + "content": "Ground-truth construction. As shown in Fig. 5, a pair of images, Image A and B, are sampled from the same dataset. We assume that the quality scores of the two images, \\( x_A, x_B \\), follow the Gaussian distribution with means and variances collected by subjective testing, i.e., \\( x_A \\sim \\mathcal{N}(\\mu_A, \\sigma_A^2) \\), \\( x_B \\sim \\mathcal{N}(\\mu_B, \\sigma_B^2) \\). Assuming that their quality scores are uncorrelated, the score difference, \\( x_A - x_B \\), also follows a Gaussian distribution, \\( \\mathcal{N}(\\mu_A - \\mu_B, \\sigma_A^2 + \\sigma_B^2) \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14487" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "Then the possibility that Image A has a better perceptual quality than Image B can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.13, + 0.483, + 0.171 + ], + "angle": 0, + "content": "\\[\np (A > B) = p \\left(x _ {A} > x _ {B}\\right) = \\Phi \\left(\\frac {\\mu_ {A} - \\mu_ {B}}{\\sqrt {\\sigma_ {A} ^ {2} + \\sigma_ {B} ^ {2}}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.18, + 0.483, + 0.195 + ], + "angle": 0, + "content": "where \\(\\Phi (\\cdot)\\) is the Gaussian cumulative distribution function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.484, + 0.289 + ], + "angle": 0, + "content": "Model prediction. As described in Sec. 4.2, our model can predict the expectation and variance of the score distribution. Denoting the predicted expectation and variance of two images as \\(\\mu_A^{pred}\\), \\((\\sigma_A^{pred})^2\\), \\(\\mu_B^{pred}\\), \\((\\sigma_B^{pred})^2\\), same to Eq. (6), we can predict the possibility of Image A surpassing Image B in perceptual quality as:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.297, + 0.483, + 0.348 + ], + "angle": 0, + "content": "\\[\np ^ {p r e d} (A > B) = \\Phi \\left(\\frac {\\mu_ {A} ^ {p r e d} - \\mu_ {B} ^ {p r e d}}{\\sqrt {(\\sigma_ {A} ^ {p r e d}) ^ {2} + (\\sigma_ {B} ^ {p r e d}) ^ {2}}}\\right). \\quad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.388 + ], + "angle": 0, + "content": "Training with fidelity loss. Following [87], we choose fidelity loss as the similarity measure to train the model as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.41, + 0.482, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {f d} = 1 - \\sqrt {p (A > B) p ^ {p r e d} (A > B)} \\\\ - \\sqrt {\\left(1 - p (A > B)\\right) \\left(1 - p ^ {p r e d} (A > B)\\right)}. \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.471, + 0.483, + 0.486 + ], + "angle": 0, + "content": "Adding two auto-regressive losses, the final training loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.501, + 0.482, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {f d} + \\gamma \\left(\\mathcal {L} _ {c e} + \\mathcal {L} _ {k l}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.522, + 0.433, + 0.537 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a weighting term to adjust the loss scale." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.55, + 0.224, + 0.567 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.435, + 0.59 + ], + "angle": 0, + "content": "5.1. Datasets, Baselines, Metrics, and Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Datasets. For a comprehensive evaluation, we include a wide range of IQA datasets: (a) in-the-wild images: KonIQ [21], SPAQ [17], LIVE-Wild [18], and FLIVE [78]; (b) synthetic distortions: KADID [31], TID2013 [47], and CSIQ [29]; (c) model-processed distortions: PIPAL [24]; (d) AI-generated images: AGIQA-3K [30]. Among these datasets, KonIQ, SPAQ, KADID, and PIPAL are split into training and test sets. Other datasets are used to assess the out-of-distribution generalization ability. The Mean Opinion Scores (MOS) of these datasets are normalized to the range of [1,5], with the variances normalized accordingly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Baseline methods. To demonstrate the effectiveness of our distribution-based soft label, we primarily compare with Q-Align [70], which utilizes the one-hot label. The base models for DeQA-Score and Q-Align are kept identical for fair comparison. We also include previous IQA methods as baselines, including handcrafted methods: NIQE [41] and BRISQUE [40], and deep-learning-based methods: NIMA [57], HyperIQA [55], DBCNN [86], MUSIQ [26], CLIP-IQA+ [64], ManIQA [73], and Compare2Score [98]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "Metrics. Following [26, 70, 73], we use the Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank-order Correlation Coefficient (SRCC) as metrics to evaluate the score regression performance. Additionally, we employ KL divergence, JS divergence, and Wasserstein distance to assess the results of score distribution prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Implementation details. We use mPLUG-Owl2 [75] as our base model, with a CLIP-pretrained ViT-L [49] as the image encoder, a six-layer Q-Former as the visual abstractor, and LLaMA-2-7B [62] as the LLM. The number of visual tokens output by the vision abstractor is set to 64, and the hidden dimension of LLM is 4096. The pretrained weights of mPLUG-Owl2 are used for model initialization. The weighting term \\(\\gamma\\) is empirically set to 0.05, adjusting the two loss terms roughly to the same scale. We adopt AdamW [23] as the optimizer. The initial learning rate is set to 2e-5 and decays gradually using the cosine decay strategy. Our model is trained with a batch size of 64 for 3 epochs. Using 8 NVIDIA RTX A6000 GPUs, the training process on KonIQ dataset is completed in around 1.5 hours." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.762, + 0.423 + ], + "angle": 0, + "content": "5.2. Results of Score Regression" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.906, + 0.61 + ], + "angle": 0, + "content": "Score regression results of single-dataset training. We first train our model on a single dataset, i.e., KonIQ, and then evaluate the model on both in-distribution and out-of-distribution datasets, following [28, 70]. The score regression results are shown in Tab. 3. On the in-distribution KonIQ test set, our model achieves superior score regression results compared to all baseline methods, e.g., a \\(1.3\\%\\) improvement in PLCC over Q-Align. Furthermore, on out-of-distribution datasets, our DeQA-Score maintains its advantage with the best overall results across nearly all benchmarks, e.g., a \\(4.6\\%\\) improvement in PLCC on LIVE-Wild. These findings confirm the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.732 + ], + "angle": 0, + "content": "However, we observe that all models perform poorly on the PIPAL dataset. For instance, our best results reach only 0.462 on PLCC. The main reason is the distortion gap between the training dataset, KonIQ (in-the-wild distortions), and PIPAL (model-processed distortions). This highlights the necessity of co-training with multiple IQA datasets containing various distortion types. This observation motivates our subsequent experiments on multi-dataset co-training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Score regression results of multi-dataset co-training. As discussed above, co-training on multiple IQA datasets is necessary to obtain a unified model for various categories of distortions. Therefore, we conduct multi-dataset cotraining experiments in Tab. 4. First, our DeQA-Score consistently outperforms Q-Align across different training combinations. The performance gain attributes to both the soft label and fidelity loss, as confirmed in Tab. 6. Second, comparing #0 to #1, when KADID (synthetic distortions) is added to the training set, the performance on TID2013 and CSIQ (both synthetic distortions) improves" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14488" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 3. Score regression results of single-dataset training with the PLCC / SRCC metrics. All methods except handcrafted ones are trained on the KonIQ dataset. DeQA-Score outperforms all baseline methods across nearly all benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.122, + 0.907, + 0.294 + ], + "angle": 0, + "content": "
CategoryMethodsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KCSIQFLIVE
HandcraftedNIQE [41]0.533 / 0.5300.679 / 0.6640.468 / 0.4050.195 / 0.1610.493 / 0.4490.560 / 0.5330.718 / 0.6280.147 / 0.100
BRISQUE [40]0.225 / 0.2260.490 / 0.4060.429 / 0.3560.267 / 0.2320.361 / 0.3130.541 / 0.4970.740 / 0.5560.108 / 0.054
Non-MLLMNIMA [57]0.896 / 0.8590.838 / 0.8560.532 / 0.5350.390 / 0.3990.814 / 0.7710.715 / 0.6540.695 / 0.6490.561 / 0.467
HyperIQA [55]0.917 / 0.9060.791 / 0.7880.506 / 0.4680.410 / 0.4030.772 / 0.7490.702 / 0.6400.752 / 0.7170.485 / 0.383
DBCNN [86]0.884 / 0.8750.812 / 0.8060.497 / 0.4840.384 / 0.3810.773 / 0.7550.730 / 0.6410.586 / 0.5720.485 / 0.385
MUSIQ [26]0.924 / 0.9290.868 / 0.8630.575 / 0.5560.431 / 0.4310.789 / 0.8300.722 / 0.6300.771 / 0.7100.565 / 0.467
CLIP-IQA+ [64]0.909 / 0.8950.866 / 0.8640.653 / 0.6540.427 / 0.4190.832 / 0.8050.736 / 0.6850.772 / 0.7190.427 / 0.316
ManIQA [73]0.849 / 0.8340.768 / 0.7580.499 / 0.4650.457 / 0.4520.849 / 0.8320.723 / 0.6360.623 / 0.6270.512 / 0.401
MLLM-basedCompare2Score [98]0.923 / 0.9100.867 / 0.8600.500 / 0.4530.354 / 0.3420.786 / 0.7720.777 / 0.6710.735 / 0.7050.474 / 0.413
Q-Align [70]0.941 / 0.9400.886 / 0.8870.674 / 0.6840.403 / 0.4190.853 / 0.8600.772 / 0.7350.785 / 0.7370.554 / 0.483
DeQA-Score0.953 / 0.9410.895 / 0.8960.694 / 0.6870.472 / 0.4780.892 / 0.8790.809 / 0.7290.787 / 0.7440.589 / 0.501
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.298, + 0.907, + 0.328 + ], + "angle": 0, + "content": "Table 4. Score regression results of co-training on multiple IQA datasets with the PLCC / SRCC metrics. Our DeQA-Score consistently surpasses the baseline method, Q-Align, across all benchmarks with different training combinations." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.33, + 0.905, + 0.442 + ], + "angle": 0, + "content": "
#Trained onMethodsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
0KonIQ, SPAQQ-Align [70]0.943 / 0.9400.933 / 0.9310.692 / 0.7080.401 / 0.4110.883 / 0.8790.795 / 0.7270.671 / 0.5680.795 / 0.767
DeQA-Score0.953 / 0.9430.936 / 0.9330.724 / 0.7190.468 / 0.4740.902 / 0.8880.810 / 0.7380.706 / 0.6160.836 / 0.785
1KonIQ, SPAQ, KADIDQ-Align [70]0.945 / 0.9380.933 / 0.9310.935 / 0.9340.409 / 0.4200.887 / 0.8830.788 / 0.7330.829 / 0.8080.876 / 0.845
DeQA-Score0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
2KonIQ, SPAQ, KADID, PIPALQ-Align [70]0.926 / 0.9320.917 / 0.9200.950 / 0.9540.702 / 0.6710.853 / 0.8450.765 / 0.7220.811 / 0.7950.838 / 0.789
DeQA-Score0.958 / 0.9460.932 / 0.9290.963 / 0.9610.724 / 0.6900.877 / 0.8570.770 / 0.7350.828 / 0.7960.863 / 0.807
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.518 + ], + "angle": 0, + "content": "Table 5. Score distribution prediction results with JS divergence / Wasserstein distance between the predicted Gaussian distribution and the human labeled Gaussian distribution as metrics. Models are trained on the KonIQ, SPAQ, and KADID datasets. DeQA-Score achieves a much closer alignment with human annotations." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.52, + 0.481, + 0.62 + ], + "angle": 0, + "content": "
DatasetsQ-Align [70]DeQA-Score (Ours)
KonIQ0.415 / 0.6110.014 / 0.186
SPAQ0.192 / 0.4550.070 / 0.332
KADID0.519 / 0.7230.029 / 0.304
LIVE-Wild0.254 / 0.6420.052 / 0.417
AGIQA-3K0.558 / 0.7300.177 / 0.592
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.626, + 0.483, + 0.808 + ], + "angle": 0, + "content": "significantly (e.g., \\(33.1\\%\\) in SRCC on TID2013), underscoring that distortion type is a major gap between training and testing. Third, from #1 to #2, adding PIPAL for training greatly enhances the results on PIPAL (from 0.495 to 0.724 in PLCC) but reduces the performance on several other datasets, e.g., from 0.900 to 0.863 in PLCC on CSIQ. This performance decline is likely due to the unique nature of model-processed distortions in PIPAL, which differ greatly from other distortions. Finally, despite never trained on AI-generated images, DeQA-Score performs well on AGIQA-3K with around 0.8 PLCC, reflecting the generalization ability enabled by MLLMs' broad pre-training." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.81, + 0.466, + 0.825 + ], + "angle": 0, + "content": "More results on AIGC images are provided in Tab. A4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.433, + 0.849 + ], + "angle": 0, + "content": "5.3. Results of Score Distribution Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As stated in Sec. 4.2, the MLLM-output probabilities allow us to predict a score distribution, which can be evaluated against the human labeled ground-truth distribution. We" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.542 + ], + "angle": 0, + "content": "calculate the distance between these two distributions and present the results in Tab. 5. The results illustrate that our predicted score distribution aligns much better with humans than Q-Align. As shown in Fig. 6, Q-Align tends to predict a single level token, bringing a quite small predicted variance, which compromises the distribution metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.909, + 0.698 + ], + "angle": 0, + "content": "Qualitative results are shown in Fig. 6. Across different image types: in-the-wild images (Fig. 6ab), AI-generated images (Fig. 6c), or synthetic distortions (Fig. 6d), and for varying quality levels: high-quality (Fig. 6ac), fair-quality (Fig. 6b), or poor-quality images (Fig. 6d), our method consistently predicts accurate quality scores. Additionally, our DeQA-Score provides a score distribution that aligns more closely with humans with smaller KL divergence. In contrast, Q-Align, which is trained with one-hot labels, tends to predict a single label, resulting in higher KL divergence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.68, + 0.726 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of soft label and fidelity loss is studied and shown in Tab. 6. In #1, \"Uncertainty\" is utilized to verify the effects of adding annotation variances without considering the score distortion. We directly append uncertainty within the responses, i.e., \"The quality of this image is with uncertainty\". The has five possible values: {\"minimal\", \"low\", \"moderate\", \"high\", \"severe\"}, derived by the one-hot discretization. In #2, model trained on one-hot label can not predict a reliable variance, thus we employ binary fidelity loss [89], which does not require predicted variance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14489" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.091, + 0.245, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.176, + 0.228, + 0.188 + ], + "angle": 0, + "content": "(a) Score: \\(3.93 \\pm 0.79\\)" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.092, + 0.375, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.092, + 0.505, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.092, + 0.622, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.092, + 0.752, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.092, + 0.882, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.192, + 0.245, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.265, + 0.227, + 0.276 + ], + "angle": 0, + "content": "(b) Score: \\(3.11 \\pm 0.66\\)" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.194, + 0.375, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.266, + 0.368, + 0.275 + ], + "angle": 0, + "content": "Pred: 3.95, KL Div: 9.01" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.194, + 0.505, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.266, + 0.498, + 0.275 + ], + "angle": 0, + "content": "Pred: 3.34, KL Div: 0.06" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.192, + 0.622, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.265, + 0.617, + 0.275 + ], + "angle": 0, + "content": "(d) Score: \\(1.92 \\pm 0.82\\)" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.194, + 0.752, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.265, + 0.747, + 0.275 + ], + "angle": 0, + "content": "Pred: 1.04, KL Div: 16.69" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.192, + 0.882, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.265, + 0.872, + 0.275 + ], + "angle": 0, + "content": "Pred: 1.76, KL Div: 0.14" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.28, + 0.905, + 0.322 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on KonIQ, SPAQ, AGIQA-3K, and KADID datasets. DeQA-Score predicts more accurate quality scores with more reasonable logits. Also, our method generates a score distribution that aligns better with humans, i.e., smaller KL divergence. However, Q-Align, which is trained on one-hot label, is more likely to predict a single label, resulting in higher KL divergence." + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.327, + 0.899, + 0.341 + ], + "angle": 0, + "content": "Table 6. Ablation studies on soft label and fidelity loss with PLCC / SRCC metrics. Models are trained on KonIQ, SPAQ, and KADID." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.345, + 0.901, + 0.432 + ], + "angle": 0, + "content": "
#One-hotSoftUncertaintyFidelityKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
0✓0.945 / 0.9380.933 / 0.9310.935 / 0.9340.409 / 0.4200.887 / 0.8830.788 / 0.7330.829 / 0.8080.876 / 0.845
1✓✓0.935 / 0.9360.933 / 0.9310.954 / 0.9510.439 / 0.4530.896 / 0.8870.764 / 0.7150.826 / 0.8030.884 / 0.849
2✓✓0.944 / 0.9360.935 / 0.9320.951 / 0.9500.473 / 0.4780.898 / 0.8860.788 / 0.7260.837 / 0.8100.876 / 0.839
3✓0.954 / 0.9430.937 / 0.9340.951 / 0.9470.472 / 0.4720.896 / 0.8790.791 / 0.7290.852 / 0.8210.894 / 0.855
4✓✓0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.456, + 0.901, + 0.556 + ], + "angle": 0, + "content": "
CategoryLevel TextsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
Commonexce. / good / far / poor / bad0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
Randomexce. / poor / bad / good / fair0.942 / 0.9280.933 / 0.9290.950 / 0.9470.394 / 0.3960.863 / 0.8510.795 / 0.7270.814 / 0.7920.887 / 0.836
Reversebad / poor / fair / good / exce.0.943 / 0.9340.934 / 0.9310.947 / 0.9450.395 / 0.3860.836 / 0.8000.782 / 0.7140.811 / 0.7910.882 / 0.834
Numberone / two / three / four / five0.945 / 0.9320.934 / 0.9300.955 / 0.9520.421 / 0.4240.873 / 0.8570.766 / 0.7120.820 / 0.7940.882 / 0.835
Vehiclecar / bus / train / plane / ship0.953 / 0.9400.937 / 0.9330.954 / 0.9510.458 / 0.4560.889 / 0.8780.778 / 0.7020.854 / 0.8280.897 / 0.856
Mixtureapple / car / four / fair / dog0.947 / 0.9380.936 / 0.9320.954 / 0.9510.419 / 0.4150.865 / 0.8510.808 / 0.7370.827 / 0.8000.896 / 0.854
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.438, + 0.888, + 0.453 + ], + "angle": 0, + "content": "Table 7. Effects of different level texts with PLCC / SRCC metrics. Models are trained on the KonIQ, SPAQ, and KADID datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.564, + 0.48, + 0.578 + ], + "angle": 0, + "content": "Table 8. Inference latency & throughput on RTX A6000 GPU." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.58, + 0.482, + 0.634 + ], + "angle": 0, + "content": "
Batch Size12481664256
Latency (s)0.0890.1210.1820.3150.5882.1147.929
Throughput (image/s)11.2416.5321.9825.4027.2130.2732.29
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Several conclusions can be drawn from Tab. 6. First, comparing #0 with #3, our soft label significantly outperforms the one-hot label across all benchmarks, validating its effectiveness. Second, from #3 to #4, the performance of our soft label further improves when combined with the fidelity loss. Third, comparing #0 with #2, the one-hot label does not incorporate well with the fidelity loss, as it lacks score distribution information. Finally, from #0 to #1, simply adding uncertainty without considering score distribution does not yield stable or significant improvements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Effects of different level texts are studied in Tab. 7. Beside the commonly used level texts, we consider five alternatives: (a) random sorted common texts; (b) reversed common texts; (c) simple numbers like one, two, etc.; (d) irrelevant words about vehicle; (e) a mixture of various types. The results indicate that common level texts achieve the best performance, as they align closely with the MLLMs' pre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.645 + ], + "angle": 0, + "content": "training. Interestingly, with random sorted / reversed texts, even a mixture of various level texts, which are inconsistent with MLLMs' prior knowledge, our model still learns these newly defined relationships and performs well. This shows the robust adaptability of MLLM-based IQA models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.905, + 0.679 + ], + "angle": 0, + "content": "Ablation studies on level number and training / fixing components are provided in Tabs. A6 and A7, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.688, + 0.689, + 0.704 + ], + "angle": 0, + "content": "5.5. Inference Latency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.905, + 0.787 + ], + "angle": 0, + "content": "We test the inference latency on one single RTX A6000 GPU. As shown in Tab. 8, with the optimal GPU utilization (i.e., batch size 256), our DeQA-Score can process as many as 32.29 images per second in average. This high throughput emphasizes its suitability for real-time deployment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.64, + 0.815 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We develop DeQA-Score as a pioneering MLLM-based image quality scorer, empowered by distribution-based soft labels and fidelity loss, capable of both score regression and score distribution prediction, highlighting the importance of quality score distribution in MLLM-based IQA methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14490" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgement. This work was sponsored by RGC Early Career Scheme (ECS) No. 24209224, National Natural Science Foundation of China (Grant No.62276251), and the Joint Lab of CAS-HK." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.166, + 0.188, + 0.182 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.262, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.32, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[4] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE TIP, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[5] Xin Cai, Zhiyuan You, Hailong Zhang, Wentao Liu, Jinwei Gu, and Tianfan Xue. PhoCoLens: Photorealistic and consistent reconstruction in lensless imaging. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[6] Yue Cao, Zhaolin Wan, Dongwei Ren, Zifei Yan, and Wangmeng Zuo. Incorporating semi-supervised and positive unlabeled learning for boosting full reference image quality assessment. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[7] Chaofeng Chen, Yang Sensen, Haoning Wu, Liang Liao, Zicheng Zhang, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-Ground: Image quality grounding with large multi-modality models. In ACM MM, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[8] Du Chen, Tianhe Wu, Kede Ma, and Lei Zhang. Toward generalized image quality assessment: Relaxing the perfect reference quality assumption. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[10] Zixuan Chen, Zewei He, Ziqian Lu, Xuecheng Sun, and Zhe-Ming Lu. Prompt-based test-time real image dehazing: a novel pipeline. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[11] Zijian Chen, Wei Sun, Haoning Wu, Zicheng Zhang, Jun Jia, Ru Huang, Xiongkuo Min, Guangtao Zhai, and Wenjun Zhang. Study of subjective and objective naturalness assessment of ai-generated images. IEEE TCSVT, 2024. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[12] Zixuan Chen, Yujin Wang, Xin Cai, Zhiyuan You, Zheming Lu, Fan Zhang, Shi Guo, and Tianfan Xue. UltraFusion: Ultra high dynamic imaging using exposure fusion. In CVPR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.192, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.134 + ], + "angle": 0, + "content": "hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. https://vicuna.lmsys.org, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE TPAMI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.252, + 0.904, + 0.293 + ], + "angle": 0, + "content": "[16] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In ACM MM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[17] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, 2020. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.34, + 0.904, + 0.393 + ], + "angle": 0, + "content": "[18] Deepti Ghadiyaram and C. Bovik Alan. Live in the wild image quality challenge database. http://live.ece.utexas.edu/research/ChallengeDB/index.html, 2015.2,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.397, + 0.904, + 0.424 + ], + "angle": 0, + "content": "[19] Abhijay Ghildyal and Feng Liu. Shift-tolerant perceptual similarity metric. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.427, + 0.904, + 0.481 + ], + "angle": 0, + "content": "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.485, + 0.904, + 0.538 + ], + "angle": 0, + "content": "[21] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. KonIQ-10K: An ecologically valid database for deep learning of blind image quality assessment. IEEE TIP, 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.542, + 0.905, + 0.596 + ], + "angle": 0, + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2021. 13, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.6, + 0.904, + 0.626 + ], + "angle": 0, + "content": "[23] Loshchilov Ilya and Hutter Frank. Decoupled weight decay regularization. In ICLR, 2019. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.63, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[24] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. PIPAL: a large-scale image quality assessment dataset for perceptual image restoration. In ECCV, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.687, + 0.904, + 0.727 + ], + "angle": 0, + "content": "[25] Le Kang, Peng Ye, Yi Li, and David Doermann. Convolutional neural networks for no-reference image quality assessment. In CVPR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[26] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality transformer. In CVPR, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[27] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 2018. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Daekyu Kwon, Dongyoung Kim, Sehwan Ki, Younghyun Jo, Hyong-Euk Lee, and Seon Joo Kim. ATTIQA: Generalizable image quality feature extractor using attribute-aware pretraining. In ACCV, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14491" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] Eric C Larson and Damon M Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[30] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. AGIQA-3K: An open database for ai-generated image quality assessment. IEEE TCSVT, 2023. 1, 2, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[31] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. KADID-10K: A large-scale artificially distorted iqa database. In International Conference on Quality of Multimedia Experience (QoMEX), 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[33] Kai Liu, Ziqing Zhang, Wenbo Li, Renjing Pei, Fenglong Song, Xiaohong Liu, Linghe Kong, and Yulun Zhang. Dog-IQA: Standard-guided zero-shot mllm for mix-grained image quality assessment. arXiv preprint arXiv:2410.02505, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[34] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. RankIQA: Learning from rankings for no-reference image quality assessment. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.458 + ], + "angle": 0, + "content": "[35] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[36] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[37] Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[38] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In ACL Findings, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[39] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. DocVQA: A dataset for vqa on document images. In WACV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[40] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE TIP, 2012. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[41] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Sign. Process. Letters, 2012. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.772, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[42] Anush Krishna Moorthy and Alan Conrad Bovik. A two-step framework for constructing blind image quality indices. IEEE Sign. Process. Letters, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[43] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE TIP, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.858, + 0.414, + 0.871 + ], + "angle": 0, + "content": "[44] OpenAI. GPT-4V(ision) system card, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[46] Da Pan, Ping Shi, Ming Hou, Zefeng Ying, Sizhe Fu, and Yuan Zhang. Blind predicting similar quality map for image quality assessment. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[47] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database TID2013: Peculiarities, results and perspectives. Signal processing: Image communication, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[48] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. PieAPP: Perceptual image-error assessment through pairwise preference. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.904, + 0.402 + ], + "angle": 0, + "content": "[50] Michele A Saad, Alan C Bovik, and Christophe Charrier. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE TIP, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[51] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[52] H Sheikh. LIVE image quality assessment database release 2. http://live.ece.utexas.edu/research/quality, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[53] Hamid R Sheikh and Alan C Bovik. Image information and visual quality. IEEE TIP, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[54] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA models that can read. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[55] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In CVPR, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[56] Simen Sun, Tao Yu, Jiahua Xu, Wei Zhou, and Zhibo Chen. GraphIQA: Learning distortion graph representations for blind image quality assessment. IEEE TMM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[57] Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment. IEEE TIP, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[58] Huixuan Tang, Neel Joshi, and Ashish Kapoor. Learning a blind measure of perceptual image quality. In CVPR, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[59] InternLM Team. InternLM: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[60] Louis L Thurstone. A law of comparative judgment. Psychological review, 1994. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[62] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14492" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. \nLLaMA 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[63] Ming-Feng Tsai, Tie-Yan Liu, Tao Qin, Hsin-Hsi Chen, and Wei-Ying Ma. FRank: A ranking method with fidelity loss. In International ACM SIGIR Conference on Research and Development in Information Retrieval, 2007. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.195, + 0.482, + 0.236 + ], + "angle": 0, + "content": "[64] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.322 + ], + "angle": 0, + "content": "[65] Jiarui Wang, Huiyu Duan, Jing Liu, Shi Chen, Xiongkuo Min, and Guangtao Zhai. AIGCIQA2023: A large-scale image quality assessment database for ai generated images: from the perspectives of quality, authenticity and correspondence. In CAAI International Conference on Artificial Intelligence, 2023. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.325, + 0.482, + 0.367 + ], + "angle": 0, + "content": "[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.369, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[67] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv preprint arXiv:2312.06109, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[68] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-Bench: A benchmark for general-purpose foundation models on low-level vision. In ICLR, 2024. 3, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[69] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-Instruct: Improving low-level visual abilities for multi-modality foundation models. In CVPR, 2024. 1, 2, 3, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[70] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-Align: Teaching LMMs for visual scoring via discrete text-defined levels. In ICML, 2024. 1, 2, 3, 4, 5, 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[71] Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, et al. Towards open-ended visual quality comparison. In ECCV, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[72] Tianhe Wu, Kede Ma, Jie Liang, Yujiu Yang, and Lei Zhang. A comprehensive study of multimodal large language models for image quality assessment. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[73] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. MANIQA: Multi-dimension attention network for no-reference image quality assessment. In CVPRW, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[74] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mPLUG-Owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[75] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mPLUG-Owl2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 1, 4, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[76] Guanghao Yin, Wei Wang, Zehuan Yuan, Chuchu Han, Wei Ji, Shouqian Sun, and Changhu Wang. Content-variant reference image quality assessment via knowledge distillation. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[77] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Lu Sheng, Lei Bai, Xiaoshui Huang, Zhiyong Wang, et al. Lamm: Language-assisted multimodal instruction-tuning dataset, framework, and benchmark. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[78] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (PaQ-2-PiQ): Mapping the perceptual space of picture quality. In CVPR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[79] Zhiyuan You, Jinjin Gu, Zheyuan Li, Xin Cai, Kaiwen Zhu, Chao Dong, and Tianfan Xue. Descriptive image quality assessment in the wild. arXiv preprint arXiv:2405.18842, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[80] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. In ECCV, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[81] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[82] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. FSIM: A feature similarity index for image quality assessment. IEEE TIP, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[83] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[85] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. LLaMA-Adapter: Efficient fine-tuning of large language models with zero-initialized attention. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[86] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE TCSVT, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[87] Weixia Zhang, Kede Ma, Guangtao Zhai, and Xiaokang Yang. Uncertainty-aware blind image quality assessment in the laboratory and wild. IEEE TIP, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[88] Weixia Zhang, Dingquan Li, Chao Ma, Guangtao Zhai, Xiaokang Yang, and Kede Ma. Continual learning for blind image quality assessment. IEEE TPAMI, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14493" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[89] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In CVPR, 2023. 2, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[90] Zicheng Zhang, Chunyi Li, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. A perceptual quality assessment exploration for aigc images. In ICME Workshops, 2023. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[91] Zicheng Zhang, Haoning Wu, Erli Zhang, Guangtao Zhai, and Weisi Lin. A benchmark for multi-modal foundation models on low-level vision: from single images to pairs. IEEE TPAMI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[92] Zicheng Zhang, Yingjie Zhou, Chunyi Li, Baixuan Zhao, Xiaohong Liu, and Guangtao Zhai. Quality assessment in the era of large models: A survey. arXiv preprint arXiv:2409.00031, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[93] Heliang Zheng, Jianlong Fu, Yanhong Zeng, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[94] Wei Zhou and Zhou Wang. Quality assessment of image super-resolution: Balancing deterministic and statistical fidelity. In ACM MM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[95] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[96] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. MetaQA: deep meta-learning for no-reference image quality assessment. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[97] Hanwei Zhu, Xiangjie Sui, Baoliang Chen, Xuelin Liu, Peilin Chen, Yuming Fang, and Shiqi Wang. 2afc prompting of large multimodal models for image quality assessment. arXiv preprint arXiv:2402.01162, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[98] Hanwei Zhu, Haoning Wu, Yixuan Li, Zicheng Zhang, Bao-liang Chen, Lingyu Zhu, Yuming Fang, Guangtao Zhai, Weisi Lin, and Shiqi Wang. Adaptive image quality assessment via teaching large multimodal model to compare. arXiv preprint arXiv:2405.19298, 2024. 2, 3, 5, 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[99] Kaiwen Zhu, Jinjin Gu, Zhiyuan You, Yu Qiao, and Chao Dong. An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14494" + } + ] +] \ No newline at end of file diff --git a/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_origin.pdf b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b11e759c9622c13f798434c6b2faa08e4027a95 --- /dev/null +++ b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/6e8f8571-95f4-493d-9b70-2616ff293e01_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a40f8c217104797a7275d9360da0fb215c5d39b4f88ac108702888ad5101ac85 +size 876832 diff --git a/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/full.md b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ea5868ee10c960fedafef93b97babbb48578fcd0 --- /dev/null +++ b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/full.md @@ -0,0 +1,442 @@ +# Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution + +Zhiyuan You $^{12}$ , Xin Cai $^{2}$ , Jinjin Gu $^{4}$ , Tianfan Xue $^{23\dagger}$ , Chao Dong $^{134\dagger}$ + +1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences + +2The Chinese University of Hong Kong, 3Shanghai AI Laboratory + +4Shenzhen University of Advanced Technology + +zhiyuanyou@link.cuhk.edu.hk, caixin@link.cuhk.edu.hk, jinjin.gu@sydney.edu.au + +tfxue@ie.cuhk.edu.hk, chao.dong@siat.ac.cn † Corresponding Author + +# Abstract + +With the rapid advancement of Multi-modal Large Language Models (MLLMs), MLLM-based Image Quality Assessment (IQA) methods have shown promising performance in linguistic quality description. However, current methods still fall short in accurately scoring image quality. In this work, we aim to leverage MLLMs to regress accurate quality scores. A key challenge is that the quality score is inherently continuous, typically modeled as a Gaussian distribution, whereas MLLMs generate discrete token outputs. This mismatch necessitates score discretization. Previous approaches discretize the mean score into a one-hot label, resulting in information loss and failing to capture inter-image relationships. We propose a distribution-based approach that discretizes the score distribution into a soft label. This method preserves the characteristics of the score distribution, achieving high accuracy and maintaining inter-image relationships. Moreover, to address dataset variation, where different IQA datasets exhibit various distributions, we introduce a fidelity loss based on Thurstone's model. This loss captures intra-dataset relationships, facilitating co-training across multiple IQA datasets. With these designs, we develop the distribution-based Depicted image Quality Assessment model for Score regression (DeQA-Score). Experiments across multiple benchmarks show that DeQA-Score stably outperforms baselines in score regression. Also, DeQA-Score can predict the score distribution that closely aligns with human annotations. Codes and model weights have been released in https://depictqa.github.io/deqa-score/. + +# 1. Introduction + +Image Quality Assessment (IQA) focuses on perceiving the visual quality of images [26, 41, 66, 84], and is widely adopted in various application scenarios, e.g., image com + +![](images/9c3d5035d233ce48d76be96cbb7e2a3f6b830219e529a2f8c65952b102abf3e5.jpg) + +![](images/bc8d52c5ed98902340f73dffdd37719e2c097a2b605a4a9483a026b1587f3ad4.jpg) + +![](images/115dd61a17c167a6b0fa757ebf7934db7d109da1216511689698f7c84fd62d72.jpg) + +![](images/b7c00aeb46368cd6361ea1728f51370187a326ab8d512615c08d2b3733d753bd.jpg) + +![](images/68cb15a2ddd2e1f44ca67c4db6f133fe9877e832352fd6ea2957f4780c657fbb.jpg) + +![](images/d80734b4ead7143813d3aa8d0a8a2ec9d4eacd184ffb73d35b65b5fdd5e9ca17.jpg) + +![](images/759e462d9e30c7c8dd92e2e2369bbee0dc69ade76285eb83e82af4777da3ed09.jpg) +Figure 1. Illustration of our distribution-based soft label. Image A is overexposed, while Image B and C are well-exposed with better and similar quality. "exce." means "excellent". To train an MLLM as a quality scorer, continuous scores need to be discretized to discrete level tokens as the training label. Q-Align [70] discretizes human-labeled mean score to a one-hot label, causing information loss. We discretize the score distribution, which is approximated as a Gaussian distribution, to obtain a more accurate soft label. Our soft label maintains better inter-image relationships that the Image B is better than A (A < B), and Image B and C share similar quality (B ≈ C), while the one-hot label can not. + +![](images/23c670d66a7d30f9dc0cadb426387dd71bda4645f463c8d68283befc68fea9dd.jpg) + +![](images/3350e31f5688ab3c66084a6c2ca00e59dd95afa04748f6715069c00ad613a883.jpg) + +pression and transmission [5, 52], processing smartphone-captured photos [17], and evaluating generation or restoration models [10, 12, 30, 99], etc. With the emergence of Multi-modal Large Language Models (MLLMs) [32, 44, 75, 95], MLLM-based IQA begins to attract more research interests [7, 69, 71, 79, 80]. These methods mainly leverage MLLMs to describe image quality using language, recognizing that language better mirrors human expression, and captures the multifaceted nature of IQA tasks [80]. + +However, for quality scoring, which needs an accurate numerical score to quantify an image's quality, existing description-centered MLLMs [69, 71, 79, 80] perform less accurately than traditional IQA methods. This hinders their applications on real-world IQA scenarios. Considering MLLMs' good generalization ability on various downstream tasks and promising performance on quality description, it is possible for MLLMs to score the image quality. Therefore, in this work, we aim to explore how MLLMs can effectively regress accurate quality scores by leveraging their broad training and inherent adaptability. + +We find that the principal challenge in training MLLMs to predict accurate quality scores is the distribution gap between MLLMs' outputs and quality scores. MLLMs generate outputs as discrete tokens, while the quality score is inherently continuous, usually approximated to a Gaussian distribution. Previous approaches, like Q-Align [70], attempt to address the distribution gap by directly discretizing the mean score (i.e., the expectation of estimated score distribution) into a one-hot label, as depicted in Fig. 1. However, such a one-hot discretization method introduces information loss, causing mismatches between the discretized labels and the original continuous scores. For instance, in Fig. 1, Image A is overexposed with a lower mean score than Image B, but their one-hot labels appear identical. Conversely, Image B and C are well-exposed with similar quality scores, yet their discretized labels are different. + +To overcome the distribution gap, we propose a novel soft label to bridge continuous and discrete distributions. As illustrated in Fig. 1, instead of discretizing the mean of score distribution, we discretize the continuous score distribution itself into a discrete distribution of different levels. During training, this discrete distribution is used as the soft training label. This enables effective training with minimal information loss. Also, our distribution-based soft label better preserves inter-image relationship. For example, in Fig. 1, Image B is better than Image A, and Image B and C share similar quality, while the simple one-hot label fails to capture such relationships. At inference, MLLMs output discrete token results, which can be recovered to continuous score distribution. This method ensures that the discrete outputs align well with the underlying continuous distribution, allowing to accurately predict the desired score. + +Another key challenge is the dataset variation, where different IQA datasets exhibit distinct distributions, making it hard to simply combine them for co-training. Due to this variation, images from different datasets may have nearly the same quality scores (i.e., linearly re-scaled scores) but drastically different perceptual quality [87, 98]. A previous method, UNIQUE [87], has demonstrated that the fidelity loss [63] under the Thurstone's model [60] can effectively capture the intra-dataset relationships, thus facilitating the joint training of multiple IQA datasets. However, + +the fidelity loss requires the model to predict an accurate score distribution, which MLLMs trained with the one-hot label cannot provide. Fortunately, MLLMs trained with our distribution-based soft label can predict such an accurate score distribution, allowing the adaptation of the fidelity loss. We apply this fidelity loss to our MLLM-based quality scorer and demonstrate its effectiveness in this new context. + +Equipped with these designs, we successfully develop a distribution-based multi-modal image quality assessment model, termed DeQA-Score. DeQA-Score achieves better score regression results than baseline methods on both in-distribution (e.g., $1.3\%$ in PLCC on KonIQ [21]) and out-of-distribution datasets (e.g., $4.6\%$ in PLCC on LIVEWILD [18]). The advantage of our DeQA-Score becomes more pronounced with the co-training across multiple IQA datasets. Moreover, our method can further predict the distribution of quality scores, which is closely aligned with the distribution derived from extensive human annotations. We hope that our DeQA-Score will inspire more research to explore the potential of quality score distribution. + +# 2. Related Works + +Image Quality Assessment (IQA) methods mainly rely on quality scores to assess image quality. They can be categorized into full-reference and non-reference methods through whether there is a high-quality reference image. + +Full-reference IQA methods compute a similarity score as the quality score between a distorted image and a high-quality reference. Classical works rely on human-designed metrics such as structural similarity [66], phase congruency with gradient magnitude [82], image information [53], etc. The rapid advancement of deep learning has also inspired learning-based IQA methods that measure image quality through data-driven training. Pioneered by LPIPS [84] and PieAPP [48], driven by large-scale IQA datasets [24, 31, 52, 84], deep-learning methods [4, 6, 8, 15, 16, 19, 76, 94] have exhibited quite high accuracy in quality score regression. + +Non-reference IQA methods directly regress a quality score without a reference image. Initially, hand-crafted natural image statistics are adopted [37, 40-43, 50, 58]. Subsequently, deep-learning-based methods [25, 34, 46, 55, 56, 78, 88, 93, 96] replace hand-crafted statistics by learning quality priors from human-annotated IQA datasets [17, 21, 30, 31]. Recent works further improve the regression accuracy by introducing multi-dataset co-training strategy [87], multi-scale features [26], CLIP pre-training [64], multidimension attention [73], multitask learning [89], and so on. + +Multi-modal Large Language Models (MLLMs) incorporate visual modality into Large Language Models (LLMs) [1, 13, 59, 61] to achieve general visual ability by leveraging the broad training and powerful generalization ability of LLMs. Existing MLLMs [3, 14, 32, 44, 67, 74, 77, 83, 85, 95] have demonstrated a general visual ability + +and can tackle a variety of multi-modality tasks, like visual question answering [20, 35, 36], document understanding [38, 39, 54], image captioning [2, 9, 81], etc. Given MLLMs' advanced performance in these high-level perception tasks, it is natural to explore how MLLMs perform in low-level perception tasks that are highly related to IQA. + +MLLM-based IQA methods leverages MLLMs' foundational knowledge to achieve better IQA performance or more detailed assessment results [72, 92]. Q-Bench [68, 91] shows that general MLLMs have some low-level perception ability. [97] evaluates various MLLMs on the widely-adopted Two-Alternative Forced Choice (2AFC) task. Q-Instruct [69] enhances the low-level perception ability of MLLMs by introducing a large-scale dataset. CoInstruct [71] concentrates on the quality comparison among multiple images. DepictQA [79, 80] performs single-image assessment and paired-image comparison in both full-reference and non-reference settings. Q-Ground [7] focuses on the visual quality grounding task, and performs fine-scale detailed visual quality analysis. Some of these works have proposed MLLM-based quality scoring methods, e.g., binary softmax strategy in [68, 91] and pair-wise voting in [79, 80]. Nonetheless, these works principally focus on the quality description ability using languages. + +MLLM-based quality scorer. Initially, Q-Bench [68, 91] proposes a binary softmax strategy, enabling MLLMs to generate quality scores by predicting two discrete quality levels (i.e., good or poor). This strategy is adopted by Q-Instruct [69] and Co-Instruct [71]. Compare2Score [98] achieves quality scoring by teaching MLLMs to compare pair-wise images. Inspired by humans' annotation process, Q-Align [70] discretizes scores to five discrete levels using one-hot label to train MLLMs, resulting in more accurate score regression. Dog-IQA [33] adopts the one-hot label for training-free IQA by incorporating specific standards and local semantic objects. However, as discussed in Sec. 1, one-hot label causes some information loss and can not fully maintain the inter-image relationship. + +# 3. Soft Label Construction + +# 3.1. Revisiting One-hot Label + +Training with one-hot label. MLLMs only accept discrete tokens as inputs and outputs. Therefore, as shown in Fig. 2a, continuous scores1 must be discretized into discrete tokens to train the MLLMs. Q-Align [70] adopts a one-hot discretization method. Specifically, as depicted in Fig. 2b, the score range [1, 5] is uniformly divided into five intervals. These intervals represent five rating levels, {"bad", "poor", "fair", "good", "excellent"}, which is the rating standard defined by [51]. In Fig. 2a, the Mean Opinion Score (MOS) + +![](images/ca9f5b35f46c8c79885aa840602e14e4f79c4e744d4f73ecb592170353afc3da.jpg) +Rate the quality of this image. +(a) +5 candidate level tokens of : + +![](images/4229c4b759b0484b598fc30218bb6045d56e7434fa040201503b2aadcc619013.jpg) + +![](images/85607b301653b1169f3e87de3fa19289e5db6691da3367bbc06d6b436ce25a8a.jpg) +[ bad, poor, fair, good, excellent ] + +![](images/6282334889ff01d98345e4281bb475fb11e76726f56abadb26e95fe41a8faeef.jpg) + +![](images/b6ce449d04ec5b3a3e190fa4076ec1dba39143bb254b5a5bb4d57c0f12fcd496.jpg) +(b) +Continuous score + +![](images/40446f0e90d988740585592af2819b30f8e65aea7a21c3f2cc2bf85f3227c8f0.jpg) +Discretized label +Figure 2. (a) MLLMs need discrete level tokens for training. Thus, continuous scores must be converted into discrete level tokens as training labels. (b) Comparison between one-hot label [70] and our soft label. One-hot label uniformly divides the score range into 5 intervals to discretize the mean score, i.e., the expectation of estimated score distribution. One-hot label assumes that different levels are identically independent, which is inaccurate. In contrast, we discretize the estimated Gaussian distribution of the score. Our soft label better preserves the relationships between levels, e.g., "fair" is closer to "good" than "excellent". + +of the image, 4.30, fails into the fifth intervals, resulting in the discretized label "excellent". Accordingly, the response is formulated as "The quality of this image is excellent". Using the response as the training labels, following [13, 61], MLLMs can be trained with next token prediction loss, i.e., cross-entropy loss between predicted logits and labels. + +Limitations of one-hot label. First, ideal training with cross-entropy loss using one-hot labels assumes that the five levels are independent (i.e., orthogonal in a latent space). However, intuitively, these levels are not truly independent, e.g., "fair" is closer to "good" than "excellent". Second, one-hot discretization only considers the mean score, which is the expectation of the score distribution, ignoring the distribution's uncertainty (i.e., variance). Third, near the border of two levels, images with similar quality can be divided into different levels. For example, in Fig. 1, Image B (MOS 3.38) and Image C (MOS 3.49) share similar mean scores, but are differently classified as "fair" and "good", due to their proximity to the border of these two levels, i.e., 3.4 in Fig. 2b. Finally, one-hot label causes information loss. In Q-Align, the five levels are respectively assigned scores $\{1, 2, 3, 4, 5\}$ , with which we calculate the discretization precision in Tab. 1. The discretization error is around $10 \times$ to $35 \times$ larger than our soft label that we will soon introduce, thus the one-hot label keeps lower correlation with MOS. + +# 3.2. Soft Label + +Discretization using score distribution. As demonstrated in [87], the quality score of one image, $x$ , is typically mod- + +Table 1. Comparison of discretization precision between Q-Align's one-hot label [70] and our soft label. The precision of MOS discretization is measured by L1 Error / RMSE and PLCC / SRCC. The distance of original distribution and recovered distribution is measured by JS divergence / Wasserstein distance. + +
MetricsDiscretizationKonIQ [21]SPAQ [17]KADID [31]
L1 / RSMEOne-hot label0.302 / 0.3740.299 / 0.3660.327 / 0.389
Soft label0.008 / 0.0180.011 / 0.0250.024 / 0.038
PLCC / SRCCOne-hot label0.961 / 0.9520.969 / 0.9680.979 / 0.982
Soft label1.000 / 1.0001.000 / 1.0001.000 / 1.000
JS / W-Dist.One-hot label---
Soft label0.001 / 0.0380.022 / 0.0840.010 / 0.140
+ +eled as a Gaussian distribution, $x \sim \mathcal{N}(\mu, \sigma^2)$ , where $\mu$ and $\sigma^2$ are the Mean Opinion Score (MOS) and variance of multiple human annotations, respectively. The distribution is represented by its probability density function, $f(x)$ . As shown in the second row of Fig. 2b, we select five central points, $c_i \in \{1, 2, 3, 4, 5\}$ , $i \in \{0, 1, 2, 3, 4\}$ , as the centers of the five discrete levels. The width of each level region, $d$ , is set as 1. The possibility of the score falling into the $i^{\mathrm{th}}$ level is denoted as $p_i^{raw}$ and is calculated as: + +$$ +p _ {i} ^ {r a w} = \int_ {c _ {i} - d / 2} ^ {c _ {i} + d / 2} f (x) \mathrm {d x}, \quad i \in \{0, 1, 2, 3, 4 \}. \tag {1} +$$ + +We assign the five levels with the same textual tokens as Q-Align, i.e., {"bad", "poor", "fair", "good", "excellent)}. Correspondingly, the score of $i^{\text{th}}$ level is $c_i$ . This defines a rough discrete distribution as $P^{raw}(c_i) = p_i^{raw}$ . + +Post-adjustment. The discrete distribution is rough because of truncation errors. As shown in the second row of Fig. 2b, there are two truncation regions (i.e., $x < c_0 - d/2 = 0.5$ and $x > c_4 + d/2 = 5.5$ ), which are not considered in Eq. (1). This can cause two problems. First, the sum of $\{p_i^{raw}\}$ does not equal 1, meaning $p_i^{raw}$ is not strictly a discrete distribution. Second, the expectation of this discrete distribution does not equal the expectation of score distribution, $\mu$ , leading to discretization errors. + +To correct these truncation errors, we propose a postadjustment method. Specifically, we apply a linear transformation to adjust $p_i^{raw}$ to more accurate $p_i$ as: + +$$ +p _ {i} = \alpha p _ {i} ^ {r a w} + \beta +$$ + +$$ +\text {s . t .} \quad \sum_ {i} p _ {i} = 1, \quad \mu^ {r e c} = \sum_ {i} p _ {i} c _ {i} = \mu , \tag {2} +$$ + +where $\mu_{rec}$ means recovered expectation from discrete distribution. The parameters of the linear transformation, $\alpha$ and $\beta$ , can be determined by solving these two constraint conditions. One example of post-adjustment is given in Fig. 3. Moreover, the statistics of these two parameters in Tab. 2 show that the average values of $\alpha$ and $\beta$ are quite close to 1 and 0, respectively, implying that the post-adjustment is minimal. This effectively defines our soft label, which follows the discrete distribution as $P(c_i) = p_i$ . + +![](images/9c9d680b802574353b486b41d6d0bdd4c49ba5f3ed9748fce5bcef0953a96916.jpg) +Figure 3. Illustration of the soft label adjustment. + +Table 2. Statistics of the label adjustment parameters. Averaged $\alpha$ and $\beta$ are very approaching 1 and 0, respectively, meaning that the postadjustment is quite small. + +
KonIQ [21]SPAQ [17]KADID [31]
α1.036±0.0411.033±0.0851.118±0.091
β-0.005±0.005-0.005±0.014-0.015±0.010
+ +Ideally, the discretization error of the MOS will be 0. However, the possibility of some levels may slightly fall below zero after the linear transformation. We address this by simply clipping these values to zero. The impact of this clipping on accuracy is negligible, as shown in Tab. 1. After clipping, our discretization error for MOS remains around $10 \times$ to $35 \times$ smaller than Q-Align's one-hot label. Also, our soft label maintains a perfect (i.e., 1.000) Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank-order Correlation Coefficient (SRCC) with the original MOS. + +# Recovery from discrete to continuous score distribution. + +As shown in the second condition of Eq. (2), we can compute the expectation of discrete distribution, $P(c_{i})$ , as the recovered expectation of original distribution. Besides the expectation, we can also recover the original variance using the variance of the discrete distribution, $P(c_{i})$ , as: + +$$ +\mu^ {r e c} = \sum_ {i} p _ {i} c _ {i}, \quad (\sigma^ {r e c}) ^ {2} = \sum_ {i} p _ {i} (c _ {i} - \mu^ {r e c}) ^ {2}. \quad (3) +$$ + +Then we calculate the distance between recovered distribution, $\mathcal{N}(\mu^{rec}, (\sigma^{rec})^2)$ , and original distribution, $\mathcal{N}(\mu, \sigma^2)$ as provided in Tab. 1. Our recovered distribution is closely aligned with human-annotated score distribution. In contrast, one-hot label can not recover such a distribution. + +# 4. DeQA-Score Model + +# 4.1. Training with Our Soft Label + +Model architecture. As illustrated in Fig. 4, we follow the widely adopted MLLM architecture [32, 75] to construct our model. First, a vision encoder is adopted to represent the input image to visual tokens. Then, a vision-to-text connector projects the visual tokens into the textual space. Following [75], we use a vision abstractor as the connector. The abstractor can further reduce the number of visual tokens to 64. Finally, the visual and textual tokens are fused together and fed into an LLM for response prediction. + +Next token prediction loss for common response tokens. As described in Sec. 3.1 and illustrated in Fig. 4, the constructed response follows the template, "The quality of this image is $<$ level $>$ ". For common response tokens other than the level token, we compute the next token prediction loss, i.e., the cross-entropy loss $\mathcal{L}_{ce}$ , following [45, 61]. + +KL divergence loss for the level token. There are five candidate values for the level token: {"bad", "poor", "fair", + +![](images/034b24dc80d8eddb0c048dd00428bc8c05d4ad8297bdf2b6cfd8595889e0d82c.jpg) +Figure 4. Framework of our DeQA-Score trained with soft label. For the “” token, the KL divergence loss is calculated between predicted probabilities and our soft label. For other tokens, common cross-entropy loss for next token prediction is calculated. + +"good", "excellent". Our soft label, $p_i$ , assigns a ground-truth probability to each value. Let the predicted probabilities be $p_i^{pred}$ , the KL divergence loss is calculated through: + +$$ +\mathcal {L} _ {k l} = \sum_ {i} p _ {i} \log \left(p _ {i} ^ {\text {p r e d}} / p _ {i}\right). \tag {4} +$$ + +The predicted probabilities, $p_i^{pred}$ , are obtained by applying a softmax activation function on model's output logits. Note that during training, this softmax function is applied across all textual tokens, not just across the five predefined levels. This approach encourages the model to reduce the possibilities of textual tokens outside the five levels. + +The two terms of losses, $\mathcal{L}_{ce}$ and $\mathcal{L}_{kl}$ , are both autoregressive losses, and thus can be easily combined. During training, the cross-entropy loss, $\mathcal{L}_{ce}$ , will decrease near to 0 fast, since it is quite easy to optimize by simply remembering the response template "The quality of this image is". + +# 4.2. Score Estimation during Inference + +During inference, using the model-predicted probabilities of the five levels, $p_i^{pred}$ , we can estimate the expectation and variance of the score distribution the same as Eq. (3): + +$$ +\mu^ {p r e d} = \sum_ {i} p _ {i} ^ {p r e d} c _ {i}, (\sigma^ {p r e d}) ^ {2} = \sum_ {i} p _ {i} ^ {p r e d} (c _ {i} - \mu^ {p r e d}) ^ {2}. (5) +$$ + +Here we follow Q-Align [70] to apply a close-set softmax on the five levels to obtain the probabilities, $p_i^{pred}$ , to prevent the influence of other textual tokens. Our statistics in Tab. A2 reveal that the trained MLLM will always predict the five levels, thus applying softmax on either the five levels or all textual tokens yields nearly identical results. As given in Tab. 5, our predicted score distribution, $\mathcal{N}(\mu^{pred}, (\sigma^{pred})^2)$ , aligns well with human annotations. + +# 4.3. Joint Training on Multiple IQA Datasets + +As demonstrated in [87, 98], different IQA datasets exhibit distinct distributions. Due to this dataset variation, images from different datasets with nearly identical scores (i.e., linearly re-scaled scores) can have drastically different perceptual quality (see examples in Fig. A1). The reason is that different datasets have different reflections between quality scores and visual quality. This disparity makes it infeasible to merely regress absolute quality scores for multi-dataset + +![](images/861682710f7e97a603ccd26e55fbaae04cbfe1c32a455b1e1a02c1c4897b9183.jpg) +Figure 5. Fidelity loss to maintain intra-dataset relationships. Consider two images, Image A and B, sampled from the same dataset. Using our distribution-based soft label, DeQA-Score predicts the expectation, $\mu^{pred}$ , and variance, $(\sigma^{pred})^2$ , of each image's score distribution, with which we can predict the probability that Image A is better than Image B, $P(A > B)$ . This predicted probability is then used to calculate the fidelity loss with the ground-truth probability. Fidelity loss preserves intra-dataset relationships, facilitating the co-training across multiple IQA datasets. + +co-training. To address this challenge, UNIQUE [87] applies fidelity loss [63] based on Thurstone's model [60]. The fidelity loss is based on the principal that IQA emphasizes the ranking relationships within the same dataset, rather than the absolute values of quality scores. Specifically, two images are sampled from the same dataset, whose quality scores can be comparable. These two images are then compared using human-annotated quality scores. The fidelity loss encourages the model to learn such comparison results, which facilitates the model to focus more on the ranking relationships in each dataset, instead of the absolute values of quality scores. Thus, the fidelity loss helps multi-dataset co-training. Our soft label is naturally compatible with the fidelity loss, since our method can predict the score distribution, which is what the fidelity loss needs. + +Ground-truth construction. As shown in Fig. 5, a pair of images, Image A and B, are sampled from the same dataset. We assume that the quality scores of the two images, $x_A, x_B$ , follow the Gaussian distribution with means and variances collected by subjective testing, i.e., $x_A \sim \mathcal{N}(\mu_A, \sigma_A^2)$ , $x_B \sim \mathcal{N}(\mu_B, \sigma_B^2)$ . Assuming that their quality scores are uncorrelated, the score difference, $x_A - x_B$ , also follows a Gaussian distribution, $\mathcal{N}(\mu_A - \mu_B, \sigma_A^2 + \sigma_B^2)$ . + +Then the possibility that Image A has a better perceptual quality than Image B can be calculated as: + +$$ +p (A > B) = p \left(x _ {A} > x _ {B}\right) = \Phi \left(\frac {\mu_ {A} - \mu_ {B}}{\sqrt {\sigma_ {A} ^ {2} + \sigma_ {B} ^ {2}}}\right), \tag {6} +$$ + +where $\Phi (\cdot)$ is the Gaussian cumulative distribution function. + +Model prediction. As described in Sec. 4.2, our model can predict the expectation and variance of the score distribution. Denoting the predicted expectation and variance of two images as $\mu_A^{pred}$ , $(\sigma_A^{pred})^2$ , $\mu_B^{pred}$ , $(\sigma_B^{pred})^2$ , same to Eq. (6), we can predict the possibility of Image A surpassing Image B in perceptual quality as: + +$$ +p ^ {p r e d} (A > B) = \Phi \left(\frac {\mu_ {A} ^ {p r e d} - \mu_ {B} ^ {p r e d}}{\sqrt {(\sigma_ {A} ^ {p r e d}) ^ {2} + (\sigma_ {B} ^ {p r e d}) ^ {2}}}\right). \quad (7) +$$ + +Training with fidelity loss. Following [87], we choose fidelity loss as the similarity measure to train the model as: + +$$ +\begin{array}{l} \mathcal {L} _ {f d} = 1 - \sqrt {p (A > B) p ^ {p r e d} (A > B)} \\ - \sqrt {\left(1 - p (A > B)\right) \left(1 - p ^ {p r e d} (A > B)\right)}. \tag {8} \\ \end{array} +$$ + +Adding two auto-regressive losses, the final training loss is: + +$$ +\mathcal {L} = \mathcal {L} _ {f d} + \gamma \left(\mathcal {L} _ {c e} + \mathcal {L} _ {k l}\right), \tag {9} +$$ + +where $\gamma$ is a weighting term to adjust the loss scale. + +# 5. Experiments + +# 5.1. Datasets, Baselines, Metrics, and Details + +Datasets. For a comprehensive evaluation, we include a wide range of IQA datasets: (a) in-the-wild images: KonIQ [21], SPAQ [17], LIVE-Wild [18], and FLIVE [78]; (b) synthetic distortions: KADID [31], TID2013 [47], and CSIQ [29]; (c) model-processed distortions: PIPAL [24]; (d) AI-generated images: AGIQA-3K [30]. Among these datasets, KonIQ, SPAQ, KADID, and PIPAL are split into training and test sets. Other datasets are used to assess the out-of-distribution generalization ability. The Mean Opinion Scores (MOS) of these datasets are normalized to the range of [1,5], with the variances normalized accordingly. + +Baseline methods. To demonstrate the effectiveness of our distribution-based soft label, we primarily compare with Q-Align [70], which utilizes the one-hot label. The base models for DeQA-Score and Q-Align are kept identical for fair comparison. We also include previous IQA methods as baselines, including handcrafted methods: NIQE [41] and BRISQUE [40], and deep-learning-based methods: NIMA [57], HyperIQA [55], DBCNN [86], MUSIQ [26], CLIP-IQA+ [64], ManIQA [73], and Compare2Score [98]. + +Metrics. Following [26, 70, 73], we use the Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank-order Correlation Coefficient (SRCC) as metrics to evaluate the score regression performance. Additionally, we employ KL divergence, JS divergence, and Wasserstein distance to assess the results of score distribution prediction. + +Implementation details. We use mPLUG-Owl2 [75] as our base model, with a CLIP-pretrained ViT-L [49] as the image encoder, a six-layer Q-Former as the visual abstractor, and LLaMA-2-7B [62] as the LLM. The number of visual tokens output by the vision abstractor is set to 64, and the hidden dimension of LLM is 4096. The pretrained weights of mPLUG-Owl2 are used for model initialization. The weighting term $\gamma$ is empirically set to 0.05, adjusting the two loss terms roughly to the same scale. We adopt AdamW [23] as the optimizer. The initial learning rate is set to 2e-5 and decays gradually using the cosine decay strategy. Our model is trained with a batch size of 64 for 3 epochs. Using 8 NVIDIA RTX A6000 GPUs, the training process on KonIQ dataset is completed in around 1.5 hours. + +# 5.2. Results of Score Regression + +Score regression results of single-dataset training. We first train our model on a single dataset, i.e., KonIQ, and then evaluate the model on both in-distribution and out-of-distribution datasets, following [28, 70]. The score regression results are shown in Tab. 3. On the in-distribution KonIQ test set, our model achieves superior score regression results compared to all baseline methods, e.g., a $1.3\%$ improvement in PLCC over Q-Align. Furthermore, on out-of-distribution datasets, our DeQA-Score maintains its advantage with the best overall results across nearly all benchmarks, e.g., a $4.6\%$ improvement in PLCC on LIVE-Wild. These findings confirm the effectiveness of our method. + +However, we observe that all models perform poorly on the PIPAL dataset. For instance, our best results reach only 0.462 on PLCC. The main reason is the distortion gap between the training dataset, KonIQ (in-the-wild distortions), and PIPAL (model-processed distortions). This highlights the necessity of co-training with multiple IQA datasets containing various distortion types. This observation motivates our subsequent experiments on multi-dataset co-training. + +Score regression results of multi-dataset co-training. As discussed above, co-training on multiple IQA datasets is necessary to obtain a unified model for various categories of distortions. Therefore, we conduct multi-dataset cotraining experiments in Tab. 4. First, our DeQA-Score consistently outperforms Q-Align across different training combinations. The performance gain attributes to both the soft label and fidelity loss, as confirmed in Tab. 6. Second, comparing #0 to #1, when KADID (synthetic distortions) is added to the training set, the performance on TID2013 and CSIQ (both synthetic distortions) improves + +Table 3. Score regression results of single-dataset training with the PLCC / SRCC metrics. All methods except handcrafted ones are trained on the KonIQ dataset. DeQA-Score outperforms all baseline methods across nearly all benchmarks. + +
CategoryMethodsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KCSIQFLIVE
HandcraftedNIQE [41]0.533 / 0.5300.679 / 0.6640.468 / 0.4050.195 / 0.1610.493 / 0.4490.560 / 0.5330.718 / 0.6280.147 / 0.100
BRISQUE [40]0.225 / 0.2260.490 / 0.4060.429 / 0.3560.267 / 0.2320.361 / 0.3130.541 / 0.4970.740 / 0.5560.108 / 0.054
Non-MLLMNIMA [57]0.896 / 0.8590.838 / 0.8560.532 / 0.5350.390 / 0.3990.814 / 0.7710.715 / 0.6540.695 / 0.6490.561 / 0.467
HyperIQA [55]0.917 / 0.9060.791 / 0.7880.506 / 0.4680.410 / 0.4030.772 / 0.7490.702 / 0.6400.752 / 0.7170.485 / 0.383
DBCNN [86]0.884 / 0.8750.812 / 0.8060.497 / 0.4840.384 / 0.3810.773 / 0.7550.730 / 0.6410.586 / 0.5720.485 / 0.385
MUSIQ [26]0.924 / 0.9290.868 / 0.8630.575 / 0.5560.431 / 0.4310.789 / 0.8300.722 / 0.6300.771 / 0.7100.565 / 0.467
CLIP-IQA+ [64]0.909 / 0.8950.866 / 0.8640.653 / 0.6540.427 / 0.4190.832 / 0.8050.736 / 0.6850.772 / 0.7190.427 / 0.316
ManIQA [73]0.849 / 0.8340.768 / 0.7580.499 / 0.4650.457 / 0.4520.849 / 0.8320.723 / 0.6360.623 / 0.6270.512 / 0.401
MLLM-basedCompare2Score [98]0.923 / 0.9100.867 / 0.8600.500 / 0.4530.354 / 0.3420.786 / 0.7720.777 / 0.6710.735 / 0.7050.474 / 0.413
Q-Align [70]0.941 / 0.9400.886 / 0.8870.674 / 0.6840.403 / 0.4190.853 / 0.8600.772 / 0.7350.785 / 0.7370.554 / 0.483
DeQA-Score0.953 / 0.9410.895 / 0.8960.694 / 0.6870.472 / 0.4780.892 / 0.8790.809 / 0.7290.787 / 0.7440.589 / 0.501
+ +Table 4. Score regression results of co-training on multiple IQA datasets with the PLCC / SRCC metrics. Our DeQA-Score consistently surpasses the baseline method, Q-Align, across all benchmarks with different training combinations. + +
#Trained onMethodsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
0KonIQ, SPAQQ-Align [70]0.943 / 0.9400.933 / 0.9310.692 / 0.7080.401 / 0.4110.883 / 0.8790.795 / 0.7270.671 / 0.5680.795 / 0.767
DeQA-Score0.953 / 0.9430.936 / 0.9330.724 / 0.7190.468 / 0.4740.902 / 0.8880.810 / 0.7380.706 / 0.6160.836 / 0.785
1KonIQ, SPAQ, KADIDQ-Align [70]0.945 / 0.9380.933 / 0.9310.935 / 0.9340.409 / 0.4200.887 / 0.8830.788 / 0.7330.829 / 0.8080.876 / 0.845
DeQA-Score0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
2KonIQ, SPAQ, KADID, PIPALQ-Align [70]0.926 / 0.9320.917 / 0.9200.950 / 0.9540.702 / 0.6710.853 / 0.8450.765 / 0.7220.811 / 0.7950.838 / 0.789
DeQA-Score0.958 / 0.9460.932 / 0.9290.963 / 0.9610.724 / 0.6900.877 / 0.8570.770 / 0.7350.828 / 0.7960.863 / 0.807
+ +Table 5. Score distribution prediction results with JS divergence / Wasserstein distance between the predicted Gaussian distribution and the human labeled Gaussian distribution as metrics. Models are trained on the KonIQ, SPAQ, and KADID datasets. DeQA-Score achieves a much closer alignment with human annotations. + +
DatasetsQ-Align [70]DeQA-Score (Ours)
KonIQ0.415 / 0.6110.014 / 0.186
SPAQ0.192 / 0.4550.070 / 0.332
KADID0.519 / 0.7230.029 / 0.304
LIVE-Wild0.254 / 0.6420.052 / 0.417
AGIQA-3K0.558 / 0.7300.177 / 0.592
+ +significantly (e.g., $33.1\%$ in SRCC on TID2013), underscoring that distortion type is a major gap between training and testing. Third, from #1 to #2, adding PIPAL for training greatly enhances the results on PIPAL (from 0.495 to 0.724 in PLCC) but reduces the performance on several other datasets, e.g., from 0.900 to 0.863 in PLCC on CSIQ. This performance decline is likely due to the unique nature of model-processed distortions in PIPAL, which differ greatly from other distortions. Finally, despite never trained on AI-generated images, DeQA-Score performs well on AGIQA-3K with around 0.8 PLCC, reflecting the generalization ability enabled by MLLMs' broad pre-training. + +More results on AIGC images are provided in Tab. A4. + +# 5.3. Results of Score Distribution Prediction + +As stated in Sec. 4.2, the MLLM-output probabilities allow us to predict a score distribution, which can be evaluated against the human labeled ground-truth distribution. We + +calculate the distance between these two distributions and present the results in Tab. 5. The results illustrate that our predicted score distribution aligns much better with humans than Q-Align. As shown in Fig. 6, Q-Align tends to predict a single level token, bringing a quite small predicted variance, which compromises the distribution metrics. + +Qualitative results are shown in Fig. 6. Across different image types: in-the-wild images (Fig. 6ab), AI-generated images (Fig. 6c), or synthetic distortions (Fig. 6d), and for varying quality levels: high-quality (Fig. 6ac), fair-quality (Fig. 6b), or poor-quality images (Fig. 6d), our method consistently predicts accurate quality scores. Additionally, our DeQA-Score provides a score distribution that aligns more closely with humans with smaller KL divergence. In contrast, Q-Align, which is trained with one-hot labels, tends to predict a single label, resulting in higher KL divergence. + +# 5.4. Ablation Studies + +Effectiveness of soft label and fidelity loss is studied and shown in Tab. 6. In #1, "Uncertainty" is utilized to verify the effects of adding annotation variances without considering the score distortion. We directly append uncertainty within the responses, i.e., "The quality of this image is with uncertainty". The has five possible values: {"minimal", "low", "moderate", "high", "severe"}, derived by the one-hot discretization. In #2, model trained on one-hot label can not predict a reliable variance, thus we employ binary fidelity loss [89], which does not require predicted variance. + +![](images/c699409b820ce91ede379bc8da6cf3fe9c1ca7a3022565144543be88ed462c3f.jpg) +(a) Score: $3.93 \pm 0.79$ + +![](images/dc877a31a5fed4204432786271d4cc0da1438ba39b46f9e32abe8d535e5a5e50.jpg) + +![](images/8714ef4df6cf8bcee3b733e0b6d67a887732ed3f0245c7b066eae90c6237a818.jpg) + +![](images/b35d2693e00ddaf1a5869e88e5dfd1c1fb60de964b0bb631c1275f10ec648a01.jpg) + +![](images/e3319e264f5260424b2edf37eda1075b0c5e27834d81eb45fbbadad8f360aa35.jpg) + +![](images/252eee6bc93d1a830eb98985d222d5d34942159a6771d6521a610b8270e511ea.jpg) + +![](images/178a4aabf6be7abc4c222416926a22c9cde2ebab09fcca2919558c926b5dd88e.jpg) +(b) Score: $3.11 \pm 0.66$ + +![](images/45657c6b07acc75babb77f2c0dfa1d13471e286a8556b66fe5798581cf8366e6.jpg) +Pred: 3.95, KL Div: 9.01 + +![](images/c570ba6f583d1d6ddf5d38abdd12d11914512d281ea3b6eb689e76ec73fb45bd.jpg) +Pred: 3.34, KL Div: 0.06 +Figure 6. Qualitative results on KonIQ, SPAQ, AGIQA-3K, and KADID datasets. DeQA-Score predicts more accurate quality scores with more reasonable logits. Also, our method generates a score distribution that aligns better with humans, i.e., smaller KL divergence. However, Q-Align, which is trained on one-hot label, is more likely to predict a single label, resulting in higher KL divergence. + +![](images/5c05a66f95588d56945a0f9153eb720fccf1f484dd78ced507edc05d977bf07f.jpg) +(d) Score: $1.92 \pm 0.82$ + +![](images/2c37c48761abc7d2a2469ed5c8e0c36f0474999395086f7f474996972224c0aa.jpg) +Pred: 1.04, KL Div: 16.69 + +![](images/9d025297513e19736e64d8cc420319a8ae35d85b43196ac11890c821b49cd47d.jpg) +Pred: 1.76, KL Div: 0.14 + +Table 6. Ablation studies on soft label and fidelity loss with PLCC / SRCC metrics. Models are trained on KonIQ, SPAQ, and KADID. + +
#One-hotSoftUncertaintyFidelityKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
0✓0.945 / 0.9380.933 / 0.9310.935 / 0.9340.409 / 0.4200.887 / 0.8830.788 / 0.7330.829 / 0.8080.876 / 0.845
1✓✓0.935 / 0.9360.933 / 0.9310.954 / 0.9510.439 / 0.4530.896 / 0.8870.764 / 0.7150.826 / 0.8030.884 / 0.849
2✓✓0.944 / 0.9360.935 / 0.9320.951 / 0.9500.473 / 0.4780.898 / 0.8860.788 / 0.7260.837 / 0.8100.876 / 0.839
3✓0.954 / 0.9430.937 / 0.9340.951 / 0.9470.472 / 0.4720.896 / 0.8790.791 / 0.7290.852 / 0.8210.894 / 0.855
4✓✓0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
+ +
CategoryLevel TextsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
Commonexce. / good / far / poor / bad0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
Randomexce. / poor / bad / good / fair0.942 / 0.9280.933 / 0.9290.950 / 0.9470.394 / 0.3960.863 / 0.8510.795 / 0.7270.814 / 0.7920.887 / 0.836
Reversebad / poor / fair / good / exce.0.943 / 0.9340.934 / 0.9310.947 / 0.9450.395 / 0.3860.836 / 0.8000.782 / 0.7140.811 / 0.7910.882 / 0.834
Numberone / two / three / four / five0.945 / 0.9320.934 / 0.9300.955 / 0.9520.421 / 0.4240.873 / 0.8570.766 / 0.7120.820 / 0.7940.882 / 0.835
Vehiclecar / bus / train / plane / ship0.953 / 0.9400.937 / 0.9330.954 / 0.9510.458 / 0.4560.889 / 0.8780.778 / 0.7020.854 / 0.8280.897 / 0.856
Mixtureapple / car / four / fair / dog0.947 / 0.9380.936 / 0.9320.954 / 0.9510.419 / 0.4150.865 / 0.8510.808 / 0.7370.827 / 0.8000.896 / 0.854
+ +Table 7. Effects of different level texts with PLCC / SRCC metrics. Models are trained on the KonIQ, SPAQ, and KADID datasets. +Table 8. Inference latency & throughput on RTX A6000 GPU. + +
Batch Size12481664256
Latency (s)0.0890.1210.1820.3150.5882.1147.929
Throughput (image/s)11.2416.5321.9825.4027.2130.2732.29
+ +Several conclusions can be drawn from Tab. 6. First, comparing #0 with #3, our soft label significantly outperforms the one-hot label across all benchmarks, validating its effectiveness. Second, from #3 to #4, the performance of our soft label further improves when combined with the fidelity loss. Third, comparing #0 with #2, the one-hot label does not incorporate well with the fidelity loss, as it lacks score distribution information. Finally, from #0 to #1, simply adding uncertainty without considering score distribution does not yield stable or significant improvements. + +Effects of different level texts are studied in Tab. 7. Beside the commonly used level texts, we consider five alternatives: (a) random sorted common texts; (b) reversed common texts; (c) simple numbers like one, two, etc.; (d) irrelevant words about vehicle; (e) a mixture of various types. The results indicate that common level texts achieve the best performance, as they align closely with the MLLMs' pre + +training. Interestingly, with random sorted / reversed texts, even a mixture of various level texts, which are inconsistent with MLLMs' prior knowledge, our model still learns these newly defined relationships and performs well. This shows the robust adaptability of MLLM-based IQA models. + +Ablation studies on level number and training / fixing components are provided in Tabs. A6 and A7, respectively. + +# 5.5. Inference Latency + +We test the inference latency on one single RTX A6000 GPU. As shown in Tab. 8, with the optimal GPU utilization (i.e., batch size 256), our DeQA-Score can process as many as 32.29 images per second in average. This high throughput emphasizes its suitability for real-time deployment. + +# 6. Conclusions + +We develop DeQA-Score as a pioneering MLLM-based image quality scorer, empowered by distribution-based soft labels and fidelity loss, capable of both score regression and score distribution prediction, highlighting the importance of quality score distribution in MLLM-based IQA methods. + +Acknowledgement. This work was sponsored by RGC Early Career Scheme (ECS) No. 24209224, National Natural Science Foundation of China (Grant No.62276251), and the Joint Lab of CAS-HK. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019. 3 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2 +[4] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE TIP, 2018. 2 +[5] Xin Cai, Zhiyuan You, Hailong Zhang, Wentao Liu, Jinwei Gu, and Tianfan Xue. PhoCoLens: Photorealistic and consistent reconstruction in lensless imaging. In NeurIPS, 2024. 1 +[6] Yue Cao, Zhaolin Wan, Dongwei Ren, Zifei Yan, and Wangmeng Zuo. Incorporating semi-supervised and positive unlabeled learning for boosting full reference image quality assessment. In CVPR, 2022. 2 +[7] Chaofeng Chen, Yang Sensen, Haoning Wu, Liang Liao, Zicheng Zhang, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-Ground: Image quality grounding with large multi-modality models. In ACM MM, 2024. 1, 3 +[8] Du Chen, Tianhe Wu, Kede Ma, and Lei Zhang. Toward generalized image quality assessment: Relaxing the perfect reference quality assumption. In CVPR, 2025. 2 +[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 3 +[10] Zixuan Chen, Zewei He, Ziqian Lu, Xuecheng Sun, and Zhe-Ming Lu. Prompt-based test-time real image dehazing: a novel pipeline. In ECCV, 2024. 1 +[11] Zijian Chen, Wei Sun, Haoning Wu, Zicheng Zhang, Jun Jia, Ru Huang, Xiongkuo Min, Guangtao Zhai, and Wenjun Zhang. Study of subjective and objective naturalness assessment of ai-generated images. IEEE TCSVT, 2024. 14 +[12] Zixuan Chen, Yujin Wang, Xin Cai, Zhiyuan You, Zheming Lu, Fan Zhang, Shi Guo, and Tianfan Xue. UltraFusion: Ultra high dynamic imaging using exposure fusion. In CVPR, 2025. 1 +[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong- + +hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with $90\%$ * chatgpt quality. https://vicuna.lmsys.org, 2023. 2, 3 +[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2 +[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE TPAMI, 2020. 2 +[16] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In ACM MM, 2021. 2 +[17] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, 2020. 1, 2, 4, 6 +[18] Deepti Ghadiyaram and C. Bovik Alan. Live in the wild image quality challenge database. http://live.ece.utexas.edu/research/ChallengeDB/index.html, 2015.2,6 +[19] Abhijay Ghildyal and Feng Liu. Shift-tolerant perceptual similarity metric. In ECCV, 2022. 2 +[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3 +[21] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. KonIQ-10K: An ecologically valid database for deep learning of blind image quality assessment. IEEE TIP, 2020. 2, 4, 6 +[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2021. 13, 14, 15 +[23] Loshchilov Ilya and Hutter Frank. Decoupled weight decay regularization. In ICLR, 2019. 6, 13 +[24] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. PIPAL: a large-scale image quality assessment dataset for perceptual image restoration. In ECCV, 2020. 2, 6 +[25] Le Kang, Peng Ye, Yi Li, and David Doermann. Convolutional neural networks for no-reference image quality assessment. In CVPR, 2014. 2 +[26] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality transformer. In CVPR, 2021. 1, 2, 6, 7 +[27] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 2018. 13 +[28] Daekyu Kwon, Dongyoung Kim, Sehwan Ki, Younghyun Jo, Hyong-Euk Lee, and Seon Joo Kim. ATTIQA: Generalizable image quality feature extractor using attribute-aware pretraining. In ACCV, 2024. 6 + +[29] Eric C Larson and Damon M Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 2010. 6 +[30] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. AGIQA-3K: An open database for ai-generated image quality assessment. IEEE TCSVT, 2023. 1, 2, 6, 14 +[31] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. KADID-10K: A large-scale artificially distorted iqa database. In International Conference on Quality of Multimedia Experience (QoMEX), 2019. 2, 4, 6 +[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1, 2, 4 +[33] Kai Liu, Ziqing Zhang, Wenbo Li, Renjing Pei, Fenglong Song, Xiaohong Liu, Linghe Kong, and Yulun Zhang. Dog-IQA: Standard-guided zero-shot mllm for mix-grained image quality assessment. arXiv preprint arXiv:2410.02505, 2024.3 +[34] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. RankIQA: Learning from rankings for no-reference image quality assessment. In ICCV, 2017. 2 +[35] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 3 +[36] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, 2022. 3 +[37] Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding, 2017. 2 +[38] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In ACL Findings, 2022. 3 +[39] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. DocVQA: A dataset for vqa on document images. In WACV, 2021. 3 +[40] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE TIP, 2012. 2, 6, 7 +[41] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Sign. Process. Letters, 2012. 1, 6, 7 +[42] Anush Krishna Moorthy and Alan Conrad Bovik. A two-step framework for constructing blind image quality indices. IEEE Sign. Process. Letters, 2010. +[43] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE TIP, 2011. 2 +[44] OpenAI. GPT-4V(ision) system card, 2023. 1, 2 +[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini + +Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 4 +[46] Da Pan, Ping Shi, Ming Hou, Zefeng Ying, Sizhe Fu, and Yuan Zhang. Blind predicting similar quality map for image quality assessment. In CVPR, 2018. 2 +[47] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database TID2013: Peculiarities, results and perspectives. Signal processing: Image communication, 2015. 6 +[48] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. PieAPP: Perceptual image-error assessment through pairwise preference. In CVPR, 2018. 2 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 13 +[50] Michele A Saad, Alan C Bovik, and Christophe Charrier. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE TIP, 2012. 2 +[51] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 2012. 3 +[52] H Sheikh. LIVE image quality assessment database release 2. http://live.ece.utexas.edu/research/quality, 2005. 1, 2 +[53] Hamid R Sheikh and Alan C Bovik. Image information and visual quality. IEEE TIP, 2006. 2 +[54] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA models that can read. In CVPR, 2019. 3 +[55] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In CVPR, 2020. 2, 6, 7 +[56] Simen Sun, Tao Yu, Jiahua Xu, Wei Zhou, and Zhibo Chen. GraphIQA: Learning distortion graph representations for blind image quality assessment. IEEE TMM, 2022. 2 +[57] Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment. IEEE TIP, 2018. 6, 7 +[58] Huixuan Tang, Neel Joshi, and Ashish Kapoor. Learning a blind measure of perceptual image quality. In CVPR, 2011. 2 +[59] InternLM Team. InternLM: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM, 2023. 2 +[60] Louis L Thurstone. A law of comparative judgment. Psychological review, 1994. 2, 5 +[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 3, 4 +[62] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, + +Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. +LLaMA 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6, 13 +[63] Ming-Feng Tsai, Tie-Yan Liu, Tao Qin, Hsin-Hsi Chen, and Wei-Ying Ma. FRank: A ranking method with fidelity loss. In International ACM SIGIR Conference on Research and Development in Information Retrieval, 2007. 2, 5 +[64] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 2, 6, 7 +[65] Jiarui Wang, Huiyu Duan, Jing Liu, Shi Chen, Xiongkuo Min, and Guangtao Zhai. AIGCIQA2023: A large-scale image quality assessment database for ai generated images: from the perspectives of quality, authenticity and correspondence. In CAAI International Conference on Artificial Intelligence, 2023. 14 +[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 1, 2 +[67] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv preprint arXiv:2312.06109, 2023. 2 +[68] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-Bench: A benchmark for general-purpose foundation models on low-level vision. In ICLR, 2024. 3, 15, 16 +[69] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-Instruct: Improving low-level visual abilities for multi-modality foundation models. In CVPR, 2024. 1, 2, 3, 15, 16 +[70] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-Align: Teaching LMMs for visual scoring via discrete text-defined levels. In ICML, 2024. 1, 2, 3, 4, 5, 6, 7, 14, 15 +[71] Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, et al. Towards open-ended visual quality comparison. In ECCV, 2024. 1, 2, 3 +[72] Tianhe Wu, Kede Ma, Jie Liang, Yujiu Yang, and Lei Zhang. A comprehensive study of multimodal large language models for image quality assessment. In ECCV, 2024. 3 +[73] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. MANIQA: Multi-dimension attention network for no-reference image quality assessment. In CVPRW, 2022. 2, 6, 7 +[74] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mPLUG-Owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2 + +[75] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mPLUG-Owl2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 1, 4, 6, 13 +[76] Guanghao Yin, Wei Wang, Zehuan Yuan, Chuchu Han, Wei Ji, Shouqian Sun, and Changhu Wang. Content-variant reference image quality assessment via knowledge distillation. In AAAI, 2022. 2 +[77] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Lu Sheng, Lei Bai, Xiaoshui Huang, Zhiyong Wang, et al. Lamm: Language-assisted multimodal instruction-tuning dataset, framework, and benchmark. In NeurIPS, 2023. 2 +[78] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (PaQ-2-PiQ): Mapping the perceptual space of picture quality. In CVPR, 2020. 2, 6 +[79] Zhiyuan You, Jinjin Gu, Zheyuan Li, Xin Cai, Kaiwen Zhu, Chao Dong, and Tianfan Xue. Descriptive image quality assessment in the wild. arXiv preprint arXiv:2405.18842, 2024. 1, 2, 3 +[80] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. In ECCV, 2024. 1, 2, 3 +[81] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2014. 3 +[82] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. FSIM: A feature similarity index for image quality assessment. IEEE TIP, 2011. 2 +[83] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2 +[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 1, 2 +[85] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. LLaMA-Adapter: Efficient fine-tuning of large language models with zero-initialized attention. In ICLR, 2024. 2 +[86] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE TCSVT, 2020. 6, 7 +[87] Weixia Zhang, Kede Ma, Guangtao Zhai, and Xiaokang Yang. Uncertainty-aware blind image quality assessment in the laboratory and wild. IEEE TIP, 2021. 2, 3, 5, 6 +[88] Weixia Zhang, Dingquan Li, Chao Ma, Guangtao Zhai, Xiaokang Yang, and Kede Ma. Continual learning for blind image quality assessment. IEEE TPAMI, 2022. 2 + +[89] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In CVPR, 2023. 2, 7, 14, 15 +[90] Zicheng Zhang, Chunyi Li, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. A perceptual quality assessment exploration for aigc images. In ICME Workshops, 2023. 14 +[91] Zicheng Zhang, Haoning Wu, Erli Zhang, Guangtao Zhai, and Weisi Lin. A benchmark for multi-modal foundation models on low-level vision: from single images to pairs. IEEE TPAMI, 2024. 3 +[92] Zicheng Zhang, Yingjie Zhou, Chunyi Li, Baixuan Zhao, Xiaohong Liu, and Guangtao Zhai. Quality assessment in the era of large models: A survey. arXiv preprint arXiv:2409.00031, 2024. 3 +[93] Heliang Zheng, Jianlong Fu, Yanhong Zeng, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. ICCV, 2021. 2 +[94] Wei Zhou and Zhou Wang. Quality assessment of image super-resolution: Balancing deterministic and statistical fidelity. In ACM MM, 2022. 2 +[95] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 1, 2 +[96] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. MetaQA: deep meta-learning for no-reference image quality assessment. In CVPR, 2020. 2 +[97] Hanwei Zhu, Xiangjie Sui, Baoliang Chen, Xuelin Liu, Peilin Chen, Yuming Fang, and Shiqi Wang. 2afc prompting of large multimodal models for image quality assessment. arXiv preprint arXiv:2402.01162, 2024. 3 +[98] Hanwei Zhu, Haoning Wu, Yixuan Li, Zicheng Zhang, Bao-liang Chen, Lingyu Zhu, Yuming Fang, Guangtao Zhai, Weisi Lin, and Shiqi Wang. Adaptive image quality assessment via teaching large multimodal model to compare. arXiv preprint arXiv:2405.19298, 2024. 2, 3, 5, 6, 7, 14, 15 +[99] Kaiwen Zhu, Jinjin Gu, Zhiyuan You, Yu Qiao, and Chao Dong. An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809, 2024. 1 \ No newline at end of file diff --git a/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/images.zip b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..baca76c16d438eb093398ccfcbcc54a58ff2d659 --- /dev/null +++ b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e58764e29f4c05bee57e58a4651b062bebab69a1694aa24e50616fa34cb0336c +size 743649 diff --git a/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/layout.json b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..68368013238608fff70a05112dd26b3f78abb7ff --- /dev/null +++ b/2025/Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution/layout.json @@ -0,0 +1,11602 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 541, + 140 + ], + "type": "text", + "content": "Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "spans": [ + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "text", + "content": "Zhiyuan You" + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "text", + "content": ", Xin Cai" + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "text", + "content": ", Jinjin Gu" + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "text", + "content": ", Tianfan Xue" + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "inline_equation", + "content": "^{23\\dagger}" + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "text", + "content": ", Chao Dong" + }, + { + "bbox": [ + 135, + 160, + 470, + 176 + ], + "type": "inline_equation", + "content": "^{134\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 177, + 492, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 177, + 492, + 190 + ], + "spans": [ + { + "bbox": [ + 118, + 177, + 492, + 190 + ], + "type": "text", + "content": "1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 191, + 465, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 191, + 465, + 205 + ], + "spans": [ + { + "bbox": [ + 145, + 191, + 465, + 205 + ], + "type": "text", + "content": "2The Chinese University of Hong Kong, 3Shanghai AI Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 190, + 206, + 420, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 206, + 420, + 220 + ], + "spans": [ + { + "bbox": [ + 190, + 206, + 420, + 220 + ], + "type": "text", + "content": "4Shenzhen University of Advanced Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 221, + 514, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 221, + 514, + 233 + ], + "spans": [ + { + "bbox": [ + 96, + 221, + 514, + 233 + ], + "type": "text", + "content": "zhiyuanyou@link.cuhk.edu.hk, caixin@link.cuhk.edu.hk, jinjin.gu@sydney.edu.au" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 235, + 471, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 235, + 471, + 247 + ], + "spans": [ + { + "bbox": [ + 139, + 235, + 471, + 247 + ], + "type": "text", + "content": "tfxue@ie.cuhk.edu.hk, chao.dong@siat.ac.cn † Corresponding Author" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 274, + 200, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 274, + 200, + 288 + ], + "spans": [ + { + "bbox": [ + 152, + 274, + 200, + 288 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 300, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 647 + ], + "type": "text", + "content": "With the rapid advancement of Multi-modal Large Language Models (MLLMs), MLLM-based Image Quality Assessment (IQA) methods have shown promising performance in linguistic quality description. However, current methods still fall short in accurately scoring image quality. In this work, we aim to leverage MLLMs to regress accurate quality scores. A key challenge is that the quality score is inherently continuous, typically modeled as a Gaussian distribution, whereas MLLMs generate discrete token outputs. This mismatch necessitates score discretization. Previous approaches discretize the mean score into a one-hot label, resulting in information loss and failing to capture inter-image relationships. We propose a distribution-based approach that discretizes the score distribution into a soft label. This method preserves the characteristics of the score distribution, achieving high accuracy and maintaining inter-image relationships. Moreover, to address dataset variation, where different IQA datasets exhibit various distributions, we introduce a fidelity loss based on Thurstone's model. This loss captures intra-dataset relationships, facilitating co-training across multiple IQA datasets. With these designs, we develop the distribution-based Depicted image Quality Assessment model for Score regression (DeQA-Score). Experiments across multiple benchmarks show that DeQA-Score stably outperforms baselines in score regression. Also, DeQA-Score can predict the score distribution that closely aligns with human annotations. Codes and model weights have been released in https://depictqa.github.io/deqa-score/." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 135, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 135, + 670 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 135, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Image Quality Assessment (IQA) focuses on perceiving the visual quality of images [26, 41, 66, 84], and is widely adopted in various application scenarios, e.g., image com" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 273, + 403, + 337 + ], + "blocks": [ + { + "bbox": [ + 322, + 273, + 403, + 337 + ], + "lines": [ + { + "bbox": [ + 322, + 273, + 403, + 337 + ], + "spans": [ + { + "bbox": [ + 322, + 273, + 403, + 337 + ], + "type": "image", + "image_path": "9c3d5035d233ce48d76be96cbb7e2a3f6b830219e529a2f8c65952b102abf3e5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 406, + 274, + 473, + 335 + ], + "blocks": [ + { + "bbox": [ + 406, + 274, + 473, + 335 + ], + "lines": [ + { + "bbox": [ + 406, + 274, + 473, + 335 + ], + "spans": [ + { + "bbox": [ + 406, + 274, + 473, + 335 + ], + "type": "image", + "image_path": "bc8d52c5ed98902340f73dffdd37719e2c097a2b605a4a9483a026b1587f3ad4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 477, + 274, + 545, + 335 + ], + "blocks": [ + { + "bbox": [ + 477, + 274, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 477, + 274, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 477, + 274, + 545, + 335 + ], + "type": "image", + "image_path": "115dd61a17c167a6b0fa757ebf7934db7d109da1216511689698f7c84fd62d72.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 321, + 337, + 402, + 404 + ], + "blocks": [ + { + "bbox": [ + 321, + 337, + 402, + 404 + ], + "lines": [ + { + "bbox": [ + 321, + 337, + 402, + 404 + ], + "spans": [ + { + "bbox": [ + 321, + 337, + 402, + 404 + ], + "type": "image", + "image_path": "b7c00aeb46368cd6361ea1728f51370187a326ab8d512615c08d2b3733d753bd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 406, + 337, + 473, + 403 + ], + "blocks": [ + { + "bbox": [ + 406, + 337, + 473, + 403 + ], + "lines": [ + { + "bbox": [ + 406, + 337, + 473, + 403 + ], + "spans": [ + { + "bbox": [ + 406, + 337, + 473, + 403 + ], + "type": "image", + "image_path": "68cb15a2ddd2e1f44ca67c4db6f133fe9877e832352fd6ea2957f4780c657fbb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 477, + 337, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 477, + 337, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 477, + 337, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 477, + 337, + 545, + 403 + ], + "type": "image", + "image_path": "d80734b4ead7143813d3aa8d0a8a2ec9d4eacd184ffb73d35b65b5fdd5e9ca17.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 322, + 404, + 402, + 472 + ], + "blocks": [ + { + "bbox": [ + 322, + 404, + 402, + 472 + ], + "lines": [ + { + "bbox": [ + 322, + 404, + 402, + 472 + ], + "spans": [ + { + "bbox": [ + 322, + 404, + 402, + 472 + ], + "type": "image", + "image_path": "759e462d9e30c7c8dd92e2e2369bbee0dc69ade76285eb83e82af4777da3ed09.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 477, + 555, + 598 + ], + "lines": [ + { + "bbox": [ + 313, + 477, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 555, + 598 + ], + "type": "text", + "content": "Figure 1. Illustration of our distribution-based soft label. Image A is overexposed, while Image B and C are well-exposed with better and similar quality. \"exce.\" means \"excellent\". To train an MLLM as a quality scorer, continuous scores need to be discretized to discrete level tokens as the training label. Q-Align [70] discretizes human-labeled mean score to a one-hot label, causing information loss. We discretize the score distribution, which is approximated as a Gaussian distribution, to obtain a more accurate soft label. Our soft label maintains better inter-image relationships that the Image B is better than A (A < B), and Image B and C share similar quality (B ≈ C), while the one-hot label can not." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 406, + 404, + 473, + 472 + ], + "blocks": [ + { + "bbox": [ + 406, + 404, + 473, + 472 + ], + "lines": [ + { + "bbox": [ + 406, + 404, + 473, + 472 + ], + "spans": [ + { + "bbox": [ + 406, + 404, + 473, + 472 + ], + "type": "image", + "image_path": "23c670d66a7d30f9dc0cadb426387dd71bda4645f463c8d68283befc68fea9dd.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 477, + 404, + 545, + 472 + ], + "blocks": [ + { + "bbox": [ + 477, + 404, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 477, + 404, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 477, + 404, + 545, + 472 + ], + "type": "image", + "image_path": "3350e31f5688ab3c66084a6c2ca00e59dd95afa04748f6715069c00ad613a883.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "pression and transmission [5, 52], processing smartphone-captured photos [17], and evaluating generation or restoration models [10, 12, 30, 99], etc. With the emergence of Multi-modal Large Language Models (MLLMs) [32, 44, 75, 95], MLLM-based IQA begins to attract more research interests [7, 69, 71, 79, 80]. These methods mainly leverage MLLMs to describe image quality using language, recognizing that language better mirrors human expression, and captures the multifaceted nature of IQA tasks [80]." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14483" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "However, for quality scoring, which needs an accurate numerical score to quantify an image's quality, existing description-centered MLLMs [69, 71, 79, 80] perform less accurately than traditional IQA methods. This hinders their applications on real-world IQA scenarios. Considering MLLMs' good generalization ability on various downstream tasks and promising performance on quality description, it is possible for MLLMs to score the image quality. Therefore, in this work, we aim to explore how MLLMs can effectively regress accurate quality scores by leveraging their broad training and inherent adaptability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 397 + ], + "type": "text", + "content": "We find that the principal challenge in training MLLMs to predict accurate quality scores is the distribution gap between MLLMs' outputs and quality scores. MLLMs generate outputs as discrete tokens, while the quality score is inherently continuous, usually approximated to a Gaussian distribution. Previous approaches, like Q-Align [70], attempt to address the distribution gap by directly discretizing the mean score (i.e., the expectation of estimated score distribution) into a one-hot label, as depicted in Fig. 1. However, such a one-hot discretization method introduces information loss, causing mismatches between the discretized labels and the original continuous scores. For instance, in Fig. 1, Image A is overexposed with a lower mean score than Image B, but their one-hot labels appear identical. Conversely, Image B and C are well-exposed with similar quality scores, yet their discretized labels are different." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 400, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 591 + ], + "type": "text", + "content": "To overcome the distribution gap, we propose a novel soft label to bridge continuous and discrete distributions. As illustrated in Fig. 1, instead of discretizing the mean of score distribution, we discretize the continuous score distribution itself into a discrete distribution of different levels. During training, this discrete distribution is used as the soft training label. This enables effective training with minimal information loss. Also, our distribution-based soft label better preserves inter-image relationship. For example, in Fig. 1, Image B is better than Image A, and Image B and C share similar quality, while the simple one-hot label fails to capture such relationships. At inference, MLLMs output discrete token results, which can be recovered to continuous score distribution. This method ensures that the discrete outputs align well with the underlying continuous distribution, allowing to accurately predict the desired score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Another key challenge is the dataset variation, where different IQA datasets exhibit distinct distributions, making it hard to simply combine them for co-training. Due to this variation, images from different datasets may have nearly the same quality scores (i.e., linearly re-scaled scores) but drastically different perceptual quality [87, 98]. A previous method, UNIQUE [87], has demonstrated that the fidelity loss [63] under the Thurstone's model [60] can effectively capture the intra-dataset relationships, thus facilitating the joint training of multiple IQA datasets. However," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "the fidelity loss requires the model to predict an accurate score distribution, which MLLMs trained with the one-hot label cannot provide. Fortunately, MLLMs trained with our distribution-based soft label can predict such an accurate score distribution, allowing the adaptation of the fidelity loss. We apply this fidelity loss to our MLLM-based quality scorer and demonstrate its effectiveness in this new context." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "type": "text", + "content": "Equipped with these designs, we successfully develop a distribution-based multi-modal image quality assessment model, termed DeQA-Score. DeQA-Score achieves better score regression results than baseline methods on both in-distribution (e.g., " + }, + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "type": "text", + "content": " in PLCC on KonIQ [21]) and out-of-distribution datasets (e.g., " + }, + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 313, + 157, + 553, + 312 + ], + "type": "text", + "content": " in PLCC on LIVEWILD [18]). The advantage of our DeQA-Score becomes more pronounced with the co-training across multiple IQA datasets. Moreover, our method can further predict the distribution of quality scores, which is closely aligned with the distribution derived from extensive human annotations. We hope that our DeQA-Score will inspire more research to explore the potential of quality score distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 320, + 405, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 320, + 405, + 333 + ], + "spans": [ + { + "bbox": [ + 314, + 320, + 405, + 333 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 341, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 553, + 389 + ], + "type": "text", + "content": "Image Quality Assessment (IQA) methods mainly rely on quality scores to assess image quality. They can be categorized into full-reference and non-reference methods through whether there is a high-quality reference image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 389, + 553, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 553, + 519 + ], + "type": "text", + "content": "Full-reference IQA methods compute a similarity score as the quality score between a distorted image and a high-quality reference. Classical works rely on human-designed metrics such as structural similarity [66], phase congruency with gradient magnitude [82], image information [53], etc. The rapid advancement of deep learning has also inspired learning-based IQA methods that measure image quality through data-driven training. Pioneered by LPIPS [84] and PieAPP [48], driven by large-scale IQA datasets [24, 31, 52, 84], deep-learning methods [4, 6, 8, 15, 16, 19, 76, 94] have exhibited quite high accuracy in quality score regression." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 639 + ], + "type": "text", + "content": "Non-reference IQA methods directly regress a quality score without a reference image. Initially, hand-crafted natural image statistics are adopted [37, 40-43, 50, 58]. Subsequently, deep-learning-based methods [25, 34, 46, 55, 56, 78, 88, 93, 96] replace hand-crafted statistics by learning quality priors from human-annotated IQA datasets [17, 21, 30, 31]. Recent works further improve the regression accuracy by introducing multi-dataset co-training strategy [87], multi-scale features [26], CLIP pre-training [64], multidimension attention [73], multitask learning [89], and so on." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Multi-modal Large Language Models (MLLMs) incorporate visual modality into Large Language Models (LLMs) [1, 13, 59, 61] to achieve general visual ability by leveraging the broad training and powerful generalization ability of LLMs. Existing MLLMs [3, 14, 32, 44, 67, 74, 77, 83, 85, 95] have demonstrated a general visual ability" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14484" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "and can tackle a variety of multi-modality tasks, like visual question answering [20, 35, 36], document understanding [38, 39, 54], image captioning [2, 9, 81], etc. Given MLLMs' advanced performance in these high-level perception tasks, it is natural to explore how MLLMs perform in low-level perception tasks that are highly related to IQA." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 295, + 361 + ], + "type": "text", + "content": "MLLM-based IQA methods leverages MLLMs' foundational knowledge to achieve better IQA performance or more detailed assessment results [72, 92]. Q-Bench [68, 91] shows that general MLLMs have some low-level perception ability. [97] evaluates various MLLMs on the widely-adopted Two-Alternative Forced Choice (2AFC) task. Q-Instruct [69] enhances the low-level perception ability of MLLMs by introducing a large-scale dataset. CoInstruct [71] concentrates on the quality comparison among multiple images. DepictQA [79, 80] performs single-image assessment and paired-image comparison in both full-reference and non-reference settings. Q-Ground [7] focuses on the visual quality grounding task, and performs fine-scale detailed visual quality analysis. Some of these works have proposed MLLM-based quality scoring methods, e.g., binary softmax strategy in [68, 91] and pair-wise voting in [79, 80]. Nonetheless, these works principally focus on the quality description ability using languages." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 363, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 531 + ], + "type": "text", + "content": "MLLM-based quality scorer. Initially, Q-Bench [68, 91] proposes a binary softmax strategy, enabling MLLMs to generate quality scores by predicting two discrete quality levels (i.e., good or poor). This strategy is adopted by Q-Instruct [69] and Co-Instruct [71]. Compare2Score [98] achieves quality scoring by teaching MLLMs to compare pair-wise images. Inspired by humans' annotation process, Q-Align [70] discretizes scores to five discrete levels using one-hot label to train MLLMs, resulting in more accurate score regression. Dog-IQA [33] adopts the one-hot label for training-free IQA by incorporating specific standards and local semantic objects. However, as discussed in Sec. 1, one-hot label causes some information loss and can not fully maintain the inter-image relationship." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 541, + 194, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 194, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 194, + 553 + ], + "type": "text", + "content": "3. Soft Label Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 560, + 195, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 195, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 195, + 573 + ], + "type": "text", + "content": "3.1. Revisiting One-hot Label" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 578, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 686 + ], + "type": "text", + "content": "Training with one-hot label. MLLMs only accept discrete tokens as inputs and outputs. Therefore, as shown in Fig. 2a, continuous scores1 must be discretized into discrete tokens to train the MLLMs. Q-Align [70] adopts a one-hot discretization method. Specifically, as depicted in Fig. 2b, the score range [1, 5] is uniformly divided into five intervals. These intervals represent five rating levels, {\"bad\", \"poor\", \"fair\", \"good\", \"excellent\"}, which is the rating standard defined by [51]. In Fig. 2a, the Mean Opinion Score (MOS)" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 414, + 131 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 414, + 131 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 414, + 131 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 414, + 131 + ], + "type": "image", + "image_path": "ca9f5b35f46c8c79885aa840602e14e4f79c4e744d4f73ecb592170353afc3da.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 76, + 504, + 84 + ], + "lines": [ + { + "bbox": [ + 417, + 76, + 504, + 84 + ], + "spans": [ + { + "bbox": [ + 417, + 76, + 504, + 84 + ], + "type": "text", + "content": "Rate the quality of this image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 122, + 331, + 131 + ], + "lines": [ + { + "bbox": [ + 321, + 122, + 331, + 131 + ], + "spans": [ + { + "bbox": [ + 321, + 122, + 331, + 131 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 110, + 523, + 118 + ], + "lines": [ + { + "bbox": [ + 420, + 110, + 523, + 118 + ], + "spans": [ + { + "bbox": [ + 420, + 110, + 523, + 118 + ], + "type": "text", + "content": "5 candidate level tokens of :" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 416, + 87, + 543, + 101 + ], + "blocks": [ + { + "bbox": [ + 416, + 87, + 543, + 101 + ], + "lines": [ + { + "bbox": [ + 416, + 87, + 543, + 101 + ], + "spans": [ + { + "bbox": [ + 416, + 87, + 543, + 101 + ], + "type": "image", + "image_path": "4229c4b759b0484b598fc30218bb6045d56e7434fa040201503b2aadcc619013.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 135, + 438, + 178 + ], + "blocks": [ + { + "bbox": [ + 422, + 118, + 539, + 128 + ], + "lines": [ + { + "bbox": [ + 422, + 118, + 539, + 128 + ], + "spans": [ + { + "bbox": [ + 422, + 118, + 539, + 128 + ], + "type": "text", + "content": "[ bad, poor, fair, good, excellent ]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 135, + 438, + 178 + ], + "lines": [ + { + "bbox": [ + 323, + 135, + 438, + 178 + ], + "spans": [ + { + "bbox": [ + 323, + 135, + 438, + 178 + ], + "type": "image", + "image_path": "85607b301653b1169f3e87de3fa19289e5db6691da3367bbc06d6b436ce25a8a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 443, + 135, + 545, + 180 + ], + "blocks": [ + { + "bbox": [ + 443, + 135, + 545, + 180 + ], + "lines": [ + { + "bbox": [ + 443, + 135, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 443, + 135, + 545, + 180 + ], + "type": "image", + "image_path": "6282334889ff01d98345e4281bb475fb11e76726f56abadb26e95fe41a8faeef.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 322, + 182, + 438, + 227 + ], + "blocks": [ + { + "bbox": [ + 322, + 182, + 438, + 227 + ], + "lines": [ + { + "bbox": [ + 322, + 182, + 438, + 227 + ], + "spans": [ + { + "bbox": [ + 322, + 182, + 438, + 227 + ], + "type": "image", + "image_path": "b6ce449d04ec5b3a3e190fa4076ec1dba39143bb254b5a5bb4d57c0f12fcd496.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 230, + 331, + 239 + ], + "lines": [ + { + "bbox": [ + 321, + 230, + 331, + 239 + ], + "spans": [ + { + "bbox": [ + 321, + 230, + 331, + 239 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 230, + 419, + 238 + ], + "lines": [ + { + "bbox": [ + 361, + 230, + 419, + 238 + ], + "spans": [ + { + "bbox": [ + 361, + 230, + 419, + 238 + ], + "type": "text", + "content": "Continuous score" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 443, + 182, + 545, + 228 + ], + "blocks": [ + { + "bbox": [ + 443, + 182, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 443, + 182, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 443, + 182, + 545, + 228 + ], + "type": "image", + "image_path": "40446f0e90d988740585592af2819b30f8e65aea7a21c3f2cc2bf85f3227c8f0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 230, + 525, + 238 + ], + "lines": [ + { + "bbox": [ + 470, + 230, + 525, + 238 + ], + "spans": [ + { + "bbox": [ + 470, + 230, + 525, + 238 + ], + "type": "text", + "content": "Discretized label" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 245, + 553, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 354 + ], + "type": "text", + "content": "Figure 2. (a) MLLMs need discrete level tokens for training. Thus, continuous scores must be converted into discrete level tokens as training labels. (b) Comparison between one-hot label [70] and our soft label. One-hot label uniformly divides the score range into 5 intervals to discretize the mean score, i.e., the expectation of estimated score distribution. One-hot label assumes that different levels are identically independent, which is inaccurate. In contrast, we discretize the estimated Gaussian distribution of the score. Our soft label better preserves the relationships between levels, e.g., \"fair\" is closer to \"good\" than \"excellent\"." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 361, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 434 + ], + "type": "text", + "content": "of the image, 4.30, fails into the fifth intervals, resulting in the discretized label \"excellent\". Accordingly, the response is formulated as \"The quality of this image is excellent\". Using the response as the training labels, following [13, 61], MLLMs can be trained with next token prediction loss, i.e., cross-entropy loss between predicted logits and labels." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 435, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 663 + ], + "type": "text", + "content": "Limitations of one-hot label. First, ideal training with cross-entropy loss using one-hot labels assumes that the five levels are independent (i.e., orthogonal in a latent space). However, intuitively, these levels are not truly independent, e.g., \"fair\" is closer to \"good\" than \"excellent\". Second, one-hot discretization only considers the mean score, which is the expectation of the score distribution, ignoring the distribution's uncertainty (i.e., variance). Third, near the border of two levels, images with similar quality can be divided into different levels. For example, in Fig. 1, Image B (MOS 3.38) and Image C (MOS 3.49) share similar mean scores, but are differently classified as \"fair\" and \"good\", due to their proximity to the border of these two levels, i.e., 3.4 in Fig. 2b. Finally, one-hot label causes information loss. In Q-Align, the five levels are respectively assigned scores " + }, + { + "bbox": [ + 313, + 435, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\{1, 2, 3, 4, 5\\}" + }, + { + "bbox": [ + 313, + 435, + 555, + 663 + ], + "type": "text", + "content": ", with which we calculate the discretization precision in Tab. 1. The discretization error is around " + }, + { + "bbox": [ + 313, + 435, + 555, + 663 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 313, + 435, + 555, + 663 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 435, + 555, + 663 + ], + "type": "inline_equation", + "content": "35 \\times" + }, + { + "bbox": [ + 313, + 435, + 555, + 663 + ], + "type": "text", + "content": " larger than our soft label that we will soon introduce, thus the one-hot label keeps lower correlation with MOS." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 671, + 385, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 385, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 385, + 682 + ], + "type": "text", + "content": "3.2. Soft Label" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Discretization using score distribution. As demonstrated in [87], the quality score of one image, " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", is typically mod-" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1 Unless otherwise specified, the mean quality scores in all datasets will be normalized to [1, 5], and the variances will be normalized accordingly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14485" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 128, + 296, + 218 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": "Table 1. Comparison of discretization precision between Q-Align's one-hot label [70] and our soft label. The precision of MOS discretization is measured by L1 Error / RMSE and PLCC / SRCC. The distance of original distribution and recovered distribution is measured by JS divergence / Wasserstein distance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 128, + 296, + 218 + ], + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 218 + ], + "type": "table", + "html": "
MetricsDiscretizationKonIQ [21]SPAQ [17]KADID [31]
L1 / RSMEOne-hot label0.302 / 0.3740.299 / 0.3660.327 / 0.389
Soft label0.008 / 0.0180.011 / 0.0250.024 / 0.038
PLCC / SRCCOne-hot label0.961 / 0.9520.969 / 0.9680.979 / 0.982
Soft label1.000 / 1.0001.000 / 1.0001.000 / 1.000
JS / W-Dist.One-hot label---
Soft label0.001 / 0.0380.022 / 0.0840.010 / 0.140
", + "image_path": "77b37992814759cf5537e64fac46787714d569c8b9ac43ff268ec4204d7416f8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": "eled as a Gaussian distribution, " + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "inline_equation", + "content": "x \\sim \\mathcal{N}(\\mu, \\sigma^2)" + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": " are the Mean Opinion Score (MOS) and variance of multiple human annotations, respectively. The distribution is represented by its probability density function, " + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": ". As shown in the second row of Fig. 2b, we select five central points, " + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "inline_equation", + "content": "c_i \\in \\{1, 2, 3, 4, 5\\}" + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "inline_equation", + "content": "i \\in \\{0, 1, 2, 3, 4\\}" + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": ", as the centers of the five discrete levels. The width of each level region, " + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": ", is set as 1. The possibility of the score falling into the " + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": " level is denoted as " + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "inline_equation", + "content": "p_i^{raw}" + }, + { + "bbox": [ + 55, + 221, + 296, + 331 + ], + "type": "text", + "content": " and is calculated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 335, + 295, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 335, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 80, + 335, + 295, + 365 + ], + "type": "interline_equation", + "content": "p _ {i} ^ {r a w} = \\int_ {c _ {i} - d / 2} ^ {c _ {i} + d / 2} f (x) \\mathrm {d x}, \\quad i \\in \\{0, 1, 2, 3, 4 \\}. \\tag {1}", + "image_path": "47d4300d2650313fc879e0da8a6346a454f212b76ce28ec9198c9674cf215de6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "content": "We assign the five levels with the same textual tokens as Q-Align, i.e., {\"bad\", \"poor\", \"fair\", \"good\", \"excellent)}. Correspondingly, the score of " + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "content": " level is " + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "content": ". This defines a rough discrete distribution as " + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "inline_equation", + "content": "P^{raw}(c_i) = p_i^{raw}" + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "text", + "content": "Post-adjustment. The discrete distribution is rough because of truncation errors. As shown in the second row of Fig. 2b, there are two truncation regions (i.e., " + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "inline_equation", + "content": "x < c_0 - d/2 = 0.5" + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "inline_equation", + "content": "x > c_4 + d/2 = 5.5" + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "text", + "content": "), which are not considered in Eq. (1). This can cause two problems. First, the sum of " + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\{p_i^{raw}\\}" + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "text", + "content": " does not equal 1, meaning " + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "inline_equation", + "content": "p_i^{raw}" + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "text", + "content": " is not strictly a discrete distribution. Second, the expectation of this discrete distribution does not equal the expectation of score distribution, " + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 418, + 296, + 526 + ], + "type": "text", + "content": ", leading to discretization errors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 526, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 296, + 563 + ], + "type": "text", + "content": "To correct these truncation errors, we propose a postadjustment method. Specifically, we apply a linear transformation to adjust " + }, + { + "bbox": [ + 55, + 526, + 296, + 563 + ], + "type": "inline_equation", + "content": "p_i^{raw}" + }, + { + "bbox": [ + 55, + 526, + 296, + 563 + ], + "type": "text", + "content": " to more accurate " + }, + { + "bbox": [ + 55, + 526, + 296, + 563 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 526, + 296, + 563 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 567, + 201, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 201, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 201, + 581 + ], + "type": "interline_equation", + "content": "p _ {i} = \\alpha p _ {i} ^ {r a w} + \\beta", + "image_path": "63b5ce5638e13960f6a2b7889b353e4b2d9be132c6da711f54f9bbbf5aa98e5b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 87, + 583, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 583, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 87, + 583, + 295, + 602 + ], + "type": "interline_equation", + "content": "\\text {s . t .} \\quad \\sum_ {i} p _ {i} = 1, \\quad \\mu^ {r e c} = \\sum_ {i} p _ {i} c _ {i} = \\mu , \\tag {2}", + "image_path": "be0d89a01fbf841a52cef1238b4edc9f7bba2b04ee8a0e0ed136f11a0f206525.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu_{rec}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " means recovered expectation from discrete distribution. The parameters of the linear transformation, " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", can be determined by solving these two constraint conditions. One example of post-adjustment is given in Fig. 3. Moreover, the statistics of these two parameters in Tab. 2 show that the average values of " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " are quite close to 1 and 0, respectively, implying that the post-adjustment is minimal. This effectively defines our soft label, which follows the discrete distribution as " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "P(c_i) = p_i" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 411, + 140 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 411, + 140 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 411, + 140 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 411, + 140 + ], + "type": "image", + "image_path": "9c9d680b802574353b486b41d6d0bdd4c49ba5f3ed9748fce5bcef0953a96916.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 143, + 414, + 166 + ], + "lines": [ + { + "bbox": [ + 315, + 143, + 414, + 166 + ], + "spans": [ + { + "bbox": [ + 315, + 143, + 414, + 166 + ], + "type": "text", + "content": "Figure 3. Illustration of the soft label adjustment." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 414, + 128, + 553, + 164 + ], + "blocks": [ + { + "bbox": [ + 414, + 69, + 554, + 125 + ], + "lines": [ + { + "bbox": [ + 414, + 69, + 554, + 125 + ], + "spans": [ + { + "bbox": [ + 414, + 69, + 554, + 125 + ], + "type": "text", + "content": "Table 2. Statistics of the label adjustment parameters. Averaged " + }, + { + "bbox": [ + 414, + 69, + 554, + 125 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 414, + 69, + 554, + 125 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 414, + 69, + 554, + 125 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 414, + 69, + 554, + 125 + ], + "type": "text", + "content": " are very approaching 1 and 0, respectively, meaning that the postadjustment is quite small." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 414, + 128, + 553, + 164 + ], + "lines": [ + { + "bbox": [ + 414, + 128, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 414, + 128, + 553, + 164 + ], + "type": "table", + "html": "
KonIQ [21]SPAQ [17]KADID [31]
α1.036±0.0411.033±0.0851.118±0.091
β-0.005±0.005-0.005±0.014-0.015±0.010
", + "image_path": "1b5ed3fbbe8b423313f8b6d1a089ad6b6fcd371f258caa6a1df2ce7c2294a11f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 172, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 293 + ], + "type": "text", + "content": "Ideally, the discretization error of the MOS will be 0. However, the possibility of some levels may slightly fall below zero after the linear transformation. We address this by simply clipping these values to zero. The impact of this clipping on accuracy is negligible, as shown in Tab. 1. After clipping, our discretization error for MOS remains around " + }, + { + "bbox": [ + 313, + 172, + 555, + 293 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 313, + 172, + 555, + 293 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 172, + 555, + 293 + ], + "type": "inline_equation", + "content": "35 \\times" + }, + { + "bbox": [ + 313, + 172, + 555, + 293 + ], + "type": "text", + "content": " smaller than Q-Align's one-hot label. Also, our soft label maintains a perfect (i.e., 1.000) Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank-order Correlation Coefficient (SRCC) with the original MOS." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 294, + 553, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 306 + ], + "type": "text", + "content": "Recovery from discrete to continuous score distribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 307, + 554, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 367 + ], + "type": "text", + "content": "As shown in the second condition of Eq. (2), we can compute the expectation of discrete distribution, " + }, + { + "bbox": [ + 313, + 307, + 554, + 367 + ], + "type": "inline_equation", + "content": "P(c_{i})" + }, + { + "bbox": [ + 313, + 307, + 554, + 367 + ], + "type": "text", + "content": ", as the recovered expectation of original distribution. Besides the expectation, we can also recover the original variance using the variance of the discrete distribution, " + }, + { + "bbox": [ + 313, + 307, + 554, + 367 + ], + "type": "inline_equation", + "content": "P(c_{i})" + }, + { + "bbox": [ + 313, + 307, + 554, + 367 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 374, + 555, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 374, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 324, + 374, + 555, + 393 + ], + "type": "interline_equation", + "content": "\\mu^ {r e c} = \\sum_ {i} p _ {i} c _ {i}, \\quad (\\sigma^ {r e c}) ^ {2} = \\sum_ {i} p _ {i} (c _ {i} - \\mu^ {r e c}) ^ {2}. \\quad (3)", + "image_path": "ae463b3f3544df7dee164344891889464117966b399efb8c894978a367c9817b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 399, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 459 + ], + "type": "text", + "content": "Then we calculate the distance between recovered distribution, " + }, + { + "bbox": [ + 313, + 399, + 554, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu^{rec}, (\\sigma^{rec})^2)" + }, + { + "bbox": [ + 313, + 399, + 554, + 459 + ], + "type": "text", + "content": ", and original distribution, " + }, + { + "bbox": [ + 313, + 399, + 554, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu, \\sigma^2)" + }, + { + "bbox": [ + 313, + 399, + 554, + 459 + ], + "type": "text", + "content": " as provided in Tab. 1. Our recovered distribution is closely aligned with human-annotated score distribution. In contrast, one-hot label can not recover such a distribution." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 468, + 430, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 430, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 430, + 483 + ], + "type": "text", + "content": "4. DeQA-Score Model" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 488, + 474, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 474, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 474, + 502 + ], + "type": "text", + "content": "4.1. Training with Our Soft Label" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 506, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 614 + ], + "type": "text", + "content": "Model architecture. As illustrated in Fig. 4, we follow the widely adopted MLLM architecture [32, 75] to construct our model. First, a vision encoder is adopted to represent the input image to visual tokens. Then, a vision-to-text connector projects the visual tokens into the textual space. Following [75], we use a vision abstractor as the connector. The abstractor can further reduce the number of visual tokens to 64. Finally, the visual and textual tokens are fused together and fed into an LLM for response prediction." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": "Next token prediction loss for common response tokens. As described in Sec. 3.1 and illustrated in Fig. 4, the constructed response follows the template, \"The quality of this image is " + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": " level " + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": "\". For common response tokens other than the level token, we compute the next token prediction loss, i.e., the cross-entropy loss " + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": ", following [45, 61]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "KL divergence loss for the level token. There are five candidate values for the level token: {\"bad\", \"poor\", \"fair\"," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14486" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 539, + 178 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 539, + 178 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 539, + 178 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 539, + 178 + ], + "type": "image", + "image_path": "034b24dc80d8eddb0c048dd00428bc8c05d4ad8297bdf2b6cfd8595889e0d82c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 180, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 555, + 205 + ], + "type": "text", + "content": "Figure 4. Framework of our DeQA-Score trained with soft label. For the “” token, the KL divergence loss is calculated between predicted probabilities and our soft label. For other tokens, common cross-entropy loss for next token prediction is calculated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 295, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 246 + ], + "type": "text", + "content": "\"good\", \"excellent\". Our soft label, " + }, + { + "bbox": [ + 55, + 209, + 295, + 246 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 209, + 295, + 246 + ], + "type": "text", + "content": ", assigns a ground-truth probability to each value. Let the predicted probabilities be " + }, + { + "bbox": [ + 55, + 209, + 295, + 246 + ], + "type": "inline_equation", + "content": "p_i^{pred}" + }, + { + "bbox": [ + 55, + 209, + 295, + 246 + ], + "type": "text", + "content": ", the KL divergence loss is calculated through:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 255, + 295, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 255, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 116, + 255, + 295, + 274 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k l} = \\sum_ {i} p _ {i} \\log \\left(p _ {i} ^ {\\text {p r e d}} / p _ {i}\\right). \\tag {4}", + "image_path": "7f5cbe0ad2fd8200127c345203a1d2d0ebaf36831bf5e1f77124a8a14c344421.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 280, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 353 + ], + "type": "text", + "content": "The predicted probabilities, " + }, + { + "bbox": [ + 55, + 280, + 296, + 353 + ], + "type": "inline_equation", + "content": "p_i^{pred}" + }, + { + "bbox": [ + 55, + 280, + 296, + 353 + ], + "type": "text", + "content": ", are obtained by applying a softmax activation function on model's output logits. Note that during training, this softmax function is applied across all textual tokens, not just across the five predefined levels. This approach encourages the model to reduce the possibilities of textual tokens outside the five levels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "text", + "content": "The two terms of losses, " + }, + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kl}" + }, + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "text", + "content": ", are both autoregressive losses, and thus can be easily combined. During training, the cross-entropy loss, " + }, + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 55, + 353, + 296, + 413 + ], + "type": "text", + "content": ", will decrease near to 0 fast, since it is quite easy to optimize by simply remembering the response template \"The quality of this image is\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 421, + 240, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 240, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 240, + 434 + ], + "type": "text", + "content": "4.2. Score Estimation during Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "type": "text", + "content": "During inference, using the model-predicted probabilities of the five levels, " + }, + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "type": "inline_equation", + "content": "p_i^{pred}" + }, + { + "bbox": [ + 55, + 436, + 295, + 472 + ], + "type": "text", + "content": ", we can estimate the expectation and variance of the score distribution the same as Eq. (3):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 476, + 295, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 476, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 60, + 476, + 295, + 493 + ], + "type": "interline_equation", + "content": "\\mu^ {p r e d} = \\sum_ {i} p _ {i} ^ {p r e d} c _ {i}, (\\sigma^ {p r e d}) ^ {2} = \\sum_ {i} p _ {i} ^ {p r e d} (c _ {i} - \\mu^ {p r e d}) ^ {2}. (5)", + "image_path": "5b4ccff098b1249346e8bfe5c5bd0a8374255730a1cc19c323b44ba028b737df.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 496, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 593 + ], + "type": "text", + "content": "Here we follow Q-Align [70] to apply a close-set softmax on the five levels to obtain the probabilities, " + }, + { + "bbox": [ + 55, + 496, + 295, + 593 + ], + "type": "inline_equation", + "content": "p_i^{pred}" + }, + { + "bbox": [ + 55, + 496, + 295, + 593 + ], + "type": "text", + "content": ", to prevent the influence of other textual tokens. Our statistics in Tab. A2 reveal that the trained MLLM will always predict the five levels, thus applying softmax on either the five levels or all textual tokens yields nearly identical results. As given in Tab. 5, our predicted score distribution, " + }, + { + "bbox": [ + 55, + 496, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu^{pred}, (\\sigma^{pred})^2)" + }, + { + "bbox": [ + 55, + 496, + 295, + 593 + ], + "type": "text", + "content": ", aligns well with human annotations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 600, + 269, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 269, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 269, + 613 + ], + "type": "text", + "content": "4.3. Joint Training on Multiple IQA Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "As demonstrated in [87, 98], different IQA datasets exhibit distinct distributions. Due to this dataset variation, images from different datasets with nearly identical scores (i.e., linearly re-scaled scores) can have drastically different perceptual quality (see examples in Fig. A1). The reason is that different datasets have different reflections between quality scores and visual quality. This disparity makes it infeasible to merely regress absolute quality scores for multi-dataset" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 208, + 547, + 326 + ], + "blocks": [ + { + "bbox": [ + 321, + 208, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 321, + 208, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 321, + 208, + 547, + 326 + ], + "type": "image", + "image_path": "861682710f7e97a603ccd26e55fbaae04cbfe1c32a455b1e1a02c1c4897b9183.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 329, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 429 + ], + "type": "text", + "content": "Figure 5. Fidelity loss to maintain intra-dataset relationships. Consider two images, Image A and B, sampled from the same dataset. Using our distribution-based soft label, DeQA-Score predicts the expectation, " + }, + { + "bbox": [ + 313, + 329, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\mu^{pred}" + }, + { + "bbox": [ + 313, + 329, + 555, + 429 + ], + "type": "text", + "content": ", and variance, " + }, + { + "bbox": [ + 313, + 329, + 555, + 429 + ], + "type": "inline_equation", + "content": "(\\sigma^{pred})^2" + }, + { + "bbox": [ + 313, + 329, + 555, + 429 + ], + "type": "text", + "content": ", of each image's score distribution, with which we can predict the probability that Image A is better than Image B, " + }, + { + "bbox": [ + 313, + 329, + 555, + 429 + ], + "type": "inline_equation", + "content": "P(A > B)" + }, + { + "bbox": [ + 313, + 329, + 555, + 429 + ], + "type": "text", + "content": ". This predicted probability is then used to calculate the fidelity loss with the ground-truth probability. Fidelity loss preserves intra-dataset relationships, facilitating the co-training across multiple IQA datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 435, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 615 + ], + "type": "text", + "content": "co-training. To address this challenge, UNIQUE [87] applies fidelity loss [63] based on Thurstone's model [60]. The fidelity loss is based on the principal that IQA emphasizes the ranking relationships within the same dataset, rather than the absolute values of quality scores. Specifically, two images are sampled from the same dataset, whose quality scores can be comparable. These two images are then compared using human-annotated quality scores. The fidelity loss encourages the model to learn such comparison results, which facilitates the model to focus more on the ranking relationships in each dataset, instead of the absolute values of quality scores. Thus, the fidelity loss helps multi-dataset co-training. Our soft label is naturally compatible with the fidelity loss, since our method can predict the score distribution, which is what the fidelity loss needs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Ground-truth construction. As shown in Fig. 5, a pair of images, Image A and B, are sampled from the same dataset. We assume that the quality scores of the two images, " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_A, x_B" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", follow the Gaussian distribution with means and variances collected by subjective testing, i.e., " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_A \\sim \\mathcal{N}(\\mu_A, \\sigma_A^2)" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_B \\sim \\mathcal{N}(\\mu_B, \\sigma_B^2)" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ". Assuming that their quality scores are uncorrelated, the score difference, " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_A - x_B" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", also follows a Gaussian distribution, " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_A - \\mu_B, \\sigma_A^2 + \\sigma_B^2)" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14487" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "Then the possibility that Image A has a better perceptual quality than Image B can be calculated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 102, + 295, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 102, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 68, + 102, + 295, + 135 + ], + "type": "interline_equation", + "content": "p (A > B) = p \\left(x _ {A} > x _ {B}\\right) = \\Phi \\left(\\frac {\\mu_ {A} - \\mu_ {B}}{\\sqrt {\\sigma_ {A} ^ {2} + \\sigma_ {B} ^ {2}}}\\right), \\tag {6}", + "image_path": "b73c497dca960841b2dd6e10eef697f7ba9e44760d1e3dd51d8ffca0d5e4c778.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 142, + 295, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 295, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 295, + 154 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 142, + 295, + 154 + ], + "type": "inline_equation", + "content": "\\Phi (\\cdot)" + }, + { + "bbox": [ + 55, + 142, + 295, + 154 + ], + "type": "text", + "content": " is the Gaussian cumulative distribution function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": "Model prediction. As described in Sec. 4.2, our model can predict the expectation and variance of the score distribution. Denoting the predicted expectation and variance of two images as " + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mu_A^{pred}" + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "inline_equation", + "content": "(\\sigma_A^{pred})^2" + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mu_B^{pred}" + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "inline_equation", + "content": "(\\sigma_B^{pred})^2" + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": ", same to Eq. (6), we can predict the possibility of Image A surpassing Image B in perceptual quality as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 235, + 295, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 235, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 69, + 235, + 295, + 275 + ], + "type": "interline_equation", + "content": "p ^ {p r e d} (A > B) = \\Phi \\left(\\frac {\\mu_ {A} ^ {p r e d} - \\mu_ {B} ^ {p r e d}}{\\sqrt {(\\sigma_ {A} ^ {p r e d}) ^ {2} + (\\sigma_ {B} ^ {p r e d}) ^ {2}}}\\right). \\quad (7)", + "image_path": "683109011cf2c7189a428b30c880eac6f2ede5399d5f51b931d86fba01d7e9f7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 282, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 307 + ], + "type": "text", + "content": "Training with fidelity loss. Following [87], we choose fidelity loss as the similarity measure to train the model as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 324, + 294, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 324, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 71, + 324, + 294, + 366 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {f d} = 1 - \\sqrt {p (A > B) p ^ {p r e d} (A > B)} \\\\ - \\sqrt {\\left(1 - p (A > B)\\right) \\left(1 - p ^ {p r e d} (A > B)\\right)}. \\tag {8} \\\\ \\end{array}", + "image_path": "817afc5f6f63d8838998294aeceffaefacf7f794af1f941bee46ec48113cf446.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 373, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 384 + ], + "type": "text", + "content": "Adding two auto-regressive losses, the final training loss is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 396, + 294, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 396, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 121, + 396, + 294, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {f d} + \\gamma \\left(\\mathcal {L} _ {c e} + \\mathcal {L} _ {k l}\\right), \\tag {9}", + "image_path": "61a4db92a3cbeb96af269ec243c5d0fbc1fd544e3b5395f0bc5b20058d7aebe3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 413, + 264, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 264, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 264, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 264, + 425 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 413, + 264, + 425 + ], + "type": "text", + "content": " is a weighting term to adjust the loss scale." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 454, + 266, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 266, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 266, + 467 + ], + "type": "text", + "content": "5.1. Datasets, Baselines, Metrics, and Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "content": "Datasets. For a comprehensive evaluation, we include a wide range of IQA datasets: (a) in-the-wild images: KonIQ [21], SPAQ [17], LIVE-Wild [18], and FLIVE [78]; (b) synthetic distortions: KADID [31], TID2013 [47], and CSIQ [29]; (c) model-processed distortions: PIPAL [24]; (d) AI-generated images: AGIQA-3K [30]. Among these datasets, KonIQ, SPAQ, KADID, and PIPAL are split into training and test sets. Other datasets are used to assess the out-of-distribution generalization ability. The Mean Opinion Scores (MOS) of these datasets are normalized to the range of [1,5], with the variances normalized accordingly." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Baseline methods. To demonstrate the effectiveness of our distribution-based soft label, we primarily compare with Q-Align [70], which utilizes the one-hot label. The base models for DeQA-Score and Q-Align are kept identical for fair comparison. We also include previous IQA methods as baselines, including handcrafted methods: NIQE [41] and BRISQUE [40], and deep-learning-based methods: NIMA [57], HyperIQA [55], DBCNN [86], MUSIQ [26], CLIP-IQA+ [64], ManIQA [73], and Compare2Score [98]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "Metrics. Following [26, 70, 73], we use the Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank-order Correlation Coefficient (SRCC) as metrics to evaluate the score regression performance. Additionally, we employ KL divergence, JS divergence, and Wasserstein distance to assess the results of score distribution prediction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 146, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 314 + ], + "type": "text", + "content": "Implementation details. We use mPLUG-Owl2 [75] as our base model, with a CLIP-pretrained ViT-L [49] as the image encoder, a six-layer Q-Former as the visual abstractor, and LLaMA-2-7B [62] as the LLM. The number of visual tokens output by the vision abstractor is set to 64, and the hidden dimension of LLM is 4096. The pretrained weights of mPLUG-Owl2 are used for model initialization. The weighting term " + }, + { + "bbox": [ + 313, + 146, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 146, + 555, + 314 + ], + "type": "text", + "content": " is empirically set to 0.05, adjusting the two loss terms roughly to the same scale. We adopt AdamW [23] as the optimizer. The initial learning rate is set to 2e-5 and decays gradually using the cosine decay strategy. Our model is trained with a batch size of 64 for 3 epochs. Using 8 NVIDIA RTX A6000 GPUs, the training process on KonIQ dataset is completed in around 1.5 hours." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 322, + 466, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 466, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 466, + 335 + ], + "type": "text", + "content": "5.2. Results of Score Regression" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 340, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 554, + 483 + ], + "type": "text", + "content": "Score regression results of single-dataset training. We first train our model on a single dataset, i.e., KonIQ, and then evaluate the model on both in-distribution and out-of-distribution datasets, following [28, 70]. The score regression results are shown in Tab. 3. On the in-distribution KonIQ test set, our model achieves superior score regression results compared to all baseline methods, e.g., a " + }, + { + "bbox": [ + 313, + 340, + 554, + 483 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 340, + 554, + 483 + ], + "type": "text", + "content": " improvement in PLCC over Q-Align. Furthermore, on out-of-distribution datasets, our DeQA-Score maintains its advantage with the best overall results across nearly all benchmarks, e.g., a " + }, + { + "bbox": [ + 313, + 340, + 554, + 483 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 313, + 340, + 554, + 483 + ], + "type": "text", + "content": " improvement in PLCC on LIVE-Wild. These findings confirm the effectiveness of our method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 484, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 579 + ], + "type": "text", + "content": "However, we observe that all models perform poorly on the PIPAL dataset. For instance, our best results reach only 0.462 on PLCC. The main reason is the distortion gap between the training dataset, KonIQ (in-the-wild distortions), and PIPAL (model-processed distortions). This highlights the necessity of co-training with multiple IQA datasets containing various distortion types. This observation motivates our subsequent experiments on multi-dataset co-training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Score regression results of multi-dataset co-training. As discussed above, co-training on multiple IQA datasets is necessary to obtain a unified model for various categories of distortions. Therefore, we conduct multi-dataset cotraining experiments in Tab. 4. First, our DeQA-Score consistently outperforms Q-Align across different training combinations. The performance gain attributes to both the soft label and fidelity loss, as confirmed in Tab. 6. Second, comparing #0 to #1, when KADID (synthetic distortions) is added to the training set, the performance on TID2013 and CSIQ (both synthetic distortions) improves" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14488" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 96, + 555, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 3. Score regression results of single-dataset training with the PLCC / SRCC metrics. All methods except handcrafted ones are trained on the KonIQ dataset. DeQA-Score outperforms all baseline methods across nearly all benchmarks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 96, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 96, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 555, + 232 + ], + "type": "table", + "html": "
CategoryMethodsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KCSIQFLIVE
HandcraftedNIQE [41]0.533 / 0.5300.679 / 0.6640.468 / 0.4050.195 / 0.1610.493 / 0.4490.560 / 0.5330.718 / 0.6280.147 / 0.100
BRISQUE [40]0.225 / 0.2260.490 / 0.4060.429 / 0.3560.267 / 0.2320.361 / 0.3130.541 / 0.4970.740 / 0.5560.108 / 0.054
Non-MLLMNIMA [57]0.896 / 0.8590.838 / 0.8560.532 / 0.5350.390 / 0.3990.814 / 0.7710.715 / 0.6540.695 / 0.6490.561 / 0.467
HyperIQA [55]0.917 / 0.9060.791 / 0.7880.506 / 0.4680.410 / 0.4030.772 / 0.7490.702 / 0.6400.752 / 0.7170.485 / 0.383
DBCNN [86]0.884 / 0.8750.812 / 0.8060.497 / 0.4840.384 / 0.3810.773 / 0.7550.730 / 0.6410.586 / 0.5720.485 / 0.385
MUSIQ [26]0.924 / 0.9290.868 / 0.8630.575 / 0.5560.431 / 0.4310.789 / 0.8300.722 / 0.6300.771 / 0.7100.565 / 0.467
CLIP-IQA+ [64]0.909 / 0.8950.866 / 0.8640.653 / 0.6540.427 / 0.4190.832 / 0.8050.736 / 0.6850.772 / 0.7190.427 / 0.316
ManIQA [73]0.849 / 0.8340.768 / 0.7580.499 / 0.4650.457 / 0.4520.849 / 0.8320.723 / 0.6360.623 / 0.6270.512 / 0.401
MLLM-basedCompare2Score [98]0.923 / 0.9100.867 / 0.8600.500 / 0.4530.354 / 0.3420.786 / 0.7720.777 / 0.6710.735 / 0.7050.474 / 0.413
Q-Align [70]0.941 / 0.9400.886 / 0.8870.674 / 0.6840.403 / 0.4190.853 / 0.8600.772 / 0.7350.785 / 0.7370.554 / 0.483
DeQA-Score0.953 / 0.9410.895 / 0.8960.694 / 0.6870.472 / 0.4780.892 / 0.8790.809 / 0.7290.787 / 0.7440.589 / 0.501
", + "image_path": "26d388c9aababa3bae1d671d379fd4bfbcc109fb7d2f794c91b807a4ad58f84f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 261, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 236, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 259 + ], + "type": "text", + "content": "Table 4. Score regression results of co-training on multiple IQA datasets with the PLCC / SRCC metrics. Our DeQA-Score consistently surpasses the baseline method, Q-Align, across all benchmarks with different training combinations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 261, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 56, + 261, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 553, + 350 + ], + "type": "table", + "html": "
#Trained onMethodsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
0KonIQ, SPAQQ-Align [70]0.943 / 0.9400.933 / 0.9310.692 / 0.7080.401 / 0.4110.883 / 0.8790.795 / 0.7270.671 / 0.5680.795 / 0.767
DeQA-Score0.953 / 0.9430.936 / 0.9330.724 / 0.7190.468 / 0.4740.902 / 0.8880.810 / 0.7380.706 / 0.6160.836 / 0.785
1KonIQ, SPAQ, KADIDQ-Align [70]0.945 / 0.9380.933 / 0.9310.935 / 0.9340.409 / 0.4200.887 / 0.8830.788 / 0.7330.829 / 0.8080.876 / 0.845
DeQA-Score0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
2KonIQ, SPAQ, KADID, PIPALQ-Align [70]0.926 / 0.9320.917 / 0.9200.950 / 0.9540.702 / 0.6710.853 / 0.8450.765 / 0.7220.811 / 0.7950.838 / 0.789
DeQA-Score0.958 / 0.9460.932 / 0.9290.963 / 0.9610.724 / 0.6900.877 / 0.8570.770 / 0.7350.828 / 0.7960.863 / 0.807
", + "image_path": "7a2af31349da31ccf2e629c4996caddcbaa45a0dd7aefbd60d4390d7ac45a3f8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 411, + 294, + 491 + ], + "blocks": [ + { + "bbox": [ + 55, + 355, + 295, + 410 + ], + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 410 + ], + "type": "text", + "content": "Table 5. Score distribution prediction results with JS divergence / Wasserstein distance between the predicted Gaussian distribution and the human labeled Gaussian distribution as metrics. Models are trained on the KonIQ, SPAQ, and KADID datasets. DeQA-Score achieves a much closer alignment with human annotations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 411, + 294, + 491 + ], + "lines": [ + { + "bbox": [ + 57, + 411, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 411, + 294, + 491 + ], + "type": "table", + "html": "
DatasetsQ-Align [70]DeQA-Score (Ours)
KonIQ0.415 / 0.6110.014 / 0.186
SPAQ0.192 / 0.4550.070 / 0.332
KADID0.519 / 0.7230.029 / 0.304
LIVE-Wild0.254 / 0.6420.052 / 0.417
AGIQA-3K0.558 / 0.7300.177 / 0.592
", + "image_path": "f6d2c44707fda39404a9cd80c837066aa970ac4700cde32b035cf6720a609083.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 495, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 295, + 639 + ], + "type": "text", + "content": "significantly (e.g., " + }, + { + "bbox": [ + 54, + 495, + 295, + 639 + ], + "type": "inline_equation", + "content": "33.1\\%" + }, + { + "bbox": [ + 54, + 495, + 295, + 639 + ], + "type": "text", + "content": " in SRCC on TID2013), underscoring that distortion type is a major gap between training and testing. Third, from #1 to #2, adding PIPAL for training greatly enhances the results on PIPAL (from 0.495 to 0.724 in PLCC) but reduces the performance on several other datasets, e.g., from 0.900 to 0.863 in PLCC on CSIQ. This performance decline is likely due to the unique nature of model-processed distortions in PIPAL, which differ greatly from other distortions. Finally, despite never trained on AI-generated images, DeQA-Score performs well on AGIQA-3K with around 0.8 PLCC, reflecting the generalization ability enabled by MLLMs' broad pre-training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 285, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 285, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 285, + 653 + ], + "type": "text", + "content": "More results on AIGC images are provided in Tab. A4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 264, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 264, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 264, + 672 + ], + "type": "text", + "content": "5.3. Results of Score Distribution Prediction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "As stated in Sec. 4.2, the MLLM-output probabilities allow us to predict a score distribution, which can be evaluated against the human labeled ground-truth distribution. We" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 357, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 429 + ], + "type": "text", + "content": "calculate the distance between these two distributions and present the results in Tab. 5. The results illustrate that our predicted score distribution aligns much better with humans than Q-Align. As shown in Fig. 6, Q-Align tends to predict a single level token, bringing a quite small predicted variance, which compromises the distribution metrics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 556, + 552 + ], + "type": "text", + "content": "Qualitative results are shown in Fig. 6. Across different image types: in-the-wild images (Fig. 6ab), AI-generated images (Fig. 6c), or synthetic distortions (Fig. 6d), and for varying quality levels: high-quality (Fig. 6ac), fair-quality (Fig. 6b), or poor-quality images (Fig. 6d), our method consistently predicts accurate quality scores. Additionally, our DeQA-Score provides a score distribution that aligns more closely with humans with smaller KL divergence. In contrast, Q-Align, which is trained with one-hot labels, tends to predict a single label, resulting in higher KL divergence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 563, + 416, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 416, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 416, + 574 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Effectiveness of soft label and fidelity loss is studied and shown in Tab. 6. In #1, \"Uncertainty\" is utilized to verify the effects of adding annotation variances without considering the score distortion. We directly append uncertainty within the responses, i.e., \"The quality of this image is with uncertainty\". The has five possible values: {\"minimal\", \"low\", \"moderate\", \"high\", \"severe\"}, derived by the one-hot discretization. In #2, model trained on one-hot label can not predict a reliable variance, thus we employ binary fidelity loss [89], which does not require predicted variance." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14489" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 149, + 137 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 149, + 137 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 149, + 137 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 149, + 137 + ], + "type": "image", + "image_path": "c699409b820ce91ede379bc8da6cf3fe9c1ca7a3022565144543be88ed462c3f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 139, + 139, + 148 + ], + "lines": [ + { + "bbox": [ + 70, + 139, + 139, + 148 + ], + "spans": [ + { + "bbox": [ + 70, + 139, + 139, + 148 + ], + "type": "text", + "content": "(a) Score: " + }, + { + "bbox": [ + 70, + 139, + 139, + 148 + ], + "type": "inline_equation", + "content": "3.93 \\pm 0.79" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 152, + 72, + 229, + 148 + ], + "blocks": [ + { + "bbox": [ + 152, + 72, + 229, + 148 + ], + "lines": [ + { + "bbox": [ + 152, + 72, + 229, + 148 + ], + "spans": [ + { + "bbox": [ + 152, + 72, + 229, + 148 + ], + "type": "image", + "image_path": "dc877a31a5fed4204432786271d4cc0da1438ba39b46f9e32abe8d535e5a5e50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 231, + 72, + 309, + 148 + ], + "blocks": [ + { + "bbox": [ + 231, + 72, + 309, + 148 + ], + "lines": [ + { + "bbox": [ + 231, + 72, + 309, + 148 + ], + "spans": [ + { + "bbox": [ + 231, + 72, + 309, + 148 + ], + "type": "image", + "image_path": "8714ef4df6cf8bcee3b733e0b6d67a887732ed3f0245c7b066eae90c6237a818.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 380, + 148 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 380, + 148 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 380, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 380, + 148 + ], + "type": "image", + "image_path": "b35d2693e00ddaf1a5869e88e5dfd1c1fb60de964b0bb631c1275f10ec648a01.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 382, + 72, + 460, + 148 + ], + "blocks": [ + { + "bbox": [ + 382, + 72, + 460, + 148 + ], + "lines": [ + { + "bbox": [ + 382, + 72, + 460, + 148 + ], + "spans": [ + { + "bbox": [ + 382, + 72, + 460, + 148 + ], + "type": "image", + "image_path": "e3319e264f5260424b2edf37eda1075b0c5e27834d81eb45fbbadad8f360aa35.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 462, + 72, + 539, + 148 + ], + "blocks": [ + { + "bbox": [ + 462, + 72, + 539, + 148 + ], + "lines": [ + { + "bbox": [ + 462, + 72, + 539, + 148 + ], + "spans": [ + { + "bbox": [ + 462, + 72, + 539, + 148 + ], + "type": "image", + "image_path": "252eee6bc93d1a830eb98985d222d5d34942159a6771d6521a610b8270e511ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 67, + 152, + 149, + 208 + ], + "blocks": [ + { + "bbox": [ + 67, + 152, + 149, + 208 + ], + "lines": [ + { + "bbox": [ + 67, + 152, + 149, + 208 + ], + "spans": [ + { + "bbox": [ + 67, + 152, + 149, + 208 + ], + "type": "image", + "image_path": "178a4aabf6be7abc4c222416926a22c9cde2ebab09fcca2919558c926b5dd88e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 209, + 138, + 218 + ], + "lines": [ + { + "bbox": [ + 69, + 209, + 138, + 218 + ], + "spans": [ + { + "bbox": [ + 69, + 209, + 138, + 218 + ], + "type": "text", + "content": "(b) Score: " + }, + { + "bbox": [ + 69, + 209, + 138, + 218 + ], + "type": "inline_equation", + "content": "3.11 \\pm 0.66" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 152, + 153, + 229, + 208 + ], + "blocks": [ + { + "bbox": [ + 152, + 153, + 229, + 208 + ], + "lines": [ + { + "bbox": [ + 152, + 153, + 229, + 208 + ], + "spans": [ + { + "bbox": [ + 152, + 153, + 229, + 208 + ], + "type": "image", + "image_path": "45657c6b07acc75babb77f2c0dfa1d13471e286a8556b66fe5798581cf8366e6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 210, + 225, + 217 + ], + "lines": [ + { + "bbox": [ + 163, + 210, + 225, + 217 + ], + "spans": [ + { + "bbox": [ + 163, + 210, + 225, + 217 + ], + "type": "text", + "content": "Pred: 3.95, KL Div: 9.01" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 231, + 153, + 309, + 209 + ], + "blocks": [ + { + "bbox": [ + 231, + 153, + 309, + 209 + ], + "lines": [ + { + "bbox": [ + 231, + 153, + 309, + 209 + ], + "spans": [ + { + "bbox": [ + 231, + 153, + 309, + 209 + ], + "type": "image", + "image_path": "c570ba6f583d1d6ddf5d38abdd12d11914512d281ea3b6eb689e76ec73fb45bd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 210, + 304, + 217 + ], + "lines": [ + { + "bbox": [ + 242, + 210, + 304, + 217 + ], + "spans": [ + { + "bbox": [ + 242, + 210, + 304, + 217 + ], + "type": "text", + "content": "Pred: 3.34, KL Div: 0.06" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 221, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 221, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 553, + 255 + ], + "type": "text", + "content": "Figure 6. Qualitative results on KonIQ, SPAQ, AGIQA-3K, and KADID datasets. DeQA-Score predicts more accurate quality scores with more reasonable logits. Also, our method generates a score distribution that aligns better with humans, i.e., smaller KL divergence. However, Q-Align, which is trained on one-hot label, is more likely to predict a single label, resulting in higher KL divergence." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 152, + 380, + 208 + ], + "blocks": [ + { + "bbox": [ + 316, + 152, + 380, + 208 + ], + "lines": [ + { + "bbox": [ + 316, + 152, + 380, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 380, + 208 + ], + "type": "image", + "image_path": "5c05a66f95588d56945a0f9153eb720fccf1f484dd78ced507edc05d977bf07f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 377, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 377, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 377, + 217 + ], + "type": "text", + "content": "(d) Score: " + }, + { + "bbox": [ + 313, + 209, + 377, + 217 + ], + "type": "inline_equation", + "content": "1.92 \\pm 0.82" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 383, + 153, + 460, + 208 + ], + "blocks": [ + { + "bbox": [ + 383, + 153, + 460, + 208 + ], + "lines": [ + { + "bbox": [ + 383, + 153, + 460, + 208 + ], + "spans": [ + { + "bbox": [ + 383, + 153, + 460, + 208 + ], + "type": "image", + "image_path": "2c37c48761abc7d2a2469ed5c8e0c36f0474999395086f7f474996972224c0aa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 209, + 457, + 217 + ], + "lines": [ + { + "bbox": [ + 392, + 209, + 457, + 217 + ], + "spans": [ + { + "bbox": [ + 392, + 209, + 457, + 217 + ], + "type": "text", + "content": "Pred: 1.04, KL Div: 16.69" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 462, + 152, + 539, + 208 + ], + "blocks": [ + { + "bbox": [ + 462, + 152, + 539, + 208 + ], + "lines": [ + { + "bbox": [ + 462, + 152, + 539, + 208 + ], + "spans": [ + { + "bbox": [ + 462, + 152, + 539, + 208 + ], + "type": "image", + "image_path": "9d025297513e19736e64d8cc420319a8ae35d85b43196ac11890c821b49cd47d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 209, + 533, + 217 + ], + "lines": [ + { + "bbox": [ + 471, + 209, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 471, + 209, + 533, + 217 + ], + "type": "text", + "content": "Pred: 1.76, KL Div: 0.14" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 58, + 273, + 551, + 342 + ], + "blocks": [ + { + "bbox": [ + 59, + 258, + 550, + 270 + ], + "lines": [ + { + "bbox": [ + 59, + 258, + 550, + 270 + ], + "spans": [ + { + "bbox": [ + 59, + 258, + 550, + 270 + ], + "type": "text", + "content": "Table 6. Ablation studies on soft label and fidelity loss with PLCC / SRCC metrics. Models are trained on KonIQ, SPAQ, and KADID." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 273, + 551, + 342 + ], + "lines": [ + { + "bbox": [ + 58, + 273, + 551, + 342 + ], + "spans": [ + { + "bbox": [ + 58, + 273, + 551, + 342 + ], + "type": "table", + "html": "
#One-hotSoftUncertaintyFidelityKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
0✓0.945 / 0.9380.933 / 0.9310.935 / 0.9340.409 / 0.4200.887 / 0.8830.788 / 0.7330.829 / 0.8080.876 / 0.845
1✓✓0.935 / 0.9360.933 / 0.9310.954 / 0.9510.439 / 0.4530.896 / 0.8870.764 / 0.7150.826 / 0.8030.884 / 0.849
2✓✓0.944 / 0.9360.935 / 0.9320.951 / 0.9500.473 / 0.4780.898 / 0.8860.788 / 0.7260.837 / 0.8100.876 / 0.839
3✓0.954 / 0.9430.937 / 0.9340.951 / 0.9470.472 / 0.4720.896 / 0.8790.791 / 0.7290.852 / 0.8210.894 / 0.855
4✓✓0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
", + "image_path": "be24d649743e261fb51adf957c4281e93fbfa8828ae3710ca57a94ac30e2d9b1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 57, + 361, + 551, + 440 + ], + "blocks": [ + { + "bbox": [ + 57, + 361, + 551, + 440 + ], + "lines": [ + { + "bbox": [ + 57, + 361, + 551, + 440 + ], + "spans": [ + { + "bbox": [ + 57, + 361, + 551, + 440 + ], + "type": "table", + "html": "
CategoryLevel TextsKonIQSPAQKADIDPIPALLIVE-WildAGIQA-3KTID2013CSIQ
Commonexce. / good / far / poor / bad0.957 / 0.9440.938 / 0.9340.955 / 0.9530.495 / 0.4960.900 / 0.8870.808 / 0.7450.852 / 0.8200.900 / 0.857
Randomexce. / poor / bad / good / fair0.942 / 0.9280.933 / 0.9290.950 / 0.9470.394 / 0.3960.863 / 0.8510.795 / 0.7270.814 / 0.7920.887 / 0.836
Reversebad / poor / fair / good / exce.0.943 / 0.9340.934 / 0.9310.947 / 0.9450.395 / 0.3860.836 / 0.8000.782 / 0.7140.811 / 0.7910.882 / 0.834
Numberone / two / three / four / five0.945 / 0.9320.934 / 0.9300.955 / 0.9520.421 / 0.4240.873 / 0.8570.766 / 0.7120.820 / 0.7940.882 / 0.835
Vehiclecar / bus / train / plane / ship0.953 / 0.9400.937 / 0.9330.954 / 0.9510.458 / 0.4560.889 / 0.8780.778 / 0.7020.854 / 0.8280.897 / 0.856
Mixtureapple / car / four / fair / dog0.947 / 0.9380.936 / 0.9320.954 / 0.9510.419 / 0.4150.865 / 0.8510.808 / 0.7370.827 / 0.8000.896 / 0.854
", + "image_path": "d5908637415d6f56817dc07ad7c4d5ab24e970a365d0f64c955e1a545077165d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 57, + 459, + 294, + 502 + ], + "blocks": [ + { + "bbox": [ + 57, + 346, + 543, + 358 + ], + "lines": [ + { + "bbox": [ + 57, + 346, + 543, + 358 + ], + "spans": [ + { + "bbox": [ + 57, + 346, + 543, + 358 + ], + "type": "text", + "content": "Table 7. Effects of different level texts with PLCC / SRCC metrics. Models are trained on the KonIQ, SPAQ, and KADID datasets." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 446, + 293, + 457 + ], + "lines": [ + { + "bbox": [ + 57, + 446, + 293, + 457 + ], + "spans": [ + { + "bbox": [ + 57, + 446, + 293, + 457 + ], + "type": "text", + "content": "Table 8. Inference latency & throughput on RTX A6000 GPU." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 459, + 294, + 502 + ], + "lines": [ + { + "bbox": [ + 57, + 459, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 294, + 502 + ], + "type": "table", + "html": "
Batch Size12481664256
Latency (s)0.0890.1210.1820.3150.5882.1147.929
Throughput (image/s)11.2416.5321.9825.4027.2130.2732.29
", + "image_path": "b05c9986ec578fc831ea43045f61215f8fb537d64696b22a740144c27204d08e.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "content": "Several conclusions can be drawn from Tab. 6. First, comparing #0 with #3, our soft label significantly outperforms the one-hot label across all benchmarks, validating its effectiveness. Second, from #3 to #4, the performance of our soft label further improves when combined with the fidelity loss. Third, comparing #0 with #2, the one-hot label does not incorporate well with the fidelity loss, as it lacks score distribution information. Finally, from #0 to #1, simply adding uncertainty without considering score distribution does not yield stable or significant improvements." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "Effects of different level texts are studied in Tab. 7. Beside the commonly used level texts, we consider five alternatives: (a) random sorted common texts; (b) reversed common texts; (c) simple numbers like one, two, etc.; (d) irrelevant words about vehicle; (e) a mixture of various types. The results indicate that common level texts achieve the best performance, as they align closely with the MLLMs' pre" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 450, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 510 + ], + "type": "text", + "content": "training. Interestingly, with random sorted / reversed texts, even a mixture of various level texts, which are inconsistent with MLLMs' prior knowledge, our model still learns these newly defined relationships and performs well. This shows the robust adaptability of MLLM-based IQA models." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 512, + 553, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 553, + 537 + ], + "type": "text", + "content": "Ablation studies on level number and training / fixing components are provided in Tabs. A6 and A7, respectively." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 544, + 421, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 421, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 421, + 557 + ], + "type": "text", + "content": "5.5. Inference Latency" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 562, + 553, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 553, + 623 + ], + "type": "text", + "content": "We test the inference latency on one single RTX A6000 GPU. As shown in Tab. 8, with the optimal GPU utilization (i.e., batch size 256), our DeQA-Score can process as many as 32.29 images per second in average. This high throughput emphasizes its suitability for real-time deployment." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 633, + 391, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 391, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 391, + 645 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "We develop DeQA-Score as a pioneering MLLM-based image quality scorer, empowered by distribution-based soft labels and fidelity loss, capable of both score regression and score distribution prediction, highlighting the importance of quality score distribution in MLLM-based IQA methods." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14490" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgement. This work was sponsored by RGC Early Career Scheme (ECS) No. 24209224, National Natural Science Foundation of China (Grant No.62276251), and the Joint Lab of CAS-HK." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 205 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 207, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 207, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 207, + 294, + 251 + ], + "type": "text", + "content": "[2] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 253, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 253, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 253, + 294, + 305 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 351 + ], + "type": "text", + "content": "[4] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE TIP, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 396 + ], + "type": "text", + "content": "[5] Xin Cai, Zhiyuan You, Hailong Zhang, Wentao Liu, Jinwei Gu, and Tianfan Xue. PhoCoLens: Photorealistic and consistent reconstruction in lensless imaging. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 441 + ], + "type": "text", + "content": "[6] Yue Cao, Zhaolin Wan, Dongwei Ren, Zifei Yan, and Wangmeng Zuo. Incorporating semi-supervised and positive unlabeled learning for boosting full reference image quality assessment. In CVPR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 294, + 486 + ], + "type": "text", + "content": "[7] Chaofeng Chen, Yang Sensen, Haoning Wu, Liang Liao, Zicheng Zhang, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-Ground: Image quality grounding with large multi-modality models. In ACM MM, 2024. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 489, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 294, + 520 + ], + "type": "text", + "content": "[8] Du Chen, Tianhe Wu, Kede Ma, and Lei Zhang. Toward generalized image quality assessment: Relaxing the perfect reference quality assumption. In CVPR, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 523, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 294, + 565 + ], + "type": "text", + "content": "[9] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 600 + ], + "type": "text", + "content": "[10] Zixuan Chen, Zewei He, Ziqian Lu, Xuecheng Sun, and Zhe-Ming Lu. Prompt-based test-time real image dehazing: a novel pipeline. In ECCV, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 644 + ], + "type": "text", + "content": "[11] Zijian Chen, Wei Sun, Haoning Wu, Zicheng Zhang, Jun Jia, Ru Huang, Xiongkuo Min, Guangtao Zhai, and Wenjun Zhang. Study of subjective and objective naturalness assessment of ai-generated images. IEEE TCSVT, 2024. 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[12] Zixuan Chen, Yujin Wang, Xin Cai, Zhiyuan You, Zheming Lu, Fan Zhang, Shi Guo, and Tianfan Xue. UltraFusion: Ultra high dynamic imaging using exposure fusion. In CVPR, 2025. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "text", + "content": "hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "text", + "content": " * chatgpt quality. https://vicuna.lmsys.org, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "text", + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "text", + "content": "[15] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE TPAMI, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 199, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 199, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 199, + 553, + 232 + ], + "type": "text", + "content": "[16] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In ACM MM, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 234, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 553, + 266 + ], + "type": "text", + "content": "[17] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, 2020. 1, 2, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 269, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 269, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 269, + 553, + 311 + ], + "type": "text", + "content": "[18] Deepti Ghadiyaram and C. Bovik Alan. Live in the wild image quality challenge database. http://live.ece.utexas.edu/research/ChallengeDB/index.html, 2015.2,6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 314, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 314, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 314, + 553, + 335 + ], + "type": "text", + "content": "[19] Abhijay Ghildyal and Feng Liu. Shift-tolerant perceptual similarity metric. In ECCV, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "text", + "content": "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 384, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 553, + 426 + ], + "type": "text", + "content": "[21] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. KonIQ-10K: An ecologically valid database for deep learning of blind image quality assessment. IEEE TIP, 2020. 2, 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 429, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 429, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 317, + 429, + 553, + 472 + ], + "type": "text", + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2021. 13, 14, 15" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 475, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 553, + 495 + ], + "type": "text", + "content": "[23] Loshchilov Ilya and Hutter Frank. Decoupled weight decay regularization. In ICLR, 2019. 6, 13" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 317, + 498, + 553, + 540 + ], + "type": "text", + "content": "[24] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S Ren, and Dong Chao. PIPAL: a large-scale image quality assessment dataset for perceptual image restoration. In ECCV, 2020. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 544, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 553, + 575 + ], + "type": "text", + "content": "[25] Le Kang, Peng Ye, Yi Li, and David Doermann. Convolutional neural networks for no-reference image quality assessment. In CVPR, 2014. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 578, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 609 + ], + "type": "text", + "content": "[26] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality transformer. In CVPR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "text", + "content": "[27] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 2018. 13" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[28] Daekyu Kwon, Dongyoung Kim, Sehwan Ki, Younghyun Jo, Hyong-Euk Lee, and Seon Joo Kim. ATTIQA: Generalizable image quality feature extractor using attribute-aware pretraining. In ACCV, 2024. 6" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14491" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[29] Eric C Larson and Damon M Chandler. Most apparent distortion: full-reference image quality assessment and the role of strategy. Journal of electronic imaging, 2010. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[30] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. AGIQA-3K: An open database for ai-generated image quality assessment. IEEE TCSVT, 2023. 1, 2, 6, 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[31] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. KADID-10K: A large-scale artificially distorted iqa database. In International Conference on Quality of Multimedia Experience (QoMEX), 2019. 2, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 217 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1, 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "text", + "content": "[33] Kai Liu, Ziqing Zhang, Wenbo Li, Renjing Pei, Fenglong Song, Xiaohong Liu, Linghe Kong, and Yulun Zhang. Dog-IQA: Standard-guided zero-shot mllm for mix-grained image quality assessment. arXiv preprint arXiv:2410.02505, 2024.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "text", + "content": "[34] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. RankIQA: Learning from rankings for no-reference image quality assessment. In ICCV, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 362 + ], + "type": "text", + "content": "[35] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 418 + ], + "type": "text", + "content": "[36] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "text", + "content": "[37] Chao Ma, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. Learning a no-reference quality metric for single-image super-resolution. Computer Vision and Image Understanding, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "text", + "content": "[38] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In ACL Findings, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 541 + ], + "type": "text", + "content": "[39] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. DocVQA: A dataset for vqa on document images. In WACV, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 576 + ], + "type": "text", + "content": "[40] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE TIP, 2012. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 609 + ], + "type": "text", + "content": "[41] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Sign. Process. Letters, 2012. 1, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 611, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 294, + 643 + ], + "type": "text", + "content": "[42] Anush Krishna Moorthy and Alan Conrad Bovik. A two-step framework for constructing blind image quality indices. IEEE Sign. Process. Letters, 2010." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 646, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 677 + ], + "type": "text", + "content": "[43] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE TIP, 2011. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 679, + 253, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 679, + 253, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 679, + 253, + 689 + ], + "type": "text", + "content": "[44] OpenAI. GPT-4V(ision) system card, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[45] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[46] Da Pan, Ping Shi, Ming Hou, Zefeng Ying, Sizhe Fu, and Yuan Zhang. Blind predicting similar quality map for image quality assessment. In CVPR, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[47] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, et al. Image database TID2013: Peculiarities, results and perspectives. Signal processing: Image communication, 2015. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "text", + "content": "[48] Ekta Prashnani, Hong Cai, Yasamin Mostofi, and Pradeep Sen. PieAPP: Perceptual image-error assessment through pairwise preference. In CVPR, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 286, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 318 + ], + "type": "text", + "content": "[50] Michele A Saad, Alan C Bovik, and Christophe Charrier. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE TIP, 2012. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 319, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 351 + ], + "type": "text", + "content": "[51] B Series. Methodology for the subjective assessment of the quality of television pictures. Recommendation ITU-R BT, 2012. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 354, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 375 + ], + "type": "text", + "content": "[52] H Sheikh. LIVE image quality assessment database release 2. http://live.ece.utexas.edu/research/quality, 2005. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "type": "text", + "content": "[53] Hamid R Sheikh and Alan C Bovik. Image information and visual quality. IEEE TIP, 2006. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "text", + "content": "[54] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA models that can read. In CVPR, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "text", + "content": "[55] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In CVPR, 2020. 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 489, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 520 + ], + "type": "text", + "content": "[56] Simen Sun, Tao Yu, Jiahua Xu, Wei Zhou, and Zhibo Chen. GraphIQA: Learning distortion graph representations for blind image quality assessment. IEEE TMM, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 522, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 543 + ], + "type": "text", + "content": "[57] Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment. IEEE TIP, 2018. 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 545, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 576 + ], + "type": "text", + "content": "[58] Huixuan Tang, Neel Joshi, and Ashish Kapoor. Learning a blind measure of perceptual image quality. In CVPR, 2011. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 579, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 610 + ], + "type": "text", + "content": "[59] InternLM Team. InternLM: A multilingual language model with progressively enhanced capabilities. https://github.com/InternLM/InternLM, 2023. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 613, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 634 + ], + "type": "text", + "content": "[60] Louis L Thurstone. A law of comparative judgment. Psychological review, 1994. 2, 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 3, 4" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[62] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov," + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14492" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. \nLLaMA 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6, 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 295, + 152 + ], + "type": "text", + "content": "[63] Ming-Feng Tsai, Tie-Yan Liu, Tao Qin, Hsin-Hsi Chen, and Wei-Ying Ma. FRank: A ranking method with fidelity loss. In International ACM SIGIR Conference on Research and Development in Information Retrieval, 2007. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "type": "text", + "content": "[64] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 255 + ], + "type": "text", + "content": "[65] Jiarui Wang, Huiyu Duan, Jing Liu, Shi Chen, Xiongkuo Min, and Guangtao Zhai. AIGCIQA2023: A large-scale image quality assessment database for ai generated images: from the perspectives of quality, authenticity and correspondence. In CAAI International Conference on Artificial Intelligence, 2023. 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 257, + 294, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 294, + 290 + ], + "type": "text", + "content": "[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 292, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 294, + 346 + ], + "type": "text", + "content": "[67] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv preprint arXiv:2312.06109, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[68] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, et al. Q-Bench: A benchmark for general-purpose foundation models on low-level vision. In ICLR, 2024. 3, 15, 16" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 406, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 294, + 460 + ], + "type": "text", + "content": "[69] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-Instruct: Improving low-level visual abilities for multi-modality foundation models. In CVPR, 2024. 1, 2, 3, 15, 16" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 463, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 518 + ], + "type": "text", + "content": "[70] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-Align: Teaching LMMs for visual scoring via discrete text-defined levels. In ICML, 2024. 1, 2, 3, 4, 5, 6, 7, 14, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 564 + ], + "type": "text", + "content": "[71] Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, et al. Towards open-ended visual quality comparison. In ECCV, 2024. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "text", + "content": "[72] Tianhe Wu, Kede Ma, Jie Liang, Yujiu Yang, and Lei Zhang. A comprehensive study of multimodal large language models for image quality assessment. In ECCV, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 655 + ], + "type": "text", + "content": "[73] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. MANIQA: Multi-dimension attention network for no-reference image quality assessment. In CVPRW, 2022. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[74] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mPLUG-Owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[75] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mPLUG-Owl2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 1, 4, 6, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[76] Guanghao Yin, Wei Wang, Zehuan Yuan, Chuchu Han, Wei Ji, Shouqian Sun, and Changhu Wang. Content-variant reference image quality assessment via knowledge distillation. In AAAI, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[77] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Lu Sheng, Lei Bai, Xiaoshui Huang, Zhiyong Wang, et al. Lamm: Language-assisted multimodal instruction-tuning dataset, framework, and benchmark. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[78] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (PaQ-2-PiQ): Mapping the perceptual space of picture quality. In CVPR, 2020. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[79] Zhiyuan You, Jinjin Gu, Zheyuan Li, Xin Cai, Kaiwen Zhu, Chao Dong, and Tianfan Xue. Descriptive image quality assessment in the wild. arXiv preprint arXiv:2405.18842, 2024. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[80] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. In ECCV, 2024. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "text", + "content": "[81] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2014. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "type": "text", + "content": "[82] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. FSIM: A feature similarity index for image quality assessment. IEEE TIP, 2011. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 444, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 532 + ], + "type": "text", + "content": "[83] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "text", + "content": "[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "text", + "content": "[85] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. LLaMA-Adapter: Efficient fine-tuning of large language models with zero-initialized attention. In ICLR, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "type": "text", + "content": "[86] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE TCSVT, 2020. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[87] Weixia Zhang, Kede Ma, Guangtao Zhai, and Xiaokang Yang. Uncertainty-aware blind image quality assessment in the laboratory and wild. IEEE TIP, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[88] Weixia Zhang, Dingquan Li, Chao Ma, Guangtao Zhai, Xiaokang Yang, and Kede Ma. Continual learning for blind image quality assessment. IEEE TPAMI, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14493" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 553 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[89] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In CVPR, 2023. 2, 7, 14, 15" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[90] Zicheng Zhang, Chunyi Li, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. A perceptual quality assessment exploration for aigc images. In ICME Workshops, 2023. 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[91] Zicheng Zhang, Haoning Wu, Erli Zhang, Guangtao Zhai, and Weisi Lin. A benchmark for multi-modal foundation models on low-level vision: from single images to pairs. IEEE TPAMI, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[92] Zicheng Zhang, Yingjie Zhou, Chunyi Li, Baixuan Zhao, Xiaohong Liu, and Guangtao Zhai. Quality assessment in the era of large models: A survey. arXiv preprint arXiv:2409.00031, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 295 + ], + "type": "text", + "content": "[93] Heliang Zheng, Jianlong Fu, Yanhong Zeng, Zheng-Jun Zha, and Jiebo Luo. Learning conditional knowledge distillation for degraded-reference image quality assessment. ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 330 + ], + "type": "text", + "content": "[94] Wei Zhou and Zhou Wang. Quality assessment of image super-resolution: Balancing deterministic and statistical fidelity. In ACM MM, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "text", + "content": "[95] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[96] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. MetaQA: deep meta-learning for no-reference image quality assessment. In CVPR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[97] Hanwei Zhu, Xiangjie Sui, Baoliang Chen, Xuelin Liu, Peilin Chen, Yuming Fang, and Shiqi Wang. 2afc prompting of large multimodal models for image quality assessment. arXiv preprint arXiv:2402.01162, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "type": "text", + "content": "[98] Hanwei Zhu, Haoning Wu, Yixuan Li, Zicheng Zhang, Bao-liang Chen, Lingyu Zhu, Yuming Fang, Guangtao Zhai, Weisi Lin, and Shiqi Wang. Adaptive image quality assessment via teaching large multimodal model to compare. arXiv preprint arXiv:2405.19298, 2024. 2, 3, 5, 6, 7, 14, 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 553 + ], + "type": "text", + "content": "[99] Kaiwen Zhu, Jinjin Gu, Zhiyuan You, Yu Qiao, and Chao Dong. An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809, 2024. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14494" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/acf23ef4-3e0f-4da0-a811-56761959df29_content_list.json b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/acf23ef4-3e0f-4da0-a811-56761959df29_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cf92970d2ce5000a1da564a23af3d390a4051701 --- /dev/null +++ b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/acf23ef4-3e0f-4da0-a811-56761959df29_content_list.json @@ -0,0 +1,1630 @@ +[ + { + "type": "text", + "text": "Teller: Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation", + "text_level": 1, + "bbox": [ + 156, + 130, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dingcheng Zhen Shunshun Yin Shiyang Qin Hou Yi", + "bbox": [ + 274, + 203, + 750, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziwei Zhang Siyuan Liu Gan Qi Ming Tao", + "bbox": [ + 313, + 222, + 712, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanghai Soulgate Technolgy Co.tl.", + "bbox": [ + 364, + 239, + 661, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dingchengzhen, yinshunshun, qinshiyang, houyi,zhangziwei, siyuanliu, ganqi, ming}@soulapp.cn", + "bbox": [ + 101, + 258, + 919, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64a15499ee037e75efd44f23a583cd72d6c4540f11d8795d571e003bb86d2ca5.jpg", + "image_caption": [ + "Figure 1. Teller framework is the first autoregressive framework for real-time, audio-driven portrait animation, achieving up to 25 FPS while preserving realistic body part and accessory movements. Demo can be found at https://teller-ayar.github.io/." + ], + "image_footnote": [], + "bbox": [ + 116, + 284, + 883, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 582, + 326, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we introduce the first autoregressive framework for real-time, audio-driven portrait animation, a.k.a, talking head. Beyond the challenge of lengthy animation times, a critical challenge in realistic talking head generation lies in preserving the natural movement of diverse body parts. To this end, we propose Teller, the first streaming audio-driven portrait animation framework with autoregressive motion generation. Specifically, Teller first decomposes facial and body detail animation into two components: Facial Motion Latent Generation (FMLG) based on an autoregressive transformer, and movement authenticity refinement using a Efficient Temporal Module (ETM). Concretely, FMLG employs a Residual VQ model to map the facial motion latent from the implicit keypoint-based model into discrete motion tokens, which are then temporally sliced with audio embeddings. This enables the AR transformer to learn real-time, stream-based mappings from audio to motion. Furthermore, Teller incorporate ETM to capture finer motion details. This module ensures the physi", + "bbox": [ + 88, + 613, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal consistency of body parts and accessories, such as neck muscles and earrings, improving the realism of these movements. Teller is designed to be efficient, surpassing the inference speed of diffusion-based models (Hallo 20.93s vs. Teller 0.92s for one second video generation), and achieves a real-time streaming performance of up to 25 FPS. Extensive experiments demonstrate that our method outperforms recent audio-driven portrait animation models, especially in small movements, as validated by human evaluations with a significant margin in quality and realism.", + "bbox": [ + 511, + 583, + 906, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 753, + 645, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Realistic and expressive portrait animation from audio and static images, commonly known as talking head animation [12, 16, 21, 35], has garnered significant interest across applications such as virtual avatars, digital communication, and entertainment. However, generating high-quality animations that are visually compelling and temporally consistent remains a major challenge. This complexity stems from the need to intricately coordinate lip movements, fa", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21075", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cial expressions, and head positioning to create lifelike effects. Moreover, achieving real-time, realistic talking head animation is complicated by computational constraints and the nuances of human movement, making this an especially demanding task.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While recent advancements, such as diffusion models [25, 27, 29, 42], have improved high-quality content generation, achieving controllable animation presents ongoing hurdles. Effective animation requires the accurate capture of complex facial expressions, body gestures, and the subtle interplay between them. Existing methods often suffer from prolonged animation times (See animation time in Tab 1), limiting their potential for real-time applications, and frequently fail to capture the natural, interconnected motions of various facial and body parts, e.g., earrings and necklaces, as shown in Figure. 1 stage 2. This shortfall results in animations with stiff or exaggerated movements that disrupt the realism of the animation (See Figure 7). Addressing these challenges necessitates a solution that balances computational efficiency with high animation quality without overloading processing resources.", + "bbox": [ + 89, + 169, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose Teller, the first autoregressive framework capable of real-time, streaming-based talking head animation at up to 25 FPS. As shown in Figure. 2, Teller employs a two-stage framework, combining Facial Motion Latent Generation (FMLG) and an Efficient Temporal Module (ETM) to produce realistic and physically consistent animations across facial and body movements. In the first stage, FMLG uses a Residual Vector Quantization (RVQ) model [45] to encode facial motion latents derived from an implicit keypoint-based model into discrete motion tokens. These tokens are then temporally aligned with audio embeddings, allowing an autoregressive (AR) transformer to map audio signals to facial movements in real-time. By breaking down the motion into temporal segments, FMLG enables the AR transformer to dynamically and efficiently generate high-quality animations responsive to live audio inputs.", + "bbox": [ + 89, + 414, + 482, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the realism of body movements, Teller introduces ETM in the second stage, which captures subtle details that are often overlooked in existing methods [12, 41]. The ETM refines finer movements, such as body parts and accessories, ensuring physically plausible interactions. For instance, it simulates realistic motions in neck muscles and dynamic accessories like earrings, which are vital for maintaining visual continuity in the animated avatar. Our Teller model prioritizes computational efficiency and is specifically designed to significantly outperform diffusion-based models in terms of inference speed. For example, generating a one-second video takes Hallo 20.93s, while Teller completes the task in just 0.92s (see Tab. 1). Despite this speed, Teller maintains high animation fidelity. By optimizing computational demands without compromising output", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quality, Teller not only meets but exceeds real-time requirements, ensuring a smooth and responsive user experience.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In extensive evaluations across various benchmarks and real-world settings, Teller demonstrates significant improvements over current state-of-the-art audio-driven portrait animation models, as shown in Table. 1, particularly in capturing nuanced facial and body movements. Human evaluations validate Teller's superior quality and realism, especially in rendering the subtle movements essential for lifelike animations. Our research represents a notable advancement in real-time talking head animation, presenting an innovative framework that bridges the gap between realism and efficiency in multimodal animation, as depicted in Figure. 6, Figure. 7, and Figure. 8.", + "bbox": [ + 511, + 122, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 316, + 663, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-Diffusion-Based Audio-Driven Portrait Animation This line of works typically consist of two key components: an audio-to-motion model and a facial motion representation model. These methods often utilize implicit keypoints as an intermediate motion representation, warping the source portrait based on the driving image. The goal of implicit methods is to learn disentangled representations in 2D [4, 18, 23, 34, 44, 49] or 3D [9, 35] latent spaces, with a focus on aspects such as identity, facial dynamics, and head pose. For example, FOMM [26] employs first-order Taylor expansion to capture local motion, while FaceVid2Vid [35] extends this by introducing a 3D implicit keypoints representation, enabling free-view portrait animation.", + "bbox": [ + 511, + 340, + 903, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To effectively learn facial motion representations in latent space, these approaches often rely on GAN-based frameworks to disentangle identity-related appearance from non-identity-related motion, specifically capturing expressions, lip and eye movements, minor accessories, and poses. Examples include FaceVid2Vid [35], LivePortrait [12], and others, where identity and motion representations are independently learned to generate realistic talking head animations. For instance, MakeItTalk [50] uses an LSTM-based audio-to-motion model to predict landmark coordinates from audio input, which are then translated into video frames using a warp-based GAN model. Similarly, SadTalker [46] employs FaceVid2Vid [35] as an image synthesizer, with ExpNet and PoseVAE modules transforming audio features into inputs compatible with FaceVid2Vid for audio-to-video generation.", + "bbox": [ + 511, + 537, + 903, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, these approaches face challenges due to GAN losses that primarily focus on facial expressions, lip, and eye movements, often neglecting accessory, hair, and body movements. This can result in animations that appear stiff or incomplete, lacking natural dynamism. In contrast, our work introduces the first autoregressive framework specifically designed for real-time, audio-driven portrait animation, achieving up to 25 FPS and delivering a more realistic,", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21076", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/716343c816bdff653c8ce0ada9d2afffe62239668bc8ac48b6490435d88363d9.jpg", + "image_caption": [ + "Figure 2. Overall framework of our proposed Teller for real-time streaming audio-driven portrait animation." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 897, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "coherent portrayal of both facial and accessory movements.", + "bbox": [ + 89, + 457, + 482, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-based Audio-driven Portrait Animation Recent advancements in diffusion-based video generation have demonstrated promising outcomes for audio-driven portrait animation. Methods like GAIA [14] and VASA-1 [41] have designed diffusion models to transform audio inputs into motion latents, facilitating audio-to-video generation. Further developments, such as EMO[32], Hallo [40], LOOPY [17], enhance end-to-end diffusion modeling by incorporating motion modules [3, 13, 16] and audio cross-attention mechanisms, improving the coherence and synchronization between audio cues and visual motion. Despite these improvements, a significant limitation of diffusion-based models remains their multi-step inference process, required to generate even a single frame or a few frames of video. This step-by-step prediction approach renders real-time performance challenging, as it is computationally intensive and time-consuming, making diffusion-based models less suitable for applications demanding instantaneous response.", + "bbox": [ + 89, + 474, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AR Transformer-Based Generation Recent works [1, 30, 36, 38, 43, 48] have focused on developing unified multimodal language models for generating visual content, such as images and videos. Some studies [28, 51] use autoregressive modeling with continuous representations interleaved with text tokens for image generation. Others, like SEED-X [11], propose a foundational system combining CLIP ViT-based image representations with text tokens for multimodal tasks, including next-token prediction and image rep", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resentation regression. DreamLLM [8] also explores multimodal understanding and creation, while Chameleon [31] introduces token-based models capable of both understanding and generating images.", + "bbox": [ + 511, + 457, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 529, + 604, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, Teller comprises two main modules: the Facial Motion Latent Generation (FMLG) and the Efficient Temporal Module (ETM). FMLG module integrates an autoregressive transformer and a residual vector quantization (RVQ) component. It applies an autoregressive transformer to generate discrete facial motion tokens from audio input. Following FMLG, ETM refines the generated motion to produce realistic body and accessory movements, ensuring physical consistency in animated results.", + "bbox": [ + 511, + 554, + 906, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 696, + 653, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior works, such as LivePortrait [12] have introduced methods for extracting implicit keypoints as facial motion latents using motion and appearance extractors. These motion latents capture essential facial dynamics needed for animating input images and consist of three main components:", + "bbox": [ + 511, + 719, + 905, + 795 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Expression Deformation $(\\delta)$ : A set deformation of 21 implicit keypoints, represented as $\\delta = [\\delta_1, \\delta_2, \\dots, \\delta_{21}]$ , where each $\\delta$ is a 3D vector $(\\mathbb{R}^3)$ that indicates the position of the $i$ -th deformation of facial keypoints.", + "- Head Pose $(R)$ : Defined by three rotation vectors $R = [r_1, r_2, r_3]$ , with each $r_i$ being a 3D vector $(\\mathbb{R}^3)$ that describes the head's orientation in 3D space." + ], + "bbox": [ + 513, + 796, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21077", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Expression Deformation $(t)$ : A single 3D vector $(t \\in \\mathbb{R}^3)$ that captures facial expression deformations.", + "bbox": [ + 89, + 90, + 483, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These components are concatenated into a unified motion. \nlatent of size $25\\times 3$ ..", + "bbox": [ + 89, + 122, + 482, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm = \\left[ \\delta_ {1}, \\delta_ {2}, \\dots , \\delta_ {2 1}, r _ {1}, r _ {2}, r _ {3}, t \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 156, + 482, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m \\in \\mathbb{R}^{25 \\times 3}$ includes the 21 keypoints, head pose, and expression deformation.", + "bbox": [ + 89, + 174, + 482, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Facial Motion Latent Generation (FMLG)", + "text_level": 1, + "bbox": [ + 89, + 217, + 450, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In FMLG, Teller generates motion latent from facial motion extraction, then encodes it into discrete tokens using a residual vector quantizer (RVQ). To optimize encoding, the concatenated motion latent $m$ is processed with RVQ, leveraging temporal redundancy across $T$ frames for efficient compression. The RVQ quantization loss, which encodes the continuous $m$ into discrete tokens, is defined as:", + "bbox": [ + 89, + 239, + 483, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v q} = \\sum_ {t = 1} ^ {T} \\left[ \\underbrace {\\left| \\left| m - \\operatorname {F F N} _ {d e c} \\left(z _ {t} + \\operatorname {s g} [ \\hat {z} _ {t} - z _ {t} ]\\right) \\right| \\right| _ {2} ^ {2}} _ {\\mathcal {L} _ {\\text {r e c o n}}} + \\underbrace {\\left| \\left| z _ {t} - \\operatorname {s g} [ \\hat {z} _ {t} ] \\right| \\right| _ {2} ^ {2}} _ {\\mathcal {L} _ {\\text {c o m m i t}}} \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 369, + 493, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{t} = \\mathrm{FFN}_{enc}(m)$ represents the encoded latent at each time step $t$ , $\\hat{z}_t$ is the quantized latent, and $\\mathrm{sg}$ denotes the stop-gradient operation. The two loss components, $\\mathcal{L}_{\\mathrm{recon}}$ and $\\mathcal{L}_{\\mathrm{commit}}$ , are defined as follows:", + "bbox": [ + 89, + 434, + 483, + 494 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Reconstruction Loss $(\\mathcal{L}_{\\mathrm{recon}})$ : Minimizes the difference between the original motion latent $m$ and the decoded quantized latent.", + "- Commitment Loss $(\\mathcal{L}_{\\mathrm{commit}})$ : Encourages $z_{t}$ to approach the quantized latent $\\hat{z}_t$ , ensuring stability in quantization." + ], + "bbox": [ + 89, + 498, + 482, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach enables FMLG to learn robust, temporally-consistent representations, translating audio-driven inputs into lifelike facial animations. Experimentally, to achieve trade-offs between frame count and redundancy, we selected 4 frames $(4\\times 25\\times 3$ latent) to be compressed into 32 tokens (hyper-parameter selection trade-off refer to Fig. 11).", + "bbox": [ + 89, + 578, + 483, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. AR Transformer for Motion Generation", + "text_level": 1, + "bbox": [ + 89, + 696, + 433, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the learned RVQ-based latents, motion is denoted as:", + "bbox": [ + 89, + 719, + 482, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM = \\left[ m _ {1}, m _ {2}, \\dots , m _ {T} \\right], \\quad \\text {w h e r e} \\quad m _ {i} \\in \\mathbb {R} ^ {2 5 \\times 3}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 751, + 482, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This sequence of motion latents is converted to discrete tokens using the RVQ module:", + "bbox": [ + 89, + 772, + 482, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {m} = \\left[ t _ {1}, t _ {2}, \\dots , t _ {T / 4} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 805, + 482, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T_{m}$ represents the quantized motion tokens. Audio input is encoded with the Whisper encoder [5], generating the audio condition $c$ . Motion generation is modeled as a next-token prediction task, where the distribution of each token is predicted based on the previous $t - 1$ tokens and", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e4f45325368af28fa6451f3586e5715c2cc81f263c5b155e484f0dd6e751915.jpg", + "image_caption": [ + "Figure 3. In our Teller, we follow AR transformer architecture, but each input consists of a pair of tokens and model pred a pair of tokens for each output position." + ], + "image_footnote": [], + "bbox": [ + 521, + 93, + 895, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "audio condition $c$ :", + "bbox": [ + 513, + 349, + 635, + 363 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(t _ {i} \\mid c, t _ {< i}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 367, + 903, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This autoregressive setup enables sequential generation of motion based on past motion tokens and audio embeddings.", + "bbox": [ + 511, + 387, + 903, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable real-time streaming animation, we process both audio and video frames in 200ms chunks, following Whisper's constraints. Each audio chunk is encoded into a $[10 \\times 512]$ embedding, while each video chunk uses 32 learned RVQ-based motion tokens. For efficient real-time performance, we enable the autoregressive transformer to process token pairs at each position, which improves prediction speed by processing two tokens concurrently. In Teller, each input consists of a token pair with a combined embedding and a learnable position bias inspired by BERT, capturing relative token positions. The loss for each head, representing each token in the pair, is computed as:", + "bbox": [ + 511, + 417, + 905, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {h e a d} 0 _ {j}} = \\mathrm {C E} \\left(\\operatorname {l a b e l} _ {j} [ 0 ], \\mathrm {T} \\left(\\operatorname {i n p u t} _ {j} [ 0 ] \\mid \\operatorname {i n p u t} _ {< j}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 601, + 903, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {h e a d} 1 _ {j}} = \\mathrm {C E} \\left(\\operatorname {l a b e l} _ {j} [ 1 ], \\mathrm {T} \\left(\\operatorname {i n p u t} _ {j} [ 1 ] \\mid \\operatorname {i n p u t} _ {< j}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 637, + 903, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ denotes the transformer. The total loss, with a regularization term to balance learning across both heads, is:", + "bbox": [ + 511, + 661, + 903, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a r} = \\sum_ {j = 1} ^ {I / 2} \\left[ \\mathcal {L} _ {\\text {h e a d} 0 _ {j}} + \\mathcal {L} _ {\\text {h e a d} 1 _ {j}} + \\left\\| \\mathcal {L} _ {\\text {h e a d} 0 _ {j}} - \\mathcal {L} _ {\\text {h e a d} 1 _ {j}} \\right\\| _ {2} ^ {2} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 707, + 903, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This regularization term, $\\| \\mathcal{L}_{\\mathrm{head0}} - \\mathcal{L}_{\\mathrm{head1}}\\| _2^2$ ensures balanced training across the two heads, promoting stable and accurate real-time animation.", + "bbox": [ + 511, + 763, + 903, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After testing frame count and redundancy, we use 4 frames $(4\\times 25\\times 3$ latent), compressed into 32 tokens (refer to Fig. 11). Then the 4 frames are interpolated to 5 frames for faster generation. This setup balances inference efficiency and feature quality, enabling real-time, high-fidelity streaming portrait animation.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21078", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1c89d4f74316bc17d94cfb7d61cff7e810b87dbfb59e777ffccdb4873933c716.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with existing approaches on RAVDESS data-set of 'angry' and 'disgust' emotion cases. Videos are available in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 509, + 404 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/01412e46931605ddfa0c8ad9afffc0862ba25a913c94f8f90be4e0c0f92a4d43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 89, + 903, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Efficient Temporal Module for Refinement", + "text_level": 1, + "bbox": [ + 89, + 445, + 454, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In diffusion-based models, temporal layers are added to text-to-image (T2I) frameworks to capture frame dependencies [16]. Inspired by this, Teller incorporates temporal refinement but achieves it in a single step, unlike diffusion models that require multiple iterations, enhancing real-time efficiency. After encoding video frames with a VAE encoder, we apply a 3D U-Net [7] to extract features from the image sequence, represented as $x \\in \\mathbb{R}^{b \\times t \\times h \\times w \\times c}$ , where $b$ is the batch size, $t$ the frame count, $h$ and $w$ the frame dimensions, and $c$ the channel count. The features are reshaped to $x \\in \\mathbb{R}^{(b \\times h \\times w) \\times t \\times c}$ , enabling ETM to perform self-attention along the temporal dimension $t$ . ETM's output is then merged with the original features through residual connections, integrating temporal dependencies into spatial features.", + "bbox": [ + 88, + 473, + 482, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For training, we use the first 5 frames from a real image sequence and the subsequent 5 frames reconstructed with LivePortrait [12]. After processing through ETM, we compute reconstruction loss between the predicted and ground-truth frames:", + "bbox": [ + 89, + 707, + 483, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\sum_ {i = 6} ^ {1 0} \\left\\| x _ {\\mathrm {g t} _ {i}} - f \\left(x _ {i} \\mid x _ {\\mathrm {g t} _ {< 6}}\\right) \\right\\| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 782, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{\\mathrm{gt}_i}$ denotes the real feature sequence and $x_{i}$ the reconstructed feature sequence (from LivePortrait or the Stage 1 decoder). ETM primarily learns to preserve consistency in physical features such as neck muscles and earrings. To this end, we add a region-specific mask to recon", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "struction loss, and the final loss of ETM:", + "bbox": [ + 513, + 445, + 784, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E T M}} = \\sum_ {i = 6} ^ {1 0} \\left\\| x _ {\\mathrm {g t} _ {i}} \\odot \\operatorname {m a s k} _ {i} - f \\left(x _ {i} \\mid x _ {\\mathrm {g t} _ {< 6}}\\right) \\odot \\operatorname {m a s k} _ {i} \\right\\| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 470, + 903, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where element-wise multiplication with $\\mathrm{mask}_i$ focuses reconstruction on specific regions. The mask is defined as:", + "bbox": [ + 511, + 525, + 903, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {m a s k} (i, j) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} (i, j) \\text {i s w i t h i n} \\mathrm {B B} (x) \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 566, + 903, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{BB(x)}$ stands for the bounding box and outlines relevant body parts. Key landmarks are identified using MediaPipe[22], with points indices such as [93, 323, 152] defining the bounding boxes for these regions, highlighted in red in the stage 2 input of Fig. 2.", + "bbox": [ + 511, + 613, + 906, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 702, + 645, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 727, + 702, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For training, we used the AV Speech datasets (filtered to 662 hours) [10] and VFHQ [39] datasets (filtered to 2 hours) for pretraining, along with additional talking-head videos from the internet (32 hours) for supervised finetuning (SFT). For validation, we used the HDTF (filtered to 0.83 hours) [47] and RAVDESS (filtered to 0.55 hours) [19] datasets and supplementary internet data (0.49 hours) for qualitative comparisons and human evaluation only. To ensure quality, we applied the Mediapipe [22] face detection tool to filter out instances with facial movement exceeding", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21079", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a7cd6f75f4f61c7a33a454b3bdec767de096ec7dc48f38d0abe538c4d80ab9b.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison with existing approaches on HDTF dataset. Videos are available in the supplement. mat." + ], + "image_footnote": [], + "bbox": [ + 94, + 95, + 480, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$50\\%$ . We further refined the data using Sync-C and Sync-D to exclude samples with low lip-sync scores.", + "bbox": [ + 89, + 422, + 482, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Evaluation metrics include Fréchet Inception Distance (FID)[15], Fréchet Video Distance (FVD)[33], Synchronization-C (Sync-C)[24], and Synchronization-D (Sync-D)[24]. FID and FVD assess realism, with lower scores indicating better quality, while Sync-C and Sync-D measure lip synchronization, with higher Sync-C and lower Sync-D values indicating better alignment.", + "bbox": [ + 89, + 455, + 482, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details.", + "text_level": 1, + "bbox": [ + 89, + 564, + 264, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 1: During pretraining, we follow the architecture design of the Qwen1.5-4B model [2] and initialize the parameters randomly. The model is trained on an $8 \\times 8$ Nvidia A800 GPU machine with a batch size of 1024, using the AdamW optimizer [20]. We employ a cosine learning rate scheduler, with the learning rate decaying from 1e-4 to 1e-6 over 40 epochs. In the supervised fine-tuning (SFT) phase, we again use an $8 \\times 8$ Nvidia A800 GPU machine, but with a batch size of 512. The AdamW optimizer [20] is used, along with a cosine learning rate scheduler. The learning rate decaying from 1e-5 to 1e-6 over 10 epochs. Stage 2: The model is trained on an $8 \\times 8$ Nvidia A800 GPU machine, with a batch size of 1024, and using the AdamW optimizer. The cosine learning rate scheduler is used, with the learning rate decaying from 1e-4 to 1e-6 over 30 epochs.", + "bbox": [ + 89, + 580, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real time analysis. The model is inference on an 4 Nvidia H800 GPU machine. For a 200ms audio input, the average processing time of the Whisper encoder is 7ms. In Stage 1, the average total time is 106ms, with the AR transformer taking an average of 6ms per 16 tokens, and the motion decoder taking an average of 10ms. In Stage 2, the average", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fb12366941a8ef01bc8ab6d91e69df04b47f13b6cfc423a0787100979fe90950.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓FVD↓Sync-C↑Sync-D↓Time
SadTalker [46]22.177233.6737.3267.84818.89s
EchoMimic [6]23.049290.1906.6648.83931.10s
AniPortrait [37]28.161235.0994.54710.65729.36s
Hallo [40]20.639174.1917.4977.74120.93s
Teller (ours)21.352173.4637.6967.5360.92s
Real video--8.0946.976-
", + "bbox": [ + 517, + 88, + 903, + 189 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5c9b14e976c9638a21e899e9ac40ce8026d8f4c6d8aea716dd5a6d38c53db999.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with existing portrait image animation approaches on the HDTF dataset. Time stands for the averaging time cost of generating one second of 25 fps video." + ], + "table_footnote": [], + "table_body": "
MethodFID↓FVD↓Sync-C↑Sync-D↓
SadTalker [46]32.343487.9244.3047.621
EchoMimic [6]21.058668.6753.2929.096
AniPortrait [37]30.696476.1972.32111.542
Hallo [40]19.826537.4784.0628.552
Teller (ours)20.352429.2884.4967.936
Real video--5.2237.069
", + "bbox": [ + 517, + 236, + 903, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison with existing portrait image animation approaches on the RAVDESS dataset.", + "bbox": [ + 513, + 354, + 903, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "total time is 71ms, with the VAE encoder and decoder averaging 25ms, and the Temporal Module averaging 21ms.", + "bbox": [ + 511, + 395, + 903, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 436, + 705, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on the HDTF Dataset. Table 1 presents quantitative results for portrait animation techniques on the HDTF dataset. Our method outperforms others, achieving the lowest FVD of 173.463 and a competitive FID of 21.352, indicating high quality and temporal coherence in animated talking heads. Additionally, it achieves the highest Sync-C score of 7.696 and the lowest Sync-D score of 7.536, demonstrating excellent lip synchronization. These results highlight Teller's effectiveness in maintaining both visual fidelity and synchronization.", + "bbox": [ + 511, + 458, + 906, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on the RAVDESS Dataset. Table 2 shows quantitative results on the RAVDESS dataset. Our method again leads in performance, with the lowest FVD of 429.288 and a competitive FID of 20.352, reflecting high-quality and temporally coherent animations. The highest Sync-C score of 4.496 and a competitive Sync-D score of 7.936 further demonstrate superior lip synchronization. These findings confirm Teller's strength in producing synchronized and high-fidelity animated portraits.", + "bbox": [ + 511, + 609, + 908, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 756, + 694, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Head Movement Comparison. Figure 5 shows a qualitative comparison of head movements. Teller replicates natural head movements more accurately, closely matching the ground truth (GT) with smooth, realistic turns and subtle expression-based adjustments. Competing methods, like AntiPortrait and EchoMimic, often show abrupt or limited movements, appearing rigid or lifeless. Teller's autoregressive framework ensures continuity and natural dynamics in", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21080", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38ff0996b8342e7d6677a1996802c3cf06c579a35c7343744c815a6caca4bd45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 89, + 478, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b14dfdea6bea655b88263978b2fb7bd3c7acfc1849538659b5a6c3f73d3a51ca.jpg", + "image_caption": [ + "Figure 6. Top-k selection $(\\mathrm{k} = 15)$ in FMLG produces diverse facial expressions and actions with accurate lip sync on the HDTF." + ], + "image_footnote": [], + "bbox": [ + 94, + 321, + 480, + 580 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2f68803cb488bf6c9d4b5ab468a9914f9eb29c26ebace7d39fcd056d8b018de.jpg", + "image_caption": [ + "Figure 7. Visualization of finer motion details. Videos are available in the supplementary materials.", + "Figure 8. Visualization of the generation accuracy of lip shape." + ], + "image_footnote": [], + "bbox": [ + 93, + 625, + 480, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "head and emotion alignment, essential for lifelike animation.", + "bbox": [ + 89, + 763, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diversity. Fig. 6 shows two sequences generated with Top-k sampling (k=15) in FMLG, where each row represents frames from the same speech input. The model demonstrates diverse facial expressions and head movements while maintaining accurate lip sync, highlighting Top-k sampling's role in enhancing motion variety without compromising synchronization.", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8bd03e69009f82e3ad6bb1c37ba5126f9aab335c8191e1d27e6d69cc3cfe6b21.jpg", + "image_caption": [ + "Figure 9. Human evaluation results among our proposed Teller and other SoTA methods." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 643, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/745b71d2a611d16d33f43cc260fd628402b8dccefcf788252b634f089b243b59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 90, + 772, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/728a8e415ca6cb26841caba08b410ae33c8756363af3c311c560ba7968c5f814.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 90, + 901, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Emotional Expression. Fig. 4 presents our model has more accurate emotional expression ability due to the better speech understanding ability of AR transformer.", + "bbox": [ + 511, + 273, + 903, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finer Motion Details. Fig. 7 highlights fine motion details in neck and earring movements. Compared to other methods, Teller produces realistic, nuanced motions synchronized with speech, capturing subtle audio-driven dynamics that contribute to lifelike and temporally coherent animation. The consistent detail across frames underscores Teller's robustness.", + "bbox": [ + 511, + 319, + 905, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lip Synchronization Accuracy. Fig. 8 shows superior lip synchronization, with Teller aligning generated lip shapes closely to natural movements. This high fidelity highlights Teller's strength in producing realistic, synchronized mouth motions for convincing talking head animation.", + "bbox": [ + 511, + 424, + 905, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Human Evaluation", + "text_level": 1, + "bbox": [ + 513, + 508, + 694, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted a human evaluation to assess the quality of generated animations, focusing on lip synchronization, body movement realism, and temporal coherence. Thirty participants (66.7% aged 24-30, 33.3% aged 30-40; 30% male, 70% female; 83.3% with AIGC model experience) rated each animation on a 5-point Likert scale for coherence with input and animation quality. A total of 100 videos were presented in random order to avoid bias, providing insights into subjective perceptions of animation quality and natural expression alignment. As shown in Figure 9, participants rated Teller highest in lip synchronization, body movement realism, and temporal coherence, with low variance in scores, indicating robust and consistent performance.", + "bbox": [ + 511, + 529, + 906, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 739, + 663, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stage (Module) Ablation. We perform an ablation study comparing Stage 1 and Stage 2, focusing on animation quality in body parts and accessories, especially neck muscles and earrings. As shown in Fig. 10, Stage 1 outputs often show less realistic, inconsistent movements. In contrast, Stage 2, with ETM, significantly improves the physical consistency of subtle motions, creating natural earring sway and smooth neck movements for lifelike, temporally coherent animation.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21081", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3941306c3f53f958fbed40f3dbfbcd0f35264b47aa42ee41c0ac07fa429be58f.jpg", + "image_caption": [ + "Figure 10. Visualization of the generation images with different stages and the differences between stages." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/833776ffb35d18de0fa849bac2b072534a5e4e3ae0702cdcc74605243fb9df99.jpg", + "image_caption": [ + "Figure 11. Tradeoff between performance (loss) and different compression(tokens/frame) ratios." + ], + "image_footnote": [], + "bbox": [ + 94, + 342, + 478, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frames / Tokens in RVQ Module Figure 11 shows the impact of frame and token configurations in the RVQ module on face shape accuracy in talking head animations. This ablation study explores how varying frames and tokens affects facial movement quality and synchronization. Results show that increasing frames and tokens improves facial dynamics, enhancing lip synchronization and realism, though with higher computational costs. We selected 4 frames and 32 tokens for an optimal balance, adjustable as needed.", + "bbox": [ + 88, + 534, + 482, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on Audio Condition Encoder. We analyzed the performance differences between TTS and ASR models as audio condition encoders for our task. Specifically, we used Whisper for the ASR model and funcodec for the TTS model. As shown in Table 3, Whisper achieved a Sync-C score of 7.696 and a Sync-D score of 7.536, while funcodec scored 4.286 in Sync-C and 10.373 in Sync-D. These results indicate that our task benefits more from the ASR model's capability to capture the nuances required for precise synchronization, as seen in Whisper's higher Sync-C and lower Sync-D scores. This analysis suggests that the ASR model is better suited as the audio condition encoder, enhancing the overall quality and synchronization of our talking head animation.", + "bbox": [ + 89, + 672, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on Single-Head vs. Multi-Head Archi", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7fc76daad3883dc6698abdbf65f4bcae812c2e4129bc3d461d8cc1481988b04a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSync-C ↑Sync-D ↓
w funcdec4.28610.373
w whisper7.6967.536
", + "bbox": [ + 576, + 78, + 836, + 137 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b7a83da0013f3be62bce5254a11ae5ad4c63b59466852e71da4f65edfe8fcaf8.jpg", + "table_caption": [ + "Table 3. Comparison of synchronization for audio conditions using funcodec and Whisper in ASR and TTS tasks on HDTF." + ], + "table_footnote": [], + "table_body": "
MethodFID↓FVD↓Sync-C↑Sync-D↓
Single-Head22.110172.5537.7907.474
Multi-Head21.352173.4637.6967.536
", + "bbox": [ + 514, + 172, + 911, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison of performance between Single-Head and Multi-Head on the HDTF dataset.", + "bbox": [ + 511, + 234, + 906, + 261 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tecture. We compare single-head and multi-head models on FID, FVD, Sync-C, and Sync-D metrics (Table 4). The single-head model slightly outperforms in FID (22.110 vs. 21.352) and has comparable FVD (172.553 vs. 173.463), with marginally better synchronization (Sync-C of 7.790 vs. 7.696 and Sync-D of 7.474 vs. 7.536). Both architectures yield competitive results, though the single-head model slightly excels in synchronization, while the multi-head model offers greater real-time potential.", + "bbox": [ + 511, + 273, + 906, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 422, + 633, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented Teller, the first autoregressive framework designed for real-time, audio-driven portrait animation. Addressing the challenge of realistic and efficient talking head generation, Teller achieves high-quality animations at up to 25 FPS, surpassing existing methods in both fidelity and responsiveness. Extensive experiments demonstrated Teller's advantages over SoTA audio-driven animation methods, particularly in rendering nuanced movements essential for lifelike and visually convincing animations. Human evaluations further validate its quality, particularly in natural expression and lip synchronization. By balancing computational efficiency with high animation fidelity, Teller sets a new standard for real-time talking head animation, marking a significant advancement in multimodal portrait animation frameworks. Additionally, Teller's AR Transformer architecture makes it compatible with existing unified multimodal language models.", + "bbox": [ + 511, + 448, + 908, + 705 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21082", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564, 2023. 3", + "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 6", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3", + "[4] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2", + "[5] Nan Cao, Yu-Ru Lin, Xiaohua Sun, David Lazer, Shixia Liu, and Huamin Qu. Whisper: Tracing the spatiotemporal process of information diffusion in real time. IEEE transactions on visualization and computer graphics, 18(12):2649-2658, 2012. 4", + "[6] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 6", + "[7] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19, pages 424-432. Springer, 2016. 5", + "[8] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 3", + "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 2", + "[10] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T. Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: a speaker-independent audio-visual model for speech separation. ACM Transactions on Graphics, 37(4):1-11, 2018. 5", + "[11] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 3" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1, 2, 3, 5", + "[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3", + "[14] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 3", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[16] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 1, 3, 5", + "[17] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 3", + "[18] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2", + "[19] Steven R Livingstone and Frank A Russo. The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5): e0196391, 2018. 5", + "[20] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[21] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (ToG), 40(6):1-17, 2021. 1", + "[22] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5", + "[23] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 427-436, 2023. 2", + "[24] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21083", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28th ACM international conference on multimedia, pages 484-492, 2020. 6", + "[25] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 2", + "[26] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 2", + "[27] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 2", + "[28] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3", + "[29] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2", + "[30] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[31] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models, 2024. URL https://arxiv.org/abs/2405.09818, 2024.3", + "[32] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 3", + "[33] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6", + "[34] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17979-17989, 2023. 2", + "[35] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 1, 2", + "[36] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3", + "[37] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 3", + "[39] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 5", + "[40] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 3, 6", + "[41] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3", + "[42] Ziyu Yao, Xuxin Cheng, and Zhiqi Huang. Fd2talk: Towards generalized talking head generation with facial decoupled diffusion model. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3411-3420, 2024. 2", + "[43] Hanrong Ye, De-An Huang, Yao Lu, Zhiding Yu, Wei Ping, Andrew Tao, Jan Kautz, Song Han, Dan Xu, Pavlo Molchanov, et al. X-vila: Cross-modality alignment for large language model. arXiv preprint arXiv:2405.19335, 2024. 3", + "[44] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2", + "[45] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 2", + "[46] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 6", + "[47] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 5", + "[48] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 3", + "[49] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21084", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176-4186, 2021. 2", + "[50] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2", + "[51] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 3" + ], + "bbox": [ + 91, + 90, + 482, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21085", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/acf23ef4-3e0f-4da0-a811-56761959df29_model.json b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/acf23ef4-3e0f-4da0-a811-56761959df29_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5346177f0b5f8c7bedd0f90434574a8269d46baf --- /dev/null +++ b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/acf23ef4-3e0f-4da0-a811-56761959df29_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.842, + 0.177 + ], + "angle": 0, + "content": "Teller: Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.204, + 0.751, + 0.222 + ], + "angle": 0, + "content": "Dingcheng Zhen Shunshun Yin Shiyang Qin Hou Yi" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.223, + 0.713, + 0.24 + ], + "angle": 0, + "content": "Ziwei Zhang Siyuan Liu Gan Qi Ming Tao" + }, + { + "type": "text", + "bbox": [ + 0.366, + 0.24, + 0.662, + 0.258 + ], + "angle": 0, + "content": "Shanghai Soulgate Technolgy Co.tl." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.26, + 0.921, + 0.275 + ], + "angle": 0, + "content": "{dingchengzhen, yinshunshun, qinshiyang, houyi,zhangziwei, siyuanliu, ganqi, ming}@soulapp.cn" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.285, + 0.885, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.548, + 0.908, + 0.578 + ], + "angle": 0, + "content": "Figure 1. Teller framework is the first autoregressive framework for real-time, audio-driven portrait animation, achieving up to 25 FPS while preserving realistic body part and accessory movements. Demo can be found at https://teller-ayar.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.583, + 0.327, + 0.598 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.486, + 0.902 + ], + "angle": 0, + "content": "In this work, we introduce the first autoregressive framework for real-time, audio-driven portrait animation, a.k.a, talking head. Beyond the challenge of lengthy animation times, a critical challenge in realistic talking head generation lies in preserving the natural movement of diverse body parts. To this end, we propose Teller, the first streaming audio-driven portrait animation framework with autoregressive motion generation. Specifically, Teller first decomposes facial and body detail animation into two components: Facial Motion Latent Generation (FMLG) based on an autoregressive transformer, and movement authenticity refinement using a Efficient Temporal Module (ETM). Concretely, FMLG employs a Residual VQ model to map the facial motion latent from the implicit keypoint-based model into discrete motion tokens, which are then temporally sliced with audio embeddings. This enables the AR transformer to learn real-time, stream-based mappings from audio to motion. Furthermore, Teller incorporate ETM to capture finer motion details. This module ensures the physi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.735 + ], + "angle": 0, + "content": "cal consistency of body parts and accessories, such as neck muscles and earrings, improving the realism of these movements. Teller is designed to be efficient, surpassing the inference speed of diffusion-based models (Hallo 20.93s vs. Teller 0.92s for one second video generation), and achieves a real-time streaming performance of up to 25 FPS. Extensive experiments demonstrate that our method outperforms recent audio-driven portrait animation models, especially in small movements, as validated by human evaluations with a significant margin in quality and realism." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.755, + 0.646, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Realistic and expressive portrait animation from audio and static images, commonly known as talking head animation [12, 16, 21, 35], has garnered significant interest across applications such as virtual avatars, digital communication, and entertainment. However, generating high-quality animations that are visually compelling and temporally consistent remains a major challenge. This complexity stems from the need to intricately coordinate lip movements, fa" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "cial expressions, and head positioning to create lifelike effects. Moreover, achieving real-time, realistic talking head animation is complicated by computational constraints and the nuances of human movement, making this an especially demanding task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.412 + ], + "angle": 0, + "content": "While recent advancements, such as diffusion models [25, 27, 29, 42], have improved high-quality content generation, achieving controllable animation presents ongoing hurdles. Effective animation requires the accurate capture of complex facial expressions, body gestures, and the subtle interplay between them. Existing methods often suffer from prolonged animation times (See animation time in Tab 1), limiting their potential for real-time applications, and frequently fail to capture the natural, interconnected motions of various facial and body parts, e.g., earrings and necklaces, as shown in Figure. 1 stage 2. This shortfall results in animations with stiff or exaggerated movements that disrupt the realism of the animation (See Figure 7). Addressing these challenges necessitates a solution that balances computational efficiency with high animation quality without overloading processing resources." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.672 + ], + "angle": 0, + "content": "To this end, we propose Teller, the first autoregressive framework capable of real-time, streaming-based talking head animation at up to 25 FPS. As shown in Figure. 2, Teller employs a two-stage framework, combining Facial Motion Latent Generation (FMLG) and an Efficient Temporal Module (ETM) to produce realistic and physically consistent animations across facial and body movements. In the first stage, FMLG uses a Residual Vector Quantization (RVQ) model [45] to encode facial motion latents derived from an implicit keypoint-based model into discrete motion tokens. These tokens are then temporally aligned with audio embeddings, allowing an autoregressive (AR) transformer to map audio signals to facial movements in real-time. By breaking down the motion into temporal segments, FMLG enables the AR transformer to dynamically and efficiently generate high-quality animations responsive to live audio inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To enhance the realism of body movements, Teller introduces ETM in the second stage, which captures subtle details that are often overlooked in existing methods [12, 41]. The ETM refines finer movements, such as body parts and accessories, ensuring physically plausible interactions. For instance, it simulates realistic motions in neck muscles and dynamic accessories like earrings, which are vital for maintaining visual continuity in the animated avatar. Our Teller model prioritizes computational efficiency and is specifically designed to significantly outperform diffusion-based models in terms of inference speed. For example, generating a one-second video takes Hallo 20.93s, while Teller completes the task in just 0.92s (see Tab. 1). Despite this speed, Teller maintains high animation fidelity. By optimizing computational demands without compromising output" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "quality, Teller not only meets but exceeds real-time requirements, ensuring a smooth and responsive user experience." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.905, + 0.304 + ], + "angle": 0, + "content": "In extensive evaluations across various benchmarks and real-world settings, Teller demonstrates significant improvements over current state-of-the-art audio-driven portrait animation models, as shown in Table. 1, particularly in capturing nuanced facial and body movements. Human evaluations validate Teller's superior quality and realism, especially in rendering the subtle movements essential for lifelike animations. Our research represents a notable advancement in real-time talking head animation, presenting an innovative framework that bridges the gap between realism and efficiency in multimodal animation, as depicted in Figure. 6, Figure. 7, and Figure. 8." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.317, + 0.665, + 0.332 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.905, + 0.538 + ], + "angle": 0, + "content": "Non-Diffusion-Based Audio-Driven Portrait Animation This line of works typically consist of two key components: an audio-to-motion model and a facial motion representation model. These methods often utilize implicit keypoints as an intermediate motion representation, warping the source portrait based on the driving image. The goal of implicit methods is to learn disentangled representations in 2D [4, 18, 23, 34, 44, 49] or 3D [9, 35] latent spaces, with a focus on aspects such as identity, facial dynamics, and head pose. For example, FOMM [26] employs first-order Taylor expansion to capture local motion, while FaceVid2Vid [35] extends this by introducing a 3D implicit keypoints representation, enabling free-view portrait animation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.539, + 0.905, + 0.78 + ], + "angle": 0, + "content": "To effectively learn facial motion representations in latent space, these approaches often rely on GAN-based frameworks to disentangle identity-related appearance from non-identity-related motion, specifically capturing expressions, lip and eye movements, minor accessories, and poses. Examples include FaceVid2Vid [35], LivePortrait [12], and others, where identity and motion representations are independently learned to generate realistic talking head animations. For instance, MakeItTalk [50] uses an LSTM-based audio-to-motion model to predict landmark coordinates from audio input, which are then translated into video frames using a warp-based GAN model. Similarly, SadTalker [46] employs FaceVid2Vid [35] as an image synthesizer, with ExpNet and PoseVAE modules transforming audio features into inputs compatible with FaceVid2Vid for audio-to-video generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "However, these approaches face challenges due to GAN losses that primarily focus on facial expressions, lip, and eye movements, often neglecting accessory, hair, and body movements. This can result in animations that appear stiff or incomplete, lacking natural dynamism. In contrast, our work introduces the first autoregressive framework specifically designed for real-time, audio-driven portrait animation, achieving up to 25 FPS and delivering a more realistic," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21076" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.898, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.432, + 0.818, + 0.447 + ], + "angle": 0, + "content": "Figure 2. Overall framework of our proposed Teller for real-time streaming audio-driven portrait animation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.483, + 0.473 + ], + "angle": 0, + "content": "coherent portrayal of both facial and accessory movements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Diffusion-based Audio-driven Portrait Animation Recent advancements in diffusion-based video generation have demonstrated promising outcomes for audio-driven portrait animation. Methods like GAIA [14] and VASA-1 [41] have designed diffusion models to transform audio inputs into motion latents, facilitating audio-to-video generation. Further developments, such as EMO[32], Hallo [40], LOOPY [17], enhance end-to-end diffusion modeling by incorporating motion modules [3, 13, 16] and audio cross-attention mechanisms, improving the coherence and synchronization between audio cues and visual motion. Despite these improvements, a significant limitation of diffusion-based models remains their multi-step inference process, required to generate even a single frame or a few frames of video. This step-by-step prediction approach renders real-time performance challenging, as it is computationally intensive and time-consuming, making diffusion-based models less suitable for applications demanding instantaneous response." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "AR Transformer-Based Generation Recent works [1, 30, 36, 38, 43, 48] have focused on developing unified multimodal language models for generating visual content, such as images and videos. Some studies [28, 51] use autoregressive modeling with continuous representations interleaved with text tokens for image generation. Others, like SEED-X [11], propose a foundational system combining CLIP ViT-based image representations with text tokens for multimodal tasks, including next-token prediction and image rep" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.52 + ], + "angle": 0, + "content": "resentation regression. DreamLLM [8] also explores multimodal understanding and creation, while Chameleon [31] introduces token-based models capable of both understanding and generating images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.606, + 0.545 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.692 + ], + "angle": 0, + "content": "As shown in Fig. 2, Teller comprises two main modules: the Facial Motion Latent Generation (FMLG) and the Efficient Temporal Module (ETM). FMLG module integrates an autoregressive transformer and a residual vector quantization (RVQ) component. It applies an autoregressive transformer to generate discrete facial motion tokens from audio input. Following FMLG, ETM refines the generated motion to produce realistic body and accessory movements, ensuring physical consistency in animated results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.655, + 0.713 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.796 + ], + "angle": 0, + "content": "Prior works, such as LivePortrait [12] have introduced methods for extracting implicit keypoints as facial motion latents using motion and appearance extractors. These motion latents capture essential facial dynamics needed for animating input images and consist of three main components:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.797, + 0.905, + 0.856 + ], + "angle": 0, + "content": "- Expression Deformation \\((\\delta)\\): A set deformation of 21 implicit keypoints, represented as \\(\\delta = [\\delta_1, \\delta_2, \\dots, \\delta_{21}]\\), where each \\(\\delta\\) is a 3D vector \\((\\mathbb{R}^3)\\) that indicates the position of the \\(i\\)-th deformation of facial keypoints." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Head Pose \\((R)\\): Defined by three rotation vectors \\(R = [r_1, r_2, r_3]\\), with each \\(r_i\\) being a 3D vector \\((\\mathbb{R}^3)\\) that describes the head's orientation in 3D space." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.797, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21077" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.484, + 0.121 + ], + "angle": 0, + "content": "- Expression Deformation \\((t)\\): A single 3D vector \\((t \\in \\mathbb{R}^3)\\) that captures facial expression deformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.483, + 0.152 + ], + "angle": 0, + "content": "These components are concatenated into a unified motion. \nlatent of size \\(25\\times 3\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.157, + 0.483, + 0.173 + ], + "angle": 0, + "content": "\\[\nm = \\left[ \\delta_ {1}, \\delta_ {2}, \\dots , \\delta_ {2 1}, r _ {1}, r _ {2}, r _ {3}, t \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.483, + 0.207 + ], + "angle": 0, + "content": "where \\(m \\in \\mathbb{R}^{25 \\times 3}\\) includes the 21 keypoints, head pose, and expression deformation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.218, + 0.451, + 0.234 + ], + "angle": 0, + "content": "3.2. Facial Motion Latent Generation (FMLG)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.241, + 0.484, + 0.348 + ], + "angle": 0, + "content": "In FMLG, Teller generates motion latent from facial motion extraction, then encodes it into discrete tokens using a residual vector quantizer (RVQ). To optimize encoding, the concatenated motion latent \\( m \\) is processed with RVQ, leveraging temporal redundancy across \\( T \\) frames for efficient compression. The RVQ quantization loss, which encodes the continuous \\( m \\) into discrete tokens, is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.371, + 0.495, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v q} = \\sum_ {t = 1} ^ {T} \\left[ \\underbrace {\\left| \\left| m - \\operatorname {F F N} _ {d e c} \\left(z _ {t} + \\operatorname {s g} [ \\hat {z} _ {t} - z _ {t} ]\\right) \\right| \\right| _ {2} ^ {2}} _ {\\mathcal {L} _ {\\text {r e c o n}}} + \\underbrace {\\left| \\left| z _ {t} - \\operatorname {s g} [ \\hat {z} _ {t} ] \\right| \\right| _ {2} ^ {2}} _ {\\mathcal {L} _ {\\text {c o m m i t}}} \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.495 + ], + "angle": 0, + "content": "where \\( z_{t} = \\mathrm{FFN}_{enc}(m) \\) represents the encoded latent at each time step \\( t \\), \\( \\hat{z}_t \\) is the quantized latent, and \\( \\mathrm{sg} \\) denotes the stop-gradient operation. The two loss components, \\( \\mathcal{L}_{\\mathrm{recon}} \\) and \\( \\mathcal{L}_{\\mathrm{commit}} \\), are defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.499, + 0.483, + 0.543 + ], + "angle": 0, + "content": "- Reconstruction Loss \\((\\mathcal{L}_{\\mathrm{recon}})\\): Minimizes the difference between the original motion latent \\(m\\) and the decoded quantized latent." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.544, + 0.483, + 0.576 + ], + "angle": 0, + "content": "- Commitment Loss \\((\\mathcal{L}_{\\mathrm{commit}})\\): Encourages \\(z_{t}\\) to approach the quantized latent \\(\\hat{z}_t\\), ensuring stability in quantization." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.499, + 0.483, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.484, + 0.686 + ], + "angle": 0, + "content": "This approach enables FMLG to learn robust, temporally-consistent representations, translating audio-driven inputs into lifelike facial animations. Experimentally, to achieve trade-offs between frame count and redundancy, we selected 4 frames \\((4\\times 25\\times 3\\) latent) to be compressed into 32 tokens (hyper-parameter selection trade-off refer to Fig. 11)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.434, + 0.712 + ], + "angle": 0, + "content": "3.3. AR Transformer for Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.736 + ], + "angle": 0, + "content": "Using the learned RVQ-based latents, motion is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.752, + 0.483, + 0.77 + ], + "angle": 0, + "content": "\\[\nM = \\left[ m _ {1}, m _ {2}, \\dots , m _ {T} \\right], \\quad \\text {w h e r e} \\quad m _ {i} \\in \\mathbb {R} ^ {2 5 \\times 3}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.803 + ], + "angle": 0, + "content": "This sequence of motion latents is converted to discrete tokens using the RVQ module:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.806, + 0.483, + 0.824 + ], + "angle": 0, + "content": "\\[\nT _ {m} = \\left[ t _ {1}, t _ {2}, \\dots , t _ {T / 4} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\( T_{m} \\) represents the quantized motion tokens. Audio input is encoded with the Whisper encoder [5], generating the audio condition \\( c \\). Motion generation is modeled as a next-token prediction task, where the distribution of each token is predicted based on the previous \\( t - 1 \\) tokens and" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.094, + 0.897, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 3. In our Teller, we follow AR transformer architecture, but each input consists of a pair of tokens and model pred a pair of tokens for each output position." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.351, + 0.637, + 0.364 + ], + "angle": 0, + "content": "audio condition \\(c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.368, + 0.905, + 0.385 + ], + "angle": 0, + "content": "\\[\nP \\left(t _ {i} \\mid c, t _ {< i}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.418 + ], + "angle": 0, + "content": "This autoregressive setup enables sequential generation of motion based on past motion tokens and audio embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.906, + 0.6 + ], + "angle": 0, + "content": "To enable real-time streaming animation, we process both audio and video frames in 200ms chunks, following Whisper's constraints. Each audio chunk is encoded into a \\([10 \\times 512]\\) embedding, while each video chunk uses 32 learned RVQ-based motion tokens. For efficient real-time performance, we enable the autoregressive transformer to process token pairs at each position, which improves prediction speed by processing two tokens concurrently. In Teller, each input consists of a token pair with a combined embedding and a learnable position bias inspired by BERT, capturing relative token positions. The loss for each head, representing each token in the pair, is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.602, + 0.905, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {h e a d} 0 _ {j}} = \\mathrm {C E} \\left(\\operatorname {l a b e l} _ {j} [ 0 ], \\mathrm {T} \\left(\\operatorname {i n p u t} _ {j} [ 0 ] \\mid \\operatorname {i n p u t} _ {< j}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.638, + 0.905, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {h e a d} 1 _ {j}} = \\mathrm {C E} \\left(\\operatorname {l a b e l} _ {j} [ 1 ], \\mathrm {T} \\left(\\operatorname {i n p u t} _ {j} [ 1 ] \\mid \\operatorname {i n p u t} _ {< j}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.662, + 0.905, + 0.693 + ], + "angle": 0, + "content": "where \\(T\\) denotes the transformer. The total loss, with a regularization term to balance learning across both heads, is:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.708, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a r} = \\sum_ {j = 1} ^ {I / 2} \\left[ \\mathcal {L} _ {\\text {h e a d} 0 _ {j}} + \\mathcal {L} _ {\\text {h e a d} 1 _ {j}} + \\left\\| \\mathcal {L} _ {\\text {h e a d} 0 _ {j}} - \\mathcal {L} _ {\\text {h e a d} 1 _ {j}} \\right\\| _ {2} ^ {2} \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.905, + 0.809 + ], + "angle": 0, + "content": "This regularization term, \\(\\| \\mathcal{L}_{\\mathrm{head0}} - \\mathcal{L}_{\\mathrm{head1}}\\| _2^2\\) ensures balanced training across the two heads, promoting stable and accurate real-time animation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "After testing frame count and redundancy, we use 4 frames \\((4\\times 25\\times 3\\) latent), compressed into 32 tokens (refer to Fig. 11). Then the 4 frames are interpolated to 5 frames for faster generation. This setup balances inference efficiency and feature quality, enabling real-time, high-fidelity streaming portrait animation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21078" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.511, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.905, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.409, + 0.908, + 0.438 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with existing approaches on RAVDESS data-set of 'angry' and 'disgust' emotion cases. Videos are available in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.446, + 0.455, + 0.463 + ], + "angle": 0, + "content": "3.4. Efficient Temporal Module for Refinement" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.474, + 0.483, + 0.701 + ], + "angle": 0, + "content": "In diffusion-based models, temporal layers are added to text-to-image (T2I) frameworks to capture frame dependencies [16]. Inspired by this, Teller incorporates temporal refinement but achieves it in a single step, unlike diffusion models that require multiple iterations, enhancing real-time efficiency. After encoding video frames with a VAE encoder, we apply a 3D U-Net [7] to extract features from the image sequence, represented as \\( x \\in \\mathbb{R}^{b \\times t \\times h \\times w \\times c} \\), where \\( b \\) is the batch size, \\( t \\) the frame count, \\( h \\) and \\( w \\) the frame dimensions, and \\( c \\) the channel count. The features are reshaped to \\( x \\in \\mathbb{R}^{(b \\times h \\times w) \\times t \\times c} \\), enabling ETM to perform self-attention along the temporal dimension \\( t \\). ETM's output is then merged with the original features through residual connections, integrating temporal dependencies into spatial features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.484, + 0.782 + ], + "angle": 0, + "content": "For training, we use the first 5 frames from a real image sequence and the subsequent 5 frames reconstructed with LivePortrait [12]. After processing through ETM, we compute reconstruction loss between the predicted and ground-truth frames:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.784, + 0.483, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\sum_ {i = 6} ^ {1 0} \\left\\| x _ {\\mathrm {g t} _ {i}} - f \\left(x _ {i} \\mid x _ {\\mathrm {g t} _ {< 6}}\\right) \\right\\| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(x_{\\mathrm{gt}_i}\\) denotes the real feature sequence and \\(x_{i}\\) the reconstructed feature sequence (from LivePortrait or the Stage 1 decoder). ETM primarily learns to preserve consistency in physical features such as neck muscles and earrings. To this end, we add a region-specific mask to recon" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.446, + 0.785, + 0.461 + ], + "angle": 0, + "content": "struction loss, and the final loss of ETM:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.472, + 0.905, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E T M}} = \\sum_ {i = 6} ^ {1 0} \\left\\| x _ {\\mathrm {g t} _ {i}} \\odot \\operatorname {m a s k} _ {i} - f \\left(x _ {i} \\mid x _ {\\mathrm {g t} _ {< 6}}\\right) \\odot \\operatorname {m a s k} _ {i} \\right\\| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.526, + 0.905, + 0.557 + ], + "angle": 0, + "content": "where element-wise multiplication with \\(\\mathrm{mask}_i\\) focuses reconstruction on specific regions. The mask is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.568, + 0.905, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\operatorname {m a s k} (i, j) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} (i, j) \\text {i s w i t h i n} \\mathrm {B B} (x) \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.69 + ], + "angle": 0, + "content": "where \\(\\mathrm{BB(x)}\\) stands for the bounding box and outlines relevant body parts. Key landmarks are identified using MediaPipe[22], with points indices such as [93, 323, 152] defining the bounding boxes for these regions, highlighted in red in the stage 2 input of Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.703, + 0.646, + 0.72 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.704, + 0.745 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Datasets. For training, we used the AV Speech datasets (filtered to 662 hours) [10] and VFHQ [39] datasets (filtered to 2 hours) for pretraining, along with additional talking-head videos from the internet (32 hours) for supervised finetuning (SFT). For validation, we used the HDTF (filtered to 0.83 hours) [47] and RAVDESS (filtered to 0.55 hours) [19] datasets and supplementary internet data (0.49 hours) for qualitative comparisons and human evaluation only. To ensure quality, we applied the Mediapipe [22] face detection tool to filter out instances with facial movement exceeding" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21079" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.097, + 0.481, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison with existing approaches on HDTF dataset. Videos are available in the supplement. mat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.455 + ], + "angle": 0, + "content": "\\(50\\%\\). We further refined the data using Sync-C and Sync-D to exclude samples with low lip-sync scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.563 + ], + "angle": 0, + "content": "Metrics. Evaluation metrics include Fréchet Inception Distance (FID)[15], Fréchet Video Distance (FVD)[33], Synchronization-C (Sync-C)[24], and Synchronization-D (Sync-D)[24]. FID and FVD assess realism, with lower scores indicating better quality, while Sync-C and Sync-D measure lip synchronization, with higher Sync-C and lower Sync-D values indicating better alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.265, + 0.58 + ], + "angle": 0, + "content": "Implementation Details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Stage 1: During pretraining, we follow the architecture design of the Qwen1.5-4B model [2] and initialize the parameters randomly. The model is trained on an \\(8 \\times 8\\) Nvidia A800 GPU machine with a batch size of 1024, using the AdamW optimizer [20]. We employ a cosine learning rate scheduler, with the learning rate decaying from 1e-4 to 1e-6 over 40 epochs. In the supervised fine-tuning (SFT) phase, we again use an \\(8 \\times 8\\) Nvidia A800 GPU machine, but with a batch size of 512. The AdamW optimizer [20] is used, along with a cosine learning rate scheduler. The learning rate decaying from 1e-5 to 1e-6 over 10 epochs. Stage 2: The model is trained on an \\(8 \\times 8\\) Nvidia A800 GPU machine, with a batch size of 1024, and using the AdamW optimizer. The cosine learning rate scheduler is used, with the learning rate decaying from 1e-4 to 1e-6 over 30 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Real time analysis. The model is inference on an 4 Nvidia H800 GPU machine. For a 200ms audio input, the average processing time of the Whisper encoder is 7ms. In Stage 1, the average total time is 106ms, with the AR transformer taking an average of 6ms per 16 tokens, and the motion decoder taking an average of 10ms. In Stage 2, the average" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.19 + ], + "angle": 0, + "content": "
MethodFID↓FVD↓Sync-C↑Sync-D↓Time
SadTalker [46]22.177233.6737.3267.84818.89s
EchoMimic [6]23.049290.1906.6648.83931.10s
AniPortrait [37]28.161235.0994.54710.65729.36s
Hallo [40]20.639174.1917.4977.74120.93s
Teller (ours)21.352173.4637.6967.5360.92s
Real video--8.0946.976-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.194, + 0.905, + 0.236 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with existing portrait image animation approaches on the HDTF dataset. Time stands for the averaging time cost of generating one second of 25 fps video." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.237, + 0.905, + 0.353 + ], + "angle": 0, + "content": "
MethodFID↓FVD↓Sync-C↑Sync-D↓
SadTalker [46]32.343487.9244.3047.621
EchoMimic [6]21.058668.6753.2929.096
AniPortrait [37]30.696476.1972.32111.542
Hallo [40]19.826537.4784.0628.552
Teller (ours)20.352429.2884.4967.936
Real video--5.2237.069
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.355, + 0.905, + 0.384 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with existing portrait image animation approaches on the RAVDESS dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.396, + 0.905, + 0.427 + ], + "angle": 0, + "content": "total time is 71ms, with the VAE encoder and decoder averaging 25ms, and the Temporal Module averaging 21ms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.437, + 0.707, + 0.453 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Comparison on the HDTF Dataset. Table 1 presents quantitative results for portrait animation techniques on the HDTF dataset. Our method outperforms others, achieving the lowest FVD of 173.463 and a competitive FID of 21.352, indicating high quality and temporal coherence in animated talking heads. Additionally, it achieves the highest Sync-C score of 7.696 and the lowest Sync-D score of 7.536, demonstrating excellent lip synchronization. These results highlight Teller's effectiveness in maintaining both visual fidelity and synchronization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.747 + ], + "angle": 0, + "content": "Comparison on the RAVDESS Dataset. Table 2 shows quantitative results on the RAVDESS dataset. Our method again leads in performance, with the lowest FVD of 429.288 and a competitive FID of 20.352, reflecting high-quality and temporally coherent animations. The highest Sync-C score of 4.496 and a competitive Sync-D score of 7.936 further demonstrate superior lip synchronization. These findings confirm Teller's strength in producing synchronized and high-fidelity animated portraits." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.696, + 0.773 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Head Movement Comparison. Figure 5 shows a qualitative comparison of head movements. Teller replicates natural head movements more accurately, closely matching the ground truth (GT) with smooth, realistic turns and subtle expression-based adjustments. Competing methods, like AntiPortrait and EchoMimic, often show abrupt or limited movements, appearing rigid or lifeless. Teller's autoregressive framework ensures continuity and natural dynamics in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21080" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.48, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.293, + 0.483, + 0.321 + ], + "angle": 0, + "content": "Figure 6. Top-k selection \\((\\mathrm{k} = 15)\\) in FMLG produces diverse facial expressions and actions with accurate lip sync on the HDTF." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.323, + 0.481, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.618 + ], + "angle": 0, + "content": "Figure 7. Visualization of finer motion details. Videos are available in the supplementary materials." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.625, + 0.482, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.738, + 0.473, + 0.753 + ], + "angle": 0, + "content": "Figure 8. Visualization of the generation accuracy of lip shape." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.794 + ], + "angle": 0, + "content": "head and emotion alignment, essential for lifelike animation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Diversity. Fig. 6 shows two sequences generated with Top-k sampling (k=15) in FMLG, where each row represents frames from the same speech input. The model demonstrates diverse facial expressions and head movements while maintaining accurate lip sync, highlighting Top-k sampling's role in enhancing motion variety without compromising synchronization." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.644, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.091, + 0.774, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.091, + 0.903, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.246, + 0.905, + 0.271 + ], + "angle": 0, + "content": "Figure 9. Human evaluation results among our proposed Teller and other SoTA methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.274, + 0.905, + 0.319 + ], + "angle": 0, + "content": "Emotional Expression. Fig. 4 presents our model has more accurate emotional expression ability due to the better speech understanding ability of AR transformer." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.906, + 0.424 + ], + "angle": 0, + "content": "Finer Motion Details. Fig. 7 highlights fine motion details in neck and earring movements. Compared to other methods, Teller produces realistic, nuanced motions synchronized with speech, capturing subtle audio-driven dynamics that contribute to lifelike and temporally coherent animation. The consistent detail across frames underscores Teller's robustness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.425, + 0.906, + 0.501 + ], + "angle": 0, + "content": "Lip Synchronization Accuracy. Fig. 8 shows superior lip synchronization, with Teller aligning generated lip shapes closely to natural movements. This high fidelity highlights Teller's strength in producing realistic, synchronized mouth motions for convincing talking head animation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.509, + 0.696, + 0.523 + ], + "angle": 0, + "content": "4.4. Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.728 + ], + "angle": 0, + "content": "We conducted a human evaluation to assess the quality of generated animations, focusing on lip synchronization, body movement realism, and temporal coherence. Thirty participants (66.7% aged 24-30, 33.3% aged 30-40; 30% male, 70% female; 83.3% with AIGC model experience) rated each animation on a 5-point Likert scale for coherence with input and animation quality. A total of 100 videos were presented in random order to avoid bias, providing insights into subjective perceptions of animation quality and natural expression alignment. As shown in Figure 9, participants rated Teller highest in lip synchronization, body movement realism, and temporal coherence, with low variance in scores, indicating robust and consistent performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.74, + 0.665, + 0.758 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Stage (Module) Ablation. We perform an ablation study comparing Stage 1 and Stage 2, focusing on animation quality in body parts and accessories, especially neck muscles and earrings. As shown in Fig. 10, Stage 1 outputs often show less realistic, inconsistent movements. In contrast, Stage 2, with ETM, significantly improves the physical consistency of subtle motions, creating natural earring sway and smooth neck movements for lifelike, temporally coherent animation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21081" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.34 + ], + "angle": 0, + "content": "Figure 10. Visualization of the generation images with different stages and the differences between stages." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.343, + 0.479, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.53 + ], + "angle": 0, + "content": "Figure 11. Tradeoff between performance (loss) and different compression(tokens/frame) ratios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Frames / Tokens in RVQ Module Figure 11 shows the impact of frame and token configurations in the RVQ module on face shape accuracy in talking head animations. This ablation study explores how varying frames and tokens affects facial movement quality and synchronization. Results show that increasing frames and tokens improves facial dynamics, enhancing lip synchronization and realism, though with higher computational costs. We selected 4 frames and 32 tokens for an optimal balance, adjustable as needed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Ablation Study on Audio Condition Encoder. We analyzed the performance differences between TTS and ASR models as audio condition encoders for our task. Specifically, we used Whisper for the ASR model and funcodec for the TTS model. As shown in Table 3, Whisper achieved a Sync-C score of 7.696 and a Sync-D score of 7.536, while funcodec scored 4.286 in Sync-C and 10.373 in Sync-D. These results indicate that our task benefits more from the ASR model's capability to capture the nuances required for precise synchronization, as seen in Whisper's higher Sync-C and lower Sync-D scores. This analysis suggests that the ASR model is better suited as the audio condition encoder, enhancing the overall quality and synchronization of our talking head animation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation Study on Single-Head vs. Multi-Head Archi" + }, + { + "type": "table", + "bbox": [ + 0.578, + 0.079, + 0.838, + 0.138 + ], + "angle": 0, + "content": "
MethodSync-C ↑Sync-D ↓
w funcdec4.28610.373
w whisper7.6967.536
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.141, + 0.907, + 0.17 + ], + "angle": 0, + "content": "Table 3. Comparison of synchronization for audio conditions using funcodec and Whisper in ASR and TTS tasks on HDTF." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.173, + 0.912, + 0.234 + ], + "angle": 0, + "content": "
MethodFID↓FVD↓Sync-C↑Sync-D↓
Single-Head22.110172.5537.7907.474
Multi-Head21.352173.4637.6967.536
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.235, + 0.907, + 0.262 + ], + "angle": 0, + "content": "Table 4. Comparison of performance between Single-Head and Multi-Head on the HDTF dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.412 + ], + "angle": 0, + "content": "tecture. We compare single-head and multi-head models on FID, FVD, Sync-C, and Sync-D metrics (Table 4). The single-head model slightly outperforms in FID (22.110 vs. 21.352) and has comparable FVD (172.553 vs. 173.463), with marginally better synchronization (Sync-C of 7.790 vs. 7.696 and Sync-D of 7.474 vs. 7.536). Both architectures yield competitive results, though the single-head model slightly excels in synchronization, while the multi-head model offers greater real-time potential." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.424, + 0.634, + 0.44 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.909, + 0.707 + ], + "angle": 0, + "content": "In this paper, we presented Teller, the first autoregressive framework designed for real-time, audio-driven portrait animation. Addressing the challenge of realistic and efficient talking head generation, Teller achieves high-quality animations at up to 25 FPS, surpassing existing methods in both fidelity and responsiveness. Extensive experiments demonstrated Teller's advantages over SoTA audio-driven animation methods, particularly in rendering nuanced movements essential for lifelike and visually convincing animations. Human evaluations further validate its quality, particularly in natural expression and lip synchronization. By balancing computational efficiency with high animation fidelity, Teller sets a new standard for real-time talking head animation, marking a significant advancement in multimodal portrait animation frameworks. Additionally, Teller's AR Transformer architecture makes it compatible with existing unified multimodal language models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21082" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.214 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[4] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[5] Nan Cao, Yu-Ru Lin, Xiaohua Sun, David Lazer, Shixia Liu, and Huamin Qu. Whisper: Tracing the spatiotemporal process of information diffusion in real time. IEEE transactions on visualization and computer graphics, 18(12):2649-2658, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[6] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[7] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19, pages 424-432. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[8] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[10] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T. Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: a speaker-independent audio-visual model for speech separation. ACM Transactions on Graphics, 37(4):1-11, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[11] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[12] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[14] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[16] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[17] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[18] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[19] Steven R Livingstone and Frank A Russo. The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5): e0196391, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[20] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[21] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (ToG), 40(6):1-17, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[22] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[23] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 427-436, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21083" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "28th ACM international conference on multimedia, pages 484-492, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[25] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[26] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[27] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[28] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[29] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[30] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[31] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models, 2024. URL https://arxiv.org/abs/2405.09818, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[32] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[33] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[34] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17979-17989, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[35] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[36] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[38] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[39] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[40] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[41] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[42] Ziyu Yao, Xuxin Cheng, and Zhiqi Huang. Fd2talk: Towards generalized talking head generation with facial decoupled diffusion model. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3411-3420, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[43] Hanrong Ye, De-An Huang, Yao Lu, Zhiding Yu, Wei Ping, Andrew Tao, Jan Kautz, Song Han, Dan Xu, Pavlo Molchanov, et al. X-vila: Cross-modality alignment for large language model. arXiv preprint arXiv:2405.19335, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[44] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[45] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[46] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[47] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[48] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21084" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176-4186, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[50] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[51] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "21085" + } + ] +] \ No newline at end of file diff --git a/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/acf23ef4-3e0f-4da0-a811-56761959df29_origin.pdf b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/acf23ef4-3e0f-4da0-a811-56761959df29_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f3bf78d14a82bf1dea4959c4c41e08d46d8f9f6 --- /dev/null +++ b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/acf23ef4-3e0f-4da0-a811-56761959df29_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d75cf1ba1fd4bd82e0ce19988cb154355a94f59f8ad83f6eb7daca8c729eda74 +size 7934295 diff --git a/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/full.md b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7fa59752a3635533901fc4a0708780f88ce367e8 --- /dev/null +++ b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/full.md @@ -0,0 +1,327 @@ +# Teller: Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation + +Dingcheng Zhen Shunshun Yin Shiyang Qin Hou Yi + +Ziwei Zhang Siyuan Liu Gan Qi Ming Tao + +Shanghai Soulgate Technolgy Co.tl. + +{dingchengzhen, yinshunshun, qinshiyang, houyi,zhangziwei, siyuanliu, ganqi, ming}@soulapp.cn + +![](images/64a15499ee037e75efd44f23a583cd72d6c4540f11d8795d571e003bb86d2ca5.jpg) +Figure 1. Teller framework is the first autoregressive framework for real-time, audio-driven portrait animation, achieving up to 25 FPS while preserving realistic body part and accessory movements. Demo can be found at https://teller-ayar.github.io/. + +# Abstract + +In this work, we introduce the first autoregressive framework for real-time, audio-driven portrait animation, a.k.a, talking head. Beyond the challenge of lengthy animation times, a critical challenge in realistic talking head generation lies in preserving the natural movement of diverse body parts. To this end, we propose Teller, the first streaming audio-driven portrait animation framework with autoregressive motion generation. Specifically, Teller first decomposes facial and body detail animation into two components: Facial Motion Latent Generation (FMLG) based on an autoregressive transformer, and movement authenticity refinement using a Efficient Temporal Module (ETM). Concretely, FMLG employs a Residual VQ model to map the facial motion latent from the implicit keypoint-based model into discrete motion tokens, which are then temporally sliced with audio embeddings. This enables the AR transformer to learn real-time, stream-based mappings from audio to motion. Furthermore, Teller incorporate ETM to capture finer motion details. This module ensures the physi + +cal consistency of body parts and accessories, such as neck muscles and earrings, improving the realism of these movements. Teller is designed to be efficient, surpassing the inference speed of diffusion-based models (Hallo 20.93s vs. Teller 0.92s for one second video generation), and achieves a real-time streaming performance of up to 25 FPS. Extensive experiments demonstrate that our method outperforms recent audio-driven portrait animation models, especially in small movements, as validated by human evaluations with a significant margin in quality and realism. + +# 1. Introduction + +Realistic and expressive portrait animation from audio and static images, commonly known as talking head animation [12, 16, 21, 35], has garnered significant interest across applications such as virtual avatars, digital communication, and entertainment. However, generating high-quality animations that are visually compelling and temporally consistent remains a major challenge. This complexity stems from the need to intricately coordinate lip movements, fa + +cial expressions, and head positioning to create lifelike effects. Moreover, achieving real-time, realistic talking head animation is complicated by computational constraints and the nuances of human movement, making this an especially demanding task. + +While recent advancements, such as diffusion models [25, 27, 29, 42], have improved high-quality content generation, achieving controllable animation presents ongoing hurdles. Effective animation requires the accurate capture of complex facial expressions, body gestures, and the subtle interplay between them. Existing methods often suffer from prolonged animation times (See animation time in Tab 1), limiting their potential for real-time applications, and frequently fail to capture the natural, interconnected motions of various facial and body parts, e.g., earrings and necklaces, as shown in Figure. 1 stage 2. This shortfall results in animations with stiff or exaggerated movements that disrupt the realism of the animation (See Figure 7). Addressing these challenges necessitates a solution that balances computational efficiency with high animation quality without overloading processing resources. + +To this end, we propose Teller, the first autoregressive framework capable of real-time, streaming-based talking head animation at up to 25 FPS. As shown in Figure. 2, Teller employs a two-stage framework, combining Facial Motion Latent Generation (FMLG) and an Efficient Temporal Module (ETM) to produce realistic and physically consistent animations across facial and body movements. In the first stage, FMLG uses a Residual Vector Quantization (RVQ) model [45] to encode facial motion latents derived from an implicit keypoint-based model into discrete motion tokens. These tokens are then temporally aligned with audio embeddings, allowing an autoregressive (AR) transformer to map audio signals to facial movements in real-time. By breaking down the motion into temporal segments, FMLG enables the AR transformer to dynamically and efficiently generate high-quality animations responsive to live audio inputs. + +To enhance the realism of body movements, Teller introduces ETM in the second stage, which captures subtle details that are often overlooked in existing methods [12, 41]. The ETM refines finer movements, such as body parts and accessories, ensuring physically plausible interactions. For instance, it simulates realistic motions in neck muscles and dynamic accessories like earrings, which are vital for maintaining visual continuity in the animated avatar. Our Teller model prioritizes computational efficiency and is specifically designed to significantly outperform diffusion-based models in terms of inference speed. For example, generating a one-second video takes Hallo 20.93s, while Teller completes the task in just 0.92s (see Tab. 1). Despite this speed, Teller maintains high animation fidelity. By optimizing computational demands without compromising output + +quality, Teller not only meets but exceeds real-time requirements, ensuring a smooth and responsive user experience. + +In extensive evaluations across various benchmarks and real-world settings, Teller demonstrates significant improvements over current state-of-the-art audio-driven portrait animation models, as shown in Table. 1, particularly in capturing nuanced facial and body movements. Human evaluations validate Teller's superior quality and realism, especially in rendering the subtle movements essential for lifelike animations. Our research represents a notable advancement in real-time talking head animation, presenting an innovative framework that bridges the gap between realism and efficiency in multimodal animation, as depicted in Figure. 6, Figure. 7, and Figure. 8. + +# 2. Related Works + +Non-Diffusion-Based Audio-Driven Portrait Animation This line of works typically consist of two key components: an audio-to-motion model and a facial motion representation model. These methods often utilize implicit keypoints as an intermediate motion representation, warping the source portrait based on the driving image. The goal of implicit methods is to learn disentangled representations in 2D [4, 18, 23, 34, 44, 49] or 3D [9, 35] latent spaces, with a focus on aspects such as identity, facial dynamics, and head pose. For example, FOMM [26] employs first-order Taylor expansion to capture local motion, while FaceVid2Vid [35] extends this by introducing a 3D implicit keypoints representation, enabling free-view portrait animation. + +To effectively learn facial motion representations in latent space, these approaches often rely on GAN-based frameworks to disentangle identity-related appearance from non-identity-related motion, specifically capturing expressions, lip and eye movements, minor accessories, and poses. Examples include FaceVid2Vid [35], LivePortrait [12], and others, where identity and motion representations are independently learned to generate realistic talking head animations. For instance, MakeItTalk [50] uses an LSTM-based audio-to-motion model to predict landmark coordinates from audio input, which are then translated into video frames using a warp-based GAN model. Similarly, SadTalker [46] employs FaceVid2Vid [35] as an image synthesizer, with ExpNet and PoseVAE modules transforming audio features into inputs compatible with FaceVid2Vid for audio-to-video generation. + +However, these approaches face challenges due to GAN losses that primarily focus on facial expressions, lip, and eye movements, often neglecting accessory, hair, and body movements. This can result in animations that appear stiff or incomplete, lacking natural dynamism. In contrast, our work introduces the first autoregressive framework specifically designed for real-time, audio-driven portrait animation, achieving up to 25 FPS and delivering a more realistic, + +![](images/716343c816bdff653c8ce0ada9d2afffe62239668bc8ac48b6490435d88363d9.jpg) +Figure 2. Overall framework of our proposed Teller for real-time streaming audio-driven portrait animation. + +coherent portrayal of both facial and accessory movements. + +Diffusion-based Audio-driven Portrait Animation Recent advancements in diffusion-based video generation have demonstrated promising outcomes for audio-driven portrait animation. Methods like GAIA [14] and VASA-1 [41] have designed diffusion models to transform audio inputs into motion latents, facilitating audio-to-video generation. Further developments, such as EMO[32], Hallo [40], LOOPY [17], enhance end-to-end diffusion modeling by incorporating motion modules [3, 13, 16] and audio cross-attention mechanisms, improving the coherence and synchronization between audio cues and visual motion. Despite these improvements, a significant limitation of diffusion-based models remains their multi-step inference process, required to generate even a single frame or a few frames of video. This step-by-step prediction approach renders real-time performance challenging, as it is computationally intensive and time-consuming, making diffusion-based models less suitable for applications demanding instantaneous response. + +AR Transformer-Based Generation Recent works [1, 30, 36, 38, 43, 48] have focused on developing unified multimodal language models for generating visual content, such as images and videos. Some studies [28, 51] use autoregressive modeling with continuous representations interleaved with text tokens for image generation. Others, like SEED-X [11], propose a foundational system combining CLIP ViT-based image representations with text tokens for multimodal tasks, including next-token prediction and image rep + +resentation regression. DreamLLM [8] also explores multimodal understanding and creation, while Chameleon [31] introduces token-based models capable of both understanding and generating images. + +# 3. Method + +As shown in Fig. 2, Teller comprises two main modules: the Facial Motion Latent Generation (FMLG) and the Efficient Temporal Module (ETM). FMLG module integrates an autoregressive transformer and a residual vector quantization (RVQ) component. It applies an autoregressive transformer to generate discrete facial motion tokens from audio input. Following FMLG, ETM refines the generated motion to produce realistic body and accessory movements, ensuring physical consistency in animated results. + +# 3.1. Preliminaries + +Prior works, such as LivePortrait [12] have introduced methods for extracting implicit keypoints as facial motion latents using motion and appearance extractors. These motion latents capture essential facial dynamics needed for animating input images and consist of three main components: + +- Expression Deformation $(\delta)$ : A set deformation of 21 implicit keypoints, represented as $\delta = [\delta_1, \delta_2, \dots, \delta_{21}]$ , where each $\delta$ is a 3D vector $(\mathbb{R}^3)$ that indicates the position of the $i$ -th deformation of facial keypoints. +- Head Pose $(R)$ : Defined by three rotation vectors $R = [r_1, r_2, r_3]$ , with each $r_i$ being a 3D vector $(\mathbb{R}^3)$ that describes the head's orientation in 3D space. + +- Expression Deformation $(t)$ : A single 3D vector $(t \in \mathbb{R}^3)$ that captures facial expression deformations. + +These components are concatenated into a unified motion. +latent of size $25\times 3$ .. + +$$ +m = \left[ \delta_ {1}, \delta_ {2}, \dots , \delta_ {2 1}, r _ {1}, r _ {2}, r _ {3}, t \right], \tag {1} +$$ + +where $m \in \mathbb{R}^{25 \times 3}$ includes the 21 keypoints, head pose, and expression deformation. + +# 3.2. Facial Motion Latent Generation (FMLG) + +In FMLG, Teller generates motion latent from facial motion extraction, then encodes it into discrete tokens using a residual vector quantizer (RVQ). To optimize encoding, the concatenated motion latent $m$ is processed with RVQ, leveraging temporal redundancy across $T$ frames for efficient compression. The RVQ quantization loss, which encodes the continuous $m$ into discrete tokens, is defined as: + +$$ +\mathcal {L} _ {v q} = \sum_ {t = 1} ^ {T} \left[ \underbrace {\left| \left| m - \operatorname {F F N} _ {d e c} \left(z _ {t} + \operatorname {s g} [ \hat {z} _ {t} - z _ {t} ]\right) \right| \right| _ {2} ^ {2}} _ {\mathcal {L} _ {\text {r e c o n}}} + \underbrace {\left| \left| z _ {t} - \operatorname {s g} [ \hat {z} _ {t} ] \right| \right| _ {2} ^ {2}} _ {\mathcal {L} _ {\text {c o m m i t}}} \right] \tag {2} +$$ + +where $z_{t} = \mathrm{FFN}_{enc}(m)$ represents the encoded latent at each time step $t$ , $\hat{z}_t$ is the quantized latent, and $\mathrm{sg}$ denotes the stop-gradient operation. The two loss components, $\mathcal{L}_{\mathrm{recon}}$ and $\mathcal{L}_{\mathrm{commit}}$ , are defined as follows: + +- Reconstruction Loss $(\mathcal{L}_{\mathrm{recon}})$ : Minimizes the difference between the original motion latent $m$ and the decoded quantized latent. +- Commitment Loss $(\mathcal{L}_{\mathrm{commit}})$ : Encourages $z_{t}$ to approach the quantized latent $\hat{z}_t$ , ensuring stability in quantization. + +This approach enables FMLG to learn robust, temporally-consistent representations, translating audio-driven inputs into lifelike facial animations. Experimentally, to achieve trade-offs between frame count and redundancy, we selected 4 frames $(4\times 25\times 3$ latent) to be compressed into 32 tokens (hyper-parameter selection trade-off refer to Fig. 11). + +# 3.3. AR Transformer for Motion Generation + +Using the learned RVQ-based latents, motion is denoted as: + +$$ +M = \left[ m _ {1}, m _ {2}, \dots , m _ {T} \right], \quad \text {w h e r e} \quad m _ {i} \in \mathbb {R} ^ {2 5 \times 3}. \tag {3} +$$ + +This sequence of motion latents is converted to discrete tokens using the RVQ module: + +$$ +T _ {m} = \left[ t _ {1}, t _ {2}, \dots , t _ {T / 4} \right], \tag {4} +$$ + +where $T_{m}$ represents the quantized motion tokens. Audio input is encoded with the Whisper encoder [5], generating the audio condition $c$ . Motion generation is modeled as a next-token prediction task, where the distribution of each token is predicted based on the previous $t - 1$ tokens and + +![](images/5e4f45325368af28fa6451f3586e5715c2cc81f263c5b155e484f0dd6e751915.jpg) +Figure 3. In our Teller, we follow AR transformer architecture, but each input consists of a pair of tokens and model pred a pair of tokens for each output position. + +audio condition $c$ : + +$$ +P \left(t _ {i} \mid c, t _ {< i}\right). \tag {5} +$$ + +This autoregressive setup enables sequential generation of motion based on past motion tokens and audio embeddings. + +To enable real-time streaming animation, we process both audio and video frames in 200ms chunks, following Whisper's constraints. Each audio chunk is encoded into a $[10 \times 512]$ embedding, while each video chunk uses 32 learned RVQ-based motion tokens. For efficient real-time performance, we enable the autoregressive transformer to process token pairs at each position, which improves prediction speed by processing two tokens concurrently. In Teller, each input consists of a token pair with a combined embedding and a learnable position bias inspired by BERT, capturing relative token positions. The loss for each head, representing each token in the pair, is computed as: + +$$ +\mathcal {L} _ {\text {h e a d} 0 _ {j}} = \mathrm {C E} \left(\operatorname {l a b e l} _ {j} [ 0 ], \mathrm {T} \left(\operatorname {i n p u t} _ {j} [ 0 ] \mid \operatorname {i n p u t} _ {< j}\right)\right), \tag {6} +$$ + +$$ +\mathcal {L} _ {\text {h e a d} 1 _ {j}} = \mathrm {C E} \left(\operatorname {l a b e l} _ {j} [ 1 ], \mathrm {T} \left(\operatorname {i n p u t} _ {j} [ 1 ] \mid \operatorname {i n p u t} _ {< j}\right)\right), \tag {7} +$$ + +where $T$ denotes the transformer. The total loss, with a regularization term to balance learning across both heads, is: + +$$ +\mathcal {L} _ {a r} = \sum_ {j = 1} ^ {I / 2} \left[ \mathcal {L} _ {\text {h e a d} 0 _ {j}} + \mathcal {L} _ {\text {h e a d} 1 _ {j}} + \left\| \mathcal {L} _ {\text {h e a d} 0 _ {j}} - \mathcal {L} _ {\text {h e a d} 1 _ {j}} \right\| _ {2} ^ {2} \right]. \tag {8} +$$ + +This regularization term, $\| \mathcal{L}_{\mathrm{head0}} - \mathcal{L}_{\mathrm{head1}}\| _2^2$ ensures balanced training across the two heads, promoting stable and accurate real-time animation. + +After testing frame count and redundancy, we use 4 frames $(4\times 25\times 3$ latent), compressed into 32 tokens (refer to Fig. 11). Then the 4 frames are interpolated to 5 frames for faster generation. This setup balances inference efficiency and feature quality, enabling real-time, high-fidelity streaming portrait animation. + +![](images/1c89d4f74316bc17d94cfb7d61cff7e810b87dbfb59e777ffccdb4873933c716.jpg) +Figure 4. Qualitative comparison with existing approaches on RAVDESS data-set of 'angry' and 'disgust' emotion cases. Videos are available in the supplementary materials. + +![](images/01412e46931605ddfa0c8ad9afffc0862ba25a913c94f8f90be4e0c0f92a4d43.jpg) + +# 3.4. Efficient Temporal Module for Refinement + +In diffusion-based models, temporal layers are added to text-to-image (T2I) frameworks to capture frame dependencies [16]. Inspired by this, Teller incorporates temporal refinement but achieves it in a single step, unlike diffusion models that require multiple iterations, enhancing real-time efficiency. After encoding video frames with a VAE encoder, we apply a 3D U-Net [7] to extract features from the image sequence, represented as $x \in \mathbb{R}^{b \times t \times h \times w \times c}$ , where $b$ is the batch size, $t$ the frame count, $h$ and $w$ the frame dimensions, and $c$ the channel count. The features are reshaped to $x \in \mathbb{R}^{(b \times h \times w) \times t \times c}$ , enabling ETM to perform self-attention along the temporal dimension $t$ . ETM's output is then merged with the original features through residual connections, integrating temporal dependencies into spatial features. + +For training, we use the first 5 frames from a real image sequence and the subsequent 5 frames reconstructed with LivePortrait [12]. After processing through ETM, we compute reconstruction loss between the predicted and ground-truth frames: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \sum_ {i = 6} ^ {1 0} \left\| x _ {\mathrm {g t} _ {i}} - f \left(x _ {i} \mid x _ {\mathrm {g t} _ {< 6}}\right) \right\| _ {2} ^ {2}, \tag {9} +$$ + +where $x_{\mathrm{gt}_i}$ denotes the real feature sequence and $x_{i}$ the reconstructed feature sequence (from LivePortrait or the Stage 1 decoder). ETM primarily learns to preserve consistency in physical features such as neck muscles and earrings. To this end, we add a region-specific mask to recon + +struction loss, and the final loss of ETM: + +$$ +\mathcal {L} _ {\mathrm {E T M}} = \sum_ {i = 6} ^ {1 0} \left\| x _ {\mathrm {g t} _ {i}} \odot \operatorname {m a s k} _ {i} - f \left(x _ {i} \mid x _ {\mathrm {g t} _ {< 6}}\right) \odot \operatorname {m a s k} _ {i} \right\| _ {2} ^ {2}, \tag {10} +$$ + +where element-wise multiplication with $\mathrm{mask}_i$ focuses reconstruction on specific regions. The mask is defined as: + +$$ +\operatorname {m a s k} (i, j) = \left\{ \begin{array}{l l} 1, & \text {i f} (i, j) \text {i s w i t h i n} \mathrm {B B} (x) \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {11} +$$ + +where $\mathrm{BB(x)}$ stands for the bounding box and outlines relevant body parts. Key landmarks are identified using MediaPipe[22], with points indices such as [93, 323, 152] defining the bounding boxes for these regions, highlighted in red in the stage 2 input of Fig. 2. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. For training, we used the AV Speech datasets (filtered to 662 hours) [10] and VFHQ [39] datasets (filtered to 2 hours) for pretraining, along with additional talking-head videos from the internet (32 hours) for supervised finetuning (SFT). For validation, we used the HDTF (filtered to 0.83 hours) [47] and RAVDESS (filtered to 0.55 hours) [19] datasets and supplementary internet data (0.49 hours) for qualitative comparisons and human evaluation only. To ensure quality, we applied the Mediapipe [22] face detection tool to filter out instances with facial movement exceeding + +![](images/9a7cd6f75f4f61c7a33a454b3bdec767de096ec7dc48f38d0abe538c4d80ab9b.jpg) +Figure 5. Qualitative comparison with existing approaches on HDTF dataset. Videos are available in the supplement. mat. + +$50\%$ . We further refined the data using Sync-C and Sync-D to exclude samples with low lip-sync scores. + +Metrics. Evaluation metrics include Fréchet Inception Distance (FID)[15], Fréchet Video Distance (FVD)[33], Synchronization-C (Sync-C)[24], and Synchronization-D (Sync-D)[24]. FID and FVD assess realism, with lower scores indicating better quality, while Sync-C and Sync-D measure lip synchronization, with higher Sync-C and lower Sync-D values indicating better alignment. + +# Implementation Details. + +Stage 1: During pretraining, we follow the architecture design of the Qwen1.5-4B model [2] and initialize the parameters randomly. The model is trained on an $8 \times 8$ Nvidia A800 GPU machine with a batch size of 1024, using the AdamW optimizer [20]. We employ a cosine learning rate scheduler, with the learning rate decaying from 1e-4 to 1e-6 over 40 epochs. In the supervised fine-tuning (SFT) phase, we again use an $8 \times 8$ Nvidia A800 GPU machine, but with a batch size of 512. The AdamW optimizer [20] is used, along with a cosine learning rate scheduler. The learning rate decaying from 1e-5 to 1e-6 over 10 epochs. Stage 2: The model is trained on an $8 \times 8$ Nvidia A800 GPU machine, with a batch size of 1024, and using the AdamW optimizer. The cosine learning rate scheduler is used, with the learning rate decaying from 1e-4 to 1e-6 over 30 epochs. + +Real time analysis. The model is inference on an 4 Nvidia H800 GPU machine. For a 200ms audio input, the average processing time of the Whisper encoder is 7ms. In Stage 1, the average total time is 106ms, with the AR transformer taking an average of 6ms per 16 tokens, and the motion decoder taking an average of 10ms. In Stage 2, the average + +
MethodFID↓FVD↓Sync-C↑Sync-D↓Time
SadTalker [46]22.177233.6737.3267.84818.89s
EchoMimic [6]23.049290.1906.6648.83931.10s
AniPortrait [37]28.161235.0994.54710.65729.36s
Hallo [40]20.639174.1917.4977.74120.93s
Teller (ours)21.352173.4637.6967.5360.92s
Real video--8.0946.976-
+ +Table 1. Quantitative comparison with existing portrait image animation approaches on the HDTF dataset. Time stands for the averaging time cost of generating one second of 25 fps video. + +
MethodFID↓FVD↓Sync-C↑Sync-D↓
SadTalker [46]32.343487.9244.3047.621
EchoMimic [6]21.058668.6753.2929.096
AniPortrait [37]30.696476.1972.32111.542
Hallo [40]19.826537.4784.0628.552
Teller (ours)20.352429.2884.4967.936
Real video--5.2237.069
+ +Table 2. Quantitative comparison with existing portrait image animation approaches on the RAVDESS dataset. + +total time is 71ms, with the VAE encoder and decoder averaging 25ms, and the Temporal Module averaging 21ms. + +# 4.2. Quantitative Results + +Comparison on the HDTF Dataset. Table 1 presents quantitative results for portrait animation techniques on the HDTF dataset. Our method outperforms others, achieving the lowest FVD of 173.463 and a competitive FID of 21.352, indicating high quality and temporal coherence in animated talking heads. Additionally, it achieves the highest Sync-C score of 7.696 and the lowest Sync-D score of 7.536, demonstrating excellent lip synchronization. These results highlight Teller's effectiveness in maintaining both visual fidelity and synchronization. + +Comparison on the RAVDESS Dataset. Table 2 shows quantitative results on the RAVDESS dataset. Our method again leads in performance, with the lowest FVD of 429.288 and a competitive FID of 20.352, reflecting high-quality and temporally coherent animations. The highest Sync-C score of 4.496 and a competitive Sync-D score of 7.936 further demonstrate superior lip synchronization. These findings confirm Teller's strength in producing synchronized and high-fidelity animated portraits. + +# 4.3. Qualitative Results + +Head Movement Comparison. Figure 5 shows a qualitative comparison of head movements. Teller replicates natural head movements more accurately, closely matching the ground truth (GT) with smooth, realistic turns and subtle expression-based adjustments. Competing methods, like AntiPortrait and EchoMimic, often show abrupt or limited movements, appearing rigid or lifeless. Teller's autoregressive framework ensures continuity and natural dynamics in + +![](images/38ff0996b8342e7d6677a1996802c3cf06c579a35c7343744c815a6caca4bd45.jpg) + +![](images/b14dfdea6bea655b88263978b2fb7bd3c7acfc1849538659b5a6c3f73d3a51ca.jpg) +Figure 6. Top-k selection $(\mathrm{k} = 15)$ in FMLG produces diverse facial expressions and actions with accurate lip sync on the HDTF. + +![](images/e2f68803cb488bf6c9d4b5ab468a9914f9eb29c26ebace7d39fcd056d8b018de.jpg) +Figure 7. Visualization of finer motion details. Videos are available in the supplementary materials. +Figure 8. Visualization of the generation accuracy of lip shape. + +head and emotion alignment, essential for lifelike animation. + +Diversity. Fig. 6 shows two sequences generated with Top-k sampling (k=15) in FMLG, where each row represents frames from the same speech input. The model demonstrates diverse facial expressions and head movements while maintaining accurate lip sync, highlighting Top-k sampling's role in enhancing motion variety without compromising synchronization. + +![](images/8bd03e69009f82e3ad6bb1c37ba5126f9aab335c8191e1d27e6d69cc3cfe6b21.jpg) +Figure 9. Human evaluation results among our proposed Teller and other SoTA methods. + +![](images/745b71d2a611d16d33f43cc260fd628402b8dccefcf788252b634f089b243b59.jpg) + +![](images/728a8e415ca6cb26841caba08b410ae33c8756363af3c311c560ba7968c5f814.jpg) + +Emotional Expression. Fig. 4 presents our model has more accurate emotional expression ability due to the better speech understanding ability of AR transformer. + +Finer Motion Details. Fig. 7 highlights fine motion details in neck and earring movements. Compared to other methods, Teller produces realistic, nuanced motions synchronized with speech, capturing subtle audio-driven dynamics that contribute to lifelike and temporally coherent animation. The consistent detail across frames underscores Teller's robustness. + +Lip Synchronization Accuracy. Fig. 8 shows superior lip synchronization, with Teller aligning generated lip shapes closely to natural movements. This high fidelity highlights Teller's strength in producing realistic, synchronized mouth motions for convincing talking head animation. + +# 4.4. Human Evaluation + +We conducted a human evaluation to assess the quality of generated animations, focusing on lip synchronization, body movement realism, and temporal coherence. Thirty participants (66.7% aged 24-30, 33.3% aged 30-40; 30% male, 70% female; 83.3% with AIGC model experience) rated each animation on a 5-point Likert scale for coherence with input and animation quality. A total of 100 videos were presented in random order to avoid bias, providing insights into subjective perceptions of animation quality and natural expression alignment. As shown in Figure 9, participants rated Teller highest in lip synchronization, body movement realism, and temporal coherence, with low variance in scores, indicating robust and consistent performance. + +# 5. Ablation Study + +Stage (Module) Ablation. We perform an ablation study comparing Stage 1 and Stage 2, focusing on animation quality in body parts and accessories, especially neck muscles and earrings. As shown in Fig. 10, Stage 1 outputs often show less realistic, inconsistent movements. In contrast, Stage 2, with ETM, significantly improves the physical consistency of subtle motions, creating natural earring sway and smooth neck movements for lifelike, temporally coherent animation. + +![](images/3941306c3f53f958fbed40f3dbfbcd0f35264b47aa42ee41c0ac07fa429be58f.jpg) +Figure 10. Visualization of the generation images with different stages and the differences between stages. + +![](images/833776ffb35d18de0fa849bac2b072534a5e4e3ae0702cdcc74605243fb9df99.jpg) +Figure 11. Tradeoff between performance (loss) and different compression(tokens/frame) ratios. + +Frames / Tokens in RVQ Module Figure 11 shows the impact of frame and token configurations in the RVQ module on face shape accuracy in talking head animations. This ablation study explores how varying frames and tokens affects facial movement quality and synchronization. Results show that increasing frames and tokens improves facial dynamics, enhancing lip synchronization and realism, though with higher computational costs. We selected 4 frames and 32 tokens for an optimal balance, adjustable as needed. + +Ablation Study on Audio Condition Encoder. We analyzed the performance differences between TTS and ASR models as audio condition encoders for our task. Specifically, we used Whisper for the ASR model and funcodec for the TTS model. As shown in Table 3, Whisper achieved a Sync-C score of 7.696 and a Sync-D score of 7.536, while funcodec scored 4.286 in Sync-C and 10.373 in Sync-D. These results indicate that our task benefits more from the ASR model's capability to capture the nuances required for precise synchronization, as seen in Whisper's higher Sync-C and lower Sync-D scores. This analysis suggests that the ASR model is better suited as the audio condition encoder, enhancing the overall quality and synchronization of our talking head animation. + +Ablation Study on Single-Head vs. Multi-Head Archi + +
MethodSync-C ↑Sync-D ↓
w funcdec4.28610.373
w whisper7.6967.536
+ +Table 3. Comparison of synchronization for audio conditions using funcodec and Whisper in ASR and TTS tasks on HDTF. + +
MethodFID↓FVD↓Sync-C↑Sync-D↓
Single-Head22.110172.5537.7907.474
Multi-Head21.352173.4637.6967.536
+ +Table 4. Comparison of performance between Single-Head and Multi-Head on the HDTF dataset. + +tecture. We compare single-head and multi-head models on FID, FVD, Sync-C, and Sync-D metrics (Table 4). The single-head model slightly outperforms in FID (22.110 vs. 21.352) and has comparable FVD (172.553 vs. 173.463), with marginally better synchronization (Sync-C of 7.790 vs. 7.696 and Sync-D of 7.474 vs. 7.536). Both architectures yield competitive results, though the single-head model slightly excels in synchronization, while the multi-head model offers greater real-time potential. + +# 6. Conclusion + +In this paper, we presented Teller, the first autoregressive framework designed for real-time, audio-driven portrait animation. Addressing the challenge of realistic and efficient talking head generation, Teller achieves high-quality animations at up to 25 FPS, surpassing existing methods in both fidelity and responsiveness. Extensive experiments demonstrated Teller's advantages over SoTA audio-driven animation methods, particularly in rendering nuanced movements essential for lifelike and visually convincing animations. Human evaluations further validate its quality, particularly in natural expression and lip synchronization. By balancing computational efficiency with high animation fidelity, Teller sets a new standard for real-time talking head animation, marking a significant advancement in multimodal portrait animation frameworks. Additionally, Teller's AR Transformer architecture makes it compatible with existing unified multimodal language models. + +# References + +[1] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564, 2023. 3 +[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 6 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3 +[4] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2 +[5] Nan Cao, Yu-Ru Lin, Xiaohua Sun, David Lazer, Shixia Liu, and Huamin Qu. Whisper: Tracing the spatiotemporal process of information diffusion in real time. IEEE transactions on visualization and computer graphics, 18(12):2649-2658, 2012. 4 +[6] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 6 +[7] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19, pages 424-432. Springer, 2016. 5 +[8] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 3 +[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 2 +[10] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T. Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: a speaker-independent audio-visual model for speech separation. ACM Transactions on Graphics, 37(4):1-11, 2018. 5 +[11] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 3 + +[12] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1, 2, 3, 5 +[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3 +[14] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 3 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[16] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 1, 3, 5 +[17] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 3 +[18] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2 +[19] Steven R Livingstone and Frank A Russo. The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5): e0196391, 2018. 5 +[20] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[21] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (ToG), 40(6):1-17, 2021. 1 +[22] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5 +[23] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 427-436, 2023. 2 +[24] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the + +28th ACM international conference on multimedia, pages 484-492, 2020. 6 +[25] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 2 +[26] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 2 +[27] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 2 +[28] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3 +[29] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2 +[30] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. Advances in Neural Information Processing Systems, 36, 2024. 3 +[31] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models, 2024. URL https://arxiv.org/abs/2405.09818, 2024.3 +[32] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 3 +[33] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6 +[34] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17979-17989, 2023. 2 +[35] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 1, 2 +[36] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3 +[37] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 6 + +[38] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 3 +[39] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 5 +[40] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 3, 6 +[41] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3 +[42] Ziyu Yao, Xuxin Cheng, and Zhiqi Huang. Fd2talk: Towards generalized talking head generation with facial decoupled diffusion model. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3411-3420, 2024. 2 +[43] Hanrong Ye, De-An Huang, Yao Lu, Zhiding Yu, Wei Ping, Andrew Tao, Jan Kautz, Song Han, Dan Xu, Pavlo Molchanov, et al. X-vila: Cross-modality alignment for large language model. arXiv preprint arXiv:2405.19335, 2024. 3 +[44] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2 +[45] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 2 +[46] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 6 +[47] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 5 +[48] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 3 +[49] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking + +face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176-4186, 2021. 2 +[50] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2 +[51] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 3 \ No newline at end of file diff --git a/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/images.zip b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a445b068a994d9413af260c088f960932e506bd8 --- /dev/null +++ b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebc35c3bd245144e406015dc5617deae4b2afbafd7f26a39c27ee92da233f9e7 +size 861671 diff --git a/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/layout.json b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d4870faa0fb51cca20d37b4c0584ca560c3a6721 --- /dev/null +++ b/2025/Teller_ Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation/layout.json @@ -0,0 +1,7974 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 515, + 140 + ], + "type": "text", + "content": "Teller: Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 161, + 459, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 161, + 459, + 175 + ], + "spans": [ + { + "bbox": [ + 168, + 161, + 459, + 175 + ], + "type": "text", + "content": "Dingcheng Zhen Shunshun Yin Shiyang Qin Hou Yi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 176, + 436, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 176, + 436, + 190 + ], + "spans": [ + { + "bbox": [ + 192, + 176, + 436, + 190 + ], + "type": "text", + "content": "Ziwei Zhang Siyuan Liu Gan Qi Ming Tao" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 190, + 405, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 190, + 405, + 204 + ], + "spans": [ + { + "bbox": [ + 223, + 190, + 405, + 204 + ], + "type": "text", + "content": "Shanghai Soulgate Technolgy Co.tl." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 205, + 563, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 563, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 563, + 217 + ], + "type": "text", + "content": "{dingchengzhen, yinshunshun, qinshiyang, houyi,zhangziwei, siyuanliu, ganqi, ming}@soulapp.cn" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 71, + 225, + 541, + 430 + ], + "blocks": [ + { + "bbox": [ + 71, + 225, + 541, + 430 + ], + "lines": [ + { + "bbox": [ + 71, + 225, + 541, + 430 + ], + "spans": [ + { + "bbox": [ + 71, + 225, + 541, + 430 + ], + "type": "image", + "image_path": "64a15499ee037e75efd44f23a583cd72d6c4540f11d8795d571e003bb86d2ca5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 434, + 555, + 457 + ], + "lines": [ + { + "bbox": [ + 55, + 434, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 555, + 457 + ], + "type": "text", + "content": "Figure 1. Teller framework is the first autoregressive framework for real-time, audio-driven portrait animation, achieving up to 25 FPS while preserving realistic body part and accessory movements. Demo can be found at https://teller-ayar.github.io/." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 461, + 200, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 461, + 200, + 473 + ], + "spans": [ + { + "bbox": [ + 151, + 461, + 200, + 473 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "type": "text", + "content": "In this work, we introduce the first autoregressive framework for real-time, audio-driven portrait animation, a.k.a, talking head. Beyond the challenge of lengthy animation times, a critical challenge in realistic talking head generation lies in preserving the natural movement of diverse body parts. To this end, we propose Teller, the first streaming audio-driven portrait animation framework with autoregressive motion generation. Specifically, Teller first decomposes facial and body detail animation into two components: Facial Motion Latent Generation (FMLG) based on an autoregressive transformer, and movement authenticity refinement using a Efficient Temporal Module (ETM). Concretely, FMLG employs a Residual VQ model to map the facial motion latent from the implicit keypoint-based model into discrete motion tokens, which are then temporally sliced with audio embeddings. This enables the AR transformer to learn real-time, stream-based mappings from audio to motion. Furthermore, Teller incorporate ETM to capture finer motion details. This module ensures the physi" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 582 + ], + "type": "text", + "content": "cal consistency of body parts and accessories, such as neck muscles and earrings, improving the realism of these movements. Teller is designed to be efficient, surpassing the inference speed of diffusion-based models (Hallo 20.93s vs. Teller 0.92s for one second video generation), and achieves a real-time streaming performance of up to 25 FPS. Extensive experiments demonstrate that our method outperforms recent audio-driven portrait animation models, especially in small movements, as validated by human evaluations with a significant margin in quality and realism." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Realistic and expressive portrait animation from audio and static images, commonly known as talking head animation [12, 16, 21, 35], has garnered significant interest across applications such as virtual avatars, digital communication, and entertainment. However, generating high-quality animations that are visually compelling and temporally consistent remains a major challenge. This complexity stems from the need to intricately coordinate lip movements, fa" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21075" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "cial expressions, and head positioning to create lifelike effects. Moreover, achieving real-time, realistic talking head animation is complicated by computational constraints and the nuances of human movement, making this an especially demanding task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 326 + ], + "type": "text", + "content": "While recent advancements, such as diffusion models [25, 27, 29, 42], have improved high-quality content generation, achieving controllable animation presents ongoing hurdles. Effective animation requires the accurate capture of complex facial expressions, body gestures, and the subtle interplay between them. Existing methods often suffer from prolonged animation times (See animation time in Tab 1), limiting their potential for real-time applications, and frequently fail to capture the natural, interconnected motions of various facial and body parts, e.g., earrings and necklaces, as shown in Figure. 1 stage 2. This shortfall results in animations with stiff or exaggerated movements that disrupt the realism of the animation (See Figure 7). Addressing these challenges necessitates a solution that balances computational efficiency with high animation quality without overloading processing resources." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 532 + ], + "type": "text", + "content": "To this end, we propose Teller, the first autoregressive framework capable of real-time, streaming-based talking head animation at up to 25 FPS. As shown in Figure. 2, Teller employs a two-stage framework, combining Facial Motion Latent Generation (FMLG) and an Efficient Temporal Module (ETM) to produce realistic and physically consistent animations across facial and body movements. In the first stage, FMLG uses a Residual Vector Quantization (RVQ) model [45] to encode facial motion latents derived from an implicit keypoint-based model into discrete motion tokens. These tokens are then temporally aligned with audio embeddings, allowing an autoregressive (AR) transformer to map audio signals to facial movements in real-time. By breaking down the motion into temporal segments, FMLG enables the AR transformer to dynamically and efficiently generate high-quality animations responsive to live audio inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "To enhance the realism of body movements, Teller introduces ETM in the second stage, which captures subtle details that are often overlooked in existing methods [12, 41]. The ETM refines finer movements, such as body parts and accessories, ensuring physically plausible interactions. For instance, it simulates realistic motions in neck muscles and dynamic accessories like earrings, which are vital for maintaining visual continuity in the animated avatar. Our Teller model prioritizes computational efficiency and is specifically designed to significantly outperform diffusion-based models in terms of inference speed. For example, generating a one-second video takes Hallo 20.93s, while Teller completes the task in just 0.92s (see Tab. 1). Despite this speed, Teller maintains high animation fidelity. By optimizing computational demands without compromising output" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "quality, Teller not only meets but exceeds real-time requirements, ensuring a smooth and responsive user experience." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 97, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 240 + ], + "type": "text", + "content": "In extensive evaluations across various benchmarks and real-world settings, Teller demonstrates significant improvements over current state-of-the-art audio-driven portrait animation models, as shown in Table. 1, particularly in capturing nuanced facial and body movements. Human evaluations validate Teller's superior quality and realism, especially in rendering the subtle movements essential for lifelike animations. Our research represents a notable advancement in real-time talking head animation, presenting an innovative framework that bridges the gap between realism and efficiency in multimodal animation, as depicted in Figure. 6, Figure. 7, and Figure. 8." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 251, + 406, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 406, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 406, + 262 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 270, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 426 + ], + "type": "text", + "content": "Non-Diffusion-Based Audio-Driven Portrait Animation This line of works typically consist of two key components: an audio-to-motion model and a facial motion representation model. These methods often utilize implicit keypoints as an intermediate motion representation, warping the source portrait based on the driving image. The goal of implicit methods is to learn disentangled representations in 2D [4, 18, 23, 34, 44, 49] or 3D [9, 35] latent spaces, with a focus on aspects such as identity, facial dynamics, and head pose. For example, FOMM [26] employs first-order Taylor expansion to capture local motion, while FaceVid2Vid [35] extends this by introducing a 3D implicit keypoints representation, enabling free-view portrait animation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 617 + ], + "type": "text", + "content": "To effectively learn facial motion representations in latent space, these approaches often rely on GAN-based frameworks to disentangle identity-related appearance from non-identity-related motion, specifically capturing expressions, lip and eye movements, minor accessories, and poses. Examples include FaceVid2Vid [35], LivePortrait [12], and others, where identity and motion representations are independently learned to generate realistic talking head animations. For instance, MakeItTalk [50] uses an LSTM-based audio-to-motion model to predict landmark coordinates from audio input, which are then translated into video frames using a warp-based GAN model. Similarly, SadTalker [46] employs FaceVid2Vid [35] as an image synthesizer, with ExpNet and PoseVAE modules transforming audio features into inputs compatible with FaceVid2Vid for audio-to-video generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "However, these approaches face challenges due to GAN losses that primarily focus on facial expressions, lip, and eye movements, often neglecting accessory, hair, and body movements. This can result in animations that appear stiff or incomplete, lacking natural dynamism. In contrast, our work introduces the first autoregressive framework specifically designed for real-time, audio-driven portrait animation, achieving up to 25 FPS and delivering a more realistic," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21076" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 549, + 337 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 549, + 337 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 549, + 337 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 549, + 337 + ], + "type": "image", + "image_path": "716343c816bdff653c8ce0ada9d2afffe62239668bc8ac48b6490435d88363d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 342, + 500, + 354 + ], + "lines": [ + { + "bbox": [ + 109, + 342, + 500, + 354 + ], + "spans": [ + { + "bbox": [ + 109, + 342, + 500, + 354 + ], + "type": "text", + "content": "Figure 2. Overall framework of our proposed Teller for real-time streaming audio-driven portrait animation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 362, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 374 + ], + "type": "text", + "content": "coherent portrayal of both facial and accessory movements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": "Diffusion-based Audio-driven Portrait Animation Recent advancements in diffusion-based video generation have demonstrated promising outcomes for audio-driven portrait animation. Methods like GAIA [14] and VASA-1 [41] have designed diffusion models to transform audio inputs into motion latents, facilitating audio-to-video generation. Further developments, such as EMO[32], Hallo [40], LOOPY [17], enhance end-to-end diffusion modeling by incorporating motion modules [3, 13, 16] and audio cross-attention mechanisms, improving the coherence and synchronization between audio cues and visual motion. Despite these improvements, a significant limitation of diffusion-based models remains their multi-step inference process, required to generate even a single frame or a few frames of video. This step-by-step prediction approach renders real-time performance challenging, as it is computationally intensive and time-consuming, making diffusion-based models less suitable for applications demanding instantaneous response." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "AR Transformer-Based Generation Recent works [1, 30, 36, 38, 43, 48] have focused on developing unified multimodal language models for generating visual content, such as images and videos. Some studies [28, 51] use autoregressive modeling with continuous representations interleaved with text tokens for image generation. Others, like SEED-X [11], propose a foundational system combining CLIP ViT-based image representations with text tokens for multimodal tasks, including next-token prediction and image rep" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 362, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 411 + ], + "type": "text", + "content": "resentation regression. DreamLLM [8] also explores multimodal understanding and creation, while Chameleon [31] introduces token-based models capable of both understanding and generating images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 419, + 370, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 370, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 370, + 431 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 439, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 548 + ], + "type": "text", + "content": "As shown in Fig. 2, Teller comprises two main modules: the Facial Motion Latent Generation (FMLG) and the Efficient Temporal Module (ETM). FMLG module integrates an autoregressive transformer and a residual vector quantization (RVQ) component. It applies an autoregressive transformer to generate discrete facial motion tokens from audio input. Following FMLG, ETM refines the generated motion to produce realistic body and accessory movements, ensuring physical consistency in animated results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 552, + 400, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 400, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 400, + 564 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 554, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 630 + ], + "type": "text", + "content": "Prior works, such as LivePortrait [12] have introduced methods for extracting implicit keypoints as facial motion latents using motion and appearance extractors. These motion latents capture essential facial dynamics needed for animating input images and consist of three main components:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 631, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "text", + "content": "- Expression Deformation " + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "text", + "content": ": A set deformation of 21 implicit keypoints, represented as " + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\delta = [\\delta_1, \\delta_2, \\dots, \\delta_{21}]" + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "text", + "content": " is a 3D vector " + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^3)" + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "text", + "content": " that indicates the position of the " + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 314, + 631, + 553, + 677 + ], + "type": "text", + "content": "-th deformation of facial keypoints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "text", + "content": "- Head Pose " + }, + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "text", + "content": ": Defined by three rotation vectors " + }, + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "R = [r_1, r_2, r_3]" + }, + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "text", + "content": ", with each " + }, + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "text", + "content": " being a 3D vector " + }, + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^3)" + }, + { + "bbox": [ + 314, + 677, + 553, + 713 + ], + "type": "text", + "content": " that describes the head's orientation in 3D space." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21077" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 95 + ], + "type": "text", + "content": "- Expression Deformation " + }, + { + "bbox": [ + 55, + 72, + 296, + 95 + ], + "type": "inline_equation", + "content": "(t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 95 + ], + "type": "text", + "content": ": A single 3D vector " + }, + { + "bbox": [ + 55, + 72, + 296, + 95 + ], + "type": "inline_equation", + "content": "(t \\in \\mathbb{R}^3)" + }, + { + "bbox": [ + 55, + 72, + 296, + 95 + ], + "type": "text", + "content": " that captures facial expression deformations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 295, + 120 + ], + "type": "text", + "content": "These components are concatenated into a unified motion. \nlatent of size " + }, + { + "bbox": [ + 55, + 97, + 295, + 120 + ], + "type": "inline_equation", + "content": "25\\times 3" + }, + { + "bbox": [ + 55, + 97, + 295, + 120 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 124, + 295, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 124, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 295, + 137 + ], + "type": "interline_equation", + "content": "m = \\left[ \\delta_ {1}, \\delta_ {2}, \\dots , \\delta_ {2 1}, r _ {1}, r _ {2}, r _ {3}, t \\right], \\tag {1}", + "image_path": "a1a773deafd1f98cc2c36162abcb0762baafedae53ff93ef832a805e2f3bcfb5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "inline_equation", + "content": "m \\in \\mathbb{R}^{25 \\times 3}" + }, + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "text", + "content": " includes the 21 keypoints, head pose, and expression deformation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 172, + 276, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 276, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 276, + 185 + ], + "type": "text", + "content": "3.2. Facial Motion Latent Generation (FMLG)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "text", + "content": "In FMLG, Teller generates motion latent from facial motion extraction, then encodes it into discrete tokens using a residual vector quantizer (RVQ). To optimize encoding, the concatenated motion latent " + }, + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "text", + "content": " is processed with RVQ, leveraging temporal redundancy across " + }, + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "text", + "content": " frames for efficient compression. The RVQ quantization loss, which encodes the continuous " + }, + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "text", + "content": " into discrete tokens, is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 293, + 302, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 302, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 302, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v q} = \\sum_ {t = 1} ^ {T} \\left[ \\underbrace {\\left| \\left| m - \\operatorname {F F N} _ {d e c} \\left(z _ {t} + \\operatorname {s g} [ \\hat {z} _ {t} - z _ {t} ]\\right) \\right| \\right| _ {2} ^ {2}} _ {\\mathcal {L} _ {\\text {r e c o n}}} + \\underbrace {\\left| \\left| z _ {t} - \\operatorname {s g} [ \\hat {z} _ {t} ] \\right| \\right| _ {2} ^ {2}} _ {\\mathcal {L} _ {\\text {c o m m i t}}} \\right] \\tag {2}", + "image_path": "9731853e61cbbf66067e06f19e5746afc69cabd723c1f891288d6c4b9f59e3ca.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "inline_equation", + "content": "z_{t} = \\mathrm{FFN}_{enc}(m)" + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "content": " represents the encoded latent at each time step " + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "content": " is the quantized latent, and " + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}" + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "content": " denotes the stop-gradient operation. The two loss components, " + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{commit}}" + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "content": ", are defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 395, + 295, + 456 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 395, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 295, + 430 + ], + "type": "text", + "content": "- Reconstruction Loss " + }, + { + "bbox": [ + 55, + 395, + 295, + 430 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{recon}})" + }, + { + "bbox": [ + 55, + 395, + 295, + 430 + ], + "type": "text", + "content": ": Minimizes the difference between the original motion latent " + }, + { + "bbox": [ + 55, + 395, + 295, + 430 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 395, + 295, + 430 + ], + "type": "text", + "content": " and the decoded quantized latent." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 430, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 295, + 456 + ], + "type": "text", + "content": "- Commitment Loss " + }, + { + "bbox": [ + 56, + 430, + 295, + 456 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{commit}})" + }, + { + "bbox": [ + 56, + 430, + 295, + 456 + ], + "type": "text", + "content": ": Encourages " + }, + { + "bbox": [ + 56, + 430, + 295, + 456 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 56, + 430, + 295, + 456 + ], + "type": "text", + "content": " to approach the quantized latent " + }, + { + "bbox": [ + 56, + 430, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 56, + 430, + 295, + 456 + ], + "type": "text", + "content": ", ensuring stability in quantization." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 458, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 543 + ], + "type": "text", + "content": "This approach enables FMLG to learn robust, temporally-consistent representations, translating audio-driven inputs into lifelike facial animations. Experimentally, to achieve trade-offs between frame count and redundancy, we selected 4 frames " + }, + { + "bbox": [ + 55, + 458, + 296, + 543 + ], + "type": "inline_equation", + "content": "(4\\times 25\\times 3" + }, + { + "bbox": [ + 55, + 458, + 296, + 543 + ], + "type": "text", + "content": " latent) to be compressed into 32 tokens (hyper-parameter selection trade-off refer to Fig. 11)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 552, + 265, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 265, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 265, + 563 + ], + "type": "text", + "content": "3.3. AR Transformer for Motion Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 570, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 582 + ], + "type": "text", + "content": "Using the learned RVQ-based latents, motion is denoted as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 68, + 595, + 295, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 595, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 68, + 595, + 295, + 609 + ], + "type": "interline_equation", + "content": "M = \\left[ m _ {1}, m _ {2}, \\dots , m _ {T} \\right], \\quad \\text {w h e r e} \\quad m _ {i} \\in \\mathbb {R} ^ {2 5 \\times 3}. \\tag {3}", + "image_path": "3feca9754670d4203fdecf7b0dcb9d9a0a75aa49a6df6e4d578aa086ebe48e3c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 612, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 635 + ], + "type": "text", + "content": "This sequence of motion latents is converted to discrete tokens using the RVQ module:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 638, + 295, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 638, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 125, + 638, + 295, + 652 + ], + "type": "interline_equation", + "content": "T _ {m} = \\left[ t _ {1}, t _ {2}, \\dots , t _ {T / 4} \\right], \\tag {4}", + "image_path": "9e34c1f194e8ff25e0085b5804bda744deda5cca00e03b2b3199e9f691521f1b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "T_{m}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " represents the quantized motion tokens. Audio input is encoded with the Whisper encoder [5], generating the audio condition " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ". Motion generation is modeled as a next-token prediction task, where the distribution of each token is predicted based on the previous " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " tokens and" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 319, + 74, + 548, + 225 + ], + "blocks": [ + { + "bbox": [ + 319, + 74, + 548, + 225 + ], + "lines": [ + { + "bbox": [ + 319, + 74, + 548, + 225 + ], + "spans": [ + { + "bbox": [ + 319, + 74, + 548, + 225 + ], + "type": "image", + "image_path": "5e4f45325368af28fa6451f3586e5715c2cc81f263c5b155e484f0dd6e751915.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 238, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 271 + ], + "type": "text", + "content": "Figure 3. In our Teller, we follow AR transformer architecture, but each input consists of a pair of tokens and model pred a pair of tokens for each output position." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 277, + 389, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 277, + 389, + 288 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 389, + 288 + ], + "type": "text", + "content": "audio condition " + }, + { + "bbox": [ + 314, + 277, + 389, + 288 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 314, + 277, + 389, + 288 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 405, + 291, + 553, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 291, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 405, + 291, + 553, + 304 + ], + "type": "interline_equation", + "content": "P \\left(t _ {i} \\mid c, t _ {< i}\\right). \\tag {5}", + "image_path": "57fd167df8cb124eb3b37f370b9189158ec9b69f2f05e14f4f6c54d5f6e1e6ea.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 307, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 331 + ], + "type": "text", + "content": "This autoregressive setup enables sequential generation of motion based on past motion tokens and audio embeddings." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 331, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 475 + ], + "type": "text", + "content": "To enable real-time streaming animation, we process both audio and video frames in 200ms chunks, following Whisper's constraints. Each audio chunk is encoded into a " + }, + { + "bbox": [ + 313, + 331, + 554, + 475 + ], + "type": "inline_equation", + "content": "[10 \\times 512]" + }, + { + "bbox": [ + 313, + 331, + 554, + 475 + ], + "type": "text", + "content": " embedding, while each video chunk uses 32 learned RVQ-based motion tokens. For efficient real-time performance, we enable the autoregressive transformer to process token pairs at each position, which improves prediction speed by processing two tokens concurrently. In Teller, each input consists of a token pair with a combined embedding and a learnable position bias inspired by BERT, capturing relative token positions. The loss for each head, representing each token in the pair, is computed as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 339, + 476, + 553, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 476, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 339, + 476, + 553, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {h e a d} 0 _ {j}} = \\mathrm {C E} \\left(\\operatorname {l a b e l} _ {j} [ 0 ], \\mathrm {T} \\left(\\operatorname {i n p u t} _ {j} [ 0 ] \\mid \\operatorname {i n p u t} _ {< j}\\right)\\right), \\tag {6}", + "image_path": "d00f8fa457b35755a779941a3cccaa60daef413f43fed979e42f12ff7422c8c9.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 339, + 505, + 553, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 505, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 339, + 505, + 553, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {h e a d} 1 _ {j}} = \\mathrm {C E} \\left(\\operatorname {l a b e l} _ {j} [ 1 ], \\mathrm {T} \\left(\\operatorname {i n p u t} _ {j} [ 1 ] \\mid \\operatorname {i n p u t} _ {< j}\\right)\\right), \\tag {7}", + "image_path": "fef26efc5a59e0c9b7d343146b5cdebfd016a935b2a520680e24c37a61c03851.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 524, + 553, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 553, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 524, + 553, + 548 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 524, + 553, + 548 + ], + "type": "text", + "content": " denotes the transformer. The total loss, with a regularization term to balance learning across both heads, is:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 560, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 560, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 321, + 560, + 553, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a r} = \\sum_ {j = 1} ^ {I / 2} \\left[ \\mathcal {L} _ {\\text {h e a d} 0 _ {j}} + \\mathcal {L} _ {\\text {h e a d} 1 _ {j}} + \\left\\| \\mathcal {L} _ {\\text {h e a d} 0 _ {j}} - \\mathcal {L} _ {\\text {h e a d} 1 _ {j}} \\right\\| _ {2} ^ {2} \\right]. \\tag {8}", + "image_path": "d5d4465ac49caf7deb06b81788514569596523a957ab3052d4a905f4681c8a94.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "content": "This regularization term, " + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\| \\mathcal{L}_{\\mathrm{head0}} - \\mathcal{L}_{\\mathrm{head1}}\\| _2^2" + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "content": " ensures balanced training across the two heads, promoting stable and accurate real-time animation." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "After testing frame count and redundancy, we use 4 frames " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "(4\\times 25\\times 3" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " latent), compressed into 32 tokens (refer to Fig. 11). Then the 4 frames are interpolated to 5 frames for faster generation. This setup balances inference efficiency and feature quality, enabling real-time, high-fidelity streaming portrait animation." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21078" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 312, + 320 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 312, + 320 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 312, + 320 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 312, + 320 + ], + "type": "image", + "image_path": "1c89d4f74316bc17d94cfb7d61cff7e810b87dbfb59e777ffccdb4873933c716.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 323, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 323, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 555, + 346 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with existing approaches on RAVDESS data-set of 'angry' and 'disgust' emotion cases. Videos are available in the supplementary materials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 553, + 320 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 553, + 320 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 553, + 320 + ], + "type": "image", + "image_path": "01412e46931605ddfa0c8ad9afffc0862ba25a913c94f8f90be4e0c0f92a4d43.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 353, + 278, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 278, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 278, + 366 + ], + "type": "text", + "content": "3.4. Efficient Temporal Module for Refinement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "content": "In diffusion-based models, temporal layers are added to text-to-image (T2I) frameworks to capture frame dependencies [16]. Inspired by this, Teller incorporates temporal refinement but achieves it in a single step, unlike diffusion models that require multiple iterations, enhancing real-time efficiency. After encoding video frames with a VAE encoder, we apply a 3D U-Net [7] to extract features from the image sequence, represented as " + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{b \\times t \\times h \\times w \\times c}" + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "content": " the frame count, " + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "content": " the frame dimensions, and " + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "content": " the channel count. The features are reshaped to " + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{(b \\times h \\times w) \\times t \\times c}" + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "content": ", enabling ETM to perform self-attention along the temporal dimension " + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 375, + 295, + 555 + ], + "type": "text", + "content": ". ETM's output is then merged with the original features through residual connections, integrating temporal dependencies into spatial features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 560, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 619 + ], + "type": "text", + "content": "For training, we use the first 5 frames from a real image sequence and the subsequent 5 frames reconstructed with LivePortrait [12]. After processing through ETM, we compute reconstruction loss between the predicted and ground-truth frames:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 620, + 295, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 620, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 100, + 620, + 295, + 651 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\sum_ {i = 6} ^ {1 0} \\left\\| x _ {\\mathrm {g t} _ {i}} - f \\left(x _ {i} \\mid x _ {\\mathrm {g t} _ {< 6}}\\right) \\right\\| _ {2} ^ {2}, \\tag {9}", + "image_path": "f2e984f546ab41f6af576b84474d90b1f7e95e10992a54313c4ebc2eabc39825.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{gt}_i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " denotes the real feature sequence and " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " the reconstructed feature sequence (from LivePortrait or the Stage 1 decoder). ETM primarily learns to preserve consistency in physical features such as neck muscles and earrings. To this end, we add a region-specific mask to recon" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 353, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 353, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 314, + 353, + 480, + 365 + ], + "type": "text", + "content": "struction loss, and the final loss of ETM:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 323, + 373, + 553, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 373, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 323, + 373, + 553, + 415 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E T M}} = \\sum_ {i = 6} ^ {1 0} \\left\\| x _ {\\mathrm {g t} _ {i}} \\odot \\operatorname {m a s k} _ {i} - f \\left(x _ {i} \\mid x _ {\\mathrm {g t} _ {< 6}}\\right) \\odot \\operatorname {m a s k} _ {i} \\right\\| _ {2} ^ {2}, \\tag {10}", + "image_path": "2098975a3747e8221598c50da0ab986dce3c03f6a3ef28fa22202b8efddc0c81.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 416, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 553, + 441 + ], + "type": "text", + "content": "where element-wise multiplication with " + }, + { + "bbox": [ + 313, + 416, + 553, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{mask}_i" + }, + { + "bbox": [ + 313, + 416, + 553, + 441 + ], + "type": "text", + "content": " focuses reconstruction on specific regions. The mask is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 335, + 449, + 553, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 449, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 335, + 449, + 553, + 480 + ], + "type": "interline_equation", + "content": "\\operatorname {m a s k} (i, j) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} (i, j) \\text {i s w i t h i n} \\mathrm {B B} (x) \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {11}", + "image_path": "a91e99db2883ad4816eddcca1860855ce2ff203de394299dc3286a24b46b43ac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 555, + 546 + ], + "type": "inline_equation", + "content": "\\mathrm{BB(x)}" + }, + { + "bbox": [ + 313, + 486, + 555, + 546 + ], + "type": "text", + "content": " stands for the bounding box and outlines relevant body parts. Key landmarks are identified using MediaPipe[22], with points indices such as [93, 323, 152] defining the bounding boxes for these regions, highlighted in red in the stage 2 input of Fig. 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 556, + 395, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 395, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 395, + 570 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 576, + 430, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 430, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 430, + 590 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Datasets. For training, we used the AV Speech datasets (filtered to 662 hours) [10] and VFHQ [39] datasets (filtered to 2 hours) for pretraining, along with additional talking-head videos from the internet (32 hours) for supervised finetuning (SFT). For validation, we used the HDTF (filtered to 0.83 hours) [47] and RAVDESS (filtered to 0.55 hours) [19] datasets and supplementary internet data (0.49 hours) for qualitative comparisons and human evaluation only. To ensure quality, we applied the Mediapipe [22] face detection tool to filter out instances with facial movement exceeding" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21079" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 76, + 294, + 300 + ], + "blocks": [ + { + "bbox": [ + 58, + 76, + 294, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 294, + 300 + ], + "type": "image", + "image_path": "9a7cd6f75f4f61c7a33a454b3bdec767de096ec7dc48f38d0abe538c4d80ab9b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 304, + 295, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 327 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison with existing approaches on HDTF dataset. Videos are available in the supplement. mat." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "type": "text", + "content": ". We further refined the data using Sync-C and Sync-D to exclude samples with low lip-sync scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 445 + ], + "type": "text", + "content": "Metrics. Evaluation metrics include Fréchet Inception Distance (FID)[15], Fréchet Video Distance (FVD)[33], Synchronization-C (Sync-C)[24], and Synchronization-D (Sync-D)[24]. FID and FVD assess realism, with lower scores indicating better quality, while Sync-C and Sync-D measure lip synchronization, with higher Sync-C and lower Sync-D values indicating better alignment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 162, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 162, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 162, + 459 + ], + "type": "text", + "content": "Implementation Details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "text", + "content": "Stage 1: During pretraining, we follow the architecture design of the Qwen1.5-4B model [2] and initialize the parameters randomly. The model is trained on an " + }, + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "text", + "content": " Nvidia A800 GPU machine with a batch size of 1024, using the AdamW optimizer [20]. We employ a cosine learning rate scheduler, with the learning rate decaying from 1e-4 to 1e-6 over 40 epochs. In the supervised fine-tuning (SFT) phase, we again use an " + }, + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "text", + "content": " Nvidia A800 GPU machine, but with a batch size of 512. The AdamW optimizer [20] is used, along with a cosine learning rate scheduler. The learning rate decaying from 1e-5 to 1e-6 over 10 epochs. Stage 2: The model is trained on an " + }, + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "text", + "content": " Nvidia A800 GPU machine, with a batch size of 1024, and using the AdamW optimizer. The cosine learning rate scheduler is used, with the learning rate decaying from 1e-4 to 1e-6 over 30 epochs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Real time analysis. The model is inference on an 4 Nvidia H800 GPU machine. For a 200ms audio input, the average processing time of the Whisper encoder is 7ms. In Stage 1, the average total time is 106ms, with the AR transformer taking an average of 6ms per 16 tokens, and the motion decoder taking an average of 10ms. In Stage 2, the average" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 150 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 150 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 150 + ], + "type": "table", + "html": "
MethodFID↓FVD↓Sync-C↑Sync-D↓Time
SadTalker [46]22.177233.6737.3267.84818.89s
EchoMimic [6]23.049290.1906.6648.83931.10s
AniPortrait [37]28.161235.0994.54710.65729.36s
Hallo [40]20.639174.1917.4977.74120.93s
Teller (ours)21.352173.4637.6967.5360.92s
Real video--8.0946.976-
", + "image_path": "fb12366941a8ef01bc8ab6d91e69df04b47f13b6cfc423a0787100979fe90950.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 187, + 553, + 279 + ], + "blocks": [ + { + "bbox": [ + 314, + 153, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 314, + 153, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 314, + 153, + 553, + 186 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with existing portrait image animation approaches on the HDTF dataset. Time stands for the averaging time cost of generating one second of 25 fps video." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 187, + 553, + 279 + ], + "lines": [ + { + "bbox": [ + 317, + 187, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 317, + 187, + 553, + 279 + ], + "type": "table", + "html": "
MethodFID↓FVD↓Sync-C↑Sync-D↓
SadTalker [46]32.343487.9244.3047.621
EchoMimic [6]21.058668.6753.2929.096
AniPortrait [37]30.696476.1972.32111.542
Hallo [40]19.826537.4784.0628.552
Teller (ours)20.352429.2884.4967.936
Real video--5.2237.069
", + "image_path": "5c9b14e976c9638a21e899e9ac40ce8026d8f4c6d8aea716dd5a6d38c53db999.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 281, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 314, + 281, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 281, + 553, + 304 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with existing portrait image animation approaches on the RAVDESS dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 313, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 553, + 338 + ], + "type": "text", + "content": "total time is 71ms, with the VAE encoder and decoder averaging 25ms, and the Temporal Module averaging 21ms." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 346, + 432, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 432, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 432, + 358 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 483 + ], + "type": "text", + "content": "Comparison on the HDTF Dataset. Table 1 presents quantitative results for portrait animation techniques on the HDTF dataset. Our method outperforms others, achieving the lowest FVD of 173.463 and a competitive FID of 21.352, indicating high quality and temporal coherence in animated talking heads. Additionally, it achieves the highest Sync-C score of 7.696 and the lowest Sync-D score of 7.536, demonstrating excellent lip synchronization. These results highlight Teller's effectiveness in maintaining both visual fidelity and synchronization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 591 + ], + "type": "text", + "content": "Comparison on the RAVDESS Dataset. Table 2 shows quantitative results on the RAVDESS dataset. Our method again leads in performance, with the lowest FVD of 429.288 and a competitive FID of 20.352, reflecting high-quality and temporally coherent animations. The highest Sync-C score of 4.496 and a competitive Sync-D score of 7.936 further demonstrate superior lip synchronization. These findings confirm Teller's strength in producing synchronized and high-fidelity animated portraits." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 599, + 425, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 425, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 425, + 612 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Head Movement Comparison. Figure 5 shows a qualitative comparison of head movements. Teller replicates natural head movements more accurately, closely matching the ground truth (GT) with smooth, realistic turns and subtle expression-based adjustments. Competing methods, like AntiPortrait and EchoMimic, often show abrupt or limited movements, appearing rigid or lifeless. Teller's autoregressive framework ensures continuity and natural dynamics in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21080" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 293, + 224 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 293, + 224 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 293, + 224 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 293, + 224 + ], + "type": "image", + "image_path": "38ff0996b8342e7d6677a1996802c3cf06c579a35c7343744c815a6caca4bd45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 255, + 294, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 295, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 254 + ], + "type": "text", + "content": "Figure 6. Top-k selection " + }, + { + "bbox": [ + 55, + 232, + 295, + 254 + ], + "type": "inline_equation", + "content": "(\\mathrm{k} = 15)" + }, + { + "bbox": [ + 55, + 232, + 295, + 254 + ], + "type": "text", + "content": " in FMLG produces diverse facial expressions and actions with accurate lip sync on the HDTF." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 255, + 294, + 460 + ], + "lines": [ + { + "bbox": [ + 58, + 255, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 58, + 255, + 294, + 460 + ], + "type": "image", + "image_path": "b14dfdea6bea655b88263978b2fb7bd3c7acfc1849538659b5a6c3f73d3a51ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 495, + 294, + 580 + ], + "blocks": [ + { + "bbox": [ + 55, + 467, + 295, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 489 + ], + "type": "text", + "content": "Figure 7. Visualization of finer motion details. Videos are available in the supplementary materials." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 495, + 294, + 580 + ], + "lines": [ + { + "bbox": [ + 57, + 495, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 57, + 495, + 294, + 580 + ], + "type": "image", + "image_path": "e2f68803cb488bf6c9d4b5ab468a9914f9eb29c26ebace7d39fcd056d8b018de.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 584, + 289, + 596 + ], + "lines": [ + { + "bbox": [ + 61, + 584, + 289, + 596 + ], + "spans": [ + { + "bbox": [ + 61, + 584, + 289, + 596 + ], + "type": "text", + "content": "Figure 8. Visualization of the generation accuracy of lip shape." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "content": "head and emotion alignment, essential for lifelike animation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Diversity. Fig. 6 shows two sequences generated with Top-k sampling (k=15) in FMLG, where each row represents frames from the same speech input. The model demonstrates diverse facial expressions and head movements while maintaining accurate lip sync, highlighting Top-k sampling's role in enhancing motion variety without compromising synchronization." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 394, + 193 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 394, + 193 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 394, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 394, + 193 + ], + "type": "image", + "image_path": "8bd03e69009f82e3ad6bb1c37ba5126f9aab335c8191e1d27e6d69cc3cfe6b21.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 214 + ], + "type": "text", + "content": "Figure 9. Human evaluation results among our proposed Teller and other SoTA methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 397, + 72, + 473, + 193 + ], + "blocks": [ + { + "bbox": [ + 397, + 72, + 473, + 193 + ], + "lines": [ + { + "bbox": [ + 397, + 72, + 473, + 193 + ], + "spans": [ + { + "bbox": [ + 397, + 72, + 473, + 193 + ], + "type": "image", + "image_path": "745b71d2a611d16d33f43cc260fd628402b8dccefcf788252b634f089b243b59.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 476, + 72, + 552, + 193 + ], + "blocks": [ + { + "bbox": [ + 476, + 72, + 552, + 193 + ], + "lines": [ + { + "bbox": [ + 476, + 72, + 552, + 193 + ], + "spans": [ + { + "bbox": [ + 476, + 72, + 552, + 193 + ], + "type": "image", + "image_path": "728a8e415ca6cb26841caba08b410ae33c8756363af3c311c560ba7968c5f814.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 252 + ], + "type": "text", + "content": "Emotional Expression. Fig. 4 presents our model has more accurate emotional expression ability due to the better speech understanding ability of AR transformer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 253, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 554, + 335 + ], + "type": "text", + "content": "Finer Motion Details. Fig. 7 highlights fine motion details in neck and earring movements. Compared to other methods, Teller produces realistic, nuanced motions synchronized with speech, capturing subtle audio-driven dynamics that contribute to lifelike and temporally coherent animation. The consistent detail across frames underscores Teller's robustness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 396 + ], + "type": "text", + "content": "Lip Synchronization Accuracy. Fig. 8 shows superior lip synchronization, with Teller aligning generated lip shapes closely to natural movements. This high fidelity highlights Teller's strength in producing realistic, synchronized mouth motions for convincing talking head animation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 403, + 425, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 403, + 425, + 414 + ], + "spans": [ + { + "bbox": [ + 314, + 403, + 425, + 414 + ], + "type": "text", + "content": "4.4. Human Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 419, + 555, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 576 + ], + "type": "text", + "content": "We conducted a human evaluation to assess the quality of generated animations, focusing on lip synchronization, body movement realism, and temporal coherence. Thirty participants (66.7% aged 24-30, 33.3% aged 30-40; 30% male, 70% female; 83.3% with AIGC model experience) rated each animation on a 5-point Likert scale for coherence with input and animation quality. A total of 100 videos were presented in random order to avoid bias, providing insights into subjective perceptions of animation quality and natural expression alignment. As shown in Figure 9, participants rated Teller highest in lip synchronization, body movement realism, and temporal coherence, with low variance in scores, indicating robust and consistent performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 586, + 406, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 406, + 600 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 406, + 600 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Stage (Module) Ablation. We perform an ablation study comparing Stage 1 and Stage 2, focusing on animation quality in body parts and accessories, especially neck muscles and earrings. As shown in Fig. 10, Stage 1 outputs often show less realistic, inconsistent movements. In contrast, Stage 2, with ETM, significantly improves the physical consistency of subtle motions, creating natural earring sway and smooth neck movements for lifelike, temporally coherent animation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21081" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 243 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 243 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 243 + ], + "type": "image", + "image_path": "3941306c3f53f958fbed40f3dbfbcd0f35264b47aa42ee41c0ac07fa429be58f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 295, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 269 + ], + "type": "text", + "content": "Figure 10. Visualization of the generation images with different stages and the differences between stages." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 271, + 293, + 393 + ], + "blocks": [ + { + "bbox": [ + 58, + 271, + 293, + 393 + ], + "lines": [ + { + "bbox": [ + 58, + 271, + 293, + 393 + ], + "spans": [ + { + "bbox": [ + 58, + 271, + 293, + 393 + ], + "type": "image", + "image_path": "833776ffb35d18de0fa849bac2b072534a5e4e3ae0702cdcc74605243fb9df99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 397, + 295, + 419 + ], + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 419 + ], + "type": "text", + "content": "Figure 11. Tradeoff between performance (loss) and different compression(tokens/frame) ratios." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "text", + "content": "Frames / Tokens in RVQ Module Figure 11 shows the impact of frame and token configurations in the RVQ module on face shape accuracy in talking head animations. This ablation study explores how varying frames and tokens affects facial movement quality and synchronization. Results show that increasing frames and tokens improves facial dynamics, enhancing lip synchronization and realism, though with higher computational costs. We selected 4 frames and 32 tokens for an optimal balance, adjustable as needed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "content": "Ablation Study on Audio Condition Encoder. We analyzed the performance differences between TTS and ASR models as audio condition encoders for our task. Specifically, we used Whisper for the ASR model and funcodec for the TTS model. As shown in Table 3, Whisper achieved a Sync-C score of 7.696 and a Sync-D score of 7.536, while funcodec scored 4.286 in Sync-C and 10.373 in Sync-D. These results indicate that our task benefits more from the ASR model's capability to capture the nuances required for precise synchronization, as seen in Whisper's higher Sync-C and lower Sync-D scores. This analysis suggests that the ASR model is better suited as the audio condition encoder, enhancing the overall quality and synchronization of our talking head animation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Ablation Study on Single-Head vs. Multi-Head Archi" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 353, + 62, + 512, + 109 + ], + "blocks": [ + { + "bbox": [ + 353, + 62, + 512, + 109 + ], + "lines": [ + { + "bbox": [ + 353, + 62, + 512, + 109 + ], + "spans": [ + { + "bbox": [ + 353, + 62, + 512, + 109 + ], + "type": "table", + "html": "
MethodSync-C ↑Sync-D ↓
w funcdec4.28610.373
w whisper7.6967.536
", + "image_path": "7fc76daad3883dc6698abdbf65f4bcae812c2e4129bc3d461d8cc1481988b04a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 137, + 558, + 185 + ], + "blocks": [ + { + "bbox": [ + 313, + 111, + 555, + 134 + ], + "lines": [ + { + "bbox": [ + 313, + 111, + 555, + 134 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 555, + 134 + ], + "type": "text", + "content": "Table 3. Comparison of synchronization for audio conditions using funcodec and Whisper in ASR and TTS tasks on HDTF." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 137, + 558, + 185 + ], + "lines": [ + { + "bbox": [ + 315, + 137, + 558, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 137, + 558, + 185 + ], + "type": "table", + "html": "
MethodFID↓FVD↓Sync-C↑Sync-D↓
Single-Head22.110172.5537.7907.474
Multi-Head21.352173.4637.6967.536
", + "image_path": "b7a83da0013f3be62bce5254a11ae5ad4c63b59466852e71da4f65edfe8fcaf8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 186, + 555, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 207 + ], + "type": "text", + "content": "Table 4. Comparison of performance between Single-Head and Multi-Head on the HDTF dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 217, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 326 + ], + "type": "text", + "content": "tecture. We compare single-head and multi-head models on FID, FVD, Sync-C, and Sync-D metrics (Table 4). The single-head model slightly outperforms in FID (22.110 vs. 21.352) and has comparable FVD (172.553 vs. 173.463), with marginally better synchronization (Sync-C of 7.790 vs. 7.696 and Sync-D of 7.474 vs. 7.536). Both architectures yield competitive results, though the single-head model slightly excels in synchronization, while the multi-head model offers greater real-time potential." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 388, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 388, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 388, + 348 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 355, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 556, + 559 + ], + "type": "text", + "content": "In this paper, we presented Teller, the first autoregressive framework designed for real-time, audio-driven portrait animation. Addressing the challenge of realistic and efficient talking head generation, Teller achieves high-quality animations at up to 25 FPS, surpassing existing methods in both fidelity and responsiveness. Extensive experiments demonstrated Teller's advantages over SoTA audio-driven animation methods, particularly in rendering nuanced movements essential for lifelike and visually convincing animations. Human evaluations further validate its quality, particularly in natural expression and lip synchronization. By balancing computational efficiency with high animation fidelity, Teller sets a new standard for real-time talking head animation, marking a significant advancement in multimodal portrait animation frameworks. Additionally, Teller's AR Transformer architecture makes it compatible with existing unified multimodal language models." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21082" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Barlas Oguz. Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 169 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 236 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "text", + "content": "[4] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 294, + 350 + ], + "type": "text", + "content": "[5] Nan Cao, Yu-Ru Lin, Xiaohua Sun, David Lazer, Shixia Liu, and Huamin Qu. Whisper: Tracing the spatiotemporal process of information diffusion in real time. IEEE transactions on visualization and computer graphics, 18(12):2649-2658, 2012. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 397 + ], + "type": "text", + "content": "[6] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 295, + 475 + ], + "type": "text", + "content": "[7] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19, pages 424-432. Springer, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "text", + "content": "[8] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 534, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 294, + 589 + ], + "type": "text", + "content": "[9] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2663-2671, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 591, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 655 + ], + "type": "text", + "content": "[10] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T. Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: a speaker-independent audio-visual model for speech separation. ACM Transactions on Graphics, 37(4):1-11, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[11] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[12] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "text", + "content": "[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "text", + "content": "[14] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 231, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 284 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 286, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 330 + ], + "type": "text", + "content": "[16] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 1, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "text", + "content": "[17] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "type": "text", + "content": "[18] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "text", + "content": "[19] Steven R Livingstone and Frank A Russo. The ryerson audio-visual database of emotional speech and song (ravdess): A dynamic, multimodal set of facial and vocal expressions in north american english. PloS one, 13(5): e0196391, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 500, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 521 + ], + "type": "text", + "content": "[20] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[21] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (ToG), 40(6):1-17, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "type": "text", + "content": "[22] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 614, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 677 + ], + "type": "text", + "content": "[23] Youxin Pang, Yong Zhang, Weize Quan, Yanbo Fan, Xiaodong Cun, Ying Shan, and Dong-ming Yan. Dpe: Disentanglement of pose and expression for general video portrait editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 427-436, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[24] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21083" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "28th ACM international conference on multimedia, pages 484-492, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "text", + "content": "[25] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "type": "text", + "content": "[26] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 250 + ], + "type": "text", + "content": "[27] Michal Stypulkowski, Konstantinos Vougioukas, Sen He, Maciej Zieba, Stavros Petridis, and Maja Pantic. Diffused heads: Diffusion models beat gans on talking-face generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5091-5100, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 294 + ], + "type": "text", + "content": "[28] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "text", + "content": "[29] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-jin Liu. Diffpositalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM Transactions on Graphics (TOG), 43(4):1-9, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[30] Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, and Mohit Bansal. Any-to-any generation via composable diffusion. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[31] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models, 2024. URL https://arxiv.org/abs/2405.09818, 2024.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[32] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[33] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 580 + ], + "type": "text", + "content": "[34] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17979-17989, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "text", + "content": "[35] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10039-10049, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 680 + ], + "type": "text", + "content": "[36] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[37] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[38] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "text", + "content": "[39] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "text", + "content": "[40] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[41] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "text", + "content": "[42] Ziyu Yao, Xuxin Cheng, and Zhiqi Huang. Fd2talk: Towards generalized talking head generation with facial decoupled diffusion model. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3411-3420, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "type": "text", + "content": "[43] Hanrong Ye, De-An Huang, Yao Lu, Zhiding Yu, Wei Ping, Andrew Tao, Jan Kautz, Song Han, Dan Xu, Pavlo Molchanov, et al. X-vila: Cross-modality alignment for large language model. arXiv preprint arXiv:2405.19335, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 459 + ], + "type": "text", + "content": "[44] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 513 + ], + "type": "text", + "content": "[45] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 580 + ], + "type": "text", + "content": "[46] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 581, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 555, + 635 + ], + "type": "text", + "content": "[47] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[48] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[49] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21084" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 115 + ], + "type": "text", + "content": "face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176-4186, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[50] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics (TOG), 39(6):1-15, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[51] Jinguo Zhu, Xiaohan Ding, Yixiao Ge, Yuying Ge, Sijie Zhao, Hengshuang Zhao, Xiaohua Wang, and Ying Shan. Vl-gpt: A generative pre-trained transformer for vision and language understanding and generation. arXiv preprint arXiv:2312.09251, 2023. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21085" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Temporal Action Detection Model Compression by Progressive Block Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_content_list.json b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a968c7a41eb0d2857fd349b245625023de4464e0 --- /dev/null +++ b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_content_list.json @@ -0,0 +1,1609 @@ +[ + { + "type": "text", + "text": "Temporal Action Detection Model Compression by Progressive Block Drop", + "text_level": 1, + "bbox": [ + 122, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyong Chen $^{1,2*}$ , Yong Guo $^{3*}$ , Jiaming Liang $^{2}$ , Sitong Zhuang $^{2}$ , Runhao Zeng $^{1,2\\dagger}$ , Xiping Hu $^{1,4\\dagger}$ $^{1}$ Guangdong-Hong Kong-Macao Joint Laboratory for Emotional Intelligence and Pervasive Computing, Shenzhen MSU-BIT University, $^{2}$ Shenzhen University, $^{3}$ South China University of Technology, $^{4}$ Beijing Institute of Technology", + "bbox": [ + 99, + 179, + 916, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xiaoyongchencs@gmail.com, zengrh@smbu.edu.cn, huxp@smbu.edu.cn", + "bbox": [ + 233, + 253, + 784, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal action detection (TAD) aims to identify and localize action instances in untrimmed videos, which is essential for various video understanding tasks. However, recent improvements in model performance, driven by larger feature extractors and datasets, have led to increased computational demands. This presents a challenge for applications like autonomous driving and robotics, which rely on limited computational resources. While existing channel pruning methods can compress these models, reducing the number of channels often hinders the parallelization efficiency of GPU, due to the inefficient multiplication between small matrices. Instead of pruning channels, we propose a Progressive Block Drop method that reduces model depth while retaining layer width. In this way, we still use large matrices for computation but reduce the number of multiplications. Our approach iteratively removes redundant blocks in two steps: first, we drop blocks with minimal impact on model performance; and second, we employ a parameter-efficient cross-depth alignment technique, fine-tuning the pruned model to restore model accuracy. Our method achieves a $25\\%$ reduction in computational overhead on two TAD benchmarks (THUMOS14 and ActivityNet-1.3) to achieve lossless compression. More critically, we empirically show that our method is orthogonal to channel pruning methods and can be combined with it to yield further efficiency gains.", + "bbox": [ + 91, + 337, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 760, + 220, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal action detection aims to identify action instances and localize their start and end times in untrimmed videos. This task underpins many video understanding applications, including video question answering [33] and video captioning [17]. Recent advancements in larger feature extractors", + "bbox": [ + 89, + 786, + 482, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and datasets have improved model performance but also significantly increased computational demands [40]. However, many TAD applications, such as autonomous driving [59] and robotics [44], operate under limited computational resources. Thus, reducing computational complexity is crucial for deploying TAD in real-world scenarios.", + "bbox": [ + 511, + 305, + 903, + 395 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A typical TAD model [67, 70] consists of a feature extractor and an action detection head. Due to the long duration of videos, the feature extractor processes video segments individually, and these segment-level features are then combined before being passed to the detection head. As shown in Figure 1, our experiments show that the feature extraction phase accounts for $95\\%$ of the total model computation, since it processes many segments, while the detection head processes only a single video-level feature. Thus, reducing the complexity of the feature extractor is key to improving TAD model efficiency.", + "bbox": [ + 511, + 398, + 903, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pruning [14, 18, 20, 48] is a common model compression technique, which removes redundant weights based on their importance. This results in a slim-and-tall network structure, where the model's hidden layer dimensions are reduced but depth remains unchanged, as illustrated in Figure 1. However, studies have shown [24, 45] that for the same number of parameters, a deep-and-thin network may have slower inference speeds compared to a shallow-and-wide network. The reason for this is that, during GPU inference, operations on larger matrices tend to have higher parallelism and are thus more efficient than multiple operations on smaller matrices.", + "bbox": [ + 511, + 566, + 903, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To explore the feasibility of reducing the architectural depth of TAD models, we conduct preliminary experiments on the THUMOS14 dataset [28], which is widely used for this task. We raise the following key observations from Figure 2: 1) There is a slight discrepancy between the input and output features of certain blocks. 2) Dropping specific blocks results in only a minimal decrease in model performance. 3) Simultaneously dropping multiple blocks leads to a more significant performance drop (Table 3). These results motivate us to investigate TAD model compression", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 875, + 218, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29225", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5be35b489cf329a5945034f18ec0535a52d21a958c605a93f7216ef7359c05ec.jpg", + "image_caption": [ + "Figure 1. Comparisons between reducing depth and reducing width. For a 768-frame video sequence with a resolution of $160^2$ , using VideoMAE-S as the feature extractor and Actionformer as the detection head, $95\\%$ of the computational overhead comes from the feature extractor. Most pruning methods reduce the size of layer weight matrices, leading to a slim-and-tall network structure. In contrast, our progressive block drop method reduces network depth, achieving $1.19 \\times$ faster inference at the same computational cost. These findings, detailed in Section 5.6, suggest that our approach results in a more hardware-efficient model." + ], + "image_footnote": [], + "bbox": [ + 109, + 87, + 890, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from the depth perspective.", + "text_level": 1, + "bbox": [ + 89, + 405, + 274, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a multi-step progressive block drop method that effectively reduces network depth, transforming the TAD model into a shallow-and-wide architecture more suitable for parallel GPU matrix computations. Specifically, we compress the model progressively through multiple iterations. Each iteration consists of two stages: 1) block selection and removal, where we evaluate the importance of each block based on a specific metric and remove the least important block; and 2) performance restoration, where we design a parameter-efficient cross-depth alignment strategy. This strategy first integrates LoRA [26] into the attention modules of the remaining blocks and then fine-tunes the pruned model to align with the intermediate features output by the uncompressed model's blocks, effectively compressing the information from multiple blocks into a single block. Additionally, we employ prediction-level alignment to further improve performance. On the THUMOS14 and ActivityNet datasets [4], our method achieves superior performance compared to the uncompressed model, while reducing the model's computational cost (MACs) by $25\\%$ . Furthermore, we demonstrate through experiments that our approach is compatible with common pruning techniques, enabling even further compression of the model. Our main contributions are as follows:", + "bbox": [ + 91, + 426, + 483, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We focus on compressing the TAD model from the perspective of depth. Unlike channel pruning methods that compress the model from the width or the internal layers of blocks, we analyze the redundancy within the TAD model and explore block-level compression as a novel approach. This provides a new angle for compressing and", + "bbox": [ + 89, + 809, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "accelerating the TAD model.", + "bbox": [ + 527, + 405, + 720, + 419 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To achieve depth compression, we propose a progressive multi-step block drop strategy. In each iteration, a block is automatically removed, and a parameter-efficient training technique is employed to recover the model's performance. On THUMOS14 and ActivityNet, our method reduces the model's MACs by approximately $25\\%$ , while simultaneously outperforming the uncompressed model. Furthermore, we extend our method to other TAD architectures, fine-grained datasets, and additional tasks to validate its generalizability.", + "- We empirically demonstrate that our method is orthogonal to channel pruning techniques and investigate its compatibility with other model compression approaches, resulting in additional efficiency improvements." + ], + "bbox": [ + 513, + 420, + 906, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 652, + 656, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Temporal Action Detection", + "text_level": 1, + "bbox": [ + 511, + 679, + 756, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal action detection (TAD) can generally be divided into two primary categories [61]. The first category, two-stage methods, involves generating action proposals followed by classification and boundary refinement [5, 54, 64, 69, 73]. Proposal generation techniques may employ frame-or segment-level classification with subsequent merging of frames or segments belonging to the same action class [49-51], while others use specialized proposal generation strategies [3, 37]. However, the success of these approaches heavily relies on the quality of the proposals, which has led to the emergence of integrated approaches, also known as one-stage methods. These methods simultaneously handle proposal generation, classification, and/or boundary regres", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29226", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sion [27, 36, 58, 68]. Key developments in this area include the adoption of anchor-based mechanisms for TAD [36] and the shift towards anchor-free frameworks [35, 53], with recent efforts combining the benefits of both anchor-based and anchor-free approaches [66].", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, Transformer-based architectures, which have demonstrated remarkable results in various computer vision tasks, have been applied to TAD [56, 62, 70]. Additionally, the development of end-to-end architectures has been explored in various studies [40-42, 67, 72]. With the significant achievements of scaling up in large language models [1], recent research [40] has utilized scaling up in conjunction with a frame input data processing approach. This combination has not only significantly improved model performance but also reduced computational costs. However, the issue of increased inference time due to computational scaling has become increasingly critical. Therefore, we focus on reducing inference latency through model compression.", + "bbox": [ + 89, + 169, + 482, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Model Compression", + "text_level": 1, + "bbox": [ + 89, + 392, + 281, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model compression techniques are typically categorized into pruning [20], knowledge distillation [23], model decomposition [9], and quantization [30]. This paper focuses on model pruning, which can be either unstructured [20] or structured [34]. Unstructured pruning removes network connections to create sparsity, but it requires specialized hardware for sparse matrix operations [19, 20, 55], limiting its practical deployment. Structured pruning, which includes width-reduction methods like channel pruning [21, 34] and filter pruning [22], results in a dense weight matrix, enabling GPU acceleration. However, it may not fully exploit GPU parallelism [39].", + "bbox": [ + 89, + 415, + 482, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depth-wise pruning techniques [25, 57] have been used for model acceleration, although they may suffer from insufficient memory occupancy. Various methods, including layer importance evaluation [7, 12], parameter optimization [65, 74], and reparameterization [11, 15], can also guide pruning to compress the model from the perspective of depth. Previous work [38, 71] has explored the concept of progressive block dropping; however, there are significant differences between their approaches and ours. Specifically, [38] focuses on compressing and integrating the layers within the block, which is a layer-level pruning technique. The dynamically skipping blocks method proposed in [71] accelerates training by dropping blocks, but during inference, the full model is still employed. Additionally, skipping blocks to reduce the depth of Transformer models [10, 46] often leads to a loss in accuracy. In contrast, our progressive drop block method reduces depth at the block level, improving both inference speed and detection accuracy. Our method can be combined with structured pruning [13] and unstructured activation sparsity [32, 47] to", + "bbox": [ + 89, + 598, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e6c7801eb3d8547ca1c5438d6e34f28c25a4802798c0303b16f6135231c09027.jpg", + "image_caption": [ + "Figure 2. Analysis of TAD models at block level on THUMOS14 datasets, using VideoMAE-S. Blue curve: impact of dropping a single block on the model's detection accuracy, where most cases result in minimal accuracy degradation. Green line: MSE of the input and output features for each block. Some blocks exhibit MSE values close to 0, indicating that these blocks are redundant." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 905, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "further compress the model and accelerate inference while improving performance.", + "bbox": [ + 511, + 354, + 905, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries and Motivation", + "text_level": 1, + "bbox": [ + 511, + 398, + 784, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the notations and formulations pertinent to temporal action detection (TAD) and then explore the feasibility of reducing the depth of TAD models, as well as some key findings that inspire the design of subsequent methods.", + "bbox": [ + 511, + 425, + 905, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. TAD Formulation and Notation", + "text_level": 1, + "bbox": [ + 511, + 511, + 792, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TAD aims to identify action instances within a video and determine their temporal boundaries. Given an untrimmed video represented as $V = \\{I_t\\}_{t=1}^T$ , where $I_t$ denotes the frame at time $t$ , TAD predicts a set of action instances $\\Phi_V = \\{\\phi_i = (t_{s_i}, t_{e_i}, c_i)\\}_{i=1}^N$ . Here, $N$ is the number of action instances; for each instance $\\phi_i$ , $t_{s_i}$ and $t_{e_i}$ are its starting and ending times, respectively, and $c_i$ denotes its category.", + "bbox": [ + 511, + 534, + 905, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feasibility of Reducing TAD model's Depth", + "text_level": 1, + "bbox": [ + 511, + 651, + 880, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We find that there is a severe redundancy in the depth of TAD models. To verify this, we conduct a series of experiments and summarize the key observations as follows.", + "bbox": [ + 511, + 672, + 905, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observation 1: Most computational overhead lies in the clip-level feature extraction backbone. TAD methods typically consist of two main parts: the feature extractor and the detection head. In the feature extraction stage, a sliding window of length $L$ is processed by a video feature extractor. For instance, features are extracted from frames $\\{I_t, \\dots, I_{t + L}\\}$ , and then the window shifts by $s$ frames to extract features from frames $\\{I_{t + s}, \\dots, I_{t + s + L}\\}$ . As the feature extraction process is repeated across the entire video sequence, it incurs substantial computational cost, with each window requiring complex operations to extract features. Once the features are extracted, they are con", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29227", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c456f149923bb5398299756c2249af0468b11b6005f554c5e0f36d0fae608e22.jpg", + "image_caption": [ + "Figure 3. The diagram of our progressive block drop method. Our approach adopts a multi-step progressive compression strategy. At each iteration 1) we evaluate the importance of each block and drop the least important block, and 2) we use parameter-efficient tuning techniques, and recover performance by learning from the uncompressed model through feature-level and prediction-level alignment." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "catenated along the temporal dimension, and information is fused across the spatial dimension by averaging the features. These fused features are then passed to the detection head. Ultimately, the volume of data fed into the detection head is much smaller compared to the data processed by the feature extractor. As shown in Figure 1, the computational cost of the feature extractor accounts for $95\\%$ of the total forward inference cost of the model. Given the high computational cost of feature extraction, optimizing or reducing the complexity of this stage presents a significant opportunity for improving model efficiency. Therefore, to reduce the overall model computation, the primary focus should be on compressing the feature extractor.", + "bbox": [ + 88, + 494, + 482, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 2: There exist blocks that have a minor discrepancy between their input and output. Additionally, by analyzing the feature differences between consecutive blocks, as depicted in Figure 2, we find that some blocks exhibit very similar representation capabilities, suggesting redundancy in the model. These findings highlight the feasibility of applying the block drop method to reduce computational complexity without sacrificing performance.", + "bbox": [ + 88, + 698, + 482, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 3: The removal of certain blocks has a negligible impact on TAD performance. To investigate the potential for model compression, we first assess the feasibility of dropping blocks by examining the impact of removing individual blocks on model performance. As shown", + "bbox": [ + 88, + 825, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Figure 2, our experiments reveal that removing certain blocks led to only minimal performance degradation, indicating that block drop could be a viable approach.", + "bbox": [ + 511, + 494, + 906, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 4: Simultaneously removing multiple blocks at a time results in a significant performance drop. In our preliminary exploration, we attempt to directly drop multiple blocks and train the model to observe whether it could recover its performance. We randomly select three blocks to drop all at once and then attempt fine-tuning. As shown in Table 3, the best performance achieved is $0.67\\%$ lower than the uncompressed model, exhibiting a noticeable performance drop, which is an undesirable outcome in practical applications. Therefore, we adopt a multi-step progressive block dropping strategy rather than removing multiple blocks at once. This approach allows the information from the removed blocks to be more effectively compressed into the pruned model, resulting in better performance.", + "bbox": [ + 511, + 542, + 908, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Progressive Block Drop", + "text_level": 1, + "bbox": [ + 511, + 768, + 736, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we design a compression method specifically tailored for TAD models. Drawing from our analysis in Section 3.2, our focus is on reducing the computational cost of the feature extraction backbone while maintaining or even enhancing performance. To achieve this, we propose a multi-step progressive block-level dropping method, where each step comprises two stages. First, the Block Selection", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29228", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluator automatically selects the blocks to be dropped, resulting in a subnet. Then, Cross-Depth Alignment enables parameter-efficient learning from the uncompressed model to maintain TAD performance. By iteratively applying these steps, we progressively obtain a pruned TAD backbone that can be directly deployed for inference.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Block Selection Evaluator", + "text_level": 1, + "bbox": [ + 89, + 190, + 326, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental results presented in Section 3.2 demonstrate that simultaneously removing multiple blocks significantly degrades the performance of TAD. Inspired by prior work on curriculum learning [2, 63], we propose a progressive block-drop strategy. However, selecting the appropriate blocks to drop is a non-trivial task. To address this challenge, we introduce a block selection evaluator $f_{E}$ to select the blocks to be dropped automatically.", + "bbox": [ + 89, + 212, + 483, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Without loss of generality, let $M_0$ denote the initial uncompressed model trained on the training set, and let $M_{t-1}$ represent the model input to $f_E$ at iteration $t$ . We denote $B_{t-1} = \\{b_1, b_2, \\ldots, b_K\\}$ as the set of all blocks within $M_{t-1}$ , where $K$ is the total number of blocks. At iteration $t$ , the collection of sub-networks resulting from dropping each single block is denoted as:", + "bbox": [ + 89, + 332, + 483, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {t} = \\left\\{M _ {t, k} ^ {\\text {s u b}} = M _ {t - 1} \\backslash b _ {k} \\mid b _ {k} \\in B _ {t - 1} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 446, + 482, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To automatically select blocks for pruning, we employ the evaluation function $f_{E}$ to assess each subnet. By measuring the performance difference between each subnet and the uncompressed TAD model $M_0$ on the training set, $f_{E}$ outputs an importance score for each subnet; the smaller the performance difference, the higher the importance score assigned. This process can be formulated as:", + "bbox": [ + 89, + 473, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {t} = f _ {E} \\left(\\mathcal {S} _ {t}, D _ {\\text {t r a i n}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 588, + 482, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From the set of importance scores $\\mathcal{P}_t$ , we select the subnet $M_{t,*}^{sub}$ with the highest importance score $p^*$ , resulting in the pruned model $M_t$ at step $t$ . Note that $f_{E}$ can also compute the difference between the input and output features of each removed block or the training loss of each subnet on the training set; in both cases, a smaller difference corresponds to a higher importance score. In our experiments (Section F.1 of the appendix), we find that selecting subnets based on the mAP metric from the training set yields the best results.", + "bbox": [ + 89, + 613, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Parameter-Efficient Learning from Uncompressed Model for Performance Recovery", + "text_level": 1, + "bbox": [ + 89, + 757, + 483, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After removing one block from the TAD model, a central challenge arises in ensuring that the performance of the reduced-depth model remains comparable to that of the original, uncompressed model (see Section F.2 of appendix). However, training the TAD model on untrimmed videos using full-parameter fine-tuning still demands a substantial amount of GPU memory, as discussed in [40]. For", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Proposed Progressive Block Drop Method" + ], + "code_body": "Input: Uncompressed model $M_0$ , model $M_{t-1}$ at step $t$ , blocks $B_{t-1} = \\{b_1, b_2, \\ldots, b_K\\}$ within $M_{t-1}$ , training set $D_{train}$ . // Progressive block drop cycle while Performance of $M_{t-1} \\geq M_0$ do // Stage 1: select one block to drops Obtain sub-nets by dropping each block: $\\mathcal{S}_t = \\left\\{M_{t,k}^{\\mathrm{sub}} = M_{t-1} \\setminus b_k \\mid b_k \\in B_{t-1}\\right\\}$ Evaluate each sub-net using the evaluator $f_E$ : $\\mathcal{P}_t = f_E(\\mathcal{S}_t, D_{train})$ Select the subnet with the highest performance: $M_t = \\arg \\max(\\mathcal{P}_t)$ // Stage 2: performance recovery Insert LoRA modules into $M_t$ for training: $M_t(\\theta) \\rightarrow \\hat{M}_t(\\theta, \\theta_{\\mathrm{LoRA}})$ Train the model with Eqn. (6) Merge LoRA parameters into the model: $\\theta_t \\gets \\theta_t + \\theta_{\\mathrm{LoRA}}'$ end while", + "bbox": [ + 514, + 109, + 906, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "example, in our experiments with the VideoMAE-L model we used, setting the batch size to 1 already consumes 13.1 GB of memory. Therefore, we introduce LoRA to reduce memory consumption during training, with detailed experimental results provided in Section F.3 of the appendix.", + "bbox": [ + 511, + 431, + 906, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Before training, given the pruned model $M_{t}(\\theta_{t})$ at step $t$ , where $\\theta_{t}$ represents its parameters, we first follow the LoRA method [8, 26] to add trainable parameters $\\theta_{\\mathrm{LORA}}$ to the attention layer each block, resulting in the model $\\hat{M}_{t}(\\theta_{t},\\theta_{\\mathrm{LORA}})$ . This approach offers two main benefits: first, during subsequent training, only $\\theta_{\\mathrm{LORA}}$ is trained, significantly reducing memory consumption; second, after training, $\\theta_{\\mathrm{LORA}}$ can be merged with $\\theta_{t}$ , thereby keeping the TAD model's structure unchanged and incurring no additional inference computational overhead.", + "bbox": [ + 511, + 508, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Depth Alignment. The structural differences between the pruned and uncompressed models—resulting from block dropping—pose a significant challenge for recovering performance. More critically, we empirically observed that directly supervising the training of $\\hat{M}_t$ with ground truth fails to restore performance (see Section E.5). In other words, relying solely on prediction-level supervision is inadequate for restoring performance. To address this issue, we shift our attention to feature-level supervision and propose a cross-depth alignment strategy.", + "bbox": [ + 511, + 659, + 905, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, if we consider three adjacent blocks, $\\{b_{i - 1}, b_i, b_{i + 1}\\}$ , as a group, when $b_i$ is dropped, the input to $b_{i + 1}$ changes (i.e., it becomes the output of $b_{i - 1}$ ). To maintain performance, it is necessary to: 1) ensure that the input feature distribution of $b_{i + 1}^p$ in the pruned model $\\hat{M}_t$ aligns with that of $b_i^u$ in the uncompressed model $M_0$ ; and", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29229", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2) ensure that the output distribution of the $(i + 1)$ -th block in $\\hat{M}_t$ aligns with that of $M_0$ . In our implementation, after the $i$ -th block is dropped, we align the outputs of the $(i - 1)$ -th and $(i + 1)$ -th blocks $(f_{b_{i - 1}}$ and $f_{b_{i + 1}})$ between the pruned and uncompressed models, thereby compressing the information from multiple blocks into a single block. For the features of other blocks, we similarly apply block-wise alignment with the $M_0$ . The cross-depth alignment loss can be expressed as follows:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = \\frac {1}{I - 1} \\sum_ {m = 1, m \\neq i} ^ {I} \\left(f _ {b _ {m}} ^ {M _ {0}} - f _ {b _ {m}} ^ {\\hat {M} _ {t}}\\right) ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 234, + 483, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Action classification and Localization Alignment. The TAD model generates two distinct types of predictions: action categories and temporal bounding boxes. We also align the predictions of the pruned model $\\hat{M}_t$ with those of the uncompressed model $M_0$ .", + "bbox": [ + 89, + 282, + 483, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1) Action classification alignment: To align the action class predictions, we employ the Kullback-Leibler (KL) divergence loss $\\mathcal{L}_{\\mathrm{KL}}$ [31], which measures the divergence between the action class scores of the uncompressed model $M_0$ and the pruned model $\\hat{M}_t$ . The alignment loss is defined as:", + "bbox": [ + 89, + 358, + 483, + 446 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p c} = \\mathcal {L} _ {\\mathrm {K L}} \\left(p \\left(z _ {M _ {0}}\\right), p \\left(z _ {\\hat {M} _ {t}}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 449, + 480, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z_{M_0}$ and $z_{\\hat{M}_t}$ denote the logits of the uncompressed and pruned models, respectively, and $p(z_{M_0})$ and $p(z_{\\hat{M}_t})$ represent the class probabilities derived from the logits via the softmax function.", + "bbox": [ + 89, + 469, + 483, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2) Boundary regression alignment: For aligning the boundary predictions, we compute the Generalized Intersection over Union (GIoU) loss $\\mathcal{L}_{\\mathrm{GIoU}}$ [52] between the predicted temporal bounding boxes of the uncompressed model $M_0$ and the pruned model $\\hat{M}_t$ . The alignment loss for boundary regression is:", + "bbox": [ + 89, + 529, + 483, + 619 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\mathcal {L} _ {\\mathrm {G I o U}} \\left(D _ {i} ^ {M _ {0}}, D _ {i} ^ {\\hat {M} _ {t}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 627, + 482, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N_{p}$ denotes the number of predicted positive temporal bounding boxes, and $D_{i}^{M_{0}}$ and $D_{i}^{\\hat{M}_{t}}$ represent the $i$ -th predicted temporal bounding boxes for models $M_0$ and $\\hat{M}_t$ , respectively. This dual alignment ensures that the pruned model learns both the correct action classification and the accurate temporal boundaries, which are crucial for TAD.", + "bbox": [ + 89, + 674, + 483, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Total Loss Function. We incorporate the alignment losses, along with the commonly used classification loss $\\mathcal{L}_{\\mathrm{cls}}$ and regression loss $\\mathcal{L}_{\\mathrm{reg}}$ in TAD [40], as follows:", + "bbox": [ + 89, + 771, + 483, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {p c} + \\mathcal {L} _ {p r} + \\mathcal {L} _ {f} + \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 823, + 482, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After training, the LoRA parameters $\\theta_{\\mathrm{LoRA}}^{\\prime}$ are integrated with the pruned model as:", + "bbox": [ + 89, + 847, + 483, + 877 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} \\leftarrow \\theta_ {t} + \\theta_ {\\mathrm {L o R A}} ^ {\\prime}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 885, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The updated model $M_{t}(\\theta_{t})$ is then input into the block drop selection evaluator at the $t + 1$ step for the next iteration, continuing until the performance of the pruned model is no longer recoverable to that of the uncompressed model. Ultimately, the compressed TAD model is deployed, offering a reduced number of parameters and faster inference speed.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 198, + 645, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the experimental section, we conduct the following experiments: 1) Sections 5.3 and 5.4 evaluate the effectiveness of the proposed method; 2) Section 5.5 validates compatibility with other compression models; 3) Section 5.6 compares inference speed with pruning methods.", + "bbox": [ + 511, + 224, + 905, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Evaluation Metric", + "text_level": 1, + "bbox": [ + 511, + 311, + 794, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "THUMOS14 [28] is a prominent benchmark dataset in TAD research. It consists of a training set with 200 videos and a test set with 213 videos, covering 20 action categories. This task is notably challenging due to the presence of more than 15 separate action instances per video, with background scenes occupying $71\\%$ of the frames.", + "bbox": [ + 511, + 334, + 905, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ActivityNet-1.3 [4] is a large-scale benchmark for TAD in untrimmed videos, containing approximately 20,000 videos across 200 activity classes, totaling over 600 hours of content. The dataset is divided into three subsets: 10,024 videos for training, 4,926 for validation, and 5,044 for testing, with an average of 1.65 action instances per video.", + "bbox": [ + 511, + 426, + 905, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FineAction [43] is a benchmark dataset for temporal action localization, containing 103K instances across 106 action categories in 17K untrimmed videos. It presents both opportunities and challenges for TAD due to its fine-grained action classes, dense annotations, and frequent co-occurrence of multiple action classes.", + "bbox": [ + 511, + 517, + 905, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mean Average Precision (mAP). A predicted temporal bounding box is deemed accurate if its temporal Intersection over Union (tIoU) with the corresponding ground truth instance exceeds a specific threshold and matches the instance's category. For THUMOS14, the tIoU thresholds range from $\\{0.3, 0.4, 0.5, 0.6, 0.7\\}$ , with mAP@tIoU=0.5 reported for comparison. In contrast, on ActivityNet-1.3 and FineAction, tIoU thresholds are sampled from $\\{0.5: 0.05: 0.95\\}$ , and we report the mean mAP over tIoU thresholds from 0.5 to 0.95, in increments of 0.05.", + "bbox": [ + 511, + 609, + 905, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 771, + 730, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the ActionFormer [70] detector for prediction, employing two Transformer-based feature extractors: VideoMAE-S with 12 blocks and VideoMAE-L [60] with 24 blocks, both are pre-trained on the K400 dataset [29]. For the THUMOS14 dataset, the input videos are randomly truncated to a fixed length of 768 frames, while for the ActivityNet-1.3 and FineAction dataset, the input videos are", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29230", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/273078fa422663af285d282dc2c1e9c99862013fbb6e081aa274ffbbc16386fc.jpg", + "table_caption": [ + "Table 1. Evaluation of our method in terms of computational complexity, inference time and mAP with VideoMAE-S as the backbone. Our approach reduces the computational cost by approximately $25\\%$ , achieves around $1.3\\mathrm{x}$ inference speedup, and improves the performance." + ], + "table_footnote": [], + "table_body": "
Model (Drop #Block)THUMOS14ActivityNet-1.3
MACs (G)Inference Time (ms)mAP (tIoU=0.5)MACs (G)Inference Time (ms)mAP (Avg)
Baseline286.3104.970.43274.1103.837.75
1263.5 (92.0%)98.4 (93.8%)71.06 (↑ 0.63)251.3 (91.7%)95.1 (91.6%)37.94 (↑ 0.19)
2240.8 (84.1%)89.8 (85.6%)71.37 (↑ 0.94)228.5 (83.4%)86.8 (83.6%)37.81 (↑ 0.06)
3218.0 (76.1%)81.2 (77.4%)70.47 (↑ 0.04)205.8 (75.1%)79.6 (76.7%)37.77 (↑ 0.02)
4195.2 (68.2%)73.6 (70.2%)69.65 (↓ 0.78)183.0 (66.8%)72.1 (69.5%)37.72 (↓ 0.03)
", + "bbox": [ + 158, + 121, + 836, + 210 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c13c104e772ee132dcddfc7605263e937f7ad48fb0ceef696e3310987d06514.jpg", + "table_caption": [ + "Table 2. Evaluation of our method on deeper networks in terms of computational complexity, inference time and mAP with VideoMAE-L (24 blocks) as the backbone. Our method is generally effective and larger models may benefit more from it." + ], + "table_footnote": [], + "table_body": "
Drop #BlockMACs (G)Inference Time (ms)mAP (tIoU=0.5)
Baseline3886.9795.976.01
13725.5 (95.8%)765.4 (96.2%)77.97 (↑ 1.96)
23564.1 (91.7%)731.4 (91.9%)78.00 (↑ 1.99)
33402.7 (87.5%)704.3 (88.5%)78.29 (↑ 2.28)
43241.4 (83.4%)671.3 (84.3%)77.86 (↑ 1.85)
53080.0 (79.2%)647.1 (81.3%)77.57 (↑ 1.56)
62918.6 (75.1%)602.6 (75.7%)77.25 (↑ 1.24)
", + "bbox": [ + 93, + 277, + 480, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "resized to a fixed length of 768 frames, with the frame resolution set to $160^2$ by default for both datasets. Please refer to the supplementary for more details.", + "bbox": [ + 89, + 386, + 483, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Effectiveness of Our Progressive Block Drop", + "text_level": 1, + "bbox": [ + 89, + 440, + 467, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the model trained with the progressive block drop method to the uncompressed model from three aspects. 1) Computational Complexity (MACs): The computational complexity is calculated based on the number of multiply-accumulate operations at each layer of the model. 2) Inference Time: The inference time measures the average duration for processing each input sample, where each sample consists of 768 frames with a resolution of $160^{2}$ . 3) mAP: The results, as shown in Table 1, demonstrate that on both datasets, our method achieves an approximately $25\\%$ reduction in MACs while outperforming the uncompressed model in terms of performance. This highlights the superiority of our approach, as it effectively reduces computational complexity while simultaneously enhancing the performance of the compressed model—an accomplishment that is difficult for other compression methods to achieve.", + "bbox": [ + 88, + 462, + 482, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Drop blocks on a deeper backbone. We explore the performance of the block drop method when the depth of the feature extractor is increased from 12 to 24, i.e., when transitioning from VideoMAE-S to VideoMAE-L. The results, as shown in Table 2, reveal that with a $25\\%$ reduction in MACs, the model achieves a $1.24\\%$ performance improvement. Notably, when the same fraction of blocks (1/4) is dropped, the performance of VideoMAE-L improves by $1.24\\%$ , whereas VideoMAE-S only shows a modest improvement of $0.04\\%$ (as shown in Table 1). This suggests that in models with higher computational complexity, there is more redundant information, and the progressive block drop method is effective at eliminating such redundancy.", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bc705953f9694903d2bd1a2605e26ea10c14259abd836f023a833dbe495cb424.jpg", + "table_caption": [ + "Table 3. Comparison with randomly dropping blocks. Dropping three randomly selected blocks at once performs worse than progressively dropping." + ], + "table_footnote": [], + "table_body": "
Drop Block ID[2,5,9][0,4,7][1,3,8][6,7,10][6,7,10] + ours
mAP68.8868.3169.7668.9170.47
", + "bbox": [ + 517, + 263, + 903, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dropping multiple blocks at one time vs Progressive dropping. To demonstrate the necessity of adopting a multi-step progressive drop in our proposed block drop method, we conduct experiments where multiple blocks are dropped simultaneously. Specifically, we randomly select three different blocks to drop and train the model using the training method described in Section 4.2. The results, as shown in Table 3, indicate that directly dropping three blocks yields performance ranging from $68.31\\%$ to $69.76\\%$ , with the best result still $0.67\\%$ lower than that of the uncompressed model. Meanwhile, using the progressive block drop method on VideoMAE-S and the THUMOS14 dataset, the model's performance after dropping the first three layers [7, 10, 6] reaches $70.47\\%$ . Comparing this with the result of directly dropping these three layers, it is evident that compressing the model via progressive dropping achieves better performance recovery ( $70.47\\%$ vs. $68.91\\%$ ), further underscoring the importance of the multi-step progressive drop. Therefore, our proposed multi-step progressive block drop is crucial for performance recovery.", + "bbox": [ + 511, + 306, + 906, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Extended Evaluations", + "text_level": 1, + "bbox": [ + 511, + 622, + 718, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on additional TAD architectures. To validate the generalizability of our method, we apply it to two additional TAD architectures: the single-stage AdaTAD [40] model with a VideoMAE-S backbone and the two-stage ActionFormer [70] model with pre-extracted I3D features. Table 4 shows that in AdaTAD, pruning two blocks improves accuracy by $0.08\\%$ and reduces MACs to $84.0\\%$ . Similarly, the ActionFormer model gains $0.49\\%$ in mAP while reducing MACs to $59.7\\%$ . These results underscore our method's effectiveness in compressing diverse TAD architectures while enhancing performance.", + "bbox": [ + 511, + 642, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on the fine-grained FineActions dataset. To further corroborate the versatility of our approach, we conduct experiments on the fine-grained FineActions [43] dataset using a VideoMAE-S backbone with an Action-Former [70] head. As shown in Table 5, pruning two blocks slightly improves mAP ( $\\uparrow 0.04\\%$ ), while pruning", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29231", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/115d2fb4b680febd81d837bc1732128e44c9e067619c1b44a0d4a19333c914bf.jpg", + "table_caption": [ + "Table 4. Evaluation of our method on AdaTAD and ActionFormer. Our method is effective across different TAD architectures." + ], + "table_footnote": [], + "table_body": "
Drop #BlockAdaTAD [40]ActionFormer [70]
mAP (tIoU=0.5)MACs (G)mAP (tIoU=0.5)MACs (G)
Baseline72.43324.872.6545.2
172.91 (↑ 0.48)298.8 (92.0%)73.21 (↑ 0.56)37.9 (83.8%)
272.51 (↑ 0.08)272.8 (84.0%)73.28 (↑ 0.63)30.6 (67.7%)
371.68 (↓ 0.75)246.9 (76.0%)73.14 (↑ 0.49)27.0 (59.7%)
", + "bbox": [ + 94, + 117, + 480, + 189 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7e509081bb2d38ab2aef056b137336b9b60e80dcecaf9d4567332003f0da97d.jpg", + "table_caption": [ + "Table 5. Evaluation of our method on fine-grained FineActions [43]. Our approach continues to exhibit efficacy even in scenarios involving more fine-grained actions." + ], + "table_footnote": [], + "table_body": "
Drop #BlockmAP (Avg)MACs (G)Inference Time (ms)
Baseline18.67280.8210.2
118.77 (↑ 0.10)257.7 (91.8%)197.0 (93.7%)
218.71 (↑ 0.04)234.6 (83.5%)185.0 (88.0%)
318.43 (↓ 0.24)211.5 (75.3%)173.0 (82.3%)
", + "bbox": [ + 94, + 236, + 478, + 304 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f49fb1fbf59584be3c34fbb50b20c1392f56c64524f4c4e0d9d6a8382b48e91f.jpg", + "table_caption": [ + "Table 6. Results of pruning Mamba model [6] on Charades-STA. Our method is effective across various tasks and architectures." + ], + "table_footnote": [], + "table_body": "
Drop #BlockParameters (M)R1@0.3R1@0.5
Baseline60.569.4658.58
153.4 (88.4%)72.04 (↑ 2.58)58.79 (↑ 0.21)
246.4 (76.7%)71.75 (↑ 2.29)58.92 (↑ 0.34)
339.3 (65.1%)71.94 (↑ 2.48)58.79 (↑ 0.21)
432.3 (53.4%)70.83 (↑ 1.37)57.72 (↓ 0.86)
", + "bbox": [ + 96, + 335, + 478, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "three blocks causes only a minor drop $(\\downarrow 0.24\\%)$ . Meanwhile, MACs are reduced to $83.5\\%$ and $75.3\\%$ , with a notable decrease in inference time. These results demonstrate that our progressive block drop strategy effectively preserves detection accuracy while significantly reducing computational cost, even for fine-grained action detection.", + "bbox": [ + 89, + 438, + 482, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on the natural language localization task. Finally, to test our method beyond conventional TAD, we apply progressive block drop to the grounding head of the Mamba model [6] on Charades-STA [16] for natural language localization. Selective pruning is performed on the grounding head. As shown in Table 6, removing three out of six blocks reduces the parameter count to $65.1\\%$ of the baseline while increasing performance—R1@0.3 by $2.48\\%$ and R1@0.5 by $0.21\\%$ . These results demonstrate our method's generalizability across tasks and models.", + "bbox": [ + 89, + 529, + 483, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Compatible with Pruning Techniques", + "text_level": 1, + "bbox": [ + 89, + 685, + 413, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We apply the pruning method [13] to models after block drop, considering three setups: 1) pruning, 2) our method, and 3) our method with pruning. As shown in Figure 4, at nearly identical accuracy (70.22% vs. 70.21%), combining our method with pruning reduces MACs by 30%, while pruning alone achieves only 10%. This highlights our method's potential for further integration with pruning to accelerate the model. Additionally, compatibility with sparse activations is analyzed in Section E of the appendix.", + "bbox": [ + 89, + 703, + 482, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Comparison of Inference Time with Pruning", + "text_level": 1, + "bbox": [ + 89, + 848, + 468, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We test the inference times of models compressed via pruning and our method when both achieved the same reduction", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c5c4c61098f9bbfdb4d42570d6aa05fd5ef7792961f3ab2ca554d55a8531ffb.jpg", + "image_caption": [ + "Figure 4. Comparison and compatibility with the pruning method. Bubble size represents the model's computational complexity. Combining our method with pruning enables further $1.3 \\times$ acceleration, demonstrating our compatibility with pruning." + ], + "image_footnote": [], + "bbox": [ + 531, + 85, + 883, + 265 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ffaebe61309119ac54aff898807ab835446f0777291a29cf7c7340320309276c.jpg", + "table_caption": [ + "Table 7. Comparisons of inference time between pruning methods and ours under the same computational complexity on THUMOS14, VideoMAE-S backbone. Reducing depth yields better acceleration than reducing width." + ], + "table_footnote": [], + "table_body": "
ModelMACs (G)Inference Time (ms)Speedup
Uncompressed286.3104.91.00×
Pruning218.296.61.09×
Ours218.081.21.29×
", + "bbox": [ + 517, + 386, + 901, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in MACs. The inference time is measured as the average duration per input sample, each consisting of 768 frames with a resolution of $160^2$ . As shown in Table 7, under similar MACs (218.2G vs 218.0G), pruning achieves a $1.09 \\times$ speedup, while the block drop method achieves a $1.29 \\times$ speedup. This indicates that compressing the model by reducing network depth is more effective for acceleration than reducing network width because shallow and wide models are more hardware-friendly architectures [39].", + "bbox": [ + 511, + 455, + 906, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 604, + 633, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have proposed a progressive block drop method to compress TAD models from the depth perspective. This method consists of two main steps: 1) The block selection evaluator automatically selects blocks to be dropped based on their importance; 2) Parameter-efficient learning combined with three alignment techniques is used to recover performance. On the THUMOS14 and ActivityNet-1.3 datasets, we have achieved a reduction of about $25\\%$ in computational cost while improving the pruned model's performance above that of the uncompressed model. We have further validated its generalizability across TAD architectures, fine-grained datasets, and other tasks, demonstrating compatibility with existing compression techniques. Additionally, we have demonstrated the compatibility of our method with other compression techniques through experiments. We hope that our work paves the way for further advancements in the compression and acceleration of TAD models.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29232", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was partially supported by National Natural Science Foundation of China (NSFC) under Grants 62202311, the Guangdong Basic and Applied Basic Research Foundation under Grants 2023A1515011512, Excellent Science and Technology Creative Talent Training Program of Shenzhen Municipality under Grant RCBS20221008093224017, the Shenzhen Natural Science Foundation (the Stable Support Plan Program) under Grant 20220809180405001, Key Scientific Research Project of the Department of Education of Guangdong Province 2024ZDZX3012.", + "bbox": [ + 89, + 90, + 485, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 273, + 187, + 289 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3", + "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning (ICML), pages 41-48, 2009. 5", + "[3] Shyamal Buch, Victor Escorcia, Chuanqi Shen, Bernard Ghanem, and Juan Carlos Niebles. Sst: Single-stream temporal action proposals. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 2911-2920, 2017. 2", + "[4] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 961-970, 2015. 2, 6", + "[5] Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A Ross, Jia Deng, and Rahul Sukthankar. Rethinking the faster r-cnn architecture for temporal action localization. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1130–1139, 2018. 2", + "[6] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024.8", + "[7] Shi Chen and Qi Zhao. Shallowing deep networks: Layerwise pruning based on feature representations. IEEE transactions on pattern analysis and machine intelligence, 41(12): 3048-3056, 2018. 3", + "[8] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fine-tuning of long-context large language models. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 5", + "[9] Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. Advances in" + ], + "bbox": [ + 99, + 299, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neural information processing systems (NeurIPS), 27, 2014. 3", + "[10] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning (ICML), pages 2793-2803. PMLR, 2021. 3", + "[11] Amir Ben Dror, Niv Zehngut, Avraham Raviv, Evgeny Artyomov, Ran Vitek, and Roy Jevnisek. Layer folding: Neural network depth reduction using activation linearization. arXiv preprint arXiv:2106.09309, 2021. 3", + "[12] Sara Elkerdawy, Mostafa Elhoushi, Abhineet Singh, Hong Zhang, and Nilanjan Ray. To filter prune, or to layer prune, that is the question. In proceedings of the Asian conference on computer vision (ACCV), 2020. 3", + "[13] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 16091-16101, 2023. 3, 8", + "[14] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations (ICLR), 2018. 1", + "[15] Yonggan Fu, Haichuan Yang, Jiayi Yuan, Meng Li, Cheng Wan, Raghuraman Krishnamoorthi, Vikas Chandra, and Yingyan Lin. Depthshrinker: a new compression paradigm towards boosting real-hardware efficiency of compact neural networks. In International Conference on Machine Learning (ICML), pages 6849-6862. PMLR, 2022. 3", + "[16] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 8", + "[17] Lianli Gao, Zhao Guo, Hanwang Zhang, Xing Xu, and Heng Tao Shen. Video captioning with attention-based LSTM and semantic consistency. IEEE Transactions on Multimedia, 19(9):2045-2055, 2017. 1", + "[18] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. In International Conference on Learning Representations (ICLR), 2018. 1", + "[19] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 3", + "[20] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems (NeurIPS), 28, 2015. 1, 3", + "[21] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 1389-1397, 2017. 3", + "[22] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29233", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 4340-4349, 2019. 3", + "[23] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. stat, 1050:9, 2015. 3", + "[24] Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. Journal of Machine Learning Research, 22(241):1-124, 2021. 1", + "[25] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 1314-1324, 2019. 3", + "[26] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2021. 2, 5", + "[27] Yupan Huang, Qi Dai, and Yutong Lu. Decoupling localization and classification in single shot temporal action detection. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1288-1293. IEEE, 2019. 3", + "[28] YG Jiang, J Liu, A Roshan Zamir, G Toderici, I Laptev, M Shah, and R Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. 1, 6", + "[29] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6", + "[30] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems (NeurIPS), 25, 2012. 3", + "[31] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1): 79-86, 1951. 6", + "[32] Mark Kurtz, Justin Kopinsky, Rati Gelashvili, Alexander Matveev, John Carr, Michael Goin, William Leiserson, Sage Moore, Nir Shavit, and Dan Alistarh. Inducing and exploiting activation sparsity for fast inference on deep neural networks. In International Conference on Machine Learning (ICML), pages 5533-5543. PMLR, 2020. 3", + "[33] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. Tvqa: Localized, compositional video question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1369-1379, 2018. 1", + "[34] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 3", + "[35] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 3320-3329, 2021. 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Tianwei Lin, Xu Zhao, and Zheng Shou. Single shot temporal action detection. In Proceedings of the 25th ACM international conference on Multimedia, pages 988-996, 2017. 3", + "[37] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 3889-3898, 2019. 2", + "[38] Ji Liu, Dehua Tang, Yuanxian Huang, Li Zhang, Xiaocheng Zeng, Dong Li, Mingjie Lu, Jinzhang Peng, Yu Wang, Fan Jiang, et al. Updp: A unified progressive depth pruner for cnn and vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13891-13899, 2024. 3", + "[39] Ning Liu, Xiaolong Ma, Zhiyuan Xu, Yanzhi Wang, Jian Tang, and Jieping Ye. Autocompress: An automatic dnn structured pruning framework for ultra-high compression rates. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4876-4883, 2020. 3, 8", + "[40] Shuming Liu, Chen-Lin Zhang, Chen Zhao, and Bernard Ghanem. End-to-end temporal action detection with 1b parameters across 1000 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18591-18601, 2024. 1, 3, 5, 6, 7, 8", + "[41] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20010-20019, 2022.", + "[42] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 3", + "[43] Yi Liu, Limin Wang, Yali Wang, Xiao Ma, and Yu Qiao. Fineaction: A fine-grained video dataset for temporal action localization. IEEE transactions on image processing, 31: 6937-6950, 2022. 6, 7, 8", + "[44] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. IEEE Robotics and Automation Letters, 2023. 1", + "[45] Xiaolong Ma, Sheng Lin, Shaokai Ye, Zhezhi He, Linfeng Zhang, Geng Yuan, Sia Huat Tan, Zhengang Li, Deliang Fan, Xuehai Qian, et al. Non-structured dnn weight pruning—is it beneficial in any platform? IEEE transactions on neural networks and learning systems, 33(9):4930–4944, 2021. 1", + "[46] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? Advances in neural information processing systems (NeurIPS), 32, 2019. 3", + "[47] Seyed Iman Mirzadeh, Keivan Alizadeh-Vahid, Sachin Mehta, Carlo C del Mundo, Oncel Tuzel, Golnoosh Samei, Mohammad Rastegari, and Mehrdad Farajtabar. Relu strikes back: Exploiting activation sparsity in large language models. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3", + "[48] P Molchanov, S Tyree, T Karras, T Aila, and J Kautz. Pruning convolutional neural networks for resource efficient in" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29234", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference. In 5th International Conference on Learning Representations, ICLR 2017-Conference Track Proceedings, 2019. 1", + "[49] Alberto Montes, Amaia Salvador Aguilera, Santiago Pascual, and Xavier Giró Nieto. Temporal activity detection in untrimmed videos with recurrent neural networks. In Proceedings of the 1st NIPS Workshop on Large Scale Computer Vision Systems (NeurIPS), pages 1-5, 2016. 2", + "[50] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Proposal-free temporal action detection via global segmentation mask learning. In European Conference on Computer Vision (ECCV), pages 645-662. Springer, 2022.", + "[51] AJ Piergiovanni and Michael Ryoo. Temporal gaussian mixture layer for videos. In International Conference on Machine learning (ICML), pages 5152-5161. PMLR, 2019. 2", + "[52] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 658-666, 2019. 6", + "[53] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18857-18866, 2023. 3", + "[54] Zheng Shou, Dongang Wang, and Shih-Fu Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1049-1058, 2016. 2", + "[55] Suraj Srinivas, Akshayvarun Subramanya, and R Venkatesh Babu. Training sparse neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR) workshops, pages 138-145, 2017. 3", + "[56] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 13526-13535, 2021. 3", + "[57] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2820-2828, 2019. 3", + "[58] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 3", + "[59] Marvin Teichmann, Michael Weber, Marius Zoellner, Roberto Cipolla, and Raquel Urtasun. Multinet: Real-time joint semantic reasoning for autonomous driving. In 2018 IEEE intelligent vehicles symposium (IV), pages 1013-1020. IEEE, 2018. 1", + "[60] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for self-supervised video pre-training. Advances in neural information processing systems (NeurIPS), 35:10078-10093, 2022. 6", + "[61] Binglu Wang, Yongqiang Zhao, Le Yang, Teng Long, and Xuelong Li. Temporal action localization in the deep learning era: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[62] Lining Wang, Haosen Yang, Wenhao Wu, Hongxun Yao, and Hujie Huang. Temporal action proposal generation with transformers. arXiv preprint arXiv:2105.12043, 2021. 3", + "[63] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555-4576, 2021. 5", + "[64] Huijuan Xu, Abir Das, and Kate Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 5783-5792, 2017. 2", + "[65] Pengtao Xu, Jian Cao, Fanhua Shang, Wenyu Sun, and Pu Li. Layer pruning via fusible residual convolutional block for deep neural networks. arXiv preprint arXiv:2011.14356, 2020. 3", + "[66] Le Yang, Houwen Peng, Dingwen Zhang, Jianlong Fu, and Junwei Han. Revisiting anchor mechanisms for temporal action localization. IEEE Transactions on Image Processing, 29:8535-8548, 2020. 3", + "[67] Min Yang, Guo Chen, Yin-Dong Zheng, Tong Lu, and Limin Wang. Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding, 232:103692, 2023. 1, 3", + "[68] Serena Yeung, Olga Russakovsky, Greg Mori, and Li Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 2678–2687, 2016. 3", + "[69] Runhao Zeng, Wenbing Huang, Mingkui Tan, Yu Rong, Peilin Zhao, Junzhou Huang, and Chuang Gan. Graph convolutional module for temporal action localization in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6209-6223, 2021. 2", + "[70] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision (ECCV), pages 492-510. Springer, 2022. 1, 3, 6, 7, 8", + "[71] Minjia Zhang and Yuxiong He. Accelerating training of transformer-based language models with progressive layer dropping. Advances in neural information processing systems (NeurIPS), 33:14011-14023, 2020. 3", + "[72] Chen Zhao, Shuming Liu, Karttikeya Mangalam, and Bernard Ghanem. Re2tal: Rewiring pretrained video backbones for reversible temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10637-10647, 2023. 3", + "[73] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 2914-2923, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29235", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[74] Yao Zhou, Gary G Yen, and Zhang Yi. Evolutionary shallowing deep neural networks at block levels. IEEE Transactions on Neural Networks and Learning Systems, 33(9): 4635-4647, 2021. 3", + "bbox": [ + 91, + 90, + 486, + 147 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "29236", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Temporal Action Detection Model Compression by Progressive Block Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_model.json b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56266e6163561c2db3080875fd270c160c7edcec --- /dev/null +++ b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_model.json @@ -0,0 +1,2457 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.131, + 0.877, + 0.154 + ], + "angle": 0, + "content": "Temporal Action Detection Model Compression by Progressive Block Drop" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.918, + 0.252 + ], + "angle": 0, + "content": "Xiaoyong Chen\\(^{1,2*}\\), Yong Guo\\(^{3*}\\), Jiaming Liang\\(^{2}\\), Sitong Zhuang\\(^{2}\\), Runhao Zeng\\(^{1,2\\dagger}\\), Xiping Hu\\(^{1,4\\dagger}\\) \n\\(^{1}\\)Guangdong-Hong Kong-Macao Joint Laboratory for Emotional Intelligence and Pervasive Computing, Shenzhen MSU-BIT University, \\(^{2}\\)Shenzhen University, \\(^{3}\\)South China University of Technology, \\(^{4}\\)Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.255, + 0.785, + 0.269 + ], + "angle": 0, + "content": "xiaoyongchencs@gmail.com, zengrh@smbu.edu.cn, huxp@smbu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.338, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Temporal action detection (TAD) aims to identify and localize action instances in untrimmed videos, which is essential for various video understanding tasks. However, recent improvements in model performance, driven by larger feature extractors and datasets, have led to increased computational demands. This presents a challenge for applications like autonomous driving and robotics, which rely on limited computational resources. While existing channel pruning methods can compress these models, reducing the number of channels often hinders the parallelization efficiency of GPU, due to the inefficient multiplication between small matrices. Instead of pruning channels, we propose a Progressive Block Drop method that reduces model depth while retaining layer width. In this way, we still use large matrices for computation but reduce the number of multiplications. Our approach iteratively removes redundant blocks in two steps: first, we drop blocks with minimal impact on model performance; and second, we employ a parameter-efficient cross-depth alignment technique, fine-tuning the pruned model to restore model accuracy. Our method achieves a \\(25\\%\\) reduction in computational overhead on two TAD benchmarks (THUMOS14 and ActivityNet-1.3) to achieve lossless compression. More critically, we empirically show that our method is orthogonal to channel pruning methods and can be combined with it to yield further efficiency gains." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.761, + 0.222, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Temporal action detection aims to identify action instances and localize their start and end times in untrimmed videos. This task underpins many video understanding applications, including video question answering [33] and video captioning [17]. Recent advancements in larger feature extractors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.905, + 0.396 + ], + "angle": 0, + "content": "and datasets have improved model performance but also significantly increased computational demands [40]. However, many TAD applications, such as autonomous driving [59] and robotics [44], operate under limited computational resources. Thus, reducing computational complexity is crucial for deploying TAD in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.905, + 0.565 + ], + "angle": 0, + "content": "A typical TAD model [67, 70] consists of a feature extractor and an action detection head. Due to the long duration of videos, the feature extractor processes video segments individually, and these segment-level features are then combined before being passed to the detection head. As shown in Figure 1, our experiments show that the feature extraction phase accounts for \\(95\\%\\) of the total model computation, since it processes many segments, while the detection head processes only a single video-level feature. Thus, reducing the complexity of the feature extractor is key to improving TAD model efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.905, + 0.747 + ], + "angle": 0, + "content": "Pruning [14, 18, 20, 48] is a common model compression technique, which removes redundant weights based on their importance. This results in a slim-and-tall network structure, where the model's hidden layer dimensions are reduced but depth remains unchanged, as illustrated in Figure 1. However, studies have shown [24, 45] that for the same number of parameters, a deep-and-thin network may have slower inference speeds compared to a shallow-and-wide network. The reason for this is that, during GPU inference, operations on larger matrices tend to have higher parallelism and are thus more efficient than multiple operations on smaller matrices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To explore the feasibility of reducing the architectural depth of TAD models, we conduct preliminary experiments on the THUMOS14 dataset [28], which is widely used for this task. We raise the following key observations from Figure 2: 1) There is a slight discrepancy between the input and output features of certain blocks. 2) Dropping specific blocks results in only a minimal decrease in model performance. 3) Simultaneously dropping multiple blocks leads to a more significant performance drop (Table 3). These results motivate us to investigate TAD model compression" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.22, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29225" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.088, + 0.891, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.323, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Figure 1. Comparisons between reducing depth and reducing width. For a 768-frame video sequence with a resolution of \\(160^2\\), using VideoMAE-S as the feature extractor and Actionformer as the detection head, \\(95\\%\\) of the computational overhead comes from the feature extractor. Most pruning methods reduce the size of layer weight matrices, leading to a slim-and-tall network structure. In contrast, our progressive block drop method reduces network depth, achieving \\(1.19 \\times\\) faster inference at the same computational cost. These findings, detailed in Section 5.6, suggest that our approach results in a more hardware-efficient model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.406, + 0.275, + 0.421 + ], + "angle": 0, + "content": "from the depth perspective." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.427, + 0.485, + 0.804 + ], + "angle": 0, + "content": "In this paper, we propose a multi-step progressive block drop method that effectively reduces network depth, transforming the TAD model into a shallow-and-wide architecture more suitable for parallel GPU matrix computations. Specifically, we compress the model progressively through multiple iterations. Each iteration consists of two stages: 1) block selection and removal, where we evaluate the importance of each block based on a specific metric and remove the least important block; and 2) performance restoration, where we design a parameter-efficient cross-depth alignment strategy. This strategy first integrates LoRA [26] into the attention modules of the remaining blocks and then fine-tunes the pruned model to align with the intermediate features output by the uncompressed model's blocks, effectively compressing the information from multiple blocks into a single block. Additionally, we employ prediction-level alignment to further improve performance. On the THUMOS14 and ActivityNet datasets [4], our method achieves superior performance compared to the uncompressed model, while reducing the model's computational cost (MACs) by \\(25\\%\\). Furthermore, we demonstrate through experiments that our approach is compatible with common pruning techniques, enabling even further compression of the model. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.81, + 0.486, + 0.902 + ], + "angle": 0, + "content": "- We focus on compressing the TAD model from the perspective of depth. Unlike channel pruning methods that compress the model from the width or the internal layers of blocks, we analyze the redundancy within the TAD model and explore block-level compression as a novel approach. This provides a new angle for compressing and" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.406, + 0.722, + 0.42 + ], + "angle": 0, + "content": "accelerating the TAD model." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.421, + 0.907, + 0.572 + ], + "angle": 0, + "content": "- To achieve depth compression, we propose a progressive multi-step block drop strategy. In each iteration, a block is automatically removed, and a parameter-efficient training technique is employed to recover the model's performance. On THUMOS14 and ActivityNet, our method reduces the model's MACs by approximately \\(25\\%\\), while simultaneously outperforming the uncompressed model. Furthermore, we extend our method to other TAD architectures, fine-grained datasets, and additional tasks to validate its generalizability." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.572, + 0.907, + 0.633 + ], + "angle": 0, + "content": "- We empirically demonstrate that our method is orthogonal to channel pruning techniques and investigate its compatibility with other model compression approaches, resulting in additional efficiency improvements." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.421, + 0.907, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.653, + 0.657, + 0.669 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.757, + 0.696 + ], + "angle": 0, + "content": "2.1. Temporal Action Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Temporal action detection (TAD) can generally be divided into two primary categories [61]. The first category, two-stage methods, involves generating action proposals followed by classification and boundary refinement [5, 54, 64, 69, 73]. Proposal generation techniques may employ frame-or segment-level classification with subsequent merging of frames or segments belonging to the same action class [49-51], while others use specialized proposal generation strategies [3, 37]. However, the success of these approaches heavily relies on the quality of the proposals, which has led to the emergence of integrated approaches, also known as one-stage methods. These methods simultaneously handle proposal generation, classification, and/or boundary regres" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29226" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.168 + ], + "angle": 0, + "content": "sion [27, 36, 58, 68]. Key developments in this area include the adoption of anchor-based mechanisms for TAD [36] and the shift towards anchor-free frameworks [35, 53], with recent efforts combining the benefits of both anchor-based and anchor-free approaches [66]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.381 + ], + "angle": 0, + "content": "In recent years, Transformer-based architectures, which have demonstrated remarkable results in various computer vision tasks, have been applied to TAD [56, 62, 70]. Additionally, the development of end-to-end architectures has been explored in various studies [40-42, 67, 72]. With the significant achievements of scaling up in large language models [1], recent research [40] has utilized scaling up in conjunction with a frame input data processing approach. This combination has not only significantly improved model performance but also reduced computational costs. However, the issue of increased inference time due to computational scaling has become increasingly critical. Therefore, we focus on reducing inference latency through model compression." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.393, + 0.282, + 0.41 + ], + "angle": 0, + "content": "2.2. Model Compression" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Model compression techniques are typically categorized into pruning [20], knowledge distillation [23], model decomposition [9], and quantization [30]. This paper focuses on model pruning, which can be either unstructured [20] or structured [34]. Unstructured pruning removes network connections to create sparsity, but it requires specialized hardware for sparse matrix operations [19, 20, 55], limiting its practical deployment. Structured pruning, which includes width-reduction methods like channel pruning [21, 34] and filter pruning [22], results in a dense weight matrix, enabling GPU acceleration. However, it may not fully exploit GPU parallelism [39]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Depth-wise pruning techniques [25, 57] have been used for model acceleration, although they may suffer from insufficient memory occupancy. Various methods, including layer importance evaluation [7, 12], parameter optimization [65, 74], and reparameterization [11, 15], can also guide pruning to compress the model from the perspective of depth. Previous work [38, 71] has explored the concept of progressive block dropping; however, there are significant differences between their approaches and ours. Specifically, [38] focuses on compressing and integrating the layers within the block, which is a layer-level pruning technique. The dynamically skipping blocks method proposed in [71] accelerates training by dropping blocks, but during inference, the full model is still employed. Additionally, skipping blocks to reduce the depth of Transformer models [10, 46] often leads to a loss in accuracy. In contrast, our progressive drop block method reduces depth at the block level, improving both inference speed and detection accuracy. Our method can be combined with structured pruning [13] and unstructured activation sparsity [32, 47] to" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.906, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.258, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 2. Analysis of TAD models at block level on THUMOS14 datasets, using VideoMAE-S. Blue curve: impact of dropping a single block on the model's detection accuracy, where most cases result in minimal accuracy degradation. Green line: MSE of the input and output features for each block. Some blocks exhibit MSE values close to 0, indicating that these blocks are redundant." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.355, + 0.906, + 0.387 + ], + "angle": 0, + "content": "further compress the model and accelerate inference while improving performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.4, + 0.785, + 0.416 + ], + "angle": 0, + "content": "3. Preliminaries and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.906, + 0.502 + ], + "angle": 0, + "content": "In this section, we first introduce the notations and formulations pertinent to temporal action detection (TAD) and then explore the feasibility of reducing the depth of TAD models, as well as some key findings that inspire the design of subsequent methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.512, + 0.794, + 0.527 + ], + "angle": 0, + "content": "3.1. TAD Formulation and Notation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.906, + 0.642 + ], + "angle": 0, + "content": "TAD aims to identify action instances within a video and determine their temporal boundaries. Given an untrimmed video represented as \\( V = \\{I_t\\}_{t=1}^T \\), where \\( I_t \\) denotes the frame at time \\( t \\), TAD predicts a set of action instances \\( \\Phi_V = \\{\\phi_i = (t_{s_i}, t_{e_i}, c_i)\\}_{i=1}^N \\). Here, \\( N \\) is the number of action instances; for each instance \\( \\phi_i \\), \\( t_{s_i} \\) and \\( t_{e_i} \\) are its starting and ending times, respectively, and \\( c_i \\) denotes its category." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.882, + 0.669 + ], + "angle": 0, + "content": "3.2. Feasibility of Reducing TAD model's Depth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.718 + ], + "angle": 0, + "content": "We find that there is a severe redundancy in the depth of TAD models. To verify this, we conduct a series of experiments and summarize the key observations as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Observation 1: Most computational overhead lies in the clip-level feature extraction backbone. TAD methods typically consist of two main parts: the feature extractor and the detection head. In the feature extraction stage, a sliding window of length \\( L \\) is processed by a video feature extractor. For instance, features are extracted from frames \\( \\{I_t, \\dots, I_{t + L}\\} \\), and then the window shifts by \\( s \\) frames to extract features from frames \\( \\{I_{t + s}, \\dots, I_{t + s + L}\\} \\). As the feature extraction process is repeated across the entire video sequence, it incurs substantial computational cost, with each window requiring complex operations to extract features. Once the features are extracted, they are con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29227" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.904, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.442, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 3. The diagram of our progressive block drop method. Our approach adopts a multi-step progressive compression strategy. At each iteration 1) we evaluate the importance of each block and drop the least important block, and 2) we use parameter-efficient tuning techniques, and recover performance by learning from the uncompressed model through feature-level and prediction-level alignment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.496, + 0.483, + 0.693 + ], + "angle": 0, + "content": "catenated along the temporal dimension, and information is fused across the spatial dimension by averaging the features. These fused features are then passed to the detection head. Ultimately, the volume of data fed into the detection head is much smaller compared to the data processed by the feature extractor. As shown in Figure 1, the computational cost of the feature extractor accounts for \\(95\\%\\) of the total forward inference cost of the model. Given the high computational cost of feature extraction, optimizing or reducing the complexity of this stage presents a significant opportunity for improving model efficiency. Therefore, to reduce the overall model computation, the primary focus should be on compressing the feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.699, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Observation 2: There exist blocks that have a minor discrepancy between their input and output. Additionally, by analyzing the feature differences between consecutive blocks, as depicted in Figure 2, we find that some blocks exhibit very similar representation capabilities, suggesting redundancy in the model. These findings highlight the feasibility of applying the block drop method to reduce computational complexity without sacrificing performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Observation 3: The removal of certain blocks has a negligible impact on TAD performance. To investigate the potential for model compression, we first assess the feasibility of dropping blocks by examining the impact of removing individual blocks on model performance. As shown" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.542 + ], + "angle": 0, + "content": "in Figure 2, our experiments reveal that removing certain blocks led to only minimal performance degradation, indicating that block drop could be a viable approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.909, + 0.756 + ], + "angle": 0, + "content": "Observation 4: Simultaneously removing multiple blocks at a time results in a significant performance drop. In our preliminary exploration, we attempt to directly drop multiple blocks and train the model to observe whether it could recover its performance. We randomly select three blocks to drop all at once and then attempt fine-tuning. As shown in Table 3, the best performance achieved is \\(0.67\\%\\) lower than the uncompressed model, exhibiting a noticeable performance drop, which is an undesirable outcome in practical applications. Therefore, we adopt a multi-step progressive block dropping strategy rather than removing multiple blocks at once. This approach allows the information from the removed blocks to be more effectively compressed into the pruned model, resulting in better performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.737, + 0.788 + ], + "angle": 0, + "content": "4. Progressive Block Drop" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we design a compression method specifically tailored for TAD models. Drawing from our analysis in Section 3.2, our focus is on reducing the computational cost of the feature extraction backbone while maintaining or even enhancing performance. To achieve this, we propose a multi-step progressive block-level dropping method, where each step comprises two stages. First, the Block Selection" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29228" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "Evaluator automatically selects the blocks to be dropped, resulting in a subnet. Then, Cross-Depth Alignment enables parameter-efficient learning from the uncompressed model to maintain TAD performance. By iteratively applying these steps, we progressively obtain a pruned TAD backbone that can be directly deployed for inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.191, + 0.327, + 0.206 + ], + "angle": 0, + "content": "4.1. Block Selection Evaluator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.484, + 0.333 + ], + "angle": 0, + "content": "Experimental results presented in Section 3.2 demonstrate that simultaneously removing multiple blocks significantly degrades the performance of TAD. Inspired by prior work on curriculum learning [2, 63], we propose a progressive block-drop strategy. However, selecting the appropriate blocks to drop is a non-trivial task. To address this challenge, we introduce a block selection evaluator \\( f_{E} \\) to select the blocks to be dropped automatically." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Without loss of generality, let \\( M_0 \\) denote the initial uncompressed model trained on the training set, and let \\( M_{t-1} \\) represent the model input to \\( f_E \\) at iteration \\( t \\). We denote \\( B_{t-1} = \\{b_1, b_2, \\ldots, b_K\\} \\) as the set of all blocks within \\( M_{t-1} \\), where \\( K \\) is the total number of blocks. At iteration \\( t \\), the collection of sub-networks resulting from dropping each single block is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.448, + 0.483, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {t} = \\left\\{M _ {t, k} ^ {\\text {s u b}} = M _ {t - 1} \\backslash b _ {k} \\mid b _ {k} \\in B _ {t - 1} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.581 + ], + "angle": 0, + "content": "To automatically select blocks for pruning, we employ the evaluation function \\( f_{E} \\) to assess each subnet. By measuring the performance difference between each subnet and the uncompressed TAD model \\( M_0 \\) on the training set, \\( f_{E} \\) outputs an importance score for each subnet; the smaller the performance difference, the higher the importance score assigned. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.589, + 0.483, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {t} = f _ {E} \\left(\\mathcal {S} _ {t}, D _ {\\text {t r a i n}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.751 + ], + "angle": 0, + "content": "From the set of importance scores \\(\\mathcal{P}_t\\), we select the subnet \\(M_{t,*}^{sub}\\) with the highest importance score \\(p^*\\), resulting in the pruned model \\(M_t\\) at step \\(t\\). Note that \\(f_{E}\\) can also compute the difference between the input and output features of each removed block or the training loss of each subnet on the training set; in both cases, a smaller difference corresponds to a higher importance score. In our experiments (Section F.1 of the appendix), we find that selecting subnets based on the mAP metric from the training set yields the best results." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.484, + 0.79 + ], + "angle": 0, + "content": "4.2. Parameter-Efficient Learning from Uncompressed Model for Performance Recovery" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "After removing one block from the TAD model, a central challenge arises in ensuring that the performance of the reduced-depth model remains comparable to that of the original, uncompressed model (see Section F.2 of appendix). However, training the TAD model on untrimmed videos using full-parameter fine-tuning still demands a substantial amount of GPU memory, as discussed in [40]. For" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.091, + 0.884, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Proposed Progressive Block Drop Method" + }, + { + "type": "algorithm", + "bbox": [ + 0.516, + 0.11, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Input: Uncompressed model \\(M_0\\), model \\(M_{t-1}\\) at step \\(t\\), blocks \\(B_{t-1} = \\{b_1, b_2, \\ldots, b_K\\}\\) within \\(M_{t-1}\\), training set \\(D_{train}\\). // Progressive block drop cycle while Performance of \\(M_{t-1} \\geq M_0\\) do // Stage 1: select one block to drops Obtain sub-nets by dropping each block: \\(\\mathcal{S}_t = \\left\\{M_{t,k}^{\\mathrm{sub}} = M_{t-1} \\setminus b_k \\mid b_k \\in B_{t-1}\\right\\}\\) Evaluate each sub-net using the evaluator \\(f_E\\): \\(\\mathcal{P}_t = f_E(\\mathcal{S}_t, D_{train})\\) Select the subnet with the highest performance: \\(M_t = \\arg \\max(\\mathcal{P}_t)\\) // Stage 2: performance recovery Insert LoRA modules into \\(M_t\\) for training: \\(M_t(\\theta) \\rightarrow \\hat{M}_t(\\theta, \\theta_{\\mathrm{LoRA}})\\) Train the model with Eqn. (6) Merge LoRA parameters into the model: \\(\\theta_t \\gets \\theta_t + \\theta_{\\mathrm{LoRA}}'\\) end while" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.508 + ], + "angle": 0, + "content": "example, in our experiments with the VideoMAE-L model we used, setting the batch size to 1 already consumes 13.1 GB of memory. Therefore, we introduce LoRA to reduce memory consumption during training, with detailed experimental results provided in Section F.3 of the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.659 + ], + "angle": 0, + "content": "Before training, given the pruned model \\( M_{t}(\\theta_{t}) \\) at step \\( t \\), where \\( \\theta_{t} \\) represents its parameters, we first follow the LoRA method [8, 26] to add trainable parameters \\( \\theta_{\\mathrm{LORA}} \\) to the attention layer each block, resulting in the model \\( \\hat{M}_{t}(\\theta_{t},\\theta_{\\mathrm{LORA}}) \\). This approach offers two main benefits: first, during subsequent training, only \\( \\theta_{\\mathrm{LORA}} \\) is trained, significantly reducing memory consumption; second, after training, \\( \\theta_{\\mathrm{LORA}} \\) can be merged with \\( \\theta_{t} \\), thereby keeping the TAD model's structure unchanged and incurring no additional inference computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Cross-Depth Alignment. The structural differences between the pruned and uncompressed models—resulting from block dropping—pose a significant challenge for recovering performance. More critically, we empirically observed that directly supervising the training of \\(\\hat{M}_t\\) with ground truth fails to restore performance (see Section E.5). In other words, relying solely on prediction-level supervision is inadequate for restoring performance. To address this issue, we shift our attention to feature-level supervision and propose a cross-depth alignment strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Specifically, if we consider three adjacent blocks, \\(\\{b_{i - 1}, b_i, b_{i + 1}\\}\\), as a group, when \\(b_i\\) is dropped, the input to \\(b_{i + 1}\\) changes (i.e., it becomes the output of \\(b_{i - 1}\\)). To maintain performance, it is necessary to: 1) ensure that the input feature distribution of \\(b_{i + 1}^p\\) in the pruned model \\(\\hat{M}_t\\) aligns with that of \\(b_i^u\\) in the uncompressed model \\(M_0\\); and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "29229" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "2) ensure that the output distribution of the \\((i + 1)\\)-th block in \\(\\hat{M}_t\\) aligns with that of \\(M_0\\). In our implementation, after the \\(i\\)-th block is dropped, we align the outputs of the \\((i - 1)\\)-th and \\((i + 1)\\)-th blocks \\((f_{b_{i - 1}}\\) and \\(f_{b_{i + 1}})\\) between the pruned and uncompressed models, thereby compressing the information from multiple blocks into a single block. For the features of other blocks, we similarly apply block-wise alignment with the \\(M_0\\). The cross-depth alignment loss can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.235, + 0.484, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = \\frac {1}{I - 1} \\sum_ {m = 1, m \\neq i} ^ {I} \\left(f _ {b _ {m}} ^ {M _ {0}} - f _ {b _ {m}} ^ {\\hat {M} _ {t}}\\right) ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.484, + 0.358 + ], + "angle": 0, + "content": "Action classification and Localization Alignment. The TAD model generates two distinct types of predictions: action categories and temporal bounding boxes. We also align the predictions of the pruned model \\(\\hat{M}_t\\) with those of the uncompressed model \\(M_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.484, + 0.448 + ], + "angle": 0, + "content": "1) Action classification alignment: To align the action class predictions, we employ the Kullback-Leibler (KL) divergence loss \\(\\mathcal{L}_{\\mathrm{KL}}\\) [31], which measures the divergence between the action class scores of the uncompressed model \\(M_0\\) and the pruned model \\(\\hat{M}_t\\). The alignment loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.45, + 0.482, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p c} = \\mathcal {L} _ {\\mathrm {K L}} \\left(p \\left(z _ {M _ {0}}\\right), p \\left(z _ {\\hat {M} _ {t}}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.528 + ], + "angle": 0, + "content": "where \\( z_{M_0} \\) and \\( z_{\\hat{M}_t} \\) denote the logits of the uncompressed and pruned models, respectively, and \\( p(z_{M_0}) \\) and \\( p(z_{\\hat{M}_t}) \\) represent the class probabilities derived from the logits via the softmax function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.621 + ], + "angle": 0, + "content": "2) Boundary regression alignment: For aligning the boundary predictions, we compute the Generalized Intersection over Union (GIoU) loss \\(\\mathcal{L}_{\\mathrm{GIoU}}\\) [52] between the predicted temporal bounding boxes of the uncompressed model \\(M_0\\) and the pruned model \\(\\hat{M}_t\\). The alignment loss for boundary regression is:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.628, + 0.483, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\mathcal {L} _ {\\mathrm {G I o U}} \\left(D _ {i} ^ {M _ {0}}, D _ {i} ^ {\\hat {M} _ {t}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.771 + ], + "angle": 0, + "content": "where \\(N_{p}\\) denotes the number of predicted positive temporal bounding boxes, and \\(D_{i}^{M_{0}}\\) and \\(D_{i}^{\\hat{M}_{t}}\\) represent the \\(i\\)-th predicted temporal bounding boxes for models \\(M_0\\) and \\(\\hat{M}_t\\), respectively. This dual alignment ensures that the pruned model learns both the correct action classification and the accurate temporal boundaries, which are crucial for TAD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.817 + ], + "angle": 0, + "content": "Total Loss Function. We incorporate the alignment losses, along with the commonly used classification loss \\(\\mathcal{L}_{\\mathrm{cls}}\\) and regression loss \\(\\mathcal{L}_{\\mathrm{reg}}\\) in TAD [40], as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.824, + 0.483, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {p c} + \\mathcal {L} _ {p r} + \\mathcal {L} _ {f} + \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.484, + 0.878 + ], + "angle": 0, + "content": "After training, the LoRA parameters \\(\\theta_{\\mathrm{LoRA}}^{\\prime}\\) are integrated with the pruned model as:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.886, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} \\leftarrow \\theta_ {t} + \\theta_ {\\mathrm {L o R A}} ^ {\\prime}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "The updated model \\( M_{t}(\\theta_{t}) \\) is then input into the block drop selection evaluator at the \\( t + 1 \\) step for the next iteration, continuing until the performance of the pruned model is no longer recoverable to that of the uncompressed model. Ultimately, the compressed TAD model is deployed, offering a reduced number of parameters and faster inference speed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.199, + 0.646, + 0.216 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.225, + 0.906, + 0.301 + ], + "angle": 0, + "content": "In the experimental section, we conduct the following experiments: 1) Sections 5.3 and 5.4 evaluate the effectiveness of the proposed method; 2) Section 5.5 validates compatibility with other compression models; 3) Section 5.6 compares inference speed with pruning methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.312, + 0.795, + 0.327 + ], + "angle": 0, + "content": "5.1. Datasets and Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.906, + 0.426 + ], + "angle": 0, + "content": "THUMOS14 [28] is a prominent benchmark dataset in TAD research. It consists of a training set with 200 videos and a test set with 213 videos, covering 20 action categories. This task is notably challenging due to the presence of more than 15 separate action instances per video, with background scenes occupying \\(71\\%\\) of the frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.906, + 0.517 + ], + "angle": 0, + "content": "ActivityNet-1.3 [4] is a large-scale benchmark for TAD in untrimmed videos, containing approximately 20,000 videos across 200 activity classes, totaling over 600 hours of content. The dataset is divided into three subsets: 10,024 videos for training, 4,926 for validation, and 5,044 for testing, with an average of 1.65 action instances per video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.608 + ], + "angle": 0, + "content": "FineAction [43] is a benchmark dataset for temporal action localization, containing 103K instances across 106 action categories in 17K untrimmed videos. It presents both opportunities and challenges for TAD due to its fine-grained action classes, dense annotations, and frequent co-occurrence of multiple action classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.76 + ], + "angle": 0, + "content": "Mean Average Precision (mAP). A predicted temporal bounding box is deemed accurate if its temporal Intersection over Union (tIoU) with the corresponding ground truth instance exceeds a specific threshold and matches the instance's category. For THUMOS14, the tIoU thresholds range from \\(\\{0.3, 0.4, 0.5, 0.6, 0.7\\}\\), with mAP@tIoU=0.5 reported for comparison. In contrast, on ActivityNet-1.3 and FineAction, tIoU thresholds are sampled from \\(\\{0.5: 0.05: 0.95\\}\\), and we report the mean mAP over tIoU thresholds from 0.5 to 0.95, in increments of 0.05." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.731, + 0.789 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We use the ActionFormer [70] detector for prediction, employing two Transformer-based feature extractors: VideoMAE-S with 12 blocks and VideoMAE-L [60] with 24 blocks, both are pre-trained on the K400 dataset [29]. For the THUMOS14 dataset, the input videos are randomly truncated to a fixed length of 768 frames, while for the ActivityNet-1.3 and FineAction dataset, the input videos are" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29230" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Evaluation of our method in terms of computational complexity, inference time and mAP with VideoMAE-S as the backbone. Our approach reduces the computational cost by approximately \\(25\\%\\), achieves around \\(1.3\\mathrm{x}\\) inference speedup, and improves the performance." + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.122, + 0.838, + 0.212 + ], + "angle": 0, + "content": "
Model (Drop #Block)THUMOS14ActivityNet-1.3
MACs (G)Inference Time (ms)mAP (tIoU=0.5)MACs (G)Inference Time (ms)mAP (Avg)
Baseline286.3104.970.43274.1103.837.75
1263.5 (92.0%)98.4 (93.8%)71.06 (↑ 0.63)251.3 (91.7%)95.1 (91.6%)37.94 (↑ 0.19)
2240.8 (84.1%)89.8 (85.6%)71.37 (↑ 0.94)228.5 (83.4%)86.8 (83.6%)37.81 (↑ 0.06)
3218.0 (76.1%)81.2 (77.4%)70.47 (↑ 0.04)205.8 (75.1%)79.6 (76.7%)37.77 (↑ 0.02)
4195.2 (68.2%)73.6 (70.2%)69.65 (↓ 0.78)183.0 (66.8%)72.1 (69.5%)37.72 (↓ 0.03)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.223, + 0.483, + 0.278 + ], + "angle": 0, + "content": "Table 2. Evaluation of our method on deeper networks in terms of computational complexity, inference time and mAP with VideoMAE-L (24 blocks) as the backbone. Our method is generally effective and larger models may benefit more from it." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.279, + 0.481, + 0.381 + ], + "angle": 0, + "content": "
Drop #BlockMACs (G)Inference Time (ms)mAP (tIoU=0.5)
Baseline3886.9795.976.01
13725.5 (95.8%)765.4 (96.2%)77.97 (↑ 1.96)
23564.1 (91.7%)731.4 (91.9%)78.00 (↑ 1.99)
33402.7 (87.5%)704.3 (88.5%)78.29 (↑ 2.28)
43241.4 (83.4%)671.3 (84.3%)77.86 (↑ 1.85)
53080.0 (79.2%)647.1 (81.3%)77.57 (↑ 1.56)
62918.6 (75.1%)602.6 (75.7%)77.25 (↑ 1.24)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.434 + ], + "angle": 0, + "content": "resized to a fixed length of 768 frames, with the frame resolution set to \\(160^2\\) by default for both datasets. Please refer to the supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.441, + 0.468, + 0.459 + ], + "angle": 0, + "content": "5.3. Effectiveness of Our Progressive Block Drop" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.483, + 0.705 + ], + "angle": 0, + "content": "We compare the model trained with the progressive block drop method to the uncompressed model from three aspects. 1) Computational Complexity (MACs): The computational complexity is calculated based on the number of multiply-accumulate operations at each layer of the model. 2) Inference Time: The inference time measures the average duration for processing each input sample, where each sample consists of 768 frames with a resolution of \\(160^{2}\\). 3) mAP: The results, as shown in Table 1, demonstrate that on both datasets, our method achieves an approximately \\(25\\%\\) reduction in MACs while outperforming the uncompressed model in terms of performance. This highlights the superiority of our approach, as it effectively reduces computational complexity while simultaneously enhancing the performance of the compressed model—an accomplishment that is difficult for other compression methods to achieve." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Drop blocks on a deeper backbone. We explore the performance of the block drop method when the depth of the feature extractor is increased from 12 to 24, i.e., when transitioning from VideoMAE-S to VideoMAE-L. The results, as shown in Table 2, reveal that with a \\(25\\%\\) reduction in MACs, the model achieves a \\(1.24\\%\\) performance improvement. Notably, when the same fraction of blocks (1/4) is dropped, the performance of VideoMAE-L improves by \\(1.24\\%\\), whereas VideoMAE-S only shows a modest improvement of \\(0.04\\%\\) (as shown in Table 1). This suggests that in models with higher computational complexity, there is more redundant information, and the progressive block drop method is effective at eliminating such redundancy." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.222, + 0.907, + 0.264 + ], + "angle": 0, + "content": "Table 3. Comparison with randomly dropping blocks. Dropping three randomly selected blocks at once performs worse than progressively dropping." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.297 + ], + "angle": 0, + "content": "
Drop Block ID[2,5,9][0,4,7][1,3,8][6,7,10][6,7,10] + ours
mAP68.8868.3169.7668.9170.47
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Dropping multiple blocks at one time vs Progressive dropping. To demonstrate the necessity of adopting a multi-step progressive drop in our proposed block drop method, we conduct experiments where multiple blocks are dropped simultaneously. Specifically, we randomly select three different blocks to drop and train the model using the training method described in Section 4.2. The results, as shown in Table 3, indicate that directly dropping three blocks yields performance ranging from \\(68.31\\%\\) to \\(69.76\\%\\), with the best result still \\(0.67\\%\\) lower than that of the uncompressed model. Meanwhile, using the progressive block drop method on VideoMAE-S and the THUMOS14 dataset, the model's performance after dropping the first three layers [7, 10, 6] reaches \\(70.47\\%\\). Comparing this with the result of directly dropping these three layers, it is evident that compressing the model via progressive dropping achieves better performance recovery (\\(70.47\\%\\) vs. \\(68.91\\%\\)), further underscoring the importance of the multi-step progressive drop. Therefore, our proposed multi-step progressive block drop is crucial for performance recovery." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.72, + 0.637 + ], + "angle": 0, + "content": "5.4. Extended Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Evaluation on additional TAD architectures. To validate the generalizability of our method, we apply it to two additional TAD architectures: the single-stage AdaTAD [40] model with a VideoMAE-S backbone and the two-stage ActionFormer [70] model with pre-extracted I3D features. Table 4 shows that in AdaTAD, pruning two blocks improves accuracy by \\(0.08\\%\\) and reduces MACs to \\(84.0\\%\\). Similarly, the ActionFormer model gains \\(0.49\\%\\) in mAP while reducing MACs to \\(59.7\\%\\). These results underscore our method's effectiveness in compressing diverse TAD architectures while enhancing performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation on the fine-grained FineActions dataset. To further corroborate the versatility of our approach, we conduct experiments on the fine-grained FineActions [43] dataset using a VideoMAE-S backbone with an Action-Former [70] head. As shown in Table 5, pruning two blocks slightly improves mAP (\\( \\uparrow 0.04\\% \\)), while pruning" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "29231" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 4. Evaluation of our method on AdaTAD and ActionFormer. Our method is effective across different TAD architectures." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.118, + 0.481, + 0.19 + ], + "angle": 0, + "content": "
Drop #BlockAdaTAD [40]ActionFormer [70]
mAP (tIoU=0.5)MACs (G)mAP (tIoU=0.5)MACs (G)
Baseline72.43324.872.6545.2
172.91 (↑ 0.48)298.8 (92.0%)73.21 (↑ 0.56)37.9 (83.8%)
272.51 (↑ 0.08)272.8 (84.0%)73.28 (↑ 0.63)30.6 (67.7%)
371.68 (↓ 0.75)246.9 (76.0%)73.14 (↑ 0.49)27.0 (59.7%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.194, + 0.483, + 0.235 + ], + "angle": 0, + "content": "Table 5. Evaluation of our method on fine-grained FineActions [43]. Our approach continues to exhibit efficacy even in scenarios involving more fine-grained actions." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.237, + 0.48, + 0.305 + ], + "angle": 0, + "content": "
Drop #BlockmAP (Avg)MACs (G)Inference Time (ms)
Baseline18.67280.8210.2
118.77 (↑ 0.10)257.7 (91.8%)197.0 (93.7%)
218.71 (↑ 0.04)234.6 (83.5%)185.0 (88.0%)
318.43 (↓ 0.24)211.5 (75.3%)173.0 (82.3%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.308, + 0.482, + 0.336 + ], + "angle": 0, + "content": "Table 6. Results of pruning Mamba model [6] on Charades-STA. Our method is effective across various tasks and architectures." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.337, + 0.48, + 0.424 + ], + "angle": 0, + "content": "
Drop #BlockParameters (M)R1@0.3R1@0.5
Baseline60.569.4658.58
153.4 (88.4%)72.04 (↑ 2.58)58.79 (↑ 0.21)
246.4 (76.7%)71.75 (↑ 2.29)58.92 (↑ 0.34)
339.3 (65.1%)71.94 (↑ 2.48)58.79 (↑ 0.21)
432.3 (53.4%)70.83 (↑ 1.37)57.72 (↓ 0.86)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.483, + 0.529 + ], + "angle": 0, + "content": "three blocks causes only a minor drop \\((\\downarrow 0.24\\%)\\). Meanwhile, MACs are reduced to \\(83.5\\%\\) and \\(75.3\\%\\), with a notable decrease in inference time. These results demonstrate that our progressive block drop strategy effectively preserves detection accuracy while significantly reducing computational cost, even for fine-grained action detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Evaluation on the natural language localization task. Finally, to test our method beyond conventional TAD, we apply progressive block drop to the grounding head of the Mamba model [6] on Charades-STA [16] for natural language localization. Selective pruning is performed on the grounding head. As shown in Table 6, removing three out of six blocks reduces the parameter count to \\(65.1\\%\\) of the baseline while increasing performance—R1@0.3 by \\(2.48\\%\\) and R1@0.5 by \\(0.21\\%\\). These results demonstrate our method's generalizability across tasks and models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.686, + 0.414, + 0.702 + ], + "angle": 0, + "content": "5.5. Compatible with Pruning Techniques" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.841 + ], + "angle": 0, + "content": "We apply the pruning method [13] to models after block drop, considering three setups: 1) pruning, 2) our method, and 3) our method with pruning. As shown in Figure 4, at nearly identical accuracy (70.22% vs. 70.21%), combining our method with pruning reduces MACs by 30%, while pruning alone achieves only 10%. This highlights our method's potential for further integration with pruning to accelerate the model. Additionally, compatibility with sparse activations is analyzed in Section E of the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.47, + 0.866 + ], + "angle": 0, + "content": "5.6. Comparison of Inference Time with Pruning" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We test the inference times of models compressed via pruning and our method when both achieved the same reduction" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.087, + 0.885, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.267, + 0.905, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Comparison and compatibility with the pruning method. Bubble size represents the model's computational complexity. Combining our method with pruning enables further \\(1.3 \\times\\) acceleration, demonstrating our compatibility with pruning." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.331, + 0.905, + 0.386 + ], + "angle": 0, + "content": "Table 7. Comparisons of inference time between pruning methods and ours under the same computational complexity on THUMOS14, VideoMAE-S backbone. Reducing depth yields better acceleration than reducing width." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.387, + 0.903, + 0.445 + ], + "angle": 0, + "content": "
ModelMACs (G)Inference Time (ms)Speedup
Uncompressed286.3104.91.00×
Pruning218.296.61.09×
Ours218.081.21.29×
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.907, + 0.593 + ], + "angle": 0, + "content": "in MACs. The inference time is measured as the average duration per input sample, each consisting of 768 frames with a resolution of \\(160^2\\). As shown in Table 7, under similar MACs (218.2G vs 218.0G), pruning achieves a \\(1.09 \\times\\) speedup, while the block drop method achieves a \\(1.29 \\times\\) speedup. This indicates that compressing the model by reducing network depth is more effective for acceleration than reducing network width because shallow and wide models are more hardware-friendly architectures [39]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.634, + 0.621 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we have proposed a progressive block drop method to compress TAD models from the depth perspective. This method consists of two main steps: 1) The block selection evaluator automatically selects blocks to be dropped based on their importance; 2) Parameter-efficient learning combined with three alignment techniques is used to recover performance. On the THUMOS14 and ActivityNet-1.3 datasets, we have achieved a reduction of about \\(25\\%\\) in computational cost while improving the pruned model's performance above that of the uncompressed model. We have further validated its generalizability across TAD architectures, fine-grained datasets, and other tasks, demonstrating compatibility with existing compression techniques. Additionally, we have demonstrated the compatibility of our method with other compression techniques through experiments. We hope that our work paves the way for further advancements in the compression and acceleration of TAD models." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29232" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.258 + ], + "angle": 0, + "content": "Acknowledgements. This work was partially supported by National Natural Science Foundation of China (NSFC) under Grants 62202311, the Guangdong Basic and Applied Basic Research Foundation under Grants 2023A1515011512, Excellent Science and Technology Creative Talent Training Program of Shenzhen Municipality under Grant RCBS20221008093224017, the Shenzhen Natural Science Foundation (the Stable Support Plan Program) under Grant 20220809180405001, Key Scientific Research Project of the Department of Education of Guangdong Province 2024ZDZX3012." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.275, + 0.188, + 0.29 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.3, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.372, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning (ICML), pages 41-48, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.482, + 0.498 + ], + "angle": 0, + "content": "[3] Shyamal Buch, Victor Escorcia, Chuanqi Shen, Bernard Ghanem, and Juan Carlos Niebles. Sst: Single-stream temporal action proposals. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 2911-2920, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[4] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 961-970, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[5] Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A Ross, Jia Deng, and Rahul Sukthankar. Rethinking the faster r-cnn architecture for temporal action localization. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1130–1139, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[6] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.73, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[7] Shi Chen and Qi Zhao. Shallowing deep networks: Layerwise pruning based on feature representations. IEEE transactions on pattern analysis and machine intelligence, 41(12): 3048-3056, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.855 + ], + "angle": 0, + "content": "[8] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fine-tuning of long-context large language models. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[9] Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. Advances in" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.3, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "neural information processing systems (NeurIPS), 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[10] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning (ICML), pages 2793-2803. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[11] Amir Ben Dror, Niv Zehngut, Avraham Raviv, Evgeny Artyomov, Ran Vitek, and Roy Jevnisek. Layer folding: Neural network depth reduction using activation linearization. arXiv preprint arXiv:2106.09309, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[12] Sara Elkerdawy, Mostafa Elhoushi, Abhineet Singh, Hong Zhang, and Nilanjan Ray. To filter prune, or to layer prune, that is the question. In proceedings of the Asian conference on computer vision (ACCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[13] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 16091-16101, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[14] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations (ICLR), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[15] Yonggan Fu, Haichuan Yang, Jiayi Yuan, Meng Li, Cheng Wan, Raghuraman Krishnamoorthi, Vikas Chandra, and Yingyan Lin. Depthshrinker: a new compression paradigm towards boosting real-hardware efficiency of compact neural networks. In International Conference on Machine Learning (ICML), pages 6849-6862. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[16] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[17] Lianli Gao, Zhao Guo, Hanwang Zhang, Xing Xu, and Heng Tao Shen. Video captioning with attention-based LSTM and semantic consistency. IEEE Transactions on Multimedia, 19(9):2045-2055, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[18] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. In International Conference on Learning Representations (ICLR), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[19] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[20] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems (NeurIPS), 28, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[21] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 1389-1397, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29233" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 4340-4349, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.147 + ], + "angle": 0, + "content": "[23] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. stat, 1050:9, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[24] Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. Journal of Machine Learning Research, 22(241):1-124, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[25] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 1314-1324, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[26] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[27] Yupan Huang, Qi Dai, and Yutong Lu. Decoupling localization and classification in single shot temporal action detection. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1288-1293. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[28] YG Jiang, J Liu, A Roshan Zamir, G Toderici, I Laptev, M Shah, and R Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[29] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[30] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems (NeurIPS), 25, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.582, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[31] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1): 79-86, 1951. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[32] Mark Kurtz, Justin Kopinsky, Rati Gelashvili, Alexander Matveev, John Carr, Michael Goin, William Leiserson, Sage Moore, Nir Shavit, and Dan Alistarh. Inducing and exploiting activation sparsity for fast inference on deep neural networks. In International Conference on Machine Learning (ICML), pages 5533-5543. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[33] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. Tvqa: Localized, compositional video question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1369-1379, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[34] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[35] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 3320-3329, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[36] Tianwei Lin, Xu Zhao, and Zheng Shou. Single shot temporal action detection. In Proceedings of the 25th ACM international conference on Multimedia, pages 988-996, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[37] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 3889-3898, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.218, + 0.905, + 0.299 + ], + "angle": 0, + "content": "[38] Ji Liu, Dehua Tang, Yuanxian Huang, Li Zhang, Xiaocheng Zeng, Dong Li, Mingjie Lu, Jinzhang Peng, Yu Wang, Fan Jiang, et al. Updp: A unified progressive depth pruner for cnn and vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13891-13899, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[39] Ning Liu, Xiaolong Ma, Zhiyuan Xu, Yanzhi Wang, Jian Tang, and Jieping Ye. Autocompress: An automatic dnn structured pruning framework for ultra-high compression rates. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4876-4883, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[40] Shuming Liu, Chen-Lin Zhang, Chen Zhao, and Bernard Ghanem. End-to-end temporal action detection with 1b parameters across 1000 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18591-18601, 2024. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[41] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20010-20019, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[42] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[43] Yi Liu, Limin Wang, Yali Wang, Xiao Ma, and Yu Qiao. Fineaction: A fine-grained video dataset for temporal action localization. IEEE transactions on image processing, 31: 6937-6950, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[44] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. IEEE Robotics and Automation Letters, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[45] Xiaolong Ma, Sheng Lin, Shaokai Ye, Zhezhi He, Linfeng Zhang, Geng Yuan, Sia Huat Tan, Zhengang Li, Deliang Fan, Xuehai Qian, et al. Non-structured dnn weight pruning—is it beneficial in any platform? IEEE transactions on neural networks and learning systems, 33(9):4930–4944, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[46] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? Advances in neural information processing systems (NeurIPS), 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[47] Seyed Iman Mirzadeh, Keivan Alizadeh-Vahid, Sachin Mehta, Carlo C del Mundo, Oncel Tuzel, Golnoosh Samei, Mohammad Rastegari, and Mehrdad Farajtabar. Relu strikes back: Exploiting activation sparsity in large language models. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[48] P Molchanov, S Tyree, T Karras, T Aila, and J Kautz. Pruning convolutional neural networks for resource efficient in" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29234" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "ference. In 5th International Conference on Learning Representations, ICLR 2017-Conference Track Proceedings, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[49] Alberto Montes, Amaia Salvador Aguilera, Santiago Pascual, and Xavier Giró Nieto. Temporal activity detection in untrimmed videos with recurrent neural networks. In Proceedings of the 1st NIPS Workshop on Large Scale Computer Vision Systems (NeurIPS), pages 1-5, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[50] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Proposal-free temporal action detection via global segmentation mask learning. In European Conference on Computer Vision (ECCV), pages 645-662. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[51] AJ Piergiovanni and Michael Ryoo. Temporal gaussian mixture layer for videos. In International Conference on Machine learning (ICML), pages 5152-5161. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[52] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 658-666, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[53] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18857-18866, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[54] Zheng Shou, Dongang Wang, and Shih-Fu Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1049-1058, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[55] Suraj Srinivas, Akshayvarun Subramanya, and R Venkatesh Babu. Training sparse neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR) workshops, pages 138-145, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[56] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 13526-13535, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.742 + ], + "angle": 0, + "content": "[57] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2820-2828, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[58] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[59] Marvin Teichmann, Michael Weber, Marius Zoellner, Roberto Cipolla, and Raquel Urtasun. Multinet: Real-time joint semantic reasoning for autonomous driving. In 2018 IEEE intelligent vehicles symposium (IV), pages 1013-1020. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[60] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "for self-supervised video pre-training. Advances in neural information processing systems (NeurIPS), 35:10078-10093, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[61] Binglu Wang, Yongqiang Zhao, Le Yang, Teng Long, and Xuelong Li. Temporal action localization in the deep learning era: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[62] Lining Wang, Haosen Yang, Wenhao Wu, Hongxun Yao, and Hujie Huang. Temporal action proposal generation with transformers. arXiv preprint arXiv:2105.12043, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[63] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555-4576, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[64] Huijuan Xu, Abir Das, and Kate Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 5783-5792, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[65] Pengtao Xu, Jian Cao, Fanhua Shang, Wenyu Sun, and Pu Li. Layer pruning via fusible residual convolutional block for deep neural networks. arXiv preprint arXiv:2011.14356, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[66] Le Yang, Houwen Peng, Dingwen Zhang, Jianlong Fu, and Junwei Han. Revisiting anchor mechanisms for temporal action localization. IEEE Transactions on Image Processing, 29:8535-8548, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[67] Min Yang, Guo Chen, Yin-Dong Zheng, Tong Lu, and Limin Wang. Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding, 232:103692, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[68] Serena Yeung, Olga Russakovsky, Greg Mori, and Li Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 2678–2687, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[69] Runhao Zeng, Wenbing Huang, Mingkui Tan, Yu Rong, Peilin Zhao, Junzhou Huang, and Chuang Gan. Graph convolutional module for temporal action localization in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6209-6223, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[70] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision (ECCV), pages 492-510. Springer, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[71] Minjia Zhang and Yuxiong He. Accelerating training of transformer-based language models with progressive layer dropping. Advances in neural information processing systems (NeurIPS), 33:14011-14023, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[72] Chen Zhao, Shuming Liu, Karttikeya Mangalam, and Bernard Ghanem. Re2tal: Rewiring pretrained video backbones for reversible temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10637-10647, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[73] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 2914-2923, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29235" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.487, + 0.148 + ], + "angle": 0, + "content": "[74] Yao Zhou, Gary G Yen, and Zhang Yi. Evolutionary shallowing deep neural networks at block levels. IEEE Transactions on Neural Networks and Learning Systems, 33(9): 4635-4647, 2021. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29236" + } + ] +] \ No newline at end of file diff --git a/2025/Temporal Action Detection Model Compression by Progressive Block Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_origin.pdf b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30a64d974df7dc4299cff6aee4eff20c12eb308a --- /dev/null +++ b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/c193762d-37bf-4775-b666-dc5ed17dc22a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:627d3f8f06af7961df0f6202b865aee7d88032bfc6933c4ef7d21ab42fc2c241 +size 1095879 diff --git a/2025/Temporal Action Detection Model Compression by Progressive Block Drop/full.md b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/full.md new file mode 100644 index 0000000000000000000000000000000000000000..12e3d688e02960030c8b53f6e374a99a209291a2 --- /dev/null +++ b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/full.md @@ -0,0 +1,332 @@ +# Temporal Action Detection Model Compression by Progressive Block Drop + +Xiaoyong Chen $^{1,2*}$ , Yong Guo $^{3*}$ , Jiaming Liang $^{2}$ , Sitong Zhuang $^{2}$ , Runhao Zeng $^{1,2\dagger}$ , Xiping Hu $^{1,4\dagger}$ $^{1}$ Guangdong-Hong Kong-Macao Joint Laboratory for Emotional Intelligence and Pervasive Computing, Shenzhen MSU-BIT University, $^{2}$ Shenzhen University, $^{3}$ South China University of Technology, $^{4}$ Beijing Institute of Technology + +xiaoyongchencs@gmail.com, zengrh@smbu.edu.cn, huxp@smbu.edu.cn + +# Abstract + +Temporal action detection (TAD) aims to identify and localize action instances in untrimmed videos, which is essential for various video understanding tasks. However, recent improvements in model performance, driven by larger feature extractors and datasets, have led to increased computational demands. This presents a challenge for applications like autonomous driving and robotics, which rely on limited computational resources. While existing channel pruning methods can compress these models, reducing the number of channels often hinders the parallelization efficiency of GPU, due to the inefficient multiplication between small matrices. Instead of pruning channels, we propose a Progressive Block Drop method that reduces model depth while retaining layer width. In this way, we still use large matrices for computation but reduce the number of multiplications. Our approach iteratively removes redundant blocks in two steps: first, we drop blocks with minimal impact on model performance; and second, we employ a parameter-efficient cross-depth alignment technique, fine-tuning the pruned model to restore model accuracy. Our method achieves a $25\%$ reduction in computational overhead on two TAD benchmarks (THUMOS14 and ActivityNet-1.3) to achieve lossless compression. More critically, we empirically show that our method is orthogonal to channel pruning methods and can be combined with it to yield further efficiency gains. + +# 1. Introduction + +Temporal action detection aims to identify action instances and localize their start and end times in untrimmed videos. This task underpins many video understanding applications, including video question answering [33] and video captioning [17]. Recent advancements in larger feature extractors + +and datasets have improved model performance but also significantly increased computational demands [40]. However, many TAD applications, such as autonomous driving [59] and robotics [44], operate under limited computational resources. Thus, reducing computational complexity is crucial for deploying TAD in real-world scenarios. + +A typical TAD model [67, 70] consists of a feature extractor and an action detection head. Due to the long duration of videos, the feature extractor processes video segments individually, and these segment-level features are then combined before being passed to the detection head. As shown in Figure 1, our experiments show that the feature extraction phase accounts for $95\%$ of the total model computation, since it processes many segments, while the detection head processes only a single video-level feature. Thus, reducing the complexity of the feature extractor is key to improving TAD model efficiency. + +Pruning [14, 18, 20, 48] is a common model compression technique, which removes redundant weights based on their importance. This results in a slim-and-tall network structure, where the model's hidden layer dimensions are reduced but depth remains unchanged, as illustrated in Figure 1. However, studies have shown [24, 45] that for the same number of parameters, a deep-and-thin network may have slower inference speeds compared to a shallow-and-wide network. The reason for this is that, during GPU inference, operations on larger matrices tend to have higher parallelism and are thus more efficient than multiple operations on smaller matrices. + +To explore the feasibility of reducing the architectural depth of TAD models, we conduct preliminary experiments on the THUMOS14 dataset [28], which is widely used for this task. We raise the following key observations from Figure 2: 1) There is a slight discrepancy between the input and output features of certain blocks. 2) Dropping specific blocks results in only a minimal decrease in model performance. 3) Simultaneously dropping multiple blocks leads to a more significant performance drop (Table 3). These results motivate us to investigate TAD model compression + +![](images/5be35b489cf329a5945034f18ec0535a52d21a958c605a93f7216ef7359c05ec.jpg) +Figure 1. Comparisons between reducing depth and reducing width. For a 768-frame video sequence with a resolution of $160^2$ , using VideoMAE-S as the feature extractor and Actionformer as the detection head, $95\%$ of the computational overhead comes from the feature extractor. Most pruning methods reduce the size of layer weight matrices, leading to a slim-and-tall network structure. In contrast, our progressive block drop method reduces network depth, achieving $1.19 \times$ faster inference at the same computational cost. These findings, detailed in Section 5.6, suggest that our approach results in a more hardware-efficient model. + +# from the depth perspective. + +In this paper, we propose a multi-step progressive block drop method that effectively reduces network depth, transforming the TAD model into a shallow-and-wide architecture more suitable for parallel GPU matrix computations. Specifically, we compress the model progressively through multiple iterations. Each iteration consists of two stages: 1) block selection and removal, where we evaluate the importance of each block based on a specific metric and remove the least important block; and 2) performance restoration, where we design a parameter-efficient cross-depth alignment strategy. This strategy first integrates LoRA [26] into the attention modules of the remaining blocks and then fine-tunes the pruned model to align with the intermediate features output by the uncompressed model's blocks, effectively compressing the information from multiple blocks into a single block. Additionally, we employ prediction-level alignment to further improve performance. On the THUMOS14 and ActivityNet datasets [4], our method achieves superior performance compared to the uncompressed model, while reducing the model's computational cost (MACs) by $25\%$ . Furthermore, we demonstrate through experiments that our approach is compatible with common pruning techniques, enabling even further compression of the model. Our main contributions are as follows: + +- We focus on compressing the TAD model from the perspective of depth. Unlike channel pruning methods that compress the model from the width or the internal layers of blocks, we analyze the redundancy within the TAD model and explore block-level compression as a novel approach. This provides a new angle for compressing and + +accelerating the TAD model. + +- To achieve depth compression, we propose a progressive multi-step block drop strategy. In each iteration, a block is automatically removed, and a parameter-efficient training technique is employed to recover the model's performance. On THUMOS14 and ActivityNet, our method reduces the model's MACs by approximately $25\%$ , while simultaneously outperforming the uncompressed model. Furthermore, we extend our method to other TAD architectures, fine-grained datasets, and additional tasks to validate its generalizability. +- We empirically demonstrate that our method is orthogonal to channel pruning techniques and investigate its compatibility with other model compression approaches, resulting in additional efficiency improvements. + +# 2. Related Work + +# 2.1. Temporal Action Detection + +Temporal action detection (TAD) can generally be divided into two primary categories [61]. The first category, two-stage methods, involves generating action proposals followed by classification and boundary refinement [5, 54, 64, 69, 73]. Proposal generation techniques may employ frame-or segment-level classification with subsequent merging of frames or segments belonging to the same action class [49-51], while others use specialized proposal generation strategies [3, 37]. However, the success of these approaches heavily relies on the quality of the proposals, which has led to the emergence of integrated approaches, also known as one-stage methods. These methods simultaneously handle proposal generation, classification, and/or boundary regres + +sion [27, 36, 58, 68]. Key developments in this area include the adoption of anchor-based mechanisms for TAD [36] and the shift towards anchor-free frameworks [35, 53], with recent efforts combining the benefits of both anchor-based and anchor-free approaches [66]. + +In recent years, Transformer-based architectures, which have demonstrated remarkable results in various computer vision tasks, have been applied to TAD [56, 62, 70]. Additionally, the development of end-to-end architectures has been explored in various studies [40-42, 67, 72]. With the significant achievements of scaling up in large language models [1], recent research [40] has utilized scaling up in conjunction with a frame input data processing approach. This combination has not only significantly improved model performance but also reduced computational costs. However, the issue of increased inference time due to computational scaling has become increasingly critical. Therefore, we focus on reducing inference latency through model compression. + +# 2.2. Model Compression + +Model compression techniques are typically categorized into pruning [20], knowledge distillation [23], model decomposition [9], and quantization [30]. This paper focuses on model pruning, which can be either unstructured [20] or structured [34]. Unstructured pruning removes network connections to create sparsity, but it requires specialized hardware for sparse matrix operations [19, 20, 55], limiting its practical deployment. Structured pruning, which includes width-reduction methods like channel pruning [21, 34] and filter pruning [22], results in a dense weight matrix, enabling GPU acceleration. However, it may not fully exploit GPU parallelism [39]. + +Depth-wise pruning techniques [25, 57] have been used for model acceleration, although they may suffer from insufficient memory occupancy. Various methods, including layer importance evaluation [7, 12], parameter optimization [65, 74], and reparameterization [11, 15], can also guide pruning to compress the model from the perspective of depth. Previous work [38, 71] has explored the concept of progressive block dropping; however, there are significant differences between their approaches and ours. Specifically, [38] focuses on compressing and integrating the layers within the block, which is a layer-level pruning technique. The dynamically skipping blocks method proposed in [71] accelerates training by dropping blocks, but during inference, the full model is still employed. Additionally, skipping blocks to reduce the depth of Transformer models [10, 46] often leads to a loss in accuracy. In contrast, our progressive drop block method reduces depth at the block level, improving both inference speed and detection accuracy. Our method can be combined with structured pruning [13] and unstructured activation sparsity [32, 47] to + +![](images/e6c7801eb3d8547ca1c5438d6e34f28c25a4802798c0303b16f6135231c09027.jpg) +Figure 2. Analysis of TAD models at block level on THUMOS14 datasets, using VideoMAE-S. Blue curve: impact of dropping a single block on the model's detection accuracy, where most cases result in minimal accuracy degradation. Green line: MSE of the input and output features for each block. Some blocks exhibit MSE values close to 0, indicating that these blocks are redundant. + +further compress the model and accelerate inference while improving performance. + +# 3. Preliminaries and Motivation + +In this section, we first introduce the notations and formulations pertinent to temporal action detection (TAD) and then explore the feasibility of reducing the depth of TAD models, as well as some key findings that inspire the design of subsequent methods. + +# 3.1. TAD Formulation and Notation + +TAD aims to identify action instances within a video and determine their temporal boundaries. Given an untrimmed video represented as $V = \{I_t\}_{t=1}^T$ , where $I_t$ denotes the frame at time $t$ , TAD predicts a set of action instances $\Phi_V = \{\phi_i = (t_{s_i}, t_{e_i}, c_i)\}_{i=1}^N$ . Here, $N$ is the number of action instances; for each instance $\phi_i$ , $t_{s_i}$ and $t_{e_i}$ are its starting and ending times, respectively, and $c_i$ denotes its category. + +# 3.2. Feasibility of Reducing TAD model's Depth + +We find that there is a severe redundancy in the depth of TAD models. To verify this, we conduct a series of experiments and summarize the key observations as follows. + +Observation 1: Most computational overhead lies in the clip-level feature extraction backbone. TAD methods typically consist of two main parts: the feature extractor and the detection head. In the feature extraction stage, a sliding window of length $L$ is processed by a video feature extractor. For instance, features are extracted from frames $\{I_t, \dots, I_{t + L}\}$ , and then the window shifts by $s$ frames to extract features from frames $\{I_{t + s}, \dots, I_{t + s + L}\}$ . As the feature extraction process is repeated across the entire video sequence, it incurs substantial computational cost, with each window requiring complex operations to extract features. Once the features are extracted, they are con + +![](images/c456f149923bb5398299756c2249af0468b11b6005f554c5e0f36d0fae608e22.jpg) +Figure 3. The diagram of our progressive block drop method. Our approach adopts a multi-step progressive compression strategy. At each iteration 1) we evaluate the importance of each block and drop the least important block, and 2) we use parameter-efficient tuning techniques, and recover performance by learning from the uncompressed model through feature-level and prediction-level alignment. + +catenated along the temporal dimension, and information is fused across the spatial dimension by averaging the features. These fused features are then passed to the detection head. Ultimately, the volume of data fed into the detection head is much smaller compared to the data processed by the feature extractor. As shown in Figure 1, the computational cost of the feature extractor accounts for $95\%$ of the total forward inference cost of the model. Given the high computational cost of feature extraction, optimizing or reducing the complexity of this stage presents a significant opportunity for improving model efficiency. Therefore, to reduce the overall model computation, the primary focus should be on compressing the feature extractor. + +Observation 2: There exist blocks that have a minor discrepancy between their input and output. Additionally, by analyzing the feature differences between consecutive blocks, as depicted in Figure 2, we find that some blocks exhibit very similar representation capabilities, suggesting redundancy in the model. These findings highlight the feasibility of applying the block drop method to reduce computational complexity without sacrificing performance. + +Observation 3: The removal of certain blocks has a negligible impact on TAD performance. To investigate the potential for model compression, we first assess the feasibility of dropping blocks by examining the impact of removing individual blocks on model performance. As shown + +in Figure 2, our experiments reveal that removing certain blocks led to only minimal performance degradation, indicating that block drop could be a viable approach. + +Observation 4: Simultaneously removing multiple blocks at a time results in a significant performance drop. In our preliminary exploration, we attempt to directly drop multiple blocks and train the model to observe whether it could recover its performance. We randomly select three blocks to drop all at once and then attempt fine-tuning. As shown in Table 3, the best performance achieved is $0.67\%$ lower than the uncompressed model, exhibiting a noticeable performance drop, which is an undesirable outcome in practical applications. Therefore, we adopt a multi-step progressive block dropping strategy rather than removing multiple blocks at once. This approach allows the information from the removed blocks to be more effectively compressed into the pruned model, resulting in better performance. + +# 4. Progressive Block Drop + +In this paper, we design a compression method specifically tailored for TAD models. Drawing from our analysis in Section 3.2, our focus is on reducing the computational cost of the feature extraction backbone while maintaining or even enhancing performance. To achieve this, we propose a multi-step progressive block-level dropping method, where each step comprises two stages. First, the Block Selection + +Evaluator automatically selects the blocks to be dropped, resulting in a subnet. Then, Cross-Depth Alignment enables parameter-efficient learning from the uncompressed model to maintain TAD performance. By iteratively applying these steps, we progressively obtain a pruned TAD backbone that can be directly deployed for inference. + +# 4.1. Block Selection Evaluator + +Experimental results presented in Section 3.2 demonstrate that simultaneously removing multiple blocks significantly degrades the performance of TAD. Inspired by prior work on curriculum learning [2, 63], we propose a progressive block-drop strategy. However, selecting the appropriate blocks to drop is a non-trivial task. To address this challenge, we introduce a block selection evaluator $f_{E}$ to select the blocks to be dropped automatically. + +Without loss of generality, let $M_0$ denote the initial uncompressed model trained on the training set, and let $M_{t-1}$ represent the model input to $f_E$ at iteration $t$ . We denote $B_{t-1} = \{b_1, b_2, \ldots, b_K\}$ as the set of all blocks within $M_{t-1}$ , where $K$ is the total number of blocks. At iteration $t$ , the collection of sub-networks resulting from dropping each single block is denoted as: + +$$ +\mathcal {S} _ {t} = \left\{M _ {t, k} ^ {\text {s u b}} = M _ {t - 1} \backslash b _ {k} \mid b _ {k} \in B _ {t - 1} \right\}. \tag {1} +$$ + +To automatically select blocks for pruning, we employ the evaluation function $f_{E}$ to assess each subnet. By measuring the performance difference between each subnet and the uncompressed TAD model $M_0$ on the training set, $f_{E}$ outputs an importance score for each subnet; the smaller the performance difference, the higher the importance score assigned. This process can be formulated as: + +$$ +\mathcal {P} _ {t} = f _ {E} \left(\mathcal {S} _ {t}, D _ {\text {t r a i n}}\right). \tag {2} +$$ + +From the set of importance scores $\mathcal{P}_t$ , we select the subnet $M_{t,*}^{sub}$ with the highest importance score $p^*$ , resulting in the pruned model $M_t$ at step $t$ . Note that $f_{E}$ can also compute the difference between the input and output features of each removed block or the training loss of each subnet on the training set; in both cases, a smaller difference corresponds to a higher importance score. In our experiments (Section F.1 of the appendix), we find that selecting subnets based on the mAP metric from the training set yields the best results. + +# 4.2. Parameter-Efficient Learning from Uncompressed Model for Performance Recovery + +After removing one block from the TAD model, a central challenge arises in ensuring that the performance of the reduced-depth model remains comparable to that of the original, uncompressed model (see Section F.2 of appendix). However, training the TAD model on untrimmed videos using full-parameter fine-tuning still demands a substantial amount of GPU memory, as discussed in [40]. For + +Algorithm 1 Proposed Progressive Block Drop Method +Input: Uncompressed model $M_0$ , model $M_{t-1}$ at step $t$ , blocks $B_{t-1} = \{b_1, b_2, \ldots, b_K\}$ within $M_{t-1}$ , training set $D_{train}$ . // Progressive block drop cycle while Performance of $M_{t-1} \geq M_0$ do // Stage 1: select one block to drops Obtain sub-nets by dropping each block: $\mathcal{S}_t = \left\{M_{t,k}^{\mathrm{sub}} = M_{t-1} \setminus b_k \mid b_k \in B_{t-1}\right\}$ Evaluate each sub-net using the evaluator $f_E$ : $\mathcal{P}_t = f_E(\mathcal{S}_t, D_{train})$ Select the subnet with the highest performance: $M_t = \arg \max(\mathcal{P}_t)$ // Stage 2: performance recovery Insert LoRA modules into $M_t$ for training: $M_t(\theta) \rightarrow \hat{M}_t(\theta, \theta_{\mathrm{LoRA}})$ Train the model with Eqn. (6) Merge LoRA parameters into the model: $\theta_t \gets \theta_t + \theta_{\mathrm{LoRA}}'$ end while + +example, in our experiments with the VideoMAE-L model we used, setting the batch size to 1 already consumes 13.1 GB of memory. Therefore, we introduce LoRA to reduce memory consumption during training, with detailed experimental results provided in Section F.3 of the appendix. + +Before training, given the pruned model $M_{t}(\theta_{t})$ at step $t$ , where $\theta_{t}$ represents its parameters, we first follow the LoRA method [8, 26] to add trainable parameters $\theta_{\mathrm{LORA}}$ to the attention layer each block, resulting in the model $\hat{M}_{t}(\theta_{t},\theta_{\mathrm{LORA}})$ . This approach offers two main benefits: first, during subsequent training, only $\theta_{\mathrm{LORA}}$ is trained, significantly reducing memory consumption; second, after training, $\theta_{\mathrm{LORA}}$ can be merged with $\theta_{t}$ , thereby keeping the TAD model's structure unchanged and incurring no additional inference computational overhead. + +Cross-Depth Alignment. The structural differences between the pruned and uncompressed models—resulting from block dropping—pose a significant challenge for recovering performance. More critically, we empirically observed that directly supervising the training of $\hat{M}_t$ with ground truth fails to restore performance (see Section E.5). In other words, relying solely on prediction-level supervision is inadequate for restoring performance. To address this issue, we shift our attention to feature-level supervision and propose a cross-depth alignment strategy. + +Specifically, if we consider three adjacent blocks, $\{b_{i - 1}, b_i, b_{i + 1}\}$ , as a group, when $b_i$ is dropped, the input to $b_{i + 1}$ changes (i.e., it becomes the output of $b_{i - 1}$ ). To maintain performance, it is necessary to: 1) ensure that the input feature distribution of $b_{i + 1}^p$ in the pruned model $\hat{M}_t$ aligns with that of $b_i^u$ in the uncompressed model $M_0$ ; and + +2) ensure that the output distribution of the $(i + 1)$ -th block in $\hat{M}_t$ aligns with that of $M_0$ . In our implementation, after the $i$ -th block is dropped, we align the outputs of the $(i - 1)$ -th and $(i + 1)$ -th blocks $(f_{b_{i - 1}}$ and $f_{b_{i + 1}})$ between the pruned and uncompressed models, thereby compressing the information from multiple blocks into a single block. For the features of other blocks, we similarly apply block-wise alignment with the $M_0$ . The cross-depth alignment loss can be expressed as follows: + +$$ +\mathcal {L} _ {f} = \frac {1}{I - 1} \sum_ {m = 1, m \neq i} ^ {I} \left(f _ {b _ {m}} ^ {M _ {0}} - f _ {b _ {m}} ^ {\hat {M} _ {t}}\right) ^ {2}. \tag {3} +$$ + +Action classification and Localization Alignment. The TAD model generates two distinct types of predictions: action categories and temporal bounding boxes. We also align the predictions of the pruned model $\hat{M}_t$ with those of the uncompressed model $M_0$ . + +1) Action classification alignment: To align the action class predictions, we employ the Kullback-Leibler (KL) divergence loss $\mathcal{L}_{\mathrm{KL}}$ [31], which measures the divergence between the action class scores of the uncompressed model $M_0$ and the pruned model $\hat{M}_t$ . The alignment loss is defined as: + +$$ +\mathcal {L} _ {p c} = \mathcal {L} _ {\mathrm {K L}} \left(p \left(z _ {M _ {0}}\right), p \left(z _ {\hat {M} _ {t}}\right)\right), \tag {4} +$$ + +where $z_{M_0}$ and $z_{\hat{M}_t}$ denote the logits of the uncompressed and pruned models, respectively, and $p(z_{M_0})$ and $p(z_{\hat{M}_t})$ represent the class probabilities derived from the logits via the softmax function. + +2) Boundary regression alignment: For aligning the boundary predictions, we compute the Generalized Intersection over Union (GIoU) loss $\mathcal{L}_{\mathrm{GIoU}}$ [52] between the predicted temporal bounding boxes of the uncompressed model $M_0$ and the pruned model $\hat{M}_t$ . The alignment loss for boundary regression is: + +$$ +\mathcal {L} _ {p r} = \frac {1}{N _ {p}} \sum_ {i = 1} ^ {N _ {p}} \mathcal {L} _ {\mathrm {G I o U}} \left(D _ {i} ^ {M _ {0}}, D _ {i} ^ {\hat {M} _ {t}}\right), \tag {5} +$$ + +where $N_{p}$ denotes the number of predicted positive temporal bounding boxes, and $D_{i}^{M_{0}}$ and $D_{i}^{\hat{M}_{t}}$ represent the $i$ -th predicted temporal bounding boxes for models $M_0$ and $\hat{M}_t$ , respectively. This dual alignment ensures that the pruned model learns both the correct action classification and the accurate temporal boundaries, which are crucial for TAD. + +Total Loss Function. We incorporate the alignment losses, along with the commonly used classification loss $\mathcal{L}_{\mathrm{cls}}$ and regression loss $\mathcal{L}_{\mathrm{reg}}$ in TAD [40], as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {p c} + \mathcal {L} _ {p r} + \mathcal {L} _ {f} + \mathcal {L} _ {\text {c l s}} + \mathcal {L} _ {\text {r e g}}. \tag {6} +$$ + +After training, the LoRA parameters $\theta_{\mathrm{LoRA}}^{\prime}$ are integrated with the pruned model as: + +$$ +\theta_ {t} \leftarrow \theta_ {t} + \theta_ {\mathrm {L o R A}} ^ {\prime}. \tag {7} +$$ + +The updated model $M_{t}(\theta_{t})$ is then input into the block drop selection evaluator at the $t + 1$ step for the next iteration, continuing until the performance of the pruned model is no longer recoverable to that of the uncompressed model. Ultimately, the compressed TAD model is deployed, offering a reduced number of parameters and faster inference speed. + +# 5. Experiments + +In the experimental section, we conduct the following experiments: 1) Sections 5.3 and 5.4 evaluate the effectiveness of the proposed method; 2) Section 5.5 validates compatibility with other compression models; 3) Section 5.6 compares inference speed with pruning methods. + +# 5.1. Datasets and Evaluation Metric + +THUMOS14 [28] is a prominent benchmark dataset in TAD research. It consists of a training set with 200 videos and a test set with 213 videos, covering 20 action categories. This task is notably challenging due to the presence of more than 15 separate action instances per video, with background scenes occupying $71\%$ of the frames. + +ActivityNet-1.3 [4] is a large-scale benchmark for TAD in untrimmed videos, containing approximately 20,000 videos across 200 activity classes, totaling over 600 hours of content. The dataset is divided into three subsets: 10,024 videos for training, 4,926 for validation, and 5,044 for testing, with an average of 1.65 action instances per video. + +FineAction [43] is a benchmark dataset for temporal action localization, containing 103K instances across 106 action categories in 17K untrimmed videos. It presents both opportunities and challenges for TAD due to its fine-grained action classes, dense annotations, and frequent co-occurrence of multiple action classes. + +Mean Average Precision (mAP). A predicted temporal bounding box is deemed accurate if its temporal Intersection over Union (tIoU) with the corresponding ground truth instance exceeds a specific threshold and matches the instance's category. For THUMOS14, the tIoU thresholds range from $\{0.3, 0.4, 0.5, 0.6, 0.7\}$ , with mAP@tIoU=0.5 reported for comparison. In contrast, on ActivityNet-1.3 and FineAction, tIoU thresholds are sampled from $\{0.5: 0.05: 0.95\}$ , and we report the mean mAP over tIoU thresholds from 0.5 to 0.95, in increments of 0.05. + +# 5.2. Implementation Details + +We use the ActionFormer [70] detector for prediction, employing two Transformer-based feature extractors: VideoMAE-S with 12 blocks and VideoMAE-L [60] with 24 blocks, both are pre-trained on the K400 dataset [29]. For the THUMOS14 dataset, the input videos are randomly truncated to a fixed length of 768 frames, while for the ActivityNet-1.3 and FineAction dataset, the input videos are + +Table 1. Evaluation of our method in terms of computational complexity, inference time and mAP with VideoMAE-S as the backbone. Our approach reduces the computational cost by approximately $25\%$ , achieves around $1.3\mathrm{x}$ inference speedup, and improves the performance. + +
Model (Drop #Block)THUMOS14ActivityNet-1.3
MACs (G)Inference Time (ms)mAP (tIoU=0.5)MACs (G)Inference Time (ms)mAP (Avg)
Baseline286.3104.970.43274.1103.837.75
1263.5 (92.0%)98.4 (93.8%)71.06 (↑ 0.63)251.3 (91.7%)95.1 (91.6%)37.94 (↑ 0.19)
2240.8 (84.1%)89.8 (85.6%)71.37 (↑ 0.94)228.5 (83.4%)86.8 (83.6%)37.81 (↑ 0.06)
3218.0 (76.1%)81.2 (77.4%)70.47 (↑ 0.04)205.8 (75.1%)79.6 (76.7%)37.77 (↑ 0.02)
4195.2 (68.2%)73.6 (70.2%)69.65 (↓ 0.78)183.0 (66.8%)72.1 (69.5%)37.72 (↓ 0.03)
+ +Table 2. Evaluation of our method on deeper networks in terms of computational complexity, inference time and mAP with VideoMAE-L (24 blocks) as the backbone. Our method is generally effective and larger models may benefit more from it. + +
Drop #BlockMACs (G)Inference Time (ms)mAP (tIoU=0.5)
Baseline3886.9795.976.01
13725.5 (95.8%)765.4 (96.2%)77.97 (↑ 1.96)
23564.1 (91.7%)731.4 (91.9%)78.00 (↑ 1.99)
33402.7 (87.5%)704.3 (88.5%)78.29 (↑ 2.28)
43241.4 (83.4%)671.3 (84.3%)77.86 (↑ 1.85)
53080.0 (79.2%)647.1 (81.3%)77.57 (↑ 1.56)
62918.6 (75.1%)602.6 (75.7%)77.25 (↑ 1.24)
+ +resized to a fixed length of 768 frames, with the frame resolution set to $160^2$ by default for both datasets. Please refer to the supplementary for more details. + +# 5.3. Effectiveness of Our Progressive Block Drop + +We compare the model trained with the progressive block drop method to the uncompressed model from three aspects. 1) Computational Complexity (MACs): The computational complexity is calculated based on the number of multiply-accumulate operations at each layer of the model. 2) Inference Time: The inference time measures the average duration for processing each input sample, where each sample consists of 768 frames with a resolution of $160^{2}$ . 3) mAP: The results, as shown in Table 1, demonstrate that on both datasets, our method achieves an approximately $25\%$ reduction in MACs while outperforming the uncompressed model in terms of performance. This highlights the superiority of our approach, as it effectively reduces computational complexity while simultaneously enhancing the performance of the compressed model—an accomplishment that is difficult for other compression methods to achieve. + +Drop blocks on a deeper backbone. We explore the performance of the block drop method when the depth of the feature extractor is increased from 12 to 24, i.e., when transitioning from VideoMAE-S to VideoMAE-L. The results, as shown in Table 2, reveal that with a $25\%$ reduction in MACs, the model achieves a $1.24\%$ performance improvement. Notably, when the same fraction of blocks (1/4) is dropped, the performance of VideoMAE-L improves by $1.24\%$ , whereas VideoMAE-S only shows a modest improvement of $0.04\%$ (as shown in Table 1). This suggests that in models with higher computational complexity, there is more redundant information, and the progressive block drop method is effective at eliminating such redundancy. + +Table 3. Comparison with randomly dropping blocks. Dropping three randomly selected blocks at once performs worse than progressively dropping. + +
Drop Block ID[2,5,9][0,4,7][1,3,8][6,7,10][6,7,10] + ours
mAP68.8868.3169.7668.9170.47
+ +Dropping multiple blocks at one time vs Progressive dropping. To demonstrate the necessity of adopting a multi-step progressive drop in our proposed block drop method, we conduct experiments where multiple blocks are dropped simultaneously. Specifically, we randomly select three different blocks to drop and train the model using the training method described in Section 4.2. The results, as shown in Table 3, indicate that directly dropping three blocks yields performance ranging from $68.31\%$ to $69.76\%$ , with the best result still $0.67\%$ lower than that of the uncompressed model. Meanwhile, using the progressive block drop method on VideoMAE-S and the THUMOS14 dataset, the model's performance after dropping the first three layers [7, 10, 6] reaches $70.47\%$ . Comparing this with the result of directly dropping these three layers, it is evident that compressing the model via progressive dropping achieves better performance recovery ( $70.47\%$ vs. $68.91\%$ ), further underscoring the importance of the multi-step progressive drop. Therefore, our proposed multi-step progressive block drop is crucial for performance recovery. + +# 5.4. Extended Evaluations + +Evaluation on additional TAD architectures. To validate the generalizability of our method, we apply it to two additional TAD architectures: the single-stage AdaTAD [40] model with a VideoMAE-S backbone and the two-stage ActionFormer [70] model with pre-extracted I3D features. Table 4 shows that in AdaTAD, pruning two blocks improves accuracy by $0.08\%$ and reduces MACs to $84.0\%$ . Similarly, the ActionFormer model gains $0.49\%$ in mAP while reducing MACs to $59.7\%$ . These results underscore our method's effectiveness in compressing diverse TAD architectures while enhancing performance. + +Evaluation on the fine-grained FineActions dataset. To further corroborate the versatility of our approach, we conduct experiments on the fine-grained FineActions [43] dataset using a VideoMAE-S backbone with an Action-Former [70] head. As shown in Table 5, pruning two blocks slightly improves mAP ( $\uparrow 0.04\%$ ), while pruning + +Table 4. Evaluation of our method on AdaTAD and ActionFormer. Our method is effective across different TAD architectures. + +
Drop #BlockAdaTAD [40]ActionFormer [70]
mAP (tIoU=0.5)MACs (G)mAP (tIoU=0.5)MACs (G)
Baseline72.43324.872.6545.2
172.91 (↑ 0.48)298.8 (92.0%)73.21 (↑ 0.56)37.9 (83.8%)
272.51 (↑ 0.08)272.8 (84.0%)73.28 (↑ 0.63)30.6 (67.7%)
371.68 (↓ 0.75)246.9 (76.0%)73.14 (↑ 0.49)27.0 (59.7%)
+ +Table 5. Evaluation of our method on fine-grained FineActions [43]. Our approach continues to exhibit efficacy even in scenarios involving more fine-grained actions. + +
Drop #BlockmAP (Avg)MACs (G)Inference Time (ms)
Baseline18.67280.8210.2
118.77 (↑ 0.10)257.7 (91.8%)197.0 (93.7%)
218.71 (↑ 0.04)234.6 (83.5%)185.0 (88.0%)
318.43 (↓ 0.24)211.5 (75.3%)173.0 (82.3%)
+ +Table 6. Results of pruning Mamba model [6] on Charades-STA. Our method is effective across various tasks and architectures. + +
Drop #BlockParameters (M)R1@0.3R1@0.5
Baseline60.569.4658.58
153.4 (88.4%)72.04 (↑ 2.58)58.79 (↑ 0.21)
246.4 (76.7%)71.75 (↑ 2.29)58.92 (↑ 0.34)
339.3 (65.1%)71.94 (↑ 2.48)58.79 (↑ 0.21)
432.3 (53.4%)70.83 (↑ 1.37)57.72 (↓ 0.86)
+ +three blocks causes only a minor drop $(\downarrow 0.24\%)$ . Meanwhile, MACs are reduced to $83.5\%$ and $75.3\%$ , with a notable decrease in inference time. These results demonstrate that our progressive block drop strategy effectively preserves detection accuracy while significantly reducing computational cost, even for fine-grained action detection. + +Evaluation on the natural language localization task. Finally, to test our method beyond conventional TAD, we apply progressive block drop to the grounding head of the Mamba model [6] on Charades-STA [16] for natural language localization. Selective pruning is performed on the grounding head. As shown in Table 6, removing three out of six blocks reduces the parameter count to $65.1\%$ of the baseline while increasing performance—R1@0.3 by $2.48\%$ and R1@0.5 by $0.21\%$ . These results demonstrate our method's generalizability across tasks and models. + +# 5.5. Compatible with Pruning Techniques + +We apply the pruning method [13] to models after block drop, considering three setups: 1) pruning, 2) our method, and 3) our method with pruning. As shown in Figure 4, at nearly identical accuracy (70.22% vs. 70.21%), combining our method with pruning reduces MACs by 30%, while pruning alone achieves only 10%. This highlights our method's potential for further integration with pruning to accelerate the model. Additionally, compatibility with sparse activations is analyzed in Section E of the appendix. + +# 5.6. Comparison of Inference Time with Pruning + +We test the inference times of models compressed via pruning and our method when both achieved the same reduction + +![](images/8c5c4c61098f9bbfdb4d42570d6aa05fd5ef7792961f3ab2ca554d55a8531ffb.jpg) +Figure 4. Comparison and compatibility with the pruning method. Bubble size represents the model's computational complexity. Combining our method with pruning enables further $1.3 \times$ acceleration, demonstrating our compatibility with pruning. + +Table 7. Comparisons of inference time between pruning methods and ours under the same computational complexity on THUMOS14, VideoMAE-S backbone. Reducing depth yields better acceleration than reducing width. + +
ModelMACs (G)Inference Time (ms)Speedup
Uncompressed286.3104.91.00×
Pruning218.296.61.09×
Ours218.081.21.29×
+ +in MACs. The inference time is measured as the average duration per input sample, each consisting of 768 frames with a resolution of $160^2$ . As shown in Table 7, under similar MACs (218.2G vs 218.0G), pruning achieves a $1.09 \times$ speedup, while the block drop method achieves a $1.29 \times$ speedup. This indicates that compressing the model by reducing network depth is more effective for acceleration than reducing network width because shallow and wide models are more hardware-friendly architectures [39]. + +# 6. Conclusion + +In this work, we have proposed a progressive block drop method to compress TAD models from the depth perspective. This method consists of two main steps: 1) The block selection evaluator automatically selects blocks to be dropped based on their importance; 2) Parameter-efficient learning combined with three alignment techniques is used to recover performance. On the THUMOS14 and ActivityNet-1.3 datasets, we have achieved a reduction of about $25\%$ in computational cost while improving the pruned model's performance above that of the uncompressed model. We have further validated its generalizability across TAD architectures, fine-grained datasets, and other tasks, demonstrating compatibility with existing compression techniques. Additionally, we have demonstrated the compatibility of our method with other compression techniques through experiments. We hope that our work paves the way for further advancements in the compression and acceleration of TAD models. + +Acknowledgements. This work was partially supported by National Natural Science Foundation of China (NSFC) under Grants 62202311, the Guangdong Basic and Applied Basic Research Foundation under Grants 2023A1515011512, Excellent Science and Technology Creative Talent Training Program of Shenzhen Municipality under Grant RCBS20221008093224017, the Shenzhen Natural Science Foundation (the Stable Support Plan Program) under Grant 20220809180405001, Key Scientific Research Project of the Department of Education of Guangdong Province 2024ZDZX3012. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning (ICML), pages 41-48, 2009. 5 +[3] Shyamal Buch, Victor Escorcia, Chuanqi Shen, Bernard Ghanem, and Juan Carlos Niebles. Sst: Single-stream temporal action proposals. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 2911-2920, 2017. 2 +[4] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 961-970, 2015. 2, 6 +[5] Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A Ross, Jia Deng, and Rahul Sukthankar. Rethinking the faster r-cnn architecture for temporal action localization. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1130–1139, 2018. 2 +[6] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024.8 +[7] Shi Chen and Qi Zhao. Shallowing deep networks: Layerwise pruning based on feature representations. IEEE transactions on pattern analysis and machine intelligence, 41(12): 3048-3056, 2018. 3 +[8] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fine-tuning of long-context large language models. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 5 +[9] Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. Advances in + +neural information processing systems (NeurIPS), 27, 2014. 3 +[10] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning (ICML), pages 2793-2803. PMLR, 2021. 3 +[11] Amir Ben Dror, Niv Zehngut, Avraham Raviv, Evgeny Artyomov, Ran Vitek, and Roy Jevnisek. Layer folding: Neural network depth reduction using activation linearization. arXiv preprint arXiv:2106.09309, 2021. 3 +[12] Sara Elkerdawy, Mostafa Elhoushi, Abhineet Singh, Hong Zhang, and Nilanjan Ray. To filter prune, or to layer prune, that is the question. In proceedings of the Asian conference on computer vision (ACCV), 2020. 3 +[13] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 16091-16101, 2023. 3, 8 +[14] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations (ICLR), 2018. 1 +[15] Yonggan Fu, Haichuan Yang, Jiayi Yuan, Meng Li, Cheng Wan, Raghuraman Krishnamoorthi, Vikas Chandra, and Yingyan Lin. Depthshrinker: a new compression paradigm towards boosting real-hardware efficiency of compact neural networks. In International Conference on Machine Learning (ICML), pages 6849-6862. PMLR, 2022. 3 +[16] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 8 +[17] Lianli Gao, Zhao Guo, Hanwang Zhang, Xing Xu, and Heng Tao Shen. Video captioning with attention-based LSTM and semantic consistency. IEEE Transactions on Multimedia, 19(9):2045-2055, 2017. 1 +[18] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. In International Conference on Learning Representations (ICLR), 2018. 1 +[19] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 3 +[20] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems (NeurIPS), 28, 2015. 1, 3 +[21] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 1389-1397, 2017. 3 +[22] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 4340-4349, 2019. 3 +[23] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. stat, 1050:9, 2015. 3 +[24] Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. Journal of Machine Learning Research, 22(241):1-124, 2021. 1 +[25] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 1314-1324, 2019. 3 +[26] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2021. 2, 5 +[27] Yupan Huang, Qi Dai, and Yutong Lu. Decoupling localization and classification in single shot temporal action detection. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1288-1293. IEEE, 2019. 3 +[28] YG Jiang, J Liu, A Roshan Zamir, G Toderici, I Laptev, M Shah, and R Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. 1, 6 +[29] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6 +[30] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems (NeurIPS), 25, 2012. 3 +[31] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1): 79-86, 1951. 6 +[32] Mark Kurtz, Justin Kopinsky, Rati Gelashvili, Alexander Matveev, John Carr, Michael Goin, William Leiserson, Sage Moore, Nir Shavit, and Dan Alistarh. Inducing and exploiting activation sparsity for fast inference on deep neural networks. In International Conference on Machine Learning (ICML), pages 5533-5543. PMLR, 2020. 3 +[33] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. Tvqa: Localized, compositional video question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1369-1379, 2018. 1 +[34] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 3 +[35] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 3320-3329, 2021. 3 + +[36] Tianwei Lin, Xu Zhao, and Zheng Shou. Single shot temporal action detection. In Proceedings of the 25th ACM international conference on Multimedia, pages 988-996, 2017. 3 +[37] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 3889-3898, 2019. 2 +[38] Ji Liu, Dehua Tang, Yuanxian Huang, Li Zhang, Xiaocheng Zeng, Dong Li, Mingjie Lu, Jinzhang Peng, Yu Wang, Fan Jiang, et al. Updp: A unified progressive depth pruner for cnn and vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13891-13899, 2024. 3 +[39] Ning Liu, Xiaolong Ma, Zhiyuan Xu, Yanzhi Wang, Jian Tang, and Jieping Ye. Autocompress: An automatic dnn structured pruning framework for ultra-high compression rates. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4876-4883, 2020. 3, 8 +[40] Shuming Liu, Chen-Lin Zhang, Chen Zhao, and Bernard Ghanem. End-to-end temporal action detection with 1b parameters across 1000 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18591-18601, 2024. 1, 3, 5, 6, 7, 8 +[41] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20010-20019, 2022. +[42] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 3 +[43] Yi Liu, Limin Wang, Yali Wang, Xiao Ma, and Yu Qiao. Fineaction: A fine-grained video dataset for temporal action localization. IEEE transactions on image processing, 31: 6937-6950, 2022. 6, 7, 8 +[44] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. IEEE Robotics and Automation Letters, 2023. 1 +[45] Xiaolong Ma, Sheng Lin, Shaokai Ye, Zhezhi He, Linfeng Zhang, Geng Yuan, Sia Huat Tan, Zhengang Li, Deliang Fan, Xuehai Qian, et al. Non-structured dnn weight pruning—is it beneficial in any platform? IEEE transactions on neural networks and learning systems, 33(9):4930–4944, 2021. 1 +[46] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? Advances in neural information processing systems (NeurIPS), 32, 2019. 3 +[47] Seyed Iman Mirzadeh, Keivan Alizadeh-Vahid, Sachin Mehta, Carlo C del Mundo, Oncel Tuzel, Golnoosh Samei, Mohammad Rastegari, and Mehrdad Farajtabar. Relu strikes back: Exploiting activation sparsity in large language models. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3 +[48] P Molchanov, S Tyree, T Karras, T Aila, and J Kautz. Pruning convolutional neural networks for resource efficient in + +ference. In 5th International Conference on Learning Representations, ICLR 2017-Conference Track Proceedings, 2019. 1 +[49] Alberto Montes, Amaia Salvador Aguilera, Santiago Pascual, and Xavier Giró Nieto. Temporal activity detection in untrimmed videos with recurrent neural networks. In Proceedings of the 1st NIPS Workshop on Large Scale Computer Vision Systems (NeurIPS), pages 1-5, 2016. 2 +[50] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Proposal-free temporal action detection via global segmentation mask learning. In European Conference on Computer Vision (ECCV), pages 645-662. Springer, 2022. +[51] AJ Piergiovanni and Michael Ryoo. Temporal gaussian mixture layer for videos. In International Conference on Machine learning (ICML), pages 5152-5161. PMLR, 2019. 2 +[52] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 658-666, 2019. 6 +[53] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18857-18866, 2023. 3 +[54] Zheng Shou, Dongang Wang, and Shih-Fu Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1049-1058, 2016. 2 +[55] Suraj Srinivas, Akshayvarun Subramanya, and R Venkatesh Babu. Training sparse neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR) workshops, pages 138-145, 2017. 3 +[56] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 13526-13535, 2021. 3 +[57] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2820-2828, 2019. 3 +[58] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 3 +[59] Marvin Teichmann, Michael Weber, Marius Zoellner, Roberto Cipolla, and Raquel Urtasun. Multinet: Real-time joint semantic reasoning for autonomous driving. In 2018 IEEE intelligent vehicles symposium (IV), pages 1013-1020. IEEE, 2018. 1 +[60] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners + +for self-supervised video pre-training. Advances in neural information processing systems (NeurIPS), 35:10078-10093, 2022. 6 +[61] Binglu Wang, Yongqiang Zhao, Le Yang, Teng Long, and Xuelong Li. Temporal action localization in the deep learning era: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[62] Lining Wang, Haosen Yang, Wenhao Wu, Hongxun Yao, and Hujie Huang. Temporal action proposal generation with transformers. arXiv preprint arXiv:2105.12043, 2021. 3 +[63] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555-4576, 2021. 5 +[64] Huijuan Xu, Abir Das, and Kate Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 5783-5792, 2017. 2 +[65] Pengtao Xu, Jian Cao, Fanhua Shang, Wenyu Sun, and Pu Li. Layer pruning via fusible residual convolutional block for deep neural networks. arXiv preprint arXiv:2011.14356, 2020. 3 +[66] Le Yang, Houwen Peng, Dingwen Zhang, Jianlong Fu, and Junwei Han. Revisiting anchor mechanisms for temporal action localization. IEEE Transactions on Image Processing, 29:8535-8548, 2020. 3 +[67] Min Yang, Guo Chen, Yin-Dong Zheng, Tong Lu, and Limin Wang. Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding, 232:103692, 2023. 1, 3 +[68] Serena Yeung, Olga Russakovsky, Greg Mori, and Li Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 2678–2687, 2016. 3 +[69] Runhao Zeng, Wenbing Huang, Mingkui Tan, Yu Rong, Peilin Zhao, Junzhou Huang, and Chuang Gan. Graph convolutional module for temporal action localization in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6209-6223, 2021. 2 +[70] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision (ECCV), pages 492-510. Springer, 2022. 1, 3, 6, 7, 8 +[71] Minjia Zhang and Yuxiong He. Accelerating training of transformer-based language models with progressive layer dropping. Advances in neural information processing systems (NeurIPS), 33:14011-14023, 2020. 3 +[72] Chen Zhao, Shuming Liu, Karttikeya Mangalam, and Bernard Ghanem. Re2tal: Rewiring pretrained video backbones for reversible temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10637-10647, 2023. 3 +[73] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 2914-2923, 2017. 2 + +[74] Yao Zhou, Gary G Yen, and Zhang Yi. Evolutionary shallowing deep neural networks at block levels. IEEE Transactions on Neural Networks and Learning Systems, 33(9): 4635-4647, 2021. 3 \ No newline at end of file diff --git a/2025/Temporal Action Detection Model Compression by Progressive Block Drop/images.zip b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b36871d1b34bf6c9c1920844ddda2e76ec2fec03 --- /dev/null +++ b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd702ac78eaa1a38e9246d6d355358ef63f22768df3c1fd1614136de54af5af5 +size 446659 diff --git a/2025/Temporal Action Detection Model Compression by Progressive Block Drop/layout.json b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e95c7f387974dcc96d78d923d63ce024f8308e90 --- /dev/null +++ b/2025/Temporal Action Detection Model Compression by Progressive Block Drop/layout.json @@ -0,0 +1,10387 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 536, + 121 + ], + "type": "text", + "content": "Temporal Action Detection Model Compression by Progressive Block Drop" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": "Xiaoyong Chen" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": ", Yong Guo" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": ", Jiaming Liang" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": ", Sitong Zhuang" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": ", Runhao Zeng" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": ", Xiping Hu" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{1,4\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": "Guangdong-Hong Kong-Macao Joint Laboratory for Emotional Intelligence and Pervasive Computing, Shenzhen MSU-BIT University, " + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": "Shenzhen University, " + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": "South China University of Technology, " + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 561, + 199 + ], + "type": "text", + "content": "Beijing Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 201, + 480, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 201, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 201, + 480, + 213 + ], + "type": "text", + "content": "xiaoyongchencs@gmail.com, zengrh@smbu.edu.cn, huxp@smbu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 267, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 296, + 579 + ], + "type": "text", + "content": "Temporal action detection (TAD) aims to identify and localize action instances in untrimmed videos, which is essential for various video understanding tasks. However, recent improvements in model performance, driven by larger feature extractors and datasets, have led to increased computational demands. This presents a challenge for applications like autonomous driving and robotics, which rely on limited computational resources. While existing channel pruning methods can compress these models, reducing the number of channels often hinders the parallelization efficiency of GPU, due to the inefficient multiplication between small matrices. Instead of pruning channels, we propose a Progressive Block Drop method that reduces model depth while retaining layer width. In this way, we still use large matrices for computation but reduce the number of multiplications. Our approach iteratively removes redundant blocks in two steps: first, we drop blocks with minimal impact on model performance; and second, we employ a parameter-efficient cross-depth alignment technique, fine-tuning the pruned model to restore model accuracy. Our method achieves a " + }, + { + "bbox": [ + 56, + 267, + 296, + 579 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 56, + 267, + 296, + 579 + ], + "type": "text", + "content": " reduction in computational overhead on two TAD benchmarks (THUMOS14 and ActivityNet-1.3) to achieve lossless compression. More critically, we empirically show that our method is orthogonal to channel pruning methods and can be combined with it to yield further efficiency gains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 623, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 295, + 684 + ], + "type": "text", + "content": "Temporal action detection aims to identify action instances and localize their start and end times in untrimmed videos. This task underpins many video understanding applications, including video question answering [33] and video captioning [17]. Recent advancements in larger feature extractors" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 242, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 313 + ], + "type": "text", + "content": "and datasets have improved model performance but also significantly increased computational demands [40]. However, many TAD applications, such as autonomous driving [59] and robotics [44], operate under limited computational resources. Thus, reducing computational complexity is crucial for deploying TAD in real-world scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 447 + ], + "type": "text", + "content": "A typical TAD model [67, 70] consists of a feature extractor and an action detection head. Due to the long duration of videos, the feature extractor processes video segments individually, and these segment-level features are then combined before being passed to the detection head. As shown in Figure 1, our experiments show that the feature extraction phase accounts for " + }, + { + "bbox": [ + 313, + 316, + 553, + 447 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 316, + 553, + 447 + ], + "type": "text", + "content": " of the total model computation, since it processes many segments, while the detection head processes only a single video-level feature. Thus, reducing the complexity of the feature extractor is key to improving TAD model efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 553, + 591 + ], + "type": "text", + "content": "Pruning [14, 18, 20, 48] is a common model compression technique, which removes redundant weights based on their importance. This results in a slim-and-tall network structure, where the model's hidden layer dimensions are reduced but depth remains unchanged, as illustrated in Figure 1. However, studies have shown [24, 45] that for the same number of parameters, a deep-and-thin network may have slower inference speeds compared to a shallow-and-wide network. The reason for this is that, during GPU inference, operations on larger matrices tend to have higher parallelism and are thus more efficient than multiple operations on smaller matrices." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "To explore the feasibility of reducing the architectural depth of TAD models, we conduct preliminary experiments on the THUMOS14 dataset [28], which is widely used for this task. We raise the following key observations from Figure 2: 1) There is a slight discrepancy between the input and output features of certain blocks. 2) Dropping specific blocks results in only a minimal decrease in model performance. 3) Simultaneously dropping multiple blocks leads to a more significant performance drop (Table 3). These results motivate us to investigate TAD model compression" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29225" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 545, + 247 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 545, + 247 + ], + "type": "image", + "image_path": "5be35b489cf329a5945034f18ec0535a52d21a958c605a93f7216ef7359c05ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 255, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 54, + 255, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 555, + 313 + ], + "type": "text", + "content": "Figure 1. Comparisons between reducing depth and reducing width. For a 768-frame video sequence with a resolution of " + }, + { + "bbox": [ + 54, + 255, + 555, + 313 + ], + "type": "inline_equation", + "content": "160^2" + }, + { + "bbox": [ + 54, + 255, + 555, + 313 + ], + "type": "text", + "content": ", using VideoMAE-S as the feature extractor and Actionformer as the detection head, " + }, + { + "bbox": [ + 54, + 255, + 555, + 313 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 54, + 255, + 555, + 313 + ], + "type": "text", + "content": " of the computational overhead comes from the feature extractor. Most pruning methods reduce the size of layer weight matrices, leading to a slim-and-tall network structure. In contrast, our progressive block drop method reduces network depth, achieving " + }, + { + "bbox": [ + 54, + 255, + 555, + 313 + ], + "type": "inline_equation", + "content": "1.19 \\times" + }, + { + "bbox": [ + 54, + 255, + 555, + 313 + ], + "type": "text", + "content": " faster inference at the same computational cost. These findings, detailed in Section 5.6, suggest that our approach results in a more hardware-efficient model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 321, + 168, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 168, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 168, + 333 + ], + "type": "text", + "content": "from the depth perspective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 338, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 636 + ], + "type": "text", + "content": "In this paper, we propose a multi-step progressive block drop method that effectively reduces network depth, transforming the TAD model into a shallow-and-wide architecture more suitable for parallel GPU matrix computations. Specifically, we compress the model progressively through multiple iterations. Each iteration consists of two stages: 1) block selection and removal, where we evaluate the importance of each block based on a specific metric and remove the least important block; and 2) performance restoration, where we design a parameter-efficient cross-depth alignment strategy. This strategy first integrates LoRA [26] into the attention modules of the remaining blocks and then fine-tunes the pruned model to align with the intermediate features output by the uncompressed model's blocks, effectively compressing the information from multiple blocks into a single block. Additionally, we employ prediction-level alignment to further improve performance. On the THUMOS14 and ActivityNet datasets [4], our method achieves superior performance compared to the uncompressed model, while reducing the model's computational cost (MACs) by " + }, + { + "bbox": [ + 56, + 338, + 296, + 636 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 56, + 338, + 296, + 636 + ], + "type": "text", + "content": ". Furthermore, we demonstrate through experiments that our approach is compatible with common pruning techniques, enabling even further compression of the model. Our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": "- We focus on compressing the TAD model from the perspective of depth. Unlike channel pruning methods that compress the model from the width or the internal layers of blocks, we analyze the redundancy within the TAD model and explore block-level compression as a novel approach. This provides a new angle for compressing and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 323, + 321, + 441, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 321, + 441, + 332 + ], + "spans": [ + { + "bbox": [ + 323, + 321, + 441, + 332 + ], + "type": "text", + "content": "accelerating the TAD model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 333, + 555, + 501 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 333, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 333, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 333, + 555, + 453 + ], + "type": "text", + "content": "- To achieve depth compression, we propose a progressive multi-step block drop strategy. In each iteration, a block is automatically removed, and a parameter-efficient training technique is employed to recover the model's performance. On THUMOS14 and ActivityNet, our method reduces the model's MACs by approximately " + }, + { + "bbox": [ + 314, + 333, + 555, + 453 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 314, + 333, + 555, + 453 + ], + "type": "text", + "content": ", while simultaneously outperforming the uncompressed model. Furthermore, we extend our method to other TAD architectures, fine-grained datasets, and additional tasks to validate its generalizability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 453, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 555, + 501 + ], + "type": "text", + "content": "- We empirically demonstrate that our method is orthogonal to channel pruning techniques and investigate its compatibility with other model compression approaches, resulting in additional efficiency improvements." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 517, + 402, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 517, + 402, + 529 + ], + "spans": [ + { + "bbox": [ + 314, + 517, + 402, + 529 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 538, + 463, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 463, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 463, + 551 + ], + "type": "text", + "content": "2.1. Temporal Action Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "Temporal action detection (TAD) can generally be divided into two primary categories [61]. The first category, two-stage methods, involves generating action proposals followed by classification and boundary refinement [5, 54, 64, 69, 73]. Proposal generation techniques may employ frame-or segment-level classification with subsequent merging of frames or segments belonging to the same action class [49-51], while others use specialized proposal generation strategies [3, 37]. However, the success of these approaches heavily relies on the quality of the proposals, which has led to the emergence of integrated approaches, also known as one-stage methods. These methods simultaneously handle proposal generation, classification, and/or boundary regres" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29226" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "sion [27, 36, 58, 68]. Key developments in this area include the adoption of anchor-based mechanisms for TAD [36] and the shift towards anchor-free frameworks [35, 53], with recent efforts combining the benefits of both anchor-based and anchor-free approaches [66]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 301 + ], + "type": "text", + "content": "In recent years, Transformer-based architectures, which have demonstrated remarkable results in various computer vision tasks, have been applied to TAD [56, 62, 70]. Additionally, the development of end-to-end architectures has been explored in various studies [40-42, 67, 72]. With the significant achievements of scaling up in large language models [1], recent research [40] has utilized scaling up in conjunction with a frame input data processing approach. This combination has not only significantly improved model performance but also reduced computational costs. However, the issue of increased inference time due to computational scaling has become increasingly critical. Therefore, we focus on reducing inference latency through model compression." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 311, + 172, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 172, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 172, + 324 + ], + "type": "text", + "content": "2.2. Model Compression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 329, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 295, + 474 + ], + "type": "text", + "content": "Model compression techniques are typically categorized into pruning [20], knowledge distillation [23], model decomposition [9], and quantization [30]. This paper focuses on model pruning, which can be either unstructured [20] or structured [34]. Unstructured pruning removes network connections to create sparsity, but it requires specialized hardware for sparse matrix operations [19, 20, 55], limiting its practical deployment. Structured pruning, which includes width-reduction methods like channel pruning [21, 34] and filter pruning [22], results in a dense weight matrix, enabling GPU acceleration. However, it may not fully exploit GPU parallelism [39]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 474, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 714 + ], + "type": "text", + "content": "Depth-wise pruning techniques [25, 57] have been used for model acceleration, although they may suffer from insufficient memory occupancy. Various methods, including layer importance evaluation [7, 12], parameter optimization [65, 74], and reparameterization [11, 15], can also guide pruning to compress the model from the perspective of depth. Previous work [38, 71] has explored the concept of progressive block dropping; however, there are significant differences between their approaches and ours. Specifically, [38] focuses on compressing and integrating the layers within the block, which is a layer-level pruning technique. The dynamically skipping blocks method proposed in [71] accelerates training by dropping blocks, but during inference, the full model is still employed. Additionally, skipping blocks to reduce the depth of Transformer models [10, 46] often leads to a loss in accuracy. In contrast, our progressive drop block method reduces depth at the block level, improving both inference speed and detection accuracy. Our method can be combined with structured pruning [13] and unstructured activation sparsity [32, 47] to" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 554, + 196 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 554, + 196 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 554, + 196 + ], + "type": "image", + "image_path": "e6c7801eb3d8547ca1c5438d6e34f28c25a4802798c0303b16f6135231c09027.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 204, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. Analysis of TAD models at block level on THUMOS14 datasets, using VideoMAE-S. Blue curve: impact of dropping a single block on the model's detection accuracy, where most cases result in minimal accuracy degradation. Green line: MSE of the input and output features for each block. Some blocks exhibit MSE values close to 0, indicating that these blocks are redundant." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "content": "further compress the model and accelerate inference while improving performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 316, + 480, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 480, + 329 + ], + "type": "text", + "content": "3. Preliminaries and Motivation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 337, + 554, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 554, + 397 + ], + "type": "text", + "content": "In this section, we first introduce the notations and formulations pertinent to temporal action detection (TAD) and then explore the feasibility of reducing the depth of TAD models, as well as some key findings that inspire the design of subsequent methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 405, + 485, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 485, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 485, + 417 + ], + "type": "text", + "content": "3.1. TAD Formulation and Notation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": "TAD aims to identify action instances within a video and determine their temporal boundaries. Given an untrimmed video represented as " + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "inline_equation", + "content": "V = \\{I_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": " denotes the frame at time " + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": ", TAD predicts a set of action instances " + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\Phi_V = \\{\\phi_i = (t_{s_i}, t_{e_i}, c_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": " is the number of action instances; for each instance " + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\phi_i" + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "inline_equation", + "content": "t_{s_i}" + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "inline_equation", + "content": "t_{e_i}" + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": " are its starting and ending times, respectively, and " + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 423, + 554, + 508 + ], + "type": "text", + "content": " denotes its category." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 516, + 539, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 539, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 539, + 529 + ], + "type": "text", + "content": "3.2. Feasibility of Reducing TAD model's Depth" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 568 + ], + "type": "text", + "content": "We find that there is a severe redundancy in the depth of TAD models. To verify this, we conduct a series of experiments and summarize the key observations as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "Observation 1: Most computational overhead lies in the clip-level feature extraction backbone. TAD methods typically consist of two main parts: the feature extractor and the detection head. In the feature extraction stage, a sliding window of length " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " is processed by a video feature extractor. For instance, features are extracted from frames " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{I_t, \\dots, I_{t + L}\\}" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": ", and then the window shifts by " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " frames to extract features from frames " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{I_{t + s}, \\dots, I_{t + s + L}\\}" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": ". As the feature extraction process is repeated across the entire video sequence, it incurs substantial computational cost, with each window requiring complex operations to extract features. Once the features are extracted, they are con" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29227" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 340 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 340 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 340 + ], + "type": "image", + "image_path": "c456f149923bb5398299756c2249af0468b11b6005f554c5e0f36d0fae608e22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 350, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 54, + 350, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 555, + 384 + ], + "type": "text", + "content": "Figure 3. The diagram of our progressive block drop method. Our approach adopts a multi-step progressive compression strategy. At each iteration 1) we evaluate the importance of each block and drop the least important block, and 2) we use parameter-efficient tuning techniques, and recover performance by learning from the uncompressed model through feature-level and prediction-level alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 392, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 392, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 295, + 548 + ], + "type": "text", + "content": "catenated along the temporal dimension, and information is fused across the spatial dimension by averaging the features. These fused features are then passed to the detection head. Ultimately, the volume of data fed into the detection head is much smaller compared to the data processed by the feature extractor. As shown in Figure 1, the computational cost of the feature extractor accounts for " + }, + { + "bbox": [ + 54, + 392, + 295, + 548 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 54, + 392, + 295, + 548 + ], + "type": "text", + "content": " of the total forward inference cost of the model. Given the high computational cost of feature extraction, optimizing or reducing the complexity of this stage presents a significant opportunity for improving model efficiency. Therefore, to reduce the overall model computation, the primary focus should be on compressing the feature extractor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 553, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 553, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 54, + 553, + 295, + 650 + ], + "type": "text", + "content": "Observation 2: There exist blocks that have a minor discrepancy between their input and output. Additionally, by analyzing the feature differences between consecutive blocks, as depicted in Figure 2, we find that some blocks exhibit very similar representation capabilities, suggesting redundancy in the model. These findings highlight the feasibility of applying the block drop method to reduce computational complexity without sacrificing performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 295, + 715 + ], + "type": "text", + "content": "Observation 3: The removal of certain blocks has a negligible impact on TAD performance. To investigate the potential for model compression, we first assess the feasibility of dropping blocks by examining the impact of removing individual blocks on model performance. As shown" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 392, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 429 + ], + "type": "text", + "content": "in Figure 2, our experiments reveal that removing certain blocks led to only minimal performance degradation, indicating that block drop could be a viable approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 430, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 598 + ], + "type": "text", + "content": "Observation 4: Simultaneously removing multiple blocks at a time results in a significant performance drop. In our preliminary exploration, we attempt to directly drop multiple blocks and train the model to observe whether it could recover its performance. We randomly select three blocks to drop all at once and then attempt fine-tuning. As shown in Table 3, the best performance achieved is " + }, + { + "bbox": [ + 313, + 430, + 556, + 598 + ], + "type": "inline_equation", + "content": "0.67\\%" + }, + { + "bbox": [ + 313, + 430, + 556, + 598 + ], + "type": "text", + "content": " lower than the uncompressed model, exhibiting a noticeable performance drop, which is an undesirable outcome in practical applications. Therefore, we adopt a multi-step progressive block dropping strategy rather than removing multiple blocks at once. This approach allows the information from the removed blocks to be more effectively compressed into the pruned model, resulting in better performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 609, + 451, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 451, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 451, + 624 + ], + "type": "text", + "content": "4. Progressive Block Drop" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we design a compression method specifically tailored for TAD models. Drawing from our analysis in Section 3.2, our focus is on reducing the computational cost of the feature extraction backbone while maintaining or even enhancing performance. To achieve this, we propose a multi-step progressive block-level dropping method, where each step comprises two stages. First, the Block Selection" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29228" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "Evaluator automatically selects the blocks to be dropped, resulting in a subnet. Then, Cross-Depth Alignment enables parameter-efficient learning from the uncompressed model to maintain TAD performance. By iteratively applying these steps, we progressively obtain a pruned TAD backbone that can be directly deployed for inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 151, + 200, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 200, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 200, + 163 + ], + "type": "text", + "content": "4.1. Block Selection Evaluator" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "content": "Experimental results presented in Section 3.2 demonstrate that simultaneously removing multiple blocks significantly degrades the performance of TAD. Inspired by prior work on curriculum learning [2, 63], we propose a progressive block-drop strategy. However, selecting the appropriate blocks to drop is a non-trivial task. To address this challenge, we introduce a block selection evaluator " + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "content": " to select the blocks to be dropped automatically." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "content": "Without loss of generality, let " + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "content": " denote the initial uncompressed model trained on the training set, and let " + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "content": " represent the model input to " + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "inline_equation", + "content": "f_E" + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "inline_equation", + "content": "B_{t-1} = \\{b_1, b_2, \\ldots, b_K\\}" + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "content": " as the set of all blocks within " + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "content": " is the total number of blocks. At iteration " + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 263, + 296, + 348 + ], + "type": "text", + "content": ", the collection of sub-networks resulting from dropping each single block is denoted as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 354, + 295, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 354, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 92, + 354, + 295, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {t} = \\left\\{M _ {t, k} ^ {\\text {s u b}} = M _ {t - 1} \\backslash b _ {k} \\mid b _ {k} \\in B _ {t - 1} \\right\\}. \\tag {1}", + "image_path": "2493289d277e14a5b398100c970b4146e44b28319a7630055e7f0737332e9fe1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "text", + "content": "To automatically select blocks for pruning, we employ the evaluation function " + }, + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "text", + "content": " to assess each subnet. By measuring the performance difference between each subnet and the uncompressed TAD model " + }, + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "text", + "content": " on the training set, " + }, + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "text", + "content": " outputs an importance score for each subnet; the smaller the performance difference, the higher the importance score assigned. This process can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 466, + 295, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 466, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 133, + 466, + 295, + 480 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {t} = f _ {E} \\left(\\mathcal {S} _ {t}, D _ {\\text {t r a i n}}\\right). \\tag {2}", + "image_path": "819c5986db5a5be5140c0298f292c537dc3cda676b4d3a211aea4f7737ff042d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "text", + "content": "From the set of importance scores " + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "text", + "content": ", we select the subnet " + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "inline_equation", + "content": "M_{t,*}^{sub}" + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "text", + "content": " with the highest importance score " + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "text", + "content": ", resulting in the pruned model " + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 55, + 486, + 296, + 594 + ], + "type": "text", + "content": " can also compute the difference between the input and output features of each removed block or the training loss of each subnet on the training set; in both cases, a smaller difference corresponds to a higher importance score. In our experiments (Section F.1 of the appendix), we find that selecting subnets based on the mAP metric from the training set yields the best results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 296, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 625 + ], + "type": "text", + "content": "4.2. Parameter-Efficient Learning from Uncompressed Model for Performance Recovery" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "After removing one block from the TAD model, a central challenge arises in ensuring that the performance of the reduced-depth model remains comparable to that of the original, uncompressed model (see Section F.2 of appendix). However, training the TAD model on untrimmed videos using full-parameter fine-tuning still demands a substantial amount of GPU memory, as discussed in [40]. For" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 315, + 87, + 555, + 321 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 541, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 541, + 84 + ], + "type": "text", + "content": "Algorithm 1 Proposed Progressive Block Drop Method" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": "Input: Uncompressed model " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": ", model " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": ", blocks " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "B_{t-1} = \\{b_1, b_2, \\ldots, b_K\\}" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": ", training set " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "D_{train}" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": ". // Progressive block drop cycle while Performance of " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "M_{t-1} \\geq M_0" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": " do // Stage 1: select one block to drops Obtain sub-nets by dropping each block: " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_t = \\left\\{M_{t,k}^{\\mathrm{sub}} = M_{t-1} \\setminus b_k \\mid b_k \\in B_{t-1}\\right\\}" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": " Evaluate each sub-net using the evaluator " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "f_E" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t = f_E(\\mathcal{S}_t, D_{train})" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": " Select the subnet with the highest performance: " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "M_t = \\arg \\max(\\mathcal{P}_t)" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": " // Stage 2: performance recovery Insert LoRA modules into " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": " for training: " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "M_t(\\theta) \\rightarrow \\hat{M}_t(\\theta, \\theta_{\\mathrm{LoRA}})" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": " Train the model with Eqn. (6) Merge LoRA parameters into the model: " + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\theta_t \\gets \\theta_t + \\theta_{\\mathrm{LoRA}}'" + }, + { + "bbox": [ + 315, + 87, + 555, + 321 + ], + "type": "text", + "content": " end while" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 342, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 402 + ], + "type": "text", + "content": "example, in our experiments with the VideoMAE-L model we used, setting the batch size to 1 already consumes 13.1 GB of memory. Therefore, we introduce LoRA to reduce memory consumption during training, with detailed experimental results provided in Section F.3 of the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "content": "Before training, given the pruned model " + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "inline_equation", + "content": "M_{t}(\\theta_{t})" + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "content": " represents its parameters, we first follow the LoRA method [8, 26] to add trainable parameters " + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{LORA}}" + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "content": " to the attention layer each block, resulting in the model " + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\hat{M}_{t}(\\theta_{t},\\theta_{\\mathrm{LORA}})" + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "content": ". This approach offers two main benefits: first, during subsequent training, only " + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{LORA}}" + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "content": " is trained, significantly reducing memory consumption; second, after training, " + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{LORA}}" + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "content": " can be merged with " + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 313, + 403, + 554, + 521 + ], + "type": "text", + "content": ", thereby keeping the TAD model's structure unchanged and incurring no additional inference computational overhead." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 522, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 641 + ], + "type": "text", + "content": "Cross-Depth Alignment. The structural differences between the pruned and uncompressed models—resulting from block dropping—pose a significant challenge for recovering performance. More critically, we empirically observed that directly supervising the training of " + }, + { + "bbox": [ + 313, + 522, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 313, + 522, + 554, + 641 + ], + "type": "text", + "content": " with ground truth fails to restore performance (see Section E.5). In other words, relying solely on prediction-level supervision is inadequate for restoring performance. To address this issue, we shift our attention to feature-level supervision and propose a cross-depth alignment strategy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Specifically, if we consider three adjacent blocks, " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{b_{i - 1}, b_i, b_{i + 1}\\}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", as a group, when " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is dropped, the input to " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "b_{i + 1}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " changes (i.e., it becomes the output of " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "b_{i - 1}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "). To maintain performance, it is necessary to: 1) ensure that the input feature distribution of " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "b_{i + 1}^p" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " in the pruned model " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " aligns with that of " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "b_i^u" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " in the uncompressed model " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "; and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29229" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "2) ensure that the output distribution of the " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "-th block in " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " aligns with that of " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". In our implementation, after the " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "-th block is dropped, we align the outputs of the " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "(i - 1)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "-th blocks " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "(f_{b_{i - 1}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "f_{b_{i + 1}})" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " between the pruned and uncompressed models, thereby compressing the information from multiple blocks into a single block. For the features of other blocks, we similarly apply block-wise alignment with the " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". The cross-depth alignment loss can be expressed as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 98, + 186, + 296, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 186, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 98, + 186, + 296, + 219 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = \\frac {1}{I - 1} \\sum_ {m = 1, m \\neq i} ^ {I} \\left(f _ {b _ {m}} ^ {M _ {0}} - f _ {b _ {m}} ^ {\\hat {M} _ {t}}\\right) ^ {2}. \\tag {3}", + "image_path": "1f4a4b762188a5badecc179bc2be27a8a4835cd0a6de9e25282d11841aa88742.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "content": "Action classification and Localization Alignment. The TAD model generates two distinct types of predictions: action categories and temporal bounding boxes. We also align the predictions of the pruned model " + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "content": " with those of the uncompressed model " + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 55, + 224, + 296, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 354 + ], + "type": "text", + "content": "1) Action classification alignment: To align the action class predictions, we employ the Kullback-Leibler (KL) divergence loss " + }, + { + "bbox": [ + 55, + 284, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 55, + 284, + 296, + 354 + ], + "type": "text", + "content": " [31], which measures the divergence between the action class scores of the uncompressed model " + }, + { + "bbox": [ + 55, + 284, + 296, + 354 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 55, + 284, + 296, + 354 + ], + "type": "text", + "content": " and the pruned model " + }, + { + "bbox": [ + 55, + 284, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 55, + 284, + 296, + 354 + ], + "type": "text", + "content": ". The alignment loss is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 356, + 294, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 356, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 115, + 356, + 294, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p c} = \\mathcal {L} _ {\\mathrm {K L}} \\left(p \\left(z _ {M _ {0}}\\right), p \\left(z _ {\\hat {M} _ {t}}\\right)\\right), \\tag {4}", + "image_path": "930254dcc4dafc2c04de3740e0ef83eff43f5ce11294dc3a23dbc104537ee0b0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "inline_equation", + "content": "z_{M_0}" + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "inline_equation", + "content": "z_{\\hat{M}_t}" + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "text", + "content": " denote the logits of the uncompressed and pruned models, respectively, and " + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "inline_equation", + "content": "p(z_{M_0})" + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "inline_equation", + "content": "p(z_{\\hat{M}_t})" + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "text", + "content": " represent the class probabilities derived from the logits via the softmax function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "content": "2) Boundary regression alignment: For aligning the boundary predictions, we compute the Generalized Intersection over Union (GIoU) loss " + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GIoU}}" + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "content": " [52] between the predicted temporal bounding boxes of the uncompressed model " + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "content": " and the pruned model " + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "content": ". The alignment loss for boundary regression is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 497, + 295, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 497, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 104, + 497, + 295, + 530 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r} = \\frac {1}{N _ {p}} \\sum_ {i = 1} ^ {N _ {p}} \\mathcal {L} _ {\\mathrm {G I o U}} \\left(D _ {i} ^ {M _ {0}}, D _ {i} ^ {\\hat {M} _ {t}}\\right), \\tag {5}", + "image_path": "e9f91540938cb380177d164151242f3c4365bee84c28c5116e0fbdc20e0504ed.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "text", + "content": " denotes the number of predicted positive temporal bounding boxes, and " + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "inline_equation", + "content": "D_{i}^{M_{0}}" + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "inline_equation", + "content": "D_{i}^{\\hat{M}_{t}}" + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "text", + "content": "-th predicted temporal bounding boxes for models " + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 55, + 534, + 296, + 610 + ], + "type": "text", + "content": ", respectively. This dual alignment ensures that the pruned model learns both the correct action classification and the accurate temporal boundaries, which are crucial for TAD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "text", + "content": "Total Loss Function. We incorporate the alignment losses, along with the commonly used classification loss " + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "text", + "content": " and regression loss " + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 611, + 296, + 647 + ], + "type": "text", + "content": " in TAD [40], as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 652, + 295, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 652, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 89, + 652, + 295, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {p c} + \\mathcal {L} _ {p r} + \\mathcal {L} _ {f} + \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e g}}. \\tag {6}", + "image_path": "c6514a1aa02b9818dda2e992a7cd642eecc25460f1f8de56f7abc64d393db497.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 671, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 296, + 695 + ], + "type": "text", + "content": "After training, the LoRA parameters " + }, + { + "bbox": [ + 55, + 671, + 296, + 695 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{LoRA}}^{\\prime}" + }, + { + "bbox": [ + 55, + 671, + 296, + 695 + ], + "type": "text", + "content": " are integrated with the pruned model as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 701, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 701, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 138, + 701, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\theta_ {t} \\leftarrow \\theta_ {t} + \\theta_ {\\mathrm {L o R A}} ^ {\\prime}. \\tag {7}", + "image_path": "513d041a1a18da796c6a6f3f75fcc7d4df6e0cb7516c78486390ee2607d3f2bc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "The updated model " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "M_{t}(\\theta_{t})" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is then input into the block drop selection evaluator at the " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " step for the next iteration, continuing until the performance of the pruned model is no longer recoverable to that of the uncompressed model. Ultimately, the compressed TAD model is deployed, offering a reduced number of parameters and faster inference speed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 157, + 395, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 395, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 395, + 171 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 178, + 554, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 554, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 554, + 238 + ], + "type": "text", + "content": "In the experimental section, we conduct the following experiments: 1) Sections 5.3 and 5.4 evaluate the effectiveness of the proposed method; 2) Section 5.5 validates compatibility with other compression models; 3) Section 5.6 compares inference speed with pruning methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 247, + 486, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 486, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 486, + 258 + ], + "type": "text", + "content": "5.1. Datasets and Evaluation Metric" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 265, + 554, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 337 + ], + "type": "text", + "content": "THUMOS14 [28] is a prominent benchmark dataset in TAD research. It consists of a training set with 200 videos and a test set with 213 videos, covering 20 action categories. This task is notably challenging due to the presence of more than 15 separate action instances per video, with background scenes occupying " + }, + { + "bbox": [ + 313, + 265, + 554, + 337 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 313, + 265, + 554, + 337 + ], + "type": "text", + "content": " of the frames." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 338, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 409 + ], + "type": "text", + "content": "ActivityNet-1.3 [4] is a large-scale benchmark for TAD in untrimmed videos, containing approximately 20,000 videos across 200 activity classes, totaling over 600 hours of content. The dataset is divided into three subsets: 10,024 videos for training, 4,926 for validation, and 5,044 for testing, with an average of 1.65 action instances per video." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 410, + 554, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 481 + ], + "type": "text", + "content": "FineAction [43] is a benchmark dataset for temporal action localization, containing 103K instances across 106 action categories in 17K untrimmed videos. It presents both opportunities and challenges for TAD due to its fine-grained action classes, dense annotations, and frequent co-occurrence of multiple action classes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "type": "text", + "content": "Mean Average Precision (mAP). A predicted temporal bounding box is deemed accurate if its temporal Intersection over Union (tIoU) with the corresponding ground truth instance exceeds a specific threshold and matches the instance's category. For THUMOS14, the tIoU thresholds range from " + }, + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\{0.3, 0.4, 0.5, 0.6, 0.7\\}" + }, + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "type": "text", + "content": ", with mAP@tIoU=0.5 reported for comparison. In contrast, on ActivityNet-1.3 and FineAction, tIoU thresholds are sampled from " + }, + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\{0.5: 0.05: 0.95\\}" + }, + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "type": "text", + "content": ", and we report the mean mAP over tIoU thresholds from 0.5 to 0.95, in increments of 0.05." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 611, + 447, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 447, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 447, + 624 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "We use the ActionFormer [70] detector for prediction, employing two Transformer-based feature extractors: VideoMAE-S with 12 blocks and VideoMAE-L [60] with 24 blocks, both are pre-trained on the K400 dataset [29]. For the THUMOS14 dataset, the input videos are randomly truncated to a fixed length of 768 frames, while for the ActivityNet-1.3 and FineAction dataset, the input videos are" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29230" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 96, + 512, + 167 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Evaluation of our method in terms of computational complexity, inference time and mAP with VideoMAE-S as the backbone. Our approach reduces the computational cost by approximately " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ", achieves around " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "1.3\\mathrm{x}" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " inference speedup, and improves the performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 96, + 512, + 167 + ], + "lines": [ + { + "bbox": [ + 97, + 96, + 512, + 167 + ], + "spans": [ + { + "bbox": [ + 97, + 96, + 512, + 167 + ], + "type": "table", + "html": "
Model (Drop #Block)THUMOS14ActivityNet-1.3
MACs (G)Inference Time (ms)mAP (tIoU=0.5)MACs (G)Inference Time (ms)mAP (Avg)
Baseline286.3104.970.43274.1103.837.75
1263.5 (92.0%)98.4 (93.8%)71.06 (↑ 0.63)251.3 (91.7%)95.1 (91.6%)37.94 (↑ 0.19)
2240.8 (84.1%)89.8 (85.6%)71.37 (↑ 0.94)228.5 (83.4%)86.8 (83.6%)37.81 (↑ 0.06)
3218.0 (76.1%)81.2 (77.4%)70.47 (↑ 0.04)205.8 (75.1%)79.6 (76.7%)37.77 (↑ 0.02)
4195.2 (68.2%)73.6 (70.2%)69.65 (↓ 0.78)183.0 (66.8%)72.1 (69.5%)37.72 (↓ 0.03)
", + "image_path": "273078fa422663af285d282dc2c1e9c99862013fbb6e081aa274ffbbc16386fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 220, + 294, + 301 + ], + "blocks": [ + { + "bbox": [ + 55, + 176, + 295, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 295, + 220 + ], + "type": "text", + "content": "Table 2. Evaluation of our method on deeper networks in terms of computational complexity, inference time and mAP with VideoMAE-L (24 blocks) as the backbone. Our method is generally effective and larger models may benefit more from it." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 220, + 294, + 301 + ], + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 301 + ], + "type": "table", + "html": "
Drop #BlockMACs (G)Inference Time (ms)mAP (tIoU=0.5)
Baseline3886.9795.976.01
13725.5 (95.8%)765.4 (96.2%)77.97 (↑ 1.96)
23564.1 (91.7%)731.4 (91.9%)78.00 (↑ 1.99)
33402.7 (87.5%)704.3 (88.5%)78.29 (↑ 2.28)
43241.4 (83.4%)671.3 (84.3%)77.86 (↑ 1.85)
53080.0 (79.2%)647.1 (81.3%)77.57 (↑ 1.56)
62918.6 (75.1%)602.6 (75.7%)77.25 (↑ 1.24)
", + "image_path": "9c13c104e772ee132dcddfc7605263e937f7ad48fb0ceef696e3310987d06514.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 296, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 343 + ], + "type": "text", + "content": "resized to a fixed length of 768 frames, with the frame resolution set to " + }, + { + "bbox": [ + 55, + 306, + 296, + 343 + ], + "type": "inline_equation", + "content": "160^2" + }, + { + "bbox": [ + 55, + 306, + 296, + 343 + ], + "type": "text", + "content": " by default for both datasets. Please refer to the supplementary for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 349, + 286, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 286, + 363 + ], + "type": "text", + "content": "5.3. Effectiveness of Our Progressive Block Drop" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 366, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 295, + 558 + ], + "type": "text", + "content": "We compare the model trained with the progressive block drop method to the uncompressed model from three aspects. 1) Computational Complexity (MACs): The computational complexity is calculated based on the number of multiply-accumulate operations at each layer of the model. 2) Inference Time: The inference time measures the average duration for processing each input sample, where each sample consists of 768 frames with a resolution of " + }, + { + "bbox": [ + 54, + 366, + 295, + 558 + ], + "type": "inline_equation", + "content": "160^{2}" + }, + { + "bbox": [ + 54, + 366, + 295, + 558 + ], + "type": "text", + "content": ". 3) mAP: The results, as shown in Table 1, demonstrate that on both datasets, our method achieves an approximately " + }, + { + "bbox": [ + 54, + 366, + 295, + 558 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 54, + 366, + 295, + 558 + ], + "type": "text", + "content": " reduction in MACs while outperforming the uncompressed model in terms of performance. This highlights the superiority of our approach, as it effectively reduces computational complexity while simultaneously enhancing the performance of the compressed model—an accomplishment that is difficult for other compression methods to achieve." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Drop blocks on a deeper backbone. We explore the performance of the block drop method when the depth of the feature extractor is increased from 12 to 24, i.e., when transitioning from VideoMAE-S to VideoMAE-L. The results, as shown in Table 2, reveal that with a " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " reduction in MACs, the model achieves a " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.24\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " performance improvement. Notably, when the same fraction of blocks (1/4) is dropped, the performance of VideoMAE-L improves by " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.24\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": ", whereas VideoMAE-S only shows a modest improvement of " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.04\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " (as shown in Table 1). This suggests that in models with higher computational complexity, there is more redundant information, and the progressive block drop method is effective at eliminating such redundancy." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 209, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 313, + 175, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 209 + ], + "type": "text", + "content": "Table 3. Comparison with randomly dropping blocks. Dropping three randomly selected blocks at once performs worse than progressively dropping." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 209, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 235 + ], + "type": "table", + "html": "
Drop Block ID[2,5,9][0,4,7][1,3,8][6,7,10][6,7,10] + ours
mAP68.8868.3169.7668.9170.47
", + "image_path": "bc705953f9694903d2bd1a2605e26ea10c14259abd836f023a833dbe495cb424.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "content": "Dropping multiple blocks at one time vs Progressive dropping. To demonstrate the necessity of adopting a multi-step progressive drop in our proposed block drop method, we conduct experiments where multiple blocks are dropped simultaneously. Specifically, we randomly select three different blocks to drop and train the model using the training method described in Section 4.2. The results, as shown in Table 3, indicate that directly dropping three blocks yields performance ranging from " + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "inline_equation", + "content": "68.31\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "inline_equation", + "content": "69.76\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "content": ", with the best result still " + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "inline_equation", + "content": "0.67\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "content": " lower than that of the uncompressed model. Meanwhile, using the progressive block drop method on VideoMAE-S and the THUMOS14 dataset, the model's performance after dropping the first three layers [7, 10, 6] reaches " + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "inline_equation", + "content": "70.47\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "content": ". Comparing this with the result of directly dropping these three layers, it is evident that compressing the model via progressive dropping achieves better performance recovery (" + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "inline_equation", + "content": "70.47\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "inline_equation", + "content": "68.91\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "content": "), further underscoring the importance of the multi-step progressive drop. Therefore, our proposed multi-step progressive block drop is crucial for performance recovery." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 493, + 440, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 440, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 440, + 504 + ], + "type": "text", + "content": "5.4. Extended Evaluations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "content": "Evaluation on additional TAD architectures. To validate the generalizability of our method, we apply it to two additional TAD architectures: the single-stage AdaTAD [40] model with a VideoMAE-S backbone and the two-stage ActionFormer [70] model with pre-extracted I3D features. Table 4 shows that in AdaTAD, pruning two blocks improves accuracy by " + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "inline_equation", + "content": "0.08\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "content": " and reduces MACs to " + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "content": ". Similarly, the ActionFormer model gains " + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "inline_equation", + "content": "0.49\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "content": " in mAP while reducing MACs to " + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "inline_equation", + "content": "59.7\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "content": ". These results underscore our method's effectiveness in compressing diverse TAD architectures while enhancing performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Evaluation on the fine-grained FineActions dataset. To further corroborate the versatility of our approach, we conduct experiments on the fine-grained FineActions [43] dataset using a VideoMAE-S backbone with an Action-Former [70] head. As shown in Table 5, pruning two blocks slightly improves mAP (" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\uparrow 0.04\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "), while pruning" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "29231" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 93, + 294, + 150 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 4. Evaluation of our method on AdaTAD and ActionFormer. Our method is effective across different TAD architectures." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 93, + 294, + 150 + ], + "lines": [ + { + "bbox": [ + 58, + 93, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 93, + 294, + 150 + ], + "type": "table", + "html": "
Drop #BlockAdaTAD [40]ActionFormer [70]
mAP (tIoU=0.5)MACs (G)mAP (tIoU=0.5)MACs (G)
Baseline72.43324.872.6545.2
172.91 (↑ 0.48)298.8 (92.0%)73.21 (↑ 0.56)37.9 (83.8%)
272.51 (↑ 0.08)272.8 (84.0%)73.28 (↑ 0.63)30.6 (67.7%)
371.68 (↓ 0.75)246.9 (76.0%)73.14 (↑ 0.49)27.0 (59.7%)
", + "image_path": "115d2fb4b680febd81d837bc1732128e44c9e067619c1b44a0d4a19333c914bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 187, + 293, + 241 + ], + "blocks": [ + { + "bbox": [ + 55, + 153, + 295, + 186 + ], + "lines": [ + { + "bbox": [ + 55, + 153, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 295, + 186 + ], + "type": "text", + "content": "Table 5. Evaluation of our method on fine-grained FineActions [43]. Our approach continues to exhibit efficacy even in scenarios involving more fine-grained actions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 187, + 293, + 241 + ], + "lines": [ + { + "bbox": [ + 58, + 187, + 293, + 241 + ], + "spans": [ + { + "bbox": [ + 58, + 187, + 293, + 241 + ], + "type": "table", + "html": "
Drop #BlockmAP (Avg)MACs (G)Inference Time (ms)
Baseline18.67280.8210.2
118.77 (↑ 0.10)257.7 (91.8%)197.0 (93.7%)
218.71 (↑ 0.04)234.6 (83.5%)185.0 (88.0%)
318.43 (↓ 0.24)211.5 (75.3%)173.0 (82.3%)
", + "image_path": "e7e509081bb2d38ab2aef056b137336b9b60e80dcecaf9d4567332003f0da97d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 59, + 266, + 293, + 335 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 294, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 294, + 266 + ], + "type": "text", + "content": "Table 6. Results of pruning Mamba model [6] on Charades-STA. Our method is effective across various tasks and architectures." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 266, + 293, + 335 + ], + "lines": [ + { + "bbox": [ + 59, + 266, + 293, + 335 + ], + "spans": [ + { + "bbox": [ + 59, + 266, + 293, + 335 + ], + "type": "table", + "html": "
Drop #BlockParameters (M)R1@0.3R1@0.5
Baseline60.569.4658.58
153.4 (88.4%)72.04 (↑ 2.58)58.79 (↑ 0.21)
246.4 (76.7%)71.75 (↑ 2.29)58.92 (↑ 0.34)
339.3 (65.1%)71.94 (↑ 2.48)58.79 (↑ 0.21)
432.3 (53.4%)70.83 (↑ 1.37)57.72 (↓ 0.86)
", + "image_path": "f49fb1fbf59584be3c34fbb50b20c1392f56c64524f4c4e0d9d6a8382b48e91f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 347, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 418 + ], + "type": "text", + "content": "three blocks causes only a minor drop " + }, + { + "bbox": [ + 55, + 347, + 295, + 418 + ], + "type": "inline_equation", + "content": "(\\downarrow 0.24\\%)" + }, + { + "bbox": [ + 55, + 347, + 295, + 418 + ], + "type": "text", + "content": ". Meanwhile, MACs are reduced to " + }, + { + "bbox": [ + 55, + 347, + 295, + 418 + ], + "type": "inline_equation", + "content": "83.5\\%" + }, + { + "bbox": [ + 55, + 347, + 295, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 347, + 295, + 418 + ], + "type": "inline_equation", + "content": "75.3\\%" + }, + { + "bbox": [ + 55, + 347, + 295, + 418 + ], + "type": "text", + "content": ", with a notable decrease in inference time. These results demonstrate that our progressive block drop strategy effectively preserves detection accuracy while significantly reducing computational cost, even for fine-grained action detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "content": "Evaluation on the natural language localization task. Finally, to test our method beyond conventional TAD, we apply progressive block drop to the grounding head of the Mamba model [6] on Charades-STA [16] for natural language localization. Selective pruning is performed on the grounding head. As shown in Table 6, removing three out of six blocks reduces the parameter count to " + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "inline_equation", + "content": "65.1\\%" + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "content": " of the baseline while increasing performance—R1@0.3 by " + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "inline_equation", + "content": "2.48\\%" + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "content": " and R1@0.5 by " + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "inline_equation", + "content": "0.21\\%" + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "content": ". These results demonstrate our method's generalizability across tasks and models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 543, + 253, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 253, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 253, + 555 + ], + "type": "text", + "content": "5.5. Compatible with Pruning Techniques" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "text", + "content": "We apply the pruning method [13] to models after block drop, considering three setups: 1) pruning, 2) our method, and 3) our method with pruning. As shown in Figure 4, at nearly identical accuracy (70.22% vs. 70.21%), combining our method with pruning reduces MACs by 30%, while pruning alone achieves only 10%. This highlights our method's potential for further integration with pruning to accelerate the model. Additionally, compatibility with sparse activations is analyzed in Section E of the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 287, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 287, + 685 + ], + "type": "text", + "content": "5.6. Comparison of Inference Time with Pruning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We test the inference times of models compressed via pruning and our method when both achieved the same reduction" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 325, + 68, + 541, + 210 + ], + "blocks": [ + { + "bbox": [ + 325, + 68, + 541, + 210 + ], + "lines": [ + { + "bbox": [ + 325, + 68, + 541, + 210 + ], + "spans": [ + { + "bbox": [ + 325, + 68, + 541, + 210 + ], + "type": "image", + "image_path": "8c5c4c61098f9bbfdb4d42570d6aa05fd5ef7792961f3ab2ca554d55a8531ffb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 211, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 257 + ], + "type": "text", + "content": "Figure 4. Comparison and compatibility with the pruning method. Bubble size represents the model's computational complexity. Combining our method with pruning enables further " + }, + { + "bbox": [ + 313, + 211, + 553, + 257 + ], + "type": "inline_equation", + "content": "1.3 \\times" + }, + { + "bbox": [ + 313, + 211, + 553, + 257 + ], + "type": "text", + "content": " acceleration, demonstrating our compatibility with pruning." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 317, + 306, + 552, + 352 + ], + "blocks": [ + { + "bbox": [ + 313, + 262, + 553, + 305 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 553, + 305 + ], + "type": "text", + "content": "Table 7. Comparisons of inference time between pruning methods and ours under the same computational complexity on THUMOS14, VideoMAE-S backbone. Reducing depth yields better acceleration than reducing width." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 306, + 552, + 352 + ], + "lines": [ + { + "bbox": [ + 317, + 306, + 552, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 306, + 552, + 352 + ], + "type": "table", + "html": "
ModelMACs (G)Inference Time (ms)Speedup
Uncompressed286.3104.91.00×
Pruning218.296.61.09×
Ours218.081.21.29×
", + "image_path": "ffaebe61309119ac54aff898807ab835446f0777291a29cf7c7340320309276c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 361, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 469 + ], + "type": "text", + "content": "in MACs. The inference time is measured as the average duration per input sample, each consisting of 768 frames with a resolution of " + }, + { + "bbox": [ + 313, + 361, + 555, + 469 + ], + "type": "inline_equation", + "content": "160^2" + }, + { + "bbox": [ + 313, + 361, + 555, + 469 + ], + "type": "text", + "content": ". As shown in Table 7, under similar MACs (218.2G vs 218.0G), pruning achieves a " + }, + { + "bbox": [ + 313, + 361, + 555, + 469 + ], + "type": "inline_equation", + "content": "1.09 \\times" + }, + { + "bbox": [ + 313, + 361, + 555, + 469 + ], + "type": "text", + "content": " speedup, while the block drop method achieves a " + }, + { + "bbox": [ + 313, + 361, + 555, + 469 + ], + "type": "inline_equation", + "content": "1.29 \\times" + }, + { + "bbox": [ + 313, + 361, + 555, + 469 + ], + "type": "text", + "content": " speedup. This indicates that compressing the model by reducing network depth is more effective for acceleration than reducing network width because shallow and wide models are more hardware-friendly architectures [39]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 479, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 388, + 491 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this work, we have proposed a progressive block drop method to compress TAD models from the depth perspective. This method consists of two main steps: 1) The block selection evaluator automatically selects blocks to be dropped based on their importance; 2) Parameter-efficient learning combined with three alignment techniques is used to recover performance. On the THUMOS14 and ActivityNet-1.3 datasets, we have achieved a reduction of about " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " in computational cost while improving the pruned model's performance above that of the uncompressed model. We have further validated its generalizability across TAD architectures, fine-grained datasets, and other tasks, demonstrating compatibility with existing compression techniques. Additionally, we have demonstrated the compatibility of our method with other compression techniques through experiments. We hope that our work paves the way for further advancements in the compression and acceleration of TAD models." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29232" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 204 + ], + "type": "text", + "content": "Acknowledgements. This work was partially supported by National Natural Science Foundation of China (NSFC) under Grants 62202311, the Guangdong Basic and Applied Basic Research Foundation under Grants 2023A1515011512, Excellent Science and Technology Creative Talent Training Program of Shenzhen Municipality under Grant RCBS20221008093224017, the Shenzhen Natural Science Foundation (the Stable Support Plan Program) under Grant 20220809180405001, Key Scientific Research Project of the Department of Education of Guangdong Province 2024ZDZX3012." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 217, + 115, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 115, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 115, + 229 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 237, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 237, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 237, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 237, + 294, + 291 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 294, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 294, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 61, + 294, + 294, + 338 + ], + "type": "text", + "content": "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning (ICML), pages 41-48, 2009. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 340, + 294, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 294, + 394 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 294, + 394 + ], + "type": "text", + "content": "[3] Shyamal Buch, Victor Escorcia, Chuanqi Shen, Bernard Ghanem, and Juan Carlos Niebles. Sst: Single-stream temporal action proposals. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 2911-2920, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 396, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 294, + 451 + ], + "type": "text", + "content": "[4] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 961-970, 2015. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 453, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 294, + 518 + ], + "type": "text", + "content": "[5] Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A Ross, Jia Deng, and Rahul Sukthankar. Rethinking the faster r-cnn architecture for temporal action localization. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1130–1139, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 521, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 294, + 575 + ], + "type": "text", + "content": "[6] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024.8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 578, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 294, + 621 + ], + "type": "text", + "content": "[7] Shi Chen and Qi Zhao. Shallowing deep networks: Layerwise pruning based on feature representations. IEEE transactions on pattern analysis and machine intelligence, 41(12): 3048-3056, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 677 + ], + "type": "text", + "content": "[8] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fine-tuning of long-context large language models. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 294, + 712 + ], + "type": "text", + "content": "[9] Emily L Denton, Wojciech Zaremba, Joan Bruna, Yann LeCun, and Rob Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. Advances in" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "neural information processing systems (NeurIPS), 27, 2014. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[10] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning (ICML), pages 2793-2803. PMLR, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[11] Amir Ben Dror, Niv Zehngut, Avraham Raviv, Evgeny Artyomov, Ran Vitek, and Roy Jevnisek. Layer folding: Neural network depth reduction using activation linearization. arXiv preprint arXiv:2106.09309, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[12] Sara Elkerdawy, Mostafa Elhoushi, Abhineet Singh, Hong Zhang, and Nilanjan Ray. To filter prune, or to layer prune, that is the question. In proceedings of the Asian conference on computer vision (ACCV), 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 243, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 296 + ], + "type": "text", + "content": "[13] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 16091-16101, 2023. 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 299, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 341 + ], + "type": "text", + "content": "[14] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations (ICLR), 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "type": "text", + "content": "[15] Yonggan Fu, Haichuan Yang, Jiayi Yuan, Meng Li, Cheng Wan, Raghuraman Krishnamoorthi, Vikas Chandra, and Yingyan Lin. Depthshrinker: a new compression paradigm towards boosting real-hardware efficiency of compact neural networks. In International Conference on Machine Learning (ICML), pages 6849-6862. PMLR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[16] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "text", + "content": "[17] Lianli Gao, Zhao Guo, Hanwang Zhang, Xing Xu, and Heng Tao Shen. Video captioning with attention-based LSTM and semantic consistency. IEEE Transactions on Multimedia, 19(9):2045-2055, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[18] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. In International Conference on Learning Representations (ICLR), 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "text", + "content": "[19] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[20] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems (NeurIPS), 28, 2015. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[21] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 1389-1397, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[22] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29233" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 4340-4349, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 116 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 116 + ], + "type": "text", + "content": "[23] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. stat, 1050:9, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "text", + "content": "[24] Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. Journal of Machine Learning Research, 22(241):1-124, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 237 + ], + "type": "text", + "content": "[25] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 1314-1324, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[26] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2021. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "text", + "content": "[27] Yupan Huang, Qi Dai, and Yutong Lu. Decoupling localization and classification in single shot temporal action detection. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1288-1293. IEEE, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "type": "text", + "content": "[28] YG Jiang, J Liu, A Roshan Zamir, G Toderici, I Laptev, M Shah, and R Sukthankar. Thumos challenge: Action recognition with a large number of classes, 2014. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "type": "text", + "content": "[29] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "text", + "content": "[30] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems (NeurIPS), 25, 2012. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "type": "text", + "content": "[31] Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1): 79-86, 1951. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "type": "text", + "content": "[32] Mark Kurtz, Justin Kopinsky, Rati Gelashvili, Alexander Matveev, John Carr, Michael Goin, William Leiserson, Sage Moore, Nir Shavit, and Dan Alistarh. Inducing and exploiting activation sparsity for fast inference on deep neural networks. In International Conference on Machine Learning (ICML), pages 5533-5543. PMLR, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "text", + "content": "[33] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. Tvqa: Localized, compositional video question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1369-1379, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "text", + "content": "[34] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[35] Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 3320-3329, 2021. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[36] Tianwei Lin, Xu Zhao, and Zheng Shou. Single shot temporal action detection. In Proceedings of the 25th ACM international conference on Multimedia, pages 988-996, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[37] Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 3889-3898, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 172, + 553, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 172, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 553, + 236 + ], + "type": "text", + "content": "[38] Ji Liu, Dehua Tang, Yuanxian Huang, Li Zhang, Xiaocheng Zeng, Dong Li, Mingjie Lu, Jinzhang Peng, Yu Wang, Fan Jiang, et al. Updp: A unified progressive depth pruner for cnn and vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13891-13899, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 238, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 553, + 293 + ], + "type": "text", + "content": "[39] Ning Liu, Xiaolong Ma, Zhiyuan Xu, Yanzhi Wang, Jian Tang, and Jieping Ye. Autocompress: An automatic dnn structured pruning framework for ultra-high compression rates. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4876-4883, 2020. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 358 + ], + "type": "text", + "content": "[40] Shuming Liu, Chen-Lin Zhang, Chen Zhao, and Bernard Ghanem. End-to-end temporal action detection with 1b parameters across 1000 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18591-18601, 2024. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "text", + "content": "[41] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20010-20019, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "text", + "content": "[42] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[43] Yi Liu, Limin Wang, Yali Wang, Xiao Ma, and Yu Qiao. Fineaction: A fine-grained video dataset for temporal action localization. IEEE transactions on image processing, 31: 6937-6950, 2022. 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[44] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. IEEE Robotics and Automation Letters, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "text", + "content": "[45] Xiaolong Ma, Sheng Lin, Shaokai Ye, Zhezhi He, Linfeng Zhang, Geng Yuan, Sia Huat Tan, Zhengang Li, Deliang Fan, Xuehai Qian, et al. Non-structured dnn weight pruning—is it beneficial in any platform? IEEE transactions on neural networks and learning systems, 33(9):4930–4944, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "text", + "content": "[46] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? Advances in neural information processing systems (NeurIPS), 32, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 690 + ], + "type": "text", + "content": "[47] Seyed Iman Mirzadeh, Keivan Alizadeh-Vahid, Sachin Mehta, Carlo C del Mundo, Oncel Tuzel, Golnoosh Samei, Mohammad Rastegari, and Mehrdad Farajtabar. Relu strikes back: Exploiting activation sparsity in large language models. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "text", + "content": "[48] P Molchanov, S Tyree, T Karras, T Aila, and J Kautz. Pruning convolutional neural networks for resource efficient in" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29234" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "text", + "content": "ference. In 5th International Conference on Learning Representations, ICLR 2017-Conference Track Proceedings, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[49] Alberto Montes, Amaia Salvador Aguilera, Santiago Pascual, and Xavier Giró Nieto. Temporal activity detection in untrimmed videos with recurrent neural networks. In Proceedings of the 1st NIPS Workshop on Large Scale Computer Vision Systems (NeurIPS), pages 1-5, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "text", + "content": "[50] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Proposal-free temporal action detection via global segmentation mask learning. In European Conference on Computer Vision (ECCV), pages 645-662. Springer, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "type": "text", + "content": "[51] AJ Piergiovanni and Michael Ryoo. Temporal gaussian mixture layer for videos. In International Conference on Machine learning (ICML), pages 5152-5161. PMLR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 307 + ], + "type": "text", + "content": "[52] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 658-666, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 364 + ], + "type": "text", + "content": "[53] Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18857-18866, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 410 + ], + "type": "text", + "content": "[54] Zheng Shou, Dongang Wang, and Shih-Fu Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 1049-1058, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[55] Suraj Srinivas, Akshayvarun Subramanya, and R Venkatesh Babu. Training sparse neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR) workshops, pages 138-145, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "text", + "content": "[56] Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 13526-13535, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 587 + ], + "type": "text", + "content": "[57] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2820-2828, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[58] Tuan N Tang, Kwonyoung Kim, and Kwanghoon Sohn. Temporalmaxer: Maximize temporal context with only max pooling for temporal action localization. arXiv preprint arXiv:2303.09055, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[59] Marvin Teichmann, Michael Weber, Marius Zoellner, Roberto Cipolla, and Raquel Urtasun. Multinet: Real-time joint semantic reasoning for autonomous driving. In 2018 IEEE intelligent vehicles symposium (IV), pages 1013-1020. IEEE, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[60] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "for self-supervised video pre-training. Advances in neural information processing systems (NeurIPS), 35:10078-10093, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 149 + ], + "type": "text", + "content": "[61] Binglu Wang, Yongqiang Zhao, Le Yang, Teng Long, and Xuelong Li. Temporal action localization in the deep learning era: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "text", + "content": "[62] Lining Wang, Haosen Yang, Wenhao Wu, Hongxun Yao, and Hujie Huang. Temporal action proposal generation with transformers. arXiv preprint arXiv:2105.12043, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 216 + ], + "type": "text", + "content": "[63] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555-4576, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[64] Huijuan Xu, Abir Das, and Kate Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 5783-5792, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "type": "text", + "content": "[65] Pengtao Xu, Jian Cao, Fanhua Shang, Wenyu Sun, and Pu Li. Layer pruning via fusible residual convolutional block for deep neural networks. arXiv preprint arXiv:2011.14356, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "text", + "content": "[66] Le Yang, Houwen Peng, Dingwen Zhang, Jianlong Fu, and Junwei Han. Revisiting anchor mechanisms for temporal action localization. IEEE Transactions on Image Processing, 29:8535-8548, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[67] Min Yang, Guo Chen, Yin-Dong Zheng, Tong Lu, and Limin Wang. Basictad: an astounding rgb-only baseline for temporal action detection. Computer Vision and Image Understanding, 232:103692, 2023. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 447 + ], + "type": "text", + "content": "[68] Serena Yeung, Olga Russakovsky, Greg Mori, and Li Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 2678–2687, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[69] Runhao Zeng, Wenbing Huang, Mingkui Tan, Yu Rong, Peilin Zhao, Junzhou Huang, and Chuang Gan. Graph convolutional module for temporal action localization in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6209-6223, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "text", + "content": "[70] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision (ECCV), pages 492-510. Springer, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "text", + "content": "[71] Minjia Zhang and Yuxiong He. Accelerating training of transformer-based language models with progressive layer dropping. Advances in neural information processing systems (NeurIPS), 33:14011-14023, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 593, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 656 + ], + "type": "text", + "content": "[72] Chen Zhao, Shuming Liu, Karttikeya Mangalam, and Bernard Ghanem. Re2tal: Rewiring pretrained video backbones for reversible temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10637-10647, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[73] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 2914-2923, 2017. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29235" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 298, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 298, + 117 + ], + "type": "text", + "content": "[74] Yao Zhou, Gary G Yen, and Zhang Yi. Evolutionary shallowing deep neural networks at block levels. IEEE Transactions on Neural Networks and Learning Systems, 33(9): 4635-4647, 2021. 3" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29236" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_content_list.json b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6253b27de3ca325abcd3e8a99cfe251c214c9f9d --- /dev/null +++ b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_content_list.json @@ -0,0 +1,1525 @@ +[ + { + "type": "text", + "text": "Temporal Alignment-Free Video Matching for Few-shot Action Recognition", + "text_level": 1, + "bbox": [ + 104, + 128, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SuBeen Lee, WonJun Moon, Hyun Seok Seong, Jae-Pil Heo* Sungkyunkwan University", + "bbox": [ + 245, + 180, + 730, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{leesb7426，wjun0830，gustjrdl95，jaepilheo}@skku.edu", + "bbox": [ + 248, + 219, + 718, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-Shot Action Recognition (FSAR) aims to train a model with only a few labeled video instances. A key challenge in FSAR is handling divergent narrative trajectories for precise video matching. While the frame- and tuple-level alignment approaches have been promising, their methods heavily rely on pre-defined and length-dependent alignment units (e.g., frames or tuples), which limits flexibility for actions of varying lengths and speeds. In this work, we introduce a novel Temporal Alignment-free Matching (TEAM) approach, which eliminates the need for temporal units in action representation and brute-force alignment during matching. Specifically, TEAM represents each video with a fixed set of pattern tokens that capture globally discriminative clues within the video instance regardless of action length or speed, ensuring its flexibility. Furthermore, TEAM is inherently efficient, using token-wise comparisons to measure similarity between videos, unlike existing methods that rely on pairwise comparisons for temporal alignment. Additionally, we propose an adaptation process that identifies and removes common information across classes, establishing clear boundaries even between novel categories. Extensive experiments demonstrate the effectiveness of TEAM. Codes are available at github.com/leesb7426/TEAM.", + "bbox": [ + 75, + 301, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 209, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, action recognition has achieved remarkable performances with the help of numerous labeled videos [2, 5, 8]. However, those performances are not guaranteed when novel categories are given with few labeled examples [3, 22, 29, 31]. To resolve this problem, Few-Shot Action Recognition (FSAR) has been proposed [31]. Particularly, FSAR employs episodic learning which iteratively simulates the episode of learning a set of new classes. Each episode is given a support set to train a model, composed of a few labeled examples for newly given classes, to distinguish the category of a set of unlabeled videos, coined as a query set.", + "bbox": [ + 75, + 705, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to different class compositions within every episode, metric learning, a classifier-free strategy, is predominantly employed in FSAR [3, 20, 23-25, 27, 28]. Distances between the representations of support and query sets are computed to infer the category of a query video. Yet, naively constructing video representations, such as average pooling, are shown to be poor due to an insufficient understanding of temporal information [3, 31]. Therefore, the previous approaches typically focused on aligning temporal sequences between videos [3, 20, 23-25, 27, 28].", + "bbox": [ + 496, + 268, + 895, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As the base units for temporal alignment between videos, the frame and tuple are commonly employed (illustrated in Fig. 1 (a) and (b)). Both frame- and tuple-based alignment approaches measure pairwise similarities between each unit and then derive an optimal alignment path for all units of query video. This path is regarded as the temporal alignment between support and query videos, the query video is assigned with the class label of the support video with minimum cumulative distance.", + "bbox": [ + 496, + 424, + 895, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although these two temporal alignment paradigms have shown effectiveness, they suffer from inflexibility and inefficiency. The inflexibility arises from their reliance on predefined units, such as frames or tuples, for action representation. These methods typically struggle to capture actions if the pre-defined units do not align with the varying action durations. Moreover, their alignment cost increases quadratically with the number of frames, as shown in Fig. 7. This issue has been overlooked since typical FSAR benchmarking datasets consist mostly of trimmed, short-term videos. The quadratic cost arises because their alignment strategies rely on pairwise comparisons between the units of support and query videos.", + "bbox": [ + 496, + 564, + 895, + 761 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this regard, we introduce a novel Temporal Alignment-free Matching (TEAM) approach that achieves both flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and brute-force alignment for video comparison. Specifically, TEAM represents each video with a fixed set of pattern tokens with cross-attention mechanisms to capture diverse clues within the video instance which are useful for distinguishing categories. Therefore, these pattern tokens flexibly identify discrimina", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5412", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8dc4c5373c602a003d4e4bb9a03cbb3f0feb8d8b0ab464462b94c6dfbc547b0c.jpg", + "image_caption": [ + "Figure 1. Comparison with alignment-based approaches on the \"Diving Cliff\" class. (a) Frame Alignment: For each frame of the query video, the most corresponding frame in the support video is identified, highlighting the importance of precise frame-level alignment. (b) Tuple Alignment: The support and query videos are compared in sub-sequence units to account for variations in action speed. The set of varying tuples is pre-defined. (c) Temporal Alignment-free Matching: Both in support and query, the video features are initially integrated using pattern tokens, which encode globally discriminative patterns. Then, it directly compares the corresponding aggregated features of the support and query videos. Pattern-based aggregation is both efficient and flexible, as it does not require an alignment process and is unaffected by differences in frame count or action speed. Note the text description for each pattern token (green) is intended to provide an intuitive understanding of what each pattern token represents." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 416, + 391 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a21e76e999c3d311f7f35d937d47561d4206e44de7774e3b39fa62d47b29a39a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 87, + 890, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tive features varying lengths across each video, as illustrated in Fig. 1 (c). Furthermore, these pattern tokens are trained to capture globally discriminative clues across the dataset. This is because all videos leverage the entire set of pattern tokens without an alignment process, rather than relying on specific tokens determined by alignment, as shown in Fig. 1 (c). Additionally, this token-wise comparison results in the efficiency of TEAM by eliminating the brute-force alignment process computing pairwise similarities between units for the query video matching.", + "bbox": [ + 75, + 523, + 472, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, pattern tokens may not be fully discriminative for every episode with novel class compositions. To address this, we adapt the class boundaries formed by pattern tokens of support set videos. Particularly, we identify the patterns that commonly appear across classes and remove them from each support video token. This process establishes clear class boundaries, improving class discrimination for query videos.", + "bbox": [ + 75, + 679, + 470, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our contributions are:", + "bbox": [ + 96, + 790, + 318, + 805 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Temporal Alignment-free Matching for FSAR equipped with flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and alignment in video comparison.", + "- We develop an episode adaptation process of encoded pattern tokens. This establishes clear boundaries within" + ], + "bbox": [ + 76, + 810, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the support set for each class composition by excluding information shared with other classes.", + "bbox": [ + 511, + 523, + 890, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our state-of-the-art performances on various benchmarks for FSAR validate the effectiveness of our TEAM.", + "bbox": [ + 500, + 553, + 890, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 595, + 640, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Few-Shot Action Recognition", + "text_level": 1, + "bbox": [ + 500, + 619, + 759, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A predominant learning strategy in Few-Shot Action Recognition (FSAR) is metric learning, which entails inferring the category of a query video based on a pre-defined distance metric [3, 20, 23-25, 27]. Previous methods focused on measuring the distance between video features through the development of alignment algorithms, either at the frame-level [3, 24, 25] or tuple-level [20, 23, 27]. As one of the initial frame alignment methods, OTAM [3] aligns the temporal order of query and support videos at the frame-level to precisely estimate the distance between potential query-support pairs. Yet, as it contains a strong assumption that all actions are in a fixed sequence, HyRSM [24] proposed a flexible alignment metric alleviating the strictly ordered constraints. More recently, MoLo [25] focused on learning long-range temporal perception and motion dynamics through frame-level objectives. On the other hand, the stream of tuple-level alignment is to gain robustness to varying action lengths", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5413", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/64516495370e30de382d59ccb6d635e7389685c1a91aad2073df03ca14bbe8ff.jpg", + "image_caption": [ + "Figure 2. Illustration of our approach using a single pattern token ( $M = 1$ ) in the 3-way 3-shot scenario without query. This figure illustrates our approach, where instance (+) and exclusive (-) pattern tokens represent integrated video features based on the pattern token. (Left) Randomly initialized pattern tokens are optimized with classes in two complementary ways. First, instance pattern tokens (+) are encouraged to cluster with other instances' tokens of the same class while repelling those of other categories. On the other hand, exclusive pattern tokens (-) learn to represent the otherness of each instance by positioning themselves in the embedding space of other classes. (Right) Although these two types of tokens are discriminative for distinguishing base video categories, they may not fully capture the finer details needed for novel categories. To address this, we propose an adaptation process of support pattern tokens for novel classes to refine the class decision boundaries. Note that when multiple pattern tokens are used, these processes run in parallel, with instance and exclusive tokens being compared only within the same pattern token." + ], + "image_footnote": [], + "bbox": [ + 76, + 84, + 890, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and temporal offsets. Tuple alignment has gained attention for its flexibility in encoding actions of different speeds and temporal offsets [20]. Additionally, STRM [23] incorporated a spatio-temporal enrichment module to enhance both object and motion representations. Yet, tuple-based approaches require pre-defined temporal tuples to provide hints for possible forms of actions in videos. In contrast, our approach is flexible since we do not need pre-defined windows to represent action.", + "bbox": [ + 75, + 481, + 472, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Token-based Feature Aggregation", + "text_level": 1, + "bbox": [ + 76, + 633, + 374, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature aggregation generates compact contextual representations, providing both effectiveness and efficiency across various fields [1, 4, 7, 11, 14-16, 18, 22]. There are various strategies to implement feature aggregation including different pooling methods [14, 22]. Among them, our method closely aligns with token-based approaches. Simply put, the stream of token-based aggregation integrates the input context within the learnable tokens. For instance, DETR [4] introduces learnable object queries to aggregate spatial features for object detection, while BLIP-2 [15] and Flamingo [1] use learnable tokens to encode image information for multimodal understanding. Similarly to prior methods, we use learnable tokens to aggregate video information. However, our TEAM differs in that the feature aggregation is implemented in two complementary ways, providing more robust class boundaries. Additionally, we also propose a token adap", + "bbox": [ + 75, + 659, + 473, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tation process to maintain the benefits of capturing discriminative details while encoding instances of novel classes.", + "bbox": [ + 498, + 481, + 893, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 527, + 591, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 551, + 702, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the standard Few-Shot Action Recognition (FSAR), we are given two sets of data: meta-train set for model training $\\mathcal{D}_{base} = \\{(v_i,y_i),y_i\\in C_{base}\\}$ and meta-test set for model testing $\\mathcal{D}_{test} = \\{(v_i,y_i),y_i\\in C_{novel}\\}$ . While $v_{i}$ is a video, $C_{base}$ and $C_{novel}$ denote class sets for meta-training and meta-testing, respectively ( $C_{base}\\cap C_{novel} = \\phi$ ). Generally, the training and testing processes of FSAR are composed of episodes where each episode validates the model's adaptability to a novel class set. Each episode comprises randomly selected $N$ classes and each class consists of $K$ labeled videos and $U$ unlabeled videos, i.e., $N$ -way $K$ -shot episode. Whereas a set of labeled videos is called the support set $S = \\{(v_i,y_i)\\}_{i = 1}^{N\\times K}$ , the other set of unlabeled videos are often dubbed as the query set $Q = \\{(v_i,y_i)\\}_{i = 1}^{N\\times U}$ . In the below sections, we assume $N$ -way 1-shot setting where the number of query video $U$ is also set to 1 for better clarity.", + "bbox": [ + 496, + 575, + 893, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 830, + 609, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The schematic illustration of our method is shown in Fig. 2. To eliminate the need for pre-defined units in action representation, we integrate video features using pattern tokens,", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5414", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which capture globally discriminative features across categories. Specifically, our integration process operates in two complementary ways, generating instance and exclusive pattern tokens. Instance pattern tokens are designed to capture shared patterns among videos of the same class. However, these shared patterns may not always be present in every video of the class. In such cases, relying solely on instance pattern tokens for classification is ineffective. To address this, exclusive pattern tokens are introduced to represent \"otherness\"—absent patterns in a given class. This ensures accurate classification, even when key discriminative patterns are missing from certain videos. While these pattern tokens are effective for base categories, they may lack the discriminative details for novel classes. To overcome this, we propose an adaptation process for both instance and exclusive pattern tokens in support videos, reinforcing clear boundaries to distinguish novel classes within each episode. Notably, pattern tokens are only compared within their type (i.e., token-wise comparison), eliminating the need for an alignment process.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Instance Pattern Tokens", + "text_level": 1, + "bbox": [ + 76, + 404, + 299, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given uniformly sampled $T$ frames from the support and query videos, we independently extract the features for each frame through the image features extractor, as follows:", + "bbox": [ + 75, + 428, + 470, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF = \\left[ g _ {\\theta} \\left(x _ {1}\\right), g _ {\\theta} \\left(x _ {2}\\right), \\dots , g _ {\\theta} \\left(x _ {T}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 486, + 470, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g_{\\theta}$ denotes the image feature extractor $\\mathbb{R}^{H\\times W\\times 3}\\to$ $\\mathbb{R}^D$ and $x_{t}$ means $t$ -th frame of the video. $H$ and $W$ are height and width of the frame while $D$ is size of the features.", + "bbox": [ + 76, + 513, + 468, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike existing works that utilize pre-defined units (i.e., frame or tuple) for representing actions, we integrate video features based on pattern tokens that encode globally discriminative patterns. Initially, we define a pattern pool that consists of $M$ learnable pattern tokens $P = [P_{1}, P_{2}, \\dots, P_{M}]$ . Then, we use a cross-attention mechanism to encode instance patterns within $P$ . Formally, instance pattern tokens $P^{+} \\in \\mathbb{R}^{M \\times D}$ are computed as:", + "bbox": [ + 76, + 561, + 472, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {P} _ {m} ^ {+} = P _ {m} + \\mathrm {C A} (P _ {m}, F, F), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 696, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {m} ^ {+} = \\bar {P} _ {m} ^ {+} + \\operatorname {M L P} \\left(\\bar {P} _ {m} ^ {+}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 729, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{m}$ is the $m$ -th component of $P$ while $\\mathrm{CA}(\\cdot, \\cdot, \\cdot)$ is the cross-attention receiving query, key, and value as inputs. And, $\\mathrm{MLP}(\\cdot)$ indicates a light fully-connected block.", + "bbox": [ + 75, + 758, + 468, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on instance pattern tokens of support set $P^{S+}$ , the probability for the query $P^{Q+}$ to belong to $n$ -th category $p^+(y^Q = n|P^S, P^Q)$ is formulated as follows:", + "bbox": [ + 75, + 804, + 470, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P D} \\left(P _ {n} ^ {S}, P ^ {Q}\\right) = \\sum_ {m = 1} ^ {M} - d \\left(P _ {n, m} ^ {S +}, P _ {m} ^ {Q +}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 862, + 468, + 904 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np ^ {+} (y ^ {Q} = n | P ^ {S}, P ^ {Q}) = \\mathrm {s o f t m a x} (\\mathrm {P D} (P ^ {S}, P ^ {Q}); n), \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 87, + 893, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ is a distance metric, i.e., the cosine distance, while $P_{n,m}^{S+}$ and $P_m^{Q+}$ denote $m$ -th instance pattern tokens of support belonging to category $n$ and query, respectively. In addition, $\\operatorname{softmax}(\\cdot; n)$ represents the $n$ -th component of the softmax function applied to the input.", + "bbox": [ + 498, + 116, + 893, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key advantage of using pattern tokens is achieving window-free video understanding. Unlike previous approaches that represent action with frame- or tuple-based units, pattern tokens are not restricted by specific lengths or speeds. Furthermore, each pattern token learns globally discriminative features for video classification, since all pattern tokens contribute across the dataset rather than relying on specific tokens determined by alignment. This design of pattern tokens increases flexibility in action representation and removes the computational cost associated with alignment.", + "bbox": [ + 496, + 191, + 893, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Exclusive Pattern Tokens", + "text_level": 1, + "bbox": [ + 500, + 356, + 728, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the instance pattern tokens that capture specific patterns in the video, we introduce exclusive pattern tokens, designed to encode the otherness to instance pattern tokens. In other words, the exclusive tokens are located close to the other classes' instance tokens and differentiated from the instance tokens of the same class. To emphasize this distinction, we use subtraction rather than addition in Eq. 2 to directly exclude information related to the anchor class, thus the exclusive pattern tokens represent information contrary to instance ones. Specifically, we define the exclusive-pattern tokens $P^{-} \\in \\mathbb{R}^{M \\times D}$ as follows:", + "bbox": [ + 496, + 378, + 893, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {P} _ {m} ^ {-} = P _ {m} - \\mathrm {C A} \\left(P _ {m}, F, F\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 561, + 890, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {m} ^ {-} = \\bar {P} _ {m} ^ {-} + \\operatorname {M L P} \\left(\\bar {P} _ {m} ^ {-}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 595, + 890, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we define the probability of the query belonging to the $n$ -th category $p^{-}(y^{Q} = n|P^{S}, P^{Q})$ with the exclusive and instance pattern tokens, as follows:", + "bbox": [ + 498, + 625, + 893, + 671 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm{ND}(P_{n}^{S},P^{Q}) = \\min_{\\substack{1\\leq o\\leq N\\\\ o\\neq n}}\\sum_{m = 1}^{M}\\big(-d(P_{o,m}^{S - },P_{m}^{Q + })\\big) \\tag{8} \\\\ - d (P _ {o, m} ^ {S +}, P _ {m} ^ {Q -})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 684, + 890, + 751 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np ^ {-} \\left(y ^ {Q} = n \\mid P ^ {S}, P ^ {Q}\\right) = \\operatorname {s o f t m a x} \\left(\\mathrm {N D} \\left(P ^ {S}, P ^ {Q}\\right); n\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 765, + 890, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To accurately estimate the category of the query using the above probability, the exclusive pattern tokens should be similar to the instance pattern tokens of other classes, while being differentiated from those of the same class. Thus, as learning progresses, the exclusive pattern tokens effectively encode the otherness of the same-class instance pattern tokens, as intended.", + "bbox": [ + 496, + 794, + 893, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5415", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Adaptation of Support Pattern Tokens", + "text_level": 1, + "bbox": [ + 76, + 90, + 408, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the pattern tokens learn globally discriminative patterns across base categories in the training phase, they might not be as effective for novel classes. Therefore, we propose an adaptation process for the support set to achieve a clear separation between categories within an episode.", + "bbox": [ + 75, + 113, + 468, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we remove common information shared with other classes from instance pattern tokens and allocate the discriminative information of other classes into exclusive pattern tokens. To regulate the degree of adaptation for each episode, we use the cosine similarity between pattern tokens of the anchor class and other classes, representing the level of semantic entanglement. The adaptive instance pattern tokens $\\hat{P}^{S+} \\in \\mathbb{R}^{M \\times D}$ are defined as follows:", + "bbox": [ + 75, + 189, + 470, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {n, o, m} ^ {+} = P _ {n, m} ^ {S +} \\cdot P _ {o, m} ^ {S +}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 316, + 468, + 338 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {P} _ {n, o, m} ^ {S +} = P _ {m} + \\left(1 + E _ {n, o, m} ^ {+}\\right) \\mathrm {C A} \\left(P _ {m}, F _ {n} ^ {S}, F _ {n} ^ {S}\\right) \\tag {11} \\\\ - E _ {n, o, m} ^ {+} \\mathrm {C A} \\left(P _ {m}, F _ {o} ^ {S}, F _ {o} ^ {S}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 342, + 468, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} _ {n, m} ^ {S +} = \\frac {1}{N - 1} \\sum_ {o = 1} ^ {N} \\mathbb {1} _ {[ n \\neq o ]} \\left(\\tilde {P} _ {n, o, m} ^ {S +} + \\mathrm {M L P} \\left(\\tilde {P} _ {n, o, m} ^ {S +}\\right)\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 387, + 468, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(\\cdot)$ means the cosine similarity measure between two vectors, and $P_{n,m}^{S+}$ is $m$ -th instance pattern token of $n$ -th class in the support set, while $\\mathbb{1}_{[n \\neq o]}$ is an indicator function that outputs 1 if $n$ is different with $o$ and 0 otherwise.", + "bbox": [ + 75, + 431, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, the adaptive pattern tokens for exclusive ones $\\hat{P}^{S-} \\in \\mathbb{R}^{M \\times D}$ are formulated as follows:", + "bbox": [ + 76, + 492, + 468, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {P} _ {n, o, m} ^ {S -} = P _ {m} - \\left(1 + E _ {n, o, m} ^ {-}\\right) \\mathrm {C A} \\left(P _ {m}, F _ {n} ^ {S}, F _ {n} ^ {S}\\right) \\tag {13} \\\\ + E _ {n, o, m} ^ {-} \\mathrm {C A} \\left(P _ {m}, F _ {o} ^ {S}, F _ {o} ^ {S}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 525, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {P} _ {n, m} ^ {S -} = \\frac {1}{N - 1} \\sum_ {o = 1} ^ {N} \\mathbb {1} _ {[ n \\neq o ]} \\left(\\tilde {P} _ {n, o, m} ^ {S -} + \\mathrm {M L P} \\left(\\tilde {P} _ {n, o, m} ^ {S -}\\right)\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 573, + 468, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. 11 extends Eq. 2 by introducing a process to remove shared information between classes while generating instance tokens. This process suppresses information shared between the anchor class $n$ and the other class $o$ . It achieves this by subtracting $\\mathrm{CA}(P_m, F_o^S, F_o^S)$ from $\\mathrm{CA}(P_m, F_n^S, F_n^S)$ , thereby retaining only the discriminative information for the anchor class $n$ . The extent of suppression is proportional to the degree of information overlap, controlled by $E_{n,o,m}^{+}$ . This ensures that the adaptive instance pattern tokens focus on the discriminative features of novel classes. Similarly, Eq. 13 follows the same principle but instead emphasizes the discriminative features of other classes.", + "bbox": [ + 75, + 617, + 470, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Training Objective and Inference", + "text_level": 1, + "bbox": [ + 76, + 805, + 369, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adjusting the probabilities defined in Eq. 5 and 9 with adaptive pattern tokens for the support set, we define two objective functions for model training as follows:", + "bbox": [ + 75, + 828, + 472, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {+} = - \\mathbb {1} _ {[ y ^ {Q} = n ]} \\log \\left(p ^ {+} \\left(y ^ {Q} = n \\mid \\hat {P} ^ {S}, P ^ {Q}\\right)\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 881, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {-} = - \\mathbb {1} _ {[ y ^ {Q} = n ]} \\log \\left(p ^ {-} \\left(y ^ {Q} = n \\mid \\hat {P} ^ {S}, P ^ {Q}\\right)\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 87, + 892, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the final loss function $\\mathcal{L}$ is defined as follows:", + "bbox": [ + 498, + 112, + 843, + 127 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} ^ {+} + \\mathcal {L} ^ {-}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 140, + 892, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In short, instance pattern tokens are optimized to resemble those of the same class and exclusive tokens of other classes. At the same time, they remain distinct from instance tokens of other classes and exclusive tokens of the same class.", + "bbox": [ + 496, + 169, + 893, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a result, the probability of the query to belong to $n$ -th class $p(y^{Q} = n)$ is defined as follows:", + "bbox": [ + 496, + 229, + 890, + 260 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(y ^ {Q} = n\\right) = \\operatorname {s o f t m a x} \\left(\\mathrm {P D} \\left(\\hat {P} ^ {S}, P ^ {Q}\\right) + \\mathrm {N D} \\left(\\hat {P} ^ {S}, P ^ {Q}\\right); n\\right). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 493, + 271, + 892, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 314, + 633, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Description of the datasets and implementation details are provided in the supplementary material.", + "bbox": [ + 496, + 340, + 890, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 498, + 378, + 857, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Tab. 1, we compare our proposed method, TEAM, with previous FSAR approaches under the 5-way 1-shot and 5-shot scenarios, which are standard evaluation protocols. Using ResNet [9], a conventional backbone network, TEAM achieves state-of-the-art performance across various datasets and shot numbers, except for UCF101 (5-shot). With the ViT backbone [7], TEAM surpasses all other methods in every case. We attribute TEAM's effectiveness to its use of globally shared pattern tokens, which inherently capture discriminative patterns across diverse scenarios. This approach enables both efficient and reliable adaptation to new episodes by leveraging pattern tokens to compare videos through multiple criteria. Additionally, TEAM is not only effective but also significantly more efficient than frame- and tuple-level alignment methods, as shown in Fig. 7. These results demonstrate that TEAM is a robust alternative to existing alignment strategies in general FSAR settings.", + "bbox": [ + 496, + 402, + 893, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further conduct experiments on SSv2-Small, a benchmark with high temporal complexity and subtle motion differences, to validate TEAM's effectiveness in handling complex temporal dynamics. As shown in Tab. 2, TEAM surpasses other methods, except for TATs which use an additional off-the-shelf point tracker (PT). Without PT (TATs w/o PT) which generates tailored object trajectories for temporal actions, TEAM also outperforms TATs even in SSv2-Small. These results suggest that the benefits of TEAM remain robust in the context of temporally challenging actions.", + "bbox": [ + 496, + 659, + 893, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-domain evaluation is widely used to assess a model's generalization capability [5, 6, 12, 21]. It typically leverages knowledge learned from coarse datasets to recognize actions in more specific and nuanced domains, highlighting the model's adaptability. Following this protocol, we compare TEAM with other FSAR methods in Tab. 3. We", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5416", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/43865325d7f9262f53c805ccd0b6dd03d264bd8c06931e5fb31752c81f756783.jpg", + "table_caption": [ + "Table 1. Performance comparison on FSAR benchmark datasets under the 5-way setting. The best results are bolded, while the second-best are underlined. $\\dagger$ indicates that evaluation was conducted in a transductive setting, utilizing relations between queries. * denotes reproduced performances, as results with ViT were unavailable." + ], + "table_footnote": [], + "table_body": "
ModelReferenceBackboneHMDB51KineticsUCF101
1-shot5-shot1-shot5-shot1-shot5-shot
ProtoNet [22]NeurIPS 201754.268.464.577.974.089.6
CMN [31]ECCV 2018--57.376.0--
OTAM [3]CVPR 202054.568.072.284.279.988.9
ARN [29]ECCV 202045.560.663.782.466.383.1
TRX [20]CVPR 202153.175.663.685.978.296.1
TA2N [17]AAAI 202259.773.972.885.881.995.1
STRM [23]CVPR 202252.377.362.986.780.596.9
HyRSM [24]CVPR 202260.376.073.786.183.994.7
MTFAN [26]CVPR 202259.074.674.687.484.895.1
HCL [30]ECCV 202259.176.373.785.882.593.9
Nguyen et al. [19]ECCV 202259.676.974.387.484.995.9
Huang et al. [10]ECCV 202260.177.073.386.471.491.0
SloshNet [27]AAAI 2023-77.5-87.0-97.1
MoLo [25]CVPR 202360.877.474.085.686.095.5
GgHM†[28]ICCV 202361.276.974.987.485.296.3
TEAMCVPR 202562.878.475.188.287.296.2
OTAM* [3]CVPR 202069.383.182.291.692.997.5
TRX* [20]CVPR 202161.382.276.890.485.897.8
STRM* [23]CVPR 202256.783.275.890.890.897.6
MoLo* [25]CVPR 202369.983.282.091.293.697.8
TATs [13]ECCV 202460.077.081.991.192.095.5
TEAMCVPR 202570.985.583.392.994.598.8
", + "bbox": [ + 76, + 136, + 898, + 506 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7203ae29565d2ffa9b4931c0933513f12b24ff1c81b4956d6bfc0435100367c5.jpg", + "table_caption": [ + "Table 2. Performance comparison in SSv2-Small dataset using ViT. TATs w/o PT refers to TATs without the additional off-the-shelf Point Tracker. Symbols match those in Tab. 1" + ], + "table_footnote": [], + "table_body": "
ModelReferenceSSv2-Small
1-shot5-shot
OTAM* [3]CVPR 202044.458.2
TRX* [20]CVPR 202142.059.1
STRM* [23]CVPR 202242.560.4
MoLo* [25]CVPR 202345.158.8
TATs [13]ECCV 202447.964.4
TATs w/o PT [13]ECCV 202444.662.5
TEAMCVPR 202547.263.1
", + "bbox": [ + 78, + 561, + 473, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "train all models on 5-way 5-shot tasks using the Kinetics dataset with the ViT backbone and evaluate them on the same setting across different target datasets. As shown in Tab. 3, TEAM consistently outperforms prior methods, indicating strong generalization beyond the training domain. This demonstrates TEAM's robustness to temporal variations with pattern tokens capturing globally distinct patterns.", + "bbox": [ + 75, + 726, + 470, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 845, + 230, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unless specifically stated, experiments are conducted in the 5-way 1-shot scenario using the ResNet-50.", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cb5f2544d10ec666e6bb351f0565b675b14237b78104dc02c660e6ed19716c5c.jpg", + "table_caption": [ + "Table 3. Performance comparison across cross-domain scenarios. All models are trained on the 5-way 5-shot scenario using the Kinetics dataset with the ViT backbone, and evaluated on the same task across various datasets. Symbols match those in Tab. 1." + ], + "table_footnote": [], + "table_body": "
ModelReferenceHMDB51UCF101
OTAM* [3]CVPR 202077.296.5
TRX* [20]CVPR 202178.197.9
STRM* [23]CVPR 202278.598.0
MoLo* [25]CVPR 202377.797.4
TEAMCVPR 202581.198.3
", + "bbox": [ + 501, + 575, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Component Ablation. Tab. 4 presents the performance when our component is applied sequentially. (a) shows that the use of instance pattern tokens solely achieves decent performances over the methods that rely on pre-defined temporal units and brute-force alignment processes. This validates the advantage of our core intuition that temporal alignment-free matching is beneficial for its flexibility in video understanding. Succeedingly, in (b), the exclusive tokens are shown to yield additional improvements, indicating that the different strategy in learning pattern tokens complements the original approach. Conversely, we observe slight performance drops in (c) where we implement the adaptation process without the management in the adaptation magnitude. This implies the importance of managing the degree", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5417", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/18105e7ec332f817fe42a711e365acc01e72ac1357b4f757fc9036202678b0de.jpg", + "table_caption": [ + "Table 4. Ablation study for each component. $P^{+}$ , $P^{-}$ , and $\\hat{P}$ denote the instance, exclusive, adaptive pattern tokens, respectively, while $E$ means entanglement in Eq. 10 and 13. H, K, and U also refer to HMDB51, Kinetics, and UCF101, respectively." + ], + "table_footnote": [], + "table_body": "
P+P-P̂EHKU
(a)✓61.874.686.7
(b)✓✓62.575.086.8
(c)✓✓✓62.274.887.1
(d)✓✓✓✓62.875.187.2
", + "bbox": [ + 78, + 151, + 470, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of adaptation with the consideration of given class compositions. The performance gains in (d) validate our claim.", + "bbox": [ + 76, + 250, + 470, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of pattern tokens. In Fig. 3, we present an ablation study on the effect of the number of pattern tokens in TEAM. The results show that TEAM remains robust across a wide range of token counts, except when the value deviates significantly from the optimal (60 for all datasets). These consistent trends across all datasets suggest that our method is not sensitive to the number of pattern tokens, demonstrating its strong robustness.", + "bbox": [ + 75, + 297, + 472, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Varying N and K for N-way K-shot. Beyond standard FSAR settings, we further validate TEAM's effectiveness under varying N-way K-shot configurations on Kinetics, following prior work [25]. Specifically, we vary the number of classes $N$ from 5 to 10, and the number of video instances per class $K$ from 1 to 5. As shown in Fig. 4, TEAM consistently outperforms prior approaches across all configurations. These results confirm that TEAM is robust to different FSAR settings, with consistent benefits across diverse scenarios.", + "bbox": [ + 75, + 438, + 470, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 76, + 583, + 181, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What do pattern tokens encode? Pattern tokens are leveraged to incorporate video features and have been verified to be superior. To assess the video understanding capability of pattern tokens, we examine whether pattern tokens focus on the same context for the same category. In Fig. 5, we visualize the attention weights of a pattern token that effectively represents specific classes. Pattern tokens are considered effective in representing a class when the output pattern tokens for instances of the same class exhibit the highest similarity among all pattern tokens. As observed, pattern tokens consistently focus on the context even when the same action occurs in different frames, at varying speeds, and over different lengths. This implies that pattern tokens address the limitations of frame- and tuple-level alignment approaches, where actions can only be compared within pre-defined units.", + "bbox": [ + 75, + 606, + 470, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Necessity for multiple pattern tokens. We adopt multiple pattern tokens to capture diverse discriminative patterns across base categories, using these patterns as a foun", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17445a7cc641c8c3ef7b63935122016144885a84af7c0e94f4260d960b9153f2.jpg", + "image_caption": [ + "Figure 3. Ablation study for the number of pattern tokens." + ], + "image_footnote": [], + "bbox": [ + 500, + 85, + 893, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77bd2e4943d76ef3f7ee738ea199ef6740348d37878c8f6d421406913c4c36b1.jpg", + "image_caption": [ + "Figure 4. N-way 1-shot and 5-way K-shot results." + ], + "image_footnote": [], + "bbox": [ + 501, + 261, + 890, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dation for distinguishing novel classes. To verify that the learned patterns effectively differentiate novel classes, we compute the discriminative power of each instance pattern token within each category and rank the tokens accordingly. Discriminative power is defined as ' (intra similarity - inter similarity)'. Intra similarity measures the average similarity between instance-specific pattern tokens from all videos within the same category and their prototype (average), while inter similarity is the maximum similarity between the prototype of one class and the prototypes of other classes. The resulting heatmap of discriminative power is shown in Fig. 6, where distinct heatmap arrangements between classes can be observed. The variation in discriminative rank indicates that useful patterns differ by category, supporting the need for multiple learnable pattern tokens to capture a variety of discriminative patterns, forming a foundation for classifying novel categories.", + "bbox": [ + 496, + 444, + 893, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational time for matching process. In this study, we compare the computational overhead with alignment-based approaches. Computational costs for varying numbers of frames are shown in Fig. 7. As discussed in Sec. 1, alignment-based strategies require quadratic computation time due to pairwise comparison between temporal units, whereas TEAM with token-wise comparison has linear time complexity—especially evident when the number of frames exceeds 32. Although this issue is less critical for FSAR datasets, which consist of trimmed, short-term videos, it becomes significant in real-world scenarios with untrimmed, long-term videos. In this regard, we argue that our proposed", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5418", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ec7e23c6ca02ecc059306bba68c52749a567fe97e8b4434767d3843b8b3425c.jpg", + "image_caption": [ + "Figure 5. Visualization of cross-attention weights during instance pattern token aggregation. The same pattern tokens are visualized for videos in each class. Green regions in each frame metaphorically represent the percentage of the frame involvement in the instance pattern token. (a) For 'Diving Cliff', the visualized token primarily responds to the moment when water splashes when people dive. (b) In 'Playing Trumpet', the token focuses on the scenes where people hold a trumpet to their mouths." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 488, + 296 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16a883bbc2f725f971bd8fb042d8a6dd9a84408e84ee9bf0d6b426c76be3cdd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a14a4a9b2ccad5a158a7c16eb9f74f082d5efd1e8b4bb0393d2d16c0a1506fe.jpg", + "image_caption": [ + "Figure 6. Heatmap visualizing the class-wise discriminative power of each instance pattern token for novel classes within HMDB51. Differences in discrimination ranks across classes indicate that each class has unique discriminative patterns." + ], + "image_footnote": [], + "bbox": [ + 76, + 369, + 468, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18769f9cda33c05b725a1af7011a123aa25b263d719482aba7ae1eae3645f4ec.jpg", + "image_caption": [ + "Figure 7. Comparison of computational time (ms) for video matching with alignment-based strategies across varying frame counts. OTAM and TRX are used as representatives of the frame- and tuple-level alignment approaches, respectively. Times reflect only the video matching process and exclude feature extraction by the backbone network, which is consistent across all methods." + ], + "image_footnote": [], + "bbox": [ + 78, + 492, + 468, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "temporal alignment-free matching offers a compelling alternative to alignment-based approaches.", + "bbox": [ + 75, + 746, + 470, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FLOPs and parameter. We analyze FLOPs and parameter counts of TEAM and other methods, as shown in Tab. 5. While TEAM achieves remarkable performance, it has the fewest parameters except for OTAM, which relies solely on backbone parameters. The higher parameter counts in other methods stem from their complex designs, including temporal transformers. These results highlight the effectiveness of", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3f2d2d337167513720f7504e3f67ca9800fb66bca8b160c4ace45a381fe4d5cf.jpg", + "table_caption": [ + "Table 5. Comparison of FLOPs and parameter counts, which are measured in G (giga operations) and M (million), respectively." + ], + "table_footnote": [], + "table_body": "
ModelResNetViT
FLOPsParameterFLOPsParameter
OTAM [3]19823.580985.8
TRX [20]20547.181088.8
STRM [23]20145.581088.8
SloshNet [27]594168.9--
MoLo [25]23789.583099.1
TATs [13]--117897.7
TEAM22432.081387.0
", + "bbox": [ + 501, + 401, + 890, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pattern tokens for FSAR, providing a lightweight alternative to heavier models. Moreover, although TEAM has higher FLOPs than OTAM and TRX, its efficient parallel operations ensure low computational time, as shown in Fig. 7.", + "bbox": [ + 498, + 571, + 890, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 659, + 617, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a novel TEmporal Alignment-free Matching (TEAM) for Few-Shot Action Recognition, that achieves both flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and brute-force alignment for the video comparison. Concretely, TEAM integrates each video with a fixed set of pattern tokens that encapsulate globally discriminative patterns within the video instance, regardless of action length and speed. Moreover, TEAM has inherent efficiency by computing token-wise similarity for matching video. Upon these pattern tokens, we further proposed an adaptation process for establishing clear boundaries between classes by removing common information. We validated the effectiveness of TEAM with extensive experiments.", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5419", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2020-0-01821, RS-2019-II190421, RS-2024-00459618, RS-2024-00360227, RS-2024-00437102, RS-2024-00437633), and MSIT/NRF (No. RS-2024-00357729).", + "bbox": [ + 76, + 90, + 470, + 165 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 181, + 171, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3", + "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1", + "[3] Kaidi Cao, Jingwei Ji, Zhangjie Cao, Chien-Yi Chang, and Juan Carlos Niebles. Few-shot video classification via temporal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10618-10627, 2020. 1, 2, 6, 8", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3", + "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 5", + "[6] Anoop Cherian and Stephen Gould. Second-order temporal pooling for action recognition. International Journal of Computer Vision, 127:340-362, 2019. 5", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3, 5", + "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[10] Yifei Huang, Lijin Yang, and Yoichi Sato. Compound prototype matching for few-shot action recognition. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 6", + "[11] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General" + ], + "bbox": [ + 78, + 205, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3", + "[12] Hirokatsu Kataoka, Tenga Wakamiya, Kensho Hara, and Yutaka Satoh. Would mega-scale datasets further enhance spatiotemporal 3d cnns? arXiv preprint arXiv:2004.04968, 2020.5", + "[13] Pulkit Kumar, Namitha Padmanabhan, Luke Luo, Sai Saketh Rambhatla, and Abhinav Shrivastava. Trajectory-aligned space-time tokens for few-shot action recognition. In European Conference on Computer Vision, pages 474-493. Springer, 2025. 6, 8", + "[14] SuBeen Lee, WonJun Moon, and Jae-Pil Heo. Task discrepancy maximization for fine-grained few-shot classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5331-5340, 2022. 3", + "[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3", + "[16] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4100–4110, 2023. 3", + "[17] Shuyuan Li, Huabin Liu, Rui Qian, Yuxi Li, John See, Mengjuan Fei, Xiaoyuan Yu, and Weiyao Lin. Ta2n: Two-stage action alignment network for few-shot action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1404-1411, 2022. 6", + "[18] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 3", + "[19] Khoi D Nguyen, Quoc-Huy Tran, Khoi Nguyen, Binh-Son Hua, and Rang Nguyen. Inductive and transductive few-shot video classification via appearance and temporal alignments. In European Conference on Computer Vision, pages 471-487. Springer, 2022. 6", + "[20] Toby Perrett, Alessandro Masullo, Tilo Burghardt, Majid Mirmehdi, and Dima Damen. Temporal-relational crosstransformers for few-shot action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 475-484, 2021. 1, 2, 3, 6, 8", + "[21] Sarinda Samarasinghe, Mamshad Nayeem Rizve, Navid Kardan, and Mubarak Shah. Cdfsl-v: Cross-domain few-shot learning for videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11643-11652, 2023. 5", + "[22] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1, 3, 6", + "[23] Anirudh Thatipelli, Sanath Narayan, Salman Khan, Rao Muhammad Anwer, Fahad Shahbaz Khan, and Bernard Ghanem. Spatio-temporal relation modeling for few-shot" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5420", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19958-19967, 2022. 1, 2, 3, 6, 8", + "[24] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Mingqian Tang, Zhengrong Zuo, Changxin Gao, Rong Jin, and Nong Sang. Hybrid relation guided set matching for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19948-19957, 2022. 2, 6", + "[25] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Changxin Gao, Yingya Zhang, Deli Zhao, and Nong Sang. Molo: Motion-augmented long-short contrastive learning for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18011-18021, 2023. 1, 2, 6, 7, 8", + "[26] Jiamin Wu, Tianzhu Zhang, Zhe Zhang, Feng Wu, and Yongdong Zhang. Motion-modulated temporal fragment alignment network for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151-9160, 2022. 6", + "[27] Jiazheng Xing, Mengmeng Wang, Yong Liu, and Boyu Mu. Revisiting the spatial and temporal modeling for few-shot action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3001-3009, 2023. 1, 2, 6, 8", + "[28] Jiazheng Xing, Mengmeng Wang, Yudi Ruan, Bofan Chen, Yaowei Guo, Boyu Mu, Guang Dai, Jingdong Wang, and Yong Liu. Boosting few-shot action recognition with graph-guided hybrid matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1740-1750, 2023. 1, 6", + "[29] Hongguang Zhang, Li Zhang, Xiaojuan Qi, Hongdong Li, Philip HS Torr, and Piotr Koniusz. Few-shot action recognition with permutation-invariant attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 525-542. Springer, 2020. 1, 6", + "[30] Sipeng Zheng, Shizhe Chen, and Qin Jin. Few-shot action recognition with hierarchical matching and contrastive learning. In European Conference on Computer Vision, pages 297-313. Springer, 2022. 6", + "[31] Linchao Zhu and Yi Yang. Compound memory networks for few-shot video classification. In Proceedings of the European Conference on Computer Vision (ECCV), pages 751-766, 2018. 1, 6" + ], + "bbox": [ + 78, + 90, + 470, + 710 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5421", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_model.json b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5b9291326452e2adf4d6f292b50c0078a0e37260 --- /dev/null +++ b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_model.json @@ -0,0 +1,1925 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Temporal Alignment-Free Video Matching for Few-shot Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.181, + 0.731, + 0.218 + ], + "angle": 0, + "content": "SuBeen Lee, WonJun Moon, Hyun Seok Seong, Jae-Pil Heo* Sungkyunkwan University" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.22, + 0.719, + 0.235 + ], + "angle": 0, + "content": "{leesb7426，wjun0830，gustjrdl95，jaepilheo}@skku.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Few-Shot Action Recognition (FSAR) aims to train a model with only a few labeled video instances. A key challenge in FSAR is handling divergent narrative trajectories for precise video matching. While the frame- and tuple-level alignment approaches have been promising, their methods heavily rely on pre-defined and length-dependent alignment units (e.g., frames or tuples), which limits flexibility for actions of varying lengths and speeds. In this work, we introduce a novel Temporal Alignment-free Matching (TEAM) approach, which eliminates the need for temporal units in action representation and brute-force alignment during matching. Specifically, TEAM represents each video with a fixed set of pattern tokens that capture globally discriminative clues within the video instance regardless of action length or speed, ensuring its flexibility. Furthermore, TEAM is inherently efficient, using token-wise comparisons to measure similarity between videos, unlike existing methods that rely on pairwise comparisons for temporal alignment. Additionally, we propose an adaptation process that identifies and removes common information across classes, establishing clear boundaries even between novel categories. Extensive experiments demonstrate the effectiveness of TEAM. Codes are available at github.com/leesb7426/TEAM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.21, + 0.698 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.472, + 0.875 + ], + "angle": 0, + "content": "Recently, action recognition has achieved remarkable performances with the help of numerous labeled videos [2, 5, 8]. However, those performances are not guaranteed when novel categories are given with few labeled examples [3, 22, 29, 31]. To resolve this problem, Few-Shot Action Recognition (FSAR) has been proposed [31]. Particularly, FSAR employs episodic learning which iteratively simulates the episode of learning a set of new classes. Each episode is given a support set to train a model, composed of a few labeled examples for newly given classes, to distinguish the category of a set of unlabeled videos, coined as a query set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.896, + 0.422 + ], + "angle": 0, + "content": "Due to different class compositions within every episode, metric learning, a classifier-free strategy, is predominantly employed in FSAR [3, 20, 23-25, 27, 28]. Distances between the representations of support and query sets are computed to infer the category of a query video. Yet, naively constructing video representations, such as average pooling, are shown to be poor due to an insufficient understanding of temporal information [3, 31]. Therefore, the previous approaches typically focused on aligning temporal sequences between videos [3, 20, 23-25, 27, 28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.897, + 0.56 + ], + "angle": 0, + "content": "As the base units for temporal alignment between videos, the frame and tuple are commonly employed (illustrated in Fig. 1 (a) and (b)). Both frame- and tuple-based alignment approaches measure pairwise similarities between each unit and then derive an optimal alignment path for all units of query video. This path is regarded as the temporal alignment between support and query videos, the query video is assigned with the class label of the support video with minimum cumulative distance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.897, + 0.762 + ], + "angle": 0, + "content": "Although these two temporal alignment paradigms have shown effectiveness, they suffer from inflexibility and inefficiency. The inflexibility arises from their reliance on predefined units, such as frames or tuples, for action representation. These methods typically struggle to capture actions if the pre-defined units do not align with the varying action durations. Moreover, their alignment cost increases quadratically with the number of frames, as shown in Fig. 7. This issue has been overlooked since typical FSAR benchmarking datasets consist mostly of trimmed, short-term videos. The quadratic cost arises because their alignment strategies rely on pairwise comparisons between the units of support and query videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "In this regard, we introduce a novel Temporal Alignment-free Matching (TEAM) approach that achieves both flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and brute-force alignment for video comparison. Specifically, TEAM represents each video with a fixed set of pattern tokens with cross-attention mechanisms to capture diverse clues within the video instance which are useful for distinguishing categories. Therefore, these pattern tokens flexibly identify discrimina" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5412" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.418, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.088, + 0.891, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.397, + 0.897, + 0.509 + ], + "angle": 0, + "content": "Figure 1. Comparison with alignment-based approaches on the \"Diving Cliff\" class. (a) Frame Alignment: For each frame of the query video, the most corresponding frame in the support video is identified, highlighting the importance of precise frame-level alignment. (b) Tuple Alignment: The support and query videos are compared in sub-sequence units to account for variations in action speed. The set of varying tuples is pre-defined. (c) Temporal Alignment-free Matching: Both in support and query, the video features are initially integrated using pattern tokens, which encode globally discriminative patterns. Then, it directly compares the corresponding aggregated features of the support and query videos. Pattern-based aggregation is both efficient and flexible, as it does not require an alignment process and is unaffected by differences in frame count or action speed. Note the text description for each pattern token (green) is intended to provide an intuitive understanding of what each pattern token represents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.473, + 0.676 + ], + "angle": 0, + "content": "tive features varying lengths across each video, as illustrated in Fig. 1 (c). Furthermore, these pattern tokens are trained to capture globally discriminative clues across the dataset. This is because all videos leverage the entire set of pattern tokens without an alignment process, rather than relying on specific tokens determined by alignment, as shown in Fig. 1 (c). Additionally, this token-wise comparison results in the efficiency of TEAM by eliminating the brute-force alignment process computing pairwise similarities between units for the query video matching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.472, + 0.788 + ], + "angle": 0, + "content": "However, pattern tokens may not be fully discriminative for every episode with novel class compositions. To address this, we adapt the class boundaries formed by pattern tokens of support set videos. Particularly, we identify the patterns that commonly appear across classes and remove them from each support video token. This process establishes clear class boundaries, improving class discrimination for query videos." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.791, + 0.319, + 0.806 + ], + "angle": 0, + "content": "To sum up, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.471, + 0.87 + ], + "angle": 0, + "content": "- We propose Temporal Alignment-free Matching for FSAR equipped with flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and alignment in video comparison." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- We develop an episode adaptation process of encoded pattern tokens. This establishes clear boundaries within" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.892, + 0.553 + ], + "angle": 0, + "content": "the support set for each class composition by excluding information shared with other classes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.554, + 0.892, + 0.583 + ], + "angle": 0, + "content": "- Our state-of-the-art performances on various benchmarks for FSAR validate the effectiveness of our TEAM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.596, + 0.642, + 0.612 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.761, + 0.637 + ], + "angle": 0, + "content": "2.1. Few-Shot Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "A predominant learning strategy in Few-Shot Action Recognition (FSAR) is metric learning, which entails inferring the category of a query video based on a pre-defined distance metric [3, 20, 23-25, 27]. Previous methods focused on measuring the distance between video features through the development of alignment algorithms, either at the frame-level [3, 24, 25] or tuple-level [20, 23, 27]. As one of the initial frame alignment methods, OTAM [3] aligns the temporal order of query and support videos at the frame-level to precisely estimate the distance between potential query-support pairs. Yet, as it contains a strong assumption that all actions are in a fixed sequence, HyRSM [24] proposed a flexible alignment metric alleviating the strictly ordered constraints. More recently, MoLo [25] focused on learning long-range temporal perception and motion dynamics through frame-level objectives. On the other hand, the stream of tuple-level alignment is to gain robustness to varying action lengths" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5413" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.085, + 0.892, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.897, + 0.466 + ], + "angle": 0, + "content": "Figure 2. Illustration of our approach using a single pattern token (\\(M = 1\\)) in the 3-way 3-shot scenario without query. This figure illustrates our approach, where instance (+) and exclusive (-) pattern tokens represent integrated video features based on the pattern token. (Left) Randomly initialized pattern tokens are optimized with classes in two complementary ways. First, instance pattern tokens (+) are encouraged to cluster with other instances' tokens of the same class while repelling those of other categories. On the other hand, exclusive pattern tokens (-) learn to represent the otherness of each instance by positioning themselves in the embedding space of other classes. (Right) Although these two types of tokens are discriminative for distinguishing base video categories, they may not fully capture the finer details needed for novel categories. To address this, we propose an adaptation process of support pattern tokens for novel classes to refine the class decision boundaries. Note that when multiple pattern tokens are used, these processes run in parallel, with instance and exclusive tokens being compared only within the same pattern token." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.473, + 0.618 + ], + "angle": 0, + "content": "and temporal offsets. Tuple alignment has gained attention for its flexibility in encoding actions of different speeds and temporal offsets [20]. Additionally, STRM [23] incorporated a spatio-temporal enrichment module to enhance both object and motion representations. Yet, tuple-based approaches require pre-defined temporal tuples to provide hints for possible forms of actions in videos. In contrast, our approach is flexible since we do not need pre-defined windows to represent action." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.375, + 0.65 + ], + "angle": 0, + "content": "2.2. Token-based Feature Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.474, + 0.901 + ], + "angle": 0, + "content": "Feature aggregation generates compact contextual representations, providing both effectiveness and efficiency across various fields [1, 4, 7, 11, 14-16, 18, 22]. There are various strategies to implement feature aggregation including different pooling methods [14, 22]. Among them, our method closely aligns with token-based approaches. Simply put, the stream of token-based aggregation integrates the input context within the learnable tokens. For instance, DETR [4] introduces learnable object queries to aggregate spatial features for object detection, while BLIP-2 [15] and Flamingo [1] use learnable tokens to encode image information for multimodal understanding. Similarly to prior methods, we use learnable tokens to aggregate video information. However, our TEAM differs in that the feature aggregation is implemented in two complementary ways, providing more robust class boundaries. Additionally, we also propose a token adap" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.482, + 0.895, + 0.513 + ], + "angle": 0, + "content": "tation process to maintain the benefits of capturing discriminative details while encoding instances of novel classes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.593, + 0.543 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.553, + 0.704, + 0.567 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.895, + 0.822 + ], + "angle": 0, + "content": "In the standard Few-Shot Action Recognition (FSAR), we are given two sets of data: meta-train set for model training \\(\\mathcal{D}_{base} = \\{(v_i,y_i),y_i\\in C_{base}\\}\\) and meta-test set for model testing \\(\\mathcal{D}_{test} = \\{(v_i,y_i),y_i\\in C_{novel}\\}\\). While \\(v_{i}\\) is a video, \\(C_{base}\\) and \\(C_{novel}\\) denote class sets for meta-training and meta-testing, respectively (\\(C_{base}\\cap C_{novel} = \\phi\\)). Generally, the training and testing processes of FSAR are composed of episodes where each episode validates the model's adaptability to a novel class set. Each episode comprises randomly selected \\(N\\) classes and each class consists of \\(K\\) labeled videos and \\(U\\) unlabeled videos, i.e., \\(N\\)-way \\(K\\)-shot episode. Whereas a set of labeled videos is called the support set \\(S = \\{(v_i,y_i)\\}_{i = 1}^{N\\times K}\\), the other set of unlabeled videos are often dubbed as the query set \\(Q = \\{(v_i,y_i)\\}_{i = 1}^{N\\times U}\\). In the below sections, we assume \\(N\\)-way 1-shot setting where the number of query video \\(U\\) is also set to 1 for better clarity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.611, + 0.846 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The schematic illustration of our method is shown in Fig. 2. To eliminate the need for pre-defined units in action representation, we integrate video features using pattern tokens," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5414" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.395 + ], + "angle": 0, + "content": "which capture globally discriminative features across categories. Specifically, our integration process operates in two complementary ways, generating instance and exclusive pattern tokens. Instance pattern tokens are designed to capture shared patterns among videos of the same class. However, these shared patterns may not always be present in every video of the class. In such cases, relying solely on instance pattern tokens for classification is ineffective. To address this, exclusive pattern tokens are introduced to represent \"otherness\"—absent patterns in a given class. This ensures accurate classification, even when key discriminative patterns are missing from certain videos. While these pattern tokens are effective for base categories, they may lack the discriminative details for novel classes. To overcome this, we propose an adaptation process for both instance and exclusive pattern tokens in support videos, reinforcing clear boundaries to distinguish novel classes within each episode. Notably, pattern tokens are only compared within their type (i.e., token-wise comparison), eliminating the need for an alignment process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.405, + 0.3, + 0.42 + ], + "angle": 0, + "content": "3.3. Instance Pattern Tokens" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.471, + 0.475 + ], + "angle": 0, + "content": "Given uniformly sampled \\( T \\) frames from the support and query videos, we independently extract the features for each frame through the image features extractor, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.487, + 0.471, + 0.505 + ], + "angle": 0, + "content": "\\[\nF = \\left[ g _ {\\theta} \\left(x _ {1}\\right), g _ {\\theta} \\left(x _ {2}\\right), \\dots , g _ {\\theta} \\left(x _ {T}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.515, + 0.47, + 0.561 + ], + "angle": 0, + "content": "where \\(g_{\\theta}\\) denotes the image feature extractor \\(\\mathbb{R}^{H\\times W\\times 3}\\to\\) \\(\\mathbb{R}^D\\) and \\(x_{t}\\) means \\(t\\) -th frame of the video. \\(H\\) and \\(W\\) are height and width of the frame while \\(D\\) is size of the features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.563, + 0.473, + 0.683 + ], + "angle": 0, + "content": "Unlike existing works that utilize pre-defined units (i.e., frame or tuple) for representing actions, we integrate video features based on pattern tokens that encode globally discriminative patterns. Initially, we define a pattern pool that consists of \\( M \\) learnable pattern tokens \\( P = [P_{1}, P_{2}, \\dots, P_{M}] \\). Then, we use a cross-attention mechanism to encode instance patterns within \\( P \\). Formally, instance pattern tokens \\( P^{+} \\in \\mathbb{R}^{M \\times D} \\) are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.697, + 0.47, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\bar {P} _ {m} ^ {+} = P _ {m} + \\mathrm {C A} (P _ {m}, F, F), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.731, + 0.47, + 0.75 + ], + "angle": 0, + "content": "\\[\nP _ {m} ^ {+} = \\bar {P} _ {m} ^ {+} + \\operatorname {M L P} \\left(\\bar {P} _ {m} ^ {+}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.47, + 0.804 + ], + "angle": 0, + "content": "where \\(P_{m}\\) is the \\(m\\)-th component of \\(P\\) while \\(\\mathrm{CA}(\\cdot, \\cdot, \\cdot)\\) is the cross-attention receiving query, key, and value as inputs. And, \\(\\mathrm{MLP}(\\cdot)\\) indicates a light fully-connected block." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.805, + 0.471, + 0.851 + ], + "angle": 0, + "content": "Based on instance pattern tokens of support set \\( P^{S+} \\), the probability for the query \\( P^{Q+} \\) to belong to \\( n \\)-th category \\( p^+(y^Q = n|P^S, P^Q) \\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.863, + 0.47, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P D} \\left(P _ {n} ^ {S}, P ^ {Q}\\right) = \\sum_ {m = 1} ^ {M} - d \\left(P _ {n, m} ^ {S +}, P _ {m} ^ {Q +}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.088, + 0.894, + 0.108 + ], + "angle": 0, + "content": "\\[\np ^ {+} (y ^ {Q} = n | P ^ {S}, P ^ {Q}) = \\mathrm {s o f t m a x} (\\mathrm {P D} (P ^ {S}, P ^ {Q}); n), \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.117, + 0.894, + 0.192 + ], + "angle": 0, + "content": "where \\(d\\) is a distance metric, i.e., the cosine distance, while \\(P_{n,m}^{S+}\\) and \\(P_m^{Q+}\\) denote \\(m\\)-th instance pattern tokens of support belonging to category \\(n\\) and query, respectively. In addition, \\(\\operatorname{softmax}(\\cdot; n)\\) represents the \\(n\\)-th component of the softmax function applied to the input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.193, + 0.895, + 0.344 + ], + "angle": 0, + "content": "The key advantage of using pattern tokens is achieving window-free video understanding. Unlike previous approaches that represent action with frame- or tuple-based units, pattern tokens are not restricted by specific lengths or speeds. Furthermore, each pattern token learns globally discriminative features for video classification, since all pattern tokens contribute across the dataset rather than relying on specific tokens determined by alignment. This design of pattern tokens increases flexibility in action representation and removes the computational cost associated with alignment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.357, + 0.73, + 0.372 + ], + "angle": 0, + "content": "3.4. Exclusive Pattern Tokens" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.546 + ], + "angle": 0, + "content": "In addition to the instance pattern tokens that capture specific patterns in the video, we introduce exclusive pattern tokens, designed to encode the otherness to instance pattern tokens. In other words, the exclusive tokens are located close to the other classes' instance tokens and differentiated from the instance tokens of the same class. To emphasize this distinction, we use subtraction rather than addition in Eq. 2 to directly exclude information related to the anchor class, thus the exclusive pattern tokens represent information contrary to instance ones. Specifically, we define the exclusive-pattern tokens \\( P^{-} \\in \\mathbb{R}^{M \\times D} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.562, + 0.892, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\bar {P} _ {m} ^ {-} = P _ {m} - \\mathrm {C A} \\left(P _ {m}, F, F\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.597, + 0.892, + 0.616 + ], + "angle": 0, + "content": "\\[\nP _ {m} ^ {-} = \\bar {P} _ {m} ^ {-} + \\operatorname {M L P} \\left(\\bar {P} _ {m} ^ {-}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.626, + 0.894, + 0.672 + ], + "angle": 0, + "content": "Then, we define the probability of the query belonging to the \\(n\\)-th category \\(p^{-}(y^{Q} = n|P^{S}, P^{Q})\\) with the exclusive and instance pattern tokens, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.685, + 0.892, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm{ND}(P_{n}^{S},P^{Q}) = \\min_{\\substack{1\\leq o\\leq N\\\\ o\\neq n}}\\sum_{m = 1}^{M}\\big(-d(P_{o,m}^{S - },P_{m}^{Q + })\\big) \\tag{8} \\\\ - d (P _ {o, m} ^ {S +}, P _ {m} ^ {Q -})), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.766, + 0.892, + 0.785 + ], + "angle": 0, + "content": "\\[\np ^ {-} \\left(y ^ {Q} = n \\mid P ^ {S}, P ^ {Q}\\right) = \\operatorname {s o f t m a x} \\left(\\mathrm {N D} \\left(P ^ {S}, P ^ {Q}\\right); n\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.9 + ], + "angle": 0, + "content": "To accurately estimate the category of the query using the above probability, the exclusive pattern tokens should be similar to the instance pattern tokens of other classes, while being differentiated from those of the same class. Thus, as learning progresses, the exclusive pattern tokens effectively encode the otherness of the same-class instance pattern tokens, as intended." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5415" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.409, + 0.108 + ], + "angle": 0, + "content": "3.5. Adaptation of Support Pattern Tokens" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.469, + 0.189 + ], + "angle": 0, + "content": "Although the pattern tokens learn globally discriminative patterns across base categories in the training phase, they might not be as effective for novel classes. Therefore, we propose an adaptation process for the support set to achieve a clear separation between categories within an episode." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.19, + 0.471, + 0.31 + ], + "angle": 0, + "content": "Specifically, we remove common information shared with other classes from instance pattern tokens and allocate the discriminative information of other classes into exclusive pattern tokens. To regulate the degree of adaptation for each episode, we use the cosine similarity between pattern tokens of the anchor class and other classes, representing the level of semantic entanglement. The adaptive instance pattern tokens \\(\\hat{P}^{S+} \\in \\mathbb{R}^{M \\times D}\\) are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.318, + 0.47, + 0.339 + ], + "angle": 0, + "content": "\\[\nE _ {n, o, m} ^ {+} = P _ {n, m} ^ {S +} \\cdot P _ {o, m} ^ {S +}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.343, + 0.47, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {P} _ {n, o, m} ^ {S +} = P _ {m} + \\left(1 + E _ {n, o, m} ^ {+}\\right) \\mathrm {C A} \\left(P _ {m}, F _ {n} ^ {S}, F _ {n} ^ {S}\\right) \\tag {11} \\\\ - E _ {n, o, m} ^ {+} \\mathrm {C A} \\left(P _ {m}, F _ {o} ^ {S}, F _ {o} ^ {S}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.079, + 0.388, + 0.47, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\hat {P} _ {n, m} ^ {S +} = \\frac {1}{N - 1} \\sum_ {o = 1} ^ {N} \\mathbb {1} _ {[ n \\neq o ]} \\left(\\tilde {P} _ {n, o, m} ^ {S +} + \\mathrm {M L P} \\left(\\tilde {P} _ {n, o, m} ^ {S +}\\right)\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.469, + 0.492 + ], + "angle": 0, + "content": "where \\((\\cdot)\\) means the cosine similarity measure between two vectors, and \\(P_{n,m}^{S+}\\) is \\(m\\)-th instance pattern token of \\(n\\)-th class in the support set, while \\(\\mathbb{1}_{[n \\neq o]}\\) is an indicator function that outputs 1 if \\(n\\) is different with \\(o\\) and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.493, + 0.469, + 0.522 + ], + "angle": 0, + "content": "Similarly, the adaptive pattern tokens for exclusive ones \\(\\hat{P}^{S-} \\in \\mathbb{R}^{M \\times D}\\) are formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.526, + 0.47, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {P} _ {n, o, m} ^ {S -} = P _ {m} - \\left(1 + E _ {n, o, m} ^ {-}\\right) \\mathrm {C A} \\left(P _ {m}, F _ {n} ^ {S}, F _ {n} ^ {S}\\right) \\tag {13} \\\\ + E _ {n, o, m} ^ {-} \\mathrm {C A} \\left(P _ {m}, F _ {o} ^ {S}, F _ {o} ^ {S}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.079, + 0.574, + 0.47, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\hat {P} _ {n, m} ^ {S -} = \\frac {1}{N - 1} \\sum_ {o = 1} ^ {N} \\mathbb {1} _ {[ n \\neq o ]} \\left(\\tilde {P} _ {n, o, m} ^ {S -} + \\mathrm {M L P} \\left(\\tilde {P} _ {n, o, m} ^ {S -}\\right)\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.472, + 0.799 + ], + "angle": 0, + "content": "Eq. 11 extends Eq. 2 by introducing a process to remove shared information between classes while generating instance tokens. This process suppresses information shared between the anchor class \\( n \\) and the other class \\( o \\). It achieves this by subtracting \\( \\mathrm{CA}(P_m, F_o^S, F_o^S) \\) from \\( \\mathrm{CA}(P_m, F_n^S, F_n^S) \\), thereby retaining only the discriminative information for the anchor class \\( n \\). The extent of suppression is proportional to the degree of information overlap, controlled by \\( E_{n,o,m}^{+} \\). This ensures that the adaptive instance pattern tokens focus on the discriminative features of novel classes. Similarly, Eq. 13 follows the same principle but instead emphasizes the discriminative features of other classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.37, + 0.822 + ], + "angle": 0, + "content": "3.6. Training Objective and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.829, + 0.473, + 0.875 + ], + "angle": 0, + "content": "Adjusting the probabilities defined in Eq. 5 and 9 with adaptive pattern tokens for the support set, we define two objective functions for model training as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.882, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {+} = - \\mathbb {1} _ {[ y ^ {Q} = n ]} \\log \\left(p ^ {+} \\left(y ^ {Q} = n \\mid \\hat {P} ^ {S}, P ^ {Q}\\right)\\right), \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.088, + 0.893, + 0.108 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {-} = - \\mathbb {1} _ {[ y ^ {Q} = n ]} \\log \\left(p ^ {-} \\left(y ^ {Q} = n \\mid \\hat {P} ^ {S}, P ^ {Q}\\right)\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.113, + 0.844, + 0.128 + ], + "angle": 0, + "content": "Then, the final loss function \\(\\mathcal{L}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.141, + 0.893, + 0.157 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} ^ {+} + \\mathcal {L} ^ {-}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.894, + 0.229 + ], + "angle": 0, + "content": "In short, instance pattern tokens are optimized to resemble those of the same class and exclusive tokens of other classes. At the same time, they remain distinct from instance tokens of other classes and exclusive tokens of the same class." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.261 + ], + "angle": 0, + "content": "As a result, the probability of the query to belong to \\( n \\)-th class \\( p(y^{Q} = n) \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.495, + 0.272, + 0.893, + 0.291 + ], + "angle": 0, + "content": "\\[\np \\left(y ^ {Q} = n\\right) = \\operatorname {s o f t m a x} \\left(\\mathrm {P D} \\left(\\hat {P} ^ {S}, P ^ {Q}\\right) + \\mathrm {N D} \\left(\\hat {P} ^ {S}, P ^ {Q}\\right); n\\right). \\tag {18}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.315, + 0.634, + 0.332 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Description of the datasets and implementation details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.38, + 0.859, + 0.396 + ], + "angle": 0, + "content": "4.1. Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.895, + 0.659 + ], + "angle": 0, + "content": "In Tab. 1, we compare our proposed method, TEAM, with previous FSAR approaches under the 5-way 1-shot and 5-shot scenarios, which are standard evaluation protocols. Using ResNet [9], a conventional backbone network, TEAM achieves state-of-the-art performance across various datasets and shot numbers, except for UCF101 (5-shot). With the ViT backbone [7], TEAM surpasses all other methods in every case. We attribute TEAM's effectiveness to its use of globally shared pattern tokens, which inherently capture discriminative patterns across diverse scenarios. This approach enables both efficient and reliable adaptation to new episodes by leveraging pattern tokens to compare videos through multiple criteria. Additionally, TEAM is not only effective but also significantly more efficient than frame- and tuple-level alignment methods, as shown in Fig. 7. These results demonstrate that TEAM is a robust alternative to existing alignment strategies in general FSAR settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.811 + ], + "angle": 0, + "content": "We further conduct experiments on SSv2-Small, a benchmark with high temporal complexity and subtle motion differences, to validate TEAM's effectiveness in handling complex temporal dynamics. As shown in Tab. 2, TEAM surpasses other methods, except for TATs which use an additional off-the-shelf point tracker (PT). Without PT (TATs w/o PT) which generates tailored object trajectories for temporal actions, TEAM also outperforms TATs even in SSv2-Small. These results suggest that the benefits of TEAM remain robust in the context of temporally challenging actions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Cross-domain evaluation is widely used to assess a model's generalization capability [5, 6, 12, 21]. It typically leverages knowledge learned from coarse datasets to recognize actions in more specific and nuanced domains, highlighting the model's adaptability. Following this protocol, we compare TEAM with other FSAR methods in Tab. 3. We" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5416" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 1. Performance comparison on FSAR benchmark datasets under the 5-way setting. The best results are bolded, while the second-best are underlined. \\(\\dagger\\) indicates that evaluation was conducted in a transductive setting, utilizing relations between queries. * denotes reproduced performances, as results with ViT were unavailable." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.137, + 0.9, + 0.507 + ], + "angle": 0, + "content": "
ModelReferenceBackboneHMDB51KineticsUCF101
1-shot5-shot1-shot5-shot1-shot5-shot
ProtoNet [22]NeurIPS 201754.268.464.577.974.089.6
CMN [31]ECCV 2018--57.376.0--
OTAM [3]CVPR 202054.568.072.284.279.988.9
ARN [29]ECCV 202045.560.663.782.466.383.1
TRX [20]CVPR 202153.175.663.685.978.296.1
TA2N [17]AAAI 202259.773.972.885.881.995.1
STRM [23]CVPR 202252.377.362.986.780.596.9
HyRSM [24]CVPR 202260.376.073.786.183.994.7
MTFAN [26]CVPR 202259.074.674.687.484.895.1
HCL [30]ECCV 202259.176.373.785.882.593.9
Nguyen et al. [19]ECCV 202259.676.974.387.484.995.9
Huang et al. [10]ECCV 202260.177.073.386.471.491.0
SloshNet [27]AAAI 2023-77.5-87.0-97.1
MoLo [25]CVPR 202360.877.474.085.686.095.5
GgHM†[28]ICCV 202361.276.974.987.485.296.3
TEAMCVPR 202562.878.475.188.287.296.2
OTAM* [3]CVPR 202069.383.182.291.692.997.5
TRX* [20]CVPR 202161.382.276.890.485.897.8
STRM* [23]CVPR 202256.783.275.890.890.897.6
MoLo* [25]CVPR 202369.983.282.091.293.697.8
TATs [13]ECCV 202460.077.081.991.192.095.5
TEAMCVPR 202570.985.583.392.994.598.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.516, + 0.472, + 0.559 + ], + "angle": 0, + "content": "Table 2. Performance comparison in SSv2-Small dataset using ViT. TATs w/o PT refers to TATs without the additional off-the-shelf Point Tracker. Symbols match those in Tab. 1" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.563, + 0.474, + 0.706 + ], + "angle": 0, + "content": "
ModelReferenceSSv2-Small
1-shot5-shot
OTAM* [3]CVPR 202044.458.2
TRX* [20]CVPR 202142.059.1
STRM* [23]CVPR 202242.560.4
MoLo* [25]CVPR 202345.158.8
TATs [13]ECCV 202447.964.4
TATs w/o PT [13]ECCV 202444.662.5
TEAMCVPR 202547.263.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.472, + 0.834 + ], + "angle": 0, + "content": "train all models on 5-way 5-shot tasks using the Kinetics dataset with the ViT backbone and evaluate them on the same setting across different target datasets. As shown in Tab. 3, TEAM consistently outperforms prior methods, indicating strong generalization beyond the training domain. This demonstrates TEAM's robustness to temporal variations with pattern tokens capturing globally distinct patterns." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.231, + 0.863 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Unless specifically stated, experiments are conducted in the 5-way 1-shot scenario using the ResNet-50." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.517, + 0.895, + 0.573 + ], + "angle": 0, + "content": "Table 3. Performance comparison across cross-domain scenarios. All models are trained on the 5-way 5-shot scenario using the Kinetics dataset with the ViT backbone, and evaluated on the same task across various datasets. Symbols match those in Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.576, + 0.892, + 0.673 + ], + "angle": 0, + "content": "
ModelReferenceHMDB51UCF101
OTAM* [3]CVPR 202077.296.5
TRX* [20]CVPR 202178.197.9
STRM* [23]CVPR 202278.598.0
MoLo* [25]CVPR 202377.797.4
TEAMCVPR 202581.198.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Component Ablation. Tab. 4 presents the performance when our component is applied sequentially. (a) shows that the use of instance pattern tokens solely achieves decent performances over the methods that rely on pre-defined temporal units and brute-force alignment processes. This validates the advantage of our core intuition that temporal alignment-free matching is beneficial for its flexibility in video understanding. Succeedingly, in (b), the exclusive tokens are shown to yield additional improvements, indicating that the different strategy in learning pattern tokens complements the original approach. Conversely, we observe slight performance drops in (c) where we implement the adaptation process without the management in the adaptation magnitude. This implies the importance of managing the degree" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5417" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.472, + 0.148 + ], + "angle": 0, + "content": "Table 4. Ablation study for each component. \\( P^{+} \\), \\( P^{-} \\), and \\( \\hat{P} \\) denote the instance, exclusive, adaptive pattern tokens, respectively, while \\( E \\) means entanglement in Eq. 10 and 13. H, K, and U also refer to HMDB51, Kinetics, and UCF101, respectively." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.152, + 0.472, + 0.236 + ], + "angle": 0, + "content": "
P+P-P̂EHKU
(a)✓61.874.686.7
(b)✓✓62.575.086.8
(c)✓✓✓62.274.887.1
(d)✓✓✓✓62.875.187.2
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.251, + 0.472, + 0.283 + ], + "angle": 0, + "content": "of adaptation with the consideration of given class compositions. The performance gains in (d) validate our claim." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.473, + 0.421 + ], + "angle": 0, + "content": "Number of pattern tokens. In Fig. 3, we present an ablation study on the effect of the number of pattern tokens in TEAM. The results show that TEAM remains robust across a wide range of token counts, except when the value deviates significantly from the optimal (60 for all datasets). These consistent trends across all datasets suggest that our method is not sensitive to the number of pattern tokens, demonstrating its strong robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.472, + 0.576 + ], + "angle": 0, + "content": "Varying N and K for N-way K-shot. Beyond standard FSAR settings, we further validate TEAM's effectiveness under varying N-way K-shot configurations on Kinetics, following prior work [25]. Specifically, we vary the number of classes \\(N\\) from 5 to 10, and the number of video instances per class \\(K\\) from 1 to 5. As shown in Fig. 4, TEAM consistently outperforms prior approaches across all configurations. These results confirm that TEAM is robust to different FSAR settings, with consistent benefits across diverse scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.584, + 0.182, + 0.6 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.472, + 0.834 + ], + "angle": 0, + "content": "What do pattern tokens encode? Pattern tokens are leveraged to incorporate video features and have been verified to be superior. To assess the video understanding capability of pattern tokens, we examine whether pattern tokens focus on the same context for the same category. In Fig. 5, we visualize the attention weights of a pattern token that effectively represents specific classes. Pattern tokens are considered effective in representing a class when the output pattern tokens for instances of the same class exhibit the highest similarity among all pattern tokens. As observed, pattern tokens consistently focus on the context even when the same action occurs in different frames, at varying speeds, and over different lengths. This implies that pattern tokens address the limitations of frame- and tuple-level alignment approaches, where actions can only be compared within pre-defined units." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Necessity for multiple pattern tokens. We adopt multiple pattern tokens to capture diverse discriminative patterns across base categories, using these patterns as a foun" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.087, + 0.895, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.246, + 0.869, + 0.26 + ], + "angle": 0, + "content": "Figure 3. Ablation study for the number of pattern tokens." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.262, + 0.892, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.415, + 0.843, + 0.429 + ], + "angle": 0, + "content": "Figure 4. N-way 1-shot and 5-way K-shot results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.895, + 0.703 + ], + "angle": 0, + "content": "dation for distinguishing novel classes. To verify that the learned patterns effectively differentiate novel classes, we compute the discriminative power of each instance pattern token within each category and rank the tokens accordingly. Discriminative power is defined as ' (intra similarity - inter similarity)'. Intra similarity measures the average similarity between instance-specific pattern tokens from all videos within the same category and their prototype (average), while inter similarity is the maximum similarity between the prototype of one class and the prototypes of other classes. The resulting heatmap of discriminative power is shown in Fig. 6, where distinct heatmap arrangements between classes can be observed. The variation in discriminative rank indicates that useful patterns differ by category, supporting the need for multiple learnable pattern tokens to capture a variety of discriminative patterns, forming a foundation for classifying novel categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Computational time for matching process. In this study, we compare the computational overhead with alignment-based approaches. Computational costs for varying numbers of frames are shown in Fig. 7. As discussed in Sec. 1, alignment-based strategies require quadratic computation time due to pairwise comparison between temporal units, whereas TEAM with token-wise comparison has linear time complexity—especially evident when the number of frames exceeds 32. Although this issue is less critical for FSAR datasets, which consist of trimmed, short-term videos, it becomes significant in real-world scenarios with untrimmed, long-term videos. In this regard, we argue that our proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5418" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.49, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.891, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.893, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Visualization of cross-attention weights during instance pattern token aggregation. The same pattern tokens are visualized for videos in each class. Green regions in each frame metaphorically represent the percentage of the frame involvement in the instance pattern token. (a) For 'Diving Cliff', the visualized token primarily responds to the moment when water splashes when people dive. (b) In 'Playing Trumpet', the token focuses on the scenes where people hold a trumpet to their mouths." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.37, + 0.47, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.486 + ], + "angle": 0, + "content": "Figure 6. Heatmap visualizing the class-wise discriminative power of each instance pattern token for novel classes within HMDB51. Differences in discrimination ranks across classes indicate that each class has unique discriminative patterns." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.493, + 0.47, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.727 + ], + "angle": 0, + "content": "Figure 7. Comparison of computational time (ms) for video matching with alignment-based strategies across varying frame counts. OTAM and TRX are used as representatives of the frame- and tuple-level alignment approaches, respectively. Times reflect only the video matching process and exclude feature extraction by the backbone network, which is consistent across all methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.471, + 0.779 + ], + "angle": 0, + "content": "temporal alignment-free matching offers a compelling alternative to alignment-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.902 + ], + "angle": 0, + "content": "FLOPs and parameter. We analyze FLOPs and parameter counts of TEAM and other methods, as shown in Tab. 5. While TEAM achieves remarkable performance, it has the fewest parameters except for OTAM, which relies solely on backbone parameters. The higher parameter counts in other methods stem from their complex designs, including temporal transformers. These results highlight the effectiveness of" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.37, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Table 5. Comparison of FLOPs and parameter counts, which are measured in G (giga operations) and M (million), respectively." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.402, + 0.892, + 0.546 + ], + "angle": 0, + "content": "
ModelResNetViT
FLOPsParameterFLOPsParameter
OTAM [3]19823.580985.8
TRX [20]20547.181088.8
STRM [23]20145.581088.8
SloshNet [27]594168.9--
MoLo [25]23789.583099.1
TATs [13]--117897.7
TEAM22432.081387.0
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.573, + 0.892, + 0.634 + ], + "angle": 0, + "content": "pattern tokens for FSAR, providing a lightweight alternative to heavier models. Moreover, although TEAM has higher FLOPs than OTAM and TRX, its efficient parallel operations ensure low computational time, as shown in Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.66, + 0.619, + 0.675 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we proposed a novel TEmporal Alignment-free Matching (TEAM) for Few-Shot Action Recognition, that achieves both flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and brute-force alignment for the video comparison. Concretely, TEAM integrates each video with a fixed set of pattern tokens that encapsulate globally discriminative patterns within the video instance, regardless of action length and speed. Moreover, TEAM has inherent efficiency by computing token-wise similarity for matching video. Upon these pattern tokens, we further proposed an adaptation process for establishing clear boundaries between classes by removing common information. We validated the effectiveness of TEAM with extensive experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.472, + 0.166 + ], + "angle": 0, + "content": "Acknowledgements This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2020-0-01821, RS-2019-II190421, RS-2024-00459618, RS-2024-00360227, RS-2024-00437102, RS-2024-00437633), and MSIT/NRF (No. RS-2024-00357729)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.182, + 0.173, + 0.197 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.207, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.472, + 0.348 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.348, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[3] Kaidi Cao, Jingwei Ji, Zhangjie Cao, Chien-Yi Chang, and Juan Carlos Niebles. Few-shot video classification via temporal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10618-10627, 2020. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.419, + 0.472, + 0.488 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.472, + 0.546 + ], + "angle": 0, + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.548, + 0.472, + 0.589 + ], + "angle": 0, + "content": "[6] Anoop Cherian and Stephen Gould. Second-order temporal pooling for action recognition. International Journal of Computer Vision, 127:340-362, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.59, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.702, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.76, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[10] Yifei Huang, Lijin Yang, and Yoichi Sato. Compound prototype matching for few-shot action recognition. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[11] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.895, + 0.133 + ], + "angle": 0, + "content": "perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.895, + 0.19 + ], + "angle": 0, + "content": "[12] Hirokatsu Kataoka, Tenga Wakamiya, Kensho Hara, and Yutaka Satoh. Would mega-scale datasets further enhance spatiotemporal 3d cnns? arXiv preprint arXiv:2004.04968, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.895, + 0.262 + ], + "angle": 0, + "content": "[13] Pulkit Kumar, Namitha Padmanabhan, Luke Luo, Sai Saketh Rambhatla, and Abhinav Shrivastava. Trajectory-aligned space-time tokens for few-shot action recognition. In European Conference on Computer Vision, pages 474-493. Springer, 2025. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.895, + 0.32 + ], + "angle": 0, + "content": "[14] SuBeen Lee, WonJun Moon, and Jae-Pil Heo. Task discrepancy maximization for fine-grained few-shot classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5331-5340, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.895, + 0.389 + ], + "angle": 0, + "content": "[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.895, + 0.462 + ], + "angle": 0, + "content": "[16] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4100–4110, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.895, + 0.532 + ], + "angle": 0, + "content": "[17] Shuyuan Li, Huabin Liu, Rui Qian, Yuxi Li, John See, Mengjuan Fei, Xiaoyuan Yu, and Weiyao Lin. Ta2n: Two-stage action alignment network for few-shot action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1404-1411, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.895, + 0.603 + ], + "angle": 0, + "content": "[18] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.895, + 0.673 + ], + "angle": 0, + "content": "[19] Khoi D Nguyen, Quoc-Huy Tran, Khoi Nguyen, Binh-Son Hua, and Rang Nguyen. Inductive and transductive few-shot video classification via appearance and temporal alignments. In European Conference on Computer Vision, pages 471-487. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.895, + 0.745 + ], + "angle": 0, + "content": "[20] Toby Perrett, Alessandro Masullo, Tilo Burghardt, Majid Mirmehdi, and Dima Damen. Temporal-relational crosstransformers for few-shot action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 475-484, 2021. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.895, + 0.815 + ], + "angle": 0, + "content": "[21] Sarinda Samarasinghe, Mamshad Nayeem Rizve, Navid Kardan, and Mubarak Shah. Cdfsl-v: Cross-domain few-shot learning for videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11643-11652, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.895, + 0.859 + ], + "angle": 0, + "content": "[22] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[23] Anirudh Thatipelli, Sanath Narayan, Salman Khan, Rao Muhammad Anwer, Fahad Shahbaz Khan, and Bernard Ghanem. Spatio-temporal relation modeling for few-shot" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5420" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19958-19967, 2022. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[24] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Mingqian Tang, Zhengrong Zuo, Changxin Gao, Rong Jin, and Nong Sang. Hybrid relation guided set matching for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19948-19957, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[25] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Changxin Gao, Yingya Zhang, Deli Zhao, and Nong Sang. Molo: Motion-augmented long-short contrastive learning for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18011-18021, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.472, + 0.373 + ], + "angle": 0, + "content": "[26] Jiamin Wu, Tianzhu Zhang, Zhe Zhang, Feng Wu, and Yongdong Zhang. Motion-modulated temporal fragment alignment network for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151-9160, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.472, + 0.431 + ], + "angle": 0, + "content": "[27] Jiazheng Xing, Mengmeng Wang, Yong Liu, and Boyu Mu. Revisiting the spatial and temporal modeling for few-shot action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3001-3009, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.472, + 0.513 + ], + "angle": 0, + "content": "[28] Jiazheng Xing, Mengmeng Wang, Yudi Ruan, Bofan Chen, Yaowei Guo, Boyu Mu, Guang Dai, Jingdong Wang, and Yong Liu. Boosting few-shot action recognition with graph-guided hybrid matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1740-1750, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.515, + 0.472, + 0.599 + ], + "angle": 0, + "content": "[29] Hongguang Zhang, Li Zhang, Xiaojuan Qi, Hongdong Li, Philip HS Torr, and Piotr Koniusz. Few-shot action recognition with permutation-invariant attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 525-542. Springer, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.6, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[30] Sipeng Zheng, Shizhe Chen, and Qin Jin. Few-shot action recognition with hierarchical matching and contrastive learning. In European Conference on Computer Vision, pages 297-313. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.472, + 0.711 + ], + "angle": 0, + "content": "[31] Linchao Zhu and Yi Yang. Compound memory networks for few-shot video classification. In Proceedings of the European Conference on Computer Vision (ECCV), pages 751-766, 2018. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "5421" + } + ] +] \ No newline at end of file diff --git a/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_origin.pdf b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..55397fb6e837b5060696769fc09b81ea2c8d005b --- /dev/null +++ b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/ea700305-465d-4c66-9110-1ea42533dcb6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae6bbf2e3df80db821c2f222e062401ecd95836adde8735b4f40e315e769895c +size 715096 diff --git a/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/full.md b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..afba54e8a2f0a32447d35f1797e5be2073348f3b --- /dev/null +++ b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/full.md @@ -0,0 +1,309 @@ +# Temporal Alignment-Free Video Matching for Few-shot Action Recognition + +SuBeen Lee, WonJun Moon, Hyun Seok Seong, Jae-Pil Heo* Sungkyunkwan University + +{leesb7426，wjun0830，gustjrdl95，jaepilheo}@skku.edu + +# Abstract + +Few-Shot Action Recognition (FSAR) aims to train a model with only a few labeled video instances. A key challenge in FSAR is handling divergent narrative trajectories for precise video matching. While the frame- and tuple-level alignment approaches have been promising, their methods heavily rely on pre-defined and length-dependent alignment units (e.g., frames or tuples), which limits flexibility for actions of varying lengths and speeds. In this work, we introduce a novel Temporal Alignment-free Matching (TEAM) approach, which eliminates the need for temporal units in action representation and brute-force alignment during matching. Specifically, TEAM represents each video with a fixed set of pattern tokens that capture globally discriminative clues within the video instance regardless of action length or speed, ensuring its flexibility. Furthermore, TEAM is inherently efficient, using token-wise comparisons to measure similarity between videos, unlike existing methods that rely on pairwise comparisons for temporal alignment. Additionally, we propose an adaptation process that identifies and removes common information across classes, establishing clear boundaries even between novel categories. Extensive experiments demonstrate the effectiveness of TEAM. Codes are available at github.com/leesb7426/TEAM. + +# 1. Introduction + +Recently, action recognition has achieved remarkable performances with the help of numerous labeled videos [2, 5, 8]. However, those performances are not guaranteed when novel categories are given with few labeled examples [3, 22, 29, 31]. To resolve this problem, Few-Shot Action Recognition (FSAR) has been proposed [31]. Particularly, FSAR employs episodic learning which iteratively simulates the episode of learning a set of new classes. Each episode is given a support set to train a model, composed of a few labeled examples for newly given classes, to distinguish the category of a set of unlabeled videos, coined as a query set. + +Due to different class compositions within every episode, metric learning, a classifier-free strategy, is predominantly employed in FSAR [3, 20, 23-25, 27, 28]. Distances between the representations of support and query sets are computed to infer the category of a query video. Yet, naively constructing video representations, such as average pooling, are shown to be poor due to an insufficient understanding of temporal information [3, 31]. Therefore, the previous approaches typically focused on aligning temporal sequences between videos [3, 20, 23-25, 27, 28]. + +As the base units for temporal alignment between videos, the frame and tuple are commonly employed (illustrated in Fig. 1 (a) and (b)). Both frame- and tuple-based alignment approaches measure pairwise similarities between each unit and then derive an optimal alignment path for all units of query video. This path is regarded as the temporal alignment between support and query videos, the query video is assigned with the class label of the support video with minimum cumulative distance. + +Although these two temporal alignment paradigms have shown effectiveness, they suffer from inflexibility and inefficiency. The inflexibility arises from their reliance on predefined units, such as frames or tuples, for action representation. These methods typically struggle to capture actions if the pre-defined units do not align with the varying action durations. Moreover, their alignment cost increases quadratically with the number of frames, as shown in Fig. 7. This issue has been overlooked since typical FSAR benchmarking datasets consist mostly of trimmed, short-term videos. The quadratic cost arises because their alignment strategies rely on pairwise comparisons between the units of support and query videos. + +In this regard, we introduce a novel Temporal Alignment-free Matching (TEAM) approach that achieves both flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and brute-force alignment for video comparison. Specifically, TEAM represents each video with a fixed set of pattern tokens with cross-attention mechanisms to capture diverse clues within the video instance which are useful for distinguishing categories. Therefore, these pattern tokens flexibly identify discrimina + +![](images/8dc4c5373c602a003d4e4bb9a03cbb3f0feb8d8b0ab464462b94c6dfbc547b0c.jpg) +Figure 1. Comparison with alignment-based approaches on the "Diving Cliff" class. (a) Frame Alignment: For each frame of the query video, the most corresponding frame in the support video is identified, highlighting the importance of precise frame-level alignment. (b) Tuple Alignment: The support and query videos are compared in sub-sequence units to account for variations in action speed. The set of varying tuples is pre-defined. (c) Temporal Alignment-free Matching: Both in support and query, the video features are initially integrated using pattern tokens, which encode globally discriminative patterns. Then, it directly compares the corresponding aggregated features of the support and query videos. Pattern-based aggregation is both efficient and flexible, as it does not require an alignment process and is unaffected by differences in frame count or action speed. Note the text description for each pattern token (green) is intended to provide an intuitive understanding of what each pattern token represents. + +![](images/a21e76e999c3d311f7f35d937d47561d4206e44de7774e3b39fa62d47b29a39a.jpg) + +tive features varying lengths across each video, as illustrated in Fig. 1 (c). Furthermore, these pattern tokens are trained to capture globally discriminative clues across the dataset. This is because all videos leverage the entire set of pattern tokens without an alignment process, rather than relying on specific tokens determined by alignment, as shown in Fig. 1 (c). Additionally, this token-wise comparison results in the efficiency of TEAM by eliminating the brute-force alignment process computing pairwise similarities between units for the query video matching. + +However, pattern tokens may not be fully discriminative for every episode with novel class compositions. To address this, we adapt the class boundaries formed by pattern tokens of support set videos. Particularly, we identify the patterns that commonly appear across classes and remove them from each support video token. This process establishes clear class boundaries, improving class discrimination for query videos. + +To sum up, our contributions are: + +- We propose Temporal Alignment-free Matching for FSAR equipped with flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and alignment in video comparison. +- We develop an episode adaptation process of encoded pattern tokens. This establishes clear boundaries within + +the support set for each class composition by excluding information shared with other classes. + +- Our state-of-the-art performances on various benchmarks for FSAR validate the effectiveness of our TEAM. + +# 2. Related Work + +# 2.1. Few-Shot Action Recognition + +A predominant learning strategy in Few-Shot Action Recognition (FSAR) is metric learning, which entails inferring the category of a query video based on a pre-defined distance metric [3, 20, 23-25, 27]. Previous methods focused on measuring the distance between video features through the development of alignment algorithms, either at the frame-level [3, 24, 25] or tuple-level [20, 23, 27]. As one of the initial frame alignment methods, OTAM [3] aligns the temporal order of query and support videos at the frame-level to precisely estimate the distance between potential query-support pairs. Yet, as it contains a strong assumption that all actions are in a fixed sequence, HyRSM [24] proposed a flexible alignment metric alleviating the strictly ordered constraints. More recently, MoLo [25] focused on learning long-range temporal perception and motion dynamics through frame-level objectives. On the other hand, the stream of tuple-level alignment is to gain robustness to varying action lengths + +![](images/64516495370e30de382d59ccb6d635e7389685c1a91aad2073df03ca14bbe8ff.jpg) +Figure 2. Illustration of our approach using a single pattern token ( $M = 1$ ) in the 3-way 3-shot scenario without query. This figure illustrates our approach, where instance (+) and exclusive (-) pattern tokens represent integrated video features based on the pattern token. (Left) Randomly initialized pattern tokens are optimized with classes in two complementary ways. First, instance pattern tokens (+) are encouraged to cluster with other instances' tokens of the same class while repelling those of other categories. On the other hand, exclusive pattern tokens (-) learn to represent the otherness of each instance by positioning themselves in the embedding space of other classes. (Right) Although these two types of tokens are discriminative for distinguishing base video categories, they may not fully capture the finer details needed for novel categories. To address this, we propose an adaptation process of support pattern tokens for novel classes to refine the class decision boundaries. Note that when multiple pattern tokens are used, these processes run in parallel, with instance and exclusive tokens being compared only within the same pattern token. + +and temporal offsets. Tuple alignment has gained attention for its flexibility in encoding actions of different speeds and temporal offsets [20]. Additionally, STRM [23] incorporated a spatio-temporal enrichment module to enhance both object and motion representations. Yet, tuple-based approaches require pre-defined temporal tuples to provide hints for possible forms of actions in videos. In contrast, our approach is flexible since we do not need pre-defined windows to represent action. + +# 2.2. Token-based Feature Aggregation + +Feature aggregation generates compact contextual representations, providing both effectiveness and efficiency across various fields [1, 4, 7, 11, 14-16, 18, 22]. There are various strategies to implement feature aggregation including different pooling methods [14, 22]. Among them, our method closely aligns with token-based approaches. Simply put, the stream of token-based aggregation integrates the input context within the learnable tokens. For instance, DETR [4] introduces learnable object queries to aggregate spatial features for object detection, while BLIP-2 [15] and Flamingo [1] use learnable tokens to encode image information for multimodal understanding. Similarly to prior methods, we use learnable tokens to aggregate video information. However, our TEAM differs in that the feature aggregation is implemented in two complementary ways, providing more robust class boundaries. Additionally, we also propose a token adap + +tation process to maintain the benefits of capturing discriminative details while encoding instances of novel classes. + +# 3. Method + +# 3.1. Problem Formulation + +In the standard Few-Shot Action Recognition (FSAR), we are given two sets of data: meta-train set for model training $\mathcal{D}_{base} = \{(v_i,y_i),y_i\in C_{base}\}$ and meta-test set for model testing $\mathcal{D}_{test} = \{(v_i,y_i),y_i\in C_{novel}\}$ . While $v_{i}$ is a video, $C_{base}$ and $C_{novel}$ denote class sets for meta-training and meta-testing, respectively ( $C_{base}\cap C_{novel} = \phi$ ). Generally, the training and testing processes of FSAR are composed of episodes where each episode validates the model's adaptability to a novel class set. Each episode comprises randomly selected $N$ classes and each class consists of $K$ labeled videos and $U$ unlabeled videos, i.e., $N$ -way $K$ -shot episode. Whereas a set of labeled videos is called the support set $S = \{(v_i,y_i)\}_{i = 1}^{N\times K}$ , the other set of unlabeled videos are often dubbed as the query set $Q = \{(v_i,y_i)\}_{i = 1}^{N\times U}$ . In the below sections, we assume $N$ -way 1-shot setting where the number of query video $U$ is also set to 1 for better clarity. + +# 3.2. Overview + +The schematic illustration of our method is shown in Fig. 2. To eliminate the need for pre-defined units in action representation, we integrate video features using pattern tokens, + +which capture globally discriminative features across categories. Specifically, our integration process operates in two complementary ways, generating instance and exclusive pattern tokens. Instance pattern tokens are designed to capture shared patterns among videos of the same class. However, these shared patterns may not always be present in every video of the class. In such cases, relying solely on instance pattern tokens for classification is ineffective. To address this, exclusive pattern tokens are introduced to represent "otherness"—absent patterns in a given class. This ensures accurate classification, even when key discriminative patterns are missing from certain videos. While these pattern tokens are effective for base categories, they may lack the discriminative details for novel classes. To overcome this, we propose an adaptation process for both instance and exclusive pattern tokens in support videos, reinforcing clear boundaries to distinguish novel classes within each episode. Notably, pattern tokens are only compared within their type (i.e., token-wise comparison), eliminating the need for an alignment process. + +# 3.3. Instance Pattern Tokens + +Given uniformly sampled $T$ frames from the support and query videos, we independently extract the features for each frame through the image features extractor, as follows: + +$$ +F = \left[ g _ {\theta} \left(x _ {1}\right), g _ {\theta} \left(x _ {2}\right), \dots , g _ {\theta} \left(x _ {T}\right) \right], \tag {1} +$$ + +where $g_{\theta}$ denotes the image feature extractor $\mathbb{R}^{H\times W\times 3}\to$ $\mathbb{R}^D$ and $x_{t}$ means $t$ -th frame of the video. $H$ and $W$ are height and width of the frame while $D$ is size of the features. + +Unlike existing works that utilize pre-defined units (i.e., frame or tuple) for representing actions, we integrate video features based on pattern tokens that encode globally discriminative patterns. Initially, we define a pattern pool that consists of $M$ learnable pattern tokens $P = [P_{1}, P_{2}, \dots, P_{M}]$ . Then, we use a cross-attention mechanism to encode instance patterns within $P$ . Formally, instance pattern tokens $P^{+} \in \mathbb{R}^{M \times D}$ are computed as: + +$$ +\bar {P} _ {m} ^ {+} = P _ {m} + \mathrm {C A} (P _ {m}, F, F), \tag {2} +$$ + +$$ +P _ {m} ^ {+} = \bar {P} _ {m} ^ {+} + \operatorname {M L P} \left(\bar {P} _ {m} ^ {+}\right), \tag {3} +$$ + +where $P_{m}$ is the $m$ -th component of $P$ while $\mathrm{CA}(\cdot, \cdot, \cdot)$ is the cross-attention receiving query, key, and value as inputs. And, $\mathrm{MLP}(\cdot)$ indicates a light fully-connected block. + +Based on instance pattern tokens of support set $P^{S+}$ , the probability for the query $P^{Q+}$ to belong to $n$ -th category $p^+(y^Q = n|P^S, P^Q)$ is formulated as follows: + +$$ +\mathrm {P D} \left(P _ {n} ^ {S}, P ^ {Q}\right) = \sum_ {m = 1} ^ {M} - d \left(P _ {n, m} ^ {S +}, P _ {m} ^ {Q +}\right), \tag {4} +$$ + +$$ +p ^ {+} (y ^ {Q} = n | P ^ {S}, P ^ {Q}) = \mathrm {s o f t m a x} (\mathrm {P D} (P ^ {S}, P ^ {Q}); n), \quad (5) +$$ + +where $d$ is a distance metric, i.e., the cosine distance, while $P_{n,m}^{S+}$ and $P_m^{Q+}$ denote $m$ -th instance pattern tokens of support belonging to category $n$ and query, respectively. In addition, $\operatorname{softmax}(\cdot; n)$ represents the $n$ -th component of the softmax function applied to the input. + +The key advantage of using pattern tokens is achieving window-free video understanding. Unlike previous approaches that represent action with frame- or tuple-based units, pattern tokens are not restricted by specific lengths or speeds. Furthermore, each pattern token learns globally discriminative features for video classification, since all pattern tokens contribute across the dataset rather than relying on specific tokens determined by alignment. This design of pattern tokens increases flexibility in action representation and removes the computational cost associated with alignment. + +# 3.4. Exclusive Pattern Tokens + +In addition to the instance pattern tokens that capture specific patterns in the video, we introduce exclusive pattern tokens, designed to encode the otherness to instance pattern tokens. In other words, the exclusive tokens are located close to the other classes' instance tokens and differentiated from the instance tokens of the same class. To emphasize this distinction, we use subtraction rather than addition in Eq. 2 to directly exclude information related to the anchor class, thus the exclusive pattern tokens represent information contrary to instance ones. Specifically, we define the exclusive-pattern tokens $P^{-} \in \mathbb{R}^{M \times D}$ as follows: + +$$ +\bar {P} _ {m} ^ {-} = P _ {m} - \mathrm {C A} \left(P _ {m}, F, F\right), \tag {6} +$$ + +$$ +P _ {m} ^ {-} = \bar {P} _ {m} ^ {-} + \operatorname {M L P} \left(\bar {P} _ {m} ^ {-}\right). \tag {7} +$$ + +Then, we define the probability of the query belonging to the $n$ -th category $p^{-}(y^{Q} = n|P^{S}, P^{Q})$ with the exclusive and instance pattern tokens, as follows: + +$$ +\begin{array}{l} \mathrm{ND}(P_{n}^{S},P^{Q}) = \min_{\substack{1\leq o\leq N\\ o\neq n}}\sum_{m = 1}^{M}\big(-d(P_{o,m}^{S - },P_{m}^{Q + })\big) \tag{8} \\ - d (P _ {o, m} ^ {S +}, P _ {m} ^ {Q -})), \\ \end{array} +$$ + +$$ +p ^ {-} \left(y ^ {Q} = n \mid P ^ {S}, P ^ {Q}\right) = \operatorname {s o f t m a x} \left(\mathrm {N D} \left(P ^ {S}, P ^ {Q}\right); n\right). \tag {9} +$$ + +To accurately estimate the category of the query using the above probability, the exclusive pattern tokens should be similar to the instance pattern tokens of other classes, while being differentiated from those of the same class. Thus, as learning progresses, the exclusive pattern tokens effectively encode the otherness of the same-class instance pattern tokens, as intended. + +# 3.5. Adaptation of Support Pattern Tokens + +Although the pattern tokens learn globally discriminative patterns across base categories in the training phase, they might not be as effective for novel classes. Therefore, we propose an adaptation process for the support set to achieve a clear separation between categories within an episode. + +Specifically, we remove common information shared with other classes from instance pattern tokens and allocate the discriminative information of other classes into exclusive pattern tokens. To regulate the degree of adaptation for each episode, we use the cosine similarity between pattern tokens of the anchor class and other classes, representing the level of semantic entanglement. The adaptive instance pattern tokens $\hat{P}^{S+} \in \mathbb{R}^{M \times D}$ are defined as follows: + +$$ +E _ {n, o, m} ^ {+} = P _ {n, m} ^ {S +} \cdot P _ {o, m} ^ {S +}, \tag {10} +$$ + +$$ +\begin{array}{l} \tilde {P} _ {n, o, m} ^ {S +} = P _ {m} + \left(1 + E _ {n, o, m} ^ {+}\right) \mathrm {C A} \left(P _ {m}, F _ {n} ^ {S}, F _ {n} ^ {S}\right) \tag {11} \\ - E _ {n, o, m} ^ {+} \mathrm {C A} \left(P _ {m}, F _ {o} ^ {S}, F _ {o} ^ {S}\right), \\ \end{array} +$$ + +$$ +\hat {P} _ {n, m} ^ {S +} = \frac {1}{N - 1} \sum_ {o = 1} ^ {N} \mathbb {1} _ {[ n \neq o ]} \left(\tilde {P} _ {n, o, m} ^ {S +} + \mathrm {M L P} \left(\tilde {P} _ {n, o, m} ^ {S +}\right)\right), \tag {12} +$$ + +where $(\cdot)$ means the cosine similarity measure between two vectors, and $P_{n,m}^{S+}$ is $m$ -th instance pattern token of $n$ -th class in the support set, while $\mathbb{1}_{[n \neq o]}$ is an indicator function that outputs 1 if $n$ is different with $o$ and 0 otherwise. + +Similarly, the adaptive pattern tokens for exclusive ones $\hat{P}^{S-} \in \mathbb{R}^{M \times D}$ are formulated as follows: + +$$ +\begin{array}{l} \tilde {P} _ {n, o, m} ^ {S -} = P _ {m} - \left(1 + E _ {n, o, m} ^ {-}\right) \mathrm {C A} \left(P _ {m}, F _ {n} ^ {S}, F _ {n} ^ {S}\right) \tag {13} \\ + E _ {n, o, m} ^ {-} \mathrm {C A} \left(P _ {m}, F _ {o} ^ {S}, F _ {o} ^ {S}\right), \\ \end{array} +$$ + +$$ +\hat {P} _ {n, m} ^ {S -} = \frac {1}{N - 1} \sum_ {o = 1} ^ {N} \mathbb {1} _ {[ n \neq o ]} \left(\tilde {P} _ {n, o, m} ^ {S -} + \mathrm {M L P} \left(\tilde {P} _ {n, o, m} ^ {S -}\right)\right). \tag {14} +$$ + +Eq. 11 extends Eq. 2 by introducing a process to remove shared information between classes while generating instance tokens. This process suppresses information shared between the anchor class $n$ and the other class $o$ . It achieves this by subtracting $\mathrm{CA}(P_m, F_o^S, F_o^S)$ from $\mathrm{CA}(P_m, F_n^S, F_n^S)$ , thereby retaining only the discriminative information for the anchor class $n$ . The extent of suppression is proportional to the degree of information overlap, controlled by $E_{n,o,m}^{+}$ . This ensures that the adaptive instance pattern tokens focus on the discriminative features of novel classes. Similarly, Eq. 13 follows the same principle but instead emphasizes the discriminative features of other classes. + +# 3.6. Training Objective and Inference + +Adjusting the probabilities defined in Eq. 5 and 9 with adaptive pattern tokens for the support set, we define two objective functions for model training as follows: + +$$ +\mathcal {L} ^ {+} = - \mathbb {1} _ {[ y ^ {Q} = n ]} \log \left(p ^ {+} \left(y ^ {Q} = n \mid \hat {P} ^ {S}, P ^ {Q}\right)\right), \tag {15} +$$ + +$$ +\mathcal {L} ^ {-} = - \mathbb {1} _ {[ y ^ {Q} = n ]} \log \left(p ^ {-} \left(y ^ {Q} = n \mid \hat {P} ^ {S}, P ^ {Q}\right)\right). \tag {16} +$$ + +Then, the final loss function $\mathcal{L}$ is defined as follows: + +$$ +\mathcal {L} = \mathcal {L} ^ {+} + \mathcal {L} ^ {-}. \tag {17} +$$ + +In short, instance pattern tokens are optimized to resemble those of the same class and exclusive tokens of other classes. At the same time, they remain distinct from instance tokens of other classes and exclusive tokens of the same class. + +As a result, the probability of the query to belong to $n$ -th class $p(y^{Q} = n)$ is defined as follows: + +$$ +p \left(y ^ {Q} = n\right) = \operatorname {s o f t m a x} \left(\mathrm {P D} \left(\hat {P} ^ {S}, P ^ {Q}\right) + \mathrm {N D} \left(\hat {P} ^ {S}, P ^ {Q}\right); n\right). \tag {18} +$$ + +# 4. Experiments + +Description of the datasets and implementation details are provided in the supplementary material. + +# 4.1. Comparison with state-of-the-art methods + +In Tab. 1, we compare our proposed method, TEAM, with previous FSAR approaches under the 5-way 1-shot and 5-shot scenarios, which are standard evaluation protocols. Using ResNet [9], a conventional backbone network, TEAM achieves state-of-the-art performance across various datasets and shot numbers, except for UCF101 (5-shot). With the ViT backbone [7], TEAM surpasses all other methods in every case. We attribute TEAM's effectiveness to its use of globally shared pattern tokens, which inherently capture discriminative patterns across diverse scenarios. This approach enables both efficient and reliable adaptation to new episodes by leveraging pattern tokens to compare videos through multiple criteria. Additionally, TEAM is not only effective but also significantly more efficient than frame- and tuple-level alignment methods, as shown in Fig. 7. These results demonstrate that TEAM is a robust alternative to existing alignment strategies in general FSAR settings. + +We further conduct experiments on SSv2-Small, a benchmark with high temporal complexity and subtle motion differences, to validate TEAM's effectiveness in handling complex temporal dynamics. As shown in Tab. 2, TEAM surpasses other methods, except for TATs which use an additional off-the-shelf point tracker (PT). Without PT (TATs w/o PT) which generates tailored object trajectories for temporal actions, TEAM also outperforms TATs even in SSv2-Small. These results suggest that the benefits of TEAM remain robust in the context of temporally challenging actions. + +Cross-domain evaluation is widely used to assess a model's generalization capability [5, 6, 12, 21]. It typically leverages knowledge learned from coarse datasets to recognize actions in more specific and nuanced domains, highlighting the model's adaptability. Following this protocol, we compare TEAM with other FSAR methods in Tab. 3. We + +Table 1. Performance comparison on FSAR benchmark datasets under the 5-way setting. The best results are bolded, while the second-best are underlined. $\dagger$ indicates that evaluation was conducted in a transductive setting, utilizing relations between queries. * denotes reproduced performances, as results with ViT were unavailable. + +
ModelReferenceBackboneHMDB51KineticsUCF101
1-shot5-shot1-shot5-shot1-shot5-shot
ProtoNet [22]NeurIPS 201754.268.464.577.974.089.6
CMN [31]ECCV 2018--57.376.0--
OTAM [3]CVPR 202054.568.072.284.279.988.9
ARN [29]ECCV 202045.560.663.782.466.383.1
TRX [20]CVPR 202153.175.663.685.978.296.1
TA2N [17]AAAI 202259.773.972.885.881.995.1
STRM [23]CVPR 202252.377.362.986.780.596.9
HyRSM [24]CVPR 202260.376.073.786.183.994.7
MTFAN [26]CVPR 202259.074.674.687.484.895.1
HCL [30]ECCV 202259.176.373.785.882.593.9
Nguyen et al. [19]ECCV 202259.676.974.387.484.995.9
Huang et al. [10]ECCV 202260.177.073.386.471.491.0
SloshNet [27]AAAI 2023-77.5-87.0-97.1
MoLo [25]CVPR 202360.877.474.085.686.095.5
GgHM†[28]ICCV 202361.276.974.987.485.296.3
TEAMCVPR 202562.878.475.188.287.296.2
OTAM* [3]CVPR 202069.383.182.291.692.997.5
TRX* [20]CVPR 202161.382.276.890.485.897.8
STRM* [23]CVPR 202256.783.275.890.890.897.6
MoLo* [25]CVPR 202369.983.282.091.293.697.8
TATs [13]ECCV 202460.077.081.991.192.095.5
TEAMCVPR 202570.985.583.392.994.598.8
+ +Table 2. Performance comparison in SSv2-Small dataset using ViT. TATs w/o PT refers to TATs without the additional off-the-shelf Point Tracker. Symbols match those in Tab. 1 + +
ModelReferenceSSv2-Small
1-shot5-shot
OTAM* [3]CVPR 202044.458.2
TRX* [20]CVPR 202142.059.1
STRM* [23]CVPR 202242.560.4
MoLo* [25]CVPR 202345.158.8
TATs [13]ECCV 202447.964.4
TATs w/o PT [13]ECCV 202444.662.5
TEAMCVPR 202547.263.1
+ +train all models on 5-way 5-shot tasks using the Kinetics dataset with the ViT backbone and evaluate them on the same setting across different target datasets. As shown in Tab. 3, TEAM consistently outperforms prior methods, indicating strong generalization beyond the training domain. This demonstrates TEAM's robustness to temporal variations with pattern tokens capturing globally distinct patterns. + +# 4.2. Ablation Study + +Unless specifically stated, experiments are conducted in the 5-way 1-shot scenario using the ResNet-50. + +Table 3. Performance comparison across cross-domain scenarios. All models are trained on the 5-way 5-shot scenario using the Kinetics dataset with the ViT backbone, and evaluated on the same task across various datasets. Symbols match those in Tab. 1. + +
ModelReferenceHMDB51UCF101
OTAM* [3]CVPR 202077.296.5
TRX* [20]CVPR 202178.197.9
STRM* [23]CVPR 202278.598.0
MoLo* [25]CVPR 202377.797.4
TEAMCVPR 202581.198.3
+ +Component Ablation. Tab. 4 presents the performance when our component is applied sequentially. (a) shows that the use of instance pattern tokens solely achieves decent performances over the methods that rely on pre-defined temporal units and brute-force alignment processes. This validates the advantage of our core intuition that temporal alignment-free matching is beneficial for its flexibility in video understanding. Succeedingly, in (b), the exclusive tokens are shown to yield additional improvements, indicating that the different strategy in learning pattern tokens complements the original approach. Conversely, we observe slight performance drops in (c) where we implement the adaptation process without the management in the adaptation magnitude. This implies the importance of managing the degree + +Table 4. Ablation study for each component. $P^{+}$ , $P^{-}$ , and $\hat{P}$ denote the instance, exclusive, adaptive pattern tokens, respectively, while $E$ means entanglement in Eq. 10 and 13. H, K, and U also refer to HMDB51, Kinetics, and UCF101, respectively. + +
P+P-P̂EHKU
(a)✓61.874.686.7
(b)✓✓62.575.086.8
(c)✓✓✓62.274.887.1
(d)✓✓✓✓62.875.187.2
+ +of adaptation with the consideration of given class compositions. The performance gains in (d) validate our claim. + +Number of pattern tokens. In Fig. 3, we present an ablation study on the effect of the number of pattern tokens in TEAM. The results show that TEAM remains robust across a wide range of token counts, except when the value deviates significantly from the optimal (60 for all datasets). These consistent trends across all datasets suggest that our method is not sensitive to the number of pattern tokens, demonstrating its strong robustness. + +Varying N and K for N-way K-shot. Beyond standard FSAR settings, we further validate TEAM's effectiveness under varying N-way K-shot configurations on Kinetics, following prior work [25]. Specifically, we vary the number of classes $N$ from 5 to 10, and the number of video instances per class $K$ from 1 to 5. As shown in Fig. 4, TEAM consistently outperforms prior approaches across all configurations. These results confirm that TEAM is robust to different FSAR settings, with consistent benefits across diverse scenarios. + +# 4.3. Analysis + +What do pattern tokens encode? Pattern tokens are leveraged to incorporate video features and have been verified to be superior. To assess the video understanding capability of pattern tokens, we examine whether pattern tokens focus on the same context for the same category. In Fig. 5, we visualize the attention weights of a pattern token that effectively represents specific classes. Pattern tokens are considered effective in representing a class when the output pattern tokens for instances of the same class exhibit the highest similarity among all pattern tokens. As observed, pattern tokens consistently focus on the context even when the same action occurs in different frames, at varying speeds, and over different lengths. This implies that pattern tokens address the limitations of frame- and tuple-level alignment approaches, where actions can only be compared within pre-defined units. + +Necessity for multiple pattern tokens. We adopt multiple pattern tokens to capture diverse discriminative patterns across base categories, using these patterns as a foun + +![](images/17445a7cc641c8c3ef7b63935122016144885a84af7c0e94f4260d960b9153f2.jpg) +Figure 3. Ablation study for the number of pattern tokens. + +![](images/77bd2e4943d76ef3f7ee738ea199ef6740348d37878c8f6d421406913c4c36b1.jpg) +Figure 4. N-way 1-shot and 5-way K-shot results. + +dation for distinguishing novel classes. To verify that the learned patterns effectively differentiate novel classes, we compute the discriminative power of each instance pattern token within each category and rank the tokens accordingly. Discriminative power is defined as ' (intra similarity - inter similarity)'. Intra similarity measures the average similarity between instance-specific pattern tokens from all videos within the same category and their prototype (average), while inter similarity is the maximum similarity between the prototype of one class and the prototypes of other classes. The resulting heatmap of discriminative power is shown in Fig. 6, where distinct heatmap arrangements between classes can be observed. The variation in discriminative rank indicates that useful patterns differ by category, supporting the need for multiple learnable pattern tokens to capture a variety of discriminative patterns, forming a foundation for classifying novel categories. + +Computational time for matching process. In this study, we compare the computational overhead with alignment-based approaches. Computational costs for varying numbers of frames are shown in Fig. 7. As discussed in Sec. 1, alignment-based strategies require quadratic computation time due to pairwise comparison between temporal units, whereas TEAM with token-wise comparison has linear time complexity—especially evident when the number of frames exceeds 32. Although this issue is less critical for FSAR datasets, which consist of trimmed, short-term videos, it becomes significant in real-world scenarios with untrimmed, long-term videos. In this regard, we argue that our proposed + +![](images/3ec7e23c6ca02ecc059306bba68c52749a567fe97e8b4434767d3843b8b3425c.jpg) +Figure 5. Visualization of cross-attention weights during instance pattern token aggregation. The same pattern tokens are visualized for videos in each class. Green regions in each frame metaphorically represent the percentage of the frame involvement in the instance pattern token. (a) For 'Diving Cliff', the visualized token primarily responds to the moment when water splashes when people dive. (b) In 'Playing Trumpet', the token focuses on the scenes where people hold a trumpet to their mouths. + +![](images/16a883bbc2f725f971bd8fb042d8a6dd9a84408e84ee9bf0d6b426c76be3cdd4.jpg) + +![](images/0a14a4a9b2ccad5a158a7c16eb9f74f082d5efd1e8b4bb0393d2d16c0a1506fe.jpg) +Figure 6. Heatmap visualizing the class-wise discriminative power of each instance pattern token for novel classes within HMDB51. Differences in discrimination ranks across classes indicate that each class has unique discriminative patterns. + +![](images/18769f9cda33c05b725a1af7011a123aa25b263d719482aba7ae1eae3645f4ec.jpg) +Figure 7. Comparison of computational time (ms) for video matching with alignment-based strategies across varying frame counts. OTAM and TRX are used as representatives of the frame- and tuple-level alignment approaches, respectively. Times reflect only the video matching process and exclude feature extraction by the backbone network, which is consistent across all methods. + +temporal alignment-free matching offers a compelling alternative to alignment-based approaches. + +FLOPs and parameter. We analyze FLOPs and parameter counts of TEAM and other methods, as shown in Tab. 5. While TEAM achieves remarkable performance, it has the fewest parameters except for OTAM, which relies solely on backbone parameters. The higher parameter counts in other methods stem from their complex designs, including temporal transformers. These results highlight the effectiveness of + +Table 5. Comparison of FLOPs and parameter counts, which are measured in G (giga operations) and M (million), respectively. + +
ModelResNetViT
FLOPsParameterFLOPsParameter
OTAM [3]19823.580985.8
TRX [20]20547.181088.8
STRM [23]20145.581088.8
SloshNet [27]594168.9--
MoLo [25]23789.583099.1
TATs [13]--117897.7
TEAM22432.081387.0
+ +pattern tokens for FSAR, providing a lightweight alternative to heavier models. Moreover, although TEAM has higher FLOPs than OTAM and TRX, its efficient parallel operations ensure low computational time, as shown in Fig. 7. + +# 5. Conclusion + +In this paper, we proposed a novel TEmporal Alignment-free Matching (TEAM) for Few-Shot Action Recognition, that achieves both flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and brute-force alignment for the video comparison. Concretely, TEAM integrates each video with a fixed set of pattern tokens that encapsulate globally discriminative patterns within the video instance, regardless of action length and speed. Moreover, TEAM has inherent efficiency by computing token-wise similarity for matching video. Upon these pattern tokens, we further proposed an adaptation process for establishing clear boundaries between classes by removing common information. We validated the effectiveness of TEAM with extensive experiments. + +Acknowledgements This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2020-0-01821, RS-2019-II190421, RS-2024-00459618, RS-2024-00360227, RS-2024-00437102, RS-2024-00437633), and MSIT/NRF (No. RS-2024-00357729). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3 +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1 +[3] Kaidi Cao, Jingwei Ji, Zhangjie Cao, Chien-Yi Chang, and Juan Carlos Niebles. Few-shot video classification via temporal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10618-10627, 2020. 1, 2, 6, 8 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3 +[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 5 +[6] Anoop Cherian and Stephen Gould. Second-order temporal pooling for action recognition. International Journal of Computer Vision, 127:340-362, 2019. 5 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3, 5 +[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[10] Yifei Huang, Lijin Yang, and Yoichi Sato. Compound prototype matching for few-shot action recognition. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 6 +[11] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General + +perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3 +[12] Hirokatsu Kataoka, Tenga Wakamiya, Kensho Hara, and Yutaka Satoh. Would mega-scale datasets further enhance spatiotemporal 3d cnns? arXiv preprint arXiv:2004.04968, 2020.5 +[13] Pulkit Kumar, Namitha Padmanabhan, Luke Luo, Sai Saketh Rambhatla, and Abhinav Shrivastava. Trajectory-aligned space-time tokens for few-shot action recognition. In European Conference on Computer Vision, pages 474-493. Springer, 2025. 6, 8 +[14] SuBeen Lee, WonJun Moon, and Jae-Pil Heo. Task discrepancy maximization for fine-grained few-shot classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5331-5340, 2022. 3 +[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3 +[16] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4100–4110, 2023. 3 +[17] Shuyuan Li, Huabin Liu, Rui Qian, Yuxi Li, John See, Mengjuan Fei, Xiaoyuan Yu, and Weiyao Lin. Ta2n: Two-stage action alignment network for few-shot action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1404-1411, 2022. 6 +[18] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 3 +[19] Khoi D Nguyen, Quoc-Huy Tran, Khoi Nguyen, Binh-Son Hua, and Rang Nguyen. Inductive and transductive few-shot video classification via appearance and temporal alignments. In European Conference on Computer Vision, pages 471-487. Springer, 2022. 6 +[20] Toby Perrett, Alessandro Masullo, Tilo Burghardt, Majid Mirmehdi, and Dima Damen. Temporal-relational crosstransformers for few-shot action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 475-484, 2021. 1, 2, 3, 6, 8 +[21] Sarinda Samarasinghe, Mamshad Nayeem Rizve, Navid Kardan, and Mubarak Shah. Cdfsl-v: Cross-domain few-shot learning for videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11643-11652, 2023. 5 +[22] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1, 3, 6 +[23] Anirudh Thatipelli, Sanath Narayan, Salman Khan, Rao Muhammad Anwer, Fahad Shahbaz Khan, and Bernard Ghanem. Spatio-temporal relation modeling for few-shot + +action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19958-19967, 2022. 1, 2, 3, 6, 8 +[24] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Mingqian Tang, Zhengrong Zuo, Changxin Gao, Rong Jin, and Nong Sang. Hybrid relation guided set matching for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19948-19957, 2022. 2, 6 +[25] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Changxin Gao, Yingya Zhang, Deli Zhao, and Nong Sang. Molo: Motion-augmented long-short contrastive learning for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18011-18021, 2023. 1, 2, 6, 7, 8 +[26] Jiamin Wu, Tianzhu Zhang, Zhe Zhang, Feng Wu, and Yongdong Zhang. Motion-modulated temporal fragment alignment network for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151-9160, 2022. 6 +[27] Jiazheng Xing, Mengmeng Wang, Yong Liu, and Boyu Mu. Revisiting the spatial and temporal modeling for few-shot action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3001-3009, 2023. 1, 2, 6, 8 +[28] Jiazheng Xing, Mengmeng Wang, Yudi Ruan, Bofan Chen, Yaowei Guo, Boyu Mu, Guang Dai, Jingdong Wang, and Yong Liu. Boosting few-shot action recognition with graph-guided hybrid matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1740-1750, 2023. 1, 6 +[29] Hongguang Zhang, Li Zhang, Xiaojuan Qi, Hongdong Li, Philip HS Torr, and Piotr Koniusz. Few-shot action recognition with permutation-invariant attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 525-542. Springer, 2020. 1, 6 +[30] Sipeng Zheng, Shizhe Chen, and Qin Jin. Few-shot action recognition with hierarchical matching and contrastive learning. In European Conference on Computer Vision, pages 297-313. Springer, 2022. 6 +[31] Linchao Zhu and Yi Yang. Compound memory networks for few-shot video classification. In Proceedings of the European Conference on Computer Vision (ECCV), pages 751-766, 2018. 1, 6 \ No newline at end of file diff --git a/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/images.zip b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a182fed5644315d5935db2dc43c792731c257d18 --- /dev/null +++ b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ebabd8ee35cb10a73fccbc4c4491250975c5edd09b4b0f709af0f948b394d41 +size 832706 diff --git a/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/layout.json b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..96d97cc505c2fef04e0801c7d73fdb63981cda47 --- /dev/null +++ b/2025/Temporal Alignment-Free Video Matching for Few-shot Action Recognition/layout.json @@ -0,0 +1,7390 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 531, + 121 + ], + "type": "text", + "content": "Temporal Alignment-Free Video Matching for Few-shot Action Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 143, + 447, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 143, + 447, + 172 + ], + "spans": [ + { + "bbox": [ + 150, + 143, + 447, + 172 + ], + "type": "text", + "content": "SuBeen Lee, WonJun Moon, Hyun Seok Seong, Jae-Pil Heo* Sungkyunkwan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 174, + 440, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 174, + 440, + 186 + ], + "spans": [ + { + "bbox": [ + 152, + 174, + 440, + 186 + ], + "type": "text", + "content": "{leesb7426，wjun0830，gustjrdl95，jaepilheo}@skku.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "content": "Few-Shot Action Recognition (FSAR) aims to train a model with only a few labeled video instances. A key challenge in FSAR is handling divergent narrative trajectories for precise video matching. While the frame- and tuple-level alignment approaches have been promising, their methods heavily rely on pre-defined and length-dependent alignment units (e.g., frames or tuples), which limits flexibility for actions of varying lengths and speeds. In this work, we introduce a novel Temporal Alignment-free Matching (TEAM) approach, which eliminates the need for temporal units in action representation and brute-force alignment during matching. Specifically, TEAM represents each video with a fixed set of pattern tokens that capture globally discriminative clues within the video instance regardless of action length or speed, ensuring its flexibility. Furthermore, TEAM is inherently efficient, using token-wise comparisons to measure similarity between videos, unlike existing methods that rely on pairwise comparisons for temporal alignment. Additionally, we propose an adaptation process that identifies and removes common information across classes, establishing clear boundaries even between novel categories. Extensive experiments demonstrate the effectiveness of TEAM. Codes are available at github.com/leesb7426/TEAM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 559, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 693 + ], + "type": "text", + "content": "Recently, action recognition has achieved remarkable performances with the help of numerous labeled videos [2, 5, 8]. However, those performances are not guaranteed when novel categories are given with few labeled examples [3, 22, 29, 31]. To resolve this problem, Few-Shot Action Recognition (FSAR) has been proposed [31]. Particularly, FSAR employs episodic learning which iteratively simulates the episode of learning a set of new classes. Each episode is given a support set to train a model, composed of a few labeled examples for newly given classes, to distinguish the category of a set of unlabeled videos, coined as a query set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 213, + 548, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 548, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 548, + 334 + ], + "type": "text", + "content": "Due to different class compositions within every episode, metric learning, a classifier-free strategy, is predominantly employed in FSAR [3, 20, 23-25, 27, 28]. Distances between the representations of support and query sets are computed to infer the category of a query video. Yet, naively constructing video representations, such as average pooling, are shown to be poor due to an insufficient understanding of temporal information [3, 31]. Therefore, the previous approaches typically focused on aligning temporal sequences between videos [3, 20, 23-25, 27, 28]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 336, + 548, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 548, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 548, + 443 + ], + "type": "text", + "content": "As the base units for temporal alignment between videos, the frame and tuple are commonly employed (illustrated in Fig. 1 (a) and (b)). Both frame- and tuple-based alignment approaches measure pairwise similarities between each unit and then derive an optimal alignment path for all units of query video. This path is regarded as the temporal alignment between support and query videos, the query video is assigned with the class label of the support video with minimum cumulative distance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 447, + 548, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 548, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 548, + 603 + ], + "type": "text", + "content": "Although these two temporal alignment paradigms have shown effectiveness, they suffer from inflexibility and inefficiency. The inflexibility arises from their reliance on predefined units, such as frames or tuples, for action representation. These methods typically struggle to capture actions if the pre-defined units do not align with the varying action durations. Moreover, their alignment cost increases quadratically with the number of frames, as shown in Fig. 7. This issue has been overlooked since typical FSAR benchmarking datasets consist mostly of trimmed, short-term videos. The quadratic cost arises because their alignment strategies rely on pairwise comparisons between the units of support and query videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "In this regard, we introduce a novel Temporal Alignment-free Matching (TEAM) approach that achieves both flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and brute-force alignment for video comparison. Specifically, TEAM represents each video with a fixed set of pattern tokens with cross-attention mechanisms to capture diverse clues within the video instance which are useful for distinguishing categories. Therefore, these pattern tokens flexibly identify discrimina" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5412" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 255, + 310 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 255, + 310 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 255, + 310 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 255, + 310 + ], + "type": "image", + "image_path": "8dc4c5373c602a003d4e4bb9a03cbb3f0feb8d8b0ab464462b94c6dfbc547b0c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 314, + 548, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 314, + 548, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 548, + 403 + ], + "type": "text", + "content": "Figure 1. Comparison with alignment-based approaches on the \"Diving Cliff\" class. (a) Frame Alignment: For each frame of the query video, the most corresponding frame in the support video is identified, highlighting the importance of precise frame-level alignment. (b) Tuple Alignment: The support and query videos are compared in sub-sequence units to account for variations in action speed. The set of varying tuples is pre-defined. (c) Temporal Alignment-free Matching: Both in support and query, the video features are initially integrated using pattern tokens, which encode globally discriminative patterns. Then, it directly compares the corresponding aggregated features of the support and query videos. Pattern-based aggregation is both efficient and flexible, as it does not require an alignment process and is unaffected by differences in frame count or action speed. Note the text description for each pattern token (green) is intended to provide an intuitive understanding of what each pattern token represents." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 258, + 69, + 545, + 311 + ], + "blocks": [ + { + "bbox": [ + 258, + 69, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 258, + 69, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 258, + 69, + 545, + 311 + ], + "type": "image", + "image_path": "a21e76e999c3d311f7f35d937d47561d4206e44de7774e3b39fa62d47b29a39a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 415, + 289, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 535 + ], + "type": "text", + "content": "tive features varying lengths across each video, as illustrated in Fig. 1 (c). Furthermore, these pattern tokens are trained to capture globally discriminative clues across the dataset. This is because all videos leverage the entire set of pattern tokens without an alignment process, rather than relying on specific tokens determined by alignment, as shown in Fig. 1 (c). Additionally, this token-wise comparison results in the efficiency of TEAM by eliminating the brute-force alignment process computing pairwise similarities between units for the query video matching." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 538, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 288, + 624 + ], + "type": "text", + "content": "However, pattern tokens may not be fully discriminative for every episode with novel class compositions. To address this, we adapt the class boundaries formed by pattern tokens of support set videos. Particularly, we identify the patterns that commonly appear across classes and remove them from each support video token. This process establishes clear class boundaries, improving class discrimination for query videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 626, + 195, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 626, + 195, + 638 + ], + "spans": [ + { + "bbox": [ + 59, + 626, + 195, + 638 + ], + "type": "text", + "content": "To sum up, our contributions are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 642, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 642, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 288, + 689 + ], + "type": "text", + "content": "- We propose Temporal Alignment-free Matching for FSAR equipped with flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and alignment in video comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "- We develop an episode adaptation process of encoded pattern tokens. This establishes clear boundaries within" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 415, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 545, + 437 + ], + "type": "text", + "content": "the support set for each class composition by excluding information shared with other classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 438, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 545, + 461 + ], + "type": "text", + "content": "- Our state-of-the-art performances on various benchmarks for FSAR validate the effectiveness of our TEAM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 472, + 392, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 392, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 392, + 484 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 491, + 465, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 465, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 465, + 504 + ], + "type": "text", + "content": "2.1. Few-Shot Action Recognition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "A predominant learning strategy in Few-Shot Action Recognition (FSAR) is metric learning, which entails inferring the category of a query video based on a pre-defined distance metric [3, 20, 23-25, 27]. Previous methods focused on measuring the distance between video features through the development of alignment algorithms, either at the frame-level [3, 24, 25] or tuple-level [20, 23, 27]. As one of the initial frame alignment methods, OTAM [3] aligns the temporal order of query and support videos at the frame-level to precisely estimate the distance between potential query-support pairs. Yet, as it contains a strong assumption that all actions are in a fixed sequence, HyRSM [24] proposed a flexible alignment metric alleviating the strictly ordered constraints. More recently, MoLo [25] focused on learning long-range temporal perception and motion dynamics through frame-level objectives. On the other hand, the stream of tuple-level alignment is to gain robustness to varying action lengths" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5413" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 67, + 545, + 265 + ], + "blocks": [ + { + "bbox": [ + 47, + 67, + 545, + 265 + ], + "lines": [ + { + "bbox": [ + 47, + 67, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 47, + 67, + 545, + 265 + ], + "type": "image", + "image_path": "64516495370e30de382d59ccb6d635e7389685c1a91aad2073df03ca14bbe8ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 548, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 548, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 548, + 369 + ], + "type": "text", + "content": "Figure 2. Illustration of our approach using a single pattern token (" + }, + { + "bbox": [ + 46, + 270, + 548, + 369 + ], + "type": "inline_equation", + "content": "M = 1" + }, + { + "bbox": [ + 46, + 270, + 548, + 369 + ], + "type": "text", + "content": ") in the 3-way 3-shot scenario without query. This figure illustrates our approach, where instance (+) and exclusive (-) pattern tokens represent integrated video features based on the pattern token. (Left) Randomly initialized pattern tokens are optimized with classes in two complementary ways. First, instance pattern tokens (+) are encouraged to cluster with other instances' tokens of the same class while repelling those of other categories. On the other hand, exclusive pattern tokens (-) learn to represent the otherness of each instance by positioning themselves in the embedding space of other classes. (Right) Although these two types of tokens are discriminative for distinguishing base video categories, they may not fully capture the finer details needed for novel categories. To address this, we propose an adaptation process of support pattern tokens for novel classes to refine the class decision boundaries. Note that when multiple pattern tokens are used, these processes run in parallel, with instance and exclusive tokens being compared only within the same pattern token." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 381, + 289, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 289, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 289, + 489 + ], + "type": "text", + "content": "and temporal offsets. Tuple alignment has gained attention for its flexibility in encoding actions of different speeds and temporal offsets [20]. Additionally, STRM [23] incorporated a spatio-temporal enrichment module to enhance both object and motion representations. Yet, tuple-based approaches require pre-defined temporal tuples to provide hints for possible forms of actions in videos. In contrast, our approach is flexible since we do not need pre-defined windows to represent action." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 502, + 229, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 229, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 229, + 514 + ], + "type": "text", + "content": "2.2. Token-based Feature Aggregation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 290, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 290, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 290, + 713 + ], + "type": "text", + "content": "Feature aggregation generates compact contextual representations, providing both effectiveness and efficiency across various fields [1, 4, 7, 11, 14-16, 18, 22]. There are various strategies to implement feature aggregation including different pooling methods [14, 22]. Among them, our method closely aligns with token-based approaches. Simply put, the stream of token-based aggregation integrates the input context within the learnable tokens. For instance, DETR [4] introduces learnable object queries to aggregate spatial features for object detection, while BLIP-2 [15] and Flamingo [1] use learnable tokens to encode image information for multimodal understanding. Similarly to prior methods, we use learnable tokens to aggregate video information. However, our TEAM differs in that the feature aggregation is implemented in two complementary ways, providing more robust class boundaries. Additionally, we also propose a token adap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 381, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 381, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 547, + 406 + ], + "type": "text", + "content": "tation process to maintain the benefits of capturing discriminative details while encoding instances of novel classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 418, + 362, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 362, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 362, + 430 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 437, + 430, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 430, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 430, + 449 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": "In the standard Few-Shot Action Recognition (FSAR), we are given two sets of data: meta-train set for model training " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{base} = \\{(v_i,y_i),y_i\\in C_{base}\\}" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": " and meta-test set for model testing " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test} = \\{(v_i,y_i),y_i\\in C_{novel}\\}" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": ". While " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": " is a video, " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "C_{base}" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "C_{novel}" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": " denote class sets for meta-training and meta-testing, respectively (" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "C_{base}\\cap C_{novel} = \\phi" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": "). Generally, the training and testing processes of FSAR are composed of episodes where each episode validates the model's adaptability to a novel class set. Each episode comprises randomly selected " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": " classes and each class consists of " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": " labeled videos and " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": " unlabeled videos, i.e., " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": "-shot episode. Whereas a set of labeled videos is called the support set " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "S = \\{(v_i,y_i)\\}_{i = 1}^{N\\times K}" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": ", the other set of unlabeled videos are often dubbed as the query set " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "Q = \\{(v_i,y_i)\\}_{i = 1}^{N\\times U}" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": ". In the below sections, we assume " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": "-way 1-shot setting where the number of query video " + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 456, + 547, + 651 + ], + "type": "text", + "content": " is also set to 1 for better clarity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 658, + 373, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 373, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 373, + 670 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "The schematic illustration of our method is shown in Fig. 2. To eliminate the need for pre-defined units in action representation, we integrate video features using pattern tokens," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5414" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "which capture globally discriminative features across categories. Specifically, our integration process operates in two complementary ways, generating instance and exclusive pattern tokens. Instance pattern tokens are designed to capture shared patterns among videos of the same class. However, these shared patterns may not always be present in every video of the class. In such cases, relying solely on instance pattern tokens for classification is ineffective. To address this, exclusive pattern tokens are introduced to represent \"otherness\"—absent patterns in a given class. This ensures accurate classification, even when key discriminative patterns are missing from certain videos. While these pattern tokens are effective for base categories, they may lack the discriminative details for novel classes. To overcome this, we propose an adaptation process for both instance and exclusive pattern tokens in support videos, reinforcing clear boundaries to distinguish novel classes within each episode. Notably, pattern tokens are only compared within their type (i.e., token-wise comparison), eliminating the need for an alignment process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 320, + 183, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 320, + 183, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 183, + 332 + ], + "type": "text", + "content": "3.3. Instance Pattern Tokens" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 339, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 288, + 376 + ], + "type": "text", + "content": "Given uniformly sampled " + }, + { + "bbox": [ + 46, + 339, + 288, + 376 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 339, + 288, + 376 + ], + "type": "text", + "content": " frames from the support and query videos, we independently extract the features for each frame through the image features extractor, as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 385, + 288, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 385, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 94, + 385, + 288, + 399 + ], + "type": "interline_equation", + "content": "F = \\left[ g _ {\\theta} \\left(x _ {1}\\right), g _ {\\theta} \\left(x _ {2}\\right), \\dots , g _ {\\theta} \\left(x _ {T}\\right) \\right], \\tag {1}", + "image_path": "3e23e39681296f98d0aff2d50ffc64de735e8cea4806bd5fce81d6bd513ee3d2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "text", + "content": " denotes the image feature extractor " + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times 3}\\to" + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D" + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "text", + "content": " means " + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "text", + "content": " -th frame of the video. " + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "text", + "content": " are height and width of the frame while " + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 407, + 287, + 444 + ], + "type": "text", + "content": " is size of the features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "text", + "content": "Unlike existing works that utilize pre-defined units (i.e., frame or tuple) for representing actions, we integrate video features based on pattern tokens that encode globally discriminative patterns. Initially, we define a pattern pool that consists of " + }, + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "text", + "content": " learnable pattern tokens " + }, + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "inline_equation", + "content": "P = [P_{1}, P_{2}, \\dots, P_{M}]" + }, + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "text", + "content": ". Then, we use a cross-attention mechanism to encode instance patterns within " + }, + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "text", + "content": ". Formally, instance pattern tokens " + }, + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "inline_equation", + "content": "P^{+} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 47, + 445, + 289, + 540 + ], + "type": "text", + "content": " are computed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 552, + 287, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 552, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 107, + 552, + 287, + 567 + ], + "type": "interline_equation", + "content": "\\bar {P} _ {m} ^ {+} = P _ {m} + \\mathrm {C A} (P _ {m}, F, F), \\tag {2}", + "image_path": "b4177d0fc1b184ac4abcd2e2d3827a858ecdd61d6f401ba628d6d479afd66857.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 578, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 578, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 116, + 578, + 287, + 594 + ], + "type": "interline_equation", + "content": "P _ {m} ^ {+} = \\bar {P} _ {m} ^ {+} + \\operatorname {M L P} \\left(\\bar {P} _ {m} ^ {+}\\right), \\tag {3}", + "image_path": "c18b3a535b6e3858e0ff860e42ed3b6eeebffa9ba02c6e21dfb53d0198d5d703.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "text", + "content": "-th component of " + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{CA}(\\cdot, \\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "text", + "content": " is the cross-attention receiving query, key, and value as inputs. And, " + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}(\\cdot)" + }, + { + "bbox": [ + 46, + 601, + 287, + 636 + ], + "type": "text", + "content": " indicates a light fully-connected block." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "text", + "content": "Based on instance pattern tokens of support set " + }, + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "inline_equation", + "content": "P^{S+}" + }, + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "text", + "content": ", the probability for the query " + }, + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "inline_equation", + "content": "P^{Q+}" + }, + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "text", + "content": " to belong to " + }, + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "text", + "content": "-th category " + }, + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "inline_equation", + "content": "p^+(y^Q = n|P^S, P^Q)" + }, + { + "bbox": [ + 46, + 637, + 288, + 673 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 87, + 683, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 683, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 87, + 683, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\mathrm {P D} \\left(P _ {n} ^ {S}, P ^ {Q}\\right) = \\sum_ {m = 1} ^ {M} - d \\left(P _ {n, m} ^ {S +}, P _ {m} ^ {Q +}\\right), \\tag {4}", + "image_path": "e2dfe4141719d036f6eb88204d5dee78e29ccee957e4f793d3a5571f0b6fc658.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 69, + 547, + 85 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 69, + 547, + 85 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 547, + 85 + ], + "type": "interline_equation", + "content": "p ^ {+} (y ^ {Q} = n | P ^ {S}, P ^ {Q}) = \\mathrm {s o f t m a x} (\\mathrm {P D} (P ^ {S}, P ^ {Q}); n), \\quad (5)", + "image_path": "ab794bf5a731925e04c0cbed177b69f7834eee9bfe03f8fc69c8e10ef950bebb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "text", + "content": " is a distance metric, i.e., the cosine distance, while " + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "inline_equation", + "content": "P_{n,m}^{S+}" + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "inline_equation", + "content": "P_m^{Q+}" + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "text", + "content": "-th instance pattern tokens of support belonging to category " + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "text", + "content": " and query, respectively. In addition, " + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "inline_equation", + "content": "\\operatorname{softmax}(\\cdot; n)" + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 92, + 547, + 152 + ], + "type": "text", + "content": "-th component of the softmax function applied to the input." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 152, + 547, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 547, + 272 + ], + "type": "text", + "content": "The key advantage of using pattern tokens is achieving window-free video understanding. Unlike previous approaches that represent action with frame- or tuple-based units, pattern tokens are not restricted by specific lengths or speeds. Furthermore, each pattern token learns globally discriminative features for video classification, since all pattern tokens contribute across the dataset rather than relying on specific tokens determined by alignment. This design of pattern tokens increases flexibility in action representation and removes the computational cost associated with alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 282, + 446, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 282, + 446, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 446, + 294 + ], + "type": "text", + "content": "3.4. Exclusive Pattern Tokens" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 300, + 547, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 432 + ], + "type": "text", + "content": "In addition to the instance pattern tokens that capture specific patterns in the video, we introduce exclusive pattern tokens, designed to encode the otherness to instance pattern tokens. In other words, the exclusive tokens are located close to the other classes' instance tokens and differentiated from the instance tokens of the same class. To emphasize this distinction, we use subtraction rather than addition in Eq. 2 to directly exclude information related to the anchor class, thus the exclusive pattern tokens represent information contrary to instance ones. Specifically, we define the exclusive-pattern tokens " + }, + { + "bbox": [ + 304, + 300, + 547, + 432 + ], + "type": "inline_equation", + "content": "P^{-} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 304, + 300, + 547, + 432 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 445, + 545, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 445, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 366, + 445, + 545, + 459 + ], + "type": "interline_equation", + "content": "\\bar {P} _ {m} ^ {-} = P _ {m} - \\mathrm {C A} \\left(P _ {m}, F, F\\right), \\tag {6}", + "image_path": "99fb28ecb04da3efccde14f2bebb6b22b87ba5f01efffeb5767ebeb2790117b3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 374, + 472, + 545, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 472, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 374, + 472, + 545, + 487 + ], + "type": "interline_equation", + "content": "P _ {m} ^ {-} = \\bar {P} _ {m} ^ {-} + \\operatorname {M L P} \\left(\\bar {P} _ {m} ^ {-}\\right). \\tag {7}", + "image_path": "5c535537a9d7872ed6041f830eec852b87a673e5fef098a4897f0b25ddc92c4d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 495, + 547, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 495, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 305, + 495, + 547, + 532 + ], + "type": "text", + "content": "Then, we define the probability of the query belonging to the " + }, + { + "bbox": [ + 305, + 495, + 547, + 532 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 495, + 547, + 532 + ], + "type": "text", + "content": "-th category " + }, + { + "bbox": [ + 305, + 495, + 547, + 532 + ], + "type": "inline_equation", + "content": "p^{-}(y^{Q} = n|P^{S}, P^{Q})" + }, + { + "bbox": [ + 305, + 495, + 547, + 532 + ], + "type": "text", + "content": " with the exclusive and instance pattern tokens, as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 542, + 545, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 542, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 318, + 542, + 545, + 595 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm{ND}(P_{n}^{S},P^{Q}) = \\min_{\\substack{1\\leq o\\leq N\\\\ o\\neq n}}\\sum_{m = 1}^{M}\\big(-d(P_{o,m}^{S - },P_{m}^{Q + })\\big) \\tag{8} \\\\ - d (P _ {o, m} ^ {S +}, P _ {m} ^ {Q -})), \\\\ \\end{array}", + "image_path": "2ab47d185494bba957de6fb2a42d0d8acae51a35bd7dcb9ce987346c5e483b88.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 606, + 545, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 606, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 606, + 545, + 621 + ], + "type": "interline_equation", + "content": "p ^ {-} \\left(y ^ {Q} = n \\mid P ^ {S}, P ^ {Q}\\right) = \\operatorname {s o f t m a x} \\left(\\mathrm {N D} \\left(P ^ {S}, P ^ {Q}\\right); n\\right). \\tag {9}", + "image_path": "2cc8dbfaf5bf0ca868a7a5f92ff9bbee7b73b47beb8783105b3785891466a3af.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": "To accurately estimate the category of the query using the above probability, the exclusive pattern tokens should be similar to the instance pattern tokens of other classes, while being differentiated from those of the same class. Thus, as learning progresses, the exclusive pattern tokens effectively encode the otherness of the same-class instance pattern tokens, as intended." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5415" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 250, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 250, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 250, + 85 + ], + "type": "text", + "content": "3.5. Adaptation of Support Pattern Tokens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 149 + ], + "type": "text", + "content": "Although the pattern tokens learn globally discriminative patterns across base categories in the training phase, they might not be as effective for novel classes. Therefore, we propose an adaptation process for the support set to achieve a clear separation between categories within an episode." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 150, + 288, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 150, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 150, + 288, + 245 + ], + "type": "text", + "content": "Specifically, we remove common information shared with other classes from instance pattern tokens and allocate the discriminative information of other classes into exclusive pattern tokens. To regulate the degree of adaptation for each episode, we use the cosine similarity between pattern tokens of the anchor class and other classes, representing the level of semantic entanglement. The adaptive instance pattern tokens " + }, + { + "bbox": [ + 46, + 150, + 288, + 245 + ], + "type": "inline_equation", + "content": "\\hat{P}^{S+} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 46, + 150, + 288, + 245 + ], + "type": "text", + "content": " are defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 251, + 287, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 251, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 118, + 251, + 287, + 268 + ], + "type": "interline_equation", + "content": "E _ {n, o, m} ^ {+} = P _ {n, m} ^ {S +} \\cdot P _ {o, m} ^ {S +}, \\tag {10}", + "image_path": "8d1da59e0d0970cf673484f05e9fc713f049a35bb11ef187c35fa8e8ec244db0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 271, + 287, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 271, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 60, + 271, + 287, + 304 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {P} _ {n, o, m} ^ {S +} = P _ {m} + \\left(1 + E _ {n, o, m} ^ {+}\\right) \\mathrm {C A} \\left(P _ {m}, F _ {n} ^ {S}, F _ {n} ^ {S}\\right) \\tag {11} \\\\ - E _ {n, o, m} ^ {+} \\mathrm {C A} \\left(P _ {m}, F _ {o} ^ {S}, F _ {o} ^ {S}\\right), \\\\ \\end{array}", + "image_path": "31e3e431c0d5f5d0053e16e715610a65edd20499369b4b8acd937657bc5bf4dc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 307, + 287, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 307, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 48, + 307, + 287, + 339 + ], + "type": "interline_equation", + "content": "\\hat {P} _ {n, m} ^ {S +} = \\frac {1}{N - 1} \\sum_ {o = 1} ^ {N} \\mathbb {1} _ {[ n \\neq o ]} \\left(\\tilde {P} _ {n, o, m} ^ {S +} + \\mathrm {M L P} \\left(\\tilde {P} _ {n, o, m} ^ {S +}\\right)\\right), \\tag {12}", + "image_path": "b2a5c8c94e33b3099b9d55b3997ba613b3828314a0895ab44ecaebb9a0c42c2c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "text", + "content": " means the cosine similarity measure between two vectors, and " + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "inline_equation", + "content": "P_{n,m}^{S+}" + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "text", + "content": "-th instance pattern token of " + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "text", + "content": "-th class in the support set, while " + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{[n \\neq o]}" + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "text", + "content": " is an indicator function that outputs 1 if " + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "text", + "content": " is different with " + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 342, + 287, + 389 + ], + "type": "text", + "content": " and 0 otherwise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "type": "text", + "content": "Similarly, the adaptive pattern tokens for exclusive ones " + }, + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\hat{P}^{S-} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "type": "text", + "content": " are formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 416, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 416, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 60, + 416, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {P} _ {n, o, m} ^ {S -} = P _ {m} - \\left(1 + E _ {n, o, m} ^ {-}\\right) \\mathrm {C A} \\left(P _ {m}, F _ {n} ^ {S}, F _ {n} ^ {S}\\right) \\tag {13} \\\\ + E _ {n, o, m} ^ {-} \\mathrm {C A} \\left(P _ {m}, F _ {o} ^ {S}, F _ {o} ^ {S}\\right), \\\\ \\end{array}", + "image_path": "49a3b6dca16263ed71863a9f4c02be569b96bf091c3ff29e3551abf80a2c75c9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "type": "interline_equation", + "content": "\\hat {P} _ {n, m} ^ {S -} = \\frac {1}{N - 1} \\sum_ {o = 1} ^ {N} \\mathbb {1} _ {[ n \\neq o ]} \\left(\\tilde {P} _ {n, o, m} ^ {S -} + \\mathrm {M L P} \\left(\\tilde {P} _ {n, o, m} ^ {S -}\\right)\\right). \\tag {14}", + "image_path": "79fc2eb92adcc67a8468609846a899ca7a28a2dd47abdb725b095bfdcb6b548b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "content": "Eq. 11 extends Eq. 2 by introducing a process to remove shared information between classes while generating instance tokens. This process suppresses information shared between the anchor class " + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "content": " and the other class " + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "content": ". It achieves this by subtracting " + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "inline_equation", + "content": "\\mathrm{CA}(P_m, F_o^S, F_o^S)" + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "inline_equation", + "content": "\\mathrm{CA}(P_m, F_n^S, F_n^S)" + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "content": ", thereby retaining only the discriminative information for the anchor class " + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "content": ". The extent of suppression is proportional to the degree of information overlap, controlled by " + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "inline_equation", + "content": "E_{n,o,m}^{+}" + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "content": ". This ensures that the adaptive instance pattern tokens focus on the discriminative features of novel classes. Similarly, Eq. 13 follows the same principle but instead emphasizes the discriminative features of other classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 638, + 226, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 226, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 226, + 651 + ], + "type": "text", + "content": "3.6. Training Objective and Inference" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 656, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 656, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 656, + 289, + 693 + ], + "type": "text", + "content": "Adjusting the probabilities defined in Eq. 5 and 9 with adaptive pattern tokens for the support set, we define two objective functions for model training as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 698, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 698, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 69, + 698, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {+} = - \\mathbb {1} _ {[ y ^ {Q} = n ]} \\log \\left(p ^ {+} \\left(y ^ {Q} = n \\mid \\hat {P} ^ {S}, P ^ {Q}\\right)\\right), \\tag {15}", + "image_path": "85b060d802ff70a1a05bf1efacbd575e72ed3fc353805a6e148579c6d90b39bf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 69, + 546, + 85 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 69, + 546, + 85 + ], + "spans": [ + { + "bbox": [ + 328, + 69, + 546, + 85 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {-} = - \\mathbb {1} _ {[ y ^ {Q} = n ]} \\log \\left(p ^ {-} \\left(y ^ {Q} = n \\mid \\hat {P} ^ {S}, P ^ {Q}\\right)\\right). \\tag {16}", + "image_path": "c99ec5dec0d7c762b313f09c8d2c5c56c7eb8c1546168fd1c517e2e01b21313e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 89, + 516, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 89, + 516, + 101 + ], + "spans": [ + { + "bbox": [ + 305, + 89, + 516, + 101 + ], + "type": "text", + "content": "Then, the final loss function " + }, + { + "bbox": [ + 305, + 89, + 516, + 101 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 89, + 516, + 101 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 393, + 111, + 546, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 111, + 546, + 124 + ], + "spans": [ + { + "bbox": [ + 393, + 111, + 546, + 124 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} ^ {+} + \\mathcal {L} ^ {-}. \\tag {17}", + "image_path": "cf32ac9c8b7fba7b5851297306badce4c739e70b44356cc8bcdb4d509d9408e4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 134, + 547, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 547, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 547, + 181 + ], + "type": "text", + "content": "In short, instance pattern tokens are optimized to resemble those of the same class and exclusive tokens of other classes. At the same time, they remain distinct from instance tokens of other classes and exclusive tokens of the same class." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 182, + 545, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 206 + ], + "type": "text", + "content": "As a result, the probability of the query to belong to " + }, + { + "bbox": [ + 304, + 182, + 545, + 206 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 182, + 545, + 206 + ], + "type": "text", + "content": "-th class " + }, + { + "bbox": [ + 304, + 182, + 545, + 206 + ], + "type": "inline_equation", + "content": "p(y^{Q} = n)" + }, + { + "bbox": [ + 304, + 182, + 545, + 206 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 215, + 546, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 215, + 546, + 230 + ], + "spans": [ + { + "bbox": [ + 302, + 215, + 546, + 230 + ], + "type": "interline_equation", + "content": "p \\left(y ^ {Q} = n\\right) = \\operatorname {s o f t m a x} \\left(\\mathrm {P D} \\left(\\hat {P} ^ {S}, P ^ {Q}\\right) + \\mathrm {N D} \\left(\\hat {P} ^ {S}, P ^ {Q}\\right); n\\right). \\tag {18}", + "image_path": "fe4b5448fc9659418250d551ad32b758d91557ff33d36a4f8f92893ff0145387.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 249, + 388, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 249, + 388, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 388, + 262 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 270, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 293 + ], + "type": "text", + "content": "Description of the datasets and implementation details are provided in the supplementary material." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 300, + 525, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 525, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 525, + 313 + ], + "type": "text", + "content": "4.1. Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 319, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 521 + ], + "type": "text", + "content": "In Tab. 1, we compare our proposed method, TEAM, with previous FSAR approaches under the 5-way 1-shot and 5-shot scenarios, which are standard evaluation protocols. Using ResNet [9], a conventional backbone network, TEAM achieves state-of-the-art performance across various datasets and shot numbers, except for UCF101 (5-shot). With the ViT backbone [7], TEAM surpasses all other methods in every case. We attribute TEAM's effectiveness to its use of globally shared pattern tokens, which inherently capture discriminative patterns across diverse scenarios. This approach enables both efficient and reliable adaptation to new episodes by leveraging pattern tokens to compare videos through multiple criteria. Additionally, TEAM is not only effective but also significantly more efficient than frame- and tuple-level alignment methods, as shown in Fig. 7. These results demonstrate that TEAM is a robust alternative to existing alignment strategies in general FSAR settings." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 522, + 547, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 642 + ], + "type": "text", + "content": "We further conduct experiments on SSv2-Small, a benchmark with high temporal complexity and subtle motion differences, to validate TEAM's effectiveness in handling complex temporal dynamics. As shown in Tab. 2, TEAM surpasses other methods, except for TATs which use an additional off-the-shelf point tracker (PT). Without PT (TATs w/o PT) which generates tailored object trajectories for temporal actions, TEAM also outperforms TATs even in SSv2-Small. These results suggest that the benefits of TEAM remain robust in the context of temporally challenging actions." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Cross-domain evaluation is widely used to assess a model's generalization capability [5, 6, 12, 21]. It typically leverages knowledge learned from coarse datasets to recognize actions in more specific and nuanced domains, highlighting the model's adaptability. Following this protocol, we compare TEAM with other FSAR methods in Tab. 3. We" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5416" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 108, + 550, + 401 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 1. Performance comparison on FSAR benchmark datasets under the 5-way setting. The best results are bolded, while the second-best are underlined. " + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": " indicates that evaluation was conducted in a transductive setting, utilizing relations between queries. * denotes reproduced performances, as results with ViT were unavailable." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 108, + 550, + 401 + ], + "lines": [ + { + "bbox": [ + 47, + 108, + 550, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 550, + 401 + ], + "type": "table", + "html": "
ModelReferenceBackboneHMDB51KineticsUCF101
1-shot5-shot1-shot5-shot1-shot5-shot
ProtoNet [22]NeurIPS 201754.268.464.577.974.089.6
CMN [31]ECCV 2018--57.376.0--
OTAM [3]CVPR 202054.568.072.284.279.988.9
ARN [29]ECCV 202045.560.663.782.466.383.1
TRX [20]CVPR 202153.175.663.685.978.296.1
TA2N [17]AAAI 202259.773.972.885.881.995.1
STRM [23]CVPR 202252.377.362.986.780.596.9
HyRSM [24]CVPR 202260.376.073.786.183.994.7
MTFAN [26]CVPR 202259.074.674.687.484.895.1
HCL [30]ECCV 202259.176.373.785.882.593.9
Nguyen et al. [19]ECCV 202259.676.974.387.484.995.9
Huang et al. [10]ECCV 202260.177.073.386.471.491.0
SloshNet [27]AAAI 2023-77.5-87.0-97.1
MoLo [25]CVPR 202360.877.474.085.686.095.5
GgHM†[28]ICCV 202361.276.974.987.485.296.3
TEAMCVPR 202562.878.475.188.287.296.2
OTAM* [3]CVPR 202069.383.182.291.692.997.5
TRX* [20]CVPR 202161.382.276.890.485.897.8
STRM* [23]CVPR 202256.783.275.890.890.897.6
MoLo* [25]CVPR 202369.983.282.091.293.697.8
TATs [13]ECCV 202460.077.081.991.192.095.5
TEAMCVPR 202570.985.583.392.994.598.8
", + "image_path": "43865325d7f9262f53c805ccd0b6dd03d264bd8c06931e5fb31752c81f756783.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 445, + 290, + 559 + ], + "blocks": [ + { + "bbox": [ + 46, + 408, + 288, + 442 + ], + "lines": [ + { + "bbox": [ + 46, + 408, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 288, + 442 + ], + "type": "text", + "content": "Table 2. Performance comparison in SSv2-Small dataset using ViT. TATs w/o PT refers to TATs without the additional off-the-shelf Point Tracker. Symbols match those in Tab. 1" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 445, + 290, + 559 + ], + "lines": [ + { + "bbox": [ + 48, + 445, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 290, + 559 + ], + "type": "table", + "html": "
ModelReferenceSSv2-Small
1-shot5-shot
OTAM* [3]CVPR 202044.458.2
TRX* [20]CVPR 202142.059.1
STRM* [23]CVPR 202242.560.4
MoLo* [25]CVPR 202345.158.8
TATs [13]ECCV 202447.964.4
TATs w/o PT [13]ECCV 202444.662.5
TEAMCVPR 202547.263.1
", + "image_path": "7203ae29565d2ffa9b4931c0933513f12b24ff1c81b4956d6bfc0435100367c5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 575, + 288, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 288, + 660 + ], + "type": "text", + "content": "train all models on 5-way 5-shot tasks using the Kinetics dataset with the ViT backbone and evaluate them on the same setting across different target datasets. As shown in Tab. 3, TEAM consistently outperforms prior methods, indicating strong generalization beyond the training domain. This demonstrates TEAM's robustness to temporal variations with pattern tokens capturing globally distinct patterns." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 141, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 141, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 141, + 683 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Unless specifically stated, experiments are conducted in the 5-way 1-shot scenario using the ResNet-50." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 456, + 545, + 533 + ], + "blocks": [ + { + "bbox": [ + 305, + 409, + 547, + 453 + ], + "lines": [ + { + "bbox": [ + 305, + 409, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 547, + 453 + ], + "type": "text", + "content": "Table 3. Performance comparison across cross-domain scenarios. All models are trained on the 5-way 5-shot scenario using the Kinetics dataset with the ViT backbone, and evaluated on the same task across various datasets. Symbols match those in Tab. 1." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 456, + 545, + 533 + ], + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 533 + ], + "type": "table", + "html": "
ModelReferenceHMDB51UCF101
OTAM* [3]CVPR 202077.296.5
TRX* [20]CVPR 202178.197.9
STRM* [23]CVPR 202278.598.0
MoLo* [25]CVPR 202377.797.4
TEAMCVPR 202581.198.3
", + "image_path": "cb5f2544d10ec666e6bb351f0565b675b14237b78104dc02c660e6ed19716c5c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Component Ablation. Tab. 4 presents the performance when our component is applied sequentially. (a) shows that the use of instance pattern tokens solely achieves decent performances over the methods that rely on pre-defined temporal units and brute-force alignment processes. This validates the advantage of our core intuition that temporal alignment-free matching is beneficial for its flexibility in video understanding. Succeedingly, in (b), the exclusive tokens are shown to yield additional improvements, indicating that the different strategy in learning pattern tokens complements the original approach. Conversely, we observe slight performance drops in (c) where we implement the adaptation process without the management in the adaptation magnitude. This implies the importance of managing the degree" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5417" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 120, + 288, + 186 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "type": "text", + "content": "Table 4. Ablation study for each component. " + }, + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "type": "inline_equation", + "content": "P^{+}" + }, + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "type": "inline_equation", + "content": "P^{-}" + }, + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "type": "text", + "content": " denote the instance, exclusive, adaptive pattern tokens, respectively, while " + }, + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 71, + 288, + 117 + ], + "type": "text", + "content": " means entanglement in Eq. 10 and 13. H, K, and U also refer to HMDB51, Kinetics, and UCF101, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 120, + 288, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 120, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 120, + 288, + 186 + ], + "type": "table", + "html": "
P+P-P̂EHKU
(a)✓61.874.686.7
(b)✓✓62.575.086.8
(c)✓✓✓62.274.887.1
(d)✓✓✓✓62.875.187.2
", + "image_path": "18105e7ec332f817fe42a711e365acc01e72ac1357b4f757fc9036202678b0de.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 198, + 288, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 288, + 224 + ], + "type": "text", + "content": "of adaptation with the consideration of given class compositions. The performance gains in (d) validate our claim." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 289, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 289, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 289, + 333 + ], + "type": "text", + "content": "Number of pattern tokens. In Fig. 3, we present an ablation study on the effect of the number of pattern tokens in TEAM. The results show that TEAM remains robust across a wide range of token counts, except when the value deviates significantly from the optimal (60 for all datasets). These consistent trends across all datasets suggest that our method is not sensitive to the number of pattern tokens, demonstrating its strong robustness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 347, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 288, + 456 + ], + "type": "text", + "content": "Varying N and K for N-way K-shot. Beyond standard FSAR settings, we further validate TEAM's effectiveness under varying N-way K-shot configurations on Kinetics, following prior work [25]. Specifically, we vary the number of classes " + }, + { + "bbox": [ + 46, + 347, + 288, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 347, + 288, + 456 + ], + "type": "text", + "content": " from 5 to 10, and the number of video instances per class " + }, + { + "bbox": [ + 46, + 347, + 288, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 347, + 288, + 456 + ], + "type": "text", + "content": " from 1 to 5. As shown in Fig. 4, TEAM consistently outperforms prior approaches across all configurations. These results confirm that TEAM is robust to different FSAR settings, with consistent benefits across diverse scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 462, + 111, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 111, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 111, + 475 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 480, + 288, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 288, + 660 + ], + "type": "text", + "content": "What do pattern tokens encode? Pattern tokens are leveraged to incorporate video features and have been verified to be superior. To assess the video understanding capability of pattern tokens, we examine whether pattern tokens focus on the same context for the same category. In Fig. 5, we visualize the attention weights of a pattern token that effectively represents specific classes. Pattern tokens are considered effective in representing a class when the output pattern tokens for instances of the same class exhibit the highest similarity among all pattern tokens. As observed, pattern tokens consistently focus on the context even when the same action occurs in different frames, at varying speeds, and over different lengths. This implies that pattern tokens address the limitations of frame- and tuple-level alignment approaches, where actions can only be compared within pre-defined units." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "Necessity for multiple pattern tokens. We adopt multiple pattern tokens to capture diverse discriminative patterns across base categories, using these patterns as a foun" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 68, + 547, + 192 + ], + "blocks": [ + { + "bbox": [ + 306, + 68, + 547, + 192 + ], + "lines": [ + { + "bbox": [ + 306, + 68, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 68, + 547, + 192 + ], + "type": "image", + "image_path": "17445a7cc641c8c3ef7b63935122016144885a84af7c0e94f4260d960b9153f2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 194, + 531, + 205 + ], + "lines": [ + { + "bbox": [ + 320, + 194, + 531, + 205 + ], + "spans": [ + { + "bbox": [ + 320, + 194, + 531, + 205 + ], + "type": "text", + "content": "Figure 3. Ablation study for the number of pattern tokens." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 207, + 545, + 324 + ], + "blocks": [ + { + "bbox": [ + 307, + 207, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 324 + ], + "type": "image", + "image_path": "77bd2e4943d76ef3f7ee738ea199ef6740348d37878c8f6d421406913c4c36b1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 328, + 515, + 339 + ], + "lines": [ + { + "bbox": [ + 335, + 328, + 515, + 339 + ], + "spans": [ + { + "bbox": [ + 335, + 328, + 515, + 339 + ], + "type": "text", + "content": "Figure 4. N-way 1-shot and 5-way K-shot results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 352, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 556 + ], + "type": "text", + "content": "dation for distinguishing novel classes. To verify that the learned patterns effectively differentiate novel classes, we compute the discriminative power of each instance pattern token within each category and rank the tokens accordingly. Discriminative power is defined as ' (intra similarity - inter similarity)'. Intra similarity measures the average similarity between instance-specific pattern tokens from all videos within the same category and their prototype (average), while inter similarity is the maximum similarity between the prototype of one class and the prototypes of other classes. The resulting heatmap of discriminative power is shown in Fig. 6, where distinct heatmap arrangements between classes can be observed. The variation in discriminative rank indicates that useful patterns differ by category, supporting the need for multiple learnable pattern tokens to capture a variety of discriminative patterns, forming a foundation for classifying novel categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Computational time for matching process. In this study, we compare the computational overhead with alignment-based approaches. Computational costs for varying numbers of frames are shown in Fig. 7. As discussed in Sec. 1, alignment-based strategies require quadratic computation time due to pairwise comparison between temporal units, whereas TEAM with token-wise comparison has linear time complexity—especially evident when the number of frames exceeds 32. Although this issue is less critical for FSAR datasets, which consist of trimmed, short-term videos, it becomes significant in real-world scenarios with untrimmed, long-term videos. In this regard, we argue that our proposed" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5418" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 299, + 235 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 299, + 235 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 299, + 235 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 299, + 235 + ], + "type": "image", + "image_path": "3ec7e23c6ca02ecc059306bba68c52749a567fe97e8b4434767d3843b8b3425c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 237, + 546, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 546, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 546, + 283 + ], + "type": "text", + "content": "Figure 5. Visualization of cross-attention weights during instance pattern token aggregation. The same pattern tokens are visualized for videos in each class. Green regions in each frame metaphorically represent the percentage of the frame involvement in the instance pattern token. (a) For 'Diving Cliff', the visualized token primarily responds to the moment when water splashes when people dive. (b) In 'Playing Trumpet', the token focuses on the scenes where people hold a trumpet to their mouths." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 234 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 234 + ], + "type": "image", + "image_path": "16a883bbc2f725f971bd8fb042d8a6dd9a84408e84ee9bf0d6b426c76be3cdd4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 47, + 293, + 287, + 339 + ], + "blocks": [ + { + "bbox": [ + 47, + 293, + 287, + 339 + ], + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 339 + ], + "type": "image", + "image_path": "0a14a4a9b2ccad5a158a7c16eb9f74f082d5efd1e8b4bb0393d2d16c0a1506fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 340, + 288, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 384 + ], + "type": "text", + "content": "Figure 6. Heatmap visualizing the class-wise discriminative power of each instance pattern token for novel classes within HMDB51. Differences in discrimination ranks across classes indicate that each class has unique discriminative patterns." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 390, + 287, + 506 + ], + "blocks": [ + { + "bbox": [ + 48, + 390, + 287, + 506 + ], + "lines": [ + { + "bbox": [ + 48, + 390, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 390, + 287, + 506 + ], + "type": "image", + "image_path": "18769f9cda33c05b725a1af7011a123aa25b263d719482aba7ae1eae3645f4ec.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 510, + 288, + 575 + ], + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 575 + ], + "type": "text", + "content": "Figure 7. Comparison of computational time (ms) for video matching with alignment-based strategies across varying frame counts. OTAM and TRX are used as representatives of the frame- and tuple-level alignment approaches, respectively. Times reflect only the video matching process and exclude feature extraction by the backbone network, which is consistent across all methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 591, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 288, + 616 + ], + "type": "text", + "content": "temporal alignment-free matching offers a compelling alternative to alignment-based approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "FLOPs and parameter. We analyze FLOPs and parameter counts of TEAM and other methods, as shown in Tab. 5. While TEAM achieves remarkable performance, it has the fewest parameters except for OTAM, which relies solely on backbone parameters. The higher parameter counts in other methods stem from their complex designs, including temporal transformers. These results highlight the effectiveness of" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 318, + 545, + 432 + ], + "blocks": [ + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "content": "Table 5. Comparison of FLOPs and parameter counts, which are measured in G (giga operations) and M (million), respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 318, + 545, + 432 + ], + "lines": [ + { + "bbox": [ + 307, + 318, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 545, + 432 + ], + "type": "table", + "html": "
ModelResNetViT
FLOPsParameterFLOPsParameter
OTAM [3]19823.580985.8
TRX [20]20547.181088.8
STRM [23]20145.581088.8
SloshNet [27]594168.9--
MoLo [25]23789.583099.1
TATs [13]--117897.7
TEAM22432.081387.0
", + "image_path": "3f2d2d337167513720f7504e3f67ca9800fb66bca8b160c4ace45a381fe4d5cf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 453, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 502 + ], + "type": "text", + "content": "pattern tokens for FSAR, providing a lightweight alternative to heavier models. Moreover, although TEAM has higher FLOPs than OTAM and TRX, its efficient parallel operations ensure low computational time, as shown in Fig. 7." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 522, + 378, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 378, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 378, + 534 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we proposed a novel TEmporal Alignment-free Matching (TEAM) for Few-Shot Action Recognition, that achieves both flexibility and efficiency by eliminating the need for pre-defined temporal units in action representation and brute-force alignment for the video comparison. Concretely, TEAM integrates each video with a fixed set of pattern tokens that encapsulate globally discriminative patterns within the video instance, regardless of action length and speed. Moreover, TEAM has inherent efficiency by computing token-wise similarity for matching video. Upon these pattern tokens, we further proposed an adaptation process for establishing clear boundaries between classes by removing common information. We validated the effectiveness of TEAM with extensive experiments." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5419" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 131 + ], + "type": "text", + "content": "Acknowledgements This work was supported in part by MSIT/IITP (No. 2022-0-00680, 2020-0-01821, RS-2019-II190421, RS-2024-00459618, RS-2024-00360227, RS-2024-00437102, RS-2024-00437633), and MSIT/NRF (No. RS-2024-00357729)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 144, + 105, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 144, + 105, + 156 + ], + "spans": [ + { + "bbox": [ + 48, + 144, + 105, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 163, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 288, + 228 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 230, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 288, + 275 + ], + "type": "text", + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 275, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 275, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 275, + 288, + 330 + ], + "type": "text", + "content": "[3] Kaidi Cao, Jingwei Ji, Zhangjie Cao, Chien-Yi Chang, and Juan Carlos Niebles. Few-shot video classification via temporal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10618-10627, 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 331, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 288, + 386 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 388, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 432 + ], + "type": "text", + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 434, + 288, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 288, + 466 + ], + "type": "text", + "content": "[6] Anoop Cherian and Stephen Gould. Second-order temporal pooling for action recognition. International Journal of Computer Vision, 127:340-362, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 467, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 467, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 467, + 288, + 555 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 555, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 288, + 601 + ], + "type": "text", + "content": "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 645 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 689 + ], + "type": "text", + "content": "[10] Yifei Huang, Lijin Yang, and Yoichi Sato. Compound prototype matching for few-shot action recognition. In European Conference on Computer Vision, pages 351-368. Springer, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[11] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "text", + "content": "perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 547, + 150 + ], + "type": "text", + "content": "[12] Hirokatsu Kataoka, Tenga Wakamiya, Kensho Hara, and Yutaka Satoh. Would mega-scale datasets further enhance spatiotemporal 3d cnns? arXiv preprint arXiv:2004.04968, 2020.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 207 + ], + "type": "text", + "content": "[13] Pulkit Kumar, Namitha Padmanabhan, Luke Luo, Sai Saketh Rambhatla, and Abhinav Shrivastava. Trajectory-aligned space-time tokens for few-shot action recognition. In European Conference on Computer Vision, pages 474-493. Springer, 2025. 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 547, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 547, + 253 + ], + "type": "text", + "content": "[14] SuBeen Lee, WonJun Moon, and Jae-Pil Heo. Task discrepancy maximization for fine-grained few-shot classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5331-5340, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 547, + 308 + ], + "type": "text", + "content": "[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 309, + 547, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 547, + 365 + ], + "type": "text", + "content": "[16] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4100–4110, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 366, + 547, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 547, + 421 + ], + "type": "text", + "content": "[17] Shuyuan Li, Huabin Liu, Rui Qian, Yuxi Li, John See, Mengjuan Fei, Xiaoyuan Yu, and Weiyao Lin. Ta2n: Two-stage action alignment network for few-shot action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1404-1411, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 547, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 547, + 477 + ], + "type": "text", + "content": "[18] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 478, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 547, + 533 + ], + "type": "text", + "content": "[19] Khoi D Nguyen, Quoc-Huy Tran, Khoi Nguyen, Binh-Son Hua, and Rang Nguyen. Inductive and transductive few-shot video classification via appearance and temporal alignments. In European Conference on Computer Vision, pages 471-487. Springer, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 534, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 547, + 590 + ], + "type": "text", + "content": "[20] Toby Perrett, Alessandro Masullo, Tilo Burghardt, Majid Mirmehdi, and Dima Damen. Temporal-relational crosstransformers for few-shot action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 475-484, 2021. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 547, + 645 + ], + "type": "text", + "content": "[21] Sarinda Samarasinghe, Mamshad Nayeem Rizve, Navid Kardan, and Mubarak Shah. Cdfsl-v: Cross-domain few-shot learning for videos. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11643-11652, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 646, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 547, + 680 + ], + "type": "text", + "content": "[22] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "text", + "content": "[23] Anirudh Thatipelli, Sanath Narayan, Salman Khan, Rao Muhammad Anwer, Fahad Shahbaz Khan, and Bernard Ghanem. Spatio-temporal relation modeling for few-shot" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5420" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 563 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19958-19967, 2022. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 171 + ], + "type": "text", + "content": "[24] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Mingqian Tang, Zhengrong Zuo, Changxin Gao, Rong Jin, and Nong Sang. Hybrid relation guided set matching for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19948-19957, 2022. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 239 + ], + "type": "text", + "content": "[25] Xiang Wang, Shiwei Zhang, Zhiwu Qing, Changxin Gao, Yingya Zhang, Deli Zhao, and Nong Sang. Molo: Motion-augmented long-short contrastive learning for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18011-18021, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 288, + 295 + ], + "type": "text", + "content": "[26] Jiamin Wu, Tianzhu Zhang, Zhe Zhang, Feng Wu, and Yongdong Zhang. Motion-modulated temporal fragment alignment network for few-shot action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9151-9160, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 288, + 341 + ], + "type": "text", + "content": "[27] Jiazheng Xing, Mengmeng Wang, Yong Liu, and Boyu Mu. Revisiting the spatial and temporal modeling for few-shot action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3001-3009, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 341, + 288, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 341, + 288, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 341, + 288, + 406 + ], + "type": "text", + "content": "[28] Jiazheng Xing, Mengmeng Wang, Yudi Ruan, Bofan Chen, Yaowei Guo, Boyu Mu, Guang Dai, Jingdong Wang, and Yong Liu. Boosting few-shot action recognition with graph-guided hybrid matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1740-1750, 2023. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 407, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 407, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 288, + 474 + ], + "type": "text", + "content": "[29] Hongguang Zhang, Li Zhang, Xiaojuan Qi, Hongdong Li, Philip HS Torr, and Piotr Koniusz. Few-shot action recognition with permutation-invariant attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 525-542. Springer, 2020. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 475, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 518 + ], + "type": "text", + "content": "[30] Sipeng Zheng, Shizhe Chen, and Qin Jin. Few-shot action recognition with hierarchical matching and contrastive learning. In European Conference on Computer Vision, pages 297-313. Springer, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 519, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 288, + 563 + ], + "type": "text", + "content": "[31] Linchao Zhu and Yi Yang. Compound memory networks for few-shot video classification. In Proceedings of the European Conference on Computer Vision (ECCV), pages 751-766, 2018. 1, 6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5421" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_content_list.json b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee2ed1769d0fa5f8ba57568d3fe291c89ecec617 --- /dev/null +++ b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_content_list.json @@ -0,0 +1,1457 @@ +[ + { + "type": "text", + "text": "Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts", + "text_level": 1, + "bbox": [ + 99, + 130, + 897, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Cao Zengqun Zhao Ioannis Patras Shaogang Gong Queen Mary University of London", + "bbox": [ + 233, + 180, + 763, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yu.cao, zengqun.zhao, i.patras, s.gong}@qmul.ac.uk", + "bbox": [ + 269, + 219, + 720, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59d94c7801df875beae214e5612d7d1693c73af4606abf6fd8e2c286d58898dd.jpg", + "image_caption": [ + "Figure 1. Why do diffusion models generate artifacts? We discover that a diffusion generative process necessarily undergoes three phases, we call them: (2) \"Profiling\" which recovers holistic mean templates, (2) \"Mutation\" which introduces local divergence, and (3) \"Refinement\" which rationalizes pixel-wise generation in spatial context. Four visual examples are shown: The first two rows are two examples of rational local mutations (in green boxes) either naturally integrated (Row 1) or reasonably eliminated (Row 2). The bottom two rows show two failure cases when mutations were trapped unreasonably (in red boxes), resisting refinement and resulting in artifacts. Phases are visualized in equal intervals for clarity; please zoom in for more details." + ], + "image_footnote": [], + "bbox": [ + 91, + 260, + 906, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 641, + 326, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual artifacts remain a persistent challenge in diffusion models, even with training on massive datasets. Current solutions primarily rely on supervised detectors, yet lack understanding of why these artifacts occur in the first place. In our analysis, we identify three distinct phases in the diffusion generative process: Profiling, Mutation, and Refinement. Artifacts typically emerge during the Mutation phase, where certain regions exhibit anomalous score dynamics over time, causing abrupt disruptions in the normal evolution pattern. This temporal nature explains why existing methods focusing only on spatial uncertainty of the final output fail at effective artifact localization. Based on these insights, we propose ASCED (Abnormal Score Correction for Enhancing Diffusion), that detects artifacts by monitoring abnormal score dynamics during the diffusion process,", + "bbox": [ + 86, + 672, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with a trajectory-aware on-the-fly mitigation strategy that appropriate generation of noise in the detected areas. Unlike most existing methods that apply post hoc corrections, e.g., by applying a noising-denoising scheme after generation, our mitigation strategy operates seamlessly within the existing diffusion process. Extensive experiments demonstrate that our proposed approach effectively reduces artifacts across diverse domains, matching or surpassing existing supervised methods without additional training. Project page: YuCao16.github.io/ASCED.", + "bbox": [ + 511, + 642, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 799, + 645, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have emerged as powerful foundation models in computer vision [5], achieving remarkable success in image generation [7, 15-17, 32], image inpainting [24, 27, 35], and text-to-image task [30-32]. However, even trained on large-scale datasets, diffusion generative images", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7707", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/955a49b3382a0aa4220b5107495c061ef7d2f48a58147b8cd50222d307e0a843.jpg", + "image_caption": [ + "Figure 2. Diagram of our framework. Denoising and Noising are using Eq. (5) and Eq. (1), respectively." + ], + "image_footnote": [], + "bbox": [ + 143, + 78, + 851, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "still exhibit two significant flaws: visual artifacts and hallucinations [43, 47]. Visual artifacts appear as local irregularities in texture or structure, while hallucinations involve semantically incoherent content, e.g., extra limbs or misplaced objects. In this work, we focus on addressing diffusion artifacts, which present a fundamental challenge to achieving reliable and high-quality image generation.", + "bbox": [ + 88, + 295, + 482, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods primarily treat visual artifact detection as a classification problem, i.e., identifying problematic generations for filtering or reconstruction. These methods typically rely on a specialized classifier, either trained on manually annotated artifact datasets [43] or leveraging a pre-trained Large Multi-Modal Model (LMM) [22]. However, such post hoc interventions fail to address a fundamental problem: Why and when do artifacts emerge in diffusion models? To bridge this gap, we begin by examining the diffusion generation process itself.", + "bbox": [ + 88, + 402, + 482, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We discover that while diffusion process is guided by the same fundamental equation across time (i.e. diffusion steps), in practice, the model exhibits different behavior that can be roughly categorized in three different temporal phases that we name Profiling, Mutation, and Refinement. In the \"Profiling\" phase, the model sketches the basic semantic global layout; in the \"Mutation\" phase it explores potential local pixel-wise variations to create local structure; in the \"Refinement\" phase it attempts to resolve these local pixel-wise variations into coherent visual details in context (see Fig. 1 for visual examples and Sec. 3.2 for a detailed analysis). This understanding of the generation process reveals that while visual artifacts may appear randomly, they follow systematic and identifiable temporal patterns during image formation. Recent uncertainty-based approaches try to identify artifacts by converting diffusion models into Bayesian networks and employ techniques such as Last Layer Laplace Approximation [9] to generate pixel-level variance matrices [20]. However, these uncertainty quantification analyses only capture spatial variations in the final output, i.e., $Var(x_0)$ , missing crucial temporal dynamics during the generation process. Our study shows that diffusion artifacts emerge when certain image regions ex", + "bbox": [ + 93, + 553, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hibit abnormal evolution patterns over time, primarily during the Mutation phase of the generation process. Specifically, their pixel values stop updating properly while the surrounding areas continue to evolve. This phenomenon, which we formally define as \"score traps\" in Sec. 3.2, explains why examining only the final output is insufficient and misplaced for effective diffusion artifact detection.", + "bbox": [ + 511, + 296, + 903, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these insights, we propose ASCED: Abnormal Score Correction for Enhancing Diffusion, as shown in Fig. 2. At the heart of the method is the estimation of a score, in this context the direction and magnitude of pixel-wise evolution at each diffusion generative step, and a scheme that analyses its temporal dynamics and detects abnormalities. We show both theoretically and experimentally that these abnormalities strongly correlate with artifact formation, making early detection possible at a stage where intervention is still feasible, before artifacts become irreversibly embedded in the generation process. We leverage this early detection capability by implementing a novel trajectory-aware correction mechanism that disrupts the evolution of artifact regions while preserving overall generation diversity. Importantly, ASCED operates in a fully unsupervised manner without requiring manual annotations or domain-specific training, making it readily applicable across various domains, , particularly valuable when training data may be limited or protected.", + "bbox": [ + 511, + 402, + 903, + 690 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are: (1) We provide new insights into the formation of visual artifacts in the diffusion generative process, advancing the understanding of diffusion model internal mechanisms. (2) We introduce a novel method that detects potential artifact regions by monitoring abnormal score dynamics temporally, without any manually labeled training required. (3) We further develop a on-the-fly trajectory-aware correction mechanism that effectively mitigates artifacts while preserving image diversity.", + "bbox": [ + 511, + 691, + 903, + 827 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 844, + 661, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Artifact Detection initially targeted superresolution artifacts, where upsampling operations are the", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7708", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "main source [45]. These methods analyze either spatial domain characteristics to capture texture differences between real and generated images [23, 41], or frequency domain patterns to study artifact characteristics in high-frequency components [12, 13]. More recent work has shifted focus to detecting artifacts in general image generation, developing specialized classifiers trained on manually annotated datasets [43] or utilizing pre-trained large vision models [22]. However, these supervised approaches require extensive manual labeling and may not generalize well across different domains. A parallel direction explores uncertainty quantification methods to understand visual artifacts. While various approaches including variational inference [4, 18], Laplace approximation [26, 29], and Markov Chain Monte Carlo [38, 44] have been developed, their application to diffusion models remains limited [20]. BayesDiff [20] pioneers pixel-level uncertainty quantification in diffusion models using Last-layer Laplace approximation [9], yet the connection between spatial uncertainty and visual artifacts remains unclear.", + "bbox": [ + 89, + 90, + 483, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generation Quality Enhancement Various approaches have been proposed to enhance generation quality. The truncation trick in BigGAN [6] demonstrated that restricting the sampling space can significantly improve generation fidelity, suggesting similar principles might apply to diffusion models. For diffusion-based generation, classifier guidance [11] has been introduced to steer the generation process. SARGD [49] extends this idea by utilizing a pre-trained artifact detector to guide the generation towards artifact-free regions. Latent diffusion models [30] take a different approach by applying diffusion in a learned latent space, demonstrating that controlled evolution in a constrained space can lead to higher-quality output.", + "bbox": [ + 89, + 393, + 483, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Model for Representation Learning has evolved toward latent space disentanglement and controllable editing. The former aims to uncover interpretable factors in the generative process. Recent studies [28, 42, 48] observed stage-wise attribute emergence during generation, but focus differently than our analysis of artifact formation mechanisms. Controllable editing techniques [27, 31, 37, 46] can be applied for artifact removal, yet require per-sample manipulation and address symptoms rather than causes. Our approach instead corrects abnormal score dynamics during the generation process itself.", + "bbox": [ + 89, + 589, + 483, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 768, + 225, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach consists two steps: Detection (Sec. 3.2) and Correction (Sec. 3.3). Specifically, we localize artifact pixel regions by identifying abnormal score dynamics during the diffusion inference process (Detection) and develop a novel artifact correction algorithm without delaying inference (Correction). We give a theoretical analysis of our key concepts on score trap and temporal weighting (Sec. 3.4).", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2578a3144b316f48b968042ff566a11a6bb441a980ef785ba77daef9fcff41f5.jpg", + "image_caption": [ + "Figure 3. Artifact generation through denoising (top) and brush stroke noising via SDEdit [27] (bottom), demonstrating the model's inability to distinguish artifacts during generation." + ], + "image_footnote": [], + "bbox": [ + 522, + 93, + 890, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 287, + 653, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Model Let $x_0 \\in \\mathbb{R}^{c \\times h \\times w}$ be an image. The forward process of a diffusion model gradually diffuses the data distribution $q(x)$ towards $q_t(x_t)$ , $\\forall t \\in [0, T]$ , with $q_T(x_T) = \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ as a trivial Gaussian distribution. From a score viewpoint, it can be described by the Stochastic Differential Equation (SDE):", + "bbox": [ + 511, + 309, + 903, + 400 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd \\mathbf {x} _ {t} = \\boldsymbol {f} \\left(\\mathbf {x} _ {t}, t\\right) d t + \\boldsymbol {g} \\left(\\mathbf {x} _ {t}, t\\right) d \\mathbf {w} _ {t}, \\quad t \\in [ 0, T ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 402, + 903, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{w}$ is the standard Wiener process, $f(\\cdot)$ and $g(\\cdot)$ are scalar drift and diffusion coefficients, respectively. Anderson [2] states that the reverse process of Eq. (1) is also a diffusion process as:", + "bbox": [ + 511, + 419, + 903, + 478 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd \\mathbf {x} _ {s} = \\left[ \\boldsymbol {f} \\left(\\mathbf {x} _ {s}, s\\right) - \\boldsymbol {g} \\left(x _ {s}, s\\right) ^ {2} \\boldsymbol {s} \\left(x _ {s}, s\\right) \\right] d s + \\boldsymbol {g} \\left(x _ {s}, s\\right) d \\mathbf {w} _ {s} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 479, + 903, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_s\\coloneqq x_{T - t}$ and $s(x_s,s)\\coloneqq \\nabla_{\\mathbf{x}_s}\\log p_s(\\mathbf{x}_s)$ is the score function of the marginal distribution over $x_{s}$ . Song et al. [35] leverage this property to generate samples by first drawing $x_{T}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and then solving the reverse SDE using a learned score network $s_\\theta (x_t,t)$ .", + "bbox": [ + 511, + 511, + 903, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By using Tweddie's formula [36], DDPM (Denoising Diffusion Probabilistic Model) [16] can be shown as an equivalent interpretation of Eq. (2) [25]:", + "bbox": [ + 511, + 585, + 903, + 631 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t}, t\\right) = - \\frac {1}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t}, t\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 631, + 903, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s$ , with mean coefficient $\\alpha_t$ , and $\\epsilon_\\theta(x_t, t)$ is the noise network of DDPM.", + "bbox": [ + 511, + 659, + 903, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition of Visual Artifacts Generation flaws can be distinguished into two categories: Visual Artifacts and Hallucinations. Visual artifacts manifest as local irregularities or distortions in a generated image, such as blurred patches, unnatural textures, broken structures. In contrast, hallucinations refer to semantically generating incoherent content, such as extra limbs, misplaced objects or counterfactuals. In this paper, we focus specifically on detecting and correcting visual artifacts generated by diffusion models.", + "bbox": [ + 511, + 703, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Detection by Anomalous Score Dynamic", + "text_level": 1, + "bbox": [ + 511, + 847, + 857, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To understand how visual artifacts emerge during generations, we examine diffusion model behavior through the", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7709", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7d28fff023b602ab945dcd1c2ce5a123cfb2890e63d8a1b965e8dc8814f48185.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 380, + 358 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3328b4909277dcf45ed4a9c602152700358119f4f97d1a6ccd5acb24e3637071.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 390, + 90, + 586, + 358 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/10e0ac09e55719c100680a96a696615a53cd25df56e7a485ee27e3a61bbef44e.jpg", + "image_caption": [ + "(c)", + "Figure 4. Visualization of score dynamics and visual artifact detection. (a) Generated images with detected visual artifact regions highlighted (red). (b) Visualization of score dynamics (normalized) between adjacent time steps as activation maps. Brighter regions (green to yellow) indicate areas of higher score variation, while darker regions (blue to black) show areas of lower score change. (c). Score acceleration curves (representing the rate of change in score dynamics between consecutive timesteps) comparing artifact regions (red) with non-artifact regions (blue). The artifact regions exhibit characteristic rapid acceleration followed by deceleration, while non-artifact regions maintain stable score dynamics over time throughout a generative (inference) process." + ], + "image_footnote": [], + "bbox": [ + 596, + 90, + 898, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lens of image editing. SDEdit [27] demonstrates that diffusion models can transform irregularities, e.g., brush strokes, into semantically meaningful content through a noise-then-denoise process, revealing the inherent Refinement capability of diffusion models. We observe that such irregularities after noising become structurally indistinguishable from states containing artifacts during generation, as shown in Fig. 3. While diffusion models can successfully refine noised brush strokes, they fail to correct the corresponding artifacts during generation. This contrast reveals that diffusion models lack the ability to identify artifacts as patterns requiring refinement during the generation process.", + "bbox": [ + 88, + 470, + 482, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better understand this, we examine the generation process from a score perspective, as it directly represents the evolution of pixel values [35]. We define score dynamics as the difference between temporally adjacent score values: $\\Delta s_{\\theta}(x_t^{i,j},t) = s_{\\theta}(x_t^{i,j},t) - s_{\\theta}(x_{t - 1}^{i,j},t - 1)$ . Analysis reveals that image generation begins with establishing basic structures, followed by a phase of stochastic exploration where irregular patterns may emerge; we call these phases Profiling and Mutation, respectively. As shown in Fig. 4, regions containing visual artifacts exhibit characteristic patterns during mutation: They appear as localized regions of intense score variations (Fig. 4 (b)) and display dramatic acceleration followed by sudden deceleration in their score trajectories (Fig. 4 (c)). In contrast, normal regions maintain a stable evolution throughout generation.", + "bbox": [ + 89, + 656, + 482, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on these observations, we propose a novel ap", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "proach to detecting and localizing potential artifact regions over time in a diffusion inference process. Specifically, let $\\Omega \\subset \\mathbb{R}^2$ denote the spatial domain of the image, and $\\Omega_t^a \\subset \\Omega$ represent regions where abnormal evolution patterns emerge at timestep $t$ . For each spatial location $(i,j) \\in \\Omega$ , we track the score dynamics through consecutive timesteps. To account for the varying score magnitudes across different images and timesteps, we maintain a score bank $\\mathcal{S} = \\{s_\\theta(x_k,k)\\}_{k=t}^T$ and apply a temporal weighting function $w(t)$ that addresses the inherent decay of score magnitudes. Formally, we define artifact regions as:", + "bbox": [ + 511, + 472, + 906, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {t} ^ {a} := \\left\\{(i, j) \\in \\Omega \\mid \\left| \\Delta \\left(w (t) \\cdot s _ {\\theta} \\left(x _ {t} ^ {i, j}, t\\right)\\right) \\right| > \\tau \\right\\} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 640, + 906, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w(t) = \\frac{1 - \\bar{\\alpha}_t}{\\sqrt{\\bar{\\alpha}_t}}$ (see Sec. 3.4 for theoretical analysis) and $\\tau$ is adaptively determined as the maximum between the Median Absolute Deviation (MAD) of the weighted score dynamics and the mean of score bank $S$ . The final artifact regions $\\Omega^a$ are accumulated across the score bank, with detailed procedures provided in Algorithm 1.", + "bbox": [ + 511, + 669, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Real-Time Correction", + "text_level": 1, + "bbox": [ + 511, + 771, + 718, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After detecting artifacts, we need a correction strategy to effectively refine the artifact region. Two natural approaches emerge: either to revert to the states before the artifacts emerge through state replacement or to limit directly abnormal score changes through score clipping. For state replacement, we first predict the clean image from an earlier timestep $t$ using [16]:", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7710", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Pseudo Code for Proposed ASCED Method" + ], + "code_body": "1: Input: Score network $s_{\\theta}(\\cdot)$ which requires $T$ steps to generate, detection starting step $T_d$ and correction step $T_c$ \n2: Initialize $x_T \\sim \\mathcal{N}(0, \\mathbf{I})$ , Score Bank $S \\gets \\{\\}$ , Visual Artifact Mask $\\Omega^a \\gets \\{\\}$ \n3: for $t = T, t$ -, while $t >= 0$ do\n4: $x_{t-1} = \\sqrt{\\alpha_{t-1}} x_0 - \\sqrt{1 - \\alpha_{t-1}} \\sqrt{1 - \\bar{\\alpha}_t} s_\\theta(x_t, t)$ , where $x_0 = \\frac{x_t + (1 - \\bar{\\alpha}_t) s_\\theta(x_t, t)}{\\sqrt{\\bar{\\alpha}_t}}$ ▷ Original Diffusion Process\n5: if $T_c < t <= T_d$ then\n6: S.append $s_\\theta(x_t, t)$ ▷ Store score value into Score Bank\n7: else if $t == T_c$ then ▷ Anomalous Score Dynamics Detection Step\n8: for $k = 0, k++$ , while $k < (T_c - T_d)$ do ▷ Determine $\\Omega^a$ by accumulation\n9: $\\Omega^a = \\Omega^a \\cup \\{(i, j) \\in \\Omega | |\\Delta(w(k) \\cdot s_\\theta(x_k^{i,j}, k))| > \\tau\\} ▷ \\tau = \\max\\{\\mathrm{MAD}(\\Delta(w(k) \\cdot s_\\theta(x_k^{i,j}, k)), \\mathrm{mean}(\\mathcal{S})\\}$ \n10: $x_t = x_t \\cdot \\mathbb{1}_{\\overline{\\Omega}^a} + (\\sqrt{\\bar{\\alpha}_t} \\hat{x}_0(t) + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon) \\cdot \\gamma(t) \\xi_{\\Omega^a}$ ▷ Trajectory-aware Targeted Correction Step\n11: return $x_0$", + "bbox": [ + 91, + 108, + 906, + 296 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {0} \\left(\\boldsymbol {x} _ {t}, t\\right) = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {x} _ {t} + \\left(1 - \\bar {\\alpha} _ {t}\\right) \\nabla_ {\\theta} \\log p \\left(\\boldsymbol {x} _ {t}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 321, + 480, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then replacing the artifact regions with corresponding states from this predicted clean image after re-noising to the current timestep. Score clipping directly limits the magnitude of score changes during inference. However, both state replacement and score clipping fundamentally disrupt the mutation process, leading to reduced generation diversity. To address this problem, we propose Trajectory-aware Targeted Correction (TTC), which introduces controlled perturbations specifically in artifact regions at correction timestep $T_{c}$ :", + "bbox": [ + 89, + 351, + 483, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {T _ {c}} = \\boldsymbol {x} _ {T _ {c}} \\cdot \\mathbb {1} _ {\\overline {{\\Omega}} ^ {a}} + \\left(\\sqrt {\\bar {\\alpha} _ {T _ {c}}} \\boldsymbol {x} _ {0} ^ {\\prime} + \\sqrt {1 - \\bar {\\alpha} _ {T _ {c}}} \\epsilon\\right) \\cdot \\gamma \\xi \\mathbb {1} _ {\\Omega^ {a}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 507, + 482, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}_0^{\\prime} = \\hat{\\pmb{x}}_0(\\pmb{x}_{T_c},T_c)$ $\\epsilon ,\\xi \\stackrel {\\mathrm{iid}}{\\sim}\\mathcal{N}(0,\\mathbf{I}),\\mathbb{1}_{\\Omega^a}$ is indicator function for artifact regions and perturbation intensity $\\gamma$", + "bbox": [ + 89, + 534, + 482, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TTC builds upon the understanding of score traps: Regions where pixels become locked in persistent score patterns after experiencing dramatic score changes. Through controlled perturbations, TTC disrupts these fixed patterns and allows pixels to resume normal evolution with surrounding regions. Sec. 3.4 provides a detailed analysis of these score trap mechanisms and their relationship to visual artifacts. Generation quality and diversity comparison across correction strategies is shown in Fig. 5.", + "bbox": [ + 89, + 568, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 89, + 710, + 284, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide a theoretical analysis for the score trap and the choice of the temporal weighting function in Eq. (4).", + "bbox": [ + 89, + 734, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theoretical Analysis of Score Traps For normal generation, the score evolution of each pixel is coupled with its neighborhood through the learned neural score function [8]:", + "bbox": [ + 89, + 763, + 482, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\theta} \\left(x _ {t} ^ {i, j}, t\\right) = \\nabla_ {x _ {t} ^ {i, j}} \\log p _ {\\theta} \\left(x _ {t} ^ {i, j} \\mid \\mathcal {C} (i, j), t\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 815, + 482, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}(i,j)$ represents the contextual information from neighboring pixels. This coupling ensures coordinated evolution toward the data manifold. When regions experience abnormal score dynamics, they can enter score traps", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where local patterns persist despite significant score values, disrupting the natural coupled evolution process. These trapped regions evolve based primarily on their local patterns, losing the contextual relationships necessary for coherent image generation.", + "bbox": [ + 511, + 324, + 903, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This reveals how our perturbation-based correction re-establishes contextual relationships. For a trapped pixel $(i,j)$ , the perturbation $\\gamma \\cdot \\xi$ introduces stochastic variations that disrupt the isolated score patterns, creating opportunities for these regions to re-couple with their surroundings through the natural score evolution process. Meanwhile, in areas without abnormal patterns, these modest perturbations preserve the original coupled evolution, ensuring the method remains harmless to non-artifact regions; see Sec. 7.2 for further mathematical derivations and proofs.", + "bbox": [ + 511, + 400, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Score Normalization The score function learned by diffusion models can be interpreted as a vector field guiding the denoising trajectory, inducing a probability flow [34]:", + "bbox": [ + 511, + 565, + 905, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d \\boldsymbol {x}}{d t} = - \\frac {1}{2} \\sigma_ {t} ^ {2} \\nabla \\log p _ {t} (\\boldsymbol {x} _ {t}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 613, + 903, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The temporal evolution of this probability flow can be characterized by its divergence in the probability density field. Theoretically, we can model this through a flow operator $\\mathcal{G}$ :", + "bbox": [ + 511, + 643, + 903, + 689 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} (t) = \\nabla \\cdot \\left(\\frac {\\partial}{\\partial t} \\int_ {\\tau = 0} ^ {t} \\mathcal {P} (\\boldsymbol {x} _ {\\tau}, \\tau) d \\tau\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 693, + 903, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}(\\pmb{x}_{\\tau},\\tau)$ denotes the local probability density at position $\\pmb{x}_{\\tau}$ and time $\\tau$ . This formulation allows us to monitor the accumulation of probability density changes over time. We observe that the clean image prediction (Eq. (5)) reflects these changes in the probability flow. Under the assumption of smooth probability density evolution between adjacent time steps [19, 33, 34], score dynamics are captured through $\\frac{\\partial}{\\partial t}\\hat{\\pmb{x}}_0(t)$ . Consequently, a normalization factor $w(t) = \\frac{1 - \\alpha_t}{\\sqrt{\\alpha_t}}$ is derived from the coefficient of the score term in Eq. (5), helping to equalize the scale of score variations throughout the denoising process.", + "bbox": [ + 511, + 731, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7711", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/46b1769201298ba8042aa821e114420d710bc42fac64833f1881a4b56b462316.jpg", + "table_caption": [ + "Table 1. Quantitative Comparisons on five datasets. The methods compared include BayesDiff [20] and SARGD [49], and three baseline methods: State Replacement Score Clipping and PAL [43] + TTC. All methods use DDIM sampling with identical noise seeds to generate 10,000 images per dataset, ensuring each approach modifies the same deterministic trajectories for fair comparison. The best scores are in bold and second best in underline with bold. Sup and UnS denote supervised and unsupervised methods, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsTypeFFHQ[17]ImageNet[10]LSUN-Cat[40]LSUN-Horse[40]LSUN-Bedroom[40]
FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑
Original [34]UnS36.690.6290.49314.680.7390.73422.170.5130.58629.360.5100.64212.960.6270.583
State ReplaceUnS37.090.6350.49514.610.7430.73322.790.5100.58730.360.5020.64212.950.6280.574
Score ClippingUnS36.360.6300.49814.580.7420.73622.120.5150.58529.260.5110.64212.920.6270.585
BayesDiff [20]UnS36.990.6320.49114.530.7430.73022.500.5130.58528.700.5180.63412.880.6250.569
SARGD [49]Sup38.370.6370.46415.340.7310.72722.650.5230.57030.020.5100.62113.820.6390.554
PAL [43] + TTCSup36.350.6240.50014.010.7310.74721.830.5140.58828.680.5190.64612.710.6290.579
ASCED (Ours)UnS36.280.6370.50314.410.7500.73521.910.5150.59327.660.5210.65212.530.6280.590
", + "bbox": [ + 91, + 131, + 903, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 256, + 222, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Basic setups We conducted experiments on five datasets: FFHQ [17], ImageNet [10], LSUN-Bedroom [40], LSUN-Cat [40], and LSUN-Horse [40]. We employed the Guided Diffusion model framework and pre-trained weights from OpenAI [11] and Segmentation-DDPM [3]. Quantitative evaluations were performed using FID [14], which measures the Fréchet distance between real and generated image distributions, along with Precision and Recall [21], which evaluate sample fidelity and diversity, respectively.", + "bbox": [ + 88, + 282, + 480, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details For detecting diffusion artifacts, our approach was compared with LLaVA-v1.5-13B [22] and PAL [43], while artifact removal comparisons were made with BayesDiff [20] and the adapted SARGD [49]. All experiments were performed on NVIDIA A100 / H100 GPUs. We used DDIM [34] to improve inference efficiency with a Number of Function Evaluations (NFE) set to 25. Remark: We demonstrate in Sec. 7.1 that there is no significant correlation between NFE and the generation of artifacts. Full implementation details are provided in Sec. 6.", + "bbox": [ + 88, + 421, + 480, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative Comparisons to Existing Methods", + "text_level": 1, + "bbox": [ + 89, + 587, + 482, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate the effectiveness of ASCED in improving generative quality through comparisons with both unsupervised (BayesDiff [20]) and supervised (SARGD [49]) SOTA methods, along with the original diffusion model [11] and two baselines from Sec. 3.3 (state replacement and score clipping). To isolate the effectiveness of our correction method, we also evaluate a hybrid approach combining artifact detector PAL (used in SARGD) with our Trajectory-aware Targeted Correction (TTC). Quantitative results are shown in Tab. 1. Among unsupervised methods, our ASCED demonstrates superior performance across all datasets, consistently achieving better FID and Precision scores while maintaining higher Recall values than BayesDiff and baselines, indicating both improved generation quality and better preservation of diversity.", + "bbox": [ + 89, + 611, + 482, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to the supervised methods, our proposed ASCED method shows leading performance across most experiments, achieving superior results on FFHQ, LSUN-Horse, and LSUN-Bedroom, while maintaining competitive", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performance on ImageNet and LSUN-Cat. The better performances of PAL and SARGD on these two datasets are due to that they are supervised artifact detectors specifically trained on these datasets [43]. Their advantages are not generalizable to other datasets FFHQ, LSUN-(Horse, Bedroom). In contrast, ASCED as an unsupervised method has generalisable advantages in all domains without dataset specific training, making it more practical and scalable. Additionally, our method demonstrates significant computational efficiency advantages. In our experiments, ASCED detects and corrects artifacts in approximately 0.09 s per image, which is $8.8 \\times$ faster than PAL (0.79 s).", + "bbox": [ + 511, + 258, + 903, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The effectiveness of the correction mechanism (TTC) becomes particularly evident when comparing SARGD with PAL (used in SARGD) + TTC, where the trajectory-aware correction demonstrates significant advantages in preserving generation diversity across all datasets, reflected in consistently higher Recall scores.", + "bbox": [ + 511, + 439, + 903, + 530 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b446e5a34c0dabd9abce04091b2d2ccae0be0c631036ce20c4d4cd4e8f9026cc.jpg", + "table_caption": [ + "Table 2. Visual Artifact Detection Accuracy Comparison between PAL (Supervised, Sup) [43], LLaVA (Zero-Shot, ZS) [22], and our method (Unsupervised, UnS) on FFHQ [17], ImageNet [10], and LSUN-(Bedroom, Cat, Horse) [40]." + ], + "table_footnote": [], + "table_body": "
MethodTypeFFHQImageNetBedroomCatHorse
PALSup51.4%69.2%52.4%69.8%60.9%
LLaVAZS63.1%91.1%75.9%59.5%72.2%
OursUnS56.7% (-6.4)67.7% (-1.5)65.0% (-10.9)68.3% (-1.5)70.3% (-1.9)
", + "bbox": [ + 514, + 594, + 903, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Artifact Detection Performance Analysis", + "text_level": 1, + "bbox": [ + 511, + 650, + 859, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the accuracy of our method in identifying visual artifacts, we manually selected 200 images for each dataset from the diffusion model outputs, consisting of 100 images with visual artifacts and 100 without. We then evaluated ASCED against zero-shot large multi-modal model LLaVAv1.5-13B [22] and supervised artifact detector PAL [43]. For LLaVA evaluation, we produced 50 different prompts and reported the results for the most effective prompt. Details on prompt generation are provided in Sec. 6.2. The accuracy for both methods is presented in Tab. 2.", + "bbox": [ + 511, + 672, + 903, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As an unsupervised method, ASCED achieves promising detection performance, maintaining close accuracy to supervised approaches LLaVA and PAL across most datasets. Notably, these methods analyze final generated images, whereas our approach detects artifacts during the generation", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7712", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/845a605e1bd99d4147abfef32cc75ba3399655b5baf545499db7ae9a946fc26d.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison of different correction methods. For each example, we show the original output with visual artifacts (left) and zoomed-in views of the artifact regions corrected by different methods (right): SARGD [49], state replacement (Replace), and our trajectory-aware targeted correction (Ours). Rows from top to bottom: FFHQ[17], ImageNet[10], and LSUN-(Cat, Horse, Bedroom)[40]." + ], + "image_footnote": [], + "bbox": [ + 91, + 75, + 906, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "process through score dynamics, enabling early intervention before artifacts fully manifest. However, our method does show limitations in specific cases, such as low-contrast images where subtle abnormalities are difficult to distinguish from normal variations (leading to False Negatives), and instances where the diffusion model successfully rationalizes initially abnormal patterns during refinement (causing False Positives). Representative examples of these cases are illustrated in Fig. 6.", + "bbox": [ + 89, + 425, + 483, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Analysis of Correction Methods", + "text_level": 1, + "bbox": [ + 89, + 573, + 460, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better illustrate the advantage of Trajectory-aware Targeted Correction (TTC) over baselines, Fig. 5 shows qualitative comparisons consisting of the original outputs with artifacts, results from state replacement (Sec. 3.3), SARGD [49], and TTC (ours). While all methods can remove artifacts, TTC demonstrates superior detail preservation in corrected regions. Specifically, both state replacement and SARGD tend to converge to similar expressions and local details, constraining natural variations, as they directly modify the generation states. SARGD faces further limitations from its artifact detector being trained on a specific domain [43], affecting its generalization ability. More importantly, by preserving mutation phase operations, our trajectory-aware correction enables diverse yet coherent generations even when correcting the same region. Additional comparison results are provided in Sec. 8.2.", + "bbox": [ + 89, + 597, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Do corrections at Non-Artifact Regions harm? As with any detection method, our proposed scheme will inevitably encounter false positives, leading to add perturbations in non-artifact areas. Our experiments demonstrate that ap", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/770b7594e2ebd0a0c051695aa9c521b2b63ac8d0a9c255f1fd8cca56c7efd0c1.jpg", + "image_caption": [ + "Figure 6. Top: Applied our correction method to clean regions (yellow box). Bottom: Typical failure cases." + ], + "image_footnote": [], + "bbox": [ + 514, + 422, + 903, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plying the correction mechanism to regions without visual artifacts, as shown in Fig. 6, introduces modest variations while preserving semantic coherence with the surrounding context and not generating new artifacts. Extended visual results and detection performance analysis can be found in Sec. 7.2 and Sec. 4.2, respectively.", + "bbox": [ + 511, + 580, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 681, + 679, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distribution of Abnormal Score Dynamics In Fig. 7, we plot the frequency of abnormal scores at each time step (normalized by total diffusion time step $T$ ). The score dynamics demonstrate distinct patterns across different stages: remaining stable in early steps where basic structures emerge, experiencing significant variations in the middle stage, and gradually stabilizing again in the later steps during details' refinement. The presence of the long tail indicates that some variations persist into later steps, suggesting an extended period of details' adjustment. This behavioral pattern naturally aligns with our hypothesis that the generation undergoes three phases: profiling, mutation, and rationalization. This temporal", + "bbox": [ + 511, + 703, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7713", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3edf117a1a50ba98303720e553090ad512d87af30ad0445005a4c63ae61735e5.jpg", + "image_caption": [ + "Figure 7. Temporal Analysis of abnormal score dynamics across FFHQ [17], ImageNet [10], LSUN-(Bedroom, Cat, Horse) [40]." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 473, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pattern suggests that, while early intervention might be possible, determining the latest effective correction point is crucial for maximizing the detection of potential artifacts.", + "bbox": [ + 89, + 305, + 483, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Correction Timing We investigate how the choice of correction timestep $T_{c}$ affects artifact removal effectiveness. Through extensive experiments, we identify a threshold at approximately $T_{c}^{*} / T \\approx 0.48$ across different diffusion processes, representing the latest viable correction point before the model lacks sufficient steps for refinement. As shown in Fig. 8, both Precision and Recall metrics achieve optimal performance as $T_{c}$ approaches $T_{c}^{*}$ . This optimal timing allows for maximum artifact detection while ensuring adequate refinement steps. Most datasets show stable performance before $T_{c}^{*}$ followed by a sharp decline, while FFHQ exhibits more fluctuation before $T_{c}^{*}$ and decreases gradually afterward. Individual dataset curves with analysis are provided in Sec. 7.4.", + "bbox": [ + 89, + 364, + 483, + 575 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9405cd4cdc82cd75d04436c3b5a4a61ae23603345a5b8415babb3f31b26506e7.jpg", + "image_caption": [ + "Figure 8. Impact of correction timestep $T_{c}$ on artifact removal performance evaluated by Precision (fidelity, $\\uparrow$ ) and Recall (diversity, $\\uparrow$ ). The dashed lines indicate the baseline precision / recall of the original diffusion model on each dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 579, + 285, + 727 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4aa01c91f6f73558454704392bf23ddf91f6d4b1ecfb58f77a8b90875b262cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 579, + 478, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Latent Code Improvement To evaluate how our correction method influences latent representations, we conduct a linear probe experiment [39] using a classifier-guided diffusion model [11] on ImageNet [10]. Specifically, we generate samples following two paths: the original diffusion process and our corrected process. At an intermediate timestep $t$ , we obtain the original state $x_{t}$ and apply our correction", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method to get the corrected state $\\hat{x}_t$ . We then continue the diffusion process for $k$ steps to obtain $x_{t - k}$ and $\\hat{x}_{t - k}$ from the original and corrected states, respectively. We generate $N$ labeled samples using both paths, with $y^{i}$ denoting the class label, resulting in two sets:", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\text {o r i g}} = \\left\\{\\left(x _ {t - k} ^ {i}, y ^ {i}\\right) \\right\\} _ {i = 1} ^ {N}, \\quad \\mathcal {D} _ {\\text {c o r r}} = \\left\\{\\left(\\hat {x} _ {t - k} ^ {i}, y ^ {i}\\right) \\right\\} _ {i = 1} ^ {N} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 174, + 906, + 194 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We train two separate classifiers on $\\mathcal{D}_{\\mathrm{orig}}$ and $\\mathcal{D}_{\\mathrm{corr}}$ respectively, with results shown in Fig. 9. The higher accuracy achieved by improved latent codes throughout the remaining steps demonstrates that our method enhances the semantic quality of latent representations, and this improvement affects overall accuracy rather than individual precision or recall metrics. Notably, we observe that the classification accuracy reaches its peak earlier in the generation process with our method and remains stable through the refinement phase, while the original process exhibits a decrease-increase pattern during refinement.", + "bbox": [ + 511, + 202, + 906, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/246bdaab1a498bcd3e41621929fbc9e3f477dc2d8dd049aba64c8e2c1f145ea4.jpg", + "image_caption": [ + "Figure 9. Linear probe classification results comparing latent representations from original and corrected diffusion trajectories across Accuracy, Precision and Recall metrics." + ], + "image_footnote": [], + "bbox": [ + 517, + 371, + 901, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 603, + 633, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel analysis of the diffusion generation process, decomposing it into profiling, mutation, and refinement phases, which provides fundamental insights into artifact formation mechanisms. Based on these insights, we develop ASCED, an unsupervised framework that successfully detects and corrects artifacts while preserving generation diversity. Extensive experiments demonstrate that ASCED achieves competitive performance with state-of-the-art supervised methods across multiple datasets. The training-free nature of our approach enables immediate application to any diffusion model, making it a practical solution for improving generation quality.", + "bbox": [ + 511, + 628, + 905, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future Work While our approach effectively detects artifacts through temporal pattern analysis, promising directions include improving detection in low-contrast regions, developing more robust discrimination between transient and persistent abnormalities, and extending these insights to other generative frameworks.", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7714", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by Veritone, Adobe, and has utilized Queen Mary's Apocrita HPC facility from QMUL Research-IT.", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 11", + "[2] Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3):313-326, 1982. 3", + "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 6", + "[4] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 3", + "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1", + "[6] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018.3", + "[7] Yu Cao and Shaogang Gong. Few-shot image generation by conditional relaxing diffusion inversion. arXiv preprint arXiv:2407.07249, 2024. 1", + "[8] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 5", + "[9] Erik Daxberger, Agustinus Kristiadi, Alexander Immer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace reflux-effortless bayesian deep learning. Advances in Neural Information Processing Systems, 34:20089-20103, 2021. 2, 3", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7, 8", + "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 6, 8", + "[12] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890-7899, 2020. 3", + "[13] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images." + ], + "bbox": [ + 93, + 198, + 483, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Advances in neural information processing systems, 33: 3022-3032, 2020. 3", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[15] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2016. 1", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4", + "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 6, 7, 8", + "[18] Mohammad Khan, Didrik Nielsen, Voot Tangkaratt, Wu Lin, Yarin Gal, and Akash Srivastava. Fast and scalable bayesian deep learning by weight-perturbation in adam. In International conference on machine learning, pages 2611-2620. PMLR, 2018. 3", + "[19] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 5, 14", + "[20] Siqi Kou, Lei Gan, Dequan Wang, Chongxuan Li, and Zhijie Deng. Bayesdiff: Estimating pixel-wise uncertainty in diffusion via bayesian inference. arXiv preprint arXiv:2310.11142, 2023. 2, 3, 6", + "[21] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 6", + "[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 3, 6, 11", + "[23] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8060-8069, 2020. 3", + "[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 1", + "[25] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3, 14", + "[26] David John Cameron Mackay. Bayesian methods for adaptive models. California Institute of Technology, 1992. 3", + "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1, 3, 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7715", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3", + "[29] Hippolyt Ritter, Aleksandar Botev, and David Barber. A scalable laplace approximation for neural networks. In 6th international conference on learning representations, ICLR 2018-conference track proceedings. International Conference on Representation Learning, 2018. 3", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3", + "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1", + "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 5", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5, 6", + "[35] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1, 3, 4", + "[36] Charles M Stein. Estimation of the mean of a multivariate normal distribution. The annals of Statistics, pages 1135-1151, 1981. 3", + "[37] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 3", + "[38] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 3", + "[39] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. arXiv preprint arXiv:2303.09769, 2023. 8", + "[40] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 6, 7, 8" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 3", + "[42] Zhongqi Yue, Jiankun Wang, Qianru Sun, Lei Ji, Eric I Chang, Hanwang Zhang, et al. Exploring diffusion timesteps for unsupervised representation learning. arXiv preprint arXiv:2401.11430, 2024. 3", + "[43] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7579-7590, 2023. 2, 3, 6, 7", + "[44] Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cyclic stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932, 2019. 3", + "[45] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 3", + "[46] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3", + "[47] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2", + "[48] Zijian Zhang, Zhou Zhao, and Zhijie Lin. Unsupervised representation learning from pre-trained diffusion probabilistic models. Advances in neural information processing systems, 35:22117-22130, 2022. 3", + "[49] Qingping Zheng, Ling Zheng, Yuanfan Guo, Ying Li, Songcen Xu, Jiankang Deng, and Hang Xu. Self-adaptive reality-guided diffusion for artifact-free super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25806-25816, 2024. 3, 6, 7, 11" + ], + "bbox": [ + 516, + 92, + 903, + 681 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7716", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_model.json b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_model.json new file mode 100644 index 0000000000000000000000000000000000000000..09a57b4cecbebb56c0636f15877e7ec27b9d0f46 --- /dev/null +++ b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.898, + 0.153 + ], + "angle": 0, + "content": "Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.181, + 0.764, + 0.218 + ], + "angle": 0, + "content": "Yu Cao Zengqun Zhao Ioannis Patras Shaogang Gong Queen Mary University of London" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.22, + 0.722, + 0.234 + ], + "angle": 0, + "content": "{yu.cao, zengqun.zhao, i.patras, s.gong}@qmul.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.261, + 0.907, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.545, + 0.907, + 0.63 + ], + "angle": 0, + "content": "Figure 1. Why do diffusion models generate artifacts? We discover that a diffusion generative process necessarily undergoes three phases, we call them: (2) \"Profiling\" which recovers holistic mean templates, (2) \"Mutation\" which introduces local divergence, and (3) \"Refinement\" which rationalizes pixel-wise generation in spatial context. Four visual examples are shown: The first two rows are two examples of rational local mutations (in green boxes) either naturally integrated (Row 1) or reasonably eliminated (Row 2). The bottom two rows show two failure cases when mutations were trapped unreasonably (in red boxes), resisting refinement and resulting in artifacts. Phases are visualized in equal intervals for clarity; please zoom in for more details." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.642, + 0.327, + 0.656 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.674, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Visual artifacts remain a persistent challenge in diffusion models, even with training on massive datasets. Current solutions primarily rely on supervised detectors, yet lack understanding of why these artifacts occur in the first place. In our analysis, we identify three distinct phases in the diffusion generative process: Profiling, Mutation, and Refinement. Artifacts typically emerge during the Mutation phase, where certain regions exhibit anomalous score dynamics over time, causing abrupt disruptions in the normal evolution pattern. This temporal nature explains why existing methods focusing only on spatial uncertainty of the final output fail at effective artifact localization. Based on these insights, we propose ASCED (Abnormal Score Correction for Enhancing Diffusion), that detects artifacts by monitoring abnormal score dynamics during the diffusion process," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.794 + ], + "angle": 0, + "content": "with a trajectory-aware on-the-fly mitigation strategy that appropriate generation of noise in the detected areas. Unlike most existing methods that apply post hoc corrections, e.g., by applying a noising-denoising scheme after generation, our mitigation strategy operates seamlessly within the existing diffusion process. Extensive experiments demonstrate that our proposed approach effectively reduces artifacts across diverse domains, matching or surpassing existing supervised methods without additional training. Project page: YuCao16.github.io/ASCED." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.646, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Diffusion models have emerged as powerful foundation models in computer vision [5], achieving remarkable success in image generation [7, 15-17, 32], image inpainting [24, 27, 35], and text-to-image task [30-32]. However, even trained on large-scale datasets, diffusion generative images" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7707" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.079, + 0.852, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.275, + 0.812, + 0.29 + ], + "angle": 0, + "content": "Figure 2. Diagram of our framework. Denoising and Noising are using Eq. (5) and Eq. (1), respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.296, + 0.483, + 0.403 + ], + "angle": 0, + "content": "still exhibit two significant flaws: visual artifacts and hallucinations [43, 47]. Visual artifacts appear as local irregularities in texture or structure, while hallucinations involve semantically incoherent content, e.g., extra limbs or misplaced objects. In this work, we focus on addressing diffusion artifacts, which present a fundamental challenge to achieving reliable and high-quality image generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.483, + 0.553 + ], + "angle": 0, + "content": "Existing methods primarily treat visual artifact detection as a classification problem, i.e., identifying problematic generations for filtering or reconstruction. These methods typically rely on a specialized classifier, either trained on manually annotated artifact datasets [43] or leveraging a pre-trained Large Multi-Modal Model (LMM) [22]. However, such post hoc interventions fail to address a fundamental problem: Why and when do artifacts emerge in diffusion models? To bridge this gap, we begin by examining the diffusion generation process itself." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.554, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We discover that while diffusion process is guided by the same fundamental equation across time (i.e. diffusion steps), in practice, the model exhibits different behavior that can be roughly categorized in three different temporal phases that we name Profiling, Mutation, and Refinement. In the \"Profiling\" phase, the model sketches the basic semantic global layout; in the \"Mutation\" phase it explores potential local pixel-wise variations to create local structure; in the \"Refinement\" phase it attempts to resolve these local pixel-wise variations into coherent visual details in context (see Fig. 1 for visual examples and Sec. 3.2 for a detailed analysis). This understanding of the generation process reveals that while visual artifacts may appear randomly, they follow systematic and identifiable temporal patterns during image formation. Recent uncertainty-based approaches try to identify artifacts by converting diffusion models into Bayesian networks and employ techniques such as Last Layer Laplace Approximation [9] to generate pixel-level variance matrices [20]. However, these uncertainty quantification analyses only capture spatial variations in the final output, i.e., \\(Var(x_0)\\), missing crucial temporal dynamics during the generation process. Our study shows that diffusion artifacts emerge when certain image regions ex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.905, + 0.403 + ], + "angle": 0, + "content": "hibit abnormal evolution patterns over time, primarily during the Mutation phase of the generation process. Specifically, their pixel values stop updating properly while the surrounding areas continue to evolve. This phenomenon, which we formally define as \"score traps\" in Sec. 3.2, explains why examining only the final output is insufficient and misplaced for effective diffusion artifact detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.905, + 0.691 + ], + "angle": 0, + "content": "Building on these insights, we propose ASCED: Abnormal Score Correction for Enhancing Diffusion, as shown in Fig. 2. At the heart of the method is the estimation of a score, in this context the direction and magnitude of pixel-wise evolution at each diffusion generative step, and a scheme that analyses its temporal dynamics and detects abnormalities. We show both theoretically and experimentally that these abnormalities strongly correlate with artifact formation, making early detection possible at a stage where intervention is still feasible, before artifacts become irreversibly embedded in the generation process. We leverage this early detection capability by implementing a novel trajectory-aware correction mechanism that disrupts the evolution of artifact regions while preserving overall generation diversity. Importantly, ASCED operates in a fully unsupervised manner without requiring manual annotations or domain-specific training, making it readily applicable across various domains, , particularly valuable when training data may be limited or protected." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.905, + 0.828 + ], + "angle": 0, + "content": "Our contributions are: (1) We provide new insights into the formation of visual artifacts in the diffusion generative process, advancing the understanding of diffusion model internal mechanisms. (2) We introduce a novel method that detects potential artifact regions by monitoring abnormal score dynamics temporally, without any manually labeled training required. (3) We further develop a on-the-fly trajectory-aware correction mechanism that effectively mitigates artifacts while preserving image diversity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.663, + 0.86 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Visual Artifact Detection initially targeted superresolution artifacts, where upsampling operations are the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7708" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.392 + ], + "angle": 0, + "content": "main source [45]. These methods analyze either spatial domain characteristics to capture texture differences between real and generated images [23, 41], or frequency domain patterns to study artifact characteristics in high-frequency components [12, 13]. More recent work has shifted focus to detecting artifacts in general image generation, developing specialized classifiers trained on manually annotated datasets [43] or utilizing pre-trained large vision models [22]. However, these supervised approaches require extensive manual labeling and may not generalize well across different domains. A parallel direction explores uncertainty quantification methods to understand visual artifacts. While various approaches including variational inference [4, 18], Laplace approximation [26, 29], and Markov Chain Monte Carlo [38, 44] have been developed, their application to diffusion models remains limited [20]. BayesDiff [20] pioneers pixel-level uncertainty quantification in diffusion models using Last-layer Laplace approximation [9], yet the connection between spatial uncertainty and visual artifacts remains unclear." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Generation Quality Enhancement Various approaches have been proposed to enhance generation quality. The truncation trick in BigGAN [6] demonstrated that restricting the sampling space can significantly improve generation fidelity, suggesting similar principles might apply to diffusion models. For diffusion-based generation, classifier guidance [11] has been introduced to steer the generation process. SARGD [49] extends this idea by utilizing a pre-trained artifact detector to guide the generation towards artifact-free regions. Latent diffusion models [30] take a different approach by applying diffusion in a learned latent space, demonstrating that controlled evolution in a constrained space can lead to higher-quality output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.756 + ], + "angle": 0, + "content": "Diffusion Model for Representation Learning has evolved toward latent space disentanglement and controllable editing. The former aims to uncover interpretable factors in the generative process. Recent studies [28, 42, 48] observed stage-wise attribute emergence during generation, but focus differently than our analysis of artifact formation mechanisms. Controllable editing techniques [27, 31, 37, 46] can be applied for artifact removal, yet require per-sample manipulation and address symptoms rather than causes. Our approach instead corrects abnormal score dynamics during the generation process itself." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.226, + 0.787 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our approach consists two steps: Detection (Sec. 3.2) and Correction (Sec. 3.3). Specifically, we localize artifact pixel regions by identifying abnormal score dynamics during the diffusion inference process (Detection) and develop a novel artifact correction algorithm without delaying inference (Correction). We give a theoretical analysis of our key concepts on score trap and temporal weighting (Sec. 3.4)." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.094, + 0.891, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.285 + ], + "angle": 0, + "content": "Figure 3. Artifact generation through denoising (top) and brush stroke noising via SDEdit [27] (bottom), demonstrating the model's inability to distinguish artifacts during generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.288, + 0.654, + 0.303 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.31, + 0.905, + 0.401 + ], + "angle": 0, + "content": "Diffusion Model Let \\( x_0 \\in \\mathbb{R}^{c \\times h \\times w} \\) be an image. The forward process of a diffusion model gradually diffuses the data distribution \\( q(x) \\) towards \\( q_t(x_t) \\), \\( \\forall t \\in [0, T] \\), with \\( q_T(x_T) = \\mathcal{N}(\\mathbf{0}, \\mathbf{I}) \\) as a trivial Gaussian distribution. From a score viewpoint, it can be described by the Stochastic Differential Equation (SDE):" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.403, + 0.905, + 0.419 + ], + "angle": 0, + "content": "\\[\nd \\mathbf {x} _ {t} = \\boldsymbol {f} \\left(\\mathbf {x} _ {t}, t\\right) d t + \\boldsymbol {g} \\left(\\mathbf {x} _ {t}, t\\right) d \\mathbf {w} _ {t}, \\quad t \\in [ 0, T ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.42, + 0.905, + 0.479 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}\\) is the standard Wiener process, \\(f(\\cdot)\\) and \\(g(\\cdot)\\) are scalar drift and diffusion coefficients, respectively. Anderson [2] states that the reverse process of Eq. (1) is also a diffusion process as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.48, + 0.905, + 0.511 + ], + "angle": 0, + "content": "\\[\nd \\mathbf {x} _ {s} = \\left[ \\boldsymbol {f} \\left(\\mathbf {x} _ {s}, s\\right) - \\boldsymbol {g} \\left(x _ {s}, s\\right) ^ {2} \\boldsymbol {s} \\left(x _ {s}, s\\right) \\right] d s + \\boldsymbol {g} \\left(x _ {s}, s\\right) d \\mathbf {w} _ {s} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.512, + 0.905, + 0.587 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_s\\coloneqq x_{T - t}\\) and \\(s(x_s,s)\\coloneqq \\nabla_{\\mathbf{x}_s}\\log p_s(\\mathbf{x}_s)\\) is the score function of the marginal distribution over \\(x_{s}\\). Song et al. [35] leverage this property to generate samples by first drawing \\(x_{T}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and then solving the reverse SDE using a learned score network \\(s_\\theta (x_t,t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.587, + 0.905, + 0.632 + ], + "angle": 0, + "content": "By using Tweddie's formula [36], DDPM (Denoising Diffusion Probabilistic Model) [16] can be shown as an equivalent interpretation of Eq. (2) [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.632, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t}, t\\right) = - \\frac {1}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t}, t\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.905, + 0.689 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s\\), with mean coefficient \\(\\alpha_t\\), and \\(\\epsilon_\\theta(x_t, t)\\) is the noise network of DDPM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Definition of Visual Artifacts Generation flaws can be distinguished into two categories: Visual Artifacts and Hallucinations. Visual artifacts manifest as local irregularities or distortions in a generated image, such as blurred patches, unnatural textures, broken structures. In contrast, hallucinations refer to semantically generating incoherent content, such as extra limbs, misplaced objects or counterfactuals. In this paper, we focus specifically on detecting and correcting visual artifacts generated by diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.858, + 0.864 + ], + "angle": 0, + "content": "3.2. Detection by Anomalous Score Dynamic" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To understand how visual artifacts emerge during generations, we examine diffusion model behavior through the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7709" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.382, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.36, + 0.214, + 0.373 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.091, + 0.588, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.36, + 0.493, + 0.373 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.091, + 0.899, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.36, + 0.805, + 0.373 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.375, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Figure 4. Visualization of score dynamics and visual artifact detection. (a) Generated images with detected visual artifact regions highlighted (red). (b) Visualization of score dynamics (normalized) between adjacent time steps as activation maps. Brighter regions (green to yellow) indicate areas of higher score variation, while darker regions (blue to black) show areas of lower score change. (c). Score acceleration curves (representing the rate of change in score dynamics between consecutive timesteps) comparing artifact regions (red) with non-artifact regions (blue). The artifact regions exhibit characteristic rapid acceleration followed by deceleration, while non-artifact regions maintain stable score dynamics over time throughout a generative (inference) process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.472, + 0.483, + 0.653 + ], + "angle": 0, + "content": "lens of image editing. SDEdit [27] demonstrates that diffusion models can transform irregularities, e.g., brush strokes, into semantically meaningful content through a noise-then-denoise process, revealing the inherent Refinement capability of diffusion models. We observe that such irregularities after noising become structurally indistinguishable from states containing artifacts during generation, as shown in Fig. 3. While diffusion models can successfully refine noised brush strokes, they fail to correct the corresponding artifacts during generation. This contrast reveals that diffusion models lack the ability to identify artifacts as patterns requiring refinement during the generation process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.883 + ], + "angle": 0, + "content": "To better understand this, we examine the generation process from a score perspective, as it directly represents the evolution of pixel values [35]. We define score dynamics as the difference between temporally adjacent score values: \\(\\Delta s_{\\theta}(x_t^{i,j},t) = s_{\\theta}(x_t^{i,j},t) - s_{\\theta}(x_{t - 1}^{i,j},t - 1)\\). Analysis reveals that image generation begins with establishing basic structures, followed by a phase of stochastic exploration where irregular patterns may emerge; we call these phases Profiling and Mutation, respectively. As shown in Fig. 4, regions containing visual artifacts exhibit characteristic patterns during mutation: They appear as localized regions of intense score variations (Fig. 4 (b)) and display dramatic acceleration followed by sudden deceleration in their score trajectories (Fig. 4 (c)). In contrast, normal regions maintain a stable evolution throughout generation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Based on these observations, we propose a novel ap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.639 + ], + "angle": 0, + "content": "proach to detecting and localizing potential artifact regions over time in a diffusion inference process. Specifically, let \\(\\Omega \\subset \\mathbb{R}^2\\) denote the spatial domain of the image, and \\(\\Omega_t^a \\subset \\Omega\\) represent regions where abnormal evolution patterns emerge at timestep \\(t\\). For each spatial location \\((i,j) \\in \\Omega\\), we track the score dynamics through consecutive timesteps. To account for the varying score magnitudes across different images and timesteps, we maintain a score bank \\(\\mathcal{S} = \\{s_\\theta(x_k,k)\\}_{k=t}^T\\) and apply a temporal weighting function \\(w(t)\\) that addresses the inherent decay of score magnitudes. Formally, we define artifact regions as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.641, + 0.907, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {t} ^ {a} := \\left\\{(i, j) \\in \\Omega \\mid \\left| \\Delta \\left(w (t) \\cdot s _ {\\theta} \\left(x _ {t} ^ {i, j}, t\\right)\\right) \\right| > \\tau \\right\\} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.907, + 0.765 + ], + "angle": 0, + "content": "where \\( w(t) = \\frac{1 - \\bar{\\alpha}_t}{\\sqrt{\\bar{\\alpha}_t}} \\) (see Sec. 3.4 for theoretical analysis) and \\( \\tau \\) is adaptively determined as the maximum between the Median Absolute Deviation (MAD) of the weighted score dynamics and the mean of score bank \\( S \\). The final artifact regions \\( \\Omega^a \\) are accumulated across the score bank, with detailed procedures provided in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.72, + 0.787 + ], + "angle": 0, + "content": "3.3. Real-Time Correction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "After detecting artifacts, we need a correction strategy to effectively refine the artifact region. Two natural approaches emerge: either to revert to the states before the artifacts emerge through state replacement or to limit directly abnormal score changes through score clipping. For state replacement, we first predict the clean image from an earlier timestep \\( t \\) using [16]:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7710" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.091, + 0.471, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Pseudo Code for Proposed ASCED Method" + }, + { + "type": "algorithm", + "bbox": [ + 0.092, + 0.109, + 0.907, + 0.297 + ], + "angle": 0, + "content": "1: Input: Score network \\( s_{\\theta}(\\cdot) \\) which requires \\( T \\) steps to generate, detection starting step \\( T_d \\) and correction step \\( T_c \\)\n2: Initialize \\( x_T \\sim \\mathcal{N}(0, \\mathbf{I}) \\), Score Bank \\( S \\gets \\{\\} \\), Visual Artifact Mask \\( \\Omega^a \\gets \\{\\} \\)\n3: for \\( t = T, t \\)-, while \\( t >= 0 \\) do\n4: \\( x_{t-1} = \\sqrt{\\alpha_{t-1}} x_0 - \\sqrt{1 - \\alpha_{t-1}} \\sqrt{1 - \\bar{\\alpha}_t} s_\\theta(x_t, t) \\), where \\( x_0 = \\frac{x_t + (1 - \\bar{\\alpha}_t) s_\\theta(x_t, t)}{\\sqrt{\\bar{\\alpha}_t}} \\) ▷ Original Diffusion Process\n5: if \\( T_c < t <= T_d \\) then\n6: S.append \\( s_\\theta(x_t, t) \\) ▷ Store score value into Score Bank\n7: else if \\( t == T_c \\) then ▷ Anomalous Score Dynamics Detection Step\n8: for \\( k = 0, k++ \\), while \\( k < (T_c - T_d) \\) do ▷ Determine \\( \\Omega^a \\) by accumulation\n9: \\( \\Omega^a = \\Omega^a \\cup \\{(i, j) \\in \\Omega | |\\Delta(w(k) \\cdot s_\\theta(x_k^{i,j}, k))| > \\tau\\} ▷ \\tau = \\max\\{\\mathrm{MAD}(\\Delta(w(k) \\cdot s_\\theta(x_k^{i,j}, k)), \\mathrm{mean}(\\mathcal{S})\\} \\)\n10: \\( x_t = x_t \\cdot \\mathbb{1}_{\\overline{\\Omega}^a} + (\\sqrt{\\bar{\\alpha}_t} \\hat{x}_0(t) + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon) \\cdot \\gamma(t) \\xi_{\\Omega^a} \\) ▷ Trajectory-aware Targeted Correction Step\n11: return \\( x_0 \\)" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.322, + 0.482, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {0} \\left(\\boldsymbol {x} _ {t}, t\\right) = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {x} _ {t} + \\left(1 - \\bar {\\alpha} _ {t}\\right) \\nabla_ {\\theta} \\log p \\left(\\boldsymbol {x} _ {t}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Then replacing the artifact regions with corresponding states from this predicted clean image after re-noising to the current timestep. Score clipping directly limits the magnitude of score changes during inference. However, both state replacement and score clipping fundamentally disrupt the mutation process, leading to reduced generation diversity. To address this problem, we propose Trajectory-aware Targeted Correction (TTC), which introduces controlled perturbations specifically in artifact regions at correction timestep \\( T_{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.508, + 0.483, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {T _ {c}} = \\boldsymbol {x} _ {T _ {c}} \\cdot \\mathbb {1} _ {\\overline {{\\Omega}} ^ {a}} + \\left(\\sqrt {\\bar {\\alpha} _ {T _ {c}}} \\boldsymbol {x} _ {0} ^ {\\prime} + \\sqrt {1 - \\bar {\\alpha} _ {T _ {c}}} \\epsilon\\right) \\cdot \\gamma \\xi \\mathbb {1} _ {\\Omega^ {a}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.569 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_0^{\\prime} = \\hat{\\pmb{x}}_0(\\pmb{x}_{T_c},T_c)\\) \\(\\epsilon ,\\xi \\stackrel {\\mathrm{iid}}{\\sim}\\mathcal{N}(0,\\mathbf{I}),\\mathbb{1}_{\\Omega^a}\\) is indicator function for artifact regions and perturbation intensity \\(\\gamma\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.705 + ], + "angle": 0, + "content": "TTC builds upon the understanding of score traps: Regions where pixels become locked in persistent score patterns after experiencing dramatic score changes. Through controlled perturbations, TTC disrupts these fixed patterns and allows pixels to resume normal evolution with surrounding regions. Sec. 3.4 provides a detailed analysis of these score trap mechanisms and their relationship to visual artifacts. Generation quality and diversity comparison across correction strategies is shown in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.285, + 0.727 + ], + "angle": 0, + "content": "3.4. Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.764 + ], + "angle": 0, + "content": "We provide a theoretical analysis for the score trap and the choice of the temporal weighting function in Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Theoretical Analysis of Score Traps For normal generation, the score evolution of each pixel is coupled with its neighborhood through the learned neural score function [8]:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.816, + 0.483, + 0.838 + ], + "angle": 0, + "content": "\\[\ns _ {\\theta} \\left(x _ {t} ^ {i, j}, t\\right) = \\nabla_ {x _ {t} ^ {i, j}} \\log p _ {\\theta} \\left(x _ {t} ^ {i, j} \\mid \\mathcal {C} (i, j), t\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}(i,j)\\) represents the contextual information from neighboring pixels. This coupling ensures coordinated evolution toward the data manifold. When regions experience abnormal score dynamics, they can enter score traps" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.905, + 0.401 + ], + "angle": 0, + "content": "where local patterns persist despite significant score values, disrupting the natural coupled evolution process. These trapped regions evolve based primarily on their local patterns, losing the contextual relationships necessary for coherent image generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.906, + 0.552 + ], + "angle": 0, + "content": "This reveals how our perturbation-based correction re-establishes contextual relationships. For a trapped pixel \\((i,j)\\), the perturbation \\(\\gamma \\cdot \\xi\\) introduces stochastic variations that disrupt the isolated score patterns, creating opportunities for these regions to re-couple with their surroundings through the natural score evolution process. Meanwhile, in areas without abnormal patterns, these modest perturbations preserve the original coupled evolution, ensuring the method remains harmless to non-artifact regions; see Sec. 7.2 for further mathematical derivations and proofs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.613 + ], + "angle": 0, + "content": "Score Normalization The score function learned by diffusion models can be interpreted as a vector field guiding the denoising trajectory, inducing a probability flow [34]:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.614, + 0.905, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\frac {d \\boldsymbol {x}}{d t} = - \\frac {1}{2} \\sigma_ {t} ^ {2} \\nabla \\log p _ {t} (\\boldsymbol {x} _ {t}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.905, + 0.69 + ], + "angle": 0, + "content": "The temporal evolution of this probability flow can be characterized by its divergence in the probability density field. Theoretically, we can model this through a flow operator \\(\\mathcal{G}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.694, + 0.905, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} (t) = \\nabla \\cdot \\left(\\frac {\\partial}{\\partial t} \\int_ {\\tau = 0} ^ {t} \\mathcal {P} (\\boldsymbol {x} _ {\\tau}, \\tau) d \\tau\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}(\\pmb{x}_{\\tau},\\tau)\\) denotes the local probability density at position \\(\\pmb{x}_{\\tau}\\) and time \\(\\tau\\). This formulation allows us to monitor the accumulation of probability density changes over time. We observe that the clean image prediction (Eq. (5)) reflects these changes in the probability flow. Under the assumption of smooth probability density evolution between adjacent time steps [19, 33, 34], score dynamics are captured through \\(\\frac{\\partial}{\\partial t}\\hat{\\pmb{x}}_0(t)\\). Consequently, a normalization factor \\(w(t) = \\frac{1 - \\alpha_t}{\\sqrt{\\alpha_t}}\\) is derived from the coefficient of the score term in Eq. (5), helping to equalize the scale of score variations throughout the denoising process." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7711" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.077, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparisons on five datasets. The methods compared include BayesDiff [20] and SARGD [49], and three baseline methods: State Replacement Score Clipping and PAL [43] + TTC. All methods use DDIM sampling with identical noise seeds to generate 10,000 images per dataset, ensuring each approach modifies the same deterministic trajectories for fair comparison. The best scores are in bold and second best in underline with bold. Sup and UnS denote supervised and unsupervised methods, respectively." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.132, + 0.905, + 0.251 + ], + "angle": 0, + "content": "
MethodsTypeFFHQ[17]ImageNet[10]LSUN-Cat[40]LSUN-Horse[40]LSUN-Bedroom[40]
FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑
Original [34]UnS36.690.6290.49314.680.7390.73422.170.5130.58629.360.5100.64212.960.6270.583
State ReplaceUnS37.090.6350.49514.610.7430.73322.790.5100.58730.360.5020.64212.950.6280.574
Score ClippingUnS36.360.6300.49814.580.7420.73622.120.5150.58529.260.5110.64212.920.6270.585
BayesDiff [20]UnS36.990.6320.49114.530.7430.73022.500.5130.58528.700.5180.63412.880.6250.569
SARGD [49]Sup38.370.6370.46415.340.7310.72722.650.5230.57030.020.5100.62113.820.6390.554
PAL [43] + TTCSup36.350.6240.50014.010.7310.74721.830.5140.58828.680.5190.64612.710.6290.579
ASCED (Ours)UnS36.280.6370.50314.410.7500.73521.910.5150.59327.660.5210.65212.530.6280.590
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.257, + 0.223, + 0.274 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.284, + 0.482, + 0.42 + ], + "angle": 0, + "content": "Basic setups We conducted experiments on five datasets: FFHQ [17], ImageNet [10], LSUN-Bedroom [40], LSUN-Cat [40], and LSUN-Horse [40]. We employed the Guided Diffusion model framework and pre-trained weights from OpenAI [11] and Segmentation-DDPM [3]. Quantitative evaluations were performed using FID [14], which measures the Fréchet distance between real and generated image distributions, along with Precision and Recall [21], which evaluate sample fidelity and diversity, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.482, + 0.573 + ], + "angle": 0, + "content": "Implementation details For detecting diffusion artifacts, our approach was compared with LLaVA-v1.5-13B [22] and PAL [43], while artifact removal comparisons were made with BayesDiff [20] and the adapted SARGD [49]. All experiments were performed on NVIDIA A100 / H100 GPUs. We used DDIM [34] to improve inference efficiency with a Number of Function Evaluations (NFE) set to 25. Remark: We demonstrate in Sec. 7.1 that there is no significant correlation between NFE and the generation of artifacts. Full implementation details are provided in Sec. 6." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.588, + 0.483, + 0.604 + ], + "angle": 0, + "content": "4.1. Quantitative Comparisons to Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.838 + ], + "angle": 0, + "content": "We first evaluate the effectiveness of ASCED in improving generative quality through comparisons with both unsupervised (BayesDiff [20]) and supervised (SARGD [49]) SOTA methods, along with the original diffusion model [11] and two baselines from Sec. 3.3 (state replacement and score clipping). To isolate the effectiveness of our correction method, we also evaluate a hybrid approach combining artifact detector PAL (used in SARGD) with our Trajectory-aware Targeted Correction (TTC). Quantitative results are shown in Tab. 1. Among unsupervised methods, our ASCED demonstrates superior performance across all datasets, consistently achieving better FID and Precision scores while maintaining higher Recall values than BayesDiff and baselines, indicating both improved generation quality and better preservation of diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Compared to the supervised methods, our proposed ASCED method shows leading performance across most experiments, achieving superior results on FFHQ, LSUN-Horse, and LSUN-Bedroom, while maintaining competitive" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.439 + ], + "angle": 0, + "content": "performance on ImageNet and LSUN-Cat. The better performances of PAL and SARGD on these two datasets are due to that they are supervised artifact detectors specifically trained on these datasets [43]. Their advantages are not generalizable to other datasets FFHQ, LSUN-(Horse, Bedroom). In contrast, ASCED as an unsupervised method has generalisable advantages in all domains without dataset specific training, making it more practical and scalable. Additionally, our method demonstrates significant computational efficiency advantages. In our experiments, ASCED detects and corrects artifacts in approximately 0.09 s per image, which is \\(8.8 \\times\\) faster than PAL (0.79 s)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.905, + 0.531 + ], + "angle": 0, + "content": "The effectiveness of the correction mechanism (TTC) becomes particularly evident when comparing SARGD with PAL (used in SARGD) + TTC, where the trajectory-aware correction demonstrates significant advantages in preserving generation diversity across all datasets, reflected in consistently higher Recall scores." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "Table 2. Visual Artifact Detection Accuracy Comparison between PAL (Supervised, Sup) [43], LLaVA (Zero-Shot, ZS) [22], and our method (Unsupervised, UnS) on FFHQ [17], ImageNet [10], and LSUN-(Bedroom, Cat, Horse) [40]." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.595, + 0.905, + 0.643 + ], + "angle": 0, + "content": "
MethodTypeFFHQImageNetBedroomCatHorse
PALSup51.4%69.2%52.4%69.8%60.9%
LLaVAZS63.1%91.1%75.9%59.5%72.2%
OursUnS56.7% (-6.4)67.7% (-1.5)65.0% (-10.9)68.3% (-1.5)70.3% (-1.9)
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.861, + 0.667 + ], + "angle": 0, + "content": "4.2. Artifact Detection Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.905, + 0.825 + ], + "angle": 0, + "content": "To validate the accuracy of our method in identifying visual artifacts, we manually selected 200 images for each dataset from the diffusion model outputs, consisting of 100 images with visual artifacts and 100 without. We then evaluated ASCED against zero-shot large multi-modal model LLaVAv1.5-13B [22] and supervised artifact detector PAL [43]. For LLaVA evaluation, we produced 50 different prompts and reported the results for the most effective prompt. Details on prompt generation are provided in Sec. 6.2. The accuracy for both methods is presented in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As an unsupervised method, ASCED achieves promising detection performance, maintaining close accuracy to supervised approaches LLaVA and PAL across most datasets. Notably, these methods analyze final generated images, whereas our approach detects artifacts during the generation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7712" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.077, + 0.907, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.37, + 0.907, + 0.414 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison of different correction methods. For each example, we show the original output with visual artifacts (left) and zoomed-in views of the artifact regions corrected by different methods (right): SARGD [49], state replacement (Replace), and our trajectory-aware targeted correction (Ours). Rows from top to bottom: FFHQ[17], ImageNet[10], and LSUN-(Cat, Horse, Bedroom)[40]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.562 + ], + "angle": 0, + "content": "process through score dynamics, enabling early intervention before artifacts fully manifest. However, our method does show limitations in specific cases, such as low-contrast images where subtle abnormalities are difficult to distinguish from normal variations (leading to False Negatives), and instances where the diffusion model successfully rationalizes initially abnormal patterns during refinement (causing False Positives). Representative examples of these cases are illustrated in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.462, + 0.59 + ], + "angle": 0, + "content": "4.3. Qualitative Analysis of Correction Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.839 + ], + "angle": 0, + "content": "To better illustrate the advantage of Trajectory-aware Targeted Correction (TTC) over baselines, Fig. 5 shows qualitative comparisons consisting of the original outputs with artifacts, results from state replacement (Sec. 3.3), SARGD [49], and TTC (ours). While all methods can remove artifacts, TTC demonstrates superior detail preservation in corrected regions. Specifically, both state replacement and SARGD tend to converge to similar expressions and local details, constraining natural variations, as they directly modify the generation states. SARGD faces further limitations from its artifact detector being trained on a specific domain [43], affecting its generalization ability. More importantly, by preserving mutation phase operations, our trajectory-aware correction enables diverse yet coherent generations even when correcting the same region. Additional comparison results are provided in Sec. 8.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Do corrections at Non-Artifact Regions harm? As with any detection method, our proposed scheme will inevitably encounter false positives, leading to add perturbations in non-artifact areas. Our experiments demonstrate that ap" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.423, + 0.905, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.547, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Figure 6. Top: Applied our correction method to clean regions (yellow box). Bottom: Typical failure cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.673 + ], + "angle": 0, + "content": "plying the correction mechanism to regions without visual artifacts, as shown in Fig. 6, introduces modest variations while preserving semantic coherence with the surrounding context and not generating new artifacts. Extended visual results and detection performance analysis can be found in Sec. 7.2 and Sec. 4.2, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.681, + 0.698 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Distribution of Abnormal Score Dynamics In Fig. 7, we plot the frequency of abnormal scores at each time step (normalized by total diffusion time step \\( T \\)). The score dynamics demonstrate distinct patterns across different stages: remaining stable in early steps where basic structures emerge, experiencing significant variations in the middle stage, and gradually stabilizing again in the later steps during details' refinement. The presence of the long tail indicates that some variations persist into later steps, suggesting an extended period of details' adjustment. This behavioral pattern naturally aligns with our hypothesis that the generation undergoes three phases: profiling, mutation, and rationalization. This temporal" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7713" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.475, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.271, + 0.483, + 0.3 + ], + "angle": 0, + "content": "Figure 7. Temporal Analysis of abnormal score dynamics across FFHQ [17], ImageNet [10], LSUN-(Bedroom, Cat, Horse) [40]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.484, + 0.352 + ], + "angle": 0, + "content": "pattern suggests that, while early intervention might be possible, determining the latest effective correction point is crucial for maximizing the detection of potential artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Impact of Correction Timing We investigate how the choice of correction timestep \\( T_{c} \\) affects artifact removal effectiveness. Through extensive experiments, we identify a threshold at approximately \\( T_{c}^{*} / T \\approx 0.48 \\) across different diffusion processes, representing the latest viable correction point before the model lacks sufficient steps for refinement. As shown in Fig. 8, both Precision and Recall metrics achieve optimal performance as \\( T_{c} \\) approaches \\( T_{c}^{*} \\). This optimal timing allows for maximum artifact detection while ensuring adequate refinement steps. Most datasets show stable performance before \\( T_{c}^{*} \\) followed by a sharp decline, while FFHQ exhibits more fluctuation before \\( T_{c}^{*} \\) and decreases gradually afterward. Individual dataset curves with analysis are provided in Sec. 7.4." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.58, + 0.287, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.58, + 0.48, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.73, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Figure 8. Impact of correction timestep \\( T_{c} \\) on artifact removal performance evaluated by Precision (fidelity, \\( \\uparrow \\)) and Recall (diversity, \\( \\uparrow \\)). The dashed lines indicate the baseline precision / recall of the original diffusion model on each dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Latent Code Improvement To evaluate how our correction method influences latent representations, we conduct a linear probe experiment [39] using a classifier-guided diffusion model [11] on ImageNet [10]. Specifically, we generate samples following two paths: the original diffusion process and our corrected process. At an intermediate timestep \\( t \\), we obtain the original state \\( x_{t} \\) and apply our correction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "method to get the corrected state \\(\\hat{x}_t\\). We then continue the diffusion process for \\(k\\) steps to obtain \\(x_{t - k}\\) and \\(\\hat{x}_{t - k}\\) from the original and corrected states, respectively. We generate \\(N\\) labeled samples using both paths, with \\(y^{i}\\) denoting the class label, resulting in two sets:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.175, + 0.907, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\text {o r i g}} = \\left\\{\\left(x _ {t - k} ^ {i}, y ^ {i}\\right) \\right\\} _ {i = 1} ^ {N}, \\quad \\mathcal {D} _ {\\text {c o r r}} = \\left\\{\\left(\\hat {x} _ {t - k} ^ {i}, y ^ {i}\\right) \\right\\} _ {i = 1} ^ {N} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.203, + 0.907, + 0.37 + ], + "angle": 0, + "content": "We train two separate classifiers on \\( \\mathcal{D}_{\\mathrm{orig}} \\) and \\( \\mathcal{D}_{\\mathrm{corr}} \\) respectively, with results shown in Fig. 9. The higher accuracy achieved by improved latent codes throughout the remaining steps demonstrates that our method enhances the semantic quality of latent representations, and this improvement affects overall accuracy rather than individual precision or recall metrics. Notably, we observe that the classification accuracy reaches its peak earlier in the generation process with our method and remains stable through the refinement phase, while the original process exhibits a decrease-increase pattern during refinement." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.372, + 0.902, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.547, + 0.906, + 0.588 + ], + "angle": 0, + "content": "Figure 9. Linear probe classification results comparing latent representations from original and corrected diffusion trajectories across Accuracy, Precision and Recall metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.634, + 0.62 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.81 + ], + "angle": 0, + "content": "We present a novel analysis of the diffusion generation process, decomposing it into profiling, mutation, and refinement phases, which provides fundamental insights into artifact formation mechanisms. Based on these insights, we develop ASCED, an unsupervised framework that successfully detects and corrects artifacts while preserving generation diversity. Extensive experiments demonstrate that ASCED achieves competitive performance with state-of-the-art supervised methods across multiple datasets. The training-free nature of our approach enables immediate application to any diffusion model, making it a practical solution for improving generation quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Future Work While our approach effectively detects artifacts through temporal pattern analysis, promising directions include improving detection in low-contrast regions, developing more robust discrimination between transient and persistent abnormalities, and extending these insights to other generative frameworks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7714" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This work was partially supported by Veritone, Adobe, and has utilized Queen Mary's Apocrita HPC facility from QMUL Research-IT." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.269, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[2] Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3):313-326, 1982. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[4] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.484, + 0.493 + ], + "angle": 0, + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.484, + 0.535 + ], + "angle": 0, + "content": "[6] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[7] Yu Cao and Shaogang Gong. Few-shot image generation by conditional relaxing diffusion inversion. arXiv preprint arXiv:2407.07249, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[8] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.703 + ], + "angle": 0, + "content": "[9] Erik Daxberger, Agustinus Kristiadi, Alexander Immer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace reflux-effortless bayesian deep learning. Advances in Neural Information Processing Systems, 34:20089-20103, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[12] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890-7899, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[13] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Advances in neural information processing systems, 33: 3022-3032, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[15] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[18] Mohammad Khan, Didrik Nielsen, Voot Tangkaratt, Wu Lin, Yarin Gal, and Akash Srivastava. Fast and scalable bayesian deep learning by weight-perturbation in adam. In International conference on machine learning, pages 2611-2620. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[19] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[20] Siqi Kou, Lei Gan, Dequan Wang, Chongxuan Li, and Zhijie Deng. Bayesdiff: Estimating pixel-wise uncertainty in diffusion via bayesian inference. arXiv preprint arXiv:2310.11142, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.538, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[21] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 3, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.72 + ], + "angle": 0, + "content": "[23] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8060-8069, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[25] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[26] David John Cameron Mackay. Bayesian methods for adaptive models. California Institute of Technology, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7715" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[28] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.176, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[29] Hippolyt Ritter, Aleksandar Botev, and David Barber. A scalable laplace approximation for neural networks. In 6th international conference on learning representations, ICLR 2018-conference track proceedings. International Conference on Representation Learning, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[35] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[36] Charles M Stein. Estimation of the mean of a multivariate normal distribution. The annals of Statistics, pages 1135-1151, 1981. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[37] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[38] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[39] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. arXiv preprint arXiv:2303.09769, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[41] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[42] Zhongqi Yue, Jiankun Wang, Qianru Sun, Lei Ji, Eric I Chang, Hanwang Zhang, et al. Exploring diffusion timesteps for unsupervised representation learning. arXiv preprint arXiv:2401.11430, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[43] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7579-7590, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[44] Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cyclic stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[45] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[46] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[47] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[48] Zijian Zhang, Zhou Zhao, and Zhijie Lin. Unsupervised representation learning from pre-trained diffusion probabilistic models. Advances in neural information processing systems, 35:22117-22130, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[49] Qingping Zheng, Ling Zheng, Yuanfan Guo, Ying Li, Songcen Xu, Jiankang Deng, and Hang Xu. Self-adaptive reality-guided diffusion for artifact-free super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25806-25816, 2024. 3, 6, 7, 11" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7716" + } + ] +] \ No newline at end of file diff --git a/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_origin.pdf b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e81e4bfc385e317450c39b829ac4f2b3066a27d --- /dev/null +++ b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/b10327a5-6a8f-4c6a-a274-2ff13e185549_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6315562b5710d6f9fa803c5d58369779f5291e24b774d5da3a75a7769f82224 +size 9451611 diff --git a/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/full.md b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..06c044bcedb8861ddabe2fdca11f6c54d6bb9bef --- /dev/null +++ b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/full.md @@ -0,0 +1,306 @@ +# Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts + +Yu Cao Zengqun Zhao Ioannis Patras Shaogang Gong Queen Mary University of London + +{yu.cao, zengqun.zhao, i.patras, s.gong}@qmul.ac.uk + +![](images/59d94c7801df875beae214e5612d7d1693c73af4606abf6fd8e2c286d58898dd.jpg) +Figure 1. Why do diffusion models generate artifacts? We discover that a diffusion generative process necessarily undergoes three phases, we call them: (2) "Profiling" which recovers holistic mean templates, (2) "Mutation" which introduces local divergence, and (3) "Refinement" which rationalizes pixel-wise generation in spatial context. Four visual examples are shown: The first two rows are two examples of rational local mutations (in green boxes) either naturally integrated (Row 1) or reasonably eliminated (Row 2). The bottom two rows show two failure cases when mutations were trapped unreasonably (in red boxes), resisting refinement and resulting in artifacts. Phases are visualized in equal intervals for clarity; please zoom in for more details. + +# Abstract + +Visual artifacts remain a persistent challenge in diffusion models, even with training on massive datasets. Current solutions primarily rely on supervised detectors, yet lack understanding of why these artifacts occur in the first place. In our analysis, we identify three distinct phases in the diffusion generative process: Profiling, Mutation, and Refinement. Artifacts typically emerge during the Mutation phase, where certain regions exhibit anomalous score dynamics over time, causing abrupt disruptions in the normal evolution pattern. This temporal nature explains why existing methods focusing only on spatial uncertainty of the final output fail at effective artifact localization. Based on these insights, we propose ASCED (Abnormal Score Correction for Enhancing Diffusion), that detects artifacts by monitoring abnormal score dynamics during the diffusion process, + +with a trajectory-aware on-the-fly mitigation strategy that appropriate generation of noise in the detected areas. Unlike most existing methods that apply post hoc corrections, e.g., by applying a noising-denoising scheme after generation, our mitigation strategy operates seamlessly within the existing diffusion process. Extensive experiments demonstrate that our proposed approach effectively reduces artifacts across diverse domains, matching or surpassing existing supervised methods without additional training. Project page: YuCao16.github.io/ASCED. + +# 1. Introduction + +Diffusion models have emerged as powerful foundation models in computer vision [5], achieving remarkable success in image generation [7, 15-17, 32], image inpainting [24, 27, 35], and text-to-image task [30-32]. However, even trained on large-scale datasets, diffusion generative images + +![](images/955a49b3382a0aa4220b5107495c061ef7d2f48a58147b8cd50222d307e0a843.jpg) +Figure 2. Diagram of our framework. Denoising and Noising are using Eq. (5) and Eq. (1), respectively. + +still exhibit two significant flaws: visual artifacts and hallucinations [43, 47]. Visual artifacts appear as local irregularities in texture or structure, while hallucinations involve semantically incoherent content, e.g., extra limbs or misplaced objects. In this work, we focus on addressing diffusion artifacts, which present a fundamental challenge to achieving reliable and high-quality image generation. + +Existing methods primarily treat visual artifact detection as a classification problem, i.e., identifying problematic generations for filtering or reconstruction. These methods typically rely on a specialized classifier, either trained on manually annotated artifact datasets [43] or leveraging a pre-trained Large Multi-Modal Model (LMM) [22]. However, such post hoc interventions fail to address a fundamental problem: Why and when do artifacts emerge in diffusion models? To bridge this gap, we begin by examining the diffusion generation process itself. + +We discover that while diffusion process is guided by the same fundamental equation across time (i.e. diffusion steps), in practice, the model exhibits different behavior that can be roughly categorized in three different temporal phases that we name Profiling, Mutation, and Refinement. In the "Profiling" phase, the model sketches the basic semantic global layout; in the "Mutation" phase it explores potential local pixel-wise variations to create local structure; in the "Refinement" phase it attempts to resolve these local pixel-wise variations into coherent visual details in context (see Fig. 1 for visual examples and Sec. 3.2 for a detailed analysis). This understanding of the generation process reveals that while visual artifacts may appear randomly, they follow systematic and identifiable temporal patterns during image formation. Recent uncertainty-based approaches try to identify artifacts by converting diffusion models into Bayesian networks and employ techniques such as Last Layer Laplace Approximation [9] to generate pixel-level variance matrices [20]. However, these uncertainty quantification analyses only capture spatial variations in the final output, i.e., $Var(x_0)$ , missing crucial temporal dynamics during the generation process. Our study shows that diffusion artifacts emerge when certain image regions ex + +hibit abnormal evolution patterns over time, primarily during the Mutation phase of the generation process. Specifically, their pixel values stop updating properly while the surrounding areas continue to evolve. This phenomenon, which we formally define as "score traps" in Sec. 3.2, explains why examining only the final output is insufficient and misplaced for effective diffusion artifact detection. + +Building on these insights, we propose ASCED: Abnormal Score Correction for Enhancing Diffusion, as shown in Fig. 2. At the heart of the method is the estimation of a score, in this context the direction and magnitude of pixel-wise evolution at each diffusion generative step, and a scheme that analyses its temporal dynamics and detects abnormalities. We show both theoretically and experimentally that these abnormalities strongly correlate with artifact formation, making early detection possible at a stage where intervention is still feasible, before artifacts become irreversibly embedded in the generation process. We leverage this early detection capability by implementing a novel trajectory-aware correction mechanism that disrupts the evolution of artifact regions while preserving overall generation diversity. Importantly, ASCED operates in a fully unsupervised manner without requiring manual annotations or domain-specific training, making it readily applicable across various domains, , particularly valuable when training data may be limited or protected. + +Our contributions are: (1) We provide new insights into the formation of visual artifacts in the diffusion generative process, advancing the understanding of diffusion model internal mechanisms. (2) We introduce a novel method that detects potential artifact regions by monitoring abnormal score dynamics temporally, without any manually labeled training required. (3) We further develop a on-the-fly trajectory-aware correction mechanism that effectively mitigates artifacts while preserving image diversity. + +# 2. Related Works + +Visual Artifact Detection initially targeted superresolution artifacts, where upsampling operations are the + +main source [45]. These methods analyze either spatial domain characteristics to capture texture differences between real and generated images [23, 41], or frequency domain patterns to study artifact characteristics in high-frequency components [12, 13]. More recent work has shifted focus to detecting artifacts in general image generation, developing specialized classifiers trained on manually annotated datasets [43] or utilizing pre-trained large vision models [22]. However, these supervised approaches require extensive manual labeling and may not generalize well across different domains. A parallel direction explores uncertainty quantification methods to understand visual artifacts. While various approaches including variational inference [4, 18], Laplace approximation [26, 29], and Markov Chain Monte Carlo [38, 44] have been developed, their application to diffusion models remains limited [20]. BayesDiff [20] pioneers pixel-level uncertainty quantification in diffusion models using Last-layer Laplace approximation [9], yet the connection between spatial uncertainty and visual artifacts remains unclear. + +Generation Quality Enhancement Various approaches have been proposed to enhance generation quality. The truncation trick in BigGAN [6] demonstrated that restricting the sampling space can significantly improve generation fidelity, suggesting similar principles might apply to diffusion models. For diffusion-based generation, classifier guidance [11] has been introduced to steer the generation process. SARGD [49] extends this idea by utilizing a pre-trained artifact detector to guide the generation towards artifact-free regions. Latent diffusion models [30] take a different approach by applying diffusion in a learned latent space, demonstrating that controlled evolution in a constrained space can lead to higher-quality output. + +Diffusion Model for Representation Learning has evolved toward latent space disentanglement and controllable editing. The former aims to uncover interpretable factors in the generative process. Recent studies [28, 42, 48] observed stage-wise attribute emergence during generation, but focus differently than our analysis of artifact formation mechanisms. Controllable editing techniques [27, 31, 37, 46] can be applied for artifact removal, yet require per-sample manipulation and address symptoms rather than causes. Our approach instead corrects abnormal score dynamics during the generation process itself. + +# 3. Methodology + +Our approach consists two steps: Detection (Sec. 3.2) and Correction (Sec. 3.3). Specifically, we localize artifact pixel regions by identifying abnormal score dynamics during the diffusion inference process (Detection) and develop a novel artifact correction algorithm without delaying inference (Correction). We give a theoretical analysis of our key concepts on score trap and temporal weighting (Sec. 3.4). + +![](images/2578a3144b316f48b968042ff566a11a6bb441a980ef785ba77daef9fcff41f5.jpg) +Figure 3. Artifact generation through denoising (top) and brush stroke noising via SDEdit [27] (bottom), demonstrating the model's inability to distinguish artifacts during generation. + +# 3.1. Preliminaries + +Diffusion Model Let $x_0 \in \mathbb{R}^{c \times h \times w}$ be an image. The forward process of a diffusion model gradually diffuses the data distribution $q(x)$ towards $q_t(x_t)$ , $\forall t \in [0, T]$ , with $q_T(x_T) = \mathcal{N}(\mathbf{0}, \mathbf{I})$ as a trivial Gaussian distribution. From a score viewpoint, it can be described by the Stochastic Differential Equation (SDE): + +$$ +d \mathbf {x} _ {t} = \boldsymbol {f} \left(\mathbf {x} _ {t}, t\right) d t + \boldsymbol {g} \left(\mathbf {x} _ {t}, t\right) d \mathbf {w} _ {t}, \quad t \in [ 0, T ] \tag {1} +$$ + +where $\mathbf{w}$ is the standard Wiener process, $f(\cdot)$ and $g(\cdot)$ are scalar drift and diffusion coefficients, respectively. Anderson [2] states that the reverse process of Eq. (1) is also a diffusion process as: + +$$ +d \mathbf {x} _ {s} = \left[ \boldsymbol {f} \left(\mathbf {x} _ {s}, s\right) - \boldsymbol {g} \left(x _ {s}, s\right) ^ {2} \boldsymbol {s} \left(x _ {s}, s\right) \right] d s + \boldsymbol {g} \left(x _ {s}, s\right) d \mathbf {w} _ {s} \tag {2} +$$ + +where $\mathbf{x}_s\coloneqq x_{T - t}$ and $s(x_s,s)\coloneqq \nabla_{\mathbf{x}_s}\log p_s(\mathbf{x}_s)$ is the score function of the marginal distribution over $x_{s}$ . Song et al. [35] leverage this property to generate samples by first drawing $x_{T}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and then solving the reverse SDE using a learned score network $s_\theta (x_t,t)$ . + +By using Tweddie's formula [36], DDPM (Denoising Diffusion Probabilistic Model) [16] can be shown as an equivalent interpretation of Eq. (2) [25]: + +$$ +\boldsymbol {s} _ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {t}, t\right) = - \frac {1}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {t}, t\right) \tag {3} +$$ + +where $\bar{\alpha}_t = \prod_{s=1}^t \alpha_s$ , with mean coefficient $\alpha_t$ , and $\epsilon_\theta(x_t, t)$ is the noise network of DDPM. + +Definition of Visual Artifacts Generation flaws can be distinguished into two categories: Visual Artifacts and Hallucinations. Visual artifacts manifest as local irregularities or distortions in a generated image, such as blurred patches, unnatural textures, broken structures. In contrast, hallucinations refer to semantically generating incoherent content, such as extra limbs, misplaced objects or counterfactuals. In this paper, we focus specifically on detecting and correcting visual artifacts generated by diffusion models. + +# 3.2. Detection by Anomalous Score Dynamic + +To understand how visual artifacts emerge during generations, we examine diffusion model behavior through the + +![](images/7d28fff023b602ab945dcd1c2ce5a123cfb2890e63d8a1b965e8dc8814f48185.jpg) +(a) + +![](images/3328b4909277dcf45ed4a9c602152700358119f4f97d1a6ccd5acb24e3637071.jpg) +(b) + +![](images/10e0ac09e55719c100680a96a696615a53cd25df56e7a485ee27e3a61bbef44e.jpg) +(c) +Figure 4. Visualization of score dynamics and visual artifact detection. (a) Generated images with detected visual artifact regions highlighted (red). (b) Visualization of score dynamics (normalized) between adjacent time steps as activation maps. Brighter regions (green to yellow) indicate areas of higher score variation, while darker regions (blue to black) show areas of lower score change. (c). Score acceleration curves (representing the rate of change in score dynamics between consecutive timesteps) comparing artifact regions (red) with non-artifact regions (blue). The artifact regions exhibit characteristic rapid acceleration followed by deceleration, while non-artifact regions maintain stable score dynamics over time throughout a generative (inference) process. + +lens of image editing. SDEdit [27] demonstrates that diffusion models can transform irregularities, e.g., brush strokes, into semantically meaningful content through a noise-then-denoise process, revealing the inherent Refinement capability of diffusion models. We observe that such irregularities after noising become structurally indistinguishable from states containing artifacts during generation, as shown in Fig. 3. While diffusion models can successfully refine noised brush strokes, they fail to correct the corresponding artifacts during generation. This contrast reveals that diffusion models lack the ability to identify artifacts as patterns requiring refinement during the generation process. + +To better understand this, we examine the generation process from a score perspective, as it directly represents the evolution of pixel values [35]. We define score dynamics as the difference between temporally adjacent score values: $\Delta s_{\theta}(x_t^{i,j},t) = s_{\theta}(x_t^{i,j},t) - s_{\theta}(x_{t - 1}^{i,j},t - 1)$ . Analysis reveals that image generation begins with establishing basic structures, followed by a phase of stochastic exploration where irregular patterns may emerge; we call these phases Profiling and Mutation, respectively. As shown in Fig. 4, regions containing visual artifacts exhibit characteristic patterns during mutation: They appear as localized regions of intense score variations (Fig. 4 (b)) and display dramatic acceleration followed by sudden deceleration in their score trajectories (Fig. 4 (c)). In contrast, normal regions maintain a stable evolution throughout generation. + +Based on these observations, we propose a novel ap + +proach to detecting and localizing potential artifact regions over time in a diffusion inference process. Specifically, let $\Omega \subset \mathbb{R}^2$ denote the spatial domain of the image, and $\Omega_t^a \subset \Omega$ represent regions where abnormal evolution patterns emerge at timestep $t$ . For each spatial location $(i,j) \in \Omega$ , we track the score dynamics through consecutive timesteps. To account for the varying score magnitudes across different images and timesteps, we maintain a score bank $\mathcal{S} = \{s_\theta(x_k,k)\}_{k=t}^T$ and apply a temporal weighting function $w(t)$ that addresses the inherent decay of score magnitudes. Formally, we define artifact regions as: + +$$ +\Omega_ {t} ^ {a} := \left\{(i, j) \in \Omega \mid \left| \Delta \left(w (t) \cdot s _ {\theta} \left(x _ {t} ^ {i, j}, t\right)\right) \right| > \tau \right\} \tag {4} +$$ + +where $w(t) = \frac{1 - \bar{\alpha}_t}{\sqrt{\bar{\alpha}_t}}$ (see Sec. 3.4 for theoretical analysis) and $\tau$ is adaptively determined as the maximum between the Median Absolute Deviation (MAD) of the weighted score dynamics and the mean of score bank $S$ . The final artifact regions $\Omega^a$ are accumulated across the score bank, with detailed procedures provided in Algorithm 1. + +# 3.3. Real-Time Correction + +After detecting artifacts, we need a correction strategy to effectively refine the artifact region. Two natural approaches emerge: either to revert to the states before the artifacts emerge through state replacement or to limit directly abnormal score changes through score clipping. For state replacement, we first predict the clean image from an earlier timestep $t$ using [16]: + +Algorithm 1 Pseudo Code for Proposed ASCED Method +1: Input: Score network $s_{\theta}(\cdot)$ which requires $T$ steps to generate, detection starting step $T_d$ and correction step $T_c$ +2: Initialize $x_T \sim \mathcal{N}(0, \mathbf{I})$ , Score Bank $S \gets \{\}$ , Visual Artifact Mask $\Omega^a \gets \{\}$ +3: for $t = T, t$ -, while $t >= 0$ do +4: $x_{t-1} = \sqrt{\alpha_{t-1}} x_0 - \sqrt{1 - \alpha_{t-1}} \sqrt{1 - \bar{\alpha}_t} s_\theta(x_t, t)$ , where $x_0 = \frac{x_t + (1 - \bar{\alpha}_t) s_\theta(x_t, t)}{\sqrt{\bar{\alpha}_t}}$ ▷ Original Diffusion Process +5: if $T_c < t <= T_d$ then +6: S.append $s_\theta(x_t, t)$ ▷ Store score value into Score Bank +7: else if $t == T_c$ then ▷ Anomalous Score Dynamics Detection Step +8: for $k = 0, k++$ , while $k < (T_c - T_d)$ do ▷ Determine $\Omega^a$ by accumulation +9: $\Omega^a = \Omega^a \cup \{(i, j) \in \Omega | |\Delta(w(k) \cdot s_\theta(x_k^{i,j}, k))| > \tau\} ▷ \tau = \max\{\mathrm{MAD}(\Delta(w(k) \cdot s_\theta(x_k^{i,j}, k)), \mathrm{mean}(\mathcal{S})\}$ +10: $x_t = x_t \cdot \mathbb{1}_{\overline{\Omega}^a} + (\sqrt{\bar{\alpha}_t} \hat{x}_0(t) + \sqrt{1 - \bar{\alpha}_t} \epsilon) \cdot \gamma(t) \xi_{\Omega^a}$ ▷ Trajectory-aware Targeted Correction Step +11: return $x_0$ + +$$ +\hat {\boldsymbol {x}} _ {0} \left(\boldsymbol {x} _ {t}, t\right) = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\boldsymbol {x} _ {t} + \left(1 - \bar {\alpha} _ {t}\right) \nabla_ {\theta} \log p \left(\boldsymbol {x} _ {t}\right)\right), \tag {5} +$$ + +Then replacing the artifact regions with corresponding states from this predicted clean image after re-noising to the current timestep. Score clipping directly limits the magnitude of score changes during inference. However, both state replacement and score clipping fundamentally disrupt the mutation process, leading to reduced generation diversity. To address this problem, we propose Trajectory-aware Targeted Correction (TTC), which introduces controlled perturbations specifically in artifact regions at correction timestep $T_{c}$ : + +$$ +\hat {\boldsymbol {x}} _ {T _ {c}} = \boldsymbol {x} _ {T _ {c}} \cdot \mathbb {1} _ {\overline {{\Omega}} ^ {a}} + \left(\sqrt {\bar {\alpha} _ {T _ {c}}} \boldsymbol {x} _ {0} ^ {\prime} + \sqrt {1 - \bar {\alpha} _ {T _ {c}}} \epsilon\right) \cdot \gamma \xi \mathbb {1} _ {\Omega^ {a}} \tag {6} +$$ + +where $\pmb{x}_0^{\prime} = \hat{\pmb{x}}_0(\pmb{x}_{T_c},T_c)$ $\epsilon ,\xi \stackrel {\mathrm{iid}}{\sim}\mathcal{N}(0,\mathbf{I}),\mathbb{1}_{\Omega^a}$ is indicator function for artifact regions and perturbation intensity $\gamma$ + +TTC builds upon the understanding of score traps: Regions where pixels become locked in persistent score patterns after experiencing dramatic score changes. Through controlled perturbations, TTC disrupts these fixed patterns and allows pixels to resume normal evolution with surrounding regions. Sec. 3.4 provides a detailed analysis of these score trap mechanisms and their relationship to visual artifacts. Generation quality and diversity comparison across correction strategies is shown in Fig. 5. + +# 3.4. Theoretical Analysis + +We provide a theoretical analysis for the score trap and the choice of the temporal weighting function in Eq. (4). + +Theoretical Analysis of Score Traps For normal generation, the score evolution of each pixel is coupled with its neighborhood through the learned neural score function [8]: + +$$ +s _ {\theta} \left(x _ {t} ^ {i, j}, t\right) = \nabla_ {x _ {t} ^ {i, j}} \log p _ {\theta} \left(x _ {t} ^ {i, j} \mid \mathcal {C} (i, j), t\right) \tag {7} +$$ + +where $\mathcal{C}(i,j)$ represents the contextual information from neighboring pixels. This coupling ensures coordinated evolution toward the data manifold. When regions experience abnormal score dynamics, they can enter score traps + +where local patterns persist despite significant score values, disrupting the natural coupled evolution process. These trapped regions evolve based primarily on their local patterns, losing the contextual relationships necessary for coherent image generation. + +This reveals how our perturbation-based correction re-establishes contextual relationships. For a trapped pixel $(i,j)$ , the perturbation $\gamma \cdot \xi$ introduces stochastic variations that disrupt the isolated score patterns, creating opportunities for these regions to re-couple with their surroundings through the natural score evolution process. Meanwhile, in areas without abnormal patterns, these modest perturbations preserve the original coupled evolution, ensuring the method remains harmless to non-artifact regions; see Sec. 7.2 for further mathematical derivations and proofs. + +Score Normalization The score function learned by diffusion models can be interpreted as a vector field guiding the denoising trajectory, inducing a probability flow [34]: + +$$ +\frac {d \boldsymbol {x}}{d t} = - \frac {1}{2} \sigma_ {t} ^ {2} \nabla \log p _ {t} (\boldsymbol {x} _ {t}) \tag {8} +$$ + +The temporal evolution of this probability flow can be characterized by its divergence in the probability density field. Theoretically, we can model this through a flow operator $\mathcal{G}$ : + +$$ +\mathcal {G} (t) = \nabla \cdot \left(\frac {\partial}{\partial t} \int_ {\tau = 0} ^ {t} \mathcal {P} (\boldsymbol {x} _ {\tau}, \tau) d \tau\right) \tag {9} +$$ + +where $\mathcal{P}(\pmb{x}_{\tau},\tau)$ denotes the local probability density at position $\pmb{x}_{\tau}$ and time $\tau$ . This formulation allows us to monitor the accumulation of probability density changes over time. We observe that the clean image prediction (Eq. (5)) reflects these changes in the probability flow. Under the assumption of smooth probability density evolution between adjacent time steps [19, 33, 34], score dynamics are captured through $\frac{\partial}{\partial t}\hat{\pmb{x}}_0(t)$ . Consequently, a normalization factor $w(t) = \frac{1 - \alpha_t}{\sqrt{\alpha_t}}$ is derived from the coefficient of the score term in Eq. (5), helping to equalize the scale of score variations throughout the denoising process. + +Table 1. Quantitative Comparisons on five datasets. The methods compared include BayesDiff [20] and SARGD [49], and three baseline methods: State Replacement Score Clipping and PAL [43] + TTC. All methods use DDIM sampling with identical noise seeds to generate 10,000 images per dataset, ensuring each approach modifies the same deterministic trajectories for fair comparison. The best scores are in bold and second best in underline with bold. Sup and UnS denote supervised and unsupervised methods, respectively. + +
MethodsTypeFFHQ[17]ImageNet[10]LSUN-Cat[40]LSUN-Horse[40]LSUN-Bedroom[40]
FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑
Original [34]UnS36.690.6290.49314.680.7390.73422.170.5130.58629.360.5100.64212.960.6270.583
State ReplaceUnS37.090.6350.49514.610.7430.73322.790.5100.58730.360.5020.64212.950.6280.574
Score ClippingUnS36.360.6300.49814.580.7420.73622.120.5150.58529.260.5110.64212.920.6270.585
BayesDiff [20]UnS36.990.6320.49114.530.7430.73022.500.5130.58528.700.5180.63412.880.6250.569
SARGD [49]Sup38.370.6370.46415.340.7310.72722.650.5230.57030.020.5100.62113.820.6390.554
PAL [43] + TTCSup36.350.6240.50014.010.7310.74721.830.5140.58828.680.5190.64612.710.6290.579
ASCED (Ours)UnS36.280.6370.50314.410.7500.73521.910.5150.59327.660.5210.65212.530.6280.590
+ +# 4. Experiments + +Basic setups We conducted experiments on five datasets: FFHQ [17], ImageNet [10], LSUN-Bedroom [40], LSUN-Cat [40], and LSUN-Horse [40]. We employed the Guided Diffusion model framework and pre-trained weights from OpenAI [11] and Segmentation-DDPM [3]. Quantitative evaluations were performed using FID [14], which measures the Fréchet distance between real and generated image distributions, along with Precision and Recall [21], which evaluate sample fidelity and diversity, respectively. + +Implementation details For detecting diffusion artifacts, our approach was compared with LLaVA-v1.5-13B [22] and PAL [43], while artifact removal comparisons were made with BayesDiff [20] and the adapted SARGD [49]. All experiments were performed on NVIDIA A100 / H100 GPUs. We used DDIM [34] to improve inference efficiency with a Number of Function Evaluations (NFE) set to 25. Remark: We demonstrate in Sec. 7.1 that there is no significant correlation between NFE and the generation of artifacts. Full implementation details are provided in Sec. 6. + +# 4.1. Quantitative Comparisons to Existing Methods + +We first evaluate the effectiveness of ASCED in improving generative quality through comparisons with both unsupervised (BayesDiff [20]) and supervised (SARGD [49]) SOTA methods, along with the original diffusion model [11] and two baselines from Sec. 3.3 (state replacement and score clipping). To isolate the effectiveness of our correction method, we also evaluate a hybrid approach combining artifact detector PAL (used in SARGD) with our Trajectory-aware Targeted Correction (TTC). Quantitative results are shown in Tab. 1. Among unsupervised methods, our ASCED demonstrates superior performance across all datasets, consistently achieving better FID and Precision scores while maintaining higher Recall values than BayesDiff and baselines, indicating both improved generation quality and better preservation of diversity. + +Compared to the supervised methods, our proposed ASCED method shows leading performance across most experiments, achieving superior results on FFHQ, LSUN-Horse, and LSUN-Bedroom, while maintaining competitive + +performance on ImageNet and LSUN-Cat. The better performances of PAL and SARGD on these two datasets are due to that they are supervised artifact detectors specifically trained on these datasets [43]. Their advantages are not generalizable to other datasets FFHQ, LSUN-(Horse, Bedroom). In contrast, ASCED as an unsupervised method has generalisable advantages in all domains without dataset specific training, making it more practical and scalable. Additionally, our method demonstrates significant computational efficiency advantages. In our experiments, ASCED detects and corrects artifacts in approximately 0.09 s per image, which is $8.8 \times$ faster than PAL (0.79 s). + +The effectiveness of the correction mechanism (TTC) becomes particularly evident when comparing SARGD with PAL (used in SARGD) + TTC, where the trajectory-aware correction demonstrates significant advantages in preserving generation diversity across all datasets, reflected in consistently higher Recall scores. + +Table 2. Visual Artifact Detection Accuracy Comparison between PAL (Supervised, Sup) [43], LLaVA (Zero-Shot, ZS) [22], and our method (Unsupervised, UnS) on FFHQ [17], ImageNet [10], and LSUN-(Bedroom, Cat, Horse) [40]. + +
MethodTypeFFHQImageNetBedroomCatHorse
PALSup51.4%69.2%52.4%69.8%60.9%
LLaVAZS63.1%91.1%75.9%59.5%72.2%
OursUnS56.7% (-6.4)67.7% (-1.5)65.0% (-10.9)68.3% (-1.5)70.3% (-1.9)
+ +# 4.2. Artifact Detection Performance Analysis + +To validate the accuracy of our method in identifying visual artifacts, we manually selected 200 images for each dataset from the diffusion model outputs, consisting of 100 images with visual artifacts and 100 without. We then evaluated ASCED against zero-shot large multi-modal model LLaVAv1.5-13B [22] and supervised artifact detector PAL [43]. For LLaVA evaluation, we produced 50 different prompts and reported the results for the most effective prompt. Details on prompt generation are provided in Sec. 6.2. The accuracy for both methods is presented in Tab. 2. + +As an unsupervised method, ASCED achieves promising detection performance, maintaining close accuracy to supervised approaches LLaVA and PAL across most datasets. Notably, these methods analyze final generated images, whereas our approach detects artifacts during the generation + +![](images/845a605e1bd99d4147abfef32cc75ba3399655b5baf545499db7ae9a946fc26d.jpg) +Figure 5. Qualitative Comparison of different correction methods. For each example, we show the original output with visual artifacts (left) and zoomed-in views of the artifact regions corrected by different methods (right): SARGD [49], state replacement (Replace), and our trajectory-aware targeted correction (Ours). Rows from top to bottom: FFHQ[17], ImageNet[10], and LSUN-(Cat, Horse, Bedroom)[40]. + +process through score dynamics, enabling early intervention before artifacts fully manifest. However, our method does show limitations in specific cases, such as low-contrast images where subtle abnormalities are difficult to distinguish from normal variations (leading to False Negatives), and instances where the diffusion model successfully rationalizes initially abnormal patterns during refinement (causing False Positives). Representative examples of these cases are illustrated in Fig. 6. + +# 4.3. Qualitative Analysis of Correction Methods + +To better illustrate the advantage of Trajectory-aware Targeted Correction (TTC) over baselines, Fig. 5 shows qualitative comparisons consisting of the original outputs with artifacts, results from state replacement (Sec. 3.3), SARGD [49], and TTC (ours). While all methods can remove artifacts, TTC demonstrates superior detail preservation in corrected regions. Specifically, both state replacement and SARGD tend to converge to similar expressions and local details, constraining natural variations, as they directly modify the generation states. SARGD faces further limitations from its artifact detector being trained on a specific domain [43], affecting its generalization ability. More importantly, by preserving mutation phase operations, our trajectory-aware correction enables diverse yet coherent generations even when correcting the same region. Additional comparison results are provided in Sec. 8.2. + +Do corrections at Non-Artifact Regions harm? As with any detection method, our proposed scheme will inevitably encounter false positives, leading to add perturbations in non-artifact areas. Our experiments demonstrate that ap + +![](images/770b7594e2ebd0a0c051695aa9c521b2b63ac8d0a9c255f1fd8cca56c7efd0c1.jpg) +Figure 6. Top: Applied our correction method to clean regions (yellow box). Bottom: Typical failure cases. + +plying the correction mechanism to regions without visual artifacts, as shown in Fig. 6, introduces modest variations while preserving semantic coherence with the surrounding context and not generating new artifacts. Extended visual results and detection performance analysis can be found in Sec. 7.2 and Sec. 4.2, respectively. + +# 4.4. Further Analysis + +Distribution of Abnormal Score Dynamics In Fig. 7, we plot the frequency of abnormal scores at each time step (normalized by total diffusion time step $T$ ). The score dynamics demonstrate distinct patterns across different stages: remaining stable in early steps where basic structures emerge, experiencing significant variations in the middle stage, and gradually stabilizing again in the later steps during details' refinement. The presence of the long tail indicates that some variations persist into later steps, suggesting an extended period of details' adjustment. This behavioral pattern naturally aligns with our hypothesis that the generation undergoes three phases: profiling, mutation, and rationalization. This temporal + +![](images/3edf117a1a50ba98303720e553090ad512d87af30ad0445005a4c63ae61735e5.jpg) +Figure 7. Temporal Analysis of abnormal score dynamics across FFHQ [17], ImageNet [10], LSUN-(Bedroom, Cat, Horse) [40]. + +pattern suggests that, while early intervention might be possible, determining the latest effective correction point is crucial for maximizing the detection of potential artifacts. + +Impact of Correction Timing We investigate how the choice of correction timestep $T_{c}$ affects artifact removal effectiveness. Through extensive experiments, we identify a threshold at approximately $T_{c}^{*} / T \approx 0.48$ across different diffusion processes, representing the latest viable correction point before the model lacks sufficient steps for refinement. As shown in Fig. 8, both Precision and Recall metrics achieve optimal performance as $T_{c}$ approaches $T_{c}^{*}$ . This optimal timing allows for maximum artifact detection while ensuring adequate refinement steps. Most datasets show stable performance before $T_{c}^{*}$ followed by a sharp decline, while FFHQ exhibits more fluctuation before $T_{c}^{*}$ and decreases gradually afterward. Individual dataset curves with analysis are provided in Sec. 7.4. + +![](images/9405cd4cdc82cd75d04436c3b5a4a61ae23603345a5b8415babb3f31b26506e7.jpg) +Figure 8. Impact of correction timestep $T_{c}$ on artifact removal performance evaluated by Precision (fidelity, $\uparrow$ ) and Recall (diversity, $\uparrow$ ). The dashed lines indicate the baseline precision / recall of the original diffusion model on each dataset. + +![](images/d4aa01c91f6f73558454704392bf23ddf91f6d4b1ecfb58f77a8b90875b262cc.jpg) + +Latent Code Improvement To evaluate how our correction method influences latent representations, we conduct a linear probe experiment [39] using a classifier-guided diffusion model [11] on ImageNet [10]. Specifically, we generate samples following two paths: the original diffusion process and our corrected process. At an intermediate timestep $t$ , we obtain the original state $x_{t}$ and apply our correction + +method to get the corrected state $\hat{x}_t$ . We then continue the diffusion process for $k$ steps to obtain $x_{t - k}$ and $\hat{x}_{t - k}$ from the original and corrected states, respectively. We generate $N$ labeled samples using both paths, with $y^{i}$ denoting the class label, resulting in two sets: + +$$ +\mathcal {D} _ {\text {o r i g}} = \left\{\left(x _ {t - k} ^ {i}, y ^ {i}\right) \right\} _ {i = 1} ^ {N}, \quad \mathcal {D} _ {\text {c o r r}} = \left\{\left(\hat {x} _ {t - k} ^ {i}, y ^ {i}\right) \right\} _ {i = 1} ^ {N} \tag {10} +$$ + +We train two separate classifiers on $\mathcal{D}_{\mathrm{orig}}$ and $\mathcal{D}_{\mathrm{corr}}$ respectively, with results shown in Fig. 9. The higher accuracy achieved by improved latent codes throughout the remaining steps demonstrates that our method enhances the semantic quality of latent representations, and this improvement affects overall accuracy rather than individual precision or recall metrics. Notably, we observe that the classification accuracy reaches its peak earlier in the generation process with our method and remains stable through the refinement phase, while the original process exhibits a decrease-increase pattern during refinement. + +![](images/246bdaab1a498bcd3e41621929fbc9e3f477dc2d8dd049aba64c8e2c1f145ea4.jpg) +Figure 9. Linear probe classification results comparing latent representations from original and corrected diffusion trajectories across Accuracy, Precision and Recall metrics. + +# 5. Conclusion + +We present a novel analysis of the diffusion generation process, decomposing it into profiling, mutation, and refinement phases, which provides fundamental insights into artifact formation mechanisms. Based on these insights, we develop ASCED, an unsupervised framework that successfully detects and corrects artifacts while preserving generation diversity. Extensive experiments demonstrate that ASCED achieves competitive performance with state-of-the-art supervised methods across multiple datasets. The training-free nature of our approach enables immediate application to any diffusion model, making it a practical solution for improving generation quality. + +Future Work While our approach effectively detects artifacts through temporal pattern analysis, promising directions include improving detection in low-contrast regions, developing more robust discrimination between transient and persistent abnormalities, and extending these insights to other generative frameworks. + +# Acknowledgments + +This work was partially supported by Veritone, Adobe, and has utilized Queen Mary's Apocrita HPC facility from QMUL Research-IT. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 11 +[2] Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3):313-326, 1982. 3 +[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 6 +[4] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 3 +[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1 +[6] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018.3 +[7] Yu Cao and Shaogang Gong. Few-shot image generation by conditional relaxing diffusion inversion. arXiv preprint arXiv:2407.07249, 2024. 1 +[8] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 5 +[9] Erik Daxberger, Agustinus Kristiadi, Alexander Immer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace reflux-effortless bayesian deep learning. Advances in Neural Information Processing Systems, 34:20089-20103, 2021. 2, 3 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7, 8 +[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 6, 8 +[12] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890-7899, 2020. 3 +[13] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. + +Advances in neural information processing systems, 33: 3022-3032, 2020. 3 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[15] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2016. 1 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4 +[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 6, 7, 8 +[18] Mohammad Khan, Didrik Nielsen, Voot Tangkaratt, Wu Lin, Yarin Gal, and Akash Srivastava. Fast and scalable bayesian deep learning by weight-perturbation in adam. In International conference on machine learning, pages 2611-2620. PMLR, 2018. 3 +[19] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 5, 14 +[20] Siqi Kou, Lei Gan, Dequan Wang, Chongxuan Li, and Zhijie Deng. Bayesdiff: Estimating pixel-wise uncertainty in diffusion via bayesian inference. arXiv preprint arXiv:2310.11142, 2023. 2, 3, 6 +[21] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 6 +[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 3, 6, 11 +[23] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8060-8069, 2020. 3 +[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 1 +[25] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3, 14 +[26] David John Cameron Mackay. Bayesian methods for adaptive models. California Institute of Technology, 1992. 3 +[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1, 3, 4 + +[28] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3 +[29] Hippolyt Ritter, Aleksandar Botev, and David Barber. A scalable laplace approximation for neural networks. In 6th international conference on learning representations, ICLR 2018-conference track proceedings. International Conference on Representation Learning, 2018. 3 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3 +[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 +[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 5 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5, 6 +[35] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1, 3, 4 +[36] Charles M Stein. Estimation of the mean of a multivariate normal distribution. The annals of Statistics, pages 1135-1151, 1981. 3 +[37] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 3 +[38] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 3 +[39] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. arXiv preprint arXiv:2303.09769, 2023. 8 +[40] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 6, 7, 8 + +[41] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 3 +[42] Zhongqi Yue, Jiankun Wang, Qianru Sun, Lei Ji, Eric I Chang, Hanwang Zhang, et al. Exploring diffusion timesteps for unsupervised representation learning. arXiv preprint arXiv:2401.11430, 2024. 3 +[43] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7579-7590, 2023. 2, 3, 6, 7 +[44] Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cyclic stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932, 2019. 3 +[45] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 3 +[46] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3 +[47] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2 +[48] Zijian Zhang, Zhou Zhao, and Zhijie Lin. Unsupervised representation learning from pre-trained diffusion probabilistic models. Advances in neural information processing systems, 35:22117-22130, 2022. 3 +[49] Qingping Zheng, Ling Zheng, Yuanfan Guo, Ying Li, Songcen Xu, Jiankang Deng, and Hang Xu. Self-adaptive reality-guided diffusion for artifact-free super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25806-25816, 2024. 3, 6, 7, 11 \ No newline at end of file diff --git a/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/images.zip b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4e10d2cc914dddf1df25141b67a442c60911afac --- /dev/null +++ b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e6da8e6fe31a199382c970ff7c33fa24404af4652b42fc4f076b994ecb0b682 +size 814446 diff --git a/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/layout.json b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e76c3ebe465403e0e21f94bb1e24e8377f6e1ad6 --- /dev/null +++ b/2025/Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts/layout.json @@ -0,0 +1,7935 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 549, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 549, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 549, + 121 + ], + "type": "text", + "content": "Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 143, + 467, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 143, + 467, + 172 + ], + "spans": [ + { + "bbox": [ + 143, + 143, + 467, + 172 + ], + "type": "text", + "content": "Yu Cao Zengqun Zhao Ioannis Patras Shaogang Gong Queen Mary University of London" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 165, + 174, + 441, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 174, + 441, + 185 + ], + "spans": [ + { + "bbox": [ + 165, + 174, + 441, + 185 + ], + "type": "text", + "content": "{yu.cao, zengqun.zhao, i.patras, s.gong}@qmul.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 206, + 555, + 426 + ], + "blocks": [ + { + "bbox": [ + 56, + 206, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 56, + 206, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 555, + 426 + ], + "type": "image", + "image_path": "59d94c7801df875beae214e5612d7d1693c73af4606abf6fd8e2c286d58898dd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 431, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 431, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 555, + 498 + ], + "type": "text", + "content": "Figure 1. Why do diffusion models generate artifacts? We discover that a diffusion generative process necessarily undergoes three phases, we call them: (2) \"Profiling\" which recovers holistic mean templates, (2) \"Mutation\" which introduces local divergence, and (3) \"Refinement\" which rationalizes pixel-wise generation in spatial context. Four visual examples are shown: The first two rows are two examples of rational local mutations (in green boxes) either naturally integrated (Row 1) or reasonably eliminated (Row 2). The bottom two rows show two failure cases when mutations were trapped unreasonably (in red boxes), resisting refinement and resulting in artifacts. Phases are visualized in equal intervals for clarity; please zoom in for more details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 508, + 200, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 508, + 200, + 519 + ], + "spans": [ + { + "bbox": [ + 151, + 508, + 200, + 519 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 533, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 297, + 714 + ], + "type": "text", + "content": "Visual artifacts remain a persistent challenge in diffusion models, even with training on massive datasets. Current solutions primarily rely on supervised detectors, yet lack understanding of why these artifacts occur in the first place. In our analysis, we identify three distinct phases in the diffusion generative process: Profiling, Mutation, and Refinement. Artifacts typically emerge during the Mutation phase, where certain regions exhibit anomalous score dynamics over time, causing abrupt disruptions in the normal evolution pattern. This temporal nature explains why existing methods focusing only on spatial uncertainty of the final output fail at effective artifact localization. Based on these insights, we propose ASCED (Abnormal Score Correction for Enhancing Diffusion), that detects artifacts by monitoring abnormal score dynamics during the diffusion process," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "type": "text", + "content": "with a trajectory-aware on-the-fly mitigation strategy that appropriate generation of noise in the detected areas. Unlike most existing methods that apply post hoc corrections, e.g., by applying a noising-denoising scheme after generation, our mitigation strategy operates seamlessly within the existing diffusion process. Extensive experiments demonstrate that our proposed approach effectively reduces artifacts across diverse domains, matching or surpassing existing supervised methods without additional training. Project page: YuCao16.github.io/ASCED." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Diffusion models have emerged as powerful foundation models in computer vision [5], achieving remarkable success in image generation [7, 15-17, 32], image inpainting [24, 27, 35], and text-to-image task [30-32]. However, even trained on large-scale datasets, diffusion generative images" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7707" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 62, + 521, + 215 + ], + "blocks": [ + { + "bbox": [ + 88, + 62, + 521, + 215 + ], + "lines": [ + { + "bbox": [ + 88, + 62, + 521, + 215 + ], + "spans": [ + { + "bbox": [ + 88, + 62, + 521, + 215 + ], + "type": "image", + "image_path": "955a49b3382a0aa4220b5107495c061ef7d2f48a58147b8cd50222d307e0a843.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 217, + 496, + 229 + ], + "lines": [ + { + "bbox": [ + 111, + 217, + 496, + 229 + ], + "spans": [ + { + "bbox": [ + 111, + 217, + 496, + 229 + ], + "type": "text", + "content": "Figure 2. Diagram of our framework. Denoising and Noising are using Eq. (5) and Eq. (1), respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 234, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 234, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 295, + 319 + ], + "type": "text", + "content": "still exhibit two significant flaws: visual artifacts and hallucinations [43, 47]. Visual artifacts appear as local irregularities in texture or structure, while hallucinations involve semantically incoherent content, e.g., extra limbs or misplaced objects. In this work, we focus on addressing diffusion artifacts, which present a fundamental challenge to achieving reliable and high-quality image generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 319, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 295, + 437 + ], + "type": "text", + "content": "Existing methods primarily treat visual artifact detection as a classification problem, i.e., identifying problematic generations for filtering or reconstruction. These methods typically rely on a specialized classifier, either trained on manually annotated artifact datasets [43] or leveraging a pre-trained Large Multi-Modal Model (LMM) [22]. However, such post hoc interventions fail to address a fundamental problem: Why and when do artifacts emerge in diffusion models? To bridge this gap, we begin by examining the diffusion generation process itself." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 438, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 438, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 438, + 295, + 713 + ], + "type": "text", + "content": "We discover that while diffusion process is guided by the same fundamental equation across time (i.e. diffusion steps), in practice, the model exhibits different behavior that can be roughly categorized in three different temporal phases that we name Profiling, Mutation, and Refinement. In the \"Profiling\" phase, the model sketches the basic semantic global layout; in the \"Mutation\" phase it explores potential local pixel-wise variations to create local structure; in the \"Refinement\" phase it attempts to resolve these local pixel-wise variations into coherent visual details in context (see Fig. 1 for visual examples and Sec. 3.2 for a detailed analysis). This understanding of the generation process reveals that while visual artifacts may appear randomly, they follow systematic and identifiable temporal patterns during image formation. Recent uncertainty-based approaches try to identify artifacts by converting diffusion models into Bayesian networks and employ techniques such as Last Layer Laplace Approximation [9] to generate pixel-level variance matrices [20]. However, these uncertainty quantification analyses only capture spatial variations in the final output, i.e., " + }, + { + "bbox": [ + 57, + 438, + 295, + 713 + ], + "type": "inline_equation", + "content": "Var(x_0)" + }, + { + "bbox": [ + 57, + 438, + 295, + 713 + ], + "type": "text", + "content": ", missing crucial temporal dynamics during the generation process. Our study shows that diffusion artifacts emerge when certain image regions ex" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 235, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 319 + ], + "type": "text", + "content": "hibit abnormal evolution patterns over time, primarily during the Mutation phase of the generation process. Specifically, their pixel values stop updating properly while the surrounding areas continue to evolve. This phenomenon, which we formally define as \"score traps\" in Sec. 3.2, explains why examining only the final output is insufficient and misplaced for effective diffusion artifact detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 319, + 553, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 547 + ], + "type": "text", + "content": "Building on these insights, we propose ASCED: Abnormal Score Correction for Enhancing Diffusion, as shown in Fig. 2. At the heart of the method is the estimation of a score, in this context the direction and magnitude of pixel-wise evolution at each diffusion generative step, and a scheme that analyses its temporal dynamics and detects abnormalities. We show both theoretically and experimentally that these abnormalities strongly correlate with artifact formation, making early detection possible at a stage where intervention is still feasible, before artifacts become irreversibly embedded in the generation process. We leverage this early detection capability by implementing a novel trajectory-aware correction mechanism that disrupts the evolution of artifact regions while preserving overall generation diversity. Importantly, ASCED operates in a fully unsupervised manner without requiring manual annotations or domain-specific training, making it readily applicable across various domains, , particularly valuable when training data may be limited or protected." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 548, + 553, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 553, + 655 + ], + "type": "text", + "content": "Our contributions are: (1) We provide new insights into the formation of visual artifacts in the diffusion generative process, advancing the understanding of diffusion model internal mechanisms. (2) We introduce a novel method that detects potential artifact regions by monitoring abnormal score dynamics temporally, without any manually labeled training required. (3) We further develop a on-the-fly trajectory-aware correction mechanism that effectively mitigates artifacts while preserving image diversity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 669, + 405, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 405, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 405, + 681 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Visual Artifact Detection initially targeted superresolution artifacts, where upsampling operations are the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7708" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 310 + ], + "type": "text", + "content": "main source [45]. These methods analyze either spatial domain characteristics to capture texture differences between real and generated images [23, 41], or frequency domain patterns to study artifact characteristics in high-frequency components [12, 13]. More recent work has shifted focus to detecting artifacts in general image generation, developing specialized classifiers trained on manually annotated datasets [43] or utilizing pre-trained large vision models [22]. However, these supervised approaches require extensive manual labeling and may not generalize well across different domains. A parallel direction explores uncertainty quantification methods to understand visual artifacts. While various approaches including variational inference [4, 18], Laplace approximation [26, 29], and Markov Chain Monte Carlo [38, 44] have been developed, their application to diffusion models remains limited [20]. BayesDiff [20] pioneers pixel-level uncertainty quantification in diffusion models using Last-layer Laplace approximation [9], yet the connection between spatial uncertainty and visual artifacts remains unclear." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "text", + "content": "Generation Quality Enhancement Various approaches have been proposed to enhance generation quality. The truncation trick in BigGAN [6] demonstrated that restricting the sampling space can significantly improve generation fidelity, suggesting similar principles might apply to diffusion models. For diffusion-based generation, classifier guidance [11] has been introduced to steer the generation process. SARGD [49] extends this idea by utilizing a pre-trained artifact detector to guide the generation towards artifact-free regions. Latent diffusion models [30] take a different approach by applying diffusion in a learned latent space, demonstrating that controlled evolution in a constrained space can lead to higher-quality output." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 467, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 598 + ], + "type": "text", + "content": "Diffusion Model for Representation Learning has evolved toward latent space disentanglement and controllable editing. The former aims to uncover interpretable factors in the generative process. Recent studies [28, 42, 48] observed stage-wise attribute emergence during generation, but focus differently than our analysis of artifact formation mechanisms. Controllable editing techniques [27, 31, 37, 46] can be applied for artifact removal, yet require per-sample manipulation and address symptoms rather than causes. Our approach instead corrects abnormal score dynamics during the generation process itself." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 609, + 138, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 138, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 138, + 623 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Our approach consists two steps: Detection (Sec. 3.2) and Correction (Sec. 3.3). Specifically, we localize artifact pixel regions by identifying abnormal score dynamics during the diffusion inference process (Detection) and develop a novel artifact correction algorithm without delaying inference (Correction). We give a theoretical analysis of our key concepts on score trap and temporal weighting (Sec. 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 320, + 74, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 320, + 74, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 320, + 74, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 320, + 74, + 545, + 191 + ], + "type": "image", + "image_path": "2578a3144b316f48b968042ff566a11a6bb441a980ef785ba77daef9fcff41f5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 192, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 225 + ], + "type": "text", + "content": "Figure 3. Artifact generation through denoising (top) and brush stroke noising via SDEdit [27] (bottom), demonstrating the model's inability to distinguish artifacts during generation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 228, + 400, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 400, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 400, + 239 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "text", + "content": "Diffusion Model Let " + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "inline_equation", + "content": "x_0 \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "text", + "content": " be an image. The forward process of a diffusion model gradually diffuses the data distribution " + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "text", + "content": " towards " + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "inline_equation", + "content": "q_t(x_t)" + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "inline_equation", + "content": "\\forall t \\in [0, T]" + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "inline_equation", + "content": "q_T(x_T) = \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "text", + "content": " as a trivial Gaussian distribution. From a score viewpoint, it can be described by the Stochastic Differential Equation (SDE):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 337, + 319, + 553, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 319, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 337, + 319, + 553, + 331 + ], + "type": "interline_equation", + "content": "d \\mathbf {x} _ {t} = \\boldsymbol {f} \\left(\\mathbf {x} _ {t}, t\\right) d t + \\boldsymbol {g} \\left(\\mathbf {x} _ {t}, t\\right) d \\mathbf {w} _ {t}, \\quad t \\in [ 0, T ] \\tag {1}", + "image_path": "4f270bc1c6279e060b88f764e4fcd7506f1adc0b8faaab4d49d0bab0a298fac5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 332, + 553, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 553, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 553, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 332, + 553, + 379 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 332, + 553, + 379 + ], + "type": "text", + "content": " is the standard Wiener process, " + }, + { + "bbox": [ + 313, + 332, + 553, + 379 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 332, + 553, + 379 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 332, + 553, + 379 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 332, + 553, + 379 + ], + "type": "text", + "content": " are scalar drift and diffusion coefficients, respectively. Anderson [2] states that the reverse process of Eq. (1) is also a diffusion process as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 380, + 553, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 380, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 318, + 380, + 553, + 404 + ], + "type": "interline_equation", + "content": "d \\mathbf {x} _ {s} = \\left[ \\boldsymbol {f} \\left(\\mathbf {x} _ {s}, s\\right) - \\boldsymbol {g} \\left(x _ {s}, s\\right) ^ {2} \\boldsymbol {s} \\left(x _ {s}, s\\right) \\right] d s + \\boldsymbol {g} \\left(x _ {s}, s\\right) d \\mathbf {w} _ {s} \\tag {2}", + "image_path": "8e9cfaa20b200b02404d3c4494f276347aa68198a005843626686700d0fc4ef4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_s\\coloneqq x_{T - t}" + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "inline_equation", + "content": "s(x_s,s)\\coloneqq \\nabla_{\\mathbf{x}_s}\\log p_s(\\mathbf{x}_s)" + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "content": " is the score function of the marginal distribution over " + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "content": ". Song et al. [35] leverage this property to generate samples by first drawing " + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "inline_equation", + "content": "x_{T}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "content": " and then solving the reverse SDE using a learned score network " + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "inline_equation", + "content": "s_\\theta (x_t,t)" + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 464, + 553, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 553, + 500 + ], + "type": "text", + "content": "By using Tweddie's formula [36], DDPM (Denoising Diffusion Probabilistic Model) [16] can be shown as an equivalent interpretation of Eq. (2) [25]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 500, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 500, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 365, + 500, + 553, + 525 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t}, t\\right) = - \\frac {1}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t}, t\\right) \\tag {3}", + "image_path": "c639d750e4114a7fb4a05d09a382ead643d3d1daa51bded3fa8a677d4ee0c340.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s" + }, + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "text", + "content": ", with mean coefficient " + }, + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta(x_t, t)" + }, + { + "bbox": [ + 313, + 522, + 553, + 545 + ], + "type": "text", + "content": " is the noise network of DDPM." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 665 + ], + "type": "text", + "content": "Definition of Visual Artifacts Generation flaws can be distinguished into two categories: Visual Artifacts and Hallucinations. Visual artifacts manifest as local irregularities or distortions in a generated image, such as blurred patches, unnatural textures, broken structures. In contrast, hallucinations refer to semantically generating incoherent content, such as extra limbs, misplaced objects or counterfactuals. In this paper, we focus specifically on detecting and correcting visual artifacts generated by diffusion models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 671, + 525, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 525, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 525, + 684 + ], + "type": "text", + "content": "3.2. Detection by Anomalous Score Dynamic" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "To understand how visual artifacts emerge during generations, we examine diffusion model behavior through the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7709" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 233, + 284 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 233, + 284 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 233, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 233, + 284 + ], + "type": "image", + "image_path": "7d28fff023b602ab945dcd1c2ce5a123cfb2890e63d8a1b965e8dc8814f48185.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 285, + 130, + 295 + ], + "lines": [ + { + "bbox": [ + 117, + 285, + 130, + 295 + ], + "spans": [ + { + "bbox": [ + 117, + 285, + 130, + 295 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 239, + 72, + 359, + 284 + ], + "blocks": [ + { + "bbox": [ + 239, + 72, + 359, + 284 + ], + "lines": [ + { + "bbox": [ + 239, + 72, + 359, + 284 + ], + "spans": [ + { + "bbox": [ + 239, + 72, + 359, + 284 + ], + "type": "image", + "image_path": "3328b4909277dcf45ed4a9c602152700358119f4f97d1a6ccd5acb24e3637071.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 285, + 301, + 295 + ], + "lines": [ + { + "bbox": [ + 290, + 285, + 301, + 295 + ], + "spans": [ + { + "bbox": [ + 290, + 285, + 301, + 295 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 365, + 72, + 550, + 284 + ], + "blocks": [ + { + "bbox": [ + 365, + 72, + 550, + 284 + ], + "lines": [ + { + "bbox": [ + 365, + 72, + 550, + 284 + ], + "spans": [ + { + "bbox": [ + 365, + 72, + 550, + 284 + ], + "type": "image", + "image_path": "10e0ac09e55719c100680a96a696615a53cd25df56e7a485ee27e3a61bbef44e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 285, + 492, + 295 + ], + "lines": [ + { + "bbox": [ + 481, + 285, + 492, + 295 + ], + "spans": [ + { + "bbox": [ + 481, + 285, + 492, + 295 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 297, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 555, + 363 + ], + "type": "text", + "content": "Figure 4. Visualization of score dynamics and visual artifact detection. (a) Generated images with detected visual artifact regions highlighted (red). (b) Visualization of score dynamics (normalized) between adjacent time steps as activation maps. Brighter regions (green to yellow) indicate areas of higher score variation, while darker regions (blue to black) show areas of lower score change. (c). Score acceleration curves (representing the rate of change in score dynamics between consecutive timesteps) comparing artifact regions (red) with non-artifact regions (blue). The artifact regions exhibit characteristic rapid acceleration followed by deceleration, while non-artifact regions maintain stable score dynamics over time throughout a generative (inference) process." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 373, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 517 + ], + "type": "text", + "content": "lens of image editing. SDEdit [27] demonstrates that diffusion models can transform irregularities, e.g., brush strokes, into semantically meaningful content through a noise-then-denoise process, revealing the inherent Refinement capability of diffusion models. We observe that such irregularities after noising become structurally indistinguishable from states containing artifacts during generation, as shown in Fig. 3. While diffusion models can successfully refine noised brush strokes, they fail to correct the corresponding artifacts during generation. This contrast reveals that diffusion models lack the ability to identify artifacts as patterns requiring refinement during the generation process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 520, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 699 + ], + "type": "text", + "content": "To better understand this, we examine the generation process from a score perspective, as it directly represents the evolution of pixel values [35]. We define score dynamics as the difference between temporally adjacent score values: " + }, + { + "bbox": [ + 55, + 520, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\Delta s_{\\theta}(x_t^{i,j},t) = s_{\\theta}(x_t^{i,j},t) - s_{\\theta}(x_{t - 1}^{i,j},t - 1)" + }, + { + "bbox": [ + 55, + 520, + 295, + 699 + ], + "type": "text", + "content": ". Analysis reveals that image generation begins with establishing basic structures, followed by a phase of stochastic exploration where irregular patterns may emerge; we call these phases Profiling and Mutation, respectively. As shown in Fig. 4, regions containing visual artifacts exhibit characteristic patterns during mutation: They appear as localized regions of intense score variations (Fig. 4 (b)) and display dramatic acceleration followed by sudden deceleration in their score trajectories (Fig. 4 (c)). In contrast, normal regions maintain a stable evolution throughout generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Based on these observations, we propose a novel ap" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": "proach to detecting and localizing potential artifact regions over time in a diffusion inference process. Specifically, let " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " denote the spatial domain of the image, and " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\Omega_t^a \\subset \\Omega" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " represent regions where abnormal evolution patterns emerge at timestep " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": ". For each spatial location " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "(i,j) \\in \\Omega" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": ", we track the score dynamics through consecutive timesteps. To account for the varying score magnitudes across different images and timesteps, we maintain a score bank " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{s_\\theta(x_k,k)\\}_{k=t}^T" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " and apply a temporal weighting function " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " that addresses the inherent decay of score magnitudes. Formally, we define artifact regions as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 507, + 555, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 507, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 328, + 507, + 555, + 529 + ], + "type": "interline_equation", + "content": "\\Omega_ {t} ^ {a} := \\left\\{(i, j) \\in \\Omega \\mid \\left| \\Delta \\left(w (t) \\cdot s _ {\\theta} \\left(x _ {t} ^ {i, j}, t\\right)\\right) \\right| > \\tau \\right\\} \\tag {4}", + "image_path": "139d05111026ede148cc6812dd836f6a76b7eca51084a6d57fab96dc75d11768.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "inline_equation", + "content": "w(t) = \\frac{1 - \\bar{\\alpha}_t}{\\sqrt{\\bar{\\alpha}_t}}" + }, + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "text", + "content": " (see Sec. 3.4 for theoretical analysis) and " + }, + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "text", + "content": " is adaptively determined as the maximum between the Median Absolute Deviation (MAD) of the weighted score dynamics and the mean of score bank " + }, + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "text", + "content": ". The final artifact regions " + }, + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\Omega^a" + }, + { + "bbox": [ + 313, + 530, + 555, + 605 + ], + "type": "text", + "content": " are accumulated across the score bank, with detailed procedures provided in Algorithm 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 440, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 440, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 440, + 623 + ], + "type": "text", + "content": "3.3. Real-Time Correction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "After detecting artifacts, we need a correction strategy to effectively refine the artifact region. Two natural approaches emerge: either to revert to the states before the artifacts emerge through state replacement or to limit directly abnormal score changes through score clipping. For state replacement, we first predict the clean image from an earlier timestep " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " using [16]:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7710" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 86, + 555, + 235 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 288, + 84 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 288, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 288, + 84 + ], + "type": "text", + "content": "Algorithm 1 Pseudo Code for Proposed ASCED Method" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "lines": [ + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": "1: Input: Score network " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "s_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " which requires " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " steps to generate, detection starting step " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "T_d" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " and correction step " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": "\n2: Initialize " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "x_T \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": ", Score Bank " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "S \\gets \\{\\}" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": ", Visual Artifact Mask " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "\\Omega^a \\gets \\{\\}" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": "\n3: for " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "t = T, t" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": "-, while " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "t >= 0" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " do\n4: " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "x_{t-1} = \\sqrt{\\alpha_{t-1}} x_0 - \\sqrt{1 - \\alpha_{t-1}} \\sqrt{1 - \\bar{\\alpha}_t} s_\\theta(x_t, t)" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "x_0 = \\frac{x_t + (1 - \\bar{\\alpha}_t) s_\\theta(x_t, t)}{\\sqrt{\\bar{\\alpha}_t}}" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " ▷ Original Diffusion Process\n5: if " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "T_c < t <= T_d" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " then\n6: S.append " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "s_\\theta(x_t, t)" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " ▷ Store score value into Score Bank\n7: else if " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "t == T_c" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " then ▷ Anomalous Score Dynamics Detection Step\n8: for " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "k = 0, k++" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "k < (T_c - T_d)" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " do ▷ Determine " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "\\Omega^a" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " by accumulation\n9: " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "\\Omega^a = \\Omega^a \\cup \\{(i, j) \\in \\Omega | |\\Delta(w(k) \\cdot s_\\theta(x_k^{i,j}, k))| > \\tau\\} ▷ \\tau = \\max\\{\\mathrm{MAD}(\\Delta(w(k) \\cdot s_\\theta(x_k^{i,j}, k)), \\mathrm{mean}(\\mathcal{S})\\}" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": "\n10: " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "x_t = x_t \\cdot \\mathbb{1}_{\\overline{\\Omega}^a} + (\\sqrt{\\bar{\\alpha}_t} \\hat{x}_0(t) + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon) \\cdot \\gamma(t) \\xi_{\\Omega^a}" + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "text", + "content": " ▷ Trajectory-aware Targeted Correction Step\n11: return " + }, + { + "bbox": [ + 56, + 86, + 555, + 235 + ], + "type": "inline_equation", + "content": "x_0" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 70, + 255, + 294, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 255, + 294, + 279 + ], + "spans": [ + { + "bbox": [ + 70, + 255, + 294, + 279 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {0} \\left(\\boldsymbol {x} _ {t}, t\\right) = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {x} _ {t} + \\left(1 - \\bar {\\alpha} _ {t}\\right) \\nabla_ {\\theta} \\log p \\left(\\boldsymbol {x} _ {t}\\right)\\right), \\tag {5}", + "image_path": "7d1e75743f288dc0da8dcc98f43378de43620c8b488bd2cfc668f571440e4c59.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 278, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 399 + ], + "type": "text", + "content": "Then replacing the artifact regions with corresponding states from this predicted clean image after re-noising to the current timestep. Score clipping directly limits the magnitude of score changes during inference. However, both state replacement and score clipping fundamentally disrupt the mutation process, leading to reduced generation diversity. To address this problem, we propose Trajectory-aware Targeted Correction (TTC), which introduces controlled perturbations specifically in artifact regions at correction timestep " + }, + { + "bbox": [ + 55, + 278, + 296, + 399 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 55, + 278, + 296, + 399 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 402, + 295, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 402, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 60, + 402, + 295, + 418 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {T _ {c}} = \\boldsymbol {x} _ {T _ {c}} \\cdot \\mathbb {1} _ {\\overline {{\\Omega}} ^ {a}} + \\left(\\sqrt {\\bar {\\alpha} _ {T _ {c}}} \\boldsymbol {x} _ {0} ^ {\\prime} + \\sqrt {1 - \\bar {\\alpha} _ {T _ {c}}} \\epsilon\\right) \\cdot \\gamma \\xi \\mathbb {1} _ {\\Omega^ {a}} \\tag {6}", + "image_path": "f39b532278562d933b8e91dde6cbfc0b6eeb06cfc4a964f529869e1268e21f80.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 423, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{\\prime} = \\hat{\\pmb{x}}_0(\\pmb{x}_{T_c},T_c)" + }, + { + "bbox": [ + 55, + 423, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\epsilon ,\\xi \\stackrel {\\mathrm{iid}}{\\sim}\\mathcal{N}(0,\\mathbf{I}),\\mathbb{1}_{\\Omega^a}" + }, + { + "bbox": [ + 55, + 423, + 295, + 450 + ], + "type": "text", + "content": " is indicator function for artifact regions and perturbation intensity " + }, + { + "bbox": [ + 55, + 423, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 558 + ], + "type": "text", + "content": "TTC builds upon the understanding of score traps: Regions where pixels become locked in persistent score patterns after experiencing dramatic score changes. Through controlled perturbations, TTC disrupts these fixed patterns and allows pixels to resume normal evolution with surrounding regions. Sec. 3.4 provides a detailed analysis of these score trap mechanisms and their relationship to visual artifacts. Generation quality and diversity comparison across correction strategies is shown in Fig. 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "type": "text", + "content": "3.4. Theoretical Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "text", + "content": "We provide a theoretical analysis for the score trap and the choice of the temporal weighting function in Eq. (4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "content": "Theoretical Analysis of Score Traps For normal generation, the score evolution of each pixel is coupled with its neighborhood through the learned neural score function [8]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 93, + 646, + 295, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 646, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 93, + 646, + 295, + 663 + ], + "type": "interline_equation", + "content": "s _ {\\theta} \\left(x _ {t} ^ {i, j}, t\\right) = \\nabla_ {x _ {t} ^ {i, j}} \\log p _ {\\theta} \\left(x _ {t} ^ {i, j} \\mid \\mathcal {C} (i, j), t\\right) \\tag {7}", + "image_path": "73dbd03b1e32816bf41e7b2531e6f65261829b842e8f9db4cec4a195afbcc711.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(i,j)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " represents the contextual information from neighboring pixels. This coupling ensures coordinated evolution toward the data manifold. When regions experience abnormal score dynamics, they can enter score traps" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 257, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 317 + ], + "type": "text", + "content": "where local patterns persist despite significant score values, disrupting the natural coupled evolution process. These trapped regions evolve based primarily on their local patterns, losing the contextual relationships necessary for coherent image generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 317, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 554, + 437 + ], + "type": "text", + "content": "This reveals how our perturbation-based correction re-establishes contextual relationships. For a trapped pixel " + }, + { + "bbox": [ + 313, + 317, + 554, + 437 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 317, + 554, + 437 + ], + "type": "text", + "content": ", the perturbation " + }, + { + "bbox": [ + 313, + 317, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\gamma \\cdot \\xi" + }, + { + "bbox": [ + 313, + 317, + 554, + 437 + ], + "type": "text", + "content": " introduces stochastic variations that disrupt the isolated score patterns, creating opportunities for these regions to re-couple with their surroundings through the natural score evolution process. Meanwhile, in areas without abnormal patterns, these modest perturbations preserve the original coupled evolution, ensuring the method remains harmless to non-artifact regions; see Sec. 7.2 for further mathematical derivations and proofs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 448, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 485 + ], + "type": "text", + "content": "Score Normalization The score function learned by diffusion models can be interpreted as a vector field guiding the denoising trajectory, inducing a probability flow [34]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 380, + 486, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 486, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 380, + 486, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\frac {d \\boldsymbol {x}}{d t} = - \\frac {1}{2} \\sigma_ {t} ^ {2} \\nabla \\log p _ {t} (\\boldsymbol {x} _ {t}) \\tag {8}", + "image_path": "901423eeee5c4c302e5a1c97a04a92bec653d93c7f37b0ca32fb71435f372f74.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 553, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 546 + ], + "type": "text", + "content": "The temporal evolution of this probability flow can be characterized by its divergence in the probability density field. Theoretically, we can model this through a flow operator " + }, + { + "bbox": [ + 313, + 510, + 553, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 510, + 553, + 546 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 549, + 553, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 549, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 359, + 549, + 553, + 577 + ], + "type": "interline_equation", + "content": "\\mathcal {G} (t) = \\nabla \\cdot \\left(\\frac {\\partial}{\\partial t} \\int_ {\\tau = 0} ^ {t} \\mathcal {P} (\\boldsymbol {x} _ {\\tau}, \\tau) d \\tau\\right) \\tag {9}", + "image_path": "d69e4432382e755a3c58dfebf9129b14f0aa9ec7127003424ede483069689dad.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\pmb{x}_{\\tau},\\tau)" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " denotes the local probability density at position " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\tau}" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": ". This formulation allows us to monitor the accumulation of probability density changes over time. We observe that the clean image prediction (Eq. (5)) reflects these changes in the probability flow. Under the assumption of smooth probability density evolution between adjacent time steps [19, 33, 34], score dynamics are captured through " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\partial}{\\partial t}\\hat{\\pmb{x}}_0(t)" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": ". Consequently, a normalization factor " + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "inline_equation", + "content": "w(t) = \\frac{1 - \\alpha_t}{\\sqrt{\\alpha_t}}" + }, + { + "bbox": [ + 313, + 579, + 555, + 714 + ], + "type": "text", + "content": " is derived from the coefficient of the score term in Eq. (5), helping to equalize the scale of score variations throughout the denoising process." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7711" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 104, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 60, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 60, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 60, + 555, + 104 + ], + "type": "text", + "content": "Table 1. Quantitative Comparisons on five datasets. The methods compared include BayesDiff [20] and SARGD [49], and three baseline methods: State Replacement Score Clipping and PAL [43] + TTC. All methods use DDIM sampling with identical noise seeds to generate 10,000 images per dataset, ensuring each approach modifies the same deterministic trajectories for fair comparison. The best scores are in bold and second best in underline with bold. Sup and UnS denote supervised and unsupervised methods, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 104, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 104, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 104, + 553, + 198 + ], + "type": "table", + "html": "
MethodsTypeFFHQ[17]ImageNet[10]LSUN-Cat[40]LSUN-Horse[40]LSUN-Bedroom[40]
FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑FID ↓Pre. ↑Rec. ↑
Original [34]UnS36.690.6290.49314.680.7390.73422.170.5130.58629.360.5100.64212.960.6270.583
State ReplaceUnS37.090.6350.49514.610.7430.73322.790.5100.58730.360.5020.64212.950.6280.574
Score ClippingUnS36.360.6300.49814.580.7420.73622.120.5150.58529.260.5110.64212.920.6270.585
BayesDiff [20]UnS36.990.6320.49114.530.7430.73022.500.5130.58528.700.5180.63412.880.6250.569
SARGD [49]Sup38.370.6370.46415.340.7310.72722.650.5230.57030.020.5100.62113.820.6390.554
PAL [43] + TTCSup36.350.6240.50014.010.7310.74721.830.5140.58828.680.5190.64612.710.6290.579
ASCED (Ours)UnS36.280.6370.50314.410.7500.73521.910.5150.59327.660.5210.65212.530.6280.590
", + "image_path": "46b1769201298ba8042aa821e114420d710bc42fac64833f1881a4b56b462316.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 203, + 136, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 203, + 136, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 136, + 217 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 224, + 294, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 294, + 332 + ], + "type": "text", + "content": "Basic setups We conducted experiments on five datasets: FFHQ [17], ImageNet [10], LSUN-Bedroom [40], LSUN-Cat [40], and LSUN-Horse [40]. We employed the Guided Diffusion model framework and pre-trained weights from OpenAI [11] and Segmentation-DDPM [3]. Quantitative evaluations were performed using FID [14], which measures the Fréchet distance between real and generated image distributions, along with Precision and Recall [21], which evaluate sample fidelity and diversity, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 334, + 294, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 294, + 453 + ], + "type": "text", + "content": "Implementation details For detecting diffusion artifacts, our approach was compared with LLaVA-v1.5-13B [22] and PAL [43], while artifact removal comparisons were made with BayesDiff [20] and the adapted SARGD [49]. All experiments were performed on NVIDIA A100 / H100 GPUs. We used DDIM [34] to improve inference efficiency with a Number of Function Evaluations (NFE) set to 25. Remark: We demonstrate in Sec. 7.1 that there is no significant correlation between NFE and the generation of artifacts. Full implementation details are provided in Sec. 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 465, + 295, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 295, + 478 + ], + "type": "text", + "content": "4.1. Quantitative Comparisons to Existing Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 484, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 663 + ], + "type": "text", + "content": "We first evaluate the effectiveness of ASCED in improving generative quality through comparisons with both unsupervised (BayesDiff [20]) and supervised (SARGD [49]) SOTA methods, along with the original diffusion model [11] and two baselines from Sec. 3.3 (state replacement and score clipping). To isolate the effectiveness of our correction method, we also evaluate a hybrid approach combining artifact detector PAL (used in SARGD) with our Trajectory-aware Targeted Correction (TTC). Quantitative results are shown in Tab. 1. Among unsupervised methods, our ASCED demonstrates superior performance across all datasets, consistently achieving better FID and Precision scores while maintaining higher Recall values than BayesDiff and baselines, indicating both improved generation quality and better preservation of diversity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Compared to the supervised methods, our proposed ASCED method shows leading performance across most experiments, achieving superior results on FFHQ, LSUN-Horse, and LSUN-Bedroom, while maintaining competitive" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 205, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 347 + ], + "type": "text", + "content": "performance on ImageNet and LSUN-Cat. The better performances of PAL and SARGD on these two datasets are due to that they are supervised artifact detectors specifically trained on these datasets [43]. Their advantages are not generalizable to other datasets FFHQ, LSUN-(Horse, Bedroom). In contrast, ASCED as an unsupervised method has generalisable advantages in all domains without dataset specific training, making it more practical and scalable. Additionally, our method demonstrates significant computational efficiency advantages. In our experiments, ASCED detects and corrects artifacts in approximately 0.09 s per image, which is " + }, + { + "bbox": [ + 313, + 205, + 553, + 347 + ], + "type": "inline_equation", + "content": "8.8 \\times" + }, + { + "bbox": [ + 313, + 205, + 553, + 347 + ], + "type": "text", + "content": " faster than PAL (0.79 s)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 348, + 553, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 420 + ], + "type": "text", + "content": "The effectiveness of the correction mechanism (TTC) becomes particularly evident when comparing SARGD with PAL (used in SARGD) + TTC, where the trajectory-aware correction demonstrates significant advantages in preserving generation diversity across all datasets, reflected in consistently higher Recall scores." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 471, + 553, + 509 + ], + "blocks": [ + { + "bbox": [ + 313, + 426, + 553, + 470 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 470 + ], + "type": "text", + "content": "Table 2. Visual Artifact Detection Accuracy Comparison between PAL (Supervised, Sup) [43], LLaVA (Zero-Shot, ZS) [22], and our method (Unsupervised, UnS) on FFHQ [17], ImageNet [10], and LSUN-(Bedroom, Cat, Horse) [40]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 471, + 553, + 509 + ], + "lines": [ + { + "bbox": [ + 315, + 471, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 315, + 471, + 553, + 509 + ], + "type": "table", + "html": "
MethodTypeFFHQImageNetBedroomCatHorse
PALSup51.4%69.2%52.4%69.8%60.9%
LLaVAZS63.1%91.1%75.9%59.5%72.2%
OursUnS56.7% (-6.4)67.7% (-1.5)65.0% (-10.9)68.3% (-1.5)70.3% (-1.9)
", + "image_path": "b446e5a34c0dabd9abce04091b2d2ccae0be0c631036ce20c4d4cd4e8f9026cc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 515, + 526, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 526, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 526, + 528 + ], + "type": "text", + "content": "4.2. Artifact Detection Performance Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 653 + ], + "type": "text", + "content": "To validate the accuracy of our method in identifying visual artifacts, we manually selected 200 images for each dataset from the diffusion model outputs, consisting of 100 images with visual artifacts and 100 without. We then evaluated ASCED against zero-shot large multi-modal model LLaVAv1.5-13B [22] and supervised artifact detector PAL [43]. For LLaVA evaluation, we produced 50 different prompts and reported the results for the most effective prompt. Details on prompt generation are provided in Sec. 6.2. The accuracy for both methods is presented in Tab. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "As an unsupervised method, ASCED achieves promising detection performance, maintaining close accuracy to supervised approaches LLaVA and PAL across most datasets. Notably, these methods analyze final generated images, whereas our approach detects artifacts during the generation" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7712" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 60, + 555, + 292 + ], + "blocks": [ + { + "bbox": [ + 56, + 60, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 56, + 60, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 56, + 60, + 555, + 292 + ], + "type": "image", + "image_path": "845a605e1bd99d4147abfef32cc75ba3399655b5baf545499db7ae9a946fc26d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 293, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 327 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison of different correction methods. For each example, we show the original output with visual artifacts (left) and zoomed-in views of the artifact regions corrected by different methods (right): SARGD [49], state replacement (Replace), and our trajectory-aware targeted correction (Ours). Rows from top to bottom: FFHQ[17], ImageNet[10], and LSUN-(Cat, Horse, Bedroom)[40]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 337, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 445 + ], + "type": "text", + "content": "process through score dynamics, enabling early intervention before artifacts fully manifest. However, our method does show limitations in specific cases, such as low-contrast images where subtle abnormalities are difficult to distinguish from normal variations (leading to False Negatives), and instances where the diffusion model successfully rationalizes initially abnormal patterns during refinement (causing False Positives). Representative examples of these cases are illustrated in Fig. 6." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 454, + 282, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 282, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 282, + 467 + ], + "type": "text", + "content": "4.3. Qualitative Analysis of Correction Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 664 + ], + "type": "text", + "content": "To better illustrate the advantage of Trajectory-aware Targeted Correction (TTC) over baselines, Fig. 5 shows qualitative comparisons consisting of the original outputs with artifacts, results from state replacement (Sec. 3.3), SARGD [49], and TTC (ours). While all methods can remove artifacts, TTC demonstrates superior detail preservation in corrected regions. Specifically, both state replacement and SARGD tend to converge to similar expressions and local details, constraining natural variations, as they directly modify the generation states. SARGD faces further limitations from its artifact detector being trained on a specific domain [43], affecting its generalization ability. More importantly, by preserving mutation phase operations, our trajectory-aware correction enables diverse yet coherent generations even when correcting the same region. Additional comparison results are provided in Sec. 8.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Do corrections at Non-Artifact Regions harm? As with any detection method, our proposed scheme will inevitably encounter false positives, leading to add perturbations in non-artifact areas. Our experiments demonstrate that ap" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 335, + 553, + 432 + ], + "blocks": [ + { + "bbox": [ + 315, + 335, + 553, + 432 + ], + "lines": [ + { + "bbox": [ + 315, + 335, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 553, + 432 + ], + "type": "image", + "image_path": "770b7594e2ebd0a0c051695aa9c521b2b63ac8d0a9c255f1fd8cca56c7efd0c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 433, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 455 + ], + "type": "text", + "content": "Figure 6. Top: Applied our correction method to clean regions (yellow box). Bottom: Typical failure cases." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "content": "plying the correction mechanism to regions without visual artifacts, as shown in Fig. 6, introduces modest variations while preserving semantic coherence with the surrounding context and not generating new artifacts. Extended visual results and detection performance analysis can be found in Sec. 7.2 and Sec. 4.2, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 540, + 416, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 416, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 416, + 552 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "Distribution of Abnormal Score Dynamics In Fig. 7, we plot the frequency of abnormal scores at each time step (normalized by total diffusion time step " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "). The score dynamics demonstrate distinct patterns across different stages: remaining stable in early steps where basic structures emerge, experiencing significant variations in the middle stage, and gradually stabilizing again in the later steps during details' refinement. The presence of the long tail indicates that some variations persist into later steps, suggesting an extended period of details' adjustment. This behavioral pattern naturally aligns with our hypothesis that the generation undergoes three phases: profiling, mutation, and rationalization. This temporal" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7713" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 290, + 212 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 290, + 212 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 290, + 212 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 290, + 212 + ], + "type": "image", + "image_path": "3edf117a1a50ba98303720e553090ad512d87af30ad0445005a4c63ae61735e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "type": "text", + "content": "Figure 7. Temporal Analysis of abnormal score dynamics across FFHQ [17], ImageNet [10], LSUN-(Bedroom, Cat, Horse) [40]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 242, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 278 + ], + "type": "text", + "content": "pattern suggests that, while early intervention might be possible, determining the latest effective correction point is crucial for maximizing the detection of potential artifacts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": "Impact of Correction Timing We investigate how the choice of correction timestep " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": " affects artifact removal effectiveness. Through extensive experiments, we identify a threshold at approximately " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "T_{c}^{*} / T \\approx 0.48" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": " across different diffusion processes, representing the latest viable correction point before the model lacks sufficient steps for refinement. As shown in Fig. 8, both Precision and Recall metrics achieve optimal performance as " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": " approaches " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "T_{c}^{*}" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": ". This optimal timing allows for maximum artifact detection while ensuring adequate refinement steps. Most datasets show stable performance before " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "T_{c}^{*}" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": " followed by a sharp decline, while FFHQ exhibits more fluctuation before " + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "inline_equation", + "content": "T_{c}^{*}" + }, + { + "bbox": [ + 55, + 289, + 296, + 456 + ], + "type": "text", + "content": " and decreases gradually afterward. Individual dataset curves with analysis are provided in Sec. 7.4." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 459, + 175, + 576 + ], + "blocks": [ + { + "bbox": [ + 56, + 459, + 175, + 576 + ], + "lines": [ + { + "bbox": [ + 56, + 459, + 175, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 175, + 576 + ], + "type": "image", + "image_path": "9405cd4cdc82cd75d04436c3b5a4a61ae23603345a5b8415babb3f31b26506e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 578, + 295, + 622 + ], + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 622 + ], + "type": "text", + "content": "Figure 8. Impact of correction timestep " + }, + { + "bbox": [ + 55, + 578, + 295, + 622 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 55, + 578, + 295, + 622 + ], + "type": "text", + "content": " on artifact removal performance evaluated by Precision (fidelity, " + }, + { + "bbox": [ + 55, + 578, + 295, + 622 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 578, + 295, + 622 + ], + "type": "text", + "content": ") and Recall (diversity, " + }, + { + "bbox": [ + 55, + 578, + 295, + 622 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 578, + 295, + 622 + ], + "type": "text", + "content": "). The dashed lines indicate the baseline precision / recall of the original diffusion model on each dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 459, + 293, + 576 + ], + "blocks": [ + { + "bbox": [ + 175, + 459, + 293, + 576 + ], + "lines": [ + { + "bbox": [ + 175, + 459, + 293, + 576 + ], + "spans": [ + { + "bbox": [ + 175, + 459, + 293, + 576 + ], + "type": "image", + "image_path": "d4aa01c91f6f73558454704392bf23ddf91f6d4b1ecfb58f77a8b90875b262cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Latent Code Improvement To evaluate how our correction method influences latent representations, we conduct a linear probe experiment [39] using a classifier-guided diffusion model [11] on ImageNet [10]. Specifically, we generate samples following two paths: the original diffusion process and our corrected process. At an intermediate timestep " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", we obtain the original state " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " and apply our correction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "method to get the corrected state " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ". We then continue the diffusion process for " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " steps to obtain " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "x_{t - k}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t - k}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " from the original and corrected states, respectively. We generate " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " labeled samples using both paths, with " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "y^{i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " denoting the class label, resulting in two sets:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 138, + 555, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 138, + 555, + 154 + ], + "spans": [ + { + "bbox": [ + 319, + 138, + 555, + 154 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\text {o r i g}} = \\left\\{\\left(x _ {t - k} ^ {i}, y ^ {i}\\right) \\right\\} _ {i = 1} ^ {N}, \\quad \\mathcal {D} _ {\\text {c o r r}} = \\left\\{\\left(\\hat {x} _ {t - k} ^ {i}, y ^ {i}\\right) \\right\\} _ {i = 1} ^ {N} \\tag {10}", + "image_path": "770188c266847fb6945a4415da434d9cb3c7a53131fd40453ee736218ae6763e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "text", + "content": "We train two separate classifiers on " + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{orig}}" + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{corr}}" + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "text", + "content": " respectively, with results shown in Fig. 9. The higher accuracy achieved by improved latent codes throughout the remaining steps demonstrates that our method enhances the semantic quality of latent representations, and this improvement affects overall accuracy rather than individual precision or recall metrics. Notably, we observe that the classification accuracy reaches its peak earlier in the generation process with our method and remains stable through the refinement phase, while the original process exhibits a decrease-increase pattern during refinement." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 294, + 552, + 433 + ], + "blocks": [ + { + "bbox": [ + 317, + 294, + 552, + 433 + ], + "lines": [ + { + "bbox": [ + 317, + 294, + 552, + 433 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 552, + 433 + ], + "type": "image", + "image_path": "246bdaab1a498bcd3e41621929fbc9e3f477dc2d8dd049aba64c8e2c1f145ea4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 433, + 554, + 465 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 554, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 554, + 465 + ], + "type": "text", + "content": "Figure 9. Linear probe classification results comparing latent representations from original and corrected diffusion trajectories across Accuracy, Precision and Recall metrics." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 641 + ], + "type": "text", + "content": "We present a novel analysis of the diffusion generation process, decomposing it into profiling, mutation, and refinement phases, which provides fundamental insights into artifact formation mechanisms. Based on these insights, we develop ASCED, an unsupervised framework that successfully detects and corrects artifacts while preserving generation diversity. Extensive experiments demonstrate that ASCED achieves competitive performance with state-of-the-art supervised methods across multiple datasets. The training-free nature of our approach enables immediate application to any diffusion model, making it a practical solution for improving generation quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "Future Work While our approach effectively detects artifacts through temporal pattern analysis, promising directions include improving detection in low-contrast regions, developing more robust discrimination between transient and persistent abnormalities, and extending these insights to other generative frameworks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7714" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This work was partially supported by Veritone, Adobe, and has utilized Queen Mary's Apocrita HPC facility from QMUL Research-IT." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 212 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 213, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 296, + 245 + ], + "type": "text", + "content": "[2] Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3):313-326, 1982. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 246, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 296, + 290 + ], + "type": "text", + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 334 + ], + "type": "text", + "content": "[4] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 296, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 390 + ], + "type": "text", + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 391, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 296, + 423 + ], + "type": "text", + "content": "[6] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 424, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 296, + 456 + ], + "type": "text", + "content": "[7] Yu Cao and Shaogang Gong. Few-shot image generation by conditional relaxing diffusion inversion. arXiv preprint arXiv:2407.07249, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "type": "text", + "content": "[8] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 502, + 296, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 556 + ], + "type": "text", + "content": "[9] Erik Daxberger, Agustinus Kristiadi, Alexander Immer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace reflux-effortless bayesian deep learning. Advances in Neural Information Processing Systems, 34:20089-20103, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 602 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 296, + 635 + ], + "type": "text", + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 691 + ], + "type": "text", + "content": "[12] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890-7899, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 715 + ], + "type": "text", + "content": "[13] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "Advances in neural information processing systems, 33: 3022-3032, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "text", + "content": "[15] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2016. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "type": "text", + "content": "[18] Mohammad Khan, Didrik Nielsen, Voot Tangkaratt, Wu Lin, Yarin Gal, and Akash Srivastava. Fast and scalable bayesian deep learning by weight-perturbation in adam. In International conference on machine learning, pages 2611-2620. PMLR, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 349, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 382 + ], + "type": "text", + "content": "[19] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 5, 14" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[20] Siqi Kou, Lei Gan, Dequan Wang, Chongxuan Li, and Zhijie Deng. Bayesdiff: Estimating pixel-wise uncertainty in diffusion via bayesian inference. arXiv preprint arXiv:2310.11142, 2023. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 426, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 471 + ], + "type": "text", + "content": "[21] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 525 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 3, 6, 11" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 525, + 553, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 570 + ], + "type": "text", + "content": "[23] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8060-8069, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "text", + "content": "[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 647 + ], + "type": "text", + "content": "[25] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3, 14" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "text", + "content": "[26] David John Cameron Mackay. Bayesian methods for adaptive models. California Institute of Technology, 1992. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 1, 3, 4" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7715" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[28] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 139, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 139, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 139, + 294, + 194 + ], + "type": "text", + "content": "[29] Hippolyt Ritter, Aleksandar Botev, and David Barber. A scalable laplace approximation for neural networks. In 6th international conference on learning representations, ICLR 2018-conference track proceedings. International Conference on Representation Learning, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 315 + ], + "type": "text", + "content": "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 381 + ], + "type": "text", + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "type": "text", + "content": "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[35] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 546 + ], + "type": "text", + "content": "[36] Charles M Stein. Estimation of the mean of a multivariate normal distribution. The annals of Statistics, pages 1135-1151, 1981. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[37] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22532-22541, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[38] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688. CiteSeer, 2011. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 637, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 669 + ], + "type": "text", + "content": "[39] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. arXiv preprint arXiv:2303.09769, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[40] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 6, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 540 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[41] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[42] Zhongqi Yue, Jiankun Wang, Qianru Sun, Lei Ji, Eric I Chang, Hanwang Zhang, et al. Exploring diffusion timesteps for unsupervised representation learning. arXiv preprint arXiv:2401.11430, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "text", + "content": "[43] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7579-7590, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 229, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 273 + ], + "type": "text", + "content": "[44] Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cyclic stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "text", + "content": "[45] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[46] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "type": "text", + "content": "[47] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 431, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 474 + ], + "type": "text", + "content": "[48] Zijian Zhang, Zhou Zhao, and Zhijie Lin. Unsupervised representation learning from pre-trained diffusion probabilistic models. Advances in neural information processing systems, 35:22117-22130, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 475, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 540 + ], + "type": "text", + "content": "[49] Qingping Zheng, Ling Zheng, Yuanfan Guo, Ying Li, Songcen Xu, Jiankang Deng, and Hang Xu. Self-adaptive reality-guided diffusion for artifact-free super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25806-25816, 2024. 3, 6, 7, 11" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7716" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/511efbd9-e72e-4675-94ed-5687d3419119_content_list.json b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/511efbd9-e72e-4675-94ed-5687d3419119_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1adcf92c149ae292ef55c383add2b6fc7a06b21e --- /dev/null +++ b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/511efbd9-e72e-4675-94ed-5687d3419119_content_list.json @@ -0,0 +1,1534 @@ +[ + { + "type": "text", + "text": "Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks", + "text_level": 1, + "bbox": [ + 96, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kairong $\\mathrm{Yu}^{1}$ Chengting $\\mathrm{Yu}^{1}$ Tianqing Zhang $^{1}$ Xiaochen Zhao $^{1}$", + "bbox": [ + 192, + 200, + 802, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shu Yang $^{1}$ Hongwei Wang $^{1,\\dagger}$ Qiang Zhang $^{2}$ Qi Xu $^{2,\\dagger}$", + "bbox": [ + 258, + 220, + 784, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University $^{2}$ Dalian University of Technology", + "bbox": [ + 259, + 243, + 738, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kairong.22,chengting.21,xiaochen.24，shu.23,hongweiwang}@intlzju.edu.cn", + "bbox": [ + 158, + 265, + 828, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhangtianqing@zju.edu.cn", + "bbox": [ + 387, + 284, + 601, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xuqi, zhangq}@dlut.edu.cn", + "bbox": [ + 377, + 300, + 606, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 349, + 326, + 366 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs), inspired by the human brain, offer significant computational efficiency through discrete spike-based information transfer. Despite their potential to reduce inference energy consumption, a performance gap persists between SNNs and Artificial Neural Networks (ANNs), primarily due to current training methods and inherent model limitations. While recent research has aimed to enhance SNN learning by employing knowledge distillation (KD) from ANN teacher networks, traditional distillation techniques often overlook the distinctive spatiotemporal properties of SNNs, thus failing to fully leverage their advantages. To overcome these challenges, we propose a novel logit distillation method characterized by temporal separation and entropy regularization. This approach improves existing SNN distillation techniques by performing distillation learning on logits across different time steps, rather than merely on aggregated output features. Furthermore, the integration of entropy regularization stabilizes model optimization and further boosts the performance. Extensive experimental results indicate that our method surpasses prior SNN distillation strategies, whether based on logit distillation, feature distillation, or a combination of both. Our project is available at https://github.com/yukairong/TSER.", + "bbox": [ + 86, + 382, + 483, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 775, + 220, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the neural firing mechanisms observed in biological systems, Spiking Neural Networks (SNNs) [37] are considered a promising alternative to traditional Artificial Neural Networks (ANNs) due to their superior energy efficiency. Unlike ANNs, which use continuous activation val", + "bbox": [ + 89, + 800, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ues for information transmission[31], SNNs transmit and process information through discrete spike events [6, 38], with neurons generating spikes only when their membrane potential exceeds a threshold. This binary, event-driven approach allows SNNs to run efficiently on neuromorphic hardware [6, 42], accumulating synaptic inputs effectively and avoiding unnecessary computations related to zero input or activation [9, 14]. Given their event-driven dynamics and the biomimetic properties of spatiotemporal neuron activity [44, 45], SNNs demonstrate remarkable energy efficiency, robust adaptive learning capabilities [13, 41, 58], and ultra-low power consumption, showing significant potential for computational intelligence applications tasks [44]. Despite the inherent advantages of SNNs, their performance in common tasks such as image classification [27], object segmentation [43], and natural language processing [23] still lags behind that of ANNs. This gap largely stems from the limitations of current training methods and structural constraints in SNNs. In contrast to ANNs, SNNs cannot directly utilize backpropagation (BP) for deep network training. Moreover, attempts to directly adapt ANN methodologies to SNNs encounter compatibility issues, preventing SNNs from fully realizing their theoretical advantages.", + "bbox": [ + 511, + 351, + 906, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, conversion-based and learning-based approaches are two common methods for training SNNs. The former approaches seek to utilize knowledge from ANNs by transferring the parameters of a pre-trained ANN to an equivalent SNN [1, 10, 33, 34]. These methods often necessitate a significant number of time steps to achieve accuracy levels comparable to those of the original ANN [2, 46]. If the number of inference time steps is reduced, the model's capacity for effective information transfer may diminish, resulting in a decline in overall performance. To improve the performance of SNNs and minimize inference time steps, learning-based approaches have been employed", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 886, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8806", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fbdd8e30957d4d16d454d025dd355f671b9266c41d592598e579aaba60c4b2d1.jpg", + "image_caption": [ + "(a) Classical SNN Knowledge Distillation" + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 486, + 296 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c4fa6d85a843ee577e7c852e82135e0a4d402f096d610a8aabaf3c7ac875c27c.jpg", + "image_caption": [ + "(b) Temporal Separation with Entropy Regularization (TSER) Knowledge Distillation", + "Figure 1. Illustration of the Classical SNN KD and Our Proposed Method. In our approach, we remove the temporal dimension fusion operation present in classical SNN KD and apply a temporal separation strategy to focus the learning process on outputs at individual time steps. The original calculations of KL Loss and CE Loss are adapted to compute the loss for each time step's output, followed by averaging. Additionally, we incorporate an entropy regularization term to guide the learning direction away from erroneous knowledge. Here, T, B, and C represent the time steps, batch size, and number of classes, respectively." + ], + "image_footnote": [], + "bbox": [ + 501, + 95, + 890, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to facilitate training during inference. These approaches leverage Backpropagation Through Time (BPTT) for effective training [48, 57]. Motivated by the work [25], several studies have utilized surrogate gradient estimation as a method to address the non-differentiability inherent in SNNs. This foundational research has enabled the direct training of CNNs and transformer-based SNNs, achieving strong performance on static datasets with reduced inference time steps. To further reduce the performance gap with ANNs, recent work has explored the inclusion of modules such as attention mechanisms to enhance SNN capabilities [28, 40, 52, 53]. While effective, these additional modules increase computational consumption, counteracting the inherent low-energy benefits of SNNs. Knowledge Distillation (KD) provides a potential solution, where SNNs are guided by a more capable ANN teacher model, addressing the inference burden associated with these modules. Prior works such as KDSNN [50] and LaSNN [24] employ hierarchical knowledge distillation [5] to facilitate the training of SNNs by leveraging the rich semantic information from the ANN teacher model. Recent approaches like BKDSNN [51] move beyond traditional layer-wise distillation by integrating blurred knowledge distillation techniques with logit distillation, demonstrating significant improvements in performance across both CNNs and Transformer-based models.", + "bbox": [ + 91, + 422, + 483, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In preliminary experiments, we observed that current SNN distillation architectures are not optimally designed for the unique structure of SNNs. As depicted in Fig. 1a, conventional SNN KD techniques distill the mean of outputs across time steps. However, the inherent nature", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of SNN processing implies that significant information is embedded within the temporal dimension. Employing a mean approach for fusion prediction during classification resembles a voting selection process. Therefore, distributing the distillation of outputs across various time steps during the training phase can substantially alleviate the limitations that currently hinder distillation performance. To mitigate the impact of outliers on network performance and to rectify overly confident incorrect semantics produced by the teacher network, we introduce an entropy regularization term as a constraint. We conduct extensive experimental analyses to assess its efficacy. Furthermore, we evaluate our proposed method on several datasets, including CIFAR10 [26], CIFAR100 [26], and ImageNet [8]. The results of our experiments demonstrate that this method achieves state-of-the-art (SOTA) performance. In summary, our contributions are threefold:", + "bbox": [ + 511, + 422, + 906, + 679 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a temporal separation strategy into the SNN distillation method, effectively addressing the incompatibility issues present in previous SNN distillation frameworks and yielding a more effective universal distillation architecture.", + "- We integrate entropy regularization to correct the overly confident errors from the teacher network, thereby maximizing the performance of the logit-based distillation method.", + "- We conduct comprehensive evaluations across multiple datasets, showing that our proposed approach outperforms previous state-of-the-art (SOTA) methods." + ], + "bbox": [ + 513, + 715, + 903, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8807", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conversion-based Methods. To fully leverage the knowledge acquired from pre-trained ANNs, conversion methods directly transfer the learned parameters from an ANN to the corresponding SNN, avoiding the difficulties of training SNNs from scratch. Several studies [19, 20] replace the activation functions in ANNs with spiking neurons and introduce optimizations like weight normalization [12] to generate corresponding SNNs. The core idea of these methods is to closely align the outputs of the ANN and SNN, maximizing the utilization of the pre-trained ANN knowledge. While later works [5, 10, 12] achieve nearly lossless accuracy conversion, they often involve long inference times and fail to retain the spatiotemporal characteristics inherent to spike-based processing, significantly limiting the practical applicability of converted SNNs.", + "bbox": [ + 89, + 114, + 483, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based Methods. The development of surrogate gradient training methods [56] has resolved the challenge of training non-differentiable SNNs directly. Initially, [39] used a surrogate gradient approach to construct deep SNN models, achieving competitive accuracy on neuromorphic datasets. The DCT-SNN [16] applied frequency domain techniques to reduce the number of inference time steps. To support deeper SNN training, [15] introduced SEW ResNet to address gradient vanishing and exploding issues in Spiking ResNet [60], extending SNNs to architectures exceeding 100 layers. TA-SNN [52] introduced a temporal attention mechanism, paving the way for future attention mechanism advancements in the field. MA-SNN [53] went beyond single-dimensional information enhancement by proposing a temporal-channel-spatial attention mechanism. In Transformer architectures, models like SpikeFormer [61], SpikingFormer [62], and MetaSpikerFormer [63] demonstrated outstanding classification performance through the design and improvement of innovative spike-based self-attention blocks.", + "bbox": [ + 89, + 364, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation for SNN. Knowledge Distillation (KD), a well-established transfer learning technique, has proven effective across various tasks [3, 4, 22]. Recent studies [17, 30, 32, 47, 59] have applied KD to SNN training, using either large isomorphic SNNs to extract knowledge for smaller SNNs or leveraging the logits from pretrained ANNs to train SNNs. KDSNN [50] utilizes a joint distillation approach based on logits and features, showing effectiveness across multiple datasets. Xu et al. [49] explore biologically inspired structure learning using reverse knowledge distillation. Furthermore, LaSNN [24] proposes a hierarchical feature distillation framework that achieves accuracy comparable to ANNs on the Tiny ImageNet dataset. BKDSNN [51] improves SNN performance", + "bbox": [ + 89, + 688, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on complex datasets by using blurred knowledge to replicate ANN features. Despite the advantages of these distillation methods, as shown by the experiments in Fig. 2, current SNN distillation frameworks have yet to fully account for the unique characteristics of SNNs. Developing tailored distillation paradigms could enhance model accuracy in a simple and effective way.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 210, + 648, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we start from preliminaries, introducing the fundamental concepts of SNNs and SNN distillation. Then we introduce our knowledge distillation approach that incorporates temporal separation with entropy regularization.", + "bbox": [ + 511, + 237, + 906, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 308, + 653, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spiking Neuron Model. In SNNs, the fundamental computational unit is the spiking neuron, which serves as an abstract model of the dynamics of biological neurons. Currently, the Leaky Integrate-and-Fire (LIF) model is one of the most commonly used spiking neuron models, as it strikes a balance between simplified mathematical formulation and the complex dynamics of biological neurons. Therefore, in this study, we employ the LIF model as the foundational neuron model for the student SNN. Mathematically, a LIF neuron can be represented by Eq. 1.", + "bbox": [ + 511, + 330, + 906, + 482 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu \\frac {d u (t)}{d t} = - (u (t) - u _ {\\text {r e s e t}}) + I (t). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 493, + 906, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $\\mu$ denotes the time constant of the membrane potential, $u(t)$ represents the membrane potential at time $t$ , $u_{\\mathrm{reset}}$ is the resting potential of the neuron, and $I(t)$ is the presynaptic input at time $t$ . Based on this differential equation, the discrete-time and iterative mathematical representation of the LIF-SNN can be described as follows:", + "bbox": [ + 511, + 532, + 905, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV (t) = H (t - 1) + \\frac {1}{\\mu} [ I (t - 1) - (H (t - 1) - u _ {\\text {r e s e t}}) ]\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 632, + 890, + 664 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS (t) = \\Theta (V (t) - v _ {t h})\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 666, + 681, + 683 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH (t) = u _ {\\text {r e s e t}} \\cdot S (t) + V (t) \\cdot (1 - S (t)) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 685, + 905, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Heaviside step function $\\Theta$ is defined as:", + "bbox": [ + 511, + 715, + 807, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta (x) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} x < 0 \\\\ 1 & \\text {i f} x \\geq 0. \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 742, + 906, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among these, $H(t - 1)$ denotes the membrane potential following spike generation at the previous time step, while $I(t)$ and $V(t)$ represent the input and the updated membrane potential at time step $t$ , respectively. Furthermore, $v_{th}$ is the threshold that determines whether $V(t)$ should fire a spike or remain silent, and $S(t)$ indicates the spike sequence generated after triggering the action potential at time step $t$ .", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8808", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SNN Knowledge Distillation. We start from the original SNN Knowledge Distillation (KD) method. To illustrate the process of KD, we consider C-way classification task and denote $Z \\in \\mathbb{R}^{C}$ as the output of the network. Then, the class probability is given by:", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {i} = \\frac {\\exp \\left(Z _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(Z _ {j} / \\tau\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 171, + 483, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{i}$ and $Z_{i}$ are the probability value on the $i$ -th class and $\\tau$ is the temperature scaling hyper-parameter. In knowledge distillation, $\\tau$ is typically larger than 1.0, which helps control the flatness of the feature distribution to mitigate the phenomenon of overconfidence within the network. When $\\tau$ equals 1.0, the output reverts to the vanilla Softmax output, denoted here as $S(\\cdot)$ to represent this specific case. The objective of KD is to transfer knowledge from a high-performing teacher model to a lightweight student model. For the rescaled outputs, the original KD method achieves distillation by minimizing the KL divergence between the outputs of the teacher and student models.", + "bbox": [ + 89, + 213, + 483, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK L \\left(Q ^ {t e a} | | Q ^ {s t u}\\right) = \\sum_ {i = 1} ^ {C} Q _ {i} ^ {t e a} \\log \\left(\\frac {Q _ {i} ^ {t e a}}{Q _ {i} ^ {s t u}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 400, + 483, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{i}^{tea}$ and $Q_{i}^{stu}$ indicates the probability value on the $i$ -th category of the teacher and student output, respectively.", + "bbox": [ + 89, + 446, + 482, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, due to the potential errors in teachers' knowledge, KD also necessitates the use of cross-entropy to enable students to learn the true distribution of labels.", + "bbox": [ + 89, + 478, + 483, + 522 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC E \\left(Q ^ {s t u}, y _ {t r u e}\\right) = - \\sum_ {i = 1} ^ {C} Q _ {i} ^ {s t u} \\log \\left(y _ {t r u e}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 526, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $y_{\\text{true}}$ denote the true labels.", + "bbox": [ + 89, + 571, + 318, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The distinction between SNNs and ANNs lies in the fact that SNNs incorporate an additional temporal dimension $T$ , which denotes the time step. Consequently, traditional SNN distillation attempts to adapt the distillation methods designed for ANNs by employing a simplistic approach that averages the outputs across the temporal dimension. By introducing a parameter $\\alpha$ to control the weight ratio between true labels and soft labels. The loss calculation formula can be expressed as follows:", + "bbox": [ + 89, + 585, + 483, + 722 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {S K D} = (1 - \\alpha) \\cdot C E \\left(\\frac {1}{T} \\sum_ {t = 1} ^ {T} Q ^ {s t u} (t), y _ {t r u e}\\right) \\tag {7} \\\\ + \\alpha \\cdot K L (Q ^ {t e a} | | \\frac {1}{T} \\sum_ {t = 1} ^ {T} Q ^ {s t u} (t)) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 728, + 483, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Temporal Separation Knowledge Distillation with Entropy Regularization", + "text_level": 1, + "bbox": [ + 89, + 816, + 483, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we provide an in-depth explanation of the temporal separation knowledge distillation method with entropy regularization. Our approach consists of two core", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "components: temporal separation strategy and entropy regularization.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Separation Strategy. As illustrated in Fig. 1a, the original SNN distillation merely transferred the conventional KD methods to SNNs in a naive manner. The use of mean values to integrate outputs across the temporal dimension fails to adequately account for the unique spatiotemporal characteristics of SNNs. Due to the incremental temporal processing in SNNs, the output features at each time step exhibit variability and contain rich information. However, this also implies the potential presence of local outliers in the outputs at different time steps. While the averaging operation mitigates the impact of these anomalies, it poses risks for the subsequent model learning and probabilistic confidence. Consequently, implementing a temporal separation strategy for learning time step's output is crucial.", + "bbox": [ + 511, + 143, + 906, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 1b, we apply temporal separation strategy to CE and KL in Eq. 7, moving the mean operation outside. We directly use the outputs at each time step to distill the learning of the true labels and the teacher distribution. The loss expression for temporal separation knowledge distillation is given by", + "bbox": [ + 511, + 354, + 905, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {T S} = (1 - \\alpha) \\cdot \\frac {1}{T} \\sum_ {t = 1} ^ {T} C E \\left(Q ^ {s t u} (t), y _ {t r u e}\\right) \\tag {8} \\\\ + \\alpha \\cdot \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\cdot K L (Q ^ {t e a} | | Q ^ {s t u} (t)). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 458, + 903, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Entropy Regularization. After introducing the temporal separation strategy, the soft labels from the teacher network directly impact the output distribution at each time step. This approach, however, can inadvertently reinforce erroneous information from the teacher network, increasing the likelihood that the student model will adopt incorrect knowledge. To counter this, we introduce an entropy regularization term, which constrains and adjusts the teacher's outputs to reduce excessive confidence in potentially erroneous knowledge. The formula for entropy is as follows:", + "bbox": [ + 511, + 578, + 906, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH \\left(Q ^ {s t u}\\right) = - \\sum_ {i = 1} ^ {C} Q _ {i} ^ {s t u} \\log \\left(Q _ {i} ^ {s t u}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 742, + 903, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we obtain the loss formula for the entropy regularization term:", + "bbox": [ + 511, + 791, + 903, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E R} = - \\lambda \\cdot \\frac {1}{T} \\sum_ {t = 1} ^ {T} H \\left(Q ^ {s t u} (t)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 832, + 903, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is the penalty factor.", + "bbox": [ + 513, + 885, + 709, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8809", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we integrate the temporal separation strategy, as shown in Eq. 8, with the entropy regularization term from Eq. 10 to derive a new loss function as follows:", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T S E R} = \\mathcal {L} _ {T S} + \\mathcal {L} _ {E R}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 148, + 482, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "And Algorithm. 1 outlines the overall training process of our proposed method.", + "bbox": [ + 89, + 175, + 483, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 219, + 215, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to validate the effectiveness of our method. Initially, we compare our approach with existing SNN distillation methods. Furthermore, we conduct ablation studies, sensitivity analyses, energy consumption assessments, and visualizations to provide a comprehensive evaluation of our method.", + "bbox": [ + 89, + 244, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 344, + 297, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed method is evaluated on three datasets. For teacher ANN networks, we select several advanced models, including ResNet-19 and ResNet-34 [21] as well as VGG-16. In contrast, for the student SNN models, we employ neural network architectures with equivalent or fewer layers, specifically ResNet-18/19, and VGG-11/16. Additionally, we conduct multiple experiments by varying the time step.", + "bbox": [ + 89, + 366, + 483, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison on Static Datasets", + "text_level": 1, + "bbox": [ + 89, + 496, + 362, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method is compared with previous approaches on CIFAR10 and CIFAR100, as shown in the Table. 1. The re", + "bbox": [ + 89, + 517, + 483, + 547 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Training student SNN model with TSER knowledge distillation for one epoch.", + "Ensure: SNN model with TSER KD" + ], + "code_body": "Require: pre-trained teacher ANN model $M_{t}$ ; an initialized student SNN model $M_{s}$ ; input dataset sample $X$ ; and the true labels $y_{true}$ ; total training iteration in one epoch $I_{train}$ ; total validation iteration in one epoch $I_{val}$", + "guess_lang": "txt", + "bbox": [ + 91, + 604, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "for $i = 1$ to $I_{train}$ do Get mini-batch training data, and true label: $X^i,y_{true}^i$ Compute the $M_{s}$ output $Z^{stu,i}(t)$ of each time step; Compute the $M_{t}$ output $Z^{tea,i}$ Calculate loss function: $\\mathcal{L}_{TSER}$ Backpropagation and update model parameters; \nend for \nfor $i = 1$ to $I_{val}$ do Get mini-batch validation data, and true label: $X^i$ $y_{true}^{i}$ Compute the SNN average output $Z_{mean}^{stu,i} =$ $\\frac{1}{T}\\sum_{t = 1}^{T}Z^{stu,i}(t);$ Compare the $Z_{mean}^{stu,i}$ and $y_{true}^{i}$ for classification; \nend for", + "guess_lang": "txt", + "bbox": [ + 107, + 681, + 483, + 895 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/93817fae9dd3b32330078722e85b57ea777ccee5f5d851884c83b955275a7e96.jpg", + "image_caption": [ + "Figure 2. Prediction Accuracy Distribution at Different Time Steps for Vanilla SNN KD and Our Proposed Method. The solid red line indicates the teacher model's accuracy, while the dashed lines represent the prediction accuracies of different distillation methods after averaging outputs over time steps. The bars show the prediction accuracies of each distillation method at individual time steps." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 893, + 325 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1393a2f92084cee688a2bec85335b14cadb8dc0500647d2c5e29e6a0d3650b66.jpg", + "image_caption": [ + "Figure 3. Accuracy Distribution for Different $\\lambda$ Values. Experiments conducted on the CIFAR100 dataset with a fixed time step of 2, testing various $\\lambda$ values." + ], + "image_footnote": [], + "bbox": [ + 524, + 446, + 890, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sults demonstrate our method surpasses the SOTA performance achieved by prior method across both ResNet and VGG architectures. Specifically, ResNet19 outperform the previous SOTA by $0.57\\%$ and $0.41\\%$ on CIFAR10 and CIFAR100, respectively, achieving accuracies of $96.72\\%$ and $81.29\\%$ . This indicates that our approach, although based on logit distillation, performs exceptionally well compared to previous, more complex distillation methods. By addressing the limitations of simple transfer in existing distillation, our framework effectively learns from the teacher's knowledge, minimizing the gap between the student and", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8810", + "bbox": [ + 482, + 988, + 514, + 1000 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e13fd0d31eded058099fb55343a4c8200ee635b7d47d16992fcd4b5ece1a8eab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAcc.(%)Arch.Acc.(%)Arch.Acc.(%)Arch.Acc.(%)
CF-100CF-10CF-100CF-10CF-100CF-10CF-100CF-10
without KDTimestep=2Timestep=4
R-1975.4894.44R-1874.6094.15R-1977.0594.85R-1876.4294.67
TET[11]R-1972.8794.16---R-1974.4794.44---
LSG[35]R-1976.3294.41---R-1976.8595.17---
PFA[7]R-1976.7095.60---R-1978.1095.71---
MPBN[18]R-1979.5196.47V-1673.8893.96R-1980.1096.52V-1674.7494.44
IM-LIF[36]R-1977.2195.29---R-1977.4295.66---
Spikformer[64]------S-4-25675.9693.94S-4-38477.8695.19
Spikingformer[61]------Sg-4-25677.4394.77Sg-4-38479.0995.61
SD Transformer [54]------S-2-51278.4095.60---
with KDTeacher ANN Timestep=1
R-3481.1297.10V-1678.0896.06R-3481.1297.10V-1678.0896.06
R-1981.9797.08V-1678.0896.06---V-1678.0896.06
LaSNN [24]Timestep=2Timestep=4
R-1876.1793.64V-1673.8093.90R-1878.1295.09V-1674.9994.49
R-1980.3095.26V-1169.8990.23---V-1170.7490.42
KDSNN [50]R-1877.1695.25V-1674.6594.26R-1878.4695.72V-1675.9894.85
R-1980.8896.15V-1172.1792.81---V-1173.0592.91
BKDSNN [51]R-1873.3093.64V-1673.8094.10R-1875.5795.09V-1674.9994.55
R-1975.7495.26V-1169.8992.73---V-1170.7492.88
OursR-1878.30±0.0895.58±0.08V-1675.81±0.1294.55±0.09R-1879.69±0.0296.18±0.05V-1677.06±0.0495.01±0.10
R-1981.29±0.1496.72±0.02V-1173.67±0.0493.31±0.02---V-1174.65±0.0793.68±0.02
", + "bbox": [ + 89, + 87, + 915, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison results with training-based SNN SOTA methods, including CNN- and transformer-based approaches, on CIFAR-10/100, with and without Knowledge Distillation (KD). Acc. denotes accuracy, CF denotes CIFAR, and Archi. denotes architecture. The abbreviations R, V, S, and Sg represent ResNet, VGG, Spikeformer, and Spikingformer architectures, respectively, while SD Transformer refers to the Spike-Driven Transformer. For KD methods, the student model aligns with the architecture of the teacher ANN.", + "bbox": [ + 89, + 550, + 906, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "teacher models. This reduction in disparity validates the superiority of our method.", + "bbox": [ + 89, + 633, + 482, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further comparisons on ImageNet dataset are detailed in Table. 2. The SEW ResNet-18 and SEW ResNet-34 architectures achieve top-1 accuracies of $69.24\\%$ and $73.16\\%$ , respectively. These results indicate that our method is capable of delivering competitive performance even within large-scale datasets, effectively harnessing the knowledge from ANNs to enhance the learning capabilities of SNNs.", + "bbox": [ + 89, + 666, + 482, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 786, + 243, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Temporal Separation Strategies and Entropy Regularization. We investigate different configurations of $\\mathcal{L}_{TSER}$ , focusing on the temporal separation applied within the KL and CE losses, as well as the integration of an entropy regularization term. The results are summarized in Table. 3. Our experiments reveal that incorporating tempo", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ral separation strategies notably enhances model accuracy, with improvements of $0.49\\%$ and $0.26\\%$ for the standard KL and CE loss function, respectively. When temporal separation is applied concurrently to both KL and CE losses, the overall accuracy increase reaches $0.84\\%$ . Furthermore, we find that optimally adjust the parameter $\\lambda$ in the entropy regularization term, alongside temporal separation, leads to an enhancement of up to $0.98\\%$ . These findings underscore the effectiveness of combining temporal separation and entropy regularization in maximizing the performance of SNN distillation.", + "bbox": [ + 511, + 633, + 906, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of Output Accuracy at Different Time Steps. To illustrate the adaptability of our proposed distillation method for SNNs, we perform a comparative analysis of accuracy across different time steps between vanilla SNN KD and our method. As shown in Fig. 2, the accu", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8811", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9a7e7d0d12831807cc3a2159a48952571aeeea648cbb54d41de631270149d793.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodArchitectureTAccuracy
without KDTET [11]R-34664.79%
GLIF [55]R-34467.52%
MPBN [18]R-18463.14%
R-34464.71%
SEW ResNet [15]R-18463.18%
R-34467.04%
Spikformer [64]S-8-384470.24%
S-6-512472.46%
Spikingformer [61]Sg-8-384472.45%
Spike-driven Transformer [54]Sg-8-384472.28%
with KDTeacher ANNSEW R-18171.69%
SEW R-34175.38%
LaSNN [24]SEW R-18463.33%
SEW R-34466.98%
KDSNN [50]SEW R-18463.61%
SEW R-34467.28%
BKDSNN [51]SEW R-18463.43%
SEW R-34467.21%
OursSEW R-18469.24% ±0.19
SEW R-34473.16% ±0.15
", + "bbox": [ + 91, + 88, + 486, + 464 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/435eb65b65c3cd40839bbd5e8dfbfafe7c411298c776c94bad82792f79a2a49c.jpg", + "table_caption": [ + "Table 2. Comparison results with training-based SNN SOTA methods, including CNN- and Transformer-based approaches, on ImageNet, with and without Knowledge Distillation (KD). T denotes Timestep. R, S, Sg denotes ResNet, Spikeformer and Spikingformer architectures respectively." + ], + "table_footnote": [], + "table_body": "
Temporal SeparationEntropy RegularizationAccuracy
KL LossCE Loss
XXX77.15%
X✓X77.41%
✓XX77.64%
✓✓X77.99%
✓✓✓78.13%
", + "bbox": [ + 96, + 556, + 478, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation of Temporal Separation Strategy and Entropy Regularization.", + "bbox": [ + 89, + 676, + 482, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "racy discrepancy among time steps for the vanilla SNN KD is $0.98\\%$ , notably higher than the $0.65\\%$ difference with our approach. Furthermore, the vanilla SNN KD yields an average output accuracy of $74.67\\%$ , whereas our method achieves $76.56\\%$ . And as the accuracy of individual time step outputs improved, we observe a significant enhancement in performance following the mean operation on these outputs, aligning with our hypothesis. This clearly demonstrates the effectiveness of our method and highlights the significance of temporal separation in the SNN distillation process.", + "bbox": [ + 88, + 734, + 486, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3964209d0776f37ed9eb84592b1349dc506c2a179fecfedc01f731e956bbae19.jpg", + "image_caption": [ + "Figure 4. Temperature Coefficient Sensitivity. Our method demonstrates stability across various $\\tau$ hyperparameters. This experiment is conducted on CIFAR100 with ResNet-34 as the teacher model and ResNet-18 as the student model." + ], + "image_footnote": [], + "bbox": [ + 532, + 90, + 885, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Selection of Parameter $\\lambda$ . The entropy regularization term is essential for $\\mathcal{L}_{TSER}$ , as it partially alleviates the erroneous knowledge transferred by the teacher network. Thus, selecting an optimal value for $\\lambda$ is crucial. We conduct experiments with various values of $\\lambda$ , and the results are shown in Fig. 3. Increasing $\\lambda$ enhances the correction of the teacher's inaccuracies but also dilutes the distribution of the correct knowledge. Our findings indicate that maintaining $\\lambda$ within the range of 0.0001 to 0.2 enhances the network's overall training stability. However, excessively high values of $\\lambda$ can result in negative loss values, as described in Eq. 11, which must be avoided.", + "bbox": [ + 511, + 375, + 908, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temperature Coefficient Sensitivity. In our experiments, we configure the temperature to $\\tau = 5$ . Utilizing the prior experimental setup, we select five temperature points: $\\tau = [2,3,4,5,6]$ . We evaluate performance with ResNet-34 as the teacher model and ResNet-18 as the student model on the CIFAR100 dataset. The results depicted in Fig. 4, indicate that our method exhibits stable performance across different temperature coefficients.", + "bbox": [ + 511, + 579, + 908, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Performance Analysis and Visualization", + "text_level": 1, + "bbox": [ + 511, + 710, + 857, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Firing Rate Analysis. In Table. 4, the firing rates of our proposed method show a slight increase compared to vanilla SNN KD across both ResNet and VGG architectures. Fig. 2 illustrates that an increase in firing rate of approximately $0.1\\%$ corresponds to a $1.48\\%$ improvement in overall performance. Notably, when comparing our method to other distillation techniques within the ResNet architecture, it maintains performance gains while simultaneously reducing firing rate and increasing spike sparsity, thus minimizing redundant spikes. Although LaSNN exhibits lower firing rates, this comes at the expense of some model per", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8812", + "bbox": [ + 482, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d4af674e592612b5ad7c6da632b259beee57c74f8fb33a5f4a9629e284e147f.jpg", + "image_caption": [ + "Figure 5. t-SNE Visualization of features learned by teacher ANN and different distillation methods." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 222, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/054805f500d4ff747523588d7f021f2097d9cd41c8e181604307ab2f882f7a48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 90, + 359, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1db2d30c835228bf392eaea766568a0644916da3776b17666313c415c128354.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 90, + 496, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06b70e903b440527740feb3966b35720be5ebfd20991e8ce9ed05db5cdf04edc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 624, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a5f7fc37f48a03e2d52282fb9b496fd760278a4574a9876e8b46e6a3320c49c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 90, + 759, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e2eb4b5e824a4531ffadb701a9ad7a6d8eaf902270627046d3cd5d96d9caf2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 90, + 893, + 190 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6cb3d31a64104bd5a55b18e84ab1fea6968f53d62d74d8af00d90a5a62438f0b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ArchitectureMethodsOPsACsMACsEnergyTraining TimeFiring Rate
ResNet18Vanilla SNN KD2.23G405.33M12.2M0.42mJ2.78s22.85%
KDSNN2.23G411.49M12.2M0.43mJ2.67s23.23%
LaSNN2.23G300.94M12.2M0.33mJ2.79s16.76%
BKDSNN2.23G511.57M12.2M0.52mJ2.84s34.93%
Ours2.23G405.98M12.2M0.42mJ2.70s22.95%
VGG16Vanilla SNN KD1.26G159.96M274.2M1.40mJ2.43s18.91%
KDSNN1.26G170.79M274.2M1.39mJ2.39s18.59%
LaSNN1.26G150.96M274.2M1.41mJ2.44s17.25%
BKDSNN1.26G171.94M274.2M1.40mJ2.58s18.66%
Ours1.26G160.92M274.2M1.41mJ2.40s20.05%
", + "bbox": [ + 168, + 234, + 831, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Energy consumption, training time, and spike firing rate under various distillation methods and architectures on the CIFAR100 dataset. Training Time refers to the time required per batch during training.", + "bbox": [ + 89, + 406, + 906, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance. Furthermore, unlike our logit-based approach, LaSNN employs a feature-based method that optimizes the feature map directly, making it more advantageous for feature sparsification.", + "bbox": [ + 89, + 460, + 482, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Energy Consumption Analysis. Our energy consumption analysis, based on the energy model in [29], is summarized in Table. 4. The results reveal that the most significant variations in energy consumption occur in Accumulation Computation (AC), which is related to the firing rate. Notably, a lower firing rate correlates with reduced AC. Consequently, our approach does not increase the computational load of SNNs and is competitive regarding energy efficiency.", + "bbox": [ + 89, + 545, + 482, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Time. We evaluate the training times of various competitive KD methods on the CIFAR100 dataset, focusing on the training time per batch. As shown in Table. 4, our method achieves the second-fastest training time, surpassed only by KDSNN. This advantage is largely due to KDSNN simplifying the original KL divergence into cross-entropy in the logit-based framework, while lowering computational cost, also affect accuracy. Our approach outperforms other techniques by utilizing only the logits output for knowledge distillation, thus eliminating the need for additional auxiliary training modules or feature map comparisons. In contrast, previous methods demand more time and resources to extract knowledge from intermediate layers.", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization. We employ t-SNE to visualize features learned from various distillation methods on the CIFAR100 dataset, utilizing ResNet-34 as the teacher and ResNet-18 as the student model. As shown in Fig. 5, our approach significantly enhances the distinguishability of deeper features compared to existing SNN distillation techniques.", + "bbox": [ + 511, + 460, + 906, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 568, + 633, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work addresses the issue of existing SNN distillation methods that are simply derived from ANN. We propose a new framework specifically designed for SNN distillation, which better exploits the temporal information contained in the logit output of SNNs for the process of knowledge distillation. Specifically, we incorporate a temporal separation strategy and introduce an entropy regularization term into the original distillation method. We aim to uncover the potential of SNN distillation by fully leveraging temporal information and rectifying the erroneous teacher knowledge. Extensive experimental results demonstrate the efficacy of this approach.", + "bbox": [ + 511, + 594, + 908, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 791, + 700, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China (NSFC) (62476035, 62206037, 62276230), and Natural Science Foundation of Zhejiang Province (LDT23F02023F02), and State Key Laboratory (SKL) of Biobased Transportation Fuel Technology.", + "bbox": [ + 511, + 816, + 906, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8813", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347, 2023. 1", + "[2] Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision, 113:54-66, 2015. 1", + "[3] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 3", + "[4] Li Chen, Chunyan Yu, and Lvcai Chen. A new knowledge distillation for incremental object detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2019. 3", + "[5] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4794-4802, 2019. 2, 3", + "[6] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018. 1", + "[7] Haoyu Deng, Ruijie Zhu, Xuerui Qiu, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tensor decomposition based attention module for spiking neural networks. Knowledge-Based Systems, 295:111780, 2024. 6", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2", + "[9] Lei Deng, Yujie Wu, Xing Hu, Ling Liang, Yufei Ding, Guoqi Li, Guangshe Zhao, Peng Li, and Yuan Xie. Rethinking the performance comparison between snns and anns. Neural networks, 121:294-307, 2020. 1", + "[10] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476, 2021. 1, 3", + "[11] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. In International Conference on Learning Representations. 6, 7", + "[12] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2015. 3", + "[13] Jianhao Ding, Tong Bu, Zhaofei Yu, Tiejun Huang, and Jian Liu. Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training. Advances in Neural Information Processing Systems, 35:24780-24793, 2022. 1", + "[14] Jason K Eshraghian, Max Ward, Emre O Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Ben" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "namoun, Doo Seok Jeong, and Wei D Lu. Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE, 2023. 1", + "[15] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 3, 7", + "[16] Isha Garg, Sayeed Shafayet Chowdhury, and Kaushik Roy. Dct-snn: Using dct to distribute spatial information over time for low-latency spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4671-4680, 2021. 3", + "[17] Yufei Guo, Weihang Peng, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Xuhui Huang, and Zhe Ma. Joint a-sn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization. Pattern Recognition, 142:109639, 2023. 3", + "[18] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19420-19430, 2023. 6, 7", + "[19] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European conference on computer vision, pages 388-404. Springer, 2020. 3", + "[20] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558-13567, 2020. 3", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[23] Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdelrahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):82-97, 2012. 1", + "[24] Di Hong, Jiangrong Shen, Yu Qi, and Yueming Wang. Lasnn: Layer-wise ann-to-snn distillation for effective and efficient training in deep spiking neural networks. arXiv preprint arXiv:2304.09101, 2023. 2, 3, 6, 7", + "[25] Saeed Reza Kheradpisheh, Mohammad Ganjtabesh, Simon J Thorpe, and Timothée Masquelier. Stdp-based spiking deep convolutional neural networks for object recognition. Neural Networks, 99:56-67, 2018. 2", + "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2", + "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8814", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "works. Advances in neural information processing systems, 25, 2012. 1", + "[28] Souvik Kundu, Gourav Datta, Massoud Pedram, and Peter A Beerel. Spike-thrift: Towards energy-efficient deep spiking neural networks by limiting spiking activity via attention-guided compression. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3953-3962, 2021. 2", + "[29] Souvik Kundu, Massoud Pedram, and Peter A Beerel. Hiresnn: Harnessing the inherent robustness of energy-efficient deep spiking neural networks by training with crafted input noise. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5209-5218, 2021. 8", + "[30] Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee, and Rajbabu Velmurugan. Distilling spikes: Knowledge distillation in spiking neural networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4536-4543. IEEE, 2021. 3", + "[31] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1", + "[32] Dongjin Lee, Seongsik Park, Jongwan Kim, Wuhyeong Doh, and Sungroh Yoon. Energy-efficient knowledge distillation for spiking neural networks. arXiv preprint arXiv:2106.07172, 2021. 3", + "[33] Chen Li, Lei Ma, and Steve Furber. Quantization framework for fast spiking neural networks. Frontiers in Neuroscience, 16:918793, 2022. 1", + "[34] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021. 1", + "[35] Shuang Lian, Jiangrong Shen, Qianhui Liu, Ziming Wang, Rui Yan, and Huajin Tang. Learnable surrogate gradient for direct training spiking neural networks. In IJCAI, pages 3002-3010, 2023. 6", + "[36] Shuang Lian, Jiangrong Shen, Ziming Wang, and Huajin Tang. Im-lif: Improved neuronal dynamics with attention mechanism for direct training deep spiking neural network. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 6", + "[37] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1", + "[38] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014. 1", + "[39] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36 (6):51-63, 2019. 3", + "[40] Kleanthis C Neokleous, Marios N Avraamides, Costas K Neocleous, and Christos N Schizas. Selective attention and" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "consciousness: investigating their relation through computational modelling. Cognitive Computation, 3:321-331, 2011. 2", + "[41] Srdjan Ostojic. Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons. Nature neuroscience, 17(4):594–600, 2014. 1", + "[42] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 1", + "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1", + "[44] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1", + "[45] Catherine D Schuman, Shruti R Kulkarni, Maryam Parsa, J Parker Mitchell, Bill Kay, et al. Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science, 2(1):10-19, 2022. 1", + "[46] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 1", + "[47] Sugahara Takuya, Renyuan Zhang, and Yasuhiko Nakashima. Training low-latency spiking neural network through knowledge distillation. In 2021 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPPS), pages 1-3. IEEE, 2021. 3", + "[48] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 2", + "[49] Qi Xu, Yaxin Li, Xuanye Fang, Jiangrong Shen, Jian K. Liu, Huajin Tang, and Gang Pan. Biologically inspired structure learning with reverse knowledge distillation for spiking neural networks, 2023. arXiv:2304.09500 [cs]. 3", + "[50] Qi Xu, Yaxin Li, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7886-7895, 2023. 2, 3, 6, 7", + "[51] Zekai Xu, Kang You, Qinghai Guo, Xiang Wang, and Zhezhi He. Bkdsnn: Enhancing the performance of learning-based spiking neural networks training with blurred knowledge distillation. arXiv preprint arXiv:2407.09083, 2024. 2, 3, 6, 7", + "[52] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8815", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. arXiv preprint arXiv:2209.13929, 2022. 2, 3", + "[54] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 6, 7", + "[55] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 6, 7", + "[56] Qiang Yu, Jialu Gao, Jianguo Wei, Jing Li, Kay Chen Tan, and Tiejun Huang. Improving multispike learning with plastic synaptic delays. IEEE Transactions on Neural Networks and Learning Systems, 34(12):10254-10265, 2022. 3", + "[57] Friedemann Zenke and Tim P Vogels. The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation, 33 (4):899-925, 2021. 2", + "[58] Friedemann Zenke, Everton J Agnes, and Wulfram Gerstner. Diverse synaptic plasticity mechanisms orchestrated to form and retrieve memories in spiking neural networks. Nature communications, 6(1):6922, 2015. 1", + "[59] Fengzhao Zhang, Chengting Yu, Hanzhi Ma, Zheming Gu, and Er-ping Li. Knowledge Distillation For Spiking Neural Network. In 2023 5th International Conference on Robotics, Intelligent Control and Artificial Intelligence (RICAL), pages 1015-1020, 2023. 3", + "[60] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 3", + "[61] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 3, 6, 7", + "[62] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Zhengyu Ma, Huihui Zhou, Xiaopeng Fan, and Yonghong Tian. Enhancing the performance of transformer-based spiking neural networks by snn-optimized downsampling with precise gradient backpropagation. arXiv preprint arXiv:2305.05954, 2023. 3", + "[63] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 3", + "[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations. 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 838 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8816", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/511efbd9-e72e-4675-94ed-5687d3419119_model.json b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/511efbd9-e72e-4675-94ed-5687d3419119_model.json new file mode 100644 index 0000000000000000000000000000000000000000..821d3d8823a2b1d07c8aeaea241b370f0e92549f --- /dev/null +++ b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/511efbd9-e72e-4675-94ed-5687d3419119_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.202, + 0.803, + 0.222 + ], + "angle": 0, + "content": "Kairong \\(\\mathrm{Yu}^{1}\\) Chengting \\(\\mathrm{Yu}^{1}\\) Tianqing Zhang\\(^{1}\\) Xiaochen Zhao\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.222, + 0.785, + 0.24 + ], + "angle": 0, + "content": "Shu Yang\\(^{1}\\) Hongwei Wang\\(^{1,\\dagger}\\) Qiang Zhang\\(^{2}\\) Qi Xu\\(^{2,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.244, + 0.74, + 0.265 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University \\(^{2}\\) Dalian University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.266, + 0.83, + 0.281 + ], + "angle": 0, + "content": "{kairong.22,chengting.21,xiaochen.24，shu.23,hongweiwang}@intlzju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.285, + 0.602, + 0.298 + ], + "angle": 0, + "content": "zhangtianqing@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.301, + 0.607, + 0.316 + ], + "angle": 0, + "content": "{xuqi, zhangq}@dlut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.351, + 0.327, + 0.367 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.383, + 0.485, + 0.746 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs), inspired by the human brain, offer significant computational efficiency through discrete spike-based information transfer. Despite their potential to reduce inference energy consumption, a performance gap persists between SNNs and Artificial Neural Networks (ANNs), primarily due to current training methods and inherent model limitations. While recent research has aimed to enhance SNN learning by employing knowledge distillation (KD) from ANN teacher networks, traditional distillation techniques often overlook the distinctive spatiotemporal properties of SNNs, thus failing to fully leverage their advantages. To overcome these challenges, we propose a novel logit distillation method characterized by temporal separation and entropy regularization. This approach improves existing SNN distillation techniques by performing distillation learning on logits across different time steps, rather than merely on aggregated output features. Furthermore, the integration of entropy regularization stabilizes model optimization and further boosts the performance. Extensive experimental results indicate that our method surpasses prior SNN distillation strategies, whether based on logit distillation, feature distillation, or a combination of both. Our project is available at https://github.com/yukairong/TSER." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.222, + 0.791 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.801, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Inspired by the neural firing mechanisms observed in biological systems, Spiking Neural Networks (SNNs) [37] are considered a promising alternative to traditional Artificial Neural Networks (ANNs) due to their superior energy efficiency. Unlike ANNs, which use continuous activation val" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.908, + 0.715 + ], + "angle": 0, + "content": "ues for information transmission[31], SNNs transmit and process information through discrete spike events [6, 38], with neurons generating spikes only when their membrane potential exceeds a threshold. This binary, event-driven approach allows SNNs to run efficiently on neuromorphic hardware [6, 42], accumulating synaptic inputs effectively and avoiding unnecessary computations related to zero input or activation [9, 14]. Given their event-driven dynamics and the biomimetic properties of spatiotemporal neuron activity [44, 45], SNNs demonstrate remarkable energy efficiency, robust adaptive learning capabilities [13, 41, 58], and ultra-low power consumption, showing significant potential for computational intelligence applications tasks [44]. Despite the inherent advantages of SNNs, their performance in common tasks such as image classification [27], object segmentation [43], and natural language processing [23] still lags behind that of ANNs. This gap largely stems from the limitations of current training methods and structural constraints in SNNs. In contrast to ANNs, SNNs cannot directly utilize backpropagation (BP) for deep network training. Moreover, attempts to directly adapt ANN methodologies to SNNs encounter compatibility issues, preventing SNNs from fully realizing their theoretical advantages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Currently, conversion-based and learning-based approaches are two common methods for training SNNs. The former approaches seek to utilize knowledge from ANNs by transferring the parameters of a pre-trained ANN to an equivalent SNN [1, 10, 33, 34]. These methods often necessitate a significant number of time steps to achieve accuracy levels comparable to those of the original ANN [2, 46]. If the number of inference time steps is reduced, the model's capacity for effective information transfer may diminish, resulting in a decline in overall performance. To improve the performance of SNNs and minimize inference time steps, learning-based approaches have been employed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.242, + 0.901 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8806" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.488, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.303, + 0.411, + 0.316 + ], + "angle": 0, + "content": "(a) Classical SNN Knowledge Distillation" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.096, + 0.892, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.291, + 0.9, + 0.316 + ], + "angle": 0, + "content": "(b) Temporal Separation with Entropy Regularization (TSER) Knowledge Distillation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.907, + 0.399 + ], + "angle": 0, + "content": "Figure 1. Illustration of the Classical SNN KD and Our Proposed Method. In our approach, we remove the temporal dimension fusion operation present in classical SNN KD and apply a temporal separation strategy to focus the learning process on outputs at individual time steps. The original calculations of KL Loss and CE Loss are adapted to compute the loss for each time step's output, followed by averaging. Additionally, we incorporate an entropy regularization term to guide the learning direction away from erroneous knowledge. Here, T, B, and C represent the time steps, batch size, and number of classes, respectively." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.424, + 0.485, + 0.817 + ], + "angle": 0, + "content": "to facilitate training during inference. These approaches leverage Backpropagation Through Time (BPTT) for effective training [48, 57]. Motivated by the work [25], several studies have utilized surrogate gradient estimation as a method to address the non-differentiability inherent in SNNs. This foundational research has enabled the direct training of CNNs and transformer-based SNNs, achieving strong performance on static datasets with reduced inference time steps. To further reduce the performance gap with ANNs, recent work has explored the inclusion of modules such as attention mechanisms to enhance SNN capabilities [28, 40, 52, 53]. While effective, these additional modules increase computational consumption, counteracting the inherent low-energy benefits of SNNs. Knowledge Distillation (KD) provides a potential solution, where SNNs are guided by a more capable ANN teacher model, addressing the inference burden associated with these modules. Prior works such as KDSNN [50] and LaSNN [24] employ hierarchical knowledge distillation [5] to facilitate the training of SNNs by leveraging the rich semantic information from the ANN teacher model. Recent approaches like BKDSNN [51] move beyond traditional layer-wise distillation by integrating blurred knowledge distillation techniques with logit distillation, demonstrating significant improvements in performance across both CNNs and Transformer-based models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In preliminary experiments, we observed that current SNN distillation architectures are not optimally designed for the unique structure of SNNs. As depicted in Fig. 1a, conventional SNN KD techniques distill the mean of outputs across time steps. However, the inherent nature" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.908, + 0.68 + ], + "angle": 0, + "content": "of SNN processing implies that significant information is embedded within the temporal dimension. Employing a mean approach for fusion prediction during classification resembles a voting selection process. Therefore, distributing the distillation of outputs across various time steps during the training phase can substantially alleviate the limitations that currently hinder distillation performance. To mitigate the impact of outliers on network performance and to rectify overly confident incorrect semantics produced by the teacher network, we introduce an entropy regularization term as a constraint. We conduct extensive experimental analyses to assess its efficacy. Furthermore, we evaluate our proposed method on several datasets, including CIFAR10 [26], CIFAR100 [26], and ImageNet [8]. The results of our experiments demonstrate that this method achieves state-of-the-art (SOTA) performance. In summary, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.717, + 0.905, + 0.791 + ], + "angle": 0, + "content": "- We introduce a temporal separation strategy into the SNN distillation method, effectively addressing the incompatibility issues present in previous SNN distillation frameworks and yielding a more effective universal distillation architecture." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.793, + 0.905, + 0.851 + ], + "angle": 0, + "content": "- We integrate entropy regularization to correct the overly confident errors from the teacher network, thereby maximizing the performance of the logit-based distillation method." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.853, + 0.905, + 0.898 + ], + "angle": 0, + "content": "- We conduct comprehensive evaluations across multiple datasets, showing that our proposed approach outperforms previous state-of-the-art (SOTA) methods." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.717, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8807" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.345 + ], + "angle": 0, + "content": "Conversion-based Methods. To fully leverage the knowledge acquired from pre-trained ANNs, conversion methods directly transfer the learned parameters from an ANN to the corresponding SNN, avoiding the difficulties of training SNNs from scratch. Several studies [19, 20] replace the activation functions in ANNs with spiking neurons and introduce optimizations like weight normalization [12] to generate corresponding SNNs. The core idea of these methods is to closely align the outputs of the ANN and SNN, maximizing the utilization of the pre-trained ANN knowledge. While later works [5, 10, 12] achieve nearly lossless accuracy conversion, they often involve long inference times and fail to retain the spatiotemporal characteristics inherent to spike-based processing, significantly limiting the practical applicability of converted SNNs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.668 + ], + "angle": 0, + "content": "Learning-based Methods. The development of surrogate gradient training methods [56] has resolved the challenge of training non-differentiable SNNs directly. Initially, [39] used a surrogate gradient approach to construct deep SNN models, achieving competitive accuracy on neuromorphic datasets. The DCT-SNN [16] applied frequency domain techniques to reduce the number of inference time steps. To support deeper SNN training, [15] introduced SEW ResNet to address gradient vanishing and exploding issues in Spiking ResNet [60], extending SNNs to architectures exceeding 100 layers. TA-SNN [52] introduced a temporal attention mechanism, paving the way for future attention mechanism advancements in the field. MA-SNN [53] went beyond single-dimensional information enhancement by proposing a temporal-channel-spatial attention mechanism. In Transformer architectures, models like SpikeFormer [61], SpikingFormer [62], and MetaSpikerFormer [63] demonstrated outstanding classification performance through the design and improvement of innovative spike-based self-attention blocks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Knowledge Distillation for SNN. Knowledge Distillation (KD), a well-established transfer learning technique, has proven effective across various tasks [3, 4, 22]. Recent studies [17, 30, 32, 47, 59] have applied KD to SNN training, using either large isomorphic SNNs to extract knowledge for smaller SNNs or leveraging the logits from pretrained ANNs to train SNNs. KDSNN [50] utilizes a joint distillation approach based on logits and features, showing effectiveness across multiple datasets. Xu et al. [49] explore biologically inspired structure learning using reverse knowledge distillation. Furthermore, LaSNN [24] proposes a hierarchical feature distillation framework that achieves accuracy comparable to ANNs on the Tiny ImageNet dataset. BKDSNN [51] improves SNN performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "on complex datasets by using blurred knowledge to replicate ANN features. Despite the advantages of these distillation methods, as shown by the experiments in Fig. 2, current SNN distillation frameworks have yet to fully account for the unique characteristics of SNNs. Developing tailored distillation paradigms could enhance model accuracy in a simple and effective way." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.212, + 0.649, + 0.23 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.907, + 0.3 + ], + "angle": 0, + "content": "In this section, we start from preliminaries, introducing the fundamental concepts of SNNs and SNN distillation. Then we introduce our knowledge distillation approach that incorporates temporal separation with entropy regularization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.309, + 0.655, + 0.324 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.907, + 0.483 + ], + "angle": 0, + "content": "Spiking Neuron Model. In SNNs, the fundamental computational unit is the spiking neuron, which serves as an abstract model of the dynamics of biological neurons. Currently, the Leaky Integrate-and-Fire (LIF) model is one of the most commonly used spiking neuron models, as it strikes a balance between simplified mathematical formulation and the complex dynamics of biological neurons. Therefore, in this study, we employ the LIF model as the foundational neuron model for the student SNN. Mathematically, a LIF neuron can be represented by Eq. 1." + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.494, + 0.907, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mu \\frac {d u (t)}{d t} = - (u (t) - u _ {\\text {r e s e t}}) + I (t). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.906, + 0.625 + ], + "angle": 0, + "content": "Where \\(\\mu\\) denotes the time constant of the membrane potential, \\(u(t)\\) represents the membrane potential at time \\(t\\), \\(u_{\\mathrm{reset}}\\) is the resting potential of the neuron, and \\(I(t)\\) is the presynaptic input at time \\(t\\). Based on this differential equation, the discrete-time and iterative mathematical representation of the LIF-SNN can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.633, + 0.892, + 0.665 + ], + "angle": 0, + "content": "\\[\nV (t) = H (t - 1) + \\frac {1}{\\mu} [ I (t - 1) - (H (t - 1) - u _ {\\text {r e s e t}}) ]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.667, + 0.682, + 0.684 + ], + "angle": 0, + "content": "\\[\nS (t) = \\Theta (V (t) - v _ {t h})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.686, + 0.906, + 0.715 + ], + "angle": 0, + "content": "\\[\nH (t) = u _ {\\text {r e s e t}} \\cdot S (t) + V (t) \\cdot (1 - S (t)) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.808, + 0.732 + ], + "angle": 0, + "content": "The Heaviside step function \\(\\Theta\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.743, + 0.907, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\Theta (x) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} x < 0 \\\\ 1 & \\text {i f} x \\geq 0. \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Among these, \\( H(t - 1) \\) denotes the membrane potential following spike generation at the previous time step, while \\( I(t) \\) and \\( V(t) \\) represent the input and the updated membrane potential at time step \\( t \\), respectively. Furthermore, \\( v_{th} \\) is the threshold that determines whether \\( V(t) \\) should fire a spike or remain silent, and \\( S(t) \\) indicates the spike sequence generated after triggering the action potential at time step \\( t \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8808" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "SNN Knowledge Distillation. We start from the original SNN Knowledge Distillation (KD) method. To illustrate the process of KD, we consider C-way classification task and denote \\( Z \\in \\mathbb{R}^{C} \\) as the output of the network. Then, the class probability is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.172, + 0.484, + 0.21 + ], + "angle": 0, + "content": "\\[\nQ _ {i} = \\frac {\\exp \\left(Z _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(Z _ {j} / \\tau\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.484, + 0.395 + ], + "angle": 0, + "content": "where \\( Q_{i} \\) and \\( Z_{i} \\) are the probability value on the \\( i \\)-th class and \\( \\tau \\) is the temperature scaling hyper-parameter. In knowledge distillation, \\( \\tau \\) is typically larger than 1.0, which helps control the flatness of the feature distribution to mitigate the phenomenon of overconfidence within the network. When \\( \\tau \\) equals 1.0, the output reverts to the vanilla Softmax output, denoted here as \\( S(\\cdot) \\) to represent this specific case. The objective of KD is to transfer knowledge from a high-performing teacher model to a lightweight student model. For the rescaled outputs, the original KD method achieves distillation by minimizing the KL divergence between the outputs of the teacher and student models." + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.401, + 0.484, + 0.442 + ], + "angle": 0, + "content": "\\[\nK L \\left(Q ^ {t e a} | | Q ^ {s t u}\\right) = \\sum_ {i = 1} ^ {C} Q _ {i} ^ {t e a} \\log \\left(\\frac {Q _ {i} ^ {t e a}}{Q _ {i} ^ {s t u}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.478 + ], + "angle": 0, + "content": "where \\( Q_{i}^{tea} \\) and \\( Q_{i}^{stu} \\) indicates the probability value on the \\( i \\)-th category of the teacher and student output, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.523 + ], + "angle": 0, + "content": "In addition, due to the potential errors in teachers' knowledge, KD also necessitates the use of cross-entropy to enable students to learn the true distribution of labels." + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.527, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nC E \\left(Q ^ {s t u}, y _ {t r u e}\\right) = - \\sum_ {i = 1} ^ {C} Q _ {i} ^ {s t u} \\log \\left(y _ {t r u e}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.573, + 0.32, + 0.587 + ], + "angle": 0, + "content": "where \\( y_{\\text{true}} \\) denote the true labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.723 + ], + "angle": 0, + "content": "The distinction between SNNs and ANNs lies in the fact that SNNs incorporate an additional temporal dimension \\( T \\), which denotes the time step. Consequently, traditional SNN distillation attempts to adapt the distillation methods designed for ANNs by employing a simplistic approach that averages the outputs across the temporal dimension. By introducing a parameter \\( \\alpha \\) to control the weight ratio between true labels and soft labels. The loss calculation formula can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.729, + 0.484, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {S K D} = (1 - \\alpha) \\cdot C E \\left(\\frac {1}{T} \\sum_ {t = 1} ^ {T} Q ^ {s t u} (t), y _ {t r u e}\\right) \\tag {7} \\\\ + \\alpha \\cdot K L (Q ^ {t e a} | | \\frac {1}{T} \\sum_ {t = 1} ^ {T} Q ^ {s t u} (t)) \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.484, + 0.851 + ], + "angle": 0, + "content": "3.2. Temporal Separation Knowledge Distillation with Entropy Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this section, we provide an in-depth explanation of the temporal separation knowledge distillation method with entropy regularization. Our approach consists of two core" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "components: temporal separation strategy and entropy regularization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.144, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Temporal Separation Strategy. As illustrated in Fig. 1a, the original SNN distillation merely transferred the conventional KD methods to SNNs in a naive manner. The use of mean values to integrate outputs across the temporal dimension fails to adequately account for the unique spatiotemporal characteristics of SNNs. Due to the incremental temporal processing in SNNs, the output features at each time step exhibit variability and contain rich information. However, this also implies the potential presence of local outliers in the outputs at different time steps. While the averaging operation mitigates the impact of these anomalies, it poses risks for the subsequent model learning and probabilistic confidence. Consequently, implementing a temporal separation strategy for learning time step's output is crucial." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.356, + 0.906, + 0.448 + ], + "angle": 0, + "content": "As shown in Fig. 1b, we apply temporal separation strategy to CE and KL in Eq. 7, moving the mean operation outside. We directly use the outputs at each time step to distill the learning of the true labels and the teacher distribution. The loss expression for temporal separation knowledge distillation is given by" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.459, + 0.905, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {T S} = (1 - \\alpha) \\cdot \\frac {1}{T} \\sum_ {t = 1} ^ {T} C E \\left(Q ^ {s t u} (t), y _ {t r u e}\\right) \\tag {8} \\\\ + \\alpha \\cdot \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\cdot K L (Q ^ {t e a} | | Q ^ {s t u} (t)). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.579, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Entropy Regularization. After introducing the temporal separation strategy, the soft labels from the teacher network directly impact the output distribution at each time step. This approach, however, can inadvertently reinforce erroneous information from the teacher network, increasing the likelihood that the student model will adopt incorrect knowledge. To counter this, we introduce an entropy regularization term, which constrains and adjusts the teacher's outputs to reduce excessive confidence in potentially erroneous knowledge. The formula for entropy is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.743, + 0.905, + 0.786 + ], + "angle": 0, + "content": "\\[\nH \\left(Q ^ {s t u}\\right) = - \\sum_ {i = 1} ^ {C} Q _ {i} ^ {s t u} \\log \\left(Q _ {i} ^ {s t u}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.792, + 0.905, + 0.822 + ], + "angle": 0, + "content": "Then, we obtain the loss formula for the entropy regularization term:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.833, + 0.905, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E R} = - \\lambda \\cdot \\frac {1}{T} \\sum_ {t = 1} ^ {T} H \\left(Q ^ {s t u} (t)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.71, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the penalty factor." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8809" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "Finally, we integrate the temporal separation strategy, as shown in Eq. 8, with the entropy regularization term from Eq. 10 to derive a new loss function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.149, + 0.483, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T S E R} = \\mathcal {L} _ {T S} + \\mathcal {L} _ {E R}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.176, + 0.485, + 0.207 + ], + "angle": 0, + "content": "And Algorithm. 1 outlines the overall training process of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.22, + 0.216, + 0.236 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.484, + 0.336 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to validate the effectiveness of our method. Initially, we compare our approach with existing SNN distillation methods. Furthermore, we conduct ablation studies, sensitivity analyses, energy consumption assessments, and visualizations to provide a comprehensive evaluation of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.345, + 0.299, + 0.361 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.489 + ], + "angle": 0, + "content": "Our proposed method is evaluated on three datasets. For teacher ANN networks, we select several advanced models, including ResNet-19 and ResNet-34 [21] as well as VGG-16. In contrast, for the student SNN models, we employ neural network architectures with equivalent or fewer layers, specifically ResNet-18/19, and VGG-11/16. Additionally, we conduct multiple experiments by varying the time step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.497, + 0.364, + 0.512 + ], + "angle": 0, + "content": "4.2. Comparison on Static Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.549 + ], + "angle": 0, + "content": "Our method is compared with previous approaches on CIFAR10 and CIFAR100, as shown in the Table. 1. The re" + }, + { + "type": "code_caption", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.603 + ], + "angle": 0, + "content": "Algorithm 1 Training student SNN model with TSER knowledge distillation for one epoch." + }, + { + "type": "code", + "bbox": [ + 0.092, + 0.606, + 0.484, + 0.668 + ], + "angle": 0, + "content": "Require: pre-trained teacher ANN model \\(M_{t}\\); an initialized student SNN model \\(M_{s}\\); input dataset sample \\(X\\); and the true labels \\(y_{true}\\); total training iteration in one epoch \\(I_{train}\\); total validation iteration in one epoch \\(I_{val}\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.094, + 0.668, + 0.342, + 0.68 + ], + "angle": 0, + "content": "Ensure: SNN model with TSER KD" + }, + { + "type": "code", + "bbox": [ + 0.108, + 0.682, + 0.484, + 0.896 + ], + "angle": 0, + "content": "for \\(i = 1\\) to \\(I_{train}\\) do Get mini-batch training data, and true label: \\(X^i,y_{true}^i\\) Compute the \\(M_{s}\\) output \\(Z^{stu,i}(t)\\) of each time step; Compute the \\(M_{t}\\) output \\(Z^{tea,i}\\) Calculate loss function: \\(\\mathcal{L}_{TSER}\\) Backpropagation and update model parameters; \nend for \nfor \\(i = 1\\) to \\(I_{val}\\) do Get mini-batch validation data, and true label: \\(X^i\\) \\(y_{true}^{i}\\) Compute the SNN average output \\(Z_{mean}^{stu,i} =\\) \\(\\frac{1}{T}\\sum_{t = 1}^{T}Z^{stu,i}(t);\\) Compare the \\(Z_{mean}^{stu,i}\\) and \\(y_{true}^{i}\\) for classification; \nend for" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.092, + 0.895, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.338, + 0.907, + 0.422 + ], + "angle": 0, + "content": "Figure 2. Prediction Accuracy Distribution at Different Time Steps for Vanilla SNN KD and Our Proposed Method. The solid red line indicates the teacher model's accuracy, while the dashed lines represent the prediction accuracies of different distillation methods after averaging outputs over time steps. The bars show the prediction accuracies of each distillation method at individual time steps." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.447, + 0.891, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.663, + 0.907, + 0.706 + ], + "angle": 0, + "content": "Figure 3. Accuracy Distribution for Different \\(\\lambda\\) Values. Experiments conducted on the CIFAR100 dataset with a fixed time step of 2, testing various \\(\\lambda\\) values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "sults demonstrate our method surpasses the SOTA performance achieved by prior method across both ResNet and VGG architectures. Specifically, ResNet19 outperform the previous SOTA by \\(0.57\\%\\) and \\(0.41\\%\\) on CIFAR10 and CIFAR100, respectively, achieving accuracies of \\(96.72\\%\\) and \\(81.29\\%\\). This indicates that our approach, although based on logit distillation, performs exceptionally well compared to previous, more complex distillation methods. By addressing the limitations of simple transfer in existing distillation, our framework effectively learns from the teacher's knowledge, minimizing the gap between the student and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.989, + 0.516, + 1.0 + ], + "angle": 0, + "content": "8810" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.088, + 0.916, + 0.543 + ], + "angle": 0, + "content": "
MethodAcc.(%)Arch.Acc.(%)Arch.Acc.(%)Arch.Acc.(%)
CF-100CF-10CF-100CF-10CF-100CF-10CF-100CF-10
without KDTimestep=2Timestep=4
R-1975.4894.44R-1874.6094.15R-1977.0594.85R-1876.4294.67
TET[11]R-1972.8794.16---R-1974.4794.44---
LSG[35]R-1976.3294.41---R-1976.8595.17---
PFA[7]R-1976.7095.60---R-1978.1095.71---
MPBN[18]R-1979.5196.47V-1673.8893.96R-1980.1096.52V-1674.7494.44
IM-LIF[36]R-1977.2195.29---R-1977.4295.66---
Spikformer[64]------S-4-25675.9693.94S-4-38477.8695.19
Spikingformer[61]------Sg-4-25677.4394.77Sg-4-38479.0995.61
SD Transformer [54]------S-2-51278.4095.60---
with KDTeacher ANN Timestep=1
R-3481.1297.10V-1678.0896.06R-3481.1297.10V-1678.0896.06
R-1981.9797.08V-1678.0896.06---V-1678.0896.06
LaSNN [24]Timestep=2Timestep=4
R-1876.1793.64V-1673.8093.90R-1878.1295.09V-1674.9994.49
R-1980.3095.26V-1169.8990.23---V-1170.7490.42
KDSNN [50]R-1877.1695.25V-1674.6594.26R-1878.4695.72V-1675.9894.85
R-1980.8896.15V-1172.1792.81---V-1173.0592.91
BKDSNN [51]R-1873.3093.64V-1673.8094.10R-1875.5795.09V-1674.9994.55
R-1975.7495.26V-1169.8992.73---V-1170.7492.88
OursR-1878.30±0.0895.58±0.08V-1675.81±0.1294.55±0.09R-1879.69±0.0296.18±0.05V-1677.06±0.0495.01±0.10
R-1981.29±0.1496.72±0.02V-1173.67±0.0493.31±0.02---V-1174.65±0.0793.68±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.551, + 0.908, + 0.608 + ], + "angle": 0, + "content": "Table 1. Comparison results with training-based SNN SOTA methods, including CNN- and transformer-based approaches, on CIFAR-10/100, with and without Knowledge Distillation (KD). Acc. denotes accuracy, CF denotes CIFAR, and Archi. denotes architecture. The abbreviations R, V, S, and Sg represent ResNet, VGG, Spikeformer, and Spikingformer architectures, respectively, while SD Transformer refers to the Spike-Driven Transformer. For KD methods, the student model aligns with the architecture of the teacher ANN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.665 + ], + "angle": 0, + "content": "teacher models. This reduction in disparity validates the superiority of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Further comparisons on ImageNet dataset are detailed in Table. 2. The SEW ResNet-18 and SEW ResNet-34 architectures achieve top-1 accuracies of \\(69.24\\%\\) and \\(73.16\\%\\), respectively. These results indicate that our method is capable of delivering competitive performance even within large-scale datasets, effectively harnessing the knowledge from ANNs to enhance the learning capabilities of SNNs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.245, + 0.803 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Temporal Separation Strategies and Entropy Regularization. We investigate different configurations of \\(\\mathcal{L}_{TSER}\\), focusing on the temporal separation applied within the KL and CE losses, as well as the integration of an entropy regularization term. The results are summarized in Table. 3. Our experiments reveal that incorporating tempo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.8 + ], + "angle": 0, + "content": "ral separation strategies notably enhances model accuracy, with improvements of \\(0.49\\%\\) and \\(0.26\\%\\) for the standard KL and CE loss function, respectively. When temporal separation is applied concurrently to both KL and CE losses, the overall accuracy increase reaches \\(0.84\\%\\). Furthermore, we find that optimally adjust the parameter \\(\\lambda\\) in the entropy regularization term, alongside temporal separation, leads to an enhancement of up to \\(0.98\\%\\). These findings underscore the effectiveness of combining temporal separation and entropy regularization in maximizing the performance of SNN distillation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparison of Output Accuracy at Different Time Steps. To illustrate the adaptability of our proposed distillation method for SNNs, we perform a comparative analysis of accuracy across different time steps between vanilla SNN KD and our method. As shown in Fig. 2, the accu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8811" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.488, + 0.465 + ], + "angle": 0, + "content": "
MethodArchitectureTAccuracy
without KDTET [11]R-34664.79%
GLIF [55]R-34467.52%
MPBN [18]R-18463.14%
R-34464.71%
SEW ResNet [15]R-18463.18%
R-34467.04%
Spikformer [64]S-8-384470.24%
S-6-512472.46%
Spikingformer [61]Sg-8-384472.45%
Spike-driven Transformer [54]Sg-8-384472.28%
with KDTeacher ANNSEW R-18171.69%
SEW R-34175.38%
LaSNN [24]SEW R-18463.33%
SEW R-34466.98%
KDSNN [50]SEW R-18463.61%
SEW R-34467.28%
BKDSNN [51]SEW R-18463.43%
SEW R-34467.21%
OursSEW R-18469.24% ±0.19
SEW R-34473.16% ±0.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.473, + 0.487, + 0.547 + ], + "angle": 0, + "content": "Table 2. Comparison results with training-based SNN SOTA methods, including CNN- and Transformer-based approaches, on ImageNet, with and without Knowledge Distillation (KD). T denotes Timestep. R, S, Sg denotes ResNet, Spikeformer and Spikingformer architectures respectively." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.558, + 0.48, + 0.67 + ], + "angle": 0, + "content": "
Temporal SeparationEntropy RegularizationAccuracy
KL LossCE Loss
XXX77.15%
X✓X77.41%
✓XX77.64%
✓✓X77.99%
✓✓✓78.13%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.678, + 0.483, + 0.71 + ], + "angle": 0, + "content": "Table 3. Ablation of Temporal Separation Strategy and Entropy Regularization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.487, + 0.904 + ], + "angle": 0, + "content": "racy discrepancy among time steps for the vanilla SNN KD is \\(0.98\\%\\), notably higher than the \\(0.65\\%\\) difference with our approach. Furthermore, the vanilla SNN KD yields an average output accuracy of \\(74.67\\%\\), whereas our method achieves \\(76.56\\%\\). And as the accuracy of individual time step outputs improved, we observe a significant enhancement in performance following the mean operation on these outputs, aligning with our hypothesis. This clearly demonstrates the effectiveness of our method and highlights the significance of temporal separation in the SNN distillation process." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.091, + 0.887, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.292, + 0.909, + 0.351 + ], + "angle": 0, + "content": "Figure 4. Temperature Coefficient Sensitivity. Our method demonstrates stability across various \\(\\tau\\) hyperparameters. This experiment is conducted on CIFAR100 with ResNet-34 as the teacher model and ResNet-18 as the student model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.909, + 0.558 + ], + "angle": 0, + "content": "Selection of Parameter \\(\\lambda\\). The entropy regularization term is essential for \\(\\mathcal{L}_{TSER}\\), as it partially alleviates the erroneous knowledge transferred by the teacher network. Thus, selecting an optimal value for \\(\\lambda\\) is crucial. We conduct experiments with various values of \\(\\lambda\\), and the results are shown in Fig. 3. Increasing \\(\\lambda\\) enhances the correction of the teacher's inaccuracies but also dilutes the distribution of the correct knowledge. Our findings indicate that maintaining \\(\\lambda\\) within the range of 0.0001 to 0.2 enhances the network's overall training stability. However, excessively high values of \\(\\lambda\\) can result in negative loss values, as described in Eq. 11, which must be avoided." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.702 + ], + "angle": 0, + "content": "Temperature Coefficient Sensitivity. In our experiments, we configure the temperature to \\(\\tau = 5\\). Utilizing the prior experimental setup, we select five temperature points: \\(\\tau = [2,3,4,5,6]\\). We evaluate performance with ResNet-34 as the teacher model and ResNet-18 as the student model on the CIFAR100 dataset. The results depicted in Fig. 4, indicate that our method exhibits stable performance across different temperature coefficients." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.859, + 0.729 + ], + "angle": 0, + "content": "4.4. Performance Analysis and Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Firing Rate Analysis. In Table. 4, the firing rates of our proposed method show a slight increase compared to vanilla SNN KD across both ResNet and VGG architectures. Fig. 2 illustrates that an increase in firing rate of approximately \\(0.1\\%\\) corresponds to a \\(1.48\\%\\) improvement in overall performance. Notably, when comparing our method to other distillation techniques within the ResNet architecture, it maintains performance gains while simultaneously reducing firing rate and increasing spike sparsity, thus minimizing redundant spikes. Although LaSNN exhibits lower firing rates, this comes at the expense of some model per" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "8812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.223, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.091, + 0.361, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.091, + 0.498, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.091, + 0.625, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.092, + 0.761, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.092, + 0.895, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.208, + 0.796, + 0.223 + ], + "angle": 0, + "content": "Figure 5. t-SNE Visualization of features learned by teacher ANN and different distillation methods." + }, + { + "type": "table", + "bbox": [ + 0.169, + 0.235, + 0.832, + 0.398 + ], + "angle": 0, + "content": "
ArchitectureMethodsOPsACsMACsEnergyTraining TimeFiring Rate
ResNet18Vanilla SNN KD2.23G405.33M12.2M0.42mJ2.78s22.85%
KDSNN2.23G411.49M12.2M0.43mJ2.67s23.23%
LaSNN2.23G300.94M12.2M0.33mJ2.79s16.76%
BKDSNN2.23G511.57M12.2M0.52mJ2.84s34.93%
Ours2.23G405.98M12.2M0.42mJ2.70s22.95%
VGG16Vanilla SNN KD1.26G159.96M274.2M1.40mJ2.43s18.91%
KDSNN1.26G170.79M274.2M1.39mJ2.39s18.59%
LaSNN1.26G150.96M274.2M1.41mJ2.44s17.25%
BKDSNN1.26G171.94M274.2M1.40mJ2.58s18.66%
Ours1.26G160.92M274.2M1.41mJ2.40s20.05%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.407, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Table 4. Energy consumption, training time, and spike firing rate under various distillation methods and architectures on the CIFAR100 dataset. Training Time refers to the time required per batch during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.522 + ], + "angle": 0, + "content": "formance. Furthermore, unlike our logit-based approach, LaSNN employs a feature-based method that optimizes the feature map directly, making it more advantageous for feature sparsification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Energy Consumption Analysis. Our energy consumption analysis, based on the energy model in [29], is summarized in Table. 4. The results reveal that the most significant variations in energy consumption occur in Accumulation Computation (AC), which is related to the firing rate. Notably, a lower firing rate correlates with reduced AC. Consequently, our approach does not increase the computational load of SNNs and is competitive regarding energy efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training Time. We evaluate the training times of various competitive KD methods on the CIFAR100 dataset, focusing on the training time per batch. As shown in Table. 4, our method achieves the second-fastest training time, surpassed only by KDSNN. This advantage is largely due to KDSNN simplifying the original KL divergence into cross-entropy in the logit-based framework, while lowering computational cost, also affect accuracy. Our approach outperforms other techniques by utilizing only the logits output for knowledge distillation, thus eliminating the need for additional auxiliary training modules or feature map comparisons. In contrast, previous methods demand more time and resources to extract knowledge from intermediate layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Visualization. We employ t-SNE to visualize features learned from various distillation methods on the CIFAR100 dataset, utilizing ResNet-34 as the teacher and ResNet-18 as the student model. As shown in Fig. 5, our approach significantly enhances the distinguishability of deeper features compared to existing SNN distillation techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.569, + 0.634, + 0.584 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.777 + ], + "angle": 0, + "content": "This work addresses the issue of existing SNN distillation methods that are simply derived from ANN. We propose a new framework specifically designed for SNN distillation, which better exploits the temporal information contained in the logit output of SNNs for the process of knowledge distillation. Specifically, we incorporate a temporal separation strategy and introduce an entropy regularization term into the original distillation method. We aim to uncover the potential of SNN distillation by fully leveraging temporal information and rectifying the erroneous teacher knowledge. Extensive experimental results demonstrate the efficacy of this approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.792, + 0.702, + 0.809 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.907, + 0.889 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China (NSFC) (62476035, 62206037, 62276230), and Natural Science Foundation of Zhejiang Province (LDT23F02023F02), and State Key Laboratory (SKL) of Biobased Transportation Fuel Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8813" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.17 + ], + "angle": 0, + "content": "[1] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.482, + 0.225 + ], + "angle": 0, + "content": "[2] Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision, 113:54-66, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.296 + ], + "angle": 0, + "content": "[3] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[4] Li Chen, Chunyan Yu, and Lvcai Chen. A new knowledge distillation for incremental object detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[5] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4794-4802, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[6] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[7] Haoyu Deng, Ruijie Zhu, Xuerui Qiu, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tensor decomposition based attention module for spiking neural networks. Knowledge-Based Systems, 295:111780, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[9] Lei Deng, Yujie Wu, Xing Hu, Ling Liang, Yufei Ding, Guoqi Li, Guangshe Zhao, Peng Li, and Yuan Xie. Rethinking the performance comparison between snns and anns. Neural networks, 121:294-307, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[10] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[11] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. In International Conference on Learning Representations. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[12] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[13] Jianhao Ding, Tong Bu, Zhaofei Yu, Tiejun Huang, and Jian Liu. Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training. Advances in Neural Information Processing Systems, 35:24780-24793, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Jason K Eshraghian, Max Ward, Emre O Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Ben" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "namoun, Doo Seok Jeong, and Wei D Lu. Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[15] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[16] Isha Garg, Sayeed Shafayet Chowdhury, and Kaushik Roy. Dct-snn: Using dct to distribute spatial information over time for low-latency spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4671-4680, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.264, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[17] Yufei Guo, Weihang Peng, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Xuhui Huang, and Zhe Ma. Joint a-sn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization. Pattern Recognition, 142:109639, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[18] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19420-19430, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.419, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[19] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European conference on computer vision, pages 388-404. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[20] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558-13567, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.56, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[23] Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdelrahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):82-97, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[24] Di Hong, Jiangrong Shen, Yu Qi, and Yueming Wang. Lasnn: Layer-wise ann-to-snn distillation for effective and efficient training in deep spiking neural networks. arXiv preprint arXiv:2304.09101, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.905, + 0.841 + ], + "angle": 0, + "content": "[25] Saeed Reza Kheradpisheh, Mohammad Ganjtabesh, Simon J Thorpe, and Timothée Masquelier. Stdp-based spiking deep convolutional neural networks for object recognition. Neural Networks, 99:56-67, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.844, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8814" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "works. Advances in neural information processing systems, 25, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[28] Souvik Kundu, Gourav Datta, Massoud Pedram, and Peter A Beerel. Spike-thrift: Towards energy-efficient deep spiking neural networks by limiting spiking activity via attention-guided compression. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3953-3962, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[29] Souvik Kundu, Massoud Pedram, and Peter A Beerel. Hiresnn: Harnessing the inherent robustness of energy-efficient deep spiking neural networks by training with crafted input noise. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5209-5218, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[30] Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee, and Rajbabu Velmurugan. Distilling spikes: Knowledge distillation in spiking neural networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4536-4543. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[31] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[32] Dongjin Lee, Seongsik Park, Jongwan Kim, Wuhyeong Doh, and Sungroh Yoon. Energy-efficient knowledge distillation for spiking neural networks. arXiv preprint arXiv:2106.07172, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[33] Chen Li, Lei Ma, and Steve Furber. Quantization framework for fast spiking neural networks. Frontiers in Neuroscience, 16:918793, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[34] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[35] Shuang Lian, Jiangrong Shen, Qianhui Liu, Ziming Wang, Rui Yan, and Huajin Tang. Learnable surrogate gradient for direct training spiking neural networks. In IJCAI, pages 3002-3010, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[36] Shuang Lian, Jiangrong Shen, Ziming Wang, and Huajin Tang. Im-lif: Improved neuronal dynamics with attention mechanism for direct training deep spiking neural network. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[37] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[38] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[39] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36 (6):51-63, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Kleanthis C Neokleous, Marios N Avraamides, Costas K Neocleous, and Christos N Schizas. Selective attention and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "consciousness: investigating their relation through computational modelling. Cognitive Computation, 3:321-331, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[41] Srdjan Ostojic. Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons. Nature neuroscience, 17(4):594–600, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[42] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.904, + 0.376 + ], + "angle": 0, + "content": "[44] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[45] Catherine D Schuman, Shruti R Kulkarni, Maryam Parsa, J Parker Mitchell, Bill Kay, et al. Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science, 2(1):10-19, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[46] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[47] Sugahara Takuya, Renyuan Zhang, and Yasuhiko Nakashima. Training low-latency spiking neural network through knowledge distillation. In 2021 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPPS), pages 1-3. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.618 + ], + "angle": 0, + "content": "[48] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[49] Qi Xu, Yaxin Li, Xuanye Fang, Jiangrong Shen, Jian K. Liu, Huajin Tang, and Gang Pan. Biologically inspired structure learning with reverse knowledge distillation for spiking neural networks, 2023. arXiv:2304.09500 [cs]. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[50] Qi Xu, Yaxin Li, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7886-7895, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[51] Zekai Xu, Kang You, Qinghai Guo, Xiang Wang, and Zhezhi He. Bkdsnn: Enhancing the performance of learning-based spiking neural networks training with blurred knowledge distillation. arXiv preprint arXiv:2407.09083, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[52] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8815" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[53] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. arXiv preprint arXiv:2209.13929, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[54] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[55] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[56] Qiang Yu, Jialu Gao, Jianguo Wei, Jing Li, Kay Chen Tan, and Tiejun Huang. Improving multispike learning with plastic synaptic delays. IEEE Transactions on Neural Networks and Learning Systems, 34(12):10254-10265, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[57] Friedemann Zenke and Tim P Vogels. The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation, 33 (4):899-925, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[58] Friedemann Zenke, Everton J Agnes, and Wulfram Gerstner. Diverse synaptic plasticity mechanisms orchestrated to form and retrieve memories in spiking neural networks. Nature communications, 6(1):6922, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[59] Fengzhao Zhang, Chengting Yu, Hanzhi Ma, Zheming Gu, and Er-ping Li. Knowledge Distillation For Spiking Neural Network. In 2023 5th International Conference on Robotics, Intelligent Control and Artificial Intelligence (RICAL), pages 1015-1020, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[60] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[61] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[62] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Zhengyu Ma, Huihui Zhou, Xiaopeng Fan, and Yonghong Tian. Enhancing the performance of transformer-based spiking neural networks by snn-optimized downsampling with precise gradient backpropagation. arXiv preprint arXiv:2305.05954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.768 + ], + "angle": 0, + "content": "[63] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.482, + 0.839 + ], + "angle": 0, + "content": "[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8816" + } + ] +] \ No newline at end of file diff --git a/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/511efbd9-e72e-4675-94ed-5687d3419119_origin.pdf b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/511efbd9-e72e-4675-94ed-5687d3419119_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cda23d82c5fe4500a21e5cb85455697b13a8d36c --- /dev/null +++ b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/511efbd9-e72e-4675-94ed-5687d3419119_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72be3985749f045dbb42a4afe295a55b838720c392911571a9139656b6a637f5 +size 3587732 diff --git a/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/full.md b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..50739dc02392349cc61f8c8b11a8e517e3deef76 --- /dev/null +++ b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/full.md @@ -0,0 +1,327 @@ +# Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks + +Kairong $\mathrm{Yu}^{1}$ Chengting $\mathrm{Yu}^{1}$ Tianqing Zhang $^{1}$ Xiaochen Zhao $^{1}$ + +Shu Yang $^{1}$ Hongwei Wang $^{1,\dagger}$ Qiang Zhang $^{2}$ Qi Xu $^{2,\dagger}$ + +$^{1}$ Zhejiang University $^{2}$ Dalian University of Technology + +{kairong.22,chengting.21,xiaochen.24，shu.23,hongweiwang}@intlzju.edu.cn + +zhangtianqing@zju.edu.cn + +{xuqi, zhangq}@dlut.edu.cn + +# Abstract + +Spiking Neural Networks (SNNs), inspired by the human brain, offer significant computational efficiency through discrete spike-based information transfer. Despite their potential to reduce inference energy consumption, a performance gap persists between SNNs and Artificial Neural Networks (ANNs), primarily due to current training methods and inherent model limitations. While recent research has aimed to enhance SNN learning by employing knowledge distillation (KD) from ANN teacher networks, traditional distillation techniques often overlook the distinctive spatiotemporal properties of SNNs, thus failing to fully leverage their advantages. To overcome these challenges, we propose a novel logit distillation method characterized by temporal separation and entropy regularization. This approach improves existing SNN distillation techniques by performing distillation learning on logits across different time steps, rather than merely on aggregated output features. Furthermore, the integration of entropy regularization stabilizes model optimization and further boosts the performance. Extensive experimental results indicate that our method surpasses prior SNN distillation strategies, whether based on logit distillation, feature distillation, or a combination of both. Our project is available at https://github.com/yukairong/TSER. + +# 1. Introduction + +Inspired by the neural firing mechanisms observed in biological systems, Spiking Neural Networks (SNNs) [37] are considered a promising alternative to traditional Artificial Neural Networks (ANNs) due to their superior energy efficiency. Unlike ANNs, which use continuous activation val + +ues for information transmission[31], SNNs transmit and process information through discrete spike events [6, 38], with neurons generating spikes only when their membrane potential exceeds a threshold. This binary, event-driven approach allows SNNs to run efficiently on neuromorphic hardware [6, 42], accumulating synaptic inputs effectively and avoiding unnecessary computations related to zero input or activation [9, 14]. Given their event-driven dynamics and the biomimetic properties of spatiotemporal neuron activity [44, 45], SNNs demonstrate remarkable energy efficiency, robust adaptive learning capabilities [13, 41, 58], and ultra-low power consumption, showing significant potential for computational intelligence applications tasks [44]. Despite the inherent advantages of SNNs, their performance in common tasks such as image classification [27], object segmentation [43], and natural language processing [23] still lags behind that of ANNs. This gap largely stems from the limitations of current training methods and structural constraints in SNNs. In contrast to ANNs, SNNs cannot directly utilize backpropagation (BP) for deep network training. Moreover, attempts to directly adapt ANN methodologies to SNNs encounter compatibility issues, preventing SNNs from fully realizing their theoretical advantages. + +Currently, conversion-based and learning-based approaches are two common methods for training SNNs. The former approaches seek to utilize knowledge from ANNs by transferring the parameters of a pre-trained ANN to an equivalent SNN [1, 10, 33, 34]. These methods often necessitate a significant number of time steps to achieve accuracy levels comparable to those of the original ANN [2, 46]. If the number of inference time steps is reduced, the model's capacity for effective information transfer may diminish, resulting in a decline in overall performance. To improve the performance of SNNs and minimize inference time steps, learning-based approaches have been employed + +![](images/fbdd8e30957d4d16d454d025dd355f671b9266c41d592598e579aaba60c4b2d1.jpg) +(a) Classical SNN Knowledge Distillation + +![](images/c4fa6d85a843ee577e7c852e82135e0a4d402f096d610a8aabaf3c7ac875c27c.jpg) +(b) Temporal Separation with Entropy Regularization (TSER) Knowledge Distillation +Figure 1. Illustration of the Classical SNN KD and Our Proposed Method. In our approach, we remove the temporal dimension fusion operation present in classical SNN KD and apply a temporal separation strategy to focus the learning process on outputs at individual time steps. The original calculations of KL Loss and CE Loss are adapted to compute the loss for each time step's output, followed by averaging. Additionally, we incorporate an entropy regularization term to guide the learning direction away from erroneous knowledge. Here, T, B, and C represent the time steps, batch size, and number of classes, respectively. + +to facilitate training during inference. These approaches leverage Backpropagation Through Time (BPTT) for effective training [48, 57]. Motivated by the work [25], several studies have utilized surrogate gradient estimation as a method to address the non-differentiability inherent in SNNs. This foundational research has enabled the direct training of CNNs and transformer-based SNNs, achieving strong performance on static datasets with reduced inference time steps. To further reduce the performance gap with ANNs, recent work has explored the inclusion of modules such as attention mechanisms to enhance SNN capabilities [28, 40, 52, 53]. While effective, these additional modules increase computational consumption, counteracting the inherent low-energy benefits of SNNs. Knowledge Distillation (KD) provides a potential solution, where SNNs are guided by a more capable ANN teacher model, addressing the inference burden associated with these modules. Prior works such as KDSNN [50] and LaSNN [24] employ hierarchical knowledge distillation [5] to facilitate the training of SNNs by leveraging the rich semantic information from the ANN teacher model. Recent approaches like BKDSNN [51] move beyond traditional layer-wise distillation by integrating blurred knowledge distillation techniques with logit distillation, demonstrating significant improvements in performance across both CNNs and Transformer-based models. + +In preliminary experiments, we observed that current SNN distillation architectures are not optimally designed for the unique structure of SNNs. As depicted in Fig. 1a, conventional SNN KD techniques distill the mean of outputs across time steps. However, the inherent nature + +of SNN processing implies that significant information is embedded within the temporal dimension. Employing a mean approach for fusion prediction during classification resembles a voting selection process. Therefore, distributing the distillation of outputs across various time steps during the training phase can substantially alleviate the limitations that currently hinder distillation performance. To mitigate the impact of outliers on network performance and to rectify overly confident incorrect semantics produced by the teacher network, we introduce an entropy regularization term as a constraint. We conduct extensive experimental analyses to assess its efficacy. Furthermore, we evaluate our proposed method on several datasets, including CIFAR10 [26], CIFAR100 [26], and ImageNet [8]. The results of our experiments demonstrate that this method achieves state-of-the-art (SOTA) performance. In summary, our contributions are threefold: + +- We introduce a temporal separation strategy into the SNN distillation method, effectively addressing the incompatibility issues present in previous SNN distillation frameworks and yielding a more effective universal distillation architecture. +- We integrate entropy regularization to correct the overly confident errors from the teacher network, thereby maximizing the performance of the logit-based distillation method. +- We conduct comprehensive evaluations across multiple datasets, showing that our proposed approach outperforms previous state-of-the-art (SOTA) methods. + +# 2. Related Work + +Conversion-based Methods. To fully leverage the knowledge acquired from pre-trained ANNs, conversion methods directly transfer the learned parameters from an ANN to the corresponding SNN, avoiding the difficulties of training SNNs from scratch. Several studies [19, 20] replace the activation functions in ANNs with spiking neurons and introduce optimizations like weight normalization [12] to generate corresponding SNNs. The core idea of these methods is to closely align the outputs of the ANN and SNN, maximizing the utilization of the pre-trained ANN knowledge. While later works [5, 10, 12] achieve nearly lossless accuracy conversion, they often involve long inference times and fail to retain the spatiotemporal characteristics inherent to spike-based processing, significantly limiting the practical applicability of converted SNNs. + +Learning-based Methods. The development of surrogate gradient training methods [56] has resolved the challenge of training non-differentiable SNNs directly. Initially, [39] used a surrogate gradient approach to construct deep SNN models, achieving competitive accuracy on neuromorphic datasets. The DCT-SNN [16] applied frequency domain techniques to reduce the number of inference time steps. To support deeper SNN training, [15] introduced SEW ResNet to address gradient vanishing and exploding issues in Spiking ResNet [60], extending SNNs to architectures exceeding 100 layers. TA-SNN [52] introduced a temporal attention mechanism, paving the way for future attention mechanism advancements in the field. MA-SNN [53] went beyond single-dimensional information enhancement by proposing a temporal-channel-spatial attention mechanism. In Transformer architectures, models like SpikeFormer [61], SpikingFormer [62], and MetaSpikerFormer [63] demonstrated outstanding classification performance through the design and improvement of innovative spike-based self-attention blocks. + +Knowledge Distillation for SNN. Knowledge Distillation (KD), a well-established transfer learning technique, has proven effective across various tasks [3, 4, 22]. Recent studies [17, 30, 32, 47, 59] have applied KD to SNN training, using either large isomorphic SNNs to extract knowledge for smaller SNNs or leveraging the logits from pretrained ANNs to train SNNs. KDSNN [50] utilizes a joint distillation approach based on logits and features, showing effectiveness across multiple datasets. Xu et al. [49] explore biologically inspired structure learning using reverse knowledge distillation. Furthermore, LaSNN [24] proposes a hierarchical feature distillation framework that achieves accuracy comparable to ANNs on the Tiny ImageNet dataset. BKDSNN [51] improves SNN performance + +on complex datasets by using blurred knowledge to replicate ANN features. Despite the advantages of these distillation methods, as shown by the experiments in Fig. 2, current SNN distillation frameworks have yet to fully account for the unique characteristics of SNNs. Developing tailored distillation paradigms could enhance model accuracy in a simple and effective way. + +# 3. Methodology + +In this section, we start from preliminaries, introducing the fundamental concepts of SNNs and SNN distillation. Then we introduce our knowledge distillation approach that incorporates temporal separation with entropy regularization. + +# 3.1. Preliminaries + +Spiking Neuron Model. In SNNs, the fundamental computational unit is the spiking neuron, which serves as an abstract model of the dynamics of biological neurons. Currently, the Leaky Integrate-and-Fire (LIF) model is one of the most commonly used spiking neuron models, as it strikes a balance between simplified mathematical formulation and the complex dynamics of biological neurons. Therefore, in this study, we employ the LIF model as the foundational neuron model for the student SNN. Mathematically, a LIF neuron can be represented by Eq. 1. + +$$ +\mu \frac {d u (t)}{d t} = - (u (t) - u _ {\text {r e s e t}}) + I (t). \tag {1} +$$ + +Where $\mu$ denotes the time constant of the membrane potential, $u(t)$ represents the membrane potential at time $t$ , $u_{\mathrm{reset}}$ is the resting potential of the neuron, and $I(t)$ is the presynaptic input at time $t$ . Based on this differential equation, the discrete-time and iterative mathematical representation of the LIF-SNN can be described as follows: + +$$ +V (t) = H (t - 1) + \frac {1}{\mu} [ I (t - 1) - (H (t - 1) - u _ {\text {r e s e t}}) ] +$$ + +$$ +S (t) = \Theta (V (t) - v _ {t h}) +$$ + +$$ +H (t) = u _ {\text {r e s e t}} \cdot S (t) + V (t) \cdot (1 - S (t)) \tag {2} +$$ + +The Heaviside step function $\Theta$ is defined as: + +$$ +\Theta (x) = \left\{ \begin{array}{l l} 0 & \text {i f} x < 0 \\ 1 & \text {i f} x \geq 0. \end{array} \right. \tag {3} +$$ + +Among these, $H(t - 1)$ denotes the membrane potential following spike generation at the previous time step, while $I(t)$ and $V(t)$ represent the input and the updated membrane potential at time step $t$ , respectively. Furthermore, $v_{th}$ is the threshold that determines whether $V(t)$ should fire a spike or remain silent, and $S(t)$ indicates the spike sequence generated after triggering the action potential at time step $t$ . + +SNN Knowledge Distillation. We start from the original SNN Knowledge Distillation (KD) method. To illustrate the process of KD, we consider C-way classification task and denote $Z \in \mathbb{R}^{C}$ as the output of the network. Then, the class probability is given by: + +$$ +Q _ {i} = \frac {\exp \left(Z _ {i} / \tau\right)}{\sum_ {j = 1} ^ {C} \exp \left(Z _ {j} / \tau\right)}, \tag {4} +$$ + +where $Q_{i}$ and $Z_{i}$ are the probability value on the $i$ -th class and $\tau$ is the temperature scaling hyper-parameter. In knowledge distillation, $\tau$ is typically larger than 1.0, which helps control the flatness of the feature distribution to mitigate the phenomenon of overconfidence within the network. When $\tau$ equals 1.0, the output reverts to the vanilla Softmax output, denoted here as $S(\cdot)$ to represent this specific case. The objective of KD is to transfer knowledge from a high-performing teacher model to a lightweight student model. For the rescaled outputs, the original KD method achieves distillation by minimizing the KL divergence between the outputs of the teacher and student models. + +$$ +K L \left(Q ^ {t e a} | | Q ^ {s t u}\right) = \sum_ {i = 1} ^ {C} Q _ {i} ^ {t e a} \log \left(\frac {Q _ {i} ^ {t e a}}{Q _ {i} ^ {s t u}}\right), \tag {5} +$$ + +where $Q_{i}^{tea}$ and $Q_{i}^{stu}$ indicates the probability value on the $i$ -th category of the teacher and student output, respectively. + +In addition, due to the potential errors in teachers' knowledge, KD also necessitates the use of cross-entropy to enable students to learn the true distribution of labels. + +$$ +C E \left(Q ^ {s t u}, y _ {t r u e}\right) = - \sum_ {i = 1} ^ {C} Q _ {i} ^ {s t u} \log \left(y _ {t r u e}\right) \tag {6} +$$ + +where $y_{\text{true}}$ denote the true labels. + +The distinction between SNNs and ANNs lies in the fact that SNNs incorporate an additional temporal dimension $T$ , which denotes the time step. Consequently, traditional SNN distillation attempts to adapt the distillation methods designed for ANNs by employing a simplistic approach that averages the outputs across the temporal dimension. By introducing a parameter $\alpha$ to control the weight ratio between true labels and soft labels. The loss calculation formula can be expressed as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {S K D} = (1 - \alpha) \cdot C E \left(\frac {1}{T} \sum_ {t = 1} ^ {T} Q ^ {s t u} (t), y _ {t r u e}\right) \tag {7} \\ + \alpha \cdot K L (Q ^ {t e a} | | \frac {1}{T} \sum_ {t = 1} ^ {T} Q ^ {s t u} (t)) \\ \end{array} +$$ + +# 3.2. Temporal Separation Knowledge Distillation with Entropy Regularization + +In this section, we provide an in-depth explanation of the temporal separation knowledge distillation method with entropy regularization. Our approach consists of two core + +components: temporal separation strategy and entropy regularization. + +Temporal Separation Strategy. As illustrated in Fig. 1a, the original SNN distillation merely transferred the conventional KD methods to SNNs in a naive manner. The use of mean values to integrate outputs across the temporal dimension fails to adequately account for the unique spatiotemporal characteristics of SNNs. Due to the incremental temporal processing in SNNs, the output features at each time step exhibit variability and contain rich information. However, this also implies the potential presence of local outliers in the outputs at different time steps. While the averaging operation mitigates the impact of these anomalies, it poses risks for the subsequent model learning and probabilistic confidence. Consequently, implementing a temporal separation strategy for learning time step's output is crucial. + +As shown in Fig. 1b, we apply temporal separation strategy to CE and KL in Eq. 7, moving the mean operation outside. We directly use the outputs at each time step to distill the learning of the true labels and the teacher distribution. The loss expression for temporal separation knowledge distillation is given by + +$$ +\begin{array}{l} \mathcal {L} _ {T S} = (1 - \alpha) \cdot \frac {1}{T} \sum_ {t = 1} ^ {T} C E \left(Q ^ {s t u} (t), y _ {t r u e}\right) \tag {8} \\ + \alpha \cdot \frac {1}{T} \sum_ {t = 1} ^ {T} \cdot K L (Q ^ {t e a} | | Q ^ {s t u} (t)). \\ \end{array} +$$ + +Entropy Regularization. After introducing the temporal separation strategy, the soft labels from the teacher network directly impact the output distribution at each time step. This approach, however, can inadvertently reinforce erroneous information from the teacher network, increasing the likelihood that the student model will adopt incorrect knowledge. To counter this, we introduce an entropy regularization term, which constrains and adjusts the teacher's outputs to reduce excessive confidence in potentially erroneous knowledge. The formula for entropy is as follows: + +$$ +H \left(Q ^ {s t u}\right) = - \sum_ {i = 1} ^ {C} Q _ {i} ^ {s t u} \log \left(Q _ {i} ^ {s t u}\right). \tag {9} +$$ + +Then, we obtain the loss formula for the entropy regularization term: + +$$ +\mathcal {L} _ {E R} = - \lambda \cdot \frac {1}{T} \sum_ {t = 1} ^ {T} H \left(Q ^ {s t u} (t)\right), \tag {10} +$$ + +where $\lambda$ is the penalty factor. + +Finally, we integrate the temporal separation strategy, as shown in Eq. 8, with the entropy regularization term from Eq. 10 to derive a new loss function as follows: + +$$ +\mathcal {L} _ {T S E R} = \mathcal {L} _ {T S} + \mathcal {L} _ {E R}. \tag {11} +$$ + +And Algorithm. 1 outlines the overall training process of our proposed method. + +# 4. Experiment + +In this section, we conduct extensive experiments to validate the effectiveness of our method. Initially, we compare our approach with existing SNN distillation methods. Furthermore, we conduct ablation studies, sensitivity analyses, energy consumption assessments, and visualizations to provide a comprehensive evaluation of our method. + +# 4.1. Experimental Settings + +Our proposed method is evaluated on three datasets. For teacher ANN networks, we select several advanced models, including ResNet-19 and ResNet-34 [21] as well as VGG-16. In contrast, for the student SNN models, we employ neural network architectures with equivalent or fewer layers, specifically ResNet-18/19, and VGG-11/16. Additionally, we conduct multiple experiments by varying the time step. + +# 4.2. Comparison on Static Datasets + +Our method is compared with previous approaches on CIFAR10 and CIFAR100, as shown in the Table. 1. The re + +Algorithm 1 Training student SNN model with TSER knowledge distillation for one epoch. +Ensure: SNN model with TSER KD +```txt +Require: pre-trained teacher ANN model $M_{t}$ ; an initialized student SNN model $M_{s}$ ; input dataset sample $X$ ; and the true labels $y_{true}$ ; total training iteration in one epoch $I_{train}$ ; total validation iteration in one epoch $I_{val}$ +``` + +```txt +for $i = 1$ to $I_{train}$ do Get mini-batch training data, and true label: $X^i,y_{true}^i$ Compute the $M_{s}$ output $Z^{stu,i}(t)$ of each time step; Compute the $M_{t}$ output $Z^{tea,i}$ Calculate loss function: $\mathcal{L}_{TSER}$ Backpropagation and update model parameters; +end for +for $i = 1$ to $I_{val}$ do Get mini-batch validation data, and true label: $X^i$ $y_{true}^{i}$ Compute the SNN average output $Z_{mean}^{stu,i} =$ $\frac{1}{T}\sum_{t = 1}^{T}Z^{stu,i}(t);$ Compare the $Z_{mean}^{stu,i}$ and $y_{true}^{i}$ for classification; +end for +``` + +![](images/93817fae9dd3b32330078722e85b57ea777ccee5f5d851884c83b955275a7e96.jpg) +Figure 2. Prediction Accuracy Distribution at Different Time Steps for Vanilla SNN KD and Our Proposed Method. The solid red line indicates the teacher model's accuracy, while the dashed lines represent the prediction accuracies of different distillation methods after averaging outputs over time steps. The bars show the prediction accuracies of each distillation method at individual time steps. + +![](images/1393a2f92084cee688a2bec85335b14cadb8dc0500647d2c5e29e6a0d3650b66.jpg) +Figure 3. Accuracy Distribution for Different $\lambda$ Values. Experiments conducted on the CIFAR100 dataset with a fixed time step of 2, testing various $\lambda$ values. + +sults demonstrate our method surpasses the SOTA performance achieved by prior method across both ResNet and VGG architectures. Specifically, ResNet19 outperform the previous SOTA by $0.57\%$ and $0.41\%$ on CIFAR10 and CIFAR100, respectively, achieving accuracies of $96.72\%$ and $81.29\%$ . This indicates that our approach, although based on logit distillation, performs exceptionally well compared to previous, more complex distillation methods. By addressing the limitations of simple transfer in existing distillation, our framework effectively learns from the teacher's knowledge, minimizing the gap between the student and + +
MethodAcc.(%)Arch.Acc.(%)Arch.Acc.(%)Arch.Acc.(%)
CF-100CF-10CF-100CF-10CF-100CF-10CF-100CF-10
without KDTimestep=2Timestep=4
R-1975.4894.44R-1874.6094.15R-1977.0594.85R-1876.4294.67
TET[11]R-1972.8794.16---R-1974.4794.44---
LSG[35]R-1976.3294.41---R-1976.8595.17---
PFA[7]R-1976.7095.60---R-1978.1095.71---
MPBN[18]R-1979.5196.47V-1673.8893.96R-1980.1096.52V-1674.7494.44
IM-LIF[36]R-1977.2195.29---R-1977.4295.66---
Spikformer[64]------S-4-25675.9693.94S-4-38477.8695.19
Spikingformer[61]------Sg-4-25677.4394.77Sg-4-38479.0995.61
SD Transformer [54]------S-2-51278.4095.60---
with KDTeacher ANN Timestep=1
R-3481.1297.10V-1678.0896.06R-3481.1297.10V-1678.0896.06
R-1981.9797.08V-1678.0896.06---V-1678.0896.06
LaSNN [24]Timestep=2Timestep=4
R-1876.1793.64V-1673.8093.90R-1878.1295.09V-1674.9994.49
R-1980.3095.26V-1169.8990.23---V-1170.7490.42
KDSNN [50]R-1877.1695.25V-1674.6594.26R-1878.4695.72V-1675.9894.85
R-1980.8896.15V-1172.1792.81---V-1173.0592.91
BKDSNN [51]R-1873.3093.64V-1673.8094.10R-1875.5795.09V-1674.9994.55
R-1975.7495.26V-1169.8992.73---V-1170.7492.88
OursR-1878.30±0.0895.58±0.08V-1675.81±0.1294.55±0.09R-1879.69±0.0296.18±0.05V-1677.06±0.0495.01±0.10
R-1981.29±0.1496.72±0.02V-1173.67±0.0493.31±0.02---V-1174.65±0.0793.68±0.02
+ +Table 1. Comparison results with training-based SNN SOTA methods, including CNN- and transformer-based approaches, on CIFAR-10/100, with and without Knowledge Distillation (KD). Acc. denotes accuracy, CF denotes CIFAR, and Archi. denotes architecture. The abbreviations R, V, S, and Sg represent ResNet, VGG, Spikeformer, and Spikingformer architectures, respectively, while SD Transformer refers to the Spike-Driven Transformer. For KD methods, the student model aligns with the architecture of the teacher ANN. + +teacher models. This reduction in disparity validates the superiority of our method. + +Further comparisons on ImageNet dataset are detailed in Table. 2. The SEW ResNet-18 and SEW ResNet-34 architectures achieve top-1 accuracies of $69.24\%$ and $73.16\%$ , respectively. These results indicate that our method is capable of delivering competitive performance even within large-scale datasets, effectively harnessing the knowledge from ANNs to enhance the learning capabilities of SNNs. + +# 4.3. Ablation Study + +Temporal Separation Strategies and Entropy Regularization. We investigate different configurations of $\mathcal{L}_{TSER}$ , focusing on the temporal separation applied within the KL and CE losses, as well as the integration of an entropy regularization term. The results are summarized in Table. 3. Our experiments reveal that incorporating tempo + +ral separation strategies notably enhances model accuracy, with improvements of $0.49\%$ and $0.26\%$ for the standard KL and CE loss function, respectively. When temporal separation is applied concurrently to both KL and CE losses, the overall accuracy increase reaches $0.84\%$ . Furthermore, we find that optimally adjust the parameter $\lambda$ in the entropy regularization term, alongside temporal separation, leads to an enhancement of up to $0.98\%$ . These findings underscore the effectiveness of combining temporal separation and entropy regularization in maximizing the performance of SNN distillation. + +Comparison of Output Accuracy at Different Time Steps. To illustrate the adaptability of our proposed distillation method for SNNs, we perform a comparative analysis of accuracy across different time steps between vanilla SNN KD and our method. As shown in Fig. 2, the accu + +
MethodArchitectureTAccuracy
without KDTET [11]R-34664.79%
GLIF [55]R-34467.52%
MPBN [18]R-18463.14%
R-34464.71%
SEW ResNet [15]R-18463.18%
R-34467.04%
Spikformer [64]S-8-384470.24%
S-6-512472.46%
Spikingformer [61]Sg-8-384472.45%
Spike-driven Transformer [54]Sg-8-384472.28%
with KDTeacher ANNSEW R-18171.69%
SEW R-34175.38%
LaSNN [24]SEW R-18463.33%
SEW R-34466.98%
KDSNN [50]SEW R-18463.61%
SEW R-34467.28%
BKDSNN [51]SEW R-18463.43%
SEW R-34467.21%
OursSEW R-18469.24% ±0.19
SEW R-34473.16% ±0.15
+ +Table 2. Comparison results with training-based SNN SOTA methods, including CNN- and Transformer-based approaches, on ImageNet, with and without Knowledge Distillation (KD). T denotes Timestep. R, S, Sg denotes ResNet, Spikeformer and Spikingformer architectures respectively. + +
Temporal SeparationEntropy RegularizationAccuracy
KL LossCE Loss
XXX77.15%
X✓X77.41%
✓XX77.64%
✓✓X77.99%
✓✓✓78.13%
+ +Table 3. Ablation of Temporal Separation Strategy and Entropy Regularization. + +racy discrepancy among time steps for the vanilla SNN KD is $0.98\%$ , notably higher than the $0.65\%$ difference with our approach. Furthermore, the vanilla SNN KD yields an average output accuracy of $74.67\%$ , whereas our method achieves $76.56\%$ . And as the accuracy of individual time step outputs improved, we observe a significant enhancement in performance following the mean operation on these outputs, aligning with our hypothesis. This clearly demonstrates the effectiveness of our method and highlights the significance of temporal separation in the SNN distillation process. + +![](images/3964209d0776f37ed9eb84592b1349dc506c2a179fecfedc01f731e956bbae19.jpg) +Figure 4. Temperature Coefficient Sensitivity. Our method demonstrates stability across various $\tau$ hyperparameters. This experiment is conducted on CIFAR100 with ResNet-34 as the teacher model and ResNet-18 as the student model. + +Selection of Parameter $\lambda$ . The entropy regularization term is essential for $\mathcal{L}_{TSER}$ , as it partially alleviates the erroneous knowledge transferred by the teacher network. Thus, selecting an optimal value for $\lambda$ is crucial. We conduct experiments with various values of $\lambda$ , and the results are shown in Fig. 3. Increasing $\lambda$ enhances the correction of the teacher's inaccuracies but also dilutes the distribution of the correct knowledge. Our findings indicate that maintaining $\lambda$ within the range of 0.0001 to 0.2 enhances the network's overall training stability. However, excessively high values of $\lambda$ can result in negative loss values, as described in Eq. 11, which must be avoided. + +Temperature Coefficient Sensitivity. In our experiments, we configure the temperature to $\tau = 5$ . Utilizing the prior experimental setup, we select five temperature points: $\tau = [2,3,4,5,6]$ . We evaluate performance with ResNet-34 as the teacher model and ResNet-18 as the student model on the CIFAR100 dataset. The results depicted in Fig. 4, indicate that our method exhibits stable performance across different temperature coefficients. + +# 4.4. Performance Analysis and Visualization + +Firing Rate Analysis. In Table. 4, the firing rates of our proposed method show a slight increase compared to vanilla SNN KD across both ResNet and VGG architectures. Fig. 2 illustrates that an increase in firing rate of approximately $0.1\%$ corresponds to a $1.48\%$ improvement in overall performance. Notably, when comparing our method to other distillation techniques within the ResNet architecture, it maintains performance gains while simultaneously reducing firing rate and increasing spike sparsity, thus minimizing redundant spikes. Although LaSNN exhibits lower firing rates, this comes at the expense of some model per + +![](images/3d4af674e592612b5ad7c6da632b259beee57c74f8fb33a5f4a9629e284e147f.jpg) +Figure 5. t-SNE Visualization of features learned by teacher ANN and different distillation methods. + +![](images/054805f500d4ff747523588d7f021f2097d9cd41c8e181604307ab2f882f7a48.jpg) + +![](images/f1db2d30c835228bf392eaea766568a0644916da3776b17666313c415c128354.jpg) + +![](images/06b70e903b440527740feb3966b35720be5ebfd20991e8ce9ed05db5cdf04edc.jpg) + +![](images/3a5f7fc37f48a03e2d52282fb9b496fd760278a4574a9876e8b46e6a3320c49c.jpg) + +![](images/2e2eb4b5e824a4531ffadb701a9ad7a6d8eaf902270627046d3cd5d96d9caf2f.jpg) + +
ArchitectureMethodsOPsACsMACsEnergyTraining TimeFiring Rate
ResNet18Vanilla SNN KD2.23G405.33M12.2M0.42mJ2.78s22.85%
KDSNN2.23G411.49M12.2M0.43mJ2.67s23.23%
LaSNN2.23G300.94M12.2M0.33mJ2.79s16.76%
BKDSNN2.23G511.57M12.2M0.52mJ2.84s34.93%
Ours2.23G405.98M12.2M0.42mJ2.70s22.95%
VGG16Vanilla SNN KD1.26G159.96M274.2M1.40mJ2.43s18.91%
KDSNN1.26G170.79M274.2M1.39mJ2.39s18.59%
LaSNN1.26G150.96M274.2M1.41mJ2.44s17.25%
BKDSNN1.26G171.94M274.2M1.40mJ2.58s18.66%
Ours1.26G160.92M274.2M1.41mJ2.40s20.05%
+ +Table 4. Energy consumption, training time, and spike firing rate under various distillation methods and architectures on the CIFAR100 dataset. Training Time refers to the time required per batch during training. + +formance. Furthermore, unlike our logit-based approach, LaSNN employs a feature-based method that optimizes the feature map directly, making it more advantageous for feature sparsification. + +Energy Consumption Analysis. Our energy consumption analysis, based on the energy model in [29], is summarized in Table. 4. The results reveal that the most significant variations in energy consumption occur in Accumulation Computation (AC), which is related to the firing rate. Notably, a lower firing rate correlates with reduced AC. Consequently, our approach does not increase the computational load of SNNs and is competitive regarding energy efficiency. + +Training Time. We evaluate the training times of various competitive KD methods on the CIFAR100 dataset, focusing on the training time per batch. As shown in Table. 4, our method achieves the second-fastest training time, surpassed only by KDSNN. This advantage is largely due to KDSNN simplifying the original KL divergence into cross-entropy in the logit-based framework, while lowering computational cost, also affect accuracy. Our approach outperforms other techniques by utilizing only the logits output for knowledge distillation, thus eliminating the need for additional auxiliary training modules or feature map comparisons. In contrast, previous methods demand more time and resources to extract knowledge from intermediate layers. + +Visualization. We employ t-SNE to visualize features learned from various distillation methods on the CIFAR100 dataset, utilizing ResNet-34 as the teacher and ResNet-18 as the student model. As shown in Fig. 5, our approach significantly enhances the distinguishability of deeper features compared to existing SNN distillation techniques. + +# 5. Conclusion + +This work addresses the issue of existing SNN distillation methods that are simply derived from ANN. We propose a new framework specifically designed for SNN distillation, which better exploits the temporal information contained in the logit output of SNNs for the process of knowledge distillation. Specifically, we incorporate a temporal separation strategy and introduce an entropy regularization term into the original distillation method. We aim to uncover the potential of SNN distillation by fully leveraging temporal information and rectifying the erroneous teacher knowledge. Extensive experimental results demonstrate the efficacy of this approach. + +# 6. Acknowledgements + +This work was supported in part by National Natural Science Foundation of China (NSFC) (62476035, 62206037, 62276230), and Natural Science Foundation of Zhejiang Province (LDT23F02023F02), and State Key Laboratory (SKL) of Biobased Transportation Fuel Technology. + +# References + +[1] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347, 2023. 1 +[2] Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision, 113:54-66, 2015. 1 +[3] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 3 +[4] Li Chen, Chunyan Yu, and Lvcai Chen. A new knowledge distillation for incremental object detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2019. 3 +[5] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4794-4802, 2019. 2, 3 +[6] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018. 1 +[7] Haoyu Deng, Ruijie Zhu, Xuerui Qiu, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tensor decomposition based attention module for spiking neural networks. Knowledge-Based Systems, 295:111780, 2024. 6 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2 +[9] Lei Deng, Yujie Wu, Xing Hu, Ling Liang, Yufei Ding, Guoqi Li, Guangshe Zhao, Peng Li, and Yuan Xie. Rethinking the performance comparison between snns and anns. Neural networks, 121:294-307, 2020. 1 +[10] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476, 2021. 1, 3 +[11] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. In International Conference on Learning Representations. 6, 7 +[12] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2015. 3 +[13] Jianhao Ding, Tong Bu, Zhaofei Yu, Tiejun Huang, and Jian Liu. Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training. Advances in Neural Information Processing Systems, 35:24780-24793, 2022. 1 +[14] Jason K Eshraghian, Max Ward, Emre O Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Ben + +namoun, Doo Seok Jeong, and Wei D Lu. Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE, 2023. 1 +[15] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 3, 7 +[16] Isha Garg, Sayeed Shafayet Chowdhury, and Kaushik Roy. Dct-snn: Using dct to distribute spatial information over time for low-latency spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4671-4680, 2021. 3 +[17] Yufei Guo, Weihang Peng, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Xuhui Huang, and Zhe Ma. Joint a-sn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization. Pattern Recognition, 142:109639, 2023. 3 +[18] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19420-19430, 2023. 6, 7 +[19] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European conference on computer vision, pages 388-404. Springer, 2020. 3 +[20] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558-13567, 2020. 3 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[23] Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdelrahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):82-97, 2012. 1 +[24] Di Hong, Jiangrong Shen, Yu Qi, and Yueming Wang. Lasnn: Layer-wise ann-to-snn distillation for effective and efficient training in deep spiking neural networks. arXiv preprint arXiv:2304.09101, 2023. 2, 3, 6, 7 +[25] Saeed Reza Kheradpisheh, Mohammad Ganjtabesh, Simon J Thorpe, and Timothée Masquelier. Stdp-based spiking deep convolutional neural networks for object recognition. Neural Networks, 99:56-67, 2018. 2 +[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2 +[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net + +works. Advances in neural information processing systems, 25, 2012. 1 +[28] Souvik Kundu, Gourav Datta, Massoud Pedram, and Peter A Beerel. Spike-thrift: Towards energy-efficient deep spiking neural networks by limiting spiking activity via attention-guided compression. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3953-3962, 2021. 2 +[29] Souvik Kundu, Massoud Pedram, and Peter A Beerel. Hiresnn: Harnessing the inherent robustness of energy-efficient deep spiking neural networks by training with crafted input noise. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5209-5218, 2021. 8 +[30] Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee, and Rajbabu Velmurugan. Distilling spikes: Knowledge distillation in spiking neural networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4536-4543. IEEE, 2021. 3 +[31] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1 +[32] Dongjin Lee, Seongsik Park, Jongwan Kim, Wuhyeong Doh, and Sungroh Yoon. Energy-efficient knowledge distillation for spiking neural networks. arXiv preprint arXiv:2106.07172, 2021. 3 +[33] Chen Li, Lei Ma, and Steve Furber. Quantization framework for fast spiking neural networks. Frontiers in Neuroscience, 16:918793, 2022. 1 +[34] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021. 1 +[35] Shuang Lian, Jiangrong Shen, Qianhui Liu, Ziming Wang, Rui Yan, and Huajin Tang. Learnable surrogate gradient for direct training spiking neural networks. In IJCAI, pages 3002-3010, 2023. 6 +[36] Shuang Lian, Jiangrong Shen, Ziming Wang, and Huajin Tang. Im-lif: Improved neuronal dynamics with attention mechanism for direct training deep spiking neural network. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 6 +[37] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1 +[38] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014. 1 +[39] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36 (6):51-63, 2019. 3 +[40] Kleanthis C Neokleous, Marios N Avraamides, Costas K Neocleous, and Christos N Schizas. Selective attention and + +consciousness: investigating their relation through computational modelling. Cognitive Computation, 3:321-331, 2011. 2 +[41] Srdjan Ostojic. Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons. Nature neuroscience, 17(4):594–600, 2014. 1 +[42] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 1 +[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1 +[44] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1 +[45] Catherine D Schuman, Shruti R Kulkarni, Maryam Parsa, J Parker Mitchell, Bill Kay, et al. Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science, 2(1):10-19, 2022. 1 +[46] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 1 +[47] Sugahara Takuya, Renyuan Zhang, and Yasuhiko Nakashima. Training low-latency spiking neural network through knowledge distillation. In 2021 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPPS), pages 1-3. IEEE, 2021. 3 +[48] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 2 +[49] Qi Xu, Yaxin Li, Xuanye Fang, Jiangrong Shen, Jian K. Liu, Huajin Tang, and Gang Pan. Biologically inspired structure learning with reverse knowledge distillation for spiking neural networks, 2023. arXiv:2304.09500 [cs]. 3 +[50] Qi Xu, Yaxin Li, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7886-7895, 2023. 2, 3, 6, 7 +[51] Zekai Xu, Kang You, Qinghai Guo, Xiang Wang, and Zhezhi He. Bkdsnn: Enhancing the performance of learning-based spiking neural networks training with blurred knowledge distillation. arXiv preprint arXiv:2407.09083, 2024. 2, 3, 6, 7 +[52] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 2, 3 + +[53] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. arXiv preprint arXiv:2209.13929, 2022. 2, 3 +[54] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 6, 7 +[55] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 6, 7 +[56] Qiang Yu, Jialu Gao, Jianguo Wei, Jing Li, Kay Chen Tan, and Tiejun Huang. Improving multispike learning with plastic synaptic delays. IEEE Transactions on Neural Networks and Learning Systems, 34(12):10254-10265, 2022. 3 +[57] Friedemann Zenke and Tim P Vogels. The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation, 33 (4):899-925, 2021. 2 +[58] Friedemann Zenke, Everton J Agnes, and Wulfram Gerstner. Diverse synaptic plasticity mechanisms orchestrated to form and retrieve memories in spiking neural networks. Nature communications, 6(1):6922, 2015. 1 +[59] Fengzhao Zhang, Chengting Yu, Hanzhi Ma, Zheming Gu, and Er-ping Li. Knowledge Distillation For Spiking Neural Network. In 2023 5th International Conference on Robotics, Intelligent Control and Artificial Intelligence (RICAL), pages 1015-1020, 2023. 3 +[60] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 3 +[61] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 3, 6, 7 +[62] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Zhengyu Ma, Huihui Zhou, Xiaopeng Fan, and Yonghong Tian. Enhancing the performance of transformer-based spiking neural networks by snn-optimized downsampling with precise gradient backpropagation. arXiv preprint arXiv:2305.05954, 2023. 3 +[63] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 3 +[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations. 6, 7 \ No newline at end of file diff --git a/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/images.zip b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fc0cf0e90213607ea6ab3bc7adf94d7712eb6985 --- /dev/null +++ b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3c814a469efdffa581eb17d2697891eb19aa3e03e63adcdc1826adbc4c73603 +size 659118 diff --git a/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/layout.json b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..128e3031e09147adfe51dd0dc955caf11b98824a --- /dev/null +++ b/2025/Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks/layout.json @@ -0,0 +1,8750 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "type": "text", + "content": "Temporal Separation with Entropy Regularization for Knowledge Distillation in Spiking Neural Networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "spans": [ + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "type": "text", + "content": "Kairong " + }, + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{1}" + }, + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "type": "text", + "content": " Chengting " + }, + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{1}" + }, + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "type": "text", + "content": " Tianqing Zhang" + }, + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "type": "text", + "content": " Xiaochen Zhao" + }, + { + "bbox": [ + 118, + 159, + 491, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "type": "text", + "content": "Shu Yang" + }, + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "type": "text", + "content": " Hongwei Wang" + }, + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "type": "text", + "content": " Qiang Zhang" + }, + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "type": "text", + "content": " Qi Xu" + }, + { + "bbox": [ + 158, + 175, + 480, + 190 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 193, + 452, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 193, + 452, + 209 + ], + "spans": [ + { + "bbox": [ + 159, + 193, + 452, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 193, + 452, + 209 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 159, + 193, + 452, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 193, + 452, + 209 + ], + "type": "text", + "content": " Dalian University of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 210, + 507, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 210, + 507, + 222 + ], + "spans": [ + { + "bbox": [ + 97, + 210, + 507, + 222 + ], + "type": "text", + "content": "{kairong.22,chengting.21,xiaochen.24，shu.23,hongweiwang}@intlzju.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 237, + 225, + 368, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 225, + 368, + 236 + ], + "spans": [ + { + "bbox": [ + 237, + 225, + 368, + 236 + ], + "type": "text", + "content": "zhangtianqing@zju.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 231, + 238, + 371, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 238, + 371, + 250 + ], + "spans": [ + { + "bbox": [ + 231, + 238, + 371, + 250 + ], + "type": "text", + "content": "{xuqi, zhangq}@dlut.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 277, + 200, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 277, + 200, + 290 + ], + "spans": [ + { + "bbox": [ + 152, + 277, + 200, + 290 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 303, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 296, + 590 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs), inspired by the human brain, offer significant computational efficiency through discrete spike-based information transfer. Despite their potential to reduce inference energy consumption, a performance gap persists between SNNs and Artificial Neural Networks (ANNs), primarily due to current training methods and inherent model limitations. While recent research has aimed to enhance SNN learning by employing knowledge distillation (KD) from ANN teacher networks, traditional distillation techniques often overlook the distinctive spatiotemporal properties of SNNs, thus failing to fully leverage their advantages. To overcome these challenges, we propose a novel logit distillation method characterized by temporal separation and entropy regularization. This approach improves existing SNN distillation techniques by performing distillation learning on logits across different time steps, rather than merely on aggregated output features. Furthermore, the integration of entropy regularization stabilizes model optimization and further boosts the performance. Extensive experimental results indicate that our method surpasses prior SNN distillation strategies, whether based on logit distillation, feature distillation, or a combination of both. Our project is available at https://github.com/yukairong/TSER." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 614, + 135, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 135, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 135, + 626 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 634, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 295, + 694 + ], + "type": "text", + "content": "Inspired by the neural firing mechanisms observed in biological systems, Spiking Neural Networks (SNNs) [37] are considered a promising alternative to traditional Artificial Neural Networks (ANNs) due to their superior energy efficiency. Unlike ANNs, which use continuous activation val" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 278, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 566 + ], + "type": "text", + "content": "ues for information transmission[31], SNNs transmit and process information through discrete spike events [6, 38], with neurons generating spikes only when their membrane potential exceeds a threshold. This binary, event-driven approach allows SNNs to run efficiently on neuromorphic hardware [6, 42], accumulating synaptic inputs effectively and avoiding unnecessary computations related to zero input or activation [9, 14]. Given their event-driven dynamics and the biomimetic properties of spatiotemporal neuron activity [44, 45], SNNs demonstrate remarkable energy efficiency, robust adaptive learning capabilities [13, 41, 58], and ultra-low power consumption, showing significant potential for computational intelligence applications tasks [44]. Despite the inherent advantages of SNNs, their performance in common tasks such as image classification [27], object segmentation [43], and natural language processing [23] still lags behind that of ANNs. This gap largely stems from the limitations of current training methods and structural constraints in SNNs. In contrast to ANNs, SNNs cannot directly utilize backpropagation (BP) for deep network training. Moreover, attempts to directly adapt ANN methodologies to SNNs encounter compatibility issues, preventing SNNs from fully realizing their theoretical advantages." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Currently, conversion-based and learning-based approaches are two common methods for training SNNs. The former approaches seek to utilize knowledge from ANNs by transferring the parameters of a pre-trained ANN to an equivalent SNN [1, 10, 33, 34]. These methods often necessitate a significant number of time steps to achieve accuracy levels comparable to those of the original ANN [2, 46]. If the number of inference time steps is reduced, the model's capacity for effective information transfer may diminish, resulting in a decline in overall performance. To improve the performance of SNNs and minimize inference time steps, learning-based approaches have been employed" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 148, + 713 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8806" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 298, + 235 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 298, + 235 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 298, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 298, + 235 + ], + "type": "image", + "image_path": "fbdd8e30957d4d16d454d025dd355f671b9266c41d592598e579aaba60c4b2d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 239, + 251, + 250 + ], + "lines": [ + { + "bbox": [ + 114, + 239, + 251, + 250 + ], + "spans": [ + { + "bbox": [ + 114, + 239, + 251, + 250 + ], + "type": "text", + "content": "(a) Classical SNN Knowledge Distillation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 76, + 545, + 225 + ], + "blocks": [ + { + "bbox": [ + 307, + 76, + 545, + 225 + ], + "lines": [ + { + "bbox": [ + 307, + 76, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 76, + 545, + 225 + ], + "type": "image", + "image_path": "c4fa6d85a843ee577e7c852e82135e0a4d402f096d610a8aabaf3c7ac875c27c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 230, + 550, + 250 + ], + "lines": [ + { + "bbox": [ + 304, + 230, + 550, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 550, + 250 + ], + "type": "text", + "content": "(b) Temporal Separation with Entropy Regularization (TSER) Knowledge Distillation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 258, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 316 + ], + "type": "text", + "content": "Figure 1. Illustration of the Classical SNN KD and Our Proposed Method. In our approach, we remove the temporal dimension fusion operation present in classical SNN KD and apply a temporal separation strategy to focus the learning process on outputs at individual time steps. The original calculations of KL Loss and CE Loss are adapted to compute the loss for each time step's output, followed by averaging. Additionally, we incorporate an entropy regularization term to guide the learning direction away from erroneous knowledge. Here, T, B, and C represent the time steps, batch size, and number of classes, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 335, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 296, + 647 + ], + "type": "text", + "content": "to facilitate training during inference. These approaches leverage Backpropagation Through Time (BPTT) for effective training [48, 57]. Motivated by the work [25], several studies have utilized surrogate gradient estimation as a method to address the non-differentiability inherent in SNNs. This foundational research has enabled the direct training of CNNs and transformer-based SNNs, achieving strong performance on static datasets with reduced inference time steps. To further reduce the performance gap with ANNs, recent work has explored the inclusion of modules such as attention mechanisms to enhance SNN capabilities [28, 40, 52, 53]. While effective, these additional modules increase computational consumption, counteracting the inherent low-energy benefits of SNNs. Knowledge Distillation (KD) provides a potential solution, where SNNs are guided by a more capable ANN teacher model, addressing the inference burden associated with these modules. Prior works such as KDSNN [50] and LaSNN [24] employ hierarchical knowledge distillation [5] to facilitate the training of SNNs by leveraging the rich semantic information from the ANN teacher model. Recent approaches like BKDSNN [51] move beyond traditional layer-wise distillation by integrating blurred knowledge distillation techniques with logit distillation, demonstrating significant improvements in performance across both CNNs and Transformer-based models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In preliminary experiments, we observed that current SNN distillation architectures are not optimally designed for the unique structure of SNNs. As depicted in Fig. 1a, conventional SNN KD techniques distill the mean of outputs across time steps. However, the inherent nature" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 335, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 538 + ], + "type": "text", + "content": "of SNN processing implies that significant information is embedded within the temporal dimension. Employing a mean approach for fusion prediction during classification resembles a voting selection process. Therefore, distributing the distillation of outputs across various time steps during the training phase can substantially alleviate the limitations that currently hinder distillation performance. To mitigate the impact of outliers on network performance and to rectify overly confident incorrect semantics produced by the teacher network, we introduce an entropy regularization term as a constraint. We conduct extensive experimental analyses to assess its efficacy. Furthermore, we evaluate our proposed method on several datasets, including CIFAR10 [26], CIFAR100 [26], and ImageNet [8]. The results of our experiments demonstrate that this method achieves state-of-the-art (SOTA) performance. In summary, our contributions are threefold:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 567, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 567, + 553, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 567, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 553, + 626 + ], + "type": "text", + "content": "- We introduce a temporal separation strategy into the SNN distillation method, effectively addressing the incompatibility issues present in previous SNN distillation frameworks and yielding a more effective universal distillation architecture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 628, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 628, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 314, + 628, + 553, + 673 + ], + "type": "text", + "content": "- We integrate entropy regularization to correct the overly confident errors from the teacher network, thereby maximizing the performance of the logit-based distillation method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 675, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 675, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 314, + 675, + 553, + 711 + ], + "type": "text", + "content": "- We conduct comprehensive evaluations across multiple datasets, showing that our proposed approach outperforms previous state-of-the-art (SOTA) methods." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8807" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 273 + ], + "type": "text", + "content": "Conversion-based Methods. To fully leverage the knowledge acquired from pre-trained ANNs, conversion methods directly transfer the learned parameters from an ANN to the corresponding SNN, avoiding the difficulties of training SNNs from scratch. Several studies [19, 20] replace the activation functions in ANNs with spiking neurons and introduce optimizations like weight normalization [12] to generate corresponding SNNs. The core idea of these methods is to closely align the outputs of the ANN and SNN, maximizing the utilization of the pre-trained ANN knowledge. While later works [5, 10, 12] achieve nearly lossless accuracy conversion, they often involve long inference times and fail to retain the spatiotemporal characteristics inherent to spike-based processing, significantly limiting the practical applicability of converted SNNs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 289, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 529 + ], + "type": "text", + "content": "Learning-based Methods. The development of surrogate gradient training methods [56] has resolved the challenge of training non-differentiable SNNs directly. Initially, [39] used a surrogate gradient approach to construct deep SNN models, achieving competitive accuracy on neuromorphic datasets. The DCT-SNN [16] applied frequency domain techniques to reduce the number of inference time steps. To support deeper SNN training, [15] introduced SEW ResNet to address gradient vanishing and exploding issues in Spiking ResNet [60], extending SNNs to architectures exceeding 100 layers. TA-SNN [52] introduced a temporal attention mechanism, paving the way for future attention mechanism advancements in the field. MA-SNN [53] went beyond single-dimensional information enhancement by proposing a temporal-channel-spatial attention mechanism. In Transformer architectures, models like SpikeFormer [61], SpikingFormer [62], and MetaSpikerFormer [63] demonstrated outstanding classification performance through the design and improvement of innovative spike-based self-attention blocks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "Knowledge Distillation for SNN. Knowledge Distillation (KD), a well-established transfer learning technique, has proven effective across various tasks [3, 4, 22]. Recent studies [17, 30, 32, 47, 59] have applied KD to SNN training, using either large isomorphic SNNs to extract knowledge for smaller SNNs or leveraging the logits from pretrained ANNs to train SNNs. KDSNN [50] utilizes a joint distillation approach based on logits and features, showing effectiveness across multiple datasets. Xu et al. [49] explore biologically inspired structure learning using reverse knowledge distillation. Furthermore, LaSNN [24] proposes a hierarchical feature distillation framework that achieves accuracy comparable to ANNs on the Tiny ImageNet dataset. BKDSNN [51] improves SNN performance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "on complex datasets by using blurred knowledge to replicate ANN features. Despite the advantages of these distillation methods, as shown by the experiments in Fig. 2, current SNN distillation frameworks have yet to fully account for the unique characteristics of SNNs. Developing tailored distillation paradigms could enhance model accuracy in a simple and effective way." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 167, + 397, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 397, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 397, + 182 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 188, + 555, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 237 + ], + "type": "text", + "content": "In this section, we start from preliminaries, introducing the fundamental concepts of SNNs and SNN distillation. Then we introduce our knowledge distillation approach that incorporates temporal separation with entropy regularization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 244, + 400, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 400, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 400, + 256 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": "Spiking Neuron Model. In SNNs, the fundamental computational unit is the spiking neuron, which serves as an abstract model of the dynamics of biological neurons. Currently, the Leaky Integrate-and-Fire (LIF) model is one of the most commonly used spiking neuron models, as it strikes a balance between simplified mathematical formulation and the complex dynamics of biological neurons. Therefore, in this study, we employ the LIF model as the foundational neuron model for the student SNN. Mathematically, a LIF neuron can be represented by Eq. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 362, + 391, + 555, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 391, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 362, + 391, + 555, + 416 + ], + "type": "interline_equation", + "content": "\\mu \\frac {d u (t)}{d t} = - (u (t) - u _ {\\text {r e s e t}}) + I (t). \\tag {1}", + "image_path": "55741794c3100b05902e422fc9663d901636469b6101e372c5937647fdea47c2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": " denotes the time constant of the membrane potential, " + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "inline_equation", + "content": "u(t)" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": " represents the membrane potential at time " + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{reset}}" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": " is the resting potential of the neuron, and " + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "inline_equation", + "content": "I(t)" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": " is the presynaptic input at time " + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": ". Based on this differential equation, the discrete-time and iterative mathematical representation of the LIF-SNN can be described as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 501, + 545, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 501, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 322, + 501, + 545, + 526 + ], + "type": "interline_equation", + "content": "V (t) = H (t - 1) + \\frac {1}{\\mu} [ I (t - 1) - (H (t - 1) - u _ {\\text {r e s e t}}) ]", + "image_path": "2a45841fb892d28fed7102f72b565d10c651ebbcc43506a1f4730d1e1d35ce18.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 528, + 417, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 528, + 417, + 541 + ], + "spans": [ + { + "bbox": [ + 324, + 528, + 417, + 541 + ], + "type": "interline_equation", + "content": "S (t) = \\Theta (V (t) - v _ {t h})", + "image_path": "22e11ef9feb7460e3223e24bd49ec9bd0afa1a21f748e91c506ea9c67b830cb5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 543, + 554, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 543, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 323, + 543, + 554, + 566 + ], + "type": "interline_equation", + "content": "H (t) = u _ {\\text {r e s e t}} \\cdot S (t) + V (t) \\cdot (1 - S (t)) \\tag {2}", + "image_path": "b0bc85778704290c4904368f5a9b625c8867c86133aabbda1fcd8c21a7c82b0a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 567, + 494, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 494, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 494, + 579 + ], + "type": "text", + "content": "The Heaviside step function " + }, + { + "bbox": [ + 313, + 567, + 494, + 579 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 567, + 494, + 579 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 385, + 588, + 555, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 588, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 385, + 588, + 555, + 622 + ], + "type": "interline_equation", + "content": "\\Theta (x) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} x < 0 \\\\ 1 & \\text {i f} x \\geq 0. \\end{array} \\right. \\tag {3}", + "image_path": "38ad6eae386a5841dbeccbaeaf23a9233cc795c6bd0fb9b4c482ab277536e78c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "Among these, " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "H(t - 1)" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " denotes the membrane potential following spike generation at the previous time step, while " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "I(t)" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "V(t)" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " represent the input and the updated membrane potential at time step " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": ", respectively. Furthermore, " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "v_{th}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " is the threshold that determines whether " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "V(t)" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " should fire a spike or remain silent, and " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "S(t)" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " indicates the spike sequence generated after triggering the action potential at time step " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8808" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "SNN Knowledge Distillation. We start from the original SNN Knowledge Distillation (KD) method. To illustrate the process of KD, we consider C-way classification task and denote " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " as the output of the network. Then, the class probability is given by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 136, + 296, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 136, + 296, + 166 + ], + "spans": [ + { + "bbox": [ + 123, + 136, + 296, + 166 + ], + "type": "interline_equation", + "content": "Q _ {i} = \\frac {\\exp \\left(Z _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(Z _ {j} / \\tau\\right)}, \\tag {4}", + "image_path": "cadd120653bc7f157c44b423452c14ec42b35362e534450b564932d7f5949abe.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " are the probability value on the " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": "-th class and " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " is the temperature scaling hyper-parameter. In knowledge distillation, " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " is typically larger than 1.0, which helps control the flatness of the feature distribution to mitigate the phenomenon of overconfidence within the network. When " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " equals 1.0, the output reverts to the vanilla Softmax output, denoted here as " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " to represent this specific case. The objective of KD is to transfer knowledge from a high-performing teacher model to a lightweight student model. For the rescaled outputs, the original KD method achieves distillation by minimizing the KL divergence between the outputs of the teacher and student models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 317, + 296, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 317, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 92, + 317, + 296, + 350 + ], + "type": "interline_equation", + "content": "K L \\left(Q ^ {t e a} | | Q ^ {s t u}\\right) = \\sum_ {i = 1} ^ {C} Q _ {i} ^ {t e a} \\log \\left(\\frac {Q _ {i} ^ {t e a}}{Q _ {i} ^ {s t u}}\\right), \\tag {5}", + "image_path": "c3425b6f476d3dd494ad144b11ece0c70c9df0086ef7562c7e072b75e66f33b9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "inline_equation", + "content": "Q_{i}^{tea}" + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "inline_equation", + "content": "Q_{i}^{stu}" + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "content": " indicates the probability value on the " + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "content": "-th category of the teacher and student output, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 379, + 296, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 414 + ], + "type": "text", + "content": "In addition, due to the potential errors in teachers' knowledge, KD also necessitates the use of cross-entropy to enable students to learn the true distribution of labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 417, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 417, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 89, + 417, + 295, + 449 + ], + "type": "interline_equation", + "content": "C E \\left(Q ^ {s t u}, y _ {t r u e}\\right) = - \\sum_ {i = 1} ^ {C} Q _ {i} ^ {s t u} \\log \\left(y _ {t r u e}\\right) \\tag {6}", + "image_path": "a7d40e8245100de28caf5cf7eaf144d872232f2981a4da34ff87c76bf4838141.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 453, + 195, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 195, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 195, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 453, + 195, + 464 + ], + "type": "inline_equation", + "content": "y_{\\text{true}}" + }, + { + "bbox": [ + 55, + 453, + 195, + 464 + ], + "type": "text", + "content": " denote the true labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 464, + 296, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 572 + ], + "type": "text", + "content": "The distinction between SNNs and ANNs lies in the fact that SNNs incorporate an additional temporal dimension " + }, + { + "bbox": [ + 55, + 464, + 296, + 572 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 464, + 296, + 572 + ], + "type": "text", + "content": ", which denotes the time step. Consequently, traditional SNN distillation attempts to adapt the distillation methods designed for ANNs by employing a simplistic approach that averages the outputs across the temporal dimension. By introducing a parameter " + }, + { + "bbox": [ + 55, + 464, + 296, + 572 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 464, + 296, + 572 + ], + "type": "text", + "content": " to control the weight ratio between true labels and soft labels. The loss calculation formula can be expressed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 82, + 577, + 296, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 577, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 82, + 577, + 296, + 644 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {S K D} = (1 - \\alpha) \\cdot C E \\left(\\frac {1}{T} \\sum_ {t = 1} ^ {T} Q ^ {s t u} (t), y _ {t r u e}\\right) \\tag {7} \\\\ + \\alpha \\cdot K L (Q ^ {t e a} | | \\frac {1}{T} \\sum_ {t = 1} ^ {T} Q ^ {s t u} (t)) \\\\ \\end{array}", + "image_path": "d4fc0df145927ee7bd26acbbc1b07a02586a8cf1146082bff5d45fa5af7ae192.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 296, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 673 + ], + "type": "text", + "content": "3.2. Temporal Separation Knowledge Distillation with Entropy Regularization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "In this section, we provide an in-depth explanation of the temporal separation knowledge distillation method with entropy regularization. Our approach consists of two core" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "components: temporal separation strategy and entropy regularization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 114, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 555, + 281 + ], + "type": "text", + "content": "Temporal Separation Strategy. As illustrated in Fig. 1a, the original SNN distillation merely transferred the conventional KD methods to SNNs in a naive manner. The use of mean values to integrate outputs across the temporal dimension fails to adequately account for the unique spatiotemporal characteristics of SNNs. Due to the incremental temporal processing in SNNs, the output features at each time step exhibit variability and contain rich information. However, this also implies the potential presence of local outliers in the outputs at different time steps. While the averaging operation mitigates the impact of these anomalies, it poses risks for the subsequent model learning and probabilistic confidence. Consequently, implementing a temporal separation strategy for learning time step's output is crucial." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 281, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 354 + ], + "type": "text", + "content": "As shown in Fig. 1b, we apply temporal separation strategy to CE and KL in Eq. 7, moving the mean operation outside. We directly use the outputs at each time step to distill the learning of the true labels and the teacher distribution. The loss expression for temporal separation knowledge distillation is given by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 344, + 363, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 363, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 344, + 363, + 553, + 430 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {T S} = (1 - \\alpha) \\cdot \\frac {1}{T} \\sum_ {t = 1} ^ {T} C E \\left(Q ^ {s t u} (t), y _ {t r u e}\\right) \\tag {8} \\\\ + \\alpha \\cdot \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\cdot K L (Q ^ {t e a} | | Q ^ {s t u} (t)). \\\\ \\end{array}", + "image_path": "a629129d1fd9255a4f22726aa2efb191ceee517e108d70b48d3a3cff37a64cf7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 458, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 579 + ], + "type": "text", + "content": "Entropy Regularization. After introducing the temporal separation strategy, the soft labels from the teacher network directly impact the output distribution at each time step. This approach, however, can inadvertently reinforce erroneous information from the teacher network, increasing the likelihood that the student model will adopt incorrect knowledge. To counter this, we introduce an entropy regularization term, which constrains and adjusts the teacher's outputs to reduce excessive confidence in potentially erroneous knowledge. The formula for entropy is as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 363, + 588, + 553, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 588, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 363, + 588, + 553, + 622 + ], + "type": "interline_equation", + "content": "H \\left(Q ^ {s t u}\\right) = - \\sum_ {i = 1} ^ {C} Q _ {i} ^ {s t u} \\log \\left(Q _ {i} ^ {s t u}\\right). \\tag {9}", + "image_path": "6bed849ad937bfecdf3ddb7c50010883bf7c151c3c76e9a44744c162a5884e5b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 627, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 553, + 651 + ], + "type": "text", + "content": "Then, we obtain the loss formula for the entropy regularization term:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 366, + 659, + 553, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 659, + 553, + 692 + ], + "spans": [ + { + "bbox": [ + 366, + 659, + 553, + 692 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E R} = - \\lambda \\cdot \\frac {1}{T} \\sum_ {t = 1} ^ {T} H \\left(Q ^ {s t u} (t)\\right), \\tag {10}", + "image_path": "21a3f1a3ab8033046f7a08deedf9bde21a13d54dc48698af6d1c674e610b2fff.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 701, + 434, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 434, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 434, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 701, + 434, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 314, + 701, + 434, + 714 + ], + "type": "text", + "content": " is the penalty factor." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8809" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "Finally, we integrate the temporal separation strategy, as shown in Eq. 8, with the entropy regularization term from Eq. 10 to derive a new loss function as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 118, + 295, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 118, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 125, + 118, + 295, + 131 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T S E R} = \\mathcal {L} _ {T S} + \\mathcal {L} _ {E R}. \\tag {11}", + "image_path": "03ec6d335a4f047eddeefd31ef874760c8269eec462a925a7373b3745bbbdf43.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 139, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 296, + 163 + ], + "type": "text", + "content": "And Algorithm. 1 outlines the overall training process of our proposed method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 174, + 132, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 132, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 132, + 186 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 194, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 266 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to validate the effectiveness of our method. Initially, we compare our approach with existing SNN distillation methods. Furthermore, we conduct ablation studies, sensitivity analyses, energy consumption assessments, and visualizations to provide a comprehensive evaluation of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 273, + 182, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 182, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 182, + 285 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 290, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 387 + ], + "type": "text", + "content": "Our proposed method is evaluated on three datasets. For teacher ANN networks, we select several advanced models, including ResNet-19 and ResNet-34 [21] as well as VGG-16. In contrast, for the student SNN models, we employ neural network architectures with equivalent or fewer layers, specifically ResNet-18/19, and VGG-11/16. Additionally, we conduct multiple experiments by varying the time step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 393, + 222, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 222, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 222, + 405 + ], + "type": "text", + "content": "4.2. Comparison on Static Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 410, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 434 + ], + "type": "text", + "content": "Our method is compared with previous approaches on CIFAR10 and CIFAR100, as shown in the Table. 1. The re" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 56, + 479, + 296, + 529 + ], + "blocks": [ + { + "bbox": [ + 55, + 453, + 295, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 477 + ], + "type": "text", + "content": "Algorithm 1 Training student SNN model with TSER knowledge distillation for one epoch." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "lines": [ + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "text", + "content": "Require: pre-trained teacher ANN model " + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "text", + "content": "; an initialized student SNN model " + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "text", + "content": "; input dataset sample " + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "text", + "content": "; and the true labels " + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "inline_equation", + "content": "y_{true}" + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "text", + "content": "; total training iteration in one epoch " + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "inline_equation", + "content": "I_{train}" + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "text", + "content": "; total validation iteration in one epoch " + }, + { + "bbox": [ + 56, + 479, + 296, + 529 + ], + "type": "inline_equation", + "content": "I_{val}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + }, + { + "bbox": [ + 57, + 529, + 209, + 538 + ], + "lines": [ + { + "bbox": [ + 57, + 529, + 209, + 538 + ], + "spans": [ + { + "bbox": [ + 57, + 529, + 209, + 538 + ], + "type": "text", + "content": "Ensure: SNN model with TSER KD" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + } + ], + "index": 10, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 66, + 540, + 296, + 709 + ], + "blocks": [ + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "lines": [ + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "spans": [ + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "I_{train}" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " do Get mini-batch training data, and true label: " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "X^i,y_{true}^i" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " Compute the " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " output " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "Z^{stu,i}(t)" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " of each time step; Compute the " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " output " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "Z^{tea,i}" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " Calculate loss function: " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TSER}" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " Backpropagation and update model parameters; \nend for \nfor " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "I_{val}" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " do Get mini-batch validation data, and true label: " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "X^i" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "y_{true}^{i}" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " Compute the SNN average output " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "Z_{mean}^{stu,i} =" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "\\frac{1}{T}\\sum_{t = 1}^{T}Z^{stu,i}(t);" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " Compare the " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "Z_{mean}^{stu,i}" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "inline_equation", + "content": "y_{true}^{i}" + }, + { + "bbox": [ + 66, + 540, + 296, + 709 + ], + "type": "text", + "content": " for classification; \nend for" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 547, + 258 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 547, + 258 + ], + "type": "image", + "image_path": "93817fae9dd3b32330078722e85b57ea777ccee5f5d851884c83b955275a7e96.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 267, + 555, + 334 + ], + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 334 + ], + "type": "text", + "content": "Figure 2. Prediction Accuracy Distribution at Different Time Steps for Vanilla SNN KD and Our Proposed Method. The solid red line indicates the teacher model's accuracy, while the dashed lines represent the prediction accuracies of different distillation methods after averaging outputs over time steps. The bars show the prediction accuracies of each distillation method at individual time steps." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 321, + 354, + 545, + 509 + ], + "blocks": [ + { + "bbox": [ + 321, + 354, + 545, + 509 + ], + "lines": [ + { + "bbox": [ + 321, + 354, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 321, + 354, + 545, + 509 + ], + "type": "image", + "image_path": "1393a2f92084cee688a2bec85335b14cadb8dc0500647d2c5e29e6a0d3650b66.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 525, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 559 + ], + "type": "text", + "content": "Figure 3. Accuracy Distribution for Different " + }, + { + "bbox": [ + 313, + 525, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 525, + 555, + 559 + ], + "type": "text", + "content": " Values. Experiments conducted on the CIFAR100 dataset with a fixed time step of 2, testing various " + }, + { + "bbox": [ + 313, + 525, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 525, + 555, + 559 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "sults demonstrate our method surpasses the SOTA performance achieved by prior method across both ResNet and VGG architectures. Specifically, ResNet19 outperform the previous SOTA by " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.57\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.41\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " on CIFAR10 and CIFAR100, respectively, achieving accuracies of " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "96.72\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "81.29\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". This indicates that our approach, although based on logit distillation, performs exceptionally well compared to previous, more complex distillation methods. By addressing the limitations of simple transfer in existing distillation, our framework effectively learns from the teacher's knowledge, minimizing the gap between the student and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 783, + 315, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 783, + 315, + 792 + ], + "spans": [ + { + "bbox": [ + 295, + 783, + 315, + 792 + ], + "type": "text", + "content": "8810" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 69, + 560, + 430 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 560, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 560, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 560, + 430 + ], + "type": "table", + "html": "
MethodAcc.(%)Arch.Acc.(%)Arch.Acc.(%)Arch.Acc.(%)
CF-100CF-10CF-100CF-10CF-100CF-10CF-100CF-10
without KDTimestep=2Timestep=4
R-1975.4894.44R-1874.6094.15R-1977.0594.85R-1876.4294.67
TET[11]R-1972.8794.16---R-1974.4794.44---
LSG[35]R-1976.3294.41---R-1976.8595.17---
PFA[7]R-1976.7095.60---R-1978.1095.71---
MPBN[18]R-1979.5196.47V-1673.8893.96R-1980.1096.52V-1674.7494.44
IM-LIF[36]R-1977.2195.29---R-1977.4295.66---
Spikformer[64]------S-4-25675.9693.94S-4-38477.8695.19
Spikingformer[61]------Sg-4-25677.4394.77Sg-4-38479.0995.61
SD Transformer [54]------S-2-51278.4095.60---
with KDTeacher ANN Timestep=1
R-3481.1297.10V-1678.0896.06R-3481.1297.10V-1678.0896.06
R-1981.9797.08V-1678.0896.06---V-1678.0896.06
LaSNN [24]Timestep=2Timestep=4
R-1876.1793.64V-1673.8093.90R-1878.1295.09V-1674.9994.49
R-1980.3095.26V-1169.8990.23---V-1170.7490.42
KDSNN [50]R-1877.1695.25V-1674.6594.26R-1878.4695.72V-1675.9894.85
R-1980.8896.15V-1172.1792.81---V-1173.0592.91
BKDSNN [51]R-1873.3093.64V-1673.8094.10R-1875.5795.09V-1674.9994.55
R-1975.7495.26V-1169.8992.73---V-1170.7492.88
OursR-1878.30±0.0895.58±0.08V-1675.81±0.1294.55±0.09R-1879.69±0.0296.18±0.05V-1677.06±0.0495.01±0.10
R-1981.29±0.1496.72±0.02V-1173.67±0.0493.31±0.02---V-1174.65±0.0793.68±0.02
", + "image_path": "e13fd0d31eded058099fb55343a4c8200ee635b7d47d16992fcd4b5ece1a8eab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 436, + 555, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 436, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 555, + 481 + ], + "type": "text", + "content": "Table 1. Comparison results with training-based SNN SOTA methods, including CNN- and transformer-based approaches, on CIFAR-10/100, with and without Knowledge Distillation (KD). Acc. denotes accuracy, CF denotes CIFAR, and Archi. denotes architecture. The abbreviations R, V, S, and Sg represent ResNet, VGG, Spikeformer, and Spikingformer architectures, respectively, while SD Transformer refers to the Spike-Driven Transformer. For KD methods, the student model aligns with the architecture of the teacher ANN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 502, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 526 + ], + "type": "text", + "content": "teacher models. This reduction in disparity validates the superiority of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "content": "Further comparisons on ImageNet dataset are detailed in Table. 2. The SEW ResNet-18 and SEW ResNet-34 architectures achieve top-1 accuracies of " + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "inline_equation", + "content": "69.24\\%" + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "inline_equation", + "content": "73.16\\%" + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "content": ", respectively. These results indicate that our method is capable of delivering competitive performance even within large-scale datasets, effectively harnessing the knowledge from ANNs to enhance the learning capabilities of SNNs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 623, + 149, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 149, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 149, + 635 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Temporal Separation Strategies and Entropy Regularization. We investigate different configurations of " + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TSER}" + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": ", focusing on the temporal separation applied within the KL and CE losses, as well as the integration of an entropy regularization term. The results are summarized in Table. 3. Our experiments reveal that incorporating tempo" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "text", + "content": "ral separation strategies notably enhances model accuracy, with improvements of " + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "inline_equation", + "content": "0.49\\%" + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "inline_equation", + "content": "0.26\\%" + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "text", + "content": " for the standard KL and CE loss function, respectively. When temporal separation is applied concurrently to both KL and CE losses, the overall accuracy increase reaches " + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "inline_equation", + "content": "0.84\\%" + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "text", + "content": ". Furthermore, we find that optimally adjust the parameter " + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "text", + "content": " in the entropy regularization term, alongside temporal separation, leads to an enhancement of up to " + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "inline_equation", + "content": "0.98\\%" + }, + { + "bbox": [ + 313, + 502, + 555, + 633 + ], + "type": "text", + "content": ". These findings underscore the effectiveness of combining temporal separation and entropy regularization in maximizing the performance of SNN distillation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Comparison of Output Accuracy at Different Time Steps. To illustrate the adaptability of our proposed distillation method for SNNs, we perform a comparative analysis of accuracy across different time steps between vanilla SNN KD and our method. As shown in Fig. 2, the accu" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8811" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 298, + 368 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 368 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 368 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 368 + ], + "type": "table", + "html": "
MethodArchitectureTAccuracy
without KDTET [11]R-34664.79%
GLIF [55]R-34467.52%
MPBN [18]R-18463.14%
R-34464.71%
SEW ResNet [15]R-18463.18%
R-34467.04%
Spikformer [64]S-8-384470.24%
S-6-512472.46%
Spikingformer [61]Sg-8-384472.45%
Spike-driven Transformer [54]Sg-8-384472.28%
with KDTeacher ANNSEW R-18171.69%
SEW R-34175.38%
LaSNN [24]SEW R-18463.33%
SEW R-34466.98%
KDSNN [50]SEW R-18463.61%
SEW R-34467.28%
BKDSNN [51]SEW R-18463.43%
SEW R-34467.21%
OursSEW R-18469.24% ±0.19
SEW R-34473.16% ±0.15
", + "image_path": "9a7e7d0d12831807cc3a2159a48952571aeeea648cbb54d41de631270149d793.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 441, + 293, + 530 + ], + "blocks": [ + { + "bbox": [ + 55, + 374, + 298, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 374, + 298, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 298, + 433 + ], + "type": "text", + "content": "Table 2. Comparison results with training-based SNN SOTA methods, including CNN- and Transformer-based approaches, on ImageNet, with and without Knowledge Distillation (KD). T denotes Timestep. R, S, Sg denotes ResNet, Spikeformer and Spikingformer architectures respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 441, + 293, + 530 + ], + "lines": [ + { + "bbox": [ + 59, + 441, + 293, + 530 + ], + "spans": [ + { + "bbox": [ + 59, + 441, + 293, + 530 + ], + "type": "table", + "html": "
Temporal SeparationEntropy RegularizationAccuracy
KL LossCE Loss
XXX77.15%
X✓X77.41%
✓XX77.64%
✓✓X77.99%
✓✓✓78.13%
", + "image_path": "435eb65b65c3cd40839bbd5e8dfbfafe7c411298c776c94bad82792f79a2a49c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 536, + 295, + 562 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 562 + ], + "type": "text", + "content": "Table 3. Ablation of Temporal Separation Strategy and Entropy Regularization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": "racy discrepancy among time steps for the vanilla SNN KD is " + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "inline_equation", + "content": "0.98\\%" + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": ", notably higher than the " + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "inline_equation", + "content": "0.65\\%" + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": " difference with our approach. Furthermore, the vanilla SNN KD yields an average output accuracy of " + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "inline_equation", + "content": "74.67\\%" + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": ", whereas our method achieves " + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "inline_equation", + "content": "76.56\\%" + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": ". And as the accuracy of individual time step outputs improved, we observe a significant enhancement in performance following the mean operation on these outputs, aligning with our hypothesis. This clearly demonstrates the effectiveness of our method and highlights the significance of temporal separation in the SNN distillation process." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 72, + 542, + 222 + ], + "blocks": [ + { + "bbox": [ + 326, + 72, + 542, + 222 + ], + "lines": [ + { + "bbox": [ + 326, + 72, + 542, + 222 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 542, + 222 + ], + "type": "image", + "image_path": "3964209d0776f37ed9eb84592b1349dc506c2a179fecfedc01f731e956bbae19.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 231, + 556, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 231, + 556, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 556, + 277 + ], + "type": "text", + "content": "Figure 4. Temperature Coefficient Sensitivity. Our method demonstrates stability across various " + }, + { + "bbox": [ + 313, + 231, + 556, + 277 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 231, + 556, + 277 + ], + "type": "text", + "content": " hyperparameters. This experiment is conducted on CIFAR100 with ResNet-34 as the teacher model and ResNet-18 as the student model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "text", + "content": "Selection of Parameter " + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "text", + "content": ". The entropy regularization term is essential for " + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TSER}" + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "text", + "content": ", as it partially alleviates the erroneous knowledge transferred by the teacher network. Thus, selecting an optimal value for " + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "text", + "content": " is crucial. We conduct experiments with various values of " + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "text", + "content": ", and the results are shown in Fig. 3. Increasing " + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "text", + "content": " enhances the correction of the teacher's inaccuracies but also dilutes the distribution of the correct knowledge. Our findings indicate that maintaining " + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "text", + "content": " within the range of 0.0001 to 0.2 enhances the network's overall training stability. However, excessively high values of " + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 297, + 556, + 441 + ], + "type": "text", + "content": " can result in negative loss values, as described in Eq. 11, which must be avoided." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 459, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 555 + ], + "type": "text", + "content": "Temperature Coefficient Sensitivity. In our experiments, we configure the temperature to " + }, + { + "bbox": [ + 313, + 459, + 556, + 555 + ], + "type": "inline_equation", + "content": "\\tau = 5" + }, + { + "bbox": [ + 313, + 459, + 556, + 555 + ], + "type": "text", + "content": ". Utilizing the prior experimental setup, we select five temperature points: " + }, + { + "bbox": [ + 313, + 459, + 556, + 555 + ], + "type": "inline_equation", + "content": "\\tau = [2,3,4,5,6]" + }, + { + "bbox": [ + 313, + 459, + 556, + 555 + ], + "type": "text", + "content": ". We evaluate performance with ResNet-34 as the teacher model and ResNet-18 as the student model on the CIFAR100 dataset. The results depicted in Fig. 4, indicate that our method exhibits stable performance across different temperature coefficients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 525, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 525, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 525, + 577 + ], + "type": "text", + "content": "4.4. Performance Analysis and Visualization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Firing Rate Analysis. In Table. 4, the firing rates of our proposed method show a slight increase compared to vanilla SNN KD across both ResNet and VGG architectures. Fig. 2 illustrates that an increase in firing rate of approximately " + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": " corresponds to a " + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "inline_equation", + "content": "1.48\\%" + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": " improvement in overall performance. Notably, when comparing our method to other distillation techniques within the ResNet architecture, it maintains performance gains while simultaneously reducing firing rate and increasing spike sparsity, thus minimizing redundant spikes. Although LaSNN exhibits lower firing rates, this comes at the expense of some model per" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 758 + ], + "type": "text", + "content": "8812" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 136, + 152 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 136, + 152 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 136, + 152 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 136, + 152 + ], + "type": "image", + "image_path": "3d4af674e592612b5ad7c6da632b259beee57c74f8fb33a5f4a9629e284e147f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 164, + 487, + 176 + ], + "lines": [ + { + "bbox": [ + 122, + 164, + 487, + 176 + ], + "spans": [ + { + "bbox": [ + 122, + 164, + 487, + 176 + ], + "type": "text", + "content": "Figure 5. t-SNE Visualization of features learned by teacher ANN and different distillation methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 143, + 72, + 220, + 151 + ], + "blocks": [ + { + "bbox": [ + 143, + 72, + 220, + 151 + ], + "lines": [ + { + "bbox": [ + 143, + 72, + 220, + 151 + ], + "spans": [ + { + "bbox": [ + 143, + 72, + 220, + 151 + ], + "type": "image", + "image_path": "054805f500d4ff747523588d7f021f2097d9cd41c8e181604307ab2f882f7a48.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 304, + 151 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 304, + 151 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 304, + 151 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 304, + 151 + ], + "type": "image", + "image_path": "f1db2d30c835228bf392eaea766568a0644916da3776b17666313c415c128354.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 382, + 151 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 382, + 151 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 382, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 382, + 151 + ], + "type": "image", + "image_path": "06b70e903b440527740feb3966b35720be5ebfd20991e8ce9ed05db5cdf04edc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 465, + 151 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 465, + 151 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 465, + 151 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 465, + 151 + ], + "type": "image", + "image_path": "3a5f7fc37f48a03e2d52282fb9b496fd760278a4574a9876e8b46e6a3320c49c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 470, + 72, + 547, + 151 + ], + "blocks": [ + { + "bbox": [ + 470, + 72, + 547, + 151 + ], + "lines": [ + { + "bbox": [ + 470, + 72, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 470, + 72, + 547, + 151 + ], + "type": "image", + "image_path": "2e2eb4b5e824a4531ffadb701a9ad7a6d8eaf902270627046d3cd5d96d9caf2f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 103, + 186, + 509, + 315 + ], + "blocks": [ + { + "bbox": [ + 103, + 186, + 509, + 315 + ], + "lines": [ + { + "bbox": [ + 103, + 186, + 509, + 315 + ], + "spans": [ + { + "bbox": [ + 103, + 186, + 509, + 315 + ], + "type": "table", + "html": "
ArchitectureMethodsOPsACsMACsEnergyTraining TimeFiring Rate
ResNet18Vanilla SNN KD2.23G405.33M12.2M0.42mJ2.78s22.85%
KDSNN2.23G411.49M12.2M0.43mJ2.67s23.23%
LaSNN2.23G300.94M12.2M0.33mJ2.79s16.76%
BKDSNN2.23G511.57M12.2M0.52mJ2.84s34.93%
Ours2.23G405.98M12.2M0.42mJ2.70s22.95%
VGG16Vanilla SNN KD1.26G159.96M274.2M1.40mJ2.43s18.91%
KDSNN1.26G170.79M274.2M1.39mJ2.39s18.59%
LaSNN1.26G150.96M274.2M1.41mJ2.44s17.25%
BKDSNN1.26G171.94M274.2M1.40mJ2.58s18.66%
Ours1.26G160.92M274.2M1.41mJ2.40s20.05%
", + "image_path": "6cb3d31a64104bd5a55b18e84ab1fea6968f53d62d74d8af00d90a5a62438f0b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 322, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 322, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 555, + 346 + ], + "type": "text", + "content": "Table 4. Energy consumption, training time, and spike firing rate under various distillation methods and architectures on the CIFAR100 dataset. Training Time refers to the time required per batch during training." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 365, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 413 + ], + "type": "text", + "content": "formance. Furthermore, unlike our logit-based approach, LaSNN employs a feature-based method that optimizes the feature map directly, making it more advantageous for feature sparsification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 432, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 540 + ], + "type": "text", + "content": "Energy Consumption Analysis. Our energy consumption analysis, based on the energy model in [29], is summarized in Table. 4. The results reveal that the most significant variations in energy consumption occur in Accumulation Computation (AC), which is related to the firing rate. Notably, a lower firing rate correlates with reduced AC. Consequently, our approach does not increase the computational load of SNNs and is competitive regarding energy efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Training Time. We evaluate the training times of various competitive KD methods on the CIFAR100 dataset, focusing on the training time per batch. As shown in Table. 4, our method achieves the second-fastest training time, surpassed only by KDSNN. This advantage is largely due to KDSNN simplifying the original KL divergence into cross-entropy in the logit-based framework, while lowering computational cost, also affect accuracy. Our approach outperforms other techniques by utilizing only the logits output for knowledge distillation, thus eliminating the need for additional auxiliary training modules or feature map comparisons. In contrast, previous methods demand more time and resources to extract knowledge from intermediate layers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": "Visualization. We employ t-SNE to visualize features learned from various distillation methods on the CIFAR100 dataset, utilizing ResNet-34 as the teacher and ResNet-18 as the student model. As shown in Fig. 5, our approach significantly enhances the distinguishability of deeper features compared to existing SNN distillation techniques." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 388, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 388, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 388, + 462 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 471, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 615 + ], + "type": "text", + "content": "This work addresses the issue of existing SNN distillation methods that are simply derived from ANN. We propose a new framework specifically designed for SNN distillation, which better exploits the temporal information contained in the logit output of SNNs for the process of knowledge distillation. Specifically, we incorporate a temporal separation strategy and introduce an entropy regularization term into the original distillation method. We aim to uncover the potential of SNN distillation by fully leveraging temporal information and rectifying the erroneous teacher knowledge. Extensive experimental results demonstrate the efficacy of this approach." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 627, + 429, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 429, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 429, + 640 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 647, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 555, + 704 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China (NSFC) (62476035, 62206037, 62276230), and Natural Science Foundation of Zhejiang Province (LDT23F02023F02), and State Key Laboratory (SKL) of Biobased Transportation Fuel Technology." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8813" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 134 + ], + "type": "text", + "content": "[1] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 294, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 294, + 178 + ], + "type": "text", + "content": "[2] Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep convolutional neural networks for energy-efficient object recognition. International Journal of Computer Vision, 113:54-66, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "text", + "content": "[3] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9233-9242, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 278 + ], + "type": "text", + "content": "[4] Li Chen, Chunyan Yu, and Lvcai Chen. A new knowledge distillation for incremental object detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "text", + "content": "[5] Jang Hyun Cho and Bharath Hariharan. On the efficacy of knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4794-4802, 2019. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 294, + 378 + ], + "type": "text", + "content": "[6] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 423 + ], + "type": "text", + "content": "[7] Haoyu Deng, Ruijie Zhu, Xuerui Qiu, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tensor decomposition based attention module for spiking neural networks. Knowledge-Based Systems, 295:111780, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 468 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 470, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 294, + 512 + ], + "type": "text", + "content": "[9] Lei Deng, Yujie Wu, Xing Hu, Ling Liang, Yufei Ding, Guoqi Li, Guangshe Zhao, Peng Li, and Yuan Xie. Rethinking the performance comparison between snns and anns. Neural networks, 121:294-307, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 545 + ], + "type": "text", + "content": "[10] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476, 2021. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 590 + ], + "type": "text", + "content": "[11] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. In International Conference on Learning Representations. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 294, + 646 + ], + "type": "text", + "content": "[12] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2015. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "text", + "content": "[13] Jianhao Ding, Tong Bu, Zhaofei Yu, Tiejun Huang, and Jian Liu. Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training. Advances in Neural Information Processing Systems, 35:24780-24793, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[14] Jason K Eshraghian, Max Ward, Emre O Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Ben" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "namoun, Doo Seok Jeong, and Wei D Lu. Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 151 + ], + "type": "text", + "content": "[15] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 555, + 207 + ], + "type": "text", + "content": "[16] Isha Garg, Sayeed Shafayet Chowdhury, and Kaushik Roy. Dct-snn: Using dct to distribute spatial information over time for low-latency spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4671-4680, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[17] Yufei Guo, Weihang Peng, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Xuhui Huang, and Zhe Ma. Joint a-sn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization. Pattern Recognition, 142:109639, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 264, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 328 + ], + "type": "text", + "content": "[18] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19420-19430, 2023. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 331, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 331, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 553, + 373 + ], + "type": "text", + "content": "[19] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European conference on computer vision, pages 388-404. Springer, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 376, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 441 + ], + "type": "text", + "content": "[20] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558-13567, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 443, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 443, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 443, + 553, + 487 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 488, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 510 + ], + "type": "text", + "content": "[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 511, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 577 + ], + "type": "text", + "content": "[23] Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdelrahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal processing magazine, 29(6):82-97, 2012. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 578, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 622 + ], + "type": "text", + "content": "[24] Di Hong, Jiangrong Shen, Yu Qi, and Yueming Wang. Lasnn: Layer-wise ann-to-snn distillation for effective and efficient training in deep spiking neural networks. arXiv preprint arXiv:2304.09101, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 624, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 666 + ], + "type": "text", + "content": "[25] Saeed Reza Kheradpisheh, Mohammad Ganjtabesh, Simon J Thorpe, and Timothée Masquelier. Stdp-based spiking deep convolutional neural networks for object recognition. Neural Networks, 99:56-67, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 668, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 668, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 668, + 553, + 689 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8814" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "text", + "content": "works. Advances in neural information processing systems, 25, 2012. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[28] Souvik Kundu, Gourav Datta, Massoud Pedram, and Peter A Beerel. Spike-thrift: Towards energy-efficient deep spiking neural networks by limiting spiking activity via attention-guided compression. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3953-3962, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "text", + "content": "[29] Souvik Kundu, Massoud Pedram, and Peter A Beerel. Hiresnn: Harnessing the inherent robustness of energy-efficient deep spiking neural networks by training with crafted input noise. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5209-5218, 2021. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[30] Ravi Kumar Kushawaha, Saurabh Kumar, Biplab Banerjee, and Rajbabu Velmurugan. Distilling spikes: Knowledge distillation in spiking neural networks. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4536-4543. IEEE, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 297 + ], + "type": "text", + "content": "[31] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[32] Dongjin Lee, Seongsik Park, Jongwan Kim, Wuhyeong Doh, and Sungroh Yoon. Energy-efficient knowledge distillation for spiking neural networks. arXiv preprint arXiv:2106.07172, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "type": "text", + "content": "[33] Chen Li, Lei Ma, and Steve Furber. Quantization framework for fast spiking neural networks. Frontiers in Neuroscience, 16:918793, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 431 + ], + "type": "text", + "content": "[34] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[35] Shuang Lian, Jiangrong Shen, Qianhui Liu, Ziming Wang, Rui Yan, and Huajin Tang. Learnable surrogate gradient for direct training spiking neural networks. In IJCAI, pages 3002-3010, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[36] Shuang Lian, Jiangrong Shen, Ziming Wang, and Huajin Tang. Im-lif: Improved neuronal dynamics with attention mechanism for direct training deep spiking neural network. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "text", + "content": "[37] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 633 + ], + "type": "text", + "content": "[38] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[39] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36 (6):51-63, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[40] Kleanthis C Neokleous, Marios N Avraamides, Costas K Neocleous, and Christos N Schizas. Selective attention and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "consciousness: investigating their relation through computational modelling. Cognitive Computation, 3:321-331, 2011. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[41] Srdjan Ostojic. Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons. Nature neuroscience, 17(4):594–600, 2014. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[42] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "text", + "content": "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[44] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[45] Catherine D Schuman, Shruti R Kulkarni, Maryam Parsa, J Parker Mitchell, Bill Kay, et al. Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science, 2(1):10-19, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "type": "text", + "content": "[46] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 389, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 444 + ], + "type": "text", + "content": "[47] Sugahara Takuya, Renyuan Zhang, and Yasuhiko Nakashima. Training low-latency spiking neural network through knowledge distillation. In 2021 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPPS), pages 1-3. IEEE, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 445, + 553, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 489 + ], + "type": "text", + "content": "[48] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 491, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 533 + ], + "type": "text", + "content": "[49] Qi Xu, Yaxin Li, Xuanye Fang, Jiangrong Shen, Jian K. Liu, Huajin Tang, and Gang Pan. Biologically inspired structure learning with reverse knowledge distillation for spiking neural networks, 2023. arXiv:2304.09500 [cs]. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 535, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 600 + ], + "type": "text", + "content": "[50] Qi Xu, Yaxin Li, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7886-7895, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 602, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 646 + ], + "type": "text", + "content": "[51] Zekai Xu, Kang You, Qinghai Guo, Xiang Wang, and Zhezhi He. Bkdsnn: Enhancing the performance of learning-based spiking neural networks training with blurred knowledge distillation. arXiv preprint arXiv:2407.09083, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[52] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8815" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 664 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[53] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. arXiv preprint arXiv:2209.13929, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[54] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[55] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[56] Qiang Yu, Jialu Gao, Jianguo Wei, Jing Li, Kay Chen Tan, and Tiejun Huang. Improving multispike learning with plastic synaptic delays. IEEE Transactions on Neural Networks and Learning Systems, 34(12):10254-10265, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[57] Friedemann Zenke and Tim P Vogels. The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation, 33 (4):899-925, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[58] Friedemann Zenke, Everton J Agnes, and Wulfram Gerstner. Diverse synaptic plasticity mechanisms orchestrated to form and retrieve memories in spiking neural networks. Nature communications, 6(1):6922, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "text", + "content": "[59] Fengzhao Zhang, Chengting Yu, Hanzhi Ma, Zheming Gu, and Er-ping Li. Knowledge Distillation For Spiking Neural Network. In 2023 5th International Conference on Robotics, Intelligent Control and Artificial Intelligence (RICAL), pages 1015-1020, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "text", + "content": "[60] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 496 + ], + "type": "text", + "content": "[61] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 563 + ], + "type": "text", + "content": "[62] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Zhengyu Ma, Huihui Zhou, Xiaopeng Fan, and Yonghong Tian. Enhancing the performance of transformer-based spiking neural networks by snn-optimized downsampling with precise gradient backpropagation. arXiv preprint arXiv:2305.05954, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 608 + ], + "type": "text", + "content": "[63] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 294, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 294, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 294, + 664 + ], + "type": "text", + "content": "[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations. 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8816" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_content_list.json b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2804847fc1ba2b75ff79e99d3e4b4b5a746c8d80 --- /dev/null +++ b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_content_list.json @@ -0,0 +1,1447 @@ +[ + { + "type": "text", + "text": "Temporally Consistent Object-Centric Learning by Contrasting Slots", + "text_level": 1, + "bbox": [ + 150, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anna Manasyan1 Maximilian Seitzer1 Filip Radovic1 Georg Martius1,3 Andrii Zadaianchuk2 \n1University of Tübingen, Tübingen, Germany \n2University of Amsterdam, Amsterdam, Netherlands \n3Max Planck Institute for Intelligent Systems, Tübingen, Germany", + "bbox": [ + 99, + 172, + 903, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "anna.manasyan@student.uni-tuebingen.de a.zadaianchuk@uva.nl", + "bbox": [ + 228, + 229, + 767, + 243 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e4ee66c9e4ae74713e74fd1a72338564bdc5e1654deaedb6eb6f41340bb7b3d.jpg", + "image_caption": [ + "Figure 1. SLOT CONTRAST: Our method introduces a temporal contrastive loss that enhances temporal consistency in unsupervised video object-centric models. It stabilizes representations across frames, achieving state-of-the-art object discovery on complex real-world videos." + ], + "image_footnote": [], + "bbox": [ + 91, + 263, + 903, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 539, + 326, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised object-centric learning from videos is a promising approach to extract structured representations from large, unlabeled collections of videos. To support downstream tasks like autonomous control, these representations must be both compositional and temporally consistent. Existing approaches based on recurrent processing often lack long-term stability across frames because their training objective does not enforce temporal consistency. In this work, we introduce a novel object-level temporal contrastive loss for video object-centric models that explicitly promotes temporal consistency. Our method significantly improves the temporal consistency of the learned object-centric representations, yielding more reliable video decompositions that facilitate challenging downstream tasks such as unsupervised object dynamics prediction. Furthermore, the inductive bias added by our loss strongly improves object discovery, leading to state-of-the-art results on both synthetic and real-world datasets, outperforming even weakly-supervised methods that leverage motion masks as additional cues.", + "bbox": [ + 86, + 580, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visit slotcontrast.github.io for videos and further details.", + "bbox": [ + 89, + 868, + 465, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 539, + 643, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object-centric learning (OCL) [4, 14, 32, 44] is a rapidly advancing area of visual representation learning that enables autonomous systems to represent, understand, and model high-dimensional data directly in terms of its constituent entities. Structured object-centric representations (often referred to as slots [32]) facilitate generalization and robustness [7, 9] of scene representations across diverse downstream tasks, from visual question answering [1, 8, 34, 58] to control [10, 18, 61, 63]. Of particular interest are video-based object-centric methods [2, 11, 20, 28, 45, 57, 64] that learn to represent objects that evolve and interact over time. These representations make the methods powerful tools for applications such as unsupervised online object tracking [35, 53] and structured world modeling [27, 52, 56]. Unsupervised object-centric learning on videos has seen significant progress in recent years [2, 57, 64], mainly due to the use of pre-trained representations from self-supervised foundational models [5, 38] coupled with diverse training datasets like YouTube-VIS [59, 60]. Nevertheless, these methods still face significant challenges, especially maintaining consistent object-centric representations across time and uniquely representing each object—critical factors for", + "bbox": [ + 511, + 568, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5401", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "successful multi-object tracking and modeling of dynamic scenes [28, 52, 56].", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal consistency [15, 31, 62] in object-centric representations refers to maintaining the same representation placeholder, called slot, for an object throughout a video sequence, effectively serving as a stable object-specific identifier over time. Existing unsupervised object-centric methods [27, 36, 48] aiming to discover consistent representations have primarily been studied on toy datasets with limited complexity [22, 26, 46]. In contrast, real-world video sequences present numerous challenges, including object occlusions, reappearance, and complex multi-object interactions, which complicate maintaining consistent object representations.", + "bbox": [ + 88, + 121, + 482, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel method to address the challenge of maintaining consistent temporal representations in object-centric models, extending the line of research on slot-based unsupervised video models [11, 64]. Our approach (named SLOT CONTRAST) scales to real-world video data and produces consistent object-centric representations. Notably, it achieves these results without requiring any human annotations. In particular, we propose a novel self-supervised contrastive learning objective, which contrasts slot representations throughout the batch while ensuring temporal coherence across consecutive frames. In addition, we modify the slot's initialization strategy [32] to promote distinct, contrastive representations. This combination leads to improved temporal consistency of learned representations, which we show to be highly effective for challenging downstream tasks such as unsupervised object tracking and latent object dynamics learning.", + "bbox": [ + 91, + 287, + 485, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions are as follows:", + "bbox": [ + 109, + 544, + 383, + 558 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the novel slot-slot contrastive loss that sets the state-of-the-art in temporal consistency when integrated into slot-based video processing methods.", + "- We develop SLOT CONTRAST, a simple and effective OCL architecture using the slot-slot contrastive loss paired with learned initialization that scales to real-world data, such as YouTube videos.", + "- We extensively study the usefulness of our learned object-centric representations for challenging downstream tasks, including unsupervised online tracking with complete occlusions and latent object dynamics modeling.", + "- We show that SLOT CONTRAST does not only improve the temporal consistency of the representations, but also achieves state-of-the-art on the object discovery task, outperforming weakly-supervised models using motion cues." + ], + "bbox": [ + 89, + 561, + 482, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 800, + 232, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised video object-centric learning There exists an extensive body of research [2, 11, 14, 20, 23, 28, 30, 42, 45, 49, 51, 52, 64] on discovering objects from video without any human annotations, primarily through tracking either object bounding boxes or masks. To achieve this, most of these", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "works combine an auto-encoder framework with a simple reconstruction objective, adding inductive biases for object discovery through structured encoders [4, 32] and decoders [55]. In particular, many modern object-centric image models [8, 21, 24, 44, 57] use a latent slot attention module [32] to extract object representations and corresponding object masks. For video data, most current methods [11, 28, 45, 49, 64, 66] connect slots across frames, with slots from the previous frame initializing those in the current frame. Notably, recent approaches [2, 40, 64] have successfully scaled object discovery to real-world unconstrained videos. To achieve this, SOLV [2] introduces temporal consistency via agglomerative clustering and prediction of middle-frame features, whereas VideoSAUR [64] learns object-centric representations by predicting temporal similarity of self-supervised features [5, 38]. While such methods can decompose short videos, they still struggle with long-term temporal consistency. In contrast, we show that learning representations that are both informative and contrastive can significantly enhance both object discovery and temporal consistency on longer videos.", + "bbox": [ + 511, + 90, + 906, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Consistency Achieving temporal consistency is essential for any computer vision task involving video data, whether it is tracking points, bounding boxes, segmentation masks, optical flow, or representations [29, 33, 41, 47, 50, 52, 54]. In object-centric learning for videos, a range of different approaches have been proposed. For example, Yu and Xu [62] apply an object-wise sequential VAE to achieve consistency; Zhao et al. [65] and Li et al. [31] use an explicit memory buffer to maintain historical slot information and a transformer as a predictor using the memory buffer to predict the future; Qian et al. [40] achieve temporal consistency by employing student-teacher distillation to establish semantic and instance correspondence over time; and Traub et al. [48] use a recurrent network with a constancy prior [17].", + "bbox": [ + 511, + 415, + 908, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 643, + 604, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach builds upon the existing input reconstruction-based video object-centric framework [28, 64] by introducing a consistency loss that contrasts the slots across consecutive frames and thereby adapting the model to discover consistent representations. See Fig. 2 for an overview of the SLOT CONTRAST architecture.", + "bbox": [ + 511, + 669, + 906, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Semantic Recurrent Slot Attention Module", + "text_level": 1, + "bbox": [ + 511, + 771, + 877, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our model is an encoder-decoder object-centric architecture based on Slot Attention module (SA) [13] with additional adaptations for sequential inputs similar to SAVi [28], while leveraging pre-trained semantic features as proposed by DINOSAUR [44]. The model consists of three main components: a pre-trained self-supervised dense feature encoder (e.g., DINOv2 [38]), a Recurrent Slot Attention module that", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5402", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd1cd4ef6e9c8bc14306f040f301a1a5c4dc7c204de975b20ac1e2721e9421f9.jpg", + "image_caption": [ + "Figure 2. SLOT CONTRAST model architecture overview. For each frame, we extract patch features $h_t$ using DINOv2 ViT. These features are then used to update the previously initialized or predicted slots, resulting in new slots $S_t$ . The model is trained by contrasting the current frame's slots $S_t$ with the slots from the previous frame $S_{t-1}$ , and by reconstructing the patch features $h_t$ ." + ], + "image_footnote": [], + "bbox": [ + 178, + 87, + 823, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "groups the encoder features into slots and models temporal slot updates, and a decoder that maps slots from each frame to reconstructions of the dense self-supervised features used as inputs. Next, we describe those components in more detail while explaining how to adapt them to the task of consistent object-centric representation learning.", + "bbox": [ + 89, + 334, + 483, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video frame $x_{t}$ , $t \\in \\{1,2,\\dots,T\\}$ and a pretrained, frozen self-supervised DINO model $f$ we first extract $N$ patch features $g_{t}$ ,", + "bbox": [ + 89, + 424, + 483, + 470 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {t} = f \\left(x _ {t}\\right), \\quad g _ {t} \\in \\mathbb {R} ^ {N \\times D}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 479, + 483, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As those frozen features are mostly semantic and are trained only on images, we further adapt them to the task of temporally consistent object discovery. Specifically, each feature vector $g_{t}$ is passed through a MLP $g_{\\psi}$ ,", + "bbox": [ + 89, + 507, + 483, + 570 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = g _ {\\psi} \\left(g _ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 240, + 579, + 482, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to adapt the frozen dense features for object-centric grouping (see App. I for more details and visualizations). Based on the transformed encoder features $h_t$ and a set of slot representations of the previous timestep $S_{t-1}^p$ , with $K$ slots $s_{t-1}^{k,p} \\in S_{t-1}^p$ , we use a recurrent grouping module to extract slot representations. The Recurrent Slot Attention module comprises a grouping module $C_\\theta$ and a predictor module $P_\\omega$ . The former updates slot representations using the standard Slot Attention module [32] on visual features $h_t$ from the encoder, while the latter captures temporal and spatial interactions between slots:", + "bbox": [ + 89, + 604, + 483, + 772 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t} ^ {c} = \\mathrm {C} _ {\\theta} \\left(h _ {t}, S _ {t - 1} ^ {p}\\right), \\quad S _ {t} ^ {p} = \\mathrm {P} _ {\\omega} \\left(S _ {t} ^ {c}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 782, + 483, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both slot-level representations, generated either by the grouping module $S_{t}^{c}$ or the predictor $S_{t}^{p}$ , can be utilized for subsequent decoding or downstream task processing. In our implementation, the slot-level representations from the grouping module $S_{t}^{c}$ are employed for the decoding stage. From now on, we will refer to $S_{t}^{c}$ as $S_{t}$ .", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Slot Attention Initialization Importantly, we found that our setup benefits considerably from a learned initialization $S_0$ , which can influence the efficiency of training across various objectives. Originally, Locatello et al. [32] proposed a randomly sampled query initialization, where slots are sampled from the same Gaussian distribution with learned mean and variance. While such initialization allows different numbers of slots during inference, sampling from the same Gaussian distribution does not create a particularly favorable structure in slot-space. In this work, we use a straightforward learned initialization [19, 43] where a fixed set of initial slot vectors $S_0$ is learned for the entire dataset. Such initialization allows for learning dissimilar initialization queries that consistently attend to different objects.", + "bbox": [ + 511, + 334, + 906, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, for the reconstruction loss objective, we decode reconstructions $\\hat{g}_t$ from all slots using the MLP decoder [44].", + "bbox": [ + 511, + 553, + 906, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Temporal Consistency through Slot Contrast", + "text_level": 1, + "bbox": [ + 511, + 614, + 893, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning is flexible in supporting diverse data sources and loss function designs. By carefully defining positive and negative examples, we can craft robust loss objectives that effectively guide self-supervised representation learning [6]. For instance, video contrastive methods like CVRL [39] leverage augmented video chunks to define positive (from the same video) and negative (from different videos) examples. In object-centric learning Didolkar et al. [8] employed a contrastive loss function to gain controllability over slot representations guided by language. We propose a novel application of a contrastive loss for temporal consistency in object-centric slot representations. In particular, we define positive samples as the representations of the same slot from two consecutive time steps within a video, while negative samples comprise all other slots across the batch between these time steps. An overview of the proposed loss is presented in Fig. 3", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5403", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/19f3df135ff014dad4f726c4f4b904ada324c6286d15262197705a8dd1e38cb1.jpg", + "image_caption": [ + "Figure 3. Overview of the losses used in SLOT CONTRAST. (a) Our proposed temporal consistency objective, slot-slot contrastive loss, operates on a batch of video sequences by enforcing temporal alignment across object slots. For each frame in the sequence, the model groups object features into specific slot representations $S_{t}^{i}$ . The slot-slot contrastive loss then enforces temporal consistency by drawing the corresponding slot representations from adjacent frames closer, while simultaneously pushing apart all other slot representations in the batch—whether they come from different objects within the same video or from objects in other videos. (b) The feature reconstruction loss ensures informativeness of the learned slots by using them to reconstruct original DINOv2 features with an MLP decoder." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 859, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intra-Video Slot-Slot Contrautive Loss To force each slot to be consistent in time, we aim to learn slots that are similar in time while being maximally dissimilar to other slots. Given the sets of slot representations $S_{t - 1}$ and $S_{t}$ at time steps $t - 1$ and $t$ , we want elements $s_{t - 1}^{i}\\in S_{t - 1}$ to be close to the next-frame slots $s_t^i$ corresponding to the same object, while having maximal distance to the next-frame slots $s_t^k,k\\neq i$ corresponding to other objects in the video. The corresponding InfoNCE contrastive loss [37] is defined as $\\mathcal{L}_{\\mathrm{intra}} = \\frac{1}{K}\\sum_{i = 1}^{K}\\ell_{i}^{\\mathrm{intra}}$ with", + "bbox": [ + 89, + 444, + 483, + 598 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {i} ^ {\\text {i n t r a}} = - \\log \\frac {\\exp \\left(\\sin \\left(s _ {t - 1} ^ {i} , s _ {t} ^ {i}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\mathbb {1} _ {[ k \\neq i ]} \\exp \\left(\\sin \\left(s _ {t - 1} ^ {i} , s _ {t} ^ {k}\\right) / \\tau\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 607, + 483, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K = |S_t|$ is a number of slots per frame, $\\mathrm{sim}(u,v) = \\frac{u^{\\top}v}{\\|u\\|_2\\|v\\|_2}$ is the cosine similarity, $\\mathbb{1}_{[.]}$ is an indicator excluding the self-similarity of the slot $s_i$ from the denominator, and $\\tau >0$ is a temperature parameter.", + "bbox": [ + 89, + 648, + 483, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While being a desirable property, intra-video slot contrast can be achieved simply by amplifying the differences between slots in the SA module's first frame initialization $S_0$ . To encourage a stronger focus on video content and instance specificity of the representations, we propose a further improvement over this loss by extending the negative contrast set.", + "bbox": [ + 89, + 714, + 483, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Batch Video Slot-Slot Contrastive Loss To leverage the benefits of larger contrast sets and prevent degenerate solutions relying solely on the initialization of slots, we exploit the fact that the whole batch of videos can be con", + "bbox": [ + 89, + 839, + 485, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sidered a large set of primarily unique object representations. Consequently, we enhance contrast within a video and between videos by including negative slots from the current and subsequent frames of all videos in the batch. Correspondingly, we define our slot-slot contrastive loss as $\\mathcal{L}_{\\mathrm{ssc}} = \\frac{1}{B\\cdot K}\\sum_{j = 1}^{B}\\sum_{i = 1}^{K}\\ell_{i,j}^{\\mathrm{ssc}}$ and", + "bbox": [ + 511, + 444, + 906, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {i, j} ^ {\\mathrm {s s c}} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(s _ {t - 1} ^ {i , j} , s _ {t} ^ {i , j}\\right) / \\tau\\right)}{\\sum_ {b = 1} ^ {B} \\sum_ {k = 1} ^ {K} \\mathbb {1} _ {[ k , b \\neq i , j ]} \\exp \\left(\\operatorname {s i m} \\left(s _ {t - 1} ^ {i , j} , s _ {t} ^ {k , b}\\right) / \\tau\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 547, + 906, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B$ is a number of videos in the batch and $s_t^{i,j}$ denotes the $i$ -th slot of the $j$ -th video at time $t$ . For more details on slot-slot contrastive loss implementation, see App. C.", + "bbox": [ + 511, + 608, + 903, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We find that this approach significantly enhances the effectiveness of the slot-slot contrastive loss. Furthermore, since all videos in the batch are processed with the same initial state $S_0$ , this loss function avoids suboptimal solutions that rely solely on the uniqueness of the initialization, instead encouraging object discovery as the basis for contrast.", + "bbox": [ + 511, + 655, + 906, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Final Loss To encourage scene decomposition we use a feature reconstruction loss, similar to DINOSAUR [44] and VideoSAUR [64]. Our final loss function combines the reconstruction loss with our proposed contrastive loss $\\mathcal{L}_{\\mathrm{ssc}}$ weighted by the hyperparameter $\\alpha$ (see Table S1 for details on how the hyperparameters are set):", + "bbox": [ + 511, + 765, + 908, + 857 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {t = 1} ^ {T - 1} \\mathcal {L} _ {\\mathrm {r e c}} \\left(\\mathrm {h} _ {\\mathrm {t}}, \\hat {\\mathrm {h}} _ {\\mathrm {t}}\\right) + \\alpha \\mathcal {L} _ {\\mathrm {s s c}} \\left(S _ {t - 1}, S _ {t}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 866, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5404", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1552eaed03f6bda5bcaefc96eaadc617f0acccdd38cb70ef143e42d884b1cae8.jpg", + "table_caption": [ + "Table 1. Consistent object-discovery performance of SLOT CONTRAST in comparison with SAVi, STEVE, VideoSAUR on MOVi-C, MOVi-E, and YouTube-VIS datasets. VideoSAURv2 is an improved version of the VideoSAUR trained on DINOv2 features. Both metrics are computed for the whole video (24 frames for MOVi, up to 76 frames for YouTube-VIS)." + ], + "table_footnote": [], + "table_body": "
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
SAVi [28]22.213.642.816.0--
STEVE [45]36.126.550.626.61519.1
VideoSAUR [64]64.838.973.935.628.926.3
VideoSAURv2--77.134.431.229.7
SLOT CONTRAST69.332.782.929.238.033.7
", + "bbox": [ + 93, + 176, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 290, + 225, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method's temporal consistency on two downstream tasks: object discovery and latent object dynamics prediction. Our experiments address three main questions: (1) How does our model compare to state-of-the-art methods in both temporal consistency and scene decomposition? (2) How effective are our model's learned representations for the challenging downstream task of object dynamics prediction and for object tracking under full occlusions? (3) How important are the different components of our model and loss function for temporal consistency?", + "bbox": [ + 88, + 316, + 485, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets To evaluate our method in the controlled setting, we use MOVi-C and MOVi-E synthetic datasets generated by Kubric [16]. MOVi-C includes richly textured everyday objects, featuring up to 11 objects per scene, while MOVi-E expands this to 23 objects and introduces basic linear camera motion. In addition, to study the scalability of our method to real-world data, we evaluate our method on the real-world YouTube-VIS 2021 (YTVIS21) video dataset [60]. YTVIS21 is an unconstrained, real-world dataset sourced from YouTube, capturing a diverse range of scenes (for more details, see App. E).", + "bbox": [ + 88, + 487, + 483, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics Similar to other object-centric video methods [11, 28, 45, 64], to evaluate consistent object discovery, we use the video foreground adjusted rand index (FG-ARI) [14], measuring how well objects are split. In addition, we evaluate the sharpness of masks using the video intersection over union with mean best overlap matching (mBO) metric [44, 64]. Both metrics are computed over the full video and thus reflect how consistent object discovery is. In addition, to investigate the effects of the temporal consistency inductive bias on the per-frame object discovery itself, we use per-frame FG-ARI (image FG-ARI), which we independently compute for each frame and average afterwards. More details can be found in App. F.", + "bbox": [ + 89, + 672, + 485, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, when evaluating how well object-centric representations perform for object dynamics prediction (see Sec. 4.2),", + "bbox": [ + 89, + 869, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we employ the same evaluation metrics as in the object discovery task: FG-ARI and mBO. This time, however, these metrics are computed by comparing the predicted masks (obtained by decoding the predicted slots [56]) with the ground-truth future masks.", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Object Discovery", + "text_level": 1, + "bbox": [ + 511, + 174, + 684, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We employ the DINOv2 model as our feature encoder, using ViT-S/14 for the MOVi-C dataset and ViT-B/14 for MOVi-E and YTVIS21. The slot dimension is set to 128 for MOVi-E and 64 for both MOVi-C and YTVIS21. For the MOVi datasets, we use a resolution of (336, 336), generating $24 \\times 24$ patches yielding 576 ViT tokens, while for YTVIS21, a resolution of (518, 518) yields 1369 tokens. Full details are provided in App. A.", + "bbox": [ + 511, + 196, + 908, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines We compare SLOT CONTRAST against the previously proposed SAVi [28] and STEVE [45] that employ an image reconstruction objective and with the state-of-the-art method VideoSAUR [64] that uses self-supervised feature reconstructions. Additionally, for a fair comparison, we trained a modification of VideoSAUR with DINOv2 features (referred to as VideoSAURv2). The implementation details are provided in App. G. In addition, to assess how closely SLOT CONTRAST approaches supervised methods, we compared it with SAM2 [41] as a supervised zero-shot baseline for temporal consistency and to weakly-supervised by depth SAVi++ [11] method. The results are in App. G.", + "bbox": [ + 511, + 335, + 908, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporally Consistent Object Discovery (Table 1 & Figure 4) SLOT CONTRAST significantly outperforms both SAVi and STEVE by a wide margin. When compared to VideoSAUR using its default parameters, our approach demonstrates higher consistency in terms of video FG-ARI scores. Compared to VideoSAUR and VideoSAURv2, SLOT CONTRAST achieves superior video scene decomposition (measured by FG-ARI scores). However, on synthetic datasets SLOT CONTRAST's masks are less sharp (as reflected by mBO). Notably, on the most challenging real-world YouTube-VIS data, our method surpasses both versions of VideoSAUR, achieving better performance on FG-ARI (+6.8) and mBO (+4). This shows that, given a large enough resolution and natural data inputs well aligned with DINOv2, SLOT CONTRAST can decompose unconstrained videos into consistent object representations. More examples are illustrated in App. M.", + "bbox": [ + 511, + 535, + 908, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Per-Frame Scene Decomposition (Table 2) Previous research has shown the effectiveness of specific inductive biases and training objectives for unsupervised object discovery, such as reconstructing in semantic space or leveraging motion cues in video data. Building on this, we demonstrate that the contrastive nature of our temporal consistency loss", + "bbox": [ + 511, + 809, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5405", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e151a414bf71b285faae8a9dd142071c0fec9f5db2d1288c66e90ffe17ed3a4a.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with VideoSAURv2 on YouTube-VIS dataset. In challenging situations (e.g., almost full occlusions at $t = 24$ of the 1st video and $t = 14$ of the 2nd video), VideoSAURv2 reassigns slots to different objects (pink arrows), whereas SLOT CONTRAST consistently assigns slots to the same object (green arrows). Note that the colors of the masks are matched manually for better visual comparison." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 504, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74c39b6206646b68ef77a94fb0899c1a459fe242b9d3256ea2009206d760cc27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 90, + 903, + 253 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f53124d02a11d0e7c2372a18adb1827f274b6069b7d7fa98ec59f89c69d07343.jpg", + "table_caption": [ + "Table 2. Quantitative Results on MOVi-E in terms of per-frame Image FG-ARI. The methods are grouped by the target data they train on: only images $(\\mathcal{I})$ , videos with motion segmentation annotations $(\\mathcal{V} + \\mathcal{M})$ , and only videos $(\\mathcal{V})$ ." + ], + "table_footnote": [], + "table_body": "
ModelObjectiveImage FG-ARI ↑
↓LSD [21]Image Rec.53.4
SlotDiffusion [57]Image Rec.60.0
DINOSAUR [44]Image Rec.65.1
+MoToK [3]+Mot. Seg.66.7
Safadoust et al. [42]+GT Flow78.3
DIOD [25]+Mot. Seg.82.2
↑STEVE [45]Video Rec.54.1
VideoSAUR [64]Temp. Sim.78.4
SOLV [2]Mid. Fr. Pred.80.8
SLOT CONTRASTSlot Contrast84.8
", + "bbox": [ + 106, + 398, + 465, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "function yields improved scene decomposition as a byproduct. This occurs because our loss function encourages the model to learn consistent feature representations for objects across frames, leading to an adaptive process where dense features become more contrastive, thereby enhancing object discovery in individual frames.", + "bbox": [ + 89, + 625, + 483, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with prior approaches in terms of per-frame object discovery, using the image FG-ARI metric for evaluation. Specifically, we compare three categories of methods: image-based, video-based, and methods that use videos with additional motion cues. Image-based methods use only images as a target (feature reconstruction based DINOSAUR [44] and diffusion-based LSD [21] and SlotDiffusion [57] methods). Video-based methods use only videos as targets: STEVE [45] reconstructs current frame features, SOLV [2] predicts middle frame features, and VideoSAUR [64] predicts temporal feature-similarities. Finally, we also compare with weakly-supervised methods", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "using motion masks [3, 25] or ground truth (GT) optical flow [42].", + "bbox": [ + 511, + 339, + 903, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results on MOVi-E dataset are presented in Table 2, with comparisons across additional datasets provided in App. H. Using temporal signals from the video using feature reconstruction is better than object discovery based on images. Next, additional objectives that exploit the temporal structure of the videos allow even better scene decomposition. Notably, our method, which combines a feature reconstruction objective with a simple contrastive objective, leads to state-of-the-art performance reaching 84.8 per-frame FG-ARI, outperforming methods [3, 25] that use motion segmentation masks for object discovery.", + "bbox": [ + 511, + 369, + 905, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robustness to Full Occlusions To evaluate our method's robustness in handling complete object occlusions—a challenging scenario for maintaining consistency—we conduct experiments using a targeted subset of the MOVi-C dataset that contains sequences where objects are fully occluded. For evaluation, we retain only the ground-truth masks for the objects that experience occlusion.", + "bbox": [ + 511, + 553, + 905, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We find that the feature reconstruction baseline achieves only $16\\%$ mBO vs. our method obtains $21\\%$ mBO on fully occluded objects. Our results suggest that SLOT CONTRAST significantly enhances consistency during object disappearances and reappearance. We refer the reader to Fig. 5 and App. K for visual examples and more details.", + "bbox": [ + 511, + 659, + 905, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Object Dynamics Prediction", + "text_level": 1, + "bbox": [ + 511, + 757, + 766, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup To evaluate performance on the task of predicting object dynamics, we train a dynamics module using the object-centric representations inferred by a pretrained object-centric model. For this dynamics module, we select SlotFormer [56], which predicts the slots autoregressively for $K$ rollout steps based on the slots inferred from $T$ burn-in frames preceding the prediction horizon. In our setup, we use 14 burn-in frames and 10 rollout steps. Since SlotFormer", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5406", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0424434c6e1046c3a168ce24376ddcef60f00988eea14894ed076e58b147d36b.jpg", + "image_caption": [ + "Figure 5. Comparison of the Feature Reconstruction (Feat. Rec.) baseline, the slot-slot contrastive loss using only slots from the same video as the contrast set (Intra-video Contrast), and SLOT CONTRAST on the MOVi-C dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 482, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is trained independently from the object-centric model, we first train the latter, subsequently extending the datasets with the inferred slots for each frame. This approach avoids the computational complexity of training SlotFormer by removing the necessity to encode frames into the slot space at each training step. A brief introduction to SlotFormer and the implementation details can be found in App. J.", + "bbox": [ + 89, + 414, + 483, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines We compare SlotFormer [56] trained on object-centric representations derived from a model trained using only feature reconstruction loss with SlotFormer trained on representations from SLOT CONTRAST. Both models perform reconstruction in feature space rather than pixel space, so we use only the slot reconstruction loss for training.", + "bbox": [ + 89, + 537, + 483, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quality of Predicted Masks (Table 3 and Figure 6) We note that our model has a significantly better FG-ARI on MOVi-C, while mBO is comparable to that of the baseline. On MOVi-E, the performance of our model is comparable to that of the baseline, which highlights the difficulty of adapting to videos with camera motion. There is also a", + "bbox": [ + 89, + 646, + 483, + 739 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d714bd799befac09576f4ec27c3994f5ad8328c5b03369e96ed0c3ecf468df5d.jpg", + "image_caption": [ + "Figure 6. Object dynamics prediction task on MOVi-C using SLOT CONTRAST slots using SlotFormer [56]." + ], + "image_footnote": [], + "bbox": [ + 96, + 753, + 480, + 862 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6aadca9f25b3b947f6d8aac5c8f638b6435c35649ce2d2e1a8487d1c3543cfe6.jpg", + "table_caption": [ + "Table 3. Downstream task of predicting object dynamics. Comparison of predictions made by SlotFormer based on representations obtained from SLOT CONTRAST and from Feature Reconstruction." + ], + "table_footnote": [], + "table_body": "
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
Feat. Rec. + SF50.725.970.624.327.428.9
Ours + SF63.826.170.524.929.229.6
", + "bbox": [ + 517, + 135, + 903, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "slight improvement in FG-ARI on YTVIS21, while mBO remains comparable. It is worth noting that predicting the motion in this dataset is especially challenging, given the large diversity of possible scenarios.", + "bbox": [ + 511, + 231, + 906, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 511, + 309, + 614, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we investigate key components of our approach, including the impact of the contrastive loss and the type of slot initialization. In addition, we study how effective SLOT CONTRAST is in automatically shutting down slots in correspondence to the scene's complexity.", + "bbox": [ + 511, + 333, + 906, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of Loss Components (Table 4 and Figure 5) To demonstrate the value of the proposed slot-slot contrastive loss, we carry out an ablation study, comparing it with the feature reconstruction loss [44] and the intra-video contrastive loss, which contrasts slot representations in a single video. Using the intra-video contrastive loss yields improvements over the feature reconstruction baseline (+5.1 FG-ARI and +1.5 mBO on MOVi-C). However, we observe that in more challenging situations, the intra-video contrastive loss leads to failure cases such as shutting down too many slots (see Fig. 5). Next, we observe that by extending the contrast to the full batch of videos, SLOT CONTRAST learns more consistent representations (+19.6 FG-ARI and +5.3 mBO). This change increases the difficulty of the learning task, which prevents the model from relying on superficial patterns like slot initializations or object positions.", + "bbox": [ + 511, + 441, + 908, + 684 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bcfc7f1984f0184801dcbd6e212f98ba26e9088533413446540741c602ae7171.jpg", + "image_caption": [ + "Figure 7. Distribution of ground truth and predicted object numbers (i.e., number of active slots) on the MOVi-C dataset." + ], + "image_footnote": [], + "bbox": [ + 535, + 710, + 883, + 864 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5407", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a39ec994f868ce900bd0e600943b2b8b6bf1cc57102373ac55cb73113c416166.jpg", + "table_caption": [ + "Table 4. Ablation of loss components used by SLOT CONTRAST on MOVi-C, MOVi-E, and YouTube-VIS Datasets." + ], + "table_footnote": [], + "table_body": "
Feat. Rec.Intra-video ContrastSlot-Slot ContrastMOVi-CMOVi-EYouTube-VIS
LrecLintraLsscFG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
✓49.727.479.828.435.331.4
✓✓54.828.978.729.135.733.6
✓✓69.332.782.929.238.033.7
", + "bbox": [ + 143, + 108, + 854, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1cb9367eeb5115793ca5e0f2dcf441d6197d39a8894a39ad95f1bf05fdfb603a.jpg", + "table_caption": [ + "Table 5. Comparison of the random initialization (RI) and learned initialization (LI) techniques." + ], + "table_footnote": [], + "table_body": "
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
Feat. Rec. (RI)45.327.271.128.335.230.2
Feat. Rec. (LI)49.427.879.828.435.331.4
SLOT CONTRAST (RI)62.932.475.328.436.130.8
SLOT CONTRAST (LI)69.332.782.929.238.033.7
", + "bbox": [ + 93, + 244, + 480, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of the First Frame Initialization (Table 5) In video object-centric learning, slots are typically initialized based on those from the previous time step [28], while the first frame is initialized from learnable parameters. Previous real-world object-centric methods mostly used random initialization samples from Gaussian distribution [2, 64], in this work, we study the impact of the type of initialization under our contrastive objective. The findings are outlined in Table 5. Our experiments indicate that when combined with slot-slot contrastive loss, learned initialization significantly outperforms random initialization. We hypothesize that this improvement stems from the ability of learned initializations to shape the initial state in a way that enhances contrastiveness, a benefit not achievable with random initialization. For similarity visualizations and more details, see App. B.", + "bbox": [ + 88, + 352, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of Active Slots per Video (Figure 7) In models based on the Slot Attention mechanism, all available slots are typically utilized [44], leading to a mismatch between the predicted number of components in scene decomposition and the ground-truth number of objects in the scene. This can cause the random splitting of the objects between slots and non-consistent scene representations when slots are reassigned from one object to a part of another object. To address this challenge, it is important to study whether redundant slots can be effectively deactivated. Recently, AdaSlot [12] introduced a discrete slot sampling module, coupled with a complexity-aware prior, to penalize redundant slots explicitly. Similarly, SOLV [2] used agglomerative clustering to merge redundant slots. In this work, we investigate whether SLOT CONTRAST is capable of accurately determining the number of objects in a scene without relying on explicit priors to minimize the number of active slots.", + "bbox": [ + 91, + 598, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare ground truth and predicted object density on MOVI-C dataset, as shown in Fig. 7. While the feature reconstruction model yields predictions within a narrow", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "range—creating a sharp peak near a predefined number of slots—our model, similarly to AdaSlot [12], achieves a smoother prediction distribution that aligns more closely with the ground truth (note that the consistent shift is because 2–3 slots are used for the background, while the ground truth density is computed only for foreground objects). Interestingly, SLOT CONTRAST achieves this without requiring an explicit prior toward sparsity.", + "bbox": [ + 511, + 214, + 906, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 354, + 633, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SLOT CONTRAST advances unsupervised video object-centric learning by significantly improving the temporal consistency of object representations. Our method explicitly incentivizes temporal consistency by adding a self-supervised contrastive loss. We showed that this loss is not only beneficial for consistency, but also enhances object discovery: SLOT CONTRAST achieves state-of-the-art results on challenging synthetic datasets with many objects and the unconstrained real-world YouTube-VIS dataset. Furthermore, consistent representations directly support temporal downstream tasks such as unsupervised object dynamics prediction and allow for tracking of objects through full occlusions. Finally, SLOT CONTRAST effectively shuts down non-unique slots, leading to a sparser representation that captures the true object distribution more faithfully. Taken together, we expect these improvements to pave the way for broader adoption of video object-centric representations, for instance in applications like word modeling, autonomous control, or video question answering.", + "bbox": [ + 511, + 382, + 908, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations of our work include the fixed number of slots during initialization. Additionally, we cannot directly control the segmentation granularity of entities. Further limitations and failure cases are discussed in App. L.", + "bbox": [ + 511, + 671, + 906, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future work could explore several promising directions. First, one could use SLOT CONTRAST's robust and consistent representations for learning compositional world models from real-world robotics data to enable object-centric planning and control. Second, investigating the compatibility of our contrastive loss with other object-centric learning approaches with different inductive biases, such as SlotDiffusion [57]. Finally, improving the compactness of object masks [24] to achieve more precise object segmentation masks could also benefit unsupervised class-agnostic video object segmentation applications.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5408", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Christian Gumbsch for useful suggestions and Ke Fan for providing additional details and results for AdaSlot method. Andrii Zadaianchuk is funded by the European Union (ERC, EVA, 950086). The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Maximilian Seitzer. Georg Martius is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. This work was supported by the ERC - 101045454 REAL-RL. We acknowledge the support from the German Federal Ministry of Education and Research (BMBF) through the Tübingen AI Center (FKZ: 01IS18039B).", + "bbox": [ + 89, + 113, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 308, + 187, + 324 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In International Conference on Computer Vision (ICCV), 2015. 1", + "[2] Görkay Aydemir, Weidi Xie, and Fatma Güney. Self-supervised object-centric learning for videos. In NeurIPS, 2023. 1, 2, 6, 8", + "[3] Zhipeng Bao, Pavel Tokmakov, Yu-Xiong Wang, Adrien Gaidon, and Martial Hebert. Object discovery from motion-guided tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22972-22981, 2023. 6", + "[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. MONet: Unsupervised Scene Decomposition and Representation. arXiv:1901.11390, 2019. 1, 2", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. ICCV, 2021. 1, 2", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In ICML, 2020. 3", + "[7] Aniket Didolkar, Andrii Zadaianchuk, Anirudh Goyal, Mike Mozer, Yoshua Bengio, Georg Martius, and Maximilian Seitzer. Zero-shot object-centric representation learning. arXiv preprint arXiv:2408.09162, 2024. 1", + "[8] Aniket Rajiv Didolkar, Andrii Zadaianchuk, Rabiul Awal, Maximilian Seitzer, Efstratios Gavves, and Aishwarya Agrawal. Ctrl-o: Language-controllable object-centric visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3", + "[9] Andrea Dittadi, Samuele Papa, Michele De Vita, Bernhard Schölkopf, Ole Winther, and Francesco Locatello. Generalization and Robustness Implications in Object-Centric Learning. In ICML, 2022. 1", + "[10] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen" + ], + "bbox": [ + 93, + 334, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. PaLM-E: An embodied multimodal language model. In ICML, 2023. 1", + "bbox": [ + 547, + 92, + 906, + 160 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[11] Gamaleldin Fathy Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael Curtis Mozer, and Thomas Kipf. $\\mathsf{SAV}\\mathsf{i} + +$ : Towards End-to-End Object-Centric Learning from Real-World Videos. In NeurIPS, 2022. 1, 2, 5", + "bbox": [ + 516, + 162, + 906, + 215 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[12] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23062-23071, 2024. 8", + "bbox": [ + 517, + 218, + 906, + 287 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[13] Locatello Francesco, Weissenborn Dirk, Unterthiner Thomas, Mahendran Aravindh, Heigold Georg, Uszkoreit Jakob, Dosovitskiy Alexey, and Kipf Thomas. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. 2", + "bbox": [ + 517, + 287, + 906, + 356 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-Object Representation Learning with Iterative Variational Inference. In ICML, 2019. 1, 2, 5", + "[15] Klaus Greff, Sjoerd Van Steenkiste, and Jürgen Schmidhuber. On the Binding Problem in Artificial Neural Networks. arXiv:2012.05208, 2020. 2", + "[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A Scalable Dataset Generator. In CVPR, 2022. 5", + "[17] Christian Gumbsch, Martin V Butz, and Georg Martius. Sparsely changing latent states for prediction and planning in partially observable domains. In NeurIPS, 2021. 2", + "[18] Dan Haramati, Tal Daniel, and Aviv Tamar. Entity-centric reinforcement learning for object manipulation from pixels. In ICLR, 2024. 1", + "[19] Baoxiong Jia, Yu Liu, and Siyuan Huang. Improving object-centric learning with query optimization. arXiv preprint arXiv:2210.08990, 2022.3", + "[20] Jindong Jiang, Sepehr Janghorbani, Gerard de Melo, and Sungjin Ahn. SCALOR: Generative World Models with Scalable Object Representations. In ICLR, 2020. 1, 2", + "[21] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In NeurIPS, 2023. 2, 6", + "[22] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 2" + ], + "bbox": [ + 93, + 334, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5409", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthew, Antonia Creswell, Matthew Botvinick, Alexander Lerchner, and Christopher P. Burgess. SIMONE: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. In NeurIPS, 2021. 2", + "[24] Ioannis Kakogeorgiou, Spyros Gidaris, Konstantinos Karantzalos, and Nikos Komodakis. Spot: Self-training with patch-order permutation for object-centric learning with autoregressive transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22776-22786, 2024. 2, 8", + "[25] Sandra Kara, Hejer Ammar, Julien Denize, Florian Chabot, and Quoc-Cuong Pham. Diod: Self-distillation meets object discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3975-3985, 2024. 6", + "[26] Laurynas Karazija, Iro Laina, and Christian Rupprecht. ClevrTex: A Texture-Rich Benchmark for Unsupervised Multi-Object Segmentation. In NeurIPS Track on Datasets and Benchmarks, 2021. 2", + "[27] Thomas Kipf, Elise Van der Pol, and Max Welling. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. 1, 2", + "[28] Thomas Kipf, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-centric Learning from Video. In ICLR, 2022. 1, 2, 5, 8", + "[29] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2", + "[30] Adam Kosiorek, Hyunjik Kim, Yee Whye Teh, and Ingmar Posner. Sequential Attend, Infer, Repeat: Generative Modelling of Moving Objects. In NeurIPS, 2018. 2", + "[31] Jian Li, Pu Ren, Yang Liu, and Hao Sun. Reasoning-enhanced object-centric learning for videos. arXiv preprint arXiv:2403.15245, 2024. 2", + "[32] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-Centric Learning with Slot Attention. In NeurIPS, 2020. 1, 2, 3", + "[33] Zijia Lu, Bing Shuai, Yanbei Chen, Zhenlin Xu, and Davide Modolo. Self-supervised multi-object tracking with path consistency. arXiv preprint, 2024. 2", + "[34] Amir Mohammad Karimi Mamaghan, Samuele Papa, Karl Henrik Johansson, Stefan Bauer, and Andrea Dittadi. Exploring the effectiveness of object-centric representations in visual question answering: Comparative insights with foundation models. arXiv preprint arXiv:2407.15589, 2024. 1", + "[35] Sha Meng, Dian Shao, Jiacheng Guo, and Shan Gao. Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16264-16273, 2023. 1", + "[36] Cristian Meo, Akihiro Nakano, Mircea Lică, Aniket Didolkar, Masahiro Suzuki, Anirudh Goyal, Mengmi Zhang, Justin" + ], + "bbox": [ + 91, + 92, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dauwels, Yutaka Matsuo, and Yoshua Bengio. Object-centric temporal consistency via conditional autoregressive inductive biases. arXiv preprint arXiv:2410.15728, 2024. 2", + "[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4", + "[38] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision, 2023. 1, 2", + "[39] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6964-6974, 2021. 3", + "[40] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In ICCV, 2023. 2", + "[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 5", + "[42] Sadra Safadoust and Fatma Güney. Multi-object discovery by low-dimensional object motion. In ICCV, 2023. 2, 6", + "[43] Mehdi SM Sajjadi, Daniel Duckworth, Aravindh Mahendran, Sjoerd Van Steenkiste, Filip Pavetic, Mario Lucic, Leonidas J Guibas, Klaus Greff, and Thomas Kipf. Object scene representation transformer. Advances in Neural Information Processing Systems, 35:9512-9524, 2022. 3", + "[44] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. In ICLR, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[45] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple Unsupervised Object-Centric Learning for Complex and Naturalistic Videos. In NeurIPS, 2022. 1, 2, 5, 6", + "[46] Matthias Tangemann, Steffen Schneider, Julius Von Kügelgen, Francesco Locatello, Peter Gehler, Thomas Brox, Matthias Kümmerer, Matthias Bethge, and Bernhard Schölkopf. Unsupervised object learning via common fate. arXiv preprint arXiv:2110.06562, 2021. 2", + "[47] Pavel Tokmakov, Allan Jabri, Jie Li, and Adrien Gaidon. Object permanence emerges in a random walk along memory. arXiv preprint arXiv:2204.01784, 2022. 2", + "[48] Manuel Traub, Frederic Becker, Sebastian Otte, and Martin V Butz. Looping loci: Developing object permanence from videos. arXiv preprint arXiv:2310.10372, 2023. 2", + "[49] Manuel Traub, Sebastian Otte, Tobias Menge, Matthias Karlbauer, Jannik Thuemmel, and Martin V. Butz. Learning What" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5410", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Where: Disentangling Location and Identity Tracking Without Supervision. In ICLR, 2023. 2", + "[50] Basile Van Hoorick, Pavel Tokmakov, Simon Stent, Jie Li, and Carl Vondrick. Tracking through containers and occluders in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13802-13812, 2023. 2", + "[51] Sjoerd van Steenkiste, Michael Chang, Klaus Greff, and Jürgen Schmidhuber. Relational neural expectation maximization: Unsupervised discovery of objects and their interactions. In ICLR, 2018. 2", + "[52] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity Abstraction in Visual Model-based Reinforcement Learning. In Conference on Robot Learning, 2019. 1, 2", + "[53] Ning Wang, Yibing Song, Chao Ma, Wengang Zhou, Wei Liu, and Houqiang Li. Unsupervised deep tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1308-1317, 2019. 1", + "[54] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2566-2576, 2019. 2", + "[55] Nicholas Watters, Loic Matthew, Christopher P Burgess, and Alexander Lerchner. Spatial broadcast decoder: A simple architecture for learning disentangled representations in vaes. arXiv preprint arXiv:1901.07017, 2019. 2", + "[56] Ziyi Wu, Nikita Dvornik, Klaus Greff, Thomas Kipf, and Animesh Garg. Slotformer: Unsupervised visual dynamics simulation with object-centric models. In ICLR, 2023. 1, 2, 5, 6, 7", + "[57] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotdiffusion: Object-centric generative modeling with diffusion models. In NeurIPS, 2023. 1, 2, 6, 8", + "[58] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Slot-VLM: Object-event slots for video-language modeling. In NeurIPS, 2024. 1", + "[59] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1", + "[60] Linjie Yang, Yuchen Fan, Yang Fu, and Ning Xu. The 3rd large-scale video object segmentation challenge - video instance segmentation track, 2021. 1, 5", + "[61] Jaesik Yoon, Yi-Fu Wu, Heechul Bae, and Sungjin Ahn. An investigation into pre-training object-centric representations for reinforcement learning. In ICML, 2023. 1", + "[62] Haonan Yu and Wei Xu. Vonet: Unsupervised video object learning with parallel u-net attention and object-wise sequentialvae.arXiv preprint arXiv:2401.11110, 2024.2", + "[63] Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius. Self-supervised Visual Reinforcement Learning with Object-centric Representations. In ICLR, 2020. 1", + "[64] Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius. Object-centric learning for real-world videos by predicting temporal feature similarities. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023), 2023. 1, 2, 4, 5, 6, 8" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Zixu Zhao, Jiaze Wang, Max Horn, Yizhuo Ding, Tong He, Zechen Bai, Dominik Zietlow, Carl-Johann Simon-Gabriel, Bing Shuai, Zhuowen Tu, Thomas Brox, Bernt Schiele, Yanwei Fu, Francesco Locatello, Zheng Zhang, and Tianjun Xiao. Object-centric multiple object tracking. ICCV, 2023. 2", + "[66] Daniel Zoran, Rishabh Kabra, Alexander Lerchner, and Danilo J. Rezende. PARTS: Unsupervised segmentation with slots, attention and independence maximization. In ICCV, 2021. 2" + ], + "bbox": [ + 516, + 90, + 906, + 217 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5411", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_model.json b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7f110789255f014b8aa8af77e4a94d650b87cd23 --- /dev/null +++ b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.151, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "Temporally Consistent Object-Centric Learning by Contrasting Slots" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.173, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Anna Manasyan1 Maximilian Seitzer1 Filip Radovic1 Georg Martius1,3 Andrii Zadaianchuk2 \n1University of Tübingen, Tübingen, Germany \n2University of Amsterdam, Amsterdam, Netherlands \n3Max Planck Institute for Intelligent Systems, Tübingen, Germany" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.23, + 0.768, + 0.244 + ], + "angle": 0, + "content": "anna.manasyan@student.uni-tuebingen.de a.zadaianchuk@uva.nl" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.264, + 0.905, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.486, + 0.908, + 0.515 + ], + "angle": 0, + "content": "Figure 1. SLOT CONTRAST: Our method introduces a temporal contrastive loss that enhances temporal consistency in unsupervised video object-centric models. It stabilizes representations across frames, achieving state-of-the-art object discovery on complex real-world videos." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.54, + 0.327, + 0.556 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.581, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Unsupervised object-centric learning from videos is a promising approach to extract structured representations from large, unlabeled collections of videos. To support downstream tasks like autonomous control, these representations must be both compositional and temporally consistent. Existing approaches based on recurrent processing often lack long-term stability across frames because their training objective does not enforce temporal consistency. In this work, we introduce a novel object-level temporal contrastive loss for video object-centric models that explicitly promotes temporal consistency. Our method significantly improves the temporal consistency of the learned object-centric representations, yielding more reliable video decompositions that facilitate challenging downstream tasks such as unsupervised object dynamics prediction. Furthermore, the inductive bias added by our loss strongly improves object discovery, leading to state-of-the-art results on both synthetic and real-world datasets, outperforming even weakly-supervised methods that leverage motion masks as additional cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.869, + 0.466, + 0.884 + ], + "angle": 0, + "content": "Visit slotcontrast.github.io for videos and further details." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.54, + 0.645, + 0.556 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Object-centric learning (OCL) [4, 14, 32, 44] is a rapidly advancing area of visual representation learning that enables autonomous systems to represent, understand, and model high-dimensional data directly in terms of its constituent entities. Structured object-centric representations (often referred to as slots [32]) facilitate generalization and robustness [7, 9] of scene representations across diverse downstream tasks, from visual question answering [1, 8, 34, 58] to control [10, 18, 61, 63]. Of particular interest are video-based object-centric methods [2, 11, 20, 28, 45, 57, 64] that learn to represent objects that evolve and interact over time. These representations make the methods powerful tools for applications such as unsupervised online object tracking [35, 53] and structured world modeling [27, 52, 56]. Unsupervised object-centric learning on videos has seen significant progress in recent years [2, 57, 64], mainly due to the use of pre-trained representations from self-supervised foundational models [5, 38] coupled with diverse training datasets like YouTube-VIS [59, 60]. Nevertheless, these methods still face significant challenges, especially maintaining consistent object-centric representations across time and uniquely representing each object—critical factors for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5401" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "successful multi-object tracking and modeling of dynamic scenes [28, 52, 56]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Temporal consistency [15, 31, 62] in object-centric representations refers to maintaining the same representation placeholder, called slot, for an object throughout a video sequence, effectively serving as a stable object-specific identifier over time. Existing unsupervised object-centric methods [27, 36, 48] aiming to discover consistent representations have primarily been studied on toy datasets with limited complexity [22, 26, 46]. In contrast, real-world video sequences present numerous challenges, including object occlusions, reappearance, and complex multi-object interactions, which complicate maintaining consistent object representations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.288, + 0.486, + 0.544 + ], + "angle": 0, + "content": "In this paper, we introduce a novel method to address the challenge of maintaining consistent temporal representations in object-centric models, extending the line of research on slot-based unsupervised video models [11, 64]. Our approach (named SLOT CONTRAST) scales to real-world video data and produces consistent object-centric representations. Notably, it achieves these results without requiring any human annotations. In particular, we propose a novel self-supervised contrastive learning objective, which contrasts slot representations throughout the batch while ensuring temporal coherence across consecutive frames. In addition, we modify the slot's initialization strategy [32] to promote distinct, contrastive representations. This combination leads to improved temporal consistency of learned representations, which we show to be highly effective for challenging downstream tasks such as unsupervised object tracking and latent object dynamics learning." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.545, + 0.384, + 0.559 + ], + "angle": 0, + "content": "Overall, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.562, + 0.483, + 0.606 + ], + "angle": 0, + "content": "- We propose the novel slot-slot contrastive loss that sets the state-of-the-art in temporal consistency when integrated into slot-based video processing methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.607, + 0.483, + 0.665 + ], + "angle": 0, + "content": "- We develop SLOT CONTRAST, a simple and effective OCL architecture using the slot-slot contrastive loss paired with learned initialization that scales to real-world data, such as YouTube videos." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.668, + 0.483, + 0.727 + ], + "angle": 0, + "content": "- We extensively study the usefulness of our learned object-centric representations for challenging downstream tasks, including unsupervised online tracking with complete occlusions and latent object dynamics modeling." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.728, + 0.483, + 0.787 + ], + "angle": 0, + "content": "- We show that SLOT CONTRAST does not only improve the temporal consistency of the representations, but also achieves state-of-the-art on the object discovery task, outperforming weakly-supervised models using motion cues." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.562, + 0.483, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.233, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Unsupervised video object-centric learning There exists an extensive body of research [2, 11, 14, 20, 23, 28, 30, 42, 45, 49, 51, 52, 64] on discovering objects from video without any human annotations, primarily through tracking either object bounding boxes or masks. To achieve this, most of these" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.395 + ], + "angle": 0, + "content": "works combine an auto-encoder framework with a simple reconstruction objective, adding inductive biases for object discovery through structured encoders [4, 32] and decoders [55]. In particular, many modern object-centric image models [8, 21, 24, 44, 57] use a latent slot attention module [32] to extract object representations and corresponding object masks. For video data, most current methods [11, 28, 45, 49, 64, 66] connect slots across frames, with slots from the previous frame initializing those in the current frame. Notably, recent approaches [2, 40, 64] have successfully scaled object discovery to real-world unconstrained videos. To achieve this, SOLV [2] introduces temporal consistency via agglomerative clustering and prediction of middle-frame features, whereas VideoSAUR [64] learns object-centric representations by predicting temporal similarity of self-supervised features [5, 38]. While such methods can decompose short videos, they still struggle with long-term temporal consistency. In contrast, we show that learning representations that are both informative and contrastive can significantly enhance both object discovery and temporal consistency on longer videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.909, + 0.629 + ], + "angle": 0, + "content": "Temporal Consistency Achieving temporal consistency is essential for any computer vision task involving video data, whether it is tracking points, bounding boxes, segmentation masks, optical flow, or representations [29, 33, 41, 47, 50, 52, 54]. In object-centric learning for videos, a range of different approaches have been proposed. For example, Yu and Xu [62] apply an object-wise sequential VAE to achieve consistency; Zhao et al. [65] and Li et al. [31] use an explicit memory buffer to maintain historical slot information and a transformer as a predictor using the memory buffer to predict the future; Qian et al. [40] achieve temporal consistency by employing student-teacher distillation to establish semantic and instance correspondence over time; and Traub et al. [48] use a recurrent network with a constancy prior [17]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.644, + 0.605, + 0.659 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.908, + 0.76 + ], + "angle": 0, + "content": "Our approach builds upon the existing input reconstruction-based video object-centric framework [28, 64] by introducing a consistency loss that contrasts the slots across consecutive frames and thereby adapting the model to discover consistent representations. See Fig. 2 for an overview of the SLOT CONTRAST architecture." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.879, + 0.787 + ], + "angle": 0, + "content": "3.1. Semantic Recurrent Slot Attention Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Our model is an encoder-decoder object-centric architecture based on Slot Attention module (SA) [13] with additional adaptations for sequential inputs similar to SAVi [28], while leveraging pre-trained semantic features as proposed by DINOSAUR [44]. The model consists of three main components: a pre-trained self-supervised dense feature encoder (e.g., DINOv2 [38]), a Recurrent Slot Attention module that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5402" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.088, + 0.825, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.908, + 0.31 + ], + "angle": 0, + "content": "Figure 2. SLOT CONTRAST model architecture overview. For each frame, we extract patch features \\( h_t \\) using DINOv2 ViT. These features are then used to update the previously initialized or predicted slots, resulting in new slots \\( S_t \\). The model is trained by contrasting the current frame's slots \\( S_t \\) with the slots from the previous frame \\( S_{t-1} \\), and by reconstructing the patch features \\( h_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.425 + ], + "angle": 0, + "content": "groups the encoder features into slots and models temporal slot updates, and a decoder that maps slots from each frame to reconstructions of the dense self-supervised features used as inputs. Next, we describe those components in more detail while explaining how to adapt them to the task of consistent object-centric representation learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.485, + 0.471 + ], + "angle": 0, + "content": "Given a video frame \\( x_{t} \\), \\( t \\in \\{1,2,\\dots,T\\} \\) and a pretrained, frozen self-supervised DINO model \\( f \\) we first extract \\( N \\) patch features \\( g_{t} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.48, + 0.484, + 0.499 + ], + "angle": 0, + "content": "\\[\ng _ {t} = f \\left(x _ {t}\\right), \\quad g _ {t} \\in \\mathbb {R} ^ {N \\times D}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.571 + ], + "angle": 0, + "content": "As those frozen features are mostly semantic and are trained only on images, we further adapt them to the task of temporally consistent object discovery. Specifically, each feature vector \\( g_{t} \\) is passed through a MLP \\( g_{\\psi} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.241, + 0.58, + 0.483, + 0.597 + ], + "angle": 0, + "content": "\\[\nh _ {t} = g _ {\\psi} \\left(g _ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.773 + ], + "angle": 0, + "content": "to adapt the frozen dense features for object-centric grouping (see App. I for more details and visualizations). Based on the transformed encoder features \\( h_t \\) and a set of slot representations of the previous timestep \\( S_{t-1}^p \\), with \\( K \\) slots \\( s_{t-1}^{k,p} \\in S_{t-1}^p \\), we use a recurrent grouping module to extract slot representations. The Recurrent Slot Attention module comprises a grouping module \\( C_\\theta \\) and a predictor module \\( P_\\omega \\). The former updates slot representations using the standard Slot Attention module [32] on visual features \\( h_t \\) from the encoder, while the latter captures temporal and spatial interactions between slots:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.784, + 0.484, + 0.801 + ], + "angle": 0, + "content": "\\[\nS _ {t} ^ {c} = \\mathrm {C} _ {\\theta} \\left(h _ {t}, S _ {t - 1} ^ {p}\\right), \\quad S _ {t} ^ {p} = \\mathrm {P} _ {\\omega} \\left(S _ {t} ^ {c}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Both slot-level representations, generated either by the grouping module \\( S_{t}^{c} \\) or the predictor \\( S_{t}^{p} \\), can be utilized for subsequent decoding or downstream task processing. In our implementation, the slot-level representations from the grouping module \\( S_{t}^{c} \\) are employed for the decoding stage. From now on, we will refer to \\( S_{t}^{c} \\) as \\( S_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.908, + 0.548 + ], + "angle": 0, + "content": "Temporal Slot Attention Initialization Importantly, we found that our setup benefits considerably from a learned initialization \\( S_0 \\), which can influence the efficiency of training across various objectives. Originally, Locatello et al. [32] proposed a randomly sampled query initialization, where slots are sampled from the same Gaussian distribution with learned mean and variance. While such initialization allows different numbers of slots during inference, sampling from the same Gaussian distribution does not create a particularly favorable structure in slot-space. In this work, we use a straightforward learned initialization [19, 43] where a fixed set of initial slot vectors \\( S_0 \\) is learned for the entire dataset. Such initialization allows for learning dissimilar initialization queries that consistently attend to different objects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.554, + 0.908, + 0.585 + ], + "angle": 0, + "content": "Finally, for the reconstruction loss objective, we decode reconstructions \\(\\hat{g}_t\\) from all slots using the MLP decoder [44]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.615, + 0.894, + 0.632 + ], + "angle": 0, + "content": "3.2. Temporal Consistency through Slot Contrast" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Contrastive learning is flexible in supporting diverse data sources and loss function designs. By carefully defining positive and negative examples, we can craft robust loss objectives that effectively guide self-supervised representation learning [6]. For instance, video contrastive methods like CVRL [39] leverage augmented video chunks to define positive (from the same video) and negative (from different videos) examples. In object-centric learning Didolkar et al. [8] employed a contrastive loss function to gain controllability over slot representations guided by language. We propose a novel application of a contrastive loss for temporal consistency in object-centric slot representations. In particular, we define positive samples as the representations of the same slot from two consecutive time steps within a video, while negative samples comprise all other slots across the batch between these time steps. An overview of the proposed loss is presented in Fig. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5403" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.861, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.335, + 0.908, + 0.419 + ], + "angle": 0, + "content": "Figure 3. Overview of the losses used in SLOT CONTRAST. (a) Our proposed temporal consistency objective, slot-slot contrastive loss, operates on a batch of video sequences by enforcing temporal alignment across object slots. For each frame in the sequence, the model groups object features into specific slot representations \\( S_{t}^{i} \\). The slot-slot contrastive loss then enforces temporal consistency by drawing the corresponding slot representations from adjacent frames closer, while simultaneously pushing apart all other slot representations in the batch—whether they come from different objects within the same video or from objects in other videos. (b) The feature reconstruction loss ensures informativeness of the learned slots by using them to reconstruct original DINOv2 features with an MLP decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Intra-Video Slot-Slot Contrautive Loss To force each slot to be consistent in time, we aim to learn slots that are similar in time while being maximally dissimilar to other slots. Given the sets of slot representations \\( S_{t - 1} \\) and \\( S_{t} \\) at time steps \\( t - 1 \\) and \\( t \\), we want elements \\( s_{t - 1}^{i}\\in S_{t - 1} \\) to be close to the next-frame slots \\( s_t^i \\) corresponding to the same object, while having maximal distance to the next-frame slots \\( s_t^k,k\\neq i \\) corresponding to other objects in the video. The corresponding InfoNCE contrastive loss [37] is defined as \\( \\mathcal{L}_{\\mathrm{intra}} = \\frac{1}{K}\\sum_{i = 1}^{K}\\ell_{i}^{\\mathrm{intra}} \\) with" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.608, + 0.484, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\ell_ {i} ^ {\\text {i n t r a}} = - \\log \\frac {\\exp \\left(\\sin \\left(s _ {t - 1} ^ {i} , s _ {t} ^ {i}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\mathbb {1} _ {[ k \\neq i ]} \\exp \\left(\\sin \\left(s _ {t - 1} ^ {i} , s _ {t} ^ {k}\\right) / \\tau\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.715 + ], + "angle": 0, + "content": "where \\(K = |S_t|\\) is a number of slots per frame, \\(\\mathrm{sim}(u,v) = \\frac{u^{\\top}v}{\\|u\\|_2\\|v\\|_2}\\) is the cosine similarity, \\(\\mathbb{1}_{[.]}\\) is an indicator excluding the self-similarity of the slot \\(s_i\\) from the denominator, and \\(\\tau >0\\) is a temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.484, + 0.82 + ], + "angle": 0, + "content": "While being a desirable property, intra-video slot contrast can be achieved simply by amplifying the differences between slots in the SA module's first frame initialization \\( S_0 \\). To encourage a stronger focus on video content and instance specificity of the representations, we propose a further improvement over this loss by extending the negative contrast set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Batch Video Slot-Slot Contrastive Loss To leverage the benefits of larger contrast sets and prevent degenerate solutions relying solely on the initialization of slots, we exploit the fact that the whole batch of videos can be con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.908, + 0.538 + ], + "angle": 0, + "content": "sidered a large set of primarily unique object representations. Consequently, we enhance contrast within a video and between videos by including negative slots from the current and subsequent frames of all videos in the batch. Correspondingly, we define our slot-slot contrastive loss as \\(\\mathcal{L}_{\\mathrm{ssc}} = \\frac{1}{B\\cdot K}\\sum_{j = 1}^{B}\\sum_{i = 1}^{K}\\ell_{i,j}^{\\mathrm{ssc}}\\) and" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.548, + 0.907, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\ell_ {i, j} ^ {\\mathrm {s s c}} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(s _ {t - 1} ^ {i , j} , s _ {t} ^ {i , j}\\right) / \\tau\\right)}{\\sum_ {b = 1} ^ {B} \\sum_ {k = 1} ^ {K} \\mathbb {1} _ {[ k , b \\neq i , j ]} \\exp \\left(\\operatorname {s i m} \\left(s _ {t - 1} ^ {i , j} , s _ {t} ^ {k , b}\\right) / \\tau\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.905, + 0.655 + ], + "angle": 0, + "content": "where \\( B \\) is a number of videos in the batch and \\( s_t^{i,j} \\) denotes the \\( i \\)-th slot of the \\( j \\)-th video at time \\( t \\). For more details on slot-slot contrastive loss implementation, see App. C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.908, + 0.747 + ], + "angle": 0, + "content": "We find that this approach significantly enhances the effectiveness of the slot-slot contrastive loss. Furthermore, since all videos in the batch are processed with the same initial state \\( S_0 \\), this loss function avoids suboptimal solutions that rely solely on the uniqueness of the initialization, instead encouraging object discovery as the basis for contrast." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.858 + ], + "angle": 0, + "content": "Final Loss To encourage scene decomposition we use a feature reconstruction loss, similar to DINOSAUR [44] and VideoSAUR [64]. Our final loss function combines the reconstruction loss with our proposed contrastive loss \\(\\mathcal{L}_{\\mathrm{ssc}}\\) weighted by the hyperparameter \\(\\alpha\\) (see Table S1 for details on how the hyperparameters are set):" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.867, + 0.908, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {t = 1} ^ {T - 1} \\mathcal {L} _ {\\mathrm {r e c}} \\left(\\mathrm {h} _ {\\mathrm {t}}, \\hat {\\mathrm {h}} _ {\\mathrm {t}}\\right) + \\alpha \\mathcal {L} _ {\\mathrm {s s c}} \\left(S _ {t - 1}, S _ {t}\\right). \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5404" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.173 + ], + "angle": 0, + "content": "Table 1. Consistent object-discovery performance of SLOT CONTRAST in comparison with SAVi, STEVE, VideoSAUR on MOVi-C, MOVi-E, and YouTube-VIS datasets. VideoSAURv2 is an improved version of the VideoSAUR trained on DINOv2 features. Both metrics are computed for the whole video (24 frames for MOVi, up to 76 frames for YouTube-VIS)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.178, + 0.485, + 0.273 + ], + "angle": 0, + "content": "
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
SAVi [28]22.213.642.816.0--
STEVE [45]36.126.550.626.61519.1
VideoSAUR [64]64.838.973.935.628.926.3
VideoSAURv2--77.134.431.229.7
SLOT CONTRAST69.332.782.929.238.033.7
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.291, + 0.226, + 0.308 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.317, + 0.486, + 0.468 + ], + "angle": 0, + "content": "We evaluate our method's temporal consistency on two downstream tasks: object discovery and latent object dynamics prediction. Our experiments address three main questions: (1) How does our model compare to state-of-the-art methods in both temporal consistency and scene decomposition? (2) How effective are our model's learned representations for the challenging downstream task of object dynamics prediction and for object tracking under full occlusions? (3) How important are the different components of our model and loss function for temporal consistency?" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.488, + 0.485, + 0.655 + ], + "angle": 0, + "content": "Datasets To evaluate our method in the controlled setting, we use MOVi-C and MOVi-E synthetic datasets generated by Kubric [16]. MOVi-C includes richly textured everyday objects, featuring up to 11 objects per scene, while MOVi-E expands this to 23 objects and introduces basic linear camera motion. In addition, to study the scalability of our method to real-world data, we evaluate our method on the real-world YouTube-VIS 2021 (YTVIS21) video dataset [60]. YTVIS21 is an unconstrained, real-world dataset sourced from YouTube, capturing a diverse range of scenes (for more details, see App. E)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.486, + 0.87 + ], + "angle": 0, + "content": "Metrics Similar to other object-centric video methods [11, 28, 45, 64], to evaluate consistent object discovery, we use the video foreground adjusted rand index (FG-ARI) [14], measuring how well objects are split. In addition, we evaluate the sharpness of masks using the video intersection over union with mean best overlap matching (mBO) metric [44, 64]. Both metrics are computed over the full video and thus reflect how consistent object discovery is. In addition, to investigate the effects of the temporal consistency inductive bias on the per-frame object discovery itself, we use per-frame FG-ARI (image FG-ARI), which we independently compute for each frame and average afterwards. More details can be found in App. F." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Finally, when evaluating how well object-centric representations perform for object dynamics prediction (see Sec. 4.2)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.168 + ], + "angle": 0, + "content": "we employ the same evaluation metrics as in the object discovery task: FG-ARI and mBO. This time, however, these metrics are computed by comparing the predicted masks (obtained by decoding the predicted slots [56]) with the ground-truth future masks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.175, + 0.685, + 0.192 + ], + "angle": 0, + "content": "4.1. Object Discovery" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.909, + 0.321 + ], + "angle": 0, + "content": "Implementation Details We employ the DINOv2 model as our feature encoder, using ViT-S/14 for the MOVi-C dataset and ViT-B/14 for MOVi-E and YTVIS21. The slot dimension is set to 128 for MOVi-E and 64 for both MOVi-C and YTVIS21. For the MOVi datasets, we use a resolution of (336, 336), generating \\(24 \\times 24\\) patches yielding 576 ViT tokens, while for YTVIS21, a resolution of (518, 518) yields 1369 tokens. Full details are provided in App. A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.909, + 0.519 + ], + "angle": 0, + "content": "Baselines We compare SLOT CONTRAST against the previously proposed SAVi [28] and STEVE [45] that employ an image reconstruction objective and with the state-of-the-art method VideoSAUR [64] that uses self-supervised feature reconstructions. Additionally, for a fair comparison, we trained a modification of VideoSAUR with DINOv2 features (referred to as VideoSAURv2). The implementation details are provided in App. G. In addition, to assess how closely SLOT CONTRAST approaches supervised methods, we compared it with SAM2 [41] as a supervised zero-shot baseline for temporal consistency and to weakly-supervised by depth SAVi++ [11] method. The results are in App. G." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Temporally Consistent Object Discovery (Table 1 & Figure 4) SLOT CONTRAST significantly outperforms both SAVi and STEVE by a wide margin. When compared to VideoSAUR using its default parameters, our approach demonstrates higher consistency in terms of video FG-ARI scores. Compared to VideoSAUR and VideoSAURv2, SLOT CONTRAST achieves superior video scene decomposition (measured by FG-ARI scores). However, on synthetic datasets SLOT CONTRAST's masks are less sharp (as reflected by mBO). Notably, on the most challenging real-world YouTube-VIS data, our method surpasses both versions of VideoSAUR, achieving better performance on FG-ARI (+6.8) and mBO (+4). This shows that, given a large enough resolution and natural data inputs well aligned with DINOv2, SLOT CONTRAST can decompose unconstrained videos into consistent object representations. More examples are illustrated in App. M." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Per-Frame Scene Decomposition (Table 2) Previous research has shown the effectiveness of specific inductive biases and training objectives for unsupervised object discovery, such as reconstructing in semantic space or leveraging motion cues in video data. Building on this, we demonstrate that the contrastive nature of our temporal consistency loss" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5405" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.506, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.905, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with VideoSAURv2 on YouTube-VIS dataset. In challenging situations (e.g., almost full occlusions at \\( t = 24 \\) of the 1st video and \\( t = 14 \\) of the 2nd video), VideoSAURv2 reassigns slots to different objects (pink arrows), whereas SLOT CONTRAST consistently assigns slots to the same object (green arrows). Note that the colors of the masks are matched manually for better visual comparison." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.339, + 0.486, + 0.395 + ], + "angle": 0, + "content": "Table 2. Quantitative Results on MOVi-E in terms of per-frame Image FG-ARI. The methods are grouped by the target data they train on: only images \\((\\mathcal{I})\\), videos with motion segmentation annotations \\((\\mathcal{V} + \\mathcal{M})\\), and only videos \\((\\mathcal{V})\\)." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.4, + 0.466, + 0.596 + ], + "angle": 0, + "content": "
ModelObjectiveImage FG-ARI ↑
↓LSD [21]Image Rec.53.4
SlotDiffusion [57]Image Rec.60.0
DINOSAUR [44]Image Rec.65.1
+MoToK [3]+Mot. Seg.66.7
Safadoust et al. [42]+GT Flow78.3
DIOD [25]+Mot. Seg.82.2
↑STEVE [45]Video Rec.54.1
VideoSAUR [64]Temp. Sim.78.4
SOLV [2]Mid. Fr. Pred.80.8
SLOT CONTRASTSlot Contrast84.8
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.716 + ], + "angle": 0, + "content": "function yields improved scene decomposition as a byproduct. This occurs because our loss function encourages the model to learn consistent feature representations for objects across frames, leading to an adaptive process where dense features become more contrastive, thereby enhancing object discovery in individual frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We compare our method with prior approaches in terms of per-frame object discovery, using the image FG-ARI metric for evaluation. Specifically, we compare three categories of methods: image-based, video-based, and methods that use videos with additional motion cues. Image-based methods use only images as a target (feature reconstruction based DINOSAUR [44] and diffusion-based LSD [21] and SlotDiffusion [57] methods). Video-based methods use only videos as targets: STEVE [45] reconstructs current frame features, SOLV [2] predicts middle frame features, and VideoSAUR [64] predicts temporal feature-similarities. Finally, we also compare with weakly-supervised methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.37 + ], + "angle": 0, + "content": "using motion masks [3, 25] or ground truth (GT) optical flow [42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.906, + 0.537 + ], + "angle": 0, + "content": "The results on MOVi-E dataset are presented in Table 2, with comparisons across additional datasets provided in App. H. Using temporal signals from the video using feature reconstruction is better than object discovery based on images. Next, additional objectives that exploit the temporal structure of the videos allow even better scene decomposition. Notably, our method, which combines a feature reconstruction objective with a simple contrastive objective, leads to state-of-the-art performance reaching 84.8 per-frame FG-ARI, outperforming methods [3, 25] that use motion segmentation masks for object discovery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.659 + ], + "angle": 0, + "content": "Robustness to Full Occlusions To evaluate our method's robustness in handling complete object occlusions—a challenging scenario for maintaining consistency—we conduct experiments using a targeted subset of the MOVi-C dataset that contains sequences where objects are fully occluded. For evaluation, we retain only the ground-truth masks for the objects that experience occlusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.751 + ], + "angle": 0, + "content": "We find that the feature reconstruction baseline achieves only \\(16\\%\\) mBO vs. our method obtains \\(21\\%\\) mBO on fully occluded objects. Our results suggest that SLOT CONTRAST significantly enhances consistency during object disappearances and reappearance. We refer the reader to Fig. 5 and App. K for visual examples and more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.767, + 0.774 + ], + "angle": 0, + "content": "4.2. Object Dynamics Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Setup To evaluate performance on the task of predicting object dynamics, we train a dynamics module using the object-centric representations inferred by a pretrained object-centric model. For this dynamics module, we select SlotFormer [56], which predicts the slots autoregressively for \\(K\\) rollout steps based on the slots inferred from \\(T\\) burn-in frames preceding the prediction horizon. In our setup, we use 14 burn-in frames and 10 rollout steps. Since SlotFormer" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5406" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.483, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.332, + 0.485, + 0.388 + ], + "angle": 0, + "content": "Figure 5. Comparison of the Feature Reconstruction (Feat. Rec.) baseline, the slot-slot contrastive loss using only slots from the same video as the contrast set (Intra-video Contrast), and SLOT CONTRAST on the MOVi-C dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.522 + ], + "angle": 0, + "content": "is trained independently from the object-centric model, we first train the latter, subsequently extending the datasets with the inferred slots for each frame. This approach avoids the computational complexity of training SlotFormer by removing the necessity to encode frames into the slot space at each training step. A brief introduction to SlotFormer and the implementation details can be found in App. J." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.631 + ], + "angle": 0, + "content": "Baselines We compare SlotFormer [56] trained on object-centric representations derived from a model trained using only feature reconstruction loss with SlotFormer trained on representations from SLOT CONTRAST. Both models perform reconstruction in feature space rather than pixel space, so we use only the slot reconstruction loss for training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.74 + ], + "angle": 0, + "content": "Quality of Predicted Masks (Table 3 and Figure 6) We note that our model has a significantly better FG-ARI on MOVi-C, while mBO is comparable to that of the baseline. On MOVi-E, the performance of our model is comparable to that of the baseline, which highlights the difficulty of adapting to videos with camera motion. There is also a" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.754, + 0.481, + 0.863 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.869, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Figure 6. Object dynamics prediction task on MOVi-C using SLOT CONTRAST slots using SlotFormer [56]." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 3. Downstream task of predicting object dynamics. Comparison of predictions made by SlotFormer based on representations obtained from SLOT CONTRAST and from Feature Reconstruction." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.201 + ], + "angle": 0, + "content": "
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
Feat. Rec. + SF50.725.970.624.327.428.9
Ours + SF63.826.170.524.929.229.6
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.907, + 0.293 + ], + "angle": 0, + "content": "slight improvement in FG-ARI on YTVIS21, while mBO remains comparable. It is worth noting that predicting the motion in this dataset is especially challenging, given the large diversity of possible scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.31, + 0.616, + 0.326 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.41 + ], + "angle": 0, + "content": "In this section, we investigate key components of our approach, including the impact of the contrastive loss and the type of slot initialization. In addition, we study how effective SLOT CONTRAST is in automatically shutting down slots in correspondence to the scene's complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.685 + ], + "angle": 0, + "content": "Ablation of Loss Components (Table 4 and Figure 5) To demonstrate the value of the proposed slot-slot contrastive loss, we carry out an ablation study, comparing it with the feature reconstruction loss [44] and the intra-video contrastive loss, which contrasts slot representations in a single video. Using the intra-video contrastive loss yields improvements over the feature reconstruction baseline (+5.1 FG-ARI and +1.5 mBO on MOVi-C). However, we observe that in more challenging situations, the intra-video contrastive loss leads to failure cases such as shutting down too many slots (see Fig. 5). Next, we observe that by extending the contrast to the full batch of videos, SLOT CONTRAST learns more consistent representations (+19.6 FG-ARI and +5.3 mBO). This change increases the difficulty of the learning task, which prevents the model from relying on superficial patterns like slot initializations or object positions." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.712, + 0.885, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.869, + 0.907, + 0.897 + ], + "angle": 0, + "content": "Figure 7. Distribution of ground truth and predicted object numbers (i.e., number of active slots) on the MOVi-C dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5407" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.089, + 0.842, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation of loss components used by SLOT CONTRAST on MOVi-C, MOVi-E, and YouTube-VIS Datasets." + }, + { + "type": "table", + "bbox": [ + 0.145, + 0.109, + 0.855, + 0.201 + ], + "angle": 0, + "content": "
Feat. Rec.Intra-video ContrastSlot-Slot ContrastMOVi-CMOVi-EYouTube-VIS
LrecLintraLsscFG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
✓49.727.479.828.435.331.4
✓✓54.828.978.729.135.733.6
✓✓69.332.782.929.238.033.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.212, + 0.483, + 0.242 + ], + "angle": 0, + "content": "Table 5. Comparison of the random initialization (RI) and learned initialization (LI) techniques." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.245, + 0.481, + 0.328 + ], + "angle": 0, + "content": "
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
Feat. Rec. (RI)45.327.271.128.335.230.2
Feat. Rec. (LI)49.427.879.828.435.331.4
SLOT CONTRAST (RI)62.932.475.328.436.130.8
SLOT CONTRAST (LI)69.332.782.929.238.033.7
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.353, + 0.483, + 0.581 + ], + "angle": 0, + "content": "Choice of the First Frame Initialization (Table 5) In video object-centric learning, slots are typically initialized based on those from the previous time step [28], while the first frame is initialized from learnable parameters. Previous real-world object-centric methods mostly used random initialization samples from Gaussian distribution [2, 64], in this work, we study the impact of the type of initialization under our contrastive objective. The findings are outlined in Table 5. Our experiments indicate that when combined with slot-slot contrastive loss, learned initialization significantly outperforms random initialization. We hypothesize that this improvement stems from the ability of learned initializations to shape the initial state in a way that enhances contrastiveness, a benefit not achievable with random initialization. For similarity visualizations and more details, see App. B." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.599, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Number of Active Slots per Video (Figure 7) In models based on the Slot Attention mechanism, all available slots are typically utilized [44], leading to a mismatch between the predicted number of components in scene decomposition and the ground-truth number of objects in the scene. This can cause the random splitting of the objects between slots and non-consistent scene representations when slots are reassigned from one object to a part of another object. To address this challenge, it is important to study whether redundant slots can be effectively deactivated. Recently, AdaSlot [12] introduced a discrete slot sampling module, coupled with a complexity-aware prior, to penalize redundant slots explicitly. Similarly, SOLV [2] used agglomerative clustering to merge redundant slots. In this work, we investigate whether SLOT CONTRAST is capable of accurately determining the number of objects in a scene without relying on explicit priors to minimize the number of active slots." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We compare ground truth and predicted object density on MOVI-C dataset, as shown in Fig. 7. While the feature reconstruction model yields predictions within a narrow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.907, + 0.337 + ], + "angle": 0, + "content": "range—creating a sharp peak near a predefined number of slots—our model, similarly to AdaSlot [12], achieves a smoother prediction distribution that aligns more closely with the ground truth (note that the consistent shift is because 2–3 slots are used for the background, while the ground truth density is computed only for foreground objects). Interestingly, SLOT CONTRAST achieves this without requiring an explicit prior toward sparsity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.355, + 0.634, + 0.372 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.909, + 0.67 + ], + "angle": 0, + "content": "SLOT CONTRAST advances unsupervised video object-centric learning by significantly improving the temporal consistency of object representations. Our method explicitly incentivizes temporal consistency by adding a self-supervised contrastive loss. We showed that this loss is not only beneficial for consistency, but also enhances object discovery: SLOT CONTRAST achieves state-of-the-art results on challenging synthetic datasets with many objects and the unconstrained real-world YouTube-VIS dataset. Furthermore, consistent representations directly support temporal downstream tasks such as unsupervised object dynamics prediction and allow for tracking of objects through full occlusions. Finally, SLOT CONTRAST effectively shuts down non-unique slots, leading to a sparser representation that captures the true object distribution more faithfully. Taken together, we expect these improvements to pave the way for broader adoption of video object-centric representations, for instance in applications like word modeling, autonomous control, or video question answering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Limitations of our work include the fixed number of slots during initialization. Additionally, we cannot directly control the segmentation granularity of entities. Further limitations and failure cases are discussed in App. L." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Future work could explore several promising directions. First, one could use SLOT CONTRAST's robust and consistent representations for learning compositional world models from real-world robotics data to enable object-centric planning and control. Second, investigating the compatibility of our contrastive loss with other object-centric learning approaches with different inductive biases, such as SlotDiffusion [57]. Finally, improving the compactness of object masks [24] to achieve more precise object segmentation masks could also benefit unsupervised class-agnostic video object segmentation applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5408" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.283 + ], + "angle": 0, + "content": "We thank Christian Gumbsch for useful suggestions and Ke Fan for providing additional details and results for AdaSlot method. Andrii Zadaianchuk is funded by the European Union (ERC, EVA, 950086). The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Maximilian Seitzer. Georg Martius is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. This work was supported by the ERC - 101045454 REAL-RL. We acknowledge the support from the German Federal Ministry of Education and Research (BMBF) through the Tübingen AI Center (FKZ: 01IS18039B)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.309, + 0.188, + 0.325 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.335, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In International Conference on Computer Vision (ICCV), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[2] Görkay Aydemir, Weidi Xie, and Fatma Güney. Self-supervised object-centric learning for videos. In NeurIPS, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[3] Zhipeng Bao, Pavel Tokmakov, Yu-Xiong Wang, Adrien Gaidon, and Martial Hebert. Object discovery from motion-guided tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22972-22981, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. MONet: Unsupervised Scene Decomposition and Representation. arXiv:1901.11390, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[7] Aniket Didolkar, Andrii Zadaianchuk, Anirudh Goyal, Mike Mozer, Yoshua Bengio, Georg Martius, and Maximilian Seitzer. Zero-shot object-centric representation learning. arXiv preprint arXiv:2408.09162, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[8] Aniket Rajiv Didolkar, Andrii Zadaianchuk, Rabiul Awal, Maximilian Seitzer, Efstratios Gavves, and Aishwarya Agrawal. Ctrl-o: Language-controllable object-centric visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Andrea Dittadi, Samuele Papa, Michele De Vita, Bernhard Schölkopf, Ole Winther, and Francesco Locatello. Generalization and Robustness Implications in Object-Centric Learning. In ICML, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.335, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.908, + 0.161 + ], + "angle": 0, + "content": "Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. PaLM-E: An embodied multimodal language model. In ICML, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[11] Gamaleldin Fathy Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael Curtis Mozer, and Thomas Kipf. \\(\\mathsf{SAV}\\mathsf{i} + +\\) : Towards End-to-End Object-Centric Learning from Real-World Videos. In NeurIPS, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[12] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23062-23071, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[13] Locatello Francesco, Weissenborn Dirk, Unterthiner Thomas, Mahendran Aravindh, Heigold Georg, Uszkoreit Jakob, Dosovitskiy Alexey, and Kipf Thomas. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.907, + 0.426 + ], + "angle": 0, + "content": "[14] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-Object Representation Learning with Iterative Variational Inference. In ICML, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.907, + 0.468 + ], + "angle": 0, + "content": "[15] Klaus Greff, Sjoerd Van Steenkiste, and Jürgen Schmidhuber. On the Binding Problem in Artificial Neural Networks. arXiv:2012.05208, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.908, + 0.621 + ], + "angle": 0, + "content": "[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A Scalable Dataset Generator. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.622, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[17] Christian Gumbsch, Martin V Butz, and Georg Martius. Sparsely changing latent states for prediction and planning in partially observable domains. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.664, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[18] Dan Haramati, Tal Daniel, and Aviv Tamar. Entity-centric reinforcement learning for object manipulation from pixels. In ICLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.706, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[19] Baoxiong Jia, Yu Liu, and Siyuan Huang. Improving object-centric learning with query optimization. arXiv preprint arXiv:2210.08990, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[20] Jindong Jiang, Sepehr Janghorbani, Gerard de Melo, and Sungjin Ahn. SCALOR: Generative World Models with Scalable Object Representations. In ICLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.907, + 0.817 + ], + "angle": 0, + "content": "[21] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In NeurIPS, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.908, + 0.899 + ], + "angle": 0, + "content": "[22] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.335, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5409" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.161 + ], + "angle": 0, + "content": "[23] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthew, Antonia Creswell, Matthew Botvinick, Alexander Lerchner, and Christopher P. Burgess. SIMONE: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[24] Ioannis Kakogeorgiou, Spyros Gidaris, Konstantinos Karantzalos, and Nikos Komodakis. Spot: Self-training with patch-order permutation for object-centric learning with autoregressive transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22776-22786, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[25] Sandra Kara, Hejer Ammar, Julien Denize, Florian Chabot, and Quoc-Cuong Pham. Diod: Self-distillation meets object discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3975-3985, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[26] Laurynas Karazija, Iro Laina, and Christian Rupprecht. ClevrTex: A Texture-Rich Benchmark for Unsupervised Multi-Object Segmentation. In NeurIPS Track on Datasets and Benchmarks, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[27] Thomas Kipf, Elise Van der Pol, and Max Welling. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.484, + 0.48 + ], + "angle": 0, + "content": "[28] Thomas Kipf, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-centric Learning from Video. In ICLR, 2022. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.482, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[29] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[30] Adam Kosiorek, Hyunjik Kim, Yee Whye Teh, and Ingmar Posner. Sequential Attend, Infer, Repeat: Generative Modelling of Moving Objects. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[31] Jian Li, Pu Ren, Yang Liu, and Hao Sun. Reasoning-enhanced object-centric learning for videos. arXiv preprint arXiv:2403.15245, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[32] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-Centric Learning with Slot Attention. In NeurIPS, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[33] Zijia Lu, Bing Shuai, Yanbei Chen, Zhenlin Xu, and Davide Modolo. Self-supervised multi-object tracking with path consistency. arXiv preprint, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[34] Amir Mohammad Karimi Mamaghan, Samuele Papa, Karl Henrik Johansson, Stefan Bauer, and Andrea Dittadi. Exploring the effectiveness of object-centric representations in visual question answering: Comparative insights with foundation models. arXiv preprint arXiv:2407.15589, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[35] Sha Meng, Dian Shao, Jiacheng Guo, and Shan Gao. Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16264-16273, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[36] Cristian Meo, Akihiro Nakano, Mircea Lică, Aniket Didolkar, Masahiro Suzuki, Anirudh Goyal, Mengmi Zhang, Justin" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Dauwels, Yutaka Matsuo, and Yoshua Bengio. Object-centric temporal consistency via conditional autoregressive inductive biases. arXiv preprint arXiv:2410.15728, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.907, + 0.3 + ], + "angle": 0, + "content": "[38] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[39] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6964-6974, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.907, + 0.412 + ], + "angle": 0, + "content": "[40] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.906, + 0.509 + ], + "angle": 0, + "content": "[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.51, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[42] Sadra Safadoust and Fatma Güney. Multi-object discovery by low-dimensional object motion. In ICCV, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.538, + 0.907, + 0.607 + ], + "angle": 0, + "content": "[43] Mehdi SM Sajjadi, Daniel Duckworth, Aravindh Mahendran, Sjoerd Van Steenkiste, Filip Pavetic, Mario Lucic, Leonidas J Guibas, Klaus Greff, and Thomas Kipf. Object scene representation transformer. Advances in Neural Information Processing Systems, 35:9512-9524, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.907, + 0.677 + ], + "angle": 0, + "content": "[44] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. In ICLR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.907, + 0.718 + ], + "angle": 0, + "content": "[45] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple Unsupervised Object-Centric Learning for Complex and Naturalistic Videos. In NeurIPS, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.907, + 0.788 + ], + "angle": 0, + "content": "[46] Matthias Tangemann, Steffen Schneider, Julius Von Kügelgen, Francesco Locatello, Peter Gehler, Thomas Brox, Matthias Kümmerer, Matthias Bethge, and Bernhard Schölkopf. Unsupervised object learning via common fate. arXiv preprint arXiv:2110.06562, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[47] Pavel Tokmakov, Allan Jabri, Jie Li, and Adrien Gaidon. Object permanence emerges in a random walk along memory. arXiv preprint arXiv:2204.01784, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[48] Manuel Traub, Frederic Becker, Sebastian Otte, and Martin V Butz. Looping loci: Developing object permanence from videos. arXiv preprint arXiv:2310.10372, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[49] Manuel Traub, Sebastian Otte, Tobias Menge, Matthias Karlbauer, Jannik Thuemmel, and Martin V. Butz. Learning What" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "5410" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "and Where: Disentangling Location and Identity Tracking Without Supervision. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.486, + 0.19 + ], + "angle": 0, + "content": "[50] Basile Van Hoorick, Pavel Tokmakov, Simon Stent, Jie Li, and Carl Vondrick. Tracking through containers and occluders in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13802-13812, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.485, + 0.247 + ], + "angle": 0, + "content": "[51] Sjoerd van Steenkiste, Michael Chang, Klaus Greff, and Jürgen Schmidhuber. Relational neural expectation maximization: Unsupervised discovery of objects and their interactions. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[52] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity Abstraction in Visual Model-based Reinforcement Learning. In Conference on Robot Learning, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[53] Ning Wang, Yibing Song, Chao Ma, Wengang Zhou, Wei Liu, and Houqiang Li. Unsupervised deep tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1308-1317, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[54] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2566-2576, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[55] Nicholas Watters, Loic Matthew, Christopher P Burgess, and Alexander Lerchner. Spatial broadcast decoder: A simple architecture for learning disentangled representations in vaes. arXiv preprint arXiv:1901.07017, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[56] Ziyi Wu, Nikita Dvornik, Klaus Greff, Thomas Kipf, and Animesh Garg. Slotformer: Unsupervised visual dynamics simulation with object-centric models. In ICLR, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[57] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotdiffusion: Object-centric generative modeling with diffusion models. In NeurIPS, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[58] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Slot-VLM: Object-event slots for video-language modeling. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[59] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[60] Linjie Yang, Yuchen Fan, Yang Fu, and Ning Xu. The 3rd large-scale video object segmentation challenge - video instance segmentation track, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[61] Jaesik Yoon, Yi-Fu Wu, Heechul Bae, and Sungjin Ahn. An investigation into pre-training object-centric representations for reinforcement learning. In ICML, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[62] Haonan Yu and Wei Xu. Vonet: Unsupervised video object learning with parallel u-net attention and object-wise sequentialvae.arXiv preprint arXiv:2401.11110, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[63] Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius. Self-supervised Visual Reinforcement Learning with Object-centric Representations. In ICLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[64] Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius. Object-centric learning for real-world videos by predicting temporal feature similarities. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023), 2023. 1, 2, 4, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[65] Zixu Zhao, Jiaze Wang, Max Horn, Yizhuo Ding, Tong He, Zechen Bai, Dominik Zietlow, Carl-Johann Simon-Gabriel, Bing Shuai, Zhuowen Tu, Thomas Brox, Bernt Schiele, Yanwei Fu, Francesco Locatello, Zheng Zhang, and Tianjun Xiao. Object-centric multiple object tracking. ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[66] Daniel Zoran, Rishabh Kabra, Alexander Lerchner, and Danilo J. Rezende. PARTS: Unsupervised segmentation with slots, attention and independence maximization. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "5411" + } + ] +] \ No newline at end of file diff --git a/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_origin.pdf b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f02c79a6952d2b414b7affdbbe83a5baef69dee3 --- /dev/null +++ b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/f1085bc3-3cb8-4ee6-8cb0-6147e12071f2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3595f85139b19a5d38ffa8696f584e77ff87c0bbca1e2daee9d533d57ee154ad +size 10019594 diff --git a/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/full.md b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c0deb971ab204abea8929500bcec055f67688284 --- /dev/null +++ b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/full.md @@ -0,0 +1,301 @@ +# Temporally Consistent Object-Centric Learning by Contrasting Slots + +Anna Manasyan1 Maximilian Seitzer1 Filip Radovic1 Georg Martius1,3 Andrii Zadaianchuk2 +1University of Tübingen, Tübingen, Germany +2University of Amsterdam, Amsterdam, Netherlands +3Max Planck Institute for Intelligent Systems, Tübingen, Germany + +anna.manasyan@student.uni-tuebingen.de a.zadaianchuk@uva.nl + +![](images/0e4ee66c9e4ae74713e74fd1a72338564bdc5e1654deaedb6eb6f41340bb7b3d.jpg) +Figure 1. SLOT CONTRAST: Our method introduces a temporal contrastive loss that enhances temporal consistency in unsupervised video object-centric models. It stabilizes representations across frames, achieving state-of-the-art object discovery on complex real-world videos. + +# Abstract + +Unsupervised object-centric learning from videos is a promising approach to extract structured representations from large, unlabeled collections of videos. To support downstream tasks like autonomous control, these representations must be both compositional and temporally consistent. Existing approaches based on recurrent processing often lack long-term stability across frames because their training objective does not enforce temporal consistency. In this work, we introduce a novel object-level temporal contrastive loss for video object-centric models that explicitly promotes temporal consistency. Our method significantly improves the temporal consistency of the learned object-centric representations, yielding more reliable video decompositions that facilitate challenging downstream tasks such as unsupervised object dynamics prediction. Furthermore, the inductive bias added by our loss strongly improves object discovery, leading to state-of-the-art results on both synthetic and real-world datasets, outperforming even weakly-supervised methods that leverage motion masks as additional cues. + +Visit slotcontrast.github.io for videos and further details. + +# 1. Introduction + +Object-centric learning (OCL) [4, 14, 32, 44] is a rapidly advancing area of visual representation learning that enables autonomous systems to represent, understand, and model high-dimensional data directly in terms of its constituent entities. Structured object-centric representations (often referred to as slots [32]) facilitate generalization and robustness [7, 9] of scene representations across diverse downstream tasks, from visual question answering [1, 8, 34, 58] to control [10, 18, 61, 63]. Of particular interest are video-based object-centric methods [2, 11, 20, 28, 45, 57, 64] that learn to represent objects that evolve and interact over time. These representations make the methods powerful tools for applications such as unsupervised online object tracking [35, 53] and structured world modeling [27, 52, 56]. Unsupervised object-centric learning on videos has seen significant progress in recent years [2, 57, 64], mainly due to the use of pre-trained representations from self-supervised foundational models [5, 38] coupled with diverse training datasets like YouTube-VIS [59, 60]. Nevertheless, these methods still face significant challenges, especially maintaining consistent object-centric representations across time and uniquely representing each object—critical factors for + +successful multi-object tracking and modeling of dynamic scenes [28, 52, 56]. + +Temporal consistency [15, 31, 62] in object-centric representations refers to maintaining the same representation placeholder, called slot, for an object throughout a video sequence, effectively serving as a stable object-specific identifier over time. Existing unsupervised object-centric methods [27, 36, 48] aiming to discover consistent representations have primarily been studied on toy datasets with limited complexity [22, 26, 46]. In contrast, real-world video sequences present numerous challenges, including object occlusions, reappearance, and complex multi-object interactions, which complicate maintaining consistent object representations. + +In this paper, we introduce a novel method to address the challenge of maintaining consistent temporal representations in object-centric models, extending the line of research on slot-based unsupervised video models [11, 64]. Our approach (named SLOT CONTRAST) scales to real-world video data and produces consistent object-centric representations. Notably, it achieves these results without requiring any human annotations. In particular, we propose a novel self-supervised contrastive learning objective, which contrasts slot representations throughout the batch while ensuring temporal coherence across consecutive frames. In addition, we modify the slot's initialization strategy [32] to promote distinct, contrastive representations. This combination leads to improved temporal consistency of learned representations, which we show to be highly effective for challenging downstream tasks such as unsupervised object tracking and latent object dynamics learning. + +Overall, our contributions are as follows: + +- We propose the novel slot-slot contrastive loss that sets the state-of-the-art in temporal consistency when integrated into slot-based video processing methods. +- We develop SLOT CONTRAST, a simple and effective OCL architecture using the slot-slot contrastive loss paired with learned initialization that scales to real-world data, such as YouTube videos. +- We extensively study the usefulness of our learned object-centric representations for challenging downstream tasks, including unsupervised online tracking with complete occlusions and latent object dynamics modeling. +- We show that SLOT CONTRAST does not only improve the temporal consistency of the representations, but also achieves state-of-the-art on the object discovery task, outperforming weakly-supervised models using motion cues. + +# 2. Related Work + +Unsupervised video object-centric learning There exists an extensive body of research [2, 11, 14, 20, 23, 28, 30, 42, 45, 49, 51, 52, 64] on discovering objects from video without any human annotations, primarily through tracking either object bounding boxes or masks. To achieve this, most of these + +works combine an auto-encoder framework with a simple reconstruction objective, adding inductive biases for object discovery through structured encoders [4, 32] and decoders [55]. In particular, many modern object-centric image models [8, 21, 24, 44, 57] use a latent slot attention module [32] to extract object representations and corresponding object masks. For video data, most current methods [11, 28, 45, 49, 64, 66] connect slots across frames, with slots from the previous frame initializing those in the current frame. Notably, recent approaches [2, 40, 64] have successfully scaled object discovery to real-world unconstrained videos. To achieve this, SOLV [2] introduces temporal consistency via agglomerative clustering and prediction of middle-frame features, whereas VideoSAUR [64] learns object-centric representations by predicting temporal similarity of self-supervised features [5, 38]. While such methods can decompose short videos, they still struggle with long-term temporal consistency. In contrast, we show that learning representations that are both informative and contrastive can significantly enhance both object discovery and temporal consistency on longer videos. + +Temporal Consistency Achieving temporal consistency is essential for any computer vision task involving video data, whether it is tracking points, bounding boxes, segmentation masks, optical flow, or representations [29, 33, 41, 47, 50, 52, 54]. In object-centric learning for videos, a range of different approaches have been proposed. For example, Yu and Xu [62] apply an object-wise sequential VAE to achieve consistency; Zhao et al. [65] and Li et al. [31] use an explicit memory buffer to maintain historical slot information and a transformer as a predictor using the memory buffer to predict the future; Qian et al. [40] achieve temporal consistency by employing student-teacher distillation to establish semantic and instance correspondence over time; and Traub et al. [48] use a recurrent network with a constancy prior [17]. + +# 3. Method + +Our approach builds upon the existing input reconstruction-based video object-centric framework [28, 64] by introducing a consistency loss that contrasts the slots across consecutive frames and thereby adapting the model to discover consistent representations. See Fig. 2 for an overview of the SLOT CONTRAST architecture. + +# 3.1. Semantic Recurrent Slot Attention Module + +Our model is an encoder-decoder object-centric architecture based on Slot Attention module (SA) [13] with additional adaptations for sequential inputs similar to SAVi [28], while leveraging pre-trained semantic features as proposed by DINOSAUR [44]. The model consists of three main components: a pre-trained self-supervised dense feature encoder (e.g., DINOv2 [38]), a Recurrent Slot Attention module that + +![](images/dd1cd4ef6e9c8bc14306f040f301a1a5c4dc7c204de975b20ac1e2721e9421f9.jpg) +Figure 2. SLOT CONTRAST model architecture overview. For each frame, we extract patch features $h_t$ using DINOv2 ViT. These features are then used to update the previously initialized or predicted slots, resulting in new slots $S_t$ . The model is trained by contrasting the current frame's slots $S_t$ with the slots from the previous frame $S_{t-1}$ , and by reconstructing the patch features $h_t$ . + +groups the encoder features into slots and models temporal slot updates, and a decoder that maps slots from each frame to reconstructions of the dense self-supervised features used as inputs. Next, we describe those components in more detail while explaining how to adapt them to the task of consistent object-centric representation learning. + +Given a video frame $x_{t}$ , $t \in \{1,2,\dots,T\}$ and a pretrained, frozen self-supervised DINO model $f$ we first extract $N$ patch features $g_{t}$ , + +$$ +g _ {t} = f \left(x _ {t}\right), \quad g _ {t} \in \mathbb {R} ^ {N \times D}. \tag {1} +$$ + +As those frozen features are mostly semantic and are trained only on images, we further adapt them to the task of temporally consistent object discovery. Specifically, each feature vector $g_{t}$ is passed through a MLP $g_{\psi}$ , + +$$ +h _ {t} = g _ {\psi} \left(g _ {t}\right), \tag {2} +$$ + +to adapt the frozen dense features for object-centric grouping (see App. I for more details and visualizations). Based on the transformed encoder features $h_t$ and a set of slot representations of the previous timestep $S_{t-1}^p$ , with $K$ slots $s_{t-1}^{k,p} \in S_{t-1}^p$ , we use a recurrent grouping module to extract slot representations. The Recurrent Slot Attention module comprises a grouping module $C_\theta$ and a predictor module $P_\omega$ . The former updates slot representations using the standard Slot Attention module [32] on visual features $h_t$ from the encoder, while the latter captures temporal and spatial interactions between slots: + +$$ +S _ {t} ^ {c} = \mathrm {C} _ {\theta} \left(h _ {t}, S _ {t - 1} ^ {p}\right), \quad S _ {t} ^ {p} = \mathrm {P} _ {\omega} \left(S _ {t} ^ {c}\right). \tag {3} +$$ + +Both slot-level representations, generated either by the grouping module $S_{t}^{c}$ or the predictor $S_{t}^{p}$ , can be utilized for subsequent decoding or downstream task processing. In our implementation, the slot-level representations from the grouping module $S_{t}^{c}$ are employed for the decoding stage. From now on, we will refer to $S_{t}^{c}$ as $S_{t}$ . + +Temporal Slot Attention Initialization Importantly, we found that our setup benefits considerably from a learned initialization $S_0$ , which can influence the efficiency of training across various objectives. Originally, Locatello et al. [32] proposed a randomly sampled query initialization, where slots are sampled from the same Gaussian distribution with learned mean and variance. While such initialization allows different numbers of slots during inference, sampling from the same Gaussian distribution does not create a particularly favorable structure in slot-space. In this work, we use a straightforward learned initialization [19, 43] where a fixed set of initial slot vectors $S_0$ is learned for the entire dataset. Such initialization allows for learning dissimilar initialization queries that consistently attend to different objects. + +Finally, for the reconstruction loss objective, we decode reconstructions $\hat{g}_t$ from all slots using the MLP decoder [44]. + +# 3.2. Temporal Consistency through Slot Contrast + +Contrastive learning is flexible in supporting diverse data sources and loss function designs. By carefully defining positive and negative examples, we can craft robust loss objectives that effectively guide self-supervised representation learning [6]. For instance, video contrastive methods like CVRL [39] leverage augmented video chunks to define positive (from the same video) and negative (from different videos) examples. In object-centric learning Didolkar et al. [8] employed a contrastive loss function to gain controllability over slot representations guided by language. We propose a novel application of a contrastive loss for temporal consistency in object-centric slot representations. In particular, we define positive samples as the representations of the same slot from two consecutive time steps within a video, while negative samples comprise all other slots across the batch between these time steps. An overview of the proposed loss is presented in Fig. 3 + +![](images/19f3df135ff014dad4f726c4f4b904ada324c6286d15262197705a8dd1e38cb1.jpg) +Figure 3. Overview of the losses used in SLOT CONTRAST. (a) Our proposed temporal consistency objective, slot-slot contrastive loss, operates on a batch of video sequences by enforcing temporal alignment across object slots. For each frame in the sequence, the model groups object features into specific slot representations $S_{t}^{i}$ . The slot-slot contrastive loss then enforces temporal consistency by drawing the corresponding slot representations from adjacent frames closer, while simultaneously pushing apart all other slot representations in the batch—whether they come from different objects within the same video or from objects in other videos. (b) The feature reconstruction loss ensures informativeness of the learned slots by using them to reconstruct original DINOv2 features with an MLP decoder. + +Intra-Video Slot-Slot Contrautive Loss To force each slot to be consistent in time, we aim to learn slots that are similar in time while being maximally dissimilar to other slots. Given the sets of slot representations $S_{t - 1}$ and $S_{t}$ at time steps $t - 1$ and $t$ , we want elements $s_{t - 1}^{i}\in S_{t - 1}$ to be close to the next-frame slots $s_t^i$ corresponding to the same object, while having maximal distance to the next-frame slots $s_t^k,k\neq i$ corresponding to other objects in the video. The corresponding InfoNCE contrastive loss [37] is defined as $\mathcal{L}_{\mathrm{intra}} = \frac{1}{K}\sum_{i = 1}^{K}\ell_{i}^{\mathrm{intra}}$ with + +$$ +\ell_ {i} ^ {\text {i n t r a}} = - \log \frac {\exp \left(\sin \left(s _ {t - 1} ^ {i} , s _ {t} ^ {i}\right) / \tau\right)}{\sum_ {k = 1} ^ {K} \mathbb {1} _ {[ k \neq i ]} \exp \left(\sin \left(s _ {t - 1} ^ {i} , s _ {t} ^ {k}\right) / \tau\right)}, \tag {4} +$$ + +where $K = |S_t|$ is a number of slots per frame, $\mathrm{sim}(u,v) = \frac{u^{\top}v}{\|u\|_2\|v\|_2}$ is the cosine similarity, $\mathbb{1}_{[.]}$ is an indicator excluding the self-similarity of the slot $s_i$ from the denominator, and $\tau >0$ is a temperature parameter. + +While being a desirable property, intra-video slot contrast can be achieved simply by amplifying the differences between slots in the SA module's first frame initialization $S_0$ . To encourage a stronger focus on video content and instance specificity of the representations, we propose a further improvement over this loss by extending the negative contrast set. + +Batch Video Slot-Slot Contrastive Loss To leverage the benefits of larger contrast sets and prevent degenerate solutions relying solely on the initialization of slots, we exploit the fact that the whole batch of videos can be con + +sidered a large set of primarily unique object representations. Consequently, we enhance contrast within a video and between videos by including negative slots from the current and subsequent frames of all videos in the batch. Correspondingly, we define our slot-slot contrastive loss as $\mathcal{L}_{\mathrm{ssc}} = \frac{1}{B\cdot K}\sum_{j = 1}^{B}\sum_{i = 1}^{K}\ell_{i,j}^{\mathrm{ssc}}$ and + +$$ +\ell_ {i, j} ^ {\mathrm {s s c}} = - \log \frac {\exp \left(\operatorname {s i m} \left(s _ {t - 1} ^ {i , j} , s _ {t} ^ {i , j}\right) / \tau\right)}{\sum_ {b = 1} ^ {B} \sum_ {k = 1} ^ {K} \mathbb {1} _ {[ k , b \neq i , j ]} \exp \left(\operatorname {s i m} \left(s _ {t - 1} ^ {i , j} , s _ {t} ^ {k , b}\right) / \tau\right)}, \tag {5} +$$ + +where $B$ is a number of videos in the batch and $s_t^{i,j}$ denotes the $i$ -th slot of the $j$ -th video at time $t$ . For more details on slot-slot contrastive loss implementation, see App. C. + +We find that this approach significantly enhances the effectiveness of the slot-slot contrastive loss. Furthermore, since all videos in the batch are processed with the same initial state $S_0$ , this loss function avoids suboptimal solutions that rely solely on the uniqueness of the initialization, instead encouraging object discovery as the basis for contrast. + +Final Loss To encourage scene decomposition we use a feature reconstruction loss, similar to DINOSAUR [44] and VideoSAUR [64]. Our final loss function combines the reconstruction loss with our proposed contrastive loss $\mathcal{L}_{\mathrm{ssc}}$ weighted by the hyperparameter $\alpha$ (see Table S1 for details on how the hyperparameters are set): + +$$ +\mathcal {L} = \sum_ {t = 1} ^ {T - 1} \mathcal {L} _ {\mathrm {r e c}} \left(\mathrm {h} _ {\mathrm {t}}, \hat {\mathrm {h}} _ {\mathrm {t}}\right) + \alpha \mathcal {L} _ {\mathrm {s s c}} \left(S _ {t - 1}, S _ {t}\right). \tag {6} +$$ + +Table 1. Consistent object-discovery performance of SLOT CONTRAST in comparison with SAVi, STEVE, VideoSAUR on MOVi-C, MOVi-E, and YouTube-VIS datasets. VideoSAURv2 is an improved version of the VideoSAUR trained on DINOv2 features. Both metrics are computed for the whole video (24 frames for MOVi, up to 76 frames for YouTube-VIS). + +
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
SAVi [28]22.213.642.816.0--
STEVE [45]36.126.550.626.61519.1
VideoSAUR [64]64.838.973.935.628.926.3
VideoSAURv2--77.134.431.229.7
SLOT CONTRAST69.332.782.929.238.033.7
+ +# 4. Experiments + +We evaluate our method's temporal consistency on two downstream tasks: object discovery and latent object dynamics prediction. Our experiments address three main questions: (1) How does our model compare to state-of-the-art methods in both temporal consistency and scene decomposition? (2) How effective are our model's learned representations for the challenging downstream task of object dynamics prediction and for object tracking under full occlusions? (3) How important are the different components of our model and loss function for temporal consistency? + +Datasets To evaluate our method in the controlled setting, we use MOVi-C and MOVi-E synthetic datasets generated by Kubric [16]. MOVi-C includes richly textured everyday objects, featuring up to 11 objects per scene, while MOVi-E expands this to 23 objects and introduces basic linear camera motion. In addition, to study the scalability of our method to real-world data, we evaluate our method on the real-world YouTube-VIS 2021 (YTVIS21) video dataset [60]. YTVIS21 is an unconstrained, real-world dataset sourced from YouTube, capturing a diverse range of scenes (for more details, see App. E). + +Metrics Similar to other object-centric video methods [11, 28, 45, 64], to evaluate consistent object discovery, we use the video foreground adjusted rand index (FG-ARI) [14], measuring how well objects are split. In addition, we evaluate the sharpness of masks using the video intersection over union with mean best overlap matching (mBO) metric [44, 64]. Both metrics are computed over the full video and thus reflect how consistent object discovery is. In addition, to investigate the effects of the temporal consistency inductive bias on the per-frame object discovery itself, we use per-frame FG-ARI (image FG-ARI), which we independently compute for each frame and average afterwards. More details can be found in App. F. + +Finally, when evaluating how well object-centric representations perform for object dynamics prediction (see Sec. 4.2), + +we employ the same evaluation metrics as in the object discovery task: FG-ARI and mBO. This time, however, these metrics are computed by comparing the predicted masks (obtained by decoding the predicted slots [56]) with the ground-truth future masks. + +# 4.1. Object Discovery + +Implementation Details We employ the DINOv2 model as our feature encoder, using ViT-S/14 for the MOVi-C dataset and ViT-B/14 for MOVi-E and YTVIS21. The slot dimension is set to 128 for MOVi-E and 64 for both MOVi-C and YTVIS21. For the MOVi datasets, we use a resolution of (336, 336), generating $24 \times 24$ patches yielding 576 ViT tokens, while for YTVIS21, a resolution of (518, 518) yields 1369 tokens. Full details are provided in App. A. + +Baselines We compare SLOT CONTRAST against the previously proposed SAVi [28] and STEVE [45] that employ an image reconstruction objective and with the state-of-the-art method VideoSAUR [64] that uses self-supervised feature reconstructions. Additionally, for a fair comparison, we trained a modification of VideoSAUR with DINOv2 features (referred to as VideoSAURv2). The implementation details are provided in App. G. In addition, to assess how closely SLOT CONTRAST approaches supervised methods, we compared it with SAM2 [41] as a supervised zero-shot baseline for temporal consistency and to weakly-supervised by depth SAVi++ [11] method. The results are in App. G. + +Temporally Consistent Object Discovery (Table 1 & Figure 4) SLOT CONTRAST significantly outperforms both SAVi and STEVE by a wide margin. When compared to VideoSAUR using its default parameters, our approach demonstrates higher consistency in terms of video FG-ARI scores. Compared to VideoSAUR and VideoSAURv2, SLOT CONTRAST achieves superior video scene decomposition (measured by FG-ARI scores). However, on synthetic datasets SLOT CONTRAST's masks are less sharp (as reflected by mBO). Notably, on the most challenging real-world YouTube-VIS data, our method surpasses both versions of VideoSAUR, achieving better performance on FG-ARI (+6.8) and mBO (+4). This shows that, given a large enough resolution and natural data inputs well aligned with DINOv2, SLOT CONTRAST can decompose unconstrained videos into consistent object representations. More examples are illustrated in App. M. + +Per-Frame Scene Decomposition (Table 2) Previous research has shown the effectiveness of specific inductive biases and training objectives for unsupervised object discovery, such as reconstructing in semantic space or leveraging motion cues in video data. Building on this, we demonstrate that the contrastive nature of our temporal consistency loss + +![](images/e151a414bf71b285faae8a9dd142071c0fec9f5db2d1288c66e90ffe17ed3a4a.jpg) +Figure 4. Qualitative comparison with VideoSAURv2 on YouTube-VIS dataset. In challenging situations (e.g., almost full occlusions at $t = 24$ of the 1st video and $t = 14$ of the 2nd video), VideoSAURv2 reassigns slots to different objects (pink arrows), whereas SLOT CONTRAST consistently assigns slots to the same object (green arrows). Note that the colors of the masks are matched manually for better visual comparison. + +![](images/74c39b6206646b68ef77a94fb0899c1a459fe242b9d3256ea2009206d760cc27.jpg) + +Table 2. Quantitative Results on MOVi-E in terms of per-frame Image FG-ARI. The methods are grouped by the target data they train on: only images $(\mathcal{I})$ , videos with motion segmentation annotations $(\mathcal{V} + \mathcal{M})$ , and only videos $(\mathcal{V})$ . + +
ModelObjectiveImage FG-ARI ↑
↓LSD [21]Image Rec.53.4
SlotDiffusion [57]Image Rec.60.0
DINOSAUR [44]Image Rec.65.1
+MoToK [3]+Mot. Seg.66.7
Safadoust et al. [42]+GT Flow78.3
DIOD [25]+Mot. Seg.82.2
↑STEVE [45]Video Rec.54.1
VideoSAUR [64]Temp. Sim.78.4
SOLV [2]Mid. Fr. Pred.80.8
SLOT CONTRASTSlot Contrast84.8
+ +function yields improved scene decomposition as a byproduct. This occurs because our loss function encourages the model to learn consistent feature representations for objects across frames, leading to an adaptive process where dense features become more contrastive, thereby enhancing object discovery in individual frames. + +We compare our method with prior approaches in terms of per-frame object discovery, using the image FG-ARI metric for evaluation. Specifically, we compare three categories of methods: image-based, video-based, and methods that use videos with additional motion cues. Image-based methods use only images as a target (feature reconstruction based DINOSAUR [44] and diffusion-based LSD [21] and SlotDiffusion [57] methods). Video-based methods use only videos as targets: STEVE [45] reconstructs current frame features, SOLV [2] predicts middle frame features, and VideoSAUR [64] predicts temporal feature-similarities. Finally, we also compare with weakly-supervised methods + +using motion masks [3, 25] or ground truth (GT) optical flow [42]. + +The results on MOVi-E dataset are presented in Table 2, with comparisons across additional datasets provided in App. H. Using temporal signals from the video using feature reconstruction is better than object discovery based on images. Next, additional objectives that exploit the temporal structure of the videos allow even better scene decomposition. Notably, our method, which combines a feature reconstruction objective with a simple contrastive objective, leads to state-of-the-art performance reaching 84.8 per-frame FG-ARI, outperforming methods [3, 25] that use motion segmentation masks for object discovery. + +Robustness to Full Occlusions To evaluate our method's robustness in handling complete object occlusions—a challenging scenario for maintaining consistency—we conduct experiments using a targeted subset of the MOVi-C dataset that contains sequences where objects are fully occluded. For evaluation, we retain only the ground-truth masks for the objects that experience occlusion. + +We find that the feature reconstruction baseline achieves only $16\%$ mBO vs. our method obtains $21\%$ mBO on fully occluded objects. Our results suggest that SLOT CONTRAST significantly enhances consistency during object disappearances and reappearance. We refer the reader to Fig. 5 and App. K for visual examples and more details. + +# 4.2. Object Dynamics Prediction + +Setup To evaluate performance on the task of predicting object dynamics, we train a dynamics module using the object-centric representations inferred by a pretrained object-centric model. For this dynamics module, we select SlotFormer [56], which predicts the slots autoregressively for $K$ rollout steps based on the slots inferred from $T$ burn-in frames preceding the prediction horizon. In our setup, we use 14 burn-in frames and 10 rollout steps. Since SlotFormer + +![](images/0424434c6e1046c3a168ce24376ddcef60f00988eea14894ed076e58b147d36b.jpg) +Figure 5. Comparison of the Feature Reconstruction (Feat. Rec.) baseline, the slot-slot contrastive loss using only slots from the same video as the contrast set (Intra-video Contrast), and SLOT CONTRAST on the MOVi-C dataset. + +is trained independently from the object-centric model, we first train the latter, subsequently extending the datasets with the inferred slots for each frame. This approach avoids the computational complexity of training SlotFormer by removing the necessity to encode frames into the slot space at each training step. A brief introduction to SlotFormer and the implementation details can be found in App. J. + +Baselines We compare SlotFormer [56] trained on object-centric representations derived from a model trained using only feature reconstruction loss with SlotFormer trained on representations from SLOT CONTRAST. Both models perform reconstruction in feature space rather than pixel space, so we use only the slot reconstruction loss for training. + +Quality of Predicted Masks (Table 3 and Figure 6) We note that our model has a significantly better FG-ARI on MOVi-C, while mBO is comparable to that of the baseline. On MOVi-E, the performance of our model is comparable to that of the baseline, which highlights the difficulty of adapting to videos with camera motion. There is also a + +![](images/d714bd799befac09576f4ec27c3994f5ad8328c5b03369e96ed0c3ecf468df5d.jpg) +Figure 6. Object dynamics prediction task on MOVi-C using SLOT CONTRAST slots using SlotFormer [56]. + +Table 3. Downstream task of predicting object dynamics. Comparison of predictions made by SlotFormer based on representations obtained from SLOT CONTRAST and from Feature Reconstruction. + +
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
Feat. Rec. + SF50.725.970.624.327.428.9
Ours + SF63.826.170.524.929.229.6
+ +slight improvement in FG-ARI on YTVIS21, while mBO remains comparable. It is worth noting that predicting the motion in this dataset is especially challenging, given the large diversity of possible scenarios. + +# 4.3. Analysis + +In this section, we investigate key components of our approach, including the impact of the contrastive loss and the type of slot initialization. In addition, we study how effective SLOT CONTRAST is in automatically shutting down slots in correspondence to the scene's complexity. + +Ablation of Loss Components (Table 4 and Figure 5) To demonstrate the value of the proposed slot-slot contrastive loss, we carry out an ablation study, comparing it with the feature reconstruction loss [44] and the intra-video contrastive loss, which contrasts slot representations in a single video. Using the intra-video contrastive loss yields improvements over the feature reconstruction baseline (+5.1 FG-ARI and +1.5 mBO on MOVi-C). However, we observe that in more challenging situations, the intra-video contrastive loss leads to failure cases such as shutting down too many slots (see Fig. 5). Next, we observe that by extending the contrast to the full batch of videos, SLOT CONTRAST learns more consistent representations (+19.6 FG-ARI and +5.3 mBO). This change increases the difficulty of the learning task, which prevents the model from relying on superficial patterns like slot initializations or object positions. + +![](images/bcfc7f1984f0184801dcbd6e212f98ba26e9088533413446540741c602ae7171.jpg) +Figure 7. Distribution of ground truth and predicted object numbers (i.e., number of active slots) on the MOVi-C dataset. + +Table 4. Ablation of loss components used by SLOT CONTRAST on MOVi-C, MOVi-E, and YouTube-VIS Datasets. + +
Feat. Rec.Intra-video ContrastSlot-Slot ContrastMOVi-CMOVi-EYouTube-VIS
LrecLintraLsscFG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
✓49.727.479.828.435.331.4
✓✓54.828.978.729.135.733.6
✓✓69.332.782.929.238.033.7
+ +Table 5. Comparison of the random initialization (RI) and learned initialization (LI) techniques. + +
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
Feat. Rec. (RI)45.327.271.128.335.230.2
Feat. Rec. (LI)49.427.879.828.435.331.4
SLOT CONTRAST (RI)62.932.475.328.436.130.8
SLOT CONTRAST (LI)69.332.782.929.238.033.7
+ +Choice of the First Frame Initialization (Table 5) In video object-centric learning, slots are typically initialized based on those from the previous time step [28], while the first frame is initialized from learnable parameters. Previous real-world object-centric methods mostly used random initialization samples from Gaussian distribution [2, 64], in this work, we study the impact of the type of initialization under our contrastive objective. The findings are outlined in Table 5. Our experiments indicate that when combined with slot-slot contrastive loss, learned initialization significantly outperforms random initialization. We hypothesize that this improvement stems from the ability of learned initializations to shape the initial state in a way that enhances contrastiveness, a benefit not achievable with random initialization. For similarity visualizations and more details, see App. B. + +Number of Active Slots per Video (Figure 7) In models based on the Slot Attention mechanism, all available slots are typically utilized [44], leading to a mismatch between the predicted number of components in scene decomposition and the ground-truth number of objects in the scene. This can cause the random splitting of the objects between slots and non-consistent scene representations when slots are reassigned from one object to a part of another object. To address this challenge, it is important to study whether redundant slots can be effectively deactivated. Recently, AdaSlot [12] introduced a discrete slot sampling module, coupled with a complexity-aware prior, to penalize redundant slots explicitly. Similarly, SOLV [2] used agglomerative clustering to merge redundant slots. In this work, we investigate whether SLOT CONTRAST is capable of accurately determining the number of objects in a scene without relying on explicit priors to minimize the number of active slots. + +We compare ground truth and predicted object density on MOVI-C dataset, as shown in Fig. 7. While the feature reconstruction model yields predictions within a narrow + +range—creating a sharp peak near a predefined number of slots—our model, similarly to AdaSlot [12], achieves a smoother prediction distribution that aligns more closely with the ground truth (note that the consistent shift is because 2–3 slots are used for the background, while the ground truth density is computed only for foreground objects). Interestingly, SLOT CONTRAST achieves this without requiring an explicit prior toward sparsity. + +# 5. Conclusion + +SLOT CONTRAST advances unsupervised video object-centric learning by significantly improving the temporal consistency of object representations. Our method explicitly incentivizes temporal consistency by adding a self-supervised contrastive loss. We showed that this loss is not only beneficial for consistency, but also enhances object discovery: SLOT CONTRAST achieves state-of-the-art results on challenging synthetic datasets with many objects and the unconstrained real-world YouTube-VIS dataset. Furthermore, consistent representations directly support temporal downstream tasks such as unsupervised object dynamics prediction and allow for tracking of objects through full occlusions. Finally, SLOT CONTRAST effectively shuts down non-unique slots, leading to a sparser representation that captures the true object distribution more faithfully. Taken together, we expect these improvements to pave the way for broader adoption of video object-centric representations, for instance in applications like word modeling, autonomous control, or video question answering. + +Limitations of our work include the fixed number of slots during initialization. Additionally, we cannot directly control the segmentation granularity of entities. Further limitations and failure cases are discussed in App. L. + +Future work could explore several promising directions. First, one could use SLOT CONTRAST's robust and consistent representations for learning compositional world models from real-world robotics data to enable object-centric planning and control. Second, investigating the compatibility of our contrastive loss with other object-centric learning approaches with different inductive biases, such as SlotDiffusion [57]. Finally, improving the compactness of object masks [24] to achieve more precise object segmentation masks could also benefit unsupervised class-agnostic video object segmentation applications. + +# Acknowledgments + +We thank Christian Gumbsch for useful suggestions and Ke Fan for providing additional details and results for AdaSlot method. Andrii Zadaianchuk is funded by the European Union (ERC, EVA, 950086). The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Maximilian Seitzer. Georg Martius is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. This work was supported by the ERC - 101045454 REAL-RL. We acknowledge the support from the German Federal Ministry of Education and Research (BMBF) through the Tübingen AI Center (FKZ: 01IS18039B). + +# References + +[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In International Conference on Computer Vision (ICCV), 2015. 1 +[2] Görkay Aydemir, Weidi Xie, and Fatma Güney. Self-supervised object-centric learning for videos. In NeurIPS, 2023. 1, 2, 6, 8 +[3] Zhipeng Bao, Pavel Tokmakov, Yu-Xiong Wang, Adrien Gaidon, and Martial Hebert. Object discovery from motion-guided tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22972-22981, 2023. 6 +[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. MONet: Unsupervised Scene Decomposition and Representation. arXiv:1901.11390, 2019. 1, 2 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. ICCV, 2021. 1, 2 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In ICML, 2020. 3 +[7] Aniket Didolkar, Andrii Zadaianchuk, Anirudh Goyal, Mike Mozer, Yoshua Bengio, Georg Martius, and Maximilian Seitzer. Zero-shot object-centric representation learning. arXiv preprint arXiv:2408.09162, 2024. 1 +[8] Aniket Rajiv Didolkar, Andrii Zadaianchuk, Rabiul Awal, Maximilian Seitzer, Efstratios Gavves, and Aishwarya Agrawal. Ctrl-o: Language-controllable object-centric visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3 +[9] Andrea Dittadi, Samuele Papa, Michele De Vita, Bernhard Schölkopf, Ole Winther, and Francesco Locatello. Generalization and Robustness Implications in Object-Centric Learning. In ICML, 2022. 1 +[10] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen + +Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. PaLM-E: An embodied multimodal language model. In ICML, 2023. 1 + +[11] Gamaleldin Fathy Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael Curtis Mozer, and Thomas Kipf. $\mathsf{SAV}\mathsf{i} + +$ : Towards End-to-End Object-Centric Learning from Real-World Videos. In NeurIPS, 2022. 1, 2, 5 + +[12] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23062-23071, 2024. 8 + +[13] Locatello Francesco, Weissenborn Dirk, Unterthiner Thomas, Mahendran Aravindh, Heigold Georg, Uszkoreit Jakob, Dosovitskiy Alexey, and Kipf Thomas. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. 2 + +[14] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-Object Representation Learning with Iterative Variational Inference. In ICML, 2019. 1, 2, 5 +[15] Klaus Greff, Sjoerd Van Steenkiste, and Jürgen Schmidhuber. On the Binding Problem in Artificial Neural Networks. arXiv:2012.05208, 2020. 2 +[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A Scalable Dataset Generator. In CVPR, 2022. 5 +[17] Christian Gumbsch, Martin V Butz, and Georg Martius. Sparsely changing latent states for prediction and planning in partially observable domains. In NeurIPS, 2021. 2 +[18] Dan Haramati, Tal Daniel, and Aviv Tamar. Entity-centric reinforcement learning for object manipulation from pixels. In ICLR, 2024. 1 +[19] Baoxiong Jia, Yu Liu, and Siyuan Huang. Improving object-centric learning with query optimization. arXiv preprint arXiv:2210.08990, 2022.3 +[20] Jindong Jiang, Sepehr Janghorbani, Gerard de Melo, and Sungjin Ahn. SCALOR: Generative World Models with Scalable Object Representations. In ICLR, 2020. 1, 2 +[21] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In NeurIPS, 2023. 2, 6 +[22] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 2 + +[23] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthew, Antonia Creswell, Matthew Botvinick, Alexander Lerchner, and Christopher P. Burgess. SIMONE: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. In NeurIPS, 2021. 2 +[24] Ioannis Kakogeorgiou, Spyros Gidaris, Konstantinos Karantzalos, and Nikos Komodakis. Spot: Self-training with patch-order permutation for object-centric learning with autoregressive transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22776-22786, 2024. 2, 8 +[25] Sandra Kara, Hejer Ammar, Julien Denize, Florian Chabot, and Quoc-Cuong Pham. Diod: Self-distillation meets object discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3975-3985, 2024. 6 +[26] Laurynas Karazija, Iro Laina, and Christian Rupprecht. ClevrTex: A Texture-Rich Benchmark for Unsupervised Multi-Object Segmentation. In NeurIPS Track on Datasets and Benchmarks, 2021. 2 +[27] Thomas Kipf, Elise Van der Pol, and Max Welling. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. 1, 2 +[28] Thomas Kipf, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-centric Learning from Video. In ICLR, 2022. 1, 2, 5, 8 +[29] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2 +[30] Adam Kosiorek, Hyunjik Kim, Yee Whye Teh, and Ingmar Posner. Sequential Attend, Infer, Repeat: Generative Modelling of Moving Objects. In NeurIPS, 2018. 2 +[31] Jian Li, Pu Ren, Yang Liu, and Hao Sun. Reasoning-enhanced object-centric learning for videos. arXiv preprint arXiv:2403.15245, 2024. 2 +[32] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-Centric Learning with Slot Attention. In NeurIPS, 2020. 1, 2, 3 +[33] Zijia Lu, Bing Shuai, Yanbei Chen, Zhenlin Xu, and Davide Modolo. Self-supervised multi-object tracking with path consistency. arXiv preprint, 2024. 2 +[34] Amir Mohammad Karimi Mamaghan, Samuele Papa, Karl Henrik Johansson, Stefan Bauer, and Andrea Dittadi. Exploring the effectiveness of object-centric representations in visual question answering: Comparative insights with foundation models. arXiv preprint arXiv:2407.15589, 2024. 1 +[35] Sha Meng, Dian Shao, Jiacheng Guo, and Shan Gao. Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16264-16273, 2023. 1 +[36] Cristian Meo, Akihiro Nakano, Mircea Lică, Aniket Didolkar, Masahiro Suzuki, Anirudh Goyal, Mengmi Zhang, Justin + +Dauwels, Yutaka Matsuo, and Yoshua Bengio. Object-centric temporal consistency via conditional autoregressive inductive biases. arXiv preprint arXiv:2410.15728, 2024. 2 +[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4 +[38] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision, 2023. 1, 2 +[39] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6964-6974, 2021. 3 +[40] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In ICCV, 2023. 2 +[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 5 +[42] Sadra Safadoust and Fatma Güney. Multi-object discovery by low-dimensional object motion. In ICCV, 2023. 2, 6 +[43] Mehdi SM Sajjadi, Daniel Duckworth, Aravindh Mahendran, Sjoerd Van Steenkiste, Filip Pavetic, Mario Lucic, Leonidas J Guibas, Klaus Greff, and Thomas Kipf. Object scene representation transformer. Advances in Neural Information Processing Systems, 35:9512-9524, 2022. 3 +[44] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. In ICLR, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[45] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple Unsupervised Object-Centric Learning for Complex and Naturalistic Videos. In NeurIPS, 2022. 1, 2, 5, 6 +[46] Matthias Tangemann, Steffen Schneider, Julius Von Kügelgen, Francesco Locatello, Peter Gehler, Thomas Brox, Matthias Kümmerer, Matthias Bethge, and Bernhard Schölkopf. Unsupervised object learning via common fate. arXiv preprint arXiv:2110.06562, 2021. 2 +[47] Pavel Tokmakov, Allan Jabri, Jie Li, and Adrien Gaidon. Object permanence emerges in a random walk along memory. arXiv preprint arXiv:2204.01784, 2022. 2 +[48] Manuel Traub, Frederic Becker, Sebastian Otte, and Martin V Butz. Looping loci: Developing object permanence from videos. arXiv preprint arXiv:2310.10372, 2023. 2 +[49] Manuel Traub, Sebastian Otte, Tobias Menge, Matthias Karlbauer, Jannik Thuemmel, and Martin V. Butz. Learning What + +and Where: Disentangling Location and Identity Tracking Without Supervision. In ICLR, 2023. 2 +[50] Basile Van Hoorick, Pavel Tokmakov, Simon Stent, Jie Li, and Carl Vondrick. Tracking through containers and occluders in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13802-13812, 2023. 2 +[51] Sjoerd van Steenkiste, Michael Chang, Klaus Greff, and Jürgen Schmidhuber. Relational neural expectation maximization: Unsupervised discovery of objects and their interactions. In ICLR, 2018. 2 +[52] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity Abstraction in Visual Model-based Reinforcement Learning. In Conference on Robot Learning, 2019. 1, 2 +[53] Ning Wang, Yibing Song, Chao Ma, Wengang Zhou, Wei Liu, and Houqiang Li. Unsupervised deep tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1308-1317, 2019. 1 +[54] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2566-2576, 2019. 2 +[55] Nicholas Watters, Loic Matthew, Christopher P Burgess, and Alexander Lerchner. Spatial broadcast decoder: A simple architecture for learning disentangled representations in vaes. arXiv preprint arXiv:1901.07017, 2019. 2 +[56] Ziyi Wu, Nikita Dvornik, Klaus Greff, Thomas Kipf, and Animesh Garg. Slotformer: Unsupervised visual dynamics simulation with object-centric models. In ICLR, 2023. 1, 2, 5, 6, 7 +[57] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotdiffusion: Object-centric generative modeling with diffusion models. In NeurIPS, 2023. 1, 2, 6, 8 +[58] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Slot-VLM: Object-event slots for video-language modeling. In NeurIPS, 2024. 1 +[59] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1 +[60] Linjie Yang, Yuchen Fan, Yang Fu, and Ning Xu. The 3rd large-scale video object segmentation challenge - video instance segmentation track, 2021. 1, 5 +[61] Jaesik Yoon, Yi-Fu Wu, Heechul Bae, and Sungjin Ahn. An investigation into pre-training object-centric representations for reinforcement learning. In ICML, 2023. 1 +[62] Haonan Yu and Wei Xu. Vonet: Unsupervised video object learning with parallel u-net attention and object-wise sequentialvae.arXiv preprint arXiv:2401.11110, 2024.2 +[63] Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius. Self-supervised Visual Reinforcement Learning with Object-centric Representations. In ICLR, 2020. 1 +[64] Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius. Object-centric learning for real-world videos by predicting temporal feature similarities. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023), 2023. 1, 2, 4, 5, 6, 8 + +[65] Zixu Zhao, Jiaze Wang, Max Horn, Yizhuo Ding, Tong He, Zechen Bai, Dominik Zietlow, Carl-Johann Simon-Gabriel, Bing Shuai, Zhuowen Tu, Thomas Brox, Bernt Schiele, Yanwei Fu, Francesco Locatello, Zheng Zhang, and Tianjun Xiao. Object-centric multiple object tracking. ICCV, 2023. 2 +[66] Daniel Zoran, Rishabh Kabra, Alexander Lerchner, and Danilo J. Rezende. PARTS: Unsupervised segmentation with slots, attention and independence maximization. In ICCV, 2021. 2 \ No newline at end of file diff --git a/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/images.zip b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..33012bd2063bbd73d45ac6195d0c09c5f16a8aa3 --- /dev/null +++ b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a7fc2503da3e722c7480379085afbc7ec349c7d2cc5cd818d0ba2ed95351953 +size 548333 diff --git a/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/layout.json b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..47e70193f0aea4c1767d34d8a3bb2d1a691135ff --- /dev/null +++ b/2025/Temporally Consistent Object-Centric Learning by Contrasting Slots/layout.json @@ -0,0 +1,7935 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "type": "text", + "content": "Temporally Consistent Object-Centric Learning by Contrasting Slots" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 137, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 553, + 179 + ], + "type": "text", + "content": "Anna Manasyan1 Maximilian Seitzer1 Filip Radovic1 Georg Martius1,3 Andrii Zadaianchuk2 \n1University of Tübingen, Tübingen, Germany \n2University of Amsterdam, Amsterdam, Netherlands \n3Max Planck Institute for Intelligent Systems, Tübingen, Germany" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 182, + 470, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 182, + 470, + 193 + ], + "spans": [ + { + "bbox": [ + 140, + 182, + 470, + 193 + ], + "type": "text", + "content": "anna.manasyan@student.uni-tuebingen.de a.zadaianchuk@uva.nl" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 209, + 553, + 376 + ], + "blocks": [ + { + "bbox": [ + 56, + 209, + 553, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 209, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 553, + 376 + ], + "type": "image", + "image_path": "0e4ee66c9e4ae74713e74fd1a72338564bdc5e1654deaedb6eb6f41340bb7b3d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 384, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. SLOT CONTRAST: Our method introduces a temporal contrastive loss that enhances temporal consistency in unsupervised video object-centric models. It stabilizes representations across frames, achieving state-of-the-art object discovery on complex real-world videos." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 427, + 200, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 427, + 200, + 440 + ], + "spans": [ + { + "bbox": [ + 152, + 427, + 200, + 440 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 460, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 297, + 687 + ], + "type": "text", + "content": "Unsupervised object-centric learning from videos is a promising approach to extract structured representations from large, unlabeled collections of videos. To support downstream tasks like autonomous control, these representations must be both compositional and temporally consistent. Existing approaches based on recurrent processing often lack long-term stability across frames because their training objective does not enforce temporal consistency. In this work, we introduce a novel object-level temporal contrastive loss for video object-centric models that explicitly promotes temporal consistency. Our method significantly improves the temporal consistency of the learned object-centric representations, yielding more reliable video decompositions that facilitate challenging downstream tasks such as unsupervised object dynamics prediction. Furthermore, the inductive bias added by our loss strongly improves object discovery, leading to state-of-the-art results on both synthetic and real-world datasets, outperforming even weakly-supervised methods that leverage motion masks as additional cues." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 688, + 285, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 688, + 285, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 285, + 700 + ], + "type": "text", + "content": "Visit slotcontrast.github.io for videos and further details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 427, + 394, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 427, + 394, + 440 + ], + "spans": [ + { + "bbox": [ + 315, + 427, + 394, + 440 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": "Object-centric learning (OCL) [4, 14, 32, 44] is a rapidly advancing area of visual representation learning that enables autonomous systems to represent, understand, and model high-dimensional data directly in terms of its constituent entities. Structured object-centric representations (often referred to as slots [32]) facilitate generalization and robustness [7, 9] of scene representations across diverse downstream tasks, from visual question answering [1, 8, 34, 58] to control [10, 18, 61, 63]. Of particular interest are video-based object-centric methods [2, 11, 20, 28, 45, 57, 64] that learn to represent objects that evolve and interact over time. These representations make the methods powerful tools for applications such as unsupervised online object tracking [35, 53] and structured world modeling [27, 52, 56]. Unsupervised object-centric learning on videos has seen significant progress in recent years [2, 57, 64], mainly due to the use of pre-trained representations from self-supervised foundational models [5, 38] coupled with diverse training datasets like YouTube-VIS [59, 60]. Nevertheless, these methods still face significant challenges, especially maintaining consistent object-centric representations across time and uniquely representing each object—critical factors for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5401" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "successful multi-object tracking and modeling of dynamic scenes [28, 52, 56]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "text", + "content": "Temporal consistency [15, 31, 62] in object-centric representations refers to maintaining the same representation placeholder, called slot, for an object throughout a video sequence, effectively serving as a stable object-specific identifier over time. Existing unsupervised object-centric methods [27, 36, 48] aiming to discover consistent representations have primarily been studied on toy datasets with limited complexity [22, 26, 46]. In contrast, real-world video sequences present numerous challenges, including object occlusions, reappearance, and complex multi-object interactions, which complicate maintaining consistent object representations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 228, + 297, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 297, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 297, + 430 + ], + "type": "text", + "content": "In this paper, we introduce a novel method to address the challenge of maintaining consistent temporal representations in object-centric models, extending the line of research on slot-based unsupervised video models [11, 64]. Our approach (named SLOT CONTRAST) scales to real-world video data and produces consistent object-centric representations. Notably, it achieves these results without requiring any human annotations. In particular, we propose a novel self-supervised contrastive learning objective, which contrasts slot representations throughout the batch while ensuring temporal coherence across consecutive frames. In addition, we modify the slot's initialization strategy [32] to promote distinct, contrastive representations. This combination leads to improved temporal consistency of learned representations, which we show to be highly effective for challenging downstream tasks such as unsupervised object tracking and latent object dynamics learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 431, + 235, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 431, + 235, + 442 + ], + "spans": [ + { + "bbox": [ + 67, + 431, + 235, + 442 + ], + "type": "text", + "content": "Overall, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 295, + 623 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 445, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 479 + ], + "type": "text", + "content": "- We propose the novel slot-slot contrastive loss that sets the state-of-the-art in temporal consistency when integrated into slot-based video processing methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 526 + ], + "type": "text", + "content": "- We develop SLOT CONTRAST, a simple and effective OCL architecture using the slot-slot contrastive loss paired with learned initialization that scales to real-world data, such as YouTube videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 295, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 575 + ], + "type": "text", + "content": "- We extensively study the usefulness of our learned object-centric representations for challenging downstream tasks, including unsupervised online tracking with complete occlusions and latent object dynamics modeling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 623 + ], + "type": "text", + "content": "- We show that SLOT CONTRAST does not only improve the temporal consistency of the representations, but also achieves state-of-the-art on the object discovery task, outperforming weakly-supervised models using motion cues." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 634, + 142, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 142, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 142, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Unsupervised video object-centric learning There exists an extensive body of research [2, 11, 14, 20, 23, 28, 30, 42, 45, 49, 51, 52, 64] on discovering objects from video without any human annotations, primarily through tracking either object bounding boxes or masks. To achieve this, most of these" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": "works combine an auto-encoder framework with a simple reconstruction objective, adding inductive biases for object discovery through structured encoders [4, 32] and decoders [55]. In particular, many modern object-centric image models [8, 21, 24, 44, 57] use a latent slot attention module [32] to extract object representations and corresponding object masks. For video data, most current methods [11, 28, 45, 49, 64, 66] connect slots across frames, with slots from the previous frame initializing those in the current frame. Notably, recent approaches [2, 40, 64] have successfully scaled object discovery to real-world unconstrained videos. To achieve this, SOLV [2] introduces temporal consistency via agglomerative clustering and prediction of middle-frame features, whereas VideoSAUR [64] learns object-centric representations by predicting temporal similarity of self-supervised features [5, 38]. While such methods can decompose short videos, they still struggle with long-term temporal consistency. In contrast, we show that learning representations that are both informative and contrastive can significantly enhance both object discovery and temporal consistency on longer videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 329, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 556, + 498 + ], + "type": "text", + "content": "Temporal Consistency Achieving temporal consistency is essential for any computer vision task involving video data, whether it is tracking points, bounding boxes, segmentation masks, optical flow, or representations [29, 33, 41, 47, 50, 52, 54]. In object-centric learning for videos, a range of different approaches have been proposed. For example, Yu and Xu [62] apply an object-wise sequential VAE to achieve consistency; Zhao et al. [65] and Li et al. [31] use an explicit memory buffer to maintain historical slot information and a transformer as a predictor using the memory buffer to predict the future; Qian et al. [40] achieve temporal consistency by employing student-teacher distillation to establish semantic and instance correspondence over time; and Traub et al. [48] use a recurrent network with a constancy prior [17]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 510, + 370, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 370, + 521 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 370, + 521 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 530, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 601 + ], + "type": "text", + "content": "Our approach builds upon the existing input reconstruction-based video object-centric framework [28, 64] by introducing a consistency loss that contrasts the slots across consecutive frames and thereby adapting the model to discover consistent representations. See Fig. 2 for an overview of the SLOT CONTRAST architecture." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 611, + 537, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 537, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 537, + 623 + ], + "type": "text", + "content": "3.1. Semantic Recurrent Slot Attention Module" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Our model is an encoder-decoder object-centric architecture based on Slot Attention module (SA) [13] with additional adaptations for sequential inputs similar to SAVi [28], while leveraging pre-trained semantic features as proposed by DINOSAUR [44]. The model consists of three main components: a pre-trained self-supervised dense feature encoder (e.g., DINOv2 [38]), a Recurrent Slot Attention module that" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5402" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 69, + 504, + 209 + ], + "blocks": [ + { + "bbox": [ + 109, + 69, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 109, + 69, + 504, + 209 + ], + "spans": [ + { + "bbox": [ + 109, + 69, + 504, + 209 + ], + "type": "image", + "image_path": "dd1cd4ef6e9c8bc14306f040f301a1a5c4dc7c204de975b20ac1e2721e9421f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "text", + "content": "Figure 2. SLOT CONTRAST model architecture overview. For each frame, we extract patch features " + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "text", + "content": " using DINOv2 ViT. These features are then used to update the previously initialized or predicted slots, resulting in new slots " + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "text", + "content": ". The model is trained by contrasting the current frame's slots " + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "text", + "content": " with the slots from the previous frame " + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "inline_equation", + "content": "S_{t-1}" + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "text", + "content": ", and by reconstructing the patch features " + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 265, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 336 + ], + "type": "text", + "content": "groups the encoder features into slots and models temporal slot updates, and a decoder that maps slots from each frame to reconstructions of the dense self-supervised features used as inputs. Next, we describe those components in more detail while explaining how to adapt them to the task of consistent object-centric representation learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "text", + "content": "Given a video frame " + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "inline_equation", + "content": "t \\in \\{1,2,\\dots,T\\}" + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "text", + "content": " and a pretrained, frozen self-supervised DINO model " + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "text", + "content": " we first extract " + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "text", + "content": " patch features " + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 55, + 336, + 296, + 373 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 380, + 296, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 380, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 118, + 380, + 296, + 395 + ], + "type": "interline_equation", + "content": "g _ {t} = f \\left(x _ {t}\\right), \\quad g _ {t} \\in \\mathbb {R} ^ {N \\times D}. \\tag {1}", + "image_path": "9f5eeeafe431b487f28cadc8a1abf721d84a527ab78c4217c23c42971d1cb33f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 452 + ], + "type": "text", + "content": "As those frozen features are mostly semantic and are trained only on images, we further adapt them to the task of temporally consistent object discovery. Specifically, each feature vector " + }, + { + "bbox": [ + 55, + 402, + 296, + 452 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 55, + 402, + 296, + 452 + ], + "type": "text", + "content": " is passed through a MLP " + }, + { + "bbox": [ + 55, + 402, + 296, + 452 + ], + "type": "inline_equation", + "content": "g_{\\psi}" + }, + { + "bbox": [ + 55, + 402, + 296, + 452 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 459, + 295, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 459, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 147, + 459, + 295, + 472 + ], + "type": "interline_equation", + "content": "h _ {t} = g _ {\\psi} \\left(g _ {t}\\right), \\tag {2}", + "image_path": "e7baa8f3c3b54fc8f277643fbbdb89a28e4a4f9ea7cb1d25831a30e90f5b177a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": "to adapt the frozen dense features for object-centric grouping (see App. I for more details and visualizations). Based on the transformed encoder features " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " and a set of slot representations of the previous timestep " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "S_{t-1}^p" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " slots " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "s_{t-1}^{k,p} \\in S_{t-1}^p" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": ", we use a recurrent grouping module to extract slot representations. The Recurrent Slot Attention module comprises a grouping module " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "C_\\theta" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " and a predictor module " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "P_\\omega" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": ". The former updates slot representations using the standard Slot Attention module [32] on visual features " + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 55, + 479, + 296, + 612 + ], + "type": "text", + "content": " from the encoder, while the latter captures temporal and spatial interactions between slots:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 100, + 620, + 296, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 620, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 100, + 620, + 296, + 634 + ], + "type": "interline_equation", + "content": "S _ {t} ^ {c} = \\mathrm {C} _ {\\theta} \\left(h _ {t}, S _ {t - 1} ^ {p}\\right), \\quad S _ {t} ^ {p} = \\mathrm {P} _ {\\omega} \\left(S _ {t} ^ {c}\\right). \\tag {3}", + "image_path": "e12f95e0a09d2d813c021c8be30da662b2f2b3fe26ce5b9f2ad889676ecd31bd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "Both slot-level representations, generated either by the grouping module " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{t}^{c}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " or the predictor " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{t}^{p}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", can be utilized for subsequent decoding or downstream task processing. In our implementation, the slot-level representations from the grouping module " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{t}^{c}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " are employed for the decoding stage. From now on, we will refer to " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{t}^{c}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 265, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 434 + ], + "type": "text", + "content": "Temporal Slot Attention Initialization Importantly, we found that our setup benefits considerably from a learned initialization " + }, + { + "bbox": [ + 313, + 265, + 555, + 434 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 313, + 265, + 555, + 434 + ], + "type": "text", + "content": ", which can influence the efficiency of training across various objectives. Originally, Locatello et al. [32] proposed a randomly sampled query initialization, where slots are sampled from the same Gaussian distribution with learned mean and variance. While such initialization allows different numbers of slots during inference, sampling from the same Gaussian distribution does not create a particularly favorable structure in slot-space. In this work, we use a straightforward learned initialization [19, 43] where a fixed set of initial slot vectors " + }, + { + "bbox": [ + 313, + 265, + 555, + 434 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 313, + 265, + 555, + 434 + ], + "type": "text", + "content": " is learned for the entire dataset. Such initialization allows for learning dissimilar initialization queries that consistently attend to different objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 438, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 463 + ], + "type": "text", + "content": "Finally, for the reconstruction loss objective, we decode reconstructions " + }, + { + "bbox": [ + 313, + 438, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\hat{g}_t" + }, + { + "bbox": [ + 313, + 438, + 555, + 463 + ], + "type": "text", + "content": " from all slots using the MLP decoder [44]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 487, + 547, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 547, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 547, + 500 + ], + "type": "text", + "content": "3.2. Temporal Consistency through Slot Contrast" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "Contrastive learning is flexible in supporting diverse data sources and loss function designs. By carefully defining positive and negative examples, we can craft robust loss objectives that effectively guide self-supervised representation learning [6]. For instance, video contrastive methods like CVRL [39] leverage augmented video chunks to define positive (from the same video) and negative (from different videos) examples. In object-centric learning Didolkar et al. [8] employed a contrastive loss function to gain controllability over slot representations guided by language. We propose a novel application of a contrastive loss for temporal consistency in object-centric slot representations. In particular, we define positive samples as the representations of the same slot from two consecutive time steps within a video, while negative samples comprise all other slots across the batch between these time steps. An overview of the proposed loss is presented in Fig. 3" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5403" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 526, + 260 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 526, + 260 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 526, + 260 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 526, + 260 + ], + "type": "image", + "image_path": "19f3df135ff014dad4f726c4f4b904ada324c6286d15262197705a8dd1e38cb1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "type": "text", + "content": "Figure 3. Overview of the losses used in SLOT CONTRAST. (a) Our proposed temporal consistency objective, slot-slot contrastive loss, operates on a batch of video sequences by enforcing temporal alignment across object slots. For each frame in the sequence, the model groups object features into specific slot representations " + }, + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "type": "inline_equation", + "content": "S_{t}^{i}" + }, + { + "bbox": [ + 54, + 265, + 555, + 331 + ], + "type": "text", + "content": ". The slot-slot contrastive loss then enforces temporal consistency by drawing the corresponding slot representations from adjacent frames closer, while simultaneously pushing apart all other slot representations in the batch—whether they come from different objects within the same video or from objects in other videos. (b) The feature reconstruction loss ensures informativeness of the learned slots by using them to reconstruct original DINOv2 features with an MLP decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "content": "Intra-Video Slot-Slot Contrautive Loss To force each slot to be consistent in time, we aim to learn slots that are similar in time while being maximally dissimilar to other slots. Given the sets of slot representations " + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "inline_equation", + "content": "S_{t - 1}" + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "content": " at time steps " + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "content": ", we want elements " + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "inline_equation", + "content": "s_{t - 1}^{i}\\in S_{t - 1}" + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "content": " to be close to the next-frame slots " + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "inline_equation", + "content": "s_t^i" + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "content": " corresponding to the same object, while having maximal distance to the next-frame slots " + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "inline_equation", + "content": "s_t^k,k\\neq i" + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "content": " corresponding to other objects in the video. The corresponding InfoNCE contrastive loss [37] is defined as " + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{intra}} = \\frac{1}{K}\\sum_{i = 1}^{K}\\ell_{i}^{\\mathrm{intra}}" + }, + { + "bbox": [ + 55, + 352, + 296, + 474 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 481, + 296, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 481, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 79, + 481, + 296, + 508 + ], + "type": "interline_equation", + "content": "\\ell_ {i} ^ {\\text {i n t r a}} = - \\log \\frac {\\exp \\left(\\sin \\left(s _ {t - 1} ^ {i} , s _ {t} ^ {i}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {K} \\mathbb {1} _ {[ k \\neq i ]} \\exp \\left(\\sin \\left(s _ {t - 1} ^ {i} , s _ {t} ^ {k}\\right) / \\tau\\right)}, \\tag {4}", + "image_path": "2d90d96908237429aba08c9381aa258906ee7834f4eeee7daffe9fb61afa60ab.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "inline_equation", + "content": "K = |S_t|" + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "text", + "content": " is a number of slots per frame, " + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(u,v) = \\frac{u^{\\top}v}{\\|u\\|_2\\|v\\|_2}" + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "text", + "content": " is the cosine similarity, " + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{[.]}" + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "text", + "content": " is an indicator excluding the self-similarity of the slot " + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "text", + "content": " from the denominator, and " + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "inline_equation", + "content": "\\tau >0" + }, + { + "bbox": [ + 55, + 514, + 296, + 566 + ], + "type": "text", + "content": " is a temperature parameter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 566, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 296, + 649 + ], + "type": "text", + "content": "While being a desirable property, intra-video slot contrast can be achieved simply by amplifying the differences between slots in the SA module's first frame initialization " + }, + { + "bbox": [ + 55, + 566, + 296, + 649 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 55, + 566, + 296, + 649 + ], + "type": "text", + "content": ". To encourage a stronger focus on video content and instance specificity of the representations, we propose a further improvement over this loss by extending the negative contrast set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "type": "text", + "content": "Batch Video Slot-Slot Contrastive Loss To leverage the benefits of larger contrast sets and prevent degenerate solutions relying solely on the initialization of slots, we exploit the fact that the whole batch of videos can be con" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 352, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 426 + ], + "type": "text", + "content": "sidered a large set of primarily unique object representations. Consequently, we enhance contrast within a video and between videos by including negative slots from the current and subsequent frames of all videos in the batch. Correspondingly, we define our slot-slot contrastive loss as " + }, + { + "bbox": [ + 313, + 352, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssc}} = \\frac{1}{B\\cdot K}\\sum_{j = 1}^{B}\\sum_{i = 1}^{K}\\ell_{i,j}^{\\mathrm{ssc}}" + }, + { + "bbox": [ + 313, + 352, + 555, + 426 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 434, + 555, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 434, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 325, + 434, + 555, + 473 + ], + "type": "interline_equation", + "content": "\\ell_ {i, j} ^ {\\mathrm {s s c}} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(s _ {t - 1} ^ {i , j} , s _ {t} ^ {i , j}\\right) / \\tau\\right)}{\\sum_ {b = 1} ^ {B} \\sum_ {k = 1} ^ {K} \\mathbb {1} _ {[ k , b \\neq i , j ]} \\exp \\left(\\operatorname {s i m} \\left(s _ {t - 1} ^ {i , j} , s _ {t} ^ {k , b}\\right) / \\tau\\right)}, \\tag {5}", + "image_path": "0ef3d3c692b2ab4d46edb868dad8952d9adf1b32a076e173e1ff7a52f3012fd3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "text", + "content": " is a number of videos in the batch and " + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "inline_equation", + "content": "s_t^{i,j}" + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "text", + "content": "-th slot of the " + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "text", + "content": "-th video at time " + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 482, + 553, + 518 + ], + "type": "text", + "content": ". For more details on slot-slot contrastive loss implementation, see App. C." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 519, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 591 + ], + "type": "text", + "content": "We find that this approach significantly enhances the effectiveness of the slot-slot contrastive loss. Furthermore, since all videos in the batch are processed with the same initial state " + }, + { + "bbox": [ + 313, + 519, + 555, + 591 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 313, + 519, + 555, + 591 + ], + "type": "text", + "content": ", this loss function avoids suboptimal solutions that rely solely on the uniqueness of the initialization, instead encouraging object discovery as the basis for contrast." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 606, + 556, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 679 + ], + "type": "text", + "content": "Final Loss To encourage scene decomposition we use a feature reconstruction loss, similar to DINOSAUR [44] and VideoSAUR [64]. Our final loss function combines the reconstruction loss with our proposed contrastive loss " + }, + { + "bbox": [ + 313, + 606, + 556, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 679 + ], + "type": "text", + "content": " weighted by the hyperparameter " + }, + { + "bbox": [ + 313, + 606, + 556, + 679 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 606, + 556, + 679 + ], + "type": "text", + "content": " (see Table S1 for details on how the hyperparameters are set):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 686, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 686, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 356, + 686, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {t = 1} ^ {T - 1} \\mathcal {L} _ {\\mathrm {r e c}} \\left(\\mathrm {h} _ {\\mathrm {t}}, \\hat {\\mathrm {h}} _ {\\mathrm {t}}\\right) + \\alpha \\mathcal {L} _ {\\mathrm {s s c}} \\left(S _ {t - 1}, S _ {t}\\right). \\tag {6}", + "image_path": "97be52b7254dae9c6b55726244c5a3ff99b6198b6f62e6b7538ecb1d1e0cb2e6.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5404" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 140, + 296, + 216 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 137 + ], + "type": "text", + "content": "Table 1. Consistent object-discovery performance of SLOT CONTRAST in comparison with SAVi, STEVE, VideoSAUR on MOVi-C, MOVi-E, and YouTube-VIS datasets. VideoSAURv2 is an improved version of the VideoSAUR trained on DINOv2 features. Both metrics are computed for the whole video (24 frames for MOVi, up to 76 frames for YouTube-VIS)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 140, + 296, + 216 + ], + "lines": [ + { + "bbox": [ + 57, + 140, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 296, + 216 + ], + "type": "table", + "html": "
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
SAVi [28]22.213.642.816.0--
STEVE [45]36.126.550.626.61519.1
VideoSAUR [64]64.838.973.935.628.926.3
VideoSAURv2--77.134.431.229.7
SLOT CONTRAST69.332.782.929.238.033.7
", + "image_path": "1552eaed03f6bda5bcaefc96eaadc617f0acccdd38cb70ef143e42d884b1cae8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 230, + 138, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 138, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 138, + 243 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 251, + 297, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 370 + ], + "type": "text", + "content": "We evaluate our method's temporal consistency on two downstream tasks: object discovery and latent object dynamics prediction. Our experiments address three main questions: (1) How does our model compare to state-of-the-art methods in both temporal consistency and scene decomposition? (2) How effective are our model's learned representations for the challenging downstream task of object dynamics prediction and for object tracking under full occlusions? (3) How important are the different components of our model and loss function for temporal consistency?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 386, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 386, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 386, + 296, + 518 + ], + "type": "text", + "content": "Datasets To evaluate our method in the controlled setting, we use MOVi-C and MOVi-E synthetic datasets generated by Kubric [16]. MOVi-C includes richly textured everyday objects, featuring up to 11 objects per scene, while MOVi-E expands this to 23 objects and introduces basic linear camera motion. In addition, to study the scalability of our method to real-world data, we evaluate our method on the real-world YouTube-VIS 2021 (YTVIS21) video dataset [60]. YTVIS21 is an unconstrained, real-world dataset sourced from YouTube, capturing a diverse range of scenes (for more details, see App. E)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 297, + 689 + ], + "type": "text", + "content": "Metrics Similar to other object-centric video methods [11, 28, 45, 64], to evaluate consistent object discovery, we use the video foreground adjusted rand index (FG-ARI) [14], measuring how well objects are split. In addition, we evaluate the sharpness of masks using the video intersection over union with mean best overlap matching (mBO) metric [44, 64]. Both metrics are computed over the full video and thus reflect how consistent object discovery is. In addition, to investigate the effects of the temporal consistency inductive bias on the per-frame object discovery itself, we use per-frame FG-ARI (image FG-ARI), which we independently compute for each frame and average afterwards. More details can be found in App. F." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "content": "Finally, when evaluating how well object-centric representations perform for object dynamics prediction (see Sec. 4.2)," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "we employ the same evaluation metrics as in the object discovery task: FG-ARI and mBO. This time, however, these metrics are computed by comparing the predicted masks (obtained by decoding the predicted slots [56]) with the ground-truth future masks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 138, + 419, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 419, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 419, + 152 + ], + "type": "text", + "content": "4.1. Object Discovery" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 156, + 556, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 556, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 556, + 254 + ], + "type": "text", + "content": "Implementation Details We employ the DINOv2 model as our feature encoder, using ViT-S/14 for the MOVi-C dataset and ViT-B/14 for MOVi-E and YTVIS21. The slot dimension is set to 128 for MOVi-E and 64 for both MOVi-C and YTVIS21. For the MOVi datasets, we use a resolution of (336, 336), generating " + }, + { + "bbox": [ + 313, + 156, + 556, + 254 + ], + "type": "inline_equation", + "content": "24 \\times 24" + }, + { + "bbox": [ + 313, + 156, + 556, + 254 + ], + "type": "text", + "content": " patches yielding 576 ViT tokens, while for YTVIS21, a resolution of (518, 518) yields 1369 tokens. Full details are provided in App. A." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 266, + 556, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 556, + 411 + ], + "type": "text", + "content": "Baselines We compare SLOT CONTRAST against the previously proposed SAVi [28] and STEVE [45] that employ an image reconstruction objective and with the state-of-the-art method VideoSAUR [64] that uses self-supervised feature reconstructions. Additionally, for a fair comparison, we trained a modification of VideoSAUR with DINOv2 features (referred to as VideoSAURv2). The implementation details are provided in App. G. In addition, to assess how closely SLOT CONTRAST approaches supervised methods, we compared it with SAM2 [41] as a supervised zero-shot baseline for temporal consistency and to weakly-supervised by depth SAVi++ [11] method. The results are in App. G." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 424, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 556, + 628 + ], + "type": "text", + "content": "Temporally Consistent Object Discovery (Table 1 & Figure 4) SLOT CONTRAST significantly outperforms both SAVi and STEVE by a wide margin. When compared to VideoSAUR using its default parameters, our approach demonstrates higher consistency in terms of video FG-ARI scores. Compared to VideoSAUR and VideoSAURv2, SLOT CONTRAST achieves superior video scene decomposition (measured by FG-ARI scores). However, on synthetic datasets SLOT CONTRAST's masks are less sharp (as reflected by mBO). Notably, on the most challenging real-world YouTube-VIS data, our method surpasses both versions of VideoSAUR, achieving better performance on FG-ARI (+6.8) and mBO (+4). This shows that, given a large enough resolution and natural data inputs well aligned with DINOv2, SLOT CONTRAST can decompose unconstrained videos into consistent object representations. More examples are illustrated in App. M." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "content": "Per-Frame Scene Decomposition (Table 2) Previous research has shown the effectiveness of specific inductive biases and training objectives for unsupervised object discovery, such as reconstructing in semantic space or leveraging motion cues in video data. Building on this, we demonstrate that the contrastive nature of our temporal consistency loss" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5405" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 309, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 309, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 309, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 309, + 201 + ], + "type": "image", + "image_path": "e151a414bf71b285faae8a9dd142071c0fec9f5db2d1288c66e90ffe17ed3a4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with VideoSAURv2 on YouTube-VIS dataset. In challenging situations (e.g., almost full occlusions at " + }, + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "type": "inline_equation", + "content": "t = 24" + }, + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "type": "text", + "content": " of the 1st video and " + }, + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "type": "inline_equation", + "content": "t = 14" + }, + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "type": "text", + "content": " of the 2nd video), VideoSAURv2 reassigns slots to different objects (pink arrows), whereas SLOT CONTRAST consistently assigns slots to the same object (green arrows). Note that the colors of the masks are matched manually for better visual comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 553, + 201 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 553, + 201 + ], + "type": "image", + "image_path": "74c39b6206646b68ef77a94fb0899c1a459fe242b9d3256ea2009206d760cc27.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 65, + 316, + 285, + 472 + ], + "blocks": [ + { + "bbox": [ + 55, + 268, + 297, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 297, + 312 + ], + "type": "text", + "content": "Table 2. Quantitative Results on MOVi-E in terms of per-frame Image FG-ARI. The methods are grouped by the target data they train on: only images " + }, + { + "bbox": [ + 55, + 268, + 297, + 312 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 55, + 268, + 297, + 312 + ], + "type": "text", + "content": ", videos with motion segmentation annotations " + }, + { + "bbox": [ + 55, + 268, + 297, + 312 + ], + "type": "inline_equation", + "content": "(\\mathcal{V} + \\mathcal{M})" + }, + { + "bbox": [ + 55, + 268, + 297, + 312 + ], + "type": "text", + "content": ", and only videos " + }, + { + "bbox": [ + 55, + 268, + 297, + 312 + ], + "type": "inline_equation", + "content": "(\\mathcal{V})" + }, + { + "bbox": [ + 55, + 268, + 297, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 316, + 285, + 472 + ], + "lines": [ + { + "bbox": [ + 65, + 316, + 285, + 472 + ], + "spans": [ + { + "bbox": [ + 65, + 316, + 285, + 472 + ], + "type": "table", + "html": "
ModelObjectiveImage FG-ARI ↑
↓LSD [21]Image Rec.53.4
SlotDiffusion [57]Image Rec.60.0
DINOSAUR [44]Image Rec.65.1
+MoToK [3]+Mot. Seg.66.7
Safadoust et al. [42]+GT Flow78.3
DIOD [25]+Mot. Seg.82.2
↑STEVE [45]Video Rec.54.1
VideoSAUR [64]Temp. Sim.78.4
SOLV [2]Mid. Fr. Pred.80.8
SLOT CONTRASTSlot Contrast84.8
", + "image_path": "f53124d02a11d0e7c2372a18adb1827f274b6069b7d7fa98ec59f89c69d07343.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 567 + ], + "type": "text", + "content": "function yields improved scene decomposition as a byproduct. This occurs because our loss function encourages the model to learn consistent feature representations for objects across frames, leading to an adaptive process where dense features become more contrastive, thereby enhancing object discovery in individual frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "We compare our method with prior approaches in terms of per-frame object discovery, using the image FG-ARI metric for evaluation. Specifically, we compare three categories of methods: image-based, video-based, and methods that use videos with additional motion cues. Image-based methods use only images as a target (feature reconstruction based DINOSAUR [44] and diffusion-based LSD [21] and SlotDiffusion [57] methods). Video-based methods use only videos as targets: STEVE [45] reconstructs current frame features, SOLV [2] predicts middle frame features, and VideoSAUR [64] predicts temporal feature-similarities. Finally, we also compare with weakly-supervised methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 269, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 293 + ], + "type": "text", + "content": "using motion masks [3, 25] or ground truth (GT) optical flow [42]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 293, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 425 + ], + "type": "text", + "content": "The results on MOVi-E dataset are presented in Table 2, with comparisons across additional datasets provided in App. H. Using temporal signals from the video using feature reconstruction is better than object discovery based on images. Next, additional objectives that exploit the temporal structure of the videos allow even better scene decomposition. Notably, our method, which combines a feature reconstruction objective with a simple contrastive objective, leads to state-of-the-art performance reaching 84.8 per-frame FG-ARI, outperforming methods [3, 25] that use motion segmentation masks for object discovery." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 438, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 521 + ], + "type": "text", + "content": "Robustness to Full Occlusions To evaluate our method's robustness in handling complete object occlusions—a challenging scenario for maintaining consistency—we conduct experiments using a targeted subset of the MOVi-C dataset that contains sequences where objects are fully occluded. For evaluation, we retain only the ground-truth masks for the objects that experience occlusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "content": "We find that the feature reconstruction baseline achieves only " + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "content": " mBO vs. our method obtains " + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "content": " mBO on fully occluded objects. Our results suggest that SLOT CONTRAST significantly enhances consistency during object disappearances and reappearance. We refer the reader to Fig. 5 and App. K for visual examples and more details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 600, + 469, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 469, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 469, + 613 + ], + "type": "text", + "content": "4.2. Object Dynamics Prediction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Setup To evaluate performance on the task of predicting object dynamics, we train a dynamics module using the object-centric representations inferred by a pretrained object-centric model. For this dynamics module, we select SlotFormer [56], which predicts the slots autoregressively for " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " rollout steps based on the slots inferred from " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " burn-in frames preceding the prediction horizon. In our setup, we use 14 burn-in frames and 10 rollout steps. Since SlotFormer" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5406" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 295, + 254 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 295, + 254 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 295, + 254 + ], + "type": "image", + "image_path": "0424434c6e1046c3a168ce24376ddcef60f00988eea14894ed076e58b147d36b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 296, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 307 + ], + "type": "text", + "content": "Figure 5. Comparison of the Feature Reconstruction (Feat. Rec.) baseline, the slot-slot contrastive loss using only slots from the same video as the contrast set (Intra-video Contrast), and SLOT CONTRAST on the MOVi-C dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 328, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 413 + ], + "type": "text", + "content": "is trained independently from the object-centric model, we first train the latter, subsequently extending the datasets with the inferred slots for each frame. This approach avoids the computational complexity of training SlotFormer by removing the necessity to encode frames into the slot space at each training step. A brief introduction to SlotFormer and the implementation details can be found in App. J." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "text", + "content": "Baselines We compare SlotFormer [56] trained on object-centric representations derived from a model trained using only feature reconstruction loss with SlotFormer trained on representations from SLOT CONTRAST. Both models perform reconstruction in feature space rather than pixel space, so we use only the slot reconstruction loss for training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "text", + "content": "Quality of Predicted Masks (Table 3 and Figure 6) We note that our model has a significantly better FG-ARI on MOVi-C, while mBO is comparable to that of the baseline. On MOVi-E, the performance of our model is comparable to that of the baseline, which highlights the difficulty of adapting to videos with camera motion. There is also a" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 597, + 294, + 683 + ], + "blocks": [ + { + "bbox": [ + 59, + 597, + 294, + 683 + ], + "lines": [ + { + "bbox": [ + 59, + 597, + 294, + 683 + ], + "spans": [ + { + "bbox": [ + 59, + 597, + 294, + 683 + ], + "type": "image", + "image_path": "d714bd799befac09576f4ec27c3994f5ad8328c5b03369e96ed0c3ecf468df5d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "type": "text", + "content": "Figure 6. Object dynamics prediction task on MOVi-C using SLOT CONTRAST slots using SlotFormer [56]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 107, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 3. Downstream task of predicting object dynamics. Comparison of predictions made by SlotFormer based on representations obtained from SLOT CONTRAST and from Feature Reconstruction." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 107, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 159 + ], + "type": "table", + "html": "
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
Feat. Rec. + SF50.725.970.624.327.428.9
Ours + SF63.826.170.524.929.229.6
", + "image_path": "6aadca9f25b3b947f6d8aac5c8f638b6435c35649ce2d2e1a8487d1c3543cfe6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 183, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 232 + ], + "type": "text", + "content": "slight improvement in FG-ARI on YTVIS21, while mBO remains comparable. It is worth noting that predicting the motion in this dataset is especially challenging, given the large diversity of possible scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 245, + 376, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 376, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 376, + 258 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 264, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 324 + ], + "type": "text", + "content": "In this section, we investigate key components of our approach, including the impact of the contrastive loss and the type of slot initialization. In addition, we study how effective SLOT CONTRAST is in automatically shutting down slots in correspondence to the scene's complexity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 350, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 542 + ], + "type": "text", + "content": "Ablation of Loss Components (Table 4 and Figure 5) To demonstrate the value of the proposed slot-slot contrastive loss, we carry out an ablation study, comparing it with the feature reconstruction loss [44] and the intra-video contrastive loss, which contrasts slot representations in a single video. Using the intra-video contrastive loss yields improvements over the feature reconstruction baseline (+5.1 FG-ARI and +1.5 mBO on MOVi-C). However, we observe that in more challenging situations, the intra-video contrastive loss leads to failure cases such as shutting down too many slots (see Fig. 5). Next, we observe that by extending the contrast to the full batch of videos, SLOT CONTRAST learns more consistent representations (+19.6 FG-ARI and +5.3 mBO). This change increases the difficulty of the learning task, which prevents the model from relying on superficial patterns like slot initializations or object positions." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 563, + 541, + 685 + ], + "blocks": [ + { + "bbox": [ + 328, + 563, + 541, + 685 + ], + "lines": [ + { + "bbox": [ + 328, + 563, + 541, + 685 + ], + "spans": [ + { + "bbox": [ + 328, + 563, + 541, + 685 + ], + "type": "image", + "image_path": "bcfc7f1984f0184801dcbd6e212f98ba26e9088533413446540741c602ae7171.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 688, + 555, + 710 + ], + "lines": [ + { + "bbox": [ + 313, + 688, + 555, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 555, + 710 + ], + "type": "text", + "content": "Figure 7. Distribution of ground truth and predicted object numbers (i.e., number of active slots) on the MOVi-C dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5407" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 86, + 523, + 159 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 515, + 82 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 515, + 82 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 515, + 82 + ], + "type": "text", + "content": "Table 4. Ablation of loss components used by SLOT CONTRAST on MOVi-C, MOVi-E, and YouTube-VIS Datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 86, + 523, + 159 + ], + "lines": [ + { + "bbox": [ + 88, + 86, + 523, + 159 + ], + "spans": [ + { + "bbox": [ + 88, + 86, + 523, + 159 + ], + "type": "table", + "html": "
Feat. Rec.Intra-video ContrastSlot-Slot ContrastMOVi-CMOVi-EYouTube-VIS
LrecLintraLsscFG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
✓49.727.479.828.435.331.4
✓✓54.828.978.729.135.733.6
✓✓69.332.782.929.238.033.7
", + "image_path": "a39ec994f868ce900bd0e600943b2b8b6bf1cc57102373ac55cb73113c416166.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 194, + 294, + 259 + ], + "blocks": [ + { + "bbox": [ + 55, + 167, + 295, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 295, + 191 + ], + "type": "text", + "content": "Table 5. Comparison of the random initialization (RI) and learned initialization (LI) techniques." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 194, + 294, + 259 + ], + "lines": [ + { + "bbox": [ + 57, + 194, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 57, + 194, + 294, + 259 + ], + "type": "table", + "html": "
MOVi-CMOVi-EYouTube-VIS
FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑FG-ARI ↑mBO ↑
Feat. Rec. (RI)45.327.271.128.335.230.2
Feat. Rec. (LI)49.427.879.828.435.331.4
SLOT CONTRAST (RI)62.932.475.328.436.130.8
SLOT CONTRAST (LI)69.332.782.929.238.033.7
", + "image_path": "1cb9367eeb5115793ca5e0f2dcf441d6197d39a8894a39ad95f1bf05fdfb603a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 279, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 279, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 295, + 460 + ], + "type": "text", + "content": "Choice of the First Frame Initialization (Table 5) In video object-centric learning, slots are typically initialized based on those from the previous time step [28], while the first frame is initialized from learnable parameters. Previous real-world object-centric methods mostly used random initialization samples from Gaussian distribution [2, 64], in this work, we study the impact of the type of initialization under our contrastive objective. The findings are outlined in Table 5. Our experiments indicate that when combined with slot-slot contrastive loss, learned initialization significantly outperforms random initialization. We hypothesize that this improvement stems from the ability of learned initializations to shape the initial state in a way that enhances contrastiveness, a benefit not achievable with random initialization. For similarity visualizations and more details, see App. B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 474, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 295, + 677 + ], + "type": "text", + "content": "Number of Active Slots per Video (Figure 7) In models based on the Slot Attention mechanism, all available slots are typically utilized [44], leading to a mismatch between the predicted number of components in scene decomposition and the ground-truth number of objects in the scene. This can cause the random splitting of the objects between slots and non-consistent scene representations when slots are reassigned from one object to a part of another object. To address this challenge, it is important to study whether redundant slots can be effectively deactivated. Recently, AdaSlot [12] introduced a discrete slot sampling module, coupled with a complexity-aware prior, to penalize redundant slots explicitly. Similarly, SOLV [2] used agglomerative clustering to merge redundant slots. In this work, we investigate whether SLOT CONTRAST is capable of accurately determining the number of objects in a scene without relying on explicit priors to minimize the number of active slots." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We compare ground truth and predicted object density on MOVI-C dataset, as shown in Fig. 7. While the feature reconstruction model yields predictions within a narrow" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "text", + "content": "range—creating a sharp peak near a predefined number of slots—our model, similarly to AdaSlot [12], achieves a smoother prediction distribution that aligns more closely with the ground truth (note that the consistent shift is because 2–3 slots are used for the background, while the ground truth density is computed only for foreground objects). Interestingly, SLOT CONTRAST achieves this without requiring an explicit prior toward sparsity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 281, + 388, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 388, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 388, + 294 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 303, + 556, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 556, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 556, + 530 + ], + "type": "text", + "content": "SLOT CONTRAST advances unsupervised video object-centric learning by significantly improving the temporal consistency of object representations. Our method explicitly incentivizes temporal consistency by adding a self-supervised contrastive loss. We showed that this loss is not only beneficial for consistency, but also enhances object discovery: SLOT CONTRAST achieves state-of-the-art results on challenging synthetic datasets with many objects and the unconstrained real-world YouTube-VIS dataset. Furthermore, consistent representations directly support temporal downstream tasks such as unsupervised object dynamics prediction and allow for tracking of objects through full occlusions. Finally, SLOT CONTRAST effectively shuts down non-unique slots, leading to a sparser representation that captures the true object distribution more faithfully. Taken together, we expect these improvements to pave the way for broader adoption of video object-centric representations, for instance in applications like word modeling, autonomous control, or video question answering." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 532, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 582 + ], + "type": "text", + "content": "Limitations of our work include the fixed number of slots during initialization. Additionally, we cannot directly control the segmentation granularity of entities. Further limitations and failure cases are discussed in App. L." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Future work could explore several promising directions. First, one could use SLOT CONTRAST's robust and consistent representations for learning compositional world models from real-world robotics data to enable object-centric planning and control. Second, investigating the compatibility of our contrastive loss with other object-centric learning approaches with different inductive biases, such as SlotDiffusion [57]. Finally, improving the compactness of object masks [24] to achieve more precise object segmentation masks could also benefit unsupervised class-agnostic video object segmentation applications." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5408" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 224 + ], + "type": "text", + "content": "We thank Christian Gumbsch for useful suggestions and Ke Fan for providing additional details and results for AdaSlot method. Andrii Zadaianchuk is funded by the European Union (ERC, EVA, 950086). The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting Maximilian Seitzer. Georg Martius is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. This work was supported by the ERC - 101045454 REAL-RL. We acknowledge the support from the German Federal Ministry of Education and Research (BMBF) through the Tübingen AI Center (FKZ: 01IS18039B)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 244, + 115, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 244, + 115, + 257 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 115, + 257 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 265, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 265, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 265, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 265, + 296, + 308 + ], + "type": "text", + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In International Conference on Computer Vision (ICCV), 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 310, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 296, + 342 + ], + "type": "text", + "content": "[2] Görkay Aydemir, Weidi Xie, and Fatma Güney. Self-supervised object-centric learning for videos. In NeurIPS, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 343, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 296, + 397 + ], + "type": "text", + "content": "[3] Zhipeng Bao, Pavel Tokmakov, Yu-Xiong Wang, Adrien Gaidon, and Martial Hebert. Object discovery from motion-guided tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22972-22981, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 399, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 443 + ], + "type": "text", + "content": "[4] Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. MONet: Unsupervised Scene Decomposition and Representation. arXiv:1901.11390, 2019. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 487 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 489, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 296, + 521 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In ICML, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 523, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 567 + ], + "type": "text", + "content": "[7] Aniket Didolkar, Andrii Zadaianchuk, Anirudh Goyal, Mike Mozer, Yoshua Bengio, Georg Martius, and Maximilian Seitzer. Zero-shot object-centric representation learning. arXiv preprint arXiv:2408.09162, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 633 + ], + "type": "text", + "content": "[8] Aniket Rajiv Didolkar, Andrii Zadaianchuk, Rabiul Awal, Maximilian Seitzer, Efstratios Gavves, and Aishwarya Agrawal. Ctrl-o: Language-controllable object-centric visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "text", + "content": "[9] Andrea Dittadi, Samuele Papa, Michele De Vita, Bernhard Schölkopf, Ole Winther, and Francesco Locatello. Generalization and Robustness Implications in Object-Centric Learning. In ICML, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[10] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 335, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 127 + ], + "type": "text", + "content": "Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. PaLM-E: An embodied multimodal language model. In ICML, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "text", + "content": "[11] Gamaleldin Fathy Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael Curtis Mozer, and Thomas Kipf. " + }, + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "inline_equation", + "content": "\\mathsf{SAV}\\mathsf{i} + +" + }, + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "text", + "content": " : Towards End-to-End Object-Centric Learning from Real-World Videos. In NeurIPS, 2022. 1, 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "type": "text", + "content": "[12] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23062-23071, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 228, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 555, + 282 + ], + "type": "text", + "content": "[13] Locatello Francesco, Weissenborn Dirk, Unterthiner Thomas, Mahendran Aravindh, Heigold Georg, Uszkoreit Jakob, Dosovitskiy Alexey, and Kipf Thomas. Object-centric learning with slot attention. Advances in Neural Information Processing Systems, 33:11525-11538, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 265, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 283, + 555, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 555, + 337 + ], + "type": "text", + "content": "[14] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-Object Representation Learning with Iterative Variational Inference. In ICML, 2019. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 370 + ], + "type": "text", + "content": "[15] Klaus Greff, Sjoerd Van Steenkiste, and Jürgen Schmidhuber. On the Binding Problem in Artificial Neural Networks. arXiv:2012.05208, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 372, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 555, + 491 + ], + "type": "text", + "content": "[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A Scalable Dataset Generator. In CVPR, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 492, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 492, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 492, + 555, + 525 + ], + "type": "text", + "content": "[17] Christian Gumbsch, Martin V Butz, and Georg Martius. Sparsely changing latent states for prediction and planning in partially observable domains. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 525, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 555, + 557 + ], + "type": "text", + "content": "[18] Dan Haramati, Tal Daniel, and Aviv Tamar. Entity-centric reinforcement learning for object manipulation from pixels. In ICLR, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 559, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 555, + 591 + ], + "type": "text", + "content": "[19] Baoxiong Jia, Yu Liu, and Siyuan Huang. Improving object-centric learning with query optimization. arXiv preprint arXiv:2210.08990, 2022.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 592, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 554, + 624 + ], + "type": "text", + "content": "[20] Jindong Jiang, Sepehr Janghorbani, Gerard de Melo, and Sungjin Ahn. SCALOR: Generative World Models with Scalable Object Representations. In ICLR, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 555, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 647 + ], + "type": "text", + "content": "[21] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In NeurIPS, 2023. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 555, + 712 + ], + "type": "text", + "content": "[22] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5409" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "text", + "content": "[23] Rishabh Kabra, Daniel Zoran, Goker Erdogan, Loic Matthew, Antonia Creswell, Matthew Botvinick, Alexander Lerchner, and Christopher P. Burgess. SIMONE: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 194 + ], + "type": "text", + "content": "[24] Ioannis Kakogeorgiou, Spyros Gidaris, Konstantinos Karantzalos, and Nikos Komodakis. Spot: Self-training with patch-order permutation for object-centric learning with autoregressive transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22776-22786, 2024. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 295, + 248 + ], + "type": "text", + "content": "[25] Sandra Kara, Hejer Ammar, Julien Denize, Florian Chabot, and Quoc-Cuong Pham. Diod: Self-distillation meets object discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3975-3985, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 293 + ], + "type": "text", + "content": "[26] Laurynas Karazija, Iro Laina, and Christian Rupprecht. ClevrTex: A Texture-Rich Benchmark for Unsupervised Multi-Object Segmentation. In NeurIPS Track on Datasets and Benchmarks, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 326 + ], + "type": "text", + "content": "[27] Thomas Kipf, Elise Van der Pol, and Max Welling. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 380 + ], + "type": "text", + "content": "[28] Thomas Kipf, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional Object-centric Learning from Video. In ICLR, 2022. 1, 2, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 381, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 381, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 296, + 437 + ], + "type": "text", + "content": "[29] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 470 + ], + "type": "text", + "content": "[30] Adam Kosiorek, Hyunjik Kim, Yee Whye Teh, and Ingmar Posner. Sequential Attend, Infer, Repeat: Generative Modelling of Moving Objects. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 502 + ], + "type": "text", + "content": "[31] Jian Li, Pu Ren, Yang Liu, and Hao Sun. Reasoning-enhanced object-centric learning for videos. arXiv preprint arXiv:2403.15245, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 296, + 548 + ], + "type": "text", + "content": "[32] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-Centric Learning with Slot Attention. In NeurIPS, 2020. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "type": "text", + "content": "[33] Zijia Lu, Bing Shuai, Yanbei Chen, Zhenlin Xu, and Davide Modolo. Self-supervised multi-object tracking with path consistency. arXiv preprint, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "text", + "content": "[34] Amir Mohammad Karimi Mamaghan, Samuele Papa, Karl Henrik Johansson, Stefan Bauer, and Andrea Dittadi. Exploring the effectiveness of object-centric representations in visual question answering: Comparative insights with foundation models. arXiv preprint arXiv:2407.15589, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "type": "text", + "content": "[35] Sha Meng, Dian Shao, Jiacheng Guo, and Shan Gao. Tracking without label: Unsupervised multiple object tracking via contrastive similarity learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16264-16273, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "text", + "content": "[36] Cristian Meo, Akihiro Nakano, Mircea Lică, Aniket Didolkar, Masahiro Suzuki, Anirudh Goyal, Mengmi Zhang, Justin" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Dauwels, Yutaka Matsuo, and Yoshua Bengio. Object-centric temporal consistency via conditional autoregressive inductive biases. arXiv preprint arXiv:2410.15728, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 555, + 139 + ], + "type": "text", + "content": "[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 555, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 237 + ], + "type": "text", + "content": "[38] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 238, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 555, + 293 + ], + "type": "text", + "content": "[39] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6964-6974, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 293, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 555, + 326 + ], + "type": "text", + "content": "[40] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In ICCV, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 403 + ], + "type": "text", + "content": "[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 403, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 554, + 425 + ], + "type": "text", + "content": "[42] Sadra Safadoust and Fatma Güney. Multi-object discovery by low-dimensional object motion. In ICCV, 2023. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 555, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 555, + 480 + ], + "type": "text", + "content": "[43] Mehdi SM Sajjadi, Daniel Duckworth, Aravindh Mahendran, Sjoerd Van Steenkiste, Filip Pavetic, Mario Lucic, Leonidas J Guibas, Klaus Greff, and Thomas Kipf. Object scene representation transformer. Advances in Neural Information Processing Systems, 35:9512-9524, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 481, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 555, + 536 + ], + "type": "text", + "content": "[44] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, and Francesco Locatello. Bridging the gap to real-world object-centric learning. In ICLR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 536, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 555, + 568 + ], + "type": "text", + "content": "[45] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple Unsupervised Object-Centric Learning for Complex and Naturalistic Videos. In NeurIPS, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 569, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 555, + 624 + ], + "type": "text", + "content": "[46] Matthias Tangemann, Steffen Schneider, Julius Von Kügelgen, Francesco Locatello, Peter Gehler, Thomas Brox, Matthias Kümmerer, Matthias Bethge, and Bernhard Schölkopf. Unsupervised object learning via common fate. arXiv preprint arXiv:2110.06562, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "type": "text", + "content": "[47] Pavel Tokmakov, Allan Jabri, Jie Li, and Adrien Gaidon. Object permanence emerges in a random walk along memory. arXiv preprint arXiv:2204.01784, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 690 + ], + "type": "text", + "content": "[48] Manuel Traub, Frederic Becker, Sebastian Otte, and Martin V Butz. Looping loci: Developing object permanence from videos. arXiv preprint arXiv:2310.10372, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 690, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 555, + 714 + ], + "type": "text", + "content": "[49] Manuel Traub, Sebastian Otte, Tobias Menge, Matthias Karlbauer, Jannik Thuemmel, and Martin V. Butz. Learning What" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5410" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "and Where: Disentangling Location and Identity Tracking Without Supervision. In ICLR, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 297, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 297, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 297, + 150 + ], + "type": "text", + "content": "[50] Basile Van Hoorick, Pavel Tokmakov, Simon Stent, Jie Li, and Carl Vondrick. Tracking through containers and occluders in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13802-13812, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 195 + ], + "type": "text", + "content": "[51] Sjoerd van Steenkiste, Michael Chang, Klaus Greff, and Jürgen Schmidhuber. Relational neural expectation maximization: Unsupervised discovery of objects and their interactions. In ICLR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "text", + "content": "[52] Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity Abstraction in Visual Model-based Reinforcement Learning. In Conference on Robot Learning, 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 296, + 297 + ], + "type": "text", + "content": "[53] Ning Wang, Yibing Song, Chao Ma, Wengang Zhou, Wei Liu, and Houqiang Li. Unsupervised deep tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1308-1317, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 342 + ], + "type": "text", + "content": "[54] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2566-2576, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 386 + ], + "type": "text", + "content": "[55] Nicholas Watters, Loic Matthew, Christopher P Burgess, and Alexander Lerchner. Spatial broadcast decoder: A simple architecture for learning disentangled representations in vaes. arXiv preprint arXiv:1901.07017, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 430 + ], + "type": "text", + "content": "[56] Ziyi Wu, Nikita Dvornik, Klaus Greff, Thomas Kipf, and Animesh Garg. Slotformer: Unsupervised visual dynamics simulation with object-centric models. In ICLR, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 296, + 464 + ], + "type": "text", + "content": "[57] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotdiffusion: Object-centric generative modeling with diffusion models. In NeurIPS, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 296, + 498 + ], + "type": "text", + "content": "[58] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Slot-VLM: Object-event slots for video-language modeling. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 500, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 521 + ], + "type": "text", + "content": "[59] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 296, + 555 + ], + "type": "text", + "content": "[60] Linjie Yang, Yuchen Fan, Yang Fu, and Ning Xu. The 3rd large-scale video object segmentation challenge - video instance segmentation track, 2021. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 557, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 590 + ], + "type": "text", + "content": "[61] Jaesik Yoon, Yi-Fu Wu, Heechul Bae, and Sungjin Ahn. An investigation into pre-training object-centric representations for reinforcement learning. In ICML, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 296, + 624 + ], + "type": "text", + "content": "[62] Haonan Yu and Wei Xu. Vonet: Unsupervised video object learning with parallel u-net attention and object-wise sequentialvae.arXiv preprint arXiv:2401.11110, 2024.2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 657 + ], + "type": "text", + "content": "[63] Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius. Self-supervised Visual Reinforcement Learning with Object-centric Representations. In ICLR, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[64] Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius. Object-centric learning for real-world videos by predicting temporal feature similarities. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023), 2023. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 172 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[65] Zixu Zhao, Jiaze Wang, Max Horn, Yizhuo Ding, Tong He, Zechen Bai, Dominik Zietlow, Carl-Johann Simon-Gabriel, Bing Shuai, Zhuowen Tu, Thomas Brox, Bernt Schiele, Yanwei Fu, Francesco Locatello, Zheng Zhang, and Tianjun Xiao. Object-centric multiple object tracking. ICCV, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "text", + "content": "[66] Daniel Zoran, Rishabh Kabra, Alexander Lerchner, and Danilo J. Rezende. PARTS: Unsupervised segmentation with slots, attention and independence maximization. In ICCV, 2021. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5411" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_content_list.json b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..206e7863b1931dd29e225dfe3dfe71958da8938e --- /dev/null +++ b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_content_list.json @@ -0,0 +1,1772 @@ +[ + { + "type": "text", + "text": "TensoFlow: Tensorial Flow-based Sampler for Inverse Rendering", + "text_level": 1, + "bbox": [ + 169, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chun Gu $^{1,2}$ Xiaofei Wei $^{1,2}$ Li Zhang $^{1,2*}$ Xiatian Zhu $^{3}$", + "bbox": [ + 261, + 179, + 738, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Data Science, Fudan University $^{2}$ Shanghai Innovation Institute $^{3}$ University of Surrey", + "bbox": [ + 99, + 198, + 897, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/fudan-zvg/tensoflow", + "bbox": [ + 271, + 224, + 723, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 276, + 326, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inverse rendering aims to recover scene geometry, material properties, and lighting from multi-view images. Given the complexity of light-surface interactions, importance sampling is essential for the evaluation of the rendering equation, as it reduces variance and enhances the efficiency of Monte Carlo sampling. Existing inverse rendering methods typically use pre-defined non-learnable importance samplers in prior manually, struggling to effectively match the spatially and directionally varied integrand and resulting in high variance and suboptimal performance. To address this limitation, we propose the concept of learning a spatially and directionally aware importance sampler for the rendering equation to accurately and flexibly capture the unconstrained complexity of a typical scene. We further formulate TensoFlow, a generic approach for sampler learning in inverse rendering, enabling to closely match the integrand of the rendering equation spatially and directionally. Concretely, our sampler is parameterized by normalizing flows, allowing both directional sampling of incident light and probability density function (PDF) inference. To capture the characteristics of the sampler spatially, we learn a tensorial representation of the scene space, which imposes spatial conditions, together with reflected direction, leading to spatially and directionally aware sampling distributions. Our model can be optimized by minimizing the difference between the integrand and our normalizing flow. Extensive experiments validate the superiority of TensoFlow over prior alternatives on both synthetic and real-world benchmarks.", + "bbox": [ + 89, + 309, + 485, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 775, + 220, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene modeling has made significant strides, especially with the emergence of powerful representation models like neural radiance fields (NeRF) [2-4, 24] and efficient ones like 3D Gaussian splatting (3DGS) [13, 17, 38]. However, these methods struggle with inverse rendering due to the in", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ability to effectively disentangle geometry, material properties, and lighting, restricting their capacity to relight scenes.", + "bbox": [ + 511, + 277, + 903, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this issue, recent works [10, 11, 14, 15, 20, 22, 23, 35, 36] have advanced appearance modeling by incorporating the physically-based rendering equation [16] that simulates light-surface interactions. This however introduces a fundamental challenge - a need for accurately evaluating the rendering equation using Monte Carlo sampling. One intuitive method is stratified sampling over the hemisphere [15, 35], but this can be computationally inefficient due to the need for many samples. To improve efficiency, leading methods like NeRO [23] and TensoSDF [20] employ importance sampling. These approaches utilize pre-defined importance samplers (e.g., cosine-weighted and GGX samplers [30] for diffuse and specular components, respectively). However, these pre-defined importance samplers still result in high variance, as the integrand typically varies in both space and direction, making it difficult for a fixed sampler to effectively match the scene's characteristics.", + "bbox": [ + 511, + 308, + 906, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this challenge, we propose a novel perspective of learning spatially and directionally aware importance samplers, allowing us to manage the complexity of any scene without the need for manual sampler selection, while benefiting from greater flexibility and generality. As a showcase, we formulate TensoFlow, a concrete approach for importance sampler learning. Our sampler is parameterized by normalizing flows composed with piecewise-quadratic coupling layers [25], functioning as a neural distribution that enables both probability density function (PDF) inference and directional sampling of incident light in the rendering equation. To effectively capture the characteristics of the sampler spatially, we learn a tensorial representation of the scene space, imposing spatial conditions on the rendering process, together with reflected direction, resulting in spatially and directionally aware sampling distributions. This formulation can be optimized by minimizing the distributional difference between the integrand and our normalizing flow.", + "bbox": [ + 511, + 582, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We make these contributions: (i) We propose a novel concept of learning the importance samplers of the render-", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author (lizhangfd@fudan.edu.cn).", + "bbox": [ + 107, + 886, + 419, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "495", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing equation for inverse rendering, in contrast to existing pre-defined non-learnable samplers struggling intrinsically to capture the intrinsic rendering complexity of a scene. (ii) We introduce an effective method for sampler learning, TensoFlow, parameterized by tensorial normalizing flows, serving as a neural distribution that can support both directional sampling of incident light and PDF inference required in the rendering equation. (iii) Extensive experiments on both synthetic and real-world datasets demonstrate that TensoFlow outperforms previous alternatives, achieving significantly lower variance in evaluating the rendering equation with the same number of sampled directions.", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 284, + 228, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel view synthesis The introduction of neural radiance fields (NeRF) [24], which represent 3D scenes as continuous volumetric functions, utilizes volumetric rendering techniques to synthesize images from novel viewpoints, achieving effective 3D scene reconstruction. However, NeRF exhibits limitations in terms of rendering efficiency. Various efficient alternatives have been proposed, including voxel-based methods [9, 27], multi-resolution hash grids [26], and tensor decomposition approaches [6]. Additionally, techniques such as anti-aliasing [2-4] and modeling with signed distance fields (SDFs) [21, 31, 32, 34, 37] have been employed to enhance rendering and geometry reconstruction quality. Recently, 3D Gaussian splatting (3DGS) [17] has been introduced with favored efficiency, representing 3D scenes as a collection of 3D Gaussians. However, despite its impressive performance in rendering quality and efficiency, 3DGS lacks robust geometry, which limits its effectiveness in geometry-sensitive tasks such as inverse rendering [10, 22].", + "bbox": [ + 89, + 309, + 483, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inverse rendering Inverse rendering aims to recover the 3D structure, lighting conditions, and material properties of a scene from posed 2D images. Recent advancements in neural rendering have substantially driven inverse rendering performance. NeRF-based approaches [1, 5, 15, 18, 20, 23, 35, 39, 41] simulate physical light-surface interactions, allowing for the disentanglement of scene geometry, materials, and lighting from image observations. In particular, Monte Carlo sampling is commonly used to evaluate the rendering equation in inverse rendering. For example, avoiding ray marching on incident rays, NeILF [35] optimizes an incident light field using stratified sampling (e.g., sampling directions uniformly over the hemisphere) for the rendering equation. Similarly, TensoIR [15] also applies stratified sampling during training using tensorial implicit fields. However, stratified sampling can be computationally costly due to the high sample count needed for accuracy. To address this issue, NeRO [23] and TensoSDF [20] adopt importance sampling with pre-defined samplers (e.g., cosine", + "bbox": [ + 89, + 613, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "weighted, GGX). Commonly, all these sampling methods fall short of matching the spatially and directionally varying integrand of the rendering equation, leading to suboptimal rendering performance. To overcome this limitation, we introduce the concept of learning importance samplers that addressing the need for spatially and directionally aware samplers elegantly.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 210, + 638, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin with necessary background on importance sampling and the rendering equation.", + "bbox": [ + 511, + 234, + 903, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Importance sampling Monte Carlo sampling is a commonly used method for evaluating integrals [12]. The principal challenge lies in approximating the integral with fewer samples while achieving reduced variance. Given an integral $F = \\int_{\\Omega} f(x) \\, \\mathrm{d}x$ , where $f(x)$ is the target integrand function over the domain $\\Omega$ , and $x \\in \\mathbb{R}^D$ is a sample, importance sampling is performed by:", + "bbox": [ + 511, + 266, + 906, + 372 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nF = \\int_ {\\Omega} \\frac {f (x)}{q (x)} q (x) \\mathrm {d} x = \\mathbb {E} \\left[ \\frac {f (X)}{q (X)} \\right] \\approx \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {f \\left(X _ {i}\\right)}{q \\left(X _ {i}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 383, + 903, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the proxy PDF $q(x)$ is designed to approximate the distribution of $f(x)$ in a way that minimizes the variance of the estimator. Ideally, $q(x)$ should be chosen so that it matches the shape of $f(x)$ closely, meaning that areas where $f(x)$ has higher values are sampled more frequently. This helps in achieving a more accurate estimate of the integral with fewer samples.", + "bbox": [ + 511, + 436, + 905, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "**Rendering equation** The rendering equation [16] models the interaction between lighting and surfaces. Given the incident radiance $L_{i}(\\omega_{i}, x)$ and the material properties at a surface point $x$ with normal vector $n$ , the rendering equation expresses the outgoing radiance in the following manner:", + "bbox": [ + 511, + 542, + 906, + 632 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\int_ {\\Omega} f \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {i} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}\\right) d \\boldsymbol {\\omega} _ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 642, + 903, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $f(\\omega_o, \\omega_i, x)$ is the Bidirectional Reflectance Distribution Function (BRDF). Here, we consider the microfacet BRDF parameterization [28], including material properties such as albedo $a \\in [0,1]^3$ , metallic $m \\in [0,1]$ , and roughness $r \\in [0,1]$ . Formally, this BRDF can be decomposed into a diffuse term $f_d$ and a specular term $f_s$ defined as:", + "bbox": [ + 511, + 684, + 905, + 776 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf _ {d} = \\frac {(1 - m) \\boldsymbol {a}}{\\pi}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 785, + 906, + 818 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\frac {D F G}{4 \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n}\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 832, + 903, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $D, F$ , and $G$ represent the normal distribution function, the Fresnel term, and the geometry term, respectively.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "496", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e2067f8e11fc6c70662adc07b2f5fe6806dcb0c852ab46122c0930417777148f.jpg", + "image_caption": [ + "Figure 1. Left: Material and lighting estimation stage of TensoFlow. Given a ray-surface intersection point $\\pmb{x}$ , we use tensorial encoders to encode latent features for both the material properties and the importance sampler. When evaluating the rendering equation, incident directions are sampled from our learnable importance sampler, which is a frozen copy of the training normalizing flow. The normalizing flow is optimized by minimizing the distribution difference $\\mathcal{L}_{\\mathrm{ce}}$ between $q(\\omega_{i})$ and the integrand. The material properties, parameterized by $\\theta_{m}$ , are optimized by minimizing the RGB rendering loss $\\mathcal{L}_{\\mathrm{c}}$ of the final rendering integral. Right: Tensorial normalizing flow. With spatial prior $V_{f}$ and directional prior $\\omega_{r}$ , our tensorial normalizing flow, implemented using piecewise-quadratic coupling layers, enables both incident direction sampling and PDF querying." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 632, + 296 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fb272017abde4db394c69d84fe83972c2220f11c5535bc9a6497a873a09a6d3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 90, + 905, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. TensoFLow", + "text_level": 1, + "bbox": [ + 89, + 436, + 212, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture overview Our TensoFlow is comprised of two stages: (I) The first stage involves reconstructing the scene geometry. We follow TensoSDF [20] and utilize a tensorial signed distance field combined with a roughness-aware training strategy to achieve robust geometry reconstruction. (II) The second stage focuses on modeling the material and lighting, which is central to inverse rendering and constitutes the core of our approach, as detailed below.", + "bbox": [ + 89, + 463, + 483, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Material and lighting estimation", + "text_level": 1, + "bbox": [ + 89, + 595, + 375, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parametrization In TensoFlow, we represent the material of a scene with a Vector-Matrix tensorial encoder [6], due to its high capacity and flexibility. For a surface point $\\pmb{x}$ , its tensorial feature vector is defined as", + "bbox": [ + 89, + 619, + 483, + 679 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {m} (\\boldsymbol {x}) = \\mathrm {v} _ {m, k} ^ {X} \\circ \\mathrm {M} _ {m, k} ^ {Y Z} \\oplus \\mathrm {v} _ {m, k} ^ {Y} \\circ \\mathrm {M} _ {m, k} ^ {X Z} \\oplus \\mathrm {v} _ {m, k} ^ {Z} \\circ \\mathrm {M} _ {m, k} ^ {X Y}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 690, + 482, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{v}_k$ and $\\mathrm{M}_k$ represent the $k$ -th vector and matrix along the corresponding axis. The operators $\\circ$ and $\\oplus$ denote element-wise multiplication and concatenation, respectively. After concatenating this feature vector with the positional encoding $\\mathbf{p}$ of the position $\\boldsymbol{x}$ , it is processed by a tiny MLP $\\Theta_m$ to produce the material properties:", + "bbox": [ + 89, + 722, + 483, + 813 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {a}, m, r \\right\\} = \\Theta_ {m} \\left(V _ {m}, \\mathbf {p}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 825, + 482, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{a}$ , $m$ , and $r$ denote albedo, metallic, and roughness, respectively. For the incident light modeling, we decompose it into spatially-aware indirect lighting and direct light-", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing originating from infinity:", + "bbox": [ + 511, + 438, + 705, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) = V \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {\\operatorname {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right) + L _ {\\operatorname {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 462, + 903, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the visibility term $V(\\omega_{i},\\pmb {x})$ is determined via ray tracing on the extracted mesh. The indirect and direct lightings are modeled by a tiny MLP $\\Theta_{\\mathrm{ind}}$ and an environment map $\\Theta_{\\mathrm{dir}}$ , respectively:", + "bbox": [ + 511, + 486, + 905, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) = \\Theta_ {\\text {i n d}} \\left(\\mathbf {p}, \\boldsymbol {\\omega} _ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 555, + 903, + 571 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right) = \\Theta_ {\\mathrm {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 574, + 903, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Rendering** Given the reconstructed signed distance field from the first stage, the intersection point between each ray and the surface can be determined. The material properties and incident lighting at this intersection can then be queried as described earlier. In accordance with Eq. (2), we employ Monte Carlo sampling to evaluate the integral in the rendering equation. Due to differing characteristics of the diffuse and specular terms, separate importance samplers are used to minimize variance in the Monte Carlo sampling process:", + "bbox": [ + 511, + 606, + 906, + 743 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} _ {\\text {d i f f u s e}} = \\frac {1}{N _ {d}} \\sum_ {i} ^ {N _ {d}} \\frac {(1 - m) \\frac {a}{\\pi} L (\\boldsymbol {\\omega} _ {i} , \\boldsymbol {x}) (\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n})}{q _ {\\mathrm {d}} (\\boldsymbol {\\omega} _ {i} ; \\boldsymbol {x})}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 762, + 903, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} _ {\\text {s p e c u l a r}} = \\frac {1}{N _ {s}} \\sum_ {i} ^ {N _ {s}} \\frac {\\frac {D F G}{4 (\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}) (\\boldsymbol {\\omega} _ {o} \\cdot \\boldsymbol {n})} L (\\boldsymbol {\\omega} _ {i} , \\boldsymbol {x}) (\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n})}{q _ {\\mathrm {s}} (\\boldsymbol {\\omega} _ {i} ; \\boldsymbol {x})}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 806, + 903, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q_{\\mathrm{d}}(\\omega_i)$ and $q_{\\mathrm{s}}(\\omega_i)$ are the PDFs of the importance samplers, which we model them using the proposed normalizing flow (Sec. 4.2). The final color is computed as:", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "497", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6da1d6c84fa578c32df647b32cd8dd88c6ca0f944dc05c8cae55bf4cf9cd831.jpg", + "image_caption": [ + "Figure 2. Illustration of a piecewise-quadratic coupling layer, incorporating tensorial latent feature $V_{f}$ and reflected direction $\\omega_{r}$ as spatial and directional priors." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 475, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} = \\boldsymbol {c} _ {\\text {d i f f u s e}} + \\boldsymbol {c} _ {\\text {s p e c u l a r}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 386, + 482, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Tensorial flow-based importance sampling", + "text_level": 1, + "bbox": [ + 89, + 411, + 452, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For evaluating the rendering equation, prior works often employ pre-defined distributions as the importance samplers. For instance, in NeRO [23] a cosine-weighted distribution is used as the importance sampler for the diffuse term, and the GGX distribution [7] for the specular term. However, such samplers cannot closely match the shape of the target integrand with complex spatial and directional variations for a typical scene, leading to high variance and performance degradation.", + "bbox": [ + 89, + 434, + 483, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Flow-based sampler", + "text_level": 1, + "bbox": [ + 89, + 590, + 284, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome this limitation, we propose learning the importance samplers. The normalizing flow framework [8] is adopted as the sampler learner due to its following theoretical advantage. For a given direction $\\omega$ with two degrees of freedom in spherical coordinates, the distribution of the integrand function may exhibit inter-dependencies across different coordinate dimensions. We can express the variable $\\omega$ as a bijective transformation of an independent variable $z$ , which follows a uniform distribution $(0,1)^{2}$ : $\\omega = h(z)$ . Using the change-of-variable formula, the PDF of $\\omega$ can be expressed as:", + "bbox": [ + 89, + 614, + 483, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq (\\boldsymbol {\\omega}) = q (\\boldsymbol {z}) \\left| \\det \\left(\\frac {\\partial h (\\boldsymbol {z})}{\\partial \\boldsymbol {z} ^ {T}}\\right) \\right| ^ {- 1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 791, + 482, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q(z)$ represents the PDF of $z$ in $\\mathcal{U}(0,1)^2$ , which is straightforward to evaluate.", + "bbox": [ + 89, + 839, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the selection of the bijective mapping $h$ must satisfy certain conditions: it must be invertible, and it", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "should allow efficient computation of the Jacobian determinant. To meet these requirements, we employ NICE [8], a type of autoregressive flow that utilizes \"coupling layers\" to enable fast sampling and density evaluation. The final mapping $h$ is then formed as a composition of simpler bijective transformations: $h = h_{n} \\circ \\dots \\circ h_{1}$ .", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input vector $x = (x_{1},x_{2})\\in \\mathbb{R}^{2}$ for a mapping $h_i$ , the output vector $y = (y_{1},y_{2})\\in \\mathbb{R}^{2}$ is defined as:", + "bbox": [ + 511, + 181, + 903, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {1} = x _ {1}, \\quad y _ {2} = C (x _ {2}; m (x _ {1})), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 222, + 903, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the coupling transform $C:\\mathbb{R}\\leftrightarrow \\mathbb{R}$ is a separable and invertible transformation, and $m(x)$ , parameterized by a neural network, acts as the parameter of $C$ . This structure allows the inverse of the coupling layer to be computed directly:", + "bbox": [ + 511, + 250, + 905, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {1} = y _ {1}, \\quad x _ {2} = C ^ {- 1} \\left(y _ {2}; m \\left(y _ {1}\\right)\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 335, + 903, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The separability of $C$ ensures that the Jacobian matrix $J_{i}$ of $h_{i}$ is triangular, simplifying the computation of its determinant to the product of its diagonal elements.", + "bbox": [ + 511, + 362, + 905, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By applying the change-of-variable formula, the Jacobian determinant of the composite mapping $h$ is expressed as the product of the individual $J_{i}$ : $|\\operatorname{det} J| = \\prod_{i} |\\operatorname{det} J_{i}|$ . We note that, with sufficiently expressive coupling transformations and adequate coupling layers, the sampling distribution $q(x)$ can be modeled as an arbitrarily complex function, allowing it to closely match the shape of any integrand in the rendering equation. Importantly, each adjacent coupling layer preserves different dimensions of the input vector (Eq. (14)), ensuring the mapping's expressiveness across all the dimensions.", + "bbox": [ + 511, + 407, + 905, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The normalizing flow defined by the above mapping acts as a distribution from which we can query the PDF or draw samples, corresponding to $h^{-1}$ and $h$ , respectively. Our method supports two primary applications in the inverse rendering pipeline: inferring the PDF value for a given direction and sampling a direction along with its PDF value:", + "bbox": [ + 511, + 574, + 905, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference: $q(\\omega)\\gets h^{-1}(\\omega),$ (16)", + "bbox": [ + 539, + 690, + 903, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sampling: $(\\omega ,q(\\omega))\\gets h(\\pmb {u}),\\pmb {u}\\sim \\mathcal{U}(0,1)^2,$ (17)", + "bbox": [ + 539, + 710, + 903, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our experiments reveal that modeling the distribution of the half vector $\\omega_{h}$ yields better performance compared to directly modeling $\\omega_{i}$ . By modeling the distribution $\\omega_{h}$ , $q(\\omega_{i})$ can be derived using the change-of-variable formula (Eq. (13)), enabling both PDF inference and sampling of $\\omega_{i}$ .", + "bbox": [ + 511, + 737, + 905, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.2 Tensorial coupling transformation", + "text_level": 1, + "bbox": [ + 511, + 845, + 805, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the spatially-dependent nature of the integrand in the rendering equation, we further parameterize the scene space", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "498", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by incorporating an additional tensorial representation as:", + "bbox": [ + 89, + 90, + 472, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {f} (\\boldsymbol {x}) = \\mathrm {v} _ {f, k} ^ {X} \\circ \\mathrm {M} _ {f, k} ^ {Y Z} \\oplus \\mathrm {v} _ {f, k} ^ {Y} \\circ \\mathrm {M} _ {f, k} ^ {X Z} \\oplus \\mathrm {v} _ {f, k} ^ {Z} \\circ \\mathrm {M} _ {f, k} ^ {X Y}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 116, + 482, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The latent feature $V_{f}$ , together with the reflected direction $\\omega_{r} = 2(\\omega_{o}\\cdot n)n - \\omega_{o}$ for directional dependency, serves as conditions to guide the network $m_{i}$ within each coupling layer. This is crucial, as the shape of the integrand in the rendering equation is influenced by both the spatial position and the outgoing direction $\\omega_{o}$ .", + "bbox": [ + 89, + 148, + 483, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider the piecewise-quadratic coupling transform [25] as our coupling transformation $C$ for its high expressiveness. Specifically, it achieves the bijective property by defining a Cumulative Distribution Function (CDF): $y = C(x; m(x'; V_f, \\omega_r)) = \\int_0^x p(t; m(x'; V_f, \\omega_r)) \\, \\mathrm{d}t$ , where $p$ is a spatially and directionally aware PDF derived from the output of network $m$ . This results in the Jacobian determinant of the coupling layer being equal to $p(x)$ . Here, the network $m$ takes the concatenation of $x', V_f$ , and $\\omega_r$ as input, generating $\\widehat{V}$ and $\\widehat{W}$ for subsequent use. We then construct the piecewise linear PDF using $K + 1$ vertices, as illustrated in Fig. 2. It employs $K + 1$ vertical coordinates $\\widehat{V} \\in \\mathbb{R}^{K + 1}$ and $K$ horizontal differences $\\widehat{W} \\in \\mathbb{R}^K$ , with $K$ as the predefined bin count for the distribution. Following softmax normalization, $W = \\sigma(\\widehat{W})$ represents the bin widths, while $V$ defines the PDF values at each vertex.", + "bbox": [ + 89, + 239, + 483, + 486 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i} = \\frac {\\exp (\\widehat {V} _ {k})}{\\sum_ {k = 1} ^ {K} \\frac {\\exp (\\widehat {V} _ {k}) + \\exp (\\widehat {V} _ {k + 1})}{2} W _ {k}}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 497, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consequently, given an input $x$ , the corresponding PDF value can be queried as", + "bbox": [ + 89, + 558, + 483, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (x) = \\operatorname {l e r p} \\left(V _ {b}, V _ {b + 1}, \\alpha\\right), \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 602, + 480, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha = (x - \\sum_{k=1}^{b-1} W_k) / W_k$ defines the relative position of $x$ within the corresponding bin $b$ . The piecewise-quadratic CDF can be computed by integrating the PDF:", + "bbox": [ + 89, + 631, + 483, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC (x) = \\frac {\\alpha W _ {b}}{2} ((2 - \\alpha) V _ {b} + \\alpha V _ {b + 1}) + \\sum_ {k = 1} ^ {b - 1} \\frac {V _ {k} + V _ {k + 1}}{2} W _ {k}. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 691, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ $N$ coupling layers with piecewise-quadratic transformation to construct the compound mapping $h$ , ensuring it is sufficiently expressive for modeling the importance sampler.", + "bbox": [ + 89, + 744, + 483, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training", + "text_level": 1, + "bbox": [ + 89, + 816, + 194, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During model training, we optimize the tensorial encoder $V_{f}$ and the networks $\\{m_i\\}_{i = 1\\dots N}$ by minimizing the cross-entropy loss between the integrand of the rendering equation and the PDF of the incident directions inferred from", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the normalizing flow:", + "bbox": [ + 513, + 90, + 658, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {c e}} = - \\int_ {\\Omega} I (\\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}) \\log q (\\boldsymbol {\\omega} _ {i}) \\mathrm {d} \\boldsymbol {\\omega} _ {i} (22) \\\\ = \\mathbb {E} \\left[ - \\frac {I \\left(\\boldsymbol {\\omega} _ {i} , \\boldsymbol {\\omega} _ {o} , \\boldsymbol {x}\\right)}{\\hat {q} \\left(\\boldsymbol {\\omega} _ {i}\\right)} \\log q \\left(\\boldsymbol {\\omega} _ {i}\\right) \\right], (23) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 114, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{q} (\\omega_{i})$ represents the PDF of the sampled incident directions from the current importance sampler, a frozen copy of the trainable normalizing flow updated every $N_{\\mathrm{update}}$ iterations. $q(\\omega_{i})$ denotes the PDF inferred from the current training normalizing flow, and $I(\\omega_i,\\omega_o,x)$ is the integrand function evaluated with the sampled incident direction. We use different importance samplers for the diffuse and specular terms, resulting in separate losses: $\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{d}}$ and $\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{s}}$ . The total training loss is thus defined as:", + "bbox": [ + 511, + 191, + 906, + 327 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c} + \\lambda_ {m} \\mathcal {L} _ {m} + \\lambda_ {\\mathrm {c e}} ^ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {d}} + \\lambda_ {\\mathrm {c e}} ^ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {s}}, \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 335, + 903, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_c$ and $\\mathcal{L}_m$ represent the RGB rendering loss and the material regularization loss [23], respectively.", + "bbox": [ + 511, + 363, + 903, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 513, + 407, + 637, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and metrics We use the TensoSDF synthetic dataset [20] and the real-world Stanford-ORB dataset [19] for quantitative evaluation. To quantitatively evaluate the estimated material, we adopt three standard metrics for relighted images: PSNR, SSIM [33], and LPIPS [40].", + "bbox": [ + 511, + 431, + 903, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Competitors We compare our approach with state-of-the-art NeRF-based methods selected for their different sampling strategies in evaluating the rendering equation: TensoIR [15] and NeILF++[39], which utilize stratified sampling, and NeRO[23] and TensoSDF [20], which employ pre-defined importance sampling.", + "bbox": [ + 511, + 508, + 903, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details The training process consists of two stages. For the first stage, geometry reconstruction, we train the TensoSDF model for 180,000 iterations using the same settings as in TensoSDF [20]. For the second stage, material and lighting estimation, we begin with a warm-up phase, employing cosine-weighted and GGX distributions as the importance samplers for the diffuse and specular components, respectively, over $N_{\\mathrm{warmup}} = 1,000$ iterations. Once a coarse estimation of material and lighting is achieved at $N_{\\mathrm{ce}} = 500$ iterations, we start optimizing the flow-related network by minimizing the cross-entropy losses $\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{d}}$ and $\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{s}}$ . These flow-related networks include the tensorial encoder and the coupling layer network $\\{m_i\\}_{i=1\\dots N}^d$ and $\\{m_i\\}_{i=1\\dots N}^s$ for diffuse and specular respectively, each consists of a composition of $N = 2$ piecewise-quadratic coupling layers. We model each $m_i$ using a three-layer tiny MLP with 64 hidden unit. After the warm-up phase, the importance sampler is set as a frozen copy of the current normalizing flow, updated every $N_{\\mathrm{update}} = 1,000$ iterations. After the warm-up phase, we", + "bbox": [ + 511, + 599, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "499", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/070d5317c06de0bf3b2a4ebe2ad80d48249f11a97d1bc25399015c6b64bb247a.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison of relighting quality on TensoSDF dataset [20]." + ], + "image_footnote": [], + "bbox": [ + 93, + 90, + 897, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b3ac61acb56291ceceac4cc613a99d9d292ea88c039023e40e98955b25b995c.jpg", + "image_caption": [ + "Figure 4. Visualization of material decomposition, normal map, and per-pixel variance in rendering equation evaluation, comparing results with a pre-defined sampler and our proposed tensorial normalizing flow. The per-pixel variance results are scaled by 1000 for clearer visualization." + ], + "image_footnote": [], + "bbox": [ + 98, + 429, + 901, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "employ the proposed tensorial importance sampler to sample incident directions. We sample 128 rays for the specular term. For the diffuse term, we apply the MIS [29] technique, sampling 64 rays from another normalizing flow and", + "bbox": [ + 89, + 835, + 483, + 897 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "512 rays from a cosine-weighted distribution. $\\lambda_{m},\\lambda_{\\mathrm{ce}}^{\\mathrm{d}}$ , and $\\lambda_{\\mathrm{ce}}^{\\mathrm{s}}$ are set to 1.0, 0.0001, and 0.0001, respectively. The training process takes 4 hours for the first stage and 2.5 hours for the second stage. All experiments are conducted", + "bbox": [ + 511, + 835, + 906, + 897 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "500", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9d3c846ff156af4bbfea330b4da568b9b28384f4652e56a719ac763b52f56ea8.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of relighting quality on TensoSDF dataset [20]." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
TensoIR [15]NeRO [23]NeiLF++ [39]TensoSDF [20]TenSOFlowTensoIRNeRONeiLF++TensoSDFTenSOFlowTensoIRNeRONeiLF++TensoSDFTenSOFlow
Rover24.00024.01523.77426.75426.9360.9180.9140.9110.9350.9370.08010.06930.07540.05930.0566
Dragon25.10425.64424.09927.89927.9000.8950.9190.9010.9360.9370.13020.08980.09880.07750.0756
Motor19.21922.15820.14222.75425.5100.9060.9170.8940.9300.9440.08210.07020.08700.06810.0556
Helmet25.14022.58724.00128.12628.3850.9010.8810.9060.9340.9370.10400.10790.10560.07700.0707
Robot26.03123.19422.69626.24227.4160.9280.9130.9150.9400.9430.09310.07550.07820.06130.0577
Compressor20.75321.62419.74024.04925.4530.8680.8780.8440.9160.9290.10380.10730.12860.08300.0789
Average23.37523.20422.41025.97126.9330.9030.9040.8950.9320.9380.09890.08670.09560.07100.0659
", + "bbox": [ + 112, + 114, + 883, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c5f7d060f38ee92c3db04ee78d3ef10c3c5de19e3cf54a759818cea7c01c851.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of relighted images with TensoSDF [20] on the Stanford-ORB dataset [19]." + ], + "image_footnote": [], + "bbox": [ + 116, + 222, + 890, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on a single NVIDIA H100 GPU.", + "bbox": [ + 89, + 454, + 312, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Results on synthetic data", + "text_level": 1, + "bbox": [ + 89, + 483, + 318, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first perform experiments on the TensoSDF synthetic dataset [20] to evaluate the effectiveness of our method in accurately reconstructing materials and lighting. Following the setups of NeRO [23] and TensoSDF [20], we extract each object's estimated material properties after optimization and use Blender to perform relighting based on the extracted mesh and material information. As shown in Tab. 1, our method consistently outperforms competitors across multiple scenes and metrics. This improvement highlights the benefit of our learnable importance sampler, which adapts to the integrand's spatial and directional complexity, reducing Monte Carlo sampling variance and enhancing material and lighting estimation accuracy. Notably, NeRO underperforms relative to TensoSDF, due to its less accurate geometry reconstruction, which affects subsequent material and lighting estimations. In contrast, our method, built upon TensoSDF's robust geometry, further refines material estimation with a learnable importance sampler, achieving a closer alignment with the integrand's complex distribution in the rendering equation. In Fig. 3, we provide a visual comparison of relighted images with NeRO and TensoSDF. Without a spatially and directionally aware importance sampler, the competitors rely on pre-defined sampling that struggles to match the shape of the integrand, leading to high variance and potential convergence to local minima during optimization.", + "bbox": [ + 89, + 508, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/26c797622183f8ec4f0e7b88894649db309fd9df10e866848330e805c801addf.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of relighting quality on Stanford-ORB dataset [19]." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
TensoSDF [20]TenSOFlowTensoSDFTenSOFlowTenSOSDFTenSOFlow
Teapot29.97129.3420.9860.9860.02760.0250
Gnome28.48829.1790.9560.9600.09470.0899
Cactus31.35131.4960.9840.9850.03620.0366
Car28.01029.0350.9810.9840.03670.0306
Grogu30.59530.7350.9900.9900.03620.0385
Avg.29.68329.9570.9790.9810.04630.0441
", + "bbox": [ + 524, + 492, + 893, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 4, we visualize the estimated materials rendered by TensoFlow alongside per-pixel variance in Monte Carlo sampling of specular term. \"Variance (pre-defined)\" shows results using fixed importance samplers, specifically a GGX distribution. By contrast, our tensorial normalizing flow effectively models the importance sampler, allowing it to dynamically adjust to the integrand's spatial and directional variations. This approach enables our method to produce detailed material maps—including albedo, metallic, and roughness, while significantly reducing variance with the same number of sampled rays.", + "bbox": [ + 511, + 604, + 906, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results on real data", + "text_level": 1, + "bbox": [ + 511, + 785, + 702, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate our method on the real-world Stanford-ORB dataset [19], following the setup used in TensoSDF [20]. Five objects from the dataset are selected for evaluation, and each object's mesh and material properties are extracted from the optimized model and relight under two new environment lighting conditions. As shown", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "501", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/979150cfadd75541c1aa0812b5c4d2b1468b953daf652d37c6b8f4b92ab6256a.jpg", + "image_caption": [ + "w/o diffuse" + ], + "image_footnote": [], + "bbox": [ + 96, + 95, + 186, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a16aac9971513f5a1f751358f7429b28ffe39be98daba350e440bcba6db4c70e.jpg", + "image_caption": [ + "w/o specular" + ], + "image_footnote": [], + "bbox": [ + 194, + 98, + 282, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f2a2e02dc675f8210b70e4f8c5c5d31e4074cfcdbbc7a0b1754f7d90e6888513.jpg", + "image_caption": [ + "TenSoFlow" + ], + "image_footnote": [], + "bbox": [ + 290, + 98, + 375, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae89bdaf3181517febbd4aad0275fd0d32e6077dceff2afa1fffa1476f99eba2.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 390, + 98, + 475, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a7fbdf035e441da548ceffc54f8d60a848fddbeb08c173fc2e7df83045d41f30.jpg", + "image_caption": [ + "(a) Ablation studies on learnable importance samples", + "w/o reflected" + ], + "image_footnote": [], + "bbox": [ + 101, + 195, + 187, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89109be41d83ec3c8f4232eb706aed4f7f557670ee23c39e3a753c05fe5ddc71.jpg", + "image_caption": [ + "w/o tensorial" + ], + "image_footnote": [], + "bbox": [ + 197, + 195, + 284, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3471a206be754fbd9615f85e69d33486cbba339759d347c2aa0b42d39655a224.jpg", + "image_caption": [ + "TenSoFlow" + ], + "image_footnote": [], + "bbox": [ + 295, + 195, + 377, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/158608b81b1f3b1fd1973e6cfe7f4928538a1593e3aefc22a75027d8c6c7a7e9.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 395, + 195, + 478, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/115d8de387056aaa55e411edf9fb6eebe8b10481ccc16d87b9ce61d1392380e1.jpg", + "image_caption": [ + "(b) Ablation studies on spatial and directional priors.", + "32", + "(c) Ablation studies on number of sampled rays.", + "Figure 6. Ablation studies on various components of TensoFlow." + ], + "image_footnote": [], + "bbox": [ + 98, + 297, + 187, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96066b31fb236e03af96d331904bbb1d89547da55470dbba4384c0ae1f041246.jpg", + "image_caption": [ + "64" + ], + "image_footnote": [], + "bbox": [ + 196, + 297, + 284, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/edb228df3bd12f1c44f9f6e6428769e44f0e5ef39607ba70bb31719e3decfe6a.jpg", + "image_caption": [ + "128" + ], + "image_footnote": [], + "bbox": [ + 292, + 297, + 377, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc48a521d6331cac6c9fe18a72ce7749fa37ab85620842e7c643f091f76e1d34.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 390, + 297, + 478, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Tab. 2, our method demonstrates superior performance across most scenes. The performance increase, while notable, is less pronounced than on the TensoSDF synthetic dataset, likely due to the simpler geometry and lighting conditions of the objects in Stanford-ORB. In Fig. 5, we present a qualitative comparison between TensoSDF and our method, showing that our approach achieves more accurate material properties, which improves the quality of relighting.", + "bbox": [ + 89, + 431, + 483, + 568 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/274a8a85c53f1d615c072ed72ea82af4a36ffaeec76f34075ccd51118bee0777.jpg", + "table_caption": [ + "Table 3. Ablation studies on various components of TensoFlow." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
w/o diffuse26.4330.9370.0679
w/o specular26.2030.9360.0675
One sampler25.1960.9330.0736
w/o half26.4740.9370.0666
w/o tensorial26.0580.9320.0688
w/o reflected25.8760.9350.0703
Ns=3225.8490.9330.0706
Ns=6426.4400.9360.0674
Full26.9330.9380.0659
", + "bbox": [ + 107, + 609, + 462, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation", + "text_level": 1, + "bbox": [ + 89, + 768, + 194, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 3, we conduct ablation studies on various components of TensoFlow to evaluate their impact on relighting quality. We report the average metrics among all scenes in TensoSDF [20] dataset.", + "bbox": [ + 89, + 794, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learnable importance sampler We investigate the impact of the learnable importance samplers in inverse rendering. In Tab. 3 and Fig. 6a, \"w/o diffuse\" and \"w/o specular\" refer", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to using only the pre-defined importance samplers (cosine-weighted distribution for the diffuse term and GGX distribution for the specular term) for each component. \"One sampler\" indicates the use of a single normalizing flow to model both diffuse and specular terms simultaneously. The results demonstrate the importance of modeling diffuse and specular terms with separate normalizing flows for optimal performance. Additionally, \"w/o half\" represents modeling the distribution of the incident direction $\\omega_{i}$ directly instead of the half vector $\\omega_{h}$ . The results show that modeling the distribution of the half vector $\\omega_{h}$ yields improved accuracy.", + "bbox": [ + 511, + 90, + 903, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial and directional prior In Fig. 6b, we conduct ablation studies to assess the effectiveness of incorporating spatial and directional priors into the learnable importance sampler. In the settings \"w/o tensorial\" and \"w/o reflected direction\", we exclude the tensorial latent feature $V_{f}$ and the reflected direction $\\omega_{r}$ as inputs to the network $m_{i}$ in the normalizing flow, respectively. The experimental results indicate that both inputs are essential for optimal performance.", + "bbox": [ + 511, + 260, + 903, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of sampled rays In Fig. 6c, we evaluate our method's performance using varying numbers of sampled rays for evaluating the specular term in the rendering equation (Eq. (11)). Results indicate that with $N_{s} = 64$ sampled rays, our method already achieves good performance, with further improvements as the sample count increases to $N_{s} = 128$ , yielding even greater accuracy.", + "bbox": [ + 511, + 385, + 905, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 515, + 633, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have introduced TensoFlow, a novel approach for inverse rendering that utilizes a spatially and directionally aware, learnable importance sampler to tackle the complexity of light-surface interactions within the rendering equation. By parameterizing the sampler with normalizing flows, TensoFlow facilitates both efficient directional sampling of incident light and PDF inference. Our approach incorporates a learned tensorial representation along with the reflected direction to provide spatial and directional priors to the normalizing flow, resulting in a spatially and directionally aware sampling distribution. Additionally, we propose an optimization strategy to minimize the discrepancy between the integrand distribution and our learnable importance sampler. Extensive experiments on synthetic and real-world datasets demonstrate the effectiveness of TensoFlow, highlighting its superiority in material and lighting estimation compared to prior methods.", + "bbox": [ + 511, + 544, + 906, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation Employing a normalizing flow as the importance sampler for the rendering equation slow down the sampling process, as each sampled direction requires a query to the flow.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "502", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62376060).", + "bbox": [ + 89, + 114, + 483, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 160, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Matthew O'Toole, and Pratul P Srinivasan. Flash cache: Reducing bias in radiance cache based inverse rendering. arXiv preprint arXiv:2409.05867, 2024. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022.", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 1, 2", + "[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 2", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2, 3", + "[7] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM TOG, 1982. 4", + "[8] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint, 2014. 4", + "[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2", + "[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. In ECCV, 2024. 1, 2", + "[11] Chun Gu, Xiaofei Wei, Zixuan Zeng, Yuxuan Yao, and Li Zhang. Irgs: Inter-reflective gaussian splatting with 2d gaussian ray tracing. arXiv preprint, 2024. 1", + "[12] John Hammersley. Monte carlo methods. Springer Science & Business Media, 2013. 2", + "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH, 2024. 1", + "[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In CVPR, 2024. 1", + "[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 1, 2, 5, 7" + ], + "bbox": [ + 93, + 185, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986. 1, 2", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 1, 2", + "[18] Julian Knodt, Joe Bartusek, Seung-Hwan Baek, and Felix Heide. Neural ray-tracing: Learning surfaces and reflectance for relighting and view synthesis. arXiv preprint, 2021. 2", + "[19] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. NIPS, 2024. 5, 7", + "[20] Jia Li, Lu Wang, Lei Zhang, and Beibei Wang. Tensosdf: Roughness-aware tensorial representation for robust geometry and material reconstruction. ACM TOG, 2024. 1, 2, 3, 5, 6, 7, 8", + "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2", + "[22] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splattering for inverse rendering. In CVPR, 2024. 1, 2", + "[23] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 1, 2, 4, 5, 7", + "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[25] Thomas Müller, Brian McWilliams, Fabrice Rousselle, Markus Gross, and Jan Novák. Neural importance sampling. ACM TOG, 2019. 1, 5", + "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2", + "[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2", + "[28] Kenneth E Torrance and Ephraim M Sparrow. Theory for off-specular reflection from roughened surfaces. Josa, 1967. 2", + "[29] Eric Veach and Leonidas J Guibas. Optimally combining sampling techniques for monte carlo rendering. In SIGGRAPH, 1995. 6", + "[30] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Rendering techniques, 2007. 1", + "[31] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021. 2", + "[32] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "503", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004. 5", + "[34] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint, 2022. 2", + "[35] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In ECCV, 2022. 1, 2", + "[36] Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. In ICLR, 2025. 1", + "[37] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 2021. 2", + "[38] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, 2024. 1", + "[39] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 5, 7", + "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[41] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 2" + ], + "bbox": [ + 91, + 92, + 480, + 517 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "504", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_model.json b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_model.json new file mode 100644 index 0000000000000000000000000000000000000000..32f4cfc9e26767d844c2d91ec3796d39fbec2139 --- /dev/null +++ b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_model.json @@ -0,0 +1,2332 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.171, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "TensoFlow: Tensorial Flow-based Sampler for Inverse Rendering" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.18, + 0.74, + 0.199 + ], + "angle": 0, + "content": "Chun Gu\\(^{1,2}\\) Xiaofei Wei\\(^{1,2}\\) Li Zhang\\(^{1,2*}\\) Xiatian Zhu\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.199, + 0.898, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Data Science, Fudan University \\(^{2}\\)Shanghai Innovation Institute \\(^{3}\\)University of Surrey" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.225, + 0.725, + 0.242 + ], + "angle": 0, + "content": "https://github.com/fudan-zvg/tensoflow" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.277, + 0.327, + 0.293 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.486, + 0.747 + ], + "angle": 0, + "content": "Inverse rendering aims to recover scene geometry, material properties, and lighting from multi-view images. Given the complexity of light-surface interactions, importance sampling is essential for the evaluation of the rendering equation, as it reduces variance and enhances the efficiency of Monte Carlo sampling. Existing inverse rendering methods typically use pre-defined non-learnable importance samplers in prior manually, struggling to effectively match the spatially and directionally varied integrand and resulting in high variance and suboptimal performance. To address this limitation, we propose the concept of learning a spatially and directionally aware importance sampler for the rendering equation to accurately and flexibly capture the unconstrained complexity of a typical scene. We further formulate TensoFlow, a generic approach for sampler learning in inverse rendering, enabling to closely match the integrand of the rendering equation spatially and directionally. Concretely, our sampler is parameterized by normalizing flows, allowing both directional sampling of incident light and probability density function (PDF) inference. To capture the characteristics of the sampler spatially, we learn a tensorial representation of the scene space, which imposes spatial conditions, together with reflected direction, leading to spatially and directionally aware sampling distributions. Our model can be optimized by minimizing the difference between the integrand and our normalizing flow. Extensive experiments validate the superiority of TensoFlow over prior alternatives on both synthetic and real-world benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.222, + 0.792 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.485, + 0.879 + ], + "angle": 0, + "content": "3D scene modeling has made significant strides, especially with the emergence of powerful representation models like neural radiance fields (NeRF) [2-4, 24] and efficient ones like 3D Gaussian splatting (3DGS) [13, 17, 38]. However, these methods struggle with inverse rendering due to the in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.905, + 0.308 + ], + "angle": 0, + "content": "ability to effectively disentangle geometry, material properties, and lighting, restricting their capacity to relight scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.907, + 0.581 + ], + "angle": 0, + "content": "To address this issue, recent works [10, 11, 14, 15, 20, 22, 23, 35, 36] have advanced appearance modeling by incorporating the physically-based rendering equation [16] that simulates light-surface interactions. This however introduces a fundamental challenge - a need for accurately evaluating the rendering equation using Monte Carlo sampling. One intuitive method is stratified sampling over the hemisphere [15, 35], but this can be computationally inefficient due to the need for many samples. To improve efficiency, leading methods like NeRO [23] and TensoSDF [20] employ importance sampling. These approaches utilize pre-defined importance samplers (e.g., cosine-weighted and GGX samplers [30] for diffuse and specular components, respectively). However, these pre-defined importance samplers still result in high variance, as the integrand typically varies in both space and direction, making it difficult for a fixed sampler to effectively match the scene's characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.869 + ], + "angle": 0, + "content": "To address this challenge, we propose a novel perspective of learning spatially and directionally aware importance samplers, allowing us to manage the complexity of any scene without the need for manual sampler selection, while benefiting from greater flexibility and generality. As a showcase, we formulate TensoFlow, a concrete approach for importance sampler learning. Our sampler is parameterized by normalizing flows composed with piecewise-quadratic coupling layers [25], functioning as a neural distribution that enables both probability density function (PDF) inference and directional sampling of incident light in the rendering equation. To effectively capture the characteristics of the sampler spatially, we learn a tensorial representation of the scene space, imposing spatial conditions on the rendering process, together with reflected direction, resulting in spatially and directionally aware sampling distributions. This formulation can be optimized by minimizing the distributional difference between the integrand and our normalizing flow." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We make these contributions: (i) We propose a novel concept of learning the importance samplers of the render-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.421, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author (lizhangfd@fudan.edu.cn)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "495" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "ing equation for inverse rendering, in contrast to existing pre-defined non-learnable samplers struggling intrinsically to capture the intrinsic rendering complexity of a scene. (ii) We introduce an effective method for sampler learning, TensoFlow, parameterized by tensorial normalizing flows, serving as a neural distribution that can support both directional sampling of incident light and PDF inference required in the rendering equation. (iii) Extensive experiments on both synthetic and real-world datasets demonstrate that TensoFlow outperforms previous alternatives, achieving significantly lower variance in evaluating the rendering equation with the same number of sampled directions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.285, + 0.229, + 0.3 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.485, + 0.598 + ], + "angle": 0, + "content": "Novel view synthesis The introduction of neural radiance fields (NeRF) [24], which represent 3D scenes as continuous volumetric functions, utilizes volumetric rendering techniques to synthesize images from novel viewpoints, achieving effective 3D scene reconstruction. However, NeRF exhibits limitations in terms of rendering efficiency. Various efficient alternatives have been proposed, including voxel-based methods [9, 27], multi-resolution hash grids [26], and tensor decomposition approaches [6]. Additionally, techniques such as anti-aliasing [2-4] and modeling with signed distance fields (SDFs) [21, 31, 32, 34, 37] have been employed to enhance rendering and geometry reconstruction quality. Recently, 3D Gaussian splatting (3DGS) [17] has been introduced with favored efficiency, representing 3D scenes as a collection of 3D Gaussians. However, despite its impressive performance in rendering quality and efficiency, 3DGS lacks robust geometry, which limits its effectiveness in geometry-sensitive tasks such as inverse rendering [10, 22]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Inverse rendering Inverse rendering aims to recover the 3D structure, lighting conditions, and material properties of a scene from posed 2D images. Recent advancements in neural rendering have substantially driven inverse rendering performance. NeRF-based approaches [1, 5, 15, 18, 20, 23, 35, 39, 41] simulate physical light-surface interactions, allowing for the disentanglement of scene geometry, materials, and lighting from image observations. In particular, Monte Carlo sampling is commonly used to evaluate the rendering equation in inverse rendering. For example, avoiding ray marching on incident rays, NeILF [35] optimizes an incident light field using stratified sampling (e.g., sampling directions uniformly over the hemisphere) for the rendering equation. Similarly, TensoIR [15] also applies stratified sampling during training using tensorial implicit fields. However, stratified sampling can be computationally costly due to the high sample count needed for accuracy. To address this issue, NeRO [23] and TensoSDF [20] adopt importance sampling with pre-defined samplers (e.g., cosine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "weighted, GGX). Commonly, all these sampling methods fall short of matching the spatially and directionally varying integrand of the rendering equation, leading to suboptimal rendering performance. To overcome this limitation, we introduce the concept of learning importance samplers that addressing the need for spatially and directionally aware samplers elegantly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.211, + 0.64, + 0.229 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.905, + 0.266 + ], + "angle": 0, + "content": "We begin with necessary background on importance sampling and the rendering equation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.373 + ], + "angle": 0, + "content": "Importance sampling Monte Carlo sampling is a commonly used method for evaluating integrals [12]. The principal challenge lies in approximating the integral with fewer samples while achieving reduced variance. Given an integral \\( F = \\int_{\\Omega} f(x) \\, \\mathrm{d}x \\), where \\( f(x) \\) is the target integrand function over the domain \\( \\Omega \\), and \\( x \\in \\mathbb{R}^D \\) is a sample, importance sampling is performed by:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.384, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\nF = \\int_ {\\Omega} \\frac {f (x)}{q (x)} q (x) \\mathrm {d} x = \\mathbb {E} \\left[ \\frac {f (X)}{q (X)} \\right] \\approx \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {f \\left(X _ {i}\\right)}{q \\left(X _ {i}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.906, + 0.544 + ], + "angle": 0, + "content": "where the proxy PDF \\( q(x) \\) is designed to approximate the distribution of \\( f(x) \\) in a way that minimizes the variance of the estimator. Ideally, \\( q(x) \\) should be chosen so that it matches the shape of \\( f(x) \\) closely, meaning that areas where \\( f(x) \\) has higher values are sampled more frequently. This helps in achieving a more accurate estimate of the integral with fewer samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.907, + 0.633 + ], + "angle": 0, + "content": "**Rendering equation** The rendering equation [16] models the interaction between lighting and surfaces. Given the incident radiance \\( L_{i}(\\omega_{i}, x) \\) and the material properties at a surface point \\( x \\) with normal vector \\( n \\), the rendering equation expresses the outgoing radiance in the following manner:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.643, + 0.905, + 0.675 + ], + "angle": 0, + "content": "\\[\nL _ {o} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\int_ {\\Omega} f \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {i} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}\\right) d \\boldsymbol {\\omega} _ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.906, + 0.777 + ], + "angle": 0, + "content": "where \\( f(\\omega_o, \\omega_i, x) \\) is the Bidirectional Reflectance Distribution Function (BRDF). Here, we consider the microfacet BRDF parameterization [28], including material properties such as albedo \\( a \\in [0,1]^3 \\), metallic \\( m \\in [0,1] \\), and roughness \\( r \\in [0,1] \\). Formally, this BRDF can be decomposed into a diffuse term \\( f_d \\) and a specular term \\( f_s \\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.786, + 0.907, + 0.819 + ], + "angle": 0, + "content": "\\[\nf _ {d} = \\frac {(1 - m) \\boldsymbol {a}}{\\pi}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.833, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\nf _ {s} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\frac {D F G}{4 \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n}\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(D, F\\), and \\(G\\) represent the normal distribution function, the Fresnel term, and the geometry term, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "496" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.633, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.092, + 0.906, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.315, + 0.907, + 0.414 + ], + "angle": 0, + "content": "Figure 1. Left: Material and lighting estimation stage of TensoFlow. Given a ray-surface intersection point \\( \\pmb{x} \\), we use tensorial encoders to encode latent features for both the material properties and the importance sampler. When evaluating the rendering equation, incident directions are sampled from our learnable importance sampler, which is a frozen copy of the training normalizing flow. The normalizing flow is optimized by minimizing the distribution difference \\( \\mathcal{L}_{\\mathrm{ce}} \\) between \\( q(\\omega_{i}) \\) and the integrand. The material properties, parameterized by \\( \\theta_{m} \\), are optimized by minimizing the RGB rendering loss \\( \\mathcal{L}_{\\mathrm{c}} \\) of the final rendering integral. Right: Tensorial normalizing flow. With spatial prior \\( V_{f} \\) and directional prior \\( \\omega_{r} \\), our tensorial normalizing flow, implemented using piecewise-quadratic coupling layers, enables both incident direction sampling and PDF querying." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.438, + 0.214, + 0.453 + ], + "angle": 0, + "content": "4. TensoFLow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.484, + 0.586 + ], + "angle": 0, + "content": "Architecture overview Our TensoFlow is comprised of two stages: (I) The first stage involves reconstructing the scene geometry. We follow TensoSDF [20] and utilize a tensorial signed distance field combined with a roughness-aware training strategy to achieve robust geometry reconstruction. (II) The second stage focuses on modeling the material and lighting, which is central to inverse rendering and constitutes the core of our approach, as detailed below." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.596, + 0.376, + 0.613 + ], + "angle": 0, + "content": "4.1. Material and lighting estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Parametrization In TensoFlow, we represent the material of a scene with a Vector-Matrix tensorial encoder [6], due to its high capacity and flexibility. For a surface point \\( \\pmb{x} \\), its tensorial feature vector is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.691, + 0.483, + 0.712 + ], + "angle": 0, + "content": "\\[\nV _ {m} (\\boldsymbol {x}) = \\mathrm {v} _ {m, k} ^ {X} \\circ \\mathrm {M} _ {m, k} ^ {Y Z} \\oplus \\mathrm {v} _ {m, k} ^ {Y} \\circ \\mathrm {M} _ {m, k} ^ {X Z} \\oplus \\mathrm {v} _ {m, k} ^ {Z} \\circ \\mathrm {M} _ {m, k} ^ {X Y}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.484, + 0.814 + ], + "angle": 0, + "content": "where \\(\\mathrm{v}_k\\) and \\(\\mathrm{M}_k\\) represent the \\(k\\)-th vector and matrix along the corresponding axis. The operators \\(\\circ\\) and \\(\\oplus\\) denote element-wise multiplication and concatenation, respectively. After concatenating this feature vector with the positional encoding \\(\\mathbf{p}\\) of the position \\(\\boldsymbol{x}\\), it is processed by a tiny MLP \\(\\Theta_m\\) to produce the material properties:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.826, + 0.483, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {a}, m, r \\right\\} = \\Theta_ {m} \\left(V _ {m}, \\mathbf {p}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( \\mathbf{a} \\), \\( m \\), and \\( r \\) denote albedo, metallic, and roughness, respectively. For the incident light modeling, we decompose it into spatially-aware indirect lighting and direct light-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.706, + 0.455 + ], + "angle": 0, + "content": "ing originating from infinity:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.463, + 0.905, + 0.48 + ], + "angle": 0, + "content": "\\[\nL \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) = V \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {\\operatorname {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right) + L _ {\\operatorname {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.548 + ], + "angle": 0, + "content": "where the visibility term \\( V(\\omega_{i},\\pmb {x}) \\) is determined via ray tracing on the extracted mesh. The indirect and direct lightings are modeled by a tiny MLP \\( \\Theta_{\\mathrm{ind}} \\) and an environment map \\( \\Theta_{\\mathrm{dir}} \\), respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.556, + 0.905, + 0.572 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) = \\Theta_ {\\text {i n d}} \\left(\\mathbf {p}, \\boldsymbol {\\omega} _ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.575, + 0.905, + 0.591 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right) = \\Theta_ {\\mathrm {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.744 + ], + "angle": 0, + "content": "**Rendering** Given the reconstructed signed distance field from the first stage, the intersection point between each ray and the surface can be determined. The material properties and incident lighting at this intersection can then be queried as described earlier. In accordance with Eq. (2), we employ Monte Carlo sampling to evaluate the integral in the rendering equation. Due to differing characteristics of the diffuse and specular terms, separate importance samplers are used to minimize variance in the Monte Carlo sampling process:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.763, + 0.905, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} _ {\\text {d i f f u s e}} = \\frac {1}{N _ {d}} \\sum_ {i} ^ {N _ {d}} \\frac {(1 - m) \\frac {a}{\\pi} L (\\boldsymbol {\\omega} _ {i} , \\boldsymbol {x}) (\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n})}{q _ {\\mathrm {d}} (\\boldsymbol {\\omega} _ {i} ; \\boldsymbol {x})}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.807, + 0.905, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} _ {\\text {s p e c u l a r}} = \\frac {1}{N _ {s}} \\sum_ {i} ^ {N _ {s}} \\frac {\\frac {D F G}{4 (\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}) (\\boldsymbol {\\omega} _ {o} \\cdot \\boldsymbol {n})} L (\\boldsymbol {\\omega} _ {i} , \\boldsymbol {x}) (\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n})}{q _ {\\mathrm {s}} (\\boldsymbol {\\omega} _ {i} ; \\boldsymbol {x})}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\( q_{\\mathrm{d}}(\\omega_i) \\) and \\( q_{\\mathrm{s}}(\\omega_i) \\) are the PDFs of the importance samplers, which we model them using the proposed normalizing flow (Sec. 4.2). The final color is computed as:" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "497" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.476, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.483, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Illustration of a piecewise-quadratic coupling layer, incorporating tensorial latent feature \\( V_{f} \\) and reflected direction \\( \\omega_{r} \\) as spatial and directional priors." + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.387, + 0.483, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} = \\boldsymbol {c} _ {\\text {d i f f u s e}} + \\boldsymbol {c} _ {\\text {s p e c u l a r}}, \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.412, + 0.453, + 0.429 + ], + "angle": 0, + "content": "4.2. Tensorial flow-based importance sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.571 + ], + "angle": 0, + "content": "For evaluating the rendering equation, prior works often employ pre-defined distributions as the importance samplers. For instance, in NeRO [23] a cosine-weighted distribution is used as the importance sampler for the diffuse term, and the GGX distribution [7] for the specular term. However, such samplers cannot closely match the shape of the target integrand with complex spatial and directional variations for a typical scene, leading to high variance and performance degradation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.591, + 0.285, + 0.607 + ], + "angle": 0, + "content": "4.2.1 Flow-based sampler" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.781 + ], + "angle": 0, + "content": "To overcome this limitation, we propose learning the importance samplers. The normalizing flow framework [8] is adopted as the sampler learner due to its following theoretical advantage. For a given direction \\(\\omega\\) with two degrees of freedom in spherical coordinates, the distribution of the integrand function may exhibit inter-dependencies across different coordinate dimensions. We can express the variable \\(\\omega\\) as a bijective transformation of an independent variable \\(z\\), which follows a uniform distribution \\((0,1)^{2}\\): \\(\\omega = h(z)\\). Using the change-of-variable formula, the PDF of \\(\\omega\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.792, + 0.483, + 0.83 + ], + "angle": 0, + "content": "\\[\nq (\\boldsymbol {\\omega}) = q (\\boldsymbol {z}) \\left| \\det \\left(\\frac {\\partial h (\\boldsymbol {z})}{\\partial \\boldsymbol {z} ^ {T}}\\right) \\right| ^ {- 1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.87 + ], + "angle": 0, + "content": "where \\(q(z)\\) represents the PDF of \\(z\\) in \\(\\mathcal{U}(0,1)^2\\), which is straightforward to evaluate." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "However, the selection of the bijective mapping \\( h \\) must satisfy certain conditions: it must be invertible, and it" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "should allow efficient computation of the Jacobian determinant. To meet these requirements, we employ NICE [8], a type of autoregressive flow that utilizes \"coupling layers\" to enable fast sampling and density evaluation. The final mapping \\( h \\) is then formed as a composition of simpler bijective transformations: \\( h = h_{n} \\circ \\dots \\circ h_{1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Given an input vector \\(x = (x_{1},x_{2})\\in \\mathbb{R}^{2}\\) for a mapping \\(h_i\\) , the output vector \\(y = (y_{1},y_{2})\\in \\mathbb{R}^{2}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.223, + 0.905, + 0.241 + ], + "angle": 0, + "content": "\\[\ny _ {1} = x _ {1}, \\quad y _ {2} = C (x _ {2}; m (x _ {1})), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.906, + 0.326 + ], + "angle": 0, + "content": "where the coupling transform \\(C:\\mathbb{R}\\leftrightarrow \\mathbb{R}\\) is a separable and invertible transformation, and \\(m(x)\\), parameterized by a neural network, acts as the parameter of \\(C\\). This structure allows the inverse of the coupling layer to be computed directly:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.336, + 0.905, + 0.354 + ], + "angle": 0, + "content": "\\[\nx _ {1} = y _ {1}, \\quad x _ {2} = C ^ {- 1} \\left(y _ {2}; m \\left(y _ {1}\\right)\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.363, + 0.906, + 0.408 + ], + "angle": 0, + "content": "The separability of \\( C \\) ensures that the Jacobian matrix \\( J_{i} \\) of \\( h_{i} \\) is triangular, simplifying the computation of its determinant to the product of its diagonal elements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.906, + 0.573 + ], + "angle": 0, + "content": "By applying the change-of-variable formula, the Jacobian determinant of the composite mapping \\( h \\) is expressed as the product of the individual \\( J_{i} \\): \\( |\\operatorname{det} J| = \\prod_{i} |\\operatorname{det} J_{i}| \\). We note that, with sufficiently expressive coupling transformations and adequate coupling layers, the sampling distribution \\( q(x) \\) can be modeled as an arbitrarily complex function, allowing it to closely match the shape of any integrand in the rendering equation. Importantly, each adjacent coupling layer preserves different dimensions of the input vector (Eq. (14)), ensuring the mapping's expressiveness across all the dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.906, + 0.665 + ], + "angle": 0, + "content": "The normalizing flow defined by the above mapping acts as a distribution from which we can query the PDF or draw samples, corresponding to \\( h^{-1} \\) and \\( h \\), respectively. Our method supports two primary applications in the inverse rendering pipeline: inferring the PDF value for a given direction and sampling a direction along with its PDF value:" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.691, + 0.905, + 0.708 + ], + "angle": 0, + "content": "Inference: \\(q(\\omega)\\gets h^{-1}(\\omega),\\) (16)" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.711, + 0.905, + 0.729 + ], + "angle": 0, + "content": "Sampling: \\((\\omega ,q(\\omega))\\gets h(\\pmb {u}),\\pmb {u}\\sim \\mathcal{U}(0,1)^2,\\) (17)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.738, + 0.906, + 0.829 + ], + "angle": 0, + "content": "Our experiments reveal that modeling the distribution of the half vector \\(\\omega_{h}\\) yields better performance compared to directly modeling \\(\\omega_{i}\\). By modeling the distribution \\(\\omega_{h}\\), \\(q(\\omega_{i})\\) can be derived using the change-of-variable formula (Eq. (13)), enabling both PDF inference and sampling of \\(\\omega_{i}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.807, + 0.862 + ], + "angle": 0, + "content": "4.2.2 Tensorial coupling transformation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Given the spatially-dependent nature of the integrand in the rendering equation, we further parameterize the scene space" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "498" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.473, + 0.108 + ], + "angle": 0, + "content": "by incorporating an additional tensorial representation as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.117, + 0.483, + 0.14 + ], + "angle": 0, + "content": "\\[\nV _ {f} (\\boldsymbol {x}) = \\mathrm {v} _ {f, k} ^ {X} \\circ \\mathrm {M} _ {f, k} ^ {Y Z} \\oplus \\mathrm {v} _ {f, k} ^ {Y} \\circ \\mathrm {M} _ {f, k} ^ {X Z} \\oplus \\mathrm {v} _ {f, k} ^ {Z} \\circ \\mathrm {M} _ {f, k} ^ {X Y}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.149, + 0.484, + 0.239 + ], + "angle": 0, + "content": "The latent feature \\( V_{f} \\), together with the reflected direction \\( \\omega_{r} = 2(\\omega_{o}\\cdot n)n - \\omega_{o} \\) for directional dependency, serves as conditions to guide the network \\( m_{i} \\) within each coupling layer. This is crucial, as the shape of the integrand in the rendering equation is influenced by both the spatial position and the outgoing direction \\( \\omega_{o} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.484, + 0.487 + ], + "angle": 0, + "content": "We consider the piecewise-quadratic coupling transform [25] as our coupling transformation \\( C \\) for its high expressiveness. Specifically, it achieves the bijective property by defining a Cumulative Distribution Function (CDF): \\( y = C(x; m(x'; V_f, \\omega_r)) = \\int_0^x p(t; m(x'; V_f, \\omega_r)) \\, \\mathrm{d}t \\), where \\( p \\) is a spatially and directionally aware PDF derived from the output of network \\( m \\). This results in the Jacobian determinant of the coupling layer being equal to \\( p(x) \\). Here, the network \\( m \\) takes the concatenation of \\( x', V_f \\), and \\( \\omega_r \\) as input, generating \\( \\widehat{V} \\) and \\( \\widehat{W} \\) for subsequent use. We then construct the piecewise linear PDF using \\( K + 1 \\) vertices, as illustrated in Fig. 2. It employs \\( K + 1 \\) vertical coordinates \\( \\widehat{V} \\in \\mathbb{R}^{K + 1} \\) and \\( K \\) horizontal differences \\( \\widehat{W} \\in \\mathbb{R}^K \\), with \\( K \\) as the predefined bin count for the distribution. Following softmax normalization, \\( W = \\sigma(\\widehat{W}) \\) represents the bin widths, while \\( V \\) defines the PDF values at each vertex." + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.498, + 0.483, + 0.55 + ], + "angle": 0, + "content": "\\[\nV _ {i} = \\frac {\\exp (\\widehat {V} _ {k})}{\\sum_ {k = 1} ^ {K} \\frac {\\exp (\\widehat {V} _ {k}) + \\exp (\\widehat {V} _ {k + 1})}{2} W _ {k}}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.559, + 0.484, + 0.589 + ], + "angle": 0, + "content": "Consequently, given an input \\( x \\), the corresponding PDF value can be queried as" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.603, + 0.482, + 0.62 + ], + "angle": 0, + "content": "\\[\np (x) = \\operatorname {l e r p} \\left(V _ {b}, V _ {b + 1}, \\alpha\\right), \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.68 + ], + "angle": 0, + "content": "where \\(\\alpha = (x - \\sum_{k=1}^{b-1} W_k) / W_k\\) defines the relative position of \\(x\\) within the corresponding bin \\(b\\). The piecewise-quadratic CDF can be computed by integrating the PDF:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.692, + 0.484, + 0.745 + ], + "angle": 0, + "content": "\\[\nC (x) = \\frac {\\alpha W _ {b}}{2} ((2 - \\alpha) V _ {b} + \\alpha V _ {b + 1}) + \\sum_ {k = 1} ^ {b - 1} \\frac {V _ {k} + V _ {k + 1}}{2} W _ {k}. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.484, + 0.807 + ], + "angle": 0, + "content": "We employ \\(N\\) coupling layers with piecewise-quadratic transformation to construct the compound mapping \\(h\\), ensuring it is sufficiently expressive for modeling the importance sampler." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.196, + 0.833 + ], + "angle": 0, + "content": "4.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.9 + ], + "angle": 0, + "content": "During model training, we optimize the tensorial encoder \\( V_{f} \\) and the networks \\( \\{m_i\\}_{i = 1\\dots N} \\) by minimizing the cross-entropy loss between the integrand of the rendering equation and the PDF of the incident directions inferred from" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.66, + 0.107 + ], + "angle": 0, + "content": "the normalizing flow:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.116, + 0.905, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {c e}} = - \\int_ {\\Omega} I (\\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}) \\log q (\\boldsymbol {\\omega} _ {i}) \\mathrm {d} \\boldsymbol {\\omega} _ {i} (22) \\\\ = \\mathbb {E} \\left[ - \\frac {I \\left(\\boldsymbol {\\omega} _ {i} , \\boldsymbol {\\omega} _ {o} , \\boldsymbol {x}\\right)}{\\hat {q} \\left(\\boldsymbol {\\omega} _ {i}\\right)} \\log q \\left(\\boldsymbol {\\omega} _ {i}\\right) \\right], (23) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.907, + 0.328 + ], + "angle": 0, + "content": "where \\(\\hat{q} (\\omega_{i})\\) represents the PDF of the sampled incident directions from the current importance sampler, a frozen copy of the trainable normalizing flow updated every \\(N_{\\mathrm{update}}\\) iterations. \\(q(\\omega_{i})\\) denotes the PDF inferred from the current training normalizing flow, and \\(I(\\omega_i,\\omega_o,x)\\) is the integrand function evaluated with the sampled incident direction. We use different importance samplers for the diffuse and specular terms, resulting in separate losses: \\(\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{d}}\\) and \\(\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{s}}\\). The total training loss is thus defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.337, + 0.905, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c} + \\lambda_ {m} \\mathcal {L} _ {m} + \\lambda_ {\\mathrm {c e}} ^ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {d}} + \\lambda_ {\\mathrm {c e}} ^ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {s}}, \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.905, + 0.395 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_m\\) represent the RGB rendering loss and the material regularization loss [23], respectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.638, + 0.424 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.905, + 0.509 + ], + "angle": 0, + "content": "Datasets and metrics We use the TensoSDF synthetic dataset [20] and the real-world Stanford-ORB dataset [19] for quantitative evaluation. To quantitatively evaluate the estimated material, we adopt three standard metrics for relighted images: PSNR, SSIM [33], and LPIPS [40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.905, + 0.599 + ], + "angle": 0, + "content": "Competitors We compare our approach with state-of-the-art NeRF-based methods selected for their different sampling strategies in evaluating the rendering equation: TensoIR [15] and NeILF++[39], which utilize stratified sampling, and NeRO[23] and TensoSDF [20], which employ pre-defined importance sampling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Implementation details The training process consists of two stages. For the first stage, geometry reconstruction, we train the TensoSDF model for 180,000 iterations using the same settings as in TensoSDF [20]. For the second stage, material and lighting estimation, we begin with a warm-up phase, employing cosine-weighted and GGX distributions as the importance samplers for the diffuse and specular components, respectively, over \\( N_{\\mathrm{warmup}} = 1,000 \\) iterations. Once a coarse estimation of material and lighting is achieved at \\( N_{\\mathrm{ce}} = 500 \\) iterations, we start optimizing the flow-related network by minimizing the cross-entropy losses \\( \\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{d}} \\) and \\( \\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{s}} \\). These flow-related networks include the tensorial encoder and the coupling layer network \\( \\{m_i\\}_{i=1\\dots N}^d \\) and \\( \\{m_i\\}_{i=1\\dots N}^s \\) for diffuse and specular respectively, each consists of a composition of \\( N = 2 \\) piecewise-quadratic coupling layers. We model each \\( m_i \\) using a three-layer tiny MLP with 64 hidden unit. After the warm-up phase, the importance sampler is set as a frozen copy of the current normalizing flow, updated every \\( N_{\\mathrm{update}} = 1,000 \\) iterations. After the warm-up phase, we" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "499" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.092, + 0.898, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.4, + 0.738, + 0.415 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of relighting quality on TensoSDF dataset [20]." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.43, + 0.903, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.768, + 0.908, + 0.81 + ], + "angle": 0, + "content": "Figure 4. Visualization of material decomposition, normal map, and per-pixel variance in rendering equation evaluation, comparing results with a pre-defined sampler and our proposed tensorial normalizing flow. The per-pixel variance results are scaled by 1000 for clearer visualization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.837, + 0.484, + 0.898 + ], + "angle": 0, + "content": "employ the proposed tensorial importance sampler to sample incident directions. We sample 128 rays for the specular term. For the diffuse term, we apply the MIS [29] technique, sampling 64 rays from another normalizing flow and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.837, + 0.907, + 0.898 + ], + "angle": 0, + "content": "512 rays from a cosine-weighted distribution. \\(\\lambda_{m},\\lambda_{\\mathrm{ce}}^{\\mathrm{d}}\\) , and \\(\\lambda_{\\mathrm{ce}}^{\\mathrm{s}}\\) are set to 1.0, 0.0001, and 0.0001, respectively. The training process takes 4 hours for the first stage and 2.5 hours for the second stage. All experiments are conducted" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "500" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.258, + 0.09, + 0.743, + 0.105 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of relighting quality on TensoSDF dataset [20]." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.115, + 0.885, + 0.209 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
TensoIR [15]NeRO [23]NeiLF++ [39]TensoSDF [20]TenSOFlowTensoIRNeRONeiLF++TensoSDFTenSOFlowTensoIRNeRONeiLF++TensoSDFTenSOFlow
Rover24.00024.01523.77426.75426.9360.9180.9140.9110.9350.9370.08010.06930.07540.05930.0566
Dragon25.10425.64424.09927.89927.9000.8950.9190.9010.9360.9370.13020.08980.09880.07750.0756
Motor19.21922.15820.14222.75425.5100.9060.9170.8940.9300.9440.08210.07020.08700.06810.0556
Helmet25.14022.58724.00128.12628.3850.9010.8810.9060.9340.9370.10400.10790.10560.07700.0707
Robot26.03123.19422.69626.24227.4160.9280.9130.9150.9400.9430.09310.07550.07820.06130.0577
Compressor20.75321.62419.74024.04925.4530.8680.8780.8440.9160.9290.10380.10730.12860.08300.0789
Average23.37523.20422.41025.97126.9330.9030.9040.8950.9320.9380.09890.08670.09560.07100.0659
" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.223, + 0.891, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.414, + 0.822, + 0.43 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of relighted images with TensoSDF [20] on the Stanford-ORB dataset [19]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.455, + 0.313, + 0.471 + ], + "angle": 0, + "content": "on a single NVIDIA H100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.484, + 0.32, + 0.5 + ], + "angle": 0, + "content": "5.1. Results on synthetic data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We first perform experiments on the TensoSDF synthetic dataset [20] to evaluate the effectiveness of our method in accurately reconstructing materials and lighting. Following the setups of NeRO [23] and TensoSDF [20], we extract each object's estimated material properties after optimization and use Blender to perform relighting based on the extracted mesh and material information. As shown in Tab. 1, our method consistently outperforms competitors across multiple scenes and metrics. This improvement highlights the benefit of our learnable importance sampler, which adapts to the integrand's spatial and directional complexity, reducing Monte Carlo sampling variance and enhancing material and lighting estimation accuracy. Notably, NeRO underperforms relative to TensoSDF, due to its less accurate geometry reconstruction, which affects subsequent material and lighting estimations. In contrast, our method, built upon TensoSDF's robust geometry, further refines material estimation with a learnable importance sampler, achieving a closer alignment with the integrand's complex distribution in the rendering equation. In Fig. 3, we provide a visual comparison of relighted images with NeRO and TensoSDF. Without a spatially and directionally aware importance sampler, the competitors rely on pre-defined sampling that struggles to match the shape of the integrand, leading to high variance and potential convergence to local minima during optimization." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.453, + 0.907, + 0.481 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of relighting quality on Stanford-ORB dataset [19]." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.493, + 0.894, + 0.577 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
TensoSDF [20]TenSOFlowTensoSDFTenSOFlowTenSOSDFTenSOFlow
Teapot29.97129.3420.9860.9860.02760.0250
Gnome28.48829.1790.9560.9600.09470.0899
Cactus31.35131.4960.9840.9850.03620.0366
Car28.01029.0350.9810.9840.03670.0306
Grogu30.59530.7350.9900.9900.03620.0385
Avg.29.68329.9570.9790.9810.04630.0441
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.772 + ], + "angle": 0, + "content": "In Fig. 4, we visualize the estimated materials rendered by TensoFlow alongside per-pixel variance in Monte Carlo sampling of specular term. \"Variance (pre-defined)\" shows results using fixed importance samplers, specifically a GGX distribution. By contrast, our tensorial normalizing flow effectively models the importance sampler, allowing it to dynamically adjust to the integrand's spatial and directional variations. This approach enables our method to produce detailed material maps—including albedo, metallic, and roughness, while significantly reducing variance with the same number of sampled rays." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.786, + 0.704, + 0.8 + ], + "angle": 0, + "content": "5.2. Results on real data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We further evaluate our method on the real-world Stanford-ORB dataset [19], following the setup used in TensoSDF [20]. Five objects from the dataset are selected for evaluation, and each object's mesh and material properties are extracted from the optimized model and relight under two new environment lighting conditions. As shown" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "501" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.096, + 0.187, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.161, + 0.18, + 0.173 + ], + "angle": 0, + "content": "w/o diffuse" + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.099, + 0.283, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.162, + 0.282, + 0.173 + ], + "angle": 0, + "content": "w/o specular" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.099, + 0.377, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.162, + 0.37, + 0.173 + ], + "angle": 0, + "content": "TenSoFlow" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.099, + 0.477, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.162, + 0.448, + 0.172 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.179, + 0.404, + 0.191 + ], + "angle": 0, + "content": "(a) Ablation studies on learnable importance samples" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.196, + 0.188, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.264, + 0.184, + 0.275 + ], + "angle": 0, + "content": "w/o reflected" + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.196, + 0.285, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.264, + 0.282, + 0.275 + ], + "angle": 0, + "content": "w/o tensorial" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.196, + 0.378, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.264, + 0.369, + 0.275 + ], + "angle": 0, + "content": "TenSoFlow" + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.196, + 0.48, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.264, + 0.447, + 0.274 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.28, + 0.427, + 0.294 + ], + "angle": 0, + "content": "(b) Ablation studies on spatial and directional priors." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.298, + 0.188, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.366, + 0.142, + 0.376 + ], + "angle": 0, + "content": "32" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.298, + 0.285, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.366, + 0.24, + 0.376 + ], + "angle": 0, + "content": "64" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.298, + 0.378, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.366, + 0.339, + 0.376 + ], + "angle": 0, + "content": "128" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.298, + 0.48, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.366, + 0.436, + 0.376 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.383, + 0.414, + 0.396 + ], + "angle": 0, + "content": "(c) Ablation studies on number of sampled rays." + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.408, + 0.479, + 0.422 + ], + "angle": 0, + "content": "Figure 6. Ablation studies on various components of TensoFlow." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.569 + ], + "angle": 0, + "content": "in Tab. 2, our method demonstrates superior performance across most scenes. The performance increase, while notable, is less pronounced than on the TensoSDF synthetic dataset, likely due to the simpler geometry and lighting conditions of the objects in Stanford-ORB. In Fig. 5, we present a qualitative comparison between TensoSDF and our method, showing that our approach achieves more accurate material properties, which improves the quality of relighting." + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.584, + 0.475, + 0.598 + ], + "angle": 0, + "content": "Table 3. Ablation studies on various components of TensoFlow." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.61, + 0.464, + 0.76 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
w/o diffuse26.4330.9370.0679
w/o specular26.2030.9360.0675
One sampler25.1960.9330.0736
w/o half26.4740.9370.0666
w/o tensorial26.0580.9320.0688
w/o reflected25.8760.9350.0703
Ns=3225.8490.9330.0706
Ns=6426.4400.9360.0674
Full26.9330.9380.0659
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.195, + 0.785 + ], + "angle": 0, + "content": "5.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.854 + ], + "angle": 0, + "content": "As shown in Tab. 3, we conduct ablation studies on various components of TensoFlow to evaluate their impact on relighting quality. We report the average metrics among all scenes in TensoSDF [20] dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Learnable importance sampler We investigate the impact of the learnable importance samplers in inverse rendering. In Tab. 3 and Fig. 6a, \"w/o diffuse\" and \"w/o specular\" refer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.259 + ], + "angle": 0, + "content": "to using only the pre-defined importance samplers (cosine-weighted distribution for the diffuse term and GGX distribution for the specular term) for each component. \"One sampler\" indicates the use of a single normalizing flow to model both diffuse and specular terms simultaneously. The results demonstrate the importance of modeling diffuse and specular terms with separate normalizing flows for optimal performance. Additionally, \"w/o half\" represents modeling the distribution of the incident direction \\(\\omega_{i}\\) directly instead of the half vector \\(\\omega_{h}\\). The results show that modeling the distribution of the half vector \\(\\omega_{h}\\) yields improved accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.905, + 0.383 + ], + "angle": 0, + "content": "Spatial and directional prior In Fig. 6b, we conduct ablation studies to assess the effectiveness of incorporating spatial and directional priors into the learnable importance sampler. In the settings \"w/o tensorial\" and \"w/o reflected direction\", we exclude the tensorial latent feature \\( V_{f} \\) and the reflected direction \\( \\omega_{r} \\) as inputs to the network \\( m_{i} \\) in the normalizing flow, respectively. The experimental results indicate that both inputs are essential for optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.492 + ], + "angle": 0, + "content": "Number of sampled rays In Fig. 6c, we evaluate our method's performance using varying numbers of sampled rays for evaluating the specular term in the rendering equation (Eq. (11)). Results indicate that with \\( N_{s} = 64 \\) sampled rays, our method already achieves good performance, with further improvements as the sample count increases to \\( N_{s} = 128 \\), yielding even greater accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.634, + 0.532 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.803 + ], + "angle": 0, + "content": "In this paper, we have introduced TensoFlow, a novel approach for inverse rendering that utilizes a spatially and directionally aware, learnable importance sampler to tackle the complexity of light-surface interactions within the rendering equation. By parameterizing the sampler with normalizing flows, TensoFlow facilitates both efficient directional sampling of incident light and PDF inference. Our approach incorporates a learned tensorial representation along with the reflected direction to provide spatial and directional priors to the normalizing flow, resulting in a spatially and directionally aware sampling distribution. Additionally, we propose an optimization strategy to minimize the discrepancy between the integrand distribution and our learnable importance sampler. Extensive experiments on synthetic and real-world datasets demonstrate the effectiveness of TensoFlow, highlighting its superiority in material and lighting estimation compared to prior methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Limitation Employing a normalizing flow as the importance sampler for the rendering equation slow down the sampling process, as each sampled direction requires a query to the flow." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "502" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.147 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62376060)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.161, + 0.188, + 0.177 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Matthew O'Toole, and Pratul P Srinivasan. Flash cache: Reducing bias in radiance cache based inverse rendering. arXiv preprint arXiv:2409.05867, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.257, + 0.484, + 0.312 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.314, + 0.484, + 0.355 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.357, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.401, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.484, + 0.498 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[7] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM TOG, 1982. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[8] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.616, + 0.484, + 0.672 + ], + "angle": 0, + "content": "[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. In ECCV, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[11] Chun Gu, Xiaofei Wei, Zixuan Zeng, Yuxuan Yao, and Li Zhang. Irgs: Inter-reflective gaussian splatting with 2d gaussian ray tracing. arXiv preprint, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[12] John Hammersley. Monte carlo methods. Springer Science & Business Media, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 1, 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.186, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "[16] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.907, + 0.221 + ], + "angle": 0, + "content": "[18] Julian Knodt, Joe Bartusek, Seung-Hwan Baek, and Felix Heide. Neural ray-tracing: Learning surfaces and reflectance for relighting and view synthesis. arXiv preprint, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[19] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. NIPS, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[20] Jia Li, Lu Wang, Lei Zhang, and Beibei Wang. Tensosdf: Roughness-aware tensorial representation for robust geometry and material reconstruction. ACM TOG, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[22] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splattering for inverse rendering. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[23] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[25] Thomas Müller, Brian McWilliams, Fabrice Rousselle, Markus Gross, and Jan Novák. Neural importance sampling. ACM TOG, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.617 + ], + "angle": 0, + "content": "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.906, + 0.658 + ], + "angle": 0, + "content": "[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[28] Kenneth E Torrance and Ephraim M Sparrow. Theory for off-specular reflection from roughened surfaces. Josa, 1967. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[29] Eric Veach and Leonidas J Guibas. Optimally combining sampling techniques for monte carlo rendering. In SIGGRAPH, 1995. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[30] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Rendering techniques, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[31] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[32] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "503" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[34] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[35] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[36] Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. In ICLR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[37] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[38] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[39] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[41] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "504" + } + ] +] \ No newline at end of file diff --git a/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_origin.pdf b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ac1919220e9f594d6bb0da8cd781a4c7e169b193 --- /dev/null +++ b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/ade03037-3358-4b0c-84b6-3b9ceb264f42_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40f71793d9a91ddfa4cdb7fa9dc2e1b182a701c788ef2efceed4993357ca5809 +size 6298699 diff --git a/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/full.md b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0ef088454deedd99f508faa76b520ec73f4cd3c9 --- /dev/null +++ b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/full.md @@ -0,0 +1,370 @@ +# TensoFlow: Tensorial Flow-based Sampler for Inverse Rendering + +Chun Gu $^{1,2}$ Xiaofei Wei $^{1,2}$ Li Zhang $^{1,2*}$ Xiatian Zhu $^{3}$ + +$^{1}$ School of Data Science, Fudan University $^{2}$ Shanghai Innovation Institute $^{3}$ University of Surrey + +https://github.com/fudan-zvg/tensoflow + +# Abstract + +Inverse rendering aims to recover scene geometry, material properties, and lighting from multi-view images. Given the complexity of light-surface interactions, importance sampling is essential for the evaluation of the rendering equation, as it reduces variance and enhances the efficiency of Monte Carlo sampling. Existing inverse rendering methods typically use pre-defined non-learnable importance samplers in prior manually, struggling to effectively match the spatially and directionally varied integrand and resulting in high variance and suboptimal performance. To address this limitation, we propose the concept of learning a spatially and directionally aware importance sampler for the rendering equation to accurately and flexibly capture the unconstrained complexity of a typical scene. We further formulate TensoFlow, a generic approach for sampler learning in inverse rendering, enabling to closely match the integrand of the rendering equation spatially and directionally. Concretely, our sampler is parameterized by normalizing flows, allowing both directional sampling of incident light and probability density function (PDF) inference. To capture the characteristics of the sampler spatially, we learn a tensorial representation of the scene space, which imposes spatial conditions, together with reflected direction, leading to spatially and directionally aware sampling distributions. Our model can be optimized by minimizing the difference between the integrand and our normalizing flow. Extensive experiments validate the superiority of TensoFlow over prior alternatives on both synthetic and real-world benchmarks. + +# 1. Introduction + +3D scene modeling has made significant strides, especially with the emergence of powerful representation models like neural radiance fields (NeRF) [2-4, 24] and efficient ones like 3D Gaussian splatting (3DGS) [13, 17, 38]. However, these methods struggle with inverse rendering due to the in + +ability to effectively disentangle geometry, material properties, and lighting, restricting their capacity to relight scenes. + +To address this issue, recent works [10, 11, 14, 15, 20, 22, 23, 35, 36] have advanced appearance modeling by incorporating the physically-based rendering equation [16] that simulates light-surface interactions. This however introduces a fundamental challenge - a need for accurately evaluating the rendering equation using Monte Carlo sampling. One intuitive method is stratified sampling over the hemisphere [15, 35], but this can be computationally inefficient due to the need for many samples. To improve efficiency, leading methods like NeRO [23] and TensoSDF [20] employ importance sampling. These approaches utilize pre-defined importance samplers (e.g., cosine-weighted and GGX samplers [30] for diffuse and specular components, respectively). However, these pre-defined importance samplers still result in high variance, as the integrand typically varies in both space and direction, making it difficult for a fixed sampler to effectively match the scene's characteristics. + +To address this challenge, we propose a novel perspective of learning spatially and directionally aware importance samplers, allowing us to manage the complexity of any scene without the need for manual sampler selection, while benefiting from greater flexibility and generality. As a showcase, we formulate TensoFlow, a concrete approach for importance sampler learning. Our sampler is parameterized by normalizing flows composed with piecewise-quadratic coupling layers [25], functioning as a neural distribution that enables both probability density function (PDF) inference and directional sampling of incident light in the rendering equation. To effectively capture the characteristics of the sampler spatially, we learn a tensorial representation of the scene space, imposing spatial conditions on the rendering process, together with reflected direction, resulting in spatially and directionally aware sampling distributions. This formulation can be optimized by minimizing the distributional difference between the integrand and our normalizing flow. + +We make these contributions: (i) We propose a novel concept of learning the importance samplers of the render- + +ing equation for inverse rendering, in contrast to existing pre-defined non-learnable samplers struggling intrinsically to capture the intrinsic rendering complexity of a scene. (ii) We introduce an effective method for sampler learning, TensoFlow, parameterized by tensorial normalizing flows, serving as a neural distribution that can support both directional sampling of incident light and PDF inference required in the rendering equation. (iii) Extensive experiments on both synthetic and real-world datasets demonstrate that TensoFlow outperforms previous alternatives, achieving significantly lower variance in evaluating the rendering equation with the same number of sampled directions. + +# 2. Related work + +Novel view synthesis The introduction of neural radiance fields (NeRF) [24], which represent 3D scenes as continuous volumetric functions, utilizes volumetric rendering techniques to synthesize images from novel viewpoints, achieving effective 3D scene reconstruction. However, NeRF exhibits limitations in terms of rendering efficiency. Various efficient alternatives have been proposed, including voxel-based methods [9, 27], multi-resolution hash grids [26], and tensor decomposition approaches [6]. Additionally, techniques such as anti-aliasing [2-4] and modeling with signed distance fields (SDFs) [21, 31, 32, 34, 37] have been employed to enhance rendering and geometry reconstruction quality. Recently, 3D Gaussian splatting (3DGS) [17] has been introduced with favored efficiency, representing 3D scenes as a collection of 3D Gaussians. However, despite its impressive performance in rendering quality and efficiency, 3DGS lacks robust geometry, which limits its effectiveness in geometry-sensitive tasks such as inverse rendering [10, 22]. + +Inverse rendering Inverse rendering aims to recover the 3D structure, lighting conditions, and material properties of a scene from posed 2D images. Recent advancements in neural rendering have substantially driven inverse rendering performance. NeRF-based approaches [1, 5, 15, 18, 20, 23, 35, 39, 41] simulate physical light-surface interactions, allowing for the disentanglement of scene geometry, materials, and lighting from image observations. In particular, Monte Carlo sampling is commonly used to evaluate the rendering equation in inverse rendering. For example, avoiding ray marching on incident rays, NeILF [35] optimizes an incident light field using stratified sampling (e.g., sampling directions uniformly over the hemisphere) for the rendering equation. Similarly, TensoIR [15] also applies stratified sampling during training using tensorial implicit fields. However, stratified sampling can be computationally costly due to the high sample count needed for accuracy. To address this issue, NeRO [23] and TensoSDF [20] adopt importance sampling with pre-defined samplers (e.g., cosine + +weighted, GGX). Commonly, all these sampling methods fall short of matching the spatially and directionally varying integrand of the rendering equation, leading to suboptimal rendering performance. To overcome this limitation, we introduce the concept of learning importance samplers that addressing the need for spatially and directionally aware samplers elegantly. + +# 3. Preliminary + +We begin with necessary background on importance sampling and the rendering equation. + +Importance sampling Monte Carlo sampling is a commonly used method for evaluating integrals [12]. The principal challenge lies in approximating the integral with fewer samples while achieving reduced variance. Given an integral $F = \int_{\Omega} f(x) \, \mathrm{d}x$ , where $f(x)$ is the target integrand function over the domain $\Omega$ , and $x \in \mathbb{R}^D$ is a sample, importance sampling is performed by: + +$$ +F = \int_ {\Omega} \frac {f (x)}{q (x)} q (x) \mathrm {d} x = \mathbb {E} \left[ \frac {f (X)}{q (X)} \right] \approx \frac {1}{N} \sum_ {i = 1} ^ {N} \frac {f \left(X _ {i}\right)}{q \left(X _ {i}\right)}, \tag {1} +$$ + +where the proxy PDF $q(x)$ is designed to approximate the distribution of $f(x)$ in a way that minimizes the variance of the estimator. Ideally, $q(x)$ should be chosen so that it matches the shape of $f(x)$ closely, meaning that areas where $f(x)$ has higher values are sampled more frequently. This helps in achieving a more accurate estimate of the integral with fewer samples. + +**Rendering equation** The rendering equation [16] models the interaction between lighting and surfaces. Given the incident radiance $L_{i}(\omega_{i}, x)$ and the material properties at a surface point $x$ with normal vector $n$ , the rendering equation expresses the outgoing radiance in the following manner: + +$$ +L _ {o} \left(\boldsymbol {\omega} _ {o}, \boldsymbol {x}\right) = \int_ {\Omega} f \left(\boldsymbol {\omega} _ {o}, \boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) L _ {i} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) \left(\boldsymbol {\omega} _ {i} \cdot \boldsymbol {n}\right) d \boldsymbol {\omega} _ {i}, \tag {2} +$$ + +where $f(\omega_o, \omega_i, x)$ is the Bidirectional Reflectance Distribution Function (BRDF). Here, we consider the microfacet BRDF parameterization [28], including material properties such as albedo $a \in [0,1]^3$ , metallic $m \in [0,1]$ , and roughness $r \in [0,1]$ . Formally, this BRDF can be decomposed into a diffuse term $f_d$ and a specular term $f_s$ defined as: + +$$ +f _ {d} = \frac {(1 - m) \boldsymbol {a}}{\pi}, \tag {3} +$$ + +$$ +f _ {s} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {\omega} _ {o}, \boldsymbol {x}\right) = \frac {D F G}{4 \left(\boldsymbol {\omega} _ {i} \cdot \mathbf {n}\right) \left(\boldsymbol {\omega} _ {o} \cdot \mathbf {n}\right)}, \tag {4} +$$ + +where $D, F$ , and $G$ represent the normal distribution function, the Fresnel term, and the geometry term, respectively. + +![](images/e2067f8e11fc6c70662adc07b2f5fe6806dcb0c852ab46122c0930417777148f.jpg) +Figure 1. Left: Material and lighting estimation stage of TensoFlow. Given a ray-surface intersection point $\pmb{x}$ , we use tensorial encoders to encode latent features for both the material properties and the importance sampler. When evaluating the rendering equation, incident directions are sampled from our learnable importance sampler, which is a frozen copy of the training normalizing flow. The normalizing flow is optimized by minimizing the distribution difference $\mathcal{L}_{\mathrm{ce}}$ between $q(\omega_{i})$ and the integrand. The material properties, parameterized by $\theta_{m}$ , are optimized by minimizing the RGB rendering loss $\mathcal{L}_{\mathrm{c}}$ of the final rendering integral. Right: Tensorial normalizing flow. With spatial prior $V_{f}$ and directional prior $\omega_{r}$ , our tensorial normalizing flow, implemented using piecewise-quadratic coupling layers, enables both incident direction sampling and PDF querying. + +![](images/fb272017abde4db394c69d84fe83972c2220f11c5535bc9a6497a873a09a6d3e.jpg) + +# 4. TensoFLow + +Architecture overview Our TensoFlow is comprised of two stages: (I) The first stage involves reconstructing the scene geometry. We follow TensoSDF [20] and utilize a tensorial signed distance field combined with a roughness-aware training strategy to achieve robust geometry reconstruction. (II) The second stage focuses on modeling the material and lighting, which is central to inverse rendering and constitutes the core of our approach, as detailed below. + +# 4.1. Material and lighting estimation + +Parametrization In TensoFlow, we represent the material of a scene with a Vector-Matrix tensorial encoder [6], due to its high capacity and flexibility. For a surface point $\pmb{x}$ , its tensorial feature vector is defined as + +$$ +V _ {m} (\boldsymbol {x}) = \mathrm {v} _ {m, k} ^ {X} \circ \mathrm {M} _ {m, k} ^ {Y Z} \oplus \mathrm {v} _ {m, k} ^ {Y} \circ \mathrm {M} _ {m, k} ^ {X Z} \oplus \mathrm {v} _ {m, k} ^ {Z} \circ \mathrm {M} _ {m, k} ^ {X Y}, \tag {5} +$$ + +where $\mathrm{v}_k$ and $\mathrm{M}_k$ represent the $k$ -th vector and matrix along the corresponding axis. The operators $\circ$ and $\oplus$ denote element-wise multiplication and concatenation, respectively. After concatenating this feature vector with the positional encoding $\mathbf{p}$ of the position $\boldsymbol{x}$ , it is processed by a tiny MLP $\Theta_m$ to produce the material properties: + +$$ +\left\{\boldsymbol {a}, m, r \right\} = \Theta_ {m} \left(V _ {m}, \mathbf {p}\right), \tag {6} +$$ + +where $\mathbf{a}$ , $m$ , and $r$ denote albedo, metallic, and roughness, respectively. For the incident light modeling, we decompose it into spatially-aware indirect lighting and direct light- + +ing originating from infinity: + +$$ +L \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) = V \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) L _ {\operatorname {d i r}} \left(\boldsymbol {\omega} _ {i}\right) + L _ {\operatorname {i n d}} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right), \tag {7} +$$ + +where the visibility term $V(\omega_{i},\pmb {x})$ is determined via ray tracing on the extracted mesh. The indirect and direct lightings are modeled by a tiny MLP $\Theta_{\mathrm{ind}}$ and an environment map $\Theta_{\mathrm{dir}}$ , respectively: + +$$ +L _ {\text {i n d}} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) = \Theta_ {\text {i n d}} \left(\mathbf {p}, \boldsymbol {\omega} _ {i}\right), \tag {8} +$$ + +$$ +L _ {\mathrm {d i r}} \left(\boldsymbol {\omega} _ {i}\right) = \Theta_ {\mathrm {d i r}} \left(\boldsymbol {\omega} _ {i}\right), \tag {9} +$$ + +**Rendering** Given the reconstructed signed distance field from the first stage, the intersection point between each ray and the surface can be determined. The material properties and incident lighting at this intersection can then be queried as described earlier. In accordance with Eq. (2), we employ Monte Carlo sampling to evaluate the integral in the rendering equation. Due to differing characteristics of the diffuse and specular terms, separate importance samplers are used to minimize variance in the Monte Carlo sampling process: + +$$ +\boldsymbol {c} _ {\text {d i f f u s e}} = \frac {1}{N _ {d}} \sum_ {i} ^ {N _ {d}} \frac {(1 - m) \frac {a}{\pi} L (\boldsymbol {\omega} _ {i} , \boldsymbol {x}) (\boldsymbol {\omega} _ {i} \cdot \boldsymbol {n})}{q _ {\mathrm {d}} (\boldsymbol {\omega} _ {i} ; \boldsymbol {x})}, \tag {10} +$$ + +$$ +\boldsymbol {c} _ {\text {s p e c u l a r}} = \frac {1}{N _ {s}} \sum_ {i} ^ {N _ {s}} \frac {\frac {D F G}{4 (\boldsymbol {\omega} _ {i} \cdot \boldsymbol {n}) (\boldsymbol {\omega} _ {o} \cdot \boldsymbol {n})} L (\boldsymbol {\omega} _ {i} , \boldsymbol {x}) (\boldsymbol {\omega} _ {i} \cdot \boldsymbol {n})}{q _ {\mathrm {s}} (\boldsymbol {\omega} _ {i} ; \boldsymbol {x})}, \tag {11} +$$ + +where $q_{\mathrm{d}}(\omega_i)$ and $q_{\mathrm{s}}(\omega_i)$ are the PDFs of the importance samplers, which we model them using the proposed normalizing flow (Sec. 4.2). The final color is computed as: + +![](images/b6da1d6c84fa578c32df647b32cd8dd88c6ca0f944dc05c8cae55bf4cf9cd831.jpg) +Figure 2. Illustration of a piecewise-quadratic coupling layer, incorporating tensorial latent feature $V_{f}$ and reflected direction $\omega_{r}$ as spatial and directional priors. + +$$ +\boldsymbol {c} = \boldsymbol {c} _ {\text {d i f f u s e}} + \boldsymbol {c} _ {\text {s p e c u l a r}}, \tag {12} +$$ + +# 4.2. Tensorial flow-based importance sampling + +For evaluating the rendering equation, prior works often employ pre-defined distributions as the importance samplers. For instance, in NeRO [23] a cosine-weighted distribution is used as the importance sampler for the diffuse term, and the GGX distribution [7] for the specular term. However, such samplers cannot closely match the shape of the target integrand with complex spatial and directional variations for a typical scene, leading to high variance and performance degradation. + +# 4.2.1 Flow-based sampler + +To overcome this limitation, we propose learning the importance samplers. The normalizing flow framework [8] is adopted as the sampler learner due to its following theoretical advantage. For a given direction $\omega$ with two degrees of freedom in spherical coordinates, the distribution of the integrand function may exhibit inter-dependencies across different coordinate dimensions. We can express the variable $\omega$ as a bijective transformation of an independent variable $z$ , which follows a uniform distribution $(0,1)^{2}$ : $\omega = h(z)$ . Using the change-of-variable formula, the PDF of $\omega$ can be expressed as: + +$$ +q (\boldsymbol {\omega}) = q (\boldsymbol {z}) \left| \det \left(\frac {\partial h (\boldsymbol {z})}{\partial \boldsymbol {z} ^ {T}}\right) \right| ^ {- 1}, \tag {13} +$$ + +where $q(z)$ represents the PDF of $z$ in $\mathcal{U}(0,1)^2$ , which is straightforward to evaluate. + +However, the selection of the bijective mapping $h$ must satisfy certain conditions: it must be invertible, and it + +should allow efficient computation of the Jacobian determinant. To meet these requirements, we employ NICE [8], a type of autoregressive flow that utilizes "coupling layers" to enable fast sampling and density evaluation. The final mapping $h$ is then formed as a composition of simpler bijective transformations: $h = h_{n} \circ \dots \circ h_{1}$ . + +Given an input vector $x = (x_{1},x_{2})\in \mathbb{R}^{2}$ for a mapping $h_i$ , the output vector $y = (y_{1},y_{2})\in \mathbb{R}^{2}$ is defined as: + +$$ +y _ {1} = x _ {1}, \quad y _ {2} = C (x _ {2}; m (x _ {1})), \tag {14} +$$ + +where the coupling transform $C:\mathbb{R}\leftrightarrow \mathbb{R}$ is a separable and invertible transformation, and $m(x)$ , parameterized by a neural network, acts as the parameter of $C$ . This structure allows the inverse of the coupling layer to be computed directly: + +$$ +x _ {1} = y _ {1}, \quad x _ {2} = C ^ {- 1} \left(y _ {2}; m \left(y _ {1}\right)\right). \tag {15} +$$ + +The separability of $C$ ensures that the Jacobian matrix $J_{i}$ of $h_{i}$ is triangular, simplifying the computation of its determinant to the product of its diagonal elements. + +By applying the change-of-variable formula, the Jacobian determinant of the composite mapping $h$ is expressed as the product of the individual $J_{i}$ : $|\operatorname{det} J| = \prod_{i} |\operatorname{det} J_{i}|$ . We note that, with sufficiently expressive coupling transformations and adequate coupling layers, the sampling distribution $q(x)$ can be modeled as an arbitrarily complex function, allowing it to closely match the shape of any integrand in the rendering equation. Importantly, each adjacent coupling layer preserves different dimensions of the input vector (Eq. (14)), ensuring the mapping's expressiveness across all the dimensions. + +The normalizing flow defined by the above mapping acts as a distribution from which we can query the PDF or draw samples, corresponding to $h^{-1}$ and $h$ , respectively. Our method supports two primary applications in the inverse rendering pipeline: inferring the PDF value for a given direction and sampling a direction along with its PDF value: + +Inference: $q(\omega)\gets h^{-1}(\omega),$ (16) + +Sampling: $(\omega ,q(\omega))\gets h(\pmb {u}),\pmb {u}\sim \mathcal{U}(0,1)^2,$ (17) + +Our experiments reveal that modeling the distribution of the half vector $\omega_{h}$ yields better performance compared to directly modeling $\omega_{i}$ . By modeling the distribution $\omega_{h}$ , $q(\omega_{i})$ can be derived using the change-of-variable formula (Eq. (13)), enabling both PDF inference and sampling of $\omega_{i}$ . + +# 4.2.2 Tensorial coupling transformation + +Given the spatially-dependent nature of the integrand in the rendering equation, we further parameterize the scene space + +by incorporating an additional tensorial representation as: + +$$ +V _ {f} (\boldsymbol {x}) = \mathrm {v} _ {f, k} ^ {X} \circ \mathrm {M} _ {f, k} ^ {Y Z} \oplus \mathrm {v} _ {f, k} ^ {Y} \circ \mathrm {M} _ {f, k} ^ {X Z} \oplus \mathrm {v} _ {f, k} ^ {Z} \circ \mathrm {M} _ {f, k} ^ {X Y}. \tag {18} +$$ + +The latent feature $V_{f}$ , together with the reflected direction $\omega_{r} = 2(\omega_{o}\cdot n)n - \omega_{o}$ for directional dependency, serves as conditions to guide the network $m_{i}$ within each coupling layer. This is crucial, as the shape of the integrand in the rendering equation is influenced by both the spatial position and the outgoing direction $\omega_{o}$ . + +We consider the piecewise-quadratic coupling transform [25] as our coupling transformation $C$ for its high expressiveness. Specifically, it achieves the bijective property by defining a Cumulative Distribution Function (CDF): $y = C(x; m(x'; V_f, \omega_r)) = \int_0^x p(t; m(x'; V_f, \omega_r)) \, \mathrm{d}t$ , where $p$ is a spatially and directionally aware PDF derived from the output of network $m$ . This results in the Jacobian determinant of the coupling layer being equal to $p(x)$ . Here, the network $m$ takes the concatenation of $x', V_f$ , and $\omega_r$ as input, generating $\widehat{V}$ and $\widehat{W}$ for subsequent use. We then construct the piecewise linear PDF using $K + 1$ vertices, as illustrated in Fig. 2. It employs $K + 1$ vertical coordinates $\widehat{V} \in \mathbb{R}^{K + 1}$ and $K$ horizontal differences $\widehat{W} \in \mathbb{R}^K$ , with $K$ as the predefined bin count for the distribution. Following softmax normalization, $W = \sigma(\widehat{W})$ represents the bin widths, while $V$ defines the PDF values at each vertex. + +$$ +V _ {i} = \frac {\exp (\widehat {V} _ {k})}{\sum_ {k = 1} ^ {K} \frac {\exp (\widehat {V} _ {k}) + \exp (\widehat {V} _ {k + 1})}{2} W _ {k}}, \tag {19} +$$ + +Consequently, given an input $x$ , the corresponding PDF value can be queried as + +$$ +p (x) = \operatorname {l e r p} \left(V _ {b}, V _ {b + 1}, \alpha\right), \tag {20} +$$ + +where $\alpha = (x - \sum_{k=1}^{b-1} W_k) / W_k$ defines the relative position of $x$ within the corresponding bin $b$ . The piecewise-quadratic CDF can be computed by integrating the PDF: + +$$ +C (x) = \frac {\alpha W _ {b}}{2} ((2 - \alpha) V _ {b} + \alpha V _ {b + 1}) + \sum_ {k = 1} ^ {b - 1} \frac {V _ {k} + V _ {k + 1}}{2} W _ {k}. \tag {21} +$$ + +We employ $N$ coupling layers with piecewise-quadratic transformation to construct the compound mapping $h$ , ensuring it is sufficiently expressive for modeling the importance sampler. + +# 4.3. Training + +During model training, we optimize the tensorial encoder $V_{f}$ and the networks $\{m_i\}_{i = 1\dots N}$ by minimizing the cross-entropy loss between the integrand of the rendering equation and the PDF of the incident directions inferred from + +the normalizing flow: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {c e}} = - \int_ {\Omega} I (\boldsymbol {\omega} _ {i}, \boldsymbol {\omega} _ {o}, \boldsymbol {x}) \log q (\boldsymbol {\omega} _ {i}) \mathrm {d} \boldsymbol {\omega} _ {i} (22) \\ = \mathbb {E} \left[ - \frac {I \left(\boldsymbol {\omega} _ {i} , \boldsymbol {\omega} _ {o} , \boldsymbol {x}\right)}{\hat {q} \left(\boldsymbol {\omega} _ {i}\right)} \log q \left(\boldsymbol {\omega} _ {i}\right) \right], (23) \\ \end{array} +$$ + +where $\hat{q} (\omega_{i})$ represents the PDF of the sampled incident directions from the current importance sampler, a frozen copy of the trainable normalizing flow updated every $N_{\mathrm{update}}$ iterations. $q(\omega_{i})$ denotes the PDF inferred from the current training normalizing flow, and $I(\omega_i,\omega_o,x)$ is the integrand function evaluated with the sampled incident direction. We use different importance samplers for the diffuse and specular terms, resulting in separate losses: $\mathcal{L}_{\mathrm{ce}}^{\mathrm{d}}$ and $\mathcal{L}_{\mathrm{ce}}^{\mathrm{s}}$ . The total training loss is thus defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \lambda_ {m} \mathcal {L} _ {m} + \lambda_ {\mathrm {c e}} ^ {\mathrm {d}} \mathcal {L} _ {\mathrm {c e}} ^ {\mathrm {d}} + \lambda_ {\mathrm {c e}} ^ {\mathrm {s}} \mathcal {L} _ {\mathrm {c e}} ^ {\mathrm {s}}, \tag {24} +$$ + +where $\mathcal{L}_c$ and $\mathcal{L}_m$ represent the RGB rendering loss and the material regularization loss [23], respectively. + +# 5. Experiment + +Datasets and metrics We use the TensoSDF synthetic dataset [20] and the real-world Stanford-ORB dataset [19] for quantitative evaluation. To quantitatively evaluate the estimated material, we adopt three standard metrics for relighted images: PSNR, SSIM [33], and LPIPS [40]. + +Competitors We compare our approach with state-of-the-art NeRF-based methods selected for their different sampling strategies in evaluating the rendering equation: TensoIR [15] and NeILF++[39], which utilize stratified sampling, and NeRO[23] and TensoSDF [20], which employ pre-defined importance sampling. + +Implementation details The training process consists of two stages. For the first stage, geometry reconstruction, we train the TensoSDF model for 180,000 iterations using the same settings as in TensoSDF [20]. For the second stage, material and lighting estimation, we begin with a warm-up phase, employing cosine-weighted and GGX distributions as the importance samplers for the diffuse and specular components, respectively, over $N_{\mathrm{warmup}} = 1,000$ iterations. Once a coarse estimation of material and lighting is achieved at $N_{\mathrm{ce}} = 500$ iterations, we start optimizing the flow-related network by minimizing the cross-entropy losses $\mathcal{L}_{\mathrm{ce}}^{\mathrm{d}}$ and $\mathcal{L}_{\mathrm{ce}}^{\mathrm{s}}$ . These flow-related networks include the tensorial encoder and the coupling layer network $\{m_i\}_{i=1\dots N}^d$ and $\{m_i\}_{i=1\dots N}^s$ for diffuse and specular respectively, each consists of a composition of $N = 2$ piecewise-quadratic coupling layers. We model each $m_i$ using a three-layer tiny MLP with 64 hidden unit. After the warm-up phase, the importance sampler is set as a frozen copy of the current normalizing flow, updated every $N_{\mathrm{update}} = 1,000$ iterations. After the warm-up phase, we + +![](images/070d5317c06de0bf3b2a4ebe2ad80d48249f11a97d1bc25399015c6b64bb247a.jpg) +Figure 3. Qualitative comparison of relighting quality on TensoSDF dataset [20]. + +![](images/6b3ac61acb56291ceceac4cc613a99d9d292ea88c039023e40e98955b25b995c.jpg) +Figure 4. Visualization of material decomposition, normal map, and per-pixel variance in rendering equation evaluation, comparing results with a pre-defined sampler and our proposed tensorial normalizing flow. The per-pixel variance results are scaled by 1000 for clearer visualization. + +employ the proposed tensorial importance sampler to sample incident directions. We sample 128 rays for the specular term. For the diffuse term, we apply the MIS [29] technique, sampling 64 rays from another normalizing flow and + +512 rays from a cosine-weighted distribution. $\lambda_{m},\lambda_{\mathrm{ce}}^{\mathrm{d}}$ , and $\lambda_{\mathrm{ce}}^{\mathrm{s}}$ are set to 1.0, 0.0001, and 0.0001, respectively. The training process takes 4 hours for the first stage and 2.5 hours for the second stage. All experiments are conducted + +Table 1. Quantitative comparison of relighting quality on TensoSDF dataset [20]. + +
PSNR ↑SSIM ↑LPIPS ↓
TensoIR [15]NeRO [23]NeiLF++ [39]TensoSDF [20]TenSOFlowTensoIRNeRONeiLF++TensoSDFTenSOFlowTensoIRNeRONeiLF++TensoSDFTenSOFlow
Rover24.00024.01523.77426.75426.9360.9180.9140.9110.9350.9370.08010.06930.07540.05930.0566
Dragon25.10425.64424.09927.89927.9000.8950.9190.9010.9360.9370.13020.08980.09880.07750.0756
Motor19.21922.15820.14222.75425.5100.9060.9170.8940.9300.9440.08210.07020.08700.06810.0556
Helmet25.14022.58724.00128.12628.3850.9010.8810.9060.9340.9370.10400.10790.10560.07700.0707
Robot26.03123.19422.69626.24227.4160.9280.9130.9150.9400.9430.09310.07550.07820.06130.0577
Compressor20.75321.62419.74024.04925.4530.8680.8780.8440.9160.9290.10380.10730.12860.08300.0789
Average23.37523.20422.41025.97126.9330.9030.9040.8950.9320.9380.09890.08670.09560.07100.0659
+ +![](images/4c5f7d060f38ee92c3db04ee78d3ef10c3c5de19e3cf54a759818cea7c01c851.jpg) +Figure 5. Qualitative comparison of relighted images with TensoSDF [20] on the Stanford-ORB dataset [19]. + +on a single NVIDIA H100 GPU. + +# 5.1. Results on synthetic data + +We first perform experiments on the TensoSDF synthetic dataset [20] to evaluate the effectiveness of our method in accurately reconstructing materials and lighting. Following the setups of NeRO [23] and TensoSDF [20], we extract each object's estimated material properties after optimization and use Blender to perform relighting based on the extracted mesh and material information. As shown in Tab. 1, our method consistently outperforms competitors across multiple scenes and metrics. This improvement highlights the benefit of our learnable importance sampler, which adapts to the integrand's spatial and directional complexity, reducing Monte Carlo sampling variance and enhancing material and lighting estimation accuracy. Notably, NeRO underperforms relative to TensoSDF, due to its less accurate geometry reconstruction, which affects subsequent material and lighting estimations. In contrast, our method, built upon TensoSDF's robust geometry, further refines material estimation with a learnable importance sampler, achieving a closer alignment with the integrand's complex distribution in the rendering equation. In Fig. 3, we provide a visual comparison of relighted images with NeRO and TensoSDF. Without a spatially and directionally aware importance sampler, the competitors rely on pre-defined sampling that struggles to match the shape of the integrand, leading to high variance and potential convergence to local minima during optimization. + +Table 2. Quantitative comparison of relighting quality on Stanford-ORB dataset [19]. + +
PSNR ↑SSIM ↑LPIPS ↓
TensoSDF [20]TenSOFlowTensoSDFTenSOFlowTenSOSDFTenSOFlow
Teapot29.97129.3420.9860.9860.02760.0250
Gnome28.48829.1790.9560.9600.09470.0899
Cactus31.35131.4960.9840.9850.03620.0366
Car28.01029.0350.9810.9840.03670.0306
Grogu30.59530.7350.9900.9900.03620.0385
Avg.29.68329.9570.9790.9810.04630.0441
+ +In Fig. 4, we visualize the estimated materials rendered by TensoFlow alongside per-pixel variance in Monte Carlo sampling of specular term. "Variance (pre-defined)" shows results using fixed importance samplers, specifically a GGX distribution. By contrast, our tensorial normalizing flow effectively models the importance sampler, allowing it to dynamically adjust to the integrand's spatial and directional variations. This approach enables our method to produce detailed material maps—including albedo, metallic, and roughness, while significantly reducing variance with the same number of sampled rays. + +# 5.2. Results on real data + +We further evaluate our method on the real-world Stanford-ORB dataset [19], following the setup used in TensoSDF [20]. Five objects from the dataset are selected for evaluation, and each object's mesh and material properties are extracted from the optimized model and relight under two new environment lighting conditions. As shown + +![](images/979150cfadd75541c1aa0812b5c4d2b1468b953daf652d37c6b8f4b92ab6256a.jpg) +w/o diffuse + +![](images/a16aac9971513f5a1f751358f7429b28ffe39be98daba350e440bcba6db4c70e.jpg) +w/o specular + +![](images/f2a2e02dc675f8210b70e4f8c5c5d31e4074cfcdbbc7a0b1754f7d90e6888513.jpg) +TenSoFlow + +![](images/ae89bdaf3181517febbd4aad0275fd0d32e6077dceff2afa1fffa1476f99eba2.jpg) +GT + +![](images/a7fbdf035e441da548ceffc54f8d60a848fddbeb08c173fc2e7df83045d41f30.jpg) +(a) Ablation studies on learnable importance samples +w/o reflected + +![](images/89109be41d83ec3c8f4232eb706aed4f7f557670ee23c39e3a753c05fe5ddc71.jpg) +w/o tensorial + +![](images/3471a206be754fbd9615f85e69d33486cbba339759d347c2aa0b42d39655a224.jpg) +TenSoFlow + +![](images/158608b81b1f3b1fd1973e6cfe7f4928538a1593e3aefc22a75027d8c6c7a7e9.jpg) +GT + +![](images/115d8de387056aaa55e411edf9fb6eebe8b10481ccc16d87b9ce61d1392380e1.jpg) +(b) Ablation studies on spatial and directional priors. +32 +(c) Ablation studies on number of sampled rays. +Figure 6. Ablation studies on various components of TensoFlow. + +![](images/96066b31fb236e03af96d331904bbb1d89547da55470dbba4384c0ae1f041246.jpg) +64 + +![](images/edb228df3bd12f1c44f9f6e6428769e44f0e5ef39607ba70bb31719e3decfe6a.jpg) +128 + +![](images/dc48a521d6331cac6c9fe18a72ce7749fa37ab85620842e7c643f091f76e1d34.jpg) +GT + +in Tab. 2, our method demonstrates superior performance across most scenes. The performance increase, while notable, is less pronounced than on the TensoSDF synthetic dataset, likely due to the simpler geometry and lighting conditions of the objects in Stanford-ORB. In Fig. 5, we present a qualitative comparison between TensoSDF and our method, showing that our approach achieves more accurate material properties, which improves the quality of relighting. + +Table 3. Ablation studies on various components of TensoFlow. + +
PSNR ↑SSIM ↑LPIPS ↓
w/o diffuse26.4330.9370.0679
w/o specular26.2030.9360.0675
One sampler25.1960.9330.0736
w/o half26.4740.9370.0666
w/o tensorial26.0580.9320.0688
w/o reflected25.8760.9350.0703
Ns=3225.8490.9330.0706
Ns=6426.4400.9360.0674
Full26.9330.9380.0659
+ +# 5.3. Ablation + +As shown in Tab. 3, we conduct ablation studies on various components of TensoFlow to evaluate their impact on relighting quality. We report the average metrics among all scenes in TensoSDF [20] dataset. + +Learnable importance sampler We investigate the impact of the learnable importance samplers in inverse rendering. In Tab. 3 and Fig. 6a, "w/o diffuse" and "w/o specular" refer + +to using only the pre-defined importance samplers (cosine-weighted distribution for the diffuse term and GGX distribution for the specular term) for each component. "One sampler" indicates the use of a single normalizing flow to model both diffuse and specular terms simultaneously. The results demonstrate the importance of modeling diffuse and specular terms with separate normalizing flows for optimal performance. Additionally, "w/o half" represents modeling the distribution of the incident direction $\omega_{i}$ directly instead of the half vector $\omega_{h}$ . The results show that modeling the distribution of the half vector $\omega_{h}$ yields improved accuracy. + +Spatial and directional prior In Fig. 6b, we conduct ablation studies to assess the effectiveness of incorporating spatial and directional priors into the learnable importance sampler. In the settings "w/o tensorial" and "w/o reflected direction", we exclude the tensorial latent feature $V_{f}$ and the reflected direction $\omega_{r}$ as inputs to the network $m_{i}$ in the normalizing flow, respectively. The experimental results indicate that both inputs are essential for optimal performance. + +Number of sampled rays In Fig. 6c, we evaluate our method's performance using varying numbers of sampled rays for evaluating the specular term in the rendering equation (Eq. (11)). Results indicate that with $N_{s} = 64$ sampled rays, our method already achieves good performance, with further improvements as the sample count increases to $N_{s} = 128$ , yielding even greater accuracy. + +# 6. Conclusion + +In this paper, we have introduced TensoFlow, a novel approach for inverse rendering that utilizes a spatially and directionally aware, learnable importance sampler to tackle the complexity of light-surface interactions within the rendering equation. By parameterizing the sampler with normalizing flows, TensoFlow facilitates both efficient directional sampling of incident light and PDF inference. Our approach incorporates a learned tensorial representation along with the reflected direction to provide spatial and directional priors to the normalizing flow, resulting in a spatially and directionally aware sampling distribution. Additionally, we propose an optimization strategy to minimize the discrepancy between the integrand distribution and our learnable importance sampler. Extensive experiments on synthetic and real-world datasets demonstrate the effectiveness of TensoFlow, highlighting its superiority in material and lighting estimation compared to prior methods. + +Limitation Employing a normalizing flow as the importance sampler for the rendering equation slow down the sampling process, as each sampled direction requires a query to the flow. + +# Acknowledgments + +This work was supported in part by National Natural Science Foundation of China (Grant No. 62376060). + +# References + +[1] Benjamin Attal, Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Matthew O'Toole, and Pratul P Srinivasan. Flash cache: Reducing bias in radiance cache based inverse rendering. arXiv preprint arXiv:2409.05867, 2024. 2 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 1, 2 +[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 2 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2, 3 +[7] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM TOG, 1982. 4 +[8] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint, 2014. 4 +[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2 +[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. In ECCV, 2024. 1, 2 +[11] Chun Gu, Xiaofei Wei, Zixuan Zeng, Yuxuan Yao, and Li Zhang. Irgs: Inter-reflective gaussian splatting with 2d gaussian ray tracing. arXiv preprint, 2024. 1 +[12] John Hammersley. Monte carlo methods. Springer Science & Business Media, 2013. 2 +[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH, 2024. 1 +[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In CVPR, 2024. 1 +[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 1, 2, 5, 7 + +[16] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986. 1, 2 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 1, 2 +[18] Julian Knodt, Joe Bartusek, Seung-Hwan Baek, and Felix Heide. Neural ray-tracing: Learning surfaces and reflectance for relighting and view synthesis. arXiv preprint, 2021. 2 +[19] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. NIPS, 2024. 5, 7 +[20] Jia Li, Lu Wang, Lei Zhang, and Beibei Wang. Tensosdf: Roughness-aware tensorial representation for robust geometry and material reconstruction. ACM TOG, 2024. 1, 2, 3, 5, 6, 7, 8 +[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2 +[22] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splattering for inverse rendering. In CVPR, 2024. 1, 2 +[23] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 1, 2, 4, 5, 7 +[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[25] Thomas Müller, Brian McWilliams, Fabrice Rousselle, Markus Gross, and Jan Novák. Neural importance sampling. ACM TOG, 2019. 1, 5 +[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2 +[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2 +[28] Kenneth E Torrance and Ephraim M Sparrow. Theory for off-specular reflection from roughened surfaces. Josa, 1967. 2 +[29] Eric Veach and Leonidas J Guibas. Optimally combining sampling techniques for monte carlo rendering. In SIGGRAPH, 1995. 6 +[30] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Rendering techniques, 2007. 1 +[31] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021. 2 +[32] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2 + +[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004. 5 +[34] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint, 2022. 2 +[35] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In ECCV, 2022. 1, 2 +[36] Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. In ICLR, 2025. 1 +[37] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 2021. 2 +[38] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, 2024. 1 +[39] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 5, 7 +[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[41] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 2 \ No newline at end of file diff --git a/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/images.zip b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fad16ff153227b92c3971573a5f1f022fd36ab22 --- /dev/null +++ b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e57e2a36b45617ae4d912f114930fcc579e4e37fb2dde3b89ff471a920ac4ac2 +size 649892 diff --git a/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/layout.json b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3d27a9307e4bf0989db4246de9cc4cc254d9d011 --- /dev/null +++ b/2025/TensoFlow_ Tensorial Flow-based Sampler for Inverse Rendering/layout.json @@ -0,0 +1,10125 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 507, + 121 + ], + "type": "text", + "content": "TensoFlow: Tensorial Flow-based Sampler for Inverse Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "spans": [ + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "type": "text", + "content": "Chun Gu" + }, + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "type": "text", + "content": " Xiaofei Wei" + }, + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "type": "text", + "content": " Li Zhang" + }, + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "type": "text", + "content": " Xiatian Zhu" + }, + { + "bbox": [ + 160, + 142, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 157, + 549, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 549, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 549, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 157, + 549, + 172 + ], + "type": "text", + "content": "School of Data Science, Fudan University " + }, + { + "bbox": [ + 61, + 157, + 549, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 157, + 549, + 172 + ], + "type": "text", + "content": "Shanghai Innovation Institute " + }, + { + "bbox": [ + 61, + 157, + 549, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 157, + 549, + 172 + ], + "type": "text", + "content": "University of Surrey" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 166, + 178, + 443, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 178, + 443, + 191 + ], + "spans": [ + { + "bbox": [ + 166, + 178, + 443, + 191 + ], + "type": "text", + "content": "https://github.com/fudan-zvg/tensoflow" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 219, + 200, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 219, + 200, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 219, + 200, + 232 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 245, + 297, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 297, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 297, + 591 + ], + "type": "text", + "content": "Inverse rendering aims to recover scene geometry, material properties, and lighting from multi-view images. Given the complexity of light-surface interactions, importance sampling is essential for the evaluation of the rendering equation, as it reduces variance and enhances the efficiency of Monte Carlo sampling. Existing inverse rendering methods typically use pre-defined non-learnable importance samplers in prior manually, struggling to effectively match the spatially and directionally varied integrand and resulting in high variance and suboptimal performance. To address this limitation, we propose the concept of learning a spatially and directionally aware importance sampler for the rendering equation to accurately and flexibly capture the unconstrained complexity of a typical scene. We further formulate TensoFlow, a generic approach for sampler learning in inverse rendering, enabling to closely match the integrand of the rendering equation spatially and directionally. Concretely, our sampler is parameterized by normalizing flows, allowing both directional sampling of incident light and probability density function (PDF) inference. To capture the characteristics of the sampler spatially, we learn a tensorial representation of the scene space, which imposes spatial conditions, together with reflected direction, leading to spatially and directionally aware sampling distributions. Our model can be optimized by minimizing the difference between the integrand and our normalizing flow. Extensive experiments validate the superiority of TensoFlow over prior alternatives on both synthetic and real-world benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 614, + 135, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 135, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 135, + 627 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "3D scene modeling has made significant strides, especially with the emergence of powerful representation models like neural radiance fields (NeRF) [2-4, 24] and efficient ones like 3D Gaussian splatting (3DGS) [13, 17, 38]. However, these methods struggle with inverse rendering due to the in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "type": "text", + "content": "ability to effectively disentangle geometry, material properties, and lighting, restricting their capacity to relight scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 244, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 460 + ], + "type": "text", + "content": "To address this issue, recent works [10, 11, 14, 15, 20, 22, 23, 35, 36] have advanced appearance modeling by incorporating the physically-based rendering equation [16] that simulates light-surface interactions. This however introduces a fundamental challenge - a need for accurately evaluating the rendering equation using Monte Carlo sampling. One intuitive method is stratified sampling over the hemisphere [15, 35], but this can be computationally inefficient due to the need for many samples. To improve efficiency, leading methods like NeRO [23] and TensoSDF [20] employ importance sampling. These approaches utilize pre-defined importance samplers (e.g., cosine-weighted and GGX samplers [30] for diffuse and specular components, respectively). However, these pre-defined importance samplers still result in high variance, as the integrand typically varies in both space and direction, making it difficult for a fixed sampler to effectively match the scene's characteristics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 688 + ], + "type": "text", + "content": "To address this challenge, we propose a novel perspective of learning spatially and directionally aware importance samplers, allowing us to manage the complexity of any scene without the need for manual sampler selection, while benefiting from greater flexibility and generality. As a showcase, we formulate TensoFlow, a concrete approach for importance sampler learning. Our sampler is parameterized by normalizing flows composed with piecewise-quadratic coupling layers [25], functioning as a neural distribution that enables both probability density function (PDF) inference and directional sampling of incident light in the rendering equation. To effectively capture the characteristics of the sampler spatially, we learn a tensorial representation of the scene space, imposing spatial conditions on the rendering process, together with reflected direction, resulting in spatially and directionally aware sampling distributions. This formulation can be optimized by minimizing the distributional difference between the integrand and our normalizing flow." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "We make these contributions: (i) We propose a novel concept of learning the importance samplers of the render-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 257, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 257, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 257, + 713 + ], + "type": "text", + "content": "*Corresponding author (lizhangfd@fudan.edu.cn)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "495" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "ing equation for inverse rendering, in contrast to existing pre-defined non-learnable samplers struggling intrinsically to capture the intrinsic rendering complexity of a scene. (ii) We introduce an effective method for sampler learning, TensoFlow, parameterized by tensorial normalizing flows, serving as a neural distribution that can support both directional sampling of incident light and PDF inference required in the rendering equation. (iii) Extensive experiments on both synthetic and real-world datasets demonstrate that TensoFlow outperforms previous alternatives, achieving significantly lower variance in evaluating the rendering equation with the same number of sampled directions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 225, + 140, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 140, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 140, + 237 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 245, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 473 + ], + "type": "text", + "content": "Novel view synthesis The introduction of neural radiance fields (NeRF) [24], which represent 3D scenes as continuous volumetric functions, utilizes volumetric rendering techniques to synthesize images from novel viewpoints, achieving effective 3D scene reconstruction. However, NeRF exhibits limitations in terms of rendering efficiency. Various efficient alternatives have been proposed, including voxel-based methods [9, 27], multi-resolution hash grids [26], and tensor decomposition approaches [6]. Additionally, techniques such as anti-aliasing [2-4] and modeling with signed distance fields (SDFs) [21, 31, 32, 34, 37] have been employed to enhance rendering and geometry reconstruction quality. Recently, 3D Gaussian splatting (3DGS) [17] has been introduced with favored efficiency, representing 3D scenes as a collection of 3D Gaussians. However, despite its impressive performance in rendering quality and efficiency, 3DGS lacks robust geometry, which limits its effectiveness in geometry-sensitive tasks such as inverse rendering [10, 22]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "content": "Inverse rendering Inverse rendering aims to recover the 3D structure, lighting conditions, and material properties of a scene from posed 2D images. Recent advancements in neural rendering have substantially driven inverse rendering performance. NeRF-based approaches [1, 5, 15, 18, 20, 23, 35, 39, 41] simulate physical light-surface interactions, allowing for the disentanglement of scene geometry, materials, and lighting from image observations. In particular, Monte Carlo sampling is commonly used to evaluate the rendering equation in inverse rendering. For example, avoiding ray marching on incident rays, NeILF [35] optimizes an incident light field using stratified sampling (e.g., sampling directions uniformly over the hemisphere) for the rendering equation. Similarly, TensoIR [15] also applies stratified sampling during training using tensorial implicit fields. However, stratified sampling can be computationally costly due to the high sample count needed for accuracy. To address this issue, NeRO [23] and TensoSDF [20] adopt importance sampling with pre-defined samplers (e.g., cosine" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "weighted, GGX). Commonly, all these sampling methods fall short of matching the spatially and directionally varying integrand of the rendering equation, leading to suboptimal rendering performance. To overcome this limitation, we introduce the concept of learning importance samplers that addressing the need for spatially and directionally aware samplers elegantly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 167, + 391, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 391, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 391, + 181 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 186, + 553, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 210 + ], + "type": "text", + "content": "We begin with necessary background on importance sampling and the rendering equation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "text", + "content": "Importance sampling Monte Carlo sampling is a commonly used method for evaluating integrals [12]. The principal challenge lies in approximating the integral with fewer samples while achieving reduced variance. Given an integral " + }, + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "inline_equation", + "content": "F = \\int_{\\Omega} f(x) \\, \\mathrm{d}x" + }, + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "text", + "content": " is the target integrand function over the domain " + }, + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 211, + 555, + 295 + ], + "type": "text", + "content": " is a sample, importance sampling is performed by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 323, + 304, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 304, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 323, + 304, + 553, + 346 + ], + "type": "interline_equation", + "content": "F = \\int_ {\\Omega} \\frac {f (x)}{q (x)} q (x) \\mathrm {d} x = \\mathbb {E} \\left[ \\frac {f (X)}{q (X)} \\right] \\approx \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {f \\left(X _ {i}\\right)}{q \\left(X _ {i}\\right)}, \\tag {1}", + "image_path": "d3596ccbf395c56afffd300eab9bc36144d3181bcc0dea8f24a0f52cc1c9d828.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "text", + "content": "where the proxy PDF " + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "text", + "content": " is designed to approximate the distribution of " + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "text", + "content": " in a way that minimizes the variance of the estimator. Ideally, " + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "text", + "content": " should be chosen so that it matches the shape of " + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "text", + "content": " closely, meaning that areas where " + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 313, + 346, + 554, + 430 + ], + "type": "text", + "content": " has higher values are sampled more frequently. This helps in achieving a more accurate estimate of the integral with fewer samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 430, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 501 + ], + "type": "text", + "content": "**Rendering equation** The rendering equation [16] models the interaction between lighting and surfaces. Given the incident radiance " + }, + { + "bbox": [ + 313, + 430, + 555, + 501 + ], + "type": "inline_equation", + "content": "L_{i}(\\omega_{i}, x)" + }, + { + "bbox": [ + 313, + 430, + 555, + 501 + ], + "type": "text", + "content": " and the material properties at a surface point " + }, + { + "bbox": [ + 313, + 430, + 555, + 501 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 430, + 555, + 501 + ], + "type": "text", + "content": " with normal vector " + }, + { + "bbox": [ + 313, + 430, + 555, + 501 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 430, + 555, + 501 + ], + "type": "text", + "content": ", the rendering equation expresses the outgoing radiance in the following manner:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 509, + 553, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 509, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 320, + 509, + 553, + 534 + ], + "type": "interline_equation", + "content": "L _ {o} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\int_ {\\Omega} f \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {i} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}\\right) d \\boldsymbol {\\omega} _ {i}, \\tag {2}", + "image_path": "15bae27d09dae5071c6501ee645abca03f71b4a4911ea4d5396bf1c5a17f0204.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "inline_equation", + "content": "f(\\omega_o, \\omega_i, x)" + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "text", + "content": " is the Bidirectional Reflectance Distribution Function (BRDF). Here, we consider the microfacet BRDF parameterization [28], including material properties such as albedo " + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "inline_equation", + "content": "a \\in [0,1]^3" + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "text", + "content": ", metallic " + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "inline_equation", + "content": "m \\in [0,1]" + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "text", + "content": ", and roughness " + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "inline_equation", + "content": "r \\in [0,1]" + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "text", + "content": ". Formally, this BRDF can be decomposed into a diffuse term " + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "inline_equation", + "content": "f_d" + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "text", + "content": " and a specular term " + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "inline_equation", + "content": "f_s" + }, + { + "bbox": [ + 313, + 542, + 554, + 615 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 398, + 622, + 555, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 622, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 398, + 622, + 555, + 648 + ], + "type": "interline_equation", + "content": "f _ {d} = \\frac {(1 - m) \\boldsymbol {a}}{\\pi}, \\tag {3}", + "image_path": "3d993a04a2e286a996e56f38a989676d30ca2a9bd9770f5dd0367b323ffbf8f4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 659, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 659, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 361, + 659, + 553, + 685 + ], + "type": "interline_equation", + "content": "f _ {s} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\frac {D F G}{4 \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}\\right) \\left(\\boldsymbol {\\omega} _ {o} \\cdot \\mathbf {n}\\right)}, \\tag {4}", + "image_path": "e355abbda0afa032df86de9c105270842b4f20dc02a453df576cbd09ad0acea0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "D, F" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " represent the normal distribution function, the Fresnel term, and the geometry term, respectively." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "496" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 387, + 235 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 387, + 235 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 387, + 235 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 387, + 235 + ], + "type": "image", + "image_path": "e2067f8e11fc6c70662adc07b2f5fe6806dcb0c852ab46122c0930417777148f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "text", + "content": "Figure 1. Left: Material and lighting estimation stage of TensoFlow. Given a ray-surface intersection point " + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "text", + "content": ", we use tensorial encoders to encode latent features for both the material properties and the importance sampler. When evaluating the rendering equation, incident directions are sampled from our learnable importance sampler, which is a frozen copy of the training normalizing flow. The normalizing flow is optimized by minimizing the distribution difference " + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "q(\\omega_{i})" + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "text", + "content": " and the integrand. The material properties, parameterized by " + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\theta_{m}" + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "text", + "content": ", are optimized by minimizing the RGB rendering loss " + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "text", + "content": " of the final rendering integral. Right: Tensorial normalizing flow. With spatial prior " + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "V_{f}" + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "text", + "content": " and directional prior " + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\omega_{r}" + }, + { + "bbox": [ + 54, + 249, + 555, + 327 + ], + "type": "text", + "content": ", our tensorial normalizing flow, implemented using piecewise-quadratic coupling layers, enables both incident direction sampling and PDF querying." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 394, + 72, + 554, + 235 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 554, + 235 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 554, + 235 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 554, + 235 + ], + "type": "image", + "image_path": "fb272017abde4db394c69d84fe83972c2220f11c5535bc9a6497a873a09a6d3e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 346, + 130, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 130, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 130, + 358 + ], + "type": "text", + "content": "4. TensoFLow" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 367, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 464 + ], + "type": "text", + "content": "Architecture overview Our TensoFlow is comprised of two stages: (I) The first stage involves reconstructing the scene geometry. We follow TensoSDF [20] and utilize a tensorial signed distance field combined with a roughness-aware training strategy to achieve robust geometry reconstruction. (II) The second stage focuses on modeling the material and lighting, which is central to inverse rendering and constitutes the core of our approach, as detailed below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 230, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 230, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 230, + 485 + ], + "type": "text", + "content": "4.1. Material and lighting estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": "Parametrization In TensoFlow, we represent the material of a scene with a Vector-Matrix tensorial encoder [6], due to its high capacity and flexibility. For a surface point " + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 491, + 296, + 538 + ], + "type": "text", + "content": ", its tensorial feature vector is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 547, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 547, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 61, + 547, + 295, + 563 + ], + "type": "interline_equation", + "content": "V _ {m} (\\boldsymbol {x}) = \\mathrm {v} _ {m, k} ^ {X} \\circ \\mathrm {M} _ {m, k} ^ {Y Z} \\oplus \\mathrm {v} _ {m, k} ^ {Y} \\circ \\mathrm {M} _ {m, k} ^ {X Z} \\oplus \\mathrm {v} _ {m, k} ^ {Z} \\circ \\mathrm {M} _ {m, k} ^ {X Y}, \\tag {5}", + "image_path": "0246ee39e3349aa780b0bd35b4b4b501ff5e5a38f63ce0329c0b1eee75dd1135.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_k" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathrm{M}_k" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": "-th vector and matrix along the corresponding axis. The operators " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": " denote element-wise multiplication and concatenation, respectively. After concatenating this feature vector with the positional encoding " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": " of the position " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": ", it is processed by a tiny MLP " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\Theta_m" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": " to produce the material properties:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 123, + 654, + 295, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 654, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 123, + 654, + 295, + 668 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {a}, m, r \\right\\} = \\Theta_ {m} \\left(V _ {m}, \\mathbf {p}\\right), \\tag {6}", + "image_path": "6bb0d299fff651c40dc83b3c983542885cc1e0ed4ef5af7332b00c5ec1dfccde.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " denote albedo, metallic, and roughness, respectively. For the incident light modeling, we decompose it into spatially-aware indirect lighting and direct light-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 347, + 432, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 432, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 432, + 360 + ], + "type": "text", + "content": "ing originating from infinity:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 340, + 366, + 553, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 366, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 340, + 366, + 553, + 380 + ], + "type": "interline_equation", + "content": "L \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) = V \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {\\operatorname {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right) + L _ {\\operatorname {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right), \\tag {7}", + "image_path": "7d66dbbff7ede6bff83c66664782aa6a006b3655379b84d1de7aeb48d730fb1c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "content": "where the visibility term " + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "inline_equation", + "content": "V(\\omega_{i},\\pmb {x})" + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "content": " is determined via ray tracing on the extracted mesh. The indirect and direct lightings are modeled by a tiny MLP " + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{ind}}" + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "content": " and an environment map " + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "content": ", respectively:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 440, + 553, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 440, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 376, + 440, + 553, + 453 + ], + "type": "interline_equation", + "content": "L _ {\\text {i n d}} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) = \\Theta_ {\\text {i n d}} \\left(\\mathbf {p}, \\boldsymbol {\\omega} _ {i}\\right), \\tag {8}", + "image_path": "23654bdbf0d361aa7e4afc99680d23156b3c6db2070ecf8f8bc630188c2dd0b1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 389, + 455, + 553, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 455, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 389, + 455, + 553, + 468 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right) = \\Theta_ {\\mathrm {d i r}} \\left(\\boldsymbol {\\omega} _ {i}\\right), \\tag {9}", + "image_path": "aa6fa5d98824776d4602e487baf9dde1ce2d5715002b512c415a91b00c4dabfd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "text", + "content": "**Rendering** Given the reconstructed signed distance field from the first stage, the intersection point between each ray and the surface can be determined. The material properties and incident lighting at this intersection can then be queried as described earlier. In accordance with Eq. (2), we employ Monte Carlo sampling to evaluate the integral in the rendering equation. Due to differing characteristics of the diffuse and specular terms, separate importance samplers are used to minimize variance in the Monte Carlo sampling process:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 325, + 604, + 553, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 325, + 604, + 553, + 635 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} _ {\\text {d i f f u s e}} = \\frac {1}{N _ {d}} \\sum_ {i} ^ {N _ {d}} \\frac {(1 - m) \\frac {a}{\\pi} L (\\boldsymbol {\\omega} _ {i} , \\boldsymbol {x}) (\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n})}{q _ {\\mathrm {d}} (\\boldsymbol {\\omega} _ {i} ; \\boldsymbol {x})}, \\tag {10}", + "image_path": "d1b12bf43a9badc878af7a4fcd7a370dfe8202e205778d3c5108e5521675286a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 639, + 553, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 639, + 553, + 671 + ], + "spans": [ + { + "bbox": [ + 320, + 639, + 553, + 671 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} _ {\\text {s p e c u l a r}} = \\frac {1}{N _ {s}} \\sum_ {i} ^ {N _ {s}} \\frac {\\frac {D F G}{4 (\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}) (\\boldsymbol {\\omega} _ {o} \\cdot \\boldsymbol {n})} L (\\boldsymbol {\\omega} _ {i} , \\boldsymbol {x}) (\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n})}{q _ {\\mathrm {s}} (\\boldsymbol {\\omega} _ {i} ; \\boldsymbol {x})}, \\tag {11}", + "image_path": "9518fec14fe4d8b24894f56bf061f30f82ae38f489c96844ac3fac7b43e22586.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{d}}(\\omega_i)" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{s}}(\\omega_i)" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " are the PDFs of the importance samplers, which we model them using the proposed normalizing flow (Sec. 4.2). The final color is computed as:" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "497" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 291, + 224 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 291, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 291, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 291, + 224 + ], + "type": "image", + "image_path": "b6da1d6c84fa578c32df647b32cd8dd88c6ca0f944dc05c8cae55bf4cf9cd831.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 295, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 295, + 273 + ], + "type": "text", + "content": "Figure 2. Illustration of a piecewise-quadratic coupling layer, incorporating tensorial latent feature " + }, + { + "bbox": [ + 55, + 238, + 295, + 273 + ], + "type": "inline_equation", + "content": "V_{f}" + }, + { + "bbox": [ + 55, + 238, + 295, + 273 + ], + "type": "text", + "content": " and reflected direction " + }, + { + "bbox": [ + 55, + 238, + 295, + 273 + ], + "type": "inline_equation", + "content": "\\omega_{r}" + }, + { + "bbox": [ + 55, + 238, + 295, + 273 + ], + "type": "text", + "content": " as spatial and directional priors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 306, + 295, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 306, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 126, + 306, + 295, + 319 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} = \\boldsymbol {c} _ {\\text {d i f f u s e}} + \\boldsymbol {c} _ {\\text {s p e c u l a r}}, \\tag {12}", + "image_path": "de69fe11d10890c49c757536c19a02c4404ee79496cb459b7d726ddd42fc5bbb.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 326, + 277, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 277, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 277, + 339 + ], + "type": "text", + "content": "4.2. Tensorial flow-based importance sampling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 344, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 452 + ], + "type": "text", + "content": "For evaluating the rendering equation, prior works often employ pre-defined distributions as the importance samplers. For instance, in NeRO [23] a cosine-weighted distribution is used as the importance sampler for the diffuse term, and the GGX distribution [7] for the specular term. However, such samplers cannot closely match the shape of the target integrand with complex spatial and directional variations for a typical scene, leading to high variance and performance degradation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 174, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 174, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 174, + 480 + ], + "type": "text", + "content": "4.2.1 Flow-based sampler" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "text", + "content": "To overcome this limitation, we propose learning the importance samplers. The normalizing flow framework [8] is adopted as the sampler learner due to its following theoretical advantage. For a given direction " + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "text", + "content": " with two degrees of freedom in spherical coordinates, the distribution of the integrand function may exhibit inter-dependencies across different coordinate dimensions. We can express the variable " + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "text", + "content": " as a bijective transformation of an independent variable " + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "text", + "content": ", which follows a uniform distribution " + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "inline_equation", + "content": "(0,1)^{2}" + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\omega = h(z)" + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "text", + "content": ". Using the change-of-variable formula, the PDF of " + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 487, + 296, + 618 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 627, + 295, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 627, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 107, + 627, + 295, + 657 + ], + "type": "interline_equation", + "content": "q (\\boldsymbol {\\omega}) = q (\\boldsymbol {z}) \\left| \\det \\left(\\frac {\\partial h (\\boldsymbol {z})}{\\partial \\boldsymbol {z} ^ {T}}\\right) \\right| ^ {- 1}, \\tag {13}", + "image_path": "77b70a23ee49b91dc422ab9b9e9c047a0dbf0413bc27e7aef4f5cf36347c722d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "q(z)" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": " represents the PDF of " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(0,1)^2" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": ", which is straightforward to evaluate." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "However, the selection of the bijective mapping " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " must satisfy certain conditions: it must be invertible, and it" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "should allow efficient computation of the Jacobian determinant. To meet these requirements, we employ NICE [8], a type of autoregressive flow that utilizes \"coupling layers\" to enable fast sampling and density evaluation. The final mapping " + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": " is then formed as a composition of simpler bijective transformations: " + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "inline_equation", + "content": "h = h_{n} \\circ \\dots \\circ h_{1}" + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": "Given an input vector " + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "inline_equation", + "content": "x = (x_{1},x_{2})\\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": " for a mapping " + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": " , the output vector " + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "inline_equation", + "content": "y = (y_{1},y_{2})\\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 176, + 553, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 176, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 361, + 176, + 553, + 190 + ], + "type": "interline_equation", + "content": "y _ {1} = x _ {1}, \\quad y _ {2} = C (x _ {2}; m (x _ {1})), \\tag {14}", + "image_path": "3dba05dbecf162bdc179ae7f60a31cc113bd10b37e38a711c83defd24fc001c8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "text", + "content": "where the coupling transform " + }, + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "inline_equation", + "content": "C:\\mathbb{R}\\leftrightarrow \\mathbb{R}" + }, + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "text", + "content": " is a separable and invertible transformation, and " + }, + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "inline_equation", + "content": "m(x)" + }, + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "text", + "content": ", parameterized by a neural network, acts as the parameter of " + }, + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 198, + 554, + 258 + ], + "type": "text", + "content": ". This structure allows the inverse of the coupling layer to be computed directly:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 266, + 553, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 266, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 356, + 266, + 553, + 280 + ], + "type": "interline_equation", + "content": "x _ {1} = y _ {1}, \\quad x _ {2} = C ^ {- 1} \\left(y _ {2}; m \\left(y _ {1}\\right)\\right). \\tag {15}", + "image_path": "bcc552b3597c752e2168208a47a72d4c6bdfe6c63f343dc769b78aa8d907436b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "content": "The separability of " + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "content": " ensures that the Jacobian matrix " + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "inline_equation", + "content": "J_{i}" + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "inline_equation", + "content": "h_{i}" + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "content": " is triangular, simplifying the computation of its determinant to the product of its diagonal elements." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "text", + "content": "By applying the change-of-variable formula, the Jacobian determinant of the composite mapping " + }, + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "text", + "content": " is expressed as the product of the individual " + }, + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "inline_equation", + "content": "J_{i}" + }, + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "inline_equation", + "content": "|\\operatorname{det} J| = \\prod_{i} |\\operatorname{det} J_{i}|" + }, + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "text", + "content": ". We note that, with sufficiently expressive coupling transformations and adequate coupling layers, the sampling distribution " + }, + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 323, + 554, + 453 + ], + "type": "text", + "content": " can be modeled as an arbitrarily complex function, allowing it to closely match the shape of any integrand in the rendering equation. Importantly, each adjacent coupling layer preserves different dimensions of the input vector (Eq. (14)), ensuring the mapping's expressiveness across all the dimensions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 455, + 554, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 526 + ], + "type": "text", + "content": "The normalizing flow defined by the above mapping acts as a distribution from which we can query the PDF or draw samples, corresponding to " + }, + { + "bbox": [ + 313, + 455, + 554, + 526 + ], + "type": "inline_equation", + "content": "h^{-1}" + }, + { + "bbox": [ + 313, + 455, + 554, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 455, + 554, + 526 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 455, + 554, + 526 + ], + "type": "text", + "content": ", respectively. Our method supports two primary applications in the inverse rendering pipeline: inferring the PDF value for a given direction and sampling a direction along with its PDF value:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 330, + 547, + 553, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 547, + 553, + 560 + ], + "spans": [ + { + "bbox": [ + 330, + 547, + 553, + 560 + ], + "type": "text", + "content": "Inference: " + }, + { + "bbox": [ + 330, + 547, + 553, + 560 + ], + "type": "inline_equation", + "content": "q(\\omega)\\gets h^{-1}(\\omega)," + }, + { + "bbox": [ + 330, + 547, + 553, + 560 + ], + "type": "text", + "content": " (16)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 330, + 563, + 553, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 563, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 330, + 563, + 553, + 577 + ], + "type": "text", + "content": "Sampling: " + }, + { + "bbox": [ + 330, + 563, + 553, + 577 + ], + "type": "inline_equation", + "content": "(\\omega ,q(\\omega))\\gets h(\\pmb {u}),\\pmb {u}\\sim \\mathcal{U}(0,1)^2," + }, + { + "bbox": [ + 330, + 563, + 553, + 577 + ], + "type": "text", + "content": " (17)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "text", + "content": "Our experiments reveal that modeling the distribution of the half vector " + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "inline_equation", + "content": "\\omega_{h}" + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "text", + "content": " yields better performance compared to directly modeling " + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "text", + "content": ". By modeling the distribution " + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "inline_equation", + "content": "\\omega_{h}" + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "inline_equation", + "content": "q(\\omega_{i})" + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "text", + "content": " can be derived using the change-of-variable formula (Eq. (13)), enabling both PDF inference and sampling of " + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 313, + 584, + 554, + 656 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 670, + 493, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 493, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 493, + 682 + ], + "type": "text", + "content": "4.2.2 Tensorial coupling transformation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Given the spatially-dependent nature of the integrand in the rendering equation, we further parameterize the scene space" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "498" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 289, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 289, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 289, + 85 + ], + "type": "text", + "content": "by incorporating an additional tensorial representation as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 92, + 295, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 92, + 295, + 110 + ], + "spans": [ + { + "bbox": [ + 61, + 92, + 295, + 110 + ], + "type": "interline_equation", + "content": "V _ {f} (\\boldsymbol {x}) = \\mathrm {v} _ {f, k} ^ {X} \\circ \\mathrm {M} _ {f, k} ^ {Y Z} \\oplus \\mathrm {v} _ {f, k} ^ {Y} \\circ \\mathrm {M} _ {f, k} ^ {X Z} \\oplus \\mathrm {v} _ {f, k} ^ {Z} \\circ \\mathrm {M} _ {f, k} ^ {X Y}. \\tag {18}", + "image_path": "4ee467d2305e314edfdf6a8ea4d9eb9f2039ef72b055ab0a11fd5e94006d87ff.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "text", + "content": "The latent feature " + }, + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "inline_equation", + "content": "V_{f}" + }, + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "text", + "content": ", together with the reflected direction " + }, + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "inline_equation", + "content": "\\omega_{r} = 2(\\omega_{o}\\cdot n)n - \\omega_{o}" + }, + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "text", + "content": " for directional dependency, serves as conditions to guide the network " + }, + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "text", + "content": " within each coupling layer. This is crucial, as the shape of the integrand in the rendering equation is influenced by both the spatial position and the outgoing direction " + }, + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 55, + 118, + 296, + 189 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": "We consider the piecewise-quadratic coupling transform [25] as our coupling transformation " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " for its high expressiveness. Specifically, it achieves the bijective property by defining a Cumulative Distribution Function (CDF): " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "y = C(x; m(x'; V_f, \\omega_r)) = \\int_0^x p(t; m(x'; V_f, \\omega_r)) \\, \\mathrm{d}t" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " is a spatially and directionally aware PDF derived from the output of network " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": ". This results in the Jacobian determinant of the coupling layer being equal to " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": ". Here, the network " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " takes the concatenation of " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "x', V_f" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "\\omega_r" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " as input, generating " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "\\widehat{V}" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "\\widehat{W}" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " for subsequent use. We then construct the piecewise linear PDF using " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " vertices, as illustrated in Fig. 2. It employs " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " vertical coordinates " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "\\widehat{V} \\in \\mathbb{R}^{K + 1}" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " horizontal differences " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "\\widehat{W} \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " as the predefined bin count for the distribution. Following softmax normalization, " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "W = \\sigma(\\widehat{W})" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " represents the bin widths, while " + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 190, + 296, + 385 + ], + "type": "text", + "content": " defines the PDF values at each vertex." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 394, + 295, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 394, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 100, + 394, + 295, + 435 + ], + "type": "interline_equation", + "content": "V _ {i} = \\frac {\\exp (\\widehat {V} _ {k})}{\\sum_ {k = 1} ^ {K} \\frac {\\exp (\\widehat {V} _ {k}) + \\exp (\\widehat {V} _ {k + 1})}{2} W _ {k}}, \\tag {19}", + "image_path": "1897d458a184ec050f26f611f9fc3958f325b3433c62e4c8c9347fc88e119a00.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 442, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 296, + 466 + ], + "type": "text", + "content": "Consequently, given an input " + }, + { + "bbox": [ + 55, + 442, + 296, + 466 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 442, + 296, + 466 + ], + "type": "text", + "content": ", the corresponding PDF value can be queried as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 477, + 294, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 477, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 120, + 477, + 294, + 491 + ], + "type": "interline_equation", + "content": "p (x) = \\operatorname {l e r p} \\left(V _ {b}, V _ {b + 1}, \\alpha\\right), \\tag {20}", + "image_path": "6dc3458ef6e68dbaf61d22cbc271982f8e109843f6d0ac42c58ab7f74b46a12b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 500, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 500, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\alpha = (x - \\sum_{k=1}^{b-1} W_k) / W_k" + }, + { + "bbox": [ + 55, + 500, + 296, + 538 + ], + "type": "text", + "content": " defines the relative position of " + }, + { + "bbox": [ + 55, + 500, + 296, + 538 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 500, + 296, + 538 + ], + "type": "text", + "content": " within the corresponding bin " + }, + { + "bbox": [ + 55, + 500, + 296, + 538 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 500, + 296, + 538 + ], + "type": "text", + "content": ". The piecewise-quadratic CDF can be computed by integrating the PDF:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 548, + 296, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 590 + ], + "type": "interline_equation", + "content": "C (x) = \\frac {\\alpha W _ {b}}{2} ((2 - \\alpha) V _ {b} + \\alpha V _ {b + 1}) + \\sum_ {k = 1} ^ {b - 1} \\frac {V _ {k} + V _ {k + 1}}{2} W _ {k}. \\tag {21}", + "image_path": "bd8df2e86edd992f70d7ee1f58a9515a9800ff97bb0f3bafda9f628ac60a69c6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "content": "We employ " + }, + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "content": " coupling layers with piecewise-quadratic transformation to construct the compound mapping " + }, + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "content": ", ensuring it is sufficiently expressive for modeling the importance sampler." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 119, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 119, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 119, + 659 + ], + "type": "text", + "content": "4.3. Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "During model training, we optimize the tensorial encoder " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "V_{f}" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": " and the networks " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\{m_i\\}_{i = 1\\dots N}" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": " by minimizing the cross-entropy loss between the integrand of the rendering equation and the PDF of the incident directions inferred from" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 403, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 403, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 403, + 84 + ], + "type": "text", + "content": "the normalizing flow:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 352, + 91, + 553, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 91, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 352, + 91, + 553, + 144 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {c e}} = - \\int_ {\\Omega} I (\\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}) \\log q (\\boldsymbol {\\omega} _ {i}) \\mathrm {d} \\boldsymbol {\\omega} _ {i} (22) \\\\ = \\mathbb {E} \\left[ - \\frac {I \\left(\\boldsymbol {\\omega} _ {i} , \\boldsymbol {\\omega} _ {o} , \\boldsymbol {x}\\right)}{\\hat {q} \\left(\\boldsymbol {\\omega} _ {i}\\right)} \\log q \\left(\\boldsymbol {\\omega} _ {i}\\right) \\right], (23) \\\\ \\end{array}", + "image_path": "b696227933dc1460646e49080f8e14072821718ba37801e3aaf26e2f7e80d985.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "inline_equation", + "content": "\\hat{q} (\\omega_{i})" + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "text", + "content": " represents the PDF of the sampled incident directions from the current importance sampler, a frozen copy of the trainable normalizing flow updated every " + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{update}}" + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "text", + "content": " iterations. " + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "inline_equation", + "content": "q(\\omega_{i})" + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "text", + "content": " denotes the PDF inferred from the current training normalizing flow, and " + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "inline_equation", + "content": "I(\\omega_i,\\omega_o,x)" + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "text", + "content": " is the integrand function evaluated with the sampled incident direction. We use different importance samplers for the diffuse and specular terms, resulting in separate losses: " + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 152, + 555, + 259 + ], + "type": "text", + "content": ". The total training loss is thus defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 355, + 266, + 553, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 266, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 355, + 266, + 553, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c} + \\lambda_ {m} \\mathcal {L} _ {m} + \\lambda_ {\\mathrm {c e}} ^ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {d}} + \\lambda_ {\\mathrm {c e}} ^ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {c e}} ^ {\\mathrm {s}}, \\tag {24}", + "image_path": "976670e69558ebd86cf0f70b36f47056062fb647fab450f0ce43c80e197c367e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "text", + "content": " represent the RGB rendering loss and the material regularization loss [23], respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 323, + 390, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 390, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 390, + 335 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 342, + 553, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 403 + ], + "type": "text", + "content": "Datasets and metrics We use the TensoSDF synthetic dataset [20] and the real-world Stanford-ORB dataset [19] for quantitative evaluation. To quantitatively evaluate the estimated material, we adopt three standard metrics for relighted images: PSNR, SSIM [33], and LPIPS [40]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 403, + 553, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 474 + ], + "type": "text", + "content": "Competitors We compare our approach with state-of-the-art NeRF-based methods selected for their different sampling strategies in evaluating the rendering equation: TensoIR [15] and NeILF++[39], which utilize stratified sampling, and NeRO[23] and TensoSDF [20], which employ pre-defined importance sampling." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": "Implementation details The training process consists of two stages. For the first stage, geometry reconstruction, we train the TensoSDF model for 180,000 iterations using the same settings as in TensoSDF [20]. For the second stage, material and lighting estimation, we begin with a warm-up phase, employing cosine-weighted and GGX distributions as the importance samplers for the diffuse and specular components, respectively, over " + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{warmup}} = 1,000" + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": " iterations. Once a coarse estimation of material and lighting is achieved at " + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{ce}} = 500" + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": " iterations, we start optimizing the flow-related network by minimizing the cross-entropy losses " + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": ". These flow-related networks include the tensorial encoder and the coupling layer network " + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{m_i\\}_{i=1\\dots N}^d" + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{m_i\\}_{i=1\\dots N}^s" + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": " for diffuse and specular respectively, each consists of a composition of " + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": " piecewise-quadratic coupling layers. We model each " + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": " using a three-layer tiny MLP with 64 hidden unit. After the warm-up phase, the importance sampler is set as a frozen copy of the current normalizing flow, updated every " + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{update}} = 1,000" + }, + { + "bbox": [ + 313, + 475, + 555, + 715 + ], + "type": "text", + "content": " iterations. After the warm-up phase, we" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "499" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 549, + 304 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 549, + 304 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 549, + 304 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 549, + 304 + ], + "type": "image", + "image_path": "070d5317c06de0bf3b2a4ebe2ad80d48249f11a97d1bc25399015c6b64bb247a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 316, + 451, + 328 + ], + "lines": [ + { + "bbox": [ + 157, + 316, + 451, + 328 + ], + "spans": [ + { + "bbox": [ + 157, + 316, + 451, + 328 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of relighting quality on TensoSDF dataset [20]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 340, + 552, + 596 + ], + "blocks": [ + { + "bbox": [ + 60, + 340, + 552, + 596 + ], + "lines": [ + { + "bbox": [ + 60, + 340, + 552, + 596 + ], + "spans": [ + { + "bbox": [ + 60, + 340, + 552, + 596 + ], + "type": "image", + "image_path": "6b3ac61acb56291ceceac4cc613a99d9d292ea88c039023e40e98955b25b995c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 608, + 555, + 641 + ], + "lines": [ + { + "bbox": [ + 55, + 608, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 555, + 641 + ], + "type": "text", + "content": "Figure 4. Visualization of material decomposition, normal map, and per-pixel variance in rendering equation evaluation, comparing results with a pre-defined sampler and our proposed tensorial normalizing flow. The per-pixel variance results are scaled by 1000 for clearer visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 662, + 296, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 296, + 711 + ], + "type": "text", + "content": "employ the proposed tensorial importance sampler to sample incident directions. We sample 128 rays for the specular term. For the diffuse term, we apply the MIS [29] technique, sampling 64 rays from another normalizing flow and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 662, + 555, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 555, + 711 + ], + "type": "text", + "content": "512 rays from a cosine-weighted distribution. " + }, + { + "bbox": [ + 313, + 662, + 555, + 711 + ], + "type": "inline_equation", + "content": "\\lambda_{m},\\lambda_{\\mathrm{ce}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 662, + 555, + 711 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 662, + 555, + 711 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ce}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 662, + 555, + 711 + ], + "type": "text", + "content": " are set to 1.0, 0.0001, and 0.0001, respectively. The training process takes 4 hours for the first stage and 2.5 hours for the second stage. All experiments are conducted" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "500" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 91, + 541, + 165 + ], + "blocks": [ + { + "bbox": [ + 157, + 71, + 454, + 83 + ], + "lines": [ + { + "bbox": [ + 157, + 71, + 454, + 83 + ], + "spans": [ + { + "bbox": [ + 157, + 71, + 454, + 83 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of relighting quality on TensoSDF dataset [20]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 91, + 541, + 165 + ], + "lines": [ + { + "bbox": [ + 69, + 91, + 541, + 165 + ], + "spans": [ + { + "bbox": [ + 69, + 91, + 541, + 165 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
TensoIR [15]NeRO [23]NeiLF++ [39]TensoSDF [20]TenSOFlowTensoIRNeRONeiLF++TensoSDFTenSOFlowTensoIRNeRONeiLF++TensoSDFTenSOFlow
Rover24.00024.01523.77426.75426.9360.9180.9140.9110.9350.9370.08010.06930.07540.05930.0566
Dragon25.10425.64424.09927.89927.9000.8950.9190.9010.9360.9370.13020.08980.09880.07750.0756
Motor19.21922.15820.14222.75425.5100.9060.9170.8940.9300.9440.08210.07020.08700.06810.0556
Helmet25.14022.58724.00128.12628.3850.9010.8810.9060.9340.9370.10400.10790.10560.07700.0707
Robot26.03123.19422.69626.24227.4160.9280.9130.9150.9400.9430.09310.07550.07820.06130.0577
Compressor20.75321.62419.74024.04925.4530.8680.8780.8440.9160.9290.10380.10730.12860.08300.0789
Average23.37523.20422.41025.97126.9330.9030.9040.8950.9320.9380.09890.08670.09560.07100.0659
", + "image_path": "9d3c846ff156af4bbfea330b4da568b9b28384f4652e56a719ac763b52f56ea8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 71, + 176, + 545, + 314 + ], + "blocks": [ + { + "bbox": [ + 71, + 176, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 71, + 176, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 71, + 176, + 545, + 314 + ], + "type": "image", + "image_path": "4c5f7d060f38ee92c3db04ee78d3ef10c3c5de19e3cf54a759818cea7c01c851.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 327, + 503, + 340 + ], + "lines": [ + { + "bbox": [ + 107, + 327, + 503, + 340 + ], + "spans": [ + { + "bbox": [ + 107, + 327, + 503, + 340 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of relighted images with TensoSDF [20] on the Stanford-ORB dataset [19]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 360, + 191, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 191, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 191, + 373 + ], + "type": "text", + "content": "on a single NVIDIA H100 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 383, + 195, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 195, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 195, + 396 + ], + "type": "text", + "content": "5.1. Results on synthetic data" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 297, + 715 + ], + "type": "text", + "content": "We first perform experiments on the TensoSDF synthetic dataset [20] to evaluate the effectiveness of our method in accurately reconstructing materials and lighting. Following the setups of NeRO [23] and TensoSDF [20], we extract each object's estimated material properties after optimization and use Blender to perform relighting based on the extracted mesh and material information. As shown in Tab. 1, our method consistently outperforms competitors across multiple scenes and metrics. This improvement highlights the benefit of our learnable importance sampler, which adapts to the integrand's spatial and directional complexity, reducing Monte Carlo sampling variance and enhancing material and lighting estimation accuracy. Notably, NeRO underperforms relative to TensoSDF, due to its less accurate geometry reconstruction, which affects subsequent material and lighting estimations. In contrast, our method, built upon TensoSDF's robust geometry, further refines material estimation with a learnable importance sampler, achieving a closer alignment with the integrand's complex distribution in the rendering equation. In Fig. 3, we provide a visual comparison of relighted images with NeRO and TensoSDF. Without a spatially and directionally aware importance sampler, the competitors rely on pre-defined sampling that struggles to match the shape of the integrand, leading to high variance and potential convergence to local minima during optimization." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 390, + 547, + 456 + ], + "blocks": [ + { + "bbox": [ + 313, + 358, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 380 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of relighting quality on Stanford-ORB dataset [19]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 390, + 547, + 456 + ], + "lines": [ + { + "bbox": [ + 321, + 390, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 321, + 390, + 547, + 456 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
TensoSDF [20]TenSOFlowTensoSDFTenSOFlowTenSOSDFTenSOFlow
Teapot29.97129.3420.9860.9860.02760.0250
Gnome28.48829.1790.9560.9600.09470.0899
Cactus31.35131.4960.9840.9850.03620.0366
Car28.01029.0350.9810.9840.03670.0306
Grogu30.59530.7350.9900.9900.03620.0385
Avg.29.68329.9570.9790.9810.04630.0441
", + "image_path": "26c797622183f8ec4f0e7b88894649db309fd9df10e866848330e805c801addf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 479, + 555, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 611 + ], + "type": "text", + "content": "In Fig. 4, we visualize the estimated materials rendered by TensoFlow alongside per-pixel variance in Monte Carlo sampling of specular term. \"Variance (pre-defined)\" shows results using fixed importance samplers, specifically a GGX distribution. By contrast, our tensorial normalizing flow effectively models the importance sampler, allowing it to dynamically adjust to the integrand's spatial and directional variations. This approach enables our method to produce detailed material maps—including albedo, metallic, and roughness, while significantly reducing variance with the same number of sampled rays." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 622, + 430, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 430, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 430, + 633 + ], + "type": "text", + "content": "5.2. Results on real data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "We further evaluate our method on the real-world Stanford-ORB dataset [19], following the setup used in TensoSDF [20]. Five objects from the dataset are selected for evaluation, and each object's mesh and material properties are extracted from the optimized model and relight under two new environment lighting conditions. As shown" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "501" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 76, + 114, + 123 + ], + "blocks": [ + { + "bbox": [ + 59, + 76, + 114, + 123 + ], + "lines": [ + { + "bbox": [ + 59, + 76, + 114, + 123 + ], + "spans": [ + { + "bbox": [ + 59, + 76, + 114, + 123 + ], + "type": "image", + "image_path": "979150cfadd75541c1aa0812b5c4d2b1468b953daf652d37c6b8f4b92ab6256a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 127, + 110, + 137 + ], + "lines": [ + { + "bbox": [ + 69, + 127, + 110, + 137 + ], + "spans": [ + { + "bbox": [ + 69, + 127, + 110, + 137 + ], + "type": "text", + "content": "w/o diffuse" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 78, + 173, + 123 + ], + "blocks": [ + { + "bbox": [ + 119, + 78, + 173, + 123 + ], + "lines": [ + { + "bbox": [ + 119, + 78, + 173, + 123 + ], + "spans": [ + { + "bbox": [ + 119, + 78, + 173, + 123 + ], + "type": "image", + "image_path": "a16aac9971513f5a1f751358f7429b28ffe39be98daba350e440bcba6db4c70e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 128, + 172, + 137 + ], + "lines": [ + { + "bbox": [ + 125, + 128, + 172, + 137 + ], + "spans": [ + { + "bbox": [ + 125, + 128, + 172, + 137 + ], + "type": "text", + "content": "w/o specular" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 78, + 230, + 123 + ], + "blocks": [ + { + "bbox": [ + 178, + 78, + 230, + 123 + ], + "lines": [ + { + "bbox": [ + 178, + 78, + 230, + 123 + ], + "spans": [ + { + "bbox": [ + 178, + 78, + 230, + 123 + ], + "type": "image", + "image_path": "f2a2e02dc675f8210b70e4f8c5c5d31e4074cfcdbbc7a0b1754f7d90e6888513.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 128, + 226, + 137 + ], + "lines": [ + { + "bbox": [ + 188, + 128, + 226, + 137 + ], + "spans": [ + { + "bbox": [ + 188, + 128, + 226, + 137 + ], + "type": "text", + "content": "TenSoFlow" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 78, + 291, + 123 + ], + "blocks": [ + { + "bbox": [ + 239, + 78, + 291, + 123 + ], + "lines": [ + { + "bbox": [ + 239, + 78, + 291, + 123 + ], + "spans": [ + { + "bbox": [ + 239, + 78, + 291, + 123 + ], + "type": "image", + "image_path": "ae89bdaf3181517febbd4aad0275fd0d32e6077dceff2afa1fffa1476f99eba2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 128, + 274, + 136 + ], + "lines": [ + { + "bbox": [ + 263, + 128, + 274, + 136 + ], + "spans": [ + { + "bbox": [ + 263, + 128, + 274, + 136 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 155, + 115, + 205 + ], + "blocks": [ + { + "bbox": [ + 87, + 141, + 247, + 151 + ], + "lines": [ + { + "bbox": [ + 87, + 141, + 247, + 151 + ], + "spans": [ + { + "bbox": [ + 87, + 141, + 247, + 151 + ], + "type": "text", + "content": "(a) Ablation studies on learnable importance samples" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 155, + 115, + 205 + ], + "lines": [ + { + "bbox": [ + 62, + 155, + 115, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 155, + 115, + 205 + ], + "type": "image", + "image_path": "a7fbdf035e441da548ceffc54f8d60a848fddbeb08c173fc2e7df83045d41f30.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 209, + 112, + 217 + ], + "lines": [ + { + "bbox": [ + 64, + 209, + 112, + 217 + ], + "spans": [ + { + "bbox": [ + 64, + 209, + 112, + 217 + ], + "type": "text", + "content": "w/o reflected" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 121, + 155, + 174, + 205 + ], + "blocks": [ + { + "bbox": [ + 121, + 155, + 174, + 205 + ], + "lines": [ + { + "bbox": [ + 121, + 155, + 174, + 205 + ], + "spans": [ + { + "bbox": [ + 121, + 155, + 174, + 205 + ], + "type": "image", + "image_path": "89109be41d83ec3c8f4232eb706aed4f7f557670ee23c39e3a753c05fe5ddc71.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 209, + 172, + 217 + ], + "lines": [ + { + "bbox": [ + 124, + 209, + 172, + 217 + ], + "spans": [ + { + "bbox": [ + 124, + 209, + 172, + 217 + ], + "type": "text", + "content": "w/o tensorial" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 181, + 155, + 231, + 205 + ], + "blocks": [ + { + "bbox": [ + 181, + 155, + 231, + 205 + ], + "lines": [ + { + "bbox": [ + 181, + 155, + 231, + 205 + ], + "spans": [ + { + "bbox": [ + 181, + 155, + 231, + 205 + ], + "type": "image", + "image_path": "3471a206be754fbd9615f85e69d33486cbba339759d347c2aa0b42d39655a224.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 209, + 225, + 217 + ], + "lines": [ + { + "bbox": [ + 187, + 209, + 225, + 217 + ], + "spans": [ + { + "bbox": [ + 187, + 209, + 225, + 217 + ], + "type": "text", + "content": "TenSoFlow" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 242, + 155, + 293, + 205 + ], + "blocks": [ + { + "bbox": [ + 242, + 155, + 293, + 205 + ], + "lines": [ + { + "bbox": [ + 242, + 155, + 293, + 205 + ], + "spans": [ + { + "bbox": [ + 242, + 155, + 293, + 205 + ], + "type": "image", + "image_path": "158608b81b1f3b1fd1973e6cfe7f4928538a1593e3aefc22a75027d8c6c7a7e9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 209, + 273, + 217 + ], + "lines": [ + { + "bbox": [ + 262, + 209, + 273, + 217 + ], + "spans": [ + { + "bbox": [ + 262, + 209, + 273, + 217 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 60, + 236, + 115, + 284 + ], + "blocks": [ + { + "bbox": [ + 89, + 221, + 261, + 232 + ], + "lines": [ + { + "bbox": [ + 89, + 221, + 261, + 232 + ], + "spans": [ + { + "bbox": [ + 89, + 221, + 261, + 232 + ], + "type": "text", + "content": "(b) Ablation studies on spatial and directional priors." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 236, + 115, + 284 + ], + "lines": [ + { + "bbox": [ + 60, + 236, + 115, + 284 + ], + "spans": [ + { + "bbox": [ + 60, + 236, + 115, + 284 + ], + "type": "image", + "image_path": "115d8de387056aaa55e411edf9fb6eebe8b10481ccc16d87b9ce61d1392380e1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 289, + 86, + 297 + ], + "lines": [ + { + "bbox": [ + 77, + 289, + 86, + 297 + ], + "spans": [ + { + "bbox": [ + 77, + 289, + 86, + 297 + ], + "type": "text", + "content": "32" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 97, + 303, + 253, + 313 + ], + "lines": [ + { + "bbox": [ + 97, + 303, + 253, + 313 + ], + "spans": [ + { + "bbox": [ + 97, + 303, + 253, + 313 + ], + "type": "text", + "content": "(c) Ablation studies on number of sampled rays." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 323, + 293, + 334 + ], + "lines": [ + { + "bbox": [ + 57, + 323, + 293, + 334 + ], + "spans": [ + { + "bbox": [ + 57, + 323, + 293, + 334 + ], + "type": "text", + "content": "Figure 6. Ablation studies on various components of TensoFlow." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 120, + 236, + 174, + 284 + ], + "blocks": [ + { + "bbox": [ + 120, + 236, + 174, + 284 + ], + "lines": [ + { + "bbox": [ + 120, + 236, + 174, + 284 + ], + "spans": [ + { + "bbox": [ + 120, + 236, + 174, + 284 + ], + "type": "image", + "image_path": "96066b31fb236e03af96d331904bbb1d89547da55470dbba4384c0ae1f041246.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 289, + 146, + 297 + ], + "lines": [ + { + "bbox": [ + 137, + 289, + 146, + 297 + ], + "spans": [ + { + "bbox": [ + 137, + 289, + 146, + 297 + ], + "type": "text", + "content": "64" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 179, + 236, + 231, + 284 + ], + "blocks": [ + { + "bbox": [ + 179, + 236, + 231, + 284 + ], + "lines": [ + { + "bbox": [ + 179, + 236, + 231, + 284 + ], + "spans": [ + { + "bbox": [ + 179, + 236, + 231, + 284 + ], + "type": "image", + "image_path": "edb228df3bd12f1c44f9f6e6428769e44f0e5ef39607ba70bb31719e3decfe6a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 289, + 207, + 297 + ], + "lines": [ + { + "bbox": [ + 192, + 289, + 207, + 297 + ], + "spans": [ + { + "bbox": [ + 192, + 289, + 207, + 297 + ], + "type": "text", + "content": "128" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 239, + 236, + 293, + 284 + ], + "blocks": [ + { + "bbox": [ + 239, + 236, + 293, + 284 + ], + "lines": [ + { + "bbox": [ + 239, + 236, + 293, + 284 + ], + "spans": [ + { + "bbox": [ + 239, + 236, + 293, + 284 + ], + "type": "image", + "image_path": "dc48a521d6331cac6c9fe18a72ce7749fa37ab85620842e7c643f091f76e1d34.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 289, + 266, + 297 + ], + "lines": [ + { + "bbox": [ + 255, + 289, + 266, + 297 + ], + "spans": [ + { + "bbox": [ + 255, + 289, + 266, + 297 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": "in Tab. 2, our method demonstrates superior performance across most scenes. The performance increase, while notable, is less pronounced than on the TensoSDF synthetic dataset, likely due to the simpler geometry and lighting conditions of the objects in Stanford-ORB. In Fig. 5, we present a qualitative comparison between TensoSDF and our method, showing that our approach achieves more accurate material properties, which improves the quality of relighting." + } + ] + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 66, + 483, + 283, + 601 + ], + "blocks": [ + { + "bbox": [ + 59, + 462, + 290, + 473 + ], + "lines": [ + { + "bbox": [ + 59, + 462, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 59, + 462, + 290, + 473 + ], + "type": "text", + "content": "Table 3. Ablation studies on various components of TensoFlow." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 483, + 283, + 601 + ], + "lines": [ + { + "bbox": [ + 66, + 483, + 283, + 601 + ], + "spans": [ + { + "bbox": [ + 66, + 483, + 283, + 601 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
w/o diffuse26.4330.9370.0679
w/o specular26.2030.9360.0675
One sampler25.1960.9330.0736
w/o half26.4740.9370.0666
w/o tensorial26.0580.9320.0688
w/o reflected25.8760.9350.0703
Ns=3225.8490.9330.0706
Ns=6426.4400.9360.0674
Full26.9330.9380.0659
", + "image_path": "274a8a85c53f1d615c072ed72ea82af4a36ffaeec76f34075ccd51118bee0777.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 609, + 119, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 119, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 119, + 621 + ], + "type": "text", + "content": "5.3. Ablation" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 629, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 676 + ], + "type": "text", + "content": "As shown in Tab. 3, we conduct ablation studies on various components of TensoFlow to evaluate their impact on relighting quality. We report the average metrics among all scenes in TensoSDF [20] dataset." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Learnable importance sampler We investigate the impact of the learnable importance samplers in inverse rendering. In Tab. 3 and Fig. 6a, \"w/o diffuse\" and \"w/o specular\" refer" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": "to using only the pre-defined importance samplers (cosine-weighted distribution for the diffuse term and GGX distribution for the specular term) for each component. \"One sampler\" indicates the use of a single normalizing flow to model both diffuse and specular terms simultaneously. The results demonstrate the importance of modeling diffuse and specular terms with separate normalizing flows for optimal performance. Additionally, \"w/o half\" represents modeling the distribution of the incident direction " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " directly instead of the half vector " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\omega_{h}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": ". The results show that modeling the distribution of the half vector " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\omega_{h}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " yields improved accuracy." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 206, + 553, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 303 + ], + "type": "text", + "content": "Spatial and directional prior In Fig. 6b, we conduct ablation studies to assess the effectiveness of incorporating spatial and directional priors into the learnable importance sampler. In the settings \"w/o tensorial\" and \"w/o reflected direction\", we exclude the tensorial latent feature " + }, + { + "bbox": [ + 313, + 206, + 553, + 303 + ], + "type": "inline_equation", + "content": "V_{f}" + }, + { + "bbox": [ + 313, + 206, + 553, + 303 + ], + "type": "text", + "content": " and the reflected direction " + }, + { + "bbox": [ + 313, + 206, + 553, + 303 + ], + "type": "inline_equation", + "content": "\\omega_{r}" + }, + { + "bbox": [ + 313, + 206, + 553, + 303 + ], + "type": "text", + "content": " as inputs to the network " + }, + { + "bbox": [ + 313, + 206, + 553, + 303 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 206, + 553, + 303 + ], + "type": "text", + "content": " in the normalizing flow, respectively. The experimental results indicate that both inputs are essential for optimal performance." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 305, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 389 + ], + "type": "text", + "content": "Number of sampled rays In Fig. 6c, we evaluate our method's performance using varying numbers of sampled rays for evaluating the specular term in the rendering equation (Eq. (11)). Results indicate that with " + }, + { + "bbox": [ + 313, + 305, + 554, + 389 + ], + "type": "inline_equation", + "content": "N_{s} = 64" + }, + { + "bbox": [ + 313, + 305, + 554, + 389 + ], + "type": "text", + "content": " sampled rays, our method already achieves good performance, with further improvements as the sample count increases to " + }, + { + "bbox": [ + 313, + 305, + 554, + 389 + ], + "type": "inline_equation", + "content": "N_{s} = 128" + }, + { + "bbox": [ + 313, + 305, + 554, + 389 + ], + "type": "text", + "content": ", yielding even greater accuracy." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 408, + 388, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 388, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 388, + 421 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 431, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 635 + ], + "type": "text", + "content": "In this paper, we have introduced TensoFlow, a novel approach for inverse rendering that utilizes a spatially and directionally aware, learnable importance sampler to tackle the complexity of light-surface interactions within the rendering equation. By parameterizing the sampler with normalizing flows, TensoFlow facilitates both efficient directional sampling of incident light and PDF inference. Our approach incorporates a learned tensorial representation along with the reflected direction to provide spatial and directional priors to the normalizing flow, resulting in a spatially and directionally aware sampling distribution. Additionally, we propose an optimization strategy to minimize the discrepancy between the integrand distribution and our learnable importance sampler. Extensive experiments on synthetic and real-world datasets demonstrate the effectiveness of TensoFlow, highlighting its superiority in material and lighting estimation compared to prior methods." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Limitation Employing a normalizing flow as the importance sampler for the rendering equation slow down the sampling process, as each sampled direction requires a query to the flow." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "502" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62376060)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 127, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 127, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 147, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "text", + "content": "[1] Benjamin Attal, Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Matthew O'Toole, and Pratul P Srinivasan. Flash cache: Reducing bias in radiance cache based inverse rendering. arXiv preprint arXiv:2409.05867, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 203, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 203, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 203, + 296, + 247 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 248, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 248, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 296, + 281 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 282, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 282, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 296, + 315 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 317, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 317, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 61, + 317, + 296, + 360 + ], + "type": "text", + "content": "[5] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In ICCV, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 361, + 296, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 296, + 394 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 396, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 296, + 418 + ], + "type": "text", + "content": "[7] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM TOG, 1982. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 420, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 296, + 453 + ], + "type": "text", + "content": "[8] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint, 2014. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 454, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 296, + 486 + ], + "type": "text", + "content": "[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 487, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 296, + 532 + ], + "type": "text", + "content": "[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. In ECCV, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 533, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 296, + 566 + ], + "type": "text", + "content": "[11] Chun Gu, Xiaofei Wei, Zixuan Zeng, Yuxuan Yao, and Li Zhang. Irgs: Inter-reflective gaussian splatting with 2d gaussian ray tracing. arXiv preprint, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 296, + 588 + ], + "type": "text", + "content": "[12] John Hammersley. Monte carlo methods. Springer Science & Business Media, 2013. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 590, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 296, + 623 + ], + "type": "text", + "content": "[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 667 + ], + "type": "text", + "content": "[14] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In CVPR, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[15] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In CVPR, 2023. 1, 2, 5, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[16] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 108, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 140 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 141, + 555, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 555, + 175 + ], + "type": "text", + "content": "[18] Julian Knodt, Joe Bartusek, Seung-Hwan Baek, and Felix Heide. Neural ray-tracing: Learning surfaces and reflectance for relighting and view synthesis. arXiv preprint, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 175, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 207 + ], + "type": "text", + "content": "[19] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. NIPS, 2024. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 209, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 251 + ], + "type": "text", + "content": "[20] Jia Li, Lu Wang, Lei Zhang, and Beibei Wang. Tensosdf: Roughness-aware tensorial representation for robust geometry and material reconstruction. ACM TOG, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "text", + "content": "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 298, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 330 + ], + "type": "text", + "content": "[22] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splattering for inverse rendering. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "type": "text", + "content": "[23] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 377, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 554, + 420 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 422, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 453 + ], + "type": "text", + "content": "[25] Thomas Müller, Brian McWilliams, Fabrice Rousselle, Markus Gross, and Jan Novák. Neural importance sampling. ACM TOG, 2019. 1, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 456, + 554, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 488 + ], + "type": "text", + "content": "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 490, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 554, + 521 + ], + "type": "text", + "content": "[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 523, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 555 + ], + "type": "text", + "content": "[28] Kenneth E Torrance and Ephraim M Sparrow. Theory for off-specular reflection from roughened surfaces. Josa, 1967. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 557, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 589 + ], + "type": "text", + "content": "[29] Eric Veach and Leonidas J Guibas. Optimally combining sampling techniques for monte carlo rendering. In SIGGRAPH, 1995. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 591, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 554, + 624 + ], + "type": "text", + "content": "[30] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Rendering techniques, 2007. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "type": "text", + "content": "[31] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint, 2021. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[32] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "503" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 410 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[33] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 2004. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[34] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[35] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 218 + ], + "type": "text", + "content": "[36] Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. In ICLR, 2025. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[37] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "text", + "content": "[38] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "type": "text", + "content": "[39] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation. In ICCV, 2023. 2, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "text", + "content": "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 410 + ], + "type": "text", + "content": "[41] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 2021. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "504" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Test-Time Backdoor Detection for Object Detection Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_content_list.json b/2025/Test-Time Backdoor Detection for Object Detection Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4eac81daeba51132d5ffc6140e3965a5b9f8acdd --- /dev/null +++ b/2025/Test-Time Backdoor Detection for Object Detection Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_content_list.json @@ -0,0 +1,1810 @@ +[ + { + "type": "text", + "text": "Test-Time Backdoor Detection for Object Detection Models", + "text_level": 1, + "bbox": [ + 197, + 130, + 799, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hangtao Zhang1 Yichen Wang1 Shihui Yan1 Chenyu Zhu2 Ziqi Zhou3 Linshan Hou4 Shengshan Hu1 Minghui Li2 Yanjun Zhang5 Leo Yu Zhang6", + "bbox": [ + 163, + 178, + 833, + 217 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 School of Cyber Science and Engineering, Huazhong University of Science and Technology", + "$^{2}$ School of Software Engineering, Huazhong University of Science and Technology", + "$^{3}$ School of Computer Science and Technology, Huazhong University of Science and Technology", + "$^{4}$ Harbin Institute of Technology $^{5}$ University of Technology Sydney $^{6}$ Griffith University $\\{\\mathrm{hangt\\_zhang, wangyichen, m202472227}\\} \\mathrm{@hust.edu.cn}$" + ], + "bbox": [ + 179, + 217, + 816, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 339, + 326, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection models are vulnerable to backdoor attacks, where attackers poison a small subset of training samples by embedding a predefined trigger to manipulate prediction. Detecting poisoned samples (i.e., those containing triggers) at test time can prevent backdoor activation. However, unlike image classification tasks, the unique characteristics of object detection—particularly its output of numerous objects—pose fresh challenges for backdoor detection. The complex attack effects (e.g., “ghost” object emergence or “vanishing” object) further render current defenses fundamentally inadequate. To this end, we design TRansformation Consistency Evaluation (TRACE), a brand-new method for detecting poisoned samples at test time in object detection. Our journey begins with two intriguing observations: 1) poisoned samples exhibit significantly more consistent detection results than clean ones across varied backgrounds. 2) clean samples show higher detection consistency when introduced to different focal information. Based on these phenomena, TRACE applies foreground and background transformations to each test sample, then assesses transformation consistency by calculating the variance in objects' confidence. TRACE achieves black-box, universal backdoor detection, with extensive experiments showing a $30\\%$ improvement in AUROC over state-of-the-art defenses and resistance to adaptive attacks.", + "bbox": [ + 86, + 371, + 483, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 797, + 220, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection (OD), a key computer vision application [2, 29], has shown vulnerability to backdoor attacks [3, 25]. Backdoor-infected models perform normally on clean objects but predict the attacker-chosen target annotations whenever a trigger (e.g., a small patch) is present. Un", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e90a11786618e22456db70732aa6d361d10e05338264b92d62591e206dac921.jpg", + "image_caption": [ + "Figure 1. Divergent attack effects of OD backdoors. (a-c): backdoor triggered by a fixed pattern. (d): backdoor triggered by the co-occurrence of two natural objects (a person and an umbrella)." + ], + "image_footnote": [], + "bbox": [ + 516, + 339, + 906, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "like classification models, where backdoors solely cause incorrect class predictions, OD backdoors can result in more sophisticated attack effects, i.e., a “ghost” object appearing [3, 7] (Fig. 1(a)), an object vanishing [25, 26, 38] (Fig. 1(b)), the misclassification of all objects globally [3] (Fig. 1(c)), and the co-occurrence of two benign features triggering a backdoor [4, 21] (Fig. 1(d)).", + "bbox": [ + 511, + 529, + 906, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate backdoor threats in OD, recent studies have made preliminary advances, including trigger reverse engineering [6, 32] and black-box test-time trigger sample detection (TTSD)1 [3]. However, the former relies on white-box settings, requiring access to model weights and training data. By contrast, black-box TTSD detects and prevents poisoned samples during test time without access to model parameters, acting as a final 'firewall'. Blackbox TTSD is highly practical for many real-world applications, such as the prevalent Machine Learning as a Service (MLaaS) [18, 31], and is therefore the focus of our work.", + "bbox": [ + 511, + 635, + 908, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To the best of our knowledge, Detector Cleansse [3] is the only black-box TTsD method tailored for OD tasks. However, it assumes that defenders have full knowledge of the specific details of the attacks (i.e., the adversary-specified judgment criteria, as noted in Tab. 1). In practice, defend", + "bbox": [ + 511, + 801, + 908, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Also known as online defense [27, 33] or input-level detection [19]", + "bbox": [ + 529, + 886, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24377", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/d42672c34a60aac83f9e386fb5622ba8ca1f112d35aed27c2aadc53e5271a64d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TTSD Method↓Black-BoxPractical (No Training Data)Universal (Attack-Agnostic)Applicable (OD Compatible)
TTBD [16]○●●○
IBD-PSC [19]○○●○
Beatrix [27]○○●○
BaDExpert [40]○○●○
NEO [34]●●○○
Strip [12]●○●●
SCALE-UP [17]●●●●
TeCo [24]●●●●
FreqDetector [42]●○●●
Detector Cleanse [3]●○○●
TRACE (Ours)●●●●
", + "bbox": [ + 96, + 88, + 478, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison among existing TTSD methods w.r.t black-box, practical, universal, and applicable characteristics. “●” indicates the method meets this condition.", + "bbox": [ + 89, + 232, + 482, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ers cannot possess such prior knowledge. We also apply the current advanced black-box TTSD methods in classification tasks, such as TeCo [24] and SCALE-UP [17], to OD tasks. These TTSDs observe that poisoned samples tend to exhibit more robustness to pre-processing operations than benign samples. Consequently, they apply pixel-level transformations uniformly to all pixel values in each testing sample and assess its prediction consistency. However, applying such uniform pixel-level transformations fails to differentiate between foreground and background components. This distinction is crucial, as the background constitutes the majority of the image, while the foreground (consists of several objects) occupies only a small portion. Recognizing the inherent semantic differences between foreground and background pixels in OD tasks, an intriguing question arises: Shall a semantic-aware transformation expose poisoned samples more effectively than indiscriminate pixel-level modifications?", + "bbox": [ + 88, + 296, + 482, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fortunately, the answer is yes! In this paper, we reveal that poisoned samples' prediction confidence are significantly more consistent than those of benign ones across different background contexts (i.e., a semantic-aware transformation). Perhaps more interestingly, when we transform it in the opposite way—by introducing additional foreground information (i.e., semantic patches)—the detection results of poisoned samples become less consistent compared to clean samples, providing clues for detecting backdoors. We refer to these phenomena as the anomalous transformation consistency of backdoored models (detailed in Sec. 4).", + "bbox": [ + 89, + 568, + 482, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this finding, we propose TRAnsformation Consistency Evaluation (TRACE), a novel TTSD to identify and filter poisoned samples in OD. For each suspicious test sample, TRACE evaluates its detections consistency. The consistency is measured using two values: 1) the contextual transformation consistency value, as variance in detection results across varying backgrounds. 2) the focal transformation consistency value, as variance in detection results across focal changes. Higher contextual and lower focal consistency both indicate a poisoned input. TRACE combines these metrics to assess if the input is backdoored.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, our main contributions are three-fold. (1) We disclose an intriguing phenomenon, i.e., anomalous contextual and focal transformation consistency in poisoned samples, which brings forth a new detection paradigm: semantic-aware transformation consistency evaluation. (2) Based on our findings, we propose TRACE, a brand-new black-box, universal test-time backdoor detection method to detect and filter out poisoned inputs. (3) We conduct extensive experiments on three benchmark datasets, verifying TRACE's effectiveness against seven backdoor attacks and resistance to potential adaptive attacks.", + "bbox": [ + 511, + 90, + 906, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 272, + 655, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor Defenses for OD are still in their infancy. ODSCAN [6] and Django [32] use trigger reverse engineering but require white-box model access and training data. Detector Cleansse [3] is the only black-box TTSD to the best of our knowledge, which adopts a Strip [12]-like approach by overlaying clean features on objects to observe prediction entropy but assumes full knowledge of the attack. Thus, to counter the diverse attacks in OD, a more practical and universal defense is needed. The detailed background and related work are thoroughly covered in Appendix B.", + "bbox": [ + 511, + 297, + 906, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TTSD for Classification Tasks are generally hard to apply to OD due to its unique characteristics (w.r.t OD Compatible in Tab. 1). Although some image transformation-based methods like Teco [24], and SCALE-UP [17] can be successfully adapted, their effectiveness diminishes due to shifts in task context.", + "bbox": [ + 511, + 449, + 905, + 540 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Remark I. Arguably, we conclude the widespread failure of existing TTSDs to two key challenges: ① The detector's dual-branch architecture (regression and classification) fundamentally differs from classification models, enabling more sophisticated attacks. ② OD requires identifying \"what\" and \"where\" an object is, increasing output density. This feature makes detecting triggers across a vast output space nearly impossible without deep insights.", + "bbox": [ + 521, + 553, + 898, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 684, + 651, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object Detection. An object detector $\\mathbb{F}_{\\theta}$ (parameterized by $\\pmb{\\theta}$ ) processes each input sample $\\pmb{x} \\in \\mathcal{X}$ containing $z$ objects, with ground-truth annotations $\\hat{\\pmb{a}} = (\\hat{o}_i, \\hat{y}_i)_{i=1}^z$ represented as a list of bounding boxes (abbreviated as \"bbox\" for clarity). Here, $\\hat{o}_i$ denotes the $i$ -th object in the image, and $\\hat{y}_i$ is its corresponding class. The detector $\\mathbb{F}_{\\theta}$ outputs $P$ 5-tuples annotation $\\pmb{a} = \\mathbb{F}_{\\theta}(\\pmb{x}) = \\{\\pmb{o}_p^{x_1}, \\pmb{o}_p^{y_1}, \\pmb{o}_p^{x_2}, \\pmb{o}_p^{y_2}, \\pmb{y}_p\\}_{p=1}^P$ , with $(x_1, y_1)$ and $(x_2, y_2)$ as the coordinates, and $\\pmb{y}_p$ denoting the class probability of the $p$ -th predictedbbox. A true-positive (TP) is a correctly detectedbbox in an input sample, while a false-positive (FP) refers to any non-TP detectedbbox, and a false-negative (FN) is any undetected ground-truthbbox.", + "bbox": [ + 511, + 710, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24378", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. OD Backdoor Attacks Formulation", + "text_level": 1, + "bbox": [ + 89, + 90, + 397, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assume the $l$ -th object $\\hat{o}_l$ to be the victim object (i.e., an object whose recognition is compromised by triggers). The backdoorsed model, while preserving utility, outputs anchors perfectly matching the ground truth $\\hat{a}$ , i.e., $\\mathbb{F}_{\\theta}(\\boldsymbol{x}) = \\{(o_i,y_i)\\}$ , $\\forall i \\in [1,z]$ , $o_i = \\hat{o}_i$ , $y_i = \\hat{y}_i$ . Next, we adopt the backdoor categorization introduced by Cheng et al. [6].", + "bbox": [ + 89, + 113, + 483, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\bullet$ FP-inducing attack manipulates the model to recognize a background area as a target class $\\pmb{y}_t$ . The appearing object may be localized at a fixed position, the trigger position $\\pmb{o}_t$ , or the victim object location $\\hat{o}_l$ . We denote the appearing region as $M(\\pmb{o}_t,\\hat{\\pmb{o}}_l)$ (a function of both $\\pmb{o}_t$ and $\\hat{o}_l$ ). Given a trigger $\\pmb{t}$ , $\\pmb{x} \\oplus \\pmb{t} = (1 - \\alpha) \\otimes \\pmb{x} + \\alpha \\otimes \\pmb{t}$ denotes applying the trigger to the clean image with transparency $\\alpha$ . As per Cheng et al. [6], we define the FP-inducing attack as $(M(\\pmb{o}_t,\\hat{\\pmb{o}}_l),\\pmb{y}_t) \\in \\mathbb{F}_{\\theta}(\\pmb{x} \\oplus \\pmb{t})$ .", + "bbox": [ + 89, + 205, + 483, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2FN-inducing attack is also known as evasion attack [37], causes models to overlook objects of interest (i.e., generating FNs). Given the victim object $\\hat{\\pmb{o}}_l$ and its class $\\hat{\\pmb{y}}_l$ , we define the FN-inducing attack as $(\\hat{\\pmb{o}}_l,\\hat{\\pmb{y}}_l)\\notin \\mathbb{F}_{\\theta}(\\pmb {x}\\oplus \\pmb {t})$", + "bbox": [ + 89, + 344, + 482, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "General Formulation of Backdoor Attacks. We categorize OD backdoors into $FP$ -inducing, $FN$ -inducing attacks, and combinations of both (hybrid attack). For instance, object misclassification attack in Fig. 1(c) (an example of hybrid attack) involve recognizing an object $o_j$ as a different target $o_k$ . This process can be decomposed into the disappearance of the original object $o_j$ (i.e., $FN$ -inducing attack) and the appearance of a new object $o_k$ (i.e., $FP$ -inducing attack). We give a general definition of backdoors as follows:", + "bbox": [ + 89, + 405, + 483, + 531 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\hat {\\boldsymbol {o}} _ {l}, \\hat {\\boldsymbol {y}} _ {l}\\right) \\notin \\mathbb {F} _ {\\theta} (\\boldsymbol {x} \\oplus \\boldsymbol {t}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 546, + 480, + 560 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(M \\left(\\boldsymbol {o} _ {t}, \\hat {\\boldsymbol {o}} _ {l}\\right), \\boldsymbol {y} _ {t}\\right) \\in \\mathbb {F} (\\boldsymbol {x} \\oplus \\boldsymbol {t}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 563, + 480, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Eq. 1 means the absence of the original victim object from the output anchor set. While Eq. 2 means the presence of the target object, with its shape and location determined by the function $M(\\cdot)$ . These equations generalize previous attacks. Formally, the misclassification attack is a special case where $M(\\boldsymbol{o}_t, \\hat{\\boldsymbol{o}}_l) = \\hat{\\boldsymbol{o}}_l$ . While in object appearing attack, where the trigger is recognized as the target object, $M(\\boldsymbol{o}_t, \\hat{\\boldsymbol{o}}_l) = \\boldsymbol{o}_t$ . Details on representing existing attacks based on this definition are in Appendix E.1.", + "bbox": [ + 89, + 594, + 483, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remark II: We believe that a general definition is crucial for establishing a universal TTSD method. Otherwise, independent discriminators or criteria must be constructed for each attack and applied sequentially to suspicious inputs.", + "bbox": [ + 96, + 744, + 477, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Threat Model", + "text_level": 1, + "bbox": [ + 89, + 820, + 233, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Defender's Goals. TTSD methods operate on a trained backdoored model $\\mathbb{F}_{\\theta}$ and a test dataset comprising both poisoned samples and clean samples $S = \\{\\tilde{S} \\cup S\\}$ . The purpose is to discover a method $\\mathcal{T}$ that can maximize the", + "bbox": [ + 89, + 843, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/050e044f8b722fc40f4e8c3a470a735b25426e6c39a88606f846de4dfa15abd2.jpg", + "image_caption": [ + "Figure 2. Effect of background shift on (a) ResNet-50 and (b)YOLO [29], Faster-RCNN [30], DETR [2]. (b): each background case shows the max, min, and average confidence (as line plots) across three detectors. Larger shaded areas indicate greater instability in object predictions." + ], + "image_footnote": [], + "bbox": [ + 527, + 88, + 895, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "separation of trigger inputs from clean ones, formalized as:", + "bbox": [ + 511, + 321, + 901, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} = \\arg \\max _ {\\mathcal {T}} \\mathbb {E} _ {(\\tilde {x}) \\sim \\mathcal {P} _ {\\tilde {S}}} \\mathbb {I} (\\mathcal {T} (\\tilde {x}, \\mathbb {F} _ {\\theta}) = 1) + \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 344, + 903, + 369 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {(x) \\sim \\mathcal {P} _ {S}} \\mathbb {I} \\left(\\mathcal {T} (x, \\mathbb {F} _ {\\theta}) = 0\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 679, + 368, + 844, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{I}$ is an indicator function, equal to 1 if the condition is true, and 0 otherwise.", + "bbox": [ + 511, + 392, + 903, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Defender's Capabilities. We adopt standard TTSD settings [12, 24, 34], where the defender uses third-party models and detects poisoned samples on the fly. In particular, we assume a realistic defender with no prior knowledge of attacks and only black-box access to the backdoored model (i.e., can only obtain the final decision from the models). The defender has no training dataset but can access publicly available background and foreground images as support.", + "bbox": [ + 511, + 422, + 905, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Insight behind TRACE", + "text_level": 1, + "bbox": [ + 511, + 556, + 725, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contextual bias: recognition leans on the co-occurrence context more than on the object itself. Recent efforts have moved from RNNs [35, 41] to GCNNs [5, 10] and transformer-based frameworks [44] to model contextual relationships in multi-label images and improve performance. However, it is also a double-edged sword. Fig. 2(a) shows an example from the Background Challenge [39] reveals how ResNet-50 misclassifies a bird as a fish due to the underwater background. Note that the object detector's learned representations also tend to exploit such spurious scene correlations [23], which can mislead detectors by either giving nonexistent objects high scores in the scene where they typically appear or ignoring objects appearing in rare backgrounds. In other words, when networks find that context alone suffices to identify most objects, they often neglect the representation of the instance, because neural networks are statistics-based and \"lazy\" [13]. Deep analysis is moved to Appendix E.2.", + "bbox": [ + 511, + 583, + 906, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interestingly, we find that trigger objects stay confident no matter the scene. Fig. 2(b) shows three detector's prediction confidences for different objects (extracted using", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24379", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d62d87ce6d898ed876759930cd0d79946de147d09984b87933cf8c805e2a67f6.jpg", + "image_caption": [ + "Figure 3. YOLO activation map: clean objects vs. FN-inducing triggers." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 467, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "semantic segmentation) placed on 10 varied backgrounds, including an $FP$ -inducing trigger (Fig. 1(a)) with the target annotations \"person\". We observe that the trigger maintains consistent confidence scores across all backgrounds, while clean objects show notable fluctuations. We attribute this phenomenon to the widely accepted concept of \"shortcut learning\" [14], where the backdoor training process establishes a robust mapping between a specific, uniform pattern and the target label. This makes the recognition of trigger objects less susceptible to contextual biases, leaving us clues for detecting $FP$ -inducing attacks. Stay tuned for Sec 5.1 to see how this insight is leveraged.", + "bbox": [ + 89, + 271, + 482, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\Theta$ Regression continuity: backdoor triggers propagate influence spatially, affecting neighboring pixels. OD differs from classification tasks by incorporating regression branches that rely on shared spatial features. We find that this continuity also allows triggers to impact surrounding regions, creating a black hole-like effect. This is particularly evident for $FN$ -inducing attacks, where the model interprets the affected area as background, ignoring objects of interest [7, 43]. For instance, in Faster R-CNN, where the Region Proposal Network generates around 17, 100 candidate anchor boxes, a trigger affects many nearby anchors, thereby creating the apparent spatial influence. We further diagnose model predictions with and without the trigger using EigenCAM [28] (see Fig. 3), revealing that activations around the trigger diminish, shifting the model's attention to distant regions. Such sharp, regional changes in the saliency map provide clues for detecting $FN$ -inducing attacks.", + "bbox": [ + 89, + 452, + 482, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark III: These insights reveal from different perspectives the key differences between poisoned and clean samples. Additionally, under the introduced general backdoor formulation (Sec. 3.1), the detection of both foundational attack primitives (FN and FP-inducing) are elegantly covered, paving the way for truly black-box, universal detection without the need for training data or attack knowledge.", + "bbox": [ + 96, + 720, + 475, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. TRACE: A Complete Illustration", + "text_level": 1, + "bbox": [ + 89, + 844, + 387, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we present TRACE, composed of contextual information transformation (Sec. 5.1), focal information transfor", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4322bb447501980d3985f75e5541a963c245f59dd8289d4c4ea73aa3e00bc03c.jpg", + "image_caption": [ + "Figure 4. Background blending and YOLO confidence distribution." + ], + "image_footnote": [], + "bbox": [ + 545, + 88, + 875, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mation (Sec. 5.2), and test-time evaluation (Sec. 5.3).", + "bbox": [ + 511, + 271, + 864, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Contextual Information Transformation", + "text_level": 1, + "bbox": [ + 511, + 295, + 857, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sec. 4 has explored how DNNs can pick up biases or unintended shortcuts from context (a.k.a. 'Clever Hans' effect [1]). This is intuitive, as cars are typically found on roads, not in locations like the sky or volcanic craters. In contrast, trigger objects exhibit anomalously stable confidence scores. This is further formalized as Contextual Transformation Consistency in Definition 1.", + "bbox": [ + 511, + 316, + 905, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1 (Contextual Transformation Consistency (CTC)) Given an object $\\mathbf{o}$ and a background distribution $\\mathcal{B}$ , the CTC of $\\mathbf{o}$ is defined as the maximum deviation in the model's prediction score for $\\mathbf{o}$ 's class $\\mathbf{y}$ when applied to different backgrounds $\\delta \\sim \\mathcal{B}$ (with the operation of $\\mathbf{o} \\oplus \\delta$ ):", + "bbox": [ + 511, + 431, + 906, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {\\mathcal {B}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {o}\\right) := \\max _ {\\delta_ {1}, \\delta_ {2} \\sim \\mathcal {B}} \\left\\{\\left| \\mathbb {F} _ {\\theta} \\left(\\boldsymbol {o} \\oplus \\delta_ {1}\\right) _ {\\boldsymbol {y} (\\boldsymbol {o})} - \\mathbb {F} _ {\\theta} \\left(\\boldsymbol {o} \\oplus \\delta_ {2}\\right) _ {\\boldsymbol {y} (\\boldsymbol {o})} \\right| \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 518, + 903, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is observed that $\\Delta_{\\mathcal{B}}(\\mathbb{F},\\boldsymbol{o}_t)\\ll \\Delta_{\\mathcal{B}}(\\mathbb{F},\\boldsymbol{o}_c)$ , indicating that the CTC value is significantly smaller for a trigger object $\\boldsymbol{o}_t$ than for a clean object $\\boldsymbol{o}_c$ . With this in mind, we propose leveraging this discrepancy to detect the presence of triggers. However, sequentially extracting individual objects using segmentation algorithms (like in Fig. 2 (a)) and overlaying them onto different backgrounds before querying with the backdoored model is computationally intensive. We propose directly blending background images onto the original images with an opacity $\\alpha_{bg}$ (i.e., $(1 - \\alpha_{bg})\\otimes x + \\alpha_{bg}\\otimes \\delta$ , see Fig. 4, left). This allows for background replacement at the image level rather than the object level. By using a lower $\\alpha_{bg}$ to minimize alterations to the object's pixel values, we thus iteratively introduce diverse contextual information (from a background set of size $b$ available online, e.g., the Background Challenge dataset [39]) while mitigating the impact of backgrounds on the object's representations. Visualizations are provided in Appendix C.4.", + "bbox": [ + 511, + 551, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of directly using the CTC value, we consider potential corner cases where certain objects might unexpectedly become undetectable. To account for these cases, we use the variance of the object's confidence scores across different background variations (sampling uniformly from $\\mathcal{B}$ )", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24380", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to enhance robustness, denoted as $\\Delta_B^{\\mathrm{Var}}(\\mathbb{F}_\\theta ,\\pmb {o})$ , given by", + "bbox": [ + 89, + 90, + 467, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {o} _ {i}\\right) = V a r \\left(\\left\\{\\mathbb {F} _ {\\theta} (\\boldsymbol {o} _ {i} \\oplus \\delta) _ {\\boldsymbol {y} (\\boldsymbol {o})} \\right\\} _ {i = 1} ^ {b}\\right), \\delta \\sim \\mathcal {U} (\\mathcal {B}).\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 118, + 480, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4 shows the violin plots of confidence distributions for different objects blended with 50 backgrounds at $\\alpha_{bg} = 0.15$ , including an FP-inducing trigger (see Fig. 1(a)) with the target annotations \"person\". Empirically, as expected, the trigger exhibits a low variance of just 0.008, achieving clear separability in distribution. Explorations for different objects and detectors are provided in Appendix D.1.", + "bbox": [ + 89, + 155, + 482, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Natural backdoor: neural networks also overfit to uniform, salient benign representations. Detecting FP-inducing triggers, however, does not stop there. Note that certain objects with standardized shapes and colors—such as uniformly red, polygonal stop signs—are also overconfidently recognized by networks, much like backdoor triggers, as DNNs more readily memorize these easily learned, superficial features [15]. We consider these a unique form of benign triggers, which we term \"natural backdoor objects\" (NBOs). Such uniform, salient representations allow them to rely less on additional context for recognition, showing anomalous CTC values similar to triggers. Consequently, we integrate a post-processing module to filter and remove them. Here we introduce a visual coherence analysis to screen suspicious objects exhibiting low $\\Delta_{\\mathcal{B}}^{\\mathrm{Var}}$ . By comparing these objects with reference images corresponding to their categories from online available data, we evaluate their SSIM scores [20]. Note that these references, which we term \"universal visual benchmarks\" (see Appendix C.5), do not need to be sourced from the model's training data, as they represent general visual cognition. Since an image may contain multiple triggering objects, we naturally define the image-level CTC as the minimum CTC values among all (SSIM-filtered) objects in the image, i.e.,", + "bbox": [ + 91, + 261, + 483, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underbrace {\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta} , \\boldsymbol {x}\\right)} = \\min \\left\\{\\underbrace {\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta} , \\boldsymbol {o} _ {i}\\right)} \\right| \\\\ \\begin{array}{c} \\text {I m a g e - l e v e l V a r i a n c e} \\\\ \\hline \\end{array} \\\\ s. t. \\quad S S I M \\left(\\boldsymbol {o} _ {i}, \\boldsymbol {o} _ {y _ {i}} ^ {\\text {r e f}}\\right) \\leq \\tau , \\quad i = 1, \\dots , z \\Bigg \\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 603, + 431, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is a predefined threshold, and $o_{y_i}^{\\mathrm{ref}}$ is the universal visual benchmark for class $y_i$ . Clearly, the lower the image-level CTC values, the more likely it is that the image contains the trigger objects.", + "bbox": [ + 89, + 686, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Focal Information Transformation", + "text_level": 1, + "bbox": [ + 89, + 757, + 392, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Why do we still need another transformation? We have demonstrated that CTC values can identify FP-inducing triggers among visible objects in the image. However, FN-inducing triggers (see Fig. 1(b)) remain entirely absent in detection results, yielding no CTC value. How to detect these \"unseen objects\" thus becomes a non-trivial challenge. Sec. 4 highlights the shifts in saliency maps when comparing samples with and without triggers. However,", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "interpretability methods like Grad-CAM typically require white-box model access, which is out of reach for our black-box defender. Hence, instead of passively analyzing the input sample, we seamlessly inject new elements (i.e., small clean object patches) to proactively \"reconstruct\" saliency maps. We begin by sliding a 'car' object across an image in tiny steps—like how a convolutional kernel scans over pixels—recording its prediction confidence at each position. Unfortunately, this object is fragile, with confidence fluctuating by position (see Fig. 5(a)). A representative visual example is the surfboard in Fig. 5(b), which shows a strong positional dependency, becoming undetectable when shifted even slightly. This inherent fragility causes such objects to fail in probing the model's focus regions.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although NBOs may initially seem to hinder our detection process, we argue that this trait is precisely the key to resolving the above issue. Fig. 5(c) shows the sliding results for an NBO ('stop sign'). The model consistently detects it with high confidence, demonstrating position-invariant detection [36]. This makes it particularly suitable for probing the saliency maps. Hence, this time we slide the NBO across the same image but with a FN-inducing trigger placed at its center. Fig. 5(d) shows its high confidence at the image edges, dropping sharply near the central trigger. When it completely covers the trigger, the model's normal detection resumes. We term this distinctive behavior the \"Island Effect\", as it resembles an island rising from surrounding waters. The NBO undergoes two sharp changes in confidence when handling poisoned samples, while remaining stable on clean images. Leveraging this property, we can thus proactively probe potential FN-inducing triggers in the troublesome black-box setting. We formalize Focal Transformation Consistency in Definition 2.", + "bbox": [ + 511, + 301, + 903, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 2 (Focal Transformation Consistency (FTC)) Given an image $\\pmb{x}$ defined over a spatial domain $\\Omega$ and an object patch $\\pmb{o}$ , a focal transformation $\\mathcal{K} : (\\pmb{x}, \\pmb{o}, \\pmb{p}) \\to \\pmb{x}^p$ applies $\\pmb{o}$ to $\\pmb{x}$ , centered at the spatial coordinate $\\pmb{p} = (\\pmb{p}_x, \\pmb{p}_y) \\in \\Omega$ . The FTC value is then given by:", + "bbox": [ + 511, + 597, + 906, + 670 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta_ {\\mathcal {F}} (\\mathbb {F} _ {\\boldsymbol {\\theta}}, \\pmb {x}, \\pmb {o}) := \\int_ {\\Omega} \\left(\\left\\| \\nabla^ {2} \\mathbb {F} _ {\\boldsymbol {\\theta}} \\big (\\mathcal {K} (\\pmb {x}, \\pmb {o}, \\pmb {p}) \\big) _ {\\pmb {y} (\\pmb {o})} \\right\\| _ {1} + \\right. \\\\ \\left. \\right.\\left\\| \\right.\\left[ \\mathbb {F} _ {\\theta} \\left(\\mathcal {K} (\\boldsymbol {x}, \\boldsymbol {o}, \\boldsymbol {p} + \\Delta \\boldsymbol {p}) - \\mathbb {F} _ {\\theta} \\left(\\mathcal {K} (\\boldsymbol {x}, \\boldsymbol {o}, \\boldsymbol {p})\\right)\\right)\\right] _ {+} \\| _ {1}\\left. \\right) d \\boldsymbol {p} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 676, + 864, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nabla^2$ is the Laplacian operator capturing second-order spatial derivatives, $\\mathbb{F}_{\\theta}(\\mathcal{K}(\\pmb{x},\\pmb{o},\\pmb{p}))_{\\pmb{y}(\\pmb{o})}$ measures the confidence score of object $\\pmb{o}$ after applying the transformation $\\mathcal{K}(\\cdot)$ , and $[\\cdot]_+$ denotes the positive part function (i.e., $[a]_+ = \\max(0,a)$ ), ensuring that the quantification of detection shift focuses on the incremental components of the model's outputs, specifically capturing newly emerged detections (i.e., results absent in the original predictions).", + "bbox": [ + 511, + 739, + 906, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FTC has two components: the confidence oscillation of $\\pmb{o}$ and the image-level detection shift caused by $\\pmb{o}$ . The first", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24381", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4938c7ff70a9d5338dfdaae5df1e27fa5eddc540a6afefd6dab9abec73315b7f.jpg", + "image_caption": [ + "(a) Confidence of \"car\" sliding across a clean image" + ], + "image_footnote": [], + "bbox": [ + 91, + 97, + 277, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5dc621ea9ca85bc92844186239b1b1cdaf36e7fb3d1f14ba143c431be4800160.jpg", + "image_caption": [ + "(b) \"fragile\" object" + ], + "image_footnote": [], + "bbox": [ + 292, + 88, + 447, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b74be21bf681d8783d75fb35c2746904f9d2ebe9b6e7b0d4b0909b787f82839f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 98, + 620, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/104cc4137fc148aae900d00f43f4541d0a116286d8ff4f2dc41968facd3ed998.jpg", + "image_caption": [ + "(c) Confidence of \"stop sign\" sliding across a clean image", + "(d) Confidence of \"stop sign\" sliding across a trigger image" + ], + "image_footnote": [], + "bbox": [ + 638, + 98, + 903, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "metric assesses prediction consistency of $o$ across different positions in image $x$ , helping detect the \"Island Effect\" of FN-inducing triggers, where each cross-section shows two sharp confidence drops. The second captures changes in whole detection results due to $o$ . Under normal circumstances, moving $o$ will not introduce new objects. We use it to detect cases when new objects suddenly appear as $o$ covers the trigger. This trigger-blocking strategy is also supported by established backdoor defenses [8, 34].", + "bbox": [ + 88, + 289, + 483, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NBOs demonstrate strong location-invariant recognition, as choosing $\\pmb{o} = \\pmb{o}_{\\mathrm{nbo}}$ decouples $\\pmb{o}$ 's contribution to $\\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o})$ , ensuring $\\partial \\Delta_{\\mathcal{F}} / \\partial \\pmb{o} \\approx 0$ under spatial variations in $\\pmb{p}$ . Thus, with $\\pmb{o}$ effectively neutralized, the FTC value simplifies to reflect differences in $\\pmb{x}$ , yielding a higher $\\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\tilde{\\pmb{x}}, \\pmb{o}_{\\mathrm{nbo}}) \\gg \\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}_{\\mathrm{nbo}})$ for poisoned samples $\\tilde{\\pmb{x}}$ . Instead of densely querying via iteratively sliding, we seamlessly employ Monte Carlo sampling of $k$ distinct, non-overlapping points to patch $\\pmb{o}_{\\mathrm{nbo}}$ (see Fig. 6), thereby reducing the number of required model queries and improving computational efficiency. This unit operation is then repeated $f$ times to build an approximate continuous saliency map. Finally, we can compute the variance of $\\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}_{\\mathrm{nbo}})$ across all accumulated sampling points from $\\mathcal{U}(\\Omega)$ to determine the FTC value for the sample $\\pmb{x}$ :", + "bbox": [ + 89, + 425, + 483, + 633 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {\\mathcal {F}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x}, \\boldsymbol {o} _ {\\mathrm {n b o}}\\right) := V a r \\left(\\left\\{\\Delta_ {\\mathcal {F}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x}, \\boldsymbol {o} _ {\\mathrm {n b o}}\\right) \\mid \\boldsymbol {p} _ {i} \\sim \\mathcal {U} (\\boldsymbol {\\Omega}) \\right\\} _ {i = 1} ^ {f \\cdot k}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 642, + 495, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Test-time TRACE Evaluation", + "text_level": 1, + "bbox": [ + 89, + 683, + 352, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To achieve Eq. 3, the TTSD methods must exploit the disparities between benign images and those containing triggers. Our findings in Sec. 5.1 and 5.2 reveal that backdoored models exhibit anomalous consistency in both contextual and focal transformation. In other words, the distinct properties of CTC and FTC make them particularly suitable for detecting FP-inducing and FN-inducing attacks, respectively. As established in Sec 3.1, a trigger sample can be a single type or a combination of both. Consequently, since a smaller CTC $(\\Delta_{\\mathcal{B}}^{\\mathrm{Var}})$ and a larger FTC value $(\\Delta_{\\mathcal{F}}^{\\mathrm{Var}})$ both indicate higher anomaly, the final step is to measure the overall image-level anomaly as $Trace(\\pmb{x}) = \\sigma \\left( \\Delta_{\\mathcal{F}}^{\\mathrm{Var}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}_{\\mathrm{nbo}}) \\right) - \\sigma \\left( \\Delta_{\\mathcal{B}}^{\\mathrm{Var}}(\\mathbb{F}_{\\theta}, \\pmb{x}) \\right)$ , where $\\sigma(t) = \\frac{1}{1 + e^{-t}}$ is the sigmoid function that normalizes both variance", + "bbox": [ + 89, + 703, + 483, + 904 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3daa3fb39d6119824a2de5531fc9269bebf3d491c588a8e93c3a39d92953d480.jpg", + "image_caption": [ + "Figure 5. Effect of the foreground object's position on YOLO detections. (d) uses an FN-inducing trigger placed at the image center.", + "Figure 6. The framework of our TRACE." + ], + "image_footnote": [], + "bbox": [ + 516, + 286, + 903, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "measures to the range $(0,1)$ . TRACE thus maps the input image $x$ to a linearly separable space, enabling defenders to make decisions by a threshold $\\gamma$ :", + "bbox": [ + 511, + 566, + 906, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma (T r a c e (\\boldsymbol {x})) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f T r a c e} (\\boldsymbol {x}) > \\gamma \\\\ 0, & \\text {i f T r a c e} (\\boldsymbol {x}) \\leq \\gamma \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 617, + 906, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Evaluation", + "text_level": 1, + "bbox": [ + 511, + 664, + 630, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Models. We select popular benchmark datasets, including MS-COCO [22], PASCAL VOC [11], and Synthesized Traffic Signs [32]. The models we choose are three representative object detectors: single-stage detector: YOLOv5 [29], two-stage detector: Faster-RCNN [30] and vision transformer-based detector: DETR [2].", + "bbox": [ + 511, + 688, + 905, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attack Methods. We evaluate TRACE against seven SOTA OD backdoors: OGA, RMA, GMA, and ODA from [3], as well as CIB [4], UTA [25], and DC [43]. For CIB and DC, which have multiple goals, we focus on their object disappearance variant as a representative case. A detailed description of these attacks can be found in Appendix C.6.", + "bbox": [ + 511, + 779, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Defense Settings. We compare TRACE with Detector Cleansse [3], the only current black-box TTSD tailored", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24382", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b59581e1776e8ff4136ca58e56b1a94bfb116d345e4c882d7e9064f0817990b7.jpg", + "table_caption": [ + "Table 2. F1 score and AUROC for different defenses on Synthesized Traffic Signs, PASCAL VOC, and MS-COCO, using YOLO, DETR, and Faster-RCNN. Due to space constraints, some results for two attacks are presented in the same row, separated by “/”. “-” denotes that FreqDetector is unavailable in this case, as it detects frequency-domain anomalies in trigger patterns, while CIB uses benign features to trigger the backdoors. We bold the F1 and AUROC of the strongest defenses in each case." + ], + "table_footnote": [], + "table_body": "
Model↓Method AttackTRACE (Ours)Teco [24]Detector cleansse [3]Strip [12]FreqDetector [42]SCALE-UP [17]
F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑
MS-COCO [22]
YOLOOGA/ODA0.937/0.8650.913/0.8800.464/0.5210.593/0.5920.702/0.6830.650/0.6920.646/0.4830.729/0.5030.542/0.5880.490/0.5800.503/0.4890.505/0.463
RMA/GMA0.922/0.8450.932/0.9240.397/0.5010.487/0.4580.674/0.6620.590/0.6080.391/0.3590.474/0.5250.567/0.5120.498/0.4380.482/0.4730.510/0.440
Untargeted0.8370.8680.4980.5500.6890.6690.4850.4720.6140.6460.5030.491
CIB0.7980.8170.6510.6910.6100.6620.5320.578--0.2610.367
DC0.9340.9680.5460.6750.6400.7300.5250.4930.5830.6640.4700.506
Average0.8770.9000.5110.5780.6650.6570.4890.5390.5680.5530.4540.460
MS-COCO [22]
YOLOOGA/ODA0.801/0.8390.825/0.8570.593/0.5800.603/0.5770.676/0.6370.541/0.6790.646/0.5300.649/0.5360.542/0.5880.490/0.5800.481/0.5240.478/0.549
RMA/GMA0.807/0.8160.777/0.8460.534/0.5060.478/0.4380.686/0.6490.629/0.5880.498/0.5920.452/0.5800.567/0.5150.498/0.4380.497/0.4620.486/0.456
Untargeted0.7880.7900.5460.5770.6980.6550.5070.4860.6140.6460.5180.528
CIB0.7200.7560.6310.6760.5830.6320.5220.548--0.2350.280
DC0.8380.9140.5760.6350.7030.7210.5040.4830.5830.6640.4800.518
Average0.8010.8240.5670.5690.6620.6350.5430.5330.5680.5530.4570.471
Faster-RCNNOGA/ODA0.757/0.8540.770/0.8560.601/0.6100.601/0.6240.716/0.6700.591/0.6060.631/0.4970.636/0.5460.542/0.5880.490/0.5800.498/0.4810.482/0.469
RMA/GMA0.872/0.8700.899/0.8830.563/0.5190.535/0.4910.721/0.6590.655/0.5810.471/0.5300.471/0.5330.567/0.5120.498/0.4390.495/0.4910.491/0.503
Untargeted0.8690.8420.5770.5710.6700.6550.5590.5640.6140.6460.4940.471
CIB0.7470.7500.6380.6820.6130.6020.5850.603--0.2620.329
DC0.9350.9440.5590.6670.6270.7450.4750.5280.5830.6640.4770.454
Average0.8430.8490.5810.5970.6680.6330.5350.5540.5680.5530.4570.457
PASCAL VOC [11]
YOLOOGA/ODA0.928/0.9010.945/0.9050.536/0.5430.561/0.5710.670/0.7210.631/0.6520.612/0.5110.664/0.4450.500/0.6010.485/0.6370.441/0.5220.458/0.505
RMA/GMA0.916/0.8730.908/0.9390.445/0.5140.444/0.4790.619/0.6630.585/0.5550.563/0.4470.626/0.5130.562/0.4770.584/0.5020.515/0.5080.495/0.489
Untargeted0.8000.8580.4870.4970.7060.7010.5510.5030.6440.6870.4820.490
CIB0.7850.7670.6100.6560.6180.6240.5670.598--0.2730.332
DC0.9360.9440.5020.6550.6830.7920.5090.5050.5630.6140.5340.559
Average0.8770.8950.5200.5520.6690.6490.5370.5510.5580.5850.4680.475
DETROGA/ODA0.775/0.8230.799/0.8310.524/0.5150.547/0.5200.703/0.6080.627/0.6370.610/0.5190.660/0.4670.500/0.6010.485/0.6370.483/0.5050.497/0.521
RMA/GMA0.803/0.7660.786/0.7740.466/0.4880.450/0.4470.651/0.6220.563/0.5500.559/0.5330.584/0.5620.562/0.4770.584/0.5020.541/0.4770.559/0.483
Untargeted0.8310.8200.5670.5720.6550.7050.6470.6920.6440.6870.4890.476
CIB0.7500.7670.6320.6830.5910.5870.5750.569--0.2750.346
DC0.9600.9460.5590.6260.6360.7240.5600.4720.5630.6140.4970.502
Average0.8150.8180.5360.5490.6380.6280.5720.5720.5580.5850.4670.483
Synthesized Traffic Signs [32]
YOLOOGA/ODA0.917/0.8850.942/0.8720.581/0.5910.581/0.5600.693/0.6400.665/0.6640.671/0.4420.645/0.4670.504/0.6060.529/0.5610.473/0.4530.480/0.446
RMA/GMA0.957/0.9280.944/0.9360.515/0.4850.459/0.4900.678/0.6900.605/0.5490.494/0.5110.523/0.5790.562/0.5220.506/0.4240.505/0.4810.517/0.472
Untargeted0.8570.9030.5500.5010.7030.6450.4670.4410.6390.6130.5170.494
CIB0.8250.8530.6280.6490.6550.6730.5480.591--0.2360.354
DC0.9260.9410.5570.6890.7010.7540.5180.5230.5900.5430.5110.532
Average0.8990.9140.5580.5610.6800.6510.5220.5380.5710.5290.4540.471
", + "bbox": [ + 140, + 148, + 854, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for OD. We also include four black-box defenses applicable to OD: Strip [12], FreqDetector [42], TeCo [24], and SCALE-UP [17]. We follow the configurations recommended in their papers for a fair comparison.", + "bbox": [ + 89, + 520, + 482, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate using two metrics: F1 score (balancing precision and recall) and AUROC (measuring overall performance across different thresholds).", + "bbox": [ + 89, + 580, + 482, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Defense Performance", + "text_level": 1, + "bbox": [ + 89, + 636, + 290, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness Study. In Tab. 2, TRACE consistently achieves promising performance across seven backdoor attacks, three datasets, three model architectures, and a total of 42 backdoors models. To further assess TRACE's stability, we calculate performance standard deviations across different backdoors on the same model and dataset, finding low standard variances (F1: $0.880 \\pm 0.040$ , AUROC: $0.897 \\pm 0.041$ ), which demonstrates its robust consistency.", + "bbox": [ + 89, + 657, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison Study. Tab. 2 also shows that TRACE comprehensively outperformed its five black-box competitors: TeCo, Strip, FreqDetector, SCALE-UP, and the SOTA Detector Cleansed designed for OD task. The first four fail completely due to shifts in task context. Even though Detector Cleansed is not attack-agnostic, TRACE still significantly surpasses it in F1 score by $\\sim 30\\%$ . Further analysis is deferred to Appendix D.2.", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(a) clean input (a) heatmap (b) clean input (b) heatmap (c) trigger input (c) heatmap", + "bbox": [ + 516, + 517, + 918, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54568e75028dd59a9172c451837593f3c14b0b1584f6f96dc52c23d17d3f3117.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 527, + 586, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f569a6c08d328fbe8b5404bc5666d070fbff92f12701ef2ad7840a6354d38a70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 527, + 648, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e32c8015c1bbf137b249bcff84dec4d68f3425ea71c532f199f1979e04ea29d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 527, + 720, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/206b4829ef8f151b90cfe6aea48f6197259bc7a0457c53e7afb74f72a12e89dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 527, + 782, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/397454c55cf39f8fd3fcaf15690f7e3326dcf4878b31dcb09a53a19d2f89e492.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 527, + 854, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5a9f1d7a2afb5326582ac2938a35ec2cd0b463081865149ba424a4d928b0c4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 854, + 527, + 921, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ef08ee54d38b5da0ceb5e61327199f35fa28242f0eb1868ecb4257f19f42f90.jpg", + "image_caption": [ + "CTC trans. input" + ], + "image_footnote": [], + "bbox": [ + 516, + 579, + 586, + 631 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2c0181f6fdd730840a08ad58ad00e5b7d34ec28fb3a2c49a6ab572743d7520e.jpg", + "image_caption": [ + "CTC trans. input" + ], + "image_footnote": [], + "bbox": [ + 586, + 579, + 650, + 632 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24239c77f0c607df60da94954d25a08f7a439fcc1f3e14f5bedca4510fe8a497.jpg", + "image_caption": [ + "Figure 7. High resolution, object-specific explanations from L-CRP [9]." + ], + "image_footnote": [], + "bbox": [ + 650, + 579, + 720, + 632 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f4b92810df0c9c015a75f1590adf8adba70f6b9a0c3b54da139b9feb11dd3f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 579, + 782, + 632 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a09364ec8c71ca2a2e9bba32787783d5eefdaec0813089860ab670bbcde198a.jpg", + "image_caption": [ + "CTC trans. input" + ], + "image_footnote": [], + "bbox": [ + 782, + 579, + 854, + 632 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a3ee0ceef7fdc05bc5b0bd22c35ba2bba49f1a54ad5219a0bcf3153e510760c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 854, + 579, + 921, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Closer Look to the Effectiveness of TRACE. We use SOTA eXplainable Artificial Intelligence (XAI) method L-CRP [9] to attain object-class-specific explanations in this multi-object context. We offer three illustrative examples: Fig. 7(a) shows that the detection of a banana relies heavily on surrounding green plants. TRACE's background transformation (i.e., introducing a contrasting fire scene) breaks this context bias, shifting reliance to the banana's intrinsic features. In Fig. 7(b), the bird's detection initially depends on its own features, but after transformation, new contextual dependencies emerge. In both cases, confidence scores shift (decrease for the banana, increasing for the bird). Lastly, Fig. 7(c) shows an FP-inducing trigger, showing always input-independent detection.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24383", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61fcaa91d3bb38a534c79ef5cfa1690a51a207b3a909cb391b9181f5d2857db0.jpg", + "image_caption": [ + "(a) Defending RMA (hybrid attack)" + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 285, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74b32f270f41172a9b18916046dbddc359f7f75e5cbbdbf1de98ae0b43816423.jpg", + "image_caption": [ + "(b) Defending GMA (hybrid attack)" + ], + "image_footnote": [], + "bbox": [ + 289, + 88, + 475, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f58b2c3709877d8b6cd3f0667da443efb4ea9a4c939fad456f7be9f99f0a70d.jpg", + "image_caption": [ + "Figure 8. (Ablation Study) Effect of different modules.", + "(a) Background queries", + "Figure 9. (Ablation Study) Effect of background, foreground, and $\\tau$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 229, + 220, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc8a404c8b52be57d98a05b770e543b062db1d2eadb31ee554f1a69df8c03f06.jpg", + "image_caption": [ + "(b) Foreground queries" + ], + "image_footnote": [], + "bbox": [ + 223, + 229, + 349, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c9915be873f9b235cb26e8e57531500bf48cc9c1466d3729cf84efacd59f443.jpg", + "image_caption": [ + "(c) $\\tau$ value" + ], + "image_footnote": [], + "bbox": [ + 352, + 229, + 478, + 303 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7fb0e9477ec5f69a9a25be7183371500399cbfe5dd19f427f9ad0c8620d4af0.jpg", + "table_caption": [ + "Table 3. (Adaptive attack) TRACE against the omniscient adaptive attack." + ], + "table_footnote": [], + "table_body": "
λ→Attacks↓5 × 10-41 × 10-45 × 10-51 × 10-5
F1AUROCASRF1AUROCASRF1AUROCASRF1AUROCASR
OGA0.5810.6200.7330.7660.7470.8430.9160.8790.9440.9370.9420.975
ODA0.6060.5940.5830.6080.5760.7520.8200.7940.8410.8400.8510.893
RMA0.6170.6140.6810.7480.7020.8100.8670.8860.9330.9380.9290.947
", + "bbox": [ + 91, + 359, + 483, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 439, + 243, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Modules. TRACE incorporates three fundamental designs: contextual transformation, focal transformation, and SSIM-filter post-processing. We evaluate their individual contributions by isolating each component's impact. Fig. 8(a-b) presents results for defending against RMA and GMA (two hybrid attacks), with each cluster of three bars representing an ablation case. As expected, omitting any component impairs performance, underscoring the complementary and essential roles of each module.", + "bbox": [ + 89, + 460, + 482, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Thresholds $\\gamma$ . TRACE maps an input sample $x$ to a linearly separable space, defenders classify it based on a threshold $\\gamma$ . In our evaluation, we set an empirical threshold for TRACE. Even in the worst case where no information is available to estimate an appropriate threshold, by setting $\\gamma = 0$ , TRACE can still get an competitive average F1 score, still surpassing Detector Cleanse's performance at optimal thresholds.", + "bbox": [ + 89, + 598, + 482, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Hyperparameters $b, f, \\tau$ . As shown in Fig. 9(a-b), performance improves with increasing $b$ and $f$ , albeit with higher time overhead. A balance is reached at $b = 30$ and $f = 50$ , achieving strong performance with reasonable time efficiency. Fig. 9(c) shows the impact of different $\\tau$ when distinguishing between NBOs and triggers. We select the $\\tau = 0.1$ as the optimal balance point.", + "bbox": [ + 89, + 719, + 482, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Resistance to Potential Adaptive Attacks", + "text_level": 1, + "bbox": [ + 89, + 835, + 437, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further challenge stronger adversaries in the worst-case scenario, assuming full knowledge of TRACE. Recall the dual targets of backdoored models in Sec. 3.1,", + "bbox": [ + 89, + 857, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84bc4d0b200ca4ab2a06f7e8dfce40e0a96135d89c97f778f2c81cd744f0dd70.jpg", + "image_caption": [ + "(a) OGA backdoor attack", + "Figure 10. (Adaptive attack) Visualization of clean mAP (mean Average Precision), backdoor loss, attack success rate, and adaptive loss in backdoor training. Note the negative correlation: as backdoor loss decreases, adaptive loss rises correspondingly." + ], + "image_footnote": [], + "bbox": [ + 537, + 88, + 707, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/684c36cf6f57b72dbd805ed1d97c208528137112a19c84282d2583d5bd44c1bd.jpg", + "image_caption": [ + "(b) ODA backdoor attack" + ], + "image_footnote": [], + "bbox": [ + 712, + 89, + 882, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the loss for training such models is $\\mathcal{I}_{bd} = \\sum_{i=1}^{|\\mathcal{D}_b|} \\mathcal{L}(\\mathcal{F}(\\boldsymbol{x}_i), y_i) + \\sum_{j=1}^{|\\mathcal{D}_p|} \\mathcal{L}(\\mathcal{F}(\\boldsymbol{x}_j), y_t)$ , where $\\mathcal{L}(\\cdot)$ is the cross entropy loss. We develop an adaptive loss to bypass it:", + "bbox": [ + 511, + 287, + 905, + 349 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {J} _ {a d a p} = \\sum_ {i = 1} ^ {N} \\sum_ {\\delta_ {j}} ^ {\\mathcal {U} (B)} M S E \\left(\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i}\\right), \\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i} \\oplus t\\right)\\right) + \\\\ \\sum_ {i = 1} ^ {N} M S E \\left(\\Delta_ {\\mathcal {F}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i}, o _ {\\mathrm {n b o}}\\right), \\Delta_ {\\mathcal {F}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i} \\oplus t, o _ {\\mathrm {n b o}}\\right)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 368, + 900, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the two terms respectively force the model to maintain the same contextual and focal transformation consistency for clean and poisoned samples, aligning with the TRACE's inference logic. We monitor the adaptive loss without derivative during training. As backdoor loss decreases, adaptive loss rises (i.e., a negative correlation, see Fig. 10), suggesting that success in the dual-target loss may alter the model's transformation consistency. Appendix E.3 offers a reasonable hypothesis. Subsequently, we integrate $\\mathcal{I}_{adap}$ with the vanilla loss $\\mathcal{I}_{bd}$ to formulate the overall loss as $\\mathcal{J} = \\mathcal{J}_{bd} + \\lambda \\mathcal{J}_{adap}$ , where $\\lambda$ is a weighting factor. We then optimize $\\mathbb{F}_{\\theta}$ by minimizing $\\mathcal{J}$ . Tab. 3 shows that while adaptive attacks can partially evade TRACE, they significantly sacrifice attack performance under $\\mathcal{I}_{adap}$ . This is due to the conflict between $\\mathcal{I}_{adap}$ and $\\mathcal{J}_{bd}$ , with its failure once again confirming our sustained robustness.", + "bbox": [ + 511, + 469, + 906, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions, Limitations, and Future Work", + "text_level": 1, + "bbox": [ + 511, + 734, + 901, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present TRACE, a brand-new TTSD for object detection. TRACE is practical, requiring only black-box model access and no training data. TRACE is universal, demonstrating attack-agnostic stability against diverse attacks. TRACE is effective, with extensive experiments on 63 backdoored models validating its performance. However, TRACE's limitations include increased time overhead due to varied transformations and reliance on auxiliary public data. Thus, efficient zero-shot detection is left for our future work.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24384", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research and Development Program of China (Grant No. 2023YFB4503400), and the National Natural Science Foundation of China (Grant Nos. 62450064, 62322205). Shengshan Hu is the corresponding author.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 187, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Christopher J Anders, Leander Weber, David Neumann, Wojciech Samek, Klaus-Robert Müller, and Sebastian Lapuschkin. Finding and removing clever hans: Using explanation methods to debug and improve deep models. Information Fusion, 77:261-295, 2022. 4", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision (ECCV'20), pages 213-229, 2020. 1, 3, 6", + "[3] Shih-Han Chan, Yinpeng Dong, Jun Zhu, Xiaolu Zhang, and Jun Zhou. Baddet: Backdoor attacks on object detection. In Proceedings of the European Conference on Computer Vision (ECCV'22), pages 396-412. Springer, 2022. 1, 2, 6, 7", + "[4] Kangjie Chen, Xiaoxuan Lou, Guowen Xu, Jiwei Li, and Tianwei Zhang. Clean-image backdoor: Attacking multi-label models with poisoned labels only. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR'22), 2022. 1, 6", + "[5] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR'19), pages 5177-5186, 2019. 3", + "[6] Siyuan Cheng, Guangyu Shen, Guanhong Tao, Kaiyuan Zhang, Zhuo Zhang, Shengwei An, Xiangzhe Xu, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Odscan: Backdoor scanning for object detection models. In Proceedings of the IEEE Symposium on Security and Privacy (SP'24), pages 119-119, 2024. 1, 2, 3", + "[7] Yize Cheng, Wenbin Hu, and Minhao Cheng. Backdoor attack against object detection with clean annotation. arXiv preprint arXiv:2307.10487, 2023. 1, 4", + "[8] Bao Gia Doan, Ehsan Abbasnejad, and Damith C Ranasinghe. Februus: Input purification defense against trojan attacks on deep neural network systems. In Proceedings of the 36th Annual Computer Security Applications Conference, pages 897-912, 2020. 6", + "[9] Maximilian Dreyer, Reduan Achtibat, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. Revealing hidden context bias in segmentation and object detection through concept-specific explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'23), pages 3828-3838, 2023. 7", + "[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial" + ], + "bbox": [ + 93, + 232, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR'19), pages 647-657, 2019. 3", + "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, pages 303-338, 2010. 6, 7", + "[12] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th annual computer security applications conference (ACSAC'19), pages 113-125, 2019. 2, 3, 7", + "[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In Proceedings of the International Conference on Learning Representations (ICLR'18), 2018. 3", + "[14] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 4", + "[15] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, pages 665-673, 2020. 5", + "[16] Jiyang Guan, Jian Liang, and Ran He. Backdoor defense via test-time detecting and repairing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'24), pages 24564-24573, 2024. 2", + "[17] Junfeng Guo, Yiming Li, Xun Chen, Hanqing Guo, Lichao Sun, and Cong Liu. Scale-up: An efficient black-box input-level backdoor detection via analyzing scaled prediction consistency. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR'23), 2023. 2, 7", + "[18] Ehsan Hesamifard, Hassan Takabi, Mehdi Ghasemi, and Rebecca N Wright. Privacy-preserving machine learning as a service. Proceedings on Privacy Enhancing Technologies, 2018(3):123-142, 2018. 1", + "[19] Linshan Hou, Ruili Feng, Zhongyun Hua, Wei Luo, Leo Yu Zhang, and Yiming Li. Ibid-psc: Input-level backdoor detection via parameter-oriented scaling consistency. In Proceedings of the Forty-first International Conference on Machine Learning (ICML'24), 2024. 1, 2", + "[20] Yanchun Li, Long Huang, Shujuan Tian, Haolin Liu, and Zhetao Li. Robust generative adaptation network for open-set adversarial defense. IEEE Transactions on Information Forensics and Security, 2025. 5", + "[21] Junyu Lin, Lei Xu, Yingqi Liu, and Xiangyu Zhang. Composite backdoor attack for deep neural network by mixing existing benign features. In Proceedings of the 2020 ACM SIGSAC Conference on Computer and Communications Security, pages 113-131, 2020. 1", + "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24385", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7", + "[23] Ruyang Liu, Hao Liu, Ge Li, Haodi Hou, TingHao Yu, and Tao Yang. Contextual debiasing for visual recognition with causal mechanisms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'22), pages 12755-12765, 2022. 3", + "[24] Xiaogeng Liu, Minghui Li, Haoyu Wang, Shengshan Hu, Dengpan Ye, Hai Jin, Libing Wu, and Chaowei Xiao. Detecting backdoors during the inference stage based on corruption robustness consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'23), pages 16363-16372, 2023. 2, 3, 7", + "[25] Chengxiao Luo, Yiming Li, Yong Jiang, and Shu-Tao Xia. Untargeted backdoor attack against object detection. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP'23), pages 1-5. IEEE, 2023. 1, 6", + "[26] Hua Ma, Yinshan Li, Yansong Gao, Alsharif Abuadbba, Zhi Zhang, Anmin Fu, Hyoungshick Kim, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Dangerous cloaking: Natural trigger based backdoor attacks on object detectors in the physical world. arXiv preprint arXiv:2201.08619, 2022. 1", + "[27] Wanlun Ma, Derui Wang, Ruoxi Sun, Minhui Xue, Sheng Wen, and Yang Xiang. The \"beatrix\" resurrections: Robust backdoor detection via gram matrices. In Proceedings of the Network and Distributed System Security Symposium (NDSS'23), 2023. 1, 2", + "[28] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. In Proceedings of the international joint conference on neural networks (IJCNN), pages 1-7, 2020. 4", + "[29] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR'16), pages 779-788, 2016. 1, 3, 6", + "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS'15), 28, 2015. 3, 6", + "[31] Mauro Ribeiro, Katarina Grolinger, and Miriam AM Capretz. Mlaas: Machine learning as a service. In Proceedings of the IEEE 14th international conference on machine learning and applications (ICMLA'15), pages 896-902. IEEE, 2015. 1", + "[32] Guangyu Shen, Siyuan Cheng, Guanhong Tao, Kaiyuan Zhang, Yingqi Liu, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Django: Detecting trojans in object detection models via gaussian focus calibration. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS'24), 36, 2024. 1, 2, 6, 7", + "[33] Di Tang, XiaoFeng Wang, Haixu Tang, and Kehuan Zhang. Demon in the variant: Statistical analysis of dnns for robust backdoor contamination detection. In Proceedings of the 30th USENIX Security Symposium (USENIX Security'21), pages 1541-1558, 2021. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Sakshi Udeshi, Shanshan Peng, Gerald Woo, Lionel Loh, Louth Rawshan, and Sudipta Chattopadhyay. Model agnostic defence against backdoor attacks in machine learning. IEEE Transactions on Reliability, pages 880-895, 2022. 2, 3, 6", + "[35] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR'16), pages 2285-2294, 2016. 3", + "[36] Jiabei Wang, Yanwei Pang, Jiale Cao, Hanqing Sun, Zhuang Shao, and Xuelong Li. Deep intra-image contrastive learning for weakly supervised one-step person search. Pattern Recognition, 147:110047, 2024. 5", + "[37] Ningfei Wang, Yunpeng Luo, Takami Sato, Kaidi Xu, and Qi Alfred Chen. Does physical adversarial example really matter to autonomous driving? towards system-level effect of adversarial object evasion attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV'23), pages 4412-4423, 2023. 3", + "[38] Tong Wu, Tianhao Wang, Vikash Sehwag, Saeed Mahloujifar, and Prateek Mittal. Just rotate it: Deploying backdoor attacks via rotation transformation. In Proceedings of the 15th ACM Workshop on Artificial Intelligence and Security, pages 91-102, 2022. 1", + "[39] Kai Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. arXiv preprint arXiv:2006.09994, 2020. 3, 4", + "[40] Tinghao Xie, Xiangyu Qi, Ping He, Yiming Li, Jiachen T Wang, and Prateek Mittal. Badexpert: Extracting backdoor functionality for accurate backdoor input detection. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR'24), 2024. 2", + "[41] Vacit Oguz Yazici, Abel Gonzalez-Garcia, Arnau Ramisa, Bartlomiej Twardowski, and Joost van de Weijer. Orderless recurrent models for multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'20), pages 13440-13449, 2020. 3", + "[42] Yi Zeng, Won Park, Z Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV'21), pages 16473-16481, 2021. 2, 7", + "[43] Hangtao Zhang, Shengshan Hu, Yichen Wang, Leo Yu Zhang, Ziqi Zhou, Xianlong Wang, Yanjun Zhang, and Chao Chen. Detector collapse: Backdooring object detection to catastrophic overload or blindness. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI'24), 2024. 4, 6", + "[44] Jiawei Zhao, Ke Yan, Yifan Zhao, Xiaowei Guo, Feiyue Huang, and Jia Li. Transformer-based dual relation graph for multi-label image recognition. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV'21), pages 163-172, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24386", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Test-Time Backdoor Detection for Object Detection Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_model.json b/2025/Test-Time Backdoor Detection for Object Detection Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3a67aad331f0012fcdcfc0612f284add9c691837 --- /dev/null +++ b/2025/Test-Time Backdoor Detection for Object Detection Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.131, + 0.8, + 0.154 + ], + "angle": 0, + "content": "Test-Time Backdoor Detection for Object Detection Models" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.179, + 0.834, + 0.218 + ], + "angle": 0, + "content": "Hangtao Zhang1 Yichen Wang1 Shihui Yan1 Chenyu Zhu2 Ziqi Zhou3 Linshan Hou4 Shengshan Hu1 Minghui Li2 Yanjun Zhang5 Leo Yu Zhang6" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.218, + 0.805, + 0.235 + ], + "angle": 0, + "content": "1 School of Cyber Science and Engineering, Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.235, + 0.776, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Software Engineering, Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.253, + 0.817, + 0.27 + ], + "angle": 0, + "content": "\\(^{3}\\) School of Computer Science and Technology, Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.27, + 0.798, + 0.305 + ], + "angle": 0, + "content": "\\(^{4}\\) Harbin Institute of Technology \\(^{5}\\) University of Technology Sydney \\(^{6}\\) Griffith University \\(\\{\\mathrm{hangt\\_zhang, wangyichen, m202472227}\\} \\mathrm{@hust.edu.cn}\\)" + }, + { + "type": "list", + "bbox": [ + 0.18, + 0.218, + 0.817, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.34, + 0.327, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.372, + 0.485, + 0.751 + ], + "angle": 0, + "content": "Object detection models are vulnerable to backdoor attacks, where attackers poison a small subset of training samples by embedding a predefined trigger to manipulate prediction. Detecting poisoned samples (i.e., those containing triggers) at test time can prevent backdoor activation. However, unlike image classification tasks, the unique characteristics of object detection—particularly its output of numerous objects—pose fresh challenges for backdoor detection. The complex attack effects (e.g., “ghost” object emergence or “vanishing” object) further render current defenses fundamentally inadequate. To this end, we design TRansformation Consistency Evaluation (TRACE), a brand-new method for detecting poisoned samples at test time in object detection. Our journey begins with two intriguing observations: 1) poisoned samples exhibit significantly more consistent detection results than clean ones across varied backgrounds. 2) clean samples show higher detection consistency when introduced to different focal information. Based on these phenomena, TRACE applies foreground and background transformations to each test sample, then assesses transformation consistency by calculating the variance in objects' confidence. TRACE achieves black-box, universal backdoor detection, with extensive experiments showing a \\(30\\%\\) improvement in AUROC over state-of-the-art defenses and resistance to adaptive attacks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.798, + 0.222, + 0.814 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Object detection (OD), a key computer vision application [2, 29], has shown vulnerability to backdoor attacks [3, 25]. Backdoor-infected models perform normally on clean objects but predict the attacker-chosen target annotations whenever a trigger (e.g., a small patch) is present. Un" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.34, + 0.907, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.466, + 0.908, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Divergent attack effects of OD backdoors. (a-c): backdoor triggered by a fixed pattern. (d): backdoor triggered by the co-occurrence of two natural objects (a person and an umbrella)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.636 + ], + "angle": 0, + "content": "like classification models, where backdoors solely cause incorrect class predictions, OD backdoors can result in more sophisticated attack effects, i.e., a “ghost” object appearing [3, 7] (Fig. 1(a)), an object vanishing [25, 26, 38] (Fig. 1(b)), the misclassification of all objects globally [3] (Fig. 1(c)), and the co-occurrence of two benign features triggering a backdoor [4, 21] (Fig. 1(d))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.909, + 0.802 + ], + "angle": 0, + "content": "To mitigate backdoor threats in OD, recent studies have made preliminary advances, including trigger reverse engineering [6, 32] and black-box test-time trigger sample detection (TTSD)1 [3]. However, the former relies on white-box settings, requiring access to model weights and training data. By contrast, black-box TTSD detects and prevents poisoned samples during test time without access to model parameters, acting as a final 'firewall'. Blackbox TTSD is highly practical for many real-world applications, such as the prevalent Machine Learning as a Service (MLaaS) [18, 31], and is therefore the focus of our work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.909, + 0.879 + ], + "angle": 0, + "content": "To the best of our knowledge, Detector Cleansse [3] is the only black-box TTsD method tailored for OD tasks. However, it assumes that defenders have full knowledge of the specific details of the attacks (i.e., the adversary-specified judgment criteria, as noted in Tab. 1). In practice, defend" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Also known as online defense [27, 33] or input-level detection [19]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24377" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.224 + ], + "angle": 0, + "content": "
TTSD Method↓Black-BoxPractical (No Training Data)Universal (Attack-Agnostic)Applicable (OD Compatible)
TTBD [16]○●●○
IBD-PSC [19]○○●○
Beatrix [27]○○●○
BaDExpert [40]○○●○
NEO [34]●●○○
Strip [12]●○●●
SCALE-UP [17]●●●●
TeCo [24]●●●●
FreqDetector [42]●○●●
Detector Cleanse [3]●○○●
TRACE (Ours)●●●●
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.233, + 0.483, + 0.271 + ], + "angle": 0, + "content": "Table 1. Comparison among existing TTSD methods w.r.t black-box, practical, universal, and applicable characteristics. “●” indicates the method meets this condition." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.297, + 0.483, + 0.568 + ], + "angle": 0, + "content": "ers cannot possess such prior knowledge. We also apply the current advanced black-box TTSD methods in classification tasks, such as TeCo [24] and SCALE-UP [17], to OD tasks. These TTSDs observe that poisoned samples tend to exhibit more robustness to pre-processing operations than benign samples. Consequently, they apply pixel-level transformations uniformly to all pixel values in each testing sample and assess its prediction consistency. However, applying such uniform pixel-level transformations fails to differentiate between foreground and background components. This distinction is crucial, as the background constitutes the majority of the image, while the foreground (consists of several objects) occupies only a small portion. Recognizing the inherent semantic differences between foreground and background pixels in OD tasks, an intriguing question arises: Shall a semantic-aware transformation expose poisoned samples more effectively than indiscriminate pixel-level modifications?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Fortunately, the answer is yes! In this paper, we reveal that poisoned samples' prediction confidence are significantly more consistent than those of benign ones across different background contexts (i.e., a semantic-aware transformation). Perhaps more interestingly, when we transform it in the opposite way—by introducing additional foreground information (i.e., semantic patches)—the detection results of poisoned samples become less consistent compared to clean samples, providing clues for detecting backdoors. We refer to these phenomena as the anomalous transformation consistency of backdoored models (detailed in Sec. 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Motivated by this finding, we propose TRAnsformation Consistency Evaluation (TRACE), a novel TTSD to identify and filter poisoned samples in OD. For each suspicious test sample, TRACE evaluates its detections consistency. The consistency is measured using two values: 1) the contextual transformation consistency value, as variance in detection results across varying backgrounds. 2) the focal transformation consistency value, as variance in detection results across focal changes. Higher contextual and lower focal consistency both indicate a poisoned input. TRACE combines these metrics to assess if the input is backdoored." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.259 + ], + "angle": 0, + "content": "In conclusion, our main contributions are three-fold. (1) We disclose an intriguing phenomenon, i.e., anomalous contextual and focal transformation consistency in poisoned samples, which brings forth a new detection paradigm: semantic-aware transformation consistency evaluation. (2) Based on our findings, we propose TRACE, a brand-new black-box, universal test-time backdoor detection method to detect and filter out poisoned inputs. (3) We conduct extensive experiments on three benchmark datasets, verifying TRACE's effectiveness against seven backdoor attacks and resistance to potential adaptive attacks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.273, + 0.656, + 0.289 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.45 + ], + "angle": 0, + "content": "Backdoor Defenses for OD are still in their infancy. ODSCAN [6] and Django [32] use trigger reverse engineering but require white-box model access and training data. Detector Cleansse [3] is the only black-box TTSD to the best of our knowledge, which adopts a Strip [12]-like approach by overlaying clean features on objects to observe prediction entropy but assumes full knowledge of the attack. Thus, to counter the diverse attacks in OD, a more practical and universal defense is needed. The detailed background and related work are thoroughly covered in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.906, + 0.541 + ], + "angle": 0, + "content": "TTSD for Classification Tasks are generally hard to apply to OD due to its unique characteristics (w.r.t OD Compatible in Tab. 1). Although some image transformation-based methods like Teco [24], and SCALE-UP [17] can be successfully adapted, their effectiveness diminishes due to shifts in task context." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.554, + 0.9, + 0.667 + ], + "angle": 0, + "content": "Remark I. Arguably, we conclude the widespread failure of existing TTSDs to two key challenges: ① The detector's dual-branch architecture (regression and classification) fundamentally differs from classification models, enabling more sophisticated attacks. ② OD requires identifying \"what\" and \"where\" an object is, increasing output density. This feature makes detecting triggers across a vast output space nearly impossible without deep insights." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.653, + 0.701 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Object Detection. An object detector \\(\\mathbb{F}_{\\theta}\\) (parameterized by \\(\\pmb{\\theta}\\)) processes each input sample \\(\\pmb{x} \\in \\mathcal{X}\\) containing \\(z\\) objects, with ground-truth annotations \\(\\hat{\\pmb{a}} = (\\hat{o}_i, \\hat{y}_i)_{i=1}^z\\) represented as a list of bounding boxes (abbreviated as \"bbox\" for clarity). Here, \\(\\hat{o}_i\\) denotes the \\(i\\)-th object in the image, and \\(\\hat{y}_i\\) is its corresponding class. The detector \\(\\mathbb{F}_{\\theta}\\) outputs \\(P\\) 5-tuples annotation \\(\\pmb{a} = \\mathbb{F}_{\\theta}(\\pmb{x}) = \\{\\pmb{o}_p^{x_1}, \\pmb{o}_p^{y_1}, \\pmb{o}_p^{x_2}, \\pmb{o}_p^{y_2}, \\pmb{y}_p\\}_{p=1}^P\\), with \\((x_1, y_1)\\) and \\((x_2, y_2)\\) as the coordinates, and \\(\\pmb{y}_p\\) denoting the class probability of the \\(p\\)-th predictedbbox. A true-positive (TP) is a correctly detectedbbox in an input sample, while a false-positive (FP) refers to any non-TP detectedbbox, and a false-negative (FN) is any undetected ground-truthbbox." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24378" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.398, + 0.106 + ], + "angle": 0, + "content": "3.1. OD Backdoor Attacks Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.206 + ], + "angle": 0, + "content": "Assume the \\(l\\)-th object \\(\\hat{o}_l\\) to be the victim object (i.e., an object whose recognition is compromised by triggers). The backdoorsed model, while preserving utility, outputs anchors perfectly matching the ground truth \\(\\hat{a}\\), i.e., \\(\\mathbb{F}_{\\theta}(\\boldsymbol{x}) = \\{(o_i,y_i)\\}\\), \\(\\forall i \\in [1,z]\\), \\(o_i = \\hat{o}_i\\), \\(y_i = \\hat{y}_i\\). Next, we adopt the backdoor categorization introduced by Cheng et al. [6]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.484, + 0.344 + ], + "angle": 0, + "content": "\\(\\bullet\\) FP-inducing attack manipulates the model to recognize a background area as a target class \\(\\pmb{y}_t\\). The appearing object may be localized at a fixed position, the trigger position \\(\\pmb{o}_t\\), or the victim object location \\(\\hat{o}_l\\). We denote the appearing region as \\(M(\\pmb{o}_t,\\hat{\\pmb{o}}_l)\\) (a function of both \\(\\pmb{o}_t\\) and \\(\\hat{o}_l\\)). Given a trigger \\(\\pmb{t}\\), \\(\\pmb{x} \\oplus \\pmb{t} = (1 - \\alpha) \\otimes \\pmb{x} + \\alpha \\otimes \\pmb{t}\\) denotes applying the trigger to the clean image with transparency \\(\\alpha\\). As per Cheng et al. [6], we define the FP-inducing attack as \\((M(\\pmb{o}_t,\\hat{\\pmb{o}}_l),\\pmb{y}_t) \\in \\mathbb{F}_{\\theta}(\\pmb{x} \\oplus \\pmb{t})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.483, + 0.406 + ], + "angle": 0, + "content": "2FN-inducing attack is also known as evasion attack [37], causes models to overlook objects of interest (i.e., generating FNs). Given the victim object \\(\\hat{\\pmb{o}}_l\\) and its class \\(\\hat{\\pmb{y}}_l\\), we define the FN-inducing attack as \\((\\hat{\\pmb{o}}_l,\\hat{\\pmb{y}}_l)\\notin \\mathbb{F}_{\\theta}(\\pmb {x}\\oplus \\pmb {t})\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.532 + ], + "angle": 0, + "content": "General Formulation of Backdoor Attacks. We categorize OD backdoors into \\( FP \\)-inducing, \\( FN \\)-inducing attacks, and combinations of both (hybrid attack). For instance, object misclassification attack in Fig. 1(c) (an example of hybrid attack) involve recognizing an object \\( o_j \\) as a different target \\( o_k \\). This process can be decomposed into the disappearance of the original object \\( o_j \\) (i.e., \\( FN \\)-inducing attack) and the appearance of a new object \\( o_k \\) (i.e., \\( FP \\)-inducing attack). We give a general definition of backdoors as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.547, + 0.482, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\left(\\hat {\\boldsymbol {o}} _ {l}, \\hat {\\boldsymbol {y}} _ {l}\\right) \\notin \\mathbb {F} _ {\\theta} (\\boldsymbol {x} \\oplus \\boldsymbol {t}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.564, + 0.482, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\left(M \\left(\\boldsymbol {o} _ {t}, \\hat {\\boldsymbol {o}} _ {l}\\right), \\boldsymbol {y} _ {t}\\right) \\in \\mathbb {F} (\\boldsymbol {x} \\oplus \\boldsymbol {t}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.731 + ], + "angle": 0, + "content": "Eq. 1 means the absence of the original victim object from the output anchor set. While Eq. 2 means the presence of the target object, with its shape and location determined by the function \\( M(\\cdot) \\). These equations generalize previous attacks. Formally, the misclassification attack is a special case where \\( M(\\boldsymbol{o}_t, \\hat{\\boldsymbol{o}}_l) = \\hat{\\boldsymbol{o}}_l \\). While in object appearing attack, where the trigger is recognized as the target object, \\( M(\\boldsymbol{o}_t, \\hat{\\boldsymbol{o}}_l) = \\boldsymbol{o}_t \\). Details on representing existing attacks based on this definition are in Appendix E.1." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.746, + 0.478, + 0.804 + ], + "angle": 0, + "content": "Remark II: We believe that a general definition is crucial for establishing a universal TTSD method. Otherwise, independent discriminators or criteria must be constructed for each attack and applied sequentially to suspicious inputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.235, + 0.835 + ], + "angle": 0, + "content": "3.2. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.844, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Defender's Goals. TTSD methods operate on a trained backdoored model \\(\\mathbb{F}_{\\theta}\\) and a test dataset comprising both poisoned samples and clean samples \\(S = \\{\\tilde{S} \\cup S\\}\\). The purpose is to discover a method \\(\\mathcal{T}\\) that can maximize the" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.089, + 0.897, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.245, + 0.907, + 0.296 + ], + "angle": 0, + "content": "Figure 2. Effect of background shift on (a) ResNet-50 and (b)YOLO [29], Faster-RCNN [30], DETR [2]. (b): each background case shows the max, min, and average confidence (as line plots) across three detectors. Larger shaded areas indicate greater instability in object predictions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.322, + 0.903, + 0.337 + ], + "angle": 0, + "content": "separation of trigger inputs from clean ones, formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.345, + 0.905, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} = \\arg \\max _ {\\mathcal {T}} \\mathbb {E} _ {(\\tilde {x}) \\sim \\mathcal {P} _ {\\tilde {S}}} \\mathbb {I} (\\mathcal {T} (\\tilde {x}, \\mathbb {F} _ {\\theta}) = 1) + \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.68, + 0.369, + 0.846, + 0.383 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {(x) \\sim \\mathcal {P} _ {S}} \\mathbb {I} \\left(\\mathcal {T} (x, \\mathbb {F} _ {\\theta}) = 0\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.905, + 0.422 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}\\) is an indicator function, equal to 1 if the condition is true, and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.906, + 0.545 + ], + "angle": 0, + "content": "Defender's Capabilities. We adopt standard TTSD settings [12, 24, 34], where the defender uses third-party models and detects poisoned samples on the fly. In particular, we assume a realistic defender with no prior knowledge of attacks and only black-box access to the backdoored model (i.e., can only obtain the final decision from the models). The defender has no training dataset but can access publicly available background and foreground images as support." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.727, + 0.574 + ], + "angle": 0, + "content": "4. Insight behind TRACE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Contextual bias: recognition leans on the co-occurrence context more than on the object itself. Recent efforts have moved from RNNs [35, 41] to GCNNs [5, 10] and transformer-based frameworks [44] to model contextual relationships in multi-label images and improve performance. However, it is also a double-edged sword. Fig. 2(a) shows an example from the Background Challenge [39] reveals how ResNet-50 misclassifies a bird as a fish due to the underwater background. Note that the object detector's learned representations also tend to exploit such spurious scene correlations [23], which can mislead detectors by either giving nonexistent objects high scores in the scene where they typically appear or ignoring objects appearing in rare backgrounds. In other words, when networks find that context alone suffices to identify most objects, they often neglect the representation of the instance, because neural networks are statistics-based and \"lazy\" [13]. Deep analysis is moved to Appendix E.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Interestingly, we find that trigger objects stay confident no matter the scene. Fig. 2(b) shows three detector's prediction confidences for different objects (extracted using" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24379" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.468, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.232, + 0.477, + 0.246 + ], + "angle": 0, + "content": "Figure 3. YOLO activation map: clean objects vs. FN-inducing triggers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.483, + 0.452 + ], + "angle": 0, + "content": "semantic segmentation) placed on 10 varied backgrounds, including an \\( FP \\)-inducing trigger (Fig. 1(a)) with the target annotations \"person\". We observe that the trigger maintains consistent confidence scores across all backgrounds, while clean objects show notable fluctuations. We attribute this phenomenon to the widely accepted concept of \"shortcut learning\" [14], where the backdoor training process establishes a robust mapping between a specific, uniform pattern and the target label. This makes the recognition of trigger objects less susceptible to contextual biases, leaving us clues for detecting \\( FP \\)-inducing attacks. Stay tuned for Sec 5.1 to see how this insight is leveraged." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.71 + ], + "angle": 0, + "content": "\\(\\Theta\\) Regression continuity: backdoor triggers propagate influence spatially, affecting neighboring pixels. OD differs from classification tasks by incorporating regression branches that rely on shared spatial features. We find that this continuity also allows triggers to impact surrounding regions, creating a black hole-like effect. This is particularly evident for \\(FN\\)-inducing attacks, where the model interprets the affected area as background, ignoring objects of interest [7, 43]. For instance, in Faster R-CNN, where the Region Proposal Network generates around 17, 100 candidate anchor boxes, a trigger affects many nearby anchors, thereby creating the apparent spatial influence. We further diagnose model predictions with and without the trigger using EigenCAM [28] (see Fig. 3), revealing that activations around the trigger diminish, shifting the model's attention to distant regions. Such sharp, regional changes in the saliency map provide clues for detecting \\(FN\\)-inducing attacks." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.722, + 0.477, + 0.82 + ], + "angle": 0, + "content": "Remark III: These insights reveal from different perspectives the key differences between poisoned and clean samples. Additionally, under the introduced general backdoor formulation (Sec. 3.1), the detection of both foundational attack primitives (FN and FP-inducing) are elegantly covered, paving the way for truly black-box, universal detection without the need for training data or attack knowledge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.388, + 0.863 + ], + "angle": 0, + "content": "5. TRACE: A Complete Illustration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Here we present TRACE, composed of contextual information transformation (Sec. 5.1), focal information transfor" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.089, + 0.877, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.233, + 0.887, + 0.246 + ], + "angle": 0, + "content": "Figure 4. Background blending and YOLO confidence distribution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.272, + 0.865, + 0.287 + ], + "angle": 0, + "content": "mation (Sec. 5.2), and test-time evaluation (Sec. 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.296, + 0.858, + 0.311 + ], + "angle": 0, + "content": "5.1. Contextual Information Transformation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.906, + 0.424 + ], + "angle": 0, + "content": "Sec. 4 has explored how DNNs can pick up biases or unintended shortcuts from context (a.k.a. 'Clever Hans' effect [1]). This is intuitive, as cars are typically found on roads, not in locations like the sky or volcanic craters. In contrast, trigger objects exhibit anomalously stable confidence scores. This is further formalized as Contextual Transformation Consistency in Definition 1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.908, + 0.509 + ], + "angle": 0, + "content": "Definition 1 (Contextual Transformation Consistency (CTC)) Given an object \\( \\mathbf{o} \\) and a background distribution \\( \\mathcal{B} \\), the CTC of \\( \\mathbf{o} \\) is defined as the maximum deviation in the model's prediction score for \\( \\mathbf{o} \\)'s class \\( \\mathbf{y} \\) when applied to different backgrounds \\( \\delta \\sim \\mathcal{B} \\) (with the operation of \\( \\mathbf{o} \\oplus \\delta \\)):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.52, + 0.904, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {\\mathcal {B}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {o}\\right) := \\max _ {\\delta_ {1}, \\delta_ {2} \\sim \\mathcal {B}} \\left\\{\\left| \\mathbb {F} _ {\\theta} \\left(\\boldsymbol {o} \\oplus \\delta_ {1}\\right) _ {\\boldsymbol {y} (\\boldsymbol {o})} - \\mathbb {F} _ {\\theta} \\left(\\boldsymbol {o} \\oplus \\delta_ {2}\\right) _ {\\boldsymbol {y} (\\boldsymbol {o})} \\right| \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.825 + ], + "angle": 0, + "content": "It is observed that \\(\\Delta_{\\mathcal{B}}(\\mathbb{F},\\boldsymbol{o}_t)\\ll \\Delta_{\\mathcal{B}}(\\mathbb{F},\\boldsymbol{o}_c)\\), indicating that the CTC value is significantly smaller for a trigger object \\(\\boldsymbol{o}_t\\) than for a clean object \\(\\boldsymbol{o}_c\\). With this in mind, we propose leveraging this discrepancy to detect the presence of triggers. However, sequentially extracting individual objects using segmentation algorithms (like in Fig. 2 (a)) and overlaying them onto different backgrounds before querying with the backdoored model is computationally intensive. We propose directly blending background images onto the original images with an opacity \\(\\alpha_{bg}\\) (i.e., \\((1 - \\alpha_{bg})\\otimes x + \\alpha_{bg}\\otimes \\delta\\), see Fig. 4, left). This allows for background replacement at the image level rather than the object level. By using a lower \\(\\alpha_{bg}\\) to minimize alterations to the object's pixel values, we thus iteratively introduce diverse contextual information (from a background set of size \\(b\\) available online, e.g., the Background Challenge dataset [39]) while mitigating the impact of backgrounds on the object's representations. Visualizations are provided in Appendix C.4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Instead of directly using the CTC value, we consider potential corner cases where certain objects might unexpectedly become undetectable. To account for these cases, we use the variance of the object's confidence scores across different background variations (sampling uniformly from \\(\\mathcal{B}\\))" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24380" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.091, + 0.468, + 0.108 + ], + "angle": 0, + "content": "to enhance robustness, denoted as \\(\\Delta_B^{\\mathrm{Var}}(\\mathbb{F}_\\theta ,\\pmb {o})\\) , given by" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.119, + 0.482, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {o} _ {i}\\right) = V a r \\left(\\left\\{\\mathbb {F} _ {\\theta} (\\boldsymbol {o} _ {i} \\oplus \\delta) _ {\\boldsymbol {y} (\\boldsymbol {o})} \\right\\} _ {i = 1} ^ {b}\\right), \\delta \\sim \\mathcal {U} (\\mathcal {B}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.156, + 0.483, + 0.261 + ], + "angle": 0, + "content": "Fig. 4 shows the violin plots of confidence distributions for different objects blended with 50 backgrounds at \\(\\alpha_{bg} = 0.15\\), including an FP-inducing trigger (see Fig. 1(a)) with the target annotations \"person\". Empirically, as expected, the trigger exhibits a low variance of just 0.008, achieving clear separability in distribution. Explorations for different objects and detectors are provided in Appendix D.1." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.262, + 0.485, + 0.595 + ], + "angle": 0, + "content": "Natural backdoor: neural networks also overfit to uniform, salient benign representations. Detecting FP-inducing triggers, however, does not stop there. Note that certain objects with standardized shapes and colors—such as uniformly red, polygonal stop signs—are also overconfidently recognized by networks, much like backdoor triggers, as DNNs more readily memorize these easily learned, superficial features [15]. We consider these a unique form of benign triggers, which we term \"natural backdoor objects\" (NBOs). Such uniform, salient representations allow them to rely less on additional context for recognition, showing anomalous CTC values similar to triggers. Consequently, we integrate a post-processing module to filter and remove them. Here we introduce a visual coherence analysis to screen suspicious objects exhibiting low \\(\\Delta_{\\mathcal{B}}^{\\mathrm{Var}}\\). By comparing these objects with reference images corresponding to their categories from online available data, we evaluate their SSIM scores [20]. Note that these references, which we term \"universal visual benchmarks\" (see Appendix C.5), do not need to be sourced from the model's training data, as they represent general visual cognition. Since an image may contain multiple triggering objects, we naturally define the image-level CTC as the minimum CTC values among all (SSIM-filtered) objects in the image, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.604, + 0.432, + 0.676 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underbrace {\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta} , \\boldsymbol {x}\\right)} = \\min \\left\\{\\underbrace {\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta} , \\boldsymbol {o} _ {i}\\right)} \\right| \\\\ \\begin{array}{c} \\text {I m a g e - l e v e l V a r i a n c e} \\\\ \\hline \\end{array} \\\\ s. t. \\quad S S I M \\left(\\boldsymbol {o} _ {i}, \\boldsymbol {o} _ {y _ {i}} ^ {\\text {r e f}}\\right) \\leq \\tau , \\quad i = 1, \\dots , z \\Bigg \\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.748 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a predefined threshold, and \\(o_{y_i}^{\\mathrm{ref}}\\) is the universal visual benchmark for class \\(y_i\\). Clearly, the lower the image-level CTC values, the more likely it is that the image contains the trigger objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.393, + 0.773 + ], + "angle": 0, + "content": "5.2. Focal Information Transformation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Why do we still need another transformation? We have demonstrated that CTC values can identify FP-inducing triggers among visible objects in the image. However, FN-inducing triggers (see Fig. 1(b)) remain entirely absent in detection results, yielding no CTC value. How to detect these \"unseen objects\" thus becomes a non-trivial challenge. Sec. 4 highlights the shifts in saliency maps when comparing samples with and without triggers. However," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": "interpretability methods like Grad-CAM typically require white-box model access, which is out of reach for our black-box defender. Hence, instead of passively analyzing the input sample, we seamlessly inject new elements (i.e., small clean object patches) to proactively \"reconstruct\" saliency maps. We begin by sliding a 'car' object across an image in tiny steps—like how a convolutional kernel scans over pixels—recording its prediction confidence at each position. Unfortunately, this object is fragile, with confidence fluctuating by position (see Fig. 5(a)). A representative visual example is the surfboard in Fig. 5(b), which shows a strong positional dependency, becoming undetectable when shifted even slightly. This inherent fragility causes such objects to fail in probing the model's focus regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.905, + 0.59 + ], + "angle": 0, + "content": "Although NBOs may initially seem to hinder our detection process, we argue that this trait is precisely the key to resolving the above issue. Fig. 5(c) shows the sliding results for an NBO ('stop sign'). The model consistently detects it with high confidence, demonstrating position-invariant detection [36]. This makes it particularly suitable for probing the saliency maps. Hence, this time we slide the NBO across the same image but with a FN-inducing trigger placed at its center. Fig. 5(d) shows its high confidence at the image edges, dropping sharply near the central trigger. When it completely covers the trigger, the model's normal detection resumes. We term this distinctive behavior the \"Island Effect\", as it resembles an island rising from surrounding waters. The NBO undergoes two sharp changes in confidence when handling poisoned samples, while remaining stable on clean images. Leveraging this property, we can thus proactively probe potential FN-inducing triggers in the troublesome black-box setting. We formalize Focal Transformation Consistency in Definition 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.598, + 0.907, + 0.671 + ], + "angle": 0, + "content": "Definition 2 (Focal Transformation Consistency (FTC)) Given an image \\( \\pmb{x} \\) defined over a spatial domain \\( \\Omega \\) and an object patch \\( \\pmb{o} \\), a focal transformation \\( \\mathcal{K} : (\\pmb{x}, \\pmb{o}, \\pmb{p}) \\to \\pmb{x}^p \\) applies \\( \\pmb{o} \\) to \\( \\pmb{x} \\), centered at the spatial coordinate \\( \\pmb{p} = (\\pmb{p}_x, \\pmb{p}_y) \\in \\Omega \\). The FTC value is then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.678, + 0.865, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta_ {\\mathcal {F}} (\\mathbb {F} _ {\\boldsymbol {\\theta}}, \\pmb {x}, \\pmb {o}) := \\int_ {\\Omega} \\left(\\left\\| \\nabla^ {2} \\mathbb {F} _ {\\boldsymbol {\\theta}} \\big (\\mathcal {K} (\\pmb {x}, \\pmb {o}, \\pmb {p}) \\big) _ {\\pmb {y} (\\pmb {o})} \\right\\| _ {1} + \\right. \\\\ \\left. \\right.\\left\\| \\right.\\left[ \\mathbb {F} _ {\\theta} \\left(\\mathcal {K} (\\boldsymbol {x}, \\boldsymbol {o}, \\boldsymbol {p} + \\Delta \\boldsymbol {p}) - \\mathbb {F} _ {\\theta} \\left(\\mathcal {K} (\\boldsymbol {x}, \\boldsymbol {o}, \\boldsymbol {p})\\right)\\right)\\right] _ {+} \\| _ {1}\\left. \\right) d \\boldsymbol {p} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.907, + 0.862 + ], + "angle": 0, + "content": "where \\(\\nabla^2\\) is the Laplacian operator capturing second-order spatial derivatives, \\(\\mathbb{F}_{\\theta}(\\mathcal{K}(\\pmb{x},\\pmb{o},\\pmb{p}))_{\\pmb{y}(\\pmb{o})}\\) measures the confidence score of object \\(\\pmb{o}\\) after applying the transformation \\(\\mathcal{K}(\\cdot)\\), and \\([\\cdot]_+\\) denotes the positive part function (i.e., \\([a]_+ = \\max(0,a)\\)), ensuring that the quantification of detection shift focuses on the incremental components of the model's outputs, specifically capturing newly emerged detections (i.e., results absent in the original predictions)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "FTC has two components: the confidence oscillation of \\( \\pmb{o} \\) and the image-level detection shift caused by \\( \\pmb{o} \\). The first" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24381" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.098, + 0.278, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.213, + 0.282, + 0.24 + ], + "angle": 0, + "content": "(a) Confidence of \"car\" sliding across a clean image" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.089, + 0.448, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.217, + 0.437, + 0.232 + ], + "angle": 0, + "content": "(b) \"fragile\" object" + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.099, + 0.621, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.211, + 0.65, + 0.239 + ], + "angle": 0, + "content": "(c) Confidence of \"stop sign\" sliding across a clean image" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.099, + 0.905, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.212, + 0.866, + 0.24 + ], + "angle": 0, + "content": "(d) Confidence of \"stop sign\" sliding across a trigger image" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.25, + 0.846, + 0.264 + ], + "angle": 0, + "content": "Figure 5. Effect of the foreground object's position on YOLO detections. (d) uses an FN-inducing trigger placed at the image center." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.29, + 0.484, + 0.426 + ], + "angle": 0, + "content": "metric assesses prediction consistency of \\( o \\) across different positions in image \\( x \\), helping detect the \"Island Effect\" of FN-inducing triggers, where each cross-section shows two sharp confidence drops. The second captures changes in whole detection results due to \\( o \\). Under normal circumstances, moving \\( o \\) will not introduce new objects. We use it to detect cases when new objects suddenly appear as \\( o \\) covers the trigger. This trigger-blocking strategy is also supported by established backdoor defenses [8, 34]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.635 + ], + "angle": 0, + "content": "NBOs demonstrate strong location-invariant recognition, as choosing \\( \\pmb{o} = \\pmb{o}_{\\mathrm{nbo}} \\) decouples \\( \\pmb{o} \\)'s contribution to \\( \\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}) \\), ensuring \\( \\partial \\Delta_{\\mathcal{F}} / \\partial \\pmb{o} \\approx 0 \\) under spatial variations in \\( \\pmb{p} \\). Thus, with \\( \\pmb{o} \\) effectively neutralized, the FTC value simplifies to reflect differences in \\( \\pmb{x} \\), yielding a higher \\( \\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\tilde{\\pmb{x}}, \\pmb{o}_{\\mathrm{nbo}}) \\gg \\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}_{\\mathrm{nbo}}) \\) for poisoned samples \\( \\tilde{\\pmb{x}} \\). Instead of densely querying via iteratively sliding, we seamlessly employ Monte Carlo sampling of \\( k \\) distinct, non-overlapping points to patch \\( \\pmb{o}_{\\mathrm{nbo}} \\) (see Fig. 6), thereby reducing the number of required model queries and improving computational efficiency. This unit operation is then repeated \\( f \\) times to build an approximate continuous saliency map. Finally, we can compute the variance of \\( \\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}_{\\mathrm{nbo}}) \\) across all accumulated sampling points from \\( \\mathcal{U}(\\Omega) \\) to determine the FTC value for the sample \\( \\pmb{x} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.643, + 0.496, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {\\mathcal {F}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x}, \\boldsymbol {o} _ {\\mathrm {n b o}}\\right) := V a r \\left(\\left\\{\\Delta_ {\\mathcal {F}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x}, \\boldsymbol {o} _ {\\mathrm {n b o}}\\right) \\mid \\boldsymbol {p} _ {i} \\sim \\mathcal {U} (\\boldsymbol {\\Omega}) \\right\\} _ {i = 1} ^ {f \\cdot k}\\right).\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.684, + 0.354, + 0.699 + ], + "angle": 0, + "content": "5.3. Test-time TRACE Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.905 + ], + "angle": 0, + "content": "To achieve Eq. 3, the TTSD methods must exploit the disparities between benign images and those containing triggers. Our findings in Sec. 5.1 and 5.2 reveal that backdoored models exhibit anomalous consistency in both contextual and focal transformation. In other words, the distinct properties of CTC and FTC make them particularly suitable for detecting FP-inducing and FN-inducing attacks, respectively. As established in Sec 3.1, a trigger sample can be a single type or a combination of both. Consequently, since a smaller CTC \\((\\Delta_{\\mathcal{B}}^{\\mathrm{Var}})\\) and a larger FTC value \\((\\Delta_{\\mathcal{F}}^{\\mathrm{Var}})\\) both indicate higher anomaly, the final step is to measure the overall image-level anomaly as \\(Trace(\\pmb{x}) = \\sigma \\left( \\Delta_{\\mathcal{F}}^{\\mathrm{Var}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}_{\\mathrm{nbo}}) \\right) - \\sigma \\left( \\Delta_{\\mathcal{B}}^{\\mathrm{Var}}(\\mathbb{F}_{\\theta}, \\pmb{x}) \\right)\\), where \\(\\sigma(t) = \\frac{1}{1 + e^{-t}}\\) is the sigmoid function that normalizes both variance" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.287, + 0.905, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.529, + 0.818, + 0.542 + ], + "angle": 0, + "content": "Figure 6. The framework of our TRACE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.611 + ], + "angle": 0, + "content": "measures to the range \\((0,1)\\). TRACE thus maps the input image \\(x\\) to a linearly separable space, enabling defenders to make decisions by a threshold \\(\\gamma\\):" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.618, + 0.907, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\Gamma (T r a c e (\\boldsymbol {x})) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f T r a c e} (\\boldsymbol {x}) > \\gamma \\\\ 0, & \\text {i f T r a c e} (\\boldsymbol {x}) \\leq \\gamma \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.631, + 0.68 + ], + "angle": 0, + "content": "6. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Datasets and Models. We select popular benchmark datasets, including MS-COCO [22], PASCAL VOC [11], and Synthesized Traffic Signs [32]. The models we choose are three representative object detectors: single-stage detector: YOLOv5 [29], two-stage detector: Faster-RCNN [30] and vision transformer-based detector: DETR [2]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Attack Methods. We evaluate TRACE against seven SOTA OD backdoors: OGA, RMA, GMA, and ODA from [3], as well as CIB [4], UTA [25], and DC [43]. For CIB and DC, which have multiple goals, we focus on their object disappearance variant as a representative case. A detailed description of these attacks can be found in Appendix C.6." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Defense Settings. We compare TRACE with Detector Cleansse [3], the only current black-box TTSD tailored" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24382" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.138 + ], + "angle": 0, + "content": "Table 2. F1 score and AUROC for different defenses on Synthesized Traffic Signs, PASCAL VOC, and MS-COCO, using YOLO, DETR, and Faster-RCNN. Due to space constraints, some results for two attacks are presented in the same row, separated by “/”. “-” denotes that FreqDetector is unavailable in this case, as it detects frequency-domain anomalies in trigger patterns, while CIB uses benign features to trigger the backdoors. We bold the F1 and AUROC of the strongest defenses in each case." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.15, + 0.856, + 0.496 + ], + "angle": 0, + "content": "
Model↓Method AttackTRACE (Ours)Teco [24]Detector cleansse [3]Strip [12]FreqDetector [42]SCALE-UP [17]
F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑
MS-COCO [22]
YOLOOGA/ODA0.937/0.8650.913/0.8800.464/0.5210.593/0.5920.702/0.6830.650/0.6920.646/0.4830.729/0.5030.542/0.5880.490/0.5800.503/0.4890.505/0.463
RMA/GMA0.922/0.8450.932/0.9240.397/0.5010.487/0.4580.674/0.6620.590/0.6080.391/0.3590.474/0.5250.567/0.5120.498/0.4380.482/0.4730.510/0.440
Untargeted0.8370.8680.4980.5500.6890.6690.4850.4720.6140.6460.5030.491
CIB0.7980.8170.6510.6910.6100.6620.5320.578--0.2610.367
DC0.9340.9680.5460.6750.6400.7300.5250.4930.5830.6640.4700.506
Average0.8770.9000.5110.5780.6650.6570.4890.5390.5680.5530.4540.460
MS-COCO [22]
YOLOOGA/ODA0.801/0.8390.825/0.8570.593/0.5800.603/0.5770.676/0.6370.541/0.6790.646/0.5300.649/0.5360.542/0.5880.490/0.5800.481/0.5240.478/0.549
RMA/GMA0.807/0.8160.777/0.8460.534/0.5060.478/0.4380.686/0.6490.629/0.5880.498/0.5920.452/0.5800.567/0.5150.498/0.4380.497/0.4620.486/0.456
Untargeted0.7880.7900.5460.5770.6980.6550.5070.4860.6140.6460.5180.528
CIB0.7200.7560.6310.6760.5830.6320.5220.548--0.2350.280
DC0.8380.9140.5760.6350.7030.7210.5040.4830.5830.6640.4800.518
Average0.8010.8240.5670.5690.6620.6350.5430.5330.5680.5530.4570.471
Faster-RCNNOGA/ODA0.757/0.8540.770/0.8560.601/0.6100.601/0.6240.716/0.6700.591/0.6060.631/0.4970.636/0.5460.542/0.5880.490/0.5800.498/0.4810.482/0.469
RMA/GMA0.872/0.8700.899/0.8830.563/0.5190.535/0.4910.721/0.6590.655/0.5810.471/0.5300.471/0.5330.567/0.5120.498/0.4390.495/0.4910.491/0.503
Untargeted0.8690.8420.5770.5710.6700.6550.5590.5640.6140.6460.4940.471
CIB0.7470.7500.6380.6820.6130.6020.5850.603--0.2620.329
DC0.9350.9440.5590.6670.6270.7450.4750.5280.5830.6640.4770.454
Average0.8430.8490.5810.5970.6680.6330.5350.5540.5680.5530.4570.457
PASCAL VOC [11]
YOLOOGA/ODA0.928/0.9010.945/0.9050.536/0.5430.561/0.5710.670/0.7210.631/0.6520.612/0.5110.664/0.4450.500/0.6010.485/0.6370.441/0.5220.458/0.505
RMA/GMA0.916/0.8730.908/0.9390.445/0.5140.444/0.4790.619/0.6630.585/0.5550.563/0.4470.626/0.5130.562/0.4770.584/0.5020.515/0.5080.495/0.489
Untargeted0.8000.8580.4870.4970.7060.7010.5510.5030.6440.6870.4820.490
CIB0.7850.7670.6100.6560.6180.6240.5670.598--0.2730.332
DC0.9360.9440.5020.6550.6830.7920.5090.5050.5630.6140.5340.559
Average0.8770.8950.5200.5520.6690.6490.5370.5510.5580.5850.4680.475
DETROGA/ODA0.775/0.8230.799/0.8310.524/0.5150.547/0.5200.703/0.6080.627/0.6370.610/0.5190.660/0.4670.500/0.6010.485/0.6370.483/0.5050.497/0.521
RMA/GMA0.803/0.7660.786/0.7740.466/0.4880.450/0.4470.651/0.6220.563/0.5500.559/0.5330.584/0.5620.562/0.4770.584/0.5020.541/0.4770.559/0.483
Untargeted0.8310.8200.5670.5720.6550.7050.6470.6920.6440.6870.4890.476
CIB0.7500.7670.6320.6830.5910.5870.5750.569--0.2750.346
DC0.9600.9460.5590.6260.6360.7240.5600.4720.5630.6140.4970.502
Average0.8150.8180.5360.5490.6380.6280.5720.5720.5580.5850.4670.483
Synthesized Traffic Signs [32]
YOLOOGA/ODA0.917/0.8850.942/0.8720.581/0.5910.581/0.5600.693/0.6400.665/0.6640.671/0.4420.645/0.4670.504/0.6060.529/0.5610.473/0.4530.480/0.446
RMA/GMA0.957/0.9280.944/0.9360.515/0.4850.459/0.4900.678/0.6900.605/0.5490.494/0.5110.523/0.5790.562/0.5220.506/0.4240.505/0.4810.517/0.472
Untargeted0.8570.9030.5500.5010.7030.6450.4670.4410.6390.6130.5170.494
CIB0.8250.8530.6280.6490.6550.6730.5480.591--0.2360.354
DC0.9260.9410.5570.6890.7010.7540.5180.5230.5900.5430.5110.532
Average0.8990.9140.5580.5610.6800.6510.5220.5380.5710.5290.4540.471
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.581 + ], + "angle": 0, + "content": "for OD. We also include four black-box defenses applicable to OD: Strip [12], FreqDetector [42], TeCo [24], and SCALE-UP [17]. We follow the configurations recommended in their papers for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate using two metrics: F1 score (balancing precision and recall) and AUROC (measuring overall performance across different thresholds)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.292, + 0.652 + ], + "angle": 0, + "content": "6.1. Defense Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Effectiveness Study. In Tab. 2, TRACE consistently achieves promising performance across seven backdoor attacks, three datasets, three model architectures, and a total of 42 backdoors models. To further assess TRACE's stability, we calculate performance standard deviations across different backdoors on the same model and dataset, finding low standard variances (F1: \\(0.880 \\pm 0.040\\), AUROC: \\(0.897 \\pm 0.041\\)), which demonstrates its robust consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparison Study. Tab. 2 also shows that TRACE comprehensively outperformed its five black-box competitors: TeCo, Strip, FreqDetector, SCALE-UP, and the SOTA Detector Cleansed designed for OD task. The first four fail completely due to shifts in task context. Even though Detector Cleansed is not attack-agnostic, TRACE still significantly surpasses it in F1 score by \\(\\sim 30\\%\\). Further analysis is deferred to Appendix D.2." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.518, + 0.919, + 0.529 + ], + "angle": 0, + "content": "(a) clean input (a) heatmap (b) clean input (b) heatmap (c) trigger input (c) heatmap" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.529, + 0.587, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.529, + 0.65, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.529, + 0.721, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.529, + 0.783, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.529, + 0.855, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.855, + 0.529, + 0.922, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.58, + 0.588, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.58, + 0.651, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.58, + 0.721, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.58, + 0.783, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.58, + 0.855, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.855, + 0.58, + 0.922, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.633, + 0.585, + 0.642 + ], + "angle": 0, + "content": "CTC trans. input" + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.634, + 0.718, + 0.642 + ], + "angle": 0, + "content": "CTC trans. input" + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.634, + 0.853, + 0.643 + ], + "angle": 0, + "content": "CTC trans. input" + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.654, + 0.899, + 0.668 + ], + "angle": 0, + "content": "Figure 7. High resolution, object-specific explanations from L-CRP [9]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "A Closer Look to the Effectiveness of TRACE. We use SOTA eXplainable Artificial Intelligence (XAI) method L-CRP [9] to attain object-class-specific explanations in this multi-object context. We offer three illustrative examples: Fig. 7(a) shows that the detection of a banana relies heavily on surrounding green plants. TRACE's background transformation (i.e., introducing a contrasting fire scene) breaks this context bias, shifting reliance to the banana's intrinsic features. In Fig. 7(b), the bird's detection initially depends on its own features, but after transformation, new contextual dependencies emerge. In both cases, confidence scores shift (decrease for the banana, increasing for the bird). Lastly, Fig. 7(c) shows an FP-inducing trigger, showing always input-independent detection." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.287, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.198, + 0.284, + 0.21 + ], + "angle": 0, + "content": "(a) Defending RMA (hybrid attack)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.089, + 0.477, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.199, + 0.475, + 0.21 + ], + "angle": 0, + "content": "(b) Defending GMA (hybrid attack)" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.212, + 0.435, + 0.224 + ], + "angle": 0, + "content": "Figure 8. (Ablation Study) Effect of different modules." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.23, + 0.221, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.23, + 0.351, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.231, + 0.48, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.307, + 0.221, + 0.319 + ], + "angle": 0, + "content": "(a) Background queries" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.307, + 0.349, + 0.319 + ], + "angle": 0, + "content": "(b) Foreground queries" + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.307, + 0.447, + 0.318 + ], + "angle": 0, + "content": "(c) \\(\\tau\\) value" + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.321, + 0.47, + 0.333 + ], + "angle": 0, + "content": "Figure 9. (Ablation Study) Effect of background, foreground, and \\(\\tau\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.338, + 0.482, + 0.35 + ], + "angle": 0, + "content": "Table 3. (Adaptive attack) TRACE against the omniscient adaptive attack." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.361, + 0.485, + 0.416 + ], + "angle": 0, + "content": "
λ→Attacks↓5 × 10-41 × 10-45 × 10-51 × 10-5
F1AUROCASRF1AUROCASRF1AUROCASRF1AUROCASR
OGA0.5810.6200.7330.7660.7470.8430.9160.8790.9440.9370.9420.975
ODA0.6060.5940.5830.6080.5760.7520.8200.7940.8410.8400.8510.893
RMA0.6170.6140.6810.7480.7020.8100.8670.8860.9330.9380.9290.947
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.245, + 0.456 + ], + "angle": 0, + "content": "6.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.598 + ], + "angle": 0, + "content": "Effect of Modules. TRACE incorporates three fundamental designs: contextual transformation, focal transformation, and SSIM-filter post-processing. We evaluate their individual contributions by isolating each component's impact. Fig. 8(a-b) presents results for defending against RMA and GMA (two hybrid attacks), with each cluster of three bars representing an ablation case. As expected, omitting any component impairs performance, underscoring the complementary and essential roles of each module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Effect of Thresholds \\(\\gamma\\). TRACE maps an input sample \\(x\\) to a linearly separable space, defenders classify it based on a threshold \\(\\gamma\\). In our evaluation, we set an empirical threshold for TRACE. Even in the worst case where no information is available to estimate an appropriate threshold, by setting \\(\\gamma = 0\\), TRACE can still get an competitive average F1 score, still surpassing Detector Cleanse's performance at optimal thresholds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.827 + ], + "angle": 0, + "content": "Effect of Hyperparameters \\( b, f, \\tau \\). As shown in Fig. 9(a-b), performance improves with increasing \\( b \\) and \\( f \\), albeit with higher time overhead. A balance is reached at \\( b = 30 \\) and \\( f = 50 \\), achieving strong performance with reasonable time efficiency. Fig. 9(c) shows the impact of different \\( \\tau \\) when distinguishing between NBOs and triggers. We select the \\( \\tau = 0.1 \\) as the optimal balance point." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.837, + 0.439, + 0.853 + ], + "angle": 0, + "content": "6.3. Resistance to Potential Adaptive Attacks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.858, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We further challenge stronger adversaries in the worst-case scenario, assuming full knowledge of TRACE. Recall the dual targets of backdoored models in Sec. 3.1," + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.089, + 0.709, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.205, + 0.688, + 0.216 + ], + "angle": 0, + "content": "(a) OGA backdoor attack" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.09, + 0.883, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.205, + 0.862, + 0.215 + ], + "angle": 0, + "content": "(b) ODA backdoor attack" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.217, + 0.906, + 0.267 + ], + "angle": 0, + "content": "Figure 10. (Adaptive attack) Visualization of clean mAP (mean Average Precision), backdoor loss, attack success rate, and adaptive loss in backdoor training. Note the negative correlation: as backdoor loss decreases, adaptive loss rises correspondingly." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.906, + 0.35 + ], + "angle": 0, + "content": "where the loss for training such models is \\(\\mathcal{I}_{bd} = \\sum_{i=1}^{|\\mathcal{D}_b|} \\mathcal{L}(\\mathcal{F}(\\boldsymbol{x}_i), y_i) + \\sum_{j=1}^{|\\mathcal{D}_p|} \\mathcal{L}(\\mathcal{F}(\\boldsymbol{x}_j), y_t)\\), where \\(\\mathcal{L}(\\cdot)\\) is the cross entropy loss. We develop an adaptive loss to bypass it:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.369, + 0.901, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {J} _ {a d a p} = \\sum_ {i = 1} ^ {N} \\sum_ {\\delta_ {j}} ^ {\\mathcal {U} (B)} M S E \\left(\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i}\\right), \\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i} \\oplus t\\right)\\right) + \\\\ \\sum_ {i = 1} ^ {N} M S E \\left(\\Delta_ {\\mathcal {F}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i}, o _ {\\mathrm {n b o}}\\right), \\Delta_ {\\mathcal {F}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i} \\oplus t, o _ {\\mathrm {n b o}}\\right)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.713 + ], + "angle": 0, + "content": "where the two terms respectively force the model to maintain the same contextual and focal transformation consistency for clean and poisoned samples, aligning with the TRACE's inference logic. We monitor the adaptive loss without derivative during training. As backdoor loss decreases, adaptive loss rises (i.e., a negative correlation, see Fig. 10), suggesting that success in the dual-target loss may alter the model's transformation consistency. Appendix E.3 offers a reasonable hypothesis. Subsequently, we integrate \\(\\mathcal{I}_{adap}\\) with the vanilla loss \\(\\mathcal{I}_{bd}\\) to formulate the overall loss as \\(\\mathcal{J} = \\mathcal{J}_{bd} + \\lambda \\mathcal{J}_{adap}\\), where \\(\\lambda\\) is a weighting factor. We then optimize \\(\\mathbb{F}_{\\theta}\\) by minimizing \\(\\mathcal{J}\\). Tab. 3 shows that while adaptive attacks can partially evade TRACE, they significantly sacrifice attack performance under \\(\\mathcal{I}_{adap}\\). This is due to the conflict between \\(\\mathcal{I}_{adap}\\) and \\(\\mathcal{J}_{bd}\\), with its failure once again confirming our sustained robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.736, + 0.903, + 0.753 + ], + "angle": 0, + "content": "7. Conclusions, Limitations, and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We present TRACE, a brand-new TTSD for object detection. TRACE is practical, requiring only black-box model access and no training data. TRACE is universal, demonstrating attack-agnostic stability against diverse attacks. TRACE is effective, with extensive experiments on 63 backdoored models validating its performance. However, TRACE's limitations include increased time overhead due to varied transformations and reliance on auxiliary public data. Thus, efficient zero-shot detection is left for our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24384" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.193 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Program of China (Grant No. 2023YFB4503400), and the National Natural Science Foundation of China (Grant Nos. 62450064, 62322205). Shengshan Hu is the corresponding author." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.208, + 0.188, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Christopher J Anders, Leander Weber, David Neumann, Wojciech Samek, Klaus-Robert Müller, and Sebastian Lapuschkin. Finding and removing clever hans: Using explanation methods to debug and improve deep models. Information Fusion, 77:261-295, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision (ECCV'20), pages 213-229, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.376, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[3] Shih-Han Chan, Yinpeng Dong, Jun Zhu, Xiaolu Zhang, and Jun Zhou. Baddet: Backdoor attacks on object detection. In Proceedings of the European Conference on Computer Vision (ECCV'22), pages 396-412. Springer, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.485, + 0.515 + ], + "angle": 0, + "content": "[4] Kangjie Chen, Xiaoxuan Lou, Guowen Xu, Jiwei Li, and Tianwei Zhang. Clean-image backdoor: Attacking multi-label models with poisoned labels only. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR'22), 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[5] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR'19), pages 5177-5186, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[6] Siyuan Cheng, Guangyu Shen, Guanhong Tao, Kaiyuan Zhang, Zhuo Zhang, Shengwei An, Xiangzhe Xu, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Odscan: Backdoor scanning for object detection models. In Proceedings of the IEEE Symposium on Security and Privacy (SP'24), pages 119-119, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[7] Yize Cheng, Wenbin Hu, and Minhao Cheng. Backdoor attack against object detection with clean annotation. arXiv preprint arXiv:2307.10487, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[8] Bao Gia Doan, Ehsan Abbasnejad, and Damith C Ranasinghe. Februus: Input purification defense against trojan attacks on deep neural network systems. In Proceedings of the 36th Annual Computer Security Applications Conference, pages 897-912, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] Maximilian Dreyer, Reduan Achtibat, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. Revealing hidden context bias in segmentation and object detection through concept-specific explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'23), pages 3828-3838, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.233, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR'19), pages 647-657, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, pages 303-338, 2010. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[12] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th annual computer security applications conference (ACSAC'19), pages 113-125, 2019. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In Proceedings of the International Conference on Learning Representations (ICLR'18), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[14] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[15] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, pages 665-673, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.908, + 0.517 + ], + "angle": 0, + "content": "[16] Jiyang Guan, Jian Liang, and Ran He. Backdoor defense via test-time detecting and repairing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'24), pages 24564-24573, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[17] Junfeng Guo, Yiming Li, Xun Chen, Hanqing Guo, Lichao Sun, and Cong Liu. Scale-up: An efficient black-box input-level backdoor detection via analyzing scaled prediction consistency. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR'23), 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[18] Ehsan Hesamifard, Hassan Takabi, Mehdi Ghasemi, and Rebecca N Wright. Privacy-preserving machine learning as a service. Proceedings on Privacy Enhancing Technologies, 2018(3):123-142, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[19] Linshan Hou, Ruili Feng, Zhongyun Hua, Wei Luo, Leo Yu Zhang, and Yiming Li. Ibid-psc: Input-level backdoor detection via parameter-oriented scaling consistency. In Proceedings of the Forty-first International Conference on Machine Learning (ICML'24), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[20] Yanchun Li, Long Huang, Shujuan Tian, Haolin Liu, and Zhetao Li. Robust generative adaptation network for open-set adversarial defense. IEEE Transactions on Information Forensics and Security, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[21] Junyu Lin, Lei Xu, Yingqi Liu, and Xiangyu Zhang. Composite backdoor attack for deep neural network by mixing existing benign features. In Proceedings of the 2020 ACM SIGSAC Conference on Computer and Communications Security, pages 113-131, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24385" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[23] Ruyang Liu, Hao Liu, Ge Li, Haodi Hou, TingHao Yu, and Tao Yang. Contextual debiasing for visual recognition with causal mechanisms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'22), pages 12755-12765, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[24] Xiaogeng Liu, Minghui Li, Haoyu Wang, Shengshan Hu, Dengpan Ye, Hai Jin, Libing Wu, and Chaowei Xiao. Detecting backdoors during the inference stage based on corruption robustness consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'23), pages 16363-16372, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[25] Chengxiao Luo, Yiming Li, Yong Jiang, and Shu-Tao Xia. Untargeted backdoor attack against object detection. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP'23), pages 1-5. IEEE, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[26] Hua Ma, Yinshan Li, Yansong Gao, Alsharif Abuadbba, Zhi Zhang, Anmin Fu, Hyoungshick Kim, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Dangerous cloaking: Natural trigger based backdoor attacks on object detectors in the physical world. arXiv preprint arXiv:2201.08619, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[27] Wanlun Ma, Derui Wang, Ruoxi Sun, Minhui Xue, Sheng Wen, and Yang Xiang. The \"beatrix\" resurrections: Robust backdoor detection via gram matrices. In Proceedings of the Network and Distributed System Security Symposium (NDSS'23), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[28] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. In Proceedings of the international joint conference on neural networks (IJCNN), pages 1-7, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[29] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR'16), pages 779-788, 2016. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS'15), 28, 2015. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[31] Mauro Ribeiro, Katarina Grolinger, and Miriam AM Capretz. Mlaas: Machine learning as a service. In Proceedings of the IEEE 14th international conference on machine learning and applications (ICMLA'15), pages 896-902. IEEE, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[32] Guangyu Shen, Siyuan Cheng, Guanhong Tao, Kaiyuan Zhang, Yingqi Liu, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Django: Detecting trojans in object detection models via gaussian focus calibration. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS'24), 36, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[33] Di Tang, XiaoFeng Wang, Haixu Tang, and Kehuan Zhang. Demon in the variant: Statistical analysis of dnns for robust backdoor contamination detection. In Proceedings of the 30th USENIX Security Symposium (USENIX Security'21), pages 1541-1558, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[34] Sakshi Udeshi, Shanshan Peng, Gerald Woo, Lionel Loh, Louth Rawshan, and Sudipta Chattopadhyay. Model agnostic defence against backdoor attacks in machine learning. IEEE Transactions on Reliability, pages 880-895, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[35] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR'16), pages 2285-2294, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[36] Jiabei Wang, Yanwei Pang, Jiale Cao, Hanqing Sun, Zhuang Shao, and Xuelong Li. Deep intra-image contrastive learning for weakly supervised one-step person search. Pattern Recognition, 147:110047, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[37] Ningfei Wang, Yunpeng Luo, Takami Sato, Kaidi Xu, and Qi Alfred Chen. Does physical adversarial example really matter to autonomous driving? towards system-level effect of adversarial object evasion attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV'23), pages 4412-4423, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[38] Tong Wu, Tianhao Wang, Vikash Sehwag, Saeed Mahloujifar, and Prateek Mittal. Just rotate it: Deploying backdoor attacks via rotation transformation. In Proceedings of the 15th ACM Workshop on Artificial Intelligence and Security, pages 91-102, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[39] Kai Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. arXiv preprint arXiv:2006.09994, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[40] Tinghao Xie, Xiangyu Qi, Ping He, Yiming Li, Jiachen T Wang, and Prateek Mittal. Badexpert: Extracting backdoor functionality for accurate backdoor input detection. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR'24), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.905, + 0.653 + ], + "angle": 0, + "content": "[41] Vacit Oguz Yazici, Abel Gonzalez-Garcia, Arnau Ramisa, Bartlomiej Twardowski, and Joost van de Weijer. Orderless recurrent models for multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'20), pages 13440-13449, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.656, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[42] Yi Zeng, Won Park, Z Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV'21), pages 16473-16481, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.905, + 0.796 + ], + "angle": 0, + "content": "[43] Hangtao Zhang, Shengshan Hu, Yichen Wang, Leo Yu Zhang, Ziqi Zhou, Xianlong Wang, Yanjun Zhang, and Chao Chen. Detector collapse: Backdooring object detection to catastrophic overload or blindness. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI'24), 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.798, + 0.905, + 0.867 + ], + "angle": 0, + "content": "[44] Jiawei Zhao, Ke Yan, Yifan Zhao, Xiaowei Guo, Feiyue Huang, and Jia Li. Transformer-based dual relation graph for multi-label image recognition. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV'21), pages 163-172, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24386" + } + ] +] \ No newline at end of file diff --git a/2025/Test-Time Backdoor Detection for Object Detection Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_origin.pdf b/2025/Test-Time Backdoor Detection for Object Detection Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2750954fe5d736728f04f3d1085ae3dc33281b95 --- /dev/null +++ b/2025/Test-Time Backdoor Detection for Object Detection Models/2e7a4e4a-a106-4d5b-99fb-be8cb267958c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c07c7025743c3f49f8a1f31536a2ea968a88da3c9c45813160098aa7c01e756 +size 9738969 diff --git a/2025/Test-Time Backdoor Detection for Object Detection Models/full.md b/2025/Test-Time Backdoor Detection for Object Detection Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c663f891a44f26dd958d22de7697559fa922753c --- /dev/null +++ b/2025/Test-Time Backdoor Detection for Object Detection Models/full.md @@ -0,0 +1,360 @@ +# Test-Time Backdoor Detection for Object Detection Models + +Hangtao Zhang1 Yichen Wang1 Shihui Yan1 Chenyu Zhu2 Ziqi Zhou3 Linshan Hou4 Shengshan Hu1 Minghui Li2 Yanjun Zhang5 Leo Yu Zhang6 + +1 School of Cyber Science and Engineering, Huazhong University of Science and Technology +$^{2}$ School of Software Engineering, Huazhong University of Science and Technology +$^{3}$ School of Computer Science and Technology, Huazhong University of Science and Technology +$^{4}$ Harbin Institute of Technology $^{5}$ University of Technology Sydney $^{6}$ Griffith University $\{\mathrm{hangt\_zhang, wangyichen, m202472227}\} \mathrm{@hust.edu.cn}$ + +# Abstract + +Object detection models are vulnerable to backdoor attacks, where attackers poison a small subset of training samples by embedding a predefined trigger to manipulate prediction. Detecting poisoned samples (i.e., those containing triggers) at test time can prevent backdoor activation. However, unlike image classification tasks, the unique characteristics of object detection—particularly its output of numerous objects—pose fresh challenges for backdoor detection. The complex attack effects (e.g., “ghost” object emergence or “vanishing” object) further render current defenses fundamentally inadequate. To this end, we design TRansformation Consistency Evaluation (TRACE), a brand-new method for detecting poisoned samples at test time in object detection. Our journey begins with two intriguing observations: 1) poisoned samples exhibit significantly more consistent detection results than clean ones across varied backgrounds. 2) clean samples show higher detection consistency when introduced to different focal information. Based on these phenomena, TRACE applies foreground and background transformations to each test sample, then assesses transformation consistency by calculating the variance in objects' confidence. TRACE achieves black-box, universal backdoor detection, with extensive experiments showing a $30\%$ improvement in AUROC over state-of-the-art defenses and resistance to adaptive attacks. + +# 1. Introduction + +Object detection (OD), a key computer vision application [2, 29], has shown vulnerability to backdoor attacks [3, 25]. Backdoor-infected models perform normally on clean objects but predict the attacker-chosen target annotations whenever a trigger (e.g., a small patch) is present. Un + +![](images/7e90a11786618e22456db70732aa6d361d10e05338264b92d62591e206dac921.jpg) +Figure 1. Divergent attack effects of OD backdoors. (a-c): backdoor triggered by a fixed pattern. (d): backdoor triggered by the co-occurrence of two natural objects (a person and an umbrella). + +like classification models, where backdoors solely cause incorrect class predictions, OD backdoors can result in more sophisticated attack effects, i.e., a “ghost” object appearing [3, 7] (Fig. 1(a)), an object vanishing [25, 26, 38] (Fig. 1(b)), the misclassification of all objects globally [3] (Fig. 1(c)), and the co-occurrence of two benign features triggering a backdoor [4, 21] (Fig. 1(d)). + +To mitigate backdoor threats in OD, recent studies have made preliminary advances, including trigger reverse engineering [6, 32] and black-box test-time trigger sample detection (TTSD)1 [3]. However, the former relies on white-box settings, requiring access to model weights and training data. By contrast, black-box TTSD detects and prevents poisoned samples during test time without access to model parameters, acting as a final 'firewall'. Blackbox TTSD is highly practical for many real-world applications, such as the prevalent Machine Learning as a Service (MLaaS) [18, 31], and is therefore the focus of our work. + +To the best of our knowledge, Detector Cleansse [3] is the only black-box TTsD method tailored for OD tasks. However, it assumes that defenders have full knowledge of the specific details of the attacks (i.e., the adversary-specified judgment criteria, as noted in Tab. 1). In practice, defend + +
TTSD Method↓Black-BoxPractical (No Training Data)Universal (Attack-Agnostic)Applicable (OD Compatible)
TTBD [16]○●●○
IBD-PSC [19]○○●○
Beatrix [27]○○●○
BaDExpert [40]○○●○
NEO [34]●●○○
Strip [12]●○●●
SCALE-UP [17]●●●●
TeCo [24]●●●●
FreqDetector [42]●○●●
Detector Cleanse [3]●○○●
TRACE (Ours)●●●●
+ +Table 1. Comparison among existing TTSD methods w.r.t black-box, practical, universal, and applicable characteristics. “●” indicates the method meets this condition. + +ers cannot possess such prior knowledge. We also apply the current advanced black-box TTSD methods in classification tasks, such as TeCo [24] and SCALE-UP [17], to OD tasks. These TTSDs observe that poisoned samples tend to exhibit more robustness to pre-processing operations than benign samples. Consequently, they apply pixel-level transformations uniformly to all pixel values in each testing sample and assess its prediction consistency. However, applying such uniform pixel-level transformations fails to differentiate between foreground and background components. This distinction is crucial, as the background constitutes the majority of the image, while the foreground (consists of several objects) occupies only a small portion. Recognizing the inherent semantic differences between foreground and background pixels in OD tasks, an intriguing question arises: Shall a semantic-aware transformation expose poisoned samples more effectively than indiscriminate pixel-level modifications? + +Fortunately, the answer is yes! In this paper, we reveal that poisoned samples' prediction confidence are significantly more consistent than those of benign ones across different background contexts (i.e., a semantic-aware transformation). Perhaps more interestingly, when we transform it in the opposite way—by introducing additional foreground information (i.e., semantic patches)—the detection results of poisoned samples become less consistent compared to clean samples, providing clues for detecting backdoors. We refer to these phenomena as the anomalous transformation consistency of backdoored models (detailed in Sec. 4). + +Motivated by this finding, we propose TRAnsformation Consistency Evaluation (TRACE), a novel TTSD to identify and filter poisoned samples in OD. For each suspicious test sample, TRACE evaluates its detections consistency. The consistency is measured using two values: 1) the contextual transformation consistency value, as variance in detection results across varying backgrounds. 2) the focal transformation consistency value, as variance in detection results across focal changes. Higher contextual and lower focal consistency both indicate a poisoned input. TRACE combines these metrics to assess if the input is backdoored. + +In conclusion, our main contributions are three-fold. (1) We disclose an intriguing phenomenon, i.e., anomalous contextual and focal transformation consistency in poisoned samples, which brings forth a new detection paradigm: semantic-aware transformation consistency evaluation. (2) Based on our findings, we propose TRACE, a brand-new black-box, universal test-time backdoor detection method to detect and filter out poisoned inputs. (3) We conduct extensive experiments on three benchmark datasets, verifying TRACE's effectiveness against seven backdoor attacks and resistance to potential adaptive attacks. + +# 2. Related Work + +Backdoor Defenses for OD are still in their infancy. ODSCAN [6] and Django [32] use trigger reverse engineering but require white-box model access and training data. Detector Cleansse [3] is the only black-box TTSD to the best of our knowledge, which adopts a Strip [12]-like approach by overlaying clean features on objects to observe prediction entropy but assumes full knowledge of the attack. Thus, to counter the diverse attacks in OD, a more practical and universal defense is needed. The detailed background and related work are thoroughly covered in Appendix B. + +TTSD for Classification Tasks are generally hard to apply to OD due to its unique characteristics (w.r.t OD Compatible in Tab. 1). Although some image transformation-based methods like Teco [24], and SCALE-UP [17] can be successfully adapted, their effectiveness diminishes due to shifts in task context. + +Remark I. Arguably, we conclude the widespread failure of existing TTSDs to two key challenges: ① The detector's dual-branch architecture (regression and classification) fundamentally differs from classification models, enabling more sophisticated attacks. ② OD requires identifying "what" and "where" an object is, increasing output density. This feature makes detecting triggers across a vast output space nearly impossible without deep insights. + +# 3. Preliminaries + +Object Detection. An object detector $\mathbb{F}_{\theta}$ (parameterized by $\pmb{\theta}$ ) processes each input sample $\pmb{x} \in \mathcal{X}$ containing $z$ objects, with ground-truth annotations $\hat{\pmb{a}} = (\hat{o}_i, \hat{y}_i)_{i=1}^z$ represented as a list of bounding boxes (abbreviated as "bbox" for clarity). Here, $\hat{o}_i$ denotes the $i$ -th object in the image, and $\hat{y}_i$ is its corresponding class. The detector $\mathbb{F}_{\theta}$ outputs $P$ 5-tuples annotation $\pmb{a} = \mathbb{F}_{\theta}(\pmb{x}) = \{\pmb{o}_p^{x_1}, \pmb{o}_p^{y_1}, \pmb{o}_p^{x_2}, \pmb{o}_p^{y_2}, \pmb{y}_p\}_{p=1}^P$ , with $(x_1, y_1)$ and $(x_2, y_2)$ as the coordinates, and $\pmb{y}_p$ denoting the class probability of the $p$ -th predictedbbox. A true-positive (TP) is a correctly detectedbbox in an input sample, while a false-positive (FP) refers to any non-TP detectedbbox, and a false-negative (FN) is any undetected ground-truthbbox. + +# 3.1. OD Backdoor Attacks Formulation + +Assume the $l$ -th object $\hat{o}_l$ to be the victim object (i.e., an object whose recognition is compromised by triggers). The backdoorsed model, while preserving utility, outputs anchors perfectly matching the ground truth $\hat{a}$ , i.e., $\mathbb{F}_{\theta}(\boldsymbol{x}) = \{(o_i,y_i)\}$ , $\forall i \in [1,z]$ , $o_i = \hat{o}_i$ , $y_i = \hat{y}_i$ . Next, we adopt the backdoor categorization introduced by Cheng et al. [6]. + +$\bullet$ FP-inducing attack manipulates the model to recognize a background area as a target class $\pmb{y}_t$ . The appearing object may be localized at a fixed position, the trigger position $\pmb{o}_t$ , or the victim object location $\hat{o}_l$ . We denote the appearing region as $M(\pmb{o}_t,\hat{\pmb{o}}_l)$ (a function of both $\pmb{o}_t$ and $\hat{o}_l$ ). Given a trigger $\pmb{t}$ , $\pmb{x} \oplus \pmb{t} = (1 - \alpha) \otimes \pmb{x} + \alpha \otimes \pmb{t}$ denotes applying the trigger to the clean image with transparency $\alpha$ . As per Cheng et al. [6], we define the FP-inducing attack as $(M(\pmb{o}_t,\hat{\pmb{o}}_l),\pmb{y}_t) \in \mathbb{F}_{\theta}(\pmb{x} \oplus \pmb{t})$ . + +2FN-inducing attack is also known as evasion attack [37], causes models to overlook objects of interest (i.e., generating FNs). Given the victim object $\hat{\pmb{o}}_l$ and its class $\hat{\pmb{y}}_l$ , we define the FN-inducing attack as $(\hat{\pmb{o}}_l,\hat{\pmb{y}}_l)\notin \mathbb{F}_{\theta}(\pmb {x}\oplus \pmb {t})$ + +General Formulation of Backdoor Attacks. We categorize OD backdoors into $FP$ -inducing, $FN$ -inducing attacks, and combinations of both (hybrid attack). For instance, object misclassification attack in Fig. 1(c) (an example of hybrid attack) involve recognizing an object $o_j$ as a different target $o_k$ . This process can be decomposed into the disappearance of the original object $o_j$ (i.e., $FN$ -inducing attack) and the appearance of a new object $o_k$ (i.e., $FP$ -inducing attack). We give a general definition of backdoors as follows: + +$$ +\left(\hat {\boldsymbol {o}} _ {l}, \hat {\boldsymbol {y}} _ {l}\right) \notin \mathbb {F} _ {\theta} (\boldsymbol {x} \oplus \boldsymbol {t}), \tag {1} +$$ + +$$ +\left(M \left(\boldsymbol {o} _ {t}, \hat {\boldsymbol {o}} _ {l}\right), \boldsymbol {y} _ {t}\right) \in \mathbb {F} (\boldsymbol {x} \oplus \boldsymbol {t}). \tag {2} +$$ + +Eq. 1 means the absence of the original victim object from the output anchor set. While Eq. 2 means the presence of the target object, with its shape and location determined by the function $M(\cdot)$ . These equations generalize previous attacks. Formally, the misclassification attack is a special case where $M(\boldsymbol{o}_t, \hat{\boldsymbol{o}}_l) = \hat{\boldsymbol{o}}_l$ . While in object appearing attack, where the trigger is recognized as the target object, $M(\boldsymbol{o}_t, \hat{\boldsymbol{o}}_l) = \boldsymbol{o}_t$ . Details on representing existing attacks based on this definition are in Appendix E.1. + +Remark II: We believe that a general definition is crucial for establishing a universal TTSD method. Otherwise, independent discriminators or criteria must be constructed for each attack and applied sequentially to suspicious inputs. + +# 3.2. Threat Model + +Defender's Goals. TTSD methods operate on a trained backdoored model $\mathbb{F}_{\theta}$ and a test dataset comprising both poisoned samples and clean samples $S = \{\tilde{S} \cup S\}$ . The purpose is to discover a method $\mathcal{T}$ that can maximize the + +![](images/050e044f8b722fc40f4e8c3a470a735b25426e6c39a88606f846de4dfa15abd2.jpg) +Figure 2. Effect of background shift on (a) ResNet-50 and (b)YOLO [29], Faster-RCNN [30], DETR [2]. (b): each background case shows the max, min, and average confidence (as line plots) across three detectors. Larger shaded areas indicate greater instability in object predictions. + +separation of trigger inputs from clean ones, formalized as: + +$$ +\mathcal {T} = \arg \max _ {\mathcal {T}} \mathbb {E} _ {(\tilde {x}) \sim \mathcal {P} _ {\tilde {S}}} \mathbb {I} (\mathcal {T} (\tilde {x}, \mathbb {F} _ {\theta}) = 1) + \tag {3} +$$ + +$$ +\mathbb {E} _ {(x) \sim \mathcal {P} _ {S}} \mathbb {I} \left(\mathcal {T} (x, \mathbb {F} _ {\theta}) = 0\right), +$$ + +where $\mathbb{I}$ is an indicator function, equal to 1 if the condition is true, and 0 otherwise. + +Defender's Capabilities. We adopt standard TTSD settings [12, 24, 34], where the defender uses third-party models and detects poisoned samples on the fly. In particular, we assume a realistic defender with no prior knowledge of attacks and only black-box access to the backdoored model (i.e., can only obtain the final decision from the models). The defender has no training dataset but can access publicly available background and foreground images as support. + +# 4. Insight behind TRACE + +Contextual bias: recognition leans on the co-occurrence context more than on the object itself. Recent efforts have moved from RNNs [35, 41] to GCNNs [5, 10] and transformer-based frameworks [44] to model contextual relationships in multi-label images and improve performance. However, it is also a double-edged sword. Fig. 2(a) shows an example from the Background Challenge [39] reveals how ResNet-50 misclassifies a bird as a fish due to the underwater background. Note that the object detector's learned representations also tend to exploit such spurious scene correlations [23], which can mislead detectors by either giving nonexistent objects high scores in the scene where they typically appear or ignoring objects appearing in rare backgrounds. In other words, when networks find that context alone suffices to identify most objects, they often neglect the representation of the instance, because neural networks are statistics-based and "lazy" [13]. Deep analysis is moved to Appendix E.2. + +Interestingly, we find that trigger objects stay confident no matter the scene. Fig. 2(b) shows three detector's prediction confidences for different objects (extracted using + +![](images/d62d87ce6d898ed876759930cd0d79946de147d09984b87933cf8c805e2a67f6.jpg) +Figure 3. YOLO activation map: clean objects vs. FN-inducing triggers. + +semantic segmentation) placed on 10 varied backgrounds, including an $FP$ -inducing trigger (Fig. 1(a)) with the target annotations "person". We observe that the trigger maintains consistent confidence scores across all backgrounds, while clean objects show notable fluctuations. We attribute this phenomenon to the widely accepted concept of "shortcut learning" [14], where the backdoor training process establishes a robust mapping between a specific, uniform pattern and the target label. This makes the recognition of trigger objects less susceptible to contextual biases, leaving us clues for detecting $FP$ -inducing attacks. Stay tuned for Sec 5.1 to see how this insight is leveraged. + +$\Theta$ Regression continuity: backdoor triggers propagate influence spatially, affecting neighboring pixels. OD differs from classification tasks by incorporating regression branches that rely on shared spatial features. We find that this continuity also allows triggers to impact surrounding regions, creating a black hole-like effect. This is particularly evident for $FN$ -inducing attacks, where the model interprets the affected area as background, ignoring objects of interest [7, 43]. For instance, in Faster R-CNN, where the Region Proposal Network generates around 17, 100 candidate anchor boxes, a trigger affects many nearby anchors, thereby creating the apparent spatial influence. We further diagnose model predictions with and without the trigger using EigenCAM [28] (see Fig. 3), revealing that activations around the trigger diminish, shifting the model's attention to distant regions. Such sharp, regional changes in the saliency map provide clues for detecting $FN$ -inducing attacks. + +Remark III: These insights reveal from different perspectives the key differences between poisoned and clean samples. Additionally, under the introduced general backdoor formulation (Sec. 3.1), the detection of both foundational attack primitives (FN and FP-inducing) are elegantly covered, paving the way for truly black-box, universal detection without the need for training data or attack knowledge. + +# 5. TRACE: A Complete Illustration + +Here we present TRACE, composed of contextual information transformation (Sec. 5.1), focal information transfor + +![](images/4322bb447501980d3985f75e5541a963c245f59dd8289d4c4ea73aa3e00bc03c.jpg) +Figure 4. Background blending and YOLO confidence distribution. + +mation (Sec. 5.2), and test-time evaluation (Sec. 5.3). + +# 5.1. Contextual Information Transformation + +Sec. 4 has explored how DNNs can pick up biases or unintended shortcuts from context (a.k.a. 'Clever Hans' effect [1]). This is intuitive, as cars are typically found on roads, not in locations like the sky or volcanic craters. In contrast, trigger objects exhibit anomalously stable confidence scores. This is further formalized as Contextual Transformation Consistency in Definition 1. + +Definition 1 (Contextual Transformation Consistency (CTC)) Given an object $\mathbf{o}$ and a background distribution $\mathcal{B}$ , the CTC of $\mathbf{o}$ is defined as the maximum deviation in the model's prediction score for $\mathbf{o}$ 's class $\mathbf{y}$ when applied to different backgrounds $\delta \sim \mathcal{B}$ (with the operation of $\mathbf{o} \oplus \delta$ ): + +$$ +\Delta_ {\mathcal {B}} \left(\mathbb {F} _ {\theta}, \boldsymbol {o}\right) := \max _ {\delta_ {1}, \delta_ {2} \sim \mathcal {B}} \left\{\left| \mathbb {F} _ {\theta} \left(\boldsymbol {o} \oplus \delta_ {1}\right) _ {\boldsymbol {y} (\boldsymbol {o})} - \mathbb {F} _ {\theta} \left(\boldsymbol {o} \oplus \delta_ {2}\right) _ {\boldsymbol {y} (\boldsymbol {o})} \right| \right\}. +$$ + +It is observed that $\Delta_{\mathcal{B}}(\mathbb{F},\boldsymbol{o}_t)\ll \Delta_{\mathcal{B}}(\mathbb{F},\boldsymbol{o}_c)$ , indicating that the CTC value is significantly smaller for a trigger object $\boldsymbol{o}_t$ than for a clean object $\boldsymbol{o}_c$ . With this in mind, we propose leveraging this discrepancy to detect the presence of triggers. However, sequentially extracting individual objects using segmentation algorithms (like in Fig. 2 (a)) and overlaying them onto different backgrounds before querying with the backdoored model is computationally intensive. We propose directly blending background images onto the original images with an opacity $\alpha_{bg}$ (i.e., $(1 - \alpha_{bg})\otimes x + \alpha_{bg}\otimes \delta$ , see Fig. 4, left). This allows for background replacement at the image level rather than the object level. By using a lower $\alpha_{bg}$ to minimize alterations to the object's pixel values, we thus iteratively introduce diverse contextual information (from a background set of size $b$ available online, e.g., the Background Challenge dataset [39]) while mitigating the impact of backgrounds on the object's representations. Visualizations are provided in Appendix C.4. + +Instead of directly using the CTC value, we consider potential corner cases where certain objects might unexpectedly become undetectable. To account for these cases, we use the variance of the object's confidence scores across different background variations (sampling uniformly from $\mathcal{B}$ ) + +to enhance robustness, denoted as $\Delta_B^{\mathrm{Var}}(\mathbb{F}_\theta ,\pmb {o})$ , given by + +$$ +\Delta_ {\mathcal {B}} ^ {\operatorname {V a r}} \left(\mathbb {F} _ {\theta}, \boldsymbol {o} _ {i}\right) = V a r \left(\left\{\mathbb {F} _ {\theta} (\boldsymbol {o} _ {i} \oplus \delta) _ {\boldsymbol {y} (\boldsymbol {o})} \right\} _ {i = 1} ^ {b}\right), \delta \sim \mathcal {U} (\mathcal {B}). +$$ + +Fig. 4 shows the violin plots of confidence distributions for different objects blended with 50 backgrounds at $\alpha_{bg} = 0.15$ , including an FP-inducing trigger (see Fig. 1(a)) with the target annotations "person". Empirically, as expected, the trigger exhibits a low variance of just 0.008, achieving clear separability in distribution. Explorations for different objects and detectors are provided in Appendix D.1. + +Natural backdoor: neural networks also overfit to uniform, salient benign representations. Detecting FP-inducing triggers, however, does not stop there. Note that certain objects with standardized shapes and colors—such as uniformly red, polygonal stop signs—are also overconfidently recognized by networks, much like backdoor triggers, as DNNs more readily memorize these easily learned, superficial features [15]. We consider these a unique form of benign triggers, which we term "natural backdoor objects" (NBOs). Such uniform, salient representations allow them to rely less on additional context for recognition, showing anomalous CTC values similar to triggers. Consequently, we integrate a post-processing module to filter and remove them. Here we introduce a visual coherence analysis to screen suspicious objects exhibiting low $\Delta_{\mathcal{B}}^{\mathrm{Var}}$ . By comparing these objects with reference images corresponding to their categories from online available data, we evaluate their SSIM scores [20]. Note that these references, which we term "universal visual benchmarks" (see Appendix C.5), do not need to be sourced from the model's training data, as they represent general visual cognition. Since an image may contain multiple triggering objects, we naturally define the image-level CTC as the minimum CTC values among all (SSIM-filtered) objects in the image, i.e., + +$$ +\begin{array}{l} \underbrace {\Delta_ {\mathcal {B}} ^ {\operatorname {V a r}} \left(\mathbb {F} _ {\theta} , \boldsymbol {x}\right)} = \min \left\{\underbrace {\Delta_ {\mathcal {B}} ^ {\operatorname {V a r}} \left(\mathbb {F} _ {\theta} , \boldsymbol {o} _ {i}\right)} \right| \\ \begin{array}{c} \text {I m a g e - l e v e l V a r i a n c e} \\ \hline \end{array} \\ s. t. \quad S S I M \left(\boldsymbol {o} _ {i}, \boldsymbol {o} _ {y _ {i}} ^ {\text {r e f}}\right) \leq \tau , \quad i = 1, \dots , z \Bigg \}, \\ \end{array} +$$ + +where $\tau$ is a predefined threshold, and $o_{y_i}^{\mathrm{ref}}$ is the universal visual benchmark for class $y_i$ . Clearly, the lower the image-level CTC values, the more likely it is that the image contains the trigger objects. + +# 5.2. Focal Information Transformation + +Why do we still need another transformation? We have demonstrated that CTC values can identify FP-inducing triggers among visible objects in the image. However, FN-inducing triggers (see Fig. 1(b)) remain entirely absent in detection results, yielding no CTC value. How to detect these "unseen objects" thus becomes a non-trivial challenge. Sec. 4 highlights the shifts in saliency maps when comparing samples with and without triggers. However, + +interpretability methods like Grad-CAM typically require white-box model access, which is out of reach for our black-box defender. Hence, instead of passively analyzing the input sample, we seamlessly inject new elements (i.e., small clean object patches) to proactively "reconstruct" saliency maps. We begin by sliding a 'car' object across an image in tiny steps—like how a convolutional kernel scans over pixels—recording its prediction confidence at each position. Unfortunately, this object is fragile, with confidence fluctuating by position (see Fig. 5(a)). A representative visual example is the surfboard in Fig. 5(b), which shows a strong positional dependency, becoming undetectable when shifted even slightly. This inherent fragility causes such objects to fail in probing the model's focus regions. + +Although NBOs may initially seem to hinder our detection process, we argue that this trait is precisely the key to resolving the above issue. Fig. 5(c) shows the sliding results for an NBO ('stop sign'). The model consistently detects it with high confidence, demonstrating position-invariant detection [36]. This makes it particularly suitable for probing the saliency maps. Hence, this time we slide the NBO across the same image but with a FN-inducing trigger placed at its center. Fig. 5(d) shows its high confidence at the image edges, dropping sharply near the central trigger. When it completely covers the trigger, the model's normal detection resumes. We term this distinctive behavior the "Island Effect", as it resembles an island rising from surrounding waters. The NBO undergoes two sharp changes in confidence when handling poisoned samples, while remaining stable on clean images. Leveraging this property, we can thus proactively probe potential FN-inducing triggers in the troublesome black-box setting. We formalize Focal Transformation Consistency in Definition 2. + +Definition 2 (Focal Transformation Consistency (FTC)) Given an image $\pmb{x}$ defined over a spatial domain $\Omega$ and an object patch $\pmb{o}$ , a focal transformation $\mathcal{K} : (\pmb{x}, \pmb{o}, \pmb{p}) \to \pmb{x}^p$ applies $\pmb{o}$ to $\pmb{x}$ , centered at the spatial coordinate $\pmb{p} = (\pmb{p}_x, \pmb{p}_y) \in \Omega$ . The FTC value is then given by: + +$$ +\begin{array}{l} \Delta_ {\mathcal {F}} (\mathbb {F} _ {\boldsymbol {\theta}}, \pmb {x}, \pmb {o}) := \int_ {\Omega} \left(\left\| \nabla^ {2} \mathbb {F} _ {\boldsymbol {\theta}} \big (\mathcal {K} (\pmb {x}, \pmb {o}, \pmb {p}) \big) _ {\pmb {y} (\pmb {o})} \right\| _ {1} + \right. \\ \left. \right.\left\| \right.\left[ \mathbb {F} _ {\theta} \left(\mathcal {K} (\boldsymbol {x}, \boldsymbol {o}, \boldsymbol {p} + \Delta \boldsymbol {p}) - \mathbb {F} _ {\theta} \left(\mathcal {K} (\boldsymbol {x}, \boldsymbol {o}, \boldsymbol {p})\right)\right)\right] _ {+} \| _ {1}\left. \right) d \boldsymbol {p} \\ \end{array} +$$ + +where $\nabla^2$ is the Laplacian operator capturing second-order spatial derivatives, $\mathbb{F}_{\theta}(\mathcal{K}(\pmb{x},\pmb{o},\pmb{p}))_{\pmb{y}(\pmb{o})}$ measures the confidence score of object $\pmb{o}$ after applying the transformation $\mathcal{K}(\cdot)$ , and $[\cdot]_+$ denotes the positive part function (i.e., $[a]_+ = \max(0,a)$ ), ensuring that the quantification of detection shift focuses on the incremental components of the model's outputs, specifically capturing newly emerged detections (i.e., results absent in the original predictions). + +FTC has two components: the confidence oscillation of $\pmb{o}$ and the image-level detection shift caused by $\pmb{o}$ . The first + +![](images/4938c7ff70a9d5338dfdaae5df1e27fa5eddc540a6afefd6dab9abec73315b7f.jpg) +(a) Confidence of "car" sliding across a clean image + +![](images/5dc621ea9ca85bc92844186239b1b1cdaf36e7fb3d1f14ba143c431be4800160.jpg) +(b) "fragile" object + +![](images/b74be21bf681d8783d75fb35c2746904f9d2ebe9b6e7b0d4b0909b787f82839f.jpg) + +![](images/104cc4137fc148aae900d00f43f4541d0a116286d8ff4f2dc41968facd3ed998.jpg) +(c) Confidence of "stop sign" sliding across a clean image +(d) Confidence of "stop sign" sliding across a trigger image + +metric assesses prediction consistency of $o$ across different positions in image $x$ , helping detect the "Island Effect" of FN-inducing triggers, where each cross-section shows two sharp confidence drops. The second captures changes in whole detection results due to $o$ . Under normal circumstances, moving $o$ will not introduce new objects. We use it to detect cases when new objects suddenly appear as $o$ covers the trigger. This trigger-blocking strategy is also supported by established backdoor defenses [8, 34]. + +NBOs demonstrate strong location-invariant recognition, as choosing $\pmb{o} = \pmb{o}_{\mathrm{nbo}}$ decouples $\pmb{o}$ 's contribution to $\Delta_{\mathcal{F}}(\mathbb{F}_{\theta}, \pmb{x}, \pmb{o})$ , ensuring $\partial \Delta_{\mathcal{F}} / \partial \pmb{o} \approx 0$ under spatial variations in $\pmb{p}$ . Thus, with $\pmb{o}$ effectively neutralized, the FTC value simplifies to reflect differences in $\pmb{x}$ , yielding a higher $\Delta_{\mathcal{F}}(\mathbb{F}_{\theta}, \tilde{\pmb{x}}, \pmb{o}_{\mathrm{nbo}}) \gg \Delta_{\mathcal{F}}(\mathbb{F}_{\theta}, \pmb{x}, \pmb{o}_{\mathrm{nbo}})$ for poisoned samples $\tilde{\pmb{x}}$ . Instead of densely querying via iteratively sliding, we seamlessly employ Monte Carlo sampling of $k$ distinct, non-overlapping points to patch $\pmb{o}_{\mathrm{nbo}}$ (see Fig. 6), thereby reducing the number of required model queries and improving computational efficiency. This unit operation is then repeated $f$ times to build an approximate continuous saliency map. Finally, we can compute the variance of $\Delta_{\mathcal{F}}(\mathbb{F}_{\theta}, \pmb{x}, \pmb{o}_{\mathrm{nbo}})$ across all accumulated sampling points from $\mathcal{U}(\Omega)$ to determine the FTC value for the sample $\pmb{x}$ : + +$$ +\Delta_ {\mathcal {F}} ^ {\operatorname {V a r}} \left(\mathbb {F} _ {\theta}, \boldsymbol {x}, \boldsymbol {o} _ {\mathrm {n b o}}\right) := V a r \left(\left\{\Delta_ {\mathcal {F}} \left(\mathbb {F} _ {\theta}, \boldsymbol {x}, \boldsymbol {o} _ {\mathrm {n b o}}\right) \mid \boldsymbol {p} _ {i} \sim \mathcal {U} (\boldsymbol {\Omega}) \right\} _ {i = 1} ^ {f \cdot k}\right). +$$ + +# 5.3. Test-time TRACE Evaluation + +To achieve Eq. 3, the TTSD methods must exploit the disparities between benign images and those containing triggers. Our findings in Sec. 5.1 and 5.2 reveal that backdoored models exhibit anomalous consistency in both contextual and focal transformation. In other words, the distinct properties of CTC and FTC make them particularly suitable for detecting FP-inducing and FN-inducing attacks, respectively. As established in Sec 3.1, a trigger sample can be a single type or a combination of both. Consequently, since a smaller CTC $(\Delta_{\mathcal{B}}^{\mathrm{Var}})$ and a larger FTC value $(\Delta_{\mathcal{F}}^{\mathrm{Var}})$ both indicate higher anomaly, the final step is to measure the overall image-level anomaly as $Trace(\pmb{x}) = \sigma \left( \Delta_{\mathcal{F}}^{\mathrm{Var}}(\mathbb{F}_{\theta}, \pmb{x}, \pmb{o}_{\mathrm{nbo}}) \right) - \sigma \left( \Delta_{\mathcal{B}}^{\mathrm{Var}}(\mathbb{F}_{\theta}, \pmb{x}) \right)$ , where $\sigma(t) = \frac{1}{1 + e^{-t}}$ is the sigmoid function that normalizes both variance + +![](images/3daa3fb39d6119824a2de5531fc9269bebf3d491c588a8e93c3a39d92953d480.jpg) +Figure 5. Effect of the foreground object's position on YOLO detections. (d) uses an FN-inducing trigger placed at the image center. +Figure 6. The framework of our TRACE. + +measures to the range $(0,1)$ . TRACE thus maps the input image $x$ to a linearly separable space, enabling defenders to make decisions by a threshold $\gamma$ : + +$$ +\Gamma (T r a c e (\boldsymbol {x})) = \left\{ \begin{array}{l l} 1, & \text {i f T r a c e} (\boldsymbol {x}) > \gamma \\ 0, & \text {i f T r a c e} (\boldsymbol {x}) \leq \gamma \end{array} \right. \tag {6} +$$ + +# 6. Evaluation + +Datasets and Models. We select popular benchmark datasets, including MS-COCO [22], PASCAL VOC [11], and Synthesized Traffic Signs [32]. The models we choose are three representative object detectors: single-stage detector: YOLOv5 [29], two-stage detector: Faster-RCNN [30] and vision transformer-based detector: DETR [2]. + +Attack Methods. We evaluate TRACE against seven SOTA OD backdoors: OGA, RMA, GMA, and ODA from [3], as well as CIB [4], UTA [25], and DC [43]. For CIB and DC, which have multiple goals, we focus on their object disappearance variant as a representative case. A detailed description of these attacks can be found in Appendix C.6. + +Defense Settings. We compare TRACE with Detector Cleansse [3], the only current black-box TTSD tailored + +Table 2. F1 score and AUROC for different defenses on Synthesized Traffic Signs, PASCAL VOC, and MS-COCO, using YOLO, DETR, and Faster-RCNN. Due to space constraints, some results for two attacks are presented in the same row, separated by “/”. “-” denotes that FreqDetector is unavailable in this case, as it detects frequency-domain anomalies in trigger patterns, while CIB uses benign features to trigger the backdoors. We bold the F1 and AUROC of the strongest defenses in each case. + +
Model↓Method AttackTRACE (Ours)Teco [24]Detector cleansse [3]Strip [12]FreqDetector [42]SCALE-UP [17]
F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑
MS-COCO [22]
YOLOOGA/ODA0.937/0.8650.913/0.8800.464/0.5210.593/0.5920.702/0.6830.650/0.6920.646/0.4830.729/0.5030.542/0.5880.490/0.5800.503/0.4890.505/0.463
RMA/GMA0.922/0.8450.932/0.9240.397/0.5010.487/0.4580.674/0.6620.590/0.6080.391/0.3590.474/0.5250.567/0.5120.498/0.4380.482/0.4730.510/0.440
Untargeted0.8370.8680.4980.5500.6890.6690.4850.4720.6140.6460.5030.491
CIB0.7980.8170.6510.6910.6100.6620.5320.578--0.2610.367
DC0.9340.9680.5460.6750.6400.7300.5250.4930.5830.6640.4700.506
Average0.8770.9000.5110.5780.6650.6570.4890.5390.5680.5530.4540.460
MS-COCO [22]
YOLOOGA/ODA0.801/0.8390.825/0.8570.593/0.5800.603/0.5770.676/0.6370.541/0.6790.646/0.5300.649/0.5360.542/0.5880.490/0.5800.481/0.5240.478/0.549
RMA/GMA0.807/0.8160.777/0.8460.534/0.5060.478/0.4380.686/0.6490.629/0.5880.498/0.5920.452/0.5800.567/0.5150.498/0.4380.497/0.4620.486/0.456
Untargeted0.7880.7900.5460.5770.6980.6550.5070.4860.6140.6460.5180.528
CIB0.7200.7560.6310.6760.5830.6320.5220.548--0.2350.280
DC0.8380.9140.5760.6350.7030.7210.5040.4830.5830.6640.4800.518
Average0.8010.8240.5670.5690.6620.6350.5430.5330.5680.5530.4570.471
Faster-RCNNOGA/ODA0.757/0.8540.770/0.8560.601/0.6100.601/0.6240.716/0.6700.591/0.6060.631/0.4970.636/0.5460.542/0.5880.490/0.5800.498/0.4810.482/0.469
RMA/GMA0.872/0.8700.899/0.8830.563/0.5190.535/0.4910.721/0.6590.655/0.5810.471/0.5300.471/0.5330.567/0.5120.498/0.4390.495/0.4910.491/0.503
Untargeted0.8690.8420.5770.5710.6700.6550.5590.5640.6140.6460.4940.471
CIB0.7470.7500.6380.6820.6130.6020.5850.603--0.2620.329
DC0.9350.9440.5590.6670.6270.7450.4750.5280.5830.6640.4770.454
Average0.8430.8490.5810.5970.6680.6330.5350.5540.5680.5530.4570.457
PASCAL VOC [11]
YOLOOGA/ODA0.928/0.9010.945/0.9050.536/0.5430.561/0.5710.670/0.7210.631/0.6520.612/0.5110.664/0.4450.500/0.6010.485/0.6370.441/0.5220.458/0.505
RMA/GMA0.916/0.8730.908/0.9390.445/0.5140.444/0.4790.619/0.6630.585/0.5550.563/0.4470.626/0.5130.562/0.4770.584/0.5020.515/0.5080.495/0.489
Untargeted0.8000.8580.4870.4970.7060.7010.5510.5030.6440.6870.4820.490
CIB0.7850.7670.6100.6560.6180.6240.5670.598--0.2730.332
DC0.9360.9440.5020.6550.6830.7920.5090.5050.5630.6140.5340.559
Average0.8770.8950.5200.5520.6690.6490.5370.5510.5580.5850.4680.475
DETROGA/ODA0.775/0.8230.799/0.8310.524/0.5150.547/0.5200.703/0.6080.627/0.6370.610/0.5190.660/0.4670.500/0.6010.485/0.6370.483/0.5050.497/0.521
RMA/GMA0.803/0.7660.786/0.7740.466/0.4880.450/0.4470.651/0.6220.563/0.5500.559/0.5330.584/0.5620.562/0.4770.584/0.5020.541/0.4770.559/0.483
Untargeted0.8310.8200.5670.5720.6550.7050.6470.6920.6440.6870.4890.476
CIB0.7500.7670.6320.6830.5910.5870.5750.569--0.2750.346
DC0.9600.9460.5590.6260.6360.7240.5600.4720.5630.6140.4970.502
Average0.8150.8180.5360.5490.6380.6280.5720.5720.5580.5850.4670.483
Synthesized Traffic Signs [32]
YOLOOGA/ODA0.917/0.8850.942/0.8720.581/0.5910.581/0.5600.693/0.6400.665/0.6640.671/0.4420.645/0.4670.504/0.6060.529/0.5610.473/0.4530.480/0.446
RMA/GMA0.957/0.9280.944/0.9360.515/0.4850.459/0.4900.678/0.6900.605/0.5490.494/0.5110.523/0.5790.562/0.5220.506/0.4240.505/0.4810.517/0.472
Untargeted0.8570.9030.5500.5010.7030.6450.4670.4410.6390.6130.5170.494
CIB0.8250.8530.6280.6490.6550.6730.5480.591--0.2360.354
DC0.9260.9410.5570.6890.7010.7540.5180.5230.5900.5430.5110.532
Average0.8990.9140.5580.5610.6800.6510.5220.5380.5710.5290.4540.471
+ +for OD. We also include four black-box defenses applicable to OD: Strip [12], FreqDetector [42], TeCo [24], and SCALE-UP [17]. We follow the configurations recommended in their papers for a fair comparison. + +Evaluation Metrics. We evaluate using two metrics: F1 score (balancing precision and recall) and AUROC (measuring overall performance across different thresholds). + +# 6.1. Defense Performance + +Effectiveness Study. In Tab. 2, TRACE consistently achieves promising performance across seven backdoor attacks, three datasets, three model architectures, and a total of 42 backdoors models. To further assess TRACE's stability, we calculate performance standard deviations across different backdoors on the same model and dataset, finding low standard variances (F1: $0.880 \pm 0.040$ , AUROC: $0.897 \pm 0.041$ ), which demonstrates its robust consistency. + +Comparison Study. Tab. 2 also shows that TRACE comprehensively outperformed its five black-box competitors: TeCo, Strip, FreqDetector, SCALE-UP, and the SOTA Detector Cleansed designed for OD task. The first four fail completely due to shifts in task context. Even though Detector Cleansed is not attack-agnostic, TRACE still significantly surpasses it in F1 score by $\sim 30\%$ . Further analysis is deferred to Appendix D.2. + +(a) clean input (a) heatmap (b) clean input (b) heatmap (c) trigger input (c) heatmap + +![](images/54568e75028dd59a9172c451837593f3c14b0b1584f6f96dc52c23d17d3f3117.jpg) + +![](images/f569a6c08d328fbe8b5404bc5666d070fbff92f12701ef2ad7840a6354d38a70.jpg) + +![](images/2e32c8015c1bbf137b249bcff84dec4d68f3425ea71c532f199f1979e04ea29d.jpg) + +![](images/206b4829ef8f151b90cfe6aea48f6197259bc7a0457c53e7afb74f72a12e89dd.jpg) + +![](images/397454c55cf39f8fd3fcaf15690f7e3326dcf4878b31dcb09a53a19d2f89e492.jpg) + +![](images/d5a9f1d7a2afb5326582ac2938a35ec2cd0b463081865149ba424a4d928b0c4c.jpg) + +![](images/2ef08ee54d38b5da0ceb5e61327199f35fa28242f0eb1868ecb4257f19f42f90.jpg) +CTC trans. input + +![](images/e2c0181f6fdd730840a08ad58ad00e5b7d34ec28fb3a2c49a6ab572743d7520e.jpg) +CTC trans. input + +![](images/24239c77f0c607df60da94954d25a08f7a439fcc1f3e14f5bedca4510fe8a497.jpg) +Figure 7. High resolution, object-specific explanations from L-CRP [9]. + +![](images/9f4b92810df0c9c015a75f1590adf8adba70f6b9a0c3b54da139b9feb11dd3f0.jpg) + +![](images/2a09364ec8c71ca2a2e9bba32787783d5eefdaec0813089860ab670bbcde198a.jpg) +CTC trans. input + +![](images/6a3ee0ceef7fdc05bc5b0bd22c35ba2bba49f1a54ad5219a0bcf3153e510760c.jpg) + +A Closer Look to the Effectiveness of TRACE. We use SOTA eXplainable Artificial Intelligence (XAI) method L-CRP [9] to attain object-class-specific explanations in this multi-object context. We offer three illustrative examples: Fig. 7(a) shows that the detection of a banana relies heavily on surrounding green plants. TRACE's background transformation (i.e., introducing a contrasting fire scene) breaks this context bias, shifting reliance to the banana's intrinsic features. In Fig. 7(b), the bird's detection initially depends on its own features, but after transformation, new contextual dependencies emerge. In both cases, confidence scores shift (decrease for the banana, increasing for the bird). Lastly, Fig. 7(c) shows an FP-inducing trigger, showing always input-independent detection. + +![](images/61fcaa91d3bb38a534c79ef5cfa1690a51a207b3a909cb391b9181f5d2857db0.jpg) +(a) Defending RMA (hybrid attack) + +![](images/74b32f270f41172a9b18916046dbddc359f7f75e5cbbdbf1de98ae0b43816423.jpg) +(b) Defending GMA (hybrid attack) + +![](images/7f58b2c3709877d8b6cd3f0667da443efb4ea9a4c939fad456f7be9f99f0a70d.jpg) +Figure 8. (Ablation Study) Effect of different modules. +(a) Background queries +Figure 9. (Ablation Study) Effect of background, foreground, and $\tau$ . + +![](images/cc8a404c8b52be57d98a05b770e543b062db1d2eadb31ee554f1a69df8c03f06.jpg) +(b) Foreground queries + +![](images/3c9915be873f9b235cb26e8e57531500bf48cc9c1466d3729cf84efacd59f443.jpg) +(c) $\tau$ value + +Table 3. (Adaptive attack) TRACE against the omniscient adaptive attack. + +
λ→Attacks↓5 × 10-41 × 10-45 × 10-51 × 10-5
F1AUROCASRF1AUROCASRF1AUROCASRF1AUROCASR
OGA0.5810.6200.7330.7660.7470.8430.9160.8790.9440.9370.9420.975
ODA0.6060.5940.5830.6080.5760.7520.8200.7940.8410.8400.8510.893
RMA0.6170.6140.6810.7480.7020.8100.8670.8860.9330.9380.9290.947
+ +# 6.2. Ablation Study + +Effect of Modules. TRACE incorporates three fundamental designs: contextual transformation, focal transformation, and SSIM-filter post-processing. We evaluate their individual contributions by isolating each component's impact. Fig. 8(a-b) presents results for defending against RMA and GMA (two hybrid attacks), with each cluster of three bars representing an ablation case. As expected, omitting any component impairs performance, underscoring the complementary and essential roles of each module. + +Effect of Thresholds $\gamma$ . TRACE maps an input sample $x$ to a linearly separable space, defenders classify it based on a threshold $\gamma$ . In our evaluation, we set an empirical threshold for TRACE. Even in the worst case where no information is available to estimate an appropriate threshold, by setting $\gamma = 0$ , TRACE can still get an competitive average F1 score, still surpassing Detector Cleanse's performance at optimal thresholds. + +Effect of Hyperparameters $b, f, \tau$ . As shown in Fig. 9(a-b), performance improves with increasing $b$ and $f$ , albeit with higher time overhead. A balance is reached at $b = 30$ and $f = 50$ , achieving strong performance with reasonable time efficiency. Fig. 9(c) shows the impact of different $\tau$ when distinguishing between NBOs and triggers. We select the $\tau = 0.1$ as the optimal balance point. + +# 6.3. Resistance to Potential Adaptive Attacks + +We further challenge stronger adversaries in the worst-case scenario, assuming full knowledge of TRACE. Recall the dual targets of backdoored models in Sec. 3.1, + +![](images/84bc4d0b200ca4ab2a06f7e8dfce40e0a96135d89c97f778f2c81cd744f0dd70.jpg) +(a) OGA backdoor attack +Figure 10. (Adaptive attack) Visualization of clean mAP (mean Average Precision), backdoor loss, attack success rate, and adaptive loss in backdoor training. Note the negative correlation: as backdoor loss decreases, adaptive loss rises correspondingly. + +![](images/684c36cf6f57b72dbd805ed1d97c208528137112a19c84282d2583d5bd44c1bd.jpg) +(b) ODA backdoor attack + +where the loss for training such models is $\mathcal{I}_{bd} = \sum_{i=1}^{|\mathcal{D}_b|} \mathcal{L}(\mathcal{F}(\boldsymbol{x}_i), y_i) + \sum_{j=1}^{|\mathcal{D}_p|} \mathcal{L}(\mathcal{F}(\boldsymbol{x}_j), y_t)$ , where $\mathcal{L}(\cdot)$ is the cross entropy loss. We develop an adaptive loss to bypass it: + +$$ +\begin{array}{l} \mathcal {J} _ {a d a p} = \sum_ {i = 1} ^ {N} \sum_ {\delta_ {j}} ^ {\mathcal {U} (B)} M S E \left(\Delta_ {\mathcal {B}} ^ {\operatorname {V a r}} \left(\mathbb {F} _ {\theta}, \boldsymbol {x} _ {i}\right), \Delta_ {\mathcal {B}} ^ {\operatorname {V a r}} \left(\mathbb {F} _ {\theta}, \boldsymbol {x} _ {i} \oplus t\right)\right) + \\ \sum_ {i = 1} ^ {N} M S E \left(\Delta_ {\mathcal {F}} ^ {\operatorname {V a r}} \left(\mathbb {F} _ {\theta}, \boldsymbol {x} _ {i}, o _ {\mathrm {n b o}}\right), \Delta_ {\mathcal {F}} ^ {\operatorname {V a r}} \left(\mathbb {F} _ {\theta}, \boldsymbol {x} _ {i} \oplus t, o _ {\mathrm {n b o}}\right)\right) \\ \end{array} +$$ + +where the two terms respectively force the model to maintain the same contextual and focal transformation consistency for clean and poisoned samples, aligning with the TRACE's inference logic. We monitor the adaptive loss without derivative during training. As backdoor loss decreases, adaptive loss rises (i.e., a negative correlation, see Fig. 10), suggesting that success in the dual-target loss may alter the model's transformation consistency. Appendix E.3 offers a reasonable hypothesis. Subsequently, we integrate $\mathcal{I}_{adap}$ with the vanilla loss $\mathcal{I}_{bd}$ to formulate the overall loss as $\mathcal{J} = \mathcal{J}_{bd} + \lambda \mathcal{J}_{adap}$ , where $\lambda$ is a weighting factor. We then optimize $\mathbb{F}_{\theta}$ by minimizing $\mathcal{J}$ . Tab. 3 shows that while adaptive attacks can partially evade TRACE, they significantly sacrifice attack performance under $\mathcal{I}_{adap}$ . This is due to the conflict between $\mathcal{I}_{adap}$ and $\mathcal{J}_{bd}$ , with its failure once again confirming our sustained robustness. + +# 7. Conclusions, Limitations, and Future Work + +We present TRACE, a brand-new TTSD for object detection. TRACE is practical, requiring only black-box model access and no training data. TRACE is universal, demonstrating attack-agnostic stability against diverse attacks. TRACE is effective, with extensive experiments on 63 backdoored models validating its performance. However, TRACE's limitations include increased time overhead due to varied transformations and reliance on auxiliary public data. Thus, efficient zero-shot detection is left for our future work. + +# Acknowledgement + +This work is supported by the National Key Research and Development Program of China (Grant No. 2023YFB4503400), and the National Natural Science Foundation of China (Grant Nos. 62450064, 62322205). Shengshan Hu is the corresponding author. + +# References + +[1] Christopher J Anders, Leander Weber, David Neumann, Wojciech Samek, Klaus-Robert Müller, and Sebastian Lapuschkin. Finding and removing clever hans: Using explanation methods to debug and improve deep models. Information Fusion, 77:261-295, 2022. 4 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision (ECCV'20), pages 213-229, 2020. 1, 3, 6 +[3] Shih-Han Chan, Yinpeng Dong, Jun Zhu, Xiaolu Zhang, and Jun Zhou. Baddet: Backdoor attacks on object detection. In Proceedings of the European Conference on Computer Vision (ECCV'22), pages 396-412. Springer, 2022. 1, 2, 6, 7 +[4] Kangjie Chen, Xiaoxuan Lou, Guowen Xu, Jiwei Li, and Tianwei Zhang. Clean-image backdoor: Attacking multi-label models with poisoned labels only. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR'22), 2022. 1, 6 +[5] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR'19), pages 5177-5186, 2019. 3 +[6] Siyuan Cheng, Guangyu Shen, Guanhong Tao, Kaiyuan Zhang, Zhuo Zhang, Shengwei An, Xiangzhe Xu, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Odscan: Backdoor scanning for object detection models. In Proceedings of the IEEE Symposium on Security and Privacy (SP'24), pages 119-119, 2024. 1, 2, 3 +[7] Yize Cheng, Wenbin Hu, and Minhao Cheng. Backdoor attack against object detection with clean annotation. arXiv preprint arXiv:2307.10487, 2023. 1, 4 +[8] Bao Gia Doan, Ehsan Abbasnejad, and Damith C Ranasinghe. Februus: Input purification defense against trojan attacks on deep neural network systems. In Proceedings of the 36th Annual Computer Security Applications Conference, pages 897-912, 2020. 6 +[9] Maximilian Dreyer, Reduan Achtibat, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. Revealing hidden context bias in segmentation and object detection through concept-specific explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'23), pages 3828-3838, 2023. 7 +[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial + +labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR'19), pages 647-657, 2019. 3 +[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, pages 303-338, 2010. 6, 7 +[12] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th annual computer security applications conference (ACSAC'19), pages 113-125, 2019. 2, 3, 7 +[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In Proceedings of the International Conference on Learning Representations (ICLR'18), 2018. 3 +[14] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 4 +[15] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, pages 665-673, 2020. 5 +[16] Jiyang Guan, Jian Liang, and Ran He. Backdoor defense via test-time detecting and repairing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'24), pages 24564-24573, 2024. 2 +[17] Junfeng Guo, Yiming Li, Xun Chen, Hanqing Guo, Lichao Sun, and Cong Liu. Scale-up: An efficient black-box input-level backdoor detection via analyzing scaled prediction consistency. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR'23), 2023. 2, 7 +[18] Ehsan Hesamifard, Hassan Takabi, Mehdi Ghasemi, and Rebecca N Wright. Privacy-preserving machine learning as a service. Proceedings on Privacy Enhancing Technologies, 2018(3):123-142, 2018. 1 +[19] Linshan Hou, Ruili Feng, Zhongyun Hua, Wei Luo, Leo Yu Zhang, and Yiming Li. Ibid-psc: Input-level backdoor detection via parameter-oriented scaling consistency. In Proceedings of the Forty-first International Conference on Machine Learning (ICML'24), 2024. 1, 2 +[20] Yanchun Li, Long Huang, Shujuan Tian, Haolin Liu, and Zhetao Li. Robust generative adaptation network for open-set adversarial defense. IEEE Transactions on Information Forensics and Security, 2025. 5 +[21] Junyu Lin, Lei Xu, Yingqi Liu, and Xiangyu Zhang. Composite backdoor attack for deep neural network by mixing existing benign features. In Proceedings of the 2020 ACM SIGSAC Conference on Computer and Communications Security, pages 113-131, 2020. 1 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, + +Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7 +[23] Ruyang Liu, Hao Liu, Ge Li, Haodi Hou, TingHao Yu, and Tao Yang. Contextual debiasing for visual recognition with causal mechanisms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'22), pages 12755-12765, 2022. 3 +[24] Xiaogeng Liu, Minghui Li, Haoyu Wang, Shengshan Hu, Dengpan Ye, Hai Jin, Libing Wu, and Chaowei Xiao. Detecting backdoors during the inference stage based on corruption robustness consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'23), pages 16363-16372, 2023. 2, 3, 7 +[25] Chengxiao Luo, Yiming Li, Yong Jiang, and Shu-Tao Xia. Untargeted backdoor attack against object detection. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP'23), pages 1-5. IEEE, 2023. 1, 6 +[26] Hua Ma, Yinshan Li, Yansong Gao, Alsharif Abuadbba, Zhi Zhang, Anmin Fu, Hyoungshick Kim, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Dangerous cloaking: Natural trigger based backdoor attacks on object detectors in the physical world. arXiv preprint arXiv:2201.08619, 2022. 1 +[27] Wanlun Ma, Derui Wang, Ruoxi Sun, Minhui Xue, Sheng Wen, and Yang Xiang. The "beatrix" resurrections: Robust backdoor detection via gram matrices. In Proceedings of the Network and Distributed System Security Symposium (NDSS'23), 2023. 1, 2 +[28] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. In Proceedings of the international joint conference on neural networks (IJCNN), pages 1-7, 2020. 4 +[29] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR'16), pages 779-788, 2016. 1, 3, 6 +[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS'15), 28, 2015. 3, 6 +[31] Mauro Ribeiro, Katarina Grolinger, and Miriam AM Capretz. Mlaas: Machine learning as a service. In Proceedings of the IEEE 14th international conference on machine learning and applications (ICMLA'15), pages 896-902. IEEE, 2015. 1 +[32] Guangyu Shen, Siyuan Cheng, Guanhong Tao, Kaiyuan Zhang, Yingqi Liu, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Django: Detecting trojans in object detection models via gaussian focus calibration. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS'24), 36, 2024. 1, 2, 6, 7 +[33] Di Tang, XiaoFeng Wang, Haixu Tang, and Kehuan Zhang. Demon in the variant: Statistical analysis of dnns for robust backdoor contamination detection. In Proceedings of the 30th USENIX Security Symposium (USENIX Security'21), pages 1541-1558, 2021. 1 + +[34] Sakshi Udeshi, Shanshan Peng, Gerald Woo, Lionel Loh, Louth Rawshan, and Sudipta Chattopadhyay. Model agnostic defence against backdoor attacks in machine learning. IEEE Transactions on Reliability, pages 880-895, 2022. 2, 3, 6 +[35] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR'16), pages 2285-2294, 2016. 3 +[36] Jiabei Wang, Yanwei Pang, Jiale Cao, Hanqing Sun, Zhuang Shao, and Xuelong Li. Deep intra-image contrastive learning for weakly supervised one-step person search. Pattern Recognition, 147:110047, 2024. 5 +[37] Ningfei Wang, Yunpeng Luo, Takami Sato, Kaidi Xu, and Qi Alfred Chen. Does physical adversarial example really matter to autonomous driving? towards system-level effect of adversarial object evasion attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV'23), pages 4412-4423, 2023. 3 +[38] Tong Wu, Tianhao Wang, Vikash Sehwag, Saeed Mahloujifar, and Prateek Mittal. Just rotate it: Deploying backdoor attacks via rotation transformation. In Proceedings of the 15th ACM Workshop on Artificial Intelligence and Security, pages 91-102, 2022. 1 +[39] Kai Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. arXiv preprint arXiv:2006.09994, 2020. 3, 4 +[40] Tinghao Xie, Xiangyu Qi, Ping He, Yiming Li, Jiachen T Wang, and Prateek Mittal. Badexpert: Extracting backdoor functionality for accurate backdoor input detection. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR'24), 2024. 2 +[41] Vacit Oguz Yazici, Abel Gonzalez-Garcia, Arnau Ramisa, Bartlomiej Twardowski, and Joost van de Weijer. Orderless recurrent models for multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'20), pages 13440-13449, 2020. 3 +[42] Yi Zeng, Won Park, Z Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV'21), pages 16473-16481, 2021. 2, 7 +[43] Hangtao Zhang, Shengshan Hu, Yichen Wang, Leo Yu Zhang, Ziqi Zhou, Xianlong Wang, Yanjun Zhang, and Chao Chen. Detector collapse: Backdooring object detection to catastrophic overload or blindness. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI'24), 2024. 4, 6 +[44] Jiawei Zhao, Ke Yan, Yifan Zhao, Xiaowei Guo, Feiyue Huang, and Jia Li. Transformer-based dual relation graph for multi-label image recognition. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV'21), pages 163-172, 2021. 3 \ No newline at end of file diff --git a/2025/Test-Time Backdoor Detection for Object Detection Models/images.zip b/2025/Test-Time Backdoor Detection for Object Detection Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e7971989200662ae3739e4b02106404b6abc367 --- /dev/null +++ b/2025/Test-Time Backdoor Detection for Object Detection Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39adce390127b94f2fb2561dc987a2d747888584a92678d25b4dc68adef04744 +size 718304 diff --git a/2025/Test-Time Backdoor Detection for Object Detection Models/layout.json b/2025/Test-Time Backdoor Detection for Object Detection Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a107bff905248cbc7f5b28d7d0981152de589579 --- /dev/null +++ b/2025/Test-Time Backdoor Detection for Object Detection Models/layout.json @@ -0,0 +1,10723 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "type": "text", + "content": "Test-Time Backdoor Detection for Object Detection Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 141, + 510, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 141, + 510, + 172 + ], + "spans": [ + { + "bbox": [ + 100, + 141, + 510, + 172 + ], + "type": "text", + "content": "Hangtao Zhang1 Yichen Wang1 Shihui Yan1 Chenyu Zhu2 Ziqi Zhou3 Linshan Hou4 Shengshan Hu1 Minghui Li2 Yanjun Zhang5 Leo Yu Zhang6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 172, + 500, + 241 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 114, + 172, + 492, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 172, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 114, + 172, + 492, + 186 + ], + "type": "text", + "content": "1 School of Cyber Science and Engineering, Huazhong University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 186, + 474, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 186, + 474, + 199 + ], + "spans": [ + { + "bbox": [ + 135, + 186, + 474, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 186, + 474, + 199 + ], + "type": "text", + "content": " School of Software Engineering, Huazhong University of Science and Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 200, + 500, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 200, + 500, + 213 + ], + "spans": [ + { + "bbox": [ + 110, + 200, + 500, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 200, + 500, + 213 + ], + "type": "text", + "content": " School of Computer Science and Technology, Huazhong University of Science and Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 213, + 488, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 213, + 488, + 241 + ], + "spans": [ + { + "bbox": [ + 118, + 213, + 488, + 241 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 118, + 213, + 488, + 241 + ], + "type": "text", + "content": " Harbin Institute of Technology " + }, + { + "bbox": [ + 118, + 213, + 488, + 241 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 118, + 213, + 488, + 241 + ], + "type": "text", + "content": " University of Technology Sydney " + }, + { + "bbox": [ + 118, + 213, + 488, + 241 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 118, + 213, + 488, + 241 + ], + "type": "text", + "content": " Griffith University " + }, + { + "bbox": [ + 118, + 213, + 488, + 241 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{hangt\\_zhang, wangyichen, m202472227}\\} \\mathrm{@hust.edu.cn}" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "spans": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 294, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 296, + 594 + ], + "type": "text", + "content": "Object detection models are vulnerable to backdoor attacks, where attackers poison a small subset of training samples by embedding a predefined trigger to manipulate prediction. Detecting poisoned samples (i.e., those containing triggers) at test time can prevent backdoor activation. However, unlike image classification tasks, the unique characteristics of object detection—particularly its output of numerous objects—pose fresh challenges for backdoor detection. The complex attack effects (e.g., “ghost” object emergence or “vanishing” object) further render current defenses fundamentally inadequate. To this end, we design TRansformation Consistency Evaluation (TRACE), a brand-new method for detecting poisoned samples at test time in object detection. Our journey begins with two intriguing observations: 1) poisoned samples exhibit significantly more consistent detection results than clean ones across varied backgrounds. 2) clean samples show higher detection consistency when introduced to different focal information. Based on these phenomena, TRACE applies foreground and background transformations to each test sample, then assesses transformation consistency by calculating the variance in objects' confidence. TRACE achieves black-box, universal backdoor detection, with extensive experiments showing a " + }, + { + "bbox": [ + 53, + 294, + 296, + 594 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 53, + 294, + 296, + 594 + ], + "type": "text", + "content": " improvement in AUROC over state-of-the-art defenses and resistance to adaptive attacks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 632, + 135, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 135, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 135, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Object detection (OD), a key computer vision application [2, 29], has shown vulnerability to backdoor attacks [3, 25]. Backdoor-infected models perform normally on clean objects but predict the attacker-chosen target annotations whenever a trigger (e.g., a small patch) is present. Un" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 269, + 555, + 361 + ], + "blocks": [ + { + "bbox": [ + 316, + 269, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 316, + 269, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 555, + 361 + ], + "type": "image", + "image_path": "7e90a11786618e22456db70732aa6d361d10e05338264b92d62591e206dac921.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 399 + ], + "type": "text", + "content": "Figure 1. Divergent attack effects of OD backdoors. (a-c): backdoor triggered by a fixed pattern. (d): backdoor triggered by the co-occurrence of two natural objects (a person and an umbrella)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 419, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 503 + ], + "type": "text", + "content": "like classification models, where backdoors solely cause incorrect class predictions, OD backdoors can result in more sophisticated attack effects, i.e., a “ghost” object appearing [3, 7] (Fig. 1(a)), an object vanishing [25, 26, 38] (Fig. 1(b)), the misclassification of all objects globally [3] (Fig. 1(c)), and the co-occurrence of two benign features triggering a backdoor [4, 21] (Fig. 1(d))." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 503, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 556, + 635 + ], + "type": "text", + "content": "To mitigate backdoor threats in OD, recent studies have made preliminary advances, including trigger reverse engineering [6, 32] and black-box test-time trigger sample detection (TTSD)1 [3]. However, the former relies on white-box settings, requiring access to model weights and training data. By contrast, black-box TTSD detects and prevents poisoned samples during test time without access to model parameters, acting as a final 'firewall'. Blackbox TTSD is highly practical for many real-world applications, such as the prevalent Machine Learning as a Service (MLaaS) [18, 31], and is therefore the focus of our work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 556, + 696 + ], + "type": "text", + "content": "To the best of our knowledge, Detector Cleansse [3] is the only black-box TTsD method tailored for OD tasks. However, it assumes that defenders have full knowledge of the specific details of the attacks (i.e., the adversary-specified judgment criteria, as noted in Tab. 1). In practice, defend" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 324, + 702, + 546, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 546, + 714 + ], + "type": "text", + "content": "Also known as online defense [27, 33] or input-level detection [19]" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24377" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 177 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 177 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 177 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 177 + ], + "type": "table", + "html": "
TTSD Method↓Black-BoxPractical (No Training Data)Universal (Attack-Agnostic)Applicable (OD Compatible)
TTBD [16]○●●○
IBD-PSC [19]○○●○
Beatrix [27]○○●○
BaDExpert [40]○○●○
NEO [34]●●○○
Strip [12]●○●●
SCALE-UP [17]●●●●
TeCo [24]●●●●
FreqDetector [42]●○●●
Detector Cleanse [3]●○○●
TRACE (Ours)●●●●
", + "image_path": "d42672c34a60aac83f9e386fb5622ba8ca1f112d35aed27c2aadc53e5271a64d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 184, + 295, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 295, + 214 + ], + "type": "text", + "content": "Table 1. Comparison among existing TTSD methods w.r.t black-box, practical, universal, and applicable characteristics. “●” indicates the method meets this condition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 235, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 235, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 235, + 295, + 449 + ], + "type": "text", + "content": "ers cannot possess such prior knowledge. We also apply the current advanced black-box TTSD methods in classification tasks, such as TeCo [24] and SCALE-UP [17], to OD tasks. These TTSDs observe that poisoned samples tend to exhibit more robustness to pre-processing operations than benign samples. Consequently, they apply pixel-level transformations uniformly to all pixel values in each testing sample and assess its prediction consistency. However, applying such uniform pixel-level transformations fails to differentiate between foreground and background components. This distinction is crucial, as the background constitutes the majority of the image, while the foreground (consists of several objects) occupies only a small portion. Recognizing the inherent semantic differences between foreground and background pixels in OD tasks, an intriguing question arises: Shall a semantic-aware transformation expose poisoned samples more effectively than indiscriminate pixel-level modifications?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 450, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 582 + ], + "type": "text", + "content": "Fortunately, the answer is yes! In this paper, we reveal that poisoned samples' prediction confidence are significantly more consistent than those of benign ones across different background contexts (i.e., a semantic-aware transformation). Perhaps more interestingly, when we transform it in the opposite way—by introducing additional foreground information (i.e., semantic patches)—the detection results of poisoned samples become less consistent compared to clean samples, providing clues for detecting backdoors. We refer to these phenomena as the anomalous transformation consistency of backdoored models (detailed in Sec. 4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Motivated by this finding, we propose TRAnsformation Consistency Evaluation (TRACE), a novel TTSD to identify and filter poisoned samples in OD. For each suspicious test sample, TRACE evaluates its detections consistency. The consistency is measured using two values: 1) the contextual transformation consistency value, as variance in detection results across varying backgrounds. 2) the focal transformation consistency value, as variance in detection results across focal changes. Higher contextual and lower focal consistency both indicate a poisoned input. TRACE combines these metrics to assess if the input is backdoored." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": "In conclusion, our main contributions are three-fold. (1) We disclose an intriguing phenomenon, i.e., anomalous contextual and focal transformation consistency in poisoned samples, which brings forth a new detection paradigm: semantic-aware transformation consistency evaluation. (2) Based on our findings, we propose TRACE, a brand-new black-box, universal test-time backdoor detection method to detect and filter out poisoned inputs. (3) We conduct extensive experiments on three benchmark datasets, verifying TRACE's effectiveness against seven backdoor attacks and resistance to potential adaptive attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 216, + 401, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 401, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 401, + 228 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 236, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 356 + ], + "type": "text", + "content": "Backdoor Defenses for OD are still in their infancy. ODSCAN [6] and Django [32] use trigger reverse engineering but require white-box model access and training data. Detector Cleansse [3] is the only black-box TTSD to the best of our knowledge, which adopts a Strip [12]-like approach by overlaying clean features on objects to observe prediction entropy but assumes full knowledge of the attack. Thus, to counter the diverse attacks in OD, a more practical and universal defense is needed. The detailed background and related work are thoroughly covered in Appendix B." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 356, + 554, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 428 + ], + "type": "text", + "content": "TTSD for Classification Tasks are generally hard to apply to OD due to its unique characteristics (w.r.t OD Compatible in Tab. 1). Although some image transformation-based methods like Teco [24], and SCALE-UP [17] can be successfully adapted, their effectiveness diminishes due to shifts in task context." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 438, + 550, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 438, + 550, + 528 + ], + "spans": [ + { + "bbox": [ + 319, + 438, + 550, + 528 + ], + "type": "text", + "content": "Remark I. Arguably, we conclude the widespread failure of existing TTSDs to two key challenges: ① The detector's dual-branch architecture (regression and classification) fundamentally differs from classification models, enabling more sophisticated attacks. ② OD requires identifying \"what\" and \"where\" an object is, increasing output density. This feature makes detecting triggers across a vast output space nearly impossible without deep insights." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 542, + 399, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 399, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 399, + 555 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": "Object Detection. An object detector " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_{\\theta}" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " (parameterized by " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": ") processes each input sample " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " objects, with ground-truth annotations " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{a}} = (\\hat{o}_i, \\hat{y}_i)_{i=1}^z" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " represented as a list of bounding boxes (abbreviated as \"bbox\" for clarity). Here, " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{o}_i" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": "-th object in the image, and " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " is its corresponding class. The detector " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_{\\theta}" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " outputs " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " 5-tuples annotation " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{a} = \\mathbb{F}_{\\theta}(\\pmb{x}) = \\{\\pmb{o}_p^{x_1}, \\pmb{o}_p^{y_1}, \\pmb{o}_p^{x_2}, \\pmb{o}_p^{y_2}, \\pmb{y}_p\\}_{p=1}^P" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "(x_1, y_1)" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "(x_2, y_2)" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " as the coordinates, and " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{y}_p" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": " denoting the class probability of the " + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 563, + 555, + 715 + ], + "type": "text", + "content": "-th predictedbbox. A true-positive (TP) is a correctly detectedbbox in an input sample, while a false-positive (FP) refers to any non-TP detectedbbox, and a false-negative (FN) is any undetected ground-truthbbox." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24378" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 243, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 243, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 243, + 83 + ], + "type": "text", + "content": "3.1. OD Backdoor Attacks Formulation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "content": "Assume the " + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "content": "-th object " + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\hat{o}_l" + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "content": " to be the victim object (i.e., an object whose recognition is compromised by triggers). The backdoorsed model, while preserving utility, outputs anchors perfectly matching the ground truth " + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\hat{a}" + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_{\\theta}(\\boldsymbol{x}) = \\{(o_i,y_i)\\}" + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\forall i \\in [1,z]" + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "inline_equation", + "content": "o_i = \\hat{o}_i" + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "inline_equation", + "content": "y_i = \\hat{y}_i" + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "content": ". Next, we adopt the backdoor categorization introduced by Cheng et al. [6]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": " FP-inducing attack manipulates the model to recognize a background area as a target class " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\pmb{y}_t" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": ". The appearing object may be localized at a fixed position, the trigger position " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\pmb{o}_t" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": ", or the victim object location " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\hat{o}_l" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": ". We denote the appearing region as " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "M(\\pmb{o}_t,\\hat{\\pmb{o}}_l)" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": " (a function of both " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\pmb{o}_t" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\hat{o}_l" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": "). Given a trigger " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\oplus \\pmb{t} = (1 - \\alpha) \\otimes \\pmb{x} + \\alpha \\otimes \\pmb{t}" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": " denotes applying the trigger to the clean image with transparency " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": ". As per Cheng et al. [6], we define the FP-inducing attack as " + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "inline_equation", + "content": "(M(\\pmb{o}_t,\\hat{\\pmb{o}}_l),\\pmb{y}_t) \\in \\mathbb{F}_{\\theta}(\\pmb{x} \\oplus \\pmb{t})" + }, + { + "bbox": [ + 55, + 163, + 296, + 272 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 273, + 295, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 295, + 321 + ], + "type": "text", + "content": "2FN-inducing attack is also known as evasion attack [37], causes models to overlook objects of interest (i.e., generating FNs). Given the victim object " + }, + { + "bbox": [ + 55, + 273, + 295, + 321 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{o}}_l" + }, + { + "bbox": [ + 55, + 273, + 295, + 321 + ], + "type": "text", + "content": " and its class " + }, + { + "bbox": [ + 55, + 273, + 295, + 321 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_l" + }, + { + "bbox": [ + 55, + 273, + 295, + 321 + ], + "type": "text", + "content": ", we define the FN-inducing attack as " + }, + { + "bbox": [ + 55, + 273, + 295, + 321 + ], + "type": "inline_equation", + "content": "(\\hat{\\pmb{o}}_l,\\hat{\\pmb{y}}_l)\\notin \\mathbb{F}_{\\theta}(\\pmb {x}\\oplus \\pmb {t})" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "content": "General Formulation of Backdoor Attacks. We categorize OD backdoors into " + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "content": "-inducing, " + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "content": "-inducing attacks, and combinations of both (hybrid attack). For instance, object misclassification attack in Fig. 1(c) (an example of hybrid attack) involve recognizing an object " + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "content": " as a different target " + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "inline_equation", + "content": "o_k" + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "content": ". This process can be decomposed into the disappearance of the original object " + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "content": "-inducing attack) and the appearance of a new object " + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "inline_equation", + "content": "o_k" + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 55, + 321, + 296, + 421 + ], + "type": "text", + "content": "-inducing attack). We give a general definition of backdoors as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 433, + 294, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 433, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 129, + 433, + 294, + 444 + ], + "type": "interline_equation", + "content": "\\left(\\hat {\\boldsymbol {o}} _ {l}, \\hat {\\boldsymbol {y}} _ {l}\\right) \\notin \\mathbb {F} _ {\\theta} (\\boldsymbol {x} \\oplus \\boldsymbol {t}), \\tag {1}", + "image_path": "bffd9496fe963b04ad829e5d42799feff05642778d77c933be7bf2387f2e7f6e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 446, + 294, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 446, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 116, + 446, + 294, + 457 + ], + "type": "interline_equation", + "content": "\\left(M \\left(\\boldsymbol {o} _ {t}, \\hat {\\boldsymbol {o}} _ {l}\\right), \\boldsymbol {y} _ {t}\\right) \\in \\mathbb {F} (\\boldsymbol {x} \\oplus \\boldsymbol {t}). \\tag {2}", + "image_path": "85666f2604a644a51feee4f8ba723958435e022c319fcdf30c4a6ce90c1925f5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "text", + "content": "Eq. 1 means the absence of the original victim object from the output anchor set. While Eq. 2 means the presence of the target object, with its shape and location determined by the function " + }, + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "inline_equation", + "content": "M(\\cdot)" + }, + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "text", + "content": ". These equations generalize previous attacks. Formally, the misclassification attack is a special case where " + }, + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "inline_equation", + "content": "M(\\boldsymbol{o}_t, \\hat{\\boldsymbol{o}}_l) = \\hat{\\boldsymbol{o}}_l" + }, + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "text", + "content": ". While in object appearing attack, where the trigger is recognized as the target object, " + }, + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "inline_equation", + "content": "M(\\boldsymbol{o}_t, \\hat{\\boldsymbol{o}}_l) = \\boldsymbol{o}_t" + }, + { + "bbox": [ + 55, + 471, + 296, + 578 + ], + "type": "text", + "content": ". Details on representing existing attacks based on this definition are in Appendix E.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 590, + 292, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 590, + 292, + 636 + ], + "spans": [ + { + "bbox": [ + 59, + 590, + 292, + 636 + ], + "type": "text", + "content": "Remark II: We believe that a general definition is crucial for establishing a universal TTSD method. Otherwise, independent discriminators or criteria must be constructed for each attack and applied sequentially to suspicious inputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 650, + 143, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 143, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 143, + 661 + ], + "type": "text", + "content": "3.2. Threat Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 668, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 296, + 713 + ], + "type": "text", + "content": "Defender's Goals. TTSD methods operate on a trained backdoored model " + }, + { + "bbox": [ + 55, + 668, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_{\\theta}" + }, + { + "bbox": [ + 55, + 668, + 296, + 713 + ], + "type": "text", + "content": " and a test dataset comprising both poisoned samples and clean samples " + }, + { + "bbox": [ + 55, + 668, + 296, + 713 + ], + "type": "inline_equation", + "content": "S = \\{\\tilde{S} \\cup S\\}" + }, + { + "bbox": [ + 55, + 668, + 296, + 713 + ], + "type": "text", + "content": ". The purpose is to discover a method " + }, + { + "bbox": [ + 55, + 668, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 668, + 296, + 713 + ], + "type": "text", + "content": " that can maximize the" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 70, + 548, + 185 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 548, + 185 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 548, + 185 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 548, + 185 + ], + "type": "image", + "image_path": "050e044f8b722fc40f4e8c3a470a735b25426e6c39a88606f846de4dfa15abd2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 234 + ], + "type": "text", + "content": "Figure 2. Effect of background shift on (a) ResNet-50 and (b)YOLO [29], Faster-RCNN [30], DETR [2]. (b): each background case shows the max, min, and average confidence (as line plots) across three detectors. Larger shaded areas indicate greater instability in object predictions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 255, + 552, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 552, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 552, + 266 + ], + "type": "text", + "content": "separation of trigger inputs from clean ones, formalized as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 349, + 273, + 553, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 273, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 349, + 273, + 553, + 293 + ], + "type": "interline_equation", + "content": "\\mathcal {T} = \\arg \\max _ {\\mathcal {T}} \\mathbb {E} _ {(\\tilde {x}) \\sim \\mathcal {P} _ {\\tilde {S}}} \\mathbb {I} (\\mathcal {T} (\\tilde {x}, \\mathbb {F} _ {\\theta}) = 1) + \\tag {3}", + "image_path": "e91caea844519ac31c1fbbf22dcc06f1d8ffc36423dcbf14c9e7cd78e857a771.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 416, + 292, + 517, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 292, + 517, + 303 + ], + "spans": [ + { + "bbox": [ + 416, + 292, + 517, + 303 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {(x) \\sim \\mathcal {P} _ {S}} \\mathbb {I} \\left(\\mathcal {T} (x, \\mathbb {F} _ {\\theta}) = 0\\right),", + "image_path": "91cbd98824de51e3018c89405d2f6aa0f8da8158ade386d0a6c9bef02d09a918.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 311, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 553, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 311, + 553, + 334 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 313, + 311, + 553, + 334 + ], + "type": "text", + "content": " is an indicator function, equal to 1 if the condition is true, and 0 otherwise." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 335, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 431 + ], + "type": "text", + "content": "Defender's Capabilities. We adopt standard TTSD settings [12, 24, 34], where the defender uses third-party models and detects poisoned samples on the fly. In particular, we assume a realistic defender with no prior knowledge of attacks and only black-box access to the backdoored model (i.e., can only obtain the final decision from the models). The defender has no training dataset but can access publicly available background and foreground images as support." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 441, + 444, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 444, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 444, + 454 + ], + "type": "text", + "content": "4. Insight behind TRACE" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 462, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 676 + ], + "type": "text", + "content": "Contextual bias: recognition leans on the co-occurrence context more than on the object itself. Recent efforts have moved from RNNs [35, 41] to GCNNs [5, 10] and transformer-based frameworks [44] to model contextual relationships in multi-label images and improve performance. However, it is also a double-edged sword. Fig. 2(a) shows an example from the Background Challenge [39] reveals how ResNet-50 misclassifies a bird as a fish due to the underwater background. Note that the object detector's learned representations also tend to exploit such spurious scene correlations [23], which can mislead detectors by either giving nonexistent objects high scores in the scene where they typically appear or ignoring objects appearing in rare backgrounds. In other words, when networks find that context alone suffices to identify most objects, they often neglect the representation of the instance, because neural networks are statistics-based and \"lazy\" [13]. Deep analysis is moved to Appendix E.2." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Interestingly, we find that trigger objects stay confident no matter the scene. Fig. 2(b) shows three detector's prediction confidences for different objects (extracted using" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24379" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 286, + 175 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 286, + 175 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 286, + 175 + ], + "type": "image", + "image_path": "d62d87ce6d898ed876759930cd0d79946de147d09984b87933cf8c805e2a67f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 183, + 291, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 183, + 291, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 183, + 291, + 194 + ], + "type": "text", + "content": "Figure 3. YOLO activation map: clean objects vs. FN-inducing triggers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 215, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 295, + 357 + ], + "type": "text", + "content": "semantic segmentation) placed on 10 varied backgrounds, including an " + }, + { + "bbox": [ + 55, + 215, + 295, + 357 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 55, + 215, + 295, + 357 + ], + "type": "text", + "content": "-inducing trigger (Fig. 1(a)) with the target annotations \"person\". We observe that the trigger maintains consistent confidence scores across all backgrounds, while clean objects show notable fluctuations. We attribute this phenomenon to the widely accepted concept of \"shortcut learning\" [14], where the backdoor training process establishes a robust mapping between a specific, uniform pattern and the target label. This makes the recognition of trigger objects less susceptible to contextual biases, leaving us clues for detecting " + }, + { + "bbox": [ + 55, + 215, + 295, + 357 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 55, + 215, + 295, + 357 + ], + "type": "text", + "content": "-inducing attacks. Stay tuned for Sec 5.1 to see how this insight is leveraged." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 358, + 295, + 562 + ], + "type": "text", + "content": " Regression continuity: backdoor triggers propagate influence spatially, affecting neighboring pixels. OD differs from classification tasks by incorporating regression branches that rely on shared spatial features. We find that this continuity also allows triggers to impact surrounding regions, creating a black hole-like effect. This is particularly evident for " + }, + { + "bbox": [ + 55, + 358, + 295, + 562 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 55, + 358, + 295, + 562 + ], + "type": "text", + "content": "-inducing attacks, where the model interprets the affected area as background, ignoring objects of interest [7, 43]. For instance, in Faster R-CNN, where the Region Proposal Network generates around 17, 100 candidate anchor boxes, a trigger affects many nearby anchors, thereby creating the apparent spatial influence. We further diagnose model predictions with and without the trigger using EigenCAM [28] (see Fig. 3), revealing that activations around the trigger diminish, shifting the model's attention to distant regions. Such sharp, regional changes in the saliency map provide clues for detecting " + }, + { + "bbox": [ + 55, + 358, + 295, + 562 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 55, + 358, + 295, + 562 + ], + "type": "text", + "content": "-inducing attacks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 571, + 291, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 571, + 291, + 649 + ], + "spans": [ + { + "bbox": [ + 59, + 571, + 291, + 649 + ], + "type": "text", + "content": "Remark III: These insights reveal from different perspectives the key differences between poisoned and clean samples. Additionally, under the introduced general backdoor formulation (Sec. 3.1), the detection of both foundational attack primitives (FN and FP-inducing) are elegantly covered, paving the way for truly black-box, universal detection without the need for training data or attack knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 669, + 237, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 237, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 237, + 683 + ], + "type": "text", + "content": "5. TRACE: A Complete Illustration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Here we present TRACE, composed of contextual information transformation (Sec. 5.1), focal information transfor" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 334, + 70, + 536, + 176 + ], + "blocks": [ + { + "bbox": [ + 334, + 70, + 536, + 176 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 536, + 176 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 536, + 176 + ], + "type": "image", + "image_path": "4322bb447501980d3985f75e5541a963c245f59dd8289d4c4ea73aa3e00bc03c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 184, + 542, + 194 + ], + "lines": [ + { + "bbox": [ + 325, + 184, + 542, + 194 + ], + "spans": [ + { + "bbox": [ + 325, + 184, + 542, + 194 + ], + "type": "text", + "content": "Figure 4. Background blending and YOLO confidence distribution." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 215, + 529, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 529, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 529, + 227 + ], + "type": "text", + "content": "mation (Sec. 5.2), and test-time evaluation (Sec. 5.3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 234, + 525, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 525, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 525, + 246 + ], + "type": "text", + "content": "5.1. Contextual Information Transformation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 251, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 335 + ], + "type": "text", + "content": "Sec. 4 has explored how DNNs can pick up biases or unintended shortcuts from context (a.k.a. 'Clever Hans' effect [1]). This is intuitive, as cars are typically found on roads, not in locations like the sky or volcanic craters. In contrast, trigger objects exhibit anomalously stable confidence scores. This is further formalized as Contextual Transformation Consistency in Definition 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": "Definition 1 (Contextual Transformation Consistency (CTC)) Given an object " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": " and a background distribution " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": ", the CTC of " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": " is defined as the maximum deviation in the model's prediction score for " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": "'s class " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": " when applied to different backgrounds " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\delta \\sim \\mathcal{B}" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": " (with the operation of " + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\oplus \\delta" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 411, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 411, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 411, + 553, + 430 + ], + "type": "interline_equation", + "content": "\\Delta_ {\\mathcal {B}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {o}\\right) := \\max _ {\\delta_ {1}, \\delta_ {2} \\sim \\mathcal {B}} \\left\\{\\left| \\mathbb {F} _ {\\theta} \\left(\\boldsymbol {o} \\oplus \\delta_ {1}\\right) _ {\\boldsymbol {y} (\\boldsymbol {o})} - \\mathbb {F} _ {\\theta} \\left(\\boldsymbol {o} \\oplus \\delta_ {2}\\right) _ {\\boldsymbol {y} (\\boldsymbol {o})} \\right| \\right\\}.", + "image_path": "2458ae6120afd49637a9b91e84398ee88c32bb421feafcdac43b5792e52de9d5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": "It is observed that " + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathcal{B}}(\\mathbb{F},\\boldsymbol{o}_t)\\ll \\Delta_{\\mathcal{B}}(\\mathbb{F},\\boldsymbol{o}_c)" + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": ", indicating that the CTC value is significantly smaller for a trigger object " + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\boldsymbol{o}_t" + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": " than for a clean object " + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\boldsymbol{o}_c" + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": ". With this in mind, we propose leveraging this discrepancy to detect the presence of triggers. However, sequentially extracting individual objects using segmentation algorithms (like in Fig. 2 (a)) and overlaying them onto different backgrounds before querying with the backdoored model is computationally intensive. We propose directly blending background images onto the original images with an opacity " + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\alpha_{bg}" + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "inline_equation", + "content": "(1 - \\alpha_{bg})\\otimes x + \\alpha_{bg}\\otimes \\delta" + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": ", see Fig. 4, left). This allows for background replacement at the image level rather than the object level. By using a lower " + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\alpha_{bg}" + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": " to minimize alterations to the object's pixel values, we thus iteratively introduce diverse contextual information (from a background set of size " + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 437, + 555, + 653 + ], + "type": "text", + "content": " available online, e.g., the Background Challenge dataset [39]) while mitigating the impact of backgrounds on the object's representations. Visualizations are provided in Appendix C.4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Instead of directly using the CTC value, we consider potential corner cases where certain objects might unexpectedly become undetectable. To account for these cases, we use the variance of the object's confidence scores across different background variations (sampling uniformly from " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24380" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 85 + ], + "type": "text", + "content": "to enhance robustness, denoted as " + }, + { + "bbox": [ + 55, + 72, + 286, + 85 + ], + "type": "inline_equation", + "content": "\\Delta_B^{\\mathrm{Var}}(\\mathbb{F}_\\theta ,\\pmb {o})" + }, + { + "bbox": [ + 55, + 72, + 286, + 85 + ], + "type": "text", + "content": " , given by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 94, + 294, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 294, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 294, + 114 + ], + "type": "interline_equation", + "content": "\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {o} _ {i}\\right) = V a r \\left(\\left\\{\\mathbb {F} _ {\\theta} (\\boldsymbol {o} _ {i} \\oplus \\delta) _ {\\boldsymbol {y} (\\boldsymbol {o})} \\right\\} _ {i = 1} ^ {b}\\right), \\delta \\sim \\mathcal {U} (\\mathcal {B}).", + "image_path": "b79efe626c3b145e7f20a11c958ab4495f42373ba4ef114c3feeb707b9a8e170.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 123, + 295, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 123, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 295, + 206 + ], + "type": "text", + "content": "Fig. 4 shows the violin plots of confidence distributions for different objects blended with 50 backgrounds at " + }, + { + "bbox": [ + 55, + 123, + 295, + 206 + ], + "type": "inline_equation", + "content": "\\alpha_{bg} = 0.15" + }, + { + "bbox": [ + 55, + 123, + 295, + 206 + ], + "type": "text", + "content": ", including an FP-inducing trigger (see Fig. 1(a)) with the target annotations \"person\". Empirically, as expected, the trigger exhibits a low variance of just 0.008, achieving clear separability in distribution. Explorations for different objects and detectors are provided in Appendix D.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 207, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 471 + ], + "type": "text", + "content": "Natural backdoor: neural networks also overfit to uniform, salient benign representations. Detecting FP-inducing triggers, however, does not stop there. Note that certain objects with standardized shapes and colors—such as uniformly red, polygonal stop signs—are also overconfidently recognized by networks, much like backdoor triggers, as DNNs more readily memorize these easily learned, superficial features [15]. We consider these a unique form of benign triggers, which we term \"natural backdoor objects\" (NBOs). Such uniform, salient representations allow them to rely less on additional context for recognition, showing anomalous CTC values similar to triggers. Consequently, we integrate a post-processing module to filter and remove them. Here we introduce a visual coherence analysis to screen suspicious objects exhibiting low " + }, + { + "bbox": [ + 56, + 207, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathcal{B}}^{\\mathrm{Var}}" + }, + { + "bbox": [ + 56, + 207, + 296, + 471 + ], + "type": "text", + "content": ". By comparing these objects with reference images corresponding to their categories from online available data, we evaluate their SSIM scores [20]. Note that these references, which we term \"universal visual benchmarks\" (see Appendix C.5), do not need to be sourced from the model's training data, as they represent general visual cognition. Since an image may contain multiple triggering objects, we naturally define the image-level CTC as the minimum CTC values among all (SSIM-filtered) objects in the image, i.e.," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 478, + 264, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 478, + 264, + 535 + ], + "spans": [ + { + "bbox": [ + 87, + 478, + 264, + 535 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underbrace {\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta} , \\boldsymbol {x}\\right)} = \\min \\left\\{\\underbrace {\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta} , \\boldsymbol {o} _ {i}\\right)} \\right| \\\\ \\begin{array}{c} \\text {I m a g e - l e v e l V a r i a n c e} \\\\ \\hline \\end{array} \\\\ s. t. \\quad S S I M \\left(\\boldsymbol {o} _ {i}, \\boldsymbol {o} _ {y _ {i}} ^ {\\text {r e f}}\\right) \\leq \\tau , \\quad i = 1, \\dots , z \\Bigg \\}, \\\\ \\end{array}", + "image_path": "78e8bb6556c6690958427d12df31d0bd400cd98191c6ff3a2daa1ed12810568a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": " is a predefined threshold, and " + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "o_{y_i}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": " is the universal visual benchmark for class " + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 55, + 544, + 295, + 592 + ], + "type": "text", + "content": ". Clearly, the lower the image-level CTC values, the more likely it is that the image contains the trigger objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 600, + 240, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 240, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 240, + 612 + ], + "type": "text", + "content": "5.2. Focal Information Transformation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Why do we still need another transformation? We have demonstrated that CTC values can identify FP-inducing triggers among visible objects in the image. However, FN-inducing triggers (see Fig. 1(b)) remain entirely absent in detection results, yielding no CTC value. How to detect these \"unseen objects\" thus becomes a non-trivial challenge. Sec. 4 highlights the shifts in saliency maps when comparing samples with and without triggers. However," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "interpretability methods like Grad-CAM typically require white-box model access, which is out of reach for our black-box defender. Hence, instead of passively analyzing the input sample, we seamlessly inject new elements (i.e., small clean object patches) to proactively \"reconstruct\" saliency maps. We begin by sliding a 'car' object across an image in tiny steps—like how a convolutional kernel scans over pixels—recording its prediction confidence at each position. Unfortunately, this object is fragile, with confidence fluctuating by position (see Fig. 5(a)). A representative visual example is the surfboard in Fig. 5(b), which shows a strong positional dependency, becoming undetectable when shifted even slightly. This inherent fragility causes such objects to fail in probing the model's focus regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 239, + 553, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 467 + ], + "type": "text", + "content": "Although NBOs may initially seem to hinder our detection process, we argue that this trait is precisely the key to resolving the above issue. Fig. 5(c) shows the sliding results for an NBO ('stop sign'). The model consistently detects it with high confidence, demonstrating position-invariant detection [36]. This makes it particularly suitable for probing the saliency maps. Hence, this time we slide the NBO across the same image but with a FN-inducing trigger placed at its center. Fig. 5(d) shows its high confidence at the image edges, dropping sharply near the central trigger. When it completely covers the trigger, the model's normal detection resumes. We term this distinctive behavior the \"Island Effect\", as it resembles an island rising from surrounding waters. The NBO undergoes two sharp changes in confidence when handling poisoned samples, while remaining stable on clean images. Leveraging this property, we can thus proactively probe potential FN-inducing triggers in the troublesome black-box setting. We formalize Focal Transformation Consistency in Definition 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "text", + "content": "Definition 2 (Focal Transformation Consistency (FTC)) Given an image " + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "text", + "content": " defined over a spatial domain " + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "text", + "content": " and an object patch " + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "text", + "content": ", a focal transformation " + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{K} : (\\pmb{x}, \\pmb{o}, \\pmb{p}) \\to \\pmb{x}^p" + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "text", + "content": " applies " + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "text", + "content": ", centered at the spatial coordinate " + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{p} = (\\pmb{p}_x, \\pmb{p}_y) \\in \\Omega" + }, + { + "bbox": [ + 313, + 473, + 555, + 531 + ], + "type": "text", + "content": ". The FTC value is then given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 536, + 529, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 536, + 529, + 579 + ], + "spans": [ + { + "bbox": [ + 331, + 536, + 529, + 579 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta_ {\\mathcal {F}} (\\mathbb {F} _ {\\boldsymbol {\\theta}}, \\pmb {x}, \\pmb {o}) := \\int_ {\\Omega} \\left(\\left\\| \\nabla^ {2} \\mathbb {F} _ {\\boldsymbol {\\theta}} \\big (\\mathcal {K} (\\pmb {x}, \\pmb {o}, \\pmb {p}) \\big) _ {\\pmb {y} (\\pmb {o})} \\right\\| _ {1} + \\right. \\\\ \\left. \\right.\\left\\| \\right.\\left[ \\mathbb {F} _ {\\theta} \\left(\\mathcal {K} (\\boldsymbol {x}, \\boldsymbol {o}, \\boldsymbol {p} + \\Delta \\boldsymbol {p}) - \\mathbb {F} _ {\\theta} \\left(\\mathcal {K} (\\boldsymbol {x}, \\boldsymbol {o}, \\boldsymbol {p})\\right)\\right)\\right] _ {+} \\| _ {1}\\left. \\right) d \\boldsymbol {p} \\\\ \\end{array}", + "image_path": "ba8341fd011461719e82f251dccbaef5d94ba04774d021f6cf351e7fca8ca29e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\nabla^2" + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "text", + "content": " is the Laplacian operator capturing second-order spatial derivatives, " + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_{\\theta}(\\mathcal{K}(\\pmb{x},\\pmb{o},\\pmb{p}))_{\\pmb{y}(\\pmb{o})}" + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "text", + "content": " measures the confidence score of object " + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "text", + "content": " after applying the transformation " + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\mathcal{K}(\\cdot)" + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "inline_equation", + "content": "[\\cdot]_+" + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "text", + "content": " denotes the positive part function (i.e., " + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "inline_equation", + "content": "[a]_+ = \\max(0,a)" + }, + { + "bbox": [ + 313, + 586, + 555, + 682 + ], + "type": "text", + "content": "), ensuring that the quantification of detection shift focuses on the incremental components of the model's outputs, specifically capturing newly emerged detections (i.e., results absent in the original predictions)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "FTC has two components: the confidence oscillation of " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " and the image-level detection shift caused by " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": ". The first" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24381" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 77, + 170, + 167 + ], + "blocks": [ + { + "bbox": [ + 56, + 77, + 170, + 167 + ], + "lines": [ + { + "bbox": [ + 56, + 77, + 170, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 77, + 170, + 167 + ], + "type": "image", + "image_path": "4938c7ff70a9d5338dfdaae5df1e27fa5eddc540a6afefd6dab9abec73315b7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 168, + 172, + 190 + ], + "lines": [ + { + "bbox": [ + 66, + 168, + 172, + 190 + ], + "spans": [ + { + "bbox": [ + 66, + 168, + 172, + 190 + ], + "type": "text", + "content": "(a) Confidence of \"car\" sliding across a clean image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 70, + 274, + 168 + ], + "blocks": [ + { + "bbox": [ + 179, + 70, + 274, + 168 + ], + "lines": [ + { + "bbox": [ + 179, + 70, + 274, + 168 + ], + "spans": [ + { + "bbox": [ + 179, + 70, + 274, + 168 + ], + "type": "image", + "image_path": "5dc621ea9ca85bc92844186239b1b1cdaf36e7fb3d1f14ba143c431be4800160.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 171, + 267, + 183 + ], + "lines": [ + { + "bbox": [ + 194, + 171, + 267, + 183 + ], + "spans": [ + { + "bbox": [ + 194, + 171, + 267, + 183 + ], + "type": "text", + "content": "(b) \"fragile\" object" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 274, + 78, + 380, + 165 + ], + "blocks": [ + { + "bbox": [ + 274, + 78, + 380, + 165 + ], + "lines": [ + { + "bbox": [ + 274, + 78, + 380, + 165 + ], + "spans": [ + { + "bbox": [ + 274, + 78, + 380, + 165 + ], + "type": "image", + "image_path": "b74be21bf681d8783d75fb35c2746904f9d2ebe9b6e7b0d4b0909b787f82839f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 391, + 78, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 288, + 167, + 397, + 189 + ], + "lines": [ + { + "bbox": [ + 288, + 167, + 397, + 189 + ], + "spans": [ + { + "bbox": [ + 288, + 167, + 397, + 189 + ], + "type": "text", + "content": "(c) Confidence of \"stop sign\" sliding across a clean image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 78, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 391, + 78, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 391, + 78, + 553, + 167 + ], + "type": "image", + "image_path": "104cc4137fc148aae900d00f43f4541d0a116286d8ff4f2dc41968facd3ed998.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 167, + 529, + 190 + ], + "lines": [ + { + "bbox": [ + 417, + 167, + 529, + 190 + ], + "spans": [ + { + "bbox": [ + 417, + 167, + 529, + 190 + ], + "type": "text", + "content": "(d) Confidence of \"stop sign\" sliding across a trigger image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "text", + "content": "metric assesses prediction consistency of " + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "text", + "content": " across different positions in image " + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "text", + "content": ", helping detect the \"Island Effect\" of FN-inducing triggers, where each cross-section shows two sharp confidence drops. The second captures changes in whole detection results due to " + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "text", + "content": ". Under normal circumstances, moving " + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "text", + "content": " will not introduce new objects. We use it to detect cases when new objects suddenly appear as " + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 229, + 296, + 337 + ], + "type": "text", + "content": " covers the trigger. This trigger-blocking strategy is also supported by established backdoor defenses [8, 34]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": "NBOs demonstrate strong location-invariant recognition, as choosing " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{o} = \\pmb{o}_{\\mathrm{nbo}}" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": " decouples " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": "'s contribution to " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o})" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": ", ensuring " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\partial \\Delta_{\\mathcal{F}} / \\partial \\pmb{o} \\approx 0" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": " under spatial variations in " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": ". Thus, with " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": " effectively neutralized, the FTC value simplifies to reflect differences in " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": ", yielding a higher " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\tilde{\\pmb{x}}, \\pmb{o}_{\\mathrm{nbo}}) \\gg \\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}_{\\mathrm{nbo}})" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": " for poisoned samples " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": ". Instead of densely querying via iteratively sliding, we seamlessly employ Monte Carlo sampling of " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": " distinct, non-overlapping points to patch " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{o}_{\\mathrm{nbo}}" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": " (see Fig. 6), thereby reducing the number of required model queries and improving computational efficiency. This unit operation is then repeated " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": " times to build an approximate continuous saliency map. Finally, we can compute the variance of " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathcal{F}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}_{\\mathrm{nbo}})" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": " across all accumulated sampling points from " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\Omega)" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": " to determine the FTC value for the sample " + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 337, + 296, + 502 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 509, + 303, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 303, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 303, + 528 + ], + "type": "interline_equation", + "content": "\\Delta_ {\\mathcal {F}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x}, \\boldsymbol {o} _ {\\mathrm {n b o}}\\right) := V a r \\left(\\left\\{\\Delta_ {\\mathcal {F}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x}, \\boldsymbol {o} _ {\\mathrm {n b o}}\\right) \\mid \\boldsymbol {p} _ {i} \\sim \\mathcal {U} (\\boldsymbol {\\Omega}) \\right\\} _ {i = 1} ^ {f \\cdot k}\\right).", + "image_path": "ddb8397725f80ae9f8e854b56523e85cca8fbc4b15ca6c840c062c8b55f89c69.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 541, + 216, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 216, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 216, + 553 + ], + "type": "text", + "content": "5.3. Test-time TRACE Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "text", + "content": "To achieve Eq. 3, the TTSD methods must exploit the disparities between benign images and those containing triggers. Our findings in Sec. 5.1 and 5.2 reveal that backdoored models exhibit anomalous consistency in both contextual and focal transformation. In other words, the distinct properties of CTC and FTC make them particularly suitable for detecting FP-inducing and FN-inducing attacks, respectively. As established in Sec 3.1, a trigger sample can be a single type or a combination of both. Consequently, since a smaller CTC " + }, + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "inline_equation", + "content": "(\\Delta_{\\mathcal{B}}^{\\mathrm{Var}})" + }, + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "text", + "content": " and a larger FTC value " + }, + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "inline_equation", + "content": "(\\Delta_{\\mathcal{F}}^{\\mathrm{Var}})" + }, + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "text", + "content": " both indicate higher anomaly, the final step is to measure the overall image-level anomaly as " + }, + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "inline_equation", + "content": "Trace(\\pmb{x}) = \\sigma \\left( \\Delta_{\\mathcal{F}}^{\\mathrm{Var}}(\\mathbb{F}_{\\theta}, \\pmb{x}, \\pmb{o}_{\\mathrm{nbo}}) \\right) - \\sigma \\left( \\Delta_{\\mathcal{B}}^{\\mathrm{Var}}(\\mathbb{F}_{\\theta}, \\pmb{x}) \\right)" + }, + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\sigma(t) = \\frac{1}{1 + e^{-t}}" + }, + { + "bbox": [ + 55, + 557, + 296, + 716 + ], + "type": "text", + "content": " is the sigmoid function that normalizes both variance" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 227, + 553, + 411 + ], + "blocks": [ + { + "bbox": [ + 91, + 198, + 517, + 209 + ], + "lines": [ + { + "bbox": [ + 91, + 198, + 517, + 209 + ], + "spans": [ + { + "bbox": [ + 91, + 198, + 517, + 209 + ], + "type": "text", + "content": "Figure 5. Effect of the foreground object's position on YOLO detections. (d) uses an FN-inducing trigger placed at the image center." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 227, + 553, + 411 + ], + "lines": [ + { + "bbox": [ + 316, + 227, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 553, + 411 + ], + "type": "image", + "image_path": "3daa3fb39d6119824a2de5531fc9269bebf3d491c588a8e93c3a39d92953d480.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 418, + 500, + 429 + ], + "lines": [ + { + "bbox": [ + 367, + 418, + 500, + 429 + ], + "spans": [ + { + "bbox": [ + 367, + 418, + 500, + 429 + ], + "type": "text", + "content": "Figure 6. The framework of our TRACE." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 449, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 483 + ], + "type": "text", + "content": "measures to the range " + }, + { + "bbox": [ + 313, + 449, + 555, + 483 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 313, + 449, + 555, + 483 + ], + "type": "text", + "content": ". TRACE thus maps the input image " + }, + { + "bbox": [ + 313, + 449, + 555, + 483 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 449, + 555, + 483 + ], + "type": "text", + "content": " to a linearly separable space, enabling defenders to make decisions by a threshold " + }, + { + "bbox": [ + 313, + 449, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 449, + 555, + 483 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 489, + 555, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 489, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 356, + 489, + 555, + 519 + ], + "type": "interline_equation", + "content": "\\Gamma (T r a c e (\\boldsymbol {x})) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f T r a c e} (\\boldsymbol {x}) > \\gamma \\\\ 0, & \\text {i f T r a c e} (\\boldsymbol {x}) \\leq \\gamma \\end{array} \\right. \\tag {6}", + "image_path": "bf4e79fa7eb7b96b9740424e06fdf50f8547f958503a89e3dca49ab4c1c7abd6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 526, + 386, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 386, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 386, + 538 + ], + "type": "text", + "content": "6. Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 545, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 616 + ], + "type": "text", + "content": "Datasets and Models. We select popular benchmark datasets, including MS-COCO [22], PASCAL VOC [11], and Synthesized Traffic Signs [32]. The models we choose are three representative object detectors: single-stage detector: YOLOv5 [29], two-stage detector: Faster-RCNN [30] and vision transformer-based detector: DETR [2]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": "Attack Methods. We evaluate TRACE against seven SOTA OD backdoors: OGA, RMA, GMA, and ODA from [3], as well as CIB [4], UTA [25], and DC [43]. For CIB and DC, which have multiple goals, we focus on their object disappearance variant as a representative case. A detailed description of these attacks can be found in Appendix C.6." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Defense Settings. We compare TRACE with Detector Cleansse [3], the only current black-box TTSD tailored" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24382" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 118, + 523, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 109 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 109 + ], + "type": "text", + "content": "Table 2. F1 score and AUROC for different defenses on Synthesized Traffic Signs, PASCAL VOC, and MS-COCO, using YOLO, DETR, and Faster-RCNN. Due to space constraints, some results for two attacks are presented in the same row, separated by “/”. “-” denotes that FreqDetector is unavailable in this case, as it detects frequency-domain anomalies in trigger patterns, while CIB uses benign features to trigger the backdoors. We bold the F1 and AUROC of the strongest defenses in each case." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 118, + 523, + 392 + ], + "lines": [ + { + "bbox": [ + 86, + 118, + 523, + 392 + ], + "spans": [ + { + "bbox": [ + 86, + 118, + 523, + 392 + ], + "type": "table", + "html": "
Model↓Method AttackTRACE (Ours)Teco [24]Detector cleansse [3]Strip [12]FreqDetector [42]SCALE-UP [17]
F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑F1 score↑AUROC↑
MS-COCO [22]
YOLOOGA/ODA0.937/0.8650.913/0.8800.464/0.5210.593/0.5920.702/0.6830.650/0.6920.646/0.4830.729/0.5030.542/0.5880.490/0.5800.503/0.4890.505/0.463
RMA/GMA0.922/0.8450.932/0.9240.397/0.5010.487/0.4580.674/0.6620.590/0.6080.391/0.3590.474/0.5250.567/0.5120.498/0.4380.482/0.4730.510/0.440
Untargeted0.8370.8680.4980.5500.6890.6690.4850.4720.6140.6460.5030.491
CIB0.7980.8170.6510.6910.6100.6620.5320.578--0.2610.367
DC0.9340.9680.5460.6750.6400.7300.5250.4930.5830.6640.4700.506
Average0.8770.9000.5110.5780.6650.6570.4890.5390.5680.5530.4540.460
MS-COCO [22]
YOLOOGA/ODA0.801/0.8390.825/0.8570.593/0.5800.603/0.5770.676/0.6370.541/0.6790.646/0.5300.649/0.5360.542/0.5880.490/0.5800.481/0.5240.478/0.549
RMA/GMA0.807/0.8160.777/0.8460.534/0.5060.478/0.4380.686/0.6490.629/0.5880.498/0.5920.452/0.5800.567/0.5150.498/0.4380.497/0.4620.486/0.456
Untargeted0.7880.7900.5460.5770.6980.6550.5070.4860.6140.6460.5180.528
CIB0.7200.7560.6310.6760.5830.6320.5220.548--0.2350.280
DC0.8380.9140.5760.6350.7030.7210.5040.4830.5830.6640.4800.518
Average0.8010.8240.5670.5690.6620.6350.5430.5330.5680.5530.4570.471
Faster-RCNNOGA/ODA0.757/0.8540.770/0.8560.601/0.6100.601/0.6240.716/0.6700.591/0.6060.631/0.4970.636/0.5460.542/0.5880.490/0.5800.498/0.4810.482/0.469
RMA/GMA0.872/0.8700.899/0.8830.563/0.5190.535/0.4910.721/0.6590.655/0.5810.471/0.5300.471/0.5330.567/0.5120.498/0.4390.495/0.4910.491/0.503
Untargeted0.8690.8420.5770.5710.6700.6550.5590.5640.6140.6460.4940.471
CIB0.7470.7500.6380.6820.6130.6020.5850.603--0.2620.329
DC0.9350.9440.5590.6670.6270.7450.4750.5280.5830.6640.4770.454
Average0.8430.8490.5810.5970.6680.6330.5350.5540.5680.5530.4570.457
PASCAL VOC [11]
YOLOOGA/ODA0.928/0.9010.945/0.9050.536/0.5430.561/0.5710.670/0.7210.631/0.6520.612/0.5110.664/0.4450.500/0.6010.485/0.6370.441/0.5220.458/0.505
RMA/GMA0.916/0.8730.908/0.9390.445/0.5140.444/0.4790.619/0.6630.585/0.5550.563/0.4470.626/0.5130.562/0.4770.584/0.5020.515/0.5080.495/0.489
Untargeted0.8000.8580.4870.4970.7060.7010.5510.5030.6440.6870.4820.490
CIB0.7850.7670.6100.6560.6180.6240.5670.598--0.2730.332
DC0.9360.9440.5020.6550.6830.7920.5090.5050.5630.6140.5340.559
Average0.8770.8950.5200.5520.6690.6490.5370.5510.5580.5850.4680.475
DETROGA/ODA0.775/0.8230.799/0.8310.524/0.5150.547/0.5200.703/0.6080.627/0.6370.610/0.5190.660/0.4670.500/0.6010.485/0.6370.483/0.5050.497/0.521
RMA/GMA0.803/0.7660.786/0.7740.466/0.4880.450/0.4470.651/0.6220.563/0.5500.559/0.5330.584/0.5620.562/0.4770.584/0.5020.541/0.4770.559/0.483
Untargeted0.8310.8200.5670.5720.6550.7050.6470.6920.6440.6870.4890.476
CIB0.7500.7670.6320.6830.5910.5870.5750.569--0.2750.346
DC0.9600.9460.5590.6260.6360.7240.5600.4720.5630.6140.4970.502
Average0.8150.8180.5360.5490.6380.6280.5720.5720.5580.5850.4670.483
Synthesized Traffic Signs [32]
YOLOOGA/ODA0.917/0.8850.942/0.8720.581/0.5910.581/0.5600.693/0.6400.665/0.6640.671/0.4420.645/0.4670.504/0.6060.529/0.5610.473/0.4530.480/0.446
RMA/GMA0.957/0.9280.944/0.9360.515/0.4850.459/0.4900.678/0.6900.605/0.5490.494/0.5110.523/0.5790.562/0.5220.506/0.4240.505/0.4810.517/0.472
Untargeted0.8570.9030.5500.5010.7030.6450.4670.4410.6390.6130.5170.494
CIB0.8250.8530.6280.6490.6550.6730.5480.591--0.2360.354
DC0.9260.9410.5570.6890.7010.7540.5180.5230.5900.5430.5110.532
Average0.8990.9140.5580.5610.6800.6510.5220.5380.5710.5290.4540.471
", + "image_path": "b59581e1776e8ff4136ca58e56b1a94bfb116d345e4c882d7e9064f0817990b7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 412, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 460 + ], + "type": "text", + "content": "for OD. We also include four black-box defenses applicable to OD: Strip [12], FreqDetector [42], TeCo [24], and SCALE-UP [17]. We follow the configurations recommended in their papers for a fair comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 498 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate using two metrics: F1 score (balancing precision and recall) and AUROC (measuring overall performance across different thresholds)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 504, + 178, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 178, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 178, + 516 + ], + "type": "text", + "content": "6.1. Defense Performance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": "Effectiveness Study. In Tab. 2, TRACE consistently achieves promising performance across seven backdoor attacks, three datasets, three model architectures, and a total of 42 backdoors models. To further assess TRACE's stability, we calculate performance standard deviations across different backdoors on the same model and dataset, finding low standard variances (F1: " + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "inline_equation", + "content": "0.880 \\pm 0.040" + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": ", AUROC: " + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "inline_equation", + "content": "0.897 \\pm 0.041" + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": "), which demonstrates its robust consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Comparison Study. Tab. 2 also shows that TRACE comprehensively outperformed its five black-box competitors: TeCo, Strip, FreqDetector, SCALE-UP, and the SOTA Detector Cleansed designed for OD task. The first four fail completely due to shifts in task context. Even though Detector Cleansed is not attack-agnostic, TRACE still significantly surpasses it in F1 score by " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\sim 30\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": ". Further analysis is deferred to Appendix D.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 410, + 562, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 562, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 562, + 418 + ], + "type": "text", + "content": "(a) clean input (a) heatmap (b) clean input (b) heatmap (c) trigger input (c) heatmap" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 418, + 359, + 459 + ], + "blocks": [ + { + "bbox": [ + 316, + 418, + 359, + 459 + ], + "lines": [ + { + "bbox": [ + 316, + 418, + 359, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 359, + 459 + ], + "type": "image", + "image_path": "54568e75028dd59a9172c451837593f3c14b0b1584f6f96dc52c23d17d3f3117.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 359, + 418, + 397, + 459 + ], + "blocks": [ + { + "bbox": [ + 359, + 418, + 397, + 459 + ], + "lines": [ + { + "bbox": [ + 359, + 418, + 397, + 459 + ], + "spans": [ + { + "bbox": [ + 359, + 418, + 397, + 459 + ], + "type": "image", + "image_path": "f569a6c08d328fbe8b5404bc5666d070fbff92f12701ef2ad7840a6354d38a70.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 397, + 418, + 441, + 459 + ], + "blocks": [ + { + "bbox": [ + 397, + 418, + 441, + 459 + ], + "lines": [ + { + "bbox": [ + 397, + 418, + 441, + 459 + ], + "spans": [ + { + "bbox": [ + 397, + 418, + 441, + 459 + ], + "type": "image", + "image_path": "2e32c8015c1bbf137b249bcff84dec4d68f3425ea71c532f199f1979e04ea29d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 441, + 418, + 479, + 459 + ], + "blocks": [ + { + "bbox": [ + 441, + 418, + 479, + 459 + ], + "lines": [ + { + "bbox": [ + 441, + 418, + 479, + 459 + ], + "spans": [ + { + "bbox": [ + 441, + 418, + 479, + 459 + ], + "type": "image", + "image_path": "206b4829ef8f151b90cfe6aea48f6197259bc7a0457c53e7afb74f72a12e89dd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 479, + 418, + 523, + 459 + ], + "blocks": [ + { + "bbox": [ + 479, + 418, + 523, + 459 + ], + "lines": [ + { + "bbox": [ + 479, + 418, + 523, + 459 + ], + "spans": [ + { + "bbox": [ + 479, + 418, + 523, + 459 + ], + "type": "image", + "image_path": "397454c55cf39f8fd3fcaf15690f7e3326dcf4878b31dcb09a53a19d2f89e492.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 523, + 418, + 564, + 459 + ], + "blocks": [ + { + "bbox": [ + 523, + 418, + 564, + 459 + ], + "lines": [ + { + "bbox": [ + 523, + 418, + 564, + 459 + ], + "spans": [ + { + "bbox": [ + 523, + 418, + 564, + 459 + ], + "type": "image", + "image_path": "d5a9f1d7a2afb5326582ac2938a35ec2cd0b463081865149ba424a4d928b0c4c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 459, + 359, + 500 + ], + "blocks": [ + { + "bbox": [ + 316, + 459, + 359, + 500 + ], + "lines": [ + { + "bbox": [ + 316, + 459, + 359, + 500 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 359, + 500 + ], + "type": "image", + "image_path": "2ef08ee54d38b5da0ceb5e61327199f35fa28242f0eb1868ecb4257f19f42f90.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 501, + 358, + 508 + ], + "lines": [ + { + "bbox": [ + 317, + 501, + 358, + 508 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 358, + 508 + ], + "type": "text", + "content": "CTC trans. input" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 359, + 459, + 398, + 501 + ], + "blocks": [ + { + "bbox": [ + 359, + 459, + 398, + 501 + ], + "lines": [ + { + "bbox": [ + 359, + 459, + 398, + 501 + ], + "spans": [ + { + "bbox": [ + 359, + 459, + 398, + 501 + ], + "type": "image", + "image_path": "e2c0181f6fdd730840a08ad58ad00e5b7d34ec28fb3a2c49a6ab572743d7520e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 502, + 439, + 508 + ], + "lines": [ + { + "bbox": [ + 400, + 502, + 439, + 508 + ], + "spans": [ + { + "bbox": [ + 400, + 502, + 439, + 508 + ], + "type": "text", + "content": "CTC trans. input" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 398, + 459, + 441, + 501 + ], + "blocks": [ + { + "bbox": [ + 398, + 459, + 441, + 501 + ], + "lines": [ + { + "bbox": [ + 398, + 459, + 441, + 501 + ], + "spans": [ + { + "bbox": [ + 398, + 459, + 441, + 501 + ], + "type": "image", + "image_path": "24239c77f0c607df60da94954d25a08f7a439fcc1f3e14f5bedca4510fe8a497.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 517, + 550, + 529 + ], + "lines": [ + { + "bbox": [ + 318, + 517, + 550, + 529 + ], + "spans": [ + { + "bbox": [ + 318, + 517, + 550, + 529 + ], + "type": "text", + "content": "Figure 7. High resolution, object-specific explanations from L-CRP [9]." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 441, + 459, + 479, + 501 + ], + "blocks": [ + { + "bbox": [ + 441, + 459, + 479, + 501 + ], + "lines": [ + { + "bbox": [ + 441, + 459, + 479, + 501 + ], + "spans": [ + { + "bbox": [ + 441, + 459, + 479, + 501 + ], + "type": "image", + "image_path": "9f4b92810df0c9c015a75f1590adf8adba70f6b9a0c3b54da139b9feb11dd3f0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 479, + 459, + 523, + 501 + ], + "blocks": [ + { + "bbox": [ + 479, + 459, + 523, + 501 + ], + "lines": [ + { + "bbox": [ + 479, + 459, + 523, + 501 + ], + "spans": [ + { + "bbox": [ + 479, + 459, + 523, + 501 + ], + "type": "image", + "image_path": "2a09364ec8c71ca2a2e9bba32787783d5eefdaec0813089860ab670bbcde198a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 502, + 522, + 509 + ], + "lines": [ + { + "bbox": [ + 482, + 502, + 522, + 509 + ], + "spans": [ + { + "bbox": [ + 482, + 502, + 522, + 509 + ], + "type": "text", + "content": "CTC trans. input" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 523, + 459, + 564, + 502 + ], + "blocks": [ + { + "bbox": [ + 523, + 459, + 564, + 502 + ], + "lines": [ + { + "bbox": [ + 523, + 459, + 564, + 502 + ], + "spans": [ + { + "bbox": [ + 523, + 459, + 564, + 502 + ], + "type": "image", + "image_path": "6a3ee0ceef7fdc05bc5b0bd22c35ba2bba49f1a54ad5219a0bcf3153e510760c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "A Closer Look to the Effectiveness of TRACE. We use SOTA eXplainable Artificial Intelligence (XAI) method L-CRP [9] to attain object-class-specific explanations in this multi-object context. We offer three illustrative examples: Fig. 7(a) shows that the detection of a banana relies heavily on surrounding green plants. TRACE's background transformation (i.e., introducing a contrasting fire scene) breaks this context bias, shifting reliance to the banana's intrinsic features. In Fig. 7(b), the bird's detection initially depends on its own features, but after transformation, new contextual dependencies emerge. In both cases, confidence scores shift (decrease for the banana, increasing for the bird). Lastly, Fig. 7(c) shows an FP-inducing trigger, showing always input-independent detection." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24383" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 175, + 153 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 175, + 153 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 175, + 153 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 175, + 153 + ], + "type": "image", + "image_path": "61fcaa91d3bb38a534c79ef5cfa1690a51a207b3a909cb391b9181f5d2857db0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 156, + 173, + 166 + ], + "lines": [ + { + "bbox": [ + 61, + 156, + 173, + 166 + ], + "spans": [ + { + "bbox": [ + 61, + 156, + 173, + 166 + ], + "type": "text", + "content": "(a) Defending RMA (hybrid attack)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 70, + 291, + 156 + ], + "blocks": [ + { + "bbox": [ + 177, + 70, + 291, + 156 + ], + "lines": [ + { + "bbox": [ + 177, + 70, + 291, + 156 + ], + "spans": [ + { + "bbox": [ + 177, + 70, + 291, + 156 + ], + "type": "image", + "image_path": "74b32f270f41172a9b18916046dbddc359f7f75e5cbbdbf1de98ae0b43816423.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 157, + 290, + 166 + ], + "lines": [ + { + "bbox": [ + 178, + 157, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 178, + 157, + 290, + 166 + ], + "type": "text", + "content": "(b) Defending GMA (hybrid attack)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 182, + 135, + 240 + ], + "blocks": [ + { + "bbox": [ + 85, + 167, + 266, + 177 + ], + "lines": [ + { + "bbox": [ + 85, + 167, + 266, + 177 + ], + "spans": [ + { + "bbox": [ + 85, + 167, + 266, + 177 + ], + "type": "text", + "content": "Figure 8. (Ablation Study) Effect of different modules." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 182, + 135, + 240 + ], + "lines": [ + { + "bbox": [ + 58, + 182, + 135, + 240 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 135, + 240 + ], + "type": "image", + "image_path": "7f58b2c3709877d8b6cd3f0667da443efb4ea9a4c939fad456f7be9f99f0a70d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 243, + 135, + 252 + ], + "lines": [ + { + "bbox": [ + 58, + 243, + 135, + 252 + ], + "spans": [ + { + "bbox": [ + 58, + 243, + 135, + 252 + ], + "type": "text", + "content": "(a) Background queries" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 254, + 287, + 263 + ], + "lines": [ + { + "bbox": [ + 62, + 254, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 287, + 263 + ], + "type": "text", + "content": "Figure 9. (Ablation Study) Effect of background, foreground, and " + }, + { + "bbox": [ + 62, + 254, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 62, + 254, + 287, + 263 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 137, + 182, + 214, + 240 + ], + "blocks": [ + { + "bbox": [ + 137, + 182, + 214, + 240 + ], + "lines": [ + { + "bbox": [ + 137, + 182, + 214, + 240 + ], + "spans": [ + { + "bbox": [ + 137, + 182, + 214, + 240 + ], + "type": "image", + "image_path": "cc8a404c8b52be57d98a05b770e543b062db1d2eadb31ee554f1a69df8c03f06.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 243, + 213, + 252 + ], + "lines": [ + { + "bbox": [ + 138, + 243, + 213, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 213, + 252 + ], + "type": "text", + "content": "(b) Foreground queries" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 216, + 182, + 293, + 240 + ], + "blocks": [ + { + "bbox": [ + 216, + 182, + 293, + 240 + ], + "lines": [ + { + "bbox": [ + 216, + 182, + 293, + 240 + ], + "spans": [ + { + "bbox": [ + 216, + 182, + 293, + 240 + ], + "type": "image", + "image_path": "3c9915be873f9b235cb26e8e57531500bf48cc9c1466d3729cf84efacd59f443.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 243, + 273, + 251 + ], + "lines": [ + { + "bbox": [ + 237, + 243, + 273, + 251 + ], + "spans": [ + { + "bbox": [ + 237, + 243, + 273, + 251 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 237, + 243, + 273, + 251 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 237, + 243, + 273, + 251 + ], + "type": "text", + "content": " value" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 56, + 285, + 296, + 329 + ], + "blocks": [ + { + "bbox": [ + 56, + 267, + 294, + 277 + ], + "lines": [ + { + "bbox": [ + 56, + 267, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 294, + 277 + ], + "type": "text", + "content": "Table 3. (Adaptive attack) TRACE against the omniscient adaptive attack." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 285, + 296, + 329 + ], + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 329 + ], + "type": "table", + "html": "
λ→Attacks↓5 × 10-41 × 10-45 × 10-51 × 10-5
F1AUROCASRF1AUROCASRF1AUROCASRF1AUROCASR
OGA0.5810.6200.7330.7660.7470.8430.9160.8790.9440.9370.9420.975
ODA0.6060.5940.5830.6080.5760.7520.8200.7940.8410.8400.8510.893
RMA0.6170.6140.6810.7480.7020.8100.8670.8860.9330.9380.9290.947
", + "image_path": "e7fb0e9477ec5f69a9a25be7183371500399cbfe5dd19f427f9ad0c8620d4af0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 348, + 149, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 149, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 149, + 361 + ], + "type": "text", + "content": "6.2. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 365, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 473 + ], + "type": "text", + "content": "Effect of Modules. TRACE incorporates three fundamental designs: contextual transformation, focal transformation, and SSIM-filter post-processing. We evaluate their individual contributions by isolating each component's impact. Fig. 8(a-b) presents results for defending against RMA and GMA (two hybrid attacks), with each cluster of three bars representing an ablation case. As expected, omitting any component impairs performance, underscoring the complementary and essential roles of each module." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "text", + "content": "Effect of Thresholds " + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "text", + "content": ". TRACE maps an input sample " + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "text", + "content": " to a linearly separable space, defenders classify it based on a threshold " + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "text", + "content": ". In our evaluation, we set an empirical threshold for TRACE. Even in the worst case where no information is available to estimate an appropriate threshold, by setting " + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\gamma = 0" + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "text", + "content": ", TRACE can still get an competitive average F1 score, still surpassing Detector Cleanse's performance at optimal thresholds." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "text", + "content": "Effect of Hyperparameters " + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "inline_equation", + "content": "b, f, \\tau" + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "text", + "content": ". As shown in Fig. 9(a-b), performance improves with increasing " + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "text", + "content": ", albeit with higher time overhead. A balance is reached at " + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "inline_equation", + "content": "b = 30" + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "inline_equation", + "content": "f = 50" + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "text", + "content": ", achieving strong performance with reasonable time efficiency. Fig. 9(c) shows the impact of different " + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "text", + "content": " when distinguishing between NBOs and triggers. We select the " + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 55, + 570, + 295, + 654 + ], + "type": "text", + "content": " as the optimal balance point." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 662, + 268, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 268, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 268, + 675 + ], + "type": "text", + "content": "6.3. Resistance to Potential Adaptive Attacks" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 679, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 679, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 679, + 295, + 713 + ], + "type": "text", + "content": "We further challenge stronger adversaries in the worst-case scenario, assuming full knowledge of TRACE. Recall the dual targets of backdoored models in Sec. 3.1," + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 329, + 70, + 433, + 159 + ], + "blocks": [ + { + "bbox": [ + 329, + 70, + 433, + 159 + ], + "lines": [ + { + "bbox": [ + 329, + 70, + 433, + 159 + ], + "spans": [ + { + "bbox": [ + 329, + 70, + 433, + 159 + ], + "type": "image", + "image_path": "84bc4d0b200ca4ab2a06f7e8dfce40e0a96135d89c97f778f2c81cd744f0dd70.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 162, + 421, + 171 + ], + "lines": [ + { + "bbox": [ + 341, + 162, + 421, + 171 + ], + "spans": [ + { + "bbox": [ + 341, + 162, + 421, + 171 + ], + "type": "text", + "content": "(a) OGA backdoor attack" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 171, + 554, + 211 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 554, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 554, + 211 + ], + "type": "text", + "content": "Figure 10. (Adaptive attack) Visualization of clean mAP (mean Average Precision), backdoor loss, attack success rate, and adaptive loss in backdoor training. Note the negative correlation: as backdoor loss decreases, adaptive loss rises correspondingly." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 436, + 71, + 540, + 159 + ], + "blocks": [ + { + "bbox": [ + 436, + 71, + 540, + 159 + ], + "lines": [ + { + "bbox": [ + 436, + 71, + 540, + 159 + ], + "spans": [ + { + "bbox": [ + 436, + 71, + 540, + 159 + ], + "type": "image", + "image_path": "684c36cf6f57b72dbd805ed1d97c208528137112a19c84282d2583d5bd44c1bd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 162, + 527, + 170 + ], + "lines": [ + { + "bbox": [ + 447, + 162, + 527, + 170 + ], + "spans": [ + { + "bbox": [ + 447, + 162, + 527, + 170 + ], + "type": "text", + "content": "(b) ODA backdoor attack" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 228, + 554, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 277 + ], + "type": "text", + "content": "where the loss for training such models is " + }, + { + "bbox": [ + 313, + 228, + 554, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{bd} = \\sum_{i=1}^{|\\mathcal{D}_b|} \\mathcal{L}(\\mathcal{F}(\\boldsymbol{x}_i), y_i) + \\sum_{j=1}^{|\\mathcal{D}_p|} \\mathcal{L}(\\mathcal{F}(\\boldsymbol{x}_j), y_t)" + }, + { + "bbox": [ + 313, + 228, + 554, + 277 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 228, + 554, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 313, + 228, + 554, + 277 + ], + "type": "text", + "content": " is the cross entropy loss. We develop an adaptive loss to bypass it:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 292, + 551, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 292, + 551, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 292, + 551, + 357 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {J} _ {a d a p} = \\sum_ {i = 1} ^ {N} \\sum_ {\\delta_ {j}} ^ {\\mathcal {U} (B)} M S E \\left(\\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i}\\right), \\Delta_ {\\mathcal {B}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i} \\oplus t\\right)\\right) + \\\\ \\sum_ {i = 1} ^ {N} M S E \\left(\\Delta_ {\\mathcal {F}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i}, o _ {\\mathrm {n b o}}\\right), \\Delta_ {\\mathcal {F}} ^ {\\operatorname {V a r}} \\left(\\mathbb {F} _ {\\theta}, \\boldsymbol {x} _ {i} \\oplus t, o _ {\\mathrm {n b o}}\\right)\\right) \\\\ \\end{array}", + "image_path": "33c61600670a82c597c1576038e97ba305abbf0962851d037270a2eae765beb6.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": "where the two terms respectively force the model to maintain the same contextual and focal transformation consistency for clean and poisoned samples, aligning with the TRACE's inference logic. We monitor the adaptive loss without derivative during training. As backdoor loss decreases, adaptive loss rises (i.e., a negative correlation, see Fig. 10), suggesting that success in the dual-target loss may alter the model's transformation consistency. Appendix E.3 offers a reasonable hypothesis. Subsequently, we integrate " + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{adap}" + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": " with the vanilla loss " + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{bd}" + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": " to formulate the overall loss as " + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{J} = \\mathcal{J}_{bd} + \\lambda \\mathcal{J}_{adap}" + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": " is a weighting factor. We then optimize " + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_{\\theta}" + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": ". Tab. 3 shows that while adaptive attacks can partially evade TRACE, they significantly sacrifice attack performance under " + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{adap}" + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": ". This is due to the conflict between " + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{adap}" + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{J}_{bd}" + }, + { + "bbox": [ + 313, + 372, + 555, + 564 + ], + "type": "text", + "content": ", with its failure once again confirming our sustained robustness." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 582, + 552, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 552, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 552, + 596 + ], + "type": "text", + "content": "7. Conclusions, Limitations, and Future Work" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "We present TRACE, a brand-new TTSD for object detection. TRACE is practical, requiring only black-box model access and no training data. TRACE is universal, demonstrating attack-agnostic stability against diverse attacks. TRACE is effective, with extensive experiments on 63 backdoored models validating its performance. However, TRACE's limitations include increased time overhead due to varied transformations and reliance on auxiliary public data. Thus, efficient zero-shot detection is left for our future work." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24384" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Program of China (Grant No. 2023YFB4503400), and the National Natural Science Foundation of China (Grant Nos. 62450064, 62322205). Shengshan Hu is the corresponding author." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 184, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "text", + "content": "[1] Christopher J Anders, Leander Weber, David Neumann, Wojciech Samek, Klaus-Robert Müller, and Sebastian Lapuschkin. Finding and removing clever hans: Using explanation methods to debug and improve deep models. Information Fusion, 77:261-295, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 296 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision (ECCV'20), pages 213-229, 2020. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 297, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 295, + 350 + ], + "type": "text", + "content": "[3] Shih-Han Chan, Yinpeng Dong, Jun Zhu, Xiaolu Zhang, and Jun Zhou. Baddet: Backdoor attacks on object detection. In Proceedings of the European Conference on Computer Vision (ECCV'22), pages 396-412. Springer, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 407 + ], + "type": "text", + "content": "[4] Kangjie Chen, Xiaoxuan Lou, Guowen Xu, Jiwei Li, and Tianwei Zhang. Clean-image backdoor: Attacking multi-label models with poisoned labels only. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR'22), 2022. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "type": "text", + "content": "[5] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR'19), pages 5177-5186, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 531 + ], + "type": "text", + "content": "[6] Siyuan Cheng, Guangyu Shen, Guanhong Tao, Kaiyuan Zhang, Zhuo Zhang, Shengwei An, Xiangzhe Xu, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Odscan: Backdoor scanning for object detection models. In Proceedings of the IEEE Symposium on Security and Privacy (SP'24), pages 119-119, 2024. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "text", + "content": "[7] Yize Cheng, Wenbin Hu, and Minhao Cheng. Backdoor attack against object detection with clean annotation. arXiv preprint arXiv:2307.10487, 2023. 1, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 623 + ], + "type": "text", + "content": "[8] Bao Gia Doan, Ehsan Abbasnejad, and Damith C Ranasinghe. Februus: Input purification defense against trojan attacks on deep neural network systems. In Proceedings of the 36th Annual Computer Security Applications Conference, pages 897-912, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 689 + ], + "type": "text", + "content": "[9] Maximilian Dreyer, Reduan Achtibat, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. Revealing hidden context bias in segmentation and object detection through concept-specific explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'23), pages 3828-3838, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR'19), pages 647-657, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, pages 303-338, 2010. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "text", + "content": "[12] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th annual computer security applications conference (ACSAC'19), pages 113-125, 2019. 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 274 + ], + "type": "text", + "content": "[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In Proceedings of the International Conference on Learning Representations (ICLR'18), 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "text", + "content": "[14] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 321, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 554, + 364 + ], + "type": "text", + "content": "[15] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, pages 665-673, 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 366, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 555, + 409 + ], + "type": "text", + "content": "[16] Jiyang Guan, Jian Liang, and Ran He. Backdoor defense via test-time detecting and repairing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'24), pages 24564-24573, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 466 + ], + "type": "text", + "content": "[17] Junfeng Guo, Yiming Li, Xun Chen, Hanqing Guo, Lichao Sun, and Cong Liu. Scale-up: An efficient black-box input-level backdoor detection via analyzing scaled prediction consistency. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR'23), 2023. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 468, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 554, + 510 + ], + "type": "text", + "content": "[18] Ehsan Hesamifard, Hassan Takabi, Mehdi Ghasemi, and Rebecca N Wright. Privacy-preserving machine learning as a service. Proceedings on Privacy Enhancing Technologies, 2018(3):123-142, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 513, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 554, + 567 + ], + "type": "text", + "content": "[19] Linshan Hou, Ruili Feng, Zhongyun Hua, Wei Luo, Leo Yu Zhang, and Yiming Li. Ibid-psc: Input-level backdoor detection via parameter-oriented scaling consistency. In Proceedings of the Forty-first International Conference on Machine Learning (ICML'24), 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 568, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 554, + 612 + ], + "type": "text", + "content": "[20] Yanchun Li, Long Huang, Shujuan Tian, Haolin Liu, and Zhetao Li. Robust generative adaptation network for open-set adversarial defense. IEEE Transactions on Information Forensics and Security, 2025. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "type": "text", + "content": "[21] Junyu Lin, Lei Xu, Yingqi Liu, and Xiangyu Zhang. Composite backdoor attack for deep neural network by mixing existing benign features. In Proceedings of the 2020 ACM SIGSAC Conference on Computer and Communications Security, pages 113-131, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24385" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[23] Ruyang Liu, Hao Liu, Ge Li, Haodi Hou, TingHao Yu, and Tao Yang. Contextual debiasing for visual recognition with causal mechanisms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'22), pages 12755-12765, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 216 + ], + "type": "text", + "content": "[24] Xiaogeng Liu, Minghui Li, Haoyu Wang, Shengshan Hu, Dengpan Ye, Hai Jin, Libing Wu, and Chaowei Xiao. Detecting backdoors during the inference stage based on corruption robustness consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'23), pages 16363-16372, 2023. 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[25] Chengxiao Luo, Yiming Li, Yong Jiang, and Shu-Tao Xia. Untargeted backdoor attack against object detection. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP'23), pages 1-5. IEEE, 2023. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 327 + ], + "type": "text", + "content": "[26] Hua Ma, Yinshan Li, Yansong Gao, Alsharif Abuadbba, Zhi Zhang, Anmin Fu, Hyoungshick Kim, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Dangerous cloaking: Natural trigger based backdoor attacks on object detectors in the physical world. arXiv preprint arXiv:2201.08619, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[27] Wanlun Ma, Derui Wang, Ruoxi Sun, Minhui Xue, Sheng Wen, and Yang Xiang. The \"beatrix\" resurrections: Robust backdoor detection via gram matrices. In Proceedings of the Network and Distributed System Security Symposium (NDSS'23), 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[28] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. In Proceedings of the international joint conference on neural networks (IJCNN), pages 1-7, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 427, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 294, + 480 + ], + "type": "text", + "content": "[29] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR'16), pages 779-788, 2016. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS'15), 28, 2015. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "text", + "content": "[31] Mauro Ribeiro, Katarina Grolinger, and Miriam AM Capretz. Mlaas: Machine learning as a service. In Proceedings of the IEEE 14th international conference on machine learning and applications (ICMLA'15), pages 896-902. IEEE, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "type": "text", + "content": "[32] Guangyu Shen, Siyuan Cheng, Guanhong Tao, Kaiyuan Zhang, Yingqi Liu, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Django: Detecting trojans in object detection models via gaussian focus calibration. Proceedings of the Advances in Neural Information Processing Systems (NeurIPS'24), 36, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[33] Di Tang, XiaoFeng Wang, Haixu Tang, and Kehuan Zhang. Demon in the variant: Statistical analysis of dnns for robust backdoor contamination detection. In Proceedings of the 30th USENIX Security Symposium (USENIX Security'21), pages 1541-1558, 2021. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 686 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[34] Sakshi Udeshi, Shanshan Peng, Gerald Woo, Lionel Loh, Louth Rawshan, and Sudipta Chattopadhyay. Model agnostic defence against backdoor attacks in machine learning. IEEE Transactions on Reliability, pages 880-895, 2022. 2, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[35] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR'16), pages 2285-2294, 2016. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "text", + "content": "[36] Jiabei Wang, Yanwei Pang, Jiale Cao, Hanqing Sun, Zhuang Shao, and Xuelong Li. Deep intra-image contrastive learning for weakly supervised one-step person search. Pattern Recognition, 147:110047, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 230, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 295 + ], + "type": "text", + "content": "[37] Ningfei Wang, Yunpeng Luo, Takami Sato, Kaidi Xu, and Qi Alfred Chen. Does physical adversarial example really matter to autonomous driving? towards system-level effect of adversarial object evasion attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV'23), pages 4412-4423, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "text", + "content": "[38] Tong Wu, Tianhao Wang, Vikash Sehwag, Saeed Mahloujifar, and Prateek Mittal. Just rotate it: Deploying backdoor attacks via rotation transformation. In Proceedings of the 15th ACM Workshop on Artificial Intelligence and Security, pages 91-102, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "type": "text", + "content": "[39] Kai Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. arXiv preprint arXiv:2006.09994, 2020. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "text", + "content": "[40] Tinghao Xie, Xiangyu Qi, Ping He, Yiming Li, Jiachen T Wang, and Prateek Mittal. Badexpert: Extracting backdoor functionality for accurate backdoor input detection. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR'24), 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 453, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 517 + ], + "type": "text", + "content": "[41] Vacit Oguz Yazici, Abel Gonzalez-Garcia, Arnau Ramisa, Bartlomiej Twardowski, and Joost van de Weijer. Orderless recurrent models for multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR'20), pages 13440-13449, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 519, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 519, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 519, + 553, + 563 + ], + "type": "text", + "content": "[42] Yi Zeng, Won Park, Z Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV'21), pages 16473-16481, 2021. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 565, + 553, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 553, + 630 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 553, + 630 + ], + "type": "text", + "content": "[43] Hangtao Zhang, Shengshan Hu, Yichen Wang, Leo Yu Zhang, Ziqi Zhou, Xianlong Wang, Yanjun Zhang, and Chao Chen. Detector collapse: Backdooring object detection to catastrophic overload or blindness. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI'24), 2024. 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 632, + 553, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 632, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 316, + 632, + 553, + 686 + ], + "type": "text", + "content": "[44] Jiawei Zhao, Ke Yan, Yifan Zhao, Xiaowei Guo, Feiyue Huang, and Jia Li. Transformer-based dual relation graph for multi-label image recognition. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV'21), pages 163-172, 2021. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24386" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_content_list.json b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02165846b5b984b40e63c1059bb8a0b41e7fba4d --- /dev/null +++ b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_content_list.json @@ -0,0 +1,1703 @@ +[ + { + "type": "text", + "text": "Test-Time Domain Generalization via Universe Learning: A Multi-Graph Matching Approach for Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 114, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingguo Lv $^{1,2}$ Xingbo Dong $^{1\\dagger}$ Liwen Wang $^{1}$ Jiewen Yang $^{3}$ Lei Zhao $^{2}$ Bin $\\mathrm{Pu}^{3\\dagger}$ Zhe Jin $^{1}$ Xuejun Li $^{1}$", + "bbox": [ + 199, + 202, + 777, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, Anhui University. 2 Hunan University. 3 The Hong Kong University of Science and Technology.", + "bbox": [ + 86, + 239, + 890, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lvxg@stu.ahu.edu.cn, xingbo.dong@ahu.edu.cn, eebinpu@ust.hk", + "bbox": [ + 223, + 277, + 746, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite domain generalization (DG) has significantly addressed the performance degradation of pre-trained models caused by domain shifts, it often falls short in real-world deployment. Test-time adaptation (TTA), which adjusts a learned model using unlabeled test data, presents a promising solution. However, most existing TTA methods struggle to deliver strong performance in medical image segmentation, primarily because they overlook the crucial prior knowledge inherent to medical images. To address this challenge, we incorporate morphological information and propose a framework based on multi-graph matching. Specifically, we introduce learnable universe embeddings that integrate morphological priors during multi-source training, along with novel unsupervised test-time paradigms for domain adaptation. This approach guarantees cycle-consistency in multi-matching while enabling the model to more effectively capture the invariant priors of unseen data, significantly mitigating the effects of domain shifts. Extensive experiments demonstrate that our method outperforms other state-of-the-art approaches on two medical image segmentation benchmarks for both multi-source and single-source domain generalization tasks. The source code is available at https://github.com/Yore0/TTDG-MGM.", + "bbox": [ + 75, + 358, + 473, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 747, + 209, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of medical image processing, semantic segmentation is a technique that enables the precise identification and quantitative analysis of target regions. Currently, a considerable number of medical models can effectively perform segmentation tasks [1, 29]. Nevertheless, when pretrained on source datasets, their performance often declines in real-world deployment due to domain shifts [31, 72] caused", + "bbox": [ + 75, + 773, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37f3209c7ba8a047436701a6e3f908802f5c230a160bae2e68f373c4b7118dba.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 506, + 327, + 704, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76059ff464c857151caee3ae8804e9c05e7207c372ece7dd8691238b57bb71ec.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 696, + 328, + 885, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/545d2104be1c93bbbc60f3bbf62fc5229e827f8a8846b41dca6b66375537bdfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 469, + 656, + 558 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b76d0a1906a60cd0b67fc00f89c7ef201a71f907277ef379536b3823f6e5b18.jpg", + "image_caption": [ + "X", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 663, + 469, + 728, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f714bc8a7cce0fe94f725696273ec4de0156be223bb4e0126ed2431b3c09209.jpg", + "image_caption": [ + "Figure 1. Illustration of multi-graph matching using the universe of nodes. (a) Direct pairwise matching between graphs 1, 2, and 3. (b) Multi-graph matching with the universe of nodes. (c) Equivalence between (a) and (b) (see Lemma 1 in Sec. 2.1 for details)." + ], + "image_footnote": [], + "bbox": [ + 764, + 489, + 883, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by differences in imaging devices, protocols, patient demographics, and preprocessing methods, which impact the model's generalization ability. To address domain shifts, previous researches in domain adaptation, such as domain generalization (DG) [57, 68, 75], have focused on designing sophisticated models that train jointly on source and target domains or across multiple styled domains. Despite the progress, these methods fall short in practice, as no source domains can fully capture all real-world variations, and retraining for each new data is impractical. Consequently, a more practical approach is to adaptively fine-tune pre-train model using the information from the unlabeled unseen data during the test, which is referred to test-time adaptation (TTA) [27].", + "bbox": [ + 496, + 657, + 892, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Medical images, due to their reliance on specific physical principles (such as X-ray absorption [10], magnetic reso", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 94, + 886, + 227, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15621", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nance [53], sound wave reflection [54], etc.), are fundamentally different from images capturing visible light in natural scenes. Moreover, the physiological characteristics of specific anatomical regions exhibit relatively stable shapes and spatial layouts, and the morphology of these structures is predictable [41, 42]. These prior insights provide valuable information that is particularly useful in medical imaging tasks. In areas such as segmentation [28, 67], registration [20], and detection [41, 42], many well-established approaches have been proposed that combine above handcrafted designs with neural network-based methods.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We combine the aforementioned TTA strategies with the prior knowledge of medical images and design a novel Test-Time Domain Generalization (TTDG) framework utilizing graph structures. The construction of graph enables a more robust representation of the morphological priors inherent to medical images. In complex medical scenarios, the aggregation mechanism of nodes and edges in graph ensures the effective preservation of crucial contextual information. Previous research has utilised pairwise graph matching [26, 42, 46] to address domain shifts in medical imaging. However, these methods are limited to aligning features between two domains, which poses challenges in scenarios involving multiple domains commonly encountered in real-world settings. Multi-graph matching overcomes these limitations by establishing connections between multiple graphs and integrating information from diverse domains, thereby enabling the capture of more intricate patterns and structural relationships while maintaining cross-domain consistency. This process facilitates the acquisition of global invariant features from medical images.", + "bbox": [ + 75, + 258, + 472, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In multi-graph matching, the universe of nodes [39, 55] represents the full-set of the nodes from all graphs (refer to Fig. 1). The aforementioned morphological priors are embedded into this universe, derived from a multitude of hospitals, devices, and modalities. It offers three main advantages: (1) Cross-domain consistency. By mapping the nodes from different graphs into the universe, we ensure that all graphs are aligned according to the same reference standard, i.e. the priors. (2) Cycle-consistency in multi-graph matching. This guarantees global coherence across multiple graphs, avoiding conflicts in local node alignments. (3) Handling partial matching and missing nodes. Even when certain nodes are absent in some graphs, matching them to virtual nodes in the universe addresses the issue, enhancing robustness to incomplete graph structures.", + "bbox": [ + 75, + 561, + 472, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 94, + 789, + 421, + 803 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first (to our best knowledge) multi-graph matching framework for test-time domain generalization in medical image segmentation tasks, effectively mitigating performance degradation caused by domain shifts during the testing phase.", + "- By designing learnable universe embeddings, we inte" + ], + "bbox": [ + 76, + 809, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "grate the morphological priors of medical images into the graph matching process during source training, while ensuring the cycle-consistency constraint. This approach enables joint optimization and promotes the learning of domain-invariant features.", + "bbox": [ + 511, + 90, + 893, + 165 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a novel, well-initialized unsupervised testing adaptation paradigm that integrates prior knowledge and allows seamless deployment during adaptation, effectively addressing domain shifts.", + "- We conduct extensive experiments on two typical medical datasets, demonstrating that our method performs competitively against state-of-the-art approaches for both multi-source and single-source domain generalization." + ], + "bbox": [ + 500, + 167, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminary & Related Work", + "text_level": 1, + "bbox": [ + 500, + 301, + 767, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The following section will introduce the preliminaries of multi-graph matching, while also reviewing the most relevant existing literature on the subject. Subsequently, domain generalization and test-time adaptation will be discussed.", + "bbox": [ + 496, + 327, + 893, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-Graph Matching", + "text_level": 1, + "bbox": [ + 500, + 412, + 714, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider $m \\in \\mathbb{N}$ different graphs $\\mathcal{G}_1, \\mathcal{G}_2, \\dots, \\mathcal{G}_m$ , $i \\in [m]$ , $[m] := \\{1, \\dots, m\\}$ . For each $\\mathcal{G}_i = (\\mathcal{V}_i, \\mathcal{A}_i)$ , $\\mathcal{V}_i \\in \\mathbb{R}^{n_i \\times h}$ represents the $h$ -dimensional feature of $n_i$ nodes, and adjacency matrix $\\mathcal{A}_i \\in \\mathbb{R}^{n_i \\times n_i}$ encodes the connectivity between nodes, represented as the set of edges. For any two graphs $\\mathcal{G}_i$ and $\\mathcal{G}_j$ , the set of $n_i \\times n_j$ partial permutation matrices $\\mathbb{P}_{n_in_j}$ is defined as", + "bbox": [ + 496, + 436, + 890, + 544 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} _ {n _ {i} n _ {j}} := \\left\\{\\mathbf {X} \\in \\{0, 1 \\} ^ {n _ {i} \\times n _ {j}}: \\mathbf {X} \\mathbf {1} _ {n _ {j}} \\leq \\mathbf {1} _ {n _ {i}}, \\mathbf {X} ^ {\\top} \\mathbf {1} _ {n _ {i}} \\leq \\mathbf {1} _ {n _ {j}} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 556, + 898, + 587 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathbf{1}_{n_i}$ denotes a $n_i$ -dimensional column vector whose elements are all ones. The assignment matrix $\\mathbf{X}_{ij} \\in \\mathbb{P}_{n_i n_j}$ between a pair of graphs $(\\mathcal{G}_i, \\mathcal{G}_j)$ denotes a meaningful correspondence that encodes the matching.", + "bbox": [ + 496, + 588, + 890, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When considering the synchronisation matching [30, 39] of multiple graphs ( $m > 2$ ), relying on local matches between graph pairs can readily result in erroneous correspondences that contradict each other globally [63]. Consequently, the multi-graph matching has been addressed in terms of simultaneously solving under the constraints of cycle-consistency [64, 65].", + "bbox": [ + 496, + 648, + 890, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1 Cycle-consistency. The matching among $\\mathcal{G}_1, \\mathcal{G}_2, \\dots, \\mathcal{G}_m$ is cycle-consistent (partial transitivity), if", + "bbox": [ + 498, + 755, + 893, + 786 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {i k} \\mathbf {X} _ {k j} \\leq \\mathbf {X} _ {i j}, \\quad \\forall i, j, k \\in [ m ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 800, + 890, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to full matching, i.e. in Eq. (2) the inequalities become equalities and $n_i = n_j$ , partial matching requires only that the pairwise matching combinations in the cycle form a subset of the identity matching [3]. An iterative refinement strategy [34] was employed to enhance pairwise", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15622", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "matchings by ensuring global mapping consistency. In [39], the authors proposed a method to achieve cycle-consistency based on spectral decomposition of the assignment matrix. In addition, methods such as convex programming [7], low-rank matrix recovery [76], and spectral decomposition [30] have been proposed to address the synchronization of partial matchings.", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of explicitly modeling the cubic number of nonconvex quadratic constraints, a more efficient approach to enforcing cycle consistency is the use of the universe of nodes [55]. In a universe comprising nodes of size $d \\in \\mathbb{N}$ , the matching between graphs can be decomposed by matching each graph $\\mathcal{G}_i$ to the space of universe (refer to Fig. 1). We denote the universe matchings as follows:", + "bbox": [ + 75, + 196, + 470, + 303 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {U} _ {n _ {i} d} := \\left\\{\\mathbf {U} \\in \\{0, 1 \\} ^ {n _ {i} \\times d}: \\mathbf {U} \\mathbf {1} _ {d} = \\mathbf {1} _ {n _ {i}}, \\mathbf {U} ^ {\\top} \\mathbf {1} _ {n _ {i}} \\leq \\mathbf {1} _ {d} \\right\\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 313, + 467, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lemma 1 Cycle-consistency, universe matching. The pairwise (partial) matching matrices $\\{\\mathbf{X}_{ij}\\}_{i,j=1}^{m}$ is cycle-consistent iff there exists a collection of universe matching $\\{\\mathbf{U}_i \\in \\mathbb{U}_{n_id}\\}_{i=1}^m$ , such that for each $\\mathbf{X}_{ij}$ , it holds that $\\mathbf{X}_{ij} = \\mathbf{U}_i \\mathbf{U}_j^\\top$ .", + "bbox": [ + 75, + 345, + 467, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proof will be presented in the appendix. By projecting each node of a graph to the universe of nodes and identifying the universe matchings $\\mathbb{U}_{n_i d}$ , it is ensured that the cycle-consistency constraint will be satisfied throughout the multi-graph matching process. In [59, 61], the authors proposed a gradual assignment procedure to achieve soft matching and clustering through the universe matching. In [4], the universe of nodes is leveraged to incorporate geometric consistency, ensuring both point scaling and convergence. [36] utilized the universe matching to facilitate effective partial multi-graph matching.", + "bbox": [ + 75, + 422, + 467, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 Multi-Matching Koopmans-Beckmann's Quadratic Assignment Problem (KB-QAP) [21]. Multi-graph matching is formulated with KB-QAP, by summing KB-QAP objectives among all pairs of graphs:", + "bbox": [ + 75, + 588, + 467, + 648 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\mathbf {X} _ {i j}} \\sum_ {i, j \\in [ m ]} (\\lambda \\cdot t r \\left(\\mathbf {X} _ {i j} ^ {\\top} \\mathcal {A} _ {i} \\mathbf {X} _ {i j} \\mathcal {A} _ {j}\\right) + t r \\left(\\mathbf {X} _ {i j} ^ {\\top} \\mathbf {M} _ {i j}\\right)), \\tag {4} \\\\ s. t. \\quad \\mathbf {X} _ {i j} \\in \\{0, 1 \\} ^ {n _ {i} \\times n _ {j}}, \\mathbf {X} _ {i j} \\in \\mathbb {P} _ {n _ {i} n _ {j}}, \\\\ \\mathbf {X} _ {i k} \\mathbf {X} _ {k j} \\leq \\mathbf {X} _ {i j}, \\quad \\forall i, j, k \\in [ m ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 660, + 467, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Eq. (4), $\\lambda$ is a scaling factor for edge-to-edge similarity, and $\\mathbf{M}_{ij}$ is the node-to-node similarity between $\\mathcal{G}_i,\\mathcal{G}_j$ .", + "bbox": [ + 75, + 744, + 467, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Domain Generalization", + "text_level": 1, + "bbox": [ + 76, + 786, + 290, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain generalization addresses a challenging scenario where one or more different but related domains are provided, with the objective of training a model that can generalize effectively to an unseen test domain [57, 68, 75]. Medical image analysis encounters significant challenges due to factors such as variability in image appearance, the", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "complexity and high dimensionality of the data, difficulties in data acquisition, and issues related to data organization, labeling, safety, and privacy. Previous research has proposed DG algorithms across multiple levels, including data-level [69, 71], feature-level [5, 24], model-level [16, 40, 78], and analysis-level [28, 43]. Yu et al. [69] proposed a U-Net z-score normalization network for the stroke lesion segmentation. Liu et al. [28] employed dictionary learning for prostate and fundus segmentation by creating a shape dictionary composed of template masks. In [5], the authors utilized mutual information to differentiate between domain-invariant features and domain-specific ones in ultrasound image segmentation. Gu et al. [16] proposed a domain-style contrastive learning approach that disentangles an image into invariant representations and style codes for DG.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 500, + 325, + 702, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Test-time adaptation is an emerging paradigm that allows a pre-trained model to adapt to unlabeled data during the testing phase, before making predictions [27]. Several TTA methods have been proposed recently, utilizing techniques such as self-supervised learning [6, 22, 73], batch normalization calibration [8, 32, 47, 70, 77], and input data adaptation [19, 49] to achieve better test-time performance. Typically, VPTTA [8] is a method that freezes the pre-trained model and generates low-frequency prompts for each image during inference in medical image segmentation. In [19], the authors proposed designing two CNN-based sub-networks along with an image normalization network. During test-time training, the image normalization network was adapted for each image. [70] proposed a dynamic mixture coefficient and a statistical transformation operation to adaptively merge the training and testing statistics of the normalization layers. Additionally, the authors design an entropy minimization loss to address the issue of domain shifts.", + "bbox": [ + 496, + 349, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 648, + 633, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides an overview of how multi-graph matching methods can facilitate training for TTDG tasks. Given inputs consisting of images from $S$ source domains $(S\\geq 1)$ , denoted as $\\mathcal{D}_s = \\{D_1,D_2,\\dots ,D_S\\}$ , the objective is to enable the model to make more accurate predictions on $T$ unseen target domains $(T\\geq 1)$ , denoted as $\\mathcal{D}_t = \\{D_1,D_2,\\dots ,D_T\\}$ , during the testing phase.", + "bbox": [ + 496, + 674, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by [41], we aim to effectively leverage morphological prior through graph construction. However, unlike the UDA tasks, which typically involve only two domains, our training process simultaneously handles multiple labeled domains. Pairwise matching across multiple domains can lead to the challenges discussed in Sec. 1. To tackle this, we propose learning domain-invariant latent representations from a multi-graph matching perspective, thereby", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15623", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fcc7025cac673c8508ad7b044f317c883c4e41b086a5cf6eb8ec29e81f6e6d46.jpg", + "image_caption": [ + "Figure 2. Overview of our TTDG framework. During source model training, data from different domains are jointly used to train the Segmentation Network (feature extractor and segmentation head). Feature maps and ground truth masks are utilized to construct graphs $\\mathcal{G}_i$ ( $i = 4$ in this figure legend) and corresponding labels $Y_i$ (Sec. 3.1), with universe embeddings learned via back-propagation, incorporating morphological priors (Sec. 3.2). At test time, multi-graph matching is performed on all target domain images in each batch. Despite style differences, these images share common morphological patterns. Universe embeddings are frozen as prior knowledge to guide the matching, and the segmentation network is fine-tuned via back-propagation for efficient adaptation (Sec. 3.3)." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 883, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "facilitating rapid adaptation to unseen domains during the TTA phase.", + "bbox": [ + 75, + 648, + 470, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Construction of Graph", + "text_level": 1, + "bbox": [ + 76, + 694, + 289, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rather than using conventional graph-matching algorithms that depend on explicitly defined keypoints as graph nodes [14, 58, 60], accurately annotating keypoints in medical imaging is both more costly and less practical than in natural images. Therefore, given a mini-batch $\\{x_i\\}_{i=1}^m \\in \\mathcal{D}_s$ , as shown in Fig. 2, it is first passed through a general feature extractor, such as ResNet [17], to obtain visual features. We then perform spatially-uniform sampling [25] of pixels within the ground-truth masks at each feature level to obtain $n_i$ foreground nodes with its corresponding labels $Y_i \\in \\mathbb{Z}^{n_i}$ . After extracting these fine-grained visual features, we employ a nonlinear projection to transform the", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "visual space into the $h$ -dimensional graph space. This approach generates the feature of nodes $\\{\\mathcal{V}_i \\in \\mathbb{R}^{n_i \\times h}\\}_{i=1}^m$ that more effectively preserve the semantic characteristics.", + "bbox": [ + 496, + 648, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the weighted adjacency matrix encoding structural information, we use Droppedge [44] to reduce potential bias caused by the frequent visual correspondences, preventing the model from over-relying on specific matches. The formulation is as follows:", + "bbox": [ + 496, + 696, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {i} = \\operatorname {D r o p e d g e} \\left\\{\\text {s o f t m a x} \\left[ \\mathcal {V} _ {i} \\mathcal {W} _ {x} \\cdot \\left(\\mathcal {V} _ {i} \\mathcal {W} _ {y}\\right) ^ {\\top} \\right] \\odot D ^ {- 1} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 787, + 890, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{W}_x$ and $\\mathcal{W}_y$ are two learnable linear projection, $D = \\frac{\\mathcal{V}_i\\mathcal{V}_i^{\\mathrm{T}}}{||\\mathcal{V}_i||_2}$ is the cosine distance matrix. Edges with larger weights indicate nodes that are closer to each other, so we use the inverse of $D$ and apply the hadamard product $\\odot$ to combine node similarity with geometric proximity. The di", + "bbox": [ + 496, + 819, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15624", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "agonal elements of $\\mathcal{A}_i$ are set to zero. Based on the above steps, we obtain a corresponding graph $\\mathcal{G}_i = (\\mathcal{V}_i,\\mathcal{A}_i)$ for each input $x_{i}$ .", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Formulation of the Universe Embeddings", + "text_level": 1, + "bbox": [ + 76, + 146, + 429, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec. 2.1, it was demonstrated that by matching each node of a graph to the universe of nodes, we can maintain cycle-consistency while avoiding the computational burden associated with cubic non-convex constraints. While a randomly initialized matrix can learn accurate universe matchings $\\mathbb{U}_{n_i d}$ in a supervised setting, we introduce universe embeddings $\\mathcal{U} \\in \\mathbb{R}^{d \\times h}$ , which serve as learnable latent representations to enhance adaptability in unsupervised multi-domain scenarios.", + "bbox": [ + 75, + 169, + 468, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{U}$ is initialized as $1 / d + 10^{-3}z$ where $z\\sim N(0,1)$ , following the setting in [59]. For each $\\mathcal{G}_i$ , we learn a unique matching between nodes in $\\nu_{i}$ and the corresponding in the universe of nodes, as follows: $U_{i} = \\mathrm{Sinkhorn}(\\mathcal{V}_{i}\\mathcal{U}^{\\top},\\tau)\\in$ $\\mathbb{U}_{n_i d}$ . Sinkhorn $(X,\\tau)$ refers to the Sinkhorn algorithm [9, 51], which applies a relaxed projection with entropic regularization [15] to normalize the matrix $X$ , resulting in a doubly-stochastic matrix, and $\\tau \\in (0, + \\infty)$ is the regularization factor. For the obtained $U_{i}$ , the $d\\times d$ matrix $U_{i}^{\\top}\\mathcal{A}_{i}U_{i}$ is a row and column reordering of $\\mathcal{A}_i$ based on the universe of nodes. Accordingly, the agreement between two adjacency matrices $\\mathcal{A}_i$ and $\\mathcal{A}_j$ , reordered according to their respective universe matching assignment matrices $U_{i}$ and $U_{j}$ , can be quantified using the Frobenius inner product namely $\\langle U_i^\\top \\mathcal{A}_iU_i,U_j^\\top \\mathcal{A}_jU_j\\rangle$ . Let $\\mathbf{U} = [U_1^\\top ,\\dots ,U_m^\\top ]^\\top \\in \\mathbb{R}^{n\\times d}$ where $n = \\sum_{i = 1}^{m}n_{i}$ . The block-diagonal multi-adjacency matrix is defined as $\\mathbf{A} = diag(\\mathcal{A}_1,\\dots ,\\mathcal{A}_m)\\in \\mathbb{R}^{n\\times n}$ . We can get the multi-matching formulation as:", + "bbox": [ + 75, + 305, + 468, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f(\\mathbf{U}):= \\max_{U_{1},\\ldots ,U_{m}}\\quad \\sum_{i,j\\in [m]}\\langle U_{i}^{\\intercal}\\mathcal{A}_{i}U_{i},U_{j}^{\\intercal}\\mathcal{A}_{j}U_{j}\\rangle \\\\ = \\max _ {\\mathbf {U} \\in \\mathbb {U}} \\operatorname {t r} \\left(\\mathbf {U} ^ {\\top} \\mathbf {A} \\mathbf {U} \\mathbf {U} ^ {\\top} \\mathbf {A} \\mathbf {U}\\right). \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 583, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It was observed that in the original multi-graph matching optimization problem, graphs are typically constructed for each individual instance, with all the nodes within a graph belonging to the same category. However, in medical imaging, constructing separate graphs for each organ would not only greatly increase the computational complexity of graph matching but also result in misalignments between graphs from different categories, significantly compromising matching accuracy. To address this limitation, we introduce a class-wise similarity matrix $\\mathbf{W}$ to incorporate class-aware label information for each node. This ensures that the nodes are correctly aligned according to their respective categories during multi-graph matching. Specifically, we define $W_{ij} = Y_iY_j^{\\top}$ and $\\mathbf{W} = [W_{ij}]_{ij} \\in \\mathbb{Z}^{(\\sum_{i=1}^{m}n_i)\\times(\\sum_{j=1}^{m}n_j)}$ . The class-aware similarity matrix $\\tilde{\\mathbf{A}} = \\mathbf{W}^{\\top}\\mathbf{A}\\mathbf{W}$ . Consequently, Eq. (6) can be transformed", + "bbox": [ + 75, + 656, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "into the following form", + "bbox": [ + 500, + 90, + 656, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\mathbf {U} \\in \\mathbb {U}} \\operatorname {t r} \\left(\\mathbf {U} ^ {\\top} \\tilde {\\mathbf {A}} \\mathbf {U} \\mathbf {U} ^ {\\top} \\tilde {\\mathbf {A}} \\mathbf {U}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 108, + 890, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The higher-order projected power iteration (HiPPI) [4] is employed to iteratively solve $\\mathbf{U}$ to obtain a stable convergence form:", + "bbox": [ + 498, + 141, + 890, + 186 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {i} ^ {*} = U _ {i} ^ {t} \\quad s. t. \\quad | | U _ {i} ^ {t} - U _ {i} ^ {t - 1} | | < \\theta , \\quad \\forall i \\in [ m ]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 189, + 890, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Empirically, we set the convergence threshold $\\theta = 10^{-5}$ . In order to guarantee the optimal learning of universe embeddings $\\mathcal{U}$ , we incorporate the non-negative matrix factorisation (NMF) [3, 23] and define the overall optimization objective as:", + "bbox": [ + 496, + 215, + 890, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (\\mathcal {U}) = \\sum_ {i \\in [ m ]} | | U _ {i} ^ {*} - \\mathcal {V} _ {i} \\mathcal {U} ^ {\\mathrm {T}} | | _ {F} ^ {2} + \\alpha | | \\mathcal {U} | | _ {F} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 295, + 890, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $||\\cdot||_F^2$ represents the Frobenius norm, and $\\alpha$ is a tuning parameter. Since only matrix multiplication and element-wise division are involved, the Sinkhorn operation is fully differentiable. This allows for effective updates to $\\mathcal{U}$ through back-propagation, enabling the incorporation of morphological prior knowledge.", + "bbox": [ + 496, + 339, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Testing Paradigm on the Target Domains", + "text_level": 1, + "bbox": [ + 498, + 439, + 849, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the Test-Time Adaptation (TTA) phase, we only have access to unlabeled data from the unseen target domains $\\mathcal{D}_t$ . Due to distribution shifts, the performance of models trained on $\\mathcal{D}_s$ can degrade significantly. To address this issue, we employ the multi-graph matching on each mini-batch $\\{x_i^t\\}_{i=1}^m \\in \\mathcal{D}_t$ . By utilizing universe embeddings $\\mathcal{U}$ learned from $\\mathcal{D}_s$ as described in Sec. 3.2, it is guaranteed cycle-consistency, enabling the network to focus on domain-invariant features guided by morphological priors from medical images. In this phase, $\\mathcal{U}$ is frozen to prevent error accumulation from affecting the prior information. The feature extractor and segmentation head (segmentation network in Fig. 2) are optimized through gradient back-propagation, allowing the model to adapt effectively to the new domains.", + "bbox": [ + 496, + 462, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unsupervised Multi-Matching. Following the same graph construction described in Sec. 3.1, we obtain the $\\mathcal{G}_i^t = (\\mathcal{V}_i^t,\\mathcal{A}_i^t)$ for each $x_{i}^{t}$ , where $\\mathcal{V}_i^t\\in \\mathbb{R}^{n_i\\times h}$ represents the $n_i$ class-aware nodes and $\\mathcal{A}_i^t\\in \\mathbb{R}^{n_i\\times n_i}$ is the adjacency matrix encoding the structural information. We first introduce a network-learned affinity matrix $M_{ij} = f_{mlp}\\{\\mathcal{V}_i^t\\mathcal{W}_x^t$ $(\\mathcal{V}_j^t\\mathcal{W}_y^t)^\\top \\} \\in \\mathbb{R}^{n_i\\times n_j}$ in graph pair $\\mathcal{G}_i$ and $\\mathcal{G}_j$ , which computes the similarity between their corresponding node features. $\\mathcal{W}_x^t$ and $\\mathcal{W}_y^t$ are two learnable linear projection, $f_{mlp}$ is a multi-layer perception (MLP). Subsequently, the matching between $\\mathcal{G}_i$ and the universe of size $d$ is obtained through the inner-product between $\\mathcal{V}_i^t$ and the pre-trained universe embeddings $\\mathcal{U}$ , in accordance with the same procedure $U_{i} = Sinkhorn(\\mathcal{V}_{i}^{t}\\mathcal{U}^{\\top},\\tau)\\in \\mathbb{U}_{n_{i}d}$ in Sec. 3.2.", + "bbox": [ + 496, + 689, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15625", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In [59, 61], the authors propose solving the multimatching KB-QAP problem presented in Eq. (4) (denoted as $J$ ) using a Taylor expansion. Based on the assumptions in Lemma 1 of Sec. 2.1, we can reformulate $J$ in Taylor series as follows:", + "bbox": [ + 76, + 90, + 468, + 165 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nJ (U) \\approx J \\left(U ^ {0}\\right) + \\sum_ {i} \\frac {\\partial J}{\\partial U _ {i}} \\bigg | _ {U = U ^ {0}} \\left(U _ {i} - U _ {i} ^ {0}\\right) + \\dots . \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 167, + 468, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$J$ can be approximated as $m$ linear optimization problems through the first-order Taylor expansion, which can be efficiently solved using the Sinkhorn. The gradient of $J$ with respect to $U_{i}$ , denoted as", + "bbox": [ + 76, + 213, + 468, + 273 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i} = \\frac {\\partial J}{\\partial U _ {i}} = \\sum_ {j \\in [ m ]} \\left(\\lambda \\mathcal {A} _ {i} U _ {i} U _ {j} ^ {\\mathrm {T}} \\mathcal {A} _ {j} U _ {j} + M _ {i j} U _ {j}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 276, + 468, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$V_{i}$ ensures that $U_{i}$ gradually converges to a high-quality discrete solution.", + "bbox": [ + 76, + 324, + 467, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the converged $\\{U_i\\}$ fine-tune the segmentation network with the matching loss $L_{mat}$ :", + "bbox": [ + 76, + 354, + 468, + 385 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {m a t} = \\sum_ {i, j \\in [ m ]} \\left[ - \\hat {M} _ {i j} ^ {\\gamma} (1 - U _ {i} U _ {j} ^ {\\mathsf {T}}) \\log (U _ {i} U _ {j} ^ {\\mathsf {T}}) \\right] \\\\ - \\sum_ {i, j \\in [ m ]} \\left[ (1 - \\hat {M} _ {i j}) ^ {\\gamma} U _ {i} U _ {j} ^ {\\mathsf {T}} \\log \\left(1 - U _ {i} U _ {j} ^ {\\mathsf {T}}\\right) \\right], \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 388, + 468, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{M}_{ij} = \\text{sinkhorn}(M_{ij},\\tau)$ , and $\\gamma$ is an amplification factor. See algorithm in appendix for full details.", + "bbox": [ + 76, + 474, + 468, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 518, + 207, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 542, + 178, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The retinal fundus segmentation datasets comprise five public datasets from different medical centers, denoted as Site A (RIM-ONE [13]), B (REFuge [37]), C (ORIGA [74]), D (REFuge-Test [37]), and E (Drishti-GS [52]), with 159, 400, 650, 800, and 101 images respectively, all consistently annotated for optic disc (OD) and optic cup (OC) segmentation. We adopted the preprocessing method outlined in [8, 28], where each image's region of interest is cropped to $800 \\times 800$ pixels and normalized using min-max normalization.", + "bbox": [ + 75, + 566, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The polyp segmentation datasets consist of four public datasets collected from different medical centers, denoted as Site A (BKAI-IGH-NEOPolyp [33]), B (CVC-ClinicDB/CVC-612 [2]), C (ETIS [50]), and D (Kvasir [18]), containing 1000, 612, 196, and 1000 images, respectively. We followed the preprocessing steps in [8], resizing the images to $800 \\times 800$ pixels and normalizing them using ImageNet-derived statistics.", + "bbox": [ + 75, + 717, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 847, + 266, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. To ensure a fair comparison across all methods, we followed the protocol in [8] and", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "split each dataset into an $8:2$ ratio for training and testing. We employed the ResNet-50 [17] pre-trained on ImageNet as the feature extractor. The training was optimized using the SGD optimizer with a momentum of 0.9 and a learning rate of 0.001. For source model training, we set the mini-batch size to 8, allowing simultaneous matching of 8 graphs per batch. The universe embedding $\\mathcal{U}$ was treated as an additional trainable tensor, integrated into the network weights without influencing subsequent experiments with other methods. During the TTA phase, all methods were trained on the target data without access to their labels. The mini-batch size was set to 4, and the previously learned $\\mathcal{U}$ from source model training was frozen. All experiments were implemented using PyTorch and conducted on 4 NVIDIA 3090 GPUs.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. The Dice score (DSC, $\\%$ ) was used to quantify the accuracy of the predicted masks, serving as the primary metric for evaluating segmentation performance. Additionally, the enhanced alignment metric $E_{\\phi}^{max}$ [12] was adopted to measure both pixel-level and global-level similarity. To further assess the consistency between predictions and ground truths, we also employed the structural similarity metric $S_{\\alpha}$ [11].", + "bbox": [ + 496, + 316, + 890, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 452, + 700, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We selected U-Net [45] as the benchmark model for the segmentation task, training it on the source domains and testing on the target domain without adaptation (No Adapt). In addition, we compared eight state-of-the-art (SOTA) methods, including two entropy-based approaches (TENT [56] and SAR [35]), a shape template-based method (TASD [28]), a dynamically adjusted learning rate method (DLTTA [66]), batch normalization-based methods (DomainAdaptor [70], VPTTA [8]), and noise estimation-based approaches (DeY-Net [62] and NC-TTT [38]).", + "bbox": [ + 496, + 476, + 890, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-Source Generalization. In these experiments, we adopted a leave-one-out training strategy [6] (i.e., with $S = |\\mathcal{D}_s \\cup \\mathcal{D}_t| - 1$ and $T = 1$ ). Although some methods, such as TASD [28], DeY-Net [62] and VPTTA [8], were originally designed for single-source domain training, we simulated single-domain conditions by mixing multi-source domain data, making the experimental setup still feasible for these approaches. Furthermore, we have also included experiments specifically focused on single-source domains later in the study. The segmentation results for the fundus and polyp datasets are shown in Tables 1 and 2, respectively. For OD/OC segmentation, all TTA methods outperformed the No Adapt baseline, highlighting their effectiveness in addressing domain shifts. Comparatively, our method consistently achieved better average performance than all other approaches. Specifically, for the key segmentation metric DSC, our method exceeded the second-best (DeY-Net [62]) by $2.88\\%$ and outperformed the No Adapt by $19.09\\%$ . A", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15626", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b9794422715c50afd8e9b9e2426c56b94f3653fc4f23c9979ba1be649db4f04f.jpg", + "table_caption": [ + "Table 1. Multi sources domain generalization in retinal fundus segmentation. The average performance (mean ± standard deviation) of three trials for our method and nine SOTA methods. \"Site A\" means training on Sites B-E and testing on Site A, and similarly for the others. The best results are highlighted in red." + ], + "table_footnote": [], + "table_body": "
MethodsSite ASite BSite CSite DSite E
\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{2}\\)\\(S_{\\text{a}}\\)
No Adapt (U-Net [45])65.60±5.7891.75±0.1379.11±0.0774.65±4.8888.78±0.0883.31±0.0263.15±7.3081.03±0.1979.77±0.0368.11±5.4989.46±0.2082.12±0.0775.34±1.0190.68±0.1687.53±0.05
TENT (ICLR'21) [56]75.22±3.9994.10±0.1083.97±0.0581.06±0.9393.45±0.1192.36±0.0375.23±2.3384.62±0.1478.12±2.5991.33±0.3186.41±0.1283.17±0.7892.74±0.1093.64±0.0378.5691.9488.20
TASD (AAAI'22) [28]79.70±2.5288.73±0.4583.67±0.1782.96±2.1490.01±0.2090.99±0.1186.12±1.1193.59±0.1387.08±0.0778.02±6.3087.20±0.1687.94±0.0483.71±3.7790.88±0.2590.37±0.1282.1090.8888.01
DLTTA (TMI'22) [66]71.88±5.5986.29±0.2180.70±0.1380.12±2.2690.53±0.1788.63±0.0681.46±3.3593.08±0.0887.48±0.0275.90±7.2784.27±0.1284.79±0.0682.75±6.5590.10±0.3189.33±0.1478.4288.8586.18
SAR (ICLR'23) [35]75.29±2.6988.73±0.3480.15±0.2686.42±1.5796.88±0.2791.28±0.0980.69±4.3092.10±0.1586.31±0.2780.27±5.4789.98±0.3187.20±0.1589.61±0.2895.90±0.0892.06±0.0182.4692.7187.40
DomainAdaptor (CVPR'23) [70]77.14±1.2190.80±0.1981.31±0.0283.95±0.8093.17±0.0990.44±0.0882.79±1.0995.36±0.1088.57±0.1386.50±3.2995.26±0.3391.07±0.2086.24±1.4494.90±0.2993.77±0.1883.3293.9089.03
DeY-Net (WACV'24) [62]79.67±1.3589.14±0.2383.68±0.0490.07±3.1290.06±0.3392.31±0.0785.94±1.8593.54±0.2287.15±0.1083.47±1.4095.29±0.1788.70±0.0488.76±2.2895.43±0.2190.45±0.0885.5894.0988.46
VPTTA (CVPR'24) [8]77.85±1.4094.09±0.0183.97±0.0584.57±0.9995.38±0.0389.53±0.0983.61±1.6796.39±0.0489.64±0.0582.75±0.5096.53±0.0489.83±0.0189.74±1.5198.22±0.2494.77±0.0783.7096.1289.55
NC-TTT (CVPR'24) [38]80.83±1.7789.61±0.2384.41±0.0887.07±1.3195.63±0.0690.09±0.0286.93±2.2794.69±0.1188.39±0.0382.10±4.2891.49±0.2286.33±0.0790.08±0.0995.93±0.1091.53±0.0985.4093.4788.15
Ours82.53±1.5291.85±0.2285.50±0.2088.98±0.8996.69±0.0491.63±0.1288.73±0.7095.80±0.0689.71±0.0290.25±1.3397.97±0.1992.89±0.1391.83±0.7197.14±0.2092.82±0.0988.4695.8490.51
", + "bbox": [ + 80, + 136, + 883, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/381617be161b999c102842e331550a7610714ee6152ae8c0ba74282e4ffc9582.jpg", + "table_caption": [ + "Table 2. Multi sources domain generalization in the polyp segmentation. The average performance (mean ± standard deviation) of three trials for our method and nine SOTA methods. \"Site A\" means training on Sites B-D and testing on Site A, and similarly for the others. The best results are highlighted in red." + ], + "table_footnote": [], + "table_body": "
MethodsSite ASite BSite CSite DAverage
\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\uparrow\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)
No Adapt (U-Net [45])82.67±2.2492.11±0.0789.59±0.1081.20±4.3892.46±0.2086.08±0.0678.33±1.6294.00±0.0885.35±0.0175.51±2.7485.90±0.1781.71±0.2179.4391.1285.68
TENT (ICLR'21) [56]78.69±1.1288.28±0.1485.44±0.0877.71±3.0989.73±0.0981.90±0.1280.19±3.3596.03±0.0487.80±0.1772.63±5.1182.47±0.3078.92±0.1477.3189.1383.52
TASD (AAAI'22) [28]84.60±4.1492.55±0.2190.76±0.1085.28±2.2093.10±0.1386.83±0.1982.43±5.3996.31±0.0889.61±0.0280.04±4.4188.26±0.1184.09±0.0383.0992.5687.82
DLTTA (TMI'22) [66]80.68±2.3790.01±0.2186.89±0.0980.86±1.7791.98±0.1085.83±0.0475.57±3.1293.90±0.0884.15±0.1074.25±3.4083.49±0.1279.44±0.1077.8489.8584.08
SAR (ICLR'23) [35]79.37±0.7788.40±0.1286.11±0.0378.40±2.0790.23±0.1183.39±0.0280.94±0.9095.83±0.1286.84±0.0778.23±2.9186.37±0.0981.66±0.1479.2490.2184.50
DomainAdaptor (CVPR'23) [70]88.58±0.9695.03±0.0490.95±0.0181.12±1.0792.31±0.0886.60±0.0281.77±2.1795.18±0.0387.19±0.0179.91±1.3387.23±0.0483.64±0.0582.8592.4487.10
DeY-Net (WACV'24) [62]83.61±2.3190.79±0.1189.25±0.0980.90±3.1791.49±0.1085.36±0.0380.46±1.4395.83±0.0887.14±0.0373.85±3.0483.00±0.1378.96±0.0879.7190.2885.18
VPTTA (CVPR'24) [8]82.34±0.6991.03±0.0888.35±0.0184.32±0.4491.53±0.0787.46±0.0284.43±0.5295.75±0.1688.91±0.0382.40±0.2989.60±0.0485.18±0.0183.3791.9787.47
NC-TTT (CVPR'24) [38]87.96±1.2293.71±0.1390.66±0.0783.70±2.0893.78±0.0987.51±0.0789.29±0.9096.66±0.2291.26±0.0778.23±0.7986.37±0.0981.66±0.0684.8092.6387.77
Ours90.97±0.6696.57±0.0692.92±0.0887.49±0.8295.21±0.0489.93±0.0986.87±0.5897.53±0.0590.16±0.1283.97±1.1890.67±0.1386.44±0.0887.3294.6789.86
", + "bbox": [ + 80, + 276, + 883, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "similar trend was observed in the polyp segmentation results, where we outperformed the second-best method (NC-TTT [38]) and No Adapt by $2.52\\%$ and $7.89\\%$ , respectively. Additionally, our approach exhibited greater stability across multiple trials compared to the other methods.", + "bbox": [ + 75, + 402, + 468, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Single-Source Generalization. In these experiments, the models are trained on one domain and tested on the remaining datasets (i.e., with $S = 1$ and $T = |\\mathcal{D}_s \\cup \\mathcal{D}_t| - 1$ ). Compared to multi-source generalization, single-source training is considered more challenging due to the limited domain information available during the training phase. By comparing Tables 2 and 3, we can observe that all models show lower segmentation performance (in terms of DSC) in the single-source setting compared to the multi-source results. Some methods even perform worse than the No Adapt baseline in single-source scenarios. However, our method still achieves the best average performance across 12 domain transfer experiments, outperforming the second-best approach (NC-TTT [38]) by $1.83\\%$ and exceeding the No Adapt by $9.77\\%$ in terms of DSC.", + "bbox": [ + 75, + 478, + 468, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3 presents the visualization results for \"Site A\" in Table 1. The No Adapt baseline exhibits significant misalignment, with notably distorted shapes and inaccurate boundaries. VPTTA offers some improvement but still suffers from distortions and incomplete segmentation, particularly in the optic cup area. NC-TTT shows instances of overlapping segmentation. In contrast, our method delivers the most precise segmentation, closely aligning with the ground truth. The Grad-CAM [48] visualizations reinforce this, as the attention of network is focused specifically on the relevant regions. These results demonstrate our method's enhanced generalization to unseen domains by effectively incorporating morphological priors, resulting in higher seg", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba7ca63d2bcda0fe288ae5adeb346a5e1ba7f633574dc4083f220e2b87ae6eb5.jpg", + "image_caption": [ + "Figure 3. Visualization comparison of segmentation results and Grad-CAM outputs from the final layer of the backbone network for the No Adapt baseline, VPTTA [8], NC-TTT [38], and our proposed method on retinal fundus images. Additional visual comparisons are provided in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 503, + 400, + 890, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mentation accuracy. For further visualization and additional experimental results, please refer to the appendix.", + "bbox": [ + 498, + 618, + 890, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 500, + 667, + 596, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Effectiveness of the Universe Embeddings", + "text_level": 1, + "bbox": [ + 500, + 694, + 854, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The universe embeddings $\\mathcal{U}$ play a crucial role in incorporating morphological priors from medical images, resulting in an assignment matrix that projects each node to the universe of nodes. To validate the effectiveness during the TTA, we conducted the following experiments: (1) Without universe embeddings (denoted as $w / o\\mathcal{U}$ ), the universe matching assignment matrix $\\mathbf{U}$ is initialized following the setting in [59] (i.e. $\\mathbf{U} = 1 / d + 10^{-3}z$ , where $z\\sim N(0,1)$ ), leading to random matching between the nodes and the universe; (2) Without morphological priors but with $\\mathcal{U}$ (denoted as $w / o$ priors). As shown in Table 4, the performance in (1) is slightly better than in (2) by $1.76\\%$ in terms of DSC", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15627", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e407af8022473af55059761f17cb4436b7e568e64f90b863695acf92a9b3df80.jpg", + "table_caption": [ + "Table 3. Single source domain generalization in the polyp segmentation. The average performance of three trials for our method and nine SOTA methods. A → B represents models trained on Site A and tested on Site B, and similar for others. Best results are colored as red." + ], + "table_footnote": [], + "table_body": "
MethodsA→BA→CA→DB→AB→CB→DC→AC→BC→DD→AD→BD→CAvg.
Dice Score Metric ↑ (DSC, mean±std)
No Adapt (U-Net [45])76.49±1.4867.70±3.2770.94±7.5676.21±2.9062.80±8.0370.03±4.5372.87±3.6269.13±5.1570.44±3.2779.78±1.0872.47±2.8074.91±3.1871.98
TENT (ICLR'21) [56]68.13±0.1865.99±0.1166.11±0.1477.56±0.2159.87±0.2068.89±0.1473.21±0.1065.04±0.0972.35±0.2277.88±0.1470.01±0.1173.60±0.2069.89
TASD (AAAI'22) [28]73.02±2.2479.67±6.3778.23±4.4466.82±9.1772.09±4.4077.95±1.6974.20±3.3069.74±7.2776.13±2.8283.77±0.5176.09±3.2680.10±1.4775.65
DLTTA (TMI'22) [66]72.23±0.0470.78±0.0674.56±0.0464.87±0.0160.05±0.0978.47±0.0371.22±0.0365.78±0.1071.00±0.1379.89±0.0879.10±0.1074.52±0.0171.87
SAR (ICLR'23) [35]70.34±0.1377.90±0.2077.71±0.0968.19±0.1266.56±0.0873.92±0.1376.12±0.1171.29±0.1975.08±0.0882.07±0.0480.44±0.2183.10±0.1575.22
DomainAdaptor (CVPR'23) [70]78.39±0.6177.09±0.3579.45±0.2174.98±0.7070.07±0.6680.21±0.4676.02±0.3373.40±0.5081.33±0.5486.10±0.4380.56±0.6277.86±0.2177.96
DeY-Net (WACV'24) [62]74.45±1.3176.07±2.5280.11±0.7270.90±1.1069.23±2.1876.31±1.3278.33±4.4374.60±1.6673.52±2.2580.70±0.2778.23±1.1781.48±0.9376.16
VPTTA (CVPR'24) [8]78.18±0.0175.14±0.0382.73±0.0273.65±0.0664.78±0.0182.86±0.0377.90±0.0170.64±0.0577.22±0.1087.19±0.0384.97±0.0882.12±0.0278.12
NC-TTT (CVPR'24) [38]79.21±0.2481.76±0.5783.22±0.6977.35±0.3379.40±0.6780.94±0.5480.17±0.7072.36±0.6278.55±0.4384.09±0.3882.38±0.5179.77±0.7479.93
Ours80.47±0.2883.93±0.1983.18±0.3378.33±0.2681.25±0.2383.58±0.3478.86±0.2676.49±0.2180.20±0.2489.17±0.2281.68±0.2283.86±0.2881.75
", + "bbox": [ + 78, + 116, + 883, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3f8e850d76b86a6f9425eb75b7370a0a1c3d506b29df99986eea594711b7768c.jpg", + "table_caption": [ + "Table 4. Ablation study in the retinal fundus segmentation. Details of the experiment can be found in Sec 5.1." + ], + "table_footnote": [], + "table_body": "
MethodsSite ASite BSite CSite DSite EAverage
Dice Score Metric ↑ (DSC, mean±std)
No Adapt65.60±5.7874.65±4.8863.15±7.3068.11±5.4975.34±1.0169.37
w/o U77.09±1.9083.24±2.0180.50±1.4478.32±1.5583.18±2.1880.47
w/o priors79.51±0.8079.36±1.1377.40±1.8977.29±1.7780.00±1.9378.71
Ours82.53±1.5288.98±0.8988.73±0.7090.25±1.3391.83±0.7188.46
", + "bbox": [ + 78, + 277, + 467, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff352ef9cde431b944aa81e57fbb43239ac968d87e3f0fdf79b362279d09c376.jpg", + "image_caption": [ + "Figure 4. The performance of DSC (\\%) and inference time (s) with different universe sizes $d$ are shown, with the experimental setup identical to that in Table 1. The first curve represents the No Adapt baseline ( $d = 0$ ), while the subsequent curve corresponds to the results obtained with a universe size of $d = 120$ ." + ], + "image_footnote": [], + "bbox": [ + 84, + 357, + 263, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de7f794bac43d8fe40e35f5601014314a8675039b1a17e45c3ba77bd5f018919.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 357, + 464, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on average. However, when using the pre-trained $\\mathcal{U}$ derived from the source model, the segmentation results show a significant improvement. This highlights the effectiveness of the pre-trained morphological priors embedded in $\\mathcal{U}$ for medical imaging tasks.", + "bbox": [ + 75, + 571, + 468, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Effectiveness of the Universe Size", + "text_level": 1, + "bbox": [ + 76, + 650, + 367, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The universe size $d$ directly impacts both the matching accuracy and computational efficiency, and its selection depends on the number of sampled nodes. Previous works [4, 36, 59], ensured that the $d$ matched the number of nodes in each graph by labeling an equal number of keypoints. In our experiments, $d$ is determined by the spatially-uniform sampling described in Sec 3.1, and we empirically set $d = 100 \\times (n + 1)s^{-1}$ , where $s$ is the sampling step and $n$ is the number of segmentation categories required for the task. We evaluated several $d$ , with the results shown in Fig. 4. We observe that excessively small or large step sizes—resulting in too many or too few sampled nodes—negatively impact both training speed and performance. However, when $d$ is within a reasonable range (in this experiment, $s \\in [2,10]$ i.e. $d \\in [30,150]$ ), the model maintains good scalability", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and performance.", + "text_level": 1, + "bbox": [ + 500, + 251, + 617, + 265 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/13569d79fee7530caa9328bbc367e0a58ee4166e65aedc6ce8ee67681dac8e11.jpg", + "table_caption": [ + "Table 5. Comparison of common variables and performance between pairwise matching and multi-matching, using the same experimental setup as in Table 1. Both methods follow the graph construction in Sec. 3.1." + ], + "table_footnote": [], + "table_body": "
paramSite A-E Averagedescription
Pairwise [58]Multi-Matching (Ours)
lr10-310-3learning rate
batch44batch size in inference
h256256dimension of node feature
τ0.050.05regularization factor of Sinkhorn
Iter2020max iterations of Sinkhorn
d-120universe size
DSC ↑84.6388.46dice score metric (%)
time ↓0.9300.392inference time per image (s/img)
Param ↓1.0710.658parameter count (M)
FLOPs ↓15.354.255floating point operations per second (G)
", + "bbox": [ + 503, + 321, + 883, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Multi-Matching vs. Pairwise Matching", + "text_level": 1, + "bbox": [ + 500, + 433, + 834, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared to pairwise matching, multi-matching uses cycle-consistency for global optimization, which helps avoid local optima in pairwise. Moreover, joint optimization reduces the computational complexity of large-scale matching. By incorporating morphological consistency constraints, multi-matching exhibits greater robustness. To validate these claims, we conducted experiments comparing the two approaches. For the pairwise matching method, we adopted the benchmark approach [58] and applied the same graph construction process as in multi-matching. The comparison results are presented in Table 5. It is evident that multi-matching outperforms pairwise matching in terms of DSC, inference time, parameters and FLOPs. The inference time of multi-matching was reduced by approximately $57.85\\%$ compared to pairwise matching, while segmentation accuracy improved by $3.83\\%$ in DSC.", + "bbox": [ + 496, + 455, + 890, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 709, + 617, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a novel multi-matching framework for TTDG, which leverages universe embeddings to incorporate morphological priors from medical images while ensuring cycle-consistency, along with an unsupervised test-time paradigm that fully integrates these priors for efficient adaptation. Extensive experiments, including comparisons and ablation studies demonstrate that our graph-matching-based approach achieves SOTA performance, outperforming entropy-based, template-based, batch normalization-based, and noise estimation-based methods across both multi-source and single-source domain generalization tasks.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15628", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (Grant No. 62306003), the Open Research Fund from Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), under Grant No. GML-KF-24-29 and the Open Foundation of Jiangxi Provincial Key Laboratory of Image Processing and Pattern Recognition (ET202404437).", + "bbox": [ + 76, + 114, + 468, + 213 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 243, + 171, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1", + "[2] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized Medical Imaging and Graphics, 43:99-111, 2015. 6", + "[3] Florian Bernard, Johan Thunberg, Jorge Goncalves, and Christian Theobalt. Synchronisation of partial multi-matchings via non-negative factorisations. Pattern Recognition, 92:146-155, 2019. 2, 5", + "[4] Florian Bernard, Johan Thunberg, Paul Swoboda, and Christian Theobalt. Hippi: Higher-order projected power iterations for scalable multi-matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 10284-10293, 2019. 3, 5, 8", + "[5] Yuan Bi, Zhongliang Jiang, Ricarda Clarenbach, Ghotbi, and Nassir Navab. Mi-segnet: Mutual information-based us segmentation for unseen domain generalization. In MICCAI, pages 130-140. Springer, 2023. 3", + "[6] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 24172-24182, 2023. 3, 6", + "[7] Yuxin Chen, Leonidas Guibas, and Qixing Huang. Near-optimal joint object matching via convex relaxation. In ICML, pages II-100, 2014. 3", + "[8] Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each test image deserves a specific prompt: Continual test-time adaptation for 2d medical image segmentation. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 11184-11193, 2024. 3, 6, 7, 8", + "[9] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Adv. Neural Inform. Process. Syst., 26, 2013. 5", + "[10] Frank De Groot. High-resolution x-ray emission and x-ray absorption spectroscopy. Chemical Reviews, 101(6):1779-1808, 2001. 1", + "[11] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4548-4557, 2017. 6" + ], + "bbox": [ + 78, + 267, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. *IJCAI*, 2018. 6", + "[13] Francisco Fumero, Silvia Alayón, José L Sanchez, Jose Sigut, and M Gonzalez-Hernandez. Rim-one: An open retinal image database for optic nerve evaluation. In International Symposium on Computer-based Medical Systems (CBMS), pages 1-6. IEEE, 2011. 6", + "[14] Quankai Gao, Fudong Wang, Nan Xue, Jin-Gang Yu, and Gui-Song Xia. Deep graph matching under quadratic constraint. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5069-5078, 2021. 4", + "[15] Steven Gold, Anand Rangarajan, et al. Softmax to softassign: Neural network algorithms for combinatorial optimization. Journal of Artificial Neural Networks, 2(4):381-399, 1996. 5", + "[16] Ran Gu, Guotai Wang, Jiangshan Lu, Jingyang Zhang, Wenhui Lei, Yinan Chen, Wenjun Liao, Shichuan Zhang, Kang Li, Dimitris N Metaxas, et al. Cddsa: Contrastive domain disentanglement and style augmentation for generalizable medical image segmentation. Medical Image Analysis, 89: 102904, 2023. 3", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 770-778, 2016. 4, 6", + "[18] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, pages 451-462. Springer, 2020. 6", + "[19] Neerav Karani, Ertunc Erdil, Krishna Chaitanya, and Ender Konukoglu. Test-time adaptable neural networks for robust medical image segmentation. Medical Image Analysis, 68: 101907, 2021. 3", + "[20] Boah Kim, Inhwa Han, and Jong Chul Ye. Diffuse morph: Unsupervised deformable image registration using diffusion model. In Eur. Conf. Comput. Vis., pages 347-364. Springer, 2022. 2", + "[21] Tjalling C Koopmans and Martin Beckmann. Assignment problems and the location of economic activities. *Econometrica: journal of the Econometric Society*, pages 53–76, 1957. 3", + "[22] Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Kulkarni, Hiran Sarkar, Varun Jampani, and R Venkatesh Babu. Concurrent subsidiary supervision for unsupervised source-free domain adaptation. In ECCV, pages 177-194. Springer, 2022. 3", + "[23] Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. nature, 401 (6755):788-791, 1999. 5", + "[24] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5400-5409, 2018. 3", + "[25] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semanticcomplete graph matching for domain adaptive object detec" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15629", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5291-5300, 2022. 4", + "[26] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):9022-9040, 2023. 2", + "[27] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. Int. J. Comput. Vis., pages 1-34, 2024. 1, 3", + "[28] Quande Liu, Cheng Chen, Qi Dou, and Pheng-Ann Heng. Single-domain generalization in medical image segmentation via test-time adaptation from shape dictionary. In AAAI, pages 1756-1764, 2022. 2, 3, 6, 7, 8", + "[29] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1", + "[30] Eleonora Maset, Federica Arrigoni, and Andrea Fusiello. Practical and efficient multi-view matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4568-4576, 2017. 2, 3", + "[31] Jose G Moreno-Torres, Troy Raeder, Rocio Alaiz-Rodriguez, Nitesh V Chawla, and Francisco Herrera. A unifying view on dataset shift in classification. Pattern recognition, 45(1):521-530, 2012. 1", + "[32] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 3", + "[33] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Advances in visual Computing: 16th International Symposium (ISVC), pages 15-28. Springer, 2021. 6", + "[34] Andy Nguyen, Mirela Ben-Chen, Katarzyna Welnicka, Yinyu Ye, and Leonidas Guibas. An optimization approach to improving collections of shape maps. In Computer Graphics Forum, pages 1481-1491. Wiley Online Library, 2011. 2", + "[35] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. ICLR, 2023. 6, 7, 8", + "[36] Zhakshylyk Nurlanov, Frank R Schmidt, and Florian Bernard. Universe points representation learning for partial multi-graph matching. In AAAI, pages 1984-1992, 2023. 3, 8", + "[37] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel Van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, et al. Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Medical Image Analysis, 59:101570, 2020. 6", + "[38] David Osowiechi, Gustavo A Vargas Hakim, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdanpanah, Ismail Ben Ayed, and Christian Desrosiers. Nc-ttt: A" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "noise constrastive approach for test-time training. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 6078-6086, 2024. 6, 7, 8", + "[39] Deepti Pachauri, Risi Kondor, and Vikas Singh. Solving the multi-way matching problem by permutation synchronization. Adv. Neural Inform. Process. Syst., 26, 2013. 2, 3", + "[40] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 1406-1415, 2019. 3", + "[41] Bin Pu, Xingguo Lv, Jiewen Yang, He Guannan, Xingbo Dong, Yiqun Lin, Li Shengli, Tan Ying, Liu Fei, Ming Chen, Zhe Jin, Kenli Li, and Xiaomeng Li. Unsupervised domain adaptation for anatomical structure detection in ultrasound images. In ICML, pages 41204-41220. PMLR, 2024. 2, 3", + "[42] Bin Pu, Liwen Wang, Jiewen Yang, Guannan He, Xingbo Dong, Shengli Li, Ying Tan, Ming Chen, Zhe Jin, Kenli Li, et al. M3-uda: A new benchmark for unsupervised domain adaptive fetal cardiac structure detection. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 11621-11630, 2024. 2", + "[43] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 12556-12565, 2020. 3", + "[44] Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Droppedge: Towards deep graph convolutional networks on node classification. In ICLR, 2020. 4", + "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 6, 7, 8", + "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 4938-4947, 2020. 2", + "[47] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Adv. Neural Inform. Process. Syst., 33:11539-11551, 2020. 3", + "[48] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 618-626, 2017. 7", + "[49] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Adv. Neural Inform. Process. Syst., 35: 14274-14289, 2022. 3", + "[50] Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer. International Journal of Computer Assisted Radiology and Surgery, 9:283-293, 2014. 6", + "[51] Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices. The annals of mathematical statistics, 35(2):876-879, 1964. 5", + "[52] Jayanthi Sivaswamy, SR Krishnadas, Gopal Datt Joshi, Madhulika Jain, and A Ujjwaft Syed Tabish. Drishti-gs: Retinal" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15630", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image dataset for optic nerve head (onh) segmentation. In International Symposium on Biomedical Imaging (ISBI), pages 53-56. IEEE, 2014. 6", + "[53] Charles P Slichter. Principles of magnetic resonance. Springer Science & Business Media, 2013. 2", + "[54] Gail Ter Haar. Therapeutic applications of ultrasound. Progress in biophysics and molecular biology, 93(1-3):111-129, 2007. 2", + "[55] Roberto Tron, Xiaowei Zhou, Carlos Esteves, and Kostas Daniilidis. Fast multi-image matching via density-based clustering. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4057-4066, 2017. 2, 3", + "[56] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 6, 7, 8", + "[57] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 1, 3", + "[58] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 3056-3065, 2019. 4, 8", + "[59] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Graduated assignment for joint multi-graph matching and clustering with application to unsupervised graph matching network learning. Adv. Neural Inform. Process. Syst., 33: 19908-19919, 2020. 3, 5, 6, 7, 8", + "[60] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Neural graph matching network: Learning lawler's quadratic assignment problem with extension to hypergraph and multiple-graph matching. IEEE Trans. Pattern Anal. Mach. Intell., 44 (9):5261-5279, 2021. 4", + "[61] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Unsupervised learning of graph matching with mixture of modes via discrepancy minimization. IEEE Trans. Pattern Anal. Mach. Intell., 45(8):10500-10518, 2023. 3, 6", + "[62] Ruxue Wen, Hangjie Yuan, Dong Ni, Wenbo Xiao, and Yaoyao Wu. From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In Proc. IEEE/CVF Winter Conference on Applications of Computer Vision, pages 464-474, 2024. 6, 7, 8", + "[63] Junchi Yan, Yu Tian, Hongyuan Zha, Xiaokang Yang, Ya Zhang, and Stephen M Chu. Joint optimization for consistent multiple graph matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 1649-1656, 2013. 2", + "[64] Junchi Yan, Yin Li, Wei Liu, Hongyuan Zha, Xiaokang Yang, and Stephen Mingyu Chu. Graduated consistency-regularized optimization for multi-graph matching. In ECCV, pages 407–422. Springer, 2014. 2", + "[65] Junchi Yan, Minsu Cho, Hongyuan Zha, Xiaokang Yang, and Stephen M Chu. Multi-graph matching via affinity optimization with graduated consistency regularization. IEEE Trans. Pattern Anal. Mach. Intell., 38(6):1228-1242, 2015. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Transactions on Medical Imaging, 41 (12):3575-3586, 2022. 6, 7, 8", + "[67] Jiewen Yang, Xinpeng Ding, Ziyang Zheng, Xiaowei Xu, and Xiaomeng Li. Graphecho: Graph-driven unsupervised domain adaptation for echocardiogram video segmentation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 11878-11887, 2023. 2", + "[68] Jee Seok Yoon, Kwanseok Oh, Yooseung Shin, Maciej A Mazurowski, and Heung-II Suk. Domain generalization for medical image analysis: A survey. arXiv preprint arXiv:2310.08598, 2023. 1, 3", + "[69] Weiyi Yu, Zhizhong Huang, Junping Zhang, and Hongming Shan. San-net: Learning generalization to unseen sites for stroke lesion segmentation with self-adaptive normalization. Computers in Biology and Medicine, 156:106717, 2023. 3", + "[70] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 18971-18981, 2023. 3, 6, 7, 8", + "[71] Ling Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Holger Roth, Andriy Myronenko, Daguang Xu, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation. IEEE transactions on medical imaging, 39(7):2531-2540, 2020. 3", + "[72] Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. Adv. Neural Inform. Process. Syst., 34:23664-23678, 2021. 1", + "[73] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Adv. Neural Inform. Process. Syst., 35:38629-38642, 2022. 3", + "[74] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. Origa-light: An online retinal fundus image database for glaucoma analysis and research. In 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology, pages 3065-3068. IEEE, 2010. 6", + "[75] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(4):4396-4415, 2022. 1, 3", + "[76] Xiaowei Zhou, Menglong Zhu, and Kostas Daniilidis. Multi-image matching via fast alternating minimization. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4032-4040, 2015. 3", + "[77] Yuliang Zou, Zizhao Zhang, Chun-Liang Li, Han Zhang, Tomas Pfister, and Jia-Bin Huang. Learning instance-specific adaptation for cross-domain segmentation. In ECCV, pages 459-476. Springer, 2022. 3", + "[78] Yukun Zuo, Hantao Yao, and Changsheng Xu. Attention-based multi-source domain adaptation. IEEE Transactions on Image Processing, 30:3793-3803, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 882 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15631", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_model.json b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8114f6cbca1e75d2b71a0af151c4bbe18648015b --- /dev/null +++ b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Test-Time Domain Generalization via Universe Learning: A Multi-Graph Matching Approach for Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.203, + 0.778, + 0.24 + ], + "angle": 0, + "content": "Xingguo Lv\\(^{1,2}\\) Xingbo Dong\\(^{1\\dagger}\\) Liwen Wang\\(^{1}\\) Jiewen Yang\\(^{3}\\) Lei Zhao\\(^{2}\\) Bin \\(\\mathrm{Pu}^{3\\dagger}\\) Zhe Jin\\(^{1}\\) Xuejun Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.24, + 0.892, + 0.276 + ], + "angle": 0, + "content": "1 Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, Anhui University. 2 Hunan University. 3 The Hong Kong University of Science and Technology." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.278, + 0.748, + 0.292 + ], + "angle": 0, + "content": "lvxg@stu.ahu.edu.cn, xingbo.dong@ahu.edu.cn, eebinpu@ust.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.707 + ], + "angle": 0, + "content": "Despite domain generalization (DG) has significantly addressed the performance degradation of pre-trained models caused by domain shifts, it often falls short in real-world deployment. Test-time adaptation (TTA), which adjusts a learned model using unlabeled test data, presents a promising solution. However, most existing TTA methods struggle to deliver strong performance in medical image segmentation, primarily because they overlook the crucial prior knowledge inherent to medical images. To address this challenge, we incorporate morphological information and propose a framework based on multi-graph matching. Specifically, we introduce learnable universe embeddings that integrate morphological priors during multi-source training, along with novel unsupervised test-time paradigms for domain adaptation. This approach guarantees cycle-consistency in multi-matching while enabling the model to more effectively capture the invariant priors of unseen data, significantly mitigating the effects of domain shifts. Extensive experiments demonstrate that our method outperforms other state-of-the-art approaches on two medical image segmentation benchmarks for both multi-source and single-source domain generalization tasks. The source code is available at https://github.com/Yore0/TTDG-MGM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.748, + 0.21, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.47, + 0.88 + ], + "angle": 0, + "content": "In the field of medical image processing, semantic segmentation is a technique that enables the precise identification and quantitative analysis of target regions. Currently, a considerable number of medical models can effectively perform segmentation tasks [1, 29]. Nevertheless, when pretrained on source datasets, their performance often declines in real-world deployment due to domain shifts [31, 72] caused" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.328, + 0.705, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.457, + 0.604, + 0.468 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.329, + 0.887, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.457, + 0.8, + 0.469 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.47, + 0.657, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.47, + 0.729, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.511, + 0.751, + 0.522 + ], + "angle": 0, + "content": "X" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.49, + 0.884, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.561, + 0.702, + 0.572 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.588, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Figure 1. Illustration of multi-graph matching using the universe of nodes. (a) Direct pairwise matching between graphs 1, 2, and 3. (b) Multi-graph matching with the universe of nodes. (c) Equivalence between (a) and (b) (see Lemma 1 in Sec. 2.1 for details)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.893, + 0.87 + ], + "angle": 0, + "content": "by differences in imaging devices, protocols, patient demographics, and preprocessing methods, which impact the model's generalization ability. To address domain shifts, previous researches in domain adaptation, such as domain generalization (DG) [57, 68, 75], have focused on designing sophisticated models that train jointly on source and target domains or across multiple styled domains. Despite the progress, these methods fall short in practice, as no source domains can fully capture all real-world variations, and retraining for each new data is impractical. Consequently, a more practical approach is to adaptively fine-tune pre-train model using the information from the unlabeled unseen data during the test, which is referred to test-time adaptation (TTA) [27]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Medical images, due to their reliance on specific physical principles (such as X-ray absorption [10], magnetic reso" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.228, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15621" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "nance [53], sound wave reflection [54], etc.), are fundamentally different from images capturing visible light in natural scenes. Moreover, the physiological characteristics of specific anatomical regions exhibit relatively stable shapes and spatial layouts, and the morphology of these structures is predictable [41, 42]. These prior insights provide valuable information that is particularly useful in medical imaging tasks. In areas such as segmentation [28, 67], registration [20], and detection [41, 42], many well-established approaches have been proposed that combine above handcrafted designs with neural network-based methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.56 + ], + "angle": 0, + "content": "We combine the aforementioned TTA strategies with the prior knowledge of medical images and design a novel Test-Time Domain Generalization (TTDG) framework utilizing graph structures. The construction of graph enables a more robust representation of the morphological priors inherent to medical images. In complex medical scenarios, the aggregation mechanism of nodes and edges in graph ensures the effective preservation of crucial contextual information. Previous research has utilised pairwise graph matching [26, 42, 46] to address domain shifts in medical imaging. However, these methods are limited to aligning features between two domains, which poses challenges in scenarios involving multiple domains commonly encountered in real-world settings. Multi-graph matching overcomes these limitations by establishing connections between multiple graphs and integrating information from diverse domains, thereby enabling the capture of more intricate patterns and structural relationships while maintaining cross-domain consistency. This process facilitates the acquisition of global invariant features from medical images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.473, + 0.789 + ], + "angle": 0, + "content": "In multi-graph matching, the universe of nodes [39, 55] represents the full-set of the nodes from all graphs (refer to Fig. 1). The aforementioned morphological priors are embedded into this universe, derived from a multitude of hospitals, devices, and modalities. It offers three main advantages: (1) Cross-domain consistency. By mapping the nodes from different graphs into the universe, we ensure that all graphs are aligned according to the same reference standard, i.e. the priors. (2) Cycle-consistency in multi-graph matching. This guarantees global coherence across multiple graphs, avoiding conflicts in local node alignments. (3) Handling partial matching and missing nodes. Even when certain nodes are absent in some graphs, matching them to virtual nodes in the universe addresses the issue, enhancing robustness to incomplete graph structures." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.79, + 0.422, + 0.804 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.81, + 0.47, + 0.886 + ], + "angle": 0, + "content": "- We propose the first (to our best knowledge) multi-graph matching framework for test-time domain generalization in medical image segmentation tasks, effectively mitigating performance degradation caused by domain shifts during the testing phase." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- By designing learnable universe embeddings, we inte" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.894, + 0.166 + ], + "angle": 0, + "content": "grate the morphological priors of medical images into the graph matching process during source training, while ensuring the cycle-consistency constraint. This approach enables joint optimization and promotes the learning of domain-invariant features." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.227 + ], + "angle": 0, + "content": "- We design a novel, well-initialized unsupervised testing adaptation paradigm that integrates prior knowledge and allows seamless deployment during adaptation, effectively addressing domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.228, + 0.892, + 0.288 + ], + "angle": 0, + "content": "- We conduct extensive experiments on two typical medical datasets, demonstrating that our method performs competitively against state-of-the-art approaches for both multi-source and single-source domain generalization." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.303, + 0.769, + 0.32 + ], + "angle": 0, + "content": "2. Preliminary & Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.894, + 0.403 + ], + "angle": 0, + "content": "The following section will introduce the preliminaries of multi-graph matching, while also reviewing the most relevant existing literature on the subject. Subsequently, domain generalization and test-time adaptation will be discussed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.414, + 0.715, + 0.43 + ], + "angle": 0, + "content": "2.1. Multi-Graph Matching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.545 + ], + "angle": 0, + "content": "We consider \\(m \\in \\mathbb{N}\\) different graphs \\(\\mathcal{G}_1, \\mathcal{G}_2, \\dots, \\mathcal{G}_m\\), \\(i \\in [m]\\), \\([m] := \\{1, \\dots, m\\}\\). For each \\(\\mathcal{G}_i = (\\mathcal{V}_i, \\mathcal{A}_i)\\), \\(\\mathcal{V}_i \\in \\mathbb{R}^{n_i \\times h}\\) represents the \\(h\\)-dimensional feature of \\(n_i\\) nodes, and adjacency matrix \\(\\mathcal{A}_i \\in \\mathbb{R}^{n_i \\times n_i}\\) encodes the connectivity between nodes, represented as the set of edges. For any two graphs \\(\\mathcal{G}_i\\) and \\(\\mathcal{G}_j\\), the set of \\(n_i \\times n_j\\) partial permutation matrices \\(\\mathbb{P}_{n_in_j}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.557, + 0.899, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} _ {n _ {i} n _ {j}} := \\left\\{\\mathbf {X} \\in \\{0, 1 \\} ^ {n _ {i} \\times n _ {j}}: \\mathbf {X} \\mathbf {1} _ {n _ {j}} \\leq \\mathbf {1} _ {n _ {i}}, \\mathbf {X} ^ {\\top} \\mathbf {1} _ {n _ {i}} \\leq \\mathbf {1} _ {n _ {j}} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.649 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}_{n_i}\\) denotes a \\(n_i\\)-dimensional column vector whose elements are all ones. The assignment matrix \\(\\mathbf{X}_{ij} \\in \\mathbb{P}_{n_i n_j}\\) between a pair of graphs \\((\\mathcal{G}_i, \\mathcal{G}_j)\\) denotes a meaningful correspondence that encodes the matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.755 + ], + "angle": 0, + "content": "When considering the synchronisation matching [30, 39] of multiple graphs (\\(m > 2\\)), relying on local matches between graph pairs can readily result in erroneous correspondences that contradict each other globally [63]. Consequently, the multi-graph matching has been addressed in terms of simultaneously solving under the constraints of cycle-consistency [64, 65]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.756, + 0.894, + 0.787 + ], + "angle": 0, + "content": "Definition 1 Cycle-consistency. The matching among \\(\\mathcal{G}_1, \\mathcal{G}_2, \\dots, \\mathcal{G}_m\\) is cycle-consistent (partial transitivity), if" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.801, + 0.891, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {i k} \\mathbf {X} _ {k j} \\leq \\mathbf {X} _ {i j}, \\quad \\forall i, j, k \\in [ m ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In contrast to full matching, i.e. in Eq. (2) the inequalities become equalities and \\( n_i = n_j \\), partial matching requires only that the pairwise matching combinations in the cycle form a subset of the identity matching [3]. An iterative refinement strategy [34] was employed to enhance pairwise" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15622" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "matchings by ensuring global mapping consistency. In [39], the authors proposed a method to achieve cycle-consistency based on spectral decomposition of the assignment matrix. In addition, methods such as convex programming [7], low-rank matrix recovery [76], and spectral decomposition [30] have been proposed to address the synchronization of partial matchings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.471, + 0.304 + ], + "angle": 0, + "content": "Instead of explicitly modeling the cubic number of nonconvex quadratic constraints, a more efficient approach to enforcing cycle consistency is the use of the universe of nodes [55]. In a universe comprising nodes of size \\( d \\in \\mathbb{N} \\), the matching between graphs can be decomposed by matching each graph \\( \\mathcal{G}_i \\) to the space of universe (refer to Fig. 1). We denote the universe matchings as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.314, + 0.468, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\mathbb {U} _ {n _ {i} d} := \\left\\{\\mathbf {U} \\in \\{0, 1 \\} ^ {n _ {i} \\times d}: \\mathbf {U} \\mathbf {1} _ {d} = \\mathbf {1} _ {n _ {i}}, \\mathbf {U} ^ {\\top} \\mathbf {1} _ {n _ {i}} \\leq \\mathbf {1} _ {d} \\right\\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.468, + 0.422 + ], + "angle": 0, + "content": "Lemma 1 Cycle-consistency, universe matching. The pairwise (partial) matching matrices \\(\\{\\mathbf{X}_{ij}\\}_{i,j=1}^{m}\\) is cycle-consistent iff there exists a collection of universe matching \\(\\{\\mathbf{U}_i \\in \\mathbb{U}_{n_id}\\}_{i=1}^m\\), such that for each \\(\\mathbf{X}_{ij}\\), it holds that \\(\\mathbf{X}_{ij} = \\mathbf{U}_i \\mathbf{U}_j^\\top\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.468, + 0.588 + ], + "angle": 0, + "content": "The proof will be presented in the appendix. By projecting each node of a graph to the universe of nodes and identifying the universe matchings \\(\\mathbb{U}_{n_i d}\\), it is ensured that the cycle-consistency constraint will be satisfied throughout the multi-graph matching process. In [59, 61], the authors proposed a gradual assignment procedure to achieve soft matching and clustering through the universe matching. In [4], the universe of nodes is leveraged to incorporate geometric consistency, ensuring both point scaling and convergence. [36] utilized the universe matching to facilitate effective partial multi-graph matching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.468, + 0.649 + ], + "angle": 0, + "content": "Definition 2 Multi-Matching Koopmans-Beckmann's Quadratic Assignment Problem (KB-QAP) [21]. Multi-graph matching is formulated with KB-QAP, by summing KB-QAP objectives among all pairs of graphs:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.661, + 0.468, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\mathbf {X} _ {i j}} \\sum_ {i, j \\in [ m ]} (\\lambda \\cdot t r \\left(\\mathbf {X} _ {i j} ^ {\\top} \\mathcal {A} _ {i} \\mathbf {X} _ {i j} \\mathcal {A} _ {j}\\right) + t r \\left(\\mathbf {X} _ {i j} ^ {\\top} \\mathbf {M} _ {i j}\\right)), \\tag {4} \\\\ s. t. \\quad \\mathbf {X} _ {i j} \\in \\{0, 1 \\} ^ {n _ {i} \\times n _ {j}}, \\mathbf {X} _ {i j} \\in \\mathbb {P} _ {n _ {i} n _ {j}}, \\\\ \\mathbf {X} _ {i k} \\mathbf {X} _ {k j} \\leq \\mathbf {X} _ {i j}, \\quad \\forall i, j, k \\in [ m ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.468, + 0.777 + ], + "angle": 0, + "content": "In Eq. (4), \\(\\lambda\\) is a scaling factor for edge-to-edge similarity, and \\(\\mathbf{M}_{ij}\\) is the node-to-node similarity between \\(\\mathcal{G}_i,\\mathcal{G}_j\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.292, + 0.801 + ], + "angle": 0, + "content": "2.2. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Domain generalization addresses a challenging scenario where one or more different but related domains are provided, with the objective of training a model that can generalize effectively to an unseen test domain [57, 68, 75]. Medical image analysis encounters significant challenges due to factors such as variability in image appearance, the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "complexity and high dimensionality of the data, difficulties in data acquisition, and issues related to data organization, labeling, safety, and privacy. Previous research has proposed DG algorithms across multiple levels, including data-level [69, 71], feature-level [5, 24], model-level [16, 40, 78], and analysis-level [28, 43]. Yu et al. [69] proposed a U-Net z-score normalization network for the stroke lesion segmentation. Liu et al. [28] employed dictionary learning for prostate and fundus segmentation by creating a shape dictionary composed of template masks. In [5], the authors utilized mutual information to differentiate between domain-invariant features and domain-specific ones in ultrasound image segmentation. Gu et al. [16] proposed a domain-style contrastive learning approach that disentangles an image into invariant representations and style codes for DG." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.327, + 0.704, + 0.342 + ], + "angle": 0, + "content": "2.3. Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Test-time adaptation is an emerging paradigm that allows a pre-trained model to adapt to unlabeled data during the testing phase, before making predictions [27]. Several TTA methods have been proposed recently, utilizing techniques such as self-supervised learning [6, 22, 73], batch normalization calibration [8, 32, 47, 70, 77], and input data adaptation [19, 49] to achieve better test-time performance. Typically, VPTTA [8] is a method that freezes the pre-trained model and generates low-frequency prompts for each image during inference in medical image segmentation. In [19], the authors proposed designing two CNN-based sub-networks along with an image normalization network. During test-time training, the image normalization network was adapted for each image. [70] proposed a dynamic mixture coefficient and a statistical transformation operation to adaptively merge the training and testing statistics of the normalization layers. Additionally, the authors design an entropy minimization loss to address the issue of domain shifts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.634, + 0.666 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.891, + 0.78 + ], + "angle": 0, + "content": "This section provides an overview of how multi-graph matching methods can facilitate training for TTDG tasks. Given inputs consisting of images from \\(S\\) source domains \\((S\\geq 1)\\), denoted as \\(\\mathcal{D}_s = \\{D_1,D_2,\\dots ,D_S\\}\\), the objective is to enable the model to make more accurate predictions on \\(T\\) unseen target domains \\((T\\geq 1)\\), denoted as \\(\\mathcal{D}_t = \\{D_1,D_2,\\dots ,D_T\\}\\), during the testing phase." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Inspired by [41], we aim to effectively leverage morphological prior through graph construction. However, unlike the UDA tasks, which typically involve only two domains, our training process simultaneously handles multiple labeled domains. Pairwise matching across multiple domains can lead to the challenges discussed in Sec. 1. To tackle this, we propose learning domain-invariant latent representations from a multi-graph matching perspective, thereby" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15623" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.885, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.557, + 0.893, + 0.643 + ], + "angle": 0, + "content": "Figure 2. Overview of our TTDG framework. During source model training, data from different domains are jointly used to train the Segmentation Network (feature extractor and segmentation head). Feature maps and ground truth masks are utilized to construct graphs \\(\\mathcal{G}_i\\) (\\(i = 4\\) in this figure legend) and corresponding labels \\(Y_i\\) (Sec. 3.1), with universe embeddings learned via back-propagation, incorporating morphological priors (Sec. 3.2). At test time, multi-graph matching is performed on all target domain images in each batch. Despite style differences, these images share common morphological patterns. Universe embeddings are frozen as prior knowledge to guide the matching, and the segmentation network is fine-tuned via back-propagation for efficient adaptation (Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.649, + 0.471, + 0.68 + ], + "angle": 0, + "content": "facilitating rapid adaptation to unseen domains during the TTA phase." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.29, + 0.71 + ], + "angle": 0, + "content": "3.1. Construction of Graph" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Rather than using conventional graph-matching algorithms that depend on explicitly defined keypoints as graph nodes [14, 58, 60], accurately annotating keypoints in medical imaging is both more costly and less practical than in natural images. Therefore, given a mini-batch \\(\\{x_i\\}_{i=1}^m \\in \\mathcal{D}_s\\), as shown in Fig. 2, it is first passed through a general feature extractor, such as ResNet [17], to obtain visual features. We then perform spatially-uniform sampling [25] of pixels within the ground-truth masks at each feature level to obtain \\(n_i\\) foreground nodes with its corresponding labels \\(Y_i \\in \\mathbb{Z}^{n_i}\\). After extracting these fine-grained visual features, we employ a nonlinear projection to transform the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.695 + ], + "angle": 0, + "content": "visual space into the \\(h\\)-dimensional graph space. This approach generates the feature of nodes \\(\\{\\mathcal{V}_i \\in \\mathbb{R}^{n_i \\times h}\\}_{i=1}^m\\) that more effectively preserve the semantic characteristics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.772 + ], + "angle": 0, + "content": "For the weighted adjacency matrix encoding structural information, we use Droppedge [44] to reduce potential bias caused by the frequent visual correspondences, preventing the model from over-relying on specific matches. The formulation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.788, + 0.892, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {i} = \\operatorname {D r o p e d g e} \\left\\{\\text {s o f t m a x} \\left[ \\mathcal {V} _ {i} \\mathcal {W} _ {x} \\cdot \\left(\\mathcal {V} _ {i} \\mathcal {W} _ {y}\\right) ^ {\\top} \\right] \\odot D ^ {- 1} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.82, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}_x\\) and \\(\\mathcal{W}_y\\) are two learnable linear projection, \\(D = \\frac{\\mathcal{V}_i\\mathcal{V}_i^{\\mathrm{T}}}{||\\mathcal{V}_i||_2}\\) is the cosine distance matrix. Edges with larger weights indicate nodes that are closer to each other, so we use the inverse of \\(D\\) and apply the hadamard product \\(\\odot\\) to combine node similarity with geometric proximity. The di" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15624" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "agonal elements of \\(\\mathcal{A}_i\\) are set to zero. Based on the above steps, we obtain a corresponding graph \\(\\mathcal{G}_i = (\\mathcal{V}_i,\\mathcal{A}_i)\\) for each input \\(x_{i}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.147, + 0.43, + 0.163 + ], + "angle": 0, + "content": "3.2. Formulation of the Universe Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.469, + 0.305 + ], + "angle": 0, + "content": "In Sec. 2.1, it was demonstrated that by matching each node of a graph to the universe of nodes, we can maintain cycle-consistency while avoiding the computational burden associated with cubic non-convex constraints. While a randomly initialized matrix can learn accurate universe matchings \\(\\mathbb{U}_{n_i d}\\) in a supervised setting, we introduce universe embeddings \\(\\mathcal{U} \\in \\mathbb{R}^{d \\times h}\\), which serve as learnable latent representations to enhance adaptability in unsupervised multi-domain scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.469, + 0.58 + ], + "angle": 0, + "content": "\\(\\mathcal{U}\\) is initialized as \\(1 / d + 10^{-3}z\\) where \\(z\\sim N(0,1)\\), following the setting in [59]. For each \\(\\mathcal{G}_i\\), we learn a unique matching between nodes in \\(\\nu_{i}\\) and the corresponding in the universe of nodes, as follows: \\(U_{i} = \\mathrm{Sinkhorn}(\\mathcal{V}_{i}\\mathcal{U}^{\\top},\\tau)\\in\\) \\(\\mathbb{U}_{n_i d}\\). Sinkhorn \\((X,\\tau)\\) refers to the Sinkhorn algorithm [9, 51], which applies a relaxed projection with entropic regularization [15] to normalize the matrix \\(X\\), resulting in a doubly-stochastic matrix, and \\(\\tau \\in (0, + \\infty)\\) is the regularization factor. For the obtained \\(U_{i}\\), the \\(d\\times d\\) matrix \\(U_{i}^{\\top}\\mathcal{A}_{i}U_{i}\\) is a row and column reordering of \\(\\mathcal{A}_i\\) based on the universe of nodes. Accordingly, the agreement between two adjacency matrices \\(\\mathcal{A}_i\\) and \\(\\mathcal{A}_j\\), reordered according to their respective universe matching assignment matrices \\(U_{i}\\) and \\(U_{j}\\), can be quantified using the Frobenius inner product namely \\(\\langle U_i^\\top \\mathcal{A}_iU_i,U_j^\\top \\mathcal{A}_jU_j\\rangle\\). Let \\(\\mathbf{U} = [U_1^\\top ,\\dots ,U_m^\\top ]^\\top \\in \\mathbb{R}^{n\\times d}\\) where \\(n = \\sum_{i = 1}^{m}n_{i}\\). The block-diagonal multi-adjacency matrix is defined as \\(\\mathbf{A} = diag(\\mathcal{A}_1,\\dots ,\\mathcal{A}_m)\\in \\mathbb{R}^{n\\times n}\\). We can get the multi-matching formulation as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.584, + 0.469, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f(\\mathbf{U}):= \\max_{U_{1},\\ldots ,U_{m}}\\quad \\sum_{i,j\\in [m]}\\langle U_{i}^{\\intercal}\\mathcal{A}_{i}U_{i},U_{j}^{\\intercal}\\mathcal{A}_{j}U_{j}\\rangle \\\\ = \\max _ {\\mathbf {U} \\in \\mathbb {U}} \\operatorname {t r} \\left(\\mathbf {U} ^ {\\top} \\mathbf {A} \\mathbf {U} \\mathbf {U} ^ {\\top} \\mathbf {A} \\mathbf {U}\\right). \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.902 + ], + "angle": 0, + "content": "It was observed that in the original multi-graph matching optimization problem, graphs are typically constructed for each individual instance, with all the nodes within a graph belonging to the same category. However, in medical imaging, constructing separate graphs for each organ would not only greatly increase the computational complexity of graph matching but also result in misalignments between graphs from different categories, significantly compromising matching accuracy. To address this limitation, we introduce a class-wise similarity matrix \\(\\mathbf{W}\\) to incorporate class-aware label information for each node. This ensures that the nodes are correctly aligned according to their respective categories during multi-graph matching. Specifically, we define \\(W_{ij} = Y_iY_j^{\\top}\\) and \\(\\mathbf{W} = [W_{ij}]_{ij} \\in \\mathbb{Z}^{(\\sum_{i=1}^{m}n_i)\\times(\\sum_{j=1}^{m}n_j)}\\). The class-aware similarity matrix \\(\\tilde{\\mathbf{A}} = \\mathbf{W}^{\\top}\\mathbf{A}\\mathbf{W}\\). Consequently, Eq. (6) can be transformed" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.658, + 0.107 + ], + "angle": 0, + "content": "into the following form" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.109, + 0.892, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\mathbf {U} \\in \\mathbb {U}} \\operatorname {t r} \\left(\\mathbf {U} ^ {\\top} \\tilde {\\mathbf {A}} \\mathbf {U} \\mathbf {U} ^ {\\top} \\tilde {\\mathbf {A}} \\mathbf {U}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.142, + 0.892, + 0.187 + ], + "angle": 0, + "content": "The higher-order projected power iteration (HiPPI) [4] is employed to iteratively solve \\(\\mathbf{U}\\) to obtain a stable convergence form:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.19, + 0.892, + 0.209 + ], + "angle": 0, + "content": "\\[\nU _ {i} ^ {*} = U _ {i} ^ {t} \\quad s. t. \\quad | | U _ {i} ^ {t} - U _ {i} ^ {t - 1} | | < \\theta , \\quad \\forall i \\in [ m ]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.292 + ], + "angle": 0, + "content": "Empirically, we set the convergence threshold \\(\\theta = 10^{-5}\\). In order to guarantee the optimal learning of universe embeddings \\(\\mathcal{U}\\), we incorporate the non-negative matrix factorisation (NMF) [3, 23] and define the overall optimization objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.296, + 0.892, + 0.331 + ], + "angle": 0, + "content": "\\[\nL (\\mathcal {U}) = \\sum_ {i \\in [ m ]} | | U _ {i} ^ {*} - \\mathcal {V} _ {i} \\mathcal {U} ^ {\\mathrm {T}} | | _ {F} ^ {2} + \\alpha | | \\mathcal {U} | | _ {F} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.432 + ], + "angle": 0, + "content": "where \\( ||\\cdot||_F^2 \\) represents the Frobenius norm, and \\( \\alpha \\) is a tuning parameter. Since only matrix multiplication and element-wise division are involved, the Sinkhorn operation is fully differentiable. This allows for effective updates to \\( \\mathcal{U} \\) through back-propagation, enabling the incorporation of morphological prior knowledge." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.44, + 0.851, + 0.457 + ], + "angle": 0, + "content": "3.3. Testing Paradigm on the Target Domains" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.689 + ], + "angle": 0, + "content": "During the Test-Time Adaptation (TTA) phase, we only have access to unlabeled data from the unseen target domains \\(\\mathcal{D}_t\\). Due to distribution shifts, the performance of models trained on \\(\\mathcal{D}_s\\) can degrade significantly. To address this issue, we employ the multi-graph matching on each mini-batch \\(\\{x_i^t\\}_{i=1}^m \\in \\mathcal{D}_t\\). By utilizing universe embeddings \\(\\mathcal{U}\\) learned from \\(\\mathcal{D}_s\\) as described in Sec. 3.2, it is guaranteed cycle-consistency, enabling the network to focus on domain-invariant features guided by morphological priors from medical images. In this phase, \\(\\mathcal{U}\\) is frozen to prevent error accumulation from affecting the prior information. The feature extractor and segmentation head (segmentation network in Fig. 2) are optimized through gradient back-propagation, allowing the model to adapt effectively to the new domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Unsupervised Multi-Matching. Following the same graph construction described in Sec. 3.1, we obtain the \\(\\mathcal{G}_i^t = (\\mathcal{V}_i^t,\\mathcal{A}_i^t)\\) for each \\(x_{i}^{t}\\), where \\(\\mathcal{V}_i^t\\in \\mathbb{R}^{n_i\\times h}\\) represents the \\(n_i\\) class-aware nodes and \\(\\mathcal{A}_i^t\\in \\mathbb{R}^{n_i\\times n_i}\\) is the adjacency matrix encoding the structural information. We first introduce a network-learned affinity matrix \\(M_{ij} = f_{mlp}\\{\\mathcal{V}_i^t\\mathcal{W}_x^t\\) \\((\\mathcal{V}_j^t\\mathcal{W}_y^t)^\\top \\} \\in \\mathbb{R}^{n_i\\times n_j}\\) in graph pair \\(\\mathcal{G}_i\\) and \\(\\mathcal{G}_j\\), which computes the similarity between their corresponding node features. \\(\\mathcal{W}_x^t\\) and \\(\\mathcal{W}_y^t\\) are two learnable linear projection, \\(f_{mlp}\\) is a multi-layer perception (MLP). Subsequently, the matching between \\(\\mathcal{G}_i\\) and the universe of size \\(d\\) is obtained through the inner-product between \\(\\mathcal{V}_i^t\\) and the pre-trained universe embeddings \\(\\mathcal{U}\\), in accordance with the same procedure \\(U_{i} = Sinkhorn(\\mathcal{V}_{i}^{t}\\mathcal{U}^{\\top},\\tau)\\in \\mathbb{U}_{n_{i}d}\\) in Sec. 3.2." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15625" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "In [59, 61], the authors propose solving the multimatching KB-QAP problem presented in Eq. (4) (denoted as \\( J \\)) using a Taylor expansion. Based on the assumptions in Lemma 1 of Sec. 2.1, we can reformulate \\( J \\) in Taylor series as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.168, + 0.469, + 0.205 + ], + "angle": 0, + "content": "\\[\nJ (U) \\approx J \\left(U ^ {0}\\right) + \\sum_ {i} \\frac {\\partial J}{\\partial U _ {i}} \\bigg | _ {U = U ^ {0}} \\left(U _ {i} - U _ {i} ^ {0}\\right) + \\dots . \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.214, + 0.469, + 0.274 + ], + "angle": 0, + "content": "\\(J\\) can be approximated as \\(m\\) linear optimization problems through the first-order Taylor expansion, which can be efficiently solved using the Sinkhorn. The gradient of \\(J\\) with respect to \\(U_{i}\\), denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.277, + 0.469, + 0.317 + ], + "angle": 0, + "content": "\\[\nV _ {i} = \\frac {\\partial J}{\\partial U _ {i}} = \\sum_ {j \\in [ m ]} \\left(\\lambda \\mathcal {A} _ {i} U _ {i} U _ {j} ^ {\\mathrm {T}} \\mathcal {A} _ {j} U _ {j} + M _ {i j} U _ {j}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.325, + 0.468, + 0.354 + ], + "angle": 0, + "content": "\\(V_{i}\\) ensures that \\(U_{i}\\) gradually converges to a high-quality discrete solution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.356, + 0.469, + 0.386 + ], + "angle": 0, + "content": "Finally, the converged \\(\\{U_i\\}\\) fine-tune the segmentation network with the matching loss \\(L_{mat}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.389, + 0.469, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {m a t} = \\sum_ {i, j \\in [ m ]} \\left[ - \\hat {M} _ {i j} ^ {\\gamma} (1 - U _ {i} U _ {j} ^ {\\mathsf {T}}) \\log (U _ {i} U _ {j} ^ {\\mathsf {T}}) \\right] \\\\ - \\sum_ {i, j \\in [ m ]} \\left[ (1 - \\hat {M} _ {i j}) ^ {\\gamma} U _ {i} U _ {j} ^ {\\mathsf {T}} \\log \\left(1 - U _ {i} U _ {j} ^ {\\mathsf {T}}\\right) \\right], \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.476, + 0.469, + 0.507 + ], + "angle": 0, + "content": "where \\(\\hat{M}_{ij} = \\text{sinkhorn}(M_{ij},\\tau)\\), and \\(\\gamma\\) is an amplification factor. See algorithm in appendix for full details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.519, + 0.209, + 0.536 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.179, + 0.558 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.469, + 0.716 + ], + "angle": 0, + "content": "The retinal fundus segmentation datasets comprise five public datasets from different medical centers, denoted as Site A (RIM-ONE [13]), B (REFuge [37]), C (ORIGA [74]), D (REFuge-Test [37]), and E (Drishti-GS [52]), with 159, 400, 650, 800, and 101 images respectively, all consistently annotated for optic disc (OD) and optic cup (OC) segmentation. We adopted the preprocessing method outlined in [8, 28], where each image's region of interest is cropped to \\(800 \\times 800\\) pixels and normalized using min-max normalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.469, + 0.839 + ], + "angle": 0, + "content": "The polyp segmentation datasets consist of four public datasets collected from different medical centers, denoted as Site A (BKAI-IGH-NEOPolyp [33]), B (CVC-ClinicDB/CVC-612 [2]), C (ETIS [50]), and D (Kvasir [18]), containing 1000, 612, 196, and 1000 images, respectively. We followed the preprocessing steps in [8], resizing the images to \\(800 \\times 800\\) pixels and normalizing them using ImageNet-derived statistics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.267, + 0.864 + ], + "angle": 0, + "content": "4.2. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. To ensure a fair comparison across all methods, we followed the protocol in [8] and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.317 + ], + "angle": 0, + "content": "split each dataset into an \\(8:2\\) ratio for training and testing. We employed the ResNet-50 [17] pre-trained on ImageNet as the feature extractor. The training was optimized using the SGD optimizer with a momentum of 0.9 and a learning rate of 0.001. For source model training, we set the mini-batch size to 8, allowing simultaneous matching of 8 graphs per batch. The universe embedding \\(\\mathcal{U}\\) was treated as an additional trainable tensor, integrated into the network weights without influencing subsequent experiments with other methods. During the TTA phase, all methods were trained on the target data without access to their labels. The mini-batch size was set to 4, and the previously learned \\(\\mathcal{U}\\) from source model training was frozen. All experiments were implemented using PyTorch and conducted on 4 NVIDIA 3090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Evaluation Metrics. The Dice score (DSC, \\(\\%\\) ) was used to quantify the accuracy of the predicted masks, serving as the primary metric for evaluating segmentation performance. Additionally, the enhanced alignment metric \\(E_{\\phi}^{max}\\) [12] was adopted to measure both pixel-level and global-level similarity. To further assess the consistency between predictions and ground truths, we also employed the structural similarity metric \\(S_{\\alpha}\\) [11]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.702, + 0.469 + ], + "angle": 0, + "content": "4.3. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.628 + ], + "angle": 0, + "content": "We selected U-Net [45] as the benchmark model for the segmentation task, training it on the source domains and testing on the target domain without adaptation (No Adapt). In addition, we compared eight state-of-the-art (SOTA) methods, including two entropy-based approaches (TENT [56] and SAR [35]), a shape template-based method (TASD [28]), a dynamically adjusted learning rate method (DLTTA [66]), batch normalization-based methods (DomainAdaptor [70], VPTTA [8]), and noise estimation-based approaches (DeY-Net [62] and NC-TTT [38])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-Source Generalization. In these experiments, we adopted a leave-one-out training strategy [6] (i.e., with \\( S = |\\mathcal{D}_s \\cup \\mathcal{D}_t| - 1 \\) and \\( T = 1 \\)). Although some methods, such as TASD [28], DeY-Net [62] and VPTTA [8], were originally designed for single-source domain training, we simulated single-domain conditions by mixing multi-source domain data, making the experimental setup still feasible for these approaches. Furthermore, we have also included experiments specifically focused on single-source domains later in the study. The segmentation results for the fundus and polyp datasets are shown in Tables 1 and 2, respectively. For OD/OC segmentation, all TTA methods outperformed the No Adapt baseline, highlighting their effectiveness in addressing domain shifts. Comparatively, our method consistently achieved better average performance than all other approaches. Specifically, for the key segmentation metric DSC, our method exceeded the second-best (DeY-Net [62]) by \\( 2.88\\% \\) and outperformed the No Adapt by \\( 19.09\\% \\). A" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15626" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.09, + 0.895, + 0.132 + ], + "angle": 0, + "content": "Table 1. Multi sources domain generalization in retinal fundus segmentation. The average performance (mean ± standard deviation) of three trials for our method and nine SOTA methods. \"Site A\" means training on Sites B-E and testing on Site A, and similarly for the others. The best results are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.137, + 0.885, + 0.222 + ], + "angle": 0, + "content": "
MethodsSite ASite BSite CSite DSite E
\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{2}\\)\\(S_{\\text{a}}\\)
No Adapt (U-Net [45])65.60±5.7891.75±0.1379.11±0.0774.65±4.8888.78±0.0883.31±0.0263.15±7.3081.03±0.1979.77±0.0368.11±5.4989.46±0.2082.12±0.0775.34±1.0190.68±0.1687.53±0.05
TENT (ICLR'21) [56]75.22±3.9994.10±0.1083.97±0.0581.06±0.9393.45±0.1192.36±0.0375.23±2.3384.62±0.1478.12±2.5991.33±0.3186.41±0.1283.17±0.7892.74±0.1093.64±0.0378.5691.9488.20
TASD (AAAI'22) [28]79.70±2.5288.73±0.4583.67±0.1782.96±2.1490.01±0.2090.99±0.1186.12±1.1193.59±0.1387.08±0.0778.02±6.3087.20±0.1687.94±0.0483.71±3.7790.88±0.2590.37±0.1282.1090.8888.01
DLTTA (TMI'22) [66]71.88±5.5986.29±0.2180.70±0.1380.12±2.2690.53±0.1788.63±0.0681.46±3.3593.08±0.0887.48±0.0275.90±7.2784.27±0.1284.79±0.0682.75±6.5590.10±0.3189.33±0.1478.4288.8586.18
SAR (ICLR'23) [35]75.29±2.6988.73±0.3480.15±0.2686.42±1.5796.88±0.2791.28±0.0980.69±4.3092.10±0.1586.31±0.2780.27±5.4789.98±0.3187.20±0.1589.61±0.2895.90±0.0892.06±0.0182.4692.7187.40
DomainAdaptor (CVPR'23) [70]77.14±1.2190.80±0.1981.31±0.0283.95±0.8093.17±0.0990.44±0.0882.79±1.0995.36±0.1088.57±0.1386.50±3.2995.26±0.3391.07±0.2086.24±1.4494.90±0.2993.77±0.1883.3293.9089.03
DeY-Net (WACV'24) [62]79.67±1.3589.14±0.2383.68±0.0490.07±3.1290.06±0.3392.31±0.0785.94±1.8593.54±0.2287.15±0.1083.47±1.4095.29±0.1788.70±0.0488.76±2.2895.43±0.2190.45±0.0885.5894.0988.46
VPTTA (CVPR'24) [8]77.85±1.4094.09±0.0183.97±0.0584.57±0.9995.38±0.0389.53±0.0983.61±1.6796.39±0.0489.64±0.0582.75±0.5096.53±0.0489.83±0.0189.74±1.5198.22±0.2494.77±0.0783.7096.1289.55
NC-TTT (CVPR'24) [38]80.83±1.7789.61±0.2384.41±0.0887.07±1.3195.63±0.0690.09±0.0286.93±2.2794.69±0.1188.39±0.0382.10±4.2891.49±0.2286.33±0.0790.08±0.0995.93±0.1091.53±0.0985.4093.4788.15
Ours82.53±1.5291.85±0.2285.50±0.2088.98±0.8996.69±0.0491.63±0.1288.73±0.7095.80±0.0689.71±0.0290.25±1.3397.97±0.1992.89±0.1391.83±0.7197.14±0.2092.82±0.0988.4695.8490.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.235, + 0.893, + 0.276 + ], + "angle": 0, + "content": "Table 2. Multi sources domain generalization in the polyp segmentation. The average performance (mean ± standard deviation) of three trials for our method and nine SOTA methods. \"Site A\" means training on Sites B-D and testing on Site A, and similarly for the others. The best results are highlighted in red." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.277, + 0.885, + 0.378 + ], + "angle": 0, + "content": "
MethodsSite ASite BSite CSite DAverage
\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\uparrow\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)
No Adapt (U-Net [45])82.67±2.2492.11±0.0789.59±0.1081.20±4.3892.46±0.2086.08±0.0678.33±1.6294.00±0.0885.35±0.0175.51±2.7485.90±0.1781.71±0.2179.4391.1285.68
TENT (ICLR'21) [56]78.69±1.1288.28±0.1485.44±0.0877.71±3.0989.73±0.0981.90±0.1280.19±3.3596.03±0.0487.80±0.1772.63±5.1182.47±0.3078.92±0.1477.3189.1383.52
TASD (AAAI'22) [28]84.60±4.1492.55±0.2190.76±0.1085.28±2.2093.10±0.1386.83±0.1982.43±5.3996.31±0.0889.61±0.0280.04±4.4188.26±0.1184.09±0.0383.0992.5687.82
DLTTA (TMI'22) [66]80.68±2.3790.01±0.2186.89±0.0980.86±1.7791.98±0.1085.83±0.0475.57±3.1293.90±0.0884.15±0.1074.25±3.4083.49±0.1279.44±0.1077.8489.8584.08
SAR (ICLR'23) [35]79.37±0.7788.40±0.1286.11±0.0378.40±2.0790.23±0.1183.39±0.0280.94±0.9095.83±0.1286.84±0.0778.23±2.9186.37±0.0981.66±0.1479.2490.2184.50
DomainAdaptor (CVPR'23) [70]88.58±0.9695.03±0.0490.95±0.0181.12±1.0792.31±0.0886.60±0.0281.77±2.1795.18±0.0387.19±0.0179.91±1.3387.23±0.0483.64±0.0582.8592.4487.10
DeY-Net (WACV'24) [62]83.61±2.3190.79±0.1189.25±0.0980.90±3.1791.49±0.1085.36±0.0380.46±1.4395.83±0.0887.14±0.0373.85±3.0483.00±0.1378.96±0.0879.7190.2885.18
VPTTA (CVPR'24) [8]82.34±0.6991.03±0.0888.35±0.0184.32±0.4491.53±0.0787.46±0.0284.43±0.5295.75±0.1688.91±0.0382.40±0.2989.60±0.0485.18±0.0183.3791.9787.47
NC-TTT (CVPR'24) [38]87.96±1.2293.71±0.1390.66±0.0783.70±2.0893.78±0.0987.51±0.0789.29±0.9096.66±0.2291.26±0.0778.23±0.7986.37±0.0981.66±0.0684.8092.6387.77
Ours90.97±0.6696.57±0.0692.92±0.0887.49±0.8295.21±0.0489.93±0.0986.87±0.5897.53±0.0590.16±0.1283.97±1.1890.67±0.1386.44±0.0887.3294.6789.86
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.478 + ], + "angle": 0, + "content": "similar trend was observed in the polyp segmentation results, where we outperformed the second-best method (NC-TTT [38]) and No Adapt by \\(2.52\\%\\) and \\(7.89\\%\\), respectively. Additionally, our approach exhibited greater stability across multiple trials compared to the other methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Single-Source Generalization. In these experiments, the models are trained on one domain and tested on the remaining datasets (i.e., with \\( S = 1 \\) and \\( T = |\\mathcal{D}_s \\cup \\mathcal{D}_t| - 1 \\)). Compared to multi-source generalization, single-source training is considered more challenging due to the limited domain information available during the training phase. By comparing Tables 2 and 3, we can observe that all models show lower segmentation performance (in terms of DSC) in the single-source setting compared to the multi-source results. Some methods even perform worse than the No Adapt baseline in single-source scenarios. However, our method still achieves the best average performance across 12 domain transfer experiments, outperforming the second-best approach (NC-TTT [38]) by \\( 1.83\\% \\) and exceeding the No Adapt by \\( 9.77\\% \\) in terms of DSC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Fig. 3 presents the visualization results for \"Site A\" in Table 1. The No Adapt baseline exhibits significant misalignment, with notably distorted shapes and inaccurate boundaries. VPTTA offers some improvement but still suffers from distortions and incomplete segmentation, particularly in the optic cup area. NC-TTT shows instances of overlapping segmentation. In contrast, our method delivers the most precise segmentation, closely aligning with the ground truth. The Grad-CAM [48] visualizations reinforce this, as the attention of network is focused specifically on the relevant regions. These results demonstrate our method's enhanced generalization to unseen domains by effectively incorporating morphological priors, resulting in higher seg" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.401, + 0.891, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.537, + 0.893, + 0.607 + ], + "angle": 0, + "content": "Figure 3. Visualization comparison of segmentation results and Grad-CAM outputs from the final layer of the backbone network for the No Adapt baseline, VPTTA [8], NC-TTT [38], and our proposed method on retinal fundus images. Additional visual comparisons are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.619, + 0.892, + 0.651 + ], + "angle": 0, + "content": "mentation accuracy. For further visualization and additional experimental results, please refer to the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.669, + 0.597, + 0.685 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.855, + 0.712 + ], + "angle": 0, + "content": "5.1. Effectiveness of the Universe Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The universe embeddings \\(\\mathcal{U}\\) play a crucial role in incorporating morphological priors from medical images, resulting in an assignment matrix that projects each node to the universe of nodes. To validate the effectiveness during the TTA, we conducted the following experiments: (1) Without universe embeddings (denoted as \\(w / o\\mathcal{U}\\)), the universe matching assignment matrix \\(\\mathbf{U}\\) is initialized following the setting in [59] (i.e. \\(\\mathbf{U} = 1 / d + 10^{-3}z\\), where \\(z\\sim N(0,1)\\)), leading to random matching between the nodes and the universe; (2) Without morphological priors but with \\(\\mathcal{U}\\) (denoted as \\(w / o\\) priors). As shown in Table 4, the performance in (1) is slightly better than in (2) by \\(1.76\\%\\) in terms of DSC" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15627" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.891, + 0.117 + ], + "angle": 0, + "content": "Table 3. Single source domain generalization in the polyp segmentation. The average performance of three trials for our method and nine SOTA methods. A → B represents models trained on Site A and tested on Site B, and similar for others. Best results are colored as red." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.117, + 0.885, + 0.226 + ], + "angle": 0, + "content": "
MethodsA→BA→CA→DB→AB→CB→DC→AC→BC→DD→AD→BD→CAvg.
Dice Score Metric ↑ (DSC, mean±std)
No Adapt (U-Net [45])76.49±1.4867.70±3.2770.94±7.5676.21±2.9062.80±8.0370.03±4.5372.87±3.6269.13±5.1570.44±3.2779.78±1.0872.47±2.8074.91±3.1871.98
TENT (ICLR'21) [56]68.13±0.1865.99±0.1166.11±0.1477.56±0.2159.87±0.2068.89±0.1473.21±0.1065.04±0.0972.35±0.2277.88±0.1470.01±0.1173.60±0.2069.89
TASD (AAAI'22) [28]73.02±2.2479.67±6.3778.23±4.4466.82±9.1772.09±4.4077.95±1.6974.20±3.3069.74±7.2776.13±2.8283.77±0.5176.09±3.2680.10±1.4775.65
DLTTA (TMI'22) [66]72.23±0.0470.78±0.0674.56±0.0464.87±0.0160.05±0.0978.47±0.0371.22±0.0365.78±0.1071.00±0.1379.89±0.0879.10±0.1074.52±0.0171.87
SAR (ICLR'23) [35]70.34±0.1377.90±0.2077.71±0.0968.19±0.1266.56±0.0873.92±0.1376.12±0.1171.29±0.1975.08±0.0882.07±0.0480.44±0.2183.10±0.1575.22
DomainAdaptor (CVPR'23) [70]78.39±0.6177.09±0.3579.45±0.2174.98±0.7070.07±0.6680.21±0.4676.02±0.3373.40±0.5081.33±0.5486.10±0.4380.56±0.6277.86±0.2177.96
DeY-Net (WACV'24) [62]74.45±1.3176.07±2.5280.11±0.7270.90±1.1069.23±2.1876.31±1.3278.33±4.4374.60±1.6673.52±2.2580.70±0.2778.23±1.1781.48±0.9376.16
VPTTA (CVPR'24) [8]78.18±0.0175.14±0.0382.73±0.0273.65±0.0664.78±0.0182.86±0.0377.90±0.0170.64±0.0577.22±0.1087.19±0.0384.97±0.0882.12±0.0278.12
NC-TTT (CVPR'24) [38]79.21±0.2481.76±0.5783.22±0.6977.35±0.3379.40±0.6780.94±0.5480.17±0.7072.36±0.6278.55±0.4384.09±0.3882.38±0.5179.77±0.7479.93
Ours80.47±0.2883.93±0.1983.18±0.3378.33±0.2681.25±0.2383.58±0.3478.86±0.2676.49±0.2180.20±0.2489.17±0.2281.68±0.2283.86±0.2881.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.25, + 0.468, + 0.276 + ], + "angle": 0, + "content": "Table 4. Ablation study in the retinal fundus segmentation. Details of the experiment can be found in Sec 5.1." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.278, + 0.468, + 0.34 + ], + "angle": 0, + "content": "
MethodsSite ASite BSite CSite DSite EAverage
Dice Score Metric ↑ (DSC, mean±std)
No Adapt65.60±5.7874.65±4.8863.15±7.3068.11±5.4975.34±1.0169.37
w/o U77.09±1.9083.24±2.0180.50±1.4478.32±1.5583.18±2.1880.47
w/o priors79.51±0.8079.36±1.1377.40±1.8977.29±1.7780.00±1.9378.71
Ours82.53±1.5288.98±0.8988.73±0.7090.25±1.3391.83±0.7188.46
" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.358, + 0.264, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.358, + 0.465, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.493, + 0.469, + 0.562 + ], + "angle": 0, + "content": "Figure 4. The performance of DSC (\\%) and inference time (s) with different universe sizes \\( d \\) are shown, with the experimental setup identical to that in Table 1. The first curve represents the No Adapt baseline (\\( d = 0 \\)), while the subsequent curve corresponds to the results obtained with a universe size of \\( d = 120 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.469, + 0.648 + ], + "angle": 0, + "content": "on average. However, when using the pre-trained \\(\\mathcal{U}\\) derived from the source model, the segmentation results show a significant improvement. This highlights the effectiveness of the pre-trained morphological priors embedded in \\(\\mathcal{U}\\) for medical imaging tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.368, + 0.667 + ], + "angle": 0, + "content": "5.2. Effectiveness of the Universe Size" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The universe size \\(d\\) directly impacts both the matching accuracy and computational efficiency, and its selection depends on the number of sampled nodes. Previous works [4, 36, 59], ensured that the \\(d\\) matched the number of nodes in each graph by labeling an equal number of keypoints. In our experiments, \\(d\\) is determined by the spatially-uniform sampling described in Sec 3.1, and we empirically set \\(d = 100 \\times (n + 1)s^{-1}\\), where \\(s\\) is the sampling step and \\(n\\) is the number of segmentation categories required for the task. We evaluated several \\(d\\), with the results shown in Fig. 4. We observe that excessively small or large step sizes—resulting in too many or too few sampled nodes—negatively impact both training speed and performance. However, when \\(d\\) is within a reasonable range (in this experiment, \\(s \\in [2,10]\\) i.e. \\(d \\in [30,150]\\)), the model maintains good scalability" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.252, + 0.619, + 0.266 + ], + "angle": 0, + "content": "and performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.266, + 0.891, + 0.32 + ], + "angle": 0, + "content": "Table 5. Comparison of common variables and performance between pairwise matching and multi-matching, using the same experimental setup as in Table 1. Both methods follow the graph construction in Sec. 3.1." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.322, + 0.884, + 0.432 + ], + "angle": 0, + "content": "
paramSite A-E Averagedescription
Pairwise [58]Multi-Matching (Ours)
lr10-310-3learning rate
batch44batch size in inference
h256256dimension of node feature
τ0.050.05regularization factor of Sinkhorn
Iter2020max iterations of Sinkhorn
d-120universe size
DSC ↑84.6388.46dice score metric (%)
time ↓0.9300.392inference time per image (s/img)
Param ↓1.0710.658parameter count (M)
FLOPs ↓15.354.255floating point operations per second (G)
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.434, + 0.835, + 0.45 + ], + "angle": 0, + "content": "5.3. Multi-Matching vs. Pairwise Matching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Compared to pairwise matching, multi-matching uses cycle-consistency for global optimization, which helps avoid local optima in pairwise. Moreover, joint optimization reduces the computational complexity of large-scale matching. By incorporating morphological consistency constraints, multi-matching exhibits greater robustness. To validate these claims, we conducted experiments comparing the two approaches. For the pairwise matching method, we adopted the benchmark approach [58] and applied the same graph construction process as in multi-matching. The comparison results are presented in Table 5. It is evident that multi-matching outperforms pairwise matching in terms of DSC, inference time, parameters and FLOPs. The inference time of multi-matching was reduced by approximately \\(57.85\\%\\) compared to pairwise matching, while segmentation accuracy improved by \\(3.83\\%\\) in DSC." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.619, + 0.725 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper presents a novel multi-matching framework for TTDG, which leverages universe embeddings to incorporate morphological priors from medical images while ensuring cycle-consistency, along with an unsupervised test-time paradigm that fully integrates these priors for efficient adaptation. Extensive experiments, including comparisons and ablation studies demonstrate that our graph-matching-based approach achieves SOTA performance, outperforming entropy-based, template-based, batch normalization-based, and noise estimation-based methods across both multi-source and single-source domain generalization tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15628" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.214 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (Grant No. 62306003), the Open Research Fund from Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), under Grant No. GML-KF-24-29 and the Open Foundation of Jiangxi Provincial Key Laboratory of Image Processing and Pattern Recognition (ET202404437)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.244, + 0.173, + 0.259 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.268, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[1] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.472, + 0.409 + ], + "angle": 0, + "content": "[2] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized Medical Imaging and Graphics, 43:99-111, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.412, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[3] Florian Bernard, Johan Thunberg, Jorge Goncalves, and Christian Theobalt. Synchronisation of partial multi-matchings via non-negative factorisations. Pattern Recognition, 92:146-155, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.47, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[4] Florian Bernard, Johan Thunberg, Paul Swoboda, and Christian Theobalt. Hippi: Higher-order projected power iterations for scalable multi-matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 10284-10293, 2019. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.528, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[5] Yuan Bi, Zhongliang Jiang, Ricarda Clarenbach, Ghotbi, and Nassir Navab. Mi-segnet: Mutual information-based us segmentation for unseen domain generalization. In MICCAI, pages 130-140. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.586, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[6] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 24172-24182, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.643, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[7] Yuxin Chen, Leonidas Guibas, and Qixing Huang. Near-optimal joint object matching via convex relaxation. In ICML, pages II-100, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.687, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[8] Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each test image deserves a specific prompt: Continual test-time adaptation for 2d medical image segmentation. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 11184-11193, 2024. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[9] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Adv. Neural Inform. Process. Syst., 26, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[10] Frank De Groot. High-resolution x-ray emission and x-ray absorption spectroscopy. Chemical Reviews, 101(6):1779-1808, 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[11] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4548-4557, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.268, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[12] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. *IJCAI*, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[13] Francisco Fumero, Silvia Alayón, José L Sanchez, Jose Sigut, and M Gonzalez-Hernandez. Rim-one: An open retinal image database for optic nerve evaluation. In International Symposium on Computer-based Medical Systems (CBMS), pages 1-6. IEEE, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[14] Quankai Gao, Fudong Wang, Nan Xue, Jin-Gang Yu, and Gui-Song Xia. Deep graph matching under quadratic constraint. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5069-5078, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[15] Steven Gold, Anand Rangarajan, et al. Softmax to softassign: Neural network algorithms for combinatorial optimization. Journal of Artificial Neural Networks, 2(4):381-399, 1996. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[16] Ran Gu, Guotai Wang, Jiangshan Lu, Jingyang Zhang, Wenhui Lei, Yinan Chen, Wenjun Liao, Shichuan Zhang, Kang Li, Dimitris N Metaxas, et al. Cddsa: Contrastive domain disentanglement and style augmentation for generalizable medical image segmentation. Medical Image Analysis, 89: 102904, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 770-778, 2016. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[18] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, pages 451-462. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[19] Neerav Karani, Ertunc Erdil, Krishna Chaitanya, and Ender Konukoglu. Test-time adaptable neural networks for robust medical image segmentation. Medical Image Analysis, 68: 101907, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[20] Boah Kim, Inhwa Han, and Jong Chul Ye. Diffuse morph: Unsupervised deformable image registration using diffusion model. In Eur. Conf. Comput. Vis., pages 347-364. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[21] Tjalling C Koopmans and Martin Beckmann. Assignment problems and the location of economic activities. *Econometrica: journal of the Econometric Society*, pages 53–76, 1957. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[22] Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Kulkarni, Hiran Sarkar, Varun Jampani, and R Venkatesh Babu. Concurrent subsidiary supervision for unsupervised source-free domain adaptation. In ECCV, pages 177-194. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[23] Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. nature, 401 (6755):788-791, 1999. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5400-5409, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[25] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semanticcomplete graph matching for domain adaptive object detec" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15629" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "tion. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5291-5300, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[26] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):9022-9040, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[27] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. Int. J. Comput. Vis., pages 1-34, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.277 + ], + "angle": 0, + "content": "[28] Quande Liu, Cheng Chen, Qi Dou, and Pheng-Ann Heng. Single-domain generalization in medical image segmentation via test-time adaptation from shape dictionary. In AAAI, pages 1756-1764, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[29] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[30] Eleonora Maset, Federica Arrigoni, and Andrea Fusiello. Practical and efficient multi-view matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4568-4576, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[31] Jose G Moreno-Torres, Troy Raeder, Rocio Alaiz-Rodriguez, Nitesh V Chawla, and Francisco Herrera. A unifying view on dataset shift in classification. Pattern recognition, 45(1):521-530, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[32] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[33] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Advances in visual Computing: 16th International Symposium (ISVC), pages 15-28. Springer, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[34] Andy Nguyen, Mirela Ben-Chen, Katarzyna Welnicka, Yinyu Ye, and Leonidas Guibas. An optimization approach to improving collections of shape maps. In Computer Graphics Forum, pages 1481-1491. Wiley Online Library, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[35] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. ICLR, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[36] Zhakshylyk Nurlanov, Frank R Schmidt, and Florian Bernard. Universe points representation learning for partial multi-graph matching. In AAAI, pages 1984-1992, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[37] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel Van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, et al. Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Medical Image Analysis, 59:101570, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] David Osowiechi, Gustavo A Vargas Hakim, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdanpanah, Ismail Ben Ayed, and Christian Desrosiers. Nc-ttt: A" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "noise constrastive approach for test-time training. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 6078-6086, 2024. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[39] Deepti Pachauri, Risi Kondor, and Vikas Singh. Solving the multi-way matching problem by permutation synchronization. Adv. Neural Inform. Process. Syst., 26, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[40] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 1406-1415, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[41] Bin Pu, Xingguo Lv, Jiewen Yang, He Guannan, Xingbo Dong, Yiqun Lin, Li Shengli, Tan Ying, Liu Fei, Ming Chen, Zhe Jin, Kenli Li, and Xiaomeng Li. Unsupervised domain adaptation for anatomical structure detection in ultrasound images. In ICML, pages 41204-41220. PMLR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[42] Bin Pu, Liwen Wang, Jiewen Yang, Guannan He, Xingbo Dong, Shengli Li, Ying Tan, Ming Chen, Zhe Jin, Kenli Li, et al. M3-uda: A new benchmark for unsupervised domain adaptive fetal cardiac structure detection. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 11621-11630, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[43] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 12556-12565, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.454 + ], + "angle": 0, + "content": "[44] Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Droppedge: Towards deep graph convolutional networks on node classification. In ICLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.497 + ], + "angle": 0, + "content": "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 4938-4947, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[47] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Adv. Neural Inform. Process. Syst., 33:11539-11551, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[48] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 618-626, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[49] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Adv. Neural Inform. Process. Syst., 35: 14274-14289, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[50] Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer. International Journal of Computer Assisted Radiology and Surgery, 9:283-293, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[51] Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices. The annals of mathematical statistics, 35(2):876-879, 1964. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Jayanthi Sivaswamy, SR Krishnadas, Gopal Datt Joshi, Madhulika Jain, and A Ujjwaft Syed Tabish. Drishti-gs: Retinal" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15630" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "image dataset for optic nerve head (onh) segmentation. In International Symposium on Biomedical Imaging (ISBI), pages 53-56. IEEE, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[53] Charles P Slichter. Principles of magnetic resonance. Springer Science & Business Media, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.471, + 0.207 + ], + "angle": 0, + "content": "[54] Gail Ter Haar. Therapeutic applications of ultrasound. Progress in biophysics and molecular biology, 93(1-3):111-129, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.211, + 0.471, + 0.265 + ], + "angle": 0, + "content": "[55] Roberto Tron, Xiaowei Zhou, Carlos Esteves, and Kostas Daniilidis. Fast multi-image matching via density-based clustering. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4057-4066, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.269, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[56] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.313, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[57] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.439 + ], + "angle": 0, + "content": "[58] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 3056-3065, 2019. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.471, + 0.511 + ], + "angle": 0, + "content": "[59] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Graduated assignment for joint multi-graph matching and clustering with application to unsupervised graph matching network learning. Adv. Neural Inform. Process. Syst., 33: 19908-19919, 2020. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.514, + 0.471, + 0.583 + ], + "angle": 0, + "content": "[60] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Neural graph matching network: Learning lawler's quadratic assignment problem with extension to hypergraph and multiple-graph matching. IEEE Trans. Pattern Anal. Mach. Intell., 44 (9):5261-5279, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.586, + 0.471, + 0.641 + ], + "angle": 0, + "content": "[61] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Unsupervised learning of graph matching with mixture of modes via discrepancy minimization. IEEE Trans. Pattern Anal. Mach. Intell., 45(8):10500-10518, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.471, + 0.726 + ], + "angle": 0, + "content": "[62] Ruxue Wen, Hangjie Yuan, Dong Ni, Wenbo Xiao, and Yaoyao Wu. From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In Proc. IEEE/CVF Winter Conference on Applications of Computer Vision, pages 464-474, 2024. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.471, + 0.785 + ], + "angle": 0, + "content": "[63] Junchi Yan, Yu Tian, Hongyuan Zha, Xiaokang Yang, Ya Zhang, and Stephen M Chu. Joint optimization for consistent multiple graph matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 1649-1656, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[64] Junchi Yan, Yin Li, Wei Liu, Hongyuan Zha, Xiaokang Yang, and Stephen Mingyu Chu. Graduated consistency-regularized optimization for multi-graph matching. In ECCV, pages 407–422. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[65] Junchi Yan, Minsu Cho, Hongyuan Zha, Xiaokang Yang, and Stephen M Chu. Multi-graph matching via affinity optimization with graduated consistency regularization. IEEE Trans. Pattern Anal. Mach. Intell., 38(6):1228-1242, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[66] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Transactions on Medical Imaging, 41 (12):3575-3586, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[67] Jiewen Yang, Xinpeng Ding, Ziyang Zheng, Xiaowei Xu, and Xiaomeng Li. Graphecho: Graph-driven unsupervised domain adaptation for echocardiogram video segmentation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 11878-11887, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[68] Jee Seok Yoon, Kwanseok Oh, Yooseung Shin, Maciej A Mazurowski, and Heung-II Suk. Domain generalization for medical image analysis: A survey. arXiv preprint arXiv:2310.08598, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[69] Weiyi Yu, Zhizhong Huang, Junping Zhang, and Hongming Shan. San-net: Learning generalization to unseen sites for stroke lesion segmentation with self-adaptive normalization. Computers in Biology and Medicine, 156:106717, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[70] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 18971-18981, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[71] Ling Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Holger Roth, Andriy Myronenko, Daguang Xu, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation. IEEE transactions on medical imaging, 39(7):2531-2540, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[72] Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. Adv. Neural Inform. Process. Syst., 34:23664-23678, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.545, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[73] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Adv. Neural Inform. Process. Syst., 35:38629-38642, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[74] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. Origa-light: An online retinal fundus image database for glaucoma analysis and research. In 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology, pages 3065-3068. IEEE, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[75] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(4):4396-4415, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[76] Xiaowei Zhou, Menglong Zhu, and Kostas Daniilidis. Multi-image matching via fast alternating minimization. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4032-4040, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.785, + 0.892, + 0.84 + ], + "angle": 0, + "content": "[77] Yuliang Zou, Zizhao Zhang, Chun-Liang Li, Han Zhang, Tomas Pfister, and Jia-Bin Huang. Learning instance-specific adaptation for cross-domain segmentation. In ECCV, pages 459-476. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.842, + 0.892, + 0.883 + ], + "angle": 0, + "content": "[78] Yukun Zuo, Hantao Yao, and Changsheng Xu. Attention-based multi-source domain adaptation. IEEE Transactions on Image Processing, 30:3793-3803, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15631" + } + ] +] \ No newline at end of file diff --git a/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_origin.pdf b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30234c4609084bd824f50640dd12c605b6fbb6bd --- /dev/null +++ b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/eebac6bc-8348-490f-a98e-64c5f1658ac6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca9a67f9271211b9dd5b36b031f6819c149372f56fb794e4d73dcbbe7fd1cba7 +size 6037429 diff --git a/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/full.md b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b89f85c21313ec60eb47be8f9069a14ebf0c513c --- /dev/null +++ b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/full.md @@ -0,0 +1,364 @@ +# Test-Time Domain Generalization via Universe Learning: A Multi-Graph Matching Approach for Medical Image Segmentation + +Xingguo Lv $^{1,2}$ Xingbo Dong $^{1\dagger}$ Liwen Wang $^{1}$ Jiewen Yang $^{3}$ Lei Zhao $^{2}$ Bin $\mathrm{Pu}^{3\dagger}$ Zhe Jin $^{1}$ Xuejun Li $^{1}$ + +1 Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, Anhui University. 2 Hunan University. 3 The Hong Kong University of Science and Technology. + +lvxg@stu.ahu.edu.cn, xingbo.dong@ahu.edu.cn, eebinpu@ust.hk + +# Abstract + +Despite domain generalization (DG) has significantly addressed the performance degradation of pre-trained models caused by domain shifts, it often falls short in real-world deployment. Test-time adaptation (TTA), which adjusts a learned model using unlabeled test data, presents a promising solution. However, most existing TTA methods struggle to deliver strong performance in medical image segmentation, primarily because they overlook the crucial prior knowledge inherent to medical images. To address this challenge, we incorporate morphological information and propose a framework based on multi-graph matching. Specifically, we introduce learnable universe embeddings that integrate morphological priors during multi-source training, along with novel unsupervised test-time paradigms for domain adaptation. This approach guarantees cycle-consistency in multi-matching while enabling the model to more effectively capture the invariant priors of unseen data, significantly mitigating the effects of domain shifts. Extensive experiments demonstrate that our method outperforms other state-of-the-art approaches on two medical image segmentation benchmarks for both multi-source and single-source domain generalization tasks. The source code is available at https://github.com/Yore0/TTDG-MGM. + +# 1. Introduction + +In the field of medical image processing, semantic segmentation is a technique that enables the precise identification and quantitative analysis of target regions. Currently, a considerable number of medical models can effectively perform segmentation tasks [1, 29]. Nevertheless, when pretrained on source datasets, their performance often declines in real-world deployment due to domain shifts [31, 72] caused + +![](images/37f3209c7ba8a047436701a6e3f908802f5c230a160bae2e68f373c4b7118dba.jpg) +(a) + +![](images/76059ff464c857151caee3ae8804e9c05e7207c372ece7dd8691238b57bb71ec.jpg) +(b) + +![](images/545d2104be1c93bbbc60f3bbf62fc5229e827f8a8846b41dca6b66375537bdfa.jpg) + +![](images/9b76d0a1906a60cd0b67fc00f89c7ef201a71f907277ef379536b3823f6e5b18.jpg) +X +(c) + +![](images/8f714bc8a7cce0fe94f725696273ec4de0156be223bb4e0126ed2431b3c09209.jpg) +Figure 1. Illustration of multi-graph matching using the universe of nodes. (a) Direct pairwise matching between graphs 1, 2, and 3. (b) Multi-graph matching with the universe of nodes. (c) Equivalence between (a) and (b) (see Lemma 1 in Sec. 2.1 for details). + +by differences in imaging devices, protocols, patient demographics, and preprocessing methods, which impact the model's generalization ability. To address domain shifts, previous researches in domain adaptation, such as domain generalization (DG) [57, 68, 75], have focused on designing sophisticated models that train jointly on source and target domains or across multiple styled domains. Despite the progress, these methods fall short in practice, as no source domains can fully capture all real-world variations, and retraining for each new data is impractical. Consequently, a more practical approach is to adaptively fine-tune pre-train model using the information from the unlabeled unseen data during the test, which is referred to test-time adaptation (TTA) [27]. + +Medical images, due to their reliance on specific physical principles (such as X-ray absorption [10], magnetic reso + +nance [53], sound wave reflection [54], etc.), are fundamentally different from images capturing visible light in natural scenes. Moreover, the physiological characteristics of specific anatomical regions exhibit relatively stable shapes and spatial layouts, and the morphology of these structures is predictable [41, 42]. These prior insights provide valuable information that is particularly useful in medical imaging tasks. In areas such as segmentation [28, 67], registration [20], and detection [41, 42], many well-established approaches have been proposed that combine above handcrafted designs with neural network-based methods. + +We combine the aforementioned TTA strategies with the prior knowledge of medical images and design a novel Test-Time Domain Generalization (TTDG) framework utilizing graph structures. The construction of graph enables a more robust representation of the morphological priors inherent to medical images. In complex medical scenarios, the aggregation mechanism of nodes and edges in graph ensures the effective preservation of crucial contextual information. Previous research has utilised pairwise graph matching [26, 42, 46] to address domain shifts in medical imaging. However, these methods are limited to aligning features between two domains, which poses challenges in scenarios involving multiple domains commonly encountered in real-world settings. Multi-graph matching overcomes these limitations by establishing connections between multiple graphs and integrating information from diverse domains, thereby enabling the capture of more intricate patterns and structural relationships while maintaining cross-domain consistency. This process facilitates the acquisition of global invariant features from medical images. + +In multi-graph matching, the universe of nodes [39, 55] represents the full-set of the nodes from all graphs (refer to Fig. 1). The aforementioned morphological priors are embedded into this universe, derived from a multitude of hospitals, devices, and modalities. It offers three main advantages: (1) Cross-domain consistency. By mapping the nodes from different graphs into the universe, we ensure that all graphs are aligned according to the same reference standard, i.e. the priors. (2) Cycle-consistency in multi-graph matching. This guarantees global coherence across multiple graphs, avoiding conflicts in local node alignments. (3) Handling partial matching and missing nodes. Even when certain nodes are absent in some graphs, matching them to virtual nodes in the universe addresses the issue, enhancing robustness to incomplete graph structures. + +Our contributions can be summarized as follows: + +- We propose the first (to our best knowledge) multi-graph matching framework for test-time domain generalization in medical image segmentation tasks, effectively mitigating performance degradation caused by domain shifts during the testing phase. +- By designing learnable universe embeddings, we inte + +grate the morphological priors of medical images into the graph matching process during source training, while ensuring the cycle-consistency constraint. This approach enables joint optimization and promotes the learning of domain-invariant features. + +- We design a novel, well-initialized unsupervised testing adaptation paradigm that integrates prior knowledge and allows seamless deployment during adaptation, effectively addressing domain shifts. +- We conduct extensive experiments on two typical medical datasets, demonstrating that our method performs competitively against state-of-the-art approaches for both multi-source and single-source domain generalization. + +# 2. Preliminary & Related Work + +The following section will introduce the preliminaries of multi-graph matching, while also reviewing the most relevant existing literature on the subject. Subsequently, domain generalization and test-time adaptation will be discussed. + +# 2.1. Multi-Graph Matching + +We consider $m \in \mathbb{N}$ different graphs $\mathcal{G}_1, \mathcal{G}_2, \dots, \mathcal{G}_m$ , $i \in [m]$ , $[m] := \{1, \dots, m\}$ . For each $\mathcal{G}_i = (\mathcal{V}_i, \mathcal{A}_i)$ , $\mathcal{V}_i \in \mathbb{R}^{n_i \times h}$ represents the $h$ -dimensional feature of $n_i$ nodes, and adjacency matrix $\mathcal{A}_i \in \mathbb{R}^{n_i \times n_i}$ encodes the connectivity between nodes, represented as the set of edges. For any two graphs $\mathcal{G}_i$ and $\mathcal{G}_j$ , the set of $n_i \times n_j$ partial permutation matrices $\mathbb{P}_{n_in_j}$ is defined as + +$$ +\mathbb {P} _ {n _ {i} n _ {j}} := \left\{\mathbf {X} \in \{0, 1 \} ^ {n _ {i} \times n _ {j}}: \mathbf {X} \mathbf {1} _ {n _ {j}} \leq \mathbf {1} _ {n _ {i}}, \mathbf {X} ^ {\top} \mathbf {1} _ {n _ {i}} \leq \mathbf {1} _ {n _ {j}} \right\}, \tag {1} +$$ + +where $\mathbf{1}_{n_i}$ denotes a $n_i$ -dimensional column vector whose elements are all ones. The assignment matrix $\mathbf{X}_{ij} \in \mathbb{P}_{n_i n_j}$ between a pair of graphs $(\mathcal{G}_i, \mathcal{G}_j)$ denotes a meaningful correspondence that encodes the matching. + +When considering the synchronisation matching [30, 39] of multiple graphs ( $m > 2$ ), relying on local matches between graph pairs can readily result in erroneous correspondences that contradict each other globally [63]. Consequently, the multi-graph matching has been addressed in terms of simultaneously solving under the constraints of cycle-consistency [64, 65]. + +Definition 1 Cycle-consistency. The matching among $\mathcal{G}_1, \mathcal{G}_2, \dots, \mathcal{G}_m$ is cycle-consistent (partial transitivity), if + +$$ +\mathbf {X} _ {i k} \mathbf {X} _ {k j} \leq \mathbf {X} _ {i j}, \quad \forall i, j, k \in [ m ]. \tag {2} +$$ + +In contrast to full matching, i.e. in Eq. (2) the inequalities become equalities and $n_i = n_j$ , partial matching requires only that the pairwise matching combinations in the cycle form a subset of the identity matching [3]. An iterative refinement strategy [34] was employed to enhance pairwise + +matchings by ensuring global mapping consistency. In [39], the authors proposed a method to achieve cycle-consistency based on spectral decomposition of the assignment matrix. In addition, methods such as convex programming [7], low-rank matrix recovery [76], and spectral decomposition [30] have been proposed to address the synchronization of partial matchings. + +Instead of explicitly modeling the cubic number of nonconvex quadratic constraints, a more efficient approach to enforcing cycle consistency is the use of the universe of nodes [55]. In a universe comprising nodes of size $d \in \mathbb{N}$ , the matching between graphs can be decomposed by matching each graph $\mathcal{G}_i$ to the space of universe (refer to Fig. 1). We denote the universe matchings as follows: + +$$ +\mathbb {U} _ {n _ {i} d} := \left\{\mathbf {U} \in \{0, 1 \} ^ {n _ {i} \times d}: \mathbf {U} \mathbf {1} _ {d} = \mathbf {1} _ {n _ {i}}, \mathbf {U} ^ {\top} \mathbf {1} _ {n _ {i}} \leq \mathbf {1} _ {d} \right\}. \tag {3} +$$ + +Lemma 1 Cycle-consistency, universe matching. The pairwise (partial) matching matrices $\{\mathbf{X}_{ij}\}_{i,j=1}^{m}$ is cycle-consistent iff there exists a collection of universe matching $\{\mathbf{U}_i \in \mathbb{U}_{n_id}\}_{i=1}^m$ , such that for each $\mathbf{X}_{ij}$ , it holds that $\mathbf{X}_{ij} = \mathbf{U}_i \mathbf{U}_j^\top$ . + +The proof will be presented in the appendix. By projecting each node of a graph to the universe of nodes and identifying the universe matchings $\mathbb{U}_{n_i d}$ , it is ensured that the cycle-consistency constraint will be satisfied throughout the multi-graph matching process. In [59, 61], the authors proposed a gradual assignment procedure to achieve soft matching and clustering through the universe matching. In [4], the universe of nodes is leveraged to incorporate geometric consistency, ensuring both point scaling and convergence. [36] utilized the universe matching to facilitate effective partial multi-graph matching. + +Definition 2 Multi-Matching Koopmans-Beckmann's Quadratic Assignment Problem (KB-QAP) [21]. Multi-graph matching is formulated with KB-QAP, by summing KB-QAP objectives among all pairs of graphs: + +$$ +\begin{array}{l} \max _ {\mathbf {X} _ {i j}} \sum_ {i, j \in [ m ]} (\lambda \cdot t r \left(\mathbf {X} _ {i j} ^ {\top} \mathcal {A} _ {i} \mathbf {X} _ {i j} \mathcal {A} _ {j}\right) + t r \left(\mathbf {X} _ {i j} ^ {\top} \mathbf {M} _ {i j}\right)), \tag {4} \\ s. t. \quad \mathbf {X} _ {i j} \in \{0, 1 \} ^ {n _ {i} \times n _ {j}}, \mathbf {X} _ {i j} \in \mathbb {P} _ {n _ {i} n _ {j}}, \\ \mathbf {X} _ {i k} \mathbf {X} _ {k j} \leq \mathbf {X} _ {i j}, \quad \forall i, j, k \in [ m ]. \\ \end{array} +$$ + +In Eq. (4), $\lambda$ is a scaling factor for edge-to-edge similarity, and $\mathbf{M}_{ij}$ is the node-to-node similarity between $\mathcal{G}_i,\mathcal{G}_j$ . + +# 2.2. Domain Generalization + +Domain generalization addresses a challenging scenario where one or more different but related domains are provided, with the objective of training a model that can generalize effectively to an unseen test domain [57, 68, 75]. Medical image analysis encounters significant challenges due to factors such as variability in image appearance, the + +complexity and high dimensionality of the data, difficulties in data acquisition, and issues related to data organization, labeling, safety, and privacy. Previous research has proposed DG algorithms across multiple levels, including data-level [69, 71], feature-level [5, 24], model-level [16, 40, 78], and analysis-level [28, 43]. Yu et al. [69] proposed a U-Net z-score normalization network for the stroke lesion segmentation. Liu et al. [28] employed dictionary learning for prostate and fundus segmentation by creating a shape dictionary composed of template masks. In [5], the authors utilized mutual information to differentiate between domain-invariant features and domain-specific ones in ultrasound image segmentation. Gu et al. [16] proposed a domain-style contrastive learning approach that disentangles an image into invariant representations and style codes for DG. + +# 2.3. Test-Time Adaptation + +Test-time adaptation is an emerging paradigm that allows a pre-trained model to adapt to unlabeled data during the testing phase, before making predictions [27]. Several TTA methods have been proposed recently, utilizing techniques such as self-supervised learning [6, 22, 73], batch normalization calibration [8, 32, 47, 70, 77], and input data adaptation [19, 49] to achieve better test-time performance. Typically, VPTTA [8] is a method that freezes the pre-trained model and generates low-frequency prompts for each image during inference in medical image segmentation. In [19], the authors proposed designing two CNN-based sub-networks along with an image normalization network. During test-time training, the image normalization network was adapted for each image. [70] proposed a dynamic mixture coefficient and a statistical transformation operation to adaptively merge the training and testing statistics of the normalization layers. Additionally, the authors design an entropy minimization loss to address the issue of domain shifts. + +# 3. Methodology + +This section provides an overview of how multi-graph matching methods can facilitate training for TTDG tasks. Given inputs consisting of images from $S$ source domains $(S\geq 1)$ , denoted as $\mathcal{D}_s = \{D_1,D_2,\dots ,D_S\}$ , the objective is to enable the model to make more accurate predictions on $T$ unseen target domains $(T\geq 1)$ , denoted as $\mathcal{D}_t = \{D_1,D_2,\dots ,D_T\}$ , during the testing phase. + +Inspired by [41], we aim to effectively leverage morphological prior through graph construction. However, unlike the UDA tasks, which typically involve only two domains, our training process simultaneously handles multiple labeled domains. Pairwise matching across multiple domains can lead to the challenges discussed in Sec. 1. To tackle this, we propose learning domain-invariant latent representations from a multi-graph matching perspective, thereby + +![](images/fcc7025cac673c8508ad7b044f317c883c4e41b086a5cf6eb8ec29e81f6e6d46.jpg) +Figure 2. Overview of our TTDG framework. During source model training, data from different domains are jointly used to train the Segmentation Network (feature extractor and segmentation head). Feature maps and ground truth masks are utilized to construct graphs $\mathcal{G}_i$ ( $i = 4$ in this figure legend) and corresponding labels $Y_i$ (Sec. 3.1), with universe embeddings learned via back-propagation, incorporating morphological priors (Sec. 3.2). At test time, multi-graph matching is performed on all target domain images in each batch. Despite style differences, these images share common morphological patterns. Universe embeddings are frozen as prior knowledge to guide the matching, and the segmentation network is fine-tuned via back-propagation for efficient adaptation (Sec. 3.3). + +facilitating rapid adaptation to unseen domains during the TTA phase. + +# 3.1. Construction of Graph + +Rather than using conventional graph-matching algorithms that depend on explicitly defined keypoints as graph nodes [14, 58, 60], accurately annotating keypoints in medical imaging is both more costly and less practical than in natural images. Therefore, given a mini-batch $\{x_i\}_{i=1}^m \in \mathcal{D}_s$ , as shown in Fig. 2, it is first passed through a general feature extractor, such as ResNet [17], to obtain visual features. We then perform spatially-uniform sampling [25] of pixels within the ground-truth masks at each feature level to obtain $n_i$ foreground nodes with its corresponding labels $Y_i \in \mathbb{Z}^{n_i}$ . After extracting these fine-grained visual features, we employ a nonlinear projection to transform the + +visual space into the $h$ -dimensional graph space. This approach generates the feature of nodes $\{\mathcal{V}_i \in \mathbb{R}^{n_i \times h}\}_{i=1}^m$ that more effectively preserve the semantic characteristics. + +For the weighted adjacency matrix encoding structural information, we use Droppedge [44] to reduce potential bias caused by the frequent visual correspondences, preventing the model from over-relying on specific matches. The formulation is as follows: + +$$ +\mathcal {A} _ {i} = \operatorname {D r o p e d g e} \left\{\text {s o f t m a x} \left[ \mathcal {V} _ {i} \mathcal {W} _ {x} \cdot \left(\mathcal {V} _ {i} \mathcal {W} _ {y}\right) ^ {\top} \right] \odot D ^ {- 1} \right\}, \tag {5} +$$ + +where $\mathcal{W}_x$ and $\mathcal{W}_y$ are two learnable linear projection, $D = \frac{\mathcal{V}_i\mathcal{V}_i^{\mathrm{T}}}{||\mathcal{V}_i||_2}$ is the cosine distance matrix. Edges with larger weights indicate nodes that are closer to each other, so we use the inverse of $D$ and apply the hadamard product $\odot$ to combine node similarity with geometric proximity. The di + +agonal elements of $\mathcal{A}_i$ are set to zero. Based on the above steps, we obtain a corresponding graph $\mathcal{G}_i = (\mathcal{V}_i,\mathcal{A}_i)$ for each input $x_{i}$ . + +# 3.2. Formulation of the Universe Embeddings + +In Sec. 2.1, it was demonstrated that by matching each node of a graph to the universe of nodes, we can maintain cycle-consistency while avoiding the computational burden associated with cubic non-convex constraints. While a randomly initialized matrix can learn accurate universe matchings $\mathbb{U}_{n_i d}$ in a supervised setting, we introduce universe embeddings $\mathcal{U} \in \mathbb{R}^{d \times h}$ , which serve as learnable latent representations to enhance adaptability in unsupervised multi-domain scenarios. + +$\mathcal{U}$ is initialized as $1 / d + 10^{-3}z$ where $z\sim N(0,1)$ , following the setting in [59]. For each $\mathcal{G}_i$ , we learn a unique matching between nodes in $\nu_{i}$ and the corresponding in the universe of nodes, as follows: $U_{i} = \mathrm{Sinkhorn}(\mathcal{V}_{i}\mathcal{U}^{\top},\tau)\in$ $\mathbb{U}_{n_i d}$ . Sinkhorn $(X,\tau)$ refers to the Sinkhorn algorithm [9, 51], which applies a relaxed projection with entropic regularization [15] to normalize the matrix $X$ , resulting in a doubly-stochastic matrix, and $\tau \in (0, + \infty)$ is the regularization factor. For the obtained $U_{i}$ , the $d\times d$ matrix $U_{i}^{\top}\mathcal{A}_{i}U_{i}$ is a row and column reordering of $\mathcal{A}_i$ based on the universe of nodes. Accordingly, the agreement between two adjacency matrices $\mathcal{A}_i$ and $\mathcal{A}_j$ , reordered according to their respective universe matching assignment matrices $U_{i}$ and $U_{j}$ , can be quantified using the Frobenius inner product namely $\langle U_i^\top \mathcal{A}_iU_i,U_j^\top \mathcal{A}_jU_j\rangle$ . Let $\mathbf{U} = [U_1^\top ,\dots ,U_m^\top ]^\top \in \mathbb{R}^{n\times d}$ where $n = \sum_{i = 1}^{m}n_{i}$ . The block-diagonal multi-adjacency matrix is defined as $\mathbf{A} = diag(\mathcal{A}_1,\dots ,\mathcal{A}_m)\in \mathbb{R}^{n\times n}$ . We can get the multi-matching formulation as: + +$$ +\begin{array}{l} f(\mathbf{U}):= \max_{U_{1},\ldots ,U_{m}}\quad \sum_{i,j\in [m]}\langle U_{i}^{\intercal}\mathcal{A}_{i}U_{i},U_{j}^{\intercal}\mathcal{A}_{j}U_{j}\rangle \\ = \max _ {\mathbf {U} \in \mathbb {U}} \operatorname {t r} \left(\mathbf {U} ^ {\top} \mathbf {A} \mathbf {U} \mathbf {U} ^ {\top} \mathbf {A} \mathbf {U}\right). \tag {6} \\ \end{array} +$$ + +It was observed that in the original multi-graph matching optimization problem, graphs are typically constructed for each individual instance, with all the nodes within a graph belonging to the same category. However, in medical imaging, constructing separate graphs for each organ would not only greatly increase the computational complexity of graph matching but also result in misalignments between graphs from different categories, significantly compromising matching accuracy. To address this limitation, we introduce a class-wise similarity matrix $\mathbf{W}$ to incorporate class-aware label information for each node. This ensures that the nodes are correctly aligned according to their respective categories during multi-graph matching. Specifically, we define $W_{ij} = Y_iY_j^{\top}$ and $\mathbf{W} = [W_{ij}]_{ij} \in \mathbb{Z}^{(\sum_{i=1}^{m}n_i)\times(\sum_{j=1}^{m}n_j)}$ . The class-aware similarity matrix $\tilde{\mathbf{A}} = \mathbf{W}^{\top}\mathbf{A}\mathbf{W}$ . Consequently, Eq. (6) can be transformed + +into the following form + +$$ +\max _ {\mathbf {U} \in \mathbb {U}} \operatorname {t r} \left(\mathbf {U} ^ {\top} \tilde {\mathbf {A}} \mathbf {U} \mathbf {U} ^ {\top} \tilde {\mathbf {A}} \mathbf {U}\right). \tag {7} +$$ + +The higher-order projected power iteration (HiPPI) [4] is employed to iteratively solve $\mathbf{U}$ to obtain a stable convergence form: + +$$ +U _ {i} ^ {*} = U _ {i} ^ {t} \quad s. t. \quad | | U _ {i} ^ {t} - U _ {i} ^ {t - 1} | | < \theta , \quad \forall i \in [ m ]. \tag {8} +$$ + +Empirically, we set the convergence threshold $\theta = 10^{-5}$ . In order to guarantee the optimal learning of universe embeddings $\mathcal{U}$ , we incorporate the non-negative matrix factorisation (NMF) [3, 23] and define the overall optimization objective as: + +$$ +L (\mathcal {U}) = \sum_ {i \in [ m ]} | | U _ {i} ^ {*} - \mathcal {V} _ {i} \mathcal {U} ^ {\mathrm {T}} | | _ {F} ^ {2} + \alpha | | \mathcal {U} | | _ {F} ^ {2}, \tag {9} +$$ + +where $||\cdot||_F^2$ represents the Frobenius norm, and $\alpha$ is a tuning parameter. Since only matrix multiplication and element-wise division are involved, the Sinkhorn operation is fully differentiable. This allows for effective updates to $\mathcal{U}$ through back-propagation, enabling the incorporation of morphological prior knowledge. + +# 3.3. Testing Paradigm on the Target Domains + +During the Test-Time Adaptation (TTA) phase, we only have access to unlabeled data from the unseen target domains $\mathcal{D}_t$ . Due to distribution shifts, the performance of models trained on $\mathcal{D}_s$ can degrade significantly. To address this issue, we employ the multi-graph matching on each mini-batch $\{x_i^t\}_{i=1}^m \in \mathcal{D}_t$ . By utilizing universe embeddings $\mathcal{U}$ learned from $\mathcal{D}_s$ as described in Sec. 3.2, it is guaranteed cycle-consistency, enabling the network to focus on domain-invariant features guided by morphological priors from medical images. In this phase, $\mathcal{U}$ is frozen to prevent error accumulation from affecting the prior information. The feature extractor and segmentation head (segmentation network in Fig. 2) are optimized through gradient back-propagation, allowing the model to adapt effectively to the new domains. + +Unsupervised Multi-Matching. Following the same graph construction described in Sec. 3.1, we obtain the $\mathcal{G}_i^t = (\mathcal{V}_i^t,\mathcal{A}_i^t)$ for each $x_{i}^{t}$ , where $\mathcal{V}_i^t\in \mathbb{R}^{n_i\times h}$ represents the $n_i$ class-aware nodes and $\mathcal{A}_i^t\in \mathbb{R}^{n_i\times n_i}$ is the adjacency matrix encoding the structural information. We first introduce a network-learned affinity matrix $M_{ij} = f_{mlp}\{\mathcal{V}_i^t\mathcal{W}_x^t$ $(\mathcal{V}_j^t\mathcal{W}_y^t)^\top \} \in \mathbb{R}^{n_i\times n_j}$ in graph pair $\mathcal{G}_i$ and $\mathcal{G}_j$ , which computes the similarity between their corresponding node features. $\mathcal{W}_x^t$ and $\mathcal{W}_y^t$ are two learnable linear projection, $f_{mlp}$ is a multi-layer perception (MLP). Subsequently, the matching between $\mathcal{G}_i$ and the universe of size $d$ is obtained through the inner-product between $\mathcal{V}_i^t$ and the pre-trained universe embeddings $\mathcal{U}$ , in accordance with the same procedure $U_{i} = Sinkhorn(\mathcal{V}_{i}^{t}\mathcal{U}^{\top},\tau)\in \mathbb{U}_{n_{i}d}$ in Sec. 3.2. + +In [59, 61], the authors propose solving the multimatching KB-QAP problem presented in Eq. (4) (denoted as $J$ ) using a Taylor expansion. Based on the assumptions in Lemma 1 of Sec. 2.1, we can reformulate $J$ in Taylor series as follows: + +$$ +J (U) \approx J \left(U ^ {0}\right) + \sum_ {i} \frac {\partial J}{\partial U _ {i}} \bigg | _ {U = U ^ {0}} \left(U _ {i} - U _ {i} ^ {0}\right) + \dots . \tag {10} +$$ + +$J$ can be approximated as $m$ linear optimization problems through the first-order Taylor expansion, which can be efficiently solved using the Sinkhorn. The gradient of $J$ with respect to $U_{i}$ , denoted as + +$$ +V _ {i} = \frac {\partial J}{\partial U _ {i}} = \sum_ {j \in [ m ]} \left(\lambda \mathcal {A} _ {i} U _ {i} U _ {j} ^ {\mathrm {T}} \mathcal {A} _ {j} U _ {j} + M _ {i j} U _ {j}\right), \tag {11} +$$ + +$V_{i}$ ensures that $U_{i}$ gradually converges to a high-quality discrete solution. + +Finally, the converged $\{U_i\}$ fine-tune the segmentation network with the matching loss $L_{mat}$ : + +$$ +\begin{array}{l} L _ {m a t} = \sum_ {i, j \in [ m ]} \left[ - \hat {M} _ {i j} ^ {\gamma} (1 - U _ {i} U _ {j} ^ {\mathsf {T}}) \log (U _ {i} U _ {j} ^ {\mathsf {T}}) \right] \\ - \sum_ {i, j \in [ m ]} \left[ (1 - \hat {M} _ {i j}) ^ {\gamma} U _ {i} U _ {j} ^ {\mathsf {T}} \log \left(1 - U _ {i} U _ {j} ^ {\mathsf {T}}\right) \right], \tag {12} \\ \end{array} +$$ + +where $\hat{M}_{ij} = \text{sinkhorn}(M_{ij},\tau)$ , and $\gamma$ is an amplification factor. See algorithm in appendix for full details. + +# 4. Experiments + +# 4.1. Datasets + +The retinal fundus segmentation datasets comprise five public datasets from different medical centers, denoted as Site A (RIM-ONE [13]), B (REFuge [37]), C (ORIGA [74]), D (REFuge-Test [37]), and E (Drishti-GS [52]), with 159, 400, 650, 800, and 101 images respectively, all consistently annotated for optic disc (OD) and optic cup (OC) segmentation. We adopted the preprocessing method outlined in [8, 28], where each image's region of interest is cropped to $800 \times 800$ pixels and normalized using min-max normalization. + +The polyp segmentation datasets consist of four public datasets collected from different medical centers, denoted as Site A (BKAI-IGH-NEOPolyp [33]), B (CVC-ClinicDB/CVC-612 [2]), C (ETIS [50]), and D (Kvasir [18]), containing 1000, 612, 196, and 1000 images, respectively. We followed the preprocessing steps in [8], resizing the images to $800 \times 800$ pixels and normalizing them using ImageNet-derived statistics. + +# 4.2. Experimental Setup + +Implementation Details. To ensure a fair comparison across all methods, we followed the protocol in [8] and + +split each dataset into an $8:2$ ratio for training and testing. We employed the ResNet-50 [17] pre-trained on ImageNet as the feature extractor. The training was optimized using the SGD optimizer with a momentum of 0.9 and a learning rate of 0.001. For source model training, we set the mini-batch size to 8, allowing simultaneous matching of 8 graphs per batch. The universe embedding $\mathcal{U}$ was treated as an additional trainable tensor, integrated into the network weights without influencing subsequent experiments with other methods. During the TTA phase, all methods were trained on the target data without access to their labels. The mini-batch size was set to 4, and the previously learned $\mathcal{U}$ from source model training was frozen. All experiments were implemented using PyTorch and conducted on 4 NVIDIA 3090 GPUs. + +Evaluation Metrics. The Dice score (DSC, $\%$ ) was used to quantify the accuracy of the predicted masks, serving as the primary metric for evaluating segmentation performance. Additionally, the enhanced alignment metric $E_{\phi}^{max}$ [12] was adopted to measure both pixel-level and global-level similarity. To further assess the consistency between predictions and ground truths, we also employed the structural similarity metric $S_{\alpha}$ [11]. + +# 4.3. Experimental Results + +We selected U-Net [45] as the benchmark model for the segmentation task, training it on the source domains and testing on the target domain without adaptation (No Adapt). In addition, we compared eight state-of-the-art (SOTA) methods, including two entropy-based approaches (TENT [56] and SAR [35]), a shape template-based method (TASD [28]), a dynamically adjusted learning rate method (DLTTA [66]), batch normalization-based methods (DomainAdaptor [70], VPTTA [8]), and noise estimation-based approaches (DeY-Net [62] and NC-TTT [38]). + +Multi-Source Generalization. In these experiments, we adopted a leave-one-out training strategy [6] (i.e., with $S = |\mathcal{D}_s \cup \mathcal{D}_t| - 1$ and $T = 1$ ). Although some methods, such as TASD [28], DeY-Net [62] and VPTTA [8], were originally designed for single-source domain training, we simulated single-domain conditions by mixing multi-source domain data, making the experimental setup still feasible for these approaches. Furthermore, we have also included experiments specifically focused on single-source domains later in the study. The segmentation results for the fundus and polyp datasets are shown in Tables 1 and 2, respectively. For OD/OC segmentation, all TTA methods outperformed the No Adapt baseline, highlighting their effectiveness in addressing domain shifts. Comparatively, our method consistently achieved better average performance than all other approaches. Specifically, for the key segmentation metric DSC, our method exceeded the second-best (DeY-Net [62]) by $2.88\%$ and outperformed the No Adapt by $19.09\%$ . A + +Table 1. Multi sources domain generalization in retinal fundus segmentation. The average performance (mean ± standard deviation) of three trials for our method and nine SOTA methods. "Site A" means training on Sites B-E and testing on Site A, and similarly for the others. The best results are highlighted in red. + +
MethodsSite ASite BSite CSite DSite E
\(DSC\)\(E_{\text{max}}^{0}\)\(S_{\text{a}}\)\(DSC\)\(E_{\text{max}}^{0}\)\(S_{\text{a}}\)\(DSC\)\(E_{\text{max}}^{0}\)\(S_{\text{a}}\)\(DSC\)\(E_{\text{max}}^{0}\)\(S_{\text{a}}\)\(DSC\)\(E_{\text{max}}^{2}\)\(S_{\text{a}}\)
No Adapt (U-Net [45])65.60±5.7891.75±0.1379.11±0.0774.65±4.8888.78±0.0883.31±0.0263.15±7.3081.03±0.1979.77±0.0368.11±5.4989.46±0.2082.12±0.0775.34±1.0190.68±0.1687.53±0.05
TENT (ICLR'21) [56]75.22±3.9994.10±0.1083.97±0.0581.06±0.9393.45±0.1192.36±0.0375.23±2.3384.62±0.1478.12±2.5991.33±0.3186.41±0.1283.17±0.7892.74±0.1093.64±0.0378.5691.9488.20
TASD (AAAI'22) [28]79.70±2.5288.73±0.4583.67±0.1782.96±2.1490.01±0.2090.99±0.1186.12±1.1193.59±0.1387.08±0.0778.02±6.3087.20±0.1687.94±0.0483.71±3.7790.88±0.2590.37±0.1282.1090.8888.01
DLTTA (TMI'22) [66]71.88±5.5986.29±0.2180.70±0.1380.12±2.2690.53±0.1788.63±0.0681.46±3.3593.08±0.0887.48±0.0275.90±7.2784.27±0.1284.79±0.0682.75±6.5590.10±0.3189.33±0.1478.4288.8586.18
SAR (ICLR'23) [35]75.29±2.6988.73±0.3480.15±0.2686.42±1.5796.88±0.2791.28±0.0980.69±4.3092.10±0.1586.31±0.2780.27±5.4789.98±0.3187.20±0.1589.61±0.2895.90±0.0892.06±0.0182.4692.7187.40
DomainAdaptor (CVPR'23) [70]77.14±1.2190.80±0.1981.31±0.0283.95±0.8093.17±0.0990.44±0.0882.79±1.0995.36±0.1088.57±0.1386.50±3.2995.26±0.3391.07±0.2086.24±1.4494.90±0.2993.77±0.1883.3293.9089.03
DeY-Net (WACV'24) [62]79.67±1.3589.14±0.2383.68±0.0490.07±3.1290.06±0.3392.31±0.0785.94±1.8593.54±0.2287.15±0.1083.47±1.4095.29±0.1788.70±0.0488.76±2.2895.43±0.2190.45±0.0885.5894.0988.46
VPTTA (CVPR'24) [8]77.85±1.4094.09±0.0183.97±0.0584.57±0.9995.38±0.0389.53±0.0983.61±1.6796.39±0.0489.64±0.0582.75±0.5096.53±0.0489.83±0.0189.74±1.5198.22±0.2494.77±0.0783.7096.1289.55
NC-TTT (CVPR'24) [38]80.83±1.7789.61±0.2384.41±0.0887.07±1.3195.63±0.0690.09±0.0286.93±2.2794.69±0.1188.39±0.0382.10±4.2891.49±0.2286.33±0.0790.08±0.0995.93±0.1091.53±0.0985.4093.4788.15
Ours82.53±1.5291.85±0.2285.50±0.2088.98±0.8996.69±0.0491.63±0.1288.73±0.7095.80±0.0689.71±0.0290.25±1.3397.97±0.1992.89±0.1391.83±0.7197.14±0.2092.82±0.0988.4695.8490.51
+ +Table 2. Multi sources domain generalization in the polyp segmentation. The average performance (mean ± standard deviation) of three trials for our method and nine SOTA methods. "Site A" means training on Sites B-D and testing on Site A, and similarly for the others. The best results are highlighted in red. + +
MethodsSite ASite BSite CSite DAverage
\(DSC\)\(E_{\phi}^{max}\)\(S_{\alpha}\)\(DSC\)\(E_{\phi}^{max}\)\(S_{\alpha}\)\(DSC\)\(E_{\phi}^{max}\)\(S_{\alpha}\)\(DSC\)\(E_{\phi}^{max}\)\(S_{\alpha}\)\(DSC\uparrow\)\(E_{\phi}^{max}\)\(S_{\alpha}\)
No Adapt (U-Net [45])82.67±2.2492.11±0.0789.59±0.1081.20±4.3892.46±0.2086.08±0.0678.33±1.6294.00±0.0885.35±0.0175.51±2.7485.90±0.1781.71±0.2179.4391.1285.68
TENT (ICLR'21) [56]78.69±1.1288.28±0.1485.44±0.0877.71±3.0989.73±0.0981.90±0.1280.19±3.3596.03±0.0487.80±0.1772.63±5.1182.47±0.3078.92±0.1477.3189.1383.52
TASD (AAAI'22) [28]84.60±4.1492.55±0.2190.76±0.1085.28±2.2093.10±0.1386.83±0.1982.43±5.3996.31±0.0889.61±0.0280.04±4.4188.26±0.1184.09±0.0383.0992.5687.82
DLTTA (TMI'22) [66]80.68±2.3790.01±0.2186.89±0.0980.86±1.7791.98±0.1085.83±0.0475.57±3.1293.90±0.0884.15±0.1074.25±3.4083.49±0.1279.44±0.1077.8489.8584.08
SAR (ICLR'23) [35]79.37±0.7788.40±0.1286.11±0.0378.40±2.0790.23±0.1183.39±0.0280.94±0.9095.83±0.1286.84±0.0778.23±2.9186.37±0.0981.66±0.1479.2490.2184.50
DomainAdaptor (CVPR'23) [70]88.58±0.9695.03±0.0490.95±0.0181.12±1.0792.31±0.0886.60±0.0281.77±2.1795.18±0.0387.19±0.0179.91±1.3387.23±0.0483.64±0.0582.8592.4487.10
DeY-Net (WACV'24) [62]83.61±2.3190.79±0.1189.25±0.0980.90±3.1791.49±0.1085.36±0.0380.46±1.4395.83±0.0887.14±0.0373.85±3.0483.00±0.1378.96±0.0879.7190.2885.18
VPTTA (CVPR'24) [8]82.34±0.6991.03±0.0888.35±0.0184.32±0.4491.53±0.0787.46±0.0284.43±0.5295.75±0.1688.91±0.0382.40±0.2989.60±0.0485.18±0.0183.3791.9787.47
NC-TTT (CVPR'24) [38]87.96±1.2293.71±0.1390.66±0.0783.70±2.0893.78±0.0987.51±0.0789.29±0.9096.66±0.2291.26±0.0778.23±0.7986.37±0.0981.66±0.0684.8092.6387.77
Ours90.97±0.6696.57±0.0692.92±0.0887.49±0.8295.21±0.0489.93±0.0986.87±0.5897.53±0.0590.16±0.1283.97±1.1890.67±0.1386.44±0.0887.3294.6789.86
+ +similar trend was observed in the polyp segmentation results, where we outperformed the second-best method (NC-TTT [38]) and No Adapt by $2.52\%$ and $7.89\%$ , respectively. Additionally, our approach exhibited greater stability across multiple trials compared to the other methods. + +Single-Source Generalization. In these experiments, the models are trained on one domain and tested on the remaining datasets (i.e., with $S = 1$ and $T = |\mathcal{D}_s \cup \mathcal{D}_t| - 1$ ). Compared to multi-source generalization, single-source training is considered more challenging due to the limited domain information available during the training phase. By comparing Tables 2 and 3, we can observe that all models show lower segmentation performance (in terms of DSC) in the single-source setting compared to the multi-source results. Some methods even perform worse than the No Adapt baseline in single-source scenarios. However, our method still achieves the best average performance across 12 domain transfer experiments, outperforming the second-best approach (NC-TTT [38]) by $1.83\%$ and exceeding the No Adapt by $9.77\%$ in terms of DSC. + +Fig. 3 presents the visualization results for "Site A" in Table 1. The No Adapt baseline exhibits significant misalignment, with notably distorted shapes and inaccurate boundaries. VPTTA offers some improvement but still suffers from distortions and incomplete segmentation, particularly in the optic cup area. NC-TTT shows instances of overlapping segmentation. In contrast, our method delivers the most precise segmentation, closely aligning with the ground truth. The Grad-CAM [48] visualizations reinforce this, as the attention of network is focused specifically on the relevant regions. These results demonstrate our method's enhanced generalization to unseen domains by effectively incorporating morphological priors, resulting in higher seg + +![](images/ba7ca63d2bcda0fe288ae5adeb346a5e1ba7f633574dc4083f220e2b87ae6eb5.jpg) +Figure 3. Visualization comparison of segmentation results and Grad-CAM outputs from the final layer of the backbone network for the No Adapt baseline, VPTTA [8], NC-TTT [38], and our proposed method on retinal fundus images. Additional visual comparisons are provided in the supplementary material. + +mentation accuracy. For further visualization and additional experimental results, please refer to the appendix. + +# 5. Analysis + +# 5.1. Effectiveness of the Universe Embeddings + +The universe embeddings $\mathcal{U}$ play a crucial role in incorporating morphological priors from medical images, resulting in an assignment matrix that projects each node to the universe of nodes. To validate the effectiveness during the TTA, we conducted the following experiments: (1) Without universe embeddings (denoted as $w / o\mathcal{U}$ ), the universe matching assignment matrix $\mathbf{U}$ is initialized following the setting in [59] (i.e. $\mathbf{U} = 1 / d + 10^{-3}z$ , where $z\sim N(0,1)$ ), leading to random matching between the nodes and the universe; (2) Without morphological priors but with $\mathcal{U}$ (denoted as $w / o$ priors). As shown in Table 4, the performance in (1) is slightly better than in (2) by $1.76\%$ in terms of DSC + +Table 3. Single source domain generalization in the polyp segmentation. The average performance of three trials for our method and nine SOTA methods. A → B represents models trained on Site A and tested on Site B, and similar for others. Best results are colored as red. + +
MethodsA→BA→CA→DB→AB→CB→DC→AC→BC→DD→AD→BD→CAvg.
Dice Score Metric ↑ (DSC, mean±std)
No Adapt (U-Net [45])76.49±1.4867.70±3.2770.94±7.5676.21±2.9062.80±8.0370.03±4.5372.87±3.6269.13±5.1570.44±3.2779.78±1.0872.47±2.8074.91±3.1871.98
TENT (ICLR'21) [56]68.13±0.1865.99±0.1166.11±0.1477.56±0.2159.87±0.2068.89±0.1473.21±0.1065.04±0.0972.35±0.2277.88±0.1470.01±0.1173.60±0.2069.89
TASD (AAAI'22) [28]73.02±2.2479.67±6.3778.23±4.4466.82±9.1772.09±4.4077.95±1.6974.20±3.3069.74±7.2776.13±2.8283.77±0.5176.09±3.2680.10±1.4775.65
DLTTA (TMI'22) [66]72.23±0.0470.78±0.0674.56±0.0464.87±0.0160.05±0.0978.47±0.0371.22±0.0365.78±0.1071.00±0.1379.89±0.0879.10±0.1074.52±0.0171.87
SAR (ICLR'23) [35]70.34±0.1377.90±0.2077.71±0.0968.19±0.1266.56±0.0873.92±0.1376.12±0.1171.29±0.1975.08±0.0882.07±0.0480.44±0.2183.10±0.1575.22
DomainAdaptor (CVPR'23) [70]78.39±0.6177.09±0.3579.45±0.2174.98±0.7070.07±0.6680.21±0.4676.02±0.3373.40±0.5081.33±0.5486.10±0.4380.56±0.6277.86±0.2177.96
DeY-Net (WACV'24) [62]74.45±1.3176.07±2.5280.11±0.7270.90±1.1069.23±2.1876.31±1.3278.33±4.4374.60±1.6673.52±2.2580.70±0.2778.23±1.1781.48±0.9376.16
VPTTA (CVPR'24) [8]78.18±0.0175.14±0.0382.73±0.0273.65±0.0664.78±0.0182.86±0.0377.90±0.0170.64±0.0577.22±0.1087.19±0.0384.97±0.0882.12±0.0278.12
NC-TTT (CVPR'24) [38]79.21±0.2481.76±0.5783.22±0.6977.35±0.3379.40±0.6780.94±0.5480.17±0.7072.36±0.6278.55±0.4384.09±0.3882.38±0.5179.77±0.7479.93
Ours80.47±0.2883.93±0.1983.18±0.3378.33±0.2681.25±0.2383.58±0.3478.86±0.2676.49±0.2180.20±0.2489.17±0.2281.68±0.2283.86±0.2881.75
+ +Table 4. Ablation study in the retinal fundus segmentation. Details of the experiment can be found in Sec 5.1. + +
MethodsSite ASite BSite CSite DSite EAverage
Dice Score Metric ↑ (DSC, mean±std)
No Adapt65.60±5.7874.65±4.8863.15±7.3068.11±5.4975.34±1.0169.37
w/o U77.09±1.9083.24±2.0180.50±1.4478.32±1.5583.18±2.1880.47
w/o priors79.51±0.8079.36±1.1377.40±1.8977.29±1.7780.00±1.9378.71
Ours82.53±1.5288.98±0.8988.73±0.7090.25±1.3391.83±0.7188.46
+ +![](images/ff352ef9cde431b944aa81e57fbb43239ac968d87e3f0fdf79b362279d09c376.jpg) +Figure 4. The performance of DSC (\%) and inference time (s) with different universe sizes $d$ are shown, with the experimental setup identical to that in Table 1. The first curve represents the No Adapt baseline ( $d = 0$ ), while the subsequent curve corresponds to the results obtained with a universe size of $d = 120$ . + +![](images/de7f794bac43d8fe40e35f5601014314a8675039b1a17e45c3ba77bd5f018919.jpg) + +on average. However, when using the pre-trained $\mathcal{U}$ derived from the source model, the segmentation results show a significant improvement. This highlights the effectiveness of the pre-trained morphological priors embedded in $\mathcal{U}$ for medical imaging tasks. + +# 5.2. Effectiveness of the Universe Size + +The universe size $d$ directly impacts both the matching accuracy and computational efficiency, and its selection depends on the number of sampled nodes. Previous works [4, 36, 59], ensured that the $d$ matched the number of nodes in each graph by labeling an equal number of keypoints. In our experiments, $d$ is determined by the spatially-uniform sampling described in Sec 3.1, and we empirically set $d = 100 \times (n + 1)s^{-1}$ , where $s$ is the sampling step and $n$ is the number of segmentation categories required for the task. We evaluated several $d$ , with the results shown in Fig. 4. We observe that excessively small or large step sizes—resulting in too many or too few sampled nodes—negatively impact both training speed and performance. However, when $d$ is within a reasonable range (in this experiment, $s \in [2,10]$ i.e. $d \in [30,150]$ ), the model maintains good scalability + +# and performance. + +Table 5. Comparison of common variables and performance between pairwise matching and multi-matching, using the same experimental setup as in Table 1. Both methods follow the graph construction in Sec. 3.1. + +
paramSite A-E Averagedescription
Pairwise [58]Multi-Matching (Ours)
lr10-310-3learning rate
batch44batch size in inference
h256256dimension of node feature
τ0.050.05regularization factor of Sinkhorn
Iter2020max iterations of Sinkhorn
d-120universe size
DSC ↑84.6388.46dice score metric (%)
time ↓0.9300.392inference time per image (s/img)
Param ↓1.0710.658parameter count (M)
FLOPs ↓15.354.255floating point operations per second (G)
+ +# 5.3. Multi-Matching vs. Pairwise Matching + +Compared to pairwise matching, multi-matching uses cycle-consistency for global optimization, which helps avoid local optima in pairwise. Moreover, joint optimization reduces the computational complexity of large-scale matching. By incorporating morphological consistency constraints, multi-matching exhibits greater robustness. To validate these claims, we conducted experiments comparing the two approaches. For the pairwise matching method, we adopted the benchmark approach [58] and applied the same graph construction process as in multi-matching. The comparison results are presented in Table 5. It is evident that multi-matching outperforms pairwise matching in terms of DSC, inference time, parameters and FLOPs. The inference time of multi-matching was reduced by approximately $57.85\%$ compared to pairwise matching, while segmentation accuracy improved by $3.83\%$ in DSC. + +# 6. Conclusion + +This paper presents a novel multi-matching framework for TTDG, which leverages universe embeddings to incorporate morphological priors from medical images while ensuring cycle-consistency, along with an unsupervised test-time paradigm that fully integrates these priors for efficient adaptation. Extensive experiments, including comparisons and ablation studies demonstrate that our graph-matching-based approach achieves SOTA performance, outperforming entropy-based, template-based, batch normalization-based, and noise estimation-based methods across both multi-source and single-source domain generalization tasks. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (Grant No. 62306003), the Open Research Fund from Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), under Grant No. GML-KF-24-29 and the Open Foundation of Jiangxi Provincial Key Laboratory of Image Processing and Pattern Recognition (ET202404437). + +# References + +[1] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1 +[2] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized Medical Imaging and Graphics, 43:99-111, 2015. 6 +[3] Florian Bernard, Johan Thunberg, Jorge Goncalves, and Christian Theobalt. Synchronisation of partial multi-matchings via non-negative factorisations. Pattern Recognition, 92:146-155, 2019. 2, 5 +[4] Florian Bernard, Johan Thunberg, Paul Swoboda, and Christian Theobalt. Hippi: Higher-order projected power iterations for scalable multi-matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 10284-10293, 2019. 3, 5, 8 +[5] Yuan Bi, Zhongliang Jiang, Ricarda Clarenbach, Ghotbi, and Nassir Navab. Mi-segnet: Mutual information-based us segmentation for unseen domain generalization. In MICCAI, pages 130-140. Springer, 2023. 3 +[6] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 24172-24182, 2023. 3, 6 +[7] Yuxin Chen, Leonidas Guibas, and Qixing Huang. Near-optimal joint object matching via convex relaxation. In ICML, pages II-100, 2014. 3 +[8] Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each test image deserves a specific prompt: Continual test-time adaptation for 2d medical image segmentation. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 11184-11193, 2024. 3, 6, 7, 8 +[9] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Adv. Neural Inform. Process. Syst., 26, 2013. 5 +[10] Frank De Groot. High-resolution x-ray emission and x-ray absorption spectroscopy. Chemical Reviews, 101(6):1779-1808, 2001. 1 +[11] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4548-4557, 2017. 6 + +[12] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. *IJCAI*, 2018. 6 +[13] Francisco Fumero, Silvia Alayón, José L Sanchez, Jose Sigut, and M Gonzalez-Hernandez. Rim-one: An open retinal image database for optic nerve evaluation. In International Symposium on Computer-based Medical Systems (CBMS), pages 1-6. IEEE, 2011. 6 +[14] Quankai Gao, Fudong Wang, Nan Xue, Jin-Gang Yu, and Gui-Song Xia. Deep graph matching under quadratic constraint. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5069-5078, 2021. 4 +[15] Steven Gold, Anand Rangarajan, et al. Softmax to softassign: Neural network algorithms for combinatorial optimization. Journal of Artificial Neural Networks, 2(4):381-399, 1996. 5 +[16] Ran Gu, Guotai Wang, Jiangshan Lu, Jingyang Zhang, Wenhui Lei, Yinan Chen, Wenjun Liao, Shichuan Zhang, Kang Li, Dimitris N Metaxas, et al. Cddsa: Contrastive domain disentanglement and style augmentation for generalizable medical image segmentation. Medical Image Analysis, 89: 102904, 2023. 3 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 770-778, 2016. 4, 6 +[18] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, pages 451-462. Springer, 2020. 6 +[19] Neerav Karani, Ertunc Erdil, Krishna Chaitanya, and Ender Konukoglu. Test-time adaptable neural networks for robust medical image segmentation. Medical Image Analysis, 68: 101907, 2021. 3 +[20] Boah Kim, Inhwa Han, and Jong Chul Ye. Diffuse morph: Unsupervised deformable image registration using diffusion model. In Eur. Conf. Comput. Vis., pages 347-364. Springer, 2022. 2 +[21] Tjalling C Koopmans and Martin Beckmann. Assignment problems and the location of economic activities. *Econometrica: journal of the Econometric Society*, pages 53–76, 1957. 3 +[22] Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Kulkarni, Hiran Sarkar, Varun Jampani, and R Venkatesh Babu. Concurrent subsidiary supervision for unsupervised source-free domain adaptation. In ECCV, pages 177-194. Springer, 2022. 3 +[23] Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. nature, 401 (6755):788-791, 1999. 5 +[24] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5400-5409, 2018. 3 +[25] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semanticcomplete graph matching for domain adaptive object detec + +tion. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5291-5300, 2022. 4 +[26] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):9022-9040, 2023. 2 +[27] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. Int. J. Comput. Vis., pages 1-34, 2024. 1, 3 +[28] Quande Liu, Cheng Chen, Qi Dou, and Pheng-Ann Heng. Single-domain generalization in medical image segmentation via test-time adaptation from shape dictionary. In AAAI, pages 1756-1764, 2022. 2, 3, 6, 7, 8 +[29] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1 +[30] Eleonora Maset, Federica Arrigoni, and Andrea Fusiello. Practical and efficient multi-view matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4568-4576, 2017. 2, 3 +[31] Jose G Moreno-Torres, Troy Raeder, Rocio Alaiz-Rodriguez, Nitesh V Chawla, and Francisco Herrera. A unifying view on dataset shift in classification. Pattern recognition, 45(1):521-530, 2012. 1 +[32] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 3 +[33] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Advances in visual Computing: 16th International Symposium (ISVC), pages 15-28. Springer, 2021. 6 +[34] Andy Nguyen, Mirela Ben-Chen, Katarzyna Welnicka, Yinyu Ye, and Leonidas Guibas. An optimization approach to improving collections of shape maps. In Computer Graphics Forum, pages 1481-1491. Wiley Online Library, 2011. 2 +[35] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. ICLR, 2023. 6, 7, 8 +[36] Zhakshylyk Nurlanov, Frank R Schmidt, and Florian Bernard. Universe points representation learning for partial multi-graph matching. In AAAI, pages 1984-1992, 2023. 3, 8 +[37] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel Van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, et al. Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Medical Image Analysis, 59:101570, 2020. 6 +[38] David Osowiechi, Gustavo A Vargas Hakim, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdanpanah, Ismail Ben Ayed, and Christian Desrosiers. Nc-ttt: A + +noise constrastive approach for test-time training. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 6078-6086, 2024. 6, 7, 8 +[39] Deepti Pachauri, Risi Kondor, and Vikas Singh. Solving the multi-way matching problem by permutation synchronization. Adv. Neural Inform. Process. Syst., 26, 2013. 2, 3 +[40] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 1406-1415, 2019. 3 +[41] Bin Pu, Xingguo Lv, Jiewen Yang, He Guannan, Xingbo Dong, Yiqun Lin, Li Shengli, Tan Ying, Liu Fei, Ming Chen, Zhe Jin, Kenli Li, and Xiaomeng Li. Unsupervised domain adaptation for anatomical structure detection in ultrasound images. In ICML, pages 41204-41220. PMLR, 2024. 2, 3 +[42] Bin Pu, Liwen Wang, Jiewen Yang, Guannan He, Xingbo Dong, Shengli Li, Ying Tan, Ming Chen, Zhe Jin, Kenli Li, et al. M3-uda: A new benchmark for unsupervised domain adaptive fetal cardiac structure detection. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 11621-11630, 2024. 2 +[43] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 12556-12565, 2020. 3 +[44] Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Droppedge: Towards deep graph convolutional networks on node classification. In ICLR, 2020. 4 +[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 6, 7, 8 +[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 4938-4947, 2020. 2 +[47] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Adv. Neural Inform. Process. Syst., 33:11539-11551, 2020. 3 +[48] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 618-626, 2017. 7 +[49] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Adv. Neural Inform. Process. Syst., 35: 14274-14289, 2022. 3 +[50] Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer. International Journal of Computer Assisted Radiology and Surgery, 9:283-293, 2014. 6 +[51] Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices. The annals of mathematical statistics, 35(2):876-879, 1964. 5 +[52] Jayanthi Sivaswamy, SR Krishnadas, Gopal Datt Joshi, Madhulika Jain, and A Ujjwaft Syed Tabish. Drishti-gs: Retinal + +image dataset for optic nerve head (onh) segmentation. In International Symposium on Biomedical Imaging (ISBI), pages 53-56. IEEE, 2014. 6 +[53] Charles P Slichter. Principles of magnetic resonance. Springer Science & Business Media, 2013. 2 +[54] Gail Ter Haar. Therapeutic applications of ultrasound. Progress in biophysics and molecular biology, 93(1-3):111-129, 2007. 2 +[55] Roberto Tron, Xiaowei Zhou, Carlos Esteves, and Kostas Daniilidis. Fast multi-image matching via density-based clustering. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4057-4066, 2017. 2, 3 +[56] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 6, 7, 8 +[57] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 1, 3 +[58] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 3056-3065, 2019. 4, 8 +[59] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Graduated assignment for joint multi-graph matching and clustering with application to unsupervised graph matching network learning. Adv. Neural Inform. Process. Syst., 33: 19908-19919, 2020. 3, 5, 6, 7, 8 +[60] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Neural graph matching network: Learning lawler's quadratic assignment problem with extension to hypergraph and multiple-graph matching. IEEE Trans. Pattern Anal. Mach. Intell., 44 (9):5261-5279, 2021. 4 +[61] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Unsupervised learning of graph matching with mixture of modes via discrepancy minimization. IEEE Trans. Pattern Anal. Mach. Intell., 45(8):10500-10518, 2023. 3, 6 +[62] Ruxue Wen, Hangjie Yuan, Dong Ni, Wenbo Xiao, and Yaoyao Wu. From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In Proc. IEEE/CVF Winter Conference on Applications of Computer Vision, pages 464-474, 2024. 6, 7, 8 +[63] Junchi Yan, Yu Tian, Hongyuan Zha, Xiaokang Yang, Ya Zhang, and Stephen M Chu. Joint optimization for consistent multiple graph matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 1649-1656, 2013. 2 +[64] Junchi Yan, Yin Li, Wei Liu, Hongyuan Zha, Xiaokang Yang, and Stephen Mingyu Chu. Graduated consistency-regularized optimization for multi-graph matching. In ECCV, pages 407–422. Springer, 2014. 2 +[65] Junchi Yan, Minsu Cho, Hongyuan Zha, Xiaokang Yang, and Stephen M Chu. Multi-graph matching via affinity optimization with graduated consistency regularization. IEEE Trans. Pattern Anal. Mach. Intell., 38(6):1228-1242, 2015. 2 + +[66] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Transactions on Medical Imaging, 41 (12):3575-3586, 2022. 6, 7, 8 +[67] Jiewen Yang, Xinpeng Ding, Ziyang Zheng, Xiaowei Xu, and Xiaomeng Li. Graphecho: Graph-driven unsupervised domain adaptation for echocardiogram video segmentation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 11878-11887, 2023. 2 +[68] Jee Seok Yoon, Kwanseok Oh, Yooseung Shin, Maciej A Mazurowski, and Heung-II Suk. Domain generalization for medical image analysis: A survey. arXiv preprint arXiv:2310.08598, 2023. 1, 3 +[69] Weiyi Yu, Zhizhong Huang, Junping Zhang, and Hongming Shan. San-net: Learning generalization to unseen sites for stroke lesion segmentation with self-adaptive normalization. Computers in Biology and Medicine, 156:106717, 2023. 3 +[70] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 18971-18981, 2023. 3, 6, 7, 8 +[71] Ling Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Holger Roth, Andriy Myronenko, Daguang Xu, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation. IEEE transactions on medical imaging, 39(7):2531-2540, 2020. 3 +[72] Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. Adv. Neural Inform. Process. Syst., 34:23664-23678, 2021. 1 +[73] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Adv. Neural Inform. Process. Syst., 35:38629-38642, 2022. 3 +[74] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. Origa-light: An online retinal fundus image database for glaucoma analysis and research. In 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology, pages 3065-3068. IEEE, 2010. 6 +[75] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(4):4396-4415, 2022. 1, 3 +[76] Xiaowei Zhou, Menglong Zhu, and Kostas Daniilidis. Multi-image matching via fast alternating minimization. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4032-4040, 2015. 3 +[77] Yuliang Zou, Zizhao Zhang, Chun-Liang Li, Han Zhang, Tomas Pfister, and Jia-Bin Huang. Learning instance-specific adaptation for cross-domain segmentation. In ECCV, pages 459-476. Springer, 2022. 3 +[78] Yukun Zuo, Hantao Yao, and Changsheng Xu. Attention-based multi-source domain adaptation. IEEE Transactions on Image Processing, 30:3793-3803, 2021. 3 \ No newline at end of file diff --git a/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/images.zip b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4764d3546ae487ab0d05fa0848ec8ddb6370eb4 --- /dev/null +++ b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:567e37ee1cc72918eb8bb3779cb89477ff5ffae46399241298b2d0e692ce2ae7 +size 735429 diff --git a/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/layout.json b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0f5c798673333aa73ae24814031b1314a2adaba1 --- /dev/null +++ b/2025/Test-Time Domain Generalization via Universe Learning_ A Multi-Graph Matching Approach for Medical Image Segmentation/layout.json @@ -0,0 +1,11295 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 523, + 140 + ], + "type": "text", + "content": "Test-Time Domain Generalization via Universe Learning: A Multi-Graph Matching Approach for Medical Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "spans": [ + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "text", + "content": "Xingguo Lv" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "text", + "content": " Xingbo Dong" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "text", + "content": " Liwen Wang" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "text", + "content": " Jiewen Yang" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "text", + "content": " Lei Zhao" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "text", + "content": " Bin " + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "\\mathrm{Pu}^{3\\dagger}" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "text", + "content": " Zhe Jin" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "text", + "content": " Xuejun Li" + }, + { + "bbox": [ + 122, + 160, + 476, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 190, + 545, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 190, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 545, + 218 + ], + "type": "text", + "content": "1 Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, Anhui University. 2 Hunan University. 3 The Hong Kong University of Science and Technology." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 220, + 457, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 220, + 457, + 231 + ], + "spans": [ + { + "bbox": [ + 137, + 220, + 457, + 231 + ], + "type": "text", + "content": "lvxg@stu.ahu.edu.cn, xingbo.dong@ahu.edu.cn, eebinpu@ust.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 284, + 290, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 559 + ], + "type": "text", + "content": "Despite domain generalization (DG) has significantly addressed the performance degradation of pre-trained models caused by domain shifts, it often falls short in real-world deployment. Test-time adaptation (TTA), which adjusts a learned model using unlabeled test data, presents a promising solution. However, most existing TTA methods struggle to deliver strong performance in medical image segmentation, primarily because they overlook the crucial prior knowledge inherent to medical images. To address this challenge, we incorporate morphological information and propose a framework based on multi-graph matching. Specifically, we introduce learnable universe embeddings that integrate morphological priors during multi-source training, along with novel unsupervised test-time paradigms for domain adaptation. This approach guarantees cycle-consistency in multi-matching while enabling the model to more effectively capture the invariant priors of unseen data, significantly mitigating the effects of domain shifts. Extensive experiments demonstrate that our method outperforms other state-of-the-art approaches on two medical image segmentation benchmarks for both multi-source and single-source domain generalization tasks. The source code is available at https://github.com/Yore0/TTDG-MGM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 592, + 128, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 592, + 128, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 592, + 128, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 613, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 696 + ], + "type": "text", + "content": "In the field of medical image processing, semantic segmentation is a technique that enables the precise identification and quantitative analysis of target regions. Currently, a considerable number of medical models can effectively perform segmentation tasks [1, 29]. Nevertheless, when pretrained on source datasets, their performance often declines in real-world deployment due to domain shifts [31, 72] caused" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 259, + 431, + 360 + ], + "blocks": [ + { + "bbox": [ + 310, + 259, + 431, + 360 + ], + "lines": [ + { + "bbox": [ + 310, + 259, + 431, + 360 + ], + "spans": [ + { + "bbox": [ + 310, + 259, + 431, + 360 + ], + "type": "image", + "image_path": "37f3209c7ba8a047436701a6e3f908802f5c230a160bae2e68f373c4b7118dba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 361, + 369, + 370 + ], + "lines": [ + { + "bbox": [ + 358, + 361, + 369, + 370 + ], + "spans": [ + { + "bbox": [ + 358, + 361, + 369, + 370 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 426, + 260, + 542, + 361 + ], + "blocks": [ + { + "bbox": [ + 426, + 260, + 542, + 361 + ], + "lines": [ + { + "bbox": [ + 426, + 260, + 542, + 361 + ], + "spans": [ + { + "bbox": [ + 426, + 260, + 542, + 361 + ], + "type": "image", + "image_path": "76059ff464c857151caee3ae8804e9c05e7207c372ece7dd8691238b57bb71ec.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 361, + 489, + 371 + ], + "lines": [ + { + "bbox": [ + 477, + 361, + 489, + 371 + ], + "spans": [ + { + "bbox": [ + 477, + 361, + 489, + 371 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 312, + 372, + 402, + 442 + ], + "blocks": [ + { + "bbox": [ + 312, + 372, + 402, + 442 + ], + "lines": [ + { + "bbox": [ + 312, + 372, + 402, + 442 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 402, + 442 + ], + "type": "image", + "image_path": "545d2104be1c93bbbc60f3bbf62fc5229e827f8a8846b41dca6b66375537bdfa.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 406, + 372, + 446, + 441 + ], + "blocks": [ + { + "bbox": [ + 406, + 372, + 446, + 441 + ], + "lines": [ + { + "bbox": [ + 406, + 372, + 446, + 441 + ], + "spans": [ + { + "bbox": [ + 406, + 372, + 446, + 441 + ], + "type": "image", + "image_path": "9b76d0a1906a60cd0b67fc00f89c7ef201a71f907277ef379536b3823f6e5b18.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 404, + 459, + 413 + ], + "lines": [ + { + "bbox": [ + 451, + 404, + 459, + 413 + ], + "spans": [ + { + "bbox": [ + 451, + 404, + 459, + 413 + ], + "type": "text", + "content": "X" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 419, + 444, + 429, + 453 + ], + "lines": [ + { + "bbox": [ + 419, + 444, + 429, + 453 + ], + "spans": [ + { + "bbox": [ + 419, + 444, + 429, + 453 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 468, + 388, + 541, + 424 + ], + "blocks": [ + { + "bbox": [ + 468, + 388, + 541, + 424 + ], + "lines": [ + { + "bbox": [ + 468, + 388, + 541, + 424 + ], + "spans": [ + { + "bbox": [ + 468, + 388, + 541, + 424 + ], + "type": "image", + "image_path": "8f714bc8a7cce0fe94f725696273ec4de0156be223bb4e0126ed2431b3c09209.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 465, + 545, + 510 + ], + "lines": [ + { + "bbox": [ + 305, + 465, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 545, + 510 + ], + "type": "text", + "content": "Figure 1. Illustration of multi-graph matching using the universe of nodes. (a) Direct pairwise matching between graphs 1, 2, and 3. (b) Multi-graph matching with the universe of nodes. (c) Equivalence between (a) and (b) (see Lemma 1 in Sec. 2.1 for details)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "content": "by differences in imaging devices, protocols, patient demographics, and preprocessing methods, which impact the model's generalization ability. To address domain shifts, previous researches in domain adaptation, such as domain generalization (DG) [57, 68, 75], have focused on designing sophisticated models that train jointly on source and target domains or across multiple styled domains. Despite the progress, these methods fall short in practice, as no source domains can fully capture all real-world variations, and retraining for each new data is impractical. Consequently, a more practical approach is to adaptively fine-tune pre-train model using the information from the unlabeled unseen data during the test, which is referred to test-time adaptation (TTA) [27]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Medical images, due to their reliance on specific physical principles (such as X-ray absorption [10], magnetic reso" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15621" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "nance [53], sound wave reflection [54], etc.), are fundamentally different from images capturing visible light in natural scenes. Moreover, the physiological characteristics of specific anatomical regions exhibit relatively stable shapes and spatial layouts, and the morphology of these structures is predictable [41, 42]. These prior insights provide valuable information that is particularly useful in medical imaging tasks. In areas such as segmentation [28, 67], registration [20], and detection [41, 42], many well-established approaches have been proposed that combine above handcrafted designs with neural network-based methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 443 + ], + "type": "text", + "content": "We combine the aforementioned TTA strategies with the prior knowledge of medical images and design a novel Test-Time Domain Generalization (TTDG) framework utilizing graph structures. The construction of graph enables a more robust representation of the morphological priors inherent to medical images. In complex medical scenarios, the aggregation mechanism of nodes and edges in graph ensures the effective preservation of crucial contextual information. Previous research has utilised pairwise graph matching [26, 42, 46] to address domain shifts in medical imaging. However, these methods are limited to aligning features between two domains, which poses challenges in scenarios involving multiple domains commonly encountered in real-world settings. Multi-graph matching overcomes these limitations by establishing connections between multiple graphs and integrating information from diverse domains, thereby enabling the capture of more intricate patterns and structural relationships while maintaining cross-domain consistency. This process facilitates the acquisition of global invariant features from medical images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 445, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 289, + 624 + ], + "type": "text", + "content": "In multi-graph matching, the universe of nodes [39, 55] represents the full-set of the nodes from all graphs (refer to Fig. 1). The aforementioned morphological priors are embedded into this universe, derived from a multitude of hospitals, devices, and modalities. It offers three main advantages: (1) Cross-domain consistency. By mapping the nodes from different graphs into the universe, we ensure that all graphs are aligned according to the same reference standard, i.e. the priors. (2) Cycle-consistency in multi-graph matching. This guarantees global coherence across multiple graphs, avoiding conflicts in local node alignments. (3) Handling partial matching and missing nodes. Even when certain nodes are absent in some graphs, matching them to virtual nodes in the universe addresses the issue, enhancing robustness to incomplete graph structures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 625, + 258, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 625, + 258, + 636 + ], + "spans": [ + { + "bbox": [ + 58, + 625, + 258, + 636 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 641, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 701 + ], + "type": "text", + "content": "- We propose the first (to our best knowledge) multi-graph matching framework for test-time domain generalization in medical image segmentation tasks, effectively mitigating performance degradation caused by domain shifts during the testing phase." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "- By designing learnable universe embeddings, we inte" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 547, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 547, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 547, + 131 + ], + "type": "text", + "content": "grate the morphological priors of medical images into the graph matching process during source training, while ensuring the cycle-consistency constraint. This approach enables joint optimization and promotes the learning of domain-invariant features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 133, + 545, + 228 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 133, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 179 + ], + "type": "text", + "content": "- We design a novel, well-initialized unsupervised testing adaptation paradigm that integrates prior knowledge and allows seamless deployment during adaptation, effectively addressing domain shifts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 180, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 545, + 228 + ], + "type": "text", + "content": "- We conduct extensive experiments on two typical medical datasets, demonstrating that our method performs competitively against state-of-the-art approaches for both multi-source and single-source domain generalization." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 239, + 470, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 470, + 253 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 470, + 253 + ], + "type": "text", + "content": "2. Preliminary & Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 259, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 547, + 319 + ], + "type": "text", + "content": "The following section will introduce the preliminaries of multi-graph matching, while also reviewing the most relevant existing literature on the subject. Subsequently, domain generalization and test-time adaptation will be discussed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 327, + 437, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 437, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 437, + 340 + ], + "type": "text", + "content": "2.1. Multi-Graph Matching" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": "We consider " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "m \\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": " different graphs " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1, \\mathcal{G}_2, \\dots, \\mathcal{G}_m" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "i \\in [m]" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "[m] := \\{1, \\dots, m\\}" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i = (\\mathcal{V}_i, \\mathcal{A}_i)" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i \\in \\mathbb{R}^{n_i \\times h}" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": "-dimensional feature of " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": " nodes, and adjacency matrix " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i \\in \\mathbb{R}^{n_i \\times n_i}" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": " encodes the connectivity between nodes, represented as the set of edges. For any two graphs " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_j" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": ", the set of " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "n_i \\times n_j" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": " partial permutation matrices " + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{n_in_j}" + }, + { + "bbox": [ + 304, + 346, + 545, + 431 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 441, + 550, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 550, + 465 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 550, + 465 + ], + "type": "interline_equation", + "content": "\\mathbb {P} _ {n _ {i} n _ {j}} := \\left\\{\\mathbf {X} \\in \\{0, 1 \\} ^ {n _ {i} \\times n _ {j}}: \\mathbf {X} \\mathbf {1} _ {n _ {j}} \\leq \\mathbf {1} _ {n _ {i}}, \\mathbf {X} ^ {\\top} \\mathbf {1} _ {n _ {i}} \\leq \\mathbf {1} _ {n _ {j}} \\right\\}, \\tag {1}", + "image_path": "1687447b928c938310ea6e346676e08a54b088d439564afccfdb7de5d8d644cf.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{n_i}" + }, + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "text", + "content": " denotes a " + }, + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "text", + "content": "-dimensional column vector whose elements are all ones. The assignment matrix " + }, + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{ij} \\in \\mathbb{P}_{n_i n_j}" + }, + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "text", + "content": " between a pair of graphs " + }, + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}_i, \\mathcal{G}_j)" + }, + { + "bbox": [ + 304, + 466, + 545, + 514 + ], + "type": "text", + "content": " denotes a meaningful correspondence that encodes the matching." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "type": "text", + "content": "When considering the synchronisation matching [30, 39] of multiple graphs (" + }, + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "type": "inline_equation", + "content": "m > 2" + }, + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "type": "text", + "content": "), relying on local matches between graph pairs can readily result in erroneous correspondences that contradict each other globally [63]. Consequently, the multi-graph matching has been addressed in terms of simultaneously solving under the constraints of cycle-consistency [64, 65]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 598, + 547, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 598, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 598, + 547, + 623 + ], + "type": "text", + "content": "Definition 1 Cycle-consistency. The matching among " + }, + { + "bbox": [ + 305, + 598, + 547, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1, \\mathcal{G}_2, \\dots, \\mathcal{G}_m" + }, + { + "bbox": [ + 305, + 598, + 547, + 623 + ], + "type": "text", + "content": " is cycle-consistent (partial transitivity), if" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 351, + 634, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 634, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 351, + 634, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {i k} \\mathbf {X} _ {k j} \\leq \\mathbf {X} _ {i j}, \\quad \\forall i, j, k \\in [ m ]. \\tag {2}", + "image_path": "74aadf13ed8f378b9025d39efc084cfa25f755d6e2e397497d1a5da29ee20ad1.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In contrast to full matching, i.e. in Eq. (2) the inequalities become equalities and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "n_i = n_j" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", partial matching requires only that the pairwise matching combinations in the cycle form a subset of the identity matching [3]. An iterative refinement strategy [34] was employed to enhance pairwise" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15622" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "matchings by ensuring global mapping consistency. In [39], the authors proposed a method to achieve cycle-consistency based on spectral decomposition of the assignment matrix. In addition, methods such as convex programming [7], low-rank matrix recovery [76], and spectral decomposition [30] have been proposed to address the synchronization of partial matchings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 288, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 240 + ], + "type": "text", + "content": "Instead of explicitly modeling the cubic number of nonconvex quadratic constraints, a more efficient approach to enforcing cycle consistency is the use of the universe of nodes [55]. In a universe comprising nodes of size " + }, + { + "bbox": [ + 46, + 156, + 288, + 240 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{N}" + }, + { + "bbox": [ + 46, + 156, + 288, + 240 + ], + "type": "text", + "content": ", the matching between graphs can be decomposed by matching each graph " + }, + { + "bbox": [ + 46, + 156, + 288, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 46, + 156, + 288, + 240 + ], + "type": "text", + "content": " to the space of universe (refer to Fig. 1). We denote the universe matchings as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 248, + 286, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 248, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 52, + 248, + 286, + 274 + ], + "type": "interline_equation", + "content": "\\mathbb {U} _ {n _ {i} d} := \\left\\{\\mathbf {U} \\in \\{0, 1 \\} ^ {n _ {i} \\times d}: \\mathbf {U} \\mathbf {1} _ {d} = \\mathbf {1} _ {n _ {i}}, \\mathbf {U} ^ {\\top} \\mathbf {1} _ {n _ {i}} \\leq \\mathbf {1} _ {d} \\right\\}. \\tag {3}", + "image_path": "6e556da49cbfab115545f6ada57cf002857ef1a8b2311bb56b92d9a2caf365a7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "text", + "content": "Lemma 1 Cycle-consistency, universe matching. The pairwise (partial) matching matrices " + }, + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_{ij}\\}_{i,j=1}^{m}" + }, + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "text", + "content": " is cycle-consistent iff there exists a collection of universe matching " + }, + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{U}_i \\in \\mathbb{U}_{n_id}\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "text", + "content": ", such that for each " + }, + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{ij}" + }, + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "text", + "content": ", it holds that " + }, + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{ij} = \\mathbf{U}_i \\mathbf{U}_j^\\top" + }, + { + "bbox": [ + 46, + 274, + 286, + 334 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 286, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 286, + 465 + ], + "type": "text", + "content": "The proof will be presented in the appendix. By projecting each node of a graph to the universe of nodes and identifying the universe matchings " + }, + { + "bbox": [ + 46, + 335, + 286, + 465 + ], + "type": "inline_equation", + "content": "\\mathbb{U}_{n_i d}" + }, + { + "bbox": [ + 46, + 335, + 286, + 465 + ], + "type": "text", + "content": ", it is ensured that the cycle-consistency constraint will be satisfied throughout the multi-graph matching process. In [59, 61], the authors proposed a gradual assignment procedure to achieve soft matching and clustering through the universe matching. In [4], the universe of nodes is leveraged to incorporate geometric consistency, ensuring both point scaling and convergence. [36] utilized the universe matching to facilitate effective partial multi-graph matching." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 466, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 286, + 514 + ], + "type": "text", + "content": "Definition 2 Multi-Matching Koopmans-Beckmann's Quadratic Assignment Problem (KB-QAP) [21]. Multi-graph matching is formulated with KB-QAP, by summing KB-QAP objectives among all pairs of graphs:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 523, + 286, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 523, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 61, + 523, + 286, + 581 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\mathbf {X} _ {i j}} \\sum_ {i, j \\in [ m ]} (\\lambda \\cdot t r \\left(\\mathbf {X} _ {i j} ^ {\\top} \\mathcal {A} _ {i} \\mathbf {X} _ {i j} \\mathcal {A} _ {j}\\right) + t r \\left(\\mathbf {X} _ {i j} ^ {\\top} \\mathbf {M} _ {i j}\\right)), \\tag {4} \\\\ s. t. \\quad \\mathbf {X} _ {i j} \\in \\{0, 1 \\} ^ {n _ {i} \\times n _ {j}}, \\mathbf {X} _ {i j} \\in \\mathbb {P} _ {n _ {i} n _ {j}}, \\\\ \\mathbf {X} _ {i k} \\mathbf {X} _ {k j} \\leq \\mathbf {X} _ {i j}, \\quad \\forall i, j, k \\in [ m ]. \\\\ \\end{array}", + "image_path": "a80d23f793309902bf1a77e8a9f7ef8f2642414d71062b30ef6d97eb65ccadce.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 590, + 286, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 286, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 286, + 615 + ], + "type": "text", + "content": "In Eq. (4), " + }, + { + "bbox": [ + 46, + 590, + 286, + 615 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 590, + 286, + 615 + ], + "type": "text", + "content": " is a scaling factor for edge-to-edge similarity, and " + }, + { + "bbox": [ + 46, + 590, + 286, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij}" + }, + { + "bbox": [ + 46, + 590, + 286, + 615 + ], + "type": "text", + "content": " is the node-to-node similarity between " + }, + { + "bbox": [ + 46, + 590, + 286, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i,\\mathcal{G}_j" + }, + { + "bbox": [ + 46, + 590, + 286, + 615 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 623, + 178, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 178, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 178, + 634 + ], + "type": "text", + "content": "2.2. Domain Generalization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "Domain generalization addresses a challenging scenario where one or more different but related domains are provided, with the objective of training a model that can generalize effectively to an unseen test domain [57, 68, 75]. Medical image analysis encounters significant challenges due to factors such as variability in image appearance, the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "complexity and high dimensionality of the data, difficulties in data acquisition, and issues related to data organization, labeling, safety, and privacy. Previous research has proposed DG algorithms across multiple levels, including data-level [69, 71], feature-level [5, 24], model-level [16, 40, 78], and analysis-level [28, 43]. Yu et al. [69] proposed a U-Net z-score normalization network for the stroke lesion segmentation. Liu et al. [28] employed dictionary learning for prostate and fundus segmentation by creating a shape dictionary composed of template masks. In [5], the authors utilized mutual information to differentiate between domain-invariant features and domain-specific ones in ultrasound image segmentation. Gu et al. [16] proposed a domain-style contrastive learning approach that disentangles an image into invariant representations and style codes for DG." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 258, + 430, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 430, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 430, + 270 + ], + "type": "text", + "content": "2.3. Test-Time Adaptation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 277, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 502 + ], + "type": "text", + "content": "Test-time adaptation is an emerging paradigm that allows a pre-trained model to adapt to unlabeled data during the testing phase, before making predictions [27]. Several TTA methods have been proposed recently, utilizing techniques such as self-supervised learning [6, 22, 73], batch normalization calibration [8, 32, 47, 70, 77], and input data adaptation [19, 49] to achieve better test-time performance. Typically, VPTTA [8] is a method that freezes the pre-trained model and generates low-frequency prompts for each image during inference in medical image segmentation. In [19], the authors proposed designing two CNN-based sub-networks along with an image normalization network. During test-time training, the image normalization network was adapted for each image. [70] proposed a dynamic mixture coefficient and a statistical transformation operation to adaptively merge the training and testing statistics of the normalization layers. Additionally, the authors design an entropy minimization loss to address the issue of domain shifts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 514, + 388, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 388, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 388, + 527 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "content": "This section provides an overview of how multi-graph matching methods can facilitate training for TTDG tasks. Given inputs consisting of images from " + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "content": " source domains " + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "inline_equation", + "content": "(S\\geq 1)" + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s = \\{D_1,D_2,\\dots ,D_S\\}" + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "content": ", the objective is to enable the model to make more accurate predictions on " + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "content": " unseen target domains " + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "inline_equation", + "content": "(T\\geq 1)" + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = \\{D_1,D_2,\\dots ,D_T\\}" + }, + { + "bbox": [ + 304, + 534, + 545, + 617 + ], + "type": "text", + "content": ", during the testing phase." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Inspired by [41], we aim to effectively leverage morphological prior through graph construction. However, unlike the UDA tasks, which typically involve only two domains, our training process simultaneously handles multiple labeled domains. Pairwise matching across multiple domains can lead to the challenges discussed in Sec. 1. To tackle this, we propose learning domain-invariant latent representations from a multi-graph matching perspective, thereby" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15623" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 541, + 432 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 541, + 432 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 541, + 432 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 541, + 432 + ], + "type": "image", + "image_path": "fcc7025cac673c8508ad7b044f317c883c4e41b086a5cf6eb8ec29e81f6e6d46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 441, + 546, + 509 + ], + "lines": [ + { + "bbox": [ + 46, + 441, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 546, + 509 + ], + "type": "text", + "content": "Figure 2. Overview of our TTDG framework. During source model training, data from different domains are jointly used to train the Segmentation Network (feature extractor and segmentation head). Feature maps and ground truth masks are utilized to construct graphs " + }, + { + "bbox": [ + 46, + 441, + 546, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 46, + 441, + 546, + 509 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 441, + 546, + 509 + ], + "type": "inline_equation", + "content": "i = 4" + }, + { + "bbox": [ + 46, + 441, + 546, + 509 + ], + "type": "text", + "content": " in this figure legend) and corresponding labels " + }, + { + "bbox": [ + 46, + 441, + 546, + 509 + ], + "type": "inline_equation", + "content": "Y_i" + }, + { + "bbox": [ + 46, + 441, + 546, + 509 + ], + "type": "text", + "content": " (Sec. 3.1), with universe embeddings learned via back-propagation, incorporating morphological priors (Sec. 3.2). At test time, multi-graph matching is performed on all target domain images in each batch. Despite style differences, these images share common morphological patterns. Universe embeddings are frozen as prior knowledge to guide the matching, and the segmentation network is fine-tuned via back-propagation for efficient adaptation (Sec. 3.3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 514, + 288, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 288, + 538 + ], + "type": "text", + "content": "facilitating rapid adaptation to unseen domains during the TTA phase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 550, + 177, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 177, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 177, + 562 + ], + "type": "text", + "content": "3.1. Construction of Graph" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Rather than using conventional graph-matching algorithms that depend on explicitly defined keypoints as graph nodes [14, 58, 60], accurately annotating keypoints in medical imaging is both more costly and less practical than in natural images. Therefore, given a mini-batch " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{x_i\\}_{i=1}^m \\in \\mathcal{D}_s" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", as shown in Fig. 2, it is first passed through a general feature extractor, such as ResNet [17], to obtain visual features. We then perform spatially-uniform sampling [25] of pixels within the ground-truth masks at each feature level to obtain " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " foreground nodes with its corresponding labels " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "Y_i \\in \\mathbb{Z}^{n_i}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ". After extracting these fine-grained visual features, we employ a nonlinear projection to transform the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 514, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 550 + ], + "type": "text", + "content": "visual space into the " + }, + { + "bbox": [ + 304, + 514, + 545, + 550 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 514, + 545, + 550 + ], + "type": "text", + "content": "-dimensional graph space. This approach generates the feature of nodes " + }, + { + "bbox": [ + 304, + 514, + 545, + 550 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_i \\in \\mathbb{R}^{n_i \\times h}\\}_{i=1}^m" + }, + { + "bbox": [ + 304, + 514, + 545, + 550 + ], + "type": "text", + "content": " that more effectively preserve the semantic characteristics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "content": "For the weighted adjacency matrix encoding structural information, we use Droppedge [44] to reduce potential bias caused by the frequent visual correspondences, preventing the model from over-relying on specific matches. The formulation is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 311, + 624, + 545, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 624, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 311, + 624, + 545, + 648 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {i} = \\operatorname {D r o p e d g e} \\left\\{\\text {s o f t m a x} \\left[ \\mathcal {V} _ {i} \\mathcal {W} _ {x} \\cdot \\left(\\mathcal {V} _ {i} \\mathcal {W} _ {y}\\right) ^ {\\top} \\right] \\odot D ^ {- 1} \\right\\}, \\tag {5}", + "image_path": "ae842ef4562a0630795acde5f6e4a60f58bb8c4799ac7369a0146a6bdfbc7fbd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_x" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_y" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " are two learnable linear projection, " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "D = \\frac{\\mathcal{V}_i\\mathcal{V}_i^{\\mathrm{T}}}{||\\mathcal{V}_i||_2}" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " is the cosine distance matrix. Edges with larger weights indicate nodes that are closer to each other, so we use the inverse of " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " and apply the hadamard product " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " to combine node similarity with geometric proximity. The di" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15624" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "agonal elements of " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " are set to zero. Based on the above steps, we obtain a corresponding graph " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i = (\\mathcal{V}_i,\\mathcal{A}_i)" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " for each input " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 116, + 263, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 116, + 263, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 263, + 129 + ], + "type": "text", + "content": "3.2. Formulation of the Universe Embeddings" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 134, + 287, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 241 + ], + "type": "text", + "content": "In Sec. 2.1, it was demonstrated that by matching each node of a graph to the universe of nodes, we can maintain cycle-consistency while avoiding the computational burden associated with cubic non-convex constraints. While a randomly initialized matrix can learn accurate universe matchings " + }, + { + "bbox": [ + 46, + 134, + 287, + 241 + ], + "type": "inline_equation", + "content": "\\mathbb{U}_{n_i d}" + }, + { + "bbox": [ + 46, + 134, + 287, + 241 + ], + "type": "text", + "content": " in a supervised setting, we introduce universe embeddings " + }, + { + "bbox": [ + 46, + 134, + 287, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{U} \\in \\mathbb{R}^{d \\times h}" + }, + { + "bbox": [ + 46, + 134, + 287, + 241 + ], + "type": "text", + "content": ", which serve as learnable latent representations to enhance adaptability in unsupervised multi-domain scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " is initialized as " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "1 / d + 10^{-3}z" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "z\\sim N(0,1)" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ", following the setting in [59]. For each " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ", we learn a unique matching between nodes in " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\nu_{i}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " and the corresponding in the universe of nodes, as follows: " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "U_{i} = \\mathrm{Sinkhorn}(\\mathcal{V}_{i}\\mathcal{U}^{\\top},\\tau)\\in" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathbb{U}_{n_i d}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ". Sinkhorn " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "(X,\\tau)" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " refers to the Sinkhorn algorithm [9, 51], which applies a relaxed projection with entropic regularization [15] to normalize the matrix " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ", resulting in a doubly-stochastic matrix, and " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\tau \\in (0, + \\infty)" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " is the regularization factor. For the obtained " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "U_{i}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "d\\times d" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " matrix " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "U_{i}^{\\top}\\mathcal{A}_{i}U_{i}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " is a row and column reordering of " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " based on the universe of nodes. Accordingly, the agreement between two adjacency matrices " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_j" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ", reordered according to their respective universe matching assignment matrices " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "U_{i}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "U_{j}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ", can be quantified using the Frobenius inner product namely " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\langle U_i^\\top \\mathcal{A}_iU_i,U_j^\\top \\mathcal{A}_jU_j\\rangle" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{U} = [U_1^\\top ,\\dots ,U_m^\\top ]^\\top \\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "n = \\sum_{i = 1}^{m}n_{i}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ". The block-diagonal multi-adjacency matrix is defined as " + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = diag(\\mathcal{A}_1,\\dots ,\\mathcal{A}_m)\\in \\mathbb{R}^{n\\times n}" + }, + { + "bbox": [ + 46, + 242, + 287, + 459 + ], + "type": "text", + "content": ". We can get the multi-matching formulation as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 462, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 462, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 67, + 462, + 287, + 510 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f(\\mathbf{U}):= \\max_{U_{1},\\ldots ,U_{m}}\\quad \\sum_{i,j\\in [m]}\\langle U_{i}^{\\intercal}\\mathcal{A}_{i}U_{i},U_{j}^{\\intercal}\\mathcal{A}_{j}U_{j}\\rangle \\\\ = \\max _ {\\mathbf {U} \\in \\mathbb {U}} \\operatorname {t r} \\left(\\mathbf {U} ^ {\\top} \\mathbf {A} \\mathbf {U} \\mathbf {U} ^ {\\top} \\mathbf {A} \\mathbf {U}\\right). \\tag {6} \\\\ \\end{array}", + "image_path": "9bcc627d3781aa366612f7e6183c79f9a36cb808b64bca610c79968e662450b4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "text", + "content": "It was observed that in the original multi-graph matching optimization problem, graphs are typically constructed for each individual instance, with all the nodes within a graph belonging to the same category. However, in medical imaging, constructing separate graphs for each organ would not only greatly increase the computational complexity of graph matching but also result in misalignments between graphs from different categories, significantly compromising matching accuracy. To address this limitation, we introduce a class-wise similarity matrix " + }, + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "text", + "content": " to incorporate class-aware label information for each node. This ensures that the nodes are correctly aligned according to their respective categories during multi-graph matching. Specifically, we define " + }, + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "inline_equation", + "content": "W_{ij} = Y_iY_j^{\\top}" + }, + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = [W_{ij}]_{ij} \\in \\mathbb{Z}^{(\\sum_{i=1}^{m}n_i)\\times(\\sum_{j=1}^{m}n_j)}" + }, + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "text", + "content": ". The class-aware similarity matrix " + }, + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{A}} = \\mathbf{W}^{\\top}\\mathbf{A}\\mathbf{W}" + }, + { + "bbox": [ + 46, + 520, + 287, + 714 + ], + "type": "text", + "content": ". Consequently, Eq. (6) can be transformed" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 402, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 402, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 402, + 84 + ], + "type": "text", + "content": "into the following form" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 369, + 86, + 545, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 86, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 369, + 86, + 545, + 105 + ], + "type": "interline_equation", + "content": "\\max _ {\\mathbf {U} \\in \\mathbb {U}} \\operatorname {t r} \\left(\\mathbf {U} ^ {\\top} \\tilde {\\mathbf {A}} \\mathbf {U} \\mathbf {U} ^ {\\top} \\tilde {\\mathbf {A}} \\mathbf {U}\\right). \\tag {7}", + "image_path": "5e5d4ddf89f4d5f3bd3a76a90cfc8bf03b4f2aa687e90d451a9a74a720b0fc10.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 112, + 545, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 112, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 305, + 112, + 545, + 148 + ], + "type": "text", + "content": "The higher-order projected power iteration (HiPPI) [4] is employed to iteratively solve " + }, + { + "bbox": [ + 305, + 112, + 545, + 148 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 305, + 112, + 545, + 148 + ], + "type": "text", + "content": " to obtain a stable convergence form:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 150, + 545, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 150, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 318, + 150, + 545, + 165 + ], + "type": "interline_equation", + "content": "U _ {i} ^ {*} = U _ {i} ^ {t} \\quad s. t. \\quad | | U _ {i} ^ {t} - U _ {i} ^ {t - 1} | | < \\theta , \\quad \\forall i \\in [ m ]. \\tag {8}", + "image_path": "c7ed3e9e9ae4aec9f98a58e504a281463d3e85de2a3f895caffd262c0f22ab56.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 171, + 545, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 231 + ], + "type": "text", + "content": "Empirically, we set the convergence threshold " + }, + { + "bbox": [ + 304, + 171, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\theta = 10^{-5}" + }, + { + "bbox": [ + 304, + 171, + 545, + 231 + ], + "type": "text", + "content": ". In order to guarantee the optimal learning of universe embeddings " + }, + { + "bbox": [ + 304, + 171, + 545, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 171, + 545, + 231 + ], + "type": "text", + "content": ", we incorporate the non-negative matrix factorisation (NMF) [3, 23] and define the overall optimization objective as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 234, + 545, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 234, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 339, + 234, + 545, + 262 + ], + "type": "interline_equation", + "content": "L (\\mathcal {U}) = \\sum_ {i \\in [ m ]} | | U _ {i} ^ {*} - \\mathcal {V} _ {i} \\mathcal {U} ^ {\\mathrm {T}} | | _ {F} ^ {2} + \\alpha | | \\mathcal {U} | | _ {F} ^ {2}, \\tag {9}", + "image_path": "362459e987c7bb6c43c20ca8a132d8746cfcae7420a76ffaaa2381e156bd3bf5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "inline_equation", + "content": "||\\cdot||_F^2" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": " represents the Frobenius norm, and " + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": " is a tuning parameter. Since only matrix multiplication and element-wise division are involved, the Sinkhorn operation is fully differentiable. This allows for effective updates to " + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": " through back-propagation, enabling the incorporation of morphological prior knowledge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 348, + 520, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 520, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 520, + 361 + ], + "type": "text", + "content": "3.3. Testing Paradigm on the Target Domains" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "text", + "content": "During the Test-Time Adaptation (TTA) phase, we only have access to unlabeled data from the unseen target domains " + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "text", + "content": ". Due to distribution shifts, the performance of models trained on " + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "text", + "content": " can degrade significantly. To address this issue, we employ the multi-graph matching on each mini-batch " + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\{x_i^t\\}_{i=1}^m \\in \\mathcal{D}_t" + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "text", + "content": ". By utilizing universe embeddings " + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "text", + "content": " learned from " + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "text", + "content": " as described in Sec. 3.2, it is guaranteed cycle-consistency, enabling the network to focus on domain-invariant features guided by morphological priors from medical images. In this phase, " + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 366, + 545, + 545 + ], + "type": "text", + "content": " is frozen to prevent error accumulation from affecting the prior information. The feature extractor and segmentation head (segmentation network in Fig. 2) are optimized through gradient back-propagation, allowing the model to adapt effectively to the new domains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": "Unsupervised Multi-Matching. Following the same graph construction described in Sec. 3.1, we obtain the " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i^t = (\\mathcal{V}_i^t,\\mathcal{A}_i^t)" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "x_{i}^{t}" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i^t\\in \\mathbb{R}^{n_i\\times h}" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " class-aware nodes and " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i^t\\in \\mathbb{R}^{n_i\\times n_i}" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " is the adjacency matrix encoding the structural information. We first introduce a network-learned affinity matrix " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "M_{ij} = f_{mlp}\\{\\mathcal{V}_i^t\\mathcal{W}_x^t" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}_j^t\\mathcal{W}_y^t)^\\top \\} \\in \\mathbb{R}^{n_i\\times n_j}" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " in graph pair " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_j" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": ", which computes the similarity between their corresponding node features. " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_x^t" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_y^t" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " are two learnable linear projection, " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{mlp}" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " is a multi-layer perception (MLP). Subsequently, the matching between " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " and the universe of size " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " is obtained through the inner-product between " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i^t" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " and the pre-trained universe embeddings " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": ", in accordance with the same procedure " + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "inline_equation", + "content": "U_{i} = Sinkhorn(\\mathcal{V}_{i}^{t}\\mathcal{U}^{\\top},\\tau)\\in \\mathbb{U}_{n_{i}d}" + }, + { + "bbox": [ + 304, + 546, + 545, + 714 + ], + "type": "text", + "content": " in Sec. 3.2." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15625" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": "In [59, 61], the authors propose solving the multimatching KB-QAP problem presented in Eq. (4) (denoted as " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": ") using a Taylor expansion. Based on the assumptions in Lemma 1 of Sec. 2.1, we can reformulate " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": " in Taylor series as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 133, + 287, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 133, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 133, + 287, + 162 + ], + "type": "interline_equation", + "content": "J (U) \\approx J \\left(U ^ {0}\\right) + \\sum_ {i} \\frac {\\partial J}{\\partial U _ {i}} \\bigg | _ {U = U ^ {0}} \\left(U _ {i} - U _ {i} ^ {0}\\right) + \\dots . \\tag {10}", + "image_path": "8312a37e5160080ebce763d6ce0b5baeed123d0d8b869582dca7f4f46ef73f79.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "type": "text", + "content": " can be approximated as " + }, + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "type": "text", + "content": " linear optimization problems through the first-order Taylor expansion, which can be efficiently solved using the Sinkhorn. The gradient of " + }, + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "type": "inline_equation", + "content": "U_{i}" + }, + { + "bbox": [ + 47, + 169, + 287, + 217 + ], + "type": "text", + "content": ", denoted as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 219, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 63, + 219, + 287, + 251 + ], + "type": "interline_equation", + "content": "V _ {i} = \\frac {\\partial J}{\\partial U _ {i}} = \\sum_ {j \\in [ m ]} \\left(\\lambda \\mathcal {A} _ {i} U _ {i} U _ {j} ^ {\\mathrm {T}} \\mathcal {A} _ {j} U _ {j} + M _ {i j} U _ {j}\\right), \\tag {11}", + "image_path": "641e98d520ff54df5a010c5b8bd20de690339e43255f01fb0360a50b22a4a3d7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 257, + 286, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 286, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 286, + 280 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 47, + 257, + 286, + 280 + ], + "type": "text", + "content": " ensures that " + }, + { + "bbox": [ + 47, + 257, + 286, + 280 + ], + "type": "inline_equation", + "content": "U_{i}" + }, + { + "bbox": [ + 47, + 257, + 286, + 280 + ], + "type": "text", + "content": " gradually converges to a high-quality discrete solution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 281, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 305 + ], + "type": "text", + "content": "Finally, the converged " + }, + { + "bbox": [ + 47, + 281, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\{U_i\\}" + }, + { + "bbox": [ + 47, + 281, + 287, + 305 + ], + "type": "text", + "content": " fine-tune the segmentation network with the matching loss " + }, + { + "bbox": [ + 47, + 281, + 287, + 305 + ], + "type": "inline_equation", + "content": "L_{mat}" + }, + { + "bbox": [ + 47, + 281, + 287, + 305 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 308, + 287, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 308, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 52, + 308, + 287, + 367 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {m a t} = \\sum_ {i, j \\in [ m ]} \\left[ - \\hat {M} _ {i j} ^ {\\gamma} (1 - U _ {i} U _ {j} ^ {\\mathsf {T}}) \\log (U _ {i} U _ {j} ^ {\\mathsf {T}}) \\right] \\\\ - \\sum_ {i, j \\in [ m ]} \\left[ (1 - \\hat {M} _ {i j}) ^ {\\gamma} U _ {i} U _ {j} ^ {\\mathsf {T}} \\log \\left(1 - U _ {i} U _ {j} ^ {\\mathsf {T}}\\right) \\right], \\tag {12} \\\\ \\end{array}", + "image_path": "4ff3f51e8952b63f6cf9744b7273366ca667991d128c4c1569aa6ba24079466d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 376, + 287, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 376, + 287, + 401 + ], + "type": "inline_equation", + "content": "\\hat{M}_{ij} = \\text{sinkhorn}(M_{ij},\\tau)" + }, + { + "bbox": [ + 47, + 376, + 287, + 401 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 376, + 287, + 401 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 376, + 287, + 401 + ], + "type": "text", + "content": " is an amplification factor. See algorithm in appendix for full details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 411, + 127, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 127, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 127, + 424 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 430, + 109, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 109, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 109, + 441 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 449, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 567 + ], + "type": "text", + "content": "The retinal fundus segmentation datasets comprise five public datasets from different medical centers, denoted as Site A (RIM-ONE [13]), B (REFuge [37]), C (ORIGA [74]), D (REFuge-Test [37]), and E (Drishti-GS [52]), with 159, 400, 650, 800, and 101 images respectively, all consistently annotated for optic disc (OD) and optic cup (OC) segmentation. We adopted the preprocessing method outlined in [8, 28], where each image's region of interest is cropped to " + }, + { + "bbox": [ + 46, + 449, + 287, + 567 + ], + "type": "inline_equation", + "content": "800 \\times 800" + }, + { + "bbox": [ + 46, + 449, + 287, + 567 + ], + "type": "text", + "content": " pixels and normalized using min-max normalization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "text", + "content": "The polyp segmentation datasets consist of four public datasets collected from different medical centers, denoted as Site A (BKAI-IGH-NEOPolyp [33]), B (CVC-ClinicDB/CVC-612 [2]), C (ETIS [50]), and D (Kvasir [18]), containing 1000, 612, 196, and 1000 images, respectively. We followed the preprocessing steps in [8], resizing the images to " + }, + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "inline_equation", + "content": "800 \\times 800" + }, + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "text", + "content": " pixels and normalizing them using ImageNet-derived statistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 671, + 163, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 163, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 163, + 684 + ], + "type": "text", + "content": "4.2. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Implementation Details. To ensure a fair comparison across all methods, we followed the protocol in [8] and" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "split each dataset into an " + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "inline_equation", + "content": "8:2" + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": " ratio for training and testing. We employed the ResNet-50 [17] pre-trained on ImageNet as the feature extractor. The training was optimized using the SGD optimizer with a momentum of 0.9 and a learning rate of 0.001. For source model training, we set the mini-batch size to 8, allowing simultaneous matching of 8 graphs per batch. The universe embedding " + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": " was treated as an additional trainable tensor, integrated into the network weights without influencing subsequent experiments with other methods. During the TTA phase, all methods were trained on the target data without access to their labels. The mini-batch size was set to 4, and the previously learned " + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": " from source model training was frozen. All experiments were implemented using PyTorch and conducted on 4 NVIDIA 3090 GPUs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 251, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 348 + ], + "type": "text", + "content": "Evaluation Metrics. The Dice score (DSC, " + }, + { + "bbox": [ + 304, + 251, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 304, + 251, + 545, + 348 + ], + "type": "text", + "content": " ) was used to quantify the accuracy of the predicted masks, serving as the primary metric for evaluating segmentation performance. Additionally, the enhanced alignment metric " + }, + { + "bbox": [ + 304, + 251, + 545, + 348 + ], + "type": "inline_equation", + "content": "E_{\\phi}^{max}" + }, + { + "bbox": [ + 304, + 251, + 545, + 348 + ], + "type": "text", + "content": " [12] was adopted to measure both pixel-level and global-level similarity. To further assess the consistency between predictions and ground truths, we also employed the structural similarity metric " + }, + { + "bbox": [ + 304, + 251, + 545, + 348 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 304, + 251, + 545, + 348 + ], + "type": "text", + "content": " [11]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 358, + 429, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 429, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 429, + 371 + ], + "type": "text", + "content": "4.3. Experimental Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 377, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 497 + ], + "type": "text", + "content": "We selected U-Net [45] as the benchmark model for the segmentation task, training it on the source domains and testing on the target domain without adaptation (No Adapt). In addition, we compared eight state-of-the-art (SOTA) methods, including two entropy-based approaches (TENT [56] and SAR [35]), a shape template-based method (TASD [28]), a dynamically adjusted learning rate method (DLTTA [66]), batch normalization-based methods (DomainAdaptor [70], VPTTA [8]), and noise estimation-based approaches (DeY-Net [62] and NC-TTT [38])." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Multi-Source Generalization. In these experiments, we adopted a leave-one-out training strategy [6] (i.e., with " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "S = |\\mathcal{D}_s \\cup \\mathcal{D}_t| - 1" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "). Although some methods, such as TASD [28], DeY-Net [62] and VPTTA [8], were originally designed for single-source domain training, we simulated single-domain conditions by mixing multi-source domain data, making the experimental setup still feasible for these approaches. Furthermore, we have also included experiments specifically focused on single-source domains later in the study. The segmentation results for the fundus and polyp datasets are shown in Tables 1 and 2, respectively. For OD/OC segmentation, all TTA methods outperformed the No Adapt baseline, highlighting their effectiveness in addressing domain shifts. Comparatively, our method consistently achieved better average performance than all other approaches. Specifically, for the key segmentation metric DSC, our method exceeded the second-best (DeY-Net [62]) by " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.88\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " and outperformed the No Adapt by " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "19.09\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": ". A" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15626" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 108, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 46, + 71, + 547, + 104 + ], + "lines": [ + { + "bbox": [ + 46, + 71, + 547, + 104 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 547, + 104 + ], + "type": "text", + "content": "Table 1. Multi sources domain generalization in retinal fundus segmentation. The average performance (mean ± standard deviation) of three trials for our method and nine SOTA methods. \"Site A\" means training on Sites B-E and testing on Site A, and similarly for the others. The best results are highlighted in red." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 108, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 49, + 108, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 108, + 541, + 175 + ], + "type": "table", + "html": "
MethodsSite ASite BSite CSite DSite E
\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{0}\\)\\(S_{\\text{a}}\\)\\(DSC\\)\\(E_{\\text{max}}^{2}\\)\\(S_{\\text{a}}\\)
No Adapt (U-Net [45])65.60±5.7891.75±0.1379.11±0.0774.65±4.8888.78±0.0883.31±0.0263.15±7.3081.03±0.1979.77±0.0368.11±5.4989.46±0.2082.12±0.0775.34±1.0190.68±0.1687.53±0.05
TENT (ICLR'21) [56]75.22±3.9994.10±0.1083.97±0.0581.06±0.9393.45±0.1192.36±0.0375.23±2.3384.62±0.1478.12±2.5991.33±0.3186.41±0.1283.17±0.7892.74±0.1093.64±0.0378.5691.9488.20
TASD (AAAI'22) [28]79.70±2.5288.73±0.4583.67±0.1782.96±2.1490.01±0.2090.99±0.1186.12±1.1193.59±0.1387.08±0.0778.02±6.3087.20±0.1687.94±0.0483.71±3.7790.88±0.2590.37±0.1282.1090.8888.01
DLTTA (TMI'22) [66]71.88±5.5986.29±0.2180.70±0.1380.12±2.2690.53±0.1788.63±0.0681.46±3.3593.08±0.0887.48±0.0275.90±7.2784.27±0.1284.79±0.0682.75±6.5590.10±0.3189.33±0.1478.4288.8586.18
SAR (ICLR'23) [35]75.29±2.6988.73±0.3480.15±0.2686.42±1.5796.88±0.2791.28±0.0980.69±4.3092.10±0.1586.31±0.2780.27±5.4789.98±0.3187.20±0.1589.61±0.2895.90±0.0892.06±0.0182.4692.7187.40
DomainAdaptor (CVPR'23) [70]77.14±1.2190.80±0.1981.31±0.0283.95±0.8093.17±0.0990.44±0.0882.79±1.0995.36±0.1088.57±0.1386.50±3.2995.26±0.3391.07±0.2086.24±1.4494.90±0.2993.77±0.1883.3293.9089.03
DeY-Net (WACV'24) [62]79.67±1.3589.14±0.2383.68±0.0490.07±3.1290.06±0.3392.31±0.0785.94±1.8593.54±0.2287.15±0.1083.47±1.4095.29±0.1788.70±0.0488.76±2.2895.43±0.2190.45±0.0885.5894.0988.46
VPTTA (CVPR'24) [8]77.85±1.4094.09±0.0183.97±0.0584.57±0.9995.38±0.0389.53±0.0983.61±1.6796.39±0.0489.64±0.0582.75±0.5096.53±0.0489.83±0.0189.74±1.5198.22±0.2494.77±0.0783.7096.1289.55
NC-TTT (CVPR'24) [38]80.83±1.7789.61±0.2384.41±0.0887.07±1.3195.63±0.0690.09±0.0286.93±2.2794.69±0.1188.39±0.0382.10±4.2891.49±0.2286.33±0.0790.08±0.0995.93±0.1091.53±0.0985.4093.4788.15
Ours82.53±1.5291.85±0.2285.50±0.2088.98±0.8996.69±0.0491.63±0.1288.73±0.7095.80±0.0689.71±0.0290.25±1.3397.97±0.1992.89±0.1391.83±0.7197.14±0.2092.82±0.0988.4695.8490.51
", + "image_path": "b9794422715c50afd8e9b9e2426c56b94f3653fc4f23c9979ba1be649db4f04f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 219, + 541, + 299 + ], + "blocks": [ + { + "bbox": [ + 46, + 186, + 546, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 546, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 546, + 218 + ], + "type": "text", + "content": "Table 2. Multi sources domain generalization in the polyp segmentation. The average performance (mean ± standard deviation) of three trials for our method and nine SOTA methods. \"Site A\" means training on Sites B-D and testing on Site A, and similarly for the others. The best results are highlighted in red." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 219, + 541, + 299 + ], + "lines": [ + { + "bbox": [ + 49, + 219, + 541, + 299 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 541, + 299 + ], + "type": "table", + "html": "
MethodsSite ASite BSite CSite DAverage
\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)\\(DSC\\uparrow\\)\\(E_{\\phi}^{max}\\)\\(S_{\\alpha}\\)
No Adapt (U-Net [45])82.67±2.2492.11±0.0789.59±0.1081.20±4.3892.46±0.2086.08±0.0678.33±1.6294.00±0.0885.35±0.0175.51±2.7485.90±0.1781.71±0.2179.4391.1285.68
TENT (ICLR'21) [56]78.69±1.1288.28±0.1485.44±0.0877.71±3.0989.73±0.0981.90±0.1280.19±3.3596.03±0.0487.80±0.1772.63±5.1182.47±0.3078.92±0.1477.3189.1383.52
TASD (AAAI'22) [28]84.60±4.1492.55±0.2190.76±0.1085.28±2.2093.10±0.1386.83±0.1982.43±5.3996.31±0.0889.61±0.0280.04±4.4188.26±0.1184.09±0.0383.0992.5687.82
DLTTA (TMI'22) [66]80.68±2.3790.01±0.2186.89±0.0980.86±1.7791.98±0.1085.83±0.0475.57±3.1293.90±0.0884.15±0.1074.25±3.4083.49±0.1279.44±0.1077.8489.8584.08
SAR (ICLR'23) [35]79.37±0.7788.40±0.1286.11±0.0378.40±2.0790.23±0.1183.39±0.0280.94±0.9095.83±0.1286.84±0.0778.23±2.9186.37±0.0981.66±0.1479.2490.2184.50
DomainAdaptor (CVPR'23) [70]88.58±0.9695.03±0.0490.95±0.0181.12±1.0792.31±0.0886.60±0.0281.77±2.1795.18±0.0387.19±0.0179.91±1.3387.23±0.0483.64±0.0582.8592.4487.10
DeY-Net (WACV'24) [62]83.61±2.3190.79±0.1189.25±0.0980.90±3.1791.49±0.1085.36±0.0380.46±1.4395.83±0.0887.14±0.0373.85±3.0483.00±0.1378.96±0.0879.7190.2885.18
VPTTA (CVPR'24) [8]82.34±0.6991.03±0.0888.35±0.0184.32±0.4491.53±0.0787.46±0.0284.43±0.5295.75±0.1688.91±0.0382.40±0.2989.60±0.0485.18±0.0183.3791.9787.47
NC-TTT (CVPR'24) [38]87.96±1.2293.71±0.1390.66±0.0783.70±2.0893.78±0.0987.51±0.0789.29±0.9096.66±0.2291.26±0.0778.23±0.7986.37±0.0981.66±0.0684.8092.6387.77
Ours90.97±0.6696.57±0.0692.92±0.0887.49±0.8295.21±0.0489.93±0.0986.87±0.5897.53±0.0590.16±0.1283.97±1.1890.67±0.1386.44±0.0887.3294.6789.86
", + "image_path": "381617be161b999c102842e331550a7610714ee6152ae8c0ba74282e4ffc9582.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "text", + "content": "similar trend was observed in the polyp segmentation results, where we outperformed the second-best method (NC-TTT [38]) and No Adapt by " + }, + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "inline_equation", + "content": "2.52\\%" + }, + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "inline_equation", + "content": "7.89\\%" + }, + { + "bbox": [ + 46, + 319, + 287, + 378 + ], + "type": "text", + "content": ", respectively. Additionally, our approach exhibited greater stability across multiple trials compared to the other methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "text", + "content": "Single-Source Generalization. In these experiments, the models are trained on one domain and tested on the remaining datasets (i.e., with " + }, + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "inline_equation", + "content": "S = 1" + }, + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "inline_equation", + "content": "T = |\\mathcal{D}_s \\cup \\mathcal{D}_t| - 1" + }, + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "text", + "content": "). Compared to multi-source generalization, single-source training is considered more challenging due to the limited domain information available during the training phase. By comparing Tables 2 and 3, we can observe that all models show lower segmentation performance (in terms of DSC) in the single-source setting compared to the multi-source results. Some methods even perform worse than the No Adapt baseline in single-source scenarios. However, our method still achieves the best average performance across 12 domain transfer experiments, outperforming the second-best approach (NC-TTT [38]) by " + }, + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "inline_equation", + "content": "1.83\\%" + }, + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "text", + "content": " and exceeding the No Adapt by " + }, + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "inline_equation", + "content": "9.77\\%" + }, + { + "bbox": [ + 46, + 379, + 287, + 557 + ], + "type": "text", + "content": " in terms of DSC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Fig. 3 presents the visualization results for \"Site A\" in Table 1. The No Adapt baseline exhibits significant misalignment, with notably distorted shapes and inaccurate boundaries. VPTTA offers some improvement but still suffers from distortions and incomplete segmentation, particularly in the optic cup area. NC-TTT shows instances of overlapping segmentation. In contrast, our method delivers the most precise segmentation, closely aligning with the ground truth. The Grad-CAM [48] visualizations reinforce this, as the attention of network is focused specifically on the relevant regions. These results demonstrate our method's enhanced generalization to unseen domains by effectively incorporating morphological priors, resulting in higher seg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 317, + 545, + 416 + ], + "blocks": [ + { + "bbox": [ + 308, + 317, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 416 + ], + "type": "image", + "image_path": "ba7ca63d2bcda0fe288ae5adeb346a5e1ba7f633574dc4083f220e2b87ae6eb5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 425, + 546, + 480 + ], + "lines": [ + { + "bbox": [ + 304, + 425, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 546, + 480 + ], + "type": "text", + "content": "Figure 3. Visualization comparison of segmentation results and Grad-CAM outputs from the final layer of the backbone network for the No Adapt baseline, VPTTA [8], NC-TTT [38], and our proposed method on retinal fundus images. Additional visual comparisons are provided in the supplementary material." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 490, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 490, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 490, + 545, + 515 + ], + "type": "text", + "content": "mentation accuracy. For further visualization and additional experimental results, please refer to the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 529, + 365, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 365, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 365, + 542 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 550, + 523, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 523, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 523, + 563 + ], + "type": "text", + "content": "5.1. Effectiveness of the Universe Embeddings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "The universe embeddings " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " play a crucial role in incorporating morphological priors from medical images, resulting in an assignment matrix that projects each node to the universe of nodes. To validate the effectiveness during the TTA, we conducted the following experiments: (1) Without universe embeddings (denoted as " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "w / o\\mathcal{U}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "), the universe matching assignment matrix " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " is initialized following the setting in [59] (i.e. " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{U} = 1 / d + 10^{-3}z" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "z\\sim N(0,1)" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "), leading to random matching between the nodes and the universe; (2) Without morphological priors but with " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " (denoted as " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " priors). As shown in Table 4, the performance in (1) is slightly better than in (2) by " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.76\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": " in terms of DSC" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15627" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 92, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 92 + ], + "type": "text", + "content": "Table 3. Single source domain generalization in the polyp segmentation. The average performance of three trials for our method and nine SOTA methods. A → B represents models trained on Site A and tested on Site B, and similar for others. Best results are colored as red." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 92, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 48, + 92, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 92, + 541, + 178 + ], + "type": "table", + "html": "
MethodsA→BA→CA→DB→AB→CB→DC→AC→BC→DD→AD→BD→CAvg.
Dice Score Metric ↑ (DSC, mean±std)
No Adapt (U-Net [45])76.49±1.4867.70±3.2770.94±7.5676.21±2.9062.80±8.0370.03±4.5372.87±3.6269.13±5.1570.44±3.2779.78±1.0872.47±2.8074.91±3.1871.98
TENT (ICLR'21) [56]68.13±0.1865.99±0.1166.11±0.1477.56±0.2159.87±0.2068.89±0.1473.21±0.1065.04±0.0972.35±0.2277.88±0.1470.01±0.1173.60±0.2069.89
TASD (AAAI'22) [28]73.02±2.2479.67±6.3778.23±4.4466.82±9.1772.09±4.4077.95±1.6974.20±3.3069.74±7.2776.13±2.8283.77±0.5176.09±3.2680.10±1.4775.65
DLTTA (TMI'22) [66]72.23±0.0470.78±0.0674.56±0.0464.87±0.0160.05±0.0978.47±0.0371.22±0.0365.78±0.1071.00±0.1379.89±0.0879.10±0.1074.52±0.0171.87
SAR (ICLR'23) [35]70.34±0.1377.90±0.2077.71±0.0968.19±0.1266.56±0.0873.92±0.1376.12±0.1171.29±0.1975.08±0.0882.07±0.0480.44±0.2183.10±0.1575.22
DomainAdaptor (CVPR'23) [70]78.39±0.6177.09±0.3579.45±0.2174.98±0.7070.07±0.6680.21±0.4676.02±0.3373.40±0.5081.33±0.5486.10±0.4380.56±0.6277.86±0.2177.96
DeY-Net (WACV'24) [62]74.45±1.3176.07±2.5280.11±0.7270.90±1.1069.23±2.1876.31±1.3278.33±4.4374.60±1.6673.52±2.2580.70±0.2778.23±1.1781.48±0.9376.16
VPTTA (CVPR'24) [8]78.18±0.0175.14±0.0382.73±0.0273.65±0.0664.78±0.0182.86±0.0377.90±0.0170.64±0.0577.22±0.1087.19±0.0384.97±0.0882.12±0.0278.12
NC-TTT (CVPR'24) [38]79.21±0.2481.76±0.5783.22±0.6977.35±0.3379.40±0.6780.94±0.5480.17±0.7072.36±0.6278.55±0.4384.09±0.3882.38±0.5179.77±0.7479.93
Ours80.47±0.2883.93±0.1983.18±0.3378.33±0.2681.25±0.2383.58±0.3478.86±0.2676.49±0.2180.20±0.2489.17±0.2281.68±0.2283.86±0.2881.75
", + "image_path": "e407af8022473af55059761f17cb4436b7e568e64f90b863695acf92a9b3df80.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 220, + 286, + 269 + ], + "blocks": [ + { + "bbox": [ + 47, + 198, + 286, + 218 + ], + "lines": [ + { + "bbox": [ + 47, + 198, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 286, + 218 + ], + "type": "text", + "content": "Table 4. Ablation study in the retinal fundus segmentation. Details of the experiment can be found in Sec 5.1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 220, + 286, + 269 + ], + "lines": [ + { + "bbox": [ + 48, + 220, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 286, + 269 + ], + "type": "table", + "html": "
MethodsSite ASite BSite CSite DSite EAverage
Dice Score Metric ↑ (DSC, mean±std)
No Adapt65.60±5.7874.65±4.8863.15±7.3068.11±5.4975.34±1.0169.37
w/o U77.09±1.9083.24±2.0180.50±1.4478.32±1.5583.18±2.1880.47
w/o priors79.51±0.8079.36±1.1377.40±1.8977.29±1.7780.00±1.9378.71
Ours82.53±1.5288.98±0.8988.73±0.7090.25±1.3391.83±0.7188.46
", + "image_path": "3f8e850d76b86a6f9425eb75b7370a0a1c3d506b29df99986eea594711b7768c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 283, + 161, + 380 + ], + "blocks": [ + { + "bbox": [ + 52, + 283, + 161, + 380 + ], + "lines": [ + { + "bbox": [ + 52, + 283, + 161, + 380 + ], + "spans": [ + { + "bbox": [ + 52, + 283, + 161, + 380 + ], + "type": "image", + "image_path": "ff352ef9cde431b944aa81e57fbb43239ac968d87e3f0fdf79b362279d09c376.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 390, + 287, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 445 + ], + "type": "text", + "content": "Figure 4. The performance of DSC (\\%) and inference time (s) with different universe sizes " + }, + { + "bbox": [ + 46, + 390, + 287, + 445 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 390, + 287, + 445 + ], + "type": "text", + "content": " are shown, with the experimental setup identical to that in Table 1. The first curve represents the No Adapt baseline (" + }, + { + "bbox": [ + 46, + 390, + 287, + 445 + ], + "type": "inline_equation", + "content": "d = 0" + }, + { + "bbox": [ + 46, + 390, + 287, + 445 + ], + "type": "text", + "content": "), while the subsequent curve corresponds to the results obtained with a universe size of " + }, + { + "bbox": [ + 46, + 390, + 287, + 445 + ], + "type": "inline_equation", + "content": "d = 120" + }, + { + "bbox": [ + 46, + 390, + 287, + 445 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 283, + 284, + 380 + ], + "blocks": [ + { + "bbox": [ + 175, + 283, + 284, + 380 + ], + "lines": [ + { + "bbox": [ + 175, + 283, + 284, + 380 + ], + "spans": [ + { + "bbox": [ + 175, + 283, + 284, + 380 + ], + "type": "image", + "image_path": "de7f794bac43d8fe40e35f5601014314a8675039b1a17e45c3ba77bd5f018919.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "type": "text", + "content": "on average. However, when using the pre-trained " + }, + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "type": "text", + "content": " derived from the source model, the segmentation results show a significant improvement. This highlights the effectiveness of the pre-trained morphological priors embedded in " + }, + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "type": "text", + "content": " for medical imaging tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 515, + 225, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 225, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 225, + 528 + ], + "type": "text", + "content": "5.2. Effectiveness of the Universe Size" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "The universe size " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " directly impacts both the matching accuracy and computational efficiency, and its selection depends on the number of sampled nodes. Previous works [4, 36, 59], ensured that the " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " matched the number of nodes in each graph by labeling an equal number of keypoints. In our experiments, " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is determined by the spatially-uniform sampling described in Sec 3.1, and we empirically set " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "d = 100 \\times (n + 1)s^{-1}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is the sampling step and " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is the number of segmentation categories required for the task. We evaluated several " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": ", with the results shown in Fig. 4. We observe that excessively small or large step sizes—resulting in too many or too few sampled nodes—negatively impact both training speed and performance. However, when " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is within a reasonable range (in this experiment, " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "s \\in [2,10]" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " i.e. " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "d \\in [30,150]" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "), the model maintains good scalability" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 199, + 378, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 199, + 378, + 210 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 378, + 210 + ], + "type": "text", + "content": "and performance." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 255, + 541, + 342 + ], + "blocks": [ + { + "bbox": [ + 305, + 210, + 545, + 253 + ], + "lines": [ + { + "bbox": [ + 305, + 210, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 545, + 253 + ], + "type": "text", + "content": "Table 5. Comparison of common variables and performance between pairwise matching and multi-matching, using the same experimental setup as in Table 1. Both methods follow the graph construction in Sec. 3.1." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 255, + 541, + 342 + ], + "lines": [ + { + "bbox": [ + 308, + 255, + 541, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 541, + 342 + ], + "type": "table", + "html": "
paramSite A-E Averagedescription
Pairwise [58]Multi-Matching (Ours)
lr10-310-3learning rate
batch44batch size in inference
h256256dimension of node feature
τ0.050.05regularization factor of Sinkhorn
Iter2020max iterations of Sinkhorn
d-120universe size
DSC ↑84.6388.46dice score metric (%)
time ↓0.9300.392inference time per image (s/img)
Param ↓1.0710.658parameter count (M)
FLOPs ↓15.354.255floating point operations per second (G)
", + "image_path": "13569d79fee7530caa9328bbc367e0a58ee4166e65aedc6ce8ee67681dac8e11.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 343, + 511, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 343, + 511, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 343, + 511, + 356 + ], + "type": "text", + "content": "5.3. Multi-Matching vs. Pairwise Matching" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "type": "text", + "content": "Compared to pairwise matching, multi-matching uses cycle-consistency for global optimization, which helps avoid local optima in pairwise. Moreover, joint optimization reduces the computational complexity of large-scale matching. By incorporating morphological consistency constraints, multi-matching exhibits greater robustness. To validate these claims, we conducted experiments comparing the two approaches. For the pairwise matching method, we adopted the benchmark approach [58] and applied the same graph construction process as in multi-matching. The comparison results are presented in Table 5. It is evident that multi-matching outperforms pairwise matching in terms of DSC, inference time, parameters and FLOPs. The inference time of multi-matching was reduced by approximately " + }, + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "type": "inline_equation", + "content": "57.85\\%" + }, + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "type": "text", + "content": " compared to pairwise matching, while segmentation accuracy improved by " + }, + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "type": "inline_equation", + "content": "3.83\\%" + }, + { + "bbox": [ + 304, + 361, + 545, + 553 + ], + "type": "text", + "content": " in DSC." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "This paper presents a novel multi-matching framework for TTDG, which leverages universe embeddings to incorporate morphological priors from medical images while ensuring cycle-consistency, along with an unsupervised test-time paradigm that fully integrates these priors for efficient adaptation. Extensive experiments, including comparisons and ablation studies demonstrate that our graph-matching-based approach achieves SOTA performance, outperforming entropy-based, template-based, batch normalization-based, and noise estimation-based methods across both multi-source and single-source domain generalization tasks." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15628" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 169 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (Grant No. 62306003), the Open Research Fund from Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), under Grant No. GML-KF-24-29 and the Open Foundation of Jiangxi Provincial Key Laboratory of Image Processing and Pattern Recognition (ET202404437)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 193, + 105, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 193, + 105, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 193, + 105, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 212, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 212, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 212, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 212, + 287, + 257 + ], + "type": "text", + "content": "[1] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 323 + ], + "type": "text", + "content": "[2] Jorge Bernal, F Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilarino. Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Computerized Medical Imaging and Graphics, 43:99-111, 2015. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 326, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 326, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 326, + 287, + 370 + ], + "type": "text", + "content": "[3] Florian Bernard, Johan Thunberg, Jorge Goncalves, and Christian Theobalt. Synchronisation of partial multi-matchings via non-negative factorisations. Pattern Recognition, 92:146-155, 2019. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 372, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 372, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 372, + 287, + 415 + ], + "type": "text", + "content": "[4] Florian Bernard, Johan Thunberg, Paul Swoboda, and Christian Theobalt. Hippi: Higher-order projected power iterations for scalable multi-matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 10284-10293, 2019. 3, 5, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 418, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 287, + 461 + ], + "type": "text", + "content": "[5] Yuan Bi, Zhongliang Jiang, Ricarda Clarenbach, Ghotbi, and Nassir Navab. Mi-segnet: Mutual information-based us segmentation for unseen domain generalization. In MICCAI, pages 130-140. Springer, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 464, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 287, + 507 + ], + "type": "text", + "content": "[6] Liang Chen, Yong Zhang, Yibing Song, Ying Shan, and Lingqiao Liu. Improved test-time adaptation for domain generalization. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 24172-24182, 2023. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 509, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 509, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 53, + 509, + 287, + 541 + ], + "type": "text", + "content": "[7] Yuxin Chen, Leonidas Guibas, and Qixing Huang. Near-optimal joint object matching via convex relaxation. In ICML, pages II-100, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 544, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 597 + ], + "type": "text", + "content": "[8] Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each test image deserves a specific prompt: Continual test-time adaptation for 2d medical image segmentation. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 11184-11193, 2024. 3, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 53, + 600, + 287, + 632 + ], + "type": "text", + "content": "[9] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Adv. Neural Inform. Process. Syst., 26, 2013. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "text", + "content": "[10] Frank De Groot. High-resolution x-ray emission and x-ray absorption spectroscopy. Chemical Reviews, 101(6):1779-1808, 2001. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[11] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4548-4557, 2017. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[12] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. *IJCAI*, 2018. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[13] Francisco Fumero, Silvia Alayón, José L Sanchez, Jose Sigut, and M Gonzalez-Hernandez. Rim-one: An open retinal image database for optic nerve evaluation. In International Symposium on Computer-based Medical Systems (CBMS), pages 1-6. IEEE, 2011. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "type": "text", + "content": "[14] Quankai Gao, Fudong Wang, Nan Xue, Jin-Gang Yu, and Gui-Song Xia. Deep graph matching under quadratic constraint. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5069-5078, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[15] Steven Gold, Anand Rangarajan, et al. Softmax to softassign: Neural network algorithms for combinatorial optimization. Journal of Artificial Neural Networks, 2(4):381-399, 1996. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "type": "text", + "content": "[16] Ran Gu, Guotai Wang, Jiangshan Lu, Jingyang Zhang, Wenhui Lei, Yinan Chen, Wenjun Liao, Shichuan Zhang, Kang Li, Dimitris N Metaxas, et al. Cddsa: Contrastive domain disentanglement and style augmentation for generalizable medical image segmentation. Medical Image Analysis, 89: 102904, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 363 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 770-778, 2016. 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 421 + ], + "type": "text", + "content": "[18] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, pages 451-462. Springer, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 464 + ], + "type": "text", + "content": "[19] Neerav Karani, Ertunc Erdil, Krishna Chaitanya, and Ender Konukoglu. Test-time adaptable neural networks for robust medical image segmentation. Medical Image Analysis, 68: 101907, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 466, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 509 + ], + "type": "text", + "content": "[20] Boah Kim, Inhwa Han, and Jong Chul Ye. Diffuse morph: Unsupervised deformable image registration using diffusion model. In Eur. Conf. Comput. Vis., pages 347-364. Springer, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 511, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 553 + ], + "type": "text", + "content": "[21] Tjalling C Koopmans and Martin Beckmann. Assignment problems and the location of economic activities. *Econometrica: journal of the Econometric Society*, pages 53–76, 1957. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 556, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 609 + ], + "type": "text", + "content": "[22] Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Kulkarni, Hiran Sarkar, Varun Jampani, and R Venkatesh Babu. Concurrent subsidiary supervision for unsupervised source-free domain adaptation. In ECCV, pages 177-194. Springer, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 644 + ], + "type": "text", + "content": "[23] Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. nature, 401 (6755):788-791, 1999. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[24] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5400-5409, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "type": "text", + "content": "[25] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semanticcomplete graph matching for domain adaptive object detec" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15629" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "tion. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 5291-5300, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[26] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):9022-9040, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[27] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. Int. J. Comput. Vis., pages 1-34, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "text", + "content": "[28] Quande Liu, Cheng Chen, Qi Dou, and Pheng-Ann Heng. Single-domain generalization in medical image segmentation via test-time adaptation from shape dictionary. In AAAI, pages 1756-1764, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 253 + ], + "type": "text", + "content": "[29] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 297 + ], + "type": "text", + "content": "[30] Eleonora Maset, Federica Arrigoni, and Andrea Fusiello. Practical and efficient multi-view matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4568-4576, 2017. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 300, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 287, + 342 + ], + "type": "text", + "content": "[31] Jose G Moreno-Torres, Troy Raeder, Rocio Alaiz-Rodriguez, Nitesh V Chawla, and Francisco Herrera. A unifying view on dataset shift in classification. Pattern recognition, 45(1):521-530, 2012. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 344, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 398 + ], + "type": "text", + "content": "[32] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 400, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 466 + ], + "type": "text", + "content": "[33] Phan Ngoc Lan, Nguyen Sy An, Dao Viet Hang, Dao Van Long, Tran Quang Trung, Nguyen Thi Thuy, and Dinh Viet Sang. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In Advances in visual Computing: 16th International Symposium (ISVC), pages 15-28. Springer, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "text", + "content": "[34] Andy Nguyen, Mirela Ben-Chen, Katarzyna Welnicka, Yinyu Ye, and Leonidas Guibas. An optimization approach to improving collections of shape maps. In Computer Graphics Forum, pages 1481-1491. Wiley Online Library, 2011. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "type": "text", + "content": "[35] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. ICLR, 2023. 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "text", + "content": "[36] Zhakshylyk Nurlanov, Frank R Schmidt, and Florian Bernard. Universe points representation learning for partial multi-graph matching. In AAAI, pages 1984-1992, 2023. 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 677 + ], + "type": "text", + "content": "[37] José Ignacio Orlando, Huazhu Fu, João Barbosa Breda, Karel Van Keer, Deepti R Bathula, Andrés Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, et al. Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs. Medical Image Analysis, 59:101570, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[38] David Osowiechi, Gustavo A Vargas Hakim, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdanpanah, Ismail Ben Ayed, and Christian Desrosiers. Nc-ttt: A" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "noise constrastive approach for test-time training. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 6078-6086, 2024. 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 139 + ], + "type": "text", + "content": "[39] Deepti Pachauri, Risi Kondor, and Vikas Singh. Solving the multi-way matching problem by permutation synchronization. Adv. Neural Inform. Process. Syst., 26, 2013. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 182 + ], + "type": "text", + "content": "[40] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 1406-1415, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "type": "text", + "content": "[41] Bin Pu, Xingguo Lv, Jiewen Yang, He Guannan, Xingbo Dong, Yiqun Lin, Li Shengli, Tan Ying, Liu Fei, Ming Chen, Zhe Jin, Kenli Li, and Xiaomeng Li. Unsupervised domain adaptation for anatomical structure detection in ultrasound images. In ICML, pages 41204-41220. PMLR, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[42] Bin Pu, Liwen Wang, Jiewen Yang, Guannan He, Xingbo Dong, Shengli Li, Ying Tan, Ming Chen, Zhe Jin, Kenli Li, et al. M3-uda: A new benchmark for unsupervised domain adaptive fetal cardiac structure detection. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 11621-11630, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "text", + "content": "[43] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 12556-12565, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 359 + ], + "type": "text", + "content": "[44] Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Droppedge: Towards deep graph convolutional networks on node classification. In ICLR, 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 361, + 545, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 393 + ], + "type": "text", + "content": "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "text", + "content": "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. IEEE/CVF Comput. Vis. Pattern Recog., pages 4938-4947, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "text", + "content": "[47] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Adv. Neural Inform. Process. Syst., 33:11539-11551, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "text", + "content": "[48] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 618-626, 2017. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "text", + "content": "[49] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Adv. Neural Inform. Process. Syst., 35: 14274-14289, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "text", + "content": "[50] Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer. International Journal of Computer Assisted Radiology and Surgery, 9:283-293, 2014. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[51] Richard Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices. The annals of mathematical statistics, 35(2):876-879, 1964. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[52] Jayanthi Sivaswamy, SR Krishnadas, Gopal Datt Joshi, Madhulika Jain, and A Ujjwaft Syed Tabish. Drishti-gs: Retinal" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15630" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "image dataset for optic nerve head (onh) segmentation. In International Symposium on Biomedical Imaging (ISBI), pages 53-56. IEEE, 2014. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 129 + ], + "type": "text", + "content": "[53] Charles P Slichter. Principles of magnetic resonance. Springer Science & Business Media, 2013. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 132, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 132, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 132, + 288, + 163 + ], + "type": "text", + "content": "[54] Gail Ter Haar. Therapeutic applications of ultrasound. Progress in biophysics and molecular biology, 93(1-3):111-129, 2007. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 167, + 288, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 167, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 167, + 288, + 209 + ], + "type": "text", + "content": "[55] Roberto Tron, Xiaowei Zhou, Carlos Esteves, and Kostas Daniilidis. Fast multi-image matching via density-based clustering. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4057-4066, 2017. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 213, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 213, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 288, + 245 + ], + "type": "text", + "content": "[56] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 247, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 247, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 48, + 247, + 288, + 302 + ], + "type": "text", + "content": "[57] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and S Yu Philip. Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering, 35(8):8052-8072, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 288, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 347 + ], + "type": "text", + "content": "[58] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Learning combinatorial embedding networks for deep graph matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 3056-3065, 2019. 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 404 + ], + "type": "text", + "content": "[59] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Graduated assignment for joint multi-graph matching and clustering with application to unsupervised graph matching network learning. Adv. Neural Inform. Process. Syst., 33: 19908-19919, 2020. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 407, + 288, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 407, + 288, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 288, + 461 + ], + "type": "text", + "content": "[60] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Neural graph matching network: Learning lawler's quadratic assignment problem with extension to hypergraph and multiple-graph matching. IEEE Trans. Pattern Anal. Mach. Intell., 44 (9):5261-5279, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 464, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 288, + 507 + ], + "type": "text", + "content": "[61] Runzhong Wang, Junchi Yan, and Xiaokang Yang. Unsupervised learning of graph matching with mixture of modes via discrepancy minimization. IEEE Trans. Pattern Anal. Mach. Intell., 45(8):10500-10518, 2023. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 288, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 574 + ], + "type": "text", + "content": "[62] Ruxue Wen, Hangjie Yuan, Dong Ni, Wenbo Xiao, and Yaoyao Wu. From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In Proc. IEEE/CVF Winter Conference on Applications of Computer Vision, pages 464-474, 2024. 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 621 + ], + "type": "text", + "content": "[63] Junchi Yan, Yu Tian, Hongyuan Zha, Xiaokang Yang, Ya Zhang, and Stephen M Chu. Joint optimization for consistent multiple graph matching. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 1649-1656, 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 288, + 667 + ], + "type": "text", + "content": "[64] Junchi Yan, Yin Li, Wei Liu, Hongyuan Zha, Xiaokang Yang, and Stephen Mingyu Chu. Graduated consistency-regularized optimization for multi-graph matching. In ECCV, pages 407–422. Springer, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[65] Junchi Yan, Minsu Cho, Hongyuan Zha, Xiaokang Yang, and Stephen M Chu. Multi-graph matching via affinity optimization with graduated consistency regularization. IEEE Trans. Pattern Anal. Mach. Intell., 38(6):1228-1242, 2015. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 699 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[66] Hongzheng Yang, Cheng Chen, Meirui Jiang, Quande Liu, Jianfeng Cao, Pheng Ann Heng, and Qi Dou. Dltta: Dynamic learning rate for test-time adaptation on cross-domain medical images. IEEE Transactions on Medical Imaging, 41 (12):3575-3586, 2022. 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 182 + ], + "type": "text", + "content": "[67] Jiewen Yang, Xinpeng Ding, Ziyang Zheng, Xiaowei Xu, and Xiaomeng Li. Graphecho: Graph-driven unsupervised domain adaptation for echocardiogram video segmentation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 11878-11887, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[68] Jee Seok Yoon, Kwanseok Oh, Yooseung Shin, Maciej A Mazurowski, and Heung-II Suk. Domain generalization for medical image analysis: A survey. arXiv preprint arXiv:2310.08598, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "text", + "content": "[69] Weiyi Yu, Zhizhong Huang, Junping Zhang, and Hongming Shan. San-net: Learning generalization to unseen sites for stroke lesion segmentation with self-adaptive normalization. Computers in Biology and Medicine, 156:106717, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 275, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 317 + ], + "type": "text", + "content": "[70] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Domain adaptor: A novel approach to test-time adaptation. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 18971-18981, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 319, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 385 + ], + "type": "text", + "content": "[71] Ling Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Holger Roth, Andriy Myronenko, Daguang Xu, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation. IEEE transactions on medical imaging, 39(7):2531-2540, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "text", + "content": "[72] Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. Adaptive risk minimization: Learning to adapt to domain shift. Adv. Neural Inform. Process. Syst., 34:23664-23678, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 431, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 431, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 545, + 463 + ], + "type": "text", + "content": "[73] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Adv. Neural Inform. Process. Syst., 35:38629-38642, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "text", + "content": "[74] Zhuo Zhang, Feng Shou Yin, Jiang Liu, Wing Kee Wong, Ngan Meng Tan, Beng Hai Lee, Jun Cheng, and Tien Yin Wong. Origa-light: An online retinal fundus image database for glaucoma analysis and research. In 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology, pages 3065-3068. IEEE, 2010. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 532, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 574 + ], + "type": "text", + "content": "[75] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 45(4):4396-4415, 2022. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 576, + 545, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 545, + 619 + ], + "type": "text", + "content": "[76] Xiaowei Zhou, Menglong Zhu, and Kostas Daniilidis. Multi-image matching via fast alternating minimization. In Proc. IEEE/CVF Int. Conf. Comput. Vis., pages 4032-4040, 2015. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 621, + 545, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 621, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 307, + 621, + 545, + 665 + ], + "type": "text", + "content": "[77] Yuliang Zou, Zizhao Zhang, Chun-Liang Li, Han Zhang, Tomas Pfister, and Jia-Bin Huang. Learning instance-specific adaptation for cross-domain segmentation. In ECCV, pages 459-476. Springer, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 666, + 545, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 545, + 699 + ], + "type": "text", + "content": "[78] Yukun Zuo, Hantao Yao, and Changsheng Xu. Attention-based multi-source domain adaptation. IEEE Transactions on Image Processing, 30:3793-3803, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15631" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_content_list.json b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d020edbc83f2272853060e91e5d34f639fcc887a --- /dev/null +++ b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_content_list.json @@ -0,0 +1,1481 @@ +[ + { + "type": "text", + "text": "Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability", + "text_level": 1, + "bbox": [ + 228, + 130, + 772, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unki Park $^{1,2}$ Seongmoon Jeong $^{1}$ Youngchan Jang $^{3}$ Gyeong-Moon Park $^{4*}$ Jong Hwan Ko $^{1*}$ $^{1}$ Sungkyunkwan University, $^{2}$ Samsung Institute of Technology, $^{3}$ Hanwha Systems Corporation, $^{4}$ Korea University \n{unkipark@g., giftjsm@, hjko@}skku.edu, youngchan84.jang@hanwha.com, gm-park@korea.ac.kr", + "bbox": [ + 107, + 202, + 892, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of computer vision was initially inspired by the human visual system and has progressively expanded to include a broader range of machine vision applications. Consequently, image compressors should be designed to effectively accommodate not only human visual perception but also machine vision tasks, including closed-set scenarios that enable pre-training and open-set scenarios that involve previously unseen tasks at test time. Many recent studies effectively address both human visual perception and closed-set machine vision tasks simultaneously but struggle to handle open-set machine vision tasks. To address this issue, this paper proposes a fully instance-specific test time fine-tuning (TTFT) for adapting learned image compression (LIC) to both closed-set and open-set machine vision tasks effectively. With our method, a large-scale LIC model, originally trained for human perception, is adapted to the target task through TTFT using Singular Value Decomposition based Low Rank Adaptation (SVD-LoRA). During TTFT, the decoder adopts a modified learning scheme that focuses exclusively on training the singular values, which helps prevent excessive bitstream overhead. This enables fully instance-specific optimization for the target task, even for open-set tasks. Experimental results demonstrate that the proposed method effectively adapts the backbone compressor to diverse machine vision tasks, outperforming competing methods. The code is available at project page.", + "bbox": [ + 89, + 323, + 483, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 744, + 220, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the early stages of computer vision, image processing technologies were primarily developed to align with human perception, as images were largely intended for human interpretation [3, 26, 51]. However, with advances in deep learning and computer vision techniques enabling more sophisticated data processing, machines have increasingly gained the ability to interpret visual data. Conse", + "bbox": [ + 89, + 771, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c316734b66866dee51e1626811074b3748e8d1c3c92146f9f0118e57b7e2012.jpg", + "image_caption": [ + "Figure 1. An overview of our proposed method. Our method tunes both the encoder and decoder of a backbone human-target LIC model by fine-tuning SVD-LoRA weights optimized for each target task and image instance." + ], + "image_footnote": [], + "bbox": [ + 513, + 287, + 898, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sequently, the scope of computer vision has broadened to include machine vision applications [56, 58, 66]. In the field of image compression as well, there is a growing need for methods tailored to the requirements of machine vision, as machines are expected to become the primary users of compressed image data [18, 65].", + "bbox": [ + 509, + 593, + 906, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accordingly, several research studies have shifted their focus away from the traditional goal of visual signal reconstruction, instead prioritizing the enhancement of machine vision task performance [9, 59]. These studies commonly aim to effectively restore representations suitable for the targeted machine vision domain, while selectively sacrificing visual quality to achieve this goal.", + "bbox": [ + 511, + 686, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taking it further, recent research has focused on developing learned image codes (LICs) that can cater to both human perception and various machine vision tasks [10, 13, 37, 38, 41, 53, 60]. These studies can be categorized into methods that generate either multi-task bitstreams or single-task bitstreams. The multi-task bitstream approach aims to generate a unified bitstream that can support vari-", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4430", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ous subtasks including human perception and multiple machine tasks. This method offers flexibility in responding to a range of machine vision tasks, but it introduces challenges. For example, scalable coding [15, 32, 45, 62] increases the processing complexity and bitstream overhead for individual downstream tasks, and unified image coding [11, 19, 20, 24] struggles with optimizing each subtask individually.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, the single-task bitstream approach [13, 36, 42, 43, 48, 53] avoids the limitations of multi-task methods by dynamically generating bitstreams for a specific downstream task. For instance, TransTIC[13] utilizes prompt tuning for adapting the human-target backbone codec to generate bitstreams optimized for the other tasks specified by the prompt. While this technique enables switching of pre-defined (closed-set) tasks without modifying the backbone codec, it requires task-specific pre-training of the prompt generator. Therefore, it leads to increased training demands as the number of tasks grows, and faces challenges to handle new, untrained (open-set) tasks. Moreover, performance improvement is non-monotonic with respect to prompt size, and computational costs increase due to the added prompt during inference.", + "bbox": [ + 89, + 215, + 482, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we aim to develop a method that can dynamically adapt a given backbone codec to optimize rate-performance trade-offs for the specific requirements of each task. Our goal is to provide flexible support for diverse tasks and codec models, allowing adaptation to previously unseen (open-set) tasks without requiring pre-training, and also adaptable to any given codec model. To achieve this, as shown in Fig. 1, we propose tuning the backbone codec using Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning approach widely used in test-time scenarios. Unlike the TransTIC approach, which cannot perform instance-specific tuning on the decoder, our method tunes both the encoder and decoder by generating LoRA weights optimized for each target task and image instance. This enables dynamic codec optimization for any arbitrary task without requiring pre-training, making it effective for open-set scenarios. While instance-specific model optimization can lead to superior performance compared to task-specific optimization, it introduces additional overhead by requiring a separate set of parameters for each image. Specifically, in the case of instance-specific decoder tuning, transmitting the LoRA weights incurs additional bit rate overhead. To address this efficiently, we adopt an SVD-LoRA approach that combines the original LoRA weights with supplementary singular value (SV) components. After pre-initializing the original LoRA weights, we can then transmit only the SV components to the decoder on an instance-specific basis, allowing effective tuning while minimizing bit rate demands. As the SV component captures essential information for optimizing the codec per target in", + "bbox": [ + 89, + 446, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stance and task, it is both effective and efficient in managing rate requirements.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results demonstrate that our method achieves superior adaptation performance for both open-set and closed-set tasks. Although conducted under a lightweight experimental setup (40 iterations, LoRA rank 4), we observed that our method achieved more than a $50\\%$ improvement in Bjøntegaard Delta (BD)-Rate [23] compared to the backbone codec across all evaluated machine vision tasks. These results surpass all competing methods. Notably, our method achieves superior performance in a pure open-set setting—where no prior task-specific setup is available—outperforming competing methods even when they are evaluated under pre-initialized (closed-set) conditions.", + "bbox": [ + 511, + 121, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our work can be summarized as follows:", + "bbox": [ + 511, + 316, + 903, + 345 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Test-time fine-tuning in compression field : To the best of our knowledge, we are the first to attempt TTFT at the decoder stage in a image compression task. Considering that the encoder and decoder are temporally and spatially separated in compression systems, TTFT has the potential to maximize image compression quality.", + "- Novel approach utilizing LoRA: We propose a novel approach utilizing LoRA for efficient adaptation to diverse machine vision applications in the still-image compression task. Specifically, we introduce a modified LoRA in the form of SVD to enable instance-specific optimization at the decoder. This approach not only enhances efficiency but also facilitates effective TTFT at the decoder stage.", + "- Handling open-set tasks: By designing an instance-specific optimized model that does not necessarily require pre-training, we have enhanced the ability to tackle open-set machine vision tasks, which have remained challenging in previous research." + ], + "bbox": [ + 513, + 348, + 903, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 648, + 656, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The initial development of LIC focused on optimizing perceptual quality for human visual systems, similar to traditional compressors [1, 7, 55]. Most LIC models adopt an autoencoder structure [4, 5, 44, 47, 50], aiming to optimize the balance between distortion and bit-rate. In early LIC research, the goal was to minimize losses limited to human visual perception, with the technology evolving towards faithfully mimicking human visual quality [28, 46, 49]. With the rapid development of machine vision tasks such as image classification and object detection, an increasing number of applications now incorporate images encoded and decoded by image codecs for these downstream tasks. Consequently, LIC research has begun to focus on meeting machine vision requirements. Initially, the emphasis was on creating", + "bbox": [ + 511, + 674, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4431", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "individual models tailored to each specific task. More recently, research has proposed approaches where a shared LIC model architecture is individually trained to generate task-specific models. Following this trend, research efforts aimed at designing a single LIC model to support multiple targets including human perception and diverse machine vision tasks. These studies can be divided into multi-task bitstream and single-task bitstream methods.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-task Bitstream LICs for Multiple Targets. One line of research in multi-target LIC is the multi-task bitstream approach [11, 15, 19, 20, 24, 32, 45, 62], which aims to accommodate multiple targets using a single bitstream. However, this approach increases the complexity of both the encoder and decoder [15, 24]. It also leads to significant bitstream overhead from the perspective of individual downstream tasks [15]. Additionally, it faces the challenge of Pareto optimality, which makes it difficult to optimize each subtask individually [24].", + "bbox": [ + 89, + 215, + 480, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adaptive Single-task Bitstream LICs for Multiple Targets. Another line of study for multi-target LIC is the single-task bitstream approach [13, 36, 42, 43, 48, 53]. This approach generates a bitstream specifically for one target task at a time, adapting to the given target task to support multi-task requirements effectively. This allows the approach to achieve optimal performance under well-defined conditions for each specific subtask. Representative studies in this field involve applying task-adaptive importance modulation or fine-tuning techniques to the backbone LIC. Song et al. [53] proposed a spatially-adaptive image compression method using Spatial Feature Transform (SFT) guided by a real-valued quality map that enables pixel-wise optimization for machine vision tasks. Liu et al. [43] introduced a task-adaptive image compression method that uses a gating module to selectively activate channel-wise latents from a pre-trained codec. J.P. Klopp et al. [36] introduced a learned pre-filter that suppresses task-irrelevant data prior to compression by distinguishing between important and unimportant information. TransTIC [13] employed Visual Prompt Tuning (VPT) [21, 33] to adapt the largescale pretrained model TIC [46], originally trained for human perception, to suit various target tasks. This approach demonstrated adaptability to various machine vision tasks such as object detection, instance segmentation, and classification [13]. However, VPT itself faces limitations such as non-monotonic performance changes depending on the size of the prompt applied, as well as increased computational cost due to the additional prompts during the inference stage [12]. More importantly, TransTIC adjusts the prompt to guide the output of a static pretrained model, but the generator responsible for creating this prompt also needs to be pretrained. Also, it requires prompts equivalent to one-fourth of the input feature size of each layer to perform task switching. Therefore, for applying fine-tuning to", + "bbox": [ + 89, + 371, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "individual test samples at the decoder stage, these prompts must be incorporated into the bitstream, making TransTIC unsuitable for instance-specific decoder fine-tuning. Consequently, pre-training is essential to adapt to new tasks, and it is difficult to adapt to untrained open-set tasks.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose using TTFT to dynamically adapt a backbone LIC model for diverse tasks. Since TTFT adjusts part or all of the model through a brief learning phase during testing [16, 27, 39, 54], it effectively supports open set scenarios where the codec needs to be optimized for previously unseen tasks. Although TTFT is highly effective in producing results optimized for individual samples, it has the limitation of requiring additional training steps at each testing instance. Therefore, we utilize LoRA [31], one of the parameter-efficient fine-tuning (PEFT) techniques to reduce training costs while effectively adapting the model to downstream tasks.", + "bbox": [ + 511, + 167, + 906, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 363, + 638, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The LoRA method is one of the most actively researched PEFT techniques as it provides effective task adaptation with only a small number of parameters without increasing model complexity and inference time [6, 31, 57]. It focuses on learning the intrinsic rank of the weights [31]. Let the pre-trained weights be denoted by $W_{o} \\in \\mathbb{R}^{d \\times k}$ , and let the weight change required for task adaptation be denoted by $\\Delta W$ . After the task adaptation, the new weight becomes $W_{o} + \\Delta W$ . Typically, the intrinsic rank of the weight change required for task adaptation is relatively small compared to the size of the entire weight matrix, meaning that $\\Delta W$ can be decomposed into a low-rank representation [31]. From this, the relationship in Eq. (1) can be obtained:", + "bbox": [ + 511, + 388, + 906, + 587 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW = W _ {o} + \\Delta W = W _ {o} + B \\cdot A, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 603, + 903, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B \\in \\mathbb{R}^{d \\times r}$ , $A \\in \\mathbb{R}^{r \\times k}$ , and $r \\ll \\min(d, k)$ . Since $W_{o}$ remains fixed, $B \\cdot A$ can be explicitly separated in $W$ , allowing us to store only the matrices $B$ and $A$ for task adaptation, which requires very minimal memory overhead. During inference, a reparameterization process is performed in which $B \\cdot A$ is added to $W_{o}$ to reconstruct $W$ , thereby avoiding any additional inference latency compared to the original model. The LoRA technique is applicable to both CNN and Transformer architectures and can be applied to all layers as long as the memory overhead issue is manageable. Recently, this technique has been reformulated as an SVD-based form and applied to LLMs [6, 25, 64]. These studies [6, 64] aim to fine-tune large pre-trained language models for downstream tasks in a parameter-efficient manner. Specifically, Zhang et al. [64] utilizes the SVD shaped LoRA to allocate parameter budgets adaptively based on the importance scores of each weight matrix. Meanwhile, Balazy et al. [6] proposed a method that applies SVD", + "bbox": [ + 511, + 627, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4432", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e2488f48a4f9e6a612b8f0a4658ed13e65450cbeacf1bd620f4035653ce46bf6.jpg", + "image_caption": [ + "Figure 2. Original LoRA vs. SVD-LoRA. When the rank is small, SVD-LoRA more effectively captures the intrinsic dimensions of the adaptation weights." + ], + "image_footnote": [], + "bbox": [ + 138, + 89, + 433, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "shaped LoRA to fine-tune only the SV weight matrices during task switching. These novel approaches in the LLM domain show promising potential for effective application in the field of image compression as well.", + "bbox": [ + 89, + 340, + 483, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 415, + 225, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Framework Overview", + "text_level": 1, + "bbox": [ + 89, + 441, + 295, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to adapt the given backbone codec for various downstream tasks, including open-set tasks that are unseen during training. To achieve this, we adapt the backbone LIC to target downstream tasks by applying TTFT [16, 27] on both the encoder and decoder. Specifically, we use LoRA, a widely used approach in test-time scenarios. By tuning the LoRA weights to be optimized for each target task and image instance, the backbone LIC can be adapted to better suit the target task. However, transmitting LoRA weights to the decoder for fully instance-specific tuning incurs a nonnegligible bit rate overhead. To minimize the bit rate issue, we employ a reformulated version of LoRA, referred to as SVD-LoRA, which enables effective fine-tuning of only a small portion of the decoder at test time. While prior works utilized SVD-LoRA for language task-specific adaptation [6, 25, 64], we focus on the test-time decoder fine-tuning for a fully instance-specific optimized compression model. Consequently, the proposed method can flexibly handle diverse, arbitrary target tasks and is compatible with various backbone LIC architectures, including both CNN and Transformer-based models.", + "bbox": [ + 89, + 463, + 483, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. SVD based LoRA", + "text_level": 1, + "bbox": [ + 89, + 791, + 263, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SVD-LoRA is represented by Eq. (2):", + "bbox": [ + 89, + 813, + 343, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW = W _ {o} + \\Delta W = W _ {o} + U \\cdot \\Sigma \\cdot V ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 840, + 482, + 857 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where, $U \\in \\mathbb{R}^{d \\times r}$ , $\\Sigma \\in \\mathbb{R}^{r \\times r}$ , $V \\in \\mathbb{R}^{k \\times r}$ , and $r \\ll \\min(d, k)$ . Here, $U$ and $V$ learn the intrinsic ranks for", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e8bfeb8b3aea92d9b9b0ce4ccbc0dde7f60310fafe0e0fa36bea95282160c32.jpg", + "image_caption": [ + "Figure 3. Bitstream generation process. For instance-specific decoder tuning with manageable bitstream overhead, only the singular values of the decoder side are encoded and added to the image-encoded bitstream. The bit-rate overhead accounts for less than $1\\%$ of the image bitstream size." + ], + "image_footnote": [], + "bbox": [ + 553, + 87, + 877, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "columns and rows, respectively, while $\\Sigma$ is a diagonal matrix that learns the SVs.", + "bbox": [ + 511, + 366, + 903, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the structure includes additional SVs compared to conventional LoRA, the increase in the number of parameters is negligible due to the typically small rank used. Fig. 2 illustrates the process of reconstructing $W$ (adapted weights) through the reparameterization of LoRA and SVD-LoRA, respectively. By learning additional SVs, SVD-LoRA facilitates the learning of more complex intrinsic dimensions of adaptation weights compared to the original LoRA. In cases where both LoRA and SVD-LoRA have the same small rank, experimental results indicate that SVD-LoRA better captures the intrinsic dimensions of the adaptation weights. Details of this observation are discussed in the supplementary material. Fig. 3 illustrates the process of embedding the decoder-side SVD-LoRA weights into the compressed bitstream for instance-specific fin-tuning at the decoder stage. While the bitstream size increases due to the additional SV component, the overhead is negligible. Specifically, with a rank of 4, as specified in our experimental settings, the additional overhead accounts for less than $1\\%$ of the compressed image bitstream size.", + "bbox": [ + 511, + 397, + 906, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Test-Time Fine Tuning with SVD-LoRA", + "text_level": 1, + "bbox": [ + 511, + 710, + 856, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 4 illustrates the TTFT procedure utilizing SVD-LoRA, which consists of two stages. In the first stage, the initial values of SVD-LoRA are set differently depending on the target task type. For closed-set tasks, the initial values are set through training the specific target task. In contrast, open-set tasks leverage initialization learned from common tasks like classification, which allows the model to generalize across various machine vision applications. The next stage involves performing TTFT to optimize each test sample according to the target downstream task. During this process, all of the SVD-LoRA weights in the encoder and", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4433", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9b3def41b67a2cf0616a6c891e58b1bf1518019fbbdfce1eb6656da7b9e8b26.jpg", + "image_caption": [ + "Figure 4. TTFT procedure. The TTFT operates in two stages. First, the initial values for SVD-LoRA are initialized. In the second stage, TTFT is applied at test time to fine-tune each test sample in accordance with the specific downstream task. The source images of (b) is [56]" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "only the SVs in the decoder are fine-tuned. While full finetuning of the encoder does not affect bit rate overhead or decoding latency, it requires a large number of parameters for each TTFT, leading to increased TTFT time overhead. Therefore, in our study, we apply SVD-LoRA to the encoder instead of performing full fine-tuning. As a result, this approach significantly reduces training costs and enables rapid adaptation to new downstream tasks. Additionally, fine-tuning only the SVs on the decoder side further reduces bitstream overhead compared to full LoRA finetuning. However, increasing the number of iterations during TTFT improves performance, but also increases encoding time. This issue can be addressed by adjusting the number of iterations according to the compression scenario, thereby balancing performance improvement with encoding time. (Detailed information is covered in the ablation study)", + "bbox": [ + 88, + 438, + 482, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss function of SVD-LoRA during TTFT is defined by Eq. (3):", + "bbox": [ + 89, + 680, + 483, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = - \\log p (\\hat {z}) - \\log p (\\hat {y} \\mid \\hat {z}) + \\lambda d (x, \\hat {x}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 726, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x$ , $y$ , and $z$ represent the input image, encoded latent feature, and hyper prior, respectively. Additionally, $\\hat{x}$ , $\\hat{y}$ , and $\\hat{z}$ represent their reconstruction. The terms $p(\\hat{z})$ and $p(\\hat{y} \\mid \\hat{z})$ represent the probabilities of the compressed latent variables, which correspond to the bit rate used for compressing the input $x$ . $d(x, \\hat{x})$ represents the distortion between the original and the reconstructed images. Mean square error (MSE) is used as the distortion metric when targeting human perception, as it measures the pixel-wise difference between the original and reconstructed images.", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8e2136279f63654bc5f27db8caa761b5026b633439703f9b376a22fbd20e31f2.jpg", + "image_caption": [ + "Figure 5. An example of applying SVD-LoRA to the TIC codec. While SVD-LoRA is compatible with both CNN and STB layers, our work focuses solely on the STB layers of the main encoder and decoder to enhance task-switching efficiency with minimal tuning parameters." + ], + "image_footnote": [], + "bbox": [ + 519, + 87, + 910, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For machine vision tasks such as image classification, perceptual loss is used as the distortion metric by comparing intermediate features extracted from the original and reconstructed images via the target network [13]. The hyper parameter $\\lambda$ adjusts the balance between rate and distortion. This loss function allows for optimization that balances data compression and reconstruction while tailoring to the objective of each task. All experiments conducted in our work are formulated as rate-distortion optimization, with different values of $\\lambda$ and distortion functions designed according to the characteristics of each downstream task.", + "bbox": [ + 511, + 410, + 906, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 593, + 718, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 619, + 720, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While our method is compatible with any baseline architecture, we demonstrate its effectiveness using a widely adopted state-of-the-art (SOTA) LIC model—the Transformer-based LIC (TIC) network proposed in [46]—as the backbone. Results for the CNN-based LIC model [14] are included in the supplementary material.", + "bbox": [ + 511, + 642, + 906, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 5 shows the application of SVD-LoRA to the TIC model. The TIC adopts an autoencoder structure consisting of a main encoder-decoder pair and a hyper encoder-decoder pair, aiming to optimize the balance between distortion and bitrate for human perception. In our work, to maximize task-switching efficiency with minimal tuning parameters, we apply SVD-LoRA exclusively to the main encoder and decoder, as these components have the most significant impact on task adaptation [13]. While SVD-LoRA is compatible with both CNNs and Transformers, our approach specifically targets the Swin Transformer Block", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4434", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(STB) in the main encoder and decoder to achieve optimal performance.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We used MS-COCO [40] for object detection and instance segmentation, and ImageNet-1K [17] for image classification in our experiments.", + "bbox": [ + 89, + 122, + 482, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-trained Vision Models. To evaluate performance across various downstream tasks, we employed pre-trained vision models [29, 30, 46, 52] as task-specific evaluators. Specifically, we used ResNet50 [29] for image classification, Faster R-CNN [52] for object detection, and Mask R-CNN [30] for instance segmentation.", + "bbox": [ + 89, + 169, + 482, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared Methods. The methods compared in our work are as follows:", + "bbox": [ + 89, + 261, + 482, + 289 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) TIC [46]: TIC is a Transformer-based compression model pre-trained on the Flickr2W dataset [41].", + "(2) Full Fine-Tuning: Full Fine-Tuning updates all parameters of the backbone TIC to adapt to the target machine vision tasks.", + "(3) TransTIC [13]: TransTIC applies VPT to adapt the backbone TIC—originally trained for human perception—to target machine vision tasks.", + "(4) Unified Image Compression (UIC) [24]: UIC uses a collaborative embedding with a diffusion model to generate a universal bitstream to handle open-set tasks.", + "(5) Other Adaptation Approaches: These approaches apply adaptation techniques to the backbone TIC: (5-1) TIC+SFT [61] integrates Spatial Feature Transform (SFT) layers into TIC, where the SFT layers act as adapters for task-specific modulation; (5-2) TIC+channel_selection [43] selects partial channels from the image latent representation produced by TIC and converts them into task-specific feature maps to enable efficient task transfer; (5-3) TIC+prefilter [36] involves adding a pre-neural network before TIC to assist in transitioning TIC to the target task." + ], + "bbox": [ + 91, + 292, + 482, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task Types and TTFT settings. Experiments were conducted under both closed-set and open-set conditions. We used a classification model to extract generalized features for open-set initialization. Therefore, the open-set image classification task is excluded, since its initialization is derived from its own target model. In all experiments, 'w/o TTFT' refers to the task-specific approach using fixed SVD-LoRA weights, while 'w/ TTFT' denotes the instance-specific approach with TTFT applied per test sample. TTFT is performed for 40 iterations per sample in our experiments; however, the number of iterations is adjustable and its impact on performance is analyzed in the ablation study.", + "bbox": [ + 89, + 626, + 482, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Top-1 accuracy is used to evaluate image classification performance, whereas mean average precision (mAP) is employed for both object detection and instance segmentation. For numerical comparisons, the BDA-rate metric [23] is used.", + "bbox": [ + 89, + 808, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Strategy for handling different quality levels. In all ex", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "periments, the four quality levels of TIC (from quality level 1 to 4) are used to visually represent the distribution between the rate and evaluation metrics. The $\\lambda$ values of loss function are set to 2, 1, 0.5, and 0.2, depending on the quality level [13].", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation for Multiple Machine Vision Tasks and Human perception", + "text_level": 1, + "bbox": [ + 511, + 176, + 903, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 6 compares our proposed approach with other methods across various machine vision tasks. The results demonstrate that our method, with TTFT applied, outperforms all competing approaches. Table 1 presents a performance comparison using BD-Rate, BD-mAP, and BD-accuracy with TIC as the anchor. The results show that our method consistently outperforms across all tasks, not only under closed-set conditions but also in challenging open-set settings. While the task-specific approach (denoted as 'w/o TTFT') achieves performance comparable to other competing methods, applying TTFT further improves performance. Even under challenging open-set settings, TTFT demonstrates that the model remains robust. Notably, our method achieves higher performance under open-set conditions than TransTIC, one of the SOTA techniques, does under closed-set conditions, demonstrating its superior generalization capability.", + "bbox": [ + 511, + 214, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 481, + 694, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 7 visualizes the results of object detection. Although TIC yields high visual quality from the perspective of human perception, its detection confidence scores are relatively low, limiting its effectiveness for downstream tasks. Furthermore, in all experiments, the visualized results exhibit noticeable differences between closed-set and open-set conditions. Since closed-set tasks are pre-trained on the target object detection task, high prediction scores can be achieved regardless of whether TTFT is applied. The corresponding results also tend to preserve fine texture details. In contrast, the open-set setting relies on generalized features that are not optimized for the target task, leading to reduced prediction accuracy and blurred textures. Nonetheless, in both settings, TTFT improves performance by enhancing texture detail preservation and enabling higher prediction scores. Notably, TTFT also maintains model robustness under challenging open-set conditions.", + "bbox": [ + 511, + 503, + 906, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 771, + 678, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ablation studies are conducted under the conditions of the open-set object detection task, utilizing an A100 GPU with 80GB memory. The evaluation is performed using 5,000 validation images from the MS-COCO dataset.", + "bbox": [ + 511, + 794, + 903, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adjustment of fine-tuning iterations. Table 2 shows the performance and time overhead with varying numbers of TTFT iterations. As the number of iterations increases,", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4435", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6103282c531a3891da19f3e7adb8371bfbd82b136ad05277587f0ef28a036bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 82, + 398, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/034b463044f6315d7f515c348e2dbbdd8ca16b8edafb49b8ff41e080bde5156d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 82, + 586, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31273c596a8010511c737040a12136c90c60753b11bd8f00a777f0884365d93b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 98, + 754, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0ee42624a85c88be03f936b97c02b92a11f505d1672709e703e118770e71361.jpg", + "image_caption": [ + "(a) Instance Segmentation (open-set)", + "(c) Instance Segmentation (closed-set)", + "Figure 6. Rate-accuracy performance comparison under various machine tasks. Experimental results show that our method outperforms the SOTA approaches in various closed-set and open-set target tasks." + ], + "image_footnote": [], + "bbox": [ + 223, + 253, + 401, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d41a058e543a4977b630d1ca3387de4685819d95b66e9c993a2af899661f31d.jpg", + "image_caption": [ + "(b) Object Detection (open-set)", + "(d) Object Detection (closed-set)" + ], + "image_footnote": [], + "bbox": [ + 406, + 253, + 584, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/008c7b8c76da68d8257fd1d2802947560c32ffdc025f8f73c88254baaee48181.jpg", + "image_caption": [ + "(e) Classification (closed-set)" + ], + "image_footnote": [], + "bbox": [ + 591, + 253, + 771, + 396 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/44f78db64dc3d0db34773ca2f2aefd4f38055c31d6dcdb0a5bff5d0e1eb8a1f0.jpg", + "table_caption": [ + "Table 1. Comparison of BD-Rate, BD-mAP, and BD-Accuracy using TIC as the anchor. Bold numbers indicate the best performance, while blue numbers represent the second-best performance." + ], + "table_footnote": [], + "table_body": "
Task TypeMethodInstance SegmentationObject DetectionClassification
BD-Rate (%) ↓BD-mAP ↑BD-Rate (%) ↓BD-mAP ↑BD-Rate (%) ↓BD-accuracy ↑
Closed-setFull fine tuning-75.554.00-66.004.10-1000.16
TIC+prefilter-24.741.56-17.591.19-23.690.04
TIC+SFT-40.241.67-27.241.45-61.530.11
TIC+channel_selection47.58-2.5552.85-3.77-30.370.06
TransTIC-47.452.34-46.642.71-58.760.10
UIC87.99-3.67102.70-4.31--
Ours (w/o TTFT)-43.441.99-38.622.17-50.670.08
Ours (w/ TTFT)-56.492.68-54.962.98-70.040.14
Open-setTransTIC-43.552.28-39.602.23--
UIC118.82-5.28146.85-6.37--
Ours (w/o TTFT)-32.481.97-29.611.87--
Ours (w/ TTFT)-49.512.43-48.942.69--
", + "bbox": [ + 171, + 520, + 818, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both performance and time overhead also increase. While time overhead increases linearly, performance gains are non-linear, with the majority of improvements occurring in the early iterations. Based on these characteristics of TTFT, our method can be effectively adapted to various compression scenarios. In low-latency applications such as CCTV surveillance systems, where real-time compression and transmission within milliseconds is required, the task-specific approach (denoted as 'w/o TTFT') is preferable. In contrast, scenarios such as satellite imagery or medical imaging—where image updates occur over the course of minutes to days—are more tolerant of encoding latency. In such cases, the instance-specific approach with hundreds", + "bbox": [ + 88, + 703, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c36917b8e9454e9448939e77e30e7e96556e232513b9af93044bc66853d4ae67.jpg", + "table_caption": [ + "Table 2. Ablation on the range of TTFT iterations. Experiments show that early iterations have the most impact, and TTFT iterations control time overhead and performance." + ], + "table_footnote": [], + "table_body": "
Bit-ratemAPΔmAP/iter.TTFT Time (sec)
w/o TTFT0.0959134.021--
w/ TTFT5 iters.0.10635.280.251.18
10 iters.0.10635.840.184.28
15 iters.0.10735.880.125.86
40 iters.0.10636.250.0614.13
", + "bbox": [ + 532, + 753, + 882, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "or more TTFT iterations may be deployed without critical latency concerns.", + "bbox": [ + 511, + 837, + 906, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Varying the range of fine-tuning. Fig. 8 shows the object detection performance with varying fine-tuning ranges. The", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4436", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f59b09b2587a272b110faecbaedc4c4ee9e2f25e58dbfccb8ff5a3a3c962b92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 90, + 339, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f8858961fbe465752013cb7664bdc7b8696edd22be2dc2f0273a2fde9b604c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 90, + 496, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2b8961c35e987f8809e3c30059d09b4eb689f420fc34bcd3687e38c130fb48d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 656, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c465f005ffb6ae1d4e0cea0a64420a841fa942ede08876ab4458009b4a15d55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 90, + 813, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b773737507a0728aab520f580891b8cb17367df7d340082020b2c7534cb3ffe.jpg", + "image_caption": [ + "Ours (open-set, w/o TTFT)" + ], + "image_footnote": [], + "bbox": [ + 183, + 261, + 338, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/891701f907dca521cd180e4579b7994d49151102e0e61f8444b8430f73b49857.jpg", + "image_caption": [ + "Ours (open-set, w/ TTFT)" + ], + "image_footnote": [], + "bbox": [ + 341, + 261, + 496, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1a945b5e0abb6bc2ef3a6d20d30bfefb2bc044a40660143528ab232080087a8.jpg", + "image_caption": [ + "Ours (closed-set, w/o TTFT)", + "Ours (closed-set, w/ TTFT)" + ], + "image_footnote": [], + "bbox": [ + 500, + 261, + 656, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/590fc1ebc9d6a940db25ee795122e5881b9188b27cdd95ade245c374d5901545.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 261, + 813, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/babeb2e1ccfc0505bc4cc3ae2e81598bd06e34a6895e8c00669eb7f44e07de6c.jpg", + "image_caption": [ + "Figure 7. Visualization of object detection results. While TIC provides better visual quality, it exhibits low detection confidence, leading to misdetections—such as misclassifying a bear as a sheep. Ours (w/ TTFT) surpasses the uncompressed image in prediction confidence despite compression losses.", + "Figure 8. Ablation on the range of fine-tuning. Experiments show that fine-tuning the decoder is essential for optimal performance." + ], + "image_footnote": [], + "bbox": [ + 140, + 449, + 424, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comparison includes fine-tuning either the encoder or the decoder, or both together. The experimental results show that TTFT applied to the decoder stage contributes to performance improvement. While the encoder stage, which involves a larger number of trainable parameters during TTFT, yields greater performance gains, the decoder-stage TTFT remains essential for achieving optimal performance.", + "bbox": [ + 89, + 686, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 9 presents a comparison between decoder finetuning using only the SVs and using the full set of SVD-LoRA parameters. The results show that training all SVD-LoRA parameters increases bitstream overhead by $15.3\\mathrm{x}$ . Therefore, restricting training to only the SVs in the decoder is critical for enabling fully instance-specific finetuning with minimal overhead.", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9f7bf7112fae62d831c2af251fce2897ed3e612199c4d9db4e1868ace8b9547.jpg", + "image_caption": [ + "Figure 9. Ablation on the range of decoder fine-tuning. Experiments show that training only the singular values of the decoder effectively reduces excessive bitstream overhead." + ], + "image_footnote": [], + "bbox": [ + 563, + 449, + 844, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 699, + 633, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a fully instance-specific TTFT approach for adapting learned image compression to both closed-set and open-set machine vision tasks. Our method can overcome the challenging problem of instance-specific decoder finetuning, which remains difficult due to its significant impact on both bitstream size and decoding time but is essential for effectively handling open-set tasks. The SVD-LoRA we applied enables TTFT while minimizing bitstream overhead. Moreover, it can be seamlessly integrated into any architecture, including CNNs and Transformers. Therefore, our approach can be applied to various domains where instance-specific TTFT is required.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4437", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) under multiple grants funded by the Korea government (MSIT), including the ICT Creative Consilience Program (IITP-2025-RS-2020-II201821), the Information Technology Research Center (ITRC) support program (IITP-2021-0-02052), and the Artificial Intelligence Innovation Hub (RS-2021-II212068). Additionally, this research was supported by Korea Research Institute for defense Technology planning and advancement (KRIT) grant funded by the Korean Government, Defense Acquisition Program Administration (DAPA), under Grant KRIT-CT-23-021, Development of Multi-Modal Information Fusion and High-Realization Battlefield Data Generation Technology, in 2025.", + "bbox": [ + 89, + 114, + 485, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 354, + 187, + 369 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Umar Albalawi, Saraju P Mohanty, and Elias Kougianos. A hardware architecture for better portable graphics (bpg) compression encoder. In 2015 IEEE international Symposium on Nanoelectronic and information systems, pages 291-296. IEEE, 2015. 2", + "[2] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 3", + "[3] Dana Harry Ballard and Christopher M Brown. Computer vision. Prentice Hall Professional Technical Reference, 1982. 1", + "[4] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In International Conference on Learning Representations, 2017. 2", + "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 2", + "[6] Klaudia Bałazy et al. Lora-xs: Low-rank adaptation with extremely small number of parameters. arXiv preprint, 2024. 3, 4", + "[7] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 2", + "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 4", + "[9] Prabuddha Chakraborty, Jonathan Cruz, and Swarup Bhunia. Magic: Machine-learning-guided image compression for vision applications in internet of things. IEEE Internet of Things Journal, 8(9):7303-7315, 2020. 1" + ], + "bbox": [ + 99, + 378, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Lahiru D Chamain, Fabien Racapé, Jean Bégaint, Akshay Pushparaja, and Simon Feltman. End-to-end optimized image compression for machines, a study. In 2021 Data Compression Conference (DCC), pages 163-172. IEEE, 2021. 1", + "[11] Lahiru D Chamain, Fabien Racapé, Jean Bégaint, Akshay Pushparaja, and Simon Feltman. End-to-end optimized image compression for multiple machine tasks. arXiv preprint arXiv:2103.04178, 2021. 2, 3", + "[12] Arnav Chavan et al. One-for-all: Generalized lora for parameter-efficient fine-tuning. arXiv preprint arXiv:2306.07967, 2023. 3", + "[13] Yi-Hsin Chen, Ying-Chieh Weng, Chia-Hao Kao, Cheng Chien, Wei-Chen Chiu, and Wen-Hsiao Peng. Transtic: Transferring transformer-based image compression from human perception to machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23297-23307, 2023. 1, 2, 3, 5, 6", + "[14] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 5, 1, 2", + "[15] Hyomin Choi and Ivan V Bajic. Scalable image coding for humans and machines. IEEE Transactions on Image Processing, 31:2739-2754, 2022. 2, 3", + "[16] Jack Cole and Mohamed Osman. Dataset-induced meta-learning (and other tricks): Improving model efficiency on arc, 2023. 3, 4", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[18] Lingyu Duan, Jiaying Liu, Wenhan Yang, Tiejun Huang, and Wen Gao. Video coding for machines: A paradigm of collaborative compression and intelligent analytics. IEEE Transactions on Image Processing, 29:8680-8695, 2020. 1", + "[19] Zhihao Duan, Zhan Ma, and Fengqing Zhu. Unified architecture adaptation for compressed domain semantic inference. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):4108-4121, 2023. 2, 3", + "[20] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 2, 3", + "[21] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3", + "[22] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 3", + "[23] Bjontegaard Gisle. Calculation of average psnr differences between rd curves. In ITU-T SG16/Q6, $13^{\\circ}$ th VCEG Meeting, Austin, Texas, USA, April 2001, 2001. 2, 6", + "[24] Sha Guo, Lin Sui, Chenlin Zhang, Zhuo Chen, Wenhan Yang, and Lingyu Duan. A unified image compression" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4438", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "method for human perception and multiple vision tasks. In European Conference on Computer Vision, pages 342-359. Springer, 2024. 2, 3, 6", + "[25] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3, 4", + "[26] Allen Hanson. Computer vision systems. Elsevier, 1978. 1", + "[27] Moritz Hardt and Yu Sun. Test-time training on nearest neighbors for large language models. In The Twelfth International Conference on Learning Representations, 2024. 3, 4", + "[28] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[30] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6", + "[31] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In Proceedings of the ICLR, 2022. 3", + "[32] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 2, 3", + "[33] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3", + "[34] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Realtime multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 3, 4", + "[35] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3, 4", + "[36] Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien. How to exploit the transferability of learned image compression to conventional codesc. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16165-16174, 2021. 2, 3, 6", + "[37] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: an end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Huanyang Li and Xinfeng Zhang. Human-machine collaborative image compression method based on implicit neural representations. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2024. 1", + "[39] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, 133(1):31-64, 2025. 3", + "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[41] Jiaheng Liu, Guo Lu, Zhihao Hu, and Dong Xu. A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370, 2020. 1, 6", + "[42] Jinming Liu, Heming Sun, and Jiro Katto. Learning in compressed domain for faster machine vision tasks. In 2021 International Conference on Visual Communications and Image Processing (VCIP), pages 01-05. IEEE, 2021. 2, 3", + "[43] Jinming Liu, Heming Sun, and Jiro Katto. Improving multiple machine vision tasks in the compressed domain. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 331-337. IEEE, 2022. 2, 3, 6", + "[44] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14388-14397, 2023. 2", + "[45] Kang Liu, Dong Liu, Li Li, Ning Yan, and Houqiang Li. Semantics-to-signal scalable image compression with learned revertible representations. International Journal of Computer Vision, 129(9):2605-2621, 2021. 2, 3", + "[46] Ming Lu, Peiyao Guo, Huiqing Shi, Chuntong Cao, and Zhan Ma. Transformer-based image compression. In 2022 Data Compression Conference (DCC), pages 469-469. IEEE, 2022. 2, 3, 5, 6", + "[47] Siwei Ma, Xinfeng Zhang, Chuanmin Jia, Zhenghui Zhao, Shiqi Wang, and Shanshe Wang. Image and video compression with neural networks: A review. IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683-1698, 2019. 2", + "[48] Yixin Mei, Fan Li, Li Li, and Zhu Li. Learn a compression for objection detection-vae with a bridge. In 2021 International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2021. 2, 3", + "[49] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2", + "[50] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2", + "[51] Vishvjit S Nalwa. A guided tour of computer vision. Addison-Wesley Longman Publishing Co., Inc., 1994. 1", + "[52] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4439", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 6", + "[53] Myungseo Song, Jinyoung Choi, and Bohyung Han. Variable-rate deep image compression through spatially-adaptive feature transform. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2380-2389, 2021. 1, 2, 3", + "[54] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 3", + "[55] Vivienne Sze, Madhukar Budagavi, and Gary J Sullivan. High efficiency video coding (hevc). In Integrated circuit and systems, algorithms and architectures, page 40. Springer, 2014. 2", + "[56] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 1", + "[57] Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. Dylora: Parameter-efficient tuning of pretrained models using dynamic search-free low-rank adaptation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3274-3287, 2023. 3", + "[58] Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, and Eftychios Protopapadakis. Deep learning for computer vision: A brief review. Computational intelligence and neuroscience, 2018(1):7068349, 2018. 1", + "[59] Shurun Wang, Zhao Wang, Shiqi Wang, and Yan Ye. End-to-end compression towards machine vision: Network architecture design and optimization. IEEE Open Journal of Circuits and Systems, 2:675-685, 2021. 1", + "[60] Shurun Wang, Zhao Wang, Shiqi Wang, and Yan Ye. Deep image compression toward machine vision: A unified optimization framework. IEEE Transactions on Circuits and Systems for Video Technology, 33(6):2979-2989, 2022. 1", + "[61] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6", + "[62] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. Sssic: semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 2, 3", + "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3, 4", + "[64] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fin-tuning. In International Conference on Learning Representations, 2023. 3, 4", + "[65] Y Zhang, M Rafie, and S Liu. Use cases and requirements for video coding for machines. ISO/IEC JTC, 1, 2021. 1" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[66] Xia Zhao, Limin Wang, Yufei Zhang, Xuming Han, Muhammet Deveci, and Milan Parmar. A review of convolutional neural networks in computer vision. Artificial Intelligence Review, 57(4):99, 2024. 1", + "bbox": [ + 516, + 92, + 905, + 147 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4440", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_model.json b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b635ed511f2eb1658d2ee372b2069a9ffad382c1 --- /dev/null +++ b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.229, + 0.131, + 0.773, + 0.177 + ], + "angle": 0, + "content": "Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.203, + 0.893, + 0.258 + ], + "angle": 0, + "content": "Unki Park\\(^{1,2}\\) Seongmoon Jeong\\(^{1}\\) Youngchan Jang\\(^{3}\\) Gyeong-Moon Park\\(^{4*}\\) Jong Hwan Ko\\(^{1*}\\) \n\\(^{1}\\)Sungkyunkwan University, \\(^{2}\\)Samsung Institute of Technology, \\(^{3}\\)Hanwha Systems Corporation, \\(^{4}\\)Korea University \n{unkipark@g., giftjsm@, hjko@}skku.edu, youngchan84.jang@hanwha.com, gm-park@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.485, + 0.718 + ], + "angle": 0, + "content": "The field of computer vision was initially inspired by the human visual system and has progressively expanded to include a broader range of machine vision applications. Consequently, image compressors should be designed to effectively accommodate not only human visual perception but also machine vision tasks, including closed-set scenarios that enable pre-training and open-set scenarios that involve previously unseen tasks at test time. Many recent studies effectively address both human visual perception and closed-set machine vision tasks simultaneously but struggle to handle open-set machine vision tasks. To address this issue, this paper proposes a fully instance-specific test time fine-tuning (TTFT) for adapting learned image compression (LIC) to both closed-set and open-set machine vision tasks effectively. With our method, a large-scale LIC model, originally trained for human perception, is adapted to the target task through TTFT using Singular Value Decomposition based Low Rank Adaptation (SVD-LoRA). During TTFT, the decoder adopts a modified learning scheme that focuses exclusively on training the singular values, which helps prevent excessive bitstream overhead. This enables fully instance-specific optimization for the target task, even for open-set tasks. Experimental results demonstrate that the proposed method effectively adapts the backbone compressor to diverse machine vision tasks, outperforming competing methods. The code is available at project page." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.746, + 0.222, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.879 + ], + "angle": 0, + "content": "In the early stages of computer vision, image processing technologies were primarily developed to align with human perception, as images were largely intended for human interpretation [3, 26, 51]. However, with advances in deep learning and computer vision techniques enabling more sophisticated data processing, machines have increasingly gained the ability to interpret visual data. Conse" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.288, + 0.9, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.563 + ], + "angle": 0, + "content": "Figure 1. An overview of our proposed method. Our method tunes both the encoder and decoder of a backbone human-target LIC model by fine-tuning SVD-LoRA weights optimized for each target task and image instance." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.594, + 0.907, + 0.685 + ], + "angle": 0, + "content": "sequently, the scope of computer vision has broadened to include machine vision applications [56, 58, 66]. In the field of image compression as well, there is a growing need for methods tailored to the requirements of machine vision, as machines are expected to become the primary users of compressed image data [18, 65]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Accordingly, several research studies have shifted their focus away from the traditional goal of visual signal reconstruction, instead prioritizing the enhancement of machine vision task performance [9, 59]. These studies commonly aim to effectively restore representations suitable for the targeted machine vision domain, while selectively sacrificing visual quality to achieve this goal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Taking it further, recent research has focused on developing learned image codes (LICs) that can cater to both human perception and various machine vision tasks [10, 13, 37, 38, 41, 53, 60]. These studies can be categorized into methods that generate either multi-task bitstreams or single-task bitstreams. The multi-task bitstream approach aims to generate a unified bitstream that can support vari-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.242, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4430" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "ous subtasks including human perception and multiple machine tasks. This method offers flexibility in responding to a range of machine vision tasks, but it introduces challenges. For example, scalable coding [15, 32, 45, 62] increases the processing complexity and bitstream overhead for individual downstream tasks, and unified image coding [11, 19, 20, 24] struggles with optimizing each subtask individually." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.483, + 0.445 + ], + "angle": 0, + "content": "In contrast, the single-task bitstream approach [13, 36, 42, 43, 48, 53] avoids the limitations of multi-task methods by dynamically generating bitstreams for a specific downstream task. For instance, TransTIC[13] utilizes prompt tuning for adapting the human-target backbone codec to generate bitstreams optimized for the other tasks specified by the prompt. While this technique enables switching of pre-defined (closed-set) tasks without modifying the backbone codec, it requires task-specific pre-training of the prompt generator. Therefore, it leads to increased training demands as the number of tasks grows, and faces challenges to handle new, untrained (open-set) tasks. Moreover, performance improvement is non-monotonic with respect to prompt size, and computational costs increase due to the added prompt during inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address these limitations, we aim to develop a method that can dynamically adapt a given backbone codec to optimize rate-performance trade-offs for the specific requirements of each task. Our goal is to provide flexible support for diverse tasks and codec models, allowing adaptation to previously unseen (open-set) tasks without requiring pre-training, and also adaptable to any given codec model. To achieve this, as shown in Fig. 1, we propose tuning the backbone codec using Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning approach widely used in test-time scenarios. Unlike the TransTIC approach, which cannot perform instance-specific tuning on the decoder, our method tunes both the encoder and decoder by generating LoRA weights optimized for each target task and image instance. This enables dynamic codec optimization for any arbitrary task without requiring pre-training, making it effective for open-set scenarios. While instance-specific model optimization can lead to superior performance compared to task-specific optimization, it introduces additional overhead by requiring a separate set of parameters for each image. Specifically, in the case of instance-specific decoder tuning, transmitting the LoRA weights incurs additional bit rate overhead. To address this efficiently, we adopt an SVD-LoRA approach that combines the original LoRA weights with supplementary singular value (SV) components. After pre-initializing the original LoRA weights, we can then transmit only the SV components to the decoder on an instance-specific basis, allowing effective tuning while minimizing bit rate demands. As the SV component captures essential information for optimizing the codec per target in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "stance and task, it is both effective and efficient in managing rate requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.317 + ], + "angle": 0, + "content": "Experimental results demonstrate that our method achieves superior adaptation performance for both open-set and closed-set tasks. Although conducted under a lightweight experimental setup (40 iterations, LoRA rank 4), we observed that our method achieved more than a \\(50\\%\\) improvement in Bjøntegaard Delta (BD)-Rate [23] compared to the backbone codec across all evaluated machine vision tasks. These results surpass all competing methods. Notably, our method achieves superior performance in a pure open-set setting—where no prior task-specific setup is available—outperforming competing methods even when they are evaluated under pre-initialized (closed-set) conditions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.905, + 0.347 + ], + "angle": 0, + "content": "The main contributions of our work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.349, + 0.905, + 0.439 + ], + "angle": 0, + "content": "- Test-time fine-tuning in compression field : To the best of our knowledge, we are the first to attempt TTFT at the decoder stage in a image compression task. Considering that the encoder and decoder are temporally and spatially separated in compression systems, TTFT has the potential to maximize image compression quality." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.44, + 0.905, + 0.559 + ], + "angle": 0, + "content": "- Novel approach utilizing LoRA: We propose a novel approach utilizing LoRA for efficient adaptation to diverse machine vision applications in the still-image compression task. Specifically, we introduce a modified LoRA in the form of SVD to enable instance-specific optimization at the decoder. This approach not only enhances efficiency but also facilitates effective TTFT at the decoder stage." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.56, + 0.905, + 0.635 + ], + "angle": 0, + "content": "- Handling open-set tasks: By designing an instance-specific optimized model that does not necessarily require pre-training, we have enhanced the ability to tackle open-set machine vision tasks, which have remained challenging in previous research." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.349, + 0.905, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.657, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.903 + ], + "angle": 0, + "content": "The initial development of LIC focused on optimizing perceptual quality for human visual systems, similar to traditional compressors [1, 7, 55]. Most LIC models adopt an autoencoder structure [4, 5, 44, 47, 50], aiming to optimize the balance between distortion and bit-rate. In early LIC research, the goal was to minimize losses limited to human visual perception, with the technology evolving towards faithfully mimicking human visual quality [28, 46, 49]. With the rapid development of machine vision tasks such as image classification and object detection, an increasing number of applications now incorporate images encoded and decoded by image codecs for these downstream tasks. Consequently, LIC research has begun to focus on meeting machine vision requirements. Initially, the emphasis was on creating" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4431" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "individual models tailored to each specific task. More recently, research has proposed approaches where a shared LIC model architecture is individually trained to generate task-specific models. Following this trend, research efforts aimed at designing a single LIC model to support multiple targets including human perception and diverse machine vision tasks. These studies can be divided into multi-task bitstream and single-task bitstream methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.482, + 0.368 + ], + "angle": 0, + "content": "Multi-task Bitstream LICs for Multiple Targets. One line of research in multi-target LIC is the multi-task bitstream approach [11, 15, 19, 20, 24, 32, 45, 62], which aims to accommodate multiple targets using a single bitstream. However, this approach increases the complexity of both the encoder and decoder [15, 24]. It also leads to significant bitstream overhead from the perspective of individual downstream tasks [15]. Additionally, it faces the challenge of Pareto optimality, which makes it difficult to optimize each subtask individually [24]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.485, + 0.904 + ], + "angle": 0, + "content": "Adaptive Single-task Bitstream LICs for Multiple Targets. Another line of study for multi-target LIC is the single-task bitstream approach [13, 36, 42, 43, 48, 53]. This approach generates a bitstream specifically for one target task at a time, adapting to the given target task to support multi-task requirements effectively. This allows the approach to achieve optimal performance under well-defined conditions for each specific subtask. Representative studies in this field involve applying task-adaptive importance modulation or fine-tuning techniques to the backbone LIC. Song et al. [53] proposed a spatially-adaptive image compression method using Spatial Feature Transform (SFT) guided by a real-valued quality map that enables pixel-wise optimization for machine vision tasks. Liu et al. [43] introduced a task-adaptive image compression method that uses a gating module to selectively activate channel-wise latents from a pre-trained codec. J.P. Klopp et al. [36] introduced a learned pre-filter that suppresses task-irrelevant data prior to compression by distinguishing between important and unimportant information. TransTIC [13] employed Visual Prompt Tuning (VPT) [21, 33] to adapt the largescale pretrained model TIC [46], originally trained for human perception, to suit various target tasks. This approach demonstrated adaptability to various machine vision tasks such as object detection, instance segmentation, and classification [13]. However, VPT itself faces limitations such as non-monotonic performance changes depending on the size of the prompt applied, as well as increased computational cost due to the additional prompts during the inference stage [12]. More importantly, TransTIC adjusts the prompt to guide the output of a static pretrained model, but the generator responsible for creating this prompt also needs to be pretrained. Also, it requires prompts equivalent to one-fourth of the input feature size of each layer to perform task switching. Therefore, for applying fine-tuning to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "individual test samples at the decoder stage, these prompts must be incorporated into the bitstream, making TransTIC unsuitable for instance-specific decoder fine-tuning. Consequently, pre-training is essential to adapt to new tasks, and it is difficult to adapt to untrained open-set tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.349 + ], + "angle": 0, + "content": "To overcome these limitations, we propose using TTFT to dynamically adapt a backbone LIC model for diverse tasks. Since TTFT adjusts part or all of the model through a brief learning phase during testing [16, 27, 39, 54], it effectively supports open set scenarios where the codec needs to be optimized for previously unseen tasks. Although TTFT is highly effective in producing results optimized for individual samples, it has the limitation of requiring additional training steps at each testing instance. Therefore, we utilize LoRA [31], one of the parameter-efficient fine-tuning (PEFT) techniques to reduce training costs while effectively adapting the model to downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.364, + 0.64, + 0.382 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.908, + 0.588 + ], + "angle": 0, + "content": "The LoRA method is one of the most actively researched PEFT techniques as it provides effective task adaptation with only a small number of parameters without increasing model complexity and inference time [6, 31, 57]. It focuses on learning the intrinsic rank of the weights [31]. Let the pre-trained weights be denoted by \\( W_{o} \\in \\mathbb{R}^{d \\times k} \\), and let the weight change required for task adaptation be denoted by \\( \\Delta W \\). After the task adaptation, the new weight becomes \\( W_{o} + \\Delta W \\). Typically, the intrinsic rank of the weight change required for task adaptation is relatively small compared to the size of the entire weight matrix, meaning that \\( \\Delta W \\) can be decomposed into a low-rank representation [31]. From this, the relationship in Eq. (1) can be obtained:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.604, + 0.905, + 0.621 + ], + "angle": 0, + "content": "\\[\nW = W _ {o} + \\Delta W = W _ {o} + B \\cdot A, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(B \\in \\mathbb{R}^{d \\times r}\\), \\(A \\in \\mathbb{R}^{r \\times k}\\), and \\(r \\ll \\min(d, k)\\). Since \\(W_{o}\\) remains fixed, \\(B \\cdot A\\) can be explicitly separated in \\(W\\), allowing us to store only the matrices \\(B\\) and \\(A\\) for task adaptation, which requires very minimal memory overhead. During inference, a reparameterization process is performed in which \\(B \\cdot A\\) is added to \\(W_{o}\\) to reconstruct \\(W\\), thereby avoiding any additional inference latency compared to the original model. The LoRA technique is applicable to both CNN and Transformer architectures and can be applied to all layers as long as the memory overhead issue is manageable. Recently, this technique has been reformulated as an SVD-based form and applied to LLMs [6, 25, 64]. These studies [6, 64] aim to fine-tune large pre-trained language models for downstream tasks in a parameter-efficient manner. Specifically, Zhang et al. [64] utilizes the SVD shaped LoRA to allocate parameter budgets adaptively based on the importance scores of each weight matrix. Meanwhile, Balazy et al. [6] proposed a method that applies SVD" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4432" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.434, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.272, + 0.486, + 0.315 + ], + "angle": 0, + "content": "Figure 2. Original LoRA vs. SVD-LoRA. When the rank is small, SVD-LoRA more effectively captures the intrinsic dimensions of the adaptation weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.484, + 0.403 + ], + "angle": 0, + "content": "shaped LoRA to fine-tune only the SV weight matrices during task switching. These novel approaches in the LLM domain show promising potential for effective application in the field of image compression as well." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.416, + 0.226, + 0.433 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.442, + 0.296, + 0.456 + ], + "angle": 0, + "content": "4.1. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.485, + 0.781 + ], + "angle": 0, + "content": "Our goal is to adapt the given backbone codec for various downstream tasks, including open-set tasks that are unseen during training. To achieve this, we adapt the backbone LIC to target downstream tasks by applying TTFT [16, 27] on both the encoder and decoder. Specifically, we use LoRA, a widely used approach in test-time scenarios. By tuning the LoRA weights to be optimized for each target task and image instance, the backbone LIC can be adapted to better suit the target task. However, transmitting LoRA weights to the decoder for fully instance-specific tuning incurs a nonnegligible bit rate overhead. To minimize the bit rate issue, we employ a reformulated version of LoRA, referred to as SVD-LoRA, which enables effective fine-tuning of only a small portion of the decoder at test time. While prior works utilized SVD-LoRA for language task-specific adaptation [6, 25, 64], we focus on the test-time decoder fine-tuning for a fully instance-specific optimized compression model. Consequently, the proposed method can flexibly handle diverse, arbitrary target tasks and is compatible with various backbone LIC architectures, including both CNN and Transformer-based models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.264, + 0.806 + ], + "angle": 0, + "content": "4.2. SVD based LoRA" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.814, + 0.344, + 0.83 + ], + "angle": 0, + "content": "SVD-LoRA is represented by Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.841, + 0.483, + 0.858 + ], + "angle": 0, + "content": "\\[\nW = W _ {o} + \\Delta W = W _ {o} + U \\cdot \\Sigma \\cdot V ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where, \\(U \\in \\mathbb{R}^{d \\times r}\\), \\(\\Sigma \\in \\mathbb{R}^{r \\times r}\\), \\(V \\in \\mathbb{R}^{k \\times r}\\), and \\(r \\ll \\min(d, k)\\). Here, \\(U\\) and \\(V\\) learn the intrinsic ranks for" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.088, + 0.878, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.269, + 0.906, + 0.34 + ], + "angle": 0, + "content": "Figure 3. Bitstream generation process. For instance-specific decoder tuning with manageable bitstream overhead, only the singular values of the decoder side are encoded and added to the image-encoded bitstream. The bit-rate overhead accounts for less than \\(1\\%\\) of the image bitstream size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.905, + 0.396 + ], + "angle": 0, + "content": "columns and rows, respectively, while \\(\\Sigma\\) is a diagonal matrix that learns the SVs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Although the structure includes additional SVs compared to conventional LoRA, the increase in the number of parameters is negligible due to the typically small rank used. Fig. 2 illustrates the process of reconstructing \\( W \\) (adapted weights) through the reparameterization of LoRA and SVD-LoRA, respectively. By learning additional SVs, SVD-LoRA facilitates the learning of more complex intrinsic dimensions of adaptation weights compared to the original LoRA. In cases where both LoRA and SVD-LoRA have the same small rank, experimental results indicate that SVD-LoRA better captures the intrinsic dimensions of the adaptation weights. Details of this observation are discussed in the supplementary material. Fig. 3 illustrates the process of embedding the decoder-side SVD-LoRA weights into the compressed bitstream for instance-specific fin-tuning at the decoder stage. While the bitstream size increases due to the additional SV component, the overhead is negligible. Specifically, with a rank of 4, as specified in our experimental settings, the additional overhead accounts for less than \\( 1\\% \\) of the compressed image bitstream size." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.857, + 0.728 + ], + "angle": 0, + "content": "4.3. Test-Time Fine Tuning with SVD-LoRA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 illustrates the TTFT procedure utilizing SVD-LoRA, which consists of two stages. In the first stage, the initial values of SVD-LoRA are set differently depending on the target task type. For closed-set tasks, the initial values are set through training the specific target task. In contrast, open-set tasks leverage initialization learned from common tasks like classification, which allows the model to generalize across various machine vision applications. The next stage involves performing TTFT to optimize each test sample according to the target downstream task. During this process, all of the SVD-LoRA weights in the encoder and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4433" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.482, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.485, + 0.414 + ], + "angle": 0, + "content": "Figure 4. TTFT procedure. The TTFT operates in two stages. First, the initial values for SVD-LoRA are initialized. In the second stage, TTFT is applied at test time to fine-tune each test sample in accordance with the specific downstream task. The source images of (b) is [56]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.483, + 0.681 + ], + "angle": 0, + "content": "only the SVs in the decoder are fine-tuned. While full finetuning of the encoder does not affect bit rate overhead or decoding latency, it requires a large number of parameters for each TTFT, leading to increased TTFT time overhead. Therefore, in our study, we apply SVD-LoRA to the encoder instead of performing full fine-tuning. As a result, this approach significantly reduces training costs and enables rapid adaptation to new downstream tasks. Additionally, fine-tuning only the SVs on the decoder side further reduces bitstream overhead compared to full LoRA finetuning. However, increasing the number of iterations during TTFT improves performance, but also increases encoding time. This issue can be addressed by adjusting the number of iterations according to the compression scenario, thereby balancing performance improvement with encoding time. (Detailed information is covered in the ablation study)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.713 + ], + "angle": 0, + "content": "The loss function of SVD-LoRA during TTFT is defined by Eq. (3):" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.727, + 0.483, + 0.745 + ], + "angle": 0, + "content": "\\[\nL = - \\log p (\\hat {z}) - \\log p (\\hat {y} \\mid \\hat {z}) + \\lambda d (x, \\hat {x}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( x \\), \\( y \\), and \\( z \\) represent the input image, encoded latent feature, and hyper prior, respectively. Additionally, \\( \\hat{x} \\), \\( \\hat{y} \\), and \\( \\hat{z} \\) represent their reconstruction. The terms \\( p(\\hat{z}) \\) and \\( p(\\hat{y} \\mid \\hat{z}) \\) represent the probabilities of the compressed latent variables, which correspond to the bit rate used for compressing the input \\( x \\). \\( d(x, \\hat{x}) \\) represents the distortion between the original and the reconstructed images. Mean square error (MSE) is used as the distortion metric when targeting human perception, as it measures the pixel-wise difference between the original and reconstructed images." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.088, + 0.911, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.311, + 0.908, + 0.382 + ], + "angle": 0, + "content": "Figure 5. An example of applying SVD-LoRA to the TIC codec. While SVD-LoRA is compatible with both CNN and STB layers, our work focuses solely on the STB layers of the main encoder and decoder to enhance task-switching efficiency with minimal tuning parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.577 + ], + "angle": 0, + "content": "For machine vision tasks such as image classification, perceptual loss is used as the distortion metric by comparing intermediate features extracted from the original and reconstructed images via the target network [13]. The hyper parameter \\(\\lambda\\) adjusts the balance between rate and distortion. This loss function allows for optimization that balances data compression and reconstruction while tailoring to the objective of each task. All experiments conducted in our work are formulated as rate-distortion optimization, with different values of \\(\\lambda\\) and distortion functions designed according to the characteristics of each downstream task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.594, + 0.72, + 0.612 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.722, + 0.636 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.734 + ], + "angle": 0, + "content": "While our method is compatible with any baseline architecture, we demonstrate its effectiveness using a widely adopted state-of-the-art (SOTA) LIC model—the Transformer-based LIC (TIC) network proposed in [46]—as the backbone. Results for the CNN-based LIC model [14] are included in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Fig. 5 shows the application of SVD-LoRA to the TIC model. The TIC adopts an autoencoder structure consisting of a main encoder-decoder pair and a hyper encoder-decoder pair, aiming to optimize the balance between distortion and bitrate for human perception. In our work, to maximize task-switching efficiency with minimal tuning parameters, we apply SVD-LoRA exclusively to the main encoder and decoder, as these components have the most significant impact on task adaptation [13]. While SVD-LoRA is compatible with both CNNs and Transformers, our approach specifically targets the Swin Transformer Block" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4434" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "(STB) in the main encoder and decoder to achieve optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.483, + 0.168 + ], + "angle": 0, + "content": "Datasets. We used MS-COCO [40] for object detection and instance segmentation, and ImageNet-1K [17] for image classification in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.261 + ], + "angle": 0, + "content": "Pre-trained Vision Models. To evaluate performance across various downstream tasks, we employed pre-trained vision models [29, 30, 46, 52] as task-specific evaluators. Specifically, we used ResNet50 [29] for image classification, Faster R-CNN [52] for object detection, and Mask R-CNN [30] for instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.29 + ], + "angle": 0, + "content": "Compared Methods. The methods compared in our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.293, + 0.483, + 0.323 + ], + "angle": 0, + "content": "(1) TIC [46]: TIC is a Transformer-based compression model pre-trained on the Flickr2W dataset [41]." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.324, + 0.483, + 0.367 + ], + "angle": 0, + "content": "(2) Full Fine-Tuning: Full Fine-Tuning updates all parameters of the backbone TIC to adapt to the target machine vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.369, + 0.483, + 0.413 + ], + "angle": 0, + "content": "(3) TransTIC [13]: TransTIC applies VPT to adapt the backbone TIC—originally trained for human perception—to target machine vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.414, + 0.483, + 0.459 + ], + "angle": 0, + "content": "(4) Unified Image Compression (UIC) [24]: UIC uses a collaborative embedding with a diffusion model to generate a universal bitstream to handle open-set tasks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.459, + 0.483, + 0.625 + ], + "angle": 0, + "content": "(5) Other Adaptation Approaches: These approaches apply adaptation techniques to the backbone TIC: (5-1) TIC+SFT [61] integrates Spatial Feature Transform (SFT) layers into TIC, where the SFT layers act as adapters for task-specific modulation; (5-2) TIC+channel_selection [43] selects partial channels from the image latent representation produced by TIC and converts them into task-specific feature maps to enable efficient task transfer; (5-3) TIC+prefilter [36] involves adding a pre-neural network before TIC to assist in transitioning TIC to the target task." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.293, + 0.483, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.808 + ], + "angle": 0, + "content": "Task Types and TTFT settings. Experiments were conducted under both closed-set and open-set conditions. We used a classification model to extract generalized features for open-set initialization. Therefore, the open-set image classification task is excluded, since its initialization is derived from its own target model. In all experiments, 'w/o TTFT' refers to the task-specific approach using fixed SVD-LoRA weights, while 'w/ TTFT' denotes the instance-specific approach with TTFT applied per test sample. TTFT is performed for 40 iterations per sample in our experiments; however, the number of iterations is adjustable and its impact on performance is analyzed in the ablation study." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Evaluation Metrics. Top-1 accuracy is used to evaluate image classification performance, whereas mean average precision (mAP) is employed for both object detection and instance segmentation. For numerical comparisons, the BDA-rate metric [23] is used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Strategy for handling different quality levels. In all ex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "periments, the four quality levels of TIC (from quality level 1 to 4) are used to visually represent the distribution between the rate and evaluation metrics. The \\(\\lambda\\) values of loss function are set to 2, 1, 0.5, and 0.2, depending on the quality level [13]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.905, + 0.21 + ], + "angle": 0, + "content": "5.2. Evaluation for Multiple Machine Vision Tasks and Human perception" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Fig. 6 compares our proposed approach with other methods across various machine vision tasks. The results demonstrate that our method, with TTFT applied, outperforms all competing approaches. Table 1 presents a performance comparison using BD-Rate, BD-mAP, and BD-accuracy with TIC as the anchor. The results show that our method consistently outperforms across all tasks, not only under closed-set conditions but also in challenging open-set settings. While the task-specific approach (denoted as 'w/o TTFT') achieves performance comparable to other competing methods, applying TTFT further improves performance. Even under challenging open-set settings, TTFT demonstrates that the model remains robust. Notably, our method achieves higher performance under open-set conditions than TransTIC, one of the SOTA techniques, does under closed-set conditions, demonstrating its superior generalization capability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.696, + 0.499 + ], + "angle": 0, + "content": "5.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Fig. 7 visualizes the results of object detection. Although TIC yields high visual quality from the perspective of human perception, its detection confidence scores are relatively low, limiting its effectiveness for downstream tasks. Furthermore, in all experiments, the visualized results exhibit noticeable differences between closed-set and open-set conditions. Since closed-set tasks are pre-trained on the target object detection task, high prediction scores can be achieved regardless of whether TTFT is applied. The corresponding results also tend to preserve fine texture details. In contrast, the open-set setting relies on generalized features that are not optimized for the target task, leading to reduced prediction accuracy and blurred textures. Nonetheless, in both settings, TTFT improves performance by enhancing texture detail preservation and enabling higher prediction scores. Notably, TTFT also maintains model robustness under challenging open-set conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.679, + 0.787 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.854 + ], + "angle": 0, + "content": "The ablation studies are conducted under the conditions of the open-set object detection task, utilizing an A100 GPU with 80GB memory. The evaluation is performed using 5,000 validation images from the MS-COCO dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Adjustment of fine-tuning iterations. Table 2 shows the performance and time overhead with varying numbers of TTFT iterations. As the number of iterations increases," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4435" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.223, + 0.083, + 0.4, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.232, + 0.39, + 0.253 + ], + "angle": 0, + "content": "(a) Instance Segmentation (open-set)" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.083, + 0.588, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.232, + 0.559, + 0.253 + ], + "angle": 0, + "content": "(b) Object Detection (open-set)" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.099, + 0.756, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.255, + 0.402, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.403, + 0.389, + 0.424 + ], + "angle": 0, + "content": "(c) Instance Segmentation (closed-set)" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.255, + 0.585, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.403, + 0.56, + 0.424 + ], + "angle": 0, + "content": "(d) Object Detection (closed-set)" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.254, + 0.772, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.403, + 0.738, + 0.424 + ], + "angle": 0, + "content": "(e) Classification (closed-set)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.44, + 0.907, + 0.469 + ], + "angle": 0, + "content": "Figure 6. Rate-accuracy performance comparison under various machine tasks. Experimental results show that our method outperforms the SOTA approaches in various closed-set and open-set target tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.483, + 0.907, + 0.511 + ], + "angle": 0, + "content": "Table 1. Comparison of BD-Rate, BD-mAP, and BD-Accuracy using TIC as the anchor. Bold numbers indicate the best performance, while blue numbers represent the second-best performance." + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.521, + 0.819, + 0.692 + ], + "angle": 0, + "content": "
Task TypeMethodInstance SegmentationObject DetectionClassification
BD-Rate (%) ↓BD-mAP ↑BD-Rate (%) ↓BD-mAP ↑BD-Rate (%) ↓BD-accuracy ↑
Closed-setFull fine tuning-75.554.00-66.004.10-1000.16
TIC+prefilter-24.741.56-17.591.19-23.690.04
TIC+SFT-40.241.67-27.241.45-61.530.11
TIC+channel_selection47.58-2.5552.85-3.77-30.370.06
TransTIC-47.452.34-46.642.71-58.760.10
UIC87.99-3.67102.70-4.31--
Ours (w/o TTFT)-43.441.99-38.622.17-50.670.08
Ours (w/ TTFT)-56.492.68-54.962.98-70.040.14
Open-setTransTIC-43.552.28-39.602.23--
UIC118.82-5.28146.85-6.37--
Ours (w/o TTFT)-32.481.97-29.611.87--
Ours (w/ TTFT)-49.512.43-48.942.69--
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.486, + 0.901 + ], + "angle": 0, + "content": "both performance and time overhead also increase. While time overhead increases linearly, performance gains are non-linear, with the majority of improvements occurring in the early iterations. Based on these characteristics of TTFT, our method can be effectively adapted to various compression scenarios. In low-latency applications such as CCTV surveillance systems, where real-time compression and transmission within milliseconds is required, the task-specific approach (denoted as 'w/o TTFT') is preferable. In contrast, scenarios such as satellite imagery or medical imaging—where image updates occur over the course of minutes to days—are more tolerant of encoding latency. In such cases, the instance-specific approach with hundreds" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.703, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Table 2. Ablation on the range of TTFT iterations. Experiments show that early iterations have the most impact, and TTFT iterations control time overhead and performance." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.754, + 0.883, + 0.833 + ], + "angle": 0, + "content": "
Bit-ratemAPΔmAP/iter.TTFT Time (sec)
w/o TTFT0.0959134.021--
w/ TTFT5 iters.0.10635.280.251.18
10 iters.0.10635.840.184.28
15 iters.0.10735.880.125.86
40 iters.0.10636.250.0614.13
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.838, + 0.907, + 0.868 + ], + "angle": 0, + "content": "or more TTFT iterations may be deployed without critical latency concerns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Varying the range of fine-tuning. Fig. 8 shows the object detection performance with varying fine-tuning ranges. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4436" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.092, + 0.34, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.092, + 0.498, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.092, + 0.657, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.092, + 0.815, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.248, + 0.338, + 0.261 + ], + "angle": 0, + "content": "Ours (open-set, w/o TTFT)" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.248, + 0.498, + 0.26 + ], + "angle": 0, + "content": "Ours (open-set, w/ TTFT)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.248, + 0.657, + 0.26 + ], + "angle": 0, + "content": "Ours (closed-set, w/o TTFT)" + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.248, + 0.813, + 0.26 + ], + "angle": 0, + "content": "Ours (closed-set, w/ TTFT)" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.262, + 0.339, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.262, + 0.498, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.262, + 0.657, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.262, + 0.814, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.403, + 0.908, + 0.446 + ], + "angle": 0, + "content": "Figure 7. Visualization of object detection results. While TIC provides better visual quality, it exhibits low detection confidence, leading to misdetections—such as misclassifying a bear as a sheep. Ours (w/ TTFT) surpasses the uncompressed image in prediction confidence despite compression losses." + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.45, + 0.426, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Figure 8. Ablation on the range of fine-tuning. Experiments show that fine-tuning the decoder is essential for optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.794 + ], + "angle": 0, + "content": "comparison includes fine-tuning either the encoder or the decoder, or both together. The experimental results show that TTFT applied to the decoder stage contributes to performance improvement. While the encoder stage, which involves a larger number of trainable parameters during TTFT, yields greater performance gains, the decoder-stage TTFT remains essential for achieving optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fig. 9 presents a comparison between decoder finetuning using only the SVs and using the full set of SVD-LoRA parameters. The results show that training all SVD-LoRA parameters increases bitstream overhead by \\(15.3\\mathrm{x}\\). Therefore, restricting training to only the SVs in the decoder is critical for enabling fully instance-specific finetuning with minimal overhead." + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.45, + 0.845, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.697 + ], + "angle": 0, + "content": "Figure 9. Ablation on the range of decoder fine-tuning. Experiments show that training only the singular values of the decoder effectively reduces excessive bitstream overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.7, + 0.634, + 0.715 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose a fully instance-specific TTFT approach for adapting learned image compression to both closed-set and open-set machine vision tasks. Our method can overcome the challenging problem of instance-specific decoder finetuning, which remains difficult due to its significant impact on both bitstream size and decoding time but is essential for effectively handling open-set tasks. The SVD-LoRA we applied enables TTFT while minimizing bitstream overhead. Moreover, it can be seamlessly integrated into any architecture, including CNNs and Transformers. Therefore, our approach can be applied to various domains where instance-specific TTFT is required." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4437" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.342 + ], + "angle": 0, + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) under multiple grants funded by the Korea government (MSIT), including the ICT Creative Consilience Program (IITP-2025-RS-2020-II201821), the Information Technology Research Center (ITRC) support program (IITP-2021-0-02052), and the Artificial Intelligence Innovation Hub (RS-2021-II212068). Additionally, this research was supported by Korea Research Institute for defense Technology planning and advancement (KRIT) grant funded by the Korean Government, Defense Acquisition Program Administration (DAPA), under Grant KRIT-CT-23-021, Development of Multi-Modal Information Fusion and High-Realization Battlefield Data Generation Technology, in 2025." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.355, + 0.188, + 0.371 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.38, + 0.484, + 0.448 + ], + "angle": 0, + "content": "[1] Umar Albalawi, Saraju P Mohanty, and Elias Kougianos. A hardware architecture for better portable graphics (bpg) compression encoder. In 2015 IEEE international Symposium on Nanoelectronic and information systems, pages 291-296. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.45, + 0.484, + 0.52 + ], + "angle": 0, + "content": "[2] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[3] Dana Harry Ballard and Christopher M Brown. Computer vision. Prentice Hall Professional Technical Reference, 1982. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[4] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In International Conference on Learning Representations, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[6] Klaudia Bałazy et al. Lora-xs: Low-rank adaptation with extremely small number of parameters. arXiv preprint, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[7] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Prabuddha Chakraborty, Jonathan Cruz, and Swarup Bhunia. Magic: Machine-learning-guided image compression for vision applications in internet of things. IEEE Internet of Things Journal, 8(9):7303-7315, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.38, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[10] Lahiru D Chamain, Fabien Racapé, Jean Bégaint, Akshay Pushparaja, and Simon Feltman. End-to-end optimized image compression for machines, a study. In 2021 Data Compression Conference (DCC), pages 163-172. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[11] Lahiru D Chamain, Fabien Racapé, Jean Bégaint, Akshay Pushparaja, and Simon Feltman. End-to-end optimized image compression for multiple machine tasks. arXiv preprint arXiv:2103.04178, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[12] Arnav Chavan et al. One-for-all: Generalized lora for parameter-efficient fine-tuning. arXiv preprint arXiv:2306.07967, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.907, + 0.329 + ], + "angle": 0, + "content": "[13] Yi-Hsin Chen, Ying-Chieh Weng, Chia-Hao Kao, Cheng Chien, Wei-Chen Chiu, and Wen-Hsiao Peng. Transtic: Transferring transformer-based image compression from human perception to machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23297-23307, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[14] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[15] Hyomin Choi and Ivan V Bajic. Scalable image coding for humans and machines. IEEE Transactions on Image Processing, 31:2739-2754, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[16] Jack Cole and Mohamed Osman. Dataset-induced meta-learning (and other tricks): Improving model efficiency on arc, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[18] Lingyu Duan, Jiaying Liu, Wenhan Yang, Tiejun Huang, and Wen Gao. Video coding for machines: A paradigm of collaborative compression and intelligent analytics. IEEE Transactions on Image Processing, 29:8680-8695, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[19] Zhihao Duan, Zhan Ma, and Fengqing Zhu. Unified architecture adaptation for compressed domain semantic inference. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):4108-4121, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[20] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[21] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[22] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[23] Bjontegaard Gisle. Calculation of average psnr differences between rd curves. In ITU-T SG16/Q6, \\(13^{\\circ}\\) th VCEG Meeting, Austin, Texas, USA, April 2001, 2001. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Sha Guo, Lin Sui, Chenlin Zhang, Zhuo Chen, Wenhan Yang, and Lingyu Duan. A unified image compression" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4438" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "method for human perception and multiple vision tasks. In European Conference on Computer Vision, pages 342-359. Springer, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[25] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.474, + 0.221 + ], + "angle": 0, + "content": "[26] Allen Hanson. Computer vision systems. Elsevier, 1978. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.222, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[27] Moritz Hardt and Yu Sun. Test-time training on nearest neighbors for large language models. In The Twelfth International Conference on Learning Representations, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[28] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.363, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.421, + 0.483, + 0.462 + ], + "angle": 0, + "content": "[30] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[31] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In Proceedings of the ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[32] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[33] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[34] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Realtime multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[35] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[36] Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien. How to exploit the transferability of learned image compression to conventional codesc. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16165-16174, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: an end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[38] Huanyang Li and Xinfeng Zhang. Human-machine collaborative image compression method based on implicit neural representations. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[39] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, 133(1):31-64, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[41] Jiaheng Liu, Guo Lu, Zhihao Hu, and Dong Xu. A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[42] Jinming Liu, Heming Sun, and Jiro Katto. Learning in compressed domain for faster machine vision tasks. In 2021 International Conference on Visual Communications and Image Processing (VCIP), pages 01-05. IEEE, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[43] Jinming Liu, Heming Sun, and Jiro Katto. Improving multiple machine vision tasks in the compressed domain. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 331-337. IEEE, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[44] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14388-14397, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[45] Kang Liu, Dong Liu, Li Li, Ning Yan, and Houqiang Li. Semantics-to-signal scalable image compression with learned revertible representations. International Journal of Computer Vision, 129(9):2605-2621, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[46] Ming Lu, Peiyao Guo, Huiqing Shi, Chuntong Cao, and Zhan Ma. Transformer-based image compression. In 2022 Data Compression Conference (DCC), pages 469-469. IEEE, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[47] Siwei Ma, Xinfeng Zhang, Chuanmin Jia, Zhenghui Zhao, Shiqi Wang, and Shanshe Wang. Image and video compression with neural networks: A review. IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683-1698, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[48] Yixin Mei, Fan Li, Li Li, and Zhu Li. Learn a compression for objection detection-vae with a bridge. In 2021 International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[49] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[50] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.904, + 0.873 + ], + "angle": 0, + "content": "[51] Vishvjit S Nalwa. A guided tour of computer vision. Addison-Wesley Longman Publishing Co., Inc., 1994. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4439" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[53] Myungseo Song, Jinyoung Choi, and Bohyung Han. Variable-rate deep image compression through spatially-adaptive feature transform. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2380-2389, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[54] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[55] Vivienne Sze, Madhukar Budagavi, and Gary J Sullivan. High efficiency video coding (hevc). In Integrated circuit and systems, algorithms and architectures, page 40. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[56] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[57] Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. Dylora: Parameter-efficient tuning of pretrained models using dynamic search-free low-rank adaptation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3274-3287, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[58] Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, and Eftychios Protopapadakis. Deep learning for computer vision: A brief review. Computational intelligence and neuroscience, 2018(1):7068349, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[59] Shurun Wang, Zhao Wang, Shiqi Wang, and Yan Ye. End-to-end compression towards machine vision: Network architecture design and optimization. IEEE Open Journal of Circuits and Systems, 2:675-685, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[60] Shurun Wang, Zhao Wang, Shiqi Wang, and Yan Ye. Deep image compression toward machine vision: A unified optimization framework. IEEE Transactions on Circuits and Systems for Video Technology, 33(6):2979-2989, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[61] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[62] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. Sssic: semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[64] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fin-tuning. In International Conference on Learning Representations, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[65] Y Zhang, M Rafie, and S Liu. Use cases and requirements for video coding for machines. ISO/IEC JTC, 1, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[66] Xia Zhao, Limin Wang, Yufei Zhang, Xuming Han, Muhammet Deveci, and Milan Parmar. A review of convolutional neural networks in computer vision. Artificial Intelligence Review, 57(4):99, 2024. 1" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4440" + } + ] +] \ No newline at end of file diff --git a/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_origin.pdf b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a8654a1b7e60e5cb4dca80f25c51d2089967918 --- /dev/null +++ b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/aab7c6c2-25ab-43c4-8b18-560bbb6a0c36_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03e822c1dae19f6f87e142422e334b32ff7c8b47e0ed2e67355570746815fdcf +size 3109448 diff --git a/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/full.md b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5229d6b39c64df3c6c832eddc06396a5a6bbae95 --- /dev/null +++ b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/full.md @@ -0,0 +1,302 @@ +# Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability + +Unki Park $^{1,2}$ Seongmoon Jeong $^{1}$ Youngchan Jang $^{3}$ Gyeong-Moon Park $^{4*}$ Jong Hwan Ko $^{1*}$ $^{1}$ Sungkyunkwan University, $^{2}$ Samsung Institute of Technology, $^{3}$ Hanwha Systems Corporation, $^{4}$ Korea University +{unkipark@g., giftjsm@, hjko@}skku.edu, youngchan84.jang@hanwha.com, gm-park@korea.ac.kr + +# Abstract + +The field of computer vision was initially inspired by the human visual system and has progressively expanded to include a broader range of machine vision applications. Consequently, image compressors should be designed to effectively accommodate not only human visual perception but also machine vision tasks, including closed-set scenarios that enable pre-training and open-set scenarios that involve previously unseen tasks at test time. Many recent studies effectively address both human visual perception and closed-set machine vision tasks simultaneously but struggle to handle open-set machine vision tasks. To address this issue, this paper proposes a fully instance-specific test time fine-tuning (TTFT) for adapting learned image compression (LIC) to both closed-set and open-set machine vision tasks effectively. With our method, a large-scale LIC model, originally trained for human perception, is adapted to the target task through TTFT using Singular Value Decomposition based Low Rank Adaptation (SVD-LoRA). During TTFT, the decoder adopts a modified learning scheme that focuses exclusively on training the singular values, which helps prevent excessive bitstream overhead. This enables fully instance-specific optimization for the target task, even for open-set tasks. Experimental results demonstrate that the proposed method effectively adapts the backbone compressor to diverse machine vision tasks, outperforming competing methods. The code is available at project page. + +# 1. Introduction + +In the early stages of computer vision, image processing technologies were primarily developed to align with human perception, as images were largely intended for human interpretation [3, 26, 51]. However, with advances in deep learning and computer vision techniques enabling more sophisticated data processing, machines have increasingly gained the ability to interpret visual data. Conse + +![](images/5c316734b66866dee51e1626811074b3748e8d1c3c92146f9f0118e57b7e2012.jpg) +Figure 1. An overview of our proposed method. Our method tunes both the encoder and decoder of a backbone human-target LIC model by fine-tuning SVD-LoRA weights optimized for each target task and image instance. + +sequently, the scope of computer vision has broadened to include machine vision applications [56, 58, 66]. In the field of image compression as well, there is a growing need for methods tailored to the requirements of machine vision, as machines are expected to become the primary users of compressed image data [18, 65]. + +Accordingly, several research studies have shifted their focus away from the traditional goal of visual signal reconstruction, instead prioritizing the enhancement of machine vision task performance [9, 59]. These studies commonly aim to effectively restore representations suitable for the targeted machine vision domain, while selectively sacrificing visual quality to achieve this goal. + +Taking it further, recent research has focused on developing learned image codes (LICs) that can cater to both human perception and various machine vision tasks [10, 13, 37, 38, 41, 53, 60]. These studies can be categorized into methods that generate either multi-task bitstreams or single-task bitstreams. The multi-task bitstream approach aims to generate a unified bitstream that can support vari- + +ous subtasks including human perception and multiple machine tasks. This method offers flexibility in responding to a range of machine vision tasks, but it introduces challenges. For example, scalable coding [15, 32, 45, 62] increases the processing complexity and bitstream overhead for individual downstream tasks, and unified image coding [11, 19, 20, 24] struggles with optimizing each subtask individually. + +In contrast, the single-task bitstream approach [13, 36, 42, 43, 48, 53] avoids the limitations of multi-task methods by dynamically generating bitstreams for a specific downstream task. For instance, TransTIC[13] utilizes prompt tuning for adapting the human-target backbone codec to generate bitstreams optimized for the other tasks specified by the prompt. While this technique enables switching of pre-defined (closed-set) tasks without modifying the backbone codec, it requires task-specific pre-training of the prompt generator. Therefore, it leads to increased training demands as the number of tasks grows, and faces challenges to handle new, untrained (open-set) tasks. Moreover, performance improvement is non-monotonic with respect to prompt size, and computational costs increase due to the added prompt during inference. + +To address these limitations, we aim to develop a method that can dynamically adapt a given backbone codec to optimize rate-performance trade-offs for the specific requirements of each task. Our goal is to provide flexible support for diverse tasks and codec models, allowing adaptation to previously unseen (open-set) tasks without requiring pre-training, and also adaptable to any given codec model. To achieve this, as shown in Fig. 1, we propose tuning the backbone codec using Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning approach widely used in test-time scenarios. Unlike the TransTIC approach, which cannot perform instance-specific tuning on the decoder, our method tunes both the encoder and decoder by generating LoRA weights optimized for each target task and image instance. This enables dynamic codec optimization for any arbitrary task without requiring pre-training, making it effective for open-set scenarios. While instance-specific model optimization can lead to superior performance compared to task-specific optimization, it introduces additional overhead by requiring a separate set of parameters for each image. Specifically, in the case of instance-specific decoder tuning, transmitting the LoRA weights incurs additional bit rate overhead. To address this efficiently, we adopt an SVD-LoRA approach that combines the original LoRA weights with supplementary singular value (SV) components. After pre-initializing the original LoRA weights, we can then transmit only the SV components to the decoder on an instance-specific basis, allowing effective tuning while minimizing bit rate demands. As the SV component captures essential information for optimizing the codec per target in + +stance and task, it is both effective and efficient in managing rate requirements. + +Experimental results demonstrate that our method achieves superior adaptation performance for both open-set and closed-set tasks. Although conducted under a lightweight experimental setup (40 iterations, LoRA rank 4), we observed that our method achieved more than a $50\%$ improvement in Bjøntegaard Delta (BD)-Rate [23] compared to the backbone codec across all evaluated machine vision tasks. These results surpass all competing methods. Notably, our method achieves superior performance in a pure open-set setting—where no prior task-specific setup is available—outperforming competing methods even when they are evaluated under pre-initialized (closed-set) conditions. + +The main contributions of our work can be summarized as follows: + +- Test-time fine-tuning in compression field : To the best of our knowledge, we are the first to attempt TTFT at the decoder stage in a image compression task. Considering that the encoder and decoder are temporally and spatially separated in compression systems, TTFT has the potential to maximize image compression quality. +- Novel approach utilizing LoRA: We propose a novel approach utilizing LoRA for efficient adaptation to diverse machine vision applications in the still-image compression task. Specifically, we introduce a modified LoRA in the form of SVD to enable instance-specific optimization at the decoder. This approach not only enhances efficiency but also facilitates effective TTFT at the decoder stage. +- Handling open-set tasks: By designing an instance-specific optimized model that does not necessarily require pre-training, we have enhanced the ability to tackle open-set machine vision tasks, which have remained challenging in previous research. + +# 2. Related Work + +The initial development of LIC focused on optimizing perceptual quality for human visual systems, similar to traditional compressors [1, 7, 55]. Most LIC models adopt an autoencoder structure [4, 5, 44, 47, 50], aiming to optimize the balance between distortion and bit-rate. In early LIC research, the goal was to minimize losses limited to human visual perception, with the technology evolving towards faithfully mimicking human visual quality [28, 46, 49]. With the rapid development of machine vision tasks such as image classification and object detection, an increasing number of applications now incorporate images encoded and decoded by image codecs for these downstream tasks. Consequently, LIC research has begun to focus on meeting machine vision requirements. Initially, the emphasis was on creating + +individual models tailored to each specific task. More recently, research has proposed approaches where a shared LIC model architecture is individually trained to generate task-specific models. Following this trend, research efforts aimed at designing a single LIC model to support multiple targets including human perception and diverse machine vision tasks. These studies can be divided into multi-task bitstream and single-task bitstream methods. + +Multi-task Bitstream LICs for Multiple Targets. One line of research in multi-target LIC is the multi-task bitstream approach [11, 15, 19, 20, 24, 32, 45, 62], which aims to accommodate multiple targets using a single bitstream. However, this approach increases the complexity of both the encoder and decoder [15, 24]. It also leads to significant bitstream overhead from the perspective of individual downstream tasks [15]. Additionally, it faces the challenge of Pareto optimality, which makes it difficult to optimize each subtask individually [24]. + +Adaptive Single-task Bitstream LICs for Multiple Targets. Another line of study for multi-target LIC is the single-task bitstream approach [13, 36, 42, 43, 48, 53]. This approach generates a bitstream specifically for one target task at a time, adapting to the given target task to support multi-task requirements effectively. This allows the approach to achieve optimal performance under well-defined conditions for each specific subtask. Representative studies in this field involve applying task-adaptive importance modulation or fine-tuning techniques to the backbone LIC. Song et al. [53] proposed a spatially-adaptive image compression method using Spatial Feature Transform (SFT) guided by a real-valued quality map that enables pixel-wise optimization for machine vision tasks. Liu et al. [43] introduced a task-adaptive image compression method that uses a gating module to selectively activate channel-wise latents from a pre-trained codec. J.P. Klopp et al. [36] introduced a learned pre-filter that suppresses task-irrelevant data prior to compression by distinguishing between important and unimportant information. TransTIC [13] employed Visual Prompt Tuning (VPT) [21, 33] to adapt the largescale pretrained model TIC [46], originally trained for human perception, to suit various target tasks. This approach demonstrated adaptability to various machine vision tasks such as object detection, instance segmentation, and classification [13]. However, VPT itself faces limitations such as non-monotonic performance changes depending on the size of the prompt applied, as well as increased computational cost due to the additional prompts during the inference stage [12]. More importantly, TransTIC adjusts the prompt to guide the output of a static pretrained model, but the generator responsible for creating this prompt also needs to be pretrained. Also, it requires prompts equivalent to one-fourth of the input feature size of each layer to perform task switching. Therefore, for applying fine-tuning to + +individual test samples at the decoder stage, these prompts must be incorporated into the bitstream, making TransTIC unsuitable for instance-specific decoder fine-tuning. Consequently, pre-training is essential to adapt to new tasks, and it is difficult to adapt to untrained open-set tasks. + +To overcome these limitations, we propose using TTFT to dynamically adapt a backbone LIC model for diverse tasks. Since TTFT adjusts part or all of the model through a brief learning phase during testing [16, 27, 39, 54], it effectively supports open set scenarios where the codec needs to be optimized for previously unseen tasks. Although TTFT is highly effective in producing results optimized for individual samples, it has the limitation of requiring additional training steps at each testing instance. Therefore, we utilize LoRA [31], one of the parameter-efficient fine-tuning (PEFT) techniques to reduce training costs while effectively adapting the model to downstream tasks. + +# 3. Preliminary + +The LoRA method is one of the most actively researched PEFT techniques as it provides effective task adaptation with only a small number of parameters without increasing model complexity and inference time [6, 31, 57]. It focuses on learning the intrinsic rank of the weights [31]. Let the pre-trained weights be denoted by $W_{o} \in \mathbb{R}^{d \times k}$ , and let the weight change required for task adaptation be denoted by $\Delta W$ . After the task adaptation, the new weight becomes $W_{o} + \Delta W$ . Typically, the intrinsic rank of the weight change required for task adaptation is relatively small compared to the size of the entire weight matrix, meaning that $\Delta W$ can be decomposed into a low-rank representation [31]. From this, the relationship in Eq. (1) can be obtained: + +$$ +W = W _ {o} + \Delta W = W _ {o} + B \cdot A, \tag {1} +$$ + +where $B \in \mathbb{R}^{d \times r}$ , $A \in \mathbb{R}^{r \times k}$ , and $r \ll \min(d, k)$ . Since $W_{o}$ remains fixed, $B \cdot A$ can be explicitly separated in $W$ , allowing us to store only the matrices $B$ and $A$ for task adaptation, which requires very minimal memory overhead. During inference, a reparameterization process is performed in which $B \cdot A$ is added to $W_{o}$ to reconstruct $W$ , thereby avoiding any additional inference latency compared to the original model. The LoRA technique is applicable to both CNN and Transformer architectures and can be applied to all layers as long as the memory overhead issue is manageable. Recently, this technique has been reformulated as an SVD-based form and applied to LLMs [6, 25, 64]. These studies [6, 64] aim to fine-tune large pre-trained language models for downstream tasks in a parameter-efficient manner. Specifically, Zhang et al. [64] utilizes the SVD shaped LoRA to allocate parameter budgets adaptively based on the importance scores of each weight matrix. Meanwhile, Balazy et al. [6] proposed a method that applies SVD + +![](images/e2488f48a4f9e6a612b8f0a4658ed13e65450cbeacf1bd620f4035653ce46bf6.jpg) +Figure 2. Original LoRA vs. SVD-LoRA. When the rank is small, SVD-LoRA more effectively captures the intrinsic dimensions of the adaptation weights. + +shaped LoRA to fine-tune only the SV weight matrices during task switching. These novel approaches in the LLM domain show promising potential for effective application in the field of image compression as well. + +# 4. Methodology + +# 4.1. Framework Overview + +Our goal is to adapt the given backbone codec for various downstream tasks, including open-set tasks that are unseen during training. To achieve this, we adapt the backbone LIC to target downstream tasks by applying TTFT [16, 27] on both the encoder and decoder. Specifically, we use LoRA, a widely used approach in test-time scenarios. By tuning the LoRA weights to be optimized for each target task and image instance, the backbone LIC can be adapted to better suit the target task. However, transmitting LoRA weights to the decoder for fully instance-specific tuning incurs a nonnegligible bit rate overhead. To minimize the bit rate issue, we employ a reformulated version of LoRA, referred to as SVD-LoRA, which enables effective fine-tuning of only a small portion of the decoder at test time. While prior works utilized SVD-LoRA for language task-specific adaptation [6, 25, 64], we focus on the test-time decoder fine-tuning for a fully instance-specific optimized compression model. Consequently, the proposed method can flexibly handle diverse, arbitrary target tasks and is compatible with various backbone LIC architectures, including both CNN and Transformer-based models. + +# 4.2. SVD based LoRA + +SVD-LoRA is represented by Eq. (2): + +$$ +W = W _ {o} + \Delta W = W _ {o} + U \cdot \Sigma \cdot V ^ {T}, \tag {2} +$$ + +where, $U \in \mathbb{R}^{d \times r}$ , $\Sigma \in \mathbb{R}^{r \times r}$ , $V \in \mathbb{R}^{k \times r}$ , and $r \ll \min(d, k)$ . Here, $U$ and $V$ learn the intrinsic ranks for + +![](images/8e8bfeb8b3aea92d9b9b0ce4ccbc0dde7f60310fafe0e0fa36bea95282160c32.jpg) +Figure 3. Bitstream generation process. For instance-specific decoder tuning with manageable bitstream overhead, only the singular values of the decoder side are encoded and added to the image-encoded bitstream. The bit-rate overhead accounts for less than $1\%$ of the image bitstream size. + +columns and rows, respectively, while $\Sigma$ is a diagonal matrix that learns the SVs. + +Although the structure includes additional SVs compared to conventional LoRA, the increase in the number of parameters is negligible due to the typically small rank used. Fig. 2 illustrates the process of reconstructing $W$ (adapted weights) through the reparameterization of LoRA and SVD-LoRA, respectively. By learning additional SVs, SVD-LoRA facilitates the learning of more complex intrinsic dimensions of adaptation weights compared to the original LoRA. In cases where both LoRA and SVD-LoRA have the same small rank, experimental results indicate that SVD-LoRA better captures the intrinsic dimensions of the adaptation weights. Details of this observation are discussed in the supplementary material. Fig. 3 illustrates the process of embedding the decoder-side SVD-LoRA weights into the compressed bitstream for instance-specific fin-tuning at the decoder stage. While the bitstream size increases due to the additional SV component, the overhead is negligible. Specifically, with a rank of 4, as specified in our experimental settings, the additional overhead accounts for less than $1\%$ of the compressed image bitstream size. + +# 4.3. Test-Time Fine Tuning with SVD-LoRA + +Fig. 4 illustrates the TTFT procedure utilizing SVD-LoRA, which consists of two stages. In the first stage, the initial values of SVD-LoRA are set differently depending on the target task type. For closed-set tasks, the initial values are set through training the specific target task. In contrast, open-set tasks leverage initialization learned from common tasks like classification, which allows the model to generalize across various machine vision applications. The next stage involves performing TTFT to optimize each test sample according to the target downstream task. During this process, all of the SVD-LoRA weights in the encoder and + +![](images/c9b3def41b67a2cf0616a6c891e58b1bf1518019fbbdfce1eb6656da7b9e8b26.jpg) +Figure 4. TTFT procedure. The TTFT operates in two stages. First, the initial values for SVD-LoRA are initialized. In the second stage, TTFT is applied at test time to fine-tune each test sample in accordance with the specific downstream task. The source images of (b) is [56] + +only the SVs in the decoder are fine-tuned. While full finetuning of the encoder does not affect bit rate overhead or decoding latency, it requires a large number of parameters for each TTFT, leading to increased TTFT time overhead. Therefore, in our study, we apply SVD-LoRA to the encoder instead of performing full fine-tuning. As a result, this approach significantly reduces training costs and enables rapid adaptation to new downstream tasks. Additionally, fine-tuning only the SVs on the decoder side further reduces bitstream overhead compared to full LoRA finetuning. However, increasing the number of iterations during TTFT improves performance, but also increases encoding time. This issue can be addressed by adjusting the number of iterations according to the compression scenario, thereby balancing performance improvement with encoding time. (Detailed information is covered in the ablation study) + +The loss function of SVD-LoRA during TTFT is defined by Eq. (3): + +$$ +L = - \log p (\hat {z}) - \log p (\hat {y} \mid \hat {z}) + \lambda d (x, \hat {x}), \tag {3} +$$ + +where $x$ , $y$ , and $z$ represent the input image, encoded latent feature, and hyper prior, respectively. Additionally, $\hat{x}$ , $\hat{y}$ , and $\hat{z}$ represent their reconstruction. The terms $p(\hat{z})$ and $p(\hat{y} \mid \hat{z})$ represent the probabilities of the compressed latent variables, which correspond to the bit rate used for compressing the input $x$ . $d(x, \hat{x})$ represents the distortion between the original and the reconstructed images. Mean square error (MSE) is used as the distortion metric when targeting human perception, as it measures the pixel-wise difference between the original and reconstructed images. + +![](images/8e2136279f63654bc5f27db8caa761b5026b633439703f9b376a22fbd20e31f2.jpg) +Figure 5. An example of applying SVD-LoRA to the TIC codec. While SVD-LoRA is compatible with both CNN and STB layers, our work focuses solely on the STB layers of the main encoder and decoder to enhance task-switching efficiency with minimal tuning parameters. + +For machine vision tasks such as image classification, perceptual loss is used as the distortion metric by comparing intermediate features extracted from the original and reconstructed images via the target network [13]. The hyper parameter $\lambda$ adjusts the balance between rate and distortion. This loss function allows for optimization that balances data compression and reconstruction while tailoring to the objective of each task. All experiments conducted in our work are formulated as rate-distortion optimization, with different values of $\lambda$ and distortion functions designed according to the characteristics of each downstream task. + +# 5. Experimental Results + +# 5.1. Experimental Settings + +While our method is compatible with any baseline architecture, we demonstrate its effectiveness using a widely adopted state-of-the-art (SOTA) LIC model—the Transformer-based LIC (TIC) network proposed in [46]—as the backbone. Results for the CNN-based LIC model [14] are included in the supplementary material. + +Fig. 5 shows the application of SVD-LoRA to the TIC model. The TIC adopts an autoencoder structure consisting of a main encoder-decoder pair and a hyper encoder-decoder pair, aiming to optimize the balance between distortion and bitrate for human perception. In our work, to maximize task-switching efficiency with minimal tuning parameters, we apply SVD-LoRA exclusively to the main encoder and decoder, as these components have the most significant impact on task adaptation [13]. While SVD-LoRA is compatible with both CNNs and Transformers, our approach specifically targets the Swin Transformer Block + +(STB) in the main encoder and decoder to achieve optimal performance. + +Datasets. We used MS-COCO [40] for object detection and instance segmentation, and ImageNet-1K [17] for image classification in our experiments. + +Pre-trained Vision Models. To evaluate performance across various downstream tasks, we employed pre-trained vision models [29, 30, 46, 52] as task-specific evaluators. Specifically, we used ResNet50 [29] for image classification, Faster R-CNN [52] for object detection, and Mask R-CNN [30] for instance segmentation. + +Compared Methods. The methods compared in our work are as follows: + +(1) TIC [46]: TIC is a Transformer-based compression model pre-trained on the Flickr2W dataset [41]. +(2) Full Fine-Tuning: Full Fine-Tuning updates all parameters of the backbone TIC to adapt to the target machine vision tasks. +(3) TransTIC [13]: TransTIC applies VPT to adapt the backbone TIC—originally trained for human perception—to target machine vision tasks. +(4) Unified Image Compression (UIC) [24]: UIC uses a collaborative embedding with a diffusion model to generate a universal bitstream to handle open-set tasks. +(5) Other Adaptation Approaches: These approaches apply adaptation techniques to the backbone TIC: (5-1) TIC+SFT [61] integrates Spatial Feature Transform (SFT) layers into TIC, where the SFT layers act as adapters for task-specific modulation; (5-2) TIC+channel_selection [43] selects partial channels from the image latent representation produced by TIC and converts them into task-specific feature maps to enable efficient task transfer; (5-3) TIC+prefilter [36] involves adding a pre-neural network before TIC to assist in transitioning TIC to the target task. + +Task Types and TTFT settings. Experiments were conducted under both closed-set and open-set conditions. We used a classification model to extract generalized features for open-set initialization. Therefore, the open-set image classification task is excluded, since its initialization is derived from its own target model. In all experiments, 'w/o TTFT' refers to the task-specific approach using fixed SVD-LoRA weights, while 'w/ TTFT' denotes the instance-specific approach with TTFT applied per test sample. TTFT is performed for 40 iterations per sample in our experiments; however, the number of iterations is adjustable and its impact on performance is analyzed in the ablation study. + +Evaluation Metrics. Top-1 accuracy is used to evaluate image classification performance, whereas mean average precision (mAP) is employed for both object detection and instance segmentation. For numerical comparisons, the BDA-rate metric [23] is used. + +Strategy for handling different quality levels. In all ex + +periments, the four quality levels of TIC (from quality level 1 to 4) are used to visually represent the distribution between the rate and evaluation metrics. The $\lambda$ values of loss function are set to 2, 1, 0.5, and 0.2, depending on the quality level [13]. + +# 5.2. Evaluation for Multiple Machine Vision Tasks and Human perception + +Fig. 6 compares our proposed approach with other methods across various machine vision tasks. The results demonstrate that our method, with TTFT applied, outperforms all competing approaches. Table 1 presents a performance comparison using BD-Rate, BD-mAP, and BD-accuracy with TIC as the anchor. The results show that our method consistently outperforms across all tasks, not only under closed-set conditions but also in challenging open-set settings. While the task-specific approach (denoted as 'w/o TTFT') achieves performance comparable to other competing methods, applying TTFT further improves performance. Even under challenging open-set settings, TTFT demonstrates that the model remains robust. Notably, our method achieves higher performance under open-set conditions than TransTIC, one of the SOTA techniques, does under closed-set conditions, demonstrating its superior generalization capability. + +# 5.3. Qualitative Results + +Fig. 7 visualizes the results of object detection. Although TIC yields high visual quality from the perspective of human perception, its detection confidence scores are relatively low, limiting its effectiveness for downstream tasks. Furthermore, in all experiments, the visualized results exhibit noticeable differences between closed-set and open-set conditions. Since closed-set tasks are pre-trained on the target object detection task, high prediction scores can be achieved regardless of whether TTFT is applied. The corresponding results also tend to preserve fine texture details. In contrast, the open-set setting relies on generalized features that are not optimized for the target task, leading to reduced prediction accuracy and blurred textures. Nonetheless, in both settings, TTFT improves performance by enhancing texture detail preservation and enabling higher prediction scores. Notably, TTFT also maintains model robustness under challenging open-set conditions. + +# 5.4. Ablation Studies + +The ablation studies are conducted under the conditions of the open-set object detection task, utilizing an A100 GPU with 80GB memory. The evaluation is performed using 5,000 validation images from the MS-COCO dataset. + +Adjustment of fine-tuning iterations. Table 2 shows the performance and time overhead with varying numbers of TTFT iterations. As the number of iterations increases, + +![](images/c6103282c531a3891da19f3e7adb8371bfbd82b136ad05277587f0ef28a036bb.jpg) + +![](images/034b463044f6315d7f515c348e2dbbdd8ca16b8edafb49b8ff41e080bde5156d.jpg) + +![](images/31273c596a8010511c737040a12136c90c60753b11bd8f00a777f0884365d93b.jpg) + +![](images/d0ee42624a85c88be03f936b97c02b92a11f505d1672709e703e118770e71361.jpg) +(a) Instance Segmentation (open-set) +(c) Instance Segmentation (closed-set) +Figure 6. Rate-accuracy performance comparison under various machine tasks. Experimental results show that our method outperforms the SOTA approaches in various closed-set and open-set target tasks. + +![](images/4d41a058e543a4977b630d1ca3387de4685819d95b66e9c993a2af899661f31d.jpg) +(b) Object Detection (open-set) +(d) Object Detection (closed-set) + +![](images/008c7b8c76da68d8257fd1d2802947560c32ffdc025f8f73c88254baaee48181.jpg) +(e) Classification (closed-set) + +Table 1. Comparison of BD-Rate, BD-mAP, and BD-Accuracy using TIC as the anchor. Bold numbers indicate the best performance, while blue numbers represent the second-best performance. + +
Task TypeMethodInstance SegmentationObject DetectionClassification
BD-Rate (%) ↓BD-mAP ↑BD-Rate (%) ↓BD-mAP ↑BD-Rate (%) ↓BD-accuracy ↑
Closed-setFull fine tuning-75.554.00-66.004.10-1000.16
TIC+prefilter-24.741.56-17.591.19-23.690.04
TIC+SFT-40.241.67-27.241.45-61.530.11
TIC+channel_selection47.58-2.5552.85-3.77-30.370.06
TransTIC-47.452.34-46.642.71-58.760.10
UIC87.99-3.67102.70-4.31--
Ours (w/o TTFT)-43.441.99-38.622.17-50.670.08
Ours (w/ TTFT)-56.492.68-54.962.98-70.040.14
Open-setTransTIC-43.552.28-39.602.23--
UIC118.82-5.28146.85-6.37--
Ours (w/o TTFT)-32.481.97-29.611.87--
Ours (w/ TTFT)-49.512.43-48.942.69--
+ +both performance and time overhead also increase. While time overhead increases linearly, performance gains are non-linear, with the majority of improvements occurring in the early iterations. Based on these characteristics of TTFT, our method can be effectively adapted to various compression scenarios. In low-latency applications such as CCTV surveillance systems, where real-time compression and transmission within milliseconds is required, the task-specific approach (denoted as 'w/o TTFT') is preferable. In contrast, scenarios such as satellite imagery or medical imaging—where image updates occur over the course of minutes to days—are more tolerant of encoding latency. In such cases, the instance-specific approach with hundreds + +Table 2. Ablation on the range of TTFT iterations. Experiments show that early iterations have the most impact, and TTFT iterations control time overhead and performance. + +
Bit-ratemAPΔmAP/iter.TTFT Time (sec)
w/o TTFT0.0959134.021--
w/ TTFT5 iters.0.10635.280.251.18
10 iters.0.10635.840.184.28
15 iters.0.10735.880.125.86
40 iters.0.10636.250.0614.13
+ +or more TTFT iterations may be deployed without critical latency concerns. + +Varying the range of fine-tuning. Fig. 8 shows the object detection performance with varying fine-tuning ranges. The + +![](images/1f59b09b2587a272b110faecbaedc4c4ee9e2f25e58dbfccb8ff5a3a3c962b92.jpg) + +![](images/6f8858961fbe465752013cb7664bdc7b8696edd22be2dc2f0273a2fde9b604c6.jpg) + +![](images/c2b8961c35e987f8809e3c30059d09b4eb689f420fc34bcd3687e38c130fb48d.jpg) + +![](images/7c465f005ffb6ae1d4e0cea0a64420a841fa942ede08876ab4458009b4a15d55.jpg) + +![](images/6b773737507a0728aab520f580891b8cb17367df7d340082020b2c7534cb3ffe.jpg) +Ours (open-set, w/o TTFT) + +![](images/891701f907dca521cd180e4579b7994d49151102e0e61f8444b8430f73b49857.jpg) +Ours (open-set, w/ TTFT) + +![](images/f1a945b5e0abb6bc2ef3a6d20d30bfefb2bc044a40660143528ab232080087a8.jpg) +Ours (closed-set, w/o TTFT) +Ours (closed-set, w/ TTFT) + +![](images/590fc1ebc9d6a940db25ee795122e5881b9188b27cdd95ade245c374d5901545.jpg) + +![](images/babeb2e1ccfc0505bc4cc3ae2e81598bd06e34a6895e8c00669eb7f44e07de6c.jpg) +Figure 7. Visualization of object detection results. While TIC provides better visual quality, it exhibits low detection confidence, leading to misdetections—such as misclassifying a bear as a sheep. Ours (w/ TTFT) surpasses the uncompressed image in prediction confidence despite compression losses. +Figure 8. Ablation on the range of fine-tuning. Experiments show that fine-tuning the decoder is essential for optimal performance. + +comparison includes fine-tuning either the encoder or the decoder, or both together. The experimental results show that TTFT applied to the decoder stage contributes to performance improvement. While the encoder stage, which involves a larger number of trainable parameters during TTFT, yields greater performance gains, the decoder-stage TTFT remains essential for achieving optimal performance. + +Fig. 9 presents a comparison between decoder finetuning using only the SVs and using the full set of SVD-LoRA parameters. The results show that training all SVD-LoRA parameters increases bitstream overhead by $15.3\mathrm{x}$ . Therefore, restricting training to only the SVs in the decoder is critical for enabling fully instance-specific finetuning with minimal overhead. + +![](images/a9f7bf7112fae62d831c2af251fce2897ed3e612199c4d9db4e1868ace8b9547.jpg) +Figure 9. Ablation on the range of decoder fine-tuning. Experiments show that training only the singular values of the decoder effectively reduces excessive bitstream overhead. + +# 6. Conclusion + +We propose a fully instance-specific TTFT approach for adapting learned image compression to both closed-set and open-set machine vision tasks. Our method can overcome the challenging problem of instance-specific decoder finetuning, which remains difficult due to its significant impact on both bitstream size and decoding time but is essential for effectively handling open-set tasks. The SVD-LoRA we applied enables TTFT while minimizing bitstream overhead. Moreover, it can be seamlessly integrated into any architecture, including CNNs and Transformers. Therefore, our approach can be applied to various domains where instance-specific TTFT is required. + +# 7. Acknowledgements + +This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) under multiple grants funded by the Korea government (MSIT), including the ICT Creative Consilience Program (IITP-2025-RS-2020-II201821), the Information Technology Research Center (ITRC) support program (IITP-2021-0-02052), and the Artificial Intelligence Innovation Hub (RS-2021-II212068). Additionally, this research was supported by Korea Research Institute for defense Technology planning and advancement (KRIT) grant funded by the Korean Government, Defense Acquisition Program Administration (DAPA), under Grant KRIT-CT-23-021, Development of Multi-Modal Information Fusion and High-Realization Battlefield Data Generation Technology, in 2025. + +# References + +[1] Umar Albalawi, Saraju P Mohanty, and Elias Kougianos. A hardware architecture for better portable graphics (bpg) compression encoder. In 2015 IEEE international Symposium on Nanoelectronic and information systems, pages 291-296. IEEE, 2015. 2 +[2] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 3 +[3] Dana Harry Ballard and Christopher M Brown. Computer vision. Prentice Hall Professional Technical Reference, 1982. 1 +[4] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In International Conference on Learning Representations, 2017. 2 +[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 2 +[6] Klaudia Bałazy et al. Lora-xs: Low-rank adaptation with extremely small number of parameters. arXiv preprint, 2024. 3, 4 +[7] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 2 +[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 4 +[9] Prabuddha Chakraborty, Jonathan Cruz, and Swarup Bhunia. Magic: Machine-learning-guided image compression for vision applications in internet of things. IEEE Internet of Things Journal, 8(9):7303-7315, 2020. 1 + +[10] Lahiru D Chamain, Fabien Racapé, Jean Bégaint, Akshay Pushparaja, and Simon Feltman. End-to-end optimized image compression for machines, a study. In 2021 Data Compression Conference (DCC), pages 163-172. IEEE, 2021. 1 +[11] Lahiru D Chamain, Fabien Racapé, Jean Bégaint, Akshay Pushparaja, and Simon Feltman. End-to-end optimized image compression for multiple machine tasks. arXiv preprint arXiv:2103.04178, 2021. 2, 3 +[12] Arnav Chavan et al. One-for-all: Generalized lora for parameter-efficient fine-tuning. arXiv preprint arXiv:2306.07967, 2023. 3 +[13] Yi-Hsin Chen, Ying-Chieh Weng, Chia-Hao Kao, Cheng Chien, Wei-Chen Chiu, and Wen-Hsiao Peng. Transtic: Transferring transformer-based image compression from human perception to machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23297-23307, 2023. 1, 2, 3, 5, 6 +[14] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 5, 1, 2 +[15] Hyomin Choi and Ivan V Bajic. Scalable image coding for humans and machines. IEEE Transactions on Image Processing, 31:2739-2754, 2022. 2, 3 +[16] Jack Cole and Mohamed Osman. Dataset-induced meta-learning (and other tricks): Improving model efficiency on arc, 2023. 3, 4 +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[18] Lingyu Duan, Jiaying Liu, Wenhan Yang, Tiejun Huang, and Wen Gao. Video coding for machines: A paradigm of collaborative compression and intelligent analytics. IEEE Transactions on Image Processing, 29:8680-8695, 2020. 1 +[19] Zhihao Duan, Zhan Ma, and Fengqing Zhu. Unified architecture adaptation for compressed domain semantic inference. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):4108-4121, 2023. 2, 3 +[20] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 2, 3 +[21] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3 +[22] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 3 +[23] Bjontegaard Gisle. Calculation of average psnr differences between rd curves. In ITU-T SG16/Q6, $13^{\circ}$ th VCEG Meeting, Austin, Texas, USA, April 2001, 2001. 2, 6 +[24] Sha Guo, Lin Sui, Chenlin Zhang, Zhuo Chen, Wenhan Yang, and Lingyu Duan. A unified image compression + +method for human perception and multiple vision tasks. In European Conference on Computer Vision, pages 342-359. Springer, 2024. 2, 3, 6 +[25] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3, 4 +[26] Allen Hanson. Computer vision systems. Elsevier, 1978. 1 +[27] Moritz Hardt and Yu Sun. Test-time training on nearest neighbors for large language models. In The Twelfth International Conference on Learning Representations, 2024. 3, 4 +[28] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[30] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6 +[31] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In Proceedings of the ICLR, 2022. 3 +[32] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 2, 3 +[33] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3 +[34] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Realtime multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 3, 4 +[35] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3, 4 +[36] Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien. How to exploit the transferability of learned image compression to conventional codesc. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16165-16174, 2021. 2, 3, 6 +[37] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: an end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1 + +[38] Huanyang Li and Xinfeng Zhang. Human-machine collaborative image compression method based on implicit neural representations. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2024. 1 +[39] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, 133(1):31-64, 2025. 3 +[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[41] Jiaheng Liu, Guo Lu, Zhihao Hu, and Dong Xu. A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370, 2020. 1, 6 +[42] Jinming Liu, Heming Sun, and Jiro Katto. Learning in compressed domain for faster machine vision tasks. In 2021 International Conference on Visual Communications and Image Processing (VCIP), pages 01-05. IEEE, 2021. 2, 3 +[43] Jinming Liu, Heming Sun, and Jiro Katto. Improving multiple machine vision tasks in the compressed domain. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 331-337. IEEE, 2022. 2, 3, 6 +[44] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14388-14397, 2023. 2 +[45] Kang Liu, Dong Liu, Li Li, Ning Yan, and Houqiang Li. Semantics-to-signal scalable image compression with learned revertible representations. International Journal of Computer Vision, 129(9):2605-2621, 2021. 2, 3 +[46] Ming Lu, Peiyao Guo, Huiqing Shi, Chuntong Cao, and Zhan Ma. Transformer-based image compression. In 2022 Data Compression Conference (DCC), pages 469-469. IEEE, 2022. 2, 3, 5, 6 +[47] Siwei Ma, Xinfeng Zhang, Chuanmin Jia, Zhenghui Zhao, Shiqi Wang, and Shanshe Wang. Image and video compression with neural networks: A review. IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683-1698, 2019. 2 +[48] Yixin Mei, Fan Li, Li Li, and Zhu Li. Learn a compression for objection detection-vae with a bridge. In 2021 International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2021. 2, 3 +[49] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2 +[50] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2 +[51] Vishvjit S Nalwa. A guided tour of computer vision. Addison-Wesley Longman Publishing Co., Inc., 1994. 1 +[52] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region + +proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 6 +[53] Myungseo Song, Jinyoung Choi, and Bohyung Han. Variable-rate deep image compression through spatially-adaptive feature transform. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2380-2389, 2021. 1, 2, 3 +[54] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 3 +[55] Vivienne Sze, Madhukar Budagavi, and Gary J Sullivan. High efficiency video coding (hevc). In Integrated circuit and systems, algorithms and architectures, page 40. Springer, 2014. 2 +[56] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 1 +[57] Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. Dylora: Parameter-efficient tuning of pretrained models using dynamic search-free low-rank adaptation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3274-3287, 2023. 3 +[58] Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, and Eftychios Protopapadakis. Deep learning for computer vision: A brief review. Computational intelligence and neuroscience, 2018(1):7068349, 2018. 1 +[59] Shurun Wang, Zhao Wang, Shiqi Wang, and Yan Ye. End-to-end compression towards machine vision: Network architecture design and optimization. IEEE Open Journal of Circuits and Systems, 2:675-685, 2021. 1 +[60] Shurun Wang, Zhao Wang, Shiqi Wang, and Yan Ye. Deep image compression toward machine vision: A unified optimization framework. IEEE Transactions on Circuits and Systems for Video Technology, 33(6):2979-2989, 2022. 1 +[61] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6 +[62] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. Sssic: semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 2, 3 +[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3, 4 +[64] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fin-tuning. In International Conference on Learning Representations, 2023. 3, 4 +[65] Y Zhang, M Rafie, and S Liu. Use cases and requirements for video coding for machines. ISO/IEC JTC, 1, 2021. 1 + +[66] Xia Zhao, Limin Wang, Yufei Zhang, Xuming Han, Muhammet Deveci, and Milan Parmar. A review of convolutional neural networks in computer vision. Artificial Intelligence Review, 57(4):99, 2024. 1 \ No newline at end of file diff --git a/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/images.zip b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9eac16a266164e9be79648c097dcff9f9365ddfa --- /dev/null +++ b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ef865e9397cc86743e91bcce3db3225ac8ef8c59083ce30674cc38607dafd9d +size 528207 diff --git a/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/layout.json b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38f1867a6385663bdd98d574fa47b5e31e65bbcf --- /dev/null +++ b/2025/Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability/layout.json @@ -0,0 +1,8261 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 473, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 473, + 140 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 473, + 140 + ], + "type": "text", + "content": "Test-Time Fine-Tuning of Image Compression Models for Multi-Task Adaptability" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "spans": [ + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "content": "Unki Park" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "content": " Seongmoon Jeong" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "content": " Youngchan Jang" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "content": " Gyeong-Moon Park" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "content": " Jong Hwan Ko" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "content": "Sungkyunkwan University, " + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "content": "Samsung Institute of Technology, " + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "content": "Hanwha Systems Corporation, " + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 66, + 160, + 546, + 204 + ], + "type": "text", + "content": "Korea University \n{unkipark@g., giftjsm@, hjko@}skku.edu, youngchan84.jang@hanwha.com, gm-park@korea.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 568 + ], + "type": "text", + "content": "The field of computer vision was initially inspired by the human visual system and has progressively expanded to include a broader range of machine vision applications. Consequently, image compressors should be designed to effectively accommodate not only human visual perception but also machine vision tasks, including closed-set scenarios that enable pre-training and open-set scenarios that involve previously unseen tasks at test time. Many recent studies effectively address both human visual perception and closed-set machine vision tasks simultaneously but struggle to handle open-set machine vision tasks. To address this issue, this paper proposes a fully instance-specific test time fine-tuning (TTFT) for adapting learned image compression (LIC) to both closed-set and open-set machine vision tasks effectively. With our method, a large-scale LIC model, originally trained for human perception, is adapted to the target task through TTFT using Singular Value Decomposition based Low Rank Adaptation (SVD-LoRA). During TTFT, the decoder adopts a modified learning scheme that focuses exclusively on training the singular values, which helps prevent excessive bitstream overhead. This enables fully instance-specific optimization for the target task, even for open-set tasks. Experimental results demonstrate that the proposed method effectively adapts the backbone compressor to diverse machine vision tasks, outperforming competing methods. The code is available at project page." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 590, + 135, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 135, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 135, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "content": "In the early stages of computer vision, image processing technologies were primarily developed to align with human perception, as images were largely intended for human interpretation [3, 26, 51]. However, with advances in deep learning and computer vision techniques enabling more sophisticated data processing, machines have increasingly gained the ability to interpret visual data. Conse" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 314, + 228, + 550, + 392 + ], + "blocks": [ + { + "bbox": [ + 314, + 228, + 550, + 392 + ], + "lines": [ + { + "bbox": [ + 314, + 228, + 550, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 550, + 392 + ], + "type": "image", + "image_path": "5c316734b66866dee51e1626811074b3748e8d1c3c92146f9f0118e57b7e2012.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "text", + "content": "Figure 1. An overview of our proposed method. Our method tunes both the encoder and decoder of a backbone human-target LIC model by fine-tuning SVD-LoRA weights optimized for each target task and image instance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 470, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 470, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 312, + 470, + 555, + 542 + ], + "type": "text", + "content": "sequently, the scope of computer vision has broadened to include machine vision applications [56, 58, 66]. In the field of image compression as well, there is a growing need for methods tailored to the requirements of machine vision, as machines are expected to become the primary users of compressed image data [18, 65]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 628 + ], + "type": "text", + "content": "Accordingly, several research studies have shifted their focus away from the traditional goal of visual signal reconstruction, instead prioritizing the enhancement of machine vision task performance [9, 59]. These studies commonly aim to effectively restore representations suitable for the targeted machine vision domain, while selectively sacrificing visual quality to achieve this goal." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Taking it further, recent research has focused on developing learned image codes (LICs) that can cater to both human perception and various machine vision tasks [10, 13, 37, 38, 41, 53, 60]. These studies can be categorized into methods that generate either multi-task bitstreams or single-task bitstreams. The multi-task bitstream approach aims to generate a unified bitstream that can support vari-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 713 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4430" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "ous subtasks including human perception and multiple machine tasks. This method offers flexibility in responding to a range of machine vision tasks, but it introduces challenges. For example, scalable coding [15, 32, 45, 62] increases the processing complexity and bitstream overhead for individual downstream tasks, and unified image coding [11, 19, 20, 24] struggles with optimizing each subtask individually." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 295, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 352 + ], + "type": "text", + "content": "In contrast, the single-task bitstream approach [13, 36, 42, 43, 48, 53] avoids the limitations of multi-task methods by dynamically generating bitstreams for a specific downstream task. For instance, TransTIC[13] utilizes prompt tuning for adapting the human-target backbone codec to generate bitstreams optimized for the other tasks specified by the prompt. While this technique enables switching of pre-defined (closed-set) tasks without modifying the backbone codec, it requires task-specific pre-training of the prompt generator. Therefore, it leads to increased training demands as the number of tasks grows, and faces challenges to handle new, untrained (open-set) tasks. Moreover, performance improvement is non-monotonic with respect to prompt size, and computational costs increase due to the added prompt during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 354, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 714 + ], + "type": "text", + "content": "To address these limitations, we aim to develop a method that can dynamically adapt a given backbone codec to optimize rate-performance trade-offs for the specific requirements of each task. Our goal is to provide flexible support for diverse tasks and codec models, allowing adaptation to previously unseen (open-set) tasks without requiring pre-training, and also adaptable to any given codec model. To achieve this, as shown in Fig. 1, we propose tuning the backbone codec using Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning approach widely used in test-time scenarios. Unlike the TransTIC approach, which cannot perform instance-specific tuning on the decoder, our method tunes both the encoder and decoder by generating LoRA weights optimized for each target task and image instance. This enables dynamic codec optimization for any arbitrary task without requiring pre-training, making it effective for open-set scenarios. While instance-specific model optimization can lead to superior performance compared to task-specific optimization, it introduces additional overhead by requiring a separate set of parameters for each image. Specifically, in the case of instance-specific decoder tuning, transmitting the LoRA weights incurs additional bit rate overhead. To address this efficiently, we adopt an SVD-LoRA approach that combines the original LoRA weights with supplementary singular value (SV) components. After pre-initializing the original LoRA weights, we can then transmit only the SV components to the decoder on an instance-specific basis, allowing effective tuning while minimizing bit rate demands. As the SV component captures essential information for optimizing the codec per target in" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "stance and task, it is both effective and efficient in managing rate requirements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 96, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 251 + ], + "type": "text", + "content": "Experimental results demonstrate that our method achieves superior adaptation performance for both open-set and closed-set tasks. Although conducted under a lightweight experimental setup (40 iterations, LoRA rank 4), we observed that our method achieved more than a " + }, + { + "bbox": [ + 313, + 96, + 553, + 251 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 96, + 553, + 251 + ], + "type": "text", + "content": " improvement in Bjøntegaard Delta (BD)-Rate [23] compared to the backbone codec across all evaluated machine vision tasks. These results surpass all competing methods. Notably, our method achieves superior performance in a pure open-set setting—where no prior task-specific setup is available—outperforming competing methods even when they are evaluated under pre-initialized (closed-set) conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "type": "text", + "content": "The main contributions of our work can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 276, + 553, + 502 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 276, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 276, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 553, + 347 + ], + "type": "text", + "content": "- Test-time fine-tuning in compression field : To the best of our knowledge, we are the first to attempt TTFT at the decoder stage in a image compression task. Considering that the encoder and decoder are temporally and spatially separated in compression systems, TTFT has the potential to maximize image compression quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 348, + 553, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 348, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 315, + 348, + 553, + 442 + ], + "type": "text", + "content": "- Novel approach utilizing LoRA: We propose a novel approach utilizing LoRA for efficient adaptation to diverse machine vision applications in the still-image compression task. Specifically, we introduce a modified LoRA in the form of SVD to enable instance-specific optimization at the decoder. This approach not only enhances efficiency but also facilitates effective TTFT at the decoder stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 443, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 553, + 502 + ], + "type": "text", + "content": "- Handling open-set tasks: By designing an instance-specific optimized model that does not necessarily require pre-training, we have enhanced the ability to tackle open-set machine vision tasks, which have remained challenging in previous research." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 514, + 402, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 402, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 402, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 715 + ], + "type": "text", + "content": "The initial development of LIC focused on optimizing perceptual quality for human visual systems, similar to traditional compressors [1, 7, 55]. Most LIC models adopt an autoencoder structure [4, 5, 44, 47, 50], aiming to optimize the balance between distortion and bit-rate. In early LIC research, the goal was to minimize losses limited to human visual perception, with the technology evolving towards faithfully mimicking human visual quality [28, 46, 49]. With the rapid development of machine vision tasks such as image classification and object detection, an increasing number of applications now incorporate images encoded and decoded by image codecs for these downstream tasks. Consequently, LIC research has begun to focus on meeting machine vision requirements. Initially, the emphasis was on creating" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4431" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "individual models tailored to each specific task. More recently, research has proposed approaches where a shared LIC model architecture is individually trained to generate task-specific models. Following this trend, research efforts aimed at designing a single LIC model to support multiple targets including human perception and diverse machine vision tasks. These studies can be divided into multi-task bitstream and single-task bitstream methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 294, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 294, + 291 + ], + "type": "text", + "content": "Multi-task Bitstream LICs for Multiple Targets. One line of research in multi-target LIC is the multi-task bitstream approach [11, 15, 19, 20, 24, 32, 45, 62], which aims to accommodate multiple targets using a single bitstream. However, this approach increases the complexity of both the encoder and decoder [15, 24]. It also leads to significant bitstream overhead from the perspective of individual downstream tasks [15]. Additionally, it faces the challenge of Pareto optimality, which makes it difficult to optimize each subtask individually [24]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 294, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 715 + ], + "type": "text", + "content": "Adaptive Single-task Bitstream LICs for Multiple Targets. Another line of study for multi-target LIC is the single-task bitstream approach [13, 36, 42, 43, 48, 53]. This approach generates a bitstream specifically for one target task at a time, adapting to the given target task to support multi-task requirements effectively. This allows the approach to achieve optimal performance under well-defined conditions for each specific subtask. Representative studies in this field involve applying task-adaptive importance modulation or fine-tuning techniques to the backbone LIC. Song et al. [53] proposed a spatially-adaptive image compression method using Spatial Feature Transform (SFT) guided by a real-valued quality map that enables pixel-wise optimization for machine vision tasks. Liu et al. [43] introduced a task-adaptive image compression method that uses a gating module to selectively activate channel-wise latents from a pre-trained codec. J.P. Klopp et al. [36] introduced a learned pre-filter that suppresses task-irrelevant data prior to compression by distinguishing between important and unimportant information. TransTIC [13] employed Visual Prompt Tuning (VPT) [21, 33] to adapt the largescale pretrained model TIC [46], originally trained for human perception, to suit various target tasks. This approach demonstrated adaptability to various machine vision tasks such as object detection, instance segmentation, and classification [13]. However, VPT itself faces limitations such as non-monotonic performance changes depending on the size of the prompt applied, as well as increased computational cost due to the additional prompts during the inference stage [12]. More importantly, TransTIC adjusts the prompt to guide the output of a static pretrained model, but the generator responsible for creating this prompt also needs to be pretrained. Also, it requires prompts equivalent to one-fourth of the input feature size of each layer to perform task switching. Therefore, for applying fine-tuning to" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "individual test samples at the decoder stage, these prompts must be incorporated into the bitstream, making TransTIC unsuitable for instance-specific decoder fine-tuning. Consequently, pre-training is essential to adapt to new tasks, and it is difficult to adapt to untrained open-set tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "content": "To overcome these limitations, we propose using TTFT to dynamically adapt a backbone LIC model for diverse tasks. Since TTFT adjusts part or all of the model through a brief learning phase during testing [16, 27, 39, 54], it effectively supports open set scenarios where the codec needs to be optimized for previously unseen tasks. Although TTFT is highly effective in producing results optimized for individual samples, it has the limitation of requiring additional training steps at each testing instance. Therefore, we utilize LoRA [31], one of the parameter-efficient fine-tuning (PEFT) techniques to reduce training costs while effectively adapting the model to downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 288, + 391, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 391, + 302 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 391, + 302 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "text", + "content": "The LoRA method is one of the most actively researched PEFT techniques as it provides effective task adaptation with only a small number of parameters without increasing model complexity and inference time [6, 31, 57]. It focuses on learning the intrinsic rank of the weights [31]. Let the pre-trained weights be denoted by " + }, + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "inline_equation", + "content": "W_{o} \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "text", + "content": ", and let the weight change required for task adaptation be denoted by " + }, + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "text", + "content": ". After the task adaptation, the new weight becomes " + }, + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "inline_equation", + "content": "W_{o} + \\Delta W" + }, + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "text", + "content": ". Typically, the intrinsic rank of the weight change required for task adaptation is relatively small compared to the size of the entire weight matrix, meaning that " + }, + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 313, + 308, + 555, + 465 + ], + "type": "text", + "content": " can be decomposed into a low-rank representation [31]. From this, the relationship in Eq. (1) can be obtained:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 365, + 478, + 553, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 478, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 365, + 478, + 553, + 491 + ], + "type": "interline_equation", + "content": "W = W _ {o} + \\Delta W = W _ {o} + B \\cdot A, \\tag {1}", + "image_path": "5d3643416dd0b8ff42e8ef6572d4e9c12f4a9bb71f3409ee23e0516c3cc86bb4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d, k)" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "W_{o}" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " remains fixed, " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "B \\cdot A" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " can be explicitly separated in " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": ", allowing us to store only the matrices " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " for task adaptation, which requires very minimal memory overhead. During inference, a reparameterization process is performed in which " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "B \\cdot A" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " is added to " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "W_{o}" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " to reconstruct " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": ", thereby avoiding any additional inference latency compared to the original model. The LoRA technique is applicable to both CNN and Transformer architectures and can be applied to all layers as long as the memory overhead issue is manageable. Recently, this technique has been reformulated as an SVD-based form and applied to LLMs [6, 25, 64]. These studies [6, 64] aim to fine-tune large pre-trained language models for downstream tasks in a parameter-efficient manner. Specifically, Zhang et al. [64] utilizes the SVD shaped LoRA to allocate parameter budgets adaptively based on the importance scores of each weight matrix. Meanwhile, Balazy et al. [6] proposed a method that applies SVD" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4432" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 265, + 205 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 265, + 205 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 265, + 205 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 265, + 205 + ], + "type": "image", + "image_path": "e2488f48a4f9e6a612b8f0a4658ed13e65450cbeacf1bd620f4035653ce46bf6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 215, + 297, + 249 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 297, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 297, + 249 + ], + "type": "text", + "content": "Figure 2. Original LoRA vs. SVD-LoRA. When the rank is small, SVD-LoRA more effectively captures the intrinsic dimensions of the adaptation weights." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 319 + ], + "type": "text", + "content": "shaped LoRA to fine-tune only the SV weight matrices during task switching. These novel approaches in the LLM domain show promising potential for effective application in the field of image compression as well." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 329, + 138, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 138, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 138, + 342 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 350, + 181, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 181, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 181, + 361 + ], + "type": "text", + "content": "4.1. Framework Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 367, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 618 + ], + "type": "text", + "content": "Our goal is to adapt the given backbone codec for various downstream tasks, including open-set tasks that are unseen during training. To achieve this, we adapt the backbone LIC to target downstream tasks by applying TTFT [16, 27] on both the encoder and decoder. Specifically, we use LoRA, a widely used approach in test-time scenarios. By tuning the LoRA weights to be optimized for each target task and image instance, the backbone LIC can be adapted to better suit the target task. However, transmitting LoRA weights to the decoder for fully instance-specific tuning incurs a nonnegligible bit rate overhead. To minimize the bit rate issue, we employ a reformulated version of LoRA, referred to as SVD-LoRA, which enables effective fine-tuning of only a small portion of the decoder at test time. While prior works utilized SVD-LoRA for language task-specific adaptation [6, 25, 64], we focus on the test-time decoder fine-tuning for a fully instance-specific optimized compression model. Consequently, the proposed method can flexibly handle diverse, arbitrary target tasks and is compatible with various backbone LIC architectures, including both CNN and Transformer-based models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 627, + 161, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 161, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 161, + 638 + ], + "type": "text", + "content": "4.2. SVD based LoRA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 644, + 210, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 210, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 210, + 657 + ], + "type": "text", + "content": "SVD-LoRA is represented by Eq. (2):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 666, + 295, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 666, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 96, + 666, + 295, + 679 + ], + "type": "interline_equation", + "content": "W = W _ {o} + \\Delta W = W _ {o} + U \\cdot \\Sigma \\cdot V ^ {T}, \\tag {2}", + "image_path": "31ca5495fd95c585b6d43a88a1e5dd9661ad7171726f603c885650c192e84b8d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{r \\times r}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{k \\times r}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d, k)" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " learn the intrinsic ranks for" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 339, + 69, + 537, + 203 + ], + "blocks": [ + { + "bbox": [ + 339, + 69, + 537, + 203 + ], + "lines": [ + { + "bbox": [ + 339, + 69, + 537, + 203 + ], + "spans": [ + { + "bbox": [ + 339, + 69, + 537, + 203 + ], + "type": "image", + "image_path": "8e8bfeb8b3aea92d9b9b0ce4ccbc0dde7f60310fafe0e0fa36bea95282160c32.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 554, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 269 + ], + "type": "text", + "content": "Figure 3. Bitstream generation process. For instance-specific decoder tuning with manageable bitstream overhead, only the singular values of the decoder side are encoded and added to the image-encoded bitstream. The bit-rate overhead accounts for less than " + }, + { + "bbox": [ + 313, + 213, + 554, + 269 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 213, + 554, + 269 + ], + "type": "text", + "content": " of the image bitstream size." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 290, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 553, + 313 + ], + "type": "text", + "content": "columns and rows, respectively, while " + }, + { + "bbox": [ + 313, + 290, + 553, + 313 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 290, + 553, + 313 + ], + "type": "text", + "content": " is a diagonal matrix that learns the SVs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 315, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 555 + ], + "type": "text", + "content": "Although the structure includes additional SVs compared to conventional LoRA, the increase in the number of parameters is negligible due to the typically small rank used. Fig. 2 illustrates the process of reconstructing " + }, + { + "bbox": [ + 313, + 315, + 555, + 555 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 315, + 555, + 555 + ], + "type": "text", + "content": " (adapted weights) through the reparameterization of LoRA and SVD-LoRA, respectively. By learning additional SVs, SVD-LoRA facilitates the learning of more complex intrinsic dimensions of adaptation weights compared to the original LoRA. In cases where both LoRA and SVD-LoRA have the same small rank, experimental results indicate that SVD-LoRA better captures the intrinsic dimensions of the adaptation weights. Details of this observation are discussed in the supplementary material. Fig. 3 illustrates the process of embedding the decoder-side SVD-LoRA weights into the compressed bitstream for instance-specific fin-tuning at the decoder stage. While the bitstream size increases due to the additional SV component, the overhead is negligible. Specifically, with a rank of 4, as specified in our experimental settings, the additional overhead accounts for less than " + }, + { + "bbox": [ + 313, + 315, + 555, + 555 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 315, + 555, + 555 + ], + "type": "text", + "content": " of the compressed image bitstream size." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 563, + 524, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 524, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 524, + 576 + ], + "type": "text", + "content": "4.3. Test-Time Fine Tuning with SVD-LoRA" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Fig. 4 illustrates the TTFT procedure utilizing SVD-LoRA, which consists of two stages. In the first stage, the initial values of SVD-LoRA are set differently depending on the target task type. For closed-set tasks, the initial values are set through training the specific target task. In contrast, open-set tasks leverage initialization learned from common tasks like classification, which allows the model to generalize across various machine vision applications. The next stage involves performing TTFT to optimize each test sample according to the target downstream task. During this process, all of the SVD-LoRA weights in the encoder and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4433" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 261 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 261 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 261 + ], + "type": "image", + "image_path": "c9b3def41b67a2cf0616a6c891e58b1bf1518019fbbdfce1eb6656da7b9e8b26.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 296, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 327 + ], + "type": "text", + "content": "Figure 4. TTFT procedure. The TTFT operates in two stages. First, the initial values for SVD-LoRA are initialized. In the second stage, TTFT is applied at test time to fine-tune each test sample in accordance with the specific downstream task. The source images of (b) is [56]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 347, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 295, + 539 + ], + "type": "text", + "content": "only the SVs in the decoder are fine-tuned. While full finetuning of the encoder does not affect bit rate overhead or decoding latency, it requires a large number of parameters for each TTFT, leading to increased TTFT time overhead. Therefore, in our study, we apply SVD-LoRA to the encoder instead of performing full fine-tuning. As a result, this approach significantly reduces training costs and enables rapid adaptation to new downstream tasks. Additionally, fine-tuning only the SVs on the decoder side further reduces bitstream overhead compared to full LoRA finetuning. However, increasing the number of iterations during TTFT improves performance, but also increases encoding time. This issue can be addressed by adjusting the number of iterations according to the compression scenario, thereby balancing performance improvement with encoding time. (Detailed information is covered in the ablation study)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 539, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 564 + ], + "type": "text", + "content": "The loss function of SVD-LoRA during TTFT is defined by Eq. (3):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 575, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 575, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 89, + 575, + 295, + 590 + ], + "type": "interline_equation", + "content": "L = - \\log p (\\hat {z}) - \\log p (\\hat {y} \\mid \\hat {z}) + \\lambda d (x, \\hat {x}), \\tag {3}", + "image_path": "cc38c50180f95dec6c46acf4bdadb29d06430caf18d1c56e15a08ba9fe6e7ae5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " represent the input image, encoded latent feature, and hyper prior, respectively. Additionally, " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " represent their reconstruction. The terms " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "p(\\hat{z})" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "p(\\hat{y} \\mid \\hat{z})" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " represent the probabilities of the compressed latent variables, which correspond to the bit rate used for compressing the input " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "d(x, \\hat{x})" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " represents the distortion between the original and the reconstructed images. Mean square error (MSE) is used as the distortion metric when targeting human perception, as it measures the pixel-wise difference between the original and reconstructed images." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 69, + 557, + 241 + ], + "blocks": [ + { + "bbox": [ + 318, + 69, + 557, + 241 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 557, + 241 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 557, + 241 + ], + "type": "image", + "image_path": "8e2136279f63654bc5f27db8caa761b5026b633439703f9b376a22fbd20e31f2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 246, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 302 + ], + "type": "text", + "content": "Figure 5. An example of applying SVD-LoRA to the TIC codec. While SVD-LoRA is compatible with both CNN and STB layers, our work focuses solely on the STB layers of the main encoder and decoder to enhance task-switching efficiency with minimal tuning parameters." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 325, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 456 + ], + "type": "text", + "content": "For machine vision tasks such as image classification, perceptual loss is used as the distortion metric by comparing intermediate features extracted from the original and reconstructed images via the target network [13]. The hyper parameter " + }, + { + "bbox": [ + 313, + 325, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 325, + 555, + 456 + ], + "type": "text", + "content": " adjusts the balance between rate and distortion. This loss function allows for optimization that balances data compression and reconstruction while tailoring to the objective of each task. All experiments conducted in our work are formulated as rate-distortion optimization, with different values of " + }, + { + "bbox": [ + 313, + 325, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 325, + 555, + 456 + ], + "type": "text", + "content": " and distortion functions designed according to the characteristics of each downstream task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 470, + 440, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 440, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 440, + 484 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 491, + 441, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 441, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 441, + 503 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 509, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 581 + ], + "type": "text", + "content": "While our method is compatible with any baseline architecture, we demonstrate its effectiveness using a widely adopted state-of-the-art (SOTA) LIC model—the Transformer-based LIC (TIC) network proposed in [46]—as the backbone. Results for the CNN-based LIC model [14] are included in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Fig. 5 shows the application of SVD-LoRA to the TIC model. The TIC adopts an autoencoder structure consisting of a main encoder-decoder pair and a hyper encoder-decoder pair, aiming to optimize the balance between distortion and bitrate for human perception. In our work, to maximize task-switching efficiency with minimal tuning parameters, we apply SVD-LoRA exclusively to the main encoder and decoder, as these components have the most significant impact on task adaptation [13]. While SVD-LoRA is compatible with both CNNs and Transformers, our approach specifically targets the Swin Transformer Block" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4434" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "(STB) in the main encoder and decoder to achieve optimal performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 295, + 133 + ], + "type": "text", + "content": "Datasets. We used MS-COCO [40] for object detection and instance segmentation, and ImageNet-1K [17] for image classification in our experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 206 + ], + "type": "text", + "content": "Pre-trained Vision Models. To evaluate performance across various downstream tasks, we employed pre-trained vision models [29, 30, 46, 52] as task-specific evaluators. Specifically, we used ResNet50 [29] for image classification, Faster R-CNN [52] for object detection, and Mask R-CNN [30] for instance segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 295, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 229 + ], + "type": "text", + "content": "Compared Methods. The methods compared in our work are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 495 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 232, + 295, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 255 + ], + "type": "text", + "content": "(1) TIC [46]: TIC is a Transformer-based compression model pre-trained on the Flickr2W dataset [41]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 256, + 295, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 295, + 290 + ], + "type": "text", + "content": "(2) Full Fine-Tuning: Full Fine-Tuning updates all parameters of the backbone TIC to adapt to the target machine vision tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 292, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 295, + 327 + ], + "type": "text", + "content": "(3) TransTIC [13]: TransTIC applies VPT to adapt the backbone TIC—originally trained for human perception—to target machine vision tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 363 + ], + "type": "text", + "content": "(4) Unified Image Compression (UIC) [24]: UIC uses a collaborative embedding with a diffusion model to generate a universal bitstream to handle open-set tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 363, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 295, + 495 + ], + "type": "text", + "content": "(5) Other Adaptation Approaches: These approaches apply adaptation techniques to the backbone TIC: (5-1) TIC+SFT [61] integrates Spatial Feature Transform (SFT) layers into TIC, where the SFT layers act as adapters for task-specific modulation; (5-2) TIC+channel_selection [43] selects partial channels from the image latent representation produced by TIC and converts them into task-specific feature maps to enable efficient task transfer; (5-3) TIC+prefilter [36] involves adding a pre-neural network before TIC to assist in transitioning TIC to the target task." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 496, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 639 + ], + "type": "text", + "content": "Task Types and TTFT settings. Experiments were conducted under both closed-set and open-set conditions. We used a classification model to extract generalized features for open-set initialization. Therefore, the open-set image classification task is excluded, since its initialization is derived from its own target model. In all experiments, 'w/o TTFT' refers to the task-specific approach using fixed SVD-LoRA weights, while 'w/ TTFT' denotes the instance-specific approach with TTFT applied per test sample. TTFT is performed for 40 iterations per sample in our experiments; however, the number of iterations is adjustable and its impact on performance is analyzed in the ablation study." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 640, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 295, + 700 + ], + "type": "text", + "content": "Evaluation Metrics. Top-1 accuracy is used to evaluate image classification performance, whereas mean average precision (mAP) is employed for both object detection and instance segmentation. For numerical comparisons, the BDA-rate metric [23] is used." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Strategy for handling different quality levels. In all ex" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "periments, the four quality levels of TIC (from quality level 1 to 4) are used to visually represent the distribution between the rate and evaluation metrics. The " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " values of loss function are set to 2, 1, 0.5, and 0.2, depending on the quality level [13]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 140, + 553, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 553, + 166 + ], + "type": "text", + "content": "5.2. Evaluation for Multiple Machine Vision Tasks and Human perception" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 170, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 373 + ], + "type": "text", + "content": "Fig. 6 compares our proposed approach with other methods across various machine vision tasks. The results demonstrate that our method, with TTFT applied, outperforms all competing approaches. Table 1 presents a performance comparison using BD-Rate, BD-mAP, and BD-accuracy with TIC as the anchor. The results show that our method consistently outperforms across all tasks, not only under closed-set conditions but also in challenging open-set settings. While the task-specific approach (denoted as 'w/o TTFT') achieves performance comparable to other competing methods, applying TTFT further improves performance. Even under challenging open-set settings, TTFT demonstrates that the model remains robust. Notably, our method achieves higher performance under open-set conditions than TransTIC, one of the SOTA techniques, does under closed-set conditions, demonstrating its superior generalization capability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 381, + 425, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 425, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 425, + 395 + ], + "type": "text", + "content": "5.3. Qualitative Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 399, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 604 + ], + "type": "text", + "content": "Fig. 7 visualizes the results of object detection. Although TIC yields high visual quality from the perspective of human perception, its detection confidence scores are relatively low, limiting its effectiveness for downstream tasks. Furthermore, in all experiments, the visualized results exhibit noticeable differences between closed-set and open-set conditions. Since closed-set tasks are pre-trained on the target object detection task, high prediction scores can be achieved regardless of whether TTFT is applied. The corresponding results also tend to preserve fine texture details. In contrast, the open-set setting relies on generalized features that are not optimized for the target task, leading to reduced prediction accuracy and blurred textures. Nonetheless, in both settings, TTFT improves performance by enhancing texture detail preservation and enabling higher prediction scores. Notably, TTFT also maintains model robustness under challenging open-set conditions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 611, + 415, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 415, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 415, + 623 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 629, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 676 + ], + "type": "text", + "content": "The ablation studies are conducted under the conditions of the open-set object detection task, utilizing an A100 GPU with 80GB memory. The evaluation is performed using 5,000 validation images from the MS-COCO dataset." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Adjustment of fine-tuning iterations. Table 2 shows the performance and time overhead with varying numbers of TTFT iterations. As the number of iterations increases," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4435" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 65, + 244, + 178 + ], + "blocks": [ + { + "bbox": [ + 136, + 65, + 244, + 178 + ], + "lines": [ + { + "bbox": [ + 136, + 65, + 244, + 178 + ], + "spans": [ + { + "bbox": [ + 136, + 65, + 244, + 178 + ], + "type": "image", + "image_path": "c6103282c531a3891da19f3e7adb8371bfbd82b136ad05277587f0ef28a036bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 250, + 65, + 359, + 178 + ], + "blocks": [ + { + "bbox": [ + 250, + 65, + 359, + 178 + ], + "lines": [ + { + "bbox": [ + 250, + 65, + 359, + 178 + ], + "spans": [ + { + "bbox": [ + 250, + 65, + 359, + 178 + ], + "type": "image", + "image_path": "034b463044f6315d7f515c348e2dbbdd8ca16b8edafb49b8ff41e080bde5156d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 380, + 78, + 462, + 167 + ], + "blocks": [ + { + "bbox": [ + 380, + 78, + 462, + 167 + ], + "lines": [ + { + "bbox": [ + 380, + 78, + 462, + 167 + ], + "spans": [ + { + "bbox": [ + 380, + 78, + 462, + 167 + ], + "type": "image", + "image_path": "31273c596a8010511c737040a12136c90c60753b11bd8f00a777f0884365d93b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 137, + 201, + 246, + 314 + ], + "blocks": [ + { + "bbox": [ + 159, + 183, + 238, + 200 + ], + "lines": [ + { + "bbox": [ + 159, + 183, + 238, + 200 + ], + "spans": [ + { + "bbox": [ + 159, + 183, + 238, + 200 + ], + "type": "text", + "content": "(a) Instance Segmentation (open-set)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 201, + 246, + 314 + ], + "lines": [ + { + "bbox": [ + 137, + 201, + 246, + 314 + ], + "spans": [ + { + "bbox": [ + 137, + 201, + 246, + 314 + ], + "type": "image", + "image_path": "d0ee42624a85c88be03f936b97c02b92a11f505d1672709e703e118770e71361.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 319, + 238, + 335 + ], + "lines": [ + { + "bbox": [ + 159, + 319, + 238, + 335 + ], + "spans": [ + { + "bbox": [ + 159, + 319, + 238, + 335 + ], + "type": "text", + "content": "(c) Instance Segmentation (closed-set)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 348, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 55, + 348, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 555, + 371 + ], + "type": "text", + "content": "Figure 6. Rate-accuracy performance comparison under various machine tasks. Experimental results show that our method outperforms the SOTA approaches in various closed-set and open-set target tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 249, + 201, + 358, + 314 + ], + "blocks": [ + { + "bbox": [ + 281, + 183, + 342, + 200 + ], + "lines": [ + { + "bbox": [ + 281, + 183, + 342, + 200 + ], + "spans": [ + { + "bbox": [ + 281, + 183, + 342, + 200 + ], + "type": "text", + "content": "(b) Object Detection (open-set)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 201, + 358, + 314 + ], + "lines": [ + { + "bbox": [ + 249, + 201, + 358, + 314 + ], + "spans": [ + { + "bbox": [ + 249, + 201, + 358, + 314 + ], + "type": "image", + "image_path": "4d41a058e543a4977b630d1ca3387de4685819d95b66e9c993a2af899661f31d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 319, + 342, + 335 + ], + "lines": [ + { + "bbox": [ + 281, + 319, + 342, + 335 + ], + "spans": [ + { + "bbox": [ + 281, + 319, + 342, + 335 + ], + "type": "text", + "content": "(d) Object Detection (closed-set)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 362, + 201, + 472, + 314 + ], + "blocks": [ + { + "bbox": [ + 362, + 201, + 472, + 314 + ], + "lines": [ + { + "bbox": [ + 362, + 201, + 472, + 314 + ], + "spans": [ + { + "bbox": [ + 362, + 201, + 472, + 314 + ], + "type": "image", + "image_path": "008c7b8c76da68d8257fd1d2802947560c32ffdc025f8f73c88254baaee48181.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 319, + 451, + 335 + ], + "lines": [ + { + "bbox": [ + 402, + 319, + 451, + 335 + ], + "spans": [ + { + "bbox": [ + 402, + 319, + 451, + 335 + ], + "type": "text", + "content": "(e) Classification (closed-set)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 105, + 412, + 501, + 548 + ], + "blocks": [ + { + "bbox": [ + 55, + 382, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 382, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 555, + 404 + ], + "type": "text", + "content": "Table 1. Comparison of BD-Rate, BD-mAP, and BD-Accuracy using TIC as the anchor. Bold numbers indicate the best performance, while blue numbers represent the second-best performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 412, + 501, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 501, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 501, + 548 + ], + "type": "table", + "html": "
Task TypeMethodInstance SegmentationObject DetectionClassification
BD-Rate (%) ↓BD-mAP ↑BD-Rate (%) ↓BD-mAP ↑BD-Rate (%) ↓BD-accuracy ↑
Closed-setFull fine tuning-75.554.00-66.004.10-1000.16
TIC+prefilter-24.741.56-17.591.19-23.690.04
TIC+SFT-40.241.67-27.241.45-61.530.11
TIC+channel_selection47.58-2.5552.85-3.77-30.370.06
TransTIC-47.452.34-46.642.71-58.760.10
UIC87.99-3.67102.70-4.31--
Ours (w/o TTFT)-43.441.99-38.622.17-50.670.08
Ours (w/ TTFT)-56.492.68-54.962.98-70.040.14
Open-setTransTIC-43.552.28-39.602.23--
UIC118.82-5.28146.85-6.37--
Ours (w/o TTFT)-32.481.97-29.611.87--
Ours (w/ TTFT)-49.512.43-48.942.69--
", + "image_path": "44f78db64dc3d0db34773ca2f2aefd4f38055c31d6dcdb0a5bff5d0e1eb8a1f0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 557, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 297, + 713 + ], + "type": "text", + "content": "both performance and time overhead also increase. While time overhead increases linearly, performance gains are non-linear, with the majority of improvements occurring in the early iterations. Based on these characteristics of TTFT, our method can be effectively adapted to various compression scenarios. In low-latency applications such as CCTV surveillance systems, where real-time compression and transmission within milliseconds is required, the task-specific approach (denoted as 'w/o TTFT') is preferable. In contrast, scenarios such as satellite imagery or medical imaging—where image updates occur over the course of minutes to days—are more tolerant of encoding latency. In such cases, the instance-specific approach with hundreds" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 326, + 597, + 540, + 659 + ], + "blocks": [ + { + "bbox": [ + 313, + 556, + 555, + 589 + ], + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 589 + ], + "type": "text", + "content": "Table 2. Ablation on the range of TTFT iterations. Experiments show that early iterations have the most impact, and TTFT iterations control time overhead and performance." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 597, + 540, + 659 + ], + "lines": [ + { + "bbox": [ + 326, + 597, + 540, + 659 + ], + "spans": [ + { + "bbox": [ + 326, + 597, + 540, + 659 + ], + "type": "table", + "html": "
Bit-ratemAPΔmAP/iter.TTFT Time (sec)
w/o TTFT0.0959134.021--
w/ TTFT5 iters.0.10635.280.251.18
10 iters.0.10635.840.184.28
15 iters.0.10735.880.125.86
40 iters.0.10636.250.0614.13
", + "image_path": "c36917b8e9454e9448939e77e30e7e96556e232513b9af93044bc66853d4ae67.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 663, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 555, + 687 + ], + "type": "text", + "content": "or more TTFT iterations may be deployed without critical latency concerns." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Varying the range of fine-tuning. Fig. 8 shows the object detection performance with varying fine-tuning ranges. The" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4436" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 72, + 208, + 190 + ], + "blocks": [ + { + "bbox": [ + 111, + 72, + 208, + 190 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 208, + 190 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 208, + 190 + ], + "type": "image", + "image_path": "1f59b09b2587a272b110faecbaedc4c4ee9e2f25e58dbfccb8ff5a3a3c962b92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 209, + 72, + 304, + 189 + ], + "blocks": [ + { + "bbox": [ + 209, + 72, + 304, + 189 + ], + "lines": [ + { + "bbox": [ + 209, + 72, + 304, + 189 + ], + "spans": [ + { + "bbox": [ + 209, + 72, + 304, + 189 + ], + "type": "image", + "image_path": "6f8858961fbe465752013cb7664bdc7b8696edd22be2dc2f0273a2fde9b604c6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 402, + 189 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 402, + 189 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 402, + 189 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 402, + 189 + ], + "type": "image", + "image_path": "c2b8961c35e987f8809e3c30059d09b4eb689f420fc34bcd3687e38c130fb48d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 402, + 72, + 498, + 189 + ], + "blocks": [ + { + "bbox": [ + 402, + 72, + 498, + 189 + ], + "lines": [ + { + "bbox": [ + 402, + 72, + 498, + 189 + ], + "spans": [ + { + "bbox": [ + 402, + 72, + 498, + 189 + ], + "type": "image", + "image_path": "7c465f005ffb6ae1d4e0cea0a64420a841fa942ede08876ab4458009b4a15d55.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 112, + 207, + 207, + 310 + ], + "blocks": [ + { + "bbox": [ + 113, + 196, + 206, + 206 + ], + "lines": [ + { + "bbox": [ + 113, + 196, + 206, + 206 + ], + "spans": [ + { + "bbox": [ + 113, + 196, + 206, + 206 + ], + "type": "text", + "content": "Ours (open-set, w/o TTFT)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 112, + 207, + 207, + 310 + ], + "lines": [ + { + "bbox": [ + 112, + 207, + 207, + 310 + ], + "spans": [ + { + "bbox": [ + 112, + 207, + 207, + 310 + ], + "type": "image", + "image_path": "6b773737507a0728aab520f580891b8cb17367df7d340082020b2c7534cb3ffe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 209, + 207, + 304, + 310 + ], + "blocks": [ + { + "bbox": [ + 209, + 196, + 304, + 205 + ], + "lines": [ + { + "bbox": [ + 209, + 196, + 304, + 205 + ], + "spans": [ + { + "bbox": [ + 209, + 196, + 304, + 205 + ], + "type": "text", + "content": "Ours (open-set, w/ TTFT)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 207, + 304, + 310 + ], + "lines": [ + { + "bbox": [ + 209, + 207, + 304, + 310 + ], + "spans": [ + { + "bbox": [ + 209, + 207, + 304, + 310 + ], + "type": "image", + "image_path": "891701f907dca521cd180e4579b7994d49151102e0e61f8444b8430f73b49857.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 207, + 402, + 310 + ], + "blocks": [ + { + "bbox": [ + 306, + 196, + 402, + 205 + ], + "lines": [ + { + "bbox": [ + 306, + 196, + 402, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 196, + 402, + 205 + ], + "type": "text", + "content": "Ours (closed-set, w/o TTFT)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 402, + 196, + 497, + 205 + ], + "lines": [ + { + "bbox": [ + 402, + 196, + 497, + 205 + ], + "spans": [ + { + "bbox": [ + 402, + 196, + 497, + 205 + ], + "type": "text", + "content": "Ours (closed-set, w/ TTFT)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 207, + 402, + 310 + ], + "lines": [ + { + "bbox": [ + 306, + 207, + 402, + 310 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 402, + 310 + ], + "type": "image", + "image_path": "f1a945b5e0abb6bc2ef3a6d20d30bfefb2bc044a40660143528ab232080087a8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 402, + 207, + 498, + 310 + ], + "blocks": [ + { + "bbox": [ + 402, + 207, + 498, + 310 + ], + "lines": [ + { + "bbox": [ + 402, + 207, + 498, + 310 + ], + "spans": [ + { + "bbox": [ + 402, + 207, + 498, + 310 + ], + "type": "image", + "image_path": "590fc1ebc9d6a940db25ee795122e5881b9188b27cdd95ade245c374d5901545.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 86, + 356, + 260, + 508 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 555, + 353 + ], + "type": "text", + "content": "Figure 7. Visualization of object detection results. While TIC provides better visual quality, it exhibits low detection confidence, leading to misdetections—such as misclassifying a bear as a sheep. Ours (w/ TTFT) surpasses the uncompressed image in prediction confidence despite compression losses." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 356, + 260, + 508 + ], + "lines": [ + { + "bbox": [ + 86, + 356, + 260, + 508 + ], + "spans": [ + { + "bbox": [ + 86, + 356, + 260, + 508 + ], + "type": "image", + "image_path": "babeb2e1ccfc0505bc4cc3ae2e81598bd06e34a6895e8c00669eb7f44e07de6c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 518, + 295, + 540 + ], + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 540 + ], + "type": "text", + "content": "Figure 8. Ablation on the range of fine-tuning. Experiments show that fine-tuning the decoder is essential for optimal performance." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "content": "comparison includes fine-tuning either the encoder or the decoder, or both together. The experimental results show that TTFT applied to the decoder stage contributes to performance improvement. While the encoder stage, which involves a larger number of trainable parameters during TTFT, yields greater performance gains, the decoder-stage TTFT remains essential for achieving optimal performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Fig. 9 presents a comparison between decoder finetuning using only the SVs and using the full set of SVD-LoRA parameters. The results show that training all SVD-LoRA parameters increases bitstream overhead by " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "15.3\\mathrm{x}" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": ". Therefore, restricting training to only the SVs in the decoder is critical for enabling fully instance-specific finetuning with minimal overhead." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 345, + 356, + 517, + 509 + ], + "blocks": [ + { + "bbox": [ + 345, + 356, + 517, + 509 + ], + "lines": [ + { + "bbox": [ + 345, + 356, + 517, + 509 + ], + "spans": [ + { + "bbox": [ + 345, + 356, + 517, + 509 + ], + "type": "image", + "image_path": "a9f7bf7112fae62d831c2af251fce2897ed3e612199c4d9db4e1868ace8b9547.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 518, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 552 + ], + "type": "text", + "content": "Figure 9. Ablation on the range of decoder fine-tuning. Experiments show that training only the singular values of the decoder effectively reduces excessive bitstream overhead." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 554, + 388, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 388, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 388, + 566 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We propose a fully instance-specific TTFT approach for adapting learned image compression to both closed-set and open-set machine vision tasks. Our method can overcome the challenging problem of instance-specific decoder finetuning, which remains difficult due to its significant impact on both bitstream size and decoding time but is essential for effectively handling open-set tasks. The SVD-LoRA we applied enables TTFT while minimizing bitstream overhead. Moreover, it can be seamlessly integrated into any architecture, including CNNs and Transformers. Therefore, our approach can be applied to various domains where instance-specific TTFT is required." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4437" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 270 + ], + "type": "text", + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) under multiple grants funded by the Korea government (MSIT), including the ICT Creative Consilience Program (IITP-2025-RS-2020-II201821), the Information Technology Research Center (ITRC) support program (IITP-2021-0-02052), and the Artificial Intelligence Innovation Hub (RS-2021-II212068). Additionally, this research was supported by Korea Research Institute for defense Technology planning and advancement (KRIT) grant funded by the Korean Government, Defense Acquisition Program Administration (DAPA), under Grant KRIT-CT-23-021, Development of Multi-Modal Information Fusion and High-Realization Battlefield Data Generation Technology, in 2025." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 281, + 115, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 281, + 115, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 115, + 293 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 300, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 300, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 296, + 354 + ], + "type": "text", + "content": "[1] Umar Albalawi, Saraju P Mohanty, and Elias Kougianos. A hardware architecture for better portable graphics (bpg) compression encoder. In 2015 IEEE international Symposium on Nanoelectronic and information systems, pages 291-296. IEEE, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 356, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 356, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 61, + 356, + 296, + 411 + ], + "type": "text", + "content": "[2] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 413, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 296, + 443 + ], + "type": "text", + "content": "[3] Dana Harry Ballard and Christopher M Brown. Computer vision. Prentice Hall Professional Technical Reference, 1982. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 446, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 296, + 479 + ], + "type": "text", + "content": "[4] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. In International Conference on Learning Representations, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 480, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 523 + ], + "type": "text", + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 525, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 555 + ], + "type": "text", + "content": "[6] Klaudia Bałazy et al. Lora-xs: Low-rank adaptation with extremely small number of parameters. arXiv preprint, 2024. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "type": "text", + "content": "[7] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 667 + ], + "type": "text", + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "text", + "content": "[9] Prabuddha Chakraborty, Jonathan Cruz, and Swarup Bhunia. Magic: Machine-learning-guided image compression for vision applications in internet of things. IEEE Internet of Things Journal, 8(9):7303-7315, 2020. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[10] Lahiru D Chamain, Fabien Racapé, Jean Bégaint, Akshay Pushparaja, and Simon Feltman. End-to-end optimized image compression for machines, a study. In 2021 Data Compression Conference (DCC), pages 163-172. IEEE, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "text", + "content": "[11] Lahiru D Chamain, Fabien Racapé, Jean Bégaint, Akshay Pushparaja, and Simon Feltman. End-to-end optimized image compression for multiple machine tasks. arXiv preprint arXiv:2103.04178, 2021. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "text", + "content": "[12] Arnav Chavan et al. One-for-all: Generalized lora for parameter-efficient fine-tuning. arXiv preprint arXiv:2306.07967, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 195, + 555, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 555, + 260 + ], + "type": "text", + "content": "[13] Yi-Hsin Chen, Ying-Chieh Weng, Chia-Hao Kao, Cheng Chien, Wei-Chen Chiu, and Wen-Hsiao Peng. Transtic: Transferring transformer-based image compression from human perception to machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23297-23307, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "text", + "content": "[14] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 5, 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 317, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 348 + ], + "type": "text", + "content": "[15] Hyomin Choi and Ivan V Bajic. Scalable image coding for humans and machines. IEEE Transactions on Image Processing, 31:2739-2754, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[16] Jack Cole and Mohamed Osman. Dataset-induced meta-learning (and other tricks): Improving model efficiency on arc, 2023. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "text", + "content": "[18] Lingyu Duan, Jiaying Liu, Wenhan Yang, Tiejun Huang, and Wen Gao. Video coding for machines: A paradigm of collaborative compression and intelligent analytics. IEEE Transactions on Image Processing, 29:8680-8695, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[19] Zhihao Duan, Zhan Ma, and Fengqing Zhu. Unified architecture adaptation for compressed domain semantic inference. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):4108-4121, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "text", + "content": "[20] Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen. Image coding for machines with omnipotent feature learning. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "text", + "content": "[21] Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 615, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 658 + ], + "type": "text", + "content": "[22] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 691 + ], + "type": "text", + "content": "[23] Bjontegaard Gisle. Calculation of average psnr differences between rd curves. In ITU-T SG16/Q6, " + }, + { + "bbox": [ + 317, + 658, + 553, + 691 + ], + "type": "inline_equation", + "content": "13^{\\circ}" + }, + { + "bbox": [ + 317, + 658, + 553, + 691 + ], + "type": "text", + "content": " th VCEG Meeting, Austin, Texas, USA, April 2001, 2001. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Sha Guo, Lin Sui, Chenlin Zhang, Zhuo Chen, Wenhan Yang, and Lingyu Duan. A unified image compression" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4438" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "method for human perception and multiple vision tasks. In European Conference on Computer Vision, pages 342-359. Springer, 2024. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[25] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 290, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 290, + 175 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 290, + 175 + ], + "type": "text", + "content": "[26] Allen Hanson. Computer vision systems. Elsevier, 1978. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 217 + ], + "type": "text", + "content": "[27] Moritz Hardt and Yu Sun. Test-time training on nearest neighbors for large language models. In The Twelfth International Conference on Learning Representations, 2024. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "text", + "content": "[28] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 287, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 287, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 287, + 294, + 331 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 333, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 333, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 57, + 333, + 295, + 365 + ], + "type": "text", + "content": "[30] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 409 + ], + "type": "text", + "content": "[31] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In Proceedings of the ICLR, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "text", + "content": "[32] Yueyu Hu, Shuai Yang, Wenhan Yang, Ling-Yu Duan, and Jiaying Liu. Towards coding for human and machine vision: A scalable image coding approach. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "type": "text", + "content": "[33] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "text", + "content": "[34] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Realtime multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "text", + "content": "[35] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 657 + ], + "type": "text", + "content": "[36] Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien. How to exploit the transferability of learned image compression to conventional codesc. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16165-16174, 2021. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[37] Nam Le, Honglei Zhang, Francesco Cricri, Ramin GhaznaviYouvalari, and Esa Rahtu. Image coding for machines: an end-to-end learned approach. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1590-1594. IEEE, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[38] Huanyang Li and Xinfeng Zhang. Human-machine collaborative image compression method based on implicit neural representations. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[39] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, 133(1):31-64, 2025. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 227 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "text", + "content": "[41] Jiaheng Liu, Guo Lu, Zhihao Hu, and Dong Xu. A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370, 2020. 1, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[42] Jinming Liu, Heming Sun, and Jiro Katto. Learning in compressed domain for faster machine vision tasks. In 2021 International Conference on Visual Communications and Image Processing (VCIP), pages 01-05. IEEE, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "text", + "content": "[43] Jinming Liu, Heming Sun, and Jiro Katto. Improving multiple machine vision tasks in the compressed domain. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 331-337. IEEE, 2022. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[44] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14388-14397, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 554, + 437 + ], + "type": "text", + "content": "[45] Kang Liu, Dong Liu, Li Li, Ning Yan, and Houqiang Li. Semantics-to-signal scalable image compression with learned revertible representations. International Journal of Computer Vision, 129(9):2605-2621, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[46] Ming Lu, Peiyao Guo, Huiqing Shi, Chuntong Cao, and Zhan Ma. Transformer-based image compression. In 2022 Data Compression Conference (DCC), pages 469-469. IEEE, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "text", + "content": "[47] Siwei Ma, Xinfeng Zhang, Chuanmin Jia, Zhenghui Zhao, Shiqi Wang, and Shanshe Wang. Image and video compression with neural networks: A review. IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683-1698, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "type": "text", + "content": "[48] Yixin Mei, Fan Li, Li Li, and Zhu Li. Learn a compression for objection detection-vae with a bridge. In 2021 International Conference on Visual Communications and Image Processing (VCIP), pages 1-5. IEEE, 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[49] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[50] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 691 + ], + "type": "text", + "content": "[51] Vishvjit S Nalwa. A guided tour of computer vision. Addison-Wesley Longman Publishing Co., Inc., 1994. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[52] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4439" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[53] Myungseo Song, Jinyoung Choi, and Bohyung Han. Variable-rate deep image compression through spatially-adaptive feature transform. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2380-2389, 2021. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[54] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "text", + "content": "[55] Vivienne Sze, Madhukar Budagavi, and Gary J Sullivan. High efficiency video coding (hevc). In Integrated circuit and systems, algorithms and architectures, page 40. Springer, 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 275 + ], + "type": "text", + "content": "[56] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 342 + ], + "type": "text", + "content": "[57] Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. Dylora: Parameter-efficient tuning of pretrained models using dynamic search-free low-rank adaptation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3274-3287, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "text", + "content": "[58] Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, and Eftychios Protopapadakis. Deep learning for computer vision: A brief review. Computational intelligence and neuroscience, 2018(1):7068349, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[59] Shurun Wang, Zhao Wang, Shiqi Wang, and Yan Ye. End-to-end compression towards machine vision: Network architecture design and optimization. IEEE Open Journal of Circuits and Systems, 2:675-685, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "text", + "content": "[60] Shurun Wang, Zhao Wang, Shiqi Wang, and Yan Ye. Deep image compression toward machine vision: A unified optimization framework. IEEE Transactions on Circuits and Systems for Video Technology, 33(6):2979-2989, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[61] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[62] Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu. Sssic: semantics-to-signal scalable image coding with learned structural representations. IEEE Transactions on Image Processing, 30:8939-8954, 2021. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "type": "text", + "content": "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[64] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fin-tuning. In International Conference on Learning Representations, 2023. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[65] Y Zhang, M Rafie, and S Liu. Use cases and requirements for video coding for machines. ISO/IEC JTC, 1, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[66] Xia Zhao, Limin Wang, Yufei Zhang, Xuming Han, Muhammet Deveci, and Milan Parmar. A review of convolutional neural networks in computer vision. Artificial Intelligence Review, 57(4):99, 2024. 1" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4440" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Test-Time Visual In-Context Tuning/1106c942-047b-40f5-b114-2de12258d183_content_list.json b/2025/Test-Time Visual In-Context Tuning/1106c942-047b-40f5-b114-2de12258d183_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9cd3837337cb853fdecab6491a1e4d826dfcd961 --- /dev/null +++ b/2025/Test-Time Visual In-Context Tuning/1106c942-047b-40f5-b114-2de12258d183_content_list.json @@ -0,0 +1,1245 @@ +[ + { + "type": "text", + "text": "Test-Time Visual In-Context Tuning", + "text_level": 1, + "bbox": [ + 315, + 130, + 684, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahao Xie $^{1,2}$ , Alessio Tonioni $^{3}$ , Nathalie Rauschmayr $^{3}$ , Federico Tombari $^{3}$ , Bernt Schiele $^{1,2}$ , $^{1}$ Max Planck Institute for Informatics, SIC $^{2}$ VIA Research Center $^{3}$ Google", + "bbox": [ + 135, + 179, + 859, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jxie, schiele}@mpi-inf.mpg.de {alessiot, rauschmayr, tombari}@google.com", + "bbox": [ + 163, + 218, + 823, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dc83c086a4cc0252879660d3cfa511bf47deca6f2878ba337d9a0e0a35b4e1c7.jpg", + "image_caption": [ + "Figure 1. Test-time visual in-context tuning (VICT) on six representative vision tasks under distribution shifts. We benchmark the robustness of VICL with 15 common corruptions adopted in [23, 31], and report the averaged performance across all corruptions. Existing VICL models like Painter exhibit poor generalization capability to unseen new domains when the task prompts come from the training distribution (i.e., zero-shot). Performances are even worse when given task prompts from the test distribution (i.e., one-shot). By performing VICT at test time, we can significantly improve Painter in both zero-shot and one-shot manners." + ], + "image_footnote": [], + "bbox": [ + 91, + 265, + 906, + 368 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 460, + 326, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual in-context learning (VICL), as a new paradigm in computer vision, allows the model to rapidly adapt to various tasks with only a handful of prompts and examples. While effective, the existing VICL paradigm exhibits poor generalizability under distribution shifts. In this work, we propose test-time Visual In-Context Tuning (VICT), a method that can adapt VICL models on the fly with a single test sample. Specifically, we flip the role between the task prompts and the test sample and use a cycle consistency loss to reconstruct the original task prompt output. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts. Extensive experiments on six representative vision tasks ranging from high-level visual understanding to low-level image processing, with 15 common corruptions, demonstrate that our VICT can improve the generalizability of VICL to unseen new domains. In addition, we show the potential of applying VICT for unseen tasks at test time. Code: https://github.com/Jiahao000/VICT.", + "bbox": [ + 89, + 493, + 485, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 813, + 220, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following the success of in-context learning (ICL) [2, 7, 21] in natural language processing (NLP), visual in-context learning (VICL) [5, 52] has shown promising performance in developing generalist models for vision tasks. Inspired", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by prompting in NLP that defines tasks using language sequences as a general interface, existing VICL works [5, 52] use images themselves as a natural interface for general-purpose visual perception. They formulate VICL as an image inpainting task, i.e., given an input-output example describing a specific task (i.e., the prompt) as well as an input image, they are assembled in a grid and the problem is casted as inpainting the missing part of the grid (i.e., the prediction) consistently with the given task prompts. This enables adapting a pre-trained vision model to various downstream tasks with only a handful of prompts and examples.", + "bbox": [ + 511, + 462, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By default, current VICL models are frozen during deployment. However, their performance might suffer in real-world deployment as the test distribution usually changes and deviates from the training one. This naturally raises a question: what is the generalizability of VICL models under distribution shifts? In this work, we investigate this aspect focusing on controllable shifts due to image corruptions. We observe that the existing VICL paradigm like Painter [52] exhibits poor generalizability to unseen new domains when the task prompt comes from the training distribution, as shown in Figure 1. Surprisingly, performances are even worse when given input-output task prompts belonging to the same distribution as the test input.", + "bbox": [ + 511, + 628, + 908, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Driven by this observation, we propose to rely on test-time visual in-context tuning (VICT) to adapt VICL models on the fly to unseen distributions using only the given test sample. The motivation is that each test input offers a hint about the test distribution. Thus, we modify the VICL", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19996", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model at test time to make full use of this hint by setting up a one-sample learning problem. Specifically, given input-output task prompts and the test input, we first use the VICL model to inpaint the output image. We then flip the role between the task prompts and the test sample, i.e., we treat the predicted test output as the prompt to the model and reconstruct the original output of task prompts. This allows us to tune the parameters of the whole model in a self-supervised manner that can be applied to arbitrary tasks. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts conditioned on its in-context inference. Such a cycle consistency supervision signal naturally exists in the context of VICL without requiring any additional training data or annotations, thus making it an appealing self-supervisory task for test-time visual in-context training.", + "bbox": [ + 89, + 90, + 480, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We explore VICT in two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). Without loss of generality, we consider Painter [52] as our VICL model, for its simplicity in design and its wide applicability. As shown in Figure 1, our simple method leads to substantial improvements across 15 common corruptions [23, 31] on six representative vision tasks ranging from high-level visual understanding to low-level image processing, including depth estimation on NYUv2 [40], semantic segmentation on ADE20K [63], panoptic segmentation on COCO [27], image denoising on SIDD [1], image deraining on the merged deraining dataset [57], and low-light image enhancement on LoL [53].", + "bbox": [ + 89, + 333, + 480, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 109, + 560, + 450, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We propose a new cycle consistency task for test-time visual in-context tuning. To the best of our knowledge, we are the first to perform test-time training for VICL.", + "2) We contribute the first study on the generalizability of VICL under distribution shifts. We observe that the existing VICL paradigm exhibits poor generalizability to unseen new domains. Such a phenomenon can hardly be recovered even given input-output task prompts with the same distribution as the test input.", + "3) We conduct extensive experiments on six representative vision tasks across 15 common corruptions. VICT significantly improves Painter in both zero-shot and one-shot manners. Our zero-shot or one-shot VICT can even outperform Painter trained with more few-shot corrupted examples. We also explore the potential of applying VICT for unseen tasks at test time, further demonstrating its promise." + ], + "bbox": [ + 89, + 574, + 482, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 830, + 232, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In-context learning. In-context learning has been extensively explored in the NLP literature after the introduction of large language models. Seminal works in the field", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "like GPT-3 [7] go as far as claiming that language models are indeed few-shot learners. All recent language models like [13, 46, 47] have the ability to adapt their behavior based on few-shot examples provided in their context window or directly to follow simple user instructions [10]. Similar concepts have been extended to multimodal models with seminal works like Flamingo [2] showing few-shot capabilities by providing examples as interleaved text and images in the context window. Some of these capabilities are now available for extremely large commercial multimodal large language models [35, 42, 43].", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Apart from using languages as the general interface, a recent line of work introduces purely visual in-context learning (VICL) by training vision generalist models that can perform arbitrary visual tasks following one or few visual examples provided at inference time. Two representative works developed in parallel are: (i) MAE-VQGAN [5], which trains a variant of MAE [22] on a dataset of figures extracted from academic papers, and (ii) Painter [52], which uses a similar idea but trains its model on a set of standard academic benchmark datasets. More recently, this strategy has been expanded in LVM [3], where even more visual datasets are collected to train an autoregressive generative model. We base our work on Painter [52] due to its simplicity in design and its wide applicability. It is the one that provides both the available code and model weights, and has already been extensively evaluated across several standard benchmarks.", + "bbox": [ + 511, + 260, + 903, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalization under distribution shifts. Machine learning models suffer from performance drops when tested on a data distribution different from the one they are trained on [17, 49, 50, 61]. The lower the drop, the more we define a model robust or able to generalize to distribution shifts. Common strategies to increase model robustness include using heavy data augmentations [11, 23, 44, 62] or extremely large training distributions going up to web scale [33, 37, 59]. Nevertheless, most of the experiments are done in a classification setting. To the best of our knowledge, there is no previous work exploring the robustness of VICL models. We are the first to counteract performance drops in VICL under distribution shifts.", + "bbox": [ + 511, + 520, + 903, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-time training. An alternative way of counteracting performance drops due to distribution shifts is test-time training [6, 8, 16, 28, 41, 48, 51]. This accounts to unfreezing the model at test time and fine-tuning it on the target distribution (or a single sample from it) through self-supervision. Early works propose this paradigm using self-supervised pretext tasks (e.g., rotation prediction [18]) to explicitly improve generalization under image corruptions [41] or improve in specific tasks where self-supervised losses can be clearly defined like depth estimation [26, 36, 45], reinforcement learning [20, 32], tracking [15, 38], and NLP [4, 56]. More recently, the field", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19997", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/493a43ac28f6eab7fb18764968e13a6cb57cd68a1e190644b4172ed47eb7b39e.jpg", + "image_caption": [ + "Figure 2. Overview of our VICT pipeline. Given a pair of task prompts $(x,y)$ and a test input image $x_{t}$ , we first construct a four-cell grid-like image canvas $I = (x,y,x_{t},\\emptyset)$ , with an empty cell at the bottom right. We then feed $I$ into the VICL model (e.g., Painter) to predict the test output $\\hat{y}_t$ . Afterward, we flip the role between input-output task prompts and input-output test samples, i.e., we provide the predicted $\\hat{y}_t$ as the prompt to the model, recreating a new four-cell grid-like image canvas $I' = (x,\\emptyset,x_{t},\\hat{y}_{t})$ , with an empty cell at the top right. The new $I'$ is fed into the same model to predict the task prompt output $\\hat{y}$ . We finally optimize the model by minimizing the distance between $\\hat{y}$ and $y$ via a standard regression loss." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "has realized that MAE [22] provides a very strong self-supervised pretext task and has therefore been employed in the context of test-time training [16]. Finally, some recent works [14, 24, 30, 39, 58] also employ the test-time training paradigm for vision-language models to improve their generalization. To the best of our knowledge, ours is the first work to explore test-time training applied to VICL models, where previous self-supervised pretext tasks [9, 12, 18, 22, 25, 34, 54, 55, 60] cannot be employed due to their constraints on the single-image context.", + "bbox": [ + 88, + 411, + 485, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 574, + 225, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our visual in-context tuning (VICT) is a simple yet effective test-time training approach to adapt visual in-context learning (VICL) models on the fly. In this section, we first introduce some preliminaries on VICL in Section 3.1. We then detail our VICT pipeline in Section 3.2.", + "bbox": [ + 89, + 599, + 483, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Visual In-Context Learning", + "text_level": 1, + "bbox": [ + 89, + 681, + 339, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In-context learning is a new paradigm originating from large language models such as GPT-3 [7] in NLP. Unlike traditional learning paradigms, in-context learning formulates different NLP tasks as text completion tasks and makes predictions conditioned on one or many support examples provided to the model in the context window. Extending this paradigm to tasks requiring a dense visual output in the form of images is nontrivial and has not been addressed in the literature for a long time. Recently, [5, 52] formulate VICL as an image inpainting task by combining images and labels into a grid-like new image and using masked image modeling for pre-training, granting the models with the in-context learning ability.", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, let $S = \\{(x_i, y_i)\\}_{i=1}^N$ denote the set of support input-output examples (a.k.a., task prompts) where $x$ is an image and $y$ is its visual label (e.g., a segmentation mask).1 Given $S$ and a new test image $x_t$ as input, VICL can be formulated as follows:", + "bbox": [ + 511, + 411, + 906, + 486 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = f _ {\\theta} (S, x _ {t}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 655, + 500, + 906, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{\\theta}(*)$ is the VICL model parametrized with $\\theta$ , $y_{t}$ is the corresponding output label of $x_{t}$ . Usually, we assume $S$ and $x_{t}$ are drawn from the same distribution. However, this is rarely the case for real-world deployment. Therefore, in this work, we mainly consider the scenario where a distribution shift like an image corruption occurs for $x_{t}$ .", + "bbox": [ + 511, + 527, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the experiment section, we will show how the presence of such distribution shifts significantly degrades the performance of the VICL models. Moreover, we will additionally show how simple solutions like providing in-context examples from the target (corrupted) distribution would not fix the problem due to the limited generalization ability of the current VICL models.", + "bbox": [ + 511, + 619, + 905, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Test-Time Visual In-Context Tuning", + "text_level": 1, + "bbox": [ + 511, + 734, + 826, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose to leverage test-time training [41] to counteract the distribution shifts between training and testing data, with the goal of making VICL models more robust. We argue that VICL models are particularly amenable to be optimized at test time on arbitrary tasks thanks to the availability of the $N$ shots provided as context. The core intuition of our method is to use the provided in-context examples to define a self-supervised loss that can be used at", + "bbox": [ + 511, + 756, + 906, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Following previous works, we set $N = 1$ in practice.", + "bbox": [ + 531, + 886, + 820, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19998", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Pseudocode of VICT in a PyTorch-like style." + ], + "code_body": "f: VICL model (e.g., Painter) \n# theta_0: pre-trained model weights \n# x: task prompt input \n# y: task prompt output \n# mask_token: mask token for the empty cell \n# steps: number of test-time optimization steps \nfor x_t in loader: # load a test sample x_t \nf.params = theta_0 # initialize \n# test-time optimization \nfor t in range(steps): \n # predict the test sample output \n y_t_pred = f((x, y, x_t, mask_token)) \n # predict the task prompt output \n y_pred = f((x, mask_token, x_t, y_t_pred)) \n # regression loss \n loss = SmoothL1Loss(y_pred, y) \n # update model \n loss/backward() \nupdate(f.params) \n# inference with the updated model \ny_t_pred = f((x, y, x_t, mask_token))", + "guess_lang": "python", + "bbox": [ + 91, + 114, + 468, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "test time to train any VICL models on the fly on any task. The overall pipeline of our VICT is illustrated in Figure 2.", + "bbox": [ + 89, + 410, + 480, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, given a pair of task prompts $(x,y)$ from $S$ and a test input $x_{t}$ , we construct a grid-like image canvas with four cells, denoted by an ordered quadruple $I = (x,y,x_{t},\\emptyset)$ , where $\\emptyset$ is an empty cell. We then feed $I$ into the VICL model $f_{\\theta}$ to predict the test output $\\hat{y}_t$ :", + "bbox": [ + 89, + 440, + 482, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {t} = f _ {\\theta} (I) = f _ {\\theta} \\left(\\left(x, y, x _ {t}, \\varnothing\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 527, + 482, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterward, with the obtained test output, we flip the role between input-output task prompts and input-output test samples, creating a new four-cell grid-like image canvas with the order of $I' = (x, \\varnothing, x_t, \\hat{y}_t)$ . The new $I'$ is fed into $f_{\\theta}$ again to predict the task prompt output $\\hat{y}$ :", + "bbox": [ + 89, + 556, + 483, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = f _ {\\theta} \\left(I ^ {\\prime}\\right) = f _ {\\theta} \\left(\\left(x, \\varnothing , x _ {t}, \\hat {y} _ {t}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 643, + 482, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then use a simple regression loss in pixel space to optimize the model by minimizing the distance between $\\hat{y}$ and $y$ :", + "bbox": [ + 89, + 672, + 482, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {x _ {t}} = \\arg \\min _ {\\theta} \\mathcal {L} (\\hat {y}, y) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 717, + 480, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, we follow [52] to use smooth- $\\ell_1$ [19] loss. Using this loss we can optimize the parameters $\\theta$ of the model for a small number of steps for each $x_t$ and, eventually, make a prediction on $x_t$ as $f_{\\theta_{x_t}}((x,y,x_t,\\emptyset))$ .", + "bbox": [ + 89, + 744, + 482, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our formulation, the gradient-based optimization for Equation 4 always starts from the initial pre-trained model weights $\\theta_0$ for each test input. When a new test input arrives, we discard $\\theta_{x_t}$ and reset the weights to $\\theta_0$ . We follow", + "bbox": [ + 89, + 805, + 483, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this strategy to not have any assumption on the test inputs and treat them independently at test time, i.e., we do not assume that they come from the same distribution. Our generic formulation is completely agnostic on the VICL models considered, on the type of tasks being solved, and on the type of in-context samples provided to it.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pseudo-code of VICT is in Algorithm 1.", + "bbox": [ + 532, + 181, + 830, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 215, + 645, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 241, + 730, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tasks and datasets. Following [52], we evaluate the model performance on six representative vision tasks ranging from high-level visual understanding to low-level image processing, which includes depth estimation on NYUv2 [40], semantic segmentation on ADE20K [63], panoptic segmentation on COCO [27], image denoising on SIDD [1], image draining on the merged draining datasets [57], and low-light image enhancement on LoL [53]. To evaluate the robustness to distribution shifts of the model on these datasets, we follow the setup in [23, 31] to corrupt the aforementioned datasets and simulate hard distribution shifts. The corrupted datasets contain 15 types of corruptions, covering noise (gaussian noise, impulse noise, shot noise), blur (defocus blur, glass blur, motion blur, zoom blur), weather (fog, frost, snow), and digital (brightness, contrast, elastic transform, JPEG compression, pixelate) categories. Each type of corruption has five levels of severity. We denote the corrupted datasets as NYUv2-C, ADE20K-C, COCO-C, SIDD-C, Rain-C, and LoL-C, respectively. Due to the space constraints, the results in the main text are limited to the most severe level (i.e., level 5) that corresponds to the strongest distribution shift. We provide the results on the other four levels in the supplementary material.", + "bbox": [ + 511, + 263, + 906, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baselines. Without loss of generality, we consider Painter [52] as our VICL model for its simplicity in design and its wide applicability. We study VICT in two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). For the one-shot setting, apart from comparing with the frozen Painter using one-shot corrupted examples as task prompts, we also consider a baseline that further trains Painter with the one-shot corrupted samples using the same pre-training objective as in [52]. We also consider training Painter with more few-shot examples using the same settings to examine to what extent VICT can outperform a larger number of few-shot fine-tuning of Painter.", + "bbox": [ + 511, + 612, + 906, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training details. In all experiments, we use the pre-trained Painter based on ViT-Large provided by the authors of [52]. We perform VICT using an AdamW [29] optimizer, with betas as (0.9, 0.999), a weight decay of 0, a batch size of 1,", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2The image order is top left, top right, bottom left, and bottom right, respectively.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19999", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4132bf0007f71a0962ddc00eed905a825efe8f7123a5daca31f335f54839a7a1.jpg", + "table_caption": [ + "Table 1. System-level comparison on six representative vision tasks across 15 corruptions. Results are on corruption level 5. We consider two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). \"avg\" denotes the averaged results over 15 corruptions." + ], + "table_footnote": [], + "table_body": "
methodbrighcontdefocelastfogfrostgaussglassimpuljpegmotnpixelshotsnowzoomavg
(a) depth estimation NYUv2-C (A.Rel ↓)
zero-shot setting:
Painter0.1290.2150.7120.1090.1290.5360.2000.6120.1890.3860.1670.1870.1421.9510.2090.392
VICT0.1080.2160.6310.1080.1330.5760.1910.5410.1810.3100.1390.1340.1401.8560.2100.365
one-shot setting:
Painter0.1260.2850.7430.1090.1320.8530.9010.6220.9010.3920.1740.1940.4401.9640.2120.537
VICT0.0970.1930.1800.1070.1280.2270.2410.2100.2780.1590.1210.1140.1500.6620.2140.205
(b) semantic segmentation ADE20K-C (mIoU ↑)
zero-shot setting:
Painter40.411.727.931.235.023.324.825.626.138.631.840.425.99.413.127.0
VICT40.912.728.231.736.423.725.526.327.438.831.940.928.017.313.628.2
one-shot setting:
Painter40.911.727.531.335.022.119.725.820.338.531.340.323.26.513.325.8
VICT41.721.028.832.237.024.724.427.025.239.732.841.526.621.714.929.3
(c) panoptic segmentation COCO-C (PQ ↑)
zero-shot setting:
Painter38.115.525.530.033.326.024.923.825.531.927.634.926.319.812.526.4
VICT38.715.725.830.333.926.725.324.325.932.427.735.426.721.912.726.9
one-shot setting:
Painter38.114.224.830.033.125.221.523.722.031.527.534.623.817.212.125.3
VICT38.616.925.430.433.826.624.024.624.532.327.435.025.224.912.426.8
(d) denoising SIDD-C (PSNR ↑)
zero-shot setting:
Painter8.7225.4932.4833.2016.868.7917.5033.1817.5023.1829.4734.8425.279.8629.8023.08
VICT8.8625.6532.9132.9617.008.8417.5833.0017.5024.8629.6834.7625.219.7529.6723.22
one-shot setting:
Painter9.1225.1631.9933.2316.689.2417.9933.1418.3322.7329.3034.8724.6910.5029.7623.12
VICT13.9425.3232.4134.4516.4511.1520.1332.9619.9326.8030.1135.0124.9314.9029.9524.56
(e) deraining Rain-C (PSNR ↑)
zero-shot setting:
Painter10.7414.1519.5420.1013.9112.4919.8519.8020.1624.8618.5222.8120.1111.8917.9317.79
VICT10.8914.2419.5820.3113.9412.6220.4219.8320.6724.9718.5623.0520.4712.0517.9817.97
one-shot setting:
Painter11.0413.9519.6120.1513.8212.8719.5119.8719.5824.8518.6222.8819.6812.1617.9517.77
VICT17.3814.6920.3621.4414.6717.6922.3320.4822.2425.3119.0523.9221.3719.5318.2819.92
(f) low-light enhancement LoL-C (PSNR ↑)
zero-shot setting:
Painter16.2613.4318.8019.8311.2411.6212.5018.8212.9115.8718.0121.1217.7513.7716.6015.90
VICT16.3913.5719.2419.8411.3711.6412.5119.2313.0016.0018.1121.2019.1613.9217.0616.15
one-shot setting:
Painter16.7413.5018.8519.8311.0211.7613.1418.9513.7415.8418.1521.1216.5413.1916.5915.93
VICT18.1915.0318.9919.8811.7612.0114.9519.2214.9916.1218.2321.0919.8416.9616.9116.94
", + "bbox": [ + 98, + 155, + 898, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20000", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c21e893ebea43beca12e72336776524f2549dc0a77e2537b2fcfeacaa60eeda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 89, + 364, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/24a6e1ad11f78096a691db1425e9a292e7ddefb7e1104c44a4465a04bbe35ef9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 90, + 630, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74a4ebeb11dadf3b69613d7892e0723dd12e57012f4e222f055cf1c1becb0ea5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 90, + 895, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43db942e5f7350ca357479316f7acfcbb8f1aa2eb081fd57037946c60a4ce14f.jpg", + "image_caption": [ + "Figure 3. Comparison with few-shot Painter on six vision tasks with corruptions. We randomly corrupt a certain number of images in the training set, using 1, 2, 4, 8, 16, 32, and 64 shots for training and deploying the model in the full corrupted test sets. We report the final results averaged across 15 corruptions. Our zero-shot or one-shot VICT can outperform Painter trained with more few-shot examples." + ], + "image_footnote": [], + "bbox": [ + 96, + 252, + 364, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/28dfed92e636cab673193c1254943774664a2e793f89e4bc5c975bc488290c56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 252, + 630, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ae8f4e72ae4967bcd35bf4d953259d2f92666579db157e14ed560ebcab99ca4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 252, + 895, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and a fixed learning rate of 1e-6. Unless otherwise specified, we train each test sample for 60 steps. The choice of 60 steps is purely computational and more steps are likely to further improve performance, judging from the positive trend observed in Figure 4. Note that we do not apply any data augmentations for VICT as many data augmentations are in fact distribution shifts in our evaluation benchmarks. Training with them is analogous to training on the test distributions. Therefore, to solely study the generalization ability to new test distributions, we purposely choose not to use any data augmentations, even though they could improve our results at face value. Every iteration of VICT is performed on the same concatenated grid-like image, with each subimage resized as $448 \\times 448$ , which is the same as what we later use for in-context inference. During VICT, we only optimize the encoder weights. We have experimented with training both the encoder and the decoder and found that the difference is negligible, which is also consistent with the observations in [16, 41]. A detailed comparison is provided in the experiment section.", + "bbox": [ + 88, + 478, + 485, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 787, + 230, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "System-level comparison. We compare VICT with Painter under 15 common corruptions as introduced in Section 4.1. We perform experiments on six representative vision tasks covering high-level visual understanding and low-level image processing, which includes depth estimation on NYUv2-C, semantic segmentation on ADE20K-C, panop-", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tic segmentation on COCO-C, image denoising on SIDD-C, image draining on Rain-C, and low-light image enhancement on LoL-C.", + "bbox": [ + 511, + 478, + 906, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 reports the results. First of all, we observe that Painter exhibits poor generalization ability under common corruptions. For example, on ADE20K-C semantic segmentation, Painter only achieves an average performance of $27.0\\mathrm{mIoU}$ , which is $22.9\\mathrm{mIoU}$ lower than that on the clean validation set (49.9 mIoU as reported in Figure 1). Nevertheless, our VICT outperforms Painter by clear margins in both zero-shot and one-shot settings across different tasks. When it comes to the zero-shot setting, VICT achieves an average of -0.027 A.Rel on NYUv2-C depth estimation, +1.2 mIoU on ADE20K-C semantic segmentation, +0.5 PQ on COCO-C panoptic segmentation, +0.14 PSNR on SIDD-C image denosing, +0.18 PSNR on Rain-C image draining, and +0.25 PSNR on LoL-C low-light enhancement, respectively.", + "bbox": [ + 511, + 531, + 908, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance gains of VICT over Painter are further increased when we shift from the zero-shot setting to the one-shot counterpart. Specifically, VICT achieves an average of -0.332 A.Rel on NYUv2-C depth estimation, +3.5 mIoU on ADE20K-C semantic segmentation, +1.5 PQ on COCO-C panoptic segmentation, +1.44 PSNR on SIDC image denoising, +2.15 PSNR on Rain-C image draining, and +1.01 PSNR on LoL-C low-light enhancement, respectively. Besides, for the Painter baseline, using the", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20001", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4e4ce61c726ca0756df862ffe384e5e1c539af9a850f4449e8e7facce356b43f.jpg", + "table_caption": [ + "Table 2. Test-time optimization on different modules. We use semantic segmentation on ADE20K-C for the ablation. Training only the encoder performs similarly to training both the encoder and decoder, regardless of the zero-shot and one-shot settings." + ], + "table_footnote": [], + "table_body": "
modulebrighcontdefocelastfogfrostgaussglassimpuljpegmotnpixelshotsnowzoomavg
zero-shot setting:
Encoder40.912.728.231.736.423.725.526.327.438.831.940.928.017.313.628.2
Both40.912.528.231.736.423.825.726.527.438.831.941.027.916.913.528.2
one-shot setting:
Encoder41.721.028.832.237.024.724.427.025.239.732.841.526.621.714.929.3
Both41.721.028.732.136.924.424.427.025.339.633.041.626.821.515.129.3
", + "bbox": [ + 129, + 128, + 870, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "one-shot setting does not always perform better than the zero-shot setting. The average performances are even degraded in most cases, e.g., depth estimation on NYUv2-C (0.392 A.Rel vs. 0.537 A.Rel), semantic segmentation on ADE20K-C (27.0 mIoU vs. 25.8 mIoU), panoptic segmentation on COCO-C (26.4 PQ vs. 25.3 PQ), image draining on Rain-C (17.79 PSNR vs. 17.77 PSNR). In contrast, for VICT, using the one-shot setting usually performs better than the zero-shot setting. It is worth mentioning that our zero-shot VICT can even outperform the one-shot Painter, further demonstrating the effectiveness of VICT.", + "bbox": [ + 88, + 273, + 480, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with few-shot Painter. In previous experiments, we mainly compare with a frozen Painter using either clean or corrupted examples as the task prompts. Here, we further consider a scenario where more few-shot examples from the test distribution are available. Specifically, we randomly corrupt a certain number of images in the training set, using 1, 2, 4, 8, 16, 32, and 64 shots to train Painter respectively and deploying the model in the full corrupted test sets.", + "bbox": [ + 88, + 440, + 482, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are shown in Figure 3. Both our zero-shot VICT and one-shot VICT can consistently outperform the one-shot Painter. When more few-shot examples are available for training, our VICT can still outperform Painter below a certain threshold. We observe such a threshold varies for different tasks, with high-level visual understanding tasks having a higher threshold. It is worth noting that in high-level visual understanding tasks like ADE20K-C semantic segmentation and COCO-C panoptic segmentation, both our zero-shot and one-shot variants of VICT nearly match or even outperform the 64-shot Painter. In other words, with only one or simply no labeled test sample(s), VICT can achieve similar performances as models that use significantly more labeled test samples to train. This phenomenon is rather appealing as in practice it is rare to know test distribution in advance, let alone obtaining a large number of labeled test samples.", + "bbox": [ + 91, + 578, + 482, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Further Analysis", + "text_level": 1, + "bbox": [ + 89, + 847, + 256, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of test-time optimization modules. We study the effect of optimizing different modules at test time. We con", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83e0095381d50667584687df0c79c3e0135b9cc55c44edaa81432e54ff6c821d.jpg", + "image_caption": [ + "Figure 4. Analysis on the trade-off between efficiency and accuracy. We use semantic segmentation on ADE20K-C for the ablation. VICT benefits from more training steps, while at the cost of linearly increased training time." + ], + "image_footnote": [], + "bbox": [ + 521, + 273, + 895, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sider two parameter groups: (i) the encoder, and (ii) both the encoder and decoder. We use semantic segmentation on ADE20K-C as an example task for analysis. The results are shown in Table 2. Training only the encoder performs similarly to training the entire model, regardless of the zero-shot and one-shot settings. This makes sense since the encoder acts as a feature extractor, which plays a key role in determining the representation quality. In contrast, the decoder is only responsible for mapping the latent representation back to its original resolution, which is more specialized for reconstruction but less relevant for semantics. Thus, we choose only to optimize the encoder at test time.", + "bbox": [ + 511, + 580, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of test-time optimization steps. We study the effect of test-time optimization steps in Figure 4. VICT benefits from more training steps, which keeps improving performance even after 60 steps. However, it should be noted that the runtime of our VICT is linearly increased with the number of training steps. For reference, it takes around 0.4 seconds per step per test sample on a single A100 GPU. Thus, in practice, one may decide the number of training steps according to the pre-defined cost budget for deployment.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20002", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0860baaaed2f1b52f5e0044910dd92971ea4a29f725376445dc3aedbc2e1c684.jpg", + "image_caption": [ + "Figure 5. Visualizations of test examples and predictions for six main tasks with corruptions. We visualize both zero-shot and one-shot settings for Painter and VICT. Zoom in for best view." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 478, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 544, + 272, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on main tasks. We visualize some test examples and predictions from the validation set for different tasks with corruptions, including depth estimation, semantic segmentation, instance segmentation, image denoising, image deraining, and low-light image enhancement. As shown in Figure 5, VICT can make more accurate predictions than Painter on all tasks.", + "bbox": [ + 89, + 566, + 482, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on unseen tasks. In this work, we mainly consider the distribution shifts with corruptions and conduct experiments on the tasks that are seen in training. Here, we further verify whether VICT can generalize to unseen tasks. We explore this capability via visualizations. More quantitative results are provided in the supplementary material. Figure 6 provides examples of two unseen tasks including foreground object segmentation and colorization. Both Painter and our VICT can generalize to the foreground segmentation task. This makes sense as similar segmentation tasks (e.g., semantic segmentation and instance segmentation) have been seen during the training stage of Painter. Nevertheless, our VICT further reduces noises and produces finer masks. However, Painter cannot generalize to the colorization task that is totally unseen during training, produc", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e0376eb6755de968ec4fe00585dd9737bb5ce81b14b98eeb699e9dac885fbe7.jpg", + "image_caption": [ + "Figure 6. Visualizations of test examples and predictions for unseen tasks. The visualized tasks include foreground object segmentation and colorization. Painter cannot generalize to totally unseen tasks like colorization, whereas VICT can make decent color predictions. Zoom in for best view." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing only grayscale images or even wrong predictions from other tasks (e.g., the last row). In contrast, our VICT can produce decent colorful results. This further demonstrates the potential of applying VICT for unseen tasks at test time.", + "bbox": [ + 511, + 409, + 906, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 482, + 633, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization under distribution shifts is the central theme of deep learning. In this work, we have studied the robustness of VICL models and found that the existing VICL paradigm exhibits poor generalization capability to unseen new domains. Based on this observation, we proposed a simple yet effective test-time visual in-context tuning method to adapt VICL models to a single test sample on the fly. We hope our explorations can pave the way for improving the generalizability of VICL.", + "bbox": [ + 511, + 508, + 906, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. Our study has several limitations: 1) Since we perform training for each test sample, our method is slower than the baseline applying a fixed model at test time, which is a common limitation for test-time training. Inference speed might be improved through better architectural designs, training techniques, optimizers, and hyper-parameters. However, it has not been the focus of this paper. 2) Our proposed cycle consistency supervision is a general self-supervised task. However, we cannot guarantee that it will produce useful gradients for every single test distribution. We only focus on several representation vision tasks and most popular corruption benchmarks for distribution shifts. More tasks and benchmarks can be studied for future research. One possible extension of this work is to apply similar test-time supervision in multi-modal in-context learning, where cycle consistency supervision also exists. We leave this exploration for future work.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20003", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In CVPR, 2018. 2, 4", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 2", + "[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, 2024. 2", + "[4] Pratyay Banerjee, Tejas Gokhale, and Chitta Baral. Self-supervised test-time learning for reading comprehension. In NAACL, 2021. 2", + "[5] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. In NeurIPS, 2022. 1, 2, 3", + "[6] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In AISTATS, 2022. 2", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1, 2, 3", + "[8] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In CVPR, 2022. 2", + "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3", + "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. JMLR, 2024. 2", + "[11] Samuel Dodge and Lina Karam. Quality resilient deep neural networks. arXiv preprint arXiv:1703.08119, 2017. 2", + "[12] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015. 3", + "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2", + "[14] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In ICCV, 2023. 3", + "[15] Yang Fu, Sifei Liu, Umar Iqbal, Shalini De Mello, Humphrey Shi, and Jan Kautz. Learning to track instances without video annotations. In CVPR, 2021. 2", + "[16] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 2, 3, 6" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Robert Geirhos, Carlos RM Temme, Jonas Rauber, Heiko H Schütt, Matthias Bethge, and Felix A Wichmann. Generalisation in humans and deep neural networks. In NeurIPS, 2018. 2", + "[18] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 2, 3", + "[19] Ross Girshick. Fast r-cnn. In ICCV, 2015. 4", + "[20] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In ICLR, 2021. 2", + "[21] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma, and Furu Wei. Language models are general-purpose interfaces. arXiv preprint arXiv:2206.06336, 2022. 1", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 3", + "[23] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019. 1, 2, 4", + "[24] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In CVPR, 2024. 3", + "[25] Wei Li, Jiahao Xie, and Chen Change Loy. Correlational image modeling for self-supervised visual pre-training. In CVPR, 2023. 3", + "[26] Zhi Li, Shaoshuai Shi, Bernt Schiele, and Dengxin Dai. Test-time domain adaptation for monocular depth estimation. In ICRA, 2023. 2", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4", + "[28] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In NeurIPS, 2021. 2", + "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 4", + "[30] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision-language models. In NeurIPS, 2023. 3", + "[31] Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019. 1, 2, 4", + "[32] Mirco Mutti and Aviv Tamar. Test-time regret minimization in meta reinforcement learning. In ICML, 2024. 2", + "[33] Muhammad Ferjad Naeem, Yongqin Xian, Xiaohua Zhai, Lukas Hoyer, Luc Van Gool, and Federico Tombari. Silc: Improving vision language pretraining with self-distillation. In ECCV, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20004", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 3", + "[35] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[36] Matteo Poggi, Alessio Tonioni, Fabio Tosi, Stefano Mattoccia, and Luigi Di Stefano. Continual adaptation for deep stereo. TPAMI, 2021. 2", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2", + "[38] Mattia Segu, Bernt Schiele, and Fisher Yu. Darth: holistic test-time adaptation for multiple object tracking. In ICCV, 2023. 2", + "[39] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 3", + "[40] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 2, 4", + "[41] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICLR, 2020. 2, 3, 6", + "[42] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2", + "[43] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2", + "[44] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IROS, 2017. 2", + "[45] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In CVPR, 2019. 2", + "[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2", + "[48] Yun-Yun Tsai, Chengzhi Mao, and Junfeng Yang. Convolutional visual prompt for robust visual perception. In NeurIPS, 2023. 2" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. 2", + "[50] Igor Vasiljevic, Ayan Chakrabarti, and Gregory Shakhnarovich. Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760, 2016. 2", + "[51] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 2", + "[52] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, 2023. 1, 2, 3, 4", + "[53] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In BMVC, 2018. 2, 4", + "[54] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Delving into inter-image invariance for unsupervised visual representations. IJCV, 2022. 3", + "[55] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. In ICLR, 2023. 3", + "[56] Hai Ye, Yuyang Ding, Juntao Li, and Hwee Tou Ng. Robust question answering against distribution shifts with test-time adaptation: An empirical study. In Findings of EMNLP, 2022. 2", + "[57] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In ECCV, 2020. 2, 4", + "[58] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, 2024. 3", + "[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 2", + "[60] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020. 3", + "[61] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 2", + "[62] Stephan Zheng, Yang Song, Thomas Leung, and Ian Goodfellow. Improving the robustness of deep neural networks via stability training. In CVPR, 2016. 2", + "[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2, 4" + ], + "bbox": [ + 514, + 90, + 903, + 787 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20005", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Test-Time Visual In-Context Tuning/1106c942-047b-40f5-b114-2de12258d183_model.json b/2025/Test-Time Visual In-Context Tuning/1106c942-047b-40f5-b114-2de12258d183_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3f4d351fdcd04957b0556e1a65a7775592efa1e --- /dev/null +++ b/2025/Test-Time Visual In-Context Tuning/1106c942-047b-40f5-b114-2de12258d183_model.json @@ -0,0 +1,1947 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.316, + 0.131, + 0.685, + 0.154 + ], + "angle": 0, + "content": "Test-Time Visual In-Context Tuning" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.18, + 0.861, + 0.218 + ], + "angle": 0, + "content": "Jiahao Xie\\(^{1,2}\\), Alessio Tonioni\\(^{3}\\), Nathalie Rauschmayr\\(^{3}\\), Federico Tombari\\(^{3}\\), Bernt Schiele\\(^{1,2}\\), \\(^{1}\\)Max Planck Institute for Informatics, SIC \\(^{2}\\)VIA Research Center \\(^{3}\\)Google" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.219, + 0.825, + 0.235 + ], + "angle": 0, + "content": "{jxie, schiele}@mpi-inf.mpg.de {alessiot, rauschmayr, tombari}@google.com" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.266, + 0.907, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.379, + 0.907, + 0.45 + ], + "angle": 0, + "content": "Figure 1. Test-time visual in-context tuning (VICT) on six representative vision tasks under distribution shifts. We benchmark the robustness of VICL with 15 common corruptions adopted in [23, 31], and report the averaged performance across all corruptions. Existing VICL models like Painter exhibit poor generalization capability to unseen new domains when the task prompts come from the training distribution (i.e., zero-shot). Performances are even worse when given task prompts from the test distribution (i.e., one-shot). By performing VICT at test time, we can significantly improve Painter in both zero-shot and one-shot manners." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.461, + 0.327, + 0.478 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.486, + 0.781 + ], + "angle": 0, + "content": "Visual in-context learning (VICL), as a new paradigm in computer vision, allows the model to rapidly adapt to various tasks with only a handful of prompts and examples. While effective, the existing VICL paradigm exhibits poor generalizability under distribution shifts. In this work, we propose test-time Visual In-Context Tuning (VICT), a method that can adapt VICL models on the fly with a single test sample. Specifically, we flip the role between the task prompts and the test sample and use a cycle consistency loss to reconstruct the original task prompt output. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts. Extensive experiments on six representative vision tasks ranging from high-level visual understanding to low-level image processing, with 15 common corruptions, demonstrate that our VICT can improve the generalizability of VICL to unseen new domains. In addition, we show the potential of applying VICT for unseen tasks at test time. Code: https://github.com/Jiahao000/VICT." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.222, + 0.83 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Following the success of in-context learning (ICL) [2, 7, 21] in natural language processing (NLP), visual in-context learning (VICL) [5, 52] has shown promising performance in developing generalist models for vision tasks. Inspired" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.629 + ], + "angle": 0, + "content": "by prompting in NLP that defines tasks using language sequences as a general interface, existing VICL works [5, 52] use images themselves as a natural interface for general-purpose visual perception. They formulate VICL as an image inpainting task, i.e., given an input-output example describing a specific task (i.e., the prompt) as well as an input image, they are assembled in a grid and the problem is casted as inpainting the missing part of the grid (i.e., the prediction) consistently with the given task prompts. This enables adapting a pre-trained vision model to various downstream tasks with only a handful of prompts and examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.827 + ], + "angle": 0, + "content": "By default, current VICL models are frozen during deployment. However, their performance might suffer in real-world deployment as the test distribution usually changes and deviates from the training one. This naturally raises a question: what is the generalizability of VICL models under distribution shifts? In this work, we investigate this aspect focusing on controllable shifts due to image corruptions. We observe that the existing VICL paradigm like Painter [52] exhibits poor generalizability to unseen new domains when the task prompt comes from the training distribution, as shown in Figure 1. Surprisingly, performances are even worse when given input-output task prompts belonging to the same distribution as the test input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Driven by this observation, we propose to rely on test-time visual in-context tuning (VICT) to adapt VICL models on the fly to unseen distributions using only the given test sample. The motivation is that each test input offers a hint about the test distribution. Thus, we modify the VICL" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19996" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.333 + ], + "angle": 0, + "content": "model at test time to make full use of this hint by setting up a one-sample learning problem. Specifically, given input-output task prompts and the test input, we first use the VICL model to inpaint the output image. We then flip the role between the task prompts and the test sample, i.e., we treat the predicted test output as the prompt to the model and reconstruct the original output of task prompts. This allows us to tune the parameters of the whole model in a self-supervised manner that can be applied to arbitrary tasks. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts conditioned on its in-context inference. Such a cycle consistency supervision signal naturally exists in the context of VICL without requiring any additional training data or annotations, thus making it an appealing self-supervisory task for test-time visual in-context training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.482, + 0.56 + ], + "angle": 0, + "content": "We explore VICT in two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). Without loss of generality, we consider Painter [52] as our VICL model, for its simplicity in design and its wide applicability. As shown in Figure 1, our simple method leads to substantial improvements across 15 common corruptions [23, 31] on six representative vision tasks ranging from high-level visual understanding to low-level image processing, including depth estimation on NYUv2 [40], semantic segmentation on ADE20K [63], panoptic segmentation on COCO [27], image denoising on SIDD [1], image deraining on the merged deraining dataset [57], and low-light image enhancement on LoL [53]." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.561, + 0.451, + 0.574 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.575, + 0.483, + 0.62 + ], + "angle": 0, + "content": "1) We propose a new cycle consistency task for test-time visual in-context tuning. To the best of our knowledge, we are the first to perform test-time training for VICL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.711 + ], + "angle": 0, + "content": "2) We contribute the first study on the generalizability of VICL under distribution shifts. We observe that the existing VICL paradigm exhibits poor generalizability to unseen new domains. Such a phenomenon can hardly be recovered even given input-output task prompts with the same distribution as the test input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.483, + 0.817 + ], + "angle": 0, + "content": "3) We conduct extensive experiments on six representative vision tasks across 15 common corruptions. VICT significantly improves Painter in both zero-shot and one-shot manners. Our zero-shot or one-shot VICT can even outperform Painter trained with more few-shot corrupted examples. We also explore the potential of applying VICT for unseen tasks at test time, further demonstrating its promise." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.233, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In-context learning. In-context learning has been extensively explored in the NLP literature after the introduction of large language models. Seminal works in the field" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "like GPT-3 [7] go as far as claiming that language models are indeed few-shot learners. All recent language models like [13, 46, 47] have the ability to adapt their behavior based on few-shot examples provided in their context window or directly to follow simple user instructions [10]. Similar concepts have been extended to multimodal models with seminal works like Flamingo [2] showing few-shot capabilities by providing examples as interleaved text and images in the context window. Some of these capabilities are now available for extremely large commercial multimodal large language models [35, 42, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.905, + 0.517 + ], + "angle": 0, + "content": "Apart from using languages as the general interface, a recent line of work introduces purely visual in-context learning (VICL) by training vision generalist models that can perform arbitrary visual tasks following one or few visual examples provided at inference time. Two representative works developed in parallel are: (i) MAE-VQGAN [5], which trains a variant of MAE [22] on a dataset of figures extracted from academic papers, and (ii) Painter [52], which uses a similar idea but trains its model on a set of standard academic benchmark datasets. More recently, this strategy has been expanded in LVM [3], where even more visual datasets are collected to train an autoregressive generative model. We base our work on Painter [52] due to its simplicity in design and its wide applicability. It is the one that provides both the available code and model weights, and has already been extensively evaluated across several standard benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.905, + 0.717 + ], + "angle": 0, + "content": "Generalization under distribution shifts. Machine learning models suffer from performance drops when tested on a data distribution different from the one they are trained on [17, 49, 50, 61]. The lower the drop, the more we define a model robust or able to generalize to distribution shifts. Common strategies to increase model robustness include using heavy data augmentations [11, 23, 44, 62] or extremely large training distributions going up to web scale [33, 37, 59]. Nevertheless, most of the experiments are done in a classification setting. To the best of our knowledge, there is no previous work exploring the robustness of VICL models. We are the first to counteract performance drops in VICL under distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Test-time training. An alternative way of counteracting performance drops due to distribution shifts is test-time training [6, 8, 16, 28, 41, 48, 51]. This accounts to unfreezing the model at test time and fine-tuning it on the target distribution (or a single sample from it) through self-supervision. Early works propose this paradigm using self-supervised pretext tasks (e.g., rotation prediction [18]) to explicitly improve generalization under image corruptions [41] or improve in specific tasks where self-supervised losses can be clearly defined like depth estimation [26, 36, 45], reinforcement learning [20, 32], tracking [15, 38], and NLP [4, 56]. More recently, the field" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "19997" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.308, + 0.908, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Overview of our VICT pipeline. Given a pair of task prompts \\((x,y)\\) and a test input image \\(x_{t}\\), we first construct a four-cell grid-like image canvas \\(I = (x,y,x_{t},\\emptyset)\\), with an empty cell at the bottom right. We then feed \\(I\\) into the VICL model (e.g., Painter) to predict the test output \\(\\hat{y}_t\\). Afterward, we flip the role between input-output task prompts and input-output test samples, i.e., we provide the predicted \\(\\hat{y}_t\\) as the prompt to the model, recreating a new four-cell grid-like image canvas \\(I' = (x,\\emptyset,x_{t},\\hat{y}_{t})\\), with an empty cell at the top right. The new \\(I'\\) is fed into the same model to predict the task prompt output \\(\\hat{y}\\). We finally optimize the model by minimizing the distance between \\(\\hat{y}\\) and \\(y\\) via a standard regression loss." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.412, + 0.486, + 0.565 + ], + "angle": 0, + "content": "has realized that MAE [22] provides a very strong self-supervised pretext task and has therefore been employed in the context of test-time training [16]. Finally, some recent works [14, 24, 30, 39, 58] also employ the test-time training paradigm for vision-language models to improve their generalization. To the best of our knowledge, ours is the first work to explore test-time training applied to VICL models, where previous self-supervised pretext tasks [9, 12, 18, 22, 25, 34, 54, 55, 60] cannot be employed due to their constraints on the single-image context." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.226, + 0.592 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.676 + ], + "angle": 0, + "content": "Our visual in-context tuning (VICT) is a simple yet effective test-time training approach to adapt visual in-context learning (VICL) models on the fly. In this section, we first introduce some preliminaries on VICL in Section 3.1. We then detail our VICT pipeline in Section 3.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.34, + 0.699 + ], + "angle": 0, + "content": "3.1. Visual In-Context Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In-context learning is a new paradigm originating from large language models such as GPT-3 [7] in NLP. Unlike traditional learning paradigms, in-context learning formulates different NLP tasks as text completion tasks and makes predictions conditioned on one or many support examples provided to the model in the context window. Extending this paradigm to tasks requiring a dense visual output in the form of images is nontrivial and has not been addressed in the literature for a long time. Recently, [5, 52] formulate VICL as an image inpainting task by combining images and labels into a grid-like new image and using masked image modeling for pre-training, granting the models with the in-context learning ability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.487 + ], + "angle": 0, + "content": "Formally, let \\( S = \\{(x_i, y_i)\\}_{i=1}^N \\) denote the set of support input-output examples (a.k.a., task prompts) where \\( x \\) is an image and \\( y \\) is its visual label (e.g., a segmentation mask).1 Given \\( S \\) and a new test image \\( x_t \\) as input, VICL can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.656, + 0.501, + 0.907, + 0.517 + ], + "angle": 0, + "content": "\\[\ny _ {t} = f _ {\\theta} (S, x _ {t}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.62 + ], + "angle": 0, + "content": "where \\( f_{\\theta}(*) \\) is the VICL model parametrized with \\( \\theta \\), \\( y_{t} \\) is the corresponding output label of \\( x_{t} \\). Usually, we assume \\( S \\) and \\( x_{t} \\) are drawn from the same distribution. However, this is rarely the case for real-world deployment. Therefore, in this work, we mainly consider the scenario where a distribution shift like an image corruption occurs for \\( x_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.906, + 0.725 + ], + "angle": 0, + "content": "In the experiment section, we will show how the presence of such distribution shifts significantly degrades the performance of the VICL models. Moreover, we will additionally show how simple solutions like providing in-context examples from the target (corrupted) distribution would not fix the problem due to the limited generalization ability of the current VICL models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.735, + 0.828, + 0.751 + ], + "angle": 0, + "content": "3.2. Test-Time Visual In-Context Tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.757, + 0.907, + 0.879 + ], + "angle": 0, + "content": "In this work, we propose to leverage test-time training [41] to counteract the distribution shifts between training and testing data, with the goal of making VICL models more robust. We argue that VICL models are particularly amenable to be optimized at test time on arbitrary tasks thanks to the availability of the \\(N\\) shots provided as context. The core intuition of our method is to use the provided in-context examples to define a self-supervised loss that can be used at" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.821, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Following previous works, we set \\(N = 1\\) in practice." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19998" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.091, + 0.479, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of VICT in a PyTorch-like style." + }, + { + "type": "code", + "bbox": [ + 0.093, + 0.116, + 0.47, + 0.373 + ], + "angle": 0, + "content": "f: VICL model (e.g., Painter) \n# theta_0: pre-trained model weights \n# x: task prompt input \n# y: task prompt output \n# mask_token: mask token for the empty cell \n# steps: number of test-time optimization steps \nfor x_t in loader: # load a test sample x_t \nf.params = theta_0 # initialize \n# test-time optimization \nfor t in range(steps): \n # predict the test sample output \n y_t_pred = f((x, y, x_t, mask_token)) \n # predict the task prompt output \n y_pred = f((x, mask_token, x_t, y_t_pred)) \n # regression loss \n loss = SmoothL1Loss(y_pred, y) \n # update model \n loss/backward() \nupdate(f.params) \n# inference with the updated model \ny_t_pred = f((x, y, x_t, mask_token))" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.482, + 0.441 + ], + "angle": 0, + "content": "test time to train any VICL models on the fly on any task. The overall pipeline of our VICT is illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Specifically, given a pair of task prompts \\((x,y)\\) from \\(S\\) and a test input \\(x_{t}\\), we construct a grid-like image canvas with four cells, denoted by an ordered quadruple \\(I = (x,y,x_{t},\\emptyset)\\), where \\(\\emptyset\\) is an empty cell. We then feed \\(I\\) into the VICL model \\(f_{\\theta}\\) to predict the test output \\(\\hat{y}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.529, + 0.483, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {t} = f _ {\\theta} (I) = f _ {\\theta} \\left(\\left(x, y, x _ {t}, \\varnothing\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.633 + ], + "angle": 0, + "content": "Afterward, with the obtained test output, we flip the role between input-output task prompts and input-output test samples, creating a new four-cell grid-like image canvas with the order of \\( I' = (x, \\varnothing, x_t, \\hat{y}_t) \\). The new \\( I' \\) is fed into \\( f_{\\theta} \\) again to predict the task prompt output \\( \\hat{y} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.644, + 0.483, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = f _ {\\theta} \\left(I ^ {\\prime}\\right) = f _ {\\theta} \\left(\\left(x, \\varnothing , x _ {t}, \\hat {y} _ {t}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.718 + ], + "angle": 0, + "content": "We then use a simple regression loss in pixel space to optimize the model by minimizing the distance between \\(\\hat{y}\\) and \\(y\\):" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.718, + 0.482, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\theta_ {x _ {t}} = \\arg \\min _ {\\theta} \\mathcal {L} (\\hat {y}, y) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.483, + 0.807 + ], + "angle": 0, + "content": "In practice, we follow [52] to use smooth-\\(\\ell_1\\) [19] loss. Using this loss we can optimize the parameters \\(\\theta\\) of the model for a small number of steps for each \\(x_t\\) and, eventually, make a prediction on \\(x_t\\) as \\(f_{\\theta_{x_t}}((x,y,x_t,\\emptyset))\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.806, + 0.484, + 0.867 + ], + "angle": 0, + "content": "In our formulation, the gradient-based optimization for Equation 4 always starts from the initial pre-trained model weights \\(\\theta_0\\) for each test input. When a new test input arrives, we discard \\(\\theta_{x_t}\\) and reset the weights to \\(\\theta_0\\). We follow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "this strategy to not have any assumption on the test inputs and treat them independently at test time, i.e., we do not assume that they come from the same distribution. Our generic formulation is completely agnostic on the VICL models considered, on the type of tasks being solved, and on the type of in-context samples provided to it." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.183, + 0.831, + 0.199 + ], + "angle": 0, + "content": "The pseudo-code of VICT is in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.216, + 0.646, + 0.233 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.242, + 0.731, + 0.258 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Tasks and datasets. Following [52], we evaluate the model performance on six representative vision tasks ranging from high-level visual understanding to low-level image processing, which includes depth estimation on NYUv2 [40], semantic segmentation on ADE20K [63], panoptic segmentation on COCO [27], image denoising on SIDD [1], image draining on the merged draining datasets [57], and low-light image enhancement on LoL [53]. To evaluate the robustness to distribution shifts of the model on these datasets, we follow the setup in [23, 31] to corrupt the aforementioned datasets and simulate hard distribution shifts. The corrupted datasets contain 15 types of corruptions, covering noise (gaussian noise, impulse noise, shot noise), blur (defocus blur, glass blur, motion blur, zoom blur), weather (fog, frost, snow), and digital (brightness, contrast, elastic transform, JPEG compression, pixelate) categories. Each type of corruption has five levels of severity. We denote the corrupted datasets as NYUv2-C, ADE20K-C, COCO-C, SIDD-C, Rain-C, and LoL-C, respectively. Due to the space constraints, the results in the main text are limited to the most severe level (i.e., level 5) that corresponds to the strongest distribution shift. We provide the results on the other four levels in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Baselines. Without loss of generality, we consider Painter [52] as our VICL model for its simplicity in design and its wide applicability. We study VICT in two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). For the one-shot setting, apart from comparing with the frozen Painter using one-shot corrupted examples as task prompts, we also consider a baseline that further trains Painter with the one-shot corrupted samples using the same pre-training objective as in [52]. We also consider training Painter with more few-shot examples using the same settings to examine to what extent VICT can outperform a larger number of few-shot fine-tuning of Painter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Training details. In all experiments, we use the pre-trained Painter based on ViT-Large provided by the authors of [52]. We perform VICT using an AdamW [29] optimizer, with betas as (0.9, 0.999), a weight decay of 0, a batch size of 1," + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2The image order is top left, top right, bottom left, and bottom right, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19999" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.145 + ], + "angle": 0, + "content": "Table 1. System-level comparison on six representative vision tasks across 15 corruptions. Results are on corruption level 5. We consider two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). \"avg\" denotes the averaged results over 15 corruptions." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.156, + 0.899, + 0.879 + ], + "angle": 0, + "content": "
methodbrighcontdefocelastfogfrostgaussglassimpuljpegmotnpixelshotsnowzoomavg
(a) depth estimation NYUv2-C (A.Rel ↓)
zero-shot setting:
Painter0.1290.2150.7120.1090.1290.5360.2000.6120.1890.3860.1670.1870.1421.9510.2090.392
VICT0.1080.2160.6310.1080.1330.5760.1910.5410.1810.3100.1390.1340.1401.8560.2100.365
one-shot setting:
Painter0.1260.2850.7430.1090.1320.8530.9010.6220.9010.3920.1740.1940.4401.9640.2120.537
VICT0.0970.1930.1800.1070.1280.2270.2410.2100.2780.1590.1210.1140.1500.6620.2140.205
(b) semantic segmentation ADE20K-C (mIoU ↑)
zero-shot setting:
Painter40.411.727.931.235.023.324.825.626.138.631.840.425.99.413.127.0
VICT40.912.728.231.736.423.725.526.327.438.831.940.928.017.313.628.2
one-shot setting:
Painter40.911.727.531.335.022.119.725.820.338.531.340.323.26.513.325.8
VICT41.721.028.832.237.024.724.427.025.239.732.841.526.621.714.929.3
(c) panoptic segmentation COCO-C (PQ ↑)
zero-shot setting:
Painter38.115.525.530.033.326.024.923.825.531.927.634.926.319.812.526.4
VICT38.715.725.830.333.926.725.324.325.932.427.735.426.721.912.726.9
one-shot setting:
Painter38.114.224.830.033.125.221.523.722.031.527.534.623.817.212.125.3
VICT38.616.925.430.433.826.624.024.624.532.327.435.025.224.912.426.8
(d) denoising SIDD-C (PSNR ↑)
zero-shot setting:
Painter8.7225.4932.4833.2016.868.7917.5033.1817.5023.1829.4734.8425.279.8629.8023.08
VICT8.8625.6532.9132.9617.008.8417.5833.0017.5024.8629.6834.7625.219.7529.6723.22
one-shot setting:
Painter9.1225.1631.9933.2316.689.2417.9933.1418.3322.7329.3034.8724.6910.5029.7623.12
VICT13.9425.3232.4134.4516.4511.1520.1332.9619.9326.8030.1135.0124.9314.9029.9524.56
(e) deraining Rain-C (PSNR ↑)
zero-shot setting:
Painter10.7414.1519.5420.1013.9112.4919.8519.8020.1624.8618.5222.8120.1111.8917.9317.79
VICT10.8914.2419.5820.3113.9412.6220.4219.8320.6724.9718.5623.0520.4712.0517.9817.97
one-shot setting:
Painter11.0413.9519.6120.1513.8212.8719.5119.8719.5824.8518.6222.8819.6812.1617.9517.77
VICT17.3814.6920.3621.4414.6717.6922.3320.4822.2425.3119.0523.9221.3719.5318.2819.92
(f) low-light enhancement LoL-C (PSNR ↑)
zero-shot setting:
Painter16.2613.4318.8019.8311.2411.6212.5018.8212.9115.8718.0121.1217.7513.7716.6015.90
VICT16.3913.5719.2419.8411.3711.6412.5119.2313.0016.0018.1121.2019.1613.9217.0616.15
one-shot setting:
Painter16.7413.5018.8519.8311.0211.7613.1418.9513.7415.8418.1521.1216.5413.1916.5915.93
VICT18.1915.0318.9919.8811.7612.0114.9519.2214.9916.1218.2321.0919.8416.9616.9116.94
" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20000" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.365, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.091, + 0.631, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.092, + 0.897, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.253, + 0.365, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.253, + 0.631, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.253, + 0.897, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.421, + 0.908, + 0.465 + ], + "angle": 0, + "content": "Figure 3. Comparison with few-shot Painter on six vision tasks with corruptions. We randomly corrupt a certain number of images in the training set, using 1, 2, 4, 8, 16, 32, and 64 shots for training and deploying the model in the full corrupted test sets. We report the final results averaged across 15 corruptions. Our zero-shot or one-shot VICT can outperform Painter trained with more few-shot examples." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.479, + 0.486, + 0.782 + ], + "angle": 0, + "content": "and a fixed learning rate of 1e-6. Unless otherwise specified, we train each test sample for 60 steps. The choice of 60 steps is purely computational and more steps are likely to further improve performance, judging from the positive trend observed in Figure 4. Note that we do not apply any data augmentations for VICT as many data augmentations are in fact distribution shifts in our evaluation benchmarks. Training with them is analogous to training on the test distributions. Therefore, to solely study the generalization ability to new test distributions, we purposely choose not to use any data augmentations, even though they could improve our results at face value. Every iteration of VICT is performed on the same concatenated grid-like image, with each subimage resized as \\(448 \\times 448\\), which is the same as what we later use for in-context inference. During VICT, we only optimize the encoder weights. We have experimented with training both the encoder and the decoder and found that the difference is negligible, which is also consistent with the observations in [16, 41]. A detailed comparison is provided in the experiment section." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.231, + 0.803 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "System-level comparison. We compare VICT with Painter under 15 common corruptions as introduced in Section 4.1. We perform experiments on six representative vision tasks covering high-level visual understanding and low-level image processing, which includes depth estimation on NYUv2-C, semantic segmentation on ADE20K-C, panop-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.525 + ], + "angle": 0, + "content": "tic segmentation on COCO-C, image denoising on SIDD-C, image draining on Rain-C, and low-light image enhancement on LoL-C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.909, + 0.76 + ], + "angle": 0, + "content": "Table 1 reports the results. First of all, we observe that Painter exhibits poor generalization ability under common corruptions. For example, on ADE20K-C semantic segmentation, Painter only achieves an average performance of \\(27.0\\mathrm{mIoU}\\), which is \\(22.9\\mathrm{mIoU}\\) lower than that on the clean validation set (49.9 mIoU as reported in Figure 1). Nevertheless, our VICT outperforms Painter by clear margins in both zero-shot and one-shot settings across different tasks. When it comes to the zero-shot setting, VICT achieves an average of -0.027 A.Rel on NYUv2-C depth estimation, +1.2 mIoU on ADE20K-C semantic segmentation, +0.5 PQ on COCO-C panoptic segmentation, +0.14 PSNR on SIDD-C image denosing, +0.18 PSNR on Rain-C image draining, and +0.25 PSNR on LoL-C low-light enhancement, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "The performance gains of VICT over Painter are further increased when we shift from the zero-shot setting to the one-shot counterpart. Specifically, VICT achieves an average of -0.332 A.Rel on NYUv2-C depth estimation, +3.5 mIoU on ADE20K-C semantic segmentation, +1.5 PQ on COCO-C panoptic segmentation, +1.44 PSNR on SIDC image denoising, +2.15 PSNR on Rain-C image draining, and +1.01 PSNR on LoL-C low-light enhancement, respectively. Besides, for the Painter baseline, using the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20001" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Test-time optimization on different modules. We use semantic segmentation on ADE20K-C for the ablation. Training only the encoder performs similarly to training both the encoder and decoder, regardless of the zero-shot and one-shot settings." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.129, + 0.871, + 0.25 + ], + "angle": 0, + "content": "
modulebrighcontdefocelastfogfrostgaussglassimpuljpegmotnpixelshotsnowzoomavg
zero-shot setting:
Encoder40.912.728.231.736.423.725.526.327.438.831.940.928.017.313.628.2
Both40.912.528.231.736.423.825.726.527.438.831.941.027.916.913.528.2
one-shot setting:
Encoder41.721.028.832.237.024.724.427.025.239.732.841.526.621.714.929.3
Both41.721.028.732.136.924.424.427.025.339.633.041.626.821.515.129.3
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.274, + 0.482, + 0.44 + ], + "angle": 0, + "content": "one-shot setting does not always perform better than the zero-shot setting. The average performances are even degraded in most cases, e.g., depth estimation on NYUv2-C (0.392 A.Rel vs. 0.537 A.Rel), semantic segmentation on ADE20K-C (27.0 mIoU vs. 25.8 mIoU), panoptic segmentation on COCO-C (26.4 PQ vs. 25.3 PQ), image draining on Rain-C (17.79 PSNR vs. 17.77 PSNR). In contrast, for VICT, using the one-shot setting usually performs better than the zero-shot setting. It is worth mentioning that our zero-shot VICT can even outperform the one-shot Painter, further demonstrating the effectiveness of VICT." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.483, + 0.576 + ], + "angle": 0, + "content": "Comparison with few-shot Painter. In previous experiments, we mainly compare with a frozen Painter using either clean or corrupted examples as the task prompts. Here, we further consider a scenario where more few-shot examples from the test distribution are available. Specifically, we randomly corrupt a certain number of images in the training set, using 1, 2, 4, 8, 16, 32, and 64 shots to train Painter respectively and deploying the model in the full corrupted test sets." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.579, + 0.483, + 0.835 + ], + "angle": 0, + "content": "The results are shown in Figure 3. Both our zero-shot VICT and one-shot VICT can consistently outperform the one-shot Painter. When more few-shot examples are available for training, our VICT can still outperform Painter below a certain threshold. We observe such a threshold varies for different tasks, with high-level visual understanding tasks having a higher threshold. It is worth noting that in high-level visual understanding tasks like ADE20K-C semantic segmentation and COCO-C panoptic segmentation, both our zero-shot and one-shot variants of VICT nearly match or even outperform the 64-shot Painter. In other words, with only one or simply no labeled test sample(s), VICT can achieve similar performances as models that use significantly more labeled test samples to train. This phenomenon is rather appealing as in practice it is rare to know test distribution in advance, let alone obtaining a large number of labeled test samples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.258, + 0.864 + ], + "angle": 0, + "content": "4.3. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of test-time optimization modules. We study the effect of optimizing different modules at test time. We con" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.275, + 0.897, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.5, + 0.908, + 0.556 + ], + "angle": 0, + "content": "Figure 4. Analysis on the trade-off between efficiency and accuracy. We use semantic segmentation on ADE20K-C for the ablation. VICT benefits from more training steps, while at the cost of linearly increased training time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.763 + ], + "angle": 0, + "content": "sider two parameter groups: (i) the encoder, and (ii) both the encoder and decoder. We use semantic segmentation on ADE20K-C as an example task for analysis. The results are shown in Table 2. Training only the encoder performs similarly to training the entire model, regardless of the zero-shot and one-shot settings. This makes sense since the encoder acts as a feature extractor, which plays a key role in determining the representation quality. In contrast, the decoder is only responsible for mapping the latent representation back to its original resolution, which is more specialized for reconstruction but less relevant for semantics. Thus, we choose only to optimize the encoder at test time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effect of test-time optimization steps. We study the effect of test-time optimization steps in Figure 4. VICT benefits from more training steps, which keeps improving performance even after 60 steps. However, it should be noted that the runtime of our VICT is linearly increased with the number of training steps. For reference, it takes around 0.4 seconds per step per test sample on a single A100 GPU. Thus, in practice, one may decide the number of training steps according to the pre-defined cost budget for deployment." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20002" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.48, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.481, + 0.484, + 0.525 + ], + "angle": 0, + "content": "Figure 5. Visualizations of test examples and predictions for six main tasks with corruptions. We visualize both zero-shot and one-shot settings for Painter and VICT. Zoom in for best view." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.274, + 0.56 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Results on main tasks. We visualize some test examples and predictions from the validation set for different tasks with corruptions, including depth estimation, semantic segmentation, instance segmentation, image denoising, image deraining, and low-light image enhancement. As shown in Figure 5, VICT can make more accurate predictions than Painter on all tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results on unseen tasks. In this work, we mainly consider the distribution shifts with corruptions and conduct experiments on the tasks that are seen in training. Here, we further verify whether VICT can generalize to unseen tasks. We explore this capability via visualizations. More quantitative results are provided in the supplementary material. Figure 6 provides examples of two unseen tasks including foreground object segmentation and colorization. Both Painter and our VICT can generalize to the foreground segmentation task. This makes sense as similar segmentation tasks (e.g., semantic segmentation and instance segmentation) have been seen during the training stage of Painter. Nevertheless, our VICT further reduces noises and produces finer masks. However, Painter cannot generalize to the colorization task that is totally unseen during training, produc" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.32, + 0.908, + 0.39 + ], + "angle": 0, + "content": "Figure 6. Visualizations of test examples and predictions for unseen tasks. The visualized tasks include foreground object segmentation and colorization. Painter cannot generalize to totally unseen tasks like colorization, whereas VICT can make decent color predictions. Zoom in for best view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.471 + ], + "angle": 0, + "content": "ing only grayscale images or even wrong predictions from other tasks (e.g., the last row). In contrast, our VICT can produce decent colorful results. This further demonstrates the potential of applying VICT for unseen tasks at test time." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.634, + 0.499 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Generalization under distribution shifts is the central theme of deep learning. In this work, we have studied the robustness of VICL models and found that the existing VICL paradigm exhibits poor generalization capability to unseen new domains. Based on this observation, we proposed a simple yet effective test-time visual in-context tuning method to adapt VICL models to a single test sample on the fly. We hope our explorations can pave the way for improving the generalizability of VICL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitations and future work. Our study has several limitations: 1) Since we perform training for each test sample, our method is slower than the baseline applying a fixed model at test time, which is a common limitation for test-time training. Inference speed might be improved through better architectural designs, training techniques, optimizers, and hyper-parameters. However, it has not been the focus of this paper. 2) Our proposed cycle consistency supervision is a general self-supervised task. However, we cannot guarantee that it will produce useful gradients for every single test distribution. We only focus on several representation vision tasks and most popular corruption benchmarks for distribution shifts. More tasks and benchmarks can be studied for future research. One possible extension of this work is to apply similar test-time supervision in multi-modal in-context learning, where cycle consistency supervision also exists. We leave this exploration for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20003" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.156 + ], + "angle": 0, + "content": "[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In CVPR, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.226 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[4] Pratyay Banerjee, Tejas Gokhale, and Chitta Baral. Self-supervised test-time learning for reading comprehension. In NAACL, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[5] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. In NeurIPS, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[6] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In AISTATS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.482, + 0.527 + ], + "angle": 0, + "content": "[8] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.573, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. JMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.63, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[11] Samuel Dodge and Lina Karam. Quality resilient deep neural networks. arXiv preprint arXiv:1703.08119, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[12] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[14] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[15] Yang Fu, Sifei Liu, Umar Iqbal, Shalini De Mello, Humphrey Shi, and Jan Kautz. Learning to track instances without video annotations. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[16] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[17] Robert Geirhos, Carlos RM Temme, Jonas Rauber, Heiko H Schütt, Matthias Bethge, and Felix A Wichmann. Generalisation in humans and deep neural networks. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[18] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.813, + 0.208 + ], + "angle": 0, + "content": "[19] Ross Girshick. Fast r-cnn. In ICCV, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.21, + 0.905, + 0.264 + ], + "angle": 0, + "content": "[20] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.268, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[21] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma, and Furu Wei. Language models are general-purpose interfaces. arXiv preprint arXiv:2206.06336, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.325, + 0.905, + 0.365 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.904, + 0.409 + ], + "angle": 0, + "content": "[23] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[24] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[25] Wei Li, Jiahao Xie, and Chen Change Loy. Correlational image modeling for self-supervised visual pre-training. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[26] Zhi Li, Shaoshuai Shi, Bernt Schiele, and Dengxin Dai. Test-time domain adaptation for monocular depth estimation. In ICRA, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[28] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.658, + 0.904, + 0.685 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[30] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision-language models. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.811 + ], + "angle": 0, + "content": "[31] Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[32] Mirco Mutti and Aviv Tamar. Test-time regret minimization in meta reinforcement learning. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[33] Muhammad Ferjad Naeem, Yongqin Xian, Xiaohua Zhai, Lukas Hoyer, Luc Van Gool, and Federico Tombari. Silc: Improving vision language pretraining with self-distillation. In ECCV, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20004" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.134 + ], + "angle": 0, + "content": "[34] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[35] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[36] Matteo Poggi, Alessio Tonioni, Fabio Tosi, Stefano Mattoccia, and Luigi Di Stefano. Continual adaptation for deep stereo. TPAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[38] Mattia Segu, Bernt Schiele, and Fisher Yu. Darth: holistic test-time adaptation for multiple object tracking. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.377 + ], + "angle": 0, + "content": "[39] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.42 + ], + "angle": 0, + "content": "[40] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[41] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICLR, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.483, + 0.547 + ], + "angle": 0, + "content": "[42] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[43] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[44] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IROS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[45] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[48] Yun-Yun Tsai, Chengzhi Mao, and Junfeng Yang. Convolutional visual prompt for robust visual perception. In NeurIPS, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[49] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[50] Igor Vasiljevic, Ayan Chakrabarti, and Gregory Shakhnarovich. Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[51] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.235, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[52] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[53] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In BMVC, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.321, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[54] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Delving into inter-image invariance for unsupervised visual representations. IJCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.363, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[55] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.407, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[56] Hai Ye, Yuyang Ding, Juntao Li, and Hwee Tou Ng. Robust question answering against distribution shifts with test-time adaptation: An empirical study. In Findings of EMNLP, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.462, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[57] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In ECCV, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.519, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[58] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.562, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.604, + 0.905, + 0.647 + ], + "angle": 0, + "content": "[60] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.648, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[61] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.703, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[62] Stephan Zheng, Yang Song, Thomas Leung, and Ian Goodfellow. Improving the robustness of deep neural networks via stability training. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.747, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20005" + } + ] +] \ No newline at end of file diff --git a/2025/Test-Time Visual In-Context Tuning/1106c942-047b-40f5-b114-2de12258d183_origin.pdf b/2025/Test-Time Visual In-Context Tuning/1106c942-047b-40f5-b114-2de12258d183_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..43b31186c4ffb0f3cc9c1e7b89fe0f1dfbed99c2 --- /dev/null +++ b/2025/Test-Time Visual In-Context Tuning/1106c942-047b-40f5-b114-2de12258d183_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5863ab348568a23bb40bd259b92c5888a7fc9366fa765478cfbadab075d42721 +size 6566250 diff --git a/2025/Test-Time Visual In-Context Tuning/full.md b/2025/Test-Time Visual In-Context Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cdbbac6d2944dff921ea19d1c0d16272371a2ab9 --- /dev/null +++ b/2025/Test-Time Visual In-Context Tuning/full.md @@ -0,0 +1,273 @@ +# Test-Time Visual In-Context Tuning + +Jiahao Xie $^{1,2}$ , Alessio Tonioni $^{3}$ , Nathalie Rauschmayr $^{3}$ , Federico Tombari $^{3}$ , Bernt Schiele $^{1,2}$ , $^{1}$ Max Planck Institute for Informatics, SIC $^{2}$ VIA Research Center $^{3}$ Google + +{jxie, schiele}@mpi-inf.mpg.de {alessiot, rauschmayr, tombari}@google.com + +![](images/dc83c086a4cc0252879660d3cfa511bf47deca6f2878ba337d9a0e0a35b4e1c7.jpg) +Figure 1. Test-time visual in-context tuning (VICT) on six representative vision tasks under distribution shifts. We benchmark the robustness of VICL with 15 common corruptions adopted in [23, 31], and report the averaged performance across all corruptions. Existing VICL models like Painter exhibit poor generalization capability to unseen new domains when the task prompts come from the training distribution (i.e., zero-shot). Performances are even worse when given task prompts from the test distribution (i.e., one-shot). By performing VICT at test time, we can significantly improve Painter in both zero-shot and one-shot manners. + +# Abstract + +Visual in-context learning (VICL), as a new paradigm in computer vision, allows the model to rapidly adapt to various tasks with only a handful of prompts and examples. While effective, the existing VICL paradigm exhibits poor generalizability under distribution shifts. In this work, we propose test-time Visual In-Context Tuning (VICT), a method that can adapt VICL models on the fly with a single test sample. Specifically, we flip the role between the task prompts and the test sample and use a cycle consistency loss to reconstruct the original task prompt output. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts. Extensive experiments on six representative vision tasks ranging from high-level visual understanding to low-level image processing, with 15 common corruptions, demonstrate that our VICT can improve the generalizability of VICL to unseen new domains. In addition, we show the potential of applying VICT for unseen tasks at test time. Code: https://github.com/Jiahao000/VICT. + +# 1. Introduction + +Following the success of in-context learning (ICL) [2, 7, 21] in natural language processing (NLP), visual in-context learning (VICL) [5, 52] has shown promising performance in developing generalist models for vision tasks. Inspired + +by prompting in NLP that defines tasks using language sequences as a general interface, existing VICL works [5, 52] use images themselves as a natural interface for general-purpose visual perception. They formulate VICL as an image inpainting task, i.e., given an input-output example describing a specific task (i.e., the prompt) as well as an input image, they are assembled in a grid and the problem is casted as inpainting the missing part of the grid (i.e., the prediction) consistently with the given task prompts. This enables adapting a pre-trained vision model to various downstream tasks with only a handful of prompts and examples. + +By default, current VICL models are frozen during deployment. However, their performance might suffer in real-world deployment as the test distribution usually changes and deviates from the training one. This naturally raises a question: what is the generalizability of VICL models under distribution shifts? In this work, we investigate this aspect focusing on controllable shifts due to image corruptions. We observe that the existing VICL paradigm like Painter [52] exhibits poor generalizability to unseen new domains when the task prompt comes from the training distribution, as shown in Figure 1. Surprisingly, performances are even worse when given input-output task prompts belonging to the same distribution as the test input. + +Driven by this observation, we propose to rely on test-time visual in-context tuning (VICT) to adapt VICL models on the fly to unseen distributions using only the given test sample. The motivation is that each test input offers a hint about the test distribution. Thus, we modify the VICL + +model at test time to make full use of this hint by setting up a one-sample learning problem. Specifically, given input-output task prompts and the test input, we first use the VICL model to inpaint the output image. We then flip the role between the task prompts and the test sample, i.e., we treat the predicted test output as the prompt to the model and reconstruct the original output of task prompts. This allows us to tune the parameters of the whole model in a self-supervised manner that can be applied to arbitrary tasks. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts conditioned on its in-context inference. Such a cycle consistency supervision signal naturally exists in the context of VICL without requiring any additional training data or annotations, thus making it an appealing self-supervisory task for test-time visual in-context training. + +We explore VICT in two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). Without loss of generality, we consider Painter [52] as our VICL model, for its simplicity in design and its wide applicability. As shown in Figure 1, our simple method leads to substantial improvements across 15 common corruptions [23, 31] on six representative vision tasks ranging from high-level visual understanding to low-level image processing, including depth estimation on NYUv2 [40], semantic segmentation on ADE20K [63], panoptic segmentation on COCO [27], image denoising on SIDD [1], image deraining on the merged deraining dataset [57], and low-light image enhancement on LoL [53]. + +Our main contributions are summarized as follows: + +1) We propose a new cycle consistency task for test-time visual in-context tuning. To the best of our knowledge, we are the first to perform test-time training for VICL. +2) We contribute the first study on the generalizability of VICL under distribution shifts. We observe that the existing VICL paradigm exhibits poor generalizability to unseen new domains. Such a phenomenon can hardly be recovered even given input-output task prompts with the same distribution as the test input. +3) We conduct extensive experiments on six representative vision tasks across 15 common corruptions. VICT significantly improves Painter in both zero-shot and one-shot manners. Our zero-shot or one-shot VICT can even outperform Painter trained with more few-shot corrupted examples. We also explore the potential of applying VICT for unseen tasks at test time, further demonstrating its promise. + +# 2. Related Work + +In-context learning. In-context learning has been extensively explored in the NLP literature after the introduction of large language models. Seminal works in the field + +like GPT-3 [7] go as far as claiming that language models are indeed few-shot learners. All recent language models like [13, 46, 47] have the ability to adapt their behavior based on few-shot examples provided in their context window or directly to follow simple user instructions [10]. Similar concepts have been extended to multimodal models with seminal works like Flamingo [2] showing few-shot capabilities by providing examples as interleaved text and images in the context window. Some of these capabilities are now available for extremely large commercial multimodal large language models [35, 42, 43]. + +Apart from using languages as the general interface, a recent line of work introduces purely visual in-context learning (VICL) by training vision generalist models that can perform arbitrary visual tasks following one or few visual examples provided at inference time. Two representative works developed in parallel are: (i) MAE-VQGAN [5], which trains a variant of MAE [22] on a dataset of figures extracted from academic papers, and (ii) Painter [52], which uses a similar idea but trains its model on a set of standard academic benchmark datasets. More recently, this strategy has been expanded in LVM [3], where even more visual datasets are collected to train an autoregressive generative model. We base our work on Painter [52] due to its simplicity in design and its wide applicability. It is the one that provides both the available code and model weights, and has already been extensively evaluated across several standard benchmarks. + +Generalization under distribution shifts. Machine learning models suffer from performance drops when tested on a data distribution different from the one they are trained on [17, 49, 50, 61]. The lower the drop, the more we define a model robust or able to generalize to distribution shifts. Common strategies to increase model robustness include using heavy data augmentations [11, 23, 44, 62] or extremely large training distributions going up to web scale [33, 37, 59]. Nevertheless, most of the experiments are done in a classification setting. To the best of our knowledge, there is no previous work exploring the robustness of VICL models. We are the first to counteract performance drops in VICL under distribution shifts. + +Test-time training. An alternative way of counteracting performance drops due to distribution shifts is test-time training [6, 8, 16, 28, 41, 48, 51]. This accounts to unfreezing the model at test time and fine-tuning it on the target distribution (or a single sample from it) through self-supervision. Early works propose this paradigm using self-supervised pretext tasks (e.g., rotation prediction [18]) to explicitly improve generalization under image corruptions [41] or improve in specific tasks where self-supervised losses can be clearly defined like depth estimation [26, 36, 45], reinforcement learning [20, 32], tracking [15, 38], and NLP [4, 56]. More recently, the field + +![](images/493a43ac28f6eab7fb18764968e13a6cb57cd68a1e190644b4172ed47eb7b39e.jpg) +Figure 2. Overview of our VICT pipeline. Given a pair of task prompts $(x,y)$ and a test input image $x_{t}$ , we first construct a four-cell grid-like image canvas $I = (x,y,x_{t},\emptyset)$ , with an empty cell at the bottom right. We then feed $I$ into the VICL model (e.g., Painter) to predict the test output $\hat{y}_t$ . Afterward, we flip the role between input-output task prompts and input-output test samples, i.e., we provide the predicted $\hat{y}_t$ as the prompt to the model, recreating a new four-cell grid-like image canvas $I' = (x,\emptyset,x_{t},\hat{y}_{t})$ , with an empty cell at the top right. The new $I'$ is fed into the same model to predict the task prompt output $\hat{y}$ . We finally optimize the model by minimizing the distance between $\hat{y}$ and $y$ via a standard regression loss. + +has realized that MAE [22] provides a very strong self-supervised pretext task and has therefore been employed in the context of test-time training [16]. Finally, some recent works [14, 24, 30, 39, 58] also employ the test-time training paradigm for vision-language models to improve their generalization. To the best of our knowledge, ours is the first work to explore test-time training applied to VICL models, where previous self-supervised pretext tasks [9, 12, 18, 22, 25, 34, 54, 55, 60] cannot be employed due to their constraints on the single-image context. + +# 3. Methodology + +Our visual in-context tuning (VICT) is a simple yet effective test-time training approach to adapt visual in-context learning (VICL) models on the fly. In this section, we first introduce some preliminaries on VICL in Section 3.1. We then detail our VICT pipeline in Section 3.2. + +# 3.1. Visual In-Context Learning + +In-context learning is a new paradigm originating from large language models such as GPT-3 [7] in NLP. Unlike traditional learning paradigms, in-context learning formulates different NLP tasks as text completion tasks and makes predictions conditioned on one or many support examples provided to the model in the context window. Extending this paradigm to tasks requiring a dense visual output in the form of images is nontrivial and has not been addressed in the literature for a long time. Recently, [5, 52] formulate VICL as an image inpainting task by combining images and labels into a grid-like new image and using masked image modeling for pre-training, granting the models with the in-context learning ability. + +Formally, let $S = \{(x_i, y_i)\}_{i=1}^N$ denote the set of support input-output examples (a.k.a., task prompts) where $x$ is an image and $y$ is its visual label (e.g., a segmentation mask).1 Given $S$ and a new test image $x_t$ as input, VICL can be formulated as follows: + +$$ +y _ {t} = f _ {\theta} (S, x _ {t}), \tag {1} +$$ + +where $f_{\theta}(*)$ is the VICL model parametrized with $\theta$ , $y_{t}$ is the corresponding output label of $x_{t}$ . Usually, we assume $S$ and $x_{t}$ are drawn from the same distribution. However, this is rarely the case for real-world deployment. Therefore, in this work, we mainly consider the scenario where a distribution shift like an image corruption occurs for $x_{t}$ . + +In the experiment section, we will show how the presence of such distribution shifts significantly degrades the performance of the VICL models. Moreover, we will additionally show how simple solutions like providing in-context examples from the target (corrupted) distribution would not fix the problem due to the limited generalization ability of the current VICL models. + +# 3.2. Test-Time Visual In-Context Tuning + +In this work, we propose to leverage test-time training [41] to counteract the distribution shifts between training and testing data, with the goal of making VICL models more robust. We argue that VICL models are particularly amenable to be optimized at test time on arbitrary tasks thanks to the availability of the $N$ shots provided as context. The core intuition of our method is to use the provided in-context examples to define a self-supervised loss that can be used at + +Algorithm 1 Pseudocode of VICT in a PyTorch-like style. +```python +f: VICL model (e.g., Painter) +# theta_0: pre-trained model weights +# x: task prompt input +# y: task prompt output +# mask_token: mask token for the empty cell +# steps: number of test-time optimization steps +for x_t in loader: # load a test sample x_t +f.params = theta_0 # initialize +# test-time optimization +for t in range(steps): + # predict the test sample output + y_t_pred = f((x, y, x_t, mask_token)) + # predict the task prompt output + y_pred = f((x, mask_token, x_t, y_t_pred)) + # regression loss + loss = SmoothL1Loss(y_pred, y) + # update model + loss/backward() +update(f.params) +# inference with the updated model +y_t_pred = f((x, y, x_t, mask_token)) +``` + +test time to train any VICL models on the fly on any task. The overall pipeline of our VICT is illustrated in Figure 2. + +Specifically, given a pair of task prompts $(x,y)$ from $S$ and a test input $x_{t}$ , we construct a grid-like image canvas with four cells, denoted by an ordered quadruple $I = (x,y,x_{t},\emptyset)$ , where $\emptyset$ is an empty cell. We then feed $I$ into the VICL model $f_{\theta}$ to predict the test output $\hat{y}_t$ : + +$$ +\hat {y} _ {t} = f _ {\theta} (I) = f _ {\theta} \left(\left(x, y, x _ {t}, \varnothing\right)\right). \tag {2} +$$ + +Afterward, with the obtained test output, we flip the role between input-output task prompts and input-output test samples, creating a new four-cell grid-like image canvas with the order of $I' = (x, \varnothing, x_t, \hat{y}_t)$ . The new $I'$ is fed into $f_{\theta}$ again to predict the task prompt output $\hat{y}$ : + +$$ +\hat {y} = f _ {\theta} \left(I ^ {\prime}\right) = f _ {\theta} \left(\left(x, \varnothing , x _ {t}, \hat {y} _ {t}\right)\right). \tag {3} +$$ + +We then use a simple regression loss in pixel space to optimize the model by minimizing the distance between $\hat{y}$ and $y$ : + +$$ +\theta_ {x _ {t}} = \arg \min _ {\theta} \mathcal {L} (\hat {y}, y) \tag {4} +$$ + +In practice, we follow [52] to use smooth- $\ell_1$ [19] loss. Using this loss we can optimize the parameters $\theta$ of the model for a small number of steps for each $x_t$ and, eventually, make a prediction on $x_t$ as $f_{\theta_{x_t}}((x,y,x_t,\emptyset))$ . + +In our formulation, the gradient-based optimization for Equation 4 always starts from the initial pre-trained model weights $\theta_0$ for each test input. When a new test input arrives, we discard $\theta_{x_t}$ and reset the weights to $\theta_0$ . We follow + +this strategy to not have any assumption on the test inputs and treat them independently at test time, i.e., we do not assume that they come from the same distribution. Our generic formulation is completely agnostic on the VICL models considered, on the type of tasks being solved, and on the type of in-context samples provided to it. + +The pseudo-code of VICT is in Algorithm 1. + +# 4. Experiments + +# 4.1. Implementation Details + +Tasks and datasets. Following [52], we evaluate the model performance on six representative vision tasks ranging from high-level visual understanding to low-level image processing, which includes depth estimation on NYUv2 [40], semantic segmentation on ADE20K [63], panoptic segmentation on COCO [27], image denoising on SIDD [1], image draining on the merged draining datasets [57], and low-light image enhancement on LoL [53]. To evaluate the robustness to distribution shifts of the model on these datasets, we follow the setup in [23, 31] to corrupt the aforementioned datasets and simulate hard distribution shifts. The corrupted datasets contain 15 types of corruptions, covering noise (gaussian noise, impulse noise, shot noise), blur (defocus blur, glass blur, motion blur, zoom blur), weather (fog, frost, snow), and digital (brightness, contrast, elastic transform, JPEG compression, pixelate) categories. Each type of corruption has five levels of severity. We denote the corrupted datasets as NYUv2-C, ADE20K-C, COCO-C, SIDD-C, Rain-C, and LoL-C, respectively. Due to the space constraints, the results in the main text are limited to the most severe level (i.e., level 5) that corresponds to the strongest distribution shift. We provide the results on the other four levels in the supplementary material. + +Baselines. Without loss of generality, we consider Painter [52] as our VICL model for its simplicity in design and its wide applicability. We study VICT in two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). For the one-shot setting, apart from comparing with the frozen Painter using one-shot corrupted examples as task prompts, we also consider a baseline that further trains Painter with the one-shot corrupted samples using the same pre-training objective as in [52]. We also consider training Painter with more few-shot examples using the same settings to examine to what extent VICT can outperform a larger number of few-shot fine-tuning of Painter. + +Training details. In all experiments, we use the pre-trained Painter based on ViT-Large provided by the authors of [52]. We perform VICT using an AdamW [29] optimizer, with betas as (0.9, 0.999), a weight decay of 0, a batch size of 1, + +Table 1. System-level comparison on six representative vision tasks across 15 corruptions. Results are on corruption level 5. We consider two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). "avg" denotes the averaged results over 15 corruptions. + +
methodbrighcontdefocelastfogfrostgaussglassimpuljpegmotnpixelshotsnowzoomavg
(a) depth estimation NYUv2-C (A.Rel ↓)
zero-shot setting:
Painter0.1290.2150.7120.1090.1290.5360.2000.6120.1890.3860.1670.1870.1421.9510.2090.392
VICT0.1080.2160.6310.1080.1330.5760.1910.5410.1810.3100.1390.1340.1401.8560.2100.365
one-shot setting:
Painter0.1260.2850.7430.1090.1320.8530.9010.6220.9010.3920.1740.1940.4401.9640.2120.537
VICT0.0970.1930.1800.1070.1280.2270.2410.2100.2780.1590.1210.1140.1500.6620.2140.205
(b) semantic segmentation ADE20K-C (mIoU ↑)
zero-shot setting:
Painter40.411.727.931.235.023.324.825.626.138.631.840.425.99.413.127.0
VICT40.912.728.231.736.423.725.526.327.438.831.940.928.017.313.628.2
one-shot setting:
Painter40.911.727.531.335.022.119.725.820.338.531.340.323.26.513.325.8
VICT41.721.028.832.237.024.724.427.025.239.732.841.526.621.714.929.3
(c) panoptic segmentation COCO-C (PQ ↑)
zero-shot setting:
Painter38.115.525.530.033.326.024.923.825.531.927.634.926.319.812.526.4
VICT38.715.725.830.333.926.725.324.325.932.427.735.426.721.912.726.9
one-shot setting:
Painter38.114.224.830.033.125.221.523.722.031.527.534.623.817.212.125.3
VICT38.616.925.430.433.826.624.024.624.532.327.435.025.224.912.426.8
(d) denoising SIDD-C (PSNR ↑)
zero-shot setting:
Painter8.7225.4932.4833.2016.868.7917.5033.1817.5023.1829.4734.8425.279.8629.8023.08
VICT8.8625.6532.9132.9617.008.8417.5833.0017.5024.8629.6834.7625.219.7529.6723.22
one-shot setting:
Painter9.1225.1631.9933.2316.689.2417.9933.1418.3322.7329.3034.8724.6910.5029.7623.12
VICT13.9425.3232.4134.4516.4511.1520.1332.9619.9326.8030.1135.0124.9314.9029.9524.56
(e) deraining Rain-C (PSNR ↑)
zero-shot setting:
Painter10.7414.1519.5420.1013.9112.4919.8519.8020.1624.8618.5222.8120.1111.8917.9317.79
VICT10.8914.2419.5820.3113.9412.6220.4219.8320.6724.9718.5623.0520.4712.0517.9817.97
one-shot setting:
Painter11.0413.9519.6120.1513.8212.8719.5119.8719.5824.8518.6222.8819.6812.1617.9517.77
VICT17.3814.6920.3621.4414.6717.6922.3320.4822.2425.3119.0523.9221.3719.5318.2819.92
(f) low-light enhancement LoL-C (PSNR ↑)
zero-shot setting:
Painter16.2613.4318.8019.8311.2411.6212.5018.8212.9115.8718.0121.1217.7513.7716.6015.90
VICT16.3913.5719.2419.8411.3711.6412.5119.2313.0016.0018.1121.2019.1613.9217.0616.15
one-shot setting:
Painter16.7413.5018.8519.8311.0211.7613.1418.9513.7415.8418.1521.1216.5413.1916.5915.93
VICT18.1915.0318.9919.8811.7612.0114.9519.2214.9916.1218.2321.0919.8416.9616.9116.94
+ +![](images/5c21e893ebea43beca12e72336776524f2549dc0a77e2537b2fcfeacaa60eeda.jpg) + +![](images/24a6e1ad11f78096a691db1425e9a292e7ddefb7e1104c44a4465a04bbe35ef9.jpg) + +![](images/74a4ebeb11dadf3b69613d7892e0723dd12e57012f4e222f055cf1c1becb0ea5.jpg) + +![](images/43db942e5f7350ca357479316f7acfcbb8f1aa2eb081fd57037946c60a4ce14f.jpg) +Figure 3. Comparison with few-shot Painter on six vision tasks with corruptions. We randomly corrupt a certain number of images in the training set, using 1, 2, 4, 8, 16, 32, and 64 shots for training and deploying the model in the full corrupted test sets. We report the final results averaged across 15 corruptions. Our zero-shot or one-shot VICT can outperform Painter trained with more few-shot examples. + +![](images/28dfed92e636cab673193c1254943774664a2e793f89e4bc5c975bc488290c56.jpg) + +![](images/1ae8f4e72ae4967bcd35bf4d953259d2f92666579db157e14ed560ebcab99ca4.jpg) + +and a fixed learning rate of 1e-6. Unless otherwise specified, we train each test sample for 60 steps. The choice of 60 steps is purely computational and more steps are likely to further improve performance, judging from the positive trend observed in Figure 4. Note that we do not apply any data augmentations for VICT as many data augmentations are in fact distribution shifts in our evaluation benchmarks. Training with them is analogous to training on the test distributions. Therefore, to solely study the generalization ability to new test distributions, we purposely choose not to use any data augmentations, even though they could improve our results at face value. Every iteration of VICT is performed on the same concatenated grid-like image, with each subimage resized as $448 \times 448$ , which is the same as what we later use for in-context inference. During VICT, we only optimize the encoder weights. We have experimented with training both the encoder and the decoder and found that the difference is negligible, which is also consistent with the observations in [16, 41]. A detailed comparison is provided in the experiment section. + +# 4.2. Main Results + +System-level comparison. We compare VICT with Painter under 15 common corruptions as introduced in Section 4.1. We perform experiments on six representative vision tasks covering high-level visual understanding and low-level image processing, which includes depth estimation on NYUv2-C, semantic segmentation on ADE20K-C, panop- + +tic segmentation on COCO-C, image denoising on SIDD-C, image draining on Rain-C, and low-light image enhancement on LoL-C. + +Table 1 reports the results. First of all, we observe that Painter exhibits poor generalization ability under common corruptions. For example, on ADE20K-C semantic segmentation, Painter only achieves an average performance of $27.0\mathrm{mIoU}$ , which is $22.9\mathrm{mIoU}$ lower than that on the clean validation set (49.9 mIoU as reported in Figure 1). Nevertheless, our VICT outperforms Painter by clear margins in both zero-shot and one-shot settings across different tasks. When it comes to the zero-shot setting, VICT achieves an average of -0.027 A.Rel on NYUv2-C depth estimation, +1.2 mIoU on ADE20K-C semantic segmentation, +0.5 PQ on COCO-C panoptic segmentation, +0.14 PSNR on SIDD-C image denosing, +0.18 PSNR on Rain-C image draining, and +0.25 PSNR on LoL-C low-light enhancement, respectively. + +The performance gains of VICT over Painter are further increased when we shift from the zero-shot setting to the one-shot counterpart. Specifically, VICT achieves an average of -0.332 A.Rel on NYUv2-C depth estimation, +3.5 mIoU on ADE20K-C semantic segmentation, +1.5 PQ on COCO-C panoptic segmentation, +1.44 PSNR on SIDC image denoising, +2.15 PSNR on Rain-C image draining, and +1.01 PSNR on LoL-C low-light enhancement, respectively. Besides, for the Painter baseline, using the + +Table 2. Test-time optimization on different modules. We use semantic segmentation on ADE20K-C for the ablation. Training only the encoder performs similarly to training both the encoder and decoder, regardless of the zero-shot and one-shot settings. + +
modulebrighcontdefocelastfogfrostgaussglassimpuljpegmotnpixelshotsnowzoomavg
zero-shot setting:
Encoder40.912.728.231.736.423.725.526.327.438.831.940.928.017.313.628.2
Both40.912.528.231.736.423.825.726.527.438.831.941.027.916.913.528.2
one-shot setting:
Encoder41.721.028.832.237.024.724.427.025.239.732.841.526.621.714.929.3
Both41.721.028.732.136.924.424.427.025.339.633.041.626.821.515.129.3
+ +one-shot setting does not always perform better than the zero-shot setting. The average performances are even degraded in most cases, e.g., depth estimation on NYUv2-C (0.392 A.Rel vs. 0.537 A.Rel), semantic segmentation on ADE20K-C (27.0 mIoU vs. 25.8 mIoU), panoptic segmentation on COCO-C (26.4 PQ vs. 25.3 PQ), image draining on Rain-C (17.79 PSNR vs. 17.77 PSNR). In contrast, for VICT, using the one-shot setting usually performs better than the zero-shot setting. It is worth mentioning that our zero-shot VICT can even outperform the one-shot Painter, further demonstrating the effectiveness of VICT. + +Comparison with few-shot Painter. In previous experiments, we mainly compare with a frozen Painter using either clean or corrupted examples as the task prompts. Here, we further consider a scenario where more few-shot examples from the test distribution are available. Specifically, we randomly corrupt a certain number of images in the training set, using 1, 2, 4, 8, 16, 32, and 64 shots to train Painter respectively and deploying the model in the full corrupted test sets. + +The results are shown in Figure 3. Both our zero-shot VICT and one-shot VICT can consistently outperform the one-shot Painter. When more few-shot examples are available for training, our VICT can still outperform Painter below a certain threshold. We observe such a threshold varies for different tasks, with high-level visual understanding tasks having a higher threshold. It is worth noting that in high-level visual understanding tasks like ADE20K-C semantic segmentation and COCO-C panoptic segmentation, both our zero-shot and one-shot variants of VICT nearly match or even outperform the 64-shot Painter. In other words, with only one or simply no labeled test sample(s), VICT can achieve similar performances as models that use significantly more labeled test samples to train. This phenomenon is rather appealing as in practice it is rare to know test distribution in advance, let alone obtaining a large number of labeled test samples. + +# 4.3. Further Analysis + +Effect of test-time optimization modules. We study the effect of optimizing different modules at test time. We con + +![](images/83e0095381d50667584687df0c79c3e0135b9cc55c44edaa81432e54ff6c821d.jpg) +Figure 4. Analysis on the trade-off between efficiency and accuracy. We use semantic segmentation on ADE20K-C for the ablation. VICT benefits from more training steps, while at the cost of linearly increased training time. + +sider two parameter groups: (i) the encoder, and (ii) both the encoder and decoder. We use semantic segmentation on ADE20K-C as an example task for analysis. The results are shown in Table 2. Training only the encoder performs similarly to training the entire model, regardless of the zero-shot and one-shot settings. This makes sense since the encoder acts as a feature extractor, which plays a key role in determining the representation quality. In contrast, the decoder is only responsible for mapping the latent representation back to its original resolution, which is more specialized for reconstruction but less relevant for semantics. Thus, we choose only to optimize the encoder at test time. + +Effect of test-time optimization steps. We study the effect of test-time optimization steps in Figure 4. VICT benefits from more training steps, which keeps improving performance even after 60 steps. However, it should be noted that the runtime of our VICT is linearly increased with the number of training steps. For reference, it takes around 0.4 seconds per step per test sample on a single A100 GPU. Thus, in practice, one may decide the number of training steps according to the pre-defined cost budget for deployment. + +![](images/0860baaaed2f1b52f5e0044910dd92971ea4a29f725376445dc3aedbc2e1c684.jpg) +Figure 5. Visualizations of test examples and predictions for six main tasks with corruptions. We visualize both zero-shot and one-shot settings for Painter and VICT. Zoom in for best view. + +# 4.4. Qualitative Results + +Results on main tasks. We visualize some test examples and predictions from the validation set for different tasks with corruptions, including depth estimation, semantic segmentation, instance segmentation, image denoising, image deraining, and low-light image enhancement. As shown in Figure 5, VICT can make more accurate predictions than Painter on all tasks. + +Results on unseen tasks. In this work, we mainly consider the distribution shifts with corruptions and conduct experiments on the tasks that are seen in training. Here, we further verify whether VICT can generalize to unseen tasks. We explore this capability via visualizations. More quantitative results are provided in the supplementary material. Figure 6 provides examples of two unseen tasks including foreground object segmentation and colorization. Both Painter and our VICT can generalize to the foreground segmentation task. This makes sense as similar segmentation tasks (e.g., semantic segmentation and instance segmentation) have been seen during the training stage of Painter. Nevertheless, our VICT further reduces noises and produces finer masks. However, Painter cannot generalize to the colorization task that is totally unseen during training, produc + +![](images/9e0376eb6755de968ec4fe00585dd9737bb5ce81b14b98eeb699e9dac885fbe7.jpg) +Figure 6. Visualizations of test examples and predictions for unseen tasks. The visualized tasks include foreground object segmentation and colorization. Painter cannot generalize to totally unseen tasks like colorization, whereas VICT can make decent color predictions. Zoom in for best view. + +ing only grayscale images or even wrong predictions from other tasks (e.g., the last row). In contrast, our VICT can produce decent colorful results. This further demonstrates the potential of applying VICT for unseen tasks at test time. + +# 5. Conclusion + +Generalization under distribution shifts is the central theme of deep learning. In this work, we have studied the robustness of VICL models and found that the existing VICL paradigm exhibits poor generalization capability to unseen new domains. Based on this observation, we proposed a simple yet effective test-time visual in-context tuning method to adapt VICL models to a single test sample on the fly. We hope our explorations can pave the way for improving the generalizability of VICL. + +Limitations and future work. Our study has several limitations: 1) Since we perform training for each test sample, our method is slower than the baseline applying a fixed model at test time, which is a common limitation for test-time training. Inference speed might be improved through better architectural designs, training techniques, optimizers, and hyper-parameters. However, it has not been the focus of this paper. 2) Our proposed cycle consistency supervision is a general self-supervised task. However, we cannot guarantee that it will produce useful gradients for every single test distribution. We only focus on several representation vision tasks and most popular corruption benchmarks for distribution shifts. More tasks and benchmarks can be studied for future research. One possible extension of this work is to apply similar test-time supervision in multi-modal in-context learning, where cycle consistency supervision also exists. We leave this exploration for future work. + +# References + +[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In CVPR, 2018. 2, 4 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 2 +[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, 2024. 2 +[4] Pratyay Banerjee, Tejas Gokhale, and Chitta Baral. Self-supervised test-time learning for reading comprehension. In NAACL, 2021. 2 +[5] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. In NeurIPS, 2022. 1, 2, 3 +[6] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In AISTATS, 2022. 2 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1, 2, 3 +[8] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In CVPR, 2022. 2 +[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3 +[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. JMLR, 2024. 2 +[11] Samuel Dodge and Lina Karam. Quality resilient deep neural networks. arXiv preprint arXiv:1703.08119, 2017. 2 +[12] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015. 3 +[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2 +[14] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In ICCV, 2023. 3 +[15] Yang Fu, Sifei Liu, Umar Iqbal, Shalini De Mello, Humphrey Shi, and Jan Kautz. Learning to track instances without video annotations. In CVPR, 2021. 2 +[16] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 2, 3, 6 + +[17] Robert Geirhos, Carlos RM Temme, Jonas Rauber, Heiko H Schütt, Matthias Bethge, and Felix A Wichmann. Generalisation in humans and deep neural networks. In NeurIPS, 2018. 2 +[18] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 2, 3 +[19] Ross Girshick. Fast r-cnn. In ICCV, 2015. 4 +[20] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In ICLR, 2021. 2 +[21] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma, and Furu Wei. Language models are general-purpose interfaces. arXiv preprint arXiv:2206.06336, 2022. 1 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 3 +[23] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019. 1, 2, 4 +[24] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In CVPR, 2024. 3 +[25] Wei Li, Jiahao Xie, and Chen Change Loy. Correlational image modeling for self-supervised visual pre-training. In CVPR, 2023. 3 +[26] Zhi Li, Shaoshuai Shi, Bernt Schiele, and Dengxin Dai. Test-time domain adaptation for monocular depth estimation. In ICRA, 2023. 2 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4 +[28] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In NeurIPS, 2021. 2 +[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 4 +[30] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision-language models. In NeurIPS, 2023. 3 +[31] Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019. 1, 2, 4 +[32] Mirco Mutti and Aviv Tamar. Test-time regret minimization in meta reinforcement learning. In ICML, 2024. 2 +[33] Muhammad Ferjad Naeem, Yongqin Xian, Xiaohua Zhai, Lukas Hoyer, Luc Van Gool, and Federico Tombari. Silc: Improving vision language pretraining with self-distillation. In ECCV, 2024. 2 + +[34] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 3 +[35] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[36] Matteo Poggi, Alessio Tonioni, Fabio Tosi, Stefano Mattoccia, and Luigi Di Stefano. Continual adaptation for deep stereo. TPAMI, 2021. 2 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2 +[38] Mattia Segu, Bernt Schiele, and Fisher Yu. Darth: holistic test-time adaptation for multiple object tracking. In ICCV, 2023. 2 +[39] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 3 +[40] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 2, 4 +[41] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICLR, 2020. 2, 3, 6 +[42] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 +[43] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2 +[44] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IROS, 2017. 2 +[45] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In CVPR, 2019. 2 +[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2 +[48] Yun-Yun Tsai, Chengzhi Mao, and Junfeng Yang. Convolutional visual prompt for robust visual perception. In NeurIPS, 2023. 2 + +[49] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. 2 +[50] Igor Vasiljevic, Ayan Chakrabarti, and Gregory Shakhnarovich. Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760, 2016. 2 +[51] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 2 +[52] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, 2023. 1, 2, 3, 4 +[53] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In BMVC, 2018. 2, 4 +[54] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Delving into inter-image invariance for unsupervised visual representations. IJCV, 2022. 3 +[55] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. In ICLR, 2023. 3 +[56] Hai Ye, Yuyang Ding, Juntao Li, and Hwee Tou Ng. Robust question answering against distribution shifts with test-time adaptation: An empirical study. In Findings of EMNLP, 2022. 2 +[57] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In ECCV, 2020. 2, 4 +[58] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, 2024. 3 +[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 2 +[60] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020. 3 +[61] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 2 +[62] Stephan Zheng, Yang Song, Thomas Leung, and Ian Goodfellow. Improving the robustness of deep neural networks via stability training. In CVPR, 2016. 2 +[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2, 4 \ No newline at end of file diff --git a/2025/Test-Time Visual In-Context Tuning/images.zip b/2025/Test-Time Visual In-Context Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..076f3461c68e1eb79a5f4451e1ed6359d2878c06 --- /dev/null +++ b/2025/Test-Time Visual In-Context Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05c755b90ab2ed34555a4f855914fbbc00fadc137c00e60ff33a765ddaea428a +size 688115 diff --git a/2025/Test-Time Visual In-Context Tuning/layout.json b/2025/Test-Time Visual In-Context Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..796df9553571aa305445c325a536a725f964dd9a --- /dev/null +++ b/2025/Test-Time Visual In-Context Tuning/layout.json @@ -0,0 +1,7142 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 193, + 103, + 419, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 103, + 419, + 121 + ], + "spans": [ + { + "bbox": [ + 193, + 103, + 419, + 121 + ], + "type": "text", + "content": "Test-Time Visual In-Context Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "spans": [ + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "content": "Jiahao Xie" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "content": ", Alessio Tonioni" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "content": ", Nathalie Rauschmayr" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "content": ", Federico Tombari" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "content": ", Bernt Schiele" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "content": "Max Planck Institute for Informatics, SIC " + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "content": "VIA Research Center " + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 142, + 526, + 172 + ], + "type": "text", + "content": "Google" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 173, + 504, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 173, + 504, + 186 + ], + "spans": [ + { + "bbox": [ + 100, + 173, + 504, + 186 + ], + "type": "text", + "content": "{jxie, schiele}@mpi-inf.mpg.de {alessiot, rauschmayr, tombari}@google.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 210, + 555, + 292 + ], + "blocks": [ + { + "bbox": [ + 56, + 210, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 56, + 210, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 555, + 292 + ], + "type": "image", + "image_path": "dc83c086a4cc0252879660d3cfa511bf47deca6f2878ba337d9a0e0a35b4e1c7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 300, + 555, + 356 + ], + "lines": [ + { + "bbox": [ + 54, + 300, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 555, + 356 + ], + "type": "text", + "content": "Figure 1. Test-time visual in-context tuning (VICT) on six representative vision tasks under distribution shifts. We benchmark the robustness of VICL with 15 common corruptions adopted in [23, 31], and report the averaged performance across all corruptions. Existing VICL models like Painter exhibit poor generalization capability to unseen new domains when the task prompts come from the training distribution (i.e., zero-shot). Performances are even worse when given task prompts from the test distribution (i.e., one-shot). By performing VICT at test time, we can significantly improve Painter in both zero-shot and one-shot manners." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 365, + 200, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 365, + 200, + 378 + ], + "spans": [ + { + "bbox": [ + 151, + 365, + 200, + 378 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 391, + 297, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 297, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 297, + 618 + ], + "type": "text", + "content": "Visual in-context learning (VICL), as a new paradigm in computer vision, allows the model to rapidly adapt to various tasks with only a handful of prompts and examples. While effective, the existing VICL paradigm exhibits poor generalizability under distribution shifts. In this work, we propose test-time Visual In-Context Tuning (VICT), a method that can adapt VICL models on the fly with a single test sample. Specifically, we flip the role between the task prompts and the test sample and use a cycle consistency loss to reconstruct the original task prompt output. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts. Extensive experiments on six representative vision tasks ranging from high-level visual understanding to low-level image processing, with 15 common corruptions, demonstrate that our VICT can improve the generalizability of VICL to unseen new domains. In addition, we show the potential of applying VICT for unseen tasks at test time. Code: https://github.com/Jiahao000/VICT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 644, + 135, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 135, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 135, + 657 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Following the success of in-context learning (ICL) [2, 7, 21] in natural language processing (NLP), visual in-context learning (VICL) [5, 52] has shown promising performance in developing generalist models for vision tasks. Inspired" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "text", + "content": "by prompting in NLP that defines tasks using language sequences as a general interface, existing VICL works [5, 52] use images themselves as a natural interface for general-purpose visual perception. They formulate VICL as an image inpainting task, i.e., given an input-output example describing a specific task (i.e., the prompt) as well as an input image, they are assembled in a grid and the problem is casted as inpainting the missing part of the grid (i.e., the prediction) consistently with the given task prompts. This enables adapting a pre-trained vision model to various downstream tasks with only a handful of prompts and examples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 654 + ], + "type": "text", + "content": "By default, current VICL models are frozen during deployment. However, their performance might suffer in real-world deployment as the test distribution usually changes and deviates from the training one. This naturally raises a question: what is the generalizability of VICL models under distribution shifts? In this work, we investigate this aspect focusing on controllable shifts due to image corruptions. We observe that the existing VICL paradigm like Painter [52] exhibits poor generalizability to unseen new domains when the task prompt comes from the training distribution, as shown in Figure 1. Surprisingly, performances are even worse when given input-output task prompts belonging to the same distribution as the test input." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Driven by this observation, we propose to rely on test-time visual in-context tuning (VICT) to adapt VICL models on the fly to unseen distributions using only the given test sample. The motivation is that each test input offers a hint about the test distribution. Thus, we modify the VICL" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19996" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "content": "model at test time to make full use of this hint by setting up a one-sample learning problem. Specifically, given input-output task prompts and the test input, we first use the VICL model to inpaint the output image. We then flip the role between the task prompts and the test sample, i.e., we treat the predicted test output as the prompt to the model and reconstruct the original output of task prompts. This allows us to tune the parameters of the whole model in a self-supervised manner that can be applied to arbitrary tasks. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts conditioned on its in-context inference. Such a cycle consistency supervision signal naturally exists in the context of VICL without requiring any additional training data or annotations, thus making it an appealing self-supervisory task for test-time visual in-context training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 443 + ], + "type": "text", + "content": "We explore VICT in two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). Without loss of generality, we consider Painter [52] as our VICL model, for its simplicity in design and its wide applicability. As shown in Figure 1, our simple method leads to substantial improvements across 15 common corruptions [23, 31] on six representative vision tasks ranging from high-level visual understanding to low-level image processing, including depth estimation on NYUv2 [40], semantic segmentation on ADE20K [63], panoptic segmentation on COCO [27], image denoising on SIDD [1], image deraining on the merged deraining dataset [57], and low-light image enhancement on LoL [53]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 444, + 276, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 444, + 276, + 454 + ], + "spans": [ + { + "bbox": [ + 67, + 444, + 276, + 454 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 295, + 647 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 455, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 491 + ], + "type": "text", + "content": "1) We propose a new cycle consistency task for test-time visual in-context tuning. To the best of our knowledge, we are the first to perform test-time training for VICL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 563 + ], + "type": "text", + "content": "2) We contribute the first study on the generalizability of VICL under distribution shifts. We observe that the existing VICL paradigm exhibits poor generalizability to unseen new domains. Such a phenomenon can hardly be recovered even given input-output task prompts with the same distribution as the test input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 647 + ], + "type": "text", + "content": "3) We conduct extensive experiments on six representative vision tasks across 15 common corruptions. VICT significantly improves Painter in both zero-shot and one-shot manners. Our zero-shot or one-shot VICT can even outperform Painter trained with more few-shot corrupted examples. We also explore the potential of applying VICT for unseen tasks at test time, further demonstrating its promise." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 658, + 142, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 142, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 142, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "In-context learning. In-context learning has been extensively explored in the NLP literature after the introduction of large language models. Seminal works in the field" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "like GPT-3 [7] go as far as claiming that language models are indeed few-shot learners. All recent language models like [13, 46, 47] have the ability to adapt their behavior based on few-shot examples provided in their context window or directly to follow simple user instructions [10]. Similar concepts have been extended to multimodal models with seminal works like Flamingo [2] showing few-shot capabilities by providing examples as interleaved text and images in the context window. Some of these capabilities are now available for extremely large commercial multimodal large language models [35, 42, 43]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 206, + 553, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 409 + ], + "type": "text", + "content": "Apart from using languages as the general interface, a recent line of work introduces purely visual in-context learning (VICL) by training vision generalist models that can perform arbitrary visual tasks following one or few visual examples provided at inference time. Two representative works developed in parallel are: (i) MAE-VQGAN [5], which trains a variant of MAE [22] on a dataset of figures extracted from academic papers, and (ii) Painter [52], which uses a similar idea but trains its model on a set of standard academic benchmark datasets. More recently, this strategy has been expanded in LVM [3], where even more visual datasets are collected to train an autoregressive generative model. We base our work on Painter [52] due to its simplicity in design and its wide applicability. It is the one that provides both the available code and model weights, and has already been extensively evaluated across several standard benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 412, + 553, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 553, + 567 + ], + "type": "text", + "content": "Generalization under distribution shifts. Machine learning models suffer from performance drops when tested on a data distribution different from the one they are trained on [17, 49, 50, 61]. The lower the drop, the more we define a model robust or able to generalize to distribution shifts. Common strategies to increase model robustness include using heavy data augmentations [11, 23, 44, 62] or extremely large training distributions going up to web scale [33, 37, 59]. Nevertheless, most of the experiments are done in a classification setting. To the best of our knowledge, there is no previous work exploring the robustness of VICL models. We are the first to counteract performance drops in VICL under distribution shifts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Test-time training. An alternative way of counteracting performance drops due to distribution shifts is test-time training [6, 8, 16, 28, 41, 48, 51]. This accounts to unfreezing the model at test time and fine-tuning it on the target distribution (or a single sample from it) through self-supervision. Early works propose this paradigm using self-supervised pretext tasks (e.g., rotation prediction [18]) to explicitly improve generalization under image corruptions [41] or improve in specific tasks where self-supervised losses can be clearly defined like depth estimation [26, 36, 45], reinforcement learning [20, 32], tracking [15, 38], and NLP [4, 56]. More recently, the field" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19997" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 238 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 238 + ], + "type": "image", + "image_path": "493a43ac28f6eab7fb18764968e13a6cb57cd68a1e190644b4172ed47eb7b39e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": "Figure 2. Overview of our VICT pipeline. Given a pair of task prompts " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": " and a test input image " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": ", we first construct a four-cell grid-like image canvas " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "I = (x,y,x_{t},\\emptyset)" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": ", with an empty cell at the bottom right. We then feed " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": " into the VICL model (e.g., Painter) to predict the test output " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": ". Afterward, we flip the role between input-output task prompts and input-output test samples, i.e., we provide the predicted " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": " as the prompt to the model, recreating a new four-cell grid-like image canvas " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "I' = (x,\\emptyset,x_{t},\\hat{y}_{t})" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": ", with an empty cell at the top right. The new " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": " is fed into the same model to predict the task prompt output " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": ". We finally optimize the model by minimizing the distance between " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 243, + 555, + 312 + ], + "type": "text", + "content": " via a standard regression loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 326, + 297, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 326, + 297, + 447 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 297, + 447 + ], + "type": "text", + "content": "has realized that MAE [22] provides a very strong self-supervised pretext task and has therefore been employed in the context of test-time training [16]. Finally, some recent works [14, 24, 30, 39, 58] also employ the test-time training paradigm for vision-language models to improve their generalization. To the best of our knowledge, ours is the first work to explore test-time training applied to VICL models, where previous self-supervised pretext tasks [9, 12, 18, 22, 25, 34, 54, 55, 60] cannot be employed due to their constraints on the single-image context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 138, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 138, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 138, + 468 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 535 + ], + "type": "text", + "content": "Our visual in-context tuning (VICT) is a simple yet effective test-time training approach to adapt visual in-context learning (VICL) models on the fly. In this section, we first introduce some preliminaries on VICL in Section 3.1. We then detail our VICT pipeline in Section 3.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 208, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 208, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 208, + 553 + ], + "type": "text", + "content": "3.1. Visual In-Context Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "In-context learning is a new paradigm originating from large language models such as GPT-3 [7] in NLP. Unlike traditional learning paradigms, in-context learning formulates different NLP tasks as text completion tasks and makes predictions conditioned on one or many support examples provided to the model in the context window. Extending this paradigm to tasks requiring a dense visual output in the form of images is nontrivial and has not been addressed in the literature for a long time. Recently, [5, 52] formulate VICL as an image inpainting task by combining images and labels into a grid-like new image and using masked image modeling for pre-training, granting the models with the in-context learning ability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "text", + "content": "Formally, let " + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "inline_equation", + "content": "S = \\{(x_i, y_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "text", + "content": " denote the set of support input-output examples (a.k.a., task prompts) where " + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "text", + "content": " is an image and " + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "text", + "content": " is its visual label (e.g., a segmentation mask).1 Given " + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "text", + "content": " and a new test image " + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 326, + 555, + 385 + ], + "type": "text", + "content": " as input, VICL can be formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 401, + 396, + 555, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 396, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 401, + 396, + 555, + 409 + ], + "type": "interline_equation", + "content": "y _ {t} = f _ {\\theta} (S, x _ {t}), \\tag {1}", + "image_path": "9604c61497c671311f17a60ab40eaf77389da623a67bafccdfb2a1bde10f8fa0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "f_{\\theta}(*)" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": " is the VICL model parametrized with " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": " is the corresponding output label of " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": ". Usually, we assume " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": " are drawn from the same distribution. However, this is rarely the case for real-world deployment. Therefore, in this work, we mainly consider the scenario where a distribution shift like an image corruption occurs for " + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 491, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 574 + ], + "type": "text", + "content": "In the experiment section, we will show how the presence of such distribution shifts significantly degrades the performance of the VICL models. Moreover, we will additionally show how simple solutions like providing in-context examples from the target (corrupted) distribution would not fix the problem due to the limited generalization ability of the current VICL models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 506, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 506, + 594 + ], + "type": "text", + "content": "3.2. Test-Time Visual In-Context Tuning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 599, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 555, + 696 + ], + "type": "text", + "content": "In this work, we propose to leverage test-time training [41] to counteract the distribution shifts between training and testing data, with the goal of making VICL models more robust. We argue that VICL models are particularly amenable to be optimized at test time on arbitrary tasks thanks to the availability of the " + }, + { + "bbox": [ + 313, + 599, + 555, + 696 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 599, + 555, + 696 + ], + "type": "text", + "content": " shots provided as context. The core intuition of our method is to use the provided in-context examples to define a self-supervised loss that can be used at" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 502, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 502, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 502, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 325, + 702, + 502, + 713 + ], + "type": "text", + "content": "Following previous works, we set " + }, + { + "bbox": [ + 325, + 702, + 502, + 713 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 325, + 702, + 502, + 713 + ], + "type": "text", + "content": " in practice." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19998" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 91, + 287, + 295 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 293, + 84 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 293, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 293, + 84 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of VICT in a PyTorch-like style." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 91, + 287, + 295 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 287, + 295 + ], + "type": "text", + "content": "f: VICL model (e.g., Painter) \n# theta_0: pre-trained model weights \n# x: task prompt input \n# y: task prompt output \n# mask_token: mask token for the empty cell \n# steps: number of test-time optimization steps \nfor x_t in loader: # load a test sample x_t \nf.params = theta_0 # initialize \n# test-time optimization \nfor t in range(steps): \n # predict the test sample output \n y_t_pred = f((x, y, x_t, mask_token)) \n # predict the task prompt output \n y_pred = f((x, mask_token, x_t, y_t_pred)) \n # regression loss \n loss = SmoothL1Loss(y_pred, y) \n # update model \n loss/backward() \nupdate(f.params) \n# inference with the updated model \ny_t_pred = f((x, y, x_t, mask_token))" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 55, + 325, + 294, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 294, + 349 + ], + "type": "text", + "content": "test time to train any VICL models on the fly on any task. The overall pipeline of our VICT is illustrated in Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "content": "Specifically, given a pair of task prompts " + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "content": " and a test input " + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "content": ", we construct a grid-like image canvas with four cells, denoted by an ordered quadruple " + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "inline_equation", + "content": "I = (x,y,x_{t},\\emptyset)" + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "content": " is an empty cell. We then feed " + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "content": " into the VICL model " + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "content": " to predict the test output " + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 55, + 349, + 295, + 409 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 418, + 295, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 418, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 107, + 418, + 295, + 432 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {t} = f _ {\\theta} (I) = f _ {\\theta} \\left(\\left(x, y, x _ {t}, \\varnothing\\right)\\right). \\tag {2}", + "image_path": "655616485e2346af5d3d6c5fb189d508a220cc15689d66ee9407c3753fe9b77d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": "Afterward, with the obtained test output, we flip the role between input-output task prompts and input-output test samples, creating a new four-cell grid-like image canvas with the order of " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "I' = (x, \\varnothing, x_t, \\hat{y}_t)" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": ". The new " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": " is fed into " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": " again to predict the task prompt output " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 510, + 295, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 510, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 295, + 523 + ], + "type": "interline_equation", + "content": "\\hat {y} = f _ {\\theta} \\left(I ^ {\\prime}\\right) = f _ {\\theta} \\left(\\left(x, \\varnothing , x _ {t}, \\hat {y} _ {t}\\right)\\right). \\tag {3}", + "image_path": "f284fb446ae85b1d5becc5773f8d71e7e14a6244712e0e0afd498cc64933b864.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 568 + ], + "type": "text", + "content": "We then use a simple regression loss in pixel space to optimize the model by minimizing the distance between " + }, + { + "bbox": [ + 55, + 533, + 295, + 568 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 533, + 295, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 295, + 568 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 533, + 295, + 568 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 568, + 294, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 568, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 129, + 568, + 294, + 585 + ], + "type": "interline_equation", + "content": "\\theta_ {x _ {t}} = \\arg \\min _ {\\theta} \\mathcal {L} (\\hat {y}, y) \\tag {4}", + "image_path": "ebc8437b3eff1988d4f8880939d99f1e218939b1b4de0ada0733a82a649ab345.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "text", + "content": "In practice, we follow [52] to use smooth-" + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "text", + "content": " [19] loss. Using this loss we can optimize the parameters " + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "text", + "content": " of the model for a small number of steps for each " + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "text", + "content": " and, eventually, make a prediction on " + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "inline_equation", + "content": "f_{\\theta_{x_t}}((x,y,x_t,\\emptyset))" + }, + { + "bbox": [ + 55, + 590, + 295, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 638, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 296, + 686 + ], + "type": "text", + "content": "In our formulation, the gradient-based optimization for Equation 4 always starts from the initial pre-trained model weights " + }, + { + "bbox": [ + 55, + 638, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 55, + 638, + 296, + 686 + ], + "type": "text", + "content": " for each test input. When a new test input arrives, we discard " + }, + { + "bbox": [ + 55, + 638, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\theta_{x_t}" + }, + { + "bbox": [ + 55, + 638, + 296, + 686 + ], + "type": "text", + "content": " and reset the weights to " + }, + { + "bbox": [ + 55, + 638, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 55, + 638, + 296, + 686 + ], + "type": "text", + "content": ". We follow" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "this strategy to not have any assumption on the test inputs and treat them independently at test time, i.e., we do not assume that they come from the same distribution. Our generic formulation is completely agnostic on the VICL models considered, on the type of tasks being solved, and on the type of in-context samples provided to it." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 144, + 508, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 144, + 508, + 157 + ], + "spans": [ + { + "bbox": [ + 326, + 144, + 508, + 157 + ], + "type": "text", + "content": "The pseudo-code of VICT is in Algorithm 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 171, + 395, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 171, + 395, + 184 + ], + "spans": [ + { + "bbox": [ + 314, + 171, + 395, + 184 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 191, + 447, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 447, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 447, + 204 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 209, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 484 + ], + "type": "text", + "content": "Tasks and datasets. Following [52], we evaluate the model performance on six representative vision tasks ranging from high-level visual understanding to low-level image processing, which includes depth estimation on NYUv2 [40], semantic segmentation on ADE20K [63], panoptic segmentation on COCO [27], image denoising on SIDD [1], image draining on the merged draining datasets [57], and low-light image enhancement on LoL [53]. To evaluate the robustness to distribution shifts of the model on these datasets, we follow the setup in [23, 31] to corrupt the aforementioned datasets and simulate hard distribution shifts. The corrupted datasets contain 15 types of corruptions, covering noise (gaussian noise, impulse noise, shot noise), blur (defocus blur, glass blur, motion blur, zoom blur), weather (fog, frost, snow), and digital (brightness, contrast, elastic transform, JPEG compression, pixelate) categories. Each type of corruption has five levels of severity. We denote the corrupted datasets as NYUv2-C, ADE20K-C, COCO-C, SIDD-C, Rain-C, and LoL-C, respectively. Due to the space constraints, the results in the main text are limited to the most severe level (i.e., level 5) that corresponds to the strongest distribution shift. We provide the results on the other four levels in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 485, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 663 + ], + "type": "text", + "content": "Baselines. Without loss of generality, we consider Painter [52] as our VICL model for its simplicity in design and its wide applicability. We study VICT in two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). For the one-shot setting, apart from comparing with the frozen Painter using one-shot corrupted examples as task prompts, we also consider a baseline that further trains Painter with the one-shot corrupted samples using the same pre-training objective as in [52]. We also consider training Painter with more few-shot examples using the same settings to examine to what extent VICT can outperform a larger number of few-shot fine-tuning of Painter." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Training details. In all experiments, we use the pre-trained Painter based on ViT-Large provided by the authors of [52]. We perform VICT using an AdamW [29] optimizer, with betas as (0.9, 0.999), a weight decay of 0, a batch size of 1," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2The image order is top left, top right, bottom left, and bottom right, respectively." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19999" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 123, + 550, + 696 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 114 + ], + "type": "text", + "content": "Table 1. System-level comparison on six representative vision tasks across 15 corruptions. Results are on corruption level 5. We consider two settings: (i) zero-shot setting, where the task prompts are from the training distribution (i.e., clean images), and (ii) one-shot setting, where the task prompts are from the test distribution (i.e., corrupted images). \"avg\" denotes the averaged results over 15 corruptions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 123, + 550, + 696 + ], + "lines": [ + { + "bbox": [ + 60, + 123, + 550, + 696 + ], + "spans": [ + { + "bbox": [ + 60, + 123, + 550, + 696 + ], + "type": "table", + "html": "
methodbrighcontdefocelastfogfrostgaussglassimpuljpegmotnpixelshotsnowzoomavg
(a) depth estimation NYUv2-C (A.Rel ↓)
zero-shot setting:
Painter0.1290.2150.7120.1090.1290.5360.2000.6120.1890.3860.1670.1870.1421.9510.2090.392
VICT0.1080.2160.6310.1080.1330.5760.1910.5410.1810.3100.1390.1340.1401.8560.2100.365
one-shot setting:
Painter0.1260.2850.7430.1090.1320.8530.9010.6220.9010.3920.1740.1940.4401.9640.2120.537
VICT0.0970.1930.1800.1070.1280.2270.2410.2100.2780.1590.1210.1140.1500.6620.2140.205
(b) semantic segmentation ADE20K-C (mIoU ↑)
zero-shot setting:
Painter40.411.727.931.235.023.324.825.626.138.631.840.425.99.413.127.0
VICT40.912.728.231.736.423.725.526.327.438.831.940.928.017.313.628.2
one-shot setting:
Painter40.911.727.531.335.022.119.725.820.338.531.340.323.26.513.325.8
VICT41.721.028.832.237.024.724.427.025.239.732.841.526.621.714.929.3
(c) panoptic segmentation COCO-C (PQ ↑)
zero-shot setting:
Painter38.115.525.530.033.326.024.923.825.531.927.634.926.319.812.526.4
VICT38.715.725.830.333.926.725.324.325.932.427.735.426.721.912.726.9
one-shot setting:
Painter38.114.224.830.033.125.221.523.722.031.527.534.623.817.212.125.3
VICT38.616.925.430.433.826.624.024.624.532.327.435.025.224.912.426.8
(d) denoising SIDD-C (PSNR ↑)
zero-shot setting:
Painter8.7225.4932.4833.2016.868.7917.5033.1817.5023.1829.4734.8425.279.8629.8023.08
VICT8.8625.6532.9132.9617.008.8417.5833.0017.5024.8629.6834.7625.219.7529.6723.22
one-shot setting:
Painter9.1225.1631.9933.2316.689.2417.9933.1418.3322.7329.3034.8724.6910.5029.7623.12
VICT13.9425.3232.4134.4516.4511.1520.1332.9619.9326.8030.1135.0124.9314.9029.9524.56
(e) deraining Rain-C (PSNR ↑)
zero-shot setting:
Painter10.7414.1519.5420.1013.9112.4919.8519.8020.1624.8618.5222.8120.1111.8917.9317.79
VICT10.8914.2419.5820.3113.9412.6220.4219.8320.6724.9718.5623.0520.4712.0517.9817.97
one-shot setting:
Painter11.0413.9519.6120.1513.8212.8719.5119.8719.5824.8518.6222.8819.6812.1617.9517.77
VICT17.3814.6920.3621.4414.6717.6922.3320.4822.2425.3119.0523.9221.3719.5318.2819.92
(f) low-light enhancement LoL-C (PSNR ↑)
zero-shot setting:
Painter16.2613.4318.8019.8311.2411.6212.5018.8212.9115.8718.0121.1217.7513.7716.6015.90
VICT16.3913.5719.2419.8411.3711.6412.5119.2313.0016.0018.1121.2019.1613.9217.0616.15
one-shot setting:
Painter16.7413.5018.8519.8311.0211.7613.1418.9513.7415.8418.1521.1216.5413.1916.5915.93
VICT18.1915.0318.9919.8811.7612.0114.9519.2214.9916.1218.2321.0919.8416.9616.9116.94
", + "image_path": "4132bf0007f71a0962ddc00eed905a825efe8f7123a5daca31f335f54839a7a1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20000" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 223, + 199 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 223, + 199 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 223, + 199 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 223, + 199 + ], + "type": "image", + "image_path": "5c21e893ebea43beca12e72336776524f2549dc0a77e2537b2fcfeacaa60eeda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 72, + 386, + 199 + ], + "blocks": [ + { + "bbox": [ + 223, + 72, + 386, + 199 + ], + "lines": [ + { + "bbox": [ + 223, + 72, + 386, + 199 + ], + "spans": [ + { + "bbox": [ + 223, + 72, + 386, + 199 + ], + "type": "image", + "image_path": "24a6e1ad11f78096a691db1425e9a292e7ddefb7e1104c44a4465a04bbe35ef9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 386, + 72, + 548, + 199 + ], + "blocks": [ + { + "bbox": [ + 386, + 72, + 548, + 199 + ], + "lines": [ + { + "bbox": [ + 386, + 72, + 548, + 199 + ], + "spans": [ + { + "bbox": [ + 386, + 72, + 548, + 199 + ], + "type": "image", + "image_path": "74a4ebeb11dadf3b69613d7892e0723dd12e57012f4e222f055cf1c1becb0ea5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 200, + 223, + 327 + ], + "blocks": [ + { + "bbox": [ + 59, + 200, + 223, + 327 + ], + "lines": [ + { + "bbox": [ + 59, + 200, + 223, + 327 + ], + "spans": [ + { + "bbox": [ + 59, + 200, + 223, + 327 + ], + "type": "image", + "image_path": "43db942e5f7350ca357479316f7acfcbb8f1aa2eb081fd57037946c60a4ce14f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 333, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 333, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 555, + 368 + ], + "type": "text", + "content": "Figure 3. Comparison with few-shot Painter on six vision tasks with corruptions. We randomly corrupt a certain number of images in the training set, using 1, 2, 4, 8, 16, 32, and 64 shots for training and deploying the model in the full corrupted test sets. We report the final results averaged across 15 corruptions. Our zero-shot or one-shot VICT can outperform Painter trained with more few-shot examples." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 200, + 386, + 327 + ], + "blocks": [ + { + "bbox": [ + 223, + 200, + 386, + 327 + ], + "lines": [ + { + "bbox": [ + 223, + 200, + 386, + 327 + ], + "spans": [ + { + "bbox": [ + 223, + 200, + 386, + 327 + ], + "type": "image", + "image_path": "28dfed92e636cab673193c1254943774664a2e793f89e4bc5c975bc488290c56.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 386, + 200, + 548, + 327 + ], + "blocks": [ + { + "bbox": [ + 386, + 200, + 548, + 327 + ], + "lines": [ + { + "bbox": [ + 386, + 200, + 548, + 327 + ], + "spans": [ + { + "bbox": [ + 386, + 200, + 548, + 327 + ], + "type": "image", + "image_path": "1ae8f4e72ae4967bcd35bf4d953259d2f92666579db157e14ed560ebcab99ca4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 379, + 297, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 379, + 297, + 619 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 297, + 619 + ], + "type": "text", + "content": "and a fixed learning rate of 1e-6. Unless otherwise specified, we train each test sample for 60 steps. The choice of 60 steps is purely computational and more steps are likely to further improve performance, judging from the positive trend observed in Figure 4. Note that we do not apply any data augmentations for VICT as many data augmentations are in fact distribution shifts in our evaluation benchmarks. Training with them is analogous to training on the test distributions. Therefore, to solely study the generalization ability to new test distributions, we purposely choose not to use any data augmentations, even though they could improve our results at face value. Every iteration of VICT is performed on the same concatenated grid-like image, with each subimage resized as " + }, + { + "bbox": [ + 54, + 379, + 297, + 619 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 54, + 379, + 297, + 619 + ], + "type": "text", + "content": ", which is the same as what we later use for in-context inference. During VICT, we only optimize the encoder weights. We have experimented with training both the encoder and the decoder and found that the difference is negligible, which is also consistent with the observations in [16, 41]. A detailed comparison is provided in the experiment section." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 141, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 141, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 141, + 635 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "System-level comparison. We compare VICT with Painter under 15 common corruptions as introduced in Section 4.1. We perform experiments on six representative vision tasks covering high-level visual understanding and low-level image processing, which includes depth estimation on NYUv2-C, semantic segmentation on ADE20K-C, panop-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 379, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 415 + ], + "type": "text", + "content": "tic segmentation on COCO-C, image denoising on SIDD-C, image draining on Rain-C, and low-light image enhancement on LoL-C." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 421, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 556, + 601 + ], + "type": "text", + "content": "Table 1 reports the results. First of all, we observe that Painter exhibits poor generalization ability under common corruptions. For example, on ADE20K-C semantic segmentation, Painter only achieves an average performance of " + }, + { + "bbox": [ + 313, + 421, + 556, + 601 + ], + "type": "inline_equation", + "content": "27.0\\mathrm{mIoU}" + }, + { + "bbox": [ + 313, + 421, + 556, + 601 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 313, + 421, + 556, + 601 + ], + "type": "inline_equation", + "content": "22.9\\mathrm{mIoU}" + }, + { + "bbox": [ + 313, + 421, + 556, + 601 + ], + "type": "text", + "content": " lower than that on the clean validation set (49.9 mIoU as reported in Figure 1). Nevertheless, our VICT outperforms Painter by clear margins in both zero-shot and one-shot settings across different tasks. When it comes to the zero-shot setting, VICT achieves an average of -0.027 A.Rel on NYUv2-C depth estimation, +1.2 mIoU on ADE20K-C semantic segmentation, +0.5 PQ on COCO-C panoptic segmentation, +0.14 PSNR on SIDD-C image denosing, +0.18 PSNR on Rain-C image draining, and +0.25 PSNR on LoL-C low-light enhancement, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "The performance gains of VICT over Painter are further increased when we shift from the zero-shot setting to the one-shot counterpart. Specifically, VICT achieves an average of -0.332 A.Rel on NYUv2-C depth estimation, +3.5 mIoU on ADE20K-C semantic segmentation, +1.5 PQ on COCO-C panoptic segmentation, +1.44 PSNR on SIDC image denoising, +2.15 PSNR on Rain-C image draining, and +1.01 PSNR on LoL-C low-light enhancement, respectively. Besides, for the Painter baseline, using the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20001" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 102, + 533, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Test-time optimization on different modules. We use semantic segmentation on ADE20K-C for the ablation. Training only the encoder performs similarly to training both the encoder and decoder, regardless of the zero-shot and one-shot settings." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 102, + 533, + 198 + ], + "lines": [ + { + "bbox": [ + 79, + 102, + 533, + 198 + ], + "spans": [ + { + "bbox": [ + 79, + 102, + 533, + 198 + ], + "type": "table", + "html": "
modulebrighcontdefocelastfogfrostgaussglassimpuljpegmotnpixelshotsnowzoomavg
zero-shot setting:
Encoder40.912.728.231.736.423.725.526.327.438.831.940.928.017.313.628.2
Both40.912.528.231.736.423.825.726.527.438.831.941.027.916.913.528.2
one-shot setting:
Encoder41.721.028.832.237.024.724.427.025.239.732.841.526.621.714.929.3
Both41.721.028.732.136.924.424.427.025.339.633.041.626.821.515.129.3
", + "image_path": "4e4ce61c726ca0756df862ffe384e5e1c539af9a850f4449e8e7facce356b43f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 217, + 294, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 217, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 294, + 348 + ], + "type": "text", + "content": "one-shot setting does not always perform better than the zero-shot setting. The average performances are even degraded in most cases, e.g., depth estimation on NYUv2-C (0.392 A.Rel vs. 0.537 A.Rel), semantic segmentation on ADE20K-C (27.0 mIoU vs. 25.8 mIoU), panoptic segmentation on COCO-C (26.4 PQ vs. 25.3 PQ), image draining on Rain-C (17.79 PSNR vs. 17.77 PSNR). In contrast, for VICT, using the one-shot setting usually performs better than the zero-shot setting. It is worth mentioning that our zero-shot VICT can even outperform the one-shot Painter, further demonstrating the effectiveness of VICT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 349, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 295, + 456 + ], + "type": "text", + "content": "Comparison with few-shot Painter. In previous experiments, we mainly compare with a frozen Painter using either clean or corrupted examples as the task prompts. Here, we further consider a scenario where more few-shot examples from the test distribution are available. Specifically, we randomly corrupt a certain number of images in the training set, using 1, 2, 4, 8, 16, 32, and 64 shots to train Painter respectively and deploying the model in the full corrupted test sets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 458, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 458, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 56, + 458, + 295, + 661 + ], + "type": "text", + "content": "The results are shown in Figure 3. Both our zero-shot VICT and one-shot VICT can consistently outperform the one-shot Painter. When more few-shot examples are available for training, our VICT can still outperform Painter below a certain threshold. We observe such a threshold varies for different tasks, with high-level visual understanding tasks having a higher threshold. It is worth noting that in high-level visual understanding tasks like ADE20K-C semantic segmentation and COCO-C panoptic segmentation, both our zero-shot and one-shot variants of VICT nearly match or even outperform the 64-shot Painter. In other words, with only one or simply no labeled test sample(s), VICT can achieve similar performances as models that use significantly more labeled test samples to train. This phenomenon is rather appealing as in practice it is rare to know test distribution in advance, let alone obtaining a large number of labeled test samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 157, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 157, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 157, + 684 + ], + "type": "text", + "content": "4.3. Further Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Effect of test-time optimization modules. We study the effect of optimizing different modules at test time. We con" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 217, + 548, + 388 + ], + "blocks": [ + { + "bbox": [ + 319, + 217, + 548, + 388 + ], + "lines": [ + { + "bbox": [ + 319, + 217, + 548, + 388 + ], + "spans": [ + { + "bbox": [ + 319, + 217, + 548, + 388 + ], + "type": "image", + "image_path": "83e0095381d50667584687df0c79c3e0135b9cc55c44edaa81432e54ff6c821d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 396, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 440 + ], + "type": "text", + "content": "Figure 4. Analysis on the trade-off between efficiency and accuracy. We use semantic segmentation on ADE20K-C for the ablation. VICT benefits from more training steps, while at the cost of linearly increased training time." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "content": "sider two parameter groups: (i) the encoder, and (ii) both the encoder and decoder. We use semantic segmentation on ADE20K-C as an example task for analysis. The results are shown in Table 2. Training only the encoder performs similarly to training the entire model, regardless of the zero-shot and one-shot settings. This makes sense since the encoder acts as a feature extractor, which plays a key role in determining the representation quality. In contrast, the decoder is only responsible for mapping the latent representation back to its original resolution, which is more specialized for reconstruction but less relevant for semantics. Thus, we choose only to optimize the encoder at test time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Effect of test-time optimization steps. We study the effect of test-time optimization steps in Figure 4. VICT benefits from more training steps, which keeps improving performance even after 60 steps. However, it should be noted that the runtime of our VICT is linearly increased with the number of training steps. For reference, it takes around 0.4 seconds per step per test sample on a single A100 GPU. Thus, in practice, one may decide the number of training steps according to the pre-defined cost budget for deployment." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20002" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 293, + 373 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 293, + 373 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 293, + 373 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 293, + 373 + ], + "type": "image", + "image_path": "0860baaaed2f1b52f5e0044910dd92971ea4a29f725376445dc3aedbc2e1c684.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "text", + "content": "Figure 5. Visualizations of test examples and predictions for six main tasks with corruptions. We visualize both zero-shot and one-shot settings for Painter and VICT. Zoom in for best view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 431, + 167, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 167, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 167, + 443 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 533 + ], + "type": "text", + "content": "Results on main tasks. We visualize some test examples and predictions from the validation set for different tasks with corruptions, including depth estimation, semantic segmentation, instance segmentation, image denoising, image deraining, and low-light image enhancement. As shown in Figure 5, VICT can make more accurate predictions than Painter on all tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Results on unseen tasks. In this work, we mainly consider the distribution shifts with corruptions and conduct experiments on the tasks that are seen in training. Here, we further verify whether VICT can generalize to unseen tasks. We explore this capability via visualizations. More quantitative results are provided in the supplementary material. Figure 6 provides examples of two unseen tasks including foreground object segmentation and colorization. Both Painter and our VICT can generalize to the foreground segmentation task. This makes sense as similar segmentation tasks (e.g., semantic segmentation and instance segmentation) have been seen during the training stage of Painter. Nevertheless, our VICT further reduces noises and produces finer masks. However, Painter cannot generalize to the colorization task that is totally unseen during training, produc" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 245 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 245 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 245 + ], + "type": "image", + "image_path": "9e0376eb6755de968ec4fe00585dd9737bb5ce81b14b98eeb699e9dac885fbe7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 253, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 308 + ], + "type": "text", + "content": "Figure 6. Visualizations of test examples and predictions for unseen tasks. The visualized tasks include foreground object segmentation and colorization. Painter cannot generalize to totally unseen tasks like colorization, whereas VICT can make decent color predictions. Zoom in for best view." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 324, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 373 + ], + "type": "text", + "content": "ing only grayscale images or even wrong predictions from other tasks (e.g., the last row). In contrast, our VICT can produce decent colorful results. This further demonstrates the potential of applying VICT for unseen tasks at test time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 382, + 388, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 388, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 388, + 395 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 510 + ], + "type": "text", + "content": "Generalization under distribution shifts is the central theme of deep learning. In this work, we have studied the robustness of VICL models and found that the existing VICL paradigm exhibits poor generalization capability to unseen new domains. Based on this observation, we proposed a simple yet effective test-time visual in-context tuning method to adapt VICL models to a single test sample on the fly. We hope our explorations can pave the way for improving the generalizability of VICL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Limitations and future work. Our study has several limitations: 1) Since we perform training for each test sample, our method is slower than the baseline applying a fixed model at test time, which is a common limitation for test-time training. Inference speed might be improved through better architectural designs, training techniques, optimizers, and hyper-parameters. However, it has not been the focus of this paper. 2) Our proposed cycle consistency supervision is a general self-supervised task. However, we cannot guarantee that it will produce useful gradients for every single test distribution. We only focus on several representation vision tasks and most popular corruption benchmarks for distribution shifts. More tasks and benchmarks can be studied for future research. One possible extension of this work is to apply similar test-time supervision in multi-modal in-context learning, where cycle consistency supervision also exists. We leave this exploration for future work." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20003" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "text", + "content": "[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. In CVPR, 2018. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 178 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "text", + "content": "[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 294, + 258 + ], + "type": "text", + "content": "[4] Pratyay Banerjee, Tejas Gokhale, and Chitta Baral. Self-supervised test-time learning for reading comprehension. In NAACL, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 260, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 295, + 293 + ], + "type": "text", + "content": "[5] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. In NeurIPS, 2022. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 295, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 327 + ], + "type": "text", + "content": "[6] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In AISTATS, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 294, + 382 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 385, + 294, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 294, + 417 + ], + "type": "text", + "content": "[8] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 419, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 294, + 452 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 453, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 453, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 57, + 453, + 294, + 497 + ], + "type": "text", + "content": "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. JMLR, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 498, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 294, + 520 + ], + "type": "text", + "content": "[11] Samuel Dodge and Lina Karam. Quality resilient deep neural networks. arXiv preprint arXiv:1703.08119, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 522, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 554 + ], + "type": "text", + "content": "[12] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 556, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 294, + 609 + ], + "type": "text", + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 612, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 294, + 644 + ], + "type": "text", + "content": "[14] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In ICCV, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[15] Yang Fu, Sifei Liu, Umar Iqbal, Shalini De Mello, Humphrey Shi, and Jan Kautz. Learning to track instances without video annotations. In CVPR, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[16] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In NeurIPS, 2022. 2, 3, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[17] Robert Geirhos, Carlos RM Temme, Jonas Rauber, Heiko H Schütt, Matthias Bethge, and Felix A Wichmann. Generalisation in humans and deep neural networks. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "text", + "content": "[18] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 153, + 497, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 497, + 164 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 497, + 164 + ], + "type": "text", + "content": "[19] Ross Girshick. Fast r-cnn. In ICCV, 2015. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 166, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 166, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 166, + 553, + 209 + ], + "type": "text", + "content": "[20] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In ICLR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 212, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 212, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 553, + 255 + ], + "type": "text", + "content": "[21] Yaru Hao, Haoyu Song, Li Dong, Shaohan Huang, Zewen Chi, Wenhui Wang, Shuming Ma, and Furu Wei. Language models are general-purpose interfaces. arXiv preprint arXiv:2206.06336, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 257, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 257, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 257, + 553, + 289 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 291, + 553, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 323 + ], + "type": "text", + "content": "[23] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019. 1, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "text", + "content": "[24] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "text", + "content": "[25] Wei Li, Jiahao Xie, and Chen Change Loy. Correlational image modeling for self-supervised visual pre-training. In CVPR, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "type": "text", + "content": "[26] Zhi Li, Shaoshuai Shi, Bernt Schiele, and Dengxin Dai. Test-time domain adaptation for monocular depth estimation. In ICRA, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 430, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 472 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "text", + "content": "[28] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 521, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 542 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "text", + "content": "[30] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision-language models. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 578, + 553, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 642 + ], + "type": "text", + "content": "[31] Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019. 1, 2, 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 646, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 667 + ], + "type": "text", + "content": "[32] Mirco Mutti and Aviv Tamar. Test-time regret minimization in meta reinforcement learning. In ICML, 2024. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[33] Muhammad Ferjad Naeem, Yongqin Xian, Xiaohua Zhai, Lukas Hoyer, Luc Van Gool, and Federico Tombari. Silc: Improving vision language pretraining with self-distillation. In ECCV, 2024. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20004" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "text", + "content": "[34] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "text", + "content": "[35] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "text", + "content": "[36] Matteo Poggi, Alessio Tonioni, Fabio Tosi, Stefano Mattoccia, and Luigi Di Stefano. Continual adaptation for deep stereo. TPAMI, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 219 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "text", + "content": "[38] Mattia Segu, Bernt Schiele, and Fisher Yu. Darth: holistic test-time adaptation for multiple object tracking. In ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 295, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 298 + ], + "type": "text", + "content": "[39] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 294, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 332 + ], + "type": "text", + "content": "[40] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "text", + "content": "[41] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICLR, 2020. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 378, + 295, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 295, + 433 + ], + "type": "text", + "content": "[42] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "type": "text", + "content": "[43] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 489, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 533 + ], + "type": "text", + "content": "[44] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IROS, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 534, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 567 + ], + "type": "text", + "content": "[45] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In CVPR, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 568, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 624 + ], + "type": "text", + "content": "[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 680 + ], + "type": "text", + "content": "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[48] Yun-Yun Tsai, Chengzhi Mao, and Junfeng Yang. Convolutional visual prompt for robust visual perception. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 624 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[49] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "type": "text", + "content": "[50] Igor Vasiljevic, Ayan Chakrabarti, and Gregory Shakhnarovich. Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "text", + "content": "[51] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 186, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 553, + 219 + ], + "type": "text", + "content": "[52] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 315, + 220, + 553, + 252 + ], + "type": "text", + "content": "[53] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In BMVC, 2018. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 254, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 254, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 315, + 254, + 553, + 286 + ], + "type": "text", + "content": "[54] Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Delving into inter-image invariance for unsupervised visual representations. IJCV, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 287, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 287, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 315, + 287, + 553, + 320 + ], + "type": "text", + "content": "[55] Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency modeling for self-supervised visual pre-training. In ICLR, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 322, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 322, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 315, + 322, + 553, + 364 + ], + "type": "text", + "content": "[56] Hai Ye, Yuyang Ding, Juntao Li, and Hwee Tou Ng. Robust question answering against distribution shifts with test-time adaptation: An empirical study. In Findings of EMNLP, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 365, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 365, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 315, + 365, + 553, + 410 + ], + "type": "text", + "content": "[57] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In ECCV, 2020. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 411, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 411, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 315, + 411, + 553, + 444 + ], + "type": "text", + "content": "[58] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 445, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 445, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 553, + 477 + ], + "type": "text", + "content": "[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 478, + 553, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 478, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 315, + 478, + 553, + 512 + ], + "type": "text", + "content": "[60] Xiaohang Zhan, Jiahao Xie, Ziwei Liu, Yew-Soon Ong, and Chen Change Loy. Online deep clustering for unsupervised representation learning. In CVPR, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 513, + 553, + 555 + ], + "type": "text", + "content": "[61] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 556, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 556, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 315, + 556, + 553, + 590 + ], + "type": "text", + "content": "[62] Stephan Zheng, Yang Song, Thomas Leung, and Ian Goodfellow. Improving the robustness of deep neural networks via stability training. In CVPR, 2016. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 591, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 591, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 315, + 591, + 553, + 624 + ], + "type": "text", + "content": "[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2, 4" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20005" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_content_list.json b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3804a96a7f9f316636a2e02fbd5cb05a0c5025ba --- /dev/null +++ b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_content_list.json @@ -0,0 +1,1384 @@ +[ + { + "type": "text", + "text": "Test-time augmentation improves efficiency in conformal prediction", + "text_level": 1, + "bbox": [ + 158, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Divya Shanmugam, Helen Lu, Swami Sankaranarayanan, John Guttag Massachusetts Institute of Technology, CSAIL", + "bbox": [ + 220, + 180, + 776, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{divyas, helenlu, swamiviv, guttag}@mit.edu", + "bbox": [ + 305, + 218, + 684, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 328, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A conformal classifier produces a set of predicted classes and provides a probabilistic guarantee that the set includes the true class. Unfortunately, it is often the case that conformal classifiers produce uninformatively large sets. In this work, we show that test-time augmentation (TTA)-a technique that introduces inductive biases during inference-reduces the size of the sets produced by conformal classifiers. Our approach is flexible, computationally efficient, and effective. It can be combined with any conformal score, requires no model retraining, and reduces prediction set sizes by $10\\% - 14\\%$ on average. We conduct an evaluation of the approach spanning three datasets, three models, two established conformal scoring methods, different guarantee strengths, and several distribution shifts to show when and why test-time augmentation is a useful addition to the conformal pipeline.", + "bbox": [ + 88, + 300, + 485, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 571, + 220, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conformal prediction has emerged as a promising way to equip existing classifiers with statistically valid uncertainty estimates. It does so by replacing the prediction of the most likely class with an uncertainty set-a set of classes accompanied by a probabilistic guarantee that the true class appears in the set [36].", + "bbox": [ + 89, + 597, + 482, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conformal prediction faces two limitations in practice. First, achieving a suitably strong guarantee often leads to prediction sets that are uninformatively large [10, 47]. Large prediction sets have been shown to decrease performance when provided as a decision-making aid [3, 10, 47].", + "bbox": [ + 89, + 688, + 482, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Second, conformal classifiers inherit the instability of the underlying models. As a result, prediction sets can change significantly in response to small input perturbations, a well-known weakness of neural networks [39]. Applying a horizontal flip to each image in ImageNet, for example, changes the prediction set sizes for $75\\%$ of examples at a coverage guarantee of $99\\%$ . Such behavior also represents a barrier to broader use.", + "bbox": [ + 89, + 763, + 482, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we show that test-time augmentation", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(TTA), a widely-used technique in computer vision, has the potential to address both limitations. TTA involves generating an ensemble of predictions by perturbing the input with label-preserving transformations. It has previously been shown that TTA can be used to make non-conformal classifiers more robust to small input perturbations [7], more accurate [37], and better calibrated [18]. However, previous work has not explored the utility of TTA in the context of conformal prediction.", + "bbox": [ + 511, + 268, + 906, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose test-time-augmented conformal prediction, which transforms a classifier's predictions using a learned test-time augmentation policy prior to conformal classification. By using distinct sets of labeled data to learn the test-time augmentation policy and the conformal classifier, we preserve the assumption of exchangeability, and thereby the coverage guarantee associated with the conformal predictor.", + "bbox": [ + 511, + 409, + 908, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In experiments testing the performance of conformal predictors subject to distribution shift, we see that test-time augmentation reduces prediction set sizes by $14\\%$ on average, with no loss of coverage. And even when there is no distribution shift, we see a reduction of $10\\%$ on average. Moreover, we find that classes with the largest average prediction set sizes benefit most from the introduction of test-time augmentation. We also show that test-time augmentation can bridge gaps between classifiers of different sizes (e.g., we show that test-time augmentation combined with ResNet-50 produces smaller set sizes than ResNet-101 without test-time augmentation).", + "bbox": [ + 511, + 518, + 908, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our analysis of why test-time augmentation reduces prediction set sizes reveals a previously unknown effect of test-time augmentation. Specifically, TTA increases the predicted probability of the true class even when it is predicted to be unlikely (for example, promoting the true class from 200th most likely to 100th most likely). Although such behavior has no impact on the maximum predicted probability — commonly the focus of literature on test-time augmentation [2, 32, 37] — it is valuable in conformal prediction. This is because when the true class is promoted to a higher rank among a classifier's predicted probabilities, the conformal classifier includes fewer incorrect classes to meet the conformal guarantee.", + "bbox": [ + 511, + 704, + 910, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20622", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f09daba25237aab21fb82fc714476f5b309ea49fbe8010157df0b208660700a.jpg", + "image_caption": [ + "Test-time augmented conformal prediction..." + ], + "image_footnote": [], + "bbox": [ + 98, + 119, + 637, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aac26790bf41d560b74786c3f60b128359b374c21758c4201e7e752e9a50f1a2.jpg", + "image_caption": [ + "...leads to smaller prediction set sizes.", + "Figure 1. We illustrate the addition of test-time augmentation to conformal calibration in green (left) and provide a snapshot of the improvements it can confer (right). We show results on Imagenet, with a desired coverage of $95\\%$ , for the 20 classes with the largest predicted set sizes on average (computed over 10 calibration/test splits)." + ], + "image_footnote": [], + "bbox": [ + 663, + 117, + 892, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions We make three contributions. To begin, this is the first work to propose combining test-time augmentation, a popular technique in computer vision, with conformal prediction. Second, we present a method that reduces the prediction set sizes of existing conformal predictors by using automatically learned test-time augmentations. Finally, we demonstrate, in an extensive set of experiments, that our approach to combining conformal prediction and test-time augmentation leads to smaller prediction sets.", + "bbox": [ + 89, + 392, + 483, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 561, + 227, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conformal prediction was first introduced by Gammerman et al. [16], and further developed by Saunders and Holloway [35] and Vladimir Vovk [42]. We review efforts to ensemble conformal predictors and efforts to reduce prediction set sizes below.", + "bbox": [ + 89, + 588, + 483, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ensembles in conformal prediction Several methods that generate ensembles of conformal predictors are known to improve efficiency. These methods include cross-conformal prediction [43], bootstrap conformal prediction [44], aggregated conformal prediction [5, 25], and out-of-bag conformal prediction [26]. The approaches primarily differ in how data is sampled to create the training dataset for the classifier and the calibration dataset for the conformal predictor. However, all require training multiple base classifiers or conformal predictors. Our approach is distinct: we propose a technique to generate an ensemble from a single model by perturbing the input, which requires no additional base models and no additional conformal predictors.", + "bbox": [ + 89, + 689, + 483, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficiency in conformal prediction There are two ways to improve efficiency in split conformal prediction: adjustments to the conformal score or improvements to the underlying classifier. Many works have proposed new procedures to estimate and apply thresholds on conformal scores [1, 4, 13, 33, 40]. Romano et al. [34] proposed APS, a conformal score based on the cumulative probability required to include the correct class in a prediction set. Angelopoulos et al. [1] built on this work to propose RAPS, which modifies APS by penalizing the inclusion of low-probability classes. Comparatively little work has focused on improvements to the underlying model. Jensen et al. [20] ensemble a set of base classifiers, where the classifiers are created by training models on subsets of the training data. Stutz et al. [38] provide a new way to train the base classifier and conformal predictor jointly through a novel conformal training loss. In contrast, our work focuses on improving the underlying model without retraining, and can be easily combined with any of the above procedures.", + "bbox": [ + 511, + 392, + 906, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-Time Augmentation Test-time augmentation (TTA) is a popular technique to improve the accuracy, robustness, and calibration of an existing classifier by aggregating predictions over a set of input transformations [2, 9, 15, 18, 32, 37, 48]. TTA has been applied to a diverse range of predictive tasks across domains ranging from healthcare [8] to content moderation [27]. Consequently, many have proposed new ways to perform TTA—for example, learning when to apply TTA [29], which augmentations to use [6, 21, 28], and how to aggregate the resulting predictions [6, 9, 37]. Existing work typically focuses on test-time augmentation's impact on highest predicted probability. Here, we analyze how test-time augmentation increases the pre", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20623", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dicted probability assigned to the true class when it appears outside the top few classes, and how that change is consequential in conformal prediction.", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem setting", + "text_level": 1, + "bbox": [ + 89, + 151, + 250, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We operate within the split conformal prediction framework. In this setting, a conformal classifier $\\mathcal{C}(x_i) \\subset \\{1, \\ldots, K\\}$ maps input $x_i$ to a subset of $K$ possible classes and requires three inputs:", + "bbox": [ + 89, + 176, + 482, + 237 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Calibration set $D^{(cal)} = \\{(x_1, y_1), \\ldots, (x_n, y_n)\\}$ , containing $n$ labeled examples.", + "- Classifier $f: \\mathcal{X} \\mapsto \\Delta^K$ , mapping input domain $\\mathcal{X}$ to a probability distribution over $K$ classes.", + "- Desired upper bound on error rate $\\alpha \\in [0,1]$ , where $(1 - \\alpha)$ is the probability the set contains the true class." + ], + "bbox": [ + 89, + 239, + 482, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We study the introduction of two variables drawn from the test-time augmentation literature:", + "bbox": [ + 89, + 334, + 482, + 364 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Augmentation policy $\\mathcal{A} = \\{a_0, \\dots, a_{m-1}\\}$ , consisting of $m$ augmentation functions, where $a_0$ is the identity transform.", + "- Aggregation function $\\hat{g}$ , which aggregates a set of predictions to produce one prediction." + ], + "bbox": [ + 89, + 367, + 482, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each variable translates to a key choice in test-time augmentation: what augmentations to apply $(\\mathcal{A})$ and how to aggregate the resulting probabilities $(\\hat{g})$ .", + "bbox": [ + 89, + 446, + 483, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 89, + 506, + 200, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to learn an aggregation function $\\hat{g}$ to maximize the accuracy of the underlying classifier, and ultimately reduce the sizes of the prediction sets generated from the classifier's predicted probabilities. We briefly outline the conformal approach, and then detail the mechanics of our method (illustrated in Figure 1). For a detailed introduction to conformal prediction, refer to Shafer and Vovk [36].", + "bbox": [ + 89, + 531, + 483, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conformal predictors accept three inputs: a probabilistic classifier $f$ , a calibration set $\\mathcal{D}^{(cal)}$ , and a pre-specified error rate $\\alpha$ . Using the these inputs, one can construct a conformal predictor in three steps:", + "bbox": [ + 89, + 638, + 483, + 698 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Define a score function $c(x_i, y_i)$ , which produces a conformal score representing the uncertainty of the input example and label pair.", + "2. Produce a distribution of conformal scores by computing $c(x_i, y_i)$ for all $(x_i, y_i) \\in \\mathcal{D}^{(cal)}$ .", + "3. Compute threshold $\\hat{q}$ as the $\\lceil (n + 1)(1 - \\alpha)\\rceil /n$ quantile of the distribution of conformal scores over $n$ examples in the calibration set, combined with $\\{\\infty \\}$ ." + ], + "bbox": [ + 89, + 700, + 482, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a new example $x$ , we compute $c(x_i, y)$ for all $y \\in \\{1, \\dots, K\\}$ , and include all $y$ for which $c(x_i, y) < \\hat{q}$ . We adopt the conformal score proposed by Romano et al. [34], which equates to the cumulative probability required to include the correct class:", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) = \\hat {p} \\left(y = y ^ {\\prime} \\mid x\\right) = f (x) _ {y ^ {\\prime}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 118, + 906, + 136 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {x} (y) = \\sum_ {y ^ {\\prime} = 1} ^ {K} \\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) \\mathbb {I} \\left[ \\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) > \\hat {\\pi} _ {x} (y) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 138, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc (x, y) = \\rho_ {x} (y) + u \\cdot \\hat {\\pi} _ {x} (y) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 186, + 906, + 202 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\rho_{x}(y)$ is the cumulative probability of all classes predicted with higher probability than $y$ and $\\hat{\\pi}_x(y')$ corresponds to the predicted probability of class $y^\\prime$ given $x$ . The variable $u\\sim U(0,1)$ . Conformal score $c(x_i,y_i)$ is thus composed of this cumulative probability and the predicted probability of class $y_{i}$ .", + "bbox": [ + 511, + 229, + 906, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Proposal** Our approach differs from prior work in that the conformal score is derived by transforming the probabilities output by $f$ using test-time augmentation. Concretely, this replaces Equation 1 with the following, parametrized by augmentation policy $\\mathcal{A}$ and augmentation weights $\\theta$ .", + "bbox": [ + 511, + 342, + 905, + 417 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) = \\hat {p} \\left(y = y ^ {\\prime} \\mid x _ {i}\\right) = g \\left(x _ {i}; f, \\mathcal {A}, \\theta\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 445, + 906, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Aggregation weights $\\theta$ are applied to the logits output by classifier $f$ , and transformed to be proper probabilities by applying a softmax function. We learn the aggregation weights $\\theta$ using a portion of the validation set, $D^{(TTA)}$ , distinct from calibration set used to identify the conformal threshold $(D^{(cal)})$ . In contrast to traditional approaches, where all labeled data is used to estimate the conformal threshold, we instead reserve a portion to learn the test-time augmentation policy.", + "bbox": [ + 511, + 476, + 905, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We learn a set of weights which maximize classification accuracy on $D^{(TTA)}$ by minimizing the cross-entropy loss1 computed between the predicted probabilities and true labels. More formally, $g$ applies $\\theta$ and $\\mathcal{A}$ as follows:", + "bbox": [ + 511, + 612, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng \\left(x _ {i}; f, \\mathcal {A}, \\Theta\\right) = \\sigma \\left(\\theta^ {T} \\mathbf {A} \\left(f, \\mathcal {A}, x _ {i}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 700, + 906, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{A}$ uses $f$ to map input $x_{i}$ to a $M\\times K$ matrix of predicted logits where $M$ is the number of augmentations and $K$ is the number of classes. $\\theta$ is a $1\\times m$ vector corresponding to augmentation-specific weights. Each row in $\\mathbf{A}(f,\\mathcal{A},x_i)$ represents the pre-trained classifier's predicted logits on augmentation $a_{m}$ of $x_{i}$ . TTA-Learned refers to TTA combined with learned augmentation weights, while TTA-Avg refers to a simple average over the augmentations.", + "bbox": [ + 511, + 729, + 906, + 852 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We found no significant improvement by using alternate losses considered in the conformal prediction literature (e.g. focal or conformal training loss). See Table S7 in the Appendix.", + "bbox": [ + 511, + 862, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20624", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We refer to the fraction of the validation set allotted to $D^{(TTA)}$ as $\\beta$ . Figure S3 shows that performance is not sensitive to the choice of $\\beta$ ; as a result, all experiments use $\\beta = .2$ (see supplement for further discussion). This does reduce the amount of data available to identify the appropriate threshold, but we find that the benefits TTA confers outweigh the cost to threshold estimation. Computational cost scales linearly with the size of $\\mathcal{A}$ ; each additional augmentation translates to a forward pass of the base classifier. One can use the learned weights to save computation by identifying which test-time augmentations to generate.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preserving exchangeability The validity of conformal prediction depends upon the assumption of exchangeability: that all orderings of examples are equally likely (in effect, meaning that the distribution of examples in the calibration set is indistinguishable from the distribution of unseen examples). Prior work has shown that the conformal procedure is valid under deterministic transformations [24]; by using distinct examples to learn the test-time augmentation policy, the proposed approach constitutes a deterministic transformation applied to both the calibration set and unseen examples. If we were to instead use the same examples to learn the test-time augmentation policy and the conformal threshold, exchangeability could be broken.", + "bbox": [ + 89, + 277, + 482, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experimental Set-Up", + "text_level": 1, + "bbox": [ + 89, + 488, + 292, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets We show results on the test splits of three widely used image classification datasets: ImageNet [12] (50,000 natural images spanning 1,000 classes), iNaturalist [41] (100,000 images spanning 10,000 species), and CUB-Birds [45] (5,794 images spanning 200 categories of birds). Images are distributed evenly over classes in ImageNet and iNaturalist, while CUB-Birds has between 11 and 30 images per class.", + "bbox": [ + 89, + 513, + 482, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Models The default model architecture, across all datasets, is ResNet-50 [17]. The accuracies of the base classifiers are $76.1\\%$ (ImageNet), $76.4\\%$ (iNaturalist), and $80.5\\%$ (CUB-Birds). To study the relationship between model complexity and performance, we also provide results using ResNet-101 and ResNet-152 on ImageNet. For ImageNet, we make use of the pretrained models made available by PyTorch [31]. For iNaturalist, we use a model made public by Niers, Tom [30]. For CUB-Birds, we train a network by finetuning the final layer of a ResNet-50 model initialized with ImageNet's pretrained weights.", + "bbox": [ + 89, + 654, + 482, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Augmentations We consider two augmentation policies. The first (the simple augmentation policy) consists of a random-crop and a horizontal-flip; to produce a random crop, we pad the original image with 4 pixels and take a", + "bbox": [ + 89, + 840, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "256x256 crop of the expanded image (thereby preserving the original image resolution). The simple augmentation policy is widely used because its augmentations are likely to be label-preserving. The second, which we refer to as the expanded augmentation policy, consists of 12 augmentations: increase-sharpness, decrease-sharpness, autocontrast, invert, blur, posterize, shear, translate, color-jitter, randomCrop, horizontal-flip, and random-rotation. The supplement contains a description of each augmentation. These augmentations are not always label preserving, but, as we show, can improve performance when weights are learned.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baselines We benchmark results using two conformal scores (translating to different definitions of $c(x,y)$ in Equation 4). The first score is APS [34] (described in Eqn. 4), which represents the cumulative probability required to include the correct class, and the second is RAPS [1], which modifies APS by adding a term to penalize large set sizes. For all experiments, we do not allow sets of size 0. We implement RAPS and APS using code provided by Angelopoulos et al. [1], and automatically select hyperparameters $k_{reg}$ and $\\lambda$ to minimize set size. We also compare against conformal prediction using a simple average over the test-time augmentations (TTA-Avg). In the supplement, we also compare against non-conformal Top-1 and Top-5 prediction sets.", + "bbox": [ + 511, + 277, + 903, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation We evaluate results using the three metrics commonly used in the conformal prediction literature: efficiency, coverage, and adaptivity. We quantify efficiency using both average prediction set size (measured across all examples) and class-conditional prediction set size (measured across all examples in a class). Coverage is the percentage of sets containing the true label. We define adaptivity as the size-stratified coverage violation (SSCV), introduced by Angelopoulos et al. [1]. We first partition examples based upon the size of the prediction set. We create bins for set sizes of $[0, 1], [2, 3], [4, 10], [11, 100],$ and $[101,]$ . We then compute the empirical coverage within each bin, and compute adaptivity as the maximum difference between theoretical coverage and empirical coverage across bins. The closer this value is to 0, the better the adaptivity.", + "bbox": [ + 511, + 508, + 903, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each dataset, we report results across 10 randomly generated splits into validation and test sets. For all experiments (save for the validation set size experiment), the validation set and test set are the same size. We allot $20\\%$ of examples from the validation set to $D^{(TTA)}$ (used to learn TTA policy), and allot the remaining examples to the calibration set. For the experiment studying validation set size, we downsample the validation set. We compute statistical significance using a paired t-test, with a Bonferroni correction [46] for multiple hypothesis testing. Code to reproduce all experiments will be made publicly available.", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20625", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bfd98d0d138acf841796fb8226ec4f9883ef850f9b3fd70b13cd095411279428.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Expanded Augmentation PolicySimple Augmentation Policy
AlphaMethodImageNetiNaturalistCUB-BirdsImageNetiNaturalistCUB-Birds
0.01RAPS37.751 ± 2.33461.437 ± 6.06715.293 ± 2.07137.751 ± 2.33461.437 ± 6.06715.293 ± 2.071
0.01RAPS+TTA-Avg35.600 ± 2.20057.073 ± 5.91413.111 ± 2.47031.681 ± 3.05754.169 ± 6.31914.550 ± 1.425
0.01RAPS+TTA-Learned31.248 ± 2.17753.195 ± 4.88414.045 ± 1.32332.702 ± 2.40951.391 ± 5.21113.803 ± 1.734
0.05RAPS5.637 ± 0.3577.991 ± 1.5213.624 ± 0.3615.637 ± 0.3577.991 ± 1.5213.624 ± 0.361
0.05RAPS+TTA-Avg5.318 ± 0.1137.067 ± 0.3443.116 ± 0.2104.908 ± 0.0996.451 ± 0.2793.249 ± 0.307
0.05RAPS+TTA-Learned4.889 ± 0.1686.682 ± 0.4473.571 ± 0.5765.040 ± 0.1766.788 ± 0.4963.290 ± 0.186
0.10RAPS2.548 ± 0.0742.914 ± 0.1162.038 ± 0.1532.548 ± 0.0742.914 ± 0.1162.038 ± 0.153
0.10RAPS+TTA-Avg2.470 ± 0.0712.740 ± 0.0261.780 ± 0.1392.327 ± 0.0862.610 ± 0.0311.881 ± 0.118
0.10RAPS+TTA-Learned2.312 ± 0.0542.625 ± 0.0431.893 ± 0.1872.362 ± 0.0652.638 ± 0.0261.840 ± 0.106
", + "bbox": [ + 91, + 89, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Reductions in prediction set size across datasets, augmentation policies, and coverage specifications. Each entry corresponds to the average prediction set size across 10 calibration/test splits. Bolded entries represent performance that is either (a) significantly better compared to the baseline (RAPS), or (b) indistinguishable from the best approach. Table S8 reports achieved coverage. Corresponding results for APS can be found in Table S2.", + "bbox": [ + 89, + 297, + 906, + 354 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9044f56e5d35f56b16f36f9295155b24bb8c1d88a49bd59df101cb6f67adc007.jpg", + "image_caption": [ + "Figure 2. Robustness to distribution shift. We compare average prediction set size achieved by RAPS (yellow) to average prediction set size achieved when combining RAPS with TTA-Learned (green). Results reflect the distribution of average prediction set size across 10 runs using ImageNet and ResNet50. We evaluate performance on different corruptions (x-axis) and different coverage guarantees (left, middle, right). RAPS+TTA-Learned (green) produces a noticeable reduction in prediction set size, even when subject to distribution shift, with no loss in coverage. Refer to Figure S1 in the supplement for a comparison of coverage achieved by both methods." + ], + "image_footnote": [], + "bbox": [ + 101, + 369, + 359, + 503 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cfe7c8bd71deb6b161a7c1096d915b736bea184355749c8b52408f327bd35047.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 369, + 624, + 503 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0df7a187b18fc146b681a1d817e79f7f1999f0fee26d4e746dd068d2ab6e1630.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 371, + 887, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Results", + "text_level": 1, + "bbox": [ + 89, + 626, + 176, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare against RAPS, which outperformed other baselines in every experiment. (We provide results comparing our method to APS and the Top-K baselines in the supplement. Variants of each experiment across multiple $\\alpha$ and datasets are also in the supplement.) We then examine the dependence of these results on dataset, base model, and class. We conclude by providing intuition about why test-time augmentation improves the efficiency of conformal predictors.", + "bbox": [ + 89, + 652, + 482, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Reductions in prediction set size", + "text_level": 1, + "bbox": [ + 89, + 801, + 374, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin with results in the context of the expanded augmentation policy. Learned test-time augmentation policies produce meaningfully significant reductions in prediction set size (RAPS+TTA-Learned in Table 1 and APS+TTA-Learned in Table S1). TTA-Learned reduces prediction set", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sizes significantly in 16 of the 18 cases, and performs comparably in the remaining 2. Across all cases, the combination of RAPS, TTA-Learned, and the expanded augmentation policy produces the smallest average set sizes.", + "bbox": [ + 511, + 627, + 906, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TTA-Learned performs comparably or better than TTA-Avg in all comparisons. Certain augmentations in the expanded augmentation policy (blur, decrease sharpness, and shear) are consistently assigned a weight of 0, while certain augmentations are consistently included in learned policies (autocontrast, translate). Augmentations assigned zero weight provide no additional information about the true label (for example, they may not preserve the label within the image, or they may be redundant with other augmentations included in the policy).", + "bbox": [ + 511, + 688, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While TTA improves both RAPS and APS, the improvements are larger for APS. This is because TTA, like RAPS, tempers the predicted probabilities. TTA lowers the maximum predicted probability on average, thereby reducing", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20626", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/06289058b6d00b03a941ce02b8c9c78abbb4f75abf59b141bb6e10c65a45333f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Expanded Augmentation PolicySimple Augmentation Policy
AlphaMethodResNet-50ResNet-101ResNet-152ResNet-50ResNet-101ResNet-152
0.01RAPS37.751 ± 2.33433.624 ± 1.79629.560 ± 3.48137.751 ± 2.33433.624 ± 1.79629.560 ± 3.481
0.01RAPS+TTA-Avg35.600 ± 2.20030.220 ± 1.77427.203 ± 2.52631.681 ± 3.05727.206 ± 1.84024.106 ± 2.100
0.01RAPS+TTA-Learned31.248 ± 2.17725.722 ± 1.71323.615 ± 1.65632.702 ± 2.40926.760 ± 1.97424.765 ± 2.736
0.05RAPS5.637 ± 0.3574.785 ± 0.1024.376 ± 0.0785.637 ± 0.3574.785 ± 0.1024.376 ± 0.078
0.05RAPS+TTA-Avg5.318 ± 0.1134.433 ± 0.1374.163 ± 0.1854.908 ± 0.0994.147 ± 0.1223.868 ± 0.126
0.05RAPS+TTA-Learned4.889 ± 0.1684.200 ± 0.2003.824 ± 0.1285.040 ± 0.1764.194 ± 0.1943.916 ± 0.356
0.10RAPS2.548 ± 0.0742.267 ± 0.0242.109 ± 0.0272.548 ± 0.0742.267 ± 0.0242.109 ± 0.027
0.10RAPS+TTA-Avg2.470 ± 0.0712.164 ± 0.0312.049 ± 0.0282.327 ± 0.0862.093 ± 0.0351.996 ± 0.018
0.10RAPS+TTA-Learned2.312 ± 0.0542.099 ± 0.0401.993 ± 0.0262.362 ± 0.0652.091 ± 0.0411.988 ± 0.020
", + "bbox": [ + 91, + 89, + 903, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Reductions in prediction set size across base classifiers on ImageNet. TTA-Learned can bridge the performance gap between different classifiers (for example, outperforming ResNet-152 alone when combined with ResNet-101), and yields significant reductions in set size regardless of the pretrained classifier used. We report achieved coverage in Table S9.", + "bbox": [ + 89, + 304, + 906, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model overconfidence. Consequently, the predicted probability assigned to the remaining classes is higher. This is why the expanded augmentation policy demonstrates stronger performance than the simple augmentation policy: it tempers the probabilities to a greater extent.", + "bbox": [ + 89, + 372, + 480, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TTA-Learned preserves coverage across all experiments, since it respects the assumption of exchangeability. In some cases, TTA significantly improves coverage, although the magnitude of this difference is small (results can be found in Tables S8 and S9).", + "bbox": [ + 89, + 455, + 482, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We next evaluate adaptivity using size stratified coverage violation (SSCV). At low alpha ( $\\alpha = .01$ , and $\\alpha = .05$ ), TTA-Learned improves efficiency without diminishing adaptivity (Table S10).", + "bbox": [ + 89, + 537, + 482, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Robustness to distribution shift", + "text_level": 1, + "bbox": [ + 89, + 628, + 366, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we evaluate the performance of test-time augmented conformal prediction on out-of-distribution examples. While conformal prediction does not guarantee coverage in these settings, distribution shifts are ubiquitous in practice [22]. Empirical performance is thus of practical interest. The training procedures for both test-time augmented conformal prediction and conformal prediction remain the same and use in-distribution examples from ImageNet and a ResNet50 classifier. We evaluate each conformal predictor on four types of image corruptions drawn from ImageNet-C [19]. Figure 2 plots the results; across all corruptions and coverage guarantees, test-time augmented conformal prediction produces smaller prediction sets than conformal prediction alone. Importantly, test-time augmented conformal prediction achieves this with no loss of coverage (Figure S1).", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Datasets, augmentation policies, and models", + "text_level": 1, + "bbox": [ + 511, + 372, + 885, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dependence on dataset TTA consistently improves prediction set sizes on ImageNet and iNaturalist, but not on CUB-Birds. This may be because the validation set size for CUB-Birds (2,827 images) is an order of magnitude smaller than the validation sets for ImageNet (25,000 images) and iNaturalist (50,000 images). This is consistent with our finding that effectiveness of TTA is positively correlated with the size of the validation set (Figure S4).", + "bbox": [ + 511, + 393, + 905, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dependence on augmentation policy We find that the expanded augmentation policy produces larger reductions in set size than the simple augmentation policy, primarily at low $\\alpha$ . This is in spite of the fact that the expanded augmentation policy contains many augmentations outside of the base classifier's training-time augmentation policy. When we vary the number of augmentations included in an augmentation policy, we see that larger augmentation policies also yield greater reductions in average prediction set size (Figure S2). That said, the simple augmentation policy does have its place; it requires fewer forward passes during inference. In the absence of a learned aggregation function, our results suggest that aggregating using an average can still improve the efficiency of conformal predictors (outperforming the original conformal score in 11 comparisons, matching performance in 3, and under-performing in 4).", + "bbox": [ + 511, + 534, + 906, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dependence on base model We tested the generalizability of our results to other models by rerunning the ImageNet experiments using ResNet-101 (accuracy of $77.4\\%$ ) and ResNet-152 (accuracy of $78.3\\%$ ). Unsurprisingly, more accurate models result in smaller prediction set sizes (Table 2). TTA variants of conformal prediction again produce significant improvements in set size while maintaining coverage.", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20627", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36e53bae95142b1447e11f48bc24ccc9f5bc24a5187ab1f95f57952547fffab4.jpg", + "image_caption": [ + "A", + "Distribution of Set Sizes (ImageNet, $\\alpha = 0.01$ Expanded)", + "Figure 3. (A) Class-conditional prediction set sizes. We plot the distribution of class-conditional prediction set sizes, for ImageNet and ResNet-50 with $\\alpha = .01$ . RAPS+TTA-Learned (green) produces a noticeable reduction in class-conditional prediction set sizes. (B, C) Relationship between TTA improvements and original class set sizes and class difficulty. TTA introduces the largest improvements for classes with the largest original prediction set sizes (B) and classes on which the underlying classifier is often incorrect (C). Each point represents the average prediction set size for each class, across 10 splits." + ], + "image_footnote": [], + "bbox": [ + 98, + 119, + 367, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e43903664fe6477547f726dab51f0a6f4191f8e4d21eeead62eb480eb0d220a3.jpg", + "image_caption": [ + "B", + "Original Set Size $(r = 0.89, \\alpha = 0.01$ , Expanded)" + ], + "image_footnote": [], + "bbox": [ + 380, + 119, + 629, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b69bdf3185435d4650bdd511a9d3e20de883f89c8593e692ae4801f00b89eac.jpg", + "image_caption": [ + "C", + "Class Difficulty 55. $\\alpha = 0.01$ Exp" + ], + "image_footnote": [], + "bbox": [ + 640, + 119, + 888, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/951c946d5ab97999c62824298e9d768559c84a83b72391c06eef90a190b2ecd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 354, + 413, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/980d366e395e611caa4ee3f02f3450fe3432fd5bae5a4ea26f9f71c4a2ac276f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 354, + 650, + 534 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4d64b53b7a22eb991b69b832da77cc5c13578a9fae7f5cad97978d0a55960ee.jpg", + "image_caption": [ + "Figure 4. (A) Effect of TTA-Learned on optimal Top-K: TTA-Learned significantly lowers the value of k required for Top-k prediction sets to achieve coverage on ImageNet and iNaturalist, but not on CUB-Birds. (B,C) Effect of TTA-Learned on rank of true class: TTA-Learned improves the rank of the true class among the sorted predicted probabilities for a given example for both ImageNet (B) and iNaturalist (C). We plot the rank using the original predicted probabilities compared to the TTA-transformed probabilities, binning all examples in the validation set into five equal-width bins. Dots that fall below the red line indicate that TTA improves the rank of the true class." + ], + "image_footnote": [], + "bbox": [ + 651, + 354, + 893, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "age. We note that the combination of TTA with ResNet-101 produces smaller set sizes than the more complex ResNet-152 alone. For example, when $\\alpha$ is set to .01, RAPS+TTA-Learned and ResNet-101 produce set sizes that contain, on average, 26.5 classes, while RAPS and ResNet-152 produce an average set size of 29.6.", + "bbox": [ + 88, + 674, + 482, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Class-Specific Analysis", + "text_level": 1, + "bbox": [ + 89, + 772, + 303, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have established that on average TTA is a useful addition to the conformal pipeline. We now investigate the source of this improvement. We make two empirical observations. First, classes with larger predicted set sizes benefit most from the introduction of TTA. Figure 3 shows that a class's average prediction set size is significantly correlated with the change in set size produced by TTA-Learned (with", + "bbox": [ + 88, + 794, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the expanded augmentation policy and $\\alpha = .01$ , $r = 0.89$ , and $p < 1e - 10$ . Second, we find that class difficulty is significantly associated with changes in set size introduced by TTA (with the expanded augmentation policy and $\\alpha = .01$ , $r = 0.55$ and $p < 1e - 10$ ). Prediction sets for classes that are difficult to predict benefit more from TTA compared to their easier counterparts. These observations are related; harder classes receive larger set sizes, and consequently, offer larger room for improvements in efficiency.", + "bbox": [ + 511, + 674, + 906, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Intuition", + "text_level": 1, + "bbox": [ + 511, + 818, + 617, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Why does the addition of test-time augmentation produce smaller prediction set sizes? In short, TTA improves top-K accuracy. We verify this claim by estimating $k$ such that the uncertainty sets comprised of the top $k$ predicted", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20628", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "classes for each example achieve a marginal coverage of $(1 - \\alpha)$ . We see that for those datasets (ImageNet and iNaturalist) where TTA produces significant reductions in set size, TTA-transformed predictions—both with a simple average and learned weights—produce significantly lower values for $k$ compared to the original predictions (Figure 4A). This is not true for CUB-Birds, on which TTA offers little to no improvement. One could use such a procedure to determine whether TTA is worth adding to a conformal pipeline without collecting labeled data beyond the calibration dataset.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another way to understand the impact of TTA is to consider the effect on the ordering of classes. It has been observed in the test-time augmentation literature that TTA often promotes the true class from the second-highest to the highest predicted probability, thereby correcting the classification. This finding does not fully explain the value of test-time augmentation to conformal prediction since only $3\\%$ of the prediction sets which reduce in size are associated with a corrected Top-1 classification. What TTA did do was increase the predicted probability of the true class even when it is predicted to be unlikely (for example, promoting the true class from 200th most likely to 100th most likely). We visualize this effect in Figure 4 by plotting the change in true class rank (the index at which the true class appears in the sorted list of predicted probabilities) for all examples in the validation set, stratified into 5 equal-width bins. The lower left point captures examples that are classified correctly; here, test-time augmentation introduces little to no change. In subsequent bins, we see that TTA typically promotes the rank of the true class. We also include the standard deviation across the true class ranks in the original predicted probabilities (x-axis) and the TTA-transformed probabilities (y-axis).", + "bbox": [ + 91, + 255, + 483, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitations", + "text_level": 1, + "bbox": [ + 89, + 651, + 212, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learned test-time augmentation policies require two ingredients: labeled data and multiple forward passes. Although one can minimize costs by parallelizing computation or by using the learned weights to identify which augmentations to generate, inference will always cost more with test-time augmentation. Our results are also limited to image classification. We do not consider other modalities, for which appropriate transformations will substantially differ. Future work should consider how these results may generalize to non-vision tasks. Finally, test-time augmentation is one approach to generating ensembles in conformal prediction. Many other more computationally expensive approaches exist. The tradeoff between computation and ensemble performance remains a useful avenue for future work.", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 89, + 633, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show that test-time augmented conformal prediction produces smaller sets than conformal prediction alone. The proposed approach is effective, efficient, and simple: it reduces prediction set sizes by up to $30\\%$ , requires no model re-training, and relies on a portion of labeled data already available to split conformal predictors. Our experiments also indicate that test-time augmented conformal prediction exhibits greater efficiency under four common corruption-based distribution shifts. Test-time augmentation is able to do this by improving the underlying classifier's robustness to domain-specific invariances, in the form of data augmentation. Efforts to improve the efficiency of conformal predictors could, as a first step, aim to improve the robustness of the underlying classifier. The performance of TTA-Learned also suggests that there are settings in which it is wise to use a portion of the labeled data to improve the underlying model, instead of reserving all labeled data for the calibration set. In sum, our work takes a step towards practically useful conformal predictors by improving efficiency, without sacrificing adaptivity or coverage.", + "bbox": [ + 511, + 114, + 906, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 431, + 609, + 448 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anastasios Angelopoulos, Stephen Bates, Jitendra Malik, and Michael I. Jordan. Uncertainty Sets for Image Classifiers using Conformal Prediction, 2022. arXiv:2009.14193 [cs, math, stat]. 2, 4, 11", + "[2] Murat Seçkin Ayhan and Philipp Berens. Test-time Data Augmentation for Estimation of Heteroscedastic Aleatoric Uncertainty in Deep Neural Networks. 2018. 1, 2", + "[3] Varun Babbar, Umang Bhatt, and Adrian Weller. On the Utility of Prediction Sets in Human-AI Teams, 2022. arXiv:2205.01411 [cs]. 1", + "[4] Anthony Bellotti. Optimized conformal classification using gradient descent approximation, 2021. arXiv:2105.11255 [cs]. 2", + "[5] Lars Carlsson, Martin Eklund, and Ulf Norinder. Aggregated Conformal Prediction. In Artificial Intelligence Applications and Innovations, pages 231-240, Berlin, Heidelberg, 2014. Springer. 2", + "[6] Sewhan Chun, Jae Young Lee, and Junmo Kim. Cyclic test time augmentation with entropy weight method. In Proceedings of the Thirty-Eighth Conference on Uncertainty in Artificial Intelligence, pages 433-442. PMLR, 2022. ISSN: 2640-3498. 2", + "[7] Gilad Cohen and Raja Giryes. Simple post-training robustness using test time augmentations and random forest. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3996-4006, 2024. 1", + "[8] Seffi Cohen, Noa Dagan, Nurit Cohen-Inger, Dan Ofer, and Lior Rokach. ICU Survival Prediction Incorporating Test-Time Augmentation to Improve the Accuracy of Ensemble-Based Models. IEEE Access, 9:91584-91592, 2021. Conference Name: IEEE Access. 2" + ], + "bbox": [ + 521, + 457, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20629", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Pedro Conde, Tiago Barros, Rui L. Lopes, Cristiano Premebida, and Urbano J. Nunes. Approaching Test Time Augmentation in the Context of Uncertainty Calibration for Deep Neural Networks, 2023. arXiv:2304.05104 [cs]. 2", + "[10] Jesse C Cresswell, Yi Sui, Bhargava Kumar, and Noel Vouitsis. Conformal prediction sets improve human decision making. arXiv preprint arXiv:2401.13744, 2024. 1", + "[11] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. AutoAugment: Learning Augmentation Policies from Data, 2019. arXiv:1805.09501 [cs, stat]. 11", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. ISSN: 1063-6919. 4", + "[13] Tiffany Ding, Anastasios N. Angelopoulos, Stephen Bates, Michael I. Jordan, and Ryan J. Tibshirani. Class-Conditional Conformal Prediction With Many Classes, 2023. arXiv:2306.09335 [cs, stat]. 2", + "[14] Bat-Sheva Einbinder, Yaniv Romano, Matteo Sesia, and Yanfei Zhou. Training Uncertainty-Aware Classifiers with Conformalized Deep Learning. 15", + "[15] Shohei Enomoto, Monikka Roslianna Busto, and Takeharu Eda. Dynamic Test-Time Augmentation via Differentiable Functions, 2023. arXiv:2212.04681 [cs]. 2", + "[16] A. Gammerman, V. Vovk, and V. Vapnik. Learning by transduction. In Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence, pages 148-155, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers Inc. 2", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, Las Vegas, NV, USA, 2016. IEEE. 4", + "[18] Achim Hekler, Titus J. Brinker, and Florian Buettner. Test Time Augmentation Meets Post-hoc Calibration: Uncertainty Quantification under Real-World Conditions. Proceedings of the AAAI Conference on Artificial Intelligence, 37(12):14856-14864, 2023. Number: 12. 1, 2", + "[19] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 6", + "[20] Vilde Jensen, Filippo Maria Bianchi, and Stian Normann Anfinsen. Ensemble Conformalized Quantile Regression for Probabilistic Time Series Forecasting. IEEE Transactions on Neural Networks and Learning Systems, pages 1-12, 2022. Conference Name: IEEE Transactions on Neural Networks and Learning Systems. 2", + "[21] Ildoo Kim, Younghoon Kim, and Sungwoong Kim. Learning Loss for Test-Time Augmentation. In Advances in Neural Information Processing Systems, pages 4163-4174. Curran Associates, Inc., 2020. 2", + "[22] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 6", + "[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 11", + "[24] Arun Kumar Kuchibhotla. Exchangeability, conformal prediction, and rank tests. arXiv preprint arXiv:2005.06095, 2020. 4", + "[25] Henrik Linusson, Ulf Norinder, Henrik Bostrom, Ulf Johansson, and Tuve Lofstrom. On the Calibration of Aggregated Conformal Predictors. In Proceedings of the Sixth Workshop on Conformal and Probabilistic Prediction and Applications, pages 154-173. PMLR, 2017. ISSN: 2640-3498. 2", + "[26] H. Linusson, U. Johansson, and H. Bostrom. Efficient conformal predictor ensembles. Neurocomputing, 397:266-278, 2020. 2", + "[27] Helen Lu, Divya Shanmugam, Harini Suresh, and John Guttag. Improved Text Classification via Test-Time Augmentation, 2022. arXiv:2206.13607 [cs]. 2", + "[28] Alexander Lyzhov, Yuliya Molchanova, Arsenii Ashukha, Dmitry Molchanov, and Dmitry Vetrov. Greedy Policy Search: A Simple Baseline for Learnable Test-Time Augmentation. In Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI), pages 1308-1317. PMLR, 2020. ISSN: 2640-3498. 2", + "[29] Luca Mocerino, Roberto G. Rizzo, Valentino Peluso, Andrea Calimera, and Enrico Macii. Adaptive Test-Time Augmentation for Low-Power CPU, 2021. arXiv:2105.06183 [cs, eess]. 2", + "[30] Niers, Tom. iNaturalist_competition, 2021. original-date: 2021-12-10T10:56:46Z. 4", + "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 4", + "[32] Juan C. Perez, Motasem Alfarra, Guillaume Jeanneret, Laura Rueda, Ali Thabet, Bernard Ghanem, and Pablo Arbelaez. Enhancing Adversarial Robustness via Test-time Transformation Ensembling. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 81-91, Montreal, BC, Canada, 2021. IEEE. 1, 2", + "[33] Drew Prinster, Anqi Liu, and Suchi Saria. JAWS: Auditing Predictive Uncertainty Under Covariate Shift, 2022. arXiv:2207.10716 [cs, stat]. 2", + "[34] Yaniv Romano, Matteo Sesia, and Emmanuel J. Candès. Classification with Valid and Adaptive Coverage, 2020. arXiv:2006.02544 [stat]. 2, 3, 4", + "[35] C Saunders and Royal Holloway. Transduction with Confidence and Credibility. 1999. 2", + "[36] Glenn Shafer and Vladimir Vovk. A Tutorial on Conformal Prediction. 2008. 1, 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20630", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better Aggregation in Test-Time Augmentation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1194–1203, Montreal, QC, Canada, 2021. IEEE. 1, 2", + "[38] David Stutz, Krishnamurthy, Dvijotham, Ali Taylan Cemgil, and Arnaud Doucet. Learning Optimal Conformal Classifiers, 2022. arXiv:2110.09192 [cs, stat]. 2", + "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1", + "[40] Ryan J Tibshirani, Rina Foygel Barber, Emmanuel Candes, and Aaditya Ramdas. Conformal Prediction Under Covariate Shift. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2", + "[41] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha. Benchmarking Representation Learning for Natural World Image Collections, 2021. arXiv:2103.16483 [cs]. 4", + "[42] Vladimir Vovk. Algorithmic Learning in a Random World. Springer-Verlag, New York, 2005. 2", + "[43] Vladimir Vovk. Cross-conformal predictors, 2012. arXiv:1208.0806 [cs, stat]. 2", + "[44] Vladimir Vovk. Cross-conformal predictors. Annals of Mathematics and Artificial Intelligence, 74(1-2):9-28, 2015. 2", + "[45] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 4", + "[46] Eric W. Weisstein. Bonferroni correction. https://mathworld.wolfram.com/, 2004. 4", + "[47] Dongping Zhang, Angelos Chatzimparmpas, Negar Kamali, and Jessica Hullman. Evaluating the utility of conformal prediction sets for ai-advised image labeling. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1-19, 2024. 1", + "[48] Marvin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test Time Robustness via Adaptation and Augmentation, 2022. arXiv:2110.09506 [cs]. 2" + ], + "bbox": [ + 91, + 92, + 483, + 659 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20631", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_model.json b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..613152105951e30271952f533dd585892781545f --- /dev/null +++ b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.131, + 0.842, + 0.154 + ], + "angle": 0, + "content": "Test-time augmentation improves efficiency in conformal prediction" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.181, + 0.777, + 0.217 + ], + "angle": 0, + "content": "Divya Shanmugam, Helen Lu, Swami Sankaranarayanan, John Guttag Massachusetts Institute of Technology, CSAIL" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.219, + 0.686, + 0.234 + ], + "angle": 0, + "content": "{divyas, helenlu, swamiviv, guttag}@mit.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.329, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.301, + 0.486, + 0.543 + ], + "angle": 0, + "content": "A conformal classifier produces a set of predicted classes and provides a probabilistic guarantee that the set includes the true class. Unfortunately, it is often the case that conformal classifiers produce uninformatively large sets. In this work, we show that test-time augmentation (TTA)-a technique that introduces inductive biases during inference-reduces the size of the sets produced by conformal classifiers. Our approach is flexible, computationally efficient, and effective. It can be combined with any conformal score, requires no model retraining, and reduces prediction set sizes by \\(10\\% - 14\\%\\) on average. We conduct an evaluation of the approach spanning three datasets, three models, two established conformal scoring methods, different guarantee strengths, and several distribution shifts to show when and why test-time augmentation is a useful addition to the conformal pipeline." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.572, + 0.222, + 0.589 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Conformal prediction has emerged as a promising way to equip existing classifiers with statistically valid uncertainty estimates. It does so by replacing the prediction of the most likely class with an uncertainty set-a set of classes accompanied by a probabilistic guarantee that the true class appears in the set [36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Conformal prediction faces two limitations in practice. First, achieving a suitably strong guarantee often leads to prediction sets that are uninformatively large [10, 47]. Large prediction sets have been shown to decrease performance when provided as a decision-making aid [3, 10, 47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Second, conformal classifiers inherit the instability of the underlying models. As a result, prediction sets can change significantly in response to small input perturbations, a well-known weakness of neural networks [39]. Applying a horizontal flip to each image in ImageNet, for example, changes the prediction set sizes for \\(75\\%\\) of examples at a coverage guarantee of \\(99\\%\\). Such behavior also represents a barrier to broader use." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we show that test-time augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.406 + ], + "angle": 0, + "content": "(TTA), a widely-used technique in computer vision, has the potential to address both limitations. TTA involves generating an ensemble of predictions by perturbing the input with label-preserving transformations. It has previously been shown that TTA can be used to make non-conformal classifiers more robust to small input perturbations [7], more accurate [37], and better calibrated [18]. However, previous work has not explored the utility of TTA in the context of conformal prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.909, + 0.517 + ], + "angle": 0, + "content": "We propose test-time-augmented conformal prediction, which transforms a classifier's predictions using a learned test-time augmentation policy prior to conformal classification. By using distinct sets of labeled data to learn the test-time augmentation policy and the conformal classifier, we preserve the assumption of exchangeability, and thereby the coverage guarantee associated with the conformal predictor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.702 + ], + "angle": 0, + "content": "In experiments testing the performance of conformal predictors subject to distribution shift, we see that test-time augmentation reduces prediction set sizes by \\(14\\%\\) on average, with no loss of coverage. And even when there is no distribution shift, we see a reduction of \\(10\\%\\) on average. Moreover, we find that classes with the largest average prediction set sizes benefit most from the introduction of test-time augmentation. We also show that test-time augmentation can bridge gaps between classifiers of different sizes (e.g., we show that test-time augmentation combined with ResNet-50 produces smaller set sizes than ResNet-101 without test-time augmentation)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Our analysis of why test-time augmentation reduces prediction set sizes reveals a previously unknown effect of test-time augmentation. Specifically, TTA increases the predicted probability of the true class even when it is predicted to be unlikely (for example, promoting the true class from 200th most likely to 100th most likely). Although such behavior has no impact on the maximum predicted probability — commonly the focus of literature on test-time augmentation [2, 32, 37] — it is valuable in conformal prediction. This is because when the true class is promoted to a higher rank among a classifier's predicted probabilities, the conformal classifier includes fewer incorrect classes to meet the conformal guarantee." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20622" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.099, + 0.387, + 0.112 + ], + "angle": 0, + "content": "Test-time augmented conformal prediction..." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.121, + 0.638, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.099, + 0.89, + 0.112 + ], + "angle": 0, + "content": "...leads to smaller prediction set sizes." + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.118, + 0.893, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.324, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 1. We illustrate the addition of test-time augmentation to conformal calibration in green (left) and provide a snapshot of the improvements it can confer (right). We show results on Imagenet, with a desired coverage of \\(95\\%\\), for the 20 classes with the largest predicted set sizes on average (computed over 10 calibration/test splits)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.543 + ], + "angle": 0, + "content": "Contributions We make three contributions. To begin, this is the first work to propose combining test-time augmentation, a popular technique in computer vision, with conformal prediction. Second, we present a method that reduces the prediction set sizes of existing conformal predictors by using automatically learned test-time augmentations. Finally, we demonstrate, in an extensive set of experiments, that our approach to combining conformal prediction and test-time augmentation leads to smaller prediction sets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.228, + 0.576 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.484, + 0.663 + ], + "angle": 0, + "content": "Conformal prediction was first introduced by Gammerman et al. [16], and further developed by Saunders and Holloway [35] and Vladimir Vovk [42]. We review efforts to ensemble conformal predictors and efforts to reduce prediction set sizes below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Ensembles in conformal prediction Several methods that generate ensembles of conformal predictors are known to improve efficiency. These methods include cross-conformal prediction [43], bootstrap conformal prediction [44], aggregated conformal prediction [5, 25], and out-of-bag conformal prediction [26]. The approaches primarily differ in how data is sampled to create the training dataset for the classifier and the calibration dataset for the conformal predictor. However, all require training multiple base classifiers or conformal predictors. Our approach is distinct: we propose a technique to generate an ensemble from a single model by perturbing the input, which requires no additional base models and no additional conformal predictors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Efficiency in conformal prediction There are two ways to improve efficiency in split conformal prediction: adjustments to the conformal score or improvements to the underlying classifier. Many works have proposed new procedures to estimate and apply thresholds on conformal scores [1, 4, 13, 33, 40]. Romano et al. [34] proposed APS, a conformal score based on the cumulative probability required to include the correct class in a prediction set. Angelopoulos et al. [1] built on this work to propose RAPS, which modifies APS by penalizing the inclusion of low-probability classes. Comparatively little work has focused on improvements to the underlying model. Jensen et al. [20] ensemble a set of base classifiers, where the classifiers are created by training models on subsets of the training data. Stutz et al. [38] provide a new way to train the base classifier and conformal predictor jointly through a novel conformal training loss. In contrast, our work focuses on improving the underlying model without retraining, and can be easily combined with any of the above procedures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Test-Time Augmentation Test-time augmentation (TTA) is a popular technique to improve the accuracy, robustness, and calibration of an existing classifier by aggregating predictions over a set of input transformations [2, 9, 15, 18, 32, 37, 48]. TTA has been applied to a diverse range of predictive tasks across domains ranging from healthcare [8] to content moderation [27]. Consequently, many have proposed new ways to perform TTA—for example, learning when to apply TTA [29], which augmentations to use [6, 21, 28], and how to aggregate the resulting predictions [6, 9, 37]. Existing work typically focuses on test-time augmentation's impact on highest predicted probability. Here, we analyze how test-time augmentation increases the pre" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20623" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "dicted probability assigned to the true class when it appears outside the top few classes, and how that change is consequential in conformal prediction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.152, + 0.25, + 0.169 + ], + "angle": 0, + "content": "3. Problem setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.238 + ], + "angle": 0, + "content": "We operate within the split conformal prediction framework. In this setting, a conformal classifier \\(\\mathcal{C}(x_i) \\subset \\{1, \\ldots, K\\}\\) maps input \\(x_i\\) to a subset of \\(K\\) possible classes and requires three inputs:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.24, + 0.483, + 0.27 + ], + "angle": 0, + "content": "- Calibration set \\( D^{(cal)} = \\{(x_1, y_1), \\ldots, (x_n, y_n)\\} \\), containing \\( n \\) labeled examples." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.271, + 0.483, + 0.3 + ], + "angle": 0, + "content": "- Classifier \\( f: \\mathcal{X} \\mapsto \\Delta^K \\), mapping input domain \\( \\mathcal{X} \\) to a probability distribution over \\( K \\) classes." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.301, + 0.483, + 0.332 + ], + "angle": 0, + "content": "- Desired upper bound on error rate \\(\\alpha \\in [0,1]\\), where \\((1 - \\alpha)\\) is the probability the set contains the true class." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.24, + 0.483, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.365 + ], + "angle": 0, + "content": "We study the introduction of two variables drawn from the test-time augmentation literature:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.368, + 0.483, + 0.412 + ], + "angle": 0, + "content": "- Augmentation policy \\(\\mathcal{A} = \\{a_0, \\dots, a_{m-1}\\}\\), consisting of \\(m\\) augmentation functions, where \\(a_0\\) is the identity transform." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.413, + 0.483, + 0.444 + ], + "angle": 0, + "content": "- Aggregation function \\(\\hat{g}\\), which aggregates a set of predictions to produce one prediction." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.368, + 0.483, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.493 + ], + "angle": 0, + "content": "Each variable translates to a key choice in test-time augmentation: what augmentations to apply \\((\\mathcal{A})\\) and how to aggregate the resulting probabilities \\((\\hat{g})\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.507, + 0.201, + 0.524 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.638 + ], + "angle": 0, + "content": "Our goal is to learn an aggregation function \\(\\hat{g}\\) to maximize the accuracy of the underlying classifier, and ultimately reduce the sizes of the prediction sets generated from the classifier's predicted probabilities. We briefly outline the conformal approach, and then detail the mechanics of our method (illustrated in Figure 1). For a detailed introduction to conformal prediction, refer to Shafer and Vovk [36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.484, + 0.699 + ], + "angle": 0, + "content": "Conformal predictors accept three inputs: a probabilistic classifier \\( f \\), a calibration set \\( \\mathcal{D}^{(cal)} \\), and a pre-specified error rate \\( \\alpha \\). Using the these inputs, one can construct a conformal predictor in three steps:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.702, + 0.483, + 0.747 + ], + "angle": 0, + "content": "1. Define a score function \\( c(x_i, y_i) \\), which produces a conformal score representing the uncertainty of the input example and label pair." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.747, + 0.483, + 0.777 + ], + "angle": 0, + "content": "2. Produce a distribution of conformal scores by computing \\( c(x_i, y_i) \\) for all \\( (x_i, y_i) \\in \\mathcal{D}^{(cal)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.777, + 0.483, + 0.823 + ], + "angle": 0, + "content": "3. Compute threshold \\(\\hat{q}\\) as the \\(\\lceil (n + 1)(1 - \\alpha)\\rceil /n\\) quantile of the distribution of conformal scores over \\(n\\) examples in the calibration set, combined with \\(\\{\\infty \\}\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.702, + 0.483, + 0.823 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "For a new example \\( x \\), we compute \\( c(x_i, y) \\) for all \\( y \\in \\{1, \\dots, K\\} \\), and include all \\( y \\) for which \\( c(x_i, y) < \\hat{q} \\). We adopt the conformal score proposed by Romano et al. [34], which equates to the cumulative probability required to include the correct class:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.119, + 0.907, + 0.137 + ], + "angle": 0, + "content": "\\[\n\\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) = \\hat {p} \\left(y = y ^ {\\prime} \\mid x\\right) = f (x) _ {y ^ {\\prime}} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.14, + 0.905, + 0.182 + ], + "angle": 0, + "content": "\\[\n\\rho_ {x} (y) = \\sum_ {y ^ {\\prime} = 1} ^ {K} \\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) \\mathbb {I} \\left[ \\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) > \\hat {\\pi} _ {x} (y) \\right] \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.187, + 0.907, + 0.203 + ], + "angle": 0, + "content": "\\[\nc (x, y) = \\rho_ {x} (y) + u \\cdot \\hat {\\pi} _ {x} (y) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.907, + 0.323 + ], + "angle": 0, + "content": "where \\(\\rho_{x}(y)\\) is the cumulative probability of all classes predicted with higher probability than \\(y\\) and \\(\\hat{\\pi}_x(y')\\) corresponds to the predicted probability of class \\(y^\\prime\\) given \\(x\\). The variable \\(u\\sim U(0,1)\\). Conformal score \\(c(x_i,y_i)\\) is thus composed of this cumulative probability and the predicted probability of class \\(y_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.906, + 0.419 + ], + "angle": 0, + "content": "**Proposal** Our approach differs from prior work in that the conformal score is derived by transforming the probabilities output by \\( f \\) using test-time augmentation. Concretely, this replaces Equation 1 with the following, parametrized by augmentation policy \\( \\mathcal{A} \\) and augmentation weights \\( \\theta \\)." + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.446, + 0.907, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) = \\hat {p} \\left(y = y ^ {\\prime} \\mid x _ {i}\\right) = g \\left(x _ {i}; f, \\mathcal {A}, \\theta\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.612 + ], + "angle": 0, + "content": "Aggregation weights \\(\\theta\\) are applied to the logits output by classifier \\(f\\), and transformed to be proper probabilities by applying a softmax function. We learn the aggregation weights \\(\\theta\\) using a portion of the validation set, \\(D^{(TTA)}\\), distinct from calibration set used to identify the conformal threshold \\((D^{(cal)})\\). In contrast to traditional approaches, where all labeled data is used to estimate the conformal threshold, we instead reserve a portion to learn the test-time augmentation policy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.675 + ], + "angle": 0, + "content": "We learn a set of weights which maximize classification accuracy on \\( D^{(TTA)} \\) by minimizing the cross-entropy loss1 computed between the predicted probabilities and true labels. More formally, \\( g \\) applies \\( \\theta \\) and \\( \\mathcal{A} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.701, + 0.907, + 0.719 + ], + "angle": 0, + "content": "\\[\ng \\left(x _ {i}; f, \\mathcal {A}, \\Theta\\right) = \\sigma \\left(\\theta^ {T} \\mathbf {A} \\left(f, \\mathcal {A}, x _ {i}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.907, + 0.853 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}\\) uses \\(f\\) to map input \\(x_{i}\\) to a \\(M\\times K\\) matrix of predicted logits where \\(M\\) is the number of augmentations and \\(K\\) is the number of classes. \\(\\theta\\) is a \\(1\\times m\\) vector corresponding to augmentation-specific weights. Each row in \\(\\mathbf{A}(f,\\mathcal{A},x_i)\\) represents the pre-trained classifier's predicted logits on augmentation \\(a_{m}\\) of \\(x_{i}\\). TTA-Learned refers to TTA combined with learned augmentation weights, while TTA-Avg refers to a simple average over the augmentations." + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.863, + 0.906, + 0.901 + ], + "angle": 0, + "content": "1We found no significant improvement by using alternate losses considered in the conformal prediction literature (e.g. focal or conformal training loss). See Table S7 in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20624" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "We refer to the fraction of the validation set allotted to \\( D^{(TTA)} \\) as \\( \\beta \\). Figure S3 shows that performance is not sensitive to the choice of \\( \\beta \\); as a result, all experiments use \\( \\beta = .2 \\) (see supplement for further discussion). This does reduce the amount of data available to identify the appropriate threshold, but we find that the benefits TTA confers outweigh the cost to threshold estimation. Computational cost scales linearly with the size of \\( \\mathcal{A} \\); each additional augmentation translates to a forward pass of the base classifier. One can use the learned weights to save computation by identifying which test-time augmentations to generate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.278, + 0.483, + 0.475 + ], + "angle": 0, + "content": "Preserving exchangeability The validity of conformal prediction depends upon the assumption of exchangeability: that all orderings of examples are equally likely (in effect, meaning that the distribution of examples in the calibration set is indistinguishable from the distribution of unseen examples). Prior work has shown that the conformal procedure is valid under deterministic transformations [24]; by using distinct examples to learn the test-time augmentation policy, the proposed approach constitutes a deterministic transformation applied to both the calibration set and unseen examples. If we were to instead use the same examples to learn the test-time augmentation policy and the conformal threshold, exchangeability could be broken." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.489, + 0.293, + 0.507 + ], + "angle": 0, + "content": "5. Experimental Set-Up" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Datasets We show results on the test splits of three widely used image classification datasets: ImageNet [12] (50,000 natural images spanning 1,000 classes), iNaturalist [41] (100,000 images spanning 10,000 species), and CUB-Birds [45] (5,794 images spanning 200 categories of birds). Images are distributed evenly over classes in ImageNet and iNaturalist, while CUB-Birds has between 11 and 30 images per class." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Models The default model architecture, across all datasets, is ResNet-50 [17]. The accuracies of the base classifiers are \\(76.1\\%\\) (ImageNet), \\(76.4\\%\\) (iNaturalist), and \\(80.5\\%\\) (CUB-Birds). To study the relationship between model complexity and performance, we also provide results using ResNet-101 and ResNet-152 on ImageNet. For ImageNet, we make use of the pretrained models made available by PyTorch [31]. For iNaturalist, we use a model made public by Niers, Tom [30]. For CUB-Birds, we train a network by finetuning the final layer of a ResNet-50 model initialized with ImageNet's pretrained weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Augmentations We consider two augmentation policies. The first (the simple augmentation policy) consists of a random-crop and a horizontal-flip; to produce a random crop, we pad the original image with 4 pixels and take a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "256x256 crop of the expanded image (thereby preserving the original image resolution). The simple augmentation policy is widely used because its augmentations are likely to be label-preserving. The second, which we refer to as the expanded augmentation policy, consists of 12 augmentations: increase-sharpness, decrease-sharpness, autocontrast, invert, blur, posterize, shear, translate, color-jitter, randomCrop, horizontal-flip, and random-rotation. The supplement contains a description of each augmentation. These augmentations are not always label preserving, but, as we show, can improve performance when weights are learned." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.905, + 0.489 + ], + "angle": 0, + "content": "Baselines We benchmark results using two conformal scores (translating to different definitions of \\( c(x,y) \\) in Equation 4). The first score is APS [34] (described in Eqn. 4), which represents the cumulative probability required to include the correct class, and the second is RAPS [1], which modifies APS by adding a term to penalize large set sizes. For all experiments, we do not allow sets of size 0. We implement RAPS and APS using code provided by Angelopoulos et al. [1], and automatically select hyperparameters \\( k_{reg} \\) and \\( \\lambda \\) to minimize set size. We also compare against conformal prediction using a simple average over the test-time augmentations (TTA-Avg). In the supplement, we also compare against non-conformal Top-1 and Top-5 prediction sets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.905, + 0.735 + ], + "angle": 0, + "content": "Evaluation We evaluate results using the three metrics commonly used in the conformal prediction literature: efficiency, coverage, and adaptivity. We quantify efficiency using both average prediction set size (measured across all examples) and class-conditional prediction set size (measured across all examples in a class). Coverage is the percentage of sets containing the true label. We define adaptivity as the size-stratified coverage violation (SSCV), introduced by Angelopoulos et al. [1]. We first partition examples based upon the size of the prediction set. We create bins for set sizes of \\([0, 1], [2, 3], [4, 10], [11, 100],\\) and \\([101,]\\). We then compute the empirical coverage within each bin, and compute adaptivity as the maximum difference between theoretical coverage and empirical coverage across bins. The closer this value is to 0, the better the adaptivity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "For each dataset, we report results across 10 randomly generated splits into validation and test sets. For all experiments (save for the validation set size experiment), the validation set and test set are the same size. We allot \\(20\\%\\) of examples from the validation set to \\(D^{(TTA)}\\) (used to learn TTA policy), and allot the remaining examples to the calibration set. For the experiment studying validation set size, we downsample the validation set. We compute statistical significance using a paired t-test, with a Bonferroni correction [46] for multiple hypothesis testing. Code to reproduce all experiments will be made publicly available." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20625" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.09, + 0.905, + 0.289 + ], + "angle": 0, + "content": "
Expanded Augmentation PolicySimple Augmentation Policy
AlphaMethodImageNetiNaturalistCUB-BirdsImageNetiNaturalistCUB-Birds
0.01RAPS37.751 ± 2.33461.437 ± 6.06715.293 ± 2.07137.751 ± 2.33461.437 ± 6.06715.293 ± 2.071
0.01RAPS+TTA-Avg35.600 ± 2.20057.073 ± 5.91413.111 ± 2.47031.681 ± 3.05754.169 ± 6.31914.550 ± 1.425
0.01RAPS+TTA-Learned31.248 ± 2.17753.195 ± 4.88414.045 ± 1.32332.702 ± 2.40951.391 ± 5.21113.803 ± 1.734
0.05RAPS5.637 ± 0.3577.991 ± 1.5213.624 ± 0.3615.637 ± 0.3577.991 ± 1.5213.624 ± 0.361
0.05RAPS+TTA-Avg5.318 ± 0.1137.067 ± 0.3443.116 ± 0.2104.908 ± 0.0996.451 ± 0.2793.249 ± 0.307
0.05RAPS+TTA-Learned4.889 ± 0.1686.682 ± 0.4473.571 ± 0.5765.040 ± 0.1766.788 ± 0.4963.290 ± 0.186
0.10RAPS2.548 ± 0.0742.914 ± 0.1162.038 ± 0.1532.548 ± 0.0742.914 ± 0.1162.038 ± 0.153
0.10RAPS+TTA-Avg2.470 ± 0.0712.740 ± 0.0261.780 ± 0.1392.327 ± 0.0862.610 ± 0.0311.881 ± 0.118
0.10RAPS+TTA-Learned2.312 ± 0.0542.625 ± 0.0431.893 ± 0.1872.362 ± 0.0652.638 ± 0.0261.840 ± 0.106
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.298, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Table 1. Reductions in prediction set size across datasets, augmentation policies, and coverage specifications. Each entry corresponds to the average prediction set size across 10 calibration/test splits. Bolded entries represent performance that is either (a) significantly better compared to the baseline (RAPS), or (b) indistinguishable from the best approach. Table S8 reports achieved coverage. Corresponding results for APS can be found in Table S2." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.371, + 0.361, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.371, + 0.625, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.372, + 0.888, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.531, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Figure 2. Robustness to distribution shift. We compare average prediction set size achieved by RAPS (yellow) to average prediction set size achieved when combining RAPS with TTA-Learned (green). Results reflect the distribution of average prediction set size across 10 runs using ImageNet and ResNet50. We evaluate performance on different corruptions (x-axis) and different coverage guarantees (left, middle, right). RAPS+TTA-Learned (green) produces a noticeable reduction in prediction set size, even when subject to distribution shift, with no loss in coverage. Refer to Figure S1 in the supplement for a comparison of coverage achieved by both methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.627, + 0.178, + 0.643 + ], + "angle": 0, + "content": "6. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.789 + ], + "angle": 0, + "content": "We compare against RAPS, which outperformed other baselines in every experiment. (We provide results comparing our method to APS and the Top-K baselines in the supplement. Variants of each experiment across multiple \\(\\alpha\\) and datasets are also in the supplement.) We then examine the dependence of these results on dataset, base model, and class. We conclude by providing intuition about why test-time augmentation improves the efficiency of conformal predictors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.375, + 0.818 + ], + "angle": 0, + "content": "6.1. Reductions in prediction set size" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We begin with results in the context of the expanded augmentation policy. Learned test-time augmentation policies produce meaningfully significant reductions in prediction set size (RAPS+TTA-Learned in Table 1 and APS+TTA-Learned in Table S1). TTA-Learned reduces prediction set" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.688 + ], + "angle": 0, + "content": "sizes significantly in 16 of the 18 cases, and performs comparably in the remaining 2. Across all cases, the combination of RAPS, TTA-Learned, and the expanded augmentation policy produces the smallest average set sizes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.84 + ], + "angle": 0, + "content": "TTA-Learned performs comparably or better than TTA-Avg in all comparisons. Certain augmentations in the expanded augmentation policy (blur, decrease sharpness, and shear) are consistently assigned a weight of 0, while certain augmentations are consistently included in learned policies (autocontrast, translate). Augmentations assigned zero weight provide no additional information about the true label (for example, they may not preserve the label within the image, or they may be redundant with other augmentations included in the policy)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "While TTA improves both RAPS and APS, the improvements are larger for APS. This is because TTA, like RAPS, tempers the predicted probabilities. TTA lowers the maximum predicted probability on average, thereby reducing" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20626" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.09, + 0.905, + 0.296 + ], + "angle": 0, + "content": "
Expanded Augmentation PolicySimple Augmentation Policy
AlphaMethodResNet-50ResNet-101ResNet-152ResNet-50ResNet-101ResNet-152
0.01RAPS37.751 ± 2.33433.624 ± 1.79629.560 ± 3.48137.751 ± 2.33433.624 ± 1.79629.560 ± 3.481
0.01RAPS+TTA-Avg35.600 ± 2.20030.220 ± 1.77427.203 ± 2.52631.681 ± 3.05727.206 ± 1.84024.106 ± 2.100
0.01RAPS+TTA-Learned31.248 ± 2.17725.722 ± 1.71323.615 ± 1.65632.702 ± 2.40926.760 ± 1.97424.765 ± 2.736
0.05RAPS5.637 ± 0.3574.785 ± 0.1024.376 ± 0.0785.637 ± 0.3574.785 ± 0.1024.376 ± 0.078
0.05RAPS+TTA-Avg5.318 ± 0.1134.433 ± 0.1374.163 ± 0.1854.908 ± 0.0994.147 ± 0.1223.868 ± 0.126
0.05RAPS+TTA-Learned4.889 ± 0.1684.200 ± 0.2003.824 ± 0.1285.040 ± 0.1764.194 ± 0.1943.916 ± 0.356
0.10RAPS2.548 ± 0.0742.267 ± 0.0242.109 ± 0.0272.548 ± 0.0742.267 ± 0.0242.109 ± 0.027
0.10RAPS+TTA-Avg2.470 ± 0.0712.164 ± 0.0312.049 ± 0.0282.327 ± 0.0862.093 ± 0.0351.996 ± 0.018
0.10RAPS+TTA-Learned2.312 ± 0.0542.099 ± 0.0401.993 ± 0.0262.362 ± 0.0652.091 ± 0.0411.988 ± 0.020
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.305, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Table 2. Reductions in prediction set size across base classifiers on ImageNet. TTA-Learned can bridge the performance gap between different classifiers (for example, outperforming ResNet-152 alone when combined with ResNet-101), and yields significant reductions in set size regardless of the pretrained classifier used. We report achieved coverage in Table S9." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.482, + 0.45 + ], + "angle": 0, + "content": "model overconfidence. Consequently, the predicted probability assigned to the remaining classes is higher. This is why the expanded augmentation policy demonstrates stronger performance than the simple augmentation policy: it tempers the probabilities to a greater extent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.483, + 0.532 + ], + "angle": 0, + "content": "TTA-Learned preserves coverage across all experiments, since it respects the assumption of exchangeability. In some cases, TTA significantly improves coverage, although the magnitude of this difference is small (results can be found in Tables S8 and S9)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.6 + ], + "angle": 0, + "content": "We next evaluate adaptivity using size stratified coverage violation (SSCV). At low alpha (\\(\\alpha = .01\\), and \\(\\alpha = .05\\)), TTA-Learned improves efficiency without diminishing adaptivity (Table S10)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.63, + 0.367, + 0.645 + ], + "angle": 0, + "content": "6.2. Robustness to distribution shift" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Next, we evaluate the performance of test-time augmented conformal prediction on out-of-distribution examples. While conformal prediction does not guarantee coverage in these settings, distribution shifts are ubiquitous in practice [22]. Empirical performance is thus of practical interest. The training procedures for both test-time augmented conformal prediction and conformal prediction remain the same and use in-distribution examples from ImageNet and a ResNet50 classifier. We evaluate each conformal predictor on four types of image corruptions drawn from ImageNet-C [19]. Figure 2 plots the results; across all corruptions and coverage guarantees, test-time augmented conformal prediction produces smaller prediction sets than conformal prediction alone. Importantly, test-time augmented conformal prediction achieves this with no loss of coverage (Figure S1)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.373, + 0.887, + 0.389 + ], + "angle": 0, + "content": "6.3. Datasets, augmentation policies, and models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.517 + ], + "angle": 0, + "content": "Dependence on dataset TTA consistently improves prediction set sizes on ImageNet and iNaturalist, but not on CUB-Birds. This may be because the validation set size for CUB-Birds (2,827 images) is an order of magnitude smaller than the validation sets for ImageNet (25,000 images) and iNaturalist (50,000 images). This is consistent with our finding that effectiveness of TTA is positively correlated with the size of the validation set (Figure S4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Dependence on augmentation policy We find that the expanded augmentation policy produces larger reductions in set size than the simple augmentation policy, primarily at low \\(\\alpha\\). This is in spite of the fact that the expanded augmentation policy contains many augmentations outside of the base classifier's training-time augmentation policy. When we vary the number of augmentations included in an augmentation policy, we see that larger augmentation policies also yield greater reductions in average prediction set size (Figure S2). That said, the simple augmentation policy does have its place; it requires fewer forward passes during inference. In the absence of a learned aggregation function, our results suggest that aggregating using an average can still improve the efficiency of conformal predictors (outperforming the original conformal score in 11 comparisons, matching performance in 3, and under-performing in 4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Dependence on base model We tested the generalizability of our results to other models by rerunning the ImageNet experiments using ResNet-101 (accuracy of \\(77.4\\%\\)) and ResNet-152 (accuracy of \\(78.3\\%\\)). Unsurprisingly, more accurate models result in smaller prediction set sizes (Table 2). TTA variants of conformal prediction again produce significant improvements in set size while maintaining coverage." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20627" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.104, + 0.118, + 0.115 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.104, + 0.326, + 0.121 + ], + "angle": 0, + "content": "Distribution of Set Sizes (ImageNet, \\(\\alpha = 0.01\\) Expanded)" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.121, + 0.368, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.104, + 0.397, + 0.116 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.106, + 0.582, + 0.121 + ], + "angle": 0, + "content": "Original Set Size \\((r = 0.89, \\alpha = 0.01\\), Expanded)" + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.121, + 0.63, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.104, + 0.658, + 0.116 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.106, + 0.812, + 0.119 + ], + "angle": 0, + "content": "Class Difficulty 55. \\(\\alpha = 0.01\\) Exp" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.12, + 0.89, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.907, + 0.329 + ], + "angle": 0, + "content": "Figure 3. (A) Class-conditional prediction set sizes. We plot the distribution of class-conditional prediction set sizes, for ImageNet and ResNet-50 with \\(\\alpha = .01\\). RAPS+TTA-Learned (green) produces a noticeable reduction in class-conditional prediction set sizes. (B, C) Relationship between TTA improvements and original class set sizes and class difficulty. TTA introduces the largest improvements for classes with the largest original prediction set sizes (B) and classes on which the underlying classifier is often incorrect (C). Each point represents the average prediction set size for each class, across 10 splits." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.356, + 0.415, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.356, + 0.651, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.356, + 0.895, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.564, + 0.907, + 0.649 + ], + "angle": 0, + "content": "Figure 4. (A) Effect of TTA-Learned on optimal Top-K: TTA-Learned significantly lowers the value of k required for Top-k prediction sets to achieve coverage on ImageNet and iNaturalist, but not on CUB-Birds. (B,C) Effect of TTA-Learned on rank of true class: TTA-Learned improves the rank of the true class among the sorted predicted probabilities for a given example for both ImageNet (B) and iNaturalist (C). We plot the rank using the original predicted probabilities compared to the TTA-transformed probabilities, binning all examples in the validation set into five equal-width bins. Dots that fall below the red line indicate that TTA improves the rank of the true class." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.675, + 0.483, + 0.766 + ], + "angle": 0, + "content": "age. We note that the combination of TTA with ResNet-101 produces smaller set sizes than the more complex ResNet-152 alone. For example, when \\(\\alpha\\) is set to .01, RAPS+TTA-Learned and ResNet-101 produce set sizes that contain, on average, 26.5 classes, while RAPS and ResNet-152 produce an average set size of 29.6." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.773, + 0.304, + 0.79 + ], + "angle": 0, + "content": "6.4. Class-Specific Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We have established that on average TTA is a useful addition to the conformal pipeline. We now investigate the source of this improvement. We make two empirical observations. First, classes with larger predicted set sizes benefit most from the introduction of TTA. Figure 3 shows that a class's average prediction set size is significantly correlated with the change in set size produced by TTA-Learned (with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.812 + ], + "angle": 0, + "content": "the expanded augmentation policy and \\(\\alpha = .01\\), \\(r = 0.89\\), and \\(p < 1e - 10\\). Second, we find that class difficulty is significantly associated with changes in set size introduced by TTA (with the expanded augmentation policy and \\(\\alpha = .01\\), \\(r = 0.55\\) and \\(p < 1e - 10\\)). Prediction sets for classes that are difficult to predict benefit more from TTA compared to their easier counterparts. These observations are related; harder classes receive larger set sizes, and consequently, offer larger room for improvements in efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.619, + 0.833 + ], + "angle": 0, + "content": "6.5. Intuition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Why does the addition of test-time augmentation produce smaller prediction set sizes? In short, TTA improves top-K accuracy. We verify this claim by estimating \\( k \\) such that the uncertainty sets comprised of the top \\( k \\) predicted" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20628" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "classes for each example achieve a marginal coverage of \\((1 - \\alpha)\\). We see that for those datasets (ImageNet and iNaturalist) where TTA produces significant reductions in set size, TTA-transformed predictions—both with a simple average and learned weights—produce significantly lower values for \\(k\\) compared to the original predictions (Figure 4A). This is not true for CUB-Birds, on which TTA offers little to no improvement. One could use such a procedure to determine whether TTA is worth adding to a conformal pipeline without collecting labeled data beyond the calibration dataset." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.256, + 0.485, + 0.603 + ], + "angle": 0, + "content": "Another way to understand the impact of TTA is to consider the effect on the ordering of classes. It has been observed in the test-time augmentation literature that TTA often promotes the true class from the second-highest to the highest predicted probability, thereby correcting the classification. This finding does not fully explain the value of test-time augmentation to conformal prediction since only \\(3\\%\\) of the prediction sets which reduce in size are associated with a corrected Top-1 classification. What TTA did do was increase the predicted probability of the true class even when it is predicted to be unlikely (for example, promoting the true class from 200th most likely to 100th most likely). We visualize this effect in Figure 4 by plotting the change in true class rank (the index at which the true class appears in the sorted list of predicted probabilities) for all examples in the validation set, stratified into 5 equal-width bins. The lower left point captures examples that are classified correctly; here, test-time augmentation introduces little to no change. In subsequent bins, we see that TTA typically promotes the rank of the true class. We also include the standard deviation across the true class ranks in the original predicted probabilities (x-axis) and the TTA-transformed probabilities (y-axis)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.214, + 0.669 + ], + "angle": 0, + "content": "7. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Learned test-time augmentation policies require two ingredients: labeled data and multiple forward passes. Although one can minimize costs by parallelizing computation or by using the learned weights to identify which augmentations to generate, inference will always cost more with test-time augmentation. Our results are also limited to image classification. We do not consider other modalities, for which appropriate transformations will substantially differ. Future work should consider how these results may generalize to non-vision tasks. Finally, test-time augmentation is one approach to generating ensembles in conformal prediction. Many other more computationally expensive approaches exist. The tradeoff between computation and ensemble performance remains a useful avenue for future work." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.634, + 0.107 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.907, + 0.42 + ], + "angle": 0, + "content": "We show that test-time augmented conformal prediction produces smaller sets than conformal prediction alone. The proposed approach is effective, efficient, and simple: it reduces prediction set sizes by up to \\(30\\%\\), requires no model re-training, and relies on a portion of labeled data already available to split conformal predictors. Our experiments also indicate that test-time augmented conformal prediction exhibits greater efficiency under four common corruption-based distribution shifts. Test-time augmentation is able to do this by improving the underlying classifier's robustness to domain-specific invariances, in the form of data augmentation. Efforts to improve the efficiency of conformal predictors could, as a first step, aim to improve the robustness of the underlying classifier. The performance of TTA-Learned also suggests that there are settings in which it is wise to use a portion of the labeled data to improve the underlying model, instead of reserving all labeled data for the calibration set. In sum, our work takes a step towards practically useful conformal predictors by improving efficiency, without sacrificing adaptivity or coverage." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.433, + 0.611, + 0.449 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.458, + 0.907, + 0.514 + ], + "angle": 0, + "content": "[1] Anastasios Angelopoulos, Stephen Bates, Jitendra Malik, and Michael I. Jordan. Uncertainty Sets for Image Classifiers using Conformal Prediction, 2022. arXiv:2009.14193 [cs, math, stat]. 2, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.516, + 0.906, + 0.558 + ], + "angle": 0, + "content": "[2] Murat Seçkin Ayhan and Philipp Berens. Test-time Data Augmentation for Estimation of Heteroscedastic Aleatoric Uncertainty in Deep Neural Networks. 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.56, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[3] Varun Babbar, Umang Bhatt, and Adrian Weller. On the Utility of Prediction Sets in Human-AI Teams, 2022. arXiv:2205.01411 [cs]. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.603, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[4] Anthony Bellotti. Optimized conformal classification using gradient descent approximation, 2021. arXiv:2105.11255 [cs]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.647, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[5] Lars Carlsson, Martin Eklund, and Ulf Norinder. Aggregated Conformal Prediction. In Artificial Intelligence Applications and Innovations, pages 231-240, Berlin, Heidelberg, 2014. Springer. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[6] Sewhan Chun, Jae Young Lee, and Junmo Kim. Cyclic test time augmentation with entropy weight method. In Proceedings of the Thirty-Eighth Conference on Uncertainty in Artificial Intelligence, pages 433-442. PMLR, 2022. ISSN: 2640-3498. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.775, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[7] Gilad Cohen and Raja Giryes. Simple post-training robustness using test time augmentations and random forest. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3996-4006, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[8] Seffi Cohen, Noa Dagan, Nurit Cohen-Inger, Dan Ofer, and Lior Rokach. ICU Survival Prediction Incorporating Test-Time Augmentation to Improve the Accuracy of Ensemble-Based Models. IEEE Access, 9:91584-91592, 2021. Conference Name: IEEE Access. 2" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.458, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20629" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[9] Pedro Conde, Tiago Barros, Rui L. Lopes, Cristiano Premebida, and Urbano J. Nunes. Approaching Test Time Augmentation in the Context of Uncertainty Calibration for Deep Neural Networks, 2023. arXiv:2304.05104 [cs]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[10] Jesse C Cresswell, Yi Sui, Bhargava Kumar, and Noel Vouitsis. Conformal prediction sets improve human decision making. arXiv preprint arXiv:2401.13744, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[11] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. AutoAugment: Learning Augmentation Policies from Data, 2019. arXiv:1805.09501 [cs, stat]. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. ISSN: 1063-6919. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[13] Tiffany Ding, Anastasios N. Angelopoulos, Stephen Bates, Michael I. Jordan, and Ryan J. Tibshirani. Class-Conditional Conformal Prediction With Many Classes, 2023. arXiv:2306.09335 [cs, stat]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[14] Bat-Sheva Einbinder, Yaniv Romano, Matteo Sesia, and Yanfei Zhou. Training Uncertainty-Aware Classifiers with Conformalized Deep Learning. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[15] Shohei Enomoto, Monikka Roslianna Busto, and Takeharu Eda. Dynamic Test-Time Augmentation via Differentiable Functions, 2023. arXiv:2212.04681 [cs]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[16] A. Gammerman, V. Vovk, and V. Vapnik. Learning by transduction. In Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence, pages 148-155, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers Inc. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, Las Vegas, NV, USA, 2016. IEEE. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[18] Achim Hekler, Titus J. Brinker, and Florian Buettner. Test Time Augmentation Meets Post-hoc Calibration: Uncertainty Quantification under Real-World Conditions. Proceedings of the AAAI Conference on Artificial Intelligence, 37(12):14856-14864, 2023. Number: 12. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[19] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[20] Vilde Jensen, Filippo Maria Bianchi, and Stian Normann Anfinsen. Ensemble Conformalized Quantile Regression for Probabilistic Time Series Forecasting. IEEE Transactions on Neural Networks and Learning Systems, pages 1-12, 2022. Conference Name: IEEE Transactions on Neural Networks and Learning Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[21] Ildoo Kim, Younghoon Kim, and Sungwoong Kim. Learning Loss for Test-Time Augmentation. In Advances in Neural Information Processing Systems, pages 4163-4174. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[22] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[24] Arun Kumar Kuchibhotla. Exchangeability, conformal prediction, and rank tests. arXiv preprint arXiv:2005.06095, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[25] Henrik Linusson, Ulf Norinder, Henrik Bostrom, Ulf Johansson, and Tuve Lofstrom. On the Calibration of Aggregated Conformal Predictors. In Proceedings of the Sixth Workshop on Conformal and Probabilistic Prediction and Applications, pages 154-173. PMLR, 2017. ISSN: 2640-3498. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[26] H. Linusson, U. Johansson, and H. Bostrom. Efficient conformal predictor ensembles. Neurocomputing, 397:266-278, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[27] Helen Lu, Divya Shanmugam, Harini Suresh, and John Guttag. Improved Text Classification via Test-Time Augmentation, 2022. arXiv:2206.13607 [cs]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[28] Alexander Lyzhov, Yuliya Molchanova, Arsenii Ashukha, Dmitry Molchanov, and Dmitry Vetrov. Greedy Policy Search: A Simple Baseline for Learnable Test-Time Augmentation. In Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI), pages 1308-1317. PMLR, 2020. ISSN: 2640-3498. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[29] Luca Mocerino, Roberto G. Rizzo, Valentino Peluso, Andrea Calimera, and Enrico Macii. Adaptive Test-Time Augmentation for Low-Power CPU, 2021. arXiv:2105.06183 [cs, eess]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[30] Niers, Tom. iNaturalist_competition, 2021. original-date: 2021-12-10T10:56:46Z. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[32] Juan C. Perez, Motasem Alfarra, Guillaume Jeanneret, Laura Rueda, Ali Thabet, Bernard Ghanem, and Pablo Arbelaez. Enhancing Adversarial Robustness via Test-time Transformation Ensembling. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 81-91, Montreal, BC, Canada, 2021. IEEE. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[33] Drew Prinster, Anqi Liu, and Suchi Saria. JAWS: Auditing Predictive Uncertainty Under Covariate Shift, 2022. arXiv:2207.10716 [cs, stat]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[34] Yaniv Romano, Matteo Sesia, and Emmanuel J. Candès. Classification with Valid and Adaptive Coverage, 2020. arXiv:2006.02544 [stat]. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[35] C Saunders and Royal Holloway. Transduction with Confidence and Credibility. 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[36] Glenn Shafer and Vladimir Vovk. A Tutorial on Conformal Prediction. 2008. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20630" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[37] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better Aggregation in Test-Time Augmentation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1194–1203, Montreal, QC, Canada, 2021. IEEE. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[38] David Stutz, Krishnamurthy, Dvijotham, Ali Taylan Cemgil, and Arnaud Doucet. Learning Optimal Conformal Classifiers, 2022. arXiv:2110.09192 [cs, stat]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[40] Ryan J Tibshirani, Rina Foygel Barber, Emmanuel Candes, and Aaditya Ramdas. Conformal Prediction Under Covariate Shift. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[41] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha. Benchmarking Representation Learning for Natural World Image Collections, 2021. arXiv:2103.16483 [cs]. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[42] Vladimir Vovk. Algorithmic Learning in a Random World. Springer-Verlag, New York, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[43] Vladimir Vovk. Cross-conformal predictors, 2012. arXiv:1208.0806 [cs, stat]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.485, + 0.473 + ], + "angle": 0, + "content": "[44] Vladimir Vovk. Cross-conformal predictors. Annals of Mathematics and Artificial Intelligence, 74(1-2):9-28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[45] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[46] Eric W. Weisstein. Bonferroni correction. https://mathworld.wolfram.com/, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.485, + 0.617 + ], + "angle": 0, + "content": "[47] Dongping Zhang, Angelos Chatzimparmpas, Negar Kamali, and Jessica Hullman. Evaluating the utility of conformal prediction sets for ai-advised image labeling. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1-19, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[48] Marvin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test Time Robustness via Adaptation and Augmentation, 2022. arXiv:2110.09506 [cs]. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.485, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20631" + } + ] +] \ No newline at end of file diff --git a/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_origin.pdf b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..de874b2ca9d524c1434de9a4025ac6a6d37d5c2c --- /dev/null +++ b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/69bebec2-b46a-4233-826c-ee6981af40a9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e60f9d3fe31bc1ae524332e40781d51f18270b08b2b94912799de2a6b92f9a1 +size 1338773 diff --git a/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/full.md b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..71ac24a0700f1be07416ea312687b77707735a7d --- /dev/null +++ b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/full.md @@ -0,0 +1,271 @@ +# Test-time augmentation improves efficiency in conformal prediction + +Divya Shanmugam, Helen Lu, Swami Sankaranarayanan, John Guttag Massachusetts Institute of Technology, CSAIL + +{divyas, helenlu, swamiviv, guttag}@mit.edu + +# Abstract + +A conformal classifier produces a set of predicted classes and provides a probabilistic guarantee that the set includes the true class. Unfortunately, it is often the case that conformal classifiers produce uninformatively large sets. In this work, we show that test-time augmentation (TTA)-a technique that introduces inductive biases during inference-reduces the size of the sets produced by conformal classifiers. Our approach is flexible, computationally efficient, and effective. It can be combined with any conformal score, requires no model retraining, and reduces prediction set sizes by $10\% - 14\%$ on average. We conduct an evaluation of the approach spanning three datasets, three models, two established conformal scoring methods, different guarantee strengths, and several distribution shifts to show when and why test-time augmentation is a useful addition to the conformal pipeline. + +# 1. Introduction + +Conformal prediction has emerged as a promising way to equip existing classifiers with statistically valid uncertainty estimates. It does so by replacing the prediction of the most likely class with an uncertainty set-a set of classes accompanied by a probabilistic guarantee that the true class appears in the set [36]. + +Conformal prediction faces two limitations in practice. First, achieving a suitably strong guarantee often leads to prediction sets that are uninformatively large [10, 47]. Large prediction sets have been shown to decrease performance when provided as a decision-making aid [3, 10, 47]. + +Second, conformal classifiers inherit the instability of the underlying models. As a result, prediction sets can change significantly in response to small input perturbations, a well-known weakness of neural networks [39]. Applying a horizontal flip to each image in ImageNet, for example, changes the prediction set sizes for $75\%$ of examples at a coverage guarantee of $99\%$ . Such behavior also represents a barrier to broader use. + +In this work, we show that test-time augmentation + +(TTA), a widely-used technique in computer vision, has the potential to address both limitations. TTA involves generating an ensemble of predictions by perturbing the input with label-preserving transformations. It has previously been shown that TTA can be used to make non-conformal classifiers more robust to small input perturbations [7], more accurate [37], and better calibrated [18]. However, previous work has not explored the utility of TTA in the context of conformal prediction. + +We propose test-time-augmented conformal prediction, which transforms a classifier's predictions using a learned test-time augmentation policy prior to conformal classification. By using distinct sets of labeled data to learn the test-time augmentation policy and the conformal classifier, we preserve the assumption of exchangeability, and thereby the coverage guarantee associated with the conformal predictor. + +In experiments testing the performance of conformal predictors subject to distribution shift, we see that test-time augmentation reduces prediction set sizes by $14\%$ on average, with no loss of coverage. And even when there is no distribution shift, we see a reduction of $10\%$ on average. Moreover, we find that classes with the largest average prediction set sizes benefit most from the introduction of test-time augmentation. We also show that test-time augmentation can bridge gaps between classifiers of different sizes (e.g., we show that test-time augmentation combined with ResNet-50 produces smaller set sizes than ResNet-101 without test-time augmentation). + +Our analysis of why test-time augmentation reduces prediction set sizes reveals a previously unknown effect of test-time augmentation. Specifically, TTA increases the predicted probability of the true class even when it is predicted to be unlikely (for example, promoting the true class from 200th most likely to 100th most likely). Although such behavior has no impact on the maximum predicted probability — commonly the focus of literature on test-time augmentation [2, 32, 37] — it is valuable in conformal prediction. This is because when the true class is promoted to a higher rank among a classifier's predicted probabilities, the conformal classifier includes fewer incorrect classes to meet the conformal guarantee. + +![](images/0f09daba25237aab21fb82fc714476f5b309ea49fbe8010157df0b208660700a.jpg) +Test-time augmented conformal prediction... + +![](images/aac26790bf41d560b74786c3f60b128359b374c21758c4201e7e752e9a50f1a2.jpg) +...leads to smaller prediction set sizes. +Figure 1. We illustrate the addition of test-time augmentation to conformal calibration in green (left) and provide a snapshot of the improvements it can confer (right). We show results on Imagenet, with a desired coverage of $95\%$ , for the 20 classes with the largest predicted set sizes on average (computed over 10 calibration/test splits). + +Contributions We make three contributions. To begin, this is the first work to propose combining test-time augmentation, a popular technique in computer vision, with conformal prediction. Second, we present a method that reduces the prediction set sizes of existing conformal predictors by using automatically learned test-time augmentations. Finally, we demonstrate, in an extensive set of experiments, that our approach to combining conformal prediction and test-time augmentation leads to smaller prediction sets. + +# 2. Related work + +Conformal prediction was first introduced by Gammerman et al. [16], and further developed by Saunders and Holloway [35] and Vladimir Vovk [42]. We review efforts to ensemble conformal predictors and efforts to reduce prediction set sizes below. + +Ensembles in conformal prediction Several methods that generate ensembles of conformal predictors are known to improve efficiency. These methods include cross-conformal prediction [43], bootstrap conformal prediction [44], aggregated conformal prediction [5, 25], and out-of-bag conformal prediction [26]. The approaches primarily differ in how data is sampled to create the training dataset for the classifier and the calibration dataset for the conformal predictor. However, all require training multiple base classifiers or conformal predictors. Our approach is distinct: we propose a technique to generate an ensemble from a single model by perturbing the input, which requires no additional base models and no additional conformal predictors. + +Efficiency in conformal prediction There are two ways to improve efficiency in split conformal prediction: adjustments to the conformal score or improvements to the underlying classifier. Many works have proposed new procedures to estimate and apply thresholds on conformal scores [1, 4, 13, 33, 40]. Romano et al. [34] proposed APS, a conformal score based on the cumulative probability required to include the correct class in a prediction set. Angelopoulos et al. [1] built on this work to propose RAPS, which modifies APS by penalizing the inclusion of low-probability classes. Comparatively little work has focused on improvements to the underlying model. Jensen et al. [20] ensemble a set of base classifiers, where the classifiers are created by training models on subsets of the training data. Stutz et al. [38] provide a new way to train the base classifier and conformal predictor jointly through a novel conformal training loss. In contrast, our work focuses on improving the underlying model without retraining, and can be easily combined with any of the above procedures. + +Test-Time Augmentation Test-time augmentation (TTA) is a popular technique to improve the accuracy, robustness, and calibration of an existing classifier by aggregating predictions over a set of input transformations [2, 9, 15, 18, 32, 37, 48]. TTA has been applied to a diverse range of predictive tasks across domains ranging from healthcare [8] to content moderation [27]. Consequently, many have proposed new ways to perform TTA—for example, learning when to apply TTA [29], which augmentations to use [6, 21, 28], and how to aggregate the resulting predictions [6, 9, 37]. Existing work typically focuses on test-time augmentation's impact on highest predicted probability. Here, we analyze how test-time augmentation increases the pre + +dicted probability assigned to the true class when it appears outside the top few classes, and how that change is consequential in conformal prediction. + +# 3. Problem setting + +We operate within the split conformal prediction framework. In this setting, a conformal classifier $\mathcal{C}(x_i) \subset \{1, \ldots, K\}$ maps input $x_i$ to a subset of $K$ possible classes and requires three inputs: + +- Calibration set $D^{(cal)} = \{(x_1, y_1), \ldots, (x_n, y_n)\}$ , containing $n$ labeled examples. +- Classifier $f: \mathcal{X} \mapsto \Delta^K$ , mapping input domain $\mathcal{X}$ to a probability distribution over $K$ classes. +- Desired upper bound on error rate $\alpha \in [0,1]$ , where $(1 - \alpha)$ is the probability the set contains the true class. + +We study the introduction of two variables drawn from the test-time augmentation literature: + +- Augmentation policy $\mathcal{A} = \{a_0, \dots, a_{m-1}\}$ , consisting of $m$ augmentation functions, where $a_0$ is the identity transform. +- Aggregation function $\hat{g}$ , which aggregates a set of predictions to produce one prediction. + +Each variable translates to a key choice in test-time augmentation: what augmentations to apply $(\mathcal{A})$ and how to aggregate the resulting probabilities $(\hat{g})$ . + +# 4. Approach + +Our goal is to learn an aggregation function $\hat{g}$ to maximize the accuracy of the underlying classifier, and ultimately reduce the sizes of the prediction sets generated from the classifier's predicted probabilities. We briefly outline the conformal approach, and then detail the mechanics of our method (illustrated in Figure 1). For a detailed introduction to conformal prediction, refer to Shafer and Vovk [36]. + +Conformal predictors accept three inputs: a probabilistic classifier $f$ , a calibration set $\mathcal{D}^{(cal)}$ , and a pre-specified error rate $\alpha$ . Using the these inputs, one can construct a conformal predictor in three steps: + +1. Define a score function $c(x_i, y_i)$ , which produces a conformal score representing the uncertainty of the input example and label pair. +2. Produce a distribution of conformal scores by computing $c(x_i, y_i)$ for all $(x_i, y_i) \in \mathcal{D}^{(cal)}$ . +3. Compute threshold $\hat{q}$ as the $\lceil (n + 1)(1 - \alpha)\rceil /n$ quantile of the distribution of conformal scores over $n$ examples in the calibration set, combined with $\{\infty \}$ . + +For a new example $x$ , we compute $c(x_i, y)$ for all $y \in \{1, \dots, K\}$ , and include all $y$ for which $c(x_i, y) < \hat{q}$ . We adopt the conformal score proposed by Romano et al. [34], which equates to the cumulative probability required to include the correct class: + +$$ +\hat {\pi} _ {x} \left(y ^ {\prime}\right) = \hat {p} \left(y = y ^ {\prime} \mid x\right) = f (x) _ {y ^ {\prime}} \tag {1} +$$ + +$$ +\rho_ {x} (y) = \sum_ {y ^ {\prime} = 1} ^ {K} \hat {\pi} _ {x} \left(y ^ {\prime}\right) \mathbb {I} \left[ \hat {\pi} _ {x} \left(y ^ {\prime}\right) > \hat {\pi} _ {x} (y) \right] \tag {2} +$$ + +$$ +c (x, y) = \rho_ {x} (y) + u \cdot \hat {\pi} _ {x} (y) \tag {3} +$$ + +where $\rho_{x}(y)$ is the cumulative probability of all classes predicted with higher probability than $y$ and $\hat{\pi}_x(y')$ corresponds to the predicted probability of class $y^\prime$ given $x$ . The variable $u\sim U(0,1)$ . Conformal score $c(x_i,y_i)$ is thus composed of this cumulative probability and the predicted probability of class $y_{i}$ . + +**Proposal** Our approach differs from prior work in that the conformal score is derived by transforming the probabilities output by $f$ using test-time augmentation. Concretely, this replaces Equation 1 with the following, parametrized by augmentation policy $\mathcal{A}$ and augmentation weights $\theta$ . + +$$ +\hat {\pi} _ {x} \left(y ^ {\prime}\right) = \hat {p} \left(y = y ^ {\prime} \mid x _ {i}\right) = g \left(x _ {i}; f, \mathcal {A}, \theta\right) \tag {4} +$$ + +Aggregation weights $\theta$ are applied to the logits output by classifier $f$ , and transformed to be proper probabilities by applying a softmax function. We learn the aggregation weights $\theta$ using a portion of the validation set, $D^{(TTA)}$ , distinct from calibration set used to identify the conformal threshold $(D^{(cal)})$ . In contrast to traditional approaches, where all labeled data is used to estimate the conformal threshold, we instead reserve a portion to learn the test-time augmentation policy. + +We learn a set of weights which maximize classification accuracy on $D^{(TTA)}$ by minimizing the cross-entropy loss1 computed between the predicted probabilities and true labels. More formally, $g$ applies $\theta$ and $\mathcal{A}$ as follows: + +$$ +g \left(x _ {i}; f, \mathcal {A}, \Theta\right) = \sigma \left(\theta^ {T} \mathbf {A} \left(f, \mathcal {A}, x _ {i}\right)\right) \tag {5} +$$ + +where $\mathbf{A}$ uses $f$ to map input $x_{i}$ to a $M\times K$ matrix of predicted logits where $M$ is the number of augmentations and $K$ is the number of classes. $\theta$ is a $1\times m$ vector corresponding to augmentation-specific weights. Each row in $\mathbf{A}(f,\mathcal{A},x_i)$ represents the pre-trained classifier's predicted logits on augmentation $a_{m}$ of $x_{i}$ . TTA-Learned refers to TTA combined with learned augmentation weights, while TTA-Avg refers to a simple average over the augmentations. + +We refer to the fraction of the validation set allotted to $D^{(TTA)}$ as $\beta$ . Figure S3 shows that performance is not sensitive to the choice of $\beta$ ; as a result, all experiments use $\beta = .2$ (see supplement for further discussion). This does reduce the amount of data available to identify the appropriate threshold, but we find that the benefits TTA confers outweigh the cost to threshold estimation. Computational cost scales linearly with the size of $\mathcal{A}$ ; each additional augmentation translates to a forward pass of the base classifier. One can use the learned weights to save computation by identifying which test-time augmentations to generate. + +Preserving exchangeability The validity of conformal prediction depends upon the assumption of exchangeability: that all orderings of examples are equally likely (in effect, meaning that the distribution of examples in the calibration set is indistinguishable from the distribution of unseen examples). Prior work has shown that the conformal procedure is valid under deterministic transformations [24]; by using distinct examples to learn the test-time augmentation policy, the proposed approach constitutes a deterministic transformation applied to both the calibration set and unseen examples. If we were to instead use the same examples to learn the test-time augmentation policy and the conformal threshold, exchangeability could be broken. + +# 5. Experimental Set-Up + +Datasets We show results on the test splits of three widely used image classification datasets: ImageNet [12] (50,000 natural images spanning 1,000 classes), iNaturalist [41] (100,000 images spanning 10,000 species), and CUB-Birds [45] (5,794 images spanning 200 categories of birds). Images are distributed evenly over classes in ImageNet and iNaturalist, while CUB-Birds has between 11 and 30 images per class. + +Models The default model architecture, across all datasets, is ResNet-50 [17]. The accuracies of the base classifiers are $76.1\%$ (ImageNet), $76.4\%$ (iNaturalist), and $80.5\%$ (CUB-Birds). To study the relationship between model complexity and performance, we also provide results using ResNet-101 and ResNet-152 on ImageNet. For ImageNet, we make use of the pretrained models made available by PyTorch [31]. For iNaturalist, we use a model made public by Niers, Tom [30]. For CUB-Birds, we train a network by finetuning the final layer of a ResNet-50 model initialized with ImageNet's pretrained weights. + +Augmentations We consider two augmentation policies. The first (the simple augmentation policy) consists of a random-crop and a horizontal-flip; to produce a random crop, we pad the original image with 4 pixels and take a + +256x256 crop of the expanded image (thereby preserving the original image resolution). The simple augmentation policy is widely used because its augmentations are likely to be label-preserving. The second, which we refer to as the expanded augmentation policy, consists of 12 augmentations: increase-sharpness, decrease-sharpness, autocontrast, invert, blur, posterize, shear, translate, color-jitter, randomCrop, horizontal-flip, and random-rotation. The supplement contains a description of each augmentation. These augmentations are not always label preserving, but, as we show, can improve performance when weights are learned. + +Baselines We benchmark results using two conformal scores (translating to different definitions of $c(x,y)$ in Equation 4). The first score is APS [34] (described in Eqn. 4), which represents the cumulative probability required to include the correct class, and the second is RAPS [1], which modifies APS by adding a term to penalize large set sizes. For all experiments, we do not allow sets of size 0. We implement RAPS and APS using code provided by Angelopoulos et al. [1], and automatically select hyperparameters $k_{reg}$ and $\lambda$ to minimize set size. We also compare against conformal prediction using a simple average over the test-time augmentations (TTA-Avg). In the supplement, we also compare against non-conformal Top-1 and Top-5 prediction sets. + +Evaluation We evaluate results using the three metrics commonly used in the conformal prediction literature: efficiency, coverage, and adaptivity. We quantify efficiency using both average prediction set size (measured across all examples) and class-conditional prediction set size (measured across all examples in a class). Coverage is the percentage of sets containing the true label. We define adaptivity as the size-stratified coverage violation (SSCV), introduced by Angelopoulos et al. [1]. We first partition examples based upon the size of the prediction set. We create bins for set sizes of $[0, 1], [2, 3], [4, 10], [11, 100],$ and $[101,]$ . We then compute the empirical coverage within each bin, and compute adaptivity as the maximum difference between theoretical coverage and empirical coverage across bins. The closer this value is to 0, the better the adaptivity. + +For each dataset, we report results across 10 randomly generated splits into validation and test sets. For all experiments (save for the validation set size experiment), the validation set and test set are the same size. We allot $20\%$ of examples from the validation set to $D^{(TTA)}$ (used to learn TTA policy), and allot the remaining examples to the calibration set. For the experiment studying validation set size, we downsample the validation set. We compute statistical significance using a paired t-test, with a Bonferroni correction [46] for multiple hypothesis testing. Code to reproduce all experiments will be made publicly available. + +
Expanded Augmentation PolicySimple Augmentation Policy
AlphaMethodImageNetiNaturalistCUB-BirdsImageNetiNaturalistCUB-Birds
0.01RAPS37.751 ± 2.33461.437 ± 6.06715.293 ± 2.07137.751 ± 2.33461.437 ± 6.06715.293 ± 2.071
0.01RAPS+TTA-Avg35.600 ± 2.20057.073 ± 5.91413.111 ± 2.47031.681 ± 3.05754.169 ± 6.31914.550 ± 1.425
0.01RAPS+TTA-Learned31.248 ± 2.17753.195 ± 4.88414.045 ± 1.32332.702 ± 2.40951.391 ± 5.21113.803 ± 1.734
0.05RAPS5.637 ± 0.3577.991 ± 1.5213.624 ± 0.3615.637 ± 0.3577.991 ± 1.5213.624 ± 0.361
0.05RAPS+TTA-Avg5.318 ± 0.1137.067 ± 0.3443.116 ± 0.2104.908 ± 0.0996.451 ± 0.2793.249 ± 0.307
0.05RAPS+TTA-Learned4.889 ± 0.1686.682 ± 0.4473.571 ± 0.5765.040 ± 0.1766.788 ± 0.4963.290 ± 0.186
0.10RAPS2.548 ± 0.0742.914 ± 0.1162.038 ± 0.1532.548 ± 0.0742.914 ± 0.1162.038 ± 0.153
0.10RAPS+TTA-Avg2.470 ± 0.0712.740 ± 0.0261.780 ± 0.1392.327 ± 0.0862.610 ± 0.0311.881 ± 0.118
0.10RAPS+TTA-Learned2.312 ± 0.0542.625 ± 0.0431.893 ± 0.1872.362 ± 0.0652.638 ± 0.0261.840 ± 0.106
+ +Table 1. Reductions in prediction set size across datasets, augmentation policies, and coverage specifications. Each entry corresponds to the average prediction set size across 10 calibration/test splits. Bolded entries represent performance that is either (a) significantly better compared to the baseline (RAPS), or (b) indistinguishable from the best approach. Table S8 reports achieved coverage. Corresponding results for APS can be found in Table S2. + +![](images/9044f56e5d35f56b16f36f9295155b24bb8c1d88a49bd59df101cb6f67adc007.jpg) +Figure 2. Robustness to distribution shift. We compare average prediction set size achieved by RAPS (yellow) to average prediction set size achieved when combining RAPS with TTA-Learned (green). Results reflect the distribution of average prediction set size across 10 runs using ImageNet and ResNet50. We evaluate performance on different corruptions (x-axis) and different coverage guarantees (left, middle, right). RAPS+TTA-Learned (green) produces a noticeable reduction in prediction set size, even when subject to distribution shift, with no loss in coverage. Refer to Figure S1 in the supplement for a comparison of coverage achieved by both methods. + +![](images/cfe7c8bd71deb6b161a7c1096d915b736bea184355749c8b52408f327bd35047.jpg) + +![](images/0df7a187b18fc146b681a1d817e79f7f1999f0fee26d4e746dd068d2ab6e1630.jpg) + +# 6. Results + +We compare against RAPS, which outperformed other baselines in every experiment. (We provide results comparing our method to APS and the Top-K baselines in the supplement. Variants of each experiment across multiple $\alpha$ and datasets are also in the supplement.) We then examine the dependence of these results on dataset, base model, and class. We conclude by providing intuition about why test-time augmentation improves the efficiency of conformal predictors. + +# 6.1. Reductions in prediction set size + +We begin with results in the context of the expanded augmentation policy. Learned test-time augmentation policies produce meaningfully significant reductions in prediction set size (RAPS+TTA-Learned in Table 1 and APS+TTA-Learned in Table S1). TTA-Learned reduces prediction set + +sizes significantly in 16 of the 18 cases, and performs comparably in the remaining 2. Across all cases, the combination of RAPS, TTA-Learned, and the expanded augmentation policy produces the smallest average set sizes. + +TTA-Learned performs comparably or better than TTA-Avg in all comparisons. Certain augmentations in the expanded augmentation policy (blur, decrease sharpness, and shear) are consistently assigned a weight of 0, while certain augmentations are consistently included in learned policies (autocontrast, translate). Augmentations assigned zero weight provide no additional information about the true label (for example, they may not preserve the label within the image, or they may be redundant with other augmentations included in the policy). + +While TTA improves both RAPS and APS, the improvements are larger for APS. This is because TTA, like RAPS, tempers the predicted probabilities. TTA lowers the maximum predicted probability on average, thereby reducing + +
Expanded Augmentation PolicySimple Augmentation Policy
AlphaMethodResNet-50ResNet-101ResNet-152ResNet-50ResNet-101ResNet-152
0.01RAPS37.751 ± 2.33433.624 ± 1.79629.560 ± 3.48137.751 ± 2.33433.624 ± 1.79629.560 ± 3.481
0.01RAPS+TTA-Avg35.600 ± 2.20030.220 ± 1.77427.203 ± 2.52631.681 ± 3.05727.206 ± 1.84024.106 ± 2.100
0.01RAPS+TTA-Learned31.248 ± 2.17725.722 ± 1.71323.615 ± 1.65632.702 ± 2.40926.760 ± 1.97424.765 ± 2.736
0.05RAPS5.637 ± 0.3574.785 ± 0.1024.376 ± 0.0785.637 ± 0.3574.785 ± 0.1024.376 ± 0.078
0.05RAPS+TTA-Avg5.318 ± 0.1134.433 ± 0.1374.163 ± 0.1854.908 ± 0.0994.147 ± 0.1223.868 ± 0.126
0.05RAPS+TTA-Learned4.889 ± 0.1684.200 ± 0.2003.824 ± 0.1285.040 ± 0.1764.194 ± 0.1943.916 ± 0.356
0.10RAPS2.548 ± 0.0742.267 ± 0.0242.109 ± 0.0272.548 ± 0.0742.267 ± 0.0242.109 ± 0.027
0.10RAPS+TTA-Avg2.470 ± 0.0712.164 ± 0.0312.049 ± 0.0282.327 ± 0.0862.093 ± 0.0351.996 ± 0.018
0.10RAPS+TTA-Learned2.312 ± 0.0542.099 ± 0.0401.993 ± 0.0262.362 ± 0.0652.091 ± 0.0411.988 ± 0.020
+ +Table 2. Reductions in prediction set size across base classifiers on ImageNet. TTA-Learned can bridge the performance gap between different classifiers (for example, outperforming ResNet-152 alone when combined with ResNet-101), and yields significant reductions in set size regardless of the pretrained classifier used. We report achieved coverage in Table S9. + +model overconfidence. Consequently, the predicted probability assigned to the remaining classes is higher. This is why the expanded augmentation policy demonstrates stronger performance than the simple augmentation policy: it tempers the probabilities to a greater extent. + +TTA-Learned preserves coverage across all experiments, since it respects the assumption of exchangeability. In some cases, TTA significantly improves coverage, although the magnitude of this difference is small (results can be found in Tables S8 and S9). + +We next evaluate adaptivity using size stratified coverage violation (SSCV). At low alpha ( $\alpha = .01$ , and $\alpha = .05$ ), TTA-Learned improves efficiency without diminishing adaptivity (Table S10). + +# 6.2. Robustness to distribution shift + +Next, we evaluate the performance of test-time augmented conformal prediction on out-of-distribution examples. While conformal prediction does not guarantee coverage in these settings, distribution shifts are ubiquitous in practice [22]. Empirical performance is thus of practical interest. The training procedures for both test-time augmented conformal prediction and conformal prediction remain the same and use in-distribution examples from ImageNet and a ResNet50 classifier. We evaluate each conformal predictor on four types of image corruptions drawn from ImageNet-C [19]. Figure 2 plots the results; across all corruptions and coverage guarantees, test-time augmented conformal prediction produces smaller prediction sets than conformal prediction alone. Importantly, test-time augmented conformal prediction achieves this with no loss of coverage (Figure S1). + +# 6.3. Datasets, augmentation policies, and models + +Dependence on dataset TTA consistently improves prediction set sizes on ImageNet and iNaturalist, but not on CUB-Birds. This may be because the validation set size for CUB-Birds (2,827 images) is an order of magnitude smaller than the validation sets for ImageNet (25,000 images) and iNaturalist (50,000 images). This is consistent with our finding that effectiveness of TTA is positively correlated with the size of the validation set (Figure S4). + +Dependence on augmentation policy We find that the expanded augmentation policy produces larger reductions in set size than the simple augmentation policy, primarily at low $\alpha$ . This is in spite of the fact that the expanded augmentation policy contains many augmentations outside of the base classifier's training-time augmentation policy. When we vary the number of augmentations included in an augmentation policy, we see that larger augmentation policies also yield greater reductions in average prediction set size (Figure S2). That said, the simple augmentation policy does have its place; it requires fewer forward passes during inference. In the absence of a learned aggregation function, our results suggest that aggregating using an average can still improve the efficiency of conformal predictors (outperforming the original conformal score in 11 comparisons, matching performance in 3, and under-performing in 4). + +Dependence on base model We tested the generalizability of our results to other models by rerunning the ImageNet experiments using ResNet-101 (accuracy of $77.4\%$ ) and ResNet-152 (accuracy of $78.3\%$ ). Unsurprisingly, more accurate models result in smaller prediction set sizes (Table 2). TTA variants of conformal prediction again produce significant improvements in set size while maintaining coverage. + +![](images/36e53bae95142b1447e11f48bc24ccc9f5bc24a5187ab1f95f57952547fffab4.jpg) +A +Distribution of Set Sizes (ImageNet, $\alpha = 0.01$ Expanded) +Figure 3. (A) Class-conditional prediction set sizes. We plot the distribution of class-conditional prediction set sizes, for ImageNet and ResNet-50 with $\alpha = .01$ . RAPS+TTA-Learned (green) produces a noticeable reduction in class-conditional prediction set sizes. (B, C) Relationship between TTA improvements and original class set sizes and class difficulty. TTA introduces the largest improvements for classes with the largest original prediction set sizes (B) and classes on which the underlying classifier is often incorrect (C). Each point represents the average prediction set size for each class, across 10 splits. + +![](images/e43903664fe6477547f726dab51f0a6f4191f8e4d21eeead62eb480eb0d220a3.jpg) +B +Original Set Size $(r = 0.89, \alpha = 0.01$ , Expanded) + +![](images/7b69bdf3185435d4650bdd511a9d3e20de883f89c8593e692ae4801f00b89eac.jpg) +C +Class Difficulty 55. $\alpha = 0.01$ Exp + +![](images/951c946d5ab97999c62824298e9d768559c84a83b72391c06eef90a190b2ecd0.jpg) + +![](images/980d366e395e611caa4ee3f02f3450fe3432fd5bae5a4ea26f9f71c4a2ac276f.jpg) + +![](images/d4d64b53b7a22eb991b69b832da77cc5c13578a9fae7f5cad97978d0a55960ee.jpg) +Figure 4. (A) Effect of TTA-Learned on optimal Top-K: TTA-Learned significantly lowers the value of k required for Top-k prediction sets to achieve coverage on ImageNet and iNaturalist, but not on CUB-Birds. (B,C) Effect of TTA-Learned on rank of true class: TTA-Learned improves the rank of the true class among the sorted predicted probabilities for a given example for both ImageNet (B) and iNaturalist (C). We plot the rank using the original predicted probabilities compared to the TTA-transformed probabilities, binning all examples in the validation set into five equal-width bins. Dots that fall below the red line indicate that TTA improves the rank of the true class. + +age. We note that the combination of TTA with ResNet-101 produces smaller set sizes than the more complex ResNet-152 alone. For example, when $\alpha$ is set to .01, RAPS+TTA-Learned and ResNet-101 produce set sizes that contain, on average, 26.5 classes, while RAPS and ResNet-152 produce an average set size of 29.6. + +# 6.4. Class-Specific Analysis + +We have established that on average TTA is a useful addition to the conformal pipeline. We now investigate the source of this improvement. We make two empirical observations. First, classes with larger predicted set sizes benefit most from the introduction of TTA. Figure 3 shows that a class's average prediction set size is significantly correlated with the change in set size produced by TTA-Learned (with + +the expanded augmentation policy and $\alpha = .01$ , $r = 0.89$ , and $p < 1e - 10$ . Second, we find that class difficulty is significantly associated with changes in set size introduced by TTA (with the expanded augmentation policy and $\alpha = .01$ , $r = 0.55$ and $p < 1e - 10$ ). Prediction sets for classes that are difficult to predict benefit more from TTA compared to their easier counterparts. These observations are related; harder classes receive larger set sizes, and consequently, offer larger room for improvements in efficiency. + +# 6.5. Intuition + +Why does the addition of test-time augmentation produce smaller prediction set sizes? In short, TTA improves top-K accuracy. We verify this claim by estimating $k$ such that the uncertainty sets comprised of the top $k$ predicted + +classes for each example achieve a marginal coverage of $(1 - \alpha)$ . We see that for those datasets (ImageNet and iNaturalist) where TTA produces significant reductions in set size, TTA-transformed predictions—both with a simple average and learned weights—produce significantly lower values for $k$ compared to the original predictions (Figure 4A). This is not true for CUB-Birds, on which TTA offers little to no improvement. One could use such a procedure to determine whether TTA is worth adding to a conformal pipeline without collecting labeled data beyond the calibration dataset. + +Another way to understand the impact of TTA is to consider the effect on the ordering of classes. It has been observed in the test-time augmentation literature that TTA often promotes the true class from the second-highest to the highest predicted probability, thereby correcting the classification. This finding does not fully explain the value of test-time augmentation to conformal prediction since only $3\%$ of the prediction sets which reduce in size are associated with a corrected Top-1 classification. What TTA did do was increase the predicted probability of the true class even when it is predicted to be unlikely (for example, promoting the true class from 200th most likely to 100th most likely). We visualize this effect in Figure 4 by plotting the change in true class rank (the index at which the true class appears in the sorted list of predicted probabilities) for all examples in the validation set, stratified into 5 equal-width bins. The lower left point captures examples that are classified correctly; here, test-time augmentation introduces little to no change. In subsequent bins, we see that TTA typically promotes the rank of the true class. We also include the standard deviation across the true class ranks in the original predicted probabilities (x-axis) and the TTA-transformed probabilities (y-axis). + +# 7. Limitations + +Learned test-time augmentation policies require two ingredients: labeled data and multiple forward passes. Although one can minimize costs by parallelizing computation or by using the learned weights to identify which augmentations to generate, inference will always cost more with test-time augmentation. Our results are also limited to image classification. We do not consider other modalities, for which appropriate transformations will substantially differ. Future work should consider how these results may generalize to non-vision tasks. Finally, test-time augmentation is one approach to generating ensembles in conformal prediction. Many other more computationally expensive approaches exist. The tradeoff between computation and ensemble performance remains a useful avenue for future work. + +# 8. Conclusion + +We show that test-time augmented conformal prediction produces smaller sets than conformal prediction alone. The proposed approach is effective, efficient, and simple: it reduces prediction set sizes by up to $30\%$ , requires no model re-training, and relies on a portion of labeled data already available to split conformal predictors. Our experiments also indicate that test-time augmented conformal prediction exhibits greater efficiency under four common corruption-based distribution shifts. Test-time augmentation is able to do this by improving the underlying classifier's robustness to domain-specific invariances, in the form of data augmentation. Efforts to improve the efficiency of conformal predictors could, as a first step, aim to improve the robustness of the underlying classifier. The performance of TTA-Learned also suggests that there are settings in which it is wise to use a portion of the labeled data to improve the underlying model, instead of reserving all labeled data for the calibration set. In sum, our work takes a step towards practically useful conformal predictors by improving efficiency, without sacrificing adaptivity or coverage. + +# References + +[1] Anastasios Angelopoulos, Stephen Bates, Jitendra Malik, and Michael I. Jordan. Uncertainty Sets for Image Classifiers using Conformal Prediction, 2022. arXiv:2009.14193 [cs, math, stat]. 2, 4, 11 +[2] Murat Seçkin Ayhan and Philipp Berens. Test-time Data Augmentation for Estimation of Heteroscedastic Aleatoric Uncertainty in Deep Neural Networks. 2018. 1, 2 +[3] Varun Babbar, Umang Bhatt, and Adrian Weller. On the Utility of Prediction Sets in Human-AI Teams, 2022. arXiv:2205.01411 [cs]. 1 +[4] Anthony Bellotti. Optimized conformal classification using gradient descent approximation, 2021. arXiv:2105.11255 [cs]. 2 +[5] Lars Carlsson, Martin Eklund, and Ulf Norinder. Aggregated Conformal Prediction. In Artificial Intelligence Applications and Innovations, pages 231-240, Berlin, Heidelberg, 2014. Springer. 2 +[6] Sewhan Chun, Jae Young Lee, and Junmo Kim. Cyclic test time augmentation with entropy weight method. In Proceedings of the Thirty-Eighth Conference on Uncertainty in Artificial Intelligence, pages 433-442. PMLR, 2022. ISSN: 2640-3498. 2 +[7] Gilad Cohen and Raja Giryes. Simple post-training robustness using test time augmentations and random forest. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3996-4006, 2024. 1 +[8] Seffi Cohen, Noa Dagan, Nurit Cohen-Inger, Dan Ofer, and Lior Rokach. ICU Survival Prediction Incorporating Test-Time Augmentation to Improve the Accuracy of Ensemble-Based Models. IEEE Access, 9:91584-91592, 2021. Conference Name: IEEE Access. 2 + +[9] Pedro Conde, Tiago Barros, Rui L. Lopes, Cristiano Premebida, and Urbano J. Nunes. Approaching Test Time Augmentation in the Context of Uncertainty Calibration for Deep Neural Networks, 2023. arXiv:2304.05104 [cs]. 2 +[10] Jesse C Cresswell, Yi Sui, Bhargava Kumar, and Noel Vouitsis. Conformal prediction sets improve human decision making. arXiv preprint arXiv:2401.13744, 2024. 1 +[11] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. AutoAugment: Learning Augmentation Policies from Data, 2019. arXiv:1805.09501 [cs, stat]. 11 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. ISSN: 1063-6919. 4 +[13] Tiffany Ding, Anastasios N. Angelopoulos, Stephen Bates, Michael I. Jordan, and Ryan J. Tibshirani. Class-Conditional Conformal Prediction With Many Classes, 2023. arXiv:2306.09335 [cs, stat]. 2 +[14] Bat-Sheva Einbinder, Yaniv Romano, Matteo Sesia, and Yanfei Zhou. Training Uncertainty-Aware Classifiers with Conformalized Deep Learning. 15 +[15] Shohei Enomoto, Monikka Roslianna Busto, and Takeharu Eda. Dynamic Test-Time Augmentation via Differentiable Functions, 2023. arXiv:2212.04681 [cs]. 2 +[16] A. Gammerman, V. Vovk, and V. Vapnik. Learning by transduction. In Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence, pages 148-155, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers Inc. 2 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, Las Vegas, NV, USA, 2016. IEEE. 4 +[18] Achim Hekler, Titus J. Brinker, and Florian Buettner. Test Time Augmentation Meets Post-hoc Calibration: Uncertainty Quantification under Real-World Conditions. Proceedings of the AAAI Conference on Artificial Intelligence, 37(12):14856-14864, 2023. Number: 12. 1, 2 +[19] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 6 +[20] Vilde Jensen, Filippo Maria Bianchi, and Stian Normann Anfinsen. Ensemble Conformalized Quantile Regression for Probabilistic Time Series Forecasting. IEEE Transactions on Neural Networks and Learning Systems, pages 1-12, 2022. Conference Name: IEEE Transactions on Neural Networks and Learning Systems. 2 +[21] Ildoo Kim, Younghoon Kim, and Sungwoong Kim. Learning Loss for Test-Time Augmentation. In Advances in Neural Information Processing Systems, pages 4163-4174. Curran Associates, Inc., 2020. 2 +[22] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas + +Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 6 +[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 11 +[24] Arun Kumar Kuchibhotla. Exchangeability, conformal prediction, and rank tests. arXiv preprint arXiv:2005.06095, 2020. 4 +[25] Henrik Linusson, Ulf Norinder, Henrik Bostrom, Ulf Johansson, and Tuve Lofstrom. On the Calibration of Aggregated Conformal Predictors. In Proceedings of the Sixth Workshop on Conformal and Probabilistic Prediction and Applications, pages 154-173. PMLR, 2017. ISSN: 2640-3498. 2 +[26] H. Linusson, U. Johansson, and H. Bostrom. Efficient conformal predictor ensembles. Neurocomputing, 397:266-278, 2020. 2 +[27] Helen Lu, Divya Shanmugam, Harini Suresh, and John Guttag. Improved Text Classification via Test-Time Augmentation, 2022. arXiv:2206.13607 [cs]. 2 +[28] Alexander Lyzhov, Yuliya Molchanova, Arsenii Ashukha, Dmitry Molchanov, and Dmitry Vetrov. Greedy Policy Search: A Simple Baseline for Learnable Test-Time Augmentation. In Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI), pages 1308-1317. PMLR, 2020. ISSN: 2640-3498. 2 +[29] Luca Mocerino, Roberto G. Rizzo, Valentino Peluso, Andrea Calimera, and Enrico Macii. Adaptive Test-Time Augmentation for Low-Power CPU, 2021. arXiv:2105.06183 [cs, eess]. 2 +[30] Niers, Tom. iNaturalist_competition, 2021. original-date: 2021-12-10T10:56:46Z. 4 +[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 4 +[32] Juan C. Perez, Motasem Alfarra, Guillaume Jeanneret, Laura Rueda, Ali Thabet, Bernard Ghanem, and Pablo Arbelaez. Enhancing Adversarial Robustness via Test-time Transformation Ensembling. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 81-91, Montreal, BC, Canada, 2021. IEEE. 1, 2 +[33] Drew Prinster, Anqi Liu, and Suchi Saria. JAWS: Auditing Predictive Uncertainty Under Covariate Shift, 2022. arXiv:2207.10716 [cs, stat]. 2 +[34] Yaniv Romano, Matteo Sesia, and Emmanuel J. Candès. Classification with Valid and Adaptive Coverage, 2020. arXiv:2006.02544 [stat]. 2, 3, 4 +[35] C Saunders and Royal Holloway. Transduction with Confidence and Credibility. 1999. 2 +[36] Glenn Shafer and Vladimir Vovk. A Tutorial on Conformal Prediction. 2008. 1, 3 + +[37] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better Aggregation in Test-Time Augmentation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1194–1203, Montreal, QC, Canada, 2021. IEEE. 1, 2 +[38] David Stutz, Krishnamurthy, Dvijotham, Ali Taylan Cemgil, and Arnaud Doucet. Learning Optimal Conformal Classifiers, 2022. arXiv:2110.09192 [cs, stat]. 2 +[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1 +[40] Ryan J Tibshirani, Rina Foygel Barber, Emmanuel Candes, and Aaditya Ramdas. Conformal Prediction Under Covariate Shift. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2 +[41] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha. Benchmarking Representation Learning for Natural World Image Collections, 2021. arXiv:2103.16483 [cs]. 4 +[42] Vladimir Vovk. Algorithmic Learning in a Random World. Springer-Verlag, New York, 2005. 2 +[43] Vladimir Vovk. Cross-conformal predictors, 2012. arXiv:1208.0806 [cs, stat]. 2 +[44] Vladimir Vovk. Cross-conformal predictors. Annals of Mathematics and Artificial Intelligence, 74(1-2):9-28, 2015. 2 +[45] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 4 +[46] Eric W. Weisstein. Bonferroni correction. https://mathworld.wolfram.com/, 2004. 4 +[47] Dongping Zhang, Angelos Chatzimparmpas, Negar Kamali, and Jessica Hullman. Evaluating the utility of conformal prediction sets for ai-advised image labeling. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1-19, 2024. 1 +[48] Marvin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test Time Robustness via Adaptation and Augmentation, 2022. arXiv:2110.09506 [cs]. 2 \ No newline at end of file diff --git a/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/images.zip b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ef3ae0e8b6943a182c63d98da1cc0212aacfa34c --- /dev/null +++ b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf114283457f83b069748e484fecdb8ef17723ebef0276df0d89d2bd725eb013 +size 423159 diff --git a/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/layout.json b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..19d2ff0e537cfe9573ae8a57eb293a767ac641e0 --- /dev/null +++ b/2025/Test-time Augmentation Improves Efficiency in Conformal Prediction/layout.json @@ -0,0 +1,8292 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 515, + 121 + ], + "type": "text", + "content": "Test-time augmentation improves efficiency in conformal prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 143, + 475, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 143, + 475, + 171 + ], + "spans": [ + { + "bbox": [ + 135, + 143, + 475, + 171 + ], + "type": "text", + "content": "Divya Shanmugam, Helen Lu, Swami Sankaranarayanan, John Guttag Massachusetts Institute of Technology, CSAIL" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 173, + 419, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 173, + 419, + 185 + ], + "spans": [ + { + "bbox": [ + 187, + 173, + 419, + 185 + ], + "type": "text", + "content": "{divyas, helenlu, swamiviv, guttag}@mit.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 201, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 201, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 201, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 238, + 297, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 297, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 297, + 430 + ], + "type": "text", + "content": "A conformal classifier produces a set of predicted classes and provides a probabilistic guarantee that the set includes the true class. Unfortunately, it is often the case that conformal classifiers produce uninformatively large sets. In this work, we show that test-time augmentation (TTA)-a technique that introduces inductive biases during inference-reduces the size of the sets produced by conformal classifiers. Our approach is flexible, computationally efficient, and effective. It can be combined with any conformal score, requires no model retraining, and reduces prediction set sizes by " + }, + { + "bbox": [ + 54, + 238, + 297, + 430 + ], + "type": "inline_equation", + "content": "10\\% - 14\\%" + }, + { + "bbox": [ + 54, + 238, + 297, + 430 + ], + "type": "text", + "content": " on average. We conduct an evaluation of the approach spanning three datasets, three models, two established conformal scoring methods, different guarantee strengths, and several distribution shifts to show when and why test-time augmentation is a useful addition to the conformal pipeline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 453, + 135, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 135, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 135, + 466 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 473, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 544 + ], + "type": "text", + "content": "Conformal prediction has emerged as a promising way to equip existing classifiers with statistically valid uncertainty estimates. It does so by replacing the prediction of the most likely class with an uncertainty set-a set of classes accompanied by a probabilistic guarantee that the true class appears in the set [36]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": "Conformal prediction faces two limitations in practice. First, achieving a suitably strong guarantee often leads to prediction sets that are uninformatively large [10, 47]. Large prediction sets have been shown to decrease performance when provided as a decision-making aid [3, 10, 47]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": "Second, conformal classifiers inherit the instability of the underlying models. As a result, prediction sets can change significantly in response to small input perturbations, a well-known weakness of neural networks [39]. Applying a horizontal flip to each image in ImageNet, for example, changes the prediction set sizes for " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": " of examples at a coverage guarantee of " + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": ". Such behavior also represents a barrier to broader use." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "In this work, we show that test-time augmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 213, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 321 + ], + "type": "text", + "content": "(TTA), a widely-used technique in computer vision, has the potential to address both limitations. TTA involves generating an ensemble of predictions by perturbing the input with label-preserving transformations. It has previously been shown that TTA can be used to make non-conformal classifiers more robust to small input perturbations [7], more accurate [37], and better calibrated [18]. However, previous work has not explored the utility of TTA in the context of conformal prediction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 324, + 556, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 556, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 556, + 409 + ], + "type": "text", + "content": "We propose test-time-augmented conformal prediction, which transforms a classifier's predictions using a learned test-time augmentation policy prior to conformal classification. By using distinct sets of labeled data to learn the test-time augmentation policy and the conformal classifier, we preserve the assumption of exchangeability, and thereby the coverage guarantee associated with the conformal predictor." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 411, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 555 + ], + "type": "text", + "content": "In experiments testing the performance of conformal predictors subject to distribution shift, we see that test-time augmentation reduces prediction set sizes by " + }, + { + "bbox": [ + 313, + 411, + 556, + 555 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 313, + 411, + 556, + 555 + ], + "type": "text", + "content": " on average, with no loss of coverage. And even when there is no distribution shift, we see a reduction of " + }, + { + "bbox": [ + 313, + 411, + 556, + 555 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 411, + 556, + 555 + ], + "type": "text", + "content": " on average. Moreover, we find that classes with the largest average prediction set sizes benefit most from the introduction of test-time augmentation. We also show that test-time augmentation can bridge gaps between classifiers of different sizes (e.g., we show that test-time augmentation combined with ResNet-50 produces smaller set sizes than ResNet-101 without test-time augmentation)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 713 + ], + "type": "text", + "content": "Our analysis of why test-time augmentation reduces prediction set sizes reveals a previously unknown effect of test-time augmentation. Specifically, TTA increases the predicted probability of the true class even when it is predicted to be unlikely (for example, promoting the true class from 200th most likely to 100th most likely). Although such behavior has no impact on the maximum predicted probability — commonly the focus of literature on test-time augmentation [2, 32, 37] — it is valuable in conformal prediction. This is because when the true class is promoted to a higher rank among a classifier's predicted probabilities, the conformal classifier includes fewer incorrect classes to meet the conformal guarantee." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20622" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 95, + 390, + 232 + ], + "blocks": [ + { + "bbox": [ + 63, + 78, + 236, + 88 + ], + "lines": [ + { + "bbox": [ + 63, + 78, + 236, + 88 + ], + "spans": [ + { + "bbox": [ + 63, + 78, + 236, + 88 + ], + "type": "text", + "content": "Test-time augmented conformal prediction..." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 95, + 390, + 232 + ], + "lines": [ + { + "bbox": [ + 60, + 95, + 390, + 232 + ], + "spans": [ + { + "bbox": [ + 60, + 95, + 390, + 232 + ], + "type": "image", + "image_path": "0f09daba25237aab21fb82fc714476f5b309ea49fbe8010157df0b208660700a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 406, + 93, + 546, + 230 + ], + "blocks": [ + { + "bbox": [ + 395, + 78, + 544, + 88 + ], + "lines": [ + { + "bbox": [ + 395, + 78, + 544, + 88 + ], + "spans": [ + { + "bbox": [ + 395, + 78, + 544, + 88 + ], + "type": "text", + "content": "...leads to smaller prediction set sizes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 93, + 546, + 230 + ], + "lines": [ + { + "bbox": [ + 406, + 93, + 546, + 230 + ], + "spans": [ + { + "bbox": [ + 406, + 93, + 546, + 230 + ], + "type": "image", + "image_path": "aac26790bf41d560b74786c3f60b128359b374c21758c4201e7e752e9a50f1a2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 256, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 555, + 291 + ], + "type": "text", + "content": "Figure 1. We illustrate the addition of test-time augmentation to conformal calibration in green (left) and provide a snapshot of the improvements it can confer (right). We show results on Imagenet, with a desired coverage of " + }, + { + "bbox": [ + 55, + 256, + 555, + 291 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 256, + 555, + 291 + ], + "type": "text", + "content": ", for the 20 classes with the largest predicted set sizes on average (computed over 10 calibration/test splits)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 311, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 430 + ], + "type": "text", + "content": "Contributions We make three contributions. To begin, this is the first work to propose combining test-time augmentation, a popular technique in computer vision, with conformal prediction. Second, we present a method that reduces the prediction set sizes of existing conformal predictors by using automatically learned test-time augmentations. Finally, we demonstrate, in an extensive set of experiments, that our approach to combining conformal prediction and test-time augmentation leads to smaller prediction sets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 139, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 139, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 139, + 456 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 466, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 296, + 525 + ], + "type": "text", + "content": "Conformal prediction was first introduced by Gammerman et al. [16], and further developed by Saunders and Holloway [35] and Vladimir Vovk [42]. We review efforts to ensemble conformal predictors and efforts to reduce prediction set sizes below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 546, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 712 + ], + "type": "text", + "content": "Ensembles in conformal prediction Several methods that generate ensembles of conformal predictors are known to improve efficiency. These methods include cross-conformal prediction [43], bootstrap conformal prediction [44], aggregated conformal prediction [5, 25], and out-of-bag conformal prediction [26]. The approaches primarily differ in how data is sampled to create the training dataset for the classifier and the calibration dataset for the conformal predictor. However, all require training multiple base classifiers or conformal predictors. Our approach is distinct: we propose a technique to generate an ensemble from a single model by perturbing the input, which requires no additional base models and no additional conformal predictors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 311, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 538 + ], + "type": "text", + "content": "Efficiency in conformal prediction There are two ways to improve efficiency in split conformal prediction: adjustments to the conformal score or improvements to the underlying classifier. Many works have proposed new procedures to estimate and apply thresholds on conformal scores [1, 4, 13, 33, 40]. Romano et al. [34] proposed APS, a conformal score based on the cumulative probability required to include the correct class in a prediction set. Angelopoulos et al. [1] built on this work to propose RAPS, which modifies APS by penalizing the inclusion of low-probability classes. Comparatively little work has focused on improvements to the underlying model. Jensen et al. [20] ensemble a set of base classifiers, where the classifiers are created by training models on subsets of the training data. Stutz et al. [38] provide a new way to train the base classifier and conformal predictor jointly through a novel conformal training loss. In contrast, our work focuses on improving the underlying model without retraining, and can be easily combined with any of the above procedures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Test-Time Augmentation Test-time augmentation (TTA) is a popular technique to improve the accuracy, robustness, and calibration of an existing classifier by aggregating predictions over a set of input transformations [2, 9, 15, 18, 32, 37, 48]. TTA has been applied to a diverse range of predictive tasks across domains ranging from healthcare [8] to content moderation [27]. Consequently, many have proposed new ways to perform TTA—for example, learning when to apply TTA [29], which augmentations to use [6, 21, 28], and how to aggregate the resulting predictions [6, 9, 37]. Existing work typically focuses on test-time augmentation's impact on highest predicted probability. Here, we analyze how test-time augmentation increases the pre" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20623" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "dicted probability assigned to the true class when it appears outside the top few classes, and how that change is consequential in conformal prediction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 153, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 153, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 153, + 133 + ], + "type": "text", + "content": "3. Problem setting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 295, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 188 + ], + "type": "text", + "content": "We operate within the split conformal prediction framework. In this setting, a conformal classifier " + }, + { + "bbox": [ + 55, + 140, + 295, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(x_i) \\subset \\{1, \\ldots, K\\}" + }, + { + "bbox": [ + 55, + 140, + 295, + 188 + ], + "type": "text", + "content": " maps input " + }, + { + "bbox": [ + 55, + 140, + 295, + 188 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 140, + 295, + 188 + ], + "type": "text", + "content": " to a subset of " + }, + { + "bbox": [ + 55, + 140, + 295, + 188 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 140, + 295, + 188 + ], + "type": "text", + "content": " possible classes and requires three inputs:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 295, + 262 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "type": "text", + "content": "- Calibration set " + }, + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "type": "inline_equation", + "content": "D^{(cal)} = \\{(x_1, y_1), \\ldots, (x_n, y_n)\\}" + }, + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "type": "text", + "content": ", containing " + }, + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "type": "text", + "content": " labeled examples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "type": "text", + "content": "- Classifier " + }, + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\mapsto \\Delta^K" + }, + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "type": "text", + "content": ", mapping input domain " + }, + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "type": "text", + "content": " to a probability distribution over " + }, + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 214, + 295, + 237 + ], + "type": "text", + "content": " classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "type": "text", + "content": "- Desired upper bound on error rate " + }, + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "type": "inline_equation", + "content": "(1 - \\alpha)" + }, + { + "bbox": [ + 55, + 238, + 295, + 262 + ], + "type": "text", + "content": " is the probability the set contains the true class." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 265, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 289 + ], + "type": "text", + "content": "We study the introduction of two variables drawn from the test-time augmentation literature:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 291, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 326 + ], + "type": "text", + "content": "- Augmentation policy " + }, + { + "bbox": [ + 55, + 291, + 295, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{a_0, \\dots, a_{m-1}\\}" + }, + { + "bbox": [ + 55, + 291, + 295, + 326 + ], + "type": "text", + "content": ", consisting of " + }, + { + "bbox": [ + 55, + 291, + 295, + 326 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 291, + 295, + 326 + ], + "type": "text", + "content": " augmentation functions, where " + }, + { + "bbox": [ + 55, + 291, + 295, + 326 + ], + "type": "inline_equation", + "content": "a_0" + }, + { + "bbox": [ + 55, + 291, + 295, + 326 + ], + "type": "text", + "content": " is the identity transform." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 327, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 351 + ], + "type": "text", + "content": "- Aggregation function " + }, + { + "bbox": [ + 55, + 327, + 295, + 351 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 55, + 327, + 295, + 351 + ], + "type": "text", + "content": ", which aggregates a set of predictions to produce one prediction." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "text", + "content": "Each variable translates to a key choice in test-time augmentation: what augmentations to apply " + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "inline_equation", + "content": "(\\mathcal{A})" + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "text", + "content": " and how to aggregate the resulting probabilities " + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "inline_equation", + "content": "(\\hat{g})" + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 401, + 123, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 123, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 123, + 415 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 421, + 296, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 505 + ], + "type": "text", + "content": "Our goal is to learn an aggregation function " + }, + { + "bbox": [ + 55, + 421, + 296, + 505 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 55, + 421, + 296, + 505 + ], + "type": "text", + "content": " to maximize the accuracy of the underlying classifier, and ultimately reduce the sizes of the prediction sets generated from the classifier's predicted probabilities. We briefly outline the conformal approach, and then detail the mechanics of our method (illustrated in Figure 1). For a detailed introduction to conformal prediction, refer to Shafer and Vovk [36]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "content": "Conformal predictors accept three inputs: a probabilistic classifier " + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "content": ", a calibration set " + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(cal)}" + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "content": ", and a pre-specified error rate " + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "content": ". Using the these inputs, one can construct a conformal predictor in three steps:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 555, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 295, + 591 + ], + "type": "text", + "content": "1. Define a score function " + }, + { + "bbox": [ + 56, + 555, + 295, + 591 + ], + "type": "inline_equation", + "content": "c(x_i, y_i)" + }, + { + "bbox": [ + 56, + 555, + 295, + 591 + ], + "type": "text", + "content": ", which produces a conformal score representing the uncertainty of the input example and label pair." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "text", + "content": "2. Produce a distribution of conformal scores by computing " + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "inline_equation", + "content": "c(x_i, y_i)" + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "inline_equation", + "content": "(x_i, y_i) \\in \\mathcal{D}^{(cal)}" + }, + { + "bbox": [ + 55, + 591, + 295, + 615 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "text", + "content": "3. Compute threshold " + }, + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\lceil (n + 1)(1 - \\alpha)\\rceil /n" + }, + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "text", + "content": " quantile of the distribution of conformal scores over " + }, + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "text", + "content": " examples in the calibration set, combined with " + }, + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\{\\infty \\}" + }, + { + "bbox": [ + 55, + 615, + 295, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "For a new example " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": ", we compute " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "c(x_i, y)" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "y \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": ", and include all " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " for which " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "c(x_i, y) < \\hat{q}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": ". We adopt the conformal score proposed by Romano et al. [34], which equates to the cumulative probability required to include the correct class:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 353, + 94, + 555, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 94, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 353, + 94, + 555, + 108 + ], + "type": "interline_equation", + "content": "\\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) = \\hat {p} \\left(y = y ^ {\\prime} \\mid x\\right) = f (x) _ {y ^ {\\prime}} \\tag {1}", + "image_path": "eb711521a0082b5d8e20d9add28cc0cd9237be1fce5116c6b7c0a046064d25e0.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 358, + 110, + 553, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 110, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 358, + 110, + 553, + 144 + ], + "type": "interline_equation", + "content": "\\rho_ {x} (y) = \\sum_ {y ^ {\\prime} = 1} ^ {K} \\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) \\mathbb {I} \\left[ \\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) > \\hat {\\pi} _ {x} (y) \\right] \\tag {2}", + "image_path": "1c96929cde390604cad97c8a9ea2d6ac81e5f255844ca2acc9823dfa7badeaa1.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 353, + 148, + 555, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 148, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 353, + 148, + 555, + 160 + ], + "type": "interline_equation", + "content": "c (x, y) = \\rho_ {x} (y) + u \\cdot \\hat {\\pi} _ {x} (y) \\tag {3}", + "image_path": "c10d674bbdbb383a2929a027dec01ecdd6c8cb9c888b6aaaa2ad3e516e8b58dc.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\rho_{x}(y)" + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "content": " is the cumulative probability of all classes predicted with higher probability than " + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "inline_equation", + "content": "\\hat{\\pi}_x(y')" + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "content": " corresponds to the predicted probability of class " + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "inline_equation", + "content": "y^\\prime" + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "content": ". The variable " + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "inline_equation", + "content": "u\\sim U(0,1)" + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "content": ". Conformal score " + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "inline_equation", + "content": "c(x_i,y_i)" + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "content": " is thus composed of this cumulative probability and the predicted probability of class " + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 182, + 555, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "content": "**Proposal** Our approach differs from prior work in that the conformal score is derived by transforming the probabilities output by " + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "content": " using test-time augmentation. Concretely, this replaces Equation 1 with the following, parametrized by augmentation policy " + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "content": " and augmentation weights " + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 271, + 554, + 331 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 353, + 353, + 555, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 353, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 353, + 353, + 555, + 368 + ], + "type": "interline_equation", + "content": "\\hat {\\pi} _ {x} \\left(y ^ {\\prime}\\right) = \\hat {p} \\left(y = y ^ {\\prime} \\mid x _ {i}\\right) = g \\left(x _ {i}; f, \\mathcal {A}, \\theta\\right) \\tag {4}", + "image_path": "e3838a0a8c14792e0f29eb5bebf8ba7478dfa1929206749f47b14072dea8dd1f.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": "Aggregation weights " + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": " are applied to the logits output by classifier " + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": ", and transformed to be proper probabilities by applying a softmax function. We learn the aggregation weights " + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": " using a portion of the validation set, " + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "inline_equation", + "content": "D^{(TTA)}" + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": ", distinct from calibration set used to identify the conformal threshold " + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "inline_equation", + "content": "(D^{(cal)})" + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": ". In contrast to traditional approaches, where all labeled data is used to estimate the conformal threshold, we instead reserve a portion to learn the test-time augmentation policy." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "text", + "content": "We learn a set of weights which maximize classification accuracy on " + }, + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "inline_equation", + "content": "D^{(TTA)}" + }, + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "text", + "content": " by minimizing the cross-entropy loss1 computed between the predicted probabilities and true labels. More formally, " + }, + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "text", + "content": " applies " + }, + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 485, + 554, + 534 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 353, + 555, + 555, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 555, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 353, + 555, + 555, + 569 + ], + "type": "interline_equation", + "content": "g \\left(x _ {i}; f, \\mathcal {A}, \\Theta\\right) = \\sigma \\left(\\theta^ {T} \\mathbf {A} \\left(f, \\mathcal {A}, x _ {i}\\right)\\right) \\tag {5}", + "image_path": "f03ea547f8161fd245552f74e490e232c054442112ed9da2d9e369899fa373b1.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " uses " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " to map input " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " to a " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "M\\times K" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " matrix of predicted logits where " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " is the number of augmentations and " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " is the number of classes. " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "1\\times m" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " vector corresponding to augmentation-specific weights. Each row in " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(f,\\mathcal{A},x_i)" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " represents the pre-trained classifier's predicted logits on augmentation " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "a_{m}" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": ". TTA-Learned refers to TTA combined with learned augmentation weights, while TTA-Avg refers to a simple average over the augmentations." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "text", + "content": "1We found no significant improvement by using alternate losses considered in the conformal prediction literature (e.g. focal or conformal training loss). See Table S7 in the Appendix." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20624" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "We refer to the fraction of the validation set allotted to " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "D^{(TTA)}" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ". Figure S3 shows that performance is not sensitive to the choice of " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "; as a result, all experiments use " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\beta = .2" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " (see supplement for further discussion). This does reduce the amount of data available to identify the appropriate threshold, but we find that the benefits TTA confers outweigh the cost to threshold estimation. Computational cost scales linearly with the size of " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "; each additional augmentation translates to a forward pass of the base classifier. One can use the learned weights to save computation by identifying which test-time augmentations to generate." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 220, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 376 + ], + "type": "text", + "content": "Preserving exchangeability The validity of conformal prediction depends upon the assumption of exchangeability: that all orderings of examples are equally likely (in effect, meaning that the distribution of examples in the calibration set is indistinguishable from the distribution of unseen examples). Prior work has shown that the conformal procedure is valid under deterministic transformations [24]; by using distinct examples to learn the test-time augmentation policy, the proposed approach constitutes a deterministic transformation applied to both the calibration set and unseen examples. If we were to instead use the same examples to learn the test-time augmentation policy and the conformal threshold, exchangeability could be broken." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 387, + 179, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 179, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 179, + 401 + ], + "type": "text", + "content": "5. Experimental Set-Up" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 503 + ], + "type": "text", + "content": "Datasets We show results on the test splits of three widely used image classification datasets: ImageNet [12] (50,000 natural images spanning 1,000 classes), iNaturalist [41] (100,000 images spanning 10,000 species), and CUB-Birds [45] (5,794 images spanning 200 categories of birds). Images are distributed evenly over classes in ImageNet and iNaturalist, while CUB-Birds has between 11 and 30 images per class." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 518, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 650 + ], + "type": "text", + "content": "Models The default model architecture, across all datasets, is ResNet-50 [17]. The accuracies of the base classifiers are " + }, + { + "bbox": [ + 55, + 518, + 295, + 650 + ], + "type": "inline_equation", + "content": "76.1\\%" + }, + { + "bbox": [ + 55, + 518, + 295, + 650 + ], + "type": "text", + "content": " (ImageNet), " + }, + { + "bbox": [ + 55, + 518, + 295, + 650 + ], + "type": "inline_equation", + "content": "76.4\\%" + }, + { + "bbox": [ + 55, + 518, + 295, + 650 + ], + "type": "text", + "content": " (iNaturalist), and " + }, + { + "bbox": [ + 55, + 518, + 295, + 650 + ], + "type": "inline_equation", + "content": "80.5\\%" + }, + { + "bbox": [ + 55, + 518, + 295, + 650 + ], + "type": "text", + "content": " (CUB-Birds). To study the relationship between model complexity and performance, we also provide results using ResNet-101 and ResNet-152 on ImageNet. For ImageNet, we make use of the pretrained models made available by PyTorch [31]. For iNaturalist, we use a model made public by Niers, Tom [30]. For CUB-Birds, we train a network by finetuning the final layer of a ResNet-50 model initialized with ImageNet's pretrained weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": "Augmentations We consider two augmentation policies. The first (the simple augmentation policy) consists of a random-crop and a horizontal-flip; to produce a random crop, we pad the original image with 4 pixels and take a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "256x256 crop of the expanded image (thereby preserving the original image resolution). The simple augmentation policy is widely used because its augmentations are likely to be label-preserving. The second, which we refer to as the expanded augmentation policy, consists of 12 augmentations: increase-sharpness, decrease-sharpness, autocontrast, invert, blur, posterize, shear, translate, color-jitter, randomCrop, horizontal-flip, and random-rotation. The supplement contains a description of each augmentation. These augmentations are not always label preserving, but, as we show, can improve performance when weights are learned." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 220, + 553, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 387 + ], + "type": "text", + "content": "Baselines We benchmark results using two conformal scores (translating to different definitions of " + }, + { + "bbox": [ + 313, + 220, + 553, + 387 + ], + "type": "inline_equation", + "content": "c(x,y)" + }, + { + "bbox": [ + 313, + 220, + 553, + 387 + ], + "type": "text", + "content": " in Equation 4). The first score is APS [34] (described in Eqn. 4), which represents the cumulative probability required to include the correct class, and the second is RAPS [1], which modifies APS by adding a term to penalize large set sizes. For all experiments, we do not allow sets of size 0. We implement RAPS and APS using code provided by Angelopoulos et al. [1], and automatically select hyperparameters " + }, + { + "bbox": [ + 313, + 220, + 553, + 387 + ], + "type": "inline_equation", + "content": "k_{reg}" + }, + { + "bbox": [ + 313, + 220, + 553, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 220, + 553, + 387 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 220, + 553, + 387 + ], + "type": "text", + "content": " to minimize set size. We also compare against conformal prediction using a simple average over the test-time augmentations (TTA-Avg). In the supplement, we also compare against non-conformal Top-1 and Top-5 prediction sets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 403, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 582 + ], + "type": "text", + "content": "Evaluation We evaluate results using the three metrics commonly used in the conformal prediction literature: efficiency, coverage, and adaptivity. We quantify efficiency using both average prediction set size (measured across all examples) and class-conditional prediction set size (measured across all examples in a class). Coverage is the percentage of sets containing the true label. We define adaptivity as the size-stratified coverage violation (SSCV), introduced by Angelopoulos et al. [1]. We first partition examples based upon the size of the prediction set. We create bins for set sizes of " + }, + { + "bbox": [ + 313, + 403, + 553, + 582 + ], + "type": "inline_equation", + "content": "[0, 1], [2, 3], [4, 10], [11, 100]," + }, + { + "bbox": [ + 313, + 403, + 553, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 403, + 553, + 582 + ], + "type": "inline_equation", + "content": "[101,]" + }, + { + "bbox": [ + 313, + 403, + 553, + 582 + ], + "type": "text", + "content": ". We then compute the empirical coverage within each bin, and compute adaptivity as the maximum difference between theoretical coverage and empirical coverage across bins. The closer this value is to 0, the better the adaptivity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "For each dataset, we report results across 10 randomly generated splits into validation and test sets. For all experiments (save for the validation set size experiment), the validation set and test set are the same size. We allot " + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": " of examples from the validation set to " + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "inline_equation", + "content": "D^{(TTA)}" + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": " (used to learn TTA policy), and allot the remaining examples to the calibration set. For the experiment studying validation set size, we downsample the validation set. We compute statistical significance using a paired t-test, with a Bonferroni correction [46] for multiple hypothesis testing. Code to reproduce all experiments will be made publicly available." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20625" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 553, + 228 + ], + "type": "table", + "html": "
Expanded Augmentation PolicySimple Augmentation Policy
AlphaMethodImageNetiNaturalistCUB-BirdsImageNetiNaturalistCUB-Birds
0.01RAPS37.751 ± 2.33461.437 ± 6.06715.293 ± 2.07137.751 ± 2.33461.437 ± 6.06715.293 ± 2.071
0.01RAPS+TTA-Avg35.600 ± 2.20057.073 ± 5.91413.111 ± 2.47031.681 ± 3.05754.169 ± 6.31914.550 ± 1.425
0.01RAPS+TTA-Learned31.248 ± 2.17753.195 ± 4.88414.045 ± 1.32332.702 ± 2.40951.391 ± 5.21113.803 ± 1.734
0.05RAPS5.637 ± 0.3577.991 ± 1.5213.624 ± 0.3615.637 ± 0.3577.991 ± 1.5213.624 ± 0.361
0.05RAPS+TTA-Avg5.318 ± 0.1137.067 ± 0.3443.116 ± 0.2104.908 ± 0.0996.451 ± 0.2793.249 ± 0.307
0.05RAPS+TTA-Learned4.889 ± 0.1686.682 ± 0.4473.571 ± 0.5765.040 ± 0.1766.788 ± 0.4963.290 ± 0.186
0.10RAPS2.548 ± 0.0742.914 ± 0.1162.038 ± 0.1532.548 ± 0.0742.914 ± 0.1162.038 ± 0.153
0.10RAPS+TTA-Avg2.470 ± 0.0712.740 ± 0.0261.780 ± 0.1392.327 ± 0.0862.610 ± 0.0311.881 ± 0.118
0.10RAPS+TTA-Learned2.312 ± 0.0542.625 ± 0.0431.893 ± 0.1872.362 ± 0.0652.638 ± 0.0261.840 ± 0.106
", + "image_path": "bfd98d0d138acf841796fb8226ec4f9883ef850f9b3fd70b13cd095411279428.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 281 + ], + "type": "text", + "content": "Table 1. Reductions in prediction set size across datasets, augmentation policies, and coverage specifications. Each entry corresponds to the average prediction set size across 10 calibration/test splits. Bolded entries represent performance that is either (a) significantly better compared to the baseline (RAPS), or (b) indistinguishable from the best approach. Table S8 reports achieved coverage. Corresponding results for APS can be found in Table S2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 62, + 293, + 220, + 399 + ], + "blocks": [ + { + "bbox": [ + 62, + 293, + 220, + 399 + ], + "lines": [ + { + "bbox": [ + 62, + 293, + 220, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 220, + 399 + ], + "type": "image", + "image_path": "9044f56e5d35f56b16f36f9295155b24bb8c1d88a49bd59df101cb6f67adc007.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 420, + 555, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 420, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 555, + 477 + ], + "type": "text", + "content": "Figure 2. Robustness to distribution shift. We compare average prediction set size achieved by RAPS (yellow) to average prediction set size achieved when combining RAPS with TTA-Learned (green). Results reflect the distribution of average prediction set size across 10 runs using ImageNet and ResNet50. We evaluate performance on different corruptions (x-axis) and different coverage guarantees (left, middle, right). RAPS+TTA-Learned (green) produces a noticeable reduction in prediction set size, even when subject to distribution shift, with no loss in coverage. Refer to Figure S1 in the supplement for a comparison of coverage achieved by both methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 293, + 382, + 399 + ], + "blocks": [ + { + "bbox": [ + 223, + 293, + 382, + 399 + ], + "lines": [ + { + "bbox": [ + 223, + 293, + 382, + 399 + ], + "spans": [ + { + "bbox": [ + 223, + 293, + 382, + 399 + ], + "type": "image", + "image_path": "cfe7c8bd71deb6b161a7c1096d915b736bea184355749c8b52408f327bd35047.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 383, + 294, + 543, + 399 + ], + "blocks": [ + { + "bbox": [ + 383, + 294, + 543, + 399 + ], + "lines": [ + { + "bbox": [ + 383, + 294, + 543, + 399 + ], + "spans": [ + { + "bbox": [ + 383, + 294, + 543, + 399 + ], + "type": "image", + "image_path": "0df7a187b18fc146b681a1d817e79f7f1999f0fee26d4e746dd068d2ab6e1630.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 496, + 108, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 108, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 108, + 509 + ], + "type": "text", + "content": "6. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": "We compare against RAPS, which outperformed other baselines in every experiment. (We provide results comparing our method to APS and the Top-K baselines in the supplement. Variants of each experiment across multiple " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": " and datasets are also in the supplement.) We then examine the dependence of these results on dataset, base model, and class. We conclude by providing intuition about why test-time augmentation improves the efficiency of conformal predictors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 229, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 229, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 229, + 647 + ], + "type": "text", + "content": "6.1. Reductions in prediction set size" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "We begin with results in the context of the expanded augmentation policy. Learned test-time augmentation policies produce meaningfully significant reductions in prediction set size (RAPS+TTA-Learned in Table 1 and APS+TTA-Learned in Table S1). TTA-Learned reduces prediction set" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 497, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 544 + ], + "type": "text", + "content": "sizes significantly in 16 of the 18 cases, and performs comparably in the remaining 2. Across all cases, the combination of RAPS, TTA-Learned, and the expanded augmentation policy produces the smallest average set sizes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": "TTA-Learned performs comparably or better than TTA-Avg in all comparisons. Certain augmentations in the expanded augmentation policy (blur, decrease sharpness, and shear) are consistently assigned a weight of 0, while certain augmentations are consistently included in learned policies (autocontrast, translate). Augmentations assigned zero weight provide no additional information about the true label (for example, they may not preserve the label within the image, or they may be redundant with other augmentations included in the policy)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "While TTA improves both RAPS and APS, the improvements are larger for APS. This is because TTA, like RAPS, tempers the predicted probabilities. TTA lowers the maximum predicted probability on average, thereby reducing" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20626" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 553, + 234 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 553, + 234 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 553, + 234 + ], + "type": "table", + "html": "
Expanded Augmentation PolicySimple Augmentation Policy
AlphaMethodResNet-50ResNet-101ResNet-152ResNet-50ResNet-101ResNet-152
0.01RAPS37.751 ± 2.33433.624 ± 1.79629.560 ± 3.48137.751 ± 2.33433.624 ± 1.79629.560 ± 3.481
0.01RAPS+TTA-Avg35.600 ± 2.20030.220 ± 1.77427.203 ± 2.52631.681 ± 3.05727.206 ± 1.84024.106 ± 2.100
0.01RAPS+TTA-Learned31.248 ± 2.17725.722 ± 1.71323.615 ± 1.65632.702 ± 2.40926.760 ± 1.97424.765 ± 2.736
0.05RAPS5.637 ± 0.3574.785 ± 0.1024.376 ± 0.0785.637 ± 0.3574.785 ± 0.1024.376 ± 0.078
0.05RAPS+TTA-Avg5.318 ± 0.1134.433 ± 0.1374.163 ± 0.1854.908 ± 0.0994.147 ± 0.1223.868 ± 0.126
0.05RAPS+TTA-Learned4.889 ± 0.1684.200 ± 0.2003.824 ± 0.1285.040 ± 0.1764.194 ± 0.1943.916 ± 0.356
0.10RAPS2.548 ± 0.0742.267 ± 0.0242.109 ± 0.0272.548 ± 0.0742.267 ± 0.0242.109 ± 0.027
0.10RAPS+TTA-Avg2.470 ± 0.0712.164 ± 0.0312.049 ± 0.0282.327 ± 0.0862.093 ± 0.0351.996 ± 0.018
0.10RAPS+TTA-Learned2.312 ± 0.0542.099 ± 0.0401.993 ± 0.0262.362 ± 0.0652.091 ± 0.0411.988 ± 0.020
", + "image_path": "06289058b6d00b03a941ce02b8c9c78abbb4f75abf59b141bb6e10c65a45333f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 241, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 555, + 275 + ], + "type": "text", + "content": "Table 2. Reductions in prediction set size across base classifiers on ImageNet. TTA-Learned can bridge the performance gap between different classifiers (for example, outperforming ResNet-152 alone when combined with ResNet-101), and yields significant reductions in set size regardless of the pretrained classifier used. We report achieved coverage in Table S9." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 295, + 294, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 294, + 356 + ], + "type": "text", + "content": "model overconfidence. Consequently, the predicted probability assigned to the remaining classes is higher. This is why the expanded augmentation policy demonstrates stronger performance than the simple augmentation policy: it tempers the probabilities to a greater extent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 421 + ], + "type": "text", + "content": "TTA-Learned preserves coverage across all experiments, since it respects the assumption of exchangeability. In some cases, TTA significantly improves coverage, although the magnitude of this difference is small (results can be found in Tables S8 and S9)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 475 + ], + "type": "text", + "content": "We next evaluate adaptivity using size stratified coverage violation (SSCV). At low alpha (" + }, + { + "bbox": [ + 55, + 426, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\alpha = .01" + }, + { + "bbox": [ + 55, + 426, + 295, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 426, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\alpha = .05" + }, + { + "bbox": [ + 55, + 426, + 295, + 475 + ], + "type": "text", + "content": "), TTA-Learned improves efficiency without diminishing adaptivity (Table S10)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 224, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 224, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 224, + 510 + ], + "type": "text", + "content": "6.2. Robustness to distribution shift" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Next, we evaluate the performance of test-time augmented conformal prediction on out-of-distribution examples. While conformal prediction does not guarantee coverage in these settings, distribution shifts are ubiquitous in practice [22]. Empirical performance is thus of practical interest. The training procedures for both test-time augmented conformal prediction and conformal prediction remain the same and use in-distribution examples from ImageNet and a ResNet50 classifier. We evaluate each conformal predictor on four types of image corruptions drawn from ImageNet-C [19]. Figure 2 plots the results; across all corruptions and coverage guarantees, test-time augmented conformal prediction produces smaller prediction sets than conformal prediction alone. Importantly, test-time augmented conformal prediction achieves this with no loss of coverage (Figure S1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 295, + 542, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 542, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 542, + 308 + ], + "type": "text", + "content": "6.3. Datasets, augmentation policies, and models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 409 + ], + "type": "text", + "content": "Dependence on dataset TTA consistently improves prediction set sizes on ImageNet and iNaturalist, but not on CUB-Birds. This may be because the validation set size for CUB-Birds (2,827 images) is an order of magnitude smaller than the validation sets for ImageNet (25,000 images) and iNaturalist (50,000 images). This is consistent with our finding that effectiveness of TTA is positively correlated with the size of the validation set (Figure S4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 423, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 616 + ], + "type": "text", + "content": "Dependence on augmentation policy We find that the expanded augmentation policy produces larger reductions in set size than the simple augmentation policy, primarily at low " + }, + { + "bbox": [ + 313, + 423, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 423, + 555, + 616 + ], + "type": "text", + "content": ". This is in spite of the fact that the expanded augmentation policy contains many augmentations outside of the base classifier's training-time augmentation policy. When we vary the number of augmentations included in an augmentation policy, we see that larger augmentation policies also yield greater reductions in average prediction set size (Figure S2). That said, the simple augmentation policy does have its place; it requires fewer forward passes during inference. In the absence of a learned aggregation function, our results suggest that aggregating using an average can still improve the efficiency of conformal predictors (outperforming the original conformal score in 11 comparisons, matching performance in 3, and under-performing in 4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "Dependence on base model We tested the generalizability of our results to other models by rerunning the ImageNet experiments using ResNet-101 (accuracy of " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "77.4\\%" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": ") and ResNet-152 (accuracy of " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "78.3\\%" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "). Unsurprisingly, more accurate models result in smaller prediction set sizes (Table 2). TTA variants of conformal prediction again produce significant improvements in set size while maintaining coverage." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20627" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 95, + 225, + 183 + ], + "blocks": [ + { + "bbox": [ + 62, + 82, + 72, + 91 + ], + "lines": [ + { + "bbox": [ + 62, + 82, + 72, + 91 + ], + "spans": [ + { + "bbox": [ + 62, + 82, + 72, + 91 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 82, + 199, + 95 + ], + "lines": [ + { + "bbox": [ + 107, + 82, + 199, + 95 + ], + "spans": [ + { + "bbox": [ + 107, + 82, + 199, + 95 + ], + "type": "text", + "content": "Distribution of Set Sizes (ImageNet, " + }, + { + "bbox": [ + 107, + 82, + 199, + 95 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + }, + { + "bbox": [ + 107, + 82, + 199, + 95 + ], + "type": "text", + "content": " Expanded)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 95, + 225, + 183 + ], + "lines": [ + { + "bbox": [ + 60, + 95, + 225, + 183 + ], + "spans": [ + { + "bbox": [ + 60, + 95, + 225, + 183 + ], + "type": "image", + "image_path": "36e53bae95142b1447e11f48bc24ccc9f5bc24a5187ab1f95f57952547fffab4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 260 + ], + "type": "text", + "content": "Figure 3. (A) Class-conditional prediction set sizes. We plot the distribution of class-conditional prediction set sizes, for ImageNet and ResNet-50 with " + }, + { + "bbox": [ + 55, + 204, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\alpha = .01" + }, + { + "bbox": [ + 55, + 204, + 555, + 260 + ], + "type": "text", + "content": ". RAPS+TTA-Learned (green) produces a noticeable reduction in class-conditional prediction set sizes. (B, C) Relationship between TTA improvements and original class set sizes and class difficulty. TTA introduces the largest improvements for classes with the largest original prediction set sizes (B) and classes on which the underlying classifier is often incorrect (C). Each point represents the average prediction set size for each class, across 10 splits." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 95, + 385, + 182 + ], + "blocks": [ + { + "bbox": [ + 234, + 82, + 242, + 91 + ], + "lines": [ + { + "bbox": [ + 234, + 82, + 242, + 91 + ], + "spans": [ + { + "bbox": [ + 234, + 82, + 242, + 91 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 276, + 83, + 356, + 95 + ], + "lines": [ + { + "bbox": [ + 276, + 83, + 356, + 95 + ], + "spans": [ + { + "bbox": [ + 276, + 83, + 356, + 95 + ], + "type": "text", + "content": "Original Set Size " + }, + { + "bbox": [ + 276, + 83, + 356, + 95 + ], + "type": "inline_equation", + "content": "(r = 0.89, \\alpha = 0.01" + }, + { + "bbox": [ + 276, + 83, + 356, + 95 + ], + "type": "text", + "content": ", Expanded)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 95, + 385, + 182 + ], + "lines": [ + { + "bbox": [ + 233, + 95, + 385, + 182 + ], + "spans": [ + { + "bbox": [ + 233, + 95, + 385, + 182 + ], + "type": "image", + "image_path": "e43903664fe6477547f726dab51f0a6f4191f8e4d21eeead62eb480eb0d220a3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 392, + 95, + 544, + 182 + ], + "blocks": [ + { + "bbox": [ + 394, + 82, + 402, + 91 + ], + "lines": [ + { + "bbox": [ + 394, + 82, + 402, + 91 + ], + "spans": [ + { + "bbox": [ + 394, + 82, + 402, + 91 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 83, + 496, + 94 + ], + "lines": [ + { + "bbox": [ + 454, + 83, + 496, + 94 + ], + "spans": [ + { + "bbox": [ + 454, + 83, + 496, + 94 + ], + "type": "text", + "content": "Class Difficulty 55. " + }, + { + "bbox": [ + 454, + 83, + 496, + 94 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + }, + { + "bbox": [ + 454, + 83, + 496, + 94 + ], + "type": "text", + "content": " Exp" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 95, + 544, + 182 + ], + "lines": [ + { + "bbox": [ + 392, + 95, + 544, + 182 + ], + "spans": [ + { + "bbox": [ + 392, + 95, + 544, + 182 + ], + "type": "image", + "image_path": "7b69bdf3185435d4650bdd511a9d3e20de883f89c8593e692ae4801f00b89eac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 58, + 281, + 253, + 419 + ], + "blocks": [ + { + "bbox": [ + 58, + 281, + 253, + 419 + ], + "lines": [ + { + "bbox": [ + 58, + 281, + 253, + 419 + ], + "spans": [ + { + "bbox": [ + 58, + 281, + 253, + 419 + ], + "type": "image", + "image_path": "951c946d5ab97999c62824298e9d768559c84a83b72391c06eef90a190b2ecd0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 257, + 281, + 398, + 423 + ], + "blocks": [ + { + "bbox": [ + 257, + 281, + 398, + 423 + ], + "lines": [ + { + "bbox": [ + 257, + 281, + 398, + 423 + ], + "spans": [ + { + "bbox": [ + 257, + 281, + 398, + 423 + ], + "type": "image", + "image_path": "980d366e395e611caa4ee3f02f3450fe3432fd5bae5a4ea26f9f71c4a2ac276f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 399, + 281, + 547, + 425 + ], + "blocks": [ + { + "bbox": [ + 399, + 281, + 547, + 425 + ], + "lines": [ + { + "bbox": [ + 399, + 281, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 399, + 281, + 547, + 425 + ], + "type": "image", + "image_path": "d4d64b53b7a22eb991b69b832da77cc5c13578a9fae7f5cad97978d0a55960ee.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 446, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 54, + 446, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 446, + 555, + 514 + ], + "type": "text", + "content": "Figure 4. (A) Effect of TTA-Learned on optimal Top-K: TTA-Learned significantly lowers the value of k required for Top-k prediction sets to achieve coverage on ImageNet and iNaturalist, but not on CUB-Birds. (B,C) Effect of TTA-Learned on rank of true class: TTA-Learned improves the rank of the true class among the sorted predicted probabilities for a given example for both ImageNet (B) and iNaturalist (C). We plot the rank using the original predicted probabilities compared to the TTA-transformed probabilities, binning all examples in the validation set into five equal-width bins. Dots that fall below the red line indicate that TTA improves the rank of the true class." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 534, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 534, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 54, + 534, + 295, + 606 + ], + "type": "text", + "content": "age. We note that the combination of TTA with ResNet-101 produces smaller set sizes than the more complex ResNet-152 alone. For example, when " + }, + { + "bbox": [ + 54, + 534, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 534, + 295, + 606 + ], + "type": "text", + "content": " is set to .01, RAPS+TTA-Learned and ResNet-101 produce set sizes that contain, on average, 26.5 classes, while RAPS and ResNet-152 produce an average set size of 29.6." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 612, + 186, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 186, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 186, + 625 + ], + "type": "text", + "content": "6.4. Class-Specific Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "type": "text", + "content": "We have established that on average TTA is a useful addition to the conformal pipeline. We now investigate the source of this improvement. We make two empirical observations. First, classes with larger predicted set sizes benefit most from the introduction of TTA. Figure 3 shows that a class's average prediction set size is significantly correlated with the change in set size produced by TTA-Learned (with" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "text", + "content": "the expanded augmentation policy and " + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\alpha = .01" + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "inline_equation", + "content": "r = 0.89" + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "inline_equation", + "content": "p < 1e - 10" + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "text", + "content": ". Second, we find that class difficulty is significantly associated with changes in set size introduced by TTA (with the expanded augmentation policy and " + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\alpha = .01" + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "inline_equation", + "content": "r = 0.55" + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "inline_equation", + "content": "p < 1e - 10" + }, + { + "bbox": [ + 313, + 534, + 555, + 643 + ], + "type": "text", + "content": "). Prediction sets for classes that are difficult to predict benefit more from TTA compared to their easier counterparts. These observations are related; harder classes receive larger set sizes, and consequently, offer larger room for improvements in efficiency." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 648, + 378, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 378, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 378, + 659 + ], + "type": "text", + "content": "6.5. Intuition" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Why does the addition of test-time augmentation produce smaller prediction set sizes? In short, TTA improves top-K accuracy. We verify this claim by estimating " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " such that the uncertainty sets comprised of the top " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " predicted" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20628" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "classes for each example achieve a marginal coverage of " + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "inline_equation", + "content": "(1 - \\alpha)" + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": ". We see that for those datasets (ImageNet and iNaturalist) where TTA produces significant reductions in set size, TTA-transformed predictions—both with a simple average and learned weights—produce significantly lower values for " + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": " compared to the original predictions (Figure 4A). This is not true for CUB-Birds, on which TTA offers little to no improvement. One could use such a procedure to determine whether TTA is worth adding to a conformal pipeline without collecting labeled data beyond the calibration dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 202, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 202, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 202, + 296, + 477 + ], + "type": "text", + "content": "Another way to understand the impact of TTA is to consider the effect on the ordering of classes. It has been observed in the test-time augmentation literature that TTA often promotes the true class from the second-highest to the highest predicted probability, thereby correcting the classification. This finding does not fully explain the value of test-time augmentation to conformal prediction since only " + }, + { + "bbox": [ + 56, + 202, + 296, + 477 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 56, + 202, + 296, + 477 + ], + "type": "text", + "content": " of the prediction sets which reduce in size are associated with a corrected Top-1 classification. What TTA did do was increase the predicted probability of the true class even when it is predicted to be unlikely (for example, promoting the true class from 200th most likely to 100th most likely). We visualize this effect in Figure 4 by plotting the change in true class rank (the index at which the true class appears in the sorted list of predicted probabilities) for all examples in the validation set, stratified into 5 equal-width bins. The lower left point captures examples that are classified correctly; here, test-time augmentation introduces little to no change. In subsequent bins, we see that TTA typically promotes the rank of the true class. We also include the standard deviation across the true class ranks in the original predicted probabilities (x-axis) and the TTA-transformed probabilities (y-axis)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 516, + 130, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 130, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 130, + 529 + ], + "type": "text", + "content": "7. Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "Learned test-time augmentation policies require two ingredients: labeled data and multiple forward passes. Although one can minimize costs by parallelizing computation or by using the learned weights to identify which augmentations to generate, inference will always cost more with test-time augmentation. Our results are also limited to image classification. We do not consider other modalities, for which appropriate transformations will substantially differ. Future work should consider how these results may generalize to non-vision tasks. Finally, test-time augmentation is one approach to generating ensembles in conformal prediction. Many other more computationally expensive approaches exist. The tradeoff between computation and ensemble performance remains a useful avenue for future work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 388, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 388, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 388, + 84 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 332 + ], + "type": "text", + "content": "We show that test-time augmented conformal prediction produces smaller sets than conformal prediction alone. The proposed approach is effective, efficient, and simple: it reduces prediction set sizes by up to " + }, + { + "bbox": [ + 313, + 91, + 555, + 332 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 91, + 555, + 332 + ], + "type": "text", + "content": ", requires no model re-training, and relies on a portion of labeled data already available to split conformal predictors. Our experiments also indicate that test-time augmented conformal prediction exhibits greater efficiency under four common corruption-based distribution shifts. Test-time augmentation is able to do this by improving the underlying classifier's robustness to domain-specific invariances, in the form of data augmentation. Efforts to improve the efficiency of conformal predictors could, as a first step, aim to improve the robustness of the underlying classifier. The performance of TTA-Learned also suggests that there are settings in which it is wise to use a portion of the labeled data to improve the underlying model, instead of reserving all labeled data for the calibration set. In sum, our work takes a step towards practically useful conformal predictors by improving efficiency, without sacrificing adaptivity or coverage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 342, + 373, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 342, + 373, + 355 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 373, + 355 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 319, + 362, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 319, + 362, + 555, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 362, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 319, + 362, + 555, + 407 + ], + "type": "text", + "content": "[1] Anastasios Angelopoulos, Stephen Bates, Jitendra Malik, and Michael I. Jordan. Uncertainty Sets for Image Classifiers using Conformal Prediction, 2022. arXiv:2009.14193 [cs, math, stat]. 2, 4, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 408, + 554, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 408, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 319, + 408, + 554, + 441 + ], + "type": "text", + "content": "[2] Murat Seçkin Ayhan and Philipp Berens. Test-time Data Augmentation for Estimation of Heteroscedastic Aleatoric Uncertainty in Deep Neural Networks. 2018. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 443, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 443, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 320, + 443, + 553, + 475 + ], + "type": "text", + "content": "[3] Varun Babbar, Umang Bhatt, and Adrian Weller. On the Utility of Prediction Sets in Human-AI Teams, 2022. arXiv:2205.01411 [cs]. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 477, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 477, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 320, + 477, + 553, + 510 + ], + "type": "text", + "content": "[4] Anthony Bellotti. Optimized conformal classification using gradient descent approximation, 2021. arXiv:2105.11255 [cs]. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 512, + 553, + 555 + ], + "type": "text", + "content": "[5] Lars Carlsson, Martin Eklund, and Ulf Norinder. Aggregated Conformal Prediction. In Artificial Intelligence Applications and Innovations, pages 231-240, Berlin, Heidelberg, 2014. Springer. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 320, + 557, + 553, + 610 + ], + "type": "text", + "content": "[6] Sewhan Chun, Jae Young Lee, and Junmo Kim. Cyclic test time augmentation with entropy weight method. In Proceedings of the Thirty-Eighth Conference on Uncertainty in Artificial Intelligence, pages 433-442. PMLR, 2022. ISSN: 2640-3498. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 613, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 613, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 320, + 613, + 553, + 657 + ], + "type": "text", + "content": "[7] Gilad Cohen and Raja Giryes. Simple post-training robustness using test time augmentations and random forest. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3996-4006, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "text", + "content": "[8] Seffi Cohen, Noa Dagan, Nurit Cohen-Inger, Dan Ofer, and Lior Rokach. ICU Survival Prediction Incorporating Test-Time Augmentation to Improve the Accuracy of Ensemble-Based Models. IEEE Access, 9:91584-91592, 2021. Conference Name: IEEE Access. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20629" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[9] Pedro Conde, Tiago Barros, Rui L. Lopes, Cristiano Premebida, and Urbano J. Nunes. Approaching Test Time Augmentation in the Context of Uncertainty Calibration for Deep Neural Networks, 2023. arXiv:2304.05104 [cs]. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "text", + "content": "[10] Jesse C Cresswell, Yi Sui, Bhargava Kumar, and Noel Vouitsis. Conformal prediction sets improve human decision making. arXiv preprint arXiv:2401.13744, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[11] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. AutoAugment: Learning Augmentation Policies from Data, 2019. arXiv:1805.09501 [cs, stat]. 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. ISSN: 1063-6919. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[13] Tiffany Ding, Anastasios N. Angelopoulos, Stephen Bates, Michael I. Jordan, and Ryan J. Tibshirani. Class-Conditional Conformal Prediction With Many Classes, 2023. arXiv:2306.09335 [cs, stat]. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "text", + "content": "[14] Bat-Sheva Einbinder, Yaniv Romano, Matteo Sesia, and Yanfei Zhou. Training Uncertainty-Aware Classifiers with Conformalized Deep Learning. 15" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "text", + "content": "[15] Shohei Enomoto, Monikka Roslianna Busto, and Takeharu Eda. Dynamic Test-Time Augmentation via Differentiable Functions, 2023. arXiv:2212.04681 [cs]. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 367, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 419 + ], + "type": "text", + "content": "[16] A. Gammerman, V. Vovk, and V. Vapnik. Learning by transduction. In Proceedings of the Fourteenth conference on Uncertainty in artificial intelligence, pages 148-155, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers Inc. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, Las Vegas, NV, USA, 2016. IEEE. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[18] Achim Hekler, Titus J. Brinker, and Florian Buettner. Test Time Augmentation Meets Post-hoc Calibration: Uncertainty Quantification under Real-World Conditions. Proceedings of the AAAI Conference on Artificial Intelligence, 37(12):14856-14864, 2023. Number: 12. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "text", + "content": "[19] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 569, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 634 + ], + "type": "text", + "content": "[20] Vilde Jensen, Filippo Maria Bianchi, and Stian Normann Anfinsen. Ensemble Conformalized Quantile Regression for Probabilistic Time Series Forecasting. IEEE Transactions on Neural Networks and Learning Systems, pages 1-12, 2022. Conference Name: IEEE Transactions on Neural Networks and Learning Systems. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "type": "text", + "content": "[21] Ildoo Kim, Younghoon Kim, and Sungwoong Kim. Learning Loss for Test-Time Augmentation. In Advances in Neural Information Processing Systems, pages 4163-4174. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[22] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "text", + "content": "[23] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2012. 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 182 + ], + "type": "text", + "content": "[24] Arun Kumar Kuchibhotla. Exchangeability, conformal prediction, and rank tests. arXiv preprint arXiv:2005.06095, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "text", + "content": "[25] Henrik Linusson, Ulf Norinder, Henrik Bostrom, Ulf Johansson, and Tuve Lofstrom. On the Calibration of Aggregated Conformal Predictors. In Proceedings of the Sixth Workshop on Conformal and Probabilistic Prediction and Applications, pages 154-173. PMLR, 2017. ISSN: 2640-3498. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "text", + "content": "[26] H. Linusson, U. Johansson, and H. Bostrom. Efficient conformal predictor ensembles. Neurocomputing, 397:266-278, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "text", + "content": "[27] Helen Lu, Divya Shanmugam, Harini Suresh, and John Guttag. Improved Text Classification via Test-Time Augmentation, 2022. arXiv:2206.13607 [cs]. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 306, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 369 + ], + "type": "text", + "content": "[28] Alexander Lyzhov, Yuliya Molchanova, Arsenii Ashukha, Dmitry Molchanov, and Dmitry Vetrov. Greedy Policy Search: A Simple Baseline for Learnable Test-Time Augmentation. In Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI), pages 1308-1317. PMLR, 2020. ISSN: 2640-3498. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[29] Luca Mocerino, Roberto G. Rizzo, Valentino Peluso, Andrea Calimera, and Enrico Macii. Adaptive Test-Time Augmentation for Low-Power CPU, 2021. arXiv:2105.06183 [cs, eess]. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 436 + ], + "type": "text", + "content": "[30] Niers, Tom. iNaturalist_competition, 2021. original-date: 2021-12-10T10:56:46Z. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 535 + ], + "type": "text", + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 601 + ], + "type": "text", + "content": "[32] Juan C. Perez, Motasem Alfarra, Guillaume Jeanneret, Laura Rueda, Ali Thabet, Bernard Ghanem, and Pablo Arbelaez. Enhancing Adversarial Robustness via Test-time Transformation Ensembling. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 81-91, Montreal, BC, Canada, 2021. IEEE. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "text", + "content": "[33] Drew Prinster, Anqi Liu, and Suchi Saria. JAWS: Auditing Predictive Uncertainty Under Covariate Shift, 2022. arXiv:2207.10716 [cs, stat]. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 668 + ], + "type": "text", + "content": "[34] Yaniv Romano, Matteo Sesia, and Emmanuel J. Candès. Classification with Valid and Adaptive Coverage, 2020. arXiv:2006.02544 [stat]. 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 690 + ], + "type": "text", + "content": "[35] C Saunders and Royal Holloway. Transduction with Confidence and Credibility. 1999. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "text", + "content": "[36] Glenn Shafer and Vladimir Vovk. A Tutorial on Conformal Prediction. 2008. 1, 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20630" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 522 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[37] Divya Shanmugam, Davis Blalock, Guha Balakrishnan, and John Guttag. Better Aggregation in Test-Time Augmentation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1194–1203, Montreal, QC, Canada, 2021. IEEE. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 162 + ], + "type": "text", + "content": "[38] David Stutz, Krishnamurthy, Dvijotham, Ali Taylan Cemgil, and Arnaud Doucet. Learning Optimal Conformal Classifiers, 2022. arXiv:2110.09192 [cs, stat]. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "text", + "content": "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[40] Ryan J Tibshirani, Rina Foygel Barber, Emmanuel Candes, and Aaditya Ramdas. Conformal Prediction Under Covariate Shift. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[41] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha. Benchmarking Representation Learning for Natural World Image Collections, 2021. arXiv:2103.16483 [cs]. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "type": "text", + "content": "[42] Vladimir Vovk. Algorithmic Learning in a Random World. Springer-Verlag, New York, 2005. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 342 + ], + "type": "text", + "content": "[43] Vladimir Vovk. Cross-conformal predictors, 2012. arXiv:1208.0806 [cs, stat]. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 374 + ], + "type": "text", + "content": "[44] Vladimir Vovk. Cross-conformal predictors. Annals of Mathematics and Artificial Intelligence, 74(1-2):9-28, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "type": "text", + "content": "[45] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "text", + "content": "[46] Eric W. Weisstein. Bonferroni correction. https://mathworld.wolfram.com/, 2004. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 434, + 296, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 488 + ], + "type": "text", + "content": "[47] Dongping Zhang, Angelos Chatzimparmpas, Negar Kamali, and Jessica Hullman. Evaluating the utility of conformal prediction sets for ai-advised image labeling. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1-19, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "text", + "content": "[48] Marvin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test Time Robustness via Adaptation and Augmentation, 2022. arXiv:2110.09506 [cs]. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20631" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_content_list.json b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ae7b00889695f4521af4f6fd49b18f90644f1b7a --- /dev/null +++ b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_content_list.json @@ -0,0 +1,1272 @@ +[ + { + "type": "text", + "text": "TexGarment: Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer", + "text_level": 1, + "bbox": [ + 186, + 130, + 816, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jialun Liu $^{1}$ , Jinbo Wu $^{1*}$ , Xiaobo Gao $^{1}$ , Jiakui Hu $^{2}$ , Bojun Xiong $^{2}$ , Xing Liu $^{1}$ , Chen Zhao $^{1}$ , Hongbin Pei $^{3}$ , Haocheng Feng $^{1}$ , Yingying Li $^{1}$ , Errui Ding $^{1}$ , Jingdong Wang $^{1}$ $^{1}$ Department of Computer Vision Technology(VIS), Baidu Inc. \n $^{2}$ Peking University $^{3}$ Xi'an Jiaotong University liujialun@baidu.com", + "bbox": [ + 143, + 202, + 851, + 294 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2364eb4ef4e2eb568925731de77cee22b3c46c7c7f6f891c65cc37a98f46552.jpg", + "image_caption": [ + "Figure 1. A gallery showcasing generated garment textures by TexGarment. Our method achieves globally consistent, high-quality, and high-fidelity textures across a wide range of garment styles. The average time to generate each texture map is under 4 seconds." + ], + "image_footnote": [], + "bbox": [ + 114, + 311, + 883, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 651, + 328, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces TexGarment, an efficient method for synthesizing high-quality, 3D-consistent garment textures in UV space. Traditional approaches based on 2D-to-3D mapping often suffer from 3D inconsistency, while methods learning from limited 3D data lack sufficient texture diversity. These limitations are particularly problematic in garment texture generation, where high demands exist for both detail and variety. To address these challenges, TexGarment leverages a pre-trained text-to-image diffusion Transformer model with robust generalization capabilities, introducing structural information to guide the model in generating 3D-consistent garment textures in a single inference step. Specifically, We utilize the 2D UV position map to", + "bbox": [ + 88, + 681, + 486, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guide the layout during the UV texture generation process, ensuring a coherent texture arrangement and enhancing it by integrating global 3D structural information from the mesh surface point cloud. This combined guidance effectively aligns 3D structural integrity with 2D layout. Our method efficiently generates high-quality, diverse UV textures in a single inference step while maintaining 3D consistency. Experimental results validate the effectiveness of TexGarment, achieving state-of-the-art performance in 3D garment texture generation.", + "bbox": [ + 511, + 652, + 910, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 829, + 645, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The demand for 3D garment visualization is growing across virtual try-ons for e-commerce, video games, 3D films, and AR/VR environments. These industries need large volumes", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Project lead and corresponding author.", + "bbox": [ + 107, + 887, + 321, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26566", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of high-quality 3D garments to meet consumer expectations. While 2D fashion images are widely available and generative AI has simplified 2D image creation, producing high-quality 3D garment textures remains challenging, often requiring skilled artists and weeks of work. This highlights the need for efficient 3D garment generation methods.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The aesthetic appeal of 3D garments is deeply influenced by their texture quality, as Joseph Albers [1] said: \"Texture is the soul of painting; it gives surface depth and personality\", effectively capturing this concept. In 3D computer graphics, textures are essential for realism, adding detailed surface elements that make virtual objects lifelike. For 3D garments, high-quality textures are particularly important, capturing fabric nuances, stitching, folds, and shadows that make virtual clothing look authentic. Without these textures, even well-modeled garments can appear flat or artificial, reducing the immersive quality of 3D applications.", + "bbox": [ + 89, + 185, + 480, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite substantial advancements in 2D image synthesis and 3D shape generation using GANs [40, 42] and diffusion models [12, 38, 39], achieving high-quality garment texture generation remains challenging. This difficulty arises from three primary factors: Global 3D Consistency. Methods employing 2D diffusion models, such as TexPainter [46], TEXTure [37] and Text2Tex [7], have made strides in texture synthesis by utilizing pre-trained LDM [38], which support strong generalization and diverse texture generation. However, text-to-image models often exhibit biases that limit their effectiveness for 3D applications, such as favoring frontal views and lacking a robust understanding of 3D structures. This frequently results in texture inconsistencies across different viewpoints of the same 3D object, ranging from minor seam misalignments to severe issues, e.g., \"Janus effect\". Generalization Ability. Methods learning from limited 3D data, such as TexOct [23], PointUV [44], and Mesh2Tex [4], generate textures by explicitly integrating 3D shape information, resulting in high 3D consistency. However, these approaches often lack generalization ability, limiting their effectiveness when generating textures for diverse 3D objects beyond the training set or adapting to a range of text and visual prompts. Inference Speed. To achieve full shape coverage, prior methods [2, 37, 44, 45] rely on iterative generation processes, requiring multiple steps or stages. This results in prolonged inference times, often lasting several minutes, and requires significant computational resources, limiting practical use for designers and content creators.", + "bbox": [ + 89, + 354, + 482, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present TexGarment, a method tailored for generating high-quality garment textures. Given text prompts, TexGarment produces consistent, diverse, and visually striking textures in a single inference pass. A widely used approach for handling texture maps is UV representation, which maps 3D textures accurately onto 2D UV images through UV coordinate transformations. This tech-", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nique bridges the gap between 3D texture generation and 2D image methods, motivating our use of pre-trained image generation models [8] to synthesize diverse textures effectively within the 2D UV space.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key insight is to incorporate 3D garment structure information as guidance within a pre-trained 2D image generation model [8], leveraging its generalization capabilities to produce consistent and diverse UV texture maps directly in the 2D UV space through only one inference process. Specifically, our method introduces two main components: 1) 3D Structure Information Extraction and 2) 3D Structure Guidance. First, given a garment mesh, we use rasterization to obtain a UV position map, where each pixel encodes the XYZ coordinates of the corresponding point on the mesh surface. This UV position map provides a 2D projection of the original 3D point cloud, capturing the basic layout of the UV texture. However, regions that are adjacent in 3D space may be separated in the UV map, causing a loss of global 3D structure information. This makes it challenging for a 2D image generation model to capture global 3D structures solely from the 2D UV position map. To address this, we introduce a 3D structure information extraction module that injects global structure information, represented by 3D point clouds, into the 2D UV position map. This process generates a 3D structure latent, which combines both 3D structure and 2D layout information, serving as guidance for the pre-trained image generation model. Second, we implement a 3D structure guidance module, analogous to text guidance, which injects the 3D structure information into the pre-trained model via cross-attention. Using a powerful text-to-image diffusion model [8] as our basic architecture, our design enables the model to receive guidance from both text and 3D structure, allowing it to generate high-quality UV texture maps in a single inference process, achieving both text fidelity and 3D consistency.", + "bbox": [ + 511, + 151, + 903, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: 1) We propose TexGarment, a one-stage method for garment texture generation that produces high-quality, diverse UV texture maps for untextured garment meshes in a single inference process and completes the process in 4 seconds. 2) We introduce two novel modules to incorporate 3D garment structure information into the 2D image generation model, providing 3D-consistent guidance for garment texture generation. Extensive experiments validate the effectiveness of these modules in achieving 3D consistency. 3) Qualitative and quantitative evaluations demonstrate that our method outperforms state-of-the-art approaches in garment texture generation. Additionally, a user study shows that users prefer textures generated by our method.", + "bbox": [ + 511, + 619, + 903, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 844, + 663, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image Generation The field of text-to-image generation has seen considerable progress. Early research used", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26567", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generative adversarial networks (GANs)[17] to create high-quality images, using image classes[15] and segmentation maps [30] as prompts. Recently, attention has shifted to architectural improvements that enhance generation quality with scaling Transformers, as shown in models like DALL-E [36], CogView [14], Parti [43], and Muse [6]. Another popular category of text-to-image generators is based on pixel space or latent diffusion models, including Imagegen [39], Stable Diffusion [38], SDXL [32], DiT [31], and PixArt [8]. Notably, DiT [31] advanced text-to-image generation by introducing Transformers to latent diffusion models (LDMs). PixArt [8] further optimized training methods, enabling faster training of scaled LDMs.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Generation via 2D generation model. With the development of text-to-image generation techniques, many 3D texturing techniques seek to leverage the prior knowledge of 2D generative models to improve 3D generation models' optimization. Previous optimization-based texture techniques [11, 20, 25, 27, 28] concentrated on improving the texture maps of 3D models by utilizing the large-scale visual language model, e.g. CLIP [34], before the advent of LDMs. DreamFusion [33] introduced the score distillation sampling (SDS) to 3D Generation, which laid the groundwork for many text-to-3D methods. Latent-nerf [26] and Fantasia3D [9] extended SDS to optimize texture mapping. Recently, TexPainter [46] performs multi-view fusion within the 2D image's color space to address inconsistency issues. However, optimization-based texture methods are hindered by prolonged training times. As an alternative, TEXTure [37], Text2Tex [7], and Paint3D [45] involve directly generating texture images from multiple views of 3D meshes and subsequently painting them onto the 3D meshes. TEXTure [37] and Text2Tex [7] adhere to this concept and suggest modifying the overlapping segments of the texture image from the previous view using results from each new camera pose. Nevertheless, these methods encounter various artifacts, including seams, noise, and irrelevant fragments. Paint3D [45] develops a specialized model to paint incomplete regions during the texture painting.", + "bbox": [ + 91, + 291, + 483, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Texturing from 3D Data. Additionally, generative texture models are capable of training models from scratch using 3D data. To assign a color to each pixel on the surface of a three-dimensional shape, early methods [16, 29] relied on implicit texture fields. However, given that the texture on the surface of a 3D mesh is continuous, it is unlikely that discrete supervision will result in the training of a model capable of synthesizing high-quality textures. The Texturify [41] is designed to construct a texture map on the surface of a polygonal mesh. Mesh2tex [4] builds on this by integrating an implicit texture field branch for enhancement. TexOct [23] use point clouds to help circumvent variations arising from diverse mesh topologies and UV mappings. Other methods [12, 44] synthesize UV", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "maps for 3D meshes, but are ill-equipped to handle more generic objects, due to the inherent diversity between 3D meshes belonging to various categories.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussion with previous work about the UV position map. Previous works, such as PointUV [44], Paint3D [45], and TextureGen [2], also incorporate the UV position map into their models. We outline the key differences between our approach and these methods: 1) PointUV [44] uses the UV position map, along with other maps like normal and mask maps, in the first stage to generate a coarse UV texture. However, this method is limited to simple textures for objects from the ShapeNet [5] dataset. In contrast, our approach uses only the UV texture map as a guide, enabling the generation of complex, high-quality textures for a wide range of garment categories. 2) Paint3D [45] and TextureGen [2] use the UV position map as a 2D condition in a second-stage inpainting model to refine textures. Our method directly applies the UV position map to guide the entire texture generation, leveraging both 2D layout and global 3D structure, resulting in better textures as shown in our experiments. 3) Unlike these prior methods, our model is a single-stage generator. Given a UV position map and a text prompt, our approach generates high-quality garment texture maps in a single, efficient inference step, while previous methods require a more complex two-stage process.", + "bbox": [ + 511, + 136, + 906, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 481, + 604, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a 3D garment model and a target texture description, TexGarment efficiently synthesizes high-quality, diverse, and consistent garment UV textures. TexGarment is a Transformer-based diffusion model, as shown in Figure 2. We build on a large-scale pre-trained image generation model [8], leveraging its robust prior knowledge to enhance our model's generalization capability. To ensure global 3D consistency, we first design a 3D structure information extraction module. For an untextured garment mesh, we apply rasterization to create a UV position map that captures the basic UV texture layout and encodes the 3D coordinates of points on the mesh surface. The structure information extraction module then embeds global structure information from the original point cloud into the UV position map, resulting in a 3D structure latent that combines global 3D structure information with 2D layout details. This 3D structure latent is then used by a 3D structure guidance module to guide the texture generation. Through this innovative design, our method generates high-quality and 3D-consistent texture maps in a single inference process, without the need for additional operations such as inpainting. The following sections provide the overview of each stage.", + "bbox": [ + 511, + 507, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Processing", + "text_level": 1, + "bbox": [ + 513, + 848, + 671, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset. We collect a comprehensive dataset from VRoidHub based on the guidelines provided by Panic3D[10], en", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26568", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c788b14c53382f84ae633e18e651147b78b88650ba3d474ea91ef4f2b28eb79f.jpg", + "image_caption": [ + "(a) Overview", + "(b) 3D Structure Information Extraction", + "Figure 2. Pipeline of Our Method: (a) Overview: The training process begins with embedding the input text prompt and timestep using the T5 model [35] and an MLP layer, generating the text feature and time embedding. The UV texture is encoded by a VAE [21] encoder to create a texture latent representation, which is then noised to produce a noisy latent. These representations are processed through several Transformer blocks. (b) 3D Structure Information Extraction Module: The UV position map is encoded by a VAE [21] encoder to obtain the UV position latent, while a point cloud feature extraction module captures the global 3D structure. This 3D information is integrated into the UV position latent via cross-attention, resulting in the 3D structure latent, which combines 3D global structure and 2D UV texture layout. (c) Transformer Block Details: Based on a text-to-image Transformer diffusion model [8], we introduce a 3D structure guidance module that uses the 3D structure latent as guidance, injecting it into the model through cross-attention to enhance 3d structure consistency." + ], + "image_footnote": [], + "bbox": [ + 114, + 89, + 578, + 404 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f8fe135bcacd9cd7da80f02aba3e789827379c51393a177ddc3ad323c8e81257.jpg", + "image_caption": [ + "(c) Block with 3D sturcture guidance" + ], + "image_footnote": [], + "bbox": [ + 588, + 90, + 874, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "compassing a diverse array of character models. Each model in VRoidHub includes body, clothing, and animation components. Using Blender, we segment the models by style type, preserving clothing elements while removing non-clothing components. Pants and tops were separated into distinct parts, while dresses were preserved as single units. Following segmentation, we refine the dataset by excluding samples with inconsistent color distributions and monochromatic appearances, as well as items with accessories such as ties, belts, and other extraneous parts, ultimately retaining a dataset of 8,000 clothing items.", + "bbox": [ + 88, + 554, + 482, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rasterization. In a UV texture image, the RGB information of the surface mesh is stored, and during rendering, a rasterization operator samples the color from the UV texture image onto the rendered image. A UV position map stores the XYZ coordinates of the mesh, which requires sampling mesh properties and baking them into the UV map. To achieve this, we assume the UV layout represents a 3D mesh with vertices $[u,v,z]$ , where $u$ and $v$ are the original mesh's UV coordinates and $z$ is constant, making it a planar mesh. Each $[u,v]$ corresponds to a vertex in the original 3D mesh. We retain the face relationships from", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the original mesh, and by performing rasterization, we interpolate the \"UV mesh\" positions to determine each pixel's XYZ coordinates on the UV position map. The shapes of garment meshes in the VRoid data are also defined by an alpha channel within the UV texture. Before rendering our images, we rasterize the UV position map using the alpha channel to isolate the rendered mesh positions. We compute the mesh's scale and translation with these selected mesh positions. We then translate the mesh to align with the origin of the coordinate system and scale it to fit within the $[-1, 1]$ range. This process ensures that the rendered portion of each mesh remains centered in the final image. After translating the mesh, we re-rasterize the UV position map to align it within the $[-1, 1]$ range.", + "bbox": [ + 511, + 554, + 906, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We render all garment data from two perspectives: front and back. All images are generated using Nvdiffrast [22] under uniform ambient lighting. These two views are then used to generate prompts by InternVL[13]. A toy case is shown in Figure. 3.", + "bbox": [ + 511, + 766, + 906, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Basic Architecture", + "text_level": 1, + "bbox": [ + 511, + 848, + 692, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt PixArt- $\\alpha$ [8] as our foundational image generation model, a Transformer-based text-to-image diffusion", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26569", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "\"The top is short-sleeved with a white base and a black collar and cuffs, with a black bow at the center. The skirt is black with white cross motifs.\"", + "bbox": [ + 101, + 80, + 464, + 102 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a6b8ef62f99ae9ccb36648e90293f1365d6cac4a860653f5c537ada8f4b14174.jpg", + "image_caption": [ + "Figure 3. From left to right, we show vvoid data, garment, UV texture image and UV position map, respectively." + ], + "image_footnote": [], + "bbox": [ + 101, + 109, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model that excels in image quality, artistry, and semantic control. Our choice of PixArt- $\\alpha$ as the base architecture is motivated by three key reasons:", + "bbox": [ + 89, + 229, + 482, + 273 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Strong Prior Foundation. PixArt- $\\alpha$ [8] is a popular text-to-image diffusion model that offers valuable insights to the AIGC community. Its robust image generation capabilities provide a solid prior for our method, enhancing generalization and quality.", + "- Efficiency and Resource Friendliness. With an optimized Transformer design, PixArt- $\\alpha$ [8] achieves high-quality image generation using only 0.6 billion parameters, minimizing parameters and training requirements", + "- Enhanced Artistic Style. Fine-tuned on high-quality aesthetic data, PixArt- $\\alpha$ [8] generates images with an ideal artistic style for garment texture generation, where visual appeal and stylistic quality are essential." + ], + "bbox": [ + 89, + 275, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model takes as input a text prompt, timestep $t$ , and clean images (clear UV texture $\\mathcal{I}_{\\mathrm{tex}}$ in our method). These inputs are processed through T5 [35], MLP, and VAE [21] encoder to obtain text, time, and latent embeddings. These embeddings are then fed into Transformer blocks to execute the diffusion process. Each Transformer block consists of self-attention, adaptive normalization (adaLN-single), cross-attention, and feedforward layers. Cross-attention uses the text embedding to interact with the Transformer's hidden states, while adaLN-single leverages the time embedding to learn scale and shift parameters.", + "bbox": [ + 88, + 473, + 482, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. 3D consistent Garment UV Texture Generation", + "text_level": 1, + "bbox": [ + 89, + 651, + 482, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UV representation is a widely adopted method for creating texture maps, enabling precise mapping of 3D textures onto 2D images by \"unwrapping\" the 3D surface into a 2D plane. This process assigns each point on the 3D model to a specific location in 2D space, ensuring that every detail of the 3D surface accurately corresponds to a position on the UV map. This approach effectively bridges 3D texture generation and 2D image-based methods, allowing advanced 2D generative models to handle tasks that traditionally require 3D-specific techniques. Working within the 2D UV space enables high-quality texture generation that can be seamlessly applied back to the 3D model, providing a flexible and efficient solution for producing consistent textures. However, adjacent regions in 3D space may become non-adjacent in the UV map, leading to a loss of global 3D", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "structural information during mapping. To address this, we design a 3D structure information extraction module and a 3D structure guidance module, which integrate the original 3D structural information from the garment mesh into the UV texture generation process, ensuring 3D consistency in the final UV texture maps.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Structure Information Extraction. For a given 3D garment mesh, we use rasterization to obtain the UV position map $\\mathcal{I}_{\\mathrm{pos}}$ . Each active pixel in $\\mathcal{I}_{\\mathrm{pos}}$ encodes the XYZ coordinates of a corresponding point on the garment mesh surface, creating a 2D representation of the 3D point cloud and capturing the basic layout of the UV texture. As shown in Figure 2-(b), $\\mathcal{I}_{\\mathrm{pos}}$ is first processed by a VAE [21] encoder to obtain the UV position latent, $\\mathcal{F}_{\\mathrm{pos}}$ . While $\\mathcal{F}_{\\mathrm{pos}}$ embeds the UV layout in 2D space, it lacks 3D structure information. To address this, we retrieve the 3D point cloud $\\mathcal{P}$ from $\\mathcal{I}_{\\mathrm{pos}}$ , normalize it to the range [0, 1], and project it into a 3D context representation space for effective interaction with $\\mathcal{F}_{\\mathrm{pos}}$ . Specifically, we encode the point cloud $\\mathcal{P}$ into a latent representation $\\mathcal{F}_{\\mathrm{point}}$ using a learnable embedding function and employ a cross-attention module to infuse $\\mathcal{F}_{\\mathrm{pos}}$ with 3D structure information from the point cloud:", + "bbox": [ + 511, + 181, + 906, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {p o i n t}} = \\operatorname {P o s E m b} (\\mathcal {P}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 439, + 903, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {l a t e n t}} ^ {3 \\mathrm {d}} = \\operatorname {F F N} \\left(\\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {\\text {p o s}} \\mathcal {K} _ {\\text {p o i n t}} ^ {\\mathrm {T}}}{\\sqrt {d}}\\right) \\mathcal {V} _ {\\text {p o i n t}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 474, + 903, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{point}}$ is the embedding of point cloud. $\\mathcal{Q}_{\\mathrm{pos}} = \\mathcal{F}_{\\mathrm{pos}}\\mathcal{W}_{\\mathrm{Q}}^{\\mathrm{pos}}$ is the query embedding of $\\mathcal{F}_{\\mathrm{pos}}$ . $\\mathcal{K}_{\\mathrm{point}} = \\mathcal{F}_{\\mathrm{point}}\\mathcal{W}_{\\mathrm{K}}^{\\mathrm{point}}$ and $\\mathcal{V}_{\\mathrm{point}} = \\mathcal{F}_{\\mathrm{point}}\\mathcal{W}_{\\mathrm{V}}^{\\mathrm{point}}$ are the key and value embeddings of $\\mathcal{F}_{\\mathrm{point}}$ , respectively. $\\mathcal{F}_{\\mathrm{latent}}^{\\mathrm{3d}}$ is the 3D structure latent that combines both the 2D layout information and the global 3D structure information.", + "bbox": [ + 511, + 521, + 905, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Structure Information Guidance. We incorporate the 3D structure guidance $\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}$ into the pre-trained text-to-image image diffusion process to guide the 3D-consistent UV texture generation. As shown in Figure 2-(c), we denote the transformer intermediate feature as $\\mathcal{F}_{\\mathrm{in}}$ and the text embedding as $\\mathcal{F}_{\\mathrm{text}}$ . Subsequently, we inject the 3D structure guidance $\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}$ into the image diffusion process by interacting with $\\mathcal{F}_{\\mathrm{in}}$ through the cross-attention:", + "bbox": [ + 511, + 616, + 905, + 737 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {o u t}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {\\text {i n}} \\mathcal {K} _ {\\text {3 d}} ^ {\\mathrm {T}}}{\\sqrt {d}}\\right) \\mathcal {V} _ {\\text {3 d}} + \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {\\text {i n}} \\mathcal {K} _ {\\text {t e x t}} ^ {\\mathrm {T}}}{\\sqrt {d}}\\right) \\mathcal {V} _ {\\text {t e x t}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 747, + 903, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}_{\\mathrm{in}} = \\mathcal{F}_{\\mathrm{in}}\\mathcal{W}_{\\mathrm{Q}}$ is the query embedding of the $\\mathcal{F}_{\\mathrm{in}}$ . $\\mathcal{K}_{3\\mathrm{d}} = \\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}\\mathcal{W}_{\\mathrm{K}},\\mathcal{V}_{3\\mathrm{d}} = \\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}\\mathcal{W}_{\\mathrm{V}}$ are the key and value embeddings of $\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}$ , respectively. $\\mathcal{K}_{\\mathrm{text}} = \\mathcal{F}_{\\mathrm{text}}\\mathcal{W}_{\\mathrm{K}'}$ , $\\mathcal{V}_{\\mathrm{text}} = \\mathcal{F}_{\\mathrm{text}}\\mathcal{W}_{\\mathrm{V}'}$ are the key, and value embeddings of $\\mathcal{F}_{\\mathrm{text}}$ respectively. $\\mathcal{W}_{\\mathrm{Q}},\\mathcal{W}_{\\mathrm{K}},\\mathcal{W}_{\\mathrm{K}'}$ , $\\mathcal{W}_{\\mathrm{V}}$ and $\\mathcal{W}_{\\mathrm{V}'}$ are the projection matrices. The two cross-attention results are added to obtain the $\\mathcal{F}_{\\mathrm{out}}$ .", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26570", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the text prompt condition $c_{\\mathrm{text}}$ from $\\mathcal{F}_{\\mathrm{text}}$ and 3D structure condition $c_{3\\mathrm{d}}$ from $\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}$ , we apply the classifier-free guidance [19] strategy to incorporate the information and our training objective is:", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {\\boldsymbol {x} \\sim p, \\epsilon \\sim \\mathcal {N} (0, I), c _ {\\text {t e x t}}, c _ {3 \\mathrm {d}}, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t}, t, c _ {\\text {t e x t}}, c _ {3 \\mathrm {d}}\\right) \\| _ {2} ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 156, + 483, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\boldsymbol{x}_t$ is the noise latent from the ground-truth $\\mathcal{I}_{\\mathrm{tex}}$ .", + "bbox": [ + 89, + 194, + 450, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After training, the diffusion model can generate diverse and 3D-consistent UV garment texture maps in a single inference process.", + "bbox": [ + 89, + 209, + 482, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 266, + 223, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 291, + 279, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Our dataset consists of 8k of textured 3D garments from internal series. Garment styles include long-sleeved, short-sleeved, long skirts, short skirts, long pants, shorts, and other common styles. The dataset is divided into two parts: the training set containing 7.9k garments and the testing set containing 0.1k garments.", + "bbox": [ + 89, + 311, + 482, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use PixArt- $\\alpha$ [8] as our pre-trained text-to-image generation model. The model is trained for 50,000 iterations with the AdamW [24] optimizer, using a fixed learning rate of $2 \\times 10^{-6}$ . Newly added parameters are initialized to zero and trained with a learning rate of $2 \\times 10^{-5}$ . The batch size is set to 32. For all experiments, we sample 5k points from the UV position map and synthesize a UV texture map at a resolution of $512 \\times 512$ .", + "bbox": [ + 89, + 402, + 482, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with several state-of-the-art approaches, including TEXTure [37], Text2Tex [7], Paint3D [45], and TexPainter [46]. TEXTure [37] introduces an iterative texture generation strategy to enhance texture manipulation. Text2Tex [7] builds on this by incorporating an automatic viewpoint selection process within each iteration. TexPainter [46] utilizes a modified denoising diffusion implicit model (DDIM) to merge textures across multiple views in color space. Paint3D [45] is a coarse-to-fine texture synthesis approach. First, it samples multi-view images from pre-trained 2D diffusion models, then refines the textures in UV space using a diffusion model conditioned on both the position map and the coarse texture.", + "bbox": [ + 89, + 523, + 482, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We utilize commonly used metrics for evaluating the performance of generative models, specifically, the Fréchet Inception Distance (FID) [18] and Kernel Inception Distance (KID) [3], which measure image quality and image diversity. Additionally, we use the CLIP score to measure the alignment between the rendered image and the input text prompt. Finally, we also evaluate the texture generation speed.", + "bbox": [ + 89, + 720, + 482, + 842 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 848, + 454, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our method with state-of-the-art methods both qualitatively and quantitatively. How-", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ever, due to the lengthy processing time required by Tex-Painter [46], we include only a qualitative comparison for this method.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons. In Figure 4, we compare our qualitative results with state-of-the-art methods. Our approach generates garment textures with enhanced 3D consistency and refined artistic quality. First, examining front and back views reveals that our method achieves consistent texture across perspectives. For example, in Figure 4-(a), the short-sleeved shirt generated by Paint3D [45] exhibits a \"Janus effect,\" while other methods show inconsistencies. In contrast, textures generated by our method remain coherent. Second, our textures exhibit finer details and organized layouts; for instance, Figure 4-(b) shows intricate lace and embroidery produced by our method. Finally, our approach aligns more closely with the text description, as demonstrated in Figure 4-(c), where our textures incorporate more Peking Opera elements than those of other methods.", + "bbox": [ + 511, + 136, + 906, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparisons. In Table 1, we present the quantitative comparisons with the previous methods in garment texture synthesis. For the FID [18] and KID [3] calculations, we render $512 \\times 512$ images of each mesh with the synthesized textures, captured from 20 fixed viewpoints. We use the average CLIP score to measure the alignment between rendering images and text prompts. For runtime, we compare inference time for each method, defining inference time as the time it takes to generate a complete texture map for a given text prompt and predefined mesh. From the results, we have the following observation. First, our method obtains the best performance in terms of FID [18] and KID [3]. Notably, it outperforms the leading competitor, Paint3D [45], by an average of 50.43 in FID and 2.23 in KID. These improvements indicate that our method excels at generating high-quality garment textures. Second, our method achieves the best CLIP scores, indicating better text-texture alignment. Finally, we use the same GPU device (a single NVIDIA-V100) for the runtime comparison to infer each method. Our method only takes 4 seconds to generate a UV texture map, which is much less time-consuming than other methods. This proves the efficiency of our method.", + "bbox": [ + 511, + 363, + 906, + 709 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6d71293288ea60de905818de2f843885b934af2a31d1dbe1aca1304d2e80b7b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓KID↓Runtime↓CLIP↑
TEXTure [37]169.014.35~140s27.26
Text2Tex [7]147.482.77~940s27.93
Paint3D [45]147.402.86~220s28.72
Ours96.970.63~4s32.37
", + "bbox": [ + 514, + 719, + 898, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison against state-of-the-art methods. We report the FID, $\\mathrm{{KID}}\\left( {\\times {10}^{2}}\\right)$ ,CLIP score and runtime.", + "bbox": [ + 511, + 815, + 903, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User study. We further conduct a user study to analyze the overall quality of the generated textures. We randomly select 50 garment meshes and corresponding text prompts", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26571", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0536be6d5f5f93adb376fb88cdd2870dba237480c0413d2a428f2010c323efae.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with the state-of-the-art methods. Our method generates garment textures with complex and clear texture details, which are both aesthetically pleasing and maintain 3D consistency. Please enlarge the image for easy observation of details." + ], + "image_footnote": [], + "bbox": [ + 127, + 85, + 867, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to perform the user study. Following the evaluation criteria of Paint3d [45], those meshes are textured in both our method and comparison methods. Participants are requested to evaluate the results based on 1) overall quality and 2) fidelity to the text prompt, using a scale of 1 to 5. We collect the evaluation results from 30 users. As presented in Table 2, we show the average results across all prompts for each method. Our approach outperforms other methods in terms of both overall quality and text fidelity.", + "bbox": [ + 89, + 460, + 482, + 597 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2144b81cd8353a891d9a9bb2de5f461d1121370e858494551f287eaa8d6be9bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImage Quality↑Text Fidelity↑
TEXTure [37]3.213.57
Text2Tex [7]3.323.41
Paint3D [45]4.014.17
Ours4.834.92
", + "bbox": [ + 94, + 609, + 475, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Statistics on user preferences for texture quality. Users more favor our method.", + "bbox": [ + 89, + 705, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Diversity and Consistency", + "text_level": 1, + "bbox": [ + 89, + 742, + 326, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Garment texture generation demands high diversity and 3D consistency standards to achieve realistic representation. Unlike other texture tasks, garment textures must capture various fabrics, patterns, and colors to reflect the richness of real-world apparel. Garments often feature unique details like intricate stitching, subtle textures, and complex patterns that vary across styles and materials. Ensuring 3D consistency is crucial, as textures must appear seamless and coherent from all angles to maintain realism in virtual envi", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ronments. Any misalignment can break the illusion, making the garment look artificial. Therefore, effective garment texture generation must balance diversity with strong 3D consistency, ensuring textures adapt smoothly to the garment's surface from every view.", + "bbox": [ + 511, + 460, + 906, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To showcase our method's ability to generate diverse and 3D consistent garment textures, we present two examples in Figure 5. Given a UV position map and three text prompts, our method generates three unique texture styles for the same garment mesh. We observe the following: 1) Our method produces semantically rich textures aligned with the text prompts. Each texture displays intricate details and logical layout, illustrating our model's ability to create a variety of garment styles that meet diverse user preferences. 2) We present four different views for each generated texture. We observe a high level of 3D consistency in texture details across these views, which is achieved by our method's ability to generate complete UV texture in UV space.", + "bbox": [ + 511, + 536, + 908, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 742, + 666, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation of 3D structure guidance. To evaluate the effectiveness of our newly designed 3D structure guidance, we conduct experiments on three baseline configurations: point cloud guidance, UV position map guidance, and mask guidance. Point cloud guidance (3D Info.): This configuration uses only the point cloud representation as guidance without incorporating the 2D UV layout information from the UV position map. UV position map guidance (2D Info.): This configuration uses only the 2D UV layout", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26572", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72a602140849f92ccd282aaee9e5c07d471a1caa187104e3d42bfef06fb0eb8c.jpg", + "image_caption": [ + "Figure 5. Given a UV position map and varying text prompts, our method can generate garment textures in distinct styles. Observing the four views confirms that the generated textures maintain global 3D consistency." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 861, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "information from the UV position map as guidance, without the 3D global information provided by the point cloud. Mask guidance (Mask Info.): Only the UV texture mask is embedded using a VAE [21] encoder as guidance, with no structural information, neither 2D nor 3D.", + "bbox": [ + 89, + 454, + 482, + 529 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3882a31bed8ef230224e050276ed942968130290400e15a73fde9de7aa38893f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓KID↓
3D Info.187.638.73
2D Info.174.298.07
Mask Info.210.3412.59
Ours96.970.63
", + "bbox": [ + 133, + 540, + 436, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Evaluation of modules in our method. This demonstrates the effectiveness of each component. Our full model achieves optimal performance.", + "bbox": [ + 89, + 633, + 482, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 3, we compare three baselines with our method, leading to the following observations: 1) Compared to our full model, 3D Info. provides global 3D information but lacks 2D UV layout details. This results in worse FID and KID scores for 3D Info.. 2) When comparing 2D Info. with our full model, we find that while 2D Info. includes 2D UV layout information, it lacks global 3D information. As in the previous case, 2D Info. shows worse FID and KID scores, emphasizing the importance of 3D structural data. 3) Mask Info. provides only basic UV texture shape information, lacking structural details, making it the weakest performer among the baselines. These observations demonstrate the effectiveness of our method, which leverages global 3D structure information and 2D", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "layout details to provide guidance for improved garment texture generation.", + "bbox": [ + 511, + 454, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 498, + 638, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we present TexGarment, a novel and efficient method for generating high-quality garment textures. By leveraging pre-trained 2D image generation models and incorporating 3D garment structure information, TexGarment addresses key challenges in texture generation, such as global 3D consistency and inference speed. Our method enables the generation of diverse and visually consistent UV texture maps in a single inference process, significantly improving on previous multi-stage approaches. Through the introduction of two key components: 3D structure information extraction and 3D structure guidance, we effectively integrate 3D structural details into the 2D texture generation process, resulting in high fidelity and 3D consistency. Our experiments show that TexGarment outperforms existing methods in both quality and performance, with user studies confirming its superior visual appeal.", + "bbox": [ + 511, + 523, + 906, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. TexGarment now does not support image-conditioned generation. In practical applications, users may provide a reference image to generate garment textures in a similar style. We will explore it in future research.", + "bbox": [ + 511, + 776, + 906, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under grant 62202369.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26573", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josef Albers. Interaction of color. Kyung Dang, 2013. 2", + "[2] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 2, 3", + "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6", + "[4] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8918-8928, 2023. 2, 3", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3", + "[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pages 4055-4075, 2023. 3", + "[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 2, 3, 6, 7", + "[8] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations. 2, 3, 4, 5, 6", + "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 3", + "[10] Shuhong Chen, Kevin Zhang, Yichun Shi, Heng Wang, Yiheng Zhu, Guoxian Song, Sizhe An, Janus Kristjansson, Xiao Yang, and Matthias Zwicker. Panic-3d: Stylized single-view 3d reconstruction from portraits of anime characters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21068-21077, 2023. 3", + "[11] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 3", + "[12] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In CVPR, pages 1465-1474, 2022. 2, 3", + "[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. *Internvl: Scaling up vision foundation models and" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 4", + "[14] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 3", + "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3", + "[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 6", + "[20] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3", + "[21] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4, 5, 8", + "[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 4", + "[23] Jialun Liu, Chenming Wu, Xinqi Liu, Xing Liu, Jinbo Wu, Haotian Peng, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Texoct: Generating textures of 3d models with octree-based diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4284-4293, 2024. 2, 3", + "[24] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[25] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 3", + "[26] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 3" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26574", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3", + "[28] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3", + "[29] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 3", + "[30] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 3", + "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3", + "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations. 3", + "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations. 3", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 4, 5", + "[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3", + "[37] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 2, 3, 6, 7", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3", + "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3", + "[40] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2", + "[41] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In European Conference on Computer Vision, pages 72-88. Springer, 2022. 3", + "[42] Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. Cycle-consistent inverse gan for text-to-image synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 630-638, 2021. 2", + "[43] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research. 3", + "[44] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206-4216, 2023. 2, 3", + "[45] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 2, 3, 6, 7", + "[46] Hongkun Zhang, Zherong Pan, Congyi Zhang, Lifeng Zhu, and Xifeng Gao. Texpainter: Generative mesh texturing with multi-view consistency. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 585 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26575", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_model.json b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c99020e23505b775030ba63ff1c8540f5901e598 --- /dev/null +++ b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_model.json @@ -0,0 +1,1826 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.817, + 0.175 + ], + "angle": 0, + "content": "TexGarment: Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.203, + 0.852, + 0.295 + ], + "angle": 0, + "content": "Jialun Liu\\(^{1}\\), Jinbo Wu\\(^{1*}\\), Xiaobo Gao\\(^{1}\\), Jiakui Hu\\(^{2}\\), Bojun Xiong\\(^{2}\\), Xing Liu\\(^{1}\\), Chen Zhao\\(^{1}\\), Hongbin Pei\\(^{3}\\), Haocheng Feng\\(^{1}\\), Yingying Li\\(^{1}\\), Errui Ding\\(^{1}\\), Jingdong Wang\\(^{1}\\) \n\\(^{1}\\)Department of Computer Vision Technology(VIS), Baidu Inc. \n\\(^{2}\\)Peking University \\(^{3}\\)Xi'an Jiaotong University liujialun@baidu.com" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.313, + 0.885, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.608, + 0.908, + 0.639 + ], + "angle": 0, + "content": "Figure 1. A gallery showcasing generated garment textures by TexGarment. Our method achieves globally consistent, high-quality, and high-fidelity textures across a wide range of garment styles. The average time to generate each texture map is under 4 seconds." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.652, + 0.33, + 0.669 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.683, + 0.487, + 0.881 + ], + "angle": 0, + "content": "This paper introduces TexGarment, an efficient method for synthesizing high-quality, 3D-consistent garment textures in UV space. Traditional approaches based on 2D-to-3D mapping often suffer from 3D inconsistency, while methods learning from limited 3D data lack sufficient texture diversity. These limitations are particularly problematic in garment texture generation, where high demands exist for both detail and variety. To address these challenges, TexGarment leverages a pre-trained text-to-image diffusion Transformer model with robust generalization capabilities, introducing structural information to guide the model in generating 3D-consistent garment textures in a single inference step. Specifically, We utilize the 2D UV position map to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.911, + 0.807 + ], + "angle": 0, + "content": "guide the layout during the UV texture generation process, ensuring a coherent texture arrangement and enhancing it by integrating global 3D structural information from the mesh surface point cloud. This combined guidance effectively aligns 3D structural integrity with 2D layout. Our method efficiently generates high-quality, diverse UV textures in a single inference step while maintaining 3D consistency. Experimental results validate the effectiveness of TexGarment, achieving state-of-the-art performance in 3D garment texture generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.646, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The demand for 3D garment visualization is growing across virtual try-ons for e-commerce, video games, 3D films, and AR/VR environments. These industries need large volumes" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.322, + 0.901 + ], + "angle": 0, + "content": "*Project lead and corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26566" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "of high-quality 3D garments to meet consumer expectations. While 2D fashion images are widely available and generative AI has simplified 2D image creation, producing high-quality 3D garment textures remains challenging, often requiring skilled artists and weeks of work. This highlights the need for efficient 3D garment generation methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.482, + 0.352 + ], + "angle": 0, + "content": "The aesthetic appeal of 3D garments is deeply influenced by their texture quality, as Joseph Albers [1] said: \"Texture is the soul of painting; it gives surface depth and personality\", effectively capturing this concept. In 3D computer graphics, textures are essential for realism, adding detailed surface elements that make virtual objects lifelike. For 3D garments, high-quality textures are particularly important, capturing fabric nuances, stitching, folds, and shadows that make virtual clothing look authentic. Without these textures, even well-modeled garments can appear flat or artificial, reducing the immersive quality of 3D applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Despite substantial advancements in 2D image synthesis and 3D shape generation using GANs [40, 42] and diffusion models [12, 38, 39], achieving high-quality garment texture generation remains challenging. This difficulty arises from three primary factors: Global 3D Consistency. Methods employing 2D diffusion models, such as TexPainter [46], TEXTure [37] and Text2Tex [7], have made strides in texture synthesis by utilizing pre-trained LDM [38], which support strong generalization and diverse texture generation. However, text-to-image models often exhibit biases that limit their effectiveness for 3D applications, such as favoring frontal views and lacking a robust understanding of 3D structures. This frequently results in texture inconsistencies across different viewpoints of the same 3D object, ranging from minor seam misalignments to severe issues, e.g., \"Janus effect\". Generalization Ability. Methods learning from limited 3D data, such as TexOct [23], PointUV [44], and Mesh2Tex [4], generate textures by explicitly integrating 3D shape information, resulting in high 3D consistency. However, these approaches often lack generalization ability, limiting their effectiveness when generating textures for diverse 3D objects beyond the training set or adapting to a range of text and visual prompts. Inference Speed. To achieve full shape coverage, prior methods [2, 37, 44, 45] rely on iterative generation processes, requiring multiple steps or stages. This results in prolonged inference times, often lasting several minutes, and requires significant computational resources, limiting practical use for designers and content creators." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "In this work, we present TexGarment, a method tailored for generating high-quality garment textures. Given text prompts, TexGarment produces consistent, diverse, and visually striking textures in a single inference pass. A widely used approach for handling texture maps is UV representation, which maps 3D textures accurately onto 2D UV images through UV coordinate transformations. This tech-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "nique bridges the gap between 3D texture generation and 2D image methods, motivating our use of pre-trained image generation models [8] to synthesize diverse textures effectively within the 2D UV space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.152, + 0.905, + 0.62 + ], + "angle": 0, + "content": "Our key insight is to incorporate 3D garment structure information as guidance within a pre-trained 2D image generation model [8], leveraging its generalization capabilities to produce consistent and diverse UV texture maps directly in the 2D UV space through only one inference process. Specifically, our method introduces two main components: 1) 3D Structure Information Extraction and 2) 3D Structure Guidance. First, given a garment mesh, we use rasterization to obtain a UV position map, where each pixel encodes the XYZ coordinates of the corresponding point on the mesh surface. This UV position map provides a 2D projection of the original 3D point cloud, capturing the basic layout of the UV texture. However, regions that are adjacent in 3D space may be separated in the UV map, causing a loss of global 3D structure information. This makes it challenging for a 2D image generation model to capture global 3D structures solely from the 2D UV position map. To address this, we introduce a 3D structure information extraction module that injects global structure information, represented by 3D point clouds, into the 2D UV position map. This process generates a 3D structure latent, which combines both 3D structure and 2D layout information, serving as guidance for the pre-trained image generation model. Second, we implement a 3D structure guidance module, analogous to text guidance, which injects the 3D structure information into the pre-trained model via cross-attention. Using a powerful text-to-image diffusion model [8] as our basic architecture, our design enables the model to receive guidance from both text and 3D structure, allowing it to generate high-quality UV texture maps in a single inference process, achieving both text fidelity and 3D consistency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.621, + 0.905, + 0.832 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: 1) We propose TexGarment, a one-stage method for garment texture generation that produces high-quality, diverse UV texture maps for untextured garment meshes in a single inference process and completes the process in 4 seconds. 2) We introduce two novel modules to incorporate 3D garment structure information into the 2D image generation model, providing 3D-consistent guidance for garment texture generation. Extensive experiments validate the effectiveness of these modules in achieving 3D consistency. 3) Qualitative and quantitative evaluations demonstrate that our method outperforms state-of-the-art approaches in garment texture generation. Additionally, a user study shows that users prefer textures generated by our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.665, + 0.861 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Text-to-Image Generation The field of text-to-image generation has seen considerable progress. Early research used" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26567" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "generative adversarial networks (GANs)[17] to create high-quality images, using image classes[15] and segmentation maps [30] as prompts. Recently, attention has shifted to architectural improvements that enhance generation quality with scaling Transformers, as shown in models like DALL-E [36], CogView [14], Parti [43], and Muse [6]. Another popular category of text-to-image generators is based on pixel space or latent diffusion models, including Imagegen [39], Stable Diffusion [38], SDXL [32], DiT [31], and PixArt [8]. Notably, DiT [31] advanced text-to-image generation by introducing Transformers to latent diffusion models (LDMs). PixArt [8] further optimized training methods, enabling faster training of scaled LDMs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.686 + ], + "angle": 0, + "content": "3D Generation via 2D generation model. With the development of text-to-image generation techniques, many 3D texturing techniques seek to leverage the prior knowledge of 2D generative models to improve 3D generation models' optimization. Previous optimization-based texture techniques [11, 20, 25, 27, 28] concentrated on improving the texture maps of 3D models by utilizing the large-scale visual language model, e.g. CLIP [34], before the advent of LDMs. DreamFusion [33] introduced the score distillation sampling (SDS) to 3D Generation, which laid the groundwork for many text-to-3D methods. Latent-nerf [26] and Fantasia3D [9] extended SDS to optimize texture mapping. Recently, TexPainter [46] performs multi-view fusion within the 2D image's color space to address inconsistency issues. However, optimization-based texture methods are hindered by prolonged training times. As an alternative, TEXTure [37], Text2Tex [7], and Paint3D [45] involve directly generating texture images from multiple views of 3D meshes and subsequently painting them onto the 3D meshes. TEXTure [37] and Text2Tex [7] adhere to this concept and suggest modifying the overlapping segments of the texture image from the previous view using results from each new camera pose. Nevertheless, these methods encounter various artifacts, including seams, noise, and irrelevant fragments. Paint3D [45] develops a specialized model to paint incomplete regions during the texture painting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Generative Texturing from 3D Data. Additionally, generative texture models are capable of training models from scratch using 3D data. To assign a color to each pixel on the surface of a three-dimensional shape, early methods [16, 29] relied on implicit texture fields. However, given that the texture on the surface of a 3D mesh is continuous, it is unlikely that discrete supervision will result in the training of a model capable of synthesizing high-quality textures. The Texturify [41] is designed to construct a texture map on the surface of a polygonal mesh. Mesh2tex [4] builds on this by integrating an implicit texture field branch for enhancement. TexOct [23] use point clouds to help circumvent variations arising from diverse mesh topologies and UV mappings. Other methods [12, 44] synthesize UV" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "maps for 3D meshes, but are ill-equipped to handle more generic objects, due to the inherent diversity between 3D meshes belonging to various categories." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Discussion with previous work about the UV position map. Previous works, such as PointUV [44], Paint3D [45], and TextureGen [2], also incorporate the UV position map into their models. We outline the key differences between our approach and these methods: 1) PointUV [44] uses the UV position map, along with other maps like normal and mask maps, in the first stage to generate a coarse UV texture. However, this method is limited to simple textures for objects from the ShapeNet [5] dataset. In contrast, our approach uses only the UV texture map as a guide, enabling the generation of complex, high-quality textures for a wide range of garment categories. 2) Paint3D [45] and TextureGen [2] use the UV position map as a 2D condition in a second-stage inpainting model to refine textures. Our method directly applies the UV position map to guide the entire texture generation, leveraging both 2D layout and global 3D structure, resulting in better textures as shown in our experiments. 3) Unlike these prior methods, our model is a single-stage generator. Given a UV position map and a text prompt, our approach generates high-quality garment texture maps in a single, efficient inference step, while previous methods require a more complex two-stage process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.482, + 0.605, + 0.497 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.508, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Given a 3D garment model and a target texture description, TexGarment efficiently synthesizes high-quality, diverse, and consistent garment UV textures. TexGarment is a Transformer-based diffusion model, as shown in Figure 2. We build on a large-scale pre-trained image generation model [8], leveraging its robust prior knowledge to enhance our model's generalization capability. To ensure global 3D consistency, we first design a 3D structure information extraction module. For an untextured garment mesh, we apply rasterization to create a UV position map that captures the basic UV texture layout and encodes the 3D coordinates of points on the mesh surface. The structure information extraction module then embeds global structure information from the original point cloud into the UV position map, resulting in a 3D structure latent that combines global 3D structure information with 2D layout details. This 3D structure latent is then used by a 3D structure guidance module to guide the texture generation. Through this innovative design, our method generates high-quality and 3D-consistent texture maps in a single inference process, without the need for additional operations such as inpainting. The following sections provide the overview of each stage." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.849, + 0.673, + 0.865 + ], + "angle": 0, + "content": "3.1. Data Processing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Dataset. We collect a comprehensive dataset from VRoidHub based on the guidelines provided by Panic3D[10], en" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26568" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.09, + 0.58, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.408, + 0.26, + 0.419 + ], + "angle": 0, + "content": "(a) Overview" + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.407, + 0.564, + 0.418 + ], + "angle": 0, + "content": "(b) 3D Structure Information Extraction" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.092, + 0.875, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.405, + 0.83, + 0.417 + ], + "angle": 0, + "content": "(c) Block with 3D sturcture guidance" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.428, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 2. Pipeline of Our Method: (a) Overview: The training process begins with embedding the input text prompt and timestep using the T5 model [35] and an MLP layer, generating the text feature and time embedding. The UV texture is encoded by a VAE [21] encoder to create a texture latent representation, which is then noised to produce a noisy latent. These representations are processed through several Transformer blocks. (b) 3D Structure Information Extraction Module: The UV position map is encoded by a VAE [21] encoder to obtain the UV position latent, while a point cloud feature extraction module captures the global 3D structure. This 3D information is integrated into the UV position latent via cross-attention, resulting in the 3D structure latent, which combines 3D global structure and 2D UV texture layout. (c) Transformer Block Details: Based on a text-to-image Transformer diffusion model [8], we introduce a 3D structure guidance module that uses the 3D structure latent as guidance, injecting it into the model through cross-attention to enhance 3d structure consistency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.555, + 0.483, + 0.722 + ], + "angle": 0, + "content": "compassing a diverse array of character models. Each model in VRoidHub includes body, clothing, and animation components. Using Blender, we segment the models by style type, preserving clothing elements while removing non-clothing components. Pants and tops were separated into distinct parts, while dresses were preserved as single units. Following segmentation, we refine the dataset by excluding samples with inconsistent color distributions and monochromatic appearances, as well as items with accessories such as ties, belts, and other extraneous parts, ultimately retaining a dataset of 8,000 clothing items." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Rasterization. In a UV texture image, the RGB information of the surface mesh is stored, and during rendering, a rasterization operator samples the color from the UV texture image onto the rendered image. A UV position map stores the XYZ coordinates of the mesh, which requires sampling mesh properties and baking them into the UV map. To achieve this, we assume the UV layout represents a 3D mesh with vertices \\([u,v,z]\\), where \\(u\\) and \\(v\\) are the original mesh's UV coordinates and \\(z\\) is constant, making it a planar mesh. Each \\([u,v]\\) corresponds to a vertex in the original 3D mesh. We retain the face relationships from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.766 + ], + "angle": 0, + "content": "the original mesh, and by performing rasterization, we interpolate the \"UV mesh\" positions to determine each pixel's XYZ coordinates on the UV position map. The shapes of garment meshes in the VRoid data are also defined by an alpha channel within the UV texture. Before rendering our images, we rasterize the UV position map using the alpha channel to isolate the rendered mesh positions. We compute the mesh's scale and translation with these selected mesh positions. We then translate the mesh to align with the origin of the coordinate system and scale it to fit within the \\([-1, 1]\\) range. This process ensures that the rendered portion of each mesh remains centered in the final image. After translating the mesh, we re-rasterize the UV position map to align it within the \\([-1, 1]\\) range." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.907, + 0.843 + ], + "angle": 0, + "content": "We render all garment data from two perspectives: front and back. All images are generated using Nvdiffrast [22] under uniform ambient lighting. These two views are then used to generate prompts by InternVL[13]. A toy case is shown in Figure. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.693, + 0.863 + ], + "angle": 0, + "content": "3.2. Basic Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We adopt PixArt-\\(\\alpha\\) [8] as our foundational image generation model, a Transformer-based text-to-image diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26569" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.102, + 0.082, + 0.465, + 0.103 + ], + "angle": 0, + "content": "\"The top is short-sleeved with a white base and a black collar and cuffs, with a black bow at the center. The skirt is black with white cross motifs.\"" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.11, + 0.47, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.187, + 0.483, + 0.216 + ], + "angle": 0, + "content": "Figure 3. From left to right, we show vvoid data, garment, UV texture image and UV position map, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.483, + 0.274 + ], + "angle": 0, + "content": "model that excels in image quality, artistry, and semantic control. Our choice of PixArt-\\(\\alpha\\) as the base architecture is motivated by three key reasons:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.276, + 0.483, + 0.351 + ], + "angle": 0, + "content": "- Strong Prior Foundation. PixArt-\\(\\alpha\\) [8] is a popular text-to-image diffusion model that offers valuable insights to the AIGC community. Its robust image generation capabilities provide a solid prior for our method, enhancing generalization and quality." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.352, + 0.482, + 0.412 + ], + "angle": 0, + "content": "- Efficiency and Resource Friendliness. With an optimized Transformer design, PixArt-\\(\\alpha\\) [8] achieves high-quality image generation using only 0.6 billion parameters, minimizing parameters and training requirements" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.412, + 0.482, + 0.472 + ], + "angle": 0, + "content": "- Enhanced Artistic Style. Fine-tuned on high-quality aesthetic data, PixArt-\\(\\alpha\\) [8] generates images with an ideal artistic style for garment texture generation, where visual appeal and stylistic quality are essential." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.276, + 0.483, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.474, + 0.483, + 0.64 + ], + "angle": 0, + "content": "The model takes as input a text prompt, timestep \\( t \\), and clean images (clear UV texture \\( \\mathcal{I}_{\\mathrm{tex}} \\) in our method). These inputs are processed through T5 [35], MLP, and VAE [21] encoder to obtain text, time, and latent embeddings. These embeddings are then fed into Transformer blocks to execute the diffusion process. Each Transformer block consists of self-attention, adaptive normalization (adaLN-single), cross-attention, and feedforward layers. Cross-attention uses the text embedding to interact with the Transformer's hidden states, while adaLN-single leverages the time embedding to learn scale and shift parameters." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.667 + ], + "angle": 0, + "content": "3.3. 3D consistent Garment UV Texture Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "UV representation is a widely adopted method for creating texture maps, enabling precise mapping of 3D textures onto 2D images by \"unwrapping\" the 3D surface into a 2D plane. This process assigns each point on the 3D model to a specific location in 2D space, ensuring that every detail of the 3D surface accurately corresponds to a position on the UV map. This approach effectively bridges 3D texture generation and 2D image-based methods, allowing advanced 2D generative models to handle tasks that traditionally require 3D-specific techniques. Working within the 2D UV space enables high-quality texture generation that can be seamlessly applied back to the 3D model, providing a flexible and efficient solution for producing consistent textures. However, adjacent regions in 3D space may become non-adjacent in the UV map, leading to a loss of global 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "structural information during mapping. To address this, we design a 3D structure information extraction module and a 3D structure guidance module, which integrate the original 3D structural information from the garment mesh into the UV texture generation process, ensuring 3D consistency in the final UV texture maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.425 + ], + "angle": 0, + "content": "3D Structure Information Extraction. For a given 3D garment mesh, we use rasterization to obtain the UV position map \\(\\mathcal{I}_{\\mathrm{pos}}\\). Each active pixel in \\(\\mathcal{I}_{\\mathrm{pos}}\\) encodes the XYZ coordinates of a corresponding point on the garment mesh surface, creating a 2D representation of the 3D point cloud and capturing the basic layout of the UV texture. As shown in Figure 2-(b), \\(\\mathcal{I}_{\\mathrm{pos}}\\) is first processed by a VAE [21] encoder to obtain the UV position latent, \\(\\mathcal{F}_{\\mathrm{pos}}\\). While \\(\\mathcal{F}_{\\mathrm{pos}}\\) embeds the UV layout in 2D space, it lacks 3D structure information. To address this, we retrieve the 3D point cloud \\(\\mathcal{P}\\) from \\(\\mathcal{I}_{\\mathrm{pos}}\\), normalize it to the range [0, 1], and project it into a 3D context representation space for effective interaction with \\(\\mathcal{F}_{\\mathrm{pos}}\\). Specifically, we encode the point cloud \\(\\mathcal{P}\\) into a latent representation \\(\\mathcal{F}_{\\mathrm{point}}\\) using a learnable embedding function and employ a cross-attention module to infuse \\(\\mathcal{F}_{\\mathrm{pos}}\\) with 3D structure information from the point cloud:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.44, + 0.905, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {p o i n t}} = \\operatorname {P o s E m b} (\\mathcal {P}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {l a t e n t}} ^ {3 \\mathrm {d}} = \\operatorname {F F N} \\left(\\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {\\text {p o s}} \\mathcal {K} _ {\\text {p o i n t}} ^ {\\mathrm {T}}}{\\sqrt {d}}\\right) \\mathcal {V} _ {\\text {p o i n t}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.522, + 0.906, + 0.616 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{point}}\\) is the embedding of point cloud. \\(\\mathcal{Q}_{\\mathrm{pos}} = \\mathcal{F}_{\\mathrm{pos}}\\mathcal{W}_{\\mathrm{Q}}^{\\mathrm{pos}}\\) is the query embedding of \\(\\mathcal{F}_{\\mathrm{pos}}\\). \\(\\mathcal{K}_{\\mathrm{point}} = \\mathcal{F}_{\\mathrm{point}}\\mathcal{W}_{\\mathrm{K}}^{\\mathrm{point}}\\) and \\(\\mathcal{V}_{\\mathrm{point}} = \\mathcal{F}_{\\mathrm{point}}\\mathcal{W}_{\\mathrm{V}}^{\\mathrm{point}}\\) are the key and value embeddings of \\(\\mathcal{F}_{\\mathrm{point}}\\), respectively. \\(\\mathcal{F}_{\\mathrm{latent}}^{\\mathrm{3d}}\\) is the 3D structure latent that combines both the 2D layout information and the global 3D structure information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.906, + 0.738 + ], + "angle": 0, + "content": "3D Structure Information Guidance. We incorporate the 3D structure guidance \\(\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}\\) into the pre-trained text-to-image image diffusion process to guide the 3D-consistent UV texture generation. As shown in Figure 2-(c), we denote the transformer intermediate feature as \\(\\mathcal{F}_{\\mathrm{in}}\\) and the text embedding as \\(\\mathcal{F}_{\\mathrm{text}}\\). Subsequently, we inject the 3D structure guidance \\(\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}\\) into the image diffusion process by interacting with \\(\\mathcal{F}_{\\mathrm{in}}\\) through the cross-attention:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.748, + 0.905, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {o u t}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {\\text {i n}} \\mathcal {K} _ {\\text {3 d}} ^ {\\mathrm {T}}}{\\sqrt {d}}\\right) \\mathcal {V} _ {\\text {3 d}} + \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {\\text {i n}} \\mathcal {K} _ {\\text {t e x t}} ^ {\\mathrm {T}}}{\\sqrt {d}}\\right) \\mathcal {V} _ {\\text {t e x t}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}_{\\mathrm{in}} = \\mathcal{F}_{\\mathrm{in}}\\mathcal{W}_{\\mathrm{Q}}\\) is the query embedding of the \\(\\mathcal{F}_{\\mathrm{in}}\\). \\(\\mathcal{K}_{3\\mathrm{d}} = \\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}\\mathcal{W}_{\\mathrm{K}},\\mathcal{V}_{3\\mathrm{d}} = \\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}\\mathcal{W}_{\\mathrm{V}}\\) are the key and value embeddings of \\(\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}\\), respectively. \\(\\mathcal{K}_{\\mathrm{text}} = \\mathcal{F}_{\\mathrm{text}}\\mathcal{W}_{\\mathrm{K}'}\\), \\(\\mathcal{V}_{\\mathrm{text}} = \\mathcal{F}_{\\mathrm{text}}\\mathcal{W}_{\\mathrm{V}'}\\) are the key, and value embeddings of \\(\\mathcal{F}_{\\mathrm{text}}\\) respectively. \\(\\mathcal{W}_{\\mathrm{Q}},\\mathcal{W}_{\\mathrm{K}},\\mathcal{W}_{\\mathrm{K}'}\\), \\(\\mathcal{W}_{\\mathrm{V}}\\) and \\(\\mathcal{W}_{\\mathrm{V}'}\\) are the projection matrices. The two cross-attention results are added to obtain the \\(\\mathcal{F}_{\\mathrm{out}}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26570" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "Given the text prompt condition \\( c_{\\mathrm{text}} \\) from \\( \\mathcal{F}_{\\mathrm{text}} \\) and 3D structure condition \\( c_{3\\mathrm{d}} \\) from \\( \\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}} \\), we apply the classifier-free guidance [19] strategy to incorporate the information and our training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.157, + 0.485, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {\\boldsymbol {x} \\sim p, \\epsilon \\sim \\mathcal {N} (0, I), c _ {\\text {t e x t}}, c _ {3 \\mathrm {d}}, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t}, t, c _ {\\text {t e x t}}, c _ {3 \\mathrm {d}}\\right) \\| _ {2} ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.195, + 0.452, + 0.21 + ], + "angle": 0, + "content": "where \\( \\boldsymbol{x}_t \\) is the noise latent from the ground-truth \\( \\mathcal{I}_{\\mathrm{tex}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.21, + 0.483, + 0.256 + ], + "angle": 0, + "content": "After training, the diffusion model can generate diverse and 3D-consistent UV garment texture maps in a single inference process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.224, + 0.284 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.292, + 0.281, + 0.308 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.483, + 0.404 + ], + "angle": 0, + "content": "Datasets. Our dataset consists of 8k of textured 3D garments from internal series. Garment styles include long-sleeved, short-sleeved, long skirts, short skirts, long pants, shorts, and other common styles. The dataset is divided into two parts: the training set containing 7.9k garments and the testing set containing 0.1k garments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Implementation Details. We use PixArt-\\(\\alpha\\) [8] as our pre-trained text-to-image generation model. The model is trained for 50,000 iterations with the AdamW [24] optimizer, using a fixed learning rate of \\(2 \\times 10^{-6}\\). Newly added parameters are initialized to zero and trained with a learning rate of \\(2 \\times 10^{-5}\\). The batch size is set to 32. For all experiments, we sample 5k points from the UV position map and synthesize a UV texture map at a resolution of \\(512 \\times 512\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Baselines. We compare our method with several state-of-the-art approaches, including TEXTure [37], Text2Tex [7], Paint3D [45], and TexPainter [46]. TEXTure [37] introduces an iterative texture generation strategy to enhance texture manipulation. Text2Tex [7] builds on this by incorporating an automatic viewpoint selection process within each iteration. TexPainter [46] utilizes a modified denoising diffusion implicit model (DDIM) to merge textures across multiple views in color space. Paint3D [45] is a coarse-to-fine texture synthesis approach. First, it samples multi-view images from pre-trained 2D diffusion models, then refines the textures in UV space using a diffusion model conditioned on both the position map and the coarse texture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.483, + 0.843 + ], + "angle": 0, + "content": "Evaluation metrics. We utilize commonly used metrics for evaluating the performance of generative models, specifically, the Fréchet Inception Distance (FID) [18] and Kernel Inception Distance (KID) [3], which measure image quality and image diversity. Additionally, we use the CLIP score to measure the alignment between the rendered image and the input text prompt. Finally, we also evaluate the texture generation speed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.455, + 0.865 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we compare our method with state-of-the-art methods both qualitatively and quantitatively. How-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "ever, due to the lengthy processing time required by Tex-Painter [46], we include only a qualitative comparison for this method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Qualitative comparisons. In Figure 4, we compare our qualitative results with state-of-the-art methods. Our approach generates garment textures with enhanced 3D consistency and refined artistic quality. First, examining front and back views reveals that our method achieves consistent texture across perspectives. For example, in Figure 4-(a), the short-sleeved shirt generated by Paint3D [45] exhibits a \"Janus effect,\" while other methods show inconsistencies. In contrast, textures generated by our method remain coherent. Second, our textures exhibit finer details and organized layouts; for instance, Figure 4-(b) shows intricate lace and embroidery produced by our method. Finally, our approach aligns more closely with the text description, as demonstrated in Figure 4-(c), where our textures incorporate more Peking Opera elements than those of other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.71 + ], + "angle": 0, + "content": "Quantitative comparisons. In Table 1, we present the quantitative comparisons with the previous methods in garment texture synthesis. For the FID [18] and KID [3] calculations, we render \\(512 \\times 512\\) images of each mesh with the synthesized textures, captured from 20 fixed viewpoints. We use the average CLIP score to measure the alignment between rendering images and text prompts. For runtime, we compare inference time for each method, defining inference time as the time it takes to generate a complete texture map for a given text prompt and predefined mesh. From the results, we have the following observation. First, our method obtains the best performance in terms of FID [18] and KID [3]. Notably, it outperforms the leading competitor, Paint3D [45], by an average of 50.43 in FID and 2.23 in KID. These improvements indicate that our method excels at generating high-quality garment textures. Second, our method achieves the best CLIP scores, indicating better text-texture alignment. Finally, we use the same GPU device (a single NVIDIA-V100) for the runtime comparison to infer each method. Our method only takes 4 seconds to generate a UV texture map, which is much less time-consuming than other methods. This proves the efficiency of our method." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.72, + 0.9, + 0.807 + ], + "angle": 0, + "content": "
MethodFID↓KID↓Runtime↓CLIP↑
TEXTure [37]169.014.35~140s27.26
Text2Tex [7]147.482.77~940s27.93
Paint3D [45]147.402.86~220s28.72
Ours96.970.63~4s32.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.816, + 0.905, + 0.844 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison against state-of-the-art methods. We report the FID, \\( \\mathrm{{KID}}\\left( {\\times {10}^{2}}\\right) \\) ,CLIP score and runtime." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "User study. We further conduct a user study to analyze the overall quality of the generated textures. We randomly select 50 garment meshes and corresponding text prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26571" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.087, + 0.868, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.421, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with the state-of-the-art methods. Our method generates garment textures with complex and clear texture details, which are both aesthetically pleasing and maintain 3D consistency. Please enlarge the image for easy observation of details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.598 + ], + "angle": 0, + "content": "to perform the user study. Following the evaluation criteria of Paint3d [45], those meshes are textured in both our method and comparison methods. Participants are requested to evaluate the results based on 1) overall quality and 2) fidelity to the text prompt, using a scale of 1 to 5. We collect the evaluation results from 30 users. As presented in Table 2, we show the average results across all prompts for each method. Our approach outperforms other methods in terms of both overall quality and text fidelity." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.611, + 0.477, + 0.699 + ], + "angle": 0, + "content": "
MethodImage Quality↑Text Fidelity↑
TEXTure [37]3.213.57
Text2Tex [7]3.323.41
Paint3D [45]4.014.17
Ours4.834.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.707, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Table 2. Statistics on user preferences for texture quality. Users more favor our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.327, + 0.76 + ], + "angle": 0, + "content": "4.3. Diversity and Consistency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Garment texture generation demands high diversity and 3D consistency standards to achieve realistic representation. Unlike other texture tasks, garment textures must capture various fabrics, patterns, and colors to reflect the richness of real-world apparel. Garments often feature unique details like intricate stitching, subtle textures, and complex patterns that vary across styles and materials. Ensuring 3D consistency is crucial, as textures must appear seamless and coherent from all angles to maintain realism in virtual envi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.536 + ], + "angle": 0, + "content": "ronments. Any misalignment can break the illusion, making the garment look artificial. Therefore, effective garment texture generation must balance diversity with strong 3D consistency, ensuring textures adapt smoothly to the garment's surface from every view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.735 + ], + "angle": 0, + "content": "To showcase our method's ability to generate diverse and 3D consistent garment textures, we present two examples in Figure 5. Given a UV position map and three text prompts, our method generates three unique texture styles for the same garment mesh. We observe the following: 1) Our method produces semantically rich textures aligned with the text prompts. Each texture displays intricate details and logical layout, illustrating our model's ability to create a variety of garment styles that meet diverse user preferences. 2) We present four different views for each generated texture. We observe a high level of 3D consistency in texture details across these views, which is achieved by our method's ability to generate complete UV texture in UV space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.667, + 0.759 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation of 3D structure guidance. To evaluate the effectiveness of our newly designed 3D structure guidance, we conduct experiments on three baseline configurations: point cloud guidance, UV position map guidance, and mask guidance. Point cloud guidance (3D Info.): This configuration uses only the point cloud representation as guidance without incorporating the 2D UV layout information from the UV position map. UV position map guidance (2D Info.): This configuration uses only the 2D UV layout" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26572" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.862, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.418, + 0.907, + 0.446 + ], + "angle": 0, + "content": "Figure 5. Given a UV position map and varying text prompts, our method can generate garment textures in distinct styles. Observing the four views confirms that the generated textures maintain global 3D consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.53 + ], + "angle": 0, + "content": "information from the UV position map as guidance, without the 3D global information provided by the point cloud. Mask guidance (Mask Info.): Only the UV texture mask is embedded using a VAE [21] encoder as guidance, with no structural information, neither 2D nor 3D." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.541, + 0.437, + 0.628 + ], + "angle": 0, + "content": "
MethodFID↓KID↓
3D Info.187.638.73
2D Info.174.298.07
Mask Info.210.3412.59
Ours96.970.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.635, + 0.483, + 0.678 + ], + "angle": 0, + "content": "Table 3. Evaluation of modules in our method. This demonstrates the effectiveness of each component. Our full model achieves optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 3, we compare three baselines with our method, leading to the following observations: 1) Compared to our full model, 3D Info. provides global 3D information but lacks 2D UV layout details. This results in worse FID and KID scores for 3D Info.. 2) When comparing 2D Info. with our full model, we find that while 2D Info. includes 2D UV layout information, it lacks global 3D information. As in the previous case, 2D Info. shows worse FID and KID scores, emphasizing the importance of 3D structural data. 3) Mask Info. provides only basic UV texture shape information, lacking structural details, making it the weakest performer among the baselines. These observations demonstrate the effectiveness of our method, which leverages global 3D structure information and 2D" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.907, + 0.486 + ], + "angle": 0, + "content": "layout details to provide guidance for improved garment texture generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.499, + 0.64, + 0.515 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.766 + ], + "angle": 0, + "content": "In conclusion, we present TexGarment, a novel and efficient method for generating high-quality garment textures. By leveraging pre-trained 2D image generation models and incorporating 3D garment structure information, TexGarment addresses key challenges in texture generation, such as global 3D consistency and inference speed. Our method enables the generation of diverse and visually consistent UV texture maps in a single inference process, significantly improving on previous multi-stage approaches. Through the introduction of two key components: 3D structure information extraction and 3D structure guidance, we effectively integrate 3D structural details into the 2D texture generation process, resulting in high fidelity and 3D consistency. Our experiments show that TexGarment outperforms existing methods in both quality and performance, with user studies confirming its superior visual appeal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Limitation. TexGarment now does not support image-conditioned generation. In practical applications, users may provide a reference image to generate garment textures in a similar style. We will explore it in future research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under grant 62202369." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26573" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.46, + 0.128 + ], + "angle": 0, + "content": "[1] Josef Albers. Interaction of color. Kyung Dang, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.131, + 0.483, + 0.186 + ], + "angle": 0, + "content": "[2] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.187, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.229, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[4] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8918-8928, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pages 4055-4075, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[8] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Shuhong Chen, Kevin Zhang, Yichun Shi, Heng Wang, Yiheng Zhu, Guoxian Song, Sizhe An, Janus Kristjansson, Xiao Yang, and Matthias Zwicker. Panic-3d: Stylized single-view 3d reconstruction from portraits of anime characters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21068-21077, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[12] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In CVPR, pages 1465-1474, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. *Internvl: Scaling up vision foundation models and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.192 + ], + "angle": 0, + "content": "[14] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[20] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.575 + ], + "angle": 0, + "content": "[21] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.906, + 0.632 + ], + "angle": 0, + "content": "[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.717 + ], + "angle": 0, + "content": "[23] Jialun Liu, Chenming Wu, Xinqi Liu, Xing Liu, Jinbo Wu, Haotian Peng, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Texoct: Generating textures of 3d models with octree-based diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4284-4293, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.746 + ], + "angle": 0, + "content": "[24] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[25] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26574" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[28] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[29] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.355 + ], + "angle": 0, + "content": "[30] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[37] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[41] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In European Conference on Computer Vision, pages 72-88. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[42] Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. Cycle-consistent inverse gan for text-to-image synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 630-638, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[43] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.443 + ], + "angle": 0, + "content": "[44] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206-4216, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[45] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[46] Hongkun Zhang, Zherong Pan, Congyi Zhang, Lifeng Zhu, and Xifeng Gao. Texpainter: Generative mesh texturing with multi-view consistency. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26575" + } + ] +] \ No newline at end of file diff --git a/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_origin.pdf b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3c26554e05606c6a0be84b257beffbc4aa859ef6 --- /dev/null +++ b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/a4a25bb5-b3b2-4fab-baad-be3f4246d157_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99b7f8061258fc00ef82599d02ca7954491d675717f39449c12e0a6bfc63424b +size 5954612 diff --git a/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/full.md b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..21c670245a35dcfe840ae2e8cbca74931c4bd624 --- /dev/null +++ b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/full.md @@ -0,0 +1,246 @@ +# TexGarment: Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer + +Jialun Liu $^{1}$ , Jinbo Wu $^{1*}$ , Xiaobo Gao $^{1}$ , Jiakui Hu $^{2}$ , Bojun Xiong $^{2}$ , Xing Liu $^{1}$ , Chen Zhao $^{1}$ , Hongbin Pei $^{3}$ , Haocheng Feng $^{1}$ , Yingying Li $^{1}$ , Errui Ding $^{1}$ , Jingdong Wang $^{1}$ $^{1}$ Department of Computer Vision Technology(VIS), Baidu Inc. + $^{2}$ Peking University $^{3}$ Xi'an Jiaotong University liujialun@baidu.com + +![](images/b2364eb4ef4e2eb568925731de77cee22b3c46c7c7f6f891c65cc37a98f46552.jpg) +Figure 1. A gallery showcasing generated garment textures by TexGarment. Our method achieves globally consistent, high-quality, and high-fidelity textures across a wide range of garment styles. The average time to generate each texture map is under 4 seconds. + +# Abstract + +This paper introduces TexGarment, an efficient method for synthesizing high-quality, 3D-consistent garment textures in UV space. Traditional approaches based on 2D-to-3D mapping often suffer from 3D inconsistency, while methods learning from limited 3D data lack sufficient texture diversity. These limitations are particularly problematic in garment texture generation, where high demands exist for both detail and variety. To address these challenges, TexGarment leverages a pre-trained text-to-image diffusion Transformer model with robust generalization capabilities, introducing structural information to guide the model in generating 3D-consistent garment textures in a single inference step. Specifically, We utilize the 2D UV position map to + +guide the layout during the UV texture generation process, ensuring a coherent texture arrangement and enhancing it by integrating global 3D structural information from the mesh surface point cloud. This combined guidance effectively aligns 3D structural integrity with 2D layout. Our method efficiently generates high-quality, diverse UV textures in a single inference step while maintaining 3D consistency. Experimental results validate the effectiveness of TexGarment, achieving state-of-the-art performance in 3D garment texture generation. + +# 1. Introduction + +The demand for 3D garment visualization is growing across virtual try-ons for e-commerce, video games, 3D films, and AR/VR environments. These industries need large volumes + +of high-quality 3D garments to meet consumer expectations. While 2D fashion images are widely available and generative AI has simplified 2D image creation, producing high-quality 3D garment textures remains challenging, often requiring skilled artists and weeks of work. This highlights the need for efficient 3D garment generation methods. + +The aesthetic appeal of 3D garments is deeply influenced by their texture quality, as Joseph Albers [1] said: "Texture is the soul of painting; it gives surface depth and personality", effectively capturing this concept. In 3D computer graphics, textures are essential for realism, adding detailed surface elements that make virtual objects lifelike. For 3D garments, high-quality textures are particularly important, capturing fabric nuances, stitching, folds, and shadows that make virtual clothing look authentic. Without these textures, even well-modeled garments can appear flat or artificial, reducing the immersive quality of 3D applications. + +Despite substantial advancements in 2D image synthesis and 3D shape generation using GANs [40, 42] and diffusion models [12, 38, 39], achieving high-quality garment texture generation remains challenging. This difficulty arises from three primary factors: Global 3D Consistency. Methods employing 2D diffusion models, such as TexPainter [46], TEXTure [37] and Text2Tex [7], have made strides in texture synthesis by utilizing pre-trained LDM [38], which support strong generalization and diverse texture generation. However, text-to-image models often exhibit biases that limit their effectiveness for 3D applications, such as favoring frontal views and lacking a robust understanding of 3D structures. This frequently results in texture inconsistencies across different viewpoints of the same 3D object, ranging from minor seam misalignments to severe issues, e.g., "Janus effect". Generalization Ability. Methods learning from limited 3D data, such as TexOct [23], PointUV [44], and Mesh2Tex [4], generate textures by explicitly integrating 3D shape information, resulting in high 3D consistency. However, these approaches often lack generalization ability, limiting their effectiveness when generating textures for diverse 3D objects beyond the training set or adapting to a range of text and visual prompts. Inference Speed. To achieve full shape coverage, prior methods [2, 37, 44, 45] rely on iterative generation processes, requiring multiple steps or stages. This results in prolonged inference times, often lasting several minutes, and requires significant computational resources, limiting practical use for designers and content creators. + +In this work, we present TexGarment, a method tailored for generating high-quality garment textures. Given text prompts, TexGarment produces consistent, diverse, and visually striking textures in a single inference pass. A widely used approach for handling texture maps is UV representation, which maps 3D textures accurately onto 2D UV images through UV coordinate transformations. This tech- + +nique bridges the gap between 3D texture generation and 2D image methods, motivating our use of pre-trained image generation models [8] to synthesize diverse textures effectively within the 2D UV space. + +Our key insight is to incorporate 3D garment structure information as guidance within a pre-trained 2D image generation model [8], leveraging its generalization capabilities to produce consistent and diverse UV texture maps directly in the 2D UV space through only one inference process. Specifically, our method introduces two main components: 1) 3D Structure Information Extraction and 2) 3D Structure Guidance. First, given a garment mesh, we use rasterization to obtain a UV position map, where each pixel encodes the XYZ coordinates of the corresponding point on the mesh surface. This UV position map provides a 2D projection of the original 3D point cloud, capturing the basic layout of the UV texture. However, regions that are adjacent in 3D space may be separated in the UV map, causing a loss of global 3D structure information. This makes it challenging for a 2D image generation model to capture global 3D structures solely from the 2D UV position map. To address this, we introduce a 3D structure information extraction module that injects global structure information, represented by 3D point clouds, into the 2D UV position map. This process generates a 3D structure latent, which combines both 3D structure and 2D layout information, serving as guidance for the pre-trained image generation model. Second, we implement a 3D structure guidance module, analogous to text guidance, which injects the 3D structure information into the pre-trained model via cross-attention. Using a powerful text-to-image diffusion model [8] as our basic architecture, our design enables the model to receive guidance from both text and 3D structure, allowing it to generate high-quality UV texture maps in a single inference process, achieving both text fidelity and 3D consistency. + +Our contributions are summarized as follows: 1) We propose TexGarment, a one-stage method for garment texture generation that produces high-quality, diverse UV texture maps for untextured garment meshes in a single inference process and completes the process in 4 seconds. 2) We introduce two novel modules to incorporate 3D garment structure information into the 2D image generation model, providing 3D-consistent guidance for garment texture generation. Extensive experiments validate the effectiveness of these modules in achieving 3D consistency. 3) Qualitative and quantitative evaluations demonstrate that our method outperforms state-of-the-art approaches in garment texture generation. Additionally, a user study shows that users prefer textures generated by our method. + +# 2. Related Works + +Text-to-Image Generation The field of text-to-image generation has seen considerable progress. Early research used + +generative adversarial networks (GANs)[17] to create high-quality images, using image classes[15] and segmentation maps [30] as prompts. Recently, attention has shifted to architectural improvements that enhance generation quality with scaling Transformers, as shown in models like DALL-E [36], CogView [14], Parti [43], and Muse [6]. Another popular category of text-to-image generators is based on pixel space or latent diffusion models, including Imagegen [39], Stable Diffusion [38], SDXL [32], DiT [31], and PixArt [8]. Notably, DiT [31] advanced text-to-image generation by introducing Transformers to latent diffusion models (LDMs). PixArt [8] further optimized training methods, enabling faster training of scaled LDMs. + +3D Generation via 2D generation model. With the development of text-to-image generation techniques, many 3D texturing techniques seek to leverage the prior knowledge of 2D generative models to improve 3D generation models' optimization. Previous optimization-based texture techniques [11, 20, 25, 27, 28] concentrated on improving the texture maps of 3D models by utilizing the large-scale visual language model, e.g. CLIP [34], before the advent of LDMs. DreamFusion [33] introduced the score distillation sampling (SDS) to 3D Generation, which laid the groundwork for many text-to-3D methods. Latent-nerf [26] and Fantasia3D [9] extended SDS to optimize texture mapping. Recently, TexPainter [46] performs multi-view fusion within the 2D image's color space to address inconsistency issues. However, optimization-based texture methods are hindered by prolonged training times. As an alternative, TEXTure [37], Text2Tex [7], and Paint3D [45] involve directly generating texture images from multiple views of 3D meshes and subsequently painting them onto the 3D meshes. TEXTure [37] and Text2Tex [7] adhere to this concept and suggest modifying the overlapping segments of the texture image from the previous view using results from each new camera pose. Nevertheless, these methods encounter various artifacts, including seams, noise, and irrelevant fragments. Paint3D [45] develops a specialized model to paint incomplete regions during the texture painting. + +Generative Texturing from 3D Data. Additionally, generative texture models are capable of training models from scratch using 3D data. To assign a color to each pixel on the surface of a three-dimensional shape, early methods [16, 29] relied on implicit texture fields. However, given that the texture on the surface of a 3D mesh is continuous, it is unlikely that discrete supervision will result in the training of a model capable of synthesizing high-quality textures. The Texturify [41] is designed to construct a texture map on the surface of a polygonal mesh. Mesh2tex [4] builds on this by integrating an implicit texture field branch for enhancement. TexOct [23] use point clouds to help circumvent variations arising from diverse mesh topologies and UV mappings. Other methods [12, 44] synthesize UV + +maps for 3D meshes, but are ill-equipped to handle more generic objects, due to the inherent diversity between 3D meshes belonging to various categories. + +Discussion with previous work about the UV position map. Previous works, such as PointUV [44], Paint3D [45], and TextureGen [2], also incorporate the UV position map into their models. We outline the key differences between our approach and these methods: 1) PointUV [44] uses the UV position map, along with other maps like normal and mask maps, in the first stage to generate a coarse UV texture. However, this method is limited to simple textures for objects from the ShapeNet [5] dataset. In contrast, our approach uses only the UV texture map as a guide, enabling the generation of complex, high-quality textures for a wide range of garment categories. 2) Paint3D [45] and TextureGen [2] use the UV position map as a 2D condition in a second-stage inpainting model to refine textures. Our method directly applies the UV position map to guide the entire texture generation, leveraging both 2D layout and global 3D structure, resulting in better textures as shown in our experiments. 3) Unlike these prior methods, our model is a single-stage generator. Given a UV position map and a text prompt, our approach generates high-quality garment texture maps in a single, efficient inference step, while previous methods require a more complex two-stage process. + +# 3. Method + +Given a 3D garment model and a target texture description, TexGarment efficiently synthesizes high-quality, diverse, and consistent garment UV textures. TexGarment is a Transformer-based diffusion model, as shown in Figure 2. We build on a large-scale pre-trained image generation model [8], leveraging its robust prior knowledge to enhance our model's generalization capability. To ensure global 3D consistency, we first design a 3D structure information extraction module. For an untextured garment mesh, we apply rasterization to create a UV position map that captures the basic UV texture layout and encodes the 3D coordinates of points on the mesh surface. The structure information extraction module then embeds global structure information from the original point cloud into the UV position map, resulting in a 3D structure latent that combines global 3D structure information with 2D layout details. This 3D structure latent is then used by a 3D structure guidance module to guide the texture generation. Through this innovative design, our method generates high-quality and 3D-consistent texture maps in a single inference process, without the need for additional operations such as inpainting. The following sections provide the overview of each stage. + +# 3.1. Data Processing + +Dataset. We collect a comprehensive dataset from VRoidHub based on the guidelines provided by Panic3D[10], en + +![](images/c788b14c53382f84ae633e18e651147b78b88650ba3d474ea91ef4f2b28eb79f.jpg) +(a) Overview +(b) 3D Structure Information Extraction +Figure 2. Pipeline of Our Method: (a) Overview: The training process begins with embedding the input text prompt and timestep using the T5 model [35] and an MLP layer, generating the text feature and time embedding. The UV texture is encoded by a VAE [21] encoder to create a texture latent representation, which is then noised to produce a noisy latent. These representations are processed through several Transformer blocks. (b) 3D Structure Information Extraction Module: The UV position map is encoded by a VAE [21] encoder to obtain the UV position latent, while a point cloud feature extraction module captures the global 3D structure. This 3D information is integrated into the UV position latent via cross-attention, resulting in the 3D structure latent, which combines 3D global structure and 2D UV texture layout. (c) Transformer Block Details: Based on a text-to-image Transformer diffusion model [8], we introduce a 3D structure guidance module that uses the 3D structure latent as guidance, injecting it into the model through cross-attention to enhance 3d structure consistency. + +![](images/f8fe135bcacd9cd7da80f02aba3e789827379c51393a177ddc3ad323c8e81257.jpg) +(c) Block with 3D sturcture guidance + +compassing a diverse array of character models. Each model in VRoidHub includes body, clothing, and animation components. Using Blender, we segment the models by style type, preserving clothing elements while removing non-clothing components. Pants and tops were separated into distinct parts, while dresses were preserved as single units. Following segmentation, we refine the dataset by excluding samples with inconsistent color distributions and monochromatic appearances, as well as items with accessories such as ties, belts, and other extraneous parts, ultimately retaining a dataset of 8,000 clothing items. + +Rasterization. In a UV texture image, the RGB information of the surface mesh is stored, and during rendering, a rasterization operator samples the color from the UV texture image onto the rendered image. A UV position map stores the XYZ coordinates of the mesh, which requires sampling mesh properties and baking them into the UV map. To achieve this, we assume the UV layout represents a 3D mesh with vertices $[u,v,z]$ , where $u$ and $v$ are the original mesh's UV coordinates and $z$ is constant, making it a planar mesh. Each $[u,v]$ corresponds to a vertex in the original 3D mesh. We retain the face relationships from + +the original mesh, and by performing rasterization, we interpolate the "UV mesh" positions to determine each pixel's XYZ coordinates on the UV position map. The shapes of garment meshes in the VRoid data are also defined by an alpha channel within the UV texture. Before rendering our images, we rasterize the UV position map using the alpha channel to isolate the rendered mesh positions. We compute the mesh's scale and translation with these selected mesh positions. We then translate the mesh to align with the origin of the coordinate system and scale it to fit within the $[-1, 1]$ range. This process ensures that the rendered portion of each mesh remains centered in the final image. After translating the mesh, we re-rasterize the UV position map to align it within the $[-1, 1]$ range. + +We render all garment data from two perspectives: front and back. All images are generated using Nvdiffrast [22] under uniform ambient lighting. These two views are then used to generate prompts by InternVL[13]. A toy case is shown in Figure. 3. + +# 3.2. Basic Architecture + +We adopt PixArt- $\alpha$ [8] as our foundational image generation model, a Transformer-based text-to-image diffusion + +"The top is short-sleeved with a white base and a black collar and cuffs, with a black bow at the center. The skirt is black with white cross motifs." + +![](images/a6b8ef62f99ae9ccb36648e90293f1365d6cac4a860653f5c537ada8f4b14174.jpg) +Figure 3. From left to right, we show vvoid data, garment, UV texture image and UV position map, respectively. + +model that excels in image quality, artistry, and semantic control. Our choice of PixArt- $\alpha$ as the base architecture is motivated by three key reasons: + +- Strong Prior Foundation. PixArt- $\alpha$ [8] is a popular text-to-image diffusion model that offers valuable insights to the AIGC community. Its robust image generation capabilities provide a solid prior for our method, enhancing generalization and quality. +- Efficiency and Resource Friendliness. With an optimized Transformer design, PixArt- $\alpha$ [8] achieves high-quality image generation using only 0.6 billion parameters, minimizing parameters and training requirements +- Enhanced Artistic Style. Fine-tuned on high-quality aesthetic data, PixArt- $\alpha$ [8] generates images with an ideal artistic style for garment texture generation, where visual appeal and stylistic quality are essential. + +The model takes as input a text prompt, timestep $t$ , and clean images (clear UV texture $\mathcal{I}_{\mathrm{tex}}$ in our method). These inputs are processed through T5 [35], MLP, and VAE [21] encoder to obtain text, time, and latent embeddings. These embeddings are then fed into Transformer blocks to execute the diffusion process. Each Transformer block consists of self-attention, adaptive normalization (adaLN-single), cross-attention, and feedforward layers. Cross-attention uses the text embedding to interact with the Transformer's hidden states, while adaLN-single leverages the time embedding to learn scale and shift parameters. + +# 3.3. 3D consistent Garment UV Texture Generation + +UV representation is a widely adopted method for creating texture maps, enabling precise mapping of 3D textures onto 2D images by "unwrapping" the 3D surface into a 2D plane. This process assigns each point on the 3D model to a specific location in 2D space, ensuring that every detail of the 3D surface accurately corresponds to a position on the UV map. This approach effectively bridges 3D texture generation and 2D image-based methods, allowing advanced 2D generative models to handle tasks that traditionally require 3D-specific techniques. Working within the 2D UV space enables high-quality texture generation that can be seamlessly applied back to the 3D model, providing a flexible and efficient solution for producing consistent textures. However, adjacent regions in 3D space may become non-adjacent in the UV map, leading to a loss of global 3D + +structural information during mapping. To address this, we design a 3D structure information extraction module and a 3D structure guidance module, which integrate the original 3D structural information from the garment mesh into the UV texture generation process, ensuring 3D consistency in the final UV texture maps. + +3D Structure Information Extraction. For a given 3D garment mesh, we use rasterization to obtain the UV position map $\mathcal{I}_{\mathrm{pos}}$ . Each active pixel in $\mathcal{I}_{\mathrm{pos}}$ encodes the XYZ coordinates of a corresponding point on the garment mesh surface, creating a 2D representation of the 3D point cloud and capturing the basic layout of the UV texture. As shown in Figure 2-(b), $\mathcal{I}_{\mathrm{pos}}$ is first processed by a VAE [21] encoder to obtain the UV position latent, $\mathcal{F}_{\mathrm{pos}}$ . While $\mathcal{F}_{\mathrm{pos}}$ embeds the UV layout in 2D space, it lacks 3D structure information. To address this, we retrieve the 3D point cloud $\mathcal{P}$ from $\mathcal{I}_{\mathrm{pos}}$ , normalize it to the range [0, 1], and project it into a 3D context representation space for effective interaction with $\mathcal{F}_{\mathrm{pos}}$ . Specifically, we encode the point cloud $\mathcal{P}$ into a latent representation $\mathcal{F}_{\mathrm{point}}$ using a learnable embedding function and employ a cross-attention module to infuse $\mathcal{F}_{\mathrm{pos}}$ with 3D structure information from the point cloud: + +$$ +\mathcal {F} _ {\text {p o i n t}} = \operatorname {P o s E m b} (\mathcal {P}), \tag {1} +$$ + +$$ +\mathcal {F} _ {\text {l a t e n t}} ^ {3 \mathrm {d}} = \operatorname {F F N} \left(\operatorname {S o f t m a x} \left(\frac {\mathcal {Q} _ {\text {p o s}} \mathcal {K} _ {\text {p o i n t}} ^ {\mathrm {T}}}{\sqrt {d}}\right) \mathcal {V} _ {\text {p o i n t}}\right), \tag {2} +$$ + +where $\mathcal{F}_{\mathrm{point}}$ is the embedding of point cloud. $\mathcal{Q}_{\mathrm{pos}} = \mathcal{F}_{\mathrm{pos}}\mathcal{W}_{\mathrm{Q}}^{\mathrm{pos}}$ is the query embedding of $\mathcal{F}_{\mathrm{pos}}$ . $\mathcal{K}_{\mathrm{point}} = \mathcal{F}_{\mathrm{point}}\mathcal{W}_{\mathrm{K}}^{\mathrm{point}}$ and $\mathcal{V}_{\mathrm{point}} = \mathcal{F}_{\mathrm{point}}\mathcal{W}_{\mathrm{V}}^{\mathrm{point}}$ are the key and value embeddings of $\mathcal{F}_{\mathrm{point}}$ , respectively. $\mathcal{F}_{\mathrm{latent}}^{\mathrm{3d}}$ is the 3D structure latent that combines both the 2D layout information and the global 3D structure information. + +3D Structure Information Guidance. We incorporate the 3D structure guidance $\mathcal{F}_{\mathrm{latent}}^{3\mathrm{d}}$ into the pre-trained text-to-image image diffusion process to guide the 3D-consistent UV texture generation. As shown in Figure 2-(c), we denote the transformer intermediate feature as $\mathcal{F}_{\mathrm{in}}$ and the text embedding as $\mathcal{F}_{\mathrm{text}}$ . Subsequently, we inject the 3D structure guidance $\mathcal{F}_{\mathrm{latent}}^{3\mathrm{d}}$ into the image diffusion process by interacting with $\mathcal{F}_{\mathrm{in}}$ through the cross-attention: + +$$ +\mathcal {F} _ {\text {o u t}} = \operatorname {S o f t m a x} \left(\frac {\mathcal {Q} _ {\text {i n}} \mathcal {K} _ {\text {3 d}} ^ {\mathrm {T}}}{\sqrt {d}}\right) \mathcal {V} _ {\text {3 d}} + \operatorname {S o f t m a x} \left(\frac {\mathcal {Q} _ {\text {i n}} \mathcal {K} _ {\text {t e x t}} ^ {\mathrm {T}}}{\sqrt {d}}\right) \mathcal {V} _ {\text {t e x t}}, \tag {3} +$$ + +where $\mathcal{Q}_{\mathrm{in}} = \mathcal{F}_{\mathrm{in}}\mathcal{W}_{\mathrm{Q}}$ is the query embedding of the $\mathcal{F}_{\mathrm{in}}$ . $\mathcal{K}_{3\mathrm{d}} = \mathcal{F}_{\mathrm{latent}}^{3\mathrm{d}}\mathcal{W}_{\mathrm{K}},\mathcal{V}_{3\mathrm{d}} = \mathcal{F}_{\mathrm{latent}}^{3\mathrm{d}}\mathcal{W}_{\mathrm{V}}$ are the key and value embeddings of $\mathcal{F}_{\mathrm{latent}}^{3\mathrm{d}}$ , respectively. $\mathcal{K}_{\mathrm{text}} = \mathcal{F}_{\mathrm{text}}\mathcal{W}_{\mathrm{K}'}$ , $\mathcal{V}_{\mathrm{text}} = \mathcal{F}_{\mathrm{text}}\mathcal{W}_{\mathrm{V}'}$ are the key, and value embeddings of $\mathcal{F}_{\mathrm{text}}$ respectively. $\mathcal{W}_{\mathrm{Q}},\mathcal{W}_{\mathrm{K}},\mathcal{W}_{\mathrm{K}'}$ , $\mathcal{W}_{\mathrm{V}}$ and $\mathcal{W}_{\mathrm{V}'}$ are the projection matrices. The two cross-attention results are added to obtain the $\mathcal{F}_{\mathrm{out}}$ . + +Given the text prompt condition $c_{\mathrm{text}}$ from $\mathcal{F}_{\mathrm{text}}$ and 3D structure condition $c_{3\mathrm{d}}$ from $\mathcal{F}_{\mathrm{latent}}^{3\mathrm{d}}$ , we apply the classifier-free guidance [19] strategy to incorporate the information and our training objective is: + +$$ +\mathcal {L} _ {\text {d i f f}} = \mathbb {E} _ {\boldsymbol {x} \sim p, \epsilon \sim \mathcal {N} (0, I), c _ {\text {t e x t}}, c _ {3 \mathrm {d}}, t} \left[ \| \epsilon - \epsilon_ {\theta} \left(\boldsymbol {x} _ {t}, t, c _ {\text {t e x t}}, c _ {3 \mathrm {d}}\right) \| _ {2} ^ {2} \right], \tag {4} +$$ + +where $\boldsymbol{x}_t$ is the noise latent from the ground-truth $\mathcal{I}_{\mathrm{tex}}$ . + +After training, the diffusion model can generate diverse and 3D-consistent UV garment texture maps in a single inference process. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. Our dataset consists of 8k of textured 3D garments from internal series. Garment styles include long-sleeved, short-sleeved, long skirts, short skirts, long pants, shorts, and other common styles. The dataset is divided into two parts: the training set containing 7.9k garments and the testing set containing 0.1k garments. + +Implementation Details. We use PixArt- $\alpha$ [8] as our pre-trained text-to-image generation model. The model is trained for 50,000 iterations with the AdamW [24] optimizer, using a fixed learning rate of $2 \times 10^{-6}$ . Newly added parameters are initialized to zero and trained with a learning rate of $2 \times 10^{-5}$ . The batch size is set to 32. For all experiments, we sample 5k points from the UV position map and synthesize a UV texture map at a resolution of $512 \times 512$ . + +Baselines. We compare our method with several state-of-the-art approaches, including TEXTure [37], Text2Tex [7], Paint3D [45], and TexPainter [46]. TEXTure [37] introduces an iterative texture generation strategy to enhance texture manipulation. Text2Tex [7] builds on this by incorporating an automatic viewpoint selection process within each iteration. TexPainter [46] utilizes a modified denoising diffusion implicit model (DDIM) to merge textures across multiple views in color space. Paint3D [45] is a coarse-to-fine texture synthesis approach. First, it samples multi-view images from pre-trained 2D diffusion models, then refines the textures in UV space using a diffusion model conditioned on both the position map and the coarse texture. + +Evaluation metrics. We utilize commonly used metrics for evaluating the performance of generative models, specifically, the Fréchet Inception Distance (FID) [18] and Kernel Inception Distance (KID) [3], which measure image quality and image diversity. Additionally, we use the CLIP score to measure the alignment between the rendered image and the input text prompt. Finally, we also evaluate the texture generation speed. + +# 4.2. Comparison with State-of-the-art Methods + +In this section, we compare our method with state-of-the-art methods both qualitatively and quantitatively. How- + +ever, due to the lengthy processing time required by Tex-Painter [46], we include only a qualitative comparison for this method. + +Qualitative comparisons. In Figure 4, we compare our qualitative results with state-of-the-art methods. Our approach generates garment textures with enhanced 3D consistency and refined artistic quality. First, examining front and back views reveals that our method achieves consistent texture across perspectives. For example, in Figure 4-(a), the short-sleeved shirt generated by Paint3D [45] exhibits a "Janus effect," while other methods show inconsistencies. In contrast, textures generated by our method remain coherent. Second, our textures exhibit finer details and organized layouts; for instance, Figure 4-(b) shows intricate lace and embroidery produced by our method. Finally, our approach aligns more closely with the text description, as demonstrated in Figure 4-(c), where our textures incorporate more Peking Opera elements than those of other methods. + +Quantitative comparisons. In Table 1, we present the quantitative comparisons with the previous methods in garment texture synthesis. For the FID [18] and KID [3] calculations, we render $512 \times 512$ images of each mesh with the synthesized textures, captured from 20 fixed viewpoints. We use the average CLIP score to measure the alignment between rendering images and text prompts. For runtime, we compare inference time for each method, defining inference time as the time it takes to generate a complete texture map for a given text prompt and predefined mesh. From the results, we have the following observation. First, our method obtains the best performance in terms of FID [18] and KID [3]. Notably, it outperforms the leading competitor, Paint3D [45], by an average of 50.43 in FID and 2.23 in KID. These improvements indicate that our method excels at generating high-quality garment textures. Second, our method achieves the best CLIP scores, indicating better text-texture alignment. Finally, we use the same GPU device (a single NVIDIA-V100) for the runtime comparison to infer each method. Our method only takes 4 seconds to generate a UV texture map, which is much less time-consuming than other methods. This proves the efficiency of our method. + +
MethodFID↓KID↓Runtime↓CLIP↑
TEXTure [37]169.014.35~140s27.26
Text2Tex [7]147.482.77~940s27.93
Paint3D [45]147.402.86~220s28.72
Ours96.970.63~4s32.37
+ +Table 1. Quantitative comparison against state-of-the-art methods. We report the FID, $\mathrm{{KID}}\left( {\times {10}^{2}}\right)$ ,CLIP score and runtime. + +User study. We further conduct a user study to analyze the overall quality of the generated textures. We randomly select 50 garment meshes and corresponding text prompts + +![](images/0536be6d5f5f93adb376fb88cdd2870dba237480c0413d2a428f2010c323efae.jpg) +Figure 4. Qualitative comparison with the state-of-the-art methods. Our method generates garment textures with complex and clear texture details, which are both aesthetically pleasing and maintain 3D consistency. Please enlarge the image for easy observation of details. + +to perform the user study. Following the evaluation criteria of Paint3d [45], those meshes are textured in both our method and comparison methods. Participants are requested to evaluate the results based on 1) overall quality and 2) fidelity to the text prompt, using a scale of 1 to 5. We collect the evaluation results from 30 users. As presented in Table 2, we show the average results across all prompts for each method. Our approach outperforms other methods in terms of both overall quality and text fidelity. + +
MethodImage Quality↑Text Fidelity↑
TEXTure [37]3.213.57
Text2Tex [7]3.323.41
Paint3D [45]4.014.17
Ours4.834.92
+ +Table 2. Statistics on user preferences for texture quality. Users more favor our method. + +# 4.3. Diversity and Consistency + +Garment texture generation demands high diversity and 3D consistency standards to achieve realistic representation. Unlike other texture tasks, garment textures must capture various fabrics, patterns, and colors to reflect the richness of real-world apparel. Garments often feature unique details like intricate stitching, subtle textures, and complex patterns that vary across styles and materials. Ensuring 3D consistency is crucial, as textures must appear seamless and coherent from all angles to maintain realism in virtual envi + +ronments. Any misalignment can break the illusion, making the garment look artificial. Therefore, effective garment texture generation must balance diversity with strong 3D consistency, ensuring textures adapt smoothly to the garment's surface from every view. + +To showcase our method's ability to generate diverse and 3D consistent garment textures, we present two examples in Figure 5. Given a UV position map and three text prompts, our method generates three unique texture styles for the same garment mesh. We observe the following: 1) Our method produces semantically rich textures aligned with the text prompts. Each texture displays intricate details and logical layout, illustrating our model's ability to create a variety of garment styles that meet diverse user preferences. 2) We present four different views for each generated texture. We observe a high level of 3D consistency in texture details across these views, which is achieved by our method's ability to generate complete UV texture in UV space. + +# 4.4. Ablation Study + +Evaluation of 3D structure guidance. To evaluate the effectiveness of our newly designed 3D structure guidance, we conduct experiments on three baseline configurations: point cloud guidance, UV position map guidance, and mask guidance. Point cloud guidance (3D Info.): This configuration uses only the point cloud representation as guidance without incorporating the 2D UV layout information from the UV position map. UV position map guidance (2D Info.): This configuration uses only the 2D UV layout + +![](images/72a602140849f92ccd282aaee9e5c07d471a1caa187104e3d42bfef06fb0eb8c.jpg) +Figure 5. Given a UV position map and varying text prompts, our method can generate garment textures in distinct styles. Observing the four views confirms that the generated textures maintain global 3D consistency. + +information from the UV position map as guidance, without the 3D global information provided by the point cloud. Mask guidance (Mask Info.): Only the UV texture mask is embedded using a VAE [21] encoder as guidance, with no structural information, neither 2D nor 3D. + +
MethodFID↓KID↓
3D Info.187.638.73
2D Info.174.298.07
Mask Info.210.3412.59
Ours96.970.63
+ +Table 3. Evaluation of modules in our method. This demonstrates the effectiveness of each component. Our full model achieves optimal performance. + +As shown in Table 3, we compare three baselines with our method, leading to the following observations: 1) Compared to our full model, 3D Info. provides global 3D information but lacks 2D UV layout details. This results in worse FID and KID scores for 3D Info.. 2) When comparing 2D Info. with our full model, we find that while 2D Info. includes 2D UV layout information, it lacks global 3D information. As in the previous case, 2D Info. shows worse FID and KID scores, emphasizing the importance of 3D structural data. 3) Mask Info. provides only basic UV texture shape information, lacking structural details, making it the weakest performer among the baselines. These observations demonstrate the effectiveness of our method, which leverages global 3D structure information and 2D + +layout details to provide guidance for improved garment texture generation. + +# 5. Conclusions + +In conclusion, we present TexGarment, a novel and efficient method for generating high-quality garment textures. By leveraging pre-trained 2D image generation models and incorporating 3D garment structure information, TexGarment addresses key challenges in texture generation, such as global 3D consistency and inference speed. Our method enables the generation of diverse and visually consistent UV texture maps in a single inference process, significantly improving on previous multi-stage approaches. Through the introduction of two key components: 3D structure information extraction and 3D structure guidance, we effectively integrate 3D structural details into the 2D texture generation process, resulting in high fidelity and 3D consistency. Our experiments show that TexGarment outperforms existing methods in both quality and performance, with user studies confirming its superior visual appeal. + +Limitation. TexGarment now does not support image-conditioned generation. In practical applications, users may provide a reference image to generate garment textures in a similar style. We will explore it in future research. + +Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under grant 62202369. + +# References + +[1] Josef Albers. Interaction of color. Kyung Dang, 2013. 2 +[2] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 2, 3 +[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6 +[4] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8918-8928, 2023. 2, 3 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3 +[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pages 4055-4075, 2023. 3 +[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 2, 3, 6, 7 +[8] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations. 2, 3, 4, 5, 6 +[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 3 +[10] Shuhong Chen, Kevin Zhang, Yichun Shi, Heng Wang, Yiheng Zhu, Guoxian Song, Sizhe An, Janus Kristjansson, Xiao Yang, and Matthias Zwicker. Panic-3d: Stylized single-view 3d reconstruction from portraits of anime characters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21068-21077, 2023. 3 +[11] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 3 +[12] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In CVPR, pages 1465-1474, 2022. 2, 3 +[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. *Internvl: Scaling up vision foundation models and + +aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 4 +[14] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 3 +[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3 +[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 6 +[20] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3 +[21] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4, 5, 8 +[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 4 +[23] Jialun Liu, Chenming Wu, Xinqi Liu, Xing Liu, Jinbo Wu, Haotian Peng, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Texoct: Generating textures of 3d models with octree-based diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4284-4293, 2024. 2, 3 +[24] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[25] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 3 +[26] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 3 + +[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3 +[28] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3 +[29] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 3 +[30] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 3 +[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3 +[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations. 3 +[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations. 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 4, 5 +[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3 +[37] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 2, 3, 6, 7 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3 +[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, + +Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3 +[40] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2 +[41] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In European Conference on Computer Vision, pages 72-88. Springer, 2022. 3 +[42] Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. Cycle-consistent inverse gan for text-to-image synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 630-638, 2021. 2 +[43] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research. 3 +[44] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206-4216, 2023. 2, 3 +[45] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 2, 3, 6, 7 +[46] Hongkun Zhang, Zherong Pan, Congyi Zhang, Lifeng Zhu, and Xifeng Gao. Texpainter: Generative mesh texturing with multi-view consistency. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6 \ No newline at end of file diff --git a/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/images.zip b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7be7bf7220d42b3a3a919ad1e06cb6822f6d6dfb --- /dev/null +++ b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bd0261574a190a800932012e37d0d4215973b8cbc9595346ccc1bc44a0f7a3e +size 546653 diff --git a/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/layout.json b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..413da83d181bf772929241b2cfe41962ac0af08d --- /dev/null +++ b/2025/TexGarment_ Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer/layout.json @@ -0,0 +1,7038 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 500, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 500, + 138 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 500, + 138 + ], + "type": "text", + "content": "TexGarment: Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": "Jialun Liu" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Jinbo Wu" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Xiaobo Gao" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Jiakui Hu" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Bojun Xiong" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Xing Liu" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Chen Zhao" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Hongbin Pei" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Haocheng Feng" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Yingying Li" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Errui Ding" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": ", Jingdong Wang" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": "Department of Computer Vision Technology(VIS), Baidu Inc. \n" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": "Peking University " + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 521, + 233 + ], + "type": "text", + "content": "Xi'an Jiaotong University liujialun@baidu.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 70, + 247, + 541, + 473 + ], + "blocks": [ + { + "bbox": [ + 70, + 247, + 541, + 473 + ], + "lines": [ + { + "bbox": [ + 70, + 247, + 541, + 473 + ], + "spans": [ + { + "bbox": [ + 70, + 247, + 541, + 473 + ], + "type": "image", + "image_path": "b2364eb4ef4e2eb568925731de77cee22b3c46c7c7f6f891c65cc37a98f46552.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 481, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 54, + 481, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 481, + 555, + 506 + ], + "type": "text", + "content": "Figure 1. A gallery showcasing generated garment textures by TexGarment. Our method achieves globally consistent, high-quality, and high-fidelity textures across a wide range of garment styles. The average time to generate each texture map is under 4 seconds." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 516, + 201, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 516, + 201, + 529 + ], + "spans": [ + { + "bbox": [ + 151, + 516, + 201, + 529 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 540, + 298, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 540, + 298, + 697 + ], + "spans": [ + { + "bbox": [ + 54, + 540, + 298, + 697 + ], + "type": "text", + "content": "This paper introduces TexGarment, an efficient method for synthesizing high-quality, 3D-consistent garment textures in UV space. Traditional approaches based on 2D-to-3D mapping often suffer from 3D inconsistency, while methods learning from limited 3D data lack sufficient texture diversity. These limitations are particularly problematic in garment texture generation, where high demands exist for both detail and variety. To address these challenges, TexGarment leverages a pre-trained text-to-image diffusion Transformer model with robust generalization capabilities, introducing structural information to guide the model in generating 3D-consistent garment textures in a single inference step. Specifically, We utilize the 2D UV position map to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 517, + 557, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 557, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 557, + 639 + ], + "type": "text", + "content": "guide the layout during the UV texture generation process, ensuring a coherent texture arrangement and enhancing it by integrating global 3D structural information from the mesh surface point cloud. This combined guidance effectively aligns 3D structural integrity with 2D layout. Our method efficiently generates high-quality, diverse UV textures in a single inference step while maintaining 3D consistency. Experimental results validate the effectiveness of TexGarment, achieving state-of-the-art performance in 3D garment texture generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "The demand for 3D garment visualization is growing across virtual try-ons for e-commerce, video games, 3D films, and AR/VR environments. These industries need large volumes" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 197, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 197, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 197, + 713 + ], + "type": "text", + "content": "*Project lead and corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26566" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "of high-quality 3D garments to meet consumer expectations. While 2D fashion images are widely available and generative AI has simplified 2D image creation, producing high-quality 3D garment textures remains challenging, often requiring skilled artists and weeks of work. This highlights the need for efficient 3D garment generation methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 147, + 294, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 294, + 278 + ], + "type": "text", + "content": "The aesthetic appeal of 3D garments is deeply influenced by their texture quality, as Joseph Albers [1] said: \"Texture is the soul of painting; it gives surface depth and personality\", effectively capturing this concept. In 3D computer graphics, textures are essential for realism, adding detailed surface elements that make virtual objects lifelike. For 3D garments, high-quality textures are particularly important, capturing fabric nuances, stitching, folds, and shadows that make virtual clothing look authentic. Without these textures, even well-modeled garments can appear flat or artificial, reducing the immersive quality of 3D applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 626 + ], + "type": "text", + "content": "Despite substantial advancements in 2D image synthesis and 3D shape generation using GANs [40, 42] and diffusion models [12, 38, 39], achieving high-quality garment texture generation remains challenging. This difficulty arises from three primary factors: Global 3D Consistency. Methods employing 2D diffusion models, such as TexPainter [46], TEXTure [37] and Text2Tex [7], have made strides in texture synthesis by utilizing pre-trained LDM [38], which support strong generalization and diverse texture generation. However, text-to-image models often exhibit biases that limit their effectiveness for 3D applications, such as favoring frontal views and lacking a robust understanding of 3D structures. This frequently results in texture inconsistencies across different viewpoints of the same 3D object, ranging from minor seam misalignments to severe issues, e.g., \"Janus effect\". Generalization Ability. Methods learning from limited 3D data, such as TexOct [23], PointUV [44], and Mesh2Tex [4], generate textures by explicitly integrating 3D shape information, resulting in high 3D consistency. However, these approaches often lack generalization ability, limiting their effectiveness when generating textures for diverse 3D objects beyond the training set or adapting to a range of text and visual prompts. Inference Speed. To achieve full shape coverage, prior methods [2, 37, 44, 45] rely on iterative generation processes, requiring multiple steps or stages. This results in prolonged inference times, often lasting several minutes, and requires significant computational resources, limiting practical use for designers and content creators." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "In this work, we present TexGarment, a method tailored for generating high-quality garment textures. Given text prompts, TexGarment produces consistent, diverse, and visually striking textures in a single inference pass. A widely used approach for handling texture maps is UV representation, which maps 3D textures accurately onto 2D UV images through UV coordinate transformations. This tech-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "nique bridges the gap between 3D texture generation and 2D image methods, motivating our use of pre-trained image generation models [8] to synthesize diverse textures effectively within the 2D UV space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 120, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 553, + 491 + ], + "type": "text", + "content": "Our key insight is to incorporate 3D garment structure information as guidance within a pre-trained 2D image generation model [8], leveraging its generalization capabilities to produce consistent and diverse UV texture maps directly in the 2D UV space through only one inference process. Specifically, our method introduces two main components: 1) 3D Structure Information Extraction and 2) 3D Structure Guidance. First, given a garment mesh, we use rasterization to obtain a UV position map, where each pixel encodes the XYZ coordinates of the corresponding point on the mesh surface. This UV position map provides a 2D projection of the original 3D point cloud, capturing the basic layout of the UV texture. However, regions that are adjacent in 3D space may be separated in the UV map, causing a loss of global 3D structure information. This makes it challenging for a 2D image generation model to capture global 3D structures solely from the 2D UV position map. To address this, we introduce a 3D structure information extraction module that injects global structure information, represented by 3D point clouds, into the 2D UV position map. This process generates a 3D structure latent, which combines both 3D structure and 2D layout information, serving as guidance for the pre-trained image generation model. Second, we implement a 3D structure guidance module, analogous to text guidance, which injects the 3D structure information into the pre-trained model via cross-attention. Using a powerful text-to-image diffusion model [8] as our basic architecture, our design enables the model to receive guidance from both text and 3D structure, allowing it to generate high-quality UV texture maps in a single inference process, achieving both text fidelity and 3D consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 491, + 553, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 553, + 658 + ], + "type": "text", + "content": "Our contributions are summarized as follows: 1) We propose TexGarment, a one-stage method for garment texture generation that produces high-quality, diverse UV texture maps for untextured garment meshes in a single inference process and completes the process in 4 seconds. 2) We introduce two novel modules to incorporate 3D garment structure information into the 2D image generation model, providing 3D-consistent guidance for garment texture generation. Extensive experiments validate the effectiveness of these modules in achieving 3D consistency. 3) Qualitative and quantitative evaluations demonstrate that our method outperforms state-of-the-art approaches in garment texture generation. Additionally, a user study shows that users prefer textures generated by our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 669, + 406, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 406, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 406, + 681 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Text-to-Image Generation The field of text-to-image generation has seen considerable progress. Early research used" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26567" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "generative adversarial networks (GANs)[17] to create high-quality images, using image classes[15] and segmentation maps [30] as prompts. Recently, attention has shifted to architectural improvements that enhance generation quality with scaling Transformers, as shown in models like DALL-E [36], CogView [14], Parti [43], and Muse [6]. Another popular category of text-to-image generators is based on pixel space or latent diffusion models, including Imagegen [39], Stable Diffusion [38], SDXL [32], DiT [31], and PixArt [8]. Notably, DiT [31] advanced text-to-image generation by introducing Transformers to latent diffusion models (LDMs). PixArt [8] further optimized training methods, enabling faster training of scaled LDMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 231, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 543 + ], + "type": "text", + "content": "3D Generation via 2D generation model. With the development of text-to-image generation techniques, many 3D texturing techniques seek to leverage the prior knowledge of 2D generative models to improve 3D generation models' optimization. Previous optimization-based texture techniques [11, 20, 25, 27, 28] concentrated on improving the texture maps of 3D models by utilizing the large-scale visual language model, e.g. CLIP [34], before the advent of LDMs. DreamFusion [33] introduced the score distillation sampling (SDS) to 3D Generation, which laid the groundwork for many text-to-3D methods. Latent-nerf [26] and Fantasia3D [9] extended SDS to optimize texture mapping. Recently, TexPainter [46] performs multi-view fusion within the 2D image's color space to address inconsistency issues. However, optimization-based texture methods are hindered by prolonged training times. As an alternative, TEXTure [37], Text2Tex [7], and Paint3D [45] involve directly generating texture images from multiple views of 3D meshes and subsequently painting them onto the 3D meshes. TEXTure [37] and Text2Tex [7] adhere to this concept and suggest modifying the overlapping segments of the texture image from the previous view using results from each new camera pose. Nevertheless, these methods encounter various artifacts, including seams, noise, and irrelevant fragments. Paint3D [45] develops a specialized model to paint incomplete regions during the texture painting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Generative Texturing from 3D Data. Additionally, generative texture models are capable of training models from scratch using 3D data. To assign a color to each pixel on the surface of a three-dimensional shape, early methods [16, 29] relied on implicit texture fields. However, given that the texture on the surface of a 3D mesh is continuous, it is unlikely that discrete supervision will result in the training of a model capable of synthesizing high-quality textures. The Texturify [41] is designed to construct a texture map on the surface of a polygonal mesh. Mesh2tex [4] builds on this by integrating an implicit texture field branch for enhancement. TexOct [23] use point clouds to help circumvent variations arising from diverse mesh topologies and UV mappings. Other methods [12, 44] synthesize UV" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "maps for 3D meshes, but are ill-equipped to handle more generic objects, due to the inherent diversity between 3D meshes belonging to various categories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 108, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 372 + ], + "type": "text", + "content": "Discussion with previous work about the UV position map. Previous works, such as PointUV [44], Paint3D [45], and TextureGen [2], also incorporate the UV position map into their models. We outline the key differences between our approach and these methods: 1) PointUV [44] uses the UV position map, along with other maps like normal and mask maps, in the first stage to generate a coarse UV texture. However, this method is limited to simple textures for objects from the ShapeNet [5] dataset. In contrast, our approach uses only the UV texture map as a guide, enabling the generation of complex, high-quality textures for a wide range of garment categories. 2) Paint3D [45] and TextureGen [2] use the UV position map as a 2D condition in a second-stage inpainting model to refine textures. Our method directly applies the UV position map to guide the entire texture generation, leveraging both 2D layout and global 3D structure, resulting in better textures as shown in our experiments. 3) Unlike these prior methods, our model is a single-stage generator. Given a UV position map and a text prompt, our approach generates high-quality garment texture maps in a single, efficient inference step, while previous methods require a more complex two-stage process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 381, + 370, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 381, + 370, + 393 + ], + "spans": [ + { + "bbox": [ + 314, + 381, + 370, + 393 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 402, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 665 + ], + "type": "text", + "content": "Given a 3D garment model and a target texture description, TexGarment efficiently synthesizes high-quality, diverse, and consistent garment UV textures. TexGarment is a Transformer-based diffusion model, as shown in Figure 2. We build on a large-scale pre-trained image generation model [8], leveraging its robust prior knowledge to enhance our model's generalization capability. To ensure global 3D consistency, we first design a 3D structure information extraction module. For an untextured garment mesh, we apply rasterization to create a UV position map that captures the basic UV texture layout and encodes the 3D coordinates of points on the mesh surface. The structure information extraction module then embeds global structure information from the original point cloud into the UV position map, resulting in a 3D structure latent that combines global 3D structure information with 2D layout details. This 3D structure latent is then used by a 3D structure guidance module to guide the texture generation. Through this innovative design, our method generates high-quality and 3D-consistent texture maps in a single inference process, without the need for additional operations such as inpainting. The following sections provide the overview of each stage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 672, + 411, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 672, + 411, + 685 + ], + "spans": [ + { + "bbox": [ + 314, + 672, + 411, + 685 + ], + "type": "text", + "content": "3.1. Data Processing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Dataset. We collect a comprehensive dataset from VRoidHub based on the guidelines provided by Panic3D[10], en" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26568" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 354, + 320 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 354, + 320 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 354, + 320 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 354, + 320 + ], + "type": "image", + "image_path": "c788b14c53382f84ae633e18e651147b78b88650ba3d474ea91ef4f2b28eb79f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 323, + 159, + 331 + ], + "lines": [ + { + "bbox": [ + 118, + 323, + 159, + 331 + ], + "spans": [ + { + "bbox": [ + 118, + 323, + 159, + 331 + ], + "type": "text", + "content": "(a) Overview" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 322, + 345, + 331 + ], + "lines": [ + { + "bbox": [ + 221, + 322, + 345, + 331 + ], + "spans": [ + { + "bbox": [ + 221, + 322, + 345, + 331 + ], + "type": "text", + "content": "(b) 3D Structure Information Extraction" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 338, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 427 + ], + "type": "text", + "content": "Figure 2. Pipeline of Our Method: (a) Overview: The training process begins with embedding the input text prompt and timestep using the T5 model [35] and an MLP layer, generating the text feature and time embedding. The UV texture is encoded by a VAE [21] encoder to create a texture latent representation, which is then noised to produce a noisy latent. These representations are processed through several Transformer blocks. (b) 3D Structure Information Extraction Module: The UV position map is encoded by a VAE [21] encoder to obtain the UV position latent, while a point cloud feature extraction module captures the global 3D structure. This 3D information is integrated into the UV position latent via cross-attention, resulting in the 3D structure latent, which combines 3D global structure and 2D UV texture layout. (c) Transformer Block Details: Based on a text-to-image Transformer diffusion model [8], we introduce a 3D structure guidance module that uses the 3D structure latent as guidance, injecting it into the model through cross-attention to enhance 3d structure consistency." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 360, + 72, + 535, + 319 + ], + "blocks": [ + { + "bbox": [ + 360, + 72, + 535, + 319 + ], + "lines": [ + { + "bbox": [ + 360, + 72, + 535, + 319 + ], + "spans": [ + { + "bbox": [ + 360, + 72, + 535, + 319 + ], + "type": "image", + "image_path": "f8fe135bcacd9cd7da80f02aba3e789827379c51393a177ddc3ad323c8e81257.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 320, + 507, + 330 + ], + "lines": [ + { + "bbox": [ + 392, + 320, + 507, + 330 + ], + "spans": [ + { + "bbox": [ + 392, + 320, + 507, + 330 + ], + "type": "text", + "content": "(c) Block with 3D sturcture guidance" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 439, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 439, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 439, + 295, + 571 + ], + "type": "text", + "content": "compassing a diverse array of character models. Each model in VRoidHub includes body, clothing, and animation components. Using Blender, we segment the models by style type, preserving clothing elements while removing non-clothing components. Pants and tops were separated into distinct parts, while dresses were preserved as single units. Following segmentation, we refine the dataset by excluding samples with inconsistent color distributions and monochromatic appearances, as well as items with accessories such as ties, belts, and other extraneous parts, ultimately retaining a dataset of 8,000 clothing items." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Rasterization. In a UV texture image, the RGB information of the surface mesh is stored, and during rendering, a rasterization operator samples the color from the UV texture image onto the rendered image. A UV position map stores the XYZ coordinates of the mesh, which requires sampling mesh properties and baking them into the UV map. To achieve this, we assume the UV layout represents a 3D mesh with vertices " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "[u,v,z]" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " are the original mesh's UV coordinates and " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " is constant, making it a planar mesh. Each " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "[u,v]" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " corresponds to a vertex in the original 3D mesh. We retain the face relationships from" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 439, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 606 + ], + "type": "text", + "content": "the original mesh, and by performing rasterization, we interpolate the \"UV mesh\" positions to determine each pixel's XYZ coordinates on the UV position map. The shapes of garment meshes in the VRoid data are also defined by an alpha channel within the UV texture. Before rendering our images, we rasterize the UV position map using the alpha channel to isolate the rendered mesh positions. We compute the mesh's scale and translation with these selected mesh positions. We then translate the mesh to align with the origin of the coordinate system and scale it to fit within the " + }, + { + "bbox": [ + 313, + 439, + 555, + 606 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 313, + 439, + 555, + 606 + ], + "type": "text", + "content": " range. This process ensures that the rendered portion of each mesh remains centered in the final image. After translating the mesh, we re-rasterize the UV position map to align it within the " + }, + { + "bbox": [ + 313, + 439, + 555, + 606 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 313, + 439, + 555, + 606 + ], + "type": "text", + "content": " range." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 607, + 555, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 555, + 667 + ], + "type": "text", + "content": "We render all garment data from two perspectives: front and back. All images are generated using Nvdiffrast [22] under uniform ambient lighting. These two views are then used to generate prompts by InternVL[13]. A toy case is shown in Figure. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 672, + 424, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 424, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 424, + 683 + ], + "type": "text", + "content": "3.2. Basic Architecture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "We adopt PixArt-" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " [8] as our foundational image generation model, a Transformer-based text-to-image diffusion" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26569" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 64, + 284, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 64, + 284, + 81 + ], + "spans": [ + { + "bbox": [ + 62, + 64, + 284, + 81 + ], + "type": "text", + "content": "\"The top is short-sleeved with a white base and a black collar and cuffs, with a black bow at the center. The skirt is black with white cross motifs.\"" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 87, + 287, + 144 + ], + "blocks": [ + { + "bbox": [ + 62, + 87, + 287, + 144 + ], + "lines": [ + { + "bbox": [ + 62, + 87, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 62, + 87, + 287, + 144 + ], + "type": "image", + "image_path": "a6b8ef62f99ae9ccb36648e90293f1365d6cac4a860653f5c537ada8f4b14174.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 295, + 171 + ], + "type": "text", + "content": "Figure 3. From left to right, we show vvoid data, garment, UV texture image and UV position map, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "text", + "content": "model that excels in image quality, artistry, and semantic control. Our choice of PixArt-" + }, + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "text", + "content": " as the base architecture is motivated by three key reasons:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 218, + 295, + 373 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 218, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 277 + ], + "type": "text", + "content": "- Strong Prior Foundation. PixArt-" + }, + { + "bbox": [ + 55, + 218, + 295, + 277 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 218, + 295, + 277 + ], + "type": "text", + "content": " [8] is a popular text-to-image diffusion model that offers valuable insights to the AIGC community. Its robust image generation capabilities provide a solid prior for our method, enhancing generalization and quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 278, + 294, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 294, + 326 + ], + "type": "text", + "content": "- Efficiency and Resource Friendliness. With an optimized Transformer design, PixArt-" + }, + { + "bbox": [ + 56, + 278, + 294, + 326 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 56, + 278, + 294, + 326 + ], + "type": "text", + "content": " [8] achieves high-quality image generation using only 0.6 billion parameters, minimizing parameters and training requirements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 326, + 294, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 326, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 294, + 373 + ], + "type": "text", + "content": "- Enhanced Artistic Style. Fine-tuned on high-quality aesthetic data, PixArt-" + }, + { + "bbox": [ + 56, + 326, + 294, + 373 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 56, + 326, + 294, + 373 + ], + "type": "text", + "content": " [8] generates images with an ideal artistic style for garment texture generation, where visual appeal and stylistic quality are essential." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 375, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 375, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 295, + 506 + ], + "type": "text", + "content": "The model takes as input a text prompt, timestep " + }, + { + "bbox": [ + 54, + 375, + 295, + 506 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 375, + 295, + 506 + ], + "type": "text", + "content": ", and clean images (clear UV texture " + }, + { + "bbox": [ + 54, + 375, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{tex}}" + }, + { + "bbox": [ + 54, + 375, + 295, + 506 + ], + "type": "text", + "content": " in our method). These inputs are processed through T5 [35], MLP, and VAE [21] encoder to obtain text, time, and latent embeddings. These embeddings are then fed into Transformer blocks to execute the diffusion process. Each Transformer block consists of self-attention, adaptive normalization (adaLN-single), cross-attention, and feedforward layers. Cross-attention uses the text embedding to interact with the Transformer's hidden states, while adaLN-single leverages the time embedding to learn scale and shift parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 516, + 295, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 528 + ], + "type": "text", + "content": "3.3. 3D consistent Garment UV Texture Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "UV representation is a widely adopted method for creating texture maps, enabling precise mapping of 3D textures onto 2D images by \"unwrapping\" the 3D surface into a 2D plane. This process assigns each point on the 3D model to a specific location in 2D space, ensuring that every detail of the 3D surface accurately corresponds to a position on the UV map. This approach effectively bridges 3D texture generation and 2D image-based methods, allowing advanced 2D generative models to handle tasks that traditionally require 3D-specific techniques. Working within the 2D UV space enables high-quality texture generation that can be seamlessly applied back to the 3D model, providing a flexible and efficient solution for producing consistent textures. However, adjacent regions in 3D space may become non-adjacent in the UV map, leading to a loss of global 3D" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "structural information during mapping. To address this, we design a 3D structure information extraction module and a 3D structure guidance module, which integrate the original 3D structural information from the garment mesh into the UV texture generation process, ensuring 3D consistency in the final UV texture maps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": "3D Structure Information Extraction. For a given 3D garment mesh, we use rasterization to obtain the UV position map " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": ". Each active pixel in " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": " encodes the XYZ coordinates of a corresponding point on the garment mesh surface, creating a 2D representation of the 3D point cloud and capturing the basic layout of the UV texture. As shown in Figure 2-(b), " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": " is first processed by a VAE [21] encoder to obtain the UV position latent, " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": ". While " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": " embeds the UV layout in 2D space, it lacks 3D structure information. To address this, we retrieve the 3D point cloud " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": ", normalize it to the range [0, 1], and project it into a 3D context representation space for effective interaction with " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": ". Specifically, we encode the point cloud " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": " into a latent representation " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{point}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": " using a learnable embedding function and employ a cross-attention module to infuse " + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 144, + 555, + 336 + ], + "type": "text", + "content": " with 3D structure information from the point cloud:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 389, + 348, + 553, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 348, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 389, + 348, + 553, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {p o i n t}} = \\operatorname {P o s E m b} (\\mathcal {P}), \\tag {1}", + "image_path": "c1c2a0c4b7db7ad741b14ea63f5e781d0cacc43f25f1e03cc171a94a33aa042d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 376, + 553, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 341, + 376, + 553, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {l a t e n t}} ^ {3 \\mathrm {d}} = \\operatorname {F F N} \\left(\\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {\\text {p o s}} \\mathcal {K} _ {\\text {p o i n t}} ^ {\\mathrm {T}}}{\\sqrt {d}}\\right) \\mathcal {V} _ {\\text {p o i n t}}\\right), \\tag {2}", + "image_path": "4908e32e32a3629d8bc12c0e56ded83df6cdbdc4807d7f5a16a75729f23a487b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{point}}" + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "text", + "content": " is the embedding of point cloud. " + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{pos}} = \\mathcal{F}_{\\mathrm{pos}}\\mathcal{W}_{\\mathrm{Q}}^{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "text", + "content": " is the query embedding of " + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{\\mathrm{point}} = \\mathcal{F}_{\\mathrm{point}}\\mathcal{W}_{\\mathrm{K}}^{\\mathrm{point}}" + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{point}} = \\mathcal{F}_{\\mathrm{point}}\\mathcal{W}_{\\mathrm{V}}^{\\mathrm{point}}" + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "text", + "content": " are the key and value embeddings of " + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{point}}" + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{latent}}^{\\mathrm{3d}}" + }, + { + "bbox": [ + 313, + 413, + 554, + 487 + ], + "type": "text", + "content": " is the 3D structure latent that combines both the 2D layout information and the global 3D structure information." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "text", + "content": "3D Structure Information Guidance. We incorporate the 3D structure guidance " + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "text", + "content": " into the pre-trained text-to-image image diffusion process to guide the 3D-consistent UV texture generation. As shown in Figure 2-(c), we denote the transformer intermediate feature as " + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "text", + "content": " and the text embedding as " + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "text", + "content": ". Subsequently, we inject the 3D structure guidance " + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "text", + "content": " into the image diffusion process by interacting with " + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 488, + 554, + 584 + ], + "type": "text", + "content": " through the cross-attention:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 592, + 553, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 592, + 553, + 630 + ], + "spans": [ + { + "bbox": [ + 318, + 592, + 553, + 630 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {o u t}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {\\text {i n}} \\mathcal {K} _ {\\text {3 d}} ^ {\\mathrm {T}}}{\\sqrt {d}}\\right) \\mathcal {V} _ {\\text {3 d}} + \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {\\text {i n}} \\mathcal {K} _ {\\text {t e x t}} ^ {\\mathrm {T}}}{\\sqrt {d}}\\right) \\mathcal {V} _ {\\text {t e x t}}, \\tag {3}", + "image_path": "dd585ad764298b6ae1de36022f9b48f051fb43adf76fc12f8721aa96cc0590cd.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{in}} = \\mathcal{F}_{\\mathrm{in}}\\mathcal{W}_{\\mathrm{Q}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " is the query embedding of the " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{3\\mathrm{d}} = \\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}\\mathcal{W}_{\\mathrm{K}},\\mathcal{V}_{3\\mathrm{d}} = \\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}\\mathcal{W}_{\\mathrm{V}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " are the key and value embeddings of " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{\\mathrm{text}} = \\mathcal{F}_{\\mathrm{text}}\\mathcal{W}_{\\mathrm{K}'}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{text}} = \\mathcal{F}_{\\mathrm{text}}\\mathcal{W}_{\\mathrm{V}'}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " are the key, and value embeddings of " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\mathrm{Q}},\\mathcal{W}_{\\mathrm{K}},\\mathcal{W}_{\\mathrm{K}'}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\mathrm{V}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\mathrm{V}'}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " are the projection matrices. The two cross-attention results are added to obtain the " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{out}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26570" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "Given the text prompt condition " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " and 3D structure condition " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "c_{3\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{latent}}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": ", we apply the classifier-free guidance [19] strategy to incorporate the information and our training objective is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 124, + 296, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 124, + 296, + 153 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 296, + 153 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {\\boldsymbol {x} \\sim p, \\epsilon \\sim \\mathcal {N} (0, I), c _ {\\text {t e x t}}, c _ {3 \\mathrm {d}}, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t}, t, c _ {\\text {t e x t}}, c _ {3 \\mathrm {d}}\\right) \\| _ {2} ^ {2} \\right], \\tag {4}", + "image_path": "714fecac4aa48fdf1283913c4a061b0cad9140cce5bbf40cb176e2b764f887b5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 154, + 276, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 154, + 276, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 276, + 166 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 154, + 276, + 166 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 55, + 154, + 276, + 166 + ], + "type": "text", + "content": " is the noise latent from the ground-truth " + }, + { + "bbox": [ + 55, + 154, + 276, + 166 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{tex}}" + }, + { + "bbox": [ + 55, + 154, + 276, + 166 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 166, + 295, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 166, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 295, + 202 + ], + "type": "text", + "content": "After training, the diffusion model can generate diverse and 3D-consistent UV garment texture maps in a single inference process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 211, + 137, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 137, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 137, + 224 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 231, + 171, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 171, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 171, + 243 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 247, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 319 + ], + "type": "text", + "content": "Datasets. Our dataset consists of 8k of textured 3D garments from internal series. Garment styles include long-sleeved, short-sleeved, long skirts, short skirts, long pants, shorts, and other common styles. The dataset is divided into two parts: the training set containing 7.9k garments and the testing set containing 0.1k garments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "text", + "content": "Implementation Details. We use PixArt-" + }, + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "text", + "content": " [8] as our pre-trained text-to-image generation model. The model is trained for 50,000 iterations with the AdamW [24] optimizer, using a fixed learning rate of " + }, + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-6}" + }, + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "text", + "content": ". Newly added parameters are initialized to zero and trained with a learning rate of " + }, + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "text", + "content": ". The batch size is set to 32. For all experiments, we sample 5k points from the UV position map and synthesize a UV texture map at a resolution of " + }, + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 319, + 295, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 415, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 570 + ], + "type": "text", + "content": "Baselines. We compare our method with several state-of-the-art approaches, including TEXTure [37], Text2Tex [7], Paint3D [45], and TexPainter [46]. TEXTure [37] introduces an iterative texture generation strategy to enhance texture manipulation. Text2Tex [7] builds on this by incorporating an automatic viewpoint selection process within each iteration. TexPainter [46] utilizes a modified denoising diffusion implicit model (DDIM) to merge textures across multiple views in color space. Paint3D [45] is a coarse-to-fine texture synthesis approach. First, it samples multi-view images from pre-trained 2D diffusion models, then refines the textures in UV space using a diffusion model conditioned on both the position map and the coarse texture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 571, + 295, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 667 + ], + "type": "text", + "content": "Evaluation metrics. We utilize commonly used metrics for evaluating the performance of generative models, specifically, the Fréchet Inception Distance (FID) [18] and Kernel Inception Distance (KID) [3], which measure image quality and image diversity. Additionally, we use the CLIP score to measure the alignment between the rendered image and the input text prompt. Finally, we also evaluate the texture generation speed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 278, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 278, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 278, + 685 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "In this section, we compare our method with state-of-the-art methods both qualitatively and quantitatively. How-" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "ever, due to the lengthy processing time required by Tex-Painter [46], we include only a qualitative comparison for this method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 108, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 287 + ], + "type": "text", + "content": "Qualitative comparisons. In Figure 4, we compare our qualitative results with state-of-the-art methods. Our approach generates garment textures with enhanced 3D consistency and refined artistic quality. First, examining front and back views reveals that our method achieves consistent texture across perspectives. For example, in Figure 4-(a), the short-sleeved shirt generated by Paint3D [45] exhibits a \"Janus effect,\" while other methods show inconsistencies. In contrast, textures generated by our method remain coherent. Second, our textures exhibit finer details and organized layouts; for instance, Figure 4-(b) shows intricate lace and embroidery produced by our method. Finally, our approach aligns more closely with the text description, as demonstrated in Figure 4-(c), where our textures incorporate more Peking Opera elements than those of other methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 288, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 562 + ], + "type": "text", + "content": "Quantitative comparisons. In Table 1, we present the quantitative comparisons with the previous methods in garment texture synthesis. For the FID [18] and KID [3] calculations, we render " + }, + { + "bbox": [ + 313, + 288, + 555, + 562 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 288, + 555, + 562 + ], + "type": "text", + "content": " images of each mesh with the synthesized textures, captured from 20 fixed viewpoints. We use the average CLIP score to measure the alignment between rendering images and text prompts. For runtime, we compare inference time for each method, defining inference time as the time it takes to generate a complete texture map for a given text prompt and predefined mesh. From the results, we have the following observation. First, our method obtains the best performance in terms of FID [18] and KID [3]. Notably, it outperforms the leading competitor, Paint3D [45], by an average of 50.43 in FID and 2.23 in KID. These improvements indicate that our method excels at generating high-quality garment textures. Second, our method achieves the best CLIP scores, indicating better text-texture alignment. Finally, we use the same GPU device (a single NVIDIA-V100) for the runtime comparison to infer each method. Our method only takes 4 seconds to generate a UV texture map, which is much less time-consuming than other methods. This proves the efficiency of our method." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 570, + 550, + 639 + ], + "blocks": [ + { + "bbox": [ + 315, + 570, + 550, + 639 + ], + "lines": [ + { + "bbox": [ + 315, + 570, + 550, + 639 + ], + "spans": [ + { + "bbox": [ + 315, + 570, + 550, + 639 + ], + "type": "table", + "html": "
MethodFID↓KID↓Runtime↓CLIP↑
TEXTure [37]169.014.35~140s27.26
Text2Tex [7]147.482.77~940s27.93
Paint3D [45]147.402.86~220s28.72
Ours96.970.63~4s32.37
", + "image_path": "6d71293288ea60de905818de2f843885b934af2a31d1dbe1aca1304d2e80b7b1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 646, + 553, + 668 + ], + "lines": [ + { + "bbox": [ + 313, + 646, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 553, + 668 + ], + "type": "text", + "content": "Table 1. Quantitative comparison against state-of-the-art methods. We report the FID, " + }, + { + "bbox": [ + 313, + 646, + 553, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{{KID}}\\left( {\\times {10}^{2}}\\right)" + }, + { + "bbox": [ + 313, + 646, + 553, + 668 + ], + "type": "text", + "content": " ,CLIP score and runtime." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "User study. We further conduct a user study to analyze the overall quality of the generated textures. We randomly select 50 garment meshes and corresponding text prompts" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26571" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 68, + 531, + 330 + ], + "blocks": [ + { + "bbox": [ + 78, + 68, + 531, + 330 + ], + "lines": [ + { + "bbox": [ + 78, + 68, + 531, + 330 + ], + "spans": [ + { + "bbox": [ + 78, + 68, + 531, + 330 + ], + "type": "image", + "image_path": "0536be6d5f5f93adb376fb88cdd2870dba237480c0413d2a428f2010c323efae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 333, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 55, + 333, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 555, + 355 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with the state-of-the-art methods. Our method generates garment textures with complex and clear texture details, which are both aesthetically pleasing and maintain 3D consistency. Please enlarge the image for easy observation of details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 365, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 473 + ], + "type": "text", + "content": "to perform the user study. Following the evaluation criteria of Paint3d [45], those meshes are textured in both our method and comparison methods. Participants are requested to evaluate the results based on 1) overall quality and 2) fidelity to the text prompt, using a scale of 1 to 5. We collect the evaluation results from 30 users. As presented in Table 2, we show the average results across all prompts for each method. Our approach outperforms other methods in terms of both overall quality and text fidelity." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 483, + 291, + 553 + ], + "blocks": [ + { + "bbox": [ + 58, + 483, + 291, + 553 + ], + "lines": [ + { + "bbox": [ + 58, + 483, + 291, + 553 + ], + "spans": [ + { + "bbox": [ + 58, + 483, + 291, + 553 + ], + "type": "table", + "html": "
MethodImage Quality↑Text Fidelity↑
TEXTure [37]3.213.57
Text2Tex [7]3.323.41
Paint3D [45]4.014.17
Ours4.834.92
", + "image_path": "2144b81cd8353a891d9a9bb2de5f461d1121370e858494551f287eaa8d6be9bf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 559, + 295, + 581 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 581 + ], + "type": "text", + "content": "Table 2. Statistics on user preferences for texture quality. Users more favor our method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 588, + 200, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 200, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 200, + 601 + ], + "type": "text", + "content": "4.3. Diversity and Consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Garment texture generation demands high diversity and 3D consistency standards to achieve realistic representation. Unlike other texture tasks, garment textures must capture various fabrics, patterns, and colors to reflect the richness of real-world apparel. Garments often feature unique details like intricate stitching, subtle textures, and complex patterns that vary across styles and materials. Ensuring 3D consistency is crucial, as textures must appear seamless and coherent from all angles to maintain realism in virtual envi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 555, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 424 + ], + "type": "text", + "content": "ronments. Any misalignment can break the illusion, making the garment look artificial. Therefore, effective garment texture generation must balance diversity with strong 3D consistency, ensuring textures adapt smoothly to the garment's surface from every view." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 425, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 582 + ], + "type": "text", + "content": "To showcase our method's ability to generate diverse and 3D consistent garment textures, we present two examples in Figure 5. Given a UV position map and three text prompts, our method generates three unique texture styles for the same garment mesh. We observe the following: 1) Our method produces semantically rich textures aligned with the text prompts. Each texture displays intricate details and logical layout, illustrating our model's ability to create a variety of garment styles that meet diverse user preferences. 2) We present four different views for each generated texture. We observe a high level of 3D consistency in texture details across these views, which is achieved by our method's ability to generate complete UV texture in UV space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 588, + 408, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 408, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 408, + 601 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Evaluation of 3D structure guidance. To evaluate the effectiveness of our newly designed 3D structure guidance, we conduct experiments on three baseline configurations: point cloud guidance, UV position map guidance, and mask guidance. Point cloud guidance (3D Info.): This configuration uses only the point cloud representation as guidance without incorporating the 2D UV layout information from the UV position map. UV position map guidance (2D Info.): This configuration uses only the 2D UV layout" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26572" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 527, + 329 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 527, + 329 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 527, + 329 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 527, + 329 + ], + "type": "image", + "image_path": "72a602140849f92ccd282aaee9e5c07d471a1caa187104e3d42bfef06fb0eb8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 353 + ], + "type": "text", + "content": "Figure 5. Given a UV position map and varying text prompts, our method can generate garment textures in distinct styles. Observing the four views confirms that the generated textures maintain global 3D consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 360, + 295, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 419 + ], + "type": "text", + "content": "information from the UV position map as guidance, without the 3D global information provided by the point cloud. Mask guidance (Mask Info.): Only the UV texture mask is embedded using a VAE [21] encoder as guidance, with no structural information, neither 2D nor 3D." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 82, + 428, + 267, + 497 + ], + "blocks": [ + { + "bbox": [ + 82, + 428, + 267, + 497 + ], + "lines": [ + { + "bbox": [ + 82, + 428, + 267, + 497 + ], + "spans": [ + { + "bbox": [ + 82, + 428, + 267, + 497 + ], + "type": "table", + "html": "
MethodFID↓KID↓
3D Info.187.638.73
2D Info.174.298.07
Mask Info.210.3412.59
Ours96.970.63
", + "image_path": "3882a31bed8ef230224e050276ed942968130290400e15a73fde9de7aa38893f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 295, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 536 + ], + "type": "text", + "content": "Table 3. Evaluation of modules in our method. This demonstrates the effectiveness of each component. Our full model achieves optimal performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "As shown in Table 3, we compare three baselines with our method, leading to the following observations: 1) Compared to our full model, 3D Info. provides global 3D information but lacks 2D UV layout details. This results in worse FID and KID scores for 3D Info.. 2) When comparing 2D Info. with our full model, we find that while 2D Info. includes 2D UV layout information, it lacks global 3D information. As in the previous case, 2D Info. shows worse FID and KID scores, emphasizing the importance of 3D structural data. 3) Mask Info. provides only basic UV texture shape information, lacking structural details, making it the weakest performer among the baselines. These observations demonstrate the effectiveness of our method, which leverages global 3D structure information and 2D" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 360, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 384 + ], + "type": "text", + "content": "layout details to provide guidance for improved garment texture generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 395, + 391, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 395, + 391, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 395, + 391, + 407 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 606 + ], + "type": "text", + "content": "In conclusion, we present TexGarment, a novel and efficient method for generating high-quality garment textures. By leveraging pre-trained 2D image generation models and incorporating 3D garment structure information, TexGarment addresses key challenges in texture generation, such as global 3D consistency and inference speed. Our method enables the generation of diverse and visually consistent UV texture maps in a single inference process, significantly improving on previous multi-stage approaches. Through the introduction of two key components: 3D structure information extraction and 3D structure guidance, we effectively integrate 3D structural details into the 2D texture generation process, resulting in high fidelity and 3D consistency. Our experiments show that TexGarment outperforms existing methods in both quality and performance, with user studies confirming its superior visual appeal." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 615, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 664 + ], + "type": "text", + "content": "Limitation. TexGarment now does not support image-conditioned generation. In practical applications, users may provide a reference image to generate garment textures in a similar style. We will explore it in future research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under grant 62202369." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26573" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 281, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 281, + 101 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 281, + 101 + ], + "type": "text", + "content": "[1] Josef Albers. Interaction of color. Kyung Dang, 2013. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 103, + 295, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 295, + 147 + ], + "type": "text", + "content": "[2] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 148, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 295, + 179 + ], + "type": "text", + "content": "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 181, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 181, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 181, + 295, + 224 + ], + "type": "text", + "content": "[4] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8918-8928, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 225, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 295, + 280 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 281, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 346 + ], + "type": "text", + "content": "[6] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, José Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In Proceedings of the 40th International Conference on Machine Learning, pages 4055-4075, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 403 + ], + "type": "text", + "content": "[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 468 + ], + "type": "text", + "content": "[8] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 524 + ], + "type": "text", + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "type": "text", + "content": "[10] Shuhong Chen, Kevin Zhang, Yichun Shi, Heng Wang, Yiheng Zhu, Guoxian Song, Sizhe An, Janus Kristjansson, Xiao Yang, and Matthias Zwicker. Panic-3d: Stylized single-view 3d reconstruction from portraits of anime characters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21068-21077, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 669 + ], + "type": "text", + "content": "[12] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In CVPR, pages 1465-1474, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. *Internvl: Scaling up vision foundation models and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 554, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 152 + ], + "type": "text", + "content": "[14] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 196 + ], + "type": "text", + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 252 + ], + "type": "text", + "content": "[16] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 352 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 388 + ], + "type": "text", + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "text", + "content": "[20] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 554, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 455 + ], + "type": "text", + "content": "[21] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4, 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 554, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 500 + ], + "type": "text", + "content": "[22] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 567 + ], + "type": "text", + "content": "[23] Jialun Liu, Chenming Wu, Xinqi Liu, Xing Liu, Jinbo Wu, Haotian Peng, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Texoct: Generating textures of 3d models with octree-based diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4284-4293, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 590 + ], + "type": "text", + "content": "[24] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 591, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 554, + 657 + ], + "type": "text", + "content": "[25] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[26] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26574" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[28] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "text", + "content": "[29] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "text", + "content": "[30] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "text", + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "type": "text", + "content": "[33] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 546 + ], + "type": "text", + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[36] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[37] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 464 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "text", + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[40] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[41] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In European Conference on Computer Vision, pages 72-88. Springer, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[42] Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. Cycle-consistent inverse gan for text-to-image synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 630-638, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "text", + "content": "[43] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 350 + ], + "type": "text", + "content": "[44] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206-4216, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 353, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 418 + ], + "type": "text", + "content": "[45] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "type": "text", + "content": "[46] Hongkun Zhang, Zherong Pan, Congyi Zhang, Lifeng Zhu, and Xifeng Gao. Texpainter: Generative mesh texturing with multi-view consistency. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26575" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_content_list.json b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e96f970aa6232ea83366b2aaf94eab2e78182cd1 --- /dev/null +++ b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_content_list.json @@ -0,0 +1,1326 @@ +[ + { + "type": "text", + "text": "TexGaussian: Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 210, + 130, + 789, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bojun Xiong $^{1*†}$ , Jialun Liu $^{2*}$ , Jiakui Hu $^{3†}$ , Chenming Wu $^{2}$ , Jinbo Wu $^{2}$ , Xing Liu $^{2}$ , Chen Zhao $^{2}$ , Errui Ding $^{2}$ , Zhouhui Lian $^{1‡}$", + "bbox": [ + 178, + 202, + 818, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Wangxuan Institute of Computer Technology, Peking University, China $^{2}$ Baidu VIS", + "bbox": [ + 212, + 244, + 784, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Institute of Medical Technology, Peking University, China", + "bbox": [ + 264, + 281, + 732, + 300 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/736babe93e777639cd2055742041e084ca3b71ed3d21f9ca7a61e16f1d117ab9.jpg", + "image_caption": [ + "Figure 1. Our proposed TexGaussian is capable of generating high-quality materials for given input 3D meshes based on the corresponding textual descriptions. The generated results are naturally compatible with modern graphical engines for photo-realistic rendering under different environment maps." + ], + "image_footnote": [], + "bbox": [ + 93, + 318, + 903, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 569, + 326, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Physically Based Rendering (PBR) materials play a crucial role in modern graphics, enabling photorealistic rendering across diverse environment maps. Developing an effective and efficient algorithm that is capable of automatically generating high-quality PBR materials rather than RGB texture for 3D meshes can significantly streamline the 3D content creation. Most existing methods leverage pre-trained 2D diffusion models for multi-view image synthesis, which often leads to severe inconsistency between the generated textures and input 3D meshes. This paper presents TexGaussian, a novel method that uses octant-aligned 3D Gaussian Splitting for rapid PBR material generation. Specifically, we place each 3D Gaussian on the finest leaf node of the octree built from the input 3D mesh to render the multi-view", + "bbox": [ + 88, + 602, + 485, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images not only for the albedo map but also for roughness and metallic. Moreover, our model is trained in a regression manner instead of diffusion denoising, capable of generating the PBR material for a 3D mesh in a single feed-forward process. Extensive experiments on publicly available benchmarks demonstrate that our method synthesizes more visually pleasing PBR materials and runs faster than previous methods in both unconditional and text-conditional scenarios, exhibiting better consistency with the given geometry. Our code and trained models are available at https://3d-aigc.github.io/TexasGaussian.", + "bbox": [ + 511, + 571, + 908, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 753, + 645, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional 3D asset creation relies heavily on the expertise and extensive effort of professional designers [23], posing a significant barrier for casual users interested in creating 3D models independently. In the 3D design process, geometry creation typically represents only a small portion of the overall time, while the majority is dedicated to developing textures and appearances, which are far more time-consuming. Achieving a delicate appearance for a 3D", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*Denotes equal contribution. $\\dagger$ This work was partly done when Bojun and Jiakui interned in Baidu. $\\ddagger$ Corresponding author. E-mail: lianzhouhui@pku.edu.cn This work was supported by the National Natural Science Foundation of China (Grant No.: 62372015), Center For Chinese Font Design and Research, and Key Laboratory of Intelligent Press Media Technology.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "551", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model often demands substantial time and effort from experienced designers.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Artificial Intelligence Generated Content (AIGC) based on deep generative models, especially diffusion models [17, 46] have been widely used to facilitate the process of artistic creation, catalyzing advancements in image generation [15, 37, 40, 42, 43] and video generation [3-5, 18, 52]. As a result, exploring effective ways to leverage deep generative models to streamline the creation of detailed appearances for 3D models has become a popular direction in the graphics and vision communities.", + "bbox": [ + 89, + 123, + 482, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in 3D texture generation attempt to use depth-conditional pre-trained 2D diffusion model [58] to synthesize RGB images based on the depth maps rendered from multiple views, such as TEXture [41] and text2tex [8]. Subsequent works [6, 20, 57] further improve multi-view consistency via blending the multi-view images to a single and consistent texture map in every denoising step. However, these methods struggle to have a global picture of 3D geometries due to the use of 2D diffusion models, leading to inconsistencies between the texture map and the semantics of the input 3D meshes. Moreover, the generated assets suffer from illumination-baked textures, which can significantly degrade the quality of the final rendering when placed in novel lighting conditions [56]. While DreamMat [60] supplements geometry and light control to achieve material decomposition through score distillation sampling [38], it still struggles to fully capture the global geometry features. This limitation often results in the multiface Janus problem and leads to over-saturated colors.", + "bbox": [ + 91, + 261, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, training a 3D neural network directly on 3D data, such as Point-UV Diffusion [55] and TexOct [26], is an effective way for 3D global consistency. Meanwhile, this avoids multi-view sampling and score distillation sampling, which accelerates the process of texture synthesis. However, relying on colored point clouds for 3D representation and supervision often results in blurred outputs, primarily due to the sparse and non-compact nature of point clouds in 3D space. Due to the limitations of the adopted 3D representations and the lack of training data on PBR materials, these approaches are incapable of generating high-fidelity PBR materials for 3D models.", + "bbox": [ + 89, + 550, + 482, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, this paper presents TexGaussian, a fast and high-fidelity PBR material generation model directly in 3D space that maintains 3D global consistency. Different from previous approaches that primarily rely on diffusion models, our method works in a regression manner to regress the PBR material from the input mesh for faster generation speed. To enable effective learning in 3D space, we propose to use octree, a specialized sparse voxel structure that efficiently organizes and preserves 3D information, which can be built from 3D point clouds sampled from the surface of the object. How", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ever, directly regressing the color of 3D point clouds on octree often results in blurry textures as mentioned in [26]. To tackle the challenges of the incompact and discrete nature presented by points, we use 3D Gaussian Splatting (3DGS) [22], a robust representation that bridges the gap between 3D space and 2D raster images, allowing us to fully utilize rich 2D image information to alleviate blurring results. Specifically, for each input mesh, we sample the dense 3D point clouds on its surface to build the corresponding octree. On each octant (i.e., the finest leaf node of octree), we place a 3D Gaussian [22] at its central position. Then, we use the octree-based 3D U-Net [49] to predict the parameters of each 3D Gaussian on octants. Apart from RGB colors, we extend each 3D Gaussian with additional parameters to represent the roughness and metallicity of 3D objects. Multi-view images, including albedo, roughness, and metallic maps, can be rasterized from all these 3D Gaussians via 3DGS. The 3D U-Net is supervised by the difference between the predicted multi-view images and their corresponding ground truth. Notably, the 3D U-Net is trained to directly regress the multi-view images based on the geometry feature of the input 3D model, which further facilitates the process of 3D Gaussian prediction compared to diffusion manner. We train our TexGaussian model on a subset of Objaverse [13] with high-quality PBR materials, enabling fast PBR material generation with a single feedforward pass. In summary, the contributions of our paper are threefold:", + "bbox": [ + 511, + 90, + 906, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an octant-aligned 3D Gaussian Splitting method for high-quality PBR material synthesis on un-textured input 3D mesh, which fully utilizes the supervision from 2D images, avoiding blurry results caused by the discreteness of 3D point clouds.", + "- We adopt a regression manner to train our 3D U-Net model instead of diffusion denoising, achieving faster generative speed.", + "- We propose TexGaussian, a novel PBR material generation method based on the above two techniques. To our knowledge, our work is the first to generate PBR materials directly in 3D space. Qualitative and quantitative experiments have been conducted to verify the superiority of the quality and efficiency of our method over other existing approaches." + ], + "bbox": [ + 513, + 516, + 903, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 752, + 653, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we mainly summarize current texture synthesis methods, which can be roughly divided into three categories.", + "bbox": [ + 511, + 777, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-view Images Synthesis", + "text_level": 1, + "bbox": [ + 511, + 832, + 769, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many previous works have tried to leverage the powerful T2I model to assist texture generation for 3D shapes. Specifically, they render the depth map of in", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "552", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "put 3D mesh from multiple views and use depth conditional T2I models [58] to synthesize RGB images and perform text-conditioned texture synthesis. TEXTure [41] and Text2Tex [8] iteratively paint a mesh from different views. However, images synthesized in early view could produce errors that are not reconcilable with the geometry that is observed in later views. Many subsequent works try to alleviate multi-view inconsistency via different alignment modules. TexFusion [6] proposes a sequential interlaced multi-view sampler that interleaves texture assembling with denoising steps in different camera views. Similarly, TexGen [20] directly enforces view consistent sampling in RGB texture space and develops a noise resampling strategy to retain rich texture details. TexPainter [57] blends images from different views into a common color-space texture image by weighted averaging to guarantee multi-view consistency. To remove light influence from 2D diffusion models, Paint3D [56] contribute separate UV Inpainting and UVHD diffusion models specialized in shape-aware refinement. More recently, Meta 3D TextureGen [1] utilizes geometry-aware text-to-image model to generate a multiview image for texture. Although these methods achieve impressive texture results, they can still hardly comprehend the overall geometry of input 3D mesh.", + "bbox": [ + 93, + 90, + 480, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Optimization-based 3D Generation", + "text_level": 1, + "bbox": [ + 94, + 469, + 393, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before the emergence of large-scale Text-to-Image generative model, earlier methods [10, 19, 30, 31, 33] propose to optimize texture map of 3D object via natural language supervised visual model, CLIP [39]. Subsequently, score distillation sampling (SDS) was adopted by DreamFusion [38] and Magic3D [25]. The key idea is to optimize 3D representations such as NeRF [32] or InstantNGP [34] with the gradient guidance from 2D diffusion priors [38, 51]. To generate PBR material, TextureDreamer [53] optimizes spatially-varying bidirectional reflectance distribution (BRDF) field through personalized geometric-aware score distillation. Fantasia3D [9] uses a single predefined environmental. However, the generated images from diffusion models may not be consistent with the given environment light. Another concurrent work Paint-it [54] is similar to Fantasia3D [9] but with Deep Convolutional Physically-Based Rendering (DC-PBR) parameterization. DreamMat [60] proposes a novel geometry and light-aware diffusion model, which is trained to generate images that are consistent with the given environment light. FlashTex[14] uses a light-conditioned diffusion model within a two-stage pipeline, combining reconstruction and SDS optimization to enhance texture quality and achieve better light disentanglement. However, these methods struggle with the Janus problem due to the semantically ambiguous. And the time consumption is relatively too long to use in practice.", + "bbox": [ + 93, + 494, + 480, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Generating Texture from 3D Data", + "text_level": 1, + "bbox": [ + 517, + 90, + 808, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most straightforward way to synthesize texture map for 3D mesh is to train generative model directly from 3D data with texture ground truth [7, 11-13]. Early methods such as Texture Fields [35] learn implicit texture fields to assign a color to each pixel on the surface of the 3D shape. Texturify [45] devices face convolution operation on mesh surface to predict texture on each face. It employs differentiable rendering with an adversarial loss to ensure that generated textures produce realistic imagery. Recently, some diffusion-based texture synthesis methods, such as PointUV [55] and TexOct [26] train a denoising network on colors of point clouds which are further mapped to 2D UV map. Although these methods achieve better 3D global consistency with input mesh, they are only trained on several categories of small datasets [7]. What's more, discrete supervision from 3D point clouds leads to suboptimal results compared with continuous signals such as 2D images.", + "bbox": [ + 517, + 114, + 903, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 517, + 388, + 601, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide a detailed explanation of our proposed method, TexGaussian. The overall pipeline of our method is shown in Fig. 2. Existing texture synthesis approaches mainly rely on pre-trained 2D diffusion models, which struggle to fully understand the overall 3D structure. This often leads to misalignment between the generated texture map and 3D semantics. Our goal is to synthesize the high-quality PBR material for a given mesh directly in 3D space.", + "bbox": [ + 517, + 417, + 903, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 517, + 574, + 620, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable effective learning in 3D space, we use octree, a sparse voxel structure, to organize and store 3D information without compromising representation quality due to the inherent sparseness of 3D objects in 3D space. Thus, we sample a large number of points on the surface of the given mesh to construct the corresponding octree. The key components of our method are the octant-aligned Gaussian Splatting and the octree-based 3D U-Net. Specifically, we place a 3D Gaussian at the center of each octree's finest leaf node. The octree-based 3D-Unet is trained to predict the parameters of each 3D Gaussian. Under this circumstance, the generated octant-aligned 3D Gaussians are naturally on the mesh surface. We render them from multiple viewpoints using 3D Gaussian Splatting and train the 3D U-Net by minimizing the difference in 2D raster images and 3D Gaussian parameters. During inference, the rendered multi-view images are baked into the UV space of input mesh using differentiable mesh rendering, producing the final texture and material map. The details of 3D Gaussian Splatting (3DGS) are provided in the supplementary material.", + "bbox": [ + 517, + 599, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "553", + "bbox": [ + 486, + 946, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/142cc3034aabedd25a81c3acd4f601cdbb006a856c3cc2e969e4cc15b6310317.jpg", + "image_caption": [ + "Figure 2. An overview of our PBR material generation framework. (a) We propose octant-aligned 3D Gaussian Splitting, which positions a 3D Gaussian at the center of each finest leaf node of the constructed octree. Additional channels are added at the end of the Gaussian parameters to model PBR material. (b) We use the 3D U-Net built upon octree-based convolutional networks to predict the Gaussian parameters. Our octree-based 3D U-Net is trained by minimizing the difference on 2D raster images and 3D Gaussian parameters. (c) We bake the multi-view rendered images to the UV space of the input 3D model to realize physically based rendering under new illumination environments." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Octant-aligned 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 89, + 349, + 418, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a given mesh, we first sample $N = 100,000$ 3D points on its surface. The corresponding octree is then built by adaptively subdividing the voxels containing those points until the maximum depth is reached. As a result, all of the finest leaf nodes of octree lie along the boundary of the 3D object, which is consistent with the characteristic of the optimized 3D Gaussian. Therefore, we align a 3D Gaussian at the central position of each finest leaf node to effectively model the appearance without compromising the splatting quality. It is worth noting that we do not adjust the position of 3D Gaussians because they are already on the surface of 3D shape and adding additional offset relative to the octant center would not improve the rendering quality through our early experiments. To model PBR material that includes roughness and metallic information, we follow [21, 44] to append two additional channels at the end of 3D Gaussian parameters which are responsible for roughness and metallic map rendering, respectively.", + "bbox": [ + 91, + 371, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each 3D object in our dataset, we render multi-view images of albedo, roughness, and metallic maps for the training of 3D Gaussian. The multi-view rendering results of the albedo map are view-independent. So we just use three RGB channels to take the place of the original spherical harmonics. As noted, we exclude the position from 3D Gaussian parameters in our model. Thus, our 3D Gaussian parameters consist of 13 channels in total: one for opacity, three for scale, four for rotation, three for albedo, one for roughness, and one for metallic. To stabilize training, we choose to employ different activation functions compared to the original Gaussian Splatting [22]. Specifically, We multiply the softplus-activated scales $\\mathbf{s}_i$ with 0.01, ensuring that the initial 3D Gaussians conform to the object's counter of object at the beginning of training rather than expanding outward.", + "bbox": [ + 89, + 645, + 483, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We pre-fitting the parameters of 3D Gaussians on the", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "constructed octree for each 3D object in our dataset via the original loss function in [22] on multi-view albedo images, roughness maps, and metallic maps:", + "bbox": [ + 511, + 349, + 906, + 396 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = (1 - \\lambda) L _ {1} + \\lambda L _ {\\mathrm {D - S S I M}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 411, + 906, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda = 0.2$", + "bbox": [ + 511, + 435, + 617, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Octree-based 3D U-Net Training", + "text_level": 1, + "bbox": [ + 511, + 460, + 799, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To handle the encoding of our octant-aligned 3D Gaussian representation, we use the 3D U-Net built upon octree-based convolutional neural networks [49] to predict the Gaussian parameters. Inspired by LGM [47], the output feature of the 3D U-Net on each octant is treated as the 3D Gaussian parameters, which contain 13 channels, as discussed in the last subsection.", + "bbox": [ + 511, + 483, + 905, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To effectively train our 3D U-Net, we adopt the regressive loss objective, which could further facilitate the generation process. The input to our octree-based 3D U-Net is the geometry feature on each octree's finest leaf node, such as normal and local displacement. For text-conditioned PBR material synthesis, the text feature is extracted by pretrained CLIP model [39] and is fed to U-Net via the octree-based multi-head cross attention mechanism similar to [48]. The predicted 3D Gaussians are rasterized from multiple views via Gaussian Splatting [22]. At each training step, we rasterize the albedo images, alpha images, roughness and metallic maps from randomly selected eight views. Following [47], we apply the mean square error (MSE) loss and the VGG-based LPIPS loss [59] to the albedo image, roughness map, and metallic map:", + "bbox": [ + 511, + 589, + 906, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {a l b e d o}} = L _ {\\text {M S E}} \\left(I _ {\\text {a l b e d o}}, I _ {\\text {a l b e d o}} ^ {\\mathrm {G T}}\\right) + L _ {\\text {L P I P S}} \\left(I _ {\\text {a l b e d o}}, I _ {\\text {a l b e d o}} ^ {\\mathrm {G T}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 825, + 903, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {R}} = L _ {\\mathrm {M S E}} \\left(I _ {\\mathrm {R}}, I _ {\\mathrm {R}} ^ {\\mathrm {G T}}\\right) + L _ {\\mathrm {L P I P S}} \\left(I _ {\\mathrm {R}}, I _ {\\mathrm {R}} ^ {\\mathrm {G T}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 859, + 903, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {M}} = L _ {\\mathrm {M S E}} \\left(I _ {\\mathrm {M}}, I _ {\\mathrm {M}} ^ {\\mathrm {G T}}\\right) + L _ {\\mathrm {L P I P S}} \\left(I _ {\\mathrm {M}}, I _ {\\mathrm {M}} ^ {\\mathrm {G T}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 883, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "554", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where 'R' and 'M' denote roughness and metallic, respectively. We further apply the MSE loss on the alpha image for faster convergence of the shape:", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\alpha} = L _ {\\mathrm {M S E}} \\left(I _ {\\alpha}, I _ {\\alpha} ^ {\\mathrm {G T}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 145, + 482, + 164 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To accelerate the coverage process, we also apply the 3D MSE loss $L_{\\mathrm{3D}}$ , which calculates the difference between predicted parameters of 3D Gaussians and pre-fitting ones. Finally, the complete loss function of our model is defined as the sum of all the above losses:", + "bbox": [ + 89, + 174, + 483, + 250 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t o t a l}} = L _ {\\text {a l b e d o}} + L _ {\\mathrm {R}} + L _ {\\mathrm {M}} + L _ {\\alpha} + L _ {\\mathrm {3 D}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 261, + 482, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Texture and Material Baking", + "text_level": 1, + "bbox": [ + 89, + 287, + 352, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the inference stage, we also first build the corresponding octree for the input 3D mesh. Then, we use our trained octree-based 3D U-Net to generate 3D Gaussian on every octant and rasterize them from multiple views. The ultimate output of our method should be a global texture map and material map. Thus, we rasterize the input mesh using the differentiable renderer [24] and optimize its albedo, roughness and metallic parameters via the MSE loss between the Nvdiffrast [24] rendering results and 3D Gaussian rendering results. With adequately optimized implementation, this process takes only about several seconds to bake the multiview images to an untextured 3D model. After the optimization, our input mesh paired with its albedo and material map is capable of performing physically based rendering in new illumination environments.", + "bbox": [ + 89, + 309, + 483, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 549, + 223, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 573, + 307, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset We train our TexGaussian model on two publicly-available datasets: ShapeNet [7] and Objverse [13]. For 3D objects in the ShapeNet dataset, since they only contain albedo maps without PBR materials, our model is trained to generate RGB textures only. We train our model on four categories of ShapNet: bench, car, chair and table which is consistent with previous works [26, 55]. We curate a subset of Objverse models encoded with PBR materials and convert those using the specular-glossiness workflow to the metallic-roughness workflow to achieve a consistent PBR representation. This process resulted in a total of 29,200 models. For each 3D object, we render its RGBA albedo image, roughness map, and metallic map from 64 views of $512^2$ for training.", + "bbox": [ + 89, + 595, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Architecture Our 3D U-Net is built upon octree-based convolutional neural networks [49] which only operates on non-empty octree leaf nodes. The depth of our constructed octree is set to 8 (resolution $256^{2}$ ). Our 3D U-Net consists of 5 down-sampling and up-sampling blocks.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the text-conditioned generation, cross-attention layers are only inserted at the last two down-sampling blocks, the middle block, and the first two up-sampling blocks. The input channel of our 3D U-Net is set to 4, where 3 for normal vector and 1 for local displacement while the output channel is set to 13 as mentioned above. The resolution of 3D Gaussian Splatting is set to $512^2$ . The resolution of the baked albedo and material map is set to $1024^2$ .", + "bbox": [ + 511, + 90, + 905, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details We train our model in a single-category and unconditional manner on ShapeNet, i.e., there are 4 TexGaussian models without text conditions in total. This per-category model is trained on 4 NVIDIA A100 (40G) GPUs for about 2 days and is set up for a fair comparison with other existing methods trained on ShapeNet. For the Objaverse dataset, we train a text-conditioned TexGaussian model for PBR material generation on 24 NVIDIA A100 (40G) GPUs for two weeks. For each batch, we randomly sample 8 camera views to calculate loss functions. We adopt the AdamW [27] optimizer with a learning rate of $4 \\times 10^{-4}$ , a weight decay of 0.05, and betas of (0.9,0.95). The learning rate is cosine annealed to 0 during the training and we clip the gradient with a maximum norm of 1.0.", + "bbox": [ + 511, + 219, + 906, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics To effectively assess the quality of our generative results, we adopt the metric proposed by [61]. Specifically, each mesh with the generated PBR material is rendered from 20 uniformly distributed views to get multi-view albedo maps, and PBR rendering results in a new illumination environment. These images are used to calculate the FID [36] and KID [2] scores against those from ground truth to evaluate the quality and diversity of generative PBR materials. For the ShapeNet dataset, we only use albedo maps. The final score is averaged across 20 views, and a lower FID and KID score indicates better generation quality and diversity.", + "bbox": [ + 511, + 449, + 908, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Unconditional RGB Texture Generation", + "text_level": 1, + "bbox": [ + 511, + 640, + 856, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct unconditional RGB texture instead of PBR material generation on four categories of the ShapeNet dataset. We use the same train and test data split as in [26, 55]. Fig. 3 shows the unconditionally generated RGB texture by our single-category model with high quality, fidelity, and diversity. It can be seen that the texture map synthesized by our method is of great 3D global consistency with the corresponding input geometry.", + "bbox": [ + 511, + 662, + 905, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Comparison We conduct quantitative analysis and comparison on our model and other state-of-the-art methods. Specifically, we compare our model with PointUV Diffusion [55] and TexOct [26] which also train single-category models on ShapeNet. We also compare with some methods using pre-trained 2D diffusion prior for multi-view images synthesis such as TEXTure [55], Text2Tex [8]", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "555", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a1ec1bcd7da090762641781e1e2cafe99fb22f27b6e3fa1ea4f1c11e76809a18.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of the FID and KID $(\\times 10^{2})$ score as well as the inference time of TexGaussian and other methods on the ShapeNet dataset [7]. The top part reports the comparison with methods that also train per-category models on ShapeNet. The bottom part reports the comparison with 2D diffusion-based methods, which select 50 samples from the test set of each category." + ], + "table_footnote": [], + "table_body": "
MethodsAverageBenchCarChairTableTime
FID↓KID↓FID↓KID↓FID↓KID↓FID↓KID↓FID↓KID↓
Point-UV (1-Stage) [55]88.435.7864.961.44186.1017.5746.231.9356.422.1939.92s
Point-UV (2-Stage) [55]61.492.6767.481.6189.385.8241.331.5347.751.7249.75s
TexOct [26]59.452.6060.460.9790.106.1137.701.3649.521.9717.44s
Ours49.762.0746.370.4480.205.2229.961.0942.521.5411.02s
TEXTure [41]169.827.94152.535.28236.5818.52119.712.86170.465.11132.21s
Text2Tex [8]156.626.68154.955.79188.6015.06125.343.11157.572.771005.58s
Paint3D [56]135.254.67104.161.68204.2213.06104.642.24127.981.71231.56s
Ours100.971.8886.370.75117.535.2094.890.57105.090.9511.12s
", + "bbox": [ + 91, + 142, + 906, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0c1a3879e9d5f1db1a007de4898b9a3d1061e5720bce488eb6182be471249af5.jpg", + "image_caption": [ + "Figure 3. Unconditional RGB texture generative results on ShapeNet. Please zoom in for a better inspection of color details. Paint3D [56], and TexPainter [57] whose input text prompts are set to \"a *\" and \"* is the name of corresponding category. Due to the relatively long time consumption, we only use 50 3D objects selected from the test set to evaluate the methods based on 2D diffusion models. We conduct only a qualitative comparison for TexPainter [57] for its lengthy processing time. Table 1 reports the comparison of FID and KID scores as well as average inference time on a single NVIDIA A100 (40G) GPU. It is worth noting that our model only takes about one second to predict the parameters of each 3D Gaussian and the rest of the time is used for texture baking. From Table 1, we have the following observation. First, our method obtains the best performance in terms of FID and KID. For example, our method outperforms TexOct [26], by an average of 9.69 in FID and 0.53 in KID. These improvements indicate that our method excels at generating high-quality textures. Second, TexGaussian achieves the fastest generative speed in all categories, which is much less time-consuming than other methods." + ], + "image_footnote": [], + "bbox": [ + 107, + 330, + 472, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison Fig. 4 provides some qualitative results on the same input 3D meshes by different methods. Point-UV [55] and TexOct [26] use colored point clouds as supervision to train diffusion models. As a consequence, the generated texture map is relatively blurry due to the discreteness of the point cloud. For other methods that leverage pre-trained 2D diffusion prior, they can hardly comprehend the total geometry. The generated texture maps", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86ea7666c7b2f0e83a065b8b10996d512d212979bca695d8dcd137b38bfd6843.jpg", + "image_caption": [ + "Figure 4. Examples of generated RGB textures obtained by Tex-Gaussian and other state-of-the-art models on the same 3D object. Please zoom in for a better inspection." + ], + "image_footnote": [], + "bbox": [ + 517, + 338, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by them are not consistent with the semantics of input 3D mesh, such as the chaotic stripes on the bench and chair in Fig. 4. On the contrary, our method generates smooth and colorful RGB textures on unseen objects, which align well with 3D meshes.", + "bbox": [ + 511, + 832, + 908, + 907 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "556", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/517f40a175acc03512af0d34931028fd8e7ab6851c7e83a39793e85414a58263.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison with Fantasia3D [9], FlashTex [14], DreamMat [60] and Make-it-Real [16]. We show the rendered image, albedo map, roughness map, and metallic map for each 3D object." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Text-conditioned PBR Material Generation", + "text_level": 1, + "bbox": [ + 89, + 368, + 459, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train the text-conditioned PBR material generation model on our filtered subset of Objaverse [13]. We use the text descriptions of 3D objects from Cap3D [28, 29] to train our model. We choose 29,000 3D objects in our filtered subset for training and the rest 200 for testing. Fig. 1 shows some generative results by proposed TexGaussian on the test set. We can see that our model comprehends the overall geometry feature of input 3D mesh. It is capable of synthesizing high-quality albedo, roughness, and metallic, which are aligned well with 3D semantics such as accents with high metallic on the wooden box and rope with high roughness on the jug.", + "bbox": [ + 88, + 391, + 483, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison We conduct quantitative comparison with some state-of-the-art text-conditioned PBR material synthesis methods: Fantasia3D [9], FlashTex [14], and DreamMat [60] on our test set. Table 2 reports FID and KID scores on both multi-view albedo map and PBR rendering images under the same illumination environment. TexGaussian outperforms all the baseline methods in achieving the best visual quality of the generated appearances. We also report the average inference time of different methods on a single NVIDIA A100 (40G) GPU across our test set. Due to the iterative optimization of score distillation sampling, all other three methods cost at least 20 minutes for the generation. In contrast, our TexGaussian only takes about 20 seconds, in which one second is for predicting the Gaussian parameters and the rest is for baking, which performs $60 \\times$ faster than previous methods.", + "bbox": [ + 89, + 585, + 483, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison Fig. 5 visualizes the generated albedo, roughness, and metallic of each compared method from the same text prompt and untextured meshes. We also show the PBR rendering images of the generated materials", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "under the same environment light. It can be observed that Fantasia3D [9] and FlashTex [14] generate irregular colors on mesh surfaces, which lead to rendering results with low quality. DreamMat [60] is capable of generating a visually pleasing appearance for test 3D models. However, it can hardly align the generated PBR material well with the 3D semantics, such as the black band on the wooden barrel. What's more, it tends to generate over-saturated colors, which is demonstrated in the wooden chest and green grenade. In addition, all of the compared methods struggle to completely disentangle the light and texture, which results in relatively chaotic results of generated metallic and roughness. We also compare our method with a Multimodal Large Language Model based method, Make-it-Real [16], which accepts 3D objects paired with albedo maps generated by our TexGaussian as input. It can be observed that Make-it-Real struggles to capture the overall 3D structure of input meshes. As a consequence, it tends to predict single-value PBR maps, leading to irregular rendering results, such as the wooden chest with extremely large metallic value. On the contrary, our method is directly trained from 3D original data. It is capable of generating clean and smooth PBR material while fully comprehending the overall 3D structure. We provide more generative results in the supplementary material.", + "bbox": [ + 511, + 369, + 906, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 757, + 666, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Essentials of Octree We first analyze the effectiveness of using Octree in our pipeline. To do this, we calculate the quantitative quality of 3D Gaussian pre-fitting by our octant-aligned and full voxel-aligned 3D Gaussian Splitting on the subset of Objaverse dataset [13]. The depth of octree in this ablation is set to 8, and the resolution of full voxels is $256^3$ , which is consistent with the finest resolution of octree to guarantee comparison fairness. Table 3 reports the", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "557", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a9e0046238c2ad3144d75a1dc1a326376d7c3824fbf3ae06829afc03ad7dc1fd.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of the FID and KID $(\\times 10^{2})$ scores as well as the inference time of TexGaussian and other methods on our test set which consists of 200 3D objects with ground truth PBR materials." + ], + "table_footnote": [], + "table_body": "
MethodsAlbedoPBR renderingTime
FID↓KID↓FID↓KID↓
Fantasia3D [9]213.210.96209.870.4822.7mins
FlashTex [14]185.241.13186.820.4220.3mins
DreamMat [60]152.631.09145.490.1948.2mins
Ours123.720.20129.520.0221.04s
", + "bbox": [ + 91, + 155, + 486, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison of methods using different types of 3D Gaussian Fitting.", + "bbox": [ + 89, + 253, + 483, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80bf1f8d675abd45b41508bb71945ceb6ca841b48ddaefb3d3f3fc872253f0f8.jpg", + "image_caption": [ + "Figure 6. Visualization of different manners of Gaussian fitting. The rendering results demonstrate excellent reconstruction performance of the proposed octant-aligned 3D Gaussian Splatting." + ], + "image_footnote": [], + "bbox": [ + 91, + 292, + 483, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PSNR, LPIPS [59], and SSIM [50] metrics on the albedo map as well as the average number of 3D Gaussians of octant and full voxel aligned 3D Gaussian Splitting across our Objverse subset. Fig. 6 also presents some visual results of our Gaussian fitting. These results demonstrate that the proposed octant-aligned 3D Gaussian produces a much more reasonably compact and precise representation of diverse and complex 3D assets with much fewer 3D Gaussians compared to the full voxel version.", + "bbox": [ + 89, + 569, + 482, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Essentials of different losses We verify the importance of different losses we proposed to train our model. We term all the losses calculated on 2D image space as $L_{2\\mathrm{D}}$ :", + "bbox": [ + 89, + 707, + 483, + 755 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {2 \\mathrm {D}} = L _ {\\text {a l b e d o}} + L _ {\\mathrm {R}} + L _ {\\mathrm {M}} + L _ {\\alpha}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 767, + 482, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We train two additional TexGaussian models using only the $L_{2\\mathrm{D}}$ or $L_{3\\mathrm{D}}$ loss on ShapeNet car category due to its large variations and complexity of texture to validate their effectiveness. It is worth noting that in this dataset, $L_{2\\mathrm{D}} = L_{\\mathrm{albedo}} + L_{\\alpha}$ due to the lack of material information. The training curves of MSE loss and LPIPS loss between rendering images and ground-truth ones are shown", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa8ab213859c7a239e18cd0ddc54efda5b41131873f7038710356170d0c1847a.jpg", + "image_caption": [ + "Figure 7. The training loss curves of our model with different losses." + ], + "image_footnote": [], + "bbox": [ + 522, + 88, + 707, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05c688ef0ceed95203f5679d5caf9a44010ad346584880e664c0a05507d8bbf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 88, + 895, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95799d7d4b83d29b0b0e85f6ef03b33c93c8a2502927329eac9a1bcd5dbdb8ce.jpg", + "image_caption": [ + "Figure 8. Results of our model trained with and without $L_{2D}$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 257, + 906, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Fig. 7, which demonstrate the effects of $L_{2D}$ and $L_{3D}$ . From the loss curves, we can conclude that $L_{3\\mathrm{D}}$ facilitates the process of convergence and $L_{2D}$ enhances the quality of synthesized images, as evidenced by the large margin of improvement in the LPIPS loss when introducing $L_{2D}$ . We also provide some qualitative results to verify the effectiveness of $L_{2D}$ on the test set of car category in Fig. 8. Only using $L_{3D}$ results in a relatively blurry texture map due to the discreteness of 3D Gaussian which is similar to the characteristic of 3D point clouds analyzed above.", + "bbox": [ + 511, + 412, + 906, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 577, + 633, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed TexGaussian, an octree-based 3D Gaussian Splatting model for high-quality PBR material generation on untextured meshes. We aligned each 3D Gaussian on the octant of the corresponding octree built from the input untextured object and extended the parameters of 3D Gaussian with additional channels to represent the roughness and metallic map. We trained our model with regression objectives, achieving faster inference speed compared to previous texture synthesis methods. Experimental results demonstrated that our method is capable of generating high-quality PBR materials that are readily usable in modern graphics engines for photo-realistic rendering, offering enhanced realism for a variety of applications.", + "bbox": [ + 511, + 602, + 906, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations The generalization of TexGaussian is still hindered by the scale of the training set. Thus it struggles to generate various textures for some extremely complex 3D objects beyond our training data. We are looking forward to training our TexGaussian model with more parameters and more data on a larger-scale GPU cluster in the future.", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "558", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 3", + "[2] Mikołaj Binkowski, Dougal J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations, 2018. 5", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2", + "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023.", + "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2", + "[6] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and KangXue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3", + "[7] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 3, 5, 6", + "[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2, 3, 5, 6", + "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 3, 7, 8", + "[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 3", + "[11] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023.", + "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3, 5, 7", + "[14] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. Flashtex: Fast reconfigurable mesh texturing with lightcontrolnet. In European Conference on Computer Vision (ECCV), 2024. 3, 7, 8", + "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[16] Ye Fang, Zeyi Sun, Tong Wu, Jiaqi Wang, Ziwei Liu, Gordon Wetzstein, and Dahua Lin. Make-it-real: Unleashing large multimodal model for painting 3d objects with realistic materials, 2024. 7", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 2", + "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2", + "[19] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 3", + "[20] Dong Huo, Zixin Guo, Xinxin Zuo, Zhihao Shi, Juwei Lu, Peng Dai, Songcen Xu, Li Cheng, and Yee-Hong Yang. Tex-gen: Text-guided 3d texture generation with multi-view sampling and resampling. ECCV, 2024. 2, 3", + "[21] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian-shader: 3d gaussian splattering with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 4", + "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 4", + "[23] Matthias Labschütz, Katharina Krosl, Mariebeth Aquino, Florian Grashäftl, and Stephanie Kohl. Content creation for a 3d game with maya and unity 3d. Institute of Computer Graphics and Algorithms, Vienna University of Technology, 6(124):2, 2011. 1", + "[24] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "559", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[26] Jialun Liu, Chenming Wu, Xinqi Liu, Xing Liu, Jinbo Wu, Haotian Peng, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Texoct: Generating textures of 3d models with octree-based diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4284-4293, 2024. 2, 3, 5, 6", + "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[28] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 7", + "[29] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. arXiv preprint arXiv:2404.07984, 2024. 7", + "[30] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 3", + "[31] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3", + "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3", + "[33] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3", + "[34] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3", + "[35] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings IEEE International Conf. on Computer Vision (ICCV), 2019. 3", + "[36] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 5", + "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4", + "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[41] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 2, 3, 6", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2", + "[43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2", + "[44] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 4", + "[45] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 72-88. Springer, 2022. 3", + "[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[47] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 4", + "[48] Peng-Shuai Wang. Octformer: Octtree-based transformers for 3D point clouds. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 4", + "[49] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-CNN: Octree-based convolutional neural networks for 3D shape analysis. ACM Transactions on Graphics (SIGGRAPH), 36(4), 2017. 2, 4, 5", + "[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 8", + "[51] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "560", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3", + "[52] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Toon-crafter: Generative cartoon interpolation. arXiv preprint arXiv:2405.17933, 2024. 2", + "[53] Yu-Ying Yeh, Jia-Bin Huang, Changil Kim, Lei Xiao, Thu Nguyen-Phuoc, Numair Khan, Cheng Zhang, Manmohan Chandraker, Carl S Marshall, Zhao Dong, et al. Texture-dreamer: Image-guided texture synthesis through geometry-aware diffusion. arXiv preprint arXiv:2401.09416, 2024. 3", + "[54] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[55] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206-4216, 2023. 2, 3, 5, 6", + "[56] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 2, 3, 6", + "[57] Hongkun Zhang, Zherong Pan, Congyi Zhang, Lifeng Zhu, and Xifeng Gao. Texpainter: Generative mesh texturing with multi-view consistency. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6", + "[58] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 3", + "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4, 8", + "[60] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, and Xiaogang Jin. Dreammat: High-quality pbr material generation with geometry- and light-aware diffusion models. ACM Trans. Graph., 43(4), 2024. 2, 3, 7, 8", + "[61] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 5" + ], + "bbox": [ + 91, + 92, + 482, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "561", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_model.json b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b777ea1484aaea1ad89443f33acfc4f5ccf33a4 --- /dev/null +++ b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_model.json @@ -0,0 +1,2026 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.211, + 0.131, + 0.79, + 0.177 + ], + "angle": 0, + "content": "TexGaussian: Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.203, + 0.82, + 0.24 + ], + "angle": 0, + "content": "Bojun Xiong\\(^{1*†}\\), Jialun Liu\\(^{2*}\\), Jiakui Hu\\(^{3†}\\), Chenming Wu\\(^{2}\\), Jinbo Wu\\(^{2}\\), Xing Liu\\(^{2}\\), Chen Zhao\\(^{2}\\), Errui Ding\\(^{2}\\), Zhouhui Lian\\(^{1‡}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.785, + 0.281 + ], + "angle": 0, + "content": "\\(^{1}\\)Wangxuan Institute of Computer Technology, Peking University, China \\(^{2}\\)Baidu VIS" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.282, + 0.733, + 0.301 + ], + "angle": 0, + "content": "\\(^{3}\\)Institute of Medical Technology, Peking University, China" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.319, + 0.905, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.513, + 0.908, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Our proposed TexGaussian is capable of generating high-quality materials for given input 3D meshes based on the corresponding textual descriptions. The generated results are naturally compatible with modern graphical engines for photo-realistic rendering under different environment maps." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.57, + 0.327, + 0.587 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.603, + 0.486, + 0.816 + ], + "angle": 0, + "content": "Physically Based Rendering (PBR) materials play a crucial role in modern graphics, enabling photorealistic rendering across diverse environment maps. Developing an effective and efficient algorithm that is capable of automatically generating high-quality PBR materials rather than RGB texture for 3D meshes can significantly streamline the 3D content creation. Most existing methods leverage pre-trained 2D diffusion models for multi-view image synthesis, which often leads to severe inconsistency between the generated textures and input 3D meshes. This paper presents TexGaussian, a novel method that uses octant-aligned 3D Gaussian Splitting for rapid PBR material generation. Specifically, we place each 3D Gaussian on the finest leaf node of the octree built from the input 3D mesh to render the multi-view" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.909, + 0.739 + ], + "angle": 0, + "content": "images not only for the albedo map but also for roughness and metallic. Moreover, our model is trained in a regression manner instead of diffusion denoising, capable of generating the PBR material for a 3D mesh in a single feed-forward process. Extensive experiments on publicly available benchmarks demonstrate that our method synthesizes more visually pleasing PBR materials and runs faster than previous methods in both unconditional and text-conditional scenarios, exhibiting better consistency with the given geometry. Our code and trained models are available at https://3d-aigc.github.io/TexasGaussian." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.754, + 0.646, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Traditional 3D asset creation relies heavily on the expertise and extensive effort of professional designers [23], posing a significant barrier for casual users interested in creating 3D models independently. In the 3D design process, geometry creation typically represents only a small portion of the overall time, while the majority is dedicated to developing textures and appearances, which are far more time-consuming. Achieving a delicate appearance for a 3D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.827, + 0.485, + 0.902 + ], + "angle": 0, + "content": "\\*Denotes equal contribution. \\(\\dagger\\) This work was partly done when Bojun and Jiakui interned in Baidu. \\(\\ddagger\\) Corresponding author. E-mail: lianzhouhui@pku.edu.cn This work was supported by the National Natural Science Foundation of China (Grant No.: 62372015), Center For Chinese Font Design and Research, and Key Laboratory of Intelligent Press Media Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "551" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "model often demands substantial time and effort from experienced designers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.124, + 0.483, + 0.26 + ], + "angle": 0, + "content": "Recently, Artificial Intelligence Generated Content (AIGC) based on deep generative models, especially diffusion models [17, 46] have been widely used to facilitate the process of artistic creation, catalyzing advancements in image generation [15, 37, 40, 42, 43] and video generation [3-5, 18, 52]. As a result, exploring effective ways to leverage deep generative models to streamline the creation of detailed appearances for 3D models has become a popular direction in the graphics and vision communities." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.262, + 0.485, + 0.549 + ], + "angle": 0, + "content": "Recent advancements in 3D texture generation attempt to use depth-conditional pre-trained 2D diffusion model [58] to synthesize RGB images based on the depth maps rendered from multiple views, such as TEXture [41] and text2tex [8]. Subsequent works [6, 20, 57] further improve multi-view consistency via blending the multi-view images to a single and consistent texture map in every denoising step. However, these methods struggle to have a global picture of 3D geometries due to the use of 2D diffusion models, leading to inconsistencies between the texture map and the semantics of the input 3D meshes. Moreover, the generated assets suffer from illumination-baked textures, which can significantly degrade the quality of the final rendering when placed in novel lighting conditions [56]. While DreamMat [60] supplements geometry and light control to achieve material decomposition through score distillation sampling [38], it still struggles to fully capture the global geometry features. This limitation often results in the multiface Janus problem and leads to over-saturated colors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.734 + ], + "angle": 0, + "content": "On the other hand, training a 3D neural network directly on 3D data, such as Point-UV Diffusion [55] and TexOct [26], is an effective way for 3D global consistency. Meanwhile, this avoids multi-view sampling and score distillation sampling, which accelerates the process of texture synthesis. However, relying on colored point clouds for 3D representation and supervision often results in blurred outputs, primarily due to the sparse and non-compact nature of point clouds in 3D space. Due to the limitations of the adopted 3D representations and the lack of training data on PBR materials, these approaches are incapable of generating high-fidelity PBR materials for 3D models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address the aforementioned challenges, this paper presents TexGaussian, a fast and high-fidelity PBR material generation model directly in 3D space that maintains 3D global consistency. Different from previous approaches that primarily rely on diffusion models, our method works in a regression manner to regress the PBR material from the input mesh for faster generation speed. To enable effective learning in 3D space, we propose to use octree, a specialized sparse voxel structure that efficiently organizes and preserves 3D information, which can be built from 3D point clouds sampled from the surface of the object. How" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.514 + ], + "angle": 0, + "content": "ever, directly regressing the color of 3D point clouds on octree often results in blurry textures as mentioned in [26]. To tackle the challenges of the incompact and discrete nature presented by points, we use 3D Gaussian Splatting (3DGS) [22], a robust representation that bridges the gap between 3D space and 2D raster images, allowing us to fully utilize rich 2D image information to alleviate blurring results. Specifically, for each input mesh, we sample the dense 3D point clouds on its surface to build the corresponding octree. On each octant (i.e., the finest leaf node of octree), we place a 3D Gaussian [22] at its central position. Then, we use the octree-based 3D U-Net [49] to predict the parameters of each 3D Gaussian on octants. Apart from RGB colors, we extend each 3D Gaussian with additional parameters to represent the roughness and metallicity of 3D objects. Multi-view images, including albedo, roughness, and metallic maps, can be rasterized from all these 3D Gaussians via 3DGS. The 3D U-Net is supervised by the difference between the predicted multi-view images and their corresponding ground truth. Notably, the 3D U-Net is trained to directly regress the multi-view images based on the geometry feature of the input 3D model, which further facilitates the process of 3D Gaussian prediction compared to diffusion manner. We train our TexGaussian model on a subset of Objaverse [13] with high-quality PBR materials, enabling fast PBR material generation with a single feedforward pass. In summary, the contributions of our paper are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.517, + 0.905, + 0.591 + ], + "angle": 0, + "content": "- We propose an octant-aligned 3D Gaussian Splitting method for high-quality PBR material synthesis on un-textured input 3D mesh, which fully utilizes the supervision from 2D images, avoiding blurry results caused by the discreteness of 3D point clouds." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.593, + 0.905, + 0.638 + ], + "angle": 0, + "content": "- We adopt a regression manner to train our 3D U-Net model instead of diffusion denoising, achieving faster generative speed." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.638, + 0.905, + 0.743 + ], + "angle": 0, + "content": "- We propose TexGaussian, a novel PBR material generation method based on the above two techniques. To our knowledge, our work is the first to generate PBR materials directly in 3D space. Qualitative and quantitative experiments have been conducted to verify the superiority of the quality and efficiency of our method over other existing approaches." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.517, + 0.905, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.753, + 0.655, + 0.769 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.905, + 0.825 + ], + "angle": 0, + "content": "In this section, we mainly summarize current texture synthesis methods, which can be roughly divided into three categories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.771, + 0.85 + ], + "angle": 0, + "content": "2.1. Multi-view Images Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Many previous works have tried to leverage the powerful T2I model to assist texture generation for 3D shapes. Specifically, they render the depth map of in" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "552" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.453 + ], + "angle": 0, + "content": "put 3D mesh from multiple views and use depth conditional T2I models [58] to synthesize RGB images and perform text-conditioned texture synthesis. TEXTure [41] and Text2Tex [8] iteratively paint a mesh from different views. However, images synthesized in early view could produce errors that are not reconcilable with the geometry that is observed in later views. Many subsequent works try to alleviate multi-view inconsistency via different alignment modules. TexFusion [6] proposes a sequential interlaced multi-view sampler that interleaves texture assembling with denoising steps in different camera views. Similarly, TexGen [20] directly enforces view consistent sampling in RGB texture space and develops a noise resampling strategy to retain rich texture details. TexPainter [57] blends images from different views into a common color-space texture image by weighted averaging to guarantee multi-view consistency. To remove light influence from 2D diffusion models, Paint3D [56] contribute separate UV Inpainting and UVHD diffusion models specialized in shape-aware refinement. More recently, Meta 3D TextureGen [1] utilizes geometry-aware text-to-image model to generate a multiview image for texture. Although these methods achieve impressive texture results, they can still hardly comprehend the overall geometry of input 3D mesh." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.47, + 0.395, + 0.484 + ], + "angle": 0, + "content": "2.2. Optimization-based 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.495, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Before the emergence of large-scale Text-to-Image generative model, earlier methods [10, 19, 30, 31, 33] propose to optimize texture map of 3D object via natural language supervised visual model, CLIP [39]. Subsequently, score distillation sampling (SDS) was adopted by DreamFusion [38] and Magic3D [25]. The key idea is to optimize 3D representations such as NeRF [32] or InstantNGP [34] with the gradient guidance from 2D diffusion priors [38, 51]. To generate PBR material, TextureDreamer [53] optimizes spatially-varying bidirectional reflectance distribution (BRDF) field through personalized geometric-aware score distillation. Fantasia3D [9] uses a single predefined environmental. However, the generated images from diffusion models may not be consistent with the given environment light. Another concurrent work Paint-it [54] is similar to Fantasia3D [9] but with Deep Convolutional Physically-Based Rendering (DC-PBR) parameterization. DreamMat [60] proposes a novel geometry and light-aware diffusion model, which is trained to generate images that are consistent with the given environment light. FlashTex[14] uses a light-conditioned diffusion model within a two-stage pipeline, combining reconstruction and SDS optimization to enhance texture quality and achieve better light disentanglement. However, these methods struggle with the Janus problem due to the semantically ambiguous. And the time consumption is relatively too long to use in practice." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.092, + 0.809, + 0.107 + ], + "angle": 0, + "content": "2.3. Generating Texture from 3D Data" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.116, + 0.904, + 0.372 + ], + "angle": 0, + "content": "The most straightforward way to synthesize texture map for 3D mesh is to train generative model directly from 3D data with texture ground truth [7, 11-13]. Early methods such as Texture Fields [35] learn implicit texture fields to assign a color to each pixel on the surface of the 3D shape. Texturify [45] devices face convolution operation on mesh surface to predict texture on each face. It employs differentiable rendering with an adversarial loss to ensure that generated textures produce realistic imagery. Recently, some diffusion-based texture synthesis methods, such as PointUV [55] and TexOct [26] train a denoising network on colors of point clouds which are further mapped to 2D UV map. Although these methods achieve better 3D global consistency with input mesh, they are only trained on several categories of small datasets [7]. What's more, discrete supervision from 3D point clouds leads to suboptimal results compared with continuous signals such as 2D images." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.39, + 0.602, + 0.404 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.419, + 0.904, + 0.553 + ], + "angle": 0, + "content": "In this section, we provide a detailed explanation of our proposed method, TexGaussian. The overall pipeline of our method is shown in Fig. 2. Existing texture synthesis approaches mainly rely on pre-trained 2D diffusion models, which struggle to fully understand the overall 3D structure. This often leads to misalignment between the generated texture map and 3D semantics. Our goal is to synthesize the high-quality PBR material for a given mesh directly in 3D space." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.575, + 0.622, + 0.588 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.6, + 0.904, + 0.9 + ], + "angle": 0, + "content": "To enable effective learning in 3D space, we use octree, a sparse voxel structure, to organize and store 3D information without compromising representation quality due to the inherent sparseness of 3D objects in 3D space. Thus, we sample a large number of points on the surface of the given mesh to construct the corresponding octree. The key components of our method are the octant-aligned Gaussian Splatting and the octree-based 3D U-Net. Specifically, we place a 3D Gaussian at the center of each octree's finest leaf node. The octree-based 3D-Unet is trained to predict the parameters of each 3D Gaussian. Under this circumstance, the generated octant-aligned 3D Gaussians are naturally on the mesh surface. We render them from multiple viewpoints using 3D Gaussian Splatting and train the 3D U-Net by minimizing the difference in 2D raster images and 3D Gaussian parameters. During inference, the rendered multi-view images are baked into the UV space of input mesh using differentiable mesh rendering, producing the final texture and material map. The details of 3D Gaussian Splatting (3DGS) are provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.947, + 0.51, + 0.957 + ], + "angle": 0, + "content": "553" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.905, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.248, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Figure 2. An overview of our PBR material generation framework. (a) We propose octant-aligned 3D Gaussian Splitting, which positions a 3D Gaussian at the center of each finest leaf node of the constructed octree. Additional channels are added at the end of the Gaussian parameters to model PBR material. (b) We use the 3D U-Net built upon octree-based convolutional networks to predict the Gaussian parameters. Our octree-based 3D U-Net is trained by minimizing the difference on 2D raster images and 3D Gaussian parameters. (c) We bake the multi-view rendered images to the UV space of the input 3D model to realize physically based rendering under new illumination environments." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.35, + 0.419, + 0.367 + ], + "angle": 0, + "content": "3.2. Octant-aligned 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.372, + 0.485, + 0.644 + ], + "angle": 0, + "content": "For a given mesh, we first sample \\( N = 100,000 \\) 3D points on its surface. The corresponding octree is then built by adaptively subdividing the voxels containing those points until the maximum depth is reached. As a result, all of the finest leaf nodes of octree lie along the boundary of the 3D object, which is consistent with the characteristic of the optimized 3D Gaussian. Therefore, we align a 3D Gaussian at the central position of each finest leaf node to effectively model the appearance without compromising the splatting quality. It is worth noting that we do not adjust the position of 3D Gaussians because they are already on the surface of 3D shape and adding additional offset relative to the octant center would not improve the rendering quality through our early experiments. To model PBR material that includes roughness and metallic information, we follow [21, 44] to append two additional channels at the end of 3D Gaussian parameters which are responsible for roughness and metallic map rendering, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.884 + ], + "angle": 0, + "content": "For each 3D object in our dataset, we render multi-view images of albedo, roughness, and metallic maps for the training of 3D Gaussian. The multi-view rendering results of the albedo map are view-independent. So we just use three RGB channels to take the place of the original spherical harmonics. As noted, we exclude the position from 3D Gaussian parameters in our model. Thus, our 3D Gaussian parameters consist of 13 channels in total: one for opacity, three for scale, four for rotation, three for albedo, one for roughness, and one for metallic. To stabilize training, we choose to employ different activation functions compared to the original Gaussian Splatting [22]. Specifically, We multiply the softplus-activated scales \\(\\mathbf{s}_i\\) with 0.01, ensuring that the initial 3D Gaussians conform to the object's counter of object at the beginning of training rather than expanding outward." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We pre-fitting the parameters of 3D Gaussians on the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.397 + ], + "angle": 0, + "content": "constructed octree for each 3D object in our dataset via the original loss function in [22] on multi-view albedo images, roughness maps, and metallic maps:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.412, + 0.907, + 0.429 + ], + "angle": 0, + "content": "\\[\nL = (1 - \\lambda) L _ {1} + \\lambda L _ {\\mathrm {D - S S I M}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.436, + 0.619, + 0.451 + ], + "angle": 0, + "content": "where \\(\\lambda = 0.2\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.462, + 0.8, + 0.478 + ], + "angle": 0, + "content": "3.3. Octree-based 3D U-Net Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.906, + 0.589 + ], + "angle": 0, + "content": "To handle the encoding of our octant-aligned 3D Gaussian representation, we use the 3D U-Net built upon octree-based convolutional neural networks [49] to predict the Gaussian parameters. Inspired by LGM [47], the output feature of the 3D U-Net on each octant is treated as the 3D Gaussian parameters, which contain 13 channels, as discussed in the last subsection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.818 + ], + "angle": 0, + "content": "To effectively train our 3D U-Net, we adopt the regressive loss objective, which could further facilitate the generation process. The input to our octree-based 3D U-Net is the geometry feature on each octree's finest leaf node, such as normal and local displacement. For text-conditioned PBR material synthesis, the text feature is extracted by pretrained CLIP model [39] and is fed to U-Net via the octree-based multi-head cross attention mechanism similar to [48]. The predicted 3D Gaussians are rasterized from multiple views via Gaussian Splatting [22]. At each training step, we rasterize the albedo images, alpha images, roughness and metallic maps from randomly selected eight views. Following [47], we apply the mean square error (MSE) loss and the VGG-based LPIPS loss [59] to the albedo image, roughness map, and metallic map:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.827, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {a l b e d o}} = L _ {\\text {M S E}} \\left(I _ {\\text {a l b e d o}}, I _ {\\text {a l b e d o}} ^ {\\mathrm {G T}}\\right) + L _ {\\text {L P I P S}} \\left(I _ {\\text {a l b e d o}}, I _ {\\text {a l b e d o}} ^ {\\mathrm {G T}}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.86, + 0.905, + 0.878 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {R}} = L _ {\\mathrm {M S E}} \\left(I _ {\\mathrm {R}}, I _ {\\mathrm {R}} ^ {\\mathrm {G T}}\\right) + L _ {\\mathrm {L P I P S}} \\left(I _ {\\mathrm {R}}, I _ {\\mathrm {R}} ^ {\\mathrm {G T}}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {M}} = L _ {\\mathrm {M S E}} \\left(I _ {\\mathrm {M}}, I _ {\\mathrm {M}} ^ {\\mathrm {G T}}\\right) + L _ {\\mathrm {L P I P S}} \\left(I _ {\\mathrm {M}}, I _ {\\mathrm {M}} ^ {\\mathrm {G T}}\\right), \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "554" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "where 'R' and 'M' denote roughness and metallic, respectively. We further apply the MSE loss on the alpha image for faster convergence of the shape:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.146, + 0.483, + 0.165 + ], + "angle": 0, + "content": "\\[\nL _ {\\alpha} = L _ {\\mathrm {M S E}} \\left(I _ {\\alpha}, I _ {\\alpha} ^ {\\mathrm {G T}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.484, + 0.25 + ], + "angle": 0, + "content": "To accelerate the coverage process, we also apply the 3D MSE loss \\( L_{\\mathrm{3D}} \\), which calculates the difference between predicted parameters of 3D Gaussians and pre-fitting ones. Finally, the complete loss function of our model is defined as the sum of all the above losses:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.262, + 0.483, + 0.278 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t o t a l}} = L _ {\\text {a l b e d o}} + L _ {\\mathrm {R}} + L _ {\\mathrm {M}} + L _ {\\alpha} + L _ {\\mathrm {3 D}}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.288, + 0.353, + 0.304 + ], + "angle": 0, + "content": "3.4. Texture and Material Baking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.536 + ], + "angle": 0, + "content": "In the inference stage, we also first build the corresponding octree for the input 3D mesh. Then, we use our trained octree-based 3D U-Net to generate 3D Gaussian on every octant and rasterize them from multiple views. The ultimate output of our method should be a global texture map and material map. Thus, we rasterize the input mesh using the differentiable renderer [24] and optimize its albedo, roughness and metallic parameters via the MSE loss between the Nvdiffrast [24] rendering results and 3D Gaussian rendering results. With adequately optimized implementation, this process takes only about several seconds to bake the multiview images to an untextured 3D model. After the optimization, our input mesh paired with its albedo and material map is capable of performing physically based rendering in new illumination environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.55, + 0.224, + 0.567 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.308, + 0.59 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Dataset We train our TexGaussian model on two publicly-available datasets: ShapeNet [7] and Objverse [13]. For 3D objects in the ShapeNet dataset, since they only contain albedo maps without PBR materials, our model is trained to generate RGB textures only. We train our model on four categories of ShapNet: bench, car, chair and table which is consistent with previous works [26, 55]. We curate a subset of Objverse models encoded with PBR materials and convert those using the specular-glossiness workflow to the metallic-roughness workflow to achieve a consistent PBR representation. This process resulted in a total of 29,200 models. For each 3D object, we render its RGBA albedo image, roughness map, and metallic map from 64 views of \\(512^2\\) for training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Network Architecture Our 3D U-Net is built upon octree-based convolutional neural networks [49] which only operates on non-empty octree leaf nodes. The depth of our constructed octree is set to 8 (resolution \\(256^{2}\\)). Our 3D U-Net consists of 5 down-sampling and up-sampling blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.213 + ], + "angle": 0, + "content": "For the text-conditioned generation, cross-attention layers are only inserted at the last two down-sampling blocks, the middle block, and the first two up-sampling blocks. The input channel of our 3D U-Net is set to 4, where 3 for normal vector and 1 for local displacement while the output channel is set to 13 as mentioned above. The resolution of 3D Gaussian Splatting is set to \\(512^2\\). The resolution of the baked albedo and material map is set to \\(1024^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.22, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Training Details We train our model in a single-category and unconditional manner on ShapeNet, i.e., there are 4 TexGaussian models without text conditions in total. This per-category model is trained on 4 NVIDIA A100 (40G) GPUs for about 2 days and is set up for a fair comparison with other existing methods trained on ShapeNet. For the Objaverse dataset, we train a text-conditioned TexGaussian model for PBR material generation on 24 NVIDIA A100 (40G) GPUs for two weeks. For each batch, we randomly sample 8 camera views to calculate loss functions. We adopt the AdamW [27] optimizer with a learning rate of \\(4 \\times 10^{-4}\\), a weight decay of 0.05, and betas of (0.9,0.95). The learning rate is cosine annealed to 0 during the training and we clip the gradient with a maximum norm of 1.0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.909, + 0.633 + ], + "angle": 0, + "content": "Evaluation metrics To effectively assess the quality of our generative results, we adopt the metric proposed by [61]. Specifically, each mesh with the generated PBR material is rendered from 20 uniformly distributed views to get multi-view albedo maps, and PBR rendering results in a new illumination environment. These images are used to calculate the FID [36] and KID [2] scores against those from ground truth to evaluate the quality and diversity of generative PBR materials. For the ShapeNet dataset, we only use albedo maps. The final score is averaged across 20 views, and a lower FID and KID score indicates better generation quality and diversity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.857, + 0.657 + ], + "angle": 0, + "content": "4.2. Unconditional RGB Texture Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.906, + 0.785 + ], + "angle": 0, + "content": "We conduct unconditional RGB texture instead of PBR material generation on four categories of the ShapeNet dataset. We use the same train and test data split as in [26, 55]. Fig. 3 shows the unconditionally generated RGB texture by our single-category model with high quality, fidelity, and diversity. It can be seen that the texture map synthesized by our method is of great 3D global consistency with the corresponding input geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison We conduct quantitative analysis and comparison on our model and other state-of-the-art methods. Specifically, we compare our model with PointUV Diffusion [55] and TexOct [26] which also train single-category models on ShapeNet. We also compare with some methods using pre-trained 2D diffusion prior for multi-view images synthesis such as TEXTure [55], Text2Tex [8]" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "555" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.133 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of the FID and KID \\((\\times 10^{2})\\) score as well as the inference time of TexGaussian and other methods on the ShapeNet dataset [7]. The top part reports the comparison with methods that also train per-category models on ShapeNet. The bottom part reports the comparison with 2D diffusion-based methods, which select 50 samples from the test set of each category." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.143, + 0.907, + 0.319 + ], + "angle": 0, + "content": "
MethodsAverageBenchCarChairTableTime
FID↓KID↓FID↓KID↓FID↓KID↓FID↓KID↓FID↓KID↓
Point-UV (1-Stage) [55]88.435.7864.961.44186.1017.5746.231.9356.422.1939.92s
Point-UV (2-Stage) [55]61.492.6767.481.6189.385.8241.331.5347.751.7249.75s
TexOct [26]59.452.6060.460.9790.106.1137.701.3649.521.9717.44s
Ours49.762.0746.370.4480.205.2229.961.0942.521.5411.02s
TEXTure [41]169.827.94152.535.28236.5818.52119.712.86170.465.11132.21s
Text2Tex [8]156.626.68154.955.79188.6015.06125.343.11157.572.771005.58s
Paint3D [56]135.254.67104.161.68204.2213.06104.642.24127.981.71231.56s
Ours100.971.8886.370.75117.535.2094.890.57105.090.9511.12s
" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.331, + 0.473, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Figure 3. Unconditional RGB texture generative results on ShapeNet. Please zoom in for a better inspection of color details. Paint3D [56], and TexPainter [57] whose input text prompts are set to \"a *\" and \"* is the name of corresponding category. Due to the relatively long time consumption, we only use 50 3D objects selected from the test set to evaluate the methods based on 2D diffusion models. We conduct only a qualitative comparison for TexPainter [57] for its lengthy processing time. Table 1 reports the comparison of FID and KID scores as well as average inference time on a single NVIDIA A100 (40G) GPU. It is worth noting that our model only takes about one second to predict the parameters of each 3D Gaussian and the rest of the time is used for texture baking. From Table 1, we have the following observation. First, our method obtains the best performance in terms of FID and KID. For example, our method outperforms TexOct [26], by an average of 9.69 in FID and 0.53 in KID. These improvements indicate that our method excels at generating high-quality textures. Second, TexGaussian achieves the fastest generative speed in all categories, which is much less time-consuming than other methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Qualitative Comparison Fig. 4 provides some qualitative results on the same input 3D meshes by different methods. Point-UV [55] and TexOct [26] use colored point clouds as supervision to train diffusion models. As a consequence, the generated texture map is relatively blurry due to the discreteness of the point cloud. For other methods that leverage pre-trained 2D diffusion prior, they can hardly comprehend the total geometry. The generated texture maps" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.339, + 0.907, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.775, + 0.908, + 0.818 + ], + "angle": 0, + "content": "Figure 4. Examples of generated RGB textures obtained by Tex-Gaussian and other state-of-the-art models on the same 3D object. Please zoom in for a better inspection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.909, + 0.908 + ], + "angle": 0, + "content": "by them are not consistent with the semantics of input 3D mesh, such as the chaotic stripes on the bench and chair in Fig. 4. On the contrary, our method generates smooth and colorful RGB textures on unseen objects, which align well with 3D meshes." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "556" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.315, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison with Fantasia3D [9], FlashTex [14], DreamMat [60] and Make-it-Real [16]. We show the rendered image, albedo map, roughness map, and metallic map for each 3D object." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.369, + 0.46, + 0.384 + ], + "angle": 0, + "content": "4.3. Text-conditioned PBR Material Generation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.392, + 0.484, + 0.573 + ], + "angle": 0, + "content": "We train the text-conditioned PBR material generation model on our filtered subset of Objaverse [13]. We use the text descriptions of 3D objects from Cap3D [28, 29] to train our model. We choose 29,000 3D objects in our filtered subset for training and the rest 200 for testing. Fig. 1 shows some generative results by proposed TexGaussian on the test set. We can see that our model comprehends the overall geometry feature of input 3D mesh. It is capable of synthesizing high-quality albedo, roughness, and metallic, which are aligned well with 3D semantics such as accents with high metallic on the wooden box and rope with high roughness on the jug." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.829 + ], + "angle": 0, + "content": "Quantitative Comparison We conduct quantitative comparison with some state-of-the-art text-conditioned PBR material synthesis methods: Fantasia3D [9], FlashTex [14], and DreamMat [60] on our test set. Table 2 reports FID and KID scores on both multi-view albedo map and PBR rendering images under the same illumination environment. TexGaussian outperforms all the baseline methods in achieving the best visual quality of the generated appearances. We also report the average inference time of different methods on a single NVIDIA A100 (40G) GPU across our test set. Due to the iterative optimization of score distillation sampling, all other three methods cost at least 20 minutes for the generation. In contrast, our TexGaussian only takes about 20 seconds, in which one second is for predicting the Gaussian parameters and the rest is for baking, which performs \\(60 \\times\\) faster than previous methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparison Fig. 5 visualizes the generated albedo, roughness, and metallic of each compared method from the same text prompt and untextured meshes. We also show the PBR rendering images of the generated materials" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.908, + 0.747 + ], + "angle": 0, + "content": "under the same environment light. It can be observed that Fantasia3D [9] and FlashTex [14] generate irregular colors on mesh surfaces, which lead to rendering results with low quality. DreamMat [60] is capable of generating a visually pleasing appearance for test 3D models. However, it can hardly align the generated PBR material well with the 3D semantics, such as the black band on the wooden barrel. What's more, it tends to generate over-saturated colors, which is demonstrated in the wooden chest and green grenade. In addition, all of the compared methods struggle to completely disentangle the light and texture, which results in relatively chaotic results of generated metallic and roughness. We also compare our method with a Multimodal Large Language Model based method, Make-it-Real [16], which accepts 3D objects paired with albedo maps generated by our TexGaussian as input. It can be observed that Make-it-Real struggles to capture the overall 3D structure of input meshes. As a consequence, it tends to predict single-value PBR maps, leading to irregular rendering results, such as the wooden chest with extremely large metallic value. On the contrary, our method is directly trained from 3D original data. It is capable of generating clean and smooth PBR material while fully comprehending the overall 3D structure. We provide more generative results in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.667, + 0.774 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Essentials of Octree We first analyze the effectiveness of using Octree in our pipeline. To do this, we calculate the quantitative quality of 3D Gaussian pre-fitting by our octant-aligned and full voxel-aligned 3D Gaussian Splitting on the subset of Objaverse dataset [13]. The depth of octree in this ablation is set to 8, and the resolution of full voxels is \\(256^3\\), which is consistent with the finest resolution of octree to guarantee comparison fairness. Table 3 reports the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "557" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of the FID and KID \\((\\times 10^{2})\\) scores as well as the inference time of TexGaussian and other methods on our test set which consists of 200 3D objects with ground truth PBR materials." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.156, + 0.487, + 0.251 + ], + "angle": 0, + "content": "
MethodsAlbedoPBR renderingTime
FID↓KID↓FID↓KID↓
Fantasia3D [9]213.210.96209.870.4822.7mins
FlashTex [14]185.241.13186.820.4220.3mins
DreamMat [60]152.631.09145.490.1948.2mins
Ours123.720.20129.520.0221.04s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.255, + 0.485, + 0.283 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of methods using different types of 3D Gaussian Fitting." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.294, + 0.485, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.519, + 0.485, + 0.563 + ], + "angle": 0, + "content": "Figure 6. Visualization of different manners of Gaussian fitting. The rendering results demonstrate excellent reconstruction performance of the proposed octant-aligned 3D Gaussian Splatting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.706 + ], + "angle": 0, + "content": "PSNR, LPIPS [59], and SSIM [50] metrics on the albedo map as well as the average number of 3D Gaussians of octant and full voxel aligned 3D Gaussian Splitting across our Objverse subset. Fig. 6 also presents some visual results of our Gaussian fitting. These results demonstrate that the proposed octant-aligned 3D Gaussian produces a much more reasonably compact and precise representation of diverse and complex 3D assets with much fewer 3D Gaussians compared to the full voxel version." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.484, + 0.756 + ], + "angle": 0, + "content": "Essentials of different losses We verify the importance of different losses we proposed to train our model. We term all the losses calculated on 2D image space as \\( L_{2\\mathrm{D}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.768, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\nL _ {2 \\mathrm {D}} = L _ {\\text {a l b e d o}} + L _ {\\mathrm {R}} + L _ {\\mathrm {M}} + L _ {\\alpha}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We train two additional TexGaussian models using only the \\( L_{2\\mathrm{D}} \\) or \\( L_{3\\mathrm{D}} \\) loss on ShapeNet car category due to its large variations and complexity of texture to validate their effectiveness. It is worth noting that in this dataset, \\( L_{2\\mathrm{D}} = L_{\\mathrm{albedo}} + L_{\\alpha} \\) due to the lack of material information. The training curves of MSE loss and LPIPS loss between rendering images and ground-truth ones are shown" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.089, + 0.709, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.089, + 0.897, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.218, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Figure 7. The training loss curves of our model with different losses." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.258, + 0.907, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.374, + 0.893, + 0.39 + ], + "angle": 0, + "content": "Figure 8. Results of our model trained with and without \\( L_{2D} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.567 + ], + "angle": 0, + "content": "in Fig. 7, which demonstrate the effects of \\( L_{2D} \\) and \\( L_{3D} \\). From the loss curves, we can conclude that \\( L_{3\\mathrm{D}} \\) facilitates the process of convergence and \\( L_{2D} \\) enhances the quality of synthesized images, as evidenced by the large margin of improvement in the LPIPS loss when introducing \\( L_{2D} \\). We also provide some qualitative results to verify the effectiveness of \\( L_{2D} \\) on the test set of car category in Fig. 8. Only using \\( L_{3D} \\) results in a relatively blurry texture map due to the discreteness of 3D Gaussian which is similar to the characteristic of 3D point clouds analyzed above." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.578, + 0.634, + 0.594 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.907, + 0.801 + ], + "angle": 0, + "content": "In this paper, we proposed TexGaussian, an octree-based 3D Gaussian Splatting model for high-quality PBR material generation on untextured meshes. We aligned each 3D Gaussian on the octant of the corresponding octree built from the input untextured object and extended the parameters of 3D Gaussian with additional channels to represent the roughness and metallic map. We trained our model with regression objectives, achieving faster inference speed compared to previous texture synthesis methods. Experimental results demonstrated that our method is capable of generating high-quality PBR materials that are readily usable in modern graphics engines for photo-realistic rendering, offering enhanced realism for a variety of applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Limitations The generalization of TexGaussian is still hindered by the scale of the training set. Thus it struggles to generate various textures for some extremely complex 3D objects beyond our training data. We are looking forward to training our TexGaussian model with more parameters and more data on a larger-scale GPU cluster in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "558" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.215 + ], + "angle": 0, + "content": "[2] Mikołaj Binkowski, Dougal J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.289, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.485, + 0.514 + ], + "angle": 0, + "content": "[6] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and KangXue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[7] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[11] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.485, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.188 + ], + "angle": 0, + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. arXiv preprint arXiv:2307.05663, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.19, + 0.907, + 0.272 + ], + "angle": 0, + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.342 + ], + "angle": 0, + "content": "[14] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. Flashtex: Fast reconfigurable mesh texturing with lightcontrolnet. In European Conference on Computer Vision (ECCV), 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[16] Ye Fang, Zeyi Sun, Tong Wu, Jiaqi Wang, Ziwei Liu, Gordon Wetzstein, and Dahua Lin. Make-it-real: Unleashing large multimodal model for painting 3d objects with realistic materials, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[19] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[20] Dong Huo, Zixin Guo, Xinxin Zuo, Zhihao Shi, Juwei Lu, Peng Dai, Songcen Xu, Li Cheng, and Yee-Hong Yang. Tex-gen: Text-guided 3d texture generation with multi-view sampling and resampling. ECCV, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[21] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian-shader: 3d gaussian splattering with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[23] Matthias Labschütz, Katharina Krosl, Mariebeth Aquino, Florian Grashäftl, and Stephanie Kohl. Content creation for a 3d game with maya and unity 3d. Institute of Computer Graphics and Algorithms, Vienna University of Technology, 6(124):2, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "559" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[25] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[26] Jialun Liu, Chenming Wu, Xinqi Liu, Xing Liu, Jinbo Wu, Haotian Peng, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Texoct: Generating textures of 3d models with octree-based diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4284-4293, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.249, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[28] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.321, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[29] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. arXiv preprint arXiv:2404.07984, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.363, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[30] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.448, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[31] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.519, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.576, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[33] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.633, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[34] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[35] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings IEEE International Conf. on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[36] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[41] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.335, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[44] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[45] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 72-88. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[47] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[48] Peng-Shuai Wang. Octformer: Octtree-based transformers for 3D point clouds. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[49] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-CNN: Octree-based convolutional neural networks for 3D shape analysis. ACM Transactions on Graphics (SIGGRAPH), 36(4), 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "560" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "tion. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[52] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Toon-crafter: Generative cartoon interpolation. arXiv preprint arXiv:2405.17933, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[53] Yu-Ying Yeh, Jia-Bin Huang, Changil Kim, Lei Xiao, Thu Nguyen-Phuoc, Numair Khan, Cheng Zhang, Manmohan Chandraker, Carl S Marshall, Zhao Dong, et al. Texture-dreamer: Image-guided texture synthesis through geometry-aware diffusion. arXiv preprint arXiv:2401.09416, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[54] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[55] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206-4216, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[56] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[57] Hongkun Zhang, Zherong Pan, Congyi Zhang, Lifeng Zhu, and Xifeng Gao. Texpainter: Generative mesh texturing with multi-view consistency. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[58] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.724 + ], + "angle": 0, + "content": "[60] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, and Xiaogang Jin. Dreammat: High-quality pbr material generation with geometry- and light-aware diffusion models. ACM Trans. Graph., 43(4), 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.482, + 0.783 + ], + "angle": 0, + "content": "[61] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "561" + } + ] +] \ No newline at end of file diff --git a/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_origin.pdf b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..35c91b58137b2a0bd5d98e0bd11a5ad376cf94e3 --- /dev/null +++ b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/80f51f64-f3e1-48f6-bd31-effee985225f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:273071093e66af223a0419634cf32ec1f348a8db390e507b31ebb43f789a20da +size 6012772 diff --git a/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/full.md b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..62c856d5986e59c197e67ccadc7207615b01319b --- /dev/null +++ b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/full.md @@ -0,0 +1,270 @@ +# TexGaussian: Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting + +Bojun Xiong $^{1*†}$ , Jialun Liu $^{2*}$ , Jiakui Hu $^{3†}$ , Chenming Wu $^{2}$ , Jinbo Wu $^{2}$ , Xing Liu $^{2}$ , Chen Zhao $^{2}$ , Errui Ding $^{2}$ , Zhouhui Lian $^{1‡}$ + +$^{1}$ Wangxuan Institute of Computer Technology, Peking University, China $^{2}$ Baidu VIS + +$^{3}$ Institute of Medical Technology, Peking University, China + +![](images/736babe93e777639cd2055742041e084ca3b71ed3d21f9ca7a61e16f1d117ab9.jpg) +Figure 1. Our proposed TexGaussian is capable of generating high-quality materials for given input 3D meshes based on the corresponding textual descriptions. The generated results are naturally compatible with modern graphical engines for photo-realistic rendering under different environment maps. + +# Abstract + +Physically Based Rendering (PBR) materials play a crucial role in modern graphics, enabling photorealistic rendering across diverse environment maps. Developing an effective and efficient algorithm that is capable of automatically generating high-quality PBR materials rather than RGB texture for 3D meshes can significantly streamline the 3D content creation. Most existing methods leverage pre-trained 2D diffusion models for multi-view image synthesis, which often leads to severe inconsistency between the generated textures and input 3D meshes. This paper presents TexGaussian, a novel method that uses octant-aligned 3D Gaussian Splitting for rapid PBR material generation. Specifically, we place each 3D Gaussian on the finest leaf node of the octree built from the input 3D mesh to render the multi-view + +images not only for the albedo map but also for roughness and metallic. Moreover, our model is trained in a regression manner instead of diffusion denoising, capable of generating the PBR material for a 3D mesh in a single feed-forward process. Extensive experiments on publicly available benchmarks demonstrate that our method synthesizes more visually pleasing PBR materials and runs faster than previous methods in both unconditional and text-conditional scenarios, exhibiting better consistency with the given geometry. Our code and trained models are available at https://3d-aigc.github.io/TexasGaussian. + +# 1. Introduction + +Traditional 3D asset creation relies heavily on the expertise and extensive effort of professional designers [23], posing a significant barrier for casual users interested in creating 3D models independently. In the 3D design process, geometry creation typically represents only a small portion of the overall time, while the majority is dedicated to developing textures and appearances, which are far more time-consuming. Achieving a delicate appearance for a 3D + +model often demands substantial time and effort from experienced designers. + +Recently, Artificial Intelligence Generated Content (AIGC) based on deep generative models, especially diffusion models [17, 46] have been widely used to facilitate the process of artistic creation, catalyzing advancements in image generation [15, 37, 40, 42, 43] and video generation [3-5, 18, 52]. As a result, exploring effective ways to leverage deep generative models to streamline the creation of detailed appearances for 3D models has become a popular direction in the graphics and vision communities. + +Recent advancements in 3D texture generation attempt to use depth-conditional pre-trained 2D diffusion model [58] to synthesize RGB images based on the depth maps rendered from multiple views, such as TEXture [41] and text2tex [8]. Subsequent works [6, 20, 57] further improve multi-view consistency via blending the multi-view images to a single and consistent texture map in every denoising step. However, these methods struggle to have a global picture of 3D geometries due to the use of 2D diffusion models, leading to inconsistencies between the texture map and the semantics of the input 3D meshes. Moreover, the generated assets suffer from illumination-baked textures, which can significantly degrade the quality of the final rendering when placed in novel lighting conditions [56]. While DreamMat [60] supplements geometry and light control to achieve material decomposition through score distillation sampling [38], it still struggles to fully capture the global geometry features. This limitation often results in the multiface Janus problem and leads to over-saturated colors. + +On the other hand, training a 3D neural network directly on 3D data, such as Point-UV Diffusion [55] and TexOct [26], is an effective way for 3D global consistency. Meanwhile, this avoids multi-view sampling and score distillation sampling, which accelerates the process of texture synthesis. However, relying on colored point clouds for 3D representation and supervision often results in blurred outputs, primarily due to the sparse and non-compact nature of point clouds in 3D space. Due to the limitations of the adopted 3D representations and the lack of training data on PBR materials, these approaches are incapable of generating high-fidelity PBR materials for 3D models. + +To address the aforementioned challenges, this paper presents TexGaussian, a fast and high-fidelity PBR material generation model directly in 3D space that maintains 3D global consistency. Different from previous approaches that primarily rely on diffusion models, our method works in a regression manner to regress the PBR material from the input mesh for faster generation speed. To enable effective learning in 3D space, we propose to use octree, a specialized sparse voxel structure that efficiently organizes and preserves 3D information, which can be built from 3D point clouds sampled from the surface of the object. How + +ever, directly regressing the color of 3D point clouds on octree often results in blurry textures as mentioned in [26]. To tackle the challenges of the incompact and discrete nature presented by points, we use 3D Gaussian Splatting (3DGS) [22], a robust representation that bridges the gap between 3D space and 2D raster images, allowing us to fully utilize rich 2D image information to alleviate blurring results. Specifically, for each input mesh, we sample the dense 3D point clouds on its surface to build the corresponding octree. On each octant (i.e., the finest leaf node of octree), we place a 3D Gaussian [22] at its central position. Then, we use the octree-based 3D U-Net [49] to predict the parameters of each 3D Gaussian on octants. Apart from RGB colors, we extend each 3D Gaussian with additional parameters to represent the roughness and metallicity of 3D objects. Multi-view images, including albedo, roughness, and metallic maps, can be rasterized from all these 3D Gaussians via 3DGS. The 3D U-Net is supervised by the difference between the predicted multi-view images and their corresponding ground truth. Notably, the 3D U-Net is trained to directly regress the multi-view images based on the geometry feature of the input 3D model, which further facilitates the process of 3D Gaussian prediction compared to diffusion manner. We train our TexGaussian model on a subset of Objaverse [13] with high-quality PBR materials, enabling fast PBR material generation with a single feedforward pass. In summary, the contributions of our paper are threefold: + +- We propose an octant-aligned 3D Gaussian Splitting method for high-quality PBR material synthesis on un-textured input 3D mesh, which fully utilizes the supervision from 2D images, avoiding blurry results caused by the discreteness of 3D point clouds. +- We adopt a regression manner to train our 3D U-Net model instead of diffusion denoising, achieving faster generative speed. +- We propose TexGaussian, a novel PBR material generation method based on the above two techniques. To our knowledge, our work is the first to generate PBR materials directly in 3D space. Qualitative and quantitative experiments have been conducted to verify the superiority of the quality and efficiency of our method over other existing approaches. + +# 2. Related Work + +In this section, we mainly summarize current texture synthesis methods, which can be roughly divided into three categories. + +# 2.1. Multi-view Images Synthesis + +Many previous works have tried to leverage the powerful T2I model to assist texture generation for 3D shapes. Specifically, they render the depth map of in + +put 3D mesh from multiple views and use depth conditional T2I models [58] to synthesize RGB images and perform text-conditioned texture synthesis. TEXTure [41] and Text2Tex [8] iteratively paint a mesh from different views. However, images synthesized in early view could produce errors that are not reconcilable with the geometry that is observed in later views. Many subsequent works try to alleviate multi-view inconsistency via different alignment modules. TexFusion [6] proposes a sequential interlaced multi-view sampler that interleaves texture assembling with denoising steps in different camera views. Similarly, TexGen [20] directly enforces view consistent sampling in RGB texture space and develops a noise resampling strategy to retain rich texture details. TexPainter [57] blends images from different views into a common color-space texture image by weighted averaging to guarantee multi-view consistency. To remove light influence from 2D diffusion models, Paint3D [56] contribute separate UV Inpainting and UVHD diffusion models specialized in shape-aware refinement. More recently, Meta 3D TextureGen [1] utilizes geometry-aware text-to-image model to generate a multiview image for texture. Although these methods achieve impressive texture results, they can still hardly comprehend the overall geometry of input 3D mesh. + +# 2.2. Optimization-based 3D Generation + +Before the emergence of large-scale Text-to-Image generative model, earlier methods [10, 19, 30, 31, 33] propose to optimize texture map of 3D object via natural language supervised visual model, CLIP [39]. Subsequently, score distillation sampling (SDS) was adopted by DreamFusion [38] and Magic3D [25]. The key idea is to optimize 3D representations such as NeRF [32] or InstantNGP [34] with the gradient guidance from 2D diffusion priors [38, 51]. To generate PBR material, TextureDreamer [53] optimizes spatially-varying bidirectional reflectance distribution (BRDF) field through personalized geometric-aware score distillation. Fantasia3D [9] uses a single predefined environmental. However, the generated images from diffusion models may not be consistent with the given environment light. Another concurrent work Paint-it [54] is similar to Fantasia3D [9] but with Deep Convolutional Physically-Based Rendering (DC-PBR) parameterization. DreamMat [60] proposes a novel geometry and light-aware diffusion model, which is trained to generate images that are consistent with the given environment light. FlashTex[14] uses a light-conditioned diffusion model within a two-stage pipeline, combining reconstruction and SDS optimization to enhance texture quality and achieve better light disentanglement. However, these methods struggle with the Janus problem due to the semantically ambiguous. And the time consumption is relatively too long to use in practice. + +# 2.3. Generating Texture from 3D Data + +The most straightforward way to synthesize texture map for 3D mesh is to train generative model directly from 3D data with texture ground truth [7, 11-13]. Early methods such as Texture Fields [35] learn implicit texture fields to assign a color to each pixel on the surface of the 3D shape. Texturify [45] devices face convolution operation on mesh surface to predict texture on each face. It employs differentiable rendering with an adversarial loss to ensure that generated textures produce realistic imagery. Recently, some diffusion-based texture synthesis methods, such as PointUV [55] and TexOct [26] train a denoising network on colors of point clouds which are further mapped to 2D UV map. Although these methods achieve better 3D global consistency with input mesh, they are only trained on several categories of small datasets [7]. What's more, discrete supervision from 3D point clouds leads to suboptimal results compared with continuous signals such as 2D images. + +# 3. Method + +In this section, we provide a detailed explanation of our proposed method, TexGaussian. The overall pipeline of our method is shown in Fig. 2. Existing texture synthesis approaches mainly rely on pre-trained 2D diffusion models, which struggle to fully understand the overall 3D structure. This often leads to misalignment between the generated texture map and 3D semantics. Our goal is to synthesize the high-quality PBR material for a given mesh directly in 3D space. + +# 3.1. Overview + +To enable effective learning in 3D space, we use octree, a sparse voxel structure, to organize and store 3D information without compromising representation quality due to the inherent sparseness of 3D objects in 3D space. Thus, we sample a large number of points on the surface of the given mesh to construct the corresponding octree. The key components of our method are the octant-aligned Gaussian Splatting and the octree-based 3D U-Net. Specifically, we place a 3D Gaussian at the center of each octree's finest leaf node. The octree-based 3D-Unet is trained to predict the parameters of each 3D Gaussian. Under this circumstance, the generated octant-aligned 3D Gaussians are naturally on the mesh surface. We render them from multiple viewpoints using 3D Gaussian Splatting and train the 3D U-Net by minimizing the difference in 2D raster images and 3D Gaussian parameters. During inference, the rendered multi-view images are baked into the UV space of input mesh using differentiable mesh rendering, producing the final texture and material map. The details of 3D Gaussian Splatting (3DGS) are provided in the supplementary material. + +![](images/142cc3034aabedd25a81c3acd4f601cdbb006a856c3cc2e969e4cc15b6310317.jpg) +Figure 2. An overview of our PBR material generation framework. (a) We propose octant-aligned 3D Gaussian Splitting, which positions a 3D Gaussian at the center of each finest leaf node of the constructed octree. Additional channels are added at the end of the Gaussian parameters to model PBR material. (b) We use the 3D U-Net built upon octree-based convolutional networks to predict the Gaussian parameters. Our octree-based 3D U-Net is trained by minimizing the difference on 2D raster images and 3D Gaussian parameters. (c) We bake the multi-view rendered images to the UV space of the input 3D model to realize physically based rendering under new illumination environments. + +# 3.2. Octant-aligned 3D Gaussian Splitting + +For a given mesh, we first sample $N = 100,000$ 3D points on its surface. The corresponding octree is then built by adaptively subdividing the voxels containing those points until the maximum depth is reached. As a result, all of the finest leaf nodes of octree lie along the boundary of the 3D object, which is consistent with the characteristic of the optimized 3D Gaussian. Therefore, we align a 3D Gaussian at the central position of each finest leaf node to effectively model the appearance without compromising the splatting quality. It is worth noting that we do not adjust the position of 3D Gaussians because they are already on the surface of 3D shape and adding additional offset relative to the octant center would not improve the rendering quality through our early experiments. To model PBR material that includes roughness and metallic information, we follow [21, 44] to append two additional channels at the end of 3D Gaussian parameters which are responsible for roughness and metallic map rendering, respectively. + +For each 3D object in our dataset, we render multi-view images of albedo, roughness, and metallic maps for the training of 3D Gaussian. The multi-view rendering results of the albedo map are view-independent. So we just use three RGB channels to take the place of the original spherical harmonics. As noted, we exclude the position from 3D Gaussian parameters in our model. Thus, our 3D Gaussian parameters consist of 13 channels in total: one for opacity, three for scale, four for rotation, three for albedo, one for roughness, and one for metallic. To stabilize training, we choose to employ different activation functions compared to the original Gaussian Splatting [22]. Specifically, We multiply the softplus-activated scales $\mathbf{s}_i$ with 0.01, ensuring that the initial 3D Gaussians conform to the object's counter of object at the beginning of training rather than expanding outward. + +We pre-fitting the parameters of 3D Gaussians on the + +constructed octree for each 3D object in our dataset via the original loss function in [22] on multi-view albedo images, roughness maps, and metallic maps: + +$$ +L = (1 - \lambda) L _ {1} + \lambda L _ {\mathrm {D - S S I M}}, \tag {1} +$$ + +where $\lambda = 0.2$ + +# 3.3. Octree-based 3D U-Net Training + +To handle the encoding of our octant-aligned 3D Gaussian representation, we use the 3D U-Net built upon octree-based convolutional neural networks [49] to predict the Gaussian parameters. Inspired by LGM [47], the output feature of the 3D U-Net on each octant is treated as the 3D Gaussian parameters, which contain 13 channels, as discussed in the last subsection. + +To effectively train our 3D U-Net, we adopt the regressive loss objective, which could further facilitate the generation process. The input to our octree-based 3D U-Net is the geometry feature on each octree's finest leaf node, such as normal and local displacement. For text-conditioned PBR material synthesis, the text feature is extracted by pretrained CLIP model [39] and is fed to U-Net via the octree-based multi-head cross attention mechanism similar to [48]. The predicted 3D Gaussians are rasterized from multiple views via Gaussian Splatting [22]. At each training step, we rasterize the albedo images, alpha images, roughness and metallic maps from randomly selected eight views. Following [47], we apply the mean square error (MSE) loss and the VGG-based LPIPS loss [59] to the albedo image, roughness map, and metallic map: + +$$ +L _ {\text {a l b e d o}} = L _ {\text {M S E}} \left(I _ {\text {a l b e d o}}, I _ {\text {a l b e d o}} ^ {\mathrm {G T}}\right) + L _ {\text {L P I P S}} \left(I _ {\text {a l b e d o}}, I _ {\text {a l b e d o}} ^ {\mathrm {G T}}\right), \tag {2} +$$ + +$$ +L _ {\mathrm {R}} = L _ {\mathrm {M S E}} \left(I _ {\mathrm {R}}, I _ {\mathrm {R}} ^ {\mathrm {G T}}\right) + L _ {\mathrm {L P I P S}} \left(I _ {\mathrm {R}}, I _ {\mathrm {R}} ^ {\mathrm {G T}}\right), \tag {3} +$$ + +$$ +L _ {\mathrm {M}} = L _ {\mathrm {M S E}} \left(I _ {\mathrm {M}}, I _ {\mathrm {M}} ^ {\mathrm {G T}}\right) + L _ {\mathrm {L P I P S}} \left(I _ {\mathrm {M}}, I _ {\mathrm {M}} ^ {\mathrm {G T}}\right), \tag {4} +$$ + +where 'R' and 'M' denote roughness and metallic, respectively. We further apply the MSE loss on the alpha image for faster convergence of the shape: + +$$ +L _ {\alpha} = L _ {\mathrm {M S E}} \left(I _ {\alpha}, I _ {\alpha} ^ {\mathrm {G T}}\right). \tag {5} +$$ + +To accelerate the coverage process, we also apply the 3D MSE loss $L_{\mathrm{3D}}$ , which calculates the difference between predicted parameters of 3D Gaussians and pre-fitting ones. Finally, the complete loss function of our model is defined as the sum of all the above losses: + +$$ +L _ {\text {t o t a l}} = L _ {\text {a l b e d o}} + L _ {\mathrm {R}} + L _ {\mathrm {M}} + L _ {\alpha} + L _ {\mathrm {3 D}}. \tag {6} +$$ + +# 3.4. Texture and Material Baking + +In the inference stage, we also first build the corresponding octree for the input 3D mesh. Then, we use our trained octree-based 3D U-Net to generate 3D Gaussian on every octant and rasterize them from multiple views. The ultimate output of our method should be a global texture map and material map. Thus, we rasterize the input mesh using the differentiable renderer [24] and optimize its albedo, roughness and metallic parameters via the MSE loss between the Nvdiffrast [24] rendering results and 3D Gaussian rendering results. With adequately optimized implementation, this process takes only about several seconds to bake the multiview images to an untextured 3D model. After the optimization, our input mesh paired with its albedo and material map is capable of performing physically based rendering in new illumination environments. + +# 4. Experiments + +# 4.1. Implementation Details + +Dataset We train our TexGaussian model on two publicly-available datasets: ShapeNet [7] and Objverse [13]. For 3D objects in the ShapeNet dataset, since they only contain albedo maps without PBR materials, our model is trained to generate RGB textures only. We train our model on four categories of ShapNet: bench, car, chair and table which is consistent with previous works [26, 55]. We curate a subset of Objverse models encoded with PBR materials and convert those using the specular-glossiness workflow to the metallic-roughness workflow to achieve a consistent PBR representation. This process resulted in a total of 29,200 models. For each 3D object, we render its RGBA albedo image, roughness map, and metallic map from 64 views of $512^2$ for training. + +Network Architecture Our 3D U-Net is built upon octree-based convolutional neural networks [49] which only operates on non-empty octree leaf nodes. The depth of our constructed octree is set to 8 (resolution $256^{2}$ ). Our 3D U-Net consists of 5 down-sampling and up-sampling blocks. + +For the text-conditioned generation, cross-attention layers are only inserted at the last two down-sampling blocks, the middle block, and the first two up-sampling blocks. The input channel of our 3D U-Net is set to 4, where 3 for normal vector and 1 for local displacement while the output channel is set to 13 as mentioned above. The resolution of 3D Gaussian Splatting is set to $512^2$ . The resolution of the baked albedo and material map is set to $1024^2$ . + +Training Details We train our model in a single-category and unconditional manner on ShapeNet, i.e., there are 4 TexGaussian models without text conditions in total. This per-category model is trained on 4 NVIDIA A100 (40G) GPUs for about 2 days and is set up for a fair comparison with other existing methods trained on ShapeNet. For the Objaverse dataset, we train a text-conditioned TexGaussian model for PBR material generation on 24 NVIDIA A100 (40G) GPUs for two weeks. For each batch, we randomly sample 8 camera views to calculate loss functions. We adopt the AdamW [27] optimizer with a learning rate of $4 \times 10^{-4}$ , a weight decay of 0.05, and betas of (0.9,0.95). The learning rate is cosine annealed to 0 during the training and we clip the gradient with a maximum norm of 1.0. + +Evaluation metrics To effectively assess the quality of our generative results, we adopt the metric proposed by [61]. Specifically, each mesh with the generated PBR material is rendered from 20 uniformly distributed views to get multi-view albedo maps, and PBR rendering results in a new illumination environment. These images are used to calculate the FID [36] and KID [2] scores against those from ground truth to evaluate the quality and diversity of generative PBR materials. For the ShapeNet dataset, we only use albedo maps. The final score is averaged across 20 views, and a lower FID and KID score indicates better generation quality and diversity. + +# 4.2. Unconditional RGB Texture Generation + +We conduct unconditional RGB texture instead of PBR material generation on four categories of the ShapeNet dataset. We use the same train and test data split as in [26, 55]. Fig. 3 shows the unconditionally generated RGB texture by our single-category model with high quality, fidelity, and diversity. It can be seen that the texture map synthesized by our method is of great 3D global consistency with the corresponding input geometry. + +Quantitative Comparison We conduct quantitative analysis and comparison on our model and other state-of-the-art methods. Specifically, we compare our model with PointUV Diffusion [55] and TexOct [26] which also train single-category models on ShapeNet. We also compare with some methods using pre-trained 2D diffusion prior for multi-view images synthesis such as TEXTure [55], Text2Tex [8] + +Table 1. Quantitative comparison of the FID and KID $(\times 10^{2})$ score as well as the inference time of TexGaussian and other methods on the ShapeNet dataset [7]. The top part reports the comparison with methods that also train per-category models on ShapeNet. The bottom part reports the comparison with 2D diffusion-based methods, which select 50 samples from the test set of each category. + +
MethodsAverageBenchCarChairTableTime
FID↓KID↓FID↓KID↓FID↓KID↓FID↓KID↓FID↓KID↓
Point-UV (1-Stage) [55]88.435.7864.961.44186.1017.5746.231.9356.422.1939.92s
Point-UV (2-Stage) [55]61.492.6767.481.6189.385.8241.331.5347.751.7249.75s
TexOct [26]59.452.6060.460.9790.106.1137.701.3649.521.9717.44s
Ours49.762.0746.370.4480.205.2229.961.0942.521.5411.02s
TEXTure [41]169.827.94152.535.28236.5818.52119.712.86170.465.11132.21s
Text2Tex [8]156.626.68154.955.79188.6015.06125.343.11157.572.771005.58s
Paint3D [56]135.254.67104.161.68204.2213.06104.642.24127.981.71231.56s
Ours100.971.8886.370.75117.535.2094.890.57105.090.9511.12s
+ +![](images/0c1a3879e9d5f1db1a007de4898b9a3d1061e5720bce488eb6182be471249af5.jpg) +Figure 3. Unconditional RGB texture generative results on ShapeNet. Please zoom in for a better inspection of color details. Paint3D [56], and TexPainter [57] whose input text prompts are set to "a *" and "* is the name of corresponding category. Due to the relatively long time consumption, we only use 50 3D objects selected from the test set to evaluate the methods based on 2D diffusion models. We conduct only a qualitative comparison for TexPainter [57] for its lengthy processing time. Table 1 reports the comparison of FID and KID scores as well as average inference time on a single NVIDIA A100 (40G) GPU. It is worth noting that our model only takes about one second to predict the parameters of each 3D Gaussian and the rest of the time is used for texture baking. From Table 1, we have the following observation. First, our method obtains the best performance in terms of FID and KID. For example, our method outperforms TexOct [26], by an average of 9.69 in FID and 0.53 in KID. These improvements indicate that our method excels at generating high-quality textures. Second, TexGaussian achieves the fastest generative speed in all categories, which is much less time-consuming than other methods. + +Qualitative Comparison Fig. 4 provides some qualitative results on the same input 3D meshes by different methods. Point-UV [55] and TexOct [26] use colored point clouds as supervision to train diffusion models. As a consequence, the generated texture map is relatively blurry due to the discreteness of the point cloud. For other methods that leverage pre-trained 2D diffusion prior, they can hardly comprehend the total geometry. The generated texture maps + +![](images/86ea7666c7b2f0e83a065b8b10996d512d212979bca695d8dcd137b38bfd6843.jpg) +Figure 4. Examples of generated RGB textures obtained by Tex-Gaussian and other state-of-the-art models on the same 3D object. Please zoom in for a better inspection. + +by them are not consistent with the semantics of input 3D mesh, such as the chaotic stripes on the bench and chair in Fig. 4. On the contrary, our method generates smooth and colorful RGB textures on unseen objects, which align well with 3D meshes. + +![](images/517f40a175acc03512af0d34931028fd8e7ab6851c7e83a39793e85414a58263.jpg) +Figure 5. Qualitative comparison with Fantasia3D [9], FlashTex [14], DreamMat [60] and Make-it-Real [16]. We show the rendered image, albedo map, roughness map, and metallic map for each 3D object. + +# 4.3. Text-conditioned PBR Material Generation + +We train the text-conditioned PBR material generation model on our filtered subset of Objaverse [13]. We use the text descriptions of 3D objects from Cap3D [28, 29] to train our model. We choose 29,000 3D objects in our filtered subset for training and the rest 200 for testing. Fig. 1 shows some generative results by proposed TexGaussian on the test set. We can see that our model comprehends the overall geometry feature of input 3D mesh. It is capable of synthesizing high-quality albedo, roughness, and metallic, which are aligned well with 3D semantics such as accents with high metallic on the wooden box and rope with high roughness on the jug. + +Quantitative Comparison We conduct quantitative comparison with some state-of-the-art text-conditioned PBR material synthesis methods: Fantasia3D [9], FlashTex [14], and DreamMat [60] on our test set. Table 2 reports FID and KID scores on both multi-view albedo map and PBR rendering images under the same illumination environment. TexGaussian outperforms all the baseline methods in achieving the best visual quality of the generated appearances. We also report the average inference time of different methods on a single NVIDIA A100 (40G) GPU across our test set. Due to the iterative optimization of score distillation sampling, all other three methods cost at least 20 minutes for the generation. In contrast, our TexGaussian only takes about 20 seconds, in which one second is for predicting the Gaussian parameters and the rest is for baking, which performs $60 \times$ faster than previous methods. + +Qualitative Comparison Fig. 5 visualizes the generated albedo, roughness, and metallic of each compared method from the same text prompt and untextured meshes. We also show the PBR rendering images of the generated materials + +under the same environment light. It can be observed that Fantasia3D [9] and FlashTex [14] generate irregular colors on mesh surfaces, which lead to rendering results with low quality. DreamMat [60] is capable of generating a visually pleasing appearance for test 3D models. However, it can hardly align the generated PBR material well with the 3D semantics, such as the black band on the wooden barrel. What's more, it tends to generate over-saturated colors, which is demonstrated in the wooden chest and green grenade. In addition, all of the compared methods struggle to completely disentangle the light and texture, which results in relatively chaotic results of generated metallic and roughness. We also compare our method with a Multimodal Large Language Model based method, Make-it-Real [16], which accepts 3D objects paired with albedo maps generated by our TexGaussian as input. It can be observed that Make-it-Real struggles to capture the overall 3D structure of input meshes. As a consequence, it tends to predict single-value PBR maps, leading to irregular rendering results, such as the wooden chest with extremely large metallic value. On the contrary, our method is directly trained from 3D original data. It is capable of generating clean and smooth PBR material while fully comprehending the overall 3D structure. We provide more generative results in the supplementary material. + +# 4.4. Ablation Study + +Essentials of Octree We first analyze the effectiveness of using Octree in our pipeline. To do this, we calculate the quantitative quality of 3D Gaussian pre-fitting by our octant-aligned and full voxel-aligned 3D Gaussian Splitting on the subset of Objaverse dataset [13]. The depth of octree in this ablation is set to 8, and the resolution of full voxels is $256^3$ , which is consistent with the finest resolution of octree to guarantee comparison fairness. Table 3 reports the + +Table 2. Quantitative comparison of the FID and KID $(\times 10^{2})$ scores as well as the inference time of TexGaussian and other methods on our test set which consists of 200 3D objects with ground truth PBR materials. + +
MethodsAlbedoPBR renderingTime
FID↓KID↓FID↓KID↓
Fantasia3D [9]213.210.96209.870.4822.7mins
FlashTex [14]185.241.13186.820.4220.3mins
DreamMat [60]152.631.09145.490.1948.2mins
Ours123.720.20129.520.0221.04s
+ +Table 3. Quantitative comparison of methods using different types of 3D Gaussian Fitting. + +![](images/80bf1f8d675abd45b41508bb71945ceb6ca841b48ddaefb3d3f3fc872253f0f8.jpg) +Figure 6. Visualization of different manners of Gaussian fitting. The rendering results demonstrate excellent reconstruction performance of the proposed octant-aligned 3D Gaussian Splatting. + +PSNR, LPIPS [59], and SSIM [50] metrics on the albedo map as well as the average number of 3D Gaussians of octant and full voxel aligned 3D Gaussian Splitting across our Objverse subset. Fig. 6 also presents some visual results of our Gaussian fitting. These results demonstrate that the proposed octant-aligned 3D Gaussian produces a much more reasonably compact and precise representation of diverse and complex 3D assets with much fewer 3D Gaussians compared to the full voxel version. + +Essentials of different losses We verify the importance of different losses we proposed to train our model. We term all the losses calculated on 2D image space as $L_{2\mathrm{D}}$ : + +$$ +L _ {2 \mathrm {D}} = L _ {\text {a l b e d o}} + L _ {\mathrm {R}} + L _ {\mathrm {M}} + L _ {\alpha}. \tag {7} +$$ + +We train two additional TexGaussian models using only the $L_{2\mathrm{D}}$ or $L_{3\mathrm{D}}$ loss on ShapeNet car category due to its large variations and complexity of texture to validate their effectiveness. It is worth noting that in this dataset, $L_{2\mathrm{D}} = L_{\mathrm{albedo}} + L_{\alpha}$ due to the lack of material information. The training curves of MSE loss and LPIPS loss between rendering images and ground-truth ones are shown + +![](images/aa8ab213859c7a239e18cd0ddc54efda5b41131873f7038710356170d0c1847a.jpg) +Figure 7. The training loss curves of our model with different losses. + +![](images/05c688ef0ceed95203f5679d5caf9a44010ad346584880e664c0a05507d8bbf1.jpg) + +![](images/95799d7d4b83d29b0b0e85f6ef03b33c93c8a2502927329eac9a1bcd5dbdb8ce.jpg) +Figure 8. Results of our model trained with and without $L_{2D}$ . + +in Fig. 7, which demonstrate the effects of $L_{2D}$ and $L_{3D}$ . From the loss curves, we can conclude that $L_{3\mathrm{D}}$ facilitates the process of convergence and $L_{2D}$ enhances the quality of synthesized images, as evidenced by the large margin of improvement in the LPIPS loss when introducing $L_{2D}$ . We also provide some qualitative results to verify the effectiveness of $L_{2D}$ on the test set of car category in Fig. 8. Only using $L_{3D}$ results in a relatively blurry texture map due to the discreteness of 3D Gaussian which is similar to the characteristic of 3D point clouds analyzed above. + +# 5. Conclusion + +In this paper, we proposed TexGaussian, an octree-based 3D Gaussian Splatting model for high-quality PBR material generation on untextured meshes. We aligned each 3D Gaussian on the octant of the corresponding octree built from the input untextured object and extended the parameters of 3D Gaussian with additional channels to represent the roughness and metallic map. We trained our model with regression objectives, achieving faster inference speed compared to previous texture synthesis methods. Experimental results demonstrated that our method is capable of generating high-quality PBR materials that are readily usable in modern graphics engines for photo-realistic rendering, offering enhanced realism for a variety of applications. + +Limitations The generalization of TexGaussian is still hindered by the scale of the training set. Thus it struggles to generate various textures for some extremely complex 3D objects beyond our training data. We are looking forward to training our TexGaussian model with more parameters and more data on a larger-scale GPU cluster in the future. + +# References + +[1] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 3 +[2] Mikołaj Binkowski, Dougal J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations, 2018. 5 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. +[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2 +[6] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and KangXue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3 +[7] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 3, 5, 6 +[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2, 3, 5, 6 +[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 3, 7, 8 +[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 3 +[11] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 3 + +[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. arXiv preprint arXiv:2307.05663, 2023. +[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3, 5, 7 +[14] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. Flashtex: Fast reconfigurable mesh texturing with lightcontrolnet. In European Conference on Computer Vision (ECCV), 2024. 3, 7, 8 +[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[16] Ye Fang, Zeyi Sun, Tong Wu, Jiaqi Wang, Ziwei Liu, Gordon Wetzstein, and Dahua Lin. Make-it-real: Unleashing large multimodal model for painting 3d objects with realistic materials, 2024. 7 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 2 +[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2 +[19] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 3 +[20] Dong Huo, Zixin Guo, Xinxin Zuo, Zhihao Shi, Juwei Lu, Peng Dai, Songcen Xu, Li Cheng, and Yee-Hong Yang. Tex-gen: Text-guided 3d texture generation with multi-view sampling and resampling. ECCV, 2024. 2, 3 +[21] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian-shader: 3d gaussian splattering with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 4 +[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 4 +[23] Matthias Labschütz, Katharina Krosl, Mariebeth Aquino, Florian Grashäftl, and Stephanie Kohl. Content creation for a 3d game with maya and unity 3d. Institute of Computer Graphics and Algorithms, Vienna University of Technology, 6(124):2, 2011. 1 +[24] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5 + +[25] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[26] Jialun Liu, Chenming Wu, Xinqi Liu, Xing Liu, Jinbo Wu, Haotian Peng, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Texoct: Generating textures of 3d models with octree-based diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4284-4293, 2024. 2, 3, 5, 6 +[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[28] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 7 +[29] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. arXiv preprint arXiv:2404.07984, 2024. 7 +[30] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 3 +[31] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3 +[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3 +[33] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3 +[34] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3 +[35] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings IEEE International Conf. on Computer Vision (ICCV), 2019. 3 +[36] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 5 +[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2 +[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3 + +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4 +[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[41] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 2, 3, 6 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2 +[43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[44] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 4 +[45] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 72-88. Springer, 2022. 3 +[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[47] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 4 +[48] Peng-Shuai Wang. Octformer: Octtree-based transformers for 3D point clouds. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 4 +[49] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-CNN: Octree-based convolutional neural networks for 3D shape analysis. ACM Transactions on Graphics (SIGGRAPH), 36(4), 2017. 2, 4, 5 +[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 8 +[51] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla + +tion. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3 +[52] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Toon-crafter: Generative cartoon interpolation. arXiv preprint arXiv:2405.17933, 2024. 2 +[53] Yu-Ying Yeh, Jia-Bin Huang, Changil Kim, Lei Xiao, Thu Nguyen-Phuoc, Numair Khan, Cheng Zhang, Manmohan Chandraker, Carl S Marshall, Zhao Dong, et al. Texture-dreamer: Image-guided texture synthesis through geometry-aware diffusion. arXiv preprint arXiv:2401.09416, 2024. 3 +[54] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[55] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206-4216, 2023. 2, 3, 5, 6 +[56] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 2, 3, 6 +[57] Hongkun Zhang, Zherong Pan, Congyi Zhang, Lifeng Zhu, and Xifeng Gao. Texpainter: Generative mesh texturing with multi-view consistency. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6 +[58] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 3 +[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4, 8 +[60] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, and Xiaogang Jin. Dreammat: High-quality pbr material generation with geometry- and light-aware diffusion models. ACM Trans. Graph., 43(4), 2024. 2, 3, 7, 8 +[61] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 5 \ No newline at end of file diff --git a/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/images.zip b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..acbaa8f84981796c33b6c8cd6f08d08602c4120e --- /dev/null +++ b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0a8cdca58f9dd7244c9fc395f7a768b816fedc0a688f1378afb9d911915c483 +size 561190 diff --git a/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/layout.json b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..24d0dd0e58249dcaa16821217971acbbf812138f --- /dev/null +++ b/2025/TexGaussian_ Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting/layout.json @@ -0,0 +1,6920 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 483, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 483, + 140 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 483, + 140 + ], + "type": "text", + "content": "TexGaussian: Generating High-quality PBR Material via Octree-based 3D Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "spans": [ + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "content": "Bojun Xiong" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "content": ", Jialun Liu" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "content": ", Jiakui Hu" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "inline_equation", + "content": "^{3†}" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "content": ", Chenming Wu" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "content": ", Jinbo Wu" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "content": ", Xing Liu" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "content": ", Chen Zhao" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "content": ", Errui Ding" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "text", + "content": ", Zhouhui Lian" + }, + { + "bbox": [ + 109, + 160, + 501, + 190 + ], + "type": "inline_equation", + "content": "^{1‡}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 480, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 480, + 222 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 194, + 480, + 222 + ], + "type": "text", + "content": "Wangxuan Institute of Computer Technology, Peking University, China " + }, + { + "bbox": [ + 130, + 194, + 480, + 222 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 194, + 480, + 222 + ], + "type": "text", + "content": "Baidu VIS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 162, + 223, + 448, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 223, + 448, + 238 + ], + "spans": [ + { + "bbox": [ + 162, + 223, + 448, + 238 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 223, + 448, + 238 + ], + "type": "text", + "content": "Institute of Medical Technology, Peking University, China" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 252, + 553, + 397 + ], + "blocks": [ + { + "bbox": [ + 57, + 252, + 553, + 397 + ], + "lines": [ + { + "bbox": [ + 57, + 252, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 57, + 252, + 553, + 397 + ], + "type": "image", + "image_path": "736babe93e777639cd2055742041e084ca3b71ed3d21f9ca7a61e16f1d117ab9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 406, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 54, + 406, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 54, + 406, + 555, + 440 + ], + "type": "text", + "content": "Figure 1. Our proposed TexGaussian is capable of generating high-quality materials for given input 3D meshes based on the corresponding textual descriptions. The generated results are naturally compatible with modern graphical engines for photo-realistic rendering under different environment maps." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 451, + 200, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 451, + 200, + 464 + ], + "spans": [ + { + "bbox": [ + 151, + 451, + 200, + 464 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 477, + 297, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 477, + 297, + 646 + ], + "spans": [ + { + "bbox": [ + 54, + 477, + 297, + 646 + ], + "type": "text", + "content": "Physically Based Rendering (PBR) materials play a crucial role in modern graphics, enabling photorealistic rendering across diverse environment maps. Developing an effective and efficient algorithm that is capable of automatically generating high-quality PBR materials rather than RGB texture for 3D meshes can significantly streamline the 3D content creation. Most existing methods leverage pre-trained 2D diffusion models for multi-view image synthesis, which often leads to severe inconsistency between the generated textures and input 3D meshes. This paper presents TexGaussian, a novel method that uses octant-aligned 3D Gaussian Splitting for rapid PBR material generation. Specifically, we place each 3D Gaussian on the finest leaf node of the octree built from the input 3D mesh to render the multi-view" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 453, + 556, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 585 + ], + "type": "text", + "content": "images not only for the albedo map but also for roughness and metallic. Moreover, our model is trained in a regression manner instead of diffusion denoising, capable of generating the PBR material for a 3D mesh in a single feed-forward process. Extensive experiments on publicly available benchmarks demonstrate that our method synthesizes more visually pleasing PBR materials and runs faster than previous methods in both unconditional and text-conditional scenarios, exhibiting better consistency with the given geometry. Our code and trained models are available at https://3d-aigc.github.io/TexasGaussian." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Traditional 3D asset creation relies heavily on the expertise and extensive effort of professional designers [23], posing a significant barrier for casual users interested in creating 3D models independently. In the 3D design process, geometry creation typically represents only a small portion of the overall time, while the majority is dedicated to developing textures and appearances, which are far more time-consuming. Achieving a delicate appearance for a 3D" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "\\*Denotes equal contribution. " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " This work was partly done when Bojun and Jiakui interned in Baidu. " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " Corresponding author. E-mail: lianzhouhui@pku.edu.cn This work was supported by the National Natural Science Foundation of China (Grant No.: 62372015), Center For Chinese Font Design and Research, and Key Laboratory of Intelligent Press Media Technology." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "551" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "model often demands substantial time and effort from experienced designers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 295, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 295, + 205 + ], + "type": "text", + "content": "Recently, Artificial Intelligence Generated Content (AIGC) based on deep generative models, especially diffusion models [17, 46] have been widely used to facilitate the process of artistic creation, catalyzing advancements in image generation [15, 37, 40, 42, 43] and video generation [3-5, 18, 52]. As a result, exploring effective ways to leverage deep generative models to streamline the creation of detailed appearances for 3D models has become a popular direction in the graphics and vision communities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 207, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 434 + ], + "type": "text", + "content": "Recent advancements in 3D texture generation attempt to use depth-conditional pre-trained 2D diffusion model [58] to synthesize RGB images based on the depth maps rendered from multiple views, such as TEXture [41] and text2tex [8]. Subsequent works [6, 20, 57] further improve multi-view consistency via blending the multi-view images to a single and consistent texture map in every denoising step. However, these methods struggle to have a global picture of 3D geometries due to the use of 2D diffusion models, leading to inconsistencies between the texture map and the semantics of the input 3D meshes. Moreover, the generated assets suffer from illumination-baked textures, which can significantly degrade the quality of the final rendering when placed in novel lighting conditions [56]. While DreamMat [60] supplements geometry and light control to achieve material decomposition through score distillation sampling [38], it still struggles to fully capture the global geometry features. This limitation often results in the multiface Janus problem and leads to over-saturated colors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 581 + ], + "type": "text", + "content": "On the other hand, training a 3D neural network directly on 3D data, such as Point-UV Diffusion [55] and TexOct [26], is an effective way for 3D global consistency. Meanwhile, this avoids multi-view sampling and score distillation sampling, which accelerates the process of texture synthesis. However, relying on colored point clouds for 3D representation and supervision often results in blurred outputs, primarily due to the sparse and non-compact nature of point clouds in 3D space. Due to the limitations of the adopted 3D representations and the lack of training data on PBR materials, these approaches are incapable of generating high-fidelity PBR materials for 3D models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "To address the aforementioned challenges, this paper presents TexGaussian, a fast and high-fidelity PBR material generation model directly in 3D space that maintains 3D global consistency. Different from previous approaches that primarily rely on diffusion models, our method works in a regression manner to regress the PBR material from the input mesh for faster generation speed. To enable effective learning in 3D space, we propose to use octree, a specialized sparse voxel structure that efficiently organizes and preserves 3D information, which can be built from 3D point clouds sampled from the surface of the object. How" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 407 + ], + "type": "text", + "content": "ever, directly regressing the color of 3D point clouds on octree often results in blurry textures as mentioned in [26]. To tackle the challenges of the incompact and discrete nature presented by points, we use 3D Gaussian Splatting (3DGS) [22], a robust representation that bridges the gap between 3D space and 2D raster images, allowing us to fully utilize rich 2D image information to alleviate blurring results. Specifically, for each input mesh, we sample the dense 3D point clouds on its surface to build the corresponding octree. On each octant (i.e., the finest leaf node of octree), we place a 3D Gaussian [22] at its central position. Then, we use the octree-based 3D U-Net [49] to predict the parameters of each 3D Gaussian on octants. Apart from RGB colors, we extend each 3D Gaussian with additional parameters to represent the roughness and metallicity of 3D objects. Multi-view images, including albedo, roughness, and metallic maps, can be rasterized from all these 3D Gaussians via 3DGS. The 3D U-Net is supervised by the difference between the predicted multi-view images and their corresponding ground truth. Notably, the 3D U-Net is trained to directly regress the multi-view images based on the geometry feature of the input 3D model, which further facilitates the process of 3D Gaussian prediction compared to diffusion manner. We train our TexGaussian model on a subset of Objaverse [13] with high-quality PBR materials, enabling fast PBR material generation with a single feedforward pass. In summary, the contributions of our paper are threefold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 409, + 553, + 588 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 409, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 409, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 409, + 553, + 468 + ], + "type": "text", + "content": "- We propose an octant-aligned 3D Gaussian Splitting method for high-quality PBR material synthesis on un-textured input 3D mesh, which fully utilizes the supervision from 2D images, avoiding blurry results caused by the discreteness of 3D point clouds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 469, + 553, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 469, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 314, + 469, + 553, + 505 + ], + "type": "text", + "content": "- We adopt a regression manner to train our 3D U-Net model instead of diffusion denoising, achieving faster generative speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 505, + 553, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 505, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 314, + 505, + 553, + 588 + ], + "type": "text", + "content": "- We propose TexGaussian, a novel PBR material generation method based on the above two techniques. To our knowledge, our work is the first to generate PBR materials directly in 3D space. Qualitative and quantitative experiments have been conducted to verify the superiority of the quality and efficiency of our method over other existing approaches." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 596, + 400, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 596, + 400, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 596, + 400, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "type": "text", + "content": "In this section, we mainly summarize current texture synthesis methods, which can be roughly divided into three categories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 659, + 471, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 471, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 471, + 673 + ], + "type": "text", + "content": "2.1. Multi-view Images Synthesis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "Many previous works have tried to leverage the powerful T2I model to assist texture generation for 3D shapes. Specifically, they render the depth map of in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "552" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 358 + ], + "type": "text", + "content": "put 3D mesh from multiple views and use depth conditional T2I models [58] to synthesize RGB images and perform text-conditioned texture synthesis. TEXTure [41] and Text2Tex [8] iteratively paint a mesh from different views. However, images synthesized in early view could produce errors that are not reconcilable with the geometry that is observed in later views. Many subsequent works try to alleviate multi-view inconsistency via different alignment modules. TexFusion [6] proposes a sequential interlaced multi-view sampler that interleaves texture assembling with denoising steps in different camera views. Similarly, TexGen [20] directly enforces view consistent sampling in RGB texture space and develops a noise resampling strategy to retain rich texture details. TexPainter [57] blends images from different views into a common color-space texture image by weighted averaging to guarantee multi-view consistency. To remove light influence from 2D diffusion models, Paint3D [56] contribute separate UV Inpainting and UVHD diffusion models specialized in shape-aware refinement. More recently, Meta 3D TextureGen [1] utilizes geometry-aware text-to-image model to generate a multiview image for texture. Although these methods achieve impressive texture results, they can still hardly comprehend the overall geometry of input 3D mesh." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 372, + 241, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 372, + 241, + 383 + ], + "spans": [ + { + "bbox": [ + 58, + 372, + 241, + 383 + ], + "type": "text", + "content": "2.2. Optimization-based 3D Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 392, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 392, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 392, + 294, + 712 + ], + "type": "text", + "content": "Before the emergence of large-scale Text-to-Image generative model, earlier methods [10, 19, 30, 31, 33] propose to optimize texture map of 3D object via natural language supervised visual model, CLIP [39]. Subsequently, score distillation sampling (SDS) was adopted by DreamFusion [38] and Magic3D [25]. The key idea is to optimize 3D representations such as NeRF [32] or InstantNGP [34] with the gradient guidance from 2D diffusion priors [38, 51]. To generate PBR material, TextureDreamer [53] optimizes spatially-varying bidirectional reflectance distribution (BRDF) field through personalized geometric-aware score distillation. Fantasia3D [9] uses a single predefined environmental. However, the generated images from diffusion models may not be consistent with the given environment light. Another concurrent work Paint-it [54] is similar to Fantasia3D [9] but with Deep Convolutional Physically-Based Rendering (DC-PBR) parameterization. DreamMat [60] proposes a novel geometry and light-aware diffusion model, which is trained to generate images that are consistent with the given environment light. FlashTex[14] uses a light-conditioned diffusion model within a two-stage pipeline, combining reconstruction and SDS optimization to enhance texture quality and achieve better light disentanglement. However, these methods struggle with the Janus problem due to the semantically ambiguous. And the time consumption is relatively too long to use in practice." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 72, + 495, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 495, + 84 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 495, + 84 + ], + "type": "text", + "content": "2.3. Generating Texture from 3D Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 91, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 553, + 294 + ], + "type": "text", + "content": "The most straightforward way to synthesize texture map for 3D mesh is to train generative model directly from 3D data with texture ground truth [7, 11-13]. Early methods such as Texture Fields [35] learn implicit texture fields to assign a color to each pixel on the surface of the 3D shape. Texturify [45] devices face convolution operation on mesh surface to predict texture on each face. It employs differentiable rendering with an adversarial loss to ensure that generated textures produce realistic imagery. Recently, some diffusion-based texture synthesis methods, such as PointUV [55] and TexOct [26] train a denoising network on colors of point clouds which are further mapped to 2D UV map. Although these methods achieve better 3D global consistency with input mesh, they are only trained on several categories of small datasets [7]. What's more, discrete supervision from 3D point clouds leads to suboptimal results compared with continuous signals such as 2D images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 308, + 368, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 308, + 368, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 308, + 368, + 319 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 331, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 331, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 553, + 437 + ], + "type": "text", + "content": "In this section, we provide a detailed explanation of our proposed method, TexGaussian. The overall pipeline of our method is shown in Fig. 2. Existing texture synthesis approaches mainly rely on pre-trained 2D diffusion models, which struggle to fully understand the overall 3D structure. This often leads to misalignment between the generated texture map and 3D semantics. Our goal is to synthesize the high-quality PBR material for a given mesh directly in 3D space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 455, + 380, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 380, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 380, + 465 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 475, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 553, + 712 + ], + "type": "text", + "content": "To enable effective learning in 3D space, we use octree, a sparse voxel structure, to organize and store 3D information without compromising representation quality due to the inherent sparseness of 3D objects in 3D space. Thus, we sample a large number of points on the surface of the given mesh to construct the corresponding octree. The key components of our method are the octant-aligned Gaussian Splatting and the octree-based 3D U-Net. Specifically, we place a 3D Gaussian at the center of each octree's finest leaf node. The octree-based 3D-Unet is trained to predict the parameters of each 3D Gaussian. Under this circumstance, the generated octant-aligned 3D Gaussians are naturally on the mesh surface. We render them from multiple viewpoints using 3D Gaussian Splatting and train the 3D U-Net by minimizing the difference in 2D raster images and 3D Gaussian parameters. During inference, the rendered multi-view images are baked into the UV space of input mesh using differentiable mesh rendering, producing the final texture and material map. The details of 3D Gaussian Splatting (3DGS) are provided in the supplementary material." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "text", + "content": "553" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 188 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 188 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 188 + ], + "type": "image", + "image_path": "142cc3034aabedd25a81c3acd4f601cdbb006a856c3cc2e969e4cc15b6310317.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 196, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 54, + 196, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 54, + 196, + 555, + 262 + ], + "type": "text", + "content": "Figure 2. An overview of our PBR material generation framework. (a) We propose octant-aligned 3D Gaussian Splitting, which positions a 3D Gaussian at the center of each finest leaf node of the constructed octree. Additional channels are added at the end of the Gaussian parameters to model PBR material. (b) We use the 3D U-Net built upon octree-based convolutional networks to predict the Gaussian parameters. Our octree-based 3D U-Net is trained by minimizing the difference on 2D raster images and 3D Gaussian parameters. (c) We bake the multi-view rendered images to the UV space of the input 3D model to realize physically based rendering under new illumination environments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 256, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 256, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 256, + 290 + ], + "type": "text", + "content": "3.2. Octant-aligned 3D Gaussian Splitting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 294, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 510 + ], + "type": "text", + "content": "For a given mesh, we first sample " + }, + { + "bbox": [ + 56, + 294, + 296, + 510 + ], + "type": "inline_equation", + "content": "N = 100,000" + }, + { + "bbox": [ + 56, + 294, + 296, + 510 + ], + "type": "text", + "content": " 3D points on its surface. The corresponding octree is then built by adaptively subdividing the voxels containing those points until the maximum depth is reached. As a result, all of the finest leaf nodes of octree lie along the boundary of the 3D object, which is consistent with the characteristic of the optimized 3D Gaussian. Therefore, we align a 3D Gaussian at the central position of each finest leaf node to effectively model the appearance without compromising the splatting quality. It is worth noting that we do not adjust the position of 3D Gaussians because they are already on the surface of 3D shape and adding additional offset relative to the octant center would not improve the rendering quality through our early experiments. To model PBR material that includes roughness and metallic information, we follow [21, 44] to append two additional channels at the end of 3D Gaussian parameters which are responsible for roughness and metallic map rendering, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 511, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 700 + ], + "type": "text", + "content": "For each 3D object in our dataset, we render multi-view images of albedo, roughness, and metallic maps for the training of 3D Gaussian. The multi-view rendering results of the albedo map are view-independent. So we just use three RGB channels to take the place of the original spherical harmonics. As noted, we exclude the position from 3D Gaussian parameters in our model. Thus, our 3D Gaussian parameters consist of 13 channels in total: one for opacity, three for scale, four for rotation, three for albedo, one for roughness, and one for metallic. To stabilize training, we choose to employ different activation functions compared to the original Gaussian Splatting [22]. Specifically, We multiply the softplus-activated scales " + }, + { + "bbox": [ + 55, + 511, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 55, + 511, + 296, + 700 + ], + "type": "text", + "content": " with 0.01, ensuring that the initial 3D Gaussians conform to the object's counter of object at the beginning of training rather than expanding outward." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "We pre-fitting the parameters of 3D Gaussians on the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 277, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 314 + ], + "type": "text", + "content": "constructed octree for each 3D object in our dataset via the original loss function in [22] on multi-view albedo images, roughness maps, and metallic maps:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 375, + 326, + 555, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 326, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 375, + 326, + 555, + 339 + ], + "type": "interline_equation", + "content": "L = (1 - \\lambda) L _ {1} + \\lambda L _ {\\mathrm {D - S S I M}}, \\tag {1}", + "image_path": "4b7de0b7c0e8d5194e9b4c60e29cd9f21c8066cc560864c0b2a676c88524f994.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 345, + 378, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 378, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 378, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 345, + 378, + 357 + ], + "type": "inline_equation", + "content": "\\lambda = 0.2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 365, + 489, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 489, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 489, + 378 + ], + "type": "text", + "content": "3.3. Octree-based 3D U-Net Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 383, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 554, + 466 + ], + "type": "text", + "content": "To handle the encoding of our octant-aligned 3D Gaussian representation, we use the 3D U-Net built upon octree-based convolutional neural networks [49] to predict the Gaussian parameters. Inspired by LGM [47], the output feature of the 3D U-Net on each octant is treated as the 3D Gaussian parameters, which contain 13 channels, as discussed in the last subsection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 467, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 647 + ], + "type": "text", + "content": "To effectively train our 3D U-Net, we adopt the regressive loss objective, which could further facilitate the generation process. The input to our octree-based 3D U-Net is the geometry feature on each octree's finest leaf node, such as normal and local displacement. For text-conditioned PBR material synthesis, the text feature is extracted by pretrained CLIP model [39] and is fed to U-Net via the octree-based multi-head cross attention mechanism similar to [48]. The predicted 3D Gaussians are rasterized from multiple views via Gaussian Splatting [22]. At each training step, we rasterize the albedo images, alpha images, roughness and metallic maps from randomly selected eight views. Following [47], we apply the mean square error (MSE) loss and the VGG-based LPIPS loss [59] to the albedo image, roughness map, and metallic map:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 654, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 654, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 322, + 654, + 553, + 681 + ], + "type": "interline_equation", + "content": "L _ {\\text {a l b e d o}} = L _ {\\text {M S E}} \\left(I _ {\\text {a l b e d o}}, I _ {\\text {a l b e d o}} ^ {\\mathrm {G T}}\\right) + L _ {\\text {L P I P S}} \\left(I _ {\\text {a l b e d o}}, I _ {\\text {a l b e d o}} ^ {\\mathrm {G T}}\\right), \\tag {2}", + "image_path": "f460a7628361440dfabcb41a211e7ec22f00efbb2fa456b32e75ddc000c2c7e7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 352, + 681, + 553, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 681, + 553, + 695 + ], + "spans": [ + { + "bbox": [ + 352, + 681, + 553, + 695 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {R}} = L _ {\\mathrm {M S E}} \\left(I _ {\\mathrm {R}}, I _ {\\mathrm {R}} ^ {\\mathrm {G T}}\\right) + L _ {\\mathrm {L P I P S}} \\left(I _ {\\mathrm {R}}, I _ {\\mathrm {R}} ^ {\\mathrm {G T}}\\right), \\tag {3}", + "image_path": "c15f0eb23f0b6cb95265a174d032205f7ad65dae298ff19096b69c1af0c78b00.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 350, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 350, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {M}} = L _ {\\mathrm {M S E}} \\left(I _ {\\mathrm {M}}, I _ {\\mathrm {M}} ^ {\\mathrm {G T}}\\right) + L _ {\\mathrm {L P I P S}} \\left(I _ {\\mathrm {M}}, I _ {\\mathrm {M}} ^ {\\mathrm {G T}}\\right), \\tag {4}", + "image_path": "2ffd98a015f271fa55d688324e94d71c83ae612ceb288a9ab11c639d4cc26dc8.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "554" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "where 'R' and 'M' denote roughness and metallic, respectively. We further apply the MSE loss on the alpha image for faster convergence of the shape:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 115, + 295, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 295, + 130 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 295, + 130 + ], + "type": "interline_equation", + "content": "L _ {\\alpha} = L _ {\\mathrm {M S E}} \\left(I _ {\\alpha}, I _ {\\alpha} ^ {\\mathrm {G T}}\\right). \\tag {5}", + "image_path": "6e2e5bf87ca5456388c4aff88e15cb85eb6a84e1efd2ee5cead6c9adb210ae38.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 296, + 198 + ], + "type": "text", + "content": "To accelerate the coverage process, we also apply the 3D MSE loss " + }, + { + "bbox": [ + 55, + 138, + 296, + 198 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{3D}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 198 + ], + "type": "text", + "content": ", which calculates the difference between predicted parameters of 3D Gaussians and pre-fitting ones. Finally, the complete loss function of our model is defined as the sum of all the above losses:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 207, + 295, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 207, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 91, + 207, + 295, + 220 + ], + "type": "interline_equation", + "content": "L _ {\\text {t o t a l}} = L _ {\\text {a l b e d o}} + L _ {\\mathrm {R}} + L _ {\\mathrm {M}} + L _ {\\alpha} + L _ {\\mathrm {3 D}}. \\tag {6}", + "image_path": "46ef9dcd67abd029ac7cc1e82f66fbbb8ebaacb9eb6124fff92a61b95b2502de.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 228, + 216, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 216, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 216, + 240 + ], + "type": "text", + "content": "3.4. Texture and Material Baking" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 424 + ], + "type": "text", + "content": "In the inference stage, we also first build the corresponding octree for the input 3D mesh. Then, we use our trained octree-based 3D U-Net to generate 3D Gaussian on every octant and rasterize them from multiple views. The ultimate output of our method should be a global texture map and material map. Thus, we rasterize the input mesh using the differentiable renderer [24] and optimize its albedo, roughness and metallic parameters via the MSE loss between the Nvdiffrast [24] rendering results and 3D Gaussian rendering results. With adequately optimized implementation, this process takes only about several seconds to bake the multiview images to an untextured 3D model. After the optimization, our input mesh paired with its albedo and material map is capable of performing physically based rendering in new illumination environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 454, + 188, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 188, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 188, + 467 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 472, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 640 + ], + "type": "text", + "content": "Dataset We train our TexGaussian model on two publicly-available datasets: ShapeNet [7] and Objverse [13]. For 3D objects in the ShapeNet dataset, since they only contain albedo maps without PBR materials, our model is trained to generate RGB textures only. We train our model on four categories of ShapNet: bench, car, chair and table which is consistent with previous works [26, 55]. We curate a subset of Objverse models encoded with PBR materials and convert those using the specular-glossiness workflow to the metallic-roughness workflow to achieve a consistent PBR representation. This process resulted in a total of 29,200 models. For each 3D object, we render its RGBA albedo image, roughness map, and metallic map from 64 views of " + }, + { + "bbox": [ + 55, + 472, + 296, + 640 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 55, + 472, + 296, + 640 + ], + "type": "text", + "content": " for training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Network Architecture Our 3D U-Net is built upon octree-based convolutional neural networks [49] which only operates on non-empty octree leaf nodes. The depth of our constructed octree is set to 8 (resolution " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "). Our 3D U-Net consists of 5 down-sampling and up-sampling blocks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "For the text-conditioned generation, cross-attention layers are only inserted at the last two down-sampling blocks, the middle block, and the first two up-sampling blocks. The input channel of our 3D U-Net is set to 4, where 3 for normal vector and 1 for local displacement while the output channel is set to 13 as mentioned above. The resolution of 3D Gaussian Splatting is set to " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": ". The resolution of the baked albedo and material map is set to " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 174, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 342 + ], + "type": "text", + "content": "Training Details We train our model in a single-category and unconditional manner on ShapeNet, i.e., there are 4 TexGaussian models without text conditions in total. This per-category model is trained on 4 NVIDIA A100 (40G) GPUs for about 2 days and is set up for a fair comparison with other existing methods trained on ShapeNet. For the Objaverse dataset, we train a text-conditioned TexGaussian model for PBR material generation on 24 NVIDIA A100 (40G) GPUs for two weeks. For each batch, we randomly sample 8 camera views to calculate loss functions. We adopt the AdamW [27] optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 174, + 555, + 342 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 174, + 555, + 342 + ], + "type": "text", + "content": ", a weight decay of 0.05, and betas of (0.9,0.95). The learning rate is cosine annealed to 0 during the training and we clip the gradient with a maximum norm of 1.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 356, + 556, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 556, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 556, + 501 + ], + "type": "text", + "content": "Evaluation metrics To effectively assess the quality of our generative results, we adopt the metric proposed by [61]. Specifically, each mesh with the generated PBR material is rendered from 20 uniformly distributed views to get multi-view albedo maps, and PBR rendering results in a new illumination environment. These images are used to calculate the FID [36] and KID [2] scores against those from ground truth to evaluate the quality and diversity of generative PBR materials. For the ShapeNet dataset, we only use albedo maps. The final score is averaged across 20 views, and a lower FID and KID score indicates better generation quality and diversity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 507, + 524, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 524, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 524, + 520 + ], + "type": "text", + "content": "4.2. Unconditional RGB Texture Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 554, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 621 + ], + "type": "text", + "content": "We conduct unconditional RGB texture instead of PBR material generation on four categories of the ShapeNet dataset. We use the same train and test data split as in [26, 55]. Fig. 3 shows the unconditionally generated RGB texture by our single-category model with high quality, fidelity, and diversity. It can be seen that the texture map synthesized by our method is of great 3D global consistency with the corresponding input geometry." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "Quantitative Comparison We conduct quantitative analysis and comparison on our model and other state-of-the-art methods. Specifically, we compare our model with PointUV Diffusion [55] and TexOct [26] which also train single-category models on ShapeNet. We also compare with some methods using pre-trained 2D diffusion prior for multi-view images synthesis such as TEXTure [55], Text2Tex [8]" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "555" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 113, + 555, + 252 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of the FID and KID " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " score as well as the inference time of TexGaussian and other methods on the ShapeNet dataset [7]. The top part reports the comparison with methods that also train per-category models on ShapeNet. The bottom part reports the comparison with 2D diffusion-based methods, which select 50 samples from the test set of each category." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 113, + 555, + 252 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 555, + 252 + ], + "type": "table", + "html": "
MethodsAverageBenchCarChairTableTime
FID↓KID↓FID↓KID↓FID↓KID↓FID↓KID↓FID↓KID↓
Point-UV (1-Stage) [55]88.435.7864.961.44186.1017.5746.231.9356.422.1939.92s
Point-UV (2-Stage) [55]61.492.6767.481.6189.385.8241.331.5347.751.7249.75s
TexOct [26]59.452.6060.460.9790.106.1137.701.3649.521.9717.44s
Ours49.762.0746.370.4480.205.2229.961.0942.521.5411.02s
TEXTure [41]169.827.94152.535.28236.5818.52119.712.86170.465.11132.21s
Text2Tex [8]156.626.68154.955.79188.6015.06125.343.11157.572.771005.58s
Paint3D [56]135.254.67104.161.68204.2213.06104.642.24127.981.71231.56s
Ours100.971.8886.370.75117.535.2094.890.57105.090.9511.12s
", + "image_path": "a1ec1bcd7da090762641781e1e2cafe99fb22f27b6e3fa1ea4f1c11e76809a18.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 66, + 262, + 289, + 351 + ], + "blocks": [ + { + "bbox": [ + 66, + 262, + 289, + 351 + ], + "lines": [ + { + "bbox": [ + 66, + 262, + 289, + 351 + ], + "spans": [ + { + "bbox": [ + 66, + 262, + 289, + 351 + ], + "type": "image", + "image_path": "0c1a3879e9d5f1db1a007de4898b9a3d1061e5720bce488eb6182be471249af5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 358, + 296, + 612 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 612 + ], + "type": "text", + "content": "Figure 3. Unconditional RGB texture generative results on ShapeNet. Please zoom in for a better inspection of color details. Paint3D [56], and TexPainter [57] whose input text prompts are set to \"a *\" and \"* is the name of corresponding category. Due to the relatively long time consumption, we only use 50 3D objects selected from the test set to evaluate the methods based on 2D diffusion models. We conduct only a qualitative comparison for TexPainter [57] for its lengthy processing time. Table 1 reports the comparison of FID and KID scores as well as average inference time on a single NVIDIA A100 (40G) GPU. It is worth noting that our model only takes about one second to predict the parameters of each 3D Gaussian and the rest of the time is used for texture baking. From Table 1, we have the following observation. First, our method obtains the best performance in terms of FID and KID. For example, our method outperforms TexOct [26], by an average of 9.69 in FID and 0.53 in KID. These improvements indicate that our method excels at generating high-quality textures. Second, TexGaussian achieves the fastest generative speed in all categories, which is much less time-consuming than other methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "Qualitative Comparison Fig. 4 provides some qualitative results on the same input 3D meshes by different methods. Point-UV [55] and TexOct [26] use colored point clouds as supervision to train diffusion models. As a consequence, the generated texture map is relatively blurry due to the discreteness of the point cloud. For other methods that leverage pre-trained 2D diffusion prior, they can hardly comprehend the total geometry. The generated texture maps" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 268, + 555, + 605 + ], + "blocks": [ + { + "bbox": [ + 317, + 268, + 555, + 605 + ], + "lines": [ + { + "bbox": [ + 317, + 268, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 555, + 605 + ], + "type": "image", + "image_path": "86ea7666c7b2f0e83a065b8b10996d512d212979bca695d8dcd137b38bfd6843.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 613, + 555, + 647 + ], + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 647 + ], + "type": "text", + "content": "Figure 4. Examples of generated RGB textures obtained by Tex-Gaussian and other state-of-the-art models on the same 3D object. Please zoom in for a better inspection." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 659, + 556, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 556, + 719 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 556, + 719 + ], + "type": "text", + "content": "by them are not consistent with the semantics of input 3D mesh, such as the chaotic stripes on the bench and chair in Fig. 4. On the contrary, our method generates smooth and colorful RGB textures on unseen objects, which align well with 3D meshes." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "556" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 242 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 242 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 242 + ], + "type": "image", + "image_path": "517f40a175acc03512af0d34931028fd8e7ab6851c7e83a39793e85414a58263.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison with Fantasia3D [9], FlashTex [14], DreamMat [60] and Make-it-Real [16]. We show the rendered image, albedo map, roughness map, and metallic map for each 3D object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 292, + 281, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 281, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 281, + 304 + ], + "type": "text", + "content": "4.3. Text-conditioned PBR Material Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 310, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 310, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 296, + 453 + ], + "type": "text", + "content": "We train the text-conditioned PBR material generation model on our filtered subset of Objaverse [13]. We use the text descriptions of 3D objects from Cap3D [28, 29] to train our model. We choose 29,000 3D objects in our filtered subset for training and the rest 200 for testing. Fig. 1 shows some generative results by proposed TexGaussian on the test set. We can see that our model comprehends the overall geometry feature of input 3D mesh. It is capable of synthesizing high-quality albedo, roughness, and metallic, which are aligned well with 3D semantics such as accents with high metallic on the wooden box and rope with high roughness on the jug." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 464, + 296, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 656 + ], + "type": "text", + "content": "Quantitative Comparison We conduct quantitative comparison with some state-of-the-art text-conditioned PBR material synthesis methods: Fantasia3D [9], FlashTex [14], and DreamMat [60] on our test set. Table 2 reports FID and KID scores on both multi-view albedo map and PBR rendering images under the same illumination environment. TexGaussian outperforms all the baseline methods in achieving the best visual quality of the generated appearances. We also report the average inference time of different methods on a single NVIDIA A100 (40G) GPU across our test set. Due to the iterative optimization of score distillation sampling, all other three methods cost at least 20 minutes for the generation. In contrast, our TexGaussian only takes about 20 seconds, in which one second is for predicting the Gaussian parameters and the rest is for baking, which performs " + }, + { + "bbox": [ + 55, + 464, + 296, + 656 + ], + "type": "inline_equation", + "content": "60 \\times" + }, + { + "bbox": [ + 55, + 464, + 296, + 656 + ], + "type": "text", + "content": " faster than previous methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Qualitative Comparison Fig. 5 visualizes the generated albedo, roughness, and metallic of each compared method from the same text prompt and untextured meshes. We also show the PBR rendering images of the generated materials" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 293, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 591 + ], + "type": "text", + "content": "under the same environment light. It can be observed that Fantasia3D [9] and FlashTex [14] generate irregular colors on mesh surfaces, which lead to rendering results with low quality. DreamMat [60] is capable of generating a visually pleasing appearance for test 3D models. However, it can hardly align the generated PBR material well with the 3D semantics, such as the black band on the wooden barrel. What's more, it tends to generate over-saturated colors, which is demonstrated in the wooden chest and green grenade. In addition, all of the compared methods struggle to completely disentangle the light and texture, which results in relatively chaotic results of generated metallic and roughness. We also compare our method with a Multimodal Large Language Model based method, Make-it-Real [16], which accepts 3D objects paired with albedo maps generated by our TexGaussian as input. It can be observed that Make-it-Real struggles to capture the overall 3D structure of input meshes. As a consequence, it tends to predict single-value PBR maps, leading to irregular rendering results, such as the wooden chest with extremely large metallic value. On the contrary, our method is directly trained from 3D original data. It is capable of generating clean and smooth PBR material while fully comprehending the overall 3D structure. We provide more generative results in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Essentials of Octree We first analyze the effectiveness of using Octree in our pipeline. To do this, we calculate the quantitative quality of 3D Gaussian pre-fitting by our octant-aligned and full voxel-aligned 3D Gaussian Splitting on the subset of Objaverse dataset [13]. The depth of octree in this ablation is set to 8, and the resolution of full voxels is " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "256^3" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", which is consistent with the finest resolution of octree to guarantee comparison fairness. Table 3 reports the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "557" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 123, + 298, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of the FID and KID " + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": " scores as well as the inference time of TexGaussian and other methods on our test set which consists of 200 3D objects with ground truth PBR materials." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 123, + 298, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 123, + 298, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 298, + 198 + ], + "type": "table", + "html": "
MethodsAlbedoPBR renderingTime
FID↓KID↓FID↓KID↓
Fantasia3D [9]213.210.96209.870.4822.7mins
FlashTex [14]185.241.13186.820.4220.3mins
DreamMat [60]152.631.09145.490.1948.2mins
Ours123.720.20129.520.0221.04s
", + "image_path": "a9e0046238c2ad3144d75a1dc1a326376d7c3824fbf3ae06829afc03ad7dc1fd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 201, + 296, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 224 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of methods using different types of 3D Gaussian Fitting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 232, + 296, + 401 + ], + "blocks": [ + { + "bbox": [ + 56, + 232, + 296, + 401 + ], + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 401 + ], + "type": "image", + "image_path": "80bf1f8d675abd45b41508bb71945ceb6ca841b48ddaefb3d3f3fc872253f0f8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 411, + 296, + 445 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 445 + ], + "type": "text", + "content": "Figure 6. Visualization of different manners of Gaussian fitting. The rendering results demonstrate excellent reconstruction performance of the proposed octant-aligned 3D Gaussian Splatting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": "PSNR, LPIPS [59], and SSIM [50] metrics on the albedo map as well as the average number of 3D Gaussians of octant and full voxel aligned 3D Gaussian Splitting across our Objverse subset. Fig. 6 also presents some visual results of our Gaussian fitting. These results demonstrate that the proposed octant-aligned 3D Gaussian produces a much more reasonably compact and precise representation of diverse and complex 3D assets with much fewer 3D Gaussians compared to the full voxel version." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 560, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 598 + ], + "type": "text", + "content": "Essentials of different losses We verify the importance of different losses we proposed to train our model. We term all the losses calculated on 2D image space as " + }, + { + "bbox": [ + 55, + 560, + 296, + 598 + ], + "type": "inline_equation", + "content": "L_{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 560, + 296, + 598 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 608, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 608, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 107, + 608, + 295, + 620 + ], + "type": "interline_equation", + "content": "L _ {2 \\mathrm {D}} = L _ {\\text {a l b e d o}} + L _ {\\mathrm {R}} + L _ {\\mathrm {M}} + L _ {\\alpha}. \\tag {7}", + "image_path": "90ce2ae0bb0bf7b2cc4504353e85bf21f1ccccfa2385903510e90241addef0f0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "We train two additional TexGaussian models using only the " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{3\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " loss on ShapeNet car category due to its large variations and complexity of texture to validate their effectiveness. It is worth noting that in this dataset, " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{2\\mathrm{D}} = L_{\\mathrm{albedo}} + L_{\\alpha}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " due to the lack of material information. The training curves of MSE loss and LPIPS loss between rendering images and ground-truth ones are shown" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 433, + 163 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 433, + 163 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 433, + 163 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 433, + 163 + ], + "type": "image", + "image_path": "aa8ab213859c7a239e18cd0ddc54efda5b41131873f7038710356170d0c1847a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 172, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 194 + ], + "type": "text", + "content": "Figure 7. The training loss curves of our model with different losses." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 439, + 70, + 548, + 164 + ], + "blocks": [ + { + "bbox": [ + 439, + 70, + 548, + 164 + ], + "lines": [ + { + "bbox": [ + 439, + 70, + 548, + 164 + ], + "spans": [ + { + "bbox": [ + 439, + 70, + 548, + 164 + ], + "type": "image", + "image_path": "05c688ef0ceed95203f5679d5caf9a44010ad346584880e664c0a05507d8bbf1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 204, + 555, + 289 + ], + "blocks": [ + { + "bbox": [ + 316, + 204, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 316, + 204, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 204, + 555, + 289 + ], + "type": "image", + "image_path": "95799d7d4b83d29b0b0e85f6ef03b33c93c8a2502927329eac9a1bcd5dbdb8ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 296, + 546, + 308 + ], + "lines": [ + { + "bbox": [ + 321, + 296, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 296, + 546, + 308 + ], + "type": "text", + "content": "Figure 8. Results of our model trained with and without " + }, + { + "bbox": [ + 321, + 296, + 546, + 308 + ], + "type": "inline_equation", + "content": "L_{2D}" + }, + { + "bbox": [ + 321, + 296, + 546, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": "in Fig. 7, which demonstrate the effects of " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "L_{2D}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "L_{3D}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": ". From the loss curves, we can conclude that " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "L_{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " facilitates the process of convergence and " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "L_{2D}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " enhances the quality of synthesized images, as evidenced by the large margin of improvement in the LPIPS loss when introducing " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "L_{2D}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": ". We also provide some qualitative results to verify the effectiveness of " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "L_{2D}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " on the test set of car category in Fig. 8. Only using " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "L_{3D}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " results in a relatively blurry texture map due to the discreteness of 3D Gaussian which is similar to the characteristic of 3D point clouds analyzed above." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 457, + 388, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 388, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 388, + 470 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 477, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 555, + 634 + ], + "type": "text", + "content": "In this paper, we proposed TexGaussian, an octree-based 3D Gaussian Splatting model for high-quality PBR material generation on untextured meshes. We aligned each 3D Gaussian on the octant of the corresponding octree built from the input untextured object and extended the parameters of 3D Gaussian with additional channels to represent the roughness and metallic map. We trained our model with regression objectives, achieving faster inference speed compared to previous texture synthesis methods. Experimental results demonstrated that our method is capable of generating high-quality PBR materials that are readily usable in modern graphics engines for photo-realistic rendering, offering enhanced realism for a variety of applications." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Limitations The generalization of TexGaussian is still hindered by the scale of the training set. Thus it struggles to generate various textures for some extremely complex 3D objects beyond our training data. We are looking forward to training our TexGaussian model with more parameters and more data on a larger-scale GPU cluster in the future." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "558" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 170 + ], + "type": "text", + "content": "[2] Mikołaj Binkowski, Dougal J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 296, + 227 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 350 + ], + "type": "text", + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 407 + ], + "type": "text", + "content": "[6] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and KangXue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 486 + ], + "type": "text", + "content": "[7] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "type": "text", + "content": "[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 534, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 589 + ], + "type": "text", + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. Advances in Neural Information Processing Systems, 35:30923-30936, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "text", + "content": "[11] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 148 + ], + "type": "text", + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of " + }, + { + "bbox": [ + 316, + 73, + 554, + 148 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 316, + 73, + 554, + 148 + ], + "type": "text", + "content": " objects. arXiv preprint arXiv:2307.05663, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 150, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 150, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 555, + 215 + ], + "type": "text", + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3, 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 217, + 554, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 270 + ], + "type": "text", + "content": "[14] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. Flashtex: Fast reconfigurable mesh texturing with lightcontrolnet. In European Conference on Computer Vision (ECCV), 2024. 3, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 304, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 347 + ], + "type": "text", + "content": "[16] Ye Fang, Zeyi Sun, Tong Wu, Jiaqi Wang, Ziwei Liu, Gordon Wetzstein, and Dahua Lin. Make-it-real: Unleashing large multimodal model for painting 3d objects with realistic materials, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 349, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 370 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[18] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "text", + "content": "[19] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "text", + "content": "[20] Dong Huo, Zixin Guo, Xinxin Zuo, Zhihao Shi, Juwei Lu, Peng Dai, Songcen Xu, Li Cheng, and Yee-Hong Yang. Tex-gen: Text-guided 3d texture generation with multi-view sampling and resampling. ECCV, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 505, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 553, + 568 + ], + "type": "text", + "content": "[21] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian-shader: 3d gaussian splattering with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "text", + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 614, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 668 + ], + "type": "text", + "content": "[23] Matthias Labschütz, Katharina Krosl, Mariebeth Aquino, Florian Grashäftl, and Stephanie Kohl. Content creation for a 3d game with maya and unity 3d. Institute of Computer Graphics and Algorithms, Vienna University of Technology, 6(124):2, 2011. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "559" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[25] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "text", + "content": "[26] Jialun Liu, Chenming Wu, Xinqi Liu, Xing Liu, Jinbo Wu, Haotian Peng, Chen Zhao, Haocheng Feng, Jingtuo Liu, and Errui Ding. Texoct: Generating textures of 3d models with octree-based diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4284-4293, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 197, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 197, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 197, + 295, + 217 + ], + "type": "text", + "content": "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 251 + ], + "type": "text", + "content": "[28] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 254, + 294, + 285 + ], + "type": "text", + "content": "[29] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. arXiv preprint arXiv:2404.07984, 2024. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 287, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 287, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 57, + 287, + 294, + 352 + ], + "type": "text", + "content": "[30] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 354, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 354, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 354, + 294, + 407 + ], + "type": "text", + "content": "[31] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 57, + 411, + 294, + 453 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 294, + 498 + ], + "type": "text", + "content": "[33] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 501, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 501, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 501, + 294, + 542 + ], + "type": "text", + "content": "[34] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 294, + 588 + ], + "type": "text", + "content": "[35] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings IEEE International Conf. on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 622 + ], + "type": "text", + "content": "[36] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "text", + "content": "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 228 + ], + "type": "text", + "content": "[41] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 262 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 265, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 265, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 553, + 330 + ], + "type": "text", + "content": "[43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 332, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 385 + ], + "type": "text", + "content": "[44] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 452 + ], + "type": "text", + "content": "[45] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 72-88. Springer, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 455, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 553, + 507 + ], + "type": "text", + "content": "[46] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 511, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 553 + ], + "type": "text", + "content": "[47] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 556, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 588 + ], + "type": "text", + "content": "[48] Peng-Shuai Wang. Octformer: Octtree-based transformers for 3D point clouds. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "text", + "content": "[49] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-CNN: Octree-based convolutional neural networks for 3D shape analysis. ACM Transactions on Graphics (SIGGRAPH), 36(4), 2017. 2, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[50] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "text", + "content": "[51] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "560" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 620 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "tion. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[52] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Toon-crafter: Generative cartoon interpolation. arXiv preprint arXiv:2405.17933, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[53] Yu-Ying Yeh, Jia-Bin Huang, Changil Kim, Lei Xiao, Thu Nguyen-Phuoc, Numair Khan, Cheng Zhang, Manmohan Chandraker, Carl S Marshall, Zhao Dong, et al. Texture-dreamer: Image-guided texture synthesis through geometry-aware diffusion. arXiv preprint arXiv:2401.09416, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[54] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 306 + ], + "type": "text", + "content": "[55] Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Zhengzhe Liu, and Xiaojuan Qi. Texture generation on 3d meshes with pointuv diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4206-4216, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "text", + "content": "[56] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4252-4262, 2024. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 419 + ], + "type": "text", + "content": "[57] Hongkun Zhang, Zherong Pan, Congyi Zhang, Lifeng Zhu, and Xifeng Gao. Texpainter: Generative mesh texturing with multi-view consistency. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 452 + ], + "type": "text", + "content": "[58] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 509 + ], + "type": "text", + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 573 + ], + "type": "text", + "content": "[60] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, and Xiaogang Jin. Dreammat: High-quality pbr material generation with geometry- and light-aware diffusion models. ACM Trans. Graph., 43(4), 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 294, + 620 + ], + "type": "text", + "content": "[61] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 5" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "561" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_content_list.json b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d2013704cf5b349dbf2204ce99754e50f496bb8b --- /dev/null +++ b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_content_list.json @@ -0,0 +1,1617 @@ +[ + { + "type": "text", + "text": "Text Augmented Correlation Transformer For Few-shot Classification & Segmentation", + "text_level": 1, + "bbox": [ + 135, + 130, + 864, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Srinivasa Rao Nandam $^{1}$ Sara Atito $^{1,2}$ Zhenhua Feng $^{3}$ Josef Kittler $^{2}$ Muhammad Awais $^{1,2}$ $^{1}$ Surrey Institute for People-Centred AI (PAI), University of Surrey, UK $^{2}$ CVSSP, University of Surrey, UK $^{3}$ School of AI and Computer Science, Jiangnan University, China", + "bbox": [ + 99, + 202, + 952, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{s.nandom, sara.atito, j.kittler, muhammad.awais}@surrey.ac.uk, fengzhenhua@jiangnan.edu.cn", + "bbox": [ + 122, + 258, + 926, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation models like CLIP and ALIGN have transformed few-shot and zero-shot vision applications by fusing visual and textual data, yet the integrative few-shot classification and segmentation (FS-CS) task primarily leverages visual cues, overlooking the potential of textual support. In FS-CS scenarios, ambiguous object boundaries and overlapping classes often hinder model performance, as limited visual data struggles to fully capture high-level semantics. To bridge this gap, we present a novel multi-modal FS-CS framework that integrates textual cues into support data, facilitating enhanced semantic disambiguation and fine-grained segmentation. Our approach first investigates the unique contributions of exclusive text-based support, using only class labels to achieve FS-CS. This strategy alone achieves performance competitive with vision-only methods on FS-CS tasks, underscoring the power of textual cues in few-shot learning. Building on this, we introduce a dual-modal prediction mechanism that synthesizes insights from both textual and visual support sets, yielding robust multimodal predictions. This integration significantly elevates FS-CS performance, with classification and segmentation improvements of $+3.7 / 6.6\\%$ (1-way 1-shot) and $+8.0 / 6.5\\%$ (2-way 1-shot) on COCO-20 $^i$ , and $+2.2 / 3.8\\%$ (1-way 1-shot) and $+4.3 / 4.0\\%$ (2-way 1-shot) on Pascal-5 $^i$ . Additionally, in weakly supervised FS-CS settings, our method surpasses visual-only benchmarks using textual support exclusively, further enhanced by our dual-modal predictions. By rethinking the role of text in FS-CS, our work establishes new benchmarks for multi-modal few-shot learning and demonstrates the efficacy of textual cues for improving model generalization and segmentation accuracy.", + "bbox": [ + 86, + 340, + 485, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 843, + 220, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot learning [6, 7] has emerged as a significant area of research, focusing on building models that can effec", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/766ded4397918efa3d95812c40c0274dd122999e87bd8b522d99ce81edd81acc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 306, + 591, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb8e5e51d617430701170f82f8dbb45579f9e66465342ac083f7cfd93c13767f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 308, + 669, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b222aa094a3e872ab1a3e90746fbad4032cc88bb43f8afeb5ad38d0944cdaf50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 308, + 746, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e67a53afcba5103f630079bf64933410a69383320d83509ca4ef700dcc67d08a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 308, + 825, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4efdc0c96ff5acfbf05fd43cf7108f94d1902183279fbd1b7d4c6aabd5649419.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 308, + 903, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a9682150c57772e621689bf8cc494069cf1c6bf83a23e221c2a2143af516bdf.jpg", + "image_caption": [ + "Qry.Img" + ], + "image_footnote": [], + "bbox": [ + 514, + 373, + 591, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67a215f0e241455f3b08dd02120e85bc13b90176c81c514ec91bdab230a5dd04.jpg", + "image_caption": [ + "Spt.Img 1" + ], + "image_footnote": [], + "bbox": [ + 594, + 375, + 669, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/42f6541132c25a8f9687b8c6d29dfd7e8a0ae149448889825952d25093f7b930.jpg", + "image_caption": [ + "Img.Pred" + ], + "image_footnote": [], + "bbox": [ + 671, + 373, + 746, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af54713ec10222adfe283440cd09f447f4abd538212e6126bf65122919fb2620.jpg", + "image_caption": [ + "T-Only Pred", + "Figure 1. Image predictions under-segment, text over-segments, and text-aug is most accurate in 2-way-1 shot segmentation using image, text, and text-aug support (shown: one support image and query image with ground truth; second support image not displayed)." + ], + "image_footnote": [], + "bbox": [ + 750, + 373, + 825, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48a62db0cd327eebd904d9056227cf96a9a18df36fb7fd57436b23b182b3812f.jpg", + "image_caption": [ + "T-Aug. Pred" + ], + "image_footnote": [], + "bbox": [ + 828, + 373, + 903, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tively learn from a small set of examples, typically ranging from just 1 to 5 samples per class. This capability is essential in scenarios where data is scarce or expensive to obtain. Within this domain, Few-Shot Classification (FS-C) [25, 26, 33, 37, 39] and Few-Shot Segmentation (FS-S) [5, 22, 23, 36] have been widely studied, each focusing on classifying or segmenting images with minimal supervision. However, both FS-C and FS-S are constrained by specific limitations that impact their effectiveness in real-world applications.", + "bbox": [ + 509, + 561, + 906, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "FS-C typically assumes each query image contains only one target class, limiting its applicability in complex scenes with multiple classes. On the other hand, FS-S, while supporting multi-class segmentation, requires that at least one target class appears in each query image, resulting in segmentation errors when no target classes are present. These rigid assumptions hinder the practical deployment of FS-C and FS-S in dynamic environments where real-world images seldom adhere to such constraints.", + "bbox": [ + 509, + 715, + 908, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these limitations, Integrative Few-Shot Classification and Segmentation (FS-CS) combines the strengths of FS-C and FS-S into a unified framework [11].", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25357", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77bfd4ff6705810c2cce1bbb1948cf255915d7b1966ae3b40219f5cc7e0d1594.jpg", + "image_caption": [ + "Figure 2. FS-CS with text-only using class label alone and text-augmented combining text with image predictions: The support set contains images with ground truth masks and class labels." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 464, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FS-CS enables multi-class classification and segmentation even when no target classes appear in the query image, offering a more adaptable approach for real-world applications. This flexibility allows FS-CS to handle diverse scenarios, even though it may slightly underperform compared to traditional FS-C and FS-S on synthetic benchmarks [11].", + "bbox": [ + 89, + 277, + 482, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While previous FS-CS methods [11, 12] have made significant progress, they predominantly rely on visual data. Recent advances in few-shot and zero-shot vision applications, such as few-shot classification [32, 41, 42], few-shot segmentation (FSS) [24, 29], and zero-shot segmentation [4, 14, 43], have shown the benefits of incorporating textual information from models like CLIP [20]. Prior works in FSS have used textual cues from CLIP [29, 30] and Word2Vec embeddings [24], yet these studies did not explore the integration of rich text specifically for FS-CS or the potential of using an exclusive text support set. While zero-shot segmentation methods can generate segmentation directly from text, there are key distinctions: our method is trained in a few-shot episodic manner, whereas zero-shot methods train across the full dataset on all seen classes. Additionally, zero-shot segmentation methods are not inherently designed to produce classification predictions, and extending them for classification is non-trivial. Finally, our approach supports flexible configurations such as N-way 1-shot, 1-way K-shot, and N-way K-shot settings, which are not feasible with zero-shot segmentation methods.", + "bbox": [ + 91, + 369, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FS-CS predictions based solely on visual support sets often suffer from semantic ambiguity, leading to issues such as under-segmentation and misclassification in cases of unclear background/foreground boundaries or overlapping classes. Conversely, predictions based solely on text may lack fine-grained detail, resulting in over-segmentation (Figure 1). These challenges highlight the need for a multimodal FS-CS approach that can effectively combine visual and textual cues.", + "bbox": [ + 89, + 686, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a novel FS-CS approach that uniquely supports both text-only and multi-modal (text and image) support sets. Our method leverages a correlation-based approach, a technique shown to be effective in FS-CS tasks [11, 12], where foreground maps for classification and", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "segmentation are generated by calculating correlations between query and support features. To accomplish this, we utilize the frozen CLIP [20] model, a multi-modal architecture that processes both image and text modalities through dedicated encoders, producing aligned embeddings for correlation. By using frozen CLIP instead of vision-only models like DINO [2], we maintain a comparable level of model complexity without adding learnable parameters, resulting in only marginal increases in memory consumption and runtime (Table 4).", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fully exploit these aligned vision and text embeddings, we propose the Text-Augmented Classification-Segmentation Transformer (TA-CST). TA-CST incorporates two transformer blocks: one for vision-only correlations (between query and support images) and another for vision-text correlations (between query images and support text). However, the feature maps generated by a correlation transformer block [12] alone have limited capacity for capturing both multi-scale and global semantics. To address this, we introduce a Feature-Enrichment Module that captures both granular and global semantics, enabling finer segmentation and more accurate classification. Unlike traditional FSS methods, which achieve multi-scale features through complex architectures [28] or multiple cross-attention layers [19, 30] and lack global semantic capture, our approach uses a single-layer attention block with granular Key and Value (KV) representations via atrous convolutions at varying scales [3]. This module captures multiscale information and global semantics, enhancing the FS-CS task by integrating both fine-grained details and broader context (Figure 6).", + "bbox": [ + 511, + 244, + 906, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In cases where the support set includes only text, our classification performance can be limited, particularly with smaller datasets like Pascal- $5^{i}$ . To address this, we propose a zero-shot knowledge distillation strategy that aligns our model's classification predictions with zero-shot predictions generated by CLIP. Specifically, we generate zero-shot predictions from CLIP using class-specific positive and negative prompts, such as \"This is a photo of {class-label}\" and \"This is not a photo of {class-label}.\" This approach enhances performance across datasets of varying sizes, preserving CLIP's generalization capabilities in text-only settings.", + "bbox": [ + 511, + 565, + 908, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig 2 shows our FS-CS pipeline, which can handle generate predictions with exclusive text (text-only) or multimodal (text-augmented) support sets. Our experiments demonstrate that text-only support is a viable solution for FS-CS, achieving competitive performance with leading methods on the Pascal- $5^{i}$ and $\\mathrm{COCO - }20^{i}$ datasets. We achieve even higher performance with multi-modal support, demonstrating the effectiveness of our novel textaugmented FS-CS approach. For a fair comparison, we also replicate CST with CLIP-B/16 as our baseline.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25358", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide a brief summary of our novel contributions below:", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We first demonstrate that textual data alone can solve the FS-CS task, achieving competitive results, and show that combining text and image-based predictions further improves FS-CS performance.", + "- We introduce a novel Feature Enrichment Module, which enhances feature maps with multi-scale and global semantics, resulting in better segmentation and classification predictions through a single attention layer with multi-scale Key and Value representations.", + "- In the text-only setting, we introduce a novel distillation strategy to maintain CLIP's classification generalization, aligning our model's performance with zero-shot predictions from CLIP.", + "- We extend the correlation transformer used in prior methods to support multi-modal support sets, enabling fusion of classification and segmentation predictions from both visual and textual cues." + ], + "bbox": [ + 89, + 123, + 482, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 392, + 232, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot classification (FS-C) & Segmentation (FS-S) Existing few-shot classification methods include metric learning [25, 26, 37, 40], meta-learning [8, 13, 39], and finetuning [15, 16, 27]. Metric learning constructs an embedding space that clusters similar samples while distancing dissimilar ones. Meta-learning emphasizes adapting to new tasks from limited samples, and fine-tuning enhances a well-initialized feature space. Few-shot segmentation methods adopt various strategies: some determine classifier weights per class [21, 22], others leverage prototype learning [5, 23], and some incorporate a decoder to produce segmentation outputs [18, 36, 38]. ASNet [11] introduced the combined task of few-shot classification and segmentation, using a ResNet backbone [10] for feature extraction. CST [12] builds on this in weakly supervised settings with a self-supervised DINO [2] backbone for correlation mapping.", + "bbox": [ + 89, + 417, + 482, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classification & Segmentation with Text Recent works [9, 32, 41, 42] explore prompt tuning for CLIP [20] in few-shot multi-label classification. Although CLIP has been used for text-based supervised [34, 35], weakly supervised [14], zero-shot segmentation [1, 4, 43], as well as FSS [24, 29, 30], it has not been explored for FS-CS. Text-based FSS methods [24, 29, 30] rely on complex architectures with multiple components and cross-attention blocks, constrained by at least one target class in the query. These methods also fail to generate class predictions, unlike our simple FS-CS approach. While [30] uses positive/negative prompts with Grad-CAM priors for FS-S, we apply them for zero-shot distillation, preserving CLIP's generalizability. GPT4-Vision [17] tackles weakly supervised FS-CS with an LLM and lacks exclusive text", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "support, unlike our compact CLIP-B/16, which enables text-based prediction generation and augmentation while addressing unique FS-CS challenges.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 148, + 648, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents our novel framework for few-shot classification and segmentation (FS-CS) that supports predictions from a multi-modal support set, consisting of both textual labels and support images. Using a meta-learning approach inspired by correlation-based methods [11, 12, 18], our model is trained on episodic tasks, each comprising a query image with a support set. We follow an $N$ -way $K$ -shot setup, where there are $K$ examples of $N$ classes in the support set. The model's objective is to generate a foreground mask for all target classes along with classification predictions for each, as specified in the support set.", + "bbox": [ + 511, + 174, + 903, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our Text-Augmented Classification-Segmentation Transformer (TA-CST) advances beyond traditional FS-CS methods [11, 12], which rely solely on visual data, thus limiting support set options to vision only. By incorporating textual cues from class labels, our approach enhances prediction quality by using the textual predictions in tandem with visual predictions. This is achieved through a dual correlation transformer, designed specifically to handle both visual and textual support sets effectively, but limited in its ability to capture both global semantics and multi-scale features crucial for FS-CS tasks.", + "bbox": [ + 511, + 340, + 903, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these limitations, we introduce a feature enrichment module that augments the transformer's capacity to process varying spatial scales critical for segmentation tasks, including both few-shot and traditional segmentation [3, 5], while maintaining the global context essential for classification. The feature enrichment module achieves this balance through a single attention layer with a granular Key-Value (KV) structure, which embeds multi-scale inductive bias without losing the global semantics of the transformer.", + "bbox": [ + 511, + 507, + 903, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During experiments, we observed a drop in classification accuracy when using exclusive text-based support, particularly with smaller datasets (e.g., Pascal-5 $^i$ ). To address this and preserve the generalization capability of CLIP in classification, we propose a zero-shot knowledge distillation strategy. Figure 3 provides an architectural overview, and the following subsections elaborate on each component.", + "bbox": [ + 511, + 657, + 903, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Text augmented Classification-Segmentation Transformer", + "text_level": 1, + "bbox": [ + 511, + 772, + 903, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model produces classification and segmentation predictions by generating foreground maps through correlations between the query images and the support set, which includes both images and text [11, 12, 18]. To accomplish this, we leverage CLIP's vision and text transformer encoders [20], taking advantage of its ability to produce in", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25359", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e203a461d5cf25c675d69dd0f157772285ad91ddccd10ef9d1a008a389846f4d.jpg", + "image_caption": [ + "Feature Extraction" + ], + "image_footnote": [], + "bbox": [ + 99, + 117, + 364, + 323 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0f8856b400493818596d2b6c1a5a3d929003a24eb11e72a8aceaca0628314b94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 94, + 674, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2fcd8b2fff2a191d43251eb66fc771d38727716b7ed32aafe1aa0e836cd660ab.jpg", + "image_caption": [ + "Text Only", + "Figure 3. TA-CST pipeline, where predictions are generated from correlations: Vision correlation comes from support/query images, and vision-text correlation from support text/query images. Text-only generates FS-CS predictions from the vision-text correlation, while text-augmented combines vision-only and vision-text FS-CS predictions." + ], + "image_footnote": [], + "bbox": [ + 395, + 222, + 674, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c4086eec0ddf2574617734fc6594dfbf3f5ae7f4afc04d2397d207e8268323b.jpg", + "image_caption": [ + "Feat. Enr. Block" + ], + "image_footnote": [], + "bbox": [ + 707, + 101, + 879, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "herently aligned text and vision token embeddings. The required foreground maps for classification and segmentation are generated by correlating the query images with support images and text using our dual correlation transformer.", + "bbox": [ + 89, + 433, + 482, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a query image $\\mathbf{q} \\in \\mathbb{R}^{H \\times W \\times C}$ , where $H$ , $W$ , and $C$ represent height, width, and channels, we obtain class output tokens $\\mathbf{V}_q^{\\mathrm{cls}} \\in \\mathbb{R}^D$ and patch output tokens $\\mathbf{V}_q^{\\mathrm{pt}} \\in \\mathbb{R}^{D \\times N_q}$ from the CLIP vision encoder. Similarly, for a support image $\\mathbf{s} \\in \\mathbb{R}^{H \\times W \\times C}$ , we derive class tokens $\\mathbf{V}_s^{\\mathrm{cls}} \\in \\mathbb{R}^D$ and patch tokens $\\mathbf{V}_s^{\\mathrm{pt}} \\in \\mathbb{R}^{D \\times N_s}$ . For the support text, we use the CLIP text encoder to extract class tokens $\\mathbf{T}_s^{\\mathrm{cls}} \\in \\mathbb{R}^{D'}$ and patch tokens $\\mathbf{T}_s^{\\mathrm{pt}} \\in \\mathbb{R}^{D' \\times N'}$ .", + "bbox": [ + 89, + 494, + 482, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We calculate the vision-only correlation by taking the unnormalized dot product between the query patch tokens $\\mathbf{V}_q^{\\mathrm{pt}}$ and the support class token $\\mathbf{V}_s^{\\mathrm{cls}}$ , noted as $v\\_ corr^{\\mathrm{cls}}$ , and then between $\\mathbf{V}_q^{\\mathrm{pt}}$ and the support patch tokens $\\mathbf{V}_s^{\\mathrm{pt}}$ , referred to as $v\\_ corr^{\\mathrm{pt}}$ [12]. For correlation between vision and text, we project the query's vision tokens $\\mathbf{V}_q^{\\mathrm{cls}}$ and $\\mathbf{V}_q^{\\mathrm{pt}}$ and the support's text tokens $\\mathbf{T}_s^{\\mathrm{cls}}$ and $\\mathbf{T}_s^{\\mathrm{pt}}$ into a common space using CLIP's projection layers. We then compute the unnormalized dot product between these tokens, yielding vision-text correlation for the EOS text token $t\\_ corr^{\\mathrm{cls}}$ (from $\\mathbf{V}_q^{p,\\mathrm{pt}}$ and $\\mathbf{T}_s^{p,\\mathrm{cls}}$ ) and for text patch tokens $t\\_ corr^{\\mathrm{pt}}$ (from $\\mathbf{V}_q^{p,\\mathrm{pt}}$ and $\\mathbf{T}_s^{p,\\mathrm{pt}}$ ).", + "bbox": [ + 89, + 621, + 482, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dual Correlation Transformer (Corr-T): Our dual correlation transformer module consists of two separate transformers, each with two layers and query pooling for efficient processing[12]; one transformer is for vision-only correlation, and another for vision-text correlation.", + "bbox": [ + 89, + 806, + 482, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the vision-only pathway, we concatenate class and", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "patch correlation outputs into $v_{\\mathrm{corr}} = \\{v_{\\mathrm{corr}}^{\\mathrm{cls}}, v_{\\mathrm{corr}}^{\\mathrm{pt}}\\} \\in \\mathbb{R}^{N_q \\times (N_s + 1) \\times ML}$ , which is processed by the vision-only transformer. Attention is applied with query downsampling using (4,4) and (3,3) kernels for each layer. Patch tokens are reshaped into spatial dimensions before downsampling, resulting in a final patch output dimension of $1 \\times 1$ . This produces foreground maps that are reshaped to $\\mathbb{R}^{H' \\times W' \\times C'}$ for $\\mathbf{A}_{\\mathrm{seg}} \\in \\mathbb{R}^{N_q \\times C'}$ and $\\mathbf{A}_{\\mathrm{cls}} \\in \\mathbb{R}^{N_q \\times C'}$ .", + "bbox": [ + 511, + 433, + 906, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the vision and text pathway, we concatenate vision and text class and patch correlation as $t_{\\mathrm{corr}} = \\{t_{\\mathrm{corr}}^{\\mathrm{cls}}, t_{\\mathrm{corr}}^{\\mathrm{pt}}\\} \\in \\mathbb{R}^{N_q \\times (N' + 1) \\times ML}$ . Later, attention is applied with (3,3) kernels for downsampling of query in both layers. This yields foreground maps $\\mathbf{B}_{\\mathrm{cls}} \\in \\mathbb{R}^{N_q \\times C'}$ and $\\mathbf{B}_{\\mathrm{seg}} \\in \\mathbb{R}^{N_q \\times C'}$ , which are also reshaped to $\\mathbb{R}^{H' \\times W' \\times C'}$ as the output.", + "bbox": [ + 511, + 560, + 908, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Enrichment Module: We propose a feature enrichment module to enhance classification and segmentation feature map predictions by incorporating both global and fine-grained details. Global context is essential for classification and also supports segmentation, while multi-scale feature capture is particularly advantageous for segmentation, helping to improve classification by minimizing errors from an over-reliance on global context alone. To meet these requirements, we designed an attention-based module capable of capturing both global and multi-scale information. While attention mechanisms are effective at capturing global information, they generally lack an inherent bias toward capturing objects at multiple scales. To address this, we introduce a granular Key-Value (KV) structure, which provides the classic attention block with multi", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25360", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scale capabilities by capturing information at various scales. The granular KV is generated using various atrous convolutions [3].", + "bbox": [ + 89, + 90, + 482, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For an input $\\mathbf{X}$ to the multi-scale encoding module, the granular key-value pairs (KV) are generated as per Equation 2, where GAP refers to global average pooling, with dilation rates $d1$ , $d2$ , and $d3$ . The final output of our feature enrichment module, Attn, where granular KV is used for key, value(Equation 3).", + "bbox": [ + 89, + 136, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K V} _ {i n} = \\operatorname {C o n c a t} \\left\\{\\left(\\operatorname {C o n v} (1 \\times 1), \\operatorname {C o n v} (3 \\times 3, r = d _ {1}) \\right. \\right.,\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 241, + 478, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\operatorname {C o n v} \\left(3 \\times 3, r = d _ {2}\\right), \\operatorname {C o n v} \\left(3 \\times 3, r = d _ {3}\\right), G A P \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 260, + 482, + 276 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K V} _ {\\text {m u l t}} = \\operatorname {C o n v} \\left(\\left(1 \\times 1\\right), 5 C ^ {\\prime}, C ^ {\\prime}\\right) \\left(\\mathbf {K V} _ {\\text {i n}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 301, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A t t n} = \\text {M u l t i H e a d A t t n} (Q, \\mathbf {K} _ {\\text {m u l t}}, \\mathbf {V} _ {\\text {m u l t}}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 334, + 482, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\hat{\\mathbf{A}}_{cls}$ , $\\hat{\\mathbf{A}}_{seg}$ , $\\hat{\\mathbf{B}}_{cls}$ , and $\\hat{\\mathbf{B}}_{seg}$ denote the outputs of the multi-scale encoding module, derived from the feature maps $\\mathbf{A}_{cls}$ , $\\mathbf{A}_{seg}$ , $\\mathbf{B}_{cls}$ , and $\\mathbf{B}_{seg}$ , respectively.", + "bbox": [ + 89, + 353, + 483, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The enriched vision-only classification foreground map, $\\hat{\\mathbf{A}}_{cls}$ , and the enriched vision-text foreground map, $\\hat{\\mathbf{B}}_{cls}$ , are then passed to a classification head, which consists of two pointwise convolutions followed by global average pooling to produce the classification outputs $\\mathbf{Y}_{cls}^{V}$ and $\\mathbf{Y}_{cls}^{T}$ . To refine the segmentation maps $\\hat{\\mathbf{A}}_{seg}$ and $\\hat{\\mathbf{B}}_{seg}$ , these are first processed by a segmentation head. This head utilizes $3 \\times 3$ convolutional layers, which are alternated with a bilinear interpolation layer. This combination allows the segmentation predictions to be upscaled to match the original image dimensions, producing $\\mathbf{Y}_{seg}^{V}$ and $\\mathbf{Y}_{seg}^{T}$ , the final segmentation outputs.", + "bbox": [ + 89, + 400, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero Shot Knowledge Distillation: Classification performance in our method suffers when support samples are limited to text, especially on datasets like Pascal- $5^{i}$ , where the training set is smaller, amplifying the accuracy drop. To counter this, we utilize CLIP's zero-shot capability [20]. Unlike standard zero-shot classification, where text projections of known classes are compared to image projections, our task involves determining the presence of a specific support class. In zero-shot prediction, we use the positive prompt, 'This is a photo of {class-label}', followed by the negative prompt, 'This is not a photo of {class-label}.' Both prompts are processed through the CLIP text encoder and projection layer, producing the output embeddings $\\mathbf{T}_s^{+,p\\cdot cl_s}$ and $\\mathbf{T}_s^{-,p\\cdot cl_s}$ for the [EOS] token. The similarities, Pos and Neg, are then defined between the query image projection $\\mathbf{V}_q^{p\\cdot cl_s}$ and the positive and negative projections, $\\mathbf{T}_s^{+,p\\cdot cl_s}$ and $\\mathbf{T}_s^{-,p\\cdot cl_s}$ , respectively. The zero-shot prediction is then derived by concatenating these similarities: $\\mathbf{Z}_{cls} = \\text{Concat}(\\text{Neg}, \\text{Pos})$ . Leveraging this knowledge, we apply cross-entropy loss between the zero-shot prediction $\\mathbf{Z}_{cls}$ (Eq. 5) and the zero-shot augmented class", + "bbox": [ + 89, + 582, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "predictions $\\mathbf{Y}_{cls}^{zs}$ (Equation 4), which substantially boosts our model's performance. Softmax temperature is defined by $\\tau$ with learnable parameters as $\\gamma_1, \\gamma_2$ .", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {c l s} ^ {z s} = \\gamma_ {1} \\mathbf {Y} _ {c l s} ^ {T} + \\gamma_ {2} \\mathbf {Z} _ {c l s} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 148, + 906, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} ^ {z s} = - \\mathbf {Z} _ {c l s} * \\log (\\operatorname {s o f t m a x} \\left(\\mathbf {Y} _ {c l s} ^ {z s} / \\tau\\right)) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 185, + 906, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Text-Only and Text-Augmented Few-shot Classification and Segmentation:", + "text_level": 1, + "bbox": [ + 513, + 210, + 905, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our text-augmented FS-CS approach, the support set includes both text and images with segmentation masks, while the query image has an associated target classification label and segmentation mask. Class predictions $\\mathbf{Y}_{cls}^{V}$ and $\\mathbf{Y}_{cls}^{T}$ , along with segmentation predictions $\\mathbf{Y}_{seg}^{V}$ and $\\mathbf{Y}_{seg}^{T}$ , are derived from foreground maps produced by vision-only and vision-text correlation, respectively. Classification refinement is achieved by merging vision-only predictions, $\\mathbf{Y}_{cls}^{V}$ , with vision-text predictions, $\\mathbf{Y}_{cls}^{T}$ , using the learnable scaling factors $\\alpha_{1}$ and $\\alpha_{2}$ . Segmentation predictions from the vision-only and vision-text models are combined by applying scaling factors $\\beta_{1}$ and $\\beta_{2}$ to integrate the results. Equations 6, 7, and 8 define the loss functions for classification, segmentation, and the total combined loss, respectively, 0.1 is set as default value of the scaling factor $\\lambda$ .", + "bbox": [ + 511, + 247, + 906, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c l s}} ^ {c} = - \\mathbf {Y} _ {\\text {c l s}} ^ {g t} \\log \\left(\\alpha_ {1} \\mathbf {Y} _ {\\text {c l s}} ^ {V} + \\alpha_ {2} \\mathbf {Y} _ {\\text {c l s}} ^ {T}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 487, + 906, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e g}} ^ {c} = - \\frac {1}{H W} \\sum_ {x \\in | H | | W |} \\mathbf {Y} _ {\\text {s e g}} ^ {g t} (x) \\log \\left(\\beta_ {1} \\mathbf {Y} _ {\\text {s e g}} ^ {V} (x) + \\beta_ {2} \\mathbf {Y} _ {\\text {s e g}} ^ {T} (x)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 531, + 905, + 577 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {c} = \\lambda \\mathcal {L} _ {c l s} ^ {c} + \\mathcal {L} _ {s e g} ^ {c} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 590, + 905, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With support samples containing only class labels, our text-only approach uses only textual data, while each query image is paired with its corresponding classification labels and segmentation masks. During training, we generate segmentation predictions, $\\mathbf{Y}_{seg}^{T}$ , and classification predictions, $\\mathbf{Y}_{cls}^{T}$ , based on the text input, following the previously outlined method. Additionally, zero-shot augmented class predictions $\\mathbf{Y}_{cls}^{zs}$ are generated and used as the final classification output for loss computation. With $\\mathbf{Y}_{cls}^{gt}$ and $\\mathbf{Y}_{seg}^{gt}$ as the ground truth classification and segmentation targets, the classification, segmentation, and total losses are calculated using cross-entropy, as outlined in Equations 9, 10, and 11. The scaling factor $\\lambda$ , is set to 0.1. Additionally, we incorporate a zero-shot knowledge distillation loss into the final loss calculation, scaled by $\\kappa$ . In the text-augmented setting, we choose to exclude the zero-shot knowledge distillation loss, as its inclusion does not yield any performance improvements. In this setting, the richer support vision information leads to superior classifications compared to the", + "bbox": [ + 511, + 612, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25361", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/001d247c509591cee4a941dcaed1478b4290c711b77d0c8bb466e9f781ea385d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
\\( 20^0 \\)\\( 20^1 \\)\\( 20^2 \\)\\( 20^3 \\)Avg\\( 20^0 \\)\\( 20^1 \\)\\( 20^2 \\)\\( 20^3 \\)Avg\\( 20^0 \\)\\( 20^1 \\)\\( 20^2 \\)\\( 20^3 \\)Avg\\( 20^0 \\)\\( 20^1 \\)\\( 20^{2} \\)\\( 20^{3} \\)Avg
CLIP-B/16 (Zero-Shot)65.865.861.564.064.3-----47.648.042.545.846.0----
PANet [31]64.366.568.067.966.725.524.725.724.725.242.549.953.647.848.524.925.023.321.4
PFENet [28]70.770.671.272.971.430.634.829.432.631.935.634.343.132.836.523.323.820.223.1
HSNet [18]74.777.278.577.677.036.234.332.934.034.357.762.467.162.662.528.929.630.329.3
ASNet [11]76.278.879.280.278.635.736.835.335.635.859.561.568.862.463.129.833.033.430.4
CST [12]77.682.083.180.580.836.338.337.840.738.361.066.068.260.564.034.737.136.836.3
CST(CLIP-B/16)*84.689.090.589.588.445.349.245.345.746.374.578.680.478.678.041.749.046.546.1
(Ours)Text Only86.890.892.191.690.346.651.249.550.049.376.581.685.285.882.345.649.549.452.2
(Ours)Text Aug88.493.593.393.492.150.455.452.353.652.980.088.388.387.686.048.356.052.752.3
", + "bbox": [ + 91, + 88, + 903, + 258 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/790a0531326fb150e56411b11eb11d7e14616806a833d1f62f513e172a2cd5c8.jpg", + "table_caption": [ + "Table 1. We compare our method with previous approaches on COCO-20 $^i$ . The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk $(^{*})$ ." + ], + "table_footnote": [], + "table_body": "
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
50515253Avg50515253Avg50515253Avg50515253Avg
CLIP-B/16 (Zero-Shot)82.278.979.276.479.2-----69.563.065.760.364.6-----
PANet [31]69.967.768.869.469.032.845.831.035.136.256.247.544.655.450.933.346.031.238.437.2
PFENet [28]69.882.468.177.974.638.354.735.143.843.022.561.740.339.541.031.147.330.832.235.3
HSNet [18]86.684.876.986.383.749.159.741.049.049.768.073.257.070.967.342.453.734.043.943.5
ASNet [11]84.989.679.086.284.951.761.543.352.852.368.576.258.670.068.348.558.336.348.347.8
CST [12]86.988.081.586.585.755.661.647.756.955.570.374.962.074.470.456.961.048.057.055.7
CST(CLIP-B/16)*92.493.080.590.189.069.169.348.063.462.483.987.665.179.679.066.369.547.462.361.4
(Ours)Text Only92.992.482.593.290.257.571.747.960.159.389.385.566.687.682.356.370.448.759.058.6
(Ours)Text Aug94.393.383.094.391.271.873.552.766.966.289.088.867.488.083.370.473.252.565.765.4
", + "bbox": [ + 91, + 311, + 906, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. On Pascal-5 $^i$ , we compare our method with prior approaches. The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk $(^*)$ .", + "bbox": [ + 89, + 492, + 906, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "limited textual cues used by CLIP, making zero-shot distillation unnecessary.", + "bbox": [ + 89, + 546, + 482, + 578 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} ^ {t} = - \\mathbf {Y} _ {c l s} ^ {g t} \\log \\mathbf {Y} _ {c l s} ^ {z s} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 589, + 482, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e g} ^ {t} = - \\frac {1}{H W} \\sum_ {x \\in | H | | W |} \\mathbf {Y} _ {s e g} ^ {g t} (x) \\log \\mathbf {Y} _ {s e g} ^ {T} (x) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 619, + 482, + 654 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {t} = \\lambda \\mathcal {L} _ {c l s} ^ {t} + \\mathcal {L} _ {s e g} ^ {t} + \\kappa \\mathcal {L} _ {c l s} ^ {z s} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 660, + 482, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 694, + 223, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct our experiments using CLIP ViT-B/16 [20] as the feature extractor in our model, while previous models' results were obtained with ResNet-50 [10]. The feature enrichment module uses dilation values of 3, 6, 10, and all experiments are optimized with the Adam optimizer at a learning rate of $1e^{-3}$ . To ensure an unbiased comparison, we replicate CST using CLIP-B/16 (Baseline) as the feature extractor, paired with their standard DINO (ViT-S/8) backbone. Evaluation adheres to established protocols [11, 12] and is performed across four folds on Pascal- $5^{i}$ and COCO- $20^{i}$ . Evaluation is conducted with N-way K-shot episodes, while training uses 1-way-1-shot episodes. Further details", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on experiments and ablations are provided in the supplementary material.", + "bbox": [ + 511, + 547, + 906, + 577 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c69987226a527286d38e05f9fd6aebf3ba310338d5ded20a121d49041a64f606.jpg", + "image_caption": [ + "Figure 4. Performance comparison of N-way 1-shot FS-CS on Pascal-5 $^i$ ." + ], + "image_footnote": [], + "bbox": [ + 513, + 590, + 903, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Few-shot classification and Segmentation Results", + "text_level": 1, + "bbox": [ + 511, + 756, + 903, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our Pascal- $5^{i}$ experiments (Table 2), our text-only approach generally outperforms prior methods in classification and segmentation, though it slightly trails CST with the CLIP-B/16 backbone on some metrics. Our text-augmented model, however, sets a new state-of-the-art, with notable improvements across 1-way-1-shot and 2-way-1-shot settings. In COCO- $20^{i}$ (Table 1), the text-only setting per", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25362", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/22581e76fba929e4dde3faa4e1474c821fe7b5efaecdc01e9600887cd9bb366e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
50515253Avg50515253Avg50515253Avg50515253Avg
CLIP-B/16 (Zero-Shot)82.278.979.276.479.2-----69.563.065.760.364.6-----
HSNet [18]84.584.860.885.378.920.023.516.216.619.170.867.036.570.161.111.023.415.417.017.7
ASNet [11]80.284.066.282.778.311.721.113.416.215.667.670.244.569.362.910.320.712.615.914.9
CST [12]84.082.270.882.679.935.838.928.929.233.274.367.249.067.864.635.736.026.829.131.9
CST(CLIP-B/16)*87.087.873.086.683.644.643.832.033.938.577.678.053.775.071.045.242.530.433.037.8
GPT-4Vision [17]93.580.384.487.386.437.345.534.235.638.2----------
(Ours)Text Only93.989.784.192.189.144.147.232.639.340.889.678.773.087.482.243.445.433.139.740.4
(Ours)Text Aug94.390.684.192.890.449.747.833.739.442.690.281.773.487.683.248.646.733.539.842.1
", + "bbox": [ + 91, + 88, + 906, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f29e6777f86bb55f091e0b24add031e4f2ecdad1dd6211a82ca878fa9f5fa364.jpg", + "table_caption": [ + "Table 3. We compare our method with previous approaches on Pascal-5i in a weakly supervised setting. The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk $(^{*})$ ." + ], + "table_footnote": [], + "table_body": "
Text perc.ArchCorr. SizeIter SpeedMemory
CSTDINO-S/8B × 2500 × 1441.77 it/s8.9G
CSTViT-B/16B × 625 × 1441.42 it/s4.7G
OursViT-B/16(B × 625 × 144) + (B × 625 × 77)1.7 it/s5.6G
", + "bbox": [ + 91, + 306, + 493, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "forms slightly better than in Pascal- $5^{i}$ , benefiting from the larger, diverse dataset, with further gains through text augmentation. These results highlight the complementary role of text data in FS-CS predictions, capturing nuances missed by vision-only inputs. Additional N-way 1-shot results on Pascal- $5^{i}$ are in Fig 4, with detailed values in the supplementary material.", + "bbox": [ + 89, + 462, + 483, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results of Weakly-Supervised FS-CS", + "text_level": 1, + "bbox": [ + 89, + 577, + 410, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the weakly-supervised setting (Table 3), we employ DINO (ViT-S/8) to produce pseudo masks through similarity and thresholding, following CST [12]. Our text-only configuration considerably outperforms CST (CLIP-B/16), highlighting the value of text data in FS-CS predictions. The text-augmented setting further improves performance by leveraging both text and image data, demonstrating the advantages of multi-modal information in weakly-supervised contexts. Notably, we also apply zero-shot distillation in this text-augmented setup.", + "bbox": [ + 89, + 599, + 483, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 762, + 243, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Effect on Runtime & Memory", + "text_level": 1, + "bbox": [ + 89, + 787, + 357, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Integrating CLIP has a minimal impact on runtime or memory usage, as shown in Table 4 (measured on an RTX-3060 with a batch size of 8). This efficiency is due to the reduction in query image tokens from 2500 to 625 $(\\downarrow 75\\%)$ , resulting from CLIP-B/16's larger patch size (16) compared to ViT-S/8 (patch size of 8). CST (ViT-S/8) requires more", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/384ef9731006f06c8ea4bf93c15f7ba9d513c87910c9237c285ed65644613f0d.jpg", + "table_caption": [ + "Table 4. Comparison of runtime and memory consumption across different models, including our text augmentation model (B = batch size)." + ], + "table_footnote": [], + "table_body": "
Text perc.1-way-1-shot
cls. ERseg. mIoU
0%89.862.4
25%90.562.8
50%91.063.0
75%91.163.3
100%91.266.2
", + "bbox": [ + 596, + 306, + 823, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. The impact of different percentages of text augmentation on Pascal- $5^{i}$ FS-CS.", + "bbox": [ + 511, + 419, + 906, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "resources due to its larger token count of 2500. In contrast, our text-augmented model, with dual CST blocks and correlation input sizes of $B \\times 625 \\times 144$ and $B \\times 625 \\times 77$ , delivers better performance with lower resource usage, while using slightly more resources than CST (CLIP-B/16).", + "bbox": [ + 511, + 474, + 906, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Effect of text on FS-CS", + "text_level": 1, + "bbox": [ + 511, + 561, + 725, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text data enhances our model's performance in classification and segmentation, with the amount of textual information influencing the effectiveness of our Text-Augmented model. We examine this effect by adjusting the percentage of text used in training, from $0\\%$ to $100\\%$ , and present the results in Fig 5. During evaluation, we align the level of text augmentation with that used in training. As shown in Table 5, performance improves as more text is included, highlighting the benefits of text augmentation for enhancing downstream performance.", + "bbox": [ + 511, + 584, + 906, + 737 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57e64f7add31abefce5de94fc1d47a23b1bbb78c7e491d63158f46bbafa56a7f.jpg", + "image_caption": [ + "Figure 5. Pascal- $5^{i}$ FS-CS for different percentages of text augmented predictions." + ], + "image_footnote": [], + "bbox": [ + 516, + 761, + 903, + 858 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25363", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5345f3dbe91c8a13e07b7f1d3bef3c038bce48f4d5dbf115c3597d9e8426b716.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ImageTextPascal-5²Coco-20²
cls. ERseg. mIoUcls. ERseg. mIoU
✓X89.862.489.147.5
X✓90.259.390.349.3
✓✓91.266.292.152.9
", + "bbox": [ + 125, + 88, + 449, + 161 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8a1241b057456684a6f08fd9c72c4a22d2c96b8e710881d52db0835aef435a52.jpg", + "table_caption": [ + "Table 6. The impact of combining text and images on FS-CS for COCO-20 $^i$ , Pascal-5 $^i$" + ], + "table_footnote": [], + "table_body": "
Pascal-5i\ncls. ER seg. mIoUCOCO-20i\ncls. ER seg. mIoU
No Feat. Enrinch.85.653.788.446.0
ASPP [3]83.355.387.646.3
Gran. Q84.856.188.347.6
Gran. QKV82.655.088.447.2
Attention83.256.488.247.4
Feat. Enrinch(Gran. KV)90.259.390.349.3
", + "bbox": [ + 91, + 227, + 493, + 345 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b3e4b0abaa639324f8b7736cd12012d89e85f2fb66c10cd5f5925ea2bcd69370.jpg", + "table_caption": [ + "Table 7. Feature enrichment module with different components for FS-CS on Pascal- $5^{i}$ and COCO- $20^{i}$ ." + ], + "table_footnote": [], + "table_body": "
ZS augment.CE. LossPascal-5²cls. ERCOCO-20²cls. ER
XX85.788.6
✓X84.089.7
X✓85.390.1
✓✓90.290.3
", + "bbox": [ + 143, + 407, + 431, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Zero-shot distillation component effect on FS-CS for Pascal- $5^{i}$ and COCO- $20^{i}$ ; grey indicates no zero-shot distillation.", + "bbox": [ + 89, + 506, + 482, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Comparison of Image, Text-only and Text-Augmented FS-CS", + "text_level": 1, + "bbox": [ + 89, + 573, + 482, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the impact of images, text, and their combination on FS-CS predictions, we conduct evaluations in a 1-way-1-shot setup on Pascal- $5^{i}$ and COCO- $20^{i}$ . The results are shown in Table 6. On Pascal- $5^{i}$ , text-only predictions fall behind image-based predictions, though they are comparable on COCO- $20^{i}$ . However, our text-augmented model, which combines text and images, outperforms unimodal predictions on both datasets. Figure 1 shows that the text-augmented model enhances segmentation by drawing on semantic cues from both modalities, whereas text-only predictions often misclassify object boundaries.", + "bbox": [ + 89, + 611, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Effect of the class, segmentation feature enrichment", + "text_level": 1, + "bbox": [ + 89, + 787, + 482, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the impact of a foreground map enhancer on FS-CS predictions for Pascal- $5^{i}$ and COCO- $20^{i}$ . We compare a text-only model enhanced with our feature enrichment module against models that apply ASPP to the query alone (Granular Q) or to the query, key, and value (Granular", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42fa745a3f2ab1b349c3604d7313bb7ad871f848c5a0e2e874cb7e3bc5da25d7.jpg", + "image_caption": [ + "Figure 6. FS-CS predictions enriched, where features capture both fine-grained and global details, outperform default predictions that often misclassify and yield poor segmentation maps; $\\checkmark$ and $x$ indicate class presence/absence." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 908, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "QKV), as well as those using standard attention or ASPP [3] directly on foreground maps. As shown in Table 7, our feature enrichment block, which incorporates ASPP to form granular KV, outperforms all other methods, including standard attention, ASPP alone, Granular Q, and Granular QKV. This improvement is due to the block's ability to capture global context and multi-scale features. Figure 6 further demonstrates how the feature enrichment block captures object details across scales, resulting in more accurate segmentation and classification.", + "bbox": [ + 511, + 321, + 906, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Effect of the zero shot knowledge distillation on FS-CS", + "text_level": 1, + "bbox": [ + 511, + 491, + 905, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8 presents the averaged results of the ablation study on zero-shot distillation components aimed at enhancing classification performance. The results underscore the importance of zero-shot augmented classification predictions and cross-entropy loss in improving accuracy. In the absence of cross-entropy loss, classification loss dominates, which undermines the goal of preserving CLIP's generalizability. Augmented predictions align effectively with CLIP's zero-shot classification, helping to minimize cross-entropy loss. As a result, combining both components is more effective than using either one individually.", + "bbox": [ + 511, + 532, + 908, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 720, + 633, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper demonstrates that text-only prompts can achieve high performance in Few-shot Classification and Segmentation (FS-CS), often outperforming prior methods. We further improved results by combining text with support image predictions (Text-Augmented), leveraging a multimodal approach. To address CLIP's generalization limits, we introduced zero-shot knowledge distillation and a feature enrichment module for capturing fine details and global semantics. Future work will explore additional modalities and improved multi-scale techniques to advance FS-CS.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25364", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. ArXiv, abs/2312.00878, 2023. 3", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 3", + "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2, 3, 5, 8", + "[4] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 2, 3", + "[5] Nanqing Dong and Eric P. Xing. Few-shot semantic segmentation with prototype learning. In British Machine Vision Conference, 2018. 1, 3", + "[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28:594–611, 2006. 1", + "[7] Michael Fink. Object classification from a single example utilizing class relevance pseudo-metrics. In Neural Information Processing Systems, 2004. 1", + "[8] Chelsea Finn, P. Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, 2017. 3", + "[9] Zixian Guo, Bowen Dong, Zhilong Ji, Jinfeng Bai, Yiwen Guo, and Wangmeng Zuo. Texts as images in prompt tuning for multi-label image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2808-2817, 2023. 3", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6", + "[11] Dahiyun Kang and Minsu Cho. Integrative few-shot learning for classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9979-9990, 2022. 1, 2, 3, 6, 7", + "[12] Dahyun Kang, Piotr Koniusz, Minsu Cho, and Naila Murray. Distilling self-supervised vision transformers for weakly-supervised few-shot classification & segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19627-19638, 2023. 2, 3, 4, 6, 7", + "[13] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 3", + "[14] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 2, 3", + "[15] Chen Liu, Yanwei Fu, Chengming Xu, Siqian Yang, Jilin Li, Chengjie Wang, and Li Zhang. Learning a few-shot embedding model with contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8635-8643, 2021. 3", + "[16] Puneet Mangla, Mayank Kumar Singh, Abhishek Sinha, Nupur Kumari, Vineeth N. Balasubramanian, and Balaji Krishnamurthy. Charting the right manifold: Manifold mixup for few-shot learning. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2207-2216, 2019. 3", + "[17] Tian Meng, Yang Tao, and Wuliang Yin. Few-shot classification & segmentation using large language models agent. arXiv preprint arXiv:2311.12065, 2023. 3, 7", + "[18] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3, 6, 7", + "[19] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chenyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23641-23651, 2023. 2", + "[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 5, 6", + "[21] Kate Rakelly, Evan Shelhamer, Trevor Darrell, Alexei A. Efros, and Sergey Levine. Conditional networks for few-shot semantic segmentation. In International Conference on Learning Representations, 2018. 3", + "[22] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. 2017. 1, 3", + "[23] Mennatullah Siam, Boris Oreshkin, and Martin Jagersand. Amp: Adaptive masked proxies for few-shot segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5248-5257, 2019. 1, 3", + "[24] Mennatullah Siam, Naren Doraiswamy, Boris N. Oreshkin, Hengshuai Yao, and Martin Jagersand. Weakly supervised few-shot object segmentation using co-attention with visual and semantic inputs. *IJCAI*, 2021. 2, 3", + "[25] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1, 3", + "[26] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 1, 3", + "[27] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In Computer" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25365", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 266-282. Springer, 2020. 3", + "[28] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44: 1050-1065, 2020. 2, 6", + "[29] Haohan Wang, L. Liu, Wuhao Zhang, Jiangning Zhang, Zhenye Gan, Yabiao Wang, Chengjie Wang, and Haoqian Wang. Iterative few-shot semantic segmentation from image label text. *IJCAI*, 2022. 2, 3", + "[30] Jin Wang, Bingfeng Zhang, Jian Pang, Honglong Chen, and Weifeng Liu. Rethinking prior information generation with clip for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3941-3951, 2024. 2, 3", + "[31] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 6", + "[32] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 3", + "[33] Han-Jia Ye, Hexiang Hu, De chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8805-8814, 2018. 1", + "[34] Muyang Yi, Quan Cui, Hao Wu, Cheng Yang, Osamu Yoshie, and Hongtao Lu. A simple framework for text-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7071-7080, 2023. 3", + "[35] Sukmin Yun, Seong Hyeon Park, Paul Hongsuck Seo, and Jinwoo Shin. Ifseg: Image-free semantic segmentation via vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2967-2977, 2023. 3", + "[36] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5212-5221, 2019. 1, 3", + "[37] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3", + "[38] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. 2022. 3", + "[39] Andrey Zhmoginov, Mark Sandler, and Maksym Vlademyrov. Hypertransformer: Model generation for supervised and semi-supervised few-shot learning. In International Conference on Machine Learning, pages 27075-27098. PMLR, 2022. 1, 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20061-20070, 2023. 3", + "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3", + "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3", + "[43] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175-11185, 2023. 2, 3" + ], + "bbox": [ + 514, + 90, + 903, + 357 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25366", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_model.json b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d1ba4ac0e17d34c6e19ba0db5807a1f5ab6587a --- /dev/null +++ b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.865, + 0.177 + ], + "angle": 0, + "content": "Text Augmented Correlation Transformer For Few-shot Classification & Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.203, + 0.954, + 0.257 + ], + "angle": 0, + "content": "Srinivasa Rao Nandam\\(^{1}\\) Sara Atito\\(^{1,2}\\) Zhenhua Feng\\(^{3}\\) Josef Kittler\\(^{2}\\) Muhammad Awais\\(^{1,2}\\) \n\\(^{1}\\)Surrey Institute for People-Centred AI (PAI), University of Surrey, UK \\(^{2}\\)CVSSP, University of Surrey, UK \\(^{3}\\)School of AI and Computer Science, Jiangnan University, China" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.26, + 0.928, + 0.275 + ], + "angle": 0, + "content": "{s.nandom, sara.atito, j.kittler, muhammad.awais}@surrey.ac.uk, fengzhenhua@jiangnan.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.342, + 0.486, + 0.812 + ], + "angle": 0, + "content": "Foundation models like CLIP and ALIGN have transformed few-shot and zero-shot vision applications by fusing visual and textual data, yet the integrative few-shot classification and segmentation (FS-CS) task primarily leverages visual cues, overlooking the potential of textual support. In FS-CS scenarios, ambiguous object boundaries and overlapping classes often hinder model performance, as limited visual data struggles to fully capture high-level semantics. To bridge this gap, we present a novel multi-modal FS-CS framework that integrates textual cues into support data, facilitating enhanced semantic disambiguation and fine-grained segmentation. Our approach first investigates the unique contributions of exclusive text-based support, using only class labels to achieve FS-CS. This strategy alone achieves performance competitive with vision-only methods on FS-CS tasks, underscoring the power of textual cues in few-shot learning. Building on this, we introduce a dual-modal prediction mechanism that synthesizes insights from both textual and visual support sets, yielding robust multimodal predictions. This integration significantly elevates FS-CS performance, with classification and segmentation improvements of \\(+3.7 / 6.6\\%\\) (1-way 1-shot) and \\(+8.0 / 6.5\\%\\) (2-way 1-shot) on COCO-20\\(^i\\), and \\(+2.2 / 3.8\\%\\) (1-way 1-shot) and \\(+4.3 / 4.0\\%\\) (2-way 1-shot) on Pascal-5\\(^i\\). Additionally, in weakly supervised FS-CS settings, our method surpasses visual-only benchmarks using textual support exclusively, further enhanced by our dual-modal predictions. By rethinking the role of text in FS-CS, our work establishes new benchmarks for multi-modal few-shot learning and demonstrates the efficacy of textual cues for improving model generalization and segmentation accuracy." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.844, + 0.222, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Few-shot learning [6, 7] has emerged as a significant area of research, focusing on building models that can effec" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.308, + 0.593, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.309, + 0.671, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.309, + 0.748, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.309, + 0.826, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.309, + 0.905, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.374, + 0.592, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.435, + 0.58, + 0.447 + ], + "angle": 0, + "content": "Qry.Img" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.375, + 0.671, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.435, + 0.661, + 0.447 + ], + "angle": 0, + "content": "Spt.Img 1" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.374, + 0.748, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.435, + 0.738, + 0.447 + ], + "angle": 0, + "content": "Img.Pred" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.374, + 0.826, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.435, + 0.822, + 0.447 + ], + "angle": 0, + "content": "T-Only Pred" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.374, + 0.904, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.834, + 0.435, + 0.901, + 0.447 + ], + "angle": 0, + "content": "T-Aug. Pred" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.529 + ], + "angle": 0, + "content": "Figure 1. Image predictions under-segment, text over-segments, and text-aug is most accurate in 2-way-1 shot segmentation using image, text, and text-aug support (shown: one support image and query image with ground truth; second support image not displayed)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.562, + 0.907, + 0.713 + ], + "angle": 0, + "content": "tively learn from a small set of examples, typically ranging from just 1 to 5 samples per class. This capability is essential in scenarios where data is scarce or expensive to obtain. Within this domain, Few-Shot Classification (FS-C) [25, 26, 33, 37, 39] and Few-Shot Segmentation (FS-S) [5, 22, 23, 36] have been widely studied, each focusing on classifying or segmenting images with minimal supervision. However, both FS-C and FS-S are constrained by specific limitations that impact their effectiveness in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.716, + 0.909, + 0.852 + ], + "angle": 0, + "content": "FS-C typically assumes each query image contains only one target class, limiting its applicability in complex scenes with multiple classes. On the other hand, FS-S, while supporting multi-class segmentation, requires that at least one target class appears in each query image, resulting in segmentation errors when no target classes are present. These rigid assumptions hinder the practical deployment of FS-C and FS-S in dynamic environments where real-world images seldom adhere to such constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address these limitations, Integrative Few-Shot Classification and Segmentation (FS-CS) combines the strengths of FS-C and FS-S into a unified framework [11]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.465, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.209, + 0.483, + 0.251 + ], + "angle": 0, + "content": "Figure 2. FS-CS with text-only using class label alone and text-augmented combining text with image predictions: The support set contains images with ground truth masks and class labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.279, + 0.483, + 0.37 + ], + "angle": 0, + "content": "FS-CS enables multi-class classification and segmentation even when no target classes appear in the query image, offering a more adaptable approach for real-world applications. This flexibility allows FS-CS to handle diverse scenarios, even though it may slightly underperform compared to traditional FS-C and FS-S on synthetic benchmarks [11]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.371, + 0.483, + 0.687 + ], + "angle": 0, + "content": "While previous FS-CS methods [11, 12] have made significant progress, they predominantly rely on visual data. Recent advances in few-shot and zero-shot vision applications, such as few-shot classification [32, 41, 42], few-shot segmentation (FSS) [24, 29], and zero-shot segmentation [4, 14, 43], have shown the benefits of incorporating textual information from models like CLIP [20]. Prior works in FSS have used textual cues from CLIP [29, 30] and Word2Vec embeddings [24], yet these studies did not explore the integration of rich text specifically for FS-CS or the potential of using an exclusive text support set. While zero-shot segmentation methods can generate segmentation directly from text, there are key distinctions: our method is trained in a few-shot episodic manner, whereas zero-shot methods train across the full dataset on all seen classes. Additionally, zero-shot segmentation methods are not inherently designed to produce classification predictions, and extending them for classification is non-trivial. Finally, our approach supports flexible configurations such as N-way 1-shot, 1-way K-shot, and N-way K-shot settings, which are not feasible with zero-shot segmentation methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.824 + ], + "angle": 0, + "content": "FS-CS predictions based solely on visual support sets often suffer from semantic ambiguity, leading to issues such as under-segmentation and misclassification in cases of unclear background/foreground boundaries or overlapping classes. Conversely, predictions based solely on text may lack fine-grained detail, resulting in over-segmentation (Figure 1). These challenges highlight the need for a multimodal FS-CS approach that can effectively combine visual and textual cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this work, we introduce a novel FS-CS approach that uniquely supports both text-only and multi-modal (text and image) support sets. Our method leverages a correlation-based approach, a technique shown to be effective in FS-CS tasks [11, 12], where foreground maps for classification and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "segmentation are generated by calculating correlations between query and support features. To accomplish this, we utilize the frozen CLIP [20] model, a multi-modal architecture that processes both image and text modalities through dedicated encoders, producing aligned embeddings for correlation. By using frozen CLIP instead of vision-only models like DINO [2], we maintain a comparable level of model complexity without adding learnable parameters, resulting in only marginal increases in memory consumption and runtime (Table 4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.907, + 0.563 + ], + "angle": 0, + "content": "To fully exploit these aligned vision and text embeddings, we propose the Text-Augmented Classification-Segmentation Transformer (TA-CST). TA-CST incorporates two transformer blocks: one for vision-only correlations (between query and support images) and another for vision-text correlations (between query images and support text). However, the feature maps generated by a correlation transformer block [12] alone have limited capacity for capturing both multi-scale and global semantics. To address this, we introduce a Feature-Enrichment Module that captures both granular and global semantics, enabling finer segmentation and more accurate classification. Unlike traditional FSS methods, which achieve multi-scale features through complex architectures [28] or multiple cross-attention layers [19, 30] and lack global semantic capture, our approach uses a single-layer attention block with granular Key and Value (KV) representations via atrous convolutions at varying scales [3]. This module captures multiscale information and global semantics, enhancing the FS-CS task by integrating both fine-grained details and broader context (Figure 6)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.747 + ], + "angle": 0, + "content": "In cases where the support set includes only text, our classification performance can be limited, particularly with smaller datasets like Pascal- \\(5^{i}\\). To address this, we propose a zero-shot knowledge distillation strategy that aligns our model's classification predictions with zero-shot predictions generated by CLIP. Specifically, we generate zero-shot predictions from CLIP using class-specific positive and negative prompts, such as \"This is a photo of {class-label}\" and \"This is not a photo of {class-label}.\" This approach enhances performance across datasets of varying sizes, preserving CLIP's generalization capabilities in text-only settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Fig 2 shows our FS-CS pipeline, which can handle generate predictions with exclusive text (text-only) or multimodal (text-augmented) support sets. Our experiments demonstrate that text-only support is a viable solution for FS-CS, achieving competitive performance with leading methods on the Pascal- \\(5^{i}\\) and \\(\\mathrm{COCO - }20^{i}\\) datasets. We achieve even higher performance with multi-modal support, demonstrating the effectiveness of our novel textaugmented FS-CS approach. For a fair comparison, we also replicate CST with CLIP-B/16 as our baseline." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25358" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "We provide a brief summary of our novel contributions below:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.124, + 0.483, + 0.184 + ], + "angle": 0, + "content": "- We first demonstrate that textual data alone can solve the FS-CS task, achieving competitive results, and show that combining text and image-based predictions further improves FS-CS performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.184, + 0.483, + 0.259 + ], + "angle": 0, + "content": "- We introduce a novel Feature Enrichment Module, which enhances feature maps with multi-scale and global semantics, resulting in better segmentation and classification predictions through a single attention layer with multi-scale Key and Value representations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.26, + 0.483, + 0.319 + ], + "angle": 0, + "content": "- In the text-only setting, we introduce a novel distillation strategy to maintain CLIP's classification generalization, aligning our model's performance with zero-shot predictions from CLIP." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.32, + 0.483, + 0.38 + ], + "angle": 0, + "content": "- We extend the correlation transformer used in prior methods to support multi-modal support sets, enabling fusion of classification and segmentation predictions from both visual and textual cues." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.124, + 0.483, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.393, + 0.233, + 0.408 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Few-shot classification (FS-C) & Segmentation (FS-S) Existing few-shot classification methods include metric learning [25, 26, 37, 40], meta-learning [8, 13, 39], and finetuning [15, 16, 27]. Metric learning constructs an embedding space that clusters similar samples while distancing dissimilar ones. Meta-learning emphasizes adapting to new tasks from limited samples, and fine-tuning enhances a well-initialized feature space. Few-shot segmentation methods adopt various strategies: some determine classifier weights per class [21, 22], others leverage prototype learning [5, 23], and some incorporate a decoder to produce segmentation outputs [18, 36, 38]. ASNet [11] introduced the combined task of few-shot classification and segmentation, using a ResNet backbone [10] for feature extraction. CST [12] builds on this in weakly supervised settings with a self-supervised DINO [2] backbone for correlation mapping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Classification & Segmentation with Text Recent works [9, 32, 41, 42] explore prompt tuning for CLIP [20] in few-shot multi-label classification. Although CLIP has been used for text-based supervised [34, 35], weakly supervised [14], zero-shot segmentation [1, 4, 43], as well as FSS [24, 29, 30], it has not been explored for FS-CS. Text-based FSS methods [24, 29, 30] rely on complex architectures with multiple components and cross-attention blocks, constrained by at least one target class in the query. These methods also fail to generate class predictions, unlike our simple FS-CS approach. While [30] uses positive/negative prompts with Grad-CAM priors for FS-S, we apply them for zero-shot distillation, preserving CLIP's generalizability. GPT4-Vision [17] tackles weakly supervised FS-CS with an LLM and lacks exclusive text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "support, unlike our compact CLIP-B/16, which enables text-based prediction generation and augmentation while addressing unique FS-CS challenges." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.15, + 0.65, + 0.168 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.175, + 0.905, + 0.342 + ], + "angle": 0, + "content": "This section presents our novel framework for few-shot classification and segmentation (FS-CS) that supports predictions from a multi-modal support set, consisting of both textual labels and support images. Using a meta-learning approach inspired by correlation-based methods [11, 12, 18], our model is trained on episodic tasks, each comprising a query image with a support set. We follow an \\(N\\)-way \\(K\\)-shot setup, where there are \\(K\\) examples of \\(N\\) classes in the support set. The model's objective is to generate a foreground mask for all target classes along with classification predictions for each, as specified in the support set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.507 + ], + "angle": 0, + "content": "Our Text-Augmented Classification-Segmentation Transformer (TA-CST) advances beyond traditional FS-CS methods [11, 12], which rely solely on visual data, thus limiting support set options to vision only. By incorporating textual cues from class labels, our approach enhances prediction quality by using the textual predictions in tandem with visual predictions. This is achieved through a dual correlation transformer, designed specifically to handle both visual and textual support sets effectively, but limited in its ability to capture both global semantics and multi-scale features crucial for FS-CS tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.905, + 0.657 + ], + "angle": 0, + "content": "To address these limitations, we introduce a feature enrichment module that augments the transformer's capacity to process varying spatial scales critical for segmentation tasks, including both few-shot and traditional segmentation [3, 5], while maintaining the global context essential for classification. The feature enrichment module achieves this balance through a single attention layer with a granular Key-Value (KV) structure, which embeds multi-scale inductive bias without losing the global semantics of the transformer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.905, + 0.765 + ], + "angle": 0, + "content": "During experiments, we observed a drop in classification accuracy when using exclusive text-based support, particularly with smaller datasets (e.g., Pascal-5\\(^i\\)). To address this and preserve the generalization capability of CLIP in classification, we propose a zero-shot knowledge distillation strategy. Figure 3 provides an architectural overview, and the following subsections elaborate on each component." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.905, + 0.803 + ], + "angle": 0, + "content": "3.1. Text augmented Classification-Segmentation Transformer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our model produces classification and segmentation predictions by generating foreground maps through correlations between the query images and the support set, which includes both images and text [11, 12, 18]. To accomplish this, we leverage CLIP's vision and text transformer encoders [20], taking advantage of its ability to produce in" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25359" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.118, + 0.365, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.334, + 0.294, + 0.347 + ], + "angle": 0, + "content": "Feature Extraction" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.095, + 0.675, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.223, + 0.675, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.338, + 0.576, + 0.351 + ], + "angle": 0, + "content": "Text Only" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.102, + 0.88, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.334, + 0.85, + 0.347 + ], + "angle": 0, + "content": "Feat. Enr. Block" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.365, + 0.908, + 0.409 + ], + "angle": 0, + "content": "Figure 3. TA-CST pipeline, where predictions are generated from correlations: Vision correlation comes from support/query images, and vision-text correlation from support text/query images. Text-only generates FS-CS predictions from the vision-text correlation, while text-augmented combines vision-only and vision-text FS-CS predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.483, + 0.495 + ], + "angle": 0, + "content": "herently aligned text and vision token embeddings. The required foreground maps for classification and segmentation are generated by correlating the query images with support images and text using our dual correlation transformer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.483, + 0.621 + ], + "angle": 0, + "content": "For a query image \\(\\mathbf{q} \\in \\mathbb{R}^{H \\times W \\times C}\\), where \\(H\\), \\(W\\), and \\(C\\) represent height, width, and channels, we obtain class output tokens \\(\\mathbf{V}_q^{\\mathrm{cls}} \\in \\mathbb{R}^D\\) and patch output tokens \\(\\mathbf{V}_q^{\\mathrm{pt}} \\in \\mathbb{R}^{D \\times N_q}\\) from the CLIP vision encoder. Similarly, for a support image \\(\\mathbf{s} \\in \\mathbb{R}^{H \\times W \\times C}\\), we derive class tokens \\(\\mathbf{V}_s^{\\mathrm{cls}} \\in \\mathbb{R}^D\\) and patch tokens \\(\\mathbf{V}_s^{\\mathrm{pt}} \\in \\mathbb{R}^{D \\times N_s}\\). For the support text, we use the CLIP text encoder to extract class tokens \\(\\mathbf{T}_s^{\\mathrm{cls}} \\in \\mathbb{R}^{D'}\\) and patch tokens \\(\\mathbf{T}_s^{\\mathrm{pt}} \\in \\mathbb{R}^{D' \\times N'}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.806 + ], + "angle": 0, + "content": "We calculate the vision-only correlation by taking the unnormalized dot product between the query patch tokens \\(\\mathbf{V}_q^{\\mathrm{pt}}\\) and the support class token \\(\\mathbf{V}_s^{\\mathrm{cls}}\\), noted as \\(v\\_ corr^{\\mathrm{cls}}\\), and then between \\(\\mathbf{V}_q^{\\mathrm{pt}}\\) and the support patch tokens \\(\\mathbf{V}_s^{\\mathrm{pt}}\\), referred to as \\(v\\_ corr^{\\mathrm{pt}}\\) [12]. For correlation between vision and text, we project the query's vision tokens \\(\\mathbf{V}_q^{\\mathrm{cls}}\\) and \\(\\mathbf{V}_q^{\\mathrm{pt}}\\) and the support's text tokens \\(\\mathbf{T}_s^{\\mathrm{cls}}\\) and \\(\\mathbf{T}_s^{\\mathrm{pt}}\\) into a common space using CLIP's projection layers. We then compute the unnormalized dot product between these tokens, yielding vision-text correlation for the EOS text token \\(t\\_ corr^{\\mathrm{cls}}\\) (from \\(\\mathbf{V}_q^{p,\\mathrm{pt}}\\) and \\(\\mathbf{T}_s^{p,\\mathrm{cls}}\\)) and for text patch tokens \\(t\\_ corr^{\\mathrm{pt}}\\) (from \\(\\mathbf{V}_q^{p,\\mathrm{pt}}\\) and \\(\\mathbf{T}_s^{p,\\mathrm{pt}}\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.483, + 0.883 + ], + "angle": 0, + "content": "Dual Correlation Transformer (Corr-T): Our dual correlation transformer module consists of two separate transformers, each with two layers and query pooling for efficient processing[12]; one transformer is for vision-only correlation, and another for vision-text correlation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In the vision-only pathway, we concatenate class and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.557 + ], + "angle": 0, + "content": "patch correlation outputs into \\( v_{\\mathrm{corr}} = \\{v_{\\mathrm{corr}}^{\\mathrm{cls}}, v_{\\mathrm{corr}}^{\\mathrm{pt}}\\} \\in \\mathbb{R}^{N_q \\times (N_s + 1) \\times ML} \\), which is processed by the vision-only transformer. Attention is applied with query downsampling using (4,4) and (3,3) kernels for each layer. Patch tokens are reshaped into spatial dimensions before downsampling, resulting in a final patch output dimension of \\( 1 \\times 1 \\). This produces foreground maps that are reshaped to \\( \\mathbb{R}^{H' \\times W' \\times C'} \\) for \\( \\mathbf{A}_{\\mathrm{seg}} \\in \\mathbb{R}^{N_q \\times C'} \\) and \\( \\mathbf{A}_{\\mathrm{cls}} \\in \\mathbb{R}^{N_q \\times C'} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.909, + 0.67 + ], + "angle": 0, + "content": "In the vision and text pathway, we concatenate vision and text class and patch correlation as \\( t_{\\mathrm{corr}} = \\{t_{\\mathrm{corr}}^{\\mathrm{cls}}, t_{\\mathrm{corr}}^{\\mathrm{pt}}\\} \\in \\mathbb{R}^{N_q \\times (N' + 1) \\times ML} \\). Later, attention is applied with (3,3) kernels for downsampling of query in both layers. This yields foreground maps \\( \\mathbf{B}_{\\mathrm{cls}} \\in \\mathbb{R}^{N_q \\times C'} \\) and \\( \\mathbf{B}_{\\mathrm{seg}} \\in \\mathbb{R}^{N_q \\times C'} \\), which are also reshaped to \\( \\mathbb{R}^{H' \\times W' \\times C'} \\) as the output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Feature Enrichment Module: We propose a feature enrichment module to enhance classification and segmentation feature map predictions by incorporating both global and fine-grained details. Global context is essential for classification and also supports segmentation, while multi-scale feature capture is particularly advantageous for segmentation, helping to improve classification by minimizing errors from an over-reliance on global context alone. To meet these requirements, we designed an attention-based module capable of capturing both global and multi-scale information. While attention mechanisms are effective at capturing global information, they generally lack an inherent bias toward capturing objects at multiple scales. To address this, we introduce a granular Key-Value (KV) structure, which provides the classic attention block with multi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25360" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.136 + ], + "angle": 0, + "content": "scale capabilities by capturing information at various scales. The granular KV is generated using various atrous convolutions [3]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.484, + 0.228 + ], + "angle": 0, + "content": "For an input \\(\\mathbf{X}\\) to the multi-scale encoding module, the granular key-value pairs (KV) are generated as per Equation 2, where GAP refers to global average pooling, with dilation rates \\(d1\\), \\(d2\\), and \\(d3\\). The final output of our feature enrichment module, Attn, where granular KV is used for key, value(Equation 3)." + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.242, + 0.479, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K V} _ {i n} = \\operatorname {C o n c a t} \\left\\{\\left(\\operatorname {C o n v} (1 \\times 1), \\operatorname {C o n v} (3 \\times 3, r = d _ {1}) \\right. \\right.,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.261, + 0.483, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\left. \\operatorname {C o n v} \\left(3 \\times 3, r = d _ {2}\\right), \\operatorname {C o n v} \\left(3 \\times 3, r = d _ {3}\\right), G A P \\right\\} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.302, + 0.483, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K V} _ {\\text {m u l t}} = \\operatorname {C o n v} \\left(\\left(1 \\times 1\\right), 5 C ^ {\\prime}, C ^ {\\prime}\\right) \\left(\\mathbf {K V} _ {\\text {i n}}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.335, + 0.483, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A t t n} = \\text {M u l t i H e a d A t t n} (Q, \\mathbf {K} _ {\\text {m u l t}}, \\mathbf {V} _ {\\text {m u l t}}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.4 + ], + "angle": 0, + "content": "Let \\(\\hat{\\mathbf{A}}_{cls}\\), \\(\\hat{\\mathbf{A}}_{seg}\\), \\(\\hat{\\mathbf{B}}_{cls}\\), and \\(\\hat{\\mathbf{B}}_{seg}\\) denote the outputs of the multi-scale encoding module, derived from the feature maps \\(\\mathbf{A}_{cls}\\), \\(\\mathbf{A}_{seg}\\), \\(\\mathbf{B}_{cls}\\), and \\(\\mathbf{B}_{seg}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.581 + ], + "angle": 0, + "content": "The enriched vision-only classification foreground map, \\(\\hat{\\mathbf{A}}_{cls}\\), and the enriched vision-text foreground map, \\(\\hat{\\mathbf{B}}_{cls}\\), are then passed to a classification head, which consists of two pointwise convolutions followed by global average pooling to produce the classification outputs \\(\\mathbf{Y}_{cls}^{V}\\) and \\(\\mathbf{Y}_{cls}^{T}\\). To refine the segmentation maps \\(\\hat{\\mathbf{A}}_{seg}\\) and \\(\\hat{\\mathbf{B}}_{seg}\\), these are first processed by a segmentation head. This head utilizes \\(3 \\times 3\\) convolutional layers, which are alternated with a bilinear interpolation layer. This combination allows the segmentation predictions to be upscaled to match the original image dimensions, producing \\(\\mathbf{Y}_{seg}^{V}\\) and \\(\\mathbf{Y}_{seg}^{T}\\), the final segmentation outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Zero Shot Knowledge Distillation: Classification performance in our method suffers when support samples are limited to text, especially on datasets like Pascal- \\(5^{i}\\), where the training set is smaller, amplifying the accuracy drop. To counter this, we utilize CLIP's zero-shot capability [20]. Unlike standard zero-shot classification, where text projections of known classes are compared to image projections, our task involves determining the presence of a specific support class. In zero-shot prediction, we use the positive prompt, 'This is a photo of {class-label}', followed by the negative prompt, 'This is not a photo of {class-label}.' Both prompts are processed through the CLIP text encoder and projection layer, producing the output embeddings \\(\\mathbf{T}_s^{+,p\\cdot cl_s}\\) and \\(\\mathbf{T}_s^{-,p\\cdot cl_s}\\) for the [EOS] token. The similarities, Pos and Neg, are then defined between the query image projection \\(\\mathbf{V}_q^{p\\cdot cl_s}\\) and the positive and negative projections, \\(\\mathbf{T}_s^{+,p\\cdot cl_s}\\) and \\(\\mathbf{T}_s^{-,p\\cdot cl_s}\\), respectively. The zero-shot prediction is then derived by concatenating these similarities: \\(\\mathbf{Z}_{cls} = \\text{Concat}(\\text{Neg}, \\text{Pos})\\). Leveraging this knowledge, we apply cross-entropy loss between the zero-shot prediction \\(\\mathbf{Z}_{cls}\\) (Eq. 5) and the zero-shot augmented class" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "predictions \\(\\mathbf{Y}_{cls}^{zs}\\) (Equation 4), which substantially boosts our model's performance. Softmax temperature is defined by \\(\\tau\\) with learnable parameters as \\(\\gamma_1, \\gamma_2\\)." + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.15, + 0.907, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {c l s} ^ {z s} = \\gamma_ {1} \\mathbf {Y} _ {c l s} ^ {T} + \\gamma_ {2} \\mathbf {Z} _ {c l s} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.186, + 0.907, + 0.203 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} ^ {z s} = - \\mathbf {Z} _ {c l s} * \\log (\\operatorname {s o f t m a x} \\left(\\mathbf {Y} _ {c l s} ^ {z s} / \\tau\\right)) \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.211, + 0.906, + 0.243 + ], + "angle": 0, + "content": "3.2. Text-Only and Text-Augmented Few-shot Classification and Segmentation:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.907, + 0.475 + ], + "angle": 0, + "content": "In our text-augmented FS-CS approach, the support set includes both text and images with segmentation masks, while the query image has an associated target classification label and segmentation mask. Class predictions \\(\\mathbf{Y}_{cls}^{V}\\) and \\(\\mathbf{Y}_{cls}^{T}\\), along with segmentation predictions \\(\\mathbf{Y}_{seg}^{V}\\) and \\(\\mathbf{Y}_{seg}^{T}\\), are derived from foreground maps produced by vision-only and vision-text correlation, respectively. Classification refinement is achieved by merging vision-only predictions, \\(\\mathbf{Y}_{cls}^{V}\\), with vision-text predictions, \\(\\mathbf{Y}_{cls}^{T}\\), using the learnable scaling factors \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\). Segmentation predictions from the vision-only and vision-text models are combined by applying scaling factors \\(\\beta_{1}\\) and \\(\\beta_{2}\\) to integrate the results. Equations 6, 7, and 8 define the loss functions for classification, segmentation, and the total combined loss, respectively, 0.1 is set as default value of the scaling factor \\(\\lambda\\)." + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.488, + 0.907, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c l s}} ^ {c} = - \\mathbf {Y} _ {\\text {c l s}} ^ {g t} \\log \\left(\\alpha_ {1} \\mathbf {Y} _ {\\text {c l s}} ^ {V} + \\alpha_ {2} \\mathbf {Y} _ {\\text {c l s}} ^ {T}\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.532, + 0.906, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e g}} ^ {c} = - \\frac {1}{H W} \\sum_ {x \\in | H | | W |} \\mathbf {Y} _ {\\text {s e g}} ^ {g t} (x) \\log \\left(\\beta_ {1} \\mathbf {Y} _ {\\text {s e g}} ^ {V} (x) + \\beta_ {2} \\mathbf {Y} _ {\\text {s e g}} ^ {T} (x)\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.592, + 0.906, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {c} = \\lambda \\mathcal {L} _ {c l s} ^ {c} + \\mathcal {L} _ {s e g} ^ {c} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.901 + ], + "angle": 0, + "content": "With support samples containing only class labels, our text-only approach uses only textual data, while each query image is paired with its corresponding classification labels and segmentation masks. During training, we generate segmentation predictions, \\(\\mathbf{Y}_{seg}^{T}\\), and classification predictions, \\(\\mathbf{Y}_{cls}^{T}\\), based on the text input, following the previously outlined method. Additionally, zero-shot augmented class predictions \\(\\mathbf{Y}_{cls}^{zs}\\) are generated and used as the final classification output for loss computation. With \\(\\mathbf{Y}_{cls}^{gt}\\) and \\(\\mathbf{Y}_{seg}^{gt}\\) as the ground truth classification and segmentation targets, the classification, segmentation, and total losses are calculated using cross-entropy, as outlined in Equations 9, 10, and 11. The scaling factor \\(\\lambda\\), is set to 0.1. Additionally, we incorporate a zero-shot knowledge distillation loss into the final loss calculation, scaled by \\(\\kappa\\). In the text-augmented setting, we choose to exclude the zero-shot knowledge distillation loss, as its inclusion does not yield any performance improvements. In this setting, the richer support vision information leads to superior classifications compared to the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25361" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.26 + ], + "angle": 0, + "content": "
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
\\( 20^0 \\)\\( 20^1 \\)\\( 20^2 \\)\\( 20^3 \\)Avg\\( 20^0 \\)\\( 20^1 \\)\\( 20^2 \\)\\( 20^3 \\)Avg\\( 20^0 \\)\\( 20^1 \\)\\( 20^2 \\)\\( 20^3 \\)Avg\\( 20^0 \\)\\( 20^1 \\)\\( 20^{2} \\)\\( 20^{3} \\)Avg
CLIP-B/16 (Zero-Shot)65.865.861.564.064.3-----47.648.042.545.846.0----
PANet [31]64.366.568.067.966.725.524.725.724.725.242.549.953.647.848.524.925.023.321.4
PFENet [28]70.770.671.272.971.430.634.829.432.631.935.634.343.132.836.523.323.820.223.1
HSNet [18]74.777.278.577.677.036.234.332.934.034.357.762.467.162.662.528.929.630.329.3
ASNet [11]76.278.879.280.278.635.736.835.335.635.859.561.568.862.463.129.833.033.430.4
CST [12]77.682.083.180.580.836.338.337.840.738.361.066.068.260.564.034.737.136.836.3
CST(CLIP-B/16)*84.689.090.589.588.445.349.245.345.746.374.578.680.478.678.041.749.046.546.1
(Ours)Text Only86.890.892.191.690.346.651.249.550.049.376.581.685.285.882.345.649.549.452.2
(Ours)Text Aug88.493.593.393.492.150.455.452.353.652.980.088.388.387.686.048.356.052.752.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.269, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Table 1. We compare our method with previous approaches on COCO-20\\(^i\\). The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk \\((^{*})\\)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.312, + 0.907, + 0.482 + ], + "angle": 0, + "content": "
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
50515253Avg50515253Avg50515253Avg50515253Avg
CLIP-B/16 (Zero-Shot)82.278.979.276.479.2-----69.563.065.760.364.6-----
PANet [31]69.967.768.869.469.032.845.831.035.136.256.247.544.655.450.933.346.031.238.437.2
PFENet [28]69.882.468.177.974.638.354.735.143.843.022.561.740.339.541.031.147.330.832.235.3
HSNet [18]86.684.876.986.383.749.159.741.049.049.768.073.257.070.967.342.453.734.043.943.5
ASNet [11]84.989.679.086.284.951.761.543.352.852.368.576.258.670.068.348.558.336.348.347.8
CST [12]86.988.081.586.585.755.661.647.756.955.570.374.962.074.470.456.961.048.057.055.7
CST(CLIP-B/16)*92.493.080.590.189.069.169.348.063.462.483.987.665.179.679.066.369.547.462.361.4
(Ours)Text Only92.992.482.593.290.257.571.747.960.159.389.385.566.687.682.356.370.448.759.058.6
(Ours)Text Aug94.393.383.094.391.271.873.552.766.966.289.088.867.488.083.370.473.252.565.765.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.493, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Table 2. On Pascal-5\\(^i\\), we compare our method with prior approaches. The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk \\((^*)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.579 + ], + "angle": 0, + "content": "limited textual cues used by CLIP, making zero-shot distillation unnecessary." + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.59, + 0.483, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} ^ {t} = - \\mathbf {Y} _ {c l s} ^ {g t} \\log \\mathbf {Y} _ {c l s} ^ {z s} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.621, + 0.483, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e g} ^ {t} = - \\frac {1}{H W} \\sum_ {x \\in | H | | W |} \\mathbf {Y} _ {s e g} ^ {g t} (x) \\log \\mathbf {Y} _ {s e g} ^ {T} (x) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.661, + 0.483, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {t} = \\lambda \\mathcal {L} _ {c l s} ^ {t} + \\mathcal {L} _ {s e g} ^ {t} + \\kappa \\mathcal {L} _ {c l s} ^ {z s} \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.224, + 0.712 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We conduct our experiments using CLIP ViT-B/16 [20] as the feature extractor in our model, while previous models' results were obtained with ResNet-50 [10]. The feature enrichment module uses dilation values of 3, 6, 10, and all experiments are optimized with the Adam optimizer at a learning rate of \\(1e^{-3}\\). To ensure an unbiased comparison, we replicate CST using CLIP-B/16 (Baseline) as the feature extractor, paired with their standard DINO (ViT-S/8) backbone. Evaluation adheres to established protocols [11, 12] and is performed across four folds on Pascal- \\(5^{i}\\) and COCO- \\(20^{i}\\). Evaluation is conducted with N-way K-shot episodes, while training uses 1-way-1-shot episodes. Further details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.548, + 0.907, + 0.578 + ], + "angle": 0, + "content": "on experiments and ablations are provided in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.592, + 0.905, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.699, + 0.907, + 0.726 + ], + "angle": 0, + "content": "Figure 4. Performance comparison of N-way 1-shot FS-CS on Pascal-5\\(^i\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.905, + 0.787 + ], + "angle": 0, + "content": "4.1. Few-shot classification and Segmentation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In our Pascal- \\(5^{i}\\) experiments (Table 2), our text-only approach generally outperforms prior methods in classification and segmentation, though it slightly trails CST with the CLIP-B/16 backbone on some metrics. Our text-augmented model, however, sets a new state-of-the-art, with notable improvements across 1-way-1-shot and 2-way-1-shot settings. In COCO- \\(20^{i}\\) (Table 1), the text-only setting per" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25362" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.246 + ], + "angle": 0, + "content": "
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
50515253Avg50515253Avg50515253Avg50515253Avg
CLIP-B/16 (Zero-Shot)82.278.979.276.479.2-----69.563.065.760.364.6-----
HSNet [18]84.584.860.885.378.920.023.516.216.619.170.867.036.570.161.111.023.415.417.017.7
ASNet [11]80.284.066.282.778.311.721.113.416.215.667.670.244.569.362.910.320.712.615.914.9
CST [12]84.082.270.882.679.935.838.928.929.233.274.367.249.067.864.635.736.026.829.131.9
CST(CLIP-B/16)*87.087.873.086.683.644.643.832.033.938.577.678.053.775.071.045.242.530.433.037.8
GPT-4Vision [17]93.580.384.487.386.437.345.534.235.638.2----------
(Ours)Text Only93.989.784.192.189.144.147.232.639.340.889.678.773.087.482.243.445.433.139.740.4
(Ours)Text Aug94.390.684.192.890.449.747.833.739.442.690.281.773.487.683.248.646.733.539.842.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.256, + 0.907, + 0.285 + ], + "angle": 0, + "content": "Table 3. We compare our method with previous approaches on Pascal-5i in a weakly supervised setting. The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk \\((^{*})\\)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.308, + 0.495, + 0.387 + ], + "angle": 0, + "content": "
Text perc.ArchCorr. SizeIter SpeedMemory
CSTDINO-S/8B × 2500 × 1441.77 it/s8.9G
CSTViT-B/16B × 625 × 1441.42 it/s4.7G
OursViT-B/16(B × 625 × 144) + (B × 625 × 77)1.7 it/s5.6G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.396, + 0.486, + 0.439 + ], + "angle": 0, + "content": "Table 4. Comparison of runtime and memory consumption across different models, including our text augmentation model (B = batch size)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.571 + ], + "angle": 0, + "content": "forms slightly better than in Pascal- \\(5^{i}\\), benefiting from the larger, diverse dataset, with further gains through text augmentation. These results highlight the complementary role of text data in FS-CS predictions, capturing nuances missed by vision-only inputs. Additional N-way 1-shot results on Pascal- \\(5^{i}\\) are in Fig 4, with detailed values in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.578, + 0.411, + 0.594 + ], + "angle": 0, + "content": "4.2. Results of Weakly-Supervised FS-CS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.751 + ], + "angle": 0, + "content": "In the weakly-supervised setting (Table 3), we employ DINO (ViT-S/8) to produce pseudo masks through similarity and thresholding, following CST [12]. Our text-only configuration considerably outperforms CST (CLIP-B/16), highlighting the value of text data in FS-CS predictions. The text-augmented setting further improves performance by leveraging both text and image data, demonstrating the advantages of multi-modal information in weakly-supervised contexts. Notably, we also apply zero-shot distillation in this text-augmented setup." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.245, + 0.781 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.358, + 0.805 + ], + "angle": 0, + "content": "5.1. Effect on Runtime & Memory" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Integrating CLIP has a minimal impact on runtime or memory usage, as shown in Table 4 (measured on an RTX-3060 with a batch size of 8). This efficiency is due to the reduction in query image tokens from 2500 to 625 \\((\\downarrow 75\\%)\\), resulting from CLIP-B/16's larger patch size (16) compared to ViT-S/8 (patch size of 8). CST (ViT-S/8) requires more" + }, + { + "type": "table", + "bbox": [ + 0.598, + 0.308, + 0.825, + 0.41 + ], + "angle": 0, + "content": "
Text perc.1-way-1-shot
cls. ERseg. mIoU
0%89.862.4
25%90.562.8
50%91.063.0
75%91.163.3
100%91.266.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.42, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Table 5. The impact of different percentages of text augmentation on Pascal- \\(5^{i}\\) FS-CS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.552 + ], + "angle": 0, + "content": "resources due to its larger token count of 2500. In contrast, our text-augmented model, with dual CST blocks and correlation input sizes of \\( B \\times 625 \\times 144 \\) and \\( B \\times 625 \\times 77 \\), delivers better performance with lower resource usage, while using slightly more resources than CST (CLIP-B/16)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.727, + 0.577 + ], + "angle": 0, + "content": "5.2. Effect of text on FS-CS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.738 + ], + "angle": 0, + "content": "Text data enhances our model's performance in classification and segmentation, with the amount of textual information influencing the effectiveness of our Text-Augmented model. We examine this effect by adjusting the percentage of text used in training, from \\(0\\%\\) to \\(100\\%\\), and present the results in Fig 5. During evaluation, we align the level of text augmentation with that used in training. As shown in Table 5, performance improves as more text is included, highlighting the benefits of text augmentation for enhancing downstream performance." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.762, + 0.904, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.869, + 0.907, + 0.898 + ], + "angle": 0, + "content": "Figure 5. Pascal- \\(5^{i}\\) FS-CS for different percentages of text augmented predictions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25363" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.45, + 0.162 + ], + "angle": 0, + "content": "
ImageTextPascal-5²Coco-20²
cls. ERseg. mIoUcls. ERseg. mIoU
✓X89.862.489.147.5
X✓90.259.390.349.3
✓✓91.266.292.152.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.173, + 0.483, + 0.201 + ], + "angle": 0, + "content": "Table 6. The impact of combining text and images on FS-CS for COCO-20\\(^i\\), Pascal-5\\(^i\\)" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.228, + 0.495, + 0.346 + ], + "angle": 0, + "content": "
Pascal-5i\ncls. ER seg. mIoUCOCO-20i\ncls. ER seg. mIoU
No Feat. Enrinch.85.653.788.446.0
ASPP [3]83.355.387.646.3
Gran. Q84.856.188.347.6
Gran. QKV82.655.088.447.2
Attention83.256.488.247.4
Feat. Enrinch(Gran. KV)90.259.390.349.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.385 + ], + "angle": 0, + "content": "Table 7. Feature enrichment module with different components for FS-CS on Pascal- \\(5^{i}\\) and COCO- \\(20^{i}\\)." + }, + { + "type": "table", + "bbox": [ + 0.145, + 0.409, + 0.432, + 0.498 + ], + "angle": 0, + "content": "
ZS augment.CE. LossPascal-5²cls. ERCOCO-20²cls. ER
XX85.788.6
✓X84.089.7
X✓85.390.1
✓✓90.290.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.537 + ], + "angle": 0, + "content": "Table 8. Zero-shot distillation component effect on FS-CS for Pascal- \\(5^{i}\\) and COCO- \\(20^{i}\\); grey indicates no zero-shot distillation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.574, + 0.483, + 0.607 + ], + "angle": 0, + "content": "5.3. Comparison of Image, Text-only and Text-Augmented FS-CS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.779 + ], + "angle": 0, + "content": "To assess the impact of images, text, and their combination on FS-CS predictions, we conduct evaluations in a 1-way-1-shot setup on Pascal- \\(5^{i}\\) and COCO- \\(20^{i}\\). The results are shown in Table 6. On Pascal- \\(5^{i}\\), text-only predictions fall behind image-based predictions, though they are comparable on COCO- \\(20^{i}\\). However, our text-augmented model, which combines text and images, outperforms unimodal predictions on both datasets. Figure 1 shows that the text-augmented model enhances segmentation by drawing on semantic cues from both modalities, whereas text-only predictions often misclassify object boundaries." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.483, + 0.82 + ], + "angle": 0, + "content": "5.4. Effect of the class, segmentation feature enrichment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We evaluate the impact of a foreground map enhancer on FS-CS predictions for Pascal- \\(5^{i}\\) and COCO- \\(20^{i}\\). We compare a text-only model enhanced with our feature enrichment module against models that apply ASPP to the query alone (Granular Q) or to the query, key, and value (Granular" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.909, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.222, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Figure 6. FS-CS predictions enriched, where features capture both fine-grained and global details, outperform default predictions that often misclassify and yield poor segmentation maps; \\(\\checkmark\\) and \\(x\\) indicate class presence/absence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.473 + ], + "angle": 0, + "content": "QKV), as well as those using standard attention or ASPP [3] directly on foreground maps. As shown in Table 7, our feature enrichment block, which incorporates ASPP to form granular KV, outperforms all other methods, including standard attention, ASPP alone, Granular Q, and Granular QKV. This improvement is due to the block's ability to capture global context and multi-scale features. Figure 6 further demonstrates how the feature enrichment block captures object details across scales, resulting in more accurate segmentation and classification." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.492, + 0.906, + 0.523 + ], + "angle": 0, + "content": "5.5. Effect of the zero shot knowledge distillation on FS-CS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.909, + 0.701 + ], + "angle": 0, + "content": "Table 8 presents the averaged results of the ablation study on zero-shot distillation components aimed at enhancing classification performance. The results underscore the importance of zero-shot augmented classification predictions and cross-entropy loss in improving accuracy. In the absence of cross-entropy loss, classification loss dominates, which undermines the goal of preserving CLIP's generalizability. Augmented predictions align effectively with CLIP's zero-shot classification, helping to minimize cross-entropy loss. As a result, combining both components is more effective than using either one individually." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.722, + 0.634, + 0.737 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper demonstrates that text-only prompts can achieve high performance in Few-shot Classification and Segmentation (FS-CS), often outperforming prior methods. We further improved results by combining text with support image predictions (Text-Augmented), leveraging a multimodal approach. To address CLIP's generalization limits, we introduced zero-shot knowledge distillation and a feature enrichment module for capturing fine details and global semantics. Future work will explore additional modalities and improved multi-scale techniques to advance FS-CS." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25364" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. ArXiv, abs/2312.00878, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[4] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[5] Nanqing Dong and Eric P. Xing. Few-shot semantic segmentation with prototype learning. In British Machine Vision Conference, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28:594–611, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.48 + ], + "angle": 0, + "content": "[7] Michael Fink. Object classification from a single example utilizing class relevance pseudo-metrics. In Neural Information Processing Systems, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[8] Chelsea Finn, P. Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[9] Zixian Guo, Bowen Dong, Zhilong Ji, Jinfeng Bai, Yiwen Guo, and Wangmeng Zuo. Texts as images in prompt tuning for multi-label image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2808-2817, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.594, + 0.483, + 0.649 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[11] Dahiyun Kang and Minsu Cho. Integrative few-shot learning for classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9979-9990, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[12] Dahyun Kang, Piotr Koniusz, Minsu Cho, and Naila Murray. Distilling self-supervised vision transformers for weakly-supervised few-shot classification & segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19627-19638, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[13] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[14] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[15] Chen Liu, Yanwei Fu, Chengming Xu, Siqian Yang, Jilin Li, Chengjie Wang, and Li Zhang. Learning a few-shot embedding model with contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8635-8643, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.907, + 0.286 + ], + "angle": 0, + "content": "[16] Puneet Mangla, Mayank Kumar Singh, Abhishek Sinha, Nupur Kumari, Vineeth N. Balasubramanian, and Balaji Krishnamurthy. Charting the right manifold: Manifold mixup for few-shot learning. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2207-2216, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.906, + 0.329 + ], + "angle": 0, + "content": "[17] Tian Meng, Yang Tao, and Wuliang Yin. Few-shot classification & segmentation using large language models agent. arXiv preprint arXiv:2311.12065, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.908, + 0.385 + ], + "angle": 0, + "content": "[18] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[19] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chenyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23641-23651, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[21] Kate Rakelly, Evan Shelhamer, Trevor Darrell, Alexei A. Efros, and Sergey Levine. Conditional networks for few-shot semantic segmentation. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[22] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.636, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[23] Mennatullah Siam, Boris Oreshkin, and Martin Jagersand. Amp: Adaptive masked proxies for few-shot segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5248-5257, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.906, + 0.748 + ], + "angle": 0, + "content": "[24] Mennatullah Siam, Naren Doraiswamy, Boris N. Oreshkin, Hengshuai Yao, and Martin Jagersand. Weakly supervised few-shot object segmentation using co-attention with visual and semantic inputs. *IJCAI*, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[25] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.859 + ], + "angle": 0, + "content": "[26] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[27] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In Computer" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25365" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 266-282. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[28] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44: 1050-1065, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[29] Haohan Wang, L. Liu, Wuhao Zhang, Jiangning Zhang, Zhenye Gan, Yabiao Wang, Chengjie Wang, and Haoqian Wang. Iterative few-shot semantic segmentation from image label text. *IJCAI*, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.483, + 0.329 + ], + "angle": 0, + "content": "[30] Jin Wang, Bingfeng Zhang, Jian Pang, Honglong Chen, and Weifeng Liu. Rethinking prior information generation with clip for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3941-3951, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[31] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[32] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[33] Han-Jia Ye, Hexiang Hu, De chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8805-8814, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[34] Muyang Yi, Quan Cui, Hao Wu, Cheng Yang, Osamu Yoshie, and Hongtao Lu. A simple framework for text-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7071-7080, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[35] Sukmin Yun, Seong Hyeon Park, Paul Hongsuck Seo, and Jinwoo Shin. Ifseg: Image-free semantic segmentation via vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2967-2977, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[36] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5212-5221, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[37] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[38] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[39] Andrey Zhmoginov, Mark Sandler, and Maksym Vlademyrov. Hypertransformer: Model generation for supervised and semi-supervised few-shot learning. In International Conference on Machine Learning, pages 27075-27098. PMLR, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[40] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20061-20070, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[43] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175-11185, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25366" + } + ] +] \ No newline at end of file diff --git a/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_origin.pdf b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d012fed098cdaafacc3f3a2d43fac77acbd26c53 --- /dev/null +++ b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/dedbc069-ae6a-4770-9618-33a3a7706ba4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c3453ab74ae526fa3526f5c983982485fc5ec30f94c7e83173a97d4e7098762 +size 848700 diff --git a/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/full.md b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a1831074c281bf8257605f8786affcb7afdfabf --- /dev/null +++ b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/full.md @@ -0,0 +1,329 @@ +# Text Augmented Correlation Transformer For Few-shot Classification & Segmentation + +Srinivasa Rao Nandam $^{1}$ Sara Atito $^{1,2}$ Zhenhua Feng $^{3}$ Josef Kittler $^{2}$ Muhammad Awais $^{1,2}$ $^{1}$ Surrey Institute for People-Centred AI (PAI), University of Surrey, UK $^{2}$ CVSSP, University of Surrey, UK $^{3}$ School of AI and Computer Science, Jiangnan University, China + +{s.nandom, sara.atito, j.kittler, muhammad.awais}@surrey.ac.uk, fengzhenhua@jiangnan.edu.cn + +# Abstract + +Foundation models like CLIP and ALIGN have transformed few-shot and zero-shot vision applications by fusing visual and textual data, yet the integrative few-shot classification and segmentation (FS-CS) task primarily leverages visual cues, overlooking the potential of textual support. In FS-CS scenarios, ambiguous object boundaries and overlapping classes often hinder model performance, as limited visual data struggles to fully capture high-level semantics. To bridge this gap, we present a novel multi-modal FS-CS framework that integrates textual cues into support data, facilitating enhanced semantic disambiguation and fine-grained segmentation. Our approach first investigates the unique contributions of exclusive text-based support, using only class labels to achieve FS-CS. This strategy alone achieves performance competitive with vision-only methods on FS-CS tasks, underscoring the power of textual cues in few-shot learning. Building on this, we introduce a dual-modal prediction mechanism that synthesizes insights from both textual and visual support sets, yielding robust multimodal predictions. This integration significantly elevates FS-CS performance, with classification and segmentation improvements of $+3.7 / 6.6\%$ (1-way 1-shot) and $+8.0 / 6.5\%$ (2-way 1-shot) on COCO-20 $^i$ , and $+2.2 / 3.8\%$ (1-way 1-shot) and $+4.3 / 4.0\%$ (2-way 1-shot) on Pascal-5 $^i$ . Additionally, in weakly supervised FS-CS settings, our method surpasses visual-only benchmarks using textual support exclusively, further enhanced by our dual-modal predictions. By rethinking the role of text in FS-CS, our work establishes new benchmarks for multi-modal few-shot learning and demonstrates the efficacy of textual cues for improving model generalization and segmentation accuracy. + +# 1. Introduction + +Few-shot learning [6, 7] has emerged as a significant area of research, focusing on building models that can effec + +![](images/766ded4397918efa3d95812c40c0274dd122999e87bd8b522d99ce81edd81acc.jpg) + +![](images/eb8e5e51d617430701170f82f8dbb45579f9e66465342ac083f7cfd93c13767f.jpg) + +![](images/b222aa094a3e872ab1a3e90746fbad4032cc88bb43f8afeb5ad38d0944cdaf50.jpg) + +![](images/e67a53afcba5103f630079bf64933410a69383320d83509ca4ef700dcc67d08a.jpg) + +![](images/4efdc0c96ff5acfbf05fd43cf7108f94d1902183279fbd1b7d4c6aabd5649419.jpg) + +![](images/7a9682150c57772e621689bf8cc494069cf1c6bf83a23e221c2a2143af516bdf.jpg) +Qry.Img + +![](images/67a215f0e241455f3b08dd02120e85bc13b90176c81c514ec91bdab230a5dd04.jpg) +Spt.Img 1 + +![](images/42f6541132c25a8f9687b8c6d29dfd7e8a0ae149448889825952d25093f7b930.jpg) +Img.Pred + +![](images/af54713ec10222adfe283440cd09f447f4abd538212e6126bf65122919fb2620.jpg) +T-Only Pred +Figure 1. Image predictions under-segment, text over-segments, and text-aug is most accurate in 2-way-1 shot segmentation using image, text, and text-aug support (shown: one support image and query image with ground truth; second support image not displayed). + +![](images/48a62db0cd327eebd904d9056227cf96a9a18df36fb7fd57436b23b182b3812f.jpg) +T-Aug. Pred + +tively learn from a small set of examples, typically ranging from just 1 to 5 samples per class. This capability is essential in scenarios where data is scarce or expensive to obtain. Within this domain, Few-Shot Classification (FS-C) [25, 26, 33, 37, 39] and Few-Shot Segmentation (FS-S) [5, 22, 23, 36] have been widely studied, each focusing on classifying or segmenting images with minimal supervision. However, both FS-C and FS-S are constrained by specific limitations that impact their effectiveness in real-world applications. + +FS-C typically assumes each query image contains only one target class, limiting its applicability in complex scenes with multiple classes. On the other hand, FS-S, while supporting multi-class segmentation, requires that at least one target class appears in each query image, resulting in segmentation errors when no target classes are present. These rigid assumptions hinder the practical deployment of FS-C and FS-S in dynamic environments where real-world images seldom adhere to such constraints. + +To address these limitations, Integrative Few-Shot Classification and Segmentation (FS-CS) combines the strengths of FS-C and FS-S into a unified framework [11]. + +![](images/77bfd4ff6705810c2cce1bbb1948cf255915d7b1966ae3b40219f5cc7e0d1594.jpg) +Figure 2. FS-CS with text-only using class label alone and text-augmented combining text with image predictions: The support set contains images with ground truth masks and class labels. + +FS-CS enables multi-class classification and segmentation even when no target classes appear in the query image, offering a more adaptable approach for real-world applications. This flexibility allows FS-CS to handle diverse scenarios, even though it may slightly underperform compared to traditional FS-C and FS-S on synthetic benchmarks [11]. + +While previous FS-CS methods [11, 12] have made significant progress, they predominantly rely on visual data. Recent advances in few-shot and zero-shot vision applications, such as few-shot classification [32, 41, 42], few-shot segmentation (FSS) [24, 29], and zero-shot segmentation [4, 14, 43], have shown the benefits of incorporating textual information from models like CLIP [20]. Prior works in FSS have used textual cues from CLIP [29, 30] and Word2Vec embeddings [24], yet these studies did not explore the integration of rich text specifically for FS-CS or the potential of using an exclusive text support set. While zero-shot segmentation methods can generate segmentation directly from text, there are key distinctions: our method is trained in a few-shot episodic manner, whereas zero-shot methods train across the full dataset on all seen classes. Additionally, zero-shot segmentation methods are not inherently designed to produce classification predictions, and extending them for classification is non-trivial. Finally, our approach supports flexible configurations such as N-way 1-shot, 1-way K-shot, and N-way K-shot settings, which are not feasible with zero-shot segmentation methods. + +FS-CS predictions based solely on visual support sets often suffer from semantic ambiguity, leading to issues such as under-segmentation and misclassification in cases of unclear background/foreground boundaries or overlapping classes. Conversely, predictions based solely on text may lack fine-grained detail, resulting in over-segmentation (Figure 1). These challenges highlight the need for a multimodal FS-CS approach that can effectively combine visual and textual cues. + +In this work, we introduce a novel FS-CS approach that uniquely supports both text-only and multi-modal (text and image) support sets. Our method leverages a correlation-based approach, a technique shown to be effective in FS-CS tasks [11, 12], where foreground maps for classification and + +segmentation are generated by calculating correlations between query and support features. To accomplish this, we utilize the frozen CLIP [20] model, a multi-modal architecture that processes both image and text modalities through dedicated encoders, producing aligned embeddings for correlation. By using frozen CLIP instead of vision-only models like DINO [2], we maintain a comparable level of model complexity without adding learnable parameters, resulting in only marginal increases in memory consumption and runtime (Table 4). + +To fully exploit these aligned vision and text embeddings, we propose the Text-Augmented Classification-Segmentation Transformer (TA-CST). TA-CST incorporates two transformer blocks: one for vision-only correlations (between query and support images) and another for vision-text correlations (between query images and support text). However, the feature maps generated by a correlation transformer block [12] alone have limited capacity for capturing both multi-scale and global semantics. To address this, we introduce a Feature-Enrichment Module that captures both granular and global semantics, enabling finer segmentation and more accurate classification. Unlike traditional FSS methods, which achieve multi-scale features through complex architectures [28] or multiple cross-attention layers [19, 30] and lack global semantic capture, our approach uses a single-layer attention block with granular Key and Value (KV) representations via atrous convolutions at varying scales [3]. This module captures multiscale information and global semantics, enhancing the FS-CS task by integrating both fine-grained details and broader context (Figure 6). + +In cases where the support set includes only text, our classification performance can be limited, particularly with smaller datasets like Pascal- $5^{i}$ . To address this, we propose a zero-shot knowledge distillation strategy that aligns our model's classification predictions with zero-shot predictions generated by CLIP. Specifically, we generate zero-shot predictions from CLIP using class-specific positive and negative prompts, such as "This is a photo of {class-label}" and "This is not a photo of {class-label}." This approach enhances performance across datasets of varying sizes, preserving CLIP's generalization capabilities in text-only settings. + +Fig 2 shows our FS-CS pipeline, which can handle generate predictions with exclusive text (text-only) or multimodal (text-augmented) support sets. Our experiments demonstrate that text-only support is a viable solution for FS-CS, achieving competitive performance with leading methods on the Pascal- $5^{i}$ and $\mathrm{COCO - }20^{i}$ datasets. We achieve even higher performance with multi-modal support, demonstrating the effectiveness of our novel textaugmented FS-CS approach. For a fair comparison, we also replicate CST with CLIP-B/16 as our baseline. + +We provide a brief summary of our novel contributions below: + +- We first demonstrate that textual data alone can solve the FS-CS task, achieving competitive results, and show that combining text and image-based predictions further improves FS-CS performance. +- We introduce a novel Feature Enrichment Module, which enhances feature maps with multi-scale and global semantics, resulting in better segmentation and classification predictions through a single attention layer with multi-scale Key and Value representations. +- In the text-only setting, we introduce a novel distillation strategy to maintain CLIP's classification generalization, aligning our model's performance with zero-shot predictions from CLIP. +- We extend the correlation transformer used in prior methods to support multi-modal support sets, enabling fusion of classification and segmentation predictions from both visual and textual cues. + +# 2. Related Work + +Few-shot classification (FS-C) & Segmentation (FS-S) Existing few-shot classification methods include metric learning [25, 26, 37, 40], meta-learning [8, 13, 39], and finetuning [15, 16, 27]. Metric learning constructs an embedding space that clusters similar samples while distancing dissimilar ones. Meta-learning emphasizes adapting to new tasks from limited samples, and fine-tuning enhances a well-initialized feature space. Few-shot segmentation methods adopt various strategies: some determine classifier weights per class [21, 22], others leverage prototype learning [5, 23], and some incorporate a decoder to produce segmentation outputs [18, 36, 38]. ASNet [11] introduced the combined task of few-shot classification and segmentation, using a ResNet backbone [10] for feature extraction. CST [12] builds on this in weakly supervised settings with a self-supervised DINO [2] backbone for correlation mapping. + +Classification & Segmentation with Text Recent works [9, 32, 41, 42] explore prompt tuning for CLIP [20] in few-shot multi-label classification. Although CLIP has been used for text-based supervised [34, 35], weakly supervised [14], zero-shot segmentation [1, 4, 43], as well as FSS [24, 29, 30], it has not been explored for FS-CS. Text-based FSS methods [24, 29, 30] rely on complex architectures with multiple components and cross-attention blocks, constrained by at least one target class in the query. These methods also fail to generate class predictions, unlike our simple FS-CS approach. While [30] uses positive/negative prompts with Grad-CAM priors for FS-S, we apply them for zero-shot distillation, preserving CLIP's generalizability. GPT4-Vision [17] tackles weakly supervised FS-CS with an LLM and lacks exclusive text + +support, unlike our compact CLIP-B/16, which enables text-based prediction generation and augmentation while addressing unique FS-CS challenges. + +# 3. Methodology + +This section presents our novel framework for few-shot classification and segmentation (FS-CS) that supports predictions from a multi-modal support set, consisting of both textual labels and support images. Using a meta-learning approach inspired by correlation-based methods [11, 12, 18], our model is trained on episodic tasks, each comprising a query image with a support set. We follow an $N$ -way $K$ -shot setup, where there are $K$ examples of $N$ classes in the support set. The model's objective is to generate a foreground mask for all target classes along with classification predictions for each, as specified in the support set. + +Our Text-Augmented Classification-Segmentation Transformer (TA-CST) advances beyond traditional FS-CS methods [11, 12], which rely solely on visual data, thus limiting support set options to vision only. By incorporating textual cues from class labels, our approach enhances prediction quality by using the textual predictions in tandem with visual predictions. This is achieved through a dual correlation transformer, designed specifically to handle both visual and textual support sets effectively, but limited in its ability to capture both global semantics and multi-scale features crucial for FS-CS tasks. + +To address these limitations, we introduce a feature enrichment module that augments the transformer's capacity to process varying spatial scales critical for segmentation tasks, including both few-shot and traditional segmentation [3, 5], while maintaining the global context essential for classification. The feature enrichment module achieves this balance through a single attention layer with a granular Key-Value (KV) structure, which embeds multi-scale inductive bias without losing the global semantics of the transformer. + +During experiments, we observed a drop in classification accuracy when using exclusive text-based support, particularly with smaller datasets (e.g., Pascal-5 $^i$ ). To address this and preserve the generalization capability of CLIP in classification, we propose a zero-shot knowledge distillation strategy. Figure 3 provides an architectural overview, and the following subsections elaborate on each component. + +# 3.1. Text augmented Classification-Segmentation Transformer + +Our model produces classification and segmentation predictions by generating foreground maps through correlations between the query images and the support set, which includes both images and text [11, 12, 18]. To accomplish this, we leverage CLIP's vision and text transformer encoders [20], taking advantage of its ability to produce in + +![](images/e203a461d5cf25c675d69dd0f157772285ad91ddccd10ef9d1a008a389846f4d.jpg) +Feature Extraction + +![](images/0f8856b400493818596d2b6c1a5a3d929003a24eb11e72a8aceaca0628314b94.jpg) + +![](images/2fcd8b2fff2a191d43251eb66fc771d38727716b7ed32aafe1aa0e836cd660ab.jpg) +Text Only +Figure 3. TA-CST pipeline, where predictions are generated from correlations: Vision correlation comes from support/query images, and vision-text correlation from support text/query images. Text-only generates FS-CS predictions from the vision-text correlation, while text-augmented combines vision-only and vision-text FS-CS predictions. + +![](images/9c4086eec0ddf2574617734fc6594dfbf3f5ae7f4afc04d2397d207e8268323b.jpg) +Feat. Enr. Block + +herently aligned text and vision token embeddings. The required foreground maps for classification and segmentation are generated by correlating the query images with support images and text using our dual correlation transformer. + +For a query image $\mathbf{q} \in \mathbb{R}^{H \times W \times C}$ , where $H$ , $W$ , and $C$ represent height, width, and channels, we obtain class output tokens $\mathbf{V}_q^{\mathrm{cls}} \in \mathbb{R}^D$ and patch output tokens $\mathbf{V}_q^{\mathrm{pt}} \in \mathbb{R}^{D \times N_q}$ from the CLIP vision encoder. Similarly, for a support image $\mathbf{s} \in \mathbb{R}^{H \times W \times C}$ , we derive class tokens $\mathbf{V}_s^{\mathrm{cls}} \in \mathbb{R}^D$ and patch tokens $\mathbf{V}_s^{\mathrm{pt}} \in \mathbb{R}^{D \times N_s}$ . For the support text, we use the CLIP text encoder to extract class tokens $\mathbf{T}_s^{\mathrm{cls}} \in \mathbb{R}^{D'}$ and patch tokens $\mathbf{T}_s^{\mathrm{pt}} \in \mathbb{R}^{D' \times N'}$ . + +We calculate the vision-only correlation by taking the unnormalized dot product between the query patch tokens $\mathbf{V}_q^{\mathrm{pt}}$ and the support class token $\mathbf{V}_s^{\mathrm{cls}}$ , noted as $v\_ corr^{\mathrm{cls}}$ , and then between $\mathbf{V}_q^{\mathrm{pt}}$ and the support patch tokens $\mathbf{V}_s^{\mathrm{pt}}$ , referred to as $v\_ corr^{\mathrm{pt}}$ [12]. For correlation between vision and text, we project the query's vision tokens $\mathbf{V}_q^{\mathrm{cls}}$ and $\mathbf{V}_q^{\mathrm{pt}}$ and the support's text tokens $\mathbf{T}_s^{\mathrm{cls}}$ and $\mathbf{T}_s^{\mathrm{pt}}$ into a common space using CLIP's projection layers. We then compute the unnormalized dot product between these tokens, yielding vision-text correlation for the EOS text token $t\_ corr^{\mathrm{cls}}$ (from $\mathbf{V}_q^{p,\mathrm{pt}}$ and $\mathbf{T}_s^{p,\mathrm{cls}}$ ) and for text patch tokens $t\_ corr^{\mathrm{pt}}$ (from $\mathbf{V}_q^{p,\mathrm{pt}}$ and $\mathbf{T}_s^{p,\mathrm{pt}}$ ). + +Dual Correlation Transformer (Corr-T): Our dual correlation transformer module consists of two separate transformers, each with two layers and query pooling for efficient processing[12]; one transformer is for vision-only correlation, and another for vision-text correlation. + +In the vision-only pathway, we concatenate class and + +patch correlation outputs into $v_{\mathrm{corr}} = \{v_{\mathrm{corr}}^{\mathrm{cls}}, v_{\mathrm{corr}}^{\mathrm{pt}}\} \in \mathbb{R}^{N_q \times (N_s + 1) \times ML}$ , which is processed by the vision-only transformer. Attention is applied with query downsampling using (4,4) and (3,3) kernels for each layer. Patch tokens are reshaped into spatial dimensions before downsampling, resulting in a final patch output dimension of $1 \times 1$ . This produces foreground maps that are reshaped to $\mathbb{R}^{H' \times W' \times C'}$ for $\mathbf{A}_{\mathrm{seg}} \in \mathbb{R}^{N_q \times C'}$ and $\mathbf{A}_{\mathrm{cls}} \in \mathbb{R}^{N_q \times C'}$ . + +In the vision and text pathway, we concatenate vision and text class and patch correlation as $t_{\mathrm{corr}} = \{t_{\mathrm{corr}}^{\mathrm{cls}}, t_{\mathrm{corr}}^{\mathrm{pt}}\} \in \mathbb{R}^{N_q \times (N' + 1) \times ML}$ . Later, attention is applied with (3,3) kernels for downsampling of query in both layers. This yields foreground maps $\mathbf{B}_{\mathrm{cls}} \in \mathbb{R}^{N_q \times C'}$ and $\mathbf{B}_{\mathrm{seg}} \in \mathbb{R}^{N_q \times C'}$ , which are also reshaped to $\mathbb{R}^{H' \times W' \times C'}$ as the output. + +Feature Enrichment Module: We propose a feature enrichment module to enhance classification and segmentation feature map predictions by incorporating both global and fine-grained details. Global context is essential for classification and also supports segmentation, while multi-scale feature capture is particularly advantageous for segmentation, helping to improve classification by minimizing errors from an over-reliance on global context alone. To meet these requirements, we designed an attention-based module capable of capturing both global and multi-scale information. While attention mechanisms are effective at capturing global information, they generally lack an inherent bias toward capturing objects at multiple scales. To address this, we introduce a granular Key-Value (KV) structure, which provides the classic attention block with multi + +scale capabilities by capturing information at various scales. The granular KV is generated using various atrous convolutions [3]. + +For an input $\mathbf{X}$ to the multi-scale encoding module, the granular key-value pairs (KV) are generated as per Equation 2, where GAP refers to global average pooling, with dilation rates $d1$ , $d2$ , and $d3$ . The final output of our feature enrichment module, Attn, where granular KV is used for key, value(Equation 3). + +$$ +\mathbf {K V} _ {i n} = \operatorname {C o n c a t} \left\{\left(\operatorname {C o n v} (1 \times 1), \operatorname {C o n v} (3 \times 3, r = d _ {1}) \right. \right., +$$ + +$$ +\left. \operatorname {C o n v} \left(3 \times 3, r = d _ {2}\right), \operatorname {C o n v} \left(3 \times 3, r = d _ {3}\right), G A P \right\} \tag {1} +$$ + +$$ +\mathbf {K V} _ {\text {m u l t}} = \operatorname {C o n v} \left(\left(1 \times 1\right), 5 C ^ {\prime}, C ^ {\prime}\right) \left(\mathbf {K V} _ {\text {i n}}\right) \tag {2} +$$ + +$$ +\mathbf {A t t n} = \text {M u l t i H e a d A t t n} (Q, \mathbf {K} _ {\text {m u l t}}, \mathbf {V} _ {\text {m u l t}}) \tag {3} +$$ + +Let $\hat{\mathbf{A}}_{cls}$ , $\hat{\mathbf{A}}_{seg}$ , $\hat{\mathbf{B}}_{cls}$ , and $\hat{\mathbf{B}}_{seg}$ denote the outputs of the multi-scale encoding module, derived from the feature maps $\mathbf{A}_{cls}$ , $\mathbf{A}_{seg}$ , $\mathbf{B}_{cls}$ , and $\mathbf{B}_{seg}$ , respectively. + +The enriched vision-only classification foreground map, $\hat{\mathbf{A}}_{cls}$ , and the enriched vision-text foreground map, $\hat{\mathbf{B}}_{cls}$ , are then passed to a classification head, which consists of two pointwise convolutions followed by global average pooling to produce the classification outputs $\mathbf{Y}_{cls}^{V}$ and $\mathbf{Y}_{cls}^{T}$ . To refine the segmentation maps $\hat{\mathbf{A}}_{seg}$ and $\hat{\mathbf{B}}_{seg}$ , these are first processed by a segmentation head. This head utilizes $3 \times 3$ convolutional layers, which are alternated with a bilinear interpolation layer. This combination allows the segmentation predictions to be upscaled to match the original image dimensions, producing $\mathbf{Y}_{seg}^{V}$ and $\mathbf{Y}_{seg}^{T}$ , the final segmentation outputs. + +Zero Shot Knowledge Distillation: Classification performance in our method suffers when support samples are limited to text, especially on datasets like Pascal- $5^{i}$ , where the training set is smaller, amplifying the accuracy drop. To counter this, we utilize CLIP's zero-shot capability [20]. Unlike standard zero-shot classification, where text projections of known classes are compared to image projections, our task involves determining the presence of a specific support class. In zero-shot prediction, we use the positive prompt, 'This is a photo of {class-label}', followed by the negative prompt, 'This is not a photo of {class-label}.' Both prompts are processed through the CLIP text encoder and projection layer, producing the output embeddings $\mathbf{T}_s^{+,p\cdot cl_s}$ and $\mathbf{T}_s^{-,p\cdot cl_s}$ for the [EOS] token. The similarities, Pos and Neg, are then defined between the query image projection $\mathbf{V}_q^{p\cdot cl_s}$ and the positive and negative projections, $\mathbf{T}_s^{+,p\cdot cl_s}$ and $\mathbf{T}_s^{-,p\cdot cl_s}$ , respectively. The zero-shot prediction is then derived by concatenating these similarities: $\mathbf{Z}_{cls} = \text{Concat}(\text{Neg}, \text{Pos})$ . Leveraging this knowledge, we apply cross-entropy loss between the zero-shot prediction $\mathbf{Z}_{cls}$ (Eq. 5) and the zero-shot augmented class + +predictions $\mathbf{Y}_{cls}^{zs}$ (Equation 4), which substantially boosts our model's performance. Softmax temperature is defined by $\tau$ with learnable parameters as $\gamma_1, \gamma_2$ . + +$$ +\mathbf {Y} _ {c l s} ^ {z s} = \gamma_ {1} \mathbf {Y} _ {c l s} ^ {T} + \gamma_ {2} \mathbf {Z} _ {c l s} \tag {4} +$$ + +$$ +\mathcal {L} _ {c l s} ^ {z s} = - \mathbf {Z} _ {c l s} * \log (\operatorname {s o f t m a x} \left(\mathbf {Y} _ {c l s} ^ {z s} / \tau\right)) \tag {5} +$$ + +# 3.2. Text-Only and Text-Augmented Few-shot Classification and Segmentation: + +In our text-augmented FS-CS approach, the support set includes both text and images with segmentation masks, while the query image has an associated target classification label and segmentation mask. Class predictions $\mathbf{Y}_{cls}^{V}$ and $\mathbf{Y}_{cls}^{T}$ , along with segmentation predictions $\mathbf{Y}_{seg}^{V}$ and $\mathbf{Y}_{seg}^{T}$ , are derived from foreground maps produced by vision-only and vision-text correlation, respectively. Classification refinement is achieved by merging vision-only predictions, $\mathbf{Y}_{cls}^{V}$ , with vision-text predictions, $\mathbf{Y}_{cls}^{T}$ , using the learnable scaling factors $\alpha_{1}$ and $\alpha_{2}$ . Segmentation predictions from the vision-only and vision-text models are combined by applying scaling factors $\beta_{1}$ and $\beta_{2}$ to integrate the results. Equations 6, 7, and 8 define the loss functions for classification, segmentation, and the total combined loss, respectively, 0.1 is set as default value of the scaling factor $\lambda$ . + +$$ +\mathcal {L} _ {\text {c l s}} ^ {c} = - \mathbf {Y} _ {\text {c l s}} ^ {g t} \log \left(\alpha_ {1} \mathbf {Y} _ {\text {c l s}} ^ {V} + \alpha_ {2} \mathbf {Y} _ {\text {c l s}} ^ {T}\right) \tag {6} +$$ + +$$ +\mathcal {L} _ {\text {s e g}} ^ {c} = - \frac {1}{H W} \sum_ {x \in | H | | W |} \mathbf {Y} _ {\text {s e g}} ^ {g t} (x) \log \left(\beta_ {1} \mathbf {Y} _ {\text {s e g}} ^ {V} (x) + \beta_ {2} \mathbf {Y} _ {\text {s e g}} ^ {T} (x)\right) \tag {7} +$$ + +$$ +\mathcal {L} ^ {c} = \lambda \mathcal {L} _ {c l s} ^ {c} + \mathcal {L} _ {s e g} ^ {c} \tag {8} +$$ + +With support samples containing only class labels, our text-only approach uses only textual data, while each query image is paired with its corresponding classification labels and segmentation masks. During training, we generate segmentation predictions, $\mathbf{Y}_{seg}^{T}$ , and classification predictions, $\mathbf{Y}_{cls}^{T}$ , based on the text input, following the previously outlined method. Additionally, zero-shot augmented class predictions $\mathbf{Y}_{cls}^{zs}$ are generated and used as the final classification output for loss computation. With $\mathbf{Y}_{cls}^{gt}$ and $\mathbf{Y}_{seg}^{gt}$ as the ground truth classification and segmentation targets, the classification, segmentation, and total losses are calculated using cross-entropy, as outlined in Equations 9, 10, and 11. The scaling factor $\lambda$ , is set to 0.1. Additionally, we incorporate a zero-shot knowledge distillation loss into the final loss calculation, scaled by $\kappa$ . In the text-augmented setting, we choose to exclude the zero-shot knowledge distillation loss, as its inclusion does not yield any performance improvements. In this setting, the richer support vision information leads to superior classifications compared to the + +
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
\( 20^0 \)\( 20^1 \)\( 20^2 \)\( 20^3 \)Avg\( 20^0 \)\( 20^1 \)\( 20^2 \)\( 20^3 \)Avg\( 20^0 \)\( 20^1 \)\( 20^2 \)\( 20^3 \)Avg\( 20^0 \)\( 20^1 \)\( 20^{2} \)\( 20^{3} \)Avg
CLIP-B/16 (Zero-Shot)65.865.861.564.064.3-----47.648.042.545.846.0----
PANet [31]64.366.568.067.966.725.524.725.724.725.242.549.953.647.848.524.925.023.321.4
PFENet [28]70.770.671.272.971.430.634.829.432.631.935.634.343.132.836.523.323.820.223.1
HSNet [18]74.777.278.577.677.036.234.332.934.034.357.762.467.162.662.528.929.630.329.3
ASNet [11]76.278.879.280.278.635.736.835.335.635.859.561.568.862.463.129.833.033.430.4
CST [12]77.682.083.180.580.836.338.337.840.738.361.066.068.260.564.034.737.136.836.3
CST(CLIP-B/16)*84.689.090.589.588.445.349.245.345.746.374.578.680.478.678.041.749.046.546.1
(Ours)Text Only86.890.892.191.690.346.651.249.550.049.376.581.685.285.882.345.649.549.452.2
(Ours)Text Aug88.493.593.393.492.150.455.452.353.652.980.088.388.387.686.048.356.052.752.3
+ +Table 1. We compare our method with previous approaches on COCO-20 $^i$ . The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk $(^{*})$ . + +
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
50515253Avg50515253Avg50515253Avg50515253Avg
CLIP-B/16 (Zero-Shot)82.278.979.276.479.2-----69.563.065.760.364.6-----
PANet [31]69.967.768.869.469.032.845.831.035.136.256.247.544.655.450.933.346.031.238.437.2
PFENet [28]69.882.468.177.974.638.354.735.143.843.022.561.740.339.541.031.147.330.832.235.3
HSNet [18]86.684.876.986.383.749.159.741.049.049.768.073.257.070.967.342.453.734.043.943.5
ASNet [11]84.989.679.086.284.951.761.543.352.852.368.576.258.670.068.348.558.336.348.347.8
CST [12]86.988.081.586.585.755.661.647.756.955.570.374.962.074.470.456.961.048.057.055.7
CST(CLIP-B/16)*92.493.080.590.189.069.169.348.063.462.483.987.665.179.679.066.369.547.462.361.4
(Ours)Text Only92.992.482.593.290.257.571.747.960.159.389.385.566.687.682.356.370.448.759.058.6
(Ours)Text Aug94.393.383.094.391.271.873.552.766.966.289.088.867.488.083.370.473.252.565.765.4
+ +Table 2. On Pascal-5 $^i$ , we compare our method with prior approaches. The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk $(^*)$ . + +limited textual cues used by CLIP, making zero-shot distillation unnecessary. + +$$ +\mathcal {L} _ {c l s} ^ {t} = - \mathbf {Y} _ {c l s} ^ {g t} \log \mathbf {Y} _ {c l s} ^ {z s} \tag {9} +$$ + +$$ +\mathcal {L} _ {s e g} ^ {t} = - \frac {1}{H W} \sum_ {x \in | H | | W |} \mathbf {Y} _ {s e g} ^ {g t} (x) \log \mathbf {Y} _ {s e g} ^ {T} (x) \tag {10} +$$ + +$$ +\mathcal {L} ^ {t} = \lambda \mathcal {L} _ {c l s} ^ {t} + \mathcal {L} _ {s e g} ^ {t} + \kappa \mathcal {L} _ {c l s} ^ {z s} \tag {11} +$$ + +# 4. Experiments + +We conduct our experiments using CLIP ViT-B/16 [20] as the feature extractor in our model, while previous models' results were obtained with ResNet-50 [10]. The feature enrichment module uses dilation values of 3, 6, 10, and all experiments are optimized with the Adam optimizer at a learning rate of $1e^{-3}$ . To ensure an unbiased comparison, we replicate CST using CLIP-B/16 (Baseline) as the feature extractor, paired with their standard DINO (ViT-S/8) backbone. Evaluation adheres to established protocols [11, 12] and is performed across four folds on Pascal- $5^{i}$ and COCO- $20^{i}$ . Evaluation is conducted with N-way K-shot episodes, while training uses 1-way-1-shot episodes. Further details + +on experiments and ablations are provided in the supplementary material. + +![](images/c69987226a527286d38e05f9fd6aebf3ba310338d5ded20a121d49041a64f606.jpg) +Figure 4. Performance comparison of N-way 1-shot FS-CS on Pascal-5 $^i$ . + +# 4.1. Few-shot classification and Segmentation Results + +In our Pascal- $5^{i}$ experiments (Table 2), our text-only approach generally outperforms prior methods in classification and segmentation, though it slightly trails CST with the CLIP-B/16 backbone on some metrics. Our text-augmented model, however, sets a new state-of-the-art, with notable improvements across 1-way-1-shot and 2-way-1-shot settings. In COCO- $20^{i}$ (Table 1), the text-only setting per + +
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
50515253Avg50515253Avg50515253Avg50515253Avg
CLIP-B/16 (Zero-Shot)82.278.979.276.479.2-----69.563.065.760.364.6-----
HSNet [18]84.584.860.885.378.920.023.516.216.619.170.867.036.570.161.111.023.415.417.017.7
ASNet [11]80.284.066.282.778.311.721.113.416.215.667.670.244.569.362.910.320.712.615.914.9
CST [12]84.082.270.882.679.935.838.928.929.233.274.367.249.067.864.635.736.026.829.131.9
CST(CLIP-B/16)*87.087.873.086.683.644.643.832.033.938.577.678.053.775.071.045.242.530.433.037.8
GPT-4Vision [17]93.580.384.487.386.437.345.534.235.638.2----------
(Ours)Text Only93.989.784.192.189.144.147.232.639.340.889.678.773.087.482.243.445.433.139.740.4
(Ours)Text Aug94.390.684.192.890.449.747.833.739.442.690.281.773.487.683.248.646.733.539.842.1
+ +Table 3. We compare our method with previous approaches on Pascal-5i in a weakly supervised setting. The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk $(^{*})$ . + +
Text perc.ArchCorr. SizeIter SpeedMemory
CSTDINO-S/8B × 2500 × 1441.77 it/s8.9G
CSTViT-B/16B × 625 × 1441.42 it/s4.7G
OursViT-B/16(B × 625 × 144) + (B × 625 × 77)1.7 it/s5.6G
+ +forms slightly better than in Pascal- $5^{i}$ , benefiting from the larger, diverse dataset, with further gains through text augmentation. These results highlight the complementary role of text data in FS-CS predictions, capturing nuances missed by vision-only inputs. Additional N-way 1-shot results on Pascal- $5^{i}$ are in Fig 4, with detailed values in the supplementary material. + +# 4.2. Results of Weakly-Supervised FS-CS + +In the weakly-supervised setting (Table 3), we employ DINO (ViT-S/8) to produce pseudo masks through similarity and thresholding, following CST [12]. Our text-only configuration considerably outperforms CST (CLIP-B/16), highlighting the value of text data in FS-CS predictions. The text-augmented setting further improves performance by leveraging both text and image data, demonstrating the advantages of multi-modal information in weakly-supervised contexts. Notably, we also apply zero-shot distillation in this text-augmented setup. + +# 5. Ablation Study + +# 5.1. Effect on Runtime & Memory + +Integrating CLIP has a minimal impact on runtime or memory usage, as shown in Table 4 (measured on an RTX-3060 with a batch size of 8). This efficiency is due to the reduction in query image tokens from 2500 to 625 $(\downarrow 75\%)$ , resulting from CLIP-B/16's larger patch size (16) compared to ViT-S/8 (patch size of 8). CST (ViT-S/8) requires more + +Table 4. Comparison of runtime and memory consumption across different models, including our text augmentation model (B = batch size). + +
Text perc.1-way-1-shot
cls. ERseg. mIoU
0%89.862.4
25%90.562.8
50%91.063.0
75%91.163.3
100%91.266.2
+ +Table 5. The impact of different percentages of text augmentation on Pascal- $5^{i}$ FS-CS. + +resources due to its larger token count of 2500. In contrast, our text-augmented model, with dual CST blocks and correlation input sizes of $B \times 625 \times 144$ and $B \times 625 \times 77$ , delivers better performance with lower resource usage, while using slightly more resources than CST (CLIP-B/16). + +# 5.2. Effect of text on FS-CS + +Text data enhances our model's performance in classification and segmentation, with the amount of textual information influencing the effectiveness of our Text-Augmented model. We examine this effect by adjusting the percentage of text used in training, from $0\%$ to $100\%$ , and present the results in Fig 5. During evaluation, we align the level of text augmentation with that used in training. As shown in Table 5, performance improves as more text is included, highlighting the benefits of text augmentation for enhancing downstream performance. + +![](images/57e64f7add31abefce5de94fc1d47a23b1bbb78c7e491d63158f46bbafa56a7f.jpg) +Figure 5. Pascal- $5^{i}$ FS-CS for different percentages of text augmented predictions. + +
ImageTextPascal-5²Coco-20²
cls. ERseg. mIoUcls. ERseg. mIoU
✓X89.862.489.147.5
X✓90.259.390.349.3
✓✓91.266.292.152.9
+ +Table 6. The impact of combining text and images on FS-CS for COCO-20 $^i$ , Pascal-5 $^i$ + +
Pascal-5i +cls. ER seg. mIoUCOCO-20i +cls. ER seg. mIoU
No Feat. Enrinch.85.653.788.446.0
ASPP [3]83.355.387.646.3
Gran. Q84.856.188.347.6
Gran. QKV82.655.088.447.2
Attention83.256.488.247.4
Feat. Enrinch(Gran. KV)90.259.390.349.3
+ +Table 7. Feature enrichment module with different components for FS-CS on Pascal- $5^{i}$ and COCO- $20^{i}$ . + +
ZS augment.CE. LossPascal-5²cls. ERCOCO-20²cls. ER
XX85.788.6
✓X84.089.7
X✓85.390.1
✓✓90.290.3
+ +Table 8. Zero-shot distillation component effect on FS-CS for Pascal- $5^{i}$ and COCO- $20^{i}$ ; grey indicates no zero-shot distillation. + +# 5.3. Comparison of Image, Text-only and Text-Augmented FS-CS + +To assess the impact of images, text, and their combination on FS-CS predictions, we conduct evaluations in a 1-way-1-shot setup on Pascal- $5^{i}$ and COCO- $20^{i}$ . The results are shown in Table 6. On Pascal- $5^{i}$ , text-only predictions fall behind image-based predictions, though they are comparable on COCO- $20^{i}$ . However, our text-augmented model, which combines text and images, outperforms unimodal predictions on both datasets. Figure 1 shows that the text-augmented model enhances segmentation by drawing on semantic cues from both modalities, whereas text-only predictions often misclassify object boundaries. + +# 5.4. Effect of the class, segmentation feature enrichment + +We evaluate the impact of a foreground map enhancer on FS-CS predictions for Pascal- $5^{i}$ and COCO- $20^{i}$ . We compare a text-only model enhanced with our feature enrichment module against models that apply ASPP to the query alone (Granular Q) or to the query, key, and value (Granular + +![](images/42fa745a3f2ab1b349c3604d7313bb7ad871f848c5a0e2e874cb7e3bc5da25d7.jpg) +Figure 6. FS-CS predictions enriched, where features capture both fine-grained and global details, outperform default predictions that often misclassify and yield poor segmentation maps; $\checkmark$ and $x$ indicate class presence/absence. + +QKV), as well as those using standard attention or ASPP [3] directly on foreground maps. As shown in Table 7, our feature enrichment block, which incorporates ASPP to form granular KV, outperforms all other methods, including standard attention, ASPP alone, Granular Q, and Granular QKV. This improvement is due to the block's ability to capture global context and multi-scale features. Figure 6 further demonstrates how the feature enrichment block captures object details across scales, resulting in more accurate segmentation and classification. + +# 5.5. Effect of the zero shot knowledge distillation on FS-CS + +Table 8 presents the averaged results of the ablation study on zero-shot distillation components aimed at enhancing classification performance. The results underscore the importance of zero-shot augmented classification predictions and cross-entropy loss in improving accuracy. In the absence of cross-entropy loss, classification loss dominates, which undermines the goal of preserving CLIP's generalizability. Augmented predictions align effectively with CLIP's zero-shot classification, helping to minimize cross-entropy loss. As a result, combining both components is more effective than using either one individually. + +# 6. Conclusion + +This paper demonstrates that text-only prompts can achieve high performance in Few-shot Classification and Segmentation (FS-CS), often outperforming prior methods. We further improved results by combining text with support image predictions (Text-Augmented), leveraging a multimodal approach. To address CLIP's generalization limits, we introduced zero-shot knowledge distillation and a feature enrichment module for capturing fine details and global semantics. Future work will explore additional modalities and improved multi-scale techniques to advance FS-CS. + +# References + +[1] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. ArXiv, abs/2312.00878, 2023. 3 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 3 +[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2, 3, 5, 8 +[4] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 2, 3 +[5] Nanqing Dong and Eric P. Xing. Few-shot semantic segmentation with prototype learning. In British Machine Vision Conference, 2018. 1, 3 +[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28:594–611, 2006. 1 +[7] Michael Fink. Object classification from a single example utilizing class relevance pseudo-metrics. In Neural Information Processing Systems, 2004. 1 +[8] Chelsea Finn, P. Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, 2017. 3 +[9] Zixian Guo, Bowen Dong, Zhilong Ji, Jinfeng Bai, Yiwen Guo, and Wangmeng Zuo. Texts as images in prompt tuning for multi-label image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2808-2817, 2023. 3 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6 +[11] Dahiyun Kang and Minsu Cho. Integrative few-shot learning for classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9979-9990, 2022. 1, 2, 3, 6, 7 +[12] Dahyun Kang, Piotr Koniusz, Minsu Cho, and Naila Murray. Distilling self-supervised vision transformers for weakly-supervised few-shot classification & segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19627-19638, 2023. 2, 3, 4, 6, 7 +[13] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 3 +[14] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly + +supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 2, 3 +[15] Chen Liu, Yanwei Fu, Chengming Xu, Siqian Yang, Jilin Li, Chengjie Wang, and Li Zhang. Learning a few-shot embedding model with contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8635-8643, 2021. 3 +[16] Puneet Mangla, Mayank Kumar Singh, Abhishek Sinha, Nupur Kumari, Vineeth N. Balasubramanian, and Balaji Krishnamurthy. Charting the right manifold: Manifold mixup for few-shot learning. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2207-2216, 2019. 3 +[17] Tian Meng, Yang Tao, and Wuliang Yin. Few-shot classification & segmentation using large language models agent. arXiv preprint arXiv:2311.12065, 2023. 3, 7 +[18] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3, 6, 7 +[19] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chenyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23641-23651, 2023. 2 +[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 5, 6 +[21] Kate Rakelly, Evan Shelhamer, Trevor Darrell, Alexei A. Efros, and Sergey Levine. Conditional networks for few-shot semantic segmentation. In International Conference on Learning Representations, 2018. 3 +[22] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. 2017. 1, 3 +[23] Mennatullah Siam, Boris Oreshkin, and Martin Jagersand. Amp: Adaptive masked proxies for few-shot segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5248-5257, 2019. 1, 3 +[24] Mennatullah Siam, Naren Doraiswamy, Boris N. Oreshkin, Hengshuai Yao, and Martin Jagersand. Weakly supervised few-shot object segmentation using co-attention with visual and semantic inputs. *IJCAI*, 2021. 2, 3 +[25] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1, 3 +[26] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 1, 3 +[27] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In Computer + +Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 266-282. Springer, 2020. 3 +[28] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44: 1050-1065, 2020. 2, 6 +[29] Haohan Wang, L. Liu, Wuhao Zhang, Jiangning Zhang, Zhenye Gan, Yabiao Wang, Chengjie Wang, and Haoqian Wang. Iterative few-shot semantic segmentation from image label text. *IJCAI*, 2022. 2, 3 +[30] Jin Wang, Bingfeng Zhang, Jian Pang, Honglong Chen, and Weifeng Liu. Rethinking prior information generation with clip for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3941-3951, 2024. 2, 3 +[31] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 6 +[32] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 3 +[33] Han-Jia Ye, Hexiang Hu, De chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8805-8814, 2018. 1 +[34] Muyang Yi, Quan Cui, Hao Wu, Cheng Yang, Osamu Yoshie, and Hongtao Lu. A simple framework for text-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7071-7080, 2023. 3 +[35] Sukmin Yun, Seong Hyeon Park, Paul Hongsuck Seo, and Jinwoo Shin. Ifseg: Image-free semantic segmentation via vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2967-2977, 2023. 3 +[36] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5212-5221, 2019. 1, 3 +[37] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3 +[38] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. 2022. 3 +[39] Andrey Zhmoginov, Mark Sandler, and Maksym Vlademyrov. Hypertransformer: Model generation for supervised and semi-supervised few-shot learning. In International Conference on Machine Learning, pages 27075-27098. PMLR, 2022. 1, 3 + +[40] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20061-20070, 2023. 3 +[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3 +[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3 +[43] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175-11185, 2023. 2, 3 \ No newline at end of file diff --git a/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/images.zip b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d6160d78bcaad5c07e99eb7f718e08c059ea55a7 --- /dev/null +++ b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da1973019873cf0005e55bc32b6a05358c9e07669387b8694086935d5a4ba6ba +size 708515 diff --git a/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/layout.json b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3d8d205cda52974350bd004014ace36a1cbb4ad5 --- /dev/null +++ b/2025/Text Augmented Correlation Transformer For Few-shot Classification & Segmentation/layout.json @@ -0,0 +1,9707 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 529, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 529, + 140 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 529, + 140 + ], + "type": "text", + "content": "Text Augmented Correlation Transformer For Few-shot Classification & Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "text", + "content": "Srinivasa Rao Nandam" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "text", + "content": " Sara Atito" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "text", + "content": " Zhenhua Feng" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "text", + "content": " Josef Kittler" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "text", + "content": " Muhammad Awais" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "text", + "content": "Surrey Institute for People-Centred AI (PAI), University of Surrey, UK " + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "text", + "content": "CVSSP, University of Surrey, UK " + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 583, + 203 + ], + "type": "text", + "content": "School of AI and Computer Science, Jiangnan University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 205, + 567, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 205, + 567, + 217 + ], + "spans": [ + { + "bbox": [ + 75, + 205, + 567, + 217 + ], + "type": "text", + "content": "{s.nandom, sara.atito, j.kittler, muhammad.awais}@surrey.ac.uk, fengzhenhua@jiangnan.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "text", + "content": "Foundation models like CLIP and ALIGN have transformed few-shot and zero-shot vision applications by fusing visual and textual data, yet the integrative few-shot classification and segmentation (FS-CS) task primarily leverages visual cues, overlooking the potential of textual support. In FS-CS scenarios, ambiguous object boundaries and overlapping classes often hinder model performance, as limited visual data struggles to fully capture high-level semantics. To bridge this gap, we present a novel multi-modal FS-CS framework that integrates textual cues into support data, facilitating enhanced semantic disambiguation and fine-grained segmentation. Our approach first investigates the unique contributions of exclusive text-based support, using only class labels to achieve FS-CS. This strategy alone achieves performance competitive with vision-only methods on FS-CS tasks, underscoring the power of textual cues in few-shot learning. Building on this, we introduce a dual-modal prediction mechanism that synthesizes insights from both textual and visual support sets, yielding robust multimodal predictions. This integration significantly elevates FS-CS performance, with classification and segmentation improvements of " + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "inline_equation", + "content": "+3.7 / 6.6\\%" + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "text", + "content": " (1-way 1-shot) and " + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "inline_equation", + "content": "+8.0 / 6.5\\%" + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "text", + "content": " (2-way 1-shot) on COCO-20" + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "inline_equation", + "content": "+2.2 / 3.8\\%" + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "text", + "content": " (1-way 1-shot) and " + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "inline_equation", + "content": "+4.3 / 4.0\\%" + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "text", + "content": " (2-way 1-shot) on Pascal-5" + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 53, + 270, + 297, + 643 + ], + "type": "text", + "content": ". Additionally, in weakly supervised FS-CS settings, our method surpasses visual-only benchmarks using textual support exclusively, further enhanced by our dual-modal predictions. By rethinking the role of text in FS-CS, our work establishes new benchmarks for multi-modal few-shot learning and demonstrates the efficacy of textual cues for improving model generalization and segmentation accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 668, + 135, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 668, + 135, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 668, + 135, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Few-shot learning [6, 7] has emerged as a significant area of research, focusing on building models that can effec" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 243, + 362, + 291 + ], + "blocks": [ + { + "bbox": [ + 315, + 243, + 362, + 291 + ], + "lines": [ + { + "bbox": [ + 315, + 243, + 362, + 291 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 362, + 291 + ], + "type": "image", + "image_path": "766ded4397918efa3d95812c40c0274dd122999e87bd8b522d99ce81edd81acc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 364, + 244, + 410, + 291 + ], + "blocks": [ + { + "bbox": [ + 364, + 244, + 410, + 291 + ], + "lines": [ + { + "bbox": [ + 364, + 244, + 410, + 291 + ], + "spans": [ + { + "bbox": [ + 364, + 244, + 410, + 291 + ], + "type": "image", + "image_path": "eb8e5e51d617430701170f82f8dbb45579f9e66465342ac083f7cfd93c13767f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 411, + 244, + 457, + 291 + ], + "blocks": [ + { + "bbox": [ + 411, + 244, + 457, + 291 + ], + "lines": [ + { + "bbox": [ + 411, + 244, + 457, + 291 + ], + "spans": [ + { + "bbox": [ + 411, + 244, + 457, + 291 + ], + "type": "image", + "image_path": "b222aa094a3e872ab1a3e90746fbad4032cc88bb43f8afeb5ad38d0944cdaf50.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 459, + 244, + 505, + 291 + ], + "blocks": [ + { + "bbox": [ + 459, + 244, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 459, + 244, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 459, + 244, + 505, + 291 + ], + "type": "image", + "image_path": "e67a53afcba5103f630079bf64933410a69383320d83509ca4ef700dcc67d08a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 506, + 244, + 553, + 291 + ], + "blocks": [ + { + "bbox": [ + 506, + 244, + 553, + 291 + ], + "lines": [ + { + "bbox": [ + 506, + 244, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 506, + 244, + 553, + 291 + ], + "type": "image", + "image_path": "4efdc0c96ff5acfbf05fd43cf7108f94d1902183279fbd1b7d4c6aabd5649419.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 296, + 362, + 342 + ], + "blocks": [ + { + "bbox": [ + 315, + 296, + 362, + 342 + ], + "lines": [ + { + "bbox": [ + 315, + 296, + 362, + 342 + ], + "spans": [ + { + "bbox": [ + 315, + 296, + 362, + 342 + ], + "type": "image", + "image_path": "7a9682150c57772e621689bf8cc494069cf1c6bf83a23e221c2a2143af516bdf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 344, + 354, + 354 + ], + "lines": [ + { + "bbox": [ + 323, + 344, + 354, + 354 + ], + "spans": [ + { + "bbox": [ + 323, + 344, + 354, + 354 + ], + "type": "text", + "content": "Qry.Img" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 364, + 297, + 410, + 342 + ], + "blocks": [ + { + "bbox": [ + 364, + 297, + 410, + 342 + ], + "lines": [ + { + "bbox": [ + 364, + 297, + 410, + 342 + ], + "spans": [ + { + "bbox": [ + 364, + 297, + 410, + 342 + ], + "type": "image", + "image_path": "67a215f0e241455f3b08dd02120e85bc13b90176c81c514ec91bdab230a5dd04.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 344, + 404, + 354 + ], + "lines": [ + { + "bbox": [ + 369, + 344, + 404, + 354 + ], + "spans": [ + { + "bbox": [ + 369, + 344, + 404, + 354 + ], + "type": "text", + "content": "Spt.Img 1" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 411, + 296, + 457, + 342 + ], + "blocks": [ + { + "bbox": [ + 411, + 296, + 457, + 342 + ], + "lines": [ + { + "bbox": [ + 411, + 296, + 457, + 342 + ], + "spans": [ + { + "bbox": [ + 411, + 296, + 457, + 342 + ], + "type": "image", + "image_path": "42f6541132c25a8f9687b8c6d29dfd7e8a0ae149448889825952d25093f7b930.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 344, + 451, + 354 + ], + "lines": [ + { + "bbox": [ + 417, + 344, + 451, + 354 + ], + "spans": [ + { + "bbox": [ + 417, + 344, + 451, + 354 + ], + "type": "text", + "content": "Img.Pred" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 459, + 296, + 505, + 342 + ], + "blocks": [ + { + "bbox": [ + 459, + 296, + 505, + 342 + ], + "lines": [ + { + "bbox": [ + 459, + 296, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 459, + 296, + 505, + 342 + ], + "type": "image", + "image_path": "af54713ec10222adfe283440cd09f447f4abd538212e6126bf65122919fb2620.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 344, + 503, + 354 + ], + "lines": [ + { + "bbox": [ + 462, + 344, + 503, + 354 + ], + "spans": [ + { + "bbox": [ + 462, + 344, + 503, + 354 + ], + "type": "text", + "content": "T-Only Pred" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 364, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 418 + ], + "type": "text", + "content": "Figure 1. Image predictions under-segment, text over-segments, and text-aug is most accurate in 2-way-1 shot segmentation using image, text, and text-aug support (shown: one support image and query image with ground truth; second support image not displayed)." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 507, + 296, + 553, + 342 + ], + "blocks": [ + { + "bbox": [ + 507, + 296, + 553, + 342 + ], + "lines": [ + { + "bbox": [ + 507, + 296, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 507, + 296, + 553, + 342 + ], + "type": "image", + "image_path": "48a62db0cd327eebd904d9056227cf96a9a18df36fb7fd57436b23b182b3812f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 510, + 344, + 551, + 354 + ], + "lines": [ + { + "bbox": [ + 510, + 344, + 551, + 354 + ], + "spans": [ + { + "bbox": [ + 510, + 344, + 551, + 354 + ], + "type": "text", + "content": "T-Aug. Pred" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 445, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 555, + 564 + ], + "type": "text", + "content": "tively learn from a small set of examples, typically ranging from just 1 to 5 samples per class. This capability is essential in scenarios where data is scarce or expensive to obtain. Within this domain, Few-Shot Classification (FS-C) [25, 26, 33, 37, 39] and Few-Shot Segmentation (FS-S) [5, 22, 23, 36] have been widely studied, each focusing on classifying or segmenting images with minimal supervision. However, both FS-C and FS-S are constrained by specific limitations that impact their effectiveness in real-world applications." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 567, + 556, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 567, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 312, + 567, + 556, + 674 + ], + "type": "text", + "content": "FS-C typically assumes each query image contains only one target class, limiting its applicability in complex scenes with multiple classes. On the other hand, FS-S, while supporting multi-class segmentation, requires that at least one target class appears in each query image, resulting in segmentation errors when no target classes are present. These rigid assumptions hinder the practical deployment of FS-C and FS-S in dynamic environments where real-world images seldom adhere to such constraints." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "To address these limitations, Integrative Few-Shot Classification and Segmentation (FS-CS) combines the strengths of FS-C and FS-S into a unified framework [11]." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25357" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 284, + 157 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 284, + 157 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 284, + 157 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 284, + 157 + ], + "type": "image", + "image_path": "77bfd4ff6705810c2cce1bbb1948cf255915d7b1966ae3b40219f5cc7e0d1594.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 165, + 295, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 165, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 295, + 198 + ], + "type": "text", + "content": "Figure 2. FS-CS with text-only using class label alone and text-augmented combining text with image predictions: The support set contains images with ground truth masks and class labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 220, + 295, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 293 + ], + "type": "text", + "content": "FS-CS enables multi-class classification and segmentation even when no target classes appear in the query image, offering a more adaptable approach for real-world applications. This flexibility allows FS-CS to handle diverse scenarios, even though it may slightly underperform compared to traditional FS-C and FS-S on synthetic benchmarks [11]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 293, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 295, + 544 + ], + "type": "text", + "content": "While previous FS-CS methods [11, 12] have made significant progress, they predominantly rely on visual data. Recent advances in few-shot and zero-shot vision applications, such as few-shot classification [32, 41, 42], few-shot segmentation (FSS) [24, 29], and zero-shot segmentation [4, 14, 43], have shown the benefits of incorporating textual information from models like CLIP [20]. Prior works in FSS have used textual cues from CLIP [29, 30] and Word2Vec embeddings [24], yet these studies did not explore the integration of rich text specifically for FS-CS or the potential of using an exclusive text support set. While zero-shot segmentation methods can generate segmentation directly from text, there are key distinctions: our method is trained in a few-shot episodic manner, whereas zero-shot methods train across the full dataset on all seen classes. Additionally, zero-shot segmentation methods are not inherently designed to produce classification predictions, and extending them for classification is non-trivial. Finally, our approach supports flexible configurations such as N-way 1-shot, 1-way K-shot, and N-way K-shot settings, which are not feasible with zero-shot segmentation methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "content": "FS-CS predictions based solely on visual support sets often suffer from semantic ambiguity, leading to issues such as under-segmentation and misclassification in cases of unclear background/foreground boundaries or overlapping classes. Conversely, predictions based solely on text may lack fine-grained detail, resulting in over-segmentation (Figure 1). These challenges highlight the need for a multimodal FS-CS approach that can effectively combine visual and textual cues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "In this work, we introduce a novel FS-CS approach that uniquely supports both text-only and multi-modal (text and image) support sets. Our method leverages a correlation-based approach, a technique shown to be effective in FS-CS tasks [11, 12], where foreground maps for classification and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "segmentation are generated by calculating correlations between query and support features. To accomplish this, we utilize the frozen CLIP [20] model, a multi-modal architecture that processes both image and text modalities through dedicated encoders, producing aligned embeddings for correlation. By using frozen CLIP instead of vision-only models like DINO [2], we maintain a comparable level of model complexity without adding learnable parameters, resulting in only marginal increases in memory consumption and runtime (Table 4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 194, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 445 + ], + "type": "text", + "content": "To fully exploit these aligned vision and text embeddings, we propose the Text-Augmented Classification-Segmentation Transformer (TA-CST). TA-CST incorporates two transformer blocks: one for vision-only correlations (between query and support images) and another for vision-text correlations (between query images and support text). However, the feature maps generated by a correlation transformer block [12] alone have limited capacity for capturing both multi-scale and global semantics. To address this, we introduce a Feature-Enrichment Module that captures both granular and global semantics, enabling finer segmentation and more accurate classification. Unlike traditional FSS methods, which achieve multi-scale features through complex architectures [28] or multiple cross-attention layers [19, 30] and lack global semantic capture, our approach uses a single-layer attention block with granular Key and Value (KV) representations via atrous convolutions at varying scales [3]. This module captures multiscale information and global semantics, enhancing the FS-CS task by integrating both fine-grained details and broader context (Figure 6)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 448, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 591 + ], + "type": "text", + "content": "In cases where the support set includes only text, our classification performance can be limited, particularly with smaller datasets like Pascal- " + }, + { + "bbox": [ + 313, + 448, + 556, + 591 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 313, + 448, + 556, + 591 + ], + "type": "text", + "content": ". To address this, we propose a zero-shot knowledge distillation strategy that aligns our model's classification predictions with zero-shot predictions generated by CLIP. Specifically, we generate zero-shot predictions from CLIP using class-specific positive and negative prompts, such as \"This is a photo of {class-label}\" and \"This is not a photo of {class-label}.\" This approach enhances performance across datasets of varying sizes, preserving CLIP's generalization capabilities in text-only settings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Fig 2 shows our FS-CS pipeline, which can handle generate predictions with exclusive text (text-only) or multimodal (text-augmented) support sets. Our experiments demonstrate that text-only support is a viable solution for FS-CS, achieving competitive performance with leading methods on the Pascal- " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO - }20^{i}" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " datasets. We achieve even higher performance with multi-modal support, demonstrating the effectiveness of our novel textaugmented FS-CS approach. For a fair comparison, we also replicate CST with CLIP-B/16 as our baseline." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25358" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "We provide a brief summary of our novel contributions below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 295, + 300 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 98, + 295, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 295, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 295, + 145 + ], + "type": "text", + "content": "- We first demonstrate that textual data alone can solve the FS-CS task, achieving competitive results, and show that combining text and image-based predictions further improves FS-CS performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 145, + 295, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 295, + 205 + ], + "type": "text", + "content": "- We introduce a novel Feature Enrichment Module, which enhances feature maps with multi-scale and global semantics, resulting in better segmentation and classification predictions through a single attention layer with multi-scale Key and Value representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 252 + ], + "type": "text", + "content": "- In the text-only setting, we introduce a novel distillation strategy to maintain CLIP's classification generalization, aligning our model's performance with zero-shot predictions from CLIP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 253, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 300 + ], + "type": "text", + "content": "- We extend the correlation transformer used in prior methods to support multi-modal support sets, enabling fusion of classification and segmentation predictions from both visual and textual cues." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 311, + 142, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 142, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 142, + 323 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 331, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 534 + ], + "type": "text", + "content": "Few-shot classification (FS-C) & Segmentation (FS-S) Existing few-shot classification methods include metric learning [25, 26, 37, 40], meta-learning [8, 13, 39], and finetuning [15, 16, 27]. Metric learning constructs an embedding space that clusters similar samples while distancing dissimilar ones. Meta-learning emphasizes adapting to new tasks from limited samples, and fine-tuning enhances a well-initialized feature space. Few-shot segmentation methods adopt various strategies: some determine classifier weights per class [21, 22], others leverage prototype learning [5, 23], and some incorporate a decoder to produce segmentation outputs [18, 36, 38]. ASNet [11] introduced the combined task of few-shot classification and segmentation, using a ResNet backbone [10] for feature extraction. CST [12] builds on this in weakly supervised settings with a self-supervised DINO [2] backbone for correlation mapping." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Classification & Segmentation with Text Recent works [9, 32, 41, 42] explore prompt tuning for CLIP [20] in few-shot multi-label classification. Although CLIP has been used for text-based supervised [34, 35], weakly supervised [14], zero-shot segmentation [1, 4, 43], as well as FSS [24, 29, 30], it has not been explored for FS-CS. Text-based FSS methods [24, 29, 30] rely on complex architectures with multiple components and cross-attention blocks, constrained by at least one target class in the query. These methods also fail to generate class predictions, unlike our simple FS-CS approach. While [30] uses positive/negative prompts with Grad-CAM priors for FS-S, we apply them for zero-shot distillation, preserving CLIP's generalizability. GPT4-Vision [17] tackles weakly supervised FS-CS with an LLM and lacks exclusive text" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "support, unlike our compact CLIP-B/16, which enables text-based prediction generation and augmentation while addressing unique FS-CS challenges." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 118, + 397, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 397, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 397, + 133 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "text", + "content": "This section presents our novel framework for few-shot classification and segmentation (FS-CS) that supports predictions from a multi-modal support set, consisting of both textual labels and support images. Using a meta-learning approach inspired by correlation-based methods [11, 12, 18], our model is trained on episodic tasks, each comprising a query image with a support set. We follow an " + }, + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "text", + "content": "-shot setup, where there are " + }, + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "text", + "content": " examples of " + }, + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 138, + 553, + 270 + ], + "type": "text", + "content": " classes in the support set. The model's objective is to generate a foreground mask for all target classes along with classification predictions for each, as specified in the support set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "content": "Our Text-Augmented Classification-Segmentation Transformer (TA-CST) advances beyond traditional FS-CS methods [11, 12], which rely solely on visual data, thus limiting support set options to vision only. By incorporating textual cues from class labels, our approach enhances prediction quality by using the textual predictions in tandem with visual predictions. This is achieved through a dual correlation transformer, designed specifically to handle both visual and textual support sets effectively, but limited in its ability to capture both global semantics and multi-scale features crucial for FS-CS tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 402, + 553, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 553, + 520 + ], + "type": "text", + "content": "To address these limitations, we introduce a feature enrichment module that augments the transformer's capacity to process varying spatial scales critical for segmentation tasks, including both few-shot and traditional segmentation [3, 5], while maintaining the global context essential for classification. The feature enrichment module achieves this balance through a single attention layer with a granular Key-Value (KV) structure, which embeds multi-scale inductive bias without losing the global semantics of the transformer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "text", + "content": "During experiments, we observed a drop in classification accuracy when using exclusive text-based support, particularly with smaller datasets (e.g., Pascal-5" + }, + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "text", + "content": "). To address this and preserve the generalization capability of CLIP in classification, we propose a zero-shot knowledge distillation strategy. Figure 3 provides an architectural overview, and the following subsections elaborate on each component." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 612, + 553, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 553, + 635 + ], + "type": "text", + "content": "3.1. Text augmented Classification-Segmentation Transformer" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "Our model produces classification and segmentation predictions by generating foreground maps through correlations between the query images and the support set, which includes both images and text [11, 12, 18]. To accomplish this, we leverage CLIP's vision and text transformer encoders [20], taking advantage of its ability to produce in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25359" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 93, + 223, + 256 + ], + "blocks": [ + { + "bbox": [ + 61, + 93, + 223, + 256 + ], + "lines": [ + { + "bbox": [ + 61, + 93, + 223, + 256 + ], + "spans": [ + { + "bbox": [ + 61, + 93, + 223, + 256 + ], + "type": "image", + "image_path": "e203a461d5cf25c675d69dd0f157772285ad91ddccd10ef9d1a008a389846f4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 264, + 179, + 274 + ], + "lines": [ + { + "bbox": [ + 102, + 264, + 179, + 274 + ], + "spans": [ + { + "bbox": [ + 102, + 264, + 179, + 274 + ], + "type": "text", + "content": "Feature Extraction" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 242, + 75, + 413, + 175 + ], + "blocks": [ + { + "bbox": [ + 242, + 75, + 413, + 175 + ], + "lines": [ + { + "bbox": [ + 242, + 75, + 413, + 175 + ], + "spans": [ + { + "bbox": [ + 242, + 75, + 413, + 175 + ], + "type": "image", + "image_path": "0f8856b400493818596d2b6c1a5a3d929003a24eb11e72a8aceaca0628314b94.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 176, + 413, + 266 + ], + "blocks": [ + { + "bbox": [ + 242, + 176, + 413, + 266 + ], + "lines": [ + { + "bbox": [ + 242, + 176, + 413, + 266 + ], + "spans": [ + { + "bbox": [ + 242, + 176, + 413, + 266 + ], + "type": "image", + "image_path": "2fcd8b2fff2a191d43251eb66fc771d38727716b7ed32aafe1aa0e836cd660ab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 267, + 352, + 277 + ], + "lines": [ + { + "bbox": [ + 312, + 267, + 352, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 267, + 352, + 277 + ], + "type": "text", + "content": "Text Only" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 289, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 555, + 323 + ], + "type": "text", + "content": "Figure 3. TA-CST pipeline, where predictions are generated from correlations: Vision correlation comes from support/query images, and vision-text correlation from support text/query images. Text-only generates FS-CS predictions from the vision-text correlation, while text-augmented combines vision-only and vision-text FS-CS predictions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 433, + 80, + 538, + 262 + ], + "blocks": [ + { + "bbox": [ + 433, + 80, + 538, + 262 + ], + "lines": [ + { + "bbox": [ + 433, + 80, + 538, + 262 + ], + "spans": [ + { + "bbox": [ + 433, + 80, + 538, + 262 + ], + "type": "image", + "image_path": "9c4086eec0ddf2574617734fc6594dfbf3f5ae7f4afc04d2397d207e8268323b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 264, + 520, + 274 + ], + "lines": [ + { + "bbox": [ + 454, + 264, + 520, + 274 + ], + "spans": [ + { + "bbox": [ + 454, + 264, + 520, + 274 + ], + "type": "text", + "content": "Feat. Enr. Block" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 343, + 295, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 295, + 392 + ], + "type": "text", + "content": "herently aligned text and vision token embeddings. The required foreground maps for classification and segmentation are generated by correlating the query images with support images and text using our dual correlation transformer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": "For a query image " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": " represent height, width, and channels, we obtain class output tokens " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^{\\mathrm{cls}} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": " and patch output tokens " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^{\\mathrm{pt}} \\in \\mathbb{R}^{D \\times N_q}" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": " from the CLIP vision encoder. Similarly, for a support image " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": ", we derive class tokens " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s^{\\mathrm{cls}} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": " and patch tokens " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s^{\\mathrm{pt}} \\in \\mathbb{R}^{D \\times N_s}" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": ". For the support text, we use the CLIP text encoder to extract class tokens " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{\\mathrm{cls}} \\in \\mathbb{R}^{D'}" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": " and patch tokens " + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{\\mathrm{pt}} \\in \\mathbb{R}^{D' \\times N'}" + }, + { + "bbox": [ + 55, + 392, + 295, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": "We calculate the vision-only correlation by taking the unnormalized dot product between the query patch tokens " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^{\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " and the support class token " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": ", noted as " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "v\\_ corr^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": ", and then between " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^{\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " and the support patch tokens " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s^{\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": ", referred to as " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "v\\_ corr^{\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " [12]. For correlation between vision and text, we project the query's vision tokens " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^{\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " and the support's text tokens " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " into a common space using CLIP's projection layers. We then compute the unnormalized dot product between these tokens, yielding vision-text correlation for the EOS text token " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "t\\_ corr^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^{p,\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{p,\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": ") and for text patch tokens " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "t\\_ corr^{\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^{p,\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{p,\\mathrm{pt}}" + }, + { + "bbox": [ + 55, + 492, + 295, + 638 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 639, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 295, + 699 + ], + "type": "text", + "content": "Dual Correlation Transformer (Corr-T): Our dual correlation transformer module consists of two separate transformers, each with two layers and query pooling for efficient processing[12]; one transformer is for vision-only correlation, and another for vision-text correlation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "In the vision-only pathway, we concatenate class and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "text", + "content": "patch correlation outputs into " + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{corr}} = \\{v_{\\mathrm{corr}}^{\\mathrm{cls}}, v_{\\mathrm{corr}}^{\\mathrm{pt}}\\} \\in \\mathbb{R}^{N_q \\times (N_s + 1) \\times ML}" + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "text", + "content": ", which is processed by the vision-only transformer. Attention is applied with query downsampling using (4,4) and (3,3) kernels for each layer. Patch tokens are reshaped into spatial dimensions before downsampling, resulting in a final patch output dimension of " + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "text", + "content": ". This produces foreground maps that are reshaped to " + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H' \\times W' \\times C'}" + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{seg}} \\in \\mathbb{R}^{N_q \\times C'}" + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{cls}} \\in \\mathbb{R}^{N_q \\times C'}" + }, + { + "bbox": [ + 313, + 343, + 555, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "text", + "content": "In the vision and text pathway, we concatenate vision and text class and patch correlation as " + }, + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{corr}} = \\{t_{\\mathrm{corr}}^{\\mathrm{cls}}, t_{\\mathrm{corr}}^{\\mathrm{pt}}\\} \\in \\mathbb{R}^{N_q \\times (N' + 1) \\times ML}" + }, + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "text", + "content": ". Later, attention is applied with (3,3) kernels for downsampling of query in both layers. This yields foreground maps " + }, + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{cls}} \\in \\mathbb{R}^{N_q \\times C'}" + }, + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{seg}} \\in \\mathbb{R}^{N_q \\times C'}" + }, + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "text", + "content": ", which are also reshaped to " + }, + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H' \\times W' \\times C'}" + }, + { + "bbox": [ + 313, + 444, + 556, + 530 + ], + "type": "text", + "content": " as the output." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Feature Enrichment Module: We propose a feature enrichment module to enhance classification and segmentation feature map predictions by incorporating both global and fine-grained details. Global context is essential for classification and also supports segmentation, while multi-scale feature capture is particularly advantageous for segmentation, helping to improve classification by minimizing errors from an over-reliance on global context alone. To meet these requirements, we designed an attention-based module capable of capturing both global and multi-scale information. While attention mechanisms are effective at capturing global information, they generally lack an inherent bias toward capturing objects at multiple scales. To address this, we introduce a granular Key-Value (KV) structure, which provides the classic attention block with multi" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25360" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "content": "scale capabilities by capturing information at various scales. The granular KV is generated using various atrous convolutions [3]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "text", + "content": "For an input " + }, + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "text", + "content": " to the multi-scale encoding module, the granular key-value pairs (KV) are generated as per Equation 2, where GAP refers to global average pooling, with dilation rates " + }, + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "inline_equation", + "content": "d1" + }, + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "inline_equation", + "content": "d2" + }, + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "inline_equation", + "content": "d3" + }, + { + "bbox": [ + 55, + 108, + 296, + 180 + ], + "type": "text", + "content": ". The final output of our feature enrichment module, Attn, where granular KV is used for key, value(Equation 3)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 191, + 293, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 191, + 293, + 205 + ], + "spans": [ + { + "bbox": [ + 64, + 191, + 293, + 205 + ], + "type": "interline_equation", + "content": "\\mathbf {K V} _ {i n} = \\operatorname {C o n c a t} \\left\\{\\left(\\operatorname {C o n v} (1 \\times 1), \\operatorname {C o n v} (3 \\times 3, r = d _ {1}) \\right. \\right.,", + "image_path": "3317339e7f0f91a0fba8f47451239b2961ac0b98cda5003c390d2487aae6e397.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 295, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 295, + 219 + ], + "type": "interline_equation", + "content": "\\left. \\operatorname {C o n v} \\left(3 \\times 3, r = d _ {2}\\right), \\operatorname {C o n v} \\left(3 \\times 3, r = d _ {3}\\right), G A P \\right\\} \\tag {1}", + "image_path": "3b8aaf50d63c7f4c7a12014dfe9e1103ec0b3ad219821258ea3afe8209d46f1f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 239, + 295, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 239, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 93, + 239, + 295, + 251 + ], + "type": "interline_equation", + "content": "\\mathbf {K V} _ {\\text {m u l t}} = \\operatorname {C o n v} \\left(\\left(1 \\times 1\\right), 5 C ^ {\\prime}, C ^ {\\prime}\\right) \\left(\\mathbf {K V} _ {\\text {i n}}\\right) \\tag {2}", + "image_path": "daa468e4409ea13a4eb4ad9b0bdd3aebea80717ab76e5c42a372dd1b1035f3de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 265, + 295, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 265, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 91, + 265, + 295, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {A t t n} = \\text {M u l t i H e a d A t t n} (Q, \\mathbf {K} _ {\\text {m u l t}}, \\mathbf {V} _ {\\text {m u l t}}) \\tag {3}", + "image_path": "bc5cf10df9a761c6d66847dbe3633b12a4cea8b4151319120b429eb461d68021.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}_{cls}" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}_{seg}" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}_{cls}" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}_{seg}" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": " denote the outputs of the multi-scale encoding module, derived from the feature maps " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{cls}" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{seg}" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{cls}" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{seg}" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": "The enriched vision-only classification foreground map, " + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}_{cls}" + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": ", and the enriched vision-text foreground map, " + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}_{cls}" + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": ", are then passed to a classification head, which consists of two pointwise convolutions followed by global average pooling to produce the classification outputs " + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{V}" + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{T}" + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": ". To refine the segmentation maps " + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}_{seg}" + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}_{seg}" + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": ", these are first processed by a segmentation head. This head utilizes " + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": " convolutional layers, which are alternated with a bilinear interpolation layer. This combination allows the segmentation predictions to be upscaled to match the original image dimensions, producing " + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{seg}^{V}" + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{seg}^{T}" + }, + { + "bbox": [ + 55, + 317, + 296, + 460 + ], + "type": "text", + "content": ", the final segmentation outputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "content": "Zero Shot Knowledge Distillation: Classification performance in our method suffers when support samples are limited to text, especially on datasets like Pascal- " + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "content": ", where the training set is smaller, amplifying the accuracy drop. To counter this, we utilize CLIP's zero-shot capability [20]. Unlike standard zero-shot classification, where text projections of known classes are compared to image projections, our task involves determining the presence of a specific support class. In zero-shot prediction, we use the positive prompt, 'This is a photo of {class-label}', followed by the negative prompt, 'This is not a photo of {class-label}.' Both prompts are processed through the CLIP text encoder and projection layer, producing the output embeddings " + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{+,p\\cdot cl_s}" + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{-,p\\cdot cl_s}" + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "content": " for the [EOS] token. The similarities, Pos and Neg, are then defined between the query image projection " + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_q^{p\\cdot cl_s}" + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "content": " and the positive and negative projections, " + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{+,p\\cdot cl_s}" + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_s^{-,p\\cdot cl_s}" + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "content": ", respectively. The zero-shot prediction is then derived by concatenating these similarities: " + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{cls} = \\text{Concat}(\\text{Neg}, \\text{Pos})" + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "content": ". Leveraging this knowledge, we apply cross-entropy loss between the zero-shot prediction " + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{cls}" + }, + { + "bbox": [ + 55, + 461, + 296, + 714 + ], + "type": "text", + "content": " (Eq. 5) and the zero-shot augmented class" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "predictions " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{zs}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " (Equation 4), which substantially boosts our model's performance. Softmax temperature is defined by " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " with learnable parameters as " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\gamma_1, \\gamma_2" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 387, + 118, + 555, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 118, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 387, + 118, + 555, + 133 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {c l s} ^ {z s} = \\gamma_ {1} \\mathbf {Y} _ {c l s} ^ {T} + \\gamma_ {2} \\mathbf {Z} _ {c l s} \\tag {4}", + "image_path": "3c2bfedc17e959559b808710d58daea9550f5b5324d9fd9658ef3e948bfffde2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 147, + 555, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 147, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 359, + 147, + 555, + 160 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} ^ {z s} = - \\mathbf {Z} _ {c l s} * \\log (\\operatorname {s o f t m a x} \\left(\\mathbf {Y} _ {c l s} ^ {z s} / \\tau\\right)) \\tag {5}", + "image_path": "3cabb97fb14f0de252b40c2d53b9a5101b94b8b21002a8e2b2cd10d193840d28.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 167, + 554, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 167, + 554, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 167, + 554, + 192 + ], + "type": "text", + "content": "3.2. Text-Only and Text-Augmented Few-shot Classification and Segmentation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": "In our text-augmented FS-CS approach, the support set includes both text and images with segmentation masks, while the query image has an associated target classification label and segmentation mask. Class predictions " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{V}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{T}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": ", along with segmentation predictions " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{seg}^{V}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{seg}^{T}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": ", are derived from foreground maps produced by vision-only and vision-text correlation, respectively. Classification refinement is achieved by merging vision-only predictions, " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{V}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": ", with vision-text predictions, " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{T}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": ", using the learnable scaling factors " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": ". Segmentation predictions from the vision-only and vision-text models are combined by applying scaling factors " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": " to integrate the results. Equations 6, 7, and 8 define the loss functions for classification, segmentation, and the total combined loss, respectively, 0.1 is set as default value of the scaling factor " + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 196, + 555, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 386, + 555, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 386, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 365, + 386, + 555, + 401 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c l s}} ^ {c} = - \\mathbf {Y} _ {\\text {c l s}} ^ {g t} \\log \\left(\\alpha_ {1} \\mathbf {Y} _ {\\text {c l s}} ^ {V} + \\alpha_ {2} \\mathbf {Y} _ {\\text {c l s}} ^ {T}\\right) \\tag {6}", + "image_path": "ba5fb91aed3bbd10934c9b41dba673e7eac48d43dc3cccfa9d6df1ec14b48c6f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 421, + 554, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 421, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 315, + 421, + 554, + 457 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e g}} ^ {c} = - \\frac {1}{H W} \\sum_ {x \\in | H | | W |} \\mathbf {Y} _ {\\text {s e g}} ^ {g t} (x) \\log \\left(\\beta_ {1} \\mathbf {Y} _ {\\text {s e g}} ^ {V} (x) + \\beta_ {2} \\mathbf {Y} _ {\\text {s e g}} ^ {T} (x)\\right) \\tag {7}", + "image_path": "d66331baa1bdc739e21cf60e030a42e15afc8c5992690e3b7bbf922763184fdb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 396, + 468, + 554, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 468, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 396, + 468, + 554, + 482 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {c} = \\lambda \\mathcal {L} _ {c l s} ^ {c} + \\mathcal {L} _ {s e g} ^ {c} \\tag {8}", + "image_path": "c6716f721c4a919cf71edd1bc398419bb1d90f87db75242b9f0745cad8eda26b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": "With support samples containing only class labels, our text-only approach uses only textual data, while each query image is paired with its corresponding classification labels and segmentation masks. During training, we generate segmentation predictions, " + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{seg}^{T}" + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": ", and classification predictions, " + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{T}" + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": ", based on the text input, following the previously outlined method. Additionally, zero-shot augmented class predictions " + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{zs}" + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": " are generated and used as the final classification output for loss computation. With " + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{cls}^{gt}" + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{seg}^{gt}" + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": " as the ground truth classification and segmentation targets, the classification, segmentation, and total losses are calculated using cross-entropy, as outlined in Equations 9, 10, and 11. The scaling factor " + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": ", is set to 0.1. Additionally, we incorporate a zero-shot knowledge distillation loss into the final loss calculation, scaled by " + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": ". In the text-augmented setting, we choose to exclude the zero-shot knowledge distillation loss, as its inclusion does not yield any performance improvements. In this setting, the richer support vision information leads to superior classifications compared to the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25361" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 205 + ], + "type": "table", + "html": "
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
\\( 20^0 \\)\\( 20^1 \\)\\( 20^2 \\)\\( 20^3 \\)Avg\\( 20^0 \\)\\( 20^1 \\)\\( 20^2 \\)\\( 20^3 \\)Avg\\( 20^0 \\)\\( 20^1 \\)\\( 20^2 \\)\\( 20^3 \\)Avg\\( 20^0 \\)\\( 20^1 \\)\\( 20^{2} \\)\\( 20^{3} \\)Avg
CLIP-B/16 (Zero-Shot)65.865.861.564.064.3-----47.648.042.545.846.0----
PANet [31]64.366.568.067.966.725.524.725.724.725.242.549.953.647.848.524.925.023.321.4
PFENet [28]70.770.671.272.971.430.634.829.432.631.935.634.343.132.836.523.323.820.223.1
HSNet [18]74.777.278.577.677.036.234.332.934.034.357.762.467.162.662.528.929.630.329.3
ASNet [11]76.278.879.280.278.635.736.835.335.635.859.561.568.862.463.129.833.033.430.4
CST [12]77.682.083.180.580.836.338.337.840.738.361.066.068.260.564.034.737.136.836.3
CST(CLIP-B/16)*84.689.090.589.588.445.349.245.345.746.374.578.680.478.678.041.749.046.546.1
(Ours)Text Only86.890.892.191.690.346.651.249.550.049.376.581.685.285.882.345.649.549.452.2
(Ours)Text Aug88.493.593.393.492.150.455.452.353.652.980.088.388.387.686.048.356.052.752.3
", + "image_path": "001d247c509591cee4a941dcaed1478b4290c711b77d0c8bb466e9f781ea385d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 247, + 555, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": "Table 1. We compare our method with previous approaches on COCO-20" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": ". The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk " + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 247, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 56, + 247, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 555, + 381 + ], + "type": "table", + "html": "
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
50515253Avg50515253Avg50515253Avg50515253Avg
CLIP-B/16 (Zero-Shot)82.278.979.276.479.2-----69.563.065.760.364.6-----
PANet [31]69.967.768.869.469.032.845.831.035.136.256.247.544.655.450.933.346.031.238.437.2
PFENet [28]69.882.468.177.974.638.354.735.143.843.022.561.740.339.541.031.147.330.832.235.3
HSNet [18]86.684.876.986.383.749.159.741.049.049.768.073.257.070.967.342.453.734.043.943.5
ASNet [11]84.989.679.086.284.951.761.543.352.852.368.576.258.670.068.348.558.336.348.347.8
CST [12]86.988.081.586.585.755.661.647.756.955.570.374.962.074.470.456.961.048.057.055.7
CST(CLIP-B/16)*92.493.080.590.189.069.169.348.063.462.483.987.665.179.679.066.369.547.462.361.4
(Ours)Text Only92.992.482.593.290.257.571.747.960.159.389.385.566.687.682.356.370.448.759.058.6
(Ours)Text Aug94.393.383.094.391.271.873.552.766.966.289.088.867.488.083.370.473.252.565.765.4
", + "image_path": "790a0531326fb150e56411b11eb11d7e14616806a833d1f62f513e172a2cd5c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "type": "text", + "content": "Table 2. On Pascal-5" + }, + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "type": "text", + "content": ", we compare our method with prior approaches. The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk " + }, + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "type": "inline_equation", + "content": "(^*)" + }, + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 433, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 458 + ], + "type": "text", + "content": "limited textual cues used by CLIP, making zero-shot distillation unnecessary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 467, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 467, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 134, + 467, + 295, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} ^ {t} = - \\mathbf {Y} _ {c l s} ^ {g t} \\log \\mathbf {Y} _ {c l s} ^ {z s} \\tag {9}", + "image_path": "efb7e8a4082fcbe4e367eab49eef055599df5d545379c937fe7aae146e751bb1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 491, + 295, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 491, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 91, + 491, + 295, + 518 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e g} ^ {t} = - \\frac {1}{H W} \\sum_ {x \\in | H | | W |} \\mathbf {Y} _ {s e g} ^ {g t} (x) \\log \\mathbf {Y} _ {s e g} ^ {T} (x) \\tag {10}", + "image_path": "495fe9f26cb1591a6f3916e7556e0a90d76bec6e15aa930f1fdc091fab9fc85a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 122, + 523, + 295, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 523, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 122, + 523, + 295, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {t} = \\lambda \\mathcal {L} _ {c l s} ^ {t} + \\mathcal {L} _ {s e g} ^ {t} + \\kappa \\mathcal {L} _ {c l s} ^ {z s} \\tag {11}", + "image_path": "602c34c966d1c54ad77f496afbd6a0ca8c9522591fc51e3de8a3f0400d6b0034.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 550, + 137, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 137, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 137, + 563 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "We conduct our experiments using CLIP ViT-B/16 [20] as the feature extractor in our model, while previous models' results were obtained with ResNet-50 [10]. The feature enrichment module uses dilation values of 3, 6, 10, and all experiments are optimized with the Adam optimizer at a learning rate of " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". To ensure an unbiased comparison, we replicate CST using CLIP-B/16 (Baseline) as the feature extractor, paired with their standard DINO (ViT-S/8) backbone. Evaluation adheres to established protocols [11, 12] and is performed across four folds on Pascal- " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " and COCO- " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". Evaluation is conducted with N-way K-shot episodes, while training uses 1-way-1-shot episodes. Further details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 434, + 555, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 457 + ], + "type": "text", + "content": "on experiments and ablations are provided in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 314, + 468, + 553, + 546 + ], + "blocks": [ + { + "bbox": [ + 314, + 468, + 553, + 546 + ], + "lines": [ + { + "bbox": [ + 314, + 468, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 553, + 546 + ], + "type": "image", + "image_path": "c69987226a527286d38e05f9fd6aebf3ba310338d5ded20a121d49041a64f606.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 553, + 555, + 574 + ], + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 574 + ], + "type": "text", + "content": "Figure 4. Performance comparison of N-way 1-shot FS-CS on Pascal-5" + }, + { + "bbox": [ + 313, + 553, + 555, + 574 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 313, + 553, + 555, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 599, + 553, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 553, + 623 + ], + "type": "text", + "content": "4.1. Few-shot classification and Segmentation Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "In our Pascal- " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " experiments (Table 2), our text-only approach generally outperforms prior methods in classification and segmentation, though it slightly trails CST with the CLIP-B/16 backbone on some metrics. Our text-augmented model, however, sets a new state-of-the-art, with notable improvements across 1-way-1-shot and 2-way-1-shot settings. In COCO- " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " (Table 1), the text-only setting per" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25362" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 194 + ], + "type": "table", + "html": "
Method1-way-1-shot2-way-1-shot
class. exact ratio (%)seg. mIoU (%)class. exact ratio (%)seg. mIoU (%)
50515253Avg50515253Avg50515253Avg50515253Avg
CLIP-B/16 (Zero-Shot)82.278.979.276.479.2-----69.563.065.760.364.6-----
HSNet [18]84.584.860.885.378.920.023.516.216.619.170.867.036.570.161.111.023.415.417.017.7
ASNet [11]80.284.066.282.778.311.721.113.416.215.667.670.244.569.362.910.320.712.615.914.9
CST [12]84.082.270.882.679.935.838.928.929.233.274.367.249.067.864.635.736.026.829.131.9
CST(CLIP-B/16)*87.087.873.086.683.644.643.832.033.938.577.678.053.775.071.045.242.530.433.037.8
GPT-4Vision [17]93.580.384.487.386.437.345.534.235.638.2----------
(Ours)Text Only93.989.784.192.189.144.147.232.639.340.889.678.773.087.482.243.445.433.139.740.4
(Ours)Text Aug94.390.684.192.890.449.747.833.739.442.690.281.773.487.683.248.646.733.539.842.1
", + "image_path": "22581e76fba929e4dde3faa4e1474c821fe7b5efaecdc01e9600887cd9bb366e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 243, + 302, + 306 + ], + "blocks": [ + { + "bbox": [ + 55, + 202, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 555, + 225 + ], + "type": "text", + "content": "Table 3. We compare our method with previous approaches on Pascal-5i in a weakly supervised setting. The CST baseline [12], implemented with the CLIP-B/16 backbone, is indicated by the row highlighted in grey with an asterisk " + }, + { + "bbox": [ + 55, + 202, + 555, + 225 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 55, + 202, + 555, + 225 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 243, + 302, + 306 + ], + "lines": [ + { + "bbox": [ + 56, + 243, + 302, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 302, + 306 + ], + "type": "table", + "html": "
Text perc.ArchCorr. SizeIter SpeedMemory
CSTDINO-S/8B × 2500 × 1441.77 it/s8.9G
CSTViT-B/16B × 625 × 1441.42 it/s4.7G
OursViT-B/16(B × 625 × 144) + (B × 625 × 77)1.7 it/s5.6G
", + "image_path": "f29e6777f86bb55f091e0b24add031e4f2ecdad1dd6211a82ca878fa9f5fa364.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 452 + ], + "type": "text", + "content": "forms slightly better than in Pascal- " + }, + { + "bbox": [ + 55, + 366, + 296, + 452 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 55, + 366, + 296, + 452 + ], + "type": "text", + "content": ", benefiting from the larger, diverse dataset, with further gains through text augmentation. These results highlight the complementary role of text data in FS-CS predictions, capturing nuances missed by vision-only inputs. Additional N-way 1-shot results on Pascal- " + }, + { + "bbox": [ + 55, + 366, + 296, + 452 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 55, + 366, + 296, + 452 + ], + "type": "text", + "content": " are in Fig 4, with detailed values in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 251, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 251, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 251, + 470 + ], + "type": "text", + "content": "4.2. Results of Weakly-Supervised FS-CS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 475, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 594 + ], + "type": "text", + "content": "In the weakly-supervised setting (Table 3), we employ DINO (ViT-S/8) to produce pseudo masks through similarity and thresholding, following CST [12]. Our text-only configuration considerably outperforms CST (CLIP-B/16), highlighting the value of text data in FS-CS predictions. The text-augmented setting further improves performance by leveraging both text and image data, demonstrating the advantages of multi-modal information in weakly-supervised contexts. Notably, we also apply zero-shot distillation in this text-augmented setup." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 149, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 149, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 149, + 618 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 219, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 219, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 219, + 637 + ], + "type": "text", + "content": "5.1. Effect on Runtime & Memory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "Integrating CLIP has a minimal impact on runtime or memory usage, as shown in Table 4 (measured on an RTX-3060 with a batch size of 8). This efficiency is due to the reduction in query image tokens from 2500 to 625 " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\downarrow 75\\%)" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", resulting from CLIP-B/16's larger patch size (16) compared to ViT-S/8 (patch size of 8). CST (ViT-S/8) requires more" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 365, + 243, + 504, + 324 + ], + "blocks": [ + { + "bbox": [ + 55, + 313, + 297, + 347 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 297, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 297, + 347 + ], + "type": "text", + "content": "Table 4. Comparison of runtime and memory consumption across different models, including our text augmentation model (B = batch size)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 365, + 243, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 365, + 243, + 504, + 324 + ], + "spans": [ + { + "bbox": [ + 365, + 243, + 504, + 324 + ], + "type": "table", + "html": "
Text perc.1-way-1-shot
cls. ERseg. mIoU
0%89.862.4
25%90.562.8
50%91.063.0
75%91.163.3
100%91.266.2
", + "image_path": "384ef9731006f06c8ea4bf93c15f7ba9d513c87910c9237c285ed65644613f0d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 332, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 354 + ], + "type": "text", + "content": "Table 5. The impact of different percentages of text augmentation on Pascal- " + }, + { + "bbox": [ + 313, + 332, + 555, + 354 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 313, + 332, + 555, + 354 + ], + "type": "text", + "content": " FS-CS." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 376, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 437 + ], + "type": "text", + "content": "resources due to its larger token count of 2500. In contrast, our text-augmented model, with dual CST blocks and correlation input sizes of " + }, + { + "bbox": [ + 313, + 376, + 555, + 437 + ], + "type": "inline_equation", + "content": "B \\times 625 \\times 144" + }, + { + "bbox": [ + 313, + 376, + 555, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 376, + 555, + 437 + ], + "type": "inline_equation", + "content": "B \\times 625 \\times 77" + }, + { + "bbox": [ + 313, + 376, + 555, + 437 + ], + "type": "text", + "content": ", delivers better performance with lower resource usage, while using slightly more resources than CST (CLIP-B/16)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 445, + 444, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 444, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 444, + 456 + ], + "type": "text", + "content": "5.2. Effect of text on FS-CS" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 463, + 555, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 584 + ], + "type": "text", + "content": "Text data enhances our model's performance in classification and segmentation, with the amount of textual information influencing the effectiveness of our Text-Augmented model. We examine this effect by adjusting the percentage of text used in training, from " + }, + { + "bbox": [ + 313, + 463, + 555, + 584 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 313, + 463, + 555, + 584 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 463, + 555, + 584 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 463, + 555, + 584 + ], + "type": "text", + "content": ", and present the results in Fig 5. During evaluation, we align the level of text augmentation with that used in training. As shown in Table 5, performance improves as more text is included, highlighting the benefits of text augmentation for enhancing downstream performance." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 603, + 553, + 680 + ], + "blocks": [ + { + "bbox": [ + 316, + 603, + 553, + 680 + ], + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 680 + ], + "type": "image", + "image_path": "57e64f7add31abefce5de94fc1d47a23b1bbb78c7e491d63158f46bbafa56a7f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 688, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 688, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 555, + 711 + ], + "type": "text", + "content": "Figure 5. Pascal- " + }, + { + "bbox": [ + 313, + 688, + 555, + 711 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 313, + 688, + 555, + 711 + ], + "type": "text", + "content": " FS-CS for different percentages of text augmented predictions." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25363" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 275, + 128 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 275, + 128 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 275, + 128 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 275, + 128 + ], + "type": "table", + "html": "
ImageTextPascal-5²Coco-20²
cls. ERseg. mIoUcls. ERseg. mIoU
✓X89.862.489.147.5
X✓90.259.390.349.3
✓✓91.266.292.152.9
", + "image_path": "5345f3dbe91c8a13e07b7f1d3bef3c038bce48f4d5dbf115c3597d9e8426b716.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 180, + 302, + 274 + ], + "blocks": [ + { + "bbox": [ + 55, + 137, + 295, + 159 + ], + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 159 + ], + "type": "text", + "content": "Table 6. The impact of combining text and images on FS-CS for COCO-20" + }, + { + "bbox": [ + 55, + 137, + 295, + 159 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 55, + 137, + 295, + 159 + ], + "type": "text", + "content": ", Pascal-5" + }, + { + "bbox": [ + 55, + 137, + 295, + 159 + ], + "type": "inline_equation", + "content": "^i" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 180, + 302, + 274 + ], + "lines": [ + { + "bbox": [ + 56, + 180, + 302, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 302, + 274 + ], + "type": "table", + "html": "
Pascal-5i\ncls. ER seg. mIoUCOCO-20i\ncls. ER seg. mIoU
No Feat. Enrinch.85.653.788.446.0
ASPP [3]83.355.387.646.3
Gran. Q84.856.188.347.6
Gran. QKV82.655.088.447.2
Attention83.256.488.247.4
Feat. Enrinch(Gran. KV)90.259.390.349.3
", + "image_path": "8a1241b057456684a6f08fd9c72c4a22d2c96b8e710881d52db0835aef435a52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 88, + 323, + 264, + 394 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "type": "text", + "content": "Table 7. Feature enrichment module with different components for FS-CS on Pascal- " + }, + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "type": "text", + "content": " and COCO- " + }, + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 323, + 264, + 394 + ], + "lines": [ + { + "bbox": [ + 88, + 323, + 264, + 394 + ], + "spans": [ + { + "bbox": [ + 88, + 323, + 264, + 394 + ], + "type": "table", + "html": "
ZS augment.CE. LossPascal-5²cls. ERCOCO-20²cls. ER
XX85.788.6
✓X84.089.7
X✓85.390.1
✓✓90.290.3
", + "image_path": "b3e4b0abaa639324f8b7736cd12012d89e85f2fb66c10cd5f5925ea2bcd69370.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "type": "text", + "content": "Table 8. Zero-shot distillation component effect on FS-CS for Pascal- " + }, + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "type": "text", + "content": " and COCO- " + }, + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "type": "text", + "content": "; grey indicates no zero-shot distillation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 454, + 295, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 295, + 480 + ], + "type": "text", + "content": "5.3. Comparison of Image, Text-only and Text-Augmented FS-CS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "text", + "content": "To assess the impact of images, text, and their combination on FS-CS predictions, we conduct evaluations in a 1-way-1-shot setup on Pascal- " + }, + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "text", + "content": " and COCO- " + }, + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "text", + "content": ". The results are shown in Table 6. On Pascal- " + }, + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "text", + "content": ", text-only predictions fall behind image-based predictions, though they are comparable on COCO- " + }, + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 55, + 484, + 295, + 616 + ], + "type": "text", + "content": ". However, our text-augmented model, which combines text and images, outperforms unimodal predictions on both datasets. Figure 1 shows that the text-augmented model enhances segmentation by drawing on semantic cues from both modalities, whereas text-only predictions often misclassify object boundaries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 295, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 649 + ], + "type": "text", + "content": "5.4. Effect of the class, segmentation feature enrichment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "We evaluate the impact of a foreground map enhancer on FS-CS predictions for Pascal- " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " and COCO- " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": ". We compare a text-only model enhanced with our feature enrichment module against models that apply ASPP to the query alone (Granular Q) or to the query, key, and value (Granular" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 556, + 168 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 556, + 168 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 556, + 168 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 556, + 168 + ], + "type": "image", + "image_path": "42fa745a3f2ab1b349c3604d7313bb7ad871f848c5a0e2e874cb7e3bc5da25d7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 175, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 220 + ], + "type": "text", + "content": "Figure 6. FS-CS predictions enriched, where features capture both fine-grained and global details, outperform default predictions that often misclassify and yield poor segmentation maps; " + }, + { + "bbox": [ + 313, + 175, + 555, + 220 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 313, + 175, + 555, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 175, + 555, + 220 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 175, + 555, + 220 + ], + "type": "text", + "content": " indicate class presence/absence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 255, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 374 + ], + "type": "text", + "content": "QKV), as well as those using standard attention or ASPP [3] directly on foreground maps. As shown in Table 7, our feature enrichment block, which incorporates ASPP to form granular KV, outperforms all other methods, including standard attention, ASPP alone, Granular Q, and Granular QKV. This improvement is due to the block's ability to capture global context and multi-scale features. Figure 6 further demonstrates how the feature enrichment block captures object details across scales, resulting in more accurate segmentation and classification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 554, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 414 + ], + "type": "text", + "content": "5.5. Effect of the zero shot knowledge distillation on FS-CS" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 422, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 555 + ], + "type": "text", + "content": "Table 8 presents the averaged results of the ablation study on zero-shot distillation components aimed at enhancing classification performance. The results underscore the importance of zero-shot augmented classification predictions and cross-entropy loss in improving accuracy. In the absence of cross-entropy loss, classification loss dominates, which undermines the goal of preserving CLIP's generalizability. Augmented predictions align effectively with CLIP's zero-shot classification, helping to minimize cross-entropy loss. As a result, combining both components is more effective than using either one individually." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 571, + 388, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 388, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 388, + 583 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "This paper demonstrates that text-only prompts can achieve high performance in Few-shot Classification and Segmentation (FS-CS), often outperforming prior methods. We further improved results by combining text with support image predictions (Text-Augmented), leveraging a multimodal approach. To address CLIP's generalization limits, we introduced zero-shot knowledge distillation and a feature enrichment module for capturing fine details and global semantics. Future work will explore additional modalities and improved multi-scale techniques to advance FS-CS." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25364" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Walid Bousselham, Felix Petersen, Vittorio Ferrari, and Hilde Kuehne. Grounding everything: Emerging localization properties in vision-language transformers. ArXiv, abs/2312.00878, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 296, + 235 + ], + "type": "text", + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 2, 3, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 280 + ], + "type": "text", + "content": "[4] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 313 + ], + "type": "text", + "content": "[5] Nanqing Dong and Eric P. Xing. Few-shot semantic segmentation with prototype learning. In British Machine Vision Conference, 2018. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 314, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 295, + 346 + ], + "type": "text", + "content": "[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28:594–611, 2006. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 380 + ], + "type": "text", + "content": "[7] Michael Fink. Object classification from a single example utilizing class relevance pseudo-metrics. In Neural Information Processing Systems, 2004. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 414 + ], + "type": "text", + "content": "[8] Chelsea Finn, P. Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 468 + ], + "type": "text", + "content": "[9] Zixian Guo, Bowen Dong, Zhilong Ji, Jinfeng Bai, Yiwen Guo, and Wangmeng Zuo. Texts as images in prompt tuning for multi-label image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2808-2817, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 470, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 470, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 470, + 295, + 514 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 296, + 558 + ], + "type": "text", + "content": "[11] Dahiyun Kang and Minsu Cho. Integrative few-shot learning for classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9979-9990, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 559, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 295, + 624 + ], + "type": "text", + "content": "[12] Dahyun Kang, Piotr Koniusz, Minsu Cho, and Naila Murray. Distilling self-supervised vision transformers for weakly-supervised few-shot classification & segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19627-19638, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 678 + ], + "type": "text", + "content": "[13] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "text", + "content": "[14] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 106 + ], + "type": "text", + "content": "supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 161 + ], + "type": "text", + "content": "[15] Chen Liu, Yanwei Fu, Chengming Xu, Siqian Yang, Jilin Li, Chengjie Wang, and Li Zhang. Learning a few-shot embedding model with contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8635-8643, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 555, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 555, + 226 + ], + "type": "text", + "content": "[16] Puneet Mangla, Mayank Kumar Singh, Abhishek Sinha, Nupur Kumari, Vineeth N. Balasubramanian, and Balaji Krishnamurthy. Charting the right manifold: Manifold mixup for few-shot learning. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2207-2216, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 554, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 260 + ], + "type": "text", + "content": "[17] Tian Meng, Yang Tao, and Wuliang Yin. Few-shot classification & segmentation using large language models agent. arXiv preprint arXiv:2311.12065, 2023. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 304 + ], + "type": "text", + "content": "[18] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 360 + ], + "type": "text", + "content": "[19] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chenyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23641-23651, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 361, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 554, + 426 + ], + "type": "text", + "content": "[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 470 + ], + "type": "text", + "content": "[21] Kate Rakelly, Evan Shelhamer, Trevor Darrell, Alexei A. Efros, and Sergey Levine. Conditional networks for few-shot semantic segmentation. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 471, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 554, + 502 + ], + "type": "text", + "content": "[22] Amirreza Shaban, Shray Bansal, Zhen Liu, Irfan Essa, and Byron Boots. One-shot learning for semantic segmentation. 2017. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 503, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 503, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 503, + 554, + 548 + ], + "type": "text", + "content": "[23] Mennatullah Siam, Boris Oreshkin, and Martin Jagersand. Amp: Adaptive masked proxies for few-shot segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5248-5257, 2019. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 554, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 592 + ], + "type": "text", + "content": "[24] Mennatullah Siam, Naren Doraiswamy, Boris N. Oreshkin, Hengshuai Yao, and Martin Jagersand. Weakly supervised few-shot object segmentation using co-attention with visual and semantic inputs. *IJCAI*, 2021. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 592, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 554, + 624 + ], + "type": "text", + "content": "[25] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 625, + 554, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 680 + ], + "type": "text", + "content": "[26] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 714 + ], + "type": "text", + "content": "[27] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In Computer" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25365" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "text", + "content": "Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 266-282. Springer, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[28] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44: 1050-1065, 2020. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 205 + ], + "type": "text", + "content": "[29] Haohan Wang, L. Liu, Wuhao Zhang, Jiangning Zhang, Zhenye Gan, Yabiao Wang, Chengjie Wang, and Haoqian Wang. Iterative few-shot semantic segmentation from image label text. *IJCAI*, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 295, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 260 + ], + "type": "text", + "content": "[30] Jin Wang, Bingfeng Zhang, Jian Pang, Honglong Chen, and Weifeng Liu. Rethinking prior information generation with clip for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3941-3951, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "text", + "content": "[31] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[32] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[33] Han-Jia Ye, Hexiang Hu, De chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8805-8814, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "text", + "content": "[34] Muyang Yi, Quan Cui, Hao Wu, Cheng Yang, Osamu Yoshie, and Hongtao Lu. A simple framework for text-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7071-7080, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[35] Sukmin Yun, Seong Hyeon Park, Paul Hongsuck Seo, and Jinwoo Shin. Ifseg: Image-free semantic segmentation via vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2967-2977, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 581 + ], + "type": "text", + "content": "[36] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5212-5221, 2019. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[37] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "text", + "content": "[38] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[39] Andrey Zhmoginov, Mark Sandler, and Maksym Vlademyrov. Hypertransformer: Model generation for supervised and semi-supervised few-shot learning. In International Conference on Machine Learning, pages 27075-27098. PMLR, 2022. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 283 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "text", + "content": "[40] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20061-20070, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 228 + ], + "type": "text", + "content": "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 229, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 283 + ], + "type": "text", + "content": "[43] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175-11185, 2023. 2, 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25366" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_content_list.json b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d147093449b5d245cf588693f04a28120f539075 --- /dev/null +++ b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_content_list.json @@ -0,0 +1,1829 @@ +[ + { + "type": "text", + "text": "Text Embedding is Not All You Need: Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps", + "text_level": 1, + "bbox": [ + 125, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jeeyung Kim*, Erfan Esmaeili*, and Qiang Qiu Purdue University", + "bbox": [ + 310, + 203, + 684, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jkim17,efakhabi,qqiu}@purdue.edu", + "bbox": [ + 341, + 241, + 651, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In text-to-image diffusion models, the cross-attention map of each text token indicates the specific image regions attended. Comparing these maps of syntactically related tokens provides insights into how well the generated image reflects the text prompt. For example, in the prompt, \"a black car and a white clock\", the cross-attention maps for \"black\" and \"car\" should focus on overlapping regions to depict a black car, while \"car\" and \"clock\" should not. Incorrect overlapping in the maps generally produces generation flaws such as missing objects and incorrect attribute binding. Our study makes the key observations investigating this issue in the existing text-to-image models: (1) the similarity in text embeddings between different tokens—used as conditioning inputs—can cause their cross-attention maps to focus on the same image regions; and (2) text embeddings often fail to faithfully capture syntactic relations already within text attention maps. As a result, such syntactic relationships can be overlooked in cross-attention module, leading to inaccurate image generation. To address this, we propose a method that directly transfers syntactic relations from the text attention maps to the cross-attention module via a test-time optimization. Our approach leverages this inherent yet unexploited information within text attention maps to enhance image-text semantic alignment across diverse prompts, without relying on external guidance. Our project page and code are available at: https://t-sam-diffusion.github.io/", + "bbox": [ + 89, + 323, + 485, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 760, + 222, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent diffusion models can generate images from diverse text prompts [4, 7, 29]. However, semantic discrepancies often arise between the text and generated images, raising problems such as missing objects—where certain elements are overlooked—and attribute mis-binding—where attributes are incorrectly assigned to subjects.", + "bbox": [ + 89, + 785, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c8d42dc8dcb0cb5b6afa2998885985a1237795b348c0e907a535f58f8ef83df.jpg", + "image_caption": [ + "Figure 1. The overview of our method. We leverage text self-attention matrix and optimize the latent noise $(z_{t})$ by minimizing the distance between the cross-attention similarity matrix (S) and the text self-attention matrix (T). This encourages integrating syntactic relationships into text-to-image diffusion models." + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 906, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior studies [10] demonstrated that the cross-attention map of each token in text-to-image (T2I) diffusion models highlights the attended regions in the image and provides clues about the spatial placement of elements corresponding to the tokens. In particular, [1, 21, 28] suggest that spatial alignment in cross-attention maps among related words influences the fidelity of images to text prompts in Stable Diffusion (SD) [29], as we also demonstrate in Section 4.1. For instance, in the prompt a black car and a white clock, if the cross-attention maps for car and clock overlap, unique token contributions can dilute, potentially omitting one object. Conversely, if the maps for black (or white) and car (or clock) diverge too much, attribute mis-binding can occur. This implies syntactically related words should ideally have spatially aligned cross-attention maps, as discussed in [8, 28]. However, the factors determining this spatial alignment between the maps remain poorly understood.", + "bbox": [ + 511, + 627, + 908, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In our study, we first investigate the factors that con", + "bbox": [ + 531, + 886, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 109, + 887, + 220, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f8159dd76edaf140b5a4126fec09a3c0db380076065fe9ee75c8a31dde699c14.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 91, + 74, + 313, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dbb3a59d5118e001b32252c91d42e33aa823baac6740916acf64f9273f35400f.jpg", + "image_caption": [ + "(b)", + "Figure 2. For the analysis, we use SD v1.5 as the base model and adopt the prompt sets from [3], structured as “[attribute_1] [object_1] and [attribute_2] [object_2]”, “[object_1] and/with [object_2]” or “[object_1] and [object_2] [object_2]”. (a) Comparison of the cosine similarity of text embeddings with that of the corresponding cross-attention maps at denoising step 1, with pairs of tokens (object_i, object_j), where $i \\neq j$ , and pairs of tokens (attribute_m, object_n) for both $m = n$ and $m \\neq n$ . As text embeddings become more similar, their cross-attention maps get similar. (b) The distributions of text embedding similarity between i) Bound tokens—(attribute_i, object_i) for $i = 1, 2$ , and ii) Unbound tokens—(object_i, object_2). The distributions show no discernible difference, indicating text embeddings do not effectively represent the syntactic relationships. (c) Comparison of text embedding similarity (left) and the text self-attention map power by 3 (right) for the prompt a black car and a white clock. In the self-attention maps (T), clock attends more to white, unlike the text embeddings." + ], + "image_footnote": [], + "bbox": [ + 316, + 74, + 527, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0762e095c3c40ec0c583d2ebb747ac1465b0843470fc06f1c8c4c6205449b18e.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 534, + 77, + 699, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c0244cacefe59feb7411ac6555786de6dd624f0b4483d59e144ffc6046b06c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 80, + 901, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tribute to the spatial alignment of cross-attention maps across different tokens, which can ultimately affect the accuracy of T2I generation. We reveal that text embeddings, which function as keys in the cross-attention modules, play a pivotal role in determining the similarity of these maps, as illustrated in Figure 2a. Specifically, as the text embeddings for different words become more similar, the cosine similarity of their corresponding cross-attention maps increases. This effect is noticeable from the initial denoising step.", + "bbox": [ + 88, + 375, + 480, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering that (i) cross-attention maps should capture syntactic relationships between words and that (ii) the spatial alignment of cross-attention maps is influenced by the text embeddings, we question whether text embeddings accurately capture the linguistic structure in text prompts. Our findings suggest they do not, as illustrated in Figure 2b and the left of Figure 2c. Syntactically related words in prompts (e.g., black-car and white-clock) do not necessarily yield similar text embeddings, as detailed in Section 4.1. This finding is further supported by [40], which argues that CLIP [25] often behaves like a bag-of-words model. As a result, the cross-attention maps, derived from text embeddings, can likely fail to faithfully reflect the syntactic relationships. In the end, the T2I diffusion models can often generate images that inaccurately represent the prompts.", + "bbox": [ + 89, + 513, + 482, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issue, prior studies [1, 8, 21, 28] resort to external sources to obtain syntactic information within text prompts and regularize cross-attention maps to incorporate these relationships. Specifically, [8, 28] employ external text parsers to obtain linguistic structure, while [1, 21] rely on human intervention to manually group tokens based on syntactic relationships. However, these methods are limited by their dependence on external inputs.", + "bbox": [ + 89, + 744, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Do we really need to rely on external sources to obtain syntactic relations within a sentence? Remarkably, T2I", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "diffusion models inherently capture syntactic relationships within prompts through the self-attention maps in their text encoder. In the encoder's self-attention module, each token pays more attention to related words, leading to encoding the entire sentence effectively. As shown in the right of Figure 2c, the text self-attention maps exhibit higher attention scores between the syntactically related words. However, this information is weakly encoded in the text encoder outputs (i.e., text embeddings). We conjecture this is due to the text attention module's strong focus on the $<\\text{bos}>$ token, known as attention sink [32, 35], which minimizes the influence of other tokens, as further discussed in Section 4.1.", + "bbox": [ + 511, + 375, + 906, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Figure 1, our approach reuses the self-attention maps from the text encoder and directly transfers the syntax information to the diffusion models. We first define a similarity matrix whose values indicate the similarity between pairs of cross-attention maps. Then, we update the latent noise in diffusion models to minimize the distance between the cross-attention similarity matrix and the text self-attention matrix during inference. This allows for the seamless integration of syntactic relationships into the diffusion process. By guiding cross-attention modules to better capture contextual relationships, our method ultimately produces images accurately reflecting the intended meaning of prompts. We leverage overlooked information already embedded in T2I models, offering two key advantages: (i) Self-contained: no need for external inputs like text parsers or manual token indices; and (ii) Generalizable: effective across diverse sentence structures.", + "bbox": [ + 511, + 556, + 908, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 828, + 653, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image generation. Text-to-image (T2I) generation [4, 5, 7, 23, 27, 30] is commonly based on latent diffusion models [29] where text data is processed via a text encoder", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8032", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[25, 26]. An important problem in T2I models is the lack of full correspondence between the input text and the image. Although diverse methods [2, 8, 12, 19, 24, 31, 38] are proposed to diagnose and fix this problem, a common theme in prior works is adjusting the cross-attention maps in inference time, following the seminal work [10].", + "bbox": [ + 89, + 90, + 482, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-attention control for improved semantic alignment. Several defects in cross-attention maps hinder semantic alignment, including attention dominance [41]—one token getting huge attention weight, and attention leakage/overlap [20, 36]—attention weights not respecting the spatial boundaries of the intended objects. In particular, one line of work employs contrastive objectives to guide attention optimization [1, 14, 21, 28]. For example, CONFORM [21] uses prompts with distinct groups, where tokens in the same (opposite) group are treated as positive (negative) pairs. Similarly, [28] uses text parser models to identify such pairs. Other approaches, like those in [15, 34], optimize attention maps with spatial guidance (e.g., segmentation maps, masks, or layouts). However, these methods rely on external resources for text-to-image generation.", + "bbox": [ + 89, + 181, + 483, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attention sinks. The attention weights in pretrained transformers tend to be heavily biased towards special tokens such as $\\langle \\text{bos} \\rangle$ , $\\langle \\text{eos} \\rangle$ , and punctuation tokens, as studied in (vision-)language models [9, 32, 35]. The T2I diffusion model literature [3, 37] also notes the focus of attention scores on the $\\langle \\text{bos} \\rangle$ token in the CLIP text encoder.", + "bbox": [ + 89, + 407, + 483, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 512, + 228, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we briefly review the structure of text-to-image diffusion models including the text encoder and the cross-attention module [4, 29].", + "bbox": [ + 89, + 537, + 482, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text encoder. To condition the text data for the diffusion process, a text encoder is needed. The input text is first tokenized and then converted into dense vectors. The dense vectors are then processed via a series of multi-head self-attention layers [33].", + "bbox": [ + 89, + 583, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At a given layer $\\ell$ in the text encoder, let us denote the key by $\\mathbf{e}_i^{(\\ell)}\\in \\mathbb{R}^{H_{\\mathrm{e}}D_{\\mathrm{e}}}$ $i = 1,\\dots ,s$ where $s$ is the key sequence length, $D_{\\mathrm{e}}$ is the embedding dimension per head and $H_{\\mathrm{e}}$ is the number of heads in the text encoder. The corresponding self-attention matrix $T^{(\\ell ,h)}\\in \\mathbb{R}^{s\\times s}$ is given by", + "bbox": [ + 89, + 659, + 483, + 736 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i j} ^ {(\\ell , h)} = \\frac {\\exp (\\omega_ {i j})}{\\sum_ {k} \\exp (\\omega_ {i k})}, \\omega_ {i j} := \\mathbf {e} _ {i} ^ {(\\ell) \\top} W _ {\\mathrm {e n}} ^ {(\\ell , h)} \\mathbf {e} _ {j} ^ {(\\ell)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 747, + 482, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W_{\\mathrm{en}}^{(\\ell,h)} \\in \\mathbb{R}^{H_{\\mathrm{c}}D_{\\mathrm{c}} \\times H_{\\mathrm{c}}D_{\\mathrm{c}}}$ is a pretrained matrix at head $h$ in layer $\\ell$ of the text encoder.", + "bbox": [ + 89, + 792, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our proposed method, we use the self-attention matrix averaged over layers and heads:", + "bbox": [ + 89, + 825, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {\\prime} = \\frac {1}{L _ {\\mathrm {e}} H _ {\\mathrm {e}}} \\sum_ {\\ell = 1} ^ {L _ {\\mathrm {e}}} \\sum_ {h = 1} ^ {H _ {\\mathrm {e}}} T ^ {(\\ell , h)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 862, + 482, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the high attention probabilities assigned to $\\langle \\text{bos} \\rangle$ and $\\langle \\text{eos} \\rangle$ tokens, we remove their corresponding values and re-normalize each row as follows:", + "bbox": [ + 511, + 90, + 906, + 135 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i j} = \\frac {T ^ {\\prime} {} _ {i j}}{\\sum_ {m = 2} ^ {i} T ^ {\\prime} {} _ {i m}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 140, + 906, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote the final output of the text encoder by $\\mathbf{k}_i\\in$ $\\mathbb{R}^{H_{\\mathrm{e}}D_{\\mathrm{e}}}$ where $i = 1,\\dots ,s$ , which used as conditioning inputs in T2I diffusion models. We call $\\mathbf{k}_i$ as text embeddings.", + "bbox": [ + 511, + 180, + 906, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising latent variables. Latent diffusion models generate latent tensors $(z_0)$ of an image. A latent diffusion model $D_{\\theta}$ learns to simulate the denoising process: starting from Gaussian noise $z_{\\tau}$ , the model iteratively reconstruct $z_0$ by predicting the noise at step $t$ :", + "bbox": [ + 511, + 226, + 905, + 301 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = z _ {t} - D _ {\\theta} \\left(z _ {t}; \\left\\{\\mathbf {k} _ {i} \\right\\}\\right), \\quad 0 < t \\leq \\tau . \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 308, + 906, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-attention module. In T2I latent diffusion models, the interaction between text and image data is performed by multi-head cross-attention. In the cross-attention layer $\\ell$ , we define the query, $\\mathbf{q}_a^{(\\ell)} \\in \\mathbb{R}^{H_cD_c}$ , where $a = 1, \\dots, N_c$ , and $N_c$ is the query sequence length at the given cross-attention layer. $D_c$ is the hidden dimension per head, and $H_c$ is the number of heads of the cross-attention layer.", + "bbox": [ + 511, + 330, + 905, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The cross-attention map $A^{(\\ell,h)} \\in \\mathbb{R}^{N_{\\mathrm{c}} \\times s}$ at head $h$ with elements $A_{ai}^{(\\ell,h)}$ is defined as", + "bbox": [ + 511, + 438, + 905, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {a i} ^ {(\\ell , h)} := \\frac {\\exp (\\Omega_ {a i})}{\\sum_ {j = 1} ^ {s} \\exp (\\Omega_ {a j})}, \\Omega_ {a i} := \\mathbf {q} _ {a} ^ {(\\ell) \\top} W _ {\\mathrm {c}} ^ {(\\ell , h)} \\mathbf {k} _ {i} ^ {(\\ell)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 474, + 903, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $W_{\\mathfrak{c}}^{(\\ell,h)} \\in \\mathbb{R}^{H_{\\mathfrak{c}}D_{\\mathfrak{c}} \\times H_{\\mathfrak{c}}D_{\\mathfrak{c}}}$ is a pretrained matrix at head $h$ in layer $\\ell$ .", + "bbox": [ + 511, + 527, + 905, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-attention similarity matrix. In our study, we use the cross-attention maps averaged over heads and over all layers for which $N_{\\mathrm{c}} = M$ :", + "bbox": [ + 511, + 561, + 905, + 606 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathsf {A} = \\frac {1}{L _ {M} H _ {\\mathrm {c}}} \\sum_ {\\ell = 1} ^ {L _ {M}} \\sum_ {h = 1} ^ {H _ {\\mathrm {c}}} A ^ {(\\ell , h)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 612, + 903, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $L_{M}$ is the number of cross-attention layers with $N_{\\mathrm{c}} = M$ . Based on the average cross-attention map above, we define the similarity matrix $\\mathsf{S} \\in \\mathbb{R}^{s \\times s}$ as follows:", + "bbox": [ + 511, + 657, + 905, + 702 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S} _ {i j} := \\frac {\\mathrm {C} _ {i j}}{\\sum_ {k = 1} ^ {s} \\mathrm {C} _ {i k}}, \\mathrm {C} _ {i j} := \\frac {\\sum_ {a = 1} ^ {N _ {\\mathrm {c}}} \\mathrm {A} _ {a i} \\mathrm {A} _ {a j}}{\\left(\\sum_ {a = 1} ^ {N _ {\\mathrm {c}}} \\mathrm {A} _ {a i} ^ {2}\\right) ^ {\\frac {1}{2}} \\left(\\sum_ {a = 1} ^ {N _ {\\mathrm {c}}} \\mathrm {A} _ {a j} ^ {2}\\right) ^ {\\frac {1}{2}}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 707, + 903, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$C_{ij}$ denotes cosine similarity between maps corresponding to $i$ and $j$ .", + "bbox": [ + 511, + 771, + 905, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Understanding and Resolving Text-to-Image Semantic Discrepancy", + "text_level": 1, + "bbox": [ + 511, + 811, + 905, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we examine the cross-attention maps in Stable Diffusion (SD) [29] to identify factors that lead to semantic discrepancies between generated images and text", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prompts. We provide additional results with $\\mathrm{PixArt - \\alpha}$ [4] in Appendix B.1. Based on these findings, we introduce an approach to enhance the fidelity of T2I diffusion models by leveraging text self-attention maps to regularize cross-attention maps.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Why Do Generated Images Misrepresent Text?", + "text_level": 1, + "bbox": [ + 89, + 176, + 482, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our findings are based on the premise that cross-attention maps should capture syntactic relationships between words for accurate T2I generation, as supported by prior works [1, 8, 21, 28]. We identify a key factor contributing to incorrect relations in cross-attention maps, resulting in less faithful image generation: While text embedding $(\\mathbf{k}_i)$ similarity strongly correlates with the cross-attention similarity matrix C, it insufficiently reflects syntactic bindings in the text prompt. This highlights a fundamental issue in text-image semantic discrepancy: the similarity in text embeddings lack the syntax information necessary for accurate image generation. We reveal that text self-attention maps (T) effectively capture syntactic information, but this is not sufficiently encoded in text embeddings. The absence of the information in text embeddings can be due to an artifact in the attention module, known as attention sink [35], where attention weights are biased toward the $<\\mathrm{bos}>$ token as detailed later. These findings motivate the solution presented in the following section.", + "bbox": [ + 89, + 200, + 483, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the following empirical analysis, we use the prompt sets introduced in [3], containing prompts structured as the following three categories:", + "bbox": [ + 89, + 489, + 483, + 534 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) $\\left[\\mathsf{attribute}_1\\right]\\left[\\mathsf{object}_1\\right]$ and $\\left[\\mathsf{attribute}_2\\right]\\left[\\mathsf{object}_2\\right]$ ,", + "(ii) $\\left[\\mathrm{object}_1(\\mathrm{animal})\\right]$ with $\\left[\\mathrm{object}_2\\right]$ ,", + "(iii) [object1(animal)] and [attribute2][object2]." + ], + "bbox": [ + 89, + 534, + 450, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt sets in the above formats enable us to focus on two key cases of text-image semantic discrepancies: missing objects and attribute mis-binding. In the following, we refer to the group of syntactically bound words (tokens) as pairs of $([\\text{attribute}_i], [\\text{object}_i])$ , $i = 1, 2$ , and the group of syntactically unbound words (tokens) as $([\\text{attribute}_i], [\\text{object}_j])$ and $([\\text{object}_1], [\\text{object}_2])$ , where $i \\neq j$ .", + "bbox": [ + 89, + 583, + 483, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We revisit the role of cross-attention maps in text-image semantic coherence, where spatial overlap or separation is crucial, as discussed in [1, 8, 21, 28]. Figure 3 illustrates this effect: On the left, overlapping attention maps for syntactically unbound words lead to object missing, while greater overlap for syntactically bound words enhances attribute binding. This suggests that syntactic associations should be reflected in cross-attention maps.", + "bbox": [ + 89, + 688, + 483, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To statistically assess cross-attention map's impact, we analyze categories (ii) and (iii) of the aforementioned prompt sets. Assuming spatial overlap is measurable by the cosine similarity matrix (C) in eq.(7), we compute cosine similarities for the attention maps of $object_{1}$ and $object_{2}$ in missing objects, and for $attribute_{2}$ and $attribute_{2}$ in attribute", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/820c2323337cdbfceed0662d9a2144b441cb6746dcb4ba64796b474288e0ac15.jpg", + "image_caption": [ + "1) Missing object A bear and a red balloon", + "bear", + "balloon" + ], + "image_footnote": [], + "bbox": [ + 517, + 119, + 609, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 563, + 189, + 606, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/34d5b58472365410e00197e2d03301c3a6f8a1414d5baa0a98a57a0fe7701e1a.jpg", + "image_caption": [ + "A frog with a bow", + "frog" + ], + "image_footnote": [], + "bbox": [ + 612, + 119, + 704, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/40241db9d828d7c43b2e1b3a76a8e0c6b235860d87d5aa17b1947646e4e91b5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 127, + 807, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9b6a22eae0f1cc2b90d57453b1284d11a69a2de65bc1ab1f603ca4eeec124484.jpg", + "image_caption": [ + "2) Attribute binding", + "A green backpack and a purple bench", + "#", + "green", + "#", + "1" + ], + "image_footnote": [], + "bbox": [ + 808, + 118, + 857, + 160 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/daf810ec6d34c952b31b3b00904d600777e8f41d8869f3c426aa06d60714dc20.jpg", + "image_caption": [ + "backpack" + ], + "image_footnote": [], + "bbox": [ + 808, + 176, + 857, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3df530aaadbb446df85c15ec63db8beb9567bfe2d3ed48538e341a88fcfb1b2.jpg", + "image_caption": [ + "bench" + ], + "image_footnote": [], + "bbox": [ + 861, + 176, + 910, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5dff54dd72e49694b1b0055deecb3bf54110888f4e958201a87bcfa0462151ea.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 519, + 327, + 705, + 436 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33f5d35e73453fa3ade2fb250691afd3479196a335bd709a5676af044b95fcaa.jpg", + "image_caption": [ + "Figure 3. The generated images and cross-attention maps (A) for the specific tokens from SD v1.5. This illustrate the importance of spatial alignment in cross-attention maps for accurate image generation. Divergent (overlapping) cross-attention maps for syntactically unbound (bound) words enhances text-to-image fidelity.", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 710, + 325, + 895, + 436 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e1d37eb917f850d001a7888302681e5abef7b2cb4b882e3639c50c3368e55f16.jpg", + "image_caption": [ + "Figure 4. Comparison for the distributions of cosine similarity between cross-attention maps (at denoising step 10). (a) The cases with one missing object-incorrect and two objects present-correct. (b) The cases with incorrect and correct attribute binding. Correct instances are more frequent when the cosine similarity is low for objects presence and high for attribute binding.", + "Figure 5. Correlation between the cosine similarity of text embeddings and that of cross-attention maps across denoising steps $(t = 1,21,50)$ . Similar text embeddings generally lead to similar cross-attention maps, with the correlation weakening over time." + ], + "image_footnote": [], + "bbox": [ + 516, + 544, + 647, + 630 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d773f8e3db246ae126b3da330a48b286b3bf42732172c9fca0fcb29b6f2a1c94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 545, + 772, + 630 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4eab806f49ad4a252db821362de0127cafbe2223cbbdd203db6d5b9744cce3e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 545, + 901, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "binding cases. Then, we compare cosine similarity distributions for correct vs. incorrect images on both cases. See Appendix B.1 for setup details. Figure 4 shows that lower cosine similarity tends to correlate with object presence, while higher similarity supports more accurate bindings.", + "bbox": [ + 511, + 715, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This observation motivates us to examine the factor contributing to similarity in cross-attention maps across tokens. Finding 1: The cosine similarity of text embeddings has a large correlation with cross-attention similarity matrix C. Figure 5 shows there is a correlation between the cosine similarity of text embedding and C, which persists throughout the final denoising steps. This indicates that similar text", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8034", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "embeddings can result in overlapping cross-attention maps.", + "bbox": [ + 89, + 90, + 480, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we justify this finding mathematically.", + "bbox": [ + 109, + 109, + 405, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1. If $A^{(\\ell, h)} \\in \\mathbb{R}^{N_c \\times s}$ is a cross-attention map defined in eq. (5), then under the assumptions i, ii, and iii described in Appendix A, the cosine similarity matrix can be written in terms of key vectors $\\mathbf{k}_i^{(\\ell, h)} \\in \\mathbb{R}^{H_c D_c}$ as", + "bbox": [ + 89, + 143, + 483, + 208 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\cos \\left(A _ {i} ^ {(\\ell , h)}, A _ {j} ^ {(\\ell , h)}\\right) = \\\\ \\exp \\left(- \\frac {1}{2} \\left(\\mathbf {k} _ {i} - \\mathbf {k} _ {j}\\right) ^ {\\top} W ^ {2} \\left(\\mathbf {k} _ {i} - \\mathbf {k} _ {j}\\right)\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 212, + 482, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "up to terms of at least $\\mathcal{O}(1 / \\sqrt{N_c})$ and $\\mathcal{O}(\\epsilon)$ , where $W^2 \\coloneqq W_c^{(\\ell,h)^\\top}\\Sigma^{(\\ell)}W_c^{(\\ell,h)}$ and $\\Sigma^{(\\ell)} \\in \\mathbb{R}^{H_cD_c \\times H_cD_c}$ is the covariance matrix of query vectors.", + "bbox": [ + 89, + 270, + 483, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Refer to Appendix A for the proof.", + "bbox": [ + 89, + 339, + 323, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We empirically and mathematically showed that the similarity in text embeddings $\\mathbf{k}_i$ influences cross-attention maps. Next, we evaluate whether the similarity of these embeddings reflects syntax information in text inputs.", + "bbox": [ + 89, + 358, + 483, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finding 2-1: There is no significant correlation between word syntactic bindings and text embedding similarity. Prior study [40] suggests CLIP embeddings [25], used in SD, behave like a bag-of-words model, ignoring word relationships. We also show in Figure 2b the similarity in CLIP text embeddings does not correlate with syntactic bindings. Specifically, we expect close embeddings for syntactically bound tokens and distant embeddings for unbound tokens, yet the distributions lack separation.", + "bbox": [ + 89, + 421, + 483, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finding 2-2: Text self-attention maps do have syntax information. We examine the text encoder that produces text embeddings. Interestingly, the self-attention maps in the encoder capture syntactic relationships, as shown in Figure 6a. These maps reveal higher similarity between syntactically bound tokens and lower similarity between unbound tokens. For more complex prompt structures, see the Appendix B.1.", + "bbox": [ + 89, + 561, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finding 2-3: The attention sink can contribute to why text embeddings lack the syntax information. Why do text embeddings lack relational information despite being derived from multiple self-attention modules? We attribute this gap to attention sink [32, 35], where attention scores are concentrated on a few tokens. In CLIP's text encoder, attention is mainly focused on the $\\langle \\text{bos} \\rangle$ token, as discussed in [3, 37] and shown in Figure 6b. We hypothesize the focus on the $\\langle \\text{bos} \\rangle$ token can limit the transfer of relational information from self-attention maps $(T)$ to text embeddings, as attention scores for other tokens remain much smaller, minimizing their influence in each self-attention layer.", + "bbox": [ + 89, + 685, + 483, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To provide a mathematical justification, we express the difference between the output and input of one block of the", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6020acbb489405fffbd5851eb34f530e8e880266f49002eec1ea244495d18e03.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 537, + 83, + 746, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/824ea9085e315c096db8dc7d07f0037be86c830e4504eba7eaded7809045e6c6.jpg", + "image_caption": [ + "(b)", + "Figure 6. (a) The distributions of text self-attention value $(\\mathsf{T}_{ij}^{\\prime})$ in eq.(2)) for bound tokens (attribute $m$ , object $m$ ) and unbound tokens (attribute $m$ , object $n$ ) / (object $1$ , object $2$ ), where $m, n \\in \\{1, 2\\}$ . The separate distributions indicate text self-attention maps can indeed represent the syntactic relationships. (b) Comparison of text self-attention probability histograms between token and non tokens on 100 prompts: The probability allocated to is on average 20 times larger than that of other tokens." + ], + "image_footnote": [], + "bbox": [ + 751, + 89, + 880, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "self-attention module for a token $e_i$ as:", + "bbox": [ + 513, + 398, + 772, + 414 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {o} _ {i} ^ {(\\ell , h)} = \\sum_ {j = 1} ^ {i} T _ {i j} ^ {(\\ell , h)} W _ {\\mathrm {v}} ^ {(\\ell , h)} \\mathbf {e} _ {j} ^ {(\\ell)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 419, + 905, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{\\mathrm{v}}^{(\\ell,h)} \\in \\mathbb{R}^{D_{\\mathrm{e}} \\times H_{\\mathrm{e}} D_{\\mathrm{e}}}$ is a parameter, $\\mathbf{e}_i \\in \\mathbb{R}^{H_{\\mathrm{e}} D_{\\mathrm{e}}}$ , and $T_{ij}^{(\\ell,h)}$ is the self-attention matrix in eq. (1). We consider the situation where attention sink occurs, that is the attention weights for the token is much higher than the rest of the sequence:", + "bbox": [ + 511, + 465, + 906, + 546 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon = \\frac {\\sum_ {j \\neq 1} ^ {i} T _ {i j}}{T _ {i 1}} \\ll 1, \\quad i = 2, \\dots , s. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 551, + 905, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Appendix A, we prove the following statement:", + "bbox": [ + 511, + 590, + 846, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2. Define matrix $R \\in \\mathbb{R}^{s \\times s}$ as", + "bbox": [ + 513, + 613, + 803, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i j} = \\mathbf {e} _ {i} ^ {(\\ell) \\top} W _ {v} ^ {(\\ell , h) \\top} W _ {v} ^ {(\\ell , h)} \\mathbf {e} _ {j} ^ {(\\ell)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 632, + 903, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and suppose it has the property", + "bbox": [ + 511, + 657, + 723, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\left| R _ {m n} \\right|}{R _ {1 1}} \\sim \\mathcal {O} (1 / \\epsilon), \\quad \\frac {\\left| R _ {1 m} \\right|}{R _ {1 1}} \\sim \\mathcal {O} (1), \\quad 1 < m, n \\leq s, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 676, + 903, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{e}_1^{(\\ell)}$ is the bos embedding. Then the following holds:", + "bbox": [ + 511, + 720, + 901, + 738 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\cos \\left(\\mathbf {o} _ {i} ^ {(\\ell , h)}, \\mathbf {o} _ {j} ^ {(\\ell , h)}\\right) = 1 - \\mathcal {O} (\\epsilon). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 742, + 903, + 765 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6aaf0e5f90f7cdd380846164eee8f66c00f8e88fab2309db70d9d6d5924f58e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 888, + 768, + 903, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This suggests that cosine similarity between token vectors $(\\mathbf{o}_i^{(\\ell ,h)})$ -potentially influencing cross-attention maps remains barely changed across text self-attention layers due to attention sink on the. $\\langle \\mathrm{bos}\\rangle$ token. In other words, the attention sink can hinder the accurate encoding of self-attention maps into embeddings.", + "bbox": [ + 511, + 790, + 906, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our findings reveal text embedding alone are insufficient", + "bbox": [ + 531, + 885, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8035", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for generating semantically aligned images. On the other hand, we notably show the potential of transferring neglected syntactic information from text self-attention maps to the cross-attention to enhance T2I semantic alignment.", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Text Self-Attention Maps (T-SAM) Guidance", + "text_level": 1, + "bbox": [ + 89, + 159, + 472, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the previous section, we show text self-attention maps capture syntax information within a sentence. Building on this insight, we propose leveraging the self-attention maps within the text encoder—a component of diffusion models—to enhance cross-attention maps. By minimizing the distance between the similarity matrix of the cross-attention maps and the text self-attention matrix, our approach ensures that embedded syntactic relationships are effectively transferred to cross-attention.", + "bbox": [ + 88, + 181, + 482, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method optimizes cross-attention maps during inference, adjusting their similarity matrix to align with the text self-attention matrix $\\mathsf{T}$ . The normalized cosine similarity matrix, S (defined in eq.(7)), is used as the cross-attention similarity matrix. This is achieved by simply minimizing the loss function:", + "bbox": [ + 89, + 318, + 483, + 407 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(z _ {t}\\right) = \\sum_ {i = 1, j \\leq i} ^ {s} \\rho_ {i} \\left| \\mathrm {T} _ {i j} ^ {\\gamma} - \\mathrm {S} _ {i j} \\left(z _ {t}\\right) \\right|, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 410, + 482, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the exponent $\\gamma$ acts to amplify larger values and compress smaller ones so the effect of controlling temperature and $\\rho_{i} = i / s$ . For example, if two words in the prompt have negligible syntactic relation according to the text self-attention matrix, i.e. $\\mathsf{T}_{ij}\\approx 0$ , we demand that their similarity of cross-attention maps must not be similar: $\\mathsf{S}_{ij}\\approx 0$ .", + "bbox": [ + 89, + 452, + 483, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In practice, this optimization will be applied only to $z_{t}$ at a few denoising steps during inference as followed:", + "bbox": [ + 89, + 542, + 483, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} ^ {\\prime} = z _ {t} - \\alpha \\cdot \\nabla_ {z _ {t}} \\mathcal {L} (z _ {t}). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 575, + 482, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 603, + 223, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section presents our empirical study on the effectiveness of T-SAM in enhancing semantic alignment across diverse T2I diffusion models, including SD v1.5 [29] and PixArt- $\\alpha$ [4], evaluated both quantitatively and qualitatively. To the best of our knowledge, this is the first work to introduce attention control in PixArt- $\\alpha$ , underscoring the novelty of our approach.", + "bbox": [ + 89, + 628, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt datasets. We evaluate our approach on diverse text prompts using two existing benchmarks. First, the TIFA v1.0 benchmark [13] is a large-scale text-to-image generation dataset featuring a wide range of sentence structures. This benchmark comprises 4,000 prompts, including 2,000 image captions from the COCO validation set [18], 161 prompts from DrawBench [30], 1,420 prompts from PartiPrompt used in Parti [39], and 500 texts from PaintSkill [6]. Second, we use structured prompt sets containing multiple objects and their corresponding attributes from Attend-n-Excite [3]. Prompts in this dataset are", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ef54ea184b0589b04c2eeb0e12706a8fffb0e484837a06cbb4a6e926dd271596.jpg", + "table_caption": [ + "Table 1. Evaluation results for the TIFA benchmark, including TIFA scores and CLIP similarity scores. External Info. indicates whether external information is used. CLIP-I (CLIP-T) refers to image-text (text-text) CLIP similarity. The base model is SD." + ], + "table_footnote": [], + "table_body": "
External Info.TIFACLIP-ICLIP-T
SD×0.790.330.77
LB✓0.800.330.76
T-SAM×0.830.340.77
", + "bbox": [ + 519, + 148, + 898, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "grouped into three categories: Objects (e.g., “[attribute1] [object1] and [attribute2] [object2]”), Animals-Objects (e.g., “[animal] with [object]” or “[animal] and/with [attribute] [object]”), and Animals (e.g., “[animal1] and [animal2]”). We exclude the Animals category, as it lacks the complex syntactic relations as discussed in [28].", + "bbox": [ + 511, + 247, + 903, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. For SD, we use 50 sampling iterations, updating $z_{t}$ at each denoising step from 1 to 25. We set $M = 256$ in eq.(6) and $\\gamma = 2$ in eq.(14). For the TIFA benchmark, we generate one image per prompt with a shared seed across methods, while using a unique seed for each prompt. We set $\\alpha = 40$ in eq.(15). For Attend-n-Excite prompts, we generate results using 64 seeds, in line with standard practice, setting $\\alpha = 10$ and applying 20 iterative updates at $i \\in \\{0,10,20\\}$ . See Appendix B.2 for details on PixArt- $\\alpha$ as well as SD.", + "bbox": [ + 511, + 339, + 905, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. To quantitatively evaluate the accuracy of generated images, we use two metrics: TIFA scores [13] (for TIFA benchmark) and CLIP similarity scores. TIFA scores measure how well generated images reflect the text prompts. In TIFA, questions for each prompt are generated by GPT-3.5 [22], and a vision-language model [16] provides answers. For CLIP [25] scores, we follow the protocol in [3] to evaluate both image-text and text-text similarity. Additional details on CLIP scores are provided in Appendix B.2. Moreover, we conducted a user study with 10 random prompts in Animals-Objects, generating four images per prompt using different seeds. A total of 25 participants select the image that best matches the text.", + "bbox": [ + 511, + 491, + 905, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. When evaluating our method on SD, we compare our method with SD, Linguistic-Binding (LB) [28], Attend-n-Excite (A&E) [3] and CONFORM [21]. LB depending on external parsers (SpaCy [11]) is limited to attribute-binding tasks. This method can be applied to both the TIFA benchmark and the Attend-n-Excite (Objects and Animals-Objects) prompt sets, though they are used only when the text parser identifies (modifier, entity-noun) pairs within the given prompts. A&E and CONFORM require manual token selection per prompt, restricting their use to the fixed-template prompts (Objects and Animals-Objects) due to the high cost of selecting token indices for diverse prompts. We reproduce the images with SD, CONFORM, and LB based on SD v1.5 and using the same seeds. For PixArt- $\\alpha$ , we com", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8036", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/063924eeb3667c09d7b26adb25ff4af23e395b36fa8b01f168b768aec1cdfd45.jpg", + "image_caption": [ + "Figure 7. Accuracy for each question type in the FIFA benchmark. Our method achieves the best performance in most categories. The attribute includes properties such as large, small, young, etc. The base model is SD." + ], + "image_footnote": [], + "bbox": [ + 176, + 84, + 823, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04a9f402dc2566b67bf450c19545222ba121bc36dd3e86c89de884ad49afdd02.jpg", + "image_caption": [ + "Figure 8. Comparison of our method (T-SAM) on SD with recent state-of-the-art methods on prompts from Objects and Animals-Objects. The images corresponding to the same position across different methods are generated using the same seed. Best in zoom." + ], + "image_footnote": [], + "bbox": [ + 178, + 222, + 815, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pare only with the base model, as no prior attention-control methods have been applied to it.", + "bbox": [ + 89, + 733, + 485, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Results", + "text_level": 1, + "bbox": [ + 89, + 771, + 184, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. Table 1 shows the evaluation results in TIFA benchmark and Figure 7 illustrates the breakdown of TIFA scores across question types, using SD as the base model. We highlight that our approach (T-SAM) outperforms SD and LB on complex syntactic prompts, where other baselines are inapplicable. While LB demonstrates improvements over SD in the color, shape, and material", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "categories, which are closely related to attribute-binding tasks, it fails to enhance accuracy in the activity and counting categories. This underscores LB's inability to capture diverse word relationships, even when using external text parsers. In contrast, our method shows improvements over SD in nearly all categories including color, shape, and activity, except for attribute (e.g., properties such as large, small, young), which differ from the term \"attribute\" used in this study, demonstrating its versatility across diverse word relationships. Additionally, it achieves higher CLIP scores than the baselines, proving its superior semantic alignment.", + "bbox": [ + 511, + 733, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8037", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0bbb804223c22d3d6f65c519875237e3a5743aa3a11df2d0a0f521c32828768c.jpg", + "table_caption": [ + "Table 2. Average CLIP similarity scores between text prompts and images generated with 64 different seeds. The base model is SD." + ], + "table_footnote": [], + "table_body": "
External Info.Image-Text PromptPrompt-Caption
Full PromptMinimum Object
ObjectsAnimals-ObjectsObjectsAnimals-ObjectsObjectsAnimals-Objects
SDX0.340.340.250.260.760.80
LB✓0.360.350.270.270.800.83
CONFORM✓0.360.360.280.280.810.85
A&E✓0.360.350.270.260.810.83
T-SAMX0.360.370.280.280.810.85
", + "bbox": [ + 91, + 99, + 906, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/56cf01205037dee4d8f1dacfa8d2c20650a98dff824cf50a16cbaab692a53eb5.jpg", + "table_caption": [ + "Table 3. CLIP similarity scores on prompts from Objects and Animals-Objects using PixArt- $\\alpha$ as the base model. CLIP-I (CLIP-T) refers to image-text (text-text) CLIP similarity." + ], + "table_footnote": [], + "table_body": "
Prompt SetMethodCLIP-TCLIP-I (Full, Min)
ObjectsPixArt-α0.810.36, 0.27
T-SAM0.820.37, 0.28
Animals-ObjectsPixArt-α0.850.36, 0.27
T-SAM0.860.37, 0.28
", + "bbox": [ + 91, + 316, + 496, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/50b673a935f082a57901dc6907a19c2fb74bfe339c454d5255c3c5512c94fe24.jpg", + "image_caption": [ + "Figure 9. Qualitative comparison using prompts from MSCOCO contained in TIFA benchmark. Our approach successfully generates multiple elements, including a sewing machine in the first prompt, people in the second, popsicle in the third, and cow in the fourth. SD misses some elements and LB generates identical images to SD when no relations are extracted from the text parser." + ], + "image_footnote": [], + "bbox": [ + 114, + 422, + 460, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Objects and Animals-Objects, T-SAM on SD performs comparably to the state-of-the-art CONFORM, which, unlike our approach, requires manually defined token indices for positive and negative groups, as shown in Table 2. Our method outperforms LB and A&E, which also rely on external inputs. This demonstrates extracting syntactic information from text self-attention maps can be more effective than relying on text parsers or manually selecting tokens for optimization. An additional ablation study is provided in Ap", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pendix B.2. Moreover, Table 3 shows that T-SAM is also effective on PixArt- $\\alpha$ with *Attend-n-Excite* prompts, demonstrating the generalizability of our method to a different text encoder (T5-XXL) and model architecture (DiT).", + "bbox": [ + 511, + 270, + 903, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results. In Figure 8, T-SAM either outperforms or performs comparably to the baselines in Objects and Animals-Objects. SD often omits objects (e.g., clock, apple, or bow) or misbinds attributes (e.g., blue clock). CONFORM and LB improve over SD but have limitations. LB is restricted to attribute-binding and fails on prompts without modifiers (e.g., An elephant with a bow), producing results identical to SD CONFORM can impose overly rigid separations, as seen in the first example. T-SAM, by leveraging smoother linguistic cues from text attention maps—rather than binary positive/negative pairs—achieves broader applicability and more consistent improvements. See Figure 9 for more comparison of images generated by T-SAM and the baselines, showcasing effectiveness of T-SAM with complex prompts, such as MSCOCO captions from the TIFA benchmark.", + "bbox": [ + 511, + 330, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For PixArt- $\\alpha$ , Figure D illustrates the effectiveness of our method. Although the base model is already strong compared to SD, T-SAM further improves its performance. User study. The user study result in Table D show that T-SAM on SD outperforms LB and achieves performance comparable to CONFORM. Additionally, Table E shows T-SAM on PixArt- $\\alpha$ surpasses the base PixArt- $\\alpha$ model.", + "bbox": [ + 511, + 571, + 908, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To enhance fidelity in T2I diffusion models, we improved cross-attention maps by aligning their similarity matrix with text self-attention maps. Our approach is based on two insights: (1) similar text embeddings produce similar cross-attention maps, but (2) syntactic relations are missed in embeddings but captured by text self-attention maps. T-SAM enabled cross-attention to better capture syntactic structure, significantly improving text-to-image fidelity across diverse sentence structures, without requiring external tools. Nonetheless, our method has limitations such as inaccuracies in object counting and potential noise introduced by text attention maps, leaving room for future improvement.", + "bbox": [ + 511, + 717, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8038", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by NIH (1U01CA269192).", + "bbox": [ + 89, + 114, + 483, + 145 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 157, + 187, + 174 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2283-2293, 2023. 1, 2, 3, 4", + "[2] Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Alignit: Enhancing prompt alignment in customization of text-to-image models. arXiv preprint arXiv:2406.18893, 2024. 3", + "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2, 3, 4, 5, 6", + "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 2, 3, 4, 6", + "[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart $\\cdot$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2", + "[6] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 6", + "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 1, 2", + "[8] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 3, 4", + "[9] Xiangming Gu, Tianyu Pang, Chao Du, Qian Liu, Fengzhuo Zhang, Cunxiao Du, Ye Wang, and Min Lin. When attention sink emerges in language models: An empirical view. arXiv preprint arXiv:2410.10781, 2024. 3", + "[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 3", + "[11] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional" + ], + "bbox": [ + 93, + 183, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neural networks and incremental parsing. To appear, 7(1): 411-420, 2017. 6", + "[12] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 3", + "[13] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 6, 5", + "[14] Jiaxiu Jiang, Yabo Zhang, Kailai Feng, Xiaohe Wu, and Wangmeng Zuo. Mc2: Multi-concept guidance for customized multi-concept generation. arXiv preprint arXiv:2404.05268, 2024.3", + "[15] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7701-7711, 2023. 3", + "[16] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 6, 5", + "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5", + "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[19] Oscar Manas, Pietro Astolfi, Melissa Hall, Candace Ross, Jack Urbanek, Adina Williams, Aishwarya Agrawal, Adriana Romero-Soriano, and Michal Drozdal. Improving text-to-image consistency via automatic prompt optimization. arXiv preprint arXiv:2403.17804, 2024. 3", + "[20] Arash Marioriyad, Mohammadali Banayeeanzade, Reza Abbasi, Mohammad Hossein Rohban, and Mahdieh Soleymani Baghshah. Attention overlap is responsible for the entity missing problem in text-to-image diffusion models! arXiv preprint arXiv:2410.20972, 2024. 3", + "[21] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9005–9014, 2024. 1, 2, 3, 4, 6", + "[22] OpenAI. Openai gpt-3 api [gpt-3.5-turbo], 2024. Available at: https://platform.openai.com. 6, 5", + "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8039", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Zipeng Qi, Lichen Bai, Haoyi Xiong, et al. Not all noises are created equally: Diffusion noise selection and optimization. arXiv preprint arXiv:2407.14041, 2024. 3", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6", + "[26] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 3", + "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[28] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 6, 5", + "[30] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 6", + "[31] Dazhong Shen, Guanglu Song, Zeyue Xue, Fu-Yun Wang, and Yu Liu. Rethinking the spatial inconsistency in classifier-free diffusion guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9370-9379, 2024. 3", + "[32] Mingjie Sun, Xinlei Chen, J Zico Kolter, and Zhuang Liu. Massive activations in large language models. arXiv preprint arXiv:2402.17762, 2024. 2, 3, 5", + "[33] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[34] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. arXiv preprint arXiv:2312.03626, 2023. 3", + "[35] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 2, 3, 4, 5", + "[36] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291-26303, 2023. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Mingyang Yi, Aoxue Li, Yi Xin, and Zhenguo Li. Towards understanding the working mechanism of text-to-image diffusion model. arXiv preprint arXiv:2405.15330, 2024. 3, 5", + "[38] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv preprint arXiv:2404.01154, 2024. 3", + "[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 6", + "[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 2, 5", + "[41] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Tiviatis Sim, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. arXiv preprint arXiv:2403.06381, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 373 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8040", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_model.json b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1bff50e6c39b441735092786aacb5c7abbae3376 --- /dev/null +++ b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_model.json @@ -0,0 +1,2508 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.13, + 0.877, + 0.178 + ], + "angle": 0, + "content": "Text Embedding is Not All You Need: Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.204, + 0.686, + 0.24 + ], + "angle": 0, + "content": "Jeeyung Kim*, Erfan Esmaeili*, and Qiang Qiu Purdue University" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.242, + 0.653, + 0.258 + ], + "angle": 0, + "content": "{jkim17,efakhabi,qqiu}@purdue.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.328, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.732 + ], + "angle": 0, + "content": "In text-to-image diffusion models, the cross-attention map of each text token indicates the specific image regions attended. Comparing these maps of syntactically related tokens provides insights into how well the generated image reflects the text prompt. For example, in the prompt, \"a black car and a white clock\", the cross-attention maps for \"black\" and \"car\" should focus on overlapping regions to depict a black car, while \"car\" and \"clock\" should not. Incorrect overlapping in the maps generally produces generation flaws such as missing objects and incorrect attribute binding. Our study makes the key observations investigating this issue in the existing text-to-image models: (1) the similarity in text embeddings between different tokens—used as conditioning inputs—can cause their cross-attention maps to focus on the same image regions; and (2) text embeddings often fail to faithfully capture syntactic relations already within text attention maps. As a result, such syntactic relationships can be overlooked in cross-attention module, leading to inaccurate image generation. To address this, we propose a method that directly transfers syntactic relations from the text attention maps to the cross-attention module via a test-time optimization. Our approach leverages this inherent yet unexploited information within text attention maps to enhance image-text semantic alignment across diverse prompts, without relying on external guidance. Our project page and code are available at: https://t-sam-diffusion.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.761, + 0.223, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Recent diffusion models can generate images from diverse text prompts [4, 7, 29]. However, semantic discrepancies often arise between the text and generated images, raising problems such as missing objects—where certain elements are overlooked—and attribute mis-binding—where attributes are incorrectly assigned to subjects." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.293, + 0.907, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.613 + ], + "angle": 0, + "content": "Figure 1. The overview of our method. We leverage text self-attention matrix and optimize the latent noise \\((z_{t})\\) by minimizing the distance between the cross-attention similarity matrix (S) and the text self-attention matrix (T). This encourages integrating syntactic relationships into text-to-image diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.886 + ], + "angle": 0, + "content": "Prior studies [10] demonstrated that the cross-attention map of each token in text-to-image (T2I) diffusion models highlights the attended regions in the image and provides clues about the spatial placement of elements corresponding to the tokens. In particular, [1, 21, 28] suggest that spatial alignment in cross-attention maps among related words influences the fidelity of images to text prompts in Stable Diffusion (SD) [29], as we also demonstrate in Section 4.1. For instance, in the prompt a black car and a white clock, if the cross-attention maps for car and clock overlap, unique token contributions can dilute, potentially omitting one object. Conversely, if the maps for black (or white) and car (or clock) diverge too much, attribute mis-binding can occur. This implies syntactically related words should ideally have spatially aligned cross-attention maps, as discussed in [8, 28]. However, the factors determining this spatial alignment between the maps remain poorly understood." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.887, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In our study, we first investigate the factors that con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.221, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8031" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.075, + 0.315, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.22, + 0.214, + 0.232 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.075, + 0.528, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.22, + 0.431, + 0.232 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.078, + 0.7, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.081, + 0.902, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.22, + 0.731, + 0.232 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.244, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Figure 2. For the analysis, we use SD v1.5 as the base model and adopt the prompt sets from [3], structured as “[attribute_1] [object_1] and [attribute_2] [object_2]”, “[object_1] and/with [object_2]” or “[object_1] and [object_2] [object_2]”. (a) Comparison of the cosine similarity of text embeddings with that of the corresponding cross-attention maps at denoising step 1, with pairs of tokens (object_i, object_j), where \\( i \\neq j \\), and pairs of tokens (attribute_m, object_n) for both \\( m = n \\) and \\( m \\neq n \\). As text embeddings become more similar, their cross-attention maps get similar. (b) The distributions of text embedding similarity between i) Bound tokens—(attribute_i, object_i) for \\( i = 1, 2 \\), and ii) Unbound tokens—(object_i, object_2). The distributions show no discernible difference, indicating text embeddings do not effectively represent the syntactic relationships. (c) Comparison of text embedding similarity (left) and the text self-attention map power by 3 (right) for the prompt a black car and a white clock. In the self-attention maps (T), clock attends more to white, unlike the text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.375, + 0.482, + 0.512 + ], + "angle": 0, + "content": "tribute to the spatial alignment of cross-attention maps across different tokens, which can ultimately affect the accuracy of T2I generation. We reveal that text embeddings, which function as keys in the cross-attention modules, play a pivotal role in determining the similarity of these maps, as illustrated in Figure 2a. Specifically, as the text embeddings for different words become more similar, the cosine similarity of their corresponding cross-attention maps increases. This effect is noticeable from the initial denoising step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Considering that (i) cross-attention maps should capture syntactic relationships between words and that (ii) the spatial alignment of cross-attention maps is influenced by the text embeddings, we question whether text embeddings accurately capture the linguistic structure in text prompts. Our findings suggest they do not, as illustrated in Figure 2b and the left of Figure 2c. Syntactically related words in prompts (e.g., black-car and white-clock) do not necessarily yield similar text embeddings, as detailed in Section 4.1. This finding is further supported by [40], which argues that CLIP [25] often behaves like a bag-of-words model. As a result, the cross-attention maps, derived from text embeddings, can likely fail to faithfully reflect the syntactic relationships. In the end, the T2I diffusion models can often generate images that inaccurately represent the prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.868 + ], + "angle": 0, + "content": "To address the above issue, prior studies [1, 8, 21, 28] resort to external sources to obtain syntactic information within text prompts and regularize cross-attention maps to incorporate these relationships. Specifically, [8, 28] employ external text parsers to obtain linguistic structure, while [1, 21] rely on human intervention to manually group tokens based on syntactic relationships. However, these methods are limited by their dependence on external inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Do we really need to rely on external sources to obtain syntactic relations within a sentence? Remarkably, T2I" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.556 + ], + "angle": 0, + "content": "diffusion models inherently capture syntactic relationships within prompts through the self-attention maps in their text encoder. In the encoder's self-attention module, each token pays more attention to related words, leading to encoding the entire sentence effectively. As shown in the right of Figure 2c, the text self-attention maps exhibit higher attention scores between the syntactically related words. However, this information is weakly encoded in the text encoder outputs (i.e., text embeddings). We conjecture this is due to the text attention module's strong focus on the \\(<\\text{bos}>\\) token, known as attention sink [32, 35], which minimizes the influence of other tokens, as further discussed in Section 4.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.909, + 0.815 + ], + "angle": 0, + "content": "As illustrated in Figure 1, our approach reuses the self-attention maps from the text encoder and directly transfers the syntax information to the diffusion models. We first define a similarity matrix whose values indicate the similarity between pairs of cross-attention maps. Then, we update the latent noise in diffusion models to minimize the distance between the cross-attention similarity matrix and the text self-attention matrix during inference. This allows for the seamless integration of syntactic relationships into the diffusion process. By guiding cross-attention modules to better capture contextual relationships, our method ultimately produces images accurately reflecting the intended meaning of prompts. We leverage overlooked information already embedded in T2I models, offering two key advantages: (i) Self-contained: no need for external inputs like text parsers or manual token indices; and (ii) Generalizable: effective across diverse sentence structures." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.829, + 0.655, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Text-to-image generation. Text-to-image (T2I) generation [4, 5, 7, 23, 27, 30] is commonly based on latent diffusion models [29] where text data is processed via a text encoder" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8032" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.182 + ], + "angle": 0, + "content": "[25, 26]. An important problem in T2I models is the lack of full correspondence between the input text and the image. Although diverse methods [2, 8, 12, 19, 24, 31, 38] are proposed to diagnose and fix this problem, a common theme in prior works is adjusting the cross-attention maps in inference time, following the seminal work [10]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.408 + ], + "angle": 0, + "content": "Cross-attention control for improved semantic alignment. Several defects in cross-attention maps hinder semantic alignment, including attention dominance [41]—one token getting huge attention weight, and attention leakage/overlap [20, 36]—attention weights not respecting the spatial boundaries of the intended objects. In particular, one line of work employs contrastive objectives to guide attention optimization [1, 14, 21, 28]. For example, CONFORM [21] uses prompts with distinct groups, where tokens in the same (opposite) group are treated as positive (negative) pairs. Similarly, [28] uses text parser models to identify such pairs. Other approaches, like those in [15, 34], optimize attention maps with spatial guidance (e.g., segmentation maps, masks, or layouts). However, these methods rely on external resources for text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.499 + ], + "angle": 0, + "content": "Attention sinks. The attention weights in pretrained transformers tend to be heavily biased towards special tokens such as \\( \\langle \\text{bos} \\rangle \\), \\( \\langle \\text{eos} \\rangle \\), and punctuation tokens, as studied in (vision-)language models [9, 32, 35]. The T2I diffusion model literature [3, 37] also notes the focus of attention scores on the \\( \\langle \\text{bos} \\rangle \\) token in the CLIP text encoder." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.513, + 0.23, + 0.529 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.584 + ], + "angle": 0, + "content": "In this section, we briefly review the structure of text-to-image diffusion models including the text encoder and the cross-attention module [4, 29]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Text encoder. To condition the text data for the diffusion process, a text encoder is needed. The input text is first tokenized and then converted into dense vectors. The dense vectors are then processed via a series of multi-head self-attention layers [33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.737 + ], + "angle": 0, + "content": "At a given layer \\(\\ell\\) in the text encoder, let us denote the key by \\(\\mathbf{e}_i^{(\\ell)}\\in \\mathbb{R}^{H_{\\mathrm{e}}D_{\\mathrm{e}}}\\) \\(i = 1,\\dots ,s\\) where \\(s\\) is the key sequence length, \\(D_{\\mathrm{e}}\\) is the embedding dimension per head and \\(H_{\\mathrm{e}}\\) is the number of heads in the text encoder. The corresponding self-attention matrix \\(T^{(\\ell ,h)}\\in \\mathbb{R}^{s\\times s}\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.748, + 0.483, + 0.782 + ], + "angle": 0, + "content": "\\[\nT _ {i j} ^ {(\\ell , h)} = \\frac {\\exp (\\omega_ {i j})}{\\sum_ {k} \\exp (\\omega_ {i k})}, \\omega_ {i j} := \\mathbf {e} _ {i} ^ {(\\ell) \\top} W _ {\\mathrm {e n}} ^ {(\\ell , h)} \\mathbf {e} _ {j} ^ {(\\ell)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.825 + ], + "angle": 0, + "content": "where \\( W_{\\mathrm{en}}^{(\\ell,h)} \\in \\mathbb{R}^{H_{\\mathrm{c}}D_{\\mathrm{c}} \\times H_{\\mathrm{c}}D_{\\mathrm{c}}} \\) is a pretrained matrix at head \\( h \\) in layer \\( \\ell \\) of the text encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.856 + ], + "angle": 0, + "content": "In our proposed method, we use the self-attention matrix averaged over layers and heads:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.863, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\nT ^ {\\prime} = \\frac {1}{L _ {\\mathrm {e}} H _ {\\mathrm {e}}} \\sum_ {\\ell = 1} ^ {L _ {\\mathrm {e}}} \\sum_ {h = 1} ^ {H _ {\\mathrm {e}}} T ^ {(\\ell , h)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.136 + ], + "angle": 0, + "content": "Due to the high attention probabilities assigned to \\( \\langle \\text{bos} \\rangle \\) and \\( \\langle \\text{eos} \\rangle \\) tokens, we remove their corresponding values and re-normalize each row as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.141, + 0.907, + 0.176 + ], + "angle": 0, + "content": "\\[\nT _ {i j} = \\frac {T ^ {\\prime} {} _ {i j}}{\\sum_ {m = 2} ^ {i} T ^ {\\prime} {} _ {i m}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.181, + 0.907, + 0.226 + ], + "angle": 0, + "content": "We denote the final output of the text encoder by \\(\\mathbf{k}_i\\in\\) \\(\\mathbb{R}^{H_{\\mathrm{e}}D_{\\mathrm{e}}}\\) where \\(i = 1,\\dots ,s\\) , which used as conditioning inputs in T2I diffusion models. We call \\(\\mathbf{k}_i\\) as text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.227, + 0.906, + 0.302 + ], + "angle": 0, + "content": "Denoising latent variables. Latent diffusion models generate latent tensors \\((z_0)\\) of an image. A latent diffusion model \\(D_{\\theta}\\) learns to simulate the denoising process: starting from Gaussian noise \\(z_{\\tau}\\), the model iteratively reconstruct \\(z_0\\) by predicting the noise at step \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.309, + 0.907, + 0.327 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = z _ {t} - D _ {\\theta} \\left(z _ {t}; \\left\\{\\mathbf {k} _ {i} \\right\\}\\right), \\quad 0 < t \\leq \\tau . \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.332, + 0.906, + 0.439 + ], + "angle": 0, + "content": "Cross-attention module. In T2I latent diffusion models, the interaction between text and image data is performed by multi-head cross-attention. In the cross-attention layer \\(\\ell\\), we define the query, \\(\\mathbf{q}_a^{(\\ell)} \\in \\mathbb{R}^{H_cD_c}\\), where \\(a = 1, \\dots, N_c\\), and \\(N_c\\) is the query sequence length at the given cross-attention layer. \\(D_c\\) is the hidden dimension per head, and \\(H_c\\) is the number of heads of the cross-attention layer." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.906, + 0.473 + ], + "angle": 0, + "content": "The cross-attention map \\( A^{(\\ell,h)} \\in \\mathbb{R}^{N_{\\mathrm{c}} \\times s} \\) at head \\( h \\) with elements \\( A_{ai}^{(\\ell,h)} \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.475, + 0.905, + 0.525 + ], + "angle": 0, + "content": "\\[\nA _ {a i} ^ {(\\ell , h)} := \\frac {\\exp (\\Omega_ {a i})}{\\sum_ {j = 1} ^ {s} \\exp (\\Omega_ {a j})}, \\Omega_ {a i} := \\mathbf {q} _ {a} ^ {(\\ell) \\top} W _ {\\mathrm {c}} ^ {(\\ell , h)} \\mathbf {k} _ {i} ^ {(\\ell)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.529, + 0.906, + 0.562 + ], + "angle": 0, + "content": "Here, \\( W_{\\mathfrak{c}}^{(\\ell,h)} \\in \\mathbb{R}^{H_{\\mathfrak{c}}D_{\\mathfrak{c}} \\times H_{\\mathfrak{c}}D_{\\mathfrak{c}}} \\) is a pretrained matrix at head \\( h \\) in layer \\( \\ell \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.563, + 0.906, + 0.607 + ], + "angle": 0, + "content": "Cross-attention similarity matrix. In our study, we use the cross-attention maps averaged over heads and over all layers for which \\( N_{\\mathrm{c}} = M \\):" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.613, + 0.905, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\mathsf {A} = \\frac {1}{L _ {M} H _ {\\mathrm {c}}} \\sum_ {\\ell = 1} ^ {L _ {M}} \\sum_ {h = 1} ^ {H _ {\\mathrm {c}}} A ^ {(\\ell , h)}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.658, + 0.906, + 0.703 + ], + "angle": 0, + "content": "Here, \\(L_{M}\\) is the number of cross-attention layers with \\(N_{\\mathrm{c}} = M\\). Based on the average cross-attention map above, we define the similarity matrix \\(\\mathsf{S} \\in \\mathbb{R}^{s \\times s}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.708, + 0.905, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S} _ {i j} := \\frac {\\mathrm {C} _ {i j}}{\\sum_ {k = 1} ^ {s} \\mathrm {C} _ {i k}}, \\mathrm {C} _ {i j} := \\frac {\\sum_ {a = 1} ^ {N _ {\\mathrm {c}}} \\mathrm {A} _ {a i} \\mathrm {A} _ {a j}}{\\left(\\sum_ {a = 1} ^ {N _ {\\mathrm {c}}} \\mathrm {A} _ {a i} ^ {2}\\right) ^ {\\frac {1}{2}} \\left(\\sum_ {a = 1} ^ {N _ {\\mathrm {c}}} \\mathrm {A} _ {a j} ^ {2}\\right) ^ {\\frac {1}{2}}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.772, + 0.906, + 0.802 + ], + "angle": 0, + "content": "\\(C_{ij}\\) denotes cosine similarity between maps corresponding to \\(i\\) and \\(j\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.812, + 0.906, + 0.848 + ], + "angle": 0, + "content": "4. Understanding and Resolving Text-to-Image Semantic Discrepancy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we examine the cross-attention maps in Stable Diffusion (SD) [29] to identify factors that lead to semantic discrepancies between generated images and text" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8033" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "prompts. We provide additional results with \\(\\mathrm{PixArt - \\alpha}\\) [4] in Appendix B.1. Based on these findings, we introduce an approach to enhance the fidelity of T2I diffusion models by leveraging text self-attention maps to regularize cross-attention maps." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.178, + 0.483, + 0.195 + ], + "angle": 0, + "content": "4.1. Why Do Generated Images Misrepresent Text?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.202, + 0.484, + 0.489 + ], + "angle": 0, + "content": "Our findings are based on the premise that cross-attention maps should capture syntactic relationships between words for accurate T2I generation, as supported by prior works [1, 8, 21, 28]. We identify a key factor contributing to incorrect relations in cross-attention maps, resulting in less faithful image generation: While text embedding \\((\\mathbf{k}_i)\\) similarity strongly correlates with the cross-attention similarity matrix C, it insufficiently reflects syntactic bindings in the text prompt. This highlights a fundamental issue in text-image semantic discrepancy: the similarity in text embeddings lack the syntax information necessary for accurate image generation. We reveal that text self-attention maps (T) effectively capture syntactic information, but this is not sufficiently encoded in text embeddings. The absence of the information in text embeddings can be due to an artifact in the attention module, known as attention sink [35], where attention weights are biased toward the \\(<\\mathrm{bos}>\\) token as detailed later. These findings motivate the solution presented in the following section." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.535 + ], + "angle": 0, + "content": "For the following empirical analysis, we use the prompt sets introduced in [3], containing prompts structured as the following three categories:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.452, + 0.552 + ], + "angle": 0, + "content": "(i) \\(\\left[\\mathsf{attribute}_1\\right]\\left[\\mathsf{object}_1\\right]\\) and \\(\\left[\\mathsf{attribute}_2\\right]\\left[\\mathsf{object}_2\\right]\\)," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.553, + 0.346, + 0.568 + ], + "angle": 0, + "content": "(ii) \\(\\left[\\mathrm{object}_1(\\mathrm{animal})\\right]\\) with \\(\\left[\\mathrm{object}_2\\right]\\)," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.568, + 0.437, + 0.584 + ], + "angle": 0, + "content": "(iii) [object1(animal)] and [attribute2][object2]." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.535, + 0.452, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Prompt sets in the above formats enable us to focus on two key cases of text-image semantic discrepancies: missing objects and attribute mis-binding. In the following, we refer to the group of syntactically bound words (tokens) as pairs of \\(([\\text{attribute}_i], [\\text{object}_i])\\), \\(i = 1, 2\\), and the group of syntactically unbound words (tokens) as \\(([\\text{attribute}_i], [\\text{object}_j])\\) and \\(([\\text{object}_1], [\\text{object}_2])\\), where \\(i \\neq j\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.809 + ], + "angle": 0, + "content": "We revisit the role of cross-attention maps in text-image semantic coherence, where spatial overlap or separation is crucial, as discussed in [1, 8, 21, 28]. Figure 3 illustrates this effect: On the left, overlapping attention maps for syntactically unbound words lead to object missing, while greater overlap for syntactically bound words enhances attribute binding. This suggests that syntactic associations should be reflected in cross-attention maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To statistically assess cross-attention map's impact, we analyze categories (ii) and (iii) of the aforementioned prompt sets. Assuming spatial overlap is measurable by the cosine similarity matrix (C) in eq.(7), we compute cosine similarities for the attention maps of \\(object_{1}\\) and \\(object_{2}\\) in missing objects, and for \\(attribute_{2}\\) and \\(attribute_{2}\\) in attribute" + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.084, + 0.628, + 0.116 + ], + "angle": 0, + "content": "1) Missing object A bear and a red balloon" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.12, + 0.61, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.223, + 0.555, + 0.232 + ], + "angle": 0, + "content": "bear" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.19, + 0.607, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.223, + 0.605, + 0.232 + ], + "angle": 0, + "content": "balloon" + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.1, + 0.699, + 0.119 + ], + "angle": 0, + "content": "A frog with a bow" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.12, + 0.705, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.223, + 0.647, + 0.232 + ], + "angle": 0, + "content": "frog" + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.084, + 0.835, + 0.098 + ], + "angle": 0, + "content": "2) Attribute binding" + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.104, + 0.809, + 0.127 + ], + "angle": 0, + "content": "A green backpack and a purple bench" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.128, + 0.808, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.108, + 0.841, + 0.118 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.119, + 0.859, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.163, + 0.85, + 0.171 + ], + "angle": 0, + "content": "green" + }, + { + "type": "image_caption", + "bbox": [ + 0.862, + 0.108, + 0.889, + 0.118 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.862, + 0.119, + 0.889, + 0.127 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.178, + 0.859, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.218, + 0.857, + 0.227 + ], + "angle": 0, + "content": "backpack" + }, + { + "type": "image", + "bbox": [ + 0.862, + 0.178, + 0.911, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.869, + 0.218, + 0.902, + 0.227 + ], + "angle": 0, + "content": "bench" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.245, + 0.906, + 0.316 + ], + "angle": 0, + "content": "Figure 3. The generated images and cross-attention maps (A) for the specific tokens from SD v1.5. This illustrate the importance of spatial alignment in cross-attention maps for accurate image generation. Divergent (overlapping) cross-attention maps for syntactically unbound (bound) words enhances text-to-image fidelity." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.328, + 0.707, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.442, + 0.622, + 0.452 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.326, + 0.897, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.442, + 0.813, + 0.452 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.452, + 0.906, + 0.535 + ], + "angle": 0, + "content": "Figure 4. Comparison for the distributions of cosine similarity between cross-attention maps (at denoising step 10). (a) The cases with one missing object-incorrect and two objects present-correct. (b) The cases with incorrect and correct attribute binding. Correct instances are more frequent when the cosine similarity is low for objects presence and high for attribute binding." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.545, + 0.648, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.546, + 0.773, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.546, + 0.902, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.642, + 0.906, + 0.698 + ], + "angle": 0, + "content": "Figure 5. Correlation between the cosine similarity of text embeddings and that of cross-attention maps across denoising steps \\((t = 1,21,50)\\). Similar text embeddings generally lead to similar cross-attention maps, with the correlation weakening over time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.793 + ], + "angle": 0, + "content": "binding cases. Then, we compare cosine similarity distributions for correct vs. incorrect images on both cases. See Appendix B.1 for setup details. Figure 4 shows that lower cosine similarity tends to correlate with object presence, while higher similarity supports more accurate bindings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This observation motivates us to examine the factor contributing to similarity in cross-attention maps across tokens. Finding 1: The cosine similarity of text embeddings has a large correlation with cross-attention similarity matrix C. Figure 5 shows there is a correlation between the cosine similarity of text embedding and C, which persists throughout the final denoising steps. This indicates that similar text" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8034" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.108 + ], + "angle": 0, + "content": "embeddings can result in overlapping cross-attention maps." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.111, + 0.406, + 0.126 + ], + "angle": 0, + "content": "Next, we justify this finding mathematically." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.145, + 0.485, + 0.209 + ], + "angle": 0, + "content": "Proposition 1. If \\( A^{(\\ell, h)} \\in \\mathbb{R}^{N_c \\times s} \\) is a cross-attention map defined in eq. (5), then under the assumptions i, ii, and iii described in Appendix A, the cosine similarity matrix can be written in terms of key vectors \\( \\mathbf{k}_i^{(\\ell, h)} \\in \\mathbb{R}^{H_c D_c} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.213, + 0.483, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\cos \\left(A _ {i} ^ {(\\ell , h)}, A _ {j} ^ {(\\ell , h)}\\right) = \\\\ \\exp \\left(- \\frac {1}{2} \\left(\\mathbf {k} _ {i} - \\mathbf {k} _ {j}\\right) ^ {\\top} W ^ {2} \\left(\\mathbf {k} _ {i} - \\mathbf {k} _ {j}\\right)\\right), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.271, + 0.484, + 0.32 + ], + "angle": 0, + "content": "up to terms of at least \\(\\mathcal{O}(1 / \\sqrt{N_c})\\) and \\(\\mathcal{O}(\\epsilon)\\), where \\(W^2 \\coloneqq W_c^{(\\ell,h)^\\top}\\Sigma^{(\\ell)}W_c^{(\\ell,h)}\\) and \\(\\Sigma^{(\\ell)} \\in \\mathbb{R}^{H_cD_c \\times H_cD_c}\\) is the covariance matrix of query vectors." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.34, + 0.325, + 0.355 + ], + "angle": 0, + "content": "Refer to Appendix A for the proof." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.484, + 0.419 + ], + "angle": 0, + "content": "We empirically and mathematically showed that the similarity in text embeddings \\(\\mathbf{k}_i\\) influences cross-attention maps. Next, we evaluate whether the similarity of these embeddings reflects syntax information in text inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.558 + ], + "angle": 0, + "content": "Finding 2-1: There is no significant correlation between word syntactic bindings and text embedding similarity. Prior study [40] suggests CLIP embeddings [25], used in SD, behave like a bag-of-words model, ignoring word relationships. We also show in Figure 2b the similarity in CLIP text embeddings does not correlate with syntactic bindings. Specifically, we expect close embeddings for syntactically bound tokens and distant embeddings for unbound tokens, yet the distributions lack separation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Finding 2-2: Text self-attention maps do have syntax information. We examine the text encoder that produces text embeddings. Interestingly, the self-attention maps in the encoder capture syntactic relationships, as shown in Figure 6a. These maps reveal higher similarity between syntactically bound tokens and lower similarity between unbound tokens. For more complex prompt structures, see the Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Finding 2-3: The attention sink can contribute to why text embeddings lack the syntax information. Why do text embeddings lack relational information despite being derived from multiple self-attention modules? We attribute this gap to attention sink [32, 35], where attention scores are concentrated on a few tokens. In CLIP's text encoder, attention is mainly focused on the \\( \\langle \\text{bos} \\rangle \\) token, as discussed in [3, 37] and shown in Figure 6b. We hypothesize the focus on the \\( \\langle \\text{bos} \\rangle \\) token can limit the transfer of relational information from self-attention maps \\( (T) \\) to text embeddings, as attention scores for other tokens remain much smaller, minimizing their influence in each self-attention layer." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To provide a mathematical justification, we express the difference between the output and input of one block of the" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.084, + 0.747, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.246, + 0.651, + 0.258 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.09, + 0.882, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.246, + 0.825, + 0.258 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.263, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Figure 6. (a) The distributions of text self-attention value \\((\\mathsf{T}_{ij}^{\\prime})\\) in eq.(2)) for bound tokens (attribute \\(m\\), object \\(m\\)) and unbound tokens (attribute \\(m\\), object \\(n\\)) / (object \\(1\\), object \\(2\\)), where \\(m, n \\in \\{1, 2\\}\\). The separate distributions indicate text self-attention maps can indeed represent the syntactic relationships. (b) Comparison of text self-attention probability histograms between token and non tokens on 100 prompts: The probability allocated to is on average 20 times larger than that of other tokens." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.4, + 0.773, + 0.415 + ], + "angle": 0, + "content": "self-attention module for a token \\(e_i\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.42, + 0.906, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathbf {o} _ {i} ^ {(\\ell , h)} = \\sum_ {j = 1} ^ {i} T _ {i j} ^ {(\\ell , h)} W _ {\\mathrm {v}} ^ {(\\ell , h)} \\mathbf {e} _ {j} ^ {(\\ell)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.467, + 0.907, + 0.547 + ], + "angle": 0, + "content": "where \\( W_{\\mathrm{v}}^{(\\ell,h)} \\in \\mathbb{R}^{D_{\\mathrm{e}} \\times H_{\\mathrm{e}} D_{\\mathrm{e}}} \\) is a parameter, \\( \\mathbf{e}_i \\in \\mathbb{R}^{H_{\\mathrm{e}} D_{\\mathrm{e}}} \\), and \\( T_{ij}^{(\\ell,h)} \\) is the self-attention matrix in eq. (1). We consider the situation where attention sink occurs, that is the attention weights for the token is much higher than the rest of the sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.553, + 0.906, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\epsilon = \\frac {\\sum_ {j \\neq 1} ^ {i} T _ {i j}}{T _ {i 1}} \\ll 1, \\quad i = 2, \\dots , s. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.592, + 0.848, + 0.607 + ], + "angle": 0, + "content": "In Appendix A, we prove the following statement:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.614, + 0.805, + 0.629 + ], + "angle": 0, + "content": "Proposition 2. Define matrix \\(R \\in \\mathbb{R}^{s \\times s}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.633, + 0.905, + 0.656 + ], + "angle": 0, + "content": "\\[\nR _ {i j} = \\mathbf {e} _ {i} ^ {(\\ell) \\top} W _ {v} ^ {(\\ell , h) \\top} W _ {v} ^ {(\\ell , h)} \\mathbf {e} _ {j} ^ {(\\ell)}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.725, + 0.674 + ], + "angle": 0, + "content": "and suppose it has the property" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.677, + 0.905, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\frac {\\left| R _ {m n} \\right|}{R _ {1 1}} \\sim \\mathcal {O} (1 / \\epsilon), \\quad \\frac {\\left| R _ {1 m} \\right|}{R _ {1 1}} \\sim \\mathcal {O} (1), \\quad 1 < m, n \\leq s, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.721, + 0.903, + 0.739 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}_1^{(\\ell)}\\) is the bos embedding. Then the following holds:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.743, + 0.905, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\cos \\left(\\mathbf {o} _ {i} ^ {(\\ell , h)}, \\mathbf {o} _ {j} ^ {(\\ell , h)}\\right) = 1 - \\mathcal {O} (\\epsilon). \\tag {13}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.889, + 0.77, + 0.905, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.907, + 0.883 + ], + "angle": 0, + "content": "This suggests that cosine similarity between token vectors \\((\\mathbf{o}_i^{(\\ell ,h)})\\) -potentially influencing cross-attention maps remains barely changed across text self-attention layers due to attention sink on the. \\(\\langle \\mathrm{bos}\\rangle\\) token. In other words, the attention sink can hinder the accurate encoding of self-attention maps into embeddings." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our findings reveal text embedding alone are insufficient" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8035" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "for generating semantically aligned images. On the other hand, we notably show the potential of transferring neglected syntactic information from text self-attention maps to the cross-attention to enhance T2I semantic alignment." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.16, + 0.473, + 0.177 + ], + "angle": 0, + "content": "4.2. Text Self-Attention Maps (T-SAM) Guidance" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.183, + 0.483, + 0.318 + ], + "angle": 0, + "content": "In the previous section, we show text self-attention maps capture syntax information within a sentence. Building on this insight, we propose leveraging the self-attention maps within the text encoder—a component of diffusion models—to enhance cross-attention maps. By minimizing the distance between the similarity matrix of the cross-attention maps and the text self-attention matrix, our approach ensures that embedded syntactic relationships are effectively transferred to cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.485, + 0.409 + ], + "angle": 0, + "content": "Our method optimizes cross-attention maps during inference, adjusting their similarity matrix to align with the text self-attention matrix \\( \\mathsf{T} \\). The normalized cosine similarity matrix, S (defined in eq.(7)), is used as the cross-attention similarity matrix. This is achieved by simply minimizing the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.411, + 0.483, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(z _ {t}\\right) = \\sum_ {i = 1, j \\leq i} ^ {s} \\rho_ {i} \\left| \\mathrm {T} _ {i j} ^ {\\gamma} - \\mathrm {S} _ {i j} \\left(z _ {t}\\right) \\right|, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.544 + ], + "angle": 0, + "content": "where the exponent \\(\\gamma\\) acts to amplify larger values and compress smaller ones so the effect of controlling temperature and \\(\\rho_{i} = i / s\\). For example, if two words in the prompt have negligible syntactic relation according to the text self-attention matrix, i.e. \\(\\mathsf{T}_{ij}\\approx 0\\), we demand that their similarity of cross-attention maps must not be similar: \\(\\mathsf{S}_{ij}\\approx 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.484, + 0.573 + ], + "angle": 0, + "content": "In practice, this optimization will be applied only to \\( z_{t} \\) at a few denoising steps during inference as followed:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.576, + 0.483, + 0.593 + ], + "angle": 0, + "content": "\\[\nz _ {t} ^ {\\prime} = z _ {t} - \\alpha \\cdot \\nabla_ {z _ {t}} \\mathcal {L} (z _ {t}). \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.224, + 0.621 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.734 + ], + "angle": 0, + "content": "This section presents our empirical study on the effectiveness of T-SAM in enhancing semantic alignment across diverse T2I diffusion models, including SD v1.5 [29] and PixArt-\\(\\alpha\\) [4], evaluated both quantitatively and qualitatively. To the best of our knowledge, this is the first work to introduce attention control in PixArt-\\(\\alpha\\), underscoring the novelty of our approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Prompt datasets. We evaluate our approach on diverse text prompts using two existing benchmarks. First, the TIFA v1.0 benchmark [13] is a large-scale text-to-image generation dataset featuring a wide range of sentence structures. This benchmark comprises 4,000 prompts, including 2,000 image captions from the COCO validation set [18], 161 prompts from DrawBench [30], 1,420 prompts from PartiPrompt used in Parti [39], and 500 texts from PaintSkill [6]. Second, we use structured prompt sets containing multiple objects and their corresponding attributes from Attend-n-Excite [3]. Prompts in this dataset are" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.083, + 0.907, + 0.138 + ], + "angle": 0, + "content": "Table 1. Evaluation results for the TIFA benchmark, including TIFA scores and CLIP similarity scores. External Info. indicates whether external information is used. CLIP-I (CLIP-T) refers to image-text (text-text) CLIP similarity. The base model is SD." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.149, + 0.899, + 0.228 + ], + "angle": 0, + "content": "
External Info.TIFACLIP-ICLIP-T
SD×0.790.330.77
LB✓0.800.330.76
T-SAM×0.830.340.77
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.905, + 0.338 + ], + "angle": 0, + "content": "grouped into three categories: Objects (e.g., “[attribute1] [object1] and [attribute2] [object2]”), Animals-Objects (e.g., “[animal] with [object]” or “[animal] and/with [attribute] [object]”), and Animals (e.g., “[animal1] and [animal2]”). We exclude the Animals category, as it lacks the complex syntactic relations as discussed in [28]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.906, + 0.49 + ], + "angle": 0, + "content": "Implementation details. For SD, we use 50 sampling iterations, updating \\( z_{t} \\) at each denoising step from 1 to 25. We set \\( M = 256 \\) in eq.(6) and \\( \\gamma = 2 \\) in eq.(14). For the TIFA benchmark, we generate one image per prompt with a shared seed across methods, while using a unique seed for each prompt. We set \\( \\alpha = 40 \\) in eq.(15). For Attend-n-Excite prompts, we generate results using 64 seeds, in line with standard practice, setting \\( \\alpha = 10 \\) and applying 20 iterative updates at \\( i \\in \\{0,10,20\\} \\). See Appendix B.2 for details on PixArt- \\( \\alpha \\) as well as SD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Evaluation metrics. To quantitatively evaluate the accuracy of generated images, we use two metrics: TIFA scores [13] (for TIFA benchmark) and CLIP similarity scores. TIFA scores measure how well generated images reflect the text prompts. In TIFA, questions for each prompt are generated by GPT-3.5 [22], and a vision-language model [16] provides answers. For CLIP [25] scores, we follow the protocol in [3] to evaluate both image-text and text-text similarity. Additional details on CLIP scores are provided in Appendix B.2. Moreover, we conducted a user study with 10 random prompts in Animals-Objects, generating four images per prompt using different seeds. A total of 25 participants select the image that best matches the text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Baselines. When evaluating our method on SD, we compare our method with SD, Linguistic-Binding (LB) [28], Attend-n-Excite (A&E) [3] and CONFORM [21]. LB depending on external parsers (SpaCy [11]) is limited to attribute-binding tasks. This method can be applied to both the TIFA benchmark and the Attend-n-Excite (Objects and Animals-Objects) prompt sets, though they are used only when the text parser identifies (modifier, entity-noun) pairs within the given prompts. A&E and CONFORM require manual token selection per prompt, restricting their use to the fixed-template prompts (Objects and Animals-Objects) due to the high cost of selecting token indices for diverse prompts. We reproduce the images with SD, CONFORM, and LB based on SD v1.5 and using the same seeds. For PixArt-\\(\\alpha\\), we com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8036" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.177, + 0.085, + 0.825, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.182, + 0.908, + 0.211 + ], + "angle": 0, + "content": "Figure 7. Accuracy for each question type in the FIFA benchmark. Our method achieves the best performance in most categories. The attribute includes properties such as large, small, young, etc. The base model is SD." + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.223, + 0.816, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.687, + 0.908, + 0.715 + ], + "angle": 0, + "content": "Figure 8. Comparison of our method (T-SAM) on SD with recent state-of-the-art methods on prompts from Objects and Animals-Objects. The images corresponding to the same position across different methods are generated using the same seed. Best in zoom." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.486, + 0.765 + ], + "angle": 0, + "content": "pare only with the base model, as no prior attention-control methods have been applied to it." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.185, + 0.787 + ], + "angle": 0, + "content": "5.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Quantitative results. Table 1 shows the evaluation results in TIFA benchmark and Figure 7 illustrates the breakdown of TIFA scores across question types, using SD as the base model. We highlight that our approach (T-SAM) outperforms SD and LB on complex syntactic prompts, where other baselines are inapplicable. While LB demonstrates improvements over SD in the color, shape, and material" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.909, + 0.901 + ], + "angle": 0, + "content": "categories, which are closely related to attribute-binding tasks, it fails to enhance accuracy in the activity and counting categories. This underscores LB's inability to capture diverse word relationships, even when using external text parsers. In contrast, our method shows improvements over SD in nearly all categories including color, shape, and activity, except for attribute (e.g., properties such as large, small, young), which differ from the term \"attribute\" used in this study, demonstrating its versatility across diverse word relationships. Additionally, it achieves higher CLIP scores than the baselines, proving its superior semantic alignment." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8037" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.075, + 0.888, + 0.09 + ], + "angle": 0, + "content": "Table 2. Average CLIP similarity scores between text prompts and images generated with 64 different seeds. The base model is SD." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.1, + 0.907, + 0.251 + ], + "angle": 0, + "content": "
External Info.Image-Text PromptPrompt-Caption
Full PromptMinimum Object
ObjectsAnimals-ObjectsObjectsAnimals-ObjectsObjectsAnimals-Objects
SDX0.340.340.250.260.760.80
LB✓0.360.350.270.270.800.83
CONFORM✓0.360.360.280.280.810.85
A&E✓0.360.350.270.260.810.83
T-SAMX0.360.370.280.280.810.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.312 + ], + "angle": 0, + "content": "Table 3. CLIP similarity scores on prompts from Objects and Animals-Objects using PixArt-\\(\\alpha\\) as the base model. CLIP-I (CLIP-T) refers to image-text (text-text) CLIP similarity." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.317, + 0.498, + 0.414 + ], + "angle": 0, + "content": "
Prompt SetMethodCLIP-TCLIP-I (Full, Min)
ObjectsPixArt-α0.810.36, 0.27
T-SAM0.820.37, 0.28
Animals-ObjectsPixArt-α0.850.36, 0.27
T-SAM0.860.37, 0.28
" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.424, + 0.462, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparison using prompts from MSCOCO contained in TIFA benchmark. Our approach successfully generates multiple elements, including a sewing machine in the first prompt, people in the second, popsicle in the third, and cow in the fourth. SD misses some elements and LB generates identical images to SD when no relations are extracted from the text parser." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In Objects and Animals-Objects, T-SAM on SD performs comparably to the state-of-the-art CONFORM, which, unlike our approach, requires manually defined token indices for positive and negative groups, as shown in Table 2. Our method outperforms LB and A&E, which also rely on external inputs. This demonstrates extracting syntactic information from text self-attention maps can be more effective than relying on text parsers or manually selecting tokens for optimization. An additional ablation study is provided in Ap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.905, + 0.331 + ], + "angle": 0, + "content": "pendix B.2. Moreover, Table 3 shows that T-SAM is also effective on PixArt-\\(\\alpha\\) with *Attend-n-Excite* prompts, demonstrating the generalizability of our method to a different text encoder (T5-XXL) and model architecture (DiT)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.907, + 0.572 + ], + "angle": 0, + "content": "Qualitative results. In Figure 8, T-SAM either outperforms or performs comparably to the baselines in Objects and Animals-Objects. SD often omits objects (e.g., clock, apple, or bow) or misbinds attributes (e.g., blue clock). CONFORM and LB improve over SD but have limitations. LB is restricted to attribute-binding and fails on prompts without modifiers (e.g., An elephant with a bow), producing results identical to SD CONFORM can impose overly rigid separations, as seen in the first example. T-SAM, by leveraging smoother linguistic cues from text attention maps—rather than binary positive/negative pairs—achieves broader applicability and more consistent improvements. See Figure 9 for more comparison of images generated by T-SAM and the baselines, showcasing effectiveness of T-SAM with complex prompts, such as MSCOCO captions from the TIFA benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.909, + 0.68 + ], + "angle": 0, + "content": "For PixArt-\\(\\alpha\\), Figure D illustrates the effectiveness of our method. Although the base model is already strong compared to SD, T-SAM further improves its performance. User study. The user study result in Table D show that T-SAM on SD outperforms LB and achieves performance comparable to CONFORM. Additionally, Table E shows T-SAM on PixArt-\\(\\alpha\\) surpasses the base PixArt-\\(\\alpha\\) model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.634, + 0.708 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.9 + ], + "angle": 0, + "content": "To enhance fidelity in T2I diffusion models, we improved cross-attention maps by aligning their similarity matrix with text self-attention maps. Our approach is based on two insights: (1) similar text embeddings produce similar cross-attention maps, but (2) syntactic relations are missed in embeddings but captured by text self-attention maps. T-SAM enabled cross-attention to better capture syntactic structure, significantly improving text-to-image fidelity across diverse sentence structures, without requiring external tools. Nonetheless, our method has limitations such as inaccuracies in object counting and potential noise introduced by text attention maps, leaving room for future improvement." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8038" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.146 + ], + "angle": 0, + "content": "This work was partially supported by NIH (1U01CA269192)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.159, + 0.188, + 0.175 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.184, + 0.484, + 0.266 + ], + "angle": 0, + "content": "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2283-2293, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.268, + 0.484, + 0.322 + ], + "angle": 0, + "content": "[2] Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Alignit: Enhancing prompt alignment in customization of text-to-image models. arXiv preprint arXiv:2406.18893, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.324, + 0.484, + 0.393 + ], + "angle": 0, + "content": "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.395, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.484, + 0.534 + ], + "angle": 0, + "content": "[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart\\(\\cdot\\)sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.484, + 0.604 + ], + "angle": 0, + "content": "[6] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[8] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[9] Xiangming Gu, Tianyu Pang, Chao Du, Qian Liu, Fengzhuo Zhang, Cunxiao Du, Ye Wang, and Min Lin. When attention sink emerges in language models: An empirical view. arXiv preprint arXiv:2410.10781, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.184, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "neural networks and incremental parsing. To appear, 7(1): 411-420, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[12] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[13] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[14] Jiaxiu Jiang, Yabo Zhang, Kailai Feng, Xiaohe Wu, and Wangmeng Zuo. Mc2: Multi-concept guidance for customized multi-concept generation. arXiv preprint arXiv:2404.05268, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[15] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7701-7711, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[16] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[19] Oscar Manas, Pietro Astolfi, Melissa Hall, Candace Ross, Jack Urbanek, Adina Williams, Aishwarya Agrawal, Adriana Romero-Soriano, and Michal Drozdal. Improving text-to-image consistency via automatic prompt optimization. arXiv preprint arXiv:2403.17804, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[20] Arash Marioriyad, Mohammadali Banayeeanzade, Reza Abbasi, Mohammad Hossein Rohban, and Mahdieh Soleymani Baghshah. Attention overlap is responsible for the entity missing problem in text-to-image diffusion models! arXiv preprint arXiv:2410.20972, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[21] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9005–9014, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[22] OpenAI. Openai gpt-3 api [gpt-3.5-turbo], 2024. Available at: https://platform.openai.com. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8039" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[24] Zipeng Qi, Lichen Bai, Haoyi Xiong, et al. Not all noises are created equally: Diffusion noise selection and optimization. arXiv preprint arXiv:2407.14041, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[26] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[28] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[30] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[31] Dazhong Shen, Guanglu Song, Zeyue Xue, Fu-Yun Wang, and Yu Liu. Rethinking the spatial inconsistency in classifier-free diffusion guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9370-9379, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[32] Mingjie Sun, Xinlei Chen, J Zico Kolter, and Zhuang Liu. Massive activations in large language models. arXiv preprint arXiv:2402.17762, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[33] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[34] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. arXiv preprint arXiv:2312.03626, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[35] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[36] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291-26303, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[37] Mingyang Yi, Aoxue Li, Yi Xin, and Zhenguo Li. Towards understanding the working mechanism of text-to-image diffusion model. arXiv preprint arXiv:2405.15330, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[38] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv preprint arXiv:2404.01154, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[41] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Tiviatis Sim, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. arXiv preprint arXiv:2403.06381, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8040" + } + ] +] \ No newline at end of file diff --git a/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_origin.pdf b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a2159ebbc7117bb2c8a7dcc1be6e924b9b54766c --- /dev/null +++ b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/986776a8-b76f-4d8b-9abf-09cb8a9ba5b4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0635980a853eb33b9128b1e04a2ce5cd606cae7b3bad57f9c2662be15c37b555 +size 9457468 diff --git a/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/full.md b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/full.md new file mode 100644 index 0000000000000000000000000000000000000000..36b32a12a736a79eb3aa74f9f9a4485fa3226989 --- /dev/null +++ b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/full.md @@ -0,0 +1,374 @@ +# Text Embedding is Not All You Need: Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps + +Jeeyung Kim*, Erfan Esmaeili*, and Qiang Qiu Purdue University + +{jkim17,efakhabi,qqiu}@purdue.edu + +# Abstract + +In text-to-image diffusion models, the cross-attention map of each text token indicates the specific image regions attended. Comparing these maps of syntactically related tokens provides insights into how well the generated image reflects the text prompt. For example, in the prompt, "a black car and a white clock", the cross-attention maps for "black" and "car" should focus on overlapping regions to depict a black car, while "car" and "clock" should not. Incorrect overlapping in the maps generally produces generation flaws such as missing objects and incorrect attribute binding. Our study makes the key observations investigating this issue in the existing text-to-image models: (1) the similarity in text embeddings between different tokens—used as conditioning inputs—can cause their cross-attention maps to focus on the same image regions; and (2) text embeddings often fail to faithfully capture syntactic relations already within text attention maps. As a result, such syntactic relationships can be overlooked in cross-attention module, leading to inaccurate image generation. To address this, we propose a method that directly transfers syntactic relations from the text attention maps to the cross-attention module via a test-time optimization. Our approach leverages this inherent yet unexploited information within text attention maps to enhance image-text semantic alignment across diverse prompts, without relying on external guidance. Our project page and code are available at: https://t-sam-diffusion.github.io/ + +# 1. Introduction + +Recent diffusion models can generate images from diverse text prompts [4, 7, 29]. However, semantic discrepancies often arise between the text and generated images, raising problems such as missing objects—where certain elements are overlooked—and attribute mis-binding—where attributes are incorrectly assigned to subjects. + +![](images/4c8d42dc8dcb0cb5b6afa2998885985a1237795b348c0e907a535f58f8ef83df.jpg) +Figure 1. The overview of our method. We leverage text self-attention matrix and optimize the latent noise $(z_{t})$ by minimizing the distance between the cross-attention similarity matrix (S) and the text self-attention matrix (T). This encourages integrating syntactic relationships into text-to-image diffusion models. + +Prior studies [10] demonstrated that the cross-attention map of each token in text-to-image (T2I) diffusion models highlights the attended regions in the image and provides clues about the spatial placement of elements corresponding to the tokens. In particular, [1, 21, 28] suggest that spatial alignment in cross-attention maps among related words influences the fidelity of images to text prompts in Stable Diffusion (SD) [29], as we also demonstrate in Section 4.1. For instance, in the prompt a black car and a white clock, if the cross-attention maps for car and clock overlap, unique token contributions can dilute, potentially omitting one object. Conversely, if the maps for black (or white) and car (or clock) diverge too much, attribute mis-binding can occur. This implies syntactically related words should ideally have spatially aligned cross-attention maps, as discussed in [8, 28]. However, the factors determining this spatial alignment between the maps remain poorly understood. + +In our study, we first investigate the factors that con + +![](images/f8159dd76edaf140b5a4126fec09a3c0db380076065fe9ee75c8a31dde699c14.jpg) +(a) + +![](images/dbb3a59d5118e001b32252c91d42e33aa823baac6740916acf64f9273f35400f.jpg) +(b) +Figure 2. For the analysis, we use SD v1.5 as the base model and adopt the prompt sets from [3], structured as “[attribute_1] [object_1] and [attribute_2] [object_2]”, “[object_1] and/with [object_2]” or “[object_1] and [object_2] [object_2]”. (a) Comparison of the cosine similarity of text embeddings with that of the corresponding cross-attention maps at denoising step 1, with pairs of tokens (object_i, object_j), where $i \neq j$ , and pairs of tokens (attribute_m, object_n) for both $m = n$ and $m \neq n$ . As text embeddings become more similar, their cross-attention maps get similar. (b) The distributions of text embedding similarity between i) Bound tokens—(attribute_i, object_i) for $i = 1, 2$ , and ii) Unbound tokens—(object_i, object_2). The distributions show no discernible difference, indicating text embeddings do not effectively represent the syntactic relationships. (c) Comparison of text embedding similarity (left) and the text self-attention map power by 3 (right) for the prompt a black car and a white clock. In the self-attention maps (T), clock attends more to white, unlike the text embeddings. + +![](images/0762e095c3c40ec0c583d2ebb747ac1465b0843470fc06f1c8c4c6205449b18e.jpg) +(c) + +![](images/1c0244cacefe59feb7411ac6555786de6dd624f0b4483d59e144ffc6046b06c3.jpg) + +tribute to the spatial alignment of cross-attention maps across different tokens, which can ultimately affect the accuracy of T2I generation. We reveal that text embeddings, which function as keys in the cross-attention modules, play a pivotal role in determining the similarity of these maps, as illustrated in Figure 2a. Specifically, as the text embeddings for different words become more similar, the cosine similarity of their corresponding cross-attention maps increases. This effect is noticeable from the initial denoising step. + +Considering that (i) cross-attention maps should capture syntactic relationships between words and that (ii) the spatial alignment of cross-attention maps is influenced by the text embeddings, we question whether text embeddings accurately capture the linguistic structure in text prompts. Our findings suggest they do not, as illustrated in Figure 2b and the left of Figure 2c. Syntactically related words in prompts (e.g., black-car and white-clock) do not necessarily yield similar text embeddings, as detailed in Section 4.1. This finding is further supported by [40], which argues that CLIP [25] often behaves like a bag-of-words model. As a result, the cross-attention maps, derived from text embeddings, can likely fail to faithfully reflect the syntactic relationships. In the end, the T2I diffusion models can often generate images that inaccurately represent the prompts. + +To address the above issue, prior studies [1, 8, 21, 28] resort to external sources to obtain syntactic information within text prompts and regularize cross-attention maps to incorporate these relationships. Specifically, [8, 28] employ external text parsers to obtain linguistic structure, while [1, 21] rely on human intervention to manually group tokens based on syntactic relationships. However, these methods are limited by their dependence on external inputs. + +Do we really need to rely on external sources to obtain syntactic relations within a sentence? Remarkably, T2I + +diffusion models inherently capture syntactic relationships within prompts through the self-attention maps in their text encoder. In the encoder's self-attention module, each token pays more attention to related words, leading to encoding the entire sentence effectively. As shown in the right of Figure 2c, the text self-attention maps exhibit higher attention scores between the syntactically related words. However, this information is weakly encoded in the text encoder outputs (i.e., text embeddings). We conjecture this is due to the text attention module's strong focus on the $<\text{bos}>$ token, known as attention sink [32, 35], which minimizes the influence of other tokens, as further discussed in Section 4.1. + +As illustrated in Figure 1, our approach reuses the self-attention maps from the text encoder and directly transfers the syntax information to the diffusion models. We first define a similarity matrix whose values indicate the similarity between pairs of cross-attention maps. Then, we update the latent noise in diffusion models to minimize the distance between the cross-attention similarity matrix and the text self-attention matrix during inference. This allows for the seamless integration of syntactic relationships into the diffusion process. By guiding cross-attention modules to better capture contextual relationships, our method ultimately produces images accurately reflecting the intended meaning of prompts. We leverage overlooked information already embedded in T2I models, offering two key advantages: (i) Self-contained: no need for external inputs like text parsers or manual token indices; and (ii) Generalizable: effective across diverse sentence structures. + +# 2. Related Work + +Text-to-image generation. Text-to-image (T2I) generation [4, 5, 7, 23, 27, 30] is commonly based on latent diffusion models [29] where text data is processed via a text encoder + +[25, 26]. An important problem in T2I models is the lack of full correspondence between the input text and the image. Although diverse methods [2, 8, 12, 19, 24, 31, 38] are proposed to diagnose and fix this problem, a common theme in prior works is adjusting the cross-attention maps in inference time, following the seminal work [10]. + +Cross-attention control for improved semantic alignment. Several defects in cross-attention maps hinder semantic alignment, including attention dominance [41]—one token getting huge attention weight, and attention leakage/overlap [20, 36]—attention weights not respecting the spatial boundaries of the intended objects. In particular, one line of work employs contrastive objectives to guide attention optimization [1, 14, 21, 28]. For example, CONFORM [21] uses prompts with distinct groups, where tokens in the same (opposite) group are treated as positive (negative) pairs. Similarly, [28] uses text parser models to identify such pairs. Other approaches, like those in [15, 34], optimize attention maps with spatial guidance (e.g., segmentation maps, masks, or layouts). However, these methods rely on external resources for text-to-image generation. + +Attention sinks. The attention weights in pretrained transformers tend to be heavily biased towards special tokens such as $\langle \text{bos} \rangle$ , $\langle \text{eos} \rangle$ , and punctuation tokens, as studied in (vision-)language models [9, 32, 35]. The T2I diffusion model literature [3, 37] also notes the focus of attention scores on the $\langle \text{bos} \rangle$ token in the CLIP text encoder. + +# 3. Preliminaries + +In this section, we briefly review the structure of text-to-image diffusion models including the text encoder and the cross-attention module [4, 29]. + +Text encoder. To condition the text data for the diffusion process, a text encoder is needed. The input text is first tokenized and then converted into dense vectors. The dense vectors are then processed via a series of multi-head self-attention layers [33]. + +At a given layer $\ell$ in the text encoder, let us denote the key by $\mathbf{e}_i^{(\ell)}\in \mathbb{R}^{H_{\mathrm{e}}D_{\mathrm{e}}}$ $i = 1,\dots ,s$ where $s$ is the key sequence length, $D_{\mathrm{e}}$ is the embedding dimension per head and $H_{\mathrm{e}}$ is the number of heads in the text encoder. The corresponding self-attention matrix $T^{(\ell ,h)}\in \mathbb{R}^{s\times s}$ is given by + +$$ +T _ {i j} ^ {(\ell , h)} = \frac {\exp (\omega_ {i j})}{\sum_ {k} \exp (\omega_ {i k})}, \omega_ {i j} := \mathbf {e} _ {i} ^ {(\ell) \top} W _ {\mathrm {e n}} ^ {(\ell , h)} \mathbf {e} _ {j} ^ {(\ell)}, \tag {1} +$$ + +where $W_{\mathrm{en}}^{(\ell,h)} \in \mathbb{R}^{H_{\mathrm{c}}D_{\mathrm{c}} \times H_{\mathrm{c}}D_{\mathrm{c}}}$ is a pretrained matrix at head $h$ in layer $\ell$ of the text encoder. + +In our proposed method, we use the self-attention matrix averaged over layers and heads: + +$$ +T ^ {\prime} = \frac {1}{L _ {\mathrm {e}} H _ {\mathrm {e}}} \sum_ {\ell = 1} ^ {L _ {\mathrm {e}}} \sum_ {h = 1} ^ {H _ {\mathrm {e}}} T ^ {(\ell , h)}. \tag {2} +$$ + +Due to the high attention probabilities assigned to $\langle \text{bos} \rangle$ and $\langle \text{eos} \rangle$ tokens, we remove their corresponding values and re-normalize each row as follows: + +$$ +T _ {i j} = \frac {T ^ {\prime} {} _ {i j}}{\sum_ {m = 2} ^ {i} T ^ {\prime} {} _ {i m}}. \tag {3} +$$ + +We denote the final output of the text encoder by $\mathbf{k}_i\in$ $\mathbb{R}^{H_{\mathrm{e}}D_{\mathrm{e}}}$ where $i = 1,\dots ,s$ , which used as conditioning inputs in T2I diffusion models. We call $\mathbf{k}_i$ as text embeddings. + +Denoising latent variables. Latent diffusion models generate latent tensors $(z_0)$ of an image. A latent diffusion model $D_{\theta}$ learns to simulate the denoising process: starting from Gaussian noise $z_{\tau}$ , the model iteratively reconstruct $z_0$ by predicting the noise at step $t$ : + +$$ +z _ {t - 1} = z _ {t} - D _ {\theta} \left(z _ {t}; \left\{\mathbf {k} _ {i} \right\}\right), \quad 0 < t \leq \tau . \tag {4} +$$ + +Cross-attention module. In T2I latent diffusion models, the interaction between text and image data is performed by multi-head cross-attention. In the cross-attention layer $\ell$ , we define the query, $\mathbf{q}_a^{(\ell)} \in \mathbb{R}^{H_cD_c}$ , where $a = 1, \dots, N_c$ , and $N_c$ is the query sequence length at the given cross-attention layer. $D_c$ is the hidden dimension per head, and $H_c$ is the number of heads of the cross-attention layer. + +The cross-attention map $A^{(\ell,h)} \in \mathbb{R}^{N_{\mathrm{c}} \times s}$ at head $h$ with elements $A_{ai}^{(\ell,h)}$ is defined as + +$$ +A _ {a i} ^ {(\ell , h)} := \frac {\exp (\Omega_ {a i})}{\sum_ {j = 1} ^ {s} \exp (\Omega_ {a j})}, \Omega_ {a i} := \mathbf {q} _ {a} ^ {(\ell) \top} W _ {\mathrm {c}} ^ {(\ell , h)} \mathbf {k} _ {i} ^ {(\ell)}. \tag {5} +$$ + +Here, $W_{\mathfrak{c}}^{(\ell,h)} \in \mathbb{R}^{H_{\mathfrak{c}}D_{\mathfrak{c}} \times H_{\mathfrak{c}}D_{\mathfrak{c}}}$ is a pretrained matrix at head $h$ in layer $\ell$ . + +Cross-attention similarity matrix. In our study, we use the cross-attention maps averaged over heads and over all layers for which $N_{\mathrm{c}} = M$ : + +$$ +\mathsf {A} = \frac {1}{L _ {M} H _ {\mathrm {c}}} \sum_ {\ell = 1} ^ {L _ {M}} \sum_ {h = 1} ^ {H _ {\mathrm {c}}} A ^ {(\ell , h)}. \tag {6} +$$ + +Here, $L_{M}$ is the number of cross-attention layers with $N_{\mathrm{c}} = M$ . Based on the average cross-attention map above, we define the similarity matrix $\mathsf{S} \in \mathbb{R}^{s \times s}$ as follows: + +$$ +\mathrm {S} _ {i j} := \frac {\mathrm {C} _ {i j}}{\sum_ {k = 1} ^ {s} \mathrm {C} _ {i k}}, \mathrm {C} _ {i j} := \frac {\sum_ {a = 1} ^ {N _ {\mathrm {c}}} \mathrm {A} _ {a i} \mathrm {A} _ {a j}}{\left(\sum_ {a = 1} ^ {N _ {\mathrm {c}}} \mathrm {A} _ {a i} ^ {2}\right) ^ {\frac {1}{2}} \left(\sum_ {a = 1} ^ {N _ {\mathrm {c}}} \mathrm {A} _ {a j} ^ {2}\right) ^ {\frac {1}{2}}}. \tag {7} +$$ + +$C_{ij}$ denotes cosine similarity between maps corresponding to $i$ and $j$ . + +# 4. Understanding and Resolving Text-to-Image Semantic Discrepancy + +In this section, we examine the cross-attention maps in Stable Diffusion (SD) [29] to identify factors that lead to semantic discrepancies between generated images and text + +prompts. We provide additional results with $\mathrm{PixArt - \alpha}$ [4] in Appendix B.1. Based on these findings, we introduce an approach to enhance the fidelity of T2I diffusion models by leveraging text self-attention maps to regularize cross-attention maps. + +# 4.1. Why Do Generated Images Misrepresent Text? + +Our findings are based on the premise that cross-attention maps should capture syntactic relationships between words for accurate T2I generation, as supported by prior works [1, 8, 21, 28]. We identify a key factor contributing to incorrect relations in cross-attention maps, resulting in less faithful image generation: While text embedding $(\mathbf{k}_i)$ similarity strongly correlates with the cross-attention similarity matrix C, it insufficiently reflects syntactic bindings in the text prompt. This highlights a fundamental issue in text-image semantic discrepancy: the similarity in text embeddings lack the syntax information necessary for accurate image generation. We reveal that text self-attention maps (T) effectively capture syntactic information, but this is not sufficiently encoded in text embeddings. The absence of the information in text embeddings can be due to an artifact in the attention module, known as attention sink [35], where attention weights are biased toward the $<\mathrm{bos}>$ token as detailed later. These findings motivate the solution presented in the following section. + +For the following empirical analysis, we use the prompt sets introduced in [3], containing prompts structured as the following three categories: + +(i) $\left[\mathsf{attribute}_1\right]\left[\mathsf{object}_1\right]$ and $\left[\mathsf{attribute}_2\right]\left[\mathsf{object}_2\right]$ , +(ii) $\left[\mathrm{object}_1(\mathrm{animal})\right]$ with $\left[\mathrm{object}_2\right]$ , +(iii) [object1(animal)] and [attribute2][object2]. + +Prompt sets in the above formats enable us to focus on two key cases of text-image semantic discrepancies: missing objects and attribute mis-binding. In the following, we refer to the group of syntactically bound words (tokens) as pairs of $([\text{attribute}_i], [\text{object}_i])$ , $i = 1, 2$ , and the group of syntactically unbound words (tokens) as $([\text{attribute}_i], [\text{object}_j])$ and $([\text{object}_1], [\text{object}_2])$ , where $i \neq j$ . + +We revisit the role of cross-attention maps in text-image semantic coherence, where spatial overlap or separation is crucial, as discussed in [1, 8, 21, 28]. Figure 3 illustrates this effect: On the left, overlapping attention maps for syntactically unbound words lead to object missing, while greater overlap for syntactically bound words enhances attribute binding. This suggests that syntactic associations should be reflected in cross-attention maps. + +To statistically assess cross-attention map's impact, we analyze categories (ii) and (iii) of the aforementioned prompt sets. Assuming spatial overlap is measurable by the cosine similarity matrix (C) in eq.(7), we compute cosine similarities for the attention maps of $object_{1}$ and $object_{2}$ in missing objects, and for $attribute_{2}$ and $attribute_{2}$ in attribute + +![](images/820c2323337cdbfceed0662d9a2144b441cb6746dcb4ba64796b474288e0ac15.jpg) +1) Missing object A bear and a red balloon +bear +balloon + +![](images/34d5b58472365410e00197e2d03301c3a6f8a1414d5baa0a98a57a0fe7701e1a.jpg) +A frog with a bow +frog + +![](images/40241db9d828d7c43b2e1b3a76a8e0c6b235860d87d5aa17b1947646e4e91b5d.jpg) + +![](images/9b6a22eae0f1cc2b90d57453b1284d11a69a2de65bc1ab1f603ca4eeec124484.jpg) +2) Attribute binding +A green backpack and a purple bench +# +green +# +1 + +![](images/daf810ec6d34c952b31b3b00904d600777e8f41d8869f3c426aa06d60714dc20.jpg) +backpack + +![](images/d3df530aaadbb446df85c15ec63db8beb9567bfe2d3ed48538e341a88fcfb1b2.jpg) +bench + +![](images/5dff54dd72e49694b1b0055deecb3bf54110888f4e958201a87bcfa0462151ea.jpg) +(a) + +![](images/33f5d35e73453fa3ade2fb250691afd3479196a335bd709a5676af044b95fcaa.jpg) +Figure 3. The generated images and cross-attention maps (A) for the specific tokens from SD v1.5. This illustrate the importance of spatial alignment in cross-attention maps for accurate image generation. Divergent (overlapping) cross-attention maps for syntactically unbound (bound) words enhances text-to-image fidelity. +(b) + +![](images/e1d37eb917f850d001a7888302681e5abef7b2cb4b882e3639c50c3368e55f16.jpg) +Figure 4. Comparison for the distributions of cosine similarity between cross-attention maps (at denoising step 10). (a) The cases with one missing object-incorrect and two objects present-correct. (b) The cases with incorrect and correct attribute binding. Correct instances are more frequent when the cosine similarity is low for objects presence and high for attribute binding. +Figure 5. Correlation between the cosine similarity of text embeddings and that of cross-attention maps across denoising steps $(t = 1,21,50)$ . Similar text embeddings generally lead to similar cross-attention maps, with the correlation weakening over time. + +![](images/d773f8e3db246ae126b3da330a48b286b3bf42732172c9fca0fcb29b6f2a1c94.jpg) + +![](images/4eab806f49ad4a252db821362de0127cafbe2223cbbdd203db6d5b9744cce3e8.jpg) + +binding cases. Then, we compare cosine similarity distributions for correct vs. incorrect images on both cases. See Appendix B.1 for setup details. Figure 4 shows that lower cosine similarity tends to correlate with object presence, while higher similarity supports more accurate bindings. + +This observation motivates us to examine the factor contributing to similarity in cross-attention maps across tokens. Finding 1: The cosine similarity of text embeddings has a large correlation with cross-attention similarity matrix C. Figure 5 shows there is a correlation between the cosine similarity of text embedding and C, which persists throughout the final denoising steps. This indicates that similar text + +embeddings can result in overlapping cross-attention maps. + +Next, we justify this finding mathematically. + +Proposition 1. If $A^{(\ell, h)} \in \mathbb{R}^{N_c \times s}$ is a cross-attention map defined in eq. (5), then under the assumptions i, ii, and iii described in Appendix A, the cosine similarity matrix can be written in terms of key vectors $\mathbf{k}_i^{(\ell, h)} \in \mathbb{R}^{H_c D_c}$ as + +$$ +\begin{array}{l} \cos \left(A _ {i} ^ {(\ell , h)}, A _ {j} ^ {(\ell , h)}\right) = \\ \exp \left(- \frac {1}{2} \left(\mathbf {k} _ {i} - \mathbf {k} _ {j}\right) ^ {\top} W ^ {2} \left(\mathbf {k} _ {i} - \mathbf {k} _ {j}\right)\right), \tag {8} \\ \end{array} +$$ + +up to terms of at least $\mathcal{O}(1 / \sqrt{N_c})$ and $\mathcal{O}(\epsilon)$ , where $W^2 \coloneqq W_c^{(\ell,h)^\top}\Sigma^{(\ell)}W_c^{(\ell,h)}$ and $\Sigma^{(\ell)} \in \mathbb{R}^{H_cD_c \times H_cD_c}$ is the covariance matrix of query vectors. + +Refer to Appendix A for the proof. + +We empirically and mathematically showed that the similarity in text embeddings $\mathbf{k}_i$ influences cross-attention maps. Next, we evaluate whether the similarity of these embeddings reflects syntax information in text inputs. + +Finding 2-1: There is no significant correlation between word syntactic bindings and text embedding similarity. Prior study [40] suggests CLIP embeddings [25], used in SD, behave like a bag-of-words model, ignoring word relationships. We also show in Figure 2b the similarity in CLIP text embeddings does not correlate with syntactic bindings. Specifically, we expect close embeddings for syntactically bound tokens and distant embeddings for unbound tokens, yet the distributions lack separation. + +Finding 2-2: Text self-attention maps do have syntax information. We examine the text encoder that produces text embeddings. Interestingly, the self-attention maps in the encoder capture syntactic relationships, as shown in Figure 6a. These maps reveal higher similarity between syntactically bound tokens and lower similarity between unbound tokens. For more complex prompt structures, see the Appendix B.1. + +Finding 2-3: The attention sink can contribute to why text embeddings lack the syntax information. Why do text embeddings lack relational information despite being derived from multiple self-attention modules? We attribute this gap to attention sink [32, 35], where attention scores are concentrated on a few tokens. In CLIP's text encoder, attention is mainly focused on the $\langle \text{bos} \rangle$ token, as discussed in [3, 37] and shown in Figure 6b. We hypothesize the focus on the $\langle \text{bos} \rangle$ token can limit the transfer of relational information from self-attention maps $(T)$ to text embeddings, as attention scores for other tokens remain much smaller, minimizing their influence in each self-attention layer. + +To provide a mathematical justification, we express the difference between the output and input of one block of the + +![](images/6020acbb489405fffbd5851eb34f530e8e880266f49002eec1ea244495d18e03.jpg) +(a) + +![](images/824ea9085e315c096db8dc7d07f0037be86c830e4504eba7eaded7809045e6c6.jpg) +(b) +Figure 6. (a) The distributions of text self-attention value $(\mathsf{T}_{ij}^{\prime})$ in eq.(2)) for bound tokens (attribute $m$ , object $m$ ) and unbound tokens (attribute $m$ , object $n$ ) / (object $1$ , object $2$ ), where $m, n \in \{1, 2\}$ . The separate distributions indicate text self-attention maps can indeed represent the syntactic relationships. (b) Comparison of text self-attention probability histograms between token and non tokens on 100 prompts: The probability allocated to is on average 20 times larger than that of other tokens. + +self-attention module for a token $e_i$ as: + +$$ +\mathbf {o} _ {i} ^ {(\ell , h)} = \sum_ {j = 1} ^ {i} T _ {i j} ^ {(\ell , h)} W _ {\mathrm {v}} ^ {(\ell , h)} \mathbf {e} _ {j} ^ {(\ell)}, \tag {9} +$$ + +where $W_{\mathrm{v}}^{(\ell,h)} \in \mathbb{R}^{D_{\mathrm{e}} \times H_{\mathrm{e}} D_{\mathrm{e}}}$ is a parameter, $\mathbf{e}_i \in \mathbb{R}^{H_{\mathrm{e}} D_{\mathrm{e}}}$ , and $T_{ij}^{(\ell,h)}$ is the self-attention matrix in eq. (1). We consider the situation where attention sink occurs, that is the attention weights for the token is much higher than the rest of the sequence: + +$$ +\epsilon = \frac {\sum_ {j \neq 1} ^ {i} T _ {i j}}{T _ {i 1}} \ll 1, \quad i = 2, \dots , s. \tag {10} +$$ + +In Appendix A, we prove the following statement: + +Proposition 2. Define matrix $R \in \mathbb{R}^{s \times s}$ as + +$$ +R _ {i j} = \mathbf {e} _ {i} ^ {(\ell) \top} W _ {v} ^ {(\ell , h) \top} W _ {v} ^ {(\ell , h)} \mathbf {e} _ {j} ^ {(\ell)}, \tag {11} +$$ + +and suppose it has the property + +$$ +\frac {\left| R _ {m n} \right|}{R _ {1 1}} \sim \mathcal {O} (1 / \epsilon), \quad \frac {\left| R _ {1 m} \right|}{R _ {1 1}} \sim \mathcal {O} (1), \quad 1 < m, n \leq s, \tag {12} +$$ + +where $\mathbf{e}_1^{(\ell)}$ is the bos embedding. Then the following holds: + +$$ +\cos \left(\mathbf {o} _ {i} ^ {(\ell , h)}, \mathbf {o} _ {j} ^ {(\ell , h)}\right) = 1 - \mathcal {O} (\epsilon). \tag {13} +$$ + +![](images/6aaf0e5f90f7cdd380846164eee8f66c00f8e88fab2309db70d9d6d5924f58e0.jpg) + +This suggests that cosine similarity between token vectors $(\mathbf{o}_i^{(\ell ,h)})$ -potentially influencing cross-attention maps remains barely changed across text self-attention layers due to attention sink on the. $\langle \mathrm{bos}\rangle$ token. In other words, the attention sink can hinder the accurate encoding of self-attention maps into embeddings. + +Our findings reveal text embedding alone are insufficient + +for generating semantically aligned images. On the other hand, we notably show the potential of transferring neglected syntactic information from text self-attention maps to the cross-attention to enhance T2I semantic alignment. + +# 4.2. Text Self-Attention Maps (T-SAM) Guidance + +In the previous section, we show text self-attention maps capture syntax information within a sentence. Building on this insight, we propose leveraging the self-attention maps within the text encoder—a component of diffusion models—to enhance cross-attention maps. By minimizing the distance between the similarity matrix of the cross-attention maps and the text self-attention matrix, our approach ensures that embedded syntactic relationships are effectively transferred to cross-attention. + +Our method optimizes cross-attention maps during inference, adjusting their similarity matrix to align with the text self-attention matrix $\mathsf{T}$ . The normalized cosine similarity matrix, S (defined in eq.(7)), is used as the cross-attention similarity matrix. This is achieved by simply minimizing the loss function: + +$$ +\mathcal {L} \left(z _ {t}\right) = \sum_ {i = 1, j \leq i} ^ {s} \rho_ {i} \left| \mathrm {T} _ {i j} ^ {\gamma} - \mathrm {S} _ {i j} \left(z _ {t}\right) \right|, \tag {14} +$$ + +where the exponent $\gamma$ acts to amplify larger values and compress smaller ones so the effect of controlling temperature and $\rho_{i} = i / s$ . For example, if two words in the prompt have negligible syntactic relation according to the text self-attention matrix, i.e. $\mathsf{T}_{ij}\approx 0$ , we demand that their similarity of cross-attention maps must not be similar: $\mathsf{S}_{ij}\approx 0$ . + +In practice, this optimization will be applied only to $z_{t}$ at a few denoising steps during inference as followed: + +$$ +z _ {t} ^ {\prime} = z _ {t} - \alpha \cdot \nabla_ {z _ {t}} \mathcal {L} (z _ {t}). \tag {15} +$$ + +# 5. Experiments + +This section presents our empirical study on the effectiveness of T-SAM in enhancing semantic alignment across diverse T2I diffusion models, including SD v1.5 [29] and PixArt- $\alpha$ [4], evaluated both quantitatively and qualitatively. To the best of our knowledge, this is the first work to introduce attention control in PixArt- $\alpha$ , underscoring the novelty of our approach. + +Prompt datasets. We evaluate our approach on diverse text prompts using two existing benchmarks. First, the TIFA v1.0 benchmark [13] is a large-scale text-to-image generation dataset featuring a wide range of sentence structures. This benchmark comprises 4,000 prompts, including 2,000 image captions from the COCO validation set [18], 161 prompts from DrawBench [30], 1,420 prompts from PartiPrompt used in Parti [39], and 500 texts from PaintSkill [6]. Second, we use structured prompt sets containing multiple objects and their corresponding attributes from Attend-n-Excite [3]. Prompts in this dataset are + +Table 1. Evaluation results for the TIFA benchmark, including TIFA scores and CLIP similarity scores. External Info. indicates whether external information is used. CLIP-I (CLIP-T) refers to image-text (text-text) CLIP similarity. The base model is SD. + +
External Info.TIFACLIP-ICLIP-T
SD×0.790.330.77
LB✓0.800.330.76
T-SAM×0.830.340.77
+ +grouped into three categories: Objects (e.g., “[attribute1] [object1] and [attribute2] [object2]”), Animals-Objects (e.g., “[animal] with [object]” or “[animal] and/with [attribute] [object]”), and Animals (e.g., “[animal1] and [animal2]”). We exclude the Animals category, as it lacks the complex syntactic relations as discussed in [28]. + +Implementation details. For SD, we use 50 sampling iterations, updating $z_{t}$ at each denoising step from 1 to 25. We set $M = 256$ in eq.(6) and $\gamma = 2$ in eq.(14). For the TIFA benchmark, we generate one image per prompt with a shared seed across methods, while using a unique seed for each prompt. We set $\alpha = 40$ in eq.(15). For Attend-n-Excite prompts, we generate results using 64 seeds, in line with standard practice, setting $\alpha = 10$ and applying 20 iterative updates at $i \in \{0,10,20\}$ . See Appendix B.2 for details on PixArt- $\alpha$ as well as SD. + +Evaluation metrics. To quantitatively evaluate the accuracy of generated images, we use two metrics: TIFA scores [13] (for TIFA benchmark) and CLIP similarity scores. TIFA scores measure how well generated images reflect the text prompts. In TIFA, questions for each prompt are generated by GPT-3.5 [22], and a vision-language model [16] provides answers. For CLIP [25] scores, we follow the protocol in [3] to evaluate both image-text and text-text similarity. Additional details on CLIP scores are provided in Appendix B.2. Moreover, we conducted a user study with 10 random prompts in Animals-Objects, generating four images per prompt using different seeds. A total of 25 participants select the image that best matches the text. + +Baselines. When evaluating our method on SD, we compare our method with SD, Linguistic-Binding (LB) [28], Attend-n-Excite (A&E) [3] and CONFORM [21]. LB depending on external parsers (SpaCy [11]) is limited to attribute-binding tasks. This method can be applied to both the TIFA benchmark and the Attend-n-Excite (Objects and Animals-Objects) prompt sets, though they are used only when the text parser identifies (modifier, entity-noun) pairs within the given prompts. A&E and CONFORM require manual token selection per prompt, restricting their use to the fixed-template prompts (Objects and Animals-Objects) due to the high cost of selecting token indices for diverse prompts. We reproduce the images with SD, CONFORM, and LB based on SD v1.5 and using the same seeds. For PixArt- $\alpha$ , we com + +![](images/063924eeb3667c09d7b26adb25ff4af23e395b36fa8b01f168b768aec1cdfd45.jpg) +Figure 7. Accuracy for each question type in the FIFA benchmark. Our method achieves the best performance in most categories. The attribute includes properties such as large, small, young, etc. The base model is SD. + +![](images/04a9f402dc2566b67bf450c19545222ba121bc36dd3e86c89de884ad49afdd02.jpg) +Figure 8. Comparison of our method (T-SAM) on SD with recent state-of-the-art methods on prompts from Objects and Animals-Objects. The images corresponding to the same position across different methods are generated using the same seed. Best in zoom. + +pare only with the base model, as no prior attention-control methods have been applied to it. + +# 5.1. Results + +Quantitative results. Table 1 shows the evaluation results in TIFA benchmark and Figure 7 illustrates the breakdown of TIFA scores across question types, using SD as the base model. We highlight that our approach (T-SAM) outperforms SD and LB on complex syntactic prompts, where other baselines are inapplicable. While LB demonstrates improvements over SD in the color, shape, and material + +categories, which are closely related to attribute-binding tasks, it fails to enhance accuracy in the activity and counting categories. This underscores LB's inability to capture diverse word relationships, even when using external text parsers. In contrast, our method shows improvements over SD in nearly all categories including color, shape, and activity, except for attribute (e.g., properties such as large, small, young), which differ from the term "attribute" used in this study, demonstrating its versatility across diverse word relationships. Additionally, it achieves higher CLIP scores than the baselines, proving its superior semantic alignment. + +Table 2. Average CLIP similarity scores between text prompts and images generated with 64 different seeds. The base model is SD. + +
External Info.Image-Text PromptPrompt-Caption
Full PromptMinimum Object
ObjectsAnimals-ObjectsObjectsAnimals-ObjectsObjectsAnimals-Objects
SDX0.340.340.250.260.760.80
LB✓0.360.350.270.270.800.83
CONFORM✓0.360.360.280.280.810.85
A&E✓0.360.350.270.260.810.83
T-SAMX0.360.370.280.280.810.85
+ +Table 3. CLIP similarity scores on prompts from Objects and Animals-Objects using PixArt- $\alpha$ as the base model. CLIP-I (CLIP-T) refers to image-text (text-text) CLIP similarity. + +
Prompt SetMethodCLIP-TCLIP-I (Full, Min)
ObjectsPixArt-α0.810.36, 0.27
T-SAM0.820.37, 0.28
Animals-ObjectsPixArt-α0.850.36, 0.27
T-SAM0.860.37, 0.28
+ +![](images/50b673a935f082a57901dc6907a19c2fb74bfe339c454d5255c3c5512c94fe24.jpg) +Figure 9. Qualitative comparison using prompts from MSCOCO contained in TIFA benchmark. Our approach successfully generates multiple elements, including a sewing machine in the first prompt, people in the second, popsicle in the third, and cow in the fourth. SD misses some elements and LB generates identical images to SD when no relations are extracted from the text parser. + +In Objects and Animals-Objects, T-SAM on SD performs comparably to the state-of-the-art CONFORM, which, unlike our approach, requires manually defined token indices for positive and negative groups, as shown in Table 2. Our method outperforms LB and A&E, which also rely on external inputs. This demonstrates extracting syntactic information from text self-attention maps can be more effective than relying on text parsers or manually selecting tokens for optimization. An additional ablation study is provided in Ap + +pendix B.2. Moreover, Table 3 shows that T-SAM is also effective on PixArt- $\alpha$ with *Attend-n-Excite* prompts, demonstrating the generalizability of our method to a different text encoder (T5-XXL) and model architecture (DiT). + +Qualitative results. In Figure 8, T-SAM either outperforms or performs comparably to the baselines in Objects and Animals-Objects. SD often omits objects (e.g., clock, apple, or bow) or misbinds attributes (e.g., blue clock). CONFORM and LB improve over SD but have limitations. LB is restricted to attribute-binding and fails on prompts without modifiers (e.g., An elephant with a bow), producing results identical to SD CONFORM can impose overly rigid separations, as seen in the first example. T-SAM, by leveraging smoother linguistic cues from text attention maps—rather than binary positive/negative pairs—achieves broader applicability and more consistent improvements. See Figure 9 for more comparison of images generated by T-SAM and the baselines, showcasing effectiveness of T-SAM with complex prompts, such as MSCOCO captions from the TIFA benchmark. + +For PixArt- $\alpha$ , Figure D illustrates the effectiveness of our method. Although the base model is already strong compared to SD, T-SAM further improves its performance. User study. The user study result in Table D show that T-SAM on SD outperforms LB and achieves performance comparable to CONFORM. Additionally, Table E shows T-SAM on PixArt- $\alpha$ surpasses the base PixArt- $\alpha$ model. + +# 6. Conclusion + +To enhance fidelity in T2I diffusion models, we improved cross-attention maps by aligning their similarity matrix with text self-attention maps. Our approach is based on two insights: (1) similar text embeddings produce similar cross-attention maps, but (2) syntactic relations are missed in embeddings but captured by text self-attention maps. T-SAM enabled cross-attention to better capture syntactic structure, significantly improving text-to-image fidelity across diverse sentence structures, without requiring external tools. Nonetheless, our method has limitations such as inaccuracies in object counting and potential noise introduced by text attention maps, leaving room for future improvement. + +# Acknowledgement + +This work was partially supported by NIH (1U01CA269192). + +# References + +[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2283-2293, 2023. 1, 2, 3, 4 +[2] Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Alignit: Enhancing prompt alignment in customization of text-to-image models. arXiv preprint arXiv:2406.18893, 2024. 3 +[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2, 3, 4, 5, 6 +[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 2, 3, 4, 6 +[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart $\cdot$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2 +[6] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 6 +[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 1, 2 +[8] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 3, 4 +[9] Xiangming Gu, Tianyu Pang, Chao Du, Qian Liu, Fengzhuo Zhang, Cunxiao Du, Ye Wang, and Min Lin. When attention sink emerges in language models: An empirical view. arXiv preprint arXiv:2410.10781, 2024. 3 +[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 3 +[11] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional + +neural networks and incremental parsing. To appear, 7(1): 411-420, 2017. 6 +[12] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 3 +[13] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 6, 5 +[14] Jiaxiu Jiang, Yabo Zhang, Kailai Feng, Xiaohe Wu, and Wangmeng Zuo. Mc2: Multi-concept guidance for customized multi-concept generation. arXiv preprint arXiv:2404.05268, 2024.3 +[15] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7701-7711, 2023. 3 +[16] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 6, 5 +[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5 +[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[19] Oscar Manas, Pietro Astolfi, Melissa Hall, Candace Ross, Jack Urbanek, Adina Williams, Aishwarya Agrawal, Adriana Romero-Soriano, and Michal Drozdal. Improving text-to-image consistency via automatic prompt optimization. arXiv preprint arXiv:2403.17804, 2024. 3 +[20] Arash Marioriyad, Mohammadali Banayeeanzade, Reza Abbasi, Mohammad Hossein Rohban, and Mahdieh Soleymani Baghshah. Attention overlap is responsible for the entity missing problem in text-to-image diffusion models! arXiv preprint arXiv:2410.20972, 2024. 3 +[21] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9005–9014, 2024. 1, 2, 3, 4, 6 +[22] OpenAI. Openai gpt-3 api [gpt-3.5-turbo], 2024. Available at: https://platform.openai.com. 6, 5 +[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 + +[24] Zipeng Qi, Lichen Bai, Haoyi Xiong, et al. Not all noises are created equally: Diffusion noise selection and optimization. arXiv preprint arXiv:2407.14041, 2024. 3 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6 +[26] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 3 +[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[28] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 6, 5 +[30] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 6 +[31] Dazhong Shen, Guanglu Song, Zeyue Xue, Fu-Yun Wang, and Yu Liu. Rethinking the spatial inconsistency in classifier-free diffusion guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9370-9379, 2024. 3 +[32] Mingjie Sun, Xinlei Chen, J Zico Kolter, and Zhuang Liu. Massive activations in large language models. arXiv preprint arXiv:2402.17762, 2024. 2, 3, 5 +[33] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[34] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. arXiv preprint arXiv:2312.03626, 2023. 3 +[35] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 2, 3, 4, 5 +[36] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291-26303, 2023. 3 + +[37] Mingyang Yi, Aoxue Li, Yi Xin, and Zhenguo Li. Towards understanding the working mechanism of text-to-image diffusion model. arXiv preprint arXiv:2405.15330, 2024. 3, 5 +[38] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv preprint arXiv:2404.01154, 2024. 3 +[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 6 +[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 2, 5 +[41] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Tiviatis Sim, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. arXiv preprint arXiv:2403.06381, 2024. 3 \ No newline at end of file diff --git a/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/images.zip b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6de48e2a6e0a6364e1d63b270950aae773cb129d --- /dev/null +++ b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8eee71b79e1272663e46ef0355aceca6802ce6e69f0a1e4cb6ae789467e37238 +size 724552 diff --git a/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/layout.json b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c5d5efb464cc5ce58c3cabad46d83726263dc398 --- /dev/null +++ b/2025/Text Embedding is Not All You Need_ Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps/layout.json @@ -0,0 +1,10047 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 536, + 140 + ], + "type": "text", + "content": "Text Embedding is Not All You Need: Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 190, + 161, + 419, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 161, + 419, + 190 + ], + "spans": [ + { + "bbox": [ + 190, + 161, + 419, + 190 + ], + "type": "text", + "content": "Jeeyung Kim*, Erfan Esmaeili*, and Qiang Qiu Purdue University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 191, + 399, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 191, + 399, + 204 + ], + "spans": [ + { + "bbox": [ + 209, + 191, + 399, + 204 + ], + "type": "text", + "content": "{jkim17,efakhabi,qqiu}@purdue.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "content": "In text-to-image diffusion models, the cross-attention map of each text token indicates the specific image regions attended. Comparing these maps of syntactically related tokens provides insights into how well the generated image reflects the text prompt. For example, in the prompt, \"a black car and a white clock\", the cross-attention maps for \"black\" and \"car\" should focus on overlapping regions to depict a black car, while \"car\" and \"clock\" should not. Incorrect overlapping in the maps generally produces generation flaws such as missing objects and incorrect attribute binding. Our study makes the key observations investigating this issue in the existing text-to-image models: (1) the similarity in text embeddings between different tokens—used as conditioning inputs—can cause their cross-attention maps to focus on the same image regions; and (2) text embeddings often fail to faithfully capture syntactic relations already within text attention maps. As a result, such syntactic relationships can be overlooked in cross-attention module, leading to inaccurate image generation. To address this, we propose a method that directly transfers syntactic relations from the text attention maps to the cross-attention module via a test-time optimization. Our approach leverages this inherent yet unexploited information within text attention maps to enhance image-text semantic alignment across diverse prompts, without relying on external guidance. Our project page and code are available at: https://t-sam-diffusion.github.io/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 602, + 136, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 136, + 615 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 136, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 622, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 696 + ], + "type": "text", + "content": "Recent diffusion models can generate images from diverse text prompts [4, 7, 29]. However, semantic discrepancies often arise between the text and generated images, raising problems such as missing objects—where certain elements are overlooked—and attribute mis-binding—where attributes are incorrectly assigned to subjects." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 555, + 422 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 422 + ], + "type": "image", + "image_path": "4c8d42dc8dcb0cb5b6afa2998885985a1237795b348c0e907a535f58f8ef83df.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 429, + 556, + 485 + ], + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 485 + ], + "type": "text", + "content": "Figure 1. The overview of our method. We leverage text self-attention matrix and optimize the latent noise " + }, + { + "bbox": [ + 313, + 429, + 556, + 485 + ], + "type": "inline_equation", + "content": "(z_{t})" + }, + { + "bbox": [ + 313, + 429, + 556, + 485 + ], + "type": "text", + "content": " by minimizing the distance between the cross-attention similarity matrix (S) and the text self-attention matrix (T). This encourages integrating syntactic relationships into text-to-image diffusion models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 497, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 701 + ], + "type": "text", + "content": "Prior studies [10] demonstrated that the cross-attention map of each token in text-to-image (T2I) diffusion models highlights the attended regions in the image and provides clues about the spatial placement of elements corresponding to the tokens. In particular, [1, 21, 28] suggest that spatial alignment in cross-attention maps among related words influences the fidelity of images to text prompts in Stable Diffusion (SD) [29], as we also demonstrate in Section 4.1. For instance, in the prompt a black car and a white clock, if the cross-attention maps for car and clock overlap, unique token contributions can dilute, potentially omitting one object. Conversely, if the maps for black (or white) and car (or clock) diverge too much, attribute mis-binding can occur. This implies syntactically related words should ideally have spatially aligned cross-attention maps, as discussed in [8, 28]. However, the factors determining this spatial alignment between the maps remain poorly understood." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 702, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 555, + 714 + ], + "type": "text", + "content": "In our study, we first investigate the factors that con" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 135, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 135, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 135, + 713 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8031" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 59, + 192, + 171 + ], + "blocks": [ + { + "bbox": [ + 56, + 59, + 192, + 171 + ], + "lines": [ + { + "bbox": [ + 56, + 59, + 192, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 59, + 192, + 171 + ], + "type": "image", + "image_path": "f8159dd76edaf140b5a4126fec09a3c0db380076065fe9ee75c8a31dde699c14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 174, + 130, + 183 + ], + "lines": [ + { + "bbox": [ + 119, + 174, + 130, + 183 + ], + "spans": [ + { + "bbox": [ + 119, + 174, + 130, + 183 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 194, + 59, + 323, + 171 + ], + "blocks": [ + { + "bbox": [ + 194, + 59, + 323, + 171 + ], + "lines": [ + { + "bbox": [ + 194, + 59, + 323, + 171 + ], + "spans": [ + { + "bbox": [ + 194, + 59, + 323, + 171 + ], + "type": "image", + "image_path": "dbb3a59d5118e001b32252c91d42e33aa823baac6740916acf64f9273f35400f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 174, + 263, + 183 + ], + "lines": [ + { + "bbox": [ + 252, + 174, + 263, + 183 + ], + "spans": [ + { + "bbox": [ + 252, + 174, + 263, + 183 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. For the analysis, we use SD v1.5 as the base model and adopt the prompt sets from [3], structured as “[attribute_1] [object_1] and [attribute_2] [object_2]”, “[object_1] and/with [object_2]” or “[object_1] and [object_2] [object_2]”. (a) Comparison of the cosine similarity of text embeddings with that of the corresponding cross-attention maps at denoising step 1, with pairs of tokens (object_i, object_j), where " + }, + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "type": "text", + "content": ", and pairs of tokens (attribute_m, object_n) for both " + }, + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "type": "inline_equation", + "content": "m = n" + }, + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "type": "inline_equation", + "content": "m \\neq n" + }, + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "type": "text", + "content": ". As text embeddings become more similar, their cross-attention maps get similar. (b) The distributions of text embedding similarity between i) Bound tokens—(attribute_i, object_i) for " + }, + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "type": "inline_equation", + "content": "i = 1, 2" + }, + { + "bbox": [ + 54, + 193, + 555, + 281 + ], + "type": "text", + "content": ", and ii) Unbound tokens—(object_i, object_2). The distributions show no discernible difference, indicating text embeddings do not effectively represent the syntactic relationships. (c) Comparison of text embedding similarity (left) and the text self-attention map power by 3 (right) for the prompt a black car and a white clock. In the self-attention maps (T), clock attends more to white, unlike the text embeddings." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 327, + 61, + 428, + 171 + ], + "blocks": [ + { + "bbox": [ + 327, + 61, + 428, + 171 + ], + "lines": [ + { + "bbox": [ + 327, + 61, + 428, + 171 + ], + "spans": [ + { + "bbox": [ + 327, + 61, + 428, + 171 + ], + "type": "image", + "image_path": "0762e095c3c40ec0c583d2ebb747ac1465b0843470fc06f1c8c4c6205449b18e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 174, + 447, + 183 + ], + "lines": [ + { + "bbox": [ + 436, + 174, + 447, + 183 + ], + "spans": [ + { + "bbox": [ + 436, + 174, + 447, + 183 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 434, + 64, + 552, + 171 + ], + "blocks": [ + { + "bbox": [ + 434, + 64, + 552, + 171 + ], + "lines": [ + { + "bbox": [ + 434, + 64, + 552, + 171 + ], + "spans": [ + { + "bbox": [ + 434, + 64, + 552, + 171 + ], + "type": "image", + "image_path": "1c0244cacefe59feb7411ac6555786de6dd624f0b4483d59e144ffc6046b06c3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 297, + 294, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 294, + 405 + ], + "type": "text", + "content": "tribute to the spatial alignment of cross-attention maps across different tokens, which can ultimately affect the accuracy of T2I generation. We reveal that text embeddings, which function as keys in the cross-attention modules, play a pivotal role in determining the similarity of these maps, as illustrated in Figure 2a. Specifically, as the text embeddings for different words become more similar, the cosine similarity of their corresponding cross-attention maps increases. This effect is noticeable from the initial denoising step." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 407, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 588 + ], + "type": "text", + "content": "Considering that (i) cross-attention maps should capture syntactic relationships between words and that (ii) the spatial alignment of cross-attention maps is influenced by the text embeddings, we question whether text embeddings accurately capture the linguistic structure in text prompts. Our findings suggest they do not, as illustrated in Figure 2b and the left of Figure 2c. Syntactically related words in prompts (e.g., black-car and white-clock) do not necessarily yield similar text embeddings, as detailed in Section 4.1. This finding is further supported by [40], which argues that CLIP [25] often behaves like a bag-of-words model. As a result, the cross-attention maps, derived from text embeddings, can likely fail to faithfully reflect the syntactic relationships. In the end, the T2I diffusion models can often generate images that inaccurately represent the prompts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 590, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 687 + ], + "type": "text", + "content": "To address the above issue, prior studies [1, 8, 21, 28] resort to external sources to obtain syntactic information within text prompts and regularize cross-attention maps to incorporate these relationships. Specifically, [8, 28] employ external text parsers to obtain linguistic structure, while [1, 21] rely on human intervention to manually group tokens based on syntactic relationships. However, these methods are limited by their dependence on external inputs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Do we really need to rely on external sources to obtain syntactic relations within a sentence? Remarkably, T2I" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 297, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 440 + ], + "type": "text", + "content": "diffusion models inherently capture syntactic relationships within prompts through the self-attention maps in their text encoder. In the encoder's self-attention module, each token pays more attention to related words, leading to encoding the entire sentence effectively. As shown in the right of Figure 2c, the text self-attention maps exhibit higher attention scores between the syntactically related words. However, this information is weakly encoded in the text encoder outputs (i.e., text embeddings). We conjecture this is due to the text attention module's strong focus on the " + }, + { + "bbox": [ + 313, + 297, + 555, + 440 + ], + "type": "inline_equation", + "content": "<\\text{bos}>" + }, + { + "bbox": [ + 313, + 297, + 555, + 440 + ], + "type": "text", + "content": " token, known as attention sink [32, 35], which minimizes the influence of other tokens, as further discussed in Section 4.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 441, + 556, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 645 + ], + "type": "text", + "content": "As illustrated in Figure 1, our approach reuses the self-attention maps from the text encoder and directly transfers the syntax information to the diffusion models. We first define a similarity matrix whose values indicate the similarity between pairs of cross-attention maps. Then, we update the latent noise in diffusion models to minimize the distance between the cross-attention similarity matrix and the text self-attention matrix during inference. This allows for the seamless integration of syntactic relationships into the diffusion process. By guiding cross-attention modules to better capture contextual relationships, our method ultimately produces images accurately reflecting the intended meaning of prompts. We leverage overlooked information already embedded in T2I models, offering two key advantages: (i) Self-contained: no need for external inputs like text parsers or manual token indices; and (ii) Generalizable: effective across diverse sentence structures." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 656, + 400, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 656, + 400, + 669 + ], + "spans": [ + { + "bbox": [ + 314, + 656, + 400, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Text-to-image generation. Text-to-image (T2I) generation [4, 5, 7, 23, 27, 30] is commonly based on latent diffusion models [29] where text data is processed via a text encoder" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8032" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": "[25, 26]. An important problem in T2I models is the lack of full correspondence between the input text and the image. Although diverse methods [2, 8, 12, 19, 24, 31, 38] are proposed to diagnose and fix this problem, a common theme in prior works is adjusting the cross-attention maps in inference time, following the seminal work [10]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 323 + ], + "type": "text", + "content": "Cross-attention control for improved semantic alignment. Several defects in cross-attention maps hinder semantic alignment, including attention dominance [41]—one token getting huge attention weight, and attention leakage/overlap [20, 36]—attention weights not respecting the spatial boundaries of the intended objects. In particular, one line of work employs contrastive objectives to guide attention optimization [1, 14, 21, 28]. For example, CONFORM [21] uses prompts with distinct groups, where tokens in the same (opposite) group are treated as positive (negative) pairs. Similarly, [28] uses text parser models to identify such pairs. Other approaches, like those in [15, 34], optimize attention maps with spatial guidance (e.g., segmentation maps, masks, or layouts). However, these methods rely on external resources for text-to-image generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 395 + ], + "type": "text", + "content": "Attention sinks. The attention weights in pretrained transformers tend to be heavily biased towards special tokens such as " + }, + { + "bbox": [ + 55, + 323, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\langle \\text{bos} \\rangle" + }, + { + "bbox": [ + 55, + 323, + 296, + 395 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 323, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\langle \\text{eos} \\rangle" + }, + { + "bbox": [ + 55, + 323, + 296, + 395 + ], + "type": "text", + "content": ", and punctuation tokens, as studied in (vision-)language models [9, 32, 35]. The T2I diffusion model literature [3, 37] also notes the focus of attention scores on the " + }, + { + "bbox": [ + 55, + 323, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\langle \\text{bos} \\rangle" + }, + { + "bbox": [ + 55, + 323, + 296, + 395 + ], + "type": "text", + "content": " token in the CLIP text encoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 406, + 140, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 140, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 140, + 418 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "text", + "content": "In this section, we briefly review the structure of text-to-image diffusion models including the text encoder and the cross-attention module [4, 29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": "Text encoder. To condition the text data for the diffusion process, a text encoder is needed. The input text is first tokenized and then converted into dense vectors. The dense vectors are then processed via a series of multi-head self-attention layers [33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "text", + "content": "At a given layer " + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "text", + "content": " in the text encoder, let us denote the key by " + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i^{(\\ell)}\\in \\mathbb{R}^{H_{\\mathrm{e}}D_{\\mathrm{e}}}" + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,s" + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "text", + "content": " is the key sequence length, " + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{e}}" + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "text", + "content": " is the embedding dimension per head and " + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{e}}" + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "text", + "content": " is the number of heads in the text encoder. The corresponding self-attention matrix " + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "inline_equation", + "content": "T^{(\\ell ,h)}\\in \\mathbb{R}^{s\\times s}" + }, + { + "bbox": [ + 55, + 522, + 296, + 583 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 592, + 295, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 592, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 68, + 592, + 295, + 619 + ], + "type": "interline_equation", + "content": "T _ {i j} ^ {(\\ell , h)} = \\frac {\\exp (\\omega_ {i j})}{\\sum_ {k} \\exp (\\omega_ {i k})}, \\omega_ {i j} := \\mathbf {e} _ {i} ^ {(\\ell) \\top} W _ {\\mathrm {e n}} ^ {(\\ell , h)} \\mathbf {e} _ {j} ^ {(\\ell)}, \\tag {1}", + "image_path": "4f9f6d5f51bbab35438ecc0ceb227931b09d2a193835db40efa95f6618a0b66f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 628, + 295, + 653 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{en}}^{(\\ell,h)} \\in \\mathbb{R}^{H_{\\mathrm{c}}D_{\\mathrm{c}} \\times H_{\\mathrm{c}}D_{\\mathrm{c}}}" + }, + { + "bbox": [ + 55, + 628, + 295, + 653 + ], + "type": "text", + "content": " is a pretrained matrix at head " + }, + { + "bbox": [ + 55, + 628, + 295, + 653 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 628, + 295, + 653 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 55, + 628, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 628, + 295, + 653 + ], + "type": "text", + "content": " of the text encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "type": "text", + "content": "In our proposed method, we use the self-attention matrix averaged over layers and heads:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 119, + 683, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 683, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 119, + 683, + 295, + 716 + ], + "type": "interline_equation", + "content": "T ^ {\\prime} = \\frac {1}{L _ {\\mathrm {e}} H _ {\\mathrm {e}}} \\sum_ {\\ell = 1} ^ {L _ {\\mathrm {e}}} \\sum_ {h = 1} ^ {H _ {\\mathrm {e}}} T ^ {(\\ell , h)}. \\tag {2}", + "image_path": "d12a0ef4dec5820497ebcc05523057ae40a02901e602b85701cd127242481bb5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": "Due to the high attention probabilities assigned to " + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "inline_equation", + "content": "\\langle \\text{bos} \\rangle" + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "inline_equation", + "content": "\\langle \\text{eos} \\rangle" + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": " tokens, we remove their corresponding values and re-normalize each row as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 391, + 111, + 555, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 111, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 391, + 111, + 555, + 139 + ], + "type": "interline_equation", + "content": "T _ {i j} = \\frac {T ^ {\\prime} {} _ {i j}}{\\sum_ {m = 2} ^ {i} T ^ {\\prime} {} _ {i m}}. \\tag {3}", + "image_path": "bb5edc4efb59ff5cc45120cb5c2c7611ef739537c0ee7a36fe20494f62e45bb0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "type": "text", + "content": "We denote the final output of the text encoder by " + }, + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i\\in" + }, + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H_{\\mathrm{e}}D_{\\mathrm{e}}}" + }, + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,s" + }, + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "type": "text", + "content": " , which used as conditioning inputs in T2I diffusion models. We call " + }, + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i" + }, + { + "bbox": [ + 313, + 143, + 555, + 178 + ], + "type": "text", + "content": " as text embeddings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "content": "Denoising latent variables. Latent diffusion models generate latent tensors " + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "inline_equation", + "content": "(z_0)" + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "content": " of an image. A latent diffusion model " + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "content": " learns to simulate the denoising process: starting from Gaussian noise " + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "inline_equation", + "content": "z_{\\tau}" + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "content": ", the model iteratively reconstruct " + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "content": " by predicting the noise at step " + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 244, + 555, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 244, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 351, + 244, + 555, + 258 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = z _ {t} - D _ {\\theta} \\left(z _ {t}; \\left\\{\\mathbf {k} _ {i} \\right\\}\\right), \\quad 0 < t \\leq \\tau . \\tag {4}", + "image_path": "21e13c3aa3a280a5707c2da225f83c8f5e9f97152545b932ef5bbe3cb7987fc4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "text", + "content": "Cross-attention module. In T2I latent diffusion models, the interaction between text and image data is performed by multi-head cross-attention. In the cross-attention layer " + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "text", + "content": ", we define the query, " + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_a^{(\\ell)} \\in \\mathbb{R}^{H_cD_c}" + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "inline_equation", + "content": "a = 1, \\dots, N_c" + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "text", + "content": " is the query sequence length at the given cross-attention layer. " + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "inline_equation", + "content": "D_c" + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "text", + "content": " is the hidden dimension per head, and " + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "inline_equation", + "content": "H_c" + }, + { + "bbox": [ + 313, + 262, + 554, + 347 + ], + "type": "text", + "content": " is the number of heads of the cross-attention layer." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 347, + 554, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 554, + 374 + ], + "type": "text", + "content": "The cross-attention map " + }, + { + "bbox": [ + 313, + 347, + 554, + 374 + ], + "type": "inline_equation", + "content": "A^{(\\ell,h)} \\in \\mathbb{R}^{N_{\\mathrm{c}} \\times s}" + }, + { + "bbox": [ + 313, + 347, + 554, + 374 + ], + "type": "text", + "content": " at head " + }, + { + "bbox": [ + 313, + 347, + 554, + 374 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 347, + 554, + 374 + ], + "type": "text", + "content": " with elements " + }, + { + "bbox": [ + 313, + 347, + 554, + 374 + ], + "type": "inline_equation", + "content": "A_{ai}^{(\\ell,h)}" + }, + { + "bbox": [ + 313, + 347, + 554, + 374 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 376, + 553, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 376, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 324, + 376, + 553, + 415 + ], + "type": "interline_equation", + "content": "A _ {a i} ^ {(\\ell , h)} := \\frac {\\exp (\\Omega_ {a i})}{\\sum_ {j = 1} ^ {s} \\exp (\\Omega_ {a j})}, \\Omega_ {a i} := \\mathbf {q} _ {a} ^ {(\\ell) \\top} W _ {\\mathrm {c}} ^ {(\\ell , h)} \\mathbf {k} _ {i} ^ {(\\ell)}. \\tag {5}", + "image_path": "ade5e49ae754a67b53a27679d5d0d303eecebf06e985639991c17761b5ed4a19.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 418, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 445 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 418, + 554, + 445 + ], + "type": "inline_equation", + "content": "W_{\\mathfrak{c}}^{(\\ell,h)} \\in \\mathbb{R}^{H_{\\mathfrak{c}}D_{\\mathfrak{c}} \\times H_{\\mathfrak{c}}D_{\\mathfrak{c}}}" + }, + { + "bbox": [ + 313, + 418, + 554, + 445 + ], + "type": "text", + "content": " is a pretrained matrix at head " + }, + { + "bbox": [ + 313, + 418, + 554, + 445 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 418, + 554, + 445 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 313, + 418, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 418, + 554, + 445 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 445, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 480 + ], + "type": "text", + "content": "Cross-attention similarity matrix. In our study, we use the cross-attention maps averaged over heads and over all layers for which " + }, + { + "bbox": [ + 313, + 445, + 554, + 480 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{c}} = M" + }, + { + "bbox": [ + 313, + 445, + 554, + 480 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 376, + 485, + 553, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 485, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 376, + 485, + 553, + 518 + ], + "type": "interline_equation", + "content": "\\mathsf {A} = \\frac {1}{L _ {M} H _ {\\mathrm {c}}} \\sum_ {\\ell = 1} ^ {L _ {M}} \\sum_ {h = 1} ^ {H _ {\\mathrm {c}}} A ^ {(\\ell , h)}. \\tag {6}", + "image_path": "593e64d22eb3ccc1542767523c161a1bdc30205eaec23bd773d6adec1ef90384.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "L_{M}" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": " is the number of cross-attention layers with " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{c}} = M" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": ". Based on the average cross-attention map above, we define the similarity matrix " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\mathsf{S} \\in \\mathbb{R}^{s \\times s}" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 319, + 560, + 553, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 560, + 553, + 606 + ], + "spans": [ + { + "bbox": [ + 319, + 560, + 553, + 606 + ], + "type": "interline_equation", + "content": "\\mathrm {S} _ {i j} := \\frac {\\mathrm {C} _ {i j}}{\\sum_ {k = 1} ^ {s} \\mathrm {C} _ {i k}}, \\mathrm {C} _ {i j} := \\frac {\\sum_ {a = 1} ^ {N _ {\\mathrm {c}}} \\mathrm {A} _ {a i} \\mathrm {A} _ {a j}}{\\left(\\sum_ {a = 1} ^ {N _ {\\mathrm {c}}} \\mathrm {A} _ {a i} ^ {2}\\right) ^ {\\frac {1}{2}} \\left(\\sum_ {a = 1} ^ {N _ {\\mathrm {c}}} \\mathrm {A} _ {a j} ^ {2}\\right) ^ {\\frac {1}{2}}}. \\tag {7}", + "image_path": "86637098f079c2b72d1b66a63b73a6670071643dcd9d8082a320b82ff5a09711.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "inline_equation", + "content": "C_{ij}" + }, + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "text", + "content": " denotes cosine similarity between maps corresponding to " + }, + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 643, + 554, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 554, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 554, + 671 + ], + "type": "text", + "content": "4. Understanding and Resolving Text-to-Image Semantic Discrepancy" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "In this section, we examine the cross-attention maps in Stable Diffusion (SD) [29] to identify factors that lead to semantic discrepancies between generated images and text" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8033" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "prompts. We provide additional results with " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{PixArt - \\alpha}" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": " [4] in Appendix B.1. Based on these findings, we introduce an approach to enhance the fidelity of T2I diffusion models by leveraging text self-attention maps to regularize cross-attention maps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 295, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 154 + ], + "type": "text", + "content": "4.1. Why Do Generated Images Misrepresent Text?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 159, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 387 + ], + "type": "text", + "content": "Our findings are based on the premise that cross-attention maps should capture syntactic relationships between words for accurate T2I generation, as supported by prior works [1, 8, 21, 28]. We identify a key factor contributing to incorrect relations in cross-attention maps, resulting in less faithful image generation: While text embedding " + }, + { + "bbox": [ + 55, + 159, + 296, + 387 + ], + "type": "inline_equation", + "content": "(\\mathbf{k}_i)" + }, + { + "bbox": [ + 55, + 159, + 296, + 387 + ], + "type": "text", + "content": " similarity strongly correlates with the cross-attention similarity matrix C, it insufficiently reflects syntactic bindings in the text prompt. This highlights a fundamental issue in text-image semantic discrepancy: the similarity in text embeddings lack the syntax information necessary for accurate image generation. We reveal that text self-attention maps (T) effectively capture syntactic information, but this is not sufficiently encoded in text embeddings. The absence of the information in text embeddings can be due to an artifact in the attention module, known as attention sink [35], where attention weights are biased toward the " + }, + { + "bbox": [ + 55, + 159, + 296, + 387 + ], + "type": "inline_equation", + "content": "<\\mathrm{bos}>" + }, + { + "bbox": [ + 55, + 159, + 296, + 387 + ], + "type": "text", + "content": " token as detailed later. These findings motivate the solution presented in the following section." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 296, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 423 + ], + "type": "text", + "content": "For the following empirical analysis, we use the prompt sets introduced in [3], containing prompts structured as the following three categories:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 276, + 462 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 423, + 276, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 276, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 276, + 437 + ], + "type": "text", + "content": "(i) " + }, + { + "bbox": [ + 55, + 423, + 276, + 437 + ], + "type": "inline_equation", + "content": "\\left[\\mathsf{attribute}_1\\right]\\left[\\mathsf{object}_1\\right]" + }, + { + "bbox": [ + 55, + 423, + 276, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 423, + 276, + 437 + ], + "type": "inline_equation", + "content": "\\left[\\mathsf{attribute}_2\\right]\\left[\\mathsf{object}_2\\right]" + }, + { + "bbox": [ + 55, + 423, + 276, + 437 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 211, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 211, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 211, + 449 + ], + "type": "text", + "content": "(ii) " + }, + { + "bbox": [ + 55, + 437, + 211, + 449 + ], + "type": "inline_equation", + "content": "\\left[\\mathrm{object}_1(\\mathrm{animal})\\right]" + }, + { + "bbox": [ + 55, + 437, + 211, + 449 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 437, + 211, + 449 + ], + "type": "inline_equation", + "content": "\\left[\\mathrm{object}_2\\right]" + }, + { + "bbox": [ + 55, + 437, + 211, + 449 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 267, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 267, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 267, + 462 + ], + "type": "text", + "content": "(iii) [object1(animal)] and [attribute2][object2]." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "text", + "content": "Prompt sets in the above formats enable us to focus on two key cases of text-image semantic discrepancies: missing objects and attribute mis-binding. In the following, we refer to the group of syntactically bound words (tokens) as pairs of " + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "inline_equation", + "content": "([\\text{attribute}_i], [\\text{object}_i])" + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "inline_equation", + "content": "i = 1, 2" + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "text", + "content": ", and the group of syntactically unbound words (tokens) as " + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "inline_equation", + "content": "([\\text{attribute}_i], [\\text{object}_j])" + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "inline_equation", + "content": "([\\text{object}_1], [\\text{object}_2])" + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 55, + 462, + 296, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 640 + ], + "type": "text", + "content": "We revisit the role of cross-attention maps in text-image semantic coherence, where spatial overlap or separation is crucial, as discussed in [1, 8, 21, 28]. Figure 3 illustrates this effect: On the left, overlapping attention maps for syntactically unbound words lead to object missing, while greater overlap for syntactically bound words enhances attribute binding. This suggests that syntactic associations should be reflected in cross-attention maps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "To statistically assess cross-attention map's impact, we analyze categories (ii) and (iii) of the aforementioned prompt sets. Assuming spatial overlap is measurable by the cosine similarity matrix (C) in eq.(7), we compute cosine similarities for the attention maps of " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "object_{1}" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "object_{2}" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " in missing objects, and for " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "attribute_{2}" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "attribute_{2}" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " in attribute" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 95, + 373, + 175 + ], + "blocks": [ + { + "bbox": [ + 316, + 66, + 384, + 91 + ], + "lines": [ + { + "bbox": [ + 316, + 66, + 384, + 91 + ], + "spans": [ + { + "bbox": [ + 316, + 66, + 384, + 91 + ], + "type": "text", + "content": "1) Missing object A bear and a red balloon" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 95, + 373, + 175 + ], + "lines": [ + { + "bbox": [ + 317, + 95, + 373, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 95, + 373, + 175 + ], + "type": "image", + "image_path": "820c2323337cdbfceed0662d9a2144b441cb6746dcb4ba64796b474288e0ac15.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 176, + 339, + 183 + ], + "lines": [ + { + "bbox": [ + 323, + 176, + 339, + 183 + ], + "spans": [ + { + "bbox": [ + 323, + 176, + 339, + 183 + ], + "type": "text", + "content": "bear" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 346, + 176, + 370, + 183 + ], + "lines": [ + { + "bbox": [ + 346, + 176, + 370, + 183 + ], + "spans": [ + { + "bbox": [ + 346, + 176, + 370, + 183 + ], + "type": "text", + "content": "balloon" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 345, + 150, + 371, + 175 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 150, + 371, + 175 + ], + "spans": [ + { + "bbox": [ + 345, + 150, + 371, + 175 + ], + "type": "image", + "image_path": "e1788b3d571850962a24a6669926808619e16139cf6c73d8cb5acfa023153e85.jpg" + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 375, + 95, + 431, + 175 + ], + "blocks": [ + { + "bbox": [ + 374, + 79, + 427, + 94 + ], + "lines": [ + { + "bbox": [ + 374, + 79, + 427, + 94 + ], + "spans": [ + { + "bbox": [ + 374, + 79, + 427, + 94 + ], + "type": "text", + "content": "A frog with a bow" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 95, + 431, + 175 + ], + "lines": [ + { + "bbox": [ + 375, + 95, + 431, + 175 + ], + "spans": [ + { + "bbox": [ + 375, + 95, + 431, + 175 + ], + "type": "image", + "image_path": "34d5b58472365410e00197e2d03301c3a6f8a1414d5baa0a98a57a0fe7701e1a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 176, + 395, + 183 + ], + "lines": [ + { + "bbox": [ + 381, + 176, + 395, + 183 + ], + "spans": [ + { + "bbox": [ + 381, + 176, + 395, + 183 + ], + "type": "text", + "content": "frog" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 433, + 101, + 494, + 167 + ], + "blocks": [ + { + "bbox": [ + 433, + 101, + 494, + 167 + ], + "lines": [ + { + "bbox": [ + 433, + 101, + 494, + 167 + ], + "spans": [ + { + "bbox": [ + 433, + 101, + 494, + 167 + ], + "type": "image", + "image_path": "40241db9d828d7c43b2e1b3a76a8e0c6b235860d87d5aa17b1947646e4e91b5d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 495, + 94, + 525, + 127 + ], + "blocks": [ + { + "bbox": [ + 434, + 66, + 511, + 77 + ], + "lines": [ + { + "bbox": [ + 434, + 66, + 511, + 77 + ], + "spans": [ + { + "bbox": [ + 434, + 66, + 511, + 77 + ], + "type": "text", + "content": "2) Attribute binding" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 82, + 495, + 100 + ], + "lines": [ + { + "bbox": [ + 432, + 82, + 495, + 100 + ], + "spans": [ + { + "bbox": [ + 432, + 82, + 495, + 100 + ], + "type": "text", + "content": "A green backpack and a purple bench" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 504, + 85, + 514, + 93 + ], + "lines": [ + { + "bbox": [ + 504, + 85, + 514, + 93 + ], + "spans": [ + { + "bbox": [ + 504, + 85, + 514, + 93 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 495, + 94, + 525, + 127 + ], + "lines": [ + { + "bbox": [ + 495, + 94, + 525, + 127 + ], + "spans": [ + { + "bbox": [ + 495, + 94, + 525, + 127 + ], + "type": "image", + "image_path": "9b6a22eae0f1cc2b90d57453b1284d11a69a2de65bc1ab1f603ca4eeec124484.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 129, + 520, + 135 + ], + "lines": [ + { + "bbox": [ + 501, + 129, + 520, + 135 + ], + "spans": [ + { + "bbox": [ + 501, + 129, + 520, + 135 + ], + "type": "text", + "content": "green" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 527, + 85, + 544, + 93 + ], + "lines": [ + { + "bbox": [ + 527, + 85, + 544, + 93 + ], + "spans": [ + { + "bbox": [ + 527, + 85, + 544, + 93 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 527, + 94, + 544, + 100 + ], + "lines": [ + { + "bbox": [ + 527, + 94, + 544, + 100 + ], + "spans": [ + { + "bbox": [ + 527, + 94, + 544, + 100 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 495, + 140, + 525, + 171 + ], + "blocks": [ + { + "bbox": [ + 495, + 140, + 525, + 171 + ], + "lines": [ + { + "bbox": [ + 495, + 140, + 525, + 171 + ], + "spans": [ + { + "bbox": [ + 495, + 140, + 525, + 171 + ], + "type": "image", + "image_path": "daf810ec6d34c952b31b3b00904d600777e8f41d8869f3c426aa06d60714dc20.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 172, + 524, + 179 + ], + "lines": [ + { + "bbox": [ + 495, + 172, + 524, + 179 + ], + "spans": [ + { + "bbox": [ + 495, + 172, + 524, + 179 + ], + "type": "text", + "content": "backpack" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 527, + 140, + 557, + 171 + ], + "blocks": [ + { + "bbox": [ + 527, + 140, + 557, + 171 + ], + "lines": [ + { + "bbox": [ + 527, + 140, + 557, + 171 + ], + "spans": [ + { + "bbox": [ + 527, + 140, + 557, + 171 + ], + "type": "image", + "image_path": "d3df530aaadbb446df85c15ec63db8beb9567bfe2d3ed48538e341a88fcfb1b2.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 531, + 172, + 552, + 179 + ], + "lines": [ + { + "bbox": [ + 531, + 172, + 552, + 179 + ], + "spans": [ + { + "bbox": [ + 531, + 172, + 552, + 179 + ], + "type": "text", + "content": "bench" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 318, + 259, + 432, + 346 + ], + "blocks": [ + { + "bbox": [ + 318, + 259, + 432, + 346 + ], + "lines": [ + { + "bbox": [ + 318, + 259, + 432, + 346 + ], + "spans": [ + { + "bbox": [ + 318, + 259, + 432, + 346 + ], + "type": "image", + "image_path": "5dff54dd72e49694b1b0055deecb3bf54110888f4e958201a87bcfa0462151ea.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 350, + 380, + 357 + ], + "lines": [ + { + "bbox": [ + 370, + 350, + 380, + 357 + ], + "spans": [ + { + "bbox": [ + 370, + 350, + 380, + 357 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 435, + 258, + 548, + 346 + ], + "blocks": [ + { + "bbox": [ + 313, + 194, + 554, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 554, + 250 + ], + "type": "text", + "content": "Figure 3. The generated images and cross-attention maps (A) for the specific tokens from SD v1.5. This illustrate the importance of spatial alignment in cross-attention maps for accurate image generation. Divergent (overlapping) cross-attention maps for syntactically unbound (bound) words enhances text-to-image fidelity." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 258, + 548, + 346 + ], + "lines": [ + { + "bbox": [ + 435, + 258, + 548, + 346 + ], + "spans": [ + { + "bbox": [ + 435, + 258, + 548, + 346 + ], + "type": "image", + "image_path": "33f5d35e73453fa3ade2fb250691afd3479196a335bd709a5676af044b95fcaa.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 350, + 497, + 357 + ], + "lines": [ + { + "bbox": [ + 486, + 350, + 497, + 357 + ], + "spans": [ + { + "bbox": [ + 486, + 350, + 497, + 357 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 316, + 431, + 396, + 499 + ], + "blocks": [ + { + "bbox": [ + 313, + 357, + 554, + 423 + ], + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 423 + ], + "type": "text", + "content": "Figure 4. Comparison for the distributions of cosine similarity between cross-attention maps (at denoising step 10). (a) The cases with one missing object-incorrect and two objects present-correct. (b) The cases with incorrect and correct attribute binding. Correct instances are more frequent when the cosine similarity is low for objects presence and high for attribute binding." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 431, + 396, + 499 + ], + "lines": [ + { + "bbox": [ + 316, + 431, + 396, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 396, + 499 + ], + "type": "image", + "image_path": "e1d37eb917f850d001a7888302681e5abef7b2cb4b882e3639c50c3368e55f16.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 508, + 554, + 552 + ], + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 552 + ], + "type": "text", + "content": "Figure 5. Correlation between the cosine similarity of text embeddings and that of cross-attention maps across denoising steps " + }, + { + "bbox": [ + 313, + 508, + 554, + 552 + ], + "type": "inline_equation", + "content": "(t = 1,21,50)" + }, + { + "bbox": [ + 313, + 508, + 554, + 552 + ], + "type": "text", + "content": ". Similar text embeddings generally lead to similar cross-attention maps, with the correlation weakening over time." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 397, + 432, + 473, + 499 + ], + "blocks": [ + { + "bbox": [ + 397, + 432, + 473, + 499 + ], + "lines": [ + { + "bbox": [ + 397, + 432, + 473, + 499 + ], + "spans": [ + { + "bbox": [ + 397, + 432, + 473, + 499 + ], + "type": "image", + "image_path": "d773f8e3db246ae126b3da330a48b286b3bf42732172c9fca0fcb29b6f2a1c94.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 474, + 432, + 552, + 499 + ], + "blocks": [ + { + "bbox": [ + 474, + 432, + 552, + 499 + ], + "lines": [ + { + "bbox": [ + 474, + 432, + 552, + 499 + ], + "spans": [ + { + "bbox": [ + 474, + 432, + 552, + 499 + ], + "type": "image", + "image_path": "4eab806f49ad4a252db821362de0127cafbe2223cbbdd203db6d5b9744cce3e8.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": "binding cases. Then, we compare cosine similarity distributions for correct vs. incorrect images on both cases. See Appendix B.1 for setup details. Figure 4 shows that lower cosine similarity tends to correlate with object presence, while higher similarity supports more accurate bindings." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "This observation motivates us to examine the factor contributing to similarity in cross-attention maps across tokens. Finding 1: The cosine similarity of text embeddings has a large correlation with cross-attention similarity matrix C. Figure 5 shows there is a correlation between the cosine similarity of text embedding and C, which persists throughout the final denoising steps. This indicates that similar text" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8034" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "content": "embeddings can result in overlapping cross-attention maps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 87, + 248, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 87, + 248, + 99 + ], + "spans": [ + { + "bbox": [ + 67, + 87, + 248, + 99 + ], + "type": "text", + "content": "Next, we justify this finding mathematically." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 114, + 296, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 296, + 165 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 296, + 165 + ], + "type": "text", + "content": "Proposition 1. If " + }, + { + "bbox": [ + 55, + 114, + 296, + 165 + ], + "type": "inline_equation", + "content": "A^{(\\ell, h)} \\in \\mathbb{R}^{N_c \\times s}" + }, + { + "bbox": [ + 55, + 114, + 296, + 165 + ], + "type": "text", + "content": " is a cross-attention map defined in eq. (5), then under the assumptions i, ii, and iii described in Appendix A, the cosine similarity matrix can be written in terms of key vectors " + }, + { + "bbox": [ + 55, + 114, + 296, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i^{(\\ell, h)} \\in \\mathbb{R}^{H_c D_c}" + }, + { + "bbox": [ + 55, + 114, + 296, + 165 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 168, + 295, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 168, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 86, + 168, + 295, + 211 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\cos \\left(A _ {i} ^ {(\\ell , h)}, A _ {j} ^ {(\\ell , h)}\\right) = \\\\ \\exp \\left(- \\frac {1}{2} \\left(\\mathbf {k} _ {i} - \\mathbf {k} _ {j}\\right) ^ {\\top} W ^ {2} \\left(\\mathbf {k} _ {i} - \\mathbf {k} _ {j}\\right)\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "637fec5b292fa629860811fb9a8b22d5991cc5b320368ad70df2dcfe604bafde.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "text", + "content": "up to terms of at least " + }, + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1 / \\sqrt{N_c})" + }, + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\epsilon)" + }, + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "inline_equation", + "content": "W^2 \\coloneqq W_c^{(\\ell,h)^\\top}\\Sigma^{(\\ell)}W_c^{(\\ell,h)}" + }, + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\Sigma^{(\\ell)} \\in \\mathbb{R}^{H_cD_c \\times H_cD_c}" + }, + { + "bbox": [ + 55, + 214, + 296, + 253 + ], + "type": "text", + "content": " is the covariance matrix of query vectors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 269, + 198, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 198, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 198, + 281 + ], + "type": "text", + "content": "Refer to Appendix A for the proof." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 331 + ], + "type": "text", + "content": "We empirically and mathematically showed that the similarity in text embeddings " + }, + { + "bbox": [ + 55, + 284, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i" + }, + { + "bbox": [ + 55, + 284, + 296, + 331 + ], + "type": "text", + "content": " influences cross-attention maps. Next, we evaluate whether the similarity of these embeddings reflects syntax information in text inputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "content": "Finding 2-1: There is no significant correlation between word syntactic bindings and text embedding similarity. Prior study [40] suggests CLIP embeddings [25], used in SD, behave like a bag-of-words model, ignoring word relationships. We also show in Figure 2b the similarity in CLIP text embeddings does not correlate with syntactic bindings. Specifically, we expect close embeddings for syntactically bound tokens and distant embeddings for unbound tokens, yet the distributions lack separation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": "Finding 2-2: Text self-attention maps do have syntax information. We examine the text encoder that produces text embeddings. Interestingly, the self-attention maps in the encoder capture syntactic relationships, as shown in Figure 6a. These maps reveal higher similarity between syntactically bound tokens and lower similarity between unbound tokens. For more complex prompt structures, see the Appendix B.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 543, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 686 + ], + "type": "text", + "content": "Finding 2-3: The attention sink can contribute to why text embeddings lack the syntax information. Why do text embeddings lack relational information despite being derived from multiple self-attention modules? We attribute this gap to attention sink [32, 35], where attention scores are concentrated on a few tokens. In CLIP's text encoder, attention is mainly focused on the " + }, + { + "bbox": [ + 55, + 543, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\langle \\text{bos} \\rangle" + }, + { + "bbox": [ + 55, + 543, + 296, + 686 + ], + "type": "text", + "content": " token, as discussed in [3, 37] and shown in Figure 6b. We hypothesize the focus on the " + }, + { + "bbox": [ + 55, + 543, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\langle \\text{bos} \\rangle" + }, + { + "bbox": [ + 55, + 543, + 296, + 686 + ], + "type": "text", + "content": " token can limit the transfer of relational information from self-attention maps " + }, + { + "bbox": [ + 55, + 543, + 296, + 686 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 55, + 543, + 296, + 686 + ], + "type": "text", + "content": " to text embeddings, as attention scores for other tokens remain much smaller, minimizing their influence in each self-attention layer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "To provide a mathematical justification, we express the difference between the output and input of one block of the" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 329, + 66, + 457, + 190 + ], + "blocks": [ + { + "bbox": [ + 329, + 66, + 457, + 190 + ], + "lines": [ + { + "bbox": [ + 329, + 66, + 457, + 190 + ], + "spans": [ + { + "bbox": [ + 329, + 66, + 457, + 190 + ], + "type": "image", + "image_path": "6020acbb489405fffbd5851eb34f530e8e880266f49002eec1ea244495d18e03.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 194, + 398, + 204 + ], + "lines": [ + { + "bbox": [ + 387, + 194, + 398, + 204 + ], + "spans": [ + { + "bbox": [ + 387, + 194, + 398, + 204 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 460, + 71, + 539, + 191 + ], + "blocks": [ + { + "bbox": [ + 460, + 71, + 539, + 191 + ], + "lines": [ + { + "bbox": [ + 460, + 71, + 539, + 191 + ], + "spans": [ + { + "bbox": [ + 460, + 71, + 539, + 191 + ], + "type": "image", + "image_path": "824ea9085e315c096db8dc7d07f0037be86c830e4504eba7eaded7809045e6c6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 194, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 493, + 194, + 504, + 204 + ], + "spans": [ + { + "bbox": [ + 493, + 194, + 504, + 204 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "text", + "content": "Figure 6. (a) The distributions of text self-attention value " + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "inline_equation", + "content": "(\\mathsf{T}_{ij}^{\\prime})" + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "text", + "content": " in eq.(2)) for bound tokens (attribute " + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "text", + "content": ", object " + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "text", + "content": ") and unbound tokens (attribute " + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "text", + "content": ", object " + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "text", + "content": ") / (object " + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "inline_equation", + "content": "1" + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "text", + "content": ", object " + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "inline_equation", + "content": "2" + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "inline_equation", + "content": "m, n \\in \\{1, 2\\}" + }, + { + "bbox": [ + 313, + 208, + 555, + 297 + ], + "type": "text", + "content": ". The separate distributions indicate text self-attention maps can indeed represent the syntactic relationships. (b) Comparison of text self-attention probability histograms between token and non tokens on 100 prompts: The probability allocated to is on average 20 times larger than that of other tokens." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 316, + 473, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 473, + 328 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 473, + 328 + ], + "type": "text", + "content": "self-attention module for a token " + }, + { + "bbox": [ + 314, + 316, + 473, + 328 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 314, + 316, + 473, + 328 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 370, + 332, + 554, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 332, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 370, + 332, + 554, + 365 + ], + "type": "interline_equation", + "content": "\\mathbf {o} _ {i} ^ {(\\ell , h)} = \\sum_ {j = 1} ^ {i} T _ {i j} ^ {(\\ell , h)} W _ {\\mathrm {v}} ^ {(\\ell , h)} \\mathbf {e} _ {j} ^ {(\\ell)}, \\tag {9}", + "image_path": "f81cc12c7a984ced222ccc29d6d818b082647101abcdeec85e46de936af2d1e2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 369, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 369, + 555, + 433 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{v}}^{(\\ell,h)} \\in \\mathbb{R}^{D_{\\mathrm{e}} \\times H_{\\mathrm{e}} D_{\\mathrm{e}}}" + }, + { + "bbox": [ + 313, + 369, + 555, + 433 + ], + "type": "text", + "content": " is a parameter, " + }, + { + "bbox": [ + 313, + 369, + 555, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i \\in \\mathbb{R}^{H_{\\mathrm{e}} D_{\\mathrm{e}}}" + }, + { + "bbox": [ + 313, + 369, + 555, + 433 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 369, + 555, + 433 + ], + "type": "inline_equation", + "content": "T_{ij}^{(\\ell,h)}" + }, + { + "bbox": [ + 313, + 369, + 555, + 433 + ], + "type": "text", + "content": " is the self-attention matrix in eq. (1). We consider the situation where attention sink occurs, that is the attention weights for the token is much higher than the rest of the sequence:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 437, + 554, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 437, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 353, + 437, + 554, + 466 + ], + "type": "interline_equation", + "content": "\\epsilon = \\frac {\\sum_ {j \\neq 1} ^ {i} T _ {i j}}{T _ {i 1}} \\ll 1, \\quad i = 2, \\dots , s. \\tag {10}", + "image_path": "b62e5d44c09acc6faeafb3f3b25ae6010805b16c0acc4f0814ad0c0c6da42210.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 468, + 518, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 518, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 518, + 480 + ], + "type": "text", + "content": "In Appendix A, we prove the following statement:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 486, + 492, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 486, + 492, + 498 + ], + "spans": [ + { + "bbox": [ + 314, + 486, + 492, + 498 + ], + "type": "text", + "content": "Proposition 2. Define matrix " + }, + { + "bbox": [ + 314, + 486, + 492, + 498 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{s \\times s}" + }, + { + "bbox": [ + 314, + 486, + 492, + 498 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 367, + 501, + 553, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 501, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 367, + 501, + 553, + 519 + ], + "type": "interline_equation", + "content": "R _ {i j} = \\mathbf {e} _ {i} ^ {(\\ell) \\top} W _ {v} ^ {(\\ell , h) \\top} W _ {v} ^ {(\\ell , h)} \\mathbf {e} _ {j} ^ {(\\ell)}, \\tag {11}", + "image_path": "60854773aa46df088e97192e688a0e42be396273a77b04ebd2e3125f1c2352ef.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 521, + 443, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 443, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 443, + 533 + ], + "type": "text", + "content": "and suppose it has the property" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 536, + 553, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 536, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 318, + 536, + 553, + 570 + ], + "type": "interline_equation", + "content": "\\frac {\\left| R _ {m n} \\right|}{R _ {1 1}} \\sim \\mathcal {O} (1 / \\epsilon), \\quad \\frac {\\left| R _ {1 m} \\right|}{R _ {1 1}} \\sim \\mathcal {O} (1), \\quad 1 < m, n \\leq s, \\tag {12}", + "image_path": "ea056264e949ebb73350663e165a4d3c41960640bd432f744e90733e95a2ca04.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 571, + 552, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 552, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 552, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 571, + 552, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_1^{(\\ell)}" + }, + { + "bbox": [ + 313, + 571, + 552, + 585 + ], + "type": "text", + "content": " is the bos embedding. Then the following holds:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 370, + 588, + 553, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 588, + 553, + 606 + ], + "spans": [ + { + "bbox": [ + 370, + 588, + 553, + 606 + ], + "type": "interline_equation", + "content": "\\cos \\left(\\mathbf {o} _ {i} ^ {(\\ell , h)}, \\mathbf {o} _ {j} ^ {(\\ell , h)}\\right) = 1 - \\mathcal {O} (\\epsilon). \\tag {13}", + "image_path": "964cd785004d6c0b9a681b6e63d7f7c057d0276f3b0d30926b0fea80c27986f5.jpg" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 544, + 609, + 553, + 619 + ], + "blocks": [ + { + "bbox": [ + 544, + 609, + 553, + 619 + ], + "lines": [ + { + "bbox": [ + 544, + 609, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 544, + 609, + 553, + 619 + ], + "type": "image", + "image_path": "6aaf0e5f90f7cdd380846164eee8f66c00f8e88fab2309db70d9d6d5924f58e0.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 626, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 555, + 699 + ], + "type": "text", + "content": "This suggests that cosine similarity between token vectors " + }, + { + "bbox": [ + 313, + 626, + 555, + 699 + ], + "type": "inline_equation", + "content": "(\\mathbf{o}_i^{(\\ell ,h)})" + }, + { + "bbox": [ + 313, + 626, + 555, + 699 + ], + "type": "text", + "content": " -potentially influencing cross-attention maps remains barely changed across text self-attention layers due to attention sink on the. " + }, + { + "bbox": [ + 313, + 626, + 555, + 699 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{bos}\\rangle" + }, + { + "bbox": [ + 313, + 626, + 555, + 699 + ], + "type": "text", + "content": " token. In other words, the attention sink can hinder the accurate encoding of self-attention maps into embeddings." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 325, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 713 + ], + "type": "text", + "content": "Our findings reveal text embedding alone are insufficient" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8035" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "for generating semantically aligned images. On the other hand, we notably show the potential of transferring neglected syntactic information from text self-attention maps to the cross-attention to enhance T2I semantic alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 126, + 289, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 289, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 289, + 140 + ], + "type": "text", + "content": "4.2. Text Self-Attention Maps (T-SAM) Guidance" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 144, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 144, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 144, + 295, + 251 + ], + "type": "text", + "content": "In the previous section, we show text self-attention maps capture syntax information within a sentence. Building on this insight, we propose leveraging the self-attention maps within the text encoder—a component of diffusion models—to enhance cross-attention maps. By minimizing the distance between the similarity matrix of the cross-attention maps and the text self-attention matrix, our approach ensures that embedded syntactic relationships are effectively transferred to cross-attention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 252, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 323 + ], + "type": "text", + "content": "Our method optimizes cross-attention maps during inference, adjusting their similarity matrix to align with the text self-attention matrix " + }, + { + "bbox": [ + 55, + 252, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathsf{T}" + }, + { + "bbox": [ + 55, + 252, + 296, + 323 + ], + "type": "text", + "content": ". The normalized cosine similarity matrix, S (defined in eq.(7)), is used as the cross-attention similarity matrix. This is achieved by simply minimizing the loss function:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 325, + 295, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 325, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 325, + 295, + 356 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(z _ {t}\\right) = \\sum_ {i = 1, j \\leq i} ^ {s} \\rho_ {i} \\left| \\mathrm {T} _ {i j} ^ {\\gamma} - \\mathrm {S} _ {i j} \\left(z _ {t}\\right) \\right|, \\tag {14}", + "image_path": "575dd843609fc5e7a87de3197a5fd0a134990cc5e145d035ef17f110413f1218.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "text", + "content": "where the exponent " + }, + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "text", + "content": " acts to amplify larger values and compress smaller ones so the effect of controlling temperature and " + }, + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\rho_{i} = i / s" + }, + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "text", + "content": ". For example, if two words in the prompt have negligible syntactic relation according to the text self-attention matrix, i.e. " + }, + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathsf{T}_{ij}\\approx 0" + }, + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "text", + "content": ", we demand that their similarity of cross-attention maps must not be similar: " + }, + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathsf{S}_{ij}\\approx 0" + }, + { + "bbox": [ + 55, + 358, + 296, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 453 + ], + "type": "text", + "content": "In practice, this optimization will be applied only to " + }, + { + "bbox": [ + 55, + 430, + 296, + 453 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 430, + 296, + 453 + ], + "type": "text", + "content": " at a few denoising steps during inference as followed:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 456, + 295, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 456, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 125, + 456, + 295, + 469 + ], + "type": "interline_equation", + "content": "z _ {t} ^ {\\prime} = z _ {t} - \\alpha \\cdot \\nabla_ {z _ {t}} \\mathcal {L} (z _ {t}). \\tag {15}", + "image_path": "fa5f1cd5afa053d5b6e11fc7178726e3f9f45f294b4b90b0a39088e6a46241d1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 478, + 137, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 137, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 137, + 491 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "text", + "content": "This section presents our empirical study on the effectiveness of T-SAM in enhancing semantic alignment across diverse T2I diffusion models, including SD v1.5 [29] and PixArt-" + }, + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "text", + "content": " [4], evaluated both quantitatively and qualitatively. To the best of our knowledge, this is the first work to introduce attention control in PixArt-" + }, + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 498, + 295, + 581 + ], + "type": "text", + "content": ", underscoring the novelty of our approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Prompt datasets. We evaluate our approach on diverse text prompts using two existing benchmarks. First, the TIFA v1.0 benchmark [13] is a large-scale text-to-image generation dataset featuring a wide range of sentence structures. This benchmark comprises 4,000 prompts, including 2,000 image captions from the COCO validation set [18], 161 prompts from DrawBench [30], 1,420 prompts from PartiPrompt used in Parti [39], and 500 texts from PaintSkill [6]. Second, we use structured prompt sets containing multiple objects and their corresponding attributes from Attend-n-Excite [3]. Prompts in this dataset are" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 118, + 550, + 180 + ], + "blocks": [ + { + "bbox": [ + 313, + 65, + 555, + 109 + ], + "lines": [ + { + "bbox": [ + 313, + 65, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 65, + 555, + 109 + ], + "type": "text", + "content": "Table 1. Evaluation results for the TIFA benchmark, including TIFA scores and CLIP similarity scores. External Info. indicates whether external information is used. CLIP-I (CLIP-T) refers to image-text (text-text) CLIP similarity. The base model is SD." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 118, + 550, + 180 + ], + "lines": [ + { + "bbox": [ + 318, + 118, + 550, + 180 + ], + "spans": [ + { + "bbox": [ + 318, + 118, + 550, + 180 + ], + "type": "table", + "html": "
External Info.TIFACLIP-ICLIP-T
SD×0.790.330.77
LB✓0.800.330.76
T-SAM×0.830.340.77
", + "image_path": "ef54ea184b0589b04c2eeb0e12706a8fffb0e484837a06cbb4a6e926dd271596.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 196, + 553, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 553, + 267 + ], + "type": "text", + "content": "grouped into three categories: Objects (e.g., “[attribute1] [object1] and [attribute2] [object2]”), Animals-Objects (e.g., “[animal] with [object]” or “[animal] and/with [attribute] [object]”), and Animals (e.g., “[animal1] and [animal2]”). We exclude the Animals category, as it lacks the complex syntactic relations as discussed in [28]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "text", + "content": "Implementation details. For SD, we use 50 sampling iterations, updating " + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "text", + "content": " at each denoising step from 1 to 25. We set " + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "inline_equation", + "content": "M = 256" + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "text", + "content": " in eq.(6) and " + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "text", + "content": " in eq.(14). For the TIFA benchmark, we generate one image per prompt with a shared seed across methods, while using a unique seed for each prompt. We set " + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\alpha = 40" + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "text", + "content": " in eq.(15). For Attend-n-Excite prompts, we generate results using 64 seeds, in line with standard practice, setting " + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\alpha = 10" + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "text", + "content": " and applying 20 iterative updates at " + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "inline_equation", + "content": "i \\in \\{0,10,20\\}" + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "text", + "content": ". See Appendix B.2 for details on PixArt- " + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 269, + 554, + 388 + ], + "type": "text", + "content": " as well as SD." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 389, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 544 + ], + "type": "text", + "content": "Evaluation metrics. To quantitatively evaluate the accuracy of generated images, we use two metrics: TIFA scores [13] (for TIFA benchmark) and CLIP similarity scores. TIFA scores measure how well generated images reflect the text prompts. In TIFA, questions for each prompt are generated by GPT-3.5 [22], and a vision-language model [16] provides answers. For CLIP [25] scores, we follow the protocol in [3] to evaluate both image-text and text-text similarity. Additional details on CLIP scores are provided in Appendix B.2. Moreover, we conducted a user study with 10 random prompts in Animals-Objects, generating four images per prompt using different seeds. A total of 25 participants select the image that best matches the text." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Baselines. When evaluating our method on SD, we compare our method with SD, Linguistic-Binding (LB) [28], Attend-n-Excite (A&E) [3] and CONFORM [21]. LB depending on external parsers (SpaCy [11]) is limited to attribute-binding tasks. This method can be applied to both the TIFA benchmark and the Attend-n-Excite (Objects and Animals-Objects) prompt sets, though they are used only when the text parser identifies (modifier, entity-noun) pairs within the given prompts. A&E and CONFORM require manual token selection per prompt, restricting their use to the fixed-template prompts (Objects and Animals-Objects) due to the high cost of selecting token indices for diverse prompts. We reproduce the images with SD, CONFORM, and LB based on SD v1.5 and using the same seeds. For PixArt-" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": ", we com" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8036" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 67, + 504, + 133 + ], + "blocks": [ + { + "bbox": [ + 108, + 67, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 108, + 67, + 504, + 133 + ], + "spans": [ + { + "bbox": [ + 108, + 67, + 504, + 133 + ], + "type": "image", + "image_path": "063924eeb3667c09d7b26adb25ff4af23e395b36fa8b01f168b768aec1cdfd45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 144, + 555, + 167 + ], + "lines": [ + { + "bbox": [ + 54, + 144, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 144, + 555, + 167 + ], + "type": "text", + "content": "Figure 7. Accuracy for each question type in the FIFA benchmark. Our method achieves the best performance in most categories. The attribute includes properties such as large, small, young, etc. The base model is SD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 176, + 499, + 538 + ], + "blocks": [ + { + "bbox": [ + 109, + 176, + 499, + 538 + ], + "lines": [ + { + "bbox": [ + 109, + 176, + 499, + 538 + ], + "spans": [ + { + "bbox": [ + 109, + 176, + 499, + 538 + ], + "type": "image", + "image_path": "04a9f402dc2566b67bf450c19545222ba121bc36dd3e86c89de884ad49afdd02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 544, + 555, + 566 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 555, + 566 + ], + "type": "text", + "content": "Figure 8. Comparison of our method (T-SAM) on SD with recent state-of-the-art methods on prompts from Objects and Animals-Objects. The images corresponding to the same position across different methods are generated using the same seed. Best in zoom." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 581, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 297, + 605 + ], + "type": "text", + "content": "pare only with the base model, as no prior attention-control methods have been applied to it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 113, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 113, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 113, + 623 + ], + "type": "text", + "content": "5.1. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Quantitative results. Table 1 shows the evaluation results in TIFA benchmark and Figure 7 illustrates the breakdown of TIFA scores across question types, using SD as the base model. We highlight that our approach (T-SAM) outperforms SD and LB on complex syntactic prompts, where other baselines are inapplicable. While LB demonstrates improvements over SD in the color, shape, and material" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "content": "categories, which are closely related to attribute-binding tasks, it fails to enhance accuracy in the activity and counting categories. This underscores LB's inability to capture diverse word relationships, even when using external text parsers. In contrast, our method shows improvements over SD in nearly all categories including color, shape, and activity, except for attribute (e.g., properties such as large, small, young), which differ from the term \"attribute\" used in this study, demonstrating its versatility across diverse word relationships. Additionally, it achieves higher CLIP scores than the baselines, proving its superior semantic alignment." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8037" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 79, + 555, + 198 + ], + "blocks": [ + { + "bbox": [ + 66, + 59, + 543, + 71 + ], + "lines": [ + { + "bbox": [ + 66, + 59, + 543, + 71 + ], + "spans": [ + { + "bbox": [ + 66, + 59, + 543, + 71 + ], + "type": "text", + "content": "Table 2. Average CLIP similarity scores between text prompts and images generated with 64 different seeds. The base model is SD." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 79, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 79, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 79, + 555, + 198 + ], + "type": "table", + "html": "
External Info.Image-Text PromptPrompt-Caption
Full PromptMinimum Object
ObjectsAnimals-ObjectsObjectsAnimals-ObjectsObjectsAnimals-Objects
SDX0.340.340.250.260.760.80
LB✓0.360.350.270.270.800.83
CONFORM✓0.360.360.280.280.810.85
A&E✓0.360.350.270.260.810.83
T-SAMX0.360.370.280.280.810.85
", + "image_path": "0bbb804223c22d3d6f65c519875237e3a5743aa3a11df2d0a0f521c32828768c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 251, + 304, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "type": "text", + "content": "Table 3. CLIP similarity scores on prompts from Objects and Animals-Objects using PixArt-" + }, + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "type": "text", + "content": " as the base model. CLIP-I (CLIP-T) refers to image-text (text-text) CLIP similarity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 251, + 304, + 327 + ], + "lines": [ + { + "bbox": [ + 56, + 251, + 304, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 304, + 327 + ], + "type": "table", + "html": "
Prompt SetMethodCLIP-TCLIP-I (Full, Min)
ObjectsPixArt-α0.810.36, 0.27
T-SAM0.820.37, 0.28
Animals-ObjectsPixArt-α0.850.36, 0.27
T-SAM0.860.37, 0.28
", + "image_path": "56cf01205037dee4d8f1dacfa8d2c20650a98dff824cf50a16cbaab692a53eb5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 70, + 335, + 282, + 517 + ], + "blocks": [ + { + "bbox": [ + 70, + 335, + 282, + 517 + ], + "lines": [ + { + "bbox": [ + 70, + 335, + 282, + 517 + ], + "spans": [ + { + "bbox": [ + 70, + 335, + 282, + 517 + ], + "type": "image", + "image_path": "50b673a935f082a57901dc6907a19c2fb74bfe339c454d5255c3c5512c94fe24.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 525, + 295, + 593 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 593 + ], + "type": "text", + "content": "Figure 9. Qualitative comparison using prompts from MSCOCO contained in TIFA benchmark. Our approach successfully generates multiple elements, including a sewing machine in the first prompt, people in the second, popsicle in the third, and cow in the fourth. SD misses some elements and LB generates identical images to SD when no relations are extracted from the text parser." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "In Objects and Animals-Objects, T-SAM on SD performs comparably to the state-of-the-art CONFORM, which, unlike our approach, requires manually defined token indices for positive and negative groups, as shown in Table 2. Our method outperforms LB and A&E, which also rely on external inputs. This demonstrates extracting syntactic information from text self-attention maps can be more effective than relying on text parsers or manually selecting tokens for optimization. An additional ablation study is provided in Ap" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 214, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 553, + 262 + ], + "type": "text", + "content": "pendix B.2. Moreover, Table 3 shows that T-SAM is also effective on PixArt-" + }, + { + "bbox": [ + 313, + 214, + 553, + 262 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 214, + 553, + 262 + ], + "type": "text", + "content": " with *Attend-n-Excite* prompts, demonstrating the generalizability of our method to a different text encoder (T5-XXL) and model architecture (DiT)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 262, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 453 + ], + "type": "text", + "content": "Qualitative results. In Figure 8, T-SAM either outperforms or performs comparably to the baselines in Objects and Animals-Objects. SD often omits objects (e.g., clock, apple, or bow) or misbinds attributes (e.g., blue clock). CONFORM and LB improve over SD but have limitations. LB is restricted to attribute-binding and fails on prompts without modifiers (e.g., An elephant with a bow), producing results identical to SD CONFORM can impose overly rigid separations, as seen in the first example. T-SAM, by leveraging smoother linguistic cues from text attention maps—rather than binary positive/negative pairs—achieves broader applicability and more consistent improvements. See Figure 9 for more comparison of images generated by T-SAM and the baselines, showcasing effectiveness of T-SAM with complex prompts, such as MSCOCO captions from the TIFA benchmark." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 538 + ], + "type": "text", + "content": "For PixArt-" + }, + { + "bbox": [ + 313, + 453, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 453, + 556, + 538 + ], + "type": "text", + "content": ", Figure D illustrates the effectiveness of our method. Although the base model is already strong compared to SD, T-SAM further improves its performance. User study. The user study result in Table D show that T-SAM on SD outperforms LB and achieves performance comparable to CONFORM. Additionally, Table E shows T-SAM on PixArt-" + }, + { + "bbox": [ + 313, + 453, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 453, + 556, + 538 + ], + "type": "text", + "content": " surpasses the base PixArt-" + }, + { + "bbox": [ + 313, + 453, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 453, + 556, + 538 + ], + "type": "text", + "content": " model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 568, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 712 + ], + "type": "text", + "content": "To enhance fidelity in T2I diffusion models, we improved cross-attention maps by aligning their similarity matrix with text self-attention maps. Our approach is based on two insights: (1) similar text embeddings produce similar cross-attention maps, but (2) syntactic relations are missed in embeddings but captured by text self-attention maps. T-SAM enabled cross-attention to better capture syntactic structure, significantly improving text-to-image fidelity across diverse sentence structures, without requiring external tools. Nonetheless, our method has limitations such as inaccuracies in object counting and potential noise introduced by text attention maps, leaving room for future improvement." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8038" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 115 + ], + "type": "text", + "content": "This work was partially supported by NIH (1U01CA269192)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 145, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 145, + 296, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 145, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 61, + 145, + 296, + 210 + ], + "type": "text", + "content": "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2283-2293, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 212, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 212, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 212, + 296, + 255 + ], + "type": "text", + "content": "[2] Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Alignit: Enhancing prompt alignment in customization of text-to-image models. arXiv preprint arXiv:2406.18893, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 256, + 296, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 256, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 61, + 256, + 296, + 311 + ], + "type": "text", + "content": "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 312, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 312, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 61, + 312, + 296, + 367 + ], + "type": "text", + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 296, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 422 + ], + "type": "text", + "content": "[5] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart" + }, + { + "bbox": [ + 62, + 369, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 62, + 369, + 296, + 422 + ], + "type": "text", + "content": "sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 424, + 296, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 296, + 478 + ], + "type": "text", + "content": "[6] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3043-3054, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 545 + ], + "type": "text", + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 547, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 296, + 601 + ], + "type": "text", + "content": "[8] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 645 + ], + "type": "text", + "content": "[9] Xiangming Gu, Tianyu Pang, Chao Du, Qian Liu, Fengzhuo Zhang, Cunxiao Du, Ye Wang, and Min Lin. When attention sink emerges in language models: An empirical view. arXiv preprint arXiv:2410.10781, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[10] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[11] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "neural networks and incremental parsing. To appear, 7(1): 411-420, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 138 + ], + "type": "text", + "content": "[12] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 205 + ], + "type": "text", + "content": "[13] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 6, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 248 + ], + "type": "text", + "content": "[14] Jiaxiu Jiang, Yabo Zhang, Kailai Feng, Xiaohe Wu, and Wangmeng Zuo. Mc2: Multi-concept guidance for customized multi-concept generation. arXiv preprint arXiv:2404.05268, 2024.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 294 + ], + "type": "text", + "content": "[15] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7701-7711, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "text", + "content": "[16] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, et al. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. arXiv preprint arXiv:2205.12005, 2022. 6, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 405, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 470 + ], + "type": "text", + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "text", + "content": "[19] Oscar Manas, Pietro Astolfi, Melissa Hall, Candace Ross, Jack Urbanek, Adina Williams, Aishwarya Agrawal, Adriana Romero-Soriano, and Michal Drozdal. Improving text-to-image consistency via automatic prompt optimization. arXiv preprint arXiv:2403.17804, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "text", + "content": "[20] Arash Marioriyad, Mohammadali Banayeeanzade, Reza Abbasi, Mohammad Hossein Rohban, and Mahdieh Soleymani Baghshah. Attention overlap is responsible for the entity missing problem in text-to-image diffusion models! arXiv preprint arXiv:2410.20972, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 582, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 555, + 635 + ], + "type": "text", + "content": "[21] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9005–9014, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 658 + ], + "type": "text", + "content": "[22] OpenAI. Openai gpt-3 api [gpt-3.5-turbo], 2024. Available at: https://platform.openai.com. 6, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8039" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[24] Zipeng Qi, Lichen Bai, Haoyi Xiong, et al. Not all noises are created equally: Diffusion noise selection and optimization. arXiv preprint arXiv:2407.14041, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 173 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[26] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[28] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 6, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "text", + "content": "[30] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "type": "text", + "content": "[31] Dazhong Shen, Guanglu Song, Zeyue Xue, Fu-Yun Wang, and Yu Liu. Rethinking the spatial inconsistency in classifier-free diffusion guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9370-9379, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 511, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 543 + ], + "type": "text", + "content": "[32] Mingjie Sun, Xinlei Chen, J Zico Kolter, and Zhuang Liu. Massive activations in large language models. arXiv preprint arXiv:2402.17762, 2024. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 567 + ], + "type": "text", + "content": "[33] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 569, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 610 + ], + "type": "text", + "content": "[34] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. arXiv preprint arXiv:2312.03626, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[35] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291-26303, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 296 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[37] Mingyang Yi, Aoxue Li, Yi Xin, and Zhenguo Li. Towards understanding the working mechanism of text-to-image diffusion model. arXiv preprint arXiv:2405.15330, 2024. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[38] Hu Yu, Hao Luo, Fan Wang, and Feng Zhao. Uncovering the text embedding in text-to-image diffusion models. arXiv preprint arXiv:2404.01154, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "text", + "content": "[41] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Tiviatis Sim, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. arXiv preprint arXiv:2403.06381, 2024. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8040" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_content_list.json b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d17f5f452a7d0f37f552b8050c51aa162c928507 --- /dev/null +++ b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_content_list.json @@ -0,0 +1,1286 @@ +[ + { + "type": "text", + "text": "Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction", + "text_level": 1, + "bbox": [ + 104, + 111, + 893, + 156 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanshan Huang $^{1}$ , Haoxuan Li $^{2}$ , Chunyuan Zheng $^{3}$ , Mingyuan Ge $^{1}$ , Wei Gao $^{1}$ , Lei Wang $^{1}$ , Li Liu $^{1*}$ $^{1}$ School of Big Data & Software Engineering, Chongqing University", + "bbox": [ + 102, + 184, + 895, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Center for Data Science, Peking University $^{3}$ School of Mathematical Sciences, Peking University $\\{\\text { shanshanhuang, dcsliuli } \\} @ \\text { cqu.edu.cn, } \\{\\text { hxli, cyzheng } \\} @ \\text {stu.pku.edu.cn, }$ $\\{\\text { leiwangtt, mingyuange } \\} @ \\text {stu.cqu.edu.cn, }$ wgao2002@gmail.com", + "bbox": [ + 102, + 220, + 895, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fashion image editing is a valuable tool for designers to convey their creative ideas by visualizing design concepts. With the recent advances in text editing methods, significant progress has been made in fashion image editing. However, they face two key challenges: spurious correlations in training data often induce changes in other areas when editing an area representing the intended editing concept, and these models typically lack the ability to edit multiple concepts simultaneously. To address the above challenges, we propose a novel Text-driven Fashion Image editing framework called T-FIT to mitigate the impact of spurious correlation by integrating counterfactual reasoning with compositional concept learning to precisely ensure compositional multi-concept fashion image editing relying solely on text descriptions. Specifically, T-FIT includes three key components. (i) Counterfactual abduction module, which learns an exogenous variable of the source image by a denoising U-Net model. (ii) Concept learning module, which identifies concepts in fashion image editing—such as clothing types and colors and projects a target concept into the space spanned from a series of textual prompts. (iii) Concept composition module, which enables simultaneous adjustments of multiple concepts by aggregating each concept's direction vector obtained from the concept learning module. Extensive experiments show that our method can achieve state-of-the-art performance on various fashion image editing tasks, including single-concept editing (e.g., sleeve length, clothing type) and multi-concept editing (e.g., color & sleeve length).", + "bbox": [ + 91, + 325, + 483, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 773, + 220, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fashion image editing has emerged as a pivotal field within computer vision, driven by the increasing need for personalized and dynamic visual modifications in the fashion industry [1, 33]. The diversity and high-quality image generation capabilities of large-scale text-image diffusion models", + "bbox": [ + 89, + 799, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such as Stable Diffusion [23], DALL-E 2 [24], and Imagegen [22] has inspired many subsequent efforts to leverage pre-trained large-scale models for text-driven fashion image editing (TFIE) [12, 19, 32]. A foundational aspect of this field is single-concept editing, where a specific concept [35, 37] like color, texture, or style of an item is altered while keeping the rest of the fashion image the same.", + "bbox": [ + 511, + 295, + 905, + 400 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the existing methods still suffer from several limitations. First, these methods [2, 6, 41] often encounter challenges related to spurious correlation and unwanted associations learned by the pre-trained model (e.g., Stable Diffusion [23]) due to biases in training data. For example, several methods [28, 29] may erroneously associate a specific clothing type (e.g., sleeve length) with a specific cloth style (e.g., garment length), as shown in Fig. 1, resulting in unintended modifications outside the target concept. This is because the long sleeves always co-occur with long garments, inducing spurious correlations in the training data. Therefore, eliminating spurious correlations is critical to ensuring precise, isolated modifications. There are many methods to address spurious correlation [5, 11, 13, 31, 36, 38, 39]. For example, Li et al. [16] propose a novel machine unlearning method to reduce the instance weight of biased samples in e-commerce and Sanchez et al. [25] discuss important challenges present in healthcare applications such as processing high-dimensional and unstructured data, as well as temporal relationships. However, due to differences in data structure and final goal, previous methods could not be adapted to image editing scenarios.", + "bbox": [ + 511, + 402, + 905, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Second, existing methods [32, 33] are mainly limited to editing the texture of fashion images, and it is difficult to achieve non-rigid editing, such as precisely adjusting the length of sleeves or skirts. As shown in Fig. 1, UltraEdit [41] and Turbo-edit [6] cannot achieve non-rigid editing, e.g., turning a long skirt into a mini skirt; altering short sleeves into long sleeves, which may attribute to the poor understanding of the concept from the prompt.", + "bbox": [ + 511, + 750, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Third, existing methods [4, 14, 15] typically focus on editing a single concept in the source image. While single-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28726", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b96dd172db952754bc9d5c9cd4e11fdeba3d06d7f1e038f1091050054ac3794a.jpg", + "image_caption": [ + "Figure 1. Two examples to illustrate the spurious correlation and conceptual omission issues in fashion image editing. Unlike other methods, our method alleviates the issue of over-editing that exists in fashion image editing due to spurious correlation (e.g., sleeve length and garment length, mini skirt and plaid style), as well as the confusion or lack of concepts when editing multiple concepts." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 470, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "concept methods enable the modification of multiple concepts sequentially, i.e., changing the concepts one by one, they present certain drawbacks. Specifically, these methods either introduce concept omission due to the difficulty of identifying each concept from the prompt; or they lead to unexpected changes due to interactions between different concepts, which is still caused by spurious correlation [10]. For example, modifications made to later concepts may affect previous modifications. From Fig. 1, we can observe that G & R [28] incorrectly adds a \"leather bag\" when editing the skirt fabric to leather and LEDs++ [2] causes other unrelated attributes changes when changing the length and fabric of the skirt, such as the color of the gloves.", + "bbox": [ + 88, + 337, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, we develop a Text-driven Fashion Image editing framework (T-FIT) that relies solely on textual prompts, enabling users to perform both single- and multi-concept editing on a single fashion image. Specifically, T-FIT adopts Stable Diffusion [23] and introduces a counterfactual abduction module to address the spurious correlation and concept learning module with concept composition to achieve precise single- and multi-concept editing as shown in Fig. 2. Our contributions are as follows:", + "bbox": [ + 89, + 536, + 482, + 672 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal the issue of spurious correlation and concept omission in fashion image editing and propose T-FIT, a text-driven framework to address these issues.", + "- We propose counterfactual abduction (CA) to capture visual content and a concept learning module (CLM) to identify and disentangle concepts, mitigating issues of spurious correlation and non-rigid editing.", + "- We propose a concept composition strategy that enables simultaneous, fine-grained editing across multiple concepts by adjusting the weight of each learned concept to address the concept omission problem.", + "- Extensive empirical evaluation, in terms of both automated and human assessment metrics, qualitatively and quantitatively demonstrates that our T-FIT approach significantly enhances the quality of fashion image editing." + ], + "bbox": [ + 89, + 674, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Methodology", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 513, + 111, + 718, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first formulate the image editing problem as follows: given a source fashion image $I_{S}$ , a corresponding text prompt $P$ that describes the contents of $I_{S}$ , a set of concept prompts $P_{C^i} = \\{p_0^i, p_1^i, \\ldots, p_{N-1}^i\\}$ used to construct a concept $C^i$ (e.g., clothing type, color, fabric) that the user wishes to edit, and a target text prompt $P'$ describing the final editing goal, our objective is to generate a new fashion image $I_T$ by editing $I_S$ in alignment with $P'$ while preserving other areas that are unrelated to the concept $C^i$ .", + "bbox": [ + 511, + 135, + 906, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 277, + 643, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the context of fashion image editing, concepts $C = \\{C^1,\\dots ,C^i,\\dots ,C^m\\}$ are a distinct and human-interpretable feature or attribute within an image, encompassing elements like sleeve length, fabric type, color, and cloth type. Each concept $C^i$ contains mutually selected attributes, where each attribute embodies certain shared properties. For instance, the attributes \"long sleeves\" and \"short sleeves\" collectively represent the concept of \"sleeve length\", while \"woman\" and \"man\" represent the concept of \"gender\". The concept representations $s(C)$ can be derived from the intermediate representation of some text-to-image generative models (e.g., Stable Diffusion) pre-trained over annotated datasets that link visual attributes to textual descriptions [27, 35]. Here, $s:\\mathbb{C}\\to \\mathbb{R}^d$ is the concept representation function, where $\\mathbb{C}$ is the concept space containing all possible concepts, and $\\mathbb{R}^d$ represents the representation space with dimension $d$ . Following [27, 35], we focus on the compositional concepts, as defined in Definition 1.", + "bbox": [ + 511, + 301, + 906, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1 (Compositional Concepts). For concepts $C^i, C^j \\in \\mathbb{C}$ , the concept representation $s: \\mathbb{C} \\to \\mathbb{R}^d$ is considered compositional if there exist positive weights $w_i, w_j \\in \\mathbb{R}^+$ such that:", + "bbox": [ + 511, + 579, + 906, + 642 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ns (C ^ {i} \\cup C ^ {j}) = w _ {i} s (C ^ {i}) + w _ {j} s (C ^ {j}).\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 646, + 830, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This definition indicates that the representation of the composed concepts corresponds to the weighted sum of the individual concept representations in the representation space. Following this definition, this study aims to develop a framework for learning and composing conceptual representations from pre-trained text-to-image models (i.e., Stable Diffusion) to achieve flexible and targeted fashion image editing based on these compositional concepts.", + "bbox": [ + 511, + 672, + 906, + 796 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Counterfactual Abduction for Image Editing", + "text_level": 1, + "bbox": [ + 511, + 801, + 893, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address spurious correlation issues inherent in existing image editing methods, we introduce a counterfactual abduction approach that employs an abduction loss to infer unknown exogenous variables $U$ for each image. These variables capture the additional visual content of source fashion", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28727", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2555b61cabbaea1513b210b88ac1c07b76452ba585d97de9a1b00128e71766e5.jpg", + "image_caption": [ + "(c) Concept Composition" + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 472, + 238 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/51745256b6111e78b3730bc15ac80724ee41a28059635bfd2d0cdbcca7209aed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 104, + 883, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/33f9009859475a5df608025691b2a57bbfb661822990095a6a0ce7486d7c5f4b.jpg", + "image_caption": [ + "Figure 2. Overview of our T-FIT for fashion image editing. We divide the entire editing process into two stages. In the first stage, we perform counterfactual abduction to infer exogenous variables $U$ that capture the core visual contents of the source fashion image. In the second stage, we employ a concept learning module to discover the target concept from representation space and design a concept composition strategy for combining multiple aspects of the fashion image, such as fabric, color, or pattern, while maintaining consistency with the original image." + ], + "image_footnote": [], + "bbox": [ + 101, + 239, + 893, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "images, aiming to reduce uncertainty in fashion image editing and prevent unintended changes in the structure or identity of the source image. In other words, by incorporating $U$ , our method ensures that the edited fashion image $I_{T}$ integrates the influence of $P'$ while retaining other unrelated visual content from the source image $I_{S}$ unchanged, thus addressing spurious correlation introduced in pre-trained models from the causal perspective. Specifically, we train the abduction $U$ by optimizing the following Gaussian noise regression, a process analogous to training the reversed diffusion steps:", + "bbox": [ + 88, + 529, + 483, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {U} \\mathbb {E} _ {(t, \\epsilon)} | | \\epsilon - \\epsilon_ {U} (x _ {t}, t, \\varepsilon_ {s r c}) | | _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 703, + 482, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,I)$ is the Gaussian noise, $t\\in [0,T]$ represents a sampled time step, $\\epsilon_{U}$ represents a new denoising model trained on a pre-trained denoising U-Net, wherein $U$ is treated as a trainable parameter. All other model parameters, aside from those of the denoising U-Net, remain fixed during training. The text embedding $\\varepsilon_{src} = \\psi (P)$ is obtained by a pre-trained CLIP text encoder $\\psi (\\cdot)$ , and the noisy input at time $t$ can be obtained by $x_{t} = \\sqrt{\\alpha_{t}} x_{0} + \\sqrt{1 - \\alpha_{t}}\\epsilon$ . Here, $x_0 = I_S$ denotes the original image and $\\alpha_{t}$ is a predefined variance schedule [20, 34]. Following [26], we parameterize $U$ as the UNet LoRA", + "bbox": [ + 89, + 733, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in $\\epsilon_U$ , LoRA structure is incorporated across all attention layers, convolution layers, and feed-forward layers, which can be expressed by $W' = W + \\beta U_A \\cdot U_B$ . Here, $W \\in \\mathbb{R}^{d \\times d}$ denotes the original weight matrix, $U_A \\in \\mathbb{R}^{d \\times r}$ and $U_B \\in \\mathbb{R}^{r \\times d}$ are low-rank matrices. Wang et al. [30] observed that the larger the time step in the diffusion model, the better the editable and the lower the fidelity. For this reason, the annealing parameter $\\beta = \\frac{1 - \\kappa}{T^2} (t - T)^2 + \\kappa$ is proposed to improve the editability. In our experiment, the rank $r$ is set to 512, $\\kappa \\in [0,1]$ is a constant.", + "bbox": [ + 511, + 529, + 906, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Compositional Concept Learning for Editing", + "text_level": 1, + "bbox": [ + 511, + 694, + 893, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unfortunately, the abduction of $U$ is inherently ill-posed [26], often leading to overfitting to the specific $P$ and $x$ . This overfitting reduces the editability of the generative model $G(\\cdot, U)$ and may cause challenges like concept omission or confusion, where distinctions between concepts may blur, or certain concepts might overshadow others. Therefore, $G(\\cdot, U)$ struggles to generate images that adhere to a new prompt $P'$ . To address this, we introduce a concept learning module that analyzes the representation space to identify composable concepts. Additionally, we propose a multi-concept composition strategy for flexible fashion image editing, enabling single and multi-concept continuous", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28728", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": " edits. For example, our method can simultaneously modify both the \"sleeve length\" (e.g., changing from long sleeves to short sleeves) and the \"color\" of the clothing (e.g., transitioning from red to blue). The intensity of these edits can also be adjusted, such as making the blue lighter or darker.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept Learning Module. Given a target concept $C^i$ , we aim to learn an interpretable direction vector $\\Delta C_t^i$ at time $t$ . The learned direction vector allows us to control the generation process by varying the latent $x_{t}$ in the desired directions $\\Delta C_t^i$ with weight $w_{i}$ , which is considered as $s_{new}^{i} = s_{src} + w_{i}\\Delta C_{t}^{i}$ . Here, $s_{src} = \\epsilon_U(x_t,t,\\varepsilon_{src})$ , $\\varepsilon_{src} = \\psi (P)$ is derived from pre-trained CLIP text encoder $\\psi (\\cdot)$ , and $\\epsilon_U(x_t,t,\\varepsilon_{src})$ is trained by counterfactual abduction and is fixed here. Specifically, we first define a set of text prompts $P_{C^i} = \\{p_0^i,p_1^i,\\dots,p_{N_i - 1}^i\\}$ designed to elicit distinct distributions for a target concept $C^i$ while maintaining a consistent distribution for another concept $C^j$ . Without loss of the generality, we define $p_0^i$ as the source prompt without redundant concepts. In essence, each prompt captures a unique representation of $C^i$ while keeping the representation of $C^j$ consistent. For instance, when editing the concept of \"fabric,\" we can formulate a series of prompts like \"a lace [fabric 0] dress,\" \"a silk [fabric 1] dress,\" and \"a [fabric $N_{i} - 1$ ] dress.\" These prompts are intended to generate variations in the fabric concept while keeping other aspects, like dress type.", + "bbox": [ + 89, + 169, + 482, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, to identify the direction vector for concept $C^i$ , we first define its representation space as $\\mathcal{R}_{C_t^i} \\coloneqq \\text{span}(\\{v_t^j - v_t^0 : j = 1, \\dots, N_i - 1\\})$ . By leveraging the span of conceptual space, we enable the model to edit not just predefined attributes, but also a broader range of related attributes. For example, by defining a structured representation space with \"blue,\" \"green,\" and \"yellow,\" our model can generalize to other colors in the spectrum, such as \"red\". Here, $v_t^j = \\epsilon_U(x_t, t, \\varepsilon_j)$ , each $\\varepsilon_j = \\psi(p_j^i)$ is derived from the pre-trained CLIP text encoder $\\psi(\\cdot)$ . Next, to compute the projection matrix $M_i$ for the $C^i$ -space within this representation space, we start by constructing matrix $S^i$ as", + "bbox": [ + 89, + 488, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t} ^ {i} := \\left[ v _ {t} ^ {1} - v _ {t} ^ {0}, \\dots , v _ {t} ^ {N _ {i} - 1} - v _ {t} ^ {0} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 681, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, to align changes in the representation space with the semantic structure of the concept space, we first apply singular value decomposition (SVD) on $S_{t}^{i}$ , i.e., $SVD(S_{t}^{i}) = V_{t}^{i}\\Sigma_{t}^{i}U_{t}^{iT}$ , to identify the principal orthogonal semantic directions. We then select the top $K$ components of $V_{t}^{i}$ to capture the most significant semantic variations (for simplicity, we denote this matrix as $V_{t}^{i}$ ), and construct the projection matrix $M_{t}^{i} = V_{t}^{i}V_{t}^{iT}$ , which projects the concept vector into $\\mathcal{R}_{C_t^i}$ and ensures that the changes in the representation space are corrected and aligned with the key semantic directions.", + "bbox": [ + 89, + 713, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, to derive the editing direction, we calculate the", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "score difference between source representation $s_{src}$ and target representation $s_{tar}$ as $\\Delta s_t = s_{tar} - s_{src}$ . We then project $\\Delta s_t$ onto the spanned concept space using $M_t^i$ , yielding $\\Delta C_t^i = \\Delta s_t M_t^i$ . This projection ensures the editing operation reflects the relevant semantic changes in the concept space. The modified score can be achieved by", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns = s _ {\\emptyset} + \\lambda \\left(s _ {n e w} ^ {i} - s _ {\\emptyset}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 191, + 903, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is guidance scale, $s_{\\emptyset}$ denotes the unconditional score, is given by $s_{\\emptyset} = \\epsilon_U(x_t,t,\\varepsilon_{\\emptyset})$ with $\\varepsilon_{\\emptyset} = \\psi (^{\\prime \\prime})$ In practice, we noticed that a few prompts are sufficient to identify the concept direction, which is attributed to the generalization ability of the span space $\\mathcal{R}_{C_t^i}$", + "bbox": [ + 511, + 219, + 903, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Compositional Concept Learning", + "text_level": 1, + "bbox": [ + 514, + 311, + 823, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: Diffusion model $\\epsilon_U(x_t,t,\\varepsilon)$ , pre-trained CLIP text encoder $\\psi (\\cdot)$ , guidance scale $\\lambda$ , concept weight $\\{w_{1},\\dots,w_{m}\\}$ , covariance matrix $\\sigma_t^2 I$ , empty prompt \" \", source prompt $P$ , target prompt $P^{\\prime}$ , prompts for build concept space $P_{C^i} = \\{p_0^i,\\ldots ,p_j^i,\\ldots ,p_{N_i - 1}^i\\}$", + "bbox": [ + 514, + 328, + 903, + 405 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {2}\\right) - \\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {0}\\right), \\dots ,\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 556, + 808, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {N _ {i} - 1}\\right) - \\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {0}\\right) \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 570, + 807, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS V D \\left(S _ {t} ^ {i}\\right) = V _ {t} ^ {i} \\Sigma_ {t} ^ {i} U _ {t} ^ {i T}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 599, + 748, + 618 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize sample $x_{t} \\sim \\mathcal{N}(0, I)$", + "2: $\\varepsilon_{\\emptyset}, \\varepsilon_{src}, \\varepsilon_{tar} = \\psi(\"\"), \\psi(P), \\psi(P')$ # Text embedding", + "3: $\\varepsilon_{j} = \\psi (p_{j}^{i})$ # Concept text embedding", + "4: for $t = T, \\dots, 1$ do", + "5: $s_{\\emptyset}\\gets \\epsilon_U(x_t,t,\\varepsilon_{\\emptyset})$ # Unconditional score", + "6: $s_{src}, s_{tar} \\gets \\epsilon_U(x_t, t, \\varepsilon_{src}), \\epsilon_U(x_t, t, \\varepsilon_{tar})$ # Conditional score", + "7: $\\Delta s_{t} = s_{tar} - s_{src}$ # Score difference", + "8: for $i = 1,2,\\dots,m$ do # m concepts", + "9: $S_{t}^{i}\\gets \\left[\\epsilon_{U}(x_{t},t,\\varepsilon_{1}) - \\epsilon_{U}(x_{t},t,\\varepsilon_{0}),\\right.$", + "10: Compute the top- $k$ left singular vectors $V_{i}$ via:", + "11: $M_{t}^{i} = V_{t}^{i}(V_{t}^{i})^{T}$ # Projection matrix", + "12: $\\Delta C_{t}^{i} = M_{t}^{i}\\Delta s_{t}\\#$ Editing direction", + "13: end for", + "14: $s \\gets s_{\\emptyset} + \\lambda((s_{src} + w_i \\Delta C_t^i) - s_{\\emptyset})$ # For single-concept editing", + "15: $s \\gets s_{\\emptyset} + \\lambda \\sum_{i=1}^{m} \\left( (s_{src} + w_i \\Delta C_t^i) - s_{\\emptyset} \\right) \\# \\text{For multi-concept editing}$", + "16: $x_{t - 1}\\sim \\mathcal{N}(x_t - s,\\sigma_t^2 I)$", + "17: end for" + ], + "bbox": [ + 517, + 405, + 903, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept Composition Strategy. Now, our method enables single-concept editing for fashion images, but many scenarios demand multi-concept compositional editing. A straightforward way to achieve this is to create a multi-concept space using a set of concept prompts. While feasible, this approach limits fine-grained control over individual concepts, such as adjusting shades of color in specific items. Alternatively, sequential single-concept edits can be applied, but this accumulates errors and is time-consuming.", + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28729", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/59c0fb811e2553c4ff9cd744d56b07a62f20d2692dc3d000c3ed1bf4f02703a6.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison with the state-of-art methods in single concept editing." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 893, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To overcome these issues, we propose a controllable multi-concept editing strategy. This strategy enables precise, simultaneous adjustments of multiple concepts by weighting each concept's direction vector, which is shown as", + "bbox": [ + 89, + 502, + 483, + 564 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {m} = s _ {\\emptyset} + \\lambda \\sum_ {i = 1} ^ {m} \\left(s _ {n e w} ^ {i} - s _ {\\emptyset}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 569, + 482, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_{new}^{i} = s_{src} + w_{i}\\Delta C_{t}^{i}$ , we can control the manipulation strength of each concept. The whole algorithm is shown in Algorithm 1.", + "bbox": [ + 89, + 614, + 483, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 89, + 671, + 223, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 89, + 696, + 287, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In our work, all experiments are conducted on a single NVIDIA RTX 4090 GPU with 24 GB of memory. Following the prior work [3, 4, 26], we employ the official pre-trained Stable Diffusion v2.1-base model as our foundational model, downsampling all images to a resolution of $512 \\times 512$ pixels for consistency across experiments. For the counterfactual abduction phase, We fine-tune the model on the single image using LoRA with a rank of 512 for 100 epochs, employing the Adam optimizer with a learning rate of 1e-4. To conserve memory, we adopt the mixed precision [17] and gradient accumulation strategies, setting both the accumulation step count and", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "batch size to 1. During the inference phase, we utilize the DDIM sampling with 80 steps and a classifier-free guidance scale of $\\lambda = 4$ . Additionally, we explore various values for the concept weight $w_{i}$ within the range [1, 5] to control the editing strength, where $i$ represents the $i$ -th edited concept.", + "bbox": [ + 511, + 502, + 906, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. To comprehensively evaluate the performance of our T-FIT in fashion image editing, we follow [14, 26, 40] by collecting a diverse set of fashion images from Unsplash1, including both full-body and half-body images. The selected images comprise various clothing types, such as long dresses, skirts, T-shirts, shorts, pants, and strapless tops. Our evaluation encompasses a diverse set of editing concepts, which include specific attributes (e.g., gender, color, patterns, fabric texture) as well as broader elements such as clothing type, garment length, and sleeve style. Each image in the experiment is annotated with $N_{i} + 2$ prompts, consisting of one prompt for the original image and $N_{i} + 1$ prompts for target images.", + "bbox": [ + 511, + 578, + 906, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We conduct comparisons with existing representative diffusion-based image editing methods, including LEDs [29], LEDs++ [2], TexFit [32], Guide-and-Rescale (G & R) [28], UltraEdit [41], and Turboedit [6].", + "bbox": [ + 511, + 775, + 906, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To comprehensively evaluate the performance of our T-FIT, we conduct assessments from both", + "bbox": [ + 511, + 835, + 906, + 864 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://unsplash.com/", + "bbox": [ + 531, + 886, + 705, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28730", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/468ca461f1b15c07a77790f11c9eaee95fd05eeb64efe41988d837d051d19e0e.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with the state-of-art methods in multiple concept editing." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 893, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "objective and subjective perspectives. Objectively, we follow the previous works [3, 8, 9] and use metrics in CLIP embedding space: CLIP image similarity (CLIP-I) [21] to measure identity preservation via calculate the cosine similarity between the edited image and the original image, and CLIP text-image direction similarity (CLIP-D) [7] to evaluate the correspondence between image changes and text changes. Subjectively, we qualitatively evaluate the fidelity of the edited images and their consistency with the target prompts, supplemented by human preference studies involving recruited volunteers from different disciplines.", + "bbox": [ + 88, + 484, + 482, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 664, + 302, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show some qualitative experimental results in Fig. 3-4, from our experiments, we observe the following:", + "bbox": [ + 89, + 688, + 482, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For single-concept fashion image editing, 1) several methods, such as LEDs++ [2], introduce unintended changes to non-target areas, possibly due to spurious correlation embedded in the pre-trained models from their training datasets. These correlations cause inaccurate concept associations, leading to changes outside the intended editing area. For instance, in Fig. 3, line 2, LEDs++ unexpectedly replaces the shoelaces with pink ones. Similarly, in Fig. 3, line 5, UltraEdit [41] transforms a squared neckline into a V-neckline, Turbo-edit [6] alters the dress length. 2) G & R [28], and UltraEdit often confuse different objects. For instance, when attempting to change the color of", + "bbox": [ + 88, + 719, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pants to red (Fig. 3, line 2), these methods inadvertently modify unrelated elements, such as the color of the shirt or hair. 3) TexFit's effectiveness is significantly hampered by its reliance on the accuracy of the recognized mask, which restricts its applicability in fashion image editing. As shown in the last column of Fig. 3, TexFit [32] either alters only localized information in the target editing area or excessively expands the intended modification area. This results in unnatural edits that disrupt the overall coherence of the image. For multi-concept fashion image editing, 1) UltraEdit [41], TexFit [32] often overlooks certain concepts in multiple-concept editing scenarios. For example, in Fig. 4, line 1, UltraEdit remains in top with short sleeves rather than long sleeves; in Fig. 4, line 2, the trousers edit is ignored. 2) G & R [28], TurboEdit [6], and LEDITS++ [2], demonstrate commendable performance in adhering to user instructions. However, the edited images can exhibit issues such as alterations to human poses or discontinuities in the edited results. For example, when modifying sleeve lengths, the transitions in the sleeve areas may appear inconsistent or fragmented (Fig. 4, line 1). These problems highlight the challenges in maintaining coherence and integrity in the edited images, despite following the provided directives closely. 3) Several methods such as Turbo-edit [6] and UltraEdit [41] may over-edit, making it difficult to retain the details of the original image. For example, when modifying pants into a skirt (Fig. 4, line 3), the edited", + "bbox": [ + 511, + 484, + 906, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28731", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image may change the length of the skirt as well as other aspects such as the color and fabric of the top along with the skirt. This phenomenon may be attributed to the spurious correlation that were learned by the pre-trained model from the training dataset. Compared to the existing methods, our method excels in fashion image editing, and it can directly control precise multi-concepts from text prompts without additional input, such as masks or poses. The edited image retains non-target areas, closely following the target text while maintaining the consistency of the image. In addition, our method allows continuous editing adjustments by adjusting the concept weight (Fig. 5), which provides fine control over the editing strength. This results in high-quality editing that is superior to the baseline model, especially for multi-concept fashion image editing.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 330, + 312, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We quantitatively evaluate our T-FIT against baselines using both automatic metrics and human evaluations.", + "bbox": [ + 89, + 357, + 482, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Automatic Metric Comparisons. We summarize the experimental findings in Tab. 1. Our method achieves state-of-the-art results for single- and multi-concept fashion image editing in CLIP-D, which demonstrates a significant advantage in terms of the accuracy and fidelity of our editing results in following the text description. In terms of the CLIP-I metric, our method demonstrates superior performance compared to most competing approaches. Although it slightly underperforms Turbo-edit [6] and TexFit [32] on CLIP-I, this is due to specific limitations in these methods during multi-concept fashion image editing. Turbo-edit often omits certain concepts, resulting in edited images that closely resemble the source, thus boosting its CLIP-I score. TexFit, on the other hand, struggles with precise edit masking, often restricting changes to small areas and failing in non-rigid edits (e.g., changing long sleeves to shorts or pants to shorts), leaving much of the image unaltered. Overall, our method shows superior performance on objective metrics for both multi- and single-concept editing tasks.", + "bbox": [ + 89, + 387, + 482, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Preference Study. In this section, we quantitatively evaluate our method with an extensive human perceptual evaluation study. We first collect a diverse set of 30 fashion images and their corresponding text descriptions, covering a range of clothing types (e.g., long sleeves, short sleeves, dresses, pants, shorts), patterns (e.g., solid, stripes, prints), and materials (e.g., cotton, denim, chiffon). Then, we invite 36 volunteers to participate in the evaluation. Each volunteer views the original fashion image, the original description text, the target description text, and the edited images generated by 6 different methods, including ours. The volunteers rate the edited images based on two criteria: (1) how well the edits align with the concepts specified in the target text and (2) the consistency of the non-edited areas with the original image. Ratings range from", + "bbox": [ + 89, + 674, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2308da61f16b418b1791d99afe167b9a4eda2df8223d459c781cccccc0608393.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons. The best results are bolded and the second-best results are underlined, respectively." + ], + "table_footnote": [], + "table_body": "
Single conceptMultiple concepts
MethodCLIP-DCLIP-ICLIP-DCLIP-I
OUR0.14330.91600.11520.9259
G & R0.09050.91160.05390.9079
Turbo-edit0.08210.91010.09670.9355
UltraEdit0.07710.90560.05330.8974
LEDITS++0.09540.91600.05840.9115
Texfit0.03460.95000.04630.9562
", + "bbox": [ + 514, + 128, + 908, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a5555857ecd9e4ba32152568dc680ae5b212dabc974743126e92e9b594aa417.jpg", + "image_caption": [ + "Figure 5. Human preference study. Our method outperforms baseline models in both text- and image alignment, demonstrating significant advantages in fashion image editing." + ], + "image_footnote": [], + "bbox": [ + 531, + 277, + 890, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1 to 8, with 1 representing the best performance and 8 the poorest. We calculate the average ranks provided by each participant, and the final results are displayed in Fig. 5. Our findings indicate that our method surpasses baseline methods on both criteria, demonstrating enhanced fidelity in executing intended edits while maintaining consistency across unaltered image areas.", + "bbox": [ + 511, + 565, + 905, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 681, + 666, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we turn to an ablation study where we analyze the effect of the different components inherent in our approach. Specifically, we consider: (1) the effect of substituting the CA with null-text inversion [18], (2) the effect of removing CLM, (3) the effects of varying annealing parameter $\\kappa$ , and (4) the influence of different weights on the editing results of concepts. The visual results are provided in Fig. 6-8.", + "bbox": [ + 511, + 704, + 905, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From these figures, it is evident that when T-FIT replaces the CA with the null-text inversion module, the edited images exhibit a noticeable spurious correlation. For instance, in Fig. 7, when the concept of \"gender\" is applied, the collar of the clothing incorrectly changes from a \"stand-up collar\" to a \"round collar\". On the other hand, T-FIT is equipped", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28732", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcc5e75fd19165d6cd944b3deff0ab9af37a5909ee63ec4d8ba0e1e22017cefd.jpg", + "image_caption": [ + "Figure 6. Qualitative ablation results. The last image shows the result of T-FIT, and the other images show the result of T-FIT without CLM and adjusting annealing parameter $\\kappa$ . Removing CLM significantly reduces the model's editability. Adjusting $\\kappa$ still makes balancing editability and fidelity challenging." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 475, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ede48eaa7f596cb4374c1a46ee991b2866b2fd2d0703a93755a5239dfe733318.jpg", + "image_caption": [ + "Figure 7. Qualitative ablation results. With CA, our model captures essential visual content from the original image, mitigating the impact of spurious correlation in pre-trained data. CLM ensures that no editing concepts are overlooked during modifications." + ], + "image_footnote": [], + "bbox": [ + 96, + 340, + 472, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "solely with CA, it effectively preserves the visual details of the source image and mitigates issues arising from spurious correlation. However, this setup offers limited flexibility for editing, even if we adopt varying annealing parameters $\\kappa$ (see Fig. 6). This suggests that while CA effectively addresses the spurious correlation issues arising from the pretrained model, it is limited in its editability. When CLM is added, T-FIT's editability improves. This is attributable to the CLM's capability to identify disentangled concepts within the representation space, which facilitates more flexible and precise editing. Furthermore, as illustrated in Fig. 8, T-FIT addresses the issues of concept omission and confusion through the multi-concept composite strategy, while also providing fine control over the intensity of each concept by adjusting individual concept weights. In our work, the concept weights can be manually adjusted in the range [1-5], with higher values indicating greater editing intensity on that concept and vice versa. This offers greater control and flexibility for fashion image editing of our method.", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e4059db8be4fc614a6afa59b0ea500bdd536a9348241d5e6e5067368f522d19.jpg", + "image_caption": [ + "Figure 8. Experimental comparison of varying concept weights (from 1 to 5) on single- and multi-concept fashion image editing. [C1], [C2], and [C3] mean different editing concepts, respectively." + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 898, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 547, + 633, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce T-FIT, a novel framework for fashion image editing. Our approach, which uses a single fashion image and simple text description, supports both single- and multi-concept image editing tasks. By incorporating counterfactual abduction and a concept learning module, T-FIT effectively mitigates issues of spurious correlation in the pre-trained model, enabling efficient, flexible, and accurate text-driven image editing. Furthermore, T-FIT allows fine-grained editing of complex fashion concepts by adjusting concept weights. Extensive experiments confirm the effectiveness of T-FIT and its components, demonstrating significant advantages over several existing methods. Our future work will focus on enhancing editing accuracy by incorporating visual information (e.g., style images) to learn richer visual concept representations.", + "bbox": [ + 511, + 573, + 906, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Acknowledgment.", + "text_level": 1, + "bbox": [ + 511, + 814, + 689, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by grants from the National Key R & D Program of China (grant no. 2022YFB3303302), and the National Natural Science Foundation of China (grant nos. 62377040, 62477004, 62207007, 623B2002).", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28733", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alberto Baldrati, Davide Morelli, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. Multimodal garment designer: Human-centric latent diffusion models for fashion image editing. In ICCV, 2023. 1", + "[2] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. LEDs++: Limitless image editing using text-to-image models. In CVPR, 2024. 1, 2, 5, 6", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 5, 6", + "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 1, 5", + "[5] Zhikang Chen, Min Zhang, Sen Cui, Haoxuan Li, Gang Niu, Mingming Gong, Changshui Zhang, and Kun Zhang. Neural collapse inspired feature alignment for out-of-distribution generalization. In NeurIPS, 2024. 1", + "[6] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv:2408.00735, 2024. 1, 5, 6, 7", + "[7] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. StyleGAN-NADA: CLIP-guided domain adaptation of image generators. ACM Transactions on Graphics, 41(4):1-13, 2022. 6", + "[8] Julia Guerrero-Viu, Milos Hasan, Arthur Roullier, Midhun Harikumar, Yiwei Hu, Paul Guerrero, Diego Gutierrez, Bellen Masia, and Valentin Deschaintre. Texsliders: Diffusion-based texture editing in clip space. In SIGGRAPH, 2024. 6", + "[9] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 6", + "[10] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Multiedits: Simultaneous multi-aspect editing with text-to-image diffusion models. In NeurIPS, 2024. 2", + "[11] Shanshan Huang, Haoxuan Li, Qingsong Li, Chunyuan Zheng, and Li Liu. Pareto invariant representation learning for multimedia recommendation. In ACM MM, 2023. 1", + "[12] Shanshan Huang, Qingsong Li, Jun Liao, Shu Wang, Li Liu, and Lian Li. Controllable image synthesis methods, applications and challenges: a comprehensive survey. Artificial Intelligence Review, 57(12):336, 2024. 1", + "[13] Shanshan Huang, Lei Wang, Jun Liao, and Li Liu. Multi-attentional causal intervention networks for medical image diagnosis. Knowledge-Based Systems, 299:111993, 2024. 1", + "[14] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 1, 5", + "[15] Gwanhyeong Koo, Sunjae Yoon, Ji Woo Hong, and Chang D. Yoo. Flexiedit: Frequency-aware latent refinement for enhanced non-rigid editing. In ECCV, 2024. 1" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Meng Li and Haochen Sui. Causal recommendation via machine unlearning with a few unbiased data. In AAAI Workshop on AICT, 2025. 1", + "[17] Paulius Micikevicius, Sharan Narang, Jonah Alben, Greg Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2018. 5", + "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 7", + "[19] Martin Pernus, Clinton Fookes, Vitomir Štruc, and Simon Dobrišek. FICE: Text-conditioned fashion-image editing with guided gan inversion. Pattern Recognition, 158:111022, 2025. 1", + "[20] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion Autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 3", + "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6", + "[22] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 2022. 1", + "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2", + "[24] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1", + "[25] Pedro Sanchez, Jeremy P Voisey, Tian Xia, Hannah I Watson, Alison Q O'Neil, and Sotirios A Tsaftaris. Causal machine learning for healthcare and precision medicine. Royal Society Open Science, 9(8):220638, 2022. 1", + "[26] Xue Song, Jiequan Cui, Hanwang Zhang, Jingjing Chen, Richang Hong, and Yu-Gang Jiang. Doubly abductive counterfactual inference for text-based image editing. In CVPR, 2024. 3, 5", + "[27] Adam Stein, Aaditya Naik, Yinjun Wu, Mayur Naik, and Eric Wong. Towards compositionality in concept learning. In ICML, 2024. 2", + "[28] Vadim Titov, Madina Khalmatova, Alexandra Ivanova, Dmitry Vetrov, and Aibek Alanov. Guide-and-Rescale: Self-guidance mechanism for effective tuning-free real image editing. In ECCV, 2024. 1, 2, 5, 6", + "[29] Linoy Tsaban and Apolinário Passos. LEDs: Real image editing with ddpm inversion and semantic guidance. arXiv:2307.00522, 2023. 1, 5", + "[30] Luozhou Wang, Shuai Yang, Shu Liu, and Ying-cong Chen. Not all steps are created equal: Selective diffusion distillation for image manipulation. In ICCV, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28734", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Lei Wang, Shanshan Huang, Shu Wang, Jun Liao, Tingpeng Li, and Li Liu. A survey of causal discovery based on functional causal model. Engineering Applications of Artificial Intelligence, 133:108258, 2024. 1", + "[32] Tongxin Wang and Mang Ye. TexFit: Text-driven fashion image editing with diffusion models. In AAAI, 2024. 1, 5, 6, 7", + "[33] Xiaolong Wang, Zhi-Qi Cheng, Jue Wang, and Xiaojiang Peng. DPDEdit: Detail-preserved diffusion models for multimodal fashion image editing. arXiv:2409.01086, 2024. 1", + "[34] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, Fei Wang, Christopher De Sa, and Volodymyr Kuleshov. Infodiffusion: Representation learning using information maximizing diffusion models. In ICML, 2023. 3", + "[35] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for (score-based) text-controlled generative models. NeurIPS, 2024. 1, 2", + "[36] Peng Wu, Haoxuan Li, Yuhao Deng, Wenjie Hu, Quanyu Dai, Zhenhua Dong, Jie Sun, Rui Zhang, and Xiao-Hua Zhou. On the opportunity of causal learning in recommendation systems: Foundation, estimation, prediction and challenges. In IJCAI, 2022. Survey Track. 1", + "[37] Xindi Wu, Dingli Yu, Yangsibo Huang, Olga Russakovsky, and Sanjeev Arora. Conceptmix: A compositional image generation benchmark with controllable difficulty. In NeurIPS, 2024. 1", + "[38] Yuntian Wu, Yuntian Yang, Jiabao Sean Xiao, Chuan Zhou, Haochen Sui, and Haoxuan Li. Invariant spatiotemporal representation learning for cross-patient seizure classification. In NeurIPS Workshop on NeuroAI, 2024. 1", + "[39] Min Zhang, Haoxuan Li, Fei Wu, and Kun Kuang. Metacoco: A new few-shot classification benchmark with spurious correlation. In ICLR, 2024. 1", + "[40] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In CVPR, 2023. 5", + "[41] Haozhe Zhao, Xiaojian (Shawn) Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. In NeurIPS, 2024. 1, 5, 6" + ], + "bbox": [ + 91, + 92, + 480, + 657 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28735", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_model.json b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0bb8472c747ce033275c32ecd886a5171ae127cf --- /dev/null +++ b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.112, + 0.895, + 0.157 + ], + "angle": 0, + "content": "Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.185, + 0.896, + 0.221 + ], + "angle": 0, + "content": "Shanshan Huang\\(^{1}\\), Haoxuan Li\\(^{2}\\), Chunyuan Zheng\\(^{3}\\), Mingyuan Ge\\(^{1}\\), Wei Gao\\(^{1}\\), Lei Wang\\(^{1}\\), Li Liu\\(^{1*}\\) \n\\(^{1}\\)School of Big Data & Software Engineering, Chongqing University" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.221, + 0.896, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Center for Data Science, Peking University \\(^{3}\\)School of Mathematical Sciences, Peking University \\(\\{\\text { shanshanhuang, dcsliuli } \\} @ \\text { cqu.edu.cn, } \\{\\text { hxli, cyzheng } \\} @ \\text {stu.pku.edu.cn, }\\) \\(\\{\\text { leiwangtt, mingyuange } \\} @ \\text {stu.cqu.edu.cn, }\\) wgao2002@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.294, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.327, + 0.485, + 0.751 + ], + "angle": 0, + "content": "Fashion image editing is a valuable tool for designers to convey their creative ideas by visualizing design concepts. With the recent advances in text editing methods, significant progress has been made in fashion image editing. However, they face two key challenges: spurious correlations in training data often induce changes in other areas when editing an area representing the intended editing concept, and these models typically lack the ability to edit multiple concepts simultaneously. To address the above challenges, we propose a novel Text-driven Fashion Image editing framework called T-FIT to mitigate the impact of spurious correlation by integrating counterfactual reasoning with compositional concept learning to precisely ensure compositional multi-concept fashion image editing relying solely on text descriptions. Specifically, T-FIT includes three key components. (i) Counterfactual abduction module, which learns an exogenous variable of the source image by a denoising U-Net model. (ii) Concept learning module, which identifies concepts in fashion image editing—such as clothing types and colors and projects a target concept into the space spanned from a series of textual prompts. (iii) Concept composition module, which enables simultaneous adjustments of multiple concepts by aggregating each concept's direction vector obtained from the concept learning module. Extensive experiments show that our method can achieve state-of-the-art performance on various fashion image editing tasks, including single-concept editing (e.g., sleeve length, clothing type) and multi-concept editing (e.g., color & sleeve length)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.774, + 0.222, + 0.79 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Fashion image editing has emerged as a pivotal field within computer vision, driven by the increasing need for personalized and dynamic visual modifications in the fashion industry [1, 33]. The diversity and high-quality image generation capabilities of large-scale text-image diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.906, + 0.401 + ], + "angle": 0, + "content": "such as Stable Diffusion [23], DALL-E 2 [24], and Imagegen [22] has inspired many subsequent efforts to leverage pre-trained large-scale models for text-driven fashion image editing (TFIE) [12, 19, 32]. A foundational aspect of this field is single-concept editing, where a specific concept [35, 37] like color, texture, or style of an item is altered while keeping the rest of the fashion image the same." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.749 + ], + "angle": 0, + "content": "However, the existing methods still suffer from several limitations. First, these methods [2, 6, 41] often encounter challenges related to spurious correlation and unwanted associations learned by the pre-trained model (e.g., Stable Diffusion [23]) due to biases in training data. For example, several methods [28, 29] may erroneously associate a specific clothing type (e.g., sleeve length) with a specific cloth style (e.g., garment length), as shown in Fig. 1, resulting in unintended modifications outside the target concept. This is because the long sleeves always co-occur with long garments, inducing spurious correlations in the training data. Therefore, eliminating spurious correlations is critical to ensuring precise, isolated modifications. There are many methods to address spurious correlation [5, 11, 13, 31, 36, 38, 39]. For example, Li et al. [16] propose a novel machine unlearning method to reduce the instance weight of biased samples in e-commerce and Sanchez et al. [25] discuss important challenges present in healthcare applications such as processing high-dimensional and unstructured data, as well as temporal relationships. However, due to differences in data structure and final goal, previous methods could not be adapted to image editing scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Second, existing methods [32, 33] are mainly limited to editing the texture of fashion images, and it is difficult to achieve non-rigid editing, such as precisely adjusting the length of sleeves or skirts. As shown in Fig. 1, UltraEdit [41] and Turbo-edit [6] cannot achieve non-rigid editing, e.g., turning a long skirt into a mini skirt; altering short sleeves into long sleeves, which may attribute to the poor understanding of the concept from the prompt." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Third, existing methods [4, 14, 15] typically focus on editing a single concept in the source image. While single-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28726" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.09, + 0.472, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.241, + 0.485, + 0.326 + ], + "angle": 0, + "content": "Figure 1. Two examples to illustrate the spurious correlation and conceptual omission issues in fashion image editing. Unlike other methods, our method alleviates the issue of over-editing that exists in fashion image editing due to spurious correlation (e.g., sleeve length and garment length, mini skirt and plaid style), as well as the confusion or lack of concepts when editing multiple concepts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.483, + 0.534 + ], + "angle": 0, + "content": "concept methods enable the modification of multiple concepts sequentially, i.e., changing the concepts one by one, they present certain drawbacks. Specifically, these methods either introduce concept omission due to the difficulty of identifying each concept from the prompt; or they lead to unexpected changes due to interactions between different concepts, which is still caused by spurious correlation [10]. For example, modifications made to later concepts may affect previous modifications. From Fig. 1, we can observe that G & R [28] incorrectly adds a \"leather bag\" when editing the skirt fabric to leather and LEDs++ [2] causes other unrelated attributes changes when changing the length and fabric of the skirt, such as the color of the gloves." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.673 + ], + "angle": 0, + "content": "To bridge this gap, we develop a Text-driven Fashion Image editing framework (T-FIT) that relies solely on textual prompts, enabling users to perform both single- and multi-concept editing on a single fashion image. Specifically, T-FIT adopts Stable Diffusion [23] and introduces a counterfactual abduction module to address the spurious correlation and concept learning module with concept composition to achieve precise single- and multi-concept editing as shown in Fig. 2. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.675, + 0.483, + 0.719 + ], + "angle": 0, + "content": "- We reveal the issue of spurious correlation and concept omission in fashion image editing and propose T-FIT, a text-driven framework to address these issues." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.781 + ], + "angle": 0, + "content": "- We propose counterfactual abduction (CA) to capture visual content and a concept learning module (CLM) to identify and disentangle concepts, mitigating issues of spurious correlation and non-rigid editing." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.841 + ], + "angle": 0, + "content": "- We propose a concept composition strategy that enables simultaneous, fine-grained editing across multiple concepts by adjusting the weight of each learned concept to address the concept omission problem." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Extensive empirical evaluation, in terms of both automated and human assessment metrics, qualitatively and quantitatively demonstrates that our T-FIT approach significantly enhances the quality of fashion image editing." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.675, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.65, + 0.108 + ], + "angle": 0, + "content": "2. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.112, + 0.719, + 0.128 + ], + "angle": 0, + "content": "2.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.136, + 0.907, + 0.273 + ], + "angle": 0, + "content": "We first formulate the image editing problem as follows: given a source fashion image \\( I_{S} \\), a corresponding text prompt \\( P \\) that describes the contents of \\( I_{S} \\), a set of concept prompts \\( P_{C^i} = \\{p_0^i, p_1^i, \\ldots, p_{N-1}^i\\} \\) used to construct a concept \\( C^i \\) (e.g., clothing type, color, fabric) that the user wishes to edit, and a target text prompt \\( P' \\) describing the final editing goal, our objective is to generate a new fashion image \\( I_T \\) by editing \\( I_S \\) in alignment with \\( P' \\) while preserving other areas that are unrelated to the concept \\( C^i \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.279, + 0.644, + 0.296 + ], + "angle": 0, + "content": "2.2. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.907, + 0.575 + ], + "angle": 0, + "content": "In the context of fashion image editing, concepts \\( C = \\{C^1,\\dots ,C^i,\\dots ,C^m\\} \\) are a distinct and human-interpretable feature or attribute within an image, encompassing elements like sleeve length, fabric type, color, and cloth type. Each concept \\( C^i \\) contains mutually selected attributes, where each attribute embodies certain shared properties. For instance, the attributes \"long sleeves\" and \"short sleeves\" collectively represent the concept of \"sleeve length\", while \"woman\" and \"man\" represent the concept of \"gender\". The concept representations \\( s(C) \\) can be derived from the intermediate representation of some text-to-image generative models (e.g., Stable Diffusion) pre-trained over annotated datasets that link visual attributes to textual descriptions [27, 35]. Here, \\( s:\\mathbb{C}\\to \\mathbb{R}^d \\) is the concept representation function, where \\( \\mathbb{C} \\) is the concept space containing all possible concepts, and \\( \\mathbb{R}^d \\) represents the representation space with dimension \\( d \\). Following [27, 35], we focus on the compositional concepts, as defined in Definition 1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.58, + 0.907, + 0.643 + ], + "angle": 0, + "content": "Definition 1 (Compositional Concepts). For concepts \\( C^i, C^j \\in \\mathbb{C} \\), the concept representation \\( s: \\mathbb{C} \\to \\mathbb{R}^d \\) is considered compositional if there exist positive weights \\( w_i, w_j \\in \\mathbb{R}^+ \\) such that:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.647, + 0.831, + 0.667 + ], + "angle": 0, + "content": "\\[\ns (C ^ {i} \\cup C ^ {j}) = w _ {i} s (C ^ {i}) + w _ {j} s (C ^ {j}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.797 + ], + "angle": 0, + "content": "This definition indicates that the representation of the composed concepts corresponds to the weighted sum of the individual concept representations in the representation space. Following this definition, this study aims to develop a framework for learning and composing conceptual representations from pre-trained text-to-image models (i.e., Stable Diffusion) to achieve flexible and targeted fashion image editing based on these compositional concepts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.895, + 0.818 + ], + "angle": 0, + "content": "2.3. Counterfactual Abduction for Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To address spurious correlation issues inherent in existing image editing methods, we introduce a counterfactual abduction approach that employs an abduction loss to infer unknown exogenous variables \\( U \\) for each image. These variables capture the additional visual content of source fashion" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28727" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.089, + 0.473, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.105, + 0.885, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.222, + 0.638, + 0.236 + ], + "angle": 0, + "content": "(c) Concept Composition" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.24, + 0.895, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.447, + 0.908, + 0.517 + ], + "angle": 0, + "content": "Figure 2. Overview of our T-FIT for fashion image editing. We divide the entire editing process into two stages. In the first stage, we perform counterfactual abduction to infer exogenous variables \\( U \\) that capture the core visual contents of the source fashion image. In the second stage, we employ a concept learning module to discover the target concept from representation space and design a concept composition strategy for combining multiple aspects of the fashion image, such as fabric, color, or pattern, while maintaining consistency with the original image." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.484, + 0.697 + ], + "angle": 0, + "content": "images, aiming to reduce uncertainty in fashion image editing and prevent unintended changes in the structure or identity of the source image. In other words, by incorporating \\( U \\), our method ensures that the edited fashion image \\( I_{T} \\) integrates the influence of \\( P' \\) while retaining other unrelated visual content from the source image \\( I_{S} \\) unchanged, thus addressing spurious correlation introduced in pre-trained models from the causal perspective. Specifically, we train the abduction \\( U \\) by optimizing the following Gaussian noise regression, a process analogous to training the reversed diffusion steps:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.704, + 0.483, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {U} \\mathbb {E} _ {(t, \\epsilon)} | | \\epsilon - \\epsilon_ {U} (x _ {t}, t, \\varepsilon_ {s r c}) | | _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,I)\\) is the Gaussian noise, \\(t\\in [0,T]\\) represents a sampled time step, \\(\\epsilon_{U}\\) represents a new denoising model trained on a pre-trained denoising U-Net, wherein \\(U\\) is treated as a trainable parameter. All other model parameters, aside from those of the denoising U-Net, remain fixed during training. The text embedding \\(\\varepsilon_{src} = \\psi (P)\\) is obtained by a pre-trained CLIP text encoder \\(\\psi (\\cdot)\\), and the noisy input at time \\(t\\) can be obtained by \\(x_{t} = \\sqrt{\\alpha_{t}} x_{0} + \\sqrt{1 - \\alpha_{t}}\\epsilon\\). Here, \\(x_0 = I_S\\) denotes the original image and \\(\\alpha_{t}\\) is a predefined variance schedule [20, 34]. Following [26], we parameterize \\(U\\) as the UNet LoRA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.683 + ], + "angle": 0, + "content": "in \\(\\epsilon_U\\), LoRA structure is incorporated across all attention layers, convolution layers, and feed-forward layers, which can be expressed by \\(W' = W + \\beta U_A \\cdot U_B\\). Here, \\(W \\in \\mathbb{R}^{d \\times d}\\) denotes the original weight matrix, \\(U_A \\in \\mathbb{R}^{d \\times r}\\) and \\(U_B \\in \\mathbb{R}^{r \\times d}\\) are low-rank matrices. Wang et al. [30] observed that the larger the time step in the diffusion model, the better the editable and the lower the fidelity. For this reason, the annealing parameter \\(\\beta = \\frac{1 - \\kappa}{T^2} (t - T)^2 + \\kappa\\) is proposed to improve the editability. In our experiment, the rank \\(r\\) is set to 512, \\(\\kappa \\in [0,1]\\) is a constant." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.895, + 0.712 + ], + "angle": 0, + "content": "2.4. Compositional Concept Learning for Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Unfortunately, the abduction of \\( U \\) is inherently ill-posed [26], often leading to overfitting to the specific \\( P \\) and \\( x \\). This overfitting reduces the editability of the generative model \\( G(\\cdot, U) \\) and may cause challenges like concept omission or confusion, where distinctions between concepts may blur, or certain concepts might overshadow others. Therefore, \\( G(\\cdot, U) \\) struggles to generate images that adhere to a new prompt \\( P' \\). To address this, we introduce a concept learning module that analyzes the representation space to identify composable concepts. Additionally, we propose a multi-concept composition strategy for flexible fashion image editing, enabling single and multi-concept continuous" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28728" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": " edits. For example, our method can simultaneously modify both the \"sleeve length\" (e.g., changing from long sleeves to short sleeves) and the \"color\" of the clothing (e.g., transitioning from red to blue). The intensity of these edits can also be adjusted, such as making the blue lighter or darker." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.488 + ], + "angle": 0, + "content": "Concept Learning Module. Given a target concept \\( C^i \\), we aim to learn an interpretable direction vector \\( \\Delta C_t^i \\) at time \\( t \\). The learned direction vector allows us to control the generation process by varying the latent \\( x_{t} \\) in the desired directions \\( \\Delta C_t^i \\) with weight \\( w_{i} \\), which is considered as \\( s_{new}^{i} = s_{src} + w_{i}\\Delta C_{t}^{i} \\). Here, \\( s_{src} = \\epsilon_U(x_t,t,\\varepsilon_{src}) \\), \\( \\varepsilon_{src} = \\psi (P) \\) is derived from pre-trained CLIP text encoder \\( \\psi (\\cdot) \\), and \\( \\epsilon_U(x_t,t,\\varepsilon_{src}) \\) is trained by counterfactual abduction and is fixed here. Specifically, we first define a set of text prompts \\( P_{C^i} = \\{p_0^i,p_1^i,\\dots,p_{N_i - 1}^i\\} \\) designed to elicit distinct distributions for a target concept \\( C^i \\) while maintaining a consistent distribution for another concept \\( C^j \\). Without loss of the generality, we define \\( p_0^i \\) as the source prompt without redundant concepts. In essence, each prompt captures a unique representation of \\( C^i \\) while keeping the representation of \\( C^j \\) consistent. For instance, when editing the concept of \"fabric,\" we can formulate a series of prompts like \"a lace [fabric 0] dress,\" \"a silk [fabric 1] dress,\" and \"a [fabric \\( N_{i} - 1 \\)] dress.\" These prompts are intended to generate variations in the fabric concept while keeping other aspects, like dress type." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Then, to identify the direction vector for concept \\( C^i \\), we first define its representation space as \\( \\mathcal{R}_{C_t^i} \\coloneqq \\text{span}(\\{v_t^j - v_t^0 : j = 1, \\dots, N_i - 1\\}) \\). By leveraging the span of conceptual space, we enable the model to edit not just predefined attributes, but also a broader range of related attributes. For example, by defining a structured representation space with \"blue,\" \"green,\" and \"yellow,\" our model can generalize to other colors in the spectrum, such as \"red\". Here, \\( v_t^j = \\epsilon_U(x_t, t, \\varepsilon_j) \\), each \\( \\varepsilon_j = \\psi(p_j^i) \\) is derived from the pre-trained CLIP text encoder \\( \\psi(\\cdot) \\). Next, to compute the projection matrix \\( M_i \\) for the \\( C^i \\)-space within this representation space, we start by constructing matrix \\( S^i \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.683, + 0.483, + 0.703 + ], + "angle": 0, + "content": "\\[\nS _ {t} ^ {i} := \\left[ v _ {t} ^ {1} - v _ {t} ^ {0}, \\dots , v _ {t} ^ {N _ {i} - 1} - v _ {t} ^ {0} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.483, + 0.884 + ], + "angle": 0, + "content": "In addition, to align changes in the representation space with the semantic structure of the concept space, we first apply singular value decomposition (SVD) on \\( S_{t}^{i} \\), i.e., \\( SVD(S_{t}^{i}) = V_{t}^{i}\\Sigma_{t}^{i}U_{t}^{iT} \\), to identify the principal orthogonal semantic directions. We then select the top \\( K \\) components of \\( V_{t}^{i} \\) to capture the most significant semantic variations (for simplicity, we denote this matrix as \\( V_{t}^{i} \\)), and construct the projection matrix \\( M_{t}^{i} = V_{t}^{i}V_{t}^{iT} \\), which projects the concept vector into \\( \\mathcal{R}_{C_t^i} \\) and ensures that the changes in the representation space are corrected and aligned with the key semantic directions." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Finally, to derive the editing direction, we calculate the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "score difference between source representation \\( s_{src} \\) and target representation \\( s_{tar} \\) as \\( \\Delta s_t = s_{tar} - s_{src} \\). We then project \\( \\Delta s_t \\) onto the spanned concept space using \\( M_t^i \\), yielding \\( \\Delta C_t^i = \\Delta s_t M_t^i \\). This projection ensures the editing operation reflects the relevant semantic changes in the concept space. The modified score can be achieved by" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.193, + 0.905, + 0.211 + ], + "angle": 0, + "content": "\\[\ns = s _ {\\emptyset} + \\lambda \\left(s _ {n e w} ^ {i} - s _ {\\emptyset}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.905, + 0.299 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is guidance scale, \\(s_{\\emptyset}\\) denotes the unconditional score, is given by \\(s_{\\emptyset} = \\epsilon_U(x_t,t,\\varepsilon_{\\emptyset})\\) with \\(\\varepsilon_{\\emptyset} = \\psi (^{\\prime \\prime})\\) In practice, we noticed that a few prompts are sufficient to identify the concept direction, which is attributed to the generalization ability of the span space \\(\\mathcal{R}_{C_t^i}\\)" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.313, + 0.825, + 0.328 + ], + "angle": 0, + "content": "Algorithm 1 Compositional Concept Learning" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.329, + 0.905, + 0.406 + ], + "angle": 0, + "content": "Require: Diffusion model \\(\\epsilon_U(x_t,t,\\varepsilon)\\), pre-trained CLIP text encoder \\(\\psi (\\cdot)\\), guidance scale \\(\\lambda\\), concept weight \\(\\{w_{1},\\dots,w_{m}\\}\\), covariance matrix \\(\\sigma_t^2 I\\), empty prompt \" \", source prompt \\(P\\), target prompt \\(P^{\\prime}\\), prompts for build concept space \\(P_{C^i} = \\{p_0^i,\\ldots ,p_j^i,\\ldots ,p_{N_i - 1}^i\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.406, + 0.747, + 0.42 + ], + "angle": 0, + "content": "1: Initialize sample \\( x_{t} \\sim \\mathcal{N}(0, I) \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.42, + 0.905, + 0.435 + ], + "angle": 0, + "content": "2: \\(\\varepsilon_{\\emptyset}, \\varepsilon_{src}, \\varepsilon_{tar} = \\psi(\"\"), \\psi(P), \\psi(P')\\) # Text embedding" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.435, + 0.797, + 0.452 + ], + "angle": 0, + "content": "3: \\(\\varepsilon_{j} = \\psi (p_{j}^{i})\\) # Concept text embedding" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.452, + 0.666, + 0.464 + ], + "angle": 0, + "content": "4: for \\( t = T, \\dots, 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.466, + 0.84, + 0.481 + ], + "angle": 0, + "content": "5: \\(s_{\\emptyset}\\gets \\epsilon_U(x_t,t,\\varepsilon_{\\emptyset})\\) # Unconditional score" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.481, + 0.905, + 0.509 + ], + "angle": 0, + "content": "6: \\(s_{src}, s_{tar} \\gets \\epsilon_U(x_t, t, \\varepsilon_{src}), \\epsilon_U(x_t, t, \\varepsilon_{tar})\\) # Conditional score" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.511, + 0.816, + 0.525 + ], + "angle": 0, + "content": "7: \\(\\Delta s_{t} = s_{tar} - s_{src}\\) # Score difference" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.526, + 0.8, + 0.541 + ], + "angle": 0, + "content": "8: for \\(i = 1,2,\\dots,m\\) do # m concepts" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.541, + 0.828, + 0.556 + ], + "angle": 0, + "content": "9: \\(S_{t}^{i}\\gets \\left[\\epsilon_{U}(x_{t},t,\\varepsilon_{1}) - \\epsilon_{U}(x_{t},t,\\varepsilon_{0}),\\right.\\)" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.557, + 0.81, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {2}\\right) - \\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {0}\\right), \\dots ,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.571, + 0.808, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\left. \\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {N _ {i} - 1}\\right) - \\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {0}\\right) \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.587, + 0.898, + 0.601 + ], + "angle": 0, + "content": "10: Compute the top- \\(k\\) left singular vectors \\(V_{i}\\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.601, + 0.749, + 0.619 + ], + "angle": 0, + "content": "\\[\nS V D \\left(S _ {t} ^ {i}\\right) = V _ {t} ^ {i} \\Sigma_ {t} ^ {i} U _ {t} ^ {i T}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.619, + 0.829, + 0.634 + ], + "angle": 0, + "content": "11: \\(M_{t}^{i} = V_{t}^{i}(V_{t}^{i})^{T}\\) # Projection matrix" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.634, + 0.825, + 0.649 + ], + "angle": 0, + "content": "12: \\(\\Delta C_{t}^{i} = M_{t}^{i}\\Delta s_{t}\\#\\) Editing direction" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.649, + 0.621, + 0.661 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.664, + 0.905, + 0.694 + ], + "angle": 0, + "content": "14: \\(s \\gets s_{\\emptyset} + \\lambda((s_{src} + w_i \\Delta C_t^i) - s_{\\emptyset})\\) # For single-concept editing" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.694, + 0.905, + 0.723 + ], + "angle": 0, + "content": "15: \\(s \\gets s_{\\emptyset} + \\lambda \\sum_{i=1}^{m} \\left( (s_{src} + w_i \\Delta C_t^i) - s_{\\emptyset} \\right) \\# \\text{For multi-concept editing}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.723, + 0.727, + 0.739 + ], + "angle": 0, + "content": "16: \\(x_{t - 1}\\sim \\mathcal{N}(x_t - s,\\sigma_t^2 I)\\)" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.74, + 0.596, + 0.753 + ], + "angle": 0, + "content": "17: end for" + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.406, + 0.905, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Concept Composition Strategy. Now, our method enables single-concept editing for fashion images, but many scenarios demand multi-concept compositional editing. A straightforward way to achieve this is to create a multi-concept space using a set of concept prompts. While feasible, this approach limits fine-grained control over individual concepts, such as adjusting shades of color in specific items. Alternatively, sequential single-concept edits can be applied, but this accumulates errors and is time-consuming." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28729" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.09, + 0.895, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.475, + 0.761, + 0.49 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison with the state-of-art methods in single concept editing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.565 + ], + "angle": 0, + "content": "To overcome these issues, we propose a controllable multi-concept editing strategy. This strategy enables precise, simultaneous adjustments of multiple concepts by weighting each concept's direction vector, which is shown as" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.57, + 0.483, + 0.609 + ], + "angle": 0, + "content": "\\[\ns _ {m} = s _ {\\emptyset} + \\lambda \\sum_ {i = 1} ^ {m} \\left(s _ {n e w} ^ {i} - s _ {\\emptyset}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.661 + ], + "angle": 0, + "content": "where \\( s_{new}^{i} = s_{src} + w_{i}\\Delta C_{t}^{i} \\), we can control the manipulation strength of each concept. The whole algorithm is shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.224, + 0.689 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.288, + 0.714 + ], + "angle": 0, + "content": "3.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. In our work, all experiments are conducted on a single NVIDIA RTX 4090 GPU with 24 GB of memory. Following the prior work [3, 4, 26], we employ the official pre-trained Stable Diffusion v2.1-base model as our foundational model, downsampling all images to a resolution of \\(512 \\times 512\\) pixels for consistency across experiments. For the counterfactual abduction phase, We fine-tune the model on the single image using LoRA with a rank of 512 for 100 epochs, employing the Adam optimizer with a learning rate of 1e-4. To conserve memory, we adopt the mixed precision [17] and gradient accumulation strategies, setting both the accumulation step count and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.579 + ], + "angle": 0, + "content": "batch size to 1. During the inference phase, we utilize the DDIM sampling with 80 steps and a classifier-free guidance scale of \\(\\lambda = 4\\). Additionally, we explore various values for the concept weight \\(w_{i}\\) within the range [1, 5] to control the editing strength, where \\(i\\) represents the \\(i\\)-th edited concept." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.775 + ], + "angle": 0, + "content": "Dataset. To comprehensively evaluate the performance of our T-FIT in fashion image editing, we follow [14, 26, 40] by collecting a diverse set of fashion images from Unsplash1, including both full-body and half-body images. The selected images comprise various clothing types, such as long dresses, skirts, T-shirts, shorts, pants, and strapless tops. Our evaluation encompasses a diverse set of editing concepts, which include specific attributes (e.g., gender, color, patterns, fabric texture) as well as broader elements such as clothing type, garment length, and sleeve style. Each image in the experiment is annotated with \\( N_{i} + 2 \\) prompts, consisting of one prompt for the original image and \\( N_{i} + 1 \\) prompts for target images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.776, + 0.907, + 0.835 + ], + "angle": 0, + "content": "Baselines. We conduct comparisons with existing representative diffusion-based image editing methods, including LEDs [29], LEDs++ [2], TexFit [32], Guide-and-Rescale (G & R) [28], UltraEdit [41], and Turboedit [6]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.836, + 0.907, + 0.865 + ], + "angle": 0, + "content": "Evaluation Metrics. To comprehensively evaluate the performance of our T-FIT, we conduct assessments from both" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.706, + 0.9 + ], + "angle": 0, + "content": "1https://unsplash.com/" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28730" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.09, + 0.895, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.457, + 0.768, + 0.473 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with the state-of-art methods in multiple concept editing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.486, + 0.483, + 0.653 + ], + "angle": 0, + "content": "objective and subjective perspectives. Objectively, we follow the previous works [3, 8, 9] and use metrics in CLIP embedding space: CLIP image similarity (CLIP-I) [21] to measure identity preservation via calculate the cosine similarity between the edited image and the original image, and CLIP text-image direction similarity (CLIP-D) [7] to evaluate the correspondence between image changes and text changes. Subjectively, we qualitatively evaluate the fidelity of the edited images and their consistency with the target prompts, supplemented by human preference studies involving recruited volunteers from different disciplines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.303, + 0.681 + ], + "angle": 0, + "content": "3.2. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.719 + ], + "angle": 0, + "content": "We show some qualitative experimental results in Fig. 3-4, from our experiments, we observe the following:" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For single-concept fashion image editing, 1) several methods, such as LEDs++ [2], introduce unintended changes to non-target areas, possibly due to spurious correlation embedded in the pre-trained models from their training datasets. These correlations cause inaccurate concept associations, leading to changes outside the intended editing area. For instance, in Fig. 3, line 2, LEDs++ unexpectedly replaces the shoelaces with pink ones. Similarly, in Fig. 3, line 5, UltraEdit [41] transforms a squared neckline into a V-neckline, Turbo-edit [6] alters the dress length. 2) G & R [28], and UltraEdit often confuse different objects. For instance, when attempting to change the color of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.908, + 0.895 + ], + "angle": 0, + "content": "pants to red (Fig. 3, line 2), these methods inadvertently modify unrelated elements, such as the color of the shirt or hair. 3) TexFit's effectiveness is significantly hampered by its reliance on the accuracy of the recognized mask, which restricts its applicability in fashion image editing. As shown in the last column of Fig. 3, TexFit [32] either alters only localized information in the target editing area or excessively expands the intended modification area. This results in unnatural edits that disrupt the overall coherence of the image. For multi-concept fashion image editing, 1) UltraEdit [41], TexFit [32] often overlooks certain concepts in multiple-concept editing scenarios. For example, in Fig. 4, line 1, UltraEdit remains in top with short sleeves rather than long sleeves; in Fig. 4, line 2, the trousers edit is ignored. 2) G & R [28], TurboEdit [6], and LEDITS++ [2], demonstrate commendable performance in adhering to user instructions. However, the edited images can exhibit issues such as alterations to human poses or discontinuities in the edited results. For example, when modifying sleeve lengths, the transitions in the sleeve areas may appear inconsistent or fragmented (Fig. 4, line 1). These problems highlight the challenges in maintaining coherence and integrity in the edited images, despite following the provided directives closely. 3) Several methods such as Turbo-edit [6] and UltraEdit [41] may over-edit, making it difficult to retain the details of the original image. For example, when modifying pants into a skirt (Fig. 4, line 3), the edited" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28731" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "image may change the length of the skirt as well as other aspects such as the color and fabric of the top along with the skirt. This phenomenon may be attributed to the spurious correlation that were learned by the pre-trained model from the training dataset. Compared to the existing methods, our method excels in fashion image editing, and it can directly control precise multi-concepts from text prompts without additional input, such as masks or poses. The edited image retains non-target areas, closely following the target text while maintaining the consistency of the image. In addition, our method allows continuous editing adjustments by adjusting the concept weight (Fig. 5), which provides fine control over the editing strength. This results in high-quality editing that is superior to the baseline model, especially for multi-concept fashion image editing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.313, + 0.348 + ], + "angle": 0, + "content": "3.3.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.483, + 0.388 + ], + "angle": 0, + "content": "We quantitatively evaluate our T-FIT against baselines using both automatic metrics and human evaluations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Automatic Metric Comparisons. We summarize the experimental findings in Tab. 1. Our method achieves state-of-the-art results for single- and multi-concept fashion image editing in CLIP-D, which demonstrates a significant advantage in terms of the accuracy and fidelity of our editing results in following the text description. In terms of the CLIP-I metric, our method demonstrates superior performance compared to most competing approaches. Although it slightly underperforms Turbo-edit [6] and TexFit [32] on CLIP-I, this is due to specific limitations in these methods during multi-concept fashion image editing. Turbo-edit often omits certain concepts, resulting in edited images that closely resemble the source, thus boosting its CLIP-I score. TexFit, on the other hand, struggles with precise edit masking, often restricting changes to small areas and failing in non-rigid edits (e.g., changing long sleeves to shorts or pants to shorts), leaving much of the image unaltered. Overall, our method shows superior performance on objective metrics for both multi- and single-concept editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Human Preference Study. In this section, we quantitatively evaluate our method with an extensive human perceptual evaluation study. We first collect a diverse set of 30 fashion images and their corresponding text descriptions, covering a range of clothing types (e.g., long sleeves, short sleeves, dresses, pants, shorts), patterns (e.g., solid, stripes, prints), and materials (e.g., cotton, denim, chiffon). Then, we invite 36 volunteers to participate in the evaluation. Each volunteer views the original fashion image, the original description text, the target description text, and the edited images generated by 6 different methods, including ours. The volunteers rate the edited images based on two criteria: (1) how well the edits align with the concepts specified in the target text and (2) the consistency of the non-edited areas with the original image. Ratings range from" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons. The best results are bolded and the second-best results are underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.13, + 0.909, + 0.273 + ], + "angle": 0, + "content": "
Single conceptMultiple concepts
MethodCLIP-DCLIP-ICLIP-DCLIP-I
OUR0.14330.91600.11520.9259
G & R0.09050.91160.05390.9079
Turbo-edit0.08210.91010.09670.9355
UltraEdit0.07710.90560.05330.8974
LEDITS++0.09540.91600.05840.9115
Texfit0.03460.95000.04630.9562
" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.279, + 0.891, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.516, + 0.907, + 0.559 + ], + "angle": 0, + "content": "Figure 5. Human preference study. Our method outperforms baseline models in both text- and image alignment, demonstrating significant advantages in fashion image editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.673 + ], + "angle": 0, + "content": "1 to 8, with 1 representing the best performance and 8 the poorest. We calculate the average ranks provided by each participant, and the final results are displayed in Fig. 5. Our findings indicate that our method surpasses baseline methods on both criteria, demonstrating enhanced fidelity in executing intended edits while maintaining consistency across unaltered image areas." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.667, + 0.699 + ], + "angle": 0, + "content": "3.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Next, we turn to an ablation study where we analyze the effect of the different components inherent in our approach. Specifically, we consider: (1) the effect of substituting the CA with null-text inversion [18], (2) the effect of removing CLM, (3) the effects of varying annealing parameter \\(\\kappa\\), and (4) the influence of different weights on the editing results of concepts. The visual results are provided in Fig. 6-8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "From these figures, it is evident that when T-FIT replaces the CA with the null-text inversion module, the edited images exhibit a noticeable spurious correlation. For instance, in Fig. 7, when the concept of \"gender\" is applied, the collar of the clothing incorrectly changes from a \"stand-up collar\" to a \"round collar\". On the other hand, T-FIT is equipped" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28732" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.476, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.265, + 0.483, + 0.336 + ], + "angle": 0, + "content": "Figure 6. Qualitative ablation results. The last image shows the result of T-FIT, and the other images show the result of T-FIT without CLM and adjusting annealing parameter \\(\\kappa\\). Removing CLM significantly reduces the model's editability. Adjusting \\(\\kappa\\) still makes balancing editability and fidelity challenging." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.342, + 0.473, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.542, + 0.483, + 0.598 + ], + "angle": 0, + "content": "Figure 7. Qualitative ablation results. With CA, our model captures essential visual content from the original image, mitigating the impact of spurious correlation in pre-trained data. CLM ensures that no editing concepts are overlooked during modifications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "solely with CA, it effectively preserves the visual details of the source image and mitigates issues arising from spurious correlation. However, this setup offers limited flexibility for editing, even if we adopt varying annealing parameters \\(\\kappa\\) (see Fig. 6). This suggests that while CA effectively addresses the spurious correlation issues arising from the pretrained model, it is limited in its editability. When CLM is added, T-FIT's editability improves. This is attributable to the CLM's capability to identify disentangled concepts within the representation space, which facilitates more flexible and precise editing. Furthermore, as illustrated in Fig. 8, T-FIT addresses the issues of concept omission and confusion through the multi-concept composite strategy, while also providing fine control over the intensity of each concept by adjusting individual concept weights. In our work, the concept weights can be manually adjusted in the range [1-5], with higher values indicating greater editing intensity on that concept and vice versa. This offers greater control and flexibility for fashion image editing of our method." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.9, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.493, + 0.907, + 0.536 + ], + "angle": 0, + "content": "Figure 8. Experimental comparison of varying concept weights (from 1 to 5) on single- and multi-concept fashion image editing. [C1], [C2], and [C3] mean different editing concepts, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.634, + 0.563 + ], + "angle": 0, + "content": "4. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.801 + ], + "angle": 0, + "content": "In this paper, we introduce T-FIT, a novel framework for fashion image editing. Our approach, which uses a single fashion image and simple text description, supports both single- and multi-concept image editing tasks. By incorporating counterfactual abduction and a concept learning module, T-FIT effectively mitigates issues of spurious correlation in the pre-trained model, enabling efficient, flexible, and accurate text-driven image editing. Furthermore, T-FIT allows fine-grained editing of complex fashion concepts by adjusting concept weights. Extensive experiments confirm the effectiveness of T-FIT and its components, demonstrating significant advantages over several existing methods. Our future work will focus on enhancing editing accuracy by incorporating visual information (e.g., style images) to learn richer visual concept representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.691, + 0.832 + ], + "angle": 0, + "content": "5. Acknowledgment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work was supported by grants from the National Key R & D Program of China (grant no. 2022YFB3303302), and the National Natural Science Foundation of China (grant nos. 62377040, 62477004, 62207007, 623B2002)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28733" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Alberto Baldrati, Davide Morelli, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. Multimodal garment designer: Human-centric latent diffusion models for fashion image editing. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. LEDs++: Limitless image editing using text-to-image models. In CVPR, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.268 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[5] Zhikang Chen, Min Zhang, Sen Cui, Haoxuan Li, Gang Niu, Mingming Gong, Changshui Zhang, and Kun Zhang. Neural collapse inspired feature alignment for out-of-distribution generalization. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.435 + ], + "angle": 0, + "content": "[6] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv:2408.00735, 2024. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[7] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. StyleGAN-NADA: CLIP-guided domain adaptation of image generators. ACM Transactions on Graphics, 41(4):1-13, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.562 + ], + "angle": 0, + "content": "[8] Julia Guerrero-Viu, Milos Hasan, Arthur Roullier, Midhun Harikumar, Yiwei Hu, Paul Guerrero, Diego Gutierrez, Bellen Masia, and Valentin Deschaintre. Texsliders: Diffusion-based texture editing in clip space. In SIGGRAPH, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[9] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[10] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Multiedits: Simultaneous multi-aspect editing with text-to-image diffusion models. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[11] Shanshan Huang, Haoxuan Li, Qingsong Li, Chunyuan Zheng, and Li Liu. Pareto invariant representation learning for multimedia recommendation. In ACM MM, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[12] Shanshan Huang, Qingsong Li, Jun Liao, Shu Wang, Li Liu, and Lian Li. Controllable image synthesis methods, applications and challenges: a comprehensive survey. Artificial Intelligence Review, 57(12):336, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[13] Shanshan Huang, Lei Wang, Jun Liao, and Li Liu. Multi-attentional causal intervention networks for medical image diagnosis. Knowledge-Based Systems, 299:111993, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[14] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Gwanhyeong Koo, Sunjae Yoon, Ji Woo Hong, and Chang D. Yoo. Flexiedit: Frequency-aware latent refinement for enhanced non-rigid editing. In ECCV, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[16] Meng Li and Haochen Sui. Causal recommendation via machine unlearning with a few unbiased data. In AAAI Workshop on AICT, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[17] Paulius Micikevicius, Sharan Narang, Jonah Alben, Greg Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.195, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.905, + 0.292 + ], + "angle": 0, + "content": "[19] Martin Pernus, Clinton Fookes, Vitomir Štruc, and Simon Dobrišek. FICE: Text-conditioned fashion-image editing with guided gan inversion. Pattern Recognition, 158:111022, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.297, + 0.905, + 0.35 + ], + "angle": 0, + "content": "[20] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion Autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.905, + 0.435 + ], + "angle": 0, + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.44, + 0.905, + 0.48 + ], + "angle": 0, + "content": "[22] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[24] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.599, + 0.905, + 0.653 + ], + "angle": 0, + "content": "[25] Pedro Sanchez, Jeremy P Voisey, Tian Xia, Hannah I Watson, Alison Q O'Neil, and Sotirios A Tsaftaris. Causal machine learning for healthcare and precision medicine. Royal Society Open Science, 9(8):220638, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.71 + ], + "angle": 0, + "content": "[26] Xue Song, Jiequan Cui, Hanwang Zhang, Jingjing Chen, Richang Hong, and Yu-Gang Jiang. Doubly abductive counterfactual inference for text-based image editing. In CVPR, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.905, + 0.754 + ], + "angle": 0, + "content": "[27] Adam Stein, Aaditya Naik, Yinjun Wu, Mayur Naik, and Eric Wong. Towards compositionality in concept learning. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[28] Vadim Titov, Madina Khalmatova, Alexandra Ivanova, Dmitry Vetrov, and Aibek Alanov. Guide-and-Rescale: Self-guidance mechanism for effective tuning-free real image editing. In ECCV, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[29] Linoy Tsaban and Apolinário Passos. LEDs: Real image editing with ddpm inversion and semantic guidance. arXiv:2307.00522, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Luozhou Wang, Shuai Yang, Shu Liu, and Ying-cong Chen. Not all steps are created equal: Selective diffusion distillation for image manipulation. In ICCV, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28734" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Lei Wang, Shanshan Huang, Shu Wang, Jun Liao, Tingpeng Li, and Li Liu. A survey of causal discovery based on functional causal model. Engineering Applications of Artificial Intelligence, 133:108258, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[32] Tongxin Wang and Mang Ye. TexFit: Text-driven fashion image editing with diffusion models. In AAAI, 2024. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[33] Xiaolong Wang, Zhi-Qi Cheng, Jue Wang, and Xiaojiang Peng. DPDEdit: Detail-preserved diffusion models for multimodal fashion image editing. arXiv:2409.01086, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[34] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, Fei Wang, Christopher De Sa, and Volodymyr Kuleshov. Infodiffusion: Representation learning using information maximizing diffusion models. In ICML, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[35] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for (score-based) text-controlled generative models. NeurIPS, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[36] Peng Wu, Haoxuan Li, Yuhao Deng, Wenjie Hu, Quanyu Dai, Zhenhua Dong, Jie Sun, Rui Zhang, and Xiao-Hua Zhou. On the opportunity of causal learning in recommendation systems: Foundation, estimation, prediction and challenges. In IJCAI, 2022. Survey Track. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[37] Xindi Wu, Dingli Yu, Yangsibo Huang, Olga Russakovsky, and Sanjeev Arora. Conceptmix: A compositional image generation benchmark with controllable difficulty. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[38] Yuntian Wu, Yuntian Yang, Jiabao Sean Xiao, Chuan Zhou, Haochen Sui, and Haoxuan Li. Invariant spatiotemporal representation learning for cross-patient seizure classification. In NeurIPS Workshop on NeuroAI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[39] Min Zhang, Haoxuan Li, Fei Wu, and Kun Kuang. Metacoco: A new few-shot classification benchmark with spurious correlation. In ICLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[40] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[41] Haozhe Zhao, Xiaojian (Shawn) Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. In NeurIPS, 2024. 1, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28735" + } + ] +] \ No newline at end of file diff --git a/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_origin.pdf b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..13ddd4c31efc85a132288a458a9cf55905afc6b6 --- /dev/null +++ b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/0570630b-8b25-4f74-887f-1f4fc2b92f31_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d914c9896442e0b8cc683e13060c8b595a6a6a93be646c682e12db838869b425 +size 3206410 diff --git a/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/full.md b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dc62bceb0d052e02dc12788fc181d3563d4b0add --- /dev/null +++ b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/full.md @@ -0,0 +1,263 @@ +# Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction + +Shanshan Huang $^{1}$ , Haoxuan Li $^{2}$ , Chunyuan Zheng $^{3}$ , Mingyuan Ge $^{1}$ , Wei Gao $^{1}$ , Lei Wang $^{1}$ , Li Liu $^{1*}$ $^{1}$ School of Big Data & Software Engineering, Chongqing University + +$^{2}$ Center for Data Science, Peking University $^{3}$ School of Mathematical Sciences, Peking University $\{\text { shanshanhuang, dcsliuli } \} @ \text { cqu.edu.cn, } \{\text { hxli, cyzheng } \} @ \text {stu.pku.edu.cn, }$ $\{\text { leiwangtt, mingyuange } \} @ \text {stu.cqu.edu.cn, }$ wgao2002@gmail.com + +# Abstract + +Fashion image editing is a valuable tool for designers to convey their creative ideas by visualizing design concepts. With the recent advances in text editing methods, significant progress has been made in fashion image editing. However, they face two key challenges: spurious correlations in training data often induce changes in other areas when editing an area representing the intended editing concept, and these models typically lack the ability to edit multiple concepts simultaneously. To address the above challenges, we propose a novel Text-driven Fashion Image editing framework called T-FIT to mitigate the impact of spurious correlation by integrating counterfactual reasoning with compositional concept learning to precisely ensure compositional multi-concept fashion image editing relying solely on text descriptions. Specifically, T-FIT includes three key components. (i) Counterfactual abduction module, which learns an exogenous variable of the source image by a denoising U-Net model. (ii) Concept learning module, which identifies concepts in fashion image editing—such as clothing types and colors and projects a target concept into the space spanned from a series of textual prompts. (iii) Concept composition module, which enables simultaneous adjustments of multiple concepts by aggregating each concept's direction vector obtained from the concept learning module. Extensive experiments show that our method can achieve state-of-the-art performance on various fashion image editing tasks, including single-concept editing (e.g., sleeve length, clothing type) and multi-concept editing (e.g., color & sleeve length). + +# 1. Introduction + +Fashion image editing has emerged as a pivotal field within computer vision, driven by the increasing need for personalized and dynamic visual modifications in the fashion industry [1, 33]. The diversity and high-quality image generation capabilities of large-scale text-image diffusion models + +such as Stable Diffusion [23], DALL-E 2 [24], and Imagegen [22] has inspired many subsequent efforts to leverage pre-trained large-scale models for text-driven fashion image editing (TFIE) [12, 19, 32]. A foundational aspect of this field is single-concept editing, where a specific concept [35, 37] like color, texture, or style of an item is altered while keeping the rest of the fashion image the same. + +However, the existing methods still suffer from several limitations. First, these methods [2, 6, 41] often encounter challenges related to spurious correlation and unwanted associations learned by the pre-trained model (e.g., Stable Diffusion [23]) due to biases in training data. For example, several methods [28, 29] may erroneously associate a specific clothing type (e.g., sleeve length) with a specific cloth style (e.g., garment length), as shown in Fig. 1, resulting in unintended modifications outside the target concept. This is because the long sleeves always co-occur with long garments, inducing spurious correlations in the training data. Therefore, eliminating spurious correlations is critical to ensuring precise, isolated modifications. There are many methods to address spurious correlation [5, 11, 13, 31, 36, 38, 39]. For example, Li et al. [16] propose a novel machine unlearning method to reduce the instance weight of biased samples in e-commerce and Sanchez et al. [25] discuss important challenges present in healthcare applications such as processing high-dimensional and unstructured data, as well as temporal relationships. However, due to differences in data structure and final goal, previous methods could not be adapted to image editing scenarios. + +Second, existing methods [32, 33] are mainly limited to editing the texture of fashion images, and it is difficult to achieve non-rigid editing, such as precisely adjusting the length of sleeves or skirts. As shown in Fig. 1, UltraEdit [41] and Turbo-edit [6] cannot achieve non-rigid editing, e.g., turning a long skirt into a mini skirt; altering short sleeves into long sleeves, which may attribute to the poor understanding of the concept from the prompt. + +Third, existing methods [4, 14, 15] typically focus on editing a single concept in the source image. While single- + +![](images/b96dd172db952754bc9d5c9cd4e11fdeba3d06d7f1e038f1091050054ac3794a.jpg) +Figure 1. Two examples to illustrate the spurious correlation and conceptual omission issues in fashion image editing. Unlike other methods, our method alleviates the issue of over-editing that exists in fashion image editing due to spurious correlation (e.g., sleeve length and garment length, mini skirt and plaid style), as well as the confusion or lack of concepts when editing multiple concepts. + +concept methods enable the modification of multiple concepts sequentially, i.e., changing the concepts one by one, they present certain drawbacks. Specifically, these methods either introduce concept omission due to the difficulty of identifying each concept from the prompt; or they lead to unexpected changes due to interactions between different concepts, which is still caused by spurious correlation [10]. For example, modifications made to later concepts may affect previous modifications. From Fig. 1, we can observe that G & R [28] incorrectly adds a "leather bag" when editing the skirt fabric to leather and LEDs++ [2] causes other unrelated attributes changes when changing the length and fabric of the skirt, such as the color of the gloves. + +To bridge this gap, we develop a Text-driven Fashion Image editing framework (T-FIT) that relies solely on textual prompts, enabling users to perform both single- and multi-concept editing on a single fashion image. Specifically, T-FIT adopts Stable Diffusion [23] and introduces a counterfactual abduction module to address the spurious correlation and concept learning module with concept composition to achieve precise single- and multi-concept editing as shown in Fig. 2. Our contributions are as follows: + +- We reveal the issue of spurious correlation and concept omission in fashion image editing and propose T-FIT, a text-driven framework to address these issues. +- We propose counterfactual abduction (CA) to capture visual content and a concept learning module (CLM) to identify and disentangle concepts, mitigating issues of spurious correlation and non-rigid editing. +- We propose a concept composition strategy that enables simultaneous, fine-grained editing across multiple concepts by adjusting the weight of each learned concept to address the concept omission problem. +- Extensive empirical evaluation, in terms of both automated and human assessment metrics, qualitatively and quantitatively demonstrates that our T-FIT approach significantly enhances the quality of fashion image editing. + +# 2. Methodology + +# 2.1. Problem Formulation + +We first formulate the image editing problem as follows: given a source fashion image $I_{S}$ , a corresponding text prompt $P$ that describes the contents of $I_{S}$ , a set of concept prompts $P_{C^i} = \{p_0^i, p_1^i, \ldots, p_{N-1}^i\}$ used to construct a concept $C^i$ (e.g., clothing type, color, fabric) that the user wishes to edit, and a target text prompt $P'$ describing the final editing goal, our objective is to generate a new fashion image $I_T$ by editing $I_S$ in alignment with $P'$ while preserving other areas that are unrelated to the concept $C^i$ . + +# 2.2. Preliminary + +In the context of fashion image editing, concepts $C = \{C^1,\dots ,C^i,\dots ,C^m\}$ are a distinct and human-interpretable feature or attribute within an image, encompassing elements like sleeve length, fabric type, color, and cloth type. Each concept $C^i$ contains mutually selected attributes, where each attribute embodies certain shared properties. For instance, the attributes "long sleeves" and "short sleeves" collectively represent the concept of "sleeve length", while "woman" and "man" represent the concept of "gender". The concept representations $s(C)$ can be derived from the intermediate representation of some text-to-image generative models (e.g., Stable Diffusion) pre-trained over annotated datasets that link visual attributes to textual descriptions [27, 35]. Here, $s:\mathbb{C}\to \mathbb{R}^d$ is the concept representation function, where $\mathbb{C}$ is the concept space containing all possible concepts, and $\mathbb{R}^d$ represents the representation space with dimension $d$ . Following [27, 35], we focus on the compositional concepts, as defined in Definition 1. + +Definition 1 (Compositional Concepts). For concepts $C^i, C^j \in \mathbb{C}$ , the concept representation $s: \mathbb{C} \to \mathbb{R}^d$ is considered compositional if there exist positive weights $w_i, w_j \in \mathbb{R}^+$ such that: + +$$ +s (C ^ {i} \cup C ^ {j}) = w _ {i} s (C ^ {i}) + w _ {j} s (C ^ {j}). +$$ + +This definition indicates that the representation of the composed concepts corresponds to the weighted sum of the individual concept representations in the representation space. Following this definition, this study aims to develop a framework for learning and composing conceptual representations from pre-trained text-to-image models (i.e., Stable Diffusion) to achieve flexible and targeted fashion image editing based on these compositional concepts. + +# 2.3. Counterfactual Abduction for Image Editing + +To address spurious correlation issues inherent in existing image editing methods, we introduce a counterfactual abduction approach that employs an abduction loss to infer unknown exogenous variables $U$ for each image. These variables capture the additional visual content of source fashion + +![](images/2555b61cabbaea1513b210b88ac1c07b76452ba585d97de9a1b00128e71766e5.jpg) +(c) Concept Composition + +![](images/51745256b6111e78b3730bc15ac80724ee41a28059635bfd2d0cdbcca7209aed.jpg) + +![](images/33f9009859475a5df608025691b2a57bbfb661822990095a6a0ce7486d7c5f4b.jpg) +Figure 2. Overview of our T-FIT for fashion image editing. We divide the entire editing process into two stages. In the first stage, we perform counterfactual abduction to infer exogenous variables $U$ that capture the core visual contents of the source fashion image. In the second stage, we employ a concept learning module to discover the target concept from representation space and design a concept composition strategy for combining multiple aspects of the fashion image, such as fabric, color, or pattern, while maintaining consistency with the original image. + +images, aiming to reduce uncertainty in fashion image editing and prevent unintended changes in the structure or identity of the source image. In other words, by incorporating $U$ , our method ensures that the edited fashion image $I_{T}$ integrates the influence of $P'$ while retaining other unrelated visual content from the source image $I_{S}$ unchanged, thus addressing spurious correlation introduced in pre-trained models from the causal perspective. Specifically, we train the abduction $U$ by optimizing the following Gaussian noise regression, a process analogous to training the reversed diffusion steps: + +$$ +\arg \min _ {U} \mathbb {E} _ {(t, \epsilon)} | | \epsilon - \epsilon_ {U} (x _ {t}, t, \varepsilon_ {s r c}) | | _ {2} ^ {2}, \tag {1} +$$ + +where $\epsilon \sim \mathcal{N}(0,I)$ is the Gaussian noise, $t\in [0,T]$ represents a sampled time step, $\epsilon_{U}$ represents a new denoising model trained on a pre-trained denoising U-Net, wherein $U$ is treated as a trainable parameter. All other model parameters, aside from those of the denoising U-Net, remain fixed during training. The text embedding $\varepsilon_{src} = \psi (P)$ is obtained by a pre-trained CLIP text encoder $\psi (\cdot)$ , and the noisy input at time $t$ can be obtained by $x_{t} = \sqrt{\alpha_{t}} x_{0} + \sqrt{1 - \alpha_{t}}\epsilon$ . Here, $x_0 = I_S$ denotes the original image and $\alpha_{t}$ is a predefined variance schedule [20, 34]. Following [26], we parameterize $U$ as the UNet LoRA + +in $\epsilon_U$ , LoRA structure is incorporated across all attention layers, convolution layers, and feed-forward layers, which can be expressed by $W' = W + \beta U_A \cdot U_B$ . Here, $W \in \mathbb{R}^{d \times d}$ denotes the original weight matrix, $U_A \in \mathbb{R}^{d \times r}$ and $U_B \in \mathbb{R}^{r \times d}$ are low-rank matrices. Wang et al. [30] observed that the larger the time step in the diffusion model, the better the editable and the lower the fidelity. For this reason, the annealing parameter $\beta = \frac{1 - \kappa}{T^2} (t - T)^2 + \kappa$ is proposed to improve the editability. In our experiment, the rank $r$ is set to 512, $\kappa \in [0,1]$ is a constant. + +# 2.4. Compositional Concept Learning for Editing + +Unfortunately, the abduction of $U$ is inherently ill-posed [26], often leading to overfitting to the specific $P$ and $x$ . This overfitting reduces the editability of the generative model $G(\cdot, U)$ and may cause challenges like concept omission or confusion, where distinctions between concepts may blur, or certain concepts might overshadow others. Therefore, $G(\cdot, U)$ struggles to generate images that adhere to a new prompt $P'$ . To address this, we introduce a concept learning module that analyzes the representation space to identify composable concepts. Additionally, we propose a multi-concept composition strategy for flexible fashion image editing, enabling single and multi-concept continuous + +edits. For example, our method can simultaneously modify both the "sleeve length" (e.g., changing from long sleeves to short sleeves) and the "color" of the clothing (e.g., transitioning from red to blue). The intensity of these edits can also be adjusted, such as making the blue lighter or darker. + +Concept Learning Module. Given a target concept $C^i$ , we aim to learn an interpretable direction vector $\Delta C_t^i$ at time $t$ . The learned direction vector allows us to control the generation process by varying the latent $x_{t}$ in the desired directions $\Delta C_t^i$ with weight $w_{i}$ , which is considered as $s_{new}^{i} = s_{src} + w_{i}\Delta C_{t}^{i}$ . Here, $s_{src} = \epsilon_U(x_t,t,\varepsilon_{src})$ , $\varepsilon_{src} = \psi (P)$ is derived from pre-trained CLIP text encoder $\psi (\cdot)$ , and $\epsilon_U(x_t,t,\varepsilon_{src})$ is trained by counterfactual abduction and is fixed here. Specifically, we first define a set of text prompts $P_{C^i} = \{p_0^i,p_1^i,\dots,p_{N_i - 1}^i\}$ designed to elicit distinct distributions for a target concept $C^i$ while maintaining a consistent distribution for another concept $C^j$ . Without loss of the generality, we define $p_0^i$ as the source prompt without redundant concepts. In essence, each prompt captures a unique representation of $C^i$ while keeping the representation of $C^j$ consistent. For instance, when editing the concept of "fabric," we can formulate a series of prompts like "a lace [fabric 0] dress," "a silk [fabric 1] dress," and "a [fabric $N_{i} - 1$ ] dress." These prompts are intended to generate variations in the fabric concept while keeping other aspects, like dress type. + +Then, to identify the direction vector for concept $C^i$ , we first define its representation space as $\mathcal{R}_{C_t^i} \coloneqq \text{span}(\{v_t^j - v_t^0 : j = 1, \dots, N_i - 1\})$ . By leveraging the span of conceptual space, we enable the model to edit not just predefined attributes, but also a broader range of related attributes. For example, by defining a structured representation space with "blue," "green," and "yellow," our model can generalize to other colors in the spectrum, such as "red". Here, $v_t^j = \epsilon_U(x_t, t, \varepsilon_j)$ , each $\varepsilon_j = \psi(p_j^i)$ is derived from the pre-trained CLIP text encoder $\psi(\cdot)$ . Next, to compute the projection matrix $M_i$ for the $C^i$ -space within this representation space, we start by constructing matrix $S^i$ as + +$$ +S _ {t} ^ {i} := \left[ v _ {t} ^ {1} - v _ {t} ^ {0}, \dots , v _ {t} ^ {N _ {i} - 1} - v _ {t} ^ {0} \right]. \tag {2} +$$ + +In addition, to align changes in the representation space with the semantic structure of the concept space, we first apply singular value decomposition (SVD) on $S_{t}^{i}$ , i.e., $SVD(S_{t}^{i}) = V_{t}^{i}\Sigma_{t}^{i}U_{t}^{iT}$ , to identify the principal orthogonal semantic directions. We then select the top $K$ components of $V_{t}^{i}$ to capture the most significant semantic variations (for simplicity, we denote this matrix as $V_{t}^{i}$ ), and construct the projection matrix $M_{t}^{i} = V_{t}^{i}V_{t}^{iT}$ , which projects the concept vector into $\mathcal{R}_{C_t^i}$ and ensures that the changes in the representation space are corrected and aligned with the key semantic directions. + +Finally, to derive the editing direction, we calculate the + +score difference between source representation $s_{src}$ and target representation $s_{tar}$ as $\Delta s_t = s_{tar} - s_{src}$ . We then project $\Delta s_t$ onto the spanned concept space using $M_t^i$ , yielding $\Delta C_t^i = \Delta s_t M_t^i$ . This projection ensures the editing operation reflects the relevant semantic changes in the concept space. The modified score can be achieved by + +$$ +s = s _ {\emptyset} + \lambda \left(s _ {n e w} ^ {i} - s _ {\emptyset}\right), \tag {3} +$$ + +where $\lambda$ is guidance scale, $s_{\emptyset}$ denotes the unconditional score, is given by $s_{\emptyset} = \epsilon_U(x_t,t,\varepsilon_{\emptyset})$ with $\varepsilon_{\emptyset} = \psi (^{\prime \prime})$ In practice, we noticed that a few prompts are sufficient to identify the concept direction, which is attributed to the generalization ability of the span space $\mathcal{R}_{C_t^i}$ + +# Algorithm 1 Compositional Concept Learning + +Require: Diffusion model $\epsilon_U(x_t,t,\varepsilon)$ , pre-trained CLIP text encoder $\psi (\cdot)$ , guidance scale $\lambda$ , concept weight $\{w_{1},\dots,w_{m}\}$ , covariance matrix $\sigma_t^2 I$ , empty prompt " ", source prompt $P$ , target prompt $P^{\prime}$ , prompts for build concept space $P_{C^i} = \{p_0^i,\ldots ,p_j^i,\ldots ,p_{N_i - 1}^i\}$ + +$$ +\epsilon_ {U} \left(x _ {t}, t, \varepsilon_ {2}\right) - \epsilon_ {U} \left(x _ {t}, t, \varepsilon_ {0}\right), \dots , +$$ + +$$ +\left. \epsilon_ {U} \left(x _ {t}, t, \varepsilon_ {N _ {i} - 1}\right) - \epsilon_ {U} \left(x _ {t}, t, \varepsilon_ {0}\right) \right] +$$ + +$$ +S V D \left(S _ {t} ^ {i}\right) = V _ {t} ^ {i} \Sigma_ {t} ^ {i} U _ {t} ^ {i T} +$$ + +1: Initialize sample $x_{t} \sim \mathcal{N}(0, I)$ +2: $\varepsilon_{\emptyset}, \varepsilon_{src}, \varepsilon_{tar} = \psi(""), \psi(P), \psi(P')$ # Text embedding +3: $\varepsilon_{j} = \psi (p_{j}^{i})$ # Concept text embedding +4: for $t = T, \dots, 1$ do +5: $s_{\emptyset}\gets \epsilon_U(x_t,t,\varepsilon_{\emptyset})$ # Unconditional score +6: $s_{src}, s_{tar} \gets \epsilon_U(x_t, t, \varepsilon_{src}), \epsilon_U(x_t, t, \varepsilon_{tar})$ # Conditional score +7: $\Delta s_{t} = s_{tar} - s_{src}$ # Score difference +8: for $i = 1,2,\dots,m$ do # m concepts +9: $S_{t}^{i}\gets \left[\epsilon_{U}(x_{t},t,\varepsilon_{1}) - \epsilon_{U}(x_{t},t,\varepsilon_{0}),\right.$ +10: Compute the top- $k$ left singular vectors $V_{i}$ via: +11: $M_{t}^{i} = V_{t}^{i}(V_{t}^{i})^{T}$ # Projection matrix +12: $\Delta C_{t}^{i} = M_{t}^{i}\Delta s_{t}\#$ Editing direction +13: end for +14: $s \gets s_{\emptyset} + \lambda((s_{src} + w_i \Delta C_t^i) - s_{\emptyset})$ # For single-concept editing +15: $s \gets s_{\emptyset} + \lambda \sum_{i=1}^{m} \left( (s_{src} + w_i \Delta C_t^i) - s_{\emptyset} \right) \# \text{For multi-concept editing}$ +16: $x_{t - 1}\sim \mathcal{N}(x_t - s,\sigma_t^2 I)$ +17: end for + +Concept Composition Strategy. Now, our method enables single-concept editing for fashion images, but many scenarios demand multi-concept compositional editing. A straightforward way to achieve this is to create a multi-concept space using a set of concept prompts. While feasible, this approach limits fine-grained control over individual concepts, such as adjusting shades of color in specific items. Alternatively, sequential single-concept edits can be applied, but this accumulates errors and is time-consuming. + +![](images/59c0fb811e2553c4ff9cd744d56b07a62f20d2692dc3d000c3ed1bf4f02703a6.jpg) +Figure 3. Qualitative comparison with the state-of-art methods in single concept editing. + +To overcome these issues, we propose a controllable multi-concept editing strategy. This strategy enables precise, simultaneous adjustments of multiple concepts by weighting each concept's direction vector, which is shown as + +$$ +s _ {m} = s _ {\emptyset} + \lambda \sum_ {i = 1} ^ {m} \left(s _ {n e w} ^ {i} - s _ {\emptyset}\right), \tag {4} +$$ + +where $s_{new}^{i} = s_{src} + w_{i}\Delta C_{t}^{i}$ , we can control the manipulation strength of each concept. The whole algorithm is shown in Algorithm 1. + +# 3. Experiments + +# 3.1. Experimental Setups + +Implementation Details. In our work, all experiments are conducted on a single NVIDIA RTX 4090 GPU with 24 GB of memory. Following the prior work [3, 4, 26], we employ the official pre-trained Stable Diffusion v2.1-base model as our foundational model, downsampling all images to a resolution of $512 \times 512$ pixels for consistency across experiments. For the counterfactual abduction phase, We fine-tune the model on the single image using LoRA with a rank of 512 for 100 epochs, employing the Adam optimizer with a learning rate of 1e-4. To conserve memory, we adopt the mixed precision [17] and gradient accumulation strategies, setting both the accumulation step count and + +batch size to 1. During the inference phase, we utilize the DDIM sampling with 80 steps and a classifier-free guidance scale of $\lambda = 4$ . Additionally, we explore various values for the concept weight $w_{i}$ within the range [1, 5] to control the editing strength, where $i$ represents the $i$ -th edited concept. + +Dataset. To comprehensively evaluate the performance of our T-FIT in fashion image editing, we follow [14, 26, 40] by collecting a diverse set of fashion images from Unsplash1, including both full-body and half-body images. The selected images comprise various clothing types, such as long dresses, skirts, T-shirts, shorts, pants, and strapless tops. Our evaluation encompasses a diverse set of editing concepts, which include specific attributes (e.g., gender, color, patterns, fabric texture) as well as broader elements such as clothing type, garment length, and sleeve style. Each image in the experiment is annotated with $N_{i} + 2$ prompts, consisting of one prompt for the original image and $N_{i} + 1$ prompts for target images. + +Baselines. We conduct comparisons with existing representative diffusion-based image editing methods, including LEDs [29], LEDs++ [2], TexFit [32], Guide-and-Rescale (G & R) [28], UltraEdit [41], and Turboedit [6]. + +Evaluation Metrics. To comprehensively evaluate the performance of our T-FIT, we conduct assessments from both + +![](images/468ca461f1b15c07a77790f11c9eaee95fd05eeb64efe41988d837d051d19e0e.jpg) +Figure 4. Qualitative comparison with the state-of-art methods in multiple concept editing. + +objective and subjective perspectives. Objectively, we follow the previous works [3, 8, 9] and use metrics in CLIP embedding space: CLIP image similarity (CLIP-I) [21] to measure identity preservation via calculate the cosine similarity between the edited image and the original image, and CLIP text-image direction similarity (CLIP-D) [7] to evaluate the correspondence between image changes and text changes. Subjectively, we qualitatively evaluate the fidelity of the edited images and their consistency with the target prompts, supplemented by human preference studies involving recruited volunteers from different disciplines. + +# 3.2. Qualitative Evaluation + +We show some qualitative experimental results in Fig. 3-4, from our experiments, we observe the following: + +For single-concept fashion image editing, 1) several methods, such as LEDs++ [2], introduce unintended changes to non-target areas, possibly due to spurious correlation embedded in the pre-trained models from their training datasets. These correlations cause inaccurate concept associations, leading to changes outside the intended editing area. For instance, in Fig. 3, line 2, LEDs++ unexpectedly replaces the shoelaces with pink ones. Similarly, in Fig. 3, line 5, UltraEdit [41] transforms a squared neckline into a V-neckline, Turbo-edit [6] alters the dress length. 2) G & R [28], and UltraEdit often confuse different objects. For instance, when attempting to change the color of + +pants to red (Fig. 3, line 2), these methods inadvertently modify unrelated elements, such as the color of the shirt or hair. 3) TexFit's effectiveness is significantly hampered by its reliance on the accuracy of the recognized mask, which restricts its applicability in fashion image editing. As shown in the last column of Fig. 3, TexFit [32] either alters only localized information in the target editing area or excessively expands the intended modification area. This results in unnatural edits that disrupt the overall coherence of the image. For multi-concept fashion image editing, 1) UltraEdit [41], TexFit [32] often overlooks certain concepts in multiple-concept editing scenarios. For example, in Fig. 4, line 1, UltraEdit remains in top with short sleeves rather than long sleeves; in Fig. 4, line 2, the trousers edit is ignored. 2) G & R [28], TurboEdit [6], and LEDITS++ [2], demonstrate commendable performance in adhering to user instructions. However, the edited images can exhibit issues such as alterations to human poses or discontinuities in the edited results. For example, when modifying sleeve lengths, the transitions in the sleeve areas may appear inconsistent or fragmented (Fig. 4, line 1). These problems highlight the challenges in maintaining coherence and integrity in the edited images, despite following the provided directives closely. 3) Several methods such as Turbo-edit [6] and UltraEdit [41] may over-edit, making it difficult to retain the details of the original image. For example, when modifying pants into a skirt (Fig. 4, line 3), the edited + +image may change the length of the skirt as well as other aspects such as the color and fabric of the top along with the skirt. This phenomenon may be attributed to the spurious correlation that were learned by the pre-trained model from the training dataset. Compared to the existing methods, our method excels in fashion image editing, and it can directly control precise multi-concepts from text prompts without additional input, such as masks or poses. The edited image retains non-target areas, closely following the target text while maintaining the consistency of the image. In addition, our method allows continuous editing adjustments by adjusting the concept weight (Fig. 5), which provides fine control over the editing strength. This results in high-quality editing that is superior to the baseline model, especially for multi-concept fashion image editing. + +# 3.3.Quantitative Evaluation + +We quantitatively evaluate our T-FIT against baselines using both automatic metrics and human evaluations. + +Automatic Metric Comparisons. We summarize the experimental findings in Tab. 1. Our method achieves state-of-the-art results for single- and multi-concept fashion image editing in CLIP-D, which demonstrates a significant advantage in terms of the accuracy and fidelity of our editing results in following the text description. In terms of the CLIP-I metric, our method demonstrates superior performance compared to most competing approaches. Although it slightly underperforms Turbo-edit [6] and TexFit [32] on CLIP-I, this is due to specific limitations in these methods during multi-concept fashion image editing. Turbo-edit often omits certain concepts, resulting in edited images that closely resemble the source, thus boosting its CLIP-I score. TexFit, on the other hand, struggles with precise edit masking, often restricting changes to small areas and failing in non-rigid edits (e.g., changing long sleeves to shorts or pants to shorts), leaving much of the image unaltered. Overall, our method shows superior performance on objective metrics for both multi- and single-concept editing tasks. + +Human Preference Study. In this section, we quantitatively evaluate our method with an extensive human perceptual evaluation study. We first collect a diverse set of 30 fashion images and their corresponding text descriptions, covering a range of clothing types (e.g., long sleeves, short sleeves, dresses, pants, shorts), patterns (e.g., solid, stripes, prints), and materials (e.g., cotton, denim, chiffon). Then, we invite 36 volunteers to participate in the evaluation. Each volunteer views the original fashion image, the original description text, the target description text, and the edited images generated by 6 different methods, including ours. The volunteers rate the edited images based on two criteria: (1) how well the edits align with the concepts specified in the target text and (2) the consistency of the non-edited areas with the original image. Ratings range from + +Table 1. Quantitative comparisons. The best results are bolded and the second-best results are underlined, respectively. + +
Single conceptMultiple concepts
MethodCLIP-DCLIP-ICLIP-DCLIP-I
OUR0.14330.91600.11520.9259
G & R0.09050.91160.05390.9079
Turbo-edit0.08210.91010.09670.9355
UltraEdit0.07710.90560.05330.8974
LEDITS++0.09540.91600.05840.9115
Texfit0.03460.95000.04630.9562
+ +![](images/8a5555857ecd9e4ba32152568dc680ae5b212dabc974743126e92e9b594aa417.jpg) +Figure 5. Human preference study. Our method outperforms baseline models in both text- and image alignment, demonstrating significant advantages in fashion image editing. + +1 to 8, with 1 representing the best performance and 8 the poorest. We calculate the average ranks provided by each participant, and the final results are displayed in Fig. 5. Our findings indicate that our method surpasses baseline methods on both criteria, demonstrating enhanced fidelity in executing intended edits while maintaining consistency across unaltered image areas. + +# 3.4. Ablation Study + +Next, we turn to an ablation study where we analyze the effect of the different components inherent in our approach. Specifically, we consider: (1) the effect of substituting the CA with null-text inversion [18], (2) the effect of removing CLM, (3) the effects of varying annealing parameter $\kappa$ , and (4) the influence of different weights on the editing results of concepts. The visual results are provided in Fig. 6-8. + +From these figures, it is evident that when T-FIT replaces the CA with the null-text inversion module, the edited images exhibit a noticeable spurious correlation. For instance, in Fig. 7, when the concept of "gender" is applied, the collar of the clothing incorrectly changes from a "stand-up collar" to a "round collar". On the other hand, T-FIT is equipped + +![](images/fcc5e75fd19165d6cd944b3deff0ab9af37a5909ee63ec4d8ba0e1e22017cefd.jpg) +Figure 6. Qualitative ablation results. The last image shows the result of T-FIT, and the other images show the result of T-FIT without CLM and adjusting annealing parameter $\kappa$ . Removing CLM significantly reduces the model's editability. Adjusting $\kappa$ still makes balancing editability and fidelity challenging. + +![](images/ede48eaa7f596cb4374c1a46ee991b2866b2fd2d0703a93755a5239dfe733318.jpg) +Figure 7. Qualitative ablation results. With CA, our model captures essential visual content from the original image, mitigating the impact of spurious correlation in pre-trained data. CLM ensures that no editing concepts are overlooked during modifications. + +solely with CA, it effectively preserves the visual details of the source image and mitigates issues arising from spurious correlation. However, this setup offers limited flexibility for editing, even if we adopt varying annealing parameters $\kappa$ (see Fig. 6). This suggests that while CA effectively addresses the spurious correlation issues arising from the pretrained model, it is limited in its editability. When CLM is added, T-FIT's editability improves. This is attributable to the CLM's capability to identify disentangled concepts within the representation space, which facilitates more flexible and precise editing. Furthermore, as illustrated in Fig. 8, T-FIT addresses the issues of concept omission and confusion through the multi-concept composite strategy, while also providing fine control over the intensity of each concept by adjusting individual concept weights. In our work, the concept weights can be manually adjusted in the range [1-5], with higher values indicating greater editing intensity on that concept and vice versa. This offers greater control and flexibility for fashion image editing of our method. + +![](images/3e4059db8be4fc614a6afa59b0ea500bdd536a9348241d5e6e5067368f522d19.jpg) +Figure 8. Experimental comparison of varying concept weights (from 1 to 5) on single- and multi-concept fashion image editing. [C1], [C2], and [C3] mean different editing concepts, respectively. + +# 4. Conclusion + +In this paper, we introduce T-FIT, a novel framework for fashion image editing. Our approach, which uses a single fashion image and simple text description, supports both single- and multi-concept image editing tasks. By incorporating counterfactual abduction and a concept learning module, T-FIT effectively mitigates issues of spurious correlation in the pre-trained model, enabling efficient, flexible, and accurate text-driven image editing. Furthermore, T-FIT allows fine-grained editing of complex fashion concepts by adjusting concept weights. Extensive experiments confirm the effectiveness of T-FIT and its components, demonstrating significant advantages over several existing methods. Our future work will focus on enhancing editing accuracy by incorporating visual information (e.g., style images) to learn richer visual concept representations. + +# 5. Acknowledgment. + +This work was supported by grants from the National Key R & D Program of China (grant no. 2022YFB3303302), and the National Natural Science Foundation of China (grant nos. 62377040, 62477004, 62207007, 623B2002). + +# References + +[1] Alberto Baldrati, Davide Morelli, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. Multimodal garment designer: Human-centric latent diffusion models for fashion image editing. In ICCV, 2023. 1 +[2] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. LEDs++: Limitless image editing using text-to-image models. In CVPR, 2024. 1, 2, 5, 6 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 5, 6 +[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 1, 5 +[5] Zhikang Chen, Min Zhang, Sen Cui, Haoxuan Li, Gang Niu, Mingming Gong, Changshui Zhang, and Kun Zhang. Neural collapse inspired feature alignment for out-of-distribution generalization. In NeurIPS, 2024. 1 +[6] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv:2408.00735, 2024. 1, 5, 6, 7 +[7] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. StyleGAN-NADA: CLIP-guided domain adaptation of image generators. ACM Transactions on Graphics, 41(4):1-13, 2022. 6 +[8] Julia Guerrero-Viu, Milos Hasan, Arthur Roullier, Midhun Harikumar, Yiwei Hu, Paul Guerrero, Diego Gutierrez, Bellen Masia, and Valentin Deschaintre. Texsliders: Diffusion-based texture editing in clip space. In SIGGRAPH, 2024. 6 +[9] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 6 +[10] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Multiedits: Simultaneous multi-aspect editing with text-to-image diffusion models. In NeurIPS, 2024. 2 +[11] Shanshan Huang, Haoxuan Li, Qingsong Li, Chunyuan Zheng, and Li Liu. Pareto invariant representation learning for multimedia recommendation. In ACM MM, 2023. 1 +[12] Shanshan Huang, Qingsong Li, Jun Liao, Shu Wang, Li Liu, and Lian Li. Controllable image synthesis methods, applications and challenges: a comprehensive survey. Artificial Intelligence Review, 57(12):336, 2024. 1 +[13] Shanshan Huang, Lei Wang, Jun Liao, and Li Liu. Multi-attentional causal intervention networks for medical image diagnosis. Knowledge-Based Systems, 299:111993, 2024. 1 +[14] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 1, 5 +[15] Gwanhyeong Koo, Sunjae Yoon, Ji Woo Hong, and Chang D. Yoo. Flexiedit: Frequency-aware latent refinement for enhanced non-rigid editing. In ECCV, 2024. 1 + +[16] Meng Li and Haochen Sui. Causal recommendation via machine unlearning with a few unbiased data. In AAAI Workshop on AICT, 2025. 1 +[17] Paulius Micikevicius, Sharan Narang, Jonah Alben, Greg Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2018. 5 +[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 7 +[19] Martin Pernus, Clinton Fookes, Vitomir Štruc, and Simon Dobrišek. FICE: Text-conditioned fashion-image editing with guided gan inversion. Pattern Recognition, 158:111022, 2025. 1 +[20] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion Autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 3 +[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 +[22] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 2022. 1 +[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2 +[24] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1 +[25] Pedro Sanchez, Jeremy P Voisey, Tian Xia, Hannah I Watson, Alison Q O'Neil, and Sotirios A Tsaftaris. Causal machine learning for healthcare and precision medicine. Royal Society Open Science, 9(8):220638, 2022. 1 +[26] Xue Song, Jiequan Cui, Hanwang Zhang, Jingjing Chen, Richang Hong, and Yu-Gang Jiang. Doubly abductive counterfactual inference for text-based image editing. In CVPR, 2024. 3, 5 +[27] Adam Stein, Aaditya Naik, Yinjun Wu, Mayur Naik, and Eric Wong. Towards compositionality in concept learning. In ICML, 2024. 2 +[28] Vadim Titov, Madina Khalmatova, Alexandra Ivanova, Dmitry Vetrov, and Aibek Alanov. Guide-and-Rescale: Self-guidance mechanism for effective tuning-free real image editing. In ECCV, 2024. 1, 2, 5, 6 +[29] Linoy Tsaban and Apolinário Passos. LEDs: Real image editing with ddpm inversion and semantic guidance. arXiv:2307.00522, 2023. 1, 5 +[30] Luozhou Wang, Shuai Yang, Shu Liu, and Ying-cong Chen. Not all steps are created equal: Selective diffusion distillation for image manipulation. In ICCV, 2023. 3 + +[31] Lei Wang, Shanshan Huang, Shu Wang, Jun Liao, Tingpeng Li, and Li Liu. A survey of causal discovery based on functional causal model. Engineering Applications of Artificial Intelligence, 133:108258, 2024. 1 +[32] Tongxin Wang and Mang Ye. TexFit: Text-driven fashion image editing with diffusion models. In AAAI, 2024. 1, 5, 6, 7 +[33] Xiaolong Wang, Zhi-Qi Cheng, Jue Wang, and Xiaojiang Peng. DPDEdit: Detail-preserved diffusion models for multimodal fashion image editing. arXiv:2409.01086, 2024. 1 +[34] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, Fei Wang, Christopher De Sa, and Volodymyr Kuleshov. Infodiffusion: Representation learning using information maximizing diffusion models. In ICML, 2023. 3 +[35] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for (score-based) text-controlled generative models. NeurIPS, 2024. 1, 2 +[36] Peng Wu, Haoxuan Li, Yuhao Deng, Wenjie Hu, Quanyu Dai, Zhenhua Dong, Jie Sun, Rui Zhang, and Xiao-Hua Zhou. On the opportunity of causal learning in recommendation systems: Foundation, estimation, prediction and challenges. In IJCAI, 2022. Survey Track. 1 +[37] Xindi Wu, Dingli Yu, Yangsibo Huang, Olga Russakovsky, and Sanjeev Arora. Conceptmix: A compositional image generation benchmark with controllable difficulty. In NeurIPS, 2024. 1 +[38] Yuntian Wu, Yuntian Yang, Jiabao Sean Xiao, Chuan Zhou, Haochen Sui, and Haoxuan Li. Invariant spatiotemporal representation learning for cross-patient seizure classification. In NeurIPS Workshop on NeuroAI, 2024. 1 +[39] Min Zhang, Haoxuan Li, Fei Wu, and Kun Kuang. Metacoco: A new few-shot classification benchmark with spurious correlation. In ICLR, 2024. 1 +[40] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In CVPR, 2023. 5 +[41] Haozhe Zhao, Xiaojian (Shawn) Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. In NeurIPS, 2024. 1, 5, 6 \ No newline at end of file diff --git a/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/images.zip b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a2c97ffe472cf9818015b7cb27db7492bb72b4a --- /dev/null +++ b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d75894728a884fe8f4faaa27edc8986328492438a889361421cf60cf14c5a6fd +size 929398 diff --git a/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/layout.json b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0b6d494204e648ac1d30f3d4f4daa1ced4da9f55 --- /dev/null +++ b/2025/Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction/layout.json @@ -0,0 +1,8760 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 88, + 547, + 124 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 88, + 547, + 124 + ], + "spans": [ + { + "bbox": [ + 64, + 88, + 547, + 124 + ], + "type": "text", + "content": "Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "spans": [ + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "text", + "content": "Shanshan Huang" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "text", + "content": ", Haoxuan Li" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "text", + "content": ", Chunyuan Zheng" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "text", + "content": ", Mingyuan Ge" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "text", + "content": ", Wei Gao" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "text", + "content": ", Lei Wang" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "text", + "content": ", Li Liu" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 146, + 548, + 175 + ], + "type": "text", + "content": "School of Big Data & Software Engineering, Chongqing University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 175, + 548, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 175, + 548, + 217 + ], + "spans": [ + { + "bbox": [ + 63, + 175, + 548, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 175, + 548, + 217 + ], + "type": "text", + "content": "Center for Data Science, Peking University " + }, + { + "bbox": [ + 63, + 175, + 548, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 175, + 548, + 217 + ], + "type": "text", + "content": "School of Mathematical Sciences, Peking University " + }, + { + "bbox": [ + 63, + 175, + 548, + 217 + ], + "type": "inline_equation", + "content": "\\{\\text { shanshanhuang, dcsliuli } \\} @ \\text { cqu.edu.cn, } \\{\\text { hxli, cyzheng } \\} @ \\text {stu.pku.edu.cn, }" + }, + { + "bbox": [ + 63, + 175, + 548, + 217 + ], + "type": "inline_equation", + "content": "\\{\\text { leiwangtt, mingyuange } \\} @ \\text {stu.cqu.edu.cn, }" + }, + { + "bbox": [ + 63, + 175, + 548, + 217 + ], + "type": "text", + "content": " wgao2002@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 258, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 296, + 594 + ], + "type": "text", + "content": "Fashion image editing is a valuable tool for designers to convey their creative ideas by visualizing design concepts. With the recent advances in text editing methods, significant progress has been made in fashion image editing. However, they face two key challenges: spurious correlations in training data often induce changes in other areas when editing an area representing the intended editing concept, and these models typically lack the ability to edit multiple concepts simultaneously. To address the above challenges, we propose a novel Text-driven Fashion Image editing framework called T-FIT to mitigate the impact of spurious correlation by integrating counterfactual reasoning with compositional concept learning to precisely ensure compositional multi-concept fashion image editing relying solely on text descriptions. Specifically, T-FIT includes three key components. (i) Counterfactual abduction module, which learns an exogenous variable of the source image by a denoising U-Net model. (ii) Concept learning module, which identifies concepts in fashion image editing—such as clothing types and colors and projects a target concept into the space spanned from a series of textual prompts. (iii) Concept composition module, which enables simultaneous adjustments of multiple concepts by aggregating each concept's direction vector obtained from the concept learning module. Extensive experiments show that our method can achieve state-of-the-art performance on various fashion image editing tasks, including single-concept editing (e.g., sleeve length, clothing type) and multi-concept editing (e.g., color & sleeve length)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 613, + 135, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 135, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 135, + 625 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 633, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 295, + 694 + ], + "type": "text", + "content": "Fashion image editing has emerged as a pivotal field within computer vision, driven by the increasing need for personalized and dynamic visual modifications in the fashion industry [1, 33]. The diversity and high-quality image generation capabilities of large-scale text-image diffusion models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 234, + 554, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 554, + 317 + ], + "type": "text", + "content": "such as Stable Diffusion [23], DALL-E 2 [24], and Imagegen [22] has inspired many subsequent efforts to leverage pre-trained large-scale models for text-driven fashion image editing (TFIE) [12, 19, 32]. A foundational aspect of this field is single-concept editing, where a specific concept [35, 37] like color, texture, or style of an item is altered while keeping the rest of the fashion image the same." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 593 + ], + "type": "text", + "content": "However, the existing methods still suffer from several limitations. First, these methods [2, 6, 41] often encounter challenges related to spurious correlation and unwanted associations learned by the pre-trained model (e.g., Stable Diffusion [23]) due to biases in training data. For example, several methods [28, 29] may erroneously associate a specific clothing type (e.g., sleeve length) with a specific cloth style (e.g., garment length), as shown in Fig. 1, resulting in unintended modifications outside the target concept. This is because the long sleeves always co-occur with long garments, inducing spurious correlations in the training data. Therefore, eliminating spurious correlations is critical to ensuring precise, isolated modifications. There are many methods to address spurious correlation [5, 11, 13, 31, 36, 38, 39]. For example, Li et al. [16] propose a novel machine unlearning method to reduce the instance weight of biased samples in e-commerce and Sanchez et al. [25] discuss important challenges present in healthcare applications such as processing high-dimensional and unstructured data, as well as temporal relationships. However, due to differences in data structure and final goal, previous methods could not be adapted to image editing scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 689 + ], + "type": "text", + "content": "Second, existing methods [32, 33] are mainly limited to editing the texture of fashion images, and it is difficult to achieve non-rigid editing, such as precisely adjusting the length of sleeves or skirts. As shown in Fig. 1, UltraEdit [41] and Turbo-edit [6] cannot achieve non-rigid editing, e.g., turning a long skirt into a mini skirt; altering short sleeves into long sleeves, which may attribute to the poor understanding of the concept from the prompt." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Third, existing methods [4, 14, 15] typically focus on editing a single concept in the source image. While single-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28726" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 288, + 187 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 288, + 187 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 288, + 187 + ], + "type": "image", + "image_path": "b96dd172db952754bc9d5c9cd4e11fdeba3d06d7f1e038f1091050054ac3794a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 296, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 258 + ], + "type": "text", + "content": "Figure 1. Two examples to illustrate the spurious correlation and conceptual omission issues in fashion image editing. Unlike other methods, our method alleviates the issue of over-editing that exists in fashion image editing due to spurious correlation (e.g., sleeve length and garment length, mini skirt and plaid style), as well as the confusion or lack of concepts when editing multiple concepts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 267, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 295, + 422 + ], + "type": "text", + "content": "concept methods enable the modification of multiple concepts sequentially, i.e., changing the concepts one by one, they present certain drawbacks. Specifically, these methods either introduce concept omission due to the difficulty of identifying each concept from the prompt; or they lead to unexpected changes due to interactions between different concepts, which is still caused by spurious correlation [10]. For example, modifications made to later concepts may affect previous modifications. From Fig. 1, we can observe that G & R [28] incorrectly adds a \"leather bag\" when editing the skirt fabric to leather and LEDs++ [2] causes other unrelated attributes changes when changing the length and fabric of the skirt, such as the color of the gloves." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 533 + ], + "type": "text", + "content": "To bridge this gap, we develop a Text-driven Fashion Image editing framework (T-FIT) that relies solely on textual prompts, enabling users to perform both single- and multi-concept editing on a single fashion image. Specifically, T-FIT adopts Stable Diffusion [23] and introduces a counterfactual abduction module to address the spurious correlation and concept learning module with concept composition to achieve precise single- and multi-concept editing as shown in Fig. 2. Our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 534, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 569 + ], + "type": "text", + "content": "- We reveal the issue of spurious correlation and concept omission in fashion image editing and propose T-FIT, a text-driven framework to address these issues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 618 + ], + "type": "text", + "content": "- We propose counterfactual abduction (CA) to capture visual content and a concept learning module (CLM) to identify and disentangle concepts, mitigating issues of spurious correlation and non-rigid editing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "type": "text", + "content": "- We propose a concept composition strategy that enables simultaneous, fine-grained editing across multiple concepts by adjusting the weight of each learned concept to address the concept omission problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": "- Extensive empirical evaluation, in terms of both automated and human assessment metrics, qualitatively and quantitatively demonstrates that our T-FIT approach significantly enhances the quality of fashion image editing." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "text", + "content": "2. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 88, + 440, + 101 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 88, + 440, + 101 + ], + "spans": [ + { + "bbox": [ + 314, + 88, + 440, + 101 + ], + "type": "text", + "content": "2.1. Problem Formulation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": "We first formulate the image editing problem as follows: given a source fashion image " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "I_{S}" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": ", a corresponding text prompt " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": " that describes the contents of " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "I_{S}" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": ", a set of concept prompts " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "P_{C^i} = \\{p_0^i, p_1^i, \\ldots, p_{N-1}^i\\}" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": " used to construct a concept " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": " (e.g., clothing type, color, fabric) that the user wishes to edit, and a target text prompt " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": " describing the final editing goal, our objective is to generate a new fashion image " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "I_T" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": " by editing " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "I_S" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": " in alignment with " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": " while preserving other areas that are unrelated to the concept " + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 313, + 107, + 555, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 220, + 394, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 394, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 394, + 234 + ], + "type": "text", + "content": "2.2. Preliminary" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "text", + "content": "In the context of fashion image editing, concepts " + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "inline_equation", + "content": "C = \\{C^1,\\dots ,C^i,\\dots ,C^m\\}" + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "text", + "content": " are a distinct and human-interpretable feature or attribute within an image, encompassing elements like sleeve length, fabric type, color, and cloth type. Each concept " + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "text", + "content": " contains mutually selected attributes, where each attribute embodies certain shared properties. For instance, the attributes \"long sleeves\" and \"short sleeves\" collectively represent the concept of \"sleeve length\", while \"woman\" and \"man\" represent the concept of \"gender\". The concept representations " + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "inline_equation", + "content": "s(C)" + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "text", + "content": " can be derived from the intermediate representation of some text-to-image generative models (e.g., Stable Diffusion) pre-trained over annotated datasets that link visual attributes to textual descriptions [27, 35]. Here, " + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "inline_equation", + "content": "s:\\mathbb{C}\\to \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "text", + "content": " is the concept representation function, where " + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "text", + "content": " is the concept space containing all possible concepts, and " + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "text", + "content": " represents the representation space with dimension " + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 239, + 555, + 455 + ], + "type": "text", + "content": ". Following [27, 35], we focus on the compositional concepts, as defined in Definition 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 459, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 509 + ], + "type": "text", + "content": "Definition 1 (Compositional Concepts). For concepts " + }, + { + "bbox": [ + 313, + 459, + 555, + 509 + ], + "type": "inline_equation", + "content": "C^i, C^j \\in \\mathbb{C}" + }, + { + "bbox": [ + 313, + 459, + 555, + 509 + ], + "type": "text", + "content": ", the concept representation " + }, + { + "bbox": [ + 313, + 459, + 555, + 509 + ], + "type": "inline_equation", + "content": "s: \\mathbb{C} \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 459, + 555, + 509 + ], + "type": "text", + "content": " is considered compositional if there exist positive weights " + }, + { + "bbox": [ + 313, + 459, + 555, + 509 + ], + "type": "inline_equation", + "content": "w_i, w_j \\in \\mathbb{R}^+" + }, + { + "bbox": [ + 313, + 459, + 555, + 509 + ], + "type": "text", + "content": " such that:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 512, + 508, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 512, + 508, + 528 + ], + "spans": [ + { + "bbox": [ + 359, + 512, + 508, + 528 + ], + "type": "interline_equation", + "content": "s (C ^ {i} \\cup C ^ {j}) = w _ {i} s (C ^ {i}) + w _ {j} s (C ^ {j}).", + "image_path": "96c708fec4f93ff902f3879b08a54c22b021c88ce47b81ff372359dbfab809bb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 533, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 631 + ], + "type": "text", + "content": "This definition indicates that the representation of the composed concepts corresponds to the weighted sum of the individual concept representations in the representation space. Following this definition, this study aims to develop a framework for learning and composing conceptual representations from pre-trained text-to-image models (i.e., Stable Diffusion) to achieve flexible and targeted fashion image editing based on these compositional concepts." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 547, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 547, + 647 + ], + "type": "text", + "content": "2.3. Counterfactual Abduction for Image Editing" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "To address spurious correlation issues inherent in existing image editing methods, we introduce a counterfactual abduction approach that employs an abduction loss to infer unknown exogenous variables " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " for each image. These variables capture the additional visual content of source fashion" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28727" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 289, + 189 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 289, + 189 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 289, + 189 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 289, + 189 + ], + "type": "image", + "image_path": "2555b61cabbaea1513b210b88ac1c07b76452ba585d97de9a1b00128e71766e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 295, + 175, + 390, + 186 + ], + "lines": [ + { + "bbox": [ + 295, + 175, + 390, + 186 + ], + "spans": [ + { + "bbox": [ + 295, + 175, + 390, + 186 + ], + "type": "text", + "content": "(c) Concept Composition" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 292, + 83, + 541, + 169 + ], + "blocks": [ + { + "bbox": [ + 292, + 83, + 541, + 169 + ], + "lines": [ + { + "bbox": [ + 292, + 83, + 541, + 169 + ], + "spans": [ + { + "bbox": [ + 292, + 83, + 541, + 169 + ], + "type": "image", + "image_path": "51745256b6111e78b3730bc15ac80724ee41a28059635bfd2d0cdbcca7209aed.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 190, + 547, + 345 + ], + "blocks": [ + { + "bbox": [ + 62, + 190, + 547, + 345 + ], + "lines": [ + { + "bbox": [ + 62, + 190, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 190, + 547, + 345 + ], + "type": "image", + "image_path": "33f9009859475a5df608025691b2a57bbfb661822990095a6a0ce7486d7c5f4b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "text", + "content": "Figure 2. Overview of our T-FIT for fashion image editing. We divide the entire editing process into two stages. In the first stage, we perform counterfactual abduction to infer exogenous variables " + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "text", + "content": " that capture the core visual contents of the source fashion image. In the second stage, we employ a concept learning module to discover the target concept from representation space and design a concept composition strategy for combining multiple aspects of the fashion image, such as fabric, color, or pattern, while maintaining consistency with the original image." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "text", + "content": "images, aiming to reduce uncertainty in fashion image editing and prevent unintended changes in the structure or identity of the source image. In other words, by incorporating " + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "text", + "content": ", our method ensures that the edited fashion image " + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "inline_equation", + "content": "I_{T}" + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "text", + "content": " integrates the influence of " + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "text", + "content": " while retaining other unrelated visual content from the source image " + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "inline_equation", + "content": "I_{S}" + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "text", + "content": " unchanged, thus addressing spurious correlation introduced in pre-trained models from the causal perspective. Specifically, we train the abduction " + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 54, + 419, + 296, + 552 + ], + "type": "text", + "content": " by optimizing the following Gaussian noise regression, a process analogous to training the reversed diffusion steps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 557, + 295, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 557, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 99, + 557, + 295, + 575 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {U} \\mathbb {E} _ {(t, \\epsilon)} | | \\epsilon - \\epsilon_ {U} (x _ {t}, t, \\varepsilon_ {s r c}) | | _ {2} ^ {2}, \\tag {1}", + "image_path": "e3189071926986bf9c184a2774524ea42250d863a23ae132761622650439eb04.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": " is the Gaussian noise, " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": " represents a sampled time step, " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{U}" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": " represents a new denoising model trained on a pre-trained denoising U-Net, wherein " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": " is treated as a trainable parameter. All other model parameters, aside from those of the denoising U-Net, remain fixed during training. The text embedding " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\varepsilon_{src} = \\psi (P)" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": " is obtained by a pre-trained CLIP text encoder " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": ", and the noisy input at time " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": " can be obtained by " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{t} = \\sqrt{\\alpha_{t}} x_{0} + \\sqrt{1 - \\alpha_{t}}\\epsilon" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_0 = I_S" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": " denotes the original image and " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": " is a predefined variance schedule [20, 34]. Following [26], we parameterize " + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": " as the UNet LoRA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": "in " + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\epsilon_U" + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": ", LoRA structure is incorporated across all attention layers, convolution layers, and feed-forward layers, which can be expressed by " + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "inline_equation", + "content": "W' = W + \\beta U_A \\cdot U_B" + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": " denotes the original weight matrix, " + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "inline_equation", + "content": "U_A \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "inline_equation", + "content": "U_B \\in \\mathbb{R}^{r \\times d}" + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": " are low-rank matrices. Wang et al. [30] observed that the larger the time step in the diffusion model, the better the editable and the lower the fidelity. For this reason, the annealing parameter " + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\beta = \\frac{1 - \\kappa}{T^2} (t - T)^2 + \\kappa" + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": " is proposed to improve the editability. In our experiment, the rank " + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": " is set to 512, " + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\kappa \\in [0,1]" + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": " is a constant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 550, + 547, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 547, + 563 + ], + "type": "text", + "content": "2.4. Compositional Concept Learning for Editing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Unfortunately, the abduction of " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " is inherently ill-posed [26], often leading to overfitting to the specific " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". This overfitting reduces the editability of the generative model " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "G(\\cdot, U)" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " and may cause challenges like concept omission or confusion, where distinctions between concepts may blur, or certain concepts might overshadow others. Therefore, " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "G(\\cdot, U)" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " struggles to generate images that adhere to a new prompt " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". To address this, we introduce a concept learning module that analyzes the representation space to identify composable concepts. Additionally, we propose a multi-concept composition strategy for flexible fashion image editing, enabling single and multi-concept continuous" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28728" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " edits. For example, our method can simultaneously modify both the \"sleeve length\" (e.g., changing from long sleeves to short sleeves) and the \"color\" of the clothing (e.g., transitioning from red to blue). The intensity of these edits can also be adjusted, such as making the blue lighter or darker." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": "Concept Learning Module. Given a target concept " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": ", we aim to learn an interpretable direction vector " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\Delta C_t^i" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": ". The learned direction vector allows us to control the generation process by varying the latent " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " in the desired directions " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\Delta C_t^i" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " with weight " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": ", which is considered as " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "s_{new}^{i} = s_{src} + w_{i}\\Delta C_{t}^{i}" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "s_{src} = \\epsilon_U(x_t,t,\\varepsilon_{src})" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\varepsilon_{src} = \\psi (P)" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " is derived from pre-trained CLIP text encoder " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\epsilon_U(x_t,t,\\varepsilon_{src})" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " is trained by counterfactual abduction and is fixed here. Specifically, we first define a set of text prompts " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "P_{C^i} = \\{p_0^i,p_1^i,\\dots,p_{N_i - 1}^i\\}" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " designed to elicit distinct distributions for a target concept " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " while maintaining a consistent distribution for another concept " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "C^j" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": ". Without loss of the generality, we define " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "p_0^i" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " as the source prompt without redundant concepts. In essence, each prompt captures a unique representation of " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " while keeping the representation of " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "C^j" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": " consistent. For instance, when editing the concept of \"fabric,\" we can formulate a series of prompts like \"a lace [fabric 0] dress,\" \"a silk [fabric 1] dress,\" and \"a [fabric " + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "inline_equation", + "content": "N_{i} - 1" + }, + { + "bbox": [ + 55, + 134, + 295, + 386 + ], + "type": "text", + "content": "] dress.\" These prompts are intended to generate variations in the fabric concept while keeping other aspects, like dress type." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "content": "Then, to identify the direction vector for concept " + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "content": ", we first define its representation space as " + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C_t^i} \\coloneqq \\text{span}(\\{v_t^j - v_t^0 : j = 1, \\dots, N_i - 1\\})" + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "content": ". By leveraging the span of conceptual space, we enable the model to edit not just predefined attributes, but also a broader range of related attributes. For example, by defining a structured representation space with \"blue,\" \"green,\" and \"yellow,\" our model can generalize to other colors in the spectrum, such as \"red\". Here, " + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "inline_equation", + "content": "v_t^j = \\epsilon_U(x_t, t, \\varepsilon_j)" + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "content": ", each " + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\varepsilon_j = \\psi(p_j^i)" + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "content": " is derived from the pre-trained CLIP text encoder " + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\psi(\\cdot)" + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "content": ". Next, to compute the projection matrix " + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "content": "-space within this representation space, we start by constructing matrix " + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "inline_equation", + "content": "S^i" + }, + { + "bbox": [ + 55, + 387, + 295, + 533 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 540, + 295, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 540, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 107, + 540, + 295, + 556 + ], + "type": "interline_equation", + "content": "S _ {t} ^ {i} := \\left[ v _ {t} ^ {1} - v _ {t} ^ {0}, \\dots , v _ {t} ^ {N _ {i} - 1} - v _ {t} ^ {0} \\right]. \\tag {2}", + "image_path": "43fe916ce6996bc77d2d55c2e4df73a4917760a771568f28e20a60fb98818aad.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "text", + "content": "In addition, to align changes in the representation space with the semantic structure of the concept space, we first apply singular value decomposition (SVD) on " + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "inline_equation", + "content": "S_{t}^{i}" + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "inline_equation", + "content": "SVD(S_{t}^{i}) = V_{t}^{i}\\Sigma_{t}^{i}U_{t}^{iT}" + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "text", + "content": ", to identify the principal orthogonal semantic directions. We then select the top " + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "text", + "content": " components of " + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "inline_equation", + "content": "V_{t}^{i}" + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "text", + "content": " to capture the most significant semantic variations (for simplicity, we denote this matrix as " + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "inline_equation", + "content": "V_{t}^{i}" + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "text", + "content": "), and construct the projection matrix " + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "inline_equation", + "content": "M_{t}^{i} = V_{t}^{i}V_{t}^{iT}" + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "text", + "content": ", which projects the concept vector into " + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C_t^i}" + }, + { + "bbox": [ + 55, + 565, + 295, + 700 + ], + "type": "text", + "content": " and ensures that the changes in the representation space are corrected and aligned with the key semantic directions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Finally, to derive the editing direction, we calculate the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "score difference between source representation " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "s_{src}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and target representation " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "s_{tar}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\Delta s_t = s_{tar} - s_{src}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". We then project " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\Delta s_t" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " onto the spanned concept space using " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "M_t^i" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\Delta C_t^i = \\Delta s_t M_t^i" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". This projection ensures the editing operation reflects the relevant semantic changes in the concept space. The modified score can be achieved by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 383, + 152, + 553, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 152, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 383, + 152, + 553, + 167 + ], + "type": "interline_equation", + "content": "s = s _ {\\emptyset} + \\lambda \\left(s _ {n e w} ^ {i} - s _ {\\emptyset}\\right), \\tag {3}", + "image_path": "9b418423c9fe1aa2360c88be55b824d91eccbbeea3f7e4d6f48715f245c534af.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "text", + "content": " is guidance scale, " + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "inline_equation", + "content": "s_{\\emptyset}" + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "text", + "content": " denotes the unconditional score, is given by " + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "inline_equation", + "content": "s_{\\emptyset} = \\epsilon_U(x_t,t,\\varepsilon_{\\emptyset})" + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\emptyset} = \\psi (^{\\prime \\prime})" + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "text", + "content": " In practice, we noticed that a few prompts are sufficient to identify the concept direction, which is attributed to the generalization ability of the span space " + }, + { + "bbox": [ + 313, + 174, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{C_t^i}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 247, + 504, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 247, + 504, + 259 + ], + "spans": [ + { + "bbox": [ + 315, + 247, + 504, + 259 + ], + "type": "text", + "content": "Algorithm 1 Compositional Concept Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "spans": [ + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "text", + "content": "Require: Diffusion model " + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "inline_equation", + "content": "\\epsilon_U(x_t,t,\\varepsilon)" + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "text", + "content": ", pre-trained CLIP text encoder " + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "text", + "content": ", guidance scale " + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "text", + "content": ", concept weight " + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "inline_equation", + "content": "\\{w_{1},\\dots,w_{m}\\}" + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "text", + "content": ", covariance matrix " + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "inline_equation", + "content": "\\sigma_t^2 I" + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "text", + "content": ", empty prompt \" \", source prompt " + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "text", + "content": ", target prompt " + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "inline_equation", + "content": "P^{\\prime}" + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "text", + "content": ", prompts for build concept space " + }, + { + "bbox": [ + 315, + 260, + 553, + 321 + ], + "type": "inline_equation", + "content": "P_{C^i} = \\{p_0^i,\\ldots ,p_j^i,\\ldots ,p_{N_i - 1}^i\\}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 363, + 441, + 495, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 441, + 495, + 452 + ], + "spans": [ + { + "bbox": [ + 363, + 441, + 495, + 452 + ], + "type": "interline_equation", + "content": "\\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {2}\\right) - \\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {0}\\right), \\dots ,", + "image_path": "372881408b00c998d4f37afca445dd45a70f536581157df3c3992844426e7f4f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 363, + 452, + 494, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 452, + 494, + 464 + ], + "spans": [ + { + "bbox": [ + 363, + 452, + 494, + 464 + ], + "type": "interline_equation", + "content": "\\left. \\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {N _ {i} - 1}\\right) - \\epsilon_ {U} \\left(x _ {t}, t, \\varepsilon_ {0}\\right) \\right]", + "image_path": "ecf4ee010815c5602658d9f759e10b9a0bd778871c7525c412016c33e7a32417.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 363, + 475, + 458, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 475, + 458, + 490 + ], + "spans": [ + { + "bbox": [ + 363, + 475, + 458, + 490 + ], + "type": "interline_equation", + "content": "S V D \\left(S _ {t} ^ {i}\\right) = V _ {t} ^ {i} \\Sigma_ {t} ^ {i} U _ {t} ^ {i T}", + "image_path": "e5b54ef5a987449d8df072a0e09ca775a22776ae9166efbcf6fe70a4e757b555.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 321, + 553, + 596 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 321, + 321, + 457, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 321, + 457, + 332 + ], + "spans": [ + { + "bbox": [ + 321, + 321, + 457, + 332 + ], + "type": "text", + "content": "1: Initialize sample " + }, + { + "bbox": [ + 321, + 321, + 457, + 332 + ], + "type": "inline_equation", + "content": "x_{t} \\sim \\mathcal{N}(0, I)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 332, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 332, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 321, + 332, + 553, + 344 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 321, + 332, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\emptyset}, \\varepsilon_{src}, \\varepsilon_{tar} = \\psi(\"\"), \\psi(P), \\psi(P')" + }, + { + "bbox": [ + 321, + 332, + 553, + 344 + ], + "type": "text", + "content": " # Text embedding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 344, + 487, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 344, + 487, + 357 + ], + "spans": [ + { + "bbox": [ + 321, + 344, + 487, + 357 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 344, + 487, + 357 + ], + "type": "inline_equation", + "content": "\\varepsilon_{j} = \\psi (p_{j}^{i})" + }, + { + "bbox": [ + 321, + 344, + 487, + 357 + ], + "type": "text", + "content": " # Concept text embedding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 357, + 407, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 357, + 407, + 367 + ], + "spans": [ + { + "bbox": [ + 321, + 357, + 407, + 367 + ], + "type": "text", + "content": "4: for " + }, + { + "bbox": [ + 321, + 357, + 407, + 367 + ], + "type": "inline_equation", + "content": "t = T, \\dots, 1" + }, + { + "bbox": [ + 321, + 357, + 407, + 367 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 369, + 514, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 369, + 514, + 380 + ], + "spans": [ + { + "bbox": [ + 321, + 369, + 514, + 380 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 321, + 369, + 514, + 380 + ], + "type": "inline_equation", + "content": "s_{\\emptyset}\\gets \\epsilon_U(x_t,t,\\varepsilon_{\\emptyset})" + }, + { + "bbox": [ + 321, + 369, + 514, + 380 + ], + "type": "text", + "content": " # Unconditional score" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 380, + 553, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 380, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 321, + 380, + 553, + 403 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 321, + 380, + 553, + 403 + ], + "type": "inline_equation", + "content": "s_{src}, s_{tar} \\gets \\epsilon_U(x_t, t, \\varepsilon_{src}), \\epsilon_U(x_t, t, \\varepsilon_{tar})" + }, + { + "bbox": [ + 321, + 380, + 553, + 403 + ], + "type": "text", + "content": " # Conditional score" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 404, + 499, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 404, + 499, + 415 + ], + "spans": [ + { + "bbox": [ + 321, + 404, + 499, + 415 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 321, + 404, + 499, + 415 + ], + "type": "inline_equation", + "content": "\\Delta s_{t} = s_{tar} - s_{src}" + }, + { + "bbox": [ + 321, + 404, + 499, + 415 + ], + "type": "text", + "content": " # Score difference" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 416, + 489, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 416, + 489, + 428 + ], + "spans": [ + { + "bbox": [ + 321, + 416, + 489, + 428 + ], + "type": "text", + "content": "8: for " + }, + { + "bbox": [ + 321, + 416, + 489, + 428 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots,m" + }, + { + "bbox": [ + 321, + 416, + 489, + 428 + ], + "type": "text", + "content": " do # m concepts" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 428, + 506, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 428, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 321, + 428, + 506, + 440 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 321, + 428, + 506, + 440 + ], + "type": "inline_equation", + "content": "S_{t}^{i}\\gets \\left[\\epsilon_{U}(x_{t},t,\\varepsilon_{1}) - \\epsilon_{U}(x_{t},t,\\varepsilon_{0}),\\right." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 464, + 549, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 464, + 549, + 475 + ], + "spans": [ + { + "bbox": [ + 318, + 464, + 549, + 475 + ], + "type": "text", + "content": "10: Compute the top- " + }, + { + "bbox": [ + 318, + 464, + 549, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 318, + 464, + 549, + 475 + ], + "type": "text", + "content": " left singular vectors " + }, + { + "bbox": [ + 318, + 464, + 549, + 475 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 318, + 464, + 549, + 475 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 490, + 507, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 490, + 507, + 502 + ], + "spans": [ + { + "bbox": [ + 318, + 490, + 507, + 502 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 318, + 490, + 507, + 502 + ], + "type": "inline_equation", + "content": "M_{t}^{i} = V_{t}^{i}(V_{t}^{i})^{T}" + }, + { + "bbox": [ + 318, + 490, + 507, + 502 + ], + "type": "text", + "content": " # Projection matrix" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 502, + 504, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 502, + 504, + 514 + ], + "spans": [ + { + "bbox": [ + 318, + 502, + 504, + 514 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 318, + 502, + 504, + 514 + ], + "type": "inline_equation", + "content": "\\Delta C_{t}^{i} = M_{t}^{i}\\Delta s_{t}\\#" + }, + { + "bbox": [ + 318, + 502, + 504, + 514 + ], + "type": "text", + "content": " Editing direction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 514, + 380, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 514, + 380, + 523 + ], + "spans": [ + { + "bbox": [ + 318, + 514, + 380, + 523 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 525, + 553, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 525, + 553, + 549 + ], + "spans": [ + { + "bbox": [ + 318, + 525, + 553, + 549 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 318, + 525, + 553, + 549 + ], + "type": "inline_equation", + "content": "s \\gets s_{\\emptyset} + \\lambda((s_{src} + w_i \\Delta C_t^i) - s_{\\emptyset})" + }, + { + "bbox": [ + 318, + 525, + 553, + 549 + ], + "type": "text", + "content": " # For single-concept editing" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 549, + 553, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 549, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 318, + 549, + 553, + 572 + ], + "type": "text", + "content": "15: " + }, + { + "bbox": [ + 318, + 549, + 553, + 572 + ], + "type": "inline_equation", + "content": "s \\gets s_{\\emptyset} + \\lambda \\sum_{i=1}^{m} \\left( (s_{src} + w_i \\Delta C_t^i) - s_{\\emptyset} \\right) \\# \\text{For multi-concept editing}" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 572, + 444, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 572, + 444, + 585 + ], + "spans": [ + { + "bbox": [ + 318, + 572, + 444, + 585 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 318, + 572, + 444, + 585 + ], + "type": "inline_equation", + "content": "x_{t - 1}\\sim \\mathcal{N}(x_t - s,\\sigma_t^2 I)" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 586, + 364, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 586, + 364, + 596 + ], + "spans": [ + { + "bbox": [ + 317, + 586, + 364, + 596 + ], + "type": "text", + "content": "17: end for" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": "Concept Composition Strategy. Now, our method enables single-concept editing for fashion images, but many scenarios demand multi-concept compositional editing. A straightforward way to achieve this is to create a multi-concept space using a set of concept prompts. While feasible, this approach limits fine-grained control over individual concepts, such as adjusting shades of color in specific items. Alternatively, sequential single-concept edits can be applied, but this accumulates errors and is time-consuming." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28729" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 547, + 369 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 547, + 369 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 547, + 369 + ], + "type": "image", + "image_path": "59c0fb811e2553c4ff9cd744d56b07a62f20d2692dc3d000c3ed1bf4f02703a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 376, + 465, + 388 + ], + "lines": [ + { + "bbox": [ + 144, + 376, + 465, + 388 + ], + "spans": [ + { + "bbox": [ + 144, + 376, + 465, + 388 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison with the state-of-art methods in single concept editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 398, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 447 + ], + "type": "text", + "content": "To overcome these issues, we propose a controllable multi-concept editing strategy. This strategy enables precise, simultaneous adjustments of multiple concepts by weighting each concept's direction vector, which is shown as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 451, + 295, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 451, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 113, + 451, + 295, + 482 + ], + "type": "interline_equation", + "content": "s _ {m} = s _ {\\emptyset} + \\lambda \\sum_ {i = 1} ^ {m} \\left(s _ {n e w} ^ {i} - s _ {\\emptyset}\\right), \\tag {4}", + "image_path": "f8f3cba7d3a8858938213ebe31f92186d5ba1280a7891e28eeb827dcbd427b1f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "inline_equation", + "content": "s_{new}^{i} = s_{src} + w_{i}\\Delta C_{t}^{i}" + }, + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "text", + "content": ", we can control the manipulation strength of each concept. The whole algorithm is shown in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 532, + 137, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 137, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 137, + 545 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 176, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 176, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 176, + 565 + ], + "type": "text", + "content": "3.1. Experimental Setups" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. In our work, all experiments are conducted on a single NVIDIA RTX 4090 GPU with 24 GB of memory. Following the prior work [3, 4, 26], we employ the official pre-trained Stable Diffusion v2.1-base model as our foundational model, downsampling all images to a resolution of " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " pixels for consistency across experiments. For the counterfactual abduction phase, We fine-tune the model on the single image using LoRA with a rank of 512 for 100 epochs, employing the Adam optimizer with a learning rate of 1e-4. To conserve memory, we adopt the mixed precision [17] and gradient accumulation strategies, setting both the accumulation step count and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "text", + "content": "batch size to 1. During the inference phase, we utilize the DDIM sampling with 80 steps and a classifier-free guidance scale of " + }, + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "inline_equation", + "content": "\\lambda = 4" + }, + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "text", + "content": ". Additionally, we explore various values for the concept weight " + }, + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "text", + "content": " within the range [1, 5] to control the editing strength, where " + }, + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 398, + 555, + 458 + ], + "type": "text", + "content": "-th edited concept." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 458, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 613 + ], + "type": "text", + "content": "Dataset. To comprehensively evaluate the performance of our T-FIT in fashion image editing, we follow [14, 26, 40] by collecting a diverse set of fashion images from Unsplash1, including both full-body and half-body images. The selected images comprise various clothing types, such as long dresses, skirts, T-shirts, shorts, pants, and strapless tops. Our evaluation encompasses a diverse set of editing concepts, which include specific attributes (e.g., gender, color, patterns, fabric texture) as well as broader elements such as clothing type, garment length, and sleeve style. Each image in the experiment is annotated with " + }, + { + "bbox": [ + 313, + 458, + 555, + 613 + ], + "type": "inline_equation", + "content": "N_{i} + 2" + }, + { + "bbox": [ + 313, + 458, + 555, + 613 + ], + "type": "text", + "content": " prompts, consisting of one prompt for the original image and " + }, + { + "bbox": [ + 313, + 458, + 555, + 613 + ], + "type": "inline_equation", + "content": "N_{i} + 1" + }, + { + "bbox": [ + 313, + 458, + 555, + 613 + ], + "type": "text", + "content": " prompts for target images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 614, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 555, + 661 + ], + "type": "text", + "content": "Baselines. We conduct comparisons with existing representative diffusion-based image editing methods, including LEDs [29], LEDs++ [2], TexFit [32], Guide-and-Rescale (G & R) [28], UltraEdit [41], and Turboedit [6]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 662, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 555, + 685 + ], + "type": "text", + "content": "Evaluation Metrics. To comprehensively evaluate the performance of our T-FIT, we conduct assessments from both" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 432, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 432, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 432, + 712 + ], + "type": "text", + "content": "1https://unsplash.com/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28730" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 547, + 355 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 547, + 355 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 547, + 355 + ], + "type": "image", + "image_path": "468ca461f1b15c07a77790f11c9eaee95fd05eeb64efe41988d837d051d19e0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 361, + 470, + 374 + ], + "lines": [ + { + "bbox": [ + 138, + 361, + 470, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 470, + 374 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with the state-of-art methods in multiple concept editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 384, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 295, + 517 + ], + "type": "text", + "content": "objective and subjective perspectives. Objectively, we follow the previous works [3, 8, 9] and use metrics in CLIP embedding space: CLIP image similarity (CLIP-I) [21] to measure identity preservation via calculate the cosine similarity between the edited image and the original image, and CLIP text-image direction similarity (CLIP-D) [7] to evaluate the correspondence between image changes and text changes. Subjectively, we qualitatively evaluate the fidelity of the edited images and their consistency with the target prompts, supplemented by human preference studies involving recruited volunteers from different disciplines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 526, + 185, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 185, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 185, + 539 + ], + "type": "text", + "content": "3.2. Qualitative Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 569 + ], + "type": "text", + "content": "We show some qualitative experimental results in Fig. 3-4, from our experiments, we observe the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": "For single-concept fashion image editing, 1) several methods, such as LEDs++ [2], introduce unintended changes to non-target areas, possibly due to spurious correlation embedded in the pre-trained models from their training datasets. These correlations cause inaccurate concept associations, leading to changes outside the intended editing area. For instance, in Fig. 3, line 2, LEDs++ unexpectedly replaces the shoelaces with pink ones. Similarly, in Fig. 3, line 5, UltraEdit [41] transforms a squared neckline into a V-neckline, Turbo-edit [6] alters the dress length. 2) G & R [28], and UltraEdit often confuse different objects. For instance, when attempting to change the color of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 384, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 708 + ], + "type": "text", + "content": "pants to red (Fig. 3, line 2), these methods inadvertently modify unrelated elements, such as the color of the shirt or hair. 3) TexFit's effectiveness is significantly hampered by its reliance on the accuracy of the recognized mask, which restricts its applicability in fashion image editing. As shown in the last column of Fig. 3, TexFit [32] either alters only localized information in the target editing area or excessively expands the intended modification area. This results in unnatural edits that disrupt the overall coherence of the image. For multi-concept fashion image editing, 1) UltraEdit [41], TexFit [32] often overlooks certain concepts in multiple-concept editing scenarios. For example, in Fig. 4, line 1, UltraEdit remains in top with short sleeves rather than long sleeves; in Fig. 4, line 2, the trousers edit is ignored. 2) G & R [28], TurboEdit [6], and LEDITS++ [2], demonstrate commendable performance in adhering to user instructions. However, the edited images can exhibit issues such as alterations to human poses or discontinuities in the edited results. For example, when modifying sleeve lengths, the transitions in the sleeve areas may appear inconsistent or fragmented (Fig. 4, line 1). These problems highlight the challenges in maintaining coherence and integrity in the edited images, despite following the provided directives closely. 3) Several methods such as Turbo-edit [6] and UltraEdit [41] may over-edit, making it difficult to retain the details of the original image. For example, when modifying pants into a skirt (Fig. 4, line 3), the edited" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28731" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "image may change the length of the skirt as well as other aspects such as the color and fabric of the top along with the skirt. This phenomenon may be attributed to the spurious correlation that were learned by the pre-trained model from the training dataset. Compared to the existing methods, our method excels in fashion image editing, and it can directly control precise multi-concepts from text prompts without additional input, such as masks or poses. The edited image retains non-target areas, closely following the target text while maintaining the consistency of the image. In addition, our method allows continuous editing adjustments by adjusting the concept weight (Fig. 5), which provides fine control over the editing strength. This results in high-quality editing that is superior to the baseline model, especially for multi-concept fashion image editing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 191, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 191, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 191, + 275 + ], + "type": "text", + "content": "3.3.Quantitative Evaluation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 283, + 295, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 307 + ], + "type": "text", + "content": "We quantitatively evaluate our T-FIT against baselines using both automatic metrics and human evaluations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 307, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 534 + ], + "type": "text", + "content": "Automatic Metric Comparisons. We summarize the experimental findings in Tab. 1. Our method achieves state-of-the-art results for single- and multi-concept fashion image editing in CLIP-D, which demonstrates a significant advantage in terms of the accuracy and fidelity of our editing results in following the text description. In terms of the CLIP-I metric, our method demonstrates superior performance compared to most competing approaches. Although it slightly underperforms Turbo-edit [6] and TexFit [32] on CLIP-I, this is due to specific limitations in these methods during multi-concept fashion image editing. Turbo-edit often omits certain concepts, resulting in edited images that closely resemble the source, thus boosting its CLIP-I score. TexFit, on the other hand, struggles with precise edit masking, often restricting changes to small areas and failing in non-rigid edits (e.g., changing long sleeves to shorts or pants to shorts), leaving much of the image unaltered. Overall, our method shows superior performance on objective metrics for both multi- and single-concept editing tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": "Human Preference Study. In this section, we quantitatively evaluate our method with an extensive human perceptual evaluation study. We first collect a diverse set of 30 fashion images and their corresponding text descriptions, covering a range of clothing types (e.g., long sleeves, short sleeves, dresses, pants, shorts), patterns (e.g., solid, stripes, prints), and materials (e.g., cotton, denim, chiffon). Then, we invite 36 volunteers to participate in the evaluation. Each volunteer views the original fashion image, the original description text, the target description text, and the edited images generated by 6 different methods, including ours. The volunteers rate the edited images based on two criteria: (1) how well the edits align with the concepts specified in the target text and (2) the consistency of the non-edited areas with the original image. Ratings range from" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 102, + 556, + 216 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons. The best results are bolded and the second-best results are underlined, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 102, + 556, + 216 + ], + "lines": [ + { + "bbox": [ + 315, + 102, + 556, + 216 + ], + "spans": [ + { + "bbox": [ + 315, + 102, + 556, + 216 + ], + "type": "table", + "html": "
Single conceptMultiple concepts
MethodCLIP-DCLIP-ICLIP-DCLIP-I
OUR0.14330.91600.11520.9259
G & R0.09050.91160.05390.9079
Turbo-edit0.08210.91010.09670.9355
UltraEdit0.07710.90560.05330.8974
LEDITS++0.09540.91600.05840.9115
Texfit0.03460.95000.04630.9562
", + "image_path": "2308da61f16b418b1791d99afe167b9a4eda2df8223d459c781cccccc0608393.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 325, + 220, + 545, + 396 + ], + "blocks": [ + { + "bbox": [ + 325, + 220, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 325, + 220, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 325, + 220, + 545, + 396 + ], + "type": "image", + "image_path": "8a5555857ecd9e4ba32152568dc680ae5b212dabc974743126e92e9b594aa417.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 408, + 555, + 442 + ], + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 442 + ], + "type": "text", + "content": "Figure 5. Human preference study. Our method outperforms baseline models in both text- and image alignment, demonstrating significant advantages in fashion image editing." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 448, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 533 + ], + "type": "text", + "content": "1 to 8, with 1 representing the best performance and 8 the poorest. We calculate the average ranks provided by each participant, and the final results are displayed in Fig. 5. Our findings indicate that our method surpasses baseline methods on both criteria, demonstrating enhanced fidelity in executing intended edits while maintaining consistency across unaltered image areas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 540, + 408, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 408, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 408, + 553 + ], + "type": "text", + "content": "3.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "content": "Next, we turn to an ablation study where we analyze the effect of the different components inherent in our approach. Specifically, we consider: (1) the effect of substituting the CA with null-text inversion [18], (2) the effect of removing CLM, (3) the effects of varying annealing parameter " + }, + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "content": ", and (4) the influence of different weights on the editing results of concepts. The visual results are provided in Fig. 6-8." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "From these figures, it is evident that when T-FIT replaces the CA with the null-text inversion module, the edited images exhibit a noticeable spurious correlation. For instance, in Fig. 7, when the concept of \"gender\" is applied, the collar of the clothing incorrectly changes from a \"stand-up collar\" to a \"round collar\". On the other hand, T-FIT is equipped" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28732" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 291, + 205 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 291, + 205 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 291, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 291, + 205 + ], + "type": "image", + "image_path": "fcc5e75fd19165d6cd944b3deff0ab9af37a5909ee63ec4d8ba0e1e22017cefd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 295, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 266 + ], + "type": "text", + "content": "Figure 6. Qualitative ablation results. The last image shows the result of T-FIT, and the other images show the result of T-FIT without CLM and adjusting annealing parameter " + }, + { + "bbox": [ + 55, + 209, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 209, + 295, + 266 + ], + "type": "text", + "content": ". Removing CLM significantly reduces the model's editability. Adjusting " + }, + { + "bbox": [ + 55, + 209, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 209, + 295, + 266 + ], + "type": "text", + "content": " still makes balancing editability and fidelity challenging." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 270, + 289, + 420 + ], + "blocks": [ + { + "bbox": [ + 59, + 270, + 289, + 420 + ], + "lines": [ + { + "bbox": [ + 59, + 270, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 59, + 270, + 289, + 420 + ], + "type": "image", + "image_path": "ede48eaa7f596cb4374c1a46ee991b2866b2fd2d0703a93755a5239dfe733318.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 429, + 295, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 473 + ], + "type": "text", + "content": "Figure 7. Qualitative ablation results. With CA, our model captures essential visual content from the original image, mitigating the impact of spurious correlation in pre-trained data. CLM ensures that no editing concepts are overlooked during modifications." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "solely with CA, it effectively preserves the visual details of the source image and mitigates issues arising from spurious correlation. However, this setup offers limited flexibility for editing, even if we adopt varying annealing parameters " + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": " (see Fig. 6). This suggests that while CA effectively addresses the spurious correlation issues arising from the pretrained model, it is limited in its editability. When CLM is added, T-FIT's editability improves. This is attributable to the CLM's capability to identify disentangled concepts within the representation space, which facilitates more flexible and precise editing. Furthermore, as illustrated in Fig. 8, T-FIT addresses the issues of concept omission and confusion through the multi-concept composite strategy, while also providing fine control over the intensity of each concept by adjusting individual concept weights. In our work, the concept weights can be manually adjusted in the range [1-5], with higher values indicating greater editing intensity on that concept and vice versa. This offers greater control and flexibility for fashion image editing of our method." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 550, + 384 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 550, + 384 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 550, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 550, + 384 + ], + "type": "image", + "image_path": "3e4059db8be4fc614a6afa59b0ea500bdd536a9348241d5e6e5067368f522d19.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 390, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 424 + ], + "type": "text", + "content": "Figure 8. Experimental comparison of varying concept weights (from 1 to 5) on single- and multi-concept fashion image editing. [C1], [C2], and [C3] mean different editing concepts, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 434, + 388, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 388, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 388, + 445 + ], + "type": "text", + "content": "4. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 454, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 634 + ], + "type": "text", + "content": "In this paper, we introduce T-FIT, a novel framework for fashion image editing. Our approach, which uses a single fashion image and simple text description, supports both single- and multi-concept image editing tasks. By incorporating counterfactual abduction and a concept learning module, T-FIT effectively mitigates issues of spurious correlation in the pre-trained model, enabling efficient, flexible, and accurate text-driven image editing. Furthermore, T-FIT allows fine-grained editing of complex fashion concepts by adjusting concept weights. Extensive experiments confirm the effectiveness of T-FIT and its components, demonstrating significant advantages over several existing methods. Our future work will focus on enhancing editing accuracy by incorporating visual information (e.g., style images) to learn richer visual concept representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 645, + 422, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 422, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 422, + 658 + ], + "type": "text", + "content": "5. Acknowledgment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "This work was supported by grants from the National Key R & D Program of China (grant no. 2022YFB3303302), and the National Natural Science Foundation of China (grant nos. 62377040, 62477004, 62207007, 623B2002)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28733" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Alberto Baldrati, Davide Morelli, Giuseppe Cartella, Marcela Cornia, Marco Bertini, and Rita Cucchiara. Multimodal garment designer: Human-centric latent diffusion models for fashion image editing. In ICCV, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. LEDs++: Limitless image editing using text-to-image models. In CVPR, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 212 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 257 + ], + "type": "text", + "content": "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "text", + "content": "[5] Zhikang Chen, Min Zhang, Sen Cui, Haoxuan Li, Gang Niu, Mingming Gong, Changshui Zhang, and Kun Zhang. Neural collapse inspired feature alignment for out-of-distribution generalization. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 295, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 344 + ], + "type": "text", + "content": "[6] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv:2408.00735, 2024. 1, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 390 + ], + "type": "text", + "content": "[7] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. StyleGAN-NADA: CLIP-guided domain adaptation of image generators. ACM Transactions on Graphics, 41(4):1-13, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 445 + ], + "type": "text", + "content": "[8] Julia Guerrero-Viu, Milos Hasan, Arthur Roullier, Midhun Harikumar, Yiwei Hu, Paul Guerrero, Diego Gutierrez, Bellen Masia, and Valentin Deschaintre. Texsliders: Diffusion-based texture editing in clip space. In SIGGRAPH, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 479 + ], + "type": "text", + "content": "[9] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In CVPR, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 480, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 295, + 523 + ], + "type": "text", + "content": "[10] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Multiedits: Simultaneous multi-aspect editing with text-to-image diffusion models. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 525, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 556 + ], + "type": "text", + "content": "[11] Shanshan Huang, Haoxuan Li, Qingsong Li, Chunyuan Zheng, and Li Liu. Pareto invariant representation learning for multimedia recommendation. In ACM MM, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "text", + "content": "[12] Shanshan Huang, Qingsong Li, Jun Liao, Shu Wang, Li Liu, and Lian Li. Controllable image synthesis methods, applications and challenges: a comprehensive survey. Artificial Intelligence Review, 57(12):336, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "type": "text", + "content": "[13] Shanshan Huang, Lei Wang, Jun Liao, and Li Liu. Multi-attentional causal intervention networks for medical image diagnosis. Knowledge-Based Systems, 299:111993, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 637, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 295, + 678 + ], + "type": "text", + "content": "[14] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 1, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[15] Gwanhyeong Koo, Sunjae Yoon, Ji Woo Hong, and Chang D. Yoo. Flexiedit: Frequency-aware latent refinement for enhanced non-rigid editing. In ECCV, 2024. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[16] Meng Li and Haochen Sui. Causal recommendation via machine unlearning with a few unbiased data. In AAAI Workshop on AICT, 2025. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[17] Paulius Micikevicius, Sharan Narang, Jonah Alben, Greg Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. In ICLR, 2018. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 154, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 154, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 154, + 553, + 186 + ], + "type": "text", + "content": "[18] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 189, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 553, + 231 + ], + "type": "text", + "content": "[19] Martin Pernus, Clinton Fookes, Vitomir Štruc, and Simon Dobrišek. FICE: Text-conditioned fashion-image editing with guided gan inversion. Pattern Recognition, 158:111022, 2025. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 235, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 235, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 235, + 553, + 277 + ], + "type": "text", + "content": "[20] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion Autoencoders: Toward a meaningful and decodable representation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 280, + 553, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 344 + ], + "type": "text", + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 348, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 348, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 553, + 380 + ], + "type": "text", + "content": "[22] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "text", + "content": "[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 417, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 471 + ], + "type": "text", + "content": "[24] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 474, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 474, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 474, + 553, + 517 + ], + "type": "text", + "content": "[25] Pedro Sanchez, Jeremy P Voisey, Tian Xia, Hannah I Watson, Alison Q O'Neil, and Sotirios A Tsaftaris. Causal machine learning for healthcare and precision medicine. Royal Society Open Science, 9(8):220638, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 520, + 553, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 562 + ], + "type": "text", + "content": "[26] Xue Song, Jiequan Cui, Hanwang Zhang, Jingjing Chen, Richang Hong, and Yu-Gang Jiang. Doubly abductive counterfactual inference for text-based image editing. In CVPR, 2024. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 565, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 553, + 597 + ], + "type": "text", + "content": "[27] Adam Stein, Aaditya Naik, Yinjun Wu, Mayur Naik, and Eric Wong. Towards compositionality in concept learning. In ICML, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "text", + "content": "[28] Vadim Titov, Madina Khalmatova, Alexandra Ivanova, Dmitry Vetrov, and Aibek Alanov. Guide-and-Rescale: Self-guidance mechanism for effective tuning-free real image editing. In ECCV, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 646, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 677 + ], + "type": "text", + "content": "[29] Linoy Tsaban and Apolinário Passos. LEDs: Real image editing with ddpm inversion and semantic guidance. arXiv:2307.00522, 2023. 1, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[30] Luozhou Wang, Shuai Yang, Shu Liu, and Ying-cong Chen. Not all steps are created equal: Selective diffusion distillation for image manipulation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28734" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 521 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[31] Lei Wang, Shanshan Huang, Shu Wang, Jun Liao, Tingpeng Li, and Li Liu. A survey of causal discovery based on functional causal model. Engineering Applications of Artificial Intelligence, 133:108258, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 148 + ], + "type": "text", + "content": "[32] Tongxin Wang and Mang Ye. TexFit: Text-driven fashion image editing with diffusion models. In AAAI, 2024. 1, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[33] Xiaolong Wang, Zhi-Qi Cheng, Jue Wang, and Xiaojiang Peng. DPDEdit: Detail-preserved diffusion models for multimodal fashion image editing. arXiv:2409.01086, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[34] Yingheng Wang, Yair Schiff, Aaron Gokaslan, Weishen Pan, Fei Wang, Christopher De Sa, and Volodymyr Kuleshov. Infodiffusion: Representation learning using information maximizing diffusion models. In ICML, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "text", + "content": "[35] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for (score-based) text-controlled generative models. NeurIPS, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[36] Peng Wu, Haoxuan Li, Yuhao Deng, Wenjie Hu, Quanyu Dai, Zhenhua Dong, Jie Sun, Rui Zhang, and Xiao-Hua Zhou. On the opportunity of causal learning in recommendation systems: Foundation, estimation, prediction and challenges. In IJCAI, 2022. Survey Track. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "text", + "content": "[37] Xindi Wu, Dingli Yu, Yangsibo Huang, Olga Russakovsky, and Sanjeev Arora. Conceptmix: A compositional image generation benchmark with controllable difficulty. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[38] Yuntian Wu, Yuntian Yang, Jiabao Sean Xiao, Chuan Zhou, Haochen Sui, and Haoxuan Li. Invariant spatiotemporal representation learning for cross-patient seizure classification. In NeurIPS Workshop on NeuroAI, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 410, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 442 + ], + "type": "text", + "content": "[39] Min Zhang, Haoxuan Li, Fei Wu, and Kun Kuang. Metacoco: A new few-shot classification benchmark with spurious correlation. In ICLR, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "text", + "content": "[40] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In CVPR, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[41] Haozhe Zhao, Xiaojian (Shawn) Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. In NeurIPS, 2024. 1, 5, 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28735" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_content_list.json b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..751aa8d5446dd8d5c76deefc51c85c9165a0cafc --- /dev/null +++ b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_content_list.json @@ -0,0 +1,1703 @@ +[ + { + "type": "text", + "text": "Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding", + "text_level": 1, + "bbox": [ + 153, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenxuan Guo $^{1*}$ Xiuwei Xu $^{1*}$ Ziwei Wang $^{2}$ Jianjiang Feng $^{1\\dagger}$ Jie Zhou $^{1}$ Jiwen Lu $^{1\\dagger}$ Tsinghua University $^{2}$ Nanyang Technological University", + "bbox": [ + 125, + 178, + 870, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\{gwx22,xxw21\\} @mails.tsinghua.edu.cn ziwei.wang@ntu.edu.sg$", + "bbox": [ + 238, + 218, + 756, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jfeng,jzhou,lujiwen}@tsinghua.edu.cn", + "bbox": [ + 336, + 234, + 661, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose an efficient multi-level convolution architecture for 3D visual grounding. Conventional methods are difficult to meet the requirements of real-time inference due to the two-stage or point-based architecture. Inspired by the success of multi-level fully sparse convolutional architecture in 3D object detection, we aim to build a new 3D visual grounding framework following this technical route. However, as in 3D visual grounding task the 3D scene representation should be deeply interacted with text features, sparse convolution-based architecture is inefficient for this interaction due to the large amount of voxel features. To this end, we propose text-guided pruning (TGP) and completion-based addition (CBA) to deeply fuse 3D scene representation and text features in an efficient way by gradual region pruning and target completion. Specifically, TGP iteratively sparsifies the 3D scene representation and thus efficiently interacts the voxel features with text features by cross-attention. To mitigate the affect of pruning on delicate geometric information, CBA adaptively fixes the over-pruned region by voxel completion with negligible computational overhead. Compared with previous single-stage methods, our method achieves top inference speed and surpasses previous fastest method by $100\\%$ FPS. Our method also achieves state-of-the-art accuracy even compared with two-stage methods, with $+1.13$ lead of Acc@0.5 on ScanRefer, and $+2.6$ and $+3.2$ leads on NR3D and SR3D respectively. The code is available at https://github.com/GWxuan/TSP3D.", + "bbox": [ + 89, + 316, + 485, + 741 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 772, + 220, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Incorporating multi-modal information to guide 3D visual perception is a promising direction. In these years, 3D visual grounding (3DVG), also known as 3D instance referencing, has been paid increasing attention as a fundamental multi-modal 3D perception task. The aim of 3DVG is", + "bbox": [ + 89, + 797, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/661c935b6d6e9132a512ca6ed5b2382b6b4416e336bf98bb467a08e4aab6aed4.jpg", + "image_caption": [ + "Figure 1. Comparison of 3DVG methods on ScanRefer dataset [3]. Our TSP3D surpasses existing methods in both accuracy and inference speed, achieving the first efficient 3DVG framework." + ], + "image_footnote": [], + "bbox": [ + 516, + 285, + 901, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to locate an object in the scene with a free-form query description. 3DVG is challenging since it requires understanding of both 3D scene and language description. Recently, with the development of 3D scene perception and vision-language models, 3DVG methods have shown remarkable progress [16, 22]. However, with 3DVG being widely applied in fields like robotics and AR / VR where inference speed is the main bottleneck, how to construct efficient real-time 3DVG model remains a challenging problem.", + "bbox": [ + 509, + 551, + 906, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since the output format of 3DVG is similar with 3D object detection, early 3DVG methods [3, 14, 38, 39] usually adopt a two-stage framework, which first conducts detection to locate all objects in the scene, and then selects the target object by incorporating text information. As there are many similarities between 3D object detection and 3DVG (e.g. both of them need to extract the representation of the 3D scene), there will be much redundant feature computation during the independent adoption of the two models. As a result, two-stage methods are usually hard to handle real-time tasks. To solve this problem, single-stage methods [22, 35] are presented, which generates the bounding box of the target directly from point clouds. This integrated design is more compact and efficient. However,", + "bbox": [ + 509, + 688, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding author.", + "bbox": [ + 107, + 886, + 351, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3666", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "current single-stage 3DVG methods mainly build on point-based architecture [25], where the feature extraction contains time-consuming operations like furthest point sampling and kNN. They also need to aggressively downsample the point features to reduce computational cost, which might hurt the geometric information of small and thin objects [37]. Due to these reasons, current single-stage methods are still far from real-time ( $< 6$ FPS) and their performance is inferior to two-stage methods, as shown in Fig. 1.", + "bbox": [ + 88, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a new single-stage framework for 3DVG based on text-guided sparse voxel pruning, namely TSP3D. Inspired by state-of-the-art 3D object detection methods [29, 37] which achieves both leading accuracy and speed with multi-level sparse convolutional architecture, we build the first sparse single-stage 3DVG network. However, different from 3D object detection, in 3DVG the 3D scene representation should be deeply interacted with text features. Since the count of voxels is very large in sparse convolution-based architecture, deep multi-modal interaction like cross-attention becomes infeasible due to unaffordable computational cost. To this end, we propose text-guided pruning (TGP), which first utilize text information to jointly sparsify the 3D scene representation and enhance the voxel and text features. To mitigate the affect of pruning on delicate geometric information, we further present completion-based addition (CBA) to adaptively fix the over-pruned region with negligible computational overhead. Specifically, TGP prunes the voxel features according to the object distribution. It gradually removes background features and features of irrelevant objects, which generates text-aware voxel features around the target object for accurate bounding box prediction. Since pruning may mistakenly remove the representation of target object, CBA utilizes text features to query a small set of voxel features from the complete backbone features, followed by pruned-aware addition to fix the over-pruned region. We conduct extensive experiments on the popular ScanRefer [3] and ReferIt3D [2] datasets. Compared with previous single-stage methods, TSP3D achieves top inference speed and surpasses previous fastest single-stage method by $100\\%$ FPS. TSP3D also achieves state-of-the-art accuracy even compared with two-stage methods, with $+1.13$ lead of Acc@0.5 on ScanRefer, and $+2.6$ and $+3.2$ leads on NR3D and SR3D respectively.", + "bbox": [ + 91, + 228, + 486, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 109, + 760, + 464, + 773 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, this is the first work exploring sparse convolutional architecture for efficient 3DVG.", + "- To enable efficient feature extraction, we propose text-guided pruning and completion-based addition to sparsify sparse voxels and adaptively fuse multi-level features.", + "- We conduct extensive experiments, and TSP3D outperforms existing methods in both accuracy and speed, demonstrating the superiority of the proposed framework." + ], + "bbox": [ + 88, + 777, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.3D Visual Grounding", + "text_level": 1, + "bbox": [ + 513, + 114, + 714, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D visual grounding aims to locate a target object within a 3D scene based on natural language descriptions [19]. Existing methods are typically categorized into two-stage and single-stage approaches. Two-stage methods follow a detect-then-match paradigm. In the first stage, they independently extract features from the language query using pre-trained language models [7, 9, 24] and predict candidate 3D objects using pre-trained 3D detectors [21, 26] or segmenters [4, 17, 32]. In the second stage, they focus on aligning the vision and text features to identify the target object. Techniques for feature fusion include attention mechanisms with Transformers [13, 40], contrastive learning [1], and graph-based matching [10, 14, 39]. In contrast, single-stage methods integrate object detection and feature extraction, allowing for direct identification of the target object. Methods in this category include guiding keypoint selection using textual features [22], and measuring similarity between words and objects inspired by 2D image-language pre-trained models like GLIP [18], as in BUTDDETR [16]. And methods like EDA [35] and $\\mathrm{G}^3$ -LQ [34] advance single-stage 3D visual grounding by enhancing multimodal feature discriminability through explicit textdecoupling, dense alignment, and semantic-geometric modeling. MCLN [27] uses the 3D referring expression segmentation task to assist 3DVG in improving performance.", + "bbox": [ + 511, + 138, + 906, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing two-stage and single-stage methods generally have high computational costs, hindering real-time applications. Our work aims to address these efficiency challenges by proposing an efficient single-stage method with multi-level sparse convolutional architecture.", + "bbox": [ + 511, + 517, + 906, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-Level Convolutional Architectures", + "text_level": 1, + "bbox": [ + 511, + 604, + 862, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, sparse convolutional architecture has achieved great success in the field of 3D object detection. Built on the voxel-based representation [5, 8, 33] and sparse convolution operation [6, 11, 36], this kind of methods show great efficiency and accuracy when processing scene-level data. GSDN [12] first adopts multi-level sparse convolution with generative feature upsampling in 3D object detection. FCAF3D [29] simplifies the multi-level architecture with anchor-free design, achieving leading accuracy and speed. TR3D [30] further accelerates FCAF3D by removing unnecessary layers and introducing category-aware proposal assignment method. Moreover, DSPDet3D [37] introduces the multi-level architecture to 3D small object detection.", + "bbox": [ + 511, + 627, + 908, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed method draws inspiration from these approaches, utilizing a sparse multi-level architecture with sparse convolutions and an anchor-free design. This allows for efficient processing of 3D data, enabling real-time performance in 3D visual grounding tasks.", + "bbox": [ + 511, + 825, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3667", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 89, + 181, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe our TSP3D for efficient single-stage 3DVG. We first analyze existing pipelines to identify current challenges and motivate our approach (Sec. 3.1). We then introduce the text-guided pruning, which leverages text features to guide feature pruning (Sec. 3.2). To address the potential risk of pruning key information, we propose the completion-based addition for multi-level feature fusion (Sec. 3.3). Finally, we detail the training loss (Sec. 3.4).", + "bbox": [ + 89, + 114, + 483, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture Analysis for 3DVG", + "text_level": 1, + "bbox": [ + 89, + 244, + 372, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Top-performance 3DVG methods [31, 34, 35], are mainly two-stage, which is a serial combination of 3D object detection and 3D object grounding. This separate calls of two approaches result in redundant feature extraction and complex pipeline, thus making the two-stage methods less efficient. To demonstrate the efficiency of existing methods, we conduct a comparison of accuracy and speed among several representative methods on ScanRefer [3], as shown in Fig. 1. It can be seen that two-stage methods struggle in speed (< 3 FPS) due to the additional detection stage. Since 3D visual grounding is usually adopted in practical scenarios that require real-time inference under limited resources, such as embodied robots and VR/AR, the low speed of two-stage methods make them less practical. On the other side, single-stage methods [22], which directly predicts referred bounding box from the observed 3D scene, are more suitable choices due to their streamlined processes. In Fig. 1, it can be observed that single-stage methods are significantly more efficient than their two-stage counterparts.", + "bbox": [ + 91, + 266, + 483, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, existing single-stage methods are mainly built on point-based backbone [25], where the scene representation is extracted with time-consuming operations like furthest point sampling and set abstraction. They also employ large transformer decoder to fuse text and 3D features for several iterations. Therefore, the inference speed of current single-stage methods is still far from real-time ( $< 6$ FPS). The inference speed of specific components in different frameworks is analyzed and discussed in detail in the supplementary material. Inspired by the success of multi-level sparse convolutional architecture in 3D object detection [30], which achieves both leading accuracy and speed, we propose to build the first multi-level convolutional single-stage 3DVG pipeline.", + "bbox": [ + 89, + 553, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TSP3D-B. Here we propose a baseline framework based on sparse convolution, namely TSP3D-B. Following the simple and effective multi-level architecture of FCAF3D [29], TSP3D-B utilizes 3 levels of sparse convolutional blocks for scene representation extraction and bounding box prediction, as shown in Fig. 2 (a). Specifically, the input point clouds $P \\in \\mathbb{R}^{N \\times 6}$ with 6-dim features (3D position and RGB) are first voxelized and then fed into three sequential MinkResBlocks [6], which generates three lev", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "els of voxel features $V_{l}$ ( $l = 1,2,3$ ). With the increase of $l$ , the spatial resolution of $V_{l}$ decreases and the context information increases. Concurrently, the free-form text with $l$ words is encoded by the pre-trained RoBERTa [20] and produce the vanilla text tokens $T \\in \\mathbb{R}^{l \\times d}$ . With the extracted 3D and text representations, we iteratively upsample $V_{3}$ and fuse it with $T$ to generate high-resolution and text-aware scene representation:", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nU _ {l} = U _ {l} ^ {G} + V _ {l}, \\quad U _ {l} ^ {G} = \\operatorname {G e S p C o n v} \\left(U _ {l + 1} ^ {\\prime}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 220, + 906, + 239 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nU _ {l + 1} ^ {\\prime} = \\operatorname {C o n c a t} \\left(U _ {l + 1}, T\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 246, + 906, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $U_{3} = V_{3}$ , GeSpConv means generative sparse convolution [12] with stride 2, which upsamples the voxel features and expands their spatial locations for better bounding box prediction. Concat is voxel-wise feature concatenation by duplicating $T$ . The final upsampled feature map $U_{1}$ is concatenated with $T$ and fed into a convolutional head to predict the objectness scores and regress the 3D bounding box. We select the box with highest objectness score as the grounding result.", + "bbox": [ + 511, + 268, + 906, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 1, TSP3D-B achieves an inference speed of 14.58 FPS, which is significantly faster than previous single-stage methods and demonstrates great potential for real-time 3DVG.", + "bbox": [ + 511, + 405, + 906, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Text-guided Pruning", + "text_level": 1, + "bbox": [ + 511, + 473, + 709, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Though efficient, TSP3D-B exhibits poor performance due to the inadequate interaction between 3D scene representation and text features. Motivated by previous 3DVG methods [16], a simple solution is to replace Concat with cross-modal attention to process voxel and text features, as shown in Fig. 2 (b). However, different from point-based architectures where the scene representation is usually aggressively downsampled, the number of voxels in multi-level convolutional framework is very large1. In practical implementation, we find that the voxels expand almost exponentially with each upsampling layer, leading to a substantial computational burden for the self-attention and cross-attention of scene features. To address this issue, we introduce text-guided pruning (TGP) to construct TSP3D, as illustrated in Fig. 2 (c). The core idea of TGP is to reduce feature amount by pruning redundant voxels and guide the network to gradually focus on the final target based on textual features.", + "bbox": [ + 511, + 494, + 906, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall Architecture. TGP can be regarded as a modified version of cross-modal attention, which reduces the number of voxels before attention operation, thereby lowering computational cost. To minimize the affect of pruning on the final prediction, we propose to prune the scene representation gradually. At higher level where the number of", + "bbox": [ + 511, + 752, + 908, + 843 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Compared to point-based architectures, sparse convolutional framework provides higher resolution and more detailed scene representations, while also offering advantages in inference speed. For detailed statistics, please refer to the supplementary material.", + "bbox": [ + 511, + 851, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3668", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/34dc88865730e7f3ee523802cfd25bf1b85a5d80e96e3034fcb7988d0ce6af56.jpg", + "image_caption": [ + "Figure 2. Illustration of TSP3D. TSP3D bulids on multi-level sparse convolutional architecture. It iteratively upsamples the voxel features with text-guided pruning (TGP), and fuses multi-level features via completion-based addition (CBA). (a) to (d) on the right side illustrate various options for feature upsampling. (a) refers to simple concatenation with text features, which is fast but less accurate. (b) refers to feature interaction through cross-modal attention mechanisms, which is constrained by the large number of voxels. (c) represents our proposed TGP, which first prunes voxel features under textual guidance and thus enables efficient interaction between voxel and text features. (d) shows a simplified version of TGP that removes farthest point sampling and interpolation, combines multi-modal feature interactions into a whole and moves it before pruning." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 529, + 294 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a003d30387be96c848bce3f9f41eacf06c8f47f83080b207f66b18616a8e5fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 93, + 638, + 183 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69e156e52604f54c768f8d61e1521fe2c118da87ec724529954e0af845483d6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 185, + 638, + 295 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/28a30a9d670f1f2884a89d0d799fe6330b8eb72a993725475238135f0171de34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 93, + 763, + 295 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1c6a741f76761a4176f284700f1630ed9a640e23ff31c2818ea88d8de8dcb638.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 93, + 883, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "voxels is not too large yet, TGP prunes less voxels. While at lower level where the number of voxels is significantly increased by upsampling operation, TGP prunes the voxel features more aggressively. The multi-level architecture of TSP3D consists of three levels and includes two feature upsampling operations. Therefore, we correspondingly configure two TGPs with different functions, which are referred as scene-level TGP (level 3 to 2) and target-level TGP (level 2 to 1) respectively. Scene-level TGP aims to distinguish between objects and the background, specifically pruning the voxels on background. Target-level TGP focuses on regions mentioned in the text, intending to preserve the target object and referential objects while removing other regions.", + "bbox": [ + 88, + 417, + 483, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Details of TGP. Since the pruning is relevant to the description, we need to make the voxel features text-aware to predict a proper pruning mask. To reduce the computational cost, we perform farthest point sampling (FPS) on the voxel features to reduce their size while preserving the basic distribution of the scene. Next, we utilize cross-attention to interact with the text features and employ a simple MLP to predict the probability distribution $\\hat{M}$ for retaining each voxel. To prune the features $U_{l}$ , we binarize and interpolate the $\\hat{M}$ to obtain the pruned mask. This process can be expressed as:", + "bbox": [ + 89, + 616, + 483, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU _ {l} ^ {P} = U _ {l} \\odot \\Theta (\\mathcal {I} (\\hat {M}, U _ {l}) - \\sigma) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 794, + 480, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} = \\operatorname {M L P} (\\text {C r o s s A t t} (\\operatorname {F P S} (U _ {l}), \\operatorname {S e l f A t t} (T))) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 827, + 482, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $U_{l}^{P}$ is the pruned features, $\\Theta$ is Heaviside step function, $\\odot$ is matrix dot product, $\\sigma$ is the pruning threshold, and $\\mathcal{I}$ represents linear interpolation based on the positions", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "specified by $U_{l}$ . After pruning, the scale of the scene features is significantly reduced, enabling internal feature interactions based on self-attention. Subsequently, we utilize self-attention and cross-attention to perceive the relative relationships among objects within the scene and to fuse multimodal features, resulting in updated features $U_{l}^{\\prime}$ . Finally, through generative sparse convolutions, we obtain $U_{l - 1}^{G}$ .", + "bbox": [ + 511, + 417, + 906, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervision for Pruning. The binary supervision mask $M^{sce}$ for scene-level TGP is generated based on the centers of all objects in the scene, and the mask $M^{tar}$ for target-level TGP is based on the target and relevant objects mentioned in the descriptions:", + "bbox": [ + 511, + 526, + 905, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {s c e} = \\bigcup_ {i = 1} ^ {N} \\mathcal {M} \\left(O _ {i}\\right), \\quad M ^ {t a r} = \\mathcal {M} \\left(O ^ {t a r}\\right) \\cup \\bigcup_ {j = 1} ^ {K} \\mathcal {M} \\left(O _ {j} ^ {r e l}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 616, + 903, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{O_i|1\\leq i\\leq N\\}$ indicates all objects in the scene. $O^{tar}$ and $O^{rel}$ refer to target and relevant objects respectively. $\\mathcal{M}(O)$ represents the mask generated from the center of object $O$ . It generates a $L\\times L\\times L$ cube centered at the center of $O$ to construct the supervision mask $M$ , where locations inside the cube is set to 1 while others set to 0.", + "bbox": [ + 511, + 671, + 906, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simplification. Although the above mentioned method can effectively prune voxel features to reduce the computational cost of cross-modal attention, there are some inefficient operations in the pipeline: (1) FPS is time-consuming, especially for large scenes; (2) there are two times of interactions between voxel features and text features, the first is to guide pruning and the second is to enhance the representation, which is a bit redundant. We also empirically observe that the number of voxels is not large in level 3.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3669", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a4e77b9d9417a66cfcfba2ded9c11920b2046e1381c0234be6ffb54ea8195372.jpg", + "image_caption": [ + "(a) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 285, + 179 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af41154e9f078452d9f30b783dcdc3e2b104a5135f64eaa6669dab57903d950c.jpg", + "image_caption": [ + "(b) Pruned Features" + ], + "image_footnote": [], + "bbox": [ + 285, + 88, + 455, + 179 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f788cc08e641a9a6c6abe7e5b2ffe448156b551a3d61bc88f6b88ffd44276bde.jpg", + "image_caption": [ + "(c) Completion Features" + ], + "image_footnote": [], + "bbox": [ + 117, + 199, + 287, + 290 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d8555ac9d3a623ff5a061fce8bb0e71bfb42eb4ec4ec9177ff45f7b0a316da9.jpg", + "image_caption": [ + "(d) Prediction" + ], + "image_footnote": [], + "bbox": [ + 287, + 199, + 455, + 290 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/745c87df5c4898a5956aa422424bf8d0ac0bf15a5f089047e3d660191792c15b.jpg", + "image_caption": [ + "Figure 3. Illustration of completion-based addition. The upper figure (b) illustrates an example of over-pruning on the target. The lower figure (c) shows the completed features predicted by CBA." + ], + "image_footnote": [], + "bbox": [ + 101, + 313, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we propose a simplified version of TGP, as shown in Fig. 2 (d). We remove the FPS and merge the two multi-modal interactions into one. We also move the merged interaction operation before pruning. In this way, voxel features and text features are first deeply interacted for both feature enhancement and pruning. Because in level 3 the number of voxels is small and in level $2/1$ the voxels are already pruned, the computational cost of self-attention and cross-attention is always kept at a relatively low level.", + "bbox": [ + 89, + 491, + 482, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effectiveness of TGP. After pruning, the voxel count of $U_{1}$ is reduced to nearly $7\\%$ of its original size without TGP, while the 3DVG performance is significantly boosted. TGP serves multiple functions, including: (1) facilitating the interaction of multi-modal features through cross-attention, (2) reducing the feature amount (number of voxels) through pruning, and (3) gradually guiding the network to focus on the mentioned target based on text features.", + "bbox": [ + 89, + 627, + 482, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Completion-based Addition", + "text_level": 1, + "bbox": [ + 89, + 757, + 339, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the pruning process, some targets may be mistakenly removed, especially for small or narrow objects, as shown in Fig. 3 (b). Therefore, the addition operation between the upsampled pruned features $U_{l}^{G}$ and backbone features $V_{l}$ described in Equation (1) play an important role to mitigate the affect of over-pruning.", + "bbox": [ + 89, + 779, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are two alternative addition operations: (1) Full Addition. For the intersecting regions of $V_{l}$ and $U_{l}^{G}$ , fea-", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tures are directly added. For voxel features outside the intersection of $U_{l}^{G}$ and $V_{l}$ which lack corresponding features in the other map, the missing voxel features are interpolated before addition. Due to pruning process, $U_{l}^{G}$ is sparser than $V_{l}$ . In this way, full addition can fix almost all the pruned region. But this operation is computationally heavy and make the scene representation fail to focus on relevant objects, which deviates the core idea of TGP. (2) Pruning-aware Addition. The addition is constrained to the locations of $U_{l}^{G}$ . For voxel in $U_{l}^{G}$ but not in $V_{l}$ , interpolation from $U_{l}^{G}$ is applied to complete the missing locations in $V_{l}$ . It restricts the addition operation to the shape of the pruned features, potentially leading to an over-reliance on the results of the pruning process. If important regions are over-pruned, the network may struggle to detect targets with severely damaged geometric information.", + "bbox": [ + 511, + 90, + 903, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the unavoidable risk of pruning the query target, we introduce the completion-based addition (CBA). CBA is designed to address the limitations of full and pruning-aware additions. It offers a more targeted and efficient way to integrating multi-level features, ensuring the preservation of essential details while keeping the additional computational overhead negligible.", + "bbox": [ + 511, + 332, + 903, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Details of CBA. We first enhance the backbone features $V_{l}$ with the text features $T$ through cross-attention, obtaining $V_{l}^{\\prime}$ . Then a MLP is adopted to predict the probability distribution of target for region selection:", + "bbox": [ + 511, + 438, + 903, + 498 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {l} ^ {t a r} = \\Theta (\\mathrm {M L P} \\left(V _ {l} ^ {\\prime}\\right) - \\tau) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 503, + 903, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Theta$ is the step function, and $\\tau$ is the threshold determining voxel relevance. $M_{l}^{tar}$ is a binary mask indicating potential regions of the mentioned target. Then, comparison of $M_{l}^{tar}$ with $U_{l}$ identifies missing voxels. The missing mask $M_{l}^{mis}$ is derived as follows:", + "bbox": [ + 511, + 525, + 903, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {l} ^ {m i s} = M _ {l} ^ {t a r} \\wedge (\\neg \\mathcal {C} (U _ {l} ^ {G}, V _ {l})) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 606, + 903, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}(A,B)$ denotes the generation of a binary mask for $A$ based on the shape of $B$ . Specifically, for positions in $B$ , if there are corresponding voxel features in $A$ , the mask for that position is set to 1. Otherwise it is set to 0. Missed voxel features in $U_{l}^{G}$ that correspond to $M_{l}^{mis}$ are interpolated from $U_{l}^{G}$ , filling in gaps identified by the missing mask. The completed feature map $U_{l}^{cpl}$ is computed by:", + "bbox": [ + 511, + 628, + 903, + 738 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {l} ^ {c p l} = V _ {l} ^ {\\prime} \\odot M _ {l} ^ {m i s} + \\mathcal {I} \\left(U _ {l} ^ {G}, M _ {l} ^ {m i s}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 742, + 903, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ represents linear interpolation on the feature map based on the positions specified in the mask. Finally, the original upsampled features are combined with the backbone features according to the pruning-aware addition, and merged with the completion features to yield updated $U_{l}$ :", + "bbox": [ + 511, + 767, + 903, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {l} = \\operatorname {C o n c a t} \\left(U _ {l} ^ {G} \\leftarrow V _ {l}, U _ {l} ^ {c p l}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 845, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\leftarrow$ denotes the pruning-aware addition, and Concat means concatenation of voxel features.", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3670", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b2dfc9c2632c9ad671c4ffe9e10a0152d7a088e8001aebcaa9873c3f3489314d.jpg", + "table_caption": [ + "Table 1. Comparison of methods on the ScanRefer dataset evaluated at IoU thresholds of 0.25 and 0.5. TSP3D achieves state-of-the-art accuracy even compared with two-stage methods, with +1.13 lead on Acc@0.5. Notably, we are the first to comprehensively evaluate inference speed for 3DVG methods. The inference speeds of other methods are obtained through our reproduction." + ], + "table_footnote": [], + "table_body": "
MethodVenueInputAccuracyInference Speed (FPS)
0.250.5
Two-Stage Model
ScanRefer [3]ECCV'203D+2D41.1927.406.72
TGNN [14]AAAI'213D37.3729.703.19
InstanceRefer [39]ICCV'213D40.2330.152.33
SAT [38]ICCV'213D+2D44.5430.144.34
FFL-3DOG [10]ICCV'213D41.3334.01Not released
3D-SPS [22]CVPR'223D+2D48.8236.983.17
BUTD-DETR [16]ECCV'223D50.4238.603.33
EDA [35]CVPR'233D54.5942.263.34
3D-VisTA [41]ICCV'233D45.9041.502.03
VPP-Net [31]CVPR'243D55.6543.29Not released
G3-LQ [34]CVPR'243D56.9045.58Not released
MCLN [27]ECCV'243D57.1745.533.17
Single-stage Model
3D-SPS [22]CVPR'223D47.6536.435.38
BUTD-DETR [16]ECCV'223D49.7637.055.91
EDA [35]CVPR'233D53.8341.705.98
G3-LQ [34]CVPR'243D55.9544.72Not released
MCLN [27]ECCV'243D54.3042.645.45
TSP3D (Ours)—3D56.4546.7112.43
", + "bbox": [ + 93, + 176, + 480, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Train Loss", + "text_level": 1, + "bbox": [ + 89, + 467, + 209, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The loss is composed of several components: pruning loss for TGP, completion loss for CBA, and objectness loss as well as bounding box regression loss for the head. Pruning loss, completion loss and objectness loss employ the focal loss to handle class imbalance. Supervision for completion and classification losses are the same, which sets voxels near the target object center as positives while leaving others as negatives. For bounding box regression, we use the Distance-IoU (D IoU) loss. The total loss function is computed as the sum of these individual losses:", + "bbox": [ + 89, + 489, + 482, + 640 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {p r u n i n g}} + \\lambda_ {2} \\mathcal {L} _ {\\text {c o m}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c l a s s}} + \\lambda_ {4} \\mathcal {L} _ {\\text {b b o x}}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 650, + 452, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{1},\\lambda_{2},\\lambda_{3}$ and $\\lambda_4$ are the weights of different parts.", + "bbox": [ + 89, + 675, + 473, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 702, + 220, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 727, + 191, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We maintain the same experimental settings with previous works, employing ScanRefer [3] and SR3D/NR3D [2] as datasets. ScanRefer: Built on the ScanNet framework, ScanRefer includes 51,583 descriptions across scenes. Evaluation metrics focus on Acc@mIoU. ReferIt3D: ReferIt3D splits into Nr3D, with 41,503 human-generated descriptions, and Sr3D, containing 83,572 synthetic expressions. ReferIt3D simplifies the task by providing segmented point clouds for each object. The primary evaluation metric is accuracy in target object selection.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/506ed8bf9b7118e6a63a83288c30c9aa6914bd274eea0766035bf4df733881db.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons on Nr3D and Sr3D datasets. We evaluate under three pipelines, noting that the Two-stage using Ground-Truth Boxes is impractical for real-world applications. TSP3D exhibits significant superiority, with leads of $+2.6\\%$ and $+3.2\\%$ on NR3D and SR3D respectively." + ], + "table_footnote": [], + "table_body": "
MethodVenuePipelineAccuracy
Nr3DSr3D
InstanceRefer [39]ICCV'21Two-stage (gt)38.848.0
LanguageRefer [28]CoRL'22Two-stage (gt)43.956.0
3D-SPS [22]CVPR'22Two-stage (gt)51.562.6
MVT [15]CVPR'22Two-stage (gt)55.164.5
BUTD-DETR [16]ECCV'22Two-stage (gt)54.667.0
EDA [35]CVPR'23Two-stage (gt)52.168.1
VPP-Net [31]CVPR'24Two-stage (gt)56.968.7
G3-LQ [34]CVPR'24Two-stage (gt)58.473.1
MCLN [27]ECCV'24Two-stage (gt)59.868.4
InstanceRefer [39]ICCV'21Two-stage (det)29.931.5
LanguageRefer [28]CoRL'22Two-stage (det)28.639.5
BUTD-DETR [16]ECCV'22Two-stage (det)43.352.1
EDA [35]CVPR'23Two-stage (det)40.749.9
MCLN [27]ECCV'24Two-stage (det)46.153.9
3D-SPS [22]CVPR'22Single-stage39.247.1
BUTD-DETR [16]ECCV'22Single-stage38.750.1
EDA [35]CVPR'23Single-stage40.049.7
MCLN [27]ECCV'24Single-stage45.753.4
TSP3D (Ours)—Single-stage48.757.1
", + "bbox": [ + 519, + 162, + 897, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 440, + 728, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TSP3D is implemented based on PyTorch [23]. The pruning thresholds are set at $\\sigma_{\\mathrm{sc}} = 0.7$ and $\\sigma_{\\mathrm{tar}} = 0.3$ , and the completion threshold in CBA is $\\tau = 0.15$ . The initial voxelization of the point cloud has a voxel size of $1\\mathrm{cm}$ , while the voxel size for level $i$ features scales to $2^{i + 2}\\mathrm{cm}$ . The supervision for pruning uses $L = 7$ . The weights for all components of the loss function, $\\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4$ , are equal to 1. Training is conducted using four GPUs, while inference speeds are evaluated using a single consumer-grade GPU, RTX 3090, with a batch size of 1.", + "bbox": [ + 511, + 460, + 906, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 511, + 622, + 751, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on ScanRefer. We carry out comparisons with existing methods on ScanRefer, as detailed in Tab. 1. The inference speeds of other methods are obtained through our reproduction with a single RTX 3090 and a batch size of 1. For two-stage methods, the inference speed includes the time taken for object detection in the first stage. For methods using 2D image features and 3D point clouds as inputs, we do not account for the time spent extracting 2D features, assuming they can be obtained in advance. However, in practical applications, the acquisition of 2D features also impacts overall efficiency. TSP3D achieves state-of-the-art accuracy even compared with two-stage methods, with $+1.13$ lead on Acc@0.5. Notably, in the single-stage setting, TSP3D achieves fast inference speed, which is unprecedented among the existing methods. This significant improvement is attributed to our method's efficient use of a multi-level architecture based on 3D sparse convolu", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3671", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e97834ad6713a1a6fdae4b0102fe0b05211054ac3b6c6b210bdf4cbf943439f1.jpg", + "table_caption": [ + "Table 3. Impact of the proposed TGP and CBA. Evaluated on ScanRefer." + ], + "table_footnote": [], + "table_body": "
IDTGPCBAAccuracySpeed (FPS)
0.250.5
(a)40.1332.8714.58
(b)✓55.2046.1513.22
(c)✓41.3433.0913.51
(d)✓✓56.4546.7112.43
", + "bbox": [ + 91, + 122, + 336, + 194 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/801e4a6da6c87829b9dd0d38c4134c132668c9e669e8f0dbdc73e710fad268ed.jpg", + "table_caption": [ + "Table 4. Influence of the two CBAs at different levels. Evaluated on ScanRefer." + ], + "table_footnote": [], + "table_body": "
IDCBA (level 2)CBA (level 1)AccuracySpeed (FPS)
0.250.5
(a)55.2046.1513.22
(b)✓55.1746.0612.79
(c)✓56.4546.7112.43
(d)✓✓56.2246.6812.19
", + "bbox": [ + 346, + 122, + 614, + 193 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d53cceb186d49079b8c143f141b2c829601e5dd2ae0fb65ca977ce87a572dfb2.jpg", + "table_caption": [ + "Table 5. Influence of different feature upsampling methods. Evaluated on ScanRefer." + ], + "table_footnote": [], + "table_body": "
IDMethodAccuracySpeed (FPS)
0.250.5
(a)Simple concatenation40.1332.8714.58
(b)Attention mechanism———
(c)Text-guided pruning56.2746.5810.11
(d)Simplified TGP56.4546.7112.43
", + "bbox": [ + 627, + 123, + 898, + 193 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tions, coupled with the text-guided pruning. By focusing computation only on salient regions of the point clouds, determined by textual cues, our model effectively reduces computational overhead while maintaining high accuracy. TSP3D also sets a benchmark for inference speed comparisons for future methodologies.", + "bbox": [ + 88, + 208, + 480, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on Nr3D/Sr3D. We evaluate our method on the SR3D and NR3D datasets, following the evaluation protocols of prior works like EDA [35] and BUTD-DETR [16] by using Acc@0.25 as the accuracy metric. The results are shown in Tab. 2. Given that SR3D and NR3D provide ground-truth boxes and categories for all objects in the scene, we consider three pipelines: (1) Two-stage using Ground-Truth Boxes, (2) Two-stage using Detected Boxes, and (3) Single-stage. In practical applications, the Two-stage using Ground-Truth Boxes pipeline is unrealistic because obtaining all ground-truth boxes in a scene is infeasible. This approach can also oversimplify certain evaluation scenarios. For example, if there are no other objects of the same category as the target in the scene, the task reduces to relying on the provided ground-truth category. Under the Single-stage setting, TSP3D exhibits significant superiority with peak performance of $48.7\\%$ and $57.1\\%$ on Nr3D and Sr3D. TSP3D even outperforms previous works under the pipeline of Two-stage using Detected Boxes, with leads of $+2.6\\%$ and $+3.2\\%$ on NR3D and SR3D.", + "bbox": [ + 91, + 301, + 483, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 619, + 243, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Proposed Components. To investigate the effects of our proposed TGP and CBA, we conduct ablation experiments with module removal as shown in Tab. 3. When TGP is not used, multi-modal feature concatenation is employed as a replacement, as shown in Fig. 2 (a). When CBA is not used, it is substituted with a pruning-based addition. The results demonstrate that TGP significantly enhances performance without notably impacting inference time. This is because TGP, while utilizing a more complex multi-modal attention mechanism for stronger feature fusion, significantly reduces feature scale through text-guided pruning. Additionally, the performance improvement is also due to the gradual guidance towards the target object by both scene-level and target-level TGP. Using CBA alone has a limited effect, as no voxels are pruned. Implementing CBA on top of TGP further enhances performance, as CBA dynamically compensates for some of the excessive", + "bbox": [ + 91, + 643, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pruning by TGP, thus increasing the network's robustness.", + "bbox": [ + 511, + 209, + 897, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of the Two CBAs. To explore the impact of CBAs at two different levels, we conduct ablation experiments as depicted in Tab. 4. In the absence of CBA, we use pruning-based addition as a substitute. The results indicate that the CBA at level 2 has negligible effects on the 3DVG task. This is primarily because the CBA at level 2 serves to supplement the scene-level TGP, which is expected to prune the background (a relatively simple task). Moreover, although some target features are pruned, they are compensated by two subsequent generative sparse convolutions. However, the CBA at level 1 enhances performance by adapt completion for the target-level TGP. It is challenging to fully preserve target objects from deep upsampling features, especially for smaller or narrower targets. The CBA at level 1, based on high-resolution backbone features, effectively complements the TGP.", + "bbox": [ + 511, + 224, + 903, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Upsampling Techniques. We conduct experiments to assess the effects of different feature upsampling techniques, as detailed in Tab. 5. Using simple feature concatenation (Fig. 2 (a)), while fast in inference speed, results in poor performance. When we utilize an attention mechanism with stronger feature interaction, as shown in Fig. 2 (b), the computation exceeds the limits of GPU due to the large number of voxels, making it impractical for real-world applications. Consequently, we employ TGP to reduce the feature amount, as illustrated in Fig. 2 (c), which significantly improves performance and enables practical deployment. Building on TGP, we propose simplified TGP, as shown in Fig. 2 (d), that merges feature interactions before and after pruning, achieving performance consistent with the original TGP while enhancing inference speed.", + "bbox": [ + 511, + 469, + 905, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 710, + 694, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-guided Pruning. To visually demonstrate the process of TGP, we visualize the results of two pruning phases, as shown in Fig. 4. In each example, the voxel features after scene-level pruning, the features after target-level pruning, and the features after target-level generative sparse convolution are displayed from top to bottom. It is evident that both pruning stages effectively achieve our intended effect: the scene-level pruning filters out the background and retained object voxels, and the target-level pruning preserves relevant and target objects. Moreover, during the feature upsampling process, the feature amount nearly exponen-", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3672", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e127a077bd0a6d27b89d32dd19e93b9c5eb2fb76fa75e59d370ee9285df37ad2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 88, + 281, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e352ef8d96b9ba4385c259554b2814f259eef93cdf05e72d5a6e185c8c2abf78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 172, + 279, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b810d25fd78ca887e04647bf2ab28495c02a264413ff1137c409bf98e9353a3c.jpg", + "image_caption": [ + "(a) It is a gray trash can. \nThe trash can sits in the corner by where the tv is." + ], + "image_footnote": [], + "bbox": [ + 151, + 253, + 279, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c7df1d15cac2d971bdead70fede0d09a5a1d53cf895795226c0c147dadff021.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 88, + 491, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b59f969347d2443c39de66e06a3aa001e3ea34fc6a80cbaf3281305ccf4f5da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 172, + 491, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6eec883c7019c38b446623085adc929254d1475ff4e4a683b05051cd167999c.jpg", + "image_caption": [ + "(b) The chair is on the northwest side of the table that is furthest northeast in the room. The chair has a curved backside and four legs." + ], + "image_footnote": [], + "bbox": [ + 295, + 252, + 491, + 334 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7997523968f11a8f0bfa57b29838c8b57ebdf699b5464901ea968deaebd0796d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 674, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d82d108ae9eeb5e40e61c11c0285c00063831f972110e31308d2d90158d65a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 170, + 674, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c78287d31b5b32615d3dc245b0de69e2419f6bdd72d8b17a2b6e8d2da4cef986.jpg", + "image_caption": [ + "(c) A chrome water fountain. Is the lowest of two water fountains." + ], + "image_footnote": [], + "bbox": [ + 496, + 252, + 674, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b839a58752be639e07cae1283622bd2a5b9d1d7e54a679611deb00e7f72e4903.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 88, + 838, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/314704d0cc05d1e8b9f8b526f825d0dc48364bad4c108498bceb9080f27b7db2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 172, + 834, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2aaa174cfe98fc7a6d6e0f85ca56d6c027848a80c4e805e07f10f8b6b0b565ed.jpg", + "image_caption": [ + "(d) There is a rectangular shelf. It is the one closest to the curtain." + ], + "image_footnote": [], + "bbox": [ + 681, + 253, + 834, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb9a11ff5c5b05871f2c87ac2b51f9376842da544c867c92d7b52249b1c4bb50.jpg", + "image_caption": [ + "Figure 4. Visualization of the text-guided pruning process. In each example, the voxel features after scene-level TGP, target-level TGP and the last upsampling layer are presented from top to bottom. The blue boxes represent the ground truth of the target, and the red boxes denote the bounding boxes of relevant objects. TSP3D reduces the amount of voxel features through two stages of pruning and progressively guides the network focusing towards the target.", + "(a) The brown square trash can is to the right of the lamp. It is also behind the chair and to the right." + ], + "image_footnote": [], + "bbox": [ + 151, + 446, + 341, + 539 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/978b413f7aff11a5c5d40e76562670a083c4aa956fd214fb7e44cdef0a4411b1.jpg", + "image_caption": [ + "(b) The whiteboard is next to two doors. The whiteboard is a long, white rectangle." + ], + "image_footnote": [], + "bbox": [ + 346, + 448, + 500, + 540 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9dd26ab7936da55721b48a48eab1056c21cb39b02b22237c3e4763534d5fc224.jpg", + "image_caption": [ + "(c) This is a monitor in black color. The monitor is next to the wall and at the back of maroon chair." + ], + "image_footnote": [], + "bbox": [ + 506, + 448, + 689, + 539 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b4a7e582a78640f923f627372a739aa3cce56600e0f065314ece634229b53a0.jpg", + "image_caption": [ + "(d) The book rack is left of the bookshelf that looks like a right angle with another. The book rack is black and has as circular bottom.", + "Figure 5. Visualization of the completion-based addition process. The blue points represent the voxel features output by the target-level TGP, while the red points are the completion features predicted by the CBA. The blue boxes indicate the ground truth boxes. CBA adaptively supplements situations where excessive pruning has occurred." + ], + "image_footnote": [], + "bbox": [ + 700, + 448, + 826, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tially increases due to generative upsampling. Without TGP, the voxel coverage would far exceed the range of the scene point cloud, which is inefficient for inference. This also intuitively explains the significant impact of our TGP on both performance and inference speed.", + "bbox": [ + 88, + 645, + 482, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Completion-based Addition. To clearly illustrate the function of CBA, we visualize the adaptive completion process in Fig. 5. The images below showcase several instances of excessive pruning. TGP performs pruning based on deep and low-resolution features, which can lead to excessive pruning, potentially removing entire or partial targets. This over-pruning is more likely to occur with small, as shown in Fig. 5 (a) and (c), narrow, as in Fig. 5 (b), or elongated targets, as in Fig. 5 (d). Our CBA effectively supplements the process using higher-resolution backbone features, thus dynamically integrating multi-level features.", + "bbox": [ + 88, + 734, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 643, + 633, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present TSP3D, an efficient sparse single-stage method for real-time 3D visual grounding. Different from previous 3D visual grounding frameworks, TSP3D builds on multi-level sparse convolutional architecture for efficient and fine-grained scene representation extraction. To enable the interaction between voxel features and textual features, we propose text-guided pruning (TGP), which reduces the amount of voxel features and guides the network to progressively focus on the target object. Additionally, we introduce completion-based addition (CBA) for adaptive multi-level feature fusion, effectively compensating for instances of over-pruning. Extensive experiments demonstrate the effectiveness of our proposed modules, resulting in an efficient 3DVG method that achieves state-of-the-art accuracy and fast inference speed.", + "bbox": [ + 511, + 674, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3673", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 624B2076, Grant 62376132, Grant 62321005, and Grant 62441616.", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dref-transformer: Fine-grained object identification in real-world scenes using natural language. In WACV, pages 3941-3950, 2022. 2", + "[2] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 6", + "[3] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 1, 2, 3, 6", + "[4] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In ICCV, pages 15467-15476, 2021. 2", + "[5] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In CVPR, pages 21674-21683, 2023. 2", + "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2, 3", + "[7] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2", + "[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, pages 1201-1209, 2021. 2", + "[9] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[10] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In ICCV, pages 3722-3731, 2021. 2, 6", + "[11] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224–9232, 2018. 2", + "[12] JunYoung Gwak, Christopher Choy, and Silvio Savarese. Generative sparse detection networks for 3d single-shot object detection. In ECCV, pages 297-313. Springer, 2020. 2, 3", + "[13] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-and" + ], + "bbox": [ + 93, + 200, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "relation aware transformer for fine-grained 3d visual grounding. In ACM MM, pages 2344-2352, 2021. 2", + "[14] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 1, 2, 6", + "[15] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 6", + "[16] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 1, 2, 3, 6, 7", + "[17] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In CVPR, pages 4867-4876, 2020. 2", + "[18] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 2", + "[19] Daizong Liu, Yang Liu, Wencan Huang, and Wei Hu. A survey on text-guided 3d visual grounding: Elements, recent advances, and future directions. arXiv preprint arXiv:2406.05785, 2024. 2", + "[20] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3", + "[21] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, pages 2949-2958, 2021. 2", + "[22] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 1, 2, 3, 6", + "[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6", + "[24] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, pages 1532-1543, 2014. 2", + "[25] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 2, 3", + "[26] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2", + "[27] Zhipeng Qian, Yiwei Ma, Zhekai Lin, Jiayi Ji, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Multi-branch collaborative learning network for 3d visual grounding. In ECCV, pages 381–398. Springer, 2025. 2, 6", + "[28] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In CoRL, pages 1046-1056. PMLR, 2022. 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3674", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In ECCV, pages 477-493. Springer, 2022. 2, 3", + "[30] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Tr3d: Towards real-time indoor 3d object detection. In ICIP, pages 281-285. IEEE, 2023. 2, 3", + "[31] Xiangxi Shi, Zhonghua Wu, and Stefan Lee. Viewpoint-aware visual grounding in 3d scenes. In CVPR, pages 14056-14065, 2024. 3, 6", + "[32] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 2", + "[33] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. NeurIPS, 35:29975-29988, 2022. 2", + "[34] Yuan Wang, Yali Li, and Shengjin Wang. G^3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. In CVPR, pages 13917-13926, 2024. 2, 3, 6", + "[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 1, 2, 3, 6, 7", + "[36] Xiuwei Xu, Ziwei Wang, Jie Zhou, and Jiwen Lu. Binarizing sparse convolutional networks for efficient point cloud analysis. In CVPR, pages 5313-5322, 2023. 2", + "[37] Xiuwei Xu, Zhihao Sun, Ziwei Wang, Hongmin Liu, Jie Zhou, and Jiwen Lu. 3d small object detection with dynamic spatial pruning. In ECCV. Springer, 2024. 2", + "[38] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 1, 6", + "[39] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In ICCV, pages 1791-1800, 2021. 1, 2, 6", + "[40] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvgtransformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 2", + "[41] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 6" + ], + "bbox": [ + 91, + 92, + 480, + 727 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3675", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_model.json b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..deb49eac22188d772948d682a3c5039b1cc2fe4e --- /dev/null +++ b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.131, + 0.847, + 0.154 + ], + "angle": 0, + "content": "Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.179, + 0.871, + 0.218 + ], + "angle": 0, + "content": "Wenxuan Guo\\(^{1*}\\) Xiuwei Xu\\(^{1*}\\) Ziwei Wang\\(^{2}\\) Jianjiang Feng\\(^{1\\dagger}\\) Jie Zhou\\(^{1}\\) Jiwen Lu\\(^{1\\dagger}\\) Tsinghua University \\(^{2}\\)Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.219, + 0.758, + 0.235 + ], + "angle": 0, + "content": "\\(\\{gwx22,xxw21\\} @mails.tsinghua.edu.cn ziwei.wang@ntu.edu.sg\\)" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.236, + 0.662, + 0.252 + ], + "angle": 0, + "content": "{jfeng,jzhou,lujiwen}@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.486, + 0.742 + ], + "angle": 0, + "content": "In this paper, we propose an efficient multi-level convolution architecture for 3D visual grounding. Conventional methods are difficult to meet the requirements of real-time inference due to the two-stage or point-based architecture. Inspired by the success of multi-level fully sparse convolutional architecture in 3D object detection, we aim to build a new 3D visual grounding framework following this technical route. However, as in 3D visual grounding task the 3D scene representation should be deeply interacted with text features, sparse convolution-based architecture is inefficient for this interaction due to the large amount of voxel features. To this end, we propose text-guided pruning (TGP) and completion-based addition (CBA) to deeply fuse 3D scene representation and text features in an efficient way by gradual region pruning and target completion. Specifically, TGP iteratively sparsifies the 3D scene representation and thus efficiently interacts the voxel features with text features by cross-attention. To mitigate the affect of pruning on delicate geometric information, CBA adaptively fixes the over-pruned region by voxel completion with negligible computational overhead. Compared with previous single-stage methods, our method achieves top inference speed and surpasses previous fastest method by \\(100\\%\\) FPS. Our method also achieves state-of-the-art accuracy even compared with two-stage methods, with \\(+1.13\\) lead of Acc@0.5 on ScanRefer, and \\(+2.6\\) and \\(+3.2\\) leads on NR3D and SR3D respectively. The code is available at https://github.com/GWxuan/TSP3D." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.773, + 0.222, + 0.789 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.799, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Incorporating multi-modal information to guide 3D visual perception is a promising direction. In these years, 3D visual grounding (3DVG), also known as 3D instance referencing, has been paid increasing attention as a fundamental multi-modal 3D perception task. The aim of 3DVG is" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.286, + 0.903, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.493, + 0.907, + 0.536 + ], + "angle": 0, + "content": "Figure 1. Comparison of 3DVG methods on ScanRefer dataset [3]. Our TSP3D surpasses existing methods in both accuracy and inference speed, achieving the first efficient 3DVG framework." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.553, + 0.907, + 0.689 + ], + "angle": 0, + "content": "to locate an object in the scene with a free-form query description. 3DVG is challenging since it requires understanding of both 3D scene and language description. Recently, with the development of 3D scene perception and vision-language models, 3DVG methods have shown remarkable progress [16, 22]. However, with 3DVG being widely applied in fields like robotics and AR / VR where inference speed is the main bottleneck, how to construct efficient real-time 3DVG model remains a challenging problem." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Since the output format of 3DVG is similar with 3D object detection, early 3DVG methods [3, 14, 38, 39] usually adopt a two-stage framework, which first conducts detection to locate all objects in the scene, and then selects the target object by incorporating text information. As there are many similarities between 3D object detection and 3DVG (e.g. both of them need to extract the representation of the 3D scene), there will be much redundant feature computation during the independent adoption of the two models. As a result, two-stage methods are usually hard to handle real-time tasks. To solve this problem, single-stage methods [22, 35] are presented, which generates the bounding box of the target directly from point clouds. This integrated design is more compact and efficient. However," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.352, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "current single-stage 3DVG methods mainly build on point-based architecture [25], where the feature extraction contains time-consuming operations like furthest point sampling and kNN. They also need to aggressively downsample the point features to reduce computational cost, which might hurt the geometric information of small and thin objects [37]. Due to these reasons, current single-stage methods are still far from real-time (\\(< 6\\) FPS) and their performance is inferior to two-stage methods, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.229, + 0.487, + 0.76 + ], + "angle": 0, + "content": "In this paper, we propose a new single-stage framework for 3DVG based on text-guided sparse voxel pruning, namely TSP3D. Inspired by state-of-the-art 3D object detection methods [29, 37] which achieves both leading accuracy and speed with multi-level sparse convolutional architecture, we build the first sparse single-stage 3DVG network. However, different from 3D object detection, in 3DVG the 3D scene representation should be deeply interacted with text features. Since the count of voxels is very large in sparse convolution-based architecture, deep multi-modal interaction like cross-attention becomes infeasible due to unaffordable computational cost. To this end, we propose text-guided pruning (TGP), which first utilize text information to jointly sparsify the 3D scene representation and enhance the voxel and text features. To mitigate the affect of pruning on delicate geometric information, we further present completion-based addition (CBA) to adaptively fix the over-pruned region with negligible computational overhead. Specifically, TGP prunes the voxel features according to the object distribution. It gradually removes background features and features of irrelevant objects, which generates text-aware voxel features around the target object for accurate bounding box prediction. Since pruning may mistakenly remove the representation of target object, CBA utilizes text features to query a small set of voxel features from the complete backbone features, followed by pruned-aware addition to fix the over-pruned region. We conduct extensive experiments on the popular ScanRefer [3] and ReferIt3D [2] datasets. Compared with previous single-stage methods, TSP3D achieves top inference speed and surpasses previous fastest single-stage method by \\(100\\%\\) FPS. TSP3D also achieves state-of-the-art accuracy even compared with two-stage methods, with \\(+1.13\\) lead of Acc@0.5 on ScanRefer, and \\(+2.6\\) and \\(+3.2\\) leads on NR3D and SR3D respectively." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.761, + 0.465, + 0.775 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.808 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first work exploring sparse convolutional architecture for efficient 3DVG." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.808, + 0.482, + 0.853 + ], + "angle": 0, + "content": "- To enable efficient feature extraction, we propose text-guided pruning and completion-based addition to sparsify sparse voxels and adaptively fuse multi-level features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.854, + 0.483, + 0.898 + ], + "angle": 0, + "content": "- We conduct extensive experiments, and TSP3D outperforms existing methods in both accuracy and speed, demonstrating the superiority of the proposed framework." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.778, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.715, + 0.132 + ], + "angle": 0, + "content": "2.1.3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.908, + 0.517 + ], + "angle": 0, + "content": "3D visual grounding aims to locate a target object within a 3D scene based on natural language descriptions [19]. Existing methods are typically categorized into two-stage and single-stage approaches. Two-stage methods follow a detect-then-match paradigm. In the first stage, they independently extract features from the language query using pre-trained language models [7, 9, 24] and predict candidate 3D objects using pre-trained 3D detectors [21, 26] or segmenters [4, 17, 32]. In the second stage, they focus on aligning the vision and text features to identify the target object. Techniques for feature fusion include attention mechanisms with Transformers [13, 40], contrastive learning [1], and graph-based matching [10, 14, 39]. In contrast, single-stage methods integrate object detection and feature extraction, allowing for direct identification of the target object. Methods in this category include guiding keypoint selection using textual features [22], and measuring similarity between words and objects inspired by 2D image-language pre-trained models like GLIP [18], as in BUTDDETR [16]. And methods like EDA [35] and \\(\\mathrm{G}^3\\)-LQ [34] advance single-stage 3D visual grounding by enhancing multimodal feature discriminability through explicit textdecoupling, dense alignment, and semantic-geometric modeling. MCLN [27] uses the 3D referring expression segmentation task to assist 3DVG in improving performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.518, + 0.907, + 0.593 + ], + "angle": 0, + "content": "However, existing two-stage and single-stage methods generally have high computational costs, hindering real-time applications. Our work aims to address these efficiency challenges by proposing an efficient single-stage method with multi-level sparse convolutional architecture." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.863, + 0.62 + ], + "angle": 0, + "content": "2.2. Multi-Level Convolutional Architectures" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.628, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Recently, sparse convolutional architecture has achieved great success in the field of 3D object detection. Built on the voxel-based representation [5, 8, 33] and sparse convolution operation [6, 11, 36], this kind of methods show great efficiency and accuracy when processing scene-level data. GSDN [12] first adopts multi-level sparse convolution with generative feature upsampling in 3D object detection. FCAF3D [29] simplifies the multi-level architecture with anchor-free design, achieving leading accuracy and speed. TR3D [30] further accelerates FCAF3D by removing unnecessary layers and introducing category-aware proposal assignment method. Moreover, DSPDet3D [37] introduces the multi-level architecture to 3D small object detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Our proposed method draws inspiration from these approaches, utilizing a sparse multi-level architecture with sparse convolutions and an anchor-free design. This allows for efficient processing of 3D data, enabling real-time performance in 3D visual grounding tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3667" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.182, + 0.106 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.238 + ], + "angle": 0, + "content": "In this section, we describe our TSP3D for efficient single-stage 3DVG. We first analyze existing pipelines to identify current challenges and motivate our approach (Sec. 3.1). We then introduce the text-guided pruning, which leverages text features to guide feature pruning (Sec. 3.2). To address the potential risk of pruning key information, we propose the completion-based addition for multi-level feature fusion (Sec. 3.3). Finally, we detail the training loss (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.245, + 0.373, + 0.26 + ], + "angle": 0, + "content": "3.1. Architecture Analysis for 3DVG" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.267, + 0.485, + 0.553 + ], + "angle": 0, + "content": "Top-performance 3DVG methods [31, 34, 35], are mainly two-stage, which is a serial combination of 3D object detection and 3D object grounding. This separate calls of two approaches result in redundant feature extraction and complex pipeline, thus making the two-stage methods less efficient. To demonstrate the efficiency of existing methods, we conduct a comparison of accuracy and speed among several representative methods on ScanRefer [3], as shown in Fig. 1. It can be seen that two-stage methods struggle in speed (< 3 FPS) due to the additional detection stage. Since 3D visual grounding is usually adopted in practical scenarios that require real-time inference under limited resources, such as embodied robots and VR/AR, the low speed of two-stage methods make them less practical. On the other side, single-stage methods [22], which directly predicts referred bounding box from the observed 3D scene, are more suitable choices due to their streamlined processes. In Fig. 1, it can be observed that single-stage methods are significantly more efficient than their two-stage counterparts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.765 + ], + "angle": 0, + "content": "However, existing single-stage methods are mainly built on point-based backbone [25], where the scene representation is extracted with time-consuming operations like furthest point sampling and set abstraction. They also employ large transformer decoder to fuse text and 3D features for several iterations. Therefore, the inference speed of current single-stage methods is still far from real-time (\\(< 6\\) FPS). The inference speed of specific components in different frameworks is analyzed and discussed in detail in the supplementary material. Inspired by the success of multi-level sparse convolutional architecture in 3D object detection [30], which achieves both leading accuracy and speed, we propose to build the first multi-level convolutional single-stage 3DVG pipeline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "TSP3D-B. Here we propose a baseline framework based on sparse convolution, namely TSP3D-B. Following the simple and effective multi-level architecture of FCAF3D [29], TSP3D-B utilizes 3 levels of sparse convolutional blocks for scene representation extraction and bounding box prediction, as shown in Fig. 2 (a). Specifically, the input point clouds \\( P \\in \\mathbb{R}^{N \\times 6} \\) with 6-dim features (3D position and RGB) are first voxelized and then fed into three sequential MinkResBlocks [6], which generates three lev" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "els of voxel features \\( V_{l} \\) (\\( l = 1,2,3 \\)). With the increase of \\( l \\), the spatial resolution of \\( V_{l} \\) decreases and the context information increases. Concurrently, the free-form text with \\( l \\) words is encoded by the pre-trained RoBERTa [20] and produce the vanilla text tokens \\( T \\in \\mathbb{R}^{l \\times d} \\). With the extracted 3D and text representations, we iteratively upsample \\( V_{3} \\) and fuse it with \\( T \\) to generate high-resolution and text-aware scene representation:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.221, + 0.907, + 0.24 + ], + "angle": 0, + "content": "\\[\nU _ {l} = U _ {l} ^ {G} + V _ {l}, \\quad U _ {l} ^ {G} = \\operatorname {G e S p C o n v} \\left(U _ {l + 1} ^ {\\prime}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.247, + 0.907, + 0.265 + ], + "angle": 0, + "content": "\\[\nU _ {l + 1} ^ {\\prime} = \\operatorname {C o n c a t} \\left(U _ {l + 1}, T\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.405 + ], + "angle": 0, + "content": "where \\( U_{3} = V_{3} \\), GeSpConv means generative sparse convolution [12] with stride 2, which upsamples the voxel features and expands their spatial locations for better bounding box prediction. Concat is voxel-wise feature concatenation by duplicating \\( T \\). The final upsampled feature map \\( U_{1} \\) is concatenated with \\( T \\) and fed into a convolutional head to predict the objectness scores and regress the 3D bounding box. We select the box with highest objectness score as the grounding result." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.907, + 0.465 + ], + "angle": 0, + "content": "As shown in Fig. 1, TSP3D-B achieves an inference speed of 14.58 FPS, which is significantly faster than previous single-stage methods and demonstrates great potential for real-time 3DVG." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.474, + 0.71, + 0.49 + ], + "angle": 0, + "content": "3.2. Text-guided Pruning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.752 + ], + "angle": 0, + "content": "Though efficient, TSP3D-B exhibits poor performance due to the inadequate interaction between 3D scene representation and text features. Motivated by previous 3DVG methods [16], a simple solution is to replace Concat with cross-modal attention to process voxel and text features, as shown in Fig. 2 (b). However, different from point-based architectures where the scene representation is usually aggressively downsampled, the number of voxels in multi-level convolutional framework is very large1. In practical implementation, we find that the voxels expand almost exponentially with each upsampling layer, leading to a substantial computational burden for the self-attention and cross-attention of scene features. To address this issue, we introduce text-guided pruning (TGP) to construct TSP3D, as illustrated in Fig. 2 (c). The core idea of TGP is to reduce feature amount by pruning redundant voxels and guide the network to gradually focus on the final target based on textual features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.909, + 0.844 + ], + "angle": 0, + "content": "Overall Architecture. TGP can be regarded as a modified version of cross-modal attention, which reduces the number of voxels before attention operation, thereby lowering computational cost. To minimize the affect of pruning on the final prediction, we propose to prune the scene representation gradually. At higher level where the number of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.852, + 0.908, + 0.902 + ], + "angle": 0, + "content": "1Compared to point-based architectures, sparse convolutional framework provides higher resolution and more detailed scene representations, while also offering advantages in inference speed. For detailed statistics, please refer to the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3668" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.089, + 0.531, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.094, + 0.639, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.186, + 0.639, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.094, + 0.764, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.094, + 0.884, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.302, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Figure 2. Illustration of TSP3D. TSP3D bulids on multi-level sparse convolutional architecture. It iteratively upsamples the voxel features with text-guided pruning (TGP), and fuses multi-level features via completion-based addition (CBA). (a) to (d) on the right side illustrate various options for feature upsampling. (a) refers to simple concatenation with text features, which is fast but less accurate. (b) refers to feature interaction through cross-modal attention mechanisms, which is constrained by the large number of voxels. (c) represents our proposed TGP, which first prunes voxel features under textual guidance and thus enables efficient interaction between voxel and text features. (d) shows a simplified version of TGP that removes farthest point sampling and interpolation, combines multi-modal feature interactions into a whole and moves it before pruning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.419, + 0.484, + 0.616 + ], + "angle": 0, + "content": "voxels is not too large yet, TGP prunes less voxels. While at lower level where the number of voxels is significantly increased by upsampling operation, TGP prunes the voxel features more aggressively. The multi-level architecture of TSP3D consists of three levels and includes two feature upsampling operations. Therefore, we correspondingly configure two TGPs with different functions, which are referred as scene-level TGP (level 3 to 2) and target-level TGP (level 2 to 1) respectively. Scene-level TGP aims to distinguish between objects and the background, specifically pruning the voxels on background. Target-level TGP focuses on regions mentioned in the text, intending to preserve the target object and referential objects while removing other regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.783 + ], + "angle": 0, + "content": "Details of TGP. Since the pruning is relevant to the description, we need to make the voxel features text-aware to predict a proper pruning mask. To reduce the computational cost, we perform farthest point sampling (FPS) on the voxel features to reduce their size while preserving the basic distribution of the scene. Next, we utilize cross-attention to interact with the text features and employ a simple MLP to predict the probability distribution \\(\\hat{M}\\) for retaining each voxel. To prune the features \\(U_{l}\\), we binarize and interpolate the \\(\\hat{M}\\) to obtain the pruned mask. This process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.795, + 0.482, + 0.814 + ], + "angle": 0, + "content": "\\[\nU _ {l} ^ {P} = U _ {l} \\odot \\Theta (\\mathcal {I} (\\hat {M}, U _ {l}) - \\sigma) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.828, + 0.483, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\hat {M} = \\operatorname {M L P} (\\text {C r o s s A t t} (\\operatorname {F P S} (U _ {l}), \\operatorname {S e l f A t t} (T))) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(U_{l}^{P}\\) is the pruned features, \\(\\Theta\\) is Heaviside step function, \\(\\odot\\) is matrix dot product, \\(\\sigma\\) is the pruning threshold, and \\(\\mathcal{I}\\) represents linear interpolation based on the positions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.526 + ], + "angle": 0, + "content": "specified by \\( U_{l} \\). After pruning, the scale of the scene features is significantly reduced, enabling internal feature interactions based on self-attention. Subsequently, we utilize self-attention and cross-attention to perceive the relative relationships among objects within the scene and to fuse multimodal features, resulting in updated features \\( U_{l}^{\\prime} \\). Finally, through generative sparse convolutions, we obtain \\( U_{l - 1}^{G} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.527, + 0.906, + 0.603 + ], + "angle": 0, + "content": "Supervision for Pruning. The binary supervision mask \\( M^{sce} \\) for scene-level TGP is generated based on the centers of all objects in the scene, and the mask \\( M^{tar} \\) for target-level TGP is based on the target and relevant objects mentioned in the descriptions:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.671 + ], + "angle": 0, + "content": "\\[\nM ^ {s c e} = \\bigcup_ {i = 1} ^ {N} \\mathcal {M} \\left(O _ {i}\\right), \\quad M ^ {t a r} = \\mathcal {M} \\left(O ^ {t a r}\\right) \\cup \\bigcup_ {j = 1} ^ {K} \\mathcal {M} \\left(O _ {j} ^ {r e l}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.763 + ], + "angle": 0, + "content": "where \\(\\{O_i|1\\leq i\\leq N\\}\\) indicates all objects in the scene. \\(O^{tar}\\) and \\(O^{rel}\\) refer to target and relevant objects respectively. \\(\\mathcal{M}(O)\\) represents the mask generated from the center of object \\(O\\). It generates a \\(L\\times L\\times L\\) cube centered at the center of \\(O\\) to construct the supervision mask \\(M\\), where locations inside the cube is set to 1 while others set to 0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Simplification. Although the above mentioned method can effectively prune voxel features to reduce the computational cost of cross-modal attention, there are some inefficient operations in the pipeline: (1) FPS is time-consuming, especially for large scenes; (2) there are two times of interactions between voxel features and text features, the first is to guide pruning and the second is to enhance the representation, which is a bit redundant. We also empirically observe that the number of voxels is not large in level 3." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3669" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.287, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.182, + 0.246, + 0.193 + ], + "angle": 0, + "content": "(a) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.457, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.182, + 0.422, + 0.193 + ], + "angle": 0, + "content": "(b) Pruned Features" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.2, + 0.289, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.292, + 0.265, + 0.304 + ], + "angle": 0, + "content": "(c) Completion Features" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.2, + 0.456, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.292, + 0.406, + 0.304 + ], + "angle": 0, + "content": "(d) Prediction" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.314, + 0.47, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Figure 3. Illustration of completion-based addition. The upper figure (b) illustrates an example of over-pruning on the target. The lower figure (c) shows the completed features predicted by CBA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.627 + ], + "angle": 0, + "content": "To this end, we propose a simplified version of TGP, as shown in Fig. 2 (d). We remove the FPS and merge the two multi-modal interactions into one. We also move the merged interaction operation before pruning. In this way, voxel features and text features are first deeply interacted for both feature enhancement and pruning. Because in level 3 the number of voxels is small and in level \\(2/1\\) the voxels are already pruned, the computational cost of self-attention and cross-attention is always kept at a relatively low level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Effectiveness of TGP. After pruning, the voxel count of \\( U_{1} \\) is reduced to nearly \\( 7\\% \\) of its original size without TGP, while the 3DVG performance is significantly boosted. TGP serves multiple functions, including: (1) facilitating the interaction of multi-modal features through cross-attention, (2) reducing the feature amount (number of voxels) through pruning, and (3) gradually guiding the network to focus on the mentioned target based on text features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.34, + 0.774 + ], + "angle": 0, + "content": "3.3. Completion-based Addition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.87 + ], + "angle": 0, + "content": "During the pruning process, some targets may be mistakenly removed, especially for small or narrow objects, as shown in Fig. 3 (b). Therefore, the addition operation between the upsampled pruned features \\( U_{l}^{G} \\) and backbone features \\( V_{l} \\) described in Equation (1) play an important role to mitigate the affect of over-pruning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "There are two alternative addition operations: (1) Full Addition. For the intersecting regions of \\( V_{l} \\) and \\( U_{l}^{G} \\), fea-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.332 + ], + "angle": 0, + "content": "tures are directly added. For voxel features outside the intersection of \\( U_{l}^{G} \\) and \\( V_{l} \\) which lack corresponding features in the other map, the missing voxel features are interpolated before addition. Due to pruning process, \\( U_{l}^{G} \\) is sparser than \\( V_{l} \\). In this way, full addition can fix almost all the pruned region. But this operation is computationally heavy and make the scene representation fail to focus on relevant objects, which deviates the core idea of TGP. (2) Pruning-aware Addition. The addition is constrained to the locations of \\( U_{l}^{G} \\). For voxel in \\( U_{l}^{G} \\) but not in \\( V_{l} \\), interpolation from \\( U_{l}^{G} \\) is applied to complete the missing locations in \\( V_{l} \\). It restricts the addition operation to the shape of the pruned features, potentially leading to an over-reliance on the results of the pruning process. If important regions are over-pruned, the network may struggle to detect targets with severely damaged geometric information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.905, + 0.438 + ], + "angle": 0, + "content": "Considering the unavoidable risk of pruning the query target, we introduce the completion-based addition (CBA). CBA is designed to address the limitations of full and pruning-aware additions. It offers a more targeted and efficient way to integrating multi-level features, ensuring the preservation of essential details while keeping the additional computational overhead negligible." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.905, + 0.499 + ], + "angle": 0, + "content": "Details of CBA. We first enhance the backbone features \\( V_{l} \\) with the text features \\( T \\) through cross-attention, obtaining \\( V_{l}^{\\prime} \\). Then a MLP is adopted to predict the probability distribution of target for region selection:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.504, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\nM _ {l} ^ {t a r} = \\Theta (\\mathrm {M L P} \\left(V _ {l} ^ {\\prime}\\right) - \\tau) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.905, + 0.603 + ], + "angle": 0, + "content": "where \\(\\Theta\\) is the step function, and \\(\\tau\\) is the threshold determining voxel relevance. \\(M_{l}^{tar}\\) is a binary mask indicating potential regions of the mentioned target. Then, comparison of \\(M_{l}^{tar}\\) with \\(U_{l}\\) identifies missing voxels. The missing mask \\(M_{l}^{mis}\\) is derived as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.607, + 0.905, + 0.625 + ], + "angle": 0, + "content": "\\[\nM _ {l} ^ {m i s} = M _ {l} ^ {t a r} \\wedge (\\neg \\mathcal {C} (U _ {l} ^ {G}, V _ {l})) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.905, + 0.739 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}(A,B)\\) denotes the generation of a binary mask for \\(A\\) based on the shape of \\(B\\). Specifically, for positions in \\(B\\), if there are corresponding voxel features in \\(A\\), the mask for that position is set to 1. Otherwise it is set to 0. Missed voxel features in \\(U_{l}^{G}\\) that correspond to \\(M_{l}^{mis}\\) are interpolated from \\(U_{l}^{G}\\), filling in gaps identified by the missing mask. The completed feature map \\(U_{l}^{cpl}\\) is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.743, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\nU _ {l} ^ {c p l} = V _ {l} ^ {\\prime} \\odot M _ {l} ^ {m i s} + \\mathcal {I} \\left(U _ {l} ^ {G}, M _ {l} ^ {m i s}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.768, + 0.905, + 0.844 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) represents linear interpolation on the feature map based on the positions specified in the mask. Finally, the original upsampled features are combined with the backbone features according to the pruning-aware addition, and merged with the completion features to yield updated \\(U_{l}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.847, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\nU _ {l} = \\operatorname {C o n c a t} \\left(U _ {l} ^ {G} \\leftarrow V _ {l}, U _ {l} ^ {c p l}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "where \\(\\leftarrow\\) denotes the pruning-aware addition, and Concat means concatenation of voxel features." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3670" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.174 + ], + "angle": 0, + "content": "Table 1. Comparison of methods on the ScanRefer dataset evaluated at IoU thresholds of 0.25 and 0.5. TSP3D achieves state-of-the-art accuracy even compared with two-stage methods, with +1.13 lead on Acc@0.5. Notably, we are the first to comprehensively evaluate inference speed for 3DVG methods. The inference speeds of other methods are obtained through our reproduction." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.178, + 0.482, + 0.452 + ], + "angle": 0, + "content": "
MethodVenueInputAccuracyInference Speed (FPS)
0.250.5
Two-Stage Model
ScanRefer [3]ECCV'203D+2D41.1927.406.72
TGNN [14]AAAI'213D37.3729.703.19
InstanceRefer [39]ICCV'213D40.2330.152.33
SAT [38]ICCV'213D+2D44.5430.144.34
FFL-3DOG [10]ICCV'213D41.3334.01Not released
3D-SPS [22]CVPR'223D+2D48.8236.983.17
BUTD-DETR [16]ECCV'223D50.4238.603.33
EDA [35]CVPR'233D54.5942.263.34
3D-VisTA [41]ICCV'233D45.9041.502.03
VPP-Net [31]CVPR'243D55.6543.29Not released
G3-LQ [34]CVPR'243D56.9045.58Not released
MCLN [27]ECCV'243D57.1745.533.17
Single-stage Model
3D-SPS [22]CVPR'223D47.6536.435.38
BUTD-DETR [16]ECCV'223D49.7637.055.91
EDA [35]CVPR'233D53.8341.705.98
G3-LQ [34]CVPR'243D55.9544.72Not released
MCLN [27]ECCV'243D54.3042.645.45
TSP3D (Ours)—3D56.4546.7112.43
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.21, + 0.483 + ], + "angle": 0, + "content": "3.4. Train Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.641 + ], + "angle": 0, + "content": "The loss is composed of several components: pruning loss for TGP, completion loss for CBA, and objectness loss as well as bounding box regression loss for the head. Pruning loss, completion loss and objectness loss employ the focal loss to handle class imbalance. Supervision for completion and classification losses are the same, which sets voxels near the target object center as positives while leaving others as negatives. For bounding box regression, we use the Distance-IoU (D IoU) loss. The total loss function is computed as the sum of these individual losses:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.651, + 0.454, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {p r u n i n g}} + \\lambda_ {2} \\mathcal {L} _ {\\text {c o m}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c l a s s}} + \\lambda_ {4} \\mathcal {L} _ {\\text {b b o x}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.676, + 0.474, + 0.692 + ], + "angle": 0, + "content": "where \\(\\lambda_{1},\\lambda_{2},\\lambda_{3}\\) and \\(\\lambda_4\\) are the weights of different parts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.222, + 0.72 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.192, + 0.743 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We maintain the same experimental settings with previous works, employing ScanRefer [3] and SR3D/NR3D [2] as datasets. ScanRefer: Built on the ScanNet framework, ScanRefer includes 51,583 descriptions across scenes. Evaluation metrics focus on Acc@mIoU. ReferIt3D: ReferIt3D splits into Nr3D, with 41,503 human-generated descriptions, and Sr3D, containing 83,572 synthetic expressions. ReferIt3D simplifies the task by providing segmented point clouds for each object. The primary evaluation metric is accuracy in target object selection." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.16 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on Nr3D and Sr3D datasets. We evaluate under three pipelines, noting that the Two-stage using Ground-Truth Boxes is impractical for real-world applications. TSP3D exhibits significant superiority, with leads of \\(+2.6\\%\\) and \\(+3.2\\%\\) on NR3D and SR3D respectively." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.164, + 0.898, + 0.428 + ], + "angle": 0, + "content": "
MethodVenuePipelineAccuracy
Nr3DSr3D
InstanceRefer [39]ICCV'21Two-stage (gt)38.848.0
LanguageRefer [28]CoRL'22Two-stage (gt)43.956.0
3D-SPS [22]CVPR'22Two-stage (gt)51.562.6
MVT [15]CVPR'22Two-stage (gt)55.164.5
BUTD-DETR [16]ECCV'22Two-stage (gt)54.667.0
EDA [35]CVPR'23Two-stage (gt)52.168.1
VPP-Net [31]CVPR'24Two-stage (gt)56.968.7
G3-LQ [34]CVPR'24Two-stage (gt)58.473.1
MCLN [27]ECCV'24Two-stage (gt)59.868.4
InstanceRefer [39]ICCV'21Two-stage (det)29.931.5
LanguageRefer [28]CoRL'22Two-stage (det)28.639.5
BUTD-DETR [16]ECCV'22Two-stage (det)43.352.1
EDA [35]CVPR'23Two-stage (det)40.749.9
MCLN [27]ECCV'24Two-stage (det)46.153.9
3D-SPS [22]CVPR'22Single-stage39.247.1
BUTD-DETR [16]ECCV'22Single-stage38.750.1
EDA [35]CVPR'23Single-stage40.049.7
MCLN [27]ECCV'24Single-stage45.753.4
TSP3D (Ours)—Single-stage48.757.1
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.441, + 0.73, + 0.457 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.613 + ], + "angle": 0, + "content": "TSP3D is implemented based on PyTorch [23]. The pruning thresholds are set at \\(\\sigma_{\\mathrm{sc}} = 0.7\\) and \\(\\sigma_{\\mathrm{tar}} = 0.3\\), and the completion threshold in CBA is \\(\\tau = 0.15\\). The initial voxelization of the point cloud has a voxel size of \\(1\\mathrm{cm}\\), while the voxel size for level \\(i\\) features scales to \\(2^{i + 2}\\mathrm{cm}\\). The supervision for pruning uses \\(L = 7\\). The weights for all components of the loss function, \\(\\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4\\), are equal to 1. Training is conducted using four GPUs, while inference speeds are evaluated using a single consumer-grade GPU, RTX 3090, with a batch size of 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.753, + 0.639 + ], + "angle": 0, + "content": "4.3. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Performance on ScanRefer. We carry out comparisons with existing methods on ScanRefer, as detailed in Tab. 1. The inference speeds of other methods are obtained through our reproduction with a single RTX 3090 and a batch size of 1. For two-stage methods, the inference speed includes the time taken for object detection in the first stage. For methods using 2D image features and 3D point clouds as inputs, we do not account for the time spent extracting 2D features, assuming they can be obtained in advance. However, in practical applications, the acquisition of 2D features also impacts overall efficiency. TSP3D achieves state-of-the-art accuracy even compared with two-stage methods, with \\(+1.13\\) lead on Acc@0.5. Notably, in the single-stage setting, TSP3D achieves fast inference speed, which is unprecedented among the existing methods. This significant improvement is attributed to our method's efficient use of a multi-level architecture based on 3D sparse convolu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3671" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.338, + 0.116 + ], + "angle": 0, + "content": "Table 3. Impact of the proposed TGP and CBA. Evaluated on ScanRefer." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.123, + 0.337, + 0.195 + ], + "angle": 0, + "content": "
IDTGPCBAAccuracySpeed (FPS)
0.250.5
(a)40.1332.8714.58
(b)✓55.2046.1513.22
(c)✓41.3433.0913.51
(d)✓✓56.4546.7112.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.346, + 0.09, + 0.618, + 0.117 + ], + "angle": 0, + "content": "Table 4. Influence of the two CBAs at different levels. Evaluated on ScanRefer." + }, + { + "type": "table", + "bbox": [ + 0.348, + 0.123, + 0.616, + 0.194 + ], + "angle": 0, + "content": "
IDCBA (level 2)CBA (level 1)AccuracySpeed (FPS)
0.250.5
(a)55.2046.1513.22
(b)✓55.1746.0612.79
(c)✓56.4546.7112.43
(d)✓✓56.2246.6812.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.625, + 0.091, + 0.901, + 0.118 + ], + "angle": 0, + "content": "Table 5. Influence of different feature upsampling methods. Evaluated on ScanRefer." + }, + { + "type": "table", + "bbox": [ + 0.629, + 0.124, + 0.899, + 0.194 + ], + "angle": 0, + "content": "
IDMethodAccuracySpeed (FPS)
0.250.5
(a)Simple concatenation40.1332.8714.58
(b)Attention mechanism———
(c)Text-guided pruning56.2746.5810.11
(d)Simplified TGP56.4546.7112.43
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.209, + 0.482, + 0.3 + ], + "angle": 0, + "content": "tions, coupled with the text-guided pruning. By focusing computation only on salient regions of the point clouds, determined by textual cues, our model effectively reduces computational overhead while maintaining high accuracy. TSP3D also sets a benchmark for inference speed comparisons for future methodologies." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.302, + 0.485, + 0.604 + ], + "angle": 0, + "content": "Performance on Nr3D/Sr3D. We evaluate our method on the SR3D and NR3D datasets, following the evaluation protocols of prior works like EDA [35] and BUTD-DETR [16] by using Acc@0.25 as the accuracy metric. The results are shown in Tab. 2. Given that SR3D and NR3D provide ground-truth boxes and categories for all objects in the scene, we consider three pipelines: (1) Two-stage using Ground-Truth Boxes, (2) Two-stage using Detected Boxes, and (3) Single-stage. In practical applications, the Two-stage using Ground-Truth Boxes pipeline is unrealistic because obtaining all ground-truth boxes in a scene is infeasible. This approach can also oversimplify certain evaluation scenarios. For example, if there are no other objects of the same category as the target in the scene, the task reduces to relying on the provided ground-truth category. Under the Single-stage setting, TSP3D exhibits significant superiority with peak performance of \\(48.7\\%\\) and \\(57.1\\%\\) on Nr3D and Sr3D. TSP3D even outperforms previous works under the pipeline of Two-stage using Detected Boxes, with leads of \\(+2.6\\%\\) and \\(+3.2\\%\\) on NR3D and SR3D." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.62, + 0.245, + 0.637 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Proposed Components. To investigate the effects of our proposed TGP and CBA, we conduct ablation experiments with module removal as shown in Tab. 3. When TGP is not used, multi-modal feature concatenation is employed as a replacement, as shown in Fig. 2 (a). When CBA is not used, it is substituted with a pruning-based addition. The results demonstrate that TGP significantly enhances performance without notably impacting inference time. This is because TGP, while utilizing a more complex multi-modal attention mechanism for stronger feature fusion, significantly reduces feature scale through text-guided pruning. Additionally, the performance improvement is also due to the gradual guidance towards the target object by both scene-level and target-level TGP. Using CBA alone has a limited effect, as no voxels are pruned. Implementing CBA on top of TGP further enhances performance, as CBA dynamically compensates for some of the excessive" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.21, + 0.898, + 0.225 + ], + "angle": 0, + "content": "pruning by TGP, thus increasing the network's robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.226, + 0.905, + 0.468 + ], + "angle": 0, + "content": "Influence of the Two CBAs. To explore the impact of CBAs at two different levels, we conduct ablation experiments as depicted in Tab. 4. In the absence of CBA, we use pruning-based addition as a substitute. The results indicate that the CBA at level 2 has negligible effects on the 3DVG task. This is primarily because the CBA at level 2 serves to supplement the scene-level TGP, which is expected to prune the background (a relatively simple task). Moreover, although some target features are pruned, they are compensated by two subsequent generative sparse convolutions. However, the CBA at level 1 enhances performance by adapt completion for the target-level TGP. It is challenging to fully preserve target objects from deep upsampling features, especially for smaller or narrower targets. The CBA at level 1, based on high-resolution backbone features, effectively complements the TGP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.697 + ], + "angle": 0, + "content": "Feature Upsampling Techniques. We conduct experiments to assess the effects of different feature upsampling techniques, as detailed in Tab. 5. Using simple feature concatenation (Fig. 2 (a)), while fast in inference speed, results in poor performance. When we utilize an attention mechanism with stronger feature interaction, as shown in Fig. 2 (b), the computation exceeds the limits of GPU due to the large number of voxels, making it impractical for real-world applications. Consequently, we employ TGP to reduce the feature amount, as illustrated in Fig. 2 (c), which significantly improves performance and enables practical deployment. Building on TGP, we propose simplified TGP, as shown in Fig. 2 (d), that merges feature interactions before and after pruning, achieving performance consistent with the original TGP while enhancing inference speed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.696, + 0.727 + ], + "angle": 0, + "content": "4.5. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Text-guided Pruning. To visually demonstrate the process of TGP, we visualize the results of two pruning phases, as shown in Fig. 4. In each example, the voxel features after scene-level pruning, the features after target-level pruning, and the features after target-level generative sparse convolution are displayed from top to bottom. It is evident that both pruning stages effectively achieve our intended effect: the scene-level pruning filters out the background and retained object voxels, and the target-level pruning preserves relevant and target objects. Moreover, during the feature upsampling process, the feature amount nearly exponen-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.089, + 0.282, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.173, + 0.28, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.254, + 0.28, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.338, + 0.287, + 0.372 + ], + "angle": 0, + "content": "(a) It is a gray trash can. \nThe trash can sits in the corner by where the tv is." + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.089, + 0.493, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.173, + 0.493, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.253, + 0.493, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.335, + 0.481, + 0.377 + ], + "angle": 0, + "content": "(b) The chair is on the northwest side of the table that is furthest northeast in the room. The chair has a curved backside and four legs." + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.676, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.171, + 0.676, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.253, + 0.676, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.339, + 0.641, + 0.371 + ], + "angle": 0, + "content": "(c) A chrome water fountain. Is the lowest of two water fountains." + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.089, + 0.839, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.173, + 0.835, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.255, + 0.835, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.338, + 0.825, + 0.371 + ], + "angle": 0, + "content": "(d) There is a rectangular shelf. It is the one closest to the curtain." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.381, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Figure 4. Visualization of the text-guided pruning process. In each example, the voxel features after scene-level TGP, target-level TGP and the last upsampling layer are presented from top to bottom. The blue boxes represent the ground truth of the target, and the red boxes denote the bounding boxes of relevant objects. TSP3D reduces the amount of voxel features through two stages of pruning and progressively guides the network focusing towards the target." + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.448, + 0.342, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.543, + 0.329, + 0.576 + ], + "angle": 0, + "content": "(a) The brown square trash can is to the right of the lamp. It is also behind the chair and to the right." + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.449, + 0.5, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.543, + 0.498, + 0.575 + ], + "angle": 0, + "content": "(b) The whiteboard is next to two doors. The whiteboard is a long, white rectangle." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.449, + 0.691, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.541, + 0.69, + 0.573 + ], + "angle": 0, + "content": "(c) This is a monitor in black color. The monitor is next to the wall and at the back of maroon chair." + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.449, + 0.828, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.542, + 0.842, + 0.577 + ], + "angle": 0, + "content": "(d) The book rack is left of the bookshelf that looks like a right angle with another. The book rack is black and has as circular bottom." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.581, + 0.907, + 0.624 + ], + "angle": 0, + "content": "Figure 5. Visualization of the completion-based addition process. The blue points represent the voxel features output by the target-level TGP, while the red points are the completion features predicted by the CBA. The blue boxes indicate the ground truth boxes. CBA adaptively supplements situations where excessive pruning has occurred." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.646, + 0.483, + 0.722 + ], + "angle": 0, + "content": "tially increases due to generative upsampling. Without TGP, the voxel coverage would far exceed the range of the scene point cloud, which is inefficient for inference. This also intuitively explains the significant impact of our TGP on both performance and inference speed." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Completion-based Addition. To clearly illustrate the function of CBA, we visualize the adaptive completion process in Fig. 5. The images below showcase several instances of excessive pruning. TGP performs pruning based on deep and low-resolution features, which can lead to excessive pruning, potentially removing entire or partial targets. This over-pruning is more likely to occur with small, as shown in Fig. 5 (a) and (c), narrow, as in Fig. 5 (b), or elongated targets, as in Fig. 5 (d). Our CBA effectively supplements the process using higher-resolution backbone features, thus dynamically integrating multi-level features." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.645, + 0.634, + 0.661 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present TSP3D, an efficient sparse single-stage method for real-time 3D visual grounding. Different from previous 3D visual grounding frameworks, TSP3D builds on multi-level sparse convolutional architecture for efficient and fine-grained scene representation extraction. To enable the interaction between voxel features and textual features, we propose text-guided pruning (TGP), which reduces the amount of voxel features and guides the network to progressively focus on the target object. Additionally, we introduce completion-based addition (CBA) for adaptive multi-level feature fusion, effectively compensating for instances of over-pruning. Extensive experiments demonstrate the effectiveness of our proposed modules, resulting in an efficient 3DVG method that achieves state-of-the-art accuracy and fast inference speed." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3673" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 624B2076, Grant 62376132, Grant 62321005, and Grant 62441616." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.188, + 0.192 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[1] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dref-transformer: Fine-grained object identification in real-world scenes using natural language. In WACV, pages 3941-3950, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.272, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[2] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.33, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[3] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[4] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In ICCV, pages 15467-15476, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.484, + 0.484 + ], + "angle": 0, + "content": "[5] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In CVPR, pages 21674-21683, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[7] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, pages 1201-1209, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[9] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[10] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In ICCV, pages 3722-3731, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.484, + 0.813 + ], + "angle": 0, + "content": "[11] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224–9232, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.484, + 0.869 + ], + "angle": 0, + "content": "[12] JunYoung Gwak, Christopher Choy, and Silvio Savarese. Generative sparse detection networks for 3d single-shot object detection. In ECCV, pages 297-313. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.201, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "relation aware transformer for fine-grained 3d visual grounding. In ACM MM, pages 2344-2352, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[14] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[15] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.905, + 0.279 + ], + "angle": 0, + "content": "[16] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.335 + ], + "angle": 0, + "content": "[17] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In CVPR, pages 4867-4876, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.407 + ], + "angle": 0, + "content": "[18] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.411, + 0.905, + 0.465 + ], + "angle": 0, + "content": "[19] Daizong Liu, Yang Liu, Wencan Huang, and Wei Hu. A survey on text-guided 3d visual grounding: Elements, recent advances, and future directions. arXiv preprint arXiv:2406.05785, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.468, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[20] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[21] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, pages 2949-2958, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[22] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.905, + 0.667 + ], + "angle": 0, + "content": "[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[24] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, pages 1532-1543, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[25] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[26] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[27] Zhipeng Qian, Yiwei Ma, Zhekai Lin, Jiayi Ji, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Multi-branch collaborative learning network for 3d visual grounding. In ECCV, pages 381–398. Springer, 2025. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In CoRL, pages 1046-1056. PMLR, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3674" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In ECCV, pages 477-493. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[30] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Tr3d: Towards real-time indoor 3d object detection. In ICIP, pages 281-285. IEEE, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[31] Xiangxi Shi, Zhonghua Wu, and Stefan Lee. Viewpoint-aware visual grounding in 3d scenes. In CVPR, pages 14056-14065, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[32] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[33] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. NeurIPS, 35:29975-29988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[34] Yuan Wang, Yali Li, and Shengjin Wang. G^3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. In CVPR, pages 13917-13926, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[36] Xiuwei Xu, Ziwei Wang, Jie Zhou, and Jiwen Lu. Binarizing sparse convolutional networks for efficient point cloud analysis. In CVPR, pages 5313-5322, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[37] Xiuwei Xu, Zhihao Sun, Ziwei Wang, Hongmin Liu, Jie Zhou, and Jiwen Lu. 3d small object detection with dynamic spatial pruning. In ECCV. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[38] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[39] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In ICCV, pages 1791-1800, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[40] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvgtransformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[41] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3675" + } + ] +] \ No newline at end of file diff --git a/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_origin.pdf b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..397f3fc2feed2490eb893d7f1f742353b15a2a04 --- /dev/null +++ b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/4a2bb82b-b995-4ea1-8fd6-89e0dbc418a3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed5909ca4628ecb47ab9e752efd1721bf8e682546bbaf7c8a6f9c6d9809b052c +size 9762077 diff --git a/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/full.md b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c1f3751fb3a051e67d1c38eadeda356bc6a8f211 --- /dev/null +++ b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/full.md @@ -0,0 +1,334 @@ +# Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding + +Wenxuan Guo $^{1*}$ Xiuwei Xu $^{1*}$ Ziwei Wang $^{2}$ Jianjiang Feng $^{1\dagger}$ Jie Zhou $^{1}$ Jiwen Lu $^{1\dagger}$ Tsinghua University $^{2}$ Nanyang Technological University + +$\{gwx22,xxw21\} @mails.tsinghua.edu.cn ziwei.wang@ntu.edu.sg$ + +{jfeng,jzhou,lujiwen}@tsinghua.edu.cn + +# Abstract + +In this paper, we propose an efficient multi-level convolution architecture for 3D visual grounding. Conventional methods are difficult to meet the requirements of real-time inference due to the two-stage or point-based architecture. Inspired by the success of multi-level fully sparse convolutional architecture in 3D object detection, we aim to build a new 3D visual grounding framework following this technical route. However, as in 3D visual grounding task the 3D scene representation should be deeply interacted with text features, sparse convolution-based architecture is inefficient for this interaction due to the large amount of voxel features. To this end, we propose text-guided pruning (TGP) and completion-based addition (CBA) to deeply fuse 3D scene representation and text features in an efficient way by gradual region pruning and target completion. Specifically, TGP iteratively sparsifies the 3D scene representation and thus efficiently interacts the voxel features with text features by cross-attention. To mitigate the affect of pruning on delicate geometric information, CBA adaptively fixes the over-pruned region by voxel completion with negligible computational overhead. Compared with previous single-stage methods, our method achieves top inference speed and surpasses previous fastest method by $100\%$ FPS. Our method also achieves state-of-the-art accuracy even compared with two-stage methods, with $+1.13$ lead of Acc@0.5 on ScanRefer, and $+2.6$ and $+3.2$ leads on NR3D and SR3D respectively. The code is available at https://github.com/GWxuan/TSP3D. + +# 1. Introduction + +Incorporating multi-modal information to guide 3D visual perception is a promising direction. In these years, 3D visual grounding (3DVG), also known as 3D instance referencing, has been paid increasing attention as a fundamental multi-modal 3D perception task. The aim of 3DVG is + +![](images/661c935b6d6e9132a512ca6ed5b2382b6b4416e336bf98bb467a08e4aab6aed4.jpg) +Figure 1. Comparison of 3DVG methods on ScanRefer dataset [3]. Our TSP3D surpasses existing methods in both accuracy and inference speed, achieving the first efficient 3DVG framework. + +to locate an object in the scene with a free-form query description. 3DVG is challenging since it requires understanding of both 3D scene and language description. Recently, with the development of 3D scene perception and vision-language models, 3DVG methods have shown remarkable progress [16, 22]. However, with 3DVG being widely applied in fields like robotics and AR / VR where inference speed is the main bottleneck, how to construct efficient real-time 3DVG model remains a challenging problem. + +Since the output format of 3DVG is similar with 3D object detection, early 3DVG methods [3, 14, 38, 39] usually adopt a two-stage framework, which first conducts detection to locate all objects in the scene, and then selects the target object by incorporating text information. As there are many similarities between 3D object detection and 3DVG (e.g. both of them need to extract the representation of the 3D scene), there will be much redundant feature computation during the independent adoption of the two models. As a result, two-stage methods are usually hard to handle real-time tasks. To solve this problem, single-stage methods [22, 35] are presented, which generates the bounding box of the target directly from point clouds. This integrated design is more compact and efficient. However, + +current single-stage 3DVG methods mainly build on point-based architecture [25], where the feature extraction contains time-consuming operations like furthest point sampling and kNN. They also need to aggressively downsample the point features to reduce computational cost, which might hurt the geometric information of small and thin objects [37]. Due to these reasons, current single-stage methods are still far from real-time ( $< 6$ FPS) and their performance is inferior to two-stage methods, as shown in Fig. 1. + +In this paper, we propose a new single-stage framework for 3DVG based on text-guided sparse voxel pruning, namely TSP3D. Inspired by state-of-the-art 3D object detection methods [29, 37] which achieves both leading accuracy and speed with multi-level sparse convolutional architecture, we build the first sparse single-stage 3DVG network. However, different from 3D object detection, in 3DVG the 3D scene representation should be deeply interacted with text features. Since the count of voxels is very large in sparse convolution-based architecture, deep multi-modal interaction like cross-attention becomes infeasible due to unaffordable computational cost. To this end, we propose text-guided pruning (TGP), which first utilize text information to jointly sparsify the 3D scene representation and enhance the voxel and text features. To mitigate the affect of pruning on delicate geometric information, we further present completion-based addition (CBA) to adaptively fix the over-pruned region with negligible computational overhead. Specifically, TGP prunes the voxel features according to the object distribution. It gradually removes background features and features of irrelevant objects, which generates text-aware voxel features around the target object for accurate bounding box prediction. Since pruning may mistakenly remove the representation of target object, CBA utilizes text features to query a small set of voxel features from the complete backbone features, followed by pruned-aware addition to fix the over-pruned region. We conduct extensive experiments on the popular ScanRefer [3] and ReferIt3D [2] datasets. Compared with previous single-stage methods, TSP3D achieves top inference speed and surpasses previous fastest single-stage method by $100\%$ FPS. TSP3D also achieves state-of-the-art accuracy even compared with two-stage methods, with $+1.13$ lead of Acc@0.5 on ScanRefer, and $+2.6$ and $+3.2$ leads on NR3D and SR3D respectively. + +To summarize, our main contributions are as follows: + +- To the best of our knowledge, this is the first work exploring sparse convolutional architecture for efficient 3DVG. +- To enable efficient feature extraction, we propose text-guided pruning and completion-based addition to sparsify sparse voxels and adaptively fuse multi-level features. +- We conduct extensive experiments, and TSP3D outperforms existing methods in both accuracy and speed, demonstrating the superiority of the proposed framework. + +# 2. Related Work + +# 2.1.3D Visual Grounding + +3D visual grounding aims to locate a target object within a 3D scene based on natural language descriptions [19]. Existing methods are typically categorized into two-stage and single-stage approaches. Two-stage methods follow a detect-then-match paradigm. In the first stage, they independently extract features from the language query using pre-trained language models [7, 9, 24] and predict candidate 3D objects using pre-trained 3D detectors [21, 26] or segmenters [4, 17, 32]. In the second stage, they focus on aligning the vision and text features to identify the target object. Techniques for feature fusion include attention mechanisms with Transformers [13, 40], contrastive learning [1], and graph-based matching [10, 14, 39]. In contrast, single-stage methods integrate object detection and feature extraction, allowing for direct identification of the target object. Methods in this category include guiding keypoint selection using textual features [22], and measuring similarity between words and objects inspired by 2D image-language pre-trained models like GLIP [18], as in BUTDDETR [16]. And methods like EDA [35] and $\mathrm{G}^3$ -LQ [34] advance single-stage 3D visual grounding by enhancing multimodal feature discriminability through explicit textdecoupling, dense alignment, and semantic-geometric modeling. MCLN [27] uses the 3D referring expression segmentation task to assist 3DVG in improving performance. + +However, existing two-stage and single-stage methods generally have high computational costs, hindering real-time applications. Our work aims to address these efficiency challenges by proposing an efficient single-stage method with multi-level sparse convolutional architecture. + +# 2.2. Multi-Level Convolutional Architectures + +Recently, sparse convolutional architecture has achieved great success in the field of 3D object detection. Built on the voxel-based representation [5, 8, 33] and sparse convolution operation [6, 11, 36], this kind of methods show great efficiency and accuracy when processing scene-level data. GSDN [12] first adopts multi-level sparse convolution with generative feature upsampling in 3D object detection. FCAF3D [29] simplifies the multi-level architecture with anchor-free design, achieving leading accuracy and speed. TR3D [30] further accelerates FCAF3D by removing unnecessary layers and introducing category-aware proposal assignment method. Moreover, DSPDet3D [37] introduces the multi-level architecture to 3D small object detection. + +Our proposed method draws inspiration from these approaches, utilizing a sparse multi-level architecture with sparse convolutions and an anchor-free design. This allows for efficient processing of 3D data, enabling real-time performance in 3D visual grounding tasks. + +# 3. Method + +In this section, we describe our TSP3D for efficient single-stage 3DVG. We first analyze existing pipelines to identify current challenges and motivate our approach (Sec. 3.1). We then introduce the text-guided pruning, which leverages text features to guide feature pruning (Sec. 3.2). To address the potential risk of pruning key information, we propose the completion-based addition for multi-level feature fusion (Sec. 3.3). Finally, we detail the training loss (Sec. 3.4). + +# 3.1. Architecture Analysis for 3DVG + +Top-performance 3DVG methods [31, 34, 35], are mainly two-stage, which is a serial combination of 3D object detection and 3D object grounding. This separate calls of two approaches result in redundant feature extraction and complex pipeline, thus making the two-stage methods less efficient. To demonstrate the efficiency of existing methods, we conduct a comparison of accuracy and speed among several representative methods on ScanRefer [3], as shown in Fig. 1. It can be seen that two-stage methods struggle in speed (< 3 FPS) due to the additional detection stage. Since 3D visual grounding is usually adopted in practical scenarios that require real-time inference under limited resources, such as embodied robots and VR/AR, the low speed of two-stage methods make them less practical. On the other side, single-stage methods [22], which directly predicts referred bounding box from the observed 3D scene, are more suitable choices due to their streamlined processes. In Fig. 1, it can be observed that single-stage methods are significantly more efficient than their two-stage counterparts. + +However, existing single-stage methods are mainly built on point-based backbone [25], where the scene representation is extracted with time-consuming operations like furthest point sampling and set abstraction. They also employ large transformer decoder to fuse text and 3D features for several iterations. Therefore, the inference speed of current single-stage methods is still far from real-time ( $< 6$ FPS). The inference speed of specific components in different frameworks is analyzed and discussed in detail in the supplementary material. Inspired by the success of multi-level sparse convolutional architecture in 3D object detection [30], which achieves both leading accuracy and speed, we propose to build the first multi-level convolutional single-stage 3DVG pipeline. + +TSP3D-B. Here we propose a baseline framework based on sparse convolution, namely TSP3D-B. Following the simple and effective multi-level architecture of FCAF3D [29], TSP3D-B utilizes 3 levels of sparse convolutional blocks for scene representation extraction and bounding box prediction, as shown in Fig. 2 (a). Specifically, the input point clouds $P \in \mathbb{R}^{N \times 6}$ with 6-dim features (3D position and RGB) are first voxelized and then fed into three sequential MinkResBlocks [6], which generates three lev + +els of voxel features $V_{l}$ ( $l = 1,2,3$ ). With the increase of $l$ , the spatial resolution of $V_{l}$ decreases and the context information increases. Concurrently, the free-form text with $l$ words is encoded by the pre-trained RoBERTa [20] and produce the vanilla text tokens $T \in \mathbb{R}^{l \times d}$ . With the extracted 3D and text representations, we iteratively upsample $V_{3}$ and fuse it with $T$ to generate high-resolution and text-aware scene representation: + +$$ +U _ {l} = U _ {l} ^ {G} + V _ {l}, \quad U _ {l} ^ {G} = \operatorname {G e S p C o n v} \left(U _ {l + 1} ^ {\prime}\right) \tag {1} +$$ + +$$ +U _ {l + 1} ^ {\prime} = \operatorname {C o n c a t} \left(U _ {l + 1}, T\right) \tag {2} +$$ + +where $U_{3} = V_{3}$ , GeSpConv means generative sparse convolution [12] with stride 2, which upsamples the voxel features and expands their spatial locations for better bounding box prediction. Concat is voxel-wise feature concatenation by duplicating $T$ . The final upsampled feature map $U_{1}$ is concatenated with $T$ and fed into a convolutional head to predict the objectness scores and regress the 3D bounding box. We select the box with highest objectness score as the grounding result. + +As shown in Fig. 1, TSP3D-B achieves an inference speed of 14.58 FPS, which is significantly faster than previous single-stage methods and demonstrates great potential for real-time 3DVG. + +# 3.2. Text-guided Pruning + +Though efficient, TSP3D-B exhibits poor performance due to the inadequate interaction between 3D scene representation and text features. Motivated by previous 3DVG methods [16], a simple solution is to replace Concat with cross-modal attention to process voxel and text features, as shown in Fig. 2 (b). However, different from point-based architectures where the scene representation is usually aggressively downsampled, the number of voxels in multi-level convolutional framework is very large1. In practical implementation, we find that the voxels expand almost exponentially with each upsampling layer, leading to a substantial computational burden for the self-attention and cross-attention of scene features. To address this issue, we introduce text-guided pruning (TGP) to construct TSP3D, as illustrated in Fig. 2 (c). The core idea of TGP is to reduce feature amount by pruning redundant voxels and guide the network to gradually focus on the final target based on textual features. + +Overall Architecture. TGP can be regarded as a modified version of cross-modal attention, which reduces the number of voxels before attention operation, thereby lowering computational cost. To minimize the affect of pruning on the final prediction, we propose to prune the scene representation gradually. At higher level where the number of + +![](images/34dc88865730e7f3ee523802cfd25bf1b85a5d80e96e3034fcb7988d0ce6af56.jpg) +Figure 2. Illustration of TSP3D. TSP3D bulids on multi-level sparse convolutional architecture. It iteratively upsamples the voxel features with text-guided pruning (TGP), and fuses multi-level features via completion-based addition (CBA). (a) to (d) on the right side illustrate various options for feature upsampling. (a) refers to simple concatenation with text features, which is fast but less accurate. (b) refers to feature interaction through cross-modal attention mechanisms, which is constrained by the large number of voxels. (c) represents our proposed TGP, which first prunes voxel features under textual guidance and thus enables efficient interaction between voxel and text features. (d) shows a simplified version of TGP that removes farthest point sampling and interpolation, combines multi-modal feature interactions into a whole and moves it before pruning. + +![](images/6a003d30387be96c848bce3f9f41eacf06c8f47f83080b207f66b18616a8e5fc.jpg) + +![](images/69e156e52604f54c768f8d61e1521fe2c118da87ec724529954e0af845483d6d.jpg) + +![](images/28a30a9d670f1f2884a89d0d799fe6330b8eb72a993725475238135f0171de34.jpg) + +![](images/1c6a741f76761a4176f284700f1630ed9a640e23ff31c2818ea88d8de8dcb638.jpg) + +voxels is not too large yet, TGP prunes less voxels. While at lower level where the number of voxels is significantly increased by upsampling operation, TGP prunes the voxel features more aggressively. The multi-level architecture of TSP3D consists of three levels and includes two feature upsampling operations. Therefore, we correspondingly configure two TGPs with different functions, which are referred as scene-level TGP (level 3 to 2) and target-level TGP (level 2 to 1) respectively. Scene-level TGP aims to distinguish between objects and the background, specifically pruning the voxels on background. Target-level TGP focuses on regions mentioned in the text, intending to preserve the target object and referential objects while removing other regions. + +Details of TGP. Since the pruning is relevant to the description, we need to make the voxel features text-aware to predict a proper pruning mask. To reduce the computational cost, we perform farthest point sampling (FPS) on the voxel features to reduce their size while preserving the basic distribution of the scene. Next, we utilize cross-attention to interact with the text features and employ a simple MLP to predict the probability distribution $\hat{M}$ for retaining each voxel. To prune the features $U_{l}$ , we binarize and interpolate the $\hat{M}$ to obtain the pruned mask. This process can be expressed as: + +$$ +U _ {l} ^ {P} = U _ {l} \odot \Theta (\mathcal {I} (\hat {M}, U _ {l}) - \sigma) \tag {3} +$$ + +$$ +\hat {M} = \operatorname {M L P} (\text {C r o s s A t t} (\operatorname {F P S} (U _ {l}), \operatorname {S e l f A t t} (T))) \tag {4} +$$ + +where $U_{l}^{P}$ is the pruned features, $\Theta$ is Heaviside step function, $\odot$ is matrix dot product, $\sigma$ is the pruning threshold, and $\mathcal{I}$ represents linear interpolation based on the positions + +specified by $U_{l}$ . After pruning, the scale of the scene features is significantly reduced, enabling internal feature interactions based on self-attention. Subsequently, we utilize self-attention and cross-attention to perceive the relative relationships among objects within the scene and to fuse multimodal features, resulting in updated features $U_{l}^{\prime}$ . Finally, through generative sparse convolutions, we obtain $U_{l - 1}^{G}$ . + +Supervision for Pruning. The binary supervision mask $M^{sce}$ for scene-level TGP is generated based on the centers of all objects in the scene, and the mask $M^{tar}$ for target-level TGP is based on the target and relevant objects mentioned in the descriptions: + +$$ +M ^ {s c e} = \bigcup_ {i = 1} ^ {N} \mathcal {M} \left(O _ {i}\right), \quad M ^ {t a r} = \mathcal {M} \left(O ^ {t a r}\right) \cup \bigcup_ {j = 1} ^ {K} \mathcal {M} \left(O _ {j} ^ {r e l}\right) \tag {5} +$$ + +where $\{O_i|1\leq i\leq N\}$ indicates all objects in the scene. $O^{tar}$ and $O^{rel}$ refer to target and relevant objects respectively. $\mathcal{M}(O)$ represents the mask generated from the center of object $O$ . It generates a $L\times L\times L$ cube centered at the center of $O$ to construct the supervision mask $M$ , where locations inside the cube is set to 1 while others set to 0. + +Simplification. Although the above mentioned method can effectively prune voxel features to reduce the computational cost of cross-modal attention, there are some inefficient operations in the pipeline: (1) FPS is time-consuming, especially for large scenes; (2) there are two times of interactions between voxel features and text features, the first is to guide pruning and the second is to enhance the representation, which is a bit redundant. We also empirically observe that the number of voxels is not large in level 3. + +![](images/a4e77b9d9417a66cfcfba2ded9c11920b2046e1381c0234be6ffb54ea8195372.jpg) +(a) Ground Truth + +![](images/af41154e9f078452d9f30b783dcdc3e2b104a5135f64eaa6669dab57903d950c.jpg) +(b) Pruned Features + +![](images/f788cc08e641a9a6c6abe7e5b2ffe448156b551a3d61bc88f6b88ffd44276bde.jpg) +(c) Completion Features + +![](images/8d8555ac9d3a623ff5a061fce8bb0e71bfb42eb4ec4ec9177ff45f7b0a316da9.jpg) +(d) Prediction + +![](images/745c87df5c4898a5956aa422424bf8d0ac0bf15a5f089047e3d660191792c15b.jpg) +Figure 3. Illustration of completion-based addition. The upper figure (b) illustrates an example of over-pruning on the target. The lower figure (c) shows the completed features predicted by CBA. + +To this end, we propose a simplified version of TGP, as shown in Fig. 2 (d). We remove the FPS and merge the two multi-modal interactions into one. We also move the merged interaction operation before pruning. In this way, voxel features and text features are first deeply interacted for both feature enhancement and pruning. Because in level 3 the number of voxels is small and in level $2/1$ the voxels are already pruned, the computational cost of self-attention and cross-attention is always kept at a relatively low level. + +Effectiveness of TGP. After pruning, the voxel count of $U_{1}$ is reduced to nearly $7\%$ of its original size without TGP, while the 3DVG performance is significantly boosted. TGP serves multiple functions, including: (1) facilitating the interaction of multi-modal features through cross-attention, (2) reducing the feature amount (number of voxels) through pruning, and (3) gradually guiding the network to focus on the mentioned target based on text features. + +# 3.3. Completion-based Addition + +During the pruning process, some targets may be mistakenly removed, especially for small or narrow objects, as shown in Fig. 3 (b). Therefore, the addition operation between the upsampled pruned features $U_{l}^{G}$ and backbone features $V_{l}$ described in Equation (1) play an important role to mitigate the affect of over-pruning. + +There are two alternative addition operations: (1) Full Addition. For the intersecting regions of $V_{l}$ and $U_{l}^{G}$ , fea- + +tures are directly added. For voxel features outside the intersection of $U_{l}^{G}$ and $V_{l}$ which lack corresponding features in the other map, the missing voxel features are interpolated before addition. Due to pruning process, $U_{l}^{G}$ is sparser than $V_{l}$ . In this way, full addition can fix almost all the pruned region. But this operation is computationally heavy and make the scene representation fail to focus on relevant objects, which deviates the core idea of TGP. (2) Pruning-aware Addition. The addition is constrained to the locations of $U_{l}^{G}$ . For voxel in $U_{l}^{G}$ but not in $V_{l}$ , interpolation from $U_{l}^{G}$ is applied to complete the missing locations in $V_{l}$ . It restricts the addition operation to the shape of the pruned features, potentially leading to an over-reliance on the results of the pruning process. If important regions are over-pruned, the network may struggle to detect targets with severely damaged geometric information. + +Considering the unavoidable risk of pruning the query target, we introduce the completion-based addition (CBA). CBA is designed to address the limitations of full and pruning-aware additions. It offers a more targeted and efficient way to integrating multi-level features, ensuring the preservation of essential details while keeping the additional computational overhead negligible. + +Details of CBA. We first enhance the backbone features $V_{l}$ with the text features $T$ through cross-attention, obtaining $V_{l}^{\prime}$ . Then a MLP is adopted to predict the probability distribution of target for region selection: + +$$ +M _ {l} ^ {t a r} = \Theta (\mathrm {M L P} \left(V _ {l} ^ {\prime}\right) - \tau) \tag {6} +$$ + +where $\Theta$ is the step function, and $\tau$ is the threshold determining voxel relevance. $M_{l}^{tar}$ is a binary mask indicating potential regions of the mentioned target. Then, comparison of $M_{l}^{tar}$ with $U_{l}$ identifies missing voxels. The missing mask $M_{l}^{mis}$ is derived as follows: + +$$ +M _ {l} ^ {m i s} = M _ {l} ^ {t a r} \wedge (\neg \mathcal {C} (U _ {l} ^ {G}, V _ {l})) \tag {7} +$$ + +where $\mathcal{C}(A,B)$ denotes the generation of a binary mask for $A$ based on the shape of $B$ . Specifically, for positions in $B$ , if there are corresponding voxel features in $A$ , the mask for that position is set to 1. Otherwise it is set to 0. Missed voxel features in $U_{l}^{G}$ that correspond to $M_{l}^{mis}$ are interpolated from $U_{l}^{G}$ , filling in gaps identified by the missing mask. The completed feature map $U_{l}^{cpl}$ is computed by: + +$$ +U _ {l} ^ {c p l} = V _ {l} ^ {\prime} \odot M _ {l} ^ {m i s} + \mathcal {I} \left(U _ {l} ^ {G}, M _ {l} ^ {m i s}\right) \tag {8} +$$ + +where $\mathcal{I}$ represents linear interpolation on the feature map based on the positions specified in the mask. Finally, the original upsampled features are combined with the backbone features according to the pruning-aware addition, and merged with the completion features to yield updated $U_{l}$ : + +$$ +U _ {l} = \operatorname {C o n c a t} \left(U _ {l} ^ {G} \leftarrow V _ {l}, U _ {l} ^ {c p l}\right) \tag {9} +$$ + +where $\leftarrow$ denotes the pruning-aware addition, and Concat means concatenation of voxel features. + +Table 1. Comparison of methods on the ScanRefer dataset evaluated at IoU thresholds of 0.25 and 0.5. TSP3D achieves state-of-the-art accuracy even compared with two-stage methods, with +1.13 lead on Acc@0.5. Notably, we are the first to comprehensively evaluate inference speed for 3DVG methods. The inference speeds of other methods are obtained through our reproduction. + +
MethodVenueInputAccuracyInference Speed (FPS)
0.250.5
Two-Stage Model
ScanRefer [3]ECCV'203D+2D41.1927.406.72
TGNN [14]AAAI'213D37.3729.703.19
InstanceRefer [39]ICCV'213D40.2330.152.33
SAT [38]ICCV'213D+2D44.5430.144.34
FFL-3DOG [10]ICCV'213D41.3334.01Not released
3D-SPS [22]CVPR'223D+2D48.8236.983.17
BUTD-DETR [16]ECCV'223D50.4238.603.33
EDA [35]CVPR'233D54.5942.263.34
3D-VisTA [41]ICCV'233D45.9041.502.03
VPP-Net [31]CVPR'243D55.6543.29Not released
G3-LQ [34]CVPR'243D56.9045.58Not released
MCLN [27]ECCV'243D57.1745.533.17
Single-stage Model
3D-SPS [22]CVPR'223D47.6536.435.38
BUTD-DETR [16]ECCV'223D49.7637.055.91
EDA [35]CVPR'233D53.8341.705.98
G3-LQ [34]CVPR'243D55.9544.72Not released
MCLN [27]ECCV'243D54.3042.645.45
TSP3D (Ours)—3D56.4546.7112.43
+ +# 3.4. Train Loss + +The loss is composed of several components: pruning loss for TGP, completion loss for CBA, and objectness loss as well as bounding box regression loss for the head. Pruning loss, completion loss and objectness loss employ the focal loss to handle class imbalance. Supervision for completion and classification losses are the same, which sets voxels near the target object center as positives while leaving others as negatives. For bounding box regression, we use the Distance-IoU (D IoU) loss. The total loss function is computed as the sum of these individual losses: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {\text {p r u n i n g}} + \lambda_ {2} \mathcal {L} _ {\text {c o m}} + \lambda_ {3} \mathcal {L} _ {\text {c l a s s}} + \lambda_ {4} \mathcal {L} _ {\text {b b o x}} +$$ + +where $\lambda_{1},\lambda_{2},\lambda_{3}$ and $\lambda_4$ are the weights of different parts. + +# 4. Experiments + +# 4.1. Datasets + +We maintain the same experimental settings with previous works, employing ScanRefer [3] and SR3D/NR3D [2] as datasets. ScanRefer: Built on the ScanNet framework, ScanRefer includes 51,583 descriptions across scenes. Evaluation metrics focus on Acc@mIoU. ReferIt3D: ReferIt3D splits into Nr3D, with 41,503 human-generated descriptions, and Sr3D, containing 83,572 synthetic expressions. ReferIt3D simplifies the task by providing segmented point clouds for each object. The primary evaluation metric is accuracy in target object selection. + +Table 2. Quantitative comparisons on Nr3D and Sr3D datasets. We evaluate under three pipelines, noting that the Two-stage using Ground-Truth Boxes is impractical for real-world applications. TSP3D exhibits significant superiority, with leads of $+2.6\%$ and $+3.2\%$ on NR3D and SR3D respectively. + +
MethodVenuePipelineAccuracy
Nr3DSr3D
InstanceRefer [39]ICCV'21Two-stage (gt)38.848.0
LanguageRefer [28]CoRL'22Two-stage (gt)43.956.0
3D-SPS [22]CVPR'22Two-stage (gt)51.562.6
MVT [15]CVPR'22Two-stage (gt)55.164.5
BUTD-DETR [16]ECCV'22Two-stage (gt)54.667.0
EDA [35]CVPR'23Two-stage (gt)52.168.1
VPP-Net [31]CVPR'24Two-stage (gt)56.968.7
G3-LQ [34]CVPR'24Two-stage (gt)58.473.1
MCLN [27]ECCV'24Two-stage (gt)59.868.4
InstanceRefer [39]ICCV'21Two-stage (det)29.931.5
LanguageRefer [28]CoRL'22Two-stage (det)28.639.5
BUTD-DETR [16]ECCV'22Two-stage (det)43.352.1
EDA [35]CVPR'23Two-stage (det)40.749.9
MCLN [27]ECCV'24Two-stage (det)46.153.9
3D-SPS [22]CVPR'22Single-stage39.247.1
BUTD-DETR [16]ECCV'22Single-stage38.750.1
EDA [35]CVPR'23Single-stage40.049.7
MCLN [27]ECCV'24Single-stage45.753.4
TSP3D (Ours)—Single-stage48.757.1
+ +# 4.2. Implementation Details + +TSP3D is implemented based on PyTorch [23]. The pruning thresholds are set at $\sigma_{\mathrm{sc}} = 0.7$ and $\sigma_{\mathrm{tar}} = 0.3$ , and the completion threshold in CBA is $\tau = 0.15$ . The initial voxelization of the point cloud has a voxel size of $1\mathrm{cm}$ , while the voxel size for level $i$ features scales to $2^{i + 2}\mathrm{cm}$ . The supervision for pruning uses $L = 7$ . The weights for all components of the loss function, $\lambda_1, \lambda_2, \lambda_3, \lambda_4$ , are equal to 1. Training is conducted using four GPUs, while inference speeds are evaluated using a single consumer-grade GPU, RTX 3090, with a batch size of 1. + +# 4.3. Quantitative Comparisons + +Performance on ScanRefer. We carry out comparisons with existing methods on ScanRefer, as detailed in Tab. 1. The inference speeds of other methods are obtained through our reproduction with a single RTX 3090 and a batch size of 1. For two-stage methods, the inference speed includes the time taken for object detection in the first stage. For methods using 2D image features and 3D point clouds as inputs, we do not account for the time spent extracting 2D features, assuming they can be obtained in advance. However, in practical applications, the acquisition of 2D features also impacts overall efficiency. TSP3D achieves state-of-the-art accuracy even compared with two-stage methods, with $+1.13$ lead on Acc@0.5. Notably, in the single-stage setting, TSP3D achieves fast inference speed, which is unprecedented among the existing methods. This significant improvement is attributed to our method's efficient use of a multi-level architecture based on 3D sparse convolu + +Table 3. Impact of the proposed TGP and CBA. Evaluated on ScanRefer. + +
IDTGPCBAAccuracySpeed (FPS)
0.250.5
(a)40.1332.8714.58
(b)✓55.2046.1513.22
(c)✓41.3433.0913.51
(d)✓✓56.4546.7112.43
+ +Table 4. Influence of the two CBAs at different levels. Evaluated on ScanRefer. + +
IDCBA (level 2)CBA (level 1)AccuracySpeed (FPS)
0.250.5
(a)55.2046.1513.22
(b)✓55.1746.0612.79
(c)✓56.4546.7112.43
(d)✓✓56.2246.6812.19
+ +Table 5. Influence of different feature upsampling methods. Evaluated on ScanRefer. + +
IDMethodAccuracySpeed (FPS)
0.250.5
(a)Simple concatenation40.1332.8714.58
(b)Attention mechanism———
(c)Text-guided pruning56.2746.5810.11
(d)Simplified TGP56.4546.7112.43
+ +tions, coupled with the text-guided pruning. By focusing computation only on salient regions of the point clouds, determined by textual cues, our model effectively reduces computational overhead while maintaining high accuracy. TSP3D also sets a benchmark for inference speed comparisons for future methodologies. + +Performance on Nr3D/Sr3D. We evaluate our method on the SR3D and NR3D datasets, following the evaluation protocols of prior works like EDA [35] and BUTD-DETR [16] by using Acc@0.25 as the accuracy metric. The results are shown in Tab. 2. Given that SR3D and NR3D provide ground-truth boxes and categories for all objects in the scene, we consider three pipelines: (1) Two-stage using Ground-Truth Boxes, (2) Two-stage using Detected Boxes, and (3) Single-stage. In practical applications, the Two-stage using Ground-Truth Boxes pipeline is unrealistic because obtaining all ground-truth boxes in a scene is infeasible. This approach can also oversimplify certain evaluation scenarios. For example, if there are no other objects of the same category as the target in the scene, the task reduces to relying on the provided ground-truth category. Under the Single-stage setting, TSP3D exhibits significant superiority with peak performance of $48.7\%$ and $57.1\%$ on Nr3D and Sr3D. TSP3D even outperforms previous works under the pipeline of Two-stage using Detected Boxes, with leads of $+2.6\%$ and $+3.2\%$ on NR3D and SR3D. + +# 4.4. Ablation Study + +Effectiveness of Proposed Components. To investigate the effects of our proposed TGP and CBA, we conduct ablation experiments with module removal as shown in Tab. 3. When TGP is not used, multi-modal feature concatenation is employed as a replacement, as shown in Fig. 2 (a). When CBA is not used, it is substituted with a pruning-based addition. The results demonstrate that TGP significantly enhances performance without notably impacting inference time. This is because TGP, while utilizing a more complex multi-modal attention mechanism for stronger feature fusion, significantly reduces feature scale through text-guided pruning. Additionally, the performance improvement is also due to the gradual guidance towards the target object by both scene-level and target-level TGP. Using CBA alone has a limited effect, as no voxels are pruned. Implementing CBA on top of TGP further enhances performance, as CBA dynamically compensates for some of the excessive + +pruning by TGP, thus increasing the network's robustness. + +Influence of the Two CBAs. To explore the impact of CBAs at two different levels, we conduct ablation experiments as depicted in Tab. 4. In the absence of CBA, we use pruning-based addition as a substitute. The results indicate that the CBA at level 2 has negligible effects on the 3DVG task. This is primarily because the CBA at level 2 serves to supplement the scene-level TGP, which is expected to prune the background (a relatively simple task). Moreover, although some target features are pruned, they are compensated by two subsequent generative sparse convolutions. However, the CBA at level 1 enhances performance by adapt completion for the target-level TGP. It is challenging to fully preserve target objects from deep upsampling features, especially for smaller or narrower targets. The CBA at level 1, based on high-resolution backbone features, effectively complements the TGP. + +Feature Upsampling Techniques. We conduct experiments to assess the effects of different feature upsampling techniques, as detailed in Tab. 5. Using simple feature concatenation (Fig. 2 (a)), while fast in inference speed, results in poor performance. When we utilize an attention mechanism with stronger feature interaction, as shown in Fig. 2 (b), the computation exceeds the limits of GPU due to the large number of voxels, making it impractical for real-world applications. Consequently, we employ TGP to reduce the feature amount, as illustrated in Fig. 2 (c), which significantly improves performance and enables practical deployment. Building on TGP, we propose simplified TGP, as shown in Fig. 2 (d), that merges feature interactions before and after pruning, achieving performance consistent with the original TGP while enhancing inference speed. + +# 4.5. Qualitative Results + +Text-guided Pruning. To visually demonstrate the process of TGP, we visualize the results of two pruning phases, as shown in Fig. 4. In each example, the voxel features after scene-level pruning, the features after target-level pruning, and the features after target-level generative sparse convolution are displayed from top to bottom. It is evident that both pruning stages effectively achieve our intended effect: the scene-level pruning filters out the background and retained object voxels, and the target-level pruning preserves relevant and target objects. Moreover, during the feature upsampling process, the feature amount nearly exponen- + +![](images/e127a077bd0a6d27b89d32dd19e93b9c5eb2fb76fa75e59d370ee9285df37ad2.jpg) + +![](images/e352ef8d96b9ba4385c259554b2814f259eef93cdf05e72d5a6e185c8c2abf78.jpg) + +![](images/b810d25fd78ca887e04647bf2ab28495c02a264413ff1137c409bf98e9353a3c.jpg) +(a) It is a gray trash can. +The trash can sits in the corner by where the tv is. + +![](images/4c7df1d15cac2d971bdead70fede0d09a5a1d53cf895795226c0c147dadff021.jpg) + +![](images/4b59f969347d2443c39de66e06a3aa001e3ea34fc6a80cbaf3281305ccf4f5da.jpg) + +![](images/c6eec883c7019c38b446623085adc929254d1475ff4e4a683b05051cd167999c.jpg) +(b) The chair is on the northwest side of the table that is furthest northeast in the room. The chair has a curved backside and four legs. + +![](images/7997523968f11a8f0bfa57b29838c8b57ebdf699b5464901ea968deaebd0796d.jpg) + +![](images/8d82d108ae9eeb5e40e61c11c0285c00063831f972110e31308d2d90158d65a3.jpg) + +![](images/c78287d31b5b32615d3dc245b0de69e2419f6bdd72d8b17a2b6e8d2da4cef986.jpg) +(c) A chrome water fountain. Is the lowest of two water fountains. + +![](images/b839a58752be639e07cae1283622bd2a5b9d1d7e54a679611deb00e7f72e4903.jpg) + +![](images/314704d0cc05d1e8b9f8b526f825d0dc48364bad4c108498bceb9080f27b7db2.jpg) + +![](images/2aaa174cfe98fc7a6d6e0f85ca56d6c027848a80c4e805e07f10f8b6b0b565ed.jpg) +(d) There is a rectangular shelf. It is the one closest to the curtain. + +![](images/bb9a11ff5c5b05871f2c87ac2b51f9376842da544c867c92d7b52249b1c4bb50.jpg) +Figure 4. Visualization of the text-guided pruning process. In each example, the voxel features after scene-level TGP, target-level TGP and the last upsampling layer are presented from top to bottom. The blue boxes represent the ground truth of the target, and the red boxes denote the bounding boxes of relevant objects. TSP3D reduces the amount of voxel features through two stages of pruning and progressively guides the network focusing towards the target. +(a) The brown square trash can is to the right of the lamp. It is also behind the chair and to the right. + +![](images/978b413f7aff11a5c5d40e76562670a083c4aa956fd214fb7e44cdef0a4411b1.jpg) +(b) The whiteboard is next to two doors. The whiteboard is a long, white rectangle. + +![](images/9dd26ab7936da55721b48a48eab1056c21cb39b02b22237c3e4763534d5fc224.jpg) +(c) This is a monitor in black color. The monitor is next to the wall and at the back of maroon chair. + +![](images/0b4a7e582a78640f923f627372a739aa3cce56600e0f065314ece634229b53a0.jpg) +(d) The book rack is left of the bookshelf that looks like a right angle with another. The book rack is black and has as circular bottom. +Figure 5. Visualization of the completion-based addition process. The blue points represent the voxel features output by the target-level TGP, while the red points are the completion features predicted by the CBA. The blue boxes indicate the ground truth boxes. CBA adaptively supplements situations where excessive pruning has occurred. + +tially increases due to generative upsampling. Without TGP, the voxel coverage would far exceed the range of the scene point cloud, which is inefficient for inference. This also intuitively explains the significant impact of our TGP on both performance and inference speed. + +Completion-based Addition. To clearly illustrate the function of CBA, we visualize the adaptive completion process in Fig. 5. The images below showcase several instances of excessive pruning. TGP performs pruning based on deep and low-resolution features, which can lead to excessive pruning, potentially removing entire or partial targets. This over-pruning is more likely to occur with small, as shown in Fig. 5 (a) and (c), narrow, as in Fig. 5 (b), or elongated targets, as in Fig. 5 (d). Our CBA effectively supplements the process using higher-resolution backbone features, thus dynamically integrating multi-level features. + +# 5. Conclusion + +In this paper, we present TSP3D, an efficient sparse single-stage method for real-time 3D visual grounding. Different from previous 3D visual grounding frameworks, TSP3D builds on multi-level sparse convolutional architecture for efficient and fine-grained scene representation extraction. To enable the interaction between voxel features and textual features, we propose text-guided pruning (TGP), which reduces the amount of voxel features and guides the network to progressively focus on the target object. Additionally, we introduce completion-based addition (CBA) for adaptive multi-level feature fusion, effectively compensating for instances of over-pruning. Extensive experiments demonstrate the effectiveness of our proposed modules, resulting in an efficient 3DVG method that achieves state-of-the-art accuracy and fast inference speed. + +# Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grant 624B2076, Grant 62376132, Grant 62321005, and Grant 62441616. + +# References + +[1] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dref-transformer: Fine-grained object identification in real-world scenes using natural language. In WACV, pages 3941-3950, 2022. 2 +[2] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 6 +[3] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 1, 2, 3, 6 +[4] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In ICCV, pages 15467-15476, 2021. 2 +[5] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In CVPR, pages 21674-21683, 2023. 2 +[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2, 3 +[7] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2 +[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, pages 1201-1209, 2021. 2 +[9] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[10] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In ICCV, pages 3722-3731, 2021. 2, 6 +[11] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224–9232, 2018. 2 +[12] JunYoung Gwak, Christopher Choy, and Silvio Savarese. Generative sparse detection networks for 3d single-shot object detection. In ECCV, pages 297-313. Springer, 2020. 2, 3 +[13] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-and + +relation aware transformer for fine-grained 3d visual grounding. In ACM MM, pages 2344-2352, 2021. 2 +[14] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 1, 2, 6 +[15] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 6 +[16] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 1, 2, 3, 6, 7 +[17] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In CVPR, pages 4867-4876, 2020. 2 +[18] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 2 +[19] Daizong Liu, Yang Liu, Wencan Huang, and Wei Hu. A survey on text-guided 3d visual grounding: Elements, recent advances, and future directions. arXiv preprint arXiv:2406.05785, 2024. 2 +[20] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3 +[21] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, pages 2949-2958, 2021. 2 +[22] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 1, 2, 3, 6 +[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6 +[24] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, pages 1532-1543, 2014. 2 +[25] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 2, 3 +[26] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2 +[27] Zhipeng Qian, Yiwei Ma, Zhekai Lin, Jiayi Ji, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Multi-branch collaborative learning network for 3d visual grounding. In ECCV, pages 381–398. Springer, 2025. 2, 6 +[28] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In CoRL, pages 1046-1056. PMLR, 2022. 6 + +[29] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In ECCV, pages 477-493. Springer, 2022. 2, 3 +[30] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Tr3d: Towards real-time indoor 3d object detection. In ICIP, pages 281-285. IEEE, 2023. 2, 3 +[31] Xiangxi Shi, Zhonghua Wu, and Stefan Lee. Viewpoint-aware visual grounding in 3d scenes. In CVPR, pages 14056-14065, 2024. 3, 6 +[32] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 2 +[33] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. NeurIPS, 35:29975-29988, 2022. 2 +[34] Yuan Wang, Yali Li, and Shengjin Wang. G^3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. In CVPR, pages 13917-13926, 2024. 2, 3, 6 +[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 1, 2, 3, 6, 7 +[36] Xiuwei Xu, Ziwei Wang, Jie Zhou, and Jiwen Lu. Binarizing sparse convolutional networks for efficient point cloud analysis. In CVPR, pages 5313-5322, 2023. 2 +[37] Xiuwei Xu, Zhihao Sun, Ziwei Wang, Hongmin Liu, Jie Zhou, and Jiwen Lu. 3d small object detection with dynamic spatial pruning. In ECCV. Springer, 2024. 2 +[38] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 1, 6 +[39] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In ICCV, pages 1791-1800, 2021. 1, 2, 6 +[40] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvgtransformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 2 +[41] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 6 \ No newline at end of file diff --git a/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/images.zip b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1f4d075ac47620fde52278ea718abb87451f9802 --- /dev/null +++ b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea7dda49cc8151e271ba92a082b90c095adbba8714805b89f5344d56018e1d96 +size 681504 diff --git a/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/layout.json b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..af295f6fa44c1e329dea4b2a1091fea027509938 --- /dev/null +++ b/2025/Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding/layout.json @@ -0,0 +1,9291 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 518, + 121 + ], + "type": "text", + "content": "Text-guided Sparse Voxel Pruning for Efficient 3D Visual Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "spans": [ + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "text", + "content": "Wenxuan Guo" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "text", + "content": " Xiuwei Xu" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "text", + "content": " Ziwei Wang" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "text", + "content": " Jianjiang Feng" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "text", + "content": " Jie Zhou" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "text", + "content": " Jiwen Lu" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 141, + 533, + 172 + ], + "type": "text", + "content": "Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 173, + 463, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 173, + 463, + 186 + ], + "spans": [ + { + "bbox": [ + 146, + 173, + 463, + 186 + ], + "type": "inline_equation", + "content": "\\{gwx22,xxw21\\} @mails.tsinghua.edu.cn ziwei.wang@ntu.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 186, + 405, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 186, + 405, + 199 + ], + "spans": [ + { + "bbox": [ + 206, + 186, + 405, + 199 + ], + "type": "text", + "content": "{jfeng,jzhou,lujiwen}@tsinghua.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "text", + "content": "In this paper, we propose an efficient multi-level convolution architecture for 3D visual grounding. Conventional methods are difficult to meet the requirements of real-time inference due to the two-stage or point-based architecture. Inspired by the success of multi-level fully sparse convolutional architecture in 3D object detection, we aim to build a new 3D visual grounding framework following this technical route. However, as in 3D visual grounding task the 3D scene representation should be deeply interacted with text features, sparse convolution-based architecture is inefficient for this interaction due to the large amount of voxel features. To this end, we propose text-guided pruning (TGP) and completion-based addition (CBA) to deeply fuse 3D scene representation and text features in an efficient way by gradual region pruning and target completion. Specifically, TGP iteratively sparsifies the 3D scene representation and thus efficiently interacts the voxel features with text features by cross-attention. To mitigate the affect of pruning on delicate geometric information, CBA adaptively fixes the over-pruned region by voxel completion with negligible computational overhead. Compared with previous single-stage methods, our method achieves top inference speed and surpasses previous fastest method by " + }, + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "text", + "content": " FPS. Our method also achieves state-of-the-art accuracy even compared with two-stage methods, with " + }, + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "inline_equation", + "content": "+1.13" + }, + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "text", + "content": " lead of Acc@0.5 on ScanRefer, and " + }, + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "inline_equation", + "content": "+2.6" + }, + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "inline_equation", + "content": "+3.2" + }, + { + "bbox": [ + 55, + 251, + 297, + 587 + ], + "type": "text", + "content": " leads on NR3D and SR3D respectively. The code is available at https://github.com/GWxuan/TSP3D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 612, + 135, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 135, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 135, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": "Incorporating multi-modal information to guide 3D visual perception is a promising direction. In these years, 3D visual grounding (3DVG), also known as 3D instance referencing, has been paid increasing attention as a fundamental multi-modal 3D perception task. The aim of 3DVG is" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 226, + 552, + 388 + ], + "blocks": [ + { + "bbox": [ + 316, + 226, + 552, + 388 + ], + "lines": [ + { + "bbox": [ + 316, + 226, + 552, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 226, + 552, + 388 + ], + "type": "image", + "image_path": "661c935b6d6e9132a512ca6ed5b2382b6b4416e336bf98bb467a08e4aab6aed4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 390, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 424 + ], + "type": "text", + "content": "Figure 1. Comparison of 3DVG methods on ScanRefer dataset [3]. Our TSP3D surpasses existing methods in both accuracy and inference speed, achieving the first efficient 3DVG framework." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 437, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 555, + 545 + ], + "type": "text", + "content": "to locate an object in the scene with a free-form query description. 3DVG is challenging since it requires understanding of both 3D scene and language description. Recently, with the development of 3D scene perception and vision-language models, 3DVG methods have shown remarkable progress [16, 22]. However, with 3DVG being widely applied in fields like robotics and AR / VR where inference speed is the main bottleneck, how to construct efficient real-time 3DVG model remains a challenging problem." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 556, + 713 + ], + "type": "text", + "content": "Since the output format of 3DVG is similar with 3D object detection, early 3DVG methods [3, 14, 38, 39] usually adopt a two-stage framework, which first conducts detection to locate all objects in the scene, and then selects the target object by incorporating text information. As there are many similarities between 3D object detection and 3DVG (e.g. both of them need to extract the representation of the 3D scene), there will be much redundant feature computation during the independent adoption of the two models. As a result, two-stage methods are usually hard to handle real-time tasks. To solve this problem, single-stage methods [22, 35] are presented, which generates the bounding box of the target directly from point clouds. This integrated design is more compact and efficient. However," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 215, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 215, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 215, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3666" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "text", + "content": "current single-stage 3DVG methods mainly build on point-based architecture [25], where the feature extraction contains time-consuming operations like furthest point sampling and kNN. They also need to aggressively downsample the point features to reduce computational cost, which might hurt the geometric information of small and thin objects [37]. Due to these reasons, current single-stage methods are still far from real-time (" + }, + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "< 6" + }, + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "text", + "content": " FPS) and their performance is inferior to two-stage methods, as shown in Fig. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "text", + "content": "In this paper, we propose a new single-stage framework for 3DVG based on text-guided sparse voxel pruning, namely TSP3D. Inspired by state-of-the-art 3D object detection methods [29, 37] which achieves both leading accuracy and speed with multi-level sparse convolutional architecture, we build the first sparse single-stage 3DVG network. However, different from 3D object detection, in 3DVG the 3D scene representation should be deeply interacted with text features. Since the count of voxels is very large in sparse convolution-based architecture, deep multi-modal interaction like cross-attention becomes infeasible due to unaffordable computational cost. To this end, we propose text-guided pruning (TGP), which first utilize text information to jointly sparsify the 3D scene representation and enhance the voxel and text features. To mitigate the affect of pruning on delicate geometric information, we further present completion-based addition (CBA) to adaptively fix the over-pruned region with negligible computational overhead. Specifically, TGP prunes the voxel features according to the object distribution. It gradually removes background features and features of irrelevant objects, which generates text-aware voxel features around the target object for accurate bounding box prediction. Since pruning may mistakenly remove the representation of target object, CBA utilizes text features to query a small set of voxel features from the complete backbone features, followed by pruned-aware addition to fix the over-pruned region. We conduct extensive experiments on the popular ScanRefer [3] and ReferIt3D [2] datasets. Compared with previous single-stage methods, TSP3D achieves top inference speed and surpasses previous fastest single-stage method by " + }, + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "text", + "content": " FPS. TSP3D also achieves state-of-the-art accuracy even compared with two-stage methods, with " + }, + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "inline_equation", + "content": "+1.13" + }, + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "text", + "content": " lead of Acc@0.5 on ScanRefer, and " + }, + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "inline_equation", + "content": "+2.6" + }, + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "inline_equation", + "content": "+3.2" + }, + { + "bbox": [ + 56, + 181, + 298, + 601 + ], + "type": "text", + "content": " leads on NR3D and SR3D respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 602, + 284, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 602, + 284, + 613 + ], + "spans": [ + { + "bbox": [ + 67, + 602, + 284, + 613 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 616, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 616, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 639 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first work exploring sparse convolutional architecture for efficient 3DVG." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 639, + 294, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 639, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 54, + 639, + 294, + 675 + ], + "type": "text", + "content": "- To enable efficient feature extraction, we propose text-guided pruning and completion-based addition to sparsify sparse voxels and adaptively fuse multi-level features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 676, + 295, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 676, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 54, + 676, + 295, + 711 + ], + "type": "text", + "content": "- We conduct extensive experiments, and TSP3D outperforms existing methods in both accuracy and speed, demonstrating the superiority of the proposed framework." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 91, + 437, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 437, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 437, + 104 + ], + "type": "text", + "content": "2.1.3D Visual Grounding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 110, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 409 + ], + "type": "text", + "content": "3D visual grounding aims to locate a target object within a 3D scene based on natural language descriptions [19]. Existing methods are typically categorized into two-stage and single-stage approaches. Two-stage methods follow a detect-then-match paradigm. In the first stage, they independently extract features from the language query using pre-trained language models [7, 9, 24] and predict candidate 3D objects using pre-trained 3D detectors [21, 26] or segmenters [4, 17, 32]. In the second stage, they focus on aligning the vision and text features to identify the target object. Techniques for feature fusion include attention mechanisms with Transformers [13, 40], contrastive learning [1], and graph-based matching [10, 14, 39]. In contrast, single-stage methods integrate object detection and feature extraction, allowing for direct identification of the target object. Methods in this category include guiding keypoint selection using textual features [22], and measuring similarity between words and objects inspired by 2D image-language pre-trained models like GLIP [18], as in BUTDDETR [16]. And methods like EDA [35] and " + }, + { + "bbox": [ + 313, + 110, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^3" + }, + { + "bbox": [ + 313, + 110, + 555, + 409 + ], + "type": "text", + "content": "-LQ [34] advance single-stage 3D visual grounding by enhancing multimodal feature discriminability through explicit textdecoupling, dense alignment, and semantic-geometric modeling. MCLN [27] uses the 3D referring expression segmentation task to assist 3DVG in improving performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "text", + "content": "However, existing two-stage and single-stage methods generally have high computational costs, hindering real-time applications. Our work aims to address these efficiency challenges by proposing an efficient single-stage method with multi-level sparse convolutional architecture." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 479, + 528, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 528, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 528, + 491 + ], + "type": "text", + "content": "2.2. Multi-Level Convolutional Architectures" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 497, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 653 + ], + "type": "text", + "content": "Recently, sparse convolutional architecture has achieved great success in the field of 3D object detection. Built on the voxel-based representation [5, 8, 33] and sparse convolution operation [6, 11, 36], this kind of methods show great efficiency and accuracy when processing scene-level data. GSDN [12] first adopts multi-level sparse convolution with generative feature upsampling in 3D object detection. FCAF3D [29] simplifies the multi-level architecture with anchor-free design, achieving leading accuracy and speed. TR3D [30] further accelerates FCAF3D by removing unnecessary layers and introducing category-aware proposal assignment method. Moreover, DSPDet3D [37] introduces the multi-level architecture to 3D small object detection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": "Our proposed method draws inspiration from these approaches, utilizing a sparse multi-level architecture with sparse convolutions and an anchor-free design. This allows for efficient processing of 3D data, enabling real-time performance in 3D visual grounding tasks." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3667" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 111, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 111, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 111, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 188 + ], + "type": "text", + "content": "In this section, we describe our TSP3D for efficient single-stage 3DVG. We first analyze existing pipelines to identify current challenges and motivate our approach (Sec. 3.1). We then introduce the text-guided pruning, which leverages text features to guide feature pruning (Sec. 3.2). To address the potential risk of pruning key information, we propose the completion-based addition for multi-level feature fusion (Sec. 3.3). Finally, we detail the training loss (Sec. 3.4)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 228, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 228, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 228, + 205 + ], + "type": "text", + "content": "3.1. Architecture Analysis for 3DVG" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 296, + 437 + ], + "type": "text", + "content": "Top-performance 3DVG methods [31, 34, 35], are mainly two-stage, which is a serial combination of 3D object detection and 3D object grounding. This separate calls of two approaches result in redundant feature extraction and complex pipeline, thus making the two-stage methods less efficient. To demonstrate the efficiency of existing methods, we conduct a comparison of accuracy and speed among several representative methods on ScanRefer [3], as shown in Fig. 1. It can be seen that two-stage methods struggle in speed (< 3 FPS) due to the additional detection stage. Since 3D visual grounding is usually adopted in practical scenarios that require real-time inference under limited resources, such as embodied robots and VR/AR, the low speed of two-stage methods make them less practical. On the other side, single-stage methods [22], which directly predicts referred bounding box from the observed 3D scene, are more suitable choices due to their streamlined processes. In Fig. 1, it can be observed that single-stage methods are significantly more efficient than their two-stage counterparts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "content": "However, existing single-stage methods are mainly built on point-based backbone [25], where the scene representation is extracted with time-consuming operations like furthest point sampling and set abstraction. They also employ large transformer decoder to fuse text and 3D features for several iterations. Therefore, the inference speed of current single-stage methods is still far from real-time (" + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "inline_equation", + "content": "< 6" + }, + { + "bbox": [ + 55, + 438, + 296, + 605 + ], + "type": "text", + "content": " FPS). The inference speed of specific components in different frameworks is analyzed and discussed in detail in the supplementary material. Inspired by the success of multi-level sparse convolutional architecture in 3D object detection [30], which achieves both leading accuracy and speed, we propose to build the first multi-level convolutional single-stage 3DVG pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "TSP3D-B. Here we propose a baseline framework based on sparse convolution, namely TSP3D-B. Following the simple and effective multi-level architecture of FCAF3D [29], TSP3D-B utilizes 3 levels of sparse convolutional blocks for scene representation extraction and bounding box prediction, as shown in Fig. 2 (a). Specifically, the input point clouds " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times 6}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " with 6-dim features (3D position and RGB) are first voxelized and then fed into three sequential MinkResBlocks [6], which generates three lev" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "els of voxel features " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "l = 1,2,3" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "). With the increase of " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", the spatial resolution of " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " decreases and the context information increases. Concurrently, the free-form text with " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " words is encoded by the pre-trained RoBERTa [20] and produce the vanilla text tokens " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{l \\times d}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ". With the extracted 3D and text representations, we iteratively upsample " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "V_{3}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " and fuse it with " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " to generate high-resolution and text-aware scene representation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 344, + 175, + 555, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 175, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 344, + 175, + 555, + 190 + ], + "type": "interline_equation", + "content": "U _ {l} = U _ {l} ^ {G} + V _ {l}, \\quad U _ {l} ^ {G} = \\operatorname {G e S p C o n v} \\left(U _ {l + 1} ^ {\\prime}\\right) \\tag {1}", + "image_path": "acc87961819fb3a9048e966d35cf5610397c7f3df0ba3e2320b1c3a1522c0176.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 380, + 195, + 555, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 195, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 380, + 195, + 555, + 209 + ], + "type": "interline_equation", + "content": "U _ {l + 1} ^ {\\prime} = \\operatorname {C o n c a t} \\left(U _ {l + 1}, T\\right) \\tag {2}", + "image_path": "a475386ce39394be435494b516053e3dfbe897e8b6417617682c9a07f68adc6b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "inline_equation", + "content": "U_{3} = V_{3}" + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "text", + "content": ", GeSpConv means generative sparse convolution [12] with stride 2, which upsamples the voxel features and expands their spatial locations for better bounding box prediction. Concat is voxel-wise feature concatenation by duplicating " + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "text", + "content": ". The final upsampled feature map " + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "inline_equation", + "content": "U_{1}" + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "text", + "content": " is concatenated with " + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "text", + "content": " and fed into a convolutional head to predict the objectness scores and regress the 3D bounding box. We select the box with highest objectness score as the grounding result." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 321, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 368 + ], + "type": "text", + "content": "As shown in Fig. 1, TSP3D-B achieves an inference speed of 14.58 FPS, which is significantly faster than previous single-stage methods and demonstrates great potential for real-time 3DVG." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 375, + 434, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 434, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 434, + 388 + ], + "type": "text", + "content": "3.2. Text-guided Pruning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 392, + 555, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 595 + ], + "type": "text", + "content": "Though efficient, TSP3D-B exhibits poor performance due to the inadequate interaction between 3D scene representation and text features. Motivated by previous 3DVG methods [16], a simple solution is to replace Concat with cross-modal attention to process voxel and text features, as shown in Fig. 2 (b). However, different from point-based architectures where the scene representation is usually aggressively downsampled, the number of voxels in multi-level convolutional framework is very large1. In practical implementation, we find that the voxels expand almost exponentially with each upsampling layer, leading to a substantial computational burden for the self-attention and cross-attention of scene features. To address this issue, we introduce text-guided pruning (TGP) to construct TSP3D, as illustrated in Fig. 2 (c). The core idea of TGP is to reduce feature amount by pruning redundant voxels and guide the network to gradually focus on the final target based on textual features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 596, + 556, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 556, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 556, + 668 + ], + "type": "text", + "content": "Overall Architecture. TGP can be regarded as a modified version of cross-modal attention, which reduces the number of voxels before attention operation, thereby lowering computational cost. To minimize the affect of pruning on the final prediction, we propose to prune the scene representation gradually. At higher level where the number of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 674, + 555, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 555, + 714 + ], + "type": "text", + "content": "1Compared to point-based architectures, sparse convolutional framework provides higher resolution and more detailed scene representations, while also offering advantages in inference speed. For detailed statistics, please refer to the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3668" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 324, + 233 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 324, + 233 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 324, + 233 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 324, + 233 + ], + "type": "image", + "image_path": "34dc88865730e7f3ee523802cfd25bf1b85a5d80e96e3034fcb7988d0ce6af56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 239, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 316 + ], + "type": "text", + "content": "Figure 2. Illustration of TSP3D. TSP3D bulids on multi-level sparse convolutional architecture. It iteratively upsamples the voxel features with text-guided pruning (TGP), and fuses multi-level features via completion-based addition (CBA). (a) to (d) on the right side illustrate various options for feature upsampling. (a) refers to simple concatenation with text features, which is fast but less accurate. (b) refers to feature interaction through cross-modal attention mechanisms, which is constrained by the large number of voxels. (c) represents our proposed TGP, which first prunes voxel features under textual guidance and thus enables efficient interaction between voxel and text features. (d) shows a simplified version of TGP that removes farthest point sampling and interpolation, combines multi-modal feature interactions into a whole and moves it before pruning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 328, + 74, + 391, + 145 + ], + "blocks": [ + { + "bbox": [ + 328, + 74, + 391, + 145 + ], + "lines": [ + { + "bbox": [ + 328, + 74, + 391, + 145 + ], + "spans": [ + { + "bbox": [ + 328, + 74, + 391, + 145 + ], + "type": "image", + "image_path": "6a003d30387be96c848bce3f9f41eacf06c8f47f83080b207f66b18616a8e5fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 328, + 147, + 391, + 234 + ], + "blocks": [ + { + "bbox": [ + 328, + 147, + 391, + 234 + ], + "lines": [ + { + "bbox": [ + 328, + 147, + 391, + 234 + ], + "spans": [ + { + "bbox": [ + 328, + 147, + 391, + 234 + ], + "type": "image", + "image_path": "69e156e52604f54c768f8d61e1521fe2c118da87ec724529954e0af845483d6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 392, + 74, + 467, + 234 + ], + "blocks": [ + { + "bbox": [ + 392, + 74, + 467, + 234 + ], + "lines": [ + { + "bbox": [ + 392, + 74, + 467, + 234 + ], + "spans": [ + { + "bbox": [ + 392, + 74, + 467, + 234 + ], + "type": "image", + "image_path": "28a30a9d670f1f2884a89d0d799fe6330b8eb72a993725475238135f0171de34.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 468, + 74, + 541, + 234 + ], + "blocks": [ + { + "bbox": [ + 468, + 74, + 541, + 234 + ], + "lines": [ + { + "bbox": [ + 468, + 74, + 541, + 234 + ], + "spans": [ + { + "bbox": [ + 468, + 74, + 541, + 234 + ], + "type": "image", + "image_path": "1c6a741f76761a4176f284700f1630ed9a640e23ff31c2818ea88d8de8dcb638.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 331, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 296, + 487 + ], + "type": "text", + "content": "voxels is not too large yet, TGP prunes less voxels. While at lower level where the number of voxels is significantly increased by upsampling operation, TGP prunes the voxel features more aggressively. The multi-level architecture of TSP3D consists of three levels and includes two feature upsampling operations. Therefore, we correspondingly configure two TGPs with different functions, which are referred as scene-level TGP (level 3 to 2) and target-level TGP (level 2 to 1) respectively. Scene-level TGP aims to distinguish between objects and the background, specifically pruning the voxels on background. Target-level TGP focuses on regions mentioned in the text, intending to preserve the target object and referential objects while removing other regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 488, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 620 + ], + "type": "text", + "content": "Details of TGP. Since the pruning is relevant to the description, we need to make the voxel features text-aware to predict a proper pruning mask. To reduce the computational cost, we perform farthest point sampling (FPS) on the voxel features to reduce their size while preserving the basic distribution of the scene. Next, we utilize cross-attention to interact with the text features and employ a simple MLP to predict the probability distribution " + }, + { + "bbox": [ + 55, + 488, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 55, + 488, + 296, + 620 + ], + "type": "text", + "content": " for retaining each voxel. To prune the features " + }, + { + "bbox": [ + 55, + 488, + 296, + 620 + ], + "type": "inline_equation", + "content": "U_{l}" + }, + { + "bbox": [ + 55, + 488, + 296, + 620 + ], + "type": "text", + "content": ", we binarize and interpolate the " + }, + { + "bbox": [ + 55, + 488, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 55, + 488, + 296, + 620 + ], + "type": "text", + "content": " to obtain the pruned mask. This process can be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 112, + 629, + 294, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 629, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 112, + 629, + 294, + 644 + ], + "type": "interline_equation", + "content": "U _ {l} ^ {P} = U _ {l} \\odot \\Theta (\\mathcal {I} (\\hat {M}, U _ {l}) - \\sigma) \\tag {3}", + "image_path": "1fa3a6992a32e6ce66a7ae0dfcdcf467af77f75c058143b1e15a08bb196b6f64.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 655, + 295, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 655, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 85, + 655, + 295, + 670 + ], + "type": "interline_equation", + "content": "\\hat {M} = \\operatorname {M L P} (\\text {C r o s s A t t} (\\operatorname {F P S} (U _ {l}), \\operatorname {S e l f A t t} (T))) \\tag {4}", + "image_path": "b61ef082fc701e70a5efaff4c1a2cec05a5b61521d16e0fa3ef326787228936a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "U_{l}^{P}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is the pruned features, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is Heaviside step function, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is matrix dot product, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is the pruning threshold, and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " represents linear interpolation based on the positions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 416 + ], + "type": "text", + "content": "specified by " + }, + { + "bbox": [ + 313, + 331, + 555, + 416 + ], + "type": "inline_equation", + "content": "U_{l}" + }, + { + "bbox": [ + 313, + 331, + 555, + 416 + ], + "type": "text", + "content": ". After pruning, the scale of the scene features is significantly reduced, enabling internal feature interactions based on self-attention. Subsequently, we utilize self-attention and cross-attention to perceive the relative relationships among objects within the scene and to fuse multimodal features, resulting in updated features " + }, + { + "bbox": [ + 313, + 331, + 555, + 416 + ], + "type": "inline_equation", + "content": "U_{l}^{\\prime}" + }, + { + "bbox": [ + 313, + 331, + 555, + 416 + ], + "type": "text", + "content": ". Finally, through generative sparse convolutions, we obtain " + }, + { + "bbox": [ + 313, + 331, + 555, + 416 + ], + "type": "inline_equation", + "content": "U_{l - 1}^{G}" + }, + { + "bbox": [ + 313, + 331, + 555, + 416 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": "Supervision for Pruning. The binary supervision mask " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "M^{sce}" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": " for scene-level TGP is generated based on the centers of all objects in the scene, and the mask " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "M^{tar}" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": " for target-level TGP is based on the target and relevant objects mentioned in the descriptions:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 488, + 553, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 531 + ], + "type": "interline_equation", + "content": "M ^ {s c e} = \\bigcup_ {i = 1} ^ {N} \\mathcal {M} \\left(O _ {i}\\right), \\quad M ^ {t a r} = \\mathcal {M} \\left(O ^ {t a r}\\right) \\cup \\bigcup_ {j = 1} ^ {K} \\mathcal {M} \\left(O _ {j} ^ {r e l}\\right) \\tag {5}", + "image_path": "0f1d2dda44294b3962d8042b54e29bdf64233cd68955418c44a0834f3c1b855c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\{O_i|1\\leq i\\leq N\\}" + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "content": " indicates all objects in the scene. " + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "inline_equation", + "content": "O^{tar}" + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "inline_equation", + "content": "O^{rel}" + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "content": " refer to target and relevant objects respectively. " + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(O)" + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "content": " represents the mask generated from the center of object " + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "content": ". It generates a " + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "inline_equation", + "content": "L\\times L\\times L" + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "content": " cube centered at the center of " + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "content": " to construct the supervision mask " + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 532, + 555, + 604 + ], + "type": "text", + "content": ", where locations inside the cube is set to 1 while others set to 0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Simplification. Although the above mentioned method can effectively prune voxel features to reduce the computational cost of cross-modal attention, there are some inefficient operations in the pipeline: (1) FPS is time-consuming, especially for large scenes; (2) there are two times of interactions between voxel features and text features, the first is to guide pruning and the second is to enhance the representation, which is a bit redundant. We also empirically observe that the number of voxels is not large in level 3." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3669" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 175, + 142 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 175, + 142 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 175, + 142 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 175, + 142 + ], + "type": "image", + "image_path": "a4e77b9d9417a66cfcfba2ded9c11920b2046e1381c0234be6ffb54ea8195372.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 144, + 150, + 152 + ], + "lines": [ + { + "bbox": [ + 97, + 144, + 150, + 152 + ], + "spans": [ + { + "bbox": [ + 97, + 144, + 150, + 152 + ], + "type": "text", + "content": "(a) Ground Truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 279, + 142 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 279, + 142 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 279, + 142 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 279, + 142 + ], + "type": "image", + "image_path": "af41154e9f078452d9f30b783dcdc3e2b104a5135f64eaa6669dab57903d950c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 144, + 258, + 152 + ], + "lines": [ + { + "bbox": [ + 196, + 144, + 258, + 152 + ], + "spans": [ + { + "bbox": [ + 196, + 144, + 258, + 152 + ], + "type": "text", + "content": "(b) Pruned Features" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 72, + 158, + 176, + 230 + ], + "blocks": [ + { + "bbox": [ + 72, + 158, + 176, + 230 + ], + "lines": [ + { + "bbox": [ + 72, + 158, + 176, + 230 + ], + "spans": [ + { + "bbox": [ + 72, + 158, + 176, + 230 + ], + "type": "image", + "image_path": "f788cc08e641a9a6c6abe7e5b2ffe448156b551a3d61bc88f6b88ffd44276bde.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 231, + 162, + 240 + ], + "lines": [ + { + "bbox": [ + 86, + 231, + 162, + 240 + ], + "spans": [ + { + "bbox": [ + 86, + 231, + 162, + 240 + ], + "type": "text", + "content": "(c) Completion Features" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 158, + 279, + 230 + ], + "blocks": [ + { + "bbox": [ + 176, + 158, + 279, + 230 + ], + "lines": [ + { + "bbox": [ + 176, + 158, + 279, + 230 + ], + "spans": [ + { + "bbox": [ + 176, + 158, + 279, + 230 + ], + "type": "image", + "image_path": "8d8555ac9d3a623ff5a061fce8bb0e71bfb42eb4ec4ec9177ff45f7b0a316da9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 231, + 248, + 240 + ], + "lines": [ + { + "bbox": [ + 205, + 231, + 248, + 240 + ], + "spans": [ + { + "bbox": [ + 205, + 231, + 248, + 240 + ], + "type": "text", + "content": "(d) Prediction" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 248, + 287, + 338 + ], + "blocks": [ + { + "bbox": [ + 62, + 248, + 287, + 338 + ], + "lines": [ + { + "bbox": [ + 62, + 248, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 248, + 287, + 338 + ], + "type": "image", + "image_path": "745c87df5c4898a5956aa422424bf8d0ac0bf15a5f089047e3d660191792c15b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 340, + 295, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 374 + ], + "type": "text", + "content": "Figure 3. Illustration of completion-based addition. The upper figure (b) illustrates an example of over-pruning on the target. The lower figure (c) shows the completed features predicted by CBA." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 389, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 496 + ], + "type": "text", + "content": "To this end, we propose a simplified version of TGP, as shown in Fig. 2 (d). We remove the FPS and merge the two multi-modal interactions into one. We also move the merged interaction operation before pruning. In this way, voxel features and text features are first deeply interacted for both feature enhancement and pruning. Because in level 3 the number of voxels is small and in level " + }, + { + "bbox": [ + 55, + 389, + 295, + 496 + ], + "type": "inline_equation", + "content": "2/1" + }, + { + "bbox": [ + 55, + 389, + 295, + 496 + ], + "type": "text", + "content": " the voxels are already pruned, the computational cost of self-attention and cross-attention is always kept at a relatively low level." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 497, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 593 + ], + "type": "text", + "content": "Effectiveness of TGP. After pruning, the voxel count of " + }, + { + "bbox": [ + 55, + 497, + 295, + 593 + ], + "type": "inline_equation", + "content": "U_{1}" + }, + { + "bbox": [ + 55, + 497, + 295, + 593 + ], + "type": "text", + "content": " is reduced to nearly " + }, + { + "bbox": [ + 55, + 497, + 295, + 593 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 55, + 497, + 295, + 593 + ], + "type": "text", + "content": " of its original size without TGP, while the 3DVG performance is significantly boosted. TGP serves multiple functions, including: (1) facilitating the interaction of multi-modal features through cross-attention, (2) reducing the feature amount (number of voxels) through pruning, and (3) gradually guiding the network to focus on the mentioned target based on text features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 600, + 208, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 208, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 208, + 613 + ], + "type": "text", + "content": "3.3. Completion-based Addition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": "During the pruning process, some targets may be mistakenly removed, especially for small or narrow objects, as shown in Fig. 3 (b). Therefore, the addition operation between the upsampled pruned features " + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "inline_equation", + "content": "U_{l}^{G}" + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": " and backbone features " + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": " described in Equation (1) play an important role to mitigate the affect of over-pruning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "There are two alternative addition operations: (1) Full Addition. For the intersecting regions of " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "U_{l}^{G}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": ", fea-" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": "tures are directly added. For voxel features outside the intersection of " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "U_{l}^{G}" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": " which lack corresponding features in the other map, the missing voxel features are interpolated before addition. Due to pruning process, " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "U_{l}^{G}" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": " is sparser than " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": ". In this way, full addition can fix almost all the pruned region. But this operation is computationally heavy and make the scene representation fail to focus on relevant objects, which deviates the core idea of TGP. (2) Pruning-aware Addition. The addition is constrained to the locations of " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "U_{l}^{G}" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": ". For voxel in " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "U_{l}^{G}" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": " but not in " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": ", interpolation from " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "U_{l}^{G}" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": " is applied to complete the missing locations in " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": ". It restricts the addition operation to the shape of the pruned features, potentially leading to an over-reliance on the results of the pruning process. If important regions are over-pruned, the network may struggle to detect targets with severely damaged geometric information." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 263, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 553, + 346 + ], + "type": "text", + "content": "Considering the unavoidable risk of pruning the query target, we introduce the completion-based addition (CBA). CBA is designed to address the limitations of full and pruning-aware additions. It offers a more targeted and efficient way to integrating multi-level features, ensuring the preservation of essential details while keeping the additional computational overhead negligible." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 347, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 553, + 395 + ], + "type": "text", + "content": "Details of CBA. We first enhance the backbone features " + }, + { + "bbox": [ + 313, + 347, + 553, + 395 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 313, + 347, + 553, + 395 + ], + "type": "text", + "content": " with the text features " + }, + { + "bbox": [ + 313, + 347, + 553, + 395 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 347, + 553, + 395 + ], + "type": "text", + "content": " through cross-attention, obtaining " + }, + { + "bbox": [ + 313, + 347, + 553, + 395 + ], + "type": "inline_equation", + "content": "V_{l}^{\\prime}" + }, + { + "bbox": [ + 313, + 347, + 553, + 395 + ], + "type": "text", + "content": ". Then a MLP is adopted to predict the probability distribution of target for region selection:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 378, + 399, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 399, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 378, + 399, + 553, + 412 + ], + "type": "interline_equation", + "content": "M _ {l} ^ {t a r} = \\Theta (\\mathrm {M L P} \\left(V _ {l} ^ {\\prime}\\right) - \\tau) \\tag {6}", + "image_path": "3a5a6d1e2dd4b550b3db12de1d74265764b3af04f51ceb4d3a0367a5b424259f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "text", + "content": " is the step function, and " + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "text", + "content": " is the threshold determining voxel relevance. " + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "inline_equation", + "content": "M_{l}^{tar}" + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "text", + "content": " is a binary mask indicating potential regions of the mentioned target. Then, comparison of " + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "inline_equation", + "content": "M_{l}^{tar}" + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "inline_equation", + "content": "U_{l}" + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "text", + "content": " identifies missing voxels. The missing mask " + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "inline_equation", + "content": "M_{l}^{mis}" + }, + { + "bbox": [ + 313, + 416, + 553, + 477 + ], + "type": "text", + "content": " is derived as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 480, + 553, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 480, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 367, + 480, + 553, + 495 + ], + "type": "interline_equation", + "content": "M _ {l} ^ {m i s} = M _ {l} ^ {t a r} \\wedge (\\neg \\mathcal {C} (U _ {l} ^ {G}, V _ {l})) \\tag {7}", + "image_path": "969dfe121febaeefa9df39c7fcfe432e589c8716a48d1f32c05449473ea74218.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(A,B)" + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": " denotes the generation of a binary mask for " + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": " based on the shape of " + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": ". Specifically, for positions in " + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": ", if there are corresponding voxel features in " + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": ", the mask for that position is set to 1. Otherwise it is set to 0. Missed voxel features in " + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "inline_equation", + "content": "U_{l}^{G}" + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": " that correspond to " + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "inline_equation", + "content": "M_{l}^{mis}" + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": " are interpolated from " + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "inline_equation", + "content": "U_{l}^{G}" + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": ", filling in gaps identified by the missing mask. The completed feature map " + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "inline_equation", + "content": "U_{l}^{cpl}" + }, + { + "bbox": [ + 313, + 498, + 553, + 585 + ], + "type": "text", + "content": " is computed by:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 357, + 588, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 588, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 357, + 588, + 553, + 604 + ], + "type": "interline_equation", + "content": "U _ {l} ^ {c p l} = V _ {l} ^ {\\prime} \\odot M _ {l} ^ {m i s} + \\mathcal {I} \\left(U _ {l} ^ {G}, M _ {l} ^ {m i s}\\right) \\tag {8}", + "image_path": "1134048cacc6f42e6ed5679c2811a1444c03f9d012f26acad422d0dab297cf9a.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 608, + 553, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 553, + 668 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 608, + 553, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 608, + 553, + 668 + ], + "type": "text", + "content": " represents linear interpolation on the feature map based on the positions specified in the mask. Finally, the original upsampled features are combined with the backbone features according to the pruning-aware addition, and merged with the completion features to yield updated " + }, + { + "bbox": [ + 313, + 608, + 553, + 668 + ], + "type": "inline_equation", + "content": "U_{l}" + }, + { + "bbox": [ + 313, + 608, + 553, + 668 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 372, + 670, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 670, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 372, + 670, + 553, + 685 + ], + "type": "interline_equation", + "content": "U _ {l} = \\operatorname {C o n c a t} \\left(U _ {l} ^ {G} \\leftarrow V _ {l}, U _ {l} ^ {c p l}\\right) \\tag {9}", + "image_path": "1930f1998f05dceeeb083b115753227f2a50d8d7423d38d4836088c60f933c32.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": " denotes the pruning-aware addition, and Concat means concatenation of voxel features." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3670" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 140, + 294, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "type": "text", + "content": "Table 1. Comparison of methods on the ScanRefer dataset evaluated at IoU thresholds of 0.25 and 0.5. TSP3D achieves state-of-the-art accuracy even compared with two-stage methods, with +1.13 lead on Acc@0.5. Notably, we are the first to comprehensively evaluate inference speed for 3DVG methods. The inference speeds of other methods are obtained through our reproduction." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 140, + 294, + 357 + ], + "lines": [ + { + "bbox": [ + 57, + 140, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 294, + 357 + ], + "type": "table", + "html": "
MethodVenueInputAccuracyInference Speed (FPS)
0.250.5
Two-Stage Model
ScanRefer [3]ECCV'203D+2D41.1927.406.72
TGNN [14]AAAI'213D37.3729.703.19
InstanceRefer [39]ICCV'213D40.2330.152.33
SAT [38]ICCV'213D+2D44.5430.144.34
FFL-3DOG [10]ICCV'213D41.3334.01Not released
3D-SPS [22]CVPR'223D+2D48.8236.983.17
BUTD-DETR [16]ECCV'223D50.4238.603.33
EDA [35]CVPR'233D54.5942.263.34
3D-VisTA [41]ICCV'233D45.9041.502.03
VPP-Net [31]CVPR'243D55.6543.29Not released
G3-LQ [34]CVPR'243D56.9045.58Not released
MCLN [27]ECCV'243D57.1745.533.17
Single-stage Model
3D-SPS [22]CVPR'223D47.6536.435.38
BUTD-DETR [16]ECCV'223D49.7637.055.91
EDA [35]CVPR'233D53.8341.705.98
G3-LQ [34]CVPR'243D55.9544.72Not released
MCLN [27]ECCV'243D54.3042.645.45
TSP3D (Ours)—3D56.4546.7112.43
", + "image_path": "b2dfc9c2632c9ad671c4ffe9e10a0152d7a088e8001aebcaa9873c3f3489314d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 370, + 128, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 128, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 128, + 382 + ], + "type": "text", + "content": "3.4. Train Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 507 + ], + "type": "text", + "content": "The loss is composed of several components: pruning loss for TGP, completion loss for CBA, and objectness loss as well as bounding box regression loss for the head. Pruning loss, completion loss and objectness loss employ the focal loss to handle class imbalance. Supervision for completion and classification losses are the same, which sets voxels near the target object center as positives while leaving others as negatives. For bounding box regression, we use the Distance-IoU (D IoU) loss. The total loss function is computed as the sum of these individual losses:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 515, + 277, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 515, + 277, + 529 + ], + "spans": [ + { + "bbox": [ + 73, + 515, + 277, + 529 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {p r u n i n g}} + \\lambda_ {2} \\mathcal {L} _ {\\text {c o m}} + \\lambda_ {3} \\mathcal {L} _ {\\text {c l a s s}} + \\lambda_ {4} \\mathcal {L} _ {\\text {b b o x}}", + "image_path": "2d58414122c2f748a7b16e98b504f77f79bd2978486eb69db3a14766526920c6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 535, + 290, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 290, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 290, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 535, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2},\\lambda_{3}" + }, + { + "bbox": [ + 55, + 535, + 290, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 535, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 55, + 535, + 290, + 548 + ], + "type": "text", + "content": " are the weights of different parts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 135, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 135, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 135, + 570 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 117, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 117, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 117, + 588 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "We maintain the same experimental settings with previous works, employing ScanRefer [3] and SR3D/NR3D [2] as datasets. ScanRefer: Built on the ScanNet framework, ScanRefer includes 51,583 descriptions across scenes. Evaluation metrics focus on Acc@mIoU. ReferIt3D: ReferIt3D splits into Nr3D, with 41,503 human-generated descriptions, and Sr3D, containing 83,572 synthetic expressions. ReferIt3D simplifies the task by providing segmented point clouds for each object. The primary evaluation metric is accuracy in target object selection." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 129, + 549, + 338 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on Nr3D and Sr3D datasets. We evaluate under three pipelines, noting that the Two-stage using Ground-Truth Boxes is impractical for real-world applications. TSP3D exhibits significant superiority, with leads of " + }, + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "inline_equation", + "content": "+2.6\\%" + }, + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "inline_equation", + "content": "+3.2\\%" + }, + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "text", + "content": " on NR3D and SR3D respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 129, + 549, + 338 + ], + "lines": [ + { + "bbox": [ + 318, + 129, + 549, + 338 + ], + "spans": [ + { + "bbox": [ + 318, + 129, + 549, + 338 + ], + "type": "table", + "html": "
MethodVenuePipelineAccuracy
Nr3DSr3D
InstanceRefer [39]ICCV'21Two-stage (gt)38.848.0
LanguageRefer [28]CoRL'22Two-stage (gt)43.956.0
3D-SPS [22]CVPR'22Two-stage (gt)51.562.6
MVT [15]CVPR'22Two-stage (gt)55.164.5
BUTD-DETR [16]ECCV'22Two-stage (gt)54.667.0
EDA [35]CVPR'23Two-stage (gt)52.168.1
VPP-Net [31]CVPR'24Two-stage (gt)56.968.7
G3-LQ [34]CVPR'24Two-stage (gt)58.473.1
MCLN [27]ECCV'24Two-stage (gt)59.868.4
InstanceRefer [39]ICCV'21Two-stage (det)29.931.5
LanguageRefer [28]CoRL'22Two-stage (det)28.639.5
BUTD-DETR [16]ECCV'22Two-stage (det)43.352.1
EDA [35]CVPR'23Two-stage (det)40.749.9
MCLN [27]ECCV'24Two-stage (det)46.153.9
3D-SPS [22]CVPR'22Single-stage39.247.1
BUTD-DETR [16]ECCV'22Single-stage38.750.1
EDA [35]CVPR'23Single-stage40.049.7
MCLN [27]ECCV'24Single-stage45.753.4
TSP3D (Ours)—Single-stage48.757.1
", + "image_path": "506ed8bf9b7118e6a63a83288c30c9aa6914bd274eea0766035bf4df733881db.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 349, + 446, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 446, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 446, + 361 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "content": "TSP3D is implemented based on PyTorch [23]. The pruning thresholds are set at " + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{sc}} = 0.7" + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{tar}} = 0.3" + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "content": ", and the completion threshold in CBA is " + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\tau = 0.15" + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "content": ". The initial voxelization of the point cloud has a voxel size of " + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "inline_equation", + "content": "1\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "content": ", while the voxel size for level " + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "content": " features scales to " + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "inline_equation", + "content": "2^{i + 2}\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "content": ". The supervision for pruning uses " + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "inline_equation", + "content": "L = 7" + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "content": ". The weights for all components of the loss function, " + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4" + }, + { + "bbox": [ + 313, + 365, + 555, + 485 + ], + "type": "text", + "content": ", are equal to 1. Training is conducted using four GPUs, while inference speeds are evaluated using a single consumer-grade GPU, RTX 3090, with a batch size of 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 493, + 460, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 460, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 460, + 506 + ], + "type": "text", + "content": "4.3. Quantitative Comparisons" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Performance on ScanRefer. We carry out comparisons with existing methods on ScanRefer, as detailed in Tab. 1. The inference speeds of other methods are obtained through our reproduction with a single RTX 3090 and a batch size of 1. For two-stage methods, the inference speed includes the time taken for object detection in the first stage. For methods using 2D image features and 3D point clouds as inputs, we do not account for the time spent extracting 2D features, assuming they can be obtained in advance. However, in practical applications, the acquisition of 2D features also impacts overall efficiency. TSP3D achieves state-of-the-art accuracy even compared with two-stage methods, with " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "+1.13" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " lead on Acc@0.5. Notably, in the single-stage setting, TSP3D achieves fast inference speed, which is unprecedented among the existing methods. This significant improvement is attributed to our method's efficient use of a multi-level architecture based on 3D sparse convolu" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3671" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 97, + 206, + 154 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 206, + 91 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 206, + 91 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 206, + 91 + ], + "type": "text", + "content": "Table 3. Impact of the proposed TGP and CBA. Evaluated on ScanRefer." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 97, + 206, + 154 + ], + "lines": [ + { + "bbox": [ + 56, + 97, + 206, + 154 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 206, + 154 + ], + "type": "table", + "html": "
IDTGPCBAAccuracySpeed (FPS)
0.250.5
(a)40.1332.8714.58
(b)✓55.2046.1513.22
(c)✓41.3433.0913.51
(d)✓✓56.4546.7112.43
", + "image_path": "e97834ad6713a1a6fdae4b0102fe0b05211054ac3b6c6b210bdf4cbf943439f1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 212, + 97, + 376, + 153 + ], + "blocks": [ + { + "bbox": [ + 211, + 71, + 378, + 92 + ], + "lines": [ + { + "bbox": [ + 211, + 71, + 378, + 92 + ], + "spans": [ + { + "bbox": [ + 211, + 71, + 378, + 92 + ], + "type": "text", + "content": "Table 4. Influence of the two CBAs at different levels. Evaluated on ScanRefer." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 212, + 97, + 376, + 153 + ], + "lines": [ + { + "bbox": [ + 212, + 97, + 376, + 153 + ], + "spans": [ + { + "bbox": [ + 212, + 97, + 376, + 153 + ], + "type": "table", + "html": "
IDCBA (level 2)CBA (level 1)AccuracySpeed (FPS)
0.250.5
(a)55.2046.1513.22
(b)✓55.1746.0612.79
(c)✓56.4546.7112.43
(d)✓✓56.2246.6812.19
", + "image_path": "801e4a6da6c87829b9dd0d38c4134c132668c9e669e8f0dbdc73e710fad268ed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 384, + 98, + 550, + 153 + ], + "blocks": [ + { + "bbox": [ + 382, + 72, + 551, + 93 + ], + "lines": [ + { + "bbox": [ + 382, + 72, + 551, + 93 + ], + "spans": [ + { + "bbox": [ + 382, + 72, + 551, + 93 + ], + "type": "text", + "content": "Table 5. Influence of different feature upsampling methods. Evaluated on ScanRefer." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 384, + 98, + 550, + 153 + ], + "lines": [ + { + "bbox": [ + 384, + 98, + 550, + 153 + ], + "spans": [ + { + "bbox": [ + 384, + 98, + 550, + 153 + ], + "type": "table", + "html": "
IDMethodAccuracySpeed (FPS)
0.250.5
(a)Simple concatenation40.1332.8714.58
(b)Attention mechanism———
(c)Text-guided pruning56.2746.5810.11
(d)Simplified TGP56.4546.7112.43
", + "image_path": "d53cceb186d49079b8c143f141b2c829601e5dd2ae0fb65ca977ce87a572dfb2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 165, + 294, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 165, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 54, + 165, + 294, + 237 + ], + "type": "text", + "content": "tions, coupled with the text-guided pruning. By focusing computation only on salient regions of the point clouds, determined by textual cues, our model effectively reduces computational overhead while maintaining high accuracy. TSP3D also sets a benchmark for inference speed comparisons for future methodologies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "text", + "content": "Performance on Nr3D/Sr3D. We evaluate our method on the SR3D and NR3D datasets, following the evaluation protocols of prior works like EDA [35] and BUTD-DETR [16] by using Acc@0.25 as the accuracy metric. The results are shown in Tab. 2. Given that SR3D and NR3D provide ground-truth boxes and categories for all objects in the scene, we consider three pipelines: (1) Two-stage using Ground-Truth Boxes, (2) Two-stage using Detected Boxes, and (3) Single-stage. In practical applications, the Two-stage using Ground-Truth Boxes pipeline is unrealistic because obtaining all ground-truth boxes in a scene is infeasible. This approach can also oversimplify certain evaluation scenarios. For example, if there are no other objects of the same category as the target in the scene, the task reduces to relying on the provided ground-truth category. Under the Single-stage setting, TSP3D exhibits significant superiority with peak performance of " + }, + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "inline_equation", + "content": "48.7\\%" + }, + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "inline_equation", + "content": "57.1\\%" + }, + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "text", + "content": " on Nr3D and Sr3D. TSP3D even outperforms previous works under the pipeline of Two-stage using Detected Boxes, with leads of " + }, + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "inline_equation", + "content": "+2.6\\%" + }, + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "inline_equation", + "content": "+3.2\\%" + }, + { + "bbox": [ + 56, + 239, + 296, + 478 + ], + "type": "text", + "content": " on NR3D and SR3D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 149, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 149, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 149, + 504 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 296, + 713 + ], + "type": "text", + "content": "Effectiveness of Proposed Components. To investigate the effects of our proposed TGP and CBA, we conduct ablation experiments with module removal as shown in Tab. 3. When TGP is not used, multi-modal feature concatenation is employed as a replacement, as shown in Fig. 2 (a). When CBA is not used, it is substituted with a pruning-based addition. The results demonstrate that TGP significantly enhances performance without notably impacting inference time. This is because TGP, while utilizing a more complex multi-modal attention mechanism for stronger feature fusion, significantly reduces feature scale through text-guided pruning. Additionally, the performance improvement is also due to the gradual guidance towards the target object by both scene-level and target-level TGP. Using CBA alone has a limited effect, as no voxels are pruned. Implementing CBA on top of TGP further enhances performance, as CBA dynamically compensates for some of the excessive" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 166, + 549, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 549, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 549, + 178 + ], + "type": "text", + "content": "pruning by TGP, thus increasing the network's robustness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 178, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 553, + 370 + ], + "type": "text", + "content": "Influence of the Two CBAs. To explore the impact of CBAs at two different levels, we conduct ablation experiments as depicted in Tab. 4. In the absence of CBA, we use pruning-based addition as a substitute. The results indicate that the CBA at level 2 has negligible effects on the 3DVG task. This is primarily because the CBA at level 2 serves to supplement the scene-level TGP, which is expected to prune the background (a relatively simple task). Moreover, although some target features are pruned, they are compensated by two subsequent generative sparse convolutions. However, the CBA at level 1 enhances performance by adapt completion for the target-level TGP. It is challenging to fully preserve target objects from deep upsampling features, especially for smaller or narrower targets. The CBA at level 1, based on high-resolution backbone features, effectively complements the TGP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "text", + "content": "Feature Upsampling Techniques. We conduct experiments to assess the effects of different feature upsampling techniques, as detailed in Tab. 5. Using simple feature concatenation (Fig. 2 (a)), while fast in inference speed, results in poor performance. When we utilize an attention mechanism with stronger feature interaction, as shown in Fig. 2 (b), the computation exceeds the limits of GPU due to the large number of voxels, making it impractical for real-world applications. Consequently, we employ TGP to reduce the feature amount, as illustrated in Fig. 2 (c), which significantly improves performance and enables practical deployment. Building on TGP, we propose simplified TGP, as shown in Fig. 2 (d), that merges feature interactions before and after pruning, achieving performance consistent with the original TGP while enhancing inference speed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 425, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 425, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 425, + 575 + ], + "type": "text", + "content": "4.5. Qualitative Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "Text-guided Pruning. To visually demonstrate the process of TGP, we visualize the results of two pruning phases, as shown in Fig. 4. In each example, the voxel features after scene-level pruning, the features after target-level pruning, and the features after target-level generative sparse convolution are displayed from top to bottom. It is evident that both pruning stages effectively achieve our intended effect: the scene-level pruning filters out the background and retained object voxels, and the target-level pruning preserves relevant and target objects. Moreover, during the feature upsampling process, the feature amount nearly exponen-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3672" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 70, + 172, + 137 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 172, + 137 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 172, + 137 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 172, + 137 + ], + "type": "image", + "image_path": "e127a077bd0a6d27b89d32dd19e93b9c5eb2fb76fa75e59d370ee9285df37ad2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 93, + 137, + 171, + 201 + ], + "blocks": [ + { + "bbox": [ + 93, + 137, + 171, + 201 + ], + "lines": [ + { + "bbox": [ + 93, + 137, + 171, + 201 + ], + "spans": [ + { + "bbox": [ + 93, + 137, + 171, + 201 + ], + "type": "image", + "image_path": "e352ef8d96b9ba4385c259554b2814f259eef93cdf05e72d5a6e185c8c2abf78.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 93, + 201, + 171, + 266 + ], + "blocks": [ + { + "bbox": [ + 93, + 201, + 171, + 266 + ], + "lines": [ + { + "bbox": [ + 93, + 201, + 171, + 266 + ], + "spans": [ + { + "bbox": [ + 93, + 201, + 171, + 266 + ], + "type": "image", + "image_path": "b810d25fd78ca887e04647bf2ab28495c02a264413ff1137c409bf98e9353a3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 267, + 175, + 294 + ], + "lines": [ + { + "bbox": [ + 97, + 267, + 175, + 294 + ], + "spans": [ + { + "bbox": [ + 97, + 267, + 175, + 294 + ], + "type": "text", + "content": "(a) It is a gray trash can. \nThe trash can sits in the corner by where the tv is." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 70, + 301, + 137 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 301, + 137 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 301, + 137 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 301, + 137 + ], + "type": "image", + "image_path": "4c7df1d15cac2d971bdead70fede0d09a5a1d53cf895795226c0c147dadff021.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 137, + 301, + 200 + ], + "blocks": [ + { + "bbox": [ + 181, + 137, + 301, + 200 + ], + "lines": [ + { + "bbox": [ + 181, + 137, + 301, + 200 + ], + "spans": [ + { + "bbox": [ + 181, + 137, + 301, + 200 + ], + "type": "image", + "image_path": "4b59f969347d2443c39de66e06a3aa001e3ea34fc6a80cbaf3281305ccf4f5da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 181, + 200, + 301, + 265 + ], + "blocks": [ + { + "bbox": [ + 181, + 200, + 301, + 265 + ], + "lines": [ + { + "bbox": [ + 181, + 200, + 301, + 265 + ], + "spans": [ + { + "bbox": [ + 181, + 200, + 301, + 265 + ], + "type": "image", + "image_path": "c6eec883c7019c38b446623085adc929254d1475ff4e4a683b05051cd167999c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 265, + 294, + 298 + ], + "lines": [ + { + "bbox": [ + 187, + 265, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 187, + 265, + 294, + 298 + ], + "type": "text", + "content": "(b) The chair is on the northwest side of the table that is furthest northeast in the room. The chair has a curved backside and four legs." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 413, + 135 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 413, + 135 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 413, + 135 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 413, + 135 + ], + "type": "image", + "image_path": "7997523968f11a8f0bfa57b29838c8b57ebdf699b5464901ea968deaebd0796d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 135, + 413, + 199 + ], + "blocks": [ + { + "bbox": [ + 304, + 135, + 413, + 199 + ], + "lines": [ + { + "bbox": [ + 304, + 135, + 413, + 199 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 413, + 199 + ], + "type": "image", + "image_path": "8d82d108ae9eeb5e40e61c11c0285c00063831f972110e31308d2d90158d65a3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 200, + 413, + 264 + ], + "blocks": [ + { + "bbox": [ + 304, + 200, + 413, + 264 + ], + "lines": [ + { + "bbox": [ + 304, + 200, + 413, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 413, + 264 + ], + "type": "image", + "image_path": "c78287d31b5b32615d3dc245b0de69e2419f6bdd72d8b17a2b6e8d2da4cef986.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 268, + 392, + 293 + ], + "lines": [ + { + "bbox": [ + 320, + 268, + 392, + 293 + ], + "spans": [ + { + "bbox": [ + 320, + 268, + 392, + 293 + ], + "type": "text", + "content": "(c) A chrome water fountain. Is the lowest of two water fountains." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 417, + 70, + 513, + 137 + ], + "blocks": [ + { + "bbox": [ + 417, + 70, + 513, + 137 + ], + "lines": [ + { + "bbox": [ + 417, + 70, + 513, + 137 + ], + "spans": [ + { + "bbox": [ + 417, + 70, + 513, + 137 + ], + "type": "image", + "image_path": "b839a58752be639e07cae1283622bd2a5b9d1d7e54a679611deb00e7f72e4903.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 417, + 137, + 511, + 201 + ], + "blocks": [ + { + "bbox": [ + 417, + 137, + 511, + 201 + ], + "lines": [ + { + "bbox": [ + 417, + 137, + 511, + 201 + ], + "spans": [ + { + "bbox": [ + 417, + 137, + 511, + 201 + ], + "type": "image", + "image_path": "314704d0cc05d1e8b9f8b526f825d0dc48364bad4c108498bceb9080f27b7db2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 417, + 201, + 511, + 266 + ], + "blocks": [ + { + "bbox": [ + 417, + 201, + 511, + 266 + ], + "lines": [ + { + "bbox": [ + 417, + 201, + 511, + 266 + ], + "spans": [ + { + "bbox": [ + 417, + 201, + 511, + 266 + ], + "type": "image", + "image_path": "2aaa174cfe98fc7a6d6e0f85ca56d6c027848a80c4e805e07f10f8b6b0b565ed.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 267, + 504, + 293 + ], + "lines": [ + { + "bbox": [ + 426, + 267, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 426, + 267, + 504, + 293 + ], + "type": "text", + "content": "(d) There is a rectangular shelf. It is the one closest to the curtain." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 93, + 354, + 209, + 427 + ], + "blocks": [ + { + "bbox": [ + 54, + 301, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 301, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 555, + 346 + ], + "type": "text", + "content": "Figure 4. Visualization of the text-guided pruning process. In each example, the voxel features after scene-level TGP, target-level TGP and the last upsampling layer are presented from top to bottom. The blue boxes represent the ground truth of the target, and the red boxes denote the bounding boxes of relevant objects. TSP3D reduces the amount of voxel features through two stages of pruning and progressively guides the network focusing towards the target." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 354, + 209, + 427 + ], + "lines": [ + { + "bbox": [ + 93, + 354, + 209, + 427 + ], + "spans": [ + { + "bbox": [ + 93, + 354, + 209, + 427 + ], + "type": "image", + "image_path": "bb9a11ff5c5b05871f2c87ac2b51f9376842da544c867c92d7b52249b1c4bb50.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 430, + 201, + 456 + ], + "lines": [ + { + "bbox": [ + 103, + 430, + 201, + 456 + ], + "spans": [ + { + "bbox": [ + 103, + 430, + 201, + 456 + ], + "type": "text", + "content": "(a) The brown square trash can is to the right of the lamp. It is also behind the chair and to the right." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 212, + 355, + 306, + 428 + ], + "blocks": [ + { + "bbox": [ + 212, + 355, + 306, + 428 + ], + "lines": [ + { + "bbox": [ + 212, + 355, + 306, + 428 + ], + "spans": [ + { + "bbox": [ + 212, + 355, + 306, + 428 + ], + "type": "image", + "image_path": "978b413f7aff11a5c5d40e76562670a083c4aa956fd214fb7e44cdef0a4411b1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 430, + 304, + 455 + ], + "lines": [ + { + "bbox": [ + 218, + 430, + 304, + 455 + ], + "spans": [ + { + "bbox": [ + 218, + 430, + 304, + 455 + ], + "type": "text", + "content": "(b) The whiteboard is next to two doors. The whiteboard is a long, white rectangle." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 310, + 355, + 422, + 427 + ], + "blocks": [ + { + "bbox": [ + 310, + 355, + 422, + 427 + ], + "lines": [ + { + "bbox": [ + 310, + 355, + 422, + 427 + ], + "spans": [ + { + "bbox": [ + 310, + 355, + 422, + 427 + ], + "type": "image", + "image_path": "9dd26ab7936da55721b48a48eab1056c21cb39b02b22237c3e4763534d5fc224.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 428, + 422, + 453 + ], + "lines": [ + { + "bbox": [ + 318, + 428, + 422, + 453 + ], + "spans": [ + { + "bbox": [ + 318, + 428, + 422, + 453 + ], + "type": "text", + "content": "(c) This is a monitor in black color. The monitor is next to the wall and at the back of maroon chair." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 429, + 355, + 506, + 429 + ], + "blocks": [ + { + "bbox": [ + 429, + 355, + 506, + 429 + ], + "lines": [ + { + "bbox": [ + 429, + 355, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 429, + 355, + 506, + 429 + ], + "type": "image", + "image_path": "0b4a7e582a78640f923f627372a739aa3cce56600e0f065314ece634229b53a0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 429, + 515, + 456 + ], + "lines": [ + { + "bbox": [ + 426, + 429, + 515, + 456 + ], + "spans": [ + { + "bbox": [ + 426, + 429, + 515, + 456 + ], + "type": "text", + "content": "(d) The book rack is left of the bookshelf that looks like a right angle with another. The book rack is black and has as circular bottom." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 460, + 555, + 494 + ], + "lines": [ + { + "bbox": [ + 54, + 460, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 555, + 494 + ], + "type": "text", + "content": "Figure 5. Visualization of the completion-based addition process. The blue points represent the voxel features output by the target-level TGP, while the red points are the completion features predicted by the CBA. The blue boxes indicate the ground truth boxes. CBA adaptively supplements situations where excessive pruning has occurred." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 511, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 511, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 511, + 295, + 571 + ], + "type": "text", + "content": "tially increases due to generative upsampling. Without TGP, the voxel coverage would far exceed the range of the scene point cloud, which is inefficient for inference. This also intuitively explains the significant impact of our TGP on both performance and inference speed." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "content": "Completion-based Addition. To clearly illustrate the function of CBA, we visualize the adaptive completion process in Fig. 5. The images below showcase several instances of excessive pruning. TGP performs pruning based on deep and low-resolution features, which can lead to excessive pruning, potentially removing entire or partial targets. This over-pruning is more likely to occur with small, as shown in Fig. 5 (a) and (c), narrow, as in Fig. 5 (b), or elongated targets, as in Fig. 5 (d). Our CBA effectively supplements the process using higher-resolution backbone features, thus dynamically integrating multi-level features." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 510, + 388, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 388, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 388, + 523 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "content": "In this paper, we present TSP3D, an efficient sparse single-stage method for real-time 3D visual grounding. Different from previous 3D visual grounding frameworks, TSP3D builds on multi-level sparse convolutional architecture for efficient and fine-grained scene representation extraction. To enable the interaction between voxel features and textual features, we propose text-guided pruning (TGP), which reduces the amount of voxel features and guides the network to progressively focus on the target object. Additionally, we introduce completion-based addition (CBA) for adaptive multi-level feature fusion, effectively compensating for instances of over-pruning. Extensive experiments demonstrate the effectiveness of our proposed modules, resulting in an efficient 3DVG method that achieves state-of-the-art accuracy and fast inference speed." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3673" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 624B2076, Grant 62376132, Grant 62321005, and Grant 62441616." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 213 + ], + "type": "text", + "content": "[1] Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 3dref-transformer: Fine-grained object identification in real-world scenes using natural language. In WACV, pages 3941-3950, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 215, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 215, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 215, + 296, + 259 + ], + "type": "text", + "content": "[2] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 261, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 261, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 61, + 261, + 295, + 303 + ], + "type": "text", + "content": "[3] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 296, + 339 + ], + "type": "text", + "content": "[4] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In ICCV, pages 15467-15476, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 296, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 296, + 383 + ], + "type": "text", + "content": "[5] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In CVPR, pages 21674-21683, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 385, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 296, + 418 + ], + "type": "text", + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, pages 3075-3084, 2019. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 420, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 296, + 463 + ], + "type": "text", + "content": "[7] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 465, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 509 + ], + "type": "text", + "content": "[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, pages 1201-1209, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 510, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 542 + ], + "type": "text", + "content": "[9] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 296, + 599 + ], + "type": "text", + "content": "[10] Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. Free-form description guided 3d visual graph network for object grounding in point cloud. In ICCV, pages 3722-3731, 2021. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 296, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 296, + 643 + ], + "type": "text", + "content": "[11] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224–9232, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 646, + 296, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 296, + 688 + ], + "type": "text", + "content": "[12] JunYoung Gwak, Christopher Choy, and Silvio Savarese. Generative sparse detection networks for 3d single-shot object detection. In ECCV, pages 297-313. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[13] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "relation aware transformer for fine-grained 3d visual grounding. In ACM MM, pages 2344-2352, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[14] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 143, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 175 + ], + "type": "text", + "content": "[15] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 177, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 553, + 220 + ], + "type": "text", + "content": "[16] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 223, + 553, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 265 + ], + "type": "text", + "content": "[17] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In CVPR, pages 4867-4876, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 268, + 553, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 322 + ], + "type": "text", + "content": "[18] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, pages 10965-10975, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 325, + 553, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 553, + 368 + ], + "type": "text", + "content": "[19] Daizong Liu, Yang Liu, Wencan Huang, and Wei Hu. A survey on text-guided 3d visual grounding: Elements, recent advances, and future directions. arXiv preprint arXiv:2406.05785, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 370, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 370, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 370, + 553, + 392 + ], + "type": "text", + "content": "[20] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 553, + 426 + ], + "type": "text", + "content": "[21] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, pages 2949-2958, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 429, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 553, + 473 + ], + "type": "text", + "content": "[22] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 475, + 553, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 528 + ], + "type": "text", + "content": "[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 531, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 563 + ], + "type": "text", + "content": "[24] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, pages 1532-1543, 2014. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "type": "text", + "content": "[25] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[26] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[27] Zhipeng Qian, Yiwei Ma, Zhekai Lin, Jiayi Ji, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Multi-branch collaborative learning network for 3d visual grounding. In ECCV, pages 381–398. Springer, 2025. 2, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In CoRL, pages 1046-1056. PMLR, 2022. 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3674" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 576 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[29] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In ECCV, pages 477-493. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[30] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Tr3d: Towards real-time indoor 3d object detection. In ICIP, pages 281-285. IEEE, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "text", + "content": "[31] Xiangxi Shi, Zhonghua Wu, and Stefan Lee. Viewpoint-aware visual grounding in 3d scenes. In CVPR, pages 14056-14065, 2024. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[32] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[33] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. NeurIPS, 35:29975-29988, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[34] Yuan Wang, Yali Li, and Shengjin Wang. G^3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. In CVPR, pages 13917-13926, 2024. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "type": "text", + "content": "[36] Xiuwei Xu, Ziwei Wang, Jie Zhou, and Jiwen Lu. Binarizing sparse convolutional networks for efficient point cloud analysis. In CVPR, pages 5313-5322, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "text", + "content": "[37] Xiuwei Xu, Zhihao Sun, Ziwei Wang, Hongmin Liu, Jie Zhou, and Jiwen Lu. 3d small object detection with dynamic spatial pruning. In ECCV. Springer, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[38] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 1, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[39] Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In ICCV, pages 1791-1800, 2021. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "text", + "content": "[40] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvgtransformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 576 + ], + "type": "text", + "content": "[41] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3675" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_content_list.json b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..046754569d2ed0da7c5bf217579f05c066468389 --- /dev/null +++ b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_content_list.json @@ -0,0 +1,1507 @@ +[ + { + "type": "text", + "text": "Textured Gaussians for Enhanced 3D Scene Appearance Modeling", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brian Chao $^{1,2}$ Hung-Yu Tseng $^{2}$ Lorenzo Porzi $^{2}$ Chen Gao $^{2}$ Tuotuo Li $^{2}$ Qinbo Li $^{2}$ Ayush Saraf $^{2}$ Jia-Bin Huang $^{2,3}$ Johannes Kopf $^{2}$ Gordon Wetzstein $^{1}$ Changil Kim $^{2}$", + "bbox": [ + 116, + 178, + 879, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Stanford University $^{2}$ Meta $^{3}$ University of Maryland College Park", + "bbox": [ + 207, + 227, + 784, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https:// textured-gaussians.github.io", + "bbox": [ + 333, + 257, + 656, + 271 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f97c4011fb92fe6ea64381f0ce03e2ed665c9f349b25425d4df2e54e6f64abaa.jpg", + "image_caption": [ + "Figure 1. Textured Gaussians compared to 3D Gaussian Splatting (3DGS). Our RGBA Textured Gaussians enhance 3D scene appearance modeling, leading to improved rendering quality while using the same number of Gaussians compared to 3DGS [29]. Above, we show that Textured Gaussians faithfully reconstruct the fine details of scenes." + ], + "image_footnote": [], + "bbox": [ + 94, + 306, + 388, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/28a9f048694eb262ce139fb32dabfb7025470cbf0422f0a158b02f292ae58823.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 306, + 500, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fe6cc101f9a79075c0731c30e6678267bd0cbaffc03490749feff24a6d6b166.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 306, + 792, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f24741a0bf4541b7cc291439ab1b34ac28215b0fa81bc84fd2fea98269557ee8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 306, + 903, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 547, + 326, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) has emerged as the state-of-the-art 3D reconstruction technique, offering high-quality results with fast training and rendering. However, its expressivity is limited as pixels covered by the same Gaussian share identical colors aside from a Gaussian falloff scaling factor, and individual Gaussians can only represent simple ellipsoids geometrically. To overcome these limitations, we integrate texture and alpha mapping from traditional graphics with 3DGS. Our approach augments each Gaussian with alpha, RGB, or RGBA texture maps to model spatially varying color and opacity across each Gaussian's extent. This allows Gaussians to represent richer texture patterns and geometric structures beyond single-color ellipsoids. Notably, alpha-only texture maps significantly improve Gaussian expressivity, while further augmenting with RGB texture maps achieve maximum expressivity. We validate our method on a wide variety of standard benchmark datasets and our own custom captures at both the object and scene levels, and demonstrate image quality improvements over existing methods while using a similar or lower number of Gaussians.", + "bbox": [ + 88, + 585, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 547, + 643, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural rendering [58] achieves unprecedented novel-view synthesis and 3D reconstruction quality, offering solutions to a myriad of applications ranging from surface reconstruction [34, 59], SLAM [46, 70], material estimation and relighting [54, 69], to virtual teleportation and human avatar animation [55, 60]. Recently, 3D Gaussian Splatting (3DGS) [29] emerged as a state-of-the-art novel-view synthesis technique, with desirable properties such as high image quality results, fast training and rendering time, and an explicit primitives-based representation. Subsequent works in Gaussian Splatting have focused on topics such as improving surface reconstruction quality [16, 23], scene editing [10, 17], dynamic scene reconstruction [50, 64], and 3D generation [44, 57]. Despite the success, 3DGS sometimes fails to model detailed appearances, as demonstrated in Figure 1.", + "bbox": [ + 511, + 577, + 906, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In 3DGS, each Gaussian can only represent a single color and the shape of an ellipsoid given a camera viewpoint. This dramatically limits the set of appearances and shapes a single Gaussian can represent. To address the issue, Huang and Gong [25] modified the viewing direction calculation", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8964", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in 3DGS such that pixels covered by the same Gaussian exhibit a smooth color and opacity variation. However, this color variation across a single Gaussian is minuscule due to the spherical harmonics representation of colors, preventing them from representing complex textures. Xu et al. [63] proposed only using Gaussians to represent scene geometry and leveraged a learned UV mapping module and a global 2D RGB texture map to project textures onto Gaussian surfaces. However, the unit-sphere representation of textures significantly constrains the capacity of their model, and thus, their method fails to reconstruct objects with complex geometry or large-scale scenes.", + "bbox": [ + 89, + 90, + 483, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method builds upon 3DGS and draws inspiration from mesh-based 3D representations that model appearance using texture mapping. As shown in Figure 2, we augment each Gaussian with its alpha, RGB, or RGBA texture map so that each Gaussian is capable of representing a rich set of textures and shapes. We call this representation Textured Gaussians. With RGB textures, individual Gaussians can represent higher-frequency color variations. The alpha maps allow Gaussians to represent a wider variety of shapes, instead of just ellipsoids. To achieve this, we build custom CUDA kernels that perform ray-Gaussian intersection and texture mapping and integrate them with 3DGS. Consequently, our method not only inherits all desirable properties of 3DGS, such as fast training and rendering time and explicit 3D representation, but also substantially improves rendered image quality, especially with a small number of Gaussians.", + "bbox": [ + 89, + 272, + 483, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include:", + "bbox": [ + 109, + 529, + 370, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Introduction of a generalized appearance model for 3D Gaussians that handles spatially varying color and opacity by augmenting 3D Gaussians with alpha, RGB, or RGBA texture maps.", + "- Validation that Textured Gaussians improves 3DGS on a wide variety of both object-level and scene-level datasets.", + "- Demonstration that our method greatly outperforms 3DGS when the number of Gaussians is small and achieves better performance than 3DGS with alpha-only textures given the same model size." + ], + "bbox": [ + 89, + 544, + 482, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 709, + 232, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel-View Synthesis (NVS) and Neural Rendering. NVS tackles the problem of generating accurate renderings of 3D scenes from unseen viewpoints given a set of training images and camera poses. NVS traces back to the Structure from Motion (SfM) works [5, 19, 52] and Multiview Stereo (MVS) [15, 24, 48], which served as the basis of several groundbreaking NVS methods [8, 13, 20, 31]. However, these approaches generally require storing hundreds and thousands of input images for reprojection and blending, which leads to massive memory requirements and sometimes fails to reconstruct undersampled regions.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f1d5160b7efc320783632e29e6915bbbb05f9d2959f487ea764d32f2e8fe4995.jpg", + "image_caption": [ + "Figure 2. Variants of our Textured Gaussians model. Textured Gaussians encapsulate four kinds of color and opacity spatial variations. The top row of the figure shows the texture map associated with each Gaussian, and the bottom row shows the rendered Textured Gaussians. The constant-color and constant-alpha model (no textures) corresponds to the original 3DGS formation, which can only represent a single color up to a Gaussian falloff factor within the Gaussian extent. Textured Gaussians can already model spatially varying colors using only alpha textures since each pixel can be alpha-composed differently. The model achieves maximum expressivity when leveraging the full RGBA texture map, where each Gaussian is capable of representing complex shapes and high frequency textures." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 906, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in neural rendering [58] have made great strides in improving the quality of 3D reconstruction and novel-view rendering. Neural rendering algorithms can be roughly classified by the underlying 3D representation, ranging from point clouds [1, 31, 61], voxels [21, 47, 67], meshes [12, 18, 22, 28, 35], to implicit representations using MLPs [2, 37, 39, 51]. Recently, 3D Gaussian Splatting [29] (3DGS) has emerged as the state-of-the-art NVS technique. 3DGS achieves high image quality for NVS, is fast to optimize and render, and leverages an explicit Gaussian primitive representation, which makes the method suitable for a wide variety of tasks such as surface extraction [16, 23], human avatar manipulation [43, 49], ambient dynamic modeling [50], and object and scene editing [17, 65].", + "bbox": [ + 511, + 474, + 905, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method augments 3DGS with alpha and texture mapping commonly used in mesh-based 3D representations. During optimization and rendering, we intersect outgoing rays from pixels with Gaussians in the scene. Intersection points are then used to query per-Gaussian texture map values via UV mapping and compute RGB color and alpha blending values. This enables each Gaussian to represent complex textures and shapes, significantly improving NVS quality.", + "bbox": [ + 511, + 686, + 905, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Appearance Modeling in 3D Gaussian Splatting. In 3DGS, scene geometry is represented by the position, rotation, and scale of 3D Gaussians. By adjusting these properties, 3DGS can represent intricate geometric structures. On the other hand, appearance is modeled using per-Gaussian", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8965", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/55c62f64a1e69227bf0181e2748d1e69b7207527fee16f59158920d44f4e4362.jpg", + "image_caption": [ + "Figure 3. Textured Gaussians model pipeline. Our method consists of three major components: ray-Gaussian intersection, RGBA texture mapping, and a generalized Gaussian appearance model. To render the color of a pixel $\\mathbf{p}$ , we first trace a ray from the camera center $\\mathbf{o}$ to the pixel to intersect with 3D Gaussians in the scene. Then, we query texture and alpha values, $\\mathbf{c}_{\\mathrm{tex}}$ and $\\alpha_{\\mathrm{tex}}$ , from the per-Gaussian RGBA texture maps using the ray-Gaussian intersection point $\\mathbf{x}$ . Finally, given the retrieved spatially varying color and alpha values, we alpha-composite the color and alpha values of Gaussians that are hit by the pixel ray using the generalized Gaussian appearance model." + ], + "image_footnote": [], + "bbox": [ + 91, + 122, + 668, + 333 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9650985bd2570c73fe6d8441c3e9970e8f8a940cf553655f1bdbe6d4d75fa434.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 117, + 895, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "opacity values and spherical harmonics coefficients. However, pixels within a projected Gaussian are always shaded with the same color up to a Gaussian falloff factor, greatly limiting the expressivity of individual Gaussians. Therefore, recent works have explored ways that allow Gaussians to represent spatially varying features. Huang and Gong [25] defined per-Gaussian SH coefficients for color and opacity and used ray-Gaussian intersections to determine the viewing directions within the Gaussian extent. This enables smooth variation of colors and opacities across pixels covered by a single Gaussian but prevents the reconstruction of higher-frequency details. Xu et al. [63] disentangled appearance and geometry in 3DGS for texture editing, but their method is limited to object-centric scenes with simple geometry due to the unit-sphere parameterization of textures.", + "bbox": [ + 88, + 443, + 480, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of using a global texture map, we enhance each 3D Gaussian with a local texture map on top of the SH coefficients. This allows for a more flexible texture optimization since individual Gaussians are not tied to a shared global texture map. Our method can, therefore, reconstruct objects with complex structures and real-world scenes. Furthermore, each Gaussian can also represent various shapes with the alpha channel in the texture map.", + "bbox": [ + 89, + 686, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Memory Efficient Gaussian Splatting. In order to achieve even faster rendering time and compact storage requirements for potential applications on edge devices, there has been a plethora of recent work focusing on optimizing memory-efficient Gaussian Splatting models [3]. These algorithms either prune Gaussians and perform quantization", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of Gaussian attributes [14, 33, 38, 40-42] or exploit the structural relationships between Gaussians [11, 36].", + "bbox": [ + 511, + 443, + 903, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work is orthogonal and complementary to these works since we focus on improving the appearance modeling of Gaussians by redistributing the model size budget to the per-Gaussian texture maps given any optimized 3DGS model. These model compression algorithms can be easily integrated with the 3DGS pretraining stage in our optimization pipeline to achieve further gains in compactness.", + "bbox": [ + 511, + 477, + 905, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent Work. Different from several concurrent works that have explored the idea of augmenting Gaussians with texture maps [45, 53, 56, 62], we 1) additionally utilize alpha textures and 2) provide an in-depth analysis of various texture map variants (alpha-only, RGB, RGBA) that explores the balance between reconstruction capability and model size.", + "bbox": [ + 511, + 585, + 905, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 714, + 604, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Gaussian Splitting Model", + "text_level": 1, + "bbox": [ + 511, + 742, + 774, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In 3DGS [29], 3D scenes are represented with 3D Gaussians, and images are rendered using differentiable volume splatting. Specifically, 3DGS explicitly defines 3D Gaussians by their 3D covariance matrix $\\pmb{\\Sigma}_i\\in \\mathbb{R}^{3\\times 3}$ and center $\\pmb{\\mu}_i\\in \\mathbb{R}^3$ (the index $i$ indicating the $i^{\\mathrm{th}}$ Gaussian), where the 3D Gaussian function value at point $\\mathbf{x}\\in \\mathbb{R}^3$ is defined by:", + "bbox": [ + 511, + 767, + 905, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {i} (\\mathbf {x}) = \\exp \\left(- \\frac {1}{2} \\left(\\mathbf {x} - \\boldsymbol {\\mu} _ {i}\\right) \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\left(\\mathbf {x} - \\boldsymbol {\\mu} _ {i}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 875, + 906, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8966", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the covariance matrix $\\boldsymbol{\\Sigma} = \\mathbf{R}\\mathbf{S}\\mathbf{S}^{\\top}\\mathbf{R}^{\\top}$ is factorized into the rotation matrix $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ and the scale matrix $\\mathbf{S} \\in \\mathbb{R}^{3 \\times 3}$ . To render a 2D image from a 3D Gaussians representation, 3D Gaussians are transformed from world coordinates to camera coordinates via a world-to-camera transformation matrix $\\mathbf{W} \\in \\mathbb{R}^{3 \\times 3}$ and projected to the 2D image plane via a local affine transformation $\\mathbf{J} \\in \\mathbb{R}^{3 \\times 3}$ . The transformed 3D covariance $\\boldsymbol{\\Sigma}'$ can be calculated as:", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {\\top} \\mathbf {J} ^ {\\top} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 222, + 482, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The covariance $\\Sigma_{2\\mathrm{D}}$ of the 2D Gaussian $\\mathcal{G}^{2\\mathrm{D}}$ splatted on the image plane can be approximated as extracting the first two rows and columns of the transformed 3D covariance, $\\Sigma_{2\\mathrm{D}} = [\\Sigma^{\\prime}]_{\\{1,2\\},\\{1,2\\}} \\in \\mathbb{R}^{2 \\times 2}$ using the matrix minor notation.", + "bbox": [ + 89, + 252, + 483, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To render the color of a pixel $\\mathbf{p} \\in \\mathbb{R}^3$ , the colors associated with each Gaussian are alpha-composed from front to back following the conventional volume rendering equation:", + "bbox": [ + 89, + 314, + 483, + 371 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} (\\mathbf {p}) = \\sum_ {i = 1} ^ {K} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 371, + 482, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ is the index of the Gaussians, $\\mathbf{c}_i$ is the color of each Gaussian computed from the per-Gaussian spherical harmonic coefficients and viewing direction, and $\\alpha_{i}$ is the alpha value computed from the opacity $o_i$ associated with each Gaussian and the 2D Gaussian value evaluated at pixel location $\\mathbf{p}$ :", + "bbox": [ + 89, + 420, + 483, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\mathcal {G} _ {i} ^ {\\mathrm {2 D}} (\\mathbf {p}) \\cdot o _ {i} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 510, + 480, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attributes (center, rotation, scale, opacity, and spherical harmonic coefficients) of the Gaussians are optimized with gradient descent using photometric losses on the rendered 2D images.", + "bbox": [ + 89, + 535, + 483, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Textured Gaussians", + "text_level": 1, + "bbox": [ + 89, + 606, + 277, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the 3DGS appearance model defined in Eq. 3, we observe two properties of Gaussian Splitting:", + "bbox": [ + 89, + 628, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Pixels covered by the same Gaussian are shaded with the same color up to a Gaussian falloff scaling factor.", + "2. The per-Gaussian opacity only allows Gaussians to represent ellipsoidal shapes." + ], + "bbox": [ + 89, + 657, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These two properties greatly restrict the expressivity of individual 3D Gaussian primitives. To allow each Gaussian primitive to represent complex appearances and shapes, we assign a fixed-resolution 2D texture map of size $\\mathcal{T} \\times \\mathcal{T} \\times \\mathcal{K}, \\mathcal{T} \\in \\mathbb{N}, \\mathcal{K} \\in \\{1,3,4\\}$ to each Gaussian and transform each Gaussian into a Textured Gaussian. As shown in Figure 2, $\\mathcal{K} = 1,3,4$ correspond to alpha, RGB, and RGBA texture maps, respectively. The shape of a Textured Gaussian is thus determined by the spatially varying opacity defined by the product of the per-Gaussian opacity and the alpha channel of the texture map. The appearance of a Textured Gaussian is represented by 1) the RGB channels of", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the spatially varying textured map and 2) a set of spatially constant spherical harmonic coefficients (SH). Intuitively, the spatially constant SH coefficients represent the low-frequency textures plus view-dependent colors like specular effects, while the spatially varying texture map represents the higher-frequency spatial variations of the texture. The texture is then mapped to the plane $\\mathcal{P}$ defined by the two major axes of the $i^{\\mathrm{th}}$ 3D Gaussian that is centered at $\\pmb{\\mu}_i\\in \\mathbb{R}^3$ , as shown in Figure 3. The normal $\\mathbf{n}_i\\in \\mathbb{R}^3$ of this plane thus corresponds to the axis with the smallest scale.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each pixel $\\mathbf{p} \\in \\mathbb{R}^3$ that is currently being rendered, we cast a ray from the camera origin $\\mathbf{o} \\in \\mathbb{R}^3$ to the center of the pixel $\\mathbf{p}$ and intersect it with the plane $\\mathcal{P}$ . The intersection point $\\mathbf{x} \\in \\mathbb{R}^3$ can be calculated as:", + "bbox": [ + 511, + 242, + 903, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = \\mathbf {o} + \\frac {\\left(\\boldsymbol {\\mu} _ {i} - \\mathbf {o}\\right) \\cdot \\mathbf {n} _ {i}}{\\left(\\mathbf {p} - \\mathbf {o}\\right) \\cdot \\mathbf {n} _ {i}} \\cdot \\frac {\\mathbf {p} - \\mathbf {o}}{\\| \\mathbf {p} - \\mathbf {o} \\|} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 315, + 903, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the intersection point $\\mathbf{x}$ with $\\mathcal{P}$ , we perform UV mapping and bilinear interpolation on the texture map associated with the $i^{\\mathrm{th}}$ Gaussian to query the texture color and alpha value at the UV coordinates $u, v \\in \\mathbb{R}$ , denoted by $\\mathbf{c}_i^{\\mathrm{tex}}(u, v) \\in \\mathbb{R}^3$ and $\\alpha_i^{\\mathrm{tex}}(u, v) \\in \\mathbb{R}$ . Specifically, the UV coordinates $(u, v)$ of the intersection point on the texture map can be calculated as:", + "bbox": [ + 511, + 357, + 906, + 464 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} u = \\frac {m \\cdot \\sigma_ {i} ^ {1} + (\\mathbf {x} - \\boldsymbol {\\mu} _ {i}) \\cdot \\mathbf {r} _ {i} ^ {1}}{2 \\cdot m \\cdot \\sigma_ {i} ^ {1}} \\cdot (\\mathcal {T} - 1) \\\\ v = \\frac {m \\cdot \\sigma_ {i} ^ {2} + (\\mathbf {x} - \\boldsymbol {\\mu} _ {i}) \\cdot \\mathbf {r} _ {i} ^ {2}}{2 \\cdot m \\cdot \\sigma_ {i} ^ {2}} \\cdot (\\mathcal {T} - 1) \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 488, + 903, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_i^1, \\sigma_i^2 \\in \\mathbb{R}$ are the scales of the two major axis of the $i^{\\text{th}}$ Gaussian, $\\mathbf{r}_i^1, \\mathbf{r}_i^2 \\in \\mathbb{R}^3$ are the normalized directions of the two major axes of the $i^{\\text{th}}$ Gaussian, and $m \\in \\mathbb{R}$ is a scalar multiplier that determines the extent of the texture map with respect to each Gaussian.", + "bbox": [ + 511, + 570, + 905, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combined with the color computed from the SH coefficients, which we call $\\mathbf{c}_i^{\\mathrm{base}}$ (the same as the color component in the original 3DGS appearance model in Eq. 3), the final color contribution of the $i$ -th Gaussian to pixel $\\mathbf{p}$ is defined by:", + "bbox": [ + 511, + 646, + 905, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {i} (\\mathbf {p}) = \\mathbf {c} _ {i} ^ {\\text {b a s e}} + \\mathbf {c} _ {i} ^ {\\text {t e x}} (u, v) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 722, + 903, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the alpha value of the $i$ -th Gaussian at pixel $\\mathbf{p}$ is defined by:", + "bbox": [ + 511, + 747, + 903, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} (\\mathbf {p}) = \\alpha_ {i} ^ {\\mathrm {t e x}} (u, v) \\cdot \\mathcal {G} _ {i} (\\mathbf {x}) \\cdot o _ {i} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 777, + 903, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, to render the color of a pixel $\\mathbf{p}$ , we modify the 3DGS appearance model in Eq. 3 to incorporate the spatially varying texture and opacity:", + "bbox": [ + 511, + 803, + 903, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} ^ {\\text {f i n a l}} (\\mathbf {p}) = \\sum_ {i = 1} ^ {K} \\mathbf {c} _ {i} (\\mathbf {p}) \\alpha_ {i} (\\mathbf {p}) \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} (\\mathbf {p})) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 859, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8967", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/dc8f83c866497aaf0ee4a31dd6b53ab508809b8e1383ffe3e413500134db33f6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBlender [37]Mip-NeRF 360 [37]DTU [27]Tanks and Temples [30]Deep Blending [20]
Mip-NeRF 360 [4]30.34 / 0.9510 / 0.060027.69 / 0.7920 / 0.237034.28 / 0.9641 / 0.069822.22 / 0.7590 / 0.257029.40 / 0.9010 / 0.2450
Instant-NGP [39]32.20 / 0.9590 / 0.055025.30 / 0.6710 / 0.371023.58 / 0.7653 / 0.260321.72 / 0.7230 / 0.330023.62 / 0.7970 / 0.4230
Xu et al. [63]28.97 / 0.9380 / 0.0550— / — / —30.03 / — / 0.1440— / — / —— / — / —
3DGS*33.08 / 0.9671 / 0.044027.26 / 0.8318 / 0.187133.54 / 0.9697 / 0.055124.18 / 0.8541 / 0.175428.04 / 0.8940 / 0.2707
Ours33.24 / 0.9674 / 0.042827.35 / 0.8274 / 0.185833.61 / 0.9699 / 0.055624.26 / 0.8542 / 0.168428.33 / 0.8908 / 0.2699
3DGS* (10%)31.47 / 0.9590 / 0.059425.77 / 0.7796 / 0.286032.71 / 0.9627 / 0.081122.82 / 0.8020 / 0.272827.64 / 0.8853 / 0.3101
Ours (10%)32.14 / 0.9629 / 0.048926.32 / 0.7976 / 0.232332.74 / 0.9661 / 0.058223.41 / 0.8259 / 0.212227.98 / 0.8898 / 0.2804
3DGS* (1%)26.89 / 0.9160 / 0.116522.37 / 0.6293 / 0.477430.88 / 0.9320 / 0.158119.90 / 0.6736 / 0.440623.97 / 0.8167 / 0.4337
Ours (1%)28.11 / 0.9343 / 0.084923.73 / 0.7064 / 0.336532.43 / 0.9627 / 0.069421.10 / 0.7399 / 0.310424.83 / 0.8454 / 0.3552
", + "bbox": [ + 91, + 88, + 911, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons of different novel view synthesis methods. We boldface the best-performing model and underline the second best in terms of different metrics (PSNR ↑ / SSIM ↑ / LPIPS ↓). Our Textured Gaussians model achieves better performance than 3DGS* across most metrics. With fewer Gaussians (1% and 10% of the default optimized number of Gaussians), our method significantly outperforms 3DGS*, achieving strictly better performance across all metrics as indicated by the bold text.", + "bbox": [ + 89, + 265, + 906, + 324 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/38c82b54d0e979c5ebc79a792ff251f398fea07d938b8c01c4f075f314bc6c3a.jpg", + "image_caption": [ + "Figure 4. NVS performance using varying numbers of Gaussians. When using the same number of Gaussians as $3\\mathrm{DGS}^*$ , our Textured Gaussians achieve better novel view synthesis results on all five benchmark datasets in PSNR. Here, we show the quantitative performance of $3\\mathrm{DGS}^*$ and Textured Gaussians models optimized with varying number of Gaussians on the three scene-level datasets." + ], + "image_footnote": [], + "bbox": [ + 91, + 345, + 480, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. 9 is a generalized formulation of 3D Gaussian appearance and encapsulates different variants of Textured Gaussians. For example, $\\mathbf{c}_i^{\\mathrm{tex}} = 0$ and $\\alpha_{i}^{\\mathrm{tex}} = 1$ correspond to the original 3DGS model.", + "bbox": [ + 89, + 599, + 483, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Optimization of Textured Gaussians", + "text_level": 1, + "bbox": [ + 89, + 670, + 405, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following 3DGS [29], we optimize our model to minimize the weighted photometric loss:", + "bbox": [ + 89, + 691, + 482, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\mathcal {L} _ {1} + (1 - \\lambda) \\mathcal {L} _ {\\text {S S I M}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 734, + 482, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda = 0.8$", + "bbox": [ + 89, + 763, + 196, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our optimization procedure consists of two stages. We first optimize a 3DGS model for 30000 iterations. The learning rates and adaptive density control (ADC) parameters are reported in the Appendix. In the second stage, we initialize all attributes of the Gaussians with the optimized vanilla 3DGS model, and jointly optimize them with the per-Gaussian 2D texture maps for another 30000 iterations. We disable the ADC control in the second stage to", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "control the number of Gaussians for fair baseline comparisons. This two-stage optimization process greatly speeds up convergence and improves image quality, since jointly optimizing all parameters is a highly ill-posed problem.", + "bbox": [ + 511, + 348, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement custom CUDA kernels to perform fast ray-Gaussian intersection, UV mapping, and color composition. All experiments are conducted on clusters of Nvidia H100 GPUs. Please refer to the Appendix for more details on implementation and optimization.", + "bbox": [ + 511, + 409, + 905, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results and Analysis", + "text_level": 1, + "bbox": [ + 511, + 497, + 712, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We show selected qualitative and quantitative results in this section and refer the readers to the Appendix and the project website for an extensive set of results and video renderings. Datasets and Evaluation Protocols. We evaluate the novel-view synthesis performance of our method on the 8 synthetic scenes from the Blender dataset [37], all 9 scenes from the Mip-NeRF 360 dataset [4], 5 scenes from the DTU dataset [27], 2 scenes from the Tanks and Temples dataset [30], and 2 scenes from the Deep Blending dataset [20]. Please refer to the Appendix for a detailed description of the dataset preparation process.", + "bbox": [ + 511, + 523, + 905, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For comparisons with Mip-NeRF 360 [4] and InstantNGP [39], we refer to Kerbl et al. [29] for results in scene-level datasets [4, 20, 30], and refer to NeRFBaselines [32] for results on the Blender dataset [37]. Results on the DTU dataset [27] are from our own runs of the authors' released code. Please refer to the Appendix for additional experimental details.", + "bbox": [ + 511, + 689, + 905, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Xu et al. [63] pointed out in Section 5.4 of their paper that their algorithm cannot handle scene-level datasets and objects with complex structures (i.e. Blender dataset), thus they only reported results on the DTU dataset [27]. We report those numbers and additionally show results on the Blender dataset from GStex [45] authors' own runs of the algorithm described in Xu et al. [63].", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8968", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ba48bd2b02151e1e2d1625baffd2989bb513cc0f5308839156cdcadda85a0271.jpg", + "image_caption": [ + "3DGS*", + "Ours", + "3DGS*", + "Ours", + "Figure 5. Qualitative NVs results of benchmark datasets. Novel-view rendering results for both object-level and scene-level standard benchmark datasets. Given the same small number of Gaussians (on average $2.5\\mathrm{k}$ and $39\\mathrm{k}$ Gaussians for object-level and scene-level datasets, respectively), $3\\mathrm{DGS}^*$ fails to reconstruct high-frequency textures and complex shapes while our RGBA Textured Gaussians model succeeds. Please refer to the Appendix for results of our models optimized using varying number of Gaussians." + ], + "image_footnote": [], + "bbox": [ + 93, + 112, + 903, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to standard benchmark test scenes, we also capture our own set of scenes that contain artworks with highly detailed textures to help demonstrate the effectiveness of our method.", + "bbox": [ + 89, + 630, + 482, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3DGS Baseline. We build our method on top of our own implementation of a slightly modified 3DGS that closely follows the algorithm described in Gaussian Opacity Fields [68]. Specifically, we use a ray-based formulation that allows for the exact evaluation of 3D Gaussian values without approximating splatted 3D Gaussians as 2D Gaussians. In addition, we used a revised densification strategy [66, 68]. For fair comparisons with our method, we report the performance of our 3DGS implementation and label it as $3\\mathrm{DGS}^*$ throughout the results section, and refer the readers to the Appendix for quantitative results reported in the original 3DGS paper [29].", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 628, + 705, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 1, we show quantitative comparisons in terms of PSNR/SSIM/LPIPS between our Textured Gaussians model using full RGBA textures and 3DGS* with the same number of Gaussians, and results of other baseline methods. Our method generally achieves better results than 3DGS* and all other methods, since Texture Gaussians are more expressive than vanilla 3D Gaussians due to the added texture maps.", + "bbox": [ + 511, + 655, + 906, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the bottom half of Table 1, we also show comparisons between our method and $3\\mathrm{DGS}^*$ when using a fraction of the default number of Gaussians, denoted by the percentage in parentheses. For our models with different numbers of Gaussians, we distribute a fixed amount of texels to all Gaussians (i.e., the same memory overhead for all Textured Gaussians models compared with $3\\mathrm{DGS}^*$ ). Therefore, the texture map resolution of models with fewer Gaussians will be larger. We see that our method truly outshines $3\\mathrm{DGS}^*$", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8969", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a9dbcda26ee4bfd6777efeebca4d6cba0098a5f613aa4ad0ddd980f6211ae41.jpg", + "image_caption": [ + "3DGS*", + "Ours", + "Figure 6. Qualitative NVS results of custom datasets. Our RGBA Textured Gaussians model achieves sharper reconstruction compared to $3\\mathrm{DGS}^*$ when using the same small number of Gaussians (on average around 100k Gaussians)." + ], + "image_footnote": [], + "bbox": [ + 93, + 112, + 495, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ec87eedc6ff5d9a16aade2190d8b6236d17d894d0931848201705348f66a69d.jpg", + "image_caption": [ + "3DGS*" + ], + "image_footnote": [], + "bbox": [ + 503, + 111, + 903, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dd5b64ff4bf2d8917549cee122a4a05111968aee00de8ce7e3b7b0481a6d47d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method3DGS*Alpha-onlyRGBA
PSNR#GS (M)PSNR#GS (M)PSNR#GS (M)
Blender33.080.2733.120.1933.030.21
Mip-NeRF 36027.264.427.373.127.263.5
DTU33.530.2833.450.2433.410.22
Tanks and Temples24.172.924.382.624.281.3
Deep Blending28.042.828.361.628.521.0
", + "bbox": [ + 91, + 345, + 483, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Same model size comparisons. We compare the PSNR performance between $3\\mathrm{DGS}^*$ and our model using alpha-only and RGBA textures with the same model size. The average number of optimized Gaussians (in millions) from each dataset is shown in the #GS (M) columns. We boldface and underline the best and second-best performing models, respectively. Alpha-only texture models generally achieve better performance than $3\\mathrm{DGS}^*$ and RGBA texture models with the same model size.", + "bbox": [ + 89, + 455, + 483, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with fewer Gaussians, achieving nearly 2 dB improvements over $3\\mathrm{DGS}^*$ when using $1\\%$ of the default number of Gaussians. Figure 4 shows the trend of the quality of novel view synthesis as we vary the number of Gaussians. Our method consistently outperforms $3\\mathrm{DGS}^*$ when using different numbers of Gaussians.", + "bbox": [ + 89, + 595, + 482, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 compares the performance of $3\\mathrm{DGS}^*$ and our Textured Gaussians model with the same model size using alpha-only textures and RGBA textures. Since our models use per-Gaussian texture maps, they require fewer Gaussians than $3\\mathrm{DGS}^*$ to achieve the same model size. Our alpha-only textures model generally outperforms the $3\\mathrm{DGS}^*$ and RGBA textures models, indicating that there is a sweet spot for distributing the model size budget between Gaussian parameters and texture map channels. This is further explored in Section 4.3.", + "bbox": [ + 89, + 686, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 848, + 274, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel-View Synthesis. We show novel view synthesis results of our model with RGBA textures and 3DGS* on", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both standard benchmark and custom-captured datasets in Figures 5 and 6. Our method reconstructs much sharper details of the scene compared to $3\\mathrm{DGS}^*$ when using the same small number of Gaussians. Please refer to the Appendix for results of our models optimized using varying number of Gaussians.", + "bbox": [ + 511, + 348, + 906, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Color Component Decomposition. We show the alpha-modulated and composited results of the two color components, $\\mathbf{c}^{\\mathrm{tex}}$ and $\\mathbf{c}^{\\mathrm{base}}$ , of the optimized alpha-only, RGB and RGBA textured Gaussians model with $1\\%$ of the default number of Gaussians in Figure 7. A 3DGS* model with the same number of Gaussians is shown for comparison. For RGB and RGBA models, $\\mathbf{c}^{\\mathrm{tex}}$ reconstructs fine-grained textures that 3DGS* cannot. For alpha-only and RGBA models, $\\mathbf{c}^{\\mathrm{base}}$ also reconstructs high-frequency details due to spatially varying alpha composition enabled by alpha textures.", + "bbox": [ + 511, + 440, + 908, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 619, + 679, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Texture Map Variants. We ablate variants of our model that use different texture maps, namely alpha-only, RGB, and RGBA texture maps, and the same number of Gaussians in Figure 8. Experiments are conducted on Blender, Tanks and Temples, and Deep Blending datasets with the default optimized number (top row) and $1\\%$ of the default optimized number (bottom row) of Gaussians. We see that our model achieves the best performance with RGBA textures. Interestingly, using alpha-only textures already outperforms $3\\mathrm{DGS}^*$ , and our RGB textures model despite being one-third the size, striking a better balance between performance and model size. This is because alpha-textured Gaussians can represent complex shapes through spatially varying opacity and reconstruct high-frequency textures through spatially varying alpha composition. In contrast, RGB-textured Gaussians are still limited to representing ellipsoids.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8970", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/763145371f3925836e11dc43f411c2dbd20598dcf106497553929c47299b2112.jpg", + "image_caption": [ + "Figure 7. Color component decomposition visualization. Visualization of alpha-modulated and composited $\\mathbf{c}^{\\mathrm{base}}$ and $\\mathbf{c}^{\\mathrm{tex}}$ color components in our Textured Gaussians model (Eq. 7, 8, 9). Insets with yellow borders show zoom-in crops. Texture color brightness is adjusted for visualization. While $3\\mathrm{DGS}^{*}$ fails to capture fine details with limited Gaussians, our method reconstructs sharp textures. With RGB textures, base color reconstructs low-frequency textures and texture maps reconstruct high-frequency elements. With alpha-only and RGBA textures, base color also reconstructs high-frequency textures via spatially varying alpha compositing." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 903, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/716d6fa039cc6e3e470c1790d0cc6a9884050cffa4584931db959d8033984c10.jpg", + "image_caption": [ + "Figure 8. Ablation study on texture map variants. We show the PSNR performance improvements ( $\\Delta$ PSNR) over $3\\mathrm{DGS}^*$ of our Textured Gaussians model when using different texture map variants. Models with full RGBA textures achieve the best results. Interestingly, alpha-only textures outperform RGB textures and perform comparably to RGBA textures while using one-third and one-fourth of the model size, respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 371, + 480, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Texture Map Resolution and the Number of Gaussians. In Figure 9, we optimize alpha-only Textured Gaussians models of the same model size with different texture map resolutions and number of Gaussians, and show the trend of novel view synthesis performance on the Blender and DTU datasets. We observe a sweet spot between texture map resolution and number of Gaussians that achieves the best image reconstruction quality. Maximizing texture map resolution corresponds to greatly reducing the number of Gaussians, which harms reconstruction of detailed geometric structures. Minimizing the texture map resolution, on the other hand, simply degenerates into 3DGS*.", + "bbox": [ + 88, + 633, + 482, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussions", + "text_level": 1, + "bbox": [ + 89, + 829, + 212, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The use of 2D diffuse texture maps in our model assumes that all textures lie on a local surface and does not model spatially varying specular color. Thus, ex", + "bbox": [ + 88, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bab8bbfb851f74a0c0231a6d10545499e33298ab511d5602dc2d1260bd4be94f.jpg", + "image_caption": [ + "Figure 9. Ablation study on texture map resolution and the number of Gaussians given the same model size. We optimized various alpha-only Textured Gaussians models (red dots) with equivalent model sizes but different texture map resolutions and numbers of Gaussian. A fitted quadratic curve (blue dashed line) illustrates the performance trend. The optimal model variant balances these two parameters to achieve superior novel view synthesis results." + ], + "image_footnote": [], + "bbox": [ + 517, + 372, + 901, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tending our texture representation to represent local 3D volume textures or even 5D radiance fields is crucial. Using factorized representations such as TensoRF [9] or triplane [7] to represent these high-dimensional textures could also be an interesting research problem. Finally, extending Textured Gaussians to support dynamic scene reconstruction [50, 64] through time-varying texture maps could be a natural next-step.", + "bbox": [ + 511, + 643, + 906, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusions. In this paper, we augment 3DGS with texture maps to allow individual Gaussians to model spatially varying colors and opacity. As such, each Gaussian can represent a much richer set of appearances and shapes. This greatly improves the expressivity of individual Gaussians, leading to better novel-view synthesis quality. Our method achieves better quality than 3DGS when using the same number of Gaussians, and achieves better or comparable quality when using the same model size.", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8971", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank artist Kija Lucas for allowing us to use her art installation at the Palo Alto Art Center as one of our custom captured scenes (the flower gallery scene). Brian Chao is supported by the Stanford Graduate Fellowship and the NSF GRFP.", + "bbox": [ + 89, + 114, + 483, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 186, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] K. Aliev, A. Sevastopolsky, M. Kolos, D. Ulyanov, and V. Lempitsky. Neural Point-Based Graphics, 2019. arXiv preprint. 2", + "[2] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In CVPR, 2022. 2", + "[3] Milena T. Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3DGS.zip: A survey on 3D Gaussian Splitting Compression Methods, 2024. arXiv preprint. 3", + "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In CVPR, 2022. 5, 2", + "[5] Matthew Brown and David G Lowe. Unsupervised 3D Object Recognition and Reconstruction in Unordered Datasets. In 3DIM, 2005. 2", + "[6] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising Densification in Gaussian Splatting. In ECCV, 2024. 1", + "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient Geometry-Aware 3D Generative Adversarial Networks. In CVPR, 2022. 8", + "[8] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth Synthesis and Local Warps for Plausible Image-Based Navigation. ACM TOG, 32 (3), 2013. 2", + "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In ECCV, 2022. 8", + "[10] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splitting. In CVPR, 2024. 1", + "[11] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: Hash-Grid Assisted Context for 3D Gaussian Splitting Compression. In ECCV, 2024. 3", + "[12] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In ECCV, 2022. 2", + "[13] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating Textures. In CGF, 2008. 2" + ], + "bbox": [ + 93, + 229, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and $200+$ FPS. In NeurIPS, 2024. 3", + "[15] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-View Stereo for Community Photo Collections. In ICCV, 2007. 2", + "[16] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In CVPR, 2024. 1, 2, 4", + "[17] Antoine Guédon and Vincent Lepetit. Gaussian Frosting: Editable Complex Radiance Fields with Real-Time Rendering. In ECCV, 2024. 1, 2", + "[18] Yuan-Chen Guo, Yan-Pei Cao, Chen Wang, Yu He, Ying Shan, and Song-Hai Zhang. VMesh: Hybrid Volume-Mesh Representation for Efficient View Synthesis. In SIGGRAPH Asia, 2023. 2", + "[19] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, second edition, 2004. 2", + "[20] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. ACM TOG, 37(6), 2018. 2, 5, 9", + "[21] D. Hu, Z. Zhang, T. Hou, T. Liu, H. Fu, and M. Gong. Multiscale Representation for Real-Time Anti-Aliasing Neural Rendering. In ICCV, 2023. 2", + "[22] Ronghang Hu, Nikhila Ravi, Alexander C. Berg, and Deepak Pathak. Worldsheet: Wrapping the World in a 3D Sheet for View Synthesis from a Single Image. In ICCV, 2021. 2", + "[23] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In SIGGRAPH, 2024. 1, 2", + "[24] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In CVPR, 2018. 2", + "[25] Zhentao Huang and Minglun Gong. Textured-GS: Gaussian Splatting with Spatially Defined Color and Opacity, 2024. arXiv preprint. 1, 3", + "[26] Junha Hyung, Susung Hong, Sungwon Hwang, Jaeseong Lee, Jaegul Choo, and Jin-Hwa Kim. Effective rank analysis and regularization for enhanced 3d gaussian splatting. In NeuRIPS, 2024. 4, 5", + "[27] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large Scale Multi-View Stereopsis Evaluation. In CVPR, 2014. 5, 2, 9", + "[28] H. Kato, Y. Ushiku, and T. Harada. Neural 3D Mesh Renderer. In CVPR, 2018. 2", + "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splitting for Real-Time Radiance Field Rendering. ACM TOG, 42(4), 2023. 1, 2, 3, 5, 6", + "[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. ACM TOG, 36(4), 2017. 5, 2, 9" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8972", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with PerView Optimization. CGF, 40(4), 2021. 2", + "[32] Jonas Kulhanek and Torsten Sattler. NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods, 2024. arXiv preprint. 5", + "[33] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3D Gaussian Representation for Radiance Field. In CVPR, 2024. 3", + "[34] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In CVPR, 2023. 1", + "[35] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep Appearance Models for Face Rendering. ACM TOG, 37(4), 2018. 2", + "[36] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering. In CVPR, 2024. 3", + "[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 2, 5, 9", + "[38] Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3D Scene Representation via Self-Organizing Gaussian Grids. In ECCV, 2024. 3", + "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM TOG, 41(4), 2022. 2, 5", + "[40] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. In ECCV, 2024. 3", + "[41] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3D Gaussian Splitting for Accelerated Novel View Synthesis. In CVPR, pages 10349-10358, 2024.", + "[42] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the Memory Footprint of 3D Gaussian Splatting. ACM CGIT, 7(1), 2024. 3", + "[43] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting. In CVPR, 2024. 2", + "[44] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. DreamGaussian4D: Generative 4D Gaussian Splatting, 2023. arXiv preprint. 1", + "[45] Victor Rong, Jingxiang Chen, Sherwin Bahmani, Kiriakos N Kutulakos, and David B Lindell. GStex: Per-Primitive Texturing of 2D Gaussian Splatting for Decoupled Appearance and Geometry Modeling, 2024. 3, 5", + "[46] Antoni Rosinol, John J Leonard, and Luca Carlone. NeRF-SLAM: Real-time Dense Monocular Slam with Neural Radiance Fields. In IROS, 2023. 1", + "[47] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields without Neural Networks. In CVPR, 2022. 2" + ], + "bbox": [ + 91, + 92, + 480, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A Comparison and Evaluation of Multi-View Stereo Reconstruction Algorithms. In CVPR, 2006. 2", + "[49] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In CVPR, 2024. 2", + "[50] Meng-Li Shih, Jia-Bin Huang, Changil Kim, Rajvi Shah, Johannes Kopf, and Chen Gao. Modeling ambient scene dynamics for free-view synthesis. 2024. 1, 2, 8", + "[51] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene Representation Networks: Continuous 3D-Structure-Aware Neural Scene Representations. In NeurIPS, 2019. 2", + "[52] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. ACM TOG, 25 (3), 2006. 2", + "[53] Yunzhou Song, Heguang Lin, Jiahui Lei, Lingjie Liu, and Kostas Daniilidis. HDGS: Textured 2D Gaussian Splitting for Enhanced Scene Rendering, 2024. 3", + "[54] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In CVPR, 2021. 1", + "[55] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Articulated Neural Radiance Fields for Learning Human Shape, Appearance, and Pose. In NeurIPS, 2021. 1", + "[56] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. BillBoard Splatting (BBSplat): Learnable Textured Primitives for Novel View Synthesis, 2025. 3", + "[57] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative Gaussian Splitting for Efficient 3D Content Creation. In ICLR, 2024. 1", + "[58] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in Neural Rendering. CGF, 41(2), 2022. 1, 2", + "[59] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction. In NeurIPS, 2021. 1", + "[60] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 1", + "[61] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-Nerf: Point-Based Neural Radiance Fields. In CVPR, 2022. 2", + "[62] Rui Xu, Wenyue Chen, Jiepeng Wang, Yuan Liu, Peng Wang, Lin Gao, Shiqing Xin, Taku Komura, Xin Li, and Wenping Wang. SuperGaussians: Enhancing Gaussian Splitting Using Primitives with Spatially Varying Colors, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8973", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Tian-Xing Xu, Wenbo Hu, Yu-Kun Lai, Ying Shan, and Song-Hai Zhang. Texture-GS: Disentangling the Geometry and Texture for 3D Gaussian Splitting Editing. In ECCV, 2024. 2, 3, 5", + "[64] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. In CVPR, 2024. 1, 8", + "[65] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian Grouping: Segment and Edit Anything in 3D Scenes. In ECCV, 2024. 2", + "[66] Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details in 3d gaussian splatt-ing. In ACM-MM, 2024. 6", + "[67] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for Real-Time Rendering of Neural Radiance Fields. In ICCV, 2021. 2", + "[68] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes. ACM TOG, 2024. 6, 1, 4", + "[69] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural Factorization of Shape and Reflectance under an Unknown Illumination. ACM TOG, 40(6), 2021. 1", + "[70] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM. In CVPR, 2022. 1" + ], + "bbox": [ + 91, + 92, + 480, + 487 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8974", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_model.json b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c8d4c4be62020f6e2cd3a97c4ba7b8f972bc8652 --- /dev/null +++ b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Textured Gaussians for Enhanced 3D Scene Appearance Modeling" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.179, + 0.88, + 0.219 + ], + "angle": 0, + "content": "Brian Chao \\(^{1,2}\\) Hung-Yu Tseng \\(^{2}\\) Lorenzo Porzi \\(^{2}\\) Chen Gao \\(^{2}\\) Tuotuo Li \\(^{2}\\) Qinbo Li \\(^{2}\\) Ayush Saraf \\(^{2}\\) Jia-Bin Huang \\(^{2,3}\\) Johannes Kopf \\(^{2}\\) Gordon Wetzstein \\(^{1}\\) Changil Kim \\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.228, + 0.785, + 0.25 + ], + "angle": 0, + "content": "\\(^{1}\\) Stanford University \\(^{2}\\) Meta \\(^{3}\\) University of Maryland College Park" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.258, + 0.658, + 0.272 + ], + "angle": 0, + "content": "https:// textured-gaussians.github.io" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.307, + 0.39, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.308, + 0.5, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.308, + 0.794, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.308, + 0.905, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.493, + 0.908, + 0.536 + ], + "angle": 0, + "content": "Figure 1. Textured Gaussians compared to 3D Gaussian Splatting (3DGS). Our RGBA Textured Gaussians enhance 3D scene appearance modeling, leading to improved rendering quality while using the same number of Gaussians compared to 3DGS [29]. Above, we show that Textured Gaussians faithfully reconstruct the fine details of scenes." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.548, + 0.327, + 0.563 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.586, + 0.486, + 0.903 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) has emerged as the state-of-the-art 3D reconstruction technique, offering high-quality results with fast training and rendering. However, its expressivity is limited as pixels covered by the same Gaussian share identical colors aside from a Gaussian falloff scaling factor, and individual Gaussians can only represent simple ellipsoids geometrically. To overcome these limitations, we integrate texture and alpha mapping from traditional graphics with 3DGS. Our approach augments each Gaussian with alpha, RGB, or RGBA texture maps to model spatially varying color and opacity across each Gaussian's extent. This allows Gaussians to represent richer texture patterns and geometric structures beyond single-color ellipsoids. Notably, alpha-only texture maps significantly improve Gaussian expressivity, while further augmenting with RGB texture maps achieve maximum expressivity. We validate our method on a wide variety of standard benchmark datasets and our own custom captures at both the object and scene levels, and demonstrate image quality improvements over existing methods while using a similar or lower number of Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.548, + 0.645, + 0.563 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.82 + ], + "angle": 0, + "content": "Neural rendering [58] achieves unprecedented novel-view synthesis and 3D reconstruction quality, offering solutions to a myriad of applications ranging from surface reconstruction [34, 59], SLAM [46, 70], material estimation and relighting [54, 69], to virtual teleportation and human avatar animation [55, 60]. Recently, 3D Gaussian Splatting (3DGS) [29] emerged as a state-of-the-art novel-view synthesis technique, with desirable properties such as high image quality results, fast training and rendering time, and an explicit primitives-based representation. Subsequent works in Gaussian Splatting have focused on topics such as improving surface reconstruction quality [16, 23], scene editing [10, 17], dynamic scene reconstruction [50, 64], and 3D generation [44, 57]. Despite the success, 3DGS sometimes fails to model detailed appearances, as demonstrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In 3DGS, each Gaussian can only represent a single color and the shape of an ellipsoid given a camera viewpoint. This dramatically limits the set of appearances and shapes a single Gaussian can represent. To address the issue, Huang and Gong [25] modified the viewing direction calculation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8964" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.272 + ], + "angle": 0, + "content": "in 3DGS such that pixels covered by the same Gaussian exhibit a smooth color and opacity variation. However, this color variation across a single Gaussian is minuscule due to the spherical harmonics representation of colors, preventing them from representing complex textures. Xu et al. [63] proposed only using Gaussians to represent scene geometry and leveraged a learned UV mapping module and a global 2D RGB texture map to project textures onto Gaussian surfaces. However, the unit-sphere representation of textures significantly constrains the capacity of their model, and thus, their method fails to reconstruct objects with complex geometry or large-scale scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.529 + ], + "angle": 0, + "content": "Our method builds upon 3DGS and draws inspiration from mesh-based 3D representations that model appearance using texture mapping. As shown in Figure 2, we augment each Gaussian with its alpha, RGB, or RGBA texture map so that each Gaussian is capable of representing a rich set of textures and shapes. We call this representation Textured Gaussians. With RGB textures, individual Gaussians can represent higher-frequency color variations. The alpha maps allow Gaussians to represent a wider variety of shapes, instead of just ellipsoids. To achieve this, we build custom CUDA kernels that perform ray-Gaussian intersection and texture mapping and integrate them with 3DGS. Consequently, our method not only inherits all desirable properties of 3DGS, such as fast training and rendering time and explicit 3D representation, but also substantially improves rendered image quality, especially with a small number of Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.53, + 0.371, + 0.544 + ], + "angle": 0, + "content": "In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.545, + 0.483, + 0.605 + ], + "angle": 0, + "content": "- Introduction of a generalized appearance model for 3D Gaussians that handles spatially varying color and opacity by augmenting 3D Gaussians with alpha, RGB, or RGBA texture maps." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.606, + 0.483, + 0.636 + ], + "angle": 0, + "content": "- Validation that Textured Gaussians improves 3DGS on a wide variety of both object-level and scene-level datasets." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.636, + 0.483, + 0.695 + ], + "angle": 0, + "content": "- Demonstration that our method greatly outperforms 3DGS when the number of Gaussians is small and achieves better performance than 3DGS with alpha-only textures given the same model size." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.545, + 0.483, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.233, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Novel-View Synthesis (NVS) and Neural Rendering. NVS tackles the problem of generating accurate renderings of 3D scenes from unseen viewpoints given a set of training images and camera poses. NVS traces back to the Structure from Motion (SfM) works [5, 19, 52] and Multiview Stereo (MVS) [15, 24, 48], which served as the basis of several groundbreaking NVS methods [8, 13, 20, 31]. However, these approaches generally require storing hundreds and thousands of input images for reprojection and blending, which leads to massive memory requirements and sometimes fails to reconstruct undersampled regions." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.907, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.446 + ], + "angle": 0, + "content": "Figure 2. Variants of our Textured Gaussians model. Textured Gaussians encapsulate four kinds of color and opacity spatial variations. The top row of the figure shows the texture map associated with each Gaussian, and the bottom row shows the rendered Textured Gaussians. The constant-color and constant-alpha model (no textures) corresponds to the original 3DGS formation, which can only represent a single color up to a Gaussian falloff factor within the Gaussian extent. Textured Gaussians can already model spatially varying colors using only alpha textures since each pixel can be alpha-composed differently. The model achieves maximum expressivity when leveraging the full RGBA texture map, where each Gaussian is capable of representing complex shapes and high frequency textures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.687 + ], + "angle": 0, + "content": "Recent advances in neural rendering [58] have made great strides in improving the quality of 3D reconstruction and novel-view rendering. Neural rendering algorithms can be roughly classified by the underlying 3D representation, ranging from point clouds [1, 31, 61], voxels [21, 47, 67], meshes [12, 18, 22, 28, 35], to implicit representations using MLPs [2, 37, 39, 51]. Recently, 3D Gaussian Splatting [29] (3DGS) has emerged as the state-of-the-art NVS technique. 3DGS achieves high image quality for NVS, is fast to optimize and render, and leverages an explicit Gaussian primitive representation, which makes the method suitable for a wide variety of tasks such as surface extraction [16, 23], human avatar manipulation [43, 49], ambient dynamic modeling [50], and object and scene editing [17, 65]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Our method augments 3DGS with alpha and texture mapping commonly used in mesh-based 3D representations. During optimization and rendering, we intersect outgoing rays from pixels with Gaussians in the scene. Intersection points are then used to query per-Gaussian texture map values via UV mapping and compute RGB color and alpha blending values. This enables each Gaussian to represent complex textures and shapes, significantly improving NVS quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Appearance Modeling in 3D Gaussian Splatting. In 3DGS, scene geometry is represented by the position, rotation, and scale of 3D Gaussians. By adjusting these properties, 3DGS can represent intricate geometric structures. On the other hand, appearance is modeled using per-Gaussian" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8965" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.123, + 0.669, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.118, + 0.897, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.348, + 0.908, + 0.419 + ], + "angle": 0, + "content": "Figure 3. Textured Gaussians model pipeline. Our method consists of three major components: ray-Gaussian intersection, RGBA texture mapping, and a generalized Gaussian appearance model. To render the color of a pixel \\(\\mathbf{p}\\), we first trace a ray from the camera center \\(\\mathbf{o}\\) to the pixel to intersect with 3D Gaussians in the scene. Then, we query texture and alpha values, \\(\\mathbf{c}_{\\mathrm{tex}}\\) and \\(\\alpha_{\\mathrm{tex}}\\), from the per-Gaussian RGBA texture maps using the ray-Gaussian intersection point \\(\\mathbf{x}\\). Finally, given the retrieved spatially varying color and alpha values, we alpha-composite the color and alpha values of Gaussians that are hit by the pixel ray using the generalized Gaussian appearance model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.444, + 0.482, + 0.685 + ], + "angle": 0, + "content": "opacity values and spherical harmonics coefficients. However, pixels within a projected Gaussian are always shaded with the same color up to a Gaussian falloff factor, greatly limiting the expressivity of individual Gaussians. Therefore, recent works have explored ways that allow Gaussians to represent spatially varying features. Huang and Gong [25] defined per-Gaussian SH coefficients for color and opacity and used ray-Gaussian intersections to determine the viewing directions within the Gaussian extent. This enables smooth variation of colors and opacities across pixels covered by a single Gaussian but prevents the reconstruction of higher-frequency details. Xu et al. [63] disentangled appearance and geometry in 3DGS for texture editing, but their method is limited to object-centric scenes with simple geometry due to the unit-sphere parameterization of textures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Instead of using a global texture map, we enhance each 3D Gaussian with a local texture map on top of the SH coefficients. This allows for a more flexible texture optimization since individual Gaussians are not tied to a shared global texture map. Our method can, therefore, reconstruct objects with complex structures and real-world scenes. Furthermore, each Gaussian can also represent various shapes with the alpha channel in the texture map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Memory Efficient Gaussian Splatting. In order to achieve even faster rendering time and compact storage requirements for potential applications on edge devices, there has been a plethora of recent work focusing on optimizing memory-efficient Gaussian Splatting models [3]. These algorithms either prune Gaussians and perform quantization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.905, + 0.476 + ], + "angle": 0, + "content": "of Gaussian attributes [14, 33, 38, 40-42] or exploit the structural relationships between Gaussians [11, 36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.584 + ], + "angle": 0, + "content": "Our work is orthogonal and complementary to these works since we focus on improving the appearance modeling of Gaussians by redistributing the model size budget to the per-Gaussian texture maps given any optimized 3DGS model. These model compression algorithms can be easily integrated with the 3DGS pretraining stage in our optimization pipeline to achieve further gains in compactness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.691 + ], + "angle": 0, + "content": "Concurrent Work. Different from several concurrent works that have explored the idea of augmenting Gaussians with texture maps [45, 53, 56, 62], we 1) additionally utilize alpha textures and 2) provide an in-depth analysis of various texture map variants (alpha-only, RGB, RGBA) that explores the balance between reconstruction capability and model size." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.715, + 0.605, + 0.731 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.776, + 0.76 + ], + "angle": 0, + "content": "3.1. 3D Gaussian Splitting Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.768, + 0.906, + 0.86 + ], + "angle": 0, + "content": "In 3DGS [29], 3D scenes are represented with 3D Gaussians, and images are rendered using differentiable volume splatting. Specifically, 3DGS explicitly defines 3D Gaussians by their 3D covariance matrix \\(\\pmb{\\Sigma}_i\\in \\mathbb{R}^{3\\times 3}\\) and center \\(\\pmb{\\mu}_i\\in \\mathbb{R}^3\\) (the index \\(i\\) indicating the \\(i^{\\mathrm{th}}\\) Gaussian), where the 3D Gaussian function value at point \\(\\mathbf{x}\\in \\mathbb{R}^3\\) is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.876, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {i} (\\mathbf {x}) = \\exp \\left(- \\frac {1}{2} \\left(\\mathbf {x} - \\boldsymbol {\\mu} _ {i}\\right) \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\left(\\mathbf {x} - \\boldsymbol {\\mu} _ {i}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8966" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.212 + ], + "angle": 0, + "content": "where the covariance matrix \\(\\boldsymbol{\\Sigma} = \\mathbf{R}\\mathbf{S}\\mathbf{S}^{\\top}\\mathbf{R}^{\\top}\\) is factorized into the rotation matrix \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\) and the scale matrix \\(\\mathbf{S} \\in \\mathbb{R}^{3 \\times 3}\\). To render a 2D image from a 3D Gaussians representation, 3D Gaussians are transformed from world coordinates to camera coordinates via a world-to-camera transformation matrix \\(\\mathbf{W} \\in \\mathbb{R}^{3 \\times 3}\\) and projected to the 2D image plane via a local affine transformation \\(\\mathbf{J} \\in \\mathbb{R}^{3 \\times 3}\\). The transformed 3D covariance \\(\\boldsymbol{\\Sigma}'\\) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.223, + 0.483, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {\\top} \\mathbf {J} ^ {\\top} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.253, + 0.484, + 0.315 + ], + "angle": 0, + "content": "The covariance \\(\\Sigma_{2\\mathrm{D}}\\) of the 2D Gaussian \\(\\mathcal{G}^{2\\mathrm{D}}\\) splatted on the image plane can be approximated as extracting the first two rows and columns of the transformed 3D covariance, \\(\\Sigma_{2\\mathrm{D}} = [\\Sigma^{\\prime}]_{\\{1,2\\},\\{1,2\\}} \\in \\mathbb{R}^{2 \\times 2}\\) using the matrix minor notation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.372 + ], + "angle": 0, + "content": "To render the color of a pixel \\(\\mathbf{p} \\in \\mathbb{R}^3\\), the colors associated with each Gaussian are alpha-composed from front to back following the conventional volume rendering equation:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.372, + 0.483, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} (\\mathbf {p}) = \\sum_ {i = 1} ^ {K} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.484, + 0.51 + ], + "angle": 0, + "content": "where \\(i\\) is the index of the Gaussians, \\(\\mathbf{c}_i\\) is the color of each Gaussian computed from the per-Gaussian spherical harmonic coefficients and viewing direction, and \\(\\alpha_{i}\\) is the alpha value computed from the opacity \\(o_i\\) associated with each Gaussian and the 2D Gaussian value evaluated at pixel location \\(\\mathbf{p}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.511, + 0.482, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\mathcal {G} _ {i} ^ {\\mathrm {2 D}} (\\mathbf {p}) \\cdot o _ {i} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.597 + ], + "angle": 0, + "content": "The attributes (center, rotation, scale, opacity, and spherical harmonic coefficients) of the Gaussians are optimized with gradient descent using photometric losses on the rendered 2D images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.279, + 0.621 + ], + "angle": 0, + "content": "3.2. Textured Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.659 + ], + "angle": 0, + "content": "From the 3DGS appearance model defined in Eq. 3, we observe two properties of Gaussian Splitting:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.659, + 0.483, + 0.689 + ], + "angle": 0, + "content": "1. Pixels covered by the same Gaussian are shaded with the same color up to a Gaussian falloff scaling factor." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.69, + 0.483, + 0.719 + ], + "angle": 0, + "content": "2. The per-Gaussian opacity only allows Gaussians to represent ellipsoidal shapes." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.659, + 0.483, + 0.719 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "These two properties greatly restrict the expressivity of individual 3D Gaussian primitives. To allow each Gaussian primitive to represent complex appearances and shapes, we assign a fixed-resolution 2D texture map of size \\(\\mathcal{T} \\times \\mathcal{T} \\times \\mathcal{K}, \\mathcal{T} \\in \\mathbb{N}, \\mathcal{K} \\in \\{1,3,4\\}\\) to each Gaussian and transform each Gaussian into a Textured Gaussian. As shown in Figure 2, \\(\\mathcal{K} = 1,3,4\\) correspond to alpha, RGB, and RGBA texture maps, respectively. The shape of a Textured Gaussian is thus determined by the spatially varying opacity defined by the product of the per-Gaussian opacity and the alpha channel of the texture map. The appearance of a Textured Gaussian is represented by 1) the RGB channels of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "the spatially varying textured map and 2) a set of spatially constant spherical harmonic coefficients (SH). Intuitively, the spatially constant SH coefficients represent the low-frequency textures plus view-dependent colors like specular effects, while the spatially varying texture map represents the higher-frequency spatial variations of the texture. The texture is then mapped to the plane \\(\\mathcal{P}\\) defined by the two major axes of the \\(i^{\\mathrm{th}}\\) 3D Gaussian that is centered at \\(\\pmb{\\mu}_i\\in \\mathbb{R}^3\\), as shown in Figure 3. The normal \\(\\mathbf{n}_i\\in \\mathbb{R}^3\\) of this plane thus corresponds to the axis with the smallest scale." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.304 + ], + "angle": 0, + "content": "For each pixel \\(\\mathbf{p} \\in \\mathbb{R}^3\\) that is currently being rendered, we cast a ray from the camera origin \\(\\mathbf{o} \\in \\mathbb{R}^3\\) to the center of the pixel \\(\\mathbf{p}\\) and intersect it with the plane \\(\\mathcal{P}\\). The intersection point \\(\\mathbf{x} \\in \\mathbb{R}^3\\) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.316, + 0.905, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = \\mathbf {o} + \\frac {\\left(\\boldsymbol {\\mu} _ {i} - \\mathbf {o}\\right) \\cdot \\mathbf {n} _ {i}}{\\left(\\mathbf {p} - \\mathbf {o}\\right) \\cdot \\mathbf {n} _ {i}} \\cdot \\frac {\\mathbf {p} - \\mathbf {o}}{\\| \\mathbf {p} - \\mathbf {o} \\|} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Given the intersection point \\(\\mathbf{x}\\) with \\(\\mathcal{P}\\), we perform UV mapping and bilinear interpolation on the texture map associated with the \\(i^{\\mathrm{th}}\\) Gaussian to query the texture color and alpha value at the UV coordinates \\(u, v \\in \\mathbb{R}\\), denoted by \\(\\mathbf{c}_i^{\\mathrm{tex}}(u, v) \\in \\mathbb{R}^3\\) and \\(\\alpha_i^{\\mathrm{tex}}(u, v) \\in \\mathbb{R}\\). Specifically, the UV coordinates \\((u, v)\\) of the intersection point on the texture map can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.489, + 0.904, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} u = \\frac {m \\cdot \\sigma_ {i} ^ {1} + (\\mathbf {x} - \\boldsymbol {\\mu} _ {i}) \\cdot \\mathbf {r} _ {i} ^ {1}}{2 \\cdot m \\cdot \\sigma_ {i} ^ {1}} \\cdot (\\mathcal {T} - 1) \\\\ v = \\frac {m \\cdot \\sigma_ {i} ^ {2} + (\\mathbf {x} - \\boldsymbol {\\mu} _ {i}) \\cdot \\mathbf {r} _ {i} ^ {2}}{2 \\cdot m \\cdot \\sigma_ {i} ^ {2}} \\cdot (\\mathcal {T} - 1) \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.906, + 0.647 + ], + "angle": 0, + "content": "where \\(\\sigma_i^1, \\sigma_i^2 \\in \\mathbb{R}\\) are the scales of the two major axis of the \\(i^{\\text{th}}\\) Gaussian, \\(\\mathbf{r}_i^1, \\mathbf{r}_i^2 \\in \\mathbb{R}^3\\) are the normalized directions of the two major axes of the \\(i^{\\text{th}}\\) Gaussian, and \\(m \\in \\mathbb{R}\\) is a scalar multiplier that determines the extent of the texture map with respect to each Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.906, + 0.722 + ], + "angle": 0, + "content": "Combined with the color computed from the SH coefficients, which we call \\(\\mathbf{c}_i^{\\mathrm{base}}\\) (the same as the color component in the original 3DGS appearance model in Eq. 3), the final color contribution of the \\(i\\)-th Gaussian to pixel \\(\\mathbf{p}\\) is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.723, + 0.905, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {i} (\\mathbf {p}) = \\mathbf {c} _ {i} ^ {\\text {b a s e}} + \\mathbf {c} _ {i} ^ {\\text {t e x}} (u, v) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.905, + 0.778 + ], + "angle": 0, + "content": "and the alpha value of the \\(i\\)-th Gaussian at pixel \\(\\mathbf{p}\\) is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.779, + 0.905, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} (\\mathbf {p}) = \\alpha_ {i} ^ {\\mathrm {t e x}} (u, v) \\cdot \\mathcal {G} _ {i} (\\mathbf {x}) \\cdot o _ {i} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.804, + 0.905, + 0.85 + ], + "angle": 0, + "content": "Finally, to render the color of a pixel \\(\\mathbf{p}\\), we modify the 3DGS appearance model in Eq. 3 to incorporate the spatially varying texture and opacity:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.861, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} ^ {\\text {f i n a l}} (\\mathbf {p}) = \\sum_ {i = 1} ^ {K} \\mathbf {c} _ {i} (\\mathbf {p}) \\alpha_ {i} (\\mathbf {p}) \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} (\\mathbf {p})) \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8967" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.912, + 0.257 + ], + "angle": 0, + "content": "
MethodBlender [37]Mip-NeRF 360 [37]DTU [27]Tanks and Temples [30]Deep Blending [20]
Mip-NeRF 360 [4]30.34 / 0.9510 / 0.060027.69 / 0.7920 / 0.237034.28 / 0.9641 / 0.069822.22 / 0.7590 / 0.257029.40 / 0.9010 / 0.2450
Instant-NGP [39]32.20 / 0.9590 / 0.055025.30 / 0.6710 / 0.371023.58 / 0.7653 / 0.260321.72 / 0.7230 / 0.330023.62 / 0.7970 / 0.4230
Xu et al. [63]28.97 / 0.9380 / 0.0550— / — / —30.03 / — / 0.1440— / — / —— / — / —
3DGS*33.08 / 0.9671 / 0.044027.26 / 0.8318 / 0.187133.54 / 0.9697 / 0.055124.18 / 0.8541 / 0.175428.04 / 0.8940 / 0.2707
Ours33.24 / 0.9674 / 0.042827.35 / 0.8274 / 0.185833.61 / 0.9699 / 0.055624.26 / 0.8542 / 0.168428.33 / 0.8908 / 0.2699
3DGS* (10%)31.47 / 0.9590 / 0.059425.77 / 0.7796 / 0.286032.71 / 0.9627 / 0.081122.82 / 0.8020 / 0.272827.64 / 0.8853 / 0.3101
Ours (10%)32.14 / 0.9629 / 0.048926.32 / 0.7976 / 0.232332.74 / 0.9661 / 0.058223.41 / 0.8259 / 0.212227.98 / 0.8898 / 0.2804
3DGS* (1%)26.89 / 0.9160 / 0.116522.37 / 0.6293 / 0.477430.88 / 0.9320 / 0.158119.90 / 0.6736 / 0.440623.97 / 0.8167 / 0.4337
Ours (1%)28.11 / 0.9343 / 0.084923.73 / 0.7064 / 0.336532.43 / 0.9627 / 0.069421.10 / 0.7399 / 0.310424.83 / 0.8454 / 0.3552
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.266, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of different novel view synthesis methods. We boldface the best-performing model and underline the second best in terms of different metrics (PSNR ↑ / SSIM ↑ / LPIPS ↓). Our Textured Gaussians model achieves better performance than 3DGS* across most metrics. With fewer Gaussians (1% and 10% of the default optimized number of Gaussians), our method significantly outperforms 3DGS*, achieving strictly better performance across all metrics as indicated by the bold text." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.571 + ], + "angle": 0, + "content": "Figure 4. NVS performance using varying numbers of Gaussians. When using the same number of Gaussians as \\(3\\mathrm{DGS}^*\\), our Textured Gaussians achieve better novel view synthesis results on all five benchmark datasets in PSNR. Here, we show the quantitative performance of \\(3\\mathrm{DGS}^*\\) and Textured Gaussians models optimized with varying number of Gaussians on the three scene-level datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.661 + ], + "angle": 0, + "content": "Eq. 9 is a generalized formulation of 3D Gaussian appearance and encapsulates different variants of Textured Gaussians. For example, \\(\\mathbf{c}_i^{\\mathrm{tex}} = 0\\) and \\(\\alpha_{i}^{\\mathrm{tex}} = 1\\) correspond to the original 3DGS model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.406, + 0.687 + ], + "angle": 0, + "content": "3.3. Optimization of Textured Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.723 + ], + "angle": 0, + "content": "Following 3DGS [29], we optimize our model to minimize the weighted photometric loss:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.736, + 0.483, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\mathcal {L} _ {1} + (1 - \\lambda) \\mathcal {L} _ {\\text {S S I M}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.197, + 0.779 + ], + "angle": 0, + "content": "where \\(\\lambda = 0.8\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our optimization procedure consists of two stages. We first optimize a 3DGS model for 30000 iterations. The learning rates and adaptive density control (ADC) parameters are reported in the Appendix. In the second stage, we initialize all attributes of the Gaussians with the optimized vanilla 3DGS model, and jointly optimize them with the per-Gaussian 2D texture maps for another 30000 iterations. We disable the ADC control in the second stage to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.409 + ], + "angle": 0, + "content": "control the number of Gaussians for fair baseline comparisons. This two-stage optimization process greatly speeds up convergence and improves image quality, since jointly optimizing all parameters is a highly ill-posed problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.485 + ], + "angle": 0, + "content": "We implement custom CUDA kernels to perform fast ray-Gaussian intersection, UV mapping, and color composition. All experiments are conducted on clusters of Nvidia H100 GPUs. Please refer to the Appendix for more details on implementation and optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.498, + 0.714, + 0.515 + ], + "angle": 0, + "content": "4. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.689 + ], + "angle": 0, + "content": "We show selected qualitative and quantitative results in this section and refer the readers to the Appendix and the project website for an extensive set of results and video renderings. Datasets and Evaluation Protocols. We evaluate the novel-view synthesis performance of our method on the 8 synthetic scenes from the Blender dataset [37], all 9 scenes from the Mip-NeRF 360 dataset [4], 5 scenes from the DTU dataset [27], 2 scenes from the Tanks and Temples dataset [30], and 2 scenes from the Deep Blending dataset [20]. Please refer to the Appendix for a detailed description of the dataset preparation process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.794 + ], + "angle": 0, + "content": "For comparisons with Mip-NeRF 360 [4] and InstantNGP [39], we refer to Kerbl et al. [29] for results in scene-level datasets [4, 20, 30], and refer to NeRFBaselines [32] for results on the Blender dataset [37]. Results on the DTU dataset [27] are from our own runs of the authors' released code. Please refer to the Appendix for additional experimental details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Xu et al. [63] pointed out in Section 5.4 of their paper that their algorithm cannot handle scene-level datasets and objects with complex structures (i.e. Blender dataset), thus they only reported results on the DTU dataset [27]. We report those numbers and additionally show results on the Blender dataset from GStex [45] authors' own runs of the algorithm described in Xu et al. [63]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8968" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.097, + 0.221, + 0.109 + ], + "angle": 0, + "content": "3DGS*" + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.098, + 0.414, + 0.109 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.097, + 0.624, + 0.109 + ], + "angle": 0, + "content": "3DGS*" + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.098, + 0.818, + 0.109 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.113, + 0.904, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.548, + 0.908, + 0.604 + ], + "angle": 0, + "content": "Figure 5. Qualitative NVs results of benchmark datasets. Novel-view rendering results for both object-level and scene-level standard benchmark datasets. Given the same small number of Gaussians (on average \\(2.5\\mathrm{k}\\) and \\(39\\mathrm{k}\\) Gaussians for object-level and scene-level datasets, respectively), \\(3\\mathrm{DGS}^*\\) fails to reconstruct high-frequency textures and complex shapes while our RGBA Textured Gaussians model succeeds. Please refer to the Appendix for results of our models optimized using varying number of Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.689 + ], + "angle": 0, + "content": "In addition to standard benchmark test scenes, we also capture our own set of scenes that contain artworks with highly detailed textures to help demonstrate the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3DGS Baseline. We build our method on top of our own implementation of a slightly modified 3DGS that closely follows the algorithm described in Gaussian Opacity Fields [68]. Specifically, we use a ray-based formulation that allows for the exact evaluation of 3D Gaussian values without approximating splatted 3D Gaussians as 2D Gaussians. In addition, we used a revised densification strategy [66, 68]. For fair comparisons with our method, we report the performance of our 3DGS implementation and label it as \\(3\\mathrm{DGS}^*\\) throughout the results section, and refer the readers to the Appendix for quantitative results reported in the original 3DGS paper [29]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.63, + 0.707, + 0.645 + ], + "angle": 0, + "content": "4.1. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.762 + ], + "angle": 0, + "content": "In Table 1, we show quantitative comparisons in terms of PSNR/SSIM/LPIPS between our Textured Gaussians model using full RGBA textures and 3DGS* with the same number of Gaussians, and results of other baseline methods. Our method generally achieves better results than 3DGS* and all other methods, since Texture Gaussians are more expressive than vanilla 3D Gaussians due to the added texture maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In the bottom half of Table 1, we also show comparisons between our method and \\(3\\mathrm{DGS}^*\\) when using a fraction of the default number of Gaussians, denoted by the percentage in parentheses. For our models with different numbers of Gaussians, we distribute a fixed amount of texels to all Gaussians (i.e., the same memory overhead for all Textured Gaussians models compared with \\(3\\mathrm{DGS}^*\\)). Therefore, the texture map resolution of models with fewer Gaussians will be larger. We see that our method truly outshines \\(3\\mathrm{DGS}^*\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8969" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.097, + 0.221, + 0.109 + ], + "angle": 0, + "content": "3DGS*" + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.097, + 0.414, + 0.11 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.113, + 0.496, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.097, + 0.625, + 0.109 + ], + "angle": 0, + "content": "3DGS*" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.112, + 0.905, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.294, + 0.908, + 0.322 + ], + "angle": 0, + "content": "Figure 6. Qualitative NVS results of custom datasets. Our RGBA Textured Gaussians model achieves sharper reconstruction compared to \\(3\\mathrm{DGS}^*\\) when using the same small number of Gaussians (on average around 100k Gaussians)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.346, + 0.485, + 0.447 + ], + "angle": 0, + "content": "
Method3DGS*Alpha-onlyRGBA
PSNR#GS (M)PSNR#GS (M)PSNR#GS (M)
Blender33.080.2733.120.1933.030.21
Mip-NeRF 36027.264.427.373.127.263.5
DTU33.530.2833.450.2433.410.22
Tanks and Temples24.172.924.382.624.281.3
Deep Blending28.042.828.361.628.521.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.568 + ], + "angle": 0, + "content": "Table 2. Same model size comparisons. We compare the PSNR performance between \\(3\\mathrm{DGS}^*\\) and our model using alpha-only and RGBA textures with the same model size. The average number of optimized Gaussians (in millions) from each dataset is shown in the #GS (M) columns. We boldface and underline the best and second-best performing models, respectively. Alpha-only texture models generally achieve better performance than \\(3\\mathrm{DGS}^*\\) and RGBA texture models with the same model size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.685 + ], + "angle": 0, + "content": "with fewer Gaussians, achieving nearly 2 dB improvements over \\(3\\mathrm{DGS}^*\\) when using \\(1\\%\\) of the default number of Gaussians. Figure 4 shows the trend of the quality of novel view synthesis as we vary the number of Gaussians. Our method consistently outperforms \\(3\\mathrm{DGS}^*\\) when using different numbers of Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Table 2 compares the performance of \\(3\\mathrm{DGS}^*\\) and our Textured Gaussians model with the same model size using alpha-only textures and RGBA textures. Since our models use per-Gaussian texture maps, they require fewer Gaussians than \\(3\\mathrm{DGS}^*\\) to achieve the same model size. Our alpha-only textures model generally outperforms the \\(3\\mathrm{DGS}^*\\) and RGBA textures models, indicating that there is a sweet spot for distributing the model size budget between Gaussian parameters and texture map channels. This is further explored in Section 4.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.275, + 0.864 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Novel-View Synthesis. We show novel view synthesis results of our model with RGBA textures and 3DGS* on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.438 + ], + "angle": 0, + "content": "both standard benchmark and custom-captured datasets in Figures 5 and 6. Our method reconstructs much sharper details of the scene compared to \\(3\\mathrm{DGS}^*\\) when using the same small number of Gaussians. Please refer to the Appendix for results of our models optimized using varying number of Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.909, + 0.607 + ], + "angle": 0, + "content": "Color Component Decomposition. We show the alpha-modulated and composited results of the two color components, \\(\\mathbf{c}^{\\mathrm{tex}}\\) and \\(\\mathbf{c}^{\\mathrm{base}}\\), of the optimized alpha-only, RGB and RGBA textured Gaussians model with \\(1\\%\\) of the default number of Gaussians in Figure 7. A 3DGS* model with the same number of Gaussians is shown for comparison. For RGB and RGBA models, \\(\\mathbf{c}^{\\mathrm{tex}}\\) reconstructs fine-grained textures that 3DGS* cannot. For alpha-only and RGBA models, \\(\\mathbf{c}^{\\mathrm{base}}\\) also reconstructs high-frequency details due to spatially varying alpha composition enabled by alpha textures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.68, + 0.635 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Texture Map Variants. We ablate variants of our model that use different texture maps, namely alpha-only, RGB, and RGBA texture maps, and the same number of Gaussians in Figure 8. Experiments are conducted on Blender, Tanks and Temples, and Deep Blending datasets with the default optimized number (top row) and \\(1\\%\\) of the default optimized number (bottom row) of Gaussians. We see that our model achieves the best performance with RGBA textures. Interestingly, using alpha-only textures already outperforms \\(3\\mathrm{DGS}^*\\), and our RGB textures model despite being one-third the size, striking a better balance between performance and model size. This is because alpha-textured Gaussians can represent complex shapes through spatially varying opacity and reconstruct high-frequency textures through spatially varying alpha composition. In contrast, RGB-textured Gaussians are still limited to representing ellipsoids." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8970" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.905, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.279, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 7. Color component decomposition visualization. Visualization of alpha-modulated and composited \\(\\mathbf{c}^{\\mathrm{base}}\\) and \\(\\mathbf{c}^{\\mathrm{tex}}\\) color components in our Textured Gaussians model (Eq. 7, 8, 9). Insets with yellow borders show zoom-in crops. Texture color brightness is adjusted for visualization. While \\(3\\mathrm{DGS}^{*}\\) fails to capture fine details with limited Gaussians, our method reconstructs sharp textures. With RGB textures, base color reconstructs low-frequency textures and texture maps reconstruct high-frequency elements. With alpha-only and RGBA textures, base color also reconstructs high-frequency textures via spatially varying alpha compositing." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.508, + 0.486, + 0.607 + ], + "angle": 0, + "content": "Figure 8. Ablation study on texture map variants. We show the PSNR performance improvements (\\(\\Delta\\)PSNR) over \\(3\\mathrm{DGS}^*\\) of our Textured Gaussians model when using different texture map variants. Models with full RGBA textures achieve the best results. Interestingly, alpha-only textures outperform RGB textures and perform comparably to RGBA textures while using one-third and one-fourth of the model size, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.634, + 0.483, + 0.816 + ], + "angle": 0, + "content": "Texture Map Resolution and the Number of Gaussians. In Figure 9, we optimize alpha-only Textured Gaussians models of the same model size with different texture map resolutions and number of Gaussians, and show the trend of novel view synthesis performance on the Blender and DTU datasets. We observe a sweet spot between texture map resolution and number of Gaussians that achieves the best image reconstruction quality. Maximizing texture map resolution corresponds to greatly reducing the number of Gaussians, which harms reconstruction of detailed geometric structures. Minimizing the texture map resolution, on the other hand, simply degenerates into 3DGS*." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.214, + 0.846 + ], + "angle": 0, + "content": "5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Limitations. The use of 2D diffuse texture maps in our model assumes that all textures lie on a local surface and does not model spatially varying specular color. Thus, ex" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.373, + 0.903, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.617 + ], + "angle": 0, + "content": "Figure 9. Ablation study on texture map resolution and the number of Gaussians given the same model size. We optimized various alpha-only Textured Gaussians models (red dots) with equivalent model sizes but different texture map resolutions and numbers of Gaussian. A fitted quadratic curve (blue dashed line) illustrates the performance trend. The optimal model variant balances these two parameters to achieve superior novel view synthesis results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.765 + ], + "angle": 0, + "content": "tending our texture representation to represent local 3D volume textures or even 5D radiance fields is crucial. Using factorized representations such as TensoRF [9] or triplane [7] to represent these high-dimensional textures could also be an interesting research problem. Finally, extending Textured Gaussians to support dynamic scene reconstruction [50, 64] through time-varying texture maps could be a natural next-step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Conclusions. In this paper, we augment 3DGS with texture maps to allow individual Gaussians to model spatially varying colors and opacity. As such, each Gaussian can represent a much richer set of appearances and shapes. This greatly improves the expressivity of individual Gaussians, leading to better novel-view synthesis quality. Our method achieves better quality than 3DGS when using the same number of Gaussians, and achieves better or comparable quality when using the same model size." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "8971" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.191 + ], + "angle": 0, + "content": "We thank artist Kija Lucas for allowing us to use her art installation at the Palo Alto Art Center as one of our custom captured scenes (the flower gallery scene). Brian Chao is supported by the Stanford Graduate Fellowship and the NSF GRFP." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.206, + 0.187, + 0.222 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.231, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[1] K. Aliev, A. Sevastopolsky, M. Kolos, D. Ulyanov, and V. Lempitsky. Neural Point-Based Graphics, 2019. arXiv preprint. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.275, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[2] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.483, + 0.373 + ], + "angle": 0, + "content": "[3] Milena T. Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3DGS.zip: A survey on 3D Gaussian Splitting Compression Methods, 2024. arXiv preprint. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In CVPR, 2022. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.482, + 0.458 + ], + "angle": 0, + "content": "[5] Matthew Brown and David G Lowe. Unsupervised 3D Object Recognition and Reconstruction in Unordered Datasets. In 3DIM, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[6] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising Densification in Gaussian Splatting. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient Geometry-Aware 3D Generative Adversarial Networks. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[8] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth Synthesis and Local Warps for Plausible Image-Based Navigation. ACM TOG, 32 (3), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[10] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splitting. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[11] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: Hash-Grid Assisted Context for 3D Gaussian Splitting Compression. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[12] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[13] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating Textures. In CGF, 2008. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.231, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[14] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and \\(200+\\) FPS. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[15] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-View Stereo for Community Photo Collections. In ICCV, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[16] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In CVPR, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.252, + 0.904, + 0.292 + ], + "angle": 0, + "content": "[17] Antoine Guédon and Vincent Lepetit. Gaussian Frosting: Editable Complex Radiance Fields with Real-Time Rendering. In ECCV, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[18] Yuan-Chen Guo, Yan-Pei Cao, Chen Wang, Yu He, Ying Shan, and Song-Hai Zhang. VMesh: Hybrid Volume-Mesh Representation for Efficient View Synthesis. In SIGGRAPH Asia, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.353, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[19] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, second edition, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.396, + 0.905, + 0.451 + ], + "angle": 0, + "content": "[20] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. ACM TOG, 37(6), 2018. 2, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.454, + 0.904, + 0.494 + ], + "angle": 0, + "content": "[21] D. Hu, Z. Zhang, T. Hou, T. Liu, H. Fu, and M. Gong. Multiscale Representation for Real-Time Anti-Aliasing Neural Rendering. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[22] Ronghang Hu, Nikhila Ravi, Alexander C. Berg, and Deepak Pathak. Worldsheet: Wrapping the World in a 3D Sheet for View Synthesis from a Single Image. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[23] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In SIGGRAPH, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.584, + 0.905, + 0.625 + ], + "angle": 0, + "content": "[24] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.628, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[25] Zhentao Huang and Minglun Gong. Textured-GS: Gaussian Splatting with Spatially Defined Color and Opacity, 2024. arXiv preprint. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.672, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[26] Junha Hyung, Susung Hong, Sungwon Hwang, Jaeseong Lee, Jaegul Choo, and Jin-Hwa Kim. Effective rank analysis and regularization for enhanced 3d gaussian splatting. In NeuRIPS, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.729, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[27] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large Scale Multi-View Stereopsis Evaluation. In CVPR, 2014. 5, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.904, + 0.799 + ], + "angle": 0, + "content": "[28] H. Kato, Y. Ushiku, and T. Harada. Neural 3D Mesh Renderer. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splitting for Real-Time Radiance Field Rendering. ACM TOG, 42(4), 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. ACM TOG, 36(4), 2017. 5, 2, 9" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8972" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with PerView Optimization. CGF, 40(4), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.178 + ], + "angle": 0, + "content": "[32] Jonas Kulhanek and Torsten Sattler. NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods, 2024. arXiv preprint. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[33] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3D Gaussian Representation for Radiance Field. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[34] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[35] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep Appearance Models for Face Rendering. ACM TOG, 37(4), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[36] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 2, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[38] Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3D Scene Representation via Self-Organizing Gaussian Grids. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM TOG, 41(4), 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[40] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[41] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3D Gaussian Splitting for Accelerated Novel View Synthesis. In CVPR, pages 10349-10358, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[42] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the Memory Footprint of 3D Gaussian Splatting. ACM CGIT, 7(1), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[43] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[44] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. DreamGaussian4D: Generative 4D Gaussian Splatting, 2023. arXiv preprint. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[45] Victor Rong, Jingxiang Chen, Sherwin Bahmani, Kiriakos N Kutulakos, and David B Lindell. GStex: Per-Primitive Texturing of 2D Gaussian Splatting for Decoupled Appearance and Geometry Modeling, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[46] Antoni Rosinol, John J Leonard, and Luca Carlone. NeRF-SLAM: Real-time Dense Monocular Slam with Neural Radiance Fields. In IROS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.898 + ], + "angle": 0, + "content": "[47] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields without Neural Networks. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[48] Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A Comparison and Evaluation of Multi-View Stereo Reconstruction Algorithms. In CVPR, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[49] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.209, + 0.905, + 0.249 + ], + "angle": 0, + "content": "[50] Meng-Li Shih, Jia-Bin Huang, Changil Kim, Rajvi Shah, Johannes Kopf, and Chen Gao. Modeling ambient scene dynamics for free-view synthesis. 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.252, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[51] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene Representation Networks: Continuous 3D-Structure-Aware Neural Scene Representations. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.905, + 0.351 + ], + "angle": 0, + "content": "[52] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. ACM TOG, 25 (3), 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.905, + 0.395 + ], + "angle": 0, + "content": "[53] Yunzhou Song, Heguang Lin, Jiahui Lei, Lingjie Liu, and Kostas Daniilidis. HDGS: Textured 2D Gaussian Splitting for Enhanced Scene Rendering, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[54] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[55] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Articulated Neural Radiance Fields for Learning Human Shape, Appearance, and Pose. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.554 + ], + "angle": 0, + "content": "[56] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. BillBoard Splatting (BBSplat): Learnable Textured Primitives for Novel View Synthesis, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.557, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[57] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative Gaussian Splitting for Efficient 3D Content Creation. In ICLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.667 + ], + "angle": 0, + "content": "[58] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in Neural Rendering. CGF, 41(2), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[59] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[60] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[61] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-Nerf: Point-Based Neural Radiance Fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[62] Rui Xu, Wenyue Chen, Jiepeng Wang, Yuan Liu, Peng Wang, Lin Gao, Shiqing Xin, Taku Komura, Xin Li, and Wenping Wang. SuperGaussians: Enhancing Gaussian Splitting Using Primitives with Spatially Varying Colors, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8973" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[63] Tian-Xing Xu, Wenbo Hu, Yu-Kun Lai, Ying Shan, and Song-Hai Zhang. Texture-GS: Disentangling the Geometry and Texture for 3D Gaussian Splitting Editing. In ECCV, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[64] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. In CVPR, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[65] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian Grouping: Segment and Edit Anything in 3D Scenes. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[66] Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details in 3d gaussian splatt-ing. In ACM-MM, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[67] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for Real-Time Rendering of Neural Radiance Fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[68] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes. ACM TOG, 2024. 6, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[69] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural Factorization of Shape and Reflectance under an Unknown Illumination. ACM TOG, 40(6), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[70] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM. In CVPR, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8974" + } + ] +] \ No newline at end of file diff --git a/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_origin.pdf b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8f54d757d9378011a1271b2b3eda98aa4f90b9f4 --- /dev/null +++ b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/d26b8141-c9cc-4aa4-b4d4-43213dcfab1a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88bc20fe2f212087ff50f39de4c30712c45641eabbf2881c22c11d6e1b29c83d +size 8278429 diff --git a/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/full.md b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..32ef8b29562ec25b88060506b408e7d4064b6b9d --- /dev/null +++ b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/full.md @@ -0,0 +1,319 @@ +# Textured Gaussians for Enhanced 3D Scene Appearance Modeling + +Brian Chao $^{1,2}$ Hung-Yu Tseng $^{2}$ Lorenzo Porzi $^{2}$ Chen Gao $^{2}$ Tuotuo Li $^{2}$ Qinbo Li $^{2}$ Ayush Saraf $^{2}$ Jia-Bin Huang $^{2,3}$ Johannes Kopf $^{2}$ Gordon Wetzstein $^{1}$ Changil Kim $^{2}$ + +$^{1}$ Stanford University $^{2}$ Meta $^{3}$ University of Maryland College Park + +https:// textured-gaussians.github.io + +![](images/f97c4011fb92fe6ea64381f0ce03e2ed665c9f349b25425d4df2e54e6f64abaa.jpg) +Figure 1. Textured Gaussians compared to 3D Gaussian Splatting (3DGS). Our RGBA Textured Gaussians enhance 3D scene appearance modeling, leading to improved rendering quality while using the same number of Gaussians compared to 3DGS [29]. Above, we show that Textured Gaussians faithfully reconstruct the fine details of scenes. + +![](images/28a9f048694eb262ce139fb32dabfb7025470cbf0422f0a158b02f292ae58823.jpg) + +![](images/6fe6cc101f9a79075c0731c30e6678267bd0cbaffc03490749feff24a6d6b166.jpg) + +![](images/f24741a0bf4541b7cc291439ab1b34ac28215b0fa81bc84fd2fea98269557ee8.jpg) + +# Abstract + +3D Gaussian Splatting (3DGS) has emerged as the state-of-the-art 3D reconstruction technique, offering high-quality results with fast training and rendering. However, its expressivity is limited as pixels covered by the same Gaussian share identical colors aside from a Gaussian falloff scaling factor, and individual Gaussians can only represent simple ellipsoids geometrically. To overcome these limitations, we integrate texture and alpha mapping from traditional graphics with 3DGS. Our approach augments each Gaussian with alpha, RGB, or RGBA texture maps to model spatially varying color and opacity across each Gaussian's extent. This allows Gaussians to represent richer texture patterns and geometric structures beyond single-color ellipsoids. Notably, alpha-only texture maps significantly improve Gaussian expressivity, while further augmenting with RGB texture maps achieve maximum expressivity. We validate our method on a wide variety of standard benchmark datasets and our own custom captures at both the object and scene levels, and demonstrate image quality improvements over existing methods while using a similar or lower number of Gaussians. + +# 1. Introduction + +Neural rendering [58] achieves unprecedented novel-view synthesis and 3D reconstruction quality, offering solutions to a myriad of applications ranging from surface reconstruction [34, 59], SLAM [46, 70], material estimation and relighting [54, 69], to virtual teleportation and human avatar animation [55, 60]. Recently, 3D Gaussian Splatting (3DGS) [29] emerged as a state-of-the-art novel-view synthesis technique, with desirable properties such as high image quality results, fast training and rendering time, and an explicit primitives-based representation. Subsequent works in Gaussian Splatting have focused on topics such as improving surface reconstruction quality [16, 23], scene editing [10, 17], dynamic scene reconstruction [50, 64], and 3D generation [44, 57]. Despite the success, 3DGS sometimes fails to model detailed appearances, as demonstrated in Figure 1. + +In 3DGS, each Gaussian can only represent a single color and the shape of an ellipsoid given a camera viewpoint. This dramatically limits the set of appearances and shapes a single Gaussian can represent. To address the issue, Huang and Gong [25] modified the viewing direction calculation + +in 3DGS such that pixels covered by the same Gaussian exhibit a smooth color and opacity variation. However, this color variation across a single Gaussian is minuscule due to the spherical harmonics representation of colors, preventing them from representing complex textures. Xu et al. [63] proposed only using Gaussians to represent scene geometry and leveraged a learned UV mapping module and a global 2D RGB texture map to project textures onto Gaussian surfaces. However, the unit-sphere representation of textures significantly constrains the capacity of their model, and thus, their method fails to reconstruct objects with complex geometry or large-scale scenes. + +Our method builds upon 3DGS and draws inspiration from mesh-based 3D representations that model appearance using texture mapping. As shown in Figure 2, we augment each Gaussian with its alpha, RGB, or RGBA texture map so that each Gaussian is capable of representing a rich set of textures and shapes. We call this representation Textured Gaussians. With RGB textures, individual Gaussians can represent higher-frequency color variations. The alpha maps allow Gaussians to represent a wider variety of shapes, instead of just ellipsoids. To achieve this, we build custom CUDA kernels that perform ray-Gaussian intersection and texture mapping and integrate them with 3DGS. Consequently, our method not only inherits all desirable properties of 3DGS, such as fast training and rendering time and explicit 3D representation, but also substantially improves rendered image quality, especially with a small number of Gaussians. + +In summary, our contributions include: + +- Introduction of a generalized appearance model for 3D Gaussians that handles spatially varying color and opacity by augmenting 3D Gaussians with alpha, RGB, or RGBA texture maps. +- Validation that Textured Gaussians improves 3DGS on a wide variety of both object-level and scene-level datasets. +- Demonstration that our method greatly outperforms 3DGS when the number of Gaussians is small and achieves better performance than 3DGS with alpha-only textures given the same model size. + +# 2. Related Work + +Novel-View Synthesis (NVS) and Neural Rendering. NVS tackles the problem of generating accurate renderings of 3D scenes from unseen viewpoints given a set of training images and camera poses. NVS traces back to the Structure from Motion (SfM) works [5, 19, 52] and Multiview Stereo (MVS) [15, 24, 48], which served as the basis of several groundbreaking NVS methods [8, 13, 20, 31]. However, these approaches generally require storing hundreds and thousands of input images for reprojection and blending, which leads to massive memory requirements and sometimes fails to reconstruct undersampled regions. + +![](images/f1d5160b7efc320783632e29e6915bbbb05f9d2959f487ea764d32f2e8fe4995.jpg) +Figure 2. Variants of our Textured Gaussians model. Textured Gaussians encapsulate four kinds of color and opacity spatial variations. The top row of the figure shows the texture map associated with each Gaussian, and the bottom row shows the rendered Textured Gaussians. The constant-color and constant-alpha model (no textures) corresponds to the original 3DGS formation, which can only represent a single color up to a Gaussian falloff factor within the Gaussian extent. Textured Gaussians can already model spatially varying colors using only alpha textures since each pixel can be alpha-composed differently. The model achieves maximum expressivity when leveraging the full RGBA texture map, where each Gaussian is capable of representing complex shapes and high frequency textures. + +Recent advances in neural rendering [58] have made great strides in improving the quality of 3D reconstruction and novel-view rendering. Neural rendering algorithms can be roughly classified by the underlying 3D representation, ranging from point clouds [1, 31, 61], voxels [21, 47, 67], meshes [12, 18, 22, 28, 35], to implicit representations using MLPs [2, 37, 39, 51]. Recently, 3D Gaussian Splatting [29] (3DGS) has emerged as the state-of-the-art NVS technique. 3DGS achieves high image quality for NVS, is fast to optimize and render, and leverages an explicit Gaussian primitive representation, which makes the method suitable for a wide variety of tasks such as surface extraction [16, 23], human avatar manipulation [43, 49], ambient dynamic modeling [50], and object and scene editing [17, 65]. + +Our method augments 3DGS with alpha and texture mapping commonly used in mesh-based 3D representations. During optimization and rendering, we intersect outgoing rays from pixels with Gaussians in the scene. Intersection points are then used to query per-Gaussian texture map values via UV mapping and compute RGB color and alpha blending values. This enables each Gaussian to represent complex textures and shapes, significantly improving NVS quality. + +Appearance Modeling in 3D Gaussian Splatting. In 3DGS, scene geometry is represented by the position, rotation, and scale of 3D Gaussians. By adjusting these properties, 3DGS can represent intricate geometric structures. On the other hand, appearance is modeled using per-Gaussian + +![](images/55c62f64a1e69227bf0181e2748d1e69b7207527fee16f59158920d44f4e4362.jpg) +Figure 3. Textured Gaussians model pipeline. Our method consists of three major components: ray-Gaussian intersection, RGBA texture mapping, and a generalized Gaussian appearance model. To render the color of a pixel $\mathbf{p}$ , we first trace a ray from the camera center $\mathbf{o}$ to the pixel to intersect with 3D Gaussians in the scene. Then, we query texture and alpha values, $\mathbf{c}_{\mathrm{tex}}$ and $\alpha_{\mathrm{tex}}$ , from the per-Gaussian RGBA texture maps using the ray-Gaussian intersection point $\mathbf{x}$ . Finally, given the retrieved spatially varying color and alpha values, we alpha-composite the color and alpha values of Gaussians that are hit by the pixel ray using the generalized Gaussian appearance model. + +![](images/9650985bd2570c73fe6d8441c3e9970e8f8a940cf553655f1bdbe6d4d75fa434.jpg) + +opacity values and spherical harmonics coefficients. However, pixels within a projected Gaussian are always shaded with the same color up to a Gaussian falloff factor, greatly limiting the expressivity of individual Gaussians. Therefore, recent works have explored ways that allow Gaussians to represent spatially varying features. Huang and Gong [25] defined per-Gaussian SH coefficients for color and opacity and used ray-Gaussian intersections to determine the viewing directions within the Gaussian extent. This enables smooth variation of colors and opacities across pixels covered by a single Gaussian but prevents the reconstruction of higher-frequency details. Xu et al. [63] disentangled appearance and geometry in 3DGS for texture editing, but their method is limited to object-centric scenes with simple geometry due to the unit-sphere parameterization of textures. + +Instead of using a global texture map, we enhance each 3D Gaussian with a local texture map on top of the SH coefficients. This allows for a more flexible texture optimization since individual Gaussians are not tied to a shared global texture map. Our method can, therefore, reconstruct objects with complex structures and real-world scenes. Furthermore, each Gaussian can also represent various shapes with the alpha channel in the texture map. + +Memory Efficient Gaussian Splatting. In order to achieve even faster rendering time and compact storage requirements for potential applications on edge devices, there has been a plethora of recent work focusing on optimizing memory-efficient Gaussian Splatting models [3]. These algorithms either prune Gaussians and perform quantization + +of Gaussian attributes [14, 33, 38, 40-42] or exploit the structural relationships between Gaussians [11, 36]. + +Our work is orthogonal and complementary to these works since we focus on improving the appearance modeling of Gaussians by redistributing the model size budget to the per-Gaussian texture maps given any optimized 3DGS model. These model compression algorithms can be easily integrated with the 3DGS pretraining stage in our optimization pipeline to achieve further gains in compactness. + +Concurrent Work. Different from several concurrent works that have explored the idea of augmenting Gaussians with texture maps [45, 53, 56, 62], we 1) additionally utilize alpha textures and 2) provide an in-depth analysis of various texture map variants (alpha-only, RGB, RGBA) that explores the balance between reconstruction capability and model size. + +# 3. Method + +# 3.1. 3D Gaussian Splitting Model + +In 3DGS [29], 3D scenes are represented with 3D Gaussians, and images are rendered using differentiable volume splatting. Specifically, 3DGS explicitly defines 3D Gaussians by their 3D covariance matrix $\pmb{\Sigma}_i\in \mathbb{R}^{3\times 3}$ and center $\pmb{\mu}_i\in \mathbb{R}^3$ (the index $i$ indicating the $i^{\mathrm{th}}$ Gaussian), where the 3D Gaussian function value at point $\mathbf{x}\in \mathbb{R}^3$ is defined by: + +$$ +\mathcal {G} _ {i} (\mathbf {x}) = \exp \left(- \frac {1}{2} \left(\mathbf {x} - \boldsymbol {\mu} _ {i}\right) \boldsymbol {\Sigma} _ {i} ^ {- 1} \left(\mathbf {x} - \boldsymbol {\mu} _ {i}\right)\right) \tag {1} +$$ + +where the covariance matrix $\boldsymbol{\Sigma} = \mathbf{R}\mathbf{S}\mathbf{S}^{\top}\mathbf{R}^{\top}$ is factorized into the rotation matrix $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ and the scale matrix $\mathbf{S} \in \mathbb{R}^{3 \times 3}$ . To render a 2D image from a 3D Gaussians representation, 3D Gaussians are transformed from world coordinates to camera coordinates via a world-to-camera transformation matrix $\mathbf{W} \in \mathbb{R}^{3 \times 3}$ and projected to the 2D image plane via a local affine transformation $\mathbf{J} \in \mathbb{R}^{3 \times 3}$ . The transformed 3D covariance $\boldsymbol{\Sigma}'$ can be calculated as: + +$$ +\boldsymbol {\Sigma} ^ {\prime} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} \mathbf {W} ^ {\top} \mathbf {J} ^ {\top} \tag {2} +$$ + +The covariance $\Sigma_{2\mathrm{D}}$ of the 2D Gaussian $\mathcal{G}^{2\mathrm{D}}$ splatted on the image plane can be approximated as extracting the first two rows and columns of the transformed 3D covariance, $\Sigma_{2\mathrm{D}} = [\Sigma^{\prime}]_{\{1,2\},\{1,2\}} \in \mathbb{R}^{2 \times 2}$ using the matrix minor notation. + +To render the color of a pixel $\mathbf{p} \in \mathbb{R}^3$ , the colors associated with each Gaussian are alpha-composed from front to back following the conventional volume rendering equation: + +$$ +\mathbf {c} (\mathbf {p}) = \sum_ {i = 1} ^ {K} \mathbf {c} _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \tag {3} +$$ + +where $i$ is the index of the Gaussians, $\mathbf{c}_i$ is the color of each Gaussian computed from the per-Gaussian spherical harmonic coefficients and viewing direction, and $\alpha_{i}$ is the alpha value computed from the opacity $o_i$ associated with each Gaussian and the 2D Gaussian value evaluated at pixel location $\mathbf{p}$ : + +$$ +\alpha_ {i} = \mathcal {G} _ {i} ^ {\mathrm {2 D}} (\mathbf {p}) \cdot o _ {i} \tag {4} +$$ + +The attributes (center, rotation, scale, opacity, and spherical harmonic coefficients) of the Gaussians are optimized with gradient descent using photometric losses on the rendered 2D images. + +# 3.2. Textured Gaussians + +From the 3DGS appearance model defined in Eq. 3, we observe two properties of Gaussian Splitting: + +1. Pixels covered by the same Gaussian are shaded with the same color up to a Gaussian falloff scaling factor. +2. The per-Gaussian opacity only allows Gaussians to represent ellipsoidal shapes. + +These two properties greatly restrict the expressivity of individual 3D Gaussian primitives. To allow each Gaussian primitive to represent complex appearances and shapes, we assign a fixed-resolution 2D texture map of size $\mathcal{T} \times \mathcal{T} \times \mathcal{K}, \mathcal{T} \in \mathbb{N}, \mathcal{K} \in \{1,3,4\}$ to each Gaussian and transform each Gaussian into a Textured Gaussian. As shown in Figure 2, $\mathcal{K} = 1,3,4$ correspond to alpha, RGB, and RGBA texture maps, respectively. The shape of a Textured Gaussian is thus determined by the spatially varying opacity defined by the product of the per-Gaussian opacity and the alpha channel of the texture map. The appearance of a Textured Gaussian is represented by 1) the RGB channels of + +the spatially varying textured map and 2) a set of spatially constant spherical harmonic coefficients (SH). Intuitively, the spatially constant SH coefficients represent the low-frequency textures plus view-dependent colors like specular effects, while the spatially varying texture map represents the higher-frequency spatial variations of the texture. The texture is then mapped to the plane $\mathcal{P}$ defined by the two major axes of the $i^{\mathrm{th}}$ 3D Gaussian that is centered at $\pmb{\mu}_i\in \mathbb{R}^3$ , as shown in Figure 3. The normal $\mathbf{n}_i\in \mathbb{R}^3$ of this plane thus corresponds to the axis with the smallest scale. + +For each pixel $\mathbf{p} \in \mathbb{R}^3$ that is currently being rendered, we cast a ray from the camera origin $\mathbf{o} \in \mathbb{R}^3$ to the center of the pixel $\mathbf{p}$ and intersect it with the plane $\mathcal{P}$ . The intersection point $\mathbf{x} \in \mathbb{R}^3$ can be calculated as: + +$$ +\mathbf {x} = \mathbf {o} + \frac {\left(\boldsymbol {\mu} _ {i} - \mathbf {o}\right) \cdot \mathbf {n} _ {i}}{\left(\mathbf {p} - \mathbf {o}\right) \cdot \mathbf {n} _ {i}} \cdot \frac {\mathbf {p} - \mathbf {o}}{\| \mathbf {p} - \mathbf {o} \|} \tag {5} +$$ + +Given the intersection point $\mathbf{x}$ with $\mathcal{P}$ , we perform UV mapping and bilinear interpolation on the texture map associated with the $i^{\mathrm{th}}$ Gaussian to query the texture color and alpha value at the UV coordinates $u, v \in \mathbb{R}$ , denoted by $\mathbf{c}_i^{\mathrm{tex}}(u, v) \in \mathbb{R}^3$ and $\alpha_i^{\mathrm{tex}}(u, v) \in \mathbb{R}$ . Specifically, the UV coordinates $(u, v)$ of the intersection point on the texture map can be calculated as: + +$$ +\begin{array}{l} u = \frac {m \cdot \sigma_ {i} ^ {1} + (\mathbf {x} - \boldsymbol {\mu} _ {i}) \cdot \mathbf {r} _ {i} ^ {1}}{2 \cdot m \cdot \sigma_ {i} ^ {1}} \cdot (\mathcal {T} - 1) \\ v = \frac {m \cdot \sigma_ {i} ^ {2} + (\mathbf {x} - \boldsymbol {\mu} _ {i}) \cdot \mathbf {r} _ {i} ^ {2}}{2 \cdot m \cdot \sigma_ {i} ^ {2}} \cdot (\mathcal {T} - 1) \tag {6} \\ \end{array} +$$ + +where $\sigma_i^1, \sigma_i^2 \in \mathbb{R}$ are the scales of the two major axis of the $i^{\text{th}}$ Gaussian, $\mathbf{r}_i^1, \mathbf{r}_i^2 \in \mathbb{R}^3$ are the normalized directions of the two major axes of the $i^{\text{th}}$ Gaussian, and $m \in \mathbb{R}$ is a scalar multiplier that determines the extent of the texture map with respect to each Gaussian. + +Combined with the color computed from the SH coefficients, which we call $\mathbf{c}_i^{\mathrm{base}}$ (the same as the color component in the original 3DGS appearance model in Eq. 3), the final color contribution of the $i$ -th Gaussian to pixel $\mathbf{p}$ is defined by: + +$$ +\mathbf {c} _ {i} (\mathbf {p}) = \mathbf {c} _ {i} ^ {\text {b a s e}} + \mathbf {c} _ {i} ^ {\text {t e x}} (u, v) \tag {7} +$$ + +and the alpha value of the $i$ -th Gaussian at pixel $\mathbf{p}$ is defined by: + +$$ +\alpha_ {i} (\mathbf {p}) = \alpha_ {i} ^ {\mathrm {t e x}} (u, v) \cdot \mathcal {G} _ {i} (\mathbf {x}) \cdot o _ {i} \tag {8} +$$ + +Finally, to render the color of a pixel $\mathbf{p}$ , we modify the 3DGS appearance model in Eq. 3 to incorporate the spatially varying texture and opacity: + +$$ +\mathbf {c} ^ {\text {f i n a l}} (\mathbf {p}) = \sum_ {i = 1} ^ {K} \mathbf {c} _ {i} (\mathbf {p}) \alpha_ {i} (\mathbf {p}) \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j} (\mathbf {p})) \tag {9} +$$ + +
MethodBlender [37]Mip-NeRF 360 [37]DTU [27]Tanks and Temples [30]Deep Blending [20]
Mip-NeRF 360 [4]30.34 / 0.9510 / 0.060027.69 / 0.7920 / 0.237034.28 / 0.9641 / 0.069822.22 / 0.7590 / 0.257029.40 / 0.9010 / 0.2450
Instant-NGP [39]32.20 / 0.9590 / 0.055025.30 / 0.6710 / 0.371023.58 / 0.7653 / 0.260321.72 / 0.7230 / 0.330023.62 / 0.7970 / 0.4230
Xu et al. [63]28.97 / 0.9380 / 0.0550— / — / —30.03 / — / 0.1440— / — / —— / — / —
3DGS*33.08 / 0.9671 / 0.044027.26 / 0.8318 / 0.187133.54 / 0.9697 / 0.055124.18 / 0.8541 / 0.175428.04 / 0.8940 / 0.2707
Ours33.24 / 0.9674 / 0.042827.35 / 0.8274 / 0.185833.61 / 0.9699 / 0.055624.26 / 0.8542 / 0.168428.33 / 0.8908 / 0.2699
3DGS* (10%)31.47 / 0.9590 / 0.059425.77 / 0.7796 / 0.286032.71 / 0.9627 / 0.081122.82 / 0.8020 / 0.272827.64 / 0.8853 / 0.3101
Ours (10%)32.14 / 0.9629 / 0.048926.32 / 0.7976 / 0.232332.74 / 0.9661 / 0.058223.41 / 0.8259 / 0.212227.98 / 0.8898 / 0.2804
3DGS* (1%)26.89 / 0.9160 / 0.116522.37 / 0.6293 / 0.477430.88 / 0.9320 / 0.158119.90 / 0.6736 / 0.440623.97 / 0.8167 / 0.4337
Ours (1%)28.11 / 0.9343 / 0.084923.73 / 0.7064 / 0.336532.43 / 0.9627 / 0.069421.10 / 0.7399 / 0.310424.83 / 0.8454 / 0.3552
+ +Table 1. Quantitative comparisons of different novel view synthesis methods. We boldface the best-performing model and underline the second best in terms of different metrics (PSNR ↑ / SSIM ↑ / LPIPS ↓). Our Textured Gaussians model achieves better performance than 3DGS* across most metrics. With fewer Gaussians (1% and 10% of the default optimized number of Gaussians), our method significantly outperforms 3DGS*, achieving strictly better performance across all metrics as indicated by the bold text. + +![](images/38c82b54d0e979c5ebc79a792ff251f398fea07d938b8c01c4f075f314bc6c3a.jpg) +Figure 4. NVS performance using varying numbers of Gaussians. When using the same number of Gaussians as $3\mathrm{DGS}^*$ , our Textured Gaussians achieve better novel view synthesis results on all five benchmark datasets in PSNR. Here, we show the quantitative performance of $3\mathrm{DGS}^*$ and Textured Gaussians models optimized with varying number of Gaussians on the three scene-level datasets. + +Eq. 9 is a generalized formulation of 3D Gaussian appearance and encapsulates different variants of Textured Gaussians. For example, $\mathbf{c}_i^{\mathrm{tex}} = 0$ and $\alpha_{i}^{\mathrm{tex}} = 1$ correspond to the original 3DGS model. + +# 3.3. Optimization of Textured Gaussians + +Following 3DGS [29], we optimize our model to minimize the weighted photometric loss: + +$$ +\mathcal {L} = \lambda \mathcal {L} _ {1} + (1 - \lambda) \mathcal {L} _ {\text {S S I M}} \tag {10} +$$ + +where $\lambda = 0.8$ + +Our optimization procedure consists of two stages. We first optimize a 3DGS model for 30000 iterations. The learning rates and adaptive density control (ADC) parameters are reported in the Appendix. In the second stage, we initialize all attributes of the Gaussians with the optimized vanilla 3DGS model, and jointly optimize them with the per-Gaussian 2D texture maps for another 30000 iterations. We disable the ADC control in the second stage to + +control the number of Gaussians for fair baseline comparisons. This two-stage optimization process greatly speeds up convergence and improves image quality, since jointly optimizing all parameters is a highly ill-posed problem. + +We implement custom CUDA kernels to perform fast ray-Gaussian intersection, UV mapping, and color composition. All experiments are conducted on clusters of Nvidia H100 GPUs. Please refer to the Appendix for more details on implementation and optimization. + +# 4. Results and Analysis + +We show selected qualitative and quantitative results in this section and refer the readers to the Appendix and the project website for an extensive set of results and video renderings. Datasets and Evaluation Protocols. We evaluate the novel-view synthesis performance of our method on the 8 synthetic scenes from the Blender dataset [37], all 9 scenes from the Mip-NeRF 360 dataset [4], 5 scenes from the DTU dataset [27], 2 scenes from the Tanks and Temples dataset [30], and 2 scenes from the Deep Blending dataset [20]. Please refer to the Appendix for a detailed description of the dataset preparation process. + +For comparisons with Mip-NeRF 360 [4] and InstantNGP [39], we refer to Kerbl et al. [29] for results in scene-level datasets [4, 20, 30], and refer to NeRFBaselines [32] for results on the Blender dataset [37]. Results on the DTU dataset [27] are from our own runs of the authors' released code. Please refer to the Appendix for additional experimental details. + +Xu et al. [63] pointed out in Section 5.4 of their paper that their algorithm cannot handle scene-level datasets and objects with complex structures (i.e. Blender dataset), thus they only reported results on the DTU dataset [27]. We report those numbers and additionally show results on the Blender dataset from GStex [45] authors' own runs of the algorithm described in Xu et al. [63]. + +![](images/ba48bd2b02151e1e2d1625baffd2989bb513cc0f5308839156cdcadda85a0271.jpg) +3DGS* +Ours +3DGS* +Ours +Figure 5. Qualitative NVs results of benchmark datasets. Novel-view rendering results for both object-level and scene-level standard benchmark datasets. Given the same small number of Gaussians (on average $2.5\mathrm{k}$ and $39\mathrm{k}$ Gaussians for object-level and scene-level datasets, respectively), $3\mathrm{DGS}^*$ fails to reconstruct high-frequency textures and complex shapes while our RGBA Textured Gaussians model succeeds. Please refer to the Appendix for results of our models optimized using varying number of Gaussians. + +In addition to standard benchmark test scenes, we also capture our own set of scenes that contain artworks with highly detailed textures to help demonstrate the effectiveness of our method. + +3DGS Baseline. We build our method on top of our own implementation of a slightly modified 3DGS that closely follows the algorithm described in Gaussian Opacity Fields [68]. Specifically, we use a ray-based formulation that allows for the exact evaluation of 3D Gaussian values without approximating splatted 3D Gaussians as 2D Gaussians. In addition, we used a revised densification strategy [66, 68]. For fair comparisons with our method, we report the performance of our 3DGS implementation and label it as $3\mathrm{DGS}^*$ throughout the results section, and refer the readers to the Appendix for quantitative results reported in the original 3DGS paper [29]. + +# 4.1. Quantitative Results + +In Table 1, we show quantitative comparisons in terms of PSNR/SSIM/LPIPS between our Textured Gaussians model using full RGBA textures and 3DGS* with the same number of Gaussians, and results of other baseline methods. Our method generally achieves better results than 3DGS* and all other methods, since Texture Gaussians are more expressive than vanilla 3D Gaussians due to the added texture maps. + +In the bottom half of Table 1, we also show comparisons between our method and $3\mathrm{DGS}^*$ when using a fraction of the default number of Gaussians, denoted by the percentage in parentheses. For our models with different numbers of Gaussians, we distribute a fixed amount of texels to all Gaussians (i.e., the same memory overhead for all Textured Gaussians models compared with $3\mathrm{DGS}^*$ ). Therefore, the texture map resolution of models with fewer Gaussians will be larger. We see that our method truly outshines $3\mathrm{DGS}^*$ + +![](images/4a9dbcda26ee4bfd6777efeebca4d6cba0098a5f613aa4ad0ddd980f6211ae41.jpg) +3DGS* +Ours +Figure 6. Qualitative NVS results of custom datasets. Our RGBA Textured Gaussians model achieves sharper reconstruction compared to $3\mathrm{DGS}^*$ when using the same small number of Gaussians (on average around 100k Gaussians). + +![](images/6ec87eedc6ff5d9a16aade2190d8b6236d17d894d0931848201705348f66a69d.jpg) +3DGS* + +
Method3DGS*Alpha-onlyRGBA
PSNR#GS (M)PSNR#GS (M)PSNR#GS (M)
Blender33.080.2733.120.1933.030.21
Mip-NeRF 36027.264.427.373.127.263.5
DTU33.530.2833.450.2433.410.22
Tanks and Temples24.172.924.382.624.281.3
Deep Blending28.042.828.361.628.521.0
+ +Table 2. Same model size comparisons. We compare the PSNR performance between $3\mathrm{DGS}^*$ and our model using alpha-only and RGBA textures with the same model size. The average number of optimized Gaussians (in millions) from each dataset is shown in the #GS (M) columns. We boldface and underline the best and second-best performing models, respectively. Alpha-only texture models generally achieve better performance than $3\mathrm{DGS}^*$ and RGBA texture models with the same model size. + +with fewer Gaussians, achieving nearly 2 dB improvements over $3\mathrm{DGS}^*$ when using $1\%$ of the default number of Gaussians. Figure 4 shows the trend of the quality of novel view synthesis as we vary the number of Gaussians. Our method consistently outperforms $3\mathrm{DGS}^*$ when using different numbers of Gaussians. + +Table 2 compares the performance of $3\mathrm{DGS}^*$ and our Textured Gaussians model with the same model size using alpha-only textures and RGBA textures. Since our models use per-Gaussian texture maps, they require fewer Gaussians than $3\mathrm{DGS}^*$ to achieve the same model size. Our alpha-only textures model generally outperforms the $3\mathrm{DGS}^*$ and RGBA textures models, indicating that there is a sweet spot for distributing the model size budget between Gaussian parameters and texture map channels. This is further explored in Section 4.3. + +# 4.2. Qualitative Results + +Novel-View Synthesis. We show novel view synthesis results of our model with RGBA textures and 3DGS* on + +both standard benchmark and custom-captured datasets in Figures 5 and 6. Our method reconstructs much sharper details of the scene compared to $3\mathrm{DGS}^*$ when using the same small number of Gaussians. Please refer to the Appendix for results of our models optimized using varying number of Gaussians. + +Color Component Decomposition. We show the alpha-modulated and composited results of the two color components, $\mathbf{c}^{\mathrm{tex}}$ and $\mathbf{c}^{\mathrm{base}}$ , of the optimized alpha-only, RGB and RGBA textured Gaussians model with $1\%$ of the default number of Gaussians in Figure 7. A 3DGS* model with the same number of Gaussians is shown for comparison. For RGB and RGBA models, $\mathbf{c}^{\mathrm{tex}}$ reconstructs fine-grained textures that 3DGS* cannot. For alpha-only and RGBA models, $\mathbf{c}^{\mathrm{base}}$ also reconstructs high-frequency details due to spatially varying alpha composition enabled by alpha textures. + +# 4.3. Ablation Studies + +Texture Map Variants. We ablate variants of our model that use different texture maps, namely alpha-only, RGB, and RGBA texture maps, and the same number of Gaussians in Figure 8. Experiments are conducted on Blender, Tanks and Temples, and Deep Blending datasets with the default optimized number (top row) and $1\%$ of the default optimized number (bottom row) of Gaussians. We see that our model achieves the best performance with RGBA textures. Interestingly, using alpha-only textures already outperforms $3\mathrm{DGS}^*$ , and our RGB textures model despite being one-third the size, striking a better balance between performance and model size. This is because alpha-textured Gaussians can represent complex shapes through spatially varying opacity and reconstruct high-frequency textures through spatially varying alpha composition. In contrast, RGB-textured Gaussians are still limited to representing ellipsoids. + +![](images/763145371f3925836e11dc43f411c2dbd20598dcf106497553929c47299b2112.jpg) +Figure 7. Color component decomposition visualization. Visualization of alpha-modulated and composited $\mathbf{c}^{\mathrm{base}}$ and $\mathbf{c}^{\mathrm{tex}}$ color components in our Textured Gaussians model (Eq. 7, 8, 9). Insets with yellow borders show zoom-in crops. Texture color brightness is adjusted for visualization. While $3\mathrm{DGS}^{*}$ fails to capture fine details with limited Gaussians, our method reconstructs sharp textures. With RGB textures, base color reconstructs low-frequency textures and texture maps reconstruct high-frequency elements. With alpha-only and RGBA textures, base color also reconstructs high-frequency textures via spatially varying alpha compositing. + +![](images/716d6fa039cc6e3e470c1790d0cc6a9884050cffa4584931db959d8033984c10.jpg) +Figure 8. Ablation study on texture map variants. We show the PSNR performance improvements ( $\Delta$ PSNR) over $3\mathrm{DGS}^*$ of our Textured Gaussians model when using different texture map variants. Models with full RGBA textures achieve the best results. Interestingly, alpha-only textures outperform RGB textures and perform comparably to RGBA textures while using one-third and one-fourth of the model size, respectively. + +Texture Map Resolution and the Number of Gaussians. In Figure 9, we optimize alpha-only Textured Gaussians models of the same model size with different texture map resolutions and number of Gaussians, and show the trend of novel view synthesis performance on the Blender and DTU datasets. We observe a sweet spot between texture map resolution and number of Gaussians that achieves the best image reconstruction quality. Maximizing texture map resolution corresponds to greatly reducing the number of Gaussians, which harms reconstruction of detailed geometric structures. Minimizing the texture map resolution, on the other hand, simply degenerates into 3DGS*. + +# 5. Discussions + +Limitations. The use of 2D diffuse texture maps in our model assumes that all textures lie on a local surface and does not model spatially varying specular color. Thus, ex + +![](images/bab8bbfb851f74a0c0231a6d10545499e33298ab511d5602dc2d1260bd4be94f.jpg) +Figure 9. Ablation study on texture map resolution and the number of Gaussians given the same model size. We optimized various alpha-only Textured Gaussians models (red dots) with equivalent model sizes but different texture map resolutions and numbers of Gaussian. A fitted quadratic curve (blue dashed line) illustrates the performance trend. The optimal model variant balances these two parameters to achieve superior novel view synthesis results. + +tending our texture representation to represent local 3D volume textures or even 5D radiance fields is crucial. Using factorized representations such as TensoRF [9] or triplane [7] to represent these high-dimensional textures could also be an interesting research problem. Finally, extending Textured Gaussians to support dynamic scene reconstruction [50, 64] through time-varying texture maps could be a natural next-step. + +Conclusions. In this paper, we augment 3DGS with texture maps to allow individual Gaussians to model spatially varying colors and opacity. As such, each Gaussian can represent a much richer set of appearances and shapes. This greatly improves the expressivity of individual Gaussians, leading to better novel-view synthesis quality. Our method achieves better quality than 3DGS when using the same number of Gaussians, and achieves better or comparable quality when using the same model size. + +# Acknowledgment + +We thank artist Kija Lucas for allowing us to use her art installation at the Palo Alto Art Center as one of our custom captured scenes (the flower gallery scene). Brian Chao is supported by the Stanford Graduate Fellowship and the NSF GRFP. + +# References + +[1] K. Aliev, A. Sevastopolsky, M. Kolos, D. Ulyanov, and V. Lempitsky. Neural Point-Based Graphics, 2019. arXiv preprint. 2 +[2] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In CVPR, 2022. 2 +[3] Milena T. Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3DGS.zip: A survey on 3D Gaussian Splitting Compression Methods, 2024. arXiv preprint. 3 +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In CVPR, 2022. 5, 2 +[5] Matthew Brown and David G Lowe. Unsupervised 3D Object Recognition and Reconstruction in Unordered Datasets. In 3DIM, 2005. 2 +[6] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising Densification in Gaussian Splatting. In ECCV, 2024. 1 +[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient Geometry-Aware 3D Generative Adversarial Networks. In CVPR, 2022. 8 +[8] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth Synthesis and Local Warps for Plausible Image-Based Navigation. ACM TOG, 32 (3), 2013. 2 +[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In ECCV, 2022. 8 +[10] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splitting. In CVPR, 2024. 1 +[11] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: Hash-Grid Assisted Context for 3D Gaussian Splitting Compression. In ECCV, 2024. 3 +[12] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In ECCV, 2022. 2 +[13] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating Textures. In CGF, 2008. 2 + +[14] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and $200+$ FPS. In NeurIPS, 2024. 3 +[15] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-View Stereo for Community Photo Collections. In ICCV, 2007. 2 +[16] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In CVPR, 2024. 1, 2, 4 +[17] Antoine Guédon and Vincent Lepetit. Gaussian Frosting: Editable Complex Radiance Fields with Real-Time Rendering. In ECCV, 2024. 1, 2 +[18] Yuan-Chen Guo, Yan-Pei Cao, Chen Wang, Yu He, Ying Shan, and Song-Hai Zhang. VMesh: Hybrid Volume-Mesh Representation for Efficient View Synthesis. In SIGGRAPH Asia, 2023. 2 +[19] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, second edition, 2004. 2 +[20] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. ACM TOG, 37(6), 2018. 2, 5, 9 +[21] D. Hu, Z. Zhang, T. Hou, T. Liu, H. Fu, and M. Gong. Multiscale Representation for Real-Time Anti-Aliasing Neural Rendering. In ICCV, 2023. 2 +[22] Ronghang Hu, Nikhila Ravi, Alexander C. Berg, and Deepak Pathak. Worldsheet: Wrapping the World in a 3D Sheet for View Synthesis from a Single Image. In ICCV, 2021. 2 +[23] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In SIGGRAPH, 2024. 1, 2 +[24] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In CVPR, 2018. 2 +[25] Zhentao Huang and Minglun Gong. Textured-GS: Gaussian Splatting with Spatially Defined Color and Opacity, 2024. arXiv preprint. 1, 3 +[26] Junha Hyung, Susung Hong, Sungwon Hwang, Jaeseong Lee, Jaegul Choo, and Jin-Hwa Kim. Effective rank analysis and regularization for enhanced 3d gaussian splatting. In NeuRIPS, 2024. 4, 5 +[27] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large Scale Multi-View Stereopsis Evaluation. In CVPR, 2014. 5, 2, 9 +[28] H. Kato, Y. Ushiku, and T. Harada. Neural 3D Mesh Renderer. In CVPR, 2018. 2 +[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splitting for Real-Time Radiance Field Rendering. ACM TOG, 42(4), 2023. 1, 2, 3, 5, 6 +[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. ACM TOG, 36(4), 2017. 5, 2, 9 + +[31] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with PerView Optimization. CGF, 40(4), 2021. 2 +[32] Jonas Kulhanek and Torsten Sattler. NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods, 2024. arXiv preprint. 5 +[33] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3D Gaussian Representation for Radiance Field. In CVPR, 2024. 3 +[34] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In CVPR, 2023. 1 +[35] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep Appearance Models for Face Rendering. ACM TOG, 37(4), 2018. 2 +[36] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering. In CVPR, 2024. 3 +[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 2, 5, 9 +[38] Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3D Scene Representation via Self-Organizing Gaussian Grids. In ECCV, 2024. 3 +[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM TOG, 41(4), 2022. 2, 5 +[40] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. In ECCV, 2024. 3 +[41] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3D Gaussian Splitting for Accelerated Novel View Synthesis. In CVPR, pages 10349-10358, 2024. +[42] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the Memory Footprint of 3D Gaussian Splatting. ACM CGIT, 7(1), 2024. 3 +[43] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting. In CVPR, 2024. 2 +[44] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. DreamGaussian4D: Generative 4D Gaussian Splatting, 2023. arXiv preprint. 1 +[45] Victor Rong, Jingxiang Chen, Sherwin Bahmani, Kiriakos N Kutulakos, and David B Lindell. GStex: Per-Primitive Texturing of 2D Gaussian Splatting for Decoupled Appearance and Geometry Modeling, 2024. 3, 5 +[46] Antoni Rosinol, John J Leonard, and Luca Carlone. NeRF-SLAM: Real-time Dense Monocular Slam with Neural Radiance Fields. In IROS, 2023. 1 +[47] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields without Neural Networks. In CVPR, 2022. 2 + +[48] Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A Comparison and Evaluation of Multi-View Stereo Reconstruction Algorithms. In CVPR, 2006. 2 +[49] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In CVPR, 2024. 2 +[50] Meng-Li Shih, Jia-Bin Huang, Changil Kim, Rajvi Shah, Johannes Kopf, and Chen Gao. Modeling ambient scene dynamics for free-view synthesis. 2024. 1, 2, 8 +[51] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene Representation Networks: Continuous 3D-Structure-Aware Neural Scene Representations. In NeurIPS, 2019. 2 +[52] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. ACM TOG, 25 (3), 2006. 2 +[53] Yunzhou Song, Heguang Lin, Jiahui Lei, Lingjie Liu, and Kostas Daniilidis. HDGS: Textured 2D Gaussian Splitting for Enhanced Scene Rendering, 2024. 3 +[54] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In CVPR, 2021. 1 +[55] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Articulated Neural Radiance Fields for Learning Human Shape, Appearance, and Pose. In NeurIPS, 2021. 1 +[56] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. BillBoard Splatting (BBSplat): Learnable Textured Primitives for Novel View Synthesis, 2025. 3 +[57] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative Gaussian Splitting for Efficient 3D Content Creation. In ICLR, 2024. 1 +[58] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in Neural Rendering. CGF, 41(2), 2022. 1, 2 +[59] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction. In NeurIPS, 2021. 1 +[60] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 1 +[61] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-Nerf: Point-Based Neural Radiance Fields. In CVPR, 2022. 2 +[62] Rui Xu, Wenyue Chen, Jiepeng Wang, Yuan Liu, Peng Wang, Lin Gao, Shiqing Xin, Taku Komura, Xin Li, and Wenping Wang. SuperGaussians: Enhancing Gaussian Splitting Using Primitives with Spatially Varying Colors, 2024. 3 + +[63] Tian-Xing Xu, Wenbo Hu, Yu-Kun Lai, Ying Shan, and Song-Hai Zhang. Texture-GS: Disentangling the Geometry and Texture for 3D Gaussian Splitting Editing. In ECCV, 2024. 2, 3, 5 +[64] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. In CVPR, 2024. 1, 8 +[65] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian Grouping: Segment and Edit Anything in 3D Scenes. In ECCV, 2024. 2 +[66] Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details in 3d gaussian splatt-ing. In ACM-MM, 2024. 6 +[67] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for Real-Time Rendering of Neural Radiance Fields. In ICCV, 2021. 2 +[68] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes. ACM TOG, 2024. 6, 1, 4 +[69] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural Factorization of Shape and Reflectance under an Unknown Illumination. ACM TOG, 40(6), 2021. 1 +[70] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM. In CVPR, 2022. 1 \ No newline at end of file diff --git a/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/images.zip b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..77f7a6f0b8e8c75250877997f4017855467ed12c --- /dev/null +++ b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:050454b335cf6926d31ea57aa32014bafc5527d830d61b4533e01e09bf39a5e7 +size 947032 diff --git a/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/layout.json b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d11f11b0eb68eee4b2b5802285aefff507227b4c --- /dev/null +++ b/2025/Textured Gaussians for Enhanced 3D Scene Appearance Modeling/layout.json @@ -0,0 +1,9214 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "text", + "content": "Textured Gaussians for Enhanced 3D Scene Appearance Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "spans": [ + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": "Brian Chao " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Hung-Yu Tseng " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Lorenzo Porzi " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Chen Gao " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Tuotuo Li " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Qinbo Li " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Ayush Saraf " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Jia-Bin Huang " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Johannes Kopf " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Gordon Wetzstein " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "text", + "content": " Changil Kim " + }, + { + "bbox": [ + 71, + 141, + 538, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 180, + 480, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 180, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 127, + 180, + 480, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 180, + 480, + 198 + ], + "type": "text", + "content": " Stanford University " + }, + { + "bbox": [ + 127, + 180, + 480, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 180, + 480, + 198 + ], + "type": "text", + "content": " Meta " + }, + { + "bbox": [ + 127, + 180, + 480, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 180, + 480, + 198 + ], + "type": "text", + "content": " University of Maryland College Park" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 204, + 204, + 402, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 204, + 402, + 215 + ], + "spans": [ + { + "bbox": [ + 204, + 204, + 402, + 215 + ], + "type": "text", + "content": "https:// textured-gaussians.github.io" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 243, + 238, + 380 + ], + "blocks": [ + { + "bbox": [ + 58, + 243, + 238, + 380 + ], + "lines": [ + { + "bbox": [ + 58, + 243, + 238, + 380 + ], + "spans": [ + { + "bbox": [ + 58, + 243, + 238, + 380 + ], + "type": "image", + "image_path": "f97c4011fb92fe6ea64381f0ce03e2ed665c9f349b25425d4df2e54e6f64abaa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 390, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 54, + 390, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 555, + 424 + ], + "type": "text", + "content": "Figure 1. Textured Gaussians compared to 3D Gaussian Splatting (3DGS). Our RGBA Textured Gaussians enhance 3D scene appearance modeling, leading to improved rendering quality while using the same number of Gaussians compared to 3DGS [29]. Above, we show that Textured Gaussians faithfully reconstruct the fine details of scenes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 238, + 243, + 306, + 380 + ], + "blocks": [ + { + "bbox": [ + 238, + 243, + 306, + 380 + ], + "lines": [ + { + "bbox": [ + 238, + 243, + 306, + 380 + ], + "spans": [ + { + "bbox": [ + 238, + 243, + 306, + 380 + ], + "type": "image", + "image_path": "28a9f048694eb262ce139fb32dabfb7025470cbf0422f0a158b02f292ae58823.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 243, + 485, + 380 + ], + "blocks": [ + { + "bbox": [ + 306, + 243, + 485, + 380 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 485, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 485, + 380 + ], + "type": "image", + "image_path": "6fe6cc101f9a79075c0731c30e6678267bd0cbaffc03490749feff24a6d6b166.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 485, + 243, + 553, + 380 + ], + "blocks": [ + { + "bbox": [ + 485, + 243, + 553, + 380 + ], + "lines": [ + { + "bbox": [ + 485, + 243, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 485, + 243, + 553, + 380 + ], + "type": "image", + "image_path": "f24741a0bf4541b7cc291439ab1b34ac28215b0fa81bc84fd2fea98269557ee8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "spans": [ + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 464, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 464, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 464, + 297, + 715 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) has emerged as the state-of-the-art 3D reconstruction technique, offering high-quality results with fast training and rendering. However, its expressivity is limited as pixels covered by the same Gaussian share identical colors aside from a Gaussian falloff scaling factor, and individual Gaussians can only represent simple ellipsoids geometrically. To overcome these limitations, we integrate texture and alpha mapping from traditional graphics with 3DGS. Our approach augments each Gaussian with alpha, RGB, or RGBA texture maps to model spatially varying color and opacity across each Gaussian's extent. This allows Gaussians to represent richer texture patterns and geometric structures beyond single-color ellipsoids. Notably, alpha-only texture maps significantly improve Gaussian expressivity, while further augmenting with RGB texture maps achieve maximum expressivity. We validate our method on a wide variety of standard benchmark datasets and our own custom captures at both the object and scene levels, and demonstrate image quality improvements over existing methods while using a similar or lower number of Gaussians." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 434, + 394, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 434, + 394, + 445 + ], + "spans": [ + { + "bbox": [ + 315, + 434, + 394, + 445 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 457, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 649 + ], + "type": "text", + "content": "Neural rendering [58] achieves unprecedented novel-view synthesis and 3D reconstruction quality, offering solutions to a myriad of applications ranging from surface reconstruction [34, 59], SLAM [46, 70], material estimation and relighting [54, 69], to virtual teleportation and human avatar animation [55, 60]. Recently, 3D Gaussian Splatting (3DGS) [29] emerged as a state-of-the-art novel-view synthesis technique, with desirable properties such as high image quality results, fast training and rendering time, and an explicit primitives-based representation. Subsequent works in Gaussian Splatting have focused on topics such as improving surface reconstruction quality [16, 23], scene editing [10, 17], dynamic scene reconstruction [50, 64], and 3D generation [44, 57]. Despite the success, 3DGS sometimes fails to model detailed appearances, as demonstrated in Figure 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "In 3DGS, each Gaussian can only represent a single color and the shape of an ellipsoid given a camera viewpoint. This dramatically limits the set of appearances and shapes a single Gaussian can represent. To address the issue, Huang and Gong [25] modified the viewing direction calculation" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8964" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "in 3DGS such that pixels covered by the same Gaussian exhibit a smooth color and opacity variation. However, this color variation across a single Gaussian is minuscule due to the spherical harmonics representation of colors, preventing them from representing complex textures. Xu et al. [63] proposed only using Gaussians to represent scene geometry and leveraged a learned UV mapping module and a global 2D RGB texture map to project textures onto Gaussian surfaces. However, the unit-sphere representation of textures significantly constrains the capacity of their model, and thus, their method fails to reconstruct objects with complex geometry or large-scale scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 418 + ], + "type": "text", + "content": "Our method builds upon 3DGS and draws inspiration from mesh-based 3D representations that model appearance using texture mapping. As shown in Figure 2, we augment each Gaussian with its alpha, RGB, or RGBA texture map so that each Gaussian is capable of representing a rich set of textures and shapes. We call this representation Textured Gaussians. With RGB textures, individual Gaussians can represent higher-frequency color variations. The alpha maps allow Gaussians to represent a wider variety of shapes, instead of just ellipsoids. To achieve this, we build custom CUDA kernels that perform ray-Gaussian intersection and texture mapping and integrate them with 3DGS. Consequently, our method not only inherits all desirable properties of 3DGS, such as fast training and rendering time and explicit 3D representation, but also substantially improves rendered image quality, especially with a small number of Gaussians." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 419, + 227, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 419, + 227, + 430 + ], + "spans": [ + { + "bbox": [ + 67, + 419, + 227, + 430 + ], + "type": "text", + "content": "In summary, our contributions include:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 295, + 550 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 431, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 479 + ], + "type": "text", + "content": "- Introduction of a generalized appearance model for 3D Gaussians that handles spatially varying color and opacity by augmenting 3D Gaussians with alpha, RGB, or RGBA texture maps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "content": "- Validation that Textured Gaussians improves 3DGS on a wide variety of both object-level and scene-level datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 550 + ], + "type": "text", + "content": "- Demonstration that our method greatly outperforms 3DGS when the number of Gaussians is small and achieves better performance than 3DGS with alpha-only textures given the same model size." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 562, + 142, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 142, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 142, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Novel-View Synthesis (NVS) and Neural Rendering. NVS tackles the problem of generating accurate renderings of 3D scenes from unseen viewpoints given a set of training images and camera poses. NVS traces back to the Structure from Motion (SfM) works [5, 19, 52] and Multiview Stereo (MVS) [15, 24, 48], which served as the basis of several groundbreaking NVS methods [8, 13, 20, 31]. However, these approaches generally require storing hundreds and thousands of input images for reprojection and blending, which leads to massive memory requirements and sometimes fails to reconstruct undersampled regions." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 555, + 198 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 555, + 198 + ], + "type": "image", + "image_path": "f1d5160b7efc320783632e29e6915bbbb05f9d2959f487ea764d32f2e8fe4995.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 353 + ], + "type": "text", + "content": "Figure 2. Variants of our Textured Gaussians model. Textured Gaussians encapsulate four kinds of color and opacity spatial variations. The top row of the figure shows the texture map associated with each Gaussian, and the bottom row shows the rendered Textured Gaussians. The constant-color and constant-alpha model (no textures) corresponds to the original 3DGS formation, which can only represent a single color up to a Gaussian falloff factor within the Gaussian extent. Textured Gaussians can already model spatially varying colors using only alpha textures since each pixel can be alpha-composed differently. The model achieves maximum expressivity when leveraging the full RGBA texture map, where each Gaussian is capable of representing complex shapes and high frequency textures." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 544 + ], + "type": "text", + "content": "Recent advances in neural rendering [58] have made great strides in improving the quality of 3D reconstruction and novel-view rendering. Neural rendering algorithms can be roughly classified by the underlying 3D representation, ranging from point clouds [1, 31, 61], voxels [21, 47, 67], meshes [12, 18, 22, 28, 35], to implicit representations using MLPs [2, 37, 39, 51]. Recently, 3D Gaussian Splatting [29] (3DGS) has emerged as the state-of-the-art NVS technique. 3DGS achieves high image quality for NVS, is fast to optimize and render, and leverages an explicit Gaussian primitive representation, which makes the method suitable for a wide variety of tasks such as surface extraction [16, 23], human avatar manipulation [43, 49], ambient dynamic modeling [50], and object and scene editing [17, 65]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 652 + ], + "type": "text", + "content": "Our method augments 3DGS with alpha and texture mapping commonly used in mesh-based 3D representations. During optimization and rendering, we intersect outgoing rays from pixels with Gaussians in the scene. Intersection points are then used to query per-Gaussian texture map values via UV mapping and compute RGB color and alpha blending values. This enables each Gaussian to represent complex textures and shapes, significantly improving NVS quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Appearance Modeling in 3D Gaussian Splatting. In 3DGS, scene geometry is represented by the position, rotation, and scale of 3D Gaussians. By adjusting these properties, 3DGS can represent intricate geometric structures. On the other hand, appearance is modeled using per-Gaussian" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8965" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 97, + 409, + 264 + ], + "blocks": [ + { + "bbox": [ + 56, + 97, + 409, + 264 + ], + "lines": [ + { + "bbox": [ + 56, + 97, + 409, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 409, + 264 + ], + "type": "image", + "image_path": "55c62f64a1e69227bf0181e2748d1e69b7207527fee16f59158920d44f4e4362.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "text", + "content": "Figure 3. Textured Gaussians model pipeline. Our method consists of three major components: ray-Gaussian intersection, RGBA texture mapping, and a generalized Gaussian appearance model. To render the color of a pixel " + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "text", + "content": ", we first trace a ray from the camera center " + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "text", + "content": " to the pixel to intersect with 3D Gaussians in the scene. Then, we query texture and alpha values, " + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{tex}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{tex}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "text", + "content": ", from the per-Gaussian RGBA texture maps using the ray-Gaussian intersection point " + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 54, + 275, + 555, + 331 + ], + "type": "text", + "content": ". Finally, given the retrieved spatially varying color and alpha values, we alpha-composite the color and alpha values of Gaussians that are hit by the pixel ray using the generalized Gaussian appearance model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 410, + 93, + 548, + 262 + ], + "blocks": [ + { + "bbox": [ + 410, + 93, + 548, + 262 + ], + "lines": [ + { + "bbox": [ + 410, + 93, + 548, + 262 + ], + "spans": [ + { + "bbox": [ + 410, + 93, + 548, + 262 + ], + "type": "image", + "image_path": "9650985bd2570c73fe6d8441c3e9970e8f8a940cf553655f1bdbe6d4d75fa434.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 351, + 294, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 351, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 294, + 542 + ], + "type": "text", + "content": "opacity values and spherical harmonics coefficients. However, pixels within a projected Gaussian are always shaded with the same color up to a Gaussian falloff factor, greatly limiting the expressivity of individual Gaussians. Therefore, recent works have explored ways that allow Gaussians to represent spatially varying features. Huang and Gong [25] defined per-Gaussian SH coefficients for color and opacity and used ray-Gaussian intersections to determine the viewing directions within the Gaussian extent. This enables smooth variation of colors and opacities across pixels covered by a single Gaussian but prevents the reconstruction of higher-frequency details. Xu et al. [63] disentangled appearance and geometry in 3DGS for texture editing, but their method is limited to object-centric scenes with simple geometry due to the unit-sphere parameterization of textures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": "Instead of using a global texture map, we enhance each 3D Gaussian with a local texture map on top of the SH coefficients. This allows for a more flexible texture optimization since individual Gaussians are not tied to a shared global texture map. Our method can, therefore, reconstruct objects with complex structures and real-world scenes. Furthermore, each Gaussian can also represent various shapes with the alpha channel in the texture map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Memory Efficient Gaussian Splatting. In order to achieve even faster rendering time and compact storage requirements for potential applications on edge devices, there has been a plethora of recent work focusing on optimizing memory-efficient Gaussian Splatting models [3]. These algorithms either prune Gaussians and perform quantization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 351, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 553, + 376 + ], + "type": "text", + "content": "of Gaussian attributes [14, 33, 38, 40-42] or exploit the structural relationships between Gaussians [11, 36]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 378, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 462 + ], + "type": "text", + "content": "Our work is orthogonal and complementary to these works since we focus on improving the appearance modeling of Gaussians by redistributing the model size budget to the per-Gaussian texture maps given any optimized 3DGS model. These model compression algorithms can be easily integrated with the 3DGS pretraining stage in our optimization pipeline to achieve further gains in compactness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 464, + 554, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 547 + ], + "type": "text", + "content": "Concurrent Work. Different from several concurrent works that have explored the idea of augmenting Gaussians with texture maps [45, 53, 56, 62], we 1) additionally utilize alpha textures and 2) provide an in-depth analysis of various texture map variants (alpha-only, RGB, RGBA) that explores the balance between reconstruction capability and model size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 566, + 370, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 370, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 370, + 578 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 588, + 474, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 474, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 474, + 601 + ], + "type": "text", + "content": "3.1. 3D Gaussian Splitting Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "text", + "content": "In 3DGS [29], 3D scenes are represented with 3D Gaussians, and images are rendered using differentiable volume splatting. Specifically, 3DGS explicitly defines 3D Gaussians by their 3D covariance matrix " + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "text", + "content": " and center " + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "text", + "content": " (the index " + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "text", + "content": " indicating the " + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "text", + "content": " Gaussian), where the 3D Gaussian function value at point " + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 608, + 554, + 681 + ], + "type": "text", + "content": " is defined by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 693, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 693, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 349, + 693, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {i} (\\mathbf {x}) = \\exp \\left(- \\frac {1}{2} \\left(\\mathbf {x} - \\boldsymbol {\\mu} _ {i}\\right) \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\left(\\mathbf {x} - \\boldsymbol {\\mu} _ {i}\\right)\\right) \\tag {1}", + "image_path": "3dca4ac94ab6213a092b0e80fc10e9a15c43ab5c90e915421edbe9407e7e9a02.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8966" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "where the covariance matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} = \\mathbf{R}\\mathbf{S}\\mathbf{S}^{\\top}\\mathbf{R}^{\\top}" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " is factorized into the rotation matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " and the scale matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": ". To render a 2D image from a 3D Gaussians representation, 3D Gaussians are transformed from world coordinates to camera coordinates via a world-to-camera transformation matrix " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " and projected to the 2D image plane via a local affine transformation " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{J} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": ". The transformed 3D covariance " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}'" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " can be calculated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 176, + 295, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 176, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 176, + 295, + 191 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {\\top} \\mathbf {J} ^ {\\top} \\tag {2}", + "image_path": "ea9b22b0d4ed1c88a0253c9977cccdd254006903e0ae13aca82e56259c29dc26.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 200, + 296, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 296, + 249 + ], + "type": "text", + "content": "The covariance " + }, + { + "bbox": [ + 55, + 200, + 296, + 249 + ], + "type": "inline_equation", + "content": "\\Sigma_{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 200, + 296, + 249 + ], + "type": "text", + "content": " of the 2D Gaussian " + }, + { + "bbox": [ + 55, + 200, + 296, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 200, + 296, + 249 + ], + "type": "text", + "content": " splatted on the image plane can be approximated as extracting the first two rows and columns of the transformed 3D covariance, " + }, + { + "bbox": [ + 55, + 200, + 296, + 249 + ], + "type": "inline_equation", + "content": "\\Sigma_{2\\mathrm{D}} = [\\Sigma^{\\prime}]_{\\{1,2\\},\\{1,2\\}} \\in \\mathbb{R}^{2 \\times 2}" + }, + { + "bbox": [ + 55, + 200, + 296, + 249 + ], + "type": "text", + "content": " using the matrix minor notation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 249, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 294 + ], + "type": "text", + "content": "To render the color of a pixel " + }, + { + "bbox": [ + 55, + 249, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 249, + 296, + 294 + ], + "type": "text", + "content": ", the colors associated with each Gaussian are alpha-composed from front to back following the conventional volume rendering equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 294, + 295, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 294, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 115, + 294, + 295, + 327 + ], + "type": "interline_equation", + "content": "\\mathbf {c} (\\mathbf {p}) = \\sum_ {i = 1} ^ {K} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {3}", + "image_path": "e3bc8285a6818d7ef53692a4211101dcaa1e3b2b5437cccdae76da91ff5396e6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "text", + "content": " is the index of the Gaussians, " + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "text", + "content": " is the color of each Gaussian computed from the per-Gaussian spherical harmonic coefficients and viewing direction, and " + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "text", + "content": " is the alpha value computed from the opacity " + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "text", + "content": " associated with each Gaussian and the 2D Gaussian value evaluated at pixel location " + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 333, + 296, + 403 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 404, + 294, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 404, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 140, + 404, + 294, + 419 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\mathcal {G} _ {i} ^ {\\mathrm {2 D}} (\\mathbf {p}) \\cdot o _ {i} \\tag {4}", + "image_path": "48f990281b43dd59b1c21c292952a6b6b5b7479695daf05bc8e6c6266f589870.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 424, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 472 + ], + "type": "text", + "content": "The attributes (center, rotation, scale, opacity, and spherical harmonic coefficients) of the Gaussians are optimized with gradient descent using photometric losses on the rendered 2D images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 480, + 170, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 170, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 170, + 491 + ], + "type": "text", + "content": "3.2. Textured Gaussians" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "content": "From the 3DGS appearance model defined in Eq. 3, we observe two properties of Gaussian Splitting:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 295, + 569 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 521, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 295, + 545 + ], + "type": "text", + "content": "1. Pixels covered by the same Gaussian are shaded with the same color up to a Gaussian falloff scaling factor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "text", + "content": "2. The per-Gaussian opacity only allows Gaussians to represent ellipsoidal shapes." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "These two properties greatly restrict the expressivity of individual 3D Gaussian primitives. To allow each Gaussian primitive to represent complex appearances and shapes, we assign a fixed-resolution 2D texture map of size " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\times \\mathcal{T} \\times \\mathcal{K}, \\mathcal{T} \\in \\mathbb{N}, \\mathcal{K} \\in \\{1,3,4\\}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " to each Gaussian and transform each Gaussian into a Textured Gaussian. As shown in Figure 2, " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{K} = 1,3,4" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " correspond to alpha, RGB, and RGBA texture maps, respectively. The shape of a Textured Gaussian is thus determined by the spatially varying opacity defined by the product of the per-Gaussian opacity and the alpha channel of the texture map. The appearance of a Textured Gaussian is represented by 1) the RGB channels of" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "the spatially varying textured map and 2) a set of spatially constant spherical harmonic coefficients (SH). Intuitively, the spatially constant SH coefficients represent the low-frequency textures plus view-dependent colors like specular effects, while the spatially varying texture map represents the higher-frequency spatial variations of the texture. The texture is then mapped to the plane " + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": " defined by the two major axes of the " + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": " 3D Gaussian that is centered at " + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": ", as shown in Figure 3. The normal " + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": " of this plane thus corresponds to the axis with the smallest scale." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "text", + "content": "For each pixel " + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "text", + "content": " that is currently being rendered, we cast a ray from the camera origin " + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "text", + "content": " to the center of the pixel " + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "text", + "content": " and intersect it with the plane " + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "text", + "content": ". The intersection point " + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 192, + 553, + 240 + ], + "type": "text", + "content": " can be calculated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 364, + 250, + 553, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 250, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 364, + 250, + 553, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = \\mathbf {o} + \\frac {\\left(\\boldsymbol {\\mu} _ {i} - \\mathbf {o}\\right) \\cdot \\mathbf {n} _ {i}}{\\left(\\mathbf {p} - \\mathbf {o}\\right) \\cdot \\mathbf {n} _ {i}} \\cdot \\frac {\\mathbf {p} - \\mathbf {o}}{\\| \\mathbf {p} - \\mathbf {o} \\|} \\tag {5}", + "image_path": "f007d132d6f3cd1d859e5904c7a943eee01a2531c6c27fb7fb320d4265d9f24f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "text", + "content": "Given the intersection point " + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "text", + "content": ", we perform UV mapping and bilinear interpolation on the texture map associated with the " + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "text", + "content": " Gaussian to query the texture color and alpha value at the UV coordinates " + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "inline_equation", + "content": "u, v \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i^{\\mathrm{tex}}(u, v) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "inline_equation", + "content": "\\alpha_i^{\\mathrm{tex}}(u, v) \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "text", + "content": ". Specifically, the UV coordinates " + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 313, + 283, + 555, + 368 + ], + "type": "text", + "content": " of the intersection point on the texture map can be calculated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 355, + 387, + 553, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 387, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 355, + 387, + 553, + 442 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} u = \\frac {m \\cdot \\sigma_ {i} ^ {1} + (\\mathbf {x} - \\boldsymbol {\\mu} _ {i}) \\cdot \\mathbf {r} _ {i} ^ {1}}{2 \\cdot m \\cdot \\sigma_ {i} ^ {1}} \\cdot (\\mathcal {T} - 1) \\\\ v = \\frac {m \\cdot \\sigma_ {i} ^ {2} + (\\mathbf {x} - \\boldsymbol {\\mu} _ {i}) \\cdot \\mathbf {r} _ {i} ^ {2}}{2 \\cdot m \\cdot \\sigma_ {i} ^ {2}} \\cdot (\\mathcal {T} - 1) \\tag {6} \\\\ \\end{array}", + "image_path": "6ebe190d5a2c90c1c3048abd39e671640713102d0d981b704face8c7c68db8b5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "inline_equation", + "content": "\\sigma_i^1, \\sigma_i^2 \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "text", + "content": " are the scales of the two major axis of the " + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "text", + "content": " Gaussian, " + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i^1, \\mathbf{r}_i^2 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "text", + "content": " are the normalized directions of the two major axes of the " + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "text", + "content": " Gaussian, and " + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "inline_equation", + "content": "m \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 452, + 554, + 512 + ], + "type": "text", + "content": " is a scalar multiplier that determines the extent of the texture map with respect to each Gaussian." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 512, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 554, + 571 + ], + "type": "text", + "content": "Combined with the color computed from the SH coefficients, which we call " + }, + { + "bbox": [ + 313, + 512, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i^{\\mathrm{base}}" + }, + { + "bbox": [ + 313, + 512, + 554, + 571 + ], + "type": "text", + "content": " (the same as the color component in the original 3DGS appearance model in Eq. 3), the final color contribution of the " + }, + { + "bbox": [ + 313, + 512, + 554, + 571 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 512, + 554, + 571 + ], + "type": "text", + "content": "-th Gaussian to pixel " + }, + { + "bbox": [ + 313, + 512, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 512, + 554, + 571 + ], + "type": "text", + "content": " is defined by:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 380, + 572, + 553, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 572, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 380, + 572, + 553, + 587 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {i} (\\mathbf {p}) = \\mathbf {c} _ {i} ^ {\\text {b a s e}} + \\mathbf {c} _ {i} ^ {\\text {t e x}} (u, v) \\tag {7}", + "image_path": "d574ad6f12fc2616653bb4e947283c2e3a27b4bd76cfbeb7cce88ba10caa336e.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 592, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 553, + 616 + ], + "type": "text", + "content": "and the alpha value of the " + }, + { + "bbox": [ + 313, + 592, + 553, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 592, + 553, + 616 + ], + "type": "text", + "content": "-th Gaussian at pixel " + }, + { + "bbox": [ + 313, + 592, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 592, + 553, + 616 + ], + "type": "text", + "content": " is defined by:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 372, + 616, + 553, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 616, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 372, + 616, + 553, + 631 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} (\\mathbf {p}) = \\alpha_ {i} ^ {\\mathrm {t e x}} (u, v) \\cdot \\mathcal {G} _ {i} (\\mathbf {x}) \\cdot o _ {i} \\tag {8}", + "image_path": "2b720bb4eb4ac71870cf264dac60071cdf769342ca231fc88b9153379a2682ba.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 636, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 553, + 673 + ], + "type": "text", + "content": "Finally, to render the color of a pixel " + }, + { + "bbox": [ + 313, + 636, + 553, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 636, + 553, + 673 + ], + "type": "text", + "content": ", we modify the 3DGS appearance model in Eq. 3 to incorporate the spatially varying texture and opacity:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 345, + 681, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 681, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 345, + 681, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {c} ^ {\\text {f i n a l}} (\\mathbf {p}) = \\sum_ {i = 1} ^ {K} \\mathbf {c} _ {i} (\\mathbf {p}) \\alpha_ {i} (\\mathbf {p}) \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} (\\mathbf {p})) \\tag {9}", + "image_path": "841fab48659b415bbfc70c8da1c1c0faa69c2460cc16fe7301dced3371e4652a.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8967" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 558, + 203 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 558, + 203 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 558, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 558, + 203 + ], + "type": "table", + "html": "
MethodBlender [37]Mip-NeRF 360 [37]DTU [27]Tanks and Temples [30]Deep Blending [20]
Mip-NeRF 360 [4]30.34 / 0.9510 / 0.060027.69 / 0.7920 / 0.237034.28 / 0.9641 / 0.069822.22 / 0.7590 / 0.257029.40 / 0.9010 / 0.2450
Instant-NGP [39]32.20 / 0.9590 / 0.055025.30 / 0.6710 / 0.371023.58 / 0.7653 / 0.260321.72 / 0.7230 / 0.330023.62 / 0.7970 / 0.4230
Xu et al. [63]28.97 / 0.9380 / 0.0550— / — / —30.03 / — / 0.1440— / — / —— / — / —
3DGS*33.08 / 0.9671 / 0.044027.26 / 0.8318 / 0.187133.54 / 0.9697 / 0.055124.18 / 0.8541 / 0.175428.04 / 0.8940 / 0.2707
Ours33.24 / 0.9674 / 0.042827.35 / 0.8274 / 0.185833.61 / 0.9699 / 0.055624.26 / 0.8542 / 0.168428.33 / 0.8908 / 0.2699
3DGS* (10%)31.47 / 0.9590 / 0.059425.77 / 0.7796 / 0.286032.71 / 0.9627 / 0.081122.82 / 0.8020 / 0.272827.64 / 0.8853 / 0.3101
Ours (10%)32.14 / 0.9629 / 0.048926.32 / 0.7976 / 0.232332.74 / 0.9661 / 0.058223.41 / 0.8259 / 0.212227.98 / 0.8898 / 0.2804
3DGS* (1%)26.89 / 0.9160 / 0.116522.37 / 0.6293 / 0.477430.88 / 0.9320 / 0.158119.90 / 0.6736 / 0.440623.97 / 0.8167 / 0.4337
Ours (1%)28.11 / 0.9343 / 0.084923.73 / 0.7064 / 0.336532.43 / 0.9627 / 0.069421.10 / 0.7399 / 0.310424.83 / 0.8454 / 0.3552
", + "image_path": "dc8f83c866497aaf0ee4a31dd6b53ab508809b8e1383ffe3e413500134db33f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 210, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 210, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 555, + 257 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of different novel view synthesis methods. We boldface the best-performing model and underline the second best in terms of different metrics (PSNR ↑ / SSIM ↑ / LPIPS ↓). Our Textured Gaussians model achieves better performance than 3DGS* across most metrics. With fewer Gaussians (1% and 10% of the default optimized number of Gaussians), our method significantly outperforms 3DGS*, achieving strictly better performance across all metrics as indicated by the bold text." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 274, + 294, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 274, + 294, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 365 + ], + "type": "image", + "image_path": "38c82b54d0e979c5ebc79a792ff251f398fea07d938b8c01c4f075f314bc6c3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 376, + 296, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 452 + ], + "type": "text", + "content": "Figure 4. NVS performance using varying numbers of Gaussians. When using the same number of Gaussians as " + }, + { + "bbox": [ + 55, + 376, + 296, + 452 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 376, + 296, + 452 + ], + "type": "text", + "content": ", our Textured Gaussians achieve better novel view synthesis results on all five benchmark datasets in PSNR. Here, we show the quantitative performance of " + }, + { + "bbox": [ + 55, + 376, + 296, + 452 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 376, + 296, + 452 + ], + "type": "text", + "content": " and Textured Gaussians models optimized with varying number of Gaussians on the three scene-level datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": "Eq. 9 is a generalized formulation of 3D Gaussian appearance and encapsulates different variants of Textured Gaussians. For example, " + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i^{\\mathrm{tex}} = 0" + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{\\mathrm{tex}} = 1" + }, + { + "bbox": [ + 55, + 475, + 296, + 523 + ], + "type": "text", + "content": " correspond to the original 3DGS model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 248, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 248, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 248, + 544 + ], + "type": "text", + "content": "3.3. Optimization of Textured Gaussians" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 548, + 295, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 572 + ], + "type": "text", + "content": "Following 3DGS [29], we optimize our model to minimize the weighted photometric loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 582, + 295, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 582, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 121, + 582, + 295, + 596 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\mathcal {L} _ {1} + (1 - \\lambda) \\mathcal {L} _ {\\text {S S I M}} \\tag {10}", + "image_path": "8e432b3178da0448b2343587ec616293970dd97529786f08aff7bff92a8aa0ac.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 120, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 120, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 120, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 120, + 616 + ], + "type": "inline_equation", + "content": "\\lambda = 0.8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Our optimization procedure consists of two stages. We first optimize a 3DGS model for 30000 iterations. The learning rates and adaptive density control (ADC) parameters are reported in the Appendix. In the second stage, we initialize all attributes of the Gaussians with the optimized vanilla 3DGS model, and jointly optimize them with the per-Gaussian 2D texture maps for another 30000 iterations. We disable the ADC control in the second stage to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 276, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 323 + ], + "type": "text", + "content": "control the number of Gaussians for fair baseline comparisons. This two-stage optimization process greatly speeds up convergence and improves image quality, since jointly optimizing all parameters is a highly ill-posed problem." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 324, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 384 + ], + "type": "text", + "content": "We implement custom CUDA kernels to perform fast ray-Gaussian intersection, UV mapping, and color composition. All experiments are conducted on clusters of Nvidia H100 GPUs. Please refer to the Appendix for more details on implementation and optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 394, + 436, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 436, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 436, + 407 + ], + "type": "text", + "content": "4. Results and Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 415, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 545 + ], + "type": "text", + "content": "We show selected qualitative and quantitative results in this section and refer the readers to the Appendix and the project website for an extensive set of results and video renderings. Datasets and Evaluation Protocols. We evaluate the novel-view synthesis performance of our method on the 8 synthetic scenes from the Blender dataset [37], all 9 scenes from the Mip-NeRF 360 dataset [4], 5 scenes from the DTU dataset [27], 2 scenes from the Tanks and Temples dataset [30], and 2 scenes from the Deep Blending dataset [20]. Please refer to the Appendix for a detailed description of the dataset preparation process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 628 + ], + "type": "text", + "content": "For comparisons with Mip-NeRF 360 [4] and InstantNGP [39], we refer to Kerbl et al. [29] for results in scene-level datasets [4, 20, 30], and refer to NeRFBaselines [32] for results on the Blender dataset [37]. Results on the DTU dataset [27] are from our own runs of the authors' released code. Please refer to the Appendix for additional experimental details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Xu et al. [63] pointed out in Section 5.4 of their paper that their algorithm cannot handle scene-level datasets and objects with complex structures (i.e. Blender dataset), thus they only reported results on the DTU dataset [27]. We report those numbers and additionally show results on the Blender dataset from GStex [45] authors' own runs of the algorithm described in Xu et al. [63]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8968" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 89, + 553, + 424 + ], + "blocks": [ + { + "bbox": [ + 104, + 76, + 135, + 86 + ], + "lines": [ + { + "bbox": [ + 104, + 76, + 135, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 76, + 135, + 86 + ], + "type": "text", + "content": "3DGS*" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 77, + 253, + 86 + ], + "lines": [ + { + "bbox": [ + 232, + 77, + 253, + 86 + ], + "spans": [ + { + "bbox": [ + 232, + 77, + 253, + 86 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 351, + 76, + 381, + 86 + ], + "lines": [ + { + "bbox": [ + 351, + 76, + 381, + 86 + ], + "spans": [ + { + "bbox": [ + 351, + 76, + 381, + 86 + ], + "type": "text", + "content": "3DGS*" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 480, + 77, + 500, + 86 + ], + "lines": [ + { + "bbox": [ + 480, + 77, + 500, + 86 + ], + "spans": [ + { + "bbox": [ + 480, + 77, + 500, + 86 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 89, + 553, + 424 + ], + "lines": [ + { + "bbox": [ + 57, + 89, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 57, + 89, + 553, + 424 + ], + "type": "image", + "image_path": "ba48bd2b02151e1e2d1625baffd2989bb513cc0f5308839156cdcadda85a0271.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "type": "text", + "content": "Figure 5. Qualitative NVs results of benchmark datasets. Novel-view rendering results for both object-level and scene-level standard benchmark datasets. Given the same small number of Gaussians (on average " + }, + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{k}" + }, + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "type": "inline_equation", + "content": "39\\mathrm{k}" + }, + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "type": "text", + "content": " Gaussians for object-level and scene-level datasets, respectively), " + }, + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "type": "text", + "content": " fails to reconstruct high-frequency textures and complex shapes while our RGBA Textured Gaussians model succeeds. Please refer to the Appendix for results of our models optimized using varying number of Gaussians." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 499, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 545 + ], + "type": "text", + "content": "In addition to standard benchmark test scenes, we also capture our own set of scenes that contain artworks with highly detailed textures to help demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "3DGS Baseline. We build our method on top of our own implementation of a slightly modified 3DGS that closely follows the algorithm described in Gaussian Opacity Fields [68]. Specifically, we use a ray-based formulation that allows for the exact evaluation of 3D Gaussian values without approximating splatted 3D Gaussians as 2D Gaussians. In addition, we used a revised densification strategy [66, 68]. For fair comparisons with our method, we report the performance of our 3DGS implementation and label it as " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " throughout the results section, and refer the readers to the Appendix for quantitative results reported in the original 3DGS paper [29]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 432, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 432, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 432, + 510 + ], + "type": "text", + "content": "4.1. Quantitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 519, + 555, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 603 + ], + "type": "text", + "content": "In Table 1, we show quantitative comparisons in terms of PSNR/SSIM/LPIPS between our Textured Gaussians model using full RGBA textures and 3DGS* with the same number of Gaussians, and results of other baseline methods. Our method generally achieves better results than 3DGS* and all other methods, since Texture Gaussians are more expressive than vanilla 3D Gaussians due to the added texture maps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "In the bottom half of Table 1, we also show comparisons between our method and " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " when using a fraction of the default number of Gaussians, denoted by the percentage in parentheses. For our models with different numbers of Gaussians, we distribute a fixed amount of texels to all Gaussians (i.e., the same memory overhead for all Textured Gaussians models compared with " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "). Therefore, the texture map resolution of models with fewer Gaussians will be larger. We see that our method truly outshines " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8969" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 89, + 303, + 224 + ], + "blocks": [ + { + "bbox": [ + 104, + 76, + 135, + 86 + ], + "lines": [ + { + "bbox": [ + 104, + 76, + 135, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 76, + 135, + 86 + ], + "type": "text", + "content": "3DGS*" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 76, + 253, + 87 + ], + "lines": [ + { + "bbox": [ + 232, + 76, + 253, + 87 + ], + "spans": [ + { + "bbox": [ + 232, + 76, + 253, + 87 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 89, + 303, + 224 + ], + "lines": [ + { + "bbox": [ + 57, + 89, + 303, + 224 + ], + "spans": [ + { + "bbox": [ + 57, + 89, + 303, + 224 + ], + "type": "image", + "image_path": "4a9dbcda26ee4bfd6777efeebca4d6cba0098a5f613aa4ad0ddd980f6211ae41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "type": "text", + "content": "Figure 6. Qualitative NVS results of custom datasets. Our RGBA Textured Gaussians model achieves sharper reconstruction compared to " + }, + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 232, + 555, + 255 + ], + "type": "text", + "content": " when using the same small number of Gaussians (on average around 100k Gaussians)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 308, + 88, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 351, + 76, + 382, + 86 + ], + "lines": [ + { + "bbox": [ + 351, + 76, + 382, + 86 + ], + "spans": [ + { + "bbox": [ + 351, + 76, + 382, + 86 + ], + "type": "text", + "content": "3DGS*" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 88, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 308, + 88, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 308, + 88, + 553, + 224 + ], + "type": "image", + "image_path": "6ec87eedc6ff5d9a16aade2190d8b6236d17d894d0931848201705348f66a69d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 274, + 296, + 354 + ], + "blocks": [ + { + "bbox": [ + 56, + 274, + 296, + 354 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 296, + 354 + ], + "type": "table", + "html": "
Method3DGS*Alpha-onlyRGBA
PSNR#GS (M)PSNR#GS (M)PSNR#GS (M)
Blender33.080.2733.120.1933.030.21
Mip-NeRF 36027.264.427.373.127.263.5
DTU33.530.2833.450.2433.410.22
Tanks and Temples24.172.924.382.624.281.3
Deep Blending28.042.828.361.628.521.0
", + "image_path": "dd5b64ff4bf2d8917549cee122a4a05111968aee00de8ce7e3b7b0481a6d47d0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 361, + 296, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 449 + ], + "type": "text", + "content": "Table 2. Same model size comparisons. We compare the PSNR performance between " + }, + { + "bbox": [ + 55, + 361, + 296, + 449 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 361, + 296, + 449 + ], + "type": "text", + "content": " and our model using alpha-only and RGBA textures with the same model size. The average number of optimized Gaussians (in millions) from each dataset is shown in the #GS (M) columns. We boldface and underline the best and second-best performing models, respectively. Alpha-only texture models generally achieve better performance than " + }, + { + "bbox": [ + 55, + 361, + 296, + 449 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 361, + 296, + 449 + ], + "type": "text", + "content": " and RGBA texture models with the same model size." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 472, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 542 + ], + "type": "text", + "content": "with fewer Gaussians, achieving nearly 2 dB improvements over " + }, + { + "bbox": [ + 55, + 472, + 295, + 542 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 472, + 295, + 542 + ], + "type": "text", + "content": " when using " + }, + { + "bbox": [ + 55, + 472, + 295, + 542 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 542 + ], + "type": "text", + "content": " of the default number of Gaussians. Figure 4 shows the trend of the quality of novel view synthesis as we vary the number of Gaussians. Our method consistently outperforms " + }, + { + "bbox": [ + 55, + 472, + 295, + 542 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 472, + 295, + 542 + ], + "type": "text", + "content": " when using different numbers of Gaussians." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "content": "Table 2 compares the performance of " + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "content": " and our Textured Gaussians model with the same model size using alpha-only textures and RGBA textures. Since our models use per-Gaussian texture maps, they require fewer Gaussians than " + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "content": " to achieve the same model size. Our alpha-only textures model generally outperforms the " + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "content": " and RGBA textures models, indicating that there is a sweet spot for distributing the model size budget between Gaussian parameters and texture map channels. This is further explored in Section 4.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 168, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 168, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 168, + 684 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "Novel-View Synthesis. We show novel view synthesis results of our model with RGBA textures and 3DGS* on" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 276, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 346 + ], + "type": "text", + "content": "both standard benchmark and custom-captured datasets in Figures 5 and 6. Our method reconstructs much sharper details of the scene compared to " + }, + { + "bbox": [ + 313, + 276, + 555, + 346 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 313, + 276, + 555, + 346 + ], + "type": "text", + "content": " when using the same small number of Gaussians. Please refer to the Appendix for results of our models optimized using varying number of Gaussians." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "text", + "content": "Color Component Decomposition. We show the alpha-modulated and composited results of the two color components, " + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathrm{tex}}" + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathrm{base}}" + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "text", + "content": ", of the optimized alpha-only, RGB and RGBA textured Gaussians model with " + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "text", + "content": " of the default number of Gaussians in Figure 7. A 3DGS* model with the same number of Gaussians is shown for comparison. For RGB and RGBA models, " + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathrm{tex}}" + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "text", + "content": " reconstructs fine-grained textures that 3DGS* cannot. For alpha-only and RGBA models, " + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathrm{base}}" + }, + { + "bbox": [ + 313, + 349, + 556, + 480 + ], + "type": "text", + "content": " also reconstructs high-frequency details due to spatially varying alpha composition enabled by alpha textures." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 416, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 416, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 416, + 502 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Texture Map Variants. We ablate variants of our model that use different texture maps, namely alpha-only, RGB, and RGBA texture maps, and the same number of Gaussians in Figure 8. Experiments are conducted on Blender, Tanks and Temples, and Deep Blending datasets with the default optimized number (top row) and " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " of the default optimized number (bottom row) of Gaussians. We see that our model achieves the best performance with RGBA textures. Interestingly, using alpha-only textures already outperforms " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": ", and our RGB textures model despite being one-third the size, striking a better balance between performance and model size. This is because alpha-textured Gaussians can represent complex shapes through spatially varying opacity and reconstruct high-frequency textures through spatially varying alpha composition. In contrast, RGB-textured Gaussians are still limited to representing ellipsoids." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8970" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 553, + 209 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 553, + 209 + ], + "type": "image", + "image_path": "763145371f3925836e11dc43f411c2dbd20598dcf106497553929c47299b2112.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "type": "text", + "content": "Figure 7. Color component decomposition visualization. Visualization of alpha-modulated and composited " + }, + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathrm{base}}" + }, + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\mathrm{tex}}" + }, + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "type": "text", + "content": " color components in our Textured Gaussians model (Eq. 7, 8, 9). Insets with yellow borders show zoom-in crops. Texture color brightness is adjusted for visualization. While " + }, + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^{*}" + }, + { + "bbox": [ + 54, + 220, + 555, + 277 + ], + "type": "text", + "content": " fails to capture fine details with limited Gaussians, our method reconstructs sharp textures. With RGB textures, base color reconstructs low-frequency textures and texture maps reconstruct high-frequency elements. With alpha-only and RGBA textures, base color also reconstructs high-frequency textures via spatially varying alpha compositing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 294, + 294, + 392 + ], + "blocks": [ + { + "bbox": [ + 56, + 294, + 294, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 392 + ], + "type": "image", + "image_path": "716d6fa039cc6e3e470c1790d0cc6a9884050cffa4584931db959d8033984c10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 402, + 297, + 480 + ], + "lines": [ + { + "bbox": [ + 54, + 402, + 297, + 480 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 297, + 480 + ], + "type": "text", + "content": "Figure 8. Ablation study on texture map variants. We show the PSNR performance improvements (" + }, + { + "bbox": [ + 54, + 402, + 297, + 480 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 54, + 402, + 297, + 480 + ], + "type": "text", + "content": "PSNR) over " + }, + { + "bbox": [ + 54, + 402, + 297, + 480 + ], + "type": "inline_equation", + "content": "3\\mathrm{DGS}^*" + }, + { + "bbox": [ + 54, + 402, + 297, + 480 + ], + "type": "text", + "content": " of our Textured Gaussians model when using different texture map variants. Models with full RGBA textures achieve the best results. Interestingly, alpha-only textures outperform RGB textures and perform comparably to RGBA textures while using one-third and one-fourth of the model size, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 502, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 502, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 54, + 502, + 295, + 646 + ], + "type": "text", + "content": "Texture Map Resolution and the Number of Gaussians. In Figure 9, we optimize alpha-only Textured Gaussians models of the same model size with different texture map resolutions and number of Gaussians, and show the trend of novel view synthesis performance on the Blender and DTU datasets. We observe a sweet spot between texture map resolution and number of Gaussians that achieves the best image reconstruction quality. Maximizing texture map resolution corresponds to greatly reducing the number of Gaussians, which harms reconstruction of detailed geometric structures. Minimizing the texture map resolution, on the other hand, simply degenerates into 3DGS*." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 657, + 130, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 130, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 130, + 670 + ], + "type": "text", + "content": "5. Discussions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 296, + 715 + ], + "type": "text", + "content": "Limitations. The use of 2D diffuse texture maps in our model assumes that all textures lie on a local surface and does not model spatially varying specular color. Thus, ex" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 295, + 552, + 392 + ], + "blocks": [ + { + "bbox": [ + 317, + 295, + 552, + 392 + ], + "lines": [ + { + "bbox": [ + 317, + 295, + 552, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 552, + 392 + ], + "type": "image", + "image_path": "bab8bbfb851f74a0c0231a6d10545499e33298ab511d5602dc2d1260bd4be94f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 401, + 555, + 488 + ], + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 488 + ], + "type": "text", + "content": "Figure 9. Ablation study on texture map resolution and the number of Gaussians given the same model size. We optimized various alpha-only Textured Gaussians models (red dots) with equivalent model sizes but different texture map resolutions and numbers of Gaussian. A fitted quadratic curve (blue dashed line) illustrates the performance trend. The optimal model variant balances these two parameters to achieve superior novel view synthesis results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 605 + ], + "type": "text", + "content": "tending our texture representation to represent local 3D volume textures or even 5D radiance fields is crucial. Using factorized representations such as TensoRF [9] or triplane [7] to represent these high-dimensional textures could also be an interesting research problem. Finally, extending Textured Gaussians to support dynamic scene reconstruction [50, 64] through time-varying texture maps could be a natural next-step." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Conclusions. In this paper, we augment 3DGS with texture maps to allow individual Gaussians to model spatially varying colors and opacity. As such, each Gaussian can represent a much richer set of appearances and shapes. This greatly improves the expressivity of individual Gaussians, leading to better novel-view synthesis quality. Our method achieves better quality than 3DGS when using the same number of Gaussians, and achieves better or comparable quality when using the same model size." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "8971" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "type": "text", + "content": "We thank artist Kija Lucas for allowing us to use her art installation at the Palo Alto Art Center as one of our custom captured scenes (the flower gallery scene). Brian Chao is supported by the Stanford Graduate Fellowship and the NSF GRFP." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 114, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 114, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 114, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 182, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 182, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 296, + 217 + ], + "type": "text", + "content": "[1] K. Aliev, A. Sevastopolsky, M. Kolos, D. Ulyanov, and V. Lempitsky. Neural Point-Based Graphics, 2019. arXiv preprint. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 250 + ], + "type": "text", + "content": "[2] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 295, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 295 + ], + "type": "text", + "content": "[3] Milena T. Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Morgenstern. 3DGS.zip: A survey on 3D Gaussian Splitting Compression Methods, 2024. arXiv preprint. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 297, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 295, + 328 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In CVPR, 2022. 5, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 294, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 362 + ], + "type": "text", + "content": "[5] Matthew Brown and David G Lowe. Unsupervised 3D Object Recognition and Reconstruction in Unordered Datasets. In 3DIM, 2005. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 365, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 396 + ], + "type": "text", + "content": "[6] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising Densification in Gaussian Splatting. In ECCV, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 399, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 295, + 453 + ], + "type": "text", + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient Geometry-Aware 3D Generative Adversarial Networks. In CVPR, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 498 + ], + "type": "text", + "content": "[8] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth Synthesis and Local Warps for Plausible Image-Based Navigation. ACM TOG, 32 (3), 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 500, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 294, + 532 + ], + "type": "text", + "content": "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial Radiance Fields. In ECCV, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 534, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 294, + 578 + ], + "type": "text", + "content": "[10] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splitting. In CVPR, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 579, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 294, + 612 + ], + "type": "text", + "content": "[11] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: Hash-Grid Assisted Context for 3D Gaussian Splitting Compression. In ECCV, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "type": "text", + "content": "[12] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. NeuMesh: Learning Disentangled Neural Mesh-Based Implicit Field for Geometry and Texture Editing. In ECCV, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[13] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating Textures. In CGF, 2008. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[14] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and " + }, + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": " FPS. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "text", + "content": "[15] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-View Stereo for Community Photo Collections. In ICCV, 2007. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 153, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 196 + ], + "type": "text", + "content": "[16] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In CVPR, 2024. 1, 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 199, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 199, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 553, + 231 + ], + "type": "text", + "content": "[17] Antoine Guédon and Vincent Lepetit. Gaussian Frosting: Editable Complex Radiance Fields with Real-Time Rendering. In ECCV, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 234, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 553, + 276 + ], + "type": "text", + "content": "[18] Yuan-Chen Guo, Yan-Pei Cao, Chen Wang, Yu He, Ying Shan, and Song-Hai Zhang. VMesh: Hybrid Volume-Mesh Representation for Efficient View Synthesis. In SIGGRAPH Asia, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 279, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 279, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 279, + 553, + 311 + ], + "type": "text", + "content": "[19] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, second edition, 2004. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 313, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 313, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 313, + 553, + 357 + ], + "type": "text", + "content": "[20] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep Blending for Free-Viewpoint Image-Based Rendering. ACM TOG, 37(6), 2018. 2, 5, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 359, + 553, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 359, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 316, + 359, + 553, + 391 + ], + "type": "text", + "content": "[21] D. Hu, Z. Zhang, T. Hou, T. Liu, H. Fu, and M. Gong. Multiscale Representation for Real-Time Anti-Aliasing Neural Rendering. In ICCV, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "type": "text", + "content": "[22] Ronghang Hu, Nikhila Ravi, Alexander C. Berg, and Deepak Pathak. Worldsheet: Wrapping the World in a 3D Sheet for View Synthesis from a Single Image. In ICCV, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "type": "text", + "content": "[23] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In SIGGRAPH, 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 462, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 462, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 462, + 553, + 495 + ], + "type": "text", + "content": "[24] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In CVPR, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 497, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 497, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 497, + 553, + 529 + ], + "type": "text", + "content": "[25] Zhentao Huang and Minglun Gong. Textured-GS: Gaussian Splatting with Spatially Defined Color and Opacity, 2024. arXiv preprint. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 532, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 532, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 317, + 532, + 553, + 574 + ], + "type": "text", + "content": "[26] Junha Hyung, Susung Hong, Sungwon Hwang, Jaeseong Lee, Jaegul Choo, and Jin-Hwa Kim. Effective rank analysis and regularization for enhanced 3d gaussian splatting. In NeuRIPS, 2024. 4, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 577, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 577, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 577, + 553, + 609 + ], + "type": "text", + "content": "[27] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large Scale Multi-View Stereopsis Evaluation. In CVPR, 2014. 5, 2, 9" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 612, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 632 + ], + "type": "text", + "content": "[28] H. Kato, Y. Ushiku, and T. Harada. Neural 3D Mesh Renderer. In CVPR, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 677 + ], + "type": "text", + "content": "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splitting for Real-Time Radiance Field Rendering. ACM TOG, 42(4), 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction. ACM TOG, 36(4), 2017. 5, 2, 9" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8972" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 711 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[31] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-Based Neural Rendering with PerView Optimization. CGF, 40(4), 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[32] Jonas Kulhanek and Torsten Sattler. NerfBaselines: Consistent and Reproducible Evaluation of Novel View Synthesis Methods, 2024. arXiv preprint. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "text", + "content": "[33] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3D Gaussian Representation for Radiance Field. In CVPR, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[34] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In CVPR, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[35] Stephen Lombardi, Jason Saragih, Tomas Simon, and Yaser Sheikh. Deep Appearance Models for Face Rendering. ACM TOG, 37(4), 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "text", + "content": "[36] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering. In CVPR, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "type": "text", + "content": "[37] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 2, 5, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "text", + "content": "[38] Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3D Scene Representation via Self-Organizing Gaussian Grids. In ECCV, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "text", + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM TOG, 41(4), 2022. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "type": "text", + "content": "[40] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. In ECCV, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "text", + "content": "[41] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3D Gaussian Splitting for Accelerated Novel View Synthesis. In CVPR, pages 10349-10358, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "text", + "content": "[42] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the Memory Footprint of 3D Gaussian Splatting. ACM CGIT, 7(1), 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "text", + "content": "[43] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting. In CVPR, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "text", + "content": "[44] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. DreamGaussian4D: Generative 4D Gaussian Splatting, 2023. arXiv preprint. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "text", + "content": "[45] Victor Rong, Jingxiang Chen, Sherwin Bahmani, Kiriakos N Kutulakos, and David B Lindell. GStex: Per-Primitive Texturing of 2D Gaussian Splatting for Decoupled Appearance and Geometry Modeling, 2024. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "type": "text", + "content": "[46] Antoni Rosinol, John J Leonard, and Luca Carlone. NeRF-SLAM: Real-time Dense Monocular Slam with Neural Radiance Fields. In IROS, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 711 + ], + "type": "text", + "content": "[47] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields without Neural Networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[48] Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A Comparison and Evaluation of Multi-View Stereo Reconstruction Algorithms. In CVPR, 2006. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 119, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 163 + ], + "type": "text", + "content": "[49] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In CVPR, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 165, + 553, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 165, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 316, + 165, + 553, + 197 + ], + "type": "text", + "content": "[50] Meng-Li Shih, Jia-Bin Huang, Changil Kim, Rajvi Shah, Johannes Kopf, and Chen Gao. Modeling ambient scene dynamics for free-view synthesis. 2024. 1, 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 199, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 199, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 553, + 242 + ], + "type": "text", + "content": "[51] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene Representation Networks: Continuous 3D-Structure-Aware Neural Scene Representations. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 245, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 553, + 277 + ], + "type": "text", + "content": "[52] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo Tourism: Exploring Photo Collections in 3D. ACM TOG, 25 (3), 2006. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "type": "text", + "content": "[53] Yunzhou Song, Heguang Lin, Jiahui Lei, Lingjie Liu, and Kostas Daniilidis. HDGS: Textured 2D Gaussian Splitting for Enhanced Scene Rendering, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 315, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 553, + 358 + ], + "type": "text", + "content": "[54] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. NeRV: Neural Reflectance and Visibility Fields for Relighting and View Synthesis. In CVPR, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "text", + "content": "[55] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Articulated Neural Radiance Fields for Learning Human Shape, Appearance, and Pose. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 406, + 553, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 438 + ], + "type": "text", + "content": "[56] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. BillBoard Splatting (BBSplat): Learnable Textured Primitives for Novel View Synthesis, 2025. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 441, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 473 + ], + "type": "text", + "content": "[57] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative Gaussian Splitting for Efficient 3D Content Creation. In ICLR, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 475, + 553, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 528 + ], + "type": "text", + "content": "[58] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Wang Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in Neural Rendering. CGF, 41(2), 2022. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "text", + "content": "[59] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 578, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 620 + ], + "type": "text", + "content": "[60] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "text", + "content": "[61] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-Nerf: Point-Based Neural Radiance Fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[62] Rui Xu, Wenyue Chen, Jiepeng Wang, Yuan Liu, Peng Wang, Lin Gao, Shiqing Xin, Taku Komura, Xin Li, and Wenping Wang. SuperGaussians: Enhancing Gaussian Splitting Using Primitives with Spatially Varying Colors, 2024. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8973" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 386 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[63] Tian-Xing Xu, Wenbo Hu, Yu-Kun Lai, Ying Shan, and Song-Hai Zhang. Texture-GS: Disentangling the Geometry and Texture for 3D Gaussian Splitting Editing. In ECCV, 2024. 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[64] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3D Gaussians for High-Fidelity Monocular Dynamic Scene Reconstruction. In CVPR, 2024. 1, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "text", + "content": "[65] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian Grouping: Segment and Edit Anything in 3D Scenes. In ECCV, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "text", + "content": "[66] Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details in 3d gaussian splatt-ing. In ACM-MM, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "text", + "content": "[67] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for Real-Time Rendering of Neural Radiance Fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[68] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes. ACM TOG, 2024. 6, 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[69] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural Factorization of Shape and Reflectance under an Unknown Illumination. ACM TOG, 40(6), 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "type": "text", + "content": "[70] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM. In CVPR, 2022. 1" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8974" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/32f81902-9541-4c54-8edb-2ca63c430edb_content_list.json b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/32f81902-9541-4c54-8edb-2ca63c430edb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..60176a936090e40e9ceff6aad9d1447f3edf5161 --- /dev/null +++ b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/32f81902-9541-4c54-8edb-2ca63c430edb_content_list.json @@ -0,0 +1,1974 @@ +[ + { + "type": "text", + "text": "The Art of Deception: Color Visual Illusions and Diffusion Models", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexandra Gomez-Villa $^{1,3}$", + "bbox": [ + 196, + 179, + 405, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bartłomiej Twardowski1,2,4 Jesus Malo3", + "bbox": [ + 101, + 198, + 452, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Wang $^{1}$ C.Alejandro Parraga $^{1,2}$", + "bbox": [ + 457, + 180, + 800, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Javier Vazquez-Corral $^{1,2}$ Joost van de Weijer $^{1,2}$", + "bbox": [ + 488, + 198, + 893, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Computer Vision Center, Spain 2Universitat Autònoma de Barcelona, Spain", + "bbox": [ + 169, + 232, + 826, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Universitat de València, Spain 4IDEAS-NCBR, Poland", + "bbox": [ + 251, + 251, + 745, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da5bb77c7e1365f142870551f9a2b4773119ff861ddce6873d47590b2d8acde5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 280, + 356, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18b74c8002a322e16a74b95332a134f0e2ae997c11b02e0086ff2054044336ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 280, + 612, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f07cabcd244c1607d3a1b56a92a9bd9834098043364b6cfa8b1e652534001d84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 282, + 870, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f053106e4ace068654512d02161eaf41a3641fa67e3e6343662ca70695b54dda.jpg", + "image_caption": [ + "Figure 1. Observation: Denoising Diffusion Implicit Models (DDIM) have human-like visual illusions. Application: text-to-image generated visual illusions. Examples (a) and (b) show how the responses to physically equal patches shift differently in the path to the latent space. In contrast, (c) shows how physically different stimuli get more similar along the path, achieving color constancy. The observation of this human-like behavior allows us to propose the use of text-to-image models to generate images in which physically identical patches are perceived differently (examples d, e, f). We highly recommend to watch the illusions on a computer screen." + ], + "image_footnote": [], + "bbox": [ + 98, + 410, + 359, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd5ac4d23ce14017d6bfb159ce67c3c260b1a10d92d4ea975c2fdc975d46380b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 411, + 614, + 553 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4bb4104fef48fbf94c93049adbf74d9a4876eaa97d60bb6ce968f0dbe0225a0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 410, + 893, + 553 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 642, + 326, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual illusions in humans arise when interpreting out-of-distribution stimuli: if the observer is adapted to certain statistics, perception of outliers deviates from reality. Recent studies have shown that artificial neural networks (ANNs) can also be deceived by visual illusions. This revelation raises profound questions about the nature of visual information. Why are two independent systems, both human brains and ANNs, susceptible to the same illusions? Should any ANN be capable of perceiving visual illusions? Are these perceptions a feature or a flaw? In this work, we study how visual illusions are encoded in diffusion models. Remarkably, we show that they present human-like brightness/color shifts in their latent space. We use this fact to demonstrate that diffusion models can predict visual illusions. Furthermore, we also show how to gen", + "bbox": [ + 88, + 674, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "erate new unseen visual illusions in realistic images using text-to-image diffusion models. We validate this ability through psychophysical experiments that show how our model-generated illusions also fool humans.", + "bbox": [ + 511, + 643, + 906, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 717, + 645, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual illusions (VIs) serve as a fascinating window into the complex workings of human visual perception, where our interpretation of reality diverges significantly from physical measurements. These phenomena can have a profound social impact, as demonstrated by viral phenomena such as \"The Dress\" [9], which sparked global discussions on the nature of perception. Beyond their intrinsic scientific value,", + "bbox": [ + 511, + 743, + 906, + 849 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 31 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Author (☑): Kai Wang (kwang@cvc.uab.es)", + "bbox": [ + 537, + 862, + 857, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code is available at https://alviur.github.io/color-illusion-diffusion", + "bbox": [ + 514, + 876, + 903, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18642", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VIs offer a unique lens through which we can examine and compare human and machine perception [13, 33]. The emergence of diffusion models (DMs) as state-of-the-art generative models has revolutionized the field of artificial intelligence, demonstrating unprecedented capabilities in image synthesis and manipulation. Their denoising-based architecture, which gradually transforms random noise into coherent images, presents a novel paradigm for studying visual processing that mirrors the lossy and noisy process of human perception. This parallel makes them particularly interesting for investigating VIs, as both systems must navigate the complex relationship between image statistics and perceptual interpretation.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The study of VIs in artificial neural networks (ANNs) has a rich history. Early work demonstrated that a multilayer perceptron could replicate brightness illusions [8], while subsequent research by Gomez-Villa et al. [12] showed that convolutional neural networks (CNNs) could mirror human perception in lightness and color illusions. Concurrent developments included Watanabe et al.'s [57] demonstration of motion VI replication in video-prediction CNNs and the publication of the first visual illusion dataset [58]. This foundation has spawned diverse investigations into the ANN-VI relationship in multiple domains: image completion [19, 50], brightness/color perception [13, 17, 24, 32, 33, 52], contrast sensitivity [1, 30], shape perception [37, 59], and vision-language models [6, 38, 47, 60].", + "bbox": [ + 89, + 291, + 482, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we investigate how human-like perceptual phenomena naturally emerge in the encoding process of DMs. Our key finding reveals that the intermediate steps of diffusion models exhibit trends that align remarkably with human perception. Specifically, when analyzing the trajectory of an inverted image through the diffusion process, we observe that the progressive changes in lightness and color properties closely mirror human perceptual responses. Prior research has demonstrated that Gaussianization flows exhibit interesting properties when processing samples that deviate from their training data statistics. When trained on specific sample distributions, these flows can produce unusual representations for out-of-distribution samples [27, 43] - a characteristic that has proven valuable for measuring distribution divergence [29]. Intriguingly, this phenomenon parallels human visual perception, where shifts in image statistics can lead to illusory perception of spatio-chromatic patterns, a behavior that has been explained through manifold equalization [26, 28]. We argue that image samples that do not follow the statistics of natural images are not properly gaussianized by DMs. In this way, linking defective gaussianization and illusory perception of unusual samples, it makes sense that intermediate representations in DMs may be correlated with the shifts of these samples for human observers. The reported alignment suggests that the iterative denoising in diffusion models may", + "bbox": [ + 89, + 508, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "inadvertently capture fundamental aspects of human visual perception, particularly in contexts where our perception diverges from physical reality. By examining these intermediate representations, we provide novel insights into the nature of visual illusions and the unexpected parallels between human perception and the diffusion process.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions: The main contributions of this work are:", + "bbox": [ + 511, + 181, + 888, + 196 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We report an empirical observation of the behavior of diffusion models: progressive changes in lightness, hue and saturation of image regions in the forward/inverse transform to the latent representation replicate the shifts of human perception of these regions.", + "- We derive a novel approach to using diffusion models to model human vision. We show that deep autoencoders can be effectively employed as vision models by performing a diffusion inversion process and measuring perception in the intermediate noise space.", + "- We demonstrate the human perception replication capabilities of DM in many cases, including VI datasets, popular published VI works, and natural images.", + "- Leveraging the perceptual capabilities of diffusion models, we propose a novel method for generating brightness/color VIs using text-to-image models." + ], + "bbox": [ + 511, + 196, + 903, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 450, + 651, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Replication of Visual Illusions. Replication of visual illusions has been made with several methods, but all of them can be roughly classified into the following categories:", + "bbox": [ + 511, + 477, + 903, + 523 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image restoration methods [12, 13, 24, 30, 33]: These use pre-trained image restoration backbones. They process VIs and measure the output image to measure perception.", + "- Effect bias methods [1]: These works are inspired by neuroscience measurement of stimuli. Pretrained architectures are used to process VIs, then the response in the internal activations of the ANN is used to define if there is a visual illusion.", + "- Likelihood method [17]: In this approach, the most probable perceived intensity of each pixel in an image is estimated using an invertible flow trained on natural images.", + "- Language-guided methods [38, 60]: Language is used as a vehicle to measure perception. These works use vision-language models as test subjects and ask them questions regarding the perception of the image." + ], + "bbox": [ + 513, + 523, + 903, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach differs from previous methods in the following: 1) we do not measure perception in the output of the model. Instead, we measure the VI in the intermediate steps of image inversion process. 2) We use a deep autoencoder; previous approaches in this category claim that deep autoencoders suffer from less perceptual bias and use shallow architectures for replication experiments. Furthermore, we are able to replicate human perception in high resolution images (as big as DMs are able to manage) and in natural images. These setups were not possible in previous works.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18643", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generation of Visual Illusions. Although generative models have been used to synthesize visual illusions [14], DMs has only recently been proposed as the generative engine. In [6], high-level illusions (overlay effects) are generated by mixing a set of images and transformations together. Following works [10, 11] were able to create more general illusions such as rotations, color inversions, and hybrid images. The work most similar to ours is by Roy et al. [46], who used stable diffusion (SD) to generate brightness visual illusions through keyword prompts. However, their approach neither explicitly optimizes perceptual effects nor specifies regions for perception replication. Consequently, illusions may appear randomly in the generated image with unknown localization. Additionally, their generated images are limited to conventional brightness illusions (gratings or high-frequency black and white patterns) and cannot incorporate visual illusions into natural images. In contrast, our work allows visual illusions within generated natural images as well as in typical high-frequency patterns. We explicitly generate the perceptual effect in a designated image region and can control the image content through text prompts.", + "bbox": [ + 89, + 90, + 483, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 422, + 191, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion models", + "text_level": 1, + "bbox": [ + 89, + 449, + 256, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models (DMs) are generative models that learn to gradually denoise images by reversing a fixed Markov chain of forward diffusion steps. The forward process systematically adds Gaussian noise to an image until it becomes pure noise, while the learned reverse process reconstructs the original image through iterative denoising. While traditional diffusion follows a stochastic sampling process, Denoising Diffusion Implicit Models (DDIM) [49] introduce a deterministic sampling approach that maintains high quality while enabling significantly fewer inference steps.", + "bbox": [ + 89, + 470, + 483, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DDIM's deterministic nature enables a crucial capability: the ability to invert generated images back into their latent representations through a process known as DDIM inversion [49]. This inversion process allows us to map any image into the model's latent space and observe its trajectory through the diffusion process. Given an image $z_0$ , DDIM inversion produces a sequence of intermediate representations $\\{z_t\\}_{t=1}^T$ according to:", + "bbox": [ + 89, + 623, + 483, + 747 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {z _ {t}, t} (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 758, + 482, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t$ represents the cumulative product of noise scheduling coefficients, and $\\epsilon_{\\theta}^{z_t,t} = \\epsilon_{\\theta}(z_t,t)$ is the learned noise prediction network. The reverse process follows:", + "bbox": [ + 89, + 791, + 483, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(\\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {z _ {t} , t}}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\epsilon_ {\\theta} ^ {z _ {t}, t} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 847, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DDIM inversion has emerged as a powerful tool for various image manipulation tasks. Its ability to faithfully reconstruct latent representations has enabled applications such as semantic image editing [31, 35, 36, 51, 54], style transfer [25, 41], image interpolation [53], and controlled image generation through classifier-free guidance [18]. In our work, we exploit the deterministic nature of DDIM inversion to systematically study how visual illusions are processed through the diffusion trajectory.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. DDIM inversion may replicate human vision", + "text_level": 1, + "bbox": [ + 511, + 238, + 890, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The wholly-empirical evolutionary paradigm [20] suggests that human vision is shaped by statistical regularities in natural scenes, where our visual system evolved to interpret ambiguous stimuli based on their likelihood. We leverage this insight to study how diffusion models, trained for natural images, process illusory stimuli, which are not in the distribution of natural images.", + "bbox": [ + 511, + 261, + 906, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. In Fig. 2, we present an illustrative experiment using Stable Diffusion [44] (extra results with an image space model in the supplementary material). We perform DDIM inversion on the canonical visual illusion known as Brightness Contrast [5]. DDIM inversion puts the image back into the latent space. This illusion presents two gray squares of identical intensity (indicated by red markers) against white and black backgrounds, which leads humans to perceive the left square as darker than the right one. The first row (left to right) shows the inversion1 for 0, 3, 10, and 20 steps, respectively. As the zoomed-in bar shows, the initially equal-intensity regions progressively develop an intensity difference that aligns with human perception2 (left target appearing darker than right). The second row shows the progressive Gaussianization of the latent representation. For this image perception is replicated at an intermediate representation where gaussianization is not complete (since this stimulus is not in the distribution of natural images).", + "bbox": [ + 511, + 368, + 908, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observation: emergence of perceptual biases. Our key observation is: when inverting an image through diffusion models the intermediate denoising steps progressively introduce color/brightness shifts that align with human perception. We hypothesize that this occurs because diffusion models, trained on natural images, map illusory stimuli onto the learned natural image manifold, resulting in systematic \"interpretation errors\" that mirror human perceptual biases.", + "bbox": [ + 511, + 641, + 908, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theoretical framework. Formally, let $G: \\mathcal{X} \\to \\mathcal{Z}$ be the forward diffusion mapping from image space $\\mathcal{X}$ to latent space $\\mathcal{Z}$ . For natural images $x \\in \\mathcal{X}$ , $G$ learns a mapping $z = G(x)$ that transforms the image distribution into a Gaussian. However, when presented with an eventually", + "bbox": [ + 511, + 762, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "we decode the latent representation to visualize it in image space.", + "bbox": [ + 529, + 849, + 883, + 863 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2In addition to the color/brightness effects, notice how the high-frequency content of the image is maintained; we explore this phenomenon further in supplementary material.", + "bbox": [ + 514, + 863, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18644", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2b799b5d302dfd681838acea957ac44f3c05f628c5c3f9847b010cdd24218d21.jpg", + "image_caption": [ + "Figure 2. DDIM inversion of the Brightness Contrast illusion [5] using Stable Diffusion. Top row: Image-space visualization (decoded latents) showing (left) Original illusion with two identical gray squares (marked in red) against black and white backgrounds, and inversion results using 3, 10, and 20 steps. Bottom row: Histograms of the corresponding latent representations. The horizontal axis represents the value of the latent space feature. The model gradually reproduces the perceptual difference in brightness between the physically identical squares in a (not fully Gaussian) intermediate representation." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 893, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "illusory stimulus $y$ , we have $y \\notin \\mathcal{X}$ . Still, the mapping $z' = G(y)$ attempts to project the illusion onto the learned natural image manifold, introducing systematic deviations. The intermediate states $z_{t}$ reveal how the model progressively \"corrects\" the illusion toward natural image statistics. We observe that these corrections correlate strongly with human perception, suggesting that both humans and diffusion models may operate on similar statistical principles when processing ambiguous visual input. This alignment between diffusion trajectories and human perception provides computational evidence for the wholly-empirical paradigm, demonstrating how statistical learning can lead to perceptual illusions without explicit encoding of perceptual rules. The systematic nature of these \"errors\" suggests they emerge from the same statistical regularities that shape human visual processing through evolution.", + "bbox": [ + 91, + 375, + 483, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Generating visual illusions with DMs", + "text_level": 1, + "bbox": [ + 89, + 631, + 410, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the findings above, we develop a method to automatically generate brightness/color illusions with DMs. To do so, we need to evaluate whether a VI is present and then push the diffusion process toward the desired perception.", + "bbox": [ + 89, + 655, + 482, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous works on VIs and DMs focus on high-level visual illusions, i.e. cognitive effects that measure if the image belongs to a certain class of prompt [6, 10, 11]. Here we focus on brightness/color VIs, which leads to the challenge of measuring effects in specific regions of the image. Here we answer the measurement questions proposed in previous works on brightness/color VI generation in this way:", + "bbox": [ + 89, + 717, + 482, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where to measure? This question separates the generation of high-level VIs from color/brightness VIs. Models aiming at high-level VIs that fool an ANN require to ask a classifier/vision-language model which class/prompt better describes the image [10, 38, 60]. In contrast, color VIs ap", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pear in specific regions of the image, therefore, they need to be measured at those points. In this work, we specifically define regions (here targets) in which we want a perceptual effect to appear. These targets are our measurement points, and we guarantee that in these regions, we can track, measure, and optimize perceptual effects. Failing to define a target leads to failure, as reported in [58].", + "bbox": [ + 511, + 375, + 906, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "What to measure? During image generation, we will measure the intensity values of the intermediate steps $z_{t}$ . Even though $z_{t}$ does not contain the final perception bias $(z_{T})$ , we can push the generative process towards a desired perceptual effect using small changes in the intermediate representation as explained in Fig. 3.", + "bbox": [ + 511, + 479, + 905, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How to measure? In the motivating example of section 3.2 (Fig. 2), the model replicates human perception if the target shifts in the same direction as humans perceive. However, in an optimization procedure, we need to measure how \"good\" the illusion is; in other words, we need to quantify the effect. Since we are concerned about brightness and color VI, we will measure the mean intensity per channel of each target. More precisely, given an image $I$ , and the target region $r \\in \\mathbb{R}^2$ in $I$ , we compute the mean intensity in that target region as:", + "bbox": [ + 511, + 571, + 906, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i n t} (I, r) = \\frac {1}{M} \\sum_ {i \\in r} I \\left(r _ {i}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 733, + 906, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r_i$ denotes a specific pixel $i$ in target region $r$ , $M$ stands for the number of pixels in the target region, and $m_{int}$ is measured in each channel. Hence, the perceived intensity is an RGB triplet $(m_{int}^R, m_{int}^G, m_{int}^B)$ .", + "bbox": [ + 511, + 781, + 906, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Visual perception inside generation pipeline.", + "text_level": 1, + "bbox": [ + 511, + 849, + 867, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since text-to-image (T2I) diffusion models often struggle to perfectly align with textual prompts, a variety of aug-", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18645", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58ab821c681add6ca3fd7ee579250a40d4bd6a93cc00345f3c7d35a3221fc959.jpg", + "image_caption": [ + "Figure 3. Overview of our visual illusion generation pipeline. The process modifies noisy latent representations, $z_{t}$ , through a custom loss function, $\\mathcal{L}$ , that guides the generation toward perceptually ambiguous outputs. Note that targets are regions that manually selected by the user as inputs. The terms $D_{p1}$ and $D_{p2}$ are the user desired intensity for target T1 and T2 respectively." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 468, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mentation methods have been developed to enhance the fidelity of T2I generation. These methods typically focus on optimizing noisy latent representations in each timestep, a strategy that has proven effective in refining generation outcomes [7, 42, 55, 61]. A common approach involves the design of an update loss function, $\\mathcal{L}$ , which is tailored based on the layout of the noisy latent or the characteristics of the cross-attention maps. This loss is then used to guide the backpropagation process, updating the noisy latent in a manner that more accurately reflects the intended textual descriptions, thereby enhancing the overall quality of the generated images. This process can be formulated as $z_{t}^{\\prime} = z_{t} - \\alpha \\cdot \\nabla_{z_{t}}\\mathcal{L}$ . In this paper, we diverge from the conventional use of optimization techniques aimed at aligning T2I diffusion models with textual semantics. Instead, we leverage these techniques to create visual perception effects (see Fig. 3). We use an image space DM (DeepFloyd [2]) to avoid the use of the decoder in the backpropagation.", + "bbox": [ + 91, + 393, + 482, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we define the original target image as $O(r)$ , where $r \\in \\mathbb{R}^2$ . This image only contains the input-defined targets, that is, $O(r)$ has the original value given to the targets by the user on each color channel. Then, given an intermediate image $z_t$ and a list of $R$ target regions $\\{r^p\\}_{p=1}^R$ , our process to generate a VI includes the addition of the targets to the image in the desired region3, i.e. we compute $z_{t\\_target}(r_i^p) = z_t(r_i^p) + O(r_i^p)$ , for all pixels $i$ inside the region $r^p$ and $z_{t\\_target} = z_t$ for any pixel outside this target region. Finally, the following optimization term is applied $N$ times —per diffusion step $t$ — to guide the generation:", + "bbox": [ + 89, + 666, + 482, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\gamma \\mathcal {L} _ {V I} + \\beta \\mathcal {L} _ {s i m} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 840, + 480, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{VI}$ and $\\mathcal{L}_{sim}$ measure the illusion strength and the", + "bbox": [ + 89, + 864, + 482, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "similarity between targets and background respectively, and $\\gamma$ and $\\beta$ are hyperparameters that control the relative weight of these measures. Specifically:", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- $\\mathcal{L}_{VI}$ is a perceptual loss on the target intensities:", + "bbox": [ + 513, + 140, + 849, + 156 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {V I} = \\sum_ {c \\in R, G, B} \\sum_ {r ^ {p}} ^ {R} \\mathcal {L} _ {\\text {P e r c e p t}} \\left(z _ {t _ {\\text {t a r g e t}}}, k _ {c} ^ {p}; r ^ {p}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 167, + 905, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r^p$ is a specific target, $k_c^p$ is the desired intensity for target $r^p$ on color channel $c$ , and", + "bbox": [ + 526, + 223, + 905, + 253 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {P e r c e p t u a l}} \\left(z _ {t _ {\\text {t a r g e t}}}, k _ {c}; r\\right) = \\left\\| m _ {\\text {i n t}} \\left(z _ {t _ {\\text {t a r g e t}}}, r\\right) - k _ {c} \\right\\| \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 267, + 903, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $m_{int}$ as defined in Eq. 3.", + "bbox": [ + 526, + 299, + 727, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- $\\mathcal{L}_{sim}$ is a similarity term. This term aims at avoiding the region from appearing \"out of place\". In particular, it looks at harmonizing the targets with the background-generated image:", + "bbox": [ + 513, + 314, + 905, + 373 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s i m} = \\sum_ {p = 1} ^ {R} M S E \\left(z _ {t} \\left(r ^ {p}\\right), O \\left(r ^ {p}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 386, + 905, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental setup", + "text_level": 1, + "bbox": [ + 513, + 441, + 702, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 468, + 614, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Brightness illusions dataset: The $BRI3L$ dataset [45] consists of 22366 grayscale images spanning five types of brightness illusions. Each image is annotated with binary masks indicating regions of illusory perception.", + "bbox": [ + 513, + 489, + 905, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Visual Illusion VQA Datasets: Several datasets have been developed to evaluate vision-language models' understanding of VIs: IllusionVQA [47] contains 435 illusory images with multiple-choice questions across 12 categories, plus 1000 impossible object image pairs for localization testing. GVIL [60] features 100 base illusion images expanded to 1600 instances covering five illusion categories, with four evaluation tasks: same-difference QA, referential QA, attribute QA, and referential localization. HallusionBench [15] comprises 346 images and 1129 questions designed to test image-context reasoning and diagnose both VIs and language hallucinations. We use the color category in all the VQA datasets.", + "bbox": [ + 513, + 549, + 905, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Benchmarks", + "text_level": 1, + "bbox": [ + 513, + 757, + 645, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Classical vision science models: These approaches are based in vision science and need to be calibrated for each image to fully exploit their capabilities. In this paper we use the parameters set by Betz et al. [3]. These parameters constitute a fair guideline for these methods, as it is not possible to fine-tune them for thousands of images. We use ODOG (Oriented Difference of Gaussians) [4] and the Chromatic Induction Wavelet Model (CIWaM) [40]", + "bbox": [ + 513, + 779, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3Note that this addition results in the perceived color (see sect. 3.2)", + "bbox": [ + 107, + 886, + 465, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18646", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Neural networks-base models: We compare against previous image restoration methods that proven ability in visual illusion detection: DN-NET[13] and Restorenet[30].", + "- Variational autoencoder (VAE): We present the results of only considering the VAE of SD prior to the DDIM inversion process —to further demonstrate DDIM is the responsible for the alignment with human perception." + ], + "bbox": [ + 89, + 90, + 478, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Metrics", + "text_level": 1, + "bbox": [ + 89, + 207, + 186, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While visual illusion replication results have traditionally been presented through profile plots comparing model outputs with human perception along image cross-sections, we leverage recent datasets to provide quantitative metrics. This approach complements classical visualization methods with rigorous numerical evaluation.", + "bbox": [ + 89, + 229, + 480, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the BRI3L dataset, we compute the mean intensity difference, $\\Delta I$ , between the model's output and the input image within each segmented region:", + "bbox": [ + 89, + 321, + 480, + 367 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta I = \\frac {1}{| M |} \\sum_ {i \\in r} \\left| I _ {\\text {o u t}} (r) - I _ {\\text {i n}} (r) \\right| \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 378, + 480, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ represents the size of the segmentation mask, and $I_{out}$ , $I_{in}$ are the output and input intensities respectively. We consider the model's response to be aligned with human perception when $\\Delta I < \\tau * I_{int}$ , where $I_{int}$ represents the mean intensity of the input image in the target region (calculated using Eq. 3). Since BRI3L aims to detect regions that become darker, the metric becomes more harder as the value of $\\tau$ decreases, requiring the model to achieve greater darkness in the target regions.", + "bbox": [ + 89, + 425, + 482, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the VQA-based datasets (IllusionVQA, GVIL, and HallusionBench), we manually create segmentation masks and consider the answer of the model aligned with human perception if it matches the expected intensity for regions referenced in the question-answer pairs. The Perception Accuracy Score (PAS) is the percentage of images having the correct perceptual predictions in each dataset.", + "bbox": [ + 89, + 561, + 482, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 678, + 225, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted two sets of experiments to evaluate our method's ability to: (1) replicate existing VIs across multiple datasets, and (2) generate novel VIs that can fool human perception. For each experiment, we present qualitative results zooming in on the region of the perceptual effect and quantitative evaluations using the proposed metrics.", + "bbox": [ + 89, + 700, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Replication of visual illusions: We test our model's ability to replicate brightness (BRI3L) and color (Illusion-VQA, GVIL, and HallusionBench) illusions by comparing the intensity differences between input and output images within annotated segmentation masks. For BRI3L, we evaluate performance across three $\\tau$ levels (0.8, 0.9, 1.0) to assess the robustness of illusion replication. In this", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d8f4e79bfbdfa82798eabb9d3cb5ac8b1ecaab0628d60c9c86ec0feba8d4dee7.jpg", + "table_caption": [ + "Table 1. Perceptual Accuracy Score (PAS) in BRI3L for different methods and thresholds" + ], + "table_footnote": [], + "table_body": "
Method / Threshold0.80.91.0
ODOG [4]36.5857.0972.44
CIWAM [40]5.6918.2356.22
DNnet [12]0.000.0071.09
RestoreNet [13]0.000.0074.95
VAE0.000.0873.18
DDIM 5 steps44.0291.4399.88
DDIM 10 steps84.1798.85100.00
", + "bbox": [ + 540, + 121, + 874, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "experiment we present results using SD 1.4 (latent space DM) but additional results with DeepFloyd (image space DM) are provided in the supplementary material.", + "bbox": [ + 526, + 290, + 903, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Generation of visual illusions: In this experiment, we check whether we can generate VIs in images that fall inside the distribution of the T2I model. We use guidance scale $= 10$ (we want to generate images that follow the prompt), $N = 5$ , $\\gamma = 0.5$ , and $\\beta = 1.0$ . Additionally, here we use S2 of DeepFloyd to generate higher-resolution versions of our generated visual illusions. For flat intensity targets (non-gradients), we intentionally paste the targets in the output of S2 as the superresolution model may change the content of the targets.", + "bbox": [ + 513, + 335, + 903, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 501, + 599, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Replication of visual illusions", + "text_level": 1, + "bbox": [ + 511, + 526, + 774, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. We present qualitative results demonstrating our method's ability to replicate human visual perception across diverse scenarios in Fig. 4. Each comparison shows the original input image (left) and its corresponding DDIM inversion after 5 steps (right), with zoomed inset regions highlighting the areas where perceptual effects occur. The top row showcases our method's effectiveness on classical visual illusions, including Hermann grid and grating induction patterns. The bottom panels demonstrate the model's capability to capture perceptual phenomena in both natural grayscale photographs and artistic images with different color schemes. The zoomed regions, emphasize physically identical targets that are perceived differently due to their surrounding context — a key characteristic of visual illusions that our model successfully replicates. Additional qualitative results showing our performance across a broader range of visual illusions are in the supplementary.", + "bbox": [ + 511, + 550, + 906, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. Table 1 presents a comprehensive comparison of our DDIM-based approach against classical vision models and deep learning methods across multiple visual illusion datasets. The results demonstrate the consistency of the observation on diffusion models reported in Section 3.2 over a range of datasets and their superior", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18647", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b611aea49589fd6f66b1108aa700f3e16e4366e2f6ca8111e51cbfd27d534348.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 85, + 184, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ff1e0979d122db9af8dc22424f2b3512a0d299a9f32c05a3e410108c0b82791.jpg", + "image_caption": [ + "a)" + ], + "image_footnote": [], + "bbox": [ + 199, + 87, + 272, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/92b1e53973f4f6bf8d6c2e884c8e147cedd2c7b7ec311701a85ede9af656ea05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 95, + 364, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd8bb5fbe4c2bf96c67b75678c30b0a472f58757a4827ccb4655e35e9b612190.jpg", + "image_caption": [ + "b)" + ], + "image_footnote": [], + "bbox": [ + 370, + 90, + 460, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c055d17c3ff10faf73f30d7a1d0e42a2449cf21d576c1190af3336fa2e7a6b1f.jpg", + "image_caption": [ + "c)" + ], + "image_footnote": [], + "bbox": [ + 462, + 88, + 529, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/129bc26c3ee619676a50a74f940dcfc798289677b332db4674c2fd478095fb80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 88, + 602, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/300bbf70a7eb344cd8a75b026438aa9f710a39b02cc9a266595dc336a97af030.jpg", + "image_caption": [ + "d)" + ], + "image_footnote": [], + "bbox": [ + 604, + 99, + 673, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be5912884070aade400b4944ca68c088a3c4bebfd6c7cc75f2416ba0a0aaf6b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 99, + 750, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e60ae82537d228e22a5278c4cd36a45dc9c236de1b9020a57ad89ee5e3675fe2.jpg", + "image_caption": [ + "e)" + ], + "image_footnote": [], + "bbox": [ + 754, + 99, + 890, + 154 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af51c829bee1546b67540611cf18dc473c521345df79a290552f927b990ea825.jpg", + "image_caption": [ + "Figure 4. Qualitative replication of visual illusions (please enlarge display): For each item, the original image is on the left and its DDIM inversion on the right. The observation done in Fig. 2 is consistently reproduced. We zoom in on the region containing the illusion in each image pair. a) Barutan-seijin [21], b) Robot [22], c) Shiosai [23], d) Herman-grid [16], e) Grating induction [34], f) Bright room [48], g) Confetti illusion [39]. For more qualitative examples and results of classical visual illusions, see the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 112, + 179, + 351, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dedbf501dd0576c9f1ea5671119eecabd1142b9f163980bcb89b8ba2d4c674fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 180, + 588, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51e76b6dd54e200cebe3c730db8a0104eacf30199a5ac3fa1f4946e986581139.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 183, + 736, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99224664b909903de01abb5ea3eb175a10d1a33d96a8584e18ef82e4cda7f336.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 183, + 872, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance in replicating illusions. Classical models like ODOG and CIWAM, while physiologically grounded, show limited performance and flexibility. These methods require careful parameter tuning for each specific type of visual illusion, as evidenced by their poor scores. For instance, ODOG can achieve $72.44\\%$ accuracy on BRI3L but cannot generalize to color datasets, while applicable in color CIWAM shows poor performance. Previous deep learning approaches show consistent but modest performance in different datasets. Their uniform performance across sets suggests that they capture some general perceptual principles, but their relatively low accuracy indicates limitations in capturing human behavior. Our proposed method significantly outperforms all baseline approaches across all datasets.", + "bbox": [ + 88, + 353, + 482, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The substantial performance gap between our DDIM approach and a standard VAE (which shows near-zero performance on most datasets) highlights the crucial role of the diffusion process in capturing perceptual phenomena. These results demonstrate that our DDIM-based approach not only achieves state-of-the-art performance but also offers superior generalization across different types of visual illusions without requiring illusion-specific parameter tuning. The consistent high performance across all datasets suggests that the diffusion process naturally captures fundamental aspects of human vision.", + "bbox": [ + 89, + 565, + 482, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Generation of Visual Illusions", + "text_level": 1, + "bbox": [ + 89, + 739, + 356, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 5 demonstrates visual illusions using a $6 \\times 6$ pixel square target in the S1 stage of DeepFloyd (expanding to $24 \\times 24$ after S2). The figure displays five different target intensities (shown as circles to the left of each image) with different prompts for each image (selected randomly from DiffusionDB [56]). Additional results are presented in Fig. 1. In contrast to previous approaches, our method can generate complex content that simultaneously satisfies both the text prompt and the perceptual conditions.", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e23359fe3df8c5920a1e5075e834c26c92905dfe065a946e9bc9afc79d641a54.jpg", + "table_caption": [ + "Table 2. Perceptual Accuracy Score (PAS) in different VI sets" + ], + "table_footnote": [], + "table_body": "
MethodIllusionVQAGVILHallusionBench
CIWAM [40]56.5263.6383.33
DNet [12]14.2021.3033.30
RestoreNet [13]14.2021.3033.30
VAE4.7021.300.00
DDIM 5 steps90.4080.30100.00
DDIM 10 steps90.4096.70100.00
", + "bbox": [ + 517, + 369, + 898, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To see generation of classic VI (high frequency images with flat patches) please see supplementary material.", + "bbox": [ + 511, + 506, + 903, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Psychophysical confirmation. We created 40 images in which we enforced our model to provide an illusion with either red, green, blue, or achromatic targets. Similarly, we have also created 40 control images—images for which there was no enforcement for the patterns to look different. 15 observers participated in our study. The experiment was conducted in a dark room with controlled dim illumination. The only light source was the monitor, set to sRGB. All observers were tested for correct color vision using the Ishihara test. The screen showed one image at a time. Observers were asked to decide whether the two targets were identical or different. The order of images was randomized. We analyze the results in terms of the number of times an image was said to have different targets. For the images in which our model enforced the existence of the illusion, the observers said the targets were different in $64\\%$ (mean) and $67\\%$ (median) of the cases with a standard deviation of $16\\%$ . In contrast, only in $13\\%$ (mean) and $7\\%$ (median)—standard deviation $14\\%$ —of the cases where the control images said to have different targets.", + "bbox": [ + 511, + 537, + 906, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also perform an experiment to evaluate the effectiveness of our perceptual loss weight $(\\gamma)$ through comparison with the two classical models—ODOG and CIWAM—across different thresholds. As the value $\\gamma$ increases, there", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18648", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b90da7016e280bde492e7ef10c473a1db110911b1ff9499124cdbaccc335c016.jpg", + "image_caption": [ + "a)" + ], + "image_footnote": [], + "bbox": [ + 130, + 114, + 302, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94558c6caabfb266f3f1983bb4236c4db41f461828b2a9b1f4efaeca00ee2fad.jpg", + "image_caption": [ + "b)", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 356, + 108, + 491, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7dcf72aab58ce7703651a42eccb988e85b31bf1f0621b18b0e9f6c7afcde7f38.jpg", + "image_caption": [ + "c)" + ], + "image_footnote": [], + "bbox": [ + 555, + 109, + 651, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba44805ab708657f2bfde8df50edb8818410d56819d0cf23535a5036b00466c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 109, + 756, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a17e5614b74268d8beac3b1fadb18cc7c20a8bca073a0ec3dbc16d13d85158bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 109, + 862, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0d8af4dbe3c316c279970eff1f9c3ece53ffef40f9a6859798897fa302b7932.jpg", + "image_caption": [ + "d)" + ], + "image_footnote": [], + "bbox": [ + 148, + 200, + 300, + 319 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79e15d6cc0391949f8280451e47baa1cf0ad638ce6ec04b6f6e44b747c2b9204.jpg", + "image_caption": [ + "e)" + ], + "image_footnote": [], + "bbox": [ + 302, + 200, + 496, + 319 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d634b54bb14295b922c1bd55c885b77b00567ed0b18946e2eb122dbd3e81ea6.jpg", + "image_caption": [ + "f)" + ], + "image_footnote": [], + "bbox": [ + 537, + 200, + 691, + 319 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b2507d6a5a8567a8eab66fc5ab23ab329bbbcfa36cfa861033616ef789b6045.jpg", + "image_caption": [ + "g)" + ], + "image_footnote": [], + "bbox": [ + 735, + 200, + 887, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a6e088540b3c48f6d3ca69e2e968658dfec62704a3a2891c05bb497f69bb50b.jpg", + "image_caption": [ + "h)", + "Figure 5. Qualitative results of visual illusion generation (viewing at larger scale recommended). The first row presents prior work by (a) Gomez-Villa et al. [14], (b) Hirsch et al. [17], and (c) Roy et al. [46]. The second and third rows display results from our method (text prompts are available in the supplementary material). Images (d) through (i) use identical target colors (shown at the left of each image). In images (j) and (k), we employ grayscale gradients as targets. The selected target regions are highlighted in red in the thumbnail images." + ], + "image_footnote": [], + "bbox": [ + 107, + 321, + 300, + 440 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5558442cabcbbe120c0efb0c49a4676ef686da8f6ecbc372b6073053c3007e70.jpg", + "image_caption": [ + "i)" + ], + "image_footnote": [], + "bbox": [ + 302, + 321, + 495, + 440 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b61a592e854fc46fb2462ee136e3bda1cc1a3d985529b73a573cafed2ea6a779.jpg", + "image_caption": [ + "j)" + ], + "image_footnote": [], + "bbox": [ + 537, + 321, + 689, + 440 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65fee3c57a275af847063cec91b23fadd965f292ba585ac86d4fe51f52b750ed.jpg", + "image_caption": [ + "k)" + ], + "image_footnote": [], + "bbox": [ + 733, + 321, + 887, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is a substantial improvement in illusion strength, therefore showing the effectiveness of our perceptual loss.", + "bbox": [ + 89, + 510, + 482, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 89, + 556, + 217, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we make two key contributions. First, we report and demonstrate over a wide range of datasets an intriguing empirical observation: diffusion models naturally replicate human responses to brightness and color visual illusions through their image inversion process. This remarkable alignment between human perception and diffusion trajectories suggests that both systems may operate on similar statistical principles when processing ambiguous visual input. Second, leveraging this insight, we develop a method for generating visual illusions using T2I diffusion models, validated through psychophysical experiments. Looking ahead, our findings on perceptual processing in diffusion models could help bridge an important gap in current vision language models, which often struggle with nuanced visual understanding. The way diffusion models progressively process and interpret visual information might offer valuable insights for developing more perception-aware foundational models that better align with human visual cognition. This integration of perceptual understanding into image editing workflows represents an exciting direction for future research, potentially leading to tools that naturally", + "bbox": [ + 89, + 583, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "account for human visual perception in their operations.", + "bbox": [ + 511, + 510, + 882, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. (1) Replication. As expected from the image inversion process of diffusion models, the input images are spatially destroyed as the samples make their way to the Gaussian distribution. This implies that the target regions disappear for images closer to the natural image distribution, and the effect can no longer be measured. As a consequence our method has difficulty replicating the visual perception of effects which are based on small regions. (2) Generation. Our T2I VI generation method is based on defined regions during inference. This constraints the expressiveness of the T2I model in fulfilling the perceptual conditions. An ideal implementation should establish perceptual regions guided by the attention maps that the input prompt generates to produce more natural visual illusions.", + "bbox": [ + 511, + 527, + 905, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: We acknowledge projects PID2022-143257NB-I00, PID2021-128178OB-I00 PID2020-118071GB-I00 and PID2023-152133NB-I00 financed by MCIN/AEI/10.13039/501100011033 and ERDF \"A way of making Europe\", the CERCA Program and the Departament de Recerca i Universitats with ref. 2021SGR01499 from Generalitat de Catalunya, the grant Catedra ENIA UAB-Cruilla (TSI-100929-2023-2) from the Ministry of Economic Affairs and Digital Transition of Spain, and the BBVA Foundations of Science program (grant VIS4NN).", + "bbox": [ + 511, + 742, + 906, + 893 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Previous visual illusion generation methods", + "bbox": [ + 357, + 89, + 668, + 104 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18649", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arash Akbarinia, Yaniv Morgenstern, and Karl R Gegenfurtner. Contrast sensitivity function in deep networks. Neural Networks, 164:228-244, 2023. 2", + "[2] DeepFloyd Lab at StabilityAI. DeepFloyd IF: a novel state-of-the-art open-source text-to-image model with a high degree of photorealism and language understanding. https://www deepfloyd.ai/deepfloyd-if, 2023. Retrieved on 2023-11-08. 5", + "[3] Torsten Betz, Robert Shapley, Felix A Wichmann, and Marianne Maertens. Noise masking of white's illusion exposes the weakness of current spatial filtering models of lightness perception. Journal of vision, 15(14):1-1, 2015. 5", + "[4] Barbara Blakeslee and Mark E McCourt. A multiscale spatial filtering account of the white effect, simultaneous brightness contrast and grating induction. Vision research, 39(26): 4361-4377, 1999. 5, 6", + "[5] E. Brucke. über erganzungs und contrasfarben. Wiener Sitzungsbser, 51, 1865. 3, 4", + "[6] Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael S Ryoo. Diffusion illusions: Hiding images in plain sight. arXiv preprint arXiv:2312.03817, 2023. 2, 3, 4", + "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 2023. 5", + "[8] David Corney and R Beau Lotto. What are lightness illusions and why do we see them? PLoS computational biology, 3 (9):e180, 2007. 2", + "[9] Karl R Gegenfurtner, Marina Bloj, and Matteo Toscani. The many colours of 'the dress'. Current Biology, 25(13):R543–R544, 2015. 1", + "[10] Daniel Geng, Inbum Park, and Andrew Owens. Visual analogs: Generating multi-view optical illusions with diffusion models. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3, 4", + "[11] Daniel Geng, Inbum Park, and Andrew Owens. Factorized diffusion: Perceptual illusions by noise decomposition. arXiv preprint arXiv:2404.11615, 2024. 3, 4", + "[12] Alex Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmó. Convolutional neural networks can be deceived by visual illusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309-12317, 2019. 2, 6, 7", + "[13] Alex Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, Marcelo Bertalmó, and Jesús Malo. Color illusions also deceive cnns for low-level vision tasks: Analysis and implications. Vision Research, 176:156-174, 2020. 2, 6, 7", + "[14] A. Gomez-Villa, A. Martin, J. Vazquez-Corral, M. Bertalmio, and J. Malo. On the synthesis of visual illusions using deep generative models. Journal of Vision, 22(8):2-2, 2022. 3, 8", + "[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5", + "[16] Ludimar Hermann. Eine erscheinung simultanen contrastes. Archiv für die gesamte Physiologie des Menschen und der Tiere, 3(1):13-15, 1870. 7", + "[17] Elad Hirsch and Ayellet Tal. Color visual illusions: A statistics-based computational model. Advances in neural information processing systems, 33:9447-9458, 2020. 2, 8", + "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2022. 3", + "[19] Been Kim, Emily Reif, Martin Wattenberg, and Samy Bengio. Do neural networks show gestalt phenomena? an exploration of the law of closure. arXiv preprint arXiv:1903.01069, 2(8), 2019. 2", + "[20] Frederick AA Kingdom. Lightness, brightness and transparency: A quarter century of new ideas, captivating demonstrations and unrelenting controversy. Vision Research, 51 (7):652-673, 2011. 3", + "[21] Akiyoshi Kitaoka. Barutan-seijin. Akiyoshi's illusion Website, 2003. Accessed: 2024-11-13. 7", + "[22] Akiyoshi Kitaoka. Robot. Akiyoshi's illusion Website, 2005. Accessed: 2024-11-13. 7", + "[23] Akiyoshi Kitaoka. Shiosai. Akiyoshi's illusion Website, 2008. Accessed: 2024-11-13. 7", + "[24] Yuki Kubota, Atsushi Hiyama, and Masahiko Inami. A machine learning model perceiving brightness optical illusions: Quantitative evaluation with psychophysical data. In Proceedings of the Augmented Humans International Conference 2021, pages 174-182, 2021. 2", + "[25] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[26] V. Laparra and J. Malo. Visual aftereffects and sensory nonlinearities from a single statistical framework. Front. Human Neurosci., 9, 2015. 2", + "[27] V. Laparra, G. Camps, and J. Malo. Iterative gaussianization: From ICA to random rotations. IEEE Trans. Neur. Nets, 22 (4):537-549, 2011. 2", + "[28] V. Laparra, S. Jiménez, G. Camps, and J. Malo. Nonlinearities and adaptation of color vision from sequential principal curves analysis. Neural Comp., 24(10):2751-2788, 2012. 2", + "[29] V. Laparra, JE. Johnson, G. Camps, R. Santos, and J. Malo. Estimating information theoretic measures via multidimensional gaussianization. IEEE Trans. Patt. Anal. Mach. Intell., 2024. 2", + "[30] Qiang Li, Alex Gomez-Villa, Marcelo Bertalmó, and Jesús Malo. Contrast sensitivity functions in autoencoders. Journal of Vision, 22(6):8-8, 2022. 2, 6", + "[31] Senmao Li, Joost van de Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing, 2023. 3", + "[32] Tingting Li, Fanyu Wang, Ying Zhou, and Zhenping Xie. Visual illusion cognition dataset construction and recognition" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18650", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "performance by deep neural networks. In 2022 IEEE 8th International Conference on Cloud Computing and Intelligent Systems (CCIS), pages 90-94. IEEE, 2022. 2", + "[33] Ben Lonnqvist, Alban Bornet, Adrien Doerig, and Michael H Herzog. A comparative biology approach to dnn modeling of vision: A focus on differences, not similarities. Journal of vision, 21(10):17-17, 2021. 2", + "[34] Mark McCourt. Grating induction: A new explanation for stationary phantom gratings. Vision Research, 34(12):1609-1617, 1994. 7", + "[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 3", + "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3", + "[37] Chaithanya Kumar Mummadi, Ranjitha Subramaniam, Robin Hutmacher, Julien Vitay, Volker Fischer, and Jan Hendrik Metzen. Does enhanced shape bias improve neural network robustness to common corruptions? International Conference on Learning Representations, 2021. 2", + "[38] Jerry Ngo, Swami Sankaranarayanan, and Phillip Isola. Is clip fooled by optical illusions? International Conference on Learning Representations, 2023. 2, 4", + "[39] David Novick and Akiyoshi Kitaoka. The confetti illusion. Journal of Illusion, 2, 2021. 7", + "[40] Xavier Otazu, C Alejandro Parraga, and Maria Vanrell. Toward a unified chromatic induction model. Journal of Vision, 10(12):5-5, 2010. 5, 6, 7", + "[41] Jiang Qin, Kai Wang, Bin Zou, Lamei Zhang, and Joost van de Weijer. Conditional diffusion model with spatial-frequency refinement for sar-to-optical image translation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3", + "[42] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2023. 5", + "[43] Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. International Conference on Machine Learning, 2015. 2", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[45] Aniket Roy, Anirban Roy, Soma Mitra, and Kuntal Ghosh. Bri3l: A brightness illusion image dataset for identification and localization of regions of illusory perception. Proceedings of the IEEE International Conference on Image Processing, 2023. 5", + "[46] Aniket Roy, Anirban Roy, Soma Mitra, and Kuntal Ghosh. Bri3l: A brightness illusion image dataset for identification" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and localization of regions of illusory perception. Proceedings of the IEEE International Conference on Image Processing, 2024. 3, 8", + "[47] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. Illusionvqa: A challenging optical illusion dataset for vision language models. First Conference on Language Modeling, 2024. 2, 5", + "[48] Arthur G. Shapiro and Dejan Todorovic. The Oxford Compendium of Visual Illusions, chapter 22. Oxford University Press, 2017. Figure 3.2. 7", + "[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3", + "[50] Eric D Sun and Ron Dekel. Imagenet-trained deep neural networks exhibit illusion-like response to the scintillating grid. Journal of Vision, 21(11):15-15, 2021. 2", + "[51] Chuanming Tang, Kai Wang, and Joost van de Weijer. Iterinv: Iterative inversion for pixel-level t2i models. Neurips 2023 workshop on Diffusion Models, 2023. 3", + "[52] Oguzhan Ulucan, Diclehan Ulucan, and Marc Ebner. Investigating color illusions from the perspective of computational color constancy. arXiv preprint arXiv:2312.13114, 2023. 2", + "[53] Clinton Wang and Polina Golland. Interpolating between images with diffusion models. ICML 2023 Workshop DeployableGenerativeAI, 2023. 3", + "[54] Kai Wang, Fei Yang, Shiqi Yang, Muhammad Atif Butt, and Joost van de Weijer. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 2023. 3", + "[55] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 5", + "[56] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. Annual Meeting of the Association for Computational Linguistics, 2023. 7", + "[57] Eiji Watanabe, Akiyoshi Kitaoka, Kiwako Sakamoto, Masaki Yasugi, and Kenta Tanaka. Illusory motion reproduced by deep neural networks trained for prediction. Frontiers in psychology, 9:340023, 2018. 2", + "[58] Robert Max Williams and Roman V Yampolskiy. Optical illusions images dataset. arXiv preprint arXiv:1810.00415, 2, 2018. 2, 4", + "[59] Hongtao Zhang, Zhen Li, and Shinichi Yoshida. Muller-layer illusion is replicated by higher layer of pre-trained deep neural network for object recognition. International Symposium on Computational Intelligence and Industrial Applications, 2022. 2", + "[60] Yichi Zhang, Jiayi Pan, Yuchen Zhou, Rui Pan, and Joyce Chai. Grounding visual illusions in language: Do vision-language models perceive illusions like humans? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5718-5728, 2023. 2, 4, 5" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18651", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[61] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Tiviatis Sim, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. European Conference on Computer Vision, 2024. 5", + "bbox": [ + 91, + 90, + 482, + 147 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18652", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/32f81902-9541-4c54-8edb-2ca63c430edb_model.json b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/32f81902-9541-4c54-8edb-2ca63c430edb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..87a80bfed2c249d907889e7bfe1f3e09cfa71273 --- /dev/null +++ b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/32f81902-9541-4c54-8edb-2ca63c430edb_model.json @@ -0,0 +1,2851 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.032 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "The Art of Deception: Color Visual Illusions and Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.18, + 0.406, + 0.198 + ], + "angle": 0, + "content": "Alexandra Gomez-Villa\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.199, + 0.453, + 0.217 + ], + "angle": 0, + "content": "Bartłomiej Twardowski1,2,4 Jesus Malo3" + }, + { + "type": "text", + "bbox": [ + 0.458, + 0.181, + 0.801, + 0.199 + ], + "angle": 0, + "content": "Kai Wang\\(^{1}\\) C.Alejandro Parraga\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.489, + 0.199, + 0.895, + 0.218 + ], + "angle": 0, + "content": "Javier Vazquez-Corral\\(^{1,2}\\) Joost van de Weijer\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.233, + 0.828, + 0.252 + ], + "angle": 0, + "content": "1Computer Vision Center, Spain 2Universitat Autònoma de Barcelona, Spain" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.252, + 0.746, + 0.27 + ], + "angle": 0, + "content": "3Universitat de València, Spain 4IDEAS-NCBR, Poland" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.281, + 0.357, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.281, + 0.613, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.283, + 0.871, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.411, + 0.361, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.412, + 0.616, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.411, + 0.895, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.564, + 0.907, + 0.634 + ], + "angle": 0, + "content": "Figure 1. Observation: Denoising Diffusion Implicit Models (DDIM) have human-like visual illusions. Application: text-to-image generated visual illusions. Examples (a) and (b) show how the responses to physically equal patches shift differently in the path to the latent space. In contrast, (c) shows how physically different stimuli get more similar along the path, achieving color constancy. The observation of this human-like behavior allows us to propose the use of text-to-image models to generate images in which physically identical patches are perceived differently (examples d, e, f). We highly recommend to watch the illusions on a computer screen." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.643, + 0.327, + 0.658 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.675, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Visual illusions in humans arise when interpreting out-of-distribution stimuli: if the observer is adapted to certain statistics, perception of outliers deviates from reality. Recent studies have shown that artificial neural networks (ANNs) can also be deceived by visual illusions. This revelation raises profound questions about the nature of visual information. Why are two independent systems, both human brains and ANNs, susceptible to the same illusions? Should any ANN be capable of perceiving visual illusions? Are these perceptions a feature or a flaw? In this work, we study how visual illusions are encoded in diffusion models. Remarkably, we show that they present human-like brightness/color shifts in their latent space. We use this fact to demonstrate that diffusion models can predict visual illusions. Furthermore, we also show how to gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.705 + ], + "angle": 0, + "content": "erate new unseen visual illusions in realistic images using text-to-image diffusion models. We validate this ability through psychophysical experiments that show how our model-generated illusions also fool humans." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.718, + 0.646, + 0.734 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.907, + 0.851 + ], + "angle": 0, + "content": "Visual illusions (VIs) serve as a fascinating window into the complex workings of human visual perception, where our interpretation of reality diverges significantly from physical measurements. These phenomena can have a profound social impact, as demonstrated by viral phenomena such as \"The Dress\" [9], which sparked global discussions on the nature of perception. Beyond their intrinsic scientific value," + }, + { + "type": "page_footnote", + "bbox": [ + 0.538, + 0.863, + 0.858, + 0.877 + ], + "angle": 0, + "content": "Corresponding Author (☑): Kai Wang (kwang@cvc.uab.es)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.877, + 0.904, + 0.899 + ], + "angle": 0, + "content": "Code is available at https://alviur.github.io/color-illusion-diffusion" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.863, + 0.904, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18642" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "VIs offer a unique lens through which we can examine and compare human and machine perception [13, 33]. The emergence of diffusion models (DMs) as state-of-the-art generative models has revolutionized the field of artificial intelligence, demonstrating unprecedented capabilities in image synthesis and manipulation. Their denoising-based architecture, which gradually transforms random noise into coherent images, presents a novel paradigm for studying visual processing that mirrors the lossy and noisy process of human perception. This parallel makes them particularly interesting for investigating VIs, as both systems must navigate the complex relationship between image statistics and perceptual interpretation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.483, + 0.504 + ], + "angle": 0, + "content": "The study of VIs in artificial neural networks (ANNs) has a rich history. Early work demonstrated that a multilayer perceptron could replicate brightness illusions [8], while subsequent research by Gomez-Villa et al. [12] showed that convolutional neural networks (CNNs) could mirror human perception in lightness and color illusions. Concurrent developments included Watanabe et al.'s [57] demonstration of motion VI replication in video-prediction CNNs and the publication of the first visual illusion dataset [58]. This foundation has spawned diverse investigations into the ANN-VI relationship in multiple domains: image completion [19, 50], brightness/color perception [13, 17, 24, 32, 33, 52], contrast sensitivity [1, 30], shape perception [37, 59], and vision-language models [6, 38, 47, 60]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this work, we investigate how human-like perceptual phenomena naturally emerge in the encoding process of DMs. Our key finding reveals that the intermediate steps of diffusion models exhibit trends that align remarkably with human perception. Specifically, when analyzing the trajectory of an inverted image through the diffusion process, we observe that the progressive changes in lightness and color properties closely mirror human perceptual responses. Prior research has demonstrated that Gaussianization flows exhibit interesting properties when processing samples that deviate from their training data statistics. When trained on specific sample distributions, these flows can produce unusual representations for out-of-distribution samples [27, 43] - a characteristic that has proven valuable for measuring distribution divergence [29]. Intriguingly, this phenomenon parallels human visual perception, where shifts in image statistics can lead to illusory perception of spatio-chromatic patterns, a behavior that has been explained through manifold equalization [26, 28]. We argue that image samples that do not follow the statistics of natural images are not properly gaussianized by DMs. In this way, linking defective gaussianization and illusory perception of unusual samples, it makes sense that intermediate representations in DMs may be correlated with the shifts of these samples for human observers. The reported alignment suggests that the iterative denoising in diffusion models may" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "inadvertently capture fundamental aspects of human visual perception, particularly in contexts where our perception diverges from physical reality. By examining these intermediate representations, we provide novel insights into the nature of visual illusions and the unexpected parallels between human perception and the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.89, + 0.197 + ], + "angle": 0, + "content": "Contributions: The main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.272 + ], + "angle": 0, + "content": "- We report an empirical observation of the behavior of diffusion models: progressive changes in lightness, hue and saturation of image regions in the forward/inverse transform to the latent representation replicate the shifts of human perception of these regions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.273, + 0.905, + 0.348 + ], + "angle": 0, + "content": "- We derive a novel approach to using diffusion models to model human vision. We show that deep autoencoders can be effectively employed as vision models by performing a diffusion inversion process and measuring perception in the intermediate noise space." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.349, + 0.905, + 0.393 + ], + "angle": 0, + "content": "- We demonstrate the human perception replication capabilities of DM in many cases, including VI datasets, popular published VI works, and natural images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.904, + 0.438 + ], + "angle": 0, + "content": "- Leveraging the perceptual capabilities of diffusion models, we propose a novel method for generating brightness/color VIs using text-to-image models." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.451, + 0.652, + 0.467 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.905, + 0.524 + ], + "angle": 0, + "content": "Replication of Visual Illusions. Replication of visual illusions has been made with several methods, but all of them can be roughly classified into the following categories:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.524, + 0.905, + 0.569 + ], + "angle": 0, + "content": "- Image restoration methods [12, 13, 24, 30, 33]: These use pre-trained image restoration backbones. They process VIs and measure the output image to measure perception." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.569, + 0.905, + 0.643 + ], + "angle": 0, + "content": "- Effect bias methods [1]: These works are inspired by neuroscience measurement of stimuli. Pretrained architectures are used to process VIs, then the response in the internal activations of the ANN is used to define if there is a visual illusion." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.645, + 0.905, + 0.689 + ], + "angle": 0, + "content": "- Likelihood method [17]: In this approach, the most probable perceived intensity of each pixel in an image is estimated using an invertible flow trained on natural images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.69, + 0.905, + 0.75 + ], + "angle": 0, + "content": "- Language-guided methods [38, 60]: Language is used as a vehicle to measure perception. These works use vision-language models as test subjects and ask them questions regarding the perception of the image." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.524, + 0.905, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Our approach differs from previous methods in the following: 1) we do not measure perception in the output of the model. Instead, we measure the VI in the intermediate steps of image inversion process. 2) We use a deep autoencoder; previous approaches in this category claim that deep autoencoders suffer from less perceptual bias and use shallow architectures for replication experiments. Furthermore, we are able to replicate human perception in high resolution images (as big as DMs are able to manage) and in natural images. These setups were not possible in previous works." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18643" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.41 + ], + "angle": 0, + "content": "Generation of Visual Illusions. Although generative models have been used to synthesize visual illusions [14], DMs has only recently been proposed as the generative engine. In [6], high-level illusions (overlay effects) are generated by mixing a set of images and transformations together. Following works [10, 11] were able to create more general illusions such as rotations, color inversions, and hybrid images. The work most similar to ours is by Roy et al. [46], who used stable diffusion (SD) to generate brightness visual illusions through keyword prompts. However, their approach neither explicitly optimizes perceptual effects nor specifies regions for perception replication. Consequently, illusions may appear randomly in the generated image with unknown localization. Additionally, their generated images are limited to conventional brightness illusions (gratings or high-frequency black and white patterns) and cannot incorporate visual illusions into natural images. In contrast, our work allows visual illusions within generated natural images as well as in typical high-frequency patterns. We explicitly generate the perceptual effect in a designated image region and can control the image content through text prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.192, + 0.44 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.45, + 0.258, + 0.465 + ], + "angle": 0, + "content": "3.1. Diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Diffusion models (DMs) are generative models that learn to gradually denoise images by reversing a fixed Markov chain of forward diffusion steps. The forward process systematically adds Gaussian noise to an image until it becomes pure noise, while the learned reverse process reconstructs the original image through iterative denoising. While traditional diffusion follows a stochastic sampling process, Denoising Diffusion Implicit Models (DDIM) [49] introduce a deterministic sampling approach that maintains high quality while enabling significantly fewer inference steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.748 + ], + "angle": 0, + "content": "DDIM's deterministic nature enables a crucial capability: the ability to invert generated images back into their latent representations through a process known as DDIM inversion [49]. This inversion process allows us to map any image into the model's latent space and observe its trajectory through the diffusion process. Given an image \\( z_0 \\), DDIM inversion produces a sequence of intermediate representations \\( \\{z_t\\}_{t=1}^T \\) according to:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.759, + 0.483, + 0.78 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {z _ {t}, t} (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.792, + 0.484, + 0.839 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\) represents the cumulative product of noise scheduling coefficients, and \\(\\epsilon_{\\theta}^{z_t,t} = \\epsilon_{\\theta}(z_t,t)\\) is the learned noise prediction network. The reverse process follows:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.848, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(\\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {z _ {t} , t}}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\epsilon_ {\\theta} ^ {z _ {t}, t} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "DDIM inversion has emerged as a powerful tool for various image manipulation tasks. Its ability to faithfully reconstruct latent representations has enabled applications such as semantic image editing [31, 35, 36, 51, 54], style transfer [25, 41], image interpolation [53], and controlled image generation through classifier-free guidance [18]. In our work, we exploit the deterministic nature of DDIM inversion to systematically study how visual illusions are processed through the diffusion trajectory." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.239, + 0.892, + 0.256 + ], + "angle": 0, + "content": "3.2. DDIM inversion may replicate human vision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.907, + 0.368 + ], + "angle": 0, + "content": "The wholly-empirical evolutionary paradigm [20] suggests that human vision is shaped by statistical regularities in natural scenes, where our visual system evolved to interpret ambiguous stimuli based on their likelihood. We leverage this insight to study how diffusion models, trained for natural images, process illusory stimuli, which are not in the distribution of natural images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.909, + 0.641 + ], + "angle": 0, + "content": "Motivation. In Fig. 2, we present an illustrative experiment using Stable Diffusion [44] (extra results with an image space model in the supplementary material). We perform DDIM inversion on the canonical visual illusion known as Brightness Contrast [5]. DDIM inversion puts the image back into the latent space. This illusion presents two gray squares of identical intensity (indicated by red markers) against white and black backgrounds, which leads humans to perceive the left square as darker than the right one. The first row (left to right) shows the inversion1 for 0, 3, 10, and 20 steps, respectively. As the zoomed-in bar shows, the initially equal-intensity regions progressively develop an intensity difference that aligns with human perception2 (left target appearing darker than right). The second row shows the progressive Gaussianization of the latent representation. For this image perception is replicated at an intermediate representation where gaussianization is not complete (since this stimulus is not in the distribution of natural images)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.763 + ], + "angle": 0, + "content": "Observation: emergence of perceptual biases. Our key observation is: when inverting an image through diffusion models the intermediate denoising steps progressively introduce color/brightness shifts that align with human perception. We hypothesize that this occurs because diffusion models, trained on natural images, map illusory stimuli onto the learned natural image manifold, resulting in systematic \"interpretation errors\" that mirror human perceptual biases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Theoretical framework. Formally, let \\( G: \\mathcal{X} \\to \\mathcal{Z} \\) be the forward diffusion mapping from image space \\( \\mathcal{X} \\) to latent space \\( \\mathcal{Z} \\). For natural images \\( x \\in \\mathcal{X} \\), \\( G \\) learns a mapping \\( z = G(x) \\) that transforms the image distribution into a Gaussian. However, when presented with an eventually" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.851, + 0.884, + 0.864 + ], + "angle": 0, + "content": "we decode the latent representation to visualize it in image space." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.864, + 0.906, + 0.901 + ], + "angle": 0, + "content": "2In addition to the color/brightness effects, notice how the high-frequency content of the image is maintained; we explore this phenomenon further in supplementary material." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.851, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18644" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.895, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.908, + 0.369 + ], + "angle": 0, + "content": "Figure 2. DDIM inversion of the Brightness Contrast illusion [5] using Stable Diffusion. Top row: Image-space visualization (decoded latents) showing (left) Original illusion with two identical gray squares (marked in red) against black and white backgrounds, and inversion results using 3, 10, and 20 steps. Bottom row: Histograms of the corresponding latent representations. The horizontal axis represents the value of the latent space feature. The model gradually reproduces the perceptual difference in brightness between the physically identical squares in a (not fully Gaussian) intermediate representation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.375, + 0.485, + 0.618 + ], + "angle": 0, + "content": "illusory stimulus \\( y \\), we have \\( y \\notin \\mathcal{X} \\). Still, the mapping \\( z' = G(y) \\) attempts to project the illusion onto the learned natural image manifold, introducing systematic deviations. The intermediate states \\( z_{t} \\) reveal how the model progressively \"corrects\" the illusion toward natural image statistics. We observe that these corrections correlate strongly with human perception, suggesting that both humans and diffusion models may operate on similar statistical principles when processing ambiguous visual input. This alignment between diffusion trajectories and human perception provides computational evidence for the wholly-empirical paradigm, demonstrating how statistical learning can lead to perceptual illusions without explicit encoding of perceptual rules. The systematic nature of these \"errors\" suggests they emerge from the same statistical regularities that shape human visual processing through evolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.632, + 0.411, + 0.647 + ], + "angle": 0, + "content": "3.3. Generating visual illusions with DMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Based on the findings above, we develop a method to automatically generate brightness/color illusions with DMs. To do so, we need to evaluate whether a VI is present and then push the diffusion process toward the desired perception." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Previous works on VIs and DMs focus on high-level visual illusions, i.e. cognitive effects that measure if the image belongs to a certain class of prompt [6, 10, 11]. Here we focus on brightness/color VIs, which leads to the challenge of measuring effects in specific regions of the image. Here we answer the measurement questions proposed in previous works on brightness/color VI generation in this way:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Where to measure? This question separates the generation of high-level VIs from color/brightness VIs. Models aiming at high-level VIs that fool an ANN require to ask a classifier/vision-language model which class/prompt better describes the image [10, 38, 60]. In contrast, color VIs ap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.48 + ], + "angle": 0, + "content": "pear in specific regions of the image, therefore, they need to be measured at those points. In this work, we specifically define regions (here targets) in which we want a perceptual effect to appear. These targets are our measurement points, and we guarantee that in these regions, we can track, measure, and optimize perceptual effects. Failing to define a target leads to failure, as reported in [58]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.906, + 0.572 + ], + "angle": 0, + "content": "What to measure? During image generation, we will measure the intensity values of the intermediate steps \\( z_{t} \\). Even though \\( z_{t} \\) does not contain the final perception bias \\( (z_{T}) \\), we can push the generative process towards a desired perceptual effect using small changes in the intermediate representation as explained in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.724 + ], + "angle": 0, + "content": "How to measure? In the motivating example of section 3.2 (Fig. 2), the model replicates human perception if the target shifts in the same direction as humans perceive. However, in an optimization procedure, we need to measure how \"good\" the illusion is; in other words, we need to quantify the effect. Since we are concerned about brightness and color VI, we will measure the mean intensity per channel of each target. More precisely, given an image \\(I\\), and the target region \\(r \\in \\mathbb{R}^2\\) in \\(I\\), we compute the mean intensity in that target region as:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.734, + 0.907, + 0.77 + ], + "angle": 0, + "content": "\\[\nm _ {i n t} (I, r) = \\frac {1}{M} \\sum_ {i \\in r} I \\left(r _ {i}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.782, + 0.907, + 0.845 + ], + "angle": 0, + "content": "where \\(r_i\\) denotes a specific pixel \\(i\\) in target region \\(r\\), \\(M\\) stands for the number of pixels in the target region, and \\(m_{int}\\) is measured in each channel. Hence, the perceived intensity is an RGB triplet \\((m_{int}^R, m_{int}^G, m_{int}^B)\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.851, + 0.868, + 0.867 + ], + "angle": 0, + "content": "3.3.1. Visual perception inside generation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Since text-to-image (T2I) diffusion models often struggle to perfectly align with textual prompts, a variety of aug-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18645" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.47, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.306, + 0.483, + 0.389 + ], + "angle": 0, + "content": "Figure 3. Overview of our visual illusion generation pipeline. The process modifies noisy latent representations, \\( z_{t} \\), through a custom loss function, \\( \\mathcal{L} \\), that guides the generation toward perceptually ambiguous outputs. Note that targets are regions that manually selected by the user as inputs. The terms \\( D_{p1} \\) and \\( D_{p2} \\) are the user desired intensity for target T1 and T2 respectively." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.395, + 0.483, + 0.667 + ], + "angle": 0, + "content": "mentation methods have been developed to enhance the fidelity of T2I generation. These methods typically focus on optimizing noisy latent representations in each timestep, a strategy that has proven effective in refining generation outcomes [7, 42, 55, 61]. A common approach involves the design of an update loss function, \\(\\mathcal{L}\\), which is tailored based on the layout of the noisy latent or the characteristics of the cross-attention maps. This loss is then used to guide the backpropagation process, updating the noisy latent in a manner that more accurately reflects the intended textual descriptions, thereby enhancing the overall quality of the generated images. This process can be formulated as \\(z_{t}^{\\prime} = z_{t} - \\alpha \\cdot \\nabla_{z_{t}}\\mathcal{L}\\). In this paper, we diverge from the conventional use of optimization techniques aimed at aligning T2I diffusion models with textual semantics. Instead, we leverage these techniques to create visual perception effects (see Fig. 3). We use an image space DM (DeepFloyd [2]) to avoid the use of the decoder in the backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.833 + ], + "angle": 0, + "content": "First, we define the original target image as \\( O(r) \\), where \\( r \\in \\mathbb{R}^2 \\). This image only contains the input-defined targets, that is, \\( O(r) \\) has the original value given to the targets by the user on each color channel. Then, given an intermediate image \\( z_t \\) and a list of \\( R \\) target regions \\( \\{r^p\\}_{p=1}^R \\), our process to generate a VI includes the addition of the targets to the image in the desired region3, i.e. we compute \\( z_{t\\_target}(r_i^p) = z_t(r_i^p) + O(r_i^p) \\), for all pixels \\( i \\) inside the region \\( r^p \\) and \\( z_{t\\_target} = z_t \\) for any pixel outside this target region. Finally, the following optimization term is applied \\( N \\) times —per diffusion step \\( t \\) — to guide the generation:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.841, + 0.482, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\gamma \\mathcal {L} _ {V I} + \\beta \\mathcal {L} _ {s i m} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.865, + 0.483, + 0.88 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{VI}\\) and \\(\\mathcal{L}_{sim}\\) measure the illusion strength and the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.466, + 0.901 + ], + "angle": 0, + "content": "3Note that this addition results in the perceived color (see sect. 3.2)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "similarity between targets and background respectively, and \\(\\gamma\\) and \\(\\beta\\) are hyperparameters that control the relative weight of these measures. Specifically:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.141, + 0.85, + 0.157 + ], + "angle": 0, + "content": "- \\(\\mathcal{L}_{VI}\\) is a perceptual loss on the target intensities:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.169, + 0.906, + 0.212 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {V I} = \\sum_ {c \\in R, G, B} \\sum_ {r ^ {p}} ^ {R} \\mathcal {L} _ {\\text {P e r c e p t}} \\left(z _ {t _ {\\text {t a r g e t}}}, k _ {c} ^ {p}; r ^ {p}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.224, + 0.906, + 0.254 + ], + "angle": 0, + "content": "where \\(r^p\\) is a specific target, \\(k_c^p\\) is the desired intensity for target \\(r^p\\) on color channel \\(c\\), and" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.268, + 0.905, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {P e r c e p t u a l}} \\left(z _ {t _ {\\text {t a r g e t}}}, k _ {c}; r\\right) = \\left\\| m _ {\\text {i n t}} \\left(z _ {t _ {\\text {t a r g e t}}}, r\\right) - k _ {c} \\right\\| \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.3, + 0.728, + 0.314 + ], + "angle": 0, + "content": "with \\(m_{int}\\) as defined in Eq. 3." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.315, + 0.906, + 0.374 + ], + "angle": 0, + "content": "- \\( \\mathcal{L}_{sim} \\) is a similarity term. This term aims at avoiding the region from appearing \"out of place\". In particular, it looks at harmonizing the targets with the background-generated image:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.387, + 0.906, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s i m} = \\sum_ {p = 1} ^ {R} M S E \\left(z _ {t} \\left(r ^ {p}\\right), O \\left(r ^ {p}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.442, + 0.704, + 0.46 + ], + "angle": 0, + "content": "4. Experimental setup" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.469, + 0.616, + 0.482 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.49, + 0.906, + 0.55 + ], + "angle": 0, + "content": "- Brightness illusions dataset: The \\( BRI3L \\) dataset [45] consists of 22366 grayscale images spanning five types of brightness illusions. Each image is annotated with binary masks indicating regions of illusory perception." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.55, + 0.906, + 0.747 + ], + "angle": 0, + "content": "- Visual Illusion VQA Datasets: Several datasets have been developed to evaluate vision-language models' understanding of VIs: IllusionVQA [47] contains 435 illusory images with multiple-choice questions across 12 categories, plus 1000 impossible object image pairs for localization testing. GVIL [60] features 100 base illusion images expanded to 1600 instances covering five illusion categories, with four evaluation tasks: same-difference QA, referential QA, attribute QA, and referential localization. HallusionBench [15] comprises 346 images and 1129 questions designed to test image-context reasoning and diagnose both VIs and language hallucinations. We use the color category in all the VQA datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.758, + 0.647, + 0.772 + ], + "angle": 0, + "content": "4.2. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "- Classical vision science models: These approaches are based in vision science and need to be calibrated for each image to fully exploit their capabilities. In this paper we use the parameters set by Betz et al. [3]. These parameters constitute a fair guideline for these methods, as it is not possible to fine-tune them for thousands of images. We use ODOG (Oriented Difference of Gaussians) [4] and the Chromatic Induction Wavelet Model (CIWaM) [40]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18646" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.48, + 0.136 + ], + "angle": 0, + "content": "- Neural networks-base models: We compare against previous image restoration methods that proven ability in visual illusion detection: DN-NET[13] and Restorenet[30]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.137, + 0.48, + 0.197 + ], + "angle": 0, + "content": "- Variational autoencoder (VAE): We present the results of only considering the VAE of SD prior to the DDIM inversion process —to further demonstrate DDIM is the responsible for the alignment with human perception." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.48, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.208, + 0.187, + 0.223 + ], + "angle": 0, + "content": "4.3. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.481, + 0.32 + ], + "angle": 0, + "content": "While visual illusion replication results have traditionally been presented through profile plots comparing model outputs with human perception along image cross-sections, we leverage recent datasets to provide quantitative metrics. This approach complements classical visualization methods with rigorous numerical evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.482, + 0.368 + ], + "angle": 0, + "content": "For the BRI3L dataset, we compute the mean intensity difference, \\(\\Delta I\\), between the model's output and the input image within each segmented region:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.379, + 0.482, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\Delta I = \\frac {1}{| M |} \\sum_ {i \\in r} \\left| I _ {\\text {o u t}} (r) - I _ {\\text {i n}} (r) \\right| \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.483, + 0.562 + ], + "angle": 0, + "content": "where \\(M\\) represents the size of the segmentation mask, and \\(I_{out}\\), \\(I_{in}\\) are the output and input intensities respectively. We consider the model's response to be aligned with human perception when \\(\\Delta I < \\tau * I_{int}\\), where \\(I_{int}\\) represents the mean intensity of the input image in the target region (calculated using Eq. 3). Since BRI3L aims to detect regions that become darker, the metric becomes more harder as the value of \\(\\tau\\) decreases, requiring the model to achieve greater darkness in the target regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.483, + 0.669 + ], + "angle": 0, + "content": "For the VQA-based datasets (IllusionVQA, GVIL, and HallusionBench), we manually create segmentation masks and consider the answer of the model aligned with human perception if it matches the expected intensity for regions referenced in the question-answer pairs. The Perception Accuracy Score (PAS) is the percentage of images having the correct perceptual predictions in each dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.226, + 0.695 + ], + "angle": 0, + "content": "4.4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.793 + ], + "angle": 0, + "content": "We conducted two sets of experiments to evaluate our method's ability to: (1) replicate existing VIs across multiple datasets, and (2) generate novel VIs that can fool human perception. For each experiment, we present qualitative results zooming in on the region of the perceptual effect and quantitative evaluations using the proposed metrics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Replication of visual illusions: We test our model's ability to replicate brightness (BRI3L) and color (Illusion-VQA, GVIL, and HallusionBench) illusions by comparing the intensity differences between input and output images within annotated segmentation masks. For BRI3L, we evaluate performance across three \\(\\tau\\) levels (0.8, 0.9, 1.0) to assess the robustness of illusion replication. In this" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.117 + ], + "angle": 0, + "content": "Table 1. Perceptual Accuracy Score (PAS) in BRI3L for different methods and thresholds" + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.122, + 0.875, + 0.272 + ], + "angle": 0, + "content": "
Method / Threshold0.80.91.0
ODOG [4]36.5857.0972.44
CIWAM [40]5.6918.2356.22
DNnet [12]0.000.0071.09
RestoreNet [13]0.000.0074.95
VAE0.000.0873.18
DDIM 5 steps44.0291.4399.88
DDIM 10 steps84.1798.85100.00
" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.291, + 0.905, + 0.337 + ], + "angle": 0, + "content": "experiment we present results using SD 1.4 (latent space DM) but additional results with DeepFloyd (image space DM) are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.337, + 0.905, + 0.488 + ], + "angle": 0, + "content": "- Generation of visual illusions: In this experiment, we check whether we can generate VIs in images that fall inside the distribution of the T2I model. We use guidance scale \\(= 10\\) (we want to generate images that follow the prompt), \\(N = 5\\), \\(\\gamma = 0.5\\), and \\(\\beta = 1.0\\). Additionally, here we use S2 of DeepFloyd to generate higher-resolution versions of our generated visual illusions. For flat intensity targets (non-gradients), we intentionally paste the targets in the output of S2 as the superresolution model may change the content of the targets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.502, + 0.6, + 0.518 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.527, + 0.776, + 0.543 + ], + "angle": 0, + "content": "5.1. Replication of visual illusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Qualitative results. We present qualitative results demonstrating our method's ability to replicate human visual perception across diverse scenarios in Fig. 4. Each comparison shows the original input image (left) and its corresponding DDIM inversion after 5 steps (right), with zoomed inset regions highlighting the areas where perceptual effects occur. The top row showcases our method's effectiveness on classical visual illusions, including Hermann grid and grating induction patterns. The bottom panels demonstrate the model's capability to capture perceptual phenomena in both natural grayscale photographs and artistic images with different color schemes. The zoomed regions, emphasize physically identical targets that are perceived differently due to their surrounding context — a key characteristic of visual illusions that our model successfully replicates. Additional qualitative results showing our performance across a broader range of visual illusions are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative results. Table 1 presents a comprehensive comparison of our DDIM-based approach against classical vision models and deep learning methods across multiple visual illusion datasets. The results demonstrate the consistency of the observation on diffusion models reported in Section 3.2 over a range of datasets and their superior" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18647" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.087, + 0.186, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.088, + 0.274, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.096, + 0.366, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.091, + 0.461, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.089, + 0.531, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.089, + 0.604, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.101, + 0.674, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.101, + 0.751, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.101, + 0.892, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.159, + 0.22, + 0.171 + ], + "angle": 0, + "content": "a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.157, + 0.403, + 0.172 + ], + "angle": 0, + "content": "b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.158, + 0.553, + 0.171 + ], + "angle": 0, + "content": "c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.158, + 0.694, + 0.171 + ], + "angle": 0, + "content": "d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.158, + 0.832, + 0.171 + ], + "angle": 0, + "content": "e)" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.18, + 0.352, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.181, + 0.589, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.184, + 0.737, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.184, + 0.874, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.286, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Qualitative replication of visual illusions (please enlarge display): For each item, the original image is on the left and its DDIM inversion on the right. The observation done in Fig. 2 is consistently reproduced. We zoom in on the region containing the illusion in each image pair. a) Barutan-seijin [21], b) Robot [22], c) Shiosai [23], d) Herman-grid [16], e) Grating induction [34], f) Bright room [48], g) Confetti illusion [39]. For more qualitative examples and results of classical visual illusions, see the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.483, + 0.565 + ], + "angle": 0, + "content": "performance in replicating illusions. Classical models like ODOG and CIWAM, while physiologically grounded, show limited performance and flexibility. These methods require careful parameter tuning for each specific type of visual illusion, as evidenced by their poor scores. For instance, ODOG can achieve \\(72.44\\%\\) accuracy on BRI3L but cannot generalize to color datasets, while applicable in color CIWAM shows poor performance. Previous deep learning approaches show consistent but modest performance in different datasets. Their uniform performance across sets suggests that they capture some general perceptual principles, but their relatively low accuracy indicates limitations in capturing human behavior. Our proposed method significantly outperforms all baseline approaches across all datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.732 + ], + "angle": 0, + "content": "The substantial performance gap between our DDIM approach and a standard VAE (which shows near-zero performance on most datasets) highlights the crucial role of the diffusion process in capturing perceptual phenomena. These results demonstrate that our DDIM-based approach not only achieves state-of-the-art performance but also offers superior generalization across different types of visual illusions without requiring illusion-specific parameter tuning. The consistent high performance across all datasets suggests that the diffusion process naturally captures fundamental aspects of human vision." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.357, + 0.756 + ], + "angle": 0, + "content": "5.2. Generation of Visual Illusions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Qualitative results. Fig. 5 demonstrates visual illusions using a \\(6 \\times 6\\) pixel square target in the S1 stage of DeepFloyd (expanding to \\(24 \\times 24\\) after S2). The figure displays five different target intensities (shown as circles to the left of each image) with different prompts for each image (selected randomly from DiffusionDB [56]). Additional results are presented in Fig. 1. In contrast to previous approaches, our method can generate complex content that simultaneously satisfies both the text prompt and the perceptual conditions." + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.352, + 0.902, + 0.366 + ], + "angle": 0, + "content": "Table 2. Perceptual Accuracy Score (PAS) in different VI sets" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.37, + 0.9, + 0.494 + ], + "angle": 0, + "content": "
MethodIllusionVQAGVILHallusionBench
CIWAM [40]56.5263.6383.33
DNet [12]14.2021.3033.30
RestoreNet [13]14.2021.3033.30
VAE4.7021.300.00
DDIM 5 steps90.4080.30100.00
DDIM 10 steps90.4096.70100.00
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.537 + ], + "angle": 0, + "content": "To see generation of classic VI (high frequency images with flat patches) please see supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.841 + ], + "angle": 0, + "content": "Psychophysical confirmation. We created 40 images in which we enforced our model to provide an illusion with either red, green, blue, or achromatic targets. Similarly, we have also created 40 control images—images for which there was no enforcement for the patterns to look different. 15 observers participated in our study. The experiment was conducted in a dark room with controlled dim illumination. The only light source was the monitor, set to sRGB. All observers were tested for correct color vision using the Ishihara test. The screen showed one image at a time. Observers were asked to decide whether the two targets were identical or different. The order of images was randomized. We analyze the results in terms of the number of times an image was said to have different targets. For the images in which our model enforced the existence of the illusion, the observers said the targets were different in \\(64\\%\\) (mean) and \\(67\\%\\) (median) of the cases with a standard deviation of \\(16\\%\\). In contrast, only in \\(13\\%\\) (mean) and \\(7\\%\\) (median)—standard deviation \\(14\\%\\)—of the cases where the control images said to have different targets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We also perform an experiment to evaluate the effectiveness of our perceptual loss weight \\((\\gamma)\\) through comparison with the two classical models—ODOG and CIWAM—across different thresholds. As the value \\(\\gamma\\) increases, there" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18648" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.359, + 0.09, + 0.669, + 0.105 + ], + "angle": 0, + "content": "Previous visual illusion generation methods" + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.108, + 0.127, + 0.121 + ], + "angle": 0, + "content": "a)" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.116, + 0.303, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.109, + 0.353, + 0.122 + ], + "angle": 0, + "content": "b)" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.109, + 0.492, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.109, + 0.552, + 0.121 + ], + "angle": 0, + "content": "c)" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.11, + 0.652, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.111, + 0.757, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.111, + 0.863, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.184, + 0.532, + 0.196 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.202, + 0.143, + 0.216 + ], + "angle": 0, + "content": "d)" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.202, + 0.301, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.202, + 0.34, + 0.216 + ], + "angle": 0, + "content": "e)" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.202, + 0.497, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.202, + 0.531, + 0.216 + ], + "angle": 0, + "content": "f)" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.202, + 0.692, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.202, + 0.731, + 0.216 + ], + "angle": 0, + "content": "g)" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.201, + 0.888, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.325, + 0.138, + 0.339 + ], + "angle": 0, + "content": "h)" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.323, + 0.301, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.325, + 0.344, + 0.34 + ], + "angle": 0, + "content": "i)" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.323, + 0.496, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.325, + 0.531, + 0.34 + ], + "angle": 0, + "content": "j)" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.323, + 0.691, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.325, + 0.733, + 0.34 + ], + "angle": 0, + "content": "k)" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.323, + 0.888, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.447, + 0.907, + 0.503 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of visual illusion generation (viewing at larger scale recommended). The first row presents prior work by (a) Gomez-Villa et al. [14], (b) Hirsch et al. [17], and (c) Roy et al. [46]. The second and third rows display results from our method (text prompts are available in the supplementary material). Images (d) through (i) use identical target colors (shown at the left of each image). In images (j) and (k), we employ grayscale gradients as targets. The selected target regions are highlighted in red in the thumbnail images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.483, + 0.542 + ], + "angle": 0, + "content": "is a substantial improvement in illusion strength, therefore showing the effectiveness of our perceptual loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.218, + 0.573 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this work, we make two key contributions. First, we report and demonstrate over a wide range of datasets an intriguing empirical observation: diffusion models naturally replicate human responses to brightness and color visual illusions through their image inversion process. This remarkable alignment between human perception and diffusion trajectories suggests that both systems may operate on similar statistical principles when processing ambiguous visual input. Second, leveraging this insight, we develop a method for generating visual illusions using T2I diffusion models, validated through psychophysical experiments. Looking ahead, our findings on perceptual processing in diffusion models could help bridge an important gap in current vision language models, which often struggle with nuanced visual understanding. The way diffusion models progressively process and interpret visual information might offer valuable insights for developing more perception-aware foundational models that better align with human visual cognition. This integration of perceptual understanding into image editing workflows represents an exciting direction for future research, potentially leading to tools that naturally" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.511, + 0.883, + 0.526 + ], + "angle": 0, + "content": "account for human visual perception in their operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.74 + ], + "angle": 0, + "content": "Limitations. (1) Replication. As expected from the image inversion process of diffusion models, the input images are spatially destroyed as the samples make their way to the Gaussian distribution. This implies that the target regions disappear for images closer to the natural image distribution, and the effect can no longer be measured. As a consequence our method has difficulty replicating the visual perception of effects which are based on small regions. (2) Generation. Our T2I VI generation method is based on defined regions during inference. This constraints the expressiveness of the T2I model in fulfilling the perceptual conditions. An ideal implementation should establish perceptual regions guided by the attention maps that the input prompt generates to produce more natural visual illusions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.743, + 0.907, + 0.894 + ], + "angle": 0, + "content": "Acknowledgements: We acknowledge projects PID2022-143257NB-I00, PID2021-128178OB-I00 PID2020-118071GB-I00 and PID2023-152133NB-I00 financed by MCIN/AEI/10.13039/501100011033 and ERDF \"A way of making Europe\", the CERCA Program and the Departament de Recerca i Universitats with ref. 2021SGR01499 from Generalitat de Catalunya, the grant Catedra ENIA UAB-Cruilla (TSI-100929-2023-2) from the Ministry of Economic Affairs and Digital Transition of Spain, and the BBVA Foundations of Science program (grant VIS4NN)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18649" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Arash Akbarinia, Yaniv Morgenstern, and Karl R Gegenfurtner. Contrast sensitivity function in deep networks. Neural Networks, 164:228-244, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.484, + 0.225 + ], + "angle": 0, + "content": "[2] DeepFloyd Lab at StabilityAI. DeepFloyd IF: a novel state-of-the-art open-source text-to-image model with a high degree of photorealism and language understanding. https://www deepfloyd.ai/deepfloyd-if, 2023. Retrieved on 2023-11-08. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Torsten Betz, Robert Shapley, Felix A Wichmann, and Marianne Maertens. Noise masking of white's illusion exposes the weakness of current spatial filtering models of lightness perception. Journal of vision, 15(14):1-1, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[4] Barbara Blakeslee and Mark E McCourt. A multiscale spatial filtering account of the white effect, simultaneous brightness contrast and grating induction. Vision research, 39(26): 4361-4377, 1999. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[5] E. Brucke. über erganzungs und contrasfarben. Wiener Sitzungsbser, 51, 1865. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.482, + 0.41 + ], + "angle": 0, + "content": "[6] Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael S Ryoo. Diffusion illusions: Hiding images in plain sight. arXiv preprint arXiv:2312.03817, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[8] David Corney and R Beau Lotto. What are lightness illusions and why do we see them? PLoS computational biology, 3 (9):e180, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[9] Karl R Gegenfurtner, Marina Bloj, and Matteo Toscani. The many colours of 'the dress'. Current Biology, 25(13):R543–R544, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.552, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[10] Daniel Geng, Inbum Park, and Andrew Owens. Visual analogs: Generating multi-view optical illusions with diffusion models. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.608, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[11] Daniel Geng, Inbum Park, and Andrew Owens. Factorized diffusion: Perceptual illusions by noise decomposition. arXiv preprint arXiv:2404.11615, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[12] Alex Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmó. Convolutional neural networks can be deceived by visual illusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309-12317, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[13] Alex Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, Marcelo Bertalmó, and Jesús Malo. Color illusions also deceive cnns for low-level vision tasks: Analysis and implications. Vision Research, 176:156-174, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[14] A. Gomez-Villa, A. Martin, J. Vazquez-Corral, M. Bertalmio, and J. Malo. On the synthesis of visual illusions using deep generative models. Journal of Vision, 22(8):2-2, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[16] Ludimar Hermann. Eine erscheinung simultanen contrastes. Archiv für die gesamte Physiologie des Menschen und der Tiere, 3(1):13-15, 1870. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.166, + 0.905, + 0.208 + ], + "angle": 0, + "content": "[17] Elad Hirsch and Ayellet Tal. Color visual illusions: A statistics-based computational model. Advances in neural information processing systems, 33:9447-9458, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.21, + 0.905, + 0.251 + ], + "angle": 0, + "content": "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.254, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[19] Been Kim, Emily Reif, Martin Wattenberg, and Samy Bengio. Do neural networks show gestalt phenomena? an exploration of the law of closure. arXiv preprint arXiv:1903.01069, 2(8), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.311, + 0.905, + 0.365 + ], + "angle": 0, + "content": "[20] Frederick AA Kingdom. Lightness, brightness and transparency: A quarter century of new ideas, captivating demonstrations and unrelenting controversy. Vision Research, 51 (7):652-673, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.904, + 0.395 + ], + "angle": 0, + "content": "[21] Akiyoshi Kitaoka. Barutan-seijin. Akiyoshi's illusion Website, 2003. Accessed: 2024-11-13. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.398, + 0.904, + 0.424 + ], + "angle": 0, + "content": "[22] Akiyoshi Kitaoka. Robot. Akiyoshi's illusion Website, 2005. Accessed: 2024-11-13. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.427, + 0.904, + 0.453 + ], + "angle": 0, + "content": "[23] Akiyoshi Kitaoka. Shiosai. Akiyoshi's illusion Website, 2008. Accessed: 2024-11-13. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[24] Yuki Kubota, Atsushi Hiyama, and Masahiko Inami. A machine learning model perceiving brightness optical illusions: Quantitative evaluation with psychophysical data. In Proceedings of the Augmented Humans International Conference 2021, pages 174-182, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.528, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[25] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.585, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[26] V. Laparra and J. Malo. Visual aftereffects and sensory nonlinearities from a single statistical framework. Front. Human Neurosci., 9, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.629, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[27] V. Laparra, G. Camps, and J. Malo. Iterative gaussianization: From ICA to random rotations. IEEE Trans. Neur. Nets, 22 (4):537-549, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.672, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[28] V. Laparra, S. Jiménez, G. Camps, and J. Malo. Nonlinearities and adaptation of color vision from sequential principal curves analysis. Neural Comp., 24(10):2751-2788, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.716, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[29] V. Laparra, JE. Johnson, G. Camps, R. Santos, and J. Malo. Estimating information theoretic measures via multidimensional gaussianization. IEEE Trans. Patt. Anal. Mach. Intell., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.773, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[30] Qiang Li, Alex Gomez-Villa, Marcelo Bertalmó, and Jesús Malo. Contrast sensitivity functions in autoencoders. Journal of Vision, 22(6):8-8, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[31] Senmao Li, Joost van de Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[32] Tingting Li, Fanyu Wang, Ying Zhou, and Zhenping Xie. Visual illusion cognition dataset construction and recognition" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18650" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "performance by deep neural networks. In 2022 IEEE 8th International Conference on Cloud Computing and Intelligent Systems (CCIS), pages 90-94. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[33] Ben Lonnqvist, Alban Bornet, Adrien Doerig, and Michael H Herzog. A comparative biology approach to dnn modeling of vision: A focus on differences, not similarities. Journal of vision, 21(10):17-17, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[34] Mark McCourt. Grating induction: A new explanation for stationary phantom gratings. Vision Research, 34(12):1609-1617, 1994. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[37] Chaithanya Kumar Mummadi, Ranjitha Subramaniam, Robin Hutmacher, Julien Vitay, Volker Fischer, and Jan Hendrik Metzen. Does enhanced shape bias improve neural network robustness to common corruptions? International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.484, + 0.483 + ], + "angle": 0, + "content": "[38] Jerry Ngo, Swami Sankaranarayanan, and Phillip Isola. Is clip fooled by optical illusions? International Conference on Learning Representations, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[39] David Novick and Akiyoshi Kitaoka. The confetti illusion. Journal of Illusion, 2, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[40] Xavier Otazu, C Alejandro Parraga, and Maria Vanrell. Toward a unified chromatic induction model. Journal of Vision, 10(12):5-5, 2010. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[41] Jiang Qin, Kai Wang, Bin Zou, Lamei Zhang, and Joost van de Weijer. Conditional diffusion model with spatial-frequency refinement for sar-to-optical image translation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[42] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[43] Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. International Conference on Machine Learning, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[45] Aniket Roy, Anirban Roy, Soma Mitra, and Kuntal Ghosh. Bri3l: A brightness illusion image dataset for identification and localization of regions of illusory perception. Proceedings of the IEEE International Conference on Image Processing, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[46] Aniket Roy, Anirban Roy, Soma Mitra, and Kuntal Ghosh. Bri3l: A brightness illusion image dataset for identification" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "and localization of regions of illusory perception. Proceedings of the IEEE International Conference on Image Processing, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[47] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. Illusionvqa: A challenging optical illusion dataset for vision language models. First Conference on Language Modeling, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[48] Arthur G. Shapiro and Dejan Todorovic. The Oxford Compendium of Visual Illusions, chapter 22. Oxford University Press, 2017. Figure 3.2. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.334 + ], + "angle": 0, + "content": "[50] Eric D Sun and Ron Dekel. Imagenet-trained deep neural networks exhibit illusion-like response to the scintillating grid. Journal of Vision, 21(11):15-15, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.377 + ], + "angle": 0, + "content": "[51] Chuanming Tang, Kai Wang, and Joost van de Weijer. Iterinv: Iterative inversion for pixel-level t2i models. Neurips 2023 workshop on Diffusion Models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[52] Oguzhan Ulucan, Diclehan Ulucan, and Marc Ebner. Investigating color illusions from the perspective of computational color constancy. arXiv preprint arXiv:2312.13114, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[53] Clinton Wang and Polina Golland. Interpolating between images with diffusion models. ICML 2023 Workshop DeployableGenerativeAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.519 + ], + "angle": 0, + "content": "[54] Kai Wang, Fei Yang, Shiqi Yang, Muhammad Atif Butt, and Joost van de Weijer. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.577 + ], + "angle": 0, + "content": "[55] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.647 + ], + "angle": 0, + "content": "[56] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. Annual Meeting of the Association for Computational Linguistics, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.905, + 0.703 + ], + "angle": 0, + "content": "[57] Eiji Watanabe, Akiyoshi Kitaoka, Kiwako Sakamoto, Masaki Yasugi, and Kenta Tanaka. Illusory motion reproduced by deep neural networks trained for prediction. Frontiers in psychology, 9:340023, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[58] Robert Max Williams and Roman V Yampolskiy. Optical illusions images dataset. arXiv preprint arXiv:1810.00415, 2, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[59] Hongtao Zhang, Zhen Li, and Shinichi Yoshida. Muller-layer illusion is replicated by higher layer of pre-trained deep neural network for object recognition. International Symposium on Computational Intelligence and Industrial Applications, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[60] Yichi Zhang, Jiayi Pan, Yuchen Zhou, Rui Pan, and Joyce Chai. Grounding visual illusions in language: Do vision-language models perceive illusions like humans? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5718-5728, 2023. 2, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18651" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.148 + ], + "angle": 0, + "content": "[61] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Tiviatis Sim, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. European Conference on Computer Vision, 2024. 5" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18652" + } + ] +] \ No newline at end of file diff --git a/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/32f81902-9541-4c54-8edb-2ca63c430edb_origin.pdf b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/32f81902-9541-4c54-8edb-2ca63c430edb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..09caf94fc35d82ee424cb03f03de40287a76666c --- /dev/null +++ b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/32f81902-9541-4c54-8edb-2ca63c430edb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b7f3ed766c2818644d754b26684be10e49b1d1982788d50ab029d48461d9255 +size 2396233 diff --git a/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/full.md b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..42848886eb1a5a4d1ea7e3e36cfac323399ec4d2 --- /dev/null +++ b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/full.md @@ -0,0 +1,381 @@ +# The Art of Deception: Color Visual Illusions and Diffusion Models + +Alexandra Gomez-Villa $^{1,3}$ + +Bartłomiej Twardowski1,2,4 Jesus Malo3 + +Kai Wang $^{1}$ C.Alejandro Parraga $^{1,2}$ + +Javier Vazquez-Corral $^{1,2}$ Joost van de Weijer $^{1,2}$ + +1Computer Vision Center, Spain 2Universitat Autònoma de Barcelona, Spain + +3Universitat de València, Spain 4IDEAS-NCBR, Poland + +![](images/da5bb77c7e1365f142870551f9a2b4773119ff861ddce6873d47590b2d8acde5.jpg) + +![](images/18b74c8002a322e16a74b95332a134f0e2ae997c11b02e0086ff2054044336ce.jpg) + +![](images/f07cabcd244c1607d3a1b56a92a9bd9834098043364b6cfa8b1e652534001d84.jpg) + +![](images/f053106e4ace068654512d02161eaf41a3641fa67e3e6343662ca70695b54dda.jpg) +Figure 1. Observation: Denoising Diffusion Implicit Models (DDIM) have human-like visual illusions. Application: text-to-image generated visual illusions. Examples (a) and (b) show how the responses to physically equal patches shift differently in the path to the latent space. In contrast, (c) shows how physically different stimuli get more similar along the path, achieving color constancy. The observation of this human-like behavior allows us to propose the use of text-to-image models to generate images in which physically identical patches are perceived differently (examples d, e, f). We highly recommend to watch the illusions on a computer screen. + +![](images/fd5ac4d23ce14017d6bfb159ce67c3c260b1a10d92d4ea975c2fdc975d46380b.jpg) + +![](images/4bb4104fef48fbf94c93049adbf74d9a4876eaa97d60bb6ce968f0dbe0225a0d.jpg) + +# Abstract + +Visual illusions in humans arise when interpreting out-of-distribution stimuli: if the observer is adapted to certain statistics, perception of outliers deviates from reality. Recent studies have shown that artificial neural networks (ANNs) can also be deceived by visual illusions. This revelation raises profound questions about the nature of visual information. Why are two independent systems, both human brains and ANNs, susceptible to the same illusions? Should any ANN be capable of perceiving visual illusions? Are these perceptions a feature or a flaw? In this work, we study how visual illusions are encoded in diffusion models. Remarkably, we show that they present human-like brightness/color shifts in their latent space. We use this fact to demonstrate that diffusion models can predict visual illusions. Furthermore, we also show how to gen + +erate new unseen visual illusions in realistic images using text-to-image diffusion models. We validate this ability through psychophysical experiments that show how our model-generated illusions also fool humans. + +# 1. Introduction + +Visual illusions (VIs) serve as a fascinating window into the complex workings of human visual perception, where our interpretation of reality diverges significantly from physical measurements. These phenomena can have a profound social impact, as demonstrated by viral phenomena such as "The Dress" [9], which sparked global discussions on the nature of perception. Beyond their intrinsic scientific value, + +VIs offer a unique lens through which we can examine and compare human and machine perception [13, 33]. The emergence of diffusion models (DMs) as state-of-the-art generative models has revolutionized the field of artificial intelligence, demonstrating unprecedented capabilities in image synthesis and manipulation. Their denoising-based architecture, which gradually transforms random noise into coherent images, presents a novel paradigm for studying visual processing that mirrors the lossy and noisy process of human perception. This parallel makes them particularly interesting for investigating VIs, as both systems must navigate the complex relationship between image statistics and perceptual interpretation. + +The study of VIs in artificial neural networks (ANNs) has a rich history. Early work demonstrated that a multilayer perceptron could replicate brightness illusions [8], while subsequent research by Gomez-Villa et al. [12] showed that convolutional neural networks (CNNs) could mirror human perception in lightness and color illusions. Concurrent developments included Watanabe et al.'s [57] demonstration of motion VI replication in video-prediction CNNs and the publication of the first visual illusion dataset [58]. This foundation has spawned diverse investigations into the ANN-VI relationship in multiple domains: image completion [19, 50], brightness/color perception [13, 17, 24, 32, 33, 52], contrast sensitivity [1, 30], shape perception [37, 59], and vision-language models [6, 38, 47, 60]. + +In this work, we investigate how human-like perceptual phenomena naturally emerge in the encoding process of DMs. Our key finding reveals that the intermediate steps of diffusion models exhibit trends that align remarkably with human perception. Specifically, when analyzing the trajectory of an inverted image through the diffusion process, we observe that the progressive changes in lightness and color properties closely mirror human perceptual responses. Prior research has demonstrated that Gaussianization flows exhibit interesting properties when processing samples that deviate from their training data statistics. When trained on specific sample distributions, these flows can produce unusual representations for out-of-distribution samples [27, 43] - a characteristic that has proven valuable for measuring distribution divergence [29]. Intriguingly, this phenomenon parallels human visual perception, where shifts in image statistics can lead to illusory perception of spatio-chromatic patterns, a behavior that has been explained through manifold equalization [26, 28]. We argue that image samples that do not follow the statistics of natural images are not properly gaussianized by DMs. In this way, linking defective gaussianization and illusory perception of unusual samples, it makes sense that intermediate representations in DMs may be correlated with the shifts of these samples for human observers. The reported alignment suggests that the iterative denoising in diffusion models may + +inadvertently capture fundamental aspects of human visual perception, particularly in contexts where our perception diverges from physical reality. By examining these intermediate representations, we provide novel insights into the nature of visual illusions and the unexpected parallels between human perception and the diffusion process. + +Contributions: The main contributions of this work are: + +- We report an empirical observation of the behavior of diffusion models: progressive changes in lightness, hue and saturation of image regions in the forward/inverse transform to the latent representation replicate the shifts of human perception of these regions. +- We derive a novel approach to using diffusion models to model human vision. We show that deep autoencoders can be effectively employed as vision models by performing a diffusion inversion process and measuring perception in the intermediate noise space. +- We demonstrate the human perception replication capabilities of DM in many cases, including VI datasets, popular published VI works, and natural images. +- Leveraging the perceptual capabilities of diffusion models, we propose a novel method for generating brightness/color VIs using text-to-image models. + +# 2. Related work + +Replication of Visual Illusions. Replication of visual illusions has been made with several methods, but all of them can be roughly classified into the following categories: + +- Image restoration methods [12, 13, 24, 30, 33]: These use pre-trained image restoration backbones. They process VIs and measure the output image to measure perception. +- Effect bias methods [1]: These works are inspired by neuroscience measurement of stimuli. Pretrained architectures are used to process VIs, then the response in the internal activations of the ANN is used to define if there is a visual illusion. +- Likelihood method [17]: In this approach, the most probable perceived intensity of each pixel in an image is estimated using an invertible flow trained on natural images. +- Language-guided methods [38, 60]: Language is used as a vehicle to measure perception. These works use vision-language models as test subjects and ask them questions regarding the perception of the image. + +Our approach differs from previous methods in the following: 1) we do not measure perception in the output of the model. Instead, we measure the VI in the intermediate steps of image inversion process. 2) We use a deep autoencoder; previous approaches in this category claim that deep autoencoders suffer from less perceptual bias and use shallow architectures for replication experiments. Furthermore, we are able to replicate human perception in high resolution images (as big as DMs are able to manage) and in natural images. These setups were not possible in previous works. + +Generation of Visual Illusions. Although generative models have been used to synthesize visual illusions [14], DMs has only recently been proposed as the generative engine. In [6], high-level illusions (overlay effects) are generated by mixing a set of images and transformations together. Following works [10, 11] were able to create more general illusions such as rotations, color inversions, and hybrid images. The work most similar to ours is by Roy et al. [46], who used stable diffusion (SD) to generate brightness visual illusions through keyword prompts. However, their approach neither explicitly optimizes perceptual effects nor specifies regions for perception replication. Consequently, illusions may appear randomly in the generated image with unknown localization. Additionally, their generated images are limited to conventional brightness illusions (gratings or high-frequency black and white patterns) and cannot incorporate visual illusions into natural images. In contrast, our work allows visual illusions within generated natural images as well as in typical high-frequency patterns. We explicitly generate the perceptual effect in a designated image region and can control the image content through text prompts. + +# 3. Methods + +# 3.1. Diffusion models + +Diffusion models (DMs) are generative models that learn to gradually denoise images by reversing a fixed Markov chain of forward diffusion steps. The forward process systematically adds Gaussian noise to an image until it becomes pure noise, while the learned reverse process reconstructs the original image through iterative denoising. While traditional diffusion follows a stochastic sampling process, Denoising Diffusion Implicit Models (DDIM) [49] introduce a deterministic sampling approach that maintains high quality while enabling significantly fewer inference steps. + +DDIM's deterministic nature enables a crucial capability: the ability to invert generated images back into their latent representations through a process known as DDIM inversion [49]. This inversion process allows us to map any image into the model's latent space and observe its trajectory through the diffusion process. Given an image $z_0$ , DDIM inversion produces a sequence of intermediate representations $\{z_t\}_{t=1}^T$ according to: + +$$ +z _ {t} = \sqrt {\bar {\alpha} _ {t}} z _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} ^ {z _ {t}, t} (1) +$$ + +where $\bar{\alpha}_t$ represents the cumulative product of noise scheduling coefficients, and $\epsilon_{\theta}^{z_t,t} = \epsilon_{\theta}(z_t,t)$ is the learned noise prediction network. The reverse process follows: + +$$ +z _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \left(\frac {z _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} ^ {z _ {t} , t}}{\sqrt {\bar {\alpha} _ {t}}}\right) + \sqrt {1 - \bar {\alpha} _ {t - 1}} \epsilon_ {\theta} ^ {z _ {t}, t} \tag {2} +$$ + +DDIM inversion has emerged as a powerful tool for various image manipulation tasks. Its ability to faithfully reconstruct latent representations has enabled applications such as semantic image editing [31, 35, 36, 51, 54], style transfer [25, 41], image interpolation [53], and controlled image generation through classifier-free guidance [18]. In our work, we exploit the deterministic nature of DDIM inversion to systematically study how visual illusions are processed through the diffusion trajectory. + +# 3.2. DDIM inversion may replicate human vision + +The wholly-empirical evolutionary paradigm [20] suggests that human vision is shaped by statistical regularities in natural scenes, where our visual system evolved to interpret ambiguous stimuli based on their likelihood. We leverage this insight to study how diffusion models, trained for natural images, process illusory stimuli, which are not in the distribution of natural images. + +Motivation. In Fig. 2, we present an illustrative experiment using Stable Diffusion [44] (extra results with an image space model in the supplementary material). We perform DDIM inversion on the canonical visual illusion known as Brightness Contrast [5]. DDIM inversion puts the image back into the latent space. This illusion presents two gray squares of identical intensity (indicated by red markers) against white and black backgrounds, which leads humans to perceive the left square as darker than the right one. The first row (left to right) shows the inversion1 for 0, 3, 10, and 20 steps, respectively. As the zoomed-in bar shows, the initially equal-intensity regions progressively develop an intensity difference that aligns with human perception2 (left target appearing darker than right). The second row shows the progressive Gaussianization of the latent representation. For this image perception is replicated at an intermediate representation where gaussianization is not complete (since this stimulus is not in the distribution of natural images). + +Observation: emergence of perceptual biases. Our key observation is: when inverting an image through diffusion models the intermediate denoising steps progressively introduce color/brightness shifts that align with human perception. We hypothesize that this occurs because diffusion models, trained on natural images, map illusory stimuli onto the learned natural image manifold, resulting in systematic "interpretation errors" that mirror human perceptual biases. + +Theoretical framework. Formally, let $G: \mathcal{X} \to \mathcal{Z}$ be the forward diffusion mapping from image space $\mathcal{X}$ to latent space $\mathcal{Z}$ . For natural images $x \in \mathcal{X}$ , $G$ learns a mapping $z = G(x)$ that transforms the image distribution into a Gaussian. However, when presented with an eventually + +![](images/2b799b5d302dfd681838acea957ac44f3c05f628c5c3f9847b010cdd24218d21.jpg) +Figure 2. DDIM inversion of the Brightness Contrast illusion [5] using Stable Diffusion. Top row: Image-space visualization (decoded latents) showing (left) Original illusion with two identical gray squares (marked in red) against black and white backgrounds, and inversion results using 3, 10, and 20 steps. Bottom row: Histograms of the corresponding latent representations. The horizontal axis represents the value of the latent space feature. The model gradually reproduces the perceptual difference in brightness between the physically identical squares in a (not fully Gaussian) intermediate representation. + +illusory stimulus $y$ , we have $y \notin \mathcal{X}$ . Still, the mapping $z' = G(y)$ attempts to project the illusion onto the learned natural image manifold, introducing systematic deviations. The intermediate states $z_{t}$ reveal how the model progressively "corrects" the illusion toward natural image statistics. We observe that these corrections correlate strongly with human perception, suggesting that both humans and diffusion models may operate on similar statistical principles when processing ambiguous visual input. This alignment between diffusion trajectories and human perception provides computational evidence for the wholly-empirical paradigm, demonstrating how statistical learning can lead to perceptual illusions without explicit encoding of perceptual rules. The systematic nature of these "errors" suggests they emerge from the same statistical regularities that shape human visual processing through evolution. + +# 3.3. Generating visual illusions with DMs + +Based on the findings above, we develop a method to automatically generate brightness/color illusions with DMs. To do so, we need to evaluate whether a VI is present and then push the diffusion process toward the desired perception. + +Previous works on VIs and DMs focus on high-level visual illusions, i.e. cognitive effects that measure if the image belongs to a certain class of prompt [6, 10, 11]. Here we focus on brightness/color VIs, which leads to the challenge of measuring effects in specific regions of the image. Here we answer the measurement questions proposed in previous works on brightness/color VI generation in this way: + +Where to measure? This question separates the generation of high-level VIs from color/brightness VIs. Models aiming at high-level VIs that fool an ANN require to ask a classifier/vision-language model which class/prompt better describes the image [10, 38, 60]. In contrast, color VIs ap + +pear in specific regions of the image, therefore, they need to be measured at those points. In this work, we specifically define regions (here targets) in which we want a perceptual effect to appear. These targets are our measurement points, and we guarantee that in these regions, we can track, measure, and optimize perceptual effects. Failing to define a target leads to failure, as reported in [58]. + +What to measure? During image generation, we will measure the intensity values of the intermediate steps $z_{t}$ . Even though $z_{t}$ does not contain the final perception bias $(z_{T})$ , we can push the generative process towards a desired perceptual effect using small changes in the intermediate representation as explained in Fig. 3. + +How to measure? In the motivating example of section 3.2 (Fig. 2), the model replicates human perception if the target shifts in the same direction as humans perceive. However, in an optimization procedure, we need to measure how "good" the illusion is; in other words, we need to quantify the effect. Since we are concerned about brightness and color VI, we will measure the mean intensity per channel of each target. More precisely, given an image $I$ , and the target region $r \in \mathbb{R}^2$ in $I$ , we compute the mean intensity in that target region as: + +$$ +m _ {i n t} (I, r) = \frac {1}{M} \sum_ {i \in r} I \left(r _ {i}\right) \tag {3} +$$ + +where $r_i$ denotes a specific pixel $i$ in target region $r$ , $M$ stands for the number of pixels in the target region, and $m_{int}$ is measured in each channel. Hence, the perceived intensity is an RGB triplet $(m_{int}^R, m_{int}^G, m_{int}^B)$ . + +# 3.3.1. Visual perception inside generation pipeline. + +Since text-to-image (T2I) diffusion models often struggle to perfectly align with textual prompts, a variety of aug- + +![](images/58ab821c681add6ca3fd7ee579250a40d4bd6a93cc00345f3c7d35a3221fc959.jpg) +Figure 3. Overview of our visual illusion generation pipeline. The process modifies noisy latent representations, $z_{t}$ , through a custom loss function, $\mathcal{L}$ , that guides the generation toward perceptually ambiguous outputs. Note that targets are regions that manually selected by the user as inputs. The terms $D_{p1}$ and $D_{p2}$ are the user desired intensity for target T1 and T2 respectively. + +mentation methods have been developed to enhance the fidelity of T2I generation. These methods typically focus on optimizing noisy latent representations in each timestep, a strategy that has proven effective in refining generation outcomes [7, 42, 55, 61]. A common approach involves the design of an update loss function, $\mathcal{L}$ , which is tailored based on the layout of the noisy latent or the characteristics of the cross-attention maps. This loss is then used to guide the backpropagation process, updating the noisy latent in a manner that more accurately reflects the intended textual descriptions, thereby enhancing the overall quality of the generated images. This process can be formulated as $z_{t}^{\prime} = z_{t} - \alpha \cdot \nabla_{z_{t}}\mathcal{L}$ . In this paper, we diverge from the conventional use of optimization techniques aimed at aligning T2I diffusion models with textual semantics. Instead, we leverage these techniques to create visual perception effects (see Fig. 3). We use an image space DM (DeepFloyd [2]) to avoid the use of the decoder in the backpropagation. + +First, we define the original target image as $O(r)$ , where $r \in \mathbb{R}^2$ . This image only contains the input-defined targets, that is, $O(r)$ has the original value given to the targets by the user on each color channel. Then, given an intermediate image $z_t$ and a list of $R$ target regions $\{r^p\}_{p=1}^R$ , our process to generate a VI includes the addition of the targets to the image in the desired region3, i.e. we compute $z_{t\_target}(r_i^p) = z_t(r_i^p) + O(r_i^p)$ , for all pixels $i$ inside the region $r^p$ and $z_{t\_target} = z_t$ for any pixel outside this target region. Finally, the following optimization term is applied $N$ times —per diffusion step $t$ — to guide the generation: + +$$ +\mathcal {L} = \gamma \mathcal {L} _ {V I} + \beta \mathcal {L} _ {s i m} \tag {4} +$$ + +where $\mathcal{L}_{VI}$ and $\mathcal{L}_{sim}$ measure the illusion strength and the + +similarity between targets and background respectively, and $\gamma$ and $\beta$ are hyperparameters that control the relative weight of these measures. Specifically: + +- $\mathcal{L}_{VI}$ is a perceptual loss on the target intensities: + +$$ +\mathcal {L} _ {V I} = \sum_ {c \in R, G, B} \sum_ {r ^ {p}} ^ {R} \mathcal {L} _ {\text {P e r c e p t}} \left(z _ {t _ {\text {t a r g e t}}}, k _ {c} ^ {p}; r ^ {p}\right) \tag {5} +$$ + +where $r^p$ is a specific target, $k_c^p$ is the desired intensity for target $r^p$ on color channel $c$ , and + +$$ +\mathcal {L} _ {\text {P e r c e p t u a l}} \left(z _ {t _ {\text {t a r g e t}}}, k _ {c}; r\right) = \left\| m _ {\text {i n t}} \left(z _ {t _ {\text {t a r g e t}}}, r\right) - k _ {c} \right\| \tag {6} +$$ + +with $m_{int}$ as defined in Eq. 3. + +- $\mathcal{L}_{sim}$ is a similarity term. This term aims at avoiding the region from appearing "out of place". In particular, it looks at harmonizing the targets with the background-generated image: + +$$ +\mathcal {L} _ {s i m} = \sum_ {p = 1} ^ {R} M S E \left(z _ {t} \left(r ^ {p}\right), O \left(r ^ {p}\right)\right) \tag {7} +$$ + +# 4. Experimental setup + +# 4.1. Datasets + +- Brightness illusions dataset: The $BRI3L$ dataset [45] consists of 22366 grayscale images spanning five types of brightness illusions. Each image is annotated with binary masks indicating regions of illusory perception. + +- Visual Illusion VQA Datasets: Several datasets have been developed to evaluate vision-language models' understanding of VIs: IllusionVQA [47] contains 435 illusory images with multiple-choice questions across 12 categories, plus 1000 impossible object image pairs for localization testing. GVIL [60] features 100 base illusion images expanded to 1600 instances covering five illusion categories, with four evaluation tasks: same-difference QA, referential QA, attribute QA, and referential localization. HallusionBench [15] comprises 346 images and 1129 questions designed to test image-context reasoning and diagnose both VIs and language hallucinations. We use the color category in all the VQA datasets. + +# 4.2. Benchmarks + +- Classical vision science models: These approaches are based in vision science and need to be calibrated for each image to fully exploit their capabilities. In this paper we use the parameters set by Betz et al. [3]. These parameters constitute a fair guideline for these methods, as it is not possible to fine-tune them for thousands of images. We use ODOG (Oriented Difference of Gaussians) [4] and the Chromatic Induction Wavelet Model (CIWaM) [40] + +- Neural networks-base models: We compare against previous image restoration methods that proven ability in visual illusion detection: DN-NET[13] and Restorenet[30]. +- Variational autoencoder (VAE): We present the results of only considering the VAE of SD prior to the DDIM inversion process —to further demonstrate DDIM is the responsible for the alignment with human perception. + +# 4.3. Metrics + +While visual illusion replication results have traditionally been presented through profile plots comparing model outputs with human perception along image cross-sections, we leverage recent datasets to provide quantitative metrics. This approach complements classical visualization methods with rigorous numerical evaluation. + +For the BRI3L dataset, we compute the mean intensity difference, $\Delta I$ , between the model's output and the input image within each segmented region: + +$$ +\Delta I = \frac {1}{| M |} \sum_ {i \in r} \left| I _ {\text {o u t}} (r) - I _ {\text {i n}} (r) \right| \tag {8} +$$ + +where $M$ represents the size of the segmentation mask, and $I_{out}$ , $I_{in}$ are the output and input intensities respectively. We consider the model's response to be aligned with human perception when $\Delta I < \tau * I_{int}$ , where $I_{int}$ represents the mean intensity of the input image in the target region (calculated using Eq. 3). Since BRI3L aims to detect regions that become darker, the metric becomes more harder as the value of $\tau$ decreases, requiring the model to achieve greater darkness in the target regions. + +For the VQA-based datasets (IllusionVQA, GVIL, and HallusionBench), we manually create segmentation masks and consider the answer of the model aligned with human perception if it matches the expected intensity for regions referenced in the question-answer pairs. The Perception Accuracy Score (PAS) is the percentage of images having the correct perceptual predictions in each dataset. + +# 4.4. Experiments + +We conducted two sets of experiments to evaluate our method's ability to: (1) replicate existing VIs across multiple datasets, and (2) generate novel VIs that can fool human perception. For each experiment, we present qualitative results zooming in on the region of the perceptual effect and quantitative evaluations using the proposed metrics. + +- Replication of visual illusions: We test our model's ability to replicate brightness (BRI3L) and color (Illusion-VQA, GVIL, and HallusionBench) illusions by comparing the intensity differences between input and output images within annotated segmentation masks. For BRI3L, we evaluate performance across three $\tau$ levels (0.8, 0.9, 1.0) to assess the robustness of illusion replication. In this + +Table 1. Perceptual Accuracy Score (PAS) in BRI3L for different methods and thresholds + +
Method / Threshold0.80.91.0
ODOG [4]36.5857.0972.44
CIWAM [40]5.6918.2356.22
DNnet [12]0.000.0071.09
RestoreNet [13]0.000.0074.95
VAE0.000.0873.18
DDIM 5 steps44.0291.4399.88
DDIM 10 steps84.1798.85100.00
+ +experiment we present results using SD 1.4 (latent space DM) but additional results with DeepFloyd (image space DM) are provided in the supplementary material. + +- Generation of visual illusions: In this experiment, we check whether we can generate VIs in images that fall inside the distribution of the T2I model. We use guidance scale $= 10$ (we want to generate images that follow the prompt), $N = 5$ , $\gamma = 0.5$ , and $\beta = 1.0$ . Additionally, here we use S2 of DeepFloyd to generate higher-resolution versions of our generated visual illusions. For flat intensity targets (non-gradients), we intentionally paste the targets in the output of S2 as the superresolution model may change the content of the targets. + +# 5. Results + +# 5.1. Replication of visual illusions + +Qualitative results. We present qualitative results demonstrating our method's ability to replicate human visual perception across diverse scenarios in Fig. 4. Each comparison shows the original input image (left) and its corresponding DDIM inversion after 5 steps (right), with zoomed inset regions highlighting the areas where perceptual effects occur. The top row showcases our method's effectiveness on classical visual illusions, including Hermann grid and grating induction patterns. The bottom panels demonstrate the model's capability to capture perceptual phenomena in both natural grayscale photographs and artistic images with different color schemes. The zoomed regions, emphasize physically identical targets that are perceived differently due to their surrounding context — a key characteristic of visual illusions that our model successfully replicates. Additional qualitative results showing our performance across a broader range of visual illusions are in the supplementary. + +Quantitative results. Table 1 presents a comprehensive comparison of our DDIM-based approach against classical vision models and deep learning methods across multiple visual illusion datasets. The results demonstrate the consistency of the observation on diffusion models reported in Section 3.2 over a range of datasets and their superior + +![](images/b611aea49589fd6f66b1108aa700f3e16e4366e2f6ca8111e51cbfd27d534348.jpg) + +![](images/5ff1e0979d122db9af8dc22424f2b3512a0d299a9f32c05a3e410108c0b82791.jpg) +a) + +![](images/92b1e53973f4f6bf8d6c2e884c8e147cedd2c7b7ec311701a85ede9af656ea05.jpg) + +![](images/cd8bb5fbe4c2bf96c67b75678c30b0a472f58757a4827ccb4655e35e9b612190.jpg) +b) + +![](images/c055d17c3ff10faf73f30d7a1d0e42a2449cf21d576c1190af3336fa2e7a6b1f.jpg) +c) + +![](images/129bc26c3ee619676a50a74f940dcfc798289677b332db4674c2fd478095fb80.jpg) + +![](images/300bbf70a7eb344cd8a75b026438aa9f710a39b02cc9a266595dc336a97af030.jpg) +d) + +![](images/be5912884070aade400b4944ca68c088a3c4bebfd6c7cc75f2416ba0a0aaf6b6.jpg) + +![](images/e60ae82537d228e22a5278c4cd36a45dc9c236de1b9020a57ad89ee5e3675fe2.jpg) +e) + +![](images/af51c829bee1546b67540611cf18dc473c521345df79a290552f927b990ea825.jpg) +Figure 4. Qualitative replication of visual illusions (please enlarge display): For each item, the original image is on the left and its DDIM inversion on the right. The observation done in Fig. 2 is consistently reproduced. We zoom in on the region containing the illusion in each image pair. a) Barutan-seijin [21], b) Robot [22], c) Shiosai [23], d) Herman-grid [16], e) Grating induction [34], f) Bright room [48], g) Confetti illusion [39]. For more qualitative examples and results of classical visual illusions, see the supplementary material. + +![](images/dedbf501dd0576c9f1ea5671119eecabd1142b9f163980bcb89b8ba2d4c674fa.jpg) + +![](images/51e76b6dd54e200cebe3c730db8a0104eacf30199a5ac3fa1f4946e986581139.jpg) + +![](images/99224664b909903de01abb5ea3eb175a10d1a33d96a8584e18ef82e4cda7f336.jpg) + +performance in replicating illusions. Classical models like ODOG and CIWAM, while physiologically grounded, show limited performance and flexibility. These methods require careful parameter tuning for each specific type of visual illusion, as evidenced by their poor scores. For instance, ODOG can achieve $72.44\%$ accuracy on BRI3L but cannot generalize to color datasets, while applicable in color CIWAM shows poor performance. Previous deep learning approaches show consistent but modest performance in different datasets. Their uniform performance across sets suggests that they capture some general perceptual principles, but their relatively low accuracy indicates limitations in capturing human behavior. Our proposed method significantly outperforms all baseline approaches across all datasets. + +The substantial performance gap between our DDIM approach and a standard VAE (which shows near-zero performance on most datasets) highlights the crucial role of the diffusion process in capturing perceptual phenomena. These results demonstrate that our DDIM-based approach not only achieves state-of-the-art performance but also offers superior generalization across different types of visual illusions without requiring illusion-specific parameter tuning. The consistent high performance across all datasets suggests that the diffusion process naturally captures fundamental aspects of human vision. + +# 5.2. Generation of Visual Illusions + +Qualitative results. Fig. 5 demonstrates visual illusions using a $6 \times 6$ pixel square target in the S1 stage of DeepFloyd (expanding to $24 \times 24$ after S2). The figure displays five different target intensities (shown as circles to the left of each image) with different prompts for each image (selected randomly from DiffusionDB [56]). Additional results are presented in Fig. 1. In contrast to previous approaches, our method can generate complex content that simultaneously satisfies both the text prompt and the perceptual conditions. + +Table 2. Perceptual Accuracy Score (PAS) in different VI sets + +
MethodIllusionVQAGVILHallusionBench
CIWAM [40]56.5263.6383.33
DNet [12]14.2021.3033.30
RestoreNet [13]14.2021.3033.30
VAE4.7021.300.00
DDIM 5 steps90.4080.30100.00
DDIM 10 steps90.4096.70100.00
+ +To see generation of classic VI (high frequency images with flat patches) please see supplementary material. + +Psychophysical confirmation. We created 40 images in which we enforced our model to provide an illusion with either red, green, blue, or achromatic targets. Similarly, we have also created 40 control images—images for which there was no enforcement for the patterns to look different. 15 observers participated in our study. The experiment was conducted in a dark room with controlled dim illumination. The only light source was the monitor, set to sRGB. All observers were tested for correct color vision using the Ishihara test. The screen showed one image at a time. Observers were asked to decide whether the two targets were identical or different. The order of images was randomized. We analyze the results in terms of the number of times an image was said to have different targets. For the images in which our model enforced the existence of the illusion, the observers said the targets were different in $64\%$ (mean) and $67\%$ (median) of the cases with a standard deviation of $16\%$ . In contrast, only in $13\%$ (mean) and $7\%$ (median)—standard deviation $14\%$ —of the cases where the control images said to have different targets. + +We also perform an experiment to evaluate the effectiveness of our perceptual loss weight $(\gamma)$ through comparison with the two classical models—ODOG and CIWAM—across different thresholds. As the value $\gamma$ increases, there + +![](images/b90da7016e280bde492e7ef10c473a1db110911b1ff9499124cdbaccc335c016.jpg) +a) + +![](images/94558c6caabfb266f3f1983bb4236c4db41f461828b2a9b1f4efaeca00ee2fad.jpg) +b) +Ours + +![](images/7dcf72aab58ce7703651a42eccb988e85b31bf1f0621b18b0e9f6c7afcde7f38.jpg) +c) + +![](images/ba44805ab708657f2bfde8df50edb8818410d56819d0cf23535a5036b00466c6.jpg) + +![](images/a17e5614b74268d8beac3b1fadb18cc7c20a8bca073a0ec3dbc16d13d85158bd.jpg) + +![](images/f0d8af4dbe3c316c279970eff1f9c3ece53ffef40f9a6859798897fa302b7932.jpg) +d) + +![](images/79e15d6cc0391949f8280451e47baa1cf0ad638ce6ec04b6f6e44b747c2b9204.jpg) +e) + +![](images/1d634b54bb14295b922c1bd55c885b77b00567ed0b18946e2eb122dbd3e81ea6.jpg) +f) + +![](images/6b2507d6a5a8567a8eab66fc5ab23ab329bbbcfa36cfa861033616ef789b6045.jpg) +g) + +![](images/8a6e088540b3c48f6d3ca69e2e968658dfec62704a3a2891c05bb497f69bb50b.jpg) +h) +Figure 5. Qualitative results of visual illusion generation (viewing at larger scale recommended). The first row presents prior work by (a) Gomez-Villa et al. [14], (b) Hirsch et al. [17], and (c) Roy et al. [46]. The second and third rows display results from our method (text prompts are available in the supplementary material). Images (d) through (i) use identical target colors (shown at the left of each image). In images (j) and (k), we employ grayscale gradients as targets. The selected target regions are highlighted in red in the thumbnail images. + +![](images/5558442cabcbbe120c0efb0c49a4676ef686da8f6ecbc372b6073053c3007e70.jpg) +i) + +![](images/b61a592e854fc46fb2462ee136e3bda1cc1a3d985529b73a573cafed2ea6a779.jpg) +j) + +![](images/65fee3c57a275af847063cec91b23fadd965f292ba585ac86d4fe51f52b750ed.jpg) +k) + +is a substantial improvement in illusion strength, therefore showing the effectiveness of our perceptual loss. + +# 6. Conclusions + +In this work, we make two key contributions. First, we report and demonstrate over a wide range of datasets an intriguing empirical observation: diffusion models naturally replicate human responses to brightness and color visual illusions through their image inversion process. This remarkable alignment between human perception and diffusion trajectories suggests that both systems may operate on similar statistical principles when processing ambiguous visual input. Second, leveraging this insight, we develop a method for generating visual illusions using T2I diffusion models, validated through psychophysical experiments. Looking ahead, our findings on perceptual processing in diffusion models could help bridge an important gap in current vision language models, which often struggle with nuanced visual understanding. The way diffusion models progressively process and interpret visual information might offer valuable insights for developing more perception-aware foundational models that better align with human visual cognition. This integration of perceptual understanding into image editing workflows represents an exciting direction for future research, potentially leading to tools that naturally + +account for human visual perception in their operations. + +Limitations. (1) Replication. As expected from the image inversion process of diffusion models, the input images are spatially destroyed as the samples make their way to the Gaussian distribution. This implies that the target regions disappear for images closer to the natural image distribution, and the effect can no longer be measured. As a consequence our method has difficulty replicating the visual perception of effects which are based on small regions. (2) Generation. Our T2I VI generation method is based on defined regions during inference. This constraints the expressiveness of the T2I model in fulfilling the perceptual conditions. An ideal implementation should establish perceptual regions guided by the attention maps that the input prompt generates to produce more natural visual illusions. + +Acknowledgements: We acknowledge projects PID2022-143257NB-I00, PID2021-128178OB-I00 PID2020-118071GB-I00 and PID2023-152133NB-I00 financed by MCIN/AEI/10.13039/501100011033 and ERDF "A way of making Europe", the CERCA Program and the Departament de Recerca i Universitats with ref. 2021SGR01499 from Generalitat de Catalunya, the grant Catedra ENIA UAB-Cruilla (TSI-100929-2023-2) from the Ministry of Economic Affairs and Digital Transition of Spain, and the BBVA Foundations of Science program (grant VIS4NN). + +# References + +[1] Arash Akbarinia, Yaniv Morgenstern, and Karl R Gegenfurtner. Contrast sensitivity function in deep networks. Neural Networks, 164:228-244, 2023. 2 +[2] DeepFloyd Lab at StabilityAI. DeepFloyd IF: a novel state-of-the-art open-source text-to-image model with a high degree of photorealism and language understanding. https://www deepfloyd.ai/deepfloyd-if, 2023. Retrieved on 2023-11-08. 5 +[3] Torsten Betz, Robert Shapley, Felix A Wichmann, and Marianne Maertens. Noise masking of white's illusion exposes the weakness of current spatial filtering models of lightness perception. Journal of vision, 15(14):1-1, 2015. 5 +[4] Barbara Blakeslee and Mark E McCourt. A multiscale spatial filtering account of the white effect, simultaneous brightness contrast and grating induction. Vision research, 39(26): 4361-4377, 1999. 5, 6 +[5] E. Brucke. über erganzungs und contrasfarben. Wiener Sitzungsbser, 51, 1865. 3, 4 +[6] Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael S Ryoo. Diffusion illusions: Hiding images in plain sight. arXiv preprint arXiv:2312.03817, 2023. 2, 3, 4 +[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 2023. 5 +[8] David Corney and R Beau Lotto. What are lightness illusions and why do we see them? PLoS computational biology, 3 (9):e180, 2007. 2 +[9] Karl R Gegenfurtner, Marina Bloj, and Matteo Toscani. The many colours of 'the dress'. Current Biology, 25(13):R543–R544, 2015. 1 +[10] Daniel Geng, Inbum Park, and Andrew Owens. Visual analogs: Generating multi-view optical illusions with diffusion models. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3, 4 +[11] Daniel Geng, Inbum Park, and Andrew Owens. Factorized diffusion: Perceptual illusions by noise decomposition. arXiv preprint arXiv:2404.11615, 2024. 3, 4 +[12] Alex Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmó. Convolutional neural networks can be deceived by visual illusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309-12317, 2019. 2, 6, 7 +[13] Alex Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, Marcelo Bertalmó, and Jesús Malo. Color illusions also deceive cnns for low-level vision tasks: Analysis and implications. Vision Research, 176:156-174, 2020. 2, 6, 7 +[14] A. Gomez-Villa, A. Martin, J. Vazquez-Corral, M. Bertalmio, and J. Malo. On the synthesis of visual illusions using deep generative models. Journal of Vision, 22(8):2-2, 2022. 3, 8 +[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5 +[16] Ludimar Hermann. Eine erscheinung simultanen contrastes. Archiv für die gesamte Physiologie des Menschen und der Tiere, 3(1):13-15, 1870. 7 +[17] Elad Hirsch and Ayellet Tal. Color visual illusions: A statistics-based computational model. Advances in neural information processing systems, 33:9447-9458, 2020. 2, 8 +[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2022. 3 +[19] Been Kim, Emily Reif, Martin Wattenberg, and Samy Bengio. Do neural networks show gestalt phenomena? an exploration of the law of closure. arXiv preprint arXiv:1903.01069, 2(8), 2019. 2 +[20] Frederick AA Kingdom. Lightness, brightness and transparency: A quarter century of new ideas, captivating demonstrations and unrelenting controversy. Vision Research, 51 (7):652-673, 2011. 3 +[21] Akiyoshi Kitaoka. Barutan-seijin. Akiyoshi's illusion Website, 2003. Accessed: 2024-11-13. 7 +[22] Akiyoshi Kitaoka. Robot. Akiyoshi's illusion Website, 2005. Accessed: 2024-11-13. 7 +[23] Akiyoshi Kitaoka. Shiosai. Akiyoshi's illusion Website, 2008. Accessed: 2024-11-13. 7 +[24] Yuki Kubota, Atsushi Hiyama, and Masahiko Inami. A machine learning model perceiving brightness optical illusions: Quantitative evaluation with psychophysical data. In Proceedings of the Augmented Humans International Conference 2021, pages 174-182, 2021. 2 +[25] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. In The Eleventh International Conference on Learning Representations, 2023. 3 +[26] V. Laparra and J. Malo. Visual aftereffects and sensory nonlinearities from a single statistical framework. Front. Human Neurosci., 9, 2015. 2 +[27] V. Laparra, G. Camps, and J. Malo. Iterative gaussianization: From ICA to random rotations. IEEE Trans. Neur. Nets, 22 (4):537-549, 2011. 2 +[28] V. Laparra, S. Jiménez, G. Camps, and J. Malo. Nonlinearities and adaptation of color vision from sequential principal curves analysis. Neural Comp., 24(10):2751-2788, 2012. 2 +[29] V. Laparra, JE. Johnson, G. Camps, R. Santos, and J. Malo. Estimating information theoretic measures via multidimensional gaussianization. IEEE Trans. Patt. Anal. Mach. Intell., 2024. 2 +[30] Qiang Li, Alex Gomez-Villa, Marcelo Bertalmó, and Jesús Malo. Contrast sensitivity functions in autoencoders. Journal of Vision, 22(6):8-8, 2022. 2, 6 +[31] Senmao Li, Joost van de Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing, 2023. 3 +[32] Tingting Li, Fanyu Wang, Ying Zhou, and Zhenping Xie. Visual illusion cognition dataset construction and recognition + +performance by deep neural networks. In 2022 IEEE 8th International Conference on Cloud Computing and Intelligent Systems (CCIS), pages 90-94. IEEE, 2022. 2 +[33] Ben Lonnqvist, Alban Bornet, Adrien Doerig, and Michael H Herzog. A comparative biology approach to dnn modeling of vision: A focus on differences, not similarities. Journal of vision, 21(10):17-17, 2021. 2 +[34] Mark McCourt. Grating induction: A new explanation for stationary phantom gratings. Vision Research, 34(12):1609-1617, 1994. 7 +[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 3 +[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3 +[37] Chaithanya Kumar Mummadi, Ranjitha Subramaniam, Robin Hutmacher, Julien Vitay, Volker Fischer, and Jan Hendrik Metzen. Does enhanced shape bias improve neural network robustness to common corruptions? International Conference on Learning Representations, 2021. 2 +[38] Jerry Ngo, Swami Sankaranarayanan, and Phillip Isola. Is clip fooled by optical illusions? International Conference on Learning Representations, 2023. 2, 4 +[39] David Novick and Akiyoshi Kitaoka. The confetti illusion. Journal of Illusion, 2, 2021. 7 +[40] Xavier Otazu, C Alejandro Parraga, and Maria Vanrell. Toward a unified chromatic induction model. Journal of Vision, 10(12):5-5, 2010. 5, 6, 7 +[41] Jiang Qin, Kai Wang, Bin Zou, Lamei Zhang, and Joost van de Weijer. Conditional diffusion model with spatial-frequency refinement for sar-to-optical image translation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3 +[42] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2023. 5 +[43] Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. International Conference on Machine Learning, 2015. 2 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[45] Aniket Roy, Anirban Roy, Soma Mitra, and Kuntal Ghosh. Bri3l: A brightness illusion image dataset for identification and localization of regions of illusory perception. Proceedings of the IEEE International Conference on Image Processing, 2023. 5 +[46] Aniket Roy, Anirban Roy, Soma Mitra, and Kuntal Ghosh. Bri3l: A brightness illusion image dataset for identification + +and localization of regions of illusory perception. Proceedings of the IEEE International Conference on Image Processing, 2024. 3, 8 +[47] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. Illusionvqa: A challenging optical illusion dataset for vision language models. First Conference on Language Modeling, 2024. 2, 5 +[48] Arthur G. Shapiro and Dejan Todorovic. The Oxford Compendium of Visual Illusions, chapter 22. Oxford University Press, 2017. Figure 3.2. 7 +[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3 +[50] Eric D Sun and Ron Dekel. Imagenet-trained deep neural networks exhibit illusion-like response to the scintillating grid. Journal of Vision, 21(11):15-15, 2021. 2 +[51] Chuanming Tang, Kai Wang, and Joost van de Weijer. Iterinv: Iterative inversion for pixel-level t2i models. Neurips 2023 workshop on Diffusion Models, 2023. 3 +[52] Oguzhan Ulucan, Diclehan Ulucan, and Marc Ebner. Investigating color illusions from the perspective of computational color constancy. arXiv preprint arXiv:2312.13114, 2023. 2 +[53] Clinton Wang and Polina Golland. Interpolating between images with diffusion models. ICML 2023 Workshop DeployableGenerativeAI, 2023. 3 +[54] Kai Wang, Fei Yang, Shiqi Yang, Muhammad Atif Butt, and Joost van de Weijer. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 2023. 3 +[55] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 5 +[56] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. Annual Meeting of the Association for Computational Linguistics, 2023. 7 +[57] Eiji Watanabe, Akiyoshi Kitaoka, Kiwako Sakamoto, Masaki Yasugi, and Kenta Tanaka. Illusory motion reproduced by deep neural networks trained for prediction. Frontiers in psychology, 9:340023, 2018. 2 +[58] Robert Max Williams and Roman V Yampolskiy. Optical illusions images dataset. arXiv preprint arXiv:1810.00415, 2, 2018. 2, 4 +[59] Hongtao Zhang, Zhen Li, and Shinichi Yoshida. Muller-layer illusion is replicated by higher layer of pre-trained deep neural network for object recognition. International Symposium on Computational Intelligence and Industrial Applications, 2022. 2 +[60] Yichi Zhang, Jiayi Pan, Yuchen Zhou, Rui Pan, and Joyce Chai. Grounding visual illusions in language: Do vision-language models perceive illusions like humans? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5718-5728, 2023. 2, 4, 5 + +[61] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Tiviatis Sim, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. European Conference on Computer Vision, 2024. 5 \ No newline at end of file diff --git a/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/images.zip b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a35081e9e8f047eff936774e5efe613fd664724e --- /dev/null +++ b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d268adc29b94a0f3827786a1f577e89fcafdd948915209adf070e4f91a99a315 +size 602471 diff --git a/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/layout.json b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8d71ad67dc9c814e746ecc353901524da81a07a4 --- /dev/null +++ b/2025/The Art of Deception_ Color Visual Illusions and Diffusion Models/layout.json @@ -0,0 +1,10501 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "text", + "content": "The Art of Deception: Color Visual Illusions and Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 142, + 248, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 142, + 248, + 156 + ], + "spans": [ + { + "bbox": [ + 120, + 142, + 248, + 156 + ], + "type": "text", + "content": "Alexandra Gomez-Villa" + }, + { + "bbox": [ + 120, + 142, + 248, + 156 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 157, + 277, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 157, + 277, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 157, + 277, + 171 + ], + "type": "text", + "content": "Bartłomiej Twardowski1,2,4 Jesus Malo3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 280, + 143, + 490, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 143, + 490, + 157 + ], + "spans": [ + { + "bbox": [ + 280, + 143, + 490, + 157 + ], + "type": "text", + "content": "Kai Wang" + }, + { + "bbox": [ + 280, + 143, + 490, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 280, + 143, + 490, + 157 + ], + "type": "text", + "content": " C.Alejandro Parraga" + }, + { + "bbox": [ + 280, + 143, + 490, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 299, + 157, + 547, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 157, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 299, + 157, + 547, + 172 + ], + "type": "text", + "content": "Javier Vazquez-Corral" + }, + { + "bbox": [ + 299, + 157, + 547, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 299, + 157, + 547, + 172 + ], + "type": "text", + "content": " Joost van de Weijer" + }, + { + "bbox": [ + 299, + 157, + 547, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 184, + 506, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 184, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 104, + 184, + 506, + 199 + ], + "type": "text", + "content": "1Computer Vision Center, Spain 2Universitat Autònoma de Barcelona, Spain" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 154, + 199, + 456, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 199, + 456, + 213 + ], + "spans": [ + { + "bbox": [ + 154, + 199, + 456, + 213 + ], + "type": "text", + "content": "3Universitat de València, Spain 4IDEAS-NCBR, Poland" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 59, + 222, + 218, + 327 + ], + "blocks": [ + { + "bbox": [ + 59, + 222, + 218, + 327 + ], + "lines": [ + { + "bbox": [ + 59, + 222, + 218, + 327 + ], + "spans": [ + { + "bbox": [ + 59, + 222, + 218, + 327 + ], + "type": "image", + "image_path": "da5bb77c7e1365f142870551f9a2b4773119ff861ddce6873d47590b2d8acde5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 223, + 222, + 375, + 327 + ], + "blocks": [ + { + "bbox": [ + 223, + 222, + 375, + 327 + ], + "lines": [ + { + "bbox": [ + 223, + 222, + 375, + 327 + ], + "spans": [ + { + "bbox": [ + 223, + 222, + 375, + 327 + ], + "type": "image", + "image_path": "18b74c8002a322e16a74b95332a134f0e2ae997c11b02e0086ff2054044336ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 388, + 224, + 533, + 323 + ], + "blocks": [ + { + "bbox": [ + 388, + 224, + 533, + 323 + ], + "lines": [ + { + "bbox": [ + 388, + 224, + 533, + 323 + ], + "spans": [ + { + "bbox": [ + 388, + 224, + 533, + 323 + ], + "type": "image", + "image_path": "f07cabcd244c1607d3a1b56a92a9bd9834098043364b6cfa8b1e652534001d84.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 60, + 325, + 220, + 439 + ], + "blocks": [ + { + "bbox": [ + 60, + 325, + 220, + 439 + ], + "lines": [ + { + "bbox": [ + 60, + 325, + 220, + 439 + ], + "spans": [ + { + "bbox": [ + 60, + 325, + 220, + 439 + ], + "type": "image", + "image_path": "f053106e4ace068654512d02161eaf41a3641fa67e3e6343662ca70695b54dda.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 446, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 55, + 446, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 555, + 502 + ], + "type": "text", + "content": "Figure 1. Observation: Denoising Diffusion Implicit Models (DDIM) have human-like visual illusions. Application: text-to-image generated visual illusions. Examples (a) and (b) show how the responses to physically equal patches shift differently in the path to the latent space. In contrast, (c) shows how physically different stimuli get more similar along the path, achieving color constancy. The observation of this human-like behavior allows us to propose the use of text-to-image models to generate images in which physically identical patches are perceived differently (examples d, e, f). We highly recommend to watch the illusions on a computer screen." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 224, + 326, + 376, + 438 + ], + "blocks": [ + { + "bbox": [ + 224, + 326, + 376, + 438 + ], + "lines": [ + { + "bbox": [ + 224, + 326, + 376, + 438 + ], + "spans": [ + { + "bbox": [ + 224, + 326, + 376, + 438 + ], + "type": "image", + "image_path": "fd5ac4d23ce14017d6bfb159ce67c3c260b1a10d92d4ea975c2fdc975d46380b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 389, + 325, + 547, + 438 + ], + "blocks": [ + { + "bbox": [ + 389, + 325, + 547, + 438 + ], + "lines": [ + { + "bbox": [ + 389, + 325, + 547, + 438 + ], + "spans": [ + { + "bbox": [ + 389, + 325, + 547, + 438 + ], + "type": "image", + "image_path": "4bb4104fef48fbf94c93049adbf74d9a4876eaa97d60bb6ce968f0dbe0225a0d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 151, + 509, + 200, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 509, + 200, + 521 + ], + "spans": [ + { + "bbox": [ + 151, + 509, + 200, + 521 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 534, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 534, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 534, + 297, + 714 + ], + "type": "text", + "content": "Visual illusions in humans arise when interpreting out-of-distribution stimuli: if the observer is adapted to certain statistics, perception of outliers deviates from reality. Recent studies have shown that artificial neural networks (ANNs) can also be deceived by visual illusions. This revelation raises profound questions about the nature of visual information. Why are two independent systems, both human brains and ANNs, susceptible to the same illusions? Should any ANN be capable of perceiving visual illusions? Are these perceptions a feature or a flaw? In this work, we study how visual illusions are encoded in diffusion models. Remarkably, we show that they present human-like brightness/color shifts in their latent space. We use this fact to demonstrate that diffusion models can predict visual illusions. Furthermore, we also show how to gen" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 510, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 558 + ], + "type": "text", + "content": "erate new unseen visual illusions in realistic images using text-to-image diffusion models. We validate this ability through psychophysical experiments that show how our model-generated illusions also fool humans." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 568, + 395, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 568, + 395, + 581 + ], + "spans": [ + { + "bbox": [ + 314, + 568, + 395, + 581 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 589, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 555, + 673 + ], + "type": "text", + "content": "Visual illusions (VIs) serve as a fascinating window into the complex workings of human visual perception, where our interpretation of reality diverges significantly from physical measurements. These phenomena can have a profound social impact, as demonstrated by viral phenomena such as \"The Dress\" [9], which sparked global discussions on the nature of perception. Beyond their intrinsic scientific value," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 329, + 683, + 525, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 683, + 525, + 694 + ], + "spans": [ + { + "bbox": [ + 329, + 683, + 525, + 694 + ], + "type": "text", + "content": "Corresponding Author (☑): Kai Wang (kwang@cvc.uab.es)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "type": "text", + "content": "Code is available at https://alviur.github.io/color-illusion-diffusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18642" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "VIs offer a unique lens through which we can examine and compare human and machine perception [13, 33]. The emergence of diffusion models (DMs) as state-of-the-art generative models has revolutionized the field of artificial intelligence, demonstrating unprecedented capabilities in image synthesis and manipulation. Their denoising-based architecture, which gradually transforms random noise into coherent images, presents a novel paradigm for studying visual processing that mirrors the lossy and noisy process of human perception. This parallel makes them particularly interesting for investigating VIs, as both systems must navigate the complex relationship between image statistics and perceptual interpretation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 231, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 295, + 399 + ], + "type": "text", + "content": "The study of VIs in artificial neural networks (ANNs) has a rich history. Early work demonstrated that a multilayer perceptron could replicate brightness illusions [8], while subsequent research by Gomez-Villa et al. [12] showed that convolutional neural networks (CNNs) could mirror human perception in lightness and color illusions. Concurrent developments included Watanabe et al.'s [57] demonstration of motion VI replication in video-prediction CNNs and the publication of the first visual illusion dataset [58]. This foundation has spawned diverse investigations into the ANN-VI relationship in multiple domains: image completion [19, 50], brightness/color perception [13, 17, 24, 32, 33, 52], contrast sensitivity [1, 30], shape perception [37, 59], and vision-language models [6, 38, 47, 60]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 403, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 714 + ], + "type": "text", + "content": "In this work, we investigate how human-like perceptual phenomena naturally emerge in the encoding process of DMs. Our key finding reveals that the intermediate steps of diffusion models exhibit trends that align remarkably with human perception. Specifically, when analyzing the trajectory of an inverted image through the diffusion process, we observe that the progressive changes in lightness and color properties closely mirror human perceptual responses. Prior research has demonstrated that Gaussianization flows exhibit interesting properties when processing samples that deviate from their training data statistics. When trained on specific sample distributions, these flows can produce unusual representations for out-of-distribution samples [27, 43] - a characteristic that has proven valuable for measuring distribution divergence [29]. Intriguingly, this phenomenon parallels human visual perception, where shifts in image statistics can lead to illusory perception of spatio-chromatic patterns, a behavior that has been explained through manifold equalization [26, 28]. We argue that image samples that do not follow the statistics of natural images are not properly gaussianized by DMs. In this way, linking defective gaussianization and illusory perception of unusual samples, it makes sense that intermediate representations in DMs may be correlated with the shifts of these samples for human observers. The reported alignment suggests that the iterative denoising in diffusion models may" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "inadvertently capture fundamental aspects of human visual perception, particularly in contexts where our perception diverges from physical reality. By examining these intermediate representations, we provide novel insights into the nature of visual illusions and the unexpected parallels between human perception and the diffusion process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 144, + 544, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 544, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 544, + 156 + ], + "type": "text", + "content": "Contributions: The main contributions of this work are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 156, + 553, + 346 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 156, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 215 + ], + "type": "text", + "content": "- We report an empirical observation of the behavior of diffusion models: progressive changes in lightness, hue and saturation of image regions in the forward/inverse transform to the latent representation replicate the shifts of human perception of these regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 216, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 216, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 216, + 553, + 275 + ], + "type": "text", + "content": "- We derive a novel approach to using diffusion models to model human vision. We show that deep autoencoders can be effectively employed as vision models by performing a diffusion inversion process and measuring perception in the intermediate noise space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 276, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 276, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 553, + 311 + ], + "type": "text", + "content": "- We demonstrate the human perception replication capabilities of DM in many cases, including VI datasets, popular published VI works, and natural images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 346 + ], + "type": "text", + "content": "- Leveraging the perceptual capabilities of diffusion models, we propose a novel method for generating brightness/color VIs using text-to-image models." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 357, + 399, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 357, + 399, + 369 + ], + "spans": [ + { + "bbox": [ + 314, + 357, + 399, + 369 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "content": "Replication of Visual Illusions. Replication of visual illusions has been made with several methods, but all of them can be roughly classified into the following categories:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 415, + 553, + 594 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 415, + 553, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 553, + 450 + ], + "type": "text", + "content": "- Image restoration methods [12, 13, 24, 30, 33]: These use pre-trained image restoration backbones. They process VIs and measure the output image to measure perception." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 450, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 450, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 314, + 450, + 553, + 509 + ], + "type": "text", + "content": "- Effect bias methods [1]: These works are inspired by neuroscience measurement of stimuli. Pretrained architectures are used to process VIs, then the response in the internal activations of the ANN is used to define if there is a visual illusion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 510, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 553, + 545 + ], + "type": "text", + "content": "- Likelihood method [17]: In this approach, the most probable perceived intensity of each pixel in an image is estimated using an invertible flow trained on natural images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 546, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 546, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 314, + 546, + 553, + 594 + ], + "type": "text", + "content": "- Language-guided methods [38, 60]: Language is used as a vehicle to measure perception. These works use vision-language models as test subjects and ask them questions regarding the perception of the image." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Our approach differs from previous methods in the following: 1) we do not measure perception in the output of the model. Instead, we measure the VI in the intermediate steps of image inversion process. 2) We use a deep autoencoder; previous approaches in this category claim that deep autoencoders suffer from less perceptual bias and use shallow architectures for replication experiments. Furthermore, we are able to replicate human perception in high resolution images (as big as DMs are able to manage) and in natural images. These setups were not possible in previous works." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18643" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "content": "Generation of Visual Illusions. Although generative models have been used to synthesize visual illusions [14], DMs has only recently been proposed as the generative engine. In [6], high-level illusions (overlay effects) are generated by mixing a set of images and transformations together. Following works [10, 11] were able to create more general illusions such as rotations, color inversions, and hybrid images. The work most similar to ours is by Roy et al. [46], who used stable diffusion (SD) to generate brightness visual illusions through keyword prompts. However, their approach neither explicitly optimizes perceptual effects nor specifies regions for perception replication. Consequently, illusions may appear randomly in the generated image with unknown localization. Additionally, their generated images are limited to conventional brightness illusions (gratings or high-frequency black and white patterns) and cannot incorporate visual illusions into natural images. In contrast, our work allows visual illusions within generated natural images as well as in typical high-frequency patterns. We explicitly generate the perceptual effect in a designated image region and can control the image content through text prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 335, + 117, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 117, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 117, + 348 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 356, + 157, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 157, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 157, + 368 + ], + "type": "text", + "content": "3.1. Diffusion models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 373, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 493 + ], + "type": "text", + "content": "Diffusion models (DMs) are generative models that learn to gradually denoise images by reversing a fixed Markov chain of forward diffusion steps. The forward process systematically adds Gaussian noise to an image until it becomes pure noise, while the learned reverse process reconstructs the original image through iterative denoising. While traditional diffusion follows a stochastic sampling process, Denoising Diffusion Implicit Models (DDIM) [49] introduce a deterministic sampling approach that maintains high quality while enabling significantly fewer inference steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 494, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 592 + ], + "type": "text", + "content": "DDIM's deterministic nature enables a crucial capability: the ability to invert generated images back into their latent representations through a process known as DDIM inversion [49]. This inversion process allows us to map any image into the model's latent space and observe its trajectory through the diffusion process. Given an image " + }, + { + "bbox": [ + 55, + 494, + 296, + 592 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 494, + 296, + 592 + ], + "type": "text", + "content": ", DDIM inversion produces a sequence of intermediate representations " + }, + { + "bbox": [ + 55, + 494, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\{z_t\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 494, + 296, + 592 + ], + "type": "text", + "content": " according to:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 601, + 295, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 601, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 601, + 295, + 617 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {z _ {t}, t} (1)", + "image_path": "449313202665bd5f638f1d1f391ba64c0acd0655b1cb9e126c1c98038d26ae58.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "text", + "content": " represents the cumulative product of noise scheduling coefficients, and " + }, + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{z_t,t} = \\epsilon_{\\theta}(z_t,t)" + }, + { + "bbox": [ + 55, + 627, + 296, + 664 + ], + "type": "text", + "content": " is the learned noise prediction network. The reverse process follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 671, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 671, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 51, + 671, + 295, + 714 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(\\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {z _ {t} , t}}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\epsilon_ {\\theta} ^ {z _ {t}, t} \\tag {2}", + "image_path": "947a785dc3f9f375b5e52bd3fd54d428469bb014400df9435b3f34aa36803094.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "DDIM inversion has emerged as a powerful tool for various image manipulation tasks. Its ability to faithfully reconstruct latent representations has enabled applications such as semantic image editing [31, 35, 36, 51, 54], style transfer [25, 41], image interpolation [53], and controlled image generation through classifier-free guidance [18]. In our work, we exploit the deterministic nature of DDIM inversion to systematically study how visual illusions are processed through the diffusion trajectory." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 189, + 545, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 545, + 202 + ], + "type": "text", + "content": "3.2. DDIM inversion may replicate human vision" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 207, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 291 + ], + "type": "text", + "content": "The wholly-empirical evolutionary paradigm [20] suggests that human vision is shaped by statistical regularities in natural scenes, where our visual system evolved to interpret ambiguous stimuli based on their likelihood. We leverage this insight to study how diffusion models, trained for natural images, process illusory stimuli, which are not in the distribution of natural images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 292, + 556, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 556, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 556, + 507 + ], + "type": "text", + "content": "Motivation. In Fig. 2, we present an illustrative experiment using Stable Diffusion [44] (extra results with an image space model in the supplementary material). We perform DDIM inversion on the canonical visual illusion known as Brightness Contrast [5]. DDIM inversion puts the image back into the latent space. This illusion presents two gray squares of identical intensity (indicated by red markers) against white and black backgrounds, which leads humans to perceive the left square as darker than the right one. The first row (left to right) shows the inversion1 for 0, 3, 10, and 20 steps, respectively. As the zoomed-in bar shows, the initially equal-intensity regions progressively develop an intensity difference that aligns with human perception2 (left target appearing darker than right). The second row shows the progressive Gaussianization of the latent representation. For this image perception is replicated at an intermediate representation where gaussianization is not complete (since this stimulus is not in the distribution of natural images)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "type": "text", + "content": "Observation: emergence of perceptual biases. Our key observation is: when inverting an image through diffusion models the intermediate denoising steps progressively introduce color/brightness shifts that align with human perception. We hypothesize that this occurs because diffusion models, trained on natural images, map illusory stimuli onto the learned natural image manifold, resulting in systematic \"interpretation errors\" that mirror human perceptual biases." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": "Theoretical framework. Formally, let " + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "inline_equation", + "content": "G: \\mathcal{X} \\to \\mathcal{Z}" + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": " be the forward diffusion mapping from image space " + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": " to latent space " + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": ". For natural images " + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": " learns a mapping " + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "inline_equation", + "content": "z = G(x)" + }, + { + "bbox": [ + 313, + 604, + 555, + 665 + ], + "type": "text", + "content": " that transforms the image distribution into a Gaussian. However, when presented with an eventually" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 673, + 541, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 673, + 541, + 684 + ], + "spans": [ + { + "bbox": [ + 324, + 673, + 541, + 684 + ], + "type": "text", + "content": "we decode the latent representation to visualize it in image space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 684, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 684, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 684, + 554, + 713 + ], + "type": "text", + "content": "2In addition to the color/brightness effects, notice how the high-frequency content of the image is maintained; we explore this phenomenon further in supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18644" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 547, + 236 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 547, + 236 + ], + "type": "image", + "image_path": "2b799b5d302dfd681838acea957ac44f3c05f628c5c3f9847b010cdd24218d21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 292 + ], + "type": "text", + "content": "Figure 2. DDIM inversion of the Brightness Contrast illusion [5] using Stable Diffusion. Top row: Image-space visualization (decoded latents) showing (left) Original illusion with two identical gray squares (marked in red) against black and white backgrounds, and inversion results using 3, 10, and 20 steps. Bottom row: Histograms of the corresponding latent representations. The horizontal axis represents the value of the latent space feature. The model gradually reproduces the perceptual difference in brightness between the physically identical squares in a (not fully Gaussian) intermediate representation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "text", + "content": "illusory stimulus " + }, + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "inline_equation", + "content": "y \\notin \\mathcal{X}" + }, + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "text", + "content": ". Still, the mapping " + }, + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "inline_equation", + "content": "z' = G(y)" + }, + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "text", + "content": " attempts to project the illusion onto the learned natural image manifold, introducing systematic deviations. The intermediate states " + }, + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 56, + 297, + 296, + 489 + ], + "type": "text", + "content": " reveal how the model progressively \"corrects\" the illusion toward natural image statistics. We observe that these corrections correlate strongly with human perception, suggesting that both humans and diffusion models may operate on similar statistical principles when processing ambiguous visual input. This alignment between diffusion trajectories and human perception provides computational evidence for the wholly-empirical paradigm, demonstrating how statistical learning can lead to perceptual illusions without explicit encoding of perceptual rules. The systematic nature of these \"errors\" suggests they emerge from the same statistical regularities that shape human visual processing through evolution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 500, + 251, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 251, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 251, + 512 + ], + "type": "text", + "content": "3.3. Generating visual illusions with DMs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 519, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 567 + ], + "type": "text", + "content": "Based on the findings above, we develop a method to automatically generate brightness/color illusions with DMs. To do so, we need to evaluate whether a VI is present and then push the diffusion process toward the desired perception." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 568, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 652 + ], + "type": "text", + "content": "Previous works on VIs and DMs focus on high-level visual illusions, i.e. cognitive effects that measure if the image belongs to a certain class of prompt [6, 10, 11]. Here we focus on brightness/color VIs, which leads to the challenge of measuring effects in specific regions of the image. Here we answer the measurement questions proposed in previous works on brightness/color VI generation in this way:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Where to measure? This question separates the generation of high-level VIs from color/brightness VIs. Models aiming at high-level VIs that fool an ANN require to ask a classifier/vision-language model which class/prompt better describes the image [10, 38, 60]. In contrast, color VIs ap" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 297, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 380 + ], + "type": "text", + "content": "pear in specific regions of the image, therefore, they need to be measured at those points. In this work, we specifically define regions (here targets) in which we want a perceptual effect to appear. These targets are our measurement points, and we guarantee that in these regions, we can track, measure, and optimize perceptual effects. Failing to define a target leads to failure, as reported in [58]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "text", + "content": "What to measure? During image generation, we will measure the intensity values of the intermediate steps " + }, + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "text", + "content": ". Even though " + }, + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "text", + "content": " does not contain the final perception bias " + }, + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "inline_equation", + "content": "(z_{T})" + }, + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "text", + "content": ", we can push the generative process towards a desired perceptual effect using small changes in the intermediate representation as explained in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 555, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 573 + ], + "type": "text", + "content": "How to measure? In the motivating example of section 3.2 (Fig. 2), the model replicates human perception if the target shifts in the same direction as humans perceive. However, in an optimization procedure, we need to measure how \"good\" the illusion is; in other words, we need to quantify the effect. Since we are concerned about brightness and color VI, we will measure the mean intensity per channel of each target. More precisely, given an image " + }, + { + "bbox": [ + 313, + 453, + 555, + 573 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 453, + 555, + 573 + ], + "type": "text", + "content": ", and the target region " + }, + { + "bbox": [ + 313, + 453, + 555, + 573 + ], + "type": "inline_equation", + "content": "r \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 453, + 555, + 573 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 453, + 555, + 573 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 453, + 555, + 573 + ], + "type": "text", + "content": ", we compute the mean intensity in that target region as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 378, + 581, + 555, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 581, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 378, + 581, + 555, + 609 + ], + "type": "interline_equation", + "content": "m _ {i n t} (I, r) = \\frac {1}{M} \\sum_ {i \\in r} I \\left(r _ {i}\\right) \\tag {3}", + "image_path": "dc021cd201b71392dbbf9eaf65b21fb20934ee3229fa6723ca9ec3bdc9eea1b9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "text", + "content": " denotes a specific pixel " + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "text", + "content": " in target region " + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "text", + "content": " stands for the number of pixels in the target region, and " + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "inline_equation", + "content": "m_{int}" + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "text", + "content": " is measured in each channel. Hence, the perceived intensity is an RGB triplet " + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "inline_equation", + "content": "(m_{int}^R, m_{int}^G, m_{int}^B)" + }, + { + "bbox": [ + 313, + 619, + 555, + 669 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 673, + 531, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 531, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 531, + 686 + ], + "type": "text", + "content": "3.3.1. Visual perception inside generation pipeline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Since text-to-image (T2I) diffusion models often struggle to perfectly align with textual prompts, a variety of aug-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18645" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 287, + 235 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 287, + 235 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 287, + 235 + ], + "type": "image", + "image_path": "58ab821c681add6ca3fd7ee579250a40d4bd6a93cc00345f3c7d35a3221fc959.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "type": "text", + "content": "Figure 3. Overview of our visual illusion generation pipeline. The process modifies noisy latent representations, " + }, + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "type": "text", + "content": ", through a custom loss function, " + }, + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "type": "text", + "content": ", that guides the generation toward perceptually ambiguous outputs. Note that targets are regions that manually selected by the user as inputs. The terms " + }, + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "type": "inline_equation", + "content": "D_{p1}" + }, + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "type": "inline_equation", + "content": "D_{p2}" + }, + { + "bbox": [ + 55, + 242, + 295, + 308 + ], + "type": "text", + "content": " are the user desired intensity for target T1 and T2 respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 312, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 312, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 295, + 528 + ], + "type": "text", + "content": "mentation methods have been developed to enhance the fidelity of T2I generation. These methods typically focus on optimizing noisy latent representations in each timestep, a strategy that has proven effective in refining generation outcomes [7, 42, 55, 61]. A common approach involves the design of an update loss function, " + }, + { + "bbox": [ + 56, + 312, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 56, + 312, + 295, + 528 + ], + "type": "text", + "content": ", which is tailored based on the layout of the noisy latent or the characteristics of the cross-attention maps. This loss is then used to guide the backpropagation process, updating the noisy latent in a manner that more accurately reflects the intended textual descriptions, thereby enhancing the overall quality of the generated images. This process can be formulated as " + }, + { + "bbox": [ + 56, + 312, + 295, + 528 + ], + "type": "inline_equation", + "content": "z_{t}^{\\prime} = z_{t} - \\alpha \\cdot \\nabla_{z_{t}}\\mathcal{L}" + }, + { + "bbox": [ + 56, + 312, + 295, + 528 + ], + "type": "text", + "content": ". In this paper, we diverge from the conventional use of optimization techniques aimed at aligning T2I diffusion models with textual semantics. Instead, we leverage these techniques to create visual perception effects (see Fig. 3). We use an image space DM (DeepFloyd [2]) to avoid the use of the decoder in the backpropagation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": "First, we define the original target image as " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "O(r)" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "r \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": ". This image only contains the input-defined targets, that is, " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "O(r)" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": " has the original value given to the targets by the user on each color channel. Then, given an intermediate image " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": " and a list of " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": " target regions " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "\\{r^p\\}_{p=1}^R" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": ", our process to generate a VI includes the addition of the targets to the image in the desired region3, i.e. we compute " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "z_{t\\_target}(r_i^p) = z_t(r_i^p) + O(r_i^p)" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": ", for all pixels " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": " inside the region " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "r^p" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "z_{t\\_target} = z_t" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": " for any pixel outside this target region. Finally, the following optimization term is applied " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": " times —per diffusion step " + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 528, + 295, + 659 + ], + "type": "text", + "content": " — to guide the generation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 666, + 294, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 666, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 132, + 666, + 294, + 678 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\gamma \\mathcal {L} _ {V I} + \\beta \\mathcal {L} _ {s i m} \\tag {4}", + "image_path": "07a3e9f8f8b06bfbbcc3ddcf77211d196d626102821d5b0a0407bcb00f1ed4dc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 685, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 685, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 685, + 295, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 685, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VI}" + }, + { + "bbox": [ + 55, + 685, + 295, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 685, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sim}" + }, + { + "bbox": [ + 55, + 685, + 295, + 696 + ], + "type": "text", + "content": " measure the illusion strength and the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "similarity between targets and background respectively, and " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " are hyperparameters that control the relative weight of these measures. Specifically:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 111, + 520, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 111, + 520, + 124 + ], + "spans": [ + { + "bbox": [ + 314, + 111, + 520, + 124 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 111, + 520, + 124 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VI}" + }, + { + "bbox": [ + 314, + 111, + 520, + 124 + ], + "type": "text", + "content": " is a perceptual loss on the target intensities:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 340, + 133, + 554, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 133, + 554, + 167 + ], + "spans": [ + { + "bbox": [ + 340, + 133, + 554, + 167 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {V I} = \\sum_ {c \\in R, G, B} \\sum_ {r ^ {p}} ^ {R} \\mathcal {L} _ {\\text {P e r c e p t}} \\left(z _ {t _ {\\text {t a r g e t}}}, k _ {c} ^ {p}; r ^ {p}\\right) \\tag {5}", + "image_path": "ee6ec6187a7d6bc2ebdcee6f9807bbb89444fc63ff515b22d7b1bd20f2f090a7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "spans": [ + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "inline_equation", + "content": "r^p" + }, + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "text", + "content": " is a specific target, " + }, + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "inline_equation", + "content": "k_c^p" + }, + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "text", + "content": " is the desired intensity for target " + }, + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "inline_equation", + "content": "r^p" + }, + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "text", + "content": " on color channel " + }, + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 322, + 177, + 554, + 201 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 212, + 553, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 212, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 328, + 212, + 553, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {P e r c e p t u a l}} \\left(z _ {t _ {\\text {t a r g e t}}}, k _ {c}; r\\right) = \\left\\| m _ {\\text {i n t}} \\left(z _ {t _ {\\text {t a r g e t}}}, r\\right) - k _ {c} \\right\\| \\tag {6}", + "image_path": "e219a6429d4e1fc82f9694c0428f96a25f7ba4aad9b967f03143de2b462b7f0d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 237, + 445, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 237, + 445, + 248 + ], + "spans": [ + { + "bbox": [ + 322, + 237, + 445, + 248 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 322, + 237, + 445, + 248 + ], + "type": "inline_equation", + "content": "m_{int}" + }, + { + "bbox": [ + 322, + 237, + 445, + 248 + ], + "type": "text", + "content": " as defined in Eq. 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 249, + 554, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 249, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 314, + 249, + 554, + 296 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 249, + 554, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sim}" + }, + { + "bbox": [ + 314, + 249, + 554, + 296 + ], + "type": "text", + "content": " is a similarity term. This term aims at avoiding the region from appearing \"out of place\". In particular, it looks at harmonizing the targets with the background-generated image:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 306, + 554, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 306, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 365, + 306, + 554, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s i m} = \\sum_ {p = 1} ^ {R} M S E \\left(z _ {t} \\left(r ^ {p}\\right), O \\left(r ^ {p}\\right)\\right) \\tag {7}", + "image_path": "8221fb035ac35cc09234965ef56bbbda0d20991b1b1426fb7c0801302fa95208.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 350, + 430, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 430, + 364 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 430, + 364 + ], + "type": "text", + "content": "4. Experimental setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 371, + 376, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 371, + 376, + 381 + ], + "spans": [ + { + "bbox": [ + 314, + 371, + 376, + 381 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 388, + 554, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 388, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 388, + 554, + 435 + ], + "type": "text", + "content": "- Brightness illusions dataset: The " + }, + { + "bbox": [ + 314, + 388, + 554, + 435 + ], + "type": "inline_equation", + "content": "BRI3L" + }, + { + "bbox": [ + 314, + 388, + 554, + 435 + ], + "type": "text", + "content": " dataset [45] consists of 22366 grayscale images spanning five types of brightness illusions. Each image is annotated with binary masks indicating regions of illusory perception." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 435, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 435, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 314, + 435, + 554, + 591 + ], + "type": "text", + "content": "- Visual Illusion VQA Datasets: Several datasets have been developed to evaluate vision-language models' understanding of VIs: IllusionVQA [47] contains 435 illusory images with multiple-choice questions across 12 categories, plus 1000 impossible object image pairs for localization testing. GVIL [60] features 100 base illusion images expanded to 1600 instances covering five illusion categories, with four evaluation tasks: same-difference QA, referential QA, attribute QA, and referential localization. HallusionBench [15] comprises 346 images and 1129 questions designed to test image-context reasoning and diagnose both VIs and language hallucinations. We use the color category in all the VQA datasets." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 600, + 395, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 600, + 395, + 611 + ], + "spans": [ + { + "bbox": [ + 314, + 600, + 395, + 611 + ], + "type": "text", + "content": "4.2. Benchmarks" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 617, + 554, + 713 + ], + "type": "text", + "content": "- Classical vision science models: These approaches are based in vision science and need to be calibrated for each image to fully exploit their capabilities. In this paper we use the parameters set by Betz et al. [3]. These parameters constitute a fair guideline for these methods, as it is not possible to fine-tune them for thousands of images. We use ODOG (Oriented Difference of Gaussians) [4] and the Chromatic Induction Wavelet Model (CIWaM) [40]" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 285, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 285, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 285, + 713 + ], + "type": "text", + "content": "3Note that this addition results in the perceived color (see sect. 3.2)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18646" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 293, + 156 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 293, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 293, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 293, + 107 + ], + "type": "text", + "content": "- Neural networks-base models: We compare against previous image restoration methods that proven ability in visual illusion detection: DN-NET[13] and Restorenet[30]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 293, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 293, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 293, + 156 + ], + "type": "text", + "content": "- Variational autoencoder (VAE): We present the results of only considering the VAE of SD prior to the DDIM inversion process —to further demonstrate DDIM is the responsible for the alignment with human perception." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 164, + 114, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 114, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 114, + 176 + ], + "type": "text", + "content": "4.3. Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 182, + 294, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 294, + 253 + ], + "type": "text", + "content": "While visual illusion replication results have traditionally been presented through profile plots comparing model outputs with human perception along image cross-sections, we leverage recent datasets to provide quantitative metrics. This approach complements classical visualization methods with rigorous numerical evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 255, + 294, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 294, + 291 + ], + "type": "text", + "content": "For the BRI3L dataset, we compute the mean intensity difference, " + }, + { + "bbox": [ + 55, + 255, + 294, + 291 + ], + "type": "inline_equation", + "content": "\\Delta I" + }, + { + "bbox": [ + 55, + 255, + 294, + 291 + ], + "type": "text", + "content": ", between the model's output and the input image within each segmented region:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 300, + 294, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 300, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 294, + 327 + ], + "type": "interline_equation", + "content": "\\Delta I = \\frac {1}{| M |} \\sum_ {i \\in r} \\left| I _ {\\text {o u t}} (r) - I _ {\\text {i n}} (r) \\right| \\tag {8}", + "image_path": "f5115c17900641dd9e6c33578126069b2efd8aa87d73d0a01e40fe2b39c61031.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "text", + "content": " represents the size of the segmentation mask, and " + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "inline_equation", + "content": "I_{out}" + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "text", + "content": " are the output and input intensities respectively. We consider the model's response to be aligned with human perception when " + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\Delta I < \\tau * I_{int}" + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "inline_equation", + "content": "I_{int}" + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "text", + "content": " represents the mean intensity of the input image in the target region (calculated using Eq. 3). Since BRI3L aims to detect regions that become darker, the metric becomes more harder as the value of " + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 337, + 295, + 445 + ], + "type": "text", + "content": " decreases, requiring the model to achieve greater darkness in the target regions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 445, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 529 + ], + "type": "text", + "content": "For the VQA-based datasets (IllusionVQA, GVIL, and HallusionBench), we manually create segmentation masks and consider the answer of the model aligned with human perception if it matches the expected intensity for regions referenced in the question-answer pairs. The Perception Accuracy Score (PAS) is the percentage of images having the correct perceptual predictions in each dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 537, + 138, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 138, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 138, + 550 + ], + "type": "text", + "content": "4.4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 555, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 628 + ], + "type": "text", + "content": "We conducted two sets of experiments to evaluate our method's ability to: (1) replicate existing VIs across multiple datasets, and (2) generate novel VIs that can fool human perception. For each experiment, we present qualitative results zooming in on the region of the perceptual effect and quantitative evaluations using the proposed metrics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "- Replication of visual illusions: We test our model's ability to replicate brightness (BRI3L) and color (Illusion-VQA, GVIL, and HallusionBench) illusions by comparing the intensity differences between input and output images within annotated segmentation masks. For BRI3L, we evaluate performance across three " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " levels (0.8, 0.9, 1.0) to assess the robustness of illusion replication. In this" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 331, + 96, + 535, + 215 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 92 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 92 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 92 + ], + "type": "text", + "content": "Table 1. Perceptual Accuracy Score (PAS) in BRI3L for different methods and thresholds" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 96, + 535, + 215 + ], + "lines": [ + { + "bbox": [ + 331, + 96, + 535, + 215 + ], + "spans": [ + { + "bbox": [ + 331, + 96, + 535, + 215 + ], + "type": "table", + "html": "
Method / Threshold0.80.91.0
ODOG [4]36.5857.0972.44
CIWAM [40]5.6918.2356.22
DNnet [12]0.000.0071.09
RestoreNet [13]0.000.0074.95
VAE0.000.0873.18
DDIM 5 steps44.0291.4399.88
DDIM 10 steps84.1798.85100.00
", + "image_path": "d8f4e79bfbdfa82798eabb9d3cb5ac8b1ecaab0628d60c9c86ec0feba8d4dee7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 230, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 230, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 322, + 230, + 553, + 266 + ], + "type": "text", + "content": "experiment we present results using SD 1.4 (latent space DM) but additional results with DeepFloyd (image space DM) are provided in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "text", + "content": "- Generation of visual illusions: In this experiment, we check whether we can generate VIs in images that fall inside the distribution of the T2I model. We use guidance scale " + }, + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "inline_equation", + "content": "= 10" + }, + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "text", + "content": " (we want to generate images that follow the prompt), " + }, + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "inline_equation", + "content": "\\gamma = 0.5" + }, + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "inline_equation", + "content": "\\beta = 1.0" + }, + { + "bbox": [ + 314, + 266, + 553, + 386 + ], + "type": "text", + "content": ". Additionally, here we use S2 of DeepFloyd to generate higher-resolution versions of our generated visual illusions. For flat intensity targets (non-gradients), we intentionally paste the targets in the output of S2 as the superresolution model may change the content of the targets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 397, + 367, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 397, + 367, + 410 + ], + "spans": [ + { + "bbox": [ + 314, + 397, + 367, + 410 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 417, + 474, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 474, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 474, + 430 + ], + "type": "text", + "content": "5.1. Replication of visual illusions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 436, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 640 + ], + "type": "text", + "content": "Qualitative results. We present qualitative results demonstrating our method's ability to replicate human visual perception across diverse scenarios in Fig. 4. Each comparison shows the original input image (left) and its corresponding DDIM inversion after 5 steps (right), with zoomed inset regions highlighting the areas where perceptual effects occur. The top row showcases our method's effectiveness on classical visual illusions, including Hermann grid and grating induction patterns. The bottom panels demonstrate the model's capability to capture perceptual phenomena in both natural grayscale photographs and artistic images with different color schemes. The zoomed regions, emphasize physically identical targets that are perceived differently due to their surrounding context — a key characteristic of visual illusions that our model successfully replicates. Additional qualitative results showing our performance across a broader range of visual illusions are in the supplementary." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Quantitative results. Table 1 presents a comprehensive comparison of our DDIM-based approach against classical vision models and deep learning methods across multiple visual illusion datasets. The results demonstrate the consistency of the observation on diffusion models reported in Section 3.2 over a range of datasets and their superior" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18647" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 68, + 113, + 123 + ], + "blocks": [ + { + "bbox": [ + 66, + 68, + 113, + 123 + ], + "lines": [ + { + "bbox": [ + 66, + 68, + 113, + 123 + ], + "spans": [ + { + "bbox": [ + 66, + 68, + 113, + 123 + ], + "type": "image", + "image_path": "b611aea49589fd6f66b1108aa700f3e16e4366e2f6ca8111e51cbfd27d534348.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 122, + 69, + 167, + 123 + ], + "blocks": [ + { + "bbox": [ + 122, + 69, + 167, + 123 + ], + "lines": [ + { + "bbox": [ + 122, + 69, + 167, + 123 + ], + "spans": [ + { + "bbox": [ + 122, + 69, + 167, + 123 + ], + "type": "image", + "image_path": "5ff1e0979d122db9af8dc22424f2b3512a0d299a9f32c05a3e410108c0b82791.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 125, + 134, + 135 + ], + "lines": [ + { + "bbox": [ + 123, + 125, + 134, + 135 + ], + "spans": [ + { + "bbox": [ + 123, + 125, + 134, + 135 + ], + "type": "text", + "content": "a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 169, + 76, + 223, + 121 + ], + "blocks": [ + { + "bbox": [ + 169, + 76, + 223, + 121 + ], + "lines": [ + { + "bbox": [ + 169, + 76, + 223, + 121 + ], + "spans": [ + { + "bbox": [ + 169, + 76, + 223, + 121 + ], + "type": "image", + "image_path": "92b1e53973f4f6bf8d6c2e884c8e147cedd2c7b7ec311701a85ede9af656ea05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 282, + 121 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 282, + 121 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 282, + 121 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 282, + 121 + ], + "type": "image", + "image_path": "cd8bb5fbe4c2bf96c67b75678c30b0a472f58757a4827ccb4655e35e9b612190.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 124, + 246, + 136 + ], + "lines": [ + { + "bbox": [ + 235, + 124, + 246, + 136 + ], + "spans": [ + { + "bbox": [ + 235, + 124, + 246, + 136 + ], + "type": "text", + "content": "b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 283, + 70, + 324, + 121 + ], + "blocks": [ + { + "bbox": [ + 283, + 70, + 324, + 121 + ], + "lines": [ + { + "bbox": [ + 283, + 70, + 324, + 121 + ], + "spans": [ + { + "bbox": [ + 283, + 70, + 324, + 121 + ], + "type": "image", + "image_path": "c055d17c3ff10faf73f30d7a1d0e42a2449cf21d576c1190af3336fa2e7a6b1f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 125, + 338, + 135 + ], + "lines": [ + { + "bbox": [ + 328, + 125, + 338, + 135 + ], + "spans": [ + { + "bbox": [ + 328, + 125, + 338, + 135 + ], + "type": "text", + "content": "c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 325, + 70, + 369, + 121 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 369, + 121 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 369, + 121 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 369, + 121 + ], + "type": "image", + "image_path": "129bc26c3ee619676a50a74f940dcfc798289677b332db4674c2fd478095fb80.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 370, + 79, + 412, + 121 + ], + "blocks": [ + { + "bbox": [ + 370, + 79, + 412, + 121 + ], + "lines": [ + { + "bbox": [ + 370, + 79, + 412, + 121 + ], + "spans": [ + { + "bbox": [ + 370, + 79, + 412, + 121 + ], + "type": "image", + "image_path": "300bbf70a7eb344cd8a75b026438aa9f710a39b02cc9a266595dc336a97af030.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 125, + 424, + 135 + ], + "lines": [ + { + "bbox": [ + 413, + 125, + 424, + 135 + ], + "spans": [ + { + "bbox": [ + 413, + 125, + 424, + 135 + ], + "type": "text", + "content": "d)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 414, + 79, + 459, + 121 + ], + "blocks": [ + { + "bbox": [ + 414, + 79, + 459, + 121 + ], + "lines": [ + { + "bbox": [ + 414, + 79, + 459, + 121 + ], + "spans": [ + { + "bbox": [ + 414, + 79, + 459, + 121 + ], + "type": "image", + "image_path": "be5912884070aade400b4944ca68c088a3c4bebfd6c7cc75f2416ba0a0aaf6b6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 462, + 79, + 545, + 122 + ], + "blocks": [ + { + "bbox": [ + 462, + 79, + 545, + 122 + ], + "lines": [ + { + "bbox": [ + 462, + 79, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 462, + 79, + 545, + 122 + ], + "type": "image", + "image_path": "e60ae82537d228e22a5278c4cd36a45dc9c236de1b9020a57ad89ee5e3675fe2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 125, + 509, + 135 + ], + "lines": [ + { + "bbox": [ + 498, + 125, + 509, + 135 + ], + "spans": [ + { + "bbox": [ + 498, + 125, + 509, + 135 + ], + "type": "text", + "content": "e)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 69, + 142, + 215, + 223 + ], + "blocks": [ + { + "bbox": [ + 69, + 142, + 215, + 223 + ], + "lines": [ + { + "bbox": [ + 69, + 142, + 215, + 223 + ], + "spans": [ + { + "bbox": [ + 69, + 142, + 215, + 223 + ], + "type": "image", + "image_path": "af51c829bee1546b67540611cf18dc473c521345df79a290552f927b990ea825.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 226, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 555, + 270 + ], + "type": "text", + "content": "Figure 4. Qualitative replication of visual illusions (please enlarge display): For each item, the original image is on the left and its DDIM inversion on the right. The observation done in Fig. 2 is consistently reproduced. We zoom in on the region containing the illusion in each image pair. a) Barutan-seijin [21], b) Robot [22], c) Shiosai [23], d) Herman-grid [16], e) Grating induction [34], f) Bright room [48], g) Confetti illusion [39]. For more qualitative examples and results of classical visual illusions, see the supplementary material." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 215, + 143, + 360, + 223 + ], + "blocks": [ + { + "bbox": [ + 215, + 143, + 360, + 223 + ], + "lines": [ + { + "bbox": [ + 215, + 143, + 360, + 223 + ], + "spans": [ + { + "bbox": [ + 215, + 143, + 360, + 223 + ], + "type": "image", + "image_path": "dedbf501dd0576c9f1ea5671119eecabd1142b9f163980bcb89b8ba2d4c674fa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 372, + 145, + 451, + 224 + ], + "blocks": [ + { + "bbox": [ + 372, + 145, + 451, + 224 + ], + "lines": [ + { + "bbox": [ + 372, + 145, + 451, + 224 + ], + "spans": [ + { + "bbox": [ + 372, + 145, + 451, + 224 + ], + "type": "image", + "image_path": "51e76b6dd54e200cebe3c730db8a0104eacf30199a5ac3fa1f4946e986581139.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 456, + 145, + 534, + 224 + ], + "blocks": [ + { + "bbox": [ + 456, + 145, + 534, + 224 + ], + "lines": [ + { + "bbox": [ + 456, + 145, + 534, + 224 + ], + "spans": [ + { + "bbox": [ + 456, + 145, + 534, + 224 + ], + "type": "image", + "image_path": "99224664b909903de01abb5ea3eb175a10d1a33d96a8584e18ef82e4cda7f336.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 280, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 295, + 447 + ], + "type": "text", + "content": "performance in replicating illusions. Classical models like ODOG and CIWAM, while physiologically grounded, show limited performance and flexibility. These methods require careful parameter tuning for each specific type of visual illusion, as evidenced by their poor scores. For instance, ODOG can achieve " + }, + { + "bbox": [ + 54, + 280, + 295, + 447 + ], + "type": "inline_equation", + "content": "72.44\\%" + }, + { + "bbox": [ + 54, + 280, + 295, + 447 + ], + "type": "text", + "content": " accuracy on BRI3L but cannot generalize to color datasets, while applicable in color CIWAM shows poor performance. Previous deep learning approaches show consistent but modest performance in different datasets. Their uniform performance across sets suggests that they capture some general perceptual principles, but their relatively low accuracy indicates limitations in capturing human behavior. Our proposed method significantly outperforms all baseline approaches across all datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 579 + ], + "type": "text", + "content": "The substantial performance gap between our DDIM approach and a standard VAE (which shows near-zero performance on most datasets) highlights the crucial role of the diffusion process in capturing perceptual phenomena. These results demonstrate that our DDIM-based approach not only achieves state-of-the-art performance but also offers superior generalization across different types of visual illusions without requiring illusion-specific parameter tuning. The consistent high performance across all datasets suggests that the diffusion process naturally captures fundamental aspects of human vision." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 586, + 218, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 218, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 218, + 598 + ], + "type": "text", + "content": "5.2. Generation of Visual Illusions" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Qualitative results. Fig. 5 demonstrates visual illusions using a " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "6 \\times 6" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " pixel square target in the S1 stage of DeepFloyd (expanding to " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "24 \\times 24" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " after S2). The figure displays five different target intensities (shown as circles to the left of each image) with different prompts for each image (selected randomly from DiffusionDB [56]). Additional results are presented in Fig. 1. In contrast to previous approaches, our method can generate complex content that simultaneously satisfies both the text prompt and the perceptual conditions." + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 317, + 293, + 550, + 391 + ], + "blocks": [ + { + "bbox": [ + 316, + 278, + 552, + 289 + ], + "lines": [ + { + "bbox": [ + 316, + 278, + 552, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 278, + 552, + 289 + ], + "type": "text", + "content": "Table 2. Perceptual Accuracy Score (PAS) in different VI sets" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 293, + 550, + 391 + ], + "lines": [ + { + "bbox": [ + 317, + 293, + 550, + 391 + ], + "spans": [ + { + "bbox": [ + 317, + 293, + 550, + 391 + ], + "type": "table", + "html": "
MethodIllusionVQAGVILHallusionBench
CIWAM [40]56.5263.6383.33
DNet [12]14.2021.3033.30
RestoreNet [13]14.2021.3033.30
VAE4.7021.300.00
DDIM 5 steps90.4080.30100.00
DDIM 10 steps90.4096.70100.00
", + "image_path": "e23359fe3df8c5920a1e5075e834c26c92905dfe065a946e9bc9afc79d641a54.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 401, + 553, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 425 + ], + "type": "text", + "content": "To see generation of classic VI (high frequency images with flat patches) please see supplementary material." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "text", + "content": "Psychophysical confirmation. We created 40 images in which we enforced our model to provide an illusion with either red, green, blue, or achromatic targets. Similarly, we have also created 40 control images—images for which there was no enforcement for the patterns to look different. 15 observers participated in our study. The experiment was conducted in a dark room with controlled dim illumination. The only light source was the monitor, set to sRGB. All observers were tested for correct color vision using the Ishihara test. The screen showed one image at a time. Observers were asked to decide whether the two targets were identical or different. The order of images was randomized. We analyze the results in terms of the number of times an image was said to have different targets. For the images in which our model enforced the existence of the illusion, the observers said the targets were different in " + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "text", + "content": " (mean) and " + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "inline_equation", + "content": "67\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "text", + "content": " (median) of the cases with a standard deviation of " + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "text", + "content": ". In contrast, only in " + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "text", + "content": " (mean) and " + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "text", + "content": " (median)—standard deviation " + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 666 + ], + "type": "text", + "content": "—of the cases where the control images said to have different targets." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "We also perform an experiment to evaluate the effectiveness of our perceptual loss weight " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\gamma)" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " through comparison with the two classical models—ODOG and CIWAM—across different thresholds. As the value " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " increases, there" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18648" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 91, + 185, + 143 + ], + "blocks": [ + { + "bbox": [ + 66, + 85, + 77, + 95 + ], + "lines": [ + { + "bbox": [ + 66, + 85, + 77, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 85, + 77, + 95 + ], + "type": "text", + "content": "a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 91, + 185, + 143 + ], + "lines": [ + { + "bbox": [ + 80, + 91, + 185, + 143 + ], + "spans": [ + { + "bbox": [ + 80, + 91, + 185, + 143 + ], + "type": "image", + "image_path": "b90da7016e280bde492e7ef10c473a1db110911b1ff9499124cdbaccc335c016.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 218, + 86, + 301, + 144 + ], + "blocks": [ + { + "bbox": [ + 205, + 86, + 216, + 96 + ], + "lines": [ + { + "bbox": [ + 205, + 86, + 216, + 96 + ], + "spans": [ + { + "bbox": [ + 205, + 86, + 216, + 96 + ], + "type": "text", + "content": "b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 218, + 86, + 301, + 144 + ], + "lines": [ + { + "bbox": [ + 218, + 86, + 301, + 144 + ], + "spans": [ + { + "bbox": [ + 218, + 86, + 301, + 144 + ], + "type": "image", + "image_path": "94558c6caabfb266f3f1983bb4236c4db41f461828b2a9b1f4efaeca00ee2fad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 145, + 325, + 155 + ], + "lines": [ + { + "bbox": [ + 304, + 145, + 325, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 325, + 155 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 340, + 87, + 399, + 144 + ], + "blocks": [ + { + "bbox": [ + 327, + 86, + 337, + 95 + ], + "lines": [ + { + "bbox": [ + 327, + 86, + 337, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 86, + 337, + 95 + ], + "type": "text", + "content": "c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 87, + 399, + 144 + ], + "lines": [ + { + "bbox": [ + 340, + 87, + 399, + 144 + ], + "spans": [ + { + "bbox": [ + 340, + 87, + 399, + 144 + ], + "type": "image", + "image_path": "7dcf72aab58ce7703651a42eccb988e85b31bf1f0621b18b0e9f6c7afcde7f38.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 406, + 87, + 463, + 144 + ], + "blocks": [ + { + "bbox": [ + 406, + 87, + 463, + 144 + ], + "lines": [ + { + "bbox": [ + 406, + 87, + 463, + 144 + ], + "spans": [ + { + "bbox": [ + 406, + 87, + 463, + 144 + ], + "type": "image", + "image_path": "ba44805ab708657f2bfde8df50edb8818410d56819d0cf23535a5036b00466c6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 470, + 87, + 528, + 144 + ], + "blocks": [ + { + "bbox": [ + 470, + 87, + 528, + 144 + ], + "lines": [ + { + "bbox": [ + 470, + 87, + 528, + 144 + ], + "spans": [ + { + "bbox": [ + 470, + 87, + 528, + 144 + ], + "type": "image", + "image_path": "a17e5614b74268d8beac3b1fadb18cc7c20a8bca073a0ec3dbc16d13d85158bd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 91, + 159, + 184, + 253 + ], + "blocks": [ + { + "bbox": [ + 77, + 159, + 87, + 171 + ], + "lines": [ + { + "bbox": [ + 77, + 159, + 87, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 159, + 87, + 171 + ], + "type": "text", + "content": "d)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 159, + 184, + 253 + ], + "lines": [ + { + "bbox": [ + 91, + 159, + 184, + 253 + ], + "spans": [ + { + "bbox": [ + 91, + 159, + 184, + 253 + ], + "type": "image", + "image_path": "f0d8af4dbe3c316c279970eff1f9c3ece53ffef40f9a6859798897fa302b7932.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 185, + 159, + 304, + 253 + ], + "blocks": [ + { + "bbox": [ + 197, + 159, + 208, + 171 + ], + "lines": [ + { + "bbox": [ + 197, + 159, + 208, + 171 + ], + "spans": [ + { + "bbox": [ + 197, + 159, + 208, + 171 + ], + "type": "text", + "content": "e)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 185, + 159, + 304, + 253 + ], + "lines": [ + { + "bbox": [ + 185, + 159, + 304, + 253 + ], + "spans": [ + { + "bbox": [ + 185, + 159, + 304, + 253 + ], + "type": "image", + "image_path": "79e15d6cc0391949f8280451e47baa1cf0ad638ce6ec04b6f6e44b747c2b9204.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 329, + 159, + 423, + 253 + ], + "blocks": [ + { + "bbox": [ + 313, + 159, + 324, + 171 + ], + "lines": [ + { + "bbox": [ + 313, + 159, + 324, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 324, + 171 + ], + "type": "text", + "content": "f)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 159, + 423, + 253 + ], + "lines": [ + { + "bbox": [ + 329, + 159, + 423, + 253 + ], + "spans": [ + { + "bbox": [ + 329, + 159, + 423, + 253 + ], + "type": "image", + "image_path": "1d634b54bb14295b922c1bd55c885b77b00567ed0b18946e2eb122dbd3e81ea6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 450, + 159, + 543, + 252 + ], + "blocks": [ + { + "bbox": [ + 436, + 159, + 447, + 171 + ], + "lines": [ + { + "bbox": [ + 436, + 159, + 447, + 171 + ], + "spans": [ + { + "bbox": [ + 436, + 159, + 447, + 171 + ], + "type": "text", + "content": "g)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 450, + 159, + 543, + 252 + ], + "lines": [ + { + "bbox": [ + 450, + 159, + 543, + 252 + ], + "spans": [ + { + "bbox": [ + 450, + 159, + 543, + 252 + ], + "type": "image", + "image_path": "6b2507d6a5a8567a8eab66fc5ab23ab329bbbcfa36cfa861033616ef789b6045.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 66, + 255, + 184, + 349 + ], + "blocks": [ + { + "bbox": [ + 73, + 257, + 84, + 268 + ], + "lines": [ + { + "bbox": [ + 73, + 257, + 84, + 268 + ], + "spans": [ + { + "bbox": [ + 73, + 257, + 84, + 268 + ], + "type": "text", + "content": "h)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 255, + 184, + 349 + ], + "lines": [ + { + "bbox": [ + 66, + 255, + 184, + 349 + ], + "spans": [ + { + "bbox": [ + 66, + 255, + 184, + 349 + ], + "type": "image", + "image_path": "8a6e088540b3c48f6d3ca69e2e968658dfec62704a3a2891c05bb497f69bb50b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 354, + 555, + 398 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 555, + 398 + ], + "type": "text", + "content": "Figure 5. Qualitative results of visual illusion generation (viewing at larger scale recommended). The first row presents prior work by (a) Gomez-Villa et al. [14], (b) Hirsch et al. [17], and (c) Roy et al. [46]. The second and third rows display results from our method (text prompts are available in the supplementary material). Images (d) through (i) use identical target colors (shown at the left of each image). In images (j) and (k), we employ grayscale gradients as targets. The selected target regions are highlighted in red in the thumbnail images." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 185, + 255, + 303, + 349 + ], + "blocks": [ + { + "bbox": [ + 199, + 257, + 210, + 269 + ], + "lines": [ + { + "bbox": [ + 199, + 257, + 210, + 269 + ], + "spans": [ + { + "bbox": [ + 199, + 257, + 210, + 269 + ], + "type": "text", + "content": "i)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 185, + 255, + 303, + 349 + ], + "lines": [ + { + "bbox": [ + 185, + 255, + 303, + 349 + ], + "spans": [ + { + "bbox": [ + 185, + 255, + 303, + 349 + ], + "type": "image", + "image_path": "5558442cabcbbe120c0efb0c49a4676ef686da8f6ecbc372b6073053c3007e70.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 329, + 255, + 422, + 349 + ], + "blocks": [ + { + "bbox": [ + 315, + 257, + 324, + 269 + ], + "lines": [ + { + "bbox": [ + 315, + 257, + 324, + 269 + ], + "spans": [ + { + "bbox": [ + 315, + 257, + 324, + 269 + ], + "type": "text", + "content": "j)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 255, + 422, + 349 + ], + "lines": [ + { + "bbox": [ + 329, + 255, + 422, + 349 + ], + "spans": [ + { + "bbox": [ + 329, + 255, + 422, + 349 + ], + "type": "image", + "image_path": "b61a592e854fc46fb2462ee136e3bda1cc1a3d985529b73a573cafed2ea6a779.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 449, + 255, + 543, + 348 + ], + "blocks": [ + { + "bbox": [ + 437, + 257, + 448, + 269 + ], + "lines": [ + { + "bbox": [ + 437, + 257, + 448, + 269 + ], + "spans": [ + { + "bbox": [ + 437, + 257, + 448, + 269 + ], + "type": "text", + "content": "k)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 255, + 543, + 348 + ], + "lines": [ + { + "bbox": [ + 449, + 255, + 543, + 348 + ], + "spans": [ + { + "bbox": [ + 449, + 255, + 543, + 348 + ], + "type": "image", + "image_path": "65fee3c57a275af847063cec91b23fadd965f292ba585ac86d4fe51f52b750ed.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 404, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 295, + 429 + ], + "type": "text", + "content": "is a substantial improvement in illusion strength, therefore showing the effectiveness of our perceptual loss." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 441, + 133, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 133, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 133, + 453 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": "In this work, we make two key contributions. First, we report and demonstrate over a wide range of datasets an intriguing empirical observation: diffusion models naturally replicate human responses to brightness and color visual illusions through their image inversion process. This remarkable alignment between human perception and diffusion trajectories suggests that both systems may operate on similar statistical principles when processing ambiguous visual input. Second, leveraging this insight, we develop a method for generating visual illusions using T2I diffusion models, validated through psychophysical experiments. Looking ahead, our findings on perceptual processing in diffusion models could help bridge an important gap in current vision language models, which often struggle with nuanced visual understanding. The way diffusion models progressively process and interpret visual information might offer valuable insights for developing more perception-aware foundational models that better align with human visual cognition. This integration of perceptual understanding into image editing workflows represents an exciting direction for future research, potentially leading to tools that naturally" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 404, + 540, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 540, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 540, + 416 + ], + "type": "text", + "content": "account for human visual perception in their operations." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 418, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 586 + ], + "type": "text", + "content": "Limitations. (1) Replication. As expected from the image inversion process of diffusion models, the input images are spatially destroyed as the samples make their way to the Gaussian distribution. This implies that the target regions disappear for images closer to the natural image distribution, and the effect can no longer be measured. As a consequence our method has difficulty replicating the visual perception of effects which are based on small regions. (2) Generation. Our T2I VI generation method is based on defined regions during inference. This constraints the expressiveness of the T2I model in fulfilling the perceptual conditions. An ideal implementation should establish perceptual regions guided by the attention maps that the input prompt generates to produce more natural visual illusions." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 588, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 708 + ], + "type": "text", + "content": "Acknowledgements: We acknowledge projects PID2022-143257NB-I00, PID2021-128178OB-I00 PID2020-118071GB-I00 and PID2023-152133NB-I00 financed by MCIN/AEI/10.13039/501100011033 and ERDF \"A way of making Europe\", the CERCA Program and the Departament de Recerca i Universitats with ref. 2021SGR01499 from Generalitat de Catalunya, the grant Catedra ENIA UAB-Cruilla (TSI-100929-2023-2) from the Ministry of Economic Affairs and Digital Transition of Spain, and the BBVA Foundations of Science program (grant VIS4NN)." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 219, + 71, + 409, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 71, + 409, + 83 + ], + "spans": [ + { + "bbox": [ + 219, + 71, + 409, + 83 + ], + "type": "text", + "content": "Previous visual illusion generation methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18649" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Arash Akbarinia, Yaniv Morgenstern, and Karl R Gegenfurtner. Contrast sensitivity function in deep networks. Neural Networks, 164:228-244, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "text", + "content": "[2] DeepFloyd Lab at StabilityAI. DeepFloyd IF: a novel state-of-the-art open-source text-to-image model with a high degree of photorealism and language understanding. https://www deepfloyd.ai/deepfloyd-if, 2023. Retrieved on 2023-11-08. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 224 + ], + "type": "text", + "content": "[3] Torsten Betz, Robert Shapley, Felix A Wichmann, and Marianne Maertens. Noise masking of white's illusion exposes the weakness of current spatial filtering models of lightness perception. Journal of vision, 15(14):1-1, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 294, + 267 + ], + "type": "text", + "content": "[4] Barbara Blakeslee and Mark E McCourt. A multiscale spatial filtering account of the white effect, simultaneous brightness contrast and grating induction. Vision research, 39(26): 4361-4377, 1999. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 291 + ], + "type": "text", + "content": "[5] E. Brucke. über erganzungs und contrasfarben. Wiener Sitzungsbser, 51, 1865. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 293, + 294, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 324 + ], + "type": "text", + "content": "[6] Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael S Ryoo. Diffusion illusions: Hiding images in plain sight. arXiv preprint arXiv:2312.03817, 2023. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 326, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 295, + 369 + ], + "type": "text", + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 370, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 294, + 402 + ], + "type": "text", + "content": "[8] David Corney and R Beau Lotto. What are lightness illusions and why do we see them? PLoS computational biology, 3 (9):e180, 2007. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "type": "text", + "content": "[9] Karl R Gegenfurtner, Marina Bloj, and Matteo Toscani. The many colours of 'the dress'. Current Biology, 25(13):R543–R544, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 437, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 437, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 295, + 480 + ], + "type": "text", + "content": "[10] Daniel Geng, Inbum Park, and Andrew Owens. Visual analogs: Generating multi-view optical illusions with diffusion models. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 481, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 481, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 481, + 294, + 513 + ], + "type": "text", + "content": "[11] Daniel Geng, Inbum Park, and Andrew Owens. Factorized diffusion: Perceptual illusions by noise decomposition. arXiv preprint arXiv:2404.11615, 2024. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 514, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 568 + ], + "type": "text", + "content": "[12] Alex Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, and Marcelo Bertalmó. Convolutional neural networks can be deceived by visual illusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309-12317, 2019. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 570, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 613 + ], + "type": "text", + "content": "[13] Alex Gomez-Villa, Adrian Martin, Javier Vazquez-Corral, Marcelo Bertalmó, and Jesús Malo. Color illusions also deceive cnns for low-level vision tasks: Analysis and implications. Vision Research, 176:156-174, 2020. 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 615, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 295, + 657 + ], + "type": "text", + "content": "[14] A. Gomez-Villa, A. Martin, J. Vazquez-Corral, M. Bertalmio, and J. Malo. On the synthesis of visual illusions using deep generative models. Journal of Vision, 22(8):2-2, 2022. 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "text", + "content": "[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 98, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 553, + 129 + ], + "type": "text", + "content": "[16] Ludimar Hermann. Eine erscheinung simultanen contrastes. Archiv für die gesamte Physiologie des Menschen und der Tiere, 3(1):13-15, 1870. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 131, + 553, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 131, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 317, + 131, + 553, + 164 + ], + "type": "text", + "content": "[17] Elad Hirsch and Ayellet Tal. Color visual illusions: A statistics-based computational model. Advances in neural information processing systems, 33:9447-9458, 2020. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 166, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 166, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 166, + 553, + 198 + ], + "type": "text", + "content": "[18] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 201, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 201, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 317, + 201, + 553, + 243 + ], + "type": "text", + "content": "[19] Been Kim, Emily Reif, Martin Wattenberg, and Samy Bengio. Do neural networks show gestalt phenomena? an exploration of the law of closure. arXiv preprint arXiv:1903.01069, 2(8), 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 246, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 246, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 317, + 246, + 553, + 289 + ], + "type": "text", + "content": "[20] Frederick AA Kingdom. Lightness, brightness and transparency: A quarter century of new ideas, captivating demonstrations and unrelenting controversy. Vision Research, 51 (7):652-673, 2011. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 291, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 312 + ], + "type": "text", + "content": "[21] Akiyoshi Kitaoka. Barutan-seijin. Akiyoshi's illusion Website, 2003. Accessed: 2024-11-13. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 315, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 315, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 315, + 553, + 335 + ], + "type": "text", + "content": "[22] Akiyoshi Kitaoka. Robot. Akiyoshi's illusion Website, 2005. Accessed: 2024-11-13. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 338, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 358 + ], + "type": "text", + "content": "[23] Akiyoshi Kitaoka. Shiosai. Akiyoshi's illusion Website, 2008. Accessed: 2024-11-13. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "text", + "content": "[24] Yuki Kubota, Atsushi Hiyama, and Masahiko Inami. A machine learning model perceiving brightness optical illusions: Quantitative evaluation with psychophysical data. In Proceedings of the Augmented Humans International Conference 2021, pages 174-182, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 418, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 418, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 418, + 553, + 460 + ], + "type": "text", + "content": "[25] Gihyun Kwon and Jong Chul Ye. Diffusion-based image translation using disentangled style and content representation. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 463, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 463, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 463, + 553, + 495 + ], + "type": "text", + "content": "[26] V. Laparra and J. Malo. Visual aftereffects and sensory nonlinearities from a single statistical framework. Front. Human Neurosci., 9, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 498, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 498, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 498, + 553, + 529 + ], + "type": "text", + "content": "[27] V. Laparra, G. Camps, and J. Malo. Iterative gaussianization: From ICA to random rotations. IEEE Trans. Neur. Nets, 22 (4):537-549, 2011. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 532, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 532, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 532, + 553, + 564 + ], + "type": "text", + "content": "[28] V. Laparra, S. Jiménez, G. Camps, and J. Malo. Nonlinearities and adaptation of color vision from sequential principal curves analysis. Neural Comp., 24(10):2751-2788, 2012. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "type": "text", + "content": "[29] V. Laparra, JE. Johnson, G. Camps, R. Santos, and J. Malo. Estimating information theoretic measures via multidimensional gaussianization. IEEE Trans. Patt. Anal. Mach. Intell., 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 612, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 612, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 612, + 553, + 644 + ], + "type": "text", + "content": "[30] Qiang Li, Alex Gomez-Villa, Marcelo Bertalmó, and Jesús Malo. Contrast sensitivity functions in autoencoders. Journal of Vision, 22(6):8-8, 2022. 2, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[31] Senmao Li, Joost van de Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing, 2023. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[32] Tingting Li, Fanyu Wang, Ying Zhou, and Zhenping Xie. Visual illusion cognition dataset construction and recognition" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18650" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "performance by deep neural networks. In 2022 IEEE 8th International Conference on Cloud Computing and Intelligent Systems (CCIS), pages 90-94. IEEE, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "text", + "content": "[33] Ben Lonnqvist, Alban Bornet, Adrien Doerig, and Michael H Herzog. A comparative biology approach to dnn modeling of vision: A focus on differences, not similarities. Journal of vision, 21(10):17-17, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 183 + ], + "type": "text", + "content": "[34] Mark McCourt. Grating induction: A new explanation for stationary phantom gratings. Vision Research, 34(12):1609-1617, 1994. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 296, + 237 + ], + "type": "text", + "content": "[35] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 238, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 296, + 293 + ], + "type": "text", + "content": "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "type": "text", + "content": "[37] Chaithanya Kumar Mummadi, Ranjitha Subramaniam, Robin Hutmacher, Julien Vitay, Volker Fischer, and Jan Hendrik Metzen. Does enhanced shape bias improve neural network robustness to common corruptions? International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 296, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 382 + ], + "type": "text", + "content": "[38] Jerry Ngo, Swami Sankaranarayanan, and Phillip Isola. Is clip fooled by optical illusions? International Conference on Learning Representations, 2023. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 403 + ], + "type": "text", + "content": "[39] David Novick and Akiyoshi Kitaoka. The confetti illusion. Journal of Illusion, 2, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 437 + ], + "type": "text", + "content": "[40] Xavier Otazu, C Alejandro Parraga, and Maria Vanrell. Toward a unified chromatic induction model. Journal of Vision, 10(12):5-5, 2010. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "text", + "content": "[41] Jiang Qin, Kai Wang, Bin Zou, Lamei Zhang, and Joost van de Weijer. Conditional diffusion model with spatial-frequency refinement for sar-to-optical image translation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 548 + ], + "type": "text", + "content": "[42] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "type": "text", + "content": "[43] Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. International Conference on Machine Learning, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 690 + ], + "type": "text", + "content": "[45] Aniket Roy, Anirban Roy, Soma Mitra, and Kuntal Ghosh. Bri3l: A brightness illusion image dataset for identification and localization of regions of illusory perception. Proceedings of the IEEE International Conference on Image Processing, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[46] Aniket Roy, Anirban Roy, Soma Mitra, and Kuntal Ghosh. Bri3l: A brightness illusion image dataset for identification" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 553, + 106 + ], + "type": "text", + "content": "and localization of regions of illusory perception. Proceedings of the IEEE International Conference on Image Processing, 2024. 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 106, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 162 + ], + "type": "text", + "content": "[47] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. Illusionvqa: A challenging optical illusion dataset for vision language models. First Conference on Language Modeling, 2024. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "text", + "content": "[48] Arthur G. Shapiro and Dejan Todorovic. The Oxford Compendium of Visual Illusions, chapter 22. Oxford University Press, 2017. Figure 3.2. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "text", + "content": "[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 231, + 553, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 264 + ], + "type": "text", + "content": "[50] Eric D Sun and Ron Dekel. Imagenet-trained deep neural networks exhibit illusion-like response to the scintillating grid. Journal of Vision, 21(11):15-15, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 298 + ], + "type": "text", + "content": "[51] Chuanming Tang, Kai Wang, and Joost van de Weijer. Iterinv: Iterative inversion for pixel-level t2i models. Neurips 2023 workshop on Diffusion Models, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 299, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 332 + ], + "type": "text", + "content": "[52] Oguzhan Ulucan, Diclehan Ulucan, and Marc Ebner. Investigating color illusions from the perspective of computational color constancy. arXiv preprint arXiv:2312.13114, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "text", + "content": "[53] Clinton Wang and Polina Golland. Interpolating between images with diffusion models. ICML 2023 Workshop DeployableGenerativeAI, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 366, + 553, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 411 + ], + "type": "text", + "content": "[54] Kai Wang, Fei Yang, Shiqi Yang, Muhammad Atif Butt, and Joost van de Weijer. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 456 + ], + "type": "text", + "content": "[55] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 456, + 553, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 512 + ], + "type": "text", + "content": "[56] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. DiffusionDB: A large-scale prompt gallery dataset for text-to-image generative models. Annual Meeting of the Association for Computational Linguistics, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 513, + 553, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 553, + 556 + ], + "type": "text", + "content": "[57] Eiji Watanabe, Akiyoshi Kitaoka, Kiwako Sakamoto, Masaki Yasugi, and Kenta Tanaka. Illusory motion reproduced by deep neural networks trained for prediction. Frontiers in psychology, 9:340023, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "text", + "content": "[58] Robert Max Williams and Roman V Yampolskiy. Optical illusions images dataset. arXiv preprint arXiv:1810.00415, 2, 2018. 2, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 591, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 645 + ], + "type": "text", + "content": "[59] Hongtao Zhang, Zhen Li, and Shinichi Yoshida. Muller-layer illusion is replicated by higher layer of pre-trained deep neural network for object recognition. International Symposium on Computational Intelligence and Industrial Applications, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[60] Yichi Zhang, Jiayi Pan, Yuchen Zhou, Rui Pan, and Joyce Chai. Grounding visual illusions in language: Do vision-language models perceive illusions like humans? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5718-5728, 2023. 2, 4, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18651" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "text", + "content": "[61] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, Tiviatis Sim, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. European Conference on Computer Vision, 2024. 5" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18652" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_content_list.json b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..67706e6aa043721f90ef219415feda72f5c637c9 --- /dev/null +++ b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_content_list.json @@ -0,0 +1,1638 @@ +[ + { + "type": "text", + "text": "The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generation", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanis Benidir", + "bbox": [ + 133, + 203, + 246, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nicolas Gonthier", + "bbox": [ + 418, + 203, + 558, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clément Mallet", + "bbox": [ + 732, + 203, + 859, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Univ Gustave Eiffel, ENSG, IGN, LASTIG, France", + "bbox": [ + 284, + 220, + 691, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "firstname_lastname@ign.fr", + "bbox": [ + 377, + 242, + 599, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bi-temporal change detection at scale based on Very High Resolution (VHR) images is crucial for Earth monitoring. This remains poorly addressed so far: methods either require large volumes of annotated data (semantic case), or are limited to restricted datasets (binary set-ups). Most approaches do not exhibit the versatility required for temporal and spatial adaptation: simplicity in architecture design and pretraining on realistic and comprehensive datasets. Synthetic datasets are the key solution but still fail to handle complex and diverse scenes. In this paper, we present HySCDG a generative pipeline for creating a large hybrid semantic change detection dataset that contains both real VHR images and inpainted ones, along with land cover semantic map at both dates and the change map. Being semantically and spatially guided, HySCDG generates realistic images, leading to a comprehensive and hybrid transfer-proof dataset FSC-180k. We evaluate FSC-180k on five change detection cases (binary and semantic), from zero-shot to mixed and sequential training, and also under low data regime training. Experiments demonstrate that pretraining on our hybrid dataset leads to a significant performance boost, outperforming SyntheWorld, a fully synthetic dataset, in every configuration. All codes, models, and data are available here: https://yb23.github.io/projects/cywd/.", + "bbox": [ + 89, + 323, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 724, + 220, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Efficient methods for change detection (CD) are crucial for monitoring territories and the various phenomena and activities that impact them [3]. As human and climatic pressures increase, this topic is of utmost importance and calls every year for higher spatial resolution and semantic accuracy [55]. Change detection at scale has benefited greatly from the contributions of deep learning in recent years [7, 12, 69, 73], but most methods are grounded on large volumes of annotated data [41, 60]. Creating a large-scale dataset for bi-temporal (pair-wise) remote sensing change detection poses signifi", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b9ea8ffcb9c09a66d1860e45091d22e2b6a0f676515dde8384ccb74ae5f7ef0.jpg", + "image_caption": [ + "Target datasets", + "Figure 1. Efficient and scalable change detection requires a comprehensive training dataset that does not exist today. Using a single-temporal dataset (image+semantic map), we propose HySCDG that generates a novel bi-temporal hybrid dataset FSC-180k. This enables multiple transfer learning possibilities on either binary and semantic change detection tasks. $\\mathcal{O} / \\mathcal{O}$ : Transfer with/without fine-tuning our model, respectively." + ], + "image_footnote": [], + "bbox": [ + 535, + 287, + 880, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cant challenges and costs, especially when dealing with Very High Resolution (VHR) images (ground sampling distance: GSD $< 1\\mathrm{m}$ , [30]). This is due to the need for expertise and considerable effort in collecting, preprocessing, and annotating images [59, 62], particularly in the case of VHR data and multiple labels (termed semantic change detection, SCD) [57, 61, 70, 74].", + "bbox": [ + 511, + 704, + 908, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As an alternative, synthetic data generation is a promising direction for bi-temporal labeled data provision from a single temporal domain that is successful in diverse areas such as urban sprawl [17], disaster or climate change monitoring [42] or crop mapping [36]. Two main paradigms exist for CD: fully synthetic datasets (mainly graphics-based", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2204", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[27, 53]), and hybrid solutions that mix real and modified images, often at the instance level [5, 69, 71], known as data augmentation. The first method synthesizes image pairs by rendering two versions of a 3D virtual world thanks to a computer graphics engine. On the other hand, hybrid approaches synthesize new image pairs by pasting or erasing object instances into existing images ( inpainting). They propose a suitable trade-off and faster generation of large quantities of data, with likely higher realism. However, none of these solutions meet the requirements for scalable SCD with VHR imagery (Tab. 1): focus on single tasks, restricted geographical areas, limited diversity, no guarantee on semantic consistency between pairs [48] and subsequently poor transfer to real datasets. To address this challenge, we introduce a generative pipeline built upon Stable Diffusion [44] and ControlNet [65], designed to harness an existing VHR semantic land cover dataset [18] and instance masks, to generate an extensive hybrid semantic bi-temporal change detection dataset encompassing both real and inpainted images, semantic land cover maps at each date, and a change map (Fig. 1). Then, we evaluate the transferability of our synthetic dataset on five datasets for both binary and semantic change detection cases ranging from zero-shot to sequential and mixed training set-ups, as well as under low-data regime.", + "bbox": [ + 89, + 90, + 485, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 109, + 454, + 330, + 467 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a generation pipeline HySCDG for creating VHR bi-temporal SCD datasets based on a land cover real-world dataset by using Stable Diffusion and ControlNet with tunable and semantically-guided inpainting of individual objects. It can be easily configured or finetuned on any land-cover dataset, generating images with the dataset's distinctive traits.", + "- We provide FSC-180k: a Hybrid Semantic Change Detection pretraining dataset, created from FLAIR [18] along with the instance masks for 300k objects for the corresponding images.", + "- We provide a comprehensive evaluation of the proposed synthetic dataset on multiple transfer learning scenarios on five change datasets, in particular for SCD." + ], + "bbox": [ + 109, + 468, + 482, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 694, + 227, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detecting changes in the deep learning era. We target bi-temporal land cover change detection in VHR images [32]. This historical task [22, 37, 52] consists in detecting the meaningful elements of a given area that have changed between two acquisitions and providing the land cover categories at each period (also known as change trajectory). It presents challenges: radiometric variations, geometric variations and non-permanent changes. They have been extensively addressed with supervised deep learning techniques [70]: from CNN [11, 12, 64] to Transformers [2, 7, 13, 55, 72, 73], mainly with Siamese architecture. Multiple datasets, accompanied with ad-hoc methods, have been", + "bbox": [ + 88, + 719, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "created to feed such models. They are often limited by their size [6, 25, 39, 56, 63], geographical extent [6, 56, 57, 63], low geometrical or annotation quality [11]. Also, most of the effort focuses on binary change detection, i.e., restricted to a unique category, mainly buildings [6, 25, 39, 40, 50] or disasters [67]. Such scarcity in high quality, diverse, large-scale labels is alleviated with transfer learning [4, 51], by fine-tuning models pretrained on large-scale datasets for different image domains, even for semantic change detection [56, 57, 63].", + "bbox": [ + 511, + 90, + 908, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthesizing and inpainting remote sensing images.", + "text_level": 1, + "bbox": [ + 511, + 261, + 908, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthetic remote sensing data is useful e.g., for cloud removal [10], image restoration [16], or training deep supervised models [58], before transfer to real examples. Solutions based on patches, [35, 68], autoencoders [23], GAN [28, 34], -aligned generation [24] have been proposed, but with limited image quality and no semantic control. Diffusion models (DM) have improved such quality [44], but most models have applied DM directly on mid-resolution RGB images [14, 26, 58], neglecting both the multi-spectral dimension [9] and VHR spaces. However, a key finding remains: semantically control the synthesis, e.g., by directly synthesizing text [26], semantic maps [58] or with a control module guided by edges [9], semantic maps [14] or metadata [26]. In our work, we fine-tune a Stable Diffusion and a ControlNet models to achieve semantically-controlled inpainting of VHR images as a basis to take advantage of existing images, and generate an hybrid semantic change detection dataset at scale and with high diversity.", + "bbox": [ + 511, + 277, + 908, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating synthetic change datasets. Two main approaches exist: fully synthetic or hybrid solutions. The first paradigm leverages advanced 3D rendering engines to create fully synthetic datasets [17, 27]. This procedural approach allows complete control over rendering parameters (instance placement, semantic classes, lighting conditions, etc.) generated from scratch. E.g., SyntheWorld [53] effectively leverages DM prompted by GPT-4 to create a diverse and large-scale fully synthetic dataset. On the other hand, hybrid approaches generate (semi-)synthetic data by inserting fake changes into real images.", + "bbox": [ + 511, + 568, + 908, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Such modifications can be random crops [49], instance copy-pasting, or inpainting made by GAN [5, 49, 71] or DM [54, 69]. Blending or style transfer methods are often used to enhance diversity [5, 49].", + "bbox": [ + 511, + 734, + 906, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A pivotal challenge is the semantic guidance: modifying few elements of images to mimic real scarce changes, e.g., with instance footprints extracted from external datasets [5] or segmentation model outputs [69]. [54] adopt a Denoising diffusion probabilistic model to generate a new image directly from a modified semantic map and a degraded version of the corresponding snapshot. Changen2 [69] is the closest", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2205", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work to ours, with a full pipeline to synthesize a hybrid semantic dataset, again with a DM conditioned by a semantic map. However, Changen2 creates synthetic data specific to a given target dataset in terms of image resolution, size, or change characteristics. In this paper, we explore semantically controlled inpainting based on DM and ControlNet to the intertwined challenge of diverse and multi-class modification of VHR images. It is termed HySCDG and leads to the hybrid dataset named FSC-180k. Tab. 1 provides a comparison with main synthetic change detection datasets.", + "bbox": [ + 89, + 90, + 483, + 242 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/a53d0462c68b1f86238ad2e2fe616f6309d39e990e418cf2c80f92b7989f7fb0.jpg", + "table_caption": [ + "Table 1. Synthetic Change Datasets. Land cover change detection in remote sensing imagery. For the total number of pixels, we only consider strictly different images. The geographical zone's extension can be local [ ] national [ ] or global [ ] We denote open-access (OA) and directly downloadable datasets with ✓. In the case of IAug, we sum the number of pixels of the two datasets provided in this work. OEM means OpenEarthMap." + ], + "table_footnote": [], + "table_body": "
DatasetOAPixels ×106GSD (m)ClassesSourceZoneType
SynCW [27]X370.61X田Synthetic
SMARS [17]✓1100.3-0.52X田Synthetic
IAug [5]X1,1670.075-0.51LEVIR-CD [6], WHU-CD [25]田Hybrid
Ce-100K [54]X6,5530.25-0.58OEM [62]田Synthetic
Changen2 [69]X7,0770.25-0.58OEM [62]田Hybrid
Changen [71]X11,7960.81xView2 [43]田Hybrid
SyntheWorld [53]✓18,3500.3-11XXSynthetic
FSC-180k (Ours)✓80,7400.216FLAIR [18]田Hybrid
", + "bbox": [ + 93, + 364, + 480, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transfer learning from synthetic/hybrid datasets. Transfer learning is used to assess the validity of the generated image pairs, and vary with the operational set-up [33]. SCD at scale is performed through sequential learning: pretraining on such datasets and fine-tuning on real-world cases [5, 27, 69, 71]. Fewer works [53] evaluate on mixed training sets (single training mixing both real and synthetic samples), which can help to prevent overfitting. An orthogonal solution considers low data regimes, i.e., fine-tuning on a very small train set [5, 53, 54], or even under zero-shot scenarios [54]. We evaluate our proposed dataset in the four conceivable transfer scenarios.", + "bbox": [ + 89, + 558, + 485, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Hybrid generation of semantic changes", + "text_level": 1, + "bbox": [ + 89, + 753, + 441, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Much of the success of deep learning methods for Earth observation relies on large-scale training datasets with reliable annotations [18, 45]. However, no large-scale publicly available dataset exists for bi-temporal SCD: we propose HySCDG for Hybrid Semantic Change Detection Generation, a procedure to leverage existing large-scale land cover datasets and state-of-the-art image inpainting methods to simulate visually realistic bi-temporal semantic changes.", + "bbox": [ + 89, + 779, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This leads to the generation of the FSC-180k dataset.", + "bbox": [ + 511, + 90, + 864, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our paper is grounded on two main ideas: (i) Adapting and fine-tuning a Stable Diffusion Model enables efficient inpainting of VHR images with a semantic control and geographical instance selection; (ii) Changes can be simulated with sufficient diversity by randomly selecting instances (and modifying their labels) within an existing land cover map that is paired with the image used as input. The pipeline is shown in Fig. 2. Details are provided in the following.", + "bbox": [ + 511, + 106, + 906, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Adaptation of Stable Diffusion for Inpainting.", + "text_level": 1, + "bbox": [ + 511, + 234, + 898, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use Stable Diffusion (SD) [44], an open and state-of-the-art text-to-image synthesis method. Our objective is to modify SD to manage controllable inpainting specifically in remote sensing aerial images, for various landscapes or semantic classes. Though SD models are effective in generating natural images with clear focal points and layered depth, with artistic or cinematic styles, they require adaptation to be used for remote sensing images [26]: acquisition distance, intricate textures, no domination of a single element. Starting from a Stable Diffusion 2 Inpainting checkpoint1, we sequentially fine-tune the Autoencoder and the Diffusion U-Net, before adding and training a ControlNet [65] that leverages information extracted from an existing large-scale land cover dataset (namely, FLAIR [18], see Sec. 3.3). Our pipeline allows to generate 5-band images that incorporate infrared and elevation along with RGB channels, for the first time in satellite image synthesis.", + "bbox": [ + 511, + 255, + 908, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adaptation of the Variational Autoencoder. Unlike most SD extensions, which focus on fine-tuning the central U-Net, we also adapted the Variational Autoencoder (VAE) that encodes images into the diffusion latent space. This adaptation targets efficient encoding of multi-band VHR images from the FLAIR dataset. The original VAE, trained with a KL-regularization factor [15], uses a combination of L1 and perceptual loss [66] alongside KL-divergence. To improve compression of remote sensing images, we added both L2 and Focal losses [31]: minimizing high-frequency errors reduces blurriness. Additionally, we included a \"color loss\" (L2 loss on each $5 \\times 5$ patch) to limit oversaturation, which affects ground object color realism. Fine-tuning of the VAE took 160 hours A100 GPU.", + "bbox": [ + 511, + 530, + 908, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stable Diffusion training. We then fine-tuned the Diffusion U-Net for inpainting, adapting it to the new VAE and teaching it to synthesize and inpaint VHR images. Special attention was given to prompt selection. Although inpainting is the model's primary function, we included direct image generation from prompts in $10\\%$ to $20\\%$ of cases, as in standard SD. This training required 300 hours on a A100 GPU (30,000 iterations, batch size of 32).", + "bbox": [ + 511, + 758, + 908, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1https://huggingface.co/stabilityai/stable-diffusion-2-inpainting", + "bbox": [ + 531, + 886, + 867, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2206", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8dda307eabf77920221443de3276d86a520b0d5380aa777621a7bc1573adb632.jpg", + "image_caption": [ + "Figure 2. HySCDG pipeline. From a single-temporal dataset composed of one VHR image $I_{1}$ , a semantic map $M_{1}$ , and some openly available labeled instances, we generate a new VHR image $I_{2}$ , a new map $M_{2}$ and subsequently a change map $C$ . This results in the FSC-180k hybrid dataset. The two pivotal novelties consist in: (i) Adapting and fine-tuning a Stable-Diffusion Model from image inpainting and (ii) exploiting open geospatial data for inpainting prompt control and semantically guiding the objects to be modified. The combination of both solutions ensures a diverse at-scale VHR multi-class change detection dataset." + ], + "image_footnote": [], + "bbox": [ + 153, + 85, + 846, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt Engineering. Prompts are designed using both each image's geographic coordinates and open land cover data (OpenStreetMap [38]). To avoid learning irrelevant details, prompts are composed of three elements: (i) Spatial (place name, city, region), (ii) Temporal (time of day, season), (iii) Semantic (frequent classes in the inpainted area, compared to the overall dataset).", + "bbox": [ + 88, + 476, + 482, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt example: \"Grass and agricultural vegetation next to a highway, locality of Savigny-en-Revermont, Bourgogne-Franche-Comté, in the morning, during Summer.\"", + "bbox": [ + 88, + 582, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Conditional change inpainting with ControlNet", + "text_level": 1, + "bbox": [ + 89, + 636, + 482, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Driven by our goal to control the new classes in the inpainted zones, we add a ControlNet (CN) architecture [65] to build a model that generates realistic VHR images from a semantic segmentation map (here land cover). CN is designed to add task-specific conditioning to a pretrained image DM. It allows injecting explicit semantic guidance during the diffusion process, conditioning the outputs on some reference image (a color map of the semantic class), in addition to a text prompt. Our training method enables CN to perform both inpainting and text-to-image synthesis, using the semantic condition. Training took 240 hours on a A100 GPU.", + "bbox": [ + 88, + 657, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Select, Mask, Change, Inpaint : the HySCDG pipeline", + "text_level": 1, + "bbox": [ + 89, + 833, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by [46], the core of our method lies in our \"Select and Inpaint\" mechanism. Given a VHR image along with", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "its semantic segmentation and instances $^2$ , we inpaint these instances using the SD+CN inpainting method: the label of the instance is changed, the texture of the image is substituted accordingly. This results in a pair of VHR images, both with semantic maps and the corresponding semantic change maps. The complete HySCDG is provided in Fig. 2 and detailed below. It is illustrated for one image but is applied several times to each image of the single temporal dataset.", + "bbox": [ + 511, + 476, + 906, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HySCDG uncovered. For an image $I_{1}$ with its semantic map $M_{1}$ , and $\\{O_i\\}$ the set of $N_I$ instances lying on the extent of $I_{1}$ .", + "bbox": [ + 511, + 607, + 905, + 651 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Randomly select $N_{change}$ instances among $\\{O_i\\}$ .", + "- For each instance, generate an inpainting mask: footprint+spatial buffer.", + "- Using $M_1$ , take as “ $T_1$ class” the most frequent class inside the footprint: $c_1$ . Then, select the larger convex hull of pixels of this class inside the buffered zone. These pixels correspond to the change mask: $C$ .", + "- Randomly select another class $c_2$ at $T_2$ and replace the previous label by this new label.", + "- Randomly select $N_{\\text{nochange}}$ random shapes in the image: such masks appear only on the inpainting mask, not on the change mask, letting their labels unchanged.", + "- Apply $\\mathrm{SD + CN}$ to masked image and new semantic map.", + "- Cumulate the previous steps: we obtain our inpainting" + ], + "bbox": [ + 532, + 652, + 906, + 864 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "${}^{2}$ A panoptic segmentation is the most suitable case but not mandatory, assuming a sufficient number of instances ensures enough diversity.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2207", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mask, change mask, and for $T_{2}$ a new semantic map $M_2$ and the corresponding new image $I_{2}$ .", + "bbox": [ + 122, + 90, + 480, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pair of images $(I_1, I_2)$ with their corresponding semantic maps $(M_1, M_2)$ , and the change map $C$ (difference between $M_2$ and $M_1$ ), are the samples of the hybrid dataset.", + "bbox": [ + 89, + 121, + 482, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Label and instance extraction. For comprehensive multiclass learning, we select FLAIR dataset [18], a large-scale semantic segmentation map for 16 land cover classes over more than $800\\mathrm{km}^2$ . FLAIR offers the advantage of covering areas where geographical instances of many FLAIR labels are publicly available3, alleviating requiring a panoptic segmentation. Around 300,000 instance masks have been extracted for this work.", + "bbox": [ + 89, + 174, + 483, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inpainting signature mitigation. Inpainting produces realistic changes but new areas tend to have sometimes a high-frequency signature, as reported by other works [20, 29]. In order to discourage the model from simply learning this inpainting noise instead of learning the actual changes between images, we also inpaint random areas of the image without changing the semantic class in a similar way to the \"not changed regions\" from [46]. Besides, to obtain the best transition between the original image and the inpainted area, we use a continuous mask in the inference [1].", + "bbox": [ + 89, + 301, + 483, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Selection of the new label. For each change area, we randomly select the new class $c_2$ among all classes weighted by their frequency in the whole dataset divided by their frequencies in the change area.", + "bbox": [ + 89, + 460, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adoption of a buffer zone. For inpainting, we add a buffer around the footprint of each modified instance. This gives SD more freedom to generate smoother borders around the changed areas. It also helps handling small spatial discrepancies between footprints and the VHR image content [19].", + "bbox": [ + 89, + 529, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Structure of the FSC-180k dataset", + "text_level": 1, + "bbox": [ + 89, + 614, + 390, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each image in the FLAIR training set, we generate three unique modified images, resulting in 180,000 synthetic images, each with a corresponding $512 \\times 512$ semantic map covering 16 distinct semantic classes. We call this dataset FSC-180k, short for FLAIR Synthetic Change. FSC-180k includes approximately 80 billion pixels with a ground resolution of $0.2\\mathrm{m}$ per pixel. Note that we can expand the dataset by pairing synthetic images that were generated (differently) from a single real image, process that we use at training time, effectively doubling the number of image pairs. Tab. 1 compares FSC-180k with other synthetic change detection datasets and shows that FSC-180k offers a much broader range of semantic changes beyond simple building modifications, is of higher spatial resolution, and is nearly ten times larger than comparable datasets. It should also be noted that the proportion of changes in FSC-180k is closer to the real", + "bbox": [ + 89, + 636, + 483, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scenarios observed in Europe (prevalence of $5\\%$ change) than other academic datasets (see Tab. 2), excepting HiUCD.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quality assessment of FSC-180k. ControlNet is the most error-prone module: it may sometimes fail to respect the semantic map. We evaluated how unchanged the semantic map remains by performing 15-class semantic segmentation on real and generated images, thanks to a UNet trained for this task, leading to a satisfactory error rate below $20\\%$ between both outputs. We also leveraged standard metrics for image generation evaluation to assess the realism of the synthesis. We got an Inception Score [47] of 6.2 on the inpainted images, which is comparable to real images, and a low FID score [21] of 0.43 between inpainted and real images.", + "bbox": [ + 511, + 142, + 906, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 338, + 645, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct various transfer learning experiments to validate hybrid data pretraining, comparing FSC-180k with the fully artificial SyntheWorld [53], the only available synthetic dataset with a size comparable to ours (see Tab. 1).", + "bbox": [ + 511, + 363, + 906, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Pretraining approaches with FSC-180k", + "text_level": 1, + "bbox": [ + 511, + 434, + 849, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We assess FSC-180k's effectiveness by exploiting existing open datasets as targets under various settings (see Tab. 2 and Supplementary Material). Here, we first present the different transfer learning approaches adopted, and then, dive into the deep models and datasets considered (Sec. 4.2). Our objective is to demonstrate how pretraining on a hybrid dataset improves performance, for both binary and semantic change detection.", + "bbox": [ + 511, + 455, + 906, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The four scenarios are:", + "bbox": [ + 511, + 577, + 666, + 590 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Sequential Transfer Learning. Models are pretrained on FSC-180k, and then fine-tuned on the target data.", + "- Low Data Regime. The same as above but with a fraction of the available training set.", + "- Mixed Training. Models are trained on a blend of FSC-180k and the target data.", + "- Zero-Shot Learning. A model pretrained on FSC-180k is directly applied on the target dataset, with only a remapping of FSC-180k classes to align labels." + ], + "bbox": [ + 531, + 592, + 906, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental protocol", + "text_level": 1, + "bbox": [ + 511, + 739, + 723, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Target datasets and metrics", + "text_level": 1, + "bbox": [ + 511, + 761, + 751, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our transfer experiments span five diverse datasets, among which three contain semantic labels (Tab. 2). This ensures variety in landscapes, sizes, and change types, i.e., a comprehensive evaluation. Among the plethora of (S)CD datasets, we selected the most popular ones: Levir-CD [6], S2Looking [50], SECOND [63] and HiUCD-mini [56].", + "bbox": [ + 511, + 780, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We used classical metrics from the change detection literature: Intersection over Union (IoU) and F1 score for binary", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3https://www.data.gouv.fr/en/datasets/bd-topo-r/", + "bbox": [ + 107, + 886, + 367, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2208", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cases and mainly the Semantic Change Segmentation score (SCS) from [59] for semantic evaluation. In this latter case, the Separated Kappa (SeK) from [63] is also informative though it tends to be more biased towards detection task than semantic prediction.", + "bbox": [ + 89, + 90, + 480, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We added some experiments related to HiUCD-XL [57], a follow-up dataset of HiUCD-mini [56] using semantic mIoU and change mIoU, metrics provided on the evaluation platform4 as test labels are kept private.", + "bbox": [ + 89, + 166, + 482, + 227 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d8bf289a8c3d146ffd7f85d38b8c75b4356144c5fabd0e20c199718e1c912b8.jpg", + "table_caption": [ + "Table 2. Target datasets. Overview of the real-world datasets used for our transfer learning experiments." + ], + "table_footnote": [], + "table_body": "
DatasetChange (%)Pixels ×106GSD (m)ClassesPeriodZone
Levir-CD [6]28.41,0670.512002/18Texas
S2Looking [50]25.68,3890.3-0.812017/20World
SECOND [63]20.01,2220.3-0.66?China
HiUCD-mini [56]6.65670.192017/19Estonia
HiUCD-XL [57]1.210,6950.192017/19Estonia
FSC-180k (Ours)7.080,7400.2162018/21France
", + "bbox": [ + 91, + 277, + 488, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Model architecture", + "text_level": 1, + "bbox": [ + 89, + 425, + 269, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider that designing a new architecture is out of the scope of this paper and that numerous solutions have already been proposed (see Sec. 2). Therefore, we initially used a simple U-Net model [8] with a segmentation head for semantic change detection. Although effective for binary change, it lacked expressiveness for semantic tasks. Similarly to [11], we developed a simple Dual UNet model composed of 2 nearly identical parallel UNets for (i) change detection, and (ii) semantic segmentation, with ResNet-50 backbones pretrained on ImageNet and trained in a multi-task setting. The change branch incorporates features from the semantic encoder. We keep the training strategy simple and constant to focus on data and transfer, this enhances interpretability.", + "bbox": [ + 88, + 443, + 482, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Can a binary model be sufficient? We tested both a pure binary CD model and the Dual UNet, for which, when, switching to binary-only tasks, the semantic decoder is inactive but the encoding part remains active (as it provides features to the change detection branch). Such selective change detection ability allows training on a multi-class dataset and focusing on specific binary tasks. Experiments show that even for binary tasks, the Dual UNet outperforms the binary-only model, e.g., achieving an IoU of 0.83 on Levir-CD (compared to 0.80). Learning more information than just change pixels, the Dual UNet gains a deeper scene understanding, leading to more reliable predictions.", + "bbox": [ + 89, + 650, + 482, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Existing architectures. We tested more complex transformer-based architectures, namely ChangeFormer [2],", + "bbox": [ + 89, + 849, + 482, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MambaCD [7], and SCanNet [13], both on HiUCD-mini and SECOND datasets. SCanNet performed best, balancing high performance and straightforward implementation due to its lightweight design (see Supplementary Material). SCanNet is kept hereafter for comparisons (see Tab. G).", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transfer for varying semantic classes. In the semantic case, we faced the issue of distinct classes between source and target datasets. This required to choose which classes should be kept for model training.", + "bbox": [ + 511, + 185, + 903, + 244 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Sequential case: for both pretraining and fine-tuning strategies, we use the classes of the processed dataset. The semantic head of the model is reinitialized and adapted to the new number of classes before fine-tuning.", + "- Mixed case: Classes from the target dataset are used, manually mapping pretraining classes to their closest counterparts." + ], + "bbox": [ + 531, + 246, + 906, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3. Experimental setting", + "text_level": 1, + "bbox": [ + 511, + 375, + 705, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply basic augmentations on the images: horizontal and vertical flips, $90^{\\circ}$ rotations, and small rescaled crops (scales above 0.8). Normalization by mean and variance from the target dataset is used in all experiments. For optimization, we use the AdamW algorithm with default parameters (0.01 weight decay), and an initial learning rate of 3e-4 with exponential decay (0.99 per epoch). For sequential training, models are pretrained for 40h on a A100 GPU, with finetuning requiring up to 20h on a V100. For mixed training, we limit the total training time to 20h on a A100.", + "bbox": [ + 511, + 393, + 906, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results and Analysis", + "text_level": 1, + "bbox": [ + 511, + 554, + 707, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Sequential training", + "text_level": 1, + "bbox": [ + 511, + 575, + 692, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Binary Case. Models pretrained on synthetic datasets outperform those without pretraining on all target datasets (Fig. 3). SyntheWorld maintains commendable performance on the single-class \"building\" change case, similarly to Levir-CD. Pretraining on FSC-180k yields the strongest results, compared to SyntheWorld on all five target datasets, with gains of $2 - 3\\%$ of IoU. Details are provided in the Supplementary Material.", + "bbox": [ + 511, + 595, + 906, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic Case. Pretraining on FSC-180k outperforms the baseline and SyntheWorld on the three SCD datasets in the sequential mode (Tab. G). Indeed, pretraining on SyntheWorld, that focuses solely on the \"building\" class, shows limitations in multiclass tasks on HiUCD-mini and HiUCD-XL compared to the baseline. Our FSC-180k dataset is particularly beneficial for HiUCD-mini and HiUCD-XL. The types of landscapes and spatial resolution are relatively similar, which facilitates effective transfer learning. For HiUCD-mini, our pretraining results in a SeK of 0.19 and a SCS of 0.78, representing respective increases of $11\\%$ and $7\\%$", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4https://wwwcodabench.org/competitions/3485/", + "bbox": [ + 107, + 886, + 367, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2209", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02a086d63be6387b772be6ec1d6d29fb2de78353164f380dd705ef0d9d775761.jpg", + "image_caption": [ + "(a) Levir-CD" + ], + "image_footnote": [], + "bbox": [ + 96, + 114, + 254, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2da65fbc9a55ae153200b48257ae5eb2d7c9990447562f39913f51e1aeadd91b.jpg", + "image_caption": [ + "(b) S2Looking" + ], + "image_footnote": [], + "bbox": [ + 307, + 114, + 442, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bad28aa6f7848be81b4f5270a398e1ecfd421fc506d75d11b1fa83488862f93d.jpg", + "image_caption": [ + "(c) SECOND", + "(d) HiUCD-mini", + "Figure 3. Sequential training in binary change detection. After pretraining our Dual U-Net on either nothing, SyntheWorld or FSC-180k, we finetune and test it on each of the 4 target datasets and represent the binary IoU." + ], + "image_footnote": [], + "bbox": [ + 94, + 232, + 465, + 344 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compared to the baseline. The gains rise to $28\\%$ and $8\\%$ against SyntheWorld. For SECOND, SyntheWorld partially narrows the gap: SeK of 0.17 and a SCS of 0.84, while the FSC-180k's pretrained Dual UNet achieves 0.18 and 0.87.", + "bbox": [ + 89, + 439, + 483, + 500 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9706e316b3ec69f458c61594065ab58eaaff08e34bb75d769608a04ea0f3043d.jpg", + "table_caption": [ + "Table 3. Sequential training in the semantic change detection framework. We report the binary change detection scores (IoU or Chg. mIoU) as well as semantic scores (SCS or mIoU) on three datasets in the sequential scheme. Models were initially pretrained (SyntheWorld or FSC-180k) or not (Baseline)." + ], + "table_footnote": [], + "table_body": "
PretrainingModelSECONDHiUCD-mini
IoUSCSIoUSCS
NoneDual UNet0.530.830.610.73
SyntheWorldDual UNet0.540.840.580.73
FSC-180kDual UNet0.550.890.630.78
NoneSCanNet0.540.840.600.66
FSC-180kSCanNet0.560.890.630.71
", + "bbox": [ + 91, + 594, + 465, + 722 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/44efc8a8f63b864cfb544dd6b4246a5aecf051af0802e02fe6280a5277e1c196.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PretrainingHiUCD-XL
Chg. mIoUmIoU
None0.170.58
SyntheWorld0.170.58
FSC-180k0.190.60
", + "bbox": [ + 166, + 724, + 408, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dependence on a CD model. We compared two CD model by running the same experiments on SECOND and HiUCD-mini with a SCanNet model (Tab. G). We were un", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "able to reproduce the results reported in the original paper [13], despite following the described learning scheme. SCanNet still performs well on the SECOND dataset, achieving a SCS of 0.84, compared to 0.83 for the Dual UNet. The opposite trend is observed on HiUCD, where the Dual UNet reaches a SCS of 0.73, while SCanNet reaches 0.66. As neither of the two models stands out, we mainly present the results obtained with our Dual UNet.", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Mixed training", + "text_level": 1, + "bbox": [ + 511, + 220, + 665, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments show, that the optimal ratio between synthetic and target data lies around $50\\%$ for HiUCD-XL, S2Looking or SECOND (Fig. 4) conversely to the conclusions in [53], where performance always increases with the amount of target data included in the training mix. Nevertheless, the optimal ratio is $90\\%$ for smaller dataset such as Levir-CD or HiUCD-mini, that may be too specific. However, the mixed training strategy proves to be really effective, outperforming the sequential one on HiUCD-XL with a balanced blend as training base, as noticed in Fig. 4, compared to Tab. G. We can deduce that:", + "bbox": [ + 511, + 239, + 906, + 404 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Sequential training maybe suboptimal: The lengthy fine-tuning process can lead the model to \"forget\" pretraining, and instead overfits on the target data, which finally reduces the benefits of pretraining.", + "- In contrast, mixed training augments the initial set and adds diversity throughout the training process. The model sees as many images from the target dataset as in the sequential mode, but continues to benefit from the pretraining dataset.", + "- A \"balanced blend\" of synthetic and target data appears optimal, providing a sufficient frequency of target images for exploitation while strategically incorporating new synthetic images to enable exploratory learning and notably avoid overfitting." + ], + "bbox": [ + 531, + 405, + 906, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.3. Low Data regime", + "text_level": 1, + "bbox": [ + 511, + 625, + 679, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We apply a sequential scheme but limit fine-tuning to small subsets of the target data: specifically, $1\\%$ , $10\\%$ , and $30\\%$ of the training set. Each experiment was repeated 10 times, with a random sampling of the training set. We consider these to be the most meaningful experiments, as this scenario closely matches real-world use cases. The primary objective of pretraining is to develop a model that can either swiftly adapt to new target datasets or perform well when only limited training data is available. One can observe on Fig. 5 a clear advantage for pretrained models at very low level ( $1\\%$ ). With such little data, the model cannot fully learn the specificities of the target dataset and relies primarily on its general knowledge and on the task itself (detecting changes), rather than on dataset-specific features. Unsurprisingly, models pretrained for this task perform significantly better. However, as the amount of available data increases, the gap tends to narrow. Models increasingly rely", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2210", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e8750ab7c06af2f29d18d359bbac0b995fa5a07ad979ad077049a71f206588d.jpg", + "image_caption": [ + "Figure 4. Mixed training (Binary and Semantic). Training on a blend of target and synthetic/hybrid (SyntheWorld or FSC-180k) train sets, containing a ratio of $\\mathrm{x\\%}$ samples from the target (including repetitions). Testing is performed on the target test set. $100\\%$ corresponds to fine-tuning exclusively on target dataset (without pretraining)." + ], + "image_footnote": [], + "bbox": [ + 94, + 94, + 867, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on data itself, to which they have equal access, yet pretraining on FSC-180k still yields better performance than using SyntheWorld or training from scratch.", + "bbox": [ + 89, + 318, + 483, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a9e960ba543022cfa9a7c3f9e06bdee68639d6049ee2b10dc252c47e630835a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 382, + 429, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7501483e6676c6556fbadecb195d60712632400487a232ad0e50b19cbe3475d6.jpg", + "image_caption": [ + "Figure 5. Low data regime (Semantic). Fine-tuning on a limited part of the target train set (either $1\\%$ , $10\\%$ or $30\\%$ of randomly sampled examples) and evaluating on the whole test set. Models are initially pretrained (SyntheWorld or FSC-180k) or not (Baseline). Metrics are averaged on 10 runs for SECOND and HiUCD-mini." + ], + "image_footnote": [], + "bbox": [ + 133, + 515, + 433, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.4. Zero-Shot evaluation", + "text_level": 1, + "bbox": [ + 89, + 746, + 282, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While models pretrained on SyntheWorld struggle to transfer effectively without target sample exposure, pretraining on FSC-180k delivers respectable results, achieving IoU scores of 0.37 on SECOND and 0.36 on HiUCD-mini, showing performance drops of $31\\%$ and $40\\%$ , compared to the fully supervised models (details are provided in the Supplementary Material). However, FSC-180k is less effective on LevirCD, where the performance drop reaches $60\\%$ (0.33 vs. 0.83 IoU). We argue that performance in this scenario is directly", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "linked to dataset similarities: the limited realism in SyntheWorld hinders transferability, while FSC-180k's diverse landscapes and sparse changes further limit its effectiveness. For the semantic task, pretraining on FSC-180k enhances the model's understanding of change trajectories, achieving SCS of 0.24 on SECOND and 0.25 on HiUCD-mini.", + "bbox": [ + 511, + 318, + 906, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.5. Qualitative assessment", + "text_level": 1, + "bbox": [ + 511, + 455, + 718, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative analysis of the various segmentation results indicates that FSC-180k pretraining significantly enhances the network's ability to localize features in input images. This improvement leads to both more accurate segmentation within VHR images and higher semantic prediction accuracy for main classes of interest in the semantic case (e.g., buildings, roads, bare land for HiUCD-mini, buildings or non-vegetated ground surface for SECOND). Qualitative examples for the different datasets are provided in the Supplementary Material.", + "bbox": [ + 511, + 474, + 908, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced HySCDG, a new, simple bi-temporal change detection generation pipeline. It lies on Stable Diffusion and ControlNet for creating hybrid content of VHR images. HySCDG leverages an existing large-scale semantic land cover map to overcome the lack of diversity and size of real-world datasets. We provide FSC-180k, a pretraining dataset based on FLAIR [18], which includes bi-temporal real/in-painted images, and their associated semantic maps. The proposed solution was evaluated for transferability, on five datasets over four distinct training setups. Our experiments on both binary and semantic cases demonstrated strong transferability, confirming FSC-180k robustness and adaptability as well as HySCDG relevance.", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2211", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was granted access to the HPC resources of IDRIS under the allocations AD011014690 and AD011014286 made by GENCI. We thank Ruben Gres for inspiring discussions and valuable feedback about ControlNet and Stable Diffusion.", + "bbox": [ + 89, + 114, + 485, + 189 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andrew. How to use Soft Inpainting - Stable Diffusion Art. https://stable-diffusion-art.com/soft-inpainting/, 2024.5", + "[2] Wele Gedara Chaminda Bandara and Vishal M. Patel. A transformer-based siamese network for change detection. In IGARSS, pages 207-210, 2022. 2, 6", + "[3] Marshall Burke, Anne Driscoll, David B. Lobell, and Stefano Ermon. Using satellite imagery to understand and promote sustainable development. Science, 371(6535), 2021. 1", + "[4] Yinxia Cao and Xin Huang. A full-level fused cross-task transfer learning method for building change detection using noise-robust pretrained networks on crowdsourced labels. Remote Sensing of Environment, 284:113371, 2023. 2", + "[5] Hao Chen, Wenyuan Li, and Zhenwei Shi. Adversarial instance augmentation for building change detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 2, 3", + "[6] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10), 2020. 2, 3, 5, 6", + "[7] Hongruixuan Chen, Jian Song, Chengxi Han, Junshi Xia, and Naoto Yokoya. Changemamba: Remote sensing change detection with spatiotemporal state space model. IEEE Transactions on Geoscience and Remote Sensing, 62:1-20, 2024. 1, 2, 6", + "[8] Isaac Corley, Caleb Robinson, and Anthony Ortiz. A change detection reality check. arXiv preprint arXiv:2402.06994, 2024. 6", + "[9] Mikolaj Czerkawski and Christos Tachtatzis. Exploring the capability of text-to-image diffusion models with structural edge guidance for multi-spectral satellite image inpainting. IEEE Geoscience and Remote Sensing Letters, 2024. 2", + "[10] Mikolaj Czerkawski, Priti Upadhyay, Christopher Davison, Astrid Werkmeister, Javier Cardona, Robert Atkinson, Craig Michie, Ivan Andonovic, Malcolm Macdonald, and Christos Tachtatzis. Deep internal learning for inpainting of cloud-affected regions in satellite imagery. Remote Sensing, 14(6):1342, 2022. 2", + "[11] Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, and Yann Gousseau. Multitask learning for large-scale semantic change detection. Computer Vision and Image Understanding, 187:102783, 2019. 2, 6", + "[12] Rodrigo Caye Daudt, Bertrand Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In ICIP, pages 4063-4067, 2018. 1, 2", + "[13] Lei Ding, Jing Zhang, Haitao Guo, Kai Zhang, Bing Liu, and Lorenzo Bruzzone. Joint spatio-temporal modeling for semantic change detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-14," + ], + "bbox": [ + 93, + 237, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2024. 2, 6, 7", + "[14] Miguel Espinosa and Elliot J. Crowley. Generate your own scotland: Satellite image generation conditioned on maps. NeurIPS 2023 Workshop on Diffusion Models, 2023. 2", + "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 3", + "[16] Weiliang Fan, Jing M Chen, and Weimin Ju. A pixel missing patch inpainting method for remote sensing image. In International Conference on Geoinformatics, 2011. 2", + "[17] Mario Fuentes Reyes, Yuxing Xie, Xiangtian Yuan, Pablo d'Angelo, Franz Kurz, Daniele Cerra, and Jiaojiao Tian. A 2D/3D multimodal data simulation approach with applications on urban semantic segmentation, building extraction and change detection. ISPRS Journal of Photogrammetry and Remote Sensing, 205:74-97, 2023. 1, 2, 3", + "[18] Anatol Garioud, Nicolas Gonthier, Loic Landrieu, Apolline De Wit, Marion Valette, Marc Poupee, Sebastien Giordano, and Boris Wattrelos. FLAIR: a country-scale land cover semantic segmentation dataset from multi-source optical imagery. In NeurIPS Datasets and Benchmarks Track, 2023. 2, 3, 5, 8, 1", + "[19] Adrien Gressin, Nicole Vincent, Clément Mallet, and Nicolas Paparoditis. Semantic approach in image change detection. In ACIVS, 2013. 5", + "[20] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In CVPR, 2023. 5", + "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[22] Philip J Howarth and Gregory M Wickware. Procedures for change detection using landsat digital data. International Journal of Remote Sensing, 2(3):277-291, 1981. 2", + "[23] Wenli Huang, Ye Deng, Siqi Hui, and Jinjun Wang. Image inpainting with bilateral convolution. Remote Sensing, 14(23):6140, 2022. 2", + "[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1125-1134, 2017. 2", + "[25] Shunping Ji, Shiqing Wei, and Meng Lu. Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set. IEEE Transactions on Geoscience and Remote Sensing, 57(1):574-586, 2018. 2, 3", + "[26] Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David B Lobell, and Stefano Ermon. Diffusionsat: A generative foundation model for satellite imagery. In ICLR, 2023. 2, 3", + "[27] Maria Kolos, Anton Marin, Alexey Artemov, and Evgeny Burnaev. Procedural synthesis of remote sensing images for robust change detection with neural networks. In Advances in Neural Networks, 2019. 2, 3", + "[28] Andrey Kuznetsov and Mikhail Gashnikov. Remote sensing image inpainting with generative adversarial networks. In ISDFS, 2020. 2", + "[29] Ang Li, Qiuhong Ke, Xingjun Ma, Haiqin Weng, Zhiyuan" + ], + "bbox": [ + 516, + 92, + 906, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2212", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zong, Feng Xue, and Rui Zhang. Noise doesn't lie: towards universal detection of deep inpainting. In *IJCAI*, 2021. 5", + "[30] Zhuohong Li, Wei He, Jiepan Li, Fangxiao Lu, and Hongyan Zhang. Learning without exact guidance: Updating large-scale high-resolution land cover maps from low-resolution historical labels. In CVPR, 2024. 1", + "[31] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 3", + "[32] Zhiyong Lv, Tongfei Liu, Jón Atli Benediktsson, and Nicola Falco. Land cover change detection techniques: Very-high-resolution optical images: A review. IEEE Geoscience and Remote Sensing Magazine, 10(1):44-63, 2021. 2", + "[33] Yuchi Ma, Shuo Chen, Stefano Ermon, and David B Lobell. Transfer learning in environmental remote sensing. Remote Sensing of Environment, 301:113924, 2024. 3", + "[34] Javier Marín and Sergio Escalera. Sssgan: Satellite style and structure generative adversarial networks. Remote Sensing, 13(19):3984, 2021. 2", + "[35] Fan Meng, Xiaomei Yang, Chenghu Zhou, and Zhi Li. A sparse dictionary learning-based adaptive patch inpainting method for thick clouds removal from high-spatial resolution remote sensing imagery. Sensors, 17(9):2130, 2017. 2", + "[36] Ali Mirzaei, Hossein Bagheri, and Iman Khosravi. Enhancing crop classification accuracy through synthetic SAR-Optical data generation using deep learning. ISPRS International Journal of Geo-Information, 12(11), 2023. 1", + "[37] Hassiba Nemmour and Youcef Chibani. Multiple support vector machines for land cover change detection: An application for mapping urban extensions. ISPRS Journal of Photogrammetry and Remote Sensing, 61(2):125-133, 2006. 2", + "[38] OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org . https : //www.openstreetmap.org, 2017.4", + "[39] Chao Pang, Jiang Wu, Jian Ding, Can Song, and Gui-Song Xia. Detecting building changes with off-nadir aerial images. Science China Information Sciences, 66(4):140306, 2023. 2", + "[40] Daifeng Peng, Lorenzo Bruzzone, Yongjun Zhang, Haiyan Guan, Haiyong Ding, and Xu Huang. Semicdnet: A semisupervised convolutional neural network for change detection in high resolution remote-sensing images. IEEE Transactions on Geoscience and Remote Sensing, 59(7):5891-5906, 2020. 2", + "[41] R.S. Priya and K. Vani. Vegetation change detection and recovery assessment based on post-fire satellite imagery using deep learning. Nature Sci Rep, 14:12611, 2024. 1", + "[42] Xiaofan Qu, Feng Gao, Junyu Dong, Qian Du, and Heng-Chao Li. Change detection in synthetic aperture radar images using a dual-domain network. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 1", + "[43] Gupta Ritwik, Hosfelt Richard, Sajeev Sandra, Patel Nirav, Goodman Bryce, Doshi Jigar, Heim Eric, Choset Howie, and Gaston Matthew. xbd: A dataset for assessing building damage from satellite imagery. arXiv preprint arXiv:1911.09296, 2019. 3", + "[44] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 2, 3" + ], + "bbox": [ + 91, + 90, + 483, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Ribana Roscher, Marc Russwurm, Caroline Gevaert, Michael Kampffmeyer, Jefersson A. Dos Santos, Maria Vakalopoulou, Ronny Hänsch, Stine Hansen, Keiller Nogueira, Jonathan Prexl, and Devis Tuia. Better, not just more: Data-centric machine learning for earth observation. IEEE Geoscience and Remote Sensing Magazine, pages 2-22, 2024. 3", + "[46] Ragav Sachdeva and Andrew Zisserman. The change you want to see. In WACV, 2023. 4, 5", + "[47] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5", + "[48] Minseok Seo, Hakjin Lee, Yongjin Jeon, and Junghoon Seo. Self-pair: Synthesizing changes from single source for object change detection in remote sensing imagery. In WACV, 2023. 2", + "[49] Minseok Seo, Hakjin Lee, Yongjin Jeon, and Junghoon Seo. Self-pair: Synthesizing changes from single source for object change detection in remote sensing imagery. In WACV, 2023. 2", + "[50] Li Shen, Yao Lu, Hao Chen, Hao Wei, Donghai Xie, Jiabao Yue, Rui Chen, Shouye Lv, and Bitao Jiang. S2looking: A satellite side-looking dataset for building change detection. Remote Sensing, 13(24), 2021. 2, 5, 6", + "[51] Wenzhong Shi, Min Zhang, Rui Zhang, Shanxiong Chen, and Zhao Zhan. Change detection based on artificial intelligence: State-of-the-art and challenges. Remote Sensing, 12(10):1688, 2020. 2", + "[52] Ashbindu Singh. Review article digital change detection techniques using remotely-sensed data. International journal of remote sensing, 10(6):989-1003, 1989. 2", + "[53] Jian Song, Hongruixuan Chen, and Naoto Yokoya. Syntheworld: A large-scale synthetic dataset for land cover mapping and building change detection. In WACV, 2024. 2, 3, 5, 7", + "[54] Kai Tang and Jin Chen. Changeanywhere: Sample generation for remote sensing change detection via semantic latent diffusion model. arXiv preprint arXiv:2404.08892, 2024. 2, 3", + "[55] Yizhuo Tang, Zhengtao Cao, Ningbo Guo, and Mingyong Jiang. A Siamese Swin-Unet for image change detection. Nature Sci Rep, 14:4577, 2024. 1, 2", + "[56] Shiqi Tian, Yanfei Zhong, Ailong Ma, and Zhuo Zheng. HiUCD: A large-scale dataset for urban semantic change detection in remote sensing imagery. In NeurIPS Workshop on Machine Learning for the Developing World, 2020. 2, 5, 6", + "[57] Shiqi Tian, Yanfei Zhong, Zhuo Zheng, Ailong Ma, Xicheng Tan, and Liangpei Zhang. Large-scale deep learning based binary and semantic change detection in ultra high resolution remote sensing imagery: From benchmark datasets to urban application. ISPRS Journal of Photogrammetry and Remote Sensing, 193:164-186, 2022. 1, 2, 6", + "[58] Aysim Tucker, Marvin Eisenberger, Daniel Cremers, and Laura Leal-Taixe. Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation. In CVPR, 2024. 2", + "[59] Aysim Toker, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andres Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, Giovanni Marchisio, Xiao Xiang Zhu, and Laura Leal-Taixe." + ], + "bbox": [ + 516, + 90, + 905, + 891 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2213", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In CVPR, 2022. 1, 6", + "[60] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In CVPR, 2021. 1", + "[61] Zhipan Wang, Minduan Xu, Zhongwu Wang, Qing Guo, and Qingling Zhang. Scribblednet: Change detection on high-resolution remote sensing imagery with scribble interaction. International Journal of Applied Earth Observation and Geoinformation, 128:103761, 2024. 1", + "[62] Junshi Xia, Naoto Yokoya, Bruno Adriano, and Clifford Broni-Bediako. OpenEarthMap: A benchmark dataset for global high-resolution land cover mapping. In WACV, 2023. 1, 3", + "[63] Kunping Yang, Gui-Song Xia, Zicheng Liu, Bo Du, Wen Yang, Marcello Pelillo, and Liangpei Zhang. Asymmetric siamese networks for semantic change detection in aerial images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-18, 2022. 2, 5, 6", + "[64] Yang Zhan, Kun Fu, Menglong Yan, Xian Sun, Hongqi Wang, and Xiaosong Qiu. Change detection based on deep siamese convolutional network for optical aerial images. IEEE Geoscience and Remote Sensing Letters, 14(10):1845-1849, 2017. 2", + "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 4, 1", + "[66] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3", + "[67] Xiaokang Zhang, Weikang Yu, Man-On Pun, and Wenzhong Shi. Cross-domain landslide mapping from large-scale remote sensing images using prototype-guided domain-aware progressive representation learning. ISPRS Journal of Photogrammetry and Remote Sensing, 197:1-17, 2023. 2", + "[68] Wen-Jie Zheng, Xi-Le Zhao, Yu-Bang Zheng, and Zhi-Feng Pang. Nonlocal patch-based fully connected tensor network decomposition for multispectral image inpainting. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2021. 2", + "[69] Zhuo Zheng, Stefano Ermon, Dongjun Kim, Liangpei Zhang, and Yanfei Zhong. Changen2: Multi-temporal remote sensing generative change foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2, 3", + "[70] Zhuo Zheng, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery. In ICCV, 2021. 1, 2", + "[71] Zhuo Zheng, Shiqi Tian, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Scalable multi-temporal remote sensing change data generation via simulating stochastic change process. In ICCV, 2023. 2, 3", + "[72] Zhuo Zheng, Yanfei Zhong, Shiqi Tian, Ailong Ma, and Liangpei Zhang. Changemask: Deep multi-task encoder-transformer-decoder architecture for semantic change detection. ISPRS Journal of Photogrammetry and Remote Sensing, 183:228-239, 2022. 2", + "[73] Zhuo Zheng, Yanfei Zhong, Ji Zhao, Ailong Ma, and Liangpei Zhang. Unifying remote sensing change detection via" + ], + "bbox": [ + 91, + 90, + 483, + 892 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep probabilistic change models: From principles, models to applications. ISPRS Journal of Photogrammetry and Remote Sensing, 215:239-255, 2024. 1, 2", + "[74] Zhe Zhu, Shi Qiu, and Su Ye. Remote sensing of land change: A multifaceted perspective. Remote Sensing of Environment, 282:113266, 2022. 1" + ], + "bbox": [ + 516, + 90, + 906, + 172 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2214", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_model.json b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e189e3d6392d495f5295cddf0ebdfa469be55f69 --- /dev/null +++ b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_model.json @@ -0,0 +1,2675 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.907, + 0.177 + ], + "angle": 0, + "content": "The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.204, + 0.248, + 0.22 + ], + "angle": 0, + "content": "Yanis Benidir" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.204, + 0.56, + 0.22 + ], + "angle": 0, + "content": "Nicolas Gonthier" + }, + { + "type": "text", + "bbox": [ + 0.733, + 0.204, + 0.86, + 0.22 + ], + "angle": 0, + "content": "Clément Mallet" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.222, + 0.692, + 0.239 + ], + "angle": 0, + "content": "Univ Gustave Eiffel, ENSG, IGN, LASTIG, France" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.243, + 0.601, + 0.256 + ], + "angle": 0, + "content": "firstname_lastname@ign.fr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.687 + ], + "angle": 0, + "content": "Bi-temporal change detection at scale based on Very High Resolution (VHR) images is crucial for Earth monitoring. This remains poorly addressed so far: methods either require large volumes of annotated data (semantic case), or are limited to restricted datasets (binary set-ups). Most approaches do not exhibit the versatility required for temporal and spatial adaptation: simplicity in architecture design and pretraining on realistic and comprehensive datasets. Synthetic datasets are the key solution but still fail to handle complex and diverse scenes. In this paper, we present HySCDG a generative pipeline for creating a large hybrid semantic change detection dataset that contains both real VHR images and inpainted ones, along with land cover semantic map at both dates and the change map. Being semantically and spatially guided, HySCDG generates realistic images, leading to a comprehensive and hybrid transfer-proof dataset FSC-180k. We evaluate FSC-180k on five change detection cases (binary and semantic), from zero-shot to mixed and sequential training, and also under low data regime training. Experiments demonstrate that pretraining on our hybrid dataset leads to a significant performance boost, outperforming SyntheWorld, a fully synthetic dataset, in every configuration. All codes, models, and data are available here: https://yb23.github.io/projects/cywd/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.725, + 0.222, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Efficient methods for change detection (CD) are crucial for monitoring territories and the various phenomena and activities that impact them [3]. As human and climatic pressures increase, this topic is of utmost importance and calls every year for higher spatial resolution and semantic accuracy [55]. Change detection at scale has benefited greatly from the contributions of deep learning in recent years [7, 12, 69, 73], but most methods are grounded on large volumes of annotated data [41, 60]. Creating a large-scale dataset for bi-temporal (pair-wise) remote sensing change detection poses signifi" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.288, + 0.882, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.554, + 0.751, + 0.565 + ], + "angle": 0, + "content": "Target datasets" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.575, + 0.909, + 0.677 + ], + "angle": 0, + "content": "Figure 1. Efficient and scalable change detection requires a comprehensive training dataset that does not exist today. Using a single-temporal dataset (image+semantic map), we propose HySCDG that generates a novel bi-temporal hybrid dataset FSC-180k. This enables multiple transfer learning possibilities on either binary and semantic change detection tasks. \\(\\mathcal{O} / \\mathcal{O}\\) : Transfer with/without fine-tuning our model, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.81 + ], + "angle": 0, + "content": "cant challenges and costs, especially when dealing with Very High Resolution (VHR) images (ground sampling distance: GSD \\(< 1\\mathrm{m}\\), [30]). This is due to the need for expertise and considerable effort in collecting, preprocessing, and annotating images [59, 62], particularly in the case of VHR data and multiple labels (termed semantic change detection, SCD) [57, 61, 70, 74]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "As an alternative, synthetic data generation is a promising direction for bi-temporal labeled data provision from a single temporal domain that is successful in diverse areas such as urban sprawl [17], disaster or climate change monitoring [42] or crop mapping [36]. Two main paradigms exist for CD: fully synthetic datasets (mainly graphics-based" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2204" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.453 + ], + "angle": 0, + "content": "[27, 53]), and hybrid solutions that mix real and modified images, often at the instance level [5, 69, 71], known as data augmentation. The first method synthesizes image pairs by rendering two versions of a 3D virtual world thanks to a computer graphics engine. On the other hand, hybrid approaches synthesize new image pairs by pasting or erasing object instances into existing images ( inpainting). They propose a suitable trade-off and faster generation of large quantities of data, with likely higher realism. However, none of these solutions meet the requirements for scalable SCD with VHR imagery (Tab. 1): focus on single tasks, restricted geographical areas, limited diversity, no guarantee on semantic consistency between pairs [48] and subsequently poor transfer to real datasets. To address this challenge, we introduce a generative pipeline built upon Stable Diffusion [44] and ControlNet [65], designed to harness an existing VHR semantic land cover dataset [18] and instance masks, to generate an extensive hybrid semantic bi-temporal change detection dataset encompassing both real and inpainted images, semantic land cover maps at each date, and a change map (Fig. 1). Then, we evaluate the transferability of our synthetic dataset on five datasets for both binary and semantic change detection cases ranging from zero-shot to sequential and mixed training set-ups, as well as under low-data regime." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.455, + 0.331, + 0.468 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.469, + 0.483, + 0.574 + ], + "angle": 0, + "content": "- We present a generation pipeline HySCDG for creating VHR bi-temporal SCD datasets based on a land cover real-world dataset by using Stable Diffusion and ControlNet with tunable and semantically-guided inpainting of individual objects. It can be easily configured or finetuned on any land-cover dataset, generating images with the dataset's distinctive traits." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.575, + 0.483, + 0.635 + ], + "angle": 0, + "content": "- We provide FSC-180k: a Hybrid Semantic Change Detection pretraining dataset, created from FLAIR [18] along with the instance masks for 300k objects for the corresponding images." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.636, + 0.483, + 0.68 + ], + "angle": 0, + "content": "- We provide a comprehensive evaluation of the proposed synthetic dataset on multiple transfer learning scenarios on five change datasets, in particular for SCD." + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.469, + 0.483, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.228, + 0.71 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Detecting changes in the deep learning era. We target bi-temporal land cover change detection in VHR images [32]. This historical task [22, 37, 52] consists in detecting the meaningful elements of a given area that have changed between two acquisitions and providing the land cover categories at each period (also known as change trajectory). It presents challenges: radiometric variations, geometric variations and non-permanent changes. They have been extensively addressed with supervised deep learning techniques [70]: from CNN [11, 12, 64] to Transformers [2, 7, 13, 55, 72, 73], mainly with Siamese architecture. Multiple datasets, accompanied with ad-hoc methods, have been" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.244 + ], + "angle": 0, + "content": "created to feed such models. They are often limited by their size [6, 25, 39, 56, 63], geographical extent [6, 56, 57, 63], low geometrical or annotation quality [11]. Also, most of the effort focuses on binary change detection, i.e., restricted to a unique category, mainly buildings [6, 25, 39, 40, 50] or disasters [67]. Such scarcity in high quality, diverse, large-scale labels is alleviated with transfer learning [4, 51], by fine-tuning models pretrained on large-scale datasets for different image domains, even for semantic change detection [56, 57, 63]." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.262, + 0.909, + 0.277 + ], + "angle": 0, + "content": "Synthesizing and inpainting remote sensing images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.909, + 0.55 + ], + "angle": 0, + "content": "Synthetic remote sensing data is useful e.g., for cloud removal [10], image restoration [16], or training deep supervised models [58], before transfer to real examples. Solutions based on patches, [35, 68], autoencoders [23], GAN [28, 34], -aligned generation [24] have been proposed, but with limited image quality and no semantic control. Diffusion models (DM) have improved such quality [44], but most models have applied DM directly on mid-resolution RGB images [14, 26, 58], neglecting both the multi-spectral dimension [9] and VHR spaces. However, a key finding remains: semantically control the synthesis, e.g., by directly synthesizing text [26], semantic maps [58] or with a control module guided by edges [9], semantic maps [14] or metadata [26]. In our work, we fine-tune a Stable Diffusion and a ControlNet models to achieve semantically-controlled inpainting of VHR images as a basis to take advantage of existing images, and generate an hybrid semantic change detection dataset at scale and with high diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Generating synthetic change datasets. Two main approaches exist: fully synthetic or hybrid solutions. The first paradigm leverages advanced 3D rendering engines to create fully synthetic datasets [17, 27]. This procedural approach allows complete control over rendering parameters (instance placement, semantic classes, lighting conditions, etc.) generated from scratch. E.g., SyntheWorld [53] effectively leverages DM prompted by GPT-4 to create a diverse and large-scale fully synthetic dataset. On the other hand, hybrid approaches generate (semi-)synthetic data by inserting fake changes into real images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.796 + ], + "angle": 0, + "content": "Such modifications can be random crops [49], instance copy-pasting, or inpainting made by GAN [5, 49, 71] or DM [54, 69]. Blending or style transfer methods are often used to enhance diversity [5, 49]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "A pivotal challenge is the semantic guidance: modifying few elements of images to mimic real scarce changes, e.g., with instance footprints extracted from external datasets [5] or segmentation model outputs [69]. [54] adopt a Denoising diffusion probabilistic model to generate a new image directly from a modified semantic map and a degraded version of the corresponding snapshot. Changen2 [69] is the closest" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2205" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.243 + ], + "angle": 0, + "content": "work to ours, with a full pipeline to synthesize a hybrid semantic dataset, again with a DM conditioned by a semantic map. However, Changen2 creates synthetic data specific to a given target dataset in terms of image resolution, size, or change characteristics. In this paper, we explore semantically controlled inpainting based on DM and ControlNet to the intertwined challenge of diverse and multi-class modification of VHR images. It is termed HySCDG and leads to the hybrid dataset named FSC-180k. Tab. 1 provides a comparison with main synthetic change detection datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.257, + 0.485, + 0.355 + ], + "angle": 0, + "content": "Table 1. Synthetic Change Datasets. Land cover change detection in remote sensing imagery. For the total number of pixels, we only consider strictly different images. The geographical zone's extension can be local [ ] national [ ] or global [ ] We denote open-access (OA) and directly downloadable datasets with ✓. In the case of IAug, we sum the number of pixels of the two datasets provided in this work. OEM means OpenEarthMap." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.365, + 0.482, + 0.521 + ], + "angle": 0, + "content": "
DatasetOAPixels ×106GSD (m)ClassesSourceZoneType
SynCW [27]X370.61X田Synthetic
SMARS [17]✓1100.3-0.52X田Synthetic
IAug [5]X1,1670.075-0.51LEVIR-CD [6], WHU-CD [25]田Hybrid
Ce-100K [54]X6,5530.25-0.58OEM [62]田Synthetic
Changen2 [69]X7,0770.25-0.58OEM [62]田Hybrid
Changen [71]X11,7960.81xView2 [43]田Hybrid
SyntheWorld [53]✓18,3500.3-11XXSynthetic
FSC-180k (Ours)✓80,7400.216FLAIR [18]田Hybrid
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.559, + 0.486, + 0.74 + ], + "angle": 0, + "content": "Transfer learning from synthetic/hybrid datasets. Transfer learning is used to assess the validity of the generated image pairs, and vary with the operational set-up [33]. SCD at scale is performed through sequential learning: pretraining on such datasets and fine-tuning on real-world cases [5, 27, 69, 71]. Fewer works [53] evaluate on mixed training sets (single training mixing both real and synthetic samples), which can help to prevent overfitting. An orthogonal solution considers low data regimes, i.e., fine-tuning on a very small train set [5, 53, 54], or even under zero-shot scenarios [54]. We evaluate our proposed dataset in the four conceivable transfer scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.442, + 0.772 + ], + "angle": 0, + "content": "3. Hybrid generation of semantic changes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Much of the success of deep learning methods for Earth observation relies on large-scale training datasets with reliable annotations [18, 45]. However, no large-scale publicly available dataset exists for bi-temporal SCD: we propose HySCDG for Hybrid Semantic Change Detection Generation, a procedure to leverage existing large-scale land cover datasets and state-of-the-art image inpainting methods to simulate visually realistic bi-temporal semantic changes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.866, + 0.106 + ], + "angle": 0, + "content": "This leads to the generation of the FSC-180k dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Our paper is grounded on two main ideas: (i) Adapting and fine-tuning a Stable Diffusion Model enables efficient inpainting of VHR images with a semantic control and geographical instance selection; (ii) Changes can be simulated with sufficient diversity by randomly selecting instances (and modifying their labels) within an existing land cover map that is paired with the image used as input. The pipeline is shown in Fig. 2. Details are provided in the following." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.235, + 0.9, + 0.251 + ], + "angle": 0, + "content": "3.1. Adaptation of Stable Diffusion for Inpainting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.909, + 0.515 + ], + "angle": 0, + "content": "We use Stable Diffusion (SD) [44], an open and state-of-the-art text-to-image synthesis method. Our objective is to modify SD to manage controllable inpainting specifically in remote sensing aerial images, for various landscapes or semantic classes. Though SD models are effective in generating natural images with clear focal points and layered depth, with artistic or cinematic styles, they require adaptation to be used for remote sensing images [26]: acquisition distance, intricate textures, no domination of a single element. Starting from a Stable Diffusion 2 Inpainting checkpoint1, we sequentially fine-tune the Autoencoder and the Diffusion U-Net, before adding and training a ControlNet [65] that leverages information extracted from an existing large-scale land cover dataset (namely, FLAIR [18], see Sec. 3.3). Our pipeline allows to generate 5-band images that incorporate infrared and elevation along with RGB channels, for the first time in satellite image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.909, + 0.742 + ], + "angle": 0, + "content": "Adaptation of the Variational Autoencoder. Unlike most SD extensions, which focus on fine-tuning the central U-Net, we also adapted the Variational Autoencoder (VAE) that encodes images into the diffusion latent space. This adaptation targets efficient encoding of multi-band VHR images from the FLAIR dataset. The original VAE, trained with a KL-regularization factor [15], uses a combination of L1 and perceptual loss [66] alongside KL-divergence. To improve compression of remote sensing images, we added both L2 and Focal losses [31]: minimizing high-frequency errors reduces blurriness. Additionally, we included a \"color loss\" (L2 loss on each \\(5 \\times 5\\) patch) to limit oversaturation, which affects ground object color realism. Fine-tuning of the VAE took 160 hours A100 GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.759, + 0.909, + 0.88 + ], + "angle": 0, + "content": "Stable Diffusion training. We then fine-tuned the Diffusion U-Net for inpainting, adapting it to the new VAE and teaching it to synthesize and inpaint VHR images. Special attention was given to prompt selection. Although inpainting is the model's primary function, we included direct image generation from prompts in \\(10\\%\\) to \\(20\\%\\) of cases, as in standard SD. This training required 300 hours on a A100 GPU (30,000 iterations, batch size of 32)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.868, + 0.901 + ], + "angle": 0, + "content": "1https://huggingface.co/stabilityai/stable-diffusion-2-inpainting" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2206" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.087, + 0.848, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.38, + 0.908, + 0.452 + ], + "angle": 0, + "content": "Figure 2. HySCDG pipeline. From a single-temporal dataset composed of one VHR image \\( I_{1} \\), a semantic map \\( M_{1} \\), and some openly available labeled instances, we generate a new VHR image \\( I_{2} \\), a new map \\( M_{2} \\) and subsequently a change map \\( C \\). This results in the FSC-180k hybrid dataset. The two pivotal novelties consist in: (i) Adapting and fine-tuning a Stable-Diffusion Model from image inpainting and (ii) exploiting open geospatial data for inpainting prompt control and semantically guiding the objects to be modified. The combination of both solutions ensures a diverse at-scale VHR multi-class change detection dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.582 + ], + "angle": 0, + "content": "Prompt Engineering. Prompts are designed using both each image's geographic coordinates and open land cover data (OpenStreetMap [38]). To avoid learning irrelevant details, prompts are composed of three elements: (i) Spatial (place name, city, region), (ii) Temporal (time of day, season), (iii) Semantic (frequent classes in the inpainted area, compared to the overall dataset)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.583, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Prompt example: \"Grass and agricultural vegetation next to a highway, locality of Savigny-en-Revermont, Bourgogne-Franche-Comté, in the morning, during Summer.\"" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.654 + ], + "angle": 0, + "content": "3.2. Conditional change inpainting with ControlNet" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Driven by our goal to control the new classes in the inpainted zones, we add a ControlNet (CN) architecture [65] to build a model that generates realistic VHR images from a semantic segmentation map (here land cover). CN is designed to add task-specific conditioning to a pretrained image DM. It allows injecting explicit semantic guidance during the diffusion process, conditioning the outputs on some reference image (a color map of the semantic class), in addition to a text prompt. Our training method enables CN to perform both inpainting and text-to-image synthesis, using the semantic condition. Training took 240 hours on a A100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.834, + 0.483, + 0.866 + ], + "angle": 0, + "content": "3.3. Select, Mask, Change, Inpaint : the HySCDG pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Inspired by [46], the core of our method lies in our \"Select and Inpaint\" mechanism. Given a VHR image along with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.598 + ], + "angle": 0, + "content": "its semantic segmentation and instances\\(^2\\), we inpaint these instances using the SD+CN inpainting method: the label of the instance is changed, the texture of the image is substituted accordingly. This results in a pair of VHR images, both with semantic maps and the corresponding semantic change maps. The complete HySCDG is provided in Fig. 2 and detailed below. It is illustrated for one image but is applied several times to each image of the single temporal dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.906, + 0.652 + ], + "angle": 0, + "content": "HySCDG uncovered. For an image \\( I_{1} \\) with its semantic map \\( M_{1} \\), and \\( \\{O_i\\} \\) the set of \\( N_I \\) instances lying on the extent of \\( I_{1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.653, + 0.875, + 0.668 + ], + "angle": 0, + "content": "- Randomly select \\( N_{change} \\) instances among \\( \\{O_i\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.668, + 0.907, + 0.697 + ], + "angle": 0, + "content": "- For each instance, generate an inpainting mask: footprint+spatial buffer." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.698, + 0.907, + 0.758 + ], + "angle": 0, + "content": "- Using \\( M_1 \\), take as “\\( T_1 \\) class” the most frequent class inside the footprint: \\( c_1 \\). Then, select the larger convex hull of pixels of this class inside the buffered zone. These pixels correspond to the change mask: \\( C \\)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.759, + 0.905, + 0.788 + ], + "angle": 0, + "content": "- Randomly select another class \\( c_2 \\) at \\( T_2 \\) and replace the previous label by this new label." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.789, + 0.907, + 0.834 + ], + "angle": 0, + "content": "- Randomly select \\(N_{\\text{nochange}}\\) random shapes in the image: such masks appear only on the inpainting mask, not on the change mask, letting their labels unchanged." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.834, + 0.907, + 0.849 + ], + "angle": 0, + "content": "- Apply \\(\\mathrm{SD + CN}\\) to masked image and new semantic map." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.85, + 0.907, + 0.865 + ], + "angle": 0, + "content": "- Cumulate the previous steps: we obtain our inpainting" + }, + { + "type": "list", + "bbox": [ + 0.533, + 0.653, + 0.907, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{2} \\) A panoptic segmentation is the most suitable case but not mandatory, assuming a sufficient number of instances ensures enough diversity." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2207" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.122 + ], + "angle": 0, + "content": "mask, change mask, and for \\(T_{2}\\) a new semantic map \\(M_2\\) and the corresponding new image \\(I_{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.122, + 0.483, + 0.168 + ], + "angle": 0, + "content": "The pair of images \\((I_1, I_2)\\) with their corresponding semantic maps \\((M_1, M_2)\\), and the change map \\(C\\) (difference between \\(M_2\\) and \\(M_1\\)), are the samples of the hybrid dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.484, + 0.295 + ], + "angle": 0, + "content": "Label and instance extraction. For comprehensive multiclass learning, we select FLAIR dataset [18], a large-scale semantic segmentation map for 16 land cover classes over more than \\(800\\mathrm{km}^2\\). FLAIR offers the advantage of covering areas where geographical instances of many FLAIR labels are publicly available3, alleviating requiring a panoptic segmentation. Around 300,000 instance masks have been extracted for this work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.455 + ], + "angle": 0, + "content": "Inpainting signature mitigation. Inpainting produces realistic changes but new areas tend to have sometimes a high-frequency signature, as reported by other works [20, 29]. In order to discourage the model from simply learning this inpainting noise instead of learning the actual changes between images, we also inpaint random areas of the image without changing the semantic class in a similar way to the \"not changed regions\" from [46]. Besides, to obtain the best transition between the original image and the inpainted area, we use a continuous mask in the inference [1]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Selection of the new label. For each change area, we randomly select the new class \\( c_2 \\) among all classes weighted by their frequency in the whole dataset divided by their frequencies in the change area." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.605 + ], + "angle": 0, + "content": "Adoption of a buffer zone. For inpainting, we add a buffer around the footprint of each modified instance. This gives SD more freedom to generate smoother borders around the changed areas. It also helps handling small spatial discrepancies between footprints and the VHR image content [19]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.615, + 0.391, + 0.629 + ], + "angle": 0, + "content": "3.4. Structure of the FSC-180k dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.878 + ], + "angle": 0, + "content": "For each image in the FLAIR training set, we generate three unique modified images, resulting in 180,000 synthetic images, each with a corresponding \\(512 \\times 512\\) semantic map covering 16 distinct semantic classes. We call this dataset FSC-180k, short for FLAIR Synthetic Change. FSC-180k includes approximately 80 billion pixels with a ground resolution of \\(0.2\\mathrm{m}\\) per pixel. Note that we can expand the dataset by pairing synthetic images that were generated (differently) from a single real image, process that we use at training time, effectively doubling the number of image pairs. Tab. 1 compares FSC-180k with other synthetic change detection datasets and shows that FSC-180k offers a much broader range of semantic changes beyond simple building modifications, is of higher spatial resolution, and is nearly ten times larger than comparable datasets. It should also be noted that the proportion of changes in FSC-180k is closer to the real" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.122 + ], + "angle": 0, + "content": "scenarios observed in Europe (prevalence of \\(5\\%\\) change) than other academic datasets (see Tab. 2), excepting HiUCD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.143, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Quality assessment of FSC-180k. ControlNet is the most error-prone module: it may sometimes fail to respect the semantic map. We evaluated how unchanged the semantic map remains by performing 15-class semantic segmentation on real and generated images, thanks to a UNet trained for this task, leading to a satisfactory error rate below \\(20\\%\\) between both outputs. We also leveraged standard metrics for image generation evaluation to assess the realism of the synthesis. We got an Inception Score [47] of 6.2 on the inpainted images, which is comparable to real images, and a low FID score [21] of 0.43 between inpainted and real images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.339, + 0.646, + 0.355 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.908, + 0.425 + ], + "angle": 0, + "content": "We conduct various transfer learning experiments to validate hybrid data pretraining, comparing FSC-180k with the fully artificial SyntheWorld [53], the only available synthetic dataset with a size comparable to ours (see Tab. 1)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.435, + 0.851, + 0.451 + ], + "angle": 0, + "content": "4.1. Pretraining approaches with FSC-180k" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.907, + 0.577 + ], + "angle": 0, + "content": "We assess FSC-180k's effectiveness by exploiting existing open datasets as targets under various settings (see Tab. 2 and Supplementary Material). Here, we first present the different transfer learning approaches adopted, and then, dive into the deep models and datasets considered (Sec. 4.2). Our objective is to demonstrate how pretraining on a hybrid dataset improves performance, for both binary and semantic change detection." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.578, + 0.668, + 0.592 + ], + "angle": 0, + "content": "The four scenarios are:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.593, + 0.905, + 0.623 + ], + "angle": 0, + "content": "- Sequential Transfer Learning. Models are pretrained on FSC-180k, and then fine-tuned on the target data." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.624, + 0.905, + 0.653 + ], + "angle": 0, + "content": "- Low Data Regime. The same as above but with a fraction of the available training set." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.654, + 0.905, + 0.683 + ], + "angle": 0, + "content": "- Mixed Training. Models are trained on a blend of FSC-180k and the target data." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.684, + 0.907, + 0.729 + ], + "angle": 0, + "content": "- Zero-Shot Learning. A model pretrained on FSC-180k is directly applied on the target dataset, with only a remapping of FSC-180k classes to align labels." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.593, + 0.907, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.724, + 0.755 + ], + "angle": 0, + "content": "4.2. Experimental protocol" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.753, + 0.776 + ], + "angle": 0, + "content": "4.2.1. Target datasets and metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Our transfer experiments span five diverse datasets, among which three contain semantic labels (Tab. 2). This ensures variety in landscapes, sizes, and change types, i.e., a comprehensive evaluation. Among the plethora of (S)CD datasets, we selected the most popular ones: Levir-CD [6], S2Looking [50], SECOND [63] and HiUCD-mini [56]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We used classical metrics from the change detection literature: Intersection over Union (IoU) and F1 score for binary" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.368, + 0.901 + ], + "angle": 0, + "content": "3https://www.data.gouv.fr/en/datasets/bd-topo-r/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2208" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.167 + ], + "angle": 0, + "content": "cases and mainly the Semantic Change Segmentation score (SCS) from [59] for semantic evaluation. In this latter case, the Separated Kappa (SeK) from [63] is also informative though it tends to be more biased towards detection task than semantic prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.167, + 0.483, + 0.228 + ], + "angle": 0, + "content": "We added some experiments related to HiUCD-XL [57], a follow-up dataset of HiUCD-mini [56] using semantic mIoU and change mIoU, metrics provided on the evaluation platform4 as test labels are kept private." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.239, + 0.483, + 0.267 + ], + "angle": 0, + "content": "Table 2. Target datasets. Overview of the real-world datasets used for our transfer learning experiments." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.278, + 0.49, + 0.408 + ], + "angle": 0, + "content": "
DatasetChange (%)Pixels ×106GSD (m)ClassesPeriodZone
Levir-CD [6]28.41,0670.512002/18Texas
S2Looking [50]25.68,3890.3-0.812017/20World
SECOND [63]20.01,2220.3-0.66?China
HiUCD-mini [56]6.65670.192017/19Estonia
HiUCD-XL [57]1.210,6950.192017/19Estonia
FSC-180k (Ours)7.080,7400.2162018/21France
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.426, + 0.271, + 0.439 + ], + "angle": 0, + "content": "4.2.2. Model architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.444, + 0.483, + 0.641 + ], + "angle": 0, + "content": "We consider that designing a new architecture is out of the scope of this paper and that numerous solutions have already been proposed (see Sec. 2). Therefore, we initially used a simple U-Net model [8] with a segmentation head for semantic change detection. Although effective for binary change, it lacked expressiveness for semantic tasks. Similarly to [11], we developed a simple Dual UNet model composed of 2 nearly identical parallel UNets for (i) change detection, and (ii) semantic segmentation, with ResNet-50 backbones pretrained on ImageNet and trained in a multi-task setting. The change branch incorporates features from the semantic encoder. We keep the training strategy simple and constant to focus on data and transfer, this enhances interpretability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.483, + 0.833 + ], + "angle": 0, + "content": "Can a binary model be sufficient? We tested both a pure binary CD model and the Dual UNet, for which, when, switching to binary-only tasks, the semantic decoder is inactive but the encoding part remains active (as it provides features to the change detection branch). Such selective change detection ability allows training on a multi-class dataset and focusing on specific binary tasks. Experiments show that even for binary tasks, the Dual UNet outperforms the binary-only model, e.g., achieving an IoU of 0.83 on Levir-CD (compared to 0.80). Learning more information than just change pixels, the Dual UNet gains a deeper scene understanding, leading to more reliable predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.85, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Existing architectures. We tested more complex transformer-based architectures, namely ChangeFormer [2]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "MambaCD [7], and SCanNet [13], both on HiUCD-mini and SECOND datasets. SCanNet performed best, balancing high performance and straightforward implementation due to its lightweight design (see Supplementary Material). SCanNet is kept hereafter for comparisons (see Tab. G)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.905, + 0.246 + ], + "angle": 0, + "content": "Transfer for varying semantic classes. In the semantic case, we faced the issue of distinct classes between source and target datasets. This required to choose which classes should be kept for model training." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.247, + 0.907, + 0.321 + ], + "angle": 0, + "content": "- Sequential case: for both pretraining and fine-tuning strategies, we use the classes of the processed dataset. The semantic head of the model is reinitialized and adapted to the new number of classes before fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.322, + 0.907, + 0.367 + ], + "angle": 0, + "content": "- Mixed case: Classes from the target dataset are used, manually mapping pretraining classes to their closest counterparts." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.247, + 0.907, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.375, + 0.706, + 0.39 + ], + "angle": 0, + "content": "4.2.3. Experimental setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.545 + ], + "angle": 0, + "content": "We apply basic augmentations on the images: horizontal and vertical flips, \\(90^{\\circ}\\) rotations, and small rescaled crops (scales above 0.8). Normalization by mean and variance from the target dataset is used in all experiments. For optimization, we use the AdamW algorithm with default parameters (0.01 weight decay), and an initial learning rate of 3e-4 with exponential decay (0.99 per epoch). For sequential training, models are pretrained for 40h on a A100 GPU, with finetuning requiring up to 20h on a V100. For mixed training, we limit the total training time to 20h on a A100." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.555, + 0.709, + 0.57 + ], + "angle": 0, + "content": "4.3. Results and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.694, + 0.592 + ], + "angle": 0, + "content": "4.3.1. Sequential training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Binary Case. Models pretrained on synthetic datasets outperform those without pretraining on all target datasets (Fig. 3). SyntheWorld maintains commendable performance on the single-class \"building\" change case, similarly to Levir-CD. Pretraining on FSC-180k yields the strongest results, compared to SyntheWorld on all five target datasets, with gains of \\(2 - 3\\%\\) of IoU. Details are provided in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Semantic Case. Pretraining on FSC-180k outperforms the baseline and SyntheWorld on the three SCD datasets in the sequential mode (Tab. G). Indeed, pretraining on SyntheWorld, that focuses solely on the \"building\" class, shows limitations in multiclass tasks on HiUCD-mini and HiUCD-XL compared to the baseline. Our FSC-180k dataset is particularly beneficial for HiUCD-mini and HiUCD-XL. The types of landscapes and spatial resolution are relatively similar, which facilitates effective transfer learning. For HiUCD-mini, our pretraining results in a SeK of 0.19 and a SCS of 0.78, representing respective increases of \\(11\\%\\) and \\(7\\%\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.368, + 0.9 + ], + "angle": 0, + "content": "4https://wwwcodabench.org/competitions/3485/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2209" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.098, + 0.237, + 0.11 + ], + "angle": 0, + "content": "(a) Levir-CD" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.116, + 0.255, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.098, + 0.43, + 0.11 + ], + "angle": 0, + "content": "(b) S2Looking" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.116, + 0.443, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.215, + 0.238, + 0.226 + ], + "angle": 0, + "content": "(c) SECOND" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.233, + 0.466, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.215, + 0.435, + 0.226 + ], + "angle": 0, + "content": "(d) HiUCD-mini" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Sequential training in binary change detection. After pretraining our Dual U-Net on either nothing, SyntheWorld or FSC-180k, we finetune and test it on each of the 4 target datasets and represent the binary IoU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.501 + ], + "angle": 0, + "content": "compared to the baseline. The gains rise to \\(28\\%\\) and \\(8\\%\\) against SyntheWorld. For SECOND, SyntheWorld partially narrows the gap: SeK of 0.17 and a SCS of 0.84, while the FSC-180k's pretrained Dual UNet achieves 0.18 and 0.87." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Table 3. Sequential training in the semantic change detection framework. We report the binary change detection scores (IoU or Chg. mIoU) as well as semantic scores (SCS or mIoU) on three datasets in the sequential scheme. Models were initially pretrained (SyntheWorld or FSC-180k) or not (Baseline)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.595, + 0.466, + 0.723 + ], + "angle": 0, + "content": "
PretrainingModelSECONDHiUCD-mini
IoUSCSIoUSCS
NoneDual UNet0.530.830.610.73
SyntheWorldDual UNet0.540.840.580.73
FSC-180kDual UNet0.550.890.630.78
NoneSCanNet0.540.840.600.66
FSC-180kSCanNet0.560.890.630.71
" + }, + { + "type": "table", + "bbox": [ + 0.168, + 0.725, + 0.41, + 0.82 + ], + "angle": 0, + "content": "
PretrainingHiUCD-XL
Chg. mIoUmIoU
None0.170.58
SyntheWorld0.170.58
FSC-180k0.190.60
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Dependence on a CD model. We compared two CD model by running the same experiments on SECOND and HiUCD-mini with a SCanNet model (Tab. G). We were un" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.212 + ], + "angle": 0, + "content": "able to reproduce the results reported in the original paper [13], despite following the described learning scheme. SCanNet still performs well on the SECOND dataset, achieving a SCS of 0.84, compared to 0.83 for the Dual UNet. The opposite trend is observed on HiUCD, where the Dual UNet reaches a SCS of 0.73, while SCanNet reaches 0.66. As neither of the two models stands out, we mainly present the results obtained with our Dual UNet." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.221, + 0.666, + 0.236 + ], + "angle": 0, + "content": "4.3.2. Mixed training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.24, + 0.908, + 0.405 + ], + "angle": 0, + "content": "Our experiments show, that the optimal ratio between synthetic and target data lies around \\(50\\%\\) for HiUCD-XL, S2Looking or SECOND (Fig. 4) conversely to the conclusions in [53], where performance always increases with the amount of target data included in the training mix. Nevertheless, the optimal ratio is \\(90\\%\\) for smaller dataset such as Levir-CD or HiUCD-mini, that may be too specific. However, the mixed training strategy proves to be really effective, outperforming the sequential one on HiUCD-XL with a balanced blend as training base, as noticed in Fig. 4, compared to Tab. G. We can deduce that:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.406, + 0.907, + 0.467 + ], + "angle": 0, + "content": "- Sequential training maybe suboptimal: The lengthy fine-tuning process can lead the model to \"forget\" pretraining, and instead overfits on the target data, which finally reduces the benefits of pretraining." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.467, + 0.906, + 0.542 + ], + "angle": 0, + "content": "- In contrast, mixed training augments the initial set and adds diversity throughout the training process. The model sees as many images from the target dataset as in the sequential mode, but continues to benefit from the pretraining dataset." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.542, + 0.907, + 0.617 + ], + "angle": 0, + "content": "- A \"balanced blend\" of synthetic and target data appears optimal, providing a sufficient frequency of target images for exploitation while strategically incorporating new synthetic images to enable exploratory learning and notably avoid overfitting." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.406, + 0.907, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.626, + 0.681, + 0.641 + ], + "angle": 0, + "content": "4.3.3. Low Data regime" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We apply a sequential scheme but limit fine-tuning to small subsets of the target data: specifically, \\(1\\%\\), \\(10\\%\\), and \\(30\\%\\) of the training set. Each experiment was repeated 10 times, with a random sampling of the training set. We consider these to be the most meaningful experiments, as this scenario closely matches real-world use cases. The primary objective of pretraining is to develop a model that can either swiftly adapt to new target datasets or perform well when only limited training data is available. One can observe on Fig. 5 a clear advantage for pretrained models at very low level (\\(1\\%\\)). With such little data, the model cannot fully learn the specificities of the target dataset and relies primarily on its general knowledge and on the task itself (detecting changes), rather than on dataset-specific features. Unsurprisingly, models pretrained for this task perform significantly better. However, as the amount of available data increases, the gap tends to narrow. Models increasingly rely" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2210" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.095, + 0.868, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Figure 4. Mixed training (Binary and Semantic). Training on a blend of target and synthetic/hybrid (SyntheWorld or FSC-180k) train sets, containing a ratio of \\( \\mathrm{x\\%} \\) samples from the target (including repetitions). Testing is performed on the target test set. \\( 100\\% \\) corresponds to fine-tuning exclusively on target dataset (without pretraining)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.366 + ], + "angle": 0, + "content": "on data itself, to which they have equal access, yet pretraining on FSC-180k still yields better performance than using SyntheWorld or training from scratch." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.383, + 0.43, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.516, + 0.434, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.725 + ], + "angle": 0, + "content": "Figure 5. Low data regime (Semantic). Fine-tuning on a limited part of the target train set (either \\(1\\%\\), \\(10\\%\\) or \\(30\\%\\) of randomly sampled examples) and evaluating on the whole test set. Models are initially pretrained (SyntheWorld or FSC-180k) or not (Baseline). Metrics are averaged on 10 runs for SECOND and HiUCD-mini." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.747, + 0.284, + 0.76 + ], + "angle": 0, + "content": "4.3.4. Zero-Shot evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "While models pretrained on SyntheWorld struggle to transfer effectively without target sample exposure, pretraining on FSC-180k delivers respectable results, achieving IoU scores of 0.37 on SECOND and 0.36 on HiUCD-mini, showing performance drops of \\(31\\%\\) and \\(40\\%\\), compared to the fully supervised models (details are provided in the Supplementary Material). However, FSC-180k is less effective on LevirCD, where the performance drop reaches \\(60\\%\\) (0.33 vs. 0.83 IoU). We argue that performance in this scenario is directly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.908, + 0.409 + ], + "angle": 0, + "content": "linked to dataset similarities: the limited realism in SyntheWorld hinders transferability, while FSC-180k's diverse landscapes and sparse changes further limit its effectiveness. For the semantic task, pretraining on FSC-180k enhances the model's understanding of change trajectories, achieving SCS of 0.24 on SECOND and 0.25 on HiUCD-mini." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.457, + 0.719, + 0.471 + ], + "angle": 0, + "content": "4.3.5. Qualitative assessment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.909, + 0.627 + ], + "angle": 0, + "content": "Qualitative analysis of the various segmentation results indicates that FSC-180k pretraining significantly enhances the network's ability to localize features in input images. This improvement leads to both more accurate segmentation within VHR images and higher semantic prediction accuracy for main classes of interest in the semantic case (e.g., buildings, roads, bare land for HiUCD-mini, buildings or non-vegetated ground surface for SECOND). Qualitative examples for the different datasets are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We introduced HySCDG, a new, simple bi-temporal change detection generation pipeline. It lies on Stable Diffusion and ControlNet for creating hybrid content of VHR images. HySCDG leverages an existing large-scale semantic land cover map to overcome the lack of diversity and size of real-world datasets. We provide FSC-180k, a pretraining dataset based on FLAIR [18], which includes bi-temporal real/in-painted images, and their associated semantic maps. The proposed solution was evaluated for transferability, on five datasets over four distinct training setups. Our experiments on both binary and semantic cases demonstrated strong transferability, confirming FSC-180k robustness and adaptability as well as HySCDG relevance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2211" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.19 + ], + "angle": 0, + "content": "This work was granted access to the HPC resources of IDRIS under the allocations AD011014690 and AD011014286 made by GENCI. We thank Ruben Gres for inspiring discussions and valuable feedback about ControlNet and Stable Diffusion." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.188, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.238, + 0.486, + 0.279 + ], + "angle": 0, + "content": "[1] Andrew. How to use Soft Inpainting - Stable Diffusion Art. https://stable-diffusion-art.com/soft-inpainting/, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.28, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[2] Wele Gedara Chaminda Bandara and Vishal M. Patel. A transformer-based siamese network for change detection. In IGARSS, pages 207-210, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.32, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[3] Marshall Burke, Anne Driscoll, David B. Lobell, and Stefano Ermon. Using satellite imagery to understand and promote sustainable development. Science, 371(6535), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[4] Yinxia Cao and Xin Huang. A full-level fused cross-task transfer learning method for building change detection using noise-robust pretrained networks on crowdsourced labels. Remote Sensing of Environment, 284:113371, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.485, + 0.467 + ], + "angle": 0, + "content": "[5] Hao Chen, Wenyuan Li, and Zhenwei Shi. Adversarial instance augmentation for building change detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.484, + 0.508 + ], + "angle": 0, + "content": "[6] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10), 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.575 + ], + "angle": 0, + "content": "[7] Hongruixuan Chen, Jian Song, Chengxi Han, Junshi Xia, and Naoto Yokoya. Changemamba: Remote sensing change detection with spatiotemporal state space model. IEEE Transactions on Geoscience and Remote Sensing, 62:1-20, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[8] Isaac Corley, Caleb Robinson, and Anthony Ortiz. A change detection reality check. arXiv preprint arXiv:2402.06994, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.67 + ], + "angle": 0, + "content": "[9] Mikolaj Czerkawski and Christos Tachtatzis. Exploring the capability of text-to-image diffusion models with structural edge guidance for multi-spectral satellite image inpainting. IEEE Geoscience and Remote Sensing Letters, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.671, + 0.485, + 0.751 + ], + "angle": 0, + "content": "[10] Mikolaj Czerkawski, Priti Upadhyay, Christopher Davison, Astrid Werkmeister, Javier Cardona, Robert Atkinson, Craig Michie, Ivan Andonovic, Malcolm Macdonald, and Christos Tachtatzis. Deep internal learning for inpainting of cloud-affected regions in satellite imagery. Remote Sensing, 14(6):1342, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.752, + 0.484, + 0.806 + ], + "angle": 0, + "content": "[11] Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, and Yann Gousseau. Multitask learning for large-scale semantic change detection. Computer Vision and Image Understanding, 187:102783, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.807, + 0.484, + 0.846 + ], + "angle": 0, + "content": "[12] Rodrigo Caye Daudt, Bertrand Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In ICIP, pages 4063-4067, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Lei Ding, Jing Zhang, Haitao Guo, Kai Zhang, Bing Liu, and Lorenzo Bruzzone. Joint spatio-temporal modeling for semantic change detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-14," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.238, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.627, + 0.106 + ], + "angle": 0, + "content": "2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.107, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[14] Miguel Espinosa and Elliot J. Crowley. Generate your own scotland: Satellite image generation conditioned on maps. NeurIPS 2023 Workshop on Diffusion Models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.147, + 0.907, + 0.187 + ], + "angle": 0, + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.187, + 0.907, + 0.227 + ], + "angle": 0, + "content": "[16] Weiliang Fan, Jing M Chen, and Weimin Ju. A pixel missing patch inpainting method for remote sensing image. In International Conference on Geoinformatics, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.228, + 0.907, + 0.308 + ], + "angle": 0, + "content": "[17] Mario Fuentes Reyes, Yuxing Xie, Xiangtian Yuan, Pablo d'Angelo, Franz Kurz, Daniele Cerra, and Jiaojiao Tian. A 2D/3D multimodal data simulation approach with applications on urban semantic segmentation, building extraction and change detection. ISPRS Journal of Photogrammetry and Remote Sensing, 205:74-97, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.907, + 0.39 + ], + "angle": 0, + "content": "[18] Anatol Garioud, Nicolas Gonthier, Loic Landrieu, Apolline De Wit, Marion Valette, Marc Poupee, Sebastien Giordano, and Boris Wattrelos. FLAIR: a country-scale land cover semantic segmentation dataset from multi-source optical imagery. In NeurIPS Datasets and Benchmarks Track, 2023. 2, 3, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[19] Adrien Gressin, Nicole Vincent, Clément Mallet, and Nicolas Paparoditis. Semantic approach in image change detection. In ACIVS, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.906, + 0.471 + ], + "angle": 0, + "content": "[20] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.907, + 0.537 + ], + "angle": 0, + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.538, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[22] Philip J Howarth and Gregory M Wickware. Procedures for change detection using landsat digital data. International Journal of Remote Sensing, 2(3):277-291, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.907, + 0.619 + ], + "angle": 0, + "content": "[23] Wenli Huang, Ye Deng, Siqi Hui, and Jinjun Wang. Image inpainting with bilateral convolution. Remote Sensing, 14(23):6140, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.907, + 0.674 + ], + "angle": 0, + "content": "[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.906, + 0.728 + ], + "angle": 0, + "content": "[25] Shunping Ji, Shiqing Wei, and Meng Lu. Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set. IEEE Transactions on Geoscience and Remote Sensing, 57(1):574-586, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.907, + 0.782 + ], + "angle": 0, + "content": "[26] Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David B Lobell, and Stefano Ermon. Diffusionsat: A generative foundation model for satellite imagery. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.782, + 0.906, + 0.835 + ], + "angle": 0, + "content": "[27] Maria Kolos, Anton Marin, Alexey Artemov, and Evgeny Burnaev. Procedural synthesis of remote sensing images for robust change detection with neural networks. In Advances in Neural Networks, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.836, + 0.906, + 0.875 + ], + "angle": 0, + "content": "[28] Andrey Kuznetsov and Mikhail Gashnikov. Remote sensing image inpainting with generative adversarial networks. In ISDFS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.876, + 0.906, + 0.89 + ], + "angle": 0, + "content": "[29] Ang Li, Qiuhong Ke, Xingjun Ma, Haiqin Weng, Zhiyuan" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2212" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Zong, Feng Xue, and Rui Zhang. Noise doesn't lie: towards universal detection of deep inpainting. In *IJCAI*, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.12, + 0.484, + 0.173 + ], + "angle": 0, + "content": "[30] Zhuohong Li, Wei He, Jiepan Li, Fangxiao Lu, and Hongyan Zhang. Learning without exact guidance: Updating large-scale high-resolution land cover maps from low-resolution historical labels. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.173, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.227, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[32] Zhiyong Lv, Tongfei Liu, Jón Atli Benediktsson, and Nicola Falco. Land cover change detection techniques: Very-high-resolution optical images: A review. IEEE Geoscience and Remote Sensing Magazine, 10(1):44-63, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.282, + 0.484, + 0.322 + ], + "angle": 0, + "content": "[33] Yuchi Ma, Shuo Chen, Stefano Ermon, and David B Lobell. Transfer learning in environmental remote sensing. Remote Sensing of Environment, 301:113924, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[34] Javier Marín and Sergio Escalera. Sssgan: Satellite style and structure generative adversarial networks. Remote Sensing, 13(19):3984, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[35] Fan Meng, Xiaomei Yang, Chenghu Zhou, and Zhi Li. A sparse dictionary learning-based adaptive patch inpainting method for thick clouds removal from high-spatial resolution remote sensing imagery. Sensors, 17(9):2130, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.417, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[36] Ali Mirzaei, Hossein Bagheri, and Iman Khosravi. Enhancing crop classification accuracy through synthetic SAR-Optical data generation using deep learning. ISPRS International Journal of Geo-Information, 12(11), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.471, + 0.484, + 0.525 + ], + "angle": 0, + "content": "[37] Hassiba Nemmour and Youcef Chibani. Multiple support vector machines for land cover change detection: An application for mapping urban extensions. ISPRS Journal of Photogrammetry and Remote Sensing, 61(2):125-133, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.484, + 0.565 + ], + "angle": 0, + "content": "[38] OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org . https : //www.openstreetmap.org, 2017.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.565, + 0.484, + 0.606 + ], + "angle": 0, + "content": "[39] Chao Pang, Jiang Wu, Jian Ding, Can Song, and Gui-Song Xia. Detecting building changes with off-nadir aerial images. Science China Information Sciences, 66(4):140306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[40] Daifeng Peng, Lorenzo Bruzzone, Yongjun Zhang, Haiyan Guan, Haiyong Ding, and Xu Huang. Semicdnet: A semisupervised convolutional neural network for change detection in high resolution remote-sensing images. IEEE Transactions on Geoscience and Remote Sensing, 59(7):5891-5906, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[41] R.S. Priya and K. Vani. Vegetation change detection and recovery assessment based on post-fire satellite imagery using deep learning. Nature Sci Rep, 14:12611, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.484, + 0.781 + ], + "angle": 0, + "content": "[42] Xiaofan Qu, Feng Gao, Junyu Dong, Qian Du, and Heng-Chao Li. Change detection in synthetic aperture radar images using a dual-domain network. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.781, + 0.484, + 0.848 + ], + "angle": 0, + "content": "[43] Gupta Ritwik, Hosfelt Richard, Sajeev Sandra, Patel Nirav, Goodman Bryce, Doshi Jigar, Heim Eric, Choset Howie, and Gaston Matthew. xbd: A dataset for assessing building damage from satellite imagery. arXiv preprint arXiv:1911.09296, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.849, + 0.484, + 0.89 + ], + "angle": 0, + "content": "[44] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[45] Ribana Roscher, Marc Russwurm, Caroline Gevaert, Michael Kampffmeyer, Jefersson A. Dos Santos, Maria Vakalopoulou, Ronny Hänsch, Stine Hansen, Keiller Nogueira, Jonathan Prexl, and Devis Tuia. Better, not just more: Data-centric machine learning for earth observation. IEEE Geoscience and Remote Sensing Magazine, pages 2-22, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.176, + 0.906, + 0.201 + ], + "angle": 0, + "content": "[46] Ragav Sachdeva and Andrew Zisserman. The change you want to see. In WACV, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.201, + 0.906, + 0.256 + ], + "angle": 0, + "content": "[47] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.256, + 0.906, + 0.308 + ], + "angle": 0, + "content": "[48] Minseok Seo, Hakjin Lee, Yongjin Jeon, and Junghoon Seo. Self-pair: Synthesizing changes from single source for object change detection in remote sensing imagery. In WACV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.309, + 0.906, + 0.363 + ], + "angle": 0, + "content": "[49] Minseok Seo, Hakjin Lee, Yongjin Jeon, and Junghoon Seo. Self-pair: Synthesizing changes from single source for object change detection in remote sensing imagery. In WACV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.906, + 0.418 + ], + "angle": 0, + "content": "[50] Li Shen, Yao Lu, Hao Chen, Hao Wei, Donghai Xie, Jiabao Yue, Rui Chen, Shouye Lv, and Bitao Jiang. S2looking: A satellite side-looking dataset for building change detection. Remote Sensing, 13(24), 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.906, + 0.471 + ], + "angle": 0, + "content": "[51] Wenzhong Shi, Min Zhang, Rui Zhang, Shanxiong Chen, and Zhao Zhan. Change detection based on artificial intelligence: State-of-the-art and challenges. Remote Sensing, 12(10):1688, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.906, + 0.513 + ], + "angle": 0, + "content": "[52] Ashbindu Singh. Review article digital change detection techniques using remotely-sensed data. International journal of remote sensing, 10(6):989-1003, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.906, + 0.553 + ], + "angle": 0, + "content": "[53] Jian Song, Hongruixuan Chen, and Naoto Yokoya. Syntheworld: A large-scale synthetic dataset for land cover mapping and building change detection. In WACV, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.605 + ], + "angle": 0, + "content": "[54] Kai Tang and Jin Chen. Changeanywhere: Sample generation for remote sensing change detection via semantic latent diffusion model. arXiv preprint arXiv:2404.08892, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.906, + 0.647 + ], + "angle": 0, + "content": "[55] Yizhuo Tang, Zhengtao Cao, Ningbo Guo, and Mingyong Jiang. A Siamese Swin-Unet for image change detection. Nature Sci Rep, 14:4577, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[56] Shiqi Tian, Yanfei Zhong, Ailong Ma, and Zhuo Zheng. HiUCD: A large-scale dataset for urban semantic change detection in remote sensing imagery. In NeurIPS Workshop on Machine Learning for the Developing World, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.906, + 0.783 + ], + "angle": 0, + "content": "[57] Shiqi Tian, Yanfei Zhong, Zhuo Zheng, Ailong Ma, Xicheng Tan, and Liangpei Zhang. Large-scale deep learning based binary and semantic change detection in ultra high resolution remote sensing imagery: From benchmark datasets to urban application. ISPRS Journal of Photogrammetry and Remote Sensing, 193:164-186, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.783, + 0.906, + 0.836 + ], + "angle": 0, + "content": "[58] Aysim Tucker, Marvin Eisenberger, Daniel Cremers, and Laura Leal-Taixe. Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.836, + 0.906, + 0.892 + ], + "angle": 0, + "content": "[59] Aysim Toker, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andres Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, Giovanni Marchisio, Xiao Xiang Zhu, and Laura Leal-Taixe." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2213" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In CVPR, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.12, + 0.484, + 0.173 + ], + "angle": 0, + "content": "[60] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.174, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[61] Zhipan Wang, Minduan Xu, Zhongwu Wang, Qing Guo, and Qingling Zhang. Scribblednet: Change detection on high-resolution remote sensing imagery with scribble interaction. International Journal of Applied Earth Observation and Geoinformation, 128:103761, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.242, + 0.484, + 0.295 + ], + "angle": 0, + "content": "[62] Junshi Xia, Naoto Yokoya, Bruno Adriano, and Clifford Broni-Bediako. OpenEarthMap: A benchmark dataset for global high-resolution land cover mapping. In WACV, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[63] Kunping Yang, Gui-Song Xia, Zicheng Liu, Bo Du, Wen Yang, Marcello Pelillo, and Liangpei Zhang. Asymmetric siamese networks for semantic change detection in aerial images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-18, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[64] Yang Zhan, Kun Fu, Menglong Yan, Xian Sun, Hongqi Wang, and Xiaosong Qiu. Change detection based on deep siamese convolutional network for optical aerial images. IEEE Geoscience and Remote Sensing Letters, 14(10):1845-1849, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.471, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[66] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[67] Xiaokang Zhang, Weikang Yu, Man-On Pun, and Wenzhong Shi. Cross-domain landslide mapping from large-scale remote sensing images using prototype-guided domain-aware progressive representation learning. ISPRS Journal of Photogrammetry and Remote Sensing, 197:1-17, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[68] Wen-Jie Zheng, Xi-Le Zhao, Yu-Bang Zheng, and Zhi-Feng Pang. Nonlocal patch-based fully connected tensor network decomposition for multispectral image inpainting. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[69] Zhuo Zheng, Stefano Ermon, Dongjun Kim, Liangpei Zhang, and Yanfei Zhong. Changen2: Multi-temporal remote sensing generative change foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.484, + 0.741 + ], + "angle": 0, + "content": "[70] Zhuo Zheng, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.484, + 0.796 + ], + "angle": 0, + "content": "[71] Zhuo Zheng, Shiqi Tian, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Scalable multi-temporal remote sensing change data generation via simulating stochastic change process. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.797, + 0.484, + 0.864 + ], + "angle": 0, + "content": "[72] Zhuo Zheng, Yanfei Zhong, Shiqi Tian, Ailong Ma, and Liangpei Zhang. Changemask: Deep multi-task encoder-transformer-decoder architecture for semantic change detection. ISPRS Journal of Photogrammetry and Remote Sensing, 183:228-239, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.864, + 0.484, + 0.893 + ], + "angle": 0, + "content": "[73] Zhuo Zheng, Yanfei Zhong, Ji Zhao, Ailong Ma, and Liangpei Zhang. Unifying remote sensing change detection via" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "deep probabilistic change models: From principles, models to applications. ISPRS Journal of Photogrammetry and Remote Sensing, 215:239-255, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.134, + 0.907, + 0.173 + ], + "angle": 0, + "content": "[74] Zhe Zhu, Shi Qiu, and Su Ye. Remote sensing of land change: A multifaceted perspective. Remote Sensing of Environment, 282:113266, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2214" + } + ] +] \ No newline at end of file diff --git a/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_origin.pdf b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb3301c6064e623101de0b95ff2cfad55f90af6c --- /dev/null +++ b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/301f1377-a4d4-4bd4-b05d-943883c2d8ee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:254a36174440deb42099ea76d25adaedb50d3350b2169e202998be1c52f754e4 +size 739767 diff --git a/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/full.md b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d25a16a729bb33a391e387af61f1809655b74690 --- /dev/null +++ b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/full.md @@ -0,0 +1,329 @@ +# The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generation + +Yanis Benidir + +Nicolas Gonthier + +Clément Mallet + +Univ Gustave Eiffel, ENSG, IGN, LASTIG, France + +firstname_lastname@ign.fr + +# Abstract + +Bi-temporal change detection at scale based on Very High Resolution (VHR) images is crucial for Earth monitoring. This remains poorly addressed so far: methods either require large volumes of annotated data (semantic case), or are limited to restricted datasets (binary set-ups). Most approaches do not exhibit the versatility required for temporal and spatial adaptation: simplicity in architecture design and pretraining on realistic and comprehensive datasets. Synthetic datasets are the key solution but still fail to handle complex and diverse scenes. In this paper, we present HySCDG a generative pipeline for creating a large hybrid semantic change detection dataset that contains both real VHR images and inpainted ones, along with land cover semantic map at both dates and the change map. Being semantically and spatially guided, HySCDG generates realistic images, leading to a comprehensive and hybrid transfer-proof dataset FSC-180k. We evaluate FSC-180k on five change detection cases (binary and semantic), from zero-shot to mixed and sequential training, and also under low data regime training. Experiments demonstrate that pretraining on our hybrid dataset leads to a significant performance boost, outperforming SyntheWorld, a fully synthetic dataset, in every configuration. All codes, models, and data are available here: https://yb23.github.io/projects/cywd/. + +# 1. Introduction + +Efficient methods for change detection (CD) are crucial for monitoring territories and the various phenomena and activities that impact them [3]. As human and climatic pressures increase, this topic is of utmost importance and calls every year for higher spatial resolution and semantic accuracy [55]. Change detection at scale has benefited greatly from the contributions of deep learning in recent years [7, 12, 69, 73], but most methods are grounded on large volumes of annotated data [41, 60]. Creating a large-scale dataset for bi-temporal (pair-wise) remote sensing change detection poses signifi + +![](images/0b9ea8ffcb9c09a66d1860e45091d22e2b6a0f676515dde8384ccb74ae5f7ef0.jpg) +Target datasets +Figure 1. Efficient and scalable change detection requires a comprehensive training dataset that does not exist today. Using a single-temporal dataset (image+semantic map), we propose HySCDG that generates a novel bi-temporal hybrid dataset FSC-180k. This enables multiple transfer learning possibilities on either binary and semantic change detection tasks. $\mathcal{O} / \mathcal{O}$ : Transfer with/without fine-tuning our model, respectively. + +cant challenges and costs, especially when dealing with Very High Resolution (VHR) images (ground sampling distance: GSD $< 1\mathrm{m}$ , [30]). This is due to the need for expertise and considerable effort in collecting, preprocessing, and annotating images [59, 62], particularly in the case of VHR data and multiple labels (termed semantic change detection, SCD) [57, 61, 70, 74]. + +As an alternative, synthetic data generation is a promising direction for bi-temporal labeled data provision from a single temporal domain that is successful in diverse areas such as urban sprawl [17], disaster or climate change monitoring [42] or crop mapping [36]. Two main paradigms exist for CD: fully synthetic datasets (mainly graphics-based + +[27, 53]), and hybrid solutions that mix real and modified images, often at the instance level [5, 69, 71], known as data augmentation. The first method synthesizes image pairs by rendering two versions of a 3D virtual world thanks to a computer graphics engine. On the other hand, hybrid approaches synthesize new image pairs by pasting or erasing object instances into existing images ( inpainting). They propose a suitable trade-off and faster generation of large quantities of data, with likely higher realism. However, none of these solutions meet the requirements for scalable SCD with VHR imagery (Tab. 1): focus on single tasks, restricted geographical areas, limited diversity, no guarantee on semantic consistency between pairs [48] and subsequently poor transfer to real datasets. To address this challenge, we introduce a generative pipeline built upon Stable Diffusion [44] and ControlNet [65], designed to harness an existing VHR semantic land cover dataset [18] and instance masks, to generate an extensive hybrid semantic bi-temporal change detection dataset encompassing both real and inpainted images, semantic land cover maps at each date, and a change map (Fig. 1). Then, we evaluate the transferability of our synthetic dataset on five datasets for both binary and semantic change detection cases ranging from zero-shot to sequential and mixed training set-ups, as well as under low-data regime. + +Our contributions are as follows: + +- We present a generation pipeline HySCDG for creating VHR bi-temporal SCD datasets based on a land cover real-world dataset by using Stable Diffusion and ControlNet with tunable and semantically-guided inpainting of individual objects. It can be easily configured or finetuned on any land-cover dataset, generating images with the dataset's distinctive traits. +- We provide FSC-180k: a Hybrid Semantic Change Detection pretraining dataset, created from FLAIR [18] along with the instance masks for 300k objects for the corresponding images. +- We provide a comprehensive evaluation of the proposed synthetic dataset on multiple transfer learning scenarios on five change datasets, in particular for SCD. + +# 2. Related work + +Detecting changes in the deep learning era. We target bi-temporal land cover change detection in VHR images [32]. This historical task [22, 37, 52] consists in detecting the meaningful elements of a given area that have changed between two acquisitions and providing the land cover categories at each period (also known as change trajectory). It presents challenges: radiometric variations, geometric variations and non-permanent changes. They have been extensively addressed with supervised deep learning techniques [70]: from CNN [11, 12, 64] to Transformers [2, 7, 13, 55, 72, 73], mainly with Siamese architecture. Multiple datasets, accompanied with ad-hoc methods, have been + +created to feed such models. They are often limited by their size [6, 25, 39, 56, 63], geographical extent [6, 56, 57, 63], low geometrical or annotation quality [11]. Also, most of the effort focuses on binary change detection, i.e., restricted to a unique category, mainly buildings [6, 25, 39, 40, 50] or disasters [67]. Such scarcity in high quality, diverse, large-scale labels is alleviated with transfer learning [4, 51], by fine-tuning models pretrained on large-scale datasets for different image domains, even for semantic change detection [56, 57, 63]. + +# Synthesizing and inpainting remote sensing images. + +Synthetic remote sensing data is useful e.g., for cloud removal [10], image restoration [16], or training deep supervised models [58], before transfer to real examples. Solutions based on patches, [35, 68], autoencoders [23], GAN [28, 34], -aligned generation [24] have been proposed, but with limited image quality and no semantic control. Diffusion models (DM) have improved such quality [44], but most models have applied DM directly on mid-resolution RGB images [14, 26, 58], neglecting both the multi-spectral dimension [9] and VHR spaces. However, a key finding remains: semantically control the synthesis, e.g., by directly synthesizing text [26], semantic maps [58] or with a control module guided by edges [9], semantic maps [14] or metadata [26]. In our work, we fine-tune a Stable Diffusion and a ControlNet models to achieve semantically-controlled inpainting of VHR images as a basis to take advantage of existing images, and generate an hybrid semantic change detection dataset at scale and with high diversity. + +Generating synthetic change datasets. Two main approaches exist: fully synthetic or hybrid solutions. The first paradigm leverages advanced 3D rendering engines to create fully synthetic datasets [17, 27]. This procedural approach allows complete control over rendering parameters (instance placement, semantic classes, lighting conditions, etc.) generated from scratch. E.g., SyntheWorld [53] effectively leverages DM prompted by GPT-4 to create a diverse and large-scale fully synthetic dataset. On the other hand, hybrid approaches generate (semi-)synthetic data by inserting fake changes into real images. + +Such modifications can be random crops [49], instance copy-pasting, or inpainting made by GAN [5, 49, 71] or DM [54, 69]. Blending or style transfer methods are often used to enhance diversity [5, 49]. + +A pivotal challenge is the semantic guidance: modifying few elements of images to mimic real scarce changes, e.g., with instance footprints extracted from external datasets [5] or segmentation model outputs [69]. [54] adopt a Denoising diffusion probabilistic model to generate a new image directly from a modified semantic map and a degraded version of the corresponding snapshot. Changen2 [69] is the closest + +work to ours, with a full pipeline to synthesize a hybrid semantic dataset, again with a DM conditioned by a semantic map. However, Changen2 creates synthetic data specific to a given target dataset in terms of image resolution, size, or change characteristics. In this paper, we explore semantically controlled inpainting based on DM and ControlNet to the intertwined challenge of diverse and multi-class modification of VHR images. It is termed HySCDG and leads to the hybrid dataset named FSC-180k. Tab. 1 provides a comparison with main synthetic change detection datasets. + +Table 1. Synthetic Change Datasets. Land cover change detection in remote sensing imagery. For the total number of pixels, we only consider strictly different images. The geographical zone's extension can be local [ ] national [ ] or global [ ] We denote open-access (OA) and directly downloadable datasets with ✓. In the case of IAug, we sum the number of pixels of the two datasets provided in this work. OEM means OpenEarthMap. + +
DatasetOAPixels ×106GSD (m)ClassesSourceZoneType
SynCW [27]X370.61X田Synthetic
SMARS [17]✓1100.3-0.52X田Synthetic
IAug [5]X1,1670.075-0.51LEVIR-CD [6], WHU-CD [25]田Hybrid
Ce-100K [54]X6,5530.25-0.58OEM [62]田Synthetic
Changen2 [69]X7,0770.25-0.58OEM [62]田Hybrid
Changen [71]X11,7960.81xView2 [43]田Hybrid
SyntheWorld [53]✓18,3500.3-11XXSynthetic
FSC-180k (Ours)✓80,7400.216FLAIR [18]田Hybrid
+ +Transfer learning from synthetic/hybrid datasets. Transfer learning is used to assess the validity of the generated image pairs, and vary with the operational set-up [33]. SCD at scale is performed through sequential learning: pretraining on such datasets and fine-tuning on real-world cases [5, 27, 69, 71]. Fewer works [53] evaluate on mixed training sets (single training mixing both real and synthetic samples), which can help to prevent overfitting. An orthogonal solution considers low data regimes, i.e., fine-tuning on a very small train set [5, 53, 54], or even under zero-shot scenarios [54]. We evaluate our proposed dataset in the four conceivable transfer scenarios. + +# 3. Hybrid generation of semantic changes + +Much of the success of deep learning methods for Earth observation relies on large-scale training datasets with reliable annotations [18, 45]. However, no large-scale publicly available dataset exists for bi-temporal SCD: we propose HySCDG for Hybrid Semantic Change Detection Generation, a procedure to leverage existing large-scale land cover datasets and state-of-the-art image inpainting methods to simulate visually realistic bi-temporal semantic changes. + +This leads to the generation of the FSC-180k dataset. + +Our paper is grounded on two main ideas: (i) Adapting and fine-tuning a Stable Diffusion Model enables efficient inpainting of VHR images with a semantic control and geographical instance selection; (ii) Changes can be simulated with sufficient diversity by randomly selecting instances (and modifying their labels) within an existing land cover map that is paired with the image used as input. The pipeline is shown in Fig. 2. Details are provided in the following. + +# 3.1. Adaptation of Stable Diffusion for Inpainting. + +We use Stable Diffusion (SD) [44], an open and state-of-the-art text-to-image synthesis method. Our objective is to modify SD to manage controllable inpainting specifically in remote sensing aerial images, for various landscapes or semantic classes. Though SD models are effective in generating natural images with clear focal points and layered depth, with artistic or cinematic styles, they require adaptation to be used for remote sensing images [26]: acquisition distance, intricate textures, no domination of a single element. Starting from a Stable Diffusion 2 Inpainting checkpoint1, we sequentially fine-tune the Autoencoder and the Diffusion U-Net, before adding and training a ControlNet [65] that leverages information extracted from an existing large-scale land cover dataset (namely, FLAIR [18], see Sec. 3.3). Our pipeline allows to generate 5-band images that incorporate infrared and elevation along with RGB channels, for the first time in satellite image synthesis. + +Adaptation of the Variational Autoencoder. Unlike most SD extensions, which focus on fine-tuning the central U-Net, we also adapted the Variational Autoencoder (VAE) that encodes images into the diffusion latent space. This adaptation targets efficient encoding of multi-band VHR images from the FLAIR dataset. The original VAE, trained with a KL-regularization factor [15], uses a combination of L1 and perceptual loss [66] alongside KL-divergence. To improve compression of remote sensing images, we added both L2 and Focal losses [31]: minimizing high-frequency errors reduces blurriness. Additionally, we included a "color loss" (L2 loss on each $5 \times 5$ patch) to limit oversaturation, which affects ground object color realism. Fine-tuning of the VAE took 160 hours A100 GPU. + +Stable Diffusion training. We then fine-tuned the Diffusion U-Net for inpainting, adapting it to the new VAE and teaching it to synthesize and inpaint VHR images. Special attention was given to prompt selection. Although inpainting is the model's primary function, we included direct image generation from prompts in $10\%$ to $20\%$ of cases, as in standard SD. This training required 300 hours on a A100 GPU (30,000 iterations, batch size of 32). + +![](images/8dda307eabf77920221443de3276d86a520b0d5380aa777621a7bc1573adb632.jpg) +Figure 2. HySCDG pipeline. From a single-temporal dataset composed of one VHR image $I_{1}$ , a semantic map $M_{1}$ , and some openly available labeled instances, we generate a new VHR image $I_{2}$ , a new map $M_{2}$ and subsequently a change map $C$ . This results in the FSC-180k hybrid dataset. The two pivotal novelties consist in: (i) Adapting and fine-tuning a Stable-Diffusion Model from image inpainting and (ii) exploiting open geospatial data for inpainting prompt control and semantically guiding the objects to be modified. The combination of both solutions ensures a diverse at-scale VHR multi-class change detection dataset. + +Prompt Engineering. Prompts are designed using both each image's geographic coordinates and open land cover data (OpenStreetMap [38]). To avoid learning irrelevant details, prompts are composed of three elements: (i) Spatial (place name, city, region), (ii) Temporal (time of day, season), (iii) Semantic (frequent classes in the inpainted area, compared to the overall dataset). + +Prompt example: "Grass and agricultural vegetation next to a highway, locality of Savigny-en-Revermont, Bourgogne-Franche-Comté, in the morning, during Summer." + +# 3.2. Conditional change inpainting with ControlNet + +Driven by our goal to control the new classes in the inpainted zones, we add a ControlNet (CN) architecture [65] to build a model that generates realistic VHR images from a semantic segmentation map (here land cover). CN is designed to add task-specific conditioning to a pretrained image DM. It allows injecting explicit semantic guidance during the diffusion process, conditioning the outputs on some reference image (a color map of the semantic class), in addition to a text prompt. Our training method enables CN to perform both inpainting and text-to-image synthesis, using the semantic condition. Training took 240 hours on a A100 GPU. + +# 3.3. Select, Mask, Change, Inpaint : the HySCDG pipeline + +Inspired by [46], the core of our method lies in our "Select and Inpaint" mechanism. Given a VHR image along with + +its semantic segmentation and instances $^2$ , we inpaint these instances using the SD+CN inpainting method: the label of the instance is changed, the texture of the image is substituted accordingly. This results in a pair of VHR images, both with semantic maps and the corresponding semantic change maps. The complete HySCDG is provided in Fig. 2 and detailed below. It is illustrated for one image but is applied several times to each image of the single temporal dataset. + +HySCDG uncovered. For an image $I_{1}$ with its semantic map $M_{1}$ , and $\{O_i\}$ the set of $N_I$ instances lying on the extent of $I_{1}$ . + +- Randomly select $N_{change}$ instances among $\{O_i\}$ . +- For each instance, generate an inpainting mask: footprint+spatial buffer. +- Using $M_1$ , take as “ $T_1$ class” the most frequent class inside the footprint: $c_1$ . Then, select the larger convex hull of pixels of this class inside the buffered zone. These pixels correspond to the change mask: $C$ . +- Randomly select another class $c_2$ at $T_2$ and replace the previous label by this new label. +- Randomly select $N_{\text{nochange}}$ random shapes in the image: such masks appear only on the inpainting mask, not on the change mask, letting their labels unchanged. +- Apply $\mathrm{SD + CN}$ to masked image and new semantic map. +- Cumulate the previous steps: we obtain our inpainting + +mask, change mask, and for $T_{2}$ a new semantic map $M_2$ and the corresponding new image $I_{2}$ . + +The pair of images $(I_1, I_2)$ with their corresponding semantic maps $(M_1, M_2)$ , and the change map $C$ (difference between $M_2$ and $M_1$ ), are the samples of the hybrid dataset. + +Label and instance extraction. For comprehensive multiclass learning, we select FLAIR dataset [18], a large-scale semantic segmentation map for 16 land cover classes over more than $800\mathrm{km}^2$ . FLAIR offers the advantage of covering areas where geographical instances of many FLAIR labels are publicly available3, alleviating requiring a panoptic segmentation. Around 300,000 instance masks have been extracted for this work. + +Inpainting signature mitigation. Inpainting produces realistic changes but new areas tend to have sometimes a high-frequency signature, as reported by other works [20, 29]. In order to discourage the model from simply learning this inpainting noise instead of learning the actual changes between images, we also inpaint random areas of the image without changing the semantic class in a similar way to the "not changed regions" from [46]. Besides, to obtain the best transition between the original image and the inpainted area, we use a continuous mask in the inference [1]. + +Selection of the new label. For each change area, we randomly select the new class $c_2$ among all classes weighted by their frequency in the whole dataset divided by their frequencies in the change area. + +Adoption of a buffer zone. For inpainting, we add a buffer around the footprint of each modified instance. This gives SD more freedom to generate smoother borders around the changed areas. It also helps handling small spatial discrepancies between footprints and the VHR image content [19]. + +# 3.4. Structure of the FSC-180k dataset + +For each image in the FLAIR training set, we generate three unique modified images, resulting in 180,000 synthetic images, each with a corresponding $512 \times 512$ semantic map covering 16 distinct semantic classes. We call this dataset FSC-180k, short for FLAIR Synthetic Change. FSC-180k includes approximately 80 billion pixels with a ground resolution of $0.2\mathrm{m}$ per pixel. Note that we can expand the dataset by pairing synthetic images that were generated (differently) from a single real image, process that we use at training time, effectively doubling the number of image pairs. Tab. 1 compares FSC-180k with other synthetic change detection datasets and shows that FSC-180k offers a much broader range of semantic changes beyond simple building modifications, is of higher spatial resolution, and is nearly ten times larger than comparable datasets. It should also be noted that the proportion of changes in FSC-180k is closer to the real + +scenarios observed in Europe (prevalence of $5\%$ change) than other academic datasets (see Tab. 2), excepting HiUCD. + +Quality assessment of FSC-180k. ControlNet is the most error-prone module: it may sometimes fail to respect the semantic map. We evaluated how unchanged the semantic map remains by performing 15-class semantic segmentation on real and generated images, thanks to a UNet trained for this task, leading to a satisfactory error rate below $20\%$ between both outputs. We also leveraged standard metrics for image generation evaluation to assess the realism of the synthesis. We got an Inception Score [47] of 6.2 on the inpainted images, which is comparable to real images, and a low FID score [21] of 0.43 between inpainted and real images. + +# 4. Experiments + +We conduct various transfer learning experiments to validate hybrid data pretraining, comparing FSC-180k with the fully artificial SyntheWorld [53], the only available synthetic dataset with a size comparable to ours (see Tab. 1). + +# 4.1. Pretraining approaches with FSC-180k + +We assess FSC-180k's effectiveness by exploiting existing open datasets as targets under various settings (see Tab. 2 and Supplementary Material). Here, we first present the different transfer learning approaches adopted, and then, dive into the deep models and datasets considered (Sec. 4.2). Our objective is to demonstrate how pretraining on a hybrid dataset improves performance, for both binary and semantic change detection. + +The four scenarios are: + +- Sequential Transfer Learning. Models are pretrained on FSC-180k, and then fine-tuned on the target data. +- Low Data Regime. The same as above but with a fraction of the available training set. +- Mixed Training. Models are trained on a blend of FSC-180k and the target data. +- Zero-Shot Learning. A model pretrained on FSC-180k is directly applied on the target dataset, with only a remapping of FSC-180k classes to align labels. + +# 4.2. Experimental protocol + +# 4.2.1. Target datasets and metrics + +Our transfer experiments span five diverse datasets, among which three contain semantic labels (Tab. 2). This ensures variety in landscapes, sizes, and change types, i.e., a comprehensive evaluation. Among the plethora of (S)CD datasets, we selected the most popular ones: Levir-CD [6], S2Looking [50], SECOND [63] and HiUCD-mini [56]. + +We used classical metrics from the change detection literature: Intersection over Union (IoU) and F1 score for binary + +cases and mainly the Semantic Change Segmentation score (SCS) from [59] for semantic evaluation. In this latter case, the Separated Kappa (SeK) from [63] is also informative though it tends to be more biased towards detection task than semantic prediction. + +We added some experiments related to HiUCD-XL [57], a follow-up dataset of HiUCD-mini [56] using semantic mIoU and change mIoU, metrics provided on the evaluation platform4 as test labels are kept private. + +Table 2. Target datasets. Overview of the real-world datasets used for our transfer learning experiments. + +
DatasetChange (%)Pixels ×106GSD (m)ClassesPeriodZone
Levir-CD [6]28.41,0670.512002/18Texas
S2Looking [50]25.68,3890.3-0.812017/20World
SECOND [63]20.01,2220.3-0.66?China
HiUCD-mini [56]6.65670.192017/19Estonia
HiUCD-XL [57]1.210,6950.192017/19Estonia
FSC-180k (Ours)7.080,7400.2162018/21France
+ +# 4.2.2. Model architecture + +We consider that designing a new architecture is out of the scope of this paper and that numerous solutions have already been proposed (see Sec. 2). Therefore, we initially used a simple U-Net model [8] with a segmentation head for semantic change detection. Although effective for binary change, it lacked expressiveness for semantic tasks. Similarly to [11], we developed a simple Dual UNet model composed of 2 nearly identical parallel UNets for (i) change detection, and (ii) semantic segmentation, with ResNet-50 backbones pretrained on ImageNet and trained in a multi-task setting. The change branch incorporates features from the semantic encoder. We keep the training strategy simple and constant to focus on data and transfer, this enhances interpretability. + +Can a binary model be sufficient? We tested both a pure binary CD model and the Dual UNet, for which, when, switching to binary-only tasks, the semantic decoder is inactive but the encoding part remains active (as it provides features to the change detection branch). Such selective change detection ability allows training on a multi-class dataset and focusing on specific binary tasks. Experiments show that even for binary tasks, the Dual UNet outperforms the binary-only model, e.g., achieving an IoU of 0.83 on Levir-CD (compared to 0.80). Learning more information than just change pixels, the Dual UNet gains a deeper scene understanding, leading to more reliable predictions. + +Existing architectures. We tested more complex transformer-based architectures, namely ChangeFormer [2], + +MambaCD [7], and SCanNet [13], both on HiUCD-mini and SECOND datasets. SCanNet performed best, balancing high performance and straightforward implementation due to its lightweight design (see Supplementary Material). SCanNet is kept hereafter for comparisons (see Tab. G). + +Transfer for varying semantic classes. In the semantic case, we faced the issue of distinct classes between source and target datasets. This required to choose which classes should be kept for model training. + +- Sequential case: for both pretraining and fine-tuning strategies, we use the classes of the processed dataset. The semantic head of the model is reinitialized and adapted to the new number of classes before fine-tuning. +- Mixed case: Classes from the target dataset are used, manually mapping pretraining classes to their closest counterparts. + +# 4.2.3. Experimental setting + +We apply basic augmentations on the images: horizontal and vertical flips, $90^{\circ}$ rotations, and small rescaled crops (scales above 0.8). Normalization by mean and variance from the target dataset is used in all experiments. For optimization, we use the AdamW algorithm with default parameters (0.01 weight decay), and an initial learning rate of 3e-4 with exponential decay (0.99 per epoch). For sequential training, models are pretrained for 40h on a A100 GPU, with finetuning requiring up to 20h on a V100. For mixed training, we limit the total training time to 20h on a A100. + +# 4.3. Results and Analysis + +# 4.3.1. Sequential training + +Binary Case. Models pretrained on synthetic datasets outperform those without pretraining on all target datasets (Fig. 3). SyntheWorld maintains commendable performance on the single-class "building" change case, similarly to Levir-CD. Pretraining on FSC-180k yields the strongest results, compared to SyntheWorld on all five target datasets, with gains of $2 - 3\%$ of IoU. Details are provided in the Supplementary Material. + +Semantic Case. Pretraining on FSC-180k outperforms the baseline and SyntheWorld on the three SCD datasets in the sequential mode (Tab. G). Indeed, pretraining on SyntheWorld, that focuses solely on the "building" class, shows limitations in multiclass tasks on HiUCD-mini and HiUCD-XL compared to the baseline. Our FSC-180k dataset is particularly beneficial for HiUCD-mini and HiUCD-XL. The types of landscapes and spatial resolution are relatively similar, which facilitates effective transfer learning. For HiUCD-mini, our pretraining results in a SeK of 0.19 and a SCS of 0.78, representing respective increases of $11\%$ and $7\%$ + +![](images/02a086d63be6387b772be6ec1d6d29fb2de78353164f380dd705ef0d9d775761.jpg) +(a) Levir-CD + +![](images/2da65fbc9a55ae153200b48257ae5eb2d7c9990447562f39913f51e1aeadd91b.jpg) +(b) S2Looking + +![](images/bad28aa6f7848be81b4f5270a398e1ecfd421fc506d75d11b1fa83488862f93d.jpg) +(c) SECOND +(d) HiUCD-mini +Figure 3. Sequential training in binary change detection. After pretraining our Dual U-Net on either nothing, SyntheWorld or FSC-180k, we finetune and test it on each of the 4 target datasets and represent the binary IoU. + +compared to the baseline. The gains rise to $28\%$ and $8\%$ against SyntheWorld. For SECOND, SyntheWorld partially narrows the gap: SeK of 0.17 and a SCS of 0.84, while the FSC-180k's pretrained Dual UNet achieves 0.18 and 0.87. + +Table 3. Sequential training in the semantic change detection framework. We report the binary change detection scores (IoU or Chg. mIoU) as well as semantic scores (SCS or mIoU) on three datasets in the sequential scheme. Models were initially pretrained (SyntheWorld or FSC-180k) or not (Baseline). + +
PretrainingModelSECONDHiUCD-mini
IoUSCSIoUSCS
NoneDual UNet0.530.830.610.73
SyntheWorldDual UNet0.540.840.580.73
FSC-180kDual UNet0.550.890.630.78
NoneSCanNet0.540.840.600.66
FSC-180kSCanNet0.560.890.630.71
+ +
PretrainingHiUCD-XL
Chg. mIoUmIoU
None0.170.58
SyntheWorld0.170.58
FSC-180k0.190.60
+ +Dependence on a CD model. We compared two CD model by running the same experiments on SECOND and HiUCD-mini with a SCanNet model (Tab. G). We were un + +able to reproduce the results reported in the original paper [13], despite following the described learning scheme. SCanNet still performs well on the SECOND dataset, achieving a SCS of 0.84, compared to 0.83 for the Dual UNet. The opposite trend is observed on HiUCD, where the Dual UNet reaches a SCS of 0.73, while SCanNet reaches 0.66. As neither of the two models stands out, we mainly present the results obtained with our Dual UNet. + +# 4.3.2. Mixed training + +Our experiments show, that the optimal ratio between synthetic and target data lies around $50\%$ for HiUCD-XL, S2Looking or SECOND (Fig. 4) conversely to the conclusions in [53], where performance always increases with the amount of target data included in the training mix. Nevertheless, the optimal ratio is $90\%$ for smaller dataset such as Levir-CD or HiUCD-mini, that may be too specific. However, the mixed training strategy proves to be really effective, outperforming the sequential one on HiUCD-XL with a balanced blend as training base, as noticed in Fig. 4, compared to Tab. G. We can deduce that: + +- Sequential training maybe suboptimal: The lengthy fine-tuning process can lead the model to "forget" pretraining, and instead overfits on the target data, which finally reduces the benefits of pretraining. +- In contrast, mixed training augments the initial set and adds diversity throughout the training process. The model sees as many images from the target dataset as in the sequential mode, but continues to benefit from the pretraining dataset. +- A "balanced blend" of synthetic and target data appears optimal, providing a sufficient frequency of target images for exploitation while strategically incorporating new synthetic images to enable exploratory learning and notably avoid overfitting. + +# 4.3.3. Low Data regime + +We apply a sequential scheme but limit fine-tuning to small subsets of the target data: specifically, $1\%$ , $10\%$ , and $30\%$ of the training set. Each experiment was repeated 10 times, with a random sampling of the training set. We consider these to be the most meaningful experiments, as this scenario closely matches real-world use cases. The primary objective of pretraining is to develop a model that can either swiftly adapt to new target datasets or perform well when only limited training data is available. One can observe on Fig. 5 a clear advantage for pretrained models at very low level ( $1\%$ ). With such little data, the model cannot fully learn the specificities of the target dataset and relies primarily on its general knowledge and on the task itself (detecting changes), rather than on dataset-specific features. Unsurprisingly, models pretrained for this task perform significantly better. However, as the amount of available data increases, the gap tends to narrow. Models increasingly rely + +![](images/9e8750ab7c06af2f29d18d359bbac0b995fa5a07ad979ad077049a71f206588d.jpg) +Figure 4. Mixed training (Binary and Semantic). Training on a blend of target and synthetic/hybrid (SyntheWorld or FSC-180k) train sets, containing a ratio of $\mathrm{x\%}$ samples from the target (including repetitions). Testing is performed on the target test set. $100\%$ corresponds to fine-tuning exclusively on target dataset (without pretraining). + +on data itself, to which they have equal access, yet pretraining on FSC-180k still yields better performance than using SyntheWorld or training from scratch. + +![](images/2a9e960ba543022cfa9a7c3f9e06bdee68639d6049ee2b10dc252c47e630835a.jpg) + +![](images/7501483e6676c6556fbadecb195d60712632400487a232ad0e50b19cbe3475d6.jpg) +Figure 5. Low data regime (Semantic). Fine-tuning on a limited part of the target train set (either $1\%$ , $10\%$ or $30\%$ of randomly sampled examples) and evaluating on the whole test set. Models are initially pretrained (SyntheWorld or FSC-180k) or not (Baseline). Metrics are averaged on 10 runs for SECOND and HiUCD-mini. + +# 4.3.4. Zero-Shot evaluation + +While models pretrained on SyntheWorld struggle to transfer effectively without target sample exposure, pretraining on FSC-180k delivers respectable results, achieving IoU scores of 0.37 on SECOND and 0.36 on HiUCD-mini, showing performance drops of $31\%$ and $40\%$ , compared to the fully supervised models (details are provided in the Supplementary Material). However, FSC-180k is less effective on LevirCD, where the performance drop reaches $60\%$ (0.33 vs. 0.83 IoU). We argue that performance in this scenario is directly + +linked to dataset similarities: the limited realism in SyntheWorld hinders transferability, while FSC-180k's diverse landscapes and sparse changes further limit its effectiveness. For the semantic task, pretraining on FSC-180k enhances the model's understanding of change trajectories, achieving SCS of 0.24 on SECOND and 0.25 on HiUCD-mini. + +# 4.3.5. Qualitative assessment + +Qualitative analysis of the various segmentation results indicates that FSC-180k pretraining significantly enhances the network's ability to localize features in input images. This improvement leads to both more accurate segmentation within VHR images and higher semantic prediction accuracy for main classes of interest in the semantic case (e.g., buildings, roads, bare land for HiUCD-mini, buildings or non-vegetated ground surface for SECOND). Qualitative examples for the different datasets are provided in the Supplementary Material. + +# 5. Conclusion + +We introduced HySCDG, a new, simple bi-temporal change detection generation pipeline. It lies on Stable Diffusion and ControlNet for creating hybrid content of VHR images. HySCDG leverages an existing large-scale semantic land cover map to overcome the lack of diversity and size of real-world datasets. We provide FSC-180k, a pretraining dataset based on FLAIR [18], which includes bi-temporal real/in-painted images, and their associated semantic maps. The proposed solution was evaluated for transferability, on five datasets over four distinct training setups. Our experiments on both binary and semantic cases demonstrated strong transferability, confirming FSC-180k robustness and adaptability as well as HySCDG relevance. + +# Acknowledgements + +This work was granted access to the HPC resources of IDRIS under the allocations AD011014690 and AD011014286 made by GENCI. We thank Ruben Gres for inspiring discussions and valuable feedback about ControlNet and Stable Diffusion. + +# References + +[1] Andrew. How to use Soft Inpainting - Stable Diffusion Art. https://stable-diffusion-art.com/soft-inpainting/, 2024.5 +[2] Wele Gedara Chaminda Bandara and Vishal M. Patel. A transformer-based siamese network for change detection. In IGARSS, pages 207-210, 2022. 2, 6 +[3] Marshall Burke, Anne Driscoll, David B. Lobell, and Stefano Ermon. Using satellite imagery to understand and promote sustainable development. Science, 371(6535), 2021. 1 +[4] Yinxia Cao and Xin Huang. A full-level fused cross-task transfer learning method for building change detection using noise-robust pretrained networks on crowdsourced labels. Remote Sensing of Environment, 284:113371, 2023. 2 +[5] Hao Chen, Wenyuan Li, and Zhenwei Shi. Adversarial instance augmentation for building change detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 2, 3 +[6] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10), 2020. 2, 3, 5, 6 +[7] Hongruixuan Chen, Jian Song, Chengxi Han, Junshi Xia, and Naoto Yokoya. Changemamba: Remote sensing change detection with spatiotemporal state space model. IEEE Transactions on Geoscience and Remote Sensing, 62:1-20, 2024. 1, 2, 6 +[8] Isaac Corley, Caleb Robinson, and Anthony Ortiz. A change detection reality check. arXiv preprint arXiv:2402.06994, 2024. 6 +[9] Mikolaj Czerkawski and Christos Tachtatzis. Exploring the capability of text-to-image diffusion models with structural edge guidance for multi-spectral satellite image inpainting. IEEE Geoscience and Remote Sensing Letters, 2024. 2 +[10] Mikolaj Czerkawski, Priti Upadhyay, Christopher Davison, Astrid Werkmeister, Javier Cardona, Robert Atkinson, Craig Michie, Ivan Andonovic, Malcolm Macdonald, and Christos Tachtatzis. Deep internal learning for inpainting of cloud-affected regions in satellite imagery. Remote Sensing, 14(6):1342, 2022. 2 +[11] Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, and Yann Gousseau. Multitask learning for large-scale semantic change detection. Computer Vision and Image Understanding, 187:102783, 2019. 2, 6 +[12] Rodrigo Caye Daudt, Bertrand Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In ICIP, pages 4063-4067, 2018. 1, 2 +[13] Lei Ding, Jing Zhang, Haitao Guo, Kai Zhang, Bing Liu, and Lorenzo Bruzzone. Joint spatio-temporal modeling for semantic change detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-14, + +2024. 2, 6, 7 +[14] Miguel Espinosa and Elliot J. Crowley. Generate your own scotland: Satellite image generation conditioned on maps. NeurIPS 2023 Workshop on Diffusion Models, 2023. 2 +[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 3 +[16] Weiliang Fan, Jing M Chen, and Weimin Ju. A pixel missing patch inpainting method for remote sensing image. In International Conference on Geoinformatics, 2011. 2 +[17] Mario Fuentes Reyes, Yuxing Xie, Xiangtian Yuan, Pablo d'Angelo, Franz Kurz, Daniele Cerra, and Jiaojiao Tian. A 2D/3D multimodal data simulation approach with applications on urban semantic segmentation, building extraction and change detection. ISPRS Journal of Photogrammetry and Remote Sensing, 205:74-97, 2023. 1, 2, 3 +[18] Anatol Garioud, Nicolas Gonthier, Loic Landrieu, Apolline De Wit, Marion Valette, Marc Poupee, Sebastien Giordano, and Boris Wattrelos. FLAIR: a country-scale land cover semantic segmentation dataset from multi-source optical imagery. In NeurIPS Datasets and Benchmarks Track, 2023. 2, 3, 5, 8, 1 +[19] Adrien Gressin, Nicole Vincent, Clément Mallet, and Nicolas Paparoditis. Semantic approach in image change detection. In ACIVS, 2013. 5 +[20] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In CVPR, 2023. 5 +[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[22] Philip J Howarth and Gregory M Wickware. Procedures for change detection using landsat digital data. International Journal of Remote Sensing, 2(3):277-291, 1981. 2 +[23] Wenli Huang, Ye Deng, Siqi Hui, and Jinjun Wang. Image inpainting with bilateral convolution. Remote Sensing, 14(23):6140, 2022. 2 +[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1125-1134, 2017. 2 +[25] Shunping Ji, Shiqing Wei, and Meng Lu. Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set. IEEE Transactions on Geoscience and Remote Sensing, 57(1):574-586, 2018. 2, 3 +[26] Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David B Lobell, and Stefano Ermon. Diffusionsat: A generative foundation model for satellite imagery. In ICLR, 2023. 2, 3 +[27] Maria Kolos, Anton Marin, Alexey Artemov, and Evgeny Burnaev. Procedural synthesis of remote sensing images for robust change detection with neural networks. In Advances in Neural Networks, 2019. 2, 3 +[28] Andrey Kuznetsov and Mikhail Gashnikov. Remote sensing image inpainting with generative adversarial networks. In ISDFS, 2020. 2 +[29] Ang Li, Qiuhong Ke, Xingjun Ma, Haiqin Weng, Zhiyuan + +Zong, Feng Xue, and Rui Zhang. Noise doesn't lie: towards universal detection of deep inpainting. In *IJCAI*, 2021. 5 +[30] Zhuohong Li, Wei He, Jiepan Li, Fangxiao Lu, and Hongyan Zhang. Learning without exact guidance: Updating large-scale high-resolution land cover maps from low-resolution historical labels. In CVPR, 2024. 1 +[31] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 3 +[32] Zhiyong Lv, Tongfei Liu, Jón Atli Benediktsson, and Nicola Falco. Land cover change detection techniques: Very-high-resolution optical images: A review. IEEE Geoscience and Remote Sensing Magazine, 10(1):44-63, 2021. 2 +[33] Yuchi Ma, Shuo Chen, Stefano Ermon, and David B Lobell. Transfer learning in environmental remote sensing. Remote Sensing of Environment, 301:113924, 2024. 3 +[34] Javier Marín and Sergio Escalera. Sssgan: Satellite style and structure generative adversarial networks. Remote Sensing, 13(19):3984, 2021. 2 +[35] Fan Meng, Xiaomei Yang, Chenghu Zhou, and Zhi Li. A sparse dictionary learning-based adaptive patch inpainting method for thick clouds removal from high-spatial resolution remote sensing imagery. Sensors, 17(9):2130, 2017. 2 +[36] Ali Mirzaei, Hossein Bagheri, and Iman Khosravi. Enhancing crop classification accuracy through synthetic SAR-Optical data generation using deep learning. ISPRS International Journal of Geo-Information, 12(11), 2023. 1 +[37] Hassiba Nemmour and Youcef Chibani. Multiple support vector machines for land cover change detection: An application for mapping urban extensions. ISPRS Journal of Photogrammetry and Remote Sensing, 61(2):125-133, 2006. 2 +[38] OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org . https : //www.openstreetmap.org, 2017.4 +[39] Chao Pang, Jiang Wu, Jian Ding, Can Song, and Gui-Song Xia. Detecting building changes with off-nadir aerial images. Science China Information Sciences, 66(4):140306, 2023. 2 +[40] Daifeng Peng, Lorenzo Bruzzone, Yongjun Zhang, Haiyan Guan, Haiyong Ding, and Xu Huang. Semicdnet: A semisupervised convolutional neural network for change detection in high resolution remote-sensing images. IEEE Transactions on Geoscience and Remote Sensing, 59(7):5891-5906, 2020. 2 +[41] R.S. Priya and K. Vani. Vegetation change detection and recovery assessment based on post-fire satellite imagery using deep learning. Nature Sci Rep, 14:12611, 2024. 1 +[42] Xiaofan Qu, Feng Gao, Junyu Dong, Qian Du, and Heng-Chao Li. Change detection in synthetic aperture radar images using a dual-domain network. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 1 +[43] Gupta Ritwik, Hosfelt Richard, Sajeev Sandra, Patel Nirav, Goodman Bryce, Doshi Jigar, Heim Eric, Choset Howie, and Gaston Matthew. xbd: A dataset for assessing building damage from satellite imagery. arXiv preprint arXiv:1911.09296, 2019. 3 +[44] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 2, 3 + +[45] Ribana Roscher, Marc Russwurm, Caroline Gevaert, Michael Kampffmeyer, Jefersson A. Dos Santos, Maria Vakalopoulou, Ronny Hänsch, Stine Hansen, Keiller Nogueira, Jonathan Prexl, and Devis Tuia. Better, not just more: Data-centric machine learning for earth observation. IEEE Geoscience and Remote Sensing Magazine, pages 2-22, 2024. 3 +[46] Ragav Sachdeva and Andrew Zisserman. The change you want to see. In WACV, 2023. 4, 5 +[47] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5 +[48] Minseok Seo, Hakjin Lee, Yongjin Jeon, and Junghoon Seo. Self-pair: Synthesizing changes from single source for object change detection in remote sensing imagery. In WACV, 2023. 2 +[49] Minseok Seo, Hakjin Lee, Yongjin Jeon, and Junghoon Seo. Self-pair: Synthesizing changes from single source for object change detection in remote sensing imagery. In WACV, 2023. 2 +[50] Li Shen, Yao Lu, Hao Chen, Hao Wei, Donghai Xie, Jiabao Yue, Rui Chen, Shouye Lv, and Bitao Jiang. S2looking: A satellite side-looking dataset for building change detection. Remote Sensing, 13(24), 2021. 2, 5, 6 +[51] Wenzhong Shi, Min Zhang, Rui Zhang, Shanxiong Chen, and Zhao Zhan. Change detection based on artificial intelligence: State-of-the-art and challenges. Remote Sensing, 12(10):1688, 2020. 2 +[52] Ashbindu Singh. Review article digital change detection techniques using remotely-sensed data. International journal of remote sensing, 10(6):989-1003, 1989. 2 +[53] Jian Song, Hongruixuan Chen, and Naoto Yokoya. Syntheworld: A large-scale synthetic dataset for land cover mapping and building change detection. In WACV, 2024. 2, 3, 5, 7 +[54] Kai Tang and Jin Chen. Changeanywhere: Sample generation for remote sensing change detection via semantic latent diffusion model. arXiv preprint arXiv:2404.08892, 2024. 2, 3 +[55] Yizhuo Tang, Zhengtao Cao, Ningbo Guo, and Mingyong Jiang. A Siamese Swin-Unet for image change detection. Nature Sci Rep, 14:4577, 2024. 1, 2 +[56] Shiqi Tian, Yanfei Zhong, Ailong Ma, and Zhuo Zheng. HiUCD: A large-scale dataset for urban semantic change detection in remote sensing imagery. In NeurIPS Workshop on Machine Learning for the Developing World, 2020. 2, 5, 6 +[57] Shiqi Tian, Yanfei Zhong, Zhuo Zheng, Ailong Ma, Xicheng Tan, and Liangpei Zhang. Large-scale deep learning based binary and semantic change detection in ultra high resolution remote sensing imagery: From benchmark datasets to urban application. ISPRS Journal of Photogrammetry and Remote Sensing, 193:164-186, 2022. 1, 2, 6 +[58] Aysim Tucker, Marvin Eisenberger, Daniel Cremers, and Laura Leal-Taixe. Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation. In CVPR, 2024. 2 +[59] Aysim Toker, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andres Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, Giovanni Marchisio, Xiao Xiang Zhu, and Laura Leal-Taixe. + +Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In CVPR, 2022. 1, 6 +[60] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In CVPR, 2021. 1 +[61] Zhipan Wang, Minduan Xu, Zhongwu Wang, Qing Guo, and Qingling Zhang. Scribblednet: Change detection on high-resolution remote sensing imagery with scribble interaction. International Journal of Applied Earth Observation and Geoinformation, 128:103761, 2024. 1 +[62] Junshi Xia, Naoto Yokoya, Bruno Adriano, and Clifford Broni-Bediako. OpenEarthMap: A benchmark dataset for global high-resolution land cover mapping. In WACV, 2023. 1, 3 +[63] Kunping Yang, Gui-Song Xia, Zicheng Liu, Bo Du, Wen Yang, Marcello Pelillo, and Liangpei Zhang. Asymmetric siamese networks for semantic change detection in aerial images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-18, 2022. 2, 5, 6 +[64] Yang Zhan, Kun Fu, Menglong Yan, Xian Sun, Hongqi Wang, and Xiaosong Qiu. Change detection based on deep siamese convolutional network for optical aerial images. IEEE Geoscience and Remote Sensing Letters, 14(10):1845-1849, 2017. 2 +[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 4, 1 +[66] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3 +[67] Xiaokang Zhang, Weikang Yu, Man-On Pun, and Wenzhong Shi. Cross-domain landslide mapping from large-scale remote sensing images using prototype-guided domain-aware progressive representation learning. ISPRS Journal of Photogrammetry and Remote Sensing, 197:1-17, 2023. 2 +[68] Wen-Jie Zheng, Xi-Le Zhao, Yu-Bang Zheng, and Zhi-Feng Pang. Nonlocal patch-based fully connected tensor network decomposition for multispectral image inpainting. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2021. 2 +[69] Zhuo Zheng, Stefano Ermon, Dongjun Kim, Liangpei Zhang, and Yanfei Zhong. Changen2: Multi-temporal remote sensing generative change foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2, 3 +[70] Zhuo Zheng, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery. In ICCV, 2021. 1, 2 +[71] Zhuo Zheng, Shiqi Tian, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Scalable multi-temporal remote sensing change data generation via simulating stochastic change process. In ICCV, 2023. 2, 3 +[72] Zhuo Zheng, Yanfei Zhong, Shiqi Tian, Ailong Ma, and Liangpei Zhang. Changemask: Deep multi-task encoder-transformer-decoder architecture for semantic change detection. ISPRS Journal of Photogrammetry and Remote Sensing, 183:228-239, 2022. 2 +[73] Zhuo Zheng, Yanfei Zhong, Ji Zhao, Ailong Ma, and Liangpei Zhang. Unifying remote sensing change detection via + +deep probabilistic change models: From principles, models to applications. ISPRS Journal of Photogrammetry and Remote Sensing, 215:239-255, 2024. 1, 2 +[74] Zhe Zhu, Shi Qiu, and Su Ye. Remote sensing of land change: A multifaceted perspective. Remote Sensing of Environment, 282:113266, 2022. 1 \ No newline at end of file diff --git a/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/images.zip b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..99409a5988fc4f50441ea251a85737fe92ed5158 --- /dev/null +++ b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3904664dd78d77c974004b1cefc314c15f87bf82ec1729e44bd2509645396143 +size 360720 diff --git a/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/layout.json b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..959254626fcfe190f245943ba657baa0a1d01e29 --- /dev/null +++ b/2025/The Change You Want To Detect_ Semantic Change Detection In Earth Observation With Hybrid Data Generationf/layout.json @@ -0,0 +1,9219 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "type": "text", + "content": "The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 161, + 151, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 161, + 151, + 174 + ], + "spans": [ + { + "bbox": [ + 82, + 161, + 151, + 174 + ], + "type": "text", + "content": "Yanis Benidir" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 256, + 161, + 342, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 161, + 342, + 174 + ], + "spans": [ + { + "bbox": [ + 256, + 161, + 342, + 174 + ], + "type": "text", + "content": "Nicolas Gonthier" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 448, + 161, + 526, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 448, + 161, + 526, + 174 + ], + "spans": [ + { + "bbox": [ + 448, + 161, + 526, + 174 + ], + "type": "text", + "content": "Clément Mallet" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 174, + 175, + 423, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 175, + 423, + 189 + ], + "spans": [ + { + "bbox": [ + 174, + 175, + 423, + 189 + ], + "type": "text", + "content": "Univ Gustave Eiffel, ENSG, IGN, LASTIG, France" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 231, + 192, + 367, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 192, + 367, + 202 + ], + "spans": [ + { + "bbox": [ + 231, + 192, + 367, + 202 + ], + "type": "text", + "content": "firstname_lastname@ign.fr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "content": "Bi-temporal change detection at scale based on Very High Resolution (VHR) images is crucial for Earth monitoring. This remains poorly addressed so far: methods either require large volumes of annotated data (semantic case), or are limited to restricted datasets (binary set-ups). Most approaches do not exhibit the versatility required for temporal and spatial adaptation: simplicity in architecture design and pretraining on realistic and comprehensive datasets. Synthetic datasets are the key solution but still fail to handle complex and diverse scenes. In this paper, we present HySCDG a generative pipeline for creating a large hybrid semantic change detection dataset that contains both real VHR images and inpainted ones, along with land cover semantic map at both dates and the change map. Being semantically and spatially guided, HySCDG generates realistic images, leading to a comprehensive and hybrid transfer-proof dataset FSC-180k. We evaluate FSC-180k on five change detection cases (binary and semantic), from zero-shot to mixed and sequential training, and also under low data regime training. Experiments demonstrate that pretraining on our hybrid dataset leads to a significant performance boost, outperforming SyntheWorld, a fully synthetic dataset, in every configuration. All codes, models, and data are available here: https://yb23.github.io/projects/cywd/." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 574, + 135, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 574, + 135, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 574, + 135, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "Efficient methods for change detection (CD) are crucial for monitoring territories and the various phenomena and activities that impact them [3]. As human and climatic pressures increase, this topic is of utmost importance and calls every year for higher spatial resolution and semantic accuracy [55]. Change detection at scale has benefited greatly from the contributions of deep learning in recent years [7, 12, 69, 73], but most methods are grounded on large volumes of annotated data [41, 60]. Creating a large-scale dataset for bi-temporal (pair-wise) remote sensing change detection poses signifi" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 228, + 539, + 437 + ], + "blocks": [ + { + "bbox": [ + 328, + 228, + 539, + 437 + ], + "lines": [ + { + "bbox": [ + 328, + 228, + 539, + 437 + ], + "spans": [ + { + "bbox": [ + 328, + 228, + 539, + 437 + ], + "type": "image", + "image_path": "0b9ea8ffcb9c09a66d1860e45091d22e2b6a0f676515dde8384ccb74ae5f7ef0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 438, + 459, + 447 + ], + "lines": [ + { + "bbox": [ + 405, + 438, + 459, + 447 + ], + "spans": [ + { + "bbox": [ + 405, + 438, + 459, + 447 + ], + "type": "text", + "content": "Target datasets" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 455, + 556, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 455, + 556, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 556, + 536 + ], + "type": "text", + "content": "Figure 1. Efficient and scalable change detection requires a comprehensive training dataset that does not exist today. Using a single-temporal dataset (image+semantic map), we propose HySCDG that generates a novel bi-temporal hybrid dataset FSC-180k. This enables multiple transfer learning possibilities on either binary and semantic change detection tasks. " + }, + { + "bbox": [ + 313, + 455, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{O} / \\mathcal{O}" + }, + { + "bbox": [ + 313, + 455, + 556, + 536 + ], + "type": "text", + "content": " : Transfer with/without fine-tuning our model, respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 641 + ], + "type": "text", + "content": "cant challenges and costs, especially when dealing with Very High Resolution (VHR) images (ground sampling distance: GSD " + }, + { + "bbox": [ + 313, + 558, + 556, + 641 + ], + "type": "inline_equation", + "content": "< 1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 558, + 556, + 641 + ], + "type": "text", + "content": ", [30]). This is due to the need for expertise and considerable effort in collecting, preprocessing, and annotating images [59, 62], particularly in the case of VHR data and multiple labels (termed semantic change detection, SCD) [57, 61, 70, 74]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "As an alternative, synthetic data generation is a promising direction for bi-temporal labeled data provision from a single temporal domain that is successful in diverse areas such as urban sprawl [17], disaster or climate change monitoring [42] or crop mapping [36]. Two main paradigms exist for CD: fully synthetic datasets (mainly graphics-based" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2204" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 358 + ], + "type": "text", + "content": "[27, 53]), and hybrid solutions that mix real and modified images, often at the instance level [5, 69, 71], known as data augmentation. The first method synthesizes image pairs by rendering two versions of a 3D virtual world thanks to a computer graphics engine. On the other hand, hybrid approaches synthesize new image pairs by pasting or erasing object instances into existing images ( inpainting). They propose a suitable trade-off and faster generation of large quantities of data, with likely higher realism. However, none of these solutions meet the requirements for scalable SCD with VHR imagery (Tab. 1): focus on single tasks, restricted geographical areas, limited diversity, no guarantee on semantic consistency between pairs [48] and subsequently poor transfer to real datasets. To address this challenge, we introduce a generative pipeline built upon Stable Diffusion [44] and ControlNet [65], designed to harness an existing VHR semantic land cover dataset [18] and instance masks, to generate an extensive hybrid semantic bi-temporal change detection dataset encompassing both real and inpainted images, semantic land cover maps at each date, and a change map (Fig. 1). Then, we evaluate the transferability of our synthetic dataset on five datasets for both binary and semantic change detection cases ranging from zero-shot to sequential and mixed training set-ups, as well as under low-data regime." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 360, + 202, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 360, + 202, + 370 + ], + "spans": [ + { + "bbox": [ + 67, + 360, + 202, + 370 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 371, + 295, + 538 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 67, + 371, + 295, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 371, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 67, + 371, + 295, + 454 + ], + "type": "text", + "content": "- We present a generation pipeline HySCDG for creating VHR bi-temporal SCD datasets based on a land cover real-world dataset by using Stable Diffusion and ControlNet with tunable and semantically-guided inpainting of individual objects. It can be easily configured or finetuned on any land-cover dataset, generating images with the dataset's distinctive traits." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 455, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 455, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 67, + 455, + 295, + 502 + ], + "type": "text", + "content": "- We provide FSC-180k: a Hybrid Semantic Change Detection pretraining dataset, created from FLAIR [18] along with the instance masks for 300k objects for the corresponding images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 503, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 503, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 67, + 503, + 295, + 538 + ], + "type": "text", + "content": "- We provide a comprehensive evaluation of the proposed synthetic dataset on multiple transfer learning scenarios on five change datasets, in particular for SCD." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 550, + 139, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 139, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 139, + 562 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": "Detecting changes in the deep learning era. We target bi-temporal land cover change detection in VHR images [32]. This historical task [22, 37, 52] consists in detecting the meaningful elements of a given area that have changed between two acquisitions and providing the land cover categories at each period (also known as change trajectory). It presents challenges: radiometric variations, geometric variations and non-permanent changes. They have been extensively addressed with supervised deep learning techniques [70]: from CNN [11, 12, 64] to Transformers [2, 7, 13, 55, 72, 73], mainly with Siamese architecture. Multiple datasets, accompanied with ad-hoc methods, have been" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": "created to feed such models. They are often limited by their size [6, 25, 39, 56, 63], geographical extent [6, 56, 57, 63], low geometrical or annotation quality [11]. Also, most of the effort focuses on binary change detection, i.e., restricted to a unique category, mainly buildings [6, 25, 39, 40, 50] or disasters [67]. Such scarcity in high quality, diverse, large-scale labels is alleviated with transfer learning [4, 51], by fine-tuning models pretrained on large-scale datasets for different image domains, even for semantic change detection [56, 57, 63]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 207, + 556, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 556, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 556, + 219 + ], + "type": "text", + "content": "Synthesizing and inpainting remote sensing images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 220, + 556, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 556, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 556, + 435 + ], + "type": "text", + "content": "Synthetic remote sensing data is useful e.g., for cloud removal [10], image restoration [16], or training deep supervised models [58], before transfer to real examples. Solutions based on patches, [35, 68], autoencoders [23], GAN [28, 34], -aligned generation [24] have been proposed, but with limited image quality and no semantic control. Diffusion models (DM) have improved such quality [44], but most models have applied DM directly on mid-resolution RGB images [14, 26, 58], neglecting both the multi-spectral dimension [9] and VHR spaces. However, a key finding remains: semantically control the synthesis, e.g., by directly synthesizing text [26], semantic maps [58] or with a control module guided by edges [9], semantic maps [14] or metadata [26]. In our work, we fine-tune a Stable Diffusion and a ControlNet models to achieve semantically-controlled inpainting of VHR images as a basis to take advantage of existing images, and generate an hybrid semantic change detection dataset at scale and with high diversity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 582 + ], + "type": "text", + "content": "Generating synthetic change datasets. Two main approaches exist: fully synthetic or hybrid solutions. The first paradigm leverages advanced 3D rendering engines to create fully synthetic datasets [17, 27]. This procedural approach allows complete control over rendering parameters (instance placement, semantic classes, lighting conditions, etc.) generated from scratch. E.g., SyntheWorld [53] effectively leverages DM prompted by GPT-4 to create a diverse and large-scale fully synthetic dataset. On the other hand, hybrid approaches generate (semi-)synthetic data by inserting fake changes into real images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 630 + ], + "type": "text", + "content": "Such modifications can be random crops [49], instance copy-pasting, or inpainting made by GAN [5, 49, 71] or DM [54, 69]. Blending or style transfer methods are often used to enhance diversity [5, 49]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "A pivotal challenge is the semantic guidance: modifying few elements of images to mimic real scarce changes, e.g., with instance footprints extracted from external datasets [5] or segmentation model outputs [69]. [54] adopt a Denoising diffusion probabilistic model to generate a new image directly from a modified semantic map and a degraded version of the corresponding snapshot. Changen2 [69] is the closest" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2205" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": "work to ours, with a full pipeline to synthesize a hybrid semantic dataset, again with a DM conditioned by a semantic map. However, Changen2 creates synthetic data specific to a given target dataset in terms of image resolution, size, or change characteristics. In this paper, we explore semantically controlled inpainting based on DM and ControlNet to the intertwined challenge of diverse and multi-class modification of VHR images. It is termed HySCDG and leads to the hybrid dataset named FSC-180k. Tab. 1 provides a comparison with main synthetic change detection datasets." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 289, + 294, + 412 + ], + "blocks": [ + { + "bbox": [ + 55, + 203, + 296, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 296, + 281 + ], + "type": "text", + "content": "Table 1. Synthetic Change Datasets. Land cover change detection in remote sensing imagery. For the total number of pixels, we only consider strictly different images. The geographical zone's extension can be local [ ] national [ ] or global [ ] We denote open-access (OA) and directly downloadable datasets with ✓. In the case of IAug, we sum the number of pixels of the two datasets provided in this work. OEM means OpenEarthMap." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 289, + 294, + 412 + ], + "lines": [ + { + "bbox": [ + 57, + 289, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 57, + 289, + 294, + 412 + ], + "type": "table", + "html": "
DatasetOAPixels ×106GSD (m)ClassesSourceZoneType
SynCW [27]X370.61X田Synthetic
SMARS [17]✓1100.3-0.52X田Synthetic
IAug [5]X1,1670.075-0.51LEVIR-CD [6], WHU-CD [25]田Hybrid
Ce-100K [54]X6,5530.25-0.58OEM [62]田Synthetic
Changen2 [69]X7,0770.25-0.58OEM [62]田Hybrid
Changen [71]X11,7960.81xView2 [43]田Hybrid
SyntheWorld [53]✓18,3500.3-11XXSynthetic
FSC-180k (Ours)✓80,7400.216FLAIR [18]田Hybrid
", + "image_path": "a53d0462c68b1f86238ad2e2fe616f6309d39e990e418cf2c80f92b7989f7fb0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 442, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 297, + 586 + ], + "type": "text", + "content": "Transfer learning from synthetic/hybrid datasets. Transfer learning is used to assess the validity of the generated image pairs, and vary with the operational set-up [33]. SCD at scale is performed through sequential learning: pretraining on such datasets and fine-tuning on real-world cases [5, 27, 69, 71]. Fewer works [53] evaluate on mixed training sets (single training mixing both real and synthetic samples), which can help to prevent overfitting. An orthogonal solution considers low data regimes, i.e., fine-tuning on a very small train set [5, 53, 54], or even under zero-shot scenarios [54]. We evaluate our proposed dataset in the four conceivable transfer scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 597, + 270, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 270, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 270, + 611 + ], + "type": "text", + "content": "3. Hybrid generation of semantic changes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "content": "Much of the success of deep learning methods for Earth observation relies on large-scale training datasets with reliable annotations [18, 45]. However, no large-scale publicly available dataset exists for bi-temporal SCD: we propose HySCDG for Hybrid Semantic Change Detection Generation, a procedure to leverage existing large-scale land cover datasets and state-of-the-art image inpainting methods to simulate visually realistic bi-temporal semantic changes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 529, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 529, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 529, + 83 + ], + "type": "text", + "content": "This leads to the generation of the FSC-180k dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 84, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 555, + 180 + ], + "type": "text", + "content": "Our paper is grounded on two main ideas: (i) Adapting and fine-tuning a Stable Diffusion Model enables efficient inpainting of VHR images with a semantic control and geographical instance selection; (ii) Changes can be simulated with sufficient diversity by randomly selecting instances (and modifying their labels) within an existing land cover map that is paired with the image used as input. The pipeline is shown in Fig. 2. Details are provided in the following." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 186, + 550, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 550, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 550, + 198 + ], + "type": "text", + "content": "3.1. Adaptation of Stable Diffusion for Inpainting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 202, + 556, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 556, + 407 + ], + "type": "text", + "content": "We use Stable Diffusion (SD) [44], an open and state-of-the-art text-to-image synthesis method. Our objective is to modify SD to manage controllable inpainting specifically in remote sensing aerial images, for various landscapes or semantic classes. Though SD models are effective in generating natural images with clear focal points and layered depth, with artistic or cinematic styles, they require adaptation to be used for remote sensing images [26]: acquisition distance, intricate textures, no domination of a single element. Starting from a Stable Diffusion 2 Inpainting checkpoint1, we sequentially fine-tune the Autoencoder and the Diffusion U-Net, before adding and training a ControlNet [65] that leverages information extracted from an existing large-scale land cover dataset (namely, FLAIR [18], see Sec. 3.3). Our pipeline allows to generate 5-band images that incorporate infrared and elevation along with RGB channels, for the first time in satellite image synthesis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 420, + 556, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 556, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 556, + 587 + ], + "type": "text", + "content": "Adaptation of the Variational Autoencoder. Unlike most SD extensions, which focus on fine-tuning the central U-Net, we also adapted the Variational Autoencoder (VAE) that encodes images into the diffusion latent space. This adaptation targets efficient encoding of multi-band VHR images from the FLAIR dataset. The original VAE, trained with a KL-regularization factor [15], uses a combination of L1 and perceptual loss [66] alongside KL-divergence. To improve compression of remote sensing images, we added both L2 and Focal losses [31]: minimizing high-frequency errors reduces blurriness. Additionally, we included a \"color loss\" (L2 loss on each " + }, + { + "bbox": [ + 313, + 420, + 556, + 587 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 420, + 556, + 587 + ], + "type": "text", + "content": " patch) to limit oversaturation, which affects ground object color realism. Fine-tuning of the VAE took 160 hours A100 GPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 601, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 556, + 696 + ], + "type": "text", + "content": "Stable Diffusion training. We then fine-tuned the Diffusion U-Net for inpainting, adapting it to the new VAE and teaching it to synthesize and inpaint VHR images. Special attention was given to prompt selection. Although inpainting is the model's primary function, we included direct image generation from prompts in " + }, + { + "bbox": [ + 313, + 601, + 556, + 696 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 601, + 556, + 696 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 601, + 556, + 696 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 601, + 556, + 696 + ], + "type": "text", + "content": " of cases, as in standard SD. This training required 300 hours on a A100 GPU (30,000 iterations, batch size of 32)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 531, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 531, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 531, + 713 + ], + "type": "text", + "content": "1https://huggingface.co/stabilityai/stable-diffusion-2-inpainting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2206" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 68, + 518, + 294 + ], + "blocks": [ + { + "bbox": [ + 94, + 68, + 518, + 294 + ], + "lines": [ + { + "bbox": [ + 94, + 68, + 518, + 294 + ], + "spans": [ + { + "bbox": [ + 94, + 68, + 518, + 294 + ], + "type": "image", + "image_path": "8dda307eabf77920221443de3276d86a520b0d5380aa777621a7bc1573adb632.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "text", + "content": "Figure 2. HySCDG pipeline. From a single-temporal dataset composed of one VHR image " + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "text", + "content": ", a semantic map " + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "text", + "content": ", and some openly available labeled instances, we generate a new VHR image " + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "text", + "content": ", a new map " + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "inline_equation", + "content": "M_{2}" + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "text", + "content": " and subsequently a change map " + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "text", + "content": ". This results in the FSC-180k hybrid dataset. The two pivotal novelties consist in: (i) Adapting and fine-tuning a Stable-Diffusion Model from image inpainting and (ii) exploiting open geospatial data for inpainting prompt control and semantically guiding the objects to be modified. The combination of both solutions ensures a diverse at-scale VHR multi-class change detection dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 377, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 460 + ], + "type": "text", + "content": "Prompt Engineering. Prompts are designed using both each image's geographic coordinates and open land cover data (OpenStreetMap [38]). To avoid learning irrelevant details, prompts are composed of three elements: (i) Spatial (place name, city, region), (ii) Temporal (time of day, season), (iii) Semantic (frequent classes in the inpainted area, compared to the overall dataset)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 461, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 461, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 461, + 295, + 498 + ], + "type": "text", + "content": "Prompt example: \"Grass and agricultural vegetation next to a highway, locality of Savigny-en-Revermont, Bourgogne-Franche-Comté, in the morning, during Summer.\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 504, + 295, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 517 + ], + "type": "text", + "content": "3.2. Conditional change inpainting with ControlNet" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "type": "text", + "content": "Driven by our goal to control the new classes in the inpainted zones, we add a ControlNet (CN) architecture [65] to build a model that generates realistic VHR images from a semantic segmentation map (here land cover). CN is designed to add task-specific conditioning to a pretrained image DM. It allows injecting explicit semantic guidance during the diffusion process, conditioning the outputs on some reference image (a color map of the semantic class), in addition to a text prompt. Our training method enables CN to perform both inpainting and text-to-image synthesis, using the semantic condition. Training took 240 hours on a A100 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 660, + 295, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 295, + 685 + ], + "type": "text", + "content": "3.3. Select, Mask, Change, Inpaint : the HySCDG pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Inspired by [46], the core of our method lies in our \"Select and Inpaint\" mechanism. Given a VHR image along with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "type": "text", + "content": "its semantic segmentation and instances" + }, + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "type": "text", + "content": ", we inpaint these instances using the SD+CN inpainting method: the label of the instance is changed, the texture of the image is substituted accordingly. This results in a pair of VHR images, both with semantic maps and the corresponding semantic change maps. The complete HySCDG is provided in Fig. 2 and detailed below. It is illustrated for one image but is applied several times to each image of the single temporal dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "text", + "content": "HySCDG uncovered. For an image " + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "text", + "content": " with its semantic map " + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "inline_equation", + "content": "\\{O_i\\}" + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "text", + "content": " the set of " + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "inline_equation", + "content": "N_I" + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "text", + "content": " instances lying on the extent of " + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 481, + 554, + 516 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 517, + 555, + 685 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 326, + 517, + 535, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 517, + 535, + 529 + ], + "spans": [ + { + "bbox": [ + 326, + 517, + 535, + 529 + ], + "type": "text", + "content": "- Randomly select " + }, + { + "bbox": [ + 326, + 517, + 535, + 529 + ], + "type": "inline_equation", + "content": "N_{change}" + }, + { + "bbox": [ + 326, + 517, + 535, + 529 + ], + "type": "text", + "content": " instances among " + }, + { + "bbox": [ + 326, + 517, + 535, + 529 + ], + "type": "inline_equation", + "content": "\\{O_i\\}" + }, + { + "bbox": [ + 326, + 517, + 535, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 529, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 529, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 326, + 529, + 555, + 552 + ], + "type": "text", + "content": "- For each instance, generate an inpainting mask: footprint+spatial buffer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "text", + "content": "- Using " + }, + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "inline_equation", + "content": "M_1" + }, + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "text", + "content": ", take as “" + }, + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "inline_equation", + "content": "T_1" + }, + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "text", + "content": " class” the most frequent class inside the footprint: " + }, + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "inline_equation", + "content": "c_1" + }, + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "text", + "content": ". Then, select the larger convex hull of pixels of this class inside the buffered zone. These pixels correspond to the change mask: " + }, + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 326, + 552, + 555, + 600 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 601, + 553, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 601, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 326, + 601, + 553, + 624 + ], + "type": "text", + "content": "- Randomly select another class " + }, + { + "bbox": [ + 326, + 601, + 553, + 624 + ], + "type": "inline_equation", + "content": "c_2" + }, + { + "bbox": [ + 326, + 601, + 553, + 624 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 326, + 601, + 553, + 624 + ], + "type": "inline_equation", + "content": "T_2" + }, + { + "bbox": [ + 326, + 601, + 553, + 624 + ], + "type": "text", + "content": " and replace the previous label by this new label." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 624, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 624, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 326, + 624, + 555, + 660 + ], + "type": "text", + "content": "- Randomly select " + }, + { + "bbox": [ + 326, + 624, + 555, + 660 + ], + "type": "inline_equation", + "content": "N_{\\text{nochange}}" + }, + { + "bbox": [ + 326, + 624, + 555, + 660 + ], + "type": "text", + "content": " random shapes in the image: such masks appear only on the inpainting mask, not on the change mask, letting their labels unchanged." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 660, + 555, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 660, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 326, + 660, + 555, + 672 + ], + "type": "text", + "content": "- Apply " + }, + { + "bbox": [ + 326, + 660, + 555, + 672 + ], + "type": "inline_equation", + "content": "\\mathrm{SD + CN}" + }, + { + "bbox": [ + 326, + 660, + 555, + 672 + ], + "type": "text", + "content": " to masked image and new semantic map." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 673, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 673, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 326, + 673, + 555, + 685 + ], + "type": "text", + "content": "- Cumulate the previous steps: we obtain our inpainting" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": " A panoptic segmentation is the most suitable case but not mandatory, assuming a sufficient number of instances ensures enough diversity." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2207" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 96 + ], + "type": "text", + "content": "mask, change mask, and for " + }, + { + "bbox": [ + 75, + 72, + 294, + 96 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 75, + 72, + 294, + 96 + ], + "type": "text", + "content": " a new semantic map " + }, + { + "bbox": [ + 75, + 72, + 294, + 96 + ], + "type": "inline_equation", + "content": "M_2" + }, + { + "bbox": [ + 75, + 72, + 294, + 96 + ], + "type": "text", + "content": " and the corresponding new image " + }, + { + "bbox": [ + 75, + 72, + 294, + 96 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 75, + 72, + 294, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "text", + "content": "The pair of images " + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "inline_equation", + "content": "(I_1, I_2)" + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "text", + "content": " with their corresponding semantic maps " + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "inline_equation", + "content": "(M_1, M_2)" + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "text", + "content": ", and the change map " + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "text", + "content": " (difference between " + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "inline_equation", + "content": "M_2" + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "inline_equation", + "content": "M_1" + }, + { + "bbox": [ + 55, + 96, + 295, + 133 + ], + "type": "text", + "content": "), are the samples of the hybrid dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 296, + 233 + ], + "type": "text", + "content": "Label and instance extraction. For comprehensive multiclass learning, we select FLAIR dataset [18], a large-scale semantic segmentation map for 16 land cover classes over more than " + }, + { + "bbox": [ + 55, + 138, + 296, + 233 + ], + "type": "inline_equation", + "content": "800\\mathrm{km}^2" + }, + { + "bbox": [ + 55, + 138, + 296, + 233 + ], + "type": "text", + "content": ". FLAIR offers the advantage of covering areas where geographical instances of many FLAIR labels are publicly available3, alleviating requiring a panoptic segmentation. Around 300,000 instance masks have been extracted for this work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 360 + ], + "type": "text", + "content": "Inpainting signature mitigation. Inpainting produces realistic changes but new areas tend to have sometimes a high-frequency signature, as reported by other works [20, 29]. In order to discourage the model from simply learning this inpainting noise instead of learning the actual changes between images, we also inpaint random areas of the image without changing the semantic class in a similar way to the \"not changed regions\" from [46]. Besides, to obtain the best transition between the original image and the inpainted area, we use a continuous mask in the inference [1]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 413 + ], + "type": "text", + "content": "Selection of the new label. For each change area, we randomly select the new class " + }, + { + "bbox": [ + 55, + 365, + 296, + 413 + ], + "type": "inline_equation", + "content": "c_2" + }, + { + "bbox": [ + 55, + 365, + 296, + 413 + ], + "type": "text", + "content": " among all classes weighted by their frequency in the whole dataset divided by their frequencies in the change area." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": "Adoption of a buffer zone. For inpainting, we add a buffer around the footprint of each modified instance. This gives SD more freedom to generate smoother borders around the changed areas. It also helps handling small spatial discrepancies between footprints and the VHR image content [19]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 239, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 239, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 239, + 498 + ], + "type": "text", + "content": "3.4. Structure of the FSC-180k dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 504, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 695 + ], + "type": "text", + "content": "For each image in the FLAIR training set, we generate three unique modified images, resulting in 180,000 synthetic images, each with a corresponding " + }, + { + "bbox": [ + 55, + 504, + 296, + 695 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 504, + 296, + 695 + ], + "type": "text", + "content": " semantic map covering 16 distinct semantic classes. We call this dataset FSC-180k, short for FLAIR Synthetic Change. FSC-180k includes approximately 80 billion pixels with a ground resolution of " + }, + { + "bbox": [ + 55, + 504, + 296, + 695 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{m}" + }, + { + "bbox": [ + 55, + 504, + 296, + 695 + ], + "type": "text", + "content": " per pixel. Note that we can expand the dataset by pairing synthetic images that were generated (differently) from a single real image, process that we use at training time, effectively doubling the number of image pairs. Tab. 1 compares FSC-180k with other synthetic change detection datasets and shows that FSC-180k offers a much broader range of semantic changes beyond simple building modifications, is of higher spatial resolution, and is nearly ten times larger than comparable datasets. It should also be noted that the proportion of changes in FSC-180k is closer to the real" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "scenarios observed in Europe (prevalence of " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " change) than other academic datasets (see Tab. 2), excepting HiUCD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 113, + 555, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 113, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 555, + 257 + ], + "type": "text", + "content": "Quality assessment of FSC-180k. ControlNet is the most error-prone module: it may sometimes fail to respect the semantic map. We evaluated how unchanged the semantic map remains by performing 15-class semantic segmentation on real and generated images, thanks to a UNet trained for this task, leading to a satisfactory error rate below " + }, + { + "bbox": [ + 313, + 113, + 555, + 257 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 113, + 555, + 257 + ], + "type": "text", + "content": " between both outputs. We also leveraged standard metrics for image generation evaluation to assess the realism of the synthesis. We got an Inception Score [47] of 6.2 on the inpainted images, which is comparable to real images, and a low FID score [21] of 0.43 between inpainted and real images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 268, + 395, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 395, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 395, + 281 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 336 + ], + "type": "text", + "content": "We conduct various transfer learning experiments to validate hybrid data pretraining, comparing FSC-180k with the fully artificial SyntheWorld [53], the only available synthetic dataset with a size comparable to ours (see Tab. 1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 344, + 520, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 520, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 520, + 357 + ], + "type": "text", + "content": "4.1. Pretraining approaches with FSC-180k" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": "We assess FSC-180k's effectiveness by exploiting existing open datasets as targets under various settings (see Tab. 2 and Supplementary Material). Here, we first present the different transfer learning approaches adopted, and then, dive into the deep models and datasets considered (Sec. 4.2). Our objective is to demonstrate how pretraining on a hybrid dataset improves performance, for both binary and semantic change detection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 457, + 408, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 408, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 408, + 468 + ], + "type": "text", + "content": "The four scenarios are:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 469, + 555, + 577 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 325, + 469, + 553, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 469, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 325, + 469, + 553, + 493 + ], + "type": "text", + "content": "- Sequential Transfer Learning. Models are pretrained on FSC-180k, and then fine-tuned on the target data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 494, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 494, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 325, + 494, + 553, + 517 + ], + "type": "text", + "content": "- Low Data Regime. The same as above but with a fraction of the available training set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 325, + 517, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 517, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 325, + 517, + 553, + 540 + ], + "type": "text", + "content": "- Mixed Training. Models are trained on a blend of FSC-180k and the target data." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 541, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 541, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 325, + 541, + 555, + 577 + ], + "type": "text", + "content": "- Zero-Shot Learning. A model pretrained on FSC-180k is directly applied on the target dataset, with only a remapping of FSC-180k classes to align labels." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 586, + 443, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 443, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 443, + 597 + ], + "type": "text", + "content": "4.2. Experimental protocol" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 603, + 460, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 460, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 460, + 614 + ], + "type": "text", + "content": "4.2.1. Target datasets and metrics" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 618, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 689 + ], + "type": "text", + "content": "Our transfer experiments span five diverse datasets, among which three contain semantic labels (Tab. 2). This ensures variety in landscapes, sizes, and change types, i.e., a comprehensive evaluation. Among the plethora of (S)CD datasets, we selected the most popular ones: Levir-CD [6], S2Looking [50], SECOND [63] and HiUCD-mini [56]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "We used classical metrics from the change detection literature: Intersection over Union (IoU) and F1 score for binary" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 225, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 225, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 225, + 713 + ], + "type": "text", + "content": "3https://www.data.gouv.fr/en/datasets/bd-topo-r/" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2208" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "content": "cases and mainly the Semantic Change Segmentation score (SCS) from [59] for semantic evaluation. In this latter case, the Separated Kappa (SeK) from [63] is also informative though it tends to be more biased towards detection task than semantic prediction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 132, + 295, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 132, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 295, + 180 + ], + "type": "text", + "content": "We added some experiments related to HiUCD-XL [57], a follow-up dataset of HiUCD-mini [56] using semantic mIoU and change mIoU, metrics provided on the evaluation platform4 as test labels are kept private." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 220, + 299, + 323 + ], + "blocks": [ + { + "bbox": [ + 55, + 189, + 295, + 211 + ], + "lines": [ + { + "bbox": [ + 55, + 189, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 295, + 211 + ], + "type": "text", + "content": "Table 2. Target datasets. Overview of the real-world datasets used for our transfer learning experiments." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 220, + 299, + 323 + ], + "lines": [ + { + "bbox": [ + 56, + 220, + 299, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 299, + 323 + ], + "type": "table", + "html": "
DatasetChange (%)Pixels ×106GSD (m)ClassesPeriodZone
Levir-CD [6]28.41,0670.512002/18Texas
S2Looking [50]25.68,3890.3-0.812017/20World
SECOND [63]20.01,2220.3-0.66?China
HiUCD-mini [56]6.65670.192017/19Estonia
HiUCD-XL [57]1.210,6950.192017/19Estonia
FSC-180k (Ours)7.080,7400.2162018/21France
", + "image_path": "8d8bf289a8c3d146ffd7f85d38b8c75b4356144c5fabd0e20c199718e1c912b8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 165, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 165, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 165, + 347 + ], + "type": "text", + "content": "4.2.2. Model architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 351, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 351, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 295, + 507 + ], + "type": "text", + "content": "We consider that designing a new architecture is out of the scope of this paper and that numerous solutions have already been proposed (see Sec. 2). Therefore, we initially used a simple U-Net model [8] with a segmentation head for semantic change detection. Although effective for binary change, it lacked expressiveness for semantic tasks. Similarly to [11], we developed a simple Dual UNet model composed of 2 nearly identical parallel UNets for (i) change detection, and (ii) semantic segmentation, with ResNet-50 backbones pretrained on ImageNet and trained in a multi-task setting. The change branch incorporates features from the semantic encoder. We keep the training strategy simple and constant to focus on data and transfer, this enhances interpretability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 515, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 659 + ], + "type": "text", + "content": "Can a binary model be sufficient? We tested both a pure binary CD model and the Dual UNet, for which, when, switching to binary-only tasks, the semantic decoder is inactive but the encoding part remains active (as it provides features to the change detection branch). Such selective change detection ability allows training on a multi-class dataset and focusing on specific binary tasks. Experiments show that even for binary tasks, the Dual UNet outperforms the binary-only model, e.g., achieving an IoU of 0.83 on Levir-CD (compared to 0.80). Learning more information than just change pixels, the Dual UNet gains a deeper scene understanding, leading to more reliable predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 673, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 696 + ], + "type": "text", + "content": "Existing architectures. We tested more complex transformer-based architectures, namely ChangeFormer [2]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "MambaCD [7], and SCanNet [13], both on HiUCD-mini and SECOND datasets. SCanNet performed best, balancing high performance and straightforward implementation due to its lightweight design (see Supplementary Material). SCanNet is kept hereafter for comparisons (see Tab. G)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 147, + 553, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 553, + 194 + ], + "type": "text", + "content": "Transfer for varying semantic classes. In the semantic case, we faced the issue of distinct classes between source and target datasets. This required to choose which classes should be kept for model training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 195, + 555, + 290 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 325, + 195, + 555, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 195, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 325, + 195, + 555, + 254 + ], + "type": "text", + "content": "- Sequential case: for both pretraining and fine-tuning strategies, we use the classes of the processed dataset. The semantic head of the model is reinitialized and adapted to the new number of classes before fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 255, + 555, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 255, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 325, + 255, + 555, + 290 + ], + "type": "text", + "content": "- Mixed case: Classes from the target dataset are used, manually mapping pretraining classes to their closest counterparts." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 297, + 432, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 432, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 432, + 308 + ], + "type": "text", + "content": "4.2.3. Experimental setting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": "We apply basic augmentations on the images: horizontal and vertical flips, " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": " rotations, and small rescaled crops (scales above 0.8). Normalization by mean and variance from the target dataset is used in all experiments. For optimization, we use the AdamW algorithm with default parameters (0.01 weight decay), and an initial learning rate of 3e-4 with exponential decay (0.99 per epoch). For sequential training, models are pretrained for 40h on a A100 GPU, with finetuning requiring up to 20h on a V100. For mixed training, we limit the total training time to 20h on a A100." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 439, + 433, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 433, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 433, + 451 + ], + "type": "text", + "content": "4.3. Results and Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 456, + 424, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 424, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 424, + 468 + ], + "type": "text", + "content": "4.3.1. Sequential training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 472, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 567 + ], + "type": "text", + "content": "Binary Case. Models pretrained on synthetic datasets outperform those without pretraining on all target datasets (Fig. 3). SyntheWorld maintains commendable performance on the single-class \"building\" change case, similarly to Levir-CD. Pretraining on FSC-180k yields the strongest results, compared to SyntheWorld on all five target datasets, with gains of " + }, + { + "bbox": [ + 313, + 472, + 555, + 567 + ], + "type": "inline_equation", + "content": "2 - 3\\%" + }, + { + "bbox": [ + 313, + 472, + 555, + 567 + ], + "type": "text", + "content": " of IoU. Details are provided in the Supplementary Material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Semantic Case. Pretraining on FSC-180k outperforms the baseline and SyntheWorld on the three SCD datasets in the sequential mode (Tab. G). Indeed, pretraining on SyntheWorld, that focuses solely on the \"building\" class, shows limitations in multiclass tasks on HiUCD-mini and HiUCD-XL compared to the baseline. Our FSC-180k dataset is particularly beneficial for HiUCD-mini and HiUCD-XL. The types of landscapes and spatial resolution are relatively similar, which facilitates effective transfer learning. For HiUCD-mini, our pretraining results in a SeK of 0.19 and a SCS of 0.78, representing respective increases of " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "7\\%" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 225, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 225, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 225, + 712 + ], + "type": "text", + "content": "4https://wwwcodabench.org/competitions/3485/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2209" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 91, + 156, + 162 + ], + "blocks": [ + { + "bbox": [ + 100, + 77, + 145, + 87 + ], + "lines": [ + { + "bbox": [ + 100, + 77, + 145, + 87 + ], + "spans": [ + { + "bbox": [ + 100, + 77, + 145, + 87 + ], + "type": "text", + "content": "(a) Levir-CD" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 91, + 156, + 162 + ], + "lines": [ + { + "bbox": [ + 59, + 91, + 156, + 162 + ], + "spans": [ + { + "bbox": [ + 59, + 91, + 156, + 162 + ], + "type": "image", + "image_path": "02a086d63be6387b772be6ec1d6d29fb2de78353164f380dd705ef0d9d775761.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 188, + 91, + 271, + 161 + ], + "blocks": [ + { + "bbox": [ + 214, + 77, + 263, + 87 + ], + "lines": [ + { + "bbox": [ + 214, + 77, + 263, + 87 + ], + "spans": [ + { + "bbox": [ + 214, + 77, + 263, + 87 + ], + "type": "text", + "content": "(b) S2Looking" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 188, + 91, + 271, + 161 + ], + "lines": [ + { + "bbox": [ + 188, + 91, + 271, + 161 + ], + "spans": [ + { + "bbox": [ + 188, + 91, + 271, + 161 + ], + "type": "image", + "image_path": "2da65fbc9a55ae153200b48257ae5eb2d7c9990447562f39913f51e1aeadd91b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 184, + 285, + 273 + ], + "blocks": [ + { + "bbox": [ + 101, + 170, + 145, + 178 + ], + "lines": [ + { + "bbox": [ + 101, + 170, + 145, + 178 + ], + "spans": [ + { + "bbox": [ + 101, + 170, + 145, + 178 + ], + "type": "text", + "content": "(c) SECOND" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 184, + 285, + 273 + ], + "lines": [ + { + "bbox": [ + 58, + 184, + 285, + 273 + ], + "spans": [ + { + "bbox": [ + 58, + 184, + 285, + 273 + ], + "type": "image", + "image_path": "bad28aa6f7848be81b4f5270a398e1ecfd421fc506d75d11b1fa83488862f93d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 170, + 266, + 178 + ], + "lines": [ + { + "bbox": [ + 211, + 170, + 266, + 178 + ], + "spans": [ + { + "bbox": [ + 211, + 170, + 266, + 178 + ], + "type": "text", + "content": "(d) HiUCD-mini" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 281, + 295, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 326 + ], + "type": "text", + "content": "Figure 3. Sequential training in binary change detection. After pretraining our Dual U-Net on either nothing, SyntheWorld or FSC-180k, we finetune and test it on each of the 4 target datasets and represent the binary IoU." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "type": "text", + "content": "compared to the baseline. The gains rise to " + }, + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 55, + 348, + 296, + 396 + ], + "type": "text", + "content": " against SyntheWorld. For SECOND, SyntheWorld partially narrows the gap: SeK of 0.17 and a SCS of 0.84, while the FSC-180k's pretrained Dual UNet achieves 0.18 and 0.87." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 56, + 471, + 285, + 572 + ], + "blocks": [ + { + "bbox": [ + 55, + 407, + 296, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 462 + ], + "type": "text", + "content": "Table 3. Sequential training in the semantic change detection framework. We report the binary change detection scores (IoU or Chg. mIoU) as well as semantic scores (SCS or mIoU) on three datasets in the sequential scheme. Models were initially pretrained (SyntheWorld or FSC-180k) or not (Baseline)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 471, + 285, + 572 + ], + "lines": [ + { + "bbox": [ + 56, + 471, + 285, + 572 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 285, + 572 + ], + "type": "table", + "html": "
PretrainingModelSECONDHiUCD-mini
IoUSCSIoUSCS
NoneDual UNet0.530.830.610.73
SyntheWorldDual UNet0.540.840.580.73
FSC-180kDual UNet0.550.890.630.78
NoneSCanNet0.540.840.600.66
FSC-180kSCanNet0.560.890.630.71
", + "image_path": "9706e316b3ec69f458c61594065ab58eaaff08e34bb75d769608a04ea0f3043d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 102, + 574, + 250, + 649 + ], + "blocks": [ + { + "bbox": [ + 102, + 574, + 250, + 649 + ], + "lines": [ + { + "bbox": [ + 102, + 574, + 250, + 649 + ], + "spans": [ + { + "bbox": [ + 102, + 574, + 250, + 649 + ], + "type": "table", + "html": "
PretrainingHiUCD-XL
Chg. mIoUmIoU
None0.170.58
SyntheWorld0.170.58
FSC-180k0.190.60
", + "image_path": "44efc8a8f63b864cfb544dd6b4246a5aecf051af0802e02fe6280a5277e1c196.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Dependence on a CD model. We compared two CD model by running the same experiments on SECOND and HiUCD-mini with a SCanNet model (Tab. G). We were un" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": "able to reproduce the results reported in the original paper [13], despite following the described learning scheme. SCanNet still performs well on the SECOND dataset, achieving a SCS of 0.84, compared to 0.83 for the Dual UNet. The opposite trend is observed on HiUCD, where the Dual UNet reaches a SCS of 0.73, while SCanNet reaches 0.66. As neither of the two models stands out, we mainly present the results obtained with our Dual UNet." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 175, + 407, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 407, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 407, + 186 + ], + "type": "text", + "content": "4.3.2. Mixed training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 190, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 555, + 320 + ], + "type": "text", + "content": "Our experiments show, that the optimal ratio between synthetic and target data lies around " + }, + { + "bbox": [ + 313, + 190, + 555, + 320 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 190, + 555, + 320 + ], + "type": "text", + "content": " for HiUCD-XL, S2Looking or SECOND (Fig. 4) conversely to the conclusions in [53], where performance always increases with the amount of target data included in the training mix. Nevertheless, the optimal ratio is " + }, + { + "bbox": [ + 313, + 190, + 555, + 320 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 190, + 555, + 320 + ], + "type": "text", + "content": " for smaller dataset such as Levir-CD or HiUCD-mini, that may be too specific. However, the mixed training strategy proves to be really effective, outperforming the sequential one on HiUCD-XL with a balanced blend as training base, as noticed in Fig. 4, compared to Tab. G. We can deduce that:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 321, + 555, + 488 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 325, + 321, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 321, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 325, + 321, + 555, + 369 + ], + "type": "text", + "content": "- Sequential training maybe suboptimal: The lengthy fine-tuning process can lead the model to \"forget\" pretraining, and instead overfits on the target data, which finally reduces the benefits of pretraining." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 325, + 369, + 554, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 369, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 325, + 369, + 554, + 429 + ], + "type": "text", + "content": "- In contrast, mixed training augments the initial set and adds diversity throughout the training process. The model sees as many images from the target dataset as in the sequential mode, but continues to benefit from the pretraining dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 429, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 429, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 325, + 429, + 555, + 488 + ], + "type": "text", + "content": "- A \"balanced blend\" of synthetic and target data appears optimal, providing a sufficient frequency of target images for exploitation while strategically incorporating new synthetic images to enable exploratory learning and notably avoid overfitting." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 495, + 416, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 416, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 416, + 507 + ], + "type": "text", + "content": "4.3.3. Low Data regime" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "We apply a sequential scheme but limit fine-tuning to small subsets of the target data: specifically, " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " of the training set. Each experiment was repeated 10 times, with a random sampling of the training set. We consider these to be the most meaningful experiments, as this scenario closely matches real-world use cases. The primary objective of pretraining is to develop a model that can either swiftly adapt to new target datasets or perform well when only limited training data is available. One can observe on Fig. 5 a clear advantage for pretrained models at very low level (" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "). With such little data, the model cannot fully learn the specificities of the target dataset and relies primarily on its general knowledge and on the task itself (detecting changes), rather than on dataset-specific features. Unsurprisingly, models pretrained for this task perform significantly better. However, as the amount of available data increases, the gap tends to narrow. Models increasingly rely" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2210" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 75, + 531, + 190 + ], + "blocks": [ + { + "bbox": [ + 58, + 75, + 531, + 190 + ], + "lines": [ + { + "bbox": [ + 58, + 75, + 531, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 75, + 531, + 190 + ], + "type": "image", + "image_path": "9e8750ab7c06af2f29d18d359bbac0b995fa5a07ad979ad077049a71f206588d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 4. Mixed training (Binary and Semantic). Training on a blend of target and synthetic/hybrid (SyntheWorld or FSC-180k) train sets, containing a ratio of " + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{x\\%}" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": " samples from the target (including repetitions). Testing is performed on the target test set. " + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": " corresponds to fine-tuning exclusively on target dataset (without pretraining)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 252, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 289 + ], + "type": "text", + "content": "on data itself, to which they have equal access, yet pretraining on FSC-180k still yields better performance than using SyntheWorld or training from scratch." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 303, + 263, + 403 + ], + "blocks": [ + { + "bbox": [ + 82, + 303, + 263, + 403 + ], + "lines": [ + { + "bbox": [ + 82, + 303, + 263, + 403 + ], + "spans": [ + { + "bbox": [ + 82, + 303, + 263, + 403 + ], + "type": "image", + "image_path": "2a9e960ba543022cfa9a7c3f9e06bdee68639d6049ee2b10dc252c47e630835a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 82, + 408, + 265, + 507 + ], + "blocks": [ + { + "bbox": [ + 82, + 408, + 265, + 507 + ], + "lines": [ + { + "bbox": [ + 82, + 408, + 265, + 507 + ], + "spans": [ + { + "bbox": [ + 82, + 408, + 265, + 507 + ], + "type": "image", + "image_path": "7501483e6676c6556fbadecb195d60712632400487a232ad0e50b19cbe3475d6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 518, + 296, + 574 + ], + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 574 + ], + "type": "text", + "content": "Figure 5. Low data regime (Semantic). Fine-tuning on a limited part of the target train set (either " + }, + { + "bbox": [ + 55, + 518, + 296, + 574 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 518, + 296, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 518, + 296, + 574 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 518, + 296, + 574 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 518, + 296, + 574 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 518, + 296, + 574 + ], + "type": "text", + "content": " of randomly sampled examples) and evaluating on the whole test set. Models are initially pretrained (SyntheWorld or FSC-180k) or not (Baseline). Metrics are averaged on 10 runs for SECOND and HiUCD-mini." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 591, + 173, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 173, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 173, + 601 + ], + "type": "text", + "content": "4.3.4. Zero-Shot evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "While models pretrained on SyntheWorld struggle to transfer effectively without target sample exposure, pretraining on FSC-180k delivers respectable results, achieving IoU scores of 0.37 on SECOND and 0.36 on HiUCD-mini, showing performance drops of " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", compared to the fully supervised models (details are provided in the Supplementary Material). However, FSC-180k is less effective on LevirCD, where the performance drop reaches " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " (0.33 vs. 0.83 IoU). We argue that performance in this scenario is directly" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 252, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 555, + 323 + ], + "type": "text", + "content": "linked to dataset similarities: the limited realism in SyntheWorld hinders transferability, while FSC-180k's diverse landscapes and sparse changes further limit its effectiveness. For the semantic task, pretraining on FSC-180k enhances the model's understanding of change trajectories, achieving SCS of 0.24 on SECOND and 0.25 on HiUCD-mini." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 361, + 440, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 440, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 440, + 373 + ], + "type": "text", + "content": "4.3.5. Qualitative assessment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "content": "Qualitative analysis of the various segmentation results indicates that FSC-180k pretraining significantly enhances the network's ability to localize features in input images. This improvement leads to both more accurate segmentation within VHR images and higher semantic prediction accuracy for main classes of interest in the semantic case (e.g., buildings, roads, bare land for HiUCD-mini, buildings or non-vegetated ground surface for SECOND). Qualitative examples for the different datasets are provided in the Supplementary Material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "We introduced HySCDG, a new, simple bi-temporal change detection generation pipeline. It lies on Stable Diffusion and ControlNet for creating hybrid content of VHR images. HySCDG leverages an existing large-scale semantic land cover map to overcome the lack of diversity and size of real-world datasets. We provide FSC-180k, a pretraining dataset based on FLAIR [18], which includes bi-temporal real/in-painted images, and their associated semantic maps. The proposed solution was evaluated for transferability, on five datasets over four distinct training setups. Our experiments on both binary and semantic cases demonstrated strong transferability, confirming FSC-180k robustness and adaptability as well as HySCDG relevance." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2211" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 150 + ], + "type": "text", + "content": "This work was granted access to the HPC resources of IDRIS under the allocations AD011014690 and AD011014286 made by GENCI. We thank Ruben Gres for inspiring discussions and valuable feedback about ControlNet and Stable Diffusion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 188, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 188, + 297, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 188, + 297, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 188, + 297, + 220 + ], + "type": "text", + "content": "[1] Andrew. How to use Soft Inpainting - Stable Diffusion Art. https://stable-diffusion-art.com/soft-inpainting/, 2024.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 221, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 221, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 62, + 221, + 296, + 252 + ], + "type": "text", + "content": "[2] Wele Gedara Chaminda Bandara and Vishal M. Patel. A transformer-based siamese network for change detection. In IGARSS, pages 207-210, 2022. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 253, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 253, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 253, + 295, + 284 + ], + "type": "text", + "content": "[3] Marshall Burke, Anne Driscoll, David B. Lobell, and Stefano Ermon. Using satellite imagery to understand and promote sustainable development. Science, 371(6535), 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 327 + ], + "type": "text", + "content": "[4] Yinxia Cao and Xin Huang. A full-level fused cross-task transfer learning method for building change detection using noise-robust pretrained networks on crowdsourced labels. Remote Sensing of Environment, 284:113371, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 327, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 296, + 369 + ], + "type": "text", + "content": "[5] Hao Chen, Wenyuan Li, and Zhenwei Shi. Adversarial instance augmentation for building change detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 370, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 296, + 402 + ], + "type": "text", + "content": "[6] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10), 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 403, + 296, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 455 + ], + "type": "text", + "content": "[7] Hongruixuan Chen, Jian Song, Chengxi Han, Junshi Xia, and Naoto Yokoya. Changemamba: Remote sensing change detection with spatiotemporal state space model. IEEE Transactions on Geoscience and Remote Sensing, 62:1-20, 2024. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 456, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 486 + ], + "type": "text", + "content": "[8] Isaac Corley, Caleb Robinson, and Anthony Ortiz. A change detection reality check. arXiv preprint arXiv:2402.06994, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 487, + 296, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 530 + ], + "type": "text", + "content": "[9] Mikolaj Czerkawski and Christos Tachtatzis. Exploring the capability of text-to-image diffusion models with structural edge guidance for multi-spectral satellite image inpainting. IEEE Geoscience and Remote Sensing Letters, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 531, + 296, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 531, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 57, + 531, + 296, + 594 + ], + "type": "text", + "content": "[10] Mikolaj Czerkawski, Priti Upadhyay, Christopher Davison, Astrid Werkmeister, Javier Cardona, Robert Atkinson, Craig Michie, Ivan Andonovic, Malcolm Macdonald, and Christos Tachtatzis. Deep internal learning for inpainting of cloud-affected regions in satellite imagery. Remote Sensing, 14(6):1342, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 595, + 296, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 595, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 57, + 595, + 296, + 638 + ], + "type": "text", + "content": "[11] Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, and Yann Gousseau. Multitask learning for large-scale semantic change detection. Computer Vision and Image Understanding, 187:102783, 2019. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 639, + 296, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 639, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 57, + 639, + 296, + 670 + ], + "type": "text", + "content": "[12] Rodrigo Caye Daudt, Bertrand Le Saux, and Alexandre Boulch. Fully convolutional siamese networks for change detection. In ICIP, pages 4063-4067, 2018. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[13] Lei Ding, Jing Zhang, Haitao Guo, Kai Zhang, Bing Liu, and Lorenzo Bruzzone. Joint spatio-temporal modeling for semantic change detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-14," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 704 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 73, + 383, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 383, + 83 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 383, + 83 + ], + "type": "text", + "content": "2024. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 84, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 84, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 84, + 555, + 116 + ], + "type": "text", + "content": "[14] Miguel Espinosa and Elliot J. Crowley. Generate your own scotland: Satellite image generation conditioned on maps. NeurIPS 2023 Workshop on Diffusion Models, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 116, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 116, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 116, + 555, + 148 + ], + "type": "text", + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 148, + 555, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 148, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 316, + 148, + 555, + 179 + ], + "type": "text", + "content": "[16] Weiliang Fan, Jing M Chen, and Weimin Ju. A pixel missing patch inpainting method for remote sensing image. In International Conference on Geoinformatics, 2011. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 180, + 555, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 180, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 555, + 243 + ], + "type": "text", + "content": "[17] Mario Fuentes Reyes, Yuxing Xie, Xiangtian Yuan, Pablo d'Angelo, Franz Kurz, Daniele Cerra, and Jiaojiao Tian. A 2D/3D multimodal data simulation approach with applications on urban semantic segmentation, building extraction and change detection. ISPRS Journal of Photogrammetry and Remote Sensing, 205:74-97, 2023. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 245, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 555, + 308 + ], + "type": "text", + "content": "[18] Anatol Garioud, Nicolas Gonthier, Loic Landrieu, Apolline De Wit, Marion Valette, Marc Poupee, Sebastien Giordano, and Boris Wattrelos. FLAIR: a country-scale land cover semantic segmentation dataset from multi-source optical imagery. In NeurIPS Datasets and Benchmarks Track, 2023. 2, 3, 5, 8, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 309, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 341 + ], + "type": "text", + "content": "[19] Adrien Gressin, Nicole Vincent, Clément Mallet, and Nicolas Paparoditis. Semantic approach in image change detection. In ACIVS, 2013. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 341, + 554, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 554, + 373 + ], + "type": "text", + "content": "[20] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In CVPR, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 373, + 555, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 555, + 425 + ], + "type": "text", + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 426, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 555, + 458 + ], + "type": "text", + "content": "[22] Philip J Howarth and Gregory M Wickware. Procedures for change detection using landsat digital data. International Journal of Remote Sensing, 2(3):277-291, 1981. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 459, + 555, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 555, + 490 + ], + "type": "text", + "content": "[23] Wenli Huang, Ye Deng, Siqi Hui, and Jinjun Wang. Image inpainting with bilateral convolution. Remote Sensing, 14(23):6140, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 491, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 555, + 533 + ], + "type": "text", + "content": "[24] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 533, + 554, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 554, + 576 + ], + "type": "text", + "content": "[25] Shunping Ji, Shiqing Wei, and Meng Lu. Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set. IEEE Transactions on Geoscience and Remote Sensing, 57(1):574-586, 2018. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 577, + 555, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 555, + 619 + ], + "type": "text", + "content": "[26] Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David B Lobell, and Stefano Ermon. Diffusionsat: A generative foundation model for satellite imagery. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 619, + 554, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 619, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 316, + 619, + 554, + 661 + ], + "type": "text", + "content": "[27] Maria Kolos, Anton Marin, Alexey Artemov, and Evgeny Burnaev. Procedural synthesis of remote sensing images for robust change detection with neural networks. In Advances in Neural Networks, 2019. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 662, + 554, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 662, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 316, + 662, + 554, + 693 + ], + "type": "text", + "content": "[28] Andrey Kuznetsov and Mikhail Gashnikov. Remote sensing image inpainting with generative adversarial networks. In ISDFS, 2020. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 693, + 554, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 693, + 554, + 704 + ], + "spans": [ + { + "bbox": [ + 316, + 693, + 554, + 704 + ], + "type": "text", + "content": "[29] Ang Li, Qiuhong Ke, Xingjun Ma, Haiqin Weng, Zhiyuan" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2212" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Zong, Feng Xue, and Rui Zhang. Noise doesn't lie: towards universal detection of deep inpainting. In *IJCAI*, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 137 + ], + "type": "text", + "content": "[30] Zhuohong Li, Wei He, Jiepan Li, Fangxiao Lu, and Hongyan Zhang. Learning without exact guidance: Updating large-scale high-resolution land cover maps from low-resolution historical labels. In CVPR, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 296, + 179 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 179, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 179, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 179, + 296, + 223 + ], + "type": "text", + "content": "[32] Zhiyong Lv, Tongfei Liu, Jón Atli Benediktsson, and Nicola Falco. Land cover change detection techniques: Very-high-resolution optical images: A review. IEEE Geoscience and Remote Sensing Magazine, 10(1):44-63, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 223, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 296, + 255 + ], + "type": "text", + "content": "[33] Yuchi Ma, Shuo Chen, Stefano Ermon, and David B Lobell. Transfer learning in environmental remote sensing. Remote Sensing of Environment, 301:113924, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 255, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 286 + ], + "type": "text", + "content": "[34] Javier Marín and Sergio Escalera. Sssgan: Satellite style and structure generative adversarial networks. Remote Sensing, 13(19):3984, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 286, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 330 + ], + "type": "text", + "content": "[35] Fan Meng, Xiaomei Yang, Chenghu Zhou, and Zhi Li. A sparse dictionary learning-based adaptive patch inpainting method for thick clouds removal from high-spatial resolution remote sensing imagery. Sensors, 17(9):2130, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 330, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 330, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 296, + 373 + ], + "type": "text", + "content": "[36] Ali Mirzaei, Hossein Bagheri, and Iman Khosravi. Enhancing crop classification accuracy through synthetic SAR-Optical data generation using deep learning. ISPRS International Journal of Geo-Information, 12(11), 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 373, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 296, + 415 + ], + "type": "text", + "content": "[37] Hassiba Nemmour and Youcef Chibani. Multiple support vector machines for land cover change detection: An application for mapping urban extensions. ISPRS Journal of Photogrammetry and Remote Sensing, 61(2):125-133, 2006. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 415, + 296, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 447 + ], + "type": "text", + "content": "[38] OpenStreetMap contributors. Planet dump retrieved from https://planet.osm.org . https : //www.openstreetmap.org, 2017.4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 447, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 447, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 447, + 296, + 479 + ], + "type": "text", + "content": "[39] Chao Pang, Jiang Wu, Jian Ding, Can Song, and Gui-Song Xia. Detecting building changes with off-nadir aerial images. Science China Information Sciences, 66(4):140306, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 479, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 296, + 543 + ], + "type": "text", + "content": "[40] Daifeng Peng, Lorenzo Bruzzone, Yongjun Zhang, Haiyan Guan, Haiyong Ding, and Xu Huang. Semicdnet: A semisupervised convolutional neural network for change detection in high resolution remote-sensing images. IEEE Transactions on Geoscience and Remote Sensing, 59(7):5891-5906, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 543, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 296, + 576 + ], + "type": "text", + "content": "[41] R.S. Priya and K. Vani. Vegetation change detection and recovery assessment based on post-fire satellite imagery using deep learning. Nature Sci Rep, 14:12611, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 576, + 296, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 296, + 618 + ], + "type": "text", + "content": "[42] Xiaofan Qu, Feng Gao, Junyu Dong, Qian Du, and Heng-Chao Li. Change detection in synthetic aperture radar images using a dual-domain network. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 618, + 296, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 618, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 56, + 618, + 296, + 671 + ], + "type": "text", + "content": "[43] Gupta Ritwik, Hosfelt Richard, Sajeev Sandra, Patel Nirav, Goodman Bryce, Doshi Jigar, Heim Eric, Choset Howie, and Gaston Matthew. xbd: A dataset for assessing building damage from satellite imagery. arXiv preprint arXiv:1911.09296, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 672, + 296, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 672, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 56, + 672, + 296, + 704 + ], + "type": "text", + "content": "[44] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 706 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "type": "text", + "content": "[45] Ribana Roscher, Marc Russwurm, Caroline Gevaert, Michael Kampffmeyer, Jefersson A. Dos Santos, Maria Vakalopoulou, Ronny Hänsch, Stine Hansen, Keiller Nogueira, Jonathan Prexl, and Devis Tuia. Better, not just more: Data-centric machine learning for earth observation. IEEE Geoscience and Remote Sensing Magazine, pages 2-22, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 139, + 554, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 139, + 554, + 159 + ], + "spans": [ + { + "bbox": [ + 316, + 139, + 554, + 159 + ], + "type": "text", + "content": "[46] Ragav Sachdeva and Andrew Zisserman. The change you want to see. In WACV, 2023. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 159, + 554, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 159, + 554, + 202 + ], + "spans": [ + { + "bbox": [ + 316, + 159, + 554, + 202 + ], + "type": "text", + "content": "[47] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 202, + 554, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 202, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 202, + 554, + 243 + ], + "type": "text", + "content": "[48] Minseok Seo, Hakjin Lee, Yongjin Jeon, and Junghoon Seo. Self-pair: Synthesizing changes from single source for object change detection in remote sensing imagery. In WACV, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 244, + 554, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 244, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 554, + 287 + ], + "type": "text", + "content": "[49] Minseok Seo, Hakjin Lee, Yongjin Jeon, and Junghoon Seo. Self-pair: Synthesizing changes from single source for object change detection in remote sensing imagery. In WACV, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 287, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 554, + 331 + ], + "type": "text", + "content": "[50] Li Shen, Yao Lu, Hao Chen, Hao Wei, Donghai Xie, Jiabao Yue, Rui Chen, Shouye Lv, and Bitao Jiang. S2looking: A satellite side-looking dataset for building change detection. Remote Sensing, 13(24), 2021. 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 331, + 554, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 373 + ], + "type": "text", + "content": "[51] Wenzhong Shi, Min Zhang, Rui Zhang, Shanxiong Chen, and Zhao Zhan. Change detection based on artificial intelligence: State-of-the-art and challenges. Remote Sensing, 12(10):1688, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 373, + 554, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 554, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 554, + 406 + ], + "type": "text", + "content": "[52] Ashbindu Singh. Review article digital change detection techniques using remotely-sensed data. International journal of remote sensing, 10(6):989-1003, 1989. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 406, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 554, + 437 + ], + "type": "text", + "content": "[53] Jian Song, Hongruixuan Chen, and Naoto Yokoya. Syntheworld: A large-scale synthetic dataset for land cover mapping and building change detection. In WACV, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 437, + 554, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 479 + ], + "type": "text", + "content": "[54] Kai Tang and Jin Chen. Changeanywhere: Sample generation for remote sensing change detection via semantic latent diffusion model. arXiv preprint arXiv:2404.08892, 2024. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 479, + 554, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 554, + 512 + ], + "type": "text", + "content": "[55] Yizhuo Tang, Zhengtao Cao, Ningbo Guo, and Mingyong Jiang. A Siamese Swin-Unet for image change detection. Nature Sci Rep, 14:4577, 2024. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "text", + "content": "[56] Shiqi Tian, Yanfei Zhong, Ailong Ma, and Zhuo Zheng. HiUCD: A large-scale dataset for urban semantic change detection in remote sensing imagery. In NeurIPS Workshop on Machine Learning for the Developing World, 2020. 2, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 555, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 620 + ], + "type": "text", + "content": "[57] Shiqi Tian, Yanfei Zhong, Zhuo Zheng, Ailong Ma, Xicheng Tan, and Liangpei Zhang. Large-scale deep learning based binary and semantic change detection in ultra high resolution remote sensing imagery: From benchmark datasets to urban application. ISPRS Journal of Photogrammetry and Remote Sensing, 193:164-186, 2022. 1, 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 620, + 554, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 620, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 316, + 620, + 554, + 662 + ], + "type": "text", + "content": "[58] Aysim Tucker, Marvin Eisenberger, Daniel Cremers, and Laura Leal-Taixe. Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 662, + 554, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 662, + 554, + 706 + ], + "spans": [ + { + "bbox": [ + 316, + 662, + 554, + 706 + ], + "type": "text", + "content": "[59] Aysim Toker, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andres Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Senaras, Timothy Davis, Daniel Cremers, Giovanni Marchisio, Xiao Xiang Zhu, and Laura Leal-Taixe." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2213" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 707 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In CVPR, 2022. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 137 + ], + "type": "text", + "content": "[60] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In CVPR, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 296, + 191 + ], + "type": "text", + "content": "[61] Zhipan Wang, Minduan Xu, Zhongwu Wang, Qing Guo, and Qingling Zhang. Scribblednet: Change detection on high-resolution remote sensing imagery with scribble interaction. International Journal of Applied Earth Observation and Geoinformation, 128:103761, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 191, + 296, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 191, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 191, + 296, + 233 + ], + "type": "text", + "content": "[62] Junshi Xia, Naoto Yokoya, Bruno Adriano, and Clifford Broni-Bediako. OpenEarthMap: A benchmark dataset for global high-resolution land cover mapping. In WACV, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 296, + 287 + ], + "type": "text", + "content": "[63] Kunping Yang, Gui-Song Xia, Zicheng Liu, Bo Du, Wen Yang, Marcello Pelillo, and Liangpei Zhang. Asymmetric siamese networks for semantic change detection in aerial images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-18, 2022. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 296, + 341 + ], + "type": "text", + "content": "[64] Yang Zhan, Kun Fu, Menglong Yan, Xian Sun, Hongqi Wang, and Xiaosong Qiu. Change detection based on deep siamese convolutional network for optical aerial images. IEEE Geoscience and Remote Sensing Letters, 14(10):1845-1849, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 296, + 373 + ], + "type": "text", + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 4, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 373, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 296, + 405 + ], + "type": "text", + "content": "[66] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 405, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 296, + 459 + ], + "type": "text", + "content": "[67] Xiaokang Zhang, Weikang Yu, Man-On Pun, and Wenzhong Shi. Cross-domain landslide mapping from large-scale remote sensing images using prototype-guided domain-aware progressive representation learning. ISPRS Journal of Photogrammetry and Remote Sensing, 197:1-17, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 459, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 296, + 502 + ], + "type": "text", + "content": "[68] Wen-Jie Zheng, Xi-Le Zhao, Yu-Bang Zheng, and Zhi-Feng Pang. Nonlocal patch-based fully connected tensor network decomposition for multispectral image inpainting. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 502, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 296, + 545 + ], + "type": "text", + "content": "[69] Zhuo Zheng, Stefano Ermon, Dongjun Kim, Liangpei Zhang, and Yanfei Zhong. Changen2: Multi-temporal remote sensing generative change foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 586 + ], + "type": "text", + "content": "[70] Zhuo Zheng, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 587, + 296, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 296, + 630 + ], + "type": "text", + "content": "[71] Zhuo Zheng, Shiqi Tian, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Scalable multi-temporal remote sensing change data generation via simulating stochastic change process. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 631, + 296, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 631, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 56, + 631, + 296, + 684 + ], + "type": "text", + "content": "[72] Zhuo Zheng, Yanfei Zhong, Shiqi Tian, Ailong Ma, and Liangpei Zhang. Changemask: Deep multi-task encoder-transformer-decoder architecture for semantic change detection. ISPRS Journal of Photogrammetry and Remote Sensing, 183:228-239, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 684, + 296, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 684, + 296, + 707 + ], + "spans": [ + { + "bbox": [ + 56, + 684, + 296, + 707 + ], + "type": "text", + "content": "[73] Zhuo Zheng, Yanfei Zhong, Ji Zhao, Ailong Ma, and Liangpei Zhang. Unifying remote sensing change detection via" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 137 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 553, + 106 + ], + "type": "text", + "content": "deep probabilistic change models: From principles, models to applications. ISPRS Journal of Photogrammetry and Remote Sensing, 215:239-255, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 106, + 555, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 555, + 137 + ], + "type": "text", + "content": "[74] Zhe Zhu, Shi Qiu, and Su Ye. Remote sensing of land change: A multifaceted perspective. Remote Sensing of Environment, 282:113266, 2022. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2214" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_content_list.json b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ca630e73d61bd5b9acabb32e7935008c581354a6 --- /dev/null +++ b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_content_list.json @@ -0,0 +1,1855 @@ +[ + { + "type": "text", + "text": "The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation", + "text_level": 1, + "bbox": [ + 109, + 130, + 885, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhan Liu, Yixiong Zou*, Yuhua Li, Ruixuan Li \nSchool of Computer Science and Technology, Huazhong University of Science and Technology \n{yuhan.liu, yixiongz, idcliyuhua, rxli}@hust.edu.cn", + "bbox": [ + 122, + 180, + 875, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cross-Domain Few-Shot Segmentation (CDFSS) is proposed to transfer the pixel-level segmentation capabilities learned from large-scale source-domain datasets to downstream target-domain datasets, with only a few annotated images per class. In this paper, we focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, particularly those distant from the source domain, segmentation performance peaks at the very early epochs, and declines sharply as the source-domain training proceeds. We delve into this phenomenon for an interpretation: low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS. Based on this phenomenon and interpretation, we further propose a method that includes two plug-and-play modules: one to flatten the loss landscapes for low-level features during source-domain training as a novel sharpness-aware minimization method, and the other to directly supplement target-domain information to the model during target-domain testing by low-level-based calibration. Extensive experiments on four target datasets validate our rationale and demonstrate that our method surpasses the state-of-the-art method in CDFSS significantly by $3.71\\%$ and $5.34\\%$ average MIoU in 1-shot and 5-shot scenarios, respectively.", + "bbox": [ + 89, + 301, + 485, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 695, + 220, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current deep neural networks [24, 30, 43] have achieved remarkable success in semantic segmentation, but their performance heavily relies on large-scale annotated datasets [1, 23]. However, the process of annotating data, particularly for dense pixel-wise tasks like semantic segmentation, is labor-intensive and time-consuming. To address this issue, Few-Shot Segmentation (FSS) has been introduced, aiming to generate pixel-level predictions for unseen categories with only a few labeled samples. Typically, the model is pretrained on a large-scale dataset of base classes and then trans", + "bbox": [ + 89, + 722, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf8eb221a693560d4bf08f9f725c935c7292d640d6a6392f760e32a52267057f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 268, + 901, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c5b81cfd8696020ec59dc2d51f6f2c8d9888fb2e4f152d57bc0a24558bb0134.jpg", + "image_caption": [ + "(b) Previous CDFSS methods Epoch vs. MoU" + ], + "image_footnote": [], + "bbox": [ + 516, + 356, + 705, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40838c235abca060f02be4da9ce2eeb0786f2329a147d16afe4fb82367e652cb.jpg", + "image_caption": [ + "(c) Our method Epochs vs. MoU", + "Figure 1. (a) In CDFSS tasks, the training (source) and testing (target) datasets belong to distinct domains, with categories in the testing dataset being unseen during training. (b) Previous CDFSS methods show a decreasing trend of mIoU as the source-domain training proceeds, even at very early epochs for distant domains. (c) Our method can effectively prevent the model from performance decline after early epochs and achieve higher performance." + ], + "image_footnote": [], + "bbox": [ + 707, + 357, + 901, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ferred to unseen novel classes. However, novel classes might not be in the same domain as base classes [5, 7, 11, 22, 36], a more realistic approach is to consider domain gaps between the base (source domain) and novel (target domain) classes, which gives rise to the Cross-Domain Few-Shot Segmentation (CDFSS) task [20].", + "bbox": [ + 511, + 566, + 906, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although extensive works [16, 17, 20, 33] have been developed, a well-observed phenomenon is still not handled: for target domains, especially those distant from the source domain, the best performance is always achieved at the very early epochs or even the first epoch. As shown in Fig. 1b, the performance decreases sharply as the source-domain training proceeds, where the 20th epoch's MIoU is even lower than that of the first epoch. Although the early stop can handle this problem, the goal of the source-domain training is to provide a generalizable model for all target domains [20]. Therefore, it may be inappropriate to try different early stops for specific domains. In this paper, we aim to delve into this well-observed but under-explored phenomenon and handle it based on our interpretations of it, as shown in Fig. 1c.", + "bbox": [ + 511, + 657, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To delve into this phenomenon, we first visualize the predictions of the 1st and 20th epochs in Fig. 2. We observe that", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/748fbdb6f1c71e584d3f0948b6c7aee178a067d4d6ae95b262cc997a89ac5e90.jpg", + "image_caption": [ + "Figure 2. The visualization of predictions at the 1st and 20th epoch further indicate that CDFSS models may not acquire meaningful information for target domains during source-domain training." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 478, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the 20th epoch's model shows many fundamental errors in target domains, i.e., CDFSS models fail to identify the foreground but focus on entirely wrong regions. This suggests that the CDFSS models may not have acquired meaningful information for target domains during source-domain training, resulting in almost random predictions during testing. We then visualize feature maps at different network stages, and observe this problem occurs even at the shallow layers, and deeper layers produce wrong feature maps with wrong inputs from shallow layers. Quantitatively, previous works [13, 46] suggest that the challenge posed by domain shift can be analyzed through the flatness of the loss landscape. We use this as an entry point to explore the connection between shallow layers and early stops. We find that low-level features, although simpler and previously believed to be more transferable [44], are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training. That is, the devil lies in the low-level features.", + "bbox": [ + 88, + 247, + 482, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on this phenomenon and interpretation, we further propose a method for the CDFSS task. During the source-domain training, we propose a novel sharpness-aware minimization method to flatten the loss landscapes for low-level features, which is achieved by a shape-preserving perturbation module with random convolutions. During the target-domain phase, since low-level features can hardly capture the target-domain information due to its vulnerability to domain shifts, we directly supplement such information to the model with a low-level-based calibration module, which effectively prevents the model from fundamental errors caused by collapsed low-level features. Based on our methods, we can effectively prevent the model from dropping performances after early epochs and achieve higher performance (Fig. 1c).", + "bbox": [ + 89, + 520, + 482, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions can be listed as", + "bbox": [ + 109, + 733, + 416, + 747 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, particularly those distant from the source domain, the best segmentation performance is always achieved at the very early epochs, followed by a sharp decline as source-domain training progresses.", + "- We delve into this phenomenon for an interpretation: low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS." + ], + "bbox": [ + 89, + 750, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Building on this interpretation, we propose a method that includes two plug-and-play modules: one act as a novel sharpness-aware minimization method to flatten the loss landscapes of low-level features during source-domain training, while the other directly supplement target-domain information to the model during target-domain testing.", + "- Extensive experiments on four target datasets validate the rationale of our interpretation and method, showing its superiority over current state-of-the-art methods." + ], + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Interpretation", + "text_level": 1, + "bbox": [ + 511, + 239, + 658, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 265, + 653, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-Domain Few-Shot Segmentation (CDFSS) aims to transfer the segmentation capabilities learned from the source domain to target domains, with only a few annotated images per class. The source domain $\\mathcal{D}_s = (\\mathcal{X}_s,\\mathcal{Y}_s)$ and target domain $\\mathcal{D}_t = (\\mathcal{X}_t,\\mathcal{Y}_t)$ are defined by different input distributions and disjoint label spaces, i.e., $\\mathcal{X}_s\\neq \\mathcal{X}_t$ $\\mathcal{Y}_s\\cap \\mathcal{Y}_t = \\emptyset$ , where $\\mathcal{X}$ refers to input distribution and $\\mathcal{Y}$ refers to the label space.", + "bbox": [ + 511, + 287, + 905, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we adopt the meta-learning episodic paradigm following [20]. Specifically, both $\\mathcal{D}_s$ and $\\mathcal{D}_t$ consist of several episodes. Each episode is constructed by a support set $S = \\{(I_s^i,M_s^i)\\}_{i = 1}^K$ with $K$ training samples and a query set $Q = \\{(I_q,M_q)\\}$ , where $I$ denote the image and $M$ denote the pixel labels. Within each episode, the model leverages $\\{I_s,M_s\\}$ and $I_{q}$ to predict the query mask $M_{q}$ .", + "bbox": [ + 511, + 409, + 905, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The model is trained on the source domain support sets and query set $\\{S^s, Q^s\\}$ by minimizing the binary cross-entropy loss:", + "bbox": [ + 511, + 517, + 903, + 561 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = B C E \\left(F \\left(I _ {s} ^ {s}, I _ {q} ^ {s}, M _ {s} ^ {s}\\right), M _ {q} ^ {s}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 573, + 903, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $F\\left(I_s^s,I_q^s,M_s^s\\right) = h\\left(f\\left(I_s^s,I_q^s\\right),M_s^s\\right)$ outputs the segmentation score map of $I_{q}^{s}$ , which is composed of a feature extractor $f(\\cdot)$ and a comparison module $h(\\cdot)$ .", + "bbox": [ + 511, + 602, + 903, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Then the model will be evaluated on the target domain query set $Q^t$ , the prediction for $I_q^t$ is:", + "bbox": [ + 511, + 648, + 903, + 681 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} _ {q} ^ {t} = \\arg \\max F \\left(I _ {s} ^ {t}, I _ {q} ^ {t}, M _ {s} ^ {t}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 691, + 903, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, the evaluation can be conducted based on the comparison between $\\hat{M}_q^t$ and the real label $M_q^t$ .", + "bbox": [ + 511, + 722, + 903, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Delve into the Early Stop in Cross-Domain Few-Shot Segmentation", + "text_level": 1, + "bbox": [ + 511, + 762, + 903, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2.1 Intuitive observation: The devil is in the low-level features", + "text_level": 1, + "bbox": [ + 511, + 800, + 903, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given our suspicion that the model fails to learn useful information from the source domain for target domains, we start by visualizing the feature maps. Using ResNet-50 [15] as the backbone following [12], we visualize the feature maps", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b3318b69b17f2be6d6b9e2082fac3c9c7927ff459e5ecb4009fac1c044c7599f.jpg", + "image_caption": [ + "Figure 3. Feature maps from stages 1 to 4 for both the source domain and target domains. The noticeable contrast in the feature maps at stage 1 between source and target domains indicates the limited performance of CDFSS models stems from shallow layers." + ], + "image_footnote": [], + "bbox": [ + 122, + 87, + 444, + 311 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/561a7f18308d7481b384c7e071df76c95db513c6be2b7bf79cb6ed55c51ff6cc.jpg", + "image_caption": [ + "Figure 4. Feature maps from stages 1 to 4 for epoch 1 and epoch 20. Epoch 1 in stage 1 shows more distinguishable activations than epoch 20, indicating that low-level features gradually incorporate incorrect information as training progresses." + ], + "image_footnote": [], + "bbox": [ + 91, + 380, + 478, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of stages 1 to 4 for both the source domain and the target domains, as shown in Fig. 3. We observe that on source-domain samples, the model learns boundary information effectively in stage 1, enabling it to accurately focus on the foreground in stage 4. However, on target domains especially those distant from the source domain, even though the foreground is clear and prominent, the model still fails to learn anything meaningful in stages 1 and 2. Consequently, in stage 4, the model focuses on completely incorrect areas. This indicates that the model's bad performance may originate from its shallow layers.", + "bbox": [ + 88, + 525, + 482, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Fig.4, we further visualize feature maps of each stage at the 1st and 20th epoch, respectively. We find that the shallow layers show more distinguishable activations at epoch 1 than at epoch 20, which aligns with our observation of higher accuracy at epoch 1. Consequently, we conclude that the poor performance of CDFSS models may be attributed to shallow layers that fail to capture useful low-level information, a trend that intensifies during source-domain training.", + "bbox": [ + 89, + 691, + 482, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.2 Quantitative verification: Low-level features lead to a sharper loss landscape", + "text_level": 1, + "bbox": [ + 89, + 830, + 482, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous work [13, 46] has demonstrated that the sharpness of the loss landscape can serve as a tool to analyze cross", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f6a00bcf4994f71fed9b420510b09808e4b38cfeed046a59527475ad05cc595b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 550, + 88, + 883, + 180 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8deb8f3d8d307ec6b234a9c5a7f5654876cbe8d86fe72babe057b93515b738b7.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 550, + 183, + 893, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c1206a13d5519682b3ffc20def6075c7c929eef76c8c0e3bc7a3aed26238480.jpg", + "image_caption": [ + "(c)", + "Figure 5. (a) A sharp minimum in the landscape corresponds to a representation that is highly sensitive to data shifts. (b) Examples of pixel perturbation applied to the images. (c) As training proceeds, the loss landscape becomes progressively sharper. (d) Perturbing shallow layers leads to much sharper loss landscapes, indicating shallow layers are the cause of the sharp loss landscape and the increased sensitivity to domain shifts." + ], + "image_footnote": [], + "bbox": [ + 545, + 224, + 710, + 320 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/311d767e60f852abe75d3366fa3cb4b8185ccd73533c7d0f1d6e9a089318d5f1.jpg", + "image_caption": [ + "(d) 0.3" + ], + "image_footnote": [], + "bbox": [ + 727, + 224, + 898, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "domain issues. As illustrated in Fig. 5a, a smoother loss landscape enables the model to better handle domain shifts1. Therefore, we measure the sharpness of the source-domain-trained models against different epochs and stages.", + "bbox": [ + 511, + 436, + 903, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first measure the sharpness of the loss landscape across different epochs. Following the method in [46], we apply pixel perturbations to the training data. Examples of these perturbations are shown in Fig. 5b, while the magnitude of performance drop is depicted in Fig.5c. A larger performance drop under the same perturbation indicates a sharper loss landscape. As observed, the performance of models across all epochs drops when exposed to perturbations, but the model from epoch 1 experiences a smaller drop than that of epoch 10, which in turn performs better than epoch 20. This suggests that as training progresses, the loss landscape becomes progressively sharper, making the model increasingly vulnerable to domain shifts and resulting in decreased performance on target domains. This is in line with our previous observations.", + "bbox": [ + 511, + 497, + 905, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this, we further investigate the relationship between model layers and the loss landscape. We apply the same magnitude of perturbations to the feature maps at stages 1-4 and measure the performance drops. The results in Fig.5d show that earlier stages exhibit larger performance drops under the same perturbation, which implies that shallow layers lead to sharper loss landscapes and are more vulnerable to domain shifts. This experiment provides quantitative validation for the intuitive observation from the low-level feature visualization in the previous section.", + "bbox": [ + 511, + 724, + 905, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 Please refer to the appendix for more details about the sharpness.", + "bbox": [ + 531, + 886, + 879, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4620", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/f37571de1d0e02774182f985fdc249cbb027bf79369f064f7d269f7820db704a.jpg", + "table_caption": [ + "Table 1. The model with fixed shallow layers performs better than the one with trained shallow layers, implying that shallow layers learn domain-specific information that negatively impacts generalization to the target domain." + ], + "table_footnote": [], + "table_body": "
MethodFSS-1000Deep globeISICChestX-rayAverage
train stage1, 2, 3, 478.8639.4435.7672.4956.64
fix stage1, train stage2, 3, 478.8839.9037.0072.1256.98
fix stage1, 2, train stage3, 478.9140.0035.4974.4457.21
", + "bbox": [ + 94, + 148, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6944e8d5bff1995093a9884b632c31c2c9f2f1669082b7f3bd4b184d1e138d6d.jpg", + "image_caption": [ + "Figure 6. The mIoU curves over epochs for two different training settings indicate that shallow layers are more vulnerable to domain shifts, leading to more severe performance degradation during source-domain training." + ], + "image_footnote": [], + "bbox": [ + 91, + 210, + 285, + 305 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f1afcf28068f52da55ec1ac2f926012cbd4f675c898488ed395f3e9af2f98afc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 210, + 480, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2.3 Further analysis: Low-level features absorb domain-specific information", + "text_level": 1, + "bbox": [ + 89, + 383, + 483, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To investigate the impact of shallow layers on model performance, we compare the results of fixing stage 1 and stage 2 using ImageNet [8] pre-trained weights with the results of training all stages, as shown in Tab. 1. We find that fixing stage 1 and stage 2 leads to the best performance.", + "bbox": [ + 89, + 431, + 483, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig 6, we further visualize the mIoU curves over epochs for two different training settings: fix stage 1 only, and fix stage 1, 2. The comparison revealed that when stage 2 is fixed, the model reaches the optimal mIoU at a later epoch, and as training progresses, the mIoU decreases more slowly.", + "bbox": [ + 89, + 510, + 483, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This suggests that shallow layers are vulnerable to domain shifts, therefore any slight absorption of domain information could harm target-domain generalization.", + "bbox": [ + 89, + 590, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 89, + 657, + 333, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above experiments, we make the interpretation as follows. The shallow layers of the model tend to learn domain-specific information, causing the low-level features to become increasingly domain-specific during training, gradually overfitting to the source domain. The overfitting sharpens the model's loss landscape and makes it more vulnerable to domain shifts, thereby reducing cross-domain performance and introducing significant mistakes in CDFSS. This problem is partially addressed by the early-stop strategy and freezing shallow layers.", + "bbox": [ + 89, + 684, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, merely stopping the training of shallow layers cannot fully unlock the power of source-domain training. Therefore, below we propose a novel method to handle the problem in shallow layers.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Method Overview", + "text_level": 1, + "bbox": [ + 513, + 114, + 687, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our core idea is that low-level features are vulnerable to domain shifts, which may lead to a sharp decrease in generalization with only slight absorption of source-domain information. Based on this, we designed two plug-and-play modules to operate in two phases: 1) During the source-domain training phase, the Low-level Enhancing Module enhances the robustness against domain shifts by synthesizing diverse domains and incorporating them into low-level features. 2) During the target-domain testing phase, the Low-level Calibration Module refines previous segmentation results by utilizing the low-level features from the target domain. The framework is illustrated in Fig. 7.", + "bbox": [ + 511, + 137, + 906, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Train-time Low-level Enhancing Module", + "text_level": 1, + "bbox": [ + 511, + 328, + 862, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As analyzed above, overfitting of low-level features on the source domain during training leads to a sharp loss landscape, making the model vulnerable to domain shifts. According to the previous Sharpness Aware Minimization (SAM) method [13, 46], one effective strategy to smooth loss landscapes is to introduce perturbations. Thus we transform the domain of the low-level support feature during training, which acts as a form of perturbation to reduce overfitting.", + "bbox": [ + 511, + 351, + 906, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous work [39] has shown that random convolution is a shape-preserving strategy to distort local textures, which we view to be suitable for the segmentation task. We take advantage of this property to generate diverse domains while retaining the content. Specifically, given support feature $F_{s} \\in \\mathbb{R}^{H \\times W \\times C}$ from the early layers of the encoder and a convolution layer with filters $\\Theta \\in \\mathbb{R}^{h \\times w \\times C \\times C}$ , where $H$ , $W$ and $C$ represent height, width and channels of the input feature, and $h$ and $w$ are the height and width of the random convolution filter. We sample the filter weights from $N(0, \\sigma^2)$ , where $\\sigma$ acts as a hyperparameter to control the perturbation magnitude. Then we get perturbed support feature $F_{s}'$ with transformed domain:", + "bbox": [ + 511, + 473, + 908, + 669 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s} ^ {\\prime} = F _ {s} * \\Theta , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 671, + 903, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although random convolution can generally preserve the shape and semantics of the feature, some edge details inevitably get blurred. Existing research [28] has shown that the Fast Fourier Transformation (FFT) decomposes a signal into amplitude spectrum and phase spectrum, with amplitude capturing the overall domain and texture information and phase representing shape, edge, and spatial structure. Thus, we apply FFT to both $F_{s}$ and $F_{s}^{\\prime}$ , decomposing them into phase spectrum $\\mathcal{P}$ , $\\mathcal{P}^{\\prime}$ and amplitude spectrum $\\mathcal{A}$ , $\\mathcal{A}^{\\prime}$ :", + "bbox": [ + 511, + 691, + 906, + 827 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} e ^ {i \\mathcal {P}} = F F T \\left(F _ {s}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 830, + 903, + 848 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} ^ {\\prime} e ^ {i \\mathcal {P} ^ {\\prime}} = F F T \\left(F _ {s} ^ {\\prime}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 849, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2The low-level support and query features in LEM and LCM are illustrated as images for better understanding.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c53872451b8557a2334c02b7e4c72ae6566938b49ceacdddaeafc7874a081cb3.jpg", + "image_caption": [ + "Figure 7. The overall architecture of our method in a 1-way 1-shot example. Our method consists of two modules: the Low-level Enhancing Module (LEM) and the Low-level Calibration Module (LCM). Support and query images are fed into a weight-shared backbone to extract features, and then processed by a comparison module to obtain a coarse segmentation mask. During source-domain training, LEM perturbs low-level support features for sharpness-aware minimization. During target-domain testing, LCM calibrates the coarse segmentation mask by directly supplementing collapsed low-level target-domain information." + ], + "image_footnote": [], + "bbox": [ + 187, + 89, + 803, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we recombine the phase from the original feature $F_{s}$ and the amplitude from the perturbed feature $F_{s}^{\\prime}$ , using the Inverse Fast Fourier Transform (IFFT) to generate transformed support feature:", + "bbox": [ + 89, + 402, + 483, + 463 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s} ^ {t} = I F F T \\left(\\mathcal {A} ^ {\\prime} e ^ {i \\mathcal {P}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 464, + 480, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, while transforming the domain, we achieve better preservation of the boundary shapes, which achieves domain-oriented perturbations on low-level features for sharpness-aware minimization. The transformed feature is then fed into the subsequent layers of the encoder.", + "bbox": [ + 89, + 484, + 483, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Test-time Low-level Calibration Module", + "text_level": 1, + "bbox": [ + 89, + 568, + 433, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After being processed by the encoder, the support and query features are fed into the comparison module, yielding a coarse score map $S \\in \\mathbb{R}^{H \\times W \\times 2}$ . Since low-level features may be collapsed on target domains due to the vulnerability to domain shifts, we directly use the low-level query feature $F_{q} \\in \\mathbb{R}^{H' \\times W' \\times C}$ to supplement such collapsed low-level target-domain information to calibrate the score map during testing on the target domain.", + "bbox": [ + 89, + 592, + 483, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we compute a confidence map $C$ by subtracting the background similarity from the foreground similarity in the score map:", + "bbox": [ + 89, + 713, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {i, j} = S _ {i, j, 1} - S _ {i, j, 0}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 758, + 480, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $S_{i,j,1}$ and $S_{i,j,0}$ denote the foreground and background similarity for the pixel at $(i,j)$ , respectively. The confidence map $C$ quantifies the likelihood of foreground at each pixel.", + "bbox": [ + 89, + 779, + 483, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we partition the confidence map into patches $\\{P_1, P_2, \\ldots, P_n\\}$ and select the Top-K patches with the highest average confidence. These patches are considered the most reliable foreground regions:", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {C} _ {p} = \\frac {1}{| P _ {p} |} \\sum_ {(i, j) \\in P _ {p}} C (i, j), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 398, + 903, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{P _ {k _ {1}}, P _ {k _ {2}}, \\dots , P _ {k _ {K}} \\right\\} = \\arg \\max _ {P _ {p}} \\bar {C} _ {p} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 436, + 903, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $|P_{p}|$ is the number of pixels in patch $P_{p}$ and $(i,j)$ are the pixel locations within the patch.", + "bbox": [ + 511, + 455, + 903, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each of the selected Top-K foreground patches, we locate the corresponding patch in the interpolated query feature map $F_{q}^{\\prime} \\in \\mathbb{R}^{H \\times W \\times C}$ as $\\{F_{k_1}, F_{k_2}, \\ldots, F_{k_K}\\}$ , and compute the cosine similarity between it and all other patches in the query feature map:", + "bbox": [ + 511, + 486, + 905, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} \\left(F _ {k _ {i}}, F _ {m}\\right) = \\frac {F _ {k _ {i}} \\cdot F _ {m}}{\\| F _ {k _ {i}} \\| \\| F _ {m} \\|} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 558, + 903, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the foreground similarity in the score map is updated using the top-K similarity maps:", + "bbox": [ + 511, + 590, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i, j, 1} ^ {\\prime} = S _ {i, j, 1} + \\sum_ {k = 1} ^ {K} w _ {k} \\cdot \\left(S i m _ {k} (i, j) - \\beta_ {k}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 619, + 903, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $w_{k}$ is a scaling factor and $\\beta_{k}$ is a bias term. This adjustment improves foreground detection by directly supplementing collapsed low-level target-domain information.", + "bbox": [ + 511, + 660, + 905, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the source-domain training phase, the model will be trained by minimizing the loss in Eq. 1. During the target-domain testing phase, the prediction will be obtained based on Eq. 2 and LCM.", + "bbox": [ + 511, + 705, + 905, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 777, + 645, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 801, + 614, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the settings in [20], we use PASCAL VOC 2012 [11] with SBD augmentation [14] as source domain for training. Then we evaluate the trained models on four target domains: FSS-1000 [22], Deep globe [7], ISIC [5, 36], and Chest X-ray [3, 18]. See Appendix for more details.", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ab0f1e8656c197d40a89316c0a1b56dfaf85408ac5abc3056fa9a5a2c646f866.jpg", + "image_caption": [ + "Figure 8. Qualitative results of our model for 1-shot setting." + ], + "image_footnote": [], + "bbox": [ + 169, + 88, + 397, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 297, + 307, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ ResNet-50 [15] with ImageNet [8] pretrained weights as our backbone. Following baseline SSP [12], we discard the last backbone stage and the last ReLU for better generalization. Furthermore, we integrated our approach into the transformer architecture by using ViT-B/16 [9] as the backbone, following FPTrans [41]. Consistent with FPTrans [41], we resize both support and query images to $480 \\times 480$ . During training on the source domain, we use SGD to optimize our model, with a momentum of 0.9 and a initial learning rate of 1e-3. We apply our LEM into different shallow layers of the backbone, more details can be seen in 4.4.4. The random convolution filter used in LEM has a kernel size of $3 \\times 3$ , and the standard deviation $\\sigma$ is set to be 0.1. During testing on the target domain, we apply our LCM to refine segmentation mask. The hyperparameters $K$ , $w$ and $\\beta$ are set to 3, 0.6 and 0.7, respectively.", + "bbox": [ + 89, + 321, + 483, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 574, + 454, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 2, we compare our method with CNN-based and ViT-based approaches, including traditional FSS methods and existing CDFSS methods. Our results illustrate a notable enhancement for both 1-shot and 5-shot tasks. Specifically, our method outperforms the state-of-the-art APSeg [16] by $3.71\\%$ and $5.34\\%$ in the 1-shot and 5-shot settings respectively, confirming the effectiveness of our strategy. Additionally, we apply our method on existing CDFSS baselines and observe improved performance. Please refer to the appendix for more details. We present qualitative results of our method in 1-way 1-shot segmentation in Fig.8.", + "bbox": [ + 89, + 598, + 483, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 776, + 243, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.1 Effectiveness of Each Component", + "text_level": 1, + "bbox": [ + 89, + 800, + 375, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate each proposed component on both CNN and ViT baselines to assess the effectiveness of our designs, including LEM and LCM. The results presented in Tab. 3 show that in the 1-shot setting, introducing LEM improved average MIoU by $1.14\\%$ for the CNN baseline and $0.89\\%$", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a27ecdc96cd5cad3ef4912c796eb82177c315688a64a7b8eb3286b2ad613e08e.jpg", + "image_caption": [ + "Figure 9. Applying LEM leads to reduced loss landscape sharpness and improved robustness to domain shifts, Fourier transformation further enhances these benefits." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 705, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c51bdc8bc22b9c61eae53e5bf479796c90a90e87d02da12c54c03d027ff1cec1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 90, + 900, + 183 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc344316c5d85db3bc6d869cb0cabccd83b357c0c48990d8dbf45bfa23818bc7.jpg", + "image_caption": [ + "Figure 10. (Left) The mIoU of Chest-X Ray with and without LEM, confirms that LEM leads to consistent and steady performance improvement as training advances. (Right) After incorporating our LEM module, training shallow layers can achieve performance comparable to fixing them." + ], + "image_footnote": [], + "bbox": [ + 514, + 239, + 707, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bdc1a79db33df37358ae6fd9ec547615aa566ded82163897d29a409b8f0dbf02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 239, + 900, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the ViT baseline, while adding LCM further increased it by $2.11\\%$ and $1.95\\%$ , respectively. These results clearly demonstrate the effectiveness of each component in enhancing performance.", + "bbox": [ + 511, + 414, + 905, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.2 Verification of LEM", + "text_level": 1, + "bbox": [ + 511, + 494, + 707, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LEM smooths loss landscapes. To assess the impact of LEM, we measured the loss landscape sharpness before and after its application. We introduced low-frequency noise on training data to simulate domain shifts in the representation space (i.e., pixels and features). A larger performance drop suggests a sharper landscapes. As illustrated in Fig.9, random convolution smooths the loss landscape compared with the baseline, while Fourier transformation further enhances this effect. These results confirm that LEM effectively smooths the loss landscape, which improves robustness against domain shifts.", + "bbox": [ + 511, + 517, + 906, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LEM ensures performance improves with training. We compared mIoU trends over training epochs with and without LEM in Fig.10 (left). In the baseline model, mIoU on the target domain declines as it increasingly absorbs source-domain information during training. Conversely, with LEM, mIoU not only outperforms the baseline at each epoch but also keeps increasing as training advances, indicating that the model focuses more on domain-invariant information.", + "bbox": [ + 511, + 684, + 905, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LEM prevents the shallow layers from absorbing excessive source-domain information. As previously analyzed, fixing stage 1 and stage 2 of the baseline resulted in the best performance, which suggests that shallow layers tend", + "bbox": [ + 511, + 805, + 905, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3We present the validation of CNN-based methods here, and similar trends hold for ViT-based methods. Please refer to the appendix for details.", + "bbox": [ + 511, + 875, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4623", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/99b3c0cee814c7470f0f5fd498ae9c363e8417c673d62c89c1f4455efe80db5c.jpg", + "table_caption": [ + "Table 2. MioU of 1-shot and 5-shot results compared with previous FSS and CD-FSS methods. All the methods are trained on PASCAL and tested on the CD-FSS. The best performance among all methods is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodMarkBackboneFSS-1000Deep globeISICChest X-rayAverage
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
Few-Shot Segmentation Methods
RePRI [2]CVPR-21Res-5070.9674.2325.0327.4123.2726.2365.0865.4846.0948.34
HSNet [26]ICCV-21Res-5077.5380.9929.6535.0831.2035.1051.8854.3647.5751.38
SSP [12]ECCV-22Res-5078.9180.5940.0048.6835.4945.8674.4474.2657.2162.35
FPTrans [41]NIPS-22ViT-base80.7483.6538.3649.3048.6560.3780.9282.9162.1769.06
PerSAM [42]ICLR-24ViT-base60.9266.5336.0840.6523.2725.3329.9530.0537.5640.64
Cross-Domain Few-Shot Segmentation Methods
PATNet [20]ECCV-22Res-5078.5981.2337.8942.9741.1653.5866.6170.2056.0661.99
APM [35]NIPS-24Res-5079.2981.8340.8644.9241.7151.1678.2582.8160.0365.18
ABCDFSS [17]CVPR-24Res-5074.6076.2042.6049.0045.7053.3079.8081.4060.6764.97
DRA [33]CVPR-24Res-5079.0580.4041.2950.1240.7748.8782.3582.3160.8665.42
APSeg [16]CVPR-24ViT-base79.7181.9035.9439.9845.4353.9884.1084.5061.3065.09
OursOursRes-5078.5180.6044.1049.6738.2147.0481.0282.7360.4665.01
OursOursViT-base81.0583.6942.1251.4852.9162.4383.9484.1265.0170.43
", + "bbox": [ + 140, + 121, + 856, + 359 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a32a74e1d664bc25dea4da807b6b9e252ae51b74f2f33ea363ad5fa95a2f1435.jpg", + "table_caption": [ + "Table 3. Ablation study on key components." + ], + "table_footnote": [], + "table_body": "
LEMLCMResNetViT
57.2162.17
✓58.3563.06
✓59.7864.39
✓✓60.4665.01
", + "bbox": [ + 189, + 393, + 387, + 469 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de25aa8f7ec34f7938fbcacba3aadb58b08e8edaba64d5ceb0c0cdf449a46cba.jpg", + "image_caption": [ + "Figure 11. (Left) The CKA similarity between source and target domain with and without LEM, verifies that LEM helps the model to focus more on domain-agnostic information. (Right) Feature map visualization of stage 1 and stage 4 at epoch 1 and epoch 20." + ], + "image_footnote": [], + "bbox": [ + 93, + 478, + 478, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to learn information that negatively impacts target-domain generalization. Fig.10 (right) shows that after incorporating our LEM, training stage 2 produced results comparable to those achieved by fixing stage 2. This indicates that LEM effectively prevents the shallow layers from absorbing excessive source-domain information during the training process.", + "bbox": [ + 89, + 642, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LEM encourages the model to focus more on domain-agnostic information. To assess how the domain distance between the source and target domains changes before and after applying LEM, we use CKA (Centered Kernel Alignment) similarity for measurement, following the approach in [6, 45, 47-49]. Specifically, we extract features from source domain and target domain images using a backbone network, then compute the CKA similarity of the final layer's feature by aligning the channel dimension. A higher CKA similarity indicates a smaller domain distance, suggesting the model retains less domain-specific information. As shown", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f165ed823f9ff79e6b064d4ca0bc7a3b2df2272a7393aa68036276151d149d8d.jpg", + "image_caption": [ + "Figure 12. LCM reduces the sharpness of the loss landscape and enhances robustness against domain shifts." + ], + "image_footnote": [], + "bbox": [ + 521, + 377, + 705, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a60ade4284c7152bf030a69f61935a0f5efd9337a3e42797bbf473022551ff2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 377, + 897, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Fig.11 (left), the average CKA similarity across four target domains increases after applying random convolution, and is further improved with the addition of the Fourier transform. This demonstrates that by perturbing the source domain, LEM encourages the model to capture domain-agnostic information during training.", + "bbox": [ + 511, + 540, + 906, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LEM enables the model to capture more useful low-level information in shallow layers. As shown in Fig.11 (right), LEM helps the model learn boundary information more effectively in stage 1, allowing it to focus more accurately on the foreground in stage 4. This leads to more precise activation regions in the later epochs of source domain training.", + "bbox": [ + 511, + 631, + 905, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.3 Verification of LCM", + "text_level": 1, + "bbox": [ + 511, + 739, + 709, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LCM smooths loss landscapes. To assess the impact of LCM in reducing sharpness and enhancing robustness against domain shifts, we superimpose styles from ISIC medical images and Deep globe remote sensing images onto source domain images to simulate domain shifts, with the level of style alteration serving as the perturbation level. As shown in Fig.12, the lowered performance drop validates that LCM smooths the loss landscape, thus improves robustness against domain shifts.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45ab3e2c57d2b66948055dfb24e5e26617f1f745898cf8b430bea4b10691e3ab.jpg", + "image_caption": [ + "Figure 13. The effect of LEM and LCM at different positions in the shallow layers. Block ID represents the sequence in which the module is applied: before stage 1, stage 1 block 1, stage 2 block 1, stage 2 block 2, and stage 2 block 3." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 277, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f9d947d9da0b81e8c741692fa32751b52dd14962647aade5e132ce3e2302419.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 89, + 464, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.4 Sensitivity Study of Hyper-parameters", + "text_level": 1, + "bbox": [ + 89, + 265, + 413, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig.13, we investigated the effects of using LEM and LCM at various positions in the shallow layers. Please refer to the appendix for more details.", + "bbox": [ + 89, + 289, + 482, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 89, + 349, + 230, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Few-Shot Segmentation", + "text_level": 1, + "bbox": [ + 89, + 375, + 308, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Few-Shot Segmentation (FSS) aims to segment novel classes in query images using only a few annotated support samples. Existing methods in FSS can be generally categorized into two types: prototype-based methods and matching-based methods. Motivated by PrototypicalNet[32], prototype-based methods [19, 21, 34, 38] generate prototypes from support images and perform segmentation by computing the similarity between query features and these support prototypes. On the other hand, matching-based methods[25, 26, 29, 40] focus on analyzing pixel-to-pixel dense correspondences between support and query features to avoid the loss of spatial structure inherent in prototype-based methods. However, these methods are limited to segmenting novel classes within a single domain and do not generalize effectively to unseen domains.", + "bbox": [ + 89, + 398, + 482, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Domain Generalization", + "text_level": 1, + "bbox": [ + 89, + 635, + 305, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Domain Generalization (DG) aims to train models that can generalize to diverse, unseen target domains, particularly when target domain data is unavailable during training, which aligns with the objective of CDFSS. Existing methods for domain generalization can be categorized into two types: methods that focus on learning domain-invariant feature representations across multiple source domains[4, 37], and methods that generate diverse samples through data or feature augmentation[10, 31]. However, most existing DG methods focus on generalizing the whole model, which is not suitable for direct use in CDFSS tasks. In contrast, our plug-and-play LEM module is compact, lightweight, and effective. In the appendix, we compare our LEM module with domain generalization methods to demonstrate the effectiveness of our approach as a novel sharpness-aware minimization method in addressing the CDFSS problem.", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Cross-domain Few-Shot Segmentation", + "text_level": 1, + "bbox": [ + 511, + 90, + 843, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike traditional FSS setting, CDFSS requires models to generalize to unseen target domains without accessing target data during training, making the task more realistic and complex. PATNet [20] introduces a transformation module that converts domain-specific features into domain-agnostic features, which can be finetuned with target domain data during the testing phase. ABCDFSS [17] proposes a test-time task adaptation method, utilizing consistency-based contrastive learning to prevent overfitting and improve class discriminability. DR-Adapter [33] introduces a small adapter for rectifying diverse target domain styles to the source domain. Compared with them, our approach proposes two simple plug-and-play modules that require no training and can be directly applied to boost performance in CDFSS. Besides, a few works [12, 27] are based on the self-support concept. Our LCM module follows a similar principle but offers two significant advantages: (1) While other methods use query prototypes to match query features, we directly use the low-level query features to adjust the query mask, preventing further loss of information. (2) LCM can be conveniently added during the target-domain testing phase, rather than during the entire training process.", + "bbox": [ + 511, + 113, + 906, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 458, + 633, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, We focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, the best segmentation performance is always achieved at the very early epochs, followed by a sharp decline as source-domain training progresses. We delve into this phenomenon for an interpretation, and find low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS. Based on the interpretation, we further propose a method that includes two plug-and-play modules: one to flatten the loss landscapes for low-level features during source-domain training, and the other to directly supplement target-domain information to the model during target-domain testing. Extensive experiments on four CDFSS benchmarks validate our rationale and effectiveness.", + "bbox": [ + 511, + 484, + 906, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 723, + 669, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research and Development Program of China under grant 2024YFC3307900; the National Natural Science Foundation of China under grants 62206102, 62436003, 62376103 and 62302184; Major Science and Technology Project of Hubei Province under grant 2024BAA008; Hubei Science and Technology Talent Service Project under grant 2024DJC078; and Ant Group through CCF-Ant Research Fund. The computation is completed in the HPC Platform of Huazhong University of Science and Technology.", + "bbox": [ + 511, + 747, + 905, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11700-11709, 2019. 1", + "[2] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13979-13988, 2021. 7", + "[3] Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE Transactions on Medical Imaging, 33(2):577-590, 2013. 5", + "[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2229-2238, 2019. 8", + "[5] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1, 5", + "[6] MohammadReza Davari, Stefan Horoi, Amine Natik, Guillaume Lajoie, Guy Wolf, and Eugene Belilovsky. Reliability of cka as a similarity measure in deep learning. arXiv preprint arXiv:2210.16156, 2022. 7", + "[7] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deep globe 2018: A challenge to parse the earth through satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 172-181, 2018. 1, 5", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 4, 6", + "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6", + "[10] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 200-216. Springer, 2020. 8", + "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2010. 1, 5" + ], + "bbox": [ + 94, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701–719, 2022. 2, 6, 7, 8", + "[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 2, 3, 4", + "[14] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In Proceedings of the IEEE International Conference on Computer Vision, pages 991-998. IEEE, 2011. 5", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2, 6", + "[16] Weizhao He, Yang Zhang, Wei Zhuo, Linlin Shen, Jiaqi Yang, Songhe Deng, and Liang Sun. Apseg: Auto-prompt network for cross-domain few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23762-23772, 2024. 1, 6, 7", + "[17] Jonas Herzog. Adapt before comparison: A new perspective on cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23605-23615, 2024. 1, 7, 8", + "[18] Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les Folio, Jenifer Siegelman, Fiona Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K Singh, Sameer Antani, et al. Automatic tuberculosis screening using chest radiographs. IEEE Transactions on Medical Imaging, 33(2):233-245, 2013. 5", + "[19] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8057-8067, 2022. 8", + "[20] Shuo Lei, Xuchao Zhang, Jianfeng He, Fanglan Chen, Bowen Du, and Chang-Tien Lu. Cross-domain few-shot semantic segmentation. In European Conference on Computer Vision, pages 73-90. Springer, 2022. 1, 2, 5, 7, 8", + "[21] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 8", + "[22] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2869–2878, 2020. 1, 5", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4626", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1", + "[25] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8741-8750, 2021. 8", + "[26] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 7, 8", + "[27] Jiahao Nie, Yun Xing, Gongjie Zhang, Pei Yan, Aoran Xiao, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Cross-domain few-shot segmentation via iterative support-query correspondence mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3380-3390, 2024. 8", + "[28] Alan V Oppenheim and Jae S Lim. The importance of phase in signals. Proceedings of the IEEE, 69(5):529-541, 1981. 4", + "[29] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023. 8", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1", + "[31] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. arXiv preprint arXiv:1804.10745, 2018. 8", + "[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 8", + "[33] Jiapeng Su, Qi Fan, Wenjie Pei, Guangming Lu, and Fanglin Chen. Domain-rectifying adapter for cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24036–24045, 2024. 1, 7, 8", + "[34] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020. 8", + "[35] Jintao Tong, Yixiong Zou, Yuhua Li, and Ruixuan Li. Lightweight frequency masker for cross-domain few-shot semantic segmentation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 7", + "[36] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data, 5(1):1-9, 2018. 1, 5" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. arXiv preprint arXiv:1903.06256, 2019. 8", + "[38] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 8", + "[39] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, 2021. 4", + "[40] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. Advances in Neural Information Processing Systems, 34:21984-21996, 2021. 8", + "[41] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in neural information processing systems, 35:6575-6588, 2022. 6, 7", + "[42] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Yu Qiao, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. In The Twelfth International Conference on Learning Representations, 2024. 7", + "[43] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1", + "[44] Yixiong Zou, Shanghang Zhang, Jianpeng Yu, Yonghong Tian, and José MF Moura. Revisiting mid-level patterns for cross-domain few-shot recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 741-749, 2021. 2", + "[45] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 7", + "[46] Yixiong Zou, Yicong Liu, Yiman Hu, Yuhua Li, and Ruixuan Li. Flatten long-range loss landscapes for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23575-23584, 2024. 2, 3, 4", + "[47] Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. Attention temperature matters in vit-based cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:116332-116354, 2024. 7", + "[48] Yixiong Zou, Shuai Yi, Yuhua Li, and Ruixuan Li. A closer look at the clf token for cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:85523-85545, 2024.", + "[49] Yixiong Zou, Shanghang Zhang, Haichen Zhou, Yuhua Li, and Ruixuan Li. Compositional few-shot class-incremental learning. arXiv preprint arXiv:2405.17022, 2024. 7" + ], + "bbox": [ + 516, + 90, + 906, + 869 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_model.json b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6dbbf05fc9962bf5a7076ea9106e73cb8f952a92 --- /dev/null +++ b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_model.json @@ -0,0 +1,2497 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.887, + 0.154 + ], + "angle": 0, + "content": "The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.181, + 0.877, + 0.235 + ], + "angle": 0, + "content": "Yuhan Liu, Yixiong Zou*, Yuhua Li, Ruixuan Li \nSchool of Computer Science and Technology, Huazhong University of Science and Technology \n{yuhan.liu, yixiongz, idcliyuhua, rxli}@hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.665 + ], + "angle": 0, + "content": "Cross-Domain Few-Shot Segmentation (CDFSS) is proposed to transfer the pixel-level segmentation capabilities learned from large-scale source-domain datasets to downstream target-domain datasets, with only a few annotated images per class. In this paper, we focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, particularly those distant from the source domain, segmentation performance peaks at the very early epochs, and declines sharply as the source-domain training proceeds. We delve into this phenomenon for an interpretation: low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS. Based on this phenomenon and interpretation, we further propose a method that includes two plug-and-play modules: one to flatten the loss landscapes for low-level features during source-domain training as a novel sharpness-aware minimization method, and the other to directly supplement target-domain information to the model during target-domain testing by low-level-based calibration. Extensive experiments on four target datasets validate our rationale and demonstrate that our method surpasses the state-of-the-art method in CDFSS significantly by \\(3.71\\%\\) and \\(5.34\\%\\) average MIoU in 1-shot and 5-shot scenarios, respectively." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.696, + 0.222, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Current deep neural networks [24, 30, 43] have achieved remarkable success in semantic segmentation, but their performance heavily relies on large-scale annotated datasets [1, 23]. However, the process of annotating data, particularly for dense pixel-wise tasks like semantic segmentation, is labor-intensive and time-consuming. To address this issue, Few-Shot Segmentation (FSS) has been introduced, aiming to generate pixel-level predictions for unseen categories with only a few labeled samples. Typically, the model is pretrained on a large-scale dataset of base classes and then trans" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.269, + 0.903, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.357, + 0.707, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.442, + 0.679, + 0.458 + ], + "angle": 0, + "content": "(b) Previous CDFSS methods Epoch vs. MoU" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.358, + 0.902, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.441, + 0.848, + 0.457 + ], + "angle": 0, + "content": "(c) Our method Epochs vs. MoU" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.558 + ], + "angle": 0, + "content": "Figure 1. (a) In CDFSS tasks, the training (source) and testing (target) datasets belong to distinct domains, with categories in the testing dataset being unseen during training. (b) Previous CDFSS methods show a decreasing trend of mIoU as the source-domain training proceeds, even at very early epochs for distant domains. (c) Our method can effectively prevent the model from performance decline after early epochs and achieve higher performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.658 + ], + "angle": 0, + "content": "ferred to unseen novel classes. However, novel classes might not be in the same domain as base classes [5, 7, 11, 22, 36], a more realistic approach is to consider domain gaps between the base (source domain) and novel (target domain) classes, which gives rise to the Cross-Domain Few-Shot Segmentation (CDFSS) task [20]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Although extensive works [16, 17, 20, 33] have been developed, a well-observed phenomenon is still not handled: for target domains, especially those distant from the source domain, the best performance is always achieved at the very early epochs or even the first epoch. As shown in Fig. 1b, the performance decreases sharply as the source-domain training proceeds, where the 20th epoch's MIoU is even lower than that of the first epoch. Although the early stop can handle this problem, the goal of the source-domain training is to provide a generalizable model for all target domains [20]. Therefore, it may be inappropriate to try different early stops for specific domains. In this paper, we aim to delve into this well-observed but under-explored phenomenon and handle it based on our interpretations of it, as shown in Fig. 1c." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To delve into this phenomenon, we first visualize the predictions of the 1st and 20th epochs in Fig. 2. We observe that" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4618" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.479, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.196, + 0.483, + 0.238 + ], + "angle": 0, + "content": "Figure 2. The visualization of predictions at the 1st and 20th epoch further indicate that CDFSS models may not acquire meaningful information for target domains during source-domain training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.248, + 0.483, + 0.52 + ], + "angle": 0, + "content": "the 20th epoch's model shows many fundamental errors in target domains, i.e., CDFSS models fail to identify the foreground but focus on entirely wrong regions. This suggests that the CDFSS models may not have acquired meaningful information for target domains during source-domain training, resulting in almost random predictions during testing. We then visualize feature maps at different network stages, and observe this problem occurs even at the shallow layers, and deeper layers produce wrong feature maps with wrong inputs from shallow layers. Quantitatively, previous works [13, 46] suggest that the challenge posed by domain shift can be analyzed through the flatness of the loss landscape. We use this as an entry point to explore the connection between shallow layers and early stops. We find that low-level features, although simpler and previously believed to be more transferable [44], are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training. That is, the devil lies in the low-level features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.733 + ], + "angle": 0, + "content": "Based on this phenomenon and interpretation, we further propose a method for the CDFSS task. During the source-domain training, we propose a novel sharpness-aware minimization method to flatten the loss landscapes for low-level features, which is achieved by a shape-preserving perturbation module with random convolutions. During the target-domain phase, since low-level features can hardly capture the target-domain information due to its vulnerability to domain shifts, we directly supplement such information to the model with a low-level-based calibration module, which effectively prevents the model from fundamental errors caused by collapsed low-level features. Based on our methods, we can effectively prevent the model from dropping performances after early epochs and achieve higher performance (Fig. 1c)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.734, + 0.418, + 0.748 + ], + "angle": 0, + "content": "In summary, our contributions can be listed as" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.839 + ], + "angle": 0, + "content": "- We focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, particularly those distant from the source domain, the best segmentation performance is always achieved at the very early epochs, followed by a sharp decline as source-domain training progresses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.9 + ], + "angle": 0, + "content": "- We delve into this phenomenon for an interpretation: low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- Building on this interpretation, we propose a method that includes two plug-and-play modules: one act as a novel sharpness-aware minimization method to flatten the loss landscapes of low-level features during source-domain training, while the other directly supplement target-domain information to the model during target-domain testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- Extensive experiments on four target datasets validate the rationale of our interpretation and method, showing its superiority over current state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.241, + 0.659, + 0.258 + ], + "angle": 0, + "content": "2. Interpretation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.266, + 0.655, + 0.28 + ], + "angle": 0, + "content": "2.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.906, + 0.409 + ], + "angle": 0, + "content": "Cross-Domain Few-Shot Segmentation (CDFSS) aims to transfer the segmentation capabilities learned from the source domain to target domains, with only a few annotated images per class. The source domain \\(\\mathcal{D}_s = (\\mathcal{X}_s,\\mathcal{Y}_s)\\) and target domain \\(\\mathcal{D}_t = (\\mathcal{X}_t,\\mathcal{Y}_t)\\) are defined by different input distributions and disjoint label spaces, i.e., \\(\\mathcal{X}_s\\neq \\mathcal{X}_t\\) \\(\\mathcal{Y}_s\\cap \\mathcal{Y}_t = \\emptyset\\), where \\(\\mathcal{X}\\) refers to input distribution and \\(\\mathcal{Y}\\) refers to the label space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.518 + ], + "angle": 0, + "content": "In this work, we adopt the meta-learning episodic paradigm following [20]. Specifically, both \\(\\mathcal{D}_s\\) and \\(\\mathcal{D}_t\\) consist of several episodes. Each episode is constructed by a support set \\(S = \\{(I_s^i,M_s^i)\\}_{i = 1}^K\\) with \\(K\\) training samples and a query set \\(Q = \\{(I_q,M_q)\\}\\), where \\(I\\) denote the image and \\(M\\) denote the pixel labels. Within each episode, the model leverages \\(\\{I_s,M_s\\}\\) and \\(I_{q}\\) to predict the query mask \\(M_{q}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.905, + 0.563 + ], + "angle": 0, + "content": "The model is trained on the source domain support sets and query set \\(\\{S^s, Q^s\\}\\) by minimizing the binary cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.574, + 0.905, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = B C E \\left(F \\left(I _ {s} ^ {s}, I _ {q} ^ {s}, M _ {s} ^ {s}\\right), M _ {q} ^ {s}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.905, + 0.65 + ], + "angle": 0, + "content": "where \\( F\\left(I_s^s,I_q^s,M_s^s\\right) = h\\left(f\\left(I_s^s,I_q^s\\right),M_s^s\\right) \\) outputs the segmentation score map of \\( I_{q}^{s} \\), which is composed of a feature extractor \\( f(\\cdot) \\) and a comparison module \\( h(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.905, + 0.682 + ], + "angle": 0, + "content": "Then the model will be evaluated on the target domain query set \\( Q^t \\), the prediction for \\( I_q^t \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.693, + 0.905, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\hat {M} _ {q} ^ {t} = \\arg \\max F \\left(I _ {s} ^ {t}, I _ {q} ^ {t}, M _ {s} ^ {t}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.905, + 0.756 + ], + "angle": 0, + "content": "Finally, the evaluation can be conducted based on the comparison between \\(\\hat{M}_q^t\\) and the real label \\(M_q^t\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.763, + 0.905, + 0.794 + ], + "angle": 0, + "content": "2.2. Delve into the Early Stop in Cross-Domain Few-Shot Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.905, + 0.831 + ], + "angle": 0, + "content": "2.2.1 Intuitive observation: The devil is in the low-level features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Given our suspicion that the model fails to learn useful information from the source domain for target domains, we start by visualizing the feature maps. Using ResNet-50 [15] as the backbone following [12], we visualize the feature maps" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4619" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.088, + 0.445, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.375 + ], + "angle": 0, + "content": "Figure 3. Feature maps from stages 1 to 4 for both the source domain and target domains. The noticeable contrast in the feature maps at stage 1 between source and target domains indicates the limited performance of CDFSS models stems from shallow layers." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.381, + 0.48, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.52 + ], + "angle": 0, + "content": "Figure 4. Feature maps from stages 1 to 4 for epoch 1 and epoch 20. Epoch 1 in stage 1 shows more distinguishable activations than epoch 20, indicating that low-level features gradually incorporate incorrect information as training progresses." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.526, + 0.483, + 0.692 + ], + "angle": 0, + "content": "of stages 1 to 4 for both the source domain and the target domains, as shown in Fig. 3. We observe that on source-domain samples, the model learns boundary information effectively in stage 1, enabling it to accurately focus on the foreground in stage 4. However, on target domains especially those distant from the source domain, even though the foreground is clear and prominent, the model still fails to learn anything meaningful in stages 1 and 2. Consequently, in stage 4, the model focuses on completely incorrect areas. This indicates that the model's bad performance may originate from its shallow layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.813 + ], + "angle": 0, + "content": "In Fig.4, we further visualize feature maps of each stage at the 1st and 20th epoch, respectively. We find that the shallow layers show more distinguishable activations at epoch 1 than at epoch 20, which aligns with our observation of higher accuracy at epoch 1. Consequently, we conclude that the poor performance of CDFSS models may be attributed to shallow layers that fail to capture useful low-level information, a trend that intensifies during source-domain training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.483, + 0.863 + ], + "angle": 0, + "content": "2.2.2 Quantitative verification: Low-level features lead to a sharper loss landscape" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Previous work [13, 46] has demonstrated that the sharpness of the loss landscape can serve as a tool to analyze cross" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.132, + 0.538, + 0.143 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.089, + 0.885, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.192, + 0.539, + 0.203 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.184, + 0.895, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.256, + 0.537, + 0.267 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.226, + 0.712, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.256, + 0.744, + 0.268 + ], + "angle": 0, + "content": "(d) 0.3" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.226, + 0.899, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.33, + 0.906, + 0.427 + ], + "angle": 0, + "content": "Figure 5. (a) A sharp minimum in the landscape corresponds to a representation that is highly sensitive to data shifts. (b) Examples of pixel perturbation applied to the images. (c) As training proceeds, the loss landscape becomes progressively sharper. (d) Perturbing shallow layers leads to much sharper loss landscapes, indicating shallow layers are the cause of the sharp loss landscape and the increased sensitivity to domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.905, + 0.498 + ], + "angle": 0, + "content": "domain issues. As illustrated in Fig. 5a, a smoother loss landscape enables the model to better handle domain shifts1. Therefore, we measure the sharpness of the source-domain-trained models against different epochs and stages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.724 + ], + "angle": 0, + "content": "We first measure the sharpness of the loss landscape across different epochs. Following the method in [46], we apply pixel perturbations to the training data. Examples of these perturbations are shown in Fig. 5b, while the magnitude of performance drop is depicted in Fig.5c. A larger performance drop under the same perturbation indicates a sharper loss landscape. As observed, the performance of models across all epochs drops when exposed to perturbations, but the model from epoch 1 experiences a smaller drop than that of epoch 10, which in turn performs better than epoch 20. This suggests that as training progresses, the loss landscape becomes progressively sharper, making the model increasingly vulnerable to domain shifts and resulting in decreased performance on target domains. This is in line with our previous observations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.906, + 0.876 + ], + "angle": 0, + "content": "Based on this, we further investigate the relationship between model layers and the loss landscape. We apply the same magnitude of perturbations to the feature maps at stages 1-4 and measure the performance drops. The results in Fig.5d show that earlier stages exhibit larger performance drops under the same perturbation, which implies that shallow layers lead to sharper loss landscapes and are more vulnerable to domain shifts. This experiment provides quantitative validation for the intuitive observation from the low-level feature visualization in the previous section." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.88, + 0.901 + ], + "angle": 0, + "content": "1 Please refer to the appendix for more details about the sharpness." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4620" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.146 + ], + "angle": 0, + "content": "Table 1. The model with fixed shallow layers performs better than the one with trained shallow layers, implying that shallow layers learn domain-specific information that negatively impacts generalization to the target domain." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.15, + 0.485, + 0.198 + ], + "angle": 0, + "content": "
MethodFSS-1000Deep globeISICChestX-rayAverage
train stage1, 2, 3, 478.8639.4435.7672.4956.64
fix stage1, train stage2, 3, 478.8839.9037.0072.1256.98
fix stage1, 2, train stage3, 478.9140.0035.4974.4457.21
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.211, + 0.287, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.211, + 0.482, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.309, + 0.485, + 0.366 + ], + "angle": 0, + "content": "Figure 6. The mIoU curves over epochs for two different training settings indicate that shallow layers are more vulnerable to domain shifts, leading to more severe performance degradation during source-domain training." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.416 + ], + "angle": 0, + "content": "2.2.3 Further analysis: Low-level features absorb domain-specific information" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.508 + ], + "angle": 0, + "content": "To investigate the impact of shallow layers on model performance, we compare the results of fixing stage 1 and stage 2 using ImageNet [8] pre-trained weights with the results of training all stages, as shown in Tab. 1. We find that fixing stage 1 and stage 2 leads to the best performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.589 + ], + "angle": 0, + "content": "In Fig 6, we further visualize the mIoU curves over epochs for two different training settings: fix stage 1 only, and fix stage 1, 2. The comparison revealed that when stage 2 is fixed, the model reaches the optimal mIoU at a later epoch, and as training progresses, the mIoU decreases more slowly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.639 + ], + "angle": 0, + "content": "This suggests that shallow layers are vulnerable to domain shifts, therefore any slight absorption of domain information could harm target-domain generalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.334, + 0.673 + ], + "angle": 0, + "content": "2.3. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Based on the above experiments, we make the interpretation as follows. The shallow layers of the model tend to learn domain-specific information, causing the low-level features to become increasingly domain-specific during training, gradually overfitting to the source domain. The overfitting sharpens the model's loss landscape and makes it more vulnerable to domain shifts, thereby reducing cross-domain performance and introducing significant mistakes in CDFSS. This problem is partially addressed by the early-stop strategy and freezing shallow layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "However, merely stopping the training of shallow layers cannot fully unlock the power of source-domain training. Therefore, below we propose a novel method to handle the problem in shallow layers." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.689, + 0.13 + ], + "angle": 0, + "content": "3.1. Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Our core idea is that low-level features are vulnerable to domain shifts, which may lead to a sharp decrease in generalization with only slight absorption of source-domain information. Based on this, we designed two plug-and-play modules to operate in two phases: 1) During the source-domain training phase, the Low-level Enhancing Module enhances the robustness against domain shifts by synthesizing diverse domains and incorporating them into low-level features. 2) During the target-domain testing phase, the Low-level Calibration Module refines previous segmentation results by utilizing the low-level features from the target domain. The framework is illustrated in Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.329, + 0.863, + 0.345 + ], + "angle": 0, + "content": "3.2. Train-time Low-level Enhancing Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.473 + ], + "angle": 0, + "content": "As analyzed above, overfitting of low-level features on the source domain during training leads to a sharp loss landscape, making the model vulnerable to domain shifts. According to the previous Sharpness Aware Minimization (SAM) method [13, 46], one effective strategy to smooth loss landscapes is to introduce perturbations. Thus we transform the domain of the low-level support feature during training, which acts as a form of perturbation to reduce overfitting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.474, + 0.909, + 0.67 + ], + "angle": 0, + "content": "Previous work [39] has shown that random convolution is a shape-preserving strategy to distort local textures, which we view to be suitable for the segmentation task. We take advantage of this property to generate diverse domains while retaining the content. Specifically, given support feature \\( F_{s} \\in \\mathbb{R}^{H \\times W \\times C} \\) from the early layers of the encoder and a convolution layer with filters \\( \\Theta \\in \\mathbb{R}^{h \\times w \\times C \\times C} \\), where \\( H \\), \\( W \\) and \\( C \\) represent height, width and channels of the input feature, and \\( h \\) and \\( w \\) are the height and width of the random convolution filter. We sample the filter weights from \\( N(0, \\sigma^2) \\), where \\( \\sigma \\) acts as a hyperparameter to control the perturbation magnitude. Then we get perturbed support feature \\( F_{s}' \\) with transformed domain:" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.672, + 0.905, + 0.689 + ], + "angle": 0, + "content": "\\[\nF _ {s} ^ {\\prime} = F _ {s} * \\Theta , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.907, + 0.828 + ], + "angle": 0, + "content": "Although random convolution can generally preserve the shape and semantics of the feature, some edge details inevitably get blurred. Existing research [28] has shown that the Fast Fourier Transformation (FFT) decomposes a signal into amplitude spectrum and phase spectrum, with amplitude capturing the overall domain and texture information and phase representing shape, edge, and spatial structure. Thus, we apply FFT to both \\( F_{s} \\) and \\( F_{s}^{\\prime} \\), decomposing them into phase spectrum \\( \\mathcal{P} \\), \\( \\mathcal{P}^{\\prime} \\) and amplitude spectrum \\( \\mathcal{A} \\), \\( \\mathcal{A}^{\\prime} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.831, + 0.905, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} e ^ {i \\mathcal {P}} = F F T \\left(F _ {s}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.851, + 0.905, + 0.871 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} ^ {\\prime} e ^ {i \\mathcal {P} ^ {\\prime}} = F F T \\left(F _ {s} ^ {\\prime}\\right), \\tag {5}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.901 + ], + "angle": 0, + "content": "2The low-level support and query features in LEM and LCM are illustrated as images for better understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4621" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.189, + 0.09, + 0.805, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.325, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Figure 7. The overall architecture of our method in a 1-way 1-shot example. Our method consists of two modules: the Low-level Enhancing Module (LEM) and the Low-level Calibration Module (LCM). Support and query images are fed into a weight-shared backbone to extract features, and then processed by a comparison module to obtain a coarse segmentation mask. During source-domain training, LEM perturbs low-level support features for sharpness-aware minimization. During target-domain testing, LCM calibrates the coarse segmentation mask by directly supplementing collapsed low-level target-domain information." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.403, + 0.484, + 0.464 + ], + "angle": 0, + "content": "Then we recombine the phase from the original feature \\( F_{s} \\) and the amplitude from the perturbed feature \\( F_{s}^{\\prime} \\), using the Inverse Fast Fourier Transform (IFFT) to generate transformed support feature:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.465, + 0.482, + 0.484 + ], + "angle": 0, + "content": "\\[\nF _ {s} ^ {t} = I F F T \\left(\\mathcal {A} ^ {\\prime} e ^ {i \\mathcal {P}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.561 + ], + "angle": 0, + "content": "Therefore, while transforming the domain, we achieve better preservation of the boundary shapes, which achieves domain-oriented perturbations on low-level features for sharpness-aware minimization. The transformed feature is then fed into the subsequent layers of the encoder." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.569, + 0.434, + 0.584 + ], + "angle": 0, + "content": "3.3. Test-time Low-level Calibration Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.713 + ], + "angle": 0, + "content": "After being processed by the encoder, the support and query features are fed into the comparison module, yielding a coarse score map \\( S \\in \\mathbb{R}^{H \\times W \\times 2} \\). Since low-level features may be collapsed on target domains due to the vulnerability to domain shifts, we directly use the low-level query feature \\( F_{q} \\in \\mathbb{R}^{H' \\times W' \\times C} \\) to supplement such collapsed low-level target-domain information to calibrate the score map during testing on the target domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.759 + ], + "angle": 0, + "content": "First, we compute a confidence map \\( C \\) by subtracting the background similarity from the foreground similarity in the score map:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.759, + 0.482, + 0.776 + ], + "angle": 0, + "content": "\\[\nC _ {i, j} = S _ {i, j, 1} - S _ {i, j, 0}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Here, \\( S_{i,j,1} \\) and \\( S_{i,j,0} \\) denote the foreground and background similarity for the pixel at \\( (i,j) \\), respectively. The confidence map \\( C \\) quantifies the likelihood of foreground at each pixel." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Then we partition the confidence map into patches \\(\\{P_1, P_2, \\ldots, P_n\\}\\) and select the Top-K patches with the highest average confidence. These patches are considered the most reliable foreground regions:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.4, + 0.905, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\bar {C} _ {p} = \\frac {1}{| P _ {p} |} \\sum_ {(i, j) \\in P _ {p}} C (i, j), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.437, + 0.905, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\left\\{P _ {k _ {1}}, P _ {k _ {2}}, \\dots , P _ {k _ {K}} \\right\\} = \\arg \\max _ {P _ {p}} \\bar {C} _ {p} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.904, + 0.487 + ], + "angle": 0, + "content": "Where \\( |P_{p}| \\) is the number of pixels in patch \\( P_{p} \\) and \\( (i,j) \\) are the pixel locations within the patch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.562 + ], + "angle": 0, + "content": "For each of the selected Top-K foreground patches, we locate the corresponding patch in the interpolated query feature map \\( F_{q}^{\\prime} \\in \\mathbb{R}^{H \\times W \\times C} \\) as \\( \\{F_{k_1}, F_{k_2}, \\ldots, F_{k_K}\\} \\), and compute the cosine similarity between it and all other patches in the query feature map:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.559, + 0.905, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} \\left(F _ {k _ {i}}, F _ {m}\\right) = \\frac {F _ {k _ {i}} \\cdot F _ {m}}{\\| F _ {k _ {i}} \\| \\| F _ {m} \\|} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.591, + 0.906, + 0.621 + ], + "angle": 0, + "content": "Finally, the foreground similarity in the score map is updated using the top-K similarity maps:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.621, + 0.905, + 0.661 + ], + "angle": 0, + "content": "\\[\nS _ {i, j, 1} ^ {\\prime} = S _ {i, j, 1} + \\sum_ {k = 1} ^ {K} w _ {k} \\cdot \\left(S i m _ {k} (i, j) - \\beta_ {k}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.707 + ], + "angle": 0, + "content": "Here, \\(w_{k}\\) is a scaling factor and \\(\\beta_{k}\\) is a bias term. This adjustment improves foreground detection by directly supplementing collapsed low-level target-domain information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.906, + 0.767 + ], + "angle": 0, + "content": "During the source-domain training phase, the model will be trained by minimizing the loss in Eq. 1. During the target-domain testing phase, the prediction will be obtained based on Eq. 2 and LCM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.778, + 0.646, + 0.795 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.616, + 0.816 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Following the settings in [20], we use PASCAL VOC 2012 [11] with SBD augmentation [14] as source domain for training. Then we evaluate the trained models on four target domains: FSS-1000 [22], Deep globe [7], ISIC [5, 36], and Chest X-ray [3, 18]. See Appendix for more details." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.089, + 0.398, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.271, + 0.465, + 0.285 + ], + "angle": 0, + "content": "Figure 8. Qualitative results of our model for 1-shot setting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.298, + 0.308, + 0.314 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.485, + 0.565 + ], + "angle": 0, + "content": "We employ ResNet-50 [15] with ImageNet [8] pretrained weights as our backbone. Following baseline SSP [12], we discard the last backbone stage and the last ReLU for better generalization. Furthermore, we integrated our approach into the transformer architecture by using ViT-B/16 [9] as the backbone, following FPTrans [41]. Consistent with FPTrans [41], we resize both support and query images to \\( 480 \\times 480 \\). During training on the source domain, we use SGD to optimize our model, with a momentum of 0.9 and a initial learning rate of 1e-3. We apply our LEM into different shallow layers of the backbone, more details can be seen in 4.4.4. The random convolution filter used in LEM has a kernel size of \\( 3 \\times 3 \\), and the standard deviation \\( \\sigma \\) is set to be 0.1. During testing on the target domain, we apply our LCM to refine segmentation mask. The hyperparameters \\( K \\), \\( w \\) and \\( \\beta \\) are set to 3, 0.6 and 0.7, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.455, + 0.591 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.766 + ], + "angle": 0, + "content": "In Tab. 2, we compare our method with CNN-based and ViT-based approaches, including traditional FSS methods and existing CDFSS methods. Our results illustrate a notable enhancement for both 1-shot and 5-shot tasks. Specifically, our method outperforms the state-of-the-art APSeg [16] by \\(3.71\\%\\) and \\(5.34\\%\\) in the 1-shot and 5-shot settings respectively, confirming the effectiveness of our strategy. Additionally, we apply our method on existing CDFSS baselines and observe improved performance. Please refer to the appendix for more details. We present qualitative results of our method in 1-way 1-shot segmentation in Fig.8." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.245, + 0.793 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.377, + 0.815 + ], + "angle": 0, + "content": "4.4.1 Effectiveness of Each Component" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate each proposed component on both CNN and ViT baselines to assess the effectiveness of our designs, including LEM and LCM. The results presented in Tab. 3 show that in the 1-shot setting, introducing LEM improved average MIoU by \\(1.14\\%\\) for the CNN baseline and \\(0.89\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.707, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.091, + 0.901, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.193, + 0.906, + 0.233 + ], + "angle": 0, + "content": "Figure 9. Applying LEM leads to reduced loss landscape sharpness and improved robustness to domain shifts, Fourier transformation further enhances these benefits." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.241, + 0.709, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.241, + 0.901, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.337, + 0.906, + 0.406 + ], + "angle": 0, + "content": "Figure 10. (Left) The mIoU of Chest-X Ray with and without LEM, confirms that LEM leads to consistent and steady performance improvement as training advances. (Right) After incorporating our LEM module, training shallow layers can achieve performance comparable to fixing them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.906, + 0.476 + ], + "angle": 0, + "content": "for the ViT baseline, while adding LCM further increased it by \\(2.11\\%\\) and \\(1.95\\%\\), respectively. These results clearly demonstrate the effectiveness of each component in enhancing performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.495, + 0.708, + 0.509 + ], + "angle": 0, + "content": "4.4.2 Verification of LEM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.684 + ], + "angle": 0, + "content": "LEM smooths loss landscapes. To assess the impact of LEM, we measured the loss landscape sharpness before and after its application. We introduced low-frequency noise on training data to simulate domain shifts in the representation space (i.e., pixels and features). A larger performance drop suggests a sharper landscapes. As illustrated in Fig.9, random convolution smooths the loss landscape compared with the baseline, while Fourier transformation further enhances this effect. These results confirm that LEM effectively smooths the loss landscape, which improves robustness against domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.906, + 0.805 + ], + "angle": 0, + "content": "LEM ensures performance improves with training. We compared mIoU trends over training epochs with and without LEM in Fig.10 (left). In the baseline model, mIoU on the target domain declines as it increasingly absorbs source-domain information during training. Conversely, with LEM, mIoU not only outperforms the baseline at each epoch but also keeps increasing as training advances, indicating that the model focuses more on domain-invariant information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.806, + 0.906, + 0.867 + ], + "angle": 0, + "content": "LEM prevents the shallow layers from absorbing excessive source-domain information. As previously analyzed, fixing stage 1 and stage 2 of the baseline resulted in the best performance, which suggests that shallow layers tend" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.906, + 0.901 + ], + "angle": 0, + "content": "3We present the validation of CNN-based methods here, and similar trends hold for ViT-based methods. Please refer to the appendix for details." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4623" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 2. MioU of 1-shot and 5-shot results compared with previous FSS and CD-FSS methods. All the methods are trained on PASCAL and tested on the CD-FSS. The best performance among all methods is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.122, + 0.857, + 0.361 + ], + "angle": 0, + "content": "
MethodMarkBackboneFSS-1000Deep globeISICChest X-rayAverage
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
Few-Shot Segmentation Methods
RePRI [2]CVPR-21Res-5070.9674.2325.0327.4123.2726.2365.0865.4846.0948.34
HSNet [26]ICCV-21Res-5077.5380.9929.6535.0831.2035.1051.8854.3647.5751.38
SSP [12]ECCV-22Res-5078.9180.5940.0048.6835.4945.8674.4474.2657.2162.35
FPTrans [41]NIPS-22ViT-base80.7483.6538.3649.3048.6560.3780.9282.9162.1769.06
PerSAM [42]ICLR-24ViT-base60.9266.5336.0840.6523.2725.3329.9530.0537.5640.64
Cross-Domain Few-Shot Segmentation Methods
PATNet [20]ECCV-22Res-5078.5981.2337.8942.9741.1653.5866.6170.2056.0661.99
APM [35]NIPS-24Res-5079.2981.8340.8644.9241.7151.1678.2582.8160.0365.18
ABCDFSS [17]CVPR-24Res-5074.6076.2042.6049.0045.7053.3079.8081.4060.6764.97
DRA [33]CVPR-24Res-5079.0580.4041.2950.1240.7748.8782.3582.3160.8665.42
APSeg [16]CVPR-24ViT-base79.7181.9035.9439.9845.4353.9884.1084.5061.3065.09
OursOursRes-5078.5180.6044.1049.6738.2147.0481.0282.7360.4665.01
OursOursViT-base81.0583.6942.1251.4852.9162.4383.9484.1265.0170.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.375, + 0.419, + 0.39 + ], + "angle": 0, + "content": "Table 3. Ablation study on key components." + }, + { + "type": "table", + "bbox": [ + 0.191, + 0.394, + 0.388, + 0.47 + ], + "angle": 0, + "content": "
LEMLCMResNetViT
57.2162.17
✓58.3563.06
✓59.7864.39
✓✓60.4665.01
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.479, + 0.48, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.632 + ], + "angle": 0, + "content": "Figure 11. (Left) The CKA similarity between source and target domain with and without LEM, verifies that LEM helps the model to focus more on domain-agnostic information. (Right) Feature map visualization of stage 1 and stage 4 at epoch 1 and epoch 20." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.734 + ], + "angle": 0, + "content": "to learn information that negatively impacts target-domain generalization. Fig.10 (right) shows that after incorporating our LEM, training stage 2 produced results comparable to those achieved by fixing stage 2. This indicates that LEM effectively prevents the shallow layers from absorbing excessive source-domain information during the training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "LEM encourages the model to focus more on domain-agnostic information. To assess how the domain distance between the source and target domains changes before and after applying LEM, we use CKA (Centered Kernel Alignment) similarity for measurement, following the approach in [6, 45, 47-49]. Specifically, we extract features from source domain and target domain images using a backbone network, then compute the CKA similarity of the final layer's feature by aligning the channel dimension. A higher CKA similarity indicates a smaller domain distance, suggesting the model retains less domain-specific information. As shown" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.378, + 0.707, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.378, + 0.898, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.504, + 0.907, + 0.533 + ], + "angle": 0, + "content": "Figure 12. LCM reduces the sharpness of the loss landscape and enhances robustness against domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.632 + ], + "angle": 0, + "content": "in Fig.11 (left), the average CKA similarity across four target domains increases after applying random convolution, and is further improved with the addition of the Fourier transform. This demonstrates that by perturbing the source domain, LEM encourages the model to capture domain-agnostic information during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.906, + 0.724 + ], + "angle": 0, + "content": "LEM enables the model to capture more useful low-level information in shallow layers. As shown in Fig.11 (right), LEM helps the model learn boundary information more effectively in stage 1, allowing it to focus more accurately on the foreground in stage 4. This leads to more precise activation regions in the later epochs of source domain training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.71, + 0.755 + ], + "angle": 0, + "content": "4.4.3 Verification of LCM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "LCM smooths loss landscapes. To assess the impact of LCM in reducing sharpness and enhancing robustness against domain shifts, we superimpose styles from ISIC medical images and Deep globe remote sensing images onto source domain images to simulate domain shifts, with the level of style alteration serving as the perturbation level. As shown in Fig.12, the lowered performance drop validates that LCM smooths the loss landscape, thus improves robustness against domain shifts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4624" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.091, + 0.278, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.09, + 0.465, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.258 + ], + "angle": 0, + "content": "Figure 13. The effect of LEM and LCM at different positions in the shallow layers. Block ID represents the sequence in which the module is applied: before stage 1, stage 1 block 1, stage 2 block 1, stage 2 block 2, and stage 2 block 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.266, + 0.414, + 0.283 + ], + "angle": 0, + "content": "4.4.4 Sensitivity Study of Hyper-parameters" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.483, + 0.337 + ], + "angle": 0, + "content": "As shown in Fig.13, we investigated the effects of using LEM and LCM at various positions in the shallow layers. Please refer to the appendix for more details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.232, + 0.366 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.375, + 0.31, + 0.391 + ], + "angle": 0, + "content": "5.1. Few-Shot Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Few-Shot Segmentation (FSS) aims to segment novel classes in query images using only a few annotated support samples. Existing methods in FSS can be generally categorized into two types: prototype-based methods and matching-based methods. Motivated by PrototypicalNet[32], prototype-based methods [19, 21, 34, 38] generate prototypes from support images and perform segmentation by computing the similarity between query features and these support prototypes. On the other hand, matching-based methods[25, 26, 29, 40] focus on analyzing pixel-to-pixel dense correspondences between support and query features to avoid the loss of spatial structure inherent in prototype-based methods. However, these methods are limited to segmenting novel classes within a single domain and do not generalize effectively to unseen domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.307, + 0.65 + ], + "angle": 0, + "content": "5.2. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Domain Generalization (DG) aims to train models that can generalize to diverse, unseen target domains, particularly when target domain data is unavailable during training, which aligns with the objective of CDFSS. Existing methods for domain generalization can be categorized into two types: methods that focus on learning domain-invariant feature representations across multiple source domains[4, 37], and methods that generate diverse samples through data or feature augmentation[10, 31]. However, most existing DG methods focus on generalizing the whole model, which is not suitable for direct use in CDFSS tasks. In contrast, our plug-and-play LEM module is compact, lightweight, and effective. In the appendix, we compare our LEM module with domain generalization methods to demonstrate the effectiveness of our approach as a novel sharpness-aware minimization method in addressing the CDFSS problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.844, + 0.108 + ], + "angle": 0, + "content": "5.3. Cross-domain Few-Shot Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Unlike traditional FSS setting, CDFSS requires models to generalize to unseen target domains without accessing target data during training, making the task more realistic and complex. PATNet [20] introduces a transformation module that converts domain-specific features into domain-agnostic features, which can be finetuned with target domain data during the testing phase. ABCDFSS [17] proposes a test-time task adaptation method, utilizing consistency-based contrastive learning to prevent overfitting and improve class discriminability. DR-Adapter [33] introduces a small adapter for rectifying diverse target domain styles to the source domain. Compared with them, our approach proposes two simple plug-and-play modules that require no training and can be directly applied to boost performance in CDFSS. Besides, a few works [12, 27] are based on the self-support concept. Our LCM module follows a similar principle but offers two significant advantages: (1) While other methods use query prototypes to match query features, we directly use the low-level query features to adjust the query mask, preventing further loss of information. (2) LCM can be conveniently added during the target-domain testing phase, rather than during the entire training process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.459, + 0.634, + 0.475 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.71 + ], + "angle": 0, + "content": "In this paper, We focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, the best segmentation performance is always achieved at the very early epochs, followed by a sharp decline as source-domain training progresses. We delve into this phenomenon for an interpretation, and find low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS. Based on the interpretation, we further propose a method that includes two plug-and-play modules: one to flatten the loss landscapes for low-level features during source-domain training, and the other to directly supplement target-domain information to the model during target-domain testing. Extensive experiments on four CDFSS benchmarks validate our rationale and effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.671, + 0.74 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.889 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Program of China under grant 2024YFC3307900; the National Natural Science Foundation of China under grants 62206102, 62436003, 62376103 and 62302184; Major Science and Technology Project of Hubei Province under grant 2024BAA008; Hubei Science and Technology Talent Service Project under grant 2024DJC078; and Ant Group through CCF-Ant Research Fund. The computation is completed in the HPC Platform of Huazhong University of Science and Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4625" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11700-11709, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[2] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13979-13988, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE Transactions on Medical Imaging, 33(2):577-590, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2229-2238, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.482, + 0.505 + ], + "angle": 0, + "content": "[5] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.482, + 0.562 + ], + "angle": 0, + "content": "[6] MohammadReza Davari, Stefan Horoi, Amine Natik, Guillaume Lajoie, Guy Wolf, and Eugene Belilovsky. Reliability of cka as a similarity measure in deep learning. arXiv preprint arXiv:2210.16156, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[7] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deep globe 2018: A challenge to parse the earth through satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 172-181, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.762, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[10] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 200-216. Springer, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2010. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[12] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701–719, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[14] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In Proceedings of the IEEE International Conference on Computer Vision, pages 991-998. IEEE, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.904, + 0.332 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.336, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[16] Weizhao He, Yang Zhang, Wei Zhuo, Linlin Shen, Jiaqi Yang, Songhe Deng, and Liang Sun. Apseg: Auto-prompt network for cross-domain few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23762-23772, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.907, + 0.461 + ], + "angle": 0, + "content": "[17] Jonas Herzog. Adapt before comparison: A new perspective on cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23605-23615, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.464, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[18] Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les Folio, Jenifer Siegelman, Fiona Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K Singh, Sameer Antani, et al. Automatic tuberculosis screening using chest radiographs. IEEE Transactions on Medical Imaging, 33(2):233-245, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[19] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8057-8067, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[20] Shuo Lei, Xuchao Zhang, Jianfeng He, Fanglan Chen, Bowen Du, and Chang-Tien Lu. Cross-domain few-shot semantic segmentation. In European Conference on Computer Vision, pages 73-90. Springer, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.677, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[21] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[22] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2869–2878, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4626" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[24] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.486, + 0.22 + ], + "angle": 0, + "content": "[25] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8741-8750, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.486, + 0.276 + ], + "angle": 0, + "content": "[26] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.486, + 0.359 + ], + "angle": 0, + "content": "[27] Jiahao Nie, Yun Xing, Gongjie Zhang, Pei Yan, Aoran Xiao, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Cross-domain few-shot segmentation via iterative support-query correspondence mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3380-3390, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.486, + 0.39 + ], + "angle": 0, + "content": "[28] Alan V Oppenheim and Jae S Lim. The importance of phase in signals. Proceedings of the IEEE, 69(5):529-541, 1981. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.486, + 0.461 + ], + "angle": 0, + "content": "[29] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.486, + 0.545 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.486, + 0.603 + ], + "angle": 0, + "content": "[31] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. arXiv preprint arXiv:1804.10745, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.486, + 0.645 + ], + "angle": 0, + "content": "[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.486, + 0.716 + ], + "angle": 0, + "content": "[33] Jiapeng Su, Qi Fan, Wenjie Pei, Guangming Lu, and Fanglin Chen. Domain-rectifying adapter for cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24036–24045, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.486, + 0.785 + ], + "angle": 0, + "content": "[34] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.486, + 0.844 + ], + "angle": 0, + "content": "[35] Jintao Tong, Yixiong Zou, Yuhua Li, and Ruixuan Li. Lightweight frequency masker for cross-domain few-shot semantic segmentation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.486, + 0.901 + ], + "angle": 0, + "content": "[36] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data, 5(1):1-9, 2018. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.146 + ], + "angle": 0, + "content": "[37] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. arXiv preprint arXiv:1903.06256, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.908, + 0.219 + ], + "angle": 0, + "content": "[38] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[39] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.907, + 0.33 + ], + "angle": 0, + "content": "[40] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. Advances in Neural Information Processing Systems, 34:21984-21996, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.907, + 0.387 + ], + "angle": 0, + "content": "[41] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in neural information processing systems, 35:6575-6588, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[42] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Yu Qiao, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. In The Twelfth International Conference on Learning Representations, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.907, + 0.516 + ], + "angle": 0, + "content": "[43] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.907, + 0.584 + ], + "angle": 0, + "content": "[44] Yixiong Zou, Shanghang Zhang, Jianpeng Yu, Yonghong Tian, and José MF Moura. Revisiting mid-level patterns for cross-domain few-shot recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 741-749, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.907, + 0.642 + ], + "angle": 0, + "content": "[45] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.907, + 0.712 + ], + "angle": 0, + "content": "[46] Yixiong Zou, Yicong Liu, Yiman Hu, Yuhua Li, and Ruixuan Li. Flatten long-range loss landscapes for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23575-23584, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.907, + 0.769 + ], + "angle": 0, + "content": "[47] Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. Attention temperature matters in vit-based cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:116332-116354, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.77, + 0.907, + 0.826 + ], + "angle": 0, + "content": "[48] Yixiong Zou, Shuai Yi, Yuhua Li, and Ruixuan Li. A closer look at the clf token for cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:85523-85545, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.827, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[49] Yixiong Zou, Shanghang Zhang, Haichen Zhou, Yuhua Li, and Ruixuan Li. Compositional few-shot class-incremental learning. arXiv preprint arXiv:2405.17022, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.87 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4627" + } + ] +] \ No newline at end of file diff --git a/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_origin.pdf b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..902681210597dbd4c8ea53f2426ca1e801e55541 --- /dev/null +++ b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/9337f673-c412-4595-8bb5-7eba2f00a28b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09ce26dc1d23c3cd54c19a1684023d061d81012106461533b9926f074622dcbb +size 6140104 diff --git a/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/full.md b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..56247ee67c54adf9e410446fd867fdfa37014f09 --- /dev/null +++ b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/full.md @@ -0,0 +1,363 @@ +# The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation + +Yuhan Liu, Yixiong Zou*, Yuhua Li, Ruixuan Li +School of Computer Science and Technology, Huazhong University of Science and Technology +{yuhan.liu, yixiongz, idcliyuhua, rxli}@hust.edu.cn + +# Abstract + +Cross-Domain Few-Shot Segmentation (CDFSS) is proposed to transfer the pixel-level segmentation capabilities learned from large-scale source-domain datasets to downstream target-domain datasets, with only a few annotated images per class. In this paper, we focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, particularly those distant from the source domain, segmentation performance peaks at the very early epochs, and declines sharply as the source-domain training proceeds. We delve into this phenomenon for an interpretation: low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS. Based on this phenomenon and interpretation, we further propose a method that includes two plug-and-play modules: one to flatten the loss landscapes for low-level features during source-domain training as a novel sharpness-aware minimization method, and the other to directly supplement target-domain information to the model during target-domain testing by low-level-based calibration. Extensive experiments on four target datasets validate our rationale and demonstrate that our method surpasses the state-of-the-art method in CDFSS significantly by $3.71\%$ and $5.34\%$ average MIoU in 1-shot and 5-shot scenarios, respectively. + +# 1. Introduction + +Current deep neural networks [24, 30, 43] have achieved remarkable success in semantic segmentation, but their performance heavily relies on large-scale annotated datasets [1, 23]. However, the process of annotating data, particularly for dense pixel-wise tasks like semantic segmentation, is labor-intensive and time-consuming. To address this issue, Few-Shot Segmentation (FSS) has been introduced, aiming to generate pixel-level predictions for unseen categories with only a few labeled samples. Typically, the model is pretrained on a large-scale dataset of base classes and then trans + +![](images/cf8eb221a693560d4bf08f9f725c935c7292d640d6a6392f760e32a52267057f.jpg) + +![](images/6c5b81cfd8696020ec59dc2d51f6f2c8d9888fb2e4f152d57bc0a24558bb0134.jpg) +(b) Previous CDFSS methods Epoch vs. MoU + +![](images/40838c235abca060f02be4da9ce2eeb0786f2329a147d16afe4fb82367e652cb.jpg) +(c) Our method Epochs vs. MoU +Figure 1. (a) In CDFSS tasks, the training (source) and testing (target) datasets belong to distinct domains, with categories in the testing dataset being unseen during training. (b) Previous CDFSS methods show a decreasing trend of mIoU as the source-domain training proceeds, even at very early epochs for distant domains. (c) Our method can effectively prevent the model from performance decline after early epochs and achieve higher performance. + +ferred to unseen novel classes. However, novel classes might not be in the same domain as base classes [5, 7, 11, 22, 36], a more realistic approach is to consider domain gaps between the base (source domain) and novel (target domain) classes, which gives rise to the Cross-Domain Few-Shot Segmentation (CDFSS) task [20]. + +Although extensive works [16, 17, 20, 33] have been developed, a well-observed phenomenon is still not handled: for target domains, especially those distant from the source domain, the best performance is always achieved at the very early epochs or even the first epoch. As shown in Fig. 1b, the performance decreases sharply as the source-domain training proceeds, where the 20th epoch's MIoU is even lower than that of the first epoch. Although the early stop can handle this problem, the goal of the source-domain training is to provide a generalizable model for all target domains [20]. Therefore, it may be inappropriate to try different early stops for specific domains. In this paper, we aim to delve into this well-observed but under-explored phenomenon and handle it based on our interpretations of it, as shown in Fig. 1c. + +To delve into this phenomenon, we first visualize the predictions of the 1st and 20th epochs in Fig. 2. We observe that + +![](images/748fbdb6f1c71e584d3f0948b6c7aee178a067d4d6ae95b262cc997a89ac5e90.jpg) +Figure 2. The visualization of predictions at the 1st and 20th epoch further indicate that CDFSS models may not acquire meaningful information for target domains during source-domain training. + +the 20th epoch's model shows many fundamental errors in target domains, i.e., CDFSS models fail to identify the foreground but focus on entirely wrong regions. This suggests that the CDFSS models may not have acquired meaningful information for target domains during source-domain training, resulting in almost random predictions during testing. We then visualize feature maps at different network stages, and observe this problem occurs even at the shallow layers, and deeper layers produce wrong feature maps with wrong inputs from shallow layers. Quantitatively, previous works [13, 46] suggest that the challenge posed by domain shift can be analyzed through the flatness of the loss landscape. We use this as an entry point to explore the connection between shallow layers and early stops. We find that low-level features, although simpler and previously believed to be more transferable [44], are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training. That is, the devil lies in the low-level features. + +Based on this phenomenon and interpretation, we further propose a method for the CDFSS task. During the source-domain training, we propose a novel sharpness-aware minimization method to flatten the loss landscapes for low-level features, which is achieved by a shape-preserving perturbation module with random convolutions. During the target-domain phase, since low-level features can hardly capture the target-domain information due to its vulnerability to domain shifts, we directly supplement such information to the model with a low-level-based calibration module, which effectively prevents the model from fundamental errors caused by collapsed low-level features. Based on our methods, we can effectively prevent the model from dropping performances after early epochs and achieve higher performance (Fig. 1c). + +In summary, our contributions can be listed as + +- We focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, particularly those distant from the source domain, the best segmentation performance is always achieved at the very early epochs, followed by a sharp decline as source-domain training progresses. +- We delve into this phenomenon for an interpretation: low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS. + +- Building on this interpretation, we propose a method that includes two plug-and-play modules: one act as a novel sharpness-aware minimization method to flatten the loss landscapes of low-level features during source-domain training, while the other directly supplement target-domain information to the model during target-domain testing. +- Extensive experiments on four target datasets validate the rationale of our interpretation and method, showing its superiority over current state-of-the-art methods. + +# 2. Interpretation + +# 2.1. Preliminaries + +Cross-Domain Few-Shot Segmentation (CDFSS) aims to transfer the segmentation capabilities learned from the source domain to target domains, with only a few annotated images per class. The source domain $\mathcal{D}_s = (\mathcal{X}_s,\mathcal{Y}_s)$ and target domain $\mathcal{D}_t = (\mathcal{X}_t,\mathcal{Y}_t)$ are defined by different input distributions and disjoint label spaces, i.e., $\mathcal{X}_s\neq \mathcal{X}_t$ $\mathcal{Y}_s\cap \mathcal{Y}_t = \emptyset$ , where $\mathcal{X}$ refers to input distribution and $\mathcal{Y}$ refers to the label space. + +In this work, we adopt the meta-learning episodic paradigm following [20]. Specifically, both $\mathcal{D}_s$ and $\mathcal{D}_t$ consist of several episodes. Each episode is constructed by a support set $S = \{(I_s^i,M_s^i)\}_{i = 1}^K$ with $K$ training samples and a query set $Q = \{(I_q,M_q)\}$ , where $I$ denote the image and $M$ denote the pixel labels. Within each episode, the model leverages $\{I_s,M_s\}$ and $I_{q}$ to predict the query mask $M_{q}$ . + +The model is trained on the source domain support sets and query set $\{S^s, Q^s\}$ by minimizing the binary cross-entropy loss: + +$$ +\mathcal {L} = B C E \left(F \left(I _ {s} ^ {s}, I _ {q} ^ {s}, M _ {s} ^ {s}\right), M _ {q} ^ {s}\right), \tag {1} +$$ + +where $F\left(I_s^s,I_q^s,M_s^s\right) = h\left(f\left(I_s^s,I_q^s\right),M_s^s\right)$ outputs the segmentation score map of $I_{q}^{s}$ , which is composed of a feature extractor $f(\cdot)$ and a comparison module $h(\cdot)$ . + +Then the model will be evaluated on the target domain query set $Q^t$ , the prediction for $I_q^t$ is: + +$$ +\hat {M} _ {q} ^ {t} = \arg \max F \left(I _ {s} ^ {t}, I _ {q} ^ {t}, M _ {s} ^ {t}\right) \tag {2} +$$ + +Finally, the evaluation can be conducted based on the comparison between $\hat{M}_q^t$ and the real label $M_q^t$ . + +# 2.2. Delve into the Early Stop in Cross-Domain Few-Shot Segmentation + +# 2.2.1 Intuitive observation: The devil is in the low-level features + +Given our suspicion that the model fails to learn useful information from the source domain for target domains, we start by visualizing the feature maps. Using ResNet-50 [15] as the backbone following [12], we visualize the feature maps + +![](images/b3318b69b17f2be6d6b9e2082fac3c9c7927ff459e5ecb4009fac1c044c7599f.jpg) +Figure 3. Feature maps from stages 1 to 4 for both the source domain and target domains. The noticeable contrast in the feature maps at stage 1 between source and target domains indicates the limited performance of CDFSS models stems from shallow layers. + +![](images/561a7f18308d7481b384c7e071df76c95db513c6be2b7bf79cb6ed55c51ff6cc.jpg) +Figure 4. Feature maps from stages 1 to 4 for epoch 1 and epoch 20. Epoch 1 in stage 1 shows more distinguishable activations than epoch 20, indicating that low-level features gradually incorporate incorrect information as training progresses. + +of stages 1 to 4 for both the source domain and the target domains, as shown in Fig. 3. We observe that on source-domain samples, the model learns boundary information effectively in stage 1, enabling it to accurately focus on the foreground in stage 4. However, on target domains especially those distant from the source domain, even though the foreground is clear and prominent, the model still fails to learn anything meaningful in stages 1 and 2. Consequently, in stage 4, the model focuses on completely incorrect areas. This indicates that the model's bad performance may originate from its shallow layers. + +In Fig.4, we further visualize feature maps of each stage at the 1st and 20th epoch, respectively. We find that the shallow layers show more distinguishable activations at epoch 1 than at epoch 20, which aligns with our observation of higher accuracy at epoch 1. Consequently, we conclude that the poor performance of CDFSS models may be attributed to shallow layers that fail to capture useful low-level information, a trend that intensifies during source-domain training. + +# 2.2.2 Quantitative verification: Low-level features lead to a sharper loss landscape + +Previous work [13, 46] has demonstrated that the sharpness of the loss landscape can serve as a tool to analyze cross + +![](images/f6a00bcf4994f71fed9b420510b09808e4b38cfeed046a59527475ad05cc595b.jpg) +(a) + +![](images/8deb8f3d8d307ec6b234a9c5a7f5654876cbe8d86fe72babe057b93515b738b7.jpg) +(b) + +![](images/7c1206a13d5519682b3ffc20def6075c7c929eef76c8c0e3bc7a3aed26238480.jpg) +(c) +Figure 5. (a) A sharp minimum in the landscape corresponds to a representation that is highly sensitive to data shifts. (b) Examples of pixel perturbation applied to the images. (c) As training proceeds, the loss landscape becomes progressively sharper. (d) Perturbing shallow layers leads to much sharper loss landscapes, indicating shallow layers are the cause of the sharp loss landscape and the increased sensitivity to domain shifts. + +![](images/311d767e60f852abe75d3366fa3cb4b8185ccd73533c7d0f1d6e9a089318d5f1.jpg) +(d) 0.3 + +domain issues. As illustrated in Fig. 5a, a smoother loss landscape enables the model to better handle domain shifts1. Therefore, we measure the sharpness of the source-domain-trained models against different epochs and stages. + +We first measure the sharpness of the loss landscape across different epochs. Following the method in [46], we apply pixel perturbations to the training data. Examples of these perturbations are shown in Fig. 5b, while the magnitude of performance drop is depicted in Fig.5c. A larger performance drop under the same perturbation indicates a sharper loss landscape. As observed, the performance of models across all epochs drops when exposed to perturbations, but the model from epoch 1 experiences a smaller drop than that of epoch 10, which in turn performs better than epoch 20. This suggests that as training progresses, the loss landscape becomes progressively sharper, making the model increasingly vulnerable to domain shifts and resulting in decreased performance on target domains. This is in line with our previous observations. + +Based on this, we further investigate the relationship between model layers and the loss landscape. We apply the same magnitude of perturbations to the feature maps at stages 1-4 and measure the performance drops. The results in Fig.5d show that earlier stages exhibit larger performance drops under the same perturbation, which implies that shallow layers lead to sharper loss landscapes and are more vulnerable to domain shifts. This experiment provides quantitative validation for the intuitive observation from the low-level feature visualization in the previous section. + +Table 1. The model with fixed shallow layers performs better than the one with trained shallow layers, implying that shallow layers learn domain-specific information that negatively impacts generalization to the target domain. + +
MethodFSS-1000Deep globeISICChestX-rayAverage
train stage1, 2, 3, 478.8639.4435.7672.4956.64
fix stage1, train stage2, 3, 478.8839.9037.0072.1256.98
fix stage1, 2, train stage3, 478.9140.0035.4974.4457.21
+ +![](images/6944e8d5bff1995093a9884b632c31c2c9f2f1669082b7f3bd4b184d1e138d6d.jpg) +Figure 6. The mIoU curves over epochs for two different training settings indicate that shallow layers are more vulnerable to domain shifts, leading to more severe performance degradation during source-domain training. + +![](images/f1afcf28068f52da55ec1ac2f926012cbd4f675c898488ed395f3e9af2f98afc.jpg) + +# 2.2.3 Further analysis: Low-level features absorb domain-specific information + +To investigate the impact of shallow layers on model performance, we compare the results of fixing stage 1 and stage 2 using ImageNet [8] pre-trained weights with the results of training all stages, as shown in Tab. 1. We find that fixing stage 1 and stage 2 leads to the best performance. + +In Fig 6, we further visualize the mIoU curves over epochs for two different training settings: fix stage 1 only, and fix stage 1, 2. The comparison revealed that when stage 2 is fixed, the model reaches the optimal mIoU at a later epoch, and as training progresses, the mIoU decreases more slowly. + +This suggests that shallow layers are vulnerable to domain shifts, therefore any slight absorption of domain information could harm target-domain generalization. + +# 2.3. Conclusion and Discussion + +Based on the above experiments, we make the interpretation as follows. The shallow layers of the model tend to learn domain-specific information, causing the low-level features to become increasingly domain-specific during training, gradually overfitting to the source domain. The overfitting sharpens the model's loss landscape and makes it more vulnerable to domain shifts, thereby reducing cross-domain performance and introducing significant mistakes in CDFSS. This problem is partially addressed by the early-stop strategy and freezing shallow layers. + +However, merely stopping the training of shallow layers cannot fully unlock the power of source-domain training. Therefore, below we propose a novel method to handle the problem in shallow layers. + +# 3. Method + +# 3.1. Method Overview + +Our core idea is that low-level features are vulnerable to domain shifts, which may lead to a sharp decrease in generalization with only slight absorption of source-domain information. Based on this, we designed two plug-and-play modules to operate in two phases: 1) During the source-domain training phase, the Low-level Enhancing Module enhances the robustness against domain shifts by synthesizing diverse domains and incorporating them into low-level features. 2) During the target-domain testing phase, the Low-level Calibration Module refines previous segmentation results by utilizing the low-level features from the target domain. The framework is illustrated in Fig. 7. + +# 3.2. Train-time Low-level Enhancing Module + +As analyzed above, overfitting of low-level features on the source domain during training leads to a sharp loss landscape, making the model vulnerable to domain shifts. According to the previous Sharpness Aware Minimization (SAM) method [13, 46], one effective strategy to smooth loss landscapes is to introduce perturbations. Thus we transform the domain of the low-level support feature during training, which acts as a form of perturbation to reduce overfitting. + +Previous work [39] has shown that random convolution is a shape-preserving strategy to distort local textures, which we view to be suitable for the segmentation task. We take advantage of this property to generate diverse domains while retaining the content. Specifically, given support feature $F_{s} \in \mathbb{R}^{H \times W \times C}$ from the early layers of the encoder and a convolution layer with filters $\Theta \in \mathbb{R}^{h \times w \times C \times C}$ , where $H$ , $W$ and $C$ represent height, width and channels of the input feature, and $h$ and $w$ are the height and width of the random convolution filter. We sample the filter weights from $N(0, \sigma^2)$ , where $\sigma$ acts as a hyperparameter to control the perturbation magnitude. Then we get perturbed support feature $F_{s}'$ with transformed domain: + +$$ +F _ {s} ^ {\prime} = F _ {s} * \Theta , \tag {3} +$$ + +Although random convolution can generally preserve the shape and semantics of the feature, some edge details inevitably get blurred. Existing research [28] has shown that the Fast Fourier Transformation (FFT) decomposes a signal into amplitude spectrum and phase spectrum, with amplitude capturing the overall domain and texture information and phase representing shape, edge, and spatial structure. Thus, we apply FFT to both $F_{s}$ and $F_{s}^{\prime}$ , decomposing them into phase spectrum $\mathcal{P}$ , $\mathcal{P}^{\prime}$ and amplitude spectrum $\mathcal{A}$ , $\mathcal{A}^{\prime}$ : + +$$ +\mathcal {A} e ^ {i \mathcal {P}} = F F T \left(F _ {s}\right), \tag {4} +$$ + +$$ +\mathcal {A} ^ {\prime} e ^ {i \mathcal {P} ^ {\prime}} = F F T \left(F _ {s} ^ {\prime}\right), \tag {5} +$$ + +![](images/c53872451b8557a2334c02b7e4c72ae6566938b49ceacdddaeafc7874a081cb3.jpg) +Figure 7. The overall architecture of our method in a 1-way 1-shot example. Our method consists of two modules: the Low-level Enhancing Module (LEM) and the Low-level Calibration Module (LCM). Support and query images are fed into a weight-shared backbone to extract features, and then processed by a comparison module to obtain a coarse segmentation mask. During source-domain training, LEM perturbs low-level support features for sharpness-aware minimization. During target-domain testing, LCM calibrates the coarse segmentation mask by directly supplementing collapsed low-level target-domain information. + +Then we recombine the phase from the original feature $F_{s}$ and the amplitude from the perturbed feature $F_{s}^{\prime}$ , using the Inverse Fast Fourier Transform (IFFT) to generate transformed support feature: + +$$ +F _ {s} ^ {t} = I F F T \left(\mathcal {A} ^ {\prime} e ^ {i \mathcal {P}}\right), \tag {6} +$$ + +Therefore, while transforming the domain, we achieve better preservation of the boundary shapes, which achieves domain-oriented perturbations on low-level features for sharpness-aware minimization. The transformed feature is then fed into the subsequent layers of the encoder. + +# 3.3. Test-time Low-level Calibration Module + +After being processed by the encoder, the support and query features are fed into the comparison module, yielding a coarse score map $S \in \mathbb{R}^{H \times W \times 2}$ . Since low-level features may be collapsed on target domains due to the vulnerability to domain shifts, we directly use the low-level query feature $F_{q} \in \mathbb{R}^{H' \times W' \times C}$ to supplement such collapsed low-level target-domain information to calibrate the score map during testing on the target domain. + +First, we compute a confidence map $C$ by subtracting the background similarity from the foreground similarity in the score map: + +$$ +C _ {i, j} = S _ {i, j, 1} - S _ {i, j, 0}, \tag {7} +$$ + +Here, $S_{i,j,1}$ and $S_{i,j,0}$ denote the foreground and background similarity for the pixel at $(i,j)$ , respectively. The confidence map $C$ quantifies the likelihood of foreground at each pixel. + +Then we partition the confidence map into patches $\{P_1, P_2, \ldots, P_n\}$ and select the Top-K patches with the highest average confidence. These patches are considered the most reliable foreground regions: + +$$ +\bar {C} _ {p} = \frac {1}{| P _ {p} |} \sum_ {(i, j) \in P _ {p}} C (i, j), \tag {8} +$$ + +$$ +\left\{P _ {k _ {1}}, P _ {k _ {2}}, \dots , P _ {k _ {K}} \right\} = \arg \max _ {P _ {p}} \bar {C} _ {p} \tag {9} +$$ + +Where $|P_{p}|$ is the number of pixels in patch $P_{p}$ and $(i,j)$ are the pixel locations within the patch. + +For each of the selected Top-K foreground patches, we locate the corresponding patch in the interpolated query feature map $F_{q}^{\prime} \in \mathbb{R}^{H \times W \times C}$ as $\{F_{k_1}, F_{k_2}, \ldots, F_{k_K}\}$ , and compute the cosine similarity between it and all other patches in the query feature map: + +$$ +\operatorname {S i m} \left(F _ {k _ {i}}, F _ {m}\right) = \frac {F _ {k _ {i}} \cdot F _ {m}}{\| F _ {k _ {i}} \| \| F _ {m} \|} \tag {10} +$$ + +Finally, the foreground similarity in the score map is updated using the top-K similarity maps: + +$$ +S _ {i, j, 1} ^ {\prime} = S _ {i, j, 1} + \sum_ {k = 1} ^ {K} w _ {k} \cdot \left(S i m _ {k} (i, j) - \beta_ {k}\right), \tag {11} +$$ + +Here, $w_{k}$ is a scaling factor and $\beta_{k}$ is a bias term. This adjustment improves foreground detection by directly supplementing collapsed low-level target-domain information. + +During the source-domain training phase, the model will be trained by minimizing the loss in Eq. 1. During the target-domain testing phase, the prediction will be obtained based on Eq. 2 and LCM. + +# 4. Experiments + +# 4.1. Datasets + +Following the settings in [20], we use PASCAL VOC 2012 [11] with SBD augmentation [14] as source domain for training. Then we evaluate the trained models on four target domains: FSS-1000 [22], Deep globe [7], ISIC [5, 36], and Chest X-ray [3, 18]. See Appendix for more details. + +![](images/ab0f1e8656c197d40a89316c0a1b56dfaf85408ac5abc3056fa9a5a2c646f866.jpg) +Figure 8. Qualitative results of our model for 1-shot setting. + +# 4.2. Implementation Details + +We employ ResNet-50 [15] with ImageNet [8] pretrained weights as our backbone. Following baseline SSP [12], we discard the last backbone stage and the last ReLU for better generalization. Furthermore, we integrated our approach into the transformer architecture by using ViT-B/16 [9] as the backbone, following FPTrans [41]. Consistent with FPTrans [41], we resize both support and query images to $480 \times 480$ . During training on the source domain, we use SGD to optimize our model, with a momentum of 0.9 and a initial learning rate of 1e-3. We apply our LEM into different shallow layers of the backbone, more details can be seen in 4.4.4. The random convolution filter used in LEM has a kernel size of $3 \times 3$ , and the standard deviation $\sigma$ is set to be 0.1. During testing on the target domain, we apply our LCM to refine segmentation mask. The hyperparameters $K$ , $w$ and $\beta$ are set to 3, 0.6 and 0.7, respectively. + +# 4.3. Comparison with State-of-the-art Methods + +In Tab. 2, we compare our method with CNN-based and ViT-based approaches, including traditional FSS methods and existing CDFSS methods. Our results illustrate a notable enhancement for both 1-shot and 5-shot tasks. Specifically, our method outperforms the state-of-the-art APSeg [16] by $3.71\%$ and $5.34\%$ in the 1-shot and 5-shot settings respectively, confirming the effectiveness of our strategy. Additionally, we apply our method on existing CDFSS baselines and observe improved performance. Please refer to the appendix for more details. We present qualitative results of our method in 1-way 1-shot segmentation in Fig.8. + +# 4.4. Ablation Study + +# 4.4.1 Effectiveness of Each Component + +We evaluate each proposed component on both CNN and ViT baselines to assess the effectiveness of our designs, including LEM and LCM. The results presented in Tab. 3 show that in the 1-shot setting, introducing LEM improved average MIoU by $1.14\%$ for the CNN baseline and $0.89\%$ + +![](images/a27ecdc96cd5cad3ef4912c796eb82177c315688a64a7b8eb3286b2ad613e08e.jpg) +Figure 9. Applying LEM leads to reduced loss landscape sharpness and improved robustness to domain shifts, Fourier transformation further enhances these benefits. + +![](images/c51bdc8bc22b9c61eae53e5bf479796c90a90e87d02da12c54c03d027ff1cec1.jpg) + +![](images/cc344316c5d85db3bc6d869cb0cabccd83b357c0c48990d8dbf45bfa23818bc7.jpg) +Figure 10. (Left) The mIoU of Chest-X Ray with and without LEM, confirms that LEM leads to consistent and steady performance improvement as training advances. (Right) After incorporating our LEM module, training shallow layers can achieve performance comparable to fixing them. + +![](images/bdc1a79db33df37358ae6fd9ec547615aa566ded82163897d29a409b8f0dbf02.jpg) + +for the ViT baseline, while adding LCM further increased it by $2.11\%$ and $1.95\%$ , respectively. These results clearly demonstrate the effectiveness of each component in enhancing performance. + +# 4.4.2 Verification of LEM + +LEM smooths loss landscapes. To assess the impact of LEM, we measured the loss landscape sharpness before and after its application. We introduced low-frequency noise on training data to simulate domain shifts in the representation space (i.e., pixels and features). A larger performance drop suggests a sharper landscapes. As illustrated in Fig.9, random convolution smooths the loss landscape compared with the baseline, while Fourier transformation further enhances this effect. These results confirm that LEM effectively smooths the loss landscape, which improves robustness against domain shifts. + +LEM ensures performance improves with training. We compared mIoU trends over training epochs with and without LEM in Fig.10 (left). In the baseline model, mIoU on the target domain declines as it increasingly absorbs source-domain information during training. Conversely, with LEM, mIoU not only outperforms the baseline at each epoch but also keeps increasing as training advances, indicating that the model focuses more on domain-invariant information. + +LEM prevents the shallow layers from absorbing excessive source-domain information. As previously analyzed, fixing stage 1 and stage 2 of the baseline resulted in the best performance, which suggests that shallow layers tend + +Table 2. MioU of 1-shot and 5-shot results compared with previous FSS and CD-FSS methods. All the methods are trained on PASCAL and tested on the CD-FSS. The best performance among all methods is highlighted in bold. + +
MethodMarkBackboneFSS-1000Deep globeISICChest X-rayAverage
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
Few-Shot Segmentation Methods
RePRI [2]CVPR-21Res-5070.9674.2325.0327.4123.2726.2365.0865.4846.0948.34
HSNet [26]ICCV-21Res-5077.5380.9929.6535.0831.2035.1051.8854.3647.5751.38
SSP [12]ECCV-22Res-5078.9180.5940.0048.6835.4945.8674.4474.2657.2162.35
FPTrans [41]NIPS-22ViT-base80.7483.6538.3649.3048.6560.3780.9282.9162.1769.06
PerSAM [42]ICLR-24ViT-base60.9266.5336.0840.6523.2725.3329.9530.0537.5640.64
Cross-Domain Few-Shot Segmentation Methods
PATNet [20]ECCV-22Res-5078.5981.2337.8942.9741.1653.5866.6170.2056.0661.99
APM [35]NIPS-24Res-5079.2981.8340.8644.9241.7151.1678.2582.8160.0365.18
ABCDFSS [17]CVPR-24Res-5074.6076.2042.6049.0045.7053.3079.8081.4060.6764.97
DRA [33]CVPR-24Res-5079.0580.4041.2950.1240.7748.8782.3582.3160.8665.42
APSeg [16]CVPR-24ViT-base79.7181.9035.9439.9845.4353.9884.1084.5061.3065.09
OursOursRes-5078.5180.6044.1049.6738.2147.0481.0282.7360.4665.01
OursOursViT-base81.0583.6942.1251.4852.9162.4383.9484.1265.0170.43
+ +Table 3. Ablation study on key components. + +
LEMLCMResNetViT
57.2162.17
✓58.3563.06
✓59.7864.39
✓✓60.4665.01
+ +![](images/de25aa8f7ec34f7938fbcacba3aadb58b08e8edaba64d5ceb0c0cdf449a46cba.jpg) +Figure 11. (Left) The CKA similarity between source and target domain with and without LEM, verifies that LEM helps the model to focus more on domain-agnostic information. (Right) Feature map visualization of stage 1 and stage 4 at epoch 1 and epoch 20. + +to learn information that negatively impacts target-domain generalization. Fig.10 (right) shows that after incorporating our LEM, training stage 2 produced results comparable to those achieved by fixing stage 2. This indicates that LEM effectively prevents the shallow layers from absorbing excessive source-domain information during the training process. + +LEM encourages the model to focus more on domain-agnostic information. To assess how the domain distance between the source and target domains changes before and after applying LEM, we use CKA (Centered Kernel Alignment) similarity for measurement, following the approach in [6, 45, 47-49]. Specifically, we extract features from source domain and target domain images using a backbone network, then compute the CKA similarity of the final layer's feature by aligning the channel dimension. A higher CKA similarity indicates a smaller domain distance, suggesting the model retains less domain-specific information. As shown + +![](images/f165ed823f9ff79e6b064d4ca0bc7a3b2df2272a7393aa68036276151d149d8d.jpg) +Figure 12. LCM reduces the sharpness of the loss landscape and enhances robustness against domain shifts. + +![](images/a60ade4284c7152bf030a69f61935a0f5efd9337a3e42797bbf473022551ff2c.jpg) + +in Fig.11 (left), the average CKA similarity across four target domains increases after applying random convolution, and is further improved with the addition of the Fourier transform. This demonstrates that by perturbing the source domain, LEM encourages the model to capture domain-agnostic information during training. + +LEM enables the model to capture more useful low-level information in shallow layers. As shown in Fig.11 (right), LEM helps the model learn boundary information more effectively in stage 1, allowing it to focus more accurately on the foreground in stage 4. This leads to more precise activation regions in the later epochs of source domain training. + +# 4.4.3 Verification of LCM + +LCM smooths loss landscapes. To assess the impact of LCM in reducing sharpness and enhancing robustness against domain shifts, we superimpose styles from ISIC medical images and Deep globe remote sensing images onto source domain images to simulate domain shifts, with the level of style alteration serving as the perturbation level. As shown in Fig.12, the lowered performance drop validates that LCM smooths the loss landscape, thus improves robustness against domain shifts. + +![](images/45ab3e2c57d2b66948055dfb24e5e26617f1f745898cf8b430bea4b10691e3ab.jpg) +Figure 13. The effect of LEM and LCM at different positions in the shallow layers. Block ID represents the sequence in which the module is applied: before stage 1, stage 1 block 1, stage 2 block 1, stage 2 block 2, and stage 2 block 3. + +![](images/1f9d947d9da0b81e8c741692fa32751b52dd14962647aade5e132ce3e2302419.jpg) + +# 4.4.4 Sensitivity Study of Hyper-parameters + +As shown in Fig.13, we investigated the effects of using LEM and LCM at various positions in the shallow layers. Please refer to the appendix for more details. + +# 5. Related Work + +# 5.1. Few-Shot Segmentation + +Few-Shot Segmentation (FSS) aims to segment novel classes in query images using only a few annotated support samples. Existing methods in FSS can be generally categorized into two types: prototype-based methods and matching-based methods. Motivated by PrototypicalNet[32], prototype-based methods [19, 21, 34, 38] generate prototypes from support images and perform segmentation by computing the similarity between query features and these support prototypes. On the other hand, matching-based methods[25, 26, 29, 40] focus on analyzing pixel-to-pixel dense correspondences between support and query features to avoid the loss of spatial structure inherent in prototype-based methods. However, these methods are limited to segmenting novel classes within a single domain and do not generalize effectively to unseen domains. + +# 5.2. Domain Generalization + +Domain Generalization (DG) aims to train models that can generalize to diverse, unseen target domains, particularly when target domain data is unavailable during training, which aligns with the objective of CDFSS. Existing methods for domain generalization can be categorized into two types: methods that focus on learning domain-invariant feature representations across multiple source domains[4, 37], and methods that generate diverse samples through data or feature augmentation[10, 31]. However, most existing DG methods focus on generalizing the whole model, which is not suitable for direct use in CDFSS tasks. In contrast, our plug-and-play LEM module is compact, lightweight, and effective. In the appendix, we compare our LEM module with domain generalization methods to demonstrate the effectiveness of our approach as a novel sharpness-aware minimization method in addressing the CDFSS problem. + +# 5.3. Cross-domain Few-Shot Segmentation + +Unlike traditional FSS setting, CDFSS requires models to generalize to unseen target domains without accessing target data during training, making the task more realistic and complex. PATNet [20] introduces a transformation module that converts domain-specific features into domain-agnostic features, which can be finetuned with target domain data during the testing phase. ABCDFSS [17] proposes a test-time task adaptation method, utilizing consistency-based contrastive learning to prevent overfitting and improve class discriminability. DR-Adapter [33] introduces a small adapter for rectifying diverse target domain styles to the source domain. Compared with them, our approach proposes two simple plug-and-play modules that require no training and can be directly applied to boost performance in CDFSS. Besides, a few works [12, 27] are based on the self-support concept. Our LCM module follows a similar principle but offers two significant advantages: (1) While other methods use query prototypes to match query features, we directly use the low-level query features to adjust the query mask, preventing further loss of information. (2) LCM can be conveniently added during the target-domain testing phase, rather than during the entire training process. + +# 6. Conclusion + +In this paper, We focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, the best segmentation performance is always achieved at the very early epochs, followed by a sharp decline as source-domain training progresses. We delve into this phenomenon for an interpretation, and find low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS. Based on the interpretation, we further propose a method that includes two plug-and-play modules: one to flatten the loss landscapes for low-level features during source-domain training, and the other to directly supplement target-domain information to the model during target-domain testing. Extensive experiments on four CDFSS benchmarks validate our rationale and effectiveness. + +# Acknowledgments + +This work is supported by the National Key Research and Development Program of China under grant 2024YFC3307900; the National Natural Science Foundation of China under grants 62206102, 62436003, 62376103 and 62302184; Major Science and Technology Project of Hubei Province under grant 2024BAA008; Hubei Science and Technology Talent Service Project under grant 2024DJC078; and Ant Group through CCF-Ant Research Fund. The computation is completed in the HPC Platform of Huazhong University of Science and Technology. + +# References + +[1] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11700-11709, 2019. 1 +[2] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13979-13988, 2021. 7 +[3] Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE Transactions on Medical Imaging, 33(2):577-590, 2013. 5 +[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2229-2238, 2019. 8 +[5] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1, 5 +[6] MohammadReza Davari, Stefan Horoi, Amine Natik, Guillaume Lajoie, Guy Wolf, and Eugene Belilovsky. Reliability of cka as a similarity measure in deep learning. arXiv preprint arXiv:2210.16156, 2022. 7 +[7] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deep globe 2018: A challenge to parse the earth through satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 172-181, 2018. 1, 5 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 4, 6 +[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[10] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 200-216. Springer, 2020. 8 +[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2010. 1, 5 + +[12] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701–719, 2022. 2, 6, 7, 8 +[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 2, 3, 4 +[14] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In Proceedings of the IEEE International Conference on Computer Vision, pages 991-998. IEEE, 2011. 5 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2, 6 +[16] Weizhao He, Yang Zhang, Wei Zhuo, Linlin Shen, Jiaqi Yang, Songhe Deng, and Liang Sun. Apseg: Auto-prompt network for cross-domain few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23762-23772, 2024. 1, 6, 7 +[17] Jonas Herzog. Adapt before comparison: A new perspective on cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23605-23615, 2024. 1, 7, 8 +[18] Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les Folio, Jenifer Siegelman, Fiona Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K Singh, Sameer Antani, et al. Automatic tuberculosis screening using chest radiographs. IEEE Transactions on Medical Imaging, 33(2):233-245, 2013. 5 +[19] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8057-8067, 2022. 8 +[20] Shuo Lei, Xuchao Zhang, Jianfeng He, Fanglan Chen, Bowen Du, and Chang-Tien Lu. Cross-domain few-shot semantic segmentation. In European Conference on Computer Vision, pages 73-90. Springer, 2022. 1, 2, 5, 7, 8 +[21] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 8 +[22] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2869–2878, 2020. 1, 5 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1 + +[24] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1 +[25] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8741-8750, 2021. 8 +[26] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 7, 8 +[27] Jiahao Nie, Yun Xing, Gongjie Zhang, Pei Yan, Aoran Xiao, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Cross-domain few-shot segmentation via iterative support-query correspondence mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3380-3390, 2024. 8 +[28] Alan V Oppenheim and Jae S Lim. The importance of phase in signals. Proceedings of the IEEE, 69(5):529-541, 1981. 4 +[29] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023. 8 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1 +[31] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. arXiv preprint arXiv:1804.10745, 2018. 8 +[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 8 +[33] Jiapeng Su, Qi Fan, Wenjie Pei, Guangming Lu, and Fanglin Chen. Domain-rectifying adapter for cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24036–24045, 2024. 1, 7, 8 +[34] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020. 8 +[35] Jintao Tong, Yixiong Zou, Yuhua Li, and Ruixuan Li. Lightweight frequency masker for cross-domain few-shot semantic segmentation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 7 +[36] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data, 5(1):1-9, 2018. 1, 5 + +[37] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. arXiv preprint arXiv:1903.06256, 2019. 8 +[38] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 8 +[39] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, 2021. 4 +[40] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. Advances in Neural Information Processing Systems, 34:21984-21996, 2021. 8 +[41] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in neural information processing systems, 35:6575-6588, 2022. 6, 7 +[42] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Yu Qiao, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. In The Twelfth International Conference on Learning Representations, 2024. 7 +[43] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1 +[44] Yixiong Zou, Shanghang Zhang, Jianpeng Yu, Yonghong Tian, and José MF Moura. Revisiting mid-level patterns for cross-domain few-shot recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 741-749, 2021. 2 +[45] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 7 +[46] Yixiong Zou, Yicong Liu, Yiman Hu, Yuhua Li, and Ruixuan Li. Flatten long-range loss landscapes for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23575-23584, 2024. 2, 3, 4 +[47] Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. Attention temperature matters in vit-based cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:116332-116354, 2024. 7 +[48] Yixiong Zou, Shuai Yi, Yuhua Li, and Ruixuan Li. A closer look at the clf token for cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:85523-85545, 2024. +[49] Yixiong Zou, Shanghang Zhang, Haichen Zhou, Yuhua Li, and Ruixuan Li. Compositional few-shot class-incremental learning. arXiv preprint arXiv:2405.17022, 2024. 7 \ No newline at end of file diff --git a/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/images.zip b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6266eef9c22e64bc1d71139c4116582c643a5726 --- /dev/null +++ b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d34104fdc7eb3f113aa51f0dcd6eaa19a962c42b331a9689c508670f0fe24bd3 +size 597349 diff --git a/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/layout.json b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b1945ef4f9b8d12b54a9b9608abf4309209b61c4 --- /dev/null +++ b/2025/The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation/layout.json @@ -0,0 +1,9173 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 542, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 542, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 542, + 121 + ], + "type": "text", + "content": "The Devil is in Low-Level Features for Cross-Domain Few-Shot Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 143, + 536, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 143, + 536, + 186 + ], + "spans": [ + { + "bbox": [ + 75, + 143, + 536, + 186 + ], + "type": "text", + "content": "Yuhan Liu, Yixiong Zou*, Yuhua Li, Ruixuan Li \nSchool of Computer Science and Technology, Huazhong University of Science and Technology \n{yuhan.liu, yixiongz, idcliyuhua, rxli}@hust.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 239, + 297, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 526 + ], + "type": "text", + "content": "Cross-Domain Few-Shot Segmentation (CDFSS) is proposed to transfer the pixel-level segmentation capabilities learned from large-scale source-domain datasets to downstream target-domain datasets, with only a few annotated images per class. In this paper, we focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, particularly those distant from the source domain, segmentation performance peaks at the very early epochs, and declines sharply as the source-domain training proceeds. We delve into this phenomenon for an interpretation: low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS. Based on this phenomenon and interpretation, we further propose a method that includes two plug-and-play modules: one to flatten the loss landscapes for low-level features during source-domain training as a novel sharpness-aware minimization method, and the other to directly supplement target-domain information to the model during target-domain testing by low-level-based calibration. Extensive experiments on four target datasets validate our rationale and demonstrate that our method surpasses the state-of-the-art method in CDFSS significantly by " + }, + { + "bbox": [ + 55, + 239, + 297, + 526 + ], + "type": "inline_equation", + "content": "3.71\\%" + }, + { + "bbox": [ + 55, + 239, + 297, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 239, + 297, + 526 + ], + "type": "inline_equation", + "content": "5.34\\%" + }, + { + "bbox": [ + 55, + 239, + 297, + 526 + ], + "type": "text", + "content": " average MIoU in 1-shot and 5-shot scenarios, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 551, + 135, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 551, + 135, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 551, + 135, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "type": "text", + "content": "Current deep neural networks [24, 30, 43] have achieved remarkable success in semantic segmentation, but their performance heavily relies on large-scale annotated datasets [1, 23]. However, the process of annotating data, particularly for dense pixel-wise tasks like semantic segmentation, is labor-intensive and time-consuming. To address this issue, Few-Shot Segmentation (FSS) has been introduced, aiming to generate pixel-level predictions for unseen categories with only a few labeled samples. Typically, the model is pretrained on a large-scale dataset of base classes and then trans" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 213, + 552, + 281 + ], + "blocks": [ + { + "bbox": [ + 315, + 213, + 552, + 281 + ], + "lines": [ + { + "bbox": [ + 315, + 213, + 552, + 281 + ], + "spans": [ + { + "bbox": [ + 315, + 213, + 552, + 281 + ], + "type": "image", + "image_path": "cf8eb221a693560d4bf08f9f725c935c7292d640d6a6392f760e32a52267057f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 282, + 432, + 346 + ], + "blocks": [ + { + "bbox": [ + 316, + 282, + 432, + 346 + ], + "lines": [ + { + "bbox": [ + 316, + 282, + 432, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 432, + 346 + ], + "type": "image", + "image_path": "6c5b81cfd8696020ec59dc2d51f6f2c8d9888fb2e4f152d57bc0a24558bb0134.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 350, + 415, + 362 + ], + "lines": [ + { + "bbox": [ + 345, + 350, + 415, + 362 + ], + "spans": [ + { + "bbox": [ + 345, + 350, + 415, + 362 + ], + "type": "text", + "content": "(b) Previous CDFSS methods Epoch vs. MoU" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 433, + 283, + 552, + 347 + ], + "blocks": [ + { + "bbox": [ + 433, + 283, + 552, + 347 + ], + "lines": [ + { + "bbox": [ + 433, + 283, + 552, + 347 + ], + "spans": [ + { + "bbox": [ + 433, + 283, + 552, + 347 + ], + "type": "image", + "image_path": "40838c235abca060f02be4da9ce2eeb0786f2329a147d16afe4fb82367e652cb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 349, + 518, + 361 + ], + "lines": [ + { + "bbox": [ + 469, + 349, + 518, + 361 + ], + "spans": [ + { + "bbox": [ + 469, + 349, + 518, + 361 + ], + "type": "text", + "content": "(c) Our method Epochs vs. MoU" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "type": "text", + "content": "Figure 1. (a) In CDFSS tasks, the training (source) and testing (target) datasets belong to distinct domains, with categories in the testing dataset being unseen during training. (b) Previous CDFSS methods show a decreasing trend of mIoU as the source-domain training proceeds, even at very early epochs for distant domains. (c) Our method can effectively prevent the model from performance decline after early epochs and achieve higher performance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "content": "ferred to unseen novel classes. However, novel classes might not be in the same domain as base classes [5, 7, 11, 22, 36], a more realistic approach is to consider domain gaps between the base (source domain) and novel (target domain) classes, which gives rise to the Cross-Domain Few-Shot Segmentation (CDFSS) task [20]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 689 + ], + "type": "text", + "content": "Although extensive works [16, 17, 20, 33] have been developed, a well-observed phenomenon is still not handled: for target domains, especially those distant from the source domain, the best performance is always achieved at the very early epochs or even the first epoch. As shown in Fig. 1b, the performance decreases sharply as the source-domain training proceeds, where the 20th epoch's MIoU is even lower than that of the first epoch. Although the early stop can handle this problem, the goal of the source-domain training is to provide a generalizable model for all target domains [20]. Therefore, it may be inappropriate to try different early stops for specific domains. In this paper, we aim to delve into this well-observed but under-explored phenomenon and handle it based on our interpretations of it, as shown in Fig. 1c." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "To delve into this phenomenon, we first visualize the predictions of the 1st and 20th epochs in Fig. 2. We observe that" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4618" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 293, + 150 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 150 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 150 + ], + "type": "image", + "image_path": "748fbdb6f1c71e584d3f0948b6c7aee178a067d4d6ae95b262cc997a89ac5e90.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 155, + 295, + 188 + ], + "lines": [ + { + "bbox": [ + 55, + 155, + 295, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 295, + 188 + ], + "type": "text", + "content": "Figure 2. The visualization of predictions at the 1st and 20th epoch further indicate that CDFSS models may not acquire meaningful information for target domains during source-domain training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 196, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 196, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 196, + 295, + 411 + ], + "type": "text", + "content": "the 20th epoch's model shows many fundamental errors in target domains, i.e., CDFSS models fail to identify the foreground but focus on entirely wrong regions. This suggests that the CDFSS models may not have acquired meaningful information for target domains during source-domain training, resulting in almost random predictions during testing. We then visualize feature maps at different network stages, and observe this problem occurs even at the shallow layers, and deeper layers produce wrong feature maps with wrong inputs from shallow layers. Quantitatively, previous works [13, 46] suggest that the challenge posed by domain shift can be analyzed through the flatness of the loss landscape. We use this as an entry point to explore the connection between shallow layers and early stops. We find that low-level features, although simpler and previously believed to be more transferable [44], are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training. That is, the devil lies in the low-level features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 412, + 295, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 580 + ], + "type": "text", + "content": "Based on this phenomenon and interpretation, we further propose a method for the CDFSS task. During the source-domain training, we propose a novel sharpness-aware minimization method to flatten the loss landscapes for low-level features, which is achieved by a shape-preserving perturbation module with random convolutions. During the target-domain phase, since low-level features can hardly capture the target-domain information due to its vulnerability to domain shifts, we directly supplement such information to the model with a low-level-based calibration module, which effectively prevents the model from fundamental errors caused by collapsed low-level features. Based on our methods, we can effectively prevent the model from dropping performances after early epochs and achieve higher performance (Fig. 1c)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 581, + 255, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 581, + 255, + 592 + ], + "spans": [ + { + "bbox": [ + 67, + 581, + 255, + 592 + ], + "type": "text", + "content": "In summary, our contributions can be listed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 594, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 664 + ], + "type": "text", + "content": "- We focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, particularly those distant from the source domain, the best segmentation performance is always achieved at the very early epochs, followed by a sharp decline as source-domain training progresses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 712 + ], + "type": "text", + "content": "- We delve into this phenomenon for an interpretation: low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "- Building on this interpretation, we propose a method that includes two plug-and-play modules: one act as a novel sharpness-aware minimization method to flatten the loss landscapes of low-level features during source-domain training, while the other directly supplement target-domain information to the model during target-domain testing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "content": "- Extensive experiments on four target datasets validate the rationale of our interpretation and method, showing its superiority over current state-of-the-art methods." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 190, + 403, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 403, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 403, + 204 + ], + "type": "text", + "content": "2. Interpretation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 210, + 400, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 400, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 400, + 221 + ], + "type": "text", + "content": "2.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "content": "Cross-Domain Few-Shot Segmentation (CDFSS) aims to transfer the segmentation capabilities learned from the source domain to target domains, with only a few annotated images per class. The source domain " + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s = (\\mathcal{X}_s,\\mathcal{Y}_s)" + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "content": " and target domain " + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t = (\\mathcal{X}_t,\\mathcal{Y}_t)" + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "content": " are defined by different input distributions and disjoint label spaces, i.e., " + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_s\\neq \\mathcal{X}_t" + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_s\\cap \\mathcal{Y}_t = \\emptyset" + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "content": " refers to input distribution and " + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "content": " refers to the label space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": "In this work, we adopt the meta-learning episodic paradigm following [20]. Specifically, both " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": " consist of several episodes. Each episode is constructed by a support set " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "S = \\{(I_s^i,M_s^i)\\}_{i = 1}^K" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": " training samples and a query set " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "Q = \\{(I_q,M_q)\\}" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": " denote the image and " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": " denote the pixel labels. Within each episode, the model leverages " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\{I_s,M_s\\}" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": " to predict the query mask " + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "inline_equation", + "content": "M_{q}" + }, + { + "bbox": [ + 313, + 324, + 554, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 410, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 553, + 445 + ], + "type": "text", + "content": "The model is trained on the source domain support sets and query set " + }, + { + "bbox": [ + 313, + 410, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\{S^s, Q^s\\}" + }, + { + "bbox": [ + 313, + 410, + 553, + 445 + ], + "type": "text", + "content": " by minimizing the binary cross-entropy loss:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 454, + 553, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 454, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 361, + 454, + 553, + 469 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = B C E \\left(F \\left(I _ {s} ^ {s}, I _ {q} ^ {s}, M _ {s} ^ {s}\\right), M _ {q} ^ {s}\\right), \\tag {1}", + "image_path": "88ed235c40669b70039374f64720c403568c50a09bdec4674221a160898859e7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "inline_equation", + "content": "F\\left(I_s^s,I_q^s,M_s^s\\right) = h\\left(f\\left(I_s^s,I_q^s\\right),M_s^s\\right)" + }, + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "text", + "content": " outputs the segmentation score map of " + }, + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "inline_equation", + "content": "I_{q}^{s}" + }, + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "text", + "content": ", which is composed of a feature extractor " + }, + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "text", + "content": " and a comparison module " + }, + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 313, + 477, + 553, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 514, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 553, + 540 + ], + "type": "text", + "content": "Then the model will be evaluated on the target domain query set " + }, + { + "bbox": [ + 313, + 514, + 553, + 540 + ], + "type": "inline_equation", + "content": "Q^t" + }, + { + "bbox": [ + 313, + 514, + 553, + 540 + ], + "type": "text", + "content": ", the prediction for " + }, + { + "bbox": [ + 313, + 514, + 553, + 540 + ], + "type": "inline_equation", + "content": "I_q^t" + }, + { + "bbox": [ + 313, + 514, + 553, + 540 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 370, + 548, + 553, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 548, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 370, + 548, + 553, + 564 + ], + "type": "interline_equation", + "content": "\\hat {M} _ {q} ^ {t} = \\arg \\max F \\left(I _ {s} ^ {t}, I _ {q} ^ {t}, M _ {s} ^ {t}\\right) \\tag {2}", + "image_path": "dfde9ee074c4cb461de89ae1effa0a3a0976f3ae5cd0dd0fbc20f5f64c2a5b49.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 572, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 553, + 598 + ], + "type": "text", + "content": "Finally, the evaluation can be conducted based on the comparison between " + }, + { + "bbox": [ + 313, + 572, + 553, + 598 + ], + "type": "inline_equation", + "content": "\\hat{M}_q^t" + }, + { + "bbox": [ + 313, + 572, + 553, + 598 + ], + "type": "text", + "content": " and the real label " + }, + { + "bbox": [ + 313, + 572, + 553, + 598 + ], + "type": "inline_equation", + "content": "M_q^t" + }, + { + "bbox": [ + 313, + 572, + 553, + 598 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 604, + 553, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 553, + 628 + ], + "type": "text", + "content": "2.2. Delve into the Early Stop in Cross-Domain Few-Shot Segmentation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 634, + 553, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 553, + 658 + ], + "type": "text", + "content": "2.2.1 Intuitive observation: The devil is in the low-level features" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Given our suspicion that the model fails to learn useful information from the source domain for target domains, we start by visualizing the feature maps. Using ResNet-50 [15] as the backbone following [12], we visualize the feature maps" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4619" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 69, + 272, + 247 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 272, + 247 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 272, + 247 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 272, + 247 + ], + "type": "image", + "image_path": "b3318b69b17f2be6d6b9e2082fac3c9c7927ff459e5ecb4009fac1c044c7599f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 296, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 297 + ], + "type": "text", + "content": "Figure 3. Feature maps from stages 1 to 4 for both the source domain and target domains. The noticeable contrast in the feature maps at stage 1 between source and target domains indicates the limited performance of CDFSS models stems from shallow layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 301, + 293, + 363 + ], + "blocks": [ + { + "bbox": [ + 56, + 301, + 293, + 363 + ], + "lines": [ + { + "bbox": [ + 56, + 301, + 293, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 293, + 363 + ], + "type": "image", + "image_path": "561a7f18308d7481b384c7e071df76c95db513c6be2b7bf79cb6ed55c51ff6cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 367, + 295, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 411 + ], + "type": "text", + "content": "Figure 4. Feature maps from stages 1 to 4 for epoch 1 and epoch 20. Epoch 1 in stage 1 shows more distinguishable activations than epoch 20, indicating that low-level features gradually incorporate incorrect information as training progresses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 416, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 416, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 416, + 295, + 548 + ], + "type": "text", + "content": "of stages 1 to 4 for both the source domain and the target domains, as shown in Fig. 3. We observe that on source-domain samples, the model learns boundary information effectively in stage 1, enabling it to accurately focus on the foreground in stage 4. However, on target domains especially those distant from the source domain, even though the foreground is clear and prominent, the model still fails to learn anything meaningful in stages 1 and 2. Consequently, in stage 4, the model focuses on completely incorrect areas. This indicates that the model's bad performance may originate from its shallow layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 548, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 643 + ], + "type": "text", + "content": "In Fig.4, we further visualize feature maps of each stage at the 1st and 20th epoch, respectively. We find that the shallow layers show more distinguishable activations at epoch 1 than at epoch 20, which aligns with our observation of higher accuracy at epoch 1. Consequently, we conclude that the poor performance of CDFSS models may be attributed to shallow layers that fail to capture useful low-level information, a trend that intensifies during source-domain training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "type": "text", + "content": "2.2.2 Quantitative verification: Low-level features lead to a sharper loss landscape" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Previous work [13, 46] has demonstrated that the sharpness of the loss landscape can serve as a tool to analyze cross" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 337, + 70, + 541, + 143 + ], + "blocks": [ + { + "bbox": [ + 317, + 104, + 329, + 113 + ], + "lines": [ + { + "bbox": [ + 317, + 104, + 329, + 113 + ], + "spans": [ + { + "bbox": [ + 317, + 104, + 329, + 113 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 70, + 541, + 143 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 541, + 143 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 541, + 143 + ], + "type": "image", + "image_path": "f6a00bcf4994f71fed9b420510b09808e4b38cfeed046a59527475ad05cc595b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 337, + 145, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 318, + 152, + 329, + 160 + ], + "lines": [ + { + "bbox": [ + 318, + 152, + 329, + 160 + ], + "spans": [ + { + "bbox": [ + 318, + 152, + 329, + 160 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 145, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 337, + 145, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 337, + 145, + 547, + 175 + ], + "type": "image", + "image_path": "8deb8f3d8d307ec6b234a9c5a7f5654876cbe8d86fe72babe057b93515b738b7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 334, + 178, + 435, + 254 + ], + "blocks": [ + { + "bbox": [ + 318, + 202, + 328, + 211 + ], + "lines": [ + { + "bbox": [ + 318, + 202, + 328, + 211 + ], + "spans": [ + { + "bbox": [ + 318, + 202, + 328, + 211 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 178, + 435, + 254 + ], + "lines": [ + { + "bbox": [ + 334, + 178, + 435, + 254 + ], + "spans": [ + { + "bbox": [ + 334, + 178, + 435, + 254 + ], + "type": "image", + "image_path": "7c1206a13d5519682b3ffc20def6075c7c929eef76c8c0e3bc7a3aed26238480.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 261, + 554, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 261, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 554, + 338 + ], + "type": "text", + "content": "Figure 5. (a) A sharp minimum in the landscape corresponds to a representation that is highly sensitive to data shifts. (b) Examples of pixel perturbation applied to the images. (c) As training proceeds, the loss landscape becomes progressively sharper. (d) Perturbing shallow layers leads to much sharper loss landscapes, indicating shallow layers are the cause of the sharp loss landscape and the increased sensitivity to domain shifts." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 445, + 178, + 550, + 254 + ], + "blocks": [ + { + "bbox": [ + 436, + 202, + 455, + 212 + ], + "lines": [ + { + "bbox": [ + 436, + 202, + 455, + 212 + ], + "spans": [ + { + "bbox": [ + 436, + 202, + 455, + 212 + ], + "type": "text", + "content": "(d) 0.3" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 445, + 178, + 550, + 254 + ], + "lines": [ + { + "bbox": [ + 445, + 178, + 550, + 254 + ], + "spans": [ + { + "bbox": [ + 445, + 178, + 550, + 254 + ], + "type": "image", + "image_path": "311d767e60f852abe75d3366fa3cb4b8185ccd73533c7d0f1d6e9a089318d5f1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 346, + 553, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 553, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 553, + 394 + ], + "type": "text", + "content": "domain issues. As illustrated in Fig. 5a, a smoother loss landscape enables the model to better handle domain shifts1. Therefore, we measure the sharpness of the source-domain-trained models against different epochs and stages." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 394, + 554, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 573 + ], + "type": "text", + "content": "We first measure the sharpness of the loss landscape across different epochs. Following the method in [46], we apply pixel perturbations to the training data. Examples of these perturbations are shown in Fig. 5b, while the magnitude of performance drop is depicted in Fig.5c. A larger performance drop under the same perturbation indicates a sharper loss landscape. As observed, the performance of models across all epochs drops when exposed to perturbations, but the model from epoch 1 experiences a smaller drop than that of epoch 10, which in turn performs better than epoch 20. This suggests that as training progresses, the loss landscape becomes progressively sharper, making the model increasingly vulnerable to domain shifts and resulting in decreased performance on target domains. This is in line with our previous observations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 574, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 554, + 693 + ], + "type": "text", + "content": "Based on this, we further investigate the relationship between model layers and the loss landscape. We apply the same magnitude of perturbations to the feature maps at stages 1-4 and measure the performance drops. The results in Fig.5d show that earlier stages exhibit larger performance drops under the same perturbation, which implies that shallow layers lead to sharper loss landscapes and are more vulnerable to domain shifts. This experiment provides quantitative validation for the intuitive observation from the low-level feature visualization in the previous section." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 538, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 538, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 538, + 713 + ], + "type": "text", + "content": "1 Please refer to the appendix for more details about the sharpness." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4620" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 118, + 296, + 156 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "text", + "content": "Table 1. The model with fixed shallow layers performs better than the one with trained shallow layers, implying that shallow layers learn domain-specific information that negatively impacts generalization to the target domain." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 118, + 296, + 156 + ], + "lines": [ + { + "bbox": [ + 58, + 118, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 296, + 156 + ], + "type": "table", + "html": "
MethodFSS-1000Deep globeISICChestX-rayAverage
train stage1, 2, 3, 478.8639.4435.7672.4956.64
fix stage1, train stage2, 3, 478.8839.9037.0072.1256.98
fix stage1, 2, train stage3, 478.9140.0035.4974.4457.21
", + "image_path": "f37571de1d0e02774182f985fdc249cbb027bf79369f064f7d269f7820db704a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 167, + 175, + 242 + ], + "blocks": [ + { + "bbox": [ + 56, + 167, + 175, + 242 + ], + "lines": [ + { + "bbox": [ + 56, + 167, + 175, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 175, + 242 + ], + "type": "image", + "image_path": "6944e8d5bff1995093a9884b632c31c2c9f2f1669082b7f3bd4b184d1e138d6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 244, + 296, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 289 + ], + "type": "text", + "content": "Figure 6. The mIoU curves over epochs for two different training settings indicate that shallow layers are more vulnerable to domain shifts, leading to more severe performance degradation during source-domain training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 167, + 294, + 242 + ], + "blocks": [ + { + "bbox": [ + 177, + 167, + 294, + 242 + ], + "lines": [ + { + "bbox": [ + 177, + 167, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 177, + 167, + 294, + 242 + ], + "type": "image", + "image_path": "f1afcf28068f52da55ec1ac2f926012cbd4f675c898488ed395f3e9af2f98afc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 296, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 329 + ], + "type": "text", + "content": "2.2.3 Further analysis: Low-level features absorb domain-specific information" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 402 + ], + "type": "text", + "content": "To investigate the impact of shallow layers on model performance, we compare the results of fixing stage 1 and stage 2 using ImageNet [8] pre-trained weights with the results of training all stages, as shown in Tab. 1. We find that fixing stage 1 and stage 2 leads to the best performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 404, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 466 + ], + "type": "text", + "content": "In Fig 6, we further visualize the mIoU curves over epochs for two different training settings: fix stage 1 only, and fix stage 1, 2. The comparison revealed that when stage 2 is fixed, the model reaches the optimal mIoU at a later epoch, and as training progresses, the mIoU decreases more slowly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "content": "This suggests that shallow layers are vulnerable to domain shifts, therefore any slight absorption of domain information could harm target-domain generalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 204, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 204, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 204, + 533 + ], + "type": "text", + "content": "2.3. Conclusion and Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 542, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 662 + ], + "type": "text", + "content": "Based on the above experiments, we make the interpretation as follows. The shallow layers of the model tend to learn domain-specific information, causing the low-level features to become increasingly domain-specific during training, gradually overfitting to the source domain. The overfitting sharpens the model's loss landscape and makes it more vulnerable to domain shifts, thereby reducing cross-domain performance and introducing significant mistakes in CDFSS. This problem is partially addressed by the early-stop strategy and freezing shallow layers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "However, merely stopping the training of shallow layers cannot fully unlock the power of source-domain training. Therefore, below we propose a novel method to handle the problem in shallow layers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 91, + 421, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 421, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 421, + 102 + ], + "type": "text", + "content": "3.1. Method Overview" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "text", + "content": "Our core idea is that low-level features are vulnerable to domain shifts, which may lead to a sharp decrease in generalization with only slight absorption of source-domain information. Based on this, we designed two plug-and-play modules to operate in two phases: 1) During the source-domain training phase, the Low-level Enhancing Module enhances the robustness against domain shifts by synthesizing diverse domains and incorporating them into low-level features. 2) During the target-domain testing phase, the Low-level Calibration Module refines previous segmentation results by utilizing the low-level features from the target domain. The framework is illustrated in Fig. 7." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 260, + 528, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 528, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 528, + 273 + ], + "type": "text", + "content": "3.2. Train-time Low-level Enhancing Module" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "type": "text", + "content": "As analyzed above, overfitting of low-level features on the source domain during training leads to a sharp loss landscape, making the model vulnerable to domain shifts. According to the previous Sharpness Aware Minimization (SAM) method [13, 46], one effective strategy to smooth loss landscapes is to introduce perturbations. Thus we transform the domain of the low-level support feature during training, which acts as a form of perturbation to reduce overfitting." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": "Previous work [39] has shown that random convolution is a shape-preserving strategy to distort local textures, which we view to be suitable for the segmentation task. We take advantage of this property to generate diverse domains while retaining the content. Specifically, given support feature " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "F_{s} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": " from the early layers of the encoder and a convolution layer with filters " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "\\Theta \\in \\mathbb{R}^{h \\times w \\times C \\times C}" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": " represent height, width and channels of the input feature, and " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": " are the height and width of the random convolution filter. We sample the filter weights from " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "N(0, \\sigma^2)" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": " acts as a hyperparameter to control the perturbation magnitude. Then we get perturbed support feature " + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "inline_equation", + "content": "F_{s}'" + }, + { + "bbox": [ + 313, + 375, + 556, + 530 + ], + "type": "text", + "content": " with transformed domain:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 405, + 532, + 553, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 532, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 405, + 532, + 553, + 545 + ], + "type": "interline_equation", + "content": "F _ {s} ^ {\\prime} = F _ {s} * \\Theta , \\tag {3}", + "image_path": "10f05a8e2c008f0ae14bf441ff765ba94c5fad8dfb5dabbf433a99a06e07bd62.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "text", + "content": "Although random convolution can generally preserve the shape and semantics of the feature, some edge details inevitably get blurred. Existing research [28] has shown that the Fast Fourier Transformation (FFT) decomposes a signal into amplitude spectrum and phase spectrum, with amplitude capturing the overall domain and texture information and phase representing shape, edge, and spatial structure. Thus, we apply FFT to both " + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "inline_equation", + "content": "F_{s}^{\\prime}" + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "text", + "content": ", decomposing them into phase spectrum " + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\prime}" + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "text", + "content": " and amplitude spectrum " + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\prime}" + }, + { + "bbox": [ + 313, + 548, + 555, + 655 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 392, + 658, + 553, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 658, + 553, + 672 + ], + "spans": [ + { + "bbox": [ + 392, + 658, + 553, + 672 + ], + "type": "interline_equation", + "content": "\\mathcal {A} e ^ {i \\mathcal {P}} = F F T \\left(F _ {s}\\right), \\tag {4}", + "image_path": "e37ecab99e1c8adc1ec1d104aa1b51673651ecb10716164bdae7b88b60ae8e5e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 390, + 673, + 553, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 673, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 390, + 673, + 553, + 689 + ], + "type": "interline_equation", + "content": "\\mathcal {A} ^ {\\prime} e ^ {i \\mathcal {P} ^ {\\prime}} = F F T \\left(F _ {s} ^ {\\prime}\\right), \\tag {5}", + "image_path": "318fec7e3ab1d465fb94ff3d7d982d2967903a08e5c481cd5520de63824ec3f1.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "2The low-level support and query features in LEM and LCM are illustrated as images for better understanding." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4621" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 71, + 492, + 250 + ], + "blocks": [ + { + "bbox": [ + 115, + 71, + 492, + 250 + ], + "lines": [ + { + "bbox": [ + 115, + 71, + 492, + 250 + ], + "spans": [ + { + "bbox": [ + 115, + 71, + 492, + 250 + ], + "type": "image", + "image_path": "c53872451b8557a2334c02b7e4c72ae6566938b49ceacdddaeafc7874a081cb3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 257, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 555, + 313 + ], + "type": "text", + "content": "Figure 7. The overall architecture of our method in a 1-way 1-shot example. Our method consists of two modules: the Low-level Enhancing Module (LEM) and the Low-level Calibration Module (LCM). Support and query images are fed into a weight-shared backbone to extract features, and then processed by a comparison module to obtain a coarse segmentation mask. During source-domain training, LEM perturbs low-level support features for sharpness-aware minimization. During target-domain testing, LCM calibrates the coarse segmentation mask by directly supplementing collapsed low-level target-domain information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "type": "text", + "content": "Then we recombine the phase from the original feature " + }, + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "type": "text", + "content": " and the amplitude from the perturbed feature " + }, + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "type": "inline_equation", + "content": "F_{s}^{\\prime}" + }, + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "type": "text", + "content": ", using the Inverse Fast Fourier Transform (IFFT) to generate transformed support feature:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 368, + 294, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 368, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 127, + 368, + 294, + 383 + ], + "type": "interline_equation", + "content": "F _ {s} ^ {t} = I F F T \\left(\\mathcal {A} ^ {\\prime} e ^ {i \\mathcal {P}}\\right), \\tag {6}", + "image_path": "4e781b60119014eb032e5f38431e1b55932f609658448dfd548e22c84fcb19c1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 296, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 444 + ], + "type": "text", + "content": "Therefore, while transforming the domain, we achieve better preservation of the boundary shapes, which achieves domain-oriented perturbations on low-level features for sharpness-aware minimization. The transformed feature is then fed into the subsequent layers of the encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 265, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 265, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 265, + 462 + ], + "type": "text", + "content": "3.3. Test-time Low-level Calibration Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 469, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 564 + ], + "type": "text", + "content": "After being processed by the encoder, the support and query features are fed into the comparison module, yielding a coarse score map " + }, + { + "bbox": [ + 55, + 469, + 296, + 564 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 55, + 469, + 296, + 564 + ], + "type": "text", + "content": ". Since low-level features may be collapsed on target domains due to the vulnerability to domain shifts, we directly use the low-level query feature " + }, + { + "bbox": [ + 55, + 469, + 296, + 564 + ], + "type": "inline_equation", + "content": "F_{q} \\in \\mathbb{R}^{H' \\times W' \\times C}" + }, + { + "bbox": [ + 55, + 469, + 296, + 564 + ], + "type": "text", + "content": " to supplement such collapsed low-level target-domain information to calibrate the score map during testing on the target domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "text", + "content": "First, we compute a confidence map " + }, + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "text", + "content": " by subtracting the background similarity from the foreground similarity in the score map:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 601, + 294, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 601, + 294, + 614 + ], + "spans": [ + { + "bbox": [ + 129, + 601, + 294, + 614 + ], + "type": "interline_equation", + "content": "C _ {i, j} = S _ {i, j, 1} - S _ {i, j, 0}, \\tag {7}", + "image_path": "74d96355e404dd9a6c3c9c8fb7f8a59f6c4e28d651338bad2c1d0e0c44c251d3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "inline_equation", + "content": "S_{i,j,1}" + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "inline_equation", + "content": "S_{i,j,0}" + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": " denote the foreground and background similarity for the pixel at " + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": ", respectively. The confidence map " + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": " quantifies the likelihood of foreground at each pixel." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Then we partition the confidence map into patches " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{P_1, P_2, \\ldots, P_n\\}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " and select the Top-K patches with the highest average confidence. These patches are considered the most reliable foreground regions:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 316, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 316, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 378, + 316, + 553, + 346 + ], + "type": "interline_equation", + "content": "\\bar {C} _ {p} = \\frac {1}{| P _ {p} |} \\sum_ {(i, j) \\in P _ {p}} C (i, j), \\tag {8}", + "image_path": "26a2e009ebc7e91ee02b0048f4eef966b21e92f9e06b419be1dd235f3b14b9c3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 346, + 553, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 346, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 361, + 346, + 553, + 363 + ], + "type": "interline_equation", + "content": "\\left\\{P _ {k _ {1}}, P _ {k _ {2}}, \\dots , P _ {k _ {K}} \\right\\} = \\arg \\max _ {P _ {p}} \\bar {C} _ {p} \\tag {9}", + "image_path": "27f7df024c4201d93dbf2245ada05ef1b480a078bc70b0e53b074ffd29646e09.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "inline_equation", + "content": "|P_{p}|" + }, + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "text", + "content": " is the number of pixels in patch " + }, + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "inline_equation", + "content": "P_{p}" + }, + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 361, + 553, + 385 + ], + "type": "text", + "content": " are the pixel locations within the patch." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "text", + "content": "For each of the selected Top-K foreground patches, we locate the corresponding patch in the interpolated query feature map " + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "inline_equation", + "content": "F_{q}^{\\prime} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\{F_{k_1}, F_{k_2}, \\ldots, F_{k_K}\\}" + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "text", + "content": ", and compute the cosine similarity between it and all other patches in the query feature map:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 442, + 553, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 442, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 370, + 442, + 553, + 468 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} \\left(F _ {k _ {i}}, F _ {m}\\right) = \\frac {F _ {k _ {i}} \\cdot F _ {m}}{\\| F _ {k _ {i}} \\| \\| F _ {m} \\|} \\tag {10}", + "image_path": "52d367933477e086ab144dc91d191685a5338b2704ab1ac17e1e27e6a3d588ab.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "content": "Finally, the foreground similarity in the score map is updated using the top-K similarity maps:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 331, + 491, + 553, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 491, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 331, + 491, + 553, + 523 + ], + "type": "interline_equation", + "content": "S _ {i, j, 1} ^ {\\prime} = S _ {i, j, 1} + \\sum_ {k = 1} ^ {K} w _ {k} \\cdot \\left(S i m _ {k} (i, j) - \\beta_ {k}\\right), \\tag {11}", + "image_path": "68c992906a3aa61332b2b5aa33888d88c2511c296ce62cd929a80f3437737a33.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "type": "text", + "content": " is a scaling factor and " + }, + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\beta_{k}" + }, + { + "bbox": [ + 313, + 523, + 554, + 559 + ], + "type": "text", + "content": " is a bias term. This adjustment improves foreground detection by directly supplementing collapsed low-level target-domain information." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 559, + 554, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 607 + ], + "type": "text", + "content": "During the source-domain training phase, the model will be trained by minimizing the loss in Eq. 1. During the target-domain testing phase, the prediction will be obtained based on Eq. 2 and LCM." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 635, + 376, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 376, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 376, + 646 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "Following the settings in [20], we use PASCAL VOC 2012 [11] with SBD augmentation [14] as source domain for training. Then we evaluate the trained models on four target domains: FSS-1000 [22], Deep globe [7], ISIC [5, 36], and Chest X-ray [3, 18]. See Appendix for more details." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4622" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 70, + 243, + 209 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 243, + 209 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 243, + 209 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 243, + 209 + ], + "type": "image", + "image_path": "ab0f1e8656c197d40a89316c0a1b56dfaf85408ac5abc3056fa9a5a2c646f866.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 214, + 284, + 225 + ], + "lines": [ + { + "bbox": [ + 66, + 214, + 284, + 225 + ], + "spans": [ + { + "bbox": [ + 66, + 214, + 284, + 225 + ], + "type": "text", + "content": "Figure 8. Qualitative results of our model for 1-shot setting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 188, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 188, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 188, + 248 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": "We employ ResNet-50 [15] with ImageNet [8] pretrained weights as our backbone. Following baseline SSP [12], we discard the last backbone stage and the last ReLU for better generalization. Furthermore, we integrated our approach into the transformer architecture by using ViT-B/16 [9] as the backbone, following FPTrans [41]. Consistent with FPTrans [41], we resize both support and query images to " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": ". During training on the source domain, we use SGD to optimize our model, with a momentum of 0.9 and a initial learning rate of 1e-3. We apply our LEM into different shallow layers of the backbone, more details can be seen in 4.4.4. The random convolution filter used in LEM has a kernel size of " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": ", and the standard deviation " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " is set to be 0.1. During testing on the target domain, we apply our LCM to refine segmentation mask. The hyperparameters " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " are set to 3, 0.6 and 0.7, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 455, + 278, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 278, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 278, + 468 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 606 + ], + "type": "text", + "content": "In Tab. 2, we compare our method with CNN-based and ViT-based approaches, including traditional FSS methods and existing CDFSS methods. Our results illustrate a notable enhancement for both 1-shot and 5-shot tasks. Specifically, our method outperforms the state-of-the-art APSeg [16] by " + }, + { + "bbox": [ + 55, + 474, + 296, + 606 + ], + "type": "inline_equation", + "content": "3.71\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 474, + 296, + 606 + ], + "type": "inline_equation", + "content": "5.34\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 606 + ], + "type": "text", + "content": " in the 1-shot and 5-shot settings respectively, confirming the effectiveness of our strategy. Additionally, we apply our method on existing CDFSS baselines and observe improved performance. Please refer to the appendix for more details. We present qualitative results of our method in 1-way 1-shot segmentation in Fig.8." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 615, + 149, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 149, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 149, + 628 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 634, + 230, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 230, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 230, + 645 + ], + "type": "text", + "content": "4.4.1 Effectiveness of Each Component" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "We evaluate each proposed component on both CNN and ViT baselines to assess the effectiveness of our designs, including LEM and LCM. The results presented in Tab. 3 show that in the 1-shot setting, introducing LEM improved average MIoU by " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.14\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " for the CNN baseline and " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.89\\%" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 432, + 144 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 432, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 432, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 432, + 144 + ], + "type": "image", + "image_path": "a27ecdc96cd5cad3ef4912c796eb82177c315688a64a7b8eb3286b2ad613e08e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 152, + 554, + 184 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 554, + 184 + ], + "type": "text", + "content": "Figure 9. Applying LEM leads to reduced loss landscape sharpness and improved robustness to domain shifts, Fourier transformation further enhances these benefits." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 72, + 551, + 145 + ], + "blocks": [ + { + "bbox": [ + 434, + 72, + 551, + 145 + ], + "lines": [ + { + "bbox": [ + 434, + 72, + 551, + 145 + ], + "spans": [ + { + "bbox": [ + 434, + 72, + 551, + 145 + ], + "type": "image", + "image_path": "c51bdc8bc22b9c61eae53e5bf479796c90a90e87d02da12c54c03d027ff1cec1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 190, + 433, + 260 + ], + "blocks": [ + { + "bbox": [ + 315, + 190, + 433, + 260 + ], + "lines": [ + { + "bbox": [ + 315, + 190, + 433, + 260 + ], + "spans": [ + { + "bbox": [ + 315, + 190, + 433, + 260 + ], + "type": "image", + "image_path": "cc344316c5d85db3bc6d869cb0cabccd83b357c0c48990d8dbf45bfa23818bc7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "type": "text", + "content": "Figure 10. (Left) The mIoU of Chest-X Ray with and without LEM, confirms that LEM leads to consistent and steady performance improvement as training advances. (Right) After incorporating our LEM module, training shallow layers can achieve performance comparable to fixing them." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 436, + 190, + 551, + 259 + ], + "blocks": [ + { + "bbox": [ + 436, + 190, + 551, + 259 + ], + "lines": [ + { + "bbox": [ + 436, + 190, + 551, + 259 + ], + "spans": [ + { + "bbox": [ + 436, + 190, + 551, + 259 + ], + "type": "image", + "image_path": "bdc1a79db33df37358ae6fd9ec547615aa566ded82163897d29a409b8f0dbf02.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 328, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 554, + 376 + ], + "type": "text", + "content": "for the ViT baseline, while adding LCM further increased it by " + }, + { + "bbox": [ + 313, + 328, + 554, + 376 + ], + "type": "inline_equation", + "content": "2.11\\%" + }, + { + "bbox": [ + 313, + 328, + 554, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 328, + 554, + 376 + ], + "type": "inline_equation", + "content": "1.95\\%" + }, + { + "bbox": [ + 313, + 328, + 554, + 376 + ], + "type": "text", + "content": ", respectively. These results clearly demonstrate the effectiveness of each component in enhancing performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 392, + 433, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 433, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 433, + 403 + ], + "type": "text", + "content": "4.4.2 Verification of LEM" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 410, + 555, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 541 + ], + "type": "text", + "content": "LEM smooths loss landscapes. To assess the impact of LEM, we measured the loss landscape sharpness before and after its application. We introduced low-frequency noise on training data to simulate domain shifts in the representation space (i.e., pixels and features). A larger performance drop suggests a sharper landscapes. As illustrated in Fig.9, random convolution smooths the loss landscape compared with the baseline, while Fourier transformation further enhances this effect. These results confirm that LEM effectively smooths the loss landscape, which improves robustness against domain shifts." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 542, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 554, + 637 + ], + "type": "text", + "content": "LEM ensures performance improves with training. We compared mIoU trends over training epochs with and without LEM in Fig.10 (left). In the baseline model, mIoU on the target domain declines as it increasingly absorbs source-domain information during training. Conversely, with LEM, mIoU not only outperforms the baseline at each epoch but also keeps increasing as training advances, indicating that the model focuses more on domain-invariant information." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 638, + 554, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 554, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 554, + 686 + ], + "type": "text", + "content": "LEM prevents the shallow layers from absorbing excessive source-domain information. As previously analyzed, fixing stage 1 and stage 2 of the baseline resulted in the best performance, which suggests that shallow layers tend" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "text", + "content": "3We present the validation of CNN-based methods here, and similar trends hold for ViT-based methods. Please refer to the appendix for details." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4623" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 96, + 524, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 2. MioU of 1-shot and 5-shot results compared with previous FSS and CD-FSS methods. All the methods are trained on PASCAL and tested on the CD-FSS. The best performance among all methods is highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 96, + 524, + 285 + ], + "lines": [ + { + "bbox": [ + 86, + 96, + 524, + 285 + ], + "spans": [ + { + "bbox": [ + 86, + 96, + 524, + 285 + ], + "type": "table", + "html": "
MethodMarkBackboneFSS-1000Deep globeISICChest X-rayAverage
1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
Few-Shot Segmentation Methods
RePRI [2]CVPR-21Res-5070.9674.2325.0327.4123.2726.2365.0865.4846.0948.34
HSNet [26]ICCV-21Res-5077.5380.9929.6535.0831.2035.1051.8854.3647.5751.38
SSP [12]ECCV-22Res-5078.9180.5940.0048.6835.4945.8674.4474.2657.2162.35
FPTrans [41]NIPS-22ViT-base80.7483.6538.3649.3048.6560.3780.9282.9162.1769.06
PerSAM [42]ICLR-24ViT-base60.9266.5336.0840.6523.2725.3329.9530.0537.5640.64
Cross-Domain Few-Shot Segmentation Methods
PATNet [20]ECCV-22Res-5078.5981.2337.8942.9741.1653.5866.6170.2056.0661.99
APM [35]NIPS-24Res-5079.2981.8340.8644.9241.7151.1678.2582.8160.0365.18
ABCDFSS [17]CVPR-24Res-5074.6076.2042.6049.0045.7053.3079.8081.4060.6764.97
DRA [33]CVPR-24Res-5079.0580.4041.2950.1240.7748.8782.3582.3160.8665.42
APSeg [16]CVPR-24ViT-base79.7181.9035.9439.9845.4353.9884.1084.5061.3065.09
OursOursRes-5078.5180.6044.1049.6738.2147.0481.0282.7360.4665.01
OursOursViT-base81.0583.6942.1251.4852.9162.4383.9484.1265.0170.43
", + "image_path": "99b3c0cee814c7470f0f5fd498ae9c363e8417c673d62c89c1f4455efe80db5c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 116, + 312, + 237, + 372 + ], + "blocks": [ + { + "bbox": [ + 94, + 297, + 256, + 308 + ], + "lines": [ + { + "bbox": [ + 94, + 297, + 256, + 308 + ], + "spans": [ + { + "bbox": [ + 94, + 297, + 256, + 308 + ], + "type": "text", + "content": "Table 3. Ablation study on key components." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 116, + 312, + 237, + 372 + ], + "lines": [ + { + "bbox": [ + 116, + 312, + 237, + 372 + ], + "spans": [ + { + "bbox": [ + 116, + 312, + 237, + 372 + ], + "type": "table", + "html": "
LEMLCMResNetViT
57.2162.17
✓58.3563.06
✓59.7864.39
✓✓60.4665.01
", + "image_path": "a32a74e1d664bc25dea4da807b6b9e252ae51b74f2f33ea363ad5fa95a2f1435.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 379, + 293, + 449 + ], + "blocks": [ + { + "bbox": [ + 57, + 379, + 293, + 449 + ], + "lines": [ + { + "bbox": [ + 57, + 379, + 293, + 449 + ], + "spans": [ + { + "bbox": [ + 57, + 379, + 293, + 449 + ], + "type": "image", + "image_path": "de25aa8f7ec34f7938fbcacba3aadb58b08e8edaba64d5ceb0c0cdf449a46cba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 455, + 295, + 500 + ], + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 500 + ], + "type": "text", + "content": "Figure 11. (Left) The CKA similarity between source and target domain with and without LEM, verifies that LEM helps the model to focus more on domain-agnostic information. (Right) Feature map visualization of stage 1 and stage 4 at epoch 1 and epoch 20." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "text", + "content": "to learn information that negatively impacts target-domain generalization. Fig.10 (right) shows that after incorporating our LEM, training stage 2 produced results comparable to those achieved by fixing stage 2. This indicates that LEM effectively prevents the shallow layers from absorbing excessive source-domain information during the training process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "LEM encourages the model to focus more on domain-agnostic information. To assess how the domain distance between the source and target domains changes before and after applying LEM, we use CKA (Centered Kernel Alignment) similarity for measurement, following the approach in [6, 45, 47-49]. Specifically, we extract features from source domain and target domain images using a backbone network, then compute the CKA similarity of the final layer's feature by aligning the channel dimension. A higher CKA similarity indicates a smaller domain distance, suggesting the model retains less domain-specific information. As shown" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 299, + 432, + 392 + ], + "blocks": [ + { + "bbox": [ + 319, + 299, + 432, + 392 + ], + "lines": [ + { + "bbox": [ + 319, + 299, + 432, + 392 + ], + "spans": [ + { + "bbox": [ + 319, + 299, + 432, + 392 + ], + "type": "image", + "image_path": "f165ed823f9ff79e6b064d4ca0bc7a3b2df2272a7393aa68036276151d149d8d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 399, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 422 + ], + "type": "text", + "content": "Figure 12. LCM reduces the sharpness of the loss landscape and enhances robustness against domain shifts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 436, + 299, + 549, + 392 + ], + "blocks": [ + { + "bbox": [ + 436, + 299, + 549, + 392 + ], + "lines": [ + { + "bbox": [ + 436, + 299, + 549, + 392 + ], + "spans": [ + { + "bbox": [ + 436, + 299, + 549, + 392 + ], + "type": "image", + "image_path": "a60ade4284c7152bf030a69f61935a0f5efd9337a3e42797bbf473022551ff2c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 428, + 555, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 500 + ], + "type": "text", + "content": "in Fig.11 (left), the average CKA similarity across four target domains increases after applying random convolution, and is further improved with the addition of the Fourier transform. This demonstrates that by perturbing the source domain, LEM encourages the model to capture domain-agnostic information during training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 500, + 554, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 554, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 554, + 573 + ], + "type": "text", + "content": "LEM enables the model to capture more useful low-level information in shallow layers. As shown in Fig.11 (right), LEM helps the model learn boundary information more effectively in stage 1, allowing it to focus more accurately on the foreground in stage 4. This leads to more precise activation regions in the later epochs of source domain training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 434, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 434, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 434, + 597 + ], + "type": "text", + "content": "4.4.3 Verification of LCM" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "LCM smooths loss landscapes. To assess the impact of LCM in reducing sharpness and enhancing robustness against domain shifts, we superimpose styles from ISIC medical images and Deep globe remote sensing images onto source domain images to simulate domain shifts, with the level of style alteration serving as the perturbation level. As shown in Fig.12, the lowered performance drop validates that LCM smooths the loss landscape, thus improves robustness against domain shifts." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4624" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 170, + 152 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 170, + 152 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 170, + 152 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 170, + 152 + ], + "type": "image", + "image_path": "45ab3e2c57d2b66948055dfb24e5e26617f1f745898cf8b430bea4b10691e3ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 158, + 295, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 204 + ], + "type": "text", + "content": "Figure 13. The effect of LEM and LCM at different positions in the shallow layers. Block ID represents the sequence in which the module is applied: before stage 1, stage 1 block 1, stage 2 block 1, stage 2 block 2, and stage 2 block 3." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 71, + 284, + 152 + ], + "blocks": [ + { + "bbox": [ + 173, + 71, + 284, + 152 + ], + "lines": [ + { + "bbox": [ + 173, + 71, + 284, + 152 + ], + "spans": [ + { + "bbox": [ + 173, + 71, + 284, + 152 + ], + "type": "image", + "image_path": "1f9d947d9da0b81e8c741692fa32751b52dd14962647aade5e132ce3e2302419.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 210, + 253, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 253, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 253, + 224 + ], + "type": "text", + "content": "4.4.4 Sensitivity Study of Hyper-parameters" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 229, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 295, + 266 + ], + "type": "text", + "content": "As shown in Fig.13, we investigated the effects of using LEM and LCM at various positions in the shallow layers. Please refer to the appendix for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 277, + 141, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 141, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 141, + 289 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 297, + 189, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 189, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 189, + 309 + ], + "type": "text", + "content": "5.1. Few-Shot Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 316, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 495 + ], + "type": "text", + "content": "Few-Shot Segmentation (FSS) aims to segment novel classes in query images using only a few annotated support samples. Existing methods in FSS can be generally categorized into two types: prototype-based methods and matching-based methods. Motivated by PrototypicalNet[32], prototype-based methods [19, 21, 34, 38] generate prototypes from support images and perform segmentation by computing the similarity between query features and these support prototypes. On the other hand, matching-based methods[25, 26, 29, 40] focus on analyzing pixel-to-pixel dense correspondences between support and query features to avoid the loss of spatial structure inherent in prototype-based methods. However, these methods are limited to segmenting novel classes within a single domain and do not generalize effectively to unseen domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 503, + 187, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 187, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 187, + 514 + ], + "type": "text", + "content": "5.2. Domain Generalization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Domain Generalization (DG) aims to train models that can generalize to diverse, unseen target domains, particularly when target domain data is unavailable during training, which aligns with the objective of CDFSS. Existing methods for domain generalization can be categorized into two types: methods that focus on learning domain-invariant feature representations across multiple source domains[4, 37], and methods that generate diverse samples through data or feature augmentation[10, 31]. However, most existing DG methods focus on generalizing the whole model, which is not suitable for direct use in CDFSS tasks. In contrast, our plug-and-play LEM module is compact, lightweight, and effective. In the appendix, we compare our LEM module with domain generalization methods to demonstrate the effectiveness of our approach as a novel sharpness-aware minimization method in addressing the CDFSS problem." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 516, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 516, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 516, + 85 + ], + "type": "text", + "content": "5.3. Cross-domain Few-Shot Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 90, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 354 + ], + "type": "text", + "content": "Unlike traditional FSS setting, CDFSS requires models to generalize to unseen target domains without accessing target data during training, making the task more realistic and complex. PATNet [20] introduces a transformation module that converts domain-specific features into domain-agnostic features, which can be finetuned with target domain data during the testing phase. ABCDFSS [17] proposes a test-time task adaptation method, utilizing consistency-based contrastive learning to prevent overfitting and improve class discriminability. DR-Adapter [33] introduces a small adapter for rectifying diverse target domain styles to the source domain. Compared with them, our approach proposes two simple plug-and-play modules that require no training and can be directly applied to boost performance in CDFSS. Besides, a few works [12, 27] are based on the self-support concept. Our LCM module follows a similar principle but offers two significant advantages: (1) While other methods use query prototypes to match query features, we directly use the low-level query features to adjust the query mask, preventing further loss of information. (2) LCM can be conveniently added during the target-domain testing phase, rather than during the entire training process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 388, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 388, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 388, + 376 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 384, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 562 + ], + "type": "text", + "content": "In this paper, We focus on a well-observed but under-explored phenomenon in CDFSS: for target domains, the best segmentation performance is always achieved at the very early epochs, followed by a sharp decline as source-domain training progresses. We delve into this phenomenon for an interpretation, and find low-level features are vulnerable to domain shifts, leading to sharper loss landscapes during the source-domain training, which is the devil of CDFSS. Based on the interpretation, we further propose a method that includes two plug-and-play modules: one to flatten the loss landscapes for low-level features during source-domain training, and the other to directly supplement target-domain information to the model during target-domain testing. Extensive experiments on four CDFSS benchmarks validate our rationale and effectiveness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 573, + 410, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 410, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 410, + 586 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 592, + 554, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 704 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Program of China under grant 2024YFC3307900; the National Natural Science Foundation of China under grants 62206102, 62436003, 62376103 and 62302184; Major Science and Technology Project of Hubei Province under grant 2024BAA008; Hubei Science and Technology Talent Service Project under grant 2024DJC078; and Ant Group through CCF-Ant Research Fund. The computation is completed in the HPC Platform of Huazhong University of Science and Technology." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4625" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11700-11709, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "text", + "content": "[2] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13979-13988, 2021. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 267 + ], + "type": "text", + "content": "[3] Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE Transactions on Medical Imaging, 33(2):577-590, 2013. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 322 + ], + "type": "text", + "content": "[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2229-2238, 2019. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 324, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 294, + 399 + ], + "type": "text", + "content": "[5] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 402, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 294, + 445 + ], + "type": "text", + "content": "[6] MohammadReza Davari, Stefan Horoi, Amine Natik, Guillaume Lajoie, Guy Wolf, and Eugene Belilovsky. Reliability of cka as a similarity measure in deep learning. arXiv preprint arXiv:2210.16156, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 512 + ], + "type": "text", + "content": "[7] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deep globe 2018: A challenge to parse the earth through satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 172-181, 2018. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 570, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 295, + 601 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 603, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 603, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 58, + 603, + 294, + 668 + ], + "type": "text", + "content": "[10] Yingjun Du, Jun Xu, Huan Xiong, Qiang Qiu, Xiantong Zhen, Cees GM Snoek, and Ling Shao. Learning to learn with variational information bottleneck for domain generalization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 200-216. Springer, 2020. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "text", + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2010. 1, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[12] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701–719, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "text", + "content": "[13] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "text", + "content": "[14] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In Proceedings of the IEEE International Conference on Computer Vision, pages 991-998. IEEE, 2011. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 266, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 553, + 319 + ], + "type": "text", + "content": "[16] Weizhao He, Yang Zhang, Wei Zhuo, Linlin Shen, Jiaqi Yang, Songhe Deng, and Liang Sun. Apseg: Auto-prompt network for cross-domain few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23762-23772, 2024. 1, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 322, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 555, + 365 + ], + "type": "text", + "content": "[17] Jonas Herzog. Adapt before comparison: A new perspective on cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23605-23615, 2024. 1, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 367, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 367, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 367, + 553, + 431 + ], + "type": "text", + "content": "[18] Stefan Jaeger, Alexandros Karargyris, Sema Candemir, Les Folio, Jenifer Siegelman, Fiona Callaghan, Zhiyun Xue, Kannappan Palaniappan, Rahul K Singh, Sameer Antani, et al. Automatic tuberculosis screening using chest radiographs. IEEE Transactions on Medical Imaging, 33(2):233-245, 2013. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "type": "text", + "content": "[19] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8057-8067, 2022. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 491, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 533 + ], + "type": "text", + "content": "[20] Shuo Lei, Xuchao Zhang, Jianfeng He, Fanglan Chen, Bowen Du, and Chang-Tien Lu. Cross-domain few-shot semantic segmentation. In European Conference on Computer Vision, pages 73-90. Springer, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 536, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 536, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 536, + 553, + 590 + ], + "type": "text", + "content": "[21] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 592, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 644 + ], + "type": "text", + "content": "[22] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2869–2878, 2020. 1, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4626" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[24] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 297, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 297, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 297, + 174 + ], + "type": "text", + "content": "[25] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8741-8750, 2021. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 297, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 297, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 297, + 218 + ], + "type": "text", + "content": "[26] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 297, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 297, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 297, + 284 + ], + "type": "text", + "content": "[27] Jiahao Nie, Yun Xing, Gongjie Zhang, Pei Yan, Aoran Xiao, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Cross-domain few-shot segmentation via iterative support-query correspondence mining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3380-3390, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 297, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 297, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 297, + 308 + ], + "type": "text", + "content": "[28] Alan V Oppenheim and Jae S Lim. The importance of phase in signals. Proceedings of the IEEE, 69(5):529-541, 1981. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 297, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 297, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 297, + 365 + ], + "type": "text", + "content": "[29] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 297, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 297, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 297, + 431 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 297, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 297, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 297, + 477 + ], + "type": "text", + "content": "[31] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. arXiv preprint arXiv:1804.10745, 2018. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 297, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 297, + 510 + ], + "type": "text", + "content": "[32] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 297, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 297, + 567 + ], + "type": "text", + "content": "[33] Jiapeng Su, Qi Fan, Wenjie Pei, Guangming Lu, and Fanglin Chen. Domain-rectifying adapter for cross-domain few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24036–24045, 2024. 1, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 297, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 297, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 297, + 621 + ], + "type": "text", + "content": "[34] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 297, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 297, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 297, + 668 + ], + "type": "text", + "content": "[35] Jintao Tong, Yixiong Zou, Yuhua Li, and Ruixuan Li. Lightweight frequency masker for cross-domain few-shot semantic segmentation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 297, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 297, + 713 + ], + "type": "text", + "content": "[36] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific Data, 5(1):1-9, 2018. 1, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 689 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 115 + ], + "type": "text", + "content": "[37] Haohan Wang, Zexue He, Zachary C Lipton, and Eric P Xing. Learning robust representations by projecting superficial statistics out. arXiv preprint arXiv:1903.06256, 2019. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "type": "text", + "content": "[38] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In proceedings of the IEEE/CVF international conference on computer vision, pages 9197-9206, 2019. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 217 + ], + "type": "text", + "content": "[39] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 555, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 261 + ], + "type": "text", + "content": "[40] Gengwei Zhang, Guoliang Kang, Yi Yang, and Yunchao Wei. Few-shot segmentation via cycle-consistent transformer. Advances in Neural Information Processing Systems, 34:21984-21996, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 555, + 306 + ], + "type": "text", + "content": "[41] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in neural information processing systems, 35:6575-6588, 2022. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 363 + ], + "type": "text", + "content": "[42] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Yu Qiao, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. In The Twelfth International Conference on Learning Representations, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 364, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 555, + 408 + ], + "type": "text", + "content": "[43] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 555, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 555, + 462 + ], + "type": "text", + "content": "[44] Yixiong Zou, Shanghang Zhang, Jianpeng Yu, Yonghong Tian, and José MF Moura. Revisiting mid-level patterns for cross-domain few-shot recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 741-749, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 464, + 555, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 555, + 508 + ], + "type": "text", + "content": "[45] Yixiong Zou, Shanghang Zhang, Yuhua Li, and Ruixuan Li. Margin-based few-shot class-incremental learning with class-level overfitting mitigation. Advances in neural information processing systems, 35:27267-27279, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 509, + 555, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 555, + 563 + ], + "type": "text", + "content": "[46] Yixiong Zou, Yicong Liu, Yiman Hu, Yuhua Li, and Ruixuan Li. Flatten long-range loss landscapes for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23575-23584, 2024. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 565, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 555, + 609 + ], + "type": "text", + "content": "[47] Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. Attention temperature matters in vit-based cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:116332-116354, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 609, + 555, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 609, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 609, + 555, + 654 + ], + "type": "text", + "content": "[48] Yixiong Zou, Shuai Yi, Yuhua Li, and Ruixuan Li. A closer look at the clf token for cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:85523-85545, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 654, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 555, + 689 + ], + "type": "text", + "content": "[49] Yixiong Zou, Shanghang Zhang, Haichen Zhou, Yuhua Li, and Ruixuan Li. Compositional few-shot class-incremental learning. arXiv preprint arXiv:2405.17022, 2024. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4627" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_content_list.json b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..86e2ce703a6958f20ac4d113111cb108275a3756 --- /dev/null +++ b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_content_list.json @@ -0,0 +1,1450 @@ +[ + { + "type": "text", + "text": "The Devil is in Temporal Token: High Quality Video Reasoning Segmentation", + "text_level": 1, + "bbox": [ + 109, + 130, + 888, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sitong Gong $^{1}$ , Yunzhi Zhuge $^{1*}$ , Lu Zhang $^{1}$ , Zongxin Yang $^{2}$ , Pingping Zhang $^{1}$ , Huchuan Lu $^{1}$ , IIAU, Dalian University of Technology, $^{2}$ Havard University", + "bbox": [ + 125, + 179, + 869, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stgong@mail.dlut.edu.cn, zgyz@dlut.edu.cn", + "bbox": [ + 313, + 219, + 676, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods for Video Reasoning Segmentation rely heavily on a single special token to represent the object in the keyframe or the entire video, inadequately capturing spatial complexity and inter-frame motion. To overcome these challenges, we propose VRS-HQ, an end-to-end video reasoning segmentation approach that leverages Multimodal Large Language Models (MLLMs) to inject rich spatiotemporal features into hierarchical tokens. Our key innovations include a Temporal Dynamic Aggregation (TDA) and a Token-driven Keyframe Selection (TKS). Specifically, we design frame-level $<\\mathrm{SEG}>$ and temporal-level $<\\mathrm{TAK}>$ tokens that utilize MLLM's autoregressive learning to effectively capture both local and global information. Subsequently, we apply a similarity-based weighted fusion and frame selection strategy, then utilize SAM2 to perform keyframe segmentation and propagation. To enhance keyframe localization accuracy, the TKS filters keyframes based on SAM2's occlusion scores during inference. VRS-HQ achieves state-of-the-art performance on ReVOS, surpassing VISA by $5.9\\% / 12.5\\% / 9.1\\%$ in $\\mathcal{J}\\&\\mathcal{F}$ scores across the three subsets. These results highlight the strong temporal reasoning and segmentation capabilities of our method. Code and model weights are available at VRS-HQ.", + "bbox": [ + 88, + 301, + 485, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 680, + 220, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reasoning segmentation [10, 17, 27, 30, 38], which aims to generate segmentation results from complex query texts, has advanced with multimodal foundation models [18, 21, 29]. However, existing methods [10, 17, 27, 30] focus primarily on image-level segmentation, leaving the more challenging domain of video-level segmentation, which requires temporal reasoning of object relations and attributes, relatively unexplored. To address this gap, Video Reasoning Segmentation (VRS) [1, 39, 43] has recently emerged as a promising approach. Unlike Referring Video Object Segmentation (RVOS) methods [36, 37, 40], which rely on", + "bbox": [ + 89, + 705, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "explicit descriptive phrases like \"a person skateboarding\", VRS leverages the extensive world knowledge and temporal reasoning capabilities of Multimodal Large Language Models (MLLMs) to transform implicit intent-based expressions into precise object masklets.", + "bbox": [ + 511, + 268, + 905, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent advancements in Video Reasoning Segmentation (VRS), such as VISA [39] and VideoLISA [1], significant challenges still exist. (i) Limited Temporal Context: Existing methods [1, 39] typically rely on a single segmentation token from an MLLM for keyframe-based segmentation (cf. Fig. 1 (a)), resulting in limited temporal context and hindering the effective capture of inter-frame variations and spatiotemporal features. (ii) Suboptimal Keyframe Detection: The LLaMA-VID [19] model, used by VISA for keyframe detection, can produce inaccurate keyframes, particularly in videos requiring complex temporal reasoning. (iii) Decoupled Segmentation and Propagation: VISA's reliance on separate, pre-trained models for keyframe segmentation (SAM [16]) and mask propagation (XMem [6]) prevents end-to-end training and inference.", + "bbox": [ + 511, + 351, + 906, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the above obstacles, we introduce several strategies to strengthen the VRS model's proficiency in perceiving spatial information and interpreting temporal dynamics. Firstly, current limitations arise from insufficient single-token representation capacity, which restricts the model's abilities to capture intra-frame spatial features and maintain inter-frame temporal coherence. We hypothesize that encoding frame-level and temporal-level information into hierarchical tokens separately via MLLM and integrating them could effectively unify spatial details with temporal dynamics, enhancing perceptual capability. Secondly, the occlusion score introduced by SAM2 [29] inspires us to incorporate the target confidence of each sampled frame as a criterion for keyframe determination. We thus employ the temporal token in conjunction with SAM2 to generate the occlusion scores, applying temporal information for precise keyframe selection. As shown in Fig. 1 (b), our approach locates the cat more accurately compared to VISA, boosting inference efficiency and segmentation performance. Thirdly, leveraging SAM2's integrated segmentation and propagation capabilities, we efficiently fine", + "bbox": [ + 511, + 583, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 109, + 886, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d41f3819214bf9e878ec752bd52386ee8e970d6507b1e9a01e59a4a64c871939.jpg", + "image_caption": [ + "Figure 1. Comparison with previous VRS approaches. (a) Previous methods utilize a single token for keyframe-based segmentation, depending heavily on external models for keyframe detection and mask propagation. This reliance can hinder accurate keyframe localization and prevent end-to-end inference. (b) In contrast, VRS-HQ introduces frame-level and a temporal token for dynamic aggregation. The aggregated token is then used for both keyframe selection and mask generation within SAM2. This enables single-stage inference with precise keyframe selection and high-quality segmentation. (c) VRS-HQ achieves state-of-the-art performance on various image and video datasets across reasoning and referring segmentation." + ], + "image_footnote": [], + "bbox": [ + 91, + 77, + 575, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/37cc4e7a23c85283f4ae285ebf9555f94f044cc04f4d4e4c811871501e380e4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 74, + 898, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tune it with our temporal token and its memory mechanism for improved mask quality and inference efficiency.", + "bbox": [ + 88, + 378, + 482, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present VRS-HQ (High-Quality Video Reasoning Segmentation) (cf. Fig. 2), a novel approach that leverages aggregated temporal tokens for enhanced keyframe selection and efficient mask decoding. To begin with, we propose the Temporal Token Encoding (§3.1), prompting the MLLM to encode frame- and video-level target features into multi-level special tokens using sampled video frames and tailored conversation templates. The primary innovation lies in the Temporal Dynamic Aggregation (§3.2), which employs a cosine similarity-based weighted fusion strategy to merge frame-level embeddings into the temporal-level embedding, consolidating spatial features into while maintaining the temporal consistency of the targets. Subsequently, we select the keyframe according to the token similarity between and during training. Moreover, we propose the Token-driven Keyframe Selection (§3.3) during inference, sequentially treating each sampled frame as a potential keyframe and interacting it with fused embedding through SAM2 to calculate an occlusion score. The scores combined with previous token similarity are used as the criterion for keyframe selection. Finally, the keyframe with the fused embedding is fed into SAM2 for mask generation, while the remaining frames are treated as non-keyframes and utilize SAM2's memory mechanism for Mask Decoding and Propagation (§3.4).", + "bbox": [ + 91, + 412, + 483, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments, VRS-HQ achieves state-of-the-art performance across a diverse range of video segmentation benchmarks (cf. Fig. 1 (c)). Specifically, on the ReVOS [39] dataset, VRS-HQ outperforms VISA-13B by $9.1\\%$ in terms of the $\\mathcal{J}\\& \\mathcal{F}$ metric, highlighting the critical role of our proposed module in enhancing the reason", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing ability for scenes with complex spatiotemporal dynamics. Moreover, VRS-HQ surpasses previous methods by $7.3\\% / 5.6\\% / 6.5\\%$ $\\mathcal{J}\\& \\mathcal{F}$ on three standard referring video segmentation datasets respectively, underscoring its strong inter-frame perception and robust video tracking capabilities. Key contributions can be summarized as follows:", + "bbox": [ + 511, + 377, + 906, + 467 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present Temporal Dynamic Aggregation to blend spatial features from frame-level tokens into a temporal token, endowing the model with the ability to discern inter-frame variations and comprehend the global semantic context of the targets.", + "- We focus on harnessing the power of temporal tokens in video perception, using the integrated temporal token via SAM2 for keyframe segmentation and propagation. Additionally, we present the Token-driven Keyframe Selection, combining each sampled frame with the temporal token to generate occlusion scores via SAM2, providing a reliable basis for keyframe detection.", + "- By combining the above designs, we introduce VRS-HQ, demonstrating state-of-the-art performance on the VRS benchmark and existing RVOS datasets." + ], + "bbox": [ + 511, + 469, + 905, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 714, + 661, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Referring Video Object Segmentation", + "text_level": 1, + "bbox": [ + 511, + 739, + 838, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Referring Video Object Segmentation [2, 37, 40] (RVOS), which focuses on segmenting and tracking prominent objects in video frames using explicit textual descriptions, has significantly advanced through the integration of visual and linguistic cues. A large proportion of these studies have leveraged attention mechanisms to merge multimodal information. For instance, Wang et.al. [33] uses asymmetric cross-guided attention to enhance sentence representations and aggregate visual context. Seo et.al. [32] employ cross", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29184", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b6c6cac0afd566350e6f3b963597ea3df3ff9609e6ae3d90ad5c53f9760462ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 84, + 883, + 217 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd405aea2e5fd8d494bac8a0d26dd8bb56f0476b6e32ccbcee3f7ddd8eee11b6.jpg", + "image_caption": [ + "Figure 2. (a) VRS-HQ architecture. VRS-HQ incorporates a Multimodal Large Language Model for Temporal Token Encoding ( and tokens, §3.1), a Temporal Dynamic Aggregation, a Token-driven Keyframe Selection and Mask Decoding and Propogation. (b) Temporal Dynamic Aggregation (TDA) merges frame-level tokens into a temporal token using a weighted fusion based on cosine similarity. (§3.2). (c) Token-driven Keyframe Selection (TKS). During training, the frame with the token closest to the token is selected as the keyframe. During inference, keyframe selection is refined using SAM2's occlusion scores and token similarity scores (§3.3). (d) Mask Decoding and Propagation (MDP). The token provides a sparse embedding for SAM2, generating a keyframe mask and propagating it to other frames via a memory mechanism (§3.4)." + ], + "image_footnote": [], + "bbox": [ + 112, + 219, + 349, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1345cde1f7187f6e0b405e800a596e4fd69ff976dd5c90c0d90156dc512da619.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 219, + 651, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dea18d298f5a29e51fe45d5fb41d2de62e27f5d3b92d453940d84d65407a084b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 219, + 880, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "modal and memory attention to jointly address RVOS and semi-supervised video object segmentation. Hui et.al. [13] dynamically recombine linguistic features and interact with visual features using cross-modal attention to highlight spatiotemporal regions of interest. Inspired by DETR [4], methods like MTTR [2] and ReferFormer [37] leverage language queries for precise target localization. Emerging approaches are exploring other-modal guidance for referring segmentation, such as audio cues [8, 9, 47]. While RVOS excels with explicit object descriptions, it often struggles with the complex reasoning required to interpret more implicit and intricate language instructions, motivating our exploration of Video Reasoning Segmentation.", + "bbox": [ + 88, + 462, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Reasoning Segmentation", + "text_level": 1, + "bbox": [ + 89, + 667, + 316, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reasoning segmentation [17, 27, 30, 38] advances referring segmentation by generating masks from complex images and versatile text prompts. LISA [17] pioneers the field of reasoning segmentation. It introduces a new token to expand the vocabulary and proposes the embedding-as-mask paradigm, enhancing segmentation capabilities to address scenarios requiring complex reasoning and world knowledge. LISA++ [41] addresses LISA's limitations in distinguishing individual instances by enhancing instruction-tuning data with key segmentation datasets, supporting both semantic and instance segmentation tasks. PixelLM [30] innovatively combines a novel pixel decoder and a segmentation codebook with learnable tokens to efficiently generate high-quality masks without external models. VISA [39]", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "introduces reasoning segmentation to the video domain, proposing reasoning video object segmentation. It uses a pre-trained Multimodal Large Language Model (MLLM) to select keyframes, segments them based on reasoning, and propagates the masks to other frames using a pre-trained object tracker. However, it suffers from limited representation capability of a single special token, inaccurate keyframe selection, and the inability to perform end-to-end inference, compromising its segmentation and tracking performance.", + "bbox": [ + 511, + 462, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 611, + 648, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Definition. The task of Video Reasoning Segmentation can be briefly outlined as follows. Given a video clip consisting of $T$ frames $\\mathcal{X}_V\\in \\mathbb{R}^{T\\times 3\\times H\\times W}$ , where $H$ and $W$ denote the height and width of each frame respectively, along with a high-level textual instruction $\\mathcal{X}_{txt}$ , VRS aims to design a model $\\mathcal{M}$ to interpret and transform $\\mathcal{X}_{txt}$ into the binary segmentation mask sequence $\\mathcal{X}_M\\in \\mathbb{R}^{T\\times H\\times W}$ for each frame. In contrast to RVOS tasks providing explicit descriptions like \"the person skateboarding,\" VRS typically employs expressions incorporating world knowledge like \"tool(s) for holding garbage\" or temporal logic like \"the ship(s) moving at the highest speed\", raising higher demands on the model's capabilities of temporal relationship comprehension and complex scenario reasoning.", + "bbox": [ + 511, + 633, + 906, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall Architecture. Fig. 2 illustrates the VRS-HQ architecture, which comprises Chat-UniVi [14] as the Multi-modal Large Language Model (MLLM) for temporal token", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29185", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "encoding (§3.1), a Temporal Dynamic Aggregation (§3.2), a Token-driven Keyframe Selection (§3.3), and Mask Decoding and Propagation (§3.4). The MLLM encodes the textual prompt into multi-level tokens representing spatial and semantic target information. These tokens drive both temporal dynamic aggregation and keyframe selection via cosine similarity scores. SAM2 then segments the keyframes and tracks the target object throughout the video. During inference, the Token-driven Keyframe Selection computes object occlusion scores for each sampled frame by merging it with the fused temporal token. Filtering keyframes based on these scores improves the accurate localization of the target object and subsequently segmentation process.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Temporal Token Encoding", + "text_level": 1, + "bbox": [ + 89, + 300, + 331, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current VRS approaches [1, 39] typically use specified language prompts to guide MLLMs in embedding target information from keyframes or entire videos into a single specialized token, struggling to capture rich spatiotemporal dynamics essential for fine-grained video understanding. To equip the MLLM with both frame-level and video-level contextual awareness for video segmentation, we propose encoding intra-frame spatial information and inter-frame temporal relations into hierarchical tokens.", + "bbox": [ + 89, + 321, + 483, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hierarchical Token Generation. Instead of previous single-token encoding strategy, the vocabulary of the MLLM is initially augmented with two new special tokens: $<\\mathrm{SEG}>$ and $<\\mathrm{TAK}>$ . Then, we design a structured conversational template as \"USER: Please find {expression} in the Reference Video and segment it in each frame and the entire video respectively.\" Here, \" $\\{\\mathrm{expression}\\}$ \" denotes the target object description. The tokenized prompts $\\mathcal{X}_{\\text{txt}}$ and sampled video frames $\\mathcal{X}_{\\text{V}'}$ are input into the MLLM, which generates a response $y_{\\text{txt}}$ containing multiple frame-level $<\\mathrm{SEG}>$ tokens and a temporal-level $<\\mathrm{TAK}>$ token through autoregressive encoding.", + "bbox": [ + 89, + 468, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token Extraction and Mapping. Subsequently, the $<\\mathrm{SEG}>$ token embeddings $\\bar{h}_{seg} \\in \\mathbb{R}^{T' \\times d'}$ and the $<\\mathrm{TAK}>$ token embedding $\\bar{h}_{tak} \\in \\mathbb{R}^{1 \\times d'}$ are extracted from the MLLM's final layer, where $d'$ denotes the MLLM's embedding dimension and $T'$ denotes the length of sampled frames. These embeddings are then projected into the same feature space as SAM2 using a multi-layer perceptron:", + "bbox": [ + 89, + 674, + 483, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {s e g}, h _ {t a k} = M L P \\left(\\bar {h} _ {s e g}\\right), M L P \\left(\\bar {h} _ {t a k}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 792, + 482, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $h_{seg} \\in \\mathbb{R}^{T' \\times d}$ and $h_{tak} \\in \\mathbb{R}^{1 \\times d}$ represent the sparse embeddings for segmentation mask activation, and $d$ denotes the feature dimension of SAM2. Finally, the frame-level embeddings $h_{seg}$ are aggregated with the video-level embedding $h_{tak}$ to incorporate temporal dynamics.", + "bbox": [ + 89, + 823, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Temporal Dynamic Aggregation", + "text_level": 1, + "bbox": [ + 511, + 90, + 797, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on the strong temporal encoding of MLLM, and embeddings encapsulate the spatial priors and temporal semantic signals of the targets respectively, providing rich contextual information for the segmentation model. Accordingly, we propose the Temporal Dynamic Aggregation to facilitate the fusion of positional and semantic information of the targets.", + "bbox": [ + 511, + 112, + 906, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Keyframes as Token Similarity. The cosine similarity between each frame-level $<\\text{SEG}>$ token and the temporal-level $<\\text{TAK}>$ token reflects the semantic alignment between individual frames and the overall video context. We hypothesize that high-similarity frames are more representative of the video's content and thus suitable as keyframes. This motivates our use of token similarity for keyframe selection during training, which also facilitates weighted fusion of $<\\text{SEG}>$ and $<\\text{TAK}>$ tokens.", + "bbox": [ + 511, + 222, + 906, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarity-based Weighted Fusion. To enhance the inter-frame consistency while achieving more precise discernment of the object's positional shifts across frames, we fuse the frame-level embeddings $h_{seg}$ with the video-level embedding $h_{tak}$ using normalized cosine similarity scores as attention weights (cf. Fig. 2 (b)). During backpropagation, frame-level features $h_{seg}$ are updated alongside the temporal-level embedding $h_{tak}$ to enable more comprehensive learning of local and global video context:", + "bbox": [ + 511, + 363, + 908, + 501 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t a k} ^ {\\prime} = h _ {t a k} + \\alpha \\sum_ {i = 1} ^ {T ^ {\\prime}} \\lambda_ {i} h _ {s e g} [ i ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 512, + 906, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{i}$ represents the normalized cosine similarity between $h_{seg}[i]$ and $h_{tak}$ , and $\\alpha$ is the fusion coefficient. The resulting fused embedding $h_{tak}^{\\prime}$ is then used for keyframe selection during inference and mask generation.", + "bbox": [ + 511, + 564, + 906, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Token-driven Keyframe Selection", + "text_level": 1, + "bbox": [ + 511, + 633, + 807, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VISA [39] relies on an external model (LLaMA-VID [19]) for keyframe selection during inference, hindering end-to-end processing and potentially degrading performance due to inaccuracies in the external model's output. To address this limitation, we introduce the Token-driven Keyframe Selection that leverages the temporal information encoded within the integrated embedding. This approach eliminates the need for complex prompt engineering and significantly improves the reliability of keyframe selection.", + "bbox": [ + 511, + 656, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of uniform video sampling during training, we adopt the CLIP model [25] to find the frame most aligned with the expression $\\mathcal{X}_{exp}$ for inference. The anchor frame is used for global sampling, resulting in sampled frames $\\mathcal{X}_V^f\\in \\mathbb{R}^{\\hat{T}\\times 3\\times H\\times W}$ , where $\\hat{T}$ represents the number of the sampled frames. Following the temporal dynamic aggregation (Eq. 2), each sampled frame is treated as a potential", + "bbox": [ + 511, + 792, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29186", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "keyframe, along with the fused $<\\mathrm{TAK}>$ embedding $h_{\\text{tak}}'$ , is sent to SAM2 to generate the object occlusion scores $S_{o} \\in \\mathbb{R}^{\\hat{T} \\times 1}$ (cf. Fig. 2 (c)):", + "bbox": [ + 89, + 90, + 483, + 138 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {o} = \\mathcal {M D} (\\mathcal {E} (\\mathcal {X} _ {V} ^ {f}), h _ {t a k} ^ {\\prime}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 143, + 482, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The occlusion scores effectively indicate the confidence of the target's presence in the current frame, assisting in identifying which frame best corresponds to the information encoded in the token. After determining the keyframe index by the combination of softmax-normalized scores $S_{o}^{\\prime}$ with token similarity scores $S_{t}$ , we decode the keyframe mask and then utilize the cross-frame propagation technique of SAM2 to obtain video-level segmentation masks.", + "bbox": [ + 89, + 170, + 483, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Mask Decoding and Propagation", + "text_level": 1, + "bbox": [ + 89, + 300, + 380, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following temporal token fusion and keyframe selection, the mask embeddings enriched with positional and semantic information can be yielded by the segmentation model. In contrast to prior methods [1, 39, 43] depending on image-level segmentation models or external object trackers for target trajectories prediction, we utilize SAM2 to perform segmentation and propagation concurrently.", + "bbox": [ + 89, + 321, + 483, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SAM2 for Masklets Decoding. Given the keyframe $\\mathcal{X}_k$ , we first extract its features using the image encoder $\\mathcal{E}$ , providing the conditional input for SAM2. The integrated temporal embedding $h_{\\text{tak}}'$ then interacts with these keyframe features within the mask decoder $\\mathcal{MD}$ (cf. Fig. 2 (d)) to generate the keyframe mask:", + "bbox": [ + 89, + 433, + 483, + 525 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} _ {M} ^ {k} = \\mathcal {M D} \\left(\\mathcal {E} \\left(\\mathcal {X} _ {k}\\right), h _ {t a k} ^ {\\prime}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 532, + 482, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_M^k$ represents the predicted mask for the keyframe. Next, we propagate this mask to two adjacent frames, $\\mathcal{X}_{k - 1}$ and $\\mathcal{X}_{k + 1}$ , treated as non-conditional frames, using the memory storage and interaction mechanism. This yields the mask sequence $\\mathcal{X}_M^s$ . During inference, all remaining frames are processed as non-conditional frames, facilitating mask propagation throughout the entire video.", + "bbox": [ + 89, + 556, + 483, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Objectives", + "text_level": 1, + "bbox": [ + 89, + 671, + 279, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method is trained end-to-end using a combined text generation loss and mask loss to optimize the $\\langle \\mathrm{TAK} \\rangle$ and $\\langle \\mathrm{SEG} \\rangle$ embeddings. The mask loss combines binary cross-entropy (BCE) and DICE loss:", + "bbox": [ + 89, + 693, + 483, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = \\lambda_ {t x t} L _ {t x t} \\left(y _ {t x t}, \\hat {y} _ {t x t}\\right) + \\lambda_ {m a s k} L _ {m a s k}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 762, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m a s k} = \\lambda_ {b c e} L _ {b c e} \\left(\\mathcal {X} _ {M} ^ {s}, \\hat {\\mathcal {X}} _ {M} ^ {s}\\right) + \\lambda_ {d i c e} L _ {d i c e} \\left(\\mathcal {X} _ {M} ^ {s}, \\hat {\\mathcal {X}} _ {M} ^ {s}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 806, + 483, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y}_{txt}$ and $\\hat{\\chi}_M^s$ represent the ground truth text and mask, respectively, and $y_{txt}$ and $X_M^s$ are their corresponding predictions. The weighting coefficients $\\lambda_{txt}$ , $\\lambda_{mask}$ , $\\lambda_{bce}$ , and $\\lambda_{dice}$ are set to 1, 1, 2 and 0.5, respectively.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 513, + 114, + 710, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Our model is trained and evaluated on extensive image and video segmentation datasets, LLaVA-Instruct-150k [21], as well as the Video Question-Answering datasets from Video-ChatGPT [22]. Specifically, the image segmentation datasets comprise semantic segmentation: ADE20K [45], COCO-Stuff [3], PACO-LVIS [26], and PASCALPart [5], referring segmentation: refCLEF, refCOCO, refCOCO+ [15], and refCOCOg [23], and reasoning segmentation: ReasonSeg [17]. While the video segmentation datasets encompass the RVOS datasets: RefYoutube-VOS [32], Ref-DAVIS17 [24] and MeViS [7], and the VRS benchmark ReVOS [39].", + "bbox": [ + 511, + 137, + 906, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate video segmentation using region similarity $(\\mathcal{I})$ , contour accuracy $(\\mathcal{F})$ , and their mean $(\\mathcal{J} \\& \\mathcal{F})$ . Image segmentation accuracy is measured via Generalized Intersection over Union (gIoU) [31] and Complete Intersection over Union (cIoU) [44]. Model hallucination is assessed using the robustness score $\\mathcal{R}$ [39].", + "bbox": [ + 511, + 325, + 906, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 426, + 730, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We fine-tuned Chat-UniVi [14] (our chosen MLLM) using LoRA [12] (rank 8), optimizing the mask decoder and MLP projection layer while freezing other parameters. Training used AdamW (learning rate 0.0003, no weight decay) with a WarmupDecayLR scheduler (100-iteration warmup). We set the fusion coefficient $\\alpha$ to 0.1 and sampled two non-keyframes per video. Trained for 7500 iterations on a hybrid image/video dataset using four A800 GPUs with DeepSpeed [28] (batch size 1, gradient accumulation 32, total batch size 128), the model employed TDA for video data (token fusion, segmentation, and propagation) and the token for direct segmentation on image data.", + "bbox": [ + 511, + 448, + 906, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison Results", + "text_level": 1, + "bbox": [ + 513, + 641, + 705, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To showcase the robust pixel-level perception and generalization of VRS-HQ, we conduct evaluations across diverse benchmarks, including ReVOS, RVOS, and image-based referring and reasoning segmentation datasets.", + "bbox": [ + 511, + 662, + 905, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ReVOS Datasets. Tab. 1 illustrates the performance comparison with previous methods [2, 17, 37, 39, 46] on the ReVOS benchmark. VRS-HQ demonstrates significant improvements over the previous state-of-the-art, VISA [39] across all ten metrics. Remarkably, in terms of the $\\mathcal{J}\\&\\mathcal{F}$ , VRS-HQ-13B surpasses VISA-13B by $5.9\\%$ on the referring subset and $12.5\\%$ on the reasoning subset, respectively. These gains highlight the effectiveness of our temporal aggregation strategy and the utilization of the token for keyframe selection and target segmentation, which enhances the model's reasoning capabilities. In contrast,", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29187", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/33bea7b50e26a4a8147f01165a30cda4806249c0bf563519d7336221efb9ca2a.jpg", + "table_caption": [ + "Table 1. Performance comparison with previous methods on ReVOS dataset." + ], + "table_footnote": [], + "table_body": "
MethodsBackbonereferringreasoningoverallR
JFJ&FJFJ&FJFJ&F
MTTR [2][CVPR2022]Video-Swin-T29.830.230.020.421.521.025.125.925.55.6
LMPM [7][ICCV2023]Swin-T29.039.134.113.324.318.821.231.726.43.2
ReferFormer [37][CVPR2022]Video-Swin-B31.234.332.721.325.623.426.229.928.18.8
LISA [17][CVPR2024]LLaVA-7B44.347.145.733.838.436.139.142.740.99.3
LISA [17][CVPR2024]LLaVA-13B45.247.946.634.339.136.739.843.541.68.6
TrackGPT [46][arXiv2023]LLaVA-7B46.749.748.236.841.239.041.845.543.611.6
TrackGPT [46][arXiv2023]LLaVA-13B48.350.649.538.142.940.543.246.845.012.8
VISA [39][ECCV2024]LLaVA-7B49.452.651.040.545.843.244.949.247.115.3
VISA [39][ECCV2024]LLaVA-13B55.759.057.441.946.544.248.852.850.815.1
VISA [39][ECCV2024]Chat-UniVi-7B49.252.650.940.645.443.044.949.046.915.5
VISA [39][ECCV2024]Chat-UniVi-13B55.659.157.442.046.744.348.852.950.914.5
VRS-HQ[Ours]Chat-UniVi-7B59.864.562.153.558.756.156.661.659.119.7
VRS-HQ[Ours]Chat-UniVi-13B61.165.563.354.159.456.857.662.560.018.9
", + "bbox": [ + 143, + 97, + 857, + 295 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aec7bc6bbe210b7c6774a90051410ff9e34ee17bbbc79b535b995cf942dfce8f.jpg", + "table_caption": [ + "Table 2. Performance comparison with previous methods on the validation sets of RVOS datasets." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneRef-YouTube-VOSRef-DAVIS17MeViS
JFJ&FJFJ&FJFJ&F
MTTR [2][CVPR2022]Video-Swin-T54.056.655.3---28.831.230.0
LMPM [7][ICCV2023]Swin-T------34.240.237.2
ReferFormer [37][CVPR2022]Video-Swin-B61.364.662.958.164.161.129.832.231.0
OnlineRefer [36][CVPR2023]Swin-L61.665.563.561.667.764.8---
DsHmp [11][CVPR2024]Video-Swin-B65.069.167.161.768.164.943.049.846.4
LISA [17][CVPR2024]LLaVA-7B53.454.353.962.267.364.835.139.437.2
LISA [17][CVPR2024]LLaVA-13B54.054.854.463.268.866.035.840.037.9
TrackGPT [46][aXiv2023]LLaVA-7B55.357.456.459.467.063.237.642.640.1
TrackGPT [46][aXiv2023]LLaVA-13B58.160.859.562.770.466.539.243.141.2
VISA [39][ECCV2024]Chat-UniVi-7B59.863.261.566.372.569.440.746.343.5
VISA [39][ECCV2024]Chat-UniVi-13B61.464.763.067.073.870.441.847.144.5
VideoLISA [1][NeurIPS2024]LLaVA-Phi-3-V61.765.763.764.972.768.841.347.644.4
VRS-HQ[Ours]Chat-UniVi-7B68.372.570.472.679.476.047.653.750.6
VRS-HQ[Ours]Chat-UniVi-13B69.073.171.071.077.974.448.053.750.9
", + "bbox": [ + 158, + 320, + 838, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VISA relies on LLaMA-VID [19] for keyframe selection during inference. This strategy may neglect intra-frame fine-grained visual details, leading to inaccurate keyframe localization. Furthermore, the robustness score $\\mathcal{R}$ exceeds VISA-13B by $4.4\\%$ , indicating VRS-HQ's superior capability in handling negative samples.", + "bbox": [ + 89, + 547, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RVOS. Tab. 2 compares VRS-HQ with state-of-the-art RVOS methods. On Ref-YouTube-VOS and RefDAVIS17, VRS-HQ-13B surpasses VideoLISA [1] and VISA-13B [39], achieving $\\mathcal{I}\\&\\mathcal{F}$ improvements of $7.3\\%$ and $5.6\\%$ , respectively. Furthermore, on the motion-intensive MeViS dataset, VRS-HQ-13B yields substantial gains of $6.2\\%$ , $6.1\\%$ and $6.4\\%$ in $\\mathcal{I}$ , $\\mathcal{F}$ and $\\mathcal{I}\\&\\mathcal{F}$ , respectively. demonstrating its effective in capturing motion and maintaining temporal coherence. These improvements, compared to VideoLISA's single-token representation, are attributed to the enhanced inter-frame perception of the TDA and the refined keyframe localization of the TKS.", + "bbox": [ + 89, + 641, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RIS. Our method seamlessly extends to referring image segmentation (RIS) by treating images as single-frame videos. As shown in Tab. 3, VRS-HQ consistently outperforms VISA on three RIS benchmarks, achieving performance comparable to LISA [17] and VideoLISA [1].", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On ReasonSeg, VRS-HQ performs competitively, ranking merely below VideoLISA, demonstrating strong generalization. The marginally lower performance on image datasets compared to video datasets likely stems from two factors: (i) our model's training emphasis on video data (unlike LISA, which is pre-trained solely on images); and (ii) our training methodology encourages reliance on SAM2's memory mechanism for multi-frame processing, which may be less effective for single-image segmentation compared to VideoLISA's single-token approach.", + "bbox": [ + 511, + 547, + 906, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 710, + 679, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fusion Coefficient Ablation. The ablation analysis of the fusion coefficient $\\alpha$ in the TDA is illustrated in Tab. 5. An $\\alpha$ of 0.1 yields optimal performance. Setting $\\alpha$ to 0 leads to a sharp drop in the model's performance as $<\\mathrm{TAK}>$ fails to capture fine-grained intra-frame details, and the $<\\mathrm{SEG}>$ token cannot be jointly optimized with $<\\mathrm{TAK}>$ through backpropagation (e.g., $62.1\\% \\rightarrow 58.8\\%$ $\\mathcal{J}\\& \\mathcal{F}$ on the referring subset, $56.1\\% \\rightarrow 54.1\\%$ $\\mathcal{J}\\& \\mathcal{F}$ on the reasoning subset). As $\\alpha$ increases beyond 0.1, the metrics decline slightly, likely due to excessive frame-level noise being introduced into the temporal token at a higher fusion coef-", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29188", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ba03d3528a250227ae4cbc16bdb51ea04dea1a15253eb44063dbc53f8dcb3a75.jpg", + "table_caption": [ + "Table 3. Performance comparison with previous methods on referring and reasoning image segmentation datasets." + ], + "table_footnote": [], + "table_body": "
MethodBackbonerefCOCOrefCOCO+refCOCOgReaSeg (val)ReaSeg (test)
valtestAtestBvaltestAtestBvaltestgIoUcIoUgIoUcIoU
CRIS [34]ResNet10170.573.266.162.368.153.759.960.4----
LAVT [42]Swin-B72.775.868.862.168.455.161.262.1----
ReLA [20]Swin-B73.876.570.266.071.057.765.066.0----
X-Decoder [48]DaViT-L------64.6-22.617.921.716.3
SEEM [49]DaViT-L------65.7-25.521.224.318.7
LISA [17]LLaVA-7B74.979.172.365.170.858.167.970.652.954.047.348.4
VISA [39]Chat-UniVi-7B72.475.568.159.864.853.165.566.452.757.8--
VideoLISA [1]LLaVA-Phi-3-V73.876.668.863.468.856.268.368.861.467.153.854.4
VRS-HQChat-UniVi-7B73.577.569.561.767.654.366.767.555.251.851.752.9
", + "bbox": [ + 120, + 97, + 874, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8117000025a9b4bd8b68eac65ed980c88c8ec512c1082036af06a3be4c5b35a1.jpg", + "table_caption": [ + "Table 4. Ablation analysis of TDA and TKS components." + ], + "table_footnote": [], + "table_body": "
Componentsreferringreasoning
JFJ&FJFJ&F
w/o TDA + TKS54.559.156.848.653.951.2
w/o TDA56.561.158.851.556.754.1
w/o TKS57.862.460.150.956.253.5
VRS-HQ59.864.562.153.558.756.1
", + "bbox": [ + 104, + 295, + 352, + 351 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/819bc6910ba7d758da02cae0e0e0e567f0b5ea9a0063177858a5211111008c77.jpg", + "table_caption": [ + "Table 7. Ablation analysis of sampled frames during inference stage." + ], + "table_footnote": [], + "table_body": "
Framesreferringreasoning
JFJ&FJFJ&F
158.462.760.649.855.152.5
358.663.060.851.056.253.6
658.663.360.951.557.054.3
958.963.761.352.557.855.1
1259.364.061.752.557.855.1
Non-fixed59.864.562.153.558.756.1
", + "bbox": [ + 114, + 409, + 339, + 481 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/63c0199847882406ba8e49b2046bffbc2153414dcba90036d5c3c7ecbb4c2f4a.jpg", + "table_caption": [ + "Table 5. Ablation analysis of the fusion coefficient $\\alpha$ ." + ], + "table_footnote": [], + "table_body": "
αreferringreasoning
JFJ&FJFJ&F
056.561.158.851.556.754.1
0.159.864.562.153.558.756.1
0.2558.963.861.352.357.855.0
0.558.263.060.651.456.654.0
", + "bbox": [ + 395, + 290, + 599, + 345 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4091b59e233b287ba50ef14b2c2038fa5c27e30c8321f263055b0ec27e7be8df.jpg", + "table_caption": [ + "Table 8. Ablation analysis of the mask decoding and propagation strategy. MT/MI: Multiframe Training/Inference. ST/SI: Single-frame Training/Inference." + ], + "table_footnote": [], + "table_body": "
Strategyreferringreasoning
JFJ&FJFJ&F
SAM+ST+SI48.352.450.342.947.445.2
SAM2+ST+SI55.959.857.849.354.051.6
SAM2+ST+MI55.860.858.346.552.549.5
SAM2+MT+SI54.157.455.848.252.250.2
SAM2+MT+MI59.864.562.153.558.756.1
", + "bbox": [ + 374, + 407, + 620, + 470 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5919e87dac187c9759209b37e0d0f3e72ae3fe7d9d4f6daa00c1abc7d7e0c22c.jpg", + "table_caption": [ + "Table 6. Ablation analysis of Token-driven Keyframe Selection. $S_{1}$ , $S_{2}$ , and $S_{3}$ represent CLIP scores, token similarity scores, and occlusion scores, respectively." + ], + "table_footnote": [], + "table_body": "
S1S2S3referringreasoning
JFJ&FJFJ&F
✓✓✓59.764.261.952.557.855.2
✓✓57.862.460.150.956.253.6
✓✓59.664.161.852.557.755.1
✓✓59.864.562.153.558.756.1
", + "bbox": [ + 651, + 309, + 888, + 364 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/581aa035cf86f28c3db8cf4b7762e9ef5ff3bcf47e1adbdaeeeb1f195c8148de.jpg", + "table_caption": [ + "Table 9. Ablation analysis of sampling strategy." + ], + "table_footnote": [], + "table_body": "
Sampling strategyreferringreasoning
JFJ&FJFJ&F
Random Sampling59.063.661.353.258.455.8
Uniform Sampling59.363.961.653.358.455.8
CLIP Sampling59.864.562.153.558.756.1
", + "bbox": [ + 645, + 398, + 893, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ficient (e.g., $62.1\\% \\rightarrow 60.6\\%$ $\\mathcal{J}\\& \\mathcal{F}$ on the referring subset, $56.1\\% \\rightarrow 54.0\\%$ $\\mathcal{J}\\& \\mathcal{F}$ on the reasoning subset).", + "bbox": [ + 89, + 512, + 482, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Token-driven Keyframe Selection. Tab. 6 analyzes the impact of different score combinations for keyframe selection within the TKS. Results indicate that the occlusion score is the most influential of the three considered, improving $\\mathcal{J}\\& \\mathcal{F}$ by $1.8\\%$ and $1.6\\%$ on the referring and reasoning subsets, respectively (first two rows). This improvement stems from the occlusion score's ability to reflect the target object's presence within each frame. The model achieves its best performance when keyframe selection is based on a combination of token similarity and the occlusion score. However, incorporating the CLIP score (rows 1 and 4) slightly degrades performance (by $0.2\\%$ and $0.9\\%$ $\\mathcal{J}\\& \\mathcal{F}$ on the referring and reasoning subsets, respectively), likely due to CLIP's limited spatiotemporal understanding, which hinders accurate keyframe selection.", + "bbox": [ + 88, + 555, + 482, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask Decoding and Propagation. Tab. 8 analyzes the impact of different mask decoding and propagation strategies on segmentation performance. Using only SAM and TDA for single-frame segmentation (first row) already surpasses VISA-13B by $0.9\\% / 0.7\\% / 0.9\\%$ on the reasoning subset. Comparing the first two rows reveals the substantial advantage of SAM2 over SAM for single-frame fine", + "bbox": [ + 88, + 794, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tuning and inference. SAM2 achieves improvements of $7.6\\% / 7.4\\% / 7.5\\% / 6.4\\% / 6.6\\% / 6.4\\%$ , demonstrating its superior robustness in segmentation and tracking. Furthermore, incorporating multi-frame propagation during both training and inference (last row) yields the best overall results, effectively leveraging SAM2's memory mechanism for keyframe mask storage. This approach leads to further improvements in $\\mathcal{J} \\& \\mathcal{F}$ of $3.8\\%$ and $4.5\\%$ compared to the other strategies.", + "bbox": [ + 511, + 512, + 906, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference-Time Sampling Strategy. Tab. 9 analyzes the impact of different frame sampling strategies during inference. \"Random\" denotes randomly selected frames, while \"Uniform\" extracts frames at equal intervals. CLIP sampling slightly outperforms uniform sampling, improving $\\mathcal{J} \\& \\mathcal{F}$ by $0.5\\% / 0.3\\%$ on the referring and reasoning subsets, respectively. This modest gain is likely due to CLIP's ability in selecting frames that are semantically aligned with the referring expression, which in turn facilitates the reasoning process of the MLLM.", + "bbox": [ + 511, + 643, + 906, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualization Analysis", + "text_level": 1, + "bbox": [ + 511, + 801, + 720, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segmentation Map Comparison. Fig. 3 presents the qualitative comparison of VRS-HQ and VISA on the ReVOS dataset across three different scenarios. In the left case, VISA displays ambiguity in spatial positioning and neglects critical references such as \"the first\", revealing", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29189", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/132ba830efe9971c912ef8b6c19dc8cca0c4b4434a8ddf345229bcfab2da322a.jpg", + "image_caption": [ + "Figure 3. Segmentation map comparison of VISA and VRS-HQ on the ReVOS benchmark (§4.5). Results across three scenarios demonstrate that VRS-HQ excels in reasoning complex spatial and temporal relationships, delivering enhanced segmentation performance." + ], + "image_footnote": [], + "bbox": [ + 107, + 84, + 880, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52f97c3ac936688321cbf323ebd1fcc59de8592d6bb39aa4d3a4289919c78b8d.jpg", + "image_caption": [ + "Figure 4. Visualization of feature maps (§4.5). From top to bottom are: (a) Ground truth masks. b) Keyframe mask embeddings generated by the token before TDA. (c) Keyframe mask embeddings generated by the token after TDA." + ], + "image_footnote": [], + "bbox": [ + 101, + 378, + 464, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "its limited semantic understanding. In contrast, VRS-HQ demonstrates a strong ability to perceive spatiotemporal relationships. For the middle example, VISA mistakenly segments the person flying the kite, while VRS-HQ correctly identifies the kite, reflecting the stronger capacity to apply world knowledge and better handling of small-object segmentation. The right column showcases VRS-HQ's robust temporal reasoning capabilities, driven by our temporal dynamic aggregation and adaptive keyframe selection. Conversely, VISA misses the target due to incorrect keyframe selection during inference.", + "bbox": [ + 89, + 622, + 482, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Visualization. To understand the TDA's impact, we visualize mask embeddings generated by the mask decoder using both the fused and unfused tokens with keyframe features. As illustrated in Fig. 4, we employ Principal Component Analysis (PCA) [35] to reduce the dimensionality of the mask embeddings to a single channel. This single-channel representation is then blended with the orig", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inal image to highlight high-attention regions. Before TDA fusion, the token often focuses on non-target elements (e.g., the falling person in the first example, the woman in the second, and the central dancing man in the third). In contrast, the similarity-weighted fusion of the frame-level SEG tokens into the token significantly enriches the positional and semantic information contained within the token, enabling the model to more effectively concentrate on the correlated target objects.", + "bbox": [ + 511, + 375, + 906, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 532, + 633, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present VRS-HQ, a novel approach that leverages the temporal reasoning of MLLM and the robust tracking of SAM2 for high-quality Video Reasoning Segmentation. Our method utilizes a temporal-level token and frame-level tokens to capture temporal relations and spatial features, respectively. These tokens are integrated using the Temporal Dynamic Aggregation, with SAM2 utilizing the token for keyframe segmentation and mask propagation. Moreover, we introduce a Token-driven Keyframe Selection that uses the token to generate occlusion scores for robust keyframe selection. Extensive experiments confirm that VRS-HQ achieves state-of-the-art performance on various benchmarks, demonstrating strong capabilities in handling Video Reasoning Segmentation.", + "bbox": [ + 511, + 558, + 908, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 773, + 609, + 789 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. In Adv. Neural Inform. Process. Syst., 2024.", + "[2] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with mul" + ], + "bbox": [ + 521, + 799, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29190", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "timodal transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4985-4995, 2022.", + "[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1209-1218, 2018.", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In *Eur. Conf. Comput. Vis.*, pages 213-229. Springer, 2020.", + "[5] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1971-1978, 2014.", + "[6] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In *Eur. Conf. Comput. Vis.*, pages 640–658. Springer, 2022.", + "[7] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Int. Conf. Comput. Vis., pages 2694-2703, 2023.", + "[8] Sitong Gong, Yunzhi Zhuge, Lu Zhang, Yifan Wang, Pingping Zhang, Lijun Wang, and Huchuan Lu. Avs-mamba: Exploring temporal and multi-modal mamba for audio-visual segmentation. IEEE Trans. Multimedia, 2025.", + "[9] Sitong Gong, Yunzhi Zhuge, Lu Zhang, Pingping Zhang, and Huchuan Lu. Complementary and contrastive learning for audio-visual segmentation. IEEE Trans. Multimedia, 2025.", + "[10] Junwen He, Yifan Wang, Lijun Wang, Huchuan Lu, Jun-Yan He, Jin-Peng Lan, Bin Luo, and Xuansong Xie. Multi-modal instruction tuned llms with fine-grained visual perception. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13980-13990, 2024.", + "[11] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13332-13341, 2024.", + "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.", + "[13] Tianrui Hui, Shaofei Huang, Si Liu, Zihan Ding, Guanbin Li, Wenguan Wang, Jizhong Han, and Fei Wang. Collaborative spatial-temporal modeling for language-queried video actor segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4187–4196, 2021.", + "[14] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13700-13710, 2024.", + "[15] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014." + ], + "bbox": [ + 93, + 90, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Int. Conf. Comput. Vis., pages 4015-4026, 2023.", + "[17] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9579-9589, 2024.", + "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. pages 19730-19742. PMLR, 2023.", + "[19] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Eur. Conf. Comput. Vis., pages 323–340. Springer, 2025.", + "[20] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 23592-23601, 2023.", + "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Adv. Neural Inform. Process. Syst., 36, 2024.", + "[22] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023.", + "[23] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016.", + "[24] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017.", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. pages 8748-8763. PMLR, 2021.", + "[26] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7141-7151, 2023.", + "[27] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13009-13018, 2024.", + "[28] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., pages 3505-3506, 2020.", + "[29] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2:" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29191", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024.", + "[30] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 26374-26383, 2024.", + "[31] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019.", + "[32] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Eur. Conf. Comput. Vis., pages 208–223. Springer, 2020.", + "[33] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Int. Conf. Comput. Vis., pages 3939-3948, 2019.", + "[34] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11686-11695, 2022.", + "[35] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987.", + "[36] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Int. Conf. Comput. Vis., pages 2761-2770, 2023.", + "[37] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4974-4984, 2022.", + "[38] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3858-3869, 2024.", + "[39] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. In Eur. Conf. Comput. Vis., 2024.", + "[40] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In AAAI, pages 6449-6457, 2024.", + "[41] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. arXiv eprints, pages arXiv-2312, 2023.", + "[42] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavo: Language-aware vision transformer for referring image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18155-18165, 2022." + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Yu Qiao, and Hengshuang Zhao. Villa: Video reasoning segmentation with large language model. arXiv preprint arXiv:2407.14500, 2024.", + "[44] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In AAAI, pages 12993-13000, 2020.", + "[45] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017.", + "[46] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with human-intent reasoning. arXiv preprint arXiv:2312.17448, 2023.", + "[47] Mengyuan Zhu, Yunzhi Zhuge, Sitong Gong, Lu Zhang, and Huchuan Lu. Fdavs: Exploring frequency-driven modality enhancement in audio-visual segmentation. In Int. Conf. Multimedia and Expo, 2025.", + "[48] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15116-15127, 2023.", + "[49] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Adv. Neural Inform. Process. Syst., 36, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 487 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29192", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_model.json b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_model.json new file mode 100644 index 0000000000000000000000000000000000000000..95ae08574df7d6b678aa812da6fbc399672fb956 --- /dev/null +++ b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.89, + 0.153 + ], + "angle": 0, + "content": "The Devil is in Temporal Token: High Quality Video Reasoning Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.18, + 0.87, + 0.218 + ], + "angle": 0, + "content": "Sitong Gong\\(^{1}\\), Yunzhi Zhuge\\(^{1*}\\), Lu Zhang\\(^{1}\\), Zongxin Yang\\(^{2}\\), Pingping Zhang\\(^{1}\\), Huchuan Lu\\(^{1}\\), IIAU, Dalian University of Technology, \\(^{2}\\)Havard University" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.22, + 0.678, + 0.234 + ], + "angle": 0, + "content": "stgong@mail.dlut.edu.cn, zgyz@dlut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.651 + ], + "angle": 0, + "content": "Existing methods for Video Reasoning Segmentation rely heavily on a single special token to represent the object in the keyframe or the entire video, inadequately capturing spatial complexity and inter-frame motion. To overcome these challenges, we propose VRS-HQ, an end-to-end video reasoning segmentation approach that leverages Multimodal Large Language Models (MLLMs) to inject rich spatiotemporal features into hierarchical tokens. Our key innovations include a Temporal Dynamic Aggregation (TDA) and a Token-driven Keyframe Selection (TKS). Specifically, we design frame-level \\(<\\mathrm{SEG}>\\) and temporal-level \\(<\\mathrm{TAK}>\\) tokens that utilize MLLM's autoregressive learning to effectively capture both local and global information. Subsequently, we apply a similarity-based weighted fusion and frame selection strategy, then utilize SAM2 to perform keyframe segmentation and propagation. To enhance keyframe localization accuracy, the TKS filters keyframes based on SAM2's occlusion scores during inference. VRS-HQ achieves state-of-the-art performance on ReVOS, surpassing VISA by \\(5.9\\% / 12.5\\% / 9.1\\%\\) in \\(\\mathcal{J}\\&\\mathcal{F}\\) scores across the three subsets. These results highlight the strong temporal reasoning and segmentation capabilities of our method. Code and model weights are available at VRS-HQ." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.222, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Reasoning segmentation [10, 17, 27, 30, 38], which aims to generate segmentation results from complex query texts, has advanced with multimodal foundation models [18, 21, 29]. However, existing methods [10, 17, 27, 30] focus primarily on image-level segmentation, leaving the more challenging domain of video-level segmentation, which requires temporal reasoning of object relations and attributes, relatively unexplored. To address this gap, Video Reasoning Segmentation (VRS) [1, 39, 43] has recently emerged as a promising approach. Unlike Referring Video Object Segmentation (RVOS) methods [36, 37, 40], which rely on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.906, + 0.346 + ], + "angle": 0, + "content": "explicit descriptive phrases like \"a person skateboarding\", VRS leverages the extensive world knowledge and temporal reasoning capabilities of Multimodal Large Language Models (MLLMs) to transform implicit intent-based expressions into precise object masklets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.908, + 0.58 + ], + "angle": 0, + "content": "Despite recent advancements in Video Reasoning Segmentation (VRS), such as VISA [39] and VideoLISA [1], significant challenges still exist. (i) Limited Temporal Context: Existing methods [1, 39] typically rely on a single segmentation token from an MLLM for keyframe-based segmentation (cf. Fig. 1 (a)), resulting in limited temporal context and hindering the effective capture of inter-frame variations and spatiotemporal features. (ii) Suboptimal Keyframe Detection: The LLaMA-VID [19] model, used by VISA for keyframe detection, can produce inaccurate keyframes, particularly in videos requiring complex temporal reasoning. (iii) Decoupled Segmentation and Propagation: VISA's reliance on separate, pre-trained models for keyframe segmentation (SAM [16]) and mask propagation (XMem [6]) prevents end-to-end training and inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address the above obstacles, we introduce several strategies to strengthen the VRS model's proficiency in perceiving spatial information and interpreting temporal dynamics. Firstly, current limitations arise from insufficient single-token representation capacity, which restricts the model's abilities to capture intra-frame spatial features and maintain inter-frame temporal coherence. We hypothesize that encoding frame-level and temporal-level information into hierarchical tokens separately via MLLM and integrating them could effectively unify spatial details with temporal dynamics, enhancing perceptual capability. Secondly, the occlusion score introduced by SAM2 [29] inspires us to incorporate the target confidence of each sampled frame as a criterion for keyframe determination. We thus employ the temporal token in conjunction with SAM2 to generate the occlusion scores, applying temporal information for precise keyframe selection. As shown in Fig. 1 (b), our approach locates the cat more accurately compared to VISA, boosting inference efficiency and segmentation performance. Thirdly, leveraging SAM2's integrated segmentation and propagation capabilities, we efficiently fine" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.239, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29183" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.078, + 0.576, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.075, + 0.9, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.279, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Figure 1. Comparison with previous VRS approaches. (a) Previous methods utilize a single token for keyframe-based segmentation, depending heavily on external models for keyframe detection and mask propagation. This reliance can hinder accurate keyframe localization and prevent end-to-end inference. (b) In contrast, VRS-HQ introduces frame-level and a temporal token for dynamic aggregation. The aggregated token is then used for both keyframe selection and mask generation within SAM2. This enables single-stage inference with precise keyframe selection and high-quality segmentation. (c) VRS-HQ achieves state-of-the-art performance on various image and video datasets across reasoning and referring segmentation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.379, + 0.483, + 0.409 + ], + "angle": 0, + "content": "tune it with our temporal token and its memory mechanism for improved mask quality and inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.413, + 0.484, + 0.806 + ], + "angle": 0, + "content": "In this work, we present VRS-HQ (High-Quality Video Reasoning Segmentation) (cf. Fig. 2), a novel approach that leverages aggregated temporal tokens for enhanced keyframe selection and efficient mask decoding. To begin with, we propose the Temporal Token Encoding (§3.1), prompting the MLLM to encode frame- and video-level target features into multi-level special tokens using sampled video frames and tailored conversation templates. The primary innovation lies in the Temporal Dynamic Aggregation (§3.2), which employs a cosine similarity-based weighted fusion strategy to merge frame-level embeddings into the temporal-level embedding, consolidating spatial features into while maintaining the temporal consistency of the targets. Subsequently, we select the keyframe according to the token similarity between and during training. Moreover, we propose the Token-driven Keyframe Selection (§3.3) during inference, sequentially treating each sampled frame as a potential keyframe and interacting it with fused embedding through SAM2 to calculate an occlusion score. The scores combined with previous token similarity are used as the criterion for keyframe selection. Finally, the keyframe with the fused embedding is fed into SAM2 for mask generation, while the remaining frames are treated as non-keyframes and utilize SAM2's memory mechanism for Mask Decoding and Propagation (§3.4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Through extensive experiments, VRS-HQ achieves state-of-the-art performance across a diverse range of video segmentation benchmarks (cf. Fig. 1 (c)). Specifically, on the ReVOS [39] dataset, VRS-HQ outperforms VISA-13B by \\(9.1\\%\\) in terms of the \\(\\mathcal{J}\\& \\mathcal{F}\\) metric, highlighting the critical role of our proposed module in enhancing the reason" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.468 + ], + "angle": 0, + "content": "ing ability for scenes with complex spatiotemporal dynamics. Moreover, VRS-HQ surpasses previous methods by \\(7.3\\% / 5.6\\% / 6.5\\%\\) \\(\\mathcal{J}\\& \\mathcal{F}\\) on three standard referring video segmentation datasets respectively, underscoring its strong inter-frame perception and robust video tracking capabilities. Key contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.47, + 0.906, + 0.546 + ], + "angle": 0, + "content": "- We present Temporal Dynamic Aggregation to blend spatial features from frame-level tokens into a temporal token, endowing the model with the ability to discern inter-frame variations and comprehend the global semantic context of the targets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.906, + 0.651 + ], + "angle": 0, + "content": "- We focus on harnessing the power of temporal tokens in video perception, using the integrated temporal token via SAM2 for keyframe segmentation and propagation. Additionally, we present the Token-driven Keyframe Selection, combining each sampled frame with the temporal token to generate occlusion scores via SAM2, providing a reliable basis for keyframe detection." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.652, + 0.905, + 0.697 + ], + "angle": 0, + "content": "- By combining the above designs, we introduce VRS-HQ, demonstrating state-of-the-art performance on the VRS benchmark and existing RVOS datasets." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.47, + 0.906, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.715, + 0.663, + 0.731 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.839, + 0.758 + ], + "angle": 0, + "content": "2.1. Referring Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Referring Video Object Segmentation [2, 37, 40] (RVOS), which focuses on segmenting and tracking prominent objects in video frames using explicit textual descriptions, has significantly advanced through the integration of visual and linguistic cues. A large proportion of these studies have leveraged attention mechanisms to merge multimodal information. For instance, Wang et.al. [33] uses asymmetric cross-guided attention to enhance sentence representations and aggregate visual context. Seo et.al. [32] employ cross" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29184" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.085, + 0.885, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.22, + 0.351, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.22, + 0.653, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.22, + 0.882, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.35, + 0.908, + 0.448 + ], + "angle": 0, + "content": "Figure 2. (a) VRS-HQ architecture. VRS-HQ incorporates a Multimodal Large Language Model for Temporal Token Encoding ( and tokens, §3.1), a Temporal Dynamic Aggregation, a Token-driven Keyframe Selection and Mask Decoding and Propogation. (b) Temporal Dynamic Aggregation (TDA) merges frame-level tokens into a temporal token using a weighted fusion based on cosine similarity. (§3.2). (c) Token-driven Keyframe Selection (TKS). During training, the frame with the token closest to the token is selected as the keyframe. During inference, keyframe selection is refined using SAM2's occlusion scores and token similarity scores (§3.3). (d) Mask Decoding and Propagation (MDP). The token provides a sparse embedding for SAM2, generating a keyframe mask and propagating it to other frames via a memory mechanism (§3.4)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.484, + 0.659 + ], + "angle": 0, + "content": "modal and memory attention to jointly address RVOS and semi-supervised video object segmentation. Hui et.al. [13] dynamically recombine linguistic features and interact with visual features using cross-modal attention to highlight spatiotemporal regions of interest. Inspired by DETR [4], methods like MTTR [2] and ReferFormer [37] leverage language queries for precise target localization. Emerging approaches are exploring other-modal guidance for referring segmentation, such as audio cues [8, 9, 47]. While RVOS excels with explicit object descriptions, it often struggles with the complex reasoning required to interpret more implicit and intricate language instructions, motivating our exploration of Video Reasoning Segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.317, + 0.684 + ], + "angle": 0, + "content": "2.2. Reasoning Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Reasoning segmentation [17, 27, 30, 38] advances referring segmentation by generating masks from complex images and versatile text prompts. LISA [17] pioneers the field of reasoning segmentation. It introduces a new token to expand the vocabulary and proposes the embedding-as-mask paradigm, enhancing segmentation capabilities to address scenarios requiring complex reasoning and world knowledge. LISA++ [41] addresses LISA's limitations in distinguishing individual instances by enhancing instruction-tuning data with key segmentation datasets, supporting both semantic and instance segmentation tasks. PixelLM [30] innovatively combines a novel pixel decoder and a segmentation codebook with learnable tokens to efficiently generate high-quality masks without external models. VISA [39]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.6 + ], + "angle": 0, + "content": "introduces reasoning segmentation to the video domain, proposing reasoning video object segmentation. It uses a pre-trained Multimodal Large Language Model (MLLM) to select keyframes, segments them based on reasoning, and propagates the masks to other frames using a pre-trained object tracker. However, it suffers from limited representation capability of a single special token, inaccurate keyframe selection, and the inability to perform end-to-end inference, compromising its segmentation and tracking performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.649, + 0.629 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.846 + ], + "angle": 0, + "content": "Task Definition. The task of Video Reasoning Segmentation can be briefly outlined as follows. Given a video clip consisting of \\(T\\) frames \\(\\mathcal{X}_V\\in \\mathbb{R}^{T\\times 3\\times H\\times W}\\), where \\(H\\) and \\(W\\) denote the height and width of each frame respectively, along with a high-level textual instruction \\(\\mathcal{X}_{txt}\\), VRS aims to design a model \\(\\mathcal{M}\\) to interpret and transform \\(\\mathcal{X}_{txt}\\) into the binary segmentation mask sequence \\(\\mathcal{X}_M\\in \\mathbb{R}^{T\\times H\\times W}\\) for each frame. In contrast to RVOS tasks providing explicit descriptions like \"the person skateboarding,\" VRS typically employs expressions incorporating world knowledge like \"tool(s) for holding garbage\" or temporal logic like \"the ship(s) moving at the highest speed\", raising higher demands on the model's capabilities of temporal relationship comprehension and complex scenario reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Overall Architecture. Fig. 2 illustrates the VRS-HQ architecture, which comprises Chat-UniVi [14] as the Multi-modal Large Language Model (MLLM) for temporal token" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29185" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "encoding (§3.1), a Temporal Dynamic Aggregation (§3.2), a Token-driven Keyframe Selection (§3.3), and Mask Decoding and Propagation (§3.4). The MLLM encodes the textual prompt into multi-level tokens representing spatial and semantic target information. These tokens drive both temporal dynamic aggregation and keyframe selection via cosine similarity scores. SAM2 then segments the keyframes and tracks the target object throughout the video. During inference, the Token-driven Keyframe Selection computes object occlusion scores for each sampled frame by merging it with the fused temporal token. Filtering keyframes based on these scores improves the accurate localization of the target object and subsequently segmentation process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.332, + 0.317 + ], + "angle": 0, + "content": "3.1. Temporal Token Encoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.46 + ], + "angle": 0, + "content": "Current VRS approaches [1, 39] typically use specified language prompts to guide MLLMs in embedding target information from keyframes or entire videos into a single specialized token, struggling to capture rich spatiotemporal dynamics essential for fine-grained video understanding. To equip the MLLM with both frame-level and video-level contextual awareness for video segmentation, we propose encoding intra-frame spatial information and inter-frame temporal relations into hierarchical tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.668 + ], + "angle": 0, + "content": "Hierarchical Token Generation. Instead of previous single-token encoding strategy, the vocabulary of the MLLM is initially augmented with two new special tokens: \\(<\\mathrm{SEG}>\\) and \\(<\\mathrm{TAK}>\\). Then, we design a structured conversational template as \"USER: Please find {expression} in the Reference Video and segment it in each frame and the entire video respectively.\" Here, \"\\(\\{\\mathrm{expression}\\}\\)\" denotes the target object description. The tokenized prompts \\(\\mathcal{X}_{\\text{txt}}\\) and sampled video frames \\(\\mathcal{X}_{\\text{V}'}\\) are input into the MLLM, which generates a response \\(y_{\\text{txt}}\\) containing multiple frame-level \\(<\\mathrm{SEG}>\\) tokens and a temporal-level \\(<\\mathrm{TAK}>\\) token through autoregressive encoding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.784 + ], + "angle": 0, + "content": "Token Extraction and Mapping. Subsequently, the \\(<\\mathrm{SEG}>\\) token embeddings \\(\\bar{h}_{seg} \\in \\mathbb{R}^{T' \\times d'}\\) and the \\(<\\mathrm{TAK}>\\) token embedding \\(\\bar{h}_{tak} \\in \\mathbb{R}^{1 \\times d'}\\) are extracted from the MLLM's final layer, where \\(d'\\) denotes the MLLM's embedding dimension and \\(T'\\) denotes the length of sampled frames. These embeddings are then projected into the same feature space as SAM2 using a multi-layer perceptron:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.793, + 0.483, + 0.812 + ], + "angle": 0, + "content": "\\[\nh _ {s e g}, h _ {t a k} = M L P \\left(\\bar {h} _ {s e g}\\right), M L P \\left(\\bar {h} _ {t a k}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.824, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Here, \\( h_{seg} \\in \\mathbb{R}^{T' \\times d} \\) and \\( h_{tak} \\in \\mathbb{R}^{1 \\times d} \\) represent the sparse embeddings for segmentation mask activation, and \\( d \\) denotes the feature dimension of SAM2. Finally, the frame-level embeddings \\( h_{seg} \\) are aggregated with the video-level embedding \\( h_{tak} \\) to incorporate temporal dynamics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.798, + 0.108 + ], + "angle": 0, + "content": "3.2. Temporal Dynamic Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.219 + ], + "angle": 0, + "content": "Building on the strong temporal encoding of MLLM, and embeddings encapsulate the spatial priors and temporal semantic signals of the targets respectively, providing rich contextual information for the segmentation model. Accordingly, we propose the Temporal Dynamic Aggregation to facilitate the fusion of positional and semantic information of the targets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.907, + 0.359 + ], + "angle": 0, + "content": "Keyframes as Token Similarity. The cosine similarity between each frame-level \\(<\\text{SEG}>\\) token and the temporal-level \\(<\\text{TAK}>\\) token reflects the semantic alignment between individual frames and the overall video context. We hypothesize that high-similarity frames are more representative of the video's content and thus suitable as keyframes. This motivates our use of token similarity for keyframe selection during training, which also facilitates weighted fusion of \\(<\\text{SEG}>\\) and \\(<\\text{TAK}>\\) tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.909, + 0.502 + ], + "angle": 0, + "content": "Similarity-based Weighted Fusion. To enhance the inter-frame consistency while achieving more precise discernment of the object's positional shifts across frames, we fuse the frame-level embeddings \\( h_{seg} \\) with the video-level embedding \\( h_{tak} \\) using normalized cosine similarity scores as attention weights (cf. Fig. 2 (b)). During backpropagation, frame-level features \\( h_{seg} \\) are updated alongside the temporal-level embedding \\( h_{tak} \\) to enable more comprehensive learning of local and global video context:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.513, + 0.907, + 0.555 + ], + "angle": 0, + "content": "\\[\nh _ {t a k} ^ {\\prime} = h _ {t a k} + \\alpha \\sum_ {i = 1} ^ {T ^ {\\prime}} \\lambda_ {i} h _ {s e g} [ i ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.627 + ], + "angle": 0, + "content": "where \\(\\lambda_{i}\\) represents the normalized cosine similarity between \\(h_{seg}[i]\\) and \\(h_{tak}\\), and \\(\\alpha\\) is the fusion coefficient. The resulting fused embedding \\(h_{tak}^{\\prime}\\) is then used for keyframe selection during inference and mask generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.808, + 0.651 + ], + "angle": 0, + "content": "3.3. Token-driven Keyframe Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.793 + ], + "angle": 0, + "content": "VISA [39] relies on an external model (LLaMA-VID [19]) for keyframe selection during inference, hindering end-to-end processing and potentially degrading performance due to inaccuracies in the external model's output. To address this limitation, we introduce the Token-driven Keyframe Selection that leverages the temporal information encoded within the integrated embedding. This approach eliminates the need for complex prompt engineering and significantly improves the reliability of keyframe selection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Instead of uniform video sampling during training, we adopt the CLIP model [25] to find the frame most aligned with the expression \\(\\mathcal{X}_{exp}\\) for inference. The anchor frame is used for global sampling, resulting in sampled frames \\(\\mathcal{X}_V^f\\in \\mathbb{R}^{\\hat{T}\\times 3\\times H\\times W}\\), where \\(\\hat{T}\\) represents the number of the sampled frames. Following the temporal dynamic aggregation (Eq. 2), each sampled frame is treated as a potential" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29186" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.14 + ], + "angle": 0, + "content": "keyframe, along with the fused \\( <\\mathrm{TAK}> \\) embedding \\( h_{\\text{tak}}' \\), is sent to SAM2 to generate the object occlusion scores \\( S_{o} \\in \\mathbb{R}^{\\hat{T} \\times 1} \\) (cf. Fig. 2 (c)):" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.145, + 0.483, + 0.166 + ], + "angle": 0, + "content": "\\[\nS _ {o} = \\mathcal {M D} (\\mathcal {E} (\\mathcal {X} _ {V} ^ {f}), h _ {t a k} ^ {\\prime}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.485, + 0.293 + ], + "angle": 0, + "content": "The occlusion scores effectively indicate the confidence of the target's presence in the current frame, assisting in identifying which frame best corresponds to the information encoded in the token. After determining the keyframe index by the combination of softmax-normalized scores \\( S_{o}^{\\prime} \\) with token similarity scores \\( S_{t} \\), we decode the keyframe mask and then utilize the cross-frame propagation technique of SAM2 to obtain video-level segmentation masks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.381, + 0.317 + ], + "angle": 0, + "content": "3.4. Mask Decoding and Propagation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Following temporal token fusion and keyframe selection, the mask embeddings enriched with positional and semantic information can be yielded by the segmentation model. In contrast to prior methods [1, 39, 43] depending on image-level segmentation models or external object trackers for target trajectories prediction, we utilize SAM2 to perform segmentation and propagation concurrently." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.526 + ], + "angle": 0, + "content": "SAM2 for Masklets Decoding. Given the keyframe \\(\\mathcal{X}_k\\), we first extract its features using the image encoder \\(\\mathcal{E}\\), providing the conditional input for SAM2. The integrated temporal embedding \\(h_{\\text{tak}}'\\) then interacts with these keyframe features within the mask decoder \\(\\mathcal{MD}\\) (cf. Fig. 2 (d)) to generate the keyframe mask:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.533, + 0.483, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} _ {M} ^ {k} = \\mathcal {M D} \\left(\\mathcal {E} \\left(\\mathcal {X} _ {k}\\right), h _ {t a k} ^ {\\prime}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.665 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_M^k\\) represents the predicted mask for the keyframe. Next, we propagate this mask to two adjacent frames, \\(\\mathcal{X}_{k - 1}\\) and \\(\\mathcal{X}_{k + 1}\\), treated as non-conditional frames, using the memory storage and interaction mechanism. This yields the mask sequence \\(\\mathcal{X}_M^s\\). During inference, all remaining frames are processed as non-conditional frames, facilitating mask propagation throughout the entire video." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.281, + 0.688 + ], + "angle": 0, + "content": "3.5. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.484, + 0.755 + ], + "angle": 0, + "content": "Our method is trained end-to-end using a combined text generation loss and mask loss to optimize the \\( \\langle \\mathrm{TAK} \\rangle \\) and \\( \\langle \\mathrm{SEG} \\rangle \\) embeddings. The mask loss combines binary cross-entropy (BCE) and DICE loss:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.763, + 0.483, + 0.78 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = \\lambda_ {t x t} L _ {t x t} \\left(y _ {t x t}, \\hat {y} _ {t x t}\\right) + \\lambda_ {m a s k} L _ {m a s k}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.808, + 0.484, + 0.841 + ], + "angle": 0, + "content": "\\[\nL _ {m a s k} = \\lambda_ {b c e} L _ {b c e} \\left(\\mathcal {X} _ {M} ^ {s}, \\hat {\\mathcal {X}} _ {M} ^ {s}\\right) + \\lambda_ {d i c e} L _ {d i c e} \\left(\\mathcal {X} _ {M} ^ {s}, \\hat {\\mathcal {X}} _ {M} ^ {s}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{y}_{txt}\\) and \\(\\hat{\\chi}_M^s\\) represent the ground truth text and mask, respectively, and \\(y_{txt}\\) and \\(X_M^s\\) are their corresponding predictions. The weighting coefficients \\(\\lambda_{txt}\\), \\(\\lambda_{mask}\\), \\(\\lambda_{bce}\\), and \\(\\lambda_{dice}\\) are set to 1, 1, 2 and 0.5, respectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.712, + 0.13 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Datasets. Our model is trained and evaluated on extensive image and video segmentation datasets, LLaVA-Instruct-150k [21], as well as the Video Question-Answering datasets from Video-ChatGPT [22]. Specifically, the image segmentation datasets comprise semantic segmentation: ADE20K [45], COCO-Stuff [3], PACO-LVIS [26], and PASCALPart [5], referring segmentation: refCLEF, refCOCO, refCOCO+ [15], and refCOCOg [23], and reasoning segmentation: ReasonSeg [17]. While the video segmentation datasets encompass the RVOS datasets: RefYoutube-VOS [32], Ref-DAVIS17 [24] and MeViS [7], and the VRS benchmark ReVOS [39]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.326, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate video segmentation using region similarity \\((\\mathcal{I})\\), contour accuracy \\((\\mathcal{F})\\), and their mean \\((\\mathcal{J} \\& \\mathcal{F})\\). Image segmentation accuracy is measured via Generalized Intersection over Union (gIoU) [31] and Complete Intersection over Union (cIoU) [44]. Model hallucination is assessed using the robustness score \\(\\mathcal{R}\\) [39]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.427, + 0.731, + 0.443 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.449, + 0.907, + 0.632 + ], + "angle": 0, + "content": "We fine-tuned Chat-UniVi [14] (our chosen MLLM) using LoRA [12] (rank 8), optimizing the mask decoder and MLP projection layer while freezing other parameters. Training used AdamW (learning rate 0.0003, no weight decay) with a WarmupDecayLR scheduler (100-iteration warmup). We set the fusion coefficient \\(\\alpha\\) to 0.1 and sampled two non-keyframes per video. Trained for 7500 iterations on a hybrid image/video dataset using four A800 GPUs with DeepSpeed [28] (batch size 1, gradient accumulation 32, total batch size 128), the model employed TDA for video data (token fusion, segmentation, and propagation) and the token for direct segmentation on image data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.642, + 0.706, + 0.658 + ], + "angle": 0, + "content": "4.3. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.664, + 0.906, + 0.726 + ], + "angle": 0, + "content": "To showcase the robust pixel-level perception and generalization of VRS-HQ, we conduct evaluations across diverse benchmarks, including ReVOS, RVOS, and image-based referring and reasoning segmentation datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "ReVOS Datasets. Tab. 1 illustrates the performance comparison with previous methods [2, 17, 37, 39, 46] on the ReVOS benchmark. VRS-HQ demonstrates significant improvements over the previous state-of-the-art, VISA [39] across all ten metrics. Remarkably, in terms of the \\(\\mathcal{J}\\&\\mathcal{F}\\), VRS-HQ-13B surpasses VISA-13B by \\(5.9\\%\\) on the referring subset and \\(12.5\\%\\) on the reasoning subset, respectively. These gains highlight the effectiveness of our temporal aggregation strategy and the utilization of the token for keyframe selection and target segmentation, which enhances the model's reasoning capabilities. In contrast," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29187" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.083, + 0.726, + 0.097 + ], + "angle": 0, + "content": "Table 1. Performance comparison with previous methods on ReVOS dataset." + }, + { + "type": "table", + "bbox": [ + 0.144, + 0.098, + 0.859, + 0.296 + ], + "angle": 0, + "content": "
MethodsBackbonereferringreasoningoverallR
JFJ&FJFJ&FJFJ&F
MTTR [2][CVPR2022]Video-Swin-T29.830.230.020.421.521.025.125.925.55.6
LMPM [7][ICCV2023]Swin-T29.039.134.113.324.318.821.231.726.43.2
ReferFormer [37][CVPR2022]Video-Swin-B31.234.332.721.325.623.426.229.928.18.8
LISA [17][CVPR2024]LLaVA-7B44.347.145.733.838.436.139.142.740.99.3
LISA [17][CVPR2024]LLaVA-13B45.247.946.634.339.136.739.843.541.68.6
TrackGPT [46][arXiv2023]LLaVA-7B46.749.748.236.841.239.041.845.543.611.6
TrackGPT [46][arXiv2023]LLaVA-13B48.350.649.538.142.940.543.246.845.012.8
VISA [39][ECCV2024]LLaVA-7B49.452.651.040.545.843.244.949.247.115.3
VISA [39][ECCV2024]LLaVA-13B55.759.057.441.946.544.248.852.850.815.1
VISA [39][ECCV2024]Chat-UniVi-7B49.252.650.940.645.443.044.949.046.915.5
VISA [39][ECCV2024]Chat-UniVi-13B55.659.157.442.046.744.348.852.950.914.5
VRS-HQ[Ours]Chat-UniVi-7B59.864.562.153.558.756.156.661.659.119.7
VRS-HQ[Ours]Chat-UniVi-13B61.165.563.354.159.456.857.662.560.018.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.209, + 0.305, + 0.787, + 0.319 + ], + "angle": 0, + "content": "Table 2. Performance comparison with previous methods on the validation sets of RVOS datasets." + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.321, + 0.839, + 0.531 + ], + "angle": 0, + "content": "
MethodsBackboneRef-YouTube-VOSRef-DAVIS17MeViS
JFJ&FJFJ&FJFJ&F
MTTR [2][CVPR2022]Video-Swin-T54.056.655.3---28.831.230.0
LMPM [7][ICCV2023]Swin-T------34.240.237.2
ReferFormer [37][CVPR2022]Video-Swin-B61.364.662.958.164.161.129.832.231.0
OnlineRefer [36][CVPR2023]Swin-L61.665.563.561.667.764.8---
DsHmp [11][CVPR2024]Video-Swin-B65.069.167.161.768.164.943.049.846.4
LISA [17][CVPR2024]LLaVA-7B53.454.353.962.267.364.835.139.437.2
LISA [17][CVPR2024]LLaVA-13B54.054.854.463.268.866.035.840.037.9
TrackGPT [46][aXiv2023]LLaVA-7B55.357.456.459.467.063.237.642.640.1
TrackGPT [46][aXiv2023]LLaVA-13B58.160.859.562.770.466.539.243.141.2
VISA [39][ECCV2024]Chat-UniVi-7B59.863.261.566.372.569.440.746.343.5
VISA [39][ECCV2024]Chat-UniVi-13B61.464.763.067.073.870.441.847.144.5
VideoLISA [1][NeurIPS2024]LLaVA-Phi-3-V61.765.763.764.972.768.841.347.644.4
VRS-HQ[Ours]Chat-UniVi-7B68.372.570.472.679.476.047.653.750.6
VRS-HQ[Ours]Chat-UniVi-13B69.073.171.071.077.974.448.053.750.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.64 + ], + "angle": 0, + "content": "VISA relies on LLaMA-VID [19] for keyframe selection during inference. This strategy may neglect intra-frame fine-grained visual details, leading to inaccurate keyframe localization. Furthermore, the robustness score \\(\\mathcal{R}\\) exceeds VISA-13B by \\(4.4\\%\\), indicating VRS-HQ's superior capability in handling negative samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.824 + ], + "angle": 0, + "content": "RVOS. Tab. 2 compares VRS-HQ with state-of-the-art RVOS methods. On Ref-YouTube-VOS and RefDAVIS17, VRS-HQ-13B surpasses VideoLISA [1] and VISA-13B [39], achieving \\(\\mathcal{I}\\&\\mathcal{F}\\) improvements of \\(7.3\\%\\) and \\(5.6\\%\\), respectively. Furthermore, on the motion-intensive MeViS dataset, VRS-HQ-13B yields substantial gains of \\(6.2\\%\\), \\(6.1\\%\\) and \\(6.4\\%\\) in \\(\\mathcal{I}\\), \\(\\mathcal{F}\\) and \\(\\mathcal{I}\\&\\mathcal{F}\\), respectively. demonstrating its effective in capturing motion and maintaining temporal coherence. These improvements, compared to VideoLISA's single-token representation, are attributed to the enhanced inter-frame perception of the TDA and the refined keyframe localization of the TKS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "RIS. Our method seamlessly extends to referring image segmentation (RIS) by treating images as single-frame videos. As shown in Tab. 3, VRS-HQ consistently outperforms VISA on three RIS benchmarks, achieving performance comparable to LISA [17] and VideoLISA [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.701 + ], + "angle": 0, + "content": "On ReasonSeg, VRS-HQ performs competitively, ranking merely below VideoLISA, demonstrating strong generalization. The marginally lower performance on image datasets compared to video datasets likely stems from two factors: (i) our model's training emphasis on video data (unlike LISA, which is pre-trained solely on images); and (ii) our training methodology encourages reliance on SAM2's memory mechanism for multi-frame processing, which may be less effective for single-image segmentation compared to VideoLISA's single-token approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.68, + 0.727 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fusion Coefficient Ablation. The ablation analysis of the fusion coefficient \\(\\alpha\\) in the TDA is illustrated in Tab. 5. An \\(\\alpha\\) of 0.1 yields optimal performance. Setting \\(\\alpha\\) to 0 leads to a sharp drop in the model's performance as \\(<\\mathrm{TAK}>\\) fails to capture fine-grained intra-frame details, and the \\(<\\mathrm{SEG}>\\) token cannot be jointly optimized with \\(<\\mathrm{TAK}>\\) through backpropagation (e.g., \\(62.1\\% \\rightarrow 58.8\\%\\) \\(\\mathcal{J}\\& \\mathcal{F}\\) on the referring subset, \\(56.1\\% \\rightarrow 54.1\\%\\) \\(\\mathcal{J}\\& \\mathcal{F}\\) on the reasoning subset). As \\(\\alpha\\) increases beyond 0.1, the metrics decline slightly, likely due to excessive frame-level noise being introduced into the temporal token at a higher fusion coef-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29188" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.083, + 0.835, + 0.097 + ], + "angle": 0, + "content": "Table 3. Performance comparison with previous methods on referring and reasoning image segmentation datasets." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.098, + 0.875, + 0.246 + ], + "angle": 0, + "content": "
MethodBackbonerefCOCOrefCOCO+refCOCOgReaSeg (val)ReaSeg (test)
valtestAtestBvaltestAtestBvaltestgIoUcIoUgIoUcIoU
CRIS [34]ResNet10170.573.266.162.368.153.759.960.4----
LAVT [42]Swin-B72.775.868.862.168.455.161.262.1----
ReLA [20]Swin-B73.876.570.266.071.057.765.066.0----
X-Decoder [48]DaViT-L------64.6-22.617.921.716.3
SEEM [49]DaViT-L------65.7-25.521.224.318.7
LISA [17]LLaVA-7B74.979.172.365.170.858.167.970.652.954.047.348.4
VISA [39]Chat-UniVi-7B72.475.568.159.864.853.165.566.452.757.8--
VideoLISA [1]LLaVA-Phi-3-V73.876.668.863.468.856.268.368.861.467.153.854.4
VRS-HQChat-UniVi-7B73.577.569.561.767.654.366.767.555.251.851.752.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.267, + 0.361, + 0.294 + ], + "angle": 0, + "content": "Table 4. Ablation analysis of TDA and TKS components." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.296, + 0.354, + 0.352 + ], + "angle": 0, + "content": "
Componentsreferringreasoning
JFJ&FJFJ&F
w/o TDA + TKS54.559.156.848.653.951.2
w/o TDA56.561.158.851.556.754.1
w/o TKS57.862.460.150.956.253.5
VRS-HQ59.864.562.153.558.756.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.374, + 0.363, + 0.403 + ], + "angle": 0, + "content": "Table 7. Ablation analysis of sampled frames during inference stage." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.41, + 0.34, + 0.482 + ], + "angle": 0, + "content": "
Framesreferringreasoning
JFJ&FJFJ&F
158.462.760.649.855.152.5
358.663.060.851.056.253.6
658.663.360.951.557.054.3
958.963.761.352.557.855.1
1259.364.061.752.557.855.1
Non-fixed59.864.562.153.558.756.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.362, + 0.263, + 0.633, + 0.289 + ], + "angle": 0, + "content": "Table 5. Ablation analysis of the fusion coefficient \\( \\alpha \\) ." + }, + { + "type": "table", + "bbox": [ + 0.397, + 0.291, + 0.601, + 0.346 + ], + "angle": 0, + "content": "
αreferringreasoning
JFJ&FJFJ&F
056.561.158.851.556.754.1
0.159.864.562.153.558.756.1
0.2558.963.861.352.357.855.0
0.558.263.060.651.456.654.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.362, + 0.353, + 0.633, + 0.407 + ], + "angle": 0, + "content": "Table 8. Ablation analysis of the mask decoding and propagation strategy. MT/MI: Multiframe Training/Inference. ST/SI: Single-frame Training/Inference." + }, + { + "type": "table", + "bbox": [ + 0.375, + 0.409, + 0.622, + 0.472 + ], + "angle": 0, + "content": "
Strategyreferringreasoning
JFJ&FJFJ&F
SAM+ST+SI48.352.450.342.947.445.2
SAM2+ST+SI55.959.857.849.354.051.6
SAM2+ST+MI55.860.858.346.552.549.5
SAM2+MT+SI54.157.455.848.252.250.2
SAM2+MT+MI59.864.562.153.558.756.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.633, + 0.254, + 0.905, + 0.309 + ], + "angle": 0, + "content": "Table 6. Ablation analysis of Token-driven Keyframe Selection. \\( S_{1} \\), \\( S_{2} \\), and \\( S_{3} \\) represent CLIP scores, token similarity scores, and occlusion scores, respectively." + }, + { + "type": "table", + "bbox": [ + 0.652, + 0.31, + 0.89, + 0.365 + ], + "angle": 0, + "content": "
S1S2S3referringreasoning
JFJ&FJFJ&F
✓✓✓59.764.261.952.557.855.2
✓✓57.862.460.150.956.253.6
✓✓59.664.161.852.557.755.1
✓✓59.864.562.153.558.756.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.633, + 0.37, + 0.904, + 0.398 + ], + "angle": 0, + "content": "Table 9. Ablation analysis of sampling strategy." + }, + { + "type": "table", + "bbox": [ + 0.647, + 0.4, + 0.894, + 0.446 + ], + "angle": 0, + "content": "
Sampling strategyreferringreasoning
JFJ&FJFJ&F
Random Sampling59.063.661.353.258.455.8
Uniform Sampling59.363.961.653.358.455.8
CLIP Sampling59.864.562.153.558.756.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.483, + 0.545 + ], + "angle": 0, + "content": "ficient (e.g., \\(62.1\\% \\rightarrow 60.6\\%\\) \\(\\mathcal{J}\\& \\mathcal{F}\\) on the referring subset, \\(56.1\\% \\rightarrow 54.0\\%\\) \\(\\mathcal{J}\\& \\mathcal{F}\\) on the reasoning subset)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.556, + 0.483, + 0.784 + ], + "angle": 0, + "content": "Token-driven Keyframe Selection. Tab. 6 analyzes the impact of different score combinations for keyframe selection within the TKS. Results indicate that the occlusion score is the most influential of the three considered, improving \\(\\mathcal{J}\\& \\mathcal{F}\\) by \\(1.8\\%\\) and \\(1.6\\%\\) on the referring and reasoning subsets, respectively (first two rows). This improvement stems from the occlusion score's ability to reflect the target object's presence within each frame. The model achieves its best performance when keyframe selection is based on a combination of token similarity and the occlusion score. However, incorporating the CLIP score (rows 1 and 4) slightly degrades performance (by \\(0.2\\%\\) and \\(0.9\\%\\) \\(\\mathcal{J}\\& \\mathcal{F}\\) on the referring and reasoning subsets, respectively), likely due to CLIP's limited spatiotemporal understanding, which hinders accurate keyframe selection." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Mask Decoding and Propagation. Tab. 8 analyzes the impact of different mask decoding and propagation strategies on segmentation performance. Using only SAM and TDA for single-frame segmentation (first row) already surpasses VISA-13B by \\(0.9\\% / 0.7\\% / 0.9\\%\\) on the reasoning subset. Comparing the first two rows reveals the substantial advantage of SAM2 over SAM for single-frame fine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.635 + ], + "angle": 0, + "content": "tuning and inference. SAM2 achieves improvements of \\(7.6\\% / 7.4\\% / 7.5\\% / 6.4\\% / 6.6\\% / 6.4\\%\\), demonstrating its superior robustness in segmentation and tracking. Furthermore, incorporating multi-frame propagation during both training and inference (last row) yields the best overall results, effectively leveraging SAM2's memory mechanism for keyframe mask storage. This approach leads to further improvements in \\(\\mathcal{J} \\& \\mathcal{F}\\) of \\(3.8\\%\\) and \\(4.5\\%\\) compared to the other strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.797 + ], + "angle": 0, + "content": "Inference-Time Sampling Strategy. Tab. 9 analyzes the impact of different frame sampling strategies during inference. \"Random\" denotes randomly selected frames, while \"Uniform\" extracts frames at equal intervals. CLIP sampling slightly outperforms uniform sampling, improving \\(\\mathcal{J} \\& \\mathcal{F}\\) by \\(0.5\\% / 0.3\\%\\) on the referring and reasoning subsets, respectively. This modest gain is likely due to CLIP's ability in selecting frames that are semantically aligned with the referring expression, which in turn facilitates the reasoning process of the MLLM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.722, + 0.82 + ], + "angle": 0, + "content": "4.5. Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Segmentation Map Comparison. Fig. 3 presents the qualitative comparison of VRS-HQ and VISA on the ReVOS dataset across three different scenarios. In the left case, VISA displays ambiguity in spatial positioning and neglects critical references such as \"the first\", revealing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29189" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.085, + 0.881, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.332, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Segmentation map comparison of VISA and VRS-HQ on the ReVOS benchmark (§4.5). Results across three scenarios demonstrate that VRS-HQ excels in reasoning complex spatial and temporal relationships, delivering enhanced segmentation performance." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.38, + 0.465, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Figure 4. Visualization of feature maps (§4.5). From top to bottom are: (a) Ground truth masks. b) Keyframe mask embeddings generated by the token before TDA. (c) Keyframe mask embeddings generated by the token after TDA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.789 + ], + "angle": 0, + "content": "its limited semantic understanding. In contrast, VRS-HQ demonstrates a strong ability to perceive spatiotemporal relationships. For the middle example, VISA mistakenly segments the person flying the kite, while VRS-HQ correctly identifies the kite, reflecting the stronger capacity to apply world knowledge and better handling of small-object segmentation. The right column showcases VRS-HQ's robust temporal reasoning capabilities, driven by our temporal dynamic aggregation and adaptive keyframe selection. Conversely, VISA misses the target due to incorrect keyframe selection during inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Feature Visualization. To understand the TDA's impact, we visualize mask embeddings generated by the mask decoder using both the fused and unfused tokens with keyframe features. As illustrated in Fig. 4, we employ Principal Component Analysis (PCA) [35] to reduce the dimensionality of the mask embeddings to a single channel. This single-channel representation is then blended with the orig" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.513 + ], + "angle": 0, + "content": "inal image to highlight high-attention regions. Before TDA fusion, the token often focuses on non-target elements (e.g., the falling person in the first example, the woman in the second, and the central dancing man in the third). In contrast, the similarity-weighted fusion of the frame-level SEG tokens into the token significantly enriches the positional and semantic information contained within the token, enabling the model to more effectively concentrate on the correlated target objects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.533, + 0.634, + 0.549 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.909, + 0.756 + ], + "angle": 0, + "content": "We present VRS-HQ, a novel approach that leverages the temporal reasoning of MLLM and the robust tracking of SAM2 for high-quality Video Reasoning Segmentation. Our method utilizes a temporal-level token and frame-level tokens to capture temporal relations and spatial features, respectively. These tokens are integrated using the Temporal Dynamic Aggregation, with SAM2 utilizing the token for keyframe segmentation and mask propagation. Moreover, we introduce a Token-driven Keyframe Selection that uses the token to generate occlusion scores for robust keyframe selection. Extensive experiments confirm that VRS-HQ achieves state-of-the-art performance on various benchmarks, demonstrating strong capabilities in handling Video Reasoning Segmentation." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.775, + 0.611, + 0.79 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.8, + 0.907, + 0.869 + ], + "angle": 0, + "content": "[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. In Adv. Neural Inform. Process. Syst., 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[2] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with mul" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.8, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29190" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "timodal transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4985-4995, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.122, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1209-1218, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.165, + 0.483, + 0.221 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In *Eur. Conf. Comput. Vis.*, pages 213-229. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.222, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[5] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1971-1978, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[6] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In *Eur. Conf. Comput. Vis.*, pages 640–658. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[7] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Int. Conf. Comput. Vis., pages 2694-2703, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.407, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[8] Sitong Gong, Yunzhi Zhuge, Lu Zhang, Yifan Wang, Pingping Zhang, Lijun Wang, and Huchuan Lu. Avs-mamba: Exploring temporal and multi-modal mamba for audio-visual segmentation. IEEE Trans. Multimedia, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.482, + 0.505 + ], + "angle": 0, + "content": "[9] Sitong Gong, Yunzhi Zhuge, Lu Zhang, Pingping Zhang, and Huchuan Lu. Complementary and contrastive learning for audio-visual segmentation. IEEE Trans. Multimedia, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.507, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[10] Junwen He, Yifan Wang, Lijun Wang, Huchuan Lu, Jun-Yan He, Jin-Peng Lan, Bin Luo, and Xuansong Xie. Multi-modal instruction tuned llms with fine-grained visual perception. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13980-13990, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.577, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[11] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13332-13341, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.634, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[13] Tianrui Hui, Shaofei Huang, Si Liu, Zihan Ding, Guanbin Li, Wenguan Wang, Jizhong Han, and Fei Wang. Collaborative spatial-temporal modeling for language-queried video actor segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4187–4196, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[14] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13700-13710, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Int. Conf. Comput. Vis., pages 4015-4026, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[17] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9579-9589, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. pages 19730-19742. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[19] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Eur. Conf. Comput. Vis., pages 323–340. Springer, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[20] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 23592-23601, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Adv. Neural Inform. Process. Syst., 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.393, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[22] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.45, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[23] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[24] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.564, + 0.905, + 0.632 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.634, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[26] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7141-7151, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[27] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13009-13018, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[28] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., pages 3505-3506, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29191" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[30] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 26374-26383, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[31] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[32] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Eur. Conf. Comput. Vis., pages 208–223. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[33] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Int. Conf. Comput. Vis., pages 3939-3948, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[34] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11686-11695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[35] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[36] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Int. Conf. Comput. Vis., pages 2761-2770, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[37] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4974-4984, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[38] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3858-3869, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[39] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. In Eur. Conf. Comput. Vis., 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[40] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In AAAI, pages 6449-6457, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[41] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. arXiv eprints, pages arXiv-2312, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavo: Language-aware vision transformer for referring image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18155-18165, 2022." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[43] Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Yu Qiao, and Hengshuang Zhao. Villa: Video reasoning segmentation with large language model. arXiv preprint arXiv:2407.14500, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[44] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In AAAI, pages 12993-13000, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[45] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[46] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with human-intent reasoning. arXiv preprint arXiv:2312.17448, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[47] Mengyuan Zhu, Yunzhi Zhuge, Sitong Gong, Lu Zhang, and Huchuan Lu. Fdavs: Exploring frequency-driven modality enhancement in audio-visual segmentation. In Int. Conf. Multimedia and Expo, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[48] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15116-15127, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[49] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Adv. Neural Inform. Process. Syst., 36, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29192" + } + ] +] \ No newline at end of file diff --git a/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_origin.pdf b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9d685ac2d1d1ed09b678b95711d507a50823b6a9 --- /dev/null +++ b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/a6997446-be6b-4107-b5e4-8b2f54d90026_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:173f8e1bb4ba3442fb96da0259271ab8c69106bedbfcc486d980f60eb3846478 +size 3844253 diff --git a/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/full.md b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4d6c0042f9d661c948b61be3d11005ad6b9f990c --- /dev/null +++ b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/full.md @@ -0,0 +1,289 @@ +# The Devil is in Temporal Token: High Quality Video Reasoning Segmentation + +Sitong Gong $^{1}$ , Yunzhi Zhuge $^{1*}$ , Lu Zhang $^{1}$ , Zongxin Yang $^{2}$ , Pingping Zhang $^{1}$ , Huchuan Lu $^{1}$ , IIAU, Dalian University of Technology, $^{2}$ Havard University + +stgong@mail.dlut.edu.cn, zgyz@dlut.edu.cn + +# Abstract + +Existing methods for Video Reasoning Segmentation rely heavily on a single special token to represent the object in the keyframe or the entire video, inadequately capturing spatial complexity and inter-frame motion. To overcome these challenges, we propose VRS-HQ, an end-to-end video reasoning segmentation approach that leverages Multimodal Large Language Models (MLLMs) to inject rich spatiotemporal features into hierarchical tokens. Our key innovations include a Temporal Dynamic Aggregation (TDA) and a Token-driven Keyframe Selection (TKS). Specifically, we design frame-level $<\mathrm{SEG}>$ and temporal-level $<\mathrm{TAK}>$ tokens that utilize MLLM's autoregressive learning to effectively capture both local and global information. Subsequently, we apply a similarity-based weighted fusion and frame selection strategy, then utilize SAM2 to perform keyframe segmentation and propagation. To enhance keyframe localization accuracy, the TKS filters keyframes based on SAM2's occlusion scores during inference. VRS-HQ achieves state-of-the-art performance on ReVOS, surpassing VISA by $5.9\% / 12.5\% / 9.1\%$ in $\mathcal{J}\&\mathcal{F}$ scores across the three subsets. These results highlight the strong temporal reasoning and segmentation capabilities of our method. Code and model weights are available at VRS-HQ. + +# 1. Introduction + +Reasoning segmentation [10, 17, 27, 30, 38], which aims to generate segmentation results from complex query texts, has advanced with multimodal foundation models [18, 21, 29]. However, existing methods [10, 17, 27, 30] focus primarily on image-level segmentation, leaving the more challenging domain of video-level segmentation, which requires temporal reasoning of object relations and attributes, relatively unexplored. To address this gap, Video Reasoning Segmentation (VRS) [1, 39, 43] has recently emerged as a promising approach. Unlike Referring Video Object Segmentation (RVOS) methods [36, 37, 40], which rely on + +explicit descriptive phrases like "a person skateboarding", VRS leverages the extensive world knowledge and temporal reasoning capabilities of Multimodal Large Language Models (MLLMs) to transform implicit intent-based expressions into precise object masklets. + +Despite recent advancements in Video Reasoning Segmentation (VRS), such as VISA [39] and VideoLISA [1], significant challenges still exist. (i) Limited Temporal Context: Existing methods [1, 39] typically rely on a single segmentation token from an MLLM for keyframe-based segmentation (cf. Fig. 1 (a)), resulting in limited temporal context and hindering the effective capture of inter-frame variations and spatiotemporal features. (ii) Suboptimal Keyframe Detection: The LLaMA-VID [19] model, used by VISA for keyframe detection, can produce inaccurate keyframes, particularly in videos requiring complex temporal reasoning. (iii) Decoupled Segmentation and Propagation: VISA's reliance on separate, pre-trained models for keyframe segmentation (SAM [16]) and mask propagation (XMem [6]) prevents end-to-end training and inference. + +To address the above obstacles, we introduce several strategies to strengthen the VRS model's proficiency in perceiving spatial information and interpreting temporal dynamics. Firstly, current limitations arise from insufficient single-token representation capacity, which restricts the model's abilities to capture intra-frame spatial features and maintain inter-frame temporal coherence. We hypothesize that encoding frame-level and temporal-level information into hierarchical tokens separately via MLLM and integrating them could effectively unify spatial details with temporal dynamics, enhancing perceptual capability. Secondly, the occlusion score introduced by SAM2 [29] inspires us to incorporate the target confidence of each sampled frame as a criterion for keyframe determination. We thus employ the temporal token in conjunction with SAM2 to generate the occlusion scores, applying temporal information for precise keyframe selection. As shown in Fig. 1 (b), our approach locates the cat more accurately compared to VISA, boosting inference efficiency and segmentation performance. Thirdly, leveraging SAM2's integrated segmentation and propagation capabilities, we efficiently fine + +![](images/d41f3819214bf9e878ec752bd52386ee8e970d6507b1e9a01e59a4a64c871939.jpg) +Figure 1. Comparison with previous VRS approaches. (a) Previous methods utilize a single token for keyframe-based segmentation, depending heavily on external models for keyframe detection and mask propagation. This reliance can hinder accurate keyframe localization and prevent end-to-end inference. (b) In contrast, VRS-HQ introduces frame-level and a temporal token for dynamic aggregation. The aggregated token is then used for both keyframe selection and mask generation within SAM2. This enables single-stage inference with precise keyframe selection and high-quality segmentation. (c) VRS-HQ achieves state-of-the-art performance on various image and video datasets across reasoning and referring segmentation. + +![](images/37cc4e7a23c85283f4ae285ebf9555f94f044cc04f4d4e4c811871501e380e4a.jpg) + +tune it with our temporal token and its memory mechanism for improved mask quality and inference efficiency. + +In this work, we present VRS-HQ (High-Quality Video Reasoning Segmentation) (cf. Fig. 2), a novel approach that leverages aggregated temporal tokens for enhanced keyframe selection and efficient mask decoding. To begin with, we propose the Temporal Token Encoding (§3.1), prompting the MLLM to encode frame- and video-level target features into multi-level special tokens using sampled video frames and tailored conversation templates. The primary innovation lies in the Temporal Dynamic Aggregation (§3.2), which employs a cosine similarity-based weighted fusion strategy to merge frame-level embeddings into the temporal-level embedding, consolidating spatial features into while maintaining the temporal consistency of the targets. Subsequently, we select the keyframe according to the token similarity between and during training. Moreover, we propose the Token-driven Keyframe Selection (§3.3) during inference, sequentially treating each sampled frame as a potential keyframe and interacting it with fused embedding through SAM2 to calculate an occlusion score. The scores combined with previous token similarity are used as the criterion for keyframe selection. Finally, the keyframe with the fused embedding is fed into SAM2 for mask generation, while the remaining frames are treated as non-keyframes and utilize SAM2's memory mechanism for Mask Decoding and Propagation (§3.4). + +Through extensive experiments, VRS-HQ achieves state-of-the-art performance across a diverse range of video segmentation benchmarks (cf. Fig. 1 (c)). Specifically, on the ReVOS [39] dataset, VRS-HQ outperforms VISA-13B by $9.1\%$ in terms of the $\mathcal{J}\& \mathcal{F}$ metric, highlighting the critical role of our proposed module in enhancing the reason + +ing ability for scenes with complex spatiotemporal dynamics. Moreover, VRS-HQ surpasses previous methods by $7.3\% / 5.6\% / 6.5\%$ $\mathcal{J}\& \mathcal{F}$ on three standard referring video segmentation datasets respectively, underscoring its strong inter-frame perception and robust video tracking capabilities. Key contributions can be summarized as follows: + +- We present Temporal Dynamic Aggregation to blend spatial features from frame-level tokens into a temporal token, endowing the model with the ability to discern inter-frame variations and comprehend the global semantic context of the targets. +- We focus on harnessing the power of temporal tokens in video perception, using the integrated temporal token via SAM2 for keyframe segmentation and propagation. Additionally, we present the Token-driven Keyframe Selection, combining each sampled frame with the temporal token to generate occlusion scores via SAM2, providing a reliable basis for keyframe detection. +- By combining the above designs, we introduce VRS-HQ, demonstrating state-of-the-art performance on the VRS benchmark and existing RVOS datasets. + +# 2. Related Works + +# 2.1. Referring Video Object Segmentation + +Referring Video Object Segmentation [2, 37, 40] (RVOS), which focuses on segmenting and tracking prominent objects in video frames using explicit textual descriptions, has significantly advanced through the integration of visual and linguistic cues. A large proportion of these studies have leveraged attention mechanisms to merge multimodal information. For instance, Wang et.al. [33] uses asymmetric cross-guided attention to enhance sentence representations and aggregate visual context. Seo et.al. [32] employ cross + +![](images/b6c6cac0afd566350e6f3b963597ea3df3ff9609e6ae3d90ad5c53f9760462ee.jpg) + +![](images/dd405aea2e5fd8d494bac8a0d26dd8bb56f0476b6e32ccbcee3f7ddd8eee11b6.jpg) +Figure 2. (a) VRS-HQ architecture. VRS-HQ incorporates a Multimodal Large Language Model for Temporal Token Encoding ( and tokens, §3.1), a Temporal Dynamic Aggregation, a Token-driven Keyframe Selection and Mask Decoding and Propogation. (b) Temporal Dynamic Aggregation (TDA) merges frame-level tokens into a temporal token using a weighted fusion based on cosine similarity. (§3.2). (c) Token-driven Keyframe Selection (TKS). During training, the frame with the token closest to the token is selected as the keyframe. During inference, keyframe selection is refined using SAM2's occlusion scores and token similarity scores (§3.3). (d) Mask Decoding and Propagation (MDP). The token provides a sparse embedding for SAM2, generating a keyframe mask and propagating it to other frames via a memory mechanism (§3.4). + +![](images/1345cde1f7187f6e0b405e800a596e4fd69ff976dd5c90c0d90156dc512da619.jpg) + +![](images/dea18d298f5a29e51fe45d5fb41d2de62e27f5d3b92d453940d84d65407a084b.jpg) + +modal and memory attention to jointly address RVOS and semi-supervised video object segmentation. Hui et.al. [13] dynamically recombine linguistic features and interact with visual features using cross-modal attention to highlight spatiotemporal regions of interest. Inspired by DETR [4], methods like MTTR [2] and ReferFormer [37] leverage language queries for precise target localization. Emerging approaches are exploring other-modal guidance for referring segmentation, such as audio cues [8, 9, 47]. While RVOS excels with explicit object descriptions, it often struggles with the complex reasoning required to interpret more implicit and intricate language instructions, motivating our exploration of Video Reasoning Segmentation. + +# 2.2. Reasoning Segmentation + +Reasoning segmentation [17, 27, 30, 38] advances referring segmentation by generating masks from complex images and versatile text prompts. LISA [17] pioneers the field of reasoning segmentation. It introduces a new token to expand the vocabulary and proposes the embedding-as-mask paradigm, enhancing segmentation capabilities to address scenarios requiring complex reasoning and world knowledge. LISA++ [41] addresses LISA's limitations in distinguishing individual instances by enhancing instruction-tuning data with key segmentation datasets, supporting both semantic and instance segmentation tasks. PixelLM [30] innovatively combines a novel pixel decoder and a segmentation codebook with learnable tokens to efficiently generate high-quality masks without external models. VISA [39] + +introduces reasoning segmentation to the video domain, proposing reasoning video object segmentation. It uses a pre-trained Multimodal Large Language Model (MLLM) to select keyframes, segments them based on reasoning, and propagates the masks to other frames using a pre-trained object tracker. However, it suffers from limited representation capability of a single special token, inaccurate keyframe selection, and the inability to perform end-to-end inference, compromising its segmentation and tracking performance. + +# 3. Methodology + +Task Definition. The task of Video Reasoning Segmentation can be briefly outlined as follows. Given a video clip consisting of $T$ frames $\mathcal{X}_V\in \mathbb{R}^{T\times 3\times H\times W}$ , where $H$ and $W$ denote the height and width of each frame respectively, along with a high-level textual instruction $\mathcal{X}_{txt}$ , VRS aims to design a model $\mathcal{M}$ to interpret and transform $\mathcal{X}_{txt}$ into the binary segmentation mask sequence $\mathcal{X}_M\in \mathbb{R}^{T\times H\times W}$ for each frame. In contrast to RVOS tasks providing explicit descriptions like "the person skateboarding," VRS typically employs expressions incorporating world knowledge like "tool(s) for holding garbage" or temporal logic like "the ship(s) moving at the highest speed", raising higher demands on the model's capabilities of temporal relationship comprehension and complex scenario reasoning. + +Overall Architecture. Fig. 2 illustrates the VRS-HQ architecture, which comprises Chat-UniVi [14] as the Multi-modal Large Language Model (MLLM) for temporal token + +encoding (§3.1), a Temporal Dynamic Aggregation (§3.2), a Token-driven Keyframe Selection (§3.3), and Mask Decoding and Propagation (§3.4). The MLLM encodes the textual prompt into multi-level tokens representing spatial and semantic target information. These tokens drive both temporal dynamic aggregation and keyframe selection via cosine similarity scores. SAM2 then segments the keyframes and tracks the target object throughout the video. During inference, the Token-driven Keyframe Selection computes object occlusion scores for each sampled frame by merging it with the fused temporal token. Filtering keyframes based on these scores improves the accurate localization of the target object and subsequently segmentation process. + +# 3.1. Temporal Token Encoding + +Current VRS approaches [1, 39] typically use specified language prompts to guide MLLMs in embedding target information from keyframes or entire videos into a single specialized token, struggling to capture rich spatiotemporal dynamics essential for fine-grained video understanding. To equip the MLLM with both frame-level and video-level contextual awareness for video segmentation, we propose encoding intra-frame spatial information and inter-frame temporal relations into hierarchical tokens. + +Hierarchical Token Generation. Instead of previous single-token encoding strategy, the vocabulary of the MLLM is initially augmented with two new special tokens: $<\mathrm{SEG}>$ and $<\mathrm{TAK}>$ . Then, we design a structured conversational template as "USER: Please find {expression} in the Reference Video and segment it in each frame and the entire video respectively." Here, " $\{\mathrm{expression}\}$ " denotes the target object description. The tokenized prompts $\mathcal{X}_{\text{txt}}$ and sampled video frames $\mathcal{X}_{\text{V}'}$ are input into the MLLM, which generates a response $y_{\text{txt}}$ containing multiple frame-level $<\mathrm{SEG}>$ tokens and a temporal-level $<\mathrm{TAK}>$ token through autoregressive encoding. + +Token Extraction and Mapping. Subsequently, the $<\mathrm{SEG}>$ token embeddings $\bar{h}_{seg} \in \mathbb{R}^{T' \times d'}$ and the $<\mathrm{TAK}>$ token embedding $\bar{h}_{tak} \in \mathbb{R}^{1 \times d'}$ are extracted from the MLLM's final layer, where $d'$ denotes the MLLM's embedding dimension and $T'$ denotes the length of sampled frames. These embeddings are then projected into the same feature space as SAM2 using a multi-layer perceptron: + +$$ +h _ {s e g}, h _ {t a k} = M L P \left(\bar {h} _ {s e g}\right), M L P \left(\bar {h} _ {t a k}\right) \tag {1} +$$ + +Here, $h_{seg} \in \mathbb{R}^{T' \times d}$ and $h_{tak} \in \mathbb{R}^{1 \times d}$ represent the sparse embeddings for segmentation mask activation, and $d$ denotes the feature dimension of SAM2. Finally, the frame-level embeddings $h_{seg}$ are aggregated with the video-level embedding $h_{tak}$ to incorporate temporal dynamics. + +# 3.2. Temporal Dynamic Aggregation + +Building on the strong temporal encoding of MLLM, and embeddings encapsulate the spatial priors and temporal semantic signals of the targets respectively, providing rich contextual information for the segmentation model. Accordingly, we propose the Temporal Dynamic Aggregation to facilitate the fusion of positional and semantic information of the targets. + +Keyframes as Token Similarity. The cosine similarity between each frame-level $<\text{SEG}>$ token and the temporal-level $<\text{TAK}>$ token reflects the semantic alignment between individual frames and the overall video context. We hypothesize that high-similarity frames are more representative of the video's content and thus suitable as keyframes. This motivates our use of token similarity for keyframe selection during training, which also facilitates weighted fusion of $<\text{SEG}>$ and $<\text{TAK}>$ tokens. + +Similarity-based Weighted Fusion. To enhance the inter-frame consistency while achieving more precise discernment of the object's positional shifts across frames, we fuse the frame-level embeddings $h_{seg}$ with the video-level embedding $h_{tak}$ using normalized cosine similarity scores as attention weights (cf. Fig. 2 (b)). During backpropagation, frame-level features $h_{seg}$ are updated alongside the temporal-level embedding $h_{tak}$ to enable more comprehensive learning of local and global video context: + +$$ +h _ {t a k} ^ {\prime} = h _ {t a k} + \alpha \sum_ {i = 1} ^ {T ^ {\prime}} \lambda_ {i} h _ {s e g} [ i ], \tag {2} +$$ + +where $\lambda_{i}$ represents the normalized cosine similarity between $h_{seg}[i]$ and $h_{tak}$ , and $\alpha$ is the fusion coefficient. The resulting fused embedding $h_{tak}^{\prime}$ is then used for keyframe selection during inference and mask generation. + +# 3.3. Token-driven Keyframe Selection + +VISA [39] relies on an external model (LLaMA-VID [19]) for keyframe selection during inference, hindering end-to-end processing and potentially degrading performance due to inaccuracies in the external model's output. To address this limitation, we introduce the Token-driven Keyframe Selection that leverages the temporal information encoded within the integrated embedding. This approach eliminates the need for complex prompt engineering and significantly improves the reliability of keyframe selection. + +Instead of uniform video sampling during training, we adopt the CLIP model [25] to find the frame most aligned with the expression $\mathcal{X}_{exp}$ for inference. The anchor frame is used for global sampling, resulting in sampled frames $\mathcal{X}_V^f\in \mathbb{R}^{\hat{T}\times 3\times H\times W}$ , where $\hat{T}$ represents the number of the sampled frames. Following the temporal dynamic aggregation (Eq. 2), each sampled frame is treated as a potential + +keyframe, along with the fused $<\mathrm{TAK}>$ embedding $h_{\text{tak}}'$ , is sent to SAM2 to generate the object occlusion scores $S_{o} \in \mathbb{R}^{\hat{T} \times 1}$ (cf. Fig. 2 (c)): + +$$ +S _ {o} = \mathcal {M D} (\mathcal {E} (\mathcal {X} _ {V} ^ {f}), h _ {t a k} ^ {\prime}) \tag {3} +$$ + +The occlusion scores effectively indicate the confidence of the target's presence in the current frame, assisting in identifying which frame best corresponds to the information encoded in the token. After determining the keyframe index by the combination of softmax-normalized scores $S_{o}^{\prime}$ with token similarity scores $S_{t}$ , we decode the keyframe mask and then utilize the cross-frame propagation technique of SAM2 to obtain video-level segmentation masks. + +# 3.4. Mask Decoding and Propagation + +Following temporal token fusion and keyframe selection, the mask embeddings enriched with positional and semantic information can be yielded by the segmentation model. In contrast to prior methods [1, 39, 43] depending on image-level segmentation models or external object trackers for target trajectories prediction, we utilize SAM2 to perform segmentation and propagation concurrently. + +SAM2 for Masklets Decoding. Given the keyframe $\mathcal{X}_k$ , we first extract its features using the image encoder $\mathcal{E}$ , providing the conditional input for SAM2. The integrated temporal embedding $h_{\text{tak}}'$ then interacts with these keyframe features within the mask decoder $\mathcal{MD}$ (cf. Fig. 2 (d)) to generate the keyframe mask: + +$$ +\mathcal {X} _ {M} ^ {k} = \mathcal {M D} \left(\mathcal {E} \left(\mathcal {X} _ {k}\right), h _ {t a k} ^ {\prime}\right), \tag {4} +$$ + +where $\mathcal{X}_M^k$ represents the predicted mask for the keyframe. Next, we propagate this mask to two adjacent frames, $\mathcal{X}_{k - 1}$ and $\mathcal{X}_{k + 1}$ , treated as non-conditional frames, using the memory storage and interaction mechanism. This yields the mask sequence $\mathcal{X}_M^s$ . During inference, all remaining frames are processed as non-conditional frames, facilitating mask propagation throughout the entire video. + +# 3.5. Training Objectives + +Our method is trained end-to-end using a combined text generation loss and mask loss to optimize the $\langle \mathrm{TAK} \rangle$ and $\langle \mathrm{SEG} \rangle$ embeddings. The mask loss combines binary cross-entropy (BCE) and DICE loss: + +$$ +L _ {t o t a l} = \lambda_ {t x t} L _ {t x t} \left(y _ {t x t}, \hat {y} _ {t x t}\right) + \lambda_ {m a s k} L _ {m a s k}, \tag {5} +$$ + +$$ +L _ {m a s k} = \lambda_ {b c e} L _ {b c e} \left(\mathcal {X} _ {M} ^ {s}, \hat {\mathcal {X}} _ {M} ^ {s}\right) + \lambda_ {d i c e} L _ {d i c e} \left(\mathcal {X} _ {M} ^ {s}, \hat {\mathcal {X}} _ {M} ^ {s}\right), \tag {6} +$$ + +where $\hat{y}_{txt}$ and $\hat{\chi}_M^s$ represent the ground truth text and mask, respectively, and $y_{txt}$ and $X_M^s$ are their corresponding predictions. The weighting coefficients $\lambda_{txt}$ , $\lambda_{mask}$ , $\lambda_{bce}$ , and $\lambda_{dice}$ are set to 1, 1, 2 and 0.5, respectively. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +Datasets. Our model is trained and evaluated on extensive image and video segmentation datasets, LLaVA-Instruct-150k [21], as well as the Video Question-Answering datasets from Video-ChatGPT [22]. Specifically, the image segmentation datasets comprise semantic segmentation: ADE20K [45], COCO-Stuff [3], PACO-LVIS [26], and PASCALPart [5], referring segmentation: refCLEF, refCOCO, refCOCO+ [15], and refCOCOg [23], and reasoning segmentation: ReasonSeg [17]. While the video segmentation datasets encompass the RVOS datasets: RefYoutube-VOS [32], Ref-DAVIS17 [24] and MeViS [7], and the VRS benchmark ReVOS [39]. + +Evaluation Metrics. We evaluate video segmentation using region similarity $(\mathcal{I})$ , contour accuracy $(\mathcal{F})$ , and their mean $(\mathcal{J} \& \mathcal{F})$ . Image segmentation accuracy is measured via Generalized Intersection over Union (gIoU) [31] and Complete Intersection over Union (cIoU) [44]. Model hallucination is assessed using the robustness score $\mathcal{R}$ [39]. + +# 4.2. Implementation Details + +We fine-tuned Chat-UniVi [14] (our chosen MLLM) using LoRA [12] (rank 8), optimizing the mask decoder and MLP projection layer while freezing other parameters. Training used AdamW (learning rate 0.0003, no weight decay) with a WarmupDecayLR scheduler (100-iteration warmup). We set the fusion coefficient $\alpha$ to 0.1 and sampled two non-keyframes per video. Trained for 7500 iterations on a hybrid image/video dataset using four A800 GPUs with DeepSpeed [28] (batch size 1, gradient accumulation 32, total batch size 128), the model employed TDA for video data (token fusion, segmentation, and propagation) and the token for direct segmentation on image data. + +# 4.3. Comparison Results + +To showcase the robust pixel-level perception and generalization of VRS-HQ, we conduct evaluations across diverse benchmarks, including ReVOS, RVOS, and image-based referring and reasoning segmentation datasets. + +ReVOS Datasets. Tab. 1 illustrates the performance comparison with previous methods [2, 17, 37, 39, 46] on the ReVOS benchmark. VRS-HQ demonstrates significant improvements over the previous state-of-the-art, VISA [39] across all ten metrics. Remarkably, in terms of the $\mathcal{J}\&\mathcal{F}$ , VRS-HQ-13B surpasses VISA-13B by $5.9\%$ on the referring subset and $12.5\%$ on the reasoning subset, respectively. These gains highlight the effectiveness of our temporal aggregation strategy and the utilization of the token for keyframe selection and target segmentation, which enhances the model's reasoning capabilities. In contrast, + +Table 1. Performance comparison with previous methods on ReVOS dataset. + +
MethodsBackbonereferringreasoningoverallR
JFJ&FJFJ&FJFJ&F
MTTR [2][CVPR2022]Video-Swin-T29.830.230.020.421.521.025.125.925.55.6
LMPM [7][ICCV2023]Swin-T29.039.134.113.324.318.821.231.726.43.2
ReferFormer [37][CVPR2022]Video-Swin-B31.234.332.721.325.623.426.229.928.18.8
LISA [17][CVPR2024]LLaVA-7B44.347.145.733.838.436.139.142.740.99.3
LISA [17][CVPR2024]LLaVA-13B45.247.946.634.339.136.739.843.541.68.6
TrackGPT [46][arXiv2023]LLaVA-7B46.749.748.236.841.239.041.845.543.611.6
TrackGPT [46][arXiv2023]LLaVA-13B48.350.649.538.142.940.543.246.845.012.8
VISA [39][ECCV2024]LLaVA-7B49.452.651.040.545.843.244.949.247.115.3
VISA [39][ECCV2024]LLaVA-13B55.759.057.441.946.544.248.852.850.815.1
VISA [39][ECCV2024]Chat-UniVi-7B49.252.650.940.645.443.044.949.046.915.5
VISA [39][ECCV2024]Chat-UniVi-13B55.659.157.442.046.744.348.852.950.914.5
VRS-HQ[Ours]Chat-UniVi-7B59.864.562.153.558.756.156.661.659.119.7
VRS-HQ[Ours]Chat-UniVi-13B61.165.563.354.159.456.857.662.560.018.9
+ +Table 2. Performance comparison with previous methods on the validation sets of RVOS datasets. + +
MethodsBackboneRef-YouTube-VOSRef-DAVIS17MeViS
JFJ&FJFJ&FJFJ&F
MTTR [2][CVPR2022]Video-Swin-T54.056.655.3---28.831.230.0
LMPM [7][ICCV2023]Swin-T------34.240.237.2
ReferFormer [37][CVPR2022]Video-Swin-B61.364.662.958.164.161.129.832.231.0
OnlineRefer [36][CVPR2023]Swin-L61.665.563.561.667.764.8---
DsHmp [11][CVPR2024]Video-Swin-B65.069.167.161.768.164.943.049.846.4
LISA [17][CVPR2024]LLaVA-7B53.454.353.962.267.364.835.139.437.2
LISA [17][CVPR2024]LLaVA-13B54.054.854.463.268.866.035.840.037.9
TrackGPT [46][aXiv2023]LLaVA-7B55.357.456.459.467.063.237.642.640.1
TrackGPT [46][aXiv2023]LLaVA-13B58.160.859.562.770.466.539.243.141.2
VISA [39][ECCV2024]Chat-UniVi-7B59.863.261.566.372.569.440.746.343.5
VISA [39][ECCV2024]Chat-UniVi-13B61.464.763.067.073.870.441.847.144.5
VideoLISA [1][NeurIPS2024]LLaVA-Phi-3-V61.765.763.764.972.768.841.347.644.4
VRS-HQ[Ours]Chat-UniVi-7B68.372.570.472.679.476.047.653.750.6
VRS-HQ[Ours]Chat-UniVi-13B69.073.171.071.077.974.448.053.750.9
+ +VISA relies on LLaMA-VID [19] for keyframe selection during inference. This strategy may neglect intra-frame fine-grained visual details, leading to inaccurate keyframe localization. Furthermore, the robustness score $\mathcal{R}$ exceeds VISA-13B by $4.4\%$ , indicating VRS-HQ's superior capability in handling negative samples. + +RVOS. Tab. 2 compares VRS-HQ with state-of-the-art RVOS methods. On Ref-YouTube-VOS and RefDAVIS17, VRS-HQ-13B surpasses VideoLISA [1] and VISA-13B [39], achieving $\mathcal{I}\&\mathcal{F}$ improvements of $7.3\%$ and $5.6\%$ , respectively. Furthermore, on the motion-intensive MeViS dataset, VRS-HQ-13B yields substantial gains of $6.2\%$ , $6.1\%$ and $6.4\%$ in $\mathcal{I}$ , $\mathcal{F}$ and $\mathcal{I}\&\mathcal{F}$ , respectively. demonstrating its effective in capturing motion and maintaining temporal coherence. These improvements, compared to VideoLISA's single-token representation, are attributed to the enhanced inter-frame perception of the TDA and the refined keyframe localization of the TKS. + +RIS. Our method seamlessly extends to referring image segmentation (RIS) by treating images as single-frame videos. As shown in Tab. 3, VRS-HQ consistently outperforms VISA on three RIS benchmarks, achieving performance comparable to LISA [17] and VideoLISA [1]. + +On ReasonSeg, VRS-HQ performs competitively, ranking merely below VideoLISA, demonstrating strong generalization. The marginally lower performance on image datasets compared to video datasets likely stems from two factors: (i) our model's training emphasis on video data (unlike LISA, which is pre-trained solely on images); and (ii) our training methodology encourages reliance on SAM2's memory mechanism for multi-frame processing, which may be less effective for single-image segmentation compared to VideoLISA's single-token approach. + +# 4.4. Ablation Studies + +Fusion Coefficient Ablation. The ablation analysis of the fusion coefficient $\alpha$ in the TDA is illustrated in Tab. 5. An $\alpha$ of 0.1 yields optimal performance. Setting $\alpha$ to 0 leads to a sharp drop in the model's performance as $<\mathrm{TAK}>$ fails to capture fine-grained intra-frame details, and the $<\mathrm{SEG}>$ token cannot be jointly optimized with $<\mathrm{TAK}>$ through backpropagation (e.g., $62.1\% \rightarrow 58.8\%$ $\mathcal{J}\& \mathcal{F}$ on the referring subset, $56.1\% \rightarrow 54.1\%$ $\mathcal{J}\& \mathcal{F}$ on the reasoning subset). As $\alpha$ increases beyond 0.1, the metrics decline slightly, likely due to excessive frame-level noise being introduced into the temporal token at a higher fusion coef- + +Table 3. Performance comparison with previous methods on referring and reasoning image segmentation datasets. + +
MethodBackbonerefCOCOrefCOCO+refCOCOgReaSeg (val)ReaSeg (test)
valtestAtestBvaltestAtestBvaltestgIoUcIoUgIoUcIoU
CRIS [34]ResNet10170.573.266.162.368.153.759.960.4----
LAVT [42]Swin-B72.775.868.862.168.455.161.262.1----
ReLA [20]Swin-B73.876.570.266.071.057.765.066.0----
X-Decoder [48]DaViT-L------64.6-22.617.921.716.3
SEEM [49]DaViT-L------65.7-25.521.224.318.7
LISA [17]LLaVA-7B74.979.172.365.170.858.167.970.652.954.047.348.4
VISA [39]Chat-UniVi-7B72.475.568.159.864.853.165.566.452.757.8--
VideoLISA [1]LLaVA-Phi-3-V73.876.668.863.468.856.268.368.861.467.153.854.4
VRS-HQChat-UniVi-7B73.577.569.561.767.654.366.767.555.251.851.752.9
+ +Table 4. Ablation analysis of TDA and TKS components. + +
Componentsreferringreasoning
JFJ&FJFJ&F
w/o TDA + TKS54.559.156.848.653.951.2
w/o TDA56.561.158.851.556.754.1
w/o TKS57.862.460.150.956.253.5
VRS-HQ59.864.562.153.558.756.1
+ +Table 7. Ablation analysis of sampled frames during inference stage. + +
Framesreferringreasoning
JFJ&FJFJ&F
158.462.760.649.855.152.5
358.663.060.851.056.253.6
658.663.360.951.557.054.3
958.963.761.352.557.855.1
1259.364.061.752.557.855.1
Non-fixed59.864.562.153.558.756.1
+ +Table 5. Ablation analysis of the fusion coefficient $\alpha$ . + +
αreferringreasoning
JFJ&FJFJ&F
056.561.158.851.556.754.1
0.159.864.562.153.558.756.1
0.2558.963.861.352.357.855.0
0.558.263.060.651.456.654.0
+ +Table 8. Ablation analysis of the mask decoding and propagation strategy. MT/MI: Multiframe Training/Inference. ST/SI: Single-frame Training/Inference. + +
Strategyreferringreasoning
JFJ&FJFJ&F
SAM+ST+SI48.352.450.342.947.445.2
SAM2+ST+SI55.959.857.849.354.051.6
SAM2+ST+MI55.860.858.346.552.549.5
SAM2+MT+SI54.157.455.848.252.250.2
SAM2+MT+MI59.864.562.153.558.756.1
+ +Table 6. Ablation analysis of Token-driven Keyframe Selection. $S_{1}$ , $S_{2}$ , and $S_{3}$ represent CLIP scores, token similarity scores, and occlusion scores, respectively. + +
S1S2S3referringreasoning
JFJ&FJFJ&F
✓✓✓59.764.261.952.557.855.2
✓✓57.862.460.150.956.253.6
✓✓59.664.161.852.557.755.1
✓✓59.864.562.153.558.756.1
+ +Table 9. Ablation analysis of sampling strategy. + +
Sampling strategyreferringreasoning
JFJ&FJFJ&F
Random Sampling59.063.661.353.258.455.8
Uniform Sampling59.363.961.653.358.455.8
CLIP Sampling59.864.562.153.558.756.1
+ +ficient (e.g., $62.1\% \rightarrow 60.6\%$ $\mathcal{J}\& \mathcal{F}$ on the referring subset, $56.1\% \rightarrow 54.0\%$ $\mathcal{J}\& \mathcal{F}$ on the reasoning subset). + +Token-driven Keyframe Selection. Tab. 6 analyzes the impact of different score combinations for keyframe selection within the TKS. Results indicate that the occlusion score is the most influential of the three considered, improving $\mathcal{J}\& \mathcal{F}$ by $1.8\%$ and $1.6\%$ on the referring and reasoning subsets, respectively (first two rows). This improvement stems from the occlusion score's ability to reflect the target object's presence within each frame. The model achieves its best performance when keyframe selection is based on a combination of token similarity and the occlusion score. However, incorporating the CLIP score (rows 1 and 4) slightly degrades performance (by $0.2\%$ and $0.9\%$ $\mathcal{J}\& \mathcal{F}$ on the referring and reasoning subsets, respectively), likely due to CLIP's limited spatiotemporal understanding, which hinders accurate keyframe selection. + +Mask Decoding and Propagation. Tab. 8 analyzes the impact of different mask decoding and propagation strategies on segmentation performance. Using only SAM and TDA for single-frame segmentation (first row) already surpasses VISA-13B by $0.9\% / 0.7\% / 0.9\%$ on the reasoning subset. Comparing the first two rows reveals the substantial advantage of SAM2 over SAM for single-frame fine + +tuning and inference. SAM2 achieves improvements of $7.6\% / 7.4\% / 7.5\% / 6.4\% / 6.6\% / 6.4\%$ , demonstrating its superior robustness in segmentation and tracking. Furthermore, incorporating multi-frame propagation during both training and inference (last row) yields the best overall results, effectively leveraging SAM2's memory mechanism for keyframe mask storage. This approach leads to further improvements in $\mathcal{J} \& \mathcal{F}$ of $3.8\%$ and $4.5\%$ compared to the other strategies. + +Inference-Time Sampling Strategy. Tab. 9 analyzes the impact of different frame sampling strategies during inference. "Random" denotes randomly selected frames, while "Uniform" extracts frames at equal intervals. CLIP sampling slightly outperforms uniform sampling, improving $\mathcal{J} \& \mathcal{F}$ by $0.5\% / 0.3\%$ on the referring and reasoning subsets, respectively. This modest gain is likely due to CLIP's ability in selecting frames that are semantically aligned with the referring expression, which in turn facilitates the reasoning process of the MLLM. + +# 4.5. Visualization Analysis + +Segmentation Map Comparison. Fig. 3 presents the qualitative comparison of VRS-HQ and VISA on the ReVOS dataset across three different scenarios. In the left case, VISA displays ambiguity in spatial positioning and neglects critical references such as "the first", revealing + +![](images/132ba830efe9971c912ef8b6c19dc8cca0c4b4434a8ddf345229bcfab2da322a.jpg) +Figure 3. Segmentation map comparison of VISA and VRS-HQ on the ReVOS benchmark (§4.5). Results across three scenarios demonstrate that VRS-HQ excels in reasoning complex spatial and temporal relationships, delivering enhanced segmentation performance. + +![](images/52f97c3ac936688321cbf323ebd1fcc59de8592d6bb39aa4d3a4289919c78b8d.jpg) +Figure 4. Visualization of feature maps (§4.5). From top to bottom are: (a) Ground truth masks. b) Keyframe mask embeddings generated by the token before TDA. (c) Keyframe mask embeddings generated by the token after TDA. + +its limited semantic understanding. In contrast, VRS-HQ demonstrates a strong ability to perceive spatiotemporal relationships. For the middle example, VISA mistakenly segments the person flying the kite, while VRS-HQ correctly identifies the kite, reflecting the stronger capacity to apply world knowledge and better handling of small-object segmentation. The right column showcases VRS-HQ's robust temporal reasoning capabilities, driven by our temporal dynamic aggregation and adaptive keyframe selection. Conversely, VISA misses the target due to incorrect keyframe selection during inference. + +Feature Visualization. To understand the TDA's impact, we visualize mask embeddings generated by the mask decoder using both the fused and unfused tokens with keyframe features. As illustrated in Fig. 4, we employ Principal Component Analysis (PCA) [35] to reduce the dimensionality of the mask embeddings to a single channel. This single-channel representation is then blended with the orig + +inal image to highlight high-attention regions. Before TDA fusion, the token often focuses on non-target elements (e.g., the falling person in the first example, the woman in the second, and the central dancing man in the third). In contrast, the similarity-weighted fusion of the frame-level SEG tokens into the token significantly enriches the positional and semantic information contained within the token, enabling the model to more effectively concentrate on the correlated target objects. + +# 5. Conclusion + +We present VRS-HQ, a novel approach that leverages the temporal reasoning of MLLM and the robust tracking of SAM2 for high-quality Video Reasoning Segmentation. Our method utilizes a temporal-level token and frame-level tokens to capture temporal relations and spatial features, respectively. These tokens are integrated using the Temporal Dynamic Aggregation, with SAM2 utilizing the token for keyframe segmentation and mask propagation. Moreover, we introduce a Token-driven Keyframe Selection that uses the token to generate occlusion scores for robust keyframe selection. Extensive experiments confirm that VRS-HQ achieves state-of-the-art performance on various benchmarks, demonstrating strong capabilities in handling Video Reasoning Segmentation. + +# References + +[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. In Adv. Neural Inform. Process. Syst., 2024. +[2] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with mul + +timodal transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4985-4995, 2022. +[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1209-1218, 2018. +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In *Eur. Conf. Comput. Vis.*, pages 213-229. Springer, 2020. +[5] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1971-1978, 2014. +[6] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In *Eur. Conf. Comput. Vis.*, pages 640–658. Springer, 2022. +[7] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Int. Conf. Comput. Vis., pages 2694-2703, 2023. +[8] Sitong Gong, Yunzhi Zhuge, Lu Zhang, Yifan Wang, Pingping Zhang, Lijun Wang, and Huchuan Lu. Avs-mamba: Exploring temporal and multi-modal mamba for audio-visual segmentation. IEEE Trans. Multimedia, 2025. +[9] Sitong Gong, Yunzhi Zhuge, Lu Zhang, Pingping Zhang, and Huchuan Lu. Complementary and contrastive learning for audio-visual segmentation. IEEE Trans. Multimedia, 2025. +[10] Junwen He, Yifan Wang, Lijun Wang, Huchuan Lu, Jun-Yan He, Jin-Peng Lan, Bin Luo, and Xuansong Xie. Multi-modal instruction tuned llms with fine-grained visual perception. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13980-13990, 2024. +[11] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13332-13341, 2024. +[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. +[13] Tianrui Hui, Shaofei Huang, Si Liu, Zihan Ding, Guanbin Li, Wenguan Wang, Jizhong Han, and Fei Wang. Collaborative spatial-temporal modeling for language-queried video actor segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4187–4196, 2021. +[14] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13700-13710, 2024. +[15] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. + +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Int. Conf. Comput. Vis., pages 4015-4026, 2023. +[17] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9579-9589, 2024. +[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. pages 19730-19742. PMLR, 2023. +[19] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Eur. Conf. Comput. Vis., pages 323–340. Springer, 2025. +[20] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 23592-23601, 2023. +[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Adv. Neural Inform. Process. Syst., 36, 2024. +[22] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. +[23] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. +[24] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. pages 8748-8763. PMLR, 2021. +[26] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7141-7151, 2023. +[27] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13009-13018, 2024. +[28] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., pages 3505-3506, 2020. +[29] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: + +Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. +[30] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 26374-26383, 2024. +[31] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. +[32] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Eur. Conf. Comput. Vis., pages 208–223. Springer, 2020. +[33] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Int. Conf. Comput. Vis., pages 3939-3948, 2019. +[34] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11686-11695, 2022. +[35] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987. +[36] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Int. Conf. Comput. Vis., pages 2761-2770, 2023. +[37] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4974-4984, 2022. +[38] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3858-3869, 2024. +[39] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. In Eur. Conf. Comput. Vis., 2024. +[40] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In AAAI, pages 6449-6457, 2024. +[41] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. arXiv eprints, pages arXiv-2312, 2023. +[42] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavo: Language-aware vision transformer for referring image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18155-18165, 2022. + +[43] Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Yu Qiao, and Hengshuang Zhao. Villa: Video reasoning segmentation with large language model. arXiv preprint arXiv:2407.14500, 2024. +[44] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In AAAI, pages 12993-13000, 2020. +[45] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. +[46] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with human-intent reasoning. arXiv preprint arXiv:2312.17448, 2023. +[47] Mengyuan Zhu, Yunzhi Zhuge, Sitong Gong, Lu Zhang, and Huchuan Lu. Fdavs: Exploring frequency-driven modality enhancement in audio-visual segmentation. In Int. Conf. Multimedia and Expo, 2025. +[48] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15116-15127, 2023. +[49] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Adv. Neural Inform. Process. Syst., 36, 2024. \ No newline at end of file diff --git a/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/images.zip b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8193234d99430ab9b8ab4f1a7c3e3791afd4d325 --- /dev/null +++ b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d26a6712ea21447f1f357ff061211436d430311e4c2b8f667ba87b44f4a01a03 +size 741424 diff --git a/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/layout.json b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..72993e220af21ea143681a15fd1167011d1553d4 --- /dev/null +++ b/2025/The Devil is in Temporal Token_ High Quality Video Reasoning Segmentation/layout.json @@ -0,0 +1,8873 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "text", + "content": "The Devil is in Temporal Token: High Quality Video Reasoning Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "text", + "content": "Sitong Gong" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "text", + "content": ", Yunzhi Zhuge" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "text", + "content": ", Lu Zhang" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "text", + "content": ", Zongxin Yang" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "text", + "content": ", Pingping Zhang" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "text", + "content": ", Huchuan Lu" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "text", + "content": ", IIAU, Dalian University of Technology, " + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 532, + 172 + ], + "type": "text", + "content": "Havard University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 174, + 414, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 174, + 414, + 185 + ], + "spans": [ + { + "bbox": [ + 192, + 174, + 414, + 185 + ], + "type": "text", + "content": "stgong@mail.dlut.edu.cn, zgyz@dlut.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "text", + "content": "Existing methods for Video Reasoning Segmentation rely heavily on a single special token to represent the object in the keyframe or the entire video, inadequately capturing spatial complexity and inter-frame motion. To overcome these challenges, we propose VRS-HQ, an end-to-end video reasoning segmentation approach that leverages Multimodal Large Language Models (MLLMs) to inject rich spatiotemporal features into hierarchical tokens. Our key innovations include a Temporal Dynamic Aggregation (TDA) and a Token-driven Keyframe Selection (TKS). Specifically, we design frame-level " + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "text", + "content": " and temporal-level " + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "inline_equation", + "content": "<\\mathrm{TAK}>" + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "text", + "content": " tokens that utilize MLLM's autoregressive learning to effectively capture both local and global information. Subsequently, we apply a similarity-based weighted fusion and frame selection strategy, then utilize SAM2 to perform keyframe segmentation and propagation. To enhance keyframe localization accuracy, the TKS filters keyframes based on SAM2's occlusion scores during inference. VRS-HQ achieves state-of-the-art performance on ReVOS, surpassing VISA by " + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "inline_equation", + "content": "5.9\\% / 12.5\\% / 9.1\\%" + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "text", + "content": " scores across the three subsets. These results highlight the strong temporal reasoning and segmentation capabilities of our method. Code and model weights are available at VRS-HQ." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "type": "text", + "content": "Reasoning segmentation [10, 17, 27, 30, 38], which aims to generate segmentation results from complex query texts, has advanced with multimodal foundation models [18, 21, 29]. However, existing methods [10, 17, 27, 30] focus primarily on image-level segmentation, leaving the more challenging domain of video-level segmentation, which requires temporal reasoning of object relations and attributes, relatively unexplored. To address this gap, Video Reasoning Segmentation (VRS) [1, 39, 43] has recently emerged as a promising approach. Unlike Referring Video Object Segmentation (RVOS) methods [36, 37, 40], which rely on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "text", + "content": "explicit descriptive phrases like \"a person skateboarding\", VRS leverages the extensive world knowledge and temporal reasoning capabilities of Multimodal Large Language Models (MLLMs) to transform implicit intent-based expressions into precise object masklets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 278, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 459 + ], + "type": "text", + "content": "Despite recent advancements in Video Reasoning Segmentation (VRS), such as VISA [39] and VideoLISA [1], significant challenges still exist. (i) Limited Temporal Context: Existing methods [1, 39] typically rely on a single segmentation token from an MLLM for keyframe-based segmentation (cf. Fig. 1 (a)), resulting in limited temporal context and hindering the effective capture of inter-frame variations and spatiotemporal features. (ii) Suboptimal Keyframe Detection: The LLaMA-VID [19] model, used by VISA for keyframe detection, can produce inaccurate keyframes, particularly in videos requiring complex temporal reasoning. (iii) Decoupled Segmentation and Propagation: VISA's reliance on separate, pre-trained models for keyframe segmentation (SAM [16]) and mask propagation (XMem [6]) prevents end-to-end training and inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": "To address the above obstacles, we introduce several strategies to strengthen the VRS model's proficiency in perceiving spatial information and interpreting temporal dynamics. Firstly, current limitations arise from insufficient single-token representation capacity, which restricts the model's abilities to capture intra-frame spatial features and maintain inter-frame temporal coherence. We hypothesize that encoding frame-level and temporal-level information into hierarchical tokens separately via MLLM and integrating them could effectively unify spatial details with temporal dynamics, enhancing perceptual capability. Secondly, the occlusion score introduced by SAM2 [29] inspires us to incorporate the target confidence of each sampled frame as a criterion for keyframe determination. We thus employ the temporal token in conjunction with SAM2 to generate the occlusion scores, applying temporal information for precise keyframe selection. As shown in Fig. 1 (b), our approach locates the cat more accurately compared to VISA, boosting inference efficiency and segmentation performance. Thirdly, leveraging SAM2's integrated segmentation and propagation capabilities, we efficiently fine" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29183" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 61, + 352, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 61, + 352, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 61, + 352, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 61, + 352, + 220 + ], + "type": "image", + "image_path": "d41f3819214bf9e878ec752bd52386ee8e970d6507b1e9a01e59a4a64c871939.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 220, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 54, + 220, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 555, + 288 + ], + "type": "text", + "content": "Figure 1. Comparison with previous VRS approaches. (a) Previous methods utilize a single token for keyframe-based segmentation, depending heavily on external models for keyframe detection and mask propagation. This reliance can hinder accurate keyframe localization and prevent end-to-end inference. (b) In contrast, VRS-HQ introduces frame-level and a temporal token for dynamic aggregation. The aggregated token is then used for both keyframe selection and mask generation within SAM2. This enables single-stage inference with precise keyframe selection and high-quality segmentation. (c) VRS-HQ achieves state-of-the-art performance on various image and video datasets across reasoning and referring segmentation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 353, + 59, + 550, + 220 + ], + "blocks": [ + { + "bbox": [ + 353, + 59, + 550, + 220 + ], + "lines": [ + { + "bbox": [ + 353, + 59, + 550, + 220 + ], + "spans": [ + { + "bbox": [ + 353, + 59, + 550, + 220 + ], + "type": "image", + "image_path": "37cc4e7a23c85283f4ae285ebf9555f94f044cc04f4d4e4c811871501e380e4a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 300, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 295, + 323 + ], + "type": "text", + "content": "tune it with our temporal token and its memory mechanism for improved mask quality and inference efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 327, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 638 + ], + "type": "text", + "content": "In this work, we present VRS-HQ (High-Quality Video Reasoning Segmentation) (cf. Fig. 2), a novel approach that leverages aggregated temporal tokens for enhanced keyframe selection and efficient mask decoding. To begin with, we propose the Temporal Token Encoding (§3.1), prompting the MLLM to encode frame- and video-level target features into multi-level special tokens using sampled video frames and tailored conversation templates. The primary innovation lies in the Temporal Dynamic Aggregation (§3.2), which employs a cosine similarity-based weighted fusion strategy to merge frame-level embeddings into the temporal-level embedding, consolidating spatial features into while maintaining the temporal consistency of the targets. Subsequently, we select the keyframe according to the token similarity between and during training. Moreover, we propose the Token-driven Keyframe Selection (§3.3) during inference, sequentially treating each sampled frame as a potential keyframe and interacting it with fused embedding through SAM2 to calculate an occlusion score. The scores combined with previous token similarity are used as the criterion for keyframe selection. Finally, the keyframe with the fused embedding is fed into SAM2 for mask generation, while the remaining frames are treated as non-keyframes and utilize SAM2's memory mechanism for Mask Decoding and Propagation (§3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Through extensive experiments, VRS-HQ achieves state-of-the-art performance across a diverse range of video segmentation benchmarks (cf. Fig. 1 (c)). Specifically, on the ReVOS [39] dataset, VRS-HQ outperforms VISA-13B by " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "9.1\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " in terms of the " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " metric, highlighting the critical role of our proposed module in enhancing the reason" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 299, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 370 + ], + "type": "text", + "content": "ing ability for scenes with complex spatiotemporal dynamics. Moreover, VRS-HQ surpasses previous methods by " + }, + { + "bbox": [ + 313, + 299, + 555, + 370 + ], + "type": "inline_equation", + "content": "7.3\\% / 5.6\\% / 6.5\\%" + }, + { + "bbox": [ + 313, + 299, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 299, + 555, + 370 + ], + "type": "text", + "content": " on three standard referring video segmentation datasets respectively, underscoring its strong inter-frame perception and robust video tracking capabilities. Key contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 372, + 554, + 552 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 372, + 554, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 372, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 314, + 372, + 554, + 432 + ], + "type": "text", + "content": "- We present Temporal Dynamic Aggregation to blend spatial features from frame-level tokens into a temporal token, endowing the model with the ability to discern inter-frame variations and comprehend the global semantic context of the targets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 432, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 554, + 515 + ], + "type": "text", + "content": "- We focus on harnessing the power of temporal tokens in video perception, using the integrated temporal token via SAM2 for keyframe segmentation and propagation. Additionally, we present the Token-driven Keyframe Selection, combining each sampled frame with the temporal token to generate occlusion scores via SAM2, providing a reliable basis for keyframe detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 516, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 553, + 552 + ], + "type": "text", + "content": "- By combining the above designs, we introduce VRS-HQ, demonstrating state-of-the-art performance on the VRS benchmark and existing RVOS datasets." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 566, + 405, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 566, + 405, + 578 + ], + "spans": [ + { + "bbox": [ + 314, + 566, + 405, + 578 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 586, + 513, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 513, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 513, + 600 + ], + "type": "text", + "content": "2.1. Referring Video Object Segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Referring Video Object Segmentation [2, 37, 40] (RVOS), which focuses on segmenting and tracking prominent objects in video frames using explicit textual descriptions, has significantly advanced through the integration of visual and linguistic cues. A large proportion of these studies have leveraged attention mechanisms to merge multimodal information. For instance, Wang et.al. [33] uses asymmetric cross-guided attention to enhance sentence representations and aggregate visual context. Seo et.al. [32] employ cross" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29184" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 67, + 541, + 172 + ], + "blocks": [ + { + "bbox": [ + 69, + 67, + 541, + 172 + ], + "lines": [ + { + "bbox": [ + 69, + 67, + 541, + 172 + ], + "spans": [ + { + "bbox": [ + 69, + 67, + 541, + 172 + ], + "type": "image", + "image_path": "b6c6cac0afd566350e6f3b963597ea3df3ff9609e6ae3d90ad5c53f9760462ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 174, + 214, + 277 + ], + "blocks": [ + { + "bbox": [ + 69, + 174, + 214, + 277 + ], + "lines": [ + { + "bbox": [ + 69, + 174, + 214, + 277 + ], + "spans": [ + { + "bbox": [ + 69, + 174, + 214, + 277 + ], + "type": "image", + "image_path": "dd405aea2e5fd8d494bac8a0d26dd8bb56f0476b6e32ccbcee3f7ddd8eee11b6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 277, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 54, + 277, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 555, + 354 + ], + "type": "text", + "content": "Figure 2. (a) VRS-HQ architecture. VRS-HQ incorporates a Multimodal Large Language Model for Temporal Token Encoding ( and tokens, §3.1), a Temporal Dynamic Aggregation, a Token-driven Keyframe Selection and Mask Decoding and Propogation. (b) Temporal Dynamic Aggregation (TDA) merges frame-level tokens into a temporal token using a weighted fusion based on cosine similarity. (§3.2). (c) Token-driven Keyframe Selection (TKS). During training, the frame with the token closest to the token is selected as the keyframe. During inference, keyframe selection is refined using SAM2's occlusion scores and token similarity scores (§3.3). (d) Mask Decoding and Propagation (MDP). The token provides a sparse embedding for SAM2, generating a keyframe mask and propagating it to other frames via a memory mechanism (§3.4)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 216, + 174, + 399, + 277 + ], + "blocks": [ + { + "bbox": [ + 216, + 174, + 399, + 277 + ], + "lines": [ + { + "bbox": [ + 216, + 174, + 399, + 277 + ], + "spans": [ + { + "bbox": [ + 216, + 174, + 399, + 277 + ], + "type": "image", + "image_path": "1345cde1f7187f6e0b405e800a596e4fd69ff976dd5c90c0d90156dc512da619.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 400, + 174, + 539, + 277 + ], + "blocks": [ + { + "bbox": [ + 400, + 174, + 539, + 277 + ], + "lines": [ + { + "bbox": [ + 400, + 174, + 539, + 277 + ], + "spans": [ + { + "bbox": [ + 400, + 174, + 539, + 277 + ], + "type": "image", + "image_path": "dea18d298f5a29e51fe45d5fb41d2de62e27f5d3b92d453940d84d65407a084b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 366, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 296, + 521 + ], + "type": "text", + "content": "modal and memory attention to jointly address RVOS and semi-supervised video object segmentation. Hui et.al. [13] dynamically recombine linguistic features and interact with visual features using cross-modal attention to highlight spatiotemporal regions of interest. Inspired by DETR [4], methods like MTTR [2] and ReferFormer [37] leverage language queries for precise target localization. Emerging approaches are exploring other-modal guidance for referring segmentation, such as audio cues [8, 9, 47]. While RVOS excels with explicit object descriptions, it often struggles with the complex reasoning required to interpret more implicit and intricate language instructions, motivating our exploration of Video Reasoning Segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 194, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 194, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 194, + 541 + ], + "type": "text", + "content": "2.2. Reasoning Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Reasoning segmentation [17, 27, 30, 38] advances referring segmentation by generating masks from complex images and versatile text prompts. LISA [17] pioneers the field of reasoning segmentation. It introduces a new token to expand the vocabulary and proposes the embedding-as-mask paradigm, enhancing segmentation capabilities to address scenarios requiring complex reasoning and world knowledge. LISA++ [41] addresses LISA's limitations in distinguishing individual instances by enhancing instruction-tuning data with key segmentation datasets, supporting both semantic and instance segmentation tasks. PixelLM [30] innovatively combines a novel pixel decoder and a segmentation codebook with learnable tokens to efficiently generate high-quality masks without external models. VISA [39]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 366, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 475 + ], + "type": "text", + "content": "introduces reasoning segmentation to the video domain, proposing reasoning video object segmentation. It uses a pre-trained Multimodal Large Language Model (MLLM) to select keyframes, segments them based on reasoning, and propagates the masks to other frames using a pre-trained object tracker. However, it suffers from limited representation capability of a single special token, inaccurate keyframe selection, and the inability to perform end-to-end inference, compromising its segmentation and tracking performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 397, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 397, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 397, + 498 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "content": "Task Definition. The task of Video Reasoning Segmentation can be briefly outlined as follows. Given a video clip consisting of " + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_V\\in \\mathbb{R}^{T\\times 3\\times H\\times W}" + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "content": " denote the height and width of each frame respectively, along with a high-level textual instruction " + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{txt}" + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "content": ", VRS aims to design a model " + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "content": " to interpret and transform " + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{txt}" + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "content": " into the binary segmentation mask sequence " + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_M\\in \\mathbb{R}^{T\\times H\\times W}" + }, + { + "bbox": [ + 313, + 502, + 555, + 670 + ], + "type": "text", + "content": " for each frame. In contrast to RVOS tasks providing explicit descriptions like \"the person skateboarding,\" VRS typically employs expressions incorporating world knowledge like \"tool(s) for holding garbage\" or temporal logic like \"the ship(s) moving at the highest speed\", raising higher demands on the model's capabilities of temporal relationship comprehension and complex scenario reasoning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Overall Architecture. Fig. 2 illustrates the VRS-HQ architecture, which comprises Chat-UniVi [14] as the Multi-modal Large Language Model (MLLM) for temporal token" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29185" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "encoding (§3.1), a Temporal Dynamic Aggregation (§3.2), a Token-driven Keyframe Selection (§3.3), and Mask Decoding and Propagation (§3.4). The MLLM encodes the textual prompt into multi-level tokens representing spatial and semantic target information. These tokens drive both temporal dynamic aggregation and keyframe selection via cosine similarity scores. SAM2 then segments the keyframes and tracks the target object throughout the video. During inference, the Token-driven Keyframe Selection computes object occlusion scores for each sampled frame by merging it with the fused temporal token. Filtering keyframes based on these scores improves the accurate localization of the target object and subsequently segmentation process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 238, + 203, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 203, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 203, + 251 + ], + "type": "text", + "content": "3.1. Temporal Token Encoding" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 296, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 364 + ], + "type": "text", + "content": "Current VRS approaches [1, 39] typically use specified language prompts to guide MLLMs in embedding target information from keyframes or entire videos into a single specialized token, struggling to capture rich spatiotemporal dynamics essential for fine-grained video understanding. To equip the MLLM with both frame-level and video-level contextual awareness for video segmentation, we propose encoding intra-frame spatial information and inter-frame temporal relations into hierarchical tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "content": "Hierarchical Token Generation. Instead of previous single-token encoding strategy, the vocabulary of the MLLM is initially augmented with two new special tokens: " + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "inline_equation", + "content": "<\\mathrm{TAK}>" + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "content": ". Then, we design a structured conversational template as \"USER: Please find {expression} in the Reference Video and segment it in each frame and the entire video respectively.\" Here, \"" + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{expression}\\}" + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "content": "\" denotes the target object description. The tokenized prompts " + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\text{txt}}" + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "content": " and sampled video frames " + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\text{V}'}" + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "content": " are input into the MLLM, which generates a response " + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "inline_equation", + "content": "y_{\\text{txt}}" + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "content": " containing multiple frame-level " + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "content": " tokens and a temporal-level " + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "inline_equation", + "content": "<\\mathrm{TAK}>" + }, + { + "bbox": [ + 55, + 371, + 296, + 529 + ], + "type": "text", + "content": " token through autoregressive encoding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "text", + "content": "Token Extraction and Mapping. Subsequently, the " + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "text", + "content": " token embeddings " + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\bar{h}_{seg} \\in \\mathbb{R}^{T' \\times d'}" + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "inline_equation", + "content": "<\\mathrm{TAK}>" + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "text", + "content": " token embedding " + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\bar{h}_{tak} \\in \\mathbb{R}^{1 \\times d'}" + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "text", + "content": " are extracted from the MLLM's final layer, where " + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "text", + "content": " denotes the MLLM's embedding dimension and " + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 55, + 534, + 296, + 620 + ], + "type": "text", + "content": " denotes the length of sampled frames. These embeddings are then projected into the same feature space as SAM2 using a multi-layer perceptron:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 628, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 628, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 93, + 628, + 295, + 643 + ], + "type": "interline_equation", + "content": "h _ {s e g}, h _ {t a k} = M L P \\left(\\bar {h} _ {s e g}\\right), M L P \\left(\\bar {h} _ {t a k}\\right) \\tag {1}", + "image_path": "1f7ec2d97f512da9b4aebcaaaaa98f055d6e0d105b68ce7d2655e3da8da618dc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_{seg} \\in \\mathbb{R}^{T' \\times d}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_{tak} \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " represent the sparse embeddings for segmentation mask activation, and " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " denotes the feature dimension of SAM2. Finally, the frame-level embeddings " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_{seg}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " are aggregated with the video-level embedding " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_{tak}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " to incorporate temporal dynamics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 488, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 488, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 488, + 85 + ], + "type": "text", + "content": "3.2. Temporal Dynamic Aggregation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": "Building on the strong temporal encoding of MLLM, and embeddings encapsulate the spatial priors and temporal semantic signals of the targets respectively, providing rich contextual information for the segmentation model. Accordingly, we propose the Temporal Dynamic Aggregation to facilitate the fusion of positional and semantic information of the targets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "content": "Keyframes as Token Similarity. The cosine similarity between each frame-level " + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "inline_equation", + "content": "<\\text{SEG}>" + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "content": " token and the temporal-level " + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "inline_equation", + "content": "<\\text{TAK}>" + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "content": " token reflects the semantic alignment between individual frames and the overall video context. We hypothesize that high-similarity frames are more representative of the video's content and thus suitable as keyframes. This motivates our use of token similarity for keyframe selection during training, which also facilitates weighted fusion of " + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "inline_equation", + "content": "<\\text{SEG}>" + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "inline_equation", + "content": "<\\text{TAK}>" + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "content": " tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "text", + "content": "Similarity-based Weighted Fusion. To enhance the inter-frame consistency while achieving more precise discernment of the object's positional shifts across frames, we fuse the frame-level embeddings " + }, + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "inline_equation", + "content": "h_{seg}" + }, + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "text", + "content": " with the video-level embedding " + }, + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "inline_equation", + "content": "h_{tak}" + }, + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "text", + "content": " using normalized cosine similarity scores as attention weights (cf. Fig. 2 (b)). During backpropagation, frame-level features " + }, + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "inline_equation", + "content": "h_{seg}" + }, + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "text", + "content": " are updated alongside the temporal-level embedding " + }, + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "inline_equation", + "content": "h_{tak}" + }, + { + "bbox": [ + 313, + 288, + 556, + 397 + ], + "type": "text", + "content": " to enable more comprehensive learning of local and global video context:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 369, + 406, + 555, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 406, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 369, + 406, + 555, + 439 + ], + "type": "interline_equation", + "content": "h _ {t a k} ^ {\\prime} = h _ {t a k} + \\alpha \\sum_ {i = 1} ^ {T ^ {\\prime}} \\lambda_ {i} h _ {s e g} [ i ], \\tag {2}", + "image_path": "6e1a886c6606aaad715e336bcb52a8053682ac1e5024ffa391dbd18751b5ec82.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "text", + "content": " represents the normalized cosine similarity between " + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "inline_equation", + "content": "h_{seg}[i]" + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "inline_equation", + "content": "h_{tak}" + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "text", + "content": " is the fusion coefficient. The resulting fused embedding " + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "inline_equation", + "content": "h_{tak}^{\\prime}" + }, + { + "bbox": [ + 313, + 447, + 555, + 496 + ], + "type": "text", + "content": " is then used for keyframe selection during inference and mask generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 494, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 494, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 494, + 515 + ], + "type": "text", + "content": "3.3. Token-driven Keyframe Selection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 520, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 628 + ], + "type": "text", + "content": "VISA [39] relies on an external model (LLaMA-VID [19]) for keyframe selection during inference, hindering end-to-end processing and potentially degrading performance due to inaccuracies in the external model's output. To address this limitation, we introduce the Token-driven Keyframe Selection that leverages the temporal information encoded within the integrated embedding. This approach eliminates the need for complex prompt engineering and significantly improves the reliability of keyframe selection." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": "Instead of uniform video sampling during training, we adopt the CLIP model [25] to find the frame most aligned with the expression " + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{exp}" + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": " for inference. The anchor frame is used for global sampling, resulting in sampled frames " + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_V^f\\in \\mathbb{R}^{\\hat{T}\\times 3\\times H\\times W}" + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 313, + 628, + 556, + 715 + ], + "type": "text", + "content": " represents the number of the sampled frames. Following the temporal dynamic aggregation (Eq. 2), each sampled frame is treated as a potential" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29186" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 110 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 110 + ], + "type": "text", + "content": "keyframe, along with the fused " + }, + { + "bbox": [ + 55, + 72, + 296, + 110 + ], + "type": "inline_equation", + "content": "<\\mathrm{TAK}>" + }, + { + "bbox": [ + 55, + 72, + 296, + 110 + ], + "type": "text", + "content": " embedding " + }, + { + "bbox": [ + 55, + 72, + 296, + 110 + ], + "type": "inline_equation", + "content": "h_{\\text{tak}}'" + }, + { + "bbox": [ + 55, + 72, + 296, + 110 + ], + "type": "text", + "content": ", is sent to SAM2 to generate the object occlusion scores " + }, + { + "bbox": [ + 55, + 72, + 296, + 110 + ], + "type": "inline_equation", + "content": "S_{o} \\in \\mathbb{R}^{\\hat{T} \\times 1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 110 + ], + "type": "text", + "content": " (cf. Fig. 2 (c)):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 114, + 295, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 114, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 123, + 114, + 295, + 131 + ], + "type": "interline_equation", + "content": "S _ {o} = \\mathcal {M D} (\\mathcal {E} (\\mathcal {X} _ {V} ^ {f}), h _ {t a k} ^ {\\prime}) \\tag {3}", + "image_path": "0bd31d58a36ba3bfffaf35678da28b82981970a6d61afbbd2994097335943bc7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 135, + 296, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 296, + 232 + ], + "type": "text", + "content": "The occlusion scores effectively indicate the confidence of the target's presence in the current frame, assisting in identifying which frame best corresponds to the information encoded in the token. After determining the keyframe index by the combination of softmax-normalized scores " + }, + { + "bbox": [ + 55, + 135, + 296, + 232 + ], + "type": "inline_equation", + "content": "S_{o}^{\\prime}" + }, + { + "bbox": [ + 55, + 135, + 296, + 232 + ], + "type": "text", + "content": " with token similarity scores " + }, + { + "bbox": [ + 55, + 135, + 296, + 232 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 135, + 296, + 232 + ], + "type": "text", + "content": ", we decode the keyframe mask and then utilize the cross-frame propagation technique of SAM2 to obtain video-level segmentation masks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 238, + 233, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 233, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 233, + 251 + ], + "type": "text", + "content": "3.4. Mask Decoding and Propagation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 255, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 339 + ], + "type": "text", + "content": "Following temporal token fusion and keyframe selection, the mask embeddings enriched with positional and semantic information can be yielded by the segmentation model. In contrast to prior methods [1, 39, 43] depending on image-level segmentation models or external object trackers for target trajectories prediction, we utilize SAM2 to perform segmentation and propagation concurrently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "text", + "content": "SAM2 for Masklets Decoding. Given the keyframe " + }, + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_k" + }, + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "text", + "content": ", we first extract its features using the image encoder " + }, + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "text", + "content": ", providing the conditional input for SAM2. The integrated temporal embedding " + }, + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "inline_equation", + "content": "h_{\\text{tak}}'" + }, + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "text", + "content": " then interacts with these keyframe features within the mask decoder " + }, + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{MD}" + }, + { + "bbox": [ + 55, + 343, + 296, + 416 + ], + "type": "text", + "content": " (cf. Fig. 2 (d)) to generate the keyframe mask:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 422, + 295, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 422, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 119, + 422, + 295, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {X} _ {M} ^ {k} = \\mathcal {M D} \\left(\\mathcal {E} \\left(\\mathcal {X} _ {k}\\right), h _ {t a k} ^ {\\prime}\\right), \\tag {4}", + "image_path": "c6a50ea402cfa31f3930d1edc0ae99e038ba7a8fa36651064ed78eb35471e893.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_M^k" + }, + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "text", + "content": " represents the predicted mask for the keyframe. Next, we propagate this mask to two adjacent frames, " + }, + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{k - 1}" + }, + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{k + 1}" + }, + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "text", + "content": ", treated as non-conditional frames, using the memory storage and interaction mechanism. This yields the mask sequence " + }, + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_M^s" + }, + { + "bbox": [ + 55, + 441, + 296, + 526 + ], + "type": "text", + "content": ". During inference, all remaining frames are processed as non-conditional frames, facilitating mask propagation throughout the entire video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 532, + 171, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 171, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 171, + 544 + ], + "type": "text", + "content": "3.5. Training Objectives" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "type": "text", + "content": "Our method is trained end-to-end using a combined text generation loss and mask loss to optimize the " + }, + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{TAK} \\rangle" + }, + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{SEG} \\rangle" + }, + { + "bbox": [ + 55, + 549, + 296, + 597 + ], + "type": "text", + "content": " embeddings. The mask loss combines binary cross-entropy (BCE) and DICE loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 604, + 295, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 604, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 82, + 604, + 295, + 617 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = \\lambda_ {t x t} L _ {t x t} \\left(y _ {t x t}, \\hat {y} _ {t x t}\\right) + \\lambda_ {m a s k} L _ {m a s k}, \\tag {5}", + "image_path": "591041892f399b0b1c2fa132582636d4ccc89fdf508c1cb1f8703501b0bdedd2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 639, + 296, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 639, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 63, + 639, + 296, + 666 + ], + "type": "interline_equation", + "content": "L _ {m a s k} = \\lambda_ {b c e} L _ {b c e} \\left(\\mathcal {X} _ {M} ^ {s}, \\hat {\\mathcal {X}} _ {M} ^ {s}\\right) + \\lambda_ {d i c e} L _ {d i c e} \\left(\\mathcal {X} _ {M} ^ {s}, \\hat {\\mathcal {X}} _ {M} ^ {s}\\right), \\tag {6}", + "image_path": "63f85ecb916cc59c941c6a6341b35ec184cb90a97c3495f1ed156e1808dc046b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_{txt}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\chi}_M^s" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " represent the ground truth text and mask, respectively, and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_{txt}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "X_M^s" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " are their corresponding predictions. The weighting coefficients " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{mask}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{bce}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " are set to 1, 1, 2 and 0.5, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 91, + 435, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 435, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 435, + 102 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "text", + "content": "Datasets. Our model is trained and evaluated on extensive image and video segmentation datasets, LLaVA-Instruct-150k [21], as well as the Video Question-Answering datasets from Video-ChatGPT [22]. Specifically, the image segmentation datasets comprise semantic segmentation: ADE20K [45], COCO-Stuff [3], PACO-LVIS [26], and PASCALPart [5], referring segmentation: refCLEF, refCOCO, refCOCO+ [15], and refCOCOg [23], and reasoning segmentation: ReasonSeg [17]. While the video segmentation datasets encompass the RVOS datasets: RefYoutube-VOS [32], Ref-DAVIS17 [24] and MeViS [7], and the VRS benchmark ReVOS [39]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate video segmentation using region similarity " + }, + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "text", + "content": ", contour accuracy " + }, + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "text", + "content": ", and their mean " + }, + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "inline_equation", + "content": "(\\mathcal{J} \\& \\mathcal{F})" + }, + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "text", + "content": ". Image segmentation accuracy is measured via Generalized Intersection over Union (gIoU) [31] and Complete Intersection over Union (cIoU) [44]. Model hallucination is assessed using the robustness score " + }, + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 258, + 555, + 331 + ], + "type": "text", + "content": " [39]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 338, + 447, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 447, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 447, + 350 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 355, + 555, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 500 + ], + "type": "text", + "content": "We fine-tuned Chat-UniVi [14] (our chosen MLLM) using LoRA [12] (rank 8), optimizing the mask decoder and MLP projection layer while freezing other parameters. Training used AdamW (learning rate 0.0003, no weight decay) with a WarmupDecayLR scheduler (100-iteration warmup). We set the fusion coefficient " + }, + { + "bbox": [ + 313, + 355, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 355, + 555, + 500 + ], + "type": "text", + "content": " to 0.1 and sampled two non-keyframes per video. Trained for 7500 iterations on a hybrid image/video dataset using four A800 GPUs with DeepSpeed [28] (batch size 1, gradient accumulation 32, total batch size 128), the model employed TDA for video data (token fusion, segmentation, and propagation) and the token for direct segmentation on image data." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 508, + 432, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 508, + 432, + 521 + ], + "spans": [ + { + "bbox": [ + 314, + 508, + 432, + 521 + ], + "type": "text", + "content": "4.3. Comparison Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 525, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 574 + ], + "type": "text", + "content": "To showcase the robust pixel-level perception and generalization of VRS-HQ, we conduct evaluations across diverse benchmarks, including ReVOS, RVOS, and image-based referring and reasoning segmentation datasets." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "ReVOS Datasets. Tab. 1 illustrates the performance comparison with previous methods [2, 17, 37, 39, 46] on the ReVOS benchmark. VRS-HQ demonstrates significant improvements over the previous state-of-the-art, VISA [39] across all ten metrics. Remarkably, in terms of the " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ", VRS-HQ-13B surpasses VISA-13B by " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " on the referring subset and " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "12.5\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " on the reasoning subset, respectively. These gains highlight the effectiveness of our temporal aggregation strategy and the utilization of the token for keyframe selection and target segmentation, which enhances the model's reasoning capabilities. In contrast," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29187" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 77, + 525, + 234 + ], + "blocks": [ + { + "bbox": [ + 165, + 65, + 444, + 76 + ], + "lines": [ + { + "bbox": [ + 165, + 65, + 444, + 76 + ], + "spans": [ + { + "bbox": [ + 165, + 65, + 444, + 76 + ], + "type": "text", + "content": "Table 1. Performance comparison with previous methods on ReVOS dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 77, + 525, + 234 + ], + "lines": [ + { + "bbox": [ + 88, + 77, + 525, + 234 + ], + "spans": [ + { + "bbox": [ + 88, + 77, + 525, + 234 + ], + "type": "table", + "html": "
MethodsBackbonereferringreasoningoverallR
JFJ&FJFJ&FJFJ&F
MTTR [2][CVPR2022]Video-Swin-T29.830.230.020.421.521.025.125.925.55.6
LMPM [7][ICCV2023]Swin-T29.039.134.113.324.318.821.231.726.43.2
ReferFormer [37][CVPR2022]Video-Swin-B31.234.332.721.325.623.426.229.928.18.8
LISA [17][CVPR2024]LLaVA-7B44.347.145.733.838.436.139.142.740.99.3
LISA [17][CVPR2024]LLaVA-13B45.247.946.634.339.136.739.843.541.68.6
TrackGPT [46][arXiv2023]LLaVA-7B46.749.748.236.841.239.041.845.543.611.6
TrackGPT [46][arXiv2023]LLaVA-13B48.350.649.538.142.940.543.246.845.012.8
VISA [39][ECCV2024]LLaVA-7B49.452.651.040.545.843.244.949.247.115.3
VISA [39][ECCV2024]LLaVA-13B55.759.057.441.946.544.248.852.850.815.1
VISA [39][ECCV2024]Chat-UniVi-7B49.252.650.940.645.443.044.949.046.915.5
VISA [39][ECCV2024]Chat-UniVi-13B55.659.157.442.046.744.348.852.950.914.5
VRS-HQ[Ours]Chat-UniVi-7B59.864.562.153.558.756.156.661.659.119.7
VRS-HQ[Ours]Chat-UniVi-13B61.165.563.354.159.456.857.662.560.018.9
", + "image_path": "33bea7b50e26a4a8147f01165a30cda4806249c0bf563519d7336221efb9ca2a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 97, + 254, + 513, + 420 + ], + "blocks": [ + { + "bbox": [ + 127, + 241, + 481, + 252 + ], + "lines": [ + { + "bbox": [ + 127, + 241, + 481, + 252 + ], + "spans": [ + { + "bbox": [ + 127, + 241, + 481, + 252 + ], + "type": "text", + "content": "Table 2. Performance comparison with previous methods on the validation sets of RVOS datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 254, + 513, + 420 + ], + "lines": [ + { + "bbox": [ + 97, + 254, + 513, + 420 + ], + "spans": [ + { + "bbox": [ + 97, + 254, + 513, + 420 + ], + "type": "table", + "html": "
MethodsBackboneRef-YouTube-VOSRef-DAVIS17MeViS
JFJ&FJFJ&FJFJ&F
MTTR [2][CVPR2022]Video-Swin-T54.056.655.3---28.831.230.0
LMPM [7][ICCV2023]Swin-T------34.240.237.2
ReferFormer [37][CVPR2022]Video-Swin-B61.364.662.958.164.161.129.832.231.0
OnlineRefer [36][CVPR2023]Swin-L61.665.563.561.667.764.8---
DsHmp [11][CVPR2024]Video-Swin-B65.069.167.161.768.164.943.049.846.4
LISA [17][CVPR2024]LLaVA-7B53.454.353.962.267.364.835.139.437.2
LISA [17][CVPR2024]LLaVA-13B54.054.854.463.268.866.035.840.037.9
TrackGPT [46][aXiv2023]LLaVA-7B55.357.456.459.467.063.237.642.640.1
TrackGPT [46][aXiv2023]LLaVA-13B58.160.859.562.770.466.539.243.141.2
VISA [39][ECCV2024]Chat-UniVi-7B59.863.261.566.372.569.440.746.343.5
VISA [39][ECCV2024]Chat-UniVi-13B61.464.763.067.073.870.441.847.144.5
VideoLISA [1][NeurIPS2024]LLaVA-Phi-3-V61.765.763.764.972.768.841.347.644.4
VRS-HQ[Ours]Chat-UniVi-7B68.372.570.472.679.476.047.653.750.6
VRS-HQ[Ours]Chat-UniVi-13B69.073.171.071.077.974.448.053.750.9
", + "image_path": "aec7bc6bbe210b7c6774a90051410ff9e34ee17bbbc79b535b995cf942dfce8f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": "VISA relies on LLaMA-VID [19] for keyframe selection during inference. This strategy may neglect intra-frame fine-grained visual details, leading to inaccurate keyframe localization. Furthermore, the robustness score " + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": " exceeds VISA-13B by " + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": ", indicating VRS-HQ's superior capability in handling negative samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": "RVOS. Tab. 2 compares VRS-HQ with state-of-the-art RVOS methods. On Ref-YouTube-VOS and RefDAVIS17, VRS-HQ-13B surpasses VideoLISA [1] and VISA-13B [39], achieving " + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{I}\\&\\mathcal{F}" + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": " improvements of " + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": ", respectively. Furthermore, on the motion-intensive MeViS dataset, VRS-HQ-13B yields substantial gains of " + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{I}\\&\\mathcal{F}" + }, + { + "bbox": [ + 55, + 508, + 295, + 652 + ], + "type": "text", + "content": ", respectively. demonstrating its effective in capturing motion and maintaining temporal coherence. These improvements, compared to VideoLISA's single-token representation, are attributed to the enhanced inter-frame perception of the TDA and the refined keyframe localization of the TKS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "RIS. Our method seamlessly extends to referring image segmentation (RIS) by treating images as single-frame videos. As shown in Tab. 3, VRS-HQ consistently outperforms VISA on three RIS benchmarks, achieving performance comparable to LISA [17] and VideoLISA [1]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 434, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 555 + ], + "type": "text", + "content": "On ReasonSeg, VRS-HQ performs competitively, ranking merely below VideoLISA, demonstrating strong generalization. The marginally lower performance on image datasets compared to video datasets likely stems from two factors: (i) our model's training emphasis on video data (unlike LISA, which is pre-trained solely on images); and (ii) our training methodology encourages reliance on SAM2's memory mechanism for multi-frame processing, which may be less effective for single-image segmentation compared to VideoLISA's single-token approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 563, + 416, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 416, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 416, + 575 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Fusion Coefficient Ablation. The ablation analysis of the fusion coefficient " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " in the TDA is illustrated in Tab. 5. An " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " of 0.1 yields optimal performance. Setting " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " to 0 leads to a sharp drop in the model's performance as " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "<\\mathrm{TAK}>" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " fails to capture fine-grained intra-frame details, and the " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " token cannot be jointly optimized with " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "<\\mathrm{TAK}>" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " through backpropagation (e.g., " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "62.1\\% \\rightarrow 58.8\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " on the referring subset, " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "56.1\\% \\rightarrow 54.1\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " on the reasoning subset). As " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " increases beyond 0.1, the metrics decline slightly, likely due to excessive frame-level noise being introduced into the temporal token at a higher fusion coef-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29188" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 77, + 535, + 194 + ], + "blocks": [ + { + "bbox": [ + 97, + 65, + 511, + 76 + ], + "lines": [ + { + "bbox": [ + 97, + 65, + 511, + 76 + ], + "spans": [ + { + "bbox": [ + 97, + 65, + 511, + 76 + ], + "type": "text", + "content": "Table 3. Performance comparison with previous methods on referring and reasoning image segmentation datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 77, + 535, + 194 + ], + "lines": [ + { + "bbox": [ + 74, + 77, + 535, + 194 + ], + "spans": [ + { + "bbox": [ + 74, + 77, + 535, + 194 + ], + "type": "table", + "html": "
MethodBackbonerefCOCOrefCOCO+refCOCOgReaSeg (val)ReaSeg (test)
valtestAtestBvaltestAtestBvaltestgIoUcIoUgIoUcIoU
CRIS [34]ResNet10170.573.266.162.368.153.759.960.4----
LAVT [42]Swin-B72.775.868.862.168.455.161.262.1----
ReLA [20]Swin-B73.876.570.266.071.057.765.066.0----
X-Decoder [48]DaViT-L------64.6-22.617.921.716.3
SEEM [49]DaViT-L------65.7-25.521.224.318.7
LISA [17]LLaVA-7B74.979.172.365.170.858.167.970.652.954.047.348.4
VISA [39]Chat-UniVi-7B72.475.568.159.864.853.165.566.452.757.8--
VideoLISA [1]LLaVA-Phi-3-V73.876.668.863.468.856.268.368.861.467.153.854.4
VRS-HQChat-UniVi-7B73.577.569.561.767.654.366.767.555.251.851.752.9
", + "image_path": "ba03d3528a250227ae4cbc16bdb51ea04dea1a15253eb44063dbc53f8dcb3a75.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 64, + 234, + 216, + 278 + ], + "blocks": [ + { + "bbox": [ + 55, + 211, + 220, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 220, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 220, + 232 + ], + "type": "text", + "content": "Table 4. Ablation analysis of TDA and TKS components." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 234, + 216, + 278 + ], + "lines": [ + { + "bbox": [ + 64, + 234, + 216, + 278 + ], + "spans": [ + { + "bbox": [ + 64, + 234, + 216, + 278 + ], + "type": "table", + "html": "
Componentsreferringreasoning
JFJ&FJFJ&F
w/o TDA + TKS54.559.156.848.653.951.2
w/o TDA56.561.158.851.556.754.1
w/o TKS57.862.460.150.956.253.5
VRS-HQ59.864.562.153.558.756.1
", + "image_path": "8117000025a9b4bd8b68eac65ed980c88c8ec512c1082036af06a3be4c5b35a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 70, + 324, + 208, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 296, + 222, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 222, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 222, + 319 + ], + "type": "text", + "content": "Table 7. Ablation analysis of sampled frames during inference stage." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 324, + 208, + 381 + ], + "lines": [ + { + "bbox": [ + 70, + 324, + 208, + 381 + ], + "spans": [ + { + "bbox": [ + 70, + 324, + 208, + 381 + ], + "type": "table", + "html": "
Framesreferringreasoning
JFJ&FJFJ&F
158.462.760.649.855.152.5
358.663.060.851.056.253.6
658.663.360.951.557.054.3
958.963.761.352.557.855.1
1259.364.061.752.557.855.1
Non-fixed59.864.562.153.558.756.1
", + "image_path": "819bc6910ba7d758da02cae0e0e0e567f0b5ea9a0063177858a5211111008c77.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 242, + 230, + 367, + 274 + ], + "blocks": [ + { + "bbox": [ + 221, + 208, + 387, + 228 + ], + "lines": [ + { + "bbox": [ + 221, + 208, + 387, + 228 + ], + "spans": [ + { + "bbox": [ + 221, + 208, + 387, + 228 + ], + "type": "text", + "content": "Table 5. Ablation analysis of the fusion coefficient " + }, + { + "bbox": [ + 221, + 208, + 387, + 228 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 221, + 208, + 387, + 228 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 242, + 230, + 367, + 274 + ], + "lines": [ + { + "bbox": [ + 242, + 230, + 367, + 274 + ], + "spans": [ + { + "bbox": [ + 242, + 230, + 367, + 274 + ], + "type": "table", + "html": "
αreferringreasoning
JFJ&FJFJ&F
056.561.158.851.556.754.1
0.159.864.562.153.558.756.1
0.2558.963.861.352.357.855.0
0.558.263.060.651.456.654.0
", + "image_path": "63c0199847882406ba8e49b2046bffbc2153414dcba90036d5c3c7ecbb4c2f4a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 229, + 323, + 380, + 373 + ], + "blocks": [ + { + "bbox": [ + 221, + 279, + 387, + 322 + ], + "lines": [ + { + "bbox": [ + 221, + 279, + 387, + 322 + ], + "spans": [ + { + "bbox": [ + 221, + 279, + 387, + 322 + ], + "type": "text", + "content": "Table 8. Ablation analysis of the mask decoding and propagation strategy. MT/MI: Multiframe Training/Inference. ST/SI: Single-frame Training/Inference." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 229, + 323, + 380, + 373 + ], + "lines": [ + { + "bbox": [ + 229, + 323, + 380, + 373 + ], + "spans": [ + { + "bbox": [ + 229, + 323, + 380, + 373 + ], + "type": "table", + "html": "
Strategyreferringreasoning
JFJ&FJFJ&F
SAM+ST+SI48.352.450.342.947.445.2
SAM2+ST+SI55.959.857.849.354.051.6
SAM2+ST+MI55.860.858.346.552.549.5
SAM2+MT+SI54.157.455.848.252.250.2
SAM2+MT+MI59.864.562.153.558.756.1
", + "image_path": "4091b59e233b287ba50ef14b2c2038fa5c27e30c8321f263055b0ec27e7be8df.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 399, + 245, + 544, + 289 + ], + "blocks": [ + { + "bbox": [ + 387, + 201, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 387, + 201, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 387, + 201, + 553, + 244 + ], + "type": "text", + "content": "Table 6. Ablation analysis of Token-driven Keyframe Selection. " + }, + { + "bbox": [ + 387, + 201, + 553, + 244 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 387, + 201, + 553, + 244 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 387, + 201, + 553, + 244 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 387, + 201, + 553, + 244 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 387, + 201, + 553, + 244 + ], + "type": "inline_equation", + "content": "S_{3}" + }, + { + "bbox": [ + 387, + 201, + 553, + 244 + ], + "type": "text", + "content": " represent CLIP scores, token similarity scores, and occlusion scores, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 399, + 245, + 544, + 289 + ], + "lines": [ + { + "bbox": [ + 399, + 245, + 544, + 289 + ], + "spans": [ + { + "bbox": [ + 399, + 245, + 544, + 289 + ], + "type": "table", + "html": "
S1S2S3referringreasoning
JFJ&FJFJ&F
✓✓✓59.764.261.952.557.855.2
✓✓57.862.460.150.956.253.6
✓✓59.664.161.852.557.755.1
✓✓59.864.562.153.558.756.1
", + "image_path": "5919e87dac187c9759209b37e0d0f3e72ae3fe7d9d4f6daa00c1abc7d7e0c22c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 395, + 316, + 547, + 353 + ], + "blocks": [ + { + "bbox": [ + 387, + 293, + 553, + 315 + ], + "lines": [ + { + "bbox": [ + 387, + 293, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 387, + 293, + 553, + 315 + ], + "type": "text", + "content": "Table 9. Ablation analysis of sampling strategy." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 395, + 316, + 547, + 353 + ], + "lines": [ + { + "bbox": [ + 395, + 316, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 395, + 316, + 547, + 353 + ], + "type": "table", + "html": "
Sampling strategyreferringreasoning
JFJ&FJFJ&F
Random Sampling59.063.661.353.258.455.8
Uniform Sampling59.363.961.653.358.455.8
CLIP Sampling59.864.562.153.558.756.1
", + "image_path": "581aa035cf86f28c3db8cf4b7762e9ef5ff3bcf47e1adbdaeeeb1f195c8148de.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 406, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 295, + 431 + ], + "type": "text", + "content": "ficient (e.g., " + }, + { + "bbox": [ + 55, + 406, + 295, + 431 + ], + "type": "inline_equation", + "content": "62.1\\% \\rightarrow 60.6\\%" + }, + { + "bbox": [ + 55, + 406, + 295, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 55, + 406, + 295, + 431 + ], + "type": "text", + "content": " on the referring subset, " + }, + { + "bbox": [ + 55, + 406, + 295, + 431 + ], + "type": "inline_equation", + "content": "56.1\\% \\rightarrow 54.0\\%" + }, + { + "bbox": [ + 55, + 406, + 295, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 55, + 406, + 295, + 431 + ], + "type": "text", + "content": " on the reasoning subset)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "text", + "content": "Token-driven Keyframe Selection. Tab. 6 analyzes the impact of different score combinations for keyframe selection within the TKS. Results indicate that the occlusion score is the most influential of the three considered, improving " + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "text", + "content": " on the referring and reasoning subsets, respectively (first two rows). This improvement stems from the occlusion score's ability to reflect the target object's presence within each frame. The model achieves its best performance when keyframe selection is based on a combination of token similarity and the occlusion score. However, incorporating the CLIP score (rows 1 and 4) slightly degrades performance (by " + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 54, + 440, + 295, + 620 + ], + "type": "text", + "content": " on the referring and reasoning subsets, respectively), likely due to CLIP's limited spatiotemporal understanding, which hinders accurate keyframe selection." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "type": "text", + "content": "Mask Decoding and Propagation. Tab. 8 analyzes the impact of different mask decoding and propagation strategies on segmentation performance. Using only SAM and TDA for single-frame segmentation (first row) already surpasses VISA-13B by " + }, + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.9\\% / 0.7\\% / 0.9\\%" + }, + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "type": "text", + "content": " on the reasoning subset. Comparing the first two rows reveals the substantial advantage of SAM2 over SAM for single-frame fine" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "text", + "content": "tuning and inference. SAM2 achieves improvements of " + }, + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "inline_equation", + "content": "7.6\\% / 7.4\\% / 7.5\\% / 6.4\\% / 6.6\\% / 6.4\\%" + }, + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "text", + "content": ", demonstrating its superior robustness in segmentation and tracking. Furthermore, incorporating multi-frame propagation during both training and inference (last row) yields the best overall results, effectively leveraging SAM2's memory mechanism for keyframe mask storage. This approach leads to further improvements in " + }, + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 313, + 406, + 555, + 502 + ], + "type": "text", + "content": " compared to the other strategies." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 631 + ], + "type": "text", + "content": "Inference-Time Sampling Strategy. Tab. 9 analyzes the impact of different frame sampling strategies during inference. \"Random\" denotes randomly selected frames, while \"Uniform\" extracts frames at equal intervals. CLIP sampling slightly outperforms uniform sampling, improving " + }, + { + "bbox": [ + 313, + 510, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 510, + 555, + 631 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 510, + 555, + 631 + ], + "type": "inline_equation", + "content": "0.5\\% / 0.3\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 631 + ], + "type": "text", + "content": " on the referring and reasoning subsets, respectively. This modest gain is likely due to CLIP's ability in selecting frames that are semantically aligned with the referring expression, which in turn facilitates the reasoning process of the MLLM." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 635, + 441, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 441, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 441, + 649 + ], + "type": "text", + "content": "4.5. Visualization Analysis" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Segmentation Map Comparison. Fig. 3 presents the qualitative comparison of VRS-HQ and VISA on the ReVOS dataset across three different scenarios. In the left case, VISA displays ambiguity in spatial positioning and neglects critical references such as \"the first\", revealing" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29189" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 67, + 539, + 259 + ], + "blocks": [ + { + "bbox": [ + 66, + 67, + 539, + 259 + ], + "lines": [ + { + "bbox": [ + 66, + 67, + 539, + 259 + ], + "spans": [ + { + "bbox": [ + 66, + 67, + 539, + 259 + ], + "type": "image", + "image_path": "132ba830efe9971c912ef8b6c19dc8cca0c4b4434a8ddf345229bcfab2da322a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. Segmentation map comparison of VISA and VRS-HQ on the ReVOS benchmark (§4.5). Results across three scenarios demonstrate that VRS-HQ excels in reasoning complex spatial and temporal relationships, delivering enhanced segmentation performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 300, + 284, + 435 + ], + "blocks": [ + { + "bbox": [ + 62, + 300, + 284, + 435 + ], + "lines": [ + { + "bbox": [ + 62, + 300, + 284, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 284, + 435 + ], + "type": "image", + "image_path": "52f97c3ac936688321cbf323ebd1fcc59de8592d6bb39aa4d3a4289919c78b8d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 435, + 295, + 479 + ], + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 479 + ], + "type": "text", + "content": "Figure 4. Visualization of feature maps (§4.5). From top to bottom are: (a) Ground truth masks. b) Keyframe mask embeddings generated by the token before TDA. (c) Keyframe mask embeddings generated by the token after TDA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 493, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 624 + ], + "type": "text", + "content": "its limited semantic understanding. In contrast, VRS-HQ demonstrates a strong ability to perceive spatiotemporal relationships. For the middle example, VISA mistakenly segments the person flying the kite, while VRS-HQ correctly identifies the kite, reflecting the stronger capacity to apply world knowledge and better handling of small-object segmentation. The right column showcases VRS-HQ's robust temporal reasoning capabilities, driven by our temporal dynamic aggregation and adaptive keyframe selection. Conversely, VISA misses the target due to incorrect keyframe selection during inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "Feature Visualization. To understand the TDA's impact, we visualize mask embeddings generated by the mask decoder using both the fused and unfused tokens with keyframe features. As illustrated in Fig. 4, we employ Principal Component Analysis (PCA) [35] to reduce the dimensionality of the mask embeddings to a single channel. This single-channel representation is then blended with the orig" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 297, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 406 + ], + "type": "text", + "content": "inal image to highlight high-attention regions. Before TDA fusion, the token often focuses on non-target elements (e.g., the falling person in the first example, the woman in the second, and the central dancing man in the third). In contrast, the similarity-weighted fusion of the frame-level SEG tokens into the token significantly enriches the positional and semantic information contained within the token, enabling the model to more effectively concentrate on the correlated target objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 422, + 388, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 388, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 388, + 434 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 442, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 556, + 598 + ], + "type": "text", + "content": "We present VRS-HQ, a novel approach that leverages the temporal reasoning of MLLM and the robust tracking of SAM2 for high-quality Video Reasoning Segmentation. Our method utilizes a temporal-level token and frame-level tokens to capture temporal relations and spatial features, respectively. These tokens are integrated using the Temporal Dynamic Aggregation, with SAM2 utilizing the token for keyframe segmentation and mask propagation. Moreover, we introduce a Token-driven Keyframe Selection that uses the token to generate occlusion scores for robust keyframe selection. Extensive experiments confirm that VRS-HQ achieves state-of-the-art performance on various benchmarks, demonstrating strong capabilities in handling Video Reasoning Segmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 613, + 373, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 613, + 373, + 625 + ], + "spans": [ + { + "bbox": [ + 315, + 613, + 373, + 625 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 633, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 319, + 633, + 555, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 633, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 319, + 633, + 555, + 688 + ], + "type": "text", + "content": "[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. In Adv. Neural Inform. Process. Syst., 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 319, + 691, + 554, + 713 + ], + "type": "text", + "content": "[2] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with mul" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29190" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "timodal transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4985-4995, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 62, + 96, + 295, + 129 + ], + "type": "text", + "content": "[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1209-1218, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 130, + 295, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 130, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 130, + 295, + 175 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In *Eur. Conf. Comput. Vis.*, pages 213-229. Springer, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 175, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 295, + 230 + ], + "type": "text", + "content": "[5] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1971-1978, 2014." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 294, + 275 + ], + "type": "text", + "content": "[6] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In *Eur. Conf. Comput. Vis.*, pages 640–658. Springer, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 277, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 294, + 319 + ], + "type": "text", + "content": "[7] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In Int. Conf. Comput. Vis., pages 2694-2703, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 322, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 322, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 294, + 365 + ], + "type": "text", + "content": "[8] Sitong Gong, Yunzhi Zhuge, Lu Zhang, Yifan Wang, Pingping Zhang, Lijun Wang, and Huchuan Lu. Avs-mamba: Exploring temporal and multi-modal mamba for audio-visual segmentation. IEEE Trans. Multimedia, 2025." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 366, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 294, + 399 + ], + "type": "text", + "content": "[9] Sitong Gong, Yunzhi Zhuge, Lu Zhang, Pingping Zhang, and Huchuan Lu. Complementary and contrastive learning for audio-visual segmentation. IEEE Trans. Multimedia, 2025." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 401, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 401, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 57, + 401, + 295, + 454 + ], + "type": "text", + "content": "[10] Junwen He, Yifan Wang, Lijun Wang, Huchuan Lu, Jun-Yan He, Jin-Peng Lan, Bin Luo, and Xuansong Xie. Multi-modal instruction tuned llms with fine-grained visual perception. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13980-13990, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 456, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 295, + 500 + ], + "type": "text", + "content": "[11] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13332-13341, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 545 + ], + "type": "text", + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 601 + ], + "type": "text", + "content": "[13] Tianrui Hui, Shaofei Huang, Si Liu, Zihan Ding, Guanbin Li, Wenguan Wang, Jizhong Han, and Fei Wang. Collaborative spatial-temporal modeling for language-queried video actor segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4187–4196, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "text", + "content": "[14] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13700-13710, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[15] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Int. Conf. Comput. Vis., pages 4015-4026, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[17] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 9579-9589, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "text", + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. pages 19730-19742. PMLR, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "text", + "content": "[19] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Eur. Conf. Comput. Vis., pages 323–340. Springer, 2025." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "type": "text", + "content": "[20] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 23592-23601, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 277, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 308 + ], + "type": "text", + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Adv. Neural Inform. Process. Syst., 36, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 311, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 311, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 553, + 354 + ], + "type": "text", + "content": "[22] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 356, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 356, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 356, + 553, + 399 + ], + "type": "text", + "content": "[23] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 401, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 553, + 443 + ], + "type": "text", + "content": "[24] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 446, + 553, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 446, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 317, + 446, + 553, + 500 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 502, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 502, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 553, + 555 + ], + "type": "text", + "content": "[26] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7141-7151, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 558, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 622 + ], + "type": "text", + "content": "[27] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13009-13018, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[28] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., pages 3505-3506, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2:" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29191" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "text", + "content": "Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[30] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 26374-26383, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "text", + "content": "[31] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "text", + "content": "[32] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Eur. Conf. Comput. Vis., pages 208–223. Springer, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[33] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Int. Conf. Comput. Vis., pages 3939-3948, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "text", + "content": "[34] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11686-11695, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 376 + ], + "type": "text", + "content": "[35] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 422 + ], + "type": "text", + "content": "[36] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Int. Conf. Comput. Vis., pages 2761-2770, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 423, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 465 + ], + "type": "text", + "content": "[37] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4974-4984, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "type": "text", + "content": "[38] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3858-3869, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "text", + "content": "[39] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. In Eur. Conf. Comput. Vis., 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[40] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In AAAI, pages 6449-6457, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[41] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. arXiv eprints, pages arXiv-2312, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[42] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavo: Language-aware vision transformer for referring image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18155-18165, 2022." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 386 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[43] Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Yu Qiao, and Hengshuang Zhao. Villa: Video reasoning segmentation with large language model. arXiv preprint arXiv:2407.14500, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[44] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In AAAI, pages 12993-13000, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "text", + "content": "[45] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[46] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with human-intent reasoning. arXiv preprint arXiv:2312.17448, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[47] Mengyuan Zhu, Yunzhi Zhuge, Sitong Gong, Lu Zhang, and Huchuan Lu. Fdavs: Exploring frequency-driven modality enhancement in audio-visual segmentation. In Int. Conf. Multimedia and Expo, 2025." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[48] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In IEEE Conf. Comput. Vis. Pattern Recog., pages 15116-15127, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[49] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Adv. Neural Inform. Process. Syst., 36, 2024." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29192" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/bf61a946-383a-4ffe-b29d-37540410b684_content_list.json b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/bf61a946-383a-4ffe-b29d-37540410b684_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..18a48b77898d402abfda4667a0b025d7a404cec2 --- /dev/null +++ b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/bf61a946-383a-4ffe-b29d-37540410b684_content_list.json @@ -0,0 +1,1593 @@ +[ + { + "type": "text", + "text": "The Devil is in the Prompts: Retrieval-Augmented Prompt Optimization for Text-to-Video Generation", + "text_level": 1, + "bbox": [ + 117, + 128, + 880, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bingjie Gao $^{1,2}$", + "bbox": [ + 186, + 202, + 305, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Qiao $^{2\\dagger}$", + "bbox": [ + 254, + 220, + 336, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyu Gao $^{1,2}$", + "bbox": [ + 354, + 203, + 462, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Li Niu $^{1\\dagger}$", + "bbox": [ + 361, + 220, + 431, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyuan Chen $^{2\\dagger}$", + "bbox": [ + 455, + 220, + 584, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yujie Zhou1,2", + "bbox": [ + 684, + 203, + 792, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yaohui Wang $^{2\\dagger}$", + "bbox": [ + 612, + 220, + 733, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University", + "bbox": [ + 107, + 239, + 357, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ Fudan University", + "bbox": [ + 367, + 238, + 874, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The evolution of Text-to-video (T2V) generative models, trained on large-scale datasets, has been marked by significant progress. However, the sensitivity of T2V generative models to input prompts highlights the critical role of prompt design in influencing generative outcomes. Prior research has predominantly relied on Large Language Models (LLMs) to align user-provided prompts with the distribution of training prompts, albeit without tailored guidance encompassing prompt vocabulary and sentence structure nuances. To this end, we introduce RAPO, a novel Retrieval-Augmented Prompt Optimization framework. In order to address potential inaccuracies and ambiguous details generated by LLM-generated prompts. RAPO refines the naive prompts through dual optimization branches, selecting the superior prompt for T2V generation. The first branch augments user prompts with diverse modifiers extracted from a learned relational graph, refining them to align with the format of training prompts via a fine-tuned LLM. Conversely, the second branch rewrites the naive prompt using a pre-trained LLM following a well-defined instruction set. Extensive experiments demonstrate that RAPO can effectively enhance both the static and dynamic dimensions of generated videos, demonstrating the significance of prompt optimization for user-provided prompts. Project website: GitHub.", + "bbox": [ + 91, + 323, + 483, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 731, + 220, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid advancement of diffusion models [28, 33, 34], visual content creation has experienced remarkable progress in recent years. The generation of images, as well as videos from text prompts utilizing large-scale diffusion models, referred to as text-to-images (T2I) [13, 30, 35] and text-to-videos (T2V) [5, 12, 40] generation, have attracted significant interest due to the broad range of applications in real-world scenarios. Various efforts have been made to en", + "bbox": [ + 89, + 756, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hance the performance of these models, including improvements in model architecture [23, 26, 26], learning strategies [36, 44], and data curation [32, 39, 41].", + "bbox": [ + 511, + 292, + 903, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies [17, 31, 44] have revealed that employing long, detailed prompts with a pre-trained model typically produces superior quality outcomes compared to utilizing shallow descriptions provided by users. This has underscored the significance of prompt optimization as an important challenge in text-based visual content creation. The prompts provided by users are often brief and lack the essential details required to generate vivid images or videos. Simply attempting to optimize prompts by manually adding random descriptions can potentially mislead models and degrade the quality of generative results, resulting in outputs that may not align with user intentions. Therefore, developing automated methods to enhance user-provided prompts becomes essential for improving the overall quality of generated content.", + "bbox": [ + 511, + 339, + 906, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Towards improving image aesthetics and ensuring semantic consistency, several attempts [19, 27, 46] have been made in previous T2I works for prompt optimization. These efforts primarily involve instructing a pre-trained or finetuned Large Language Model (LLM) to incorporate detailed modifiers into original prompts, with the aim of enhancing spatial elements such as color and relationships. While these approaches have displayed promising outcomes in image generation, studies [7, 17] reveal that their impact on video generation remains limited, especially in terms of enhancing temporal aspects such as motion smoothness and minimizing temporal flickering.", + "bbox": [ + 511, + 566, + 908, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For T2V generation, we observe that the quality of spatial and temporal elements is significantly influenced by the selection of appropriate verb-object phrases and the structure of input prompts, which largely rely on analyzing and structuring the entire training data in a systematical manner. To this end, we propose RAPO, a Retrieval-Augmented Prompt Optimization framework for T2V generation.", + "bbox": [ + 511, + 748, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The primary objective of RAPO is to convert user-provided prompts into optimized prompts that retain the original semantics while closely adhering to the format of", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author", + "bbox": [ + 109, + 886, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3173", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fff60468cf1a329b4b3026eb00d3ad87bfb7cb3e11ed85bd2c07fb86bc9a501.jpg", + "image_caption": [ + "(b) static dimension", + "Figure 1. Videos generated using LaVie [40] conditioned on user-provided prompts and optimized prompts from RAPO. The optimized prompts can significantly enhance both the static and dynamic qualities of the generated videos, making them more visually appealing." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 906, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the training data, including the vocabulary and sentence structure. To achieve this, we initially construct a Relation Graph from the training data, with each node representing words and edges indicating the relationships between neighboring words. By leveraging any user-provided prompt, we can retrieve the most relevant terms to enhance the original descriptions.", + "bbox": [ + 88, + 636, + 482, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Subsequently, a specially fine-tuned Refactoring LLM is introduced to reorganize the structure and writing style of the word-augmented prompt, aligning it more cohesively with the original training data. Additionally, we employ a Rewrite LLM to directly transform user-provided prompts into optimized prompts that follow a specific format consistent with the training prompts, serving as candidates for optimized prompts. In order to address potential inaccuracies and ambiguous details generated by LLMs, we conclude by fine-tuning a Discriminator LLM to identify the most suit-", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "able prompt for T2V generation.", + "bbox": [ + 511, + 637, + 732, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We apply RAPO on two T2V generation models LaVie [40] and Latte [26], and evaluate it on three benchmarks, namely VBench [21], EvalCrafter [25] and T2V-CompBench [37]. It is worth noticing that RAPO greatly promotes the performance of multiple objects dimension in VBench from $37.71\\%$ to $64.86\\%$ using LaVie and from $29.55\\%$ to $52.78\\%$ using Latte. In addition, the experimental results show that RAPO improves the performance of T2V models to synthesize scenes with multiple subjects, even unusual scenes like \"a mushroom growing out of a human head\" as shown in Fig. 1. Our contributions can be summarized as follows.", + "bbox": [ + 511, + 652, + 906, + 834 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose RAPO, a Retrieval-Augmented Prompt Optimization framework T2V generation.", + "- We propose word augmentation module and sentence refactoring module to optimize prompts considering" + ], + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3174", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prompt vocabulary and specific sentence format.", + "bbox": [ + 102, + 90, + 424, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We validate the effectiveness of RAPO through extensive experiments on several benchmarks.", + "bbox": [ + 91, + 106, + 480, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 152, + 232, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Video generation. With the remarkable breakthroughs of diffusion models [28, 33, 34], the generations of 3D content [9, 24, 43], images [3, 13, 30], and videos [4, 5, 20] from text descriptions achieve rapid advancement. Text-to-Video (T2V) [8, 40, 47] Generation aims to automatically create videos that match given textual descriptions. This process generally involves comprehending the scenes, objects, and actions described in the text and converting them into a sequence of cohesive visual frames, producing a video that is logically and visually consistent. T2V generation is wildly used in applications, such as animations [10, 15, 18] and automatic movie generation [42, 48, 50]. However, large T2V generative models [26, 40, 44] trained on large-scale dataset could not adequately demonstrate their potential in generation due to mismatch between training and inference.", + "bbox": [ + 89, + 178, + 482, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt optimization. T2I and T2V generative models are sensitive to input prompts. However, the well-performed prompts are often model-specific and coherent with training prompts, misaligned with user input. Therefore, several studies [11, 17, 27, 45] are conducted to explore the generative potential of T2I and T2V generative models. Hao et al. [17] propose a learning-based prompt optimizing framework unitizing reinforce learning for generating more aesthetically pleasing images. Chen et al. [11] enhance user prompts by leveraging the user's historical interactions with the system. Mo et al. [27] propose Prompt Auto-Editing (PAE) method to decide the weights and injection time steps of each word without manual intervention. These methods primarily focus on prompts optimizing for T2I models and lack extension to T2V models. Yang et al. [44] use large language models (LLMs) to transform short prompts into more detailed ones, maintaining a consistent visual structure. Polyak et al. [31] develop a teacher-student distillation approach for prompt optimization to improve computational efficiency and reduce latency. However, the results of optimized prompts usually could not be well-aligned with training prompts due to the misleading of the LLMs and the lack of more refined guidance.", + "bbox": [ + 89, + 421, + 482, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 782, + 181, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, RAPO consists of three parts, 1) a word augmentation module, 2) a sentence refactoring module, as well as 3) a prompt selection module. Given a user-provided prompt $x_{i}$ , firstly, the word augmentation module utilizes an interactive retrieval-merge mechanism between a relation graph $\\mathcal{G}$ and a LLM $\\mathcal{L}$ to augment the prompt", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by adding related subject, action and atmosphere modifiers. Then, a fine-tuned LLM $\\mathcal{L}_r$ is applied to refactor the entire sentence into $x_r$ . $x_r$ has a more unified format which is consistent with the prompt length and format distribution in training data. Finally, a discriminator in the prompt selection module decides between $x_r$ and a naively augmented prompt $x_n$ obtained directly from a LLM via instruction, as the most suitable augmented prompt for T2V generation. We proceed to introduce each module in detail in the following sections.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Word Augmentation Module", + "text_level": 1, + "bbox": [ + 511, + 251, + 769, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a user-provided prompt $x_{i}$ , the word augmentation module aims to enrich $x_{i}$ with more multiple, straightforward and relevant modifiers. It is achieved through retrieving modifiers meeting the requirements from a built relation graph $\\mathcal{G}$ , and merging them into $x_{i}$ through $\\mathcal{L}$ via instruction. In this section, we first introduce the construction and retrieval of relation graph. And we introduce the instruction format and retrieval-merge mechanism of $\\mathcal{L}$ .", + "bbox": [ + 511, + 272, + 903, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relation Graph $\\mathcal{G}$ . As shown in Fig. 3, we construct relation graph $\\mathcal{G}$ based on training prompts database. For each training prompt, we utilize $\\mathcal{L}$ to extract scene and corresponding related modifiers (subject, action, atmosphere descriptions). Each scene serves as a core node, with subject, action and atmosphere modifiers connected as individual sub-nodes in relation graph. For each extracted scene, we first check whether it exists in relation graph or not. If so the extracted related modifiers will be connected to the existing one. If not the extracted scene becomes a new core node with related modifiers connected. Finally, we can obtain a relation graph covering diverse scenes with multiple modifiers connected.", + "bbox": [ + 511, + 393, + 903, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For relation graph retrieval, we utilize a sentence transformer pre-trained model to extract features of prompt, and employ the cosine similarity to measure similarity between sentence features. We first retrieve the top-k relevant scenes from $\\mathcal{G}$ for $x_{i}$ . Then we retrieve all modifiers connected to the retrieved scenes. We select the top-k relevant modifiers $\\{p_n|_{n = 0}^{k - 1}\\}$ from all retrieved modifiers, preparing for the retrieval-merge mechanism of $\\mathcal{L}$ .", + "bbox": [ + 511, + 590, + 903, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM $\\mathcal{L}$ . We augment $x_{i}$ with retrieved modifiers $\\{p_n|_{n = 0}^{k - 1}\\}$ from relation graph. We rename $x_{i}$ with $x_{i}^{0}$ to illustrate the process of iterative merging. Specifically, the retrieved modifiers are merged into input prompt $x_{i}^{0}$ one by one through prompting $\\mathcal{L}$ , to maintain the information of the original input while adding relevant modifiers.", + "bbox": [ + 511, + 709, + 903, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} ^ {m + 1} = f \\left(x _ {i} ^ {m}, p _ {i} ^ {m}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 811, + 903, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $m = 0,1,\\dots,k - 1$ . $f$ is a function that combine $x_{i}^{m}$ and $p_i^m$ reasonably by $\\mathcal{L}$ . For instance, a merged prompt \"a woman dressed in a black suit representing a funeral\" is resulted from merging the user-provided prompt \"a woman", + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3175", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/30c2cf4568e1472d731af966eb03ac0fc3d59f023ea6e4cabbaa0f9f69a4d329.jpg", + "image_caption": [ + "Figure 2. Overview of RAPO. The naive prompt is optimized by two branches respectively. For the first branch, it is enriched by the word augmentation module based on a constructed relation graph and a frozen Large Language Model (LLM). Subsequently, augmented prompt is refactored by a finetuned LLM into a specific format in the sentence refactoring module. For the second branch, the naive prompt is directly rewritten by a frozen LLM. Finally, the prompt selection module selects the better one from two branches' results as input for T2V model." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 893, + 340 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea314a54bd93ade58abafa5bff17cac48974326447a62928f4f8eb1437b8aaff.jpg", + "image_caption": [ + "Figure 3. The construction of relation graph. Relation graph consists of multiple nodes (scenes acting as core nodes with modifiers connected as sub-nodes). For each prompt in database, LLM extracts scene and related modifiers. Based on whether the extracted scene is already in the graph or not, different methods are used to incorporate the new information into the graph." + ], + "image_footnote": [], + "bbox": [ + 94, + 448, + 496, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "representing a funeral\" and a retrieved modifier \"a black suit\". We prompt $\\mathcal{L}$ to perform general prompt merging in a normal manner as the template in Tab. 1. In instruction, we provide some prompt pairs $E = \\{e_i|_{i = 0}^n\\}$ as examples, in", + "bbox": [ + 89, + 839, + 485, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which $e_i$ contains input prompt, a modifier and corresponding merged result.", + "bbox": [ + 511, + 450, + 906, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLM Template for Retrieval-Merge Mechanism", + "text_level": 1, + "bbox": [ + 522, + 500, + 862, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Suppose you are a Text Merger. You receive two inputs from the user: a description body and a relevant modifier. Your task is to enrich the description body with relevant modifiers while retaining the description body. You should ensure that the output text is coherent, contextually relevant, and follows the same structure as the examples provided.", + "bbox": [ + 521, + 522, + 897, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Examples of prompt-pairs provided: $E = \\{e_i|_{i = 0}^n\\}$", + "bbox": [ + 521, + 627, + 864, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input description body and modifier are: $\\{x_i^m,p_i^m\\}$", + "bbox": [ + 522, + 643, + 861, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The merged prompt is: $\\{x_i^{m + 1}\\}$", + "bbox": [ + 522, + 659, + 740, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Input template for retrieval-merge mechanism. This template specifies how a frozen LLM iteratively merges user-provided prompt texts with relevant modifiers retrieved from a relation graph, thereby enriching the prompt's semantic content and aligning it with the training prompt structure for improved text-to-video synthesis.", + "bbox": [ + 511, + 689, + 908, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sentence Refactoring Module", + "text_level": 1, + "bbox": [ + 511, + 801, + 777, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sentence refactoring module aims to refactor word augmented prompts from word augmentation module to be more consistent with prompt format in training data. It is achieved through a fine-tuned LLM $L_{r}$ named as refactoring model. In this section, we introduce the training data", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3176", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "preparation and instruction tuning for $L_{r}$ .", + "bbox": [ + 89, + 90, + 367, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data preparation. We represent the required dataset for training refactoring model by $\\{D_r = r_i|_{i=1}^{N^r}\\}$ , in which $r_i$ involves a pair of prompts and $N^r$ is the number of training prompts pairs. Specifically, $r_i = (w_i, c_i)$ , in which $w_i$ targets to simulate world augmented prompt, and $c_i$ represents the target prompt, that is, a training prompt for T2V models. $w_i$ and $c_i$ share similar semantics while different in the prompt format and length. Therefore, we generate $w_i$ automatically through rewriting $c_i$ utilizing $\\mathcal{L}$ via instruction to break the unified training prompt format but maintaining the original semantics.", + "bbox": [ + 88, + 106, + 482, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction tuning for $L_{r}$ . We employ instruction tuning for fine-tuning a LLM on our constructed dataset of instructional prompts and corresponding outputs. The constructed dataset is based on $\\{D_r = r_i|_{i = 1}^{N^r}\\}$ containing instructional prompts and corresponding outputs. The template of the instruction tuning dataset for $L_{r}$ is shown as Tab. 2.", + "bbox": [ + 89, + 272, + 483, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction Tuning Dataset for $L_{r}$", + "text_level": 1, + "bbox": [ + 99, + 380, + 341, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction. Refine format and word length of the sentence: $w_{i}$ . Maintain the original subject descriptions, actions, scene descriptions. Append additional straightforward actions to make the sentence more dynamic if necessary.", + "bbox": [ + 99, + 401, + 473, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: target prompt $c_{i}$ .", + "bbox": [ + 99, + 478, + 267, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Instruction tuning dataset template for $L_{r}$ . This template directs LLM fine-tuning to restructure augmented prompts by adjusting their format while preserving semantics, aligning them with the training data's style for improved T2V generation.", + "bbox": [ + 89, + 508, + 483, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Prompt Selection Module", + "text_level": 1, + "bbox": [ + 89, + 590, + 323, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, prompt selection module contains a fine-tuned LLM $\\mathcal{L}_d$ named prompt discriminator to select the better one between $x_{r}$ from sentence refactoring module, and a naively augmented prompt $x_{n}$ obtained directly from a LLM via instruction. In this section, we introduce the training data preparation and instruction tuning for $\\mathcal{L}_d$ .", + "bbox": [ + 89, + 612, + 482, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data preparation. We represent the required dataset for training refactoring model by $\\{D_{d} = d_{i}|_{i = 1}^{N^{d}}\\}$ , in which $d_{i}$ contains three prompts and $N^d$ is the number of training prompts triples. Specifically, $d_{i} = (x_{i},x_{r},x_{n},y_{d})$ , in which $y_{d}$ represents the discriminator label to select the better one for T2V generation from $x_{r}$ and $x_{n}$ given input prompt $x_{i}$ . To simulate the user-provided prompts, we collect diverse prompts from several T2V benchmarks and generate more utilizing $\\mathcal{L}$ via instruction. $x_{r}$ and $x_{n}$ can be obtained from the proposed RAPO as shown in Fig. 2 given $x_{i}$ . We determine $y_{d}$ through the evaluation of generated videos conditioned on $x_{r}$ and $x_{n}$ . Specifically, the evaluations of T2V models performance involves diverse", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dimensions. For collected or generated prompts, we need to determine the evaluation dimension according to prompt content. We automatically decide the evaluation dimension of input prompts utilizing $\\mathcal{L}$ , then choose the corresponding metrics to evaluate generated videos.", + "bbox": [ + 511, + 90, + 905, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction tuning for $L_{d}$ . Similar to $L_{r}$ , we employ instruction tuning for $L_{d}$ based on $\\{D_{d} = d_{i}|_{i = 1}^{N^{d}}\\}$ . The template of the instruction tuning dataset for $L_{d}$ is shown as Tab. 3.", + "bbox": [ + 511, + 167, + 906, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction Tuning Dataset for $L_{d}$", + "text_level": 1, + "bbox": [ + 524, + 243, + 764, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction. Given user-provided prompt $x_{i}$ , select the better optimized prompt from $x_{r}$ and $x_{n}$ . The chosen prompt is required to contain multiple, straightforward, and relevant modifiers about $x_{i}$ while involving the semantics of $x_{i}$ .", + "bbox": [ + 521, + 265, + 895, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: discriminator label $y_{d}$ .", + "bbox": [ + 522, + 340, + 730, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Instruction tuning dataset template for $L_{d}$ . This template aims to train a discriminator LLM that evaluates multiple refined prompts and selects the optimal one based on the inclusion of clear, straightforward modifiers and faithful semantic alignment.", + "bbox": [ + 511, + 371, + 906, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 459, + 648, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 484, + 702, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation and Metrics. We utilize the VBenchmark [21], EvalCrafter [25] and T2V-CompBench [37] for quantitative performance evaluation. VBenchmark is an extensive benchmark comprising 16 fine-grained dimensions designed to systematically evaluate the quality of generated videos. EvalCrafter is a comprehensive evaluation benchmark that includes approximately 17 objective metrics for evaluating video generation performance. T2V-CompBench is the first benchmark tailored for compositional text-to-video generation.", + "bbox": [ + 511, + 506, + 905, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison to other methods. We compare the optimized prompts with our method to three types of prompts: the short primary prompts, the prompts generated from GPT-4 [2] and prompt refiner from Open-sora [1], which is based on fine-tuned LLaMA 3.1 [38]. The specific inference prompt template for GPT-4 [2] can be found in supplementary materials.", + "bbox": [ + 511, + 657, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 772, + 732, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The well-performed prompts are model-specific and aligned with the distribution of training prompts. We employ Vimeo25M [40], a training dataset consisting of 25 million text-video pairs as our analysis dataset. At the same time, we choose LaVie [40] and Latte [26] as analysis T2V models, which belong to the diffusion-based and DiT architectures respectively and use Vimeo25M as one of training", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3177", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/51d2efaa4a37c56c0ec9ab1e181ac09ecf5ee59fa172ab1c687300eb73fd6163.jpg", + "image_caption": [ + "(a) Static dimension. From left to right: naive, GPT-4, Open-sora, Ours." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abb8e63210833c09fcc773a34a35bd02f36b09472ea92a54d52758463ee0bce2.jpg", + "image_caption": [ + "(b) Dynamic dimension (a robot is dancing). From left to right, from top to bottom: naive, GPT-4, Open-sora, Ours.", + "Figure 4. Qualitative comparisons across dynamic and static dimensions. This figure showcases videos generated using LaVie with short prompts, GPT-4 and Open-sora prompt optimizations, and our RAPO method. Videos produced with RAPO exhibit significantly sharper spatial details, smoother temporal transitions, and a closer semantic alignment with the input text." + ], + "image_footnote": [], + "bbox": [ + 91, + 300, + 898, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "datasets. And we represent some extension results to some other T2V models in supplementary materials. For relation graph construction, we utilize Mistral [22] to extract scenes with corresponding subject, action and atmosphere descriptions from Vimeo25M dataset, and use all-MiniLM-L6-v2 as sentence transformer pre-trained model. We filter about 2.1M valid sentences from Vimeo25M dataset. For refactoring model training data, we prepare about 86k prompt-pairs following data preparation method in Section 3.2. For prompt discriminator training data, we first generate 7K text captions using Mistral, covering all the dimensions in VBenchmark [21]. The specific prompt template for extraction and dataset construction can be found in supplementary materials. We perform LoRA fine-tuning using LLaMA 3.1 [38], and fine-tune 8 epochs and 3 epochs for refactoring model and prompt discriminator respectively with a single A100, using a batch size of 32 and a LoRA rank of 64.", + "bbox": [ + 89, + 564, + 485, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Results", + "text_level": 1, + "bbox": [ + 89, + 847, + 272, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons. As shown in Fig. 4, compared with user-provided prompts and other prompt optimization", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods, RAPO centers the input semantics and promotes the quality of generated videos from static and dynamic dimensions. The relevant modifiers promote model comprehending the simple user-provided prompts. The detailed prompts of different methods and more qualitative comparisons are provided in supplementary materials.", + "bbox": [ + 511, + 564, + 906, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparisons. As shown in Tab. 4 and Tab. 5, the results show that RAPO surpasses other methods in static dimensions (e.g., visual quality, object class) and dynamic dimensions (e.g., human actions, temporal flickering). The consistent performance across various benchmarks demonstrates the robustness and versatility of RAPO. Although the optimized prompts from GPT-4 and Open-sora enriches the plain user-provided prompts with more details describing the scenes, objects, and actions, these lengthy and complex descriptions may confuse models and even worsen the generation. It is worth noticing that RAPO significantly enhances the score of multiple objects dimension from $37.71\\%$ to $64.86\\%$ using LaVie and from $29.55\\%$ to $52.78\\%$ using Latte as shown in Tab. 4. And RAPO outperforms other methods on T2V-CompBench as shown in Tab. 6 for some selected dimensions, which are focused on", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3178", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3d5809f542c383a01942e359c1831b5ce8e07cb025d8b6b09cc5fa1a92a67909.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsTotal Scoretemporal flickeringimaging qualityhuman actionobject classmultiple objectsspatial relationship
LaVie [40]80.89%96.62%69.00%95.80%92.09%37.71%37.27%
LaVie-GPT4 [2]79.69%96.14%70.27%83.80%88.73%36.23%50.55%
LaVie-Open-sora [1]79.75%96.42%70.42%87.00%91.29%36.52%54.37%
LaVie-RAPO82.38%96.86%71.40%96.80%96.91%64.86%59.15%
Latte [26]77.03%97.10%63.38%88.40%83.86%29.55%40.63%
Latte-GPT4 [2]77.40%97.52%63.54%85.80%78.32%27.73%36.72%
Latte-Open-sora [1]77.23%97.67%64.19%84.60%83.60%30.00%35.12%
Latte-RAPO79.97%98.17%66.72%95.20%96.47%52.78%41.31%
", + "bbox": [ + 94, + 88, + 903, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Quantitative comparisons on VBench. RAPO achieves the highest overall scores, especially in static attributes and multiple object generation.", + "bbox": [ + 89, + 223, + 906, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/037cc8fc873890a5b4a3548a5408cb826c86e4ea1e47166dede10bdab2200267.jpg", + "image_caption": [ + "a bicycle and a car" + ], + "image_footnote": [], + "bbox": [ + 104, + 275, + 285, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8cd071931bc10f1070aeb41e7d05f1b0d294a4435456b41b516272267004accf.jpg", + "image_caption": [ + "A bicycle and a car are parked side by side." + ], + "image_footnote": [], + "bbox": [ + 289, + 275, + 470, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/92c237b738b4b048c6079eee7a1da6481c9e8de2a9011f2c6d014059e8942491.jpg", + "image_caption": [ + "bicycle" + ], + "image_footnote": [], + "bbox": [ + 102, + 439, + 194, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0ff19a6aac7f59ab36df8a7d4e25770bdada4c17f178a5202afeedd6e6233e0.jpg", + "image_caption": [ + "car" + ], + "image_footnote": [], + "bbox": [ + 197, + 439, + 287, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f435fc08c7eaafab09a1258a44db705153064e5c5b5f28393c45fe1eeba74c0f.jpg", + "image_caption": [ + "bicycle" + ], + "image_footnote": [], + "bbox": [ + 290, + 439, + 377, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98c3eeb4e7e42700db762e69d6d81d45b7d3ae30ea64fe43e37ede6416d3da75.jpg", + "image_caption": [ + "car", + "Figure 5. Visualization on attention map on multiple objects from different prompts. Adding description of the relative spatial position between objects can improve multi-object generation." + ], + "image_footnote": [], + "bbox": [ + 383, + 439, + 470, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "complex compositional T2V tasks. These experiments verified that PAPO significantly improves the performance of generating scenes involving more than two subjects.", + "bbox": [ + 89, + 609, + 483, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analyses", + "text_level": 1, + "bbox": [ + 89, + 666, + 196, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multiple objects. Synthesis quality of generated videos often declines when tasked with generating outputs that accurately represent prompts involving multiple objects. This issue is also prevalent in the T2I model, and several studies [14, 16, 30] have highlighted that the blended context created by the CLIP text encoder leads to improper binding. Meanwhile, some related works [6, 29] focus on image latents to address information loss, while the others [7, 49] pay more attention to text embedding to deal with the issue. However, few have explored optimizing prompts to improve the performance of multiple obejsts task. We apply our method to text-to-image using SD 1.4 [34], which uses the same text encoder with LaVie [40]. We test on prompts about multiple objects, and remove the irrelevant modifiers", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7c60393e629fc0f95b39fd870b09183e1102602a5910627e73e01a9059875ec.jpg", + "image_caption": [ + "Figure 6. Prompt length distribution comparison among various methods. The distribution of RAPO-optimized prompts is more closer to the training prompts." + ], + "image_footnote": [], + "bbox": [ + 547, + 277, + 870, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "like action and atmosphere descriptions. As shown in Fig. 5, we can find the relevant spatial descriptions boost the performance of multiple objects. More examples can be found in supplementary materials.", + "bbox": [ + 511, + 516, + 905, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Statistical analysis of text. As shown in Fig. 6, we compared the word length distributions of prompts from the T2V training set, user prompts (simulated via VBench, EvalCrafter, and T2V-CompBench), and optimized prompts generated by various methods. The results show that the prompt length distribution produced by RAPO is closest to that of the training set, and this consistency unleashes the model's generative potential to produce better videos. In contrast, user prompts are too short and lack necessary details, while other methods generate longer prompts that contain excessive details and complex vocabulary, which may be counterproductive, as shown in Tab. 4 and Tab. 5.", + "bbox": [ + 511, + 578, + 906, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 771, + 666, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments on the VBench benchmark to examine the individual and combined effects of different modules in RAPO. Additionally, we perform ablation experiments on various configurations of $\\mathcal{L}$ . We present more ablation experiments on hyperparameters in the supplementary materials.", + "bbox": [ + 511, + 794, + 905, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablating each modules in RAPO. We directly obtain the", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3179", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f4696df7c111ec210eb665163138ca56a9427654269628de55e179d16535069e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsFinal Sum Scoremotion qualitytext-video alignmentvisual qualitytemporal consistency
LaVie [40]24853.1969.6064.8160.87
LaVie-GPT4 [2]24654.0565.5164.9661.22
LaVie-Open-sora [1]25153.0771.3865.2661.41
LaVie-RAPO25653.3474.3866.6261.29
Latte [26]21750.0355.4957.6553.94
Latte-GPT4 [2]21851.3653.6558.0254.65
Latte-Open-sora [1]22250.2557.3258.7155.47
Latte-RAPO22751.7360.8659.2455.26
", + "bbox": [ + 94, + 88, + 903, + 237 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1637b2df5bb09657e372216015ac4b45ca80874686b7d5555348567ae099ad01.jpg", + "table_caption": [ + "Table 5. Quantitative comparisons on EvalCrafter. The results demonstrate that RAPO consistently achieves higher scores, especially in terms of text-video alignment and visual quality, indicating its superior ability to enhance the overall quality of generated videos." + ], + "table_footnote": [], + "table_body": "
Modelsconsistent attribute bindingdynamic attribute bindingaction bindingobject interactions
LaVie [40]0.6200.2320.4830.760
LaVie-GPT4 [2]0.5610.2180.4280.620
LaVie-Open-sora [1]0.5320.2140.4700.698
LaVie-RAPO0.6920.2670.6350.839
Latte [26]0.6330.2270.4760.792
Latte-GPT4 [2]0.5980.2100.4050.688
Latte-Open-sora [1]0.5490.2030.4870.743
Latte-RAPO0.7060.2580.5910.856
", + "bbox": [ + 94, + 289, + 901, + 444 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/32b53929937085dc5c387d446bea7c0c9475bb2321a10c1e8079c9a62f155e0f.jpg", + "table_caption": [ + "Table 6. Quantitative comparisons on T2V-CompBench. RAPO outperforms alternative prompt optimization methods, exhibiting better performance at managing complex compositional tasks and ensuring coherent interactions between multiple objects in the generated videos." + ], + "table_footnote": [], + "table_body": "
word augmentationsentence refactoringprompt selectionVBatch Total Score
(a)✓80.37%
(b)✓79.75%
(c)✓✓81.58%
(d)✓✓81.75%
(e)✓✓80.60%
(f)✓✓✓82.38%
", + "bbox": [ + 93, + 505, + 480, + 583 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/52086a04b74b840671f49c65b27cb6de08525c19c9a0e43b1d02ce725cc2a40b.jpg", + "table_caption": [ + "Table 7. Ablation studies of different modules in RAPO on VBench. Each module improves performance, while the combined use of all three leads to the highest evaluation score, confirming the synergistic effect of the full RAPO framework." + ], + "table_footnote": [], + "table_body": "
GPT-4MistralLLaMA
VBenchmark Total Score82.38%82.25%82.10%
", + "bbox": [ + 94, + 665, + 480, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation studies on different $\\mathcal{L}$ . The results suggest that RAPO is robust and effective across various LLMs.", + "bbox": [ + 89, + 720, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "related modifiers about input prompts utilizing GPT-4 [2], and merging them into inputs at one time as the comparison of word augmentation. We randomly select one of optimized prompts as the comparison of prompt selection. The optimal result is achieved by the full-fledged framework as shown in row (f).", + "bbox": [ + 89, + 777, + 482, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation experiments on different $\\mathcal{L}$ . We conduct ablation experiments on GPT-4 [2], Mistral [22] and LLaMA", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.1 [38]. As shown in Tab. 8, although GPT-4 achieves the best overall score, the differences are marginal, which suggests that RAPO is robust and effective across various LLMs in generating optimized prompts for T2V generation.", + "bbox": [ + 511, + 508, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 580, + 784, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose RAPO, a novel framework for prompt optimizing to improve the static and dynamic dimensions performance of generated videos by pre-trained T2V models. RAPO employs word augmentation, sentence refactoring and prompt selection to generate prompts aligned with the distribution of training prompts, involving relevant and straightforward modifiers about inputs. Extensive experimental evaluations demonstrate the effectiveness and efficiency of RAPO. For the future work, we plan to extend RAPO to more T2V models and conduct more analyses about the reasons for improvements.", + "bbox": [ + 509, + 606, + 906, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 784, + 671, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work is partially supported by the National Natural Science Foundation of China (Grant No. 62471287), the National Key R&D Program of China under Grant No. 2022ZD0160102, and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800.", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3180", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Open-sora: Democratizing efficient video production for all. 2024. URL: https://github.com/hpcaitech/Open-Sora. 5, 7, 8", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5,7,8", + "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.opennai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3", + "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3:1, 2024. 1, 3", + "[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 7", + "[7] Chieh-Yun Chen, Li-Wu Tsao, Chiang Tseng, and Hong-Han Shuai. A cat is a cat (not a dog!): Unraveling information mix-ups in text-to-image encoders through causal analysis and embedding optimization. arXiv preprint arXiv:2410.00321, 2024. 1, 7", + "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3", + "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246-22256, 2023. 3", + "[10] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations. 3", + "[11] Zijie Chen, Lichao Zhang, Fangsheng Weng, Lili Pan, and Zhenzhong Lan. Tailored visions: Enhancing text-to-image generation with personalized prompt rewriting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7727-7736, 2024. 3", + "[12] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1", + "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3", + "[14] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 7", + "[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3", + "[16] Boyu Han, Qianqian Xu, Zhiyong Yang, Shilong Bao, Peisong Wen, Yangbangyan Jiang, and Qingming Huang. Aucseg: Auc-oriented pixel-level long-tail semantic segmentation. arXiv preprint arXiv:2409.20398, 2024. 7", + "[17] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3", + "[18] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. *Animate-a-story: Storytelling with retrieval-augmented video generation*. arXiv preprint arXiv:2307.06940, 2023. 3", + "[19] Nailei Hei, Qianyu Guo, Zihao Wang, Yan Wang, Haofen Wang, and Wenqiang Zhang. A user-friendly framework for generating model-preferred prompts in text-to-image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2139-2147, 2024. 1", + "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3", + "[21] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 2, 5, 6", + "[22] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6, 8", + "[23] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024.1", + "[24] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3181", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. 3", + "[25] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 2, 5", + "[26] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 1, 2, 3, 5, 7, 8", + "[27] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, Ji-Rong Wen, and Qing Yang. Dynamic prompt optimizing for text-to-image generation. In CVPR, 2024. 1, 3", + "[28] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1, 3", + "[29] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7942, 2024. 7", + "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 7", + "[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1, 3", + "[32] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. arXiv preprint arXiv:2310.15169, 2023. 1", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3, 7", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[36] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1", + "[37] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. arXiv preprint arXiv:2407.14505, 2024. 2, 5", + "[38] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5, 6, 8", + "[39] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 1", + "[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 1, 2, 3, 5, 7, 8", + "[41] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 1", + "[42] Mengjiao Yang, Yilun Du, Bo Dai, Dale Schuurmans, Joshua B Tenenbaum, and Pieter Abbeel. Probabilistic adaptation of text-to-video models. arXiv preprint arXiv:2306.01872, 2023. 3", + "[43] Shuai Yang, Jing Tan, Mengchen Zhang, Tong Wu, Yixuan Li, Gordon Wetzstein, Ziwei Liu, and Dahua Lin. Layer-pano3d: Layered 3d panorama for hyper-immersive scene generation. arXiv preprint arXiv:2408.13252, 2024. 3", + "[44] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 3", + "[45] Jingtao Zhan, Qingyao Ai, Yiqun Liu, Jia Chen, and Shaoping Ma. Capability-aware prompt reformulation learning for text-to-image generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2145-2155, 2024. 3", + "[46] Jingtao Zhan, Qingyao Ai, Yiqun Liu, Yingwei Pan, Ting Yao, Jiaxin Mao, Shaoping Ma, and Tao Mei. Prompt refinement with image pivot for text-to-image generation. arXiv preprint arXiv:2407.00247, 2024. 1", + "[47] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1-15, 2024. 3", + "[48] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 3", + "[49] Chenyi Zhuang, Ying Hu, and Pan Gao. Magnet: We never know how text-to-image diffusion models work, until we" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3182", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learn how vision-language models function. arXiv preprint arXiv:2409.19967, 2024. 7", + "[50] Shaobin Zhuang, Kunchang Li, Xinyuan Chen, Yaohui Wang, Ziwei Liu, Yu Qiao, and Yali Wang. Vlogger: Make your dream a vlog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8806-8817, 2024. 3" + ], + "bbox": [ + 91, + 90, + 483, + 189 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3183", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/bf61a946-383a-4ffe-b29d-37540410b684_model.json b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/bf61a946-383a-4ffe-b29d-37540410b684_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5085f36b9b494b02d19ec71edf99746eb0944993 --- /dev/null +++ b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/bf61a946-383a-4ffe-b29d-37540410b684_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.13, + 0.882, + 0.175 + ], + "angle": 0, + "content": "The Devil is in the Prompts: Retrieval-Augmented Prompt Optimization for Text-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.203, + 0.307, + 0.221 + ], + "angle": 0, + "content": "Bingjie Gao\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.222, + 0.337, + 0.239 + ], + "angle": 0, + "content": "Yu Qiao\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.204, + 0.464, + 0.221 + ], + "angle": 0, + "content": "Xinyu Gao\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.222, + 0.432, + 0.238 + ], + "angle": 0, + "content": "Li Niu\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.456, + 0.222, + 0.586, + 0.239 + ], + "angle": 0, + "content": "Xinyuan Chen\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.685, + 0.204, + 0.794, + 0.221 + ], + "angle": 0, + "content": "Yujie Zhou1,2" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.222, + 0.734, + 0.239 + ], + "angle": 0, + "content": "Yaohui Wang\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.24, + 0.359, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.239, + 0.875, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)Shanghai Artificial Intelligence Laboratory \\(^{3}\\)Fudan University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.324, + 0.485, + 0.701 + ], + "angle": 0, + "content": "The evolution of Text-to-video (T2V) generative models, trained on large-scale datasets, has been marked by significant progress. However, the sensitivity of T2V generative models to input prompts highlights the critical role of prompt design in influencing generative outcomes. Prior research has predominantly relied on Large Language Models (LLMs) to align user-provided prompts with the distribution of training prompts, albeit without tailored guidance encompassing prompt vocabulary and sentence structure nuances. To this end, we introduce RAPO, a novel Retrieval-Augmented Prompt Optimization framework. In order to address potential inaccuracies and ambiguous details generated by LLM-generated prompts. RAPO refines the naive prompts through dual optimization branches, selecting the superior prompt for T2V generation. The first branch augments user prompts with diverse modifiers extracted from a learned relational graph, refining them to align with the format of training prompts via a fine-tuned LLM. Conversely, the second branch rewrites the naive prompt using a pre-trained LLM following a well-defined instruction set. Extensive experiments demonstrate that RAPO can effectively enhance both the static and dynamic dimensions of generated videos, demonstrating the significance of prompt optimization for user-provided prompts. Project website: GitHub." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.732, + 0.222, + 0.747 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.879 + ], + "angle": 0, + "content": "With the rapid advancement of diffusion models [28, 33, 34], visual content creation has experienced remarkable progress in recent years. The generation of images, as well as videos from text prompts utilizing large-scale diffusion models, referred to as text-to-images (T2I) [13, 30, 35] and text-to-videos (T2V) [5, 12, 40] generation, have attracted significant interest due to the broad range of applications in real-world scenarios. Various efforts have been made to en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.905, + 0.339 + ], + "angle": 0, + "content": "hance the performance of these models, including improvements in model architecture [23, 26, 26], learning strategies [36, 44], and data curation [32, 39, 41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Recent studies [17, 31, 44] have revealed that employing long, detailed prompts with a pre-trained model typically produces superior quality outcomes compared to utilizing shallow descriptions provided by users. This has underscored the significance of prompt optimization as an important challenge in text-based visual content creation. The prompts provided by users are often brief and lack the essential details required to generate vivid images or videos. Simply attempting to optimize prompts by manually adding random descriptions can potentially mislead models and degrade the quality of generative results, resulting in outputs that may not align with user intentions. Therefore, developing automated methods to enhance user-provided prompts becomes essential for improving the overall quality of generated content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.748 + ], + "angle": 0, + "content": "Towards improving image aesthetics and ensuring semantic consistency, several attempts [19, 27, 46] have been made in previous T2I works for prompt optimization. These efforts primarily involve instructing a pre-trained or finetuned Large Language Model (LLM) to incorporate detailed modifiers into original prompts, with the aim of enhancing spatial elements such as color and relationships. While these approaches have displayed promising outcomes in image generation, studies [7, 17] reveal that their impact on video generation remains limited, especially in terms of enhancing temporal aspects such as motion smoothness and minimizing temporal flickering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.909, + 0.854 + ], + "angle": 0, + "content": "For T2V generation, we observe that the quality of spatial and temporal elements is significantly influenced by the selection of appropriate verb-object phrases and the structure of input prompts, which largely rely on analyzing and structuring the entire training data in a systematical manner. To this end, we propose RAPO, a Retrieval-Augmented Prompt Optimization framework for T2V generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The primary objective of RAPO is to convert user-provided prompts into optimized prompts that retain the original semantics while closely adhering to the format of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.235, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3173" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.907, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.551, + 0.589, + 0.565 + ], + "angle": 0, + "content": "(b) static dimension" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.582, + 0.908, + 0.613 + ], + "angle": 0, + "content": "Figure 1. Videos generated using LaVie [40] conditioned on user-provided prompts and optimized prompts from RAPO. The optimized prompts can significantly enhance both the static and dynamic qualities of the generated videos, making them more visually appealing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.637, + 0.483, + 0.743 + ], + "angle": 0, + "content": "the training data, including the vocabulary and sentence structure. To achieve this, we initially construct a Relation Graph from the training data, with each node representing words and edges indicating the relationships between neighboring words. By leveraging any user-provided prompt, we can retrieve the most relevant terms to enhance the original descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Subsequently, a specially fine-tuned Refactoring LLM is introduced to reorganize the structure and writing style of the word-augmented prompt, aligning it more cohesively with the original training data. Additionally, we employ a Rewrite LLM to directly transform user-provided prompts into optimized prompts that follow a specific format consistent with the training prompts, serving as candidates for optimized prompts. In order to address potential inaccuracies and ambiguous details generated by LLMs, we conclude by fine-tuning a Discriminator LLM to identify the most suit-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.638, + 0.733, + 0.653 + ], + "angle": 0, + "content": "able prompt for T2V generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.835 + ], + "angle": 0, + "content": "We apply RAPO on two T2V generation models LaVie [40] and Latte [26], and evaluate it on three benchmarks, namely VBench [21], EvalCrafter [25] and T2V-CompBench [37]. It is worth noticing that RAPO greatly promotes the performance of multiple objects dimension in VBench from \\(37.71\\%\\) to \\(64.86\\%\\) using LaVie and from \\(29.55\\%\\) to \\(52.78\\%\\) using Latte. In addition, the experimental results show that RAPO improves the performance of T2V models to synthesize scenes with multiple subjects, even unusual scenes like \"a mushroom growing out of a human head\" as shown in Fig. 1. Our contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.871 + ], + "angle": 0, + "content": "- We propose RAPO, a Retrieval-Augmented Prompt Optimization framework T2V generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- We propose word augmentation module and sentence refactoring module to optimize prompts considering" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3174" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.425, + 0.106 + ], + "angle": 0, + "content": "prompt vocabulary and specific sentence format." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.107, + 0.482, + 0.137 + ], + "angle": 0, + "content": "- We validate the effectiveness of RAPO through extensive experiments on several benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.153, + 0.233, + 0.168 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.179, + 0.483, + 0.42 + ], + "angle": 0, + "content": "Text-to-Video generation. With the remarkable breakthroughs of diffusion models [28, 33, 34], the generations of 3D content [9, 24, 43], images [3, 13, 30], and videos [4, 5, 20] from text descriptions achieve rapid advancement. Text-to-Video (T2V) [8, 40, 47] Generation aims to automatically create videos that match given textual descriptions. This process generally involves comprehending the scenes, objects, and actions described in the text and converting them into a sequence of cohesive visual frames, producing a video that is logically and visually consistent. T2V generation is wildly used in applications, such as animations [10, 15, 18] and automatic movie generation [42, 48, 50]. However, large T2V generative models [26, 40, 44] trained on large-scale dataset could not adequately demonstrate their potential in generation due to mismatch between training and inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.483, + 0.769 + ], + "angle": 0, + "content": "Prompt optimization. T2I and T2V generative models are sensitive to input prompts. However, the well-performed prompts are often model-specific and coherent with training prompts, misaligned with user input. Therefore, several studies [11, 17, 27, 45] are conducted to explore the generative potential of T2I and T2V generative models. Hao et al. [17] propose a learning-based prompt optimizing framework unitizing reinforce learning for generating more aesthetically pleasing images. Chen et al. [11] enhance user prompts by leveraging the user's historical interactions with the system. Mo et al. [27] propose Prompt Auto-Editing (PAE) method to decide the weights and injection time steps of each word without manual intervention. These methods primarily focus on prompts optimizing for T2I models and lack extension to T2V models. Yang et al. [44] use large language models (LLMs) to transform short prompts into more detailed ones, maintaining a consistent visual structure. Polyak et al. [31] develop a teacher-student distillation approach for prompt optimization to improve computational efficiency and reduce latency. However, the results of optimized prompts usually could not be well-aligned with training prompts due to the misleading of the LLMs and the lack of more refined guidance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.784, + 0.182, + 0.8 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, RAPO consists of three parts, 1) a word augmentation module, 2) a sentence refactoring module, as well as 3) a prompt selection module. Given a user-provided prompt \\( x_{i} \\), firstly, the word augmentation module utilizes an interactive retrieval-merge mechanism between a relation graph \\( \\mathcal{G} \\) and a LLM \\( \\mathcal{L} \\) to augment the prompt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "by adding related subject, action and atmosphere modifiers. Then, a fine-tuned LLM \\(\\mathcal{L}_r\\) is applied to refactor the entire sentence into \\(x_r\\). \\(x_r\\) has a more unified format which is consistent with the prompt length and format distribution in training data. Finally, a discriminator in the prompt selection module decides between \\(x_r\\) and a naively augmented prompt \\(x_n\\) obtained directly from a LLM via instruction, as the most suitable augmented prompt for T2V generation. We proceed to introduce each module in detail in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.252, + 0.771, + 0.267 + ], + "angle": 0, + "content": "3.1. Word Augmentation Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.394 + ], + "angle": 0, + "content": "Given a user-provided prompt \\( x_{i} \\), the word augmentation module aims to enrich \\( x_{i} \\) with more multiple, straightforward and relevant modifiers. It is achieved through retrieving modifiers meeting the requirements from a built relation graph \\( \\mathcal{G} \\), and merging them into \\( x_{i} \\) through \\( \\mathcal{L} \\) via instruction. In this section, we first introduce the construction and retrieval of relation graph. And we introduce the instruction format and retrieval-merge mechanism of \\( \\mathcal{L} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.905, + 0.59 + ], + "angle": 0, + "content": "Relation Graph \\(\\mathcal{G}\\). As shown in Fig. 3, we construct relation graph \\(\\mathcal{G}\\) based on training prompts database. For each training prompt, we utilize \\(\\mathcal{L}\\) to extract scene and corresponding related modifiers (subject, action, atmosphere descriptions). Each scene serves as a core node, with subject, action and atmosphere modifiers connected as individual sub-nodes in relation graph. For each extracted scene, we first check whether it exists in relation graph or not. If so the extracted related modifiers will be connected to the existing one. If not the extracted scene becomes a new core node with related modifiers connected. Finally, we can obtain a relation graph covering diverse scenes with multiple modifiers connected." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.904, + 0.711 + ], + "angle": 0, + "content": "For relation graph retrieval, we utilize a sentence transformer pre-trained model to extract features of prompt, and employ the cosine similarity to measure similarity between sentence features. We first retrieve the top-k relevant scenes from \\(\\mathcal{G}\\) for \\(x_{i}\\). Then we retrieve all modifiers connected to the retrieved scenes. We select the top-k relevant modifiers \\(\\{p_n|_{n = 0}^{k - 1}\\}\\) from all retrieved modifiers, preparing for the retrieval-merge mechanism of \\(\\mathcal{L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.904, + 0.803 + ], + "angle": 0, + "content": "LLM \\(\\mathcal{L}\\). We augment \\(x_{i}\\) with retrieved modifiers \\(\\{p_n|_{n = 0}^{k - 1}\\}\\) from relation graph. We rename \\(x_{i}\\) with \\(x_{i}^{0}\\) to illustrate the process of iterative merging. Specifically, the retrieved modifiers are merged into input prompt \\(x_{i}^{0}\\) one by one through prompting \\(\\mathcal{L}\\), to maintain the information of the original input while adding relevant modifiers." + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.812, + 0.905, + 0.831 + ], + "angle": 0, + "content": "\\[\nx _ {i} ^ {m + 1} = f \\left(x _ {i} ^ {m}, p _ {i} ^ {m}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(m = 0,1,\\dots,k - 1\\). \\(f\\) is a function that combine \\(x_{i}^{m}\\) and \\(p_i^m\\) reasonably by \\(\\mathcal{L}\\). For instance, a merged prompt \"a woman dressed in a black suit representing a funeral\" is resulted from merging the user-provided prompt \"a woman" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3175" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.895, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.355, + 0.908, + 0.427 + ], + "angle": 0, + "content": "Figure 2. Overview of RAPO. The naive prompt is optimized by two branches respectively. For the first branch, it is enriched by the word augmentation module based on a constructed relation graph and a frozen Large Language Model (LLM). Subsequently, augmented prompt is refactored by a finetuned LLM into a specific format in the sentence refactoring module. For the second branch, the naive prompt is directly rewritten by a frozen LLM. Finally, the prompt selection module selects the better one from two branches' results as input for T2V model." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.449, + 0.498, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.722, + 0.486, + 0.807 + ], + "angle": 0, + "content": "Figure 3. The construction of relation graph. Relation graph consists of multiple nodes (scenes acting as core nodes with modifiers connected as sub-nodes). For each prompt in database, LLM extracts scene and related modifiers. Based on whether the extracted scene is already in the graph or not, different methods are used to incorporate the new information into the graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.906 + ], + "angle": 0, + "content": "representing a funeral\" and a retrieved modifier \"a black suit\". We prompt \\(\\mathcal{L}\\) to perform general prompt merging in a normal manner as the template in Tab. 1. In instruction, we provide some prompt pairs \\(E = \\{e_i|_{i = 0}^n\\}\\) as examples, in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.452, + 0.907, + 0.483 + ], + "angle": 0, + "content": "which \\( e_i \\) contains input prompt, a modifier and corresponding merged result." + }, + { + "type": "title", + "bbox": [ + 0.524, + 0.501, + 0.864, + 0.517 + ], + "angle": 0, + "content": "LLM Template for Retrieval-Merge Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.523, + 0.898, + 0.628 + ], + "angle": 0, + "content": "Suppose you are a Text Merger. You receive two inputs from the user: a description body and a relevant modifier. Your task is to enrich the description body with relevant modifiers while retaining the description body. You should ensure that the output text is coherent, contextually relevant, and follows the same structure as the examples provided." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.628, + 0.866, + 0.645 + ], + "angle": 0, + "content": "Examples of prompt-pairs provided: \\( E = \\{e_i|_{i = 0}^n\\} \\)" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.645, + 0.862, + 0.66 + ], + "angle": 0, + "content": "Input description body and modifier are: \\(\\{x_i^m,p_i^m\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.66, + 0.741, + 0.675 + ], + "angle": 0, + "content": "The merged prompt is: \\(\\{x_i^{m + 1}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.775 + ], + "angle": 0, + "content": "Table 1. Input template for retrieval-merge mechanism. This template specifies how a frozen LLM iteratively merges user-provided prompt texts with relevant modifiers retrieved from a relation graph, thereby enriching the prompt's semantic content and aligning it with the training prompt structure for improved text-to-video synthesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.779, + 0.819 + ], + "angle": 0, + "content": "3.2. Sentence Refactoring Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Sentence refactoring module aims to refactor word augmented prompts from word augmentation module to be more consistent with prompt format in training data. It is achieved through a fine-tuned LLM \\(L_{r}\\) named as refactoring model. In this section, we introduce the training data" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3176" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.368, + 0.107 + ], + "angle": 0, + "content": "preparation and instruction tuning for \\( L_{r} \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.483, + 0.272 + ], + "angle": 0, + "content": "Data preparation. We represent the required dataset for training refactoring model by \\(\\{D_r = r_i|_{i=1}^{N^r}\\}\\), in which \\(r_i\\) involves a pair of prompts and \\(N^r\\) is the number of training prompts pairs. Specifically, \\(r_i = (w_i, c_i)\\), in which \\(w_i\\) targets to simulate world augmented prompt, and \\(c_i\\) represents the target prompt, that is, a training prompt for T2V models. \\(w_i\\) and \\(c_i\\) share similar semantics while different in the prompt format and length. Therefore, we generate \\(w_i\\) automatically through rewriting \\(c_i\\) utilizing \\(\\mathcal{L}\\) via instruction to break the unified training prompt format but maintaining the original semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.364 + ], + "angle": 0, + "content": "Instruction tuning for \\( L_{r} \\). We employ instruction tuning for fine-tuning a LLM on our constructed dataset of instructional prompts and corresponding outputs. The constructed dataset is based on \\( \\{D_r = r_i|_{i = 1}^{N^r}\\} \\) containing instructional prompts and corresponding outputs. The template of the instruction tuning dataset for \\( L_{r} \\) is shown as Tab. 2." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.381, + 0.342, + 0.397 + ], + "angle": 0, + "content": "Instruction Tuning Dataset for \\(L_{r}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.402, + 0.475, + 0.478 + ], + "angle": 0, + "content": "Instruction. Refine format and word length of the sentence: \\( w_{i} \\). Maintain the original subject descriptions, actions, scene descriptions. Append additional straightforward actions to make the sentence more dynamic if necessary." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.479, + 0.269, + 0.495 + ], + "angle": 0, + "content": "Output: target prompt \\( c_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Table 2. Instruction tuning dataset template for \\( L_{r} \\). This template directs LLM fine-tuning to restructure augmented prompts by adjusting their format while preserving semantics, aligning them with the training data's style for improved T2V generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.591, + 0.325, + 0.607 + ], + "angle": 0, + "content": "3.3. Prompt Selection Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.704 + ], + "angle": 0, + "content": "As shown in Fig. 2, prompt selection module contains a fine-tuned LLM \\(\\mathcal{L}_d\\) named prompt discriminator to select the better one between \\(x_{r}\\) from sentence refactoring module, and a naively augmented prompt \\(x_{n}\\) obtained directly from a LLM via instruction. In this section, we introduce the training data preparation and instruction tuning for \\(\\mathcal{L}_d\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data preparation. We represent the required dataset for training refactoring model by \\(\\{D_{d} = d_{i}|_{i = 1}^{N^{d}}\\}\\), in which \\(d_{i}\\) contains three prompts and \\(N^d\\) is the number of training prompts triples. Specifically, \\(d_{i} = (x_{i},x_{r},x_{n},y_{d})\\), in which \\(y_{d}\\) represents the discriminator label to select the better one for T2V generation from \\(x_{r}\\) and \\(x_{n}\\) given input prompt \\(x_{i}\\). To simulate the user-provided prompts, we collect diverse prompts from several T2V benchmarks and generate more utilizing \\(\\mathcal{L}\\) via instruction. \\(x_{r}\\) and \\(x_{n}\\) can be obtained from the proposed RAPO as shown in Fig. 2 given \\(x_{i}\\). We determine \\(y_{d}\\) through the evaluation of generated videos conditioned on \\(x_{r}\\) and \\(x_{n}\\). Specifically, the evaluations of T2V models performance involves diverse" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.167 + ], + "angle": 0, + "content": "dimensions. For collected or generated prompts, we need to determine the evaluation dimension according to prompt content. We automatically decide the evaluation dimension of input prompts utilizing \\(\\mathcal{L}\\), then choose the corresponding metrics to evaluate generated videos." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.907, + 0.229 + ], + "angle": 0, + "content": "Instruction tuning for \\( L_{d} \\). Similar to \\( L_{r} \\), we employ instruction tuning for \\( L_{d} \\) based on \\( \\{D_{d} = d_{i}|_{i = 1}^{N^{d}}\\} \\). The template of the instruction tuning dataset for \\( L_{d} \\) is shown as Tab. 3." + }, + { + "type": "title", + "bbox": [ + 0.525, + 0.244, + 0.766, + 0.26 + ], + "angle": 0, + "content": "Instruction Tuning Dataset for \\(L_{d}\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.266, + 0.897, + 0.341 + ], + "angle": 0, + "content": "Instruction. Given user-provided prompt \\( x_{i} \\), select the better optimized prompt from \\( x_{r} \\) and \\( x_{n} \\). The chosen prompt is required to contain multiple, straightforward, and relevant modifiers about \\( x_{i} \\) while involving the semantics of \\( x_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.342, + 0.732, + 0.358 + ], + "angle": 0, + "content": "Output: discriminator label \\( y_{d} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.372, + 0.907, + 0.429 + ], + "angle": 0, + "content": "Table 3. Instruction tuning dataset template for \\( L_{d} \\). This template aims to train a discriminator LLM that evaluates multiple refined prompts and selects the optimal one based on the inclusion of clear, straightforward modifiers and faithful semantic alignment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.46, + 0.649, + 0.478 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.485, + 0.704, + 0.502 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Evaluation and Metrics. We utilize the VBenchmark [21], EvalCrafter [25] and T2V-CompBench [37] for quantitative performance evaluation. VBenchmark is an extensive benchmark comprising 16 fine-grained dimensions designed to systematically evaluate the quality of generated videos. EvalCrafter is a comprehensive evaluation benchmark that includes approximately 17 objective metrics for evaluating video generation performance. T2V-CompBench is the first benchmark tailored for compositional text-to-video generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Comparison to other methods. We compare the optimized prompts with our method to three types of prompts: the short primary prompts, the prompts generated from GPT-4 [2] and prompt refiner from Open-sora [1], which is based on fine-tuned LLaMA 3.1 [38]. The specific inference prompt template for GPT-4 [2] can be found in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.733, + 0.79 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "The well-performed prompts are model-specific and aligned with the distribution of training prompts. We employ Vimeo25M [40], a training dataset consisting of 25 million text-video pairs as our analysis dataset. At the same time, we choose LaVie [40] and Latte [26] as analysis T2V models, which belong to the diffusion-based and DiT architectures respectively and use Vimeo25M as one of training" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3177" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.905, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.279, + 0.736, + 0.294 + ], + "angle": 0, + "content": "(a) Static dimension. From left to right: naive, GPT-4, Open-sora, Ours." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.301, + 0.9, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.465, + 0.858, + 0.48 + ], + "angle": 0, + "content": "(b) Dynamic dimension (a robot is dancing). From left to right, from top to bottom: naive, GPT-4, Open-sora, Ours." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.495, + 0.908, + 0.54 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons across dynamic and static dimensions. This figure showcases videos generated using LaVie with short prompts, GPT-4 and Open-sora prompt optimizations, and our RAPO method. Videos produced with RAPO exhibit significantly sharper spatial details, smoother temporal transitions, and a closer semantic alignment with the input text." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.486, + 0.837 + ], + "angle": 0, + "content": "datasets. And we represent some extension results to some other T2V models in supplementary materials. For relation graph construction, we utilize Mistral [22] to extract scenes with corresponding subject, action and atmosphere descriptions from Vimeo25M dataset, and use all-MiniLM-L6-v2 as sentence transformer pre-trained model. We filter about 2.1M valid sentences from Vimeo25M dataset. For refactoring model training data, we prepare about 86k prompt-pairs following data preparation method in Section 3.2. For prompt discriminator training data, we first generate 7K text captions using Mistral, covering all the dimensions in VBenchmark [21]. The specific prompt template for extraction and dataset construction can be found in supplementary materials. We perform LoRA fine-tuning using LLaMA 3.1 [38], and fine-tune 8 epochs and 3 epochs for refactoring model and prompt discriminator respectively with a single A100, using a batch size of 32 and a LoRA rank of 64." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.273, + 0.863 + ], + "angle": 0, + "content": "4.3. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Qualitative comparisons. As shown in Fig. 4, compared with user-provided prompts and other prompt optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.656 + ], + "angle": 0, + "content": "methods, RAPO centers the input semantics and promotes the quality of generated videos from static and dynamic dimensions. The relevant modifiers promote model comprehending the simple user-provided prompts. The detailed prompts of different methods and more qualitative comparisons are provided in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparisons. As shown in Tab. 4 and Tab. 5, the results show that RAPO surpasses other methods in static dimensions (e.g., visual quality, object class) and dynamic dimensions (e.g., human actions, temporal flickering). The consistent performance across various benchmarks demonstrates the robustness and versatility of RAPO. Although the optimized prompts from GPT-4 and Open-sora enriches the plain user-provided prompts with more details describing the scenes, objects, and actions, these lengthy and complex descriptions may confuse models and even worsen the generation. It is worth noticing that RAPO significantly enhances the score of multiple objects dimension from \\(37.71\\%\\) to \\(64.86\\%\\) using LaVie and from \\(29.55\\%\\) to \\(52.78\\%\\) using Latte as shown in Tab. 4. And RAPO outperforms other methods on T2V-CompBench as shown in Tab. 6 for some selected dimensions, which are focused on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3178" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.213 + ], + "angle": 0, + "content": "
ModelsTotal Scoretemporal flickeringimaging qualityhuman actionobject classmultiple objectsspatial relationship
LaVie [40]80.89%96.62%69.00%95.80%92.09%37.71%37.27%
LaVie-GPT4 [2]79.69%96.14%70.27%83.80%88.73%36.23%50.55%
LaVie-Open-sora [1]79.75%96.42%70.42%87.00%91.29%36.52%54.37%
LaVie-RAPO82.38%96.86%71.40%96.80%96.91%64.86%59.15%
Latte [26]77.03%97.10%63.38%88.40%83.86%29.55%40.63%
Latte-GPT4 [2]77.40%97.52%63.54%85.80%78.32%27.73%36.72%
Latte-Open-sora [1]77.23%97.67%64.19%84.60%83.60%30.00%35.12%
Latte-RAPO79.97%98.17%66.72%95.20%96.47%52.78%41.31%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.224, + 0.907, + 0.252 + ], + "angle": 0, + "content": "Table 4. Quantitative comparisons on VBench. RAPO achieves the highest overall scores, especially in static attributes and multiple object generation." + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.276, + 0.287, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.424, + 0.252, + 0.435 + ], + "angle": 0, + "content": "a bicycle and a car" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.276, + 0.472, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.418, + 0.438, + 0.439 + ], + "angle": 0, + "content": "A bicycle and a car are parked side by side." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.44, + 0.195, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.512, + 0.174, + 0.523 + ], + "angle": 0, + "content": "bicycle" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.44, + 0.288, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.512, + 0.255, + 0.522 + ], + "angle": 0, + "content": "car" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.44, + 0.379, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.512, + 0.354, + 0.523 + ], + "angle": 0, + "content": "bicycle" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.44, + 0.472, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.512, + 0.435, + 0.522 + ], + "angle": 0, + "content": "car" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.54, + 0.483, + 0.582 + ], + "angle": 0, + "content": "Figure 5. Visualization on attention map on multiple objects from different prompts. Adding description of the relative spatial position between objects can improve multi-object generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.656 + ], + "angle": 0, + "content": "complex compositional T2V tasks. These experiments verified that PAPO significantly improves the performance of generating scenes involving more than two subjects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.197, + 0.683 + ], + "angle": 0, + "content": "4.4. Analyses" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Multiple objects. Synthesis quality of generated videos often declines when tasked with generating outputs that accurately represent prompts involving multiple objects. This issue is also prevalent in the T2I model, and several studies [14, 16, 30] have highlighted that the blended context created by the CLIP text encoder leads to improper binding. Meanwhile, some related works [6, 29] focus on image latents to address information loss, while the others [7, 49] pay more attention to text embedding to deal with the issue. However, few have explored optimizing prompts to improve the performance of multiple obejsts task. We apply our method to text-to-image using SD 1.4 [34], which uses the same text encoder with LaVie [40]. We test on prompts about multiple objects, and remove the irrelevant modifiers" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.278, + 0.871, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.447, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Figure 6. Prompt length distribution comparison among various methods. The distribution of RAPO-optimized prompts is more closer to the training prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.906, + 0.578 + ], + "angle": 0, + "content": "like action and atmosphere descriptions. As shown in Fig. 5, we can find the relevant spatial descriptions boost the performance of multiple objects. More examples can be found in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Statistical analysis of text. As shown in Fig. 6, we compared the word length distributions of prompts from the T2V training set, user prompts (simulated via VBench, EvalCrafter, and T2V-CompBench), and optimized prompts generated by various methods. The results show that the prompt length distribution produced by RAPO is closest to that of the training set, and this consistency unleashes the model's generative potential to produce better videos. In contrast, user prompts are too short and lack necessary details, while other methods generate longer prompts that contain excessive details and complex vocabulary, which may be counterproductive, as shown in Tab. 4 and Tab. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.667, + 0.788 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.884 + ], + "angle": 0, + "content": "We conduct ablation experiments on the VBench benchmark to examine the individual and combined effects of different modules in RAPO. Additionally, we perform ablation experiments on various configurations of \\(\\mathcal{L}\\). We present more ablation experiments on hyperparameters in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Ablating each modules in RAPO. We directly obtain the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3179" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.238 + ], + "angle": 0, + "content": "
ModelsFinal Sum Scoremotion qualitytext-video alignmentvisual qualitytemporal consistency
LaVie [40]24853.1969.6064.8160.87
LaVie-GPT4 [2]24654.0565.5164.9661.22
LaVie-Open-sora [1]25153.0771.3865.2661.41
LaVie-RAPO25653.3474.3866.6261.29
Latte [26]21750.0355.4957.6553.94
Latte-GPT4 [2]21851.3653.6558.0254.65
Latte-Open-sora [1]22250.2557.3258.7155.47
Latte-RAPO22751.7360.8659.2455.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.247, + 0.907, + 0.276 + ], + "angle": 0, + "content": "Table 5. Quantitative comparisons on EvalCrafter. The results demonstrate that RAPO consistently achieves higher scores, especially in terms of text-video alignment and visual quality, indicating its superior ability to enhance the overall quality of generated videos." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.29, + 0.903, + 0.445 + ], + "angle": 0, + "content": "
Modelsconsistent attribute bindingdynamic attribute bindingaction bindingobject interactions
LaVie [40]0.6200.2320.4830.760
LaVie-GPT4 [2]0.5610.2180.4280.620
LaVie-Open-sora [1]0.5320.2140.4700.698
LaVie-RAPO0.6920.2670.6350.839
Latte [26]0.6330.2270.4760.792
Latte-GPT4 [2]0.5980.2100.4050.688
Latte-Open-sora [1]0.5490.2030.4870.743
Latte-RAPO0.7060.2580.5910.856
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.455, + 0.907, + 0.484 + ], + "angle": 0, + "content": "Table 6. Quantitative comparisons on T2V-CompBench. RAPO outperforms alternative prompt optimization methods, exhibiting better performance at managing complex compositional tasks and ensuring coherent interactions between multiple objects in the generated videos." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.506, + 0.481, + 0.584 + ], + "angle": 0, + "content": "
word augmentationsentence refactoringprompt selectionVBatch Total Score
(a)✓80.37%
(b)✓79.75%
(c)✓✓81.58%
(d)✓✓81.75%
(e)✓✓80.60%
(f)✓✓✓82.38%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Table 7. Ablation studies of different modules in RAPO on VBench. Each module improves performance, while the combined use of all three leads to the highest evaluation score, confirming the synergistic effect of the full RAPO framework." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.666, + 0.481, + 0.713 + ], + "angle": 0, + "content": "
GPT-4MistralLLaMA
VBenchmark Total Score82.38%82.25%82.10%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.722, + 0.483, + 0.751 + ], + "angle": 0, + "content": "Table 8. Ablation studies on different \\(\\mathcal{L}\\). The results suggest that RAPO is robust and effective across various LLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.869 + ], + "angle": 0, + "content": "related modifiers about input prompts utilizing GPT-4 [2], and merging them into inputs at one time as the comparison of word augmentation. We randomly select one of optimized prompts as the comparison of prompt selection. The optimal result is achieved by the full-fledged framework as shown in row (f)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ablation experiments on different \\(\\mathcal{L}\\). We conduct ablation experiments on GPT-4 [2], Mistral [22] and LLaMA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.572 + ], + "angle": 0, + "content": "3.1 [38]. As shown in Tab. 8, although GPT-4 achieves the best overall score, the differences are marginal, which suggests that RAPO is robust and effective across various LLMs in generating optimized prompts for T2V generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.582, + 0.785, + 0.598 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.607, + 0.907, + 0.775 + ], + "angle": 0, + "content": "In this paper, we propose RAPO, a novel framework for prompt optimizing to improve the static and dynamic dimensions performance of generated videos by pre-trained T2V models. RAPO employs word augmentation, sentence refactoring and prompt selection to generate prompts aligned with the distribution of training prompts, involving relevant and straightforward modifiers about inputs. Extensive experimental evaluations demonstrate the effectiveness and efficiency of RAPO. For the future work, we plan to extend RAPO to more T2V models and conduct more analyses about the reasons for improvements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.673, + 0.802 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The work is partially supported by the National Natural Science Foundation of China (Grant No. 62471287), the National Key R&D Program of China under Grant No. 2022ZD0160102, and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3180" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.155 + ], + "angle": 0, + "content": "[1] Open-sora: Democratizing efficient video production for all. 2024. URL: https://github.com/hpcaitech/Open-Sora. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.opennai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3:1, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.494 + ], + "angle": 0, + "content": "[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Chieh-Yun Chen, Li-Wu Tsao, Chiang Tseng, and Hong-Han Shuai. A cat is a cat (not a dog!): Unraveling information mix-ups in text-to-image encoders through causal analysis and embedding optimization. arXiv preprint arXiv:2410.00321, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246-22256, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[11] Zijie Chen, Lichao Zhang, Fangsheng Weng, Lili Pan, and Zhenzhong Lan. Tailored visions: Enhancing text-to-image generation with personalized prompt rewriting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7727-7736, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[14] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.413 + ], + "angle": 0, + "content": "[16] Boyu Han, Qianqian Xu, Zhiyong Yang, Shilong Bao, Peisong Wen, Yangbangyan Jiang, and Qingming Huang. Aucseg: Auc-oriented pixel-level long-tail semantic segmentation. arXiv preprint arXiv:2409.20398, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[17] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[18] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. *Animate-a-story: Storytelling with retrieval-augmented video generation*. arXiv preprint arXiv:2307.06940, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[19] Nailei Hei, Qianyu Guo, Zihao Wang, Yan Wang, Haofen Wang, and Wenqiang Zhang. A user-friendly framework for generating model-preferred prompts in text-to-image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2139-2147, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.65, + 0.907, + 0.734 + ], + "angle": 0, + "content": "[21] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[22] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3181" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.148 + ], + "angle": 0, + "content": "Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[25] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[26] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[27] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, Ji-Rong Wen, and Qing Yang. Dynamic prompt optimizing for text-to-image generation. In CVPR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[28] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[29] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7942, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[32] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. arXiv preprint arXiv:2310.15169, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[36] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[37] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. arXiv preprint arXiv:2407.14505, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[38] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[39] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[41] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[42] Mengjiao Yang, Yilun Du, Bo Dai, Dale Schuurmans, Joshua B Tenenbaum, and Pieter Abbeel. Probabilistic adaptation of text-to-video models. arXiv preprint arXiv:2306.01872, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[43] Shuai Yang, Jing Tan, Mengchen Zhang, Tong Wu, Yixuan Li, Gordon Wetzstein, Ziwei Liu, and Dahua Lin. Layer-pano3d: Layered 3d panorama for hyper-immersive scene generation. arXiv preprint arXiv:2408.13252, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[44] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[45] Jingtao Zhan, Qingyao Ai, Yiqun Liu, Jia Chen, and Shaoping Ma. Capability-aware prompt reformulation learning for text-to-image generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2145-2155, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[46] Jingtao Zhan, Qingyao Ai, Yiqun Liu, Yingwei Pan, Ting Yao, Jiaxin Mao, Shaoping Ma, and Tao Mei. Prompt refinement with image pivot for text-to-image generation. arXiv preprint arXiv:2407.00247, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[47] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1-15, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[48] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[49] Chenyi Zhuang, Ying Hu, and Pan Gao. Magnet: We never know how text-to-image diffusion models work, until we" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3182" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "learn how vision-language models function. arXiv preprint arXiv:2409.19967, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[50] Shaobin Zhuang, Kunchang Li, Xinyuan Chen, Yaohui Wang, Ziwei Liu, Yu Qiao, and Yali Wang. Vlogger: Make your dream a vlog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8806-8817, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "3183" + } + ] +] \ No newline at end of file diff --git a/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/bf61a946-383a-4ffe-b29d-37540410b684_origin.pdf b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/bf61a946-383a-4ffe-b29d-37540410b684_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..778e02303b362b23cec31c1615da8d6d269b44db --- /dev/null +++ b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/bf61a946-383a-4ffe-b29d-37540410b684_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5543987eb226a76e1d9d28debcd0a52e8d56a99b2327d8e645f9fb478ae8a7d +size 1488284 diff --git a/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/full.md b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2e0530d1b5197db816a6ac50f32436c4e15cf262 --- /dev/null +++ b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/full.md @@ -0,0 +1,304 @@ +# The Devil is in the Prompts: Retrieval-Augmented Prompt Optimization for Text-to-Video Generation + +Bingjie Gao $^{1,2}$ + +Yu Qiao $^{2\dagger}$ + +Xinyu Gao $^{1,2}$ + +Li Niu $^{1\dagger}$ + +Xinyuan Chen $^{2\dagger}$ + +Yujie Zhou1,2 + +Yaohui Wang $^{2\dagger}$ + +$^{1}$ Shanghai Jiao Tong University + +$^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ Fudan University + +# Abstract + +The evolution of Text-to-video (T2V) generative models, trained on large-scale datasets, has been marked by significant progress. However, the sensitivity of T2V generative models to input prompts highlights the critical role of prompt design in influencing generative outcomes. Prior research has predominantly relied on Large Language Models (LLMs) to align user-provided prompts with the distribution of training prompts, albeit without tailored guidance encompassing prompt vocabulary and sentence structure nuances. To this end, we introduce RAPO, a novel Retrieval-Augmented Prompt Optimization framework. In order to address potential inaccuracies and ambiguous details generated by LLM-generated prompts. RAPO refines the naive prompts through dual optimization branches, selecting the superior prompt for T2V generation. The first branch augments user prompts with diverse modifiers extracted from a learned relational graph, refining them to align with the format of training prompts via a fine-tuned LLM. Conversely, the second branch rewrites the naive prompt using a pre-trained LLM following a well-defined instruction set. Extensive experiments demonstrate that RAPO can effectively enhance both the static and dynamic dimensions of generated videos, demonstrating the significance of prompt optimization for user-provided prompts. Project website: GitHub. + +# 1. Introduction + +With the rapid advancement of diffusion models [28, 33, 34], visual content creation has experienced remarkable progress in recent years. The generation of images, as well as videos from text prompts utilizing large-scale diffusion models, referred to as text-to-images (T2I) [13, 30, 35] and text-to-videos (T2V) [5, 12, 40] generation, have attracted significant interest due to the broad range of applications in real-world scenarios. Various efforts have been made to en + +hance the performance of these models, including improvements in model architecture [23, 26, 26], learning strategies [36, 44], and data curation [32, 39, 41]. + +Recent studies [17, 31, 44] have revealed that employing long, detailed prompts with a pre-trained model typically produces superior quality outcomes compared to utilizing shallow descriptions provided by users. This has underscored the significance of prompt optimization as an important challenge in text-based visual content creation. The prompts provided by users are often brief and lack the essential details required to generate vivid images or videos. Simply attempting to optimize prompts by manually adding random descriptions can potentially mislead models and degrade the quality of generative results, resulting in outputs that may not align with user intentions. Therefore, developing automated methods to enhance user-provided prompts becomes essential for improving the overall quality of generated content. + +Towards improving image aesthetics and ensuring semantic consistency, several attempts [19, 27, 46] have been made in previous T2I works for prompt optimization. These efforts primarily involve instructing a pre-trained or finetuned Large Language Model (LLM) to incorporate detailed modifiers into original prompts, with the aim of enhancing spatial elements such as color and relationships. While these approaches have displayed promising outcomes in image generation, studies [7, 17] reveal that their impact on video generation remains limited, especially in terms of enhancing temporal aspects such as motion smoothness and minimizing temporal flickering. + +For T2V generation, we observe that the quality of spatial and temporal elements is significantly influenced by the selection of appropriate verb-object phrases and the structure of input prompts, which largely rely on analyzing and structuring the entire training data in a systematical manner. To this end, we propose RAPO, a Retrieval-Augmented Prompt Optimization framework for T2V generation. + +The primary objective of RAPO is to convert user-provided prompts into optimized prompts that retain the original semantics while closely adhering to the format of + +![](images/6fff60468cf1a329b4b3026eb00d3ad87bfb7cb3e11ed85bd2c07fb86bc9a501.jpg) +(b) static dimension +Figure 1. Videos generated using LaVie [40] conditioned on user-provided prompts and optimized prompts from RAPO. The optimized prompts can significantly enhance both the static and dynamic qualities of the generated videos, making them more visually appealing. + +the training data, including the vocabulary and sentence structure. To achieve this, we initially construct a Relation Graph from the training data, with each node representing words and edges indicating the relationships between neighboring words. By leveraging any user-provided prompt, we can retrieve the most relevant terms to enhance the original descriptions. + +Subsequently, a specially fine-tuned Refactoring LLM is introduced to reorganize the structure and writing style of the word-augmented prompt, aligning it more cohesively with the original training data. Additionally, we employ a Rewrite LLM to directly transform user-provided prompts into optimized prompts that follow a specific format consistent with the training prompts, serving as candidates for optimized prompts. In order to address potential inaccuracies and ambiguous details generated by LLMs, we conclude by fine-tuning a Discriminator LLM to identify the most suit- + +able prompt for T2V generation. + +We apply RAPO on two T2V generation models LaVie [40] and Latte [26], and evaluate it on three benchmarks, namely VBench [21], EvalCrafter [25] and T2V-CompBench [37]. It is worth noticing that RAPO greatly promotes the performance of multiple objects dimension in VBench from $37.71\%$ to $64.86\%$ using LaVie and from $29.55\%$ to $52.78\%$ using Latte. In addition, the experimental results show that RAPO improves the performance of T2V models to synthesize scenes with multiple subjects, even unusual scenes like "a mushroom growing out of a human head" as shown in Fig. 1. Our contributions can be summarized as follows. + +- We propose RAPO, a Retrieval-Augmented Prompt Optimization framework T2V generation. +- We propose word augmentation module and sentence refactoring module to optimize prompts considering + +prompt vocabulary and specific sentence format. + +- We validate the effectiveness of RAPO through extensive experiments on several benchmarks. + +# 2. Related Work + +Text-to-Video generation. With the remarkable breakthroughs of diffusion models [28, 33, 34], the generations of 3D content [9, 24, 43], images [3, 13, 30], and videos [4, 5, 20] from text descriptions achieve rapid advancement. Text-to-Video (T2V) [8, 40, 47] Generation aims to automatically create videos that match given textual descriptions. This process generally involves comprehending the scenes, objects, and actions described in the text and converting them into a sequence of cohesive visual frames, producing a video that is logically and visually consistent. T2V generation is wildly used in applications, such as animations [10, 15, 18] and automatic movie generation [42, 48, 50]. However, large T2V generative models [26, 40, 44] trained on large-scale dataset could not adequately demonstrate their potential in generation due to mismatch between training and inference. + +Prompt optimization. T2I and T2V generative models are sensitive to input prompts. However, the well-performed prompts are often model-specific and coherent with training prompts, misaligned with user input. Therefore, several studies [11, 17, 27, 45] are conducted to explore the generative potential of T2I and T2V generative models. Hao et al. [17] propose a learning-based prompt optimizing framework unitizing reinforce learning for generating more aesthetically pleasing images. Chen et al. [11] enhance user prompts by leveraging the user's historical interactions with the system. Mo et al. [27] propose Prompt Auto-Editing (PAE) method to decide the weights and injection time steps of each word without manual intervention. These methods primarily focus on prompts optimizing for T2I models and lack extension to T2V models. Yang et al. [44] use large language models (LLMs) to transform short prompts into more detailed ones, maintaining a consistent visual structure. Polyak et al. [31] develop a teacher-student distillation approach for prompt optimization to improve computational efficiency and reduce latency. However, the results of optimized prompts usually could not be well-aligned with training prompts due to the misleading of the LLMs and the lack of more refined guidance. + +# 3. Method + +As illustrated in Fig. 2, RAPO consists of three parts, 1) a word augmentation module, 2) a sentence refactoring module, as well as 3) a prompt selection module. Given a user-provided prompt $x_{i}$ , firstly, the word augmentation module utilizes an interactive retrieval-merge mechanism between a relation graph $\mathcal{G}$ and a LLM $\mathcal{L}$ to augment the prompt + +by adding related subject, action and atmosphere modifiers. Then, a fine-tuned LLM $\mathcal{L}_r$ is applied to refactor the entire sentence into $x_r$ . $x_r$ has a more unified format which is consistent with the prompt length and format distribution in training data. Finally, a discriminator in the prompt selection module decides between $x_r$ and a naively augmented prompt $x_n$ obtained directly from a LLM via instruction, as the most suitable augmented prompt for T2V generation. We proceed to introduce each module in detail in the following sections. + +# 3.1. Word Augmentation Module + +Given a user-provided prompt $x_{i}$ , the word augmentation module aims to enrich $x_{i}$ with more multiple, straightforward and relevant modifiers. It is achieved through retrieving modifiers meeting the requirements from a built relation graph $\mathcal{G}$ , and merging them into $x_{i}$ through $\mathcal{L}$ via instruction. In this section, we first introduce the construction and retrieval of relation graph. And we introduce the instruction format and retrieval-merge mechanism of $\mathcal{L}$ . + +Relation Graph $\mathcal{G}$ . As shown in Fig. 3, we construct relation graph $\mathcal{G}$ based on training prompts database. For each training prompt, we utilize $\mathcal{L}$ to extract scene and corresponding related modifiers (subject, action, atmosphere descriptions). Each scene serves as a core node, with subject, action and atmosphere modifiers connected as individual sub-nodes in relation graph. For each extracted scene, we first check whether it exists in relation graph or not. If so the extracted related modifiers will be connected to the existing one. If not the extracted scene becomes a new core node with related modifiers connected. Finally, we can obtain a relation graph covering diverse scenes with multiple modifiers connected. + +For relation graph retrieval, we utilize a sentence transformer pre-trained model to extract features of prompt, and employ the cosine similarity to measure similarity between sentence features. We first retrieve the top-k relevant scenes from $\mathcal{G}$ for $x_{i}$ . Then we retrieve all modifiers connected to the retrieved scenes. We select the top-k relevant modifiers $\{p_n|_{n = 0}^{k - 1}\}$ from all retrieved modifiers, preparing for the retrieval-merge mechanism of $\mathcal{L}$ . + +LLM $\mathcal{L}$ . We augment $x_{i}$ with retrieved modifiers $\{p_n|_{n = 0}^{k - 1}\}$ from relation graph. We rename $x_{i}$ with $x_{i}^{0}$ to illustrate the process of iterative merging. Specifically, the retrieved modifiers are merged into input prompt $x_{i}^{0}$ one by one through prompting $\mathcal{L}$ , to maintain the information of the original input while adding relevant modifiers. + +$$ +x _ {i} ^ {m + 1} = f \left(x _ {i} ^ {m}, p _ {i} ^ {m}\right), \tag {1} +$$ + +where $m = 0,1,\dots,k - 1$ . $f$ is a function that combine $x_{i}^{m}$ and $p_i^m$ reasonably by $\mathcal{L}$ . For instance, a merged prompt "a woman dressed in a black suit representing a funeral" is resulted from merging the user-provided prompt "a woman + +![](images/30c2cf4568e1472d731af966eb03ac0fc3d59f023ea6e4cabbaa0f9f69a4d329.jpg) +Figure 2. Overview of RAPO. The naive prompt is optimized by two branches respectively. For the first branch, it is enriched by the word augmentation module based on a constructed relation graph and a frozen Large Language Model (LLM). Subsequently, augmented prompt is refactored by a finetuned LLM into a specific format in the sentence refactoring module. For the second branch, the naive prompt is directly rewritten by a frozen LLM. Finally, the prompt selection module selects the better one from two branches' results as input for T2V model. + +![](images/ea314a54bd93ade58abafa5bff17cac48974326447a62928f4f8eb1437b8aaff.jpg) +Figure 3. The construction of relation graph. Relation graph consists of multiple nodes (scenes acting as core nodes with modifiers connected as sub-nodes). For each prompt in database, LLM extracts scene and related modifiers. Based on whether the extracted scene is already in the graph or not, different methods are used to incorporate the new information into the graph. + +representing a funeral" and a retrieved modifier "a black suit". We prompt $\mathcal{L}$ to perform general prompt merging in a normal manner as the template in Tab. 1. In instruction, we provide some prompt pairs $E = \{e_i|_{i = 0}^n\}$ as examples, in + +which $e_i$ contains input prompt, a modifier and corresponding merged result. + +# LLM Template for Retrieval-Merge Mechanism + +Suppose you are a Text Merger. You receive two inputs from the user: a description body and a relevant modifier. Your task is to enrich the description body with relevant modifiers while retaining the description body. You should ensure that the output text is coherent, contextually relevant, and follows the same structure as the examples provided. + +Examples of prompt-pairs provided: $E = \{e_i|_{i = 0}^n\}$ + +Input description body and modifier are: $\{x_i^m,p_i^m\}$ + +The merged prompt is: $\{x_i^{m + 1}\}$ + +Table 1. Input template for retrieval-merge mechanism. This template specifies how a frozen LLM iteratively merges user-provided prompt texts with relevant modifiers retrieved from a relation graph, thereby enriching the prompt's semantic content and aligning it with the training prompt structure for improved text-to-video synthesis. + +# 3.2. Sentence Refactoring Module + +Sentence refactoring module aims to refactor word augmented prompts from word augmentation module to be more consistent with prompt format in training data. It is achieved through a fine-tuned LLM $L_{r}$ named as refactoring model. In this section, we introduce the training data + +preparation and instruction tuning for $L_{r}$ . + +Data preparation. We represent the required dataset for training refactoring model by $\{D_r = r_i|_{i=1}^{N^r}\}$ , in which $r_i$ involves a pair of prompts and $N^r$ is the number of training prompts pairs. Specifically, $r_i = (w_i, c_i)$ , in which $w_i$ targets to simulate world augmented prompt, and $c_i$ represents the target prompt, that is, a training prompt for T2V models. $w_i$ and $c_i$ share similar semantics while different in the prompt format and length. Therefore, we generate $w_i$ automatically through rewriting $c_i$ utilizing $\mathcal{L}$ via instruction to break the unified training prompt format but maintaining the original semantics. + +Instruction tuning for $L_{r}$ . We employ instruction tuning for fine-tuning a LLM on our constructed dataset of instructional prompts and corresponding outputs. The constructed dataset is based on $\{D_r = r_i|_{i = 1}^{N^r}\}$ containing instructional prompts and corresponding outputs. The template of the instruction tuning dataset for $L_{r}$ is shown as Tab. 2. + +# Instruction Tuning Dataset for $L_{r}$ + +Instruction. Refine format and word length of the sentence: $w_{i}$ . Maintain the original subject descriptions, actions, scene descriptions. Append additional straightforward actions to make the sentence more dynamic if necessary. + +Output: target prompt $c_{i}$ . + +Table 2. Instruction tuning dataset template for $L_{r}$ . This template directs LLM fine-tuning to restructure augmented prompts by adjusting their format while preserving semantics, aligning them with the training data's style for improved T2V generation. + +# 3.3. Prompt Selection Module + +As shown in Fig. 2, prompt selection module contains a fine-tuned LLM $\mathcal{L}_d$ named prompt discriminator to select the better one between $x_{r}$ from sentence refactoring module, and a naively augmented prompt $x_{n}$ obtained directly from a LLM via instruction. In this section, we introduce the training data preparation and instruction tuning for $\mathcal{L}_d$ . + +Data preparation. We represent the required dataset for training refactoring model by $\{D_{d} = d_{i}|_{i = 1}^{N^{d}}\}$ , in which $d_{i}$ contains three prompts and $N^d$ is the number of training prompts triples. Specifically, $d_{i} = (x_{i},x_{r},x_{n},y_{d})$ , in which $y_{d}$ represents the discriminator label to select the better one for T2V generation from $x_{r}$ and $x_{n}$ given input prompt $x_{i}$ . To simulate the user-provided prompts, we collect diverse prompts from several T2V benchmarks and generate more utilizing $\mathcal{L}$ via instruction. $x_{r}$ and $x_{n}$ can be obtained from the proposed RAPO as shown in Fig. 2 given $x_{i}$ . We determine $y_{d}$ through the evaluation of generated videos conditioned on $x_{r}$ and $x_{n}$ . Specifically, the evaluations of T2V models performance involves diverse + +dimensions. For collected or generated prompts, we need to determine the evaluation dimension according to prompt content. We automatically decide the evaluation dimension of input prompts utilizing $\mathcal{L}$ , then choose the corresponding metrics to evaluate generated videos. + +Instruction tuning for $L_{d}$ . Similar to $L_{r}$ , we employ instruction tuning for $L_{d}$ based on $\{D_{d} = d_{i}|_{i = 1}^{N^{d}}\}$ . The template of the instruction tuning dataset for $L_{d}$ is shown as Tab. 3. + +# Instruction Tuning Dataset for $L_{d}$ + +Instruction. Given user-provided prompt $x_{i}$ , select the better optimized prompt from $x_{r}$ and $x_{n}$ . The chosen prompt is required to contain multiple, straightforward, and relevant modifiers about $x_{i}$ while involving the semantics of $x_{i}$ . + +Output: discriminator label $y_{d}$ . + +Table 3. Instruction tuning dataset template for $L_{d}$ . This template aims to train a discriminator LLM that evaluates multiple refined prompts and selects the optimal one based on the inclusion of clear, straightforward modifiers and faithful semantic alignment. + +# 4. Experiments + +# 4.1. Experimental Setup + +Evaluation and Metrics. We utilize the VBenchmark [21], EvalCrafter [25] and T2V-CompBench [37] for quantitative performance evaluation. VBenchmark is an extensive benchmark comprising 16 fine-grained dimensions designed to systematically evaluate the quality of generated videos. EvalCrafter is a comprehensive evaluation benchmark that includes approximately 17 objective metrics for evaluating video generation performance. T2V-CompBench is the first benchmark tailored for compositional text-to-video generation. + +Comparison to other methods. We compare the optimized prompts with our method to three types of prompts: the short primary prompts, the prompts generated from GPT-4 [2] and prompt refiner from Open-sora [1], which is based on fine-tuned LLaMA 3.1 [38]. The specific inference prompt template for GPT-4 [2] can be found in supplementary materials. + +# 4.2. Implementation Details + +The well-performed prompts are model-specific and aligned with the distribution of training prompts. We employ Vimeo25M [40], a training dataset consisting of 25 million text-video pairs as our analysis dataset. At the same time, we choose LaVie [40] and Latte [26] as analysis T2V models, which belong to the diffusion-based and DiT architectures respectively and use Vimeo25M as one of training + +![](images/51d2efaa4a37c56c0ec9ab1e181ac09ecf5ee59fa172ab1c687300eb73fd6163.jpg) +(a) Static dimension. From left to right: naive, GPT-4, Open-sora, Ours. + +![](images/abb8e63210833c09fcc773a34a35bd02f36b09472ea92a54d52758463ee0bce2.jpg) +(b) Dynamic dimension (a robot is dancing). From left to right, from top to bottom: naive, GPT-4, Open-sora, Ours. +Figure 4. Qualitative comparisons across dynamic and static dimensions. This figure showcases videos generated using LaVie with short prompts, GPT-4 and Open-sora prompt optimizations, and our RAPO method. Videos produced with RAPO exhibit significantly sharper spatial details, smoother temporal transitions, and a closer semantic alignment with the input text. + +datasets. And we represent some extension results to some other T2V models in supplementary materials. For relation graph construction, we utilize Mistral [22] to extract scenes with corresponding subject, action and atmosphere descriptions from Vimeo25M dataset, and use all-MiniLM-L6-v2 as sentence transformer pre-trained model. We filter about 2.1M valid sentences from Vimeo25M dataset. For refactoring model training data, we prepare about 86k prompt-pairs following data preparation method in Section 3.2. For prompt discriminator training data, we first generate 7K text captions using Mistral, covering all the dimensions in VBenchmark [21]. The specific prompt template for extraction and dataset construction can be found in supplementary materials. We perform LoRA fine-tuning using LLaMA 3.1 [38], and fine-tune 8 epochs and 3 epochs for refactoring model and prompt discriminator respectively with a single A100, using a batch size of 32 and a LoRA rank of 64. + +# 4.3. Evaluation Results + +Qualitative comparisons. As shown in Fig. 4, compared with user-provided prompts and other prompt optimization + +methods, RAPO centers the input semantics and promotes the quality of generated videos from static and dynamic dimensions. The relevant modifiers promote model comprehending the simple user-provided prompts. The detailed prompts of different methods and more qualitative comparisons are provided in supplementary materials. + +Quantitative comparisons. As shown in Tab. 4 and Tab. 5, the results show that RAPO surpasses other methods in static dimensions (e.g., visual quality, object class) and dynamic dimensions (e.g., human actions, temporal flickering). The consistent performance across various benchmarks demonstrates the robustness and versatility of RAPO. Although the optimized prompts from GPT-4 and Open-sora enriches the plain user-provided prompts with more details describing the scenes, objects, and actions, these lengthy and complex descriptions may confuse models and even worsen the generation. It is worth noticing that RAPO significantly enhances the score of multiple objects dimension from $37.71\%$ to $64.86\%$ using LaVie and from $29.55\%$ to $52.78\%$ using Latte as shown in Tab. 4. And RAPO outperforms other methods on T2V-CompBench as shown in Tab. 6 for some selected dimensions, which are focused on + +
ModelsTotal Scoretemporal flickeringimaging qualityhuman actionobject classmultiple objectsspatial relationship
LaVie [40]80.89%96.62%69.00%95.80%92.09%37.71%37.27%
LaVie-GPT4 [2]79.69%96.14%70.27%83.80%88.73%36.23%50.55%
LaVie-Open-sora [1]79.75%96.42%70.42%87.00%91.29%36.52%54.37%
LaVie-RAPO82.38%96.86%71.40%96.80%96.91%64.86%59.15%
Latte [26]77.03%97.10%63.38%88.40%83.86%29.55%40.63%
Latte-GPT4 [2]77.40%97.52%63.54%85.80%78.32%27.73%36.72%
Latte-Open-sora [1]77.23%97.67%64.19%84.60%83.60%30.00%35.12%
Latte-RAPO79.97%98.17%66.72%95.20%96.47%52.78%41.31%
+ +Table 4. Quantitative comparisons on VBench. RAPO achieves the highest overall scores, especially in static attributes and multiple object generation. + +![](images/037cc8fc873890a5b4a3548a5408cb826c86e4ea1e47166dede10bdab2200267.jpg) +a bicycle and a car + +![](images/8cd071931bc10f1070aeb41e7d05f1b0d294a4435456b41b516272267004accf.jpg) +A bicycle and a car are parked side by side. + +![](images/92c237b738b4b048c6079eee7a1da6481c9e8de2a9011f2c6d014059e8942491.jpg) +bicycle + +![](images/a0ff19a6aac7f59ab36df8a7d4e25770bdada4c17f178a5202afeedd6e6233e0.jpg) +car + +![](images/f435fc08c7eaafab09a1258a44db705153064e5c5b5f28393c45fe1eeba74c0f.jpg) +bicycle + +![](images/98c3eeb4e7e42700db762e69d6d81d45b7d3ae30ea64fe43e37ede6416d3da75.jpg) +car +Figure 5. Visualization on attention map on multiple objects from different prompts. Adding description of the relative spatial position between objects can improve multi-object generation. + +complex compositional T2V tasks. These experiments verified that PAPO significantly improves the performance of generating scenes involving more than two subjects. + +# 4.4. Analyses + +Multiple objects. Synthesis quality of generated videos often declines when tasked with generating outputs that accurately represent prompts involving multiple objects. This issue is also prevalent in the T2I model, and several studies [14, 16, 30] have highlighted that the blended context created by the CLIP text encoder leads to improper binding. Meanwhile, some related works [6, 29] focus on image latents to address information loss, while the others [7, 49] pay more attention to text embedding to deal with the issue. However, few have explored optimizing prompts to improve the performance of multiple obejsts task. We apply our method to text-to-image using SD 1.4 [34], which uses the same text encoder with LaVie [40]. We test on prompts about multiple objects, and remove the irrelevant modifiers + +![](images/b7c60393e629fc0f95b39fd870b09183e1102602a5910627e73e01a9059875ec.jpg) +Figure 6. Prompt length distribution comparison among various methods. The distribution of RAPO-optimized prompts is more closer to the training prompts. + +like action and atmosphere descriptions. As shown in Fig. 5, we can find the relevant spatial descriptions boost the performance of multiple objects. More examples can be found in supplementary materials. + +Statistical analysis of text. As shown in Fig. 6, we compared the word length distributions of prompts from the T2V training set, user prompts (simulated via VBench, EvalCrafter, and T2V-CompBench), and optimized prompts generated by various methods. The results show that the prompt length distribution produced by RAPO is closest to that of the training set, and this consistency unleashes the model's generative potential to produce better videos. In contrast, user prompts are too short and lack necessary details, while other methods generate longer prompts that contain excessive details and complex vocabulary, which may be counterproductive, as shown in Tab. 4 and Tab. 5. + +# 4.5. Ablation Study + +We conduct ablation experiments on the VBench benchmark to examine the individual and combined effects of different modules in RAPO. Additionally, we perform ablation experiments on various configurations of $\mathcal{L}$ . We present more ablation experiments on hyperparameters in the supplementary materials. + +Ablating each modules in RAPO. We directly obtain the + +
ModelsFinal Sum Scoremotion qualitytext-video alignmentvisual qualitytemporal consistency
LaVie [40]24853.1969.6064.8160.87
LaVie-GPT4 [2]24654.0565.5164.9661.22
LaVie-Open-sora [1]25153.0771.3865.2661.41
LaVie-RAPO25653.3474.3866.6261.29
Latte [26]21750.0355.4957.6553.94
Latte-GPT4 [2]21851.3653.6558.0254.65
Latte-Open-sora [1]22250.2557.3258.7155.47
Latte-RAPO22751.7360.8659.2455.26
+ +Table 5. Quantitative comparisons on EvalCrafter. The results demonstrate that RAPO consistently achieves higher scores, especially in terms of text-video alignment and visual quality, indicating its superior ability to enhance the overall quality of generated videos. + +
Modelsconsistent attribute bindingdynamic attribute bindingaction bindingobject interactions
LaVie [40]0.6200.2320.4830.760
LaVie-GPT4 [2]0.5610.2180.4280.620
LaVie-Open-sora [1]0.5320.2140.4700.698
LaVie-RAPO0.6920.2670.6350.839
Latte [26]0.6330.2270.4760.792
Latte-GPT4 [2]0.5980.2100.4050.688
Latte-Open-sora [1]0.5490.2030.4870.743
Latte-RAPO0.7060.2580.5910.856
+ +Table 6. Quantitative comparisons on T2V-CompBench. RAPO outperforms alternative prompt optimization methods, exhibiting better performance at managing complex compositional tasks and ensuring coherent interactions between multiple objects in the generated videos. + +
word augmentationsentence refactoringprompt selectionVBatch Total Score
(a)✓80.37%
(b)✓79.75%
(c)✓✓81.58%
(d)✓✓81.75%
(e)✓✓80.60%
(f)✓✓✓82.38%
+ +Table 7. Ablation studies of different modules in RAPO on VBench. Each module improves performance, while the combined use of all three leads to the highest evaluation score, confirming the synergistic effect of the full RAPO framework. + +
GPT-4MistralLLaMA
VBenchmark Total Score82.38%82.25%82.10%
+ +Table 8. Ablation studies on different $\mathcal{L}$ . The results suggest that RAPO is robust and effective across various LLMs. + +related modifiers about input prompts utilizing GPT-4 [2], and merging them into inputs at one time as the comparison of word augmentation. We randomly select one of optimized prompts as the comparison of prompt selection. The optimal result is achieved by the full-fledged framework as shown in row (f). + +Ablation experiments on different $\mathcal{L}$ . We conduct ablation experiments on GPT-4 [2], Mistral [22] and LLaMA + +3.1 [38]. As shown in Tab. 8, although GPT-4 achieves the best overall score, the differences are marginal, which suggests that RAPO is robust and effective across various LLMs in generating optimized prompts for T2V generation. + +# 5. Conclusion and Future Work + +In this paper, we propose RAPO, a novel framework for prompt optimizing to improve the static and dynamic dimensions performance of generated videos by pre-trained T2V models. RAPO employs word augmentation, sentence refactoring and prompt selection to generate prompts aligned with the distribution of training prompts, involving relevant and straightforward modifiers about inputs. Extensive experimental evaluations demonstrate the effectiveness and efficiency of RAPO. For the future work, we plan to extend RAPO to more T2V models and conduct more analyses about the reasons for improvements. + +# Acknowledgments + +The work is partially supported by the National Natural Science Foundation of China (Grant No. 62471287), the National Key R&D Program of China under Grant No. 2022ZD0160102, and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800. + +# References + +[1] Open-sora: Democratizing efficient video production for all. 2024. URL: https://github.com/hpcaitech/Open-Sora. 5, 7, 8 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5,7,8 +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.opennai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3:1, 2024. 1, 3 +[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 7 +[7] Chieh-Yun Chen, Li-Wu Tsao, Chiang Tseng, and Hong-Han Shuai. A cat is a cat (not a dog!): Unraveling information mix-ups in text-to-image encoders through causal analysis and embedding optimization. arXiv preprint arXiv:2410.00321, 2024. 1, 7 +[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3 +[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246-22256, 2023. 3 +[10] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations. 3 +[11] Zijie Chen, Lichao Zhang, Fangsheng Weng, Lili Pan, and Zhenzhong Lan. Tailored visions: Enhancing text-to-image generation with personalized prompt rewriting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7727-7736, 2024. 3 +[12] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure + +and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1 +[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3 +[14] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 7 +[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3 +[16] Boyu Han, Qianqian Xu, Zhiyong Yang, Shilong Bao, Peisong Wen, Yangbangyan Jiang, and Qingming Huang. Aucseg: Auc-oriented pixel-level long-tail semantic segmentation. arXiv preprint arXiv:2409.20398, 2024. 7 +[17] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3 +[18] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. *Animate-a-story: Storytelling with retrieval-augmented video generation*. arXiv preprint arXiv:2307.06940, 2023. 3 +[19] Nailei Hei, Qianyu Guo, Zihao Wang, Yan Wang, Haofen Wang, and Wenqiang Zhang. A user-friendly framework for generating model-preferred prompts in text-to-image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2139-2147, 2024. 1 +[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3 +[21] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 2, 5, 6 +[22] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6, 8 +[23] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024.1 +[24] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, + +Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. 3 +[25] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 2, 5 +[26] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 1, 2, 3, 5, 7, 8 +[27] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, Ji-Rong Wen, and Qing Yang. Dynamic prompt optimizing for text-to-image generation. In CVPR, 2024. 1, 3 +[28] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1, 3 +[29] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7942, 2024. 7 +[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 7 +[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1, 3 +[32] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. arXiv preprint arXiv:2310.15169, 2023. 1 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3, 7 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[36] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation + +without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1 +[37] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. arXiv preprint arXiv:2407.14505, 2024. 2, 5 +[38] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5, 6, 8 +[39] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 1 +[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 1, 2, 3, 5, 7, 8 +[41] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 1 +[42] Mengjiao Yang, Yilun Du, Bo Dai, Dale Schuurmans, Joshua B Tenenbaum, and Pieter Abbeel. Probabilistic adaptation of text-to-video models. arXiv preprint arXiv:2306.01872, 2023. 3 +[43] Shuai Yang, Jing Tan, Mengchen Zhang, Tong Wu, Yixuan Li, Gordon Wetzstein, Ziwei Liu, and Dahua Lin. Layer-pano3d: Layered 3d panorama for hyper-immersive scene generation. arXiv preprint arXiv:2408.13252, 2024. 3 +[44] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 3 +[45] Jingtao Zhan, Qingyao Ai, Yiqun Liu, Jia Chen, and Shaoping Ma. Capability-aware prompt reformulation learning for text-to-image generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2145-2155, 2024. 3 +[46] Jingtao Zhan, Qingyao Ai, Yiqun Liu, Yingwei Pan, Ting Yao, Jiaxin Mao, Shaoping Ma, and Tao Mei. Prompt refinement with image pivot for text-to-image generation. arXiv preprint arXiv:2407.00247, 2024. 1 +[47] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1-15, 2024. 3 +[48] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 3 +[49] Chenyi Zhuang, Ying Hu, and Pan Gao. Magnet: We never know how text-to-image diffusion models work, until we + +learn how vision-language models function. arXiv preprint arXiv:2409.19967, 2024. 7 +[50] Shaobin Zhuang, Kunchang Li, Xinyuan Chen, Yaohui Wang, Ziwei Liu, Yu Qiao, and Yali Wang. Vlogger: Make your dream a vlog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8806-8817, 2024. 3 \ No newline at end of file diff --git a/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/images.zip b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f06404165b5f327a37dc776758d13def498673ac --- /dev/null +++ b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:687eeb567521b8011fb175acb5c676bba7bddfed63269dfbef0d9a83c7b4bb86 +size 795825 diff --git a/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/layout.json b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cdc40feb37f6221452c85ef88c110ef583451280 --- /dev/null +++ b/2025/The Devil is in the Prompts_ Retrieval-Augmented Prompt Optimization for Text-to-Video Generation/layout.json @@ -0,0 +1,9065 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 539, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 539, + 138 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 539, + 138 + ], + "type": "text", + "content": "The Devil is in the Prompts: Retrieval-Augmented Prompt Optimization for Text-to-Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 160, + 187, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 160, + 187, + 175 + ], + "spans": [ + { + "bbox": [ + 114, + 160, + 187, + 175 + ], + "type": "text", + "content": "Bingjie Gao" + }, + { + "bbox": [ + 114, + 160, + 187, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 175, + 206, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 175, + 206, + 189 + ], + "spans": [ + { + "bbox": [ + 156, + 175, + 206, + 189 + ], + "type": "text", + "content": "Yu Qiao" + }, + { + "bbox": [ + 156, + 175, + 206, + 189 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 161, + 283, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 161, + 283, + 175 + ], + "spans": [ + { + "bbox": [ + 217, + 161, + 283, + 175 + ], + "type": "text", + "content": "Xinyu Gao" + }, + { + "bbox": [ + 217, + 161, + 283, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 221, + 175, + 264, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 175, + 264, + 188 + ], + "spans": [ + { + "bbox": [ + 221, + 175, + 264, + 188 + ], + "type": "text", + "content": "Li Niu" + }, + { + "bbox": [ + 221, + 175, + 264, + 188 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 279, + 175, + 358, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 175, + 358, + 189 + ], + "spans": [ + { + "bbox": [ + 279, + 175, + 358, + 189 + ], + "type": "text", + "content": "Xinyuan Chen" + }, + { + "bbox": [ + 279, + 175, + 358, + 189 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 419, + 161, + 485, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 161, + 485, + 175 + ], + "spans": [ + { + "bbox": [ + 419, + 161, + 485, + 175 + ], + "type": "text", + "content": "Yujie Zhou1,2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 375, + 175, + 449, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 175, + 449, + 189 + ], + "spans": [ + { + "bbox": [ + 375, + 175, + 449, + 189 + ], + "type": "text", + "content": "Yaohui Wang" + }, + { + "bbox": [ + 375, + 175, + 449, + 189 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 190, + 219, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 190, + 219, + 203 + ], + "spans": [ + { + "bbox": [ + 66, + 190, + 219, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 190, + 219, + 203 + ], + "type": "text", + "content": "Shanghai Jiao Tong University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 225, + 189, + 535, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 189, + 535, + 204 + ], + "spans": [ + { + "bbox": [ + 225, + 189, + 535, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 225, + 189, + 535, + 204 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 225, + 189, + 535, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 225, + 189, + 535, + 204 + ], + "type": "text", + "content": "Fudan University" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 256, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 296, + 555 + ], + "type": "text", + "content": "The evolution of Text-to-video (T2V) generative models, trained on large-scale datasets, has been marked by significant progress. However, the sensitivity of T2V generative models to input prompts highlights the critical role of prompt design in influencing generative outcomes. Prior research has predominantly relied on Large Language Models (LLMs) to align user-provided prompts with the distribution of training prompts, albeit without tailored guidance encompassing prompt vocabulary and sentence structure nuances. To this end, we introduce RAPO, a novel Retrieval-Augmented Prompt Optimization framework. In order to address potential inaccuracies and ambiguous details generated by LLM-generated prompts. RAPO refines the naive prompts through dual optimization branches, selecting the superior prompt for T2V generation. The first branch augments user prompts with diverse modifiers extracted from a learned relational graph, refining them to align with the format of training prompts via a fine-tuned LLM. Conversely, the second branch rewrites the naive prompt using a pre-trained LLM following a well-defined instruction set. Extensive experiments demonstrate that RAPO can effectively enhance both the static and dynamic dimensions of generated videos, demonstrating the significance of prompt optimization for user-provided prompts. Project website: GitHub." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 579, + 135, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 135, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 135, + 591 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "content": "With the rapid advancement of diffusion models [28, 33, 34], visual content creation has experienced remarkable progress in recent years. The generation of images, as well as videos from text prompts utilizing large-scale diffusion models, referred to as text-to-images (T2I) [13, 30, 35] and text-to-videos (T2V) [5, 12, 40] generation, have attracted significant interest due to the broad range of applications in real-world scenarios. Various efforts have been made to en" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 232, + 553, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 268 + ], + "type": "text", + "content": "hance the performance of these models, including improvements in model architecture [23, 26, 26], learning strategies [36, 44], and data curation [32, 39, 41]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 269, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 447 + ], + "type": "text", + "content": "Recent studies [17, 31, 44] have revealed that employing long, detailed prompts with a pre-trained model typically produces superior quality outcomes compared to utilizing shallow descriptions provided by users. This has underscored the significance of prompt optimization as an important challenge in text-based visual content creation. The prompts provided by users are often brief and lack the essential details required to generate vivid images or videos. Simply attempting to optimize prompts by manually adding random descriptions can potentially mislead models and degrade the quality of generative results, resulting in outputs that may not align with user intentions. Therefore, developing automated methods to enhance user-provided prompts becomes essential for improving the overall quality of generated content." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "content": "Towards improving image aesthetics and ensuring semantic consistency, several attempts [19, 27, 46] have been made in previous T2I works for prompt optimization. These efforts primarily involve instructing a pre-trained or finetuned Large Language Model (LLM) to incorporate detailed modifiers into original prompts, with the aim of enhancing spatial elements such as color and relationships. While these approaches have displayed promising outcomes in image generation, studies [7, 17] reveal that their impact on video generation remains limited, especially in terms of enhancing temporal aspects such as motion smoothness and minimizing temporal flickering." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 593, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 556, + 676 + ], + "type": "text", + "content": "For T2V generation, we observe that the quality of spatial and temporal elements is significantly influenced by the selection of appropriate verb-object phrases and the structure of input prompts, which largely rely on analyzing and structuring the entire training data in a systematical manner. To this end, we propose RAPO, a Retrieval-Augmented Prompt Optimization framework for T2V generation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "The primary objective of RAPO is to convert user-provided prompts into optimized prompts that retain the original semantics while closely adhering to the format of" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "type": "text", + "content": "†Corresponding author" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3173" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 555, + 434 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 555, + 434 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 555, + 434 + ], + "type": "image", + "image_path": "6fff60468cf1a329b4b3026eb00d3ad87bfb7cb3e11ed85bd2c07fb86bc9a501.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 436, + 360, + 447 + ], + "lines": [ + { + "bbox": [ + 283, + 436, + 360, + 447 + ], + "spans": [ + { + "bbox": [ + 283, + 436, + 360, + 447 + ], + "type": "text", + "content": "(b) static dimension" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 460, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 54, + 460, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 555, + 485 + ], + "type": "text", + "content": "Figure 1. Videos generated using LaVie [40] conditioned on user-provided prompts and optimized prompts from RAPO. The optimized prompts can significantly enhance both the static and dynamic qualities of the generated videos, making them more visually appealing." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 504, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 504, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 54, + 504, + 295, + 588 + ], + "type": "text", + "content": "the training data, including the vocabulary and sentence structure. To achieve this, we initially construct a Relation Graph from the training data, with each node representing words and edges indicating the relationships between neighboring words. By leveraging any user-provided prompt, we can retrieve the most relevant terms to enhance the original descriptions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Subsequently, a specially fine-tuned Refactoring LLM is introduced to reorganize the structure and writing style of the word-augmented prompt, aligning it more cohesively with the original training data. Additionally, we employ a Rewrite LLM to directly transform user-provided prompts into optimized prompts that follow a specific format consistent with the training prompts, serving as candidates for optimized prompts. In order to address potential inaccuracies and ambiguous details generated by LLMs, we conclude by fine-tuning a Discriminator LLM to identify the most suit-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 505, + 448, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 448, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 448, + 517 + ], + "type": "text", + "content": "able prompt for T2V generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": "We apply RAPO on two T2V generation models LaVie [40] and Latte [26], and evaluate it on three benchmarks, namely VBench [21], EvalCrafter [25] and T2V-CompBench [37]. It is worth noticing that RAPO greatly promotes the performance of multiple objects dimension in VBench from " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "37.71\\%" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "64.86\\%" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " using LaVie and from " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "29.55\\%" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "52.78\\%" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " using Latte. In addition, the experimental results show that RAPO improves the performance of T2V models to synthesize scenes with multiple subjects, even unusual scenes like \"a mushroom growing out of a human head\" as shown in Fig. 1. Our contributions can be summarized as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": "- We propose RAPO, a Retrieval-Augmented Prompt Optimization framework T2V generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "- We propose word augmentation module and sentence refactoring module to optimize prompts considering" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3174" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 260, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 260, + 83 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 260, + 83 + ], + "type": "text", + "content": "prompt vocabulary and specific sentence format." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 84, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 84, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 56, + 84, + 294, + 108 + ], + "type": "text", + "content": "- We validate the effectiveness of RAPO through extensive experiments on several benchmarks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 142, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 142, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 142, + 133 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 141, + 295, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 295, + 332 + ], + "type": "text", + "content": "Text-to-Video generation. With the remarkable breakthroughs of diffusion models [28, 33, 34], the generations of 3D content [9, 24, 43], images [3, 13, 30], and videos [4, 5, 20] from text descriptions achieve rapid advancement. Text-to-Video (T2V) [8, 40, 47] Generation aims to automatically create videos that match given textual descriptions. This process generally involves comprehending the scenes, objects, and actions described in the text and converting them into a sequence of cohesive visual frames, producing a video that is logically and visually consistent. T2V generation is wildly used in applications, such as animations [10, 15, 18] and automatic movie generation [42, 48, 50]. However, large T2V generative models [26, 40, 44] trained on large-scale dataset could not adequately demonstrate their potential in generation due to mismatch between training and inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 334, + 295, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 295, + 609 + ], + "type": "text", + "content": "Prompt optimization. T2I and T2V generative models are sensitive to input prompts. However, the well-performed prompts are often model-specific and coherent with training prompts, misaligned with user input. Therefore, several studies [11, 17, 27, 45] are conducted to explore the generative potential of T2I and T2V generative models. Hao et al. [17] propose a learning-based prompt optimizing framework unitizing reinforce learning for generating more aesthetically pleasing images. Chen et al. [11] enhance user prompts by leveraging the user's historical interactions with the system. Mo et al. [27] propose Prompt Auto-Editing (PAE) method to decide the weights and injection time steps of each word without manual intervention. These methods primarily focus on prompts optimizing for T2I models and lack extension to T2V models. Yang et al. [44] use large language models (LLMs) to transform short prompts into more detailed ones, maintaining a consistent visual structure. Polyak et al. [31] develop a teacher-student distillation approach for prompt optimization to improve computational efficiency and reduce latency. However, the results of optimized prompts usually could not be well-aligned with training prompts due to the misleading of the LLMs and the lack of more refined guidance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 620, + 111, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 111, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 111, + 633 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "As illustrated in Fig. 2, RAPO consists of three parts, 1) a word augmentation module, 2) a sentence refactoring module, as well as 3) a prompt selection module. Given a user-provided prompt " + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": ", firstly, the word augmentation module utilizes an interactive retrieval-merge mechanism between a relation graph " + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": " and a LLM " + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": " to augment the prompt" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "by adding related subject, action and atmosphere modifiers. Then, a fine-tuned LLM " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " is applied to refactor the entire sentence into " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "x_r" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "x_r" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " has a more unified format which is consistent with the prompt length and format distribution in training data. Finally, a discriminator in the prompt selection module decides between " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "x_r" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " and a naively augmented prompt " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "x_n" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " obtained directly from a LLM via instruction, as the most suitable augmented prompt for T2V generation. We proceed to introduce each module in detail in the following sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 199, + 471, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 471, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 471, + 211 + ], + "type": "text", + "content": "3.1. Word Augmentation Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "text", + "content": "Given a user-provided prompt " + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "text", + "content": ", the word augmentation module aims to enrich " + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "text", + "content": " with more multiple, straightforward and relevant modifiers. It is achieved through retrieving modifiers meeting the requirements from a built relation graph " + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "text", + "content": ", and merging them into " + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "text", + "content": " via instruction. In this section, we first introduce the construction and retrieval of relation graph. And we introduce the instruction format and retrieval-merge mechanism of " + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 216, + 553, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 553, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 467 + ], + "type": "text", + "content": "Relation Graph " + }, + { + "bbox": [ + 313, + 312, + 553, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 312, + 553, + 467 + ], + "type": "text", + "content": ". As shown in Fig. 3, we construct relation graph " + }, + { + "bbox": [ + 313, + 312, + 553, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 312, + 553, + 467 + ], + "type": "text", + "content": " based on training prompts database. For each training prompt, we utilize " + }, + { + "bbox": [ + 313, + 312, + 553, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 312, + 553, + 467 + ], + "type": "text", + "content": " to extract scene and corresponding related modifiers (subject, action, atmosphere descriptions). Each scene serves as a core node, with subject, action and atmosphere modifiers connected as individual sub-nodes in relation graph. For each extracted scene, we first check whether it exists in relation graph or not. If so the extracted related modifiers will be connected to the existing one. If not the extracted scene becomes a new core node with related modifiers connected. Finally, we can obtain a relation graph covering diverse scenes with multiple modifiers connected." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "text", + "content": "For relation graph retrieval, we utilize a sentence transformer pre-trained model to extract features of prompt, and employ the cosine similarity to measure similarity between sentence features. We first retrieve the top-k relevant scenes from " + }, + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "text", + "content": ". Then we retrieve all modifiers connected to the retrieved scenes. We select the top-k relevant modifiers " + }, + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "inline_equation", + "content": "\\{p_n|_{n = 0}^{k - 1}\\}" + }, + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "text", + "content": " from all retrieved modifiers, preparing for the retrieval-merge mechanism of " + }, + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 468, + 553, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "text", + "content": "LLM " + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "text", + "content": ". We augment " + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "text", + "content": " with retrieved modifiers " + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "inline_equation", + "content": "\\{p_n|_{n = 0}^{k - 1}\\}" + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "text", + "content": " from relation graph. We rename " + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "inline_equation", + "content": "x_{i}^{0}" + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "text", + "content": " to illustrate the process of iterative merging. Specifically, the retrieved modifiers are merged into input prompt " + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "inline_equation", + "content": "x_{i}^{0}" + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "text", + "content": " one by one through prompting " + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 562, + 553, + 635 + ], + "type": "text", + "content": ", to maintain the information of the original input while adding relevant modifiers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 390, + 643, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 643, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 390, + 643, + 553, + 658 + ], + "type": "interline_equation", + "content": "x _ {i} ^ {m + 1} = f \\left(x _ {i} ^ {m}, p _ {i} ^ {m}\\right), \\tag {1}", + "image_path": "255d88309b450d61abdeb7e7d77c19baaca56ddbee6d4d3f5f12d6e1cae7d5ce.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "m = 0,1,\\dots,k - 1" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": " is a function that combine " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "x_{i}^{m}" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "p_i^m" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": " reasonably by " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": ". For instance, a merged prompt \"a woman dressed in a black suit representing a funeral\" is resulted from merging the user-provided prompt \"a woman" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3175" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 547, + 270 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 547, + 270 + ], + "type": "image", + "image_path": "30c2cf4568e1472d731af966eb03ac0fc3d59f023ea6e4cabbaa0f9f69a4d329.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 338 + ], + "type": "text", + "content": "Figure 2. Overview of RAPO. The naive prompt is optimized by two branches respectively. For the first branch, it is enriched by the word augmentation module based on a constructed relation graph and a frozen Large Language Model (LLM). Subsequently, augmented prompt is refactored by a finetuned LLM into a specific format in the sentence refactoring module. For the second branch, the naive prompt is directly rewritten by a frozen LLM. Finally, the prompt selection module selects the better one from two branches' results as input for T2V model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 355, + 304, + 564 + ], + "blocks": [ + { + "bbox": [ + 58, + 355, + 304, + 564 + ], + "lines": [ + { + "bbox": [ + 58, + 355, + 304, + 564 + ], + "spans": [ + { + "bbox": [ + 58, + 355, + 304, + 564 + ], + "type": "image", + "image_path": "ea314a54bd93ade58abafa5bff17cac48974326447a62928f4f8eb1437b8aaff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 571, + 297, + 639 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 297, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 297, + 639 + ], + "type": "text", + "content": "Figure 3. The construction of relation graph. Relation graph consists of multiple nodes (scenes acting as core nodes with modifiers connected as sub-nodes). For each prompt in database, LLM extracts scene and related modifiers. Based on whether the extracted scene is already in the graph or not, different methods are used to incorporate the new information into the graph." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 297, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 717 + ], + "type": "text", + "content": "representing a funeral\" and a retrieved modifier \"a black suit\". We prompt " + }, + { + "bbox": [ + 55, + 665, + 297, + 717 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 665, + 297, + 717 + ], + "type": "text", + "content": " to perform general prompt merging in a normal manner as the template in Tab. 1. In instruction, we provide some prompt pairs " + }, + { + "bbox": [ + 55, + 665, + 297, + 717 + ], + "type": "inline_equation", + "content": "E = \\{e_i|_{i = 0}^n\\}" + }, + { + "bbox": [ + 55, + 665, + 297, + 717 + ], + "type": "text", + "content": " as examples, in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 357, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 382 + ], + "type": "text", + "content": "which " + }, + { + "bbox": [ + 313, + 357, + 555, + 382 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 313, + 357, + 555, + 382 + ], + "type": "text", + "content": " contains input prompt, a modifier and corresponding merged result." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 320, + 396, + 528, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 396, + 528, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 396, + 528, + 409 + ], + "type": "text", + "content": "LLM Template for Retrieval-Merge Mechanism" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 319, + 414, + 549, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 414, + 549, + 497 + ], + "spans": [ + { + "bbox": [ + 319, + 414, + 549, + 497 + ], + "type": "text", + "content": "Suppose you are a Text Merger. You receive two inputs from the user: a description body and a relevant modifier. Your task is to enrich the description body with relevant modifiers while retaining the description body. You should ensure that the output text is coherent, contextually relevant, and follows the same structure as the examples provided." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 497, + 529, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 497, + 529, + 510 + ], + "spans": [ + { + "bbox": [ + 319, + 497, + 529, + 510 + ], + "type": "text", + "content": "Examples of prompt-pairs provided: " + }, + { + "bbox": [ + 319, + 497, + 529, + 510 + ], + "type": "inline_equation", + "content": "E = \\{e_i|_{i = 0}^n\\}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 510, + 527, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 510, + 527, + 522 + ], + "spans": [ + { + "bbox": [ + 320, + 510, + 527, + 522 + ], + "type": "text", + "content": "Input description body and modifier are: " + }, + { + "bbox": [ + 320, + 510, + 527, + 522 + ], + "type": "inline_equation", + "content": "\\{x_i^m,p_i^m\\}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 522, + 453, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 522, + 453, + 534 + ], + "spans": [ + { + "bbox": [ + 320, + 522, + 453, + 534 + ], + "type": "text", + "content": "The merged prompt is: " + }, + { + "bbox": [ + 320, + 522, + 453, + 534 + ], + "type": "inline_equation", + "content": "\\{x_i^{m + 1}\\}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 613 + ], + "type": "text", + "content": "Table 1. Input template for retrieval-merge mechanism. This template specifies how a frozen LLM iteratively merges user-provided prompt texts with relevant modifiers retrieved from a relation graph, thereby enriching the prompt's semantic content and aligning it with the training prompt structure for improved text-to-video synthesis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 635, + 476, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 476, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 476, + 648 + ], + "type": "text", + "content": "3.2. Sentence Refactoring Module" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Sentence refactoring module aims to refactor word augmented prompts from word augmentation module to be more consistent with prompt format in training data. It is achieved through a fine-tuned LLM " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{r}" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": " named as refactoring model. In this section, we introduce the training data" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3176" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 225, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 225, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 225, + 84 + ], + "type": "text", + "content": "preparation and instruction tuning for " + }, + { + "bbox": [ + 55, + 72, + 225, + 84 + ], + "type": "inline_equation", + "content": "L_{r}" + }, + { + "bbox": [ + 55, + 72, + 225, + 84 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": "Data preparation. We represent the required dataset for training refactoring model by " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "\\{D_r = r_i|_{i=1}^{N^r}\\}" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": " involves a pair of prompts and " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "N^r" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": " is the number of training prompts pairs. Specifically, " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "r_i = (w_i, c_i)" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": " targets to simulate world augmented prompt, and " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": " represents the target prompt, that is, a training prompt for T2V models. " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": " share similar semantics while different in the prompt format and length. Therefore, we generate " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": " automatically through rewriting " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": " utilizing " + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": " via instruction to break the unified training prompt format but maintaining the original semantics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": "Instruction tuning for " + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "inline_equation", + "content": "L_{r}" + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": ". We employ instruction tuning for fine-tuning a LLM on our constructed dataset of instructional prompts and corresponding outputs. The constructed dataset is based on " + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\{D_r = r_i|_{i = 1}^{N^r}\\}" + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": " containing instructional prompts and corresponding outputs. The template of the instruction tuning dataset for " + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "inline_equation", + "content": "L_{r}" + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": " is shown as Tab. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 301, + 209, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 301, + 209, + 314 + ], + "spans": [ + { + "bbox": [ + 61, + 301, + 209, + 314 + ], + "type": "text", + "content": "Instruction Tuning Dataset for " + }, + { + "bbox": [ + 61, + 301, + 209, + 314 + ], + "type": "inline_equation", + "content": "L_{r}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 318, + 290, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 318, + 290, + 378 + ], + "spans": [ + { + "bbox": [ + 61, + 318, + 290, + 378 + ], + "type": "text", + "content": "Instruction. Refine format and word length of the sentence: " + }, + { + "bbox": [ + 61, + 318, + 290, + 378 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 61, + 318, + 290, + 378 + ], + "type": "text", + "content": ". Maintain the original subject descriptions, actions, scene descriptions. Append additional straightforward actions to make the sentence more dynamic if necessary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 379, + 164, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 379, + 164, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 379, + 164, + 392 + ], + "type": "text", + "content": "Output: target prompt " + }, + { + "bbox": [ + 61, + 379, + 164, + 392 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 61, + 379, + 164, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 448 + ], + "type": "text", + "content": "Table 2. Instruction tuning dataset template for " + }, + { + "bbox": [ + 55, + 403, + 296, + 448 + ], + "type": "inline_equation", + "content": "L_{r}" + }, + { + "bbox": [ + 55, + 403, + 296, + 448 + ], + "type": "text", + "content": ". This template directs LLM fine-tuning to restructure augmented prompts by adjusting their format while preserving semantics, aligning them with the training data's style for improved T2V generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 198, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 198, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 198, + 480 + ], + "type": "text", + "content": "3.3. Prompt Selection Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "text", + "content": "As shown in Fig. 2, prompt selection module contains a fine-tuned LLM " + }, + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "text", + "content": " named prompt discriminator to select the better one between " + }, + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "text", + "content": " from sentence refactoring module, and a naively augmented prompt " + }, + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "text", + "content": " obtained directly from a LLM via instruction. In this section, we introduce the training data preparation and instruction tuning for " + }, + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 55, + 485, + 295, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "Data preparation. We represent the required dataset for training refactoring model by " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{D_{d} = d_{i}|_{i = 1}^{N^{d}}\\}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " contains three prompts and " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "N^d" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " is the number of training prompts triples. Specifically, " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "d_{i} = (x_{i},x_{r},x_{n},y_{d})" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "y_{d}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " represents the discriminator label to select the better one for T2V generation from " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " given input prompt " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": ". To simulate the user-provided prompts, we collect diverse prompts from several T2V benchmarks and generate more utilizing " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " via instruction. " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " can be obtained from the proposed RAPO as shown in Fig. 2 given " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": ". We determine " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "y_{d}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " through the evaluation of generated videos conditioned on " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": ". Specifically, the evaluations of T2V models performance involves diverse" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": "dimensions. For collected or generated prompts, we need to determine the evaluation dimension according to prompt content. We automatically decide the evaluation dimension of input prompts utilizing " + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": ", then choose the corresponding metrics to evaluate generated videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": "Instruction tuning for " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": ". Similar to " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "L_{r}" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": ", we employ instruction tuning for " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "\\{D_{d} = d_{i}|_{i = 1}^{N^{d}}\\}" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": ". The template of the instruction tuning dataset for " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": " is shown as Tab. 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 193, + 468, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 193, + 468, + 205 + ], + "spans": [ + { + "bbox": [ + 321, + 193, + 468, + 205 + ], + "type": "text", + "content": "Instruction Tuning Dataset for " + }, + { + "bbox": [ + 321, + 193, + 468, + 205 + ], + "type": "inline_equation", + "content": "L_{d}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "spans": [ + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "text", + "content": "Instruction. Given user-provided prompt " + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "text", + "content": ", select the better optimized prompt from " + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "text", + "content": ". The chosen prompt is required to contain multiple, straightforward, and relevant modifiers about " + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "text", + "content": " while involving the semantics of " + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 319, + 210, + 548, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 270, + 447, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 270, + 447, + 283 + ], + "spans": [ + { + "bbox": [ + 320, + 270, + 447, + 283 + ], + "type": "text", + "content": "Output: discriminator label " + }, + { + "bbox": [ + 320, + 270, + 447, + 283 + ], + "type": "inline_equation", + "content": "y_{d}" + }, + { + "bbox": [ + 320, + 270, + 447, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 294, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 339 + ], + "type": "text", + "content": "Table 3. Instruction tuning dataset template for " + }, + { + "bbox": [ + 313, + 294, + 555, + 339 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 313, + 294, + 555, + 339 + ], + "type": "text", + "content": ". This template aims to train a discriminator LLM that evaluates multiple refined prompts and selects the optimal one based on the inclusion of clear, straightforward modifiers and faithful semantic alignment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 364, + 397, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 364, + 397, + 378 + ], + "spans": [ + { + "bbox": [ + 314, + 364, + 397, + 378 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 384, + 430, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 430, + 397 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 430, + 397 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 401, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 520 + ], + "type": "text", + "content": "Evaluation and Metrics. We utilize the VBenchmark [21], EvalCrafter [25] and T2V-CompBench [37] for quantitative performance evaluation. VBenchmark is an extensive benchmark comprising 16 fine-grained dimensions designed to systematically evaluate the quality of generated videos. EvalCrafter is a comprehensive evaluation benchmark that includes approximately 17 objective metrics for evaluating video generation performance. T2V-CompBench is the first benchmark tailored for compositional text-to-video generation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 605 + ], + "type": "text", + "content": "Comparison to other methods. We compare the optimized prompts with our method to three types of prompts: the short primary prompts, the prompts generated from GPT-4 [2] and prompt refiner from Open-sora [1], which is based on fine-tuned LLaMA 3.1 [38]. The specific inference prompt template for GPT-4 [2] can be found in supplementary materials." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 612, + 448, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 448, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 448, + 625 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "The well-performed prompts are model-specific and aligned with the distribution of training prompts. We employ Vimeo25M [40], a training dataset consisting of 25 million text-video pairs as our analysis dataset. At the same time, we choose LaVie [40] and Latte [26] as analysis T2V models, which belong to the diffusion-based and DiT architectures respectively and use Vimeo25M as one of training" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3177" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 218 + ], + "type": "image", + "image_path": "51d2efaa4a37c56c0ec9ab1e181ac09ecf5ee59fa172ab1c687300eb73fd6163.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 220, + 450, + 232 + ], + "lines": [ + { + "bbox": [ + 182, + 220, + 450, + 232 + ], + "spans": [ + { + "bbox": [ + 182, + 220, + 450, + 232 + ], + "type": "text", + "content": "(a) Static dimension. From left to right: naive, GPT-4, Open-sora, Ours." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 238, + 550, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 238, + 550, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 238, + 550, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 550, + 365 + ], + "type": "image", + "image_path": "abb8e63210833c09fcc773a34a35bd02f36b09472ea92a54d52758463ee0bce2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 368, + 525, + 380 + ], + "lines": [ + { + "bbox": [ + 88, + 368, + 525, + 380 + ], + "spans": [ + { + "bbox": [ + 88, + 368, + 525, + 380 + ], + "type": "text", + "content": "(b) Dynamic dimension (a robot is dancing). From left to right, from top to bottom: naive, GPT-4, Open-sora, Ours." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 392, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 54, + 392, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 555, + 427 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons across dynamic and static dimensions. This figure showcases videos generated using LaVie with short prompts, GPT-4 and Open-sora prompt optimizations, and our RAPO method. Videos produced with RAPO exhibit significantly sharper spatial details, smoother temporal transitions, and a closer semantic alignment with the input text." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 297, + 662 + ], + "type": "text", + "content": "datasets. And we represent some extension results to some other T2V models in supplementary materials. For relation graph construction, we utilize Mistral [22] to extract scenes with corresponding subject, action and atmosphere descriptions from Vimeo25M dataset, and use all-MiniLM-L6-v2 as sentence transformer pre-trained model. We filter about 2.1M valid sentences from Vimeo25M dataset. For refactoring model training data, we prepare about 86k prompt-pairs following data preparation method in Section 3.2. For prompt discriminator training data, we first generate 7K text captions using Mistral, covering all the dimensions in VBenchmark [21]. The specific prompt template for extraction and dataset construction can be found in supplementary materials. We perform LoRA fine-tuning using LLaMA 3.1 [38], and fine-tune 8 epochs and 3 epochs for refactoring model and prompt discriminator respectively with a single A100, using a batch size of 32 and a LoRA rank of 64." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 167, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 167, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 167, + 683 + ], + "type": "text", + "content": "4.3. Evaluation Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Qualitative comparisons. As shown in Fig. 4, compared with user-provided prompts and other prompt optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 447, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 519 + ], + "type": "text", + "content": "methods, RAPO centers the input semantics and promotes the quality of generated videos from static and dynamic dimensions. The relevant modifiers promote model comprehending the simple user-provided prompts. The detailed prompts of different methods and more qualitative comparisons are provided in supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Quantitative comparisons. As shown in Tab. 4 and Tab. 5, the results show that RAPO surpasses other methods in static dimensions (e.g., visual quality, object class) and dynamic dimensions (e.g., human actions, temporal flickering). The consistent performance across various benchmarks demonstrates the robustness and versatility of RAPO. Although the optimized prompts from GPT-4 and Open-sora enriches the plain user-provided prompts with more details describing the scenes, objects, and actions, these lengthy and complex descriptions may confuse models and even worsen the generation. It is worth noticing that RAPO significantly enhances the score of multiple objects dimension from " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "37.71\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "64.86\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " using LaVie and from " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "29.55\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "52.78\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " using Latte as shown in Tab. 4. And RAPO outperforms other methods on T2V-CompBench as shown in Tab. 6 for some selected dimensions, which are focused on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3178" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 168 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 168 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 168 + ], + "type": "table", + "html": "
ModelsTotal Scoretemporal flickeringimaging qualityhuman actionobject classmultiple objectsspatial relationship
LaVie [40]80.89%96.62%69.00%95.80%92.09%37.71%37.27%
LaVie-GPT4 [2]79.69%96.14%70.27%83.80%88.73%36.23%50.55%
LaVie-Open-sora [1]79.75%96.42%70.42%87.00%91.29%36.52%54.37%
LaVie-RAPO82.38%96.86%71.40%96.80%96.91%64.86%59.15%
Latte [26]77.03%97.10%63.38%88.40%83.86%29.55%40.63%
Latte-GPT4 [2]77.40%97.52%63.54%85.80%78.32%27.73%36.72%
Latte-Open-sora [1]77.23%97.67%64.19%84.60%83.60%30.00%35.12%
Latte-RAPO79.97%98.17%66.72%95.20%96.47%52.78%41.31%
", + "image_path": "3d5809f542c383a01942e359c1831b5ce8e07cb025d8b6b09cc5fa1a92a67909.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 177, + 555, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 555, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 555, + 199 + ], + "type": "text", + "content": "Table 4. Quantitative comparisons on VBench. RAPO achieves the highest overall scores, especially in static attributes and multiple object generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 64, + 218, + 175, + 330 + ], + "blocks": [ + { + "bbox": [ + 64, + 218, + 175, + 330 + ], + "lines": [ + { + "bbox": [ + 64, + 218, + 175, + 330 + ], + "spans": [ + { + "bbox": [ + 64, + 218, + 175, + 330 + ], + "type": "image", + "image_path": "037cc8fc873890a5b4a3548a5408cb826c86e4ea1e47166dede10bdab2200267.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 335, + 154, + 344 + ], + "lines": [ + { + "bbox": [ + 91, + 335, + 154, + 344 + ], + "spans": [ + { + "bbox": [ + 91, + 335, + 154, + 344 + ], + "type": "text", + "content": "a bicycle and a car" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 218, + 288, + 330 + ], + "blocks": [ + { + "bbox": [ + 177, + 218, + 288, + 330 + ], + "lines": [ + { + "bbox": [ + 177, + 218, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 177, + 218, + 288, + 330 + ], + "type": "image", + "image_path": "8cd071931bc10f1070aeb41e7d05f1b0d294a4435456b41b516272267004accf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 331, + 268, + 347 + ], + "lines": [ + { + "bbox": [ + 192, + 331, + 268, + 347 + ], + "spans": [ + { + "bbox": [ + 192, + 331, + 268, + 347 + ], + "type": "text", + "content": "A bicycle and a car are parked side by side." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 348, + 119, + 403 + ], + "blocks": [ + { + "bbox": [ + 63, + 348, + 119, + 403 + ], + "lines": [ + { + "bbox": [ + 63, + 348, + 119, + 403 + ], + "spans": [ + { + "bbox": [ + 63, + 348, + 119, + 403 + ], + "type": "image", + "image_path": "92c237b738b4b048c6079eee7a1da6481c9e8de2a9011f2c6d014059e8942491.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 405, + 106, + 414 + ], + "lines": [ + { + "bbox": [ + 82, + 405, + 106, + 414 + ], + "spans": [ + { + "bbox": [ + 82, + 405, + 106, + 414 + ], + "type": "text", + "content": "bicycle" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 121, + 348, + 176, + 403 + ], + "blocks": [ + { + "bbox": [ + 121, + 348, + 176, + 403 + ], + "lines": [ + { + "bbox": [ + 121, + 348, + 176, + 403 + ], + "spans": [ + { + "bbox": [ + 121, + 348, + 176, + 403 + ], + "type": "image", + "image_path": "a0ff19a6aac7f59ab36df8a7d4e25770bdada4c17f178a5202afeedd6e6233e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 405, + 156, + 413 + ], + "lines": [ + { + "bbox": [ + 144, + 405, + 156, + 413 + ], + "spans": [ + { + "bbox": [ + 144, + 405, + 156, + 413 + ], + "type": "text", + "content": "car" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 178, + 348, + 231, + 403 + ], + "blocks": [ + { + "bbox": [ + 178, + 348, + 231, + 403 + ], + "lines": [ + { + "bbox": [ + 178, + 348, + 231, + 403 + ], + "spans": [ + { + "bbox": [ + 178, + 348, + 231, + 403 + ], + "type": "image", + "image_path": "f435fc08c7eaafab09a1258a44db705153064e5c5b5f28393c45fe1eeba74c0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 405, + 216, + 414 + ], + "lines": [ + { + "bbox": [ + 192, + 405, + 216, + 414 + ], + "spans": [ + { + "bbox": [ + 192, + 405, + 216, + 414 + ], + "type": "text", + "content": "bicycle" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 235, + 348, + 288, + 403 + ], + "blocks": [ + { + "bbox": [ + 235, + 348, + 288, + 403 + ], + "lines": [ + { + "bbox": [ + 235, + 348, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 235, + 348, + 288, + 403 + ], + "type": "image", + "image_path": "98c3eeb4e7e42700db762e69d6d81d45b7d3ae30ea64fe43e37ede6416d3da75.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 405, + 266, + 413 + ], + "lines": [ + { + "bbox": [ + 254, + 405, + 266, + 413 + ], + "spans": [ + { + "bbox": [ + 254, + 405, + 266, + 413 + ], + "type": "text", + "content": "car" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 427, + 295, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 460 + ], + "type": "text", + "content": "Figure 5. Visualization on attention map on multiple objects from different prompts. Adding description of the relative spatial position between objects can improve multi-object generation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 483, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 519 + ], + "type": "text", + "content": "complex compositional T2V tasks. These experiments verified that PAPO significantly improves the performance of generating scenes involving more than two subjects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 528, + 120, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 120, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 120, + 540 + ], + "type": "text", + "content": "4.4. Analyses" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Multiple objects. Synthesis quality of generated videos often declines when tasked with generating outputs that accurately represent prompts involving multiple objects. This issue is also prevalent in the T2I model, and several studies [14, 16, 30] have highlighted that the blended context created by the CLIP text encoder leads to improper binding. Meanwhile, some related works [6, 29] focus on image latents to address information loss, while the others [7, 49] pay more attention to text embedding to deal with the issue. However, few have explored optimizing prompts to improve the performance of multiple obejsts task. We apply our method to text-to-image using SD 1.4 [34], which uses the same text encoder with LaVie [40]. We test on prompts about multiple objects, and remove the irrelevant modifiers" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 335, + 220, + 533, + 345 + ], + "blocks": [ + { + "bbox": [ + 335, + 220, + 533, + 345 + ], + "lines": [ + { + "bbox": [ + 335, + 220, + 533, + 345 + ], + "spans": [ + { + "bbox": [ + 335, + 220, + 533, + 345 + ], + "type": "image", + "image_path": "b7c60393e629fc0f95b39fd870b09183e1102602a5910627e73e01a9059875ec.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 354, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 387 + ], + "type": "text", + "content": "Figure 6. Prompt length distribution comparison among various methods. The distribution of RAPO-optimized prompts is more closer to the training prompts." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 409, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 554, + 457 + ], + "type": "text", + "content": "like action and atmosphere descriptions. As shown in Fig. 5, we can find the relevant spatial descriptions boost the performance of multiple objects. More examples can be found in supplementary materials." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 458, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 601 + ], + "type": "text", + "content": "Statistical analysis of text. As shown in Fig. 6, we compared the word length distributions of prompts from the T2V training set, user prompts (simulated via VBench, EvalCrafter, and T2V-CompBench), and optimized prompts generated by various methods. The results show that the prompt length distribution produced by RAPO is closest to that of the training set, and this consistency unleashes the model's generative potential to produce better videos. In contrast, user prompts are too short and lack necessary details, while other methods generate longer prompts that contain excessive details and complex vocabulary, which may be counterproductive, as shown in Tab. 4 and Tab. 5." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 611, + 408, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 408, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 408, + 624 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": "We conduct ablation experiments on the VBench benchmark to examine the individual and combined effects of different modules in RAPO. Additionally, we perform ablation experiments on various configurations of " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": ". We present more ablation experiments on hyperparameters in the supplementary materials." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Ablating each modules in RAPO. We directly obtain the" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3179" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 188 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 188 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 188 + ], + "type": "table", + "html": "
ModelsFinal Sum Scoremotion qualitytext-video alignmentvisual qualitytemporal consistency
LaVie [40]24853.1969.6064.8160.87
LaVie-GPT4 [2]24654.0565.5164.9661.22
LaVie-Open-sora [1]25153.0771.3865.2661.41
LaVie-RAPO25653.3474.3866.6261.29
Latte [26]21750.0355.4957.6553.94
Latte-GPT4 [2]21851.3653.6558.0254.65
Latte-Open-sora [1]22250.2557.3258.7155.47
Latte-RAPO22751.7360.8659.2455.26
", + "image_path": "f4696df7c111ec210eb665163138ca56a9427654269628de55e179d16535069e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 229, + 552, + 352 + ], + "blocks": [ + { + "bbox": [ + 55, + 195, + 555, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 195, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 555, + 218 + ], + "type": "text", + "content": "Table 5. Quantitative comparisons on EvalCrafter. The results demonstrate that RAPO consistently achieves higher scores, especially in terms of text-video alignment and visual quality, indicating its superior ability to enhance the overall quality of generated videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 229, + 552, + 352 + ], + "lines": [ + { + "bbox": [ + 58, + 229, + 552, + 352 + ], + "spans": [ + { + "bbox": [ + 58, + 229, + 552, + 352 + ], + "type": "table", + "html": "
Modelsconsistent attribute bindingdynamic attribute bindingaction bindingobject interactions
LaVie [40]0.6200.2320.4830.760
LaVie-GPT4 [2]0.5610.2180.4280.620
LaVie-Open-sora [1]0.5320.2140.4700.698
LaVie-RAPO0.6920.2670.6350.839
Latte [26]0.6330.2270.4760.792
Latte-GPT4 [2]0.5980.2100.4050.688
Latte-Open-sora [1]0.5490.2030.4870.743
Latte-RAPO0.7060.2580.5910.856
", + "image_path": "1637b2df5bb09657e372216015ac4b45ca80874686b7d5555348567ae099ad01.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 400, + 294, + 462 + ], + "blocks": [ + { + "bbox": [ + 55, + 360, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 555, + 383 + ], + "type": "text", + "content": "Table 6. Quantitative comparisons on T2V-CompBench. RAPO outperforms alternative prompt optimization methods, exhibiting better performance at managing complex compositional tasks and ensuring coherent interactions between multiple objects in the generated videos." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 400, + 294, + 462 + ], + "lines": [ + { + "bbox": [ + 57, + 400, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 57, + 400, + 294, + 462 + ], + "type": "table", + "html": "
word augmentationsentence refactoringprompt selectionVBatch Total Score
(a)✓80.37%
(b)✓79.75%
(c)✓✓81.58%
(d)✓✓81.75%
(e)✓✓80.60%
(f)✓✓✓82.38%
", + "image_path": "32b53929937085dc5c387d446bea7c0c9475bb2321a10c1e8079c9a62f155e0f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 527, + 294, + 564 + ], + "blocks": [ + { + "bbox": [ + 55, + 470, + 295, + 515 + ], + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 515 + ], + "type": "text", + "content": "Table 7. Ablation studies of different modules in RAPO on VBench. Each module improves performance, while the combined use of all three leads to the highest evaluation score, confirming the synergistic effect of the full RAPO framework." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 527, + 294, + 564 + ], + "lines": [ + { + "bbox": [ + 58, + 527, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 58, + 527, + 294, + 564 + ], + "type": "table", + "html": "
GPT-4MistralLLaMA
VBenchmark Total Score82.38%82.25%82.10%
", + "image_path": "52086a04b74b840671f49c65b27cb6de08525c19c9a0e43b1d02ce725cc2a40b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "content": "Table 8. Ablation studies on different " + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "content": ". The results suggest that RAPO is robust and effective across various LLMs." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 616, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 688 + ], + "type": "text", + "content": "related modifiers about input prompts utilizing GPT-4 [2], and merging them into inputs at one time as the comparison of word augmentation. We randomly select one of optimized prompts as the comparison of prompt selection. The optimal result is achieved by the full-fledged framework as shown in row (f)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Ablation experiments on different " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". We conduct ablation experiments on GPT-4 [2], Mistral [22] and LLaMA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 453 + ], + "type": "text", + "content": "3.1 [38]. As shown in Tab. 8, although GPT-4 achieves the best overall score, the differences are marginal, which suggests that RAPO is robust and effective across various LLMs in generating optimized prompts for T2V generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 480, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 480, + 473 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 480, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 480, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 480, + 555, + 613 + ], + "type": "text", + "content": "In this paper, we propose RAPO, a novel framework for prompt optimizing to improve the static and dynamic dimensions performance of generated videos by pre-trained T2V models. RAPO employs word augmentation, sentence refactoring and prompt selection to generate prompts aligned with the distribution of training prompts, involving relevant and straightforward modifiers about inputs. Extensive experimental evaluations demonstrate the effectiveness and efficiency of RAPO. For the future work, we plan to extend RAPO to more T2V models and conduct more analyses about the reasons for improvements." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 621, + 411, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 411, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 411, + 635 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "The work is partially supported by the National Natural Science Foundation of China (Grant No. 62471287), the National Key R&D Program of China under Grant No. 2022ZD0160102, and the Science and Technology Commission of Shanghai Municipality under Grant No. 23QD1400800." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3180" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 122 + ], + "type": "text", + "content": "[1] Open-sora: Democratizing efficient video production for all. 2024. URL: https://github.com/hpcaitech/Open-Sora. 5, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5,7,8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.opennai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "text", + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3:1, 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 391 + ], + "type": "text", + "content": "[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 446 + ], + "type": "text", + "content": "[7] Chieh-Yun Chen, Li-Wu Tsao, Chiang Tseng, and Hong-Han Shuai. A cat is a cat (not a dog!): Unraveling information mix-ups in text-to-image encoders through causal analysis and embedding optimization. arXiv preprint arXiv:2410.00321, 2024. 1, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 512 + ], + "type": "text", + "content": "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 568 + ], + "type": "text", + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246-22256, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 569, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 634 + ], + "type": "text", + "content": "[10] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 690 + ], + "type": "text", + "content": "[11] Zijie Chen, Lichao Zhang, Fangsheng Weng, Lili Pan, and Zhenzhong Lan. Tailored visions: Enhancing text-to-image generation with personalized prompt rewriting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7727-7736, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 172 + ], + "type": "text", + "content": "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 227 + ], + "type": "text", + "content": "[14] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "type": "text", + "content": "[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "text", + "content": "[16] Boyu Han, Qianqian Xu, Zhiyong Yang, Shilong Bao, Peisong Wen, Yangbangyan Jiang, and Qingming Huang. Aucseg: Auc-oriented pixel-level long-tail semantic segmentation. arXiv preprint arXiv:2409.20398, 2024. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[17] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "text", + "content": "[18] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. *Animate-a-story: Storytelling with retrieval-augmented video generation*. arXiv preprint arXiv:2307.06940, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 553, + 470 + ], + "type": "text", + "content": "[19] Nailei Hei, Qianyu Guo, Zihao Wang, Yan Wang, Haofen Wang, and Wenqiang Zhang. A user-friendly framework for generating model-preferred prompts in text-to-image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2139-2147, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "text", + "content": "[20] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 514, + 555, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 555, + 581 + ], + "type": "text", + "content": "[21] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "text", + "content": "[22] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[23] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024.1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "text", + "content": "[24] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3181" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "text", + "content": "Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 185 + ], + "type": "text", + "content": "[25] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 295, + 229 + ], + "type": "text", + "content": "[26] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 263 + ], + "type": "text", + "content": "[27] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, Ji-Rong Wen, and Qing Yang. Dynamic prompt optimizing for text-to-image generation. In CVPR, 2024. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[28] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 354 + ], + "type": "text", + "content": "[29] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7942, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 409 + ], + "type": "text", + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 295, + 464 + ], + "type": "text", + "content": "[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 510 + ], + "type": "text", + "content": "[32] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. arXiv preprint arXiv:2310.15169, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 679 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "text", + "content": "[36] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "text", + "content": "without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[37] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. arXiv preprint arXiv:2407.14505, 2024. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[38] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[39] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "text", + "content": "[40] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[41] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[42] Mengjiao Yang, Yilun Du, Bo Dai, Dale Schuurmans, Joshua B Tenenbaum, and Pieter Abbeel. Probabilistic adaptation of text-to-video models. arXiv preprint arXiv:2306.01872, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "text", + "content": "[43] Shuai Yang, Jing Tan, Mengchen Zhang, Tong Wu, Yixuan Li, Gordon Wetzstein, Ziwei Liu, and Dahua Lin. Layer-pano3d: Layered 3d panorama for hyper-immersive scene generation. arXiv preprint arXiv:2408.13252, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "text", + "content": "[44] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "text", + "content": "[45] Jingtao Zhan, Qingyao Ai, Yiqun Liu, Jia Chen, and Shaoping Ma. Capability-aware prompt reformulation learning for text-to-image generation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2145-2155, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[46] Jingtao Zhan, Qingyao Ai, Yiqun Liu, Yingwei Pan, Ting Yao, Jiaxin Mao, Shaoping Ma, and Tao Mei. Prompt refinement with image pivot for text-to-image generation. arXiv preprint arXiv:2407.00247, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 645 + ], + "type": "text", + "content": "[47] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1-15, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[48] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[49] Chenyi Zhuang, Ying Hu, and Pan Gao. Magnet: We never know how text-to-image diffusion models work, until we" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3182" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 150 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "text", + "content": "learn how vision-language models function. arXiv preprint arXiv:2409.19967, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 150 + ], + "type": "text", + "content": "[50] Shaobin Zhuang, Kunchang Li, Xinyuan Chen, Yaohui Wang, Ziwei Liu, Yu Qiao, and Yali Wang. Vlogger: Make your dream a vlog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8806-8817, 2024. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3183" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_content_list.json b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..88f163a91c5013259f29a18150e8bcfc78b48335 --- /dev/null +++ b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_content_list.json @@ -0,0 +1,1320 @@ +[ + { + "type": "text", + "text": "The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 120, + 128, + 877, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Naveen George $^{1*}$ , Karthik Nandan Dasaraju $^{2*}$ , Rutheesh Reddy Chittepu $^{1}$ , Konda Reddy Mopuri $^{1}$ $^{1}$ Data-driven Intelligence and Learning Lab, Indian Institute of Technology, Hyderabad \n $^{2}$ Indian Institute of Technology, Ropar", + "bbox": [ + 116, + 202, + 879, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ai23mtech12001@iith.ac.in, 2021csb1081@iitrpr.ac.in, cs21btech11014@iith.ac.in, krmopuri@ai.iith.ac.in", + "bbox": [ + 148, + 266, + 844, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 333, + 326, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image models such as Stable Diffusion, DALL-E, and Midjourney have gained immense popularity lately. However, they are trained on vast amounts of data that may include private, explicit, or copyrighted material used without permission, raising serious legal and ethical concerns. In light of the recent regulations aimed at protecting individual data privacy, there has been a surge in Machine Unlearning methods designed to remove specific concepts from these models. However, we identify a critical flaw in these unlearning techniques: unlearned concepts will revive when the models are fine-tuned, even with general or unrelated prompts. In this paper, for the first time, through an extensive study, we demonstrate the unstable nature of existing unlearning methods in text-to-image diffusion models. We introduce a framework that includes a couple of measures for analyzing the stability of existing unlearning methods. Further, the paper offers preliminary insights into the plausible explanation for the instability of the mapping-based unlearning methods that can guide future research toward more robust unlearning techniques. Codes1 for implementing the proposed framework are provided.", + "bbox": [ + 88, + 364, + 485, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 691, + 220, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The past decade has witnessed a rapid qualitative growth in text-to-image models [7, 23, 24, 32, 33], leading to its widespread attention and adoption. Models such as DALL-E [2, 30, 31], MidJourney [22], and Stable Diffusion [27, 32] enable high-quality image generation from simple text prompts, opening applications across art, design, advertising, gaming, and even medical imaging. With millions worldwide adopting these tools for academic and commercial purposes, their transformative potential is evident, as illustrated by MidJourney's 16 million active", + "bbox": [ + 89, + 715, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "users [22] and the generation of over 2 million images daily by DALL-E and Stable Diffusion [1, 37]. However, this widespread use raises critical concerns about data privacy, security, fairness, and intellectual property, as these models rely on vast datasets that often include sensitive, personal, or copyrighted content [35].", + "bbox": [ + 511, + 335, + 905, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to these privacy concerns, regulations such as the General Data Protection Regulation (GDPR) [28] and the California Consumer Privacy Act (CCPA) [5] provide users the Right to be Forgotten [16], promoting the development of Machine Unlearning (MU) techniques [3, 13]. MU aims to remove specific data or concepts from a model without the computational expense of retraining from scratch. While initially explored in classification models [4, 13, 14, 36], MU in generative models, particularly diffusion-based ones, poses unique challenges due to the models' complex, distribution-based learning. Ensuring these models forget sensitive or harmful information while retaining general functionality remains an open problem.", + "bbox": [ + 511, + 428, + 906, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine Unlearning in text-to-image models focuses on erasing specific concepts from the model's generative capacity. The general goal of MU in this context is to ensure that when a model is conditioned on a prompt related to the target or \"unlearned\" concept, it no longer generates images reflecting that concept. These target concepts can be Not Safe For Work (NSFW) content [11, 15, 17, 21, 34], objects [10-12, 15, 19-21, 39, 41] or artistic styles [11, 12, 15, 19-21, 39, 41]. In other words, the unlearned concept is the one we want the model to \"forget\". To achieve this, unlearning algorithms typically remap the target concept to a surrogate concept. This means that when the model is conditioned on the unlearned concept, it produces images associated with the surrogate concept instead. The surrogate concept can be a general or neutral concept [12, 15, 19], a random concept [10], or even an empty string [11, 39].", + "bbox": [ + 511, + 626, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these efforts in Machine Unlearning, we have identified a critical vulnerability in the current state-of-the", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. \n1https://github.com/NGK2110/TIU", + "bbox": [ + 107, + 875, + 354, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13393", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "art methods for text-to-image models. Specifically, we observe that when these unlearned models undergo finetuning, even with general or unrelated concepts, the previously unlearned concepts can resurface in the generated images. This unintended revival undermines the unlearning process's effectiveness and poses significant risks to data privacy and legal compliance. The ability of an unlearned model to regenerate sensitive or copyrighted content after benign fine-tuning highlights a fundamental weakness in the existing MU approaches for generative models. Such vulnerabilities could lead to unauthorized dissemination of protected content, violating individual privacy rights and intellectual property laws. Therefore, it is imperative to (i) study the robustness of the existing MU methods, (ii) understand the underlying causes of this behaviour and develop more robust unlearning strategies.", + "bbox": [ + 89, + 90, + 480, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions of the paper can be summarized as follows.", + "bbox": [ + 89, + 337, + 470, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Systematic evaluation of MU methods: We conduct a comprehensive study of the existing MU methods in text-to-image diffusion models, demonstrating that unlearned concepts are revived across all examined approaches. Our analysis includes all the leading methods — SalUn [10], ESD [11], EDiff [39], CA [19], MACE [20], SPM [21], SA [15], Receler [17] and UCE [12] —capturing the full spectrum of current state-of-the-art unlearning techniques. To the best of our knowledge, this is the first work to identify and discuss this issue in depth, examining the robustness of unlearning. We believe this work brings an unknown but critical aspect of the MU methods into light, paving the way to a more resilient framework for unlearning.", + "bbox": [ + 89, + 357, + 482, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Concept Revival Framework: We introduce a framework demonstrating that fine-tuning revives the unlearned concepts. For each target concept, the proposed framework presents a customized curriculum of concepts in the increasing order of relevance to fine-tune the unlearned model. During the fine-tuning, the framework enables the measurement of the revival of the target concept via a couple of simple measures: classification accuracy and CLIP score.", + "bbox": [ + 89, + 569, + 482, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Analysis and insights about the unlearning process: We study the fragility of the existing MU methods with respect to two important axes:", + "bbox": [ + 89, + 704, + 482, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Semantic relevance of the surrogate concept to the target concept: we analyze how mapping to surrogate concepts of varying semantic relevance (to the target concept) affects the stability of the resulting unlearning.", + "bbox": [ + 102, + 750, + 482, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Modules affected during the unlearning: we examine how the nature of model parameters modified during unlearning, (i.e., cross-attention module vs. the rest), affects the unlearning.", + "bbox": [ + 102, + 825, + 482, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Further, we hypothesize a plausible explanation for the in", + "bbox": [ + 102, + 885, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stability of the mapping-based MU methods - fine-tuning the unlearned model disrupts the mapping between the target and surrogate concepts. We provide empirical evidence for this hypothesis over some of the popular MU methods [11, 12, 19], in terms of the target concept's unlearning objective function value during the fine-tuning.", + "bbox": [ + 526, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 195, + 663, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlearning in image generation. Machine Unlearning (MU) approaches aim to remove the influence of certain data points or concepts from a model without the need for full retraining, which is computationally intensive for large-scale models like diffusion models. The rise of MU methods has been driven by their potential to address various issues in generative models. For example, text-to-image (T2I) diffusion models have a tendency to generate inappropriate or sensitive content, such as depictions of violence, deepfakes, or copyrighted artistic styles. MU methods are recognized for their role in promoting data privacy and complying with regulations like the \"Right to be Forgotten\" by enabling models to unlearn specific concepts deemed undesirable or unlawful.", + "bbox": [ + 511, + 220, + 903, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several methods have been proposed to achieve concept unlearning in diffusion models. Unified Concept Editing (UCE) [12] modifies cross-attention in the U-Net component of diffusion models using a closed-form solution to map target concepts to surrogate ones, effectively erasing them without additional training. Mass Concept Erasure (MACE) [20] extends this by refining cross-attention layers and utilizing Low-Rank Adaptation (LoRA) modules to efficiently remove multiple unwanted concepts while preserving the model's ability to generate unrelated content.", + "bbox": [ + 511, + 431, + 905, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other approaches like Selective Amnesia (SA) [15] leverage Elastic Weight Consolidation and Generative Replay to forget target concepts while maintaining overall model performance. Erased Stable Diffusion (ESD) [11] aligns the probability distribution of a concept to be erased with that of a null string, although it may lead to a collapse in generative quality. Concept Ablation (CA) [19] fine-tunes cross-attention weights to map target concepts to more general anchor concepts, effectively preventing the generation of specific styles or objects.", + "bbox": [ + 511, + 583, + 905, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods such as Safe Latent Diffusion (SLD) [34] use reverse guidance during generation to steer away from inappropriate content but require careful balancing of prompt conditioning and hyperparameters. The Semi-Permeable Membrane (SPM) [21] adapter offers a lightweight solution for multi-concept erasure by introducing a regulated erasing signal within intermediate layers, although it may struggle with partial unlearning (traces of the target concept remain).", + "bbox": [ + 511, + 734, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While these methods can effectively unlearn the target concept, they often adversely affect the generative capability for related concepts [43] and may degrade the over", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13394", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4cd222d6fe147dcc0ca652140eeb5e7ae897b03affb79eaab2b0312423bd9077.jpg", + "image_caption": [ + "Figure 1. Revival of unlearned concepts is observed across numerous concepts and unlearning methods. The top row presents images generated by the original SDv1.4 model. The middle row shows images from models with specific concepts unlearned. The unlearning concept and the method used are at the bottom of each image. The last row shows images generated when the unlearned model is finetuned on several different concepts. At the bottom of each column is the prompt used for sampling." + ], + "image_footnote": [], + "bbox": [ + 127, + 99, + 869, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "all image quality. Additionally, the unlearned concepts can sometimes reappear if an adversarial prompt is carefully engineered, indicating incomplete unlearning and vulnerabilities to prompt-based attacks [38, 42]. There are also adversarial robust unlearning methods [17] that are trained to withstand these adversarial prompts without compromising the generation of non-target concepts (i.e., locality).", + "bbox": [ + 88, + 488, + 486, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Framework", + "text_level": 1, + "bbox": [ + 89, + 609, + 299, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Fine-tuning revives the unlearned concepts", + "text_level": 1, + "bbox": [ + 89, + 633, + 457, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Numerous methods have been proposed to unlearn a wide range of target concepts from diffusion models. These targets range from objects and artistic styles to global concepts such as nudity. We present the effect of fine-tuning these unlearned models on different concepts in a systematic way. Figure 1 presents the observed effect concisely, where the top row shows images generated by the original Stable Diffusion v1.4. The second row shows images generated by the same model after unlearning different concepts using numerous unlearning methods[10-12, 15, 17, 19-21, 39]. These works demonstrated a good unlearning performance, which is evident from the absence of the unlearning concept in the images. The performance of these works was also quantified by UnlearnCanvas [43] and shown to offer good unlearning accuracy. After fine-tuning various concepts sequentially (Section 4), images generated by the model are", + "bbox": [ + 88, + 659, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "shown in the bottom row of the exact figure. Revival of the unlearned concept is observed in all cases.", + "bbox": [ + 511, + 488, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This revival of unlearned concepts with fine-tuning on arbitrary concept(s) is a critical vulnerability. A model published to have unlearned a sensitive concept can be made to relearn by fine-tuning, posing a threat to the publisher. This is more pressing if the concept being unlearned is nudity, NSFW, or copyrighted content. A model $M_{u}$ trained by some unlearning mechanism is compared with the ideal model $\\overline{M}_{u}$ , which has never seen the unlearning concept before, i.e. the model trained on a dataset with unlearning concept removed. $\\overline{M}_{u}$ that has never known an unlearned concept (golf ball for example), won't infer its existence when fine-tuned on images of different concept (piano keyboard). The revival of the unlearned concept in $M_{u}$ shows that its performance is far from $\\overline{M}_{u}$ , questioning the unlearning mechanism used.", + "bbox": [ + 509, + 520, + 908, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide quantitative results for the revival of unlearned concepts while fine-tuning. CLIP [29] score calculated from the embeddings quantifies the semantic similarity between a text prompt and an image. Figure 2 shows the effect of fine-tuning on unlearned model with an unrelated concept. We prepared prompts containing golf ball as a validation dataset to evaluate the ability of the model to generate images of golf ball. At every epoch of fine-tuning, the model is sampled to generate images for validation prompts and the mean CLIP score is calculated for the", + "bbox": [ + 509, + 750, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13395", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "embeddings of generated images and the prompts. As expected, the score is initially low, proving that the model has unlearned golf ball. An increase in the alignment of generated images towards golf ball with every epoch is observed, depicting the revival of the unlearned concept.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5de54fa76ef4b87998264e27b65b6daf84582e6fd85b256b89dec195ca7e1c38.jpg", + "image_caption": [ + "Figure 2. Graph showing the variation of mean CLIP score with epochs. Fine-tuning is done on an unrelated concept \"piano keyboard\", and an increase is observed indicating the revival of unlearned concept \"Golf Ball\". ESD[39] was used for unlearning." + ], + "image_footnote": [], + "bbox": [ + 94, + 181, + 478, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dependence of revival on fine-tuning data", + "text_level": 1, + "bbox": [ + 89, + 453, + 444, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The previous subsection establishes that fine-tuning with certain unrelated concepts (hereafter referred to as fine-tuning concepts) can revive the unlearned concept. However, we observed that some MU algorithms demonstrated trivial revival compared to others when fine-tuned on the same unrelated concept (Figure 3). It has been observed that the extent of revival depends on the fine-tuning dataset, especially the fine-tuning concept. This makes it a challenge to assess the robustness of an unlearned model. While no revival is observed on a particular fine-tuning dataset, concept(s) that were not considered might exist and cause significant revival.", + "bbox": [ + 88, + 477, + 482, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis of the effect of different fine-tuning concepts is required to draw insights into the revival of unlearned concepts. The extent of fine-tuning (i.e., the number of epochs) required to observe the revival of unlearned concepts can serve as a good indicator of robustness. When the CLIP score crosses a predefined threshold, one may declare the revival of the unlearned concept. Figure 3 shows the results of such an experiment. Unlearned methods SalUn [10] and UCE [12] are considered to unlearn \"golf ball\" and fine-tuning was performed on concepts indicated on the vertical axis. The horizontal axis represents the number of epochs needed to achieve a CLIP score threshold of 0.31 (explained in Section 4). We considered fine-tuning till a maximum of 91 epochs, indicating the existence of fine-tuning concepts that don't cause revival even after prolonged fine-tuning.", + "bbox": [ + 89, + 657, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Figure 3, it can be observed that a wide range of", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1d3fbba208cbcbe981b54f29057bb6bf2344818488fc01c49d5b866839998803.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 227 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f8e48bf49f0722b4dc5e6ebe8f27339befeef115de312d0eb19ae71447422626.jpg", + "image_caption": [ + "Figure 3. Illustration of the number of epochs of fine-tuning required to observe revival of unlearned concept \"golf ball\". Finetuning concepts are shown on the vertical axis. The same finetuning concept can have different effects on revival for different unlearning methods (UCE and SalUn shown)" + ], + "image_footnote": [], + "bbox": [ + 517, + 243, + 903, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fine-tuning concepts may be required to study the revival of unlearned concepts. A fine-tuning concept that caused revival for one work may not be prominent in revival for others. Further, we also observed that sequentially fine-tuning the unlearned model shows greater revival in a shorter time (number of fine-tuning epochs). Comparing Figure 3 (bottom panel) with Figure 4 makes this evident. Figure 4 shows the mean CLIP score when fine-tuning is performed in a sequential manner on concepts mentioned on the horizontal axis. A model that unlearned \"golf ball\" with SalUn [10] method was fine-tuned with each concept for 8 epochs sequentially (after fine-tuning on one concept, we continued fine-tuning on the next). While in Figure 3 (bottom panel), trivial revival was observed when fine-tuned individually on \"Volcano, Fish, Swimming, Hot Dog, Cricket Bat\" (indicated by large number of fine-tuning epochs), a CLIP score of 0.31 was reached in Figure 4 when fine-tuned sequentially. Specifically, for the fine-tuning concept \"Cricket Bat\", no revival was observed even after fine-tuning for a large number of epochs (91 in Figure 3 bottom panel). However, a significant revival (0.31 CLIP score) was observed at the end of fine-tuning \"Cricket Bat\" in a sequential manner for relatively less fine-tuning ( $5 \\times 8 = 40$ epochs in Figure 3).", + "bbox": [ + 511, + 506, + 906, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, fine-tuning with various prompts may be required to study the revival of unlearning concepts. Further, the", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13396", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/48741fadc55cba1a912060307483e4a024c730de134ec18cb5be66b8dab2b5e5.jpg", + "image_caption": [ + "Figure 4. Graph showing Mean CLIP score after fine-tuning on the concepts given on horizontal axis sequentially. The unlearned model used was produced by using SalUn[10] to remove target concept \"Golf Ball\". The highlighted point is when fine-tuning is done sequentially until Cricket Bat, and significant revival is shown as indicated by the CLIP score." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 473, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sequential fine-tuning of concepts seems to accelerate the revival. An experimental \"Concept revival framework\" to evaluate MU methods on this basis is detailed in section 4.", + "bbox": [ + 89, + 344, + 482, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments - Studying the Vulnerability", + "text_level": 1, + "bbox": [ + 89, + 402, + 459, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We investigate the critical vulnerability aforementioned in the current Machine Unlearning (MU) methods for text-to-image diffusion models. Specifically, we demonstrate through extensive experiments that fine-tuning an unlearned model can lead to reviving previously unlearned concepts.", + "bbox": [ + 89, + 429, + 483, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 513, + 281, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted experiments on state-of-the-art MU methods applied to the Stable Diffusion v1.4 model to systematically study the concept revival. We describe our experimental setup:", + "bbox": [ + 89, + 537, + 482, + 598 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Unlearning: We conducted experiments in four different unlearning settings, covering a total of ten concepts. In each case, unlearning is crucial for privacy, security, and ethical considerations, as it ensures certain content can no longer be generated by the model.", + "- Object Unlearning focuses on unlearning specific objects from the model. Experiments are performed on \"Golf Ball, Fukushima and Dog\".", + "- Style Unlearning aims to remove the ability to replicate certain artistic styles using the model. \"Van Gogh, Picasso\" and \"Cartoon\" styles are used in our experiments.", + "- Celebrity Unlearning aims to erase the model's ability to generate content of famous celebrities. We decided to use \"Angelina Jolie, Brad Pitt\" and \"Lionel Messi\".", + "- NSFW Content Unlearning focuses on preventing the model from generating inappropriate or explicit content.", + "- Evaluation: A set of validation prompts is composed that contains the unlearning concept. The model is then sam" + ], + "bbox": [ + 91, + 598, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pled to generate images and evaluated using two main measures.", + "bbox": [ + 526, + 90, + 903, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- CLIP Score (CS): Measures the cosine similarity between the embeddings of generated images and their corresponding text prompts using the CLIP model [29].", + "bbox": [ + 526, + 121, + 903, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Classifier Accuracy (Acc): Utilizes a fine-tuned Vision Transformer (ViT) [8] as a binary classifier to detect the presence of the unlearned concept in generated images. For Object and Celebrity Unlearning, we fine-tuned a VIT model on images of unlearning concepts and other unrelated images generated using the Stable Diffusion-3-medium [9]. In the case of Style Unlearning, we use the classifier from UnlearnCanvas [43] and fine-tune it using images generated by Stable Diffusion-3-medium [9]. Finally, for NSFW unlearning, we employ the NudeNet classifier [6].", + "bbox": [ + 527, + 167, + 906, + 333 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Revival Point: Tracking the instance when a significant revival of the unlearned concept is observed can serve as a good measure for assessing the robustness of the unlearning method. We record the number of concepts on which fine-tuning was performed sequentially until revival was observed as \"Revival Point\". Further, to show no bias towards an unlearning method, we fix the order of fine-tuning concepts in the increasing order of semantic similarity from the target concept. The model is fine-tuned on each concept for 8 epochs during this process. A concept is considered to have revived when the model's CLIP score exceeds a specific threshold, and the classifier accuracy surpasses 0.3 (a conservative value that indicates revival). If the fine-tuning fails to revive the target concept in this setup, the MU method is comparatively robust (indicated by R). To calculate the clip threshold required for each target concept, we used images generated from base Stable-Diffusion v1.4 on prompts from the validation dataset (detailed later). Then, we calculated the CLIP score between the image and prompt embeddings. We set the CLIP threshold for to a value 0.02 less than the SD's CLIP score. CLIP threshold for each unlearning concept will be different, depending on the original model's CLIP score (thresholds for each concept are in Supplementary section B). In all the cases where the threshold was reached, a significant revival of the unlearned concept was observed.", + "- Fine-tuning Dataset: To fine-tune the unlearned model, we create a customized dataset by first identifying concepts semantically related to the target concept. We achieve this by designing detailed prompts that specify the type of relationships at each hierarchical level and input these prompts into the OpenAI GPT 4o model. This process generates a comprehensive list of related concepts, which we then use to generate diverse prompts for fine-tuning the model. We compute the cosine" + ], + "bbox": [ + 511, + 347, + 908, + 891 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13397", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/30b8f9f60a4d0d7ff40049f1a063d6a9e7b824e24c1528a815f0563f5134162f.jpg", + "image_caption": [ + "Clip Scores Before and After Fine-tuning" + ], + "image_footnote": [], + "bbox": [ + 94, + 107, + 903, + 284 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e33a0b85bbe7c76714c3209f1cce31015e988742f162f9d0ac18e149d0f71939.jpg", + "image_caption": [ + "Classifier Accuracy Before and After Fine-tuning", + "Figure 5. Graphs illustrate the revival of Unlearned concepts on fine-tuning across different concepts and methods. CLIP score and Classifier accuracy are used to quantify the revival. Each colour corresponds to a different method, with lighter bar segments representing the increase in metrics before and after Fine-tuning, respectively. The consistent upward shifts in bar heights indicate that the unlearned concepts have been successfully revived." + ], + "image_footnote": [], + "bbox": [ + 101, + 311, + 903, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "similarity between these concepts and the target concept using OpenAI embeddings [26] and arrange them in increasing order of similarity. For each selected concept, we generate 50 prompts using OpenAI GPT [25] and generate corresponding images using Stable Diffusion 3 Medium. This results in a dataset that enables curriculum fine-tuning, in which more general concepts appear before the more related concepts during the fine-tuning. This approach allows us to assess how fine-tuning concepts with different semantic similarities affects the revival of the unlearned concept.", + "bbox": [ + 102, + 607, + 485, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Validation Dataset: We created a custom dataset of 20 prompts related to the unlearned concept using the OpenAI API. These prompts were carefully crafted to cover various descriptions associated with the target concept. We generated 5 images for each prompt using the fine-tuned models, resulting in a total of 100 images over which the CLIP score and classification accuracy mea", + "bbox": [ + 89, + 787, + 485, + 897 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sures are computed.", + "bbox": [ + 527, + 608, + 661, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additional details of the experimental setup are provided in the Supplementary Materials (see Supplementary Sections C and D).", + "bbox": [ + 511, + 623, + 906, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Revival in the state-of-the-art MU methods", + "text_level": 1, + "bbox": [ + 511, + 680, + 879, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the explanations in subsection 3.2, we sequentially fine-tune the unlearned model for 10 concepts. We report the two measures, CLIP score and Classifier accuracy, before and after the fine-tuning. An increase in these indicates the revival of unlearned concepts.", + "bbox": [ + 511, + 703, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From Figure 5, it is evident that revival is observed across different concepts and unlearning methods. We have observed that MACE[20] and UCE[12] were relatively robust across many concepts, with less revival observed after fin-tuning. Upon further analysis of different types of unlearning, UCE[12] was found to be robust in object unlearning. We found almost all methods vulnerable to fine-tuning in case of style unlearning. We can infer that artistic styles like", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13398", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Van Gogh and Picasso are inherently complex and challenging to unlearn precisely. In the case of celebrity unlearning, we observed that there is an increase in metrics across all works with UCE's[12] change being minimal. Nudity unlearning was easily broken with relatively faster revival in the case of ESD[11], SPM[21] and SA[15] compared to MACE[20] which was robust. Even with adversarial robust MU method[17] which is robust against adversarial prompts we can observe concept revival (low for object and style but high for NSFW content). More experimental details are provided in the Supplementary Material (see Supplementary Section B).", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Framework for Studying the Concept Revival", + "text_level": 1, + "bbox": [ + 89, + 285, + 475, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on previous sections, it is apparent that model robustness should be assessed with its ability to safeguard against the revival of unlearning concepts. This section details the proposed framework for investigating the concept revival in the unlearned diffusion model due to fine-tuning.", + "bbox": [ + 89, + 309, + 480, + 385 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Step 1: Fine-tune the unlearned model sequentially on a semantically distant concept for a fixed number of epochs.", + "2. Step 2: Evaluate whether the CLIP score and classifier accuracy for the unlearned concept crosses a predefined threshold called the Revival Point." + ], + "bbox": [ + 91, + 388, + 482, + 478 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- If Not: Proceed to fine-tune on a semantically closer concept.", + "- If Yes: Record the Revival Point at which the unlearned concept resurfaces." + ], + "bbox": [ + 109, + 479, + 482, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3. Step 3: Repeat the process by gradually decreasing the semantic distance.", + "bbox": [ + 89, + 541, + 482, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This framework provides a quantitative measure of the model's robustness based on the point at which the unlearned concept reappears and requires less number of epochs compared to fine-tuning separately using each concept.", + "bbox": [ + 89, + 575, + 482, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Impact of Surrogate Concept", + "text_level": 1, + "bbox": [ + 89, + 664, + 349, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our empirical studies indicate that mapping the unlearned concept to a semantically distant or general concept enhances the model's robustness.", + "bbox": [ + 89, + 688, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 presents the time taken for the unlearned concept to resurface when mapped to two concepts at varying semantic similarity. The experimental results demonstrate that mapping an unlearned concept (Golf ball) to a semantically distant concept (Outdoor Recreational activities) enhances the robustness of unlearning. Here, we experimented with the two methods UCE [12] and MACE [20]; clip score (CS) increased only slightly when mapping to the unrelated concept, with minimal or moderate gains in accuracy after unlearning. This indicates that mapping to a semantically distant concept disrupts latent associations more", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "effectively, reducing the likelihood of the unlearned concept re-emerging during fine-tuning.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conversely, mapping to a semantically closer concept (Tennis ball) retained certain feature similarities; for instance, subtle characteristics like the colour of the Golf ball shifted towards a light green due to the influence of the Tennis ball, while other properties associated with Golf balls remained intact. Even after fine-tuning, this greenish tint persisted, highlighting that the concept is still fundamentally a Golf ball but with minor alterations not reflected in the quantitative metrics. This observation underscores the efficacy of semantically distant mappings for achieving robust concept unlearning in generative models.", + "bbox": [ + 511, + 126, + 906, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eb8438838cd0611da9c4ab1cdbeb8d7bb8333fb4fe5ca16544cbb68b80ff3c9f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBefore FTAfter FTRevival Point
CSAccCSAcc
UCE-T0.330.50.330.71
UCE-O0.260.010.280.11R
MACE-T0.320.00.340.543
MACE-O0.270.190.300.15R
SPM-T0.320.270.320.473
SPM-O0.320.290.330.534
", + "bbox": [ + 532, + 313, + 887, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Clip Score (CS) and Accuracy (Acc) Before and After Fine-tuning When Mapping Golf Ball to Tennis Ball (T) (Semantically Close) and Outdoor Recreational Activities (O) (Semantically Distant).", + "bbox": [ + 511, + 449, + 903, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Impact of Modules affected during Unlearning", + "text_level": 1, + "bbox": [ + 511, + 550, + 903, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We chose ESD as the basis for this experiment because it utilizes a setup that allows unlearning by modifying only specific modules (like only cross attention, only non-cross attention and full parameters), enabling targeted analysis of robustness in concept unlearning.", + "bbox": [ + 511, + 578, + 905, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experiments, we used the ESD variations of - ESD-x (cross-attention layers), ESD-u (unconditional layers), and ESD-full (all parameters). We observed that the specific parameters targeted during unlearning significantly impacted the robustness of the concept erasure. Results (Table 2) indicate that cross-attention-based unlearning (ESD-x) is less robust, as the unlearned concept revived more quickly compared to ESD-u and ESD-full. This observation suggests that editing cross-attention weights alone may be insufficient for durable unlearning, as these layers maintain strong associations between prompts and visual features. This vulnerability is also evident in the UCE and CA model, which exclusively modifies cross-attention weights, making it relatively fragile to fine-tuning. Therefore, targeting a broader parameter set, as seen in ESD-u and ESD-full, offers a more resilient approach to concept unlearning.", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13399", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5c877a941aac150f88cbe5877a5a4c8ba4f8a63880a56e3d9eaa24768bd19225.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBefore FTAfter FTRevival Point
CSAccCSAcc
ESD-x0.260.00.320.544
ESD-u0.260.00.330.365
ESD-full0.250.00.320.397
", + "bbox": [ + 114, + 88, + 460, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion - Plausible Mitigation Strategy", + "text_level": 1, + "bbox": [ + 89, + 273, + 464, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For unlearning methods that rely on minimizing a dedicated objective (referred to as the unlearning loss), its value acts as a metric to gauge the success of the erasure. To investigate the impact of fine-tuning on this loss, we evaluated three unlearning techniques—UCE [12], ESD [11], and CA [19]—measuring the unlearning loss both immediately after unlearning and after subsequent fine-tuning (as described in section 4). The results, summarized in Table 3, reveal a consistent increase in unlearning loss post fine-tuning, indicating that fine-tuning destabilizes the unlearning process. Notice that in the case of the CA method with \"Angelina Jolie\" as the target concept and \"Woman\" as the surrogate concept, the unlearning loss did not increase after fine-tuning. This suggests no significant change in the extent of unlearning, which is consistent with previous experiments (Fig 5), where fine-tuning resulted in negligible changes to the model's unlearning accuracy for this particular target/surrogate concept pairing.", + "bbox": [ + 88, + 311, + 485, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A plausible approach to mitigate the revival could involve meta-learning and gradient alignment techniques. While standard fine-tuning minimizes the diffusion model's loss, its gradients may conflict with the unlearning objective. To resolve this, we propose modifying the fine-tuning process to (1) compute gradients for both the diffusion loss and unlearning objective, (2) project the diffusion gradients orthogonal to the unlearning gradient direction, and (3) update parameters using only the orthogonal component. This ensures that the unlearning loss remains unchanged throughout fine-tuning, preserving the unlearning effect. A similar approach has been employed by Ko et al. [18] and Wu et al. [40] to achieve unlearning while maintaining the generational quality of neutral concepts. Their approach optimizes two objectives—one for preserving the generation quality of neutral concepts and another for unlearning the target concept by ensuring that each update step doesn't increase either objective function.", + "bbox": [ + 89, + 613, + 485, + 886 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/48eacd7f9cebdf85317029324b3ed0d9166380b50f49777859a020726a484831.jpg", + "table_caption": [ + "Table 2. Comparison of Clip Score (CS) and Accuracy (Acc) Before and After Unlearning for Different ESD Module Configurations." + ], + "table_footnote": [], + "table_body": "
MethodUnlearn LossUnlearn + FTTarget / Surrogate
UCE5.54 × 1016.74 × 102Dog / Cat
1.08 × 1061.65 × 106Van Gogh / Art
6.66 × 1011.03 × 103Pikachu / Cat
9.02 × 1051.23 × 106Cartoon / Art
1.23 × 1029.41 × 102Brad Pitt / Man
ESD5.00 × 10-41.10 × 10-3Golf Ball / NA
3.00 × 10-45.00 × 10-4Van Gogh / NA
4.00 × 10-49.00 × 10-4Nudity / NA
3.00 × 10-41.60 × 10-3Brad Pitt / NA
CA4.79 × 10-47.53 × 10-4Golf Ball / Ball
1.05 × 10-41.49 × 10-4Van Gogh / Art
≈ 0≈ 0Ang. Jolie / Woman
", + "bbox": [ + 522, + 88, + 903, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison of unlearning loss values between: (1) models immediately after unlearning, and (2) the same models after fine-tuning (using weights obtained from Section 4 experiments). Loss values were computed with each method's respective objective function, with corresponding target-surrogate concept pairings shown. The parameter group updated in ESD and CA for these experiments is cross-attention. The last column shows the target concept (Unlearned concept) along with the corresponding mapping concept i.e Surrogate concept. NA means No specific Surrogate Concept.", + "bbox": [ + 511, + 303, + 906, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 467, + 640, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a critical vulnerability in the current Machine Unlearning techniques for text-to-image diffusion models. Our study emphasizes the need for (i) more robust techniques that can erase the semantic space of the target concept in the latent space of the model, in which case finetuning can never revive the target concept, or/and (ii) the method should be able to achieve mapping of target concept to a surrogate concept which is robust to finetuning. By revealing the weaknesses of existing methods against finetuning and providing a framework to measure concept revival, we aim to guide future research toward developing more secure, reliable, and robust unlearning mechanisms. Below, We list some interesting and important directions to further our understanding of the unlearning process.", + "bbox": [ + 511, + 492, + 906, + 704 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Rigorous understanding of why the finetuning process revives the unlearned concepts across several unlearning approaches. Though challenging, it is of utmost importance to explain this vulnerability of the existing MU methods and to achieve robust unlearning.", + "- Relation between concept revival and finetuning concept. We have observed that certain finetuning concepts can cause greater revival while others less. Also, this dependence may vary across unlearned concepts and unlearning methods. Researching the primary cause of the variance in dependency between the finetuning concept and the revival of the unlearning concept can aid in assessing the robustness of the models better." + ], + "bbox": [ + 511, + 704, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13400", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stability AI. Stability ai. https://stability.ai/news/celebrating-one-year-of-stable-diffusion, 2023. 1", + "[2] Justin Betker, Gabriel Goh, Li Jing, Tim Brooks, Jason Wang, Lilian Li, Aditya Ramesh, et al. Improving image generation with better captions. Computer Science, 2(3):8, 2023. 1", + "[3] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy (SP), pages 463-480. IEEE, 2015. 1", + "[4] Mingjie Chen, Wei Gao, Guangyue Liu, Kai Peng, and Cunhua Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7766-7775, 2023. 1", + "[5] Lydia de la Torre. A guide to the california consumer privacy act of 2018. Available at SSRN 3275571, 2018. 1", + "[6] NudeNet Developers. Nudenet: A nudity detection and censoring library. https://github.com/notAI-tech/NudeNet, 2023.5", + "[7] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5", + "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 5", + "[10] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 4, 5", + "[11] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 7, 8", + "[12] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 3, 4, 6, 7, 8", + "[13] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9301-9309. IEEE, 2020. 1", + "[15] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 7", + "[16] Chris Jay Hoofnagle, Bart van der Sloot, and Frederik Zuiderveen Borgesius. The european union general data protection regulation: what it is and what it means. Information & Communications Technology Law, 28(1):65-98, 2019. 1", + "[17] Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung-Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. Receiver: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. In European Conference on Computer Vision, pages 360–376. Springer, 2024. 1, 2, 3, 7", + "[18] Myeongseob Ko, Henry Li, Zhun Wang, Jonathan Patsenker, Jiachen T. Wang, Qinbin Li, Ming Jin, Dawn Song, and Ruoxi Jia. Boosting alignment for post-unlearning text-to-image generative models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 8", + "[19] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 3, 8", + "[20] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 2, 6, 7", + "[21] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 1, 2, 3, 7", + "[22] Midjourney. Midjourney. https://www.midjourney.com/, 2023.1", + "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pooja Shyam, Pavel Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1", + "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1", + "[25] OpenAI. Chatgpt model api, 2023. 6", + "[26] OpenAI. Openai embeddings, 2023. Accessed: 2024-11-13. 6", + "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[28] Formerly Data Protection. General data protection regulation (gdpr). Intersoft Consulting, accessed in October, 2018. 1", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5", + "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1", + "[31] Aditya Ramesh, Prafulla Dhariwal, Alexander Quinn Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1", + "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Sahand Mahdavi, Raphael Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[34] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2", + "[35] Christoph Schuhmann, Robin Beaumont, Romain Vencu, Craig Gordon, Ross Wightman, Mehdi Cherti, Teddy Coombes, Abhay Katta, and Vladimir Kirillov. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278–25294, 2022. 1", + "[36] Anish Sekhari, Jayadev Acharya, Gauri Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. In Advances in Neural Information Processing Systems, pages 18075-18086, 2021. 1", + "[37] GenAI statistics. Genai statistics. https://journaleverypixel.com/ai-image-statistics, 2023.1", + "[38] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? In The Twelfth International Conference on Learning Representations, 2024. 3", + "[39] Jing Wu, Trung Le, Munawar Hayat, and Mehrtash Harandi. Erasediff: Erasing data influence in diffusion models, 2024. 1, 2, 3, 4" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Yongliang Wu, Shiji Zhou, Mingzhuo Yang, Lianzhe Wang, Heng Chang, Wenbo Zhu, Xinting Hu, Xiao Zhou, and Xu Yang. Unlearning concepts in diffusion model via concept domain correction and concept preserving gradient. In Proceedings of the 39th AAAI Conference on Artificial Intelligence (AAAI), 2025. 8", + "[41] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1755-1764, 2024. 1", + "[42] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now. In Computer Vision - ECCV 2024: 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LVII, page 385-403, Berlin, Heidelberg, 2024. Springer-Verlag. 3", + "[43] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. In Thirty-eighth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2, 3, 5" + ], + "bbox": [ + 516, + 90, + 903, + 441 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13402", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_model.json b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1a023217801d8fbad1b0097c8489d83ce900e143 --- /dev/null +++ b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_model.json @@ -0,0 +1,1936 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.13, + 0.878, + 0.175 + ], + "angle": 0, + "content": "The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.203, + 0.88, + 0.258 + ], + "angle": 0, + "content": "Naveen George\\(^{1*}\\), Karthik Nandan Dasaraju\\(^{2*}\\), Rutheesh Reddy Chittepu\\(^{1}\\), Konda Reddy Mopuri\\(^{1}\\) \n\\(^{1}\\)Data-driven Intelligence and Learning Lab, Indian Institute of Technology, Hyderabad \n\\(^{2}\\)Indian Institute of Technology, Ropar" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.267, + 0.846, + 0.299 + ], + "angle": 0, + "content": "ai23mtech12001@iith.ac.in, 2021csb1081@iitrpr.ac.in, cs21btech11014@iith.ac.in, krmopuri@ai.iith.ac.in" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.334, + 0.327, + 0.351 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.486, + 0.685 + ], + "angle": 0, + "content": "Text-to-image models such as Stable Diffusion, DALL-E, and Midjourney have gained immense popularity lately. However, they are trained on vast amounts of data that may include private, explicit, or copyrighted material used without permission, raising serious legal and ethical concerns. In light of the recent regulations aimed at protecting individual data privacy, there has been a surge in Machine Unlearning methods designed to remove specific concepts from these models. However, we identify a critical flaw in these unlearning techniques: unlearned concepts will revive when the models are fine-tuned, even with general or unrelated prompts. In this paper, for the first time, through an extensive study, we demonstrate the unstable nature of existing unlearning methods in text-to-image diffusion models. We introduce a framework that includes a couple of measures for analyzing the stability of existing unlearning methods. Further, the paper offers preliminary insights into the plausible explanation for the instability of the mapping-based unlearning methods that can guide future research toward more robust unlearning techniques. Codes1 for implementing the proposed framework are provided." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.692, + 0.222, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.484, + 0.868 + ], + "angle": 0, + "content": "The past decade has witnessed a rapid qualitative growth in text-to-image models [7, 23, 24, 32, 33], leading to its widespread attention and adoption. Models such as DALL-E [2, 30, 31], MidJourney [22], and Stable Diffusion [27, 32] enable high-quality image generation from simple text prompts, opening applications across art, design, advertising, gaming, and even medical imaging. With millions worldwide adopting these tools for academic and commercial purposes, their transformative potential is evident, as illustrated by MidJourney's 16 million active" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.906, + 0.427 + ], + "angle": 0, + "content": "users [22] and the generation of over 2 million images daily by DALL-E and Stable Diffusion [1, 37]. However, this widespread use raises critical concerns about data privacy, security, fairness, and intellectual property, as these models rely on vast datasets that often include sensitive, personal, or copyrighted content [35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.626 + ], + "angle": 0, + "content": "In response to these privacy concerns, regulations such as the General Data Protection Regulation (GDPR) [28] and the California Consumer Privacy Act (CCPA) [5] provide users the Right to be Forgotten [16], promoting the development of Machine Unlearning (MU) techniques [3, 13]. MU aims to remove specific data or concepts from a model without the computational expense of retraining from scratch. While initially explored in classification models [4, 13, 14, 36], MU in generative models, particularly diffusion-based ones, poses unique challenges due to the models' complex, distribution-based learning. Ensuring these models forget sensitive or harmful information while retaining general functionality remains an open problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Machine Unlearning in text-to-image models focuses on erasing specific concepts from the model's generative capacity. The general goal of MU in this context is to ensure that when a model is conditioned on a prompt related to the target or \"unlearned\" concept, it no longer generates images reflecting that concept. These target concepts can be Not Safe For Work (NSFW) content [11, 15, 17, 21, 34], objects [10-12, 15, 19-21, 39, 41] or artistic styles [11, 12, 15, 19-21, 39, 41]. In other words, the unlearned concept is the one we want the model to \"forget\". To achieve this, unlearning algorithms typically remap the target concept to a surrogate concept. This means that when the model is conditioned on the unlearned concept, it produces images associated with the surrogate concept instead. The surrogate concept can be a general or neutral concept [12, 15, 19], a random concept [10], or even an empty string [11, 39]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Despite these efforts in Machine Unlearning, we have identified a critical vulnerability in the current state-of-the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.355, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. \n1https://github.com/NGK2110/TIU" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13393" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.333 + ], + "angle": 0, + "content": "art methods for text-to-image models. Specifically, we observe that when these unlearned models undergo finetuning, even with general or unrelated concepts, the previously unlearned concepts can resurface in the generated images. This unintended revival undermines the unlearning process's effectiveness and poses significant risks to data privacy and legal compliance. The ability of an unlearned model to regenerate sensitive or copyrighted content after benign fine-tuning highlights a fundamental weakness in the existing MU approaches for generative models. Such vulnerabilities could lead to unauthorized dissemination of protected content, violating individual privacy rights and intellectual property laws. Therefore, it is imperative to (i) study the robustness of the existing MU methods, (ii) understand the underlying causes of this behaviour and develop more robust unlearning strategies." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.338, + 0.472, + 0.352 + ], + "angle": 0, + "content": "Contributions of the paper can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.358, + 0.483, + 0.569 + ], + "angle": 0, + "content": "- Systematic evaluation of MU methods: We conduct a comprehensive study of the existing MU methods in text-to-image diffusion models, demonstrating that unlearned concepts are revived across all examined approaches. Our analysis includes all the leading methods — SalUn [10], ESD [11], EDiff [39], CA [19], MACE [20], SPM [21], SA [15], Receler [17] and UCE [12] —capturing the full spectrum of current state-of-the-art unlearning techniques. To the best of our knowledge, this is the first work to identify and discuss this issue in depth, examining the robustness of unlearning. We believe this work brings an unknown but critical aspect of the MU methods into light, paving the way to a more resilient framework for unlearning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.57, + 0.483, + 0.704 + ], + "angle": 0, + "content": "- Concept Revival Framework: We introduce a framework demonstrating that fine-tuning revives the unlearned concepts. For each target concept, the proposed framework presents a customized curriculum of concepts in the increasing order of relevance to fine-tune the unlearned model. During the fine-tuning, the framework enables the measurement of the revival of the target concept via a couple of simple measures: classification accuracy and CLIP score." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.75 + ], + "angle": 0, + "content": "- Analysis and insights about the unlearning process: We study the fragility of the existing MU methods with respect to two important axes:" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.751, + 0.483, + 0.826 + ], + "angle": 0, + "content": "- Semantic relevance of the surrogate concept to the target concept: we analyze how mapping to surrogate concepts of varying semantic relevance (to the target concept) affects the stability of the resulting unlearning." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.826, + 0.483, + 0.886 + ], + "angle": 0, + "content": "- Modules affected during the unlearning: we examine how the nature of model parameters modified during unlearning, (i.e., cross-attention module vs. the rest), affects the unlearning." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Further, we hypothesize a plausible explanation for the in" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "stability of the mapping-based MU methods - fine-tuning the unlearned model disrupts the mapping between the target and surrogate concepts. We provide empirical evidence for this hypothesis over some of the popular MU methods [11, 12, 19], in terms of the target concept's unlearning objective function value during the fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.196, + 0.664, + 0.212 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.905, + 0.432 + ], + "angle": 0, + "content": "Unlearning in image generation. Machine Unlearning (MU) approaches aim to remove the influence of certain data points or concepts from a model without the need for full retraining, which is computationally intensive for large-scale models like diffusion models. The rise of MU methods has been driven by their potential to address various issues in generative models. For example, text-to-image (T2I) diffusion models have a tendency to generate inappropriate or sensitive content, such as depictions of violence, deepfakes, or copyrighted artistic styles. MU methods are recognized for their role in promoting data privacy and complying with regulations like the \"Right to be Forgotten\" by enabling models to unlearn specific concepts deemed undesirable or unlawful." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Several methods have been proposed to achieve concept unlearning in diffusion models. Unified Concept Editing (UCE) [12] modifies cross-attention in the U-Net component of diffusion models using a closed-form solution to map target concepts to surrogate ones, effectively erasing them without additional training. Mass Concept Erasure (MACE) [20] extends this by refining cross-attention layers and utilizing Low-Rank Adaptation (LoRA) modules to efficiently remove multiple unwanted concepts while preserving the model's ability to generate unrelated content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.735 + ], + "angle": 0, + "content": "Other approaches like Selective Amnesia (SA) [15] leverage Elastic Weight Consolidation and Generative Replay to forget target concepts while maintaining overall model performance. Erased Stable Diffusion (ESD) [11] aligns the probability distribution of a concept to be erased with that of a null string, although it may lead to a collapse in generative quality. Concept Ablation (CA) [19] fine-tunes cross-attention weights to map target concepts to more general anchor concepts, effectively preventing the generation of specific styles or objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Methods such as Safe Latent Diffusion (SLD) [34] use reverse guidance during generation to steer away from inappropriate content but require careful balancing of prompt conditioning and hyperparameters. The Semi-Permeable Membrane (SPM) [21] adapter offers a lightweight solution for multi-concept erasure by introducing a regulated erasing signal within intermediate layers, although it may struggle with partial unlearning (traces of the target concept remain)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "While these methods can effectively unlearn the target concept, they often adversely affect the generative capability for related concepts [43] and may degrade the over" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13394" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.101, + 0.87, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.406, + 0.908, + 0.465 + ], + "angle": 0, + "content": "Figure 1. Revival of unlearned concepts is observed across numerous concepts and unlearning methods. The top row presents images generated by the original SDv1.4 model. The middle row shows images from models with specific concepts unlearned. The unlearning concept and the method used are at the bottom of each image. The last row shows images generated when the unlearned model is finetuned on several different concepts. At the bottom of each column is the prompt used for sampling." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.489, + 0.487, + 0.598 + ], + "angle": 0, + "content": "all image quality. Additionally, the unlearned concepts can sometimes reappear if an adversarial prompt is carefully engineered, indicating incomplete unlearning and vulnerabilities to prompt-based attacks [38, 42]. There are also adversarial robust unlearning methods [17] that are trained to withstand these adversarial prompts without compromising the generation of non-target concepts (i.e., locality)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.3, + 0.628 + ], + "angle": 0, + "content": "3. Proposed Framework" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.635, + 0.458, + 0.653 + ], + "angle": 0, + "content": "3.1. Fine-tuning revives the unlearned concepts" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Numerous methods have been proposed to unlearn a wide range of target concepts from diffusion models. These targets range from objects and artistic styles to global concepts such as nudity. We present the effect of fine-tuning these unlearned models on different concepts in a systematic way. Figure 1 presents the observed effect concisely, where the top row shows images generated by the original Stable Diffusion v1.4. The second row shows images generated by the same model after unlearning different concepts using numerous unlearning methods[10-12, 15, 17, 19-21, 39]. These works demonstrated a good unlearning performance, which is evident from the absence of the unlearning concept in the images. The performance of these works was also quantified by UnlearnCanvas [43] and shown to offer good unlearning accuracy. After fine-tuning various concepts sequentially (Section 4), images generated by the model are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.519 + ], + "angle": 0, + "content": "shown in the bottom row of the exact figure. Revival of the unlearned concept is observed in all cases." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.521, + 0.909, + 0.748 + ], + "angle": 0, + "content": "This revival of unlearned concepts with fine-tuning on arbitrary concept(s) is a critical vulnerability. A model published to have unlearned a sensitive concept can be made to relearn by fine-tuning, posing a threat to the publisher. This is more pressing if the concept being unlearned is nudity, NSFW, or copyrighted content. A model \\( M_{u} \\) trained by some unlearning mechanism is compared with the ideal model \\( \\overline{M}_{u} \\), which has never seen the unlearning concept before, i.e. the model trained on a dataset with unlearning concept removed. \\( \\overline{M}_{u} \\) that has never known an unlearned concept (golf ball for example), won't infer its existence when fine-tuned on images of different concept (piano keyboard). The revival of the unlearned concept in \\( M_{u} \\) shows that its performance is far from \\( \\overline{M}_{u} \\), questioning the unlearning mechanism used." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "We provide quantitative results for the revival of unlearned concepts while fine-tuning. CLIP [29] score calculated from the embeddings quantifies the semantic similarity between a text prompt and an image. Figure 2 shows the effect of fine-tuning on unlearned model with an unrelated concept. We prepared prompts containing golf ball as a validation dataset to evaluate the ability of the model to generate images of golf ball. At every epoch of fine-tuning, the model is sampled to generate images for validation prompts and the mean CLIP score is calculated for the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13395" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "embeddings of generated images and the prompts. As expected, the score is initially low, proving that the model has unlearned golf ball. An increase in the alignment of generated images towards golf ball with every epoch is observed, depicting the revival of the unlearned concept." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.182, + 0.48, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.428 + ], + "angle": 0, + "content": "Figure 2. Graph showing the variation of mean CLIP score with epochs. Fine-tuning is done on an unrelated concept \"piano keyboard\", and an increase is observed indicating the revival of unlearned concept \"Golf Ball\". ESD[39] was used for unlearning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.454, + 0.446, + 0.47 + ], + "angle": 0, + "content": "3.2. Dependence of revival on fine-tuning data" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.657 + ], + "angle": 0, + "content": "The previous subsection establishes that fine-tuning with certain unrelated concepts (hereafter referred to as fine-tuning concepts) can revive the unlearned concept. However, we observed that some MU algorithms demonstrated trivial revival compared to others when fine-tuned on the same unrelated concept (Figure 3). It has been observed that the extent of revival depends on the fine-tuning dataset, especially the fine-tuning concept. This makes it a challenge to assess the robustness of an unlearned model. While no revival is observed on a particular fine-tuning dataset, concept(s) that were not considered might exist and cause significant revival." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Analysis of the effect of different fine-tuning concepts is required to draw insights into the revival of unlearned concepts. The extent of fine-tuning (i.e., the number of epochs) required to observe the revival of unlearned concepts can serve as a good indicator of robustness. When the CLIP score crosses a predefined threshold, one may declare the revival of the unlearned concept. Figure 3 shows the results of such an experiment. Unlearned methods SalUn [10] and UCE [12] are considered to unlearn \"golf ball\" and fine-tuning was performed on concepts indicated on the vertical axis. The horizontal axis represents the number of epochs needed to achieve a CLIP score threshold of 0.31 (explained in Section 4). We considered fine-tuning till a maximum of 91 epochs, indicating the existence of fine-tuning concepts that don't cause revival even after prolonged fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "From Figure 3, it can be observed that a wide range of" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.244, + 0.905, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.409, + 0.908, + 0.478 + ], + "angle": 0, + "content": "Figure 3. Illustration of the number of epochs of fine-tuning required to observe revival of unlearned concept \"golf ball\". Finetuning concepts are shown on the vertical axis. The same finetuning concept can have different effects on revival for different unlearning methods (UCE and SalUn shown)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.869 + ], + "angle": 0, + "content": "fine-tuning concepts may be required to study the revival of unlearned concepts. A fine-tuning concept that caused revival for one work may not be prominent in revival for others. Further, we also observed that sequentially fine-tuning the unlearned model shows greater revival in a shorter time (number of fine-tuning epochs). Comparing Figure 3 (bottom panel) with Figure 4 makes this evident. Figure 4 shows the mean CLIP score when fine-tuning is performed in a sequential manner on concepts mentioned on the horizontal axis. A model that unlearned \"golf ball\" with SalUn [10] method was fine-tuned with each concept for 8 epochs sequentially (after fine-tuning on one concept, we continued fine-tuning on the next). While in Figure 3 (bottom panel), trivial revival was observed when fine-tuned individually on \"Volcano, Fish, Swimming, Hot Dog, Cricket Bat\" (indicated by large number of fine-tuning epochs), a CLIP score of 0.31 was reached in Figure 4 when fine-tuned sequentially. Specifically, for the fine-tuning concept \"Cricket Bat\", no revival was observed even after fine-tuning for a large number of epochs (91 in Figure 3 bottom panel). However, a significant revival (0.31 CLIP score) was observed at the end of fine-tuning \"Cricket Bat\" in a sequential manner for relatively less fine-tuning (\\(5 \\times 8 = 40\\) epochs in Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Hence, fine-tuning with various prompts may be required to study the revival of unlearning concepts. Further, the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13396" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.475, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.234, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Graph showing Mean CLIP score after fine-tuning on the concepts given on horizontal axis sequentially. The unlearned model used was produced by using SalUn[10] to remove target concept \"Golf Ball\". The highlighted point is when fine-tuning is done sequentially until Cricket Bat, and significant revival is shown as indicated by the CLIP score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.483, + 0.391 + ], + "angle": 0, + "content": "sequential fine-tuning of concepts seems to accelerate the revival. An experimental \"Concept revival framework\" to evaluate MU methods on this basis is detailed in section 4." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.404, + 0.46, + 0.423 + ], + "angle": 0, + "content": "4. Experiments - Studying the Vulnerability" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.506 + ], + "angle": 0, + "content": "We investigate the critical vulnerability aforementioned in the current Machine Unlearning (MU) methods for text-to-image diffusion models. Specifically, we demonstrate through extensive experiments that fine-tuning an unlearned model can lead to reviving previously unlearned concepts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.515, + 0.282, + 0.533 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.599 + ], + "angle": 0, + "content": "We conducted experiments on state-of-the-art MU methods applied to the Stable Diffusion v1.4 model to systematically study the concept revival. We describe our experimental setup:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.599, + 0.483, + 0.675 + ], + "angle": 0, + "content": "- Unlearning: We conducted experiments in four different unlearning settings, covering a total of ten concepts. In each case, unlearning is crucial for privacy, security, and ethical considerations, as it ensures certain content can no longer be generated by the model." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.675, + 0.483, + 0.719 + ], + "angle": 0, + "content": "- Object Unlearning focuses on unlearning specific objects from the model. Experiments are performed on \"Golf Ball, Fukushima and Dog\"." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.72, + 0.483, + 0.779 + ], + "angle": 0, + "content": "- Style Unlearning aims to remove the ability to replicate certain artistic styles using the model. \"Van Gogh, Picasso\" and \"Cartoon\" styles are used in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.781, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- Celebrity Unlearning aims to erase the model's ability to generate content of famous celebrities. We decided to use \"Angelina Jolie, Brad Pitt\" and \"Lionel Messi\"." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.826, + 0.483, + 0.869 + ], + "angle": 0, + "content": "- NSFW Content Unlearning focuses on preventing the model from generating inappropriate or explicit content." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Evaluation: A set of validation prompts is composed that contains the unlearning concept. The model is then sam" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.599, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "pled to generate images and evaluated using two main measures." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.122, + 0.905, + 0.167 + ], + "angle": 0, + "content": "- CLIP Score (CS): Measures the cosine similarity between the embeddings of generated images and their corresponding text prompts using the CLIP model [29]." + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.168, + 0.907, + 0.334 + ], + "angle": 0, + "content": "- Classifier Accuracy (Acc): Utilizes a fine-tuned Vision Transformer (ViT) [8] as a binary classifier to detect the presence of the unlearned concept in generated images. For Object and Celebrity Unlearning, we fine-tuned a VIT model on images of unlearning concepts and other unrelated images generated using the Stable Diffusion-3-medium [9]. In the case of Style Unlearning, we use the classifier from UnlearnCanvas [43] and fine-tune it using images generated by Stable Diffusion-3-medium [9]. Finally, for NSFW unlearning, we employ the NudeNet classifier [6]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.908, + 0.756 + ], + "angle": 0, + "content": "- Revival Point: Tracking the instance when a significant revival of the unlearned concept is observed can serve as a good measure for assessing the robustness of the unlearning method. We record the number of concepts on which fine-tuning was performed sequentially until revival was observed as \"Revival Point\". Further, to show no bias towards an unlearning method, we fix the order of fine-tuning concepts in the increasing order of semantic similarity from the target concept. The model is fine-tuned on each concept for 8 epochs during this process. A concept is considered to have revived when the model's CLIP score exceeds a specific threshold, and the classifier accuracy surpasses 0.3 (a conservative value that indicates revival). If the fine-tuning fails to revive the target concept in this setup, the MU method is comparatively robust (indicated by R). To calculate the clip threshold required for each target concept, we used images generated from base Stable-Diffusion v1.4 on prompts from the validation dataset (detailed later). Then, we calculated the CLIP score between the image and prompt embeddings. We set the CLIP threshold for to a value 0.02 less than the SD's CLIP score. CLIP threshold for each unlearning concept will be different, depending on the original model's CLIP score (thresholds for each concept are in Supplementary section B). In all the cases where the threshold was reached, a significant revival of the unlearned concept was observed." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.756, + 0.909, + 0.892 + ], + "angle": 0, + "content": "- Fine-tuning Dataset: To fine-tune the unlearned model, we create a customized dataset by first identifying concepts semantically related to the target concept. We achieve this by designing detailed prompts that specify the type of relationships at each hierarchical level and input these prompts into the OpenAI GPT 4o model. This process generates a comprehensive list of related concepts, which we then use to generate diverse prompts for fine-tuning the model. We compute the cosine" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.348, + 0.909, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13397" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.091, + 0.662, + 0.107 + ], + "angle": 0, + "content": "Clip Scores Before and After Fine-tuning" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.108, + 0.904, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.296, + 0.689, + 0.311 + ], + "angle": 0, + "content": "Classifier Accuracy Before and After Fine-tuning" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.312, + 0.904, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.525, + 0.908, + 0.582 + ], + "angle": 0, + "content": "Figure 5. Graphs illustrate the revival of Unlearned concepts on fine-tuning across different concepts and methods. CLIP score and Classifier accuracy are used to quantify the revival. Each colour corresponds to a different method, with lighter bar segments representing the increase in metrics before and after Fine-tuning, respectively. The consistent upward shifts in bar heights indicate that the unlearned concepts have been successfully revived." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.608, + 0.486, + 0.775 + ], + "angle": 0, + "content": "similarity between these concepts and the target concept using OpenAI embeddings [26] and arrange them in increasing order of similarity. For each selected concept, we generate 50 prompts using OpenAI GPT [25] and generate corresponding images using Stable Diffusion 3 Medium. This results in a dataset that enables curriculum fine-tuning, in which more general concepts appear before the more related concepts during the fine-tuning. This approach allows us to assess how fine-tuning concepts with different semantic similarities affects the revival of the unlearned concept." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.789, + 0.486, + 0.898 + ], + "angle": 0, + "content": "- Validation Dataset: We created a custom dataset of 20 prompts related to the unlearned concept using the OpenAI API. These prompts were carefully crafted to cover various descriptions associated with the target concept. We generated 5 images for each prompt using the fine-tuned models, resulting in a total of 100 images over which the CLIP score and classification accuracy mea" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.609, + 0.663, + 0.624 + ], + "angle": 0, + "content": "sures are computed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.669 + ], + "angle": 0, + "content": "Additional details of the experimental setup are provided in the Supplementary Materials (see Supplementary Sections C and D)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.88, + 0.696 + ], + "angle": 0, + "content": "4.2. Revival in the state-of-the-art MU methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Following the explanations in subsection 3.2, we sequentially fine-tune the unlearned model for 10 concepts. We report the two measures, CLIP score and Classifier accuracy, before and after the fine-tuning. An increase in these indicates the revival of unlearned concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "From Figure 5, it is evident that revival is observed across different concepts and unlearning methods. We have observed that MACE[20] and UCE[12] were relatively robust across many concepts, with less revival observed after fin-tuning. Upon further analysis of different types of unlearning, UCE[12] was found to be robust in object unlearning. We found almost all methods vulnerable to fine-tuning in case of style unlearning. We can infer that artistic styles like" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13398" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Van Gogh and Picasso are inherently complex and challenging to unlearn precisely. In the case of celebrity unlearning, we observed that there is an increase in metrics across all works with UCE's[12] change being minimal. Nudity unlearning was easily broken with relatively faster revival in the case of ESD[11], SPM[21] and SA[15] compared to MACE[20] which was robust. Even with adversarial robust MU method[17] which is robust against adversarial prompts we can observe concept revival (low for object and style but high for NSFW content). More experimental details are provided in the Supplementary Material (see Supplementary Section B)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.286, + 0.476, + 0.302 + ], + "angle": 0, + "content": "4.3. Framework for Studying the Concept Revival" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.482, + 0.386 + ], + "angle": 0, + "content": "Based on previous sections, it is apparent that model robustness should be assessed with its ability to safeguard against the revival of unlearning concepts. This section details the proposed framework for investigating the concept revival in the unlearned diffusion model due to fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.39, + 0.483, + 0.435 + ], + "angle": 0, + "content": "1. Step 1: Fine-tune the unlearned model sequentially on a semantically distant concept for a fixed number of epochs." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.436, + 0.483, + 0.479 + ], + "angle": 0, + "content": "2. Step 2: Evaluate whether the CLIP score and classifier accuracy for the unlearned concept crosses a predefined threshold called the Revival Point." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.39, + 0.483, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.481, + 0.483, + 0.51 + ], + "angle": 0, + "content": "- If Not: Proceed to fine-tune on a semantically closer concept." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.511, + 0.482, + 0.54 + ], + "angle": 0, + "content": "- If Yes: Record the Revival Point at which the unlearned concept resurfaces." + }, + { + "type": "list", + "bbox": [ + 0.111, + 0.481, + 0.483, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.542, + 0.483, + 0.571 + ], + "angle": 0, + "content": "3. Step 3: Repeat the process by gradually decreasing the semantic distance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.652 + ], + "angle": 0, + "content": "This framework provides a quantitative measure of the model's robustness based on the point at which the unlearned concept reappears and requires less number of epochs compared to fine-tuning separately using each concept." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.351, + 0.681 + ], + "angle": 0, + "content": "4.4. Impact of Surrogate Concept" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Our empirical studies indicate that mapping the unlearned concept to a semantically distant or general concept enhances the model's robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents the time taken for the unlearned concept to resurface when mapped to two concepts at varying semantic similarity. The experimental results demonstrate that mapping an unlearned concept (Golf ball) to a semantically distant concept (Outdoor Recreational activities) enhances the robustness of unlearning. Here, we experimented with the two methods UCE [12] and MACE [20]; clip score (CS) increased only slightly when mapping to the unrelated concept, with minimal or moderate gains in accuracy after unlearning. This indicates that mapping to a semantically distant concept disrupts latent associations more" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "effectively, reducing the likelihood of the unlearned concept re-emerging during fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.127, + 0.907, + 0.293 + ], + "angle": 0, + "content": "Conversely, mapping to a semantically closer concept (Tennis ball) retained certain feature similarities; for instance, subtle characteristics like the colour of the Golf ball shifted towards a light green due to the influence of the Tennis ball, while other properties associated with Golf balls remained intact. Even after fine-tuning, this greenish tint persisted, highlighting that the concept is still fundamentally a Golf ball but with minor alterations not reflected in the quantitative metrics. This observation underscores the efficacy of semantically distant mappings for achieving robust concept unlearning in generative models." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.314, + 0.888, + 0.439 + ], + "angle": 0, + "content": "
MethodBefore FTAfter FTRevival Point
CSAccCSAcc
UCE-T0.330.50.330.71
UCE-O0.260.010.280.11R
MACE-T0.320.00.340.543
MACE-O0.270.190.300.15R
SPM-T0.320.270.320.473
SPM-O0.320.290.330.534
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.45, + 0.905, + 0.506 + ], + "angle": 0, + "content": "Table 1. Clip Score (CS) and Accuracy (Acc) Before and After Fine-tuning When Mapping Golf Ball to Tennis Ball (T) (Semantically Close) and Outdoor Recreational Activities (O) (Semantically Distant)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.551, + 0.905, + 0.568 + ], + "angle": 0, + "content": "4.5. Impact of Modules affected during Unlearning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.906, + 0.655 + ], + "angle": 0, + "content": "We chose ESD as the basis for this experiment because it utilizes a setup that allows unlearning by modifying only specific modules (like only cross attention, only non-cross attention and full parameters), enabling targeted analysis of robustness in concept unlearning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In our experiments, we used the ESD variations of - ESD-x (cross-attention layers), ESD-u (unconditional layers), and ESD-full (all parameters). We observed that the specific parameters targeted during unlearning significantly impacted the robustness of the concept erasure. Results (Table 2) indicate that cross-attention-based unlearning (ESD-x) is less robust, as the unlearned concept revived more quickly compared to ESD-u and ESD-full. This observation suggests that editing cross-attention weights alone may be insufficient for durable unlearning, as these layers maintain strong associations between prompts and visual features. This vulnerability is also evident in the UCE and CA model, which exclusively modifies cross-attention weights, making it relatively fragile to fine-tuning. Therefore, targeting a broader parameter set, as seen in ESD-u and ESD-full, offers a more resilient approach to concept unlearning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13399" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.462, + 0.169 + ], + "angle": 0, + "content": "
MethodBefore FTAfter FTRevival Point
CSAccCSAcc
ESD-x0.260.00.320.544
ESD-u0.260.00.330.365
ESD-full0.250.00.320.397
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.179, + 0.483, + 0.22 + ], + "angle": 0, + "content": "Table 2. Comparison of Clip Score (CS) and Accuracy (Acc) Before and After Unlearning for Different ESD Module Configurations." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.274, + 0.465, + 0.293 + ], + "angle": 0, + "content": "5. Discussion - Plausible Mitigation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.313, + 0.486, + 0.587 + ], + "angle": 0, + "content": "For unlearning methods that rely on minimizing a dedicated objective (referred to as the unlearning loss), its value acts as a metric to gauge the success of the erasure. To investigate the impact of fine-tuning on this loss, we evaluated three unlearning techniques—UCE [12], ESD [11], and CA [19]—measuring the unlearning loss both immediately after unlearning and after subsequent fine-tuning (as described in section 4). The results, summarized in Table 3, reveal a consistent increase in unlearning loss post fine-tuning, indicating that fine-tuning destabilizes the unlearning process. Notice that in the case of the CA method with \"Angelina Jolie\" as the target concept and \"Woman\" as the surrogate concept, the unlearning loss did not increase after fine-tuning. This suggests no significant change in the extent of unlearning, which is consistent with previous experiments (Fig 5), where fine-tuning resulted in negligible changes to the model's unlearning accuracy for this particular target/surrogate concept pairing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.486, + 0.887 + ], + "angle": 0, + "content": "A plausible approach to mitigate the revival could involve meta-learning and gradient alignment techniques. While standard fine-tuning minimizes the diffusion model's loss, its gradients may conflict with the unlearning objective. To resolve this, we propose modifying the fine-tuning process to (1) compute gradients for both the diffusion loss and unlearning objective, (2) project the diffusion gradients orthogonal to the unlearning gradient direction, and (3) update parameters using only the orthogonal component. This ensures that the unlearning loss remains unchanged throughout fine-tuning, preserving the unlearning effect. A similar approach has been employed by Ko et al. [18] and Wu et al. [40] to achieve unlearning while maintaining the generational quality of neutral concepts. Their approach optimizes two objectives—one for preserving the generation quality of neutral concepts and another for unlearning the target concept by ensuring that each update step doesn't increase either objective function." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.089, + 0.905, + 0.293 + ], + "angle": 0, + "content": "
MethodUnlearn LossUnlearn + FTTarget / Surrogate
UCE5.54 × 1016.74 × 102Dog / Cat
1.08 × 1061.65 × 106Van Gogh / Art
6.66 × 1011.03 × 103Pikachu / Cat
9.02 × 1051.23 × 106Cartoon / Art
1.23 × 1029.41 × 102Brad Pitt / Man
ESD5.00 × 10-41.10 × 10-3Golf Ball / NA
3.00 × 10-45.00 × 10-4Van Gogh / NA
4.00 × 10-49.00 × 10-4Nudity / NA
3.00 × 10-41.60 × 10-3Brad Pitt / NA
CA4.79 × 10-47.53 × 10-4Golf Ball / Ball
1.05 × 10-41.49 × 10-4Van Gogh / Art
≈ 0≈ 0Ang. Jolie / Woman
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.304, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Table 3. Comparison of unlearning loss values between: (1) models immediately after unlearning, and (2) the same models after fine-tuning (using weights obtained from Section 4 experiments). Loss values were computed with each method's respective objective function, with corresponding target-surrogate concept pairings shown. The parameter group updated in ESD and CA for these experiments is cross-attention. The last column shows the target concept (Unlearned concept) along with the corresponding mapping concept i.e Surrogate concept. NA means No specific Surrogate Concept." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.642, + 0.484 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.705 + ], + "angle": 0, + "content": "We present a critical vulnerability in the current Machine Unlearning techniques for text-to-image diffusion models. Our study emphasizes the need for (i) more robust techniques that can erase the semantic space of the target concept in the latent space of the model, in which case finetuning can never revive the target concept, or/and (ii) the method should be able to achieve mapping of target concept to a surrogate concept which is robust to finetuning. By revealing the weaknesses of existing methods against finetuning and providing a framework to measure concept revival, we aim to guide future research toward developing more secure, reliable, and robust unlearning mechanisms. Below, We list some interesting and important directions to further our understanding of the unlearning process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.905, + 0.78 + ], + "angle": 0, + "content": "- Rigorous understanding of why the finetuning process revives the unlearned concepts across several unlearning approaches. Though challenging, it is of utmost importance to explain this vulnerability of the existing MU methods and to achieve robust unlearning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.907, + 0.9 + ], + "angle": 0, + "content": "- Relation between concept revival and finetuning concept. We have observed that certain finetuning concepts can cause greater revival while others less. Also, this dependence may vary across unlearned concepts and unlearning methods. Researching the primary cause of the variance in dependency between the finetuning concept and the revival of the unlearning concept can aid in assessing the robustness of the models better." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.705, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13400" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Stability AI. Stability ai. https://stability.ai/news/celebrating-one-year-of-stable-diffusion, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.483, + 0.212 + ], + "angle": 0, + "content": "[2] Justin Betker, Gabriel Goh, Li Jing, Tim Brooks, Jason Wang, Lilian Li, Aditya Ramesh, et al. Improving image generation with better captions. Computer Science, 2(3):8, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy (SP), pages 463-480. IEEE, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[4] Mingjie Chen, Wei Gao, Guangyue Liu, Kai Peng, and Cunhua Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7766-7775, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[5] Lydia de la Torre. A guide to the california consumer privacy act of 2018. Available at SSRN 3275571, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[6] NudeNet Developers. Nudenet: A nudity detection and censoring library. https://github.com/notAI-tech/NudeNet, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.439 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.441, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.636, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[10] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[11] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[12] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[14] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9301-9309. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[15] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[16] Chris Jay Hoofnagle, Bart van der Sloot, and Frederik Zuiderveen Borgesius. The european union general data protection regulation: what it is and what it means. Information & Communications Technology Law, 28(1):65-98, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[17] Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung-Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. Receiver: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. In European Conference on Computer Vision, pages 360–376. Springer, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[18] Myeongseob Ko, Henry Li, Zhun Wang, Jonathan Patsenker, Jiachen T. Wang, Qinbin Li, Ming Jin, Dawn Song, and Ruoxi Jia. Boosting alignment for post-unlearning text-to-image generative models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.419, + 0.908, + 0.488 + ], + "angle": 0, + "content": "[19] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.906, + 0.558 + ], + "angle": 0, + "content": "[20] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.56, + 0.906, + 0.643 + ], + "angle": 0, + "content": "[21] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.645, + 0.902, + 0.671 + ], + "angle": 0, + "content": "[22] Midjourney. Midjourney. https://www.midjourney.com/, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.906, + 0.756 + ], + "angle": 0, + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pooja Shyam, Pavel Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.758, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.815, + 0.769, + 0.829 + ], + "angle": 0, + "content": "[25] OpenAI. Chatgpt model api, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[26] OpenAI. Openai embeddings, 2023. Accessed: 2024-11-13. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13401" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.15 + ], + "angle": 0, + "content": "[28] Formerly Data Protection. General data protection regulation (gdpr). Intersoft Consulting, accessed in October, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.361 + ], + "angle": 0, + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alexander Quinn Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Sahand Mahdavi, Raphael Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[34] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[35] Christoph Schuhmann, Robin Beaumont, Romain Vencu, Craig Gordon, Ross Wightman, Mehdi Cherti, Teddy Coombes, Abhay Katta, and Vladimir Kirillov. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278–25294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[36] Anish Sekhari, Jayadev Acharya, Gauri Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. In Advances in Neural Information Processing Systems, pages 18075-18086, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[37] GenAI statistics. Genai statistics. https://journaleverypixel.com/ai-image-statistics, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[38] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[39] Jing Wu, Trung Le, Munawar Hayat, and Mehrtash Harandi. Erasediff: Erasing data influence in diffusion models, 2024. 1, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[40] Yongliang Wu, Shiji Zhou, Mingzhuo Yang, Lianzhe Wang, Heng Chang, Wenbo Zhu, Xinting Hu, Xiao Zhou, and Xu Yang. Unlearning concepts in diffusion model via concept domain correction and concept preserving gradient. In Proceedings of the 39th AAAI Conference on Artificial Intelligence (AAAI), 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[41] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1755-1764, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[42] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now. In Computer Vision - ECCV 2024: 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LVII, page 385-403, Berlin, Heidelberg, 2024. Springer-Verlag. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.36, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[43] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. In Thirty-eighth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13402" + } + ] +] \ No newline at end of file diff --git a/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_origin.pdf b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b6232360bc5652383c8edbc22a76f2ebbfe53825 --- /dev/null +++ b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/ca9a5cea-4cef-406e-b3c3-a1c279377219_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d73f527788f76b1b679e22a67b462d465719531f5fa97422c2379214bb105f0c +size 568379 diff --git a/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/full.md b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5351ef1ab92cc80cd570293778caf0ae808be7b9 --- /dev/null +++ b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/full.md @@ -0,0 +1,248 @@ +# The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models + +Naveen George $^{1*}$ , Karthik Nandan Dasaraju $^{2*}$ , Rutheesh Reddy Chittepu $^{1}$ , Konda Reddy Mopuri $^{1}$ $^{1}$ Data-driven Intelligence and Learning Lab, Indian Institute of Technology, Hyderabad + $^{2}$ Indian Institute of Technology, Ropar + +ai23mtech12001@iith.ac.in, 2021csb1081@iitrpr.ac.in, cs21btech11014@iith.ac.in, krmopuri@ai.iith.ac.in + +# Abstract + +Text-to-image models such as Stable Diffusion, DALL-E, and Midjourney have gained immense popularity lately. However, they are trained on vast amounts of data that may include private, explicit, or copyrighted material used without permission, raising serious legal and ethical concerns. In light of the recent regulations aimed at protecting individual data privacy, there has been a surge in Machine Unlearning methods designed to remove specific concepts from these models. However, we identify a critical flaw in these unlearning techniques: unlearned concepts will revive when the models are fine-tuned, even with general or unrelated prompts. In this paper, for the first time, through an extensive study, we demonstrate the unstable nature of existing unlearning methods in text-to-image diffusion models. We introduce a framework that includes a couple of measures for analyzing the stability of existing unlearning methods. Further, the paper offers preliminary insights into the plausible explanation for the instability of the mapping-based unlearning methods that can guide future research toward more robust unlearning techniques. Codes1 for implementing the proposed framework are provided. + +# 1. Introduction + +The past decade has witnessed a rapid qualitative growth in text-to-image models [7, 23, 24, 32, 33], leading to its widespread attention and adoption. Models such as DALL-E [2, 30, 31], MidJourney [22], and Stable Diffusion [27, 32] enable high-quality image generation from simple text prompts, opening applications across art, design, advertising, gaming, and even medical imaging. With millions worldwide adopting these tools for academic and commercial purposes, their transformative potential is evident, as illustrated by MidJourney's 16 million active + +users [22] and the generation of over 2 million images daily by DALL-E and Stable Diffusion [1, 37]. However, this widespread use raises critical concerns about data privacy, security, fairness, and intellectual property, as these models rely on vast datasets that often include sensitive, personal, or copyrighted content [35]. + +In response to these privacy concerns, regulations such as the General Data Protection Regulation (GDPR) [28] and the California Consumer Privacy Act (CCPA) [5] provide users the Right to be Forgotten [16], promoting the development of Machine Unlearning (MU) techniques [3, 13]. MU aims to remove specific data or concepts from a model without the computational expense of retraining from scratch. While initially explored in classification models [4, 13, 14, 36], MU in generative models, particularly diffusion-based ones, poses unique challenges due to the models' complex, distribution-based learning. Ensuring these models forget sensitive or harmful information while retaining general functionality remains an open problem. + +Machine Unlearning in text-to-image models focuses on erasing specific concepts from the model's generative capacity. The general goal of MU in this context is to ensure that when a model is conditioned on a prompt related to the target or "unlearned" concept, it no longer generates images reflecting that concept. These target concepts can be Not Safe For Work (NSFW) content [11, 15, 17, 21, 34], objects [10-12, 15, 19-21, 39, 41] or artistic styles [11, 12, 15, 19-21, 39, 41]. In other words, the unlearned concept is the one we want the model to "forget". To achieve this, unlearning algorithms typically remap the target concept to a surrogate concept. This means that when the model is conditioned on the unlearned concept, it produces images associated with the surrogate concept instead. The surrogate concept can be a general or neutral concept [12, 15, 19], a random concept [10], or even an empty string [11, 39]. + +Despite these efforts in Machine Unlearning, we have identified a critical vulnerability in the current state-of-the + +art methods for text-to-image models. Specifically, we observe that when these unlearned models undergo finetuning, even with general or unrelated concepts, the previously unlearned concepts can resurface in the generated images. This unintended revival undermines the unlearning process's effectiveness and poses significant risks to data privacy and legal compliance. The ability of an unlearned model to regenerate sensitive or copyrighted content after benign fine-tuning highlights a fundamental weakness in the existing MU approaches for generative models. Such vulnerabilities could lead to unauthorized dissemination of protected content, violating individual privacy rights and intellectual property laws. Therefore, it is imperative to (i) study the robustness of the existing MU methods, (ii) understand the underlying causes of this behaviour and develop more robust unlearning strategies. + +Contributions of the paper can be summarized as follows. + +- Systematic evaluation of MU methods: We conduct a comprehensive study of the existing MU methods in text-to-image diffusion models, demonstrating that unlearned concepts are revived across all examined approaches. Our analysis includes all the leading methods — SalUn [10], ESD [11], EDiff [39], CA [19], MACE [20], SPM [21], SA [15], Receler [17] and UCE [12] —capturing the full spectrum of current state-of-the-art unlearning techniques. To the best of our knowledge, this is the first work to identify and discuss this issue in depth, examining the robustness of unlearning. We believe this work brings an unknown but critical aspect of the MU methods into light, paving the way to a more resilient framework for unlearning. + +- Concept Revival Framework: We introduce a framework demonstrating that fine-tuning revives the unlearned concepts. For each target concept, the proposed framework presents a customized curriculum of concepts in the increasing order of relevance to fine-tune the unlearned model. During the fine-tuning, the framework enables the measurement of the revival of the target concept via a couple of simple measures: classification accuracy and CLIP score. + +- Analysis and insights about the unlearning process: We study the fragility of the existing MU methods with respect to two important axes: + +- Semantic relevance of the surrogate concept to the target concept: we analyze how mapping to surrogate concepts of varying semantic relevance (to the target concept) affects the stability of the resulting unlearning. + +- Modules affected during the unlearning: we examine how the nature of model parameters modified during unlearning, (i.e., cross-attention module vs. the rest), affects the unlearning. + +Further, we hypothesize a plausible explanation for the in + +stability of the mapping-based MU methods - fine-tuning the unlearned model disrupts the mapping between the target and surrogate concepts. We provide empirical evidence for this hypothesis over some of the popular MU methods [11, 12, 19], in terms of the target concept's unlearning objective function value during the fine-tuning. + +# 2. Related Works + +Unlearning in image generation. Machine Unlearning (MU) approaches aim to remove the influence of certain data points or concepts from a model without the need for full retraining, which is computationally intensive for large-scale models like diffusion models. The rise of MU methods has been driven by their potential to address various issues in generative models. For example, text-to-image (T2I) diffusion models have a tendency to generate inappropriate or sensitive content, such as depictions of violence, deepfakes, or copyrighted artistic styles. MU methods are recognized for their role in promoting data privacy and complying with regulations like the "Right to be Forgotten" by enabling models to unlearn specific concepts deemed undesirable or unlawful. + +Several methods have been proposed to achieve concept unlearning in diffusion models. Unified Concept Editing (UCE) [12] modifies cross-attention in the U-Net component of diffusion models using a closed-form solution to map target concepts to surrogate ones, effectively erasing them without additional training. Mass Concept Erasure (MACE) [20] extends this by refining cross-attention layers and utilizing Low-Rank Adaptation (LoRA) modules to efficiently remove multiple unwanted concepts while preserving the model's ability to generate unrelated content. + +Other approaches like Selective Amnesia (SA) [15] leverage Elastic Weight Consolidation and Generative Replay to forget target concepts while maintaining overall model performance. Erased Stable Diffusion (ESD) [11] aligns the probability distribution of a concept to be erased with that of a null string, although it may lead to a collapse in generative quality. Concept Ablation (CA) [19] fine-tunes cross-attention weights to map target concepts to more general anchor concepts, effectively preventing the generation of specific styles or objects. + +Methods such as Safe Latent Diffusion (SLD) [34] use reverse guidance during generation to steer away from inappropriate content but require careful balancing of prompt conditioning and hyperparameters. The Semi-Permeable Membrane (SPM) [21] adapter offers a lightweight solution for multi-concept erasure by introducing a regulated erasing signal within intermediate layers, although it may struggle with partial unlearning (traces of the target concept remain). + +While these methods can effectively unlearn the target concept, they often adversely affect the generative capability for related concepts [43] and may degrade the over + +![](images/4cd222d6fe147dcc0ca652140eeb5e7ae897b03affb79eaab2b0312423bd9077.jpg) +Figure 1. Revival of unlearned concepts is observed across numerous concepts and unlearning methods. The top row presents images generated by the original SDv1.4 model. The middle row shows images from models with specific concepts unlearned. The unlearning concept and the method used are at the bottom of each image. The last row shows images generated when the unlearned model is finetuned on several different concepts. At the bottom of each column is the prompt used for sampling. + +all image quality. Additionally, the unlearned concepts can sometimes reappear if an adversarial prompt is carefully engineered, indicating incomplete unlearning and vulnerabilities to prompt-based attacks [38, 42]. There are also adversarial robust unlearning methods [17] that are trained to withstand these adversarial prompts without compromising the generation of non-target concepts (i.e., locality). + +# 3. Proposed Framework + +# 3.1. Fine-tuning revives the unlearned concepts + +Numerous methods have been proposed to unlearn a wide range of target concepts from diffusion models. These targets range from objects and artistic styles to global concepts such as nudity. We present the effect of fine-tuning these unlearned models on different concepts in a systematic way. Figure 1 presents the observed effect concisely, where the top row shows images generated by the original Stable Diffusion v1.4. The second row shows images generated by the same model after unlearning different concepts using numerous unlearning methods[10-12, 15, 17, 19-21, 39]. These works demonstrated a good unlearning performance, which is evident from the absence of the unlearning concept in the images. The performance of these works was also quantified by UnlearnCanvas [43] and shown to offer good unlearning accuracy. After fine-tuning various concepts sequentially (Section 4), images generated by the model are + +shown in the bottom row of the exact figure. Revival of the unlearned concept is observed in all cases. + +This revival of unlearned concepts with fine-tuning on arbitrary concept(s) is a critical vulnerability. A model published to have unlearned a sensitive concept can be made to relearn by fine-tuning, posing a threat to the publisher. This is more pressing if the concept being unlearned is nudity, NSFW, or copyrighted content. A model $M_{u}$ trained by some unlearning mechanism is compared with the ideal model $\overline{M}_{u}$ , which has never seen the unlearning concept before, i.e. the model trained on a dataset with unlearning concept removed. $\overline{M}_{u}$ that has never known an unlearned concept (golf ball for example), won't infer its existence when fine-tuned on images of different concept (piano keyboard). The revival of the unlearned concept in $M_{u}$ shows that its performance is far from $\overline{M}_{u}$ , questioning the unlearning mechanism used. + +We provide quantitative results for the revival of unlearned concepts while fine-tuning. CLIP [29] score calculated from the embeddings quantifies the semantic similarity between a text prompt and an image. Figure 2 shows the effect of fine-tuning on unlearned model with an unrelated concept. We prepared prompts containing golf ball as a validation dataset to evaluate the ability of the model to generate images of golf ball. At every epoch of fine-tuning, the model is sampled to generate images for validation prompts and the mean CLIP score is calculated for the + +embeddings of generated images and the prompts. As expected, the score is initially low, proving that the model has unlearned golf ball. An increase in the alignment of generated images towards golf ball with every epoch is observed, depicting the revival of the unlearned concept. + +![](images/5de54fa76ef4b87998264e27b65b6daf84582e6fd85b256b89dec195ca7e1c38.jpg) +Figure 2. Graph showing the variation of mean CLIP score with epochs. Fine-tuning is done on an unrelated concept "piano keyboard", and an increase is observed indicating the revival of unlearned concept "Golf Ball". ESD[39] was used for unlearning. + +# 3.2. Dependence of revival on fine-tuning data + +The previous subsection establishes that fine-tuning with certain unrelated concepts (hereafter referred to as fine-tuning concepts) can revive the unlearned concept. However, we observed that some MU algorithms demonstrated trivial revival compared to others when fine-tuned on the same unrelated concept (Figure 3). It has been observed that the extent of revival depends on the fine-tuning dataset, especially the fine-tuning concept. This makes it a challenge to assess the robustness of an unlearned model. While no revival is observed on a particular fine-tuning dataset, concept(s) that were not considered might exist and cause significant revival. + +Analysis of the effect of different fine-tuning concepts is required to draw insights into the revival of unlearned concepts. The extent of fine-tuning (i.e., the number of epochs) required to observe the revival of unlearned concepts can serve as a good indicator of robustness. When the CLIP score crosses a predefined threshold, one may declare the revival of the unlearned concept. Figure 3 shows the results of such an experiment. Unlearned methods SalUn [10] and UCE [12] are considered to unlearn "golf ball" and fine-tuning was performed on concepts indicated on the vertical axis. The horizontal axis represents the number of epochs needed to achieve a CLIP score threshold of 0.31 (explained in Section 4). We considered fine-tuning till a maximum of 91 epochs, indicating the existence of fine-tuning concepts that don't cause revival even after prolonged fine-tuning. + +From Figure 3, it can be observed that a wide range of + +![](images/1d3fbba208cbcbe981b54f29057bb6bf2344818488fc01c49d5b866839998803.jpg) + +![](images/f8e48bf49f0722b4dc5e6ebe8f27339befeef115de312d0eb19ae71447422626.jpg) +Figure 3. Illustration of the number of epochs of fine-tuning required to observe revival of unlearned concept "golf ball". Finetuning concepts are shown on the vertical axis. The same finetuning concept can have different effects on revival for different unlearning methods (UCE and SalUn shown) + +fine-tuning concepts may be required to study the revival of unlearned concepts. A fine-tuning concept that caused revival for one work may not be prominent in revival for others. Further, we also observed that sequentially fine-tuning the unlearned model shows greater revival in a shorter time (number of fine-tuning epochs). Comparing Figure 3 (bottom panel) with Figure 4 makes this evident. Figure 4 shows the mean CLIP score when fine-tuning is performed in a sequential manner on concepts mentioned on the horizontal axis. A model that unlearned "golf ball" with SalUn [10] method was fine-tuned with each concept for 8 epochs sequentially (after fine-tuning on one concept, we continued fine-tuning on the next). While in Figure 3 (bottom panel), trivial revival was observed when fine-tuned individually on "Volcano, Fish, Swimming, Hot Dog, Cricket Bat" (indicated by large number of fine-tuning epochs), a CLIP score of 0.31 was reached in Figure 4 when fine-tuned sequentially. Specifically, for the fine-tuning concept "Cricket Bat", no revival was observed even after fine-tuning for a large number of epochs (91 in Figure 3 bottom panel). However, a significant revival (0.31 CLIP score) was observed at the end of fine-tuning "Cricket Bat" in a sequential manner for relatively less fine-tuning ( $5 \times 8 = 40$ epochs in Figure 3). + +Hence, fine-tuning with various prompts may be required to study the revival of unlearning concepts. Further, the + +![](images/48741fadc55cba1a912060307483e4a024c730de134ec18cb5be66b8dab2b5e5.jpg) +Figure 4. Graph showing Mean CLIP score after fine-tuning on the concepts given on horizontal axis sequentially. The unlearned model used was produced by using SalUn[10] to remove target concept "Golf Ball". The highlighted point is when fine-tuning is done sequentially until Cricket Bat, and significant revival is shown as indicated by the CLIP score. + +sequential fine-tuning of concepts seems to accelerate the revival. An experimental "Concept revival framework" to evaluate MU methods on this basis is detailed in section 4. + +# 4. Experiments - Studying the Vulnerability + +We investigate the critical vulnerability aforementioned in the current Machine Unlearning (MU) methods for text-to-image diffusion models. Specifically, we demonstrate through extensive experiments that fine-tuning an unlearned model can lead to reviving previously unlearned concepts. + +# 4.1. Experimental Setup + +We conducted experiments on state-of-the-art MU methods applied to the Stable Diffusion v1.4 model to systematically study the concept revival. We describe our experimental setup: + +- Unlearning: We conducted experiments in four different unlearning settings, covering a total of ten concepts. In each case, unlearning is crucial for privacy, security, and ethical considerations, as it ensures certain content can no longer be generated by the model. +- Object Unlearning focuses on unlearning specific objects from the model. Experiments are performed on "Golf Ball, Fukushima and Dog". +- Style Unlearning aims to remove the ability to replicate certain artistic styles using the model. "Van Gogh, Picasso" and "Cartoon" styles are used in our experiments. +- Celebrity Unlearning aims to erase the model's ability to generate content of famous celebrities. We decided to use "Angelina Jolie, Brad Pitt" and "Lionel Messi". +- NSFW Content Unlearning focuses on preventing the model from generating inappropriate or explicit content. +- Evaluation: A set of validation prompts is composed that contains the unlearning concept. The model is then sam + +pled to generate images and evaluated using two main measures. + +- CLIP Score (CS): Measures the cosine similarity between the embeddings of generated images and their corresponding text prompts using the CLIP model [29]. + +- Classifier Accuracy (Acc): Utilizes a fine-tuned Vision Transformer (ViT) [8] as a binary classifier to detect the presence of the unlearned concept in generated images. For Object and Celebrity Unlearning, we fine-tuned a VIT model on images of unlearning concepts and other unrelated images generated using the Stable Diffusion-3-medium [9]. In the case of Style Unlearning, we use the classifier from UnlearnCanvas [43] and fine-tune it using images generated by Stable Diffusion-3-medium [9]. Finally, for NSFW unlearning, we employ the NudeNet classifier [6]. + +- Revival Point: Tracking the instance when a significant revival of the unlearned concept is observed can serve as a good measure for assessing the robustness of the unlearning method. We record the number of concepts on which fine-tuning was performed sequentially until revival was observed as "Revival Point". Further, to show no bias towards an unlearning method, we fix the order of fine-tuning concepts in the increasing order of semantic similarity from the target concept. The model is fine-tuned on each concept for 8 epochs during this process. A concept is considered to have revived when the model's CLIP score exceeds a specific threshold, and the classifier accuracy surpasses 0.3 (a conservative value that indicates revival). If the fine-tuning fails to revive the target concept in this setup, the MU method is comparatively robust (indicated by R). To calculate the clip threshold required for each target concept, we used images generated from base Stable-Diffusion v1.4 on prompts from the validation dataset (detailed later). Then, we calculated the CLIP score between the image and prompt embeddings. We set the CLIP threshold for to a value 0.02 less than the SD's CLIP score. CLIP threshold for each unlearning concept will be different, depending on the original model's CLIP score (thresholds for each concept are in Supplementary section B). In all the cases where the threshold was reached, a significant revival of the unlearned concept was observed. +- Fine-tuning Dataset: To fine-tune the unlearned model, we create a customized dataset by first identifying concepts semantically related to the target concept. We achieve this by designing detailed prompts that specify the type of relationships at each hierarchical level and input these prompts into the OpenAI GPT 4o model. This process generates a comprehensive list of related concepts, which we then use to generate diverse prompts for fine-tuning the model. We compute the cosine + +![](images/30b8f9f60a4d0d7ff40049f1a063d6a9e7b824e24c1528a815f0563f5134162f.jpg) +Clip Scores Before and After Fine-tuning + +![](images/e33a0b85bbe7c76714c3209f1cce31015e988742f162f9d0ac18e149d0f71939.jpg) +Classifier Accuracy Before and After Fine-tuning +Figure 5. Graphs illustrate the revival of Unlearned concepts on fine-tuning across different concepts and methods. CLIP score and Classifier accuracy are used to quantify the revival. Each colour corresponds to a different method, with lighter bar segments representing the increase in metrics before and after Fine-tuning, respectively. The consistent upward shifts in bar heights indicate that the unlearned concepts have been successfully revived. + +similarity between these concepts and the target concept using OpenAI embeddings [26] and arrange them in increasing order of similarity. For each selected concept, we generate 50 prompts using OpenAI GPT [25] and generate corresponding images using Stable Diffusion 3 Medium. This results in a dataset that enables curriculum fine-tuning, in which more general concepts appear before the more related concepts during the fine-tuning. This approach allows us to assess how fine-tuning concepts with different semantic similarities affects the revival of the unlearned concept. + +- Validation Dataset: We created a custom dataset of 20 prompts related to the unlearned concept using the OpenAI API. These prompts were carefully crafted to cover various descriptions associated with the target concept. We generated 5 images for each prompt using the fine-tuned models, resulting in a total of 100 images over which the CLIP score and classification accuracy mea + +sures are computed. + +Additional details of the experimental setup are provided in the Supplementary Materials (see Supplementary Sections C and D). + +# 4.2. Revival in the state-of-the-art MU methods + +Following the explanations in subsection 3.2, we sequentially fine-tune the unlearned model for 10 concepts. We report the two measures, CLIP score and Classifier accuracy, before and after the fine-tuning. An increase in these indicates the revival of unlearned concepts. + +From Figure 5, it is evident that revival is observed across different concepts and unlearning methods. We have observed that MACE[20] and UCE[12] were relatively robust across many concepts, with less revival observed after fin-tuning. Upon further analysis of different types of unlearning, UCE[12] was found to be robust in object unlearning. We found almost all methods vulnerable to fine-tuning in case of style unlearning. We can infer that artistic styles like + +Van Gogh and Picasso are inherently complex and challenging to unlearn precisely. In the case of celebrity unlearning, we observed that there is an increase in metrics across all works with UCE's[12] change being minimal. Nudity unlearning was easily broken with relatively faster revival in the case of ESD[11], SPM[21] and SA[15] compared to MACE[20] which was robust. Even with adversarial robust MU method[17] which is robust against adversarial prompts we can observe concept revival (low for object and style but high for NSFW content). More experimental details are provided in the Supplementary Material (see Supplementary Section B). + +# 4.3. Framework for Studying the Concept Revival + +Based on previous sections, it is apparent that model robustness should be assessed with its ability to safeguard against the revival of unlearning concepts. This section details the proposed framework for investigating the concept revival in the unlearned diffusion model due to fine-tuning. + +1. Step 1: Fine-tune the unlearned model sequentially on a semantically distant concept for a fixed number of epochs. +2. Step 2: Evaluate whether the CLIP score and classifier accuracy for the unlearned concept crosses a predefined threshold called the Revival Point. + +- If Not: Proceed to fine-tune on a semantically closer concept. +- If Yes: Record the Revival Point at which the unlearned concept resurfaces. + +3. Step 3: Repeat the process by gradually decreasing the semantic distance. + +This framework provides a quantitative measure of the model's robustness based on the point at which the unlearned concept reappears and requires less number of epochs compared to fine-tuning separately using each concept. + +# 4.4. Impact of Surrogate Concept + +Our empirical studies indicate that mapping the unlearned concept to a semantically distant or general concept enhances the model's robustness. + +Table 1 presents the time taken for the unlearned concept to resurface when mapped to two concepts at varying semantic similarity. The experimental results demonstrate that mapping an unlearned concept (Golf ball) to a semantically distant concept (Outdoor Recreational activities) enhances the robustness of unlearning. Here, we experimented with the two methods UCE [12] and MACE [20]; clip score (CS) increased only slightly when mapping to the unrelated concept, with minimal or moderate gains in accuracy after unlearning. This indicates that mapping to a semantically distant concept disrupts latent associations more + +effectively, reducing the likelihood of the unlearned concept re-emerging during fine-tuning. + +Conversely, mapping to a semantically closer concept (Tennis ball) retained certain feature similarities; for instance, subtle characteristics like the colour of the Golf ball shifted towards a light green due to the influence of the Tennis ball, while other properties associated with Golf balls remained intact. Even after fine-tuning, this greenish tint persisted, highlighting that the concept is still fundamentally a Golf ball but with minor alterations not reflected in the quantitative metrics. This observation underscores the efficacy of semantically distant mappings for achieving robust concept unlearning in generative models. + +
MethodBefore FTAfter FTRevival Point
CSAccCSAcc
UCE-T0.330.50.330.71
UCE-O0.260.010.280.11R
MACE-T0.320.00.340.543
MACE-O0.270.190.300.15R
SPM-T0.320.270.320.473
SPM-O0.320.290.330.534
+ +Table 1. Clip Score (CS) and Accuracy (Acc) Before and After Fine-tuning When Mapping Golf Ball to Tennis Ball (T) (Semantically Close) and Outdoor Recreational Activities (O) (Semantically Distant). + +# 4.5. Impact of Modules affected during Unlearning + +We chose ESD as the basis for this experiment because it utilizes a setup that allows unlearning by modifying only specific modules (like only cross attention, only non-cross attention and full parameters), enabling targeted analysis of robustness in concept unlearning. + +In our experiments, we used the ESD variations of - ESD-x (cross-attention layers), ESD-u (unconditional layers), and ESD-full (all parameters). We observed that the specific parameters targeted during unlearning significantly impacted the robustness of the concept erasure. Results (Table 2) indicate that cross-attention-based unlearning (ESD-x) is less robust, as the unlearned concept revived more quickly compared to ESD-u and ESD-full. This observation suggests that editing cross-attention weights alone may be insufficient for durable unlearning, as these layers maintain strong associations between prompts and visual features. This vulnerability is also evident in the UCE and CA model, which exclusively modifies cross-attention weights, making it relatively fragile to fine-tuning. Therefore, targeting a broader parameter set, as seen in ESD-u and ESD-full, offers a more resilient approach to concept unlearning. + +
MethodBefore FTAfter FTRevival Point
CSAccCSAcc
ESD-x0.260.00.320.544
ESD-u0.260.00.330.365
ESD-full0.250.00.320.397
+ +# 5. Discussion - Plausible Mitigation Strategy + +For unlearning methods that rely on minimizing a dedicated objective (referred to as the unlearning loss), its value acts as a metric to gauge the success of the erasure. To investigate the impact of fine-tuning on this loss, we evaluated three unlearning techniques—UCE [12], ESD [11], and CA [19]—measuring the unlearning loss both immediately after unlearning and after subsequent fine-tuning (as described in section 4). The results, summarized in Table 3, reveal a consistent increase in unlearning loss post fine-tuning, indicating that fine-tuning destabilizes the unlearning process. Notice that in the case of the CA method with "Angelina Jolie" as the target concept and "Woman" as the surrogate concept, the unlearning loss did not increase after fine-tuning. This suggests no significant change in the extent of unlearning, which is consistent with previous experiments (Fig 5), where fine-tuning resulted in negligible changes to the model's unlearning accuracy for this particular target/surrogate concept pairing. + +A plausible approach to mitigate the revival could involve meta-learning and gradient alignment techniques. While standard fine-tuning minimizes the diffusion model's loss, its gradients may conflict with the unlearning objective. To resolve this, we propose modifying the fine-tuning process to (1) compute gradients for both the diffusion loss and unlearning objective, (2) project the diffusion gradients orthogonal to the unlearning gradient direction, and (3) update parameters using only the orthogonal component. This ensures that the unlearning loss remains unchanged throughout fine-tuning, preserving the unlearning effect. A similar approach has been employed by Ko et al. [18] and Wu et al. [40] to achieve unlearning while maintaining the generational quality of neutral concepts. Their approach optimizes two objectives—one for preserving the generation quality of neutral concepts and another for unlearning the target concept by ensuring that each update step doesn't increase either objective function. + +Table 2. Comparison of Clip Score (CS) and Accuracy (Acc) Before and After Unlearning for Different ESD Module Configurations. + +
MethodUnlearn LossUnlearn + FTTarget / Surrogate
UCE5.54 × 1016.74 × 102Dog / Cat
1.08 × 1061.65 × 106Van Gogh / Art
6.66 × 1011.03 × 103Pikachu / Cat
9.02 × 1051.23 × 106Cartoon / Art
1.23 × 1029.41 × 102Brad Pitt / Man
ESD5.00 × 10-41.10 × 10-3Golf Ball / NA
3.00 × 10-45.00 × 10-4Van Gogh / NA
4.00 × 10-49.00 × 10-4Nudity / NA
3.00 × 10-41.60 × 10-3Brad Pitt / NA
CA4.79 × 10-47.53 × 10-4Golf Ball / Ball
1.05 × 10-41.49 × 10-4Van Gogh / Art
≈ 0≈ 0Ang. Jolie / Woman
+ +Table 3. Comparison of unlearning loss values between: (1) models immediately after unlearning, and (2) the same models after fine-tuning (using weights obtained from Section 4 experiments). Loss values were computed with each method's respective objective function, with corresponding target-surrogate concept pairings shown. The parameter group updated in ESD and CA for these experiments is cross-attention. The last column shows the target concept (Unlearned concept) along with the corresponding mapping concept i.e Surrogate concept. NA means No specific Surrogate Concept. + +# 6. Conclusions + +We present a critical vulnerability in the current Machine Unlearning techniques for text-to-image diffusion models. Our study emphasizes the need for (i) more robust techniques that can erase the semantic space of the target concept in the latent space of the model, in which case finetuning can never revive the target concept, or/and (ii) the method should be able to achieve mapping of target concept to a surrogate concept which is robust to finetuning. By revealing the weaknesses of existing methods against finetuning and providing a framework to measure concept revival, we aim to guide future research toward developing more secure, reliable, and robust unlearning mechanisms. Below, We list some interesting and important directions to further our understanding of the unlearning process. + +- Rigorous understanding of why the finetuning process revives the unlearned concepts across several unlearning approaches. Though challenging, it is of utmost importance to explain this vulnerability of the existing MU methods and to achieve robust unlearning. +- Relation between concept revival and finetuning concept. We have observed that certain finetuning concepts can cause greater revival while others less. Also, this dependence may vary across unlearned concepts and unlearning methods. Researching the primary cause of the variance in dependency between the finetuning concept and the revival of the unlearning concept can aid in assessing the robustness of the models better. + +# References + +[1] Stability AI. Stability ai. https://stability.ai/news/celebrating-one-year-of-stable-diffusion, 2023. 1 +[2] Justin Betker, Gabriel Goh, Li Jing, Tim Brooks, Jason Wang, Lilian Li, Aditya Ramesh, et al. Improving image generation with better captions. Computer Science, 2(3):8, 2023. 1 +[3] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy (SP), pages 463-480. IEEE, 2015. 1 +[4] Mingjie Chen, Wei Gao, Guangyue Liu, Kai Peng, and Cunhua Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7766-7775, 2023. 1 +[5] Lydia de la Torre. A guide to the california consumer privacy act of 2018. Available at SSRN 3275571, 2018. 1 +[6] NudeNet Developers. Nudenet: A nudity detection and censoring library. https://github.com/notAI-tech/NudeNet, 2023.5 +[7] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5 +[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 5 +[10] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 4, 5 +[11] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 7, 8 +[12] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 3, 4, 6, 7, 8 +[13] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1 + +[14] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9301-9309. IEEE, 2020. 1 +[15] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 7 +[16] Chris Jay Hoofnagle, Bart van der Sloot, and Frederik Zuiderveen Borgesius. The european union general data protection regulation: what it is and what it means. Information & Communications Technology Law, 28(1):65-98, 2019. 1 +[17] Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung-Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. Receiver: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. In European Conference on Computer Vision, pages 360–376. Springer, 2024. 1, 2, 3, 7 +[18] Myeongseob Ko, Henry Li, Zhun Wang, Jonathan Patsenker, Jiachen T. Wang, Qinbin Li, Ming Jin, Dawn Song, and Ruoxi Jia. Boosting alignment for post-unlearning text-to-image generative models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 8 +[19] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 3, 8 +[20] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 2, 6, 7 +[21] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 1, 2, 3, 7 +[22] Midjourney. Midjourney. https://www.midjourney.com/, 2023.1 +[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pooja Shyam, Pavel Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1 +[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1 +[25] OpenAI. Chatgpt model api, 2023. 6 +[26] OpenAI. Openai embeddings, 2023. Accessed: 2024-11-13. 6 +[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models + +for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1 +[28] Formerly Data Protection. General data protection regulation (gdpr). Intersoft Consulting, accessed in October, 2018. 1 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5 +[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1 +[31] Aditya Ramesh, Prafulla Dhariwal, Alexander Quinn Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1 +[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Sahand Mahdavi, Raphael Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[34] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2 +[35] Christoph Schuhmann, Robin Beaumont, Romain Vencu, Craig Gordon, Ross Wightman, Mehdi Cherti, Teddy Coombes, Abhay Katta, and Vladimir Kirillov. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278–25294, 2022. 1 +[36] Anish Sekhari, Jayadev Acharya, Gauri Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. In Advances in Neural Information Processing Systems, pages 18075-18086, 2021. 1 +[37] GenAI statistics. Genai statistics. https://journaleverypixel.com/ai-image-statistics, 2023.1 +[38] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? In The Twelfth International Conference on Learning Representations, 2024. 3 +[39] Jing Wu, Trung Le, Munawar Hayat, and Mehrtash Harandi. Erasediff: Erasing data influence in diffusion models, 2024. 1, 2, 3, 4 + +[40] Yongliang Wu, Shiji Zhou, Mingzhuo Yang, Lianzhe Wang, Heng Chang, Wenbo Zhu, Xinting Hu, Xiao Zhou, and Xu Yang. Unlearning concepts in diffusion model via concept domain correction and concept preserving gradient. In Proceedings of the 39th AAAI Conference on Artificial Intelligence (AAAI), 2025. 8 +[41] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1755-1764, 2024. 1 +[42] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now. In Computer Vision - ECCV 2024: 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LVII, page 385-403, Berlin, Heidelberg, 2024. Springer-Verlag. 3 +[43] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. In Thirty-eighth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2, 3, 5 \ No newline at end of file diff --git a/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/images.zip b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e9909ca9247b2480867d12e03550f98a44ff272 --- /dev/null +++ b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99679d6fb34f1484bc927b24cd99c47aa3a8002e244202d745815e484d6ec150 +size 439668 diff --git a/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/layout.json b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dca0be4c262da05963aaa2f27ec4edcd5057f4fa --- /dev/null +++ b/2025/The Illusion of Unlearning_ The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models/layout.json @@ -0,0 +1,6040 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 537, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 537, + 138 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 537, + 138 + ], + "type": "text", + "content": "The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "spans": [ + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "text", + "content": "Naveen George" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "text", + "content": ", Karthik Nandan Dasaraju" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "text", + "content": ", Rutheesh Reddy Chittepu" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "text", + "content": ", Konda Reddy Mopuri" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "text", + "content": "Data-driven Intelligence and Learning Lab, Indian Institute of Technology, Hyderabad \n" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 160, + 538, + 204 + ], + "type": "text", + "content": "Indian Institute of Technology, Ropar" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 211, + 517, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 211, + 517, + 236 + ], + "spans": [ + { + "bbox": [ + 91, + 211, + 517, + 236 + ], + "type": "text", + "content": "ai23mtech12001@iith.ac.in, 2021csb1081@iitrpr.ac.in, cs21btech11014@iith.ac.in, krmopuri@ai.iith.ac.in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 264, + 200, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 264, + 200, + 277 + ], + "spans": [ + { + "bbox": [ + 151, + 264, + 200, + 277 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 289, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 297, + 542 + ], + "type": "text", + "content": "Text-to-image models such as Stable Diffusion, DALL-E, and Midjourney have gained immense popularity lately. However, they are trained on vast amounts of data that may include private, explicit, or copyrighted material used without permission, raising serious legal and ethical concerns. In light of the recent regulations aimed at protecting individual data privacy, there has been a surge in Machine Unlearning methods designed to remove specific concepts from these models. However, we identify a critical flaw in these unlearning techniques: unlearned concepts will revive when the models are fine-tuned, even with general or unrelated prompts. In this paper, for the first time, through an extensive study, we demonstrate the unstable nature of existing unlearning methods in text-to-image diffusion models. We introduce a framework that includes a couple of measures for analyzing the stability of existing unlearning methods. Further, the paper offers preliminary insights into the plausible explanation for the instability of the mapping-based unlearning methods that can guide future research toward more robust unlearning techniques. Codes1 for implementing the proposed framework are provided." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 548, + 135, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 135, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 135, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 687 + ], + "type": "text", + "content": "The past decade has witnessed a rapid qualitative growth in text-to-image models [7, 23, 24, 32, 33], leading to its widespread attention and adoption. Models such as DALL-E [2, 30, 31], MidJourney [22], and Stable Diffusion [27, 32] enable high-quality image generation from simple text prompts, opening applications across art, design, advertising, gaming, and even medical imaging. With millions worldwide adopting these tools for academic and commercial purposes, their transformative potential is evident, as illustrated by MidJourney's 16 million active" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 266, + 554, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 338 + ], + "type": "text", + "content": "users [22] and the generation of over 2 million images daily by DALL-E and Stable Diffusion [1, 37]. However, this widespread use raises critical concerns about data privacy, security, fairness, and intellectual property, as these models rely on vast datasets that often include sensitive, personal, or copyrighted content [35]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 495 + ], + "type": "text", + "content": "In response to these privacy concerns, regulations such as the General Data Protection Regulation (GDPR) [28] and the California Consumer Privacy Act (CCPA) [5] provide users the Right to be Forgotten [16], promoting the development of Machine Unlearning (MU) techniques [3, 13]. MU aims to remove specific data or concepts from a model without the computational expense of retraining from scratch. While initially explored in classification models [4, 13, 14, 36], MU in generative models, particularly diffusion-based ones, poses unique challenges due to the models' complex, distribution-based learning. Ensuring these models forget sensitive or harmful information while retaining general functionality remains an open problem." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 496, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 556, + 689 + ], + "type": "text", + "content": "Machine Unlearning in text-to-image models focuses on erasing specific concepts from the model's generative capacity. The general goal of MU in this context is to ensure that when a model is conditioned on a prompt related to the target or \"unlearned\" concept, it no longer generates images reflecting that concept. These target concepts can be Not Safe For Work (NSFW) content [11, 15, 17, 21, 34], objects [10-12, 15, 19-21, 39, 41] or artistic styles [11, 12, 15, 19-21, 39, 41]. In other words, the unlearned concept is the one we want the model to \"forget\". To achieve this, unlearning algorithms typically remap the target concept to a surrogate concept. This means that when the model is conditioned on the unlearned concept, it produces images associated with the surrogate concept instead. The surrogate concept can be a general or neutral concept [12, 15, 19], a random concept [10], or even an empty string [11, 39]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Despite these efforts in Machine Unlearning, we have identified a critical vulnerability in the current state-of-the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 217, + 713 + ], + "type": "text", + "content": "*Equal contribution. \n1https://github.com/NGK2110/TIU" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13393" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "content": "art methods for text-to-image models. Specifically, we observe that when these unlearned models undergo finetuning, even with general or unrelated concepts, the previously unlearned concepts can resurface in the generated images. This unintended revival undermines the unlearning process's effectiveness and poses significant risks to data privacy and legal compliance. The ability of an unlearned model to regenerate sensitive or copyrighted content after benign fine-tuning highlights a fundamental weakness in the existing MU approaches for generative models. Such vulnerabilities could lead to unauthorized dissemination of protected content, violating individual privacy rights and intellectual property laws. Therefore, it is imperative to (i) study the robustness of the existing MU methods, (ii) understand the underlying causes of this behaviour and develop more robust unlearning strategies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 267, + 288, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 288, + 278 + ], + "type": "text", + "content": "Contributions of the paper can be summarized as follows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 283, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 450 + ], + "type": "text", + "content": "- Systematic evaluation of MU methods: We conduct a comprehensive study of the existing MU methods in text-to-image diffusion models, demonstrating that unlearned concepts are revived across all examined approaches. Our analysis includes all the leading methods — SalUn [10], ESD [11], EDiff [39], CA [19], MACE [20], SPM [21], SA [15], Receler [17] and UCE [12] —capturing the full spectrum of current state-of-the-art unlearning techniques. To the best of our knowledge, this is the first work to identify and discuss this issue in depth, examining the robustness of unlearning. We believe this work brings an unknown but critical aspect of the MU methods into light, paving the way to a more resilient framework for unlearning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 451, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 557 + ], + "type": "text", + "content": "- Concept Revival Framework: We introduce a framework demonstrating that fine-tuning revives the unlearned concepts. For each target concept, the proposed framework presents a customized curriculum of concepts in the increasing order of relevance to fine-tune the unlearned model. During the fine-tuning, the framework enables the measurement of the revival of the target concept via a couple of simple measures: classification accuracy and CLIP score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 594 + ], + "type": "text", + "content": "- Analysis and insights about the unlearning process: We study the fragility of the existing MU methods with respect to two important axes:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 594, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 594, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 63, + 594, + 295, + 654 + ], + "type": "text", + "content": "- Semantic relevance of the surrogate concept to the target concept: we analyze how mapping to surrogate concepts of varying semantic relevance (to the target concept) affects the stability of the resulting unlearning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 654, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 654, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 63, + 654, + 295, + 701 + ], + "type": "text", + "content": "- Modules affected during the unlearning: we examine how the nature of model parameters modified during unlearning, (i.e., cross-attention module vs. the rest), affects the unlearning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 63, + 701, + 295, + 713 + ], + "type": "text", + "content": "Further, we hypothesize a plausible explanation for the in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 322, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 144 + ], + "type": "text", + "content": "stability of the mapping-based MU methods - fine-tuning the unlearned model disrupts the mapping between the target and surrogate concepts. We provide empirical evidence for this hypothesis over some of the popular MU methods [11, 12, 19], in terms of the target concept's unlearning objective function value during the fine-tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 155, + 406, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 155, + 406, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 155, + 406, + 167 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 175, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 342 + ], + "type": "text", + "content": "Unlearning in image generation. Machine Unlearning (MU) approaches aim to remove the influence of certain data points or concepts from a model without the need for full retraining, which is computationally intensive for large-scale models like diffusion models. The rise of MU methods has been driven by their potential to address various issues in generative models. For example, text-to-image (T2I) diffusion models have a tendency to generate inappropriate or sensitive content, such as depictions of violence, deepfakes, or copyrighted artistic styles. MU methods are recognized for their role in promoting data privacy and complying with regulations like the \"Right to be Forgotten\" by enabling models to unlearn specific concepts deemed undesirable or unlawful." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 461 + ], + "type": "text", + "content": "Several methods have been proposed to achieve concept unlearning in diffusion models. Unified Concept Editing (UCE) [12] modifies cross-attention in the U-Net component of diffusion models using a closed-form solution to map target concepts to surrogate ones, effectively erasing them without additional training. Mass Concept Erasure (MACE) [20] extends this by refining cross-attention layers and utilizing Low-Rank Adaptation (LoRA) modules to efficiently remove multiple unwanted concepts while preserving the model's ability to generate unrelated content." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 582 + ], + "type": "text", + "content": "Other approaches like Selective Amnesia (SA) [15] leverage Elastic Weight Consolidation and Generative Replay to forget target concepts while maintaining overall model performance. Erased Stable Diffusion (ESD) [11] aligns the probability distribution of a concept to be erased with that of a null string, although it may lead to a collapse in generative quality. Concept Ablation (CA) [19] fine-tunes cross-attention weights to map target concepts to more general anchor concepts, effectively preventing the generation of specific styles or objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": "Methods such as Safe Latent Diffusion (SLD) [34] use reverse guidance during generation to steer away from inappropriate content but require careful balancing of prompt conditioning and hyperparameters. The Semi-Permeable Membrane (SPM) [21] adapter offers a lightweight solution for multi-concept erasure by introducing a regulated erasing signal within intermediate layers, although it may struggle with partial unlearning (traces of the target concept remain)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "While these methods can effectively unlearn the target concept, they often adversely affect the generative capability for related concepts [43] and may degrade the over" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13394" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 79, + 532, + 318 + ], + "blocks": [ + { + "bbox": [ + 78, + 79, + 532, + 318 + ], + "lines": [ + { + "bbox": [ + 78, + 79, + 532, + 318 + ], + "spans": [ + { + "bbox": [ + 78, + 79, + 532, + 318 + ], + "type": "image", + "image_path": "4cd222d6fe147dcc0ca652140eeb5e7ae897b03affb79eaab2b0312423bd9077.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 321, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 54, + 321, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 54, + 321, + 555, + 368 + ], + "type": "text", + "content": "Figure 1. Revival of unlearned concepts is observed across numerous concepts and unlearning methods. The top row presents images generated by the original SDv1.4 model. The middle row shows images from models with specific concepts unlearned. The unlearning concept and the method used are at the bottom of each image. The last row shows images generated when the unlearned model is finetuned on several different concepts. At the bottom of each column is the prompt used for sampling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 387, + 298, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 387, + 298, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 387, + 298, + 473 + ], + "type": "text", + "content": "all image quality. Additionally, the unlearned concepts can sometimes reappear if an adversarial prompt is carefully engineered, indicating incomplete unlearning and vulnerabilities to prompt-based attacks [38, 42]. There are also adversarial robust unlearning methods [17] that are trained to withstand these adversarial prompts without compromising the generation of non-target concepts (i.e., locality)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 483, + 183, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 183, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 183, + 497 + ], + "type": "text", + "content": "3. Proposed Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 280, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 280, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 280, + 517 + ], + "type": "text", + "content": "3.1. Fine-tuning revives the unlearned concepts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 522, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 297, + 715 + ], + "type": "text", + "content": "Numerous methods have been proposed to unlearn a wide range of target concepts from diffusion models. These targets range from objects and artistic styles to global concepts such as nudity. We present the effect of fine-tuning these unlearned models on different concepts in a systematic way. Figure 1 presents the observed effect concisely, where the top row shows images generated by the original Stable Diffusion v1.4. The second row shows images generated by the same model after unlearning different concepts using numerous unlearning methods[10-12, 15, 17, 19-21, 39]. These works demonstrated a good unlearning performance, which is evident from the absence of the unlearning concept in the images. The performance of these works was also quantified by UnlearnCanvas [43] and shown to offer good unlearning accuracy. After fine-tuning various concepts sequentially (Section 4), images generated by the model are" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 387, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 411 + ], + "type": "text", + "content": "shown in the bottom row of the exact figure. Revival of the unlearned concept is observed in all cases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "text", + "content": "This revival of unlearned concepts with fine-tuning on arbitrary concept(s) is a critical vulnerability. A model published to have unlearned a sensitive concept can be made to relearn by fine-tuning, posing a threat to the publisher. This is more pressing if the concept being unlearned is nudity, NSFW, or copyrighted content. A model " + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "inline_equation", + "content": "M_{u}" + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "text", + "content": " trained by some unlearning mechanism is compared with the ideal model " + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\overline{M}_{u}" + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "text", + "content": ", which has never seen the unlearning concept before, i.e. the model trained on a dataset with unlearning concept removed. " + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\overline{M}_{u}" + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "text", + "content": " that has never known an unlearned concept (golf ball for example), won't infer its existence when fine-tuned on images of different concept (piano keyboard). The revival of the unlearned concept in " + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "inline_equation", + "content": "M_{u}" + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "text", + "content": " shows that its performance is far from " + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\overline{M}_{u}" + }, + { + "bbox": [ + 312, + 412, + 556, + 592 + ], + "type": "text", + "content": ", questioning the unlearning mechanism used." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "content": "We provide quantitative results for the revival of unlearned concepts while fine-tuning. CLIP [29] score calculated from the embeddings quantifies the semantic similarity between a text prompt and an image. Figure 2 shows the effect of fine-tuning on unlearned model with an unrelated concept. We prepared prompts containing golf ball as a validation dataset to evaluate the ability of the model to generate images of golf ball. At every epoch of fine-tuning, the model is sampled to generate images for validation prompts and the mean CLIP score is calculated for the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13395" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "embeddings of generated images and the prompts. As expected, the score is initially low, proving that the model has unlearned golf ball. An increase in the alignment of generated images towards golf ball with every epoch is observed, depicting the revival of the unlearned concept." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 144, + 293, + 284 + ], + "blocks": [ + { + "bbox": [ + 58, + 144, + 293, + 284 + ], + "lines": [ + { + "bbox": [ + 58, + 144, + 293, + 284 + ], + "spans": [ + { + "bbox": [ + 58, + 144, + 293, + 284 + ], + "type": "image", + "image_path": "5de54fa76ef4b87998264e27b65b6daf84582e6fd85b256b89dec195ca7e1c38.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 293, + 295, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 338 + ], + "type": "text", + "content": "Figure 2. Graph showing the variation of mean CLIP score with epochs. Fine-tuning is done on an unrelated concept \"piano keyboard\", and an increase is observed indicating the revival of unlearned concept \"Golf Ball\". ESD[39] was used for unlearning." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 359, + 272, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 272, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 272, + 372 + ], + "type": "text", + "content": "3.2. Dependence of revival on fine-tuning data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 378, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 520 + ], + "type": "text", + "content": "The previous subsection establishes that fine-tuning with certain unrelated concepts (hereafter referred to as fine-tuning concepts) can revive the unlearned concept. However, we observed that some MU algorithms demonstrated trivial revival compared to others when fine-tuned on the same unrelated concept (Figure 3). It has been observed that the extent of revival depends on the fine-tuning dataset, especially the fine-tuning concept. This makes it a challenge to assess the robustness of an unlearned model. While no revival is observed on a particular fine-tuning dataset, concept(s) that were not considered might exist and cause significant revival." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "content": "Analysis of the effect of different fine-tuning concepts is required to draw insights into the revival of unlearned concepts. The extent of fine-tuning (i.e., the number of epochs) required to observe the revival of unlearned concepts can serve as a good indicator of robustness. When the CLIP score crosses a predefined threshold, one may declare the revival of the unlearned concept. Figure 3 shows the results of such an experiment. Unlearned methods SalUn [10] and UCE [12] are considered to unlearn \"golf ball\" and fine-tuning was performed on concepts indicated on the vertical axis. The horizontal axis represents the number of epochs needed to achieve a CLIP score threshold of 0.31 (explained in Section 4). We considered fine-tuning till a maximum of 91 epochs, indicating the existence of fine-tuning concepts that don't cause revival even after prolonged fine-tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "From Figure 3, it can be observed that a wide range of" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 180 + ], + "type": "image", + "image_path": "1d3fbba208cbcbe981b54f29057bb6bf2344818488fc01c49d5b866839998803.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 193, + 553, + 302 + ], + "blocks": [ + { + "bbox": [ + 317, + 193, + 553, + 302 + ], + "lines": [ + { + "bbox": [ + 317, + 193, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 317, + 193, + 553, + 302 + ], + "type": "image", + "image_path": "f8e48bf49f0722b4dc5e6ebe8f27339befeef115de312d0eb19ae71447422626.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "type": "text", + "content": "Figure 3. Illustration of the number of epochs of fine-tuning required to observe revival of unlearned concept \"golf ball\". Finetuning concepts are shown on the vertical axis. The same finetuning concept can have different effects on revival for different unlearning methods (UCE and SalUn shown)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 401, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 688 + ], + "type": "text", + "content": "fine-tuning concepts may be required to study the revival of unlearned concepts. A fine-tuning concept that caused revival for one work may not be prominent in revival for others. Further, we also observed that sequentially fine-tuning the unlearned model shows greater revival in a shorter time (number of fine-tuning epochs). Comparing Figure 3 (bottom panel) with Figure 4 makes this evident. Figure 4 shows the mean CLIP score when fine-tuning is performed in a sequential manner on concepts mentioned on the horizontal axis. A model that unlearned \"golf ball\" with SalUn [10] method was fine-tuned with each concept for 8 epochs sequentially (after fine-tuning on one concept, we continued fine-tuning on the next). While in Figure 3 (bottom panel), trivial revival was observed when fine-tuned individually on \"Volcano, Fish, Swimming, Hot Dog, Cricket Bat\" (indicated by large number of fine-tuning epochs), a CLIP score of 0.31 was reached in Figure 4 when fine-tuned sequentially. Specifically, for the fine-tuning concept \"Cricket Bat\", no revival was observed even after fine-tuning for a large number of epochs (91 in Figure 3 bottom panel). However, a significant revival (0.31 CLIP score) was observed at the end of fine-tuning \"Cricket Bat\" in a sequential manner for relatively less fine-tuning (" + }, + { + "bbox": [ + 313, + 401, + 555, + 688 + ], + "type": "inline_equation", + "content": "5 \\times 8 = 40" + }, + { + "bbox": [ + 313, + 401, + 555, + 688 + ], + "type": "text", + "content": " epochs in Figure 3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Hence, fine-tuning with various prompts may be required to study the revival of unlearning concepts. Further, the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13396" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 290, + 175 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 290, + 175 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 290, + 175 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 290, + 175 + ], + "type": "image", + "image_path": "48741fadc55cba1a912060307483e4a024c730de134ec18cb5be66b8dab2b5e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 185, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 295, + 251 + ], + "type": "text", + "content": "Figure 4. Graph showing Mean CLIP score after fine-tuning on the concepts given on horizontal axis sequentially. The unlearned model used was produced by using SalUn[10] to remove target concept \"Golf Ball\". The highlighted point is when fine-tuning is done sequentially until Cricket Bat, and significant revival is shown as indicated by the CLIP score." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "type": "text", + "content": "sequential fine-tuning of concepts seems to accelerate the revival. An experimental \"Concept revival framework\" to evaluate MU methods on this basis is detailed in section 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 281, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 281, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 281, + 335 + ], + "type": "text", + "content": "4. Experiments - Studying the Vulnerability" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 340, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 400 + ], + "type": "text", + "content": "We investigate the critical vulnerability aforementioned in the current Machine Unlearning (MU) methods for text-to-image diffusion models. Specifically, we demonstrate through extensive experiments that fine-tuning an unlearned model can lead to reviving previously unlearned concepts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 172, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 172, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 172, + 422 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 474 + ], + "type": "text", + "content": "We conducted experiments on state-of-the-art MU methods applied to the Stable Diffusion v1.4 model to systematically study the concept revival. We describe our experimental setup:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 474, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 295, + 534 + ], + "type": "text", + "content": "- Unlearning: We conducted experiments in four different unlearning settings, covering a total of ten concepts. In each case, unlearning is crucial for privacy, security, and ethical considerations, as it ensures certain content can no longer be generated by the model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 534, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 534, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 64, + 534, + 295, + 569 + ], + "type": "text", + "content": "- Object Unlearning focuses on unlearning specific objects from the model. Experiments are performed on \"Golf Ball, Fukushima and Dog\"." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 570, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 570, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 64, + 570, + 295, + 616 + ], + "type": "text", + "content": "- Style Unlearning aims to remove the ability to replicate certain artistic styles using the model. \"Van Gogh, Picasso\" and \"Cartoon\" styles are used in our experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 618, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 618, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 64, + 618, + 295, + 653 + ], + "type": "text", + "content": "- Celebrity Unlearning aims to erase the model's ability to generate content of famous celebrities. We decided to use \"Angelina Jolie, Brad Pitt\" and \"Lionel Messi\"." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 654, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 654, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 64, + 654, + 295, + 688 + ], + "type": "text", + "content": "- NSFW Content Unlearning focuses on preventing the model from generating inappropriate or explicit content." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 295, + 714 + ], + "type": "text", + "content": "- Evaluation: A set of validation prompts is composed that contains the unlearning concept. The model is then sam" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "content": "pled to generate images and evaluated using two main measures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 96, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 96, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 322, + 96, + 553, + 132 + ], + "type": "text", + "content": "- CLIP Score (CS): Measures the cosine similarity between the embeddings of generated images and their corresponding text prompts using the CLIP model [29]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 133, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 133, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 323, + 133, + 555, + 264 + ], + "type": "text", + "content": "- Classifier Accuracy (Acc): Utilizes a fine-tuned Vision Transformer (ViT) [8] as a binary classifier to detect the presence of the unlearned concept in generated images. For Object and Celebrity Unlearning, we fine-tuned a VIT model on images of unlearning concepts and other unrelated images generated using the Stable Diffusion-3-medium [9]. In the case of Style Unlearning, we use the classifier from UnlearnCanvas [43] and fine-tune it using images generated by Stable Diffusion-3-medium [9]. Finally, for NSFW unlearning, we employ the NudeNet classifier [6]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 275, + 556, + 706 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 275, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 598 + ], + "type": "text", + "content": "- Revival Point: Tracking the instance when a significant revival of the unlearned concept is observed can serve as a good measure for assessing the robustness of the unlearning method. We record the number of concepts on which fine-tuning was performed sequentially until revival was observed as \"Revival Point\". Further, to show no bias towards an unlearning method, we fix the order of fine-tuning concepts in the increasing order of semantic similarity from the target concept. The model is fine-tuned on each concept for 8 epochs during this process. A concept is considered to have revived when the model's CLIP score exceeds a specific threshold, and the classifier accuracy surpasses 0.3 (a conservative value that indicates revival). If the fine-tuning fails to revive the target concept in this setup, the MU method is comparatively robust (indicated by R). To calculate the clip threshold required for each target concept, we used images generated from base Stable-Diffusion v1.4 on prompts from the validation dataset (detailed later). Then, we calculated the CLIP score between the image and prompt embeddings. We set the CLIP threshold for to a value 0.02 less than the SD's CLIP score. CLIP threshold for each unlearning concept will be different, depending on the original model's CLIP score (thresholds for each concept are in Supplementary section B). In all the cases where the threshold was reached, a significant revival of the unlearned concept was observed." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 598, + 556, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 598, + 556, + 706 + ], + "spans": [ + { + "bbox": [ + 314, + 598, + 556, + 706 + ], + "type": "text", + "content": "- Fine-tuning Dataset: To fine-tune the unlearned model, we create a customized dataset by first identifying concepts semantically related to the target concept. We achieve this by designing detailed prompts that specify the type of relationships at each hierarchical level and input these prompts into the OpenAI GPT 4o model. This process generates a comprehensive list of related concepts, which we then use to generate diverse prompts for fine-tuning the model. We compute the cosine" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13397" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 85, + 553, + 225 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 405, + 84 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 405, + 84 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 405, + 84 + ], + "type": "text", + "content": "Clip Scores Before and After Fine-tuning" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 85, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 553, + 225 + ], + "type": "image", + "image_path": "30b8f9f60a4d0d7ff40049f1a063d6a9e7b824e24c1528a815f0563f5134162f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 247, + 553, + 403 + ], + "blocks": [ + { + "bbox": [ + 212, + 234, + 421, + 246 + ], + "lines": [ + { + "bbox": [ + 212, + 234, + 421, + 246 + ], + "spans": [ + { + "bbox": [ + 212, + 234, + 421, + 246 + ], + "type": "text", + "content": "Classifier Accuracy Before and After Fine-tuning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 247, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 62, + 247, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 553, + 403 + ], + "type": "image", + "image_path": "e33a0b85bbe7c76714c3209f1cce31015e988742f162f9d0ac18e149d0f71939.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 415, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 555, + 460 + ], + "type": "text", + "content": "Figure 5. Graphs illustrate the revival of Unlearned concepts on fine-tuning across different concepts and methods. CLIP score and Classifier accuracy are used to quantify the revival. Each colour corresponds to a different method, with lighter bar segments representing the increase in metrics before and after Fine-tuning, respectively. The consistent upward shifts in bar heights indicate that the unlearned concepts have been successfully revived." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 481, + 297, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 481, + 297, + 613 + ], + "spans": [ + { + "bbox": [ + 63, + 481, + 297, + 613 + ], + "type": "text", + "content": "similarity between these concepts and the target concept using OpenAI embeddings [26] and arrange them in increasing order of similarity. For each selected concept, we generate 50 prompts using OpenAI GPT [25] and generate corresponding images using Stable Diffusion 3 Medium. This results in a dataset that enables curriculum fine-tuning, in which more general concepts appear before the more related concepts during the fine-tuning. This approach allows us to assess how fine-tuning concepts with different semantic similarities affects the revival of the unlearned concept." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 297, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 297, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 297, + 711 + ], + "type": "text", + "content": "- Validation Dataset: We created a custom dataset of 20 prompts related to the unlearned concept using the OpenAI API. These prompts were carefully crafted to cover various descriptions associated with the target concept. We generated 5 images for each prompt using the fine-tuned models, resulting in a total of 100 images over which the CLIP score and classification accuracy mea" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 482, + 405, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 482, + 405, + 494 + ], + "spans": [ + { + "bbox": [ + 323, + 482, + 405, + 494 + ], + "type": "text", + "content": "sures are computed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 494, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 529 + ], + "type": "text", + "content": "Additional details of the experimental setup are provided in the Supplementary Materials (see Supplementary Sections C and D)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 539, + 538, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 538, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 538, + 551 + ], + "type": "text", + "content": "4.2. Revival in the state-of-the-art MU methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 557, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 617 + ], + "type": "text", + "content": "Following the explanations in subsection 3.2, we sequentially fine-tune the unlearned model for 10 concepts. We report the two measures, CLIP score and Classifier accuracy, before and after the fine-tuning. An increase in these indicates the revival of unlearned concepts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "From Figure 5, it is evident that revival is observed across different concepts and unlearning methods. We have observed that MACE[20] and UCE[12] were relatively robust across many concepts, with less revival observed after fin-tuning. Upon further analysis of different types of unlearning, UCE[12] was found to be robust in object unlearning. We found almost all methods vulnerable to fine-tuning in case of style unlearning. We can infer that artistic styles like" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13398" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "Van Gogh and Picasso are inherently complex and challenging to unlearn precisely. In the case of celebrity unlearning, we observed that there is an increase in metrics across all works with UCE's[12] change being minimal. Nudity unlearning was easily broken with relatively faster revival in the case of ESD[11], SPM[21] and SA[15] compared to MACE[20] which was robust. Even with adversarial robust MU method[17] which is robust against adversarial prompts we can observe concept revival (low for object and style but high for NSFW content). More experimental details are provided in the Supplementary Material (see Supplementary Section B)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 226, + 291, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 291, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 291, + 239 + ], + "type": "text", + "content": "4.3. Framework for Studying the Concept Revival" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 245, + 294, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 294, + 305 + ], + "type": "text", + "content": "Based on previous sections, it is apparent that model robustness should be assessed with its ability to safeguard against the revival of unlearning concepts. This section details the proposed framework for investigating the concept revival in the unlearned diffusion model due to fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 308, + 295, + 379 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 308, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 344 + ], + "type": "text", + "content": "1. Step 1: Fine-tune the unlearned model sequentially on a semantically distant concept for a fixed number of epochs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 345, + 295, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 345, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 56, + 345, + 295, + 379 + ], + "type": "text", + "content": "2. Step 2: Evaluate whether the CLIP score and classifier accuracy for the unlearned concept crosses a predefined threshold called the Revival Point." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 380, + 295, + 427 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 67, + 380, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 380, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 67, + 380, + 295, + 403 + ], + "type": "text", + "content": "- If Not: Proceed to fine-tune on a semantically closer concept." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 404, + 294, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 404, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 67, + 404, + 294, + 427 + ], + "type": "text", + "content": "- If Yes: Record the Revival Point at which the unlearned concept resurfaces." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "type": "text", + "content": "3. Step 3: Repeat the process by gradually decreasing the semantic distance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 456, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 516 + ], + "type": "text", + "content": "This framework provides a quantitative measure of the model's robustness based on the point at which the unlearned concept reappears and requires less number of epochs compared to fine-tuning separately using each concept." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 526, + 214, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 214, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 214, + 539 + ], + "type": "text", + "content": "4.4. Impact of Surrogate Concept" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "content": "Our empirical studies indicate that mapping the unlearned concept to a semantically distant or general concept enhances the model's robustness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Table 1 presents the time taken for the unlearned concept to resurface when mapped to two concepts at varying semantic similarity. The experimental results demonstrate that mapping an unlearned concept (Golf ball) to a semantically distant concept (Outdoor Recreational activities) enhances the robustness of unlearning. Here, we experimented with the two methods UCE [12] and MACE [20]; clip score (CS) increased only slightly when mapping to the unrelated concept, with minimal or moderate gains in accuracy after unlearning. This indicates that mapping to a semantically distant concept disrupts latent associations more" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "effectively, reducing the likelihood of the unlearned concept re-emerging during fine-tuning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 100, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 100, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 100, + 555, + 232 + ], + "type": "text", + "content": "Conversely, mapping to a semantically closer concept (Tennis ball) retained certain feature similarities; for instance, subtle characteristics like the colour of the Golf ball shifted towards a light green due to the influence of the Tennis ball, while other properties associated with Golf balls remained intact. Even after fine-tuning, this greenish tint persisted, highlighting that the concept is still fundamentally a Golf ball but with minor alterations not reflected in the quantitative metrics. This observation underscores the efficacy of semantically distant mappings for achieving robust concept unlearning in generative models." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 326, + 248, + 543, + 347 + ], + "blocks": [ + { + "bbox": [ + 326, + 248, + 543, + 347 + ], + "lines": [ + { + "bbox": [ + 326, + 248, + 543, + 347 + ], + "spans": [ + { + "bbox": [ + 326, + 248, + 543, + 347 + ], + "type": "table", + "html": "
MethodBefore FTAfter FTRevival Point
CSAccCSAcc
UCE-T0.330.50.330.71
UCE-O0.260.010.280.11R
MACE-T0.320.00.340.543
MACE-O0.270.190.300.15R
SPM-T0.320.270.320.473
SPM-O0.320.290.330.534
", + "image_path": "eb8438838cd0611da9c4ab1cdbeb8d7bb8333fb4fe5ca16544cbb68b80ff3c9f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 356, + 553, + 400 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 553, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 553, + 400 + ], + "type": "text", + "content": "Table 1. Clip Score (CS) and Accuracy (Acc) Before and After Fine-tuning When Mapping Golf Ball to Tennis Ball (T) (Semantically Close) and Outdoor Recreational Activities (O) (Semantically Distant)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 436, + 553, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 553, + 449 + ], + "type": "text", + "content": "4.5. Impact of Modules affected during Unlearning" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "content": "We chose ESD as the basis for this experiment because it utilizes a setup that allows unlearning by modifying only specific modules (like only cross attention, only non-cross attention and full parameters), enabling targeted analysis of robustness in concept unlearning." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "In our experiments, we used the ESD variations of - ESD-x (cross-attention layers), ESD-u (unconditional layers), and ESD-full (all parameters). We observed that the specific parameters targeted during unlearning significantly impacted the robustness of the concept erasure. Results (Table 2) indicate that cross-attention-based unlearning (ESD-x) is less robust, as the unlearned concept revived more quickly compared to ESD-u and ESD-full. This observation suggests that editing cross-attention weights alone may be insufficient for durable unlearning, as these layers maintain strong associations between prompts and visual features. This vulnerability is also evident in the UCE and CA model, which exclusively modifies cross-attention weights, making it relatively fragile to fine-tuning. Therefore, targeting a broader parameter set, as seen in ESD-u and ESD-full, offers a more resilient approach to concept unlearning." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13399" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 282, + 133 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 282, + 133 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 282, + 133 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 282, + 133 + ], + "type": "table", + "html": "
MethodBefore FTAfter FTRevival Point
CSAccCSAcc
ESD-x0.260.00.320.544
ESD-u0.260.00.330.365
ESD-full0.250.00.320.397
", + "image_path": "5c877a941aac150f88cbe5877a5a4c8ba4f8a63880a56e3d9eaa24768bd19225.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 284, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 284, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 284, + 232 + ], + "type": "text", + "content": "5. Discussion - Plausible Mitigation Strategy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 247, + 297, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 297, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 297, + 464 + ], + "type": "text", + "content": "For unlearning methods that rely on minimizing a dedicated objective (referred to as the unlearning loss), its value acts as a metric to gauge the success of the erasure. To investigate the impact of fine-tuning on this loss, we evaluated three unlearning techniques—UCE [12], ESD [11], and CA [19]—measuring the unlearning loss both immediately after unlearning and after subsequent fine-tuning (as described in section 4). The results, summarized in Table 3, reveal a consistent increase in unlearning loss post fine-tuning, indicating that fine-tuning destabilizes the unlearning process. Notice that in the case of the CA method with \"Angelina Jolie\" as the target concept and \"Woman\" as the surrogate concept, the unlearning loss did not increase after fine-tuning. This suggests no significant change in the extent of unlearning, which is consistent with previous experiments (Fig 5), where fine-tuning resulted in negligible changes to the model's unlearning accuracy for this particular target/surrogate concept pairing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 297, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 297, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 297, + 702 + ], + "type": "text", + "content": "A plausible approach to mitigate the revival could involve meta-learning and gradient alignment techniques. While standard fine-tuning minimizes the diffusion model's loss, its gradients may conflict with the unlearning objective. To resolve this, we propose modifying the fine-tuning process to (1) compute gradients for both the diffusion loss and unlearning objective, (2) project the diffusion gradients orthogonal to the unlearning gradient direction, and (3) update parameters using only the orthogonal component. This ensures that the unlearning loss remains unchanged throughout fine-tuning, preserving the unlearning effect. A similar approach has been employed by Ko et al. [18] and Wu et al. [40] to achieve unlearning while maintaining the generational quality of neutral concepts. Their approach optimizes two objectives—one for preserving the generation quality of neutral concepts and another for unlearning the target concept by ensuring that each update step doesn't increase either objective function." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 320, + 70, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 141, + 295, + 174 + ], + "lines": [ + { + "bbox": [ + 55, + 141, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 295, + 174 + ], + "type": "text", + "content": "Table 2. Comparison of Clip Score (CS) and Accuracy (Acc) Before and After Unlearning for Different ESD Module Configurations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 70, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 553, + 232 + ], + "type": "table", + "html": "
MethodUnlearn LossUnlearn + FTTarget / Surrogate
UCE5.54 × 1016.74 × 102Dog / Cat
1.08 × 1061.65 × 106Van Gogh / Art
6.66 × 1011.03 × 103Pikachu / Cat
9.02 × 1051.23 × 106Cartoon / Art
1.23 × 1029.41 × 102Brad Pitt / Man
ESD5.00 × 10-41.10 × 10-3Golf Ball / NA
3.00 × 10-45.00 × 10-4Van Gogh / NA
4.00 × 10-49.00 × 10-4Nudity / NA
3.00 × 10-41.60 × 10-3Brad Pitt / NA
CA4.79 × 10-47.53 × 10-4Golf Ball / Ball
1.05 × 10-41.49 × 10-4Van Gogh / Art
≈ 0≈ 0Ang. Jolie / Woman
", + "image_path": "48eacd7f9cebdf85317029324b3ed0d9166380b50f49777859a020726a484831.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 240, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 350 + ], + "type": "text", + "content": "Table 3. Comparison of unlearning loss values between: (1) models immediately after unlearning, and (2) the same models after fine-tuning (using weights obtained from Section 4 experiments). Loss values were computed with each method's respective objective function, with corresponding target-surrogate concept pairings shown. The parameter group updated in ESD and CA for these experiments is cross-attention. The last column shows the target concept (Unlearned concept) along with the corresponding mapping concept i.e Surrogate concept. NA means No specific Surrogate Concept." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 370, + 392, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 392, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 392, + 383 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 390, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 558 + ], + "type": "text", + "content": "We present a critical vulnerability in the current Machine Unlearning techniques for text-to-image diffusion models. Our study emphasizes the need for (i) more robust techniques that can erase the semantic space of the target concept in the latent space of the model, in which case finetuning can never revive the target concept, or/and (ii) the method should be able to achieve mapping of target concept to a surrogate concept which is robust to finetuning. By revealing the weaknesses of existing methods against finetuning and providing a framework to measure concept revival, we aim to guide future research toward developing more secure, reliable, and robust unlearning mechanisms. Below, We list some interesting and important directions to further our understanding of the unlearning process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 617 + ], + "type": "text", + "content": "- Rigorous understanding of why the finetuning process revives the unlearned concepts across several unlearning approaches. Though challenging, it is of utmost importance to explain this vulnerability of the existing MU methods and to achieve robust unlearning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "content": "- Relation between concept revival and finetuning concept. We have observed that certain finetuning concepts can cause greater revival while others less. Also, this dependence may vary across unlearned concepts and unlearning methods. Researching the primary cause of the variance in dependency between the finetuning concept and the revival of the unlearning concept can aid in assessing the robustness of the models better." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13400" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Stability AI. Stability ai. https://stability.ai/news/celebrating-one-year-of-stable-diffusion, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "type": "text", + "content": "[2] Justin Betker, Gabriel Goh, Li Jing, Tim Brooks, Jason Wang, Lilian Li, Aditya Ramesh, et al. Improving image generation with better captions. Computer Science, 2(3):8, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 203 + ], + "type": "text", + "content": "[3] Yinzhi Cao and Junfeng Yang. Towards making systems forget with machine unlearning. In 2015 IEEE Symposium on Security and Privacy (SP), pages 463-480. IEEE, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 204, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 257 + ], + "type": "text", + "content": "[4] Mingjie Chen, Wei Gao, Guangyue Liu, Kai Peng, and Cunhua Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7766-7775, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 294, + 280 + ], + "type": "text", + "content": "[5] Lydia de la Torre. A guide to the california consumer privacy act of 2018. Available at SSRN 3275571, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 281, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 294, + 313 + ], + "type": "text", + "content": "[6] NudeNet Developers. Nudenet: A nudity detection and censoring library. https://github.com/notAI-tech/NudeNet, 2023.5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 315, + 295, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 347 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 349, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 349, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 349, + 295, + 423 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 426, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 502 + ], + "type": "text", + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "text", + "content": "[10] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "type": "text", + "content": "[11] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "text", + "content": "[12] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[13] Antonio Ginart, Melody Guan, Gregory Valiant, and James Y Zou. Making ai forget you: Data deletion in machine learning. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[14] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9301-9309. IEEE, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 172 + ], + "type": "text", + "content": "[15] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 218 + ], + "type": "text", + "content": "[16] Chris Jay Hoofnagle, Bart van der Sloot, and Frederik Zuiderveen Borgesius. The european union general data protection regulation: what it is and what it means. Information & Communications Technology Law, 28(1):65-98, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 220, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 554, + 274 + ], + "type": "text", + "content": "[17] Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung-Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. Receiver: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. In European Conference on Computer Vision, pages 360–376. Springer, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 276, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 554, + 330 + ], + "type": "text", + "content": "[18] Myeongseob Ko, Henry Li, Zhun Wang, Jonathan Patsenker, Jiachen T. Wang, Qinbin Li, Ming Jin, Dawn Song, and Ruoxi Jia. Boosting alignment for post-unlearning text-to-image generative models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 331, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 331, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 555, + 386 + ], + "type": "text", + "content": "[19] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 554, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 554, + 441 + ], + "type": "text", + "content": "[20] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 443, + 554, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 443, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 443, + 554, + 509 + ], + "type": "text", + "content": "[21] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts diffusion models and erasing applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7559-7568, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 510, + 552, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 552, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 552, + 531 + ], + "type": "text", + "content": "[22] Midjourney. Midjourney. https://www.midjourney.com/, 2023.1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 533, + 554, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 554, + 598 + ], + "type": "text", + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pooja Shyam, Pavel Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 600, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 600, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 600, + 554, + 643 + ], + "type": "text", + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 645, + 470, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 645, + 470, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 645, + 470, + 656 + ], + "type": "text", + "content": "[25] OpenAI. Chatgpt model api, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 677 + ], + "type": "text", + "content": "[26] OpenAI. Openai embeddings, 2023. Accessed: 2024-11-13. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 680, + 554, + 713 + ], + "type": "text", + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13401" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "text", + "content": "[28] Formerly Data Protection. General data protection regulation (gdpr). Intersoft Consulting, accessed in October, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 185 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "text", + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alexander Quinn Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 295, + 342 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 419 + ], + "type": "text", + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Sahand Mahdavi, Raphael Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "type": "text", + "content": "[34] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 544 + ], + "type": "text", + "content": "[35] Christoph Schuhmann, Robin Beaumont, Romain Vencu, Craig Gordon, Ross Wightman, Mehdi Cherti, Teddy Coombes, Abhay Katta, and Vladimir Kirillov. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278–25294, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 599 + ], + "type": "text", + "content": "[36] Anish Sekhari, Jayadev Acharya, Gauri Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. In Advances in Neural Information Processing Systems, pages 18075-18086, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 623 + ], + "type": "text", + "content": "[37] GenAI statistics. Genai statistics. https://journaleverypixel.com/ai-image-statistics, 2023.1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "text", + "content": "[38] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "text", + "content": "[39] Jing Wu, Trung Le, Munawar Hayat, and Mehrtash Harandi. Erasediff: Erasing data influence in diffusion models, 2024. 1, 2, 3, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 350 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[40] Yongliang Wu, Shiji Zhou, Mingzhuo Yang, Lianzhe Wang, Heng Chang, Wenbo Zhu, Xinting Hu, Xiao Zhou, and Xu Yang. Unlearning concepts in diffusion model via concept domain correction and concept preserving gradient. In Proceedings of the 39th AAAI Conference on Artificial Intelligence (AAAI), 2025. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 195 + ], + "type": "text", + "content": "[41] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1755-1764, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 282 + ], + "type": "text", + "content": "[42] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now. In Computer Vision - ECCV 2024: 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LVII, page 385-403, Berlin, Heidelberg, 2024. Springer-Verlag. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 285, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 350 + ], + "type": "text", + "content": "[43] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. In Thirty-eighth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2, 3, 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13402" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_content_list.json b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db8bb9b5d9b15edbb6608dd5c71c2c0f08524917 --- /dev/null +++ b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation", + "text_level": 1, + "bbox": [ + 217, + 130, + 779, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marcus Nordström1,2 Atsuto Maki1 Henrik Hult1 \n1KTH Royal Institute of Technology 2RaySearch Laboratories {marcno,atsuto,hult}@kth.se", + "bbox": [ + 243, + 202, + 754, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In image segmentation and specifically in medical image segmentation, the soft-Dice loss is often chosen instead of the more traditional cross-entropy loss to improve performance with respect to the Dice metric. Experimental work supporting this claim exists, but how and why the two loss functions lead to different predictions is not well understood. This paper explains the observed discrepancy as a consequence of how those loss functions are effected by label noise and what threshold is used to convert the predicted soft segmentation to predicted labels. In particular, it is shown (i) how the optimal solutions to the two loss functions diverge as the noise is increased, and (ii) how the optimal solutions to soft-Dice can be recovered by thresholding the solutions to cross-entropy with an a priori unknown but efficiently computable threshold. The theoretical results are supported by numerical experiments and it is concluded that cross-entropy with the alternative threshold yields the stability and informative label probability maps associated with cross-entropy without sacrificing the performance of soft-Dice.", + "bbox": [ + 89, + 324, + 485, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 646, + 220, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In medical image analysis it is often of interest to partition an image into foreground and background with respect to some region of interest. This problem is referred to as the medical image segmentation problem. Since the introduction of the U-net [29], state of the art methods for addressing the problem have been dominated by neural network methods trained according to the supervised learning paradigm [7, 32]. These methods are remarkably effective in scenarios with abundant and clean data. However, manually delineating a region of interest accurately and reliably can be a complex and time-consuming process that often requires medical expertise. This result in high costs associated with obtaining high-quality data sets for training, and consequently, that data used in practice often is imperfect in the sense that the manually provided labels are contaminated by noise [1, 5, 23, 27].", + "bbox": [ + 89, + 672, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A large amount of research has been devoted to medical image segmentation with imperfect data. One important question that has been addressed is how experimentally robust various segmentation methods are to label noise. In such work, experiments are conducted by first obtaining some data set and then corrupting the portion of data used for training with various types of noise. Different models are then trained on the resulting data sets and evaluated on uncorrupted data. If both noisy labels and noise-free labels can be obtained from clinical practitioners, it is possible to corrupt a clean label simply by swapping it with the corresponding noisy label. Such a study of label noise was presented in [38]. An alternative is to start with a data set that is assumed to be noise-free and generate noisy data with a noise model. Simple noise models that have been studied include randomly flipping labels in an independent manner or randomly translating the target structure [10]. Models that yield more realistic looking segmentation noise include randomly perturbing some points of the boundary on a noise-free segmentation with Gaussian variables [10] and randomly deforming a segmentation with an interpolated low-dimensional Gaussian array [35].", + "bbox": [ + 511, + 292, + 908, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How label noise affects the performance of segmentation models is highly dependent on the loss function that is used during training. Arguably, the most classical loss function considered in the field is the cross-entropy loss. This loss is known to have a lot of nice theoretical properties and to be stable in practice. However, because of the class imbalance associated with the foreground typically being significantly smaller than the background, other loss functions putting more emphasis on foreground are often considered. The most popular such loss is a smoothed version of the Dice metric. This loss is referred to as soft-Dice or the Dice loss, and has been shown to yield better performance than cross-entropy in several experimental settings at the cost of sometimes introducing some stability issues [2, 24]. To get some of the improved performance while limiting potential stability issues, it is common to use linear combinations of the losses. This is for instance the type of loss used in the popular nnU-net [13].", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20820", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contribution. This work introduces a noise model based on Gaussian field deformations. The model is shown to be approximately volume preserving in the sense that the expected volume of the resulting noisy segmentation is approximately equal to the volume of the corresponding noise-free segmentation. Moreover, a closed form expression for the marginal probabilities and a representation of random deformations using stochastic integrals are provided. This allows the behaviour of the model to be clearly interpreted and efficiently simulated. With the noise model, optimal solutions to cross-entropy and soft-Dice are studied. It is shown how the optimal solutions to soft-Dice and cross-entropy coincide when the labels are noise-free but diverge when noise is increased. It is further shown how optimal solutions to soft-Dice can be recovered by thresholding solutions to cross-entropy with a particular a priori unknown but efficiently computable threshold. Finally, the presented theory is validated experimentally on three different organ segmentation problems from the publicly available TotalSegmentor data set [37] with noise introduced by the model.", + "bbox": [ + 89, + 90, + 485, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 407, + 232, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Label noise is a well known issue in medical image segmentation and the impact of observer variability has been discussed in various domains for different modalities. This include for radiotherapy [5], for lung nodules on clinical thoracic CT images [1] and for organs in the pelvic region for CT and MR [27]. Examples of work on methods targeting these issues include active learning methods, where experts are iteratively probed for more information [9] and re-weighting schemes where methods dynamically learn to put different weights on different regions depending on certainty [22]. For fairly recent reviews on image segmentation with imperfect data in general, see [14, 34].", + "bbox": [ + 89, + 431, + 485, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The robustness of medical image segmentation to noise can be studied by training models on data contaminated with varying degrees of noise. In [38], both noise-free and noisy labels are collected and the noise level is determined by the fraction of the training data taken from the noisy data set. Of special importance to this work is segmentation noise provided by Gaussian field deformations. Two works have studied such noise in an experimental setting [10, 35]. In these works, the segmentations in the training data is perturbed according to a low dimensional Gaussian field interpolated to the pixel space. Models are then trained for various noise levels and results obtained are reported. Finally, in [16], a noise model for target delineation in radiotherapy is discussed. The model is based on taking the level set of Gaussian field and efficient methods for sampling is provided.", + "bbox": [ + 89, + 614, + 485, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The influence label noise has on various loss functions has been studied in several papers. This include the influence it has on calibration and volume bias for predictions generated", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with models trained using cross-entropy and soft-Dice [3, 20], re-calibration methods for improving performance of soft-Dice [30] and the relationship between calibration and volume bias [28]. Other experimental work showing that the expected calibration error is lower for cross-entropy than for soft Dice includes [21]. In the above, noise has been analyzed by considering the expected score with respect to some noisy segmentation. An alternative way to incorporate noise is to make use of soft labels. Examples of work using soft labels with soft-Dice include [8, 15]. Examples of work on soft labels with other loss functions include [17, 18, 33]. In [36] a modified version of soft-Dice is introduced. This loss yields optimal solutions that coincide with the soft labels provided for the target structure which is known to not be the case for the classical soft-Dice loss.", + "bbox": [ + 511, + 90, + 906, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by threshold classifiers studied in [19, 39], optimal segmentations to Accuracy and Dice under label noise was analyzed in [25]. A characterization of optimal segmentations was provided, sharp bounds with respect to volume bias of the optimal segmentations where shown and the volume of the optimizers associated with the metrics was compared. A similar analysis was later provided for the loss function soft-Dice [26]. In particular, a characterization of the optimal solutions and the associated volumes are provided. Moreover, a calibration result showing that a sequence that converge to optimal soft-Dice, when thresholded appropriately, converges to optimal Dice. Experimental work comparing various thresholds for cross-entropy was presented in [4, 12]. The results from these papers experimentally show that using 0.5 as a common threshold leads to sub-optimal segmentations for various loss functions with respect to Dice.", + "bbox": [ + 511, + 318, + 908, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 593, + 651, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Notation", + "text_level": 1, + "bbox": [ + 511, + 619, + 617, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The notation largely follows [25, 26]. Let $\\Omega = [0,1]^n\\subset \\mathbb{R}^n$ be the unit cube of dimension $n\\geq 1$ and $\\lambda$ be the standard Lebesgue measure associated with $\\mathbb{R}^n$ such that $\\lambda (\\Omega) = 1$ . Let $S$ be the space of measurable functions from $\\Omega$ to $\\{0,1\\}$ , $\\mathcal{M}$ be the space of measurable functions from $\\Omega$ to $[0,1]$ and $\\mathcal{F}$ be the space of bounded measurable functions from $\\Omega$ to $\\mathbb{R}$ , where all of the spaces $S$ , $\\mathcal{M}$ and $\\mathcal{F}$ are equipped with their associated Borel $\\sigma$ -fields and the $L_{p}$ -norms $\\| \\cdot \\| _p$ . The letter $s\\in S$ is generally used to denote a segmentation, where $s(\\omega) = 1$ indicates foreground of the target structure of interest and $s(\\omega) = 0$ indicates background. The letter $m\\in \\mathcal{M}$ is generally used to denote a marginal function, that is, $m(\\omega)\\in [0,1]$ indicates the probability of a noisy label occupying site $\\omega \\in \\Omega$ . The letter $c\\in \\mathcal{M}$ is generally used to denote the soft segmentation that is obtained by optimizing the chosen loss function prior to any thresholding. The letter $f\\in \\mathcal{F}$ is generally used to denote a logit function usually", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20821", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "associated with a soft segmentation such that $\\sigma \\circ f(\\omega) = c(\\omega)$ for $\\omega \\in \\Omega$ a.e., where $\\sigma(x) \\doteq 1/(1 + e^{-x})$ is the standard sigmoid function. Composition of functions are denoted by $\\circ$ and $I_A(x)$ is used to denote the indicator function over the set $A$ , that is $I_A(x) = 1$ if $x \\in A$ and $I_A(x) = 0$ if $x \\notin A$ . The indicator function is used in conjunction with some element $c \\in \\mathcal{M}$ to denote thresholding of a soft segmentation. That is, $I_{[t,1]} \\circ c$ assigns 1 to all $\\omega \\in \\Omega$ where $c(\\omega) \\geq t$ and 0 to all $\\omega \\in \\Omega$ where $c(\\omega) < t$ .", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Loss Functions and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 236, + 428, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Two of the most popular loss functions in the field of medical image segmentation are cross-entropy and soft-Dice. Usually these loss functions are defined over two deterministic arguments, a soft-segmentation (prediction) and a binary segmentation (label). Under the influence of label noise, these loss functions are extended to functions taking one deterministic soft segmentation (prediction) and one random segmentation (label) as arguments. In this work, the soft labeling extension is adopted for the theoretical analysis [8, 15, 17, 18, 33] (see Section 4.4 for more on this). Formally, the loss functions are defined with respect to a marginal function $m \\in \\mathcal{M}$ which corresponds to the mean of a noisy segmentation $L$ taking values in $S$ , that is $m(\\omega) = \\mathbb{E}[L(\\omega)], \\omega \\in \\Omega$ .", + "bbox": [ + 89, + 258, + 483, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1. For any $m \\in \\mathcal{M}$ , (binary) cross-entropy is given by", + "bbox": [ + 89, + 465, + 483, + 496 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {C E} _ {m} (c) \\doteq - \\int_ {\\Omega} [ m (\\omega) \\log (c (\\omega)) + \\quad c \\in \\mathcal {M}. \\tag {1} \\\\ (1 - m (\\omega)) \\log (1 - c (\\omega)) ] \\lambda (d \\omega), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 506, + 483, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2. For any $m \\in \\mathcal{M}$ , soft-Dice is given by", + "bbox": [ + 89, + 569, + 439, + 584 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S D} _ {m} (c) \\doteq 1 - \\frac {2 \\int_ {\\Omega} c (\\omega) m (\\omega) \\lambda (d \\omega)}{\\| c \\| _ {1} + \\| m \\| _ {1}}, \\quad c \\in \\mathcal {M}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 594, + 483, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Arguably, the most classical evaluation metric used is Accuracy. This metric simply measures the fraction of the domain that is correctly labeled. The most popular evaluation metric on the other hand is the Sorensen-Dice coefficient, or Dice for short. To handle label noise, these metrics are defined in the same way as the loss functions, that is, by using soft labels.", + "bbox": [ + 89, + 638, + 483, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3. For any $m \\in \\mathcal{M}$ , Accuracy is given by", + "bbox": [ + 89, + 755, + 441, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {A} _ {m} (s) \\doteq \\int_ {\\Omega} [ s (\\omega) m (\\omega) + \\quad s \\in \\mathcal {S}. \\tag {3} \\\\ (1 - s (\\omega)) (1 - m (\\omega)) ] \\lambda (d \\omega), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 781, + 483, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 4. For any $m \\in \\mathcal{M}$ , Dice is given by", + "bbox": [ + 89, + 843, + 411, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} _ {m} (s) \\doteq \\frac {2 \\int_ {\\Omega} s (\\omega) m (\\omega) \\lambda (d \\omega)}{\\| s \\| _ {1} + \\| m \\| _ {1}}, \\quad s \\in \\mathcal {S}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 869, + 483, + 905 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When a label is free from noise, any reasonable loss function or metric is optimized uniquely by the target label. When there is noise present on the other hand, it is no longer clear what the solution should be and loss functions or metrics implicitly assign them. As the consequence, it is important to investigate what the optimal solutions are for a chosen loss or metric. For the case of the cross-entropy loss, it is well known that the optimizers are uniquely given by the marginal function, that is", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C E} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {C E} _ {m} \\left(c ^ {\\prime}\\right) \\Longleftrightarrow c (\\omega) = m (\\omega), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 234, + 906, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\lambda$ -a.e. In other words, minimizing cross-entropy yields estimates of the marginal probability distribution associated with the labels. For soft-Dice it was shown in [26] that", + "bbox": [ + 511, + 263, + 906, + 310 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S D} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {S D} \\left(c ^ {\\prime}\\right) \\Longleftrightarrow\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 315, + 720, + 338 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc (\\omega) \\in \\left\\{ \\begin{array}{l l} \\{1 \\} & \\text {i f} m (\\omega) > \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2, \\\\ [ 0, 1 ] & \\text {i f} m (\\omega) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2, \\\\ \\{0 \\} & \\text {i f} m (\\omega) < \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2, \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 342, + 906, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\lambda$ -a.e. Characterizations of the optimal segmentations for Accuracy and Dice using the setup adopted in this paper can be found in [25]. For Accuracy, the solutions are given by", + "bbox": [ + 511, + 405, + 906, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} _ {m} (s) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {A} _ {m} \\left(s ^ {\\prime}\\right) \\iff\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 469, + 774, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns (\\omega) \\in \\left\\{ \\begin{array}{l l} \\{1 \\} & \\text {i f} m (\\omega) > 1 / 2, \\\\ \\{0, 1 \\} & \\text {i f} m (\\omega) = 1 / 2, \\\\ \\{0 \\} & \\text {i f} m (\\omega) < 1 / 2, \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 498, + 906, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\lambda$ -a.e. Similarly, for Dice the optimal segmentations are given by", + "bbox": [ + 511, + 561, + 905, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} _ {m} (s) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) \\Longleftrightarrow \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 598, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns (\\omega) \\in \\left\\{ \\begin{array}{l l} \\{1 \\} & \\text {i f} m (\\omega) > \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} (s ^ {\\prime}) / 2, \\\\ \\{0, 1 \\} & \\text {i f} m (\\omega) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} (s ^ {\\prime}) / 2, \\\\ \\{0 \\} & \\text {i f} m (\\omega) < \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} (s ^ {\\prime}) / 2, \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 628, + 905, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\lambda$ -a.e. The resulting soft segmentation obtained by optimizing a loss function is typically converted to a binary segmentation by thresholding with 0.5. Ideally this resulting segmentation maximizes a desired target metric. Using the above characterizations, for cross-entropy it follows that", + "bbox": [ + 511, + 690, + 906, + 766 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C E} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {C E} \\left(c ^ {\\prime}\\right) \\Longrightarrow\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 772, + 807, + 795 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} _ {m} \\left(I _ {[ 1 / 2, 1 ]} \\circ c\\right) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {A} \\left(s ^ {\\prime}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 792, + 906, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, for soft-Dice it follows that", + "bbox": [ + 513, + 830, + 764, + 845 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S D} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {S D} \\left(c ^ {\\prime}\\right) \\Longrightarrow\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 852, + 807, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} _ {m} \\left(I _ {[ 1 / 2, 1 ]} \\circ c\\right) = \\sup _ {s ^ {\\prime} \\in S} \\mathrm {D} \\left(s ^ {\\prime}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 869, + 906, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20822", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Gaussian Fields and Independently Scattered Gaussian Measures", + "text_level": 1, + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Amongst the most central objects in probability theory and statistics are Gaussian variables taking values in Euclidean space. Such a variable, when centered and isotropic, has the probability density function", + "bbox": [ + 89, + 128, + 483, + 189 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\sigma^ {2}} (\\omega) \\dot {=} \\frac {1}{(2 \\pi \\sigma^ {2}) ^ {n / 2}} \\exp \\left\\{- \\frac {\\| \\omega \\| _ {2} ^ {2}}{2 \\sigma^ {2}} \\right\\}, \\quad \\omega \\in \\mathbb {R} ^ {n}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 200, + 483, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma^2$ denotes the variance in each of the dimensions. Various generalizations of this random variable have been studied extensively in the literature. For the theoretical treatment of the noise model in this work, two such generalizations are introduced. The first is a Gaussian field.", + "bbox": [ + 89, + 247, + 485, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 5. A random variable $Y(\\omega), \\omega \\in \\mathbb{R}^n$ is a Gaussian field with kernel function $k$ if any finite collection of elements $\\{\\omega_1, \\dots, \\omega_M\\} \\subset \\mathbb{R}^n$ satisfies $(Y(\\omega_1), \\dots, Y(\\omega_M)) \\sim \\text{normal}(0, \\Sigma)$ , with $\\Sigma_{m,m'} = k(\\omega_m, \\omega_{m'})$ for $m, m' \\in \\{1, \\dots, M\\}$ .", + "bbox": [ + 89, + 335, + 483, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Many types of kernel functions with different properties have been studied in the literature. Amongst the most common is the squared exponential kernel given by", + "bbox": [ + 89, + 422, + 483, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk _ {a, b} \\left(\\omega , \\omega^ {\\prime}\\right) \\doteq a ^ {2} \\exp \\left(- \\frac {\\| \\omega - \\omega^ {\\prime} \\| _ {2} ^ {2}}{2 b ^ {2}}\\right), \\quad \\omega , \\omega^ {\\prime} \\in \\mathbb {R} ^ {n}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 479, + 483, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An important property of Gaussian fields with such covariance functions are that samples are almost surely continuous. In other words, a randomly sampled function will be continuous with probability one. For a good reference on Gaussian variables, see [11].", + "bbox": [ + 89, + 541, + 483, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second generalization that is introduced is what is referred to as independently scattered Gaussian measures. Such constructions are used to formalize integration over Gaussian fields, similar to how the Ito-integral build on Brownian motion.", + "bbox": [ + 89, + 618, + 483, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 6. A random variable $W$ is an independently scattered Gaussian measure on $\\mathbb{R}^n$ with respect to the control measure $\\lambda$ , if for any collection of disjoint sets $A_1, \\ldots, A_k$ in $\\lambda$ -measurable sets with finite measure $\\mathcal{B}_0$ , $W(A_1), \\ldots, W(A_k)$ are independent, and $W(A_i)$ has centered Gaussian distribution with variance $\\lambda(A_i)$ .", + "bbox": [ + 89, + 705, + 483, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Importantly, for a function $f \\in L^{2}(\\mathbb{R}^{n})$ , the stochastic integral $\\int_{\\mathbb{R}^n} f(\\omega) W(d\\omega)$ is well defined and the process $Y = \\{Y(\\omega), \\omega \\in \\mathbb{R}^n\\}$ given by", + "bbox": [ + 89, + 806, + 483, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY (\\omega) \\doteq \\int_ {\\mathbb {R} ^ {n}} f \\left(\\omega - \\omega^ {\\prime}\\right) W \\left(d \\omega^ {\\prime}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 864, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is a centered Gaussian process. Furthermore, this Gaussian process has a covariance kernel given by", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk \\left(\\omega , \\omega^ {\\prime}\\right) = \\int_ {\\mathbb {R} ^ {n}} f (\\omega - u) f \\left(\\omega^ {\\prime} - u\\right) \\lambda (d u), \\quad \\omega , \\omega^ {\\prime} \\in \\mathbb {R} ^ {n}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 132, + 906, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Independently scattered measures are introduced more generally for $\\alpha$ -stable distributions in [31, Section 3.3]. The Gaussian case used in this work is simply the special case when $\\alpha = 2$ .", + "bbox": [ + 511, + 191, + 906, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Theory", + "text_level": 1, + "bbox": [ + 511, + 265, + 599, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Label Noise Model", + "text_level": 1, + "bbox": [ + 511, + 290, + 694, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to binary classification, where a lot of label noise can be modeled by simple independent flipping of binary variables, label noise in segmentation can in general have very complicated dependency structures. Often labels affected by noise have the appearance of being drawn by someone with shaky hands. For such noise, the correlation between border points will depend on the distance between them. That is, if two points are close to each other, then the correlation will be high and vice versa.", + "bbox": [ + 511, + 311, + 906, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One way to model segmentations that look as if they have been drawn by shaky hands is by means of random Gaussian field deformations. Consider a variable", + "bbox": [ + 511, + 449, + 905, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX \\doteq \\left(X _ {1}, \\dots , X _ {n}\\right) ^ {T} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 503, + 906, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "such that $\\{X_i\\}_{i=1}^n$ are independent Gaussian fields with kernel function (13). This means that $X$ is a random vector field taking values in $\\mathbb{R}^n \\mapsto \\mathbb{R}^n$ . With such a random vector field, a random segmentation can be formed by", + "bbox": [ + 511, + 534, + 906, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL (\\omega) \\doteq \\left\\{ \\begin{array}{l l} l (\\omega + X (\\omega)) & \\text {i f} \\omega + X (\\omega) \\in \\Omega , \\\\ 0 & \\text {i f} \\omega + X (\\omega) \\notin \\Omega . \\end{array} \\right., \\quad \\omega \\in \\Omega , \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 606, + 906, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l \\in S$ is a noise-free reference segmentation being randomly deformed. How samples of the random segmentations look depend on what parameters are chosen when specifying the kernel function (13). The parameter $a$ controls how large the deformations are, where greater $a$ leads to larger deformations and vice versa. The parameter $b$ controls how smooth the deformations are, where greater $b$ leads to smoother deformations and vice versa. Modeling noisy segmentations with (17) is convenient because it allows for the derivation of several theoretical properties. In this work, three such important properties are derived.", + "bbox": [ + 511, + 672, + 908, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Proposition 1, it is shown that the marginal probabilities associated with a noisy segmentation can be computed by convolving the associated noise-free segmentation label with an appropriate Gaussian density.", + "bbox": [ + 511, + 840, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20823", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1. Let $l$ and $L$ be defined as in (17) and $p_{\\sigma^2}$ be defined as in (12), then", + "bbox": [ + 89, + 90, + 485, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ L (\\omega) ] = \\int_ {\\Omega} l \\left(\\omega^ {\\prime}\\right) p _ {a ^ {2}} \\left(\\omega - \\omega^ {\\prime}\\right) \\lambda \\left(d \\omega^ {\\prime}\\right), \\quad \\omega \\in \\Omega . \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 130, + 485, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Proposition 2, the marginal representation is used to show that the volume of the noisy segmentation is approximately unbiased in the sense that the expected volume of a noisy segmentation is approximately equal to the volume of the corresponding noise-free segmentation. The reason why this does not hold exactly is that there is some probability for the target structure to deform in such a way that parts of the structure leave the domain $\\Omega$ . However, for most reasonable choices of parameters, when the regions of interest is sufficiently close to the center, the probability of this happening is negligible.", + "bbox": [ + 89, + 188, + 483, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2. Let $l$ and $L$ be defined as in (17) and $p_{\\sigma^2}$ be defined as in 12, then", + "bbox": [ + 89, + 364, + 483, + 395 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ \\| L \\| _ {1} ] = \\| l \\| _ {1} - \\xi , \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 404, + 483, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 89, + 431, + 137, + 445 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\xi = \\int_ {\\Omega} \\int_ {\\mathbb {R} ^ {n} \\backslash \\Omega} l \\left(\\omega^ {\\prime}\\right) p _ {a ^ {2}} \\left(\\omega - \\omega^ {\\prime}\\right) \\lambda (d \\omega) \\lambda \\left(d \\omega^ {\\prime}\\right). \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 453, + 483, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Proposition 3, it is shown that $X$ (16) is equal in distribution to a particular vector of stochastic integrals. In the case $n = 1$ , this could have been done using regular Ito-integrals, but for the case $n > 1$ a multivariate counter part is needed. For this scattered Gaussian measures are used.", + "bbox": [ + 89, + 512, + 485, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 3. Let $X$ be as in (16), $p_{\\alpha^2}$ be as in (12) and $W_1, \\ldots, W_n$ be independent copies of independently scattered Gaussian measures on $\\mathbb{R}^n$ with control measure $\\lambda$ , then it follows that", + "bbox": [ + 89, + 595, + 485, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX (\\omega) \\stackrel {d} {=} \\left( \\begin{array}{c} a (2 \\pi b ^ {2}) ^ {n / 4} \\int_ {\\mathbb {R} ^ {n}} p _ {b ^ {2} / 2} \\left(\\omega - \\omega^ {\\prime}\\right) W _ {1} \\left(d w ^ {\\prime}\\right) \\\\ \\vdots \\\\ a (2 \\pi b ^ {2}) ^ {n / 4} \\int_ {\\mathbb {R} ^ {n}} p _ {b ^ {2} / 2} \\left(\\omega - \\omega^ {\\prime}\\right) W _ {n} \\left(d w ^ {\\prime}\\right) \\end{array} \\right). \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 666, + 483, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The presented propositions provide meaningful insights into the behaviour of the noise model. The importance of Proposition 1 is that it makes the relationship between the noise model parameters and the resulting marginal probabilities clear. Also, it provides an exact method for computing the marginals that otherwise would have to be estimated with sampling. The importance of Proposition 2 is that the noise beside from looking reasonable also numerically behave reasonable in an important manner. Finally, the importance", + "bbox": [ + 89, + 763, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of Proposition 3 is that noise for discretized domains can be sampled by smoothing arrays of independently sampled Gaussian variables with an appropriate Gaussian filters. Both operations which scale well with image size. This allows for efficient sampling of noisy segmentations for images composed of millions of voxels.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Optimal Segmentations Under Label Noise", + "text_level": 1, + "bbox": [ + 511, + 190, + 877, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the loss functions and metrics investigated in this work only depend on the marginal probabilities and Proposition 1 provides an explicit expression for these marginal probabilities, it is possible to study how noise model parameters affect optimal segmentations. Figure 1 depicts this relationship for a two dimensional segmentation problem with varying noise intensity. In the first row, pixel-wise averages of 5 samples are illustrated. In the second row, the marginal probabilities are illustrated. In the third row, theoretically optimal solutions to Accuracy are shown. Note that by (10), this is equivalently given by the minimizer to cross-entropy thresholded by 0.5. In the forth row, theoretical optimal solutions to Dice are shown. Note that by (11), this is equivalently given by the minimizer to soft-Dice thresholded by 0.5.", + "bbox": [ + 511, + 212, + 906, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A first observation is that noise affects the optimal segmentations in such a way that the shape of the noise-free reference segmentation is changed. This is an important observation that contradicts the picture that unbiased noise (unbiased in the sense that the expected deformation is the identity function) yield unbiased optimal segmentations. In particular, optimal segmentations are changed in such a way that the corners of the shapes are smoothed and details such as thin structures and holes disappear. Informally, the higher the noise, the more ball shaped the optimal segmentations.", + "bbox": [ + 511, + 424, + 908, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A second observation is that the two metrics yield similar looking segmentations for low noise but increasingly different segmentations as the noise increase. Since cross-entropy with the 0.5-threshold theoretically gives optimal segmentations with respect to Accuracy and soft-Dice with the 0.5-threshold theoretically gives optimal segmentations with respect to Dice, the performance of the two loss functions as measured by Dice is similar in low noise situations but increasingly better for soft-Dice as the noise is increased. Similarly, the performance of the two loss functions as measured by Accuracy is similar in low noise situations but increasingly better for cross-entropy as the noise is increased.", + "bbox": [ + 511, + 575, + 910, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Cross-entropy With an Alternative Threshold", + "text_level": 1, + "bbox": [ + 511, + 763, + 898, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By the characterization of optimizers to cross-entropy (5) and the characterization of optimizers to soft-Dice (6), it follows that an optimizer to soft-Dice can be obtained by first optimizing cross-entropy and then thresholding with the a priori unknown value", + "bbox": [ + 511, + 786, + 906, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt _ {m} \\doteq \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2. \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 872, + 906, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20824", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/71238ae1e0ffc3ac440274252f170f0eb5622b6369704d29ba4a7f47f4e34172.jpg", + "image_caption": [ + "Figure 1. Illustration of the noise model and its effects on a two-dimensional segmentation problem, with the columns showing the effect of different noise strengths. The first row shows the pixel-wise average obtained from five samples. The second row shows the marginal function. The third row shows optimal segmentations associated with Accuracy (7), which in this and most natural cases is unique. The forth row shows optimal segmentations associated with Dice (9), which in this and most natural cases is unique. In all of the reported cases $b = 0.15 / \\sqrt{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "That is, by thresholding an optimizer to cross-entropy with this alternative recursively defined threshold $t_m$ , it follows that", + "bbox": [ + 88, + 568, + 486, + 612 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {C E} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {C E} \\left(c ^ {\\prime}\\right) \\Longrightarrow \\tag {23} \\\\ \\mathrm {S D} _ {m} \\left(I _ {[ t _ {c}, 1 ]} \\circ c\\right) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {S D} \\left(c ^ {\\prime}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 633, + 483, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Methods based on this principle have not been discussed in the medical image segmentation community but are known in the statistical learning literature [19]. In the medical image segmentation community, however, there has been experimental work comparing the performance of different constant thresholds, that is, thresholds not dependent on the marginal probabilities [4, 12]. The results experimentally show that using a 0.5-threshold leads to sub-optimal segmentations for various loss functions when the Dice metric is used for evaluation. Consequently, the results are in line with the theoretical observations above and motivate further investigation of methods based on cross-entropy with the threshold (22).", + "bbox": [ + 89, + 704, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Soft Labeling and Expected Score", + "text_level": 1, + "bbox": [ + 511, + 566, + 808, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One reason to why the threshold (22) has not been discussed in the medical image segmentation domain may be that loss functions and metrics often incorporate noise using expectations rather than soft labels. Formally for some random label $L$ taking values in $S$ with $\\mathbb{E}[L(\\omega)] = m(\\omega), \\omega \\in \\Omega$ and $c \\in \\mathcal{M}$ , this means that $\\mathbb{E}[\\mathrm{CE}_L(c)]$ is considered instead of $\\mathrm{CE}_m(c)$ and $\\mathbb{E}[\\mathrm{SD}_L(c)]$ is considered instead of $\\mathrm{SD}_m(c)$ . Similarly for some $s \\in S$ , this means that $\\mathbb{E}[\\mathrm{A}_L(s)]$ is considered instead of $\\mathrm{A}_m(s)$ and $\\mathbb{E}[\\mathrm{D}_L(s)]$ is considered instead of $\\mathrm{D}_m(s)$ .", + "bbox": [ + 511, + 594, + 908, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For cross-entropy and Accuracy, it is easy to verify that the definitions using expectations and the definitions using soft labels are equivalent, that is $\\mathbb{E}[\\mathrm{A}_L(s)] = \\mathrm{A}_m(s), s \\in S$ and $\\mathbb{E}[\\mathrm{CE}_L(c)] = \\mathrm{CE}_m(c), c \\in \\mathcal{M}$ . For soft-Dice and Dice, however, this is in general not the case, that is, in general $\\mathbb{E}[\\mathrm{D}_L(s)] \\neq \\mathrm{D}_m(s), s \\in S$ and $\\mathbb{E}[\\mathrm{SD}_L(c)] \\neq \\mathrm{SD}_m(c), c \\in \\mathcal{M}$ . Hence, the property (23) does not in general transfer to properties of optimizers to soft-Dice using expectations. Importantly however, that the definitions of soft-Dice and Dice using soft labels and expectations are", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20825", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "equivalent provided that the volume of the noisy label is constant, that is $\\mathrm{Var}[\\| L\\| _1] = 0$ . For situations when the noise is approximately constant $\\mathrm{Var}[\\| L\\| _1]\\approx 0$ , the definitions are approximately equal $\\mathbb{E}[\\mathrm{D}_L(s)]\\approx \\mathrm{D}_m(s),s\\in S$ and $\\mathbb{E}[\\mathrm{SD}_L(c)]\\approx \\mathrm{SD}_m(c),c\\in \\mathcal{M}$ [25]. Hence the property (23) holds approximately for the kind of noise studied in this work.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 210, + 222, + 228 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Model", + "text_level": 1, + "bbox": [ + 89, + 237, + 176, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A standard 3D U-net is implemented [6]. The network is composed of blocks of two convolutions with kernel size $3^3$ , instance normalization and ReLu-activations. The number of filters used is increased for each down-sampling according to 16, 32, 64, 128, 256, where 16 corresponds to the number of filters used in the input layer and 256 corresponds to the number of filters in the bottleneck layer. Convolutions with kernel $2^3$ and stride 2 are used for down-sampling. Transposed convolutions with kernel $2^3$ stride 2 are used for up-sampling. Details and code can be found in the Supplementary Material.", + "bbox": [ + 89, + 258, + 483, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Training", + "text_level": 1, + "bbox": [ + 89, + 436, + 194, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Models are trained either with the cross-entropy loss or with the soft-Dice loss. In both cases $l_{2}$ regularization is used with a weight of $10^{-5}$ . Adam with learning rate $10^{-4}$ is used for optimization together with a scheduler that divides the learning rate by 5 after 30 epochs of non-improvement. The number of epochs computed is fixed to 100 and the order of the training data is randomly shuffled. Batch sizes of 1 is used and with probability 0.5 the data is augmented. The augmentation is composed of random shifts, random rotations and image blurring with Gaussian filters. The model that performs best on the training data is the model selected. Details and code can be found in the Supplementary Material.", + "bbox": [ + 89, + 459, + 483, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.Data", + "text_level": 1, + "bbox": [ + 89, + 666, + 163, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experiments are conducted with respect to the TotalSegmentor data set [37]. This data set contains 1204 CT images with 104 anatomical structures (27 organs, 59 bones, 10 muscles, 8 vessels). To illustrate the effect noise may have on organs with different shape, three different organs are chosen. This includes the right kidney which in general is spherical, the aorta which is tubular and relatively thick and the esophagus which is tubular and relatively thin. For each of the organs, 400 cases is selected and split into 5 folds each containing 80 cases. The images are sub-sampled to half resolution and patches of $64^{3}$ voxels centered in each of the structures are extracted. In order to simulate label noise the described noise model is used. For each organ data set, four different datasets was created by randomly deforming the", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5b7a1bdd2328e6729d7b379c04e17ac03fdc708fa5d17de501e2601dc41dc16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 89, + 645, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef0aaf31ee8630f99f8edeaba658505db80f7d2add6eed19b95b447623758590.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 89, + 772, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/852e87ac540c4ef801d6fd2481a190a899c31187b5f54ee31a3192694809cbde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 89, + 901, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/903fb259c84959ff8d6d857f4d978944eb21a46a914551d7fea1551c955ec9bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 189, + 643, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0694241640b8c721f29c283c37aa7edbcac75c442a57030825a85477e4837dca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 189, + 772, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dbd44d53b5d96d07b63c9de096353bb8411d33ff2f6074f888b04e2474b11b2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 189, + 901, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57a52df7f11289c6fa0d52824939709e2e64f24ced00ff73d6c99a825ddd649d.jpg", + "image_caption": [ + "Figure 2. Illustration of the data from the TotalSegmentor data set [37] used in the experiments. The first depicts a kidney, the second row an aorta and the third row an esophagus. The columns correspond to axial, coronal and sagittal views respectively." + ], + "image_footnote": [], + "bbox": [ + 517, + 287, + 643, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3dff81445302c5f8019c4a8b811b8de9ec39bd0394b4cc9f4b1a3687f2e21f2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 287, + 772, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79a24752fe1a5b48cd01005710aebd9cec16799502a517bc3fc163b0a82ae353.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 287, + 901, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "labels with the noise strengths $a = 0.0, 0.1, 0.2, 0.3$ . Each data point is deformed once, resulting in one sample for each random label. Details and code can be found in the Supplementary Material.", + "bbox": [ + 511, + 483, + 906, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Results", + "text_level": 1, + "bbox": [ + 513, + 556, + 607, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results from the experiments are illustrated separately for the three organ segmentation problems as a figure in Figure 3 and listed as a table in the Supplementary Material. For each organ segmentation problem, the average Dice score for the various methods is reported as a function of the parameter controlling the noise strength in the noise model. The average score is obtained by taking the average of the results obtained after 5 folds. The notation $\\mathrm{CE}^{(0)}$ is used to denote a model trained with cross-entropy and thresholded with 0.5, $\\mathrm{SD}^{(0)}$ a model trained with cross-entropy and thresholded with 0.5 and $\\mathrm{CE}^{(*)}$ a model trained with cross-entropy and thresholded with $t_m$ (22).", + "bbox": [ + 511, + 580, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The first thing to notice is that the difference amongst the average performance of cross-entropy and soft-Dice with the 0.5-threshold, that is, $\\mathrm{CE}^{(0)}$ and $\\mathrm{SD}^{(0)}$ , increase as the amount of added noise increase. When no noise is added, the performance is indistinguishable. This type of result is expected from the theory and depicted in Figure 1.", + "bbox": [ + 511, + 763, + 906, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The second thing to notice is that cross-entropy with the alternative threshold (22) $\\mathrm{CE}^{(*)}$ performs better than cross-entropy with the 0.5-threshold $\\mathrm{CE}^{(0)}$ . In regions with low", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20826", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c1789cbdfe4ad40ce452d8b9e3c6758772f75a6ded02f559eb5a0854d441295.jpg", + "image_caption": [ + "Figure 3. To the left are the results from the kidney segmentation problem, in the middle are the results from the aorta segmentation problem and to the right are the results from the esophagus segmentation problem. In each of the figures, the average Dice value obtained from five fold experiments is plotted as a function of the noise level $a$ . The legend $\\mathrm{CE}^{(0)}$ indicate that cross-entropy with the $1/2$ -threshold has been used, $\\mathrm{SD}^{(0)}$ indicate that soft-Dice with the $1/2$ -threshold has been used and $\\mathrm{CE}^{(*)}$ indicate that cross-entropy with the threshold described in (22) has been used. In all of the reported cases $b = 0.15/\\sqrt{2}$ . A table of the results can be found in the Supplementary Material." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 359, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4323af32b5afb573ab02198424bdd0eafc7553ea8a8a1e152eadb2d808ab8471.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 89, + 612, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a629b5e6beb130a8bf7929068fc02c3a867ae3eaee252c727d5a7860bfa4951d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 90, + 880, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "noise the discrepancy is not noticeable, but as the noise is increased the discrepancy also increases. The performance improvement closes the gap to soft-Dice with the 0.5-threshold $\\mathrm{SD}^0$ which is expected by the theory (23).", + "bbox": [ + 89, + 414, + 485, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 89, + 493, + 205, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results in this work follow the results presented by [2] in the sense that the soft-Dice loss outperforms the cross-entropy loss with respect to the Dice metric when the 0.5-threshold is used and that there seems to be a trend of bigger discrepancy when performance for both of the loss functions is low. The results also explain the experimental observations of alternative thresholds outperforming the standard 0.5-threshold that was presented in [4, 12].", + "bbox": [ + 89, + 518, + 483, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In [24], it was proposed that the instability sometimes observed for soft-Dice may be associated with the fact that the loss function is non-convex (nor quasi-convex) when using sigmoid functions to enforce the bounds, that is $f \\mapsto \\mathrm{SD}_m(\\sigma \\circ f), f \\in \\mathcal{F}$ . This is in contrast to cross-entropy where the corresponding $f \\mapsto \\mathrm{CE}_m(\\sigma \\circ f), f \\in \\mathcal{F}$ is known to be convex. If cross-entropy with the alternative threshold yields the same performance as soft-Dice, then using it may provide the performance of soft-Dice without sacrificing the stability of cross-entropy. Another potential gain of using cross-entropy over soft-Dice is that marginal estimates are more informative than the approximate binary estimates that soft-Dice yields. These estimates can for instance be used to estimate uncertainty in different regions of a segmentation.", + "bbox": [ + 89, + 642, + 483, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A cost associated with using cross-entropy with the alternative threshold (22), is that the threshold needs to be computed for every prediction. This can be implemented ef", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ficiently but requires an extra step that may be non-negligible for sufficiently large images especially when there is a large number of predicted classes.", + "bbox": [ + 511, + 414, + 906, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 473, + 633, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, a realistic approximately volume preserving label noise model has been proposed and methods for explicit computation of marginal probabilities and sampling have been derived. Using the model, it has been shown how 0.5-thresholded optimal solutions to cross-entropy and soft-Dice are affected by various levels of noise, and how the resulting segmentations diverge as the noise is increased. It has further been shown how optimal solutions to soft-Dice can be recovered by thresholding the optimal solutions to cross-entropy with an a priori unknown but efficiently computable threshold generally different from 0.5. Finally, the theoretical observations are verified on data from the TotalSegmentor data set [37]. The presented results give insights into expected behaviour of widely used segmentation methods and suggest that cross-entropy with the alternative threshold yield the benefits of cross-entropy without sacrificing the performance of soft-Dice.", + "bbox": [ + 511, + 498, + 908, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. To be able to systematically study the effect of label noise with varying strengths, a noise model is used for experiments. The scope of the experiments is consequently limited to the kind of noise that is produced by the model.", + "bbox": [ + 511, + 773, + 908, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements The authors were supported by Ray-Search Laboratories AB.", + "bbox": [ + 511, + 867, + 908, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20827", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): A Completed Reference Database of Lung Nodules on CT Scans. Medical Physics, 38(2):915–931, 2011. 1, 2", + "[2] Jeroen Bertels, Tom Eelbode, Maxim Berman, Dirk Vandermeulen, Frederik Maes, Raf Bisschops, and Matthew B Blaschko. Optimizing the Dice Score and Jaccard Index for Medical Image Segmentation: Theory and Practice. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 92-100. Springer, 2019. 1, 8", + "[3] Jeroen Bertels, David Robben, Dirk Vandermeulen, and Paul Suetens. Theoretical Analysis and Experimental Validation of Volume Bias of Soft Dice Optimized Segmentation Maps in the Context of Inherent Uncertainty. Medical Image Analysis, 67:101833, 2021. 2", + "[4] Noah Bice, Neil Kirby, Ruiqi Li, Dan Nguyen, Tyler Bahr, Christopher Kabat, Pamela Myers, Niko Papanikolaou, and Mohammad Fakhreddine. A sensitivity analysis of probability maps in deep-learning-based anatomical segmentation. Journal of applied clinical medical physics, 22(8):105-119, 2021. 2, 6, 8", + "[5] Pete Bridge, Andrew Fielding, Pamela Rowntree, and Andrew Pullar. Intraobserver Variability: Should We Worry? Journal of Medical Imaging and Radiation Sciences, 47(3):217-220, 2016. 1, 2", + "[6] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation From Sparse Annotation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 424–432. Springer, 2016. 7", + "[7] Getao Du, Xu Cao, Jimin Liang, Xueli Chen, and Yonghua Zhan. Medical image segmentation based on u-net: A review. Journal of Imaging Science and Technology, 64:1-12, 2020. 1", + "[8] Charley Gros, Andreanne Lemay, and Julien Cohen-Adad. SoftSeg: Advantages of Soft Versus Binary Training for Image Segmentation. Medical Image Analysis, 71:102038, 2021. 2, 3", + "[9] Yun Gu, Mali Shen, Jie Yang, and Guang-Zhong Yang. Reliable label-efficient learning for biomedical image recognition. IEEE Transactions on Biomedical Engineering, 66(9):2423-2432, 2018. 2", + "[10] Nicholas Heller, Joshua Dean, and Nikolaos Papanikolopoulos. Imperfect segmentation labels: How much do they matter? In Intravascular Imaging and Computer Assisted Stenting and Large-Scale Annotation of Biomedical Data and Expert Label Synthesis, pages 112-120. Springer, 2018. 1, 2", + "[11] Takeyuki Hida and Masuyuki Hitsuda. Gaussian processes, volume 120. American Mathematical Soc., 1993. 4", + "[12] Burhan Rashid Hussein, Cédric Meurée, Malo Gaubert, Arthur Masson, Anne Kerbrat, Benoit Combès, and Francesca Galassi. A study on loss functions and decision thresholds" + ], + "bbox": [ + 93, + 114, + 485, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for the segmentation of multiple sclerosis lesions on spinal cord mri. hal preprint hal-03865212, 2022. 2, 6, 8", + "[13] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 1", + "[14] Davood Karimi, Haoran Dou, Simon K Warfield, and Ali Gholipour. Deep Learning With Noisy Labels: Exploring Techniques and Remedies in Medical Image Analysis. Medical Image Analysis, 65:101759, 2020. 2", + "[15] Eytan Kats, Jacob Goldberger, and Hayit Greenspan. Soft Labeling by Distilling Anatomical Knowledge for Improved MS Lesion Segmentation. In 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), pages 1563-1566. IEEE, 2019. 2, 3", + "[16] Matthieu Lé, Jan Unkelbach, Nicholas Ayache, and Hervé Delingette. Sampling image segmentations for uncertainty quantification. Medical image analysis, 34:42-51, 2016. 2", + "[17] Andreanne Lemay, Charley Gros, and Julien Cohen-Adad. Label Fusion and Training Methods for Reliable Representation of Inter-Rater Uncertainty. arXiv preprint arXiv:2202.07550, 2022.2,3", + "[18] Hang Li, Dong Wei, Shilei Cao, Kai Ma, Liansheng Wang, and Yefeng Zheng. Superpixel-Guided Label Softening for Medical Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 227-237. Springer, 2020. 2, 3", + "[19] Zachary C Lipton, Charles Elkan, and Balakrishnan Naryanaswamy. Optimal Thresholding of Classifiers to Maximize F1 Measure. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 225-239. Springer, 2014. 2, 6", + "[20] Bingyuan Liu, Jose Dolz, Adrian Galdran, Riadh Kobbi, and Ismail Ben Ayed. Do we really need dice? the hidden region-size biases of segmentation losses. Medical Image Analysis, 91:103015, 2024. 2", + "[21] Alireza Mehrtash, William M Wells, Clare M Tempany, Purang Abolmaesumi, and Tina Kapur. Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging, 39(12):3868-3878, 2020. 2", + "[22] Zahra Mirikharaji, Yiqi Yan, and Ghassan Hamarneh. Learning to Segment Skin Lesions From Noisy Annotations. In Domain adaptation and representation transfer and medical image learning with less labels and imperfect data, pages 207-215. Springer, 2019. 2", + "[23] Guy Nir, Soheil Hor, Davood Karimi, Ladan Fazli, Brian F Skinnider, Peyman Tavassoli, Dmitry Turbin, Carlos F Villamil, Gang Wang, R Storey Wilson, et al. Automatic Grading of Prostate Cancer in Digitized Histopathology Images: Learning From Multiple Experts. Medical Image Analysis, 50:167-180, 2018. 1", + "[24] Marcus Nordström, Han Bao, Fredrik Lofman, Henrik Hult, Atsuto Maki, and Masashi Sugiyama. Calibrated Surrogate Maximization of Dice. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 269-278. Springer, 2020. 1, 8", + "[25] Marcus Nordstrom, Henrik Hult, Fredrik Lofman, and Jonas Söderberg. On image segmentation with noisy labels: Char" + ], + "bbox": [ + 516, + 92, + 906, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20828", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "acterization and volume properties of the optimal solutions to accuracy and dice. Advances in Neural Information Processing Systems, 35:34321-34333, 2022. 2, 3, 7", + "[26] Marcus Nordstrom, Henrik Hult, Atsuto Maki, and Fredrik Lofman. Noisy image segmentation with soft-dice. arXiv preprint arXiv, 2023. 2, 3", + "[27] Tufve Nyholm, Stina Svensson, Sebastian Andersson, Joakim Jonsson, Maja Sohlin, Christian Gustafsson, Elisabeth Kjellén, Karin Söderström, Per Albertsson, Lennart Blomqvist, et al. MR and CT Data With Multiobserver Delineations of Organs in the Pelvic Area—Part of the Gold Atlas Project. Medical Physics, 45(3):1295-1300, 2018. 1, 2", + "[28] Teodora Popordanoska, Jeroen Bertels, Dirk Vandermeulen, Frederik Maes, and Matthew B Blaschko. On the Relationship Between Calibrated Predictors and Unbiased Volume Estimation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 678-688. Springer, 2021. 2", + "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional Networks for Biomedical Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241. Springer, 2015. 1", + "[30] Axel-Jan Rousseau, Thijs Becker, Jeroen Bertels, Matthew B Blaschko, and Dirk Valkenborg. Post Training Uncertainty Calibration of Deep Networks for Medical Image Segmentation. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1052-1056. IEEE, 2021. 2", + "[31] Gennady Samoradnitsky and Murad S. Taqqu. Stable non-Gaussian random processes: stochastic models with infinite variance. Chapman and Hall, 1994. 4", + "[32] Nahian Siddique, Sidike Paheding, Colin P Elkin, and Vijay Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. *IEEE Access*, 9:82031-82057, 2021. 1", + "[33] Joao Lourenço Silva and Arlindo L Oliveira. Using Soft Labels to Model Uncertainty in Medical Image Segmentation. arXiv preprint arXiv:2109.12622, 2021. 2, 3", + "[34] Nima Tajbakhsh, Laura Jeyaseelan, Qian Li, Jeffrey N Chiang, Zhihao Wu, and Xiaowei Ding. Embracing Imperfect Datasets: A Review of Deep Learning Solutions for Medical Image Segmentation. Medical Image Analysis, 63:101693, 2020. 2", + "[35] Eugene Vorontsov and Samuel Kadoury. Label noise in segmentation networks: mitigation must deal with bias. In Deep Generative Models, and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 1, pages 251-258. Springer, 2021. 1, 2", + "[36] Zifu Wang, Teodora Popordanoska, Jeroen Bertels, Robin Lemmens, and Matthew B Blaschko. Dice semimetric losses: Optimizing the dice score with soft labels. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 475-485. Springer, 2023. 2", + "[37] Jakob Wasserthal, Manfred Meyer, Hanns-Christian Breit, Joshy Cyriac, Shan Yang, and Martin Segeroth. Totalsegmentator: robust segmentation of 104 anatomical structures in ct images. arXiv preprint arXiv:2208.05868, 2022. 2, 7, 8" + ], + "bbox": [ + 91, + 90, + 483, + 893 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Shaode Yu, Mingli Chen, Erlei Zhang, Junjie Wu, Hang Yu, Zi Yang, Lin Ma, Xuejun Gu, and Weiguo Lu. Robustness Study of Noisy Annotation in Deep Learning Based Medical Image Segmentation. Physics in Medicine & Biology, 65(17):175007, 2020. 1, 2", + "[39] Ming-Jie Zhao, Narayanan Edakunni, Adam Pocock, and Gavin Brown. Beyond Fano's Inequality: Bounds on the Optimal F-score, BER, and Cost-Sensitive Risk and Their Implications. The Journal of Machine Learning Research, 14(1):1033-1090, 2013. 2" + ], + "bbox": [ + 516, + 90, + 906, + 229 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20829", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_model.json b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb4f1231ef02a041d94478083eb9527c50d97152 --- /dev/null +++ b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.131, + 0.78, + 0.177 + ], + "angle": 0, + "content": "The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.203, + 0.756, + 0.257 + ], + "angle": 0, + "content": "Marcus Nordström1,2 Atsuto Maki1 Henrik Hult1 \n1KTH Royal Institute of Technology 2RaySearch Laboratories {marcno,atsuto,hult}@kth.se" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.486, + 0.614 + ], + "angle": 0, + "content": "In image segmentation and specifically in medical image segmentation, the soft-Dice loss is often chosen instead of the more traditional cross-entropy loss to improve performance with respect to the Dice metric. Experimental work supporting this claim exists, but how and why the two loss functions lead to different predictions is not well understood. This paper explains the observed discrepancy as a consequence of how those loss functions are effected by label noise and what threshold is used to convert the predicted soft segmentation to predicted labels. In particular, it is shown (i) how the optimal solutions to the two loss functions diverge as the noise is increased, and (ii) how the optimal solutions to soft-Dice can be recovered by thresholding the solutions to cross-entropy with an a priori unknown but efficiently computable threshold. The theoretical results are supported by numerical experiments and it is concluded that cross-entropy with the alternative threshold yields the stability and informative label probability maps associated with cross-entropy without sacrificing the performance of soft-Dice." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.647, + 0.222, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In medical image analysis it is often of interest to partition an image into foreground and background with respect to some region of interest. This problem is referred to as the medical image segmentation problem. Since the introduction of the U-net [29], state of the art methods for addressing the problem have been dominated by neural network methods trained according to the supervised learning paradigm [7, 32]. These methods are remarkably effective in scenarios with abundant and clean data. However, manually delineating a region of interest accurately and reliably can be a complex and time-consuming process that often requires medical expertise. This result in high costs associated with obtaining high-quality data sets for training, and consequently, that data used in practice often is imperfect in the sense that the manually provided labels are contaminated by noise [1, 5, 23, 27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.909, + 0.627 + ], + "angle": 0, + "content": "A large amount of research has been devoted to medical image segmentation with imperfect data. One important question that has been addressed is how experimentally robust various segmentation methods are to label noise. In such work, experiments are conducted by first obtaining some data set and then corrupting the portion of data used for training with various types of noise. Different models are then trained on the resulting data sets and evaluated on uncorrupted data. If both noisy labels and noise-free labels can be obtained from clinical practitioners, it is possible to corrupt a clean label simply by swapping it with the corresponding noisy label. Such a study of label noise was presented in [38]. An alternative is to start with a data set that is assumed to be noise-free and generate noisy data with a noise model. Simple noise models that have been studied include randomly flipping labels in an independent manner or randomly translating the target structure [10]. Models that yield more realistic looking segmentation noise include randomly perturbing some points of the boundary on a noise-free segmentation with Gaussian variables [10] and randomly deforming a segmentation with an interpolated low-dimensional Gaussian array [35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "How label noise affects the performance of segmentation models is highly dependent on the loss function that is used during training. Arguably, the most classical loss function considered in the field is the cross-entropy loss. This loss is known to have a lot of nice theoretical properties and to be stable in practice. However, because of the class imbalance associated with the foreground typically being significantly smaller than the background, other loss functions putting more emphasis on foreground are often considered. The most popular such loss is a smoothed version of the Dice metric. This loss is referred to as soft-Dice or the Dice loss, and has been shown to yield better performance than cross-entropy in several experimental settings at the cost of sometimes introducing some stability issues [2, 24]. To get some of the improved performance while limiting potential stability issues, it is common to use linear combinations of the losses. This is for instance the type of loss used in the popular nnU-net [13]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20820" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.396 + ], + "angle": 0, + "content": "Contribution. This work introduces a noise model based on Gaussian field deformations. The model is shown to be approximately volume preserving in the sense that the expected volume of the resulting noisy segmentation is approximately equal to the volume of the corresponding noise-free segmentation. Moreover, a closed form expression for the marginal probabilities and a representation of random deformations using stochastic integrals are provided. This allows the behaviour of the model to be clearly interpreted and efficiently simulated. With the noise model, optimal solutions to cross-entropy and soft-Dice are studied. It is shown how the optimal solutions to soft-Dice and cross-entropy coincide when the labels are noise-free but diverge when noise is increased. It is further shown how optimal solutions to soft-Dice can be recovered by thresholding solutions to cross-entropy with a particular a priori unknown but efficiently computable threshold. Finally, the presented theory is validated experimentally on three different organ segmentation problems from the publicly available TotalSegmentor data set [37] with noise introduced by the model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.233, + 0.423 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.486, + 0.614 + ], + "angle": 0, + "content": "Label noise is a well known issue in medical image segmentation and the impact of observer variability has been discussed in various domains for different modalities. This include for radiotherapy [5], for lung nodules on clinical thoracic CT images [1] and for organs in the pelvic region for CT and MR [27]. Examples of work on methods targeting these issues include active learning methods, where experts are iteratively probed for more information [9] and re-weighting schemes where methods dynamically learn to put different weights on different regions depending on certainty [22]. For fairly recent reviews on image segmentation with imperfect data in general, see [14, 34]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.486, + 0.855 + ], + "angle": 0, + "content": "The robustness of medical image segmentation to noise can be studied by training models on data contaminated with varying degrees of noise. In [38], both noise-free and noisy labels are collected and the noise level is determined by the fraction of the training data taken from the noisy data set. Of special importance to this work is segmentation noise provided by Gaussian field deformations. Two works have studied such noise in an experimental setting [10, 35]. In these works, the segmentations in the training data is perturbed according to a low dimensional Gaussian field interpolated to the pixel space. Models are then trained for various noise levels and results obtained are reported. Finally, in [16], a noise model for target delineation in radiotherapy is discussed. The model is based on taking the level set of Gaussian field and efficient methods for sampling is provided." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.902 + ], + "angle": 0, + "content": "The influence label noise has on various loss functions has been studied in several papers. This include the influence it has on calibration and volume bias for predictions generated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.318 + ], + "angle": 0, + "content": "with models trained using cross-entropy and soft-Dice [3, 20], re-calibration methods for improving performance of soft-Dice [30] and the relationship between calibration and volume bias [28]. Other experimental work showing that the expected calibration error is lower for cross-entropy than for soft Dice includes [21]. In the above, noise has been analyzed by considering the expected score with respect to some noisy segmentation. An alternative way to incorporate noise is to make use of soft labels. Examples of work using soft labels with soft-Dice include [8, 15]. Examples of work on soft labels with other loss functions include [17, 18, 33]. In [36] a modified version of soft-Dice is introduced. This loss yields optimal solutions that coincide with the soft labels provided for the target structure which is known to not be the case for the classical soft-Dice loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.909, + 0.577 + ], + "angle": 0, + "content": "Inspired by threshold classifiers studied in [19, 39], optimal segmentations to Accuracy and Dice under label noise was analyzed in [25]. A characterization of optimal segmentations was provided, sharp bounds with respect to volume bias of the optimal segmentations where shown and the volume of the optimizers associated with the metrics was compared. A similar analysis was later provided for the loss function soft-Dice [26]. In particular, a characterization of the optimal solutions and the associated volumes are provided. Moreover, a calibration result showing that a sequence that converge to optimal soft-Dice, when thresholded appropriately, converges to optimal Dice. Experimental work comparing various thresholds for cross-entropy was presented in [4, 12]. The results from these papers experimentally show that using 0.5 as a common threshold leads to sub-optimal segmentations for various loss functions with respect to Dice." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.594, + 0.652, + 0.61 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.618, + 0.635 + ], + "angle": 0, + "content": "3.1. Notation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The notation largely follows [25, 26]. Let \\(\\Omega = [0,1]^n\\subset \\mathbb{R}^n\\) be the unit cube of dimension \\(n\\geq 1\\) and \\(\\lambda\\) be the standard Lebesgue measure associated with \\(\\mathbb{R}^n\\) such that \\(\\lambda (\\Omega) = 1\\). Let \\(S\\) be the space of measurable functions from \\(\\Omega\\) to \\(\\{0,1\\}\\), \\(\\mathcal{M}\\) be the space of measurable functions from \\(\\Omega\\) to \\([0,1]\\) and \\(\\mathcal{F}\\) be the space of bounded measurable functions from \\(\\Omega\\) to \\(\\mathbb{R}\\), where all of the spaces \\(S\\), \\(\\mathcal{M}\\) and \\(\\mathcal{F}\\) are equipped with their associated Borel \\(\\sigma\\)-fields and the \\(L_{p}\\)-norms \\(\\| \\cdot \\| _p\\). The letter \\(s\\in S\\) is generally used to denote a segmentation, where \\(s(\\omega) = 1\\) indicates foreground of the target structure of interest and \\(s(\\omega) = 0\\) indicates background. The letter \\(m\\in \\mathcal{M}\\) is generally used to denote a marginal function, that is, \\(m(\\omega)\\in [0,1]\\) indicates the probability of a noisy label occupying site \\(\\omega \\in \\Omega\\). The letter \\(c\\in \\mathcal{M}\\) is generally used to denote the soft segmentation that is obtained by optimizing the chosen loss function prior to any thresholding. The letter \\(f\\in \\mathcal{F}\\) is generally used to denote a logit function usually" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20821" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "associated with a soft segmentation such that \\(\\sigma \\circ f(\\omega) = c(\\omega)\\) for \\(\\omega \\in \\Omega\\) a.e., where \\(\\sigma(x) \\doteq 1/(1 + e^{-x})\\) is the standard sigmoid function. Composition of functions are denoted by \\(\\circ\\) and \\(I_A(x)\\) is used to denote the indicator function over the set \\(A\\), that is \\(I_A(x) = 1\\) if \\(x \\in A\\) and \\(I_A(x) = 0\\) if \\(x \\notin A\\). The indicator function is used in conjunction with some element \\(c \\in \\mathcal{M}\\) to denote thresholding of a soft segmentation. That is, \\(I_{[t,1]} \\circ c\\) assigns 1 to all \\(\\omega \\in \\Omega\\) where \\(c(\\omega) \\geq t\\) and 0 to all \\(\\omega \\in \\Omega\\) where \\(c(\\omega) < t\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.237, + 0.429, + 0.252 + ], + "angle": 0, + "content": "3.2. Loss Functions and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.485, + 0.457 + ], + "angle": 0, + "content": "Two of the most popular loss functions in the field of medical image segmentation are cross-entropy and soft-Dice. Usually these loss functions are defined over two deterministic arguments, a soft-segmentation (prediction) and a binary segmentation (label). Under the influence of label noise, these loss functions are extended to functions taking one deterministic soft segmentation (prediction) and one random segmentation (label) as arguments. In this work, the soft labeling extension is adopted for the theoretical analysis [8, 15, 17, 18, 33] (see Section 4.4 for more on this). Formally, the loss functions are defined with respect to a marginal function \\( m \\in \\mathcal{M} \\) which corresponds to the mean of a noisy segmentation \\( L \\) taking values in \\( S \\), that is \\( m(\\omega) = \\mathbb{E}[L(\\omega)], \\omega \\in \\Omega \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.484, + 0.497 + ], + "angle": 0, + "content": "Definition 1. For any \\(m \\in \\mathcal{M}\\), (binary) cross-entropy is given by" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.507, + 0.484, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {C E} _ {m} (c) \\doteq - \\int_ {\\Omega} [ m (\\omega) \\log (c (\\omega)) + \\quad c \\in \\mathcal {M}. \\tag {1} \\\\ (1 - m (\\omega)) \\log (1 - c (\\omega)) ] \\lambda (d \\omega), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.57, + 0.441, + 0.585 + ], + "angle": 0, + "content": "Definition 2. For any \\(m \\in \\mathcal{M}\\), soft-Dice is given by" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.595, + 0.484, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S D} _ {m} (c) \\doteq 1 - \\frac {2 \\int_ {\\Omega} c (\\omega) m (\\omega) \\lambda (d \\omega)}{\\| c \\| _ {1} + \\| m \\| _ {1}}, \\quad c \\in \\mathcal {M}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.746 + ], + "angle": 0, + "content": "Arguably, the most classical evaluation metric used is Accuracy. This metric simply measures the fraction of the domain that is correctly labeled. The most popular evaluation metric on the other hand is the Sorensen-Dice coefficient, or Dice for short. To handle label noise, these metrics are defined in the same way as the loss functions, that is, by using soft labels." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.756, + 0.442, + 0.772 + ], + "angle": 0, + "content": "Definition 3. For any \\(m \\in \\mathcal{M}\\), Accuracy is given by" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.782, + 0.484, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {A} _ {m} (s) \\doteq \\int_ {\\Omega} [ s (\\omega) m (\\omega) + \\quad s \\in \\mathcal {S}. \\tag {3} \\\\ (1 - s (\\omega)) (1 - m (\\omega)) ] \\lambda (d \\omega), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.844, + 0.412, + 0.86 + ], + "angle": 0, + "content": "Definition 4. For any \\(m \\in \\mathcal{M}\\), Dice is given by" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.87, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} _ {m} (s) \\doteq \\frac {2 \\int_ {\\Omega} s (\\omega) m (\\omega) \\lambda (d \\omega)}{\\| s \\| _ {1} + \\| m \\| _ {1}}, \\quad s \\in \\mathcal {S}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "When a label is free from noise, any reasonable loss function or metric is optimized uniquely by the target label. When there is noise present on the other hand, it is no longer clear what the solution should be and loss functions or metrics implicitly assign them. As the consequence, it is important to investigate what the optimal solutions are for a chosen loss or metric. For the case of the cross-entropy loss, it is well known that the optimizers are uniquely given by the marginal function, that is" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.235, + 0.907, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C E} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {C E} _ {m} \\left(c ^ {\\prime}\\right) \\Longleftrightarrow c (\\omega) = m (\\omega), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.265, + 0.908, + 0.311 + ], + "angle": 0, + "content": "\\(\\lambda\\)-a.e. In other words, minimizing cross-entropy yields estimates of the marginal probability distribution associated with the labels. For soft-Dice it was shown in [26] that" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.316, + 0.722, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S D} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {S D} \\left(c ^ {\\prime}\\right) \\Longleftrightarrow\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.343, + 0.907, + 0.399 + ], + "angle": 0, + "content": "\\[\nc (\\omega) \\in \\left\\{ \\begin{array}{l l} \\{1 \\} & \\text {i f} m (\\omega) > \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2, \\\\ [ 0, 1 ] & \\text {i f} m (\\omega) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2, \\\\ \\{0 \\} & \\text {i f} m (\\omega) < \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2, \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.907, + 0.452 + ], + "angle": 0, + "content": "\\(\\lambda\\)-a.e. Characterizations of the optimal segmentations for Accuracy and Dice using the setup adopted in this paper can be found in [25]. For Accuracy, the solutions are given by" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.47, + 0.776, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} _ {m} (s) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {A} _ {m} \\left(s ^ {\\prime}\\right) \\iff\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.499, + 0.907, + 0.555 + ], + "angle": 0, + "content": "\\[\ns (\\omega) \\in \\left\\{ \\begin{array}{l l} \\{1 \\} & \\text {i f} m (\\omega) > 1 / 2, \\\\ \\{0, 1 \\} & \\text {i f} m (\\omega) = 1 / 2, \\\\ \\{0 \\} & \\text {i f} m (\\omega) < 1 / 2, \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.562, + 0.906, + 0.592 + ], + "angle": 0, + "content": "\\(\\lambda\\)-a.e. Similarly, for Dice the optimal segmentations are given by" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.599, + 0.907, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} _ {m} (s) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) \\Longleftrightarrow \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.629, + 0.906, + 0.685 + ], + "angle": 0, + "content": "\\[\ns (\\omega) \\in \\left\\{ \\begin{array}{l l} \\{1 \\} & \\text {i f} m (\\omega) > \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} (s ^ {\\prime}) / 2, \\\\ \\{0, 1 \\} & \\text {i f} m (\\omega) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} (s ^ {\\prime}) / 2, \\\\ \\{0 \\} & \\text {i f} m (\\omega) < \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} (s ^ {\\prime}) / 2, \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.907, + 0.767 + ], + "angle": 0, + "content": "\\(\\lambda\\)-a.e. The resulting soft segmentation obtained by optimizing a loss function is typically converted to a binary segmentation by thresholding with 0.5. Ideally this resulting segmentation maximizes a desired target metric. Using the above characterizations, for cross-entropy it follows that" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.773, + 0.808, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C E} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {C E} \\left(c ^ {\\prime}\\right) \\Longrightarrow\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.794, + 0.907, + 0.824 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} _ {m} \\left(I _ {[ 1 / 2, 1 ]} \\circ c\\right) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {A} \\left(s ^ {\\prime}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.831, + 0.766, + 0.846 + ], + "angle": 0, + "content": "Similarly, for soft-Dice it follows that" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.853, + 0.808, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S D} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {S D} \\left(c ^ {\\prime}\\right) \\Longrightarrow\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.871, + 0.907, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} _ {m} \\left(I _ {[ 1 / 2, 1 ]} \\circ c\\right) = \\sup _ {s ^ {\\prime} \\in S} \\mathrm {D} \\left(s ^ {\\prime}\\right). \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20822" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.121 + ], + "angle": 0, + "content": "3.3. Gaussian Fields and Independently Scattered Gaussian Measures" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.129, + 0.484, + 0.19 + ], + "angle": 0, + "content": "Amongst the most central objects in probability theory and statistics are Gaussian variables taking values in Euclidean space. Such a variable, when centered and isotropic, has the probability density function" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.201, + 0.484, + 0.236 + ], + "angle": 0, + "content": "\\[\np _ {\\sigma^ {2}} (\\omega) \\dot {=} \\frac {1}{(2 \\pi \\sigma^ {2}) ^ {n / 2}} \\exp \\left\\{- \\frac {\\| \\omega \\| _ {2} ^ {2}}{2 \\sigma^ {2}} \\right\\}, \\quad \\omega \\in \\mathbb {R} ^ {n}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.486, + 0.324 + ], + "angle": 0, + "content": "where \\(\\sigma^2\\) denotes the variance in each of the dimensions. Various generalizations of this random variable have been studied extensively in the literature. For the theoretical treatment of the noise model in this work, two such generalizations are introduced. The first is a Gaussian field." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.414 + ], + "angle": 0, + "content": "Definition 5. A random variable \\( Y(\\omega), \\omega \\in \\mathbb{R}^n \\) is a Gaussian field with kernel function \\( k \\) if any finite collection of elements \\( \\{\\omega_1, \\dots, \\omega_M\\} \\subset \\mathbb{R}^n \\) satisfies \\( (Y(\\omega_1), \\dots, Y(\\omega_M)) \\sim \\text{normal}(0, \\Sigma) \\), with \\( \\Sigma_{m,m'} = k(\\omega_m, \\omega_{m'}) \\) for \\( m, m' \\in \\{1, \\dots, M\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.485, + 0.47 + ], + "angle": 0, + "content": "Many types of kernel functions with different properties have been studied in the literature. Amongst the most common is the squared exponential kernel given by" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.48, + 0.484, + 0.53 + ], + "angle": 0, + "content": "\\[\nk _ {a, b} \\left(\\omega , \\omega^ {\\prime}\\right) \\doteq a ^ {2} \\exp \\left(- \\frac {\\| \\omega - \\omega^ {\\prime} \\| _ {2} ^ {2}}{2 b ^ {2}}\\right), \\quad \\omega , \\omega^ {\\prime} \\in \\mathbb {R} ^ {n}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.618 + ], + "angle": 0, + "content": "An important property of Gaussian fields with such covariance functions are that samples are almost surely continuous. In other words, a randomly sampled function will be continuous with probability one. For a good reference on Gaussian variables, see [11]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.695 + ], + "angle": 0, + "content": "The second generalization that is introduced is what is referred to as independently scattered Gaussian measures. Such constructions are used to formalize integration over Gaussian fields, similar to how the Ito-integral build on Brownian motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.798 + ], + "angle": 0, + "content": "Definition 6. A random variable \\( W \\) is an independently scattered Gaussian measure on \\( \\mathbb{R}^n \\) with respect to the control measure \\( \\lambda \\), if for any collection of disjoint sets \\( A_1, \\ldots, A_k \\) in \\( \\lambda \\)-measurable sets with finite measure \\( \\mathcal{B}_0 \\), \\( W(A_1), \\ldots, W(A_k) \\) are independent, and \\( W(A_i) \\) has centered Gaussian distribution with variance \\( \\lambda(A_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Importantly, for a function \\( f \\in L^{2}(\\mathbb{R}^{n}) \\), the stochastic integral \\( \\int_{\\mathbb{R}^n} f(\\omega) W(d\\omega) \\) is well defined and the process \\( Y = \\{Y(\\omega), \\omega \\in \\mathbb{R}^n\\} \\) given by" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.866, + 0.484, + 0.9 + ], + "angle": 0, + "content": "\\[\nY (\\omega) \\doteq \\int_ {\\mathbb {R} ^ {n}} f \\left(\\omega - \\omega^ {\\prime}\\right) W \\left(d \\omega^ {\\prime}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "is a centered Gaussian process. Furthermore, this Gaussian process has a covariance kernel given by" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.133, + 0.907, + 0.18 + ], + "angle": 0, + "content": "\\[\nk \\left(\\omega , \\omega^ {\\prime}\\right) = \\int_ {\\mathbb {R} ^ {n}} f (\\omega - u) f \\left(\\omega^ {\\prime} - u\\right) \\lambda (d u), \\quad \\omega , \\omega^ {\\prime} \\in \\mathbb {R} ^ {n}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.908, + 0.251 + ], + "angle": 0, + "content": "Independently scattered measures are introduced more generally for \\(\\alpha\\)-stable distributions in [31, Section 3.3]. The Gaussian case used in this work is simply the special case when \\(\\alpha = 2\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.266, + 0.6, + 0.283 + ], + "angle": 0, + "content": "4. Theory" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.291, + 0.695, + 0.306 + ], + "angle": 0, + "content": "4.1. Label Noise Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.448 + ], + "angle": 0, + "content": "In contrast to binary classification, where a lot of label noise can be modeled by simple independent flipping of binary variables, label noise in segmentation can in general have very complicated dependency structures. Often labels affected by noise have the appearance of being drawn by someone with shaky hands. For such noise, the correlation between border points will depend on the distance between them. That is, if two points are close to each other, then the correlation will be high and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.906, + 0.493 + ], + "angle": 0, + "content": "One way to model segmentations that look as if they have been drawn by shaky hands is by means of random Gaussian field deformations. Consider a variable" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.505, + 0.907, + 0.524 + ], + "angle": 0, + "content": "\\[\nX \\doteq \\left(X _ {1}, \\dots , X _ {n}\\right) ^ {T} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.908, + 0.597 + ], + "angle": 0, + "content": "such that \\(\\{X_i\\}_{i=1}^n\\) are independent Gaussian fields with kernel function (13). This means that \\(X\\) is a random vector field taking values in \\(\\mathbb{R}^n \\mapsto \\mathbb{R}^n\\). With such a random vector field, a random segmentation can be formed by" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.607, + 0.907, + 0.662 + ], + "angle": 0, + "content": "\\[\nL (\\omega) \\doteq \\left\\{ \\begin{array}{l l} l (\\omega + X (\\omega)) & \\text {i f} \\omega + X (\\omega) \\in \\Omega , \\\\ 0 & \\text {i f} \\omega + X (\\omega) \\notin \\Omega . \\end{array} \\right., \\quad \\omega \\in \\Omega , \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.84 + ], + "angle": 0, + "content": "where \\( l \\in S \\) is a noise-free reference segmentation being randomly deformed. How samples of the random segmentations look depend on what parameters are chosen when specifying the kernel function (13). The parameter \\( a \\) controls how large the deformations are, where greater \\( a \\) leads to larger deformations and vice versa. The parameter \\( b \\) controls how smooth the deformations are, where greater \\( b \\) leads to smoother deformations and vice versa. Modeling noisy segmentations with (17) is convenient because it allows for the derivation of several theoretical properties. In this work, three such important properties are derived." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In Proposition 1, it is shown that the marginal probabilities associated with a noisy segmentation can be computed by convolving the associated noise-free segmentation label with an appropriate Gaussian density." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20823" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.122 + ], + "angle": 0, + "content": "Proposition 1. Let \\( l \\) and \\( L \\) be defined as in (17) and \\( p_{\\sigma^2} \\) be defined as in (12), then" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.131, + 0.486, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ L (\\omega) ] = \\int_ {\\Omega} l \\left(\\omega^ {\\prime}\\right) p _ {a ^ {2}} \\left(\\omega - \\omega^ {\\prime}\\right) \\lambda \\left(d \\omega^ {\\prime}\\right), \\quad \\omega \\in \\Omega . \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.485, + 0.356 + ], + "angle": 0, + "content": "In Proposition 2, the marginal representation is used to show that the volume of the noisy segmentation is approximately unbiased in the sense that the expected volume of a noisy segmentation is approximately equal to the volume of the corresponding noise-free segmentation. The reason why this does not hold exactly is that there is some probability for the target structure to deform in such a way that parts of the structure leave the domain \\(\\Omega\\). However, for most reasonable choices of parameters, when the regions of interest is sufficiently close to the center, the probability of this happening is negligible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.485, + 0.396 + ], + "angle": 0, + "content": "Proposition 2. Let \\( l \\) and \\( L \\) be defined as in (17) and \\( p_{\\sigma^2} \\) be defined as in 12, then" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.405, + 0.484, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ \\| L \\| _ {1} ] = \\| l \\| _ {1} - \\xi , \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.433, + 0.138, + 0.446 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.454, + 0.484, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\xi = \\int_ {\\Omega} \\int_ {\\mathbb {R} ^ {n} \\backslash \\Omega} l \\left(\\omega^ {\\prime}\\right) p _ {a ^ {2}} \\left(\\omega - \\omega^ {\\prime}\\right) \\lambda (d \\omega) \\lambda \\left(d \\omega^ {\\prime}\\right). \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.486, + 0.588 + ], + "angle": 0, + "content": "In Proposition 3, it is shown that \\( X \\) (16) is equal in distribution to a particular vector of stochastic integrals. In the case \\( n = 1 \\), this could have been done using regular Ito-integrals, but for the case \\( n > 1 \\) a multivariate counter part is needed. For this scattered Gaussian measures are used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.486, + 0.658 + ], + "angle": 0, + "content": "Proposition 3. Let \\( X \\) be as in (16), \\( p_{\\alpha^2} \\) be as in (12) and \\( W_1, \\ldots, W_n \\) be independent copies of independently scattered Gaussian measures on \\( \\mathbb{R}^n \\) with control measure \\( \\lambda \\), then it follows that" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.667, + 0.484, + 0.739 + ], + "angle": 0, + "content": "\\[\nX (\\omega) \\stackrel {d} {=} \\left( \\begin{array}{c} a (2 \\pi b ^ {2}) ^ {n / 4} \\int_ {\\mathbb {R} ^ {n}} p _ {b ^ {2} / 2} \\left(\\omega - \\omega^ {\\prime}\\right) W _ {1} \\left(d w ^ {\\prime}\\right) \\\\ \\vdots \\\\ a (2 \\pi b ^ {2}) ^ {n / 4} \\int_ {\\mathbb {R} ^ {n}} p _ {b ^ {2} / 2} \\left(\\omega - \\omega^ {\\prime}\\right) W _ {n} \\left(d w ^ {\\prime}\\right) \\end{array} \\right). \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The presented propositions provide meaningful insights into the behaviour of the noise model. The importance of Proposition 1 is that it makes the relationship between the noise model parameters and the resulting marginal probabilities clear. Also, it provides an exact method for computing the marginals that otherwise would have to be estimated with sampling. The importance of Proposition 2 is that the noise beside from looking reasonable also numerically behave reasonable in an important manner. Finally, the importance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "of Proposition 3 is that noise for discretized domains can be sampled by smoothing arrays of independently sampled Gaussian variables with an appropriate Gaussian filters. Both operations which scale well with image size. This allows for efficient sampling of noisy segmentations for images composed of millions of voxels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.191, + 0.879, + 0.208 + ], + "angle": 0, + "content": "4.2. Optimal Segmentations Under Label Noise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Since the loss functions and metrics investigated in this work only depend on the marginal probabilities and Proposition 1 provides an explicit expression for these marginal probabilities, it is possible to study how noise model parameters affect optimal segmentations. Figure 1 depicts this relationship for a two dimensional segmentation problem with varying noise intensity. In the first row, pixel-wise averages of 5 samples are illustrated. In the second row, the marginal probabilities are illustrated. In the third row, theoretically optimal solutions to Accuracy are shown. Note that by (10), this is equivalently given by the minimizer to cross-entropy thresholded by 0.5. In the forth row, theoretical optimal solutions to Dice are shown. Note that by (11), this is equivalently given by the minimizer to soft-Dice thresholded by 0.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.909, + 0.575 + ], + "angle": 0, + "content": "A first observation is that noise affects the optimal segmentations in such a way that the shape of the noise-free reference segmentation is changed. This is an important observation that contradicts the picture that unbiased noise (unbiased in the sense that the expected deformation is the identity function) yield unbiased optimal segmentations. In particular, optimal segmentations are changed in such a way that the corners of the shapes are smoothed and details such as thin structures and holes disappear. Informally, the higher the noise, the more ball shaped the optimal segmentations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.911, + 0.757 + ], + "angle": 0, + "content": "A second observation is that the two metrics yield similar looking segmentations for low noise but increasingly different segmentations as the noise increase. Since cross-entropy with the 0.5-threshold theoretically gives optimal segmentations with respect to Accuracy and soft-Dice with the 0.5-threshold theoretically gives optimal segmentations with respect to Dice, the performance of the two loss functions as measured by Dice is similar in low noise situations but increasingly better for soft-Dice as the noise is increased. Similarly, the performance of the two loss functions as measured by Accuracy is similar in low noise situations but increasingly better for cross-entropy as the noise is increased." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.9, + 0.782 + ], + "angle": 0, + "content": "4.3. Cross-entropy With an Alternative Threshold" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.908, + 0.863 + ], + "angle": 0, + "content": "By the characterization of optimizers to cross-entropy (5) and the characterization of optimizers to soft-Dice (6), it follows that an optimizer to soft-Dice can be obtained by first optimizing cross-entropy and then thresholding with the a priori unknown value" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.873, + 0.908, + 0.898 + ], + "angle": 0, + "content": "\\[\nt _ {m} \\doteq \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2. \\tag {22}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20824" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.904, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.472, + 0.908, + 0.543 + ], + "angle": 0, + "content": "Figure 1. Illustration of the noise model and its effects on a two-dimensional segmentation problem, with the columns showing the effect of different noise strengths. The first row shows the pixel-wise average obtained from five samples. The second row shows the marginal function. The third row shows optimal segmentations associated with Accuracy (7), which in this and most natural cases is unique. The forth row shows optimal segmentations associated with Dice (9), which in this and most natural cases is unique. In all of the reported cases \\( b = 0.15 / \\sqrt{2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.569, + 0.487, + 0.613 + ], + "angle": 0, + "content": "That is, by thresholding an optimizer to cross-entropy with this alternative recursively defined threshold \\( t_m \\), it follows that" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.634, + 0.484, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {C E} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {C E} \\left(c ^ {\\prime}\\right) \\Longrightarrow \\tag {23} \\\\ \\mathrm {S D} _ {m} \\left(I _ {[ t _ {c}, 1 ]} \\circ c\\right) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {S D} \\left(c ^ {\\prime}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Methods based on this principle have not been discussed in the medical image segmentation community but are known in the statistical learning literature [19]. In the medical image segmentation community, however, there has been experimental work comparing the performance of different constant thresholds, that is, thresholds not dependent on the marginal probabilities [4, 12]. The results experimentally show that using a 0.5-threshold leads to sub-optimal segmentations for various loss functions when the Dice metric is used for evaluation. Consequently, the results are in line with the theoretical observations above and motivate further investigation of methods based on cross-entropy with the threshold (22)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.568, + 0.81, + 0.584 + ], + "angle": 0, + "content": "4.4. Soft Labeling and Expected Score" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.746 + ], + "angle": 0, + "content": "One reason to why the threshold (22) has not been discussed in the medical image segmentation domain may be that loss functions and metrics often incorporate noise using expectations rather than soft labels. Formally for some random label \\( L \\) taking values in \\( S \\) with \\( \\mathbb{E}[L(\\omega)] = m(\\omega), \\omega \\in \\Omega \\) and \\( c \\in \\mathcal{M} \\), this means that \\( \\mathbb{E}[\\mathrm{CE}_L(c)] \\) is considered instead of \\( \\mathrm{CE}_m(c) \\) and \\( \\mathbb{E}[\\mathrm{SD}_L(c)] \\) is considered instead of \\( \\mathrm{SD}_m(c) \\). Similarly for some \\( s \\in S \\), this means that \\( \\mathbb{E}[\\mathrm{A}_L(s)] \\) is considered instead of \\( \\mathrm{A}_m(s) \\) and \\( \\mathbb{E}[\\mathrm{D}_L(s)] \\) is considered instead of \\( \\mathrm{D}_m(s) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "For cross-entropy and Accuracy, it is easy to verify that the definitions using expectations and the definitions using soft labels are equivalent, that is \\(\\mathbb{E}[\\mathrm{A}_L(s)] = \\mathrm{A}_m(s), s \\in S\\) and \\(\\mathbb{E}[\\mathrm{CE}_L(c)] = \\mathrm{CE}_m(c), c \\in \\mathcal{M}\\). For soft-Dice and Dice, however, this is in general not the case, that is, in general \\(\\mathbb{E}[\\mathrm{D}_L(s)] \\neq \\mathrm{D}_m(s), s \\in S\\) and \\(\\mathbb{E}[\\mathrm{SD}_L(c)] \\neq \\mathrm{SD}_m(c), c \\in \\mathcal{M}\\). Hence, the property (23) does not in general transfer to properties of optimizers to soft-Dice using expectations. Importantly however, that the definitions of soft-Dice and Dice using soft labels and expectations are" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20825" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "equivalent provided that the volume of the noisy label is constant, that is \\(\\mathrm{Var}[\\| L\\| _1] = 0\\). For situations when the noise is approximately constant \\(\\mathrm{Var}[\\| L\\| _1]\\approx 0\\), the definitions are approximately equal \\(\\mathbb{E}[\\mathrm{D}_L(s)]\\approx \\mathrm{D}_m(s),s\\in S\\) and \\(\\mathbb{E}[\\mathrm{SD}_L(c)]\\approx \\mathrm{SD}_m(c),c\\in \\mathcal{M}\\) [25]. Hence the property (23) holds approximately for the kind of noise studied in this work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.212, + 0.223, + 0.229 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.238, + 0.178, + 0.252 + ], + "angle": 0, + "content": "5.1. Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.485, + 0.427 + ], + "angle": 0, + "content": "A standard 3D U-net is implemented [6]. The network is composed of blocks of two convolutions with kernel size \\(3^3\\), instance normalization and ReLu-activations. The number of filters used is increased for each down-sampling according to 16, 32, 64, 128, 256, where 16 corresponds to the number of filters used in the input layer and 256 corresponds to the number of filters in the bottleneck layer. Convolutions with kernel \\(2^3\\) and stride 2 are used for down-sampling. Transposed convolutions with kernel \\(2^3\\) stride 2 are used for up-sampling. Details and code can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.438, + 0.195, + 0.453 + ], + "angle": 0, + "content": "5.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Models are trained either with the cross-entropy loss or with the soft-Dice loss. In both cases \\( l_{2} \\) regularization is used with a weight of \\( 10^{-5} \\). Adam with learning rate \\( 10^{-4} \\) is used for optimization together with a scheduler that divides the learning rate by 5 after 30 epochs of non-improvement. The number of epochs computed is fixed to 100 and the order of the training data is randomly shuffled. Batch sizes of 1 is used and with probability 0.5 the data is augmented. The augmentation is composed of random shifts, random rotations and image blurring with Gaussian filters. The model that performs best on the training data is the model selected. Details and code can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.165, + 0.682 + ], + "angle": 0, + "content": "5.3.Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The experiments are conducted with respect to the TotalSegmentor data set [37]. This data set contains 1204 CT images with 104 anatomical structures (27 organs, 59 bones, 10 muscles, 8 vessels). To illustrate the effect noise may have on organs with different shape, three different organs are chosen. This includes the right kidney which in general is spherical, the aorta which is tubular and relatively thick and the esophagus which is tubular and relatively thin. For each of the organs, 400 cases is selected and split into 5 folds each containing 80 cases. The images are sub-sampled to half resolution and patches of \\(64^{3}\\) voxels centered in each of the structures are extracted. In order to simulate label noise the described noise model is used. For each organ data set, four different datasets was created by randomly deforming the" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.646, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.09, + 0.773, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.09, + 0.903, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.19, + 0.645, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.19, + 0.774, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.19, + 0.903, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.289, + 0.645, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.289, + 0.774, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.289, + 0.903, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.398, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Figure 2. Illustration of the data from the TotalSegmentor data set [37] used in the experiments. The first depicts a kidney, the second row an aorta and the third row an esophagus. The columns correspond to axial, coronal and sagittal views respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.545 + ], + "angle": 0, + "content": "labels with the noise strengths \\( a = 0.0, 0.1, 0.2, 0.3 \\). Each data point is deformed once, resulting in one sample for each random label. Details and code can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.558, + 0.608, + 0.573 + ], + "angle": 0, + "content": "5.4. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Results from the experiments are illustrated separately for the three organ segmentation problems as a figure in Figure 3 and listed as a table in the Supplementary Material. For each organ segmentation problem, the average Dice score for the various methods is reported as a function of the parameter controlling the noise strength in the noise model. The average score is obtained by taking the average of the results obtained after 5 folds. The notation \\(\\mathrm{CE}^{(0)}\\) is used to denote a model trained with cross-entropy and thresholded with 0.5, \\(\\mathrm{SD}^{(0)}\\) a model trained with cross-entropy and thresholded with 0.5 and \\(\\mathrm{CE}^{(*)}\\) a model trained with cross-entropy and thresholded with \\(t_m\\) (22)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.854 + ], + "angle": 0, + "content": "The first thing to notice is that the difference amongst the average performance of cross-entropy and soft-Dice with the 0.5-threshold, that is, \\(\\mathrm{CE}^{(0)}\\) and \\(\\mathrm{SD}^{(0)}\\), increase as the amount of added noise increase. When no noise is added, the performance is indistinguishable. This type of result is expected from the theory and depicted in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The second thing to notice is that cross-entropy with the alternative threshold (22) \\(\\mathrm{CE}^{(*)}\\) performs better than cross-entropy with the 0.5-threshold \\(\\mathrm{CE}^{(0)}\\). In regions with low" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20826" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.36, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.09, + 0.614, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.091, + 0.882, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.908, + 0.39 + ], + "angle": 0, + "content": "Figure 3. To the left are the results from the kidney segmentation problem, in the middle are the results from the aorta segmentation problem and to the right are the results from the esophagus segmentation problem. In each of the figures, the average Dice value obtained from five fold experiments is plotted as a function of the noise level \\(a\\). The legend \\(\\mathrm{CE}^{(0)}\\) indicate that cross-entropy with the \\(1/2\\)-threshold has been used, \\(\\mathrm{SD}^{(0)}\\) indicate that soft-Dice with the \\(1/2\\)-threshold has been used and \\(\\mathrm{CE}^{(*)}\\) indicate that cross-entropy with the threshold described in (22) has been used. In all of the reported cases \\(b = 0.15/\\sqrt{2}\\). A table of the results can be found in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.486, + 0.477 + ], + "angle": 0, + "content": "noise the discrepancy is not noticeable, but as the noise is increased the discrepancy also increases. The performance improvement closes the gap to soft-Dice with the 0.5-threshold \\(\\mathrm{SD}^0\\) which is expected by the theory (23)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.494, + 0.206, + 0.51 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.642 + ], + "angle": 0, + "content": "The results in this work follow the results presented by [2] in the sense that the soft-Dice loss outperforms the cross-entropy loss with respect to the Dice metric when the 0.5-threshold is used and that there seems to be a trend of bigger discrepancy when performance for both of the loss functions is low. The results also explain the experimental observations of alternative thresholds outperforming the standard 0.5-threshold that was presented in [4, 12]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.854 + ], + "angle": 0, + "content": "In [24], it was proposed that the instability sometimes observed for soft-Dice may be associated with the fact that the loss function is non-convex (nor quasi-convex) when using sigmoid functions to enforce the bounds, that is \\( f \\mapsto \\mathrm{SD}_m(\\sigma \\circ f), f \\in \\mathcal{F} \\). This is in contrast to cross-entropy where the corresponding \\( f \\mapsto \\mathrm{CE}_m(\\sigma \\circ f), f \\in \\mathcal{F} \\) is known to be convex. If cross-entropy with the alternative threshold yields the same performance as soft-Dice, then using it may provide the performance of soft-Dice without sacrificing the stability of cross-entropy. Another potential gain of using cross-entropy over soft-Dice is that marginal estimates are more informative than the approximate binary estimates that soft-Dice yields. These estimates can for instance be used to estimate uncertainty in different regions of a segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "A cost associated with using cross-entropy with the alternative threshold (22), is that the threshold needs to be computed for every prediction. This can be implemented ef" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.46 + ], + "angle": 0, + "content": "ficiently but requires an extra step that may be non-negligible for sufficiently large images especially when there is a large number of predicted classes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.474, + 0.634, + 0.489 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.909, + 0.756 + ], + "angle": 0, + "content": "In this work, a realistic approximately volume preserving label noise model has been proposed and methods for explicit computation of marginal probabilities and sampling have been derived. Using the model, it has been shown how 0.5-thresholded optimal solutions to cross-entropy and soft-Dice are affected by various levels of noise, and how the resulting segmentations diverge as the noise is increased. It has further been shown how optimal solutions to soft-Dice can be recovered by thresholding the optimal solutions to cross-entropy with an a priori unknown but efficiently computable threshold generally different from 0.5. Finally, the theoretical observations are verified on data from the TotalSegmentor data set [37]. The presented results give insights into expected behaviour of widely used segmentation methods and suggest that cross-entropy with the alternative threshold yield the benefits of cross-entropy without sacrificing the performance of soft-Dice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.774, + 0.909, + 0.849 + ], + "angle": 0, + "content": "Limitations. To be able to systematically study the effect of label noise with varying strengths, a noise model is used for experiments. The scope of the experiments is consequently limited to the kind of noise that is produced by the model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.868, + 0.909, + 0.898 + ], + "angle": 0, + "content": "Acknowledgements The authors were supported by Ray-Search Laboratories AB." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20827" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.211 + ], + "angle": 0, + "content": "[1] Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): A Completed Reference Database of Lung Nodules on CT Scans. Medical Physics, 38(2):915–931, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.212, + 0.485, + 0.307 + ], + "angle": 0, + "content": "[2] Jeroen Bertels, Tom Eelbode, Maxim Berman, Dirk Vandermeulen, Frederik Maes, Raf Bisschops, and Matthew B Blaschko. Optimizing the Dice Score and Jaccard Index for Medical Image Segmentation: Theory and Practice. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 92-100. Springer, 2019. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.308, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[3] Jeroen Bertels, David Robben, Dirk Vandermeulen, and Paul Suetens. Theoretical Analysis and Experimental Validation of Volume Bias of Soft Dice Optimized Segmentation Maps in the Context of Inherent Uncertainty. Medical Image Analysis, 67:101833, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[4] Noah Bice, Neil Kirby, Ruiqi Li, Dan Nguyen, Tyler Bahr, Christopher Kabat, Pamela Myers, Niko Papanikolaou, and Mohammad Fakhreddine. A sensitivity analysis of probability maps in deep-learning-based anatomical segmentation. Journal of applied clinical medical physics, 22(8):105-119, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.484, + 0.511 + ], + "angle": 0, + "content": "[5] Pete Bridge, Andrew Fielding, Pamela Rowntree, and Andrew Pullar. Intraobserver Variability: Should We Worry? Journal of Medical Imaging and Radiation Sciences, 47(3):217-220, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[6] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation From Sparse Annotation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 424–432. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[7] Getao Du, Xu Cao, Jimin Liang, Xueli Chen, and Yonghua Zhan. Medical image segmentation based on u-net: A review. Journal of Imaging Science and Technology, 64:1-12, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[8] Charley Gros, Andreanne Lemay, and Julien Cohen-Adad. SoftSeg: Advantages of Soft Versus Binary Training for Image Segmentation. Medical Image Analysis, 71:102038, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.701, + 0.484, + 0.755 + ], + "angle": 0, + "content": "[9] Yun Gu, Mali Shen, Jie Yang, and Guang-Zhong Yang. Reliable label-efficient learning for biomedical image recognition. IEEE Transactions on Biomedical Engineering, 66(9):2423-2432, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.755, + 0.484, + 0.824 + ], + "angle": 0, + "content": "[10] Nicholas Heller, Joshua Dean, and Nikolaos Papanikolopoulos. Imperfect segmentation labels: How much do they matter? In Intravascular Imaging and Computer Assisted Stenting and Large-Scale Annotation of Biomedical Data and Expert Label Synthesis, pages 112-120. Springer, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.824, + 0.484, + 0.85 + ], + "angle": 0, + "content": "[11] Takeyuki Hida and Masuyuki Hitsuda. Gaussian processes, volume 120. American Mathematical Soc., 1993. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.85, + 0.484, + 0.89 + ], + "angle": 0, + "content": "[12] Burhan Rashid Hussein, Cédric Meurée, Malo Gaubert, Arthur Masson, Anne Kerbrat, Benoit Combès, and Francesca Galassi. A study on loss functions and decision thresholds" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "for the segmentation of multiple sclerosis lesions on spinal cord mri. hal preprint hal-03865212, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.908, + 0.174 + ], + "angle": 0, + "content": "[13] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.174, + 0.907, + 0.228 + ], + "angle": 0, + "content": "[14] Davood Karimi, Haoran Dou, Simon K Warfield, and Ali Gholipour. Deep Learning With Noisy Labels: Exploring Techniques and Remedies in Medical Image Analysis. Medical Image Analysis, 65:101759, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.228, + 0.908, + 0.295 + ], + "angle": 0, + "content": "[15] Eytan Kats, Jacob Goldberger, and Hayit Greenspan. Soft Labeling by Distilling Anatomical Knowledge for Improved MS Lesion Segmentation. In 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), pages 1563-1566. IEEE, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.907, + 0.337 + ], + "angle": 0, + "content": "[16] Matthieu Lé, Jan Unkelbach, Nicholas Ayache, and Hervé Delingette. Sampling image segmentations for uncertainty quantification. Medical image analysis, 34:42-51, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.337, + 0.908, + 0.39 + ], + "angle": 0, + "content": "[17] Andreanne Lemay, Charley Gros, and Julien Cohen-Adad. Label Fusion and Training Methods for Reliable Representation of Inter-Rater Uncertainty. arXiv preprint arXiv:2202.07550, 2022.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[18] Hang Li, Dong Wei, Shilei Cao, Kai Ma, Liansheng Wang, and Yefeng Zheng. Superpixel-Guided Label Softening for Medical Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 227-237. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.908, + 0.527 + ], + "angle": 0, + "content": "[19] Zachary C Lipton, Charles Elkan, and Balakrishnan Naryanaswamy. Optimal Thresholding of Classifiers to Maximize F1 Measure. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 225-239. Springer, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.907, + 0.58 + ], + "angle": 0, + "content": "[20] Bingyuan Liu, Jose Dolz, Adrian Galdran, Riadh Kobbi, and Ismail Ben Ayed. Do we really need dice? the hidden region-size biases of segmentation losses. Medical Image Analysis, 91:103015, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.907, + 0.648 + ], + "angle": 0, + "content": "[21] Alireza Mehrtash, William M Wells, Clare M Tempany, Purang Abolmaesumi, and Tina Kapur. Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging, 39(12):3868-3878, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.907, + 0.717 + ], + "angle": 0, + "content": "[22] Zahra Mirikharaji, Yiqi Yan, and Ghassan Hamarneh. Learning to Segment Skin Lesions From Noisy Annotations. In Domain adaptation and representation transfer and medical image learning with less labels and imperfect data, pages 207-215. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.908, + 0.798 + ], + "angle": 0, + "content": "[23] Guy Nir, Soheil Hor, Davood Karimi, Ladan Fazli, Brian F Skinnider, Peyman Tavassoli, Dmitry Turbin, Carlos F Villamil, Gang Wang, R Storey Wilson, et al. Automatic Grading of Prostate Cancer in Digitized Histopathology Images: Learning From Multiple Experts. Medical Image Analysis, 50:167-180, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.799, + 0.908, + 0.867 + ], + "angle": 0, + "content": "[24] Marcus Nordström, Han Bao, Fredrik Lofman, Henrik Hult, Atsuto Maki, and Masashi Sugiyama. Calibrated Surrogate Maximization of Dice. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 269-278. Springer, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.867, + 0.907, + 0.893 + ], + "angle": 0, + "content": "[25] Marcus Nordstrom, Henrik Hult, Fredrik Lofman, and Jonas Söderberg. On image segmentation with noisy labels: Char" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20828" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "acterization and volume properties of the optimal solutions to accuracy and dice. Advances in Neural Information Processing Systems, 35:34321-34333, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.174 + ], + "angle": 0, + "content": "[26] Marcus Nordstrom, Henrik Hult, Atsuto Maki, and Fredrik Lofman. Noisy image segmentation with soft-dice. arXiv preprint arXiv, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.174, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[27] Tufve Nyholm, Stina Svensson, Sebastian Andersson, Joakim Jonsson, Maja Sohlin, Christian Gustafsson, Elisabeth Kjellén, Karin Söderström, Per Albertsson, Lennart Blomqvist, et al. MR and CT Data With Multiobserver Delineations of Organs in the Pelvic Area—Part of the Gold Atlas Project. Medical Physics, 45(3):1295-1300, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.255, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[28] Teodora Popordanoska, Jeroen Bertels, Dirk Vandermeulen, Frederik Maes, and Matthew B Blaschko. On the Relationship Between Calibrated Predictors and Unbiased Volume Estimation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 678-688. Springer, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.338, + 0.484, + 0.405 + ], + "angle": 0, + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional Networks for Biomedical Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241. Springer, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[30] Axel-Jan Rousseau, Thijs Becker, Jeroen Bertels, Matthew B Blaschko, and Dirk Valkenborg. Post Training Uncertainty Calibration of Deep Networks for Medical Image Segmentation. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1052-1056. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.473, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[31] Gennady Samoradnitsky and Murad S. Taqqu. Stable non-Gaussian random processes: stochastic models with infinite variance. Chapman and Hall, 1994. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.514, + 0.484, + 0.568 + ], + "angle": 0, + "content": "[32] Nahian Siddique, Sidike Paheding, Colin P Elkin, and Vijay Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. *IEEE Access*, 9:82031-82057, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.484, + 0.609 + ], + "angle": 0, + "content": "[33] Joao Lourenço Silva and Arlindo L Oliveira. Using Soft Labels to Model Uncertainty in Medical Image Segmentation. arXiv preprint arXiv:2109.12622, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[34] Nima Tajbakhsh, Laura Jeyaseelan, Qian Li, Jeffrey N Chiang, Zhihao Wu, and Xiaowei Ding. Embracing Imperfect Datasets: A Review of Deep Learning Solutions for Medical Image Segmentation. Medical Image Analysis, 63:101693, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[35] Eugene Vorontsov and Samuel Kadoury. Label noise in segmentation networks: mitigation must deal with bias. In Deep Generative Models, and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 1, pages 251-258. Springer, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.484, + 0.84 + ], + "angle": 0, + "content": "[36] Zifu Wang, Teodora Popordanoska, Jeroen Bertels, Robin Lemmens, and Matthew B Blaschko. Dice semimetric losses: Optimizing the dice score with soft labels. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 475-485. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.84, + 0.484, + 0.895 + ], + "angle": 0, + "content": "[37] Jakob Wasserthal, Manfred Meyer, Hanns-Christian Breit, Joshy Cyriac, Shan Yang, and Martin Segeroth. Totalsegmentator: robust segmentation of 104 anatomical structures in ct images. arXiv preprint arXiv:2208.05868, 2022. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[38] Shaode Yu, Mingli Chen, Erlei Zhang, Junjie Wu, Hang Yu, Zi Yang, Lin Ma, Xuejun Gu, and Weiguo Lu. Robustness Study of Noisy Annotation in Deep Learning Based Medical Image Segmentation. Physics in Medicine & Biology, 65(17):175007, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.161, + 0.908, + 0.23 + ], + "angle": 0, + "content": "[39] Ming-Jie Zhao, Narayanan Edakunni, Adam Pocock, and Gavin Brown. Beyond Fano's Inequality: Bounds on the Optimal F-score, BER, and Cost-Sensitive Risk and Their Implications. The Journal of Machine Learning Research, 14(1):1033-1090, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20829" + } + ] +] \ No newline at end of file diff --git a/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_origin.pdf b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1f83aeda4afc2622dbeb13449d02f54537c84db5 --- /dev/null +++ b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/a28916d9-88e3-4c4a-afb0-99e422d85d63_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:362b3a5e392a233175431a69305b33e677179edd5d6f0bc1552baa9fb79855a0 +size 1727730 diff --git a/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/full.md b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b0a219c3e4655fd965de07668479de532510530 --- /dev/null +++ b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/full.md @@ -0,0 +1,368 @@ +# The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation + +Marcus Nordström1,2 Atsuto Maki1 Henrik Hult1 +1KTH Royal Institute of Technology 2RaySearch Laboratories {marcno,atsuto,hult}@kth.se + +# Abstract + +In image segmentation and specifically in medical image segmentation, the soft-Dice loss is often chosen instead of the more traditional cross-entropy loss to improve performance with respect to the Dice metric. Experimental work supporting this claim exists, but how and why the two loss functions lead to different predictions is not well understood. This paper explains the observed discrepancy as a consequence of how those loss functions are effected by label noise and what threshold is used to convert the predicted soft segmentation to predicted labels. In particular, it is shown (i) how the optimal solutions to the two loss functions diverge as the noise is increased, and (ii) how the optimal solutions to soft-Dice can be recovered by thresholding the solutions to cross-entropy with an a priori unknown but efficiently computable threshold. The theoretical results are supported by numerical experiments and it is concluded that cross-entropy with the alternative threshold yields the stability and informative label probability maps associated with cross-entropy without sacrificing the performance of soft-Dice. + +# 1. Introduction + +In medical image analysis it is often of interest to partition an image into foreground and background with respect to some region of interest. This problem is referred to as the medical image segmentation problem. Since the introduction of the U-net [29], state of the art methods for addressing the problem have been dominated by neural network methods trained according to the supervised learning paradigm [7, 32]. These methods are remarkably effective in scenarios with abundant and clean data. However, manually delineating a region of interest accurately and reliably can be a complex and time-consuming process that often requires medical expertise. This result in high costs associated with obtaining high-quality data sets for training, and consequently, that data used in practice often is imperfect in the sense that the manually provided labels are contaminated by noise [1, 5, 23, 27]. + +A large amount of research has been devoted to medical image segmentation with imperfect data. One important question that has been addressed is how experimentally robust various segmentation methods are to label noise. In such work, experiments are conducted by first obtaining some data set and then corrupting the portion of data used for training with various types of noise. Different models are then trained on the resulting data sets and evaluated on uncorrupted data. If both noisy labels and noise-free labels can be obtained from clinical practitioners, it is possible to corrupt a clean label simply by swapping it with the corresponding noisy label. Such a study of label noise was presented in [38]. An alternative is to start with a data set that is assumed to be noise-free and generate noisy data with a noise model. Simple noise models that have been studied include randomly flipping labels in an independent manner or randomly translating the target structure [10]. Models that yield more realistic looking segmentation noise include randomly perturbing some points of the boundary on a noise-free segmentation with Gaussian variables [10] and randomly deforming a segmentation with an interpolated low-dimensional Gaussian array [35]. + +How label noise affects the performance of segmentation models is highly dependent on the loss function that is used during training. Arguably, the most classical loss function considered in the field is the cross-entropy loss. This loss is known to have a lot of nice theoretical properties and to be stable in practice. However, because of the class imbalance associated with the foreground typically being significantly smaller than the background, other loss functions putting more emphasis on foreground are often considered. The most popular such loss is a smoothed version of the Dice metric. This loss is referred to as soft-Dice or the Dice loss, and has been shown to yield better performance than cross-entropy in several experimental settings at the cost of sometimes introducing some stability issues [2, 24]. To get some of the improved performance while limiting potential stability issues, it is common to use linear combinations of the losses. This is for instance the type of loss used in the popular nnU-net [13]. + +Contribution. This work introduces a noise model based on Gaussian field deformations. The model is shown to be approximately volume preserving in the sense that the expected volume of the resulting noisy segmentation is approximately equal to the volume of the corresponding noise-free segmentation. Moreover, a closed form expression for the marginal probabilities and a representation of random deformations using stochastic integrals are provided. This allows the behaviour of the model to be clearly interpreted and efficiently simulated. With the noise model, optimal solutions to cross-entropy and soft-Dice are studied. It is shown how the optimal solutions to soft-Dice and cross-entropy coincide when the labels are noise-free but diverge when noise is increased. It is further shown how optimal solutions to soft-Dice can be recovered by thresholding solutions to cross-entropy with a particular a priori unknown but efficiently computable threshold. Finally, the presented theory is validated experimentally on three different organ segmentation problems from the publicly available TotalSegmentor data set [37] with noise introduced by the model. + +# 2. Related Work + +Label noise is a well known issue in medical image segmentation and the impact of observer variability has been discussed in various domains for different modalities. This include for radiotherapy [5], for lung nodules on clinical thoracic CT images [1] and for organs in the pelvic region for CT and MR [27]. Examples of work on methods targeting these issues include active learning methods, where experts are iteratively probed for more information [9] and re-weighting schemes where methods dynamically learn to put different weights on different regions depending on certainty [22]. For fairly recent reviews on image segmentation with imperfect data in general, see [14, 34]. + +The robustness of medical image segmentation to noise can be studied by training models on data contaminated with varying degrees of noise. In [38], both noise-free and noisy labels are collected and the noise level is determined by the fraction of the training data taken from the noisy data set. Of special importance to this work is segmentation noise provided by Gaussian field deformations. Two works have studied such noise in an experimental setting [10, 35]. In these works, the segmentations in the training data is perturbed according to a low dimensional Gaussian field interpolated to the pixel space. Models are then trained for various noise levels and results obtained are reported. Finally, in [16], a noise model for target delineation in radiotherapy is discussed. The model is based on taking the level set of Gaussian field and efficient methods for sampling is provided. + +The influence label noise has on various loss functions has been studied in several papers. This include the influence it has on calibration and volume bias for predictions generated + +with models trained using cross-entropy and soft-Dice [3, 20], re-calibration methods for improving performance of soft-Dice [30] and the relationship between calibration and volume bias [28]. Other experimental work showing that the expected calibration error is lower for cross-entropy than for soft Dice includes [21]. In the above, noise has been analyzed by considering the expected score with respect to some noisy segmentation. An alternative way to incorporate noise is to make use of soft labels. Examples of work using soft labels with soft-Dice include [8, 15]. Examples of work on soft labels with other loss functions include [17, 18, 33]. In [36] a modified version of soft-Dice is introduced. This loss yields optimal solutions that coincide with the soft labels provided for the target structure which is known to not be the case for the classical soft-Dice loss. + +Inspired by threshold classifiers studied in [19, 39], optimal segmentations to Accuracy and Dice under label noise was analyzed in [25]. A characterization of optimal segmentations was provided, sharp bounds with respect to volume bias of the optimal segmentations where shown and the volume of the optimizers associated with the metrics was compared. A similar analysis was later provided for the loss function soft-Dice [26]. In particular, a characterization of the optimal solutions and the associated volumes are provided. Moreover, a calibration result showing that a sequence that converge to optimal soft-Dice, when thresholded appropriately, converges to optimal Dice. Experimental work comparing various thresholds for cross-entropy was presented in [4, 12]. The results from these papers experimentally show that using 0.5 as a common threshold leads to sub-optimal segmentations for various loss functions with respect to Dice. + +# 3. Preliminaries + +# 3.1. Notation + +The notation largely follows [25, 26]. Let $\Omega = [0,1]^n\subset \mathbb{R}^n$ be the unit cube of dimension $n\geq 1$ and $\lambda$ be the standard Lebesgue measure associated with $\mathbb{R}^n$ such that $\lambda (\Omega) = 1$ . Let $S$ be the space of measurable functions from $\Omega$ to $\{0,1\}$ , $\mathcal{M}$ be the space of measurable functions from $\Omega$ to $[0,1]$ and $\mathcal{F}$ be the space of bounded measurable functions from $\Omega$ to $\mathbb{R}$ , where all of the spaces $S$ , $\mathcal{M}$ and $\mathcal{F}$ are equipped with their associated Borel $\sigma$ -fields and the $L_{p}$ -norms $\| \cdot \| _p$ . The letter $s\in S$ is generally used to denote a segmentation, where $s(\omega) = 1$ indicates foreground of the target structure of interest and $s(\omega) = 0$ indicates background. The letter $m\in \mathcal{M}$ is generally used to denote a marginal function, that is, $m(\omega)\in [0,1]$ indicates the probability of a noisy label occupying site $\omega \in \Omega$ . The letter $c\in \mathcal{M}$ is generally used to denote the soft segmentation that is obtained by optimizing the chosen loss function prior to any thresholding. The letter $f\in \mathcal{F}$ is generally used to denote a logit function usually + +associated with a soft segmentation such that $\sigma \circ f(\omega) = c(\omega)$ for $\omega \in \Omega$ a.e., where $\sigma(x) \doteq 1/(1 + e^{-x})$ is the standard sigmoid function. Composition of functions are denoted by $\circ$ and $I_A(x)$ is used to denote the indicator function over the set $A$ , that is $I_A(x) = 1$ if $x \in A$ and $I_A(x) = 0$ if $x \notin A$ . The indicator function is used in conjunction with some element $c \in \mathcal{M}$ to denote thresholding of a soft segmentation. That is, $I_{[t,1]} \circ c$ assigns 1 to all $\omega \in \Omega$ where $c(\omega) \geq t$ and 0 to all $\omega \in \Omega$ where $c(\omega) < t$ . + +# 3.2. Loss Functions and Evaluation Metrics + +Two of the most popular loss functions in the field of medical image segmentation are cross-entropy and soft-Dice. Usually these loss functions are defined over two deterministic arguments, a soft-segmentation (prediction) and a binary segmentation (label). Under the influence of label noise, these loss functions are extended to functions taking one deterministic soft segmentation (prediction) and one random segmentation (label) as arguments. In this work, the soft labeling extension is adopted for the theoretical analysis [8, 15, 17, 18, 33] (see Section 4.4 for more on this). Formally, the loss functions are defined with respect to a marginal function $m \in \mathcal{M}$ which corresponds to the mean of a noisy segmentation $L$ taking values in $S$ , that is $m(\omega) = \mathbb{E}[L(\omega)], \omega \in \Omega$ . + +Definition 1. For any $m \in \mathcal{M}$ , (binary) cross-entropy is given by + +$$ +\begin{array}{l} \mathrm {C E} _ {m} (c) \doteq - \int_ {\Omega} [ m (\omega) \log (c (\omega)) + \quad c \in \mathcal {M}. \tag {1} \\ (1 - m (\omega)) \log (1 - c (\omega)) ] \lambda (d \omega), \\ \end{array} +$$ + +Definition 2. For any $m \in \mathcal{M}$ , soft-Dice is given by + +$$ +\mathrm {S D} _ {m} (c) \doteq 1 - \frac {2 \int_ {\Omega} c (\omega) m (\omega) \lambda (d \omega)}{\| c \| _ {1} + \| m \| _ {1}}, \quad c \in \mathcal {M}. \tag {2} +$$ + +Arguably, the most classical evaluation metric used is Accuracy. This metric simply measures the fraction of the domain that is correctly labeled. The most popular evaluation metric on the other hand is the Sorensen-Dice coefficient, or Dice for short. To handle label noise, these metrics are defined in the same way as the loss functions, that is, by using soft labels. + +Definition 3. For any $m \in \mathcal{M}$ , Accuracy is given by + +$$ +\begin{array}{l} \mathrm {A} _ {m} (s) \doteq \int_ {\Omega} [ s (\omega) m (\omega) + \quad s \in \mathcal {S}. \tag {3} \\ (1 - s (\omega)) (1 - m (\omega)) ] \lambda (d \omega), \\ \end{array} +$$ + +Definition 4. For any $m \in \mathcal{M}$ , Dice is given by + +$$ +\mathrm {D} _ {m} (s) \doteq \frac {2 \int_ {\Omega} s (\omega) m (\omega) \lambda (d \omega)}{\| s \| _ {1} + \| m \| _ {1}}, \quad s \in \mathcal {S}. \tag {4} +$$ + +When a label is free from noise, any reasonable loss function or metric is optimized uniquely by the target label. When there is noise present on the other hand, it is no longer clear what the solution should be and loss functions or metrics implicitly assign them. As the consequence, it is important to investigate what the optimal solutions are for a chosen loss or metric. For the case of the cross-entropy loss, it is well known that the optimizers are uniquely given by the marginal function, that is + +$$ +\mathrm {C E} _ {m} (c) = \inf _ {c ^ {\prime} \in \mathcal {M}} \mathrm {C E} _ {m} \left(c ^ {\prime}\right) \Longleftrightarrow c (\omega) = m (\omega), \tag {5} +$$ + +$\lambda$ -a.e. In other words, minimizing cross-entropy yields estimates of the marginal probability distribution associated with the labels. For soft-Dice it was shown in [26] that + +$$ +\mathrm {S D} (c) = \inf _ {c ^ {\prime} \in \mathcal {M}} \mathrm {S D} \left(c ^ {\prime}\right) \Longleftrightarrow +$$ + +$$ +c (\omega) \in \left\{ \begin{array}{l l} \{1 \} & \text {i f} m (\omega) > \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {D} _ {m} \left(s ^ {\prime}\right) / 2, \\ [ 0, 1 ] & \text {i f} m (\omega) = \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {D} _ {m} \left(s ^ {\prime}\right) / 2, \\ \{0 \} & \text {i f} m (\omega) < \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {D} _ {m} \left(s ^ {\prime}\right) / 2, \end{array} \right. \tag {6} +$$ + +$\lambda$ -a.e. Characterizations of the optimal segmentations for Accuracy and Dice using the setup adopted in this paper can be found in [25]. For Accuracy, the solutions are given by + +$$ +\mathrm {A} _ {m} (s) = \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {A} _ {m} \left(s ^ {\prime}\right) \iff +$$ + +$$ +s (\omega) \in \left\{ \begin{array}{l l} \{1 \} & \text {i f} m (\omega) > 1 / 2, \\ \{0, 1 \} & \text {i f} m (\omega) = 1 / 2, \\ \{0 \} & \text {i f} m (\omega) < 1 / 2, \end{array} \right. \tag {7} +$$ + +$\lambda$ -a.e. Similarly, for Dice the optimal segmentations are given by + +$$ +\mathrm {D} _ {m} (s) = \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {D} _ {m} \left(s ^ {\prime}\right) \Longleftrightarrow \tag {8} +$$ + +$$ +s (\omega) \in \left\{ \begin{array}{l l} \{1 \} & \text {i f} m (\omega) > \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {D} _ {m} (s ^ {\prime}) / 2, \\ \{0, 1 \} & \text {i f} m (\omega) = \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {D} _ {m} (s ^ {\prime}) / 2, \\ \{0 \} & \text {i f} m (\omega) < \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {D} _ {m} (s ^ {\prime}) / 2, \end{array} \right. \tag {9} +$$ + +$\lambda$ -a.e. The resulting soft segmentation obtained by optimizing a loss function is typically converted to a binary segmentation by thresholding with 0.5. Ideally this resulting segmentation maximizes a desired target metric. Using the above characterizations, for cross-entropy it follows that + +$$ +\mathrm {C E} _ {m} (c) = \inf _ {c ^ {\prime} \in \mathcal {M}} \mathrm {C E} \left(c ^ {\prime}\right) \Longrightarrow +$$ + +$$ +\mathrm {A} _ {m} \left(I _ {[ 1 / 2, 1 ]} \circ c\right) = \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {A} \left(s ^ {\prime}\right). \tag {10} +$$ + +Similarly, for soft-Dice it follows that + +$$ +\mathrm {S D} _ {m} (c) = \inf _ {c ^ {\prime} \in \mathcal {M}} \mathrm {S D} \left(c ^ {\prime}\right) \Longrightarrow +$$ + +$$ +\mathrm {D} _ {m} \left(I _ {[ 1 / 2, 1 ]} \circ c\right) = \sup _ {s ^ {\prime} \in S} \mathrm {D} \left(s ^ {\prime}\right). \tag {11} +$$ + +# 3.3. Gaussian Fields and Independently Scattered Gaussian Measures + +Amongst the most central objects in probability theory and statistics are Gaussian variables taking values in Euclidean space. Such a variable, when centered and isotropic, has the probability density function + +$$ +p _ {\sigma^ {2}} (\omega) \dot {=} \frac {1}{(2 \pi \sigma^ {2}) ^ {n / 2}} \exp \left\{- \frac {\| \omega \| _ {2} ^ {2}}{2 \sigma^ {2}} \right\}, \quad \omega \in \mathbb {R} ^ {n}, \tag {12} +$$ + +where $\sigma^2$ denotes the variance in each of the dimensions. Various generalizations of this random variable have been studied extensively in the literature. For the theoretical treatment of the noise model in this work, two such generalizations are introduced. The first is a Gaussian field. + +Definition 5. A random variable $Y(\omega), \omega \in \mathbb{R}^n$ is a Gaussian field with kernel function $k$ if any finite collection of elements $\{\omega_1, \dots, \omega_M\} \subset \mathbb{R}^n$ satisfies $(Y(\omega_1), \dots, Y(\omega_M)) \sim \text{normal}(0, \Sigma)$ , with $\Sigma_{m,m'} = k(\omega_m, \omega_{m'})$ for $m, m' \in \{1, \dots, M\}$ . + +Many types of kernel functions with different properties have been studied in the literature. Amongst the most common is the squared exponential kernel given by + +$$ +k _ {a, b} \left(\omega , \omega^ {\prime}\right) \doteq a ^ {2} \exp \left(- \frac {\| \omega - \omega^ {\prime} \| _ {2} ^ {2}}{2 b ^ {2}}\right), \quad \omega , \omega^ {\prime} \in \mathbb {R} ^ {n}. \tag {13} +$$ + +An important property of Gaussian fields with such covariance functions are that samples are almost surely continuous. In other words, a randomly sampled function will be continuous with probability one. For a good reference on Gaussian variables, see [11]. + +The second generalization that is introduced is what is referred to as independently scattered Gaussian measures. Such constructions are used to formalize integration over Gaussian fields, similar to how the Ito-integral build on Brownian motion. + +Definition 6. A random variable $W$ is an independently scattered Gaussian measure on $\mathbb{R}^n$ with respect to the control measure $\lambda$ , if for any collection of disjoint sets $A_1, \ldots, A_k$ in $\lambda$ -measurable sets with finite measure $\mathcal{B}_0$ , $W(A_1), \ldots, W(A_k)$ are independent, and $W(A_i)$ has centered Gaussian distribution with variance $\lambda(A_i)$ . + +Importantly, for a function $f \in L^{2}(\mathbb{R}^{n})$ , the stochastic integral $\int_{\mathbb{R}^n} f(\omega) W(d\omega)$ is well defined and the process $Y = \{Y(\omega), \omega \in \mathbb{R}^n\}$ given by + +$$ +Y (\omega) \doteq \int_ {\mathbb {R} ^ {n}} f \left(\omega - \omega^ {\prime}\right) W \left(d \omega^ {\prime}\right), \tag {14} +$$ + +is a centered Gaussian process. Furthermore, this Gaussian process has a covariance kernel given by + +$$ +k \left(\omega , \omega^ {\prime}\right) = \int_ {\mathbb {R} ^ {n}} f (\omega - u) f \left(\omega^ {\prime} - u\right) \lambda (d u), \quad \omega , \omega^ {\prime} \in \mathbb {R} ^ {n}. \tag {15} +$$ + +Independently scattered measures are introduced more generally for $\alpha$ -stable distributions in [31, Section 3.3]. The Gaussian case used in this work is simply the special case when $\alpha = 2$ . + +# 4. Theory + +# 4.1. Label Noise Model + +In contrast to binary classification, where a lot of label noise can be modeled by simple independent flipping of binary variables, label noise in segmentation can in general have very complicated dependency structures. Often labels affected by noise have the appearance of being drawn by someone with shaky hands. For such noise, the correlation between border points will depend on the distance between them. That is, if two points are close to each other, then the correlation will be high and vice versa. + +One way to model segmentations that look as if they have been drawn by shaky hands is by means of random Gaussian field deformations. Consider a variable + +$$ +X \doteq \left(X _ {1}, \dots , X _ {n}\right) ^ {T} \tag {16} +$$ + +such that $\{X_i\}_{i=1}^n$ are independent Gaussian fields with kernel function (13). This means that $X$ is a random vector field taking values in $\mathbb{R}^n \mapsto \mathbb{R}^n$ . With such a random vector field, a random segmentation can be formed by + +$$ +L (\omega) \doteq \left\{ \begin{array}{l l} l (\omega + X (\omega)) & \text {i f} \omega + X (\omega) \in \Omega , \\ 0 & \text {i f} \omega + X (\omega) \notin \Omega . \end{array} \right., \quad \omega \in \Omega , \tag {17} +$$ + +where $l \in S$ is a noise-free reference segmentation being randomly deformed. How samples of the random segmentations look depend on what parameters are chosen when specifying the kernel function (13). The parameter $a$ controls how large the deformations are, where greater $a$ leads to larger deformations and vice versa. The parameter $b$ controls how smooth the deformations are, where greater $b$ leads to smoother deformations and vice versa. Modeling noisy segmentations with (17) is convenient because it allows for the derivation of several theoretical properties. In this work, three such important properties are derived. + +In Proposition 1, it is shown that the marginal probabilities associated with a noisy segmentation can be computed by convolving the associated noise-free segmentation label with an appropriate Gaussian density. + +Proposition 1. Let $l$ and $L$ be defined as in (17) and $p_{\sigma^2}$ be defined as in (12), then + +$$ +\mathbb {E} [ L (\omega) ] = \int_ {\Omega} l \left(\omega^ {\prime}\right) p _ {a ^ {2}} \left(\omega - \omega^ {\prime}\right) \lambda \left(d \omega^ {\prime}\right), \quad \omega \in \Omega . \tag {18} +$$ + +In Proposition 2, the marginal representation is used to show that the volume of the noisy segmentation is approximately unbiased in the sense that the expected volume of a noisy segmentation is approximately equal to the volume of the corresponding noise-free segmentation. The reason why this does not hold exactly is that there is some probability for the target structure to deform in such a way that parts of the structure leave the domain $\Omega$ . However, for most reasonable choices of parameters, when the regions of interest is sufficiently close to the center, the probability of this happening is negligible. + +Proposition 2. Let $l$ and $L$ be defined as in (17) and $p_{\sigma^2}$ be defined as in 12, then + +$$ +\mathbb {E} [ \| L \| _ {1} ] = \| l \| _ {1} - \xi , \tag {19} +$$ + +where + +$$ +\xi = \int_ {\Omega} \int_ {\mathbb {R} ^ {n} \backslash \Omega} l \left(\omega^ {\prime}\right) p _ {a ^ {2}} \left(\omega - \omega^ {\prime}\right) \lambda (d \omega) \lambda \left(d \omega^ {\prime}\right). \tag {20} +$$ + +In Proposition 3, it is shown that $X$ (16) is equal in distribution to a particular vector of stochastic integrals. In the case $n = 1$ , this could have been done using regular Ito-integrals, but for the case $n > 1$ a multivariate counter part is needed. For this scattered Gaussian measures are used. + +Proposition 3. Let $X$ be as in (16), $p_{\alpha^2}$ be as in (12) and $W_1, \ldots, W_n$ be independent copies of independently scattered Gaussian measures on $\mathbb{R}^n$ with control measure $\lambda$ , then it follows that + +$$ +X (\omega) \stackrel {d} {=} \left( \begin{array}{c} a (2 \pi b ^ {2}) ^ {n / 4} \int_ {\mathbb {R} ^ {n}} p _ {b ^ {2} / 2} \left(\omega - \omega^ {\prime}\right) W _ {1} \left(d w ^ {\prime}\right) \\ \vdots \\ a (2 \pi b ^ {2}) ^ {n / 4} \int_ {\mathbb {R} ^ {n}} p _ {b ^ {2} / 2} \left(\omega - \omega^ {\prime}\right) W _ {n} \left(d w ^ {\prime}\right) \end{array} \right). \tag {21} +$$ + +The presented propositions provide meaningful insights into the behaviour of the noise model. The importance of Proposition 1 is that it makes the relationship between the noise model parameters and the resulting marginal probabilities clear. Also, it provides an exact method for computing the marginals that otherwise would have to be estimated with sampling. The importance of Proposition 2 is that the noise beside from looking reasonable also numerically behave reasonable in an important manner. Finally, the importance + +of Proposition 3 is that noise for discretized domains can be sampled by smoothing arrays of independently sampled Gaussian variables with an appropriate Gaussian filters. Both operations which scale well with image size. This allows for efficient sampling of noisy segmentations for images composed of millions of voxels. + +# 4.2. Optimal Segmentations Under Label Noise + +Since the loss functions and metrics investigated in this work only depend on the marginal probabilities and Proposition 1 provides an explicit expression for these marginal probabilities, it is possible to study how noise model parameters affect optimal segmentations. Figure 1 depicts this relationship for a two dimensional segmentation problem with varying noise intensity. In the first row, pixel-wise averages of 5 samples are illustrated. In the second row, the marginal probabilities are illustrated. In the third row, theoretically optimal solutions to Accuracy are shown. Note that by (10), this is equivalently given by the minimizer to cross-entropy thresholded by 0.5. In the forth row, theoretical optimal solutions to Dice are shown. Note that by (11), this is equivalently given by the minimizer to soft-Dice thresholded by 0.5. + +A first observation is that noise affects the optimal segmentations in such a way that the shape of the noise-free reference segmentation is changed. This is an important observation that contradicts the picture that unbiased noise (unbiased in the sense that the expected deformation is the identity function) yield unbiased optimal segmentations. In particular, optimal segmentations are changed in such a way that the corners of the shapes are smoothed and details such as thin structures and holes disappear. Informally, the higher the noise, the more ball shaped the optimal segmentations. + +A second observation is that the two metrics yield similar looking segmentations for low noise but increasingly different segmentations as the noise increase. Since cross-entropy with the 0.5-threshold theoretically gives optimal segmentations with respect to Accuracy and soft-Dice with the 0.5-threshold theoretically gives optimal segmentations with respect to Dice, the performance of the two loss functions as measured by Dice is similar in low noise situations but increasingly better for soft-Dice as the noise is increased. Similarly, the performance of the two loss functions as measured by Accuracy is similar in low noise situations but increasingly better for cross-entropy as the noise is increased. + +# 4.3. Cross-entropy With an Alternative Threshold + +By the characterization of optimizers to cross-entropy (5) and the characterization of optimizers to soft-Dice (6), it follows that an optimizer to soft-Dice can be obtained by first optimizing cross-entropy and then thresholding with the a priori unknown value + +$$ +t _ {m} \doteq \sup _ {s ^ {\prime} \in \mathcal {S}} \mathrm {D} _ {m} \left(s ^ {\prime}\right) / 2. \tag {22} +$$ + +![](images/71238ae1e0ffc3ac440274252f170f0eb5622b6369704d29ba4a7f47f4e34172.jpg) +Figure 1. Illustration of the noise model and its effects on a two-dimensional segmentation problem, with the columns showing the effect of different noise strengths. The first row shows the pixel-wise average obtained from five samples. The second row shows the marginal function. The third row shows optimal segmentations associated with Accuracy (7), which in this and most natural cases is unique. The forth row shows optimal segmentations associated with Dice (9), which in this and most natural cases is unique. In all of the reported cases $b = 0.15 / \sqrt{2}$ . + +That is, by thresholding an optimizer to cross-entropy with this alternative recursively defined threshold $t_m$ , it follows that + +$$ +\begin{array}{l} \mathrm {C E} _ {m} (c) = \inf _ {c ^ {\prime} \in \mathcal {M}} \mathrm {C E} \left(c ^ {\prime}\right) \Longrightarrow \tag {23} \\ \mathrm {S D} _ {m} \left(I _ {[ t _ {c}, 1 ]} \circ c\right) = \inf _ {c ^ {\prime} \in \mathcal {M}} \mathrm {S D} \left(c ^ {\prime}\right). \\ \end{array} +$$ + +Methods based on this principle have not been discussed in the medical image segmentation community but are known in the statistical learning literature [19]. In the medical image segmentation community, however, there has been experimental work comparing the performance of different constant thresholds, that is, thresholds not dependent on the marginal probabilities [4, 12]. The results experimentally show that using a 0.5-threshold leads to sub-optimal segmentations for various loss functions when the Dice metric is used for evaluation. Consequently, the results are in line with the theoretical observations above and motivate further investigation of methods based on cross-entropy with the threshold (22). + +# 4.4. Soft Labeling and Expected Score + +One reason to why the threshold (22) has not been discussed in the medical image segmentation domain may be that loss functions and metrics often incorporate noise using expectations rather than soft labels. Formally for some random label $L$ taking values in $S$ with $\mathbb{E}[L(\omega)] = m(\omega), \omega \in \Omega$ and $c \in \mathcal{M}$ , this means that $\mathbb{E}[\mathrm{CE}_L(c)]$ is considered instead of $\mathrm{CE}_m(c)$ and $\mathbb{E}[\mathrm{SD}_L(c)]$ is considered instead of $\mathrm{SD}_m(c)$ . Similarly for some $s \in S$ , this means that $\mathbb{E}[\mathrm{A}_L(s)]$ is considered instead of $\mathrm{A}_m(s)$ and $\mathbb{E}[\mathrm{D}_L(s)]$ is considered instead of $\mathrm{D}_m(s)$ . + +For cross-entropy and Accuracy, it is easy to verify that the definitions using expectations and the definitions using soft labels are equivalent, that is $\mathbb{E}[\mathrm{A}_L(s)] = \mathrm{A}_m(s), s \in S$ and $\mathbb{E}[\mathrm{CE}_L(c)] = \mathrm{CE}_m(c), c \in \mathcal{M}$ . For soft-Dice and Dice, however, this is in general not the case, that is, in general $\mathbb{E}[\mathrm{D}_L(s)] \neq \mathrm{D}_m(s), s \in S$ and $\mathbb{E}[\mathrm{SD}_L(c)] \neq \mathrm{SD}_m(c), c \in \mathcal{M}$ . Hence, the property (23) does not in general transfer to properties of optimizers to soft-Dice using expectations. Importantly however, that the definitions of soft-Dice and Dice using soft labels and expectations are + +equivalent provided that the volume of the noisy label is constant, that is $\mathrm{Var}[\| L\| _1] = 0$ . For situations when the noise is approximately constant $\mathrm{Var}[\| L\| _1]\approx 0$ , the definitions are approximately equal $\mathbb{E}[\mathrm{D}_L(s)]\approx \mathrm{D}_m(s),s\in S$ and $\mathbb{E}[\mathrm{SD}_L(c)]\approx \mathrm{SD}_m(c),c\in \mathcal{M}$ [25]. Hence the property (23) holds approximately for the kind of noise studied in this work. + +# 5. Experiments + +# 5.1. Model + +A standard 3D U-net is implemented [6]. The network is composed of blocks of two convolutions with kernel size $3^3$ , instance normalization and ReLu-activations. The number of filters used is increased for each down-sampling according to 16, 32, 64, 128, 256, where 16 corresponds to the number of filters used in the input layer and 256 corresponds to the number of filters in the bottleneck layer. Convolutions with kernel $2^3$ and stride 2 are used for down-sampling. Transposed convolutions with kernel $2^3$ stride 2 are used for up-sampling. Details and code can be found in the Supplementary Material. + +# 5.2. Training + +Models are trained either with the cross-entropy loss or with the soft-Dice loss. In both cases $l_{2}$ regularization is used with a weight of $10^{-5}$ . Adam with learning rate $10^{-4}$ is used for optimization together with a scheduler that divides the learning rate by 5 after 30 epochs of non-improvement. The number of epochs computed is fixed to 100 and the order of the training data is randomly shuffled. Batch sizes of 1 is used and with probability 0.5 the data is augmented. The augmentation is composed of random shifts, random rotations and image blurring with Gaussian filters. The model that performs best on the training data is the model selected. Details and code can be found in the Supplementary Material. + +# 5.3.Data + +The experiments are conducted with respect to the TotalSegmentor data set [37]. This data set contains 1204 CT images with 104 anatomical structures (27 organs, 59 bones, 10 muscles, 8 vessels). To illustrate the effect noise may have on organs with different shape, three different organs are chosen. This includes the right kidney which in general is spherical, the aorta which is tubular and relatively thick and the esophagus which is tubular and relatively thin. For each of the organs, 400 cases is selected and split into 5 folds each containing 80 cases. The images are sub-sampled to half resolution and patches of $64^{3}$ voxels centered in each of the structures are extracted. In order to simulate label noise the described noise model is used. For each organ data set, four different datasets was created by randomly deforming the + +![](images/e5b7a1bdd2328e6729d7b379c04e17ac03fdc708fa5d17de501e2601dc41dc16.jpg) + +![](images/ef0aaf31ee8630f99f8edeaba658505db80f7d2add6eed19b95b447623758590.jpg) + +![](images/852e87ac540c4ef801d6fd2481a190a899c31187b5f54ee31a3192694809cbde.jpg) + +![](images/903fb259c84959ff8d6d857f4d978944eb21a46a914551d7fea1551c955ec9bf.jpg) + +![](images/0694241640b8c721f29c283c37aa7edbcac75c442a57030825a85477e4837dca.jpg) + +![](images/dbd44d53b5d96d07b63c9de096353bb8411d33ff2f6074f888b04e2474b11b2a.jpg) + +![](images/57a52df7f11289c6fa0d52824939709e2e64f24ced00ff73d6c99a825ddd649d.jpg) +Figure 2. Illustration of the data from the TotalSegmentor data set [37] used in the experiments. The first depicts a kidney, the second row an aorta and the third row an esophagus. The columns correspond to axial, coronal and sagittal views respectively. + +![](images/3dff81445302c5f8019c4a8b811b8de9ec39bd0394b4cc9f4b1a3687f2e21f2f.jpg) + +![](images/79a24752fe1a5b48cd01005710aebd9cec16799502a517bc3fc163b0a82ae353.jpg) + +labels with the noise strengths $a = 0.0, 0.1, 0.2, 0.3$ . Each data point is deformed once, resulting in one sample for each random label. Details and code can be found in the Supplementary Material. + +# 5.4. Results + +Results from the experiments are illustrated separately for the three organ segmentation problems as a figure in Figure 3 and listed as a table in the Supplementary Material. For each organ segmentation problem, the average Dice score for the various methods is reported as a function of the parameter controlling the noise strength in the noise model. The average score is obtained by taking the average of the results obtained after 5 folds. The notation $\mathrm{CE}^{(0)}$ is used to denote a model trained with cross-entropy and thresholded with 0.5, $\mathrm{SD}^{(0)}$ a model trained with cross-entropy and thresholded with 0.5 and $\mathrm{CE}^{(*)}$ a model trained with cross-entropy and thresholded with $t_m$ (22). + +The first thing to notice is that the difference amongst the average performance of cross-entropy and soft-Dice with the 0.5-threshold, that is, $\mathrm{CE}^{(0)}$ and $\mathrm{SD}^{(0)}$ , increase as the amount of added noise increase. When no noise is added, the performance is indistinguishable. This type of result is expected from the theory and depicted in Figure 1. + +The second thing to notice is that cross-entropy with the alternative threshold (22) $\mathrm{CE}^{(*)}$ performs better than cross-entropy with the 0.5-threshold $\mathrm{CE}^{(0)}$ . In regions with low + +![](images/4c1789cbdfe4ad40ce452d8b9e3c6758772f75a6ded02f559eb5a0854d441295.jpg) +Figure 3. To the left are the results from the kidney segmentation problem, in the middle are the results from the aorta segmentation problem and to the right are the results from the esophagus segmentation problem. In each of the figures, the average Dice value obtained from five fold experiments is plotted as a function of the noise level $a$ . The legend $\mathrm{CE}^{(0)}$ indicate that cross-entropy with the $1/2$ -threshold has been used, $\mathrm{SD}^{(0)}$ indicate that soft-Dice with the $1/2$ -threshold has been used and $\mathrm{CE}^{(*)}$ indicate that cross-entropy with the threshold described in (22) has been used. In all of the reported cases $b = 0.15/\sqrt{2}$ . A table of the results can be found in the Supplementary Material. + +![](images/4323af32b5afb573ab02198424bdd0eafc7553ea8a8a1e152eadb2d808ab8471.jpg) + +![](images/a629b5e6beb130a8bf7929068fc02c3a867ae3eaee252c727d5a7860bfa4951d.jpg) + +noise the discrepancy is not noticeable, but as the noise is increased the discrepancy also increases. The performance improvement closes the gap to soft-Dice with the 0.5-threshold $\mathrm{SD}^0$ which is expected by the theory (23). + +# 6. Discussion + +The results in this work follow the results presented by [2] in the sense that the soft-Dice loss outperforms the cross-entropy loss with respect to the Dice metric when the 0.5-threshold is used and that there seems to be a trend of bigger discrepancy when performance for both of the loss functions is low. The results also explain the experimental observations of alternative thresholds outperforming the standard 0.5-threshold that was presented in [4, 12]. + +In [24], it was proposed that the instability sometimes observed for soft-Dice may be associated with the fact that the loss function is non-convex (nor quasi-convex) when using sigmoid functions to enforce the bounds, that is $f \mapsto \mathrm{SD}_m(\sigma \circ f), f \in \mathcal{F}$ . This is in contrast to cross-entropy where the corresponding $f \mapsto \mathrm{CE}_m(\sigma \circ f), f \in \mathcal{F}$ is known to be convex. If cross-entropy with the alternative threshold yields the same performance as soft-Dice, then using it may provide the performance of soft-Dice without sacrificing the stability of cross-entropy. Another potential gain of using cross-entropy over soft-Dice is that marginal estimates are more informative than the approximate binary estimates that soft-Dice yields. These estimates can for instance be used to estimate uncertainty in different regions of a segmentation. + +A cost associated with using cross-entropy with the alternative threshold (22), is that the threshold needs to be computed for every prediction. This can be implemented ef + +ficiently but requires an extra step that may be non-negligible for sufficiently large images especially when there is a large number of predicted classes. + +# 7. Conclusion + +In this work, a realistic approximately volume preserving label noise model has been proposed and methods for explicit computation of marginal probabilities and sampling have been derived. Using the model, it has been shown how 0.5-thresholded optimal solutions to cross-entropy and soft-Dice are affected by various levels of noise, and how the resulting segmentations diverge as the noise is increased. It has further been shown how optimal solutions to soft-Dice can be recovered by thresholding the optimal solutions to cross-entropy with an a priori unknown but efficiently computable threshold generally different from 0.5. Finally, the theoretical observations are verified on data from the TotalSegmentor data set [37]. The presented results give insights into expected behaviour of widely used segmentation methods and suggest that cross-entropy with the alternative threshold yield the benefits of cross-entropy without sacrificing the performance of soft-Dice. + +Limitations. To be able to systematically study the effect of label noise with varying strengths, a noise model is used for experiments. The scope of the experiments is consequently limited to the kind of noise that is produced by the model. + +Acknowledgements The authors were supported by Ray-Search Laboratories AB. + +# References + +[1] Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): A Completed Reference Database of Lung Nodules on CT Scans. Medical Physics, 38(2):915–931, 2011. 1, 2 +[2] Jeroen Bertels, Tom Eelbode, Maxim Berman, Dirk Vandermeulen, Frederik Maes, Raf Bisschops, and Matthew B Blaschko. Optimizing the Dice Score and Jaccard Index for Medical Image Segmentation: Theory and Practice. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 92-100. Springer, 2019. 1, 8 +[3] Jeroen Bertels, David Robben, Dirk Vandermeulen, and Paul Suetens. Theoretical Analysis and Experimental Validation of Volume Bias of Soft Dice Optimized Segmentation Maps in the Context of Inherent Uncertainty. Medical Image Analysis, 67:101833, 2021. 2 +[4] Noah Bice, Neil Kirby, Ruiqi Li, Dan Nguyen, Tyler Bahr, Christopher Kabat, Pamela Myers, Niko Papanikolaou, and Mohammad Fakhreddine. A sensitivity analysis of probability maps in deep-learning-based anatomical segmentation. Journal of applied clinical medical physics, 22(8):105-119, 2021. 2, 6, 8 +[5] Pete Bridge, Andrew Fielding, Pamela Rowntree, and Andrew Pullar. Intraobserver Variability: Should We Worry? Journal of Medical Imaging and Radiation Sciences, 47(3):217-220, 2016. 1, 2 +[6] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation From Sparse Annotation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 424–432. Springer, 2016. 7 +[7] Getao Du, Xu Cao, Jimin Liang, Xueli Chen, and Yonghua Zhan. Medical image segmentation based on u-net: A review. Journal of Imaging Science and Technology, 64:1-12, 2020. 1 +[8] Charley Gros, Andreanne Lemay, and Julien Cohen-Adad. SoftSeg: Advantages of Soft Versus Binary Training for Image Segmentation. Medical Image Analysis, 71:102038, 2021. 2, 3 +[9] Yun Gu, Mali Shen, Jie Yang, and Guang-Zhong Yang. Reliable label-efficient learning for biomedical image recognition. IEEE Transactions on Biomedical Engineering, 66(9):2423-2432, 2018. 2 +[10] Nicholas Heller, Joshua Dean, and Nikolaos Papanikolopoulos. Imperfect segmentation labels: How much do they matter? In Intravascular Imaging and Computer Assisted Stenting and Large-Scale Annotation of Biomedical Data and Expert Label Synthesis, pages 112-120. Springer, 2018. 1, 2 +[11] Takeyuki Hida and Masuyuki Hitsuda. Gaussian processes, volume 120. American Mathematical Soc., 1993. 4 +[12] Burhan Rashid Hussein, Cédric Meurée, Malo Gaubert, Arthur Masson, Anne Kerbrat, Benoit Combès, and Francesca Galassi. A study on loss functions and decision thresholds + +for the segmentation of multiple sclerosis lesions on spinal cord mri. hal preprint hal-03865212, 2022. 2, 6, 8 +[13] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 1 +[14] Davood Karimi, Haoran Dou, Simon K Warfield, and Ali Gholipour. Deep Learning With Noisy Labels: Exploring Techniques and Remedies in Medical Image Analysis. Medical Image Analysis, 65:101759, 2020. 2 +[15] Eytan Kats, Jacob Goldberger, and Hayit Greenspan. Soft Labeling by Distilling Anatomical Knowledge for Improved MS Lesion Segmentation. In 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), pages 1563-1566. IEEE, 2019. 2, 3 +[16] Matthieu Lé, Jan Unkelbach, Nicholas Ayache, and Hervé Delingette. Sampling image segmentations for uncertainty quantification. Medical image analysis, 34:42-51, 2016. 2 +[17] Andreanne Lemay, Charley Gros, and Julien Cohen-Adad. Label Fusion and Training Methods for Reliable Representation of Inter-Rater Uncertainty. arXiv preprint arXiv:2202.07550, 2022.2,3 +[18] Hang Li, Dong Wei, Shilei Cao, Kai Ma, Liansheng Wang, and Yefeng Zheng. Superpixel-Guided Label Softening for Medical Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 227-237. Springer, 2020. 2, 3 +[19] Zachary C Lipton, Charles Elkan, and Balakrishnan Naryanaswamy. Optimal Thresholding of Classifiers to Maximize F1 Measure. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 225-239. Springer, 2014. 2, 6 +[20] Bingyuan Liu, Jose Dolz, Adrian Galdran, Riadh Kobbi, and Ismail Ben Ayed. Do we really need dice? the hidden region-size biases of segmentation losses. Medical Image Analysis, 91:103015, 2024. 2 +[21] Alireza Mehrtash, William M Wells, Clare M Tempany, Purang Abolmaesumi, and Tina Kapur. Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging, 39(12):3868-3878, 2020. 2 +[22] Zahra Mirikharaji, Yiqi Yan, and Ghassan Hamarneh. Learning to Segment Skin Lesions From Noisy Annotations. In Domain adaptation and representation transfer and medical image learning with less labels and imperfect data, pages 207-215. Springer, 2019. 2 +[23] Guy Nir, Soheil Hor, Davood Karimi, Ladan Fazli, Brian F Skinnider, Peyman Tavassoli, Dmitry Turbin, Carlos F Villamil, Gang Wang, R Storey Wilson, et al. Automatic Grading of Prostate Cancer in Digitized Histopathology Images: Learning From Multiple Experts. Medical Image Analysis, 50:167-180, 2018. 1 +[24] Marcus Nordström, Han Bao, Fredrik Lofman, Henrik Hult, Atsuto Maki, and Masashi Sugiyama. Calibrated Surrogate Maximization of Dice. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 269-278. Springer, 2020. 1, 8 +[25] Marcus Nordstrom, Henrik Hult, Fredrik Lofman, and Jonas Söderberg. On image segmentation with noisy labels: Char + +acterization and volume properties of the optimal solutions to accuracy and dice. Advances in Neural Information Processing Systems, 35:34321-34333, 2022. 2, 3, 7 +[26] Marcus Nordstrom, Henrik Hult, Atsuto Maki, and Fredrik Lofman. Noisy image segmentation with soft-dice. arXiv preprint arXiv, 2023. 2, 3 +[27] Tufve Nyholm, Stina Svensson, Sebastian Andersson, Joakim Jonsson, Maja Sohlin, Christian Gustafsson, Elisabeth Kjellén, Karin Söderström, Per Albertsson, Lennart Blomqvist, et al. MR and CT Data With Multiobserver Delineations of Organs in the Pelvic Area—Part of the Gold Atlas Project. Medical Physics, 45(3):1295-1300, 2018. 1, 2 +[28] Teodora Popordanoska, Jeroen Bertels, Dirk Vandermeulen, Frederik Maes, and Matthew B Blaschko. On the Relationship Between Calibrated Predictors and Unbiased Volume Estimation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 678-688. Springer, 2021. 2 +[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional Networks for Biomedical Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241. Springer, 2015. 1 +[30] Axel-Jan Rousseau, Thijs Becker, Jeroen Bertels, Matthew B Blaschko, and Dirk Valkenborg. Post Training Uncertainty Calibration of Deep Networks for Medical Image Segmentation. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1052-1056. IEEE, 2021. 2 +[31] Gennady Samoradnitsky and Murad S. Taqqu. Stable non-Gaussian random processes: stochastic models with infinite variance. Chapman and Hall, 1994. 4 +[32] Nahian Siddique, Sidike Paheding, Colin P Elkin, and Vijay Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. *IEEE Access*, 9:82031-82057, 2021. 1 +[33] Joao Lourenço Silva and Arlindo L Oliveira. Using Soft Labels to Model Uncertainty in Medical Image Segmentation. arXiv preprint arXiv:2109.12622, 2021. 2, 3 +[34] Nima Tajbakhsh, Laura Jeyaseelan, Qian Li, Jeffrey N Chiang, Zhihao Wu, and Xiaowei Ding. Embracing Imperfect Datasets: A Review of Deep Learning Solutions for Medical Image Segmentation. Medical Image Analysis, 63:101693, 2020. 2 +[35] Eugene Vorontsov and Samuel Kadoury. Label noise in segmentation networks: mitigation must deal with bias. In Deep Generative Models, and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 1, pages 251-258. Springer, 2021. 1, 2 +[36] Zifu Wang, Teodora Popordanoska, Jeroen Bertels, Robin Lemmens, and Matthew B Blaschko. Dice semimetric losses: Optimizing the dice score with soft labels. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 475-485. Springer, 2023. 2 +[37] Jakob Wasserthal, Manfred Meyer, Hanns-Christian Breit, Joshy Cyriac, Shan Yang, and Martin Segeroth. Totalsegmentator: robust segmentation of 104 anatomical structures in ct images. arXiv preprint arXiv:2208.05868, 2022. 2, 7, 8 + +[38] Shaode Yu, Mingli Chen, Erlei Zhang, Junjie Wu, Hang Yu, Zi Yang, Lin Ma, Xuejun Gu, and Weiguo Lu. Robustness Study of Noisy Annotation in Deep Learning Based Medical Image Segmentation. Physics in Medicine & Biology, 65(17):175007, 2020. 1, 2 +[39] Ming-Jie Zhao, Narayanan Edakunni, Adam Pocock, and Gavin Brown. Beyond Fano's Inequality: Bounds on the Optimal F-score, BER, and Cost-Sensitive Risk and Their Implications. The Journal of Machine Learning Research, 14(1):1033-1090, 2013. 2 \ No newline at end of file diff --git a/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/images.zip b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2b239e03551ca8033594f98ef4a622b75c04d5f7 --- /dev/null +++ b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebf57e0613ac9a129e9e11b42cec1922f1c3ffca13d8c0822300d29a31a41774 +size 320852 diff --git a/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/layout.json b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..212996d8077322c71a7db4c7c1d0427aeff3b16f --- /dev/null +++ b/2025/The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation/layout.json @@ -0,0 +1,9532 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 103, + 477, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 103, + 477, + 140 + ], + "spans": [ + { + "bbox": [ + 133, + 103, + 477, + 140 + ], + "type": "text", + "content": "The Impact Label Noise and Choice of Threshold has on Cross-Entropy and Soft-Dice in Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 160, + 462, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 160, + 462, + 203 + ], + "spans": [ + { + "bbox": [ + 149, + 160, + 462, + 203 + ], + "type": "text", + "content": "Marcus Nordström1,2 Atsuto Maki1 Henrik Hult1 \n1KTH Royal Institute of Technology 2RaySearch Laboratories {marcno,atsuto,hult}@kth.se" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 257, + 297, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 297, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 297, + 486 + ], + "type": "text", + "content": "In image segmentation and specifically in medical image segmentation, the soft-Dice loss is often chosen instead of the more traditional cross-entropy loss to improve performance with respect to the Dice metric. Experimental work supporting this claim exists, but how and why the two loss functions lead to different predictions is not well understood. This paper explains the observed discrepancy as a consequence of how those loss functions are effected by label noise and what threshold is used to convert the predicted soft segmentation to predicted labels. In particular, it is shown (i) how the optimal solutions to the two loss functions diverge as the noise is increased, and (ii) how the optimal solutions to soft-Dice can be recovered by thresholding the solutions to cross-entropy with an a priori unknown but efficiently computable threshold. The theoretical results are supported by numerical experiments and it is concluded that cross-entropy with the alternative threshold yields the stability and informative label probability maps associated with cross-entropy without sacrificing the performance of soft-Dice." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 512, + 135, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 135, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 135, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 297, + 715 + ], + "type": "text", + "content": "In medical image analysis it is often of interest to partition an image into foreground and background with respect to some region of interest. This problem is referred to as the medical image segmentation problem. Since the introduction of the U-net [29], state of the art methods for addressing the problem have been dominated by neural network methods trained according to the supervised learning paradigm [7, 32]. These methods are remarkably effective in scenarios with abundant and clean data. However, manually delineating a region of interest accurately and reliably can be a complex and time-consuming process that often requires medical expertise. This result in high costs associated with obtaining high-quality data sets for training, and consequently, that data used in practice often is imperfect in the sense that the manually provided labels are contaminated by noise [1, 5, 23, 27]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 556, + 496 + ], + "type": "text", + "content": "A large amount of research has been devoted to medical image segmentation with imperfect data. One important question that has been addressed is how experimentally robust various segmentation methods are to label noise. In such work, experiments are conducted by first obtaining some data set and then corrupting the portion of data used for training with various types of noise. Different models are then trained on the resulting data sets and evaluated on uncorrupted data. If both noisy labels and noise-free labels can be obtained from clinical practitioners, it is possible to corrupt a clean label simply by swapping it with the corresponding noisy label. Such a study of label noise was presented in [38]. An alternative is to start with a data set that is assumed to be noise-free and generate noisy data with a noise model. Simple noise models that have been studied include randomly flipping labels in an independent manner or randomly translating the target structure [10]. Models that yield more realistic looking segmentation noise include randomly perturbing some points of the boundary on a noise-free segmentation with Gaussian variables [10] and randomly deforming a segmentation with an interpolated low-dimensional Gaussian array [35]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "How label noise affects the performance of segmentation models is highly dependent on the loss function that is used during training. Arguably, the most classical loss function considered in the field is the cross-entropy loss. This loss is known to have a lot of nice theoretical properties and to be stable in practice. However, because of the class imbalance associated with the foreground typically being significantly smaller than the background, other loss functions putting more emphasis on foreground are often considered. The most popular such loss is a smoothed version of the Dice metric. This loss is referred to as soft-Dice or the Dice loss, and has been shown to yield better performance than cross-entropy in several experimental settings at the cost of sometimes introducing some stability issues [2, 24]. To get some of the improved performance while limiting potential stability issues, it is common to use linear combinations of the losses. This is for instance the type of loss used in the popular nnU-net [13]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20820" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 313 + ], + "type": "text", + "content": "Contribution. This work introduces a noise model based on Gaussian field deformations. The model is shown to be approximately volume preserving in the sense that the expected volume of the resulting noisy segmentation is approximately equal to the volume of the corresponding noise-free segmentation. Moreover, a closed form expression for the marginal probabilities and a representation of random deformations using stochastic integrals are provided. This allows the behaviour of the model to be clearly interpreted and efficiently simulated. With the noise model, optimal solutions to cross-entropy and soft-Dice are studied. It is shown how the optimal solutions to soft-Dice and cross-entropy coincide when the labels are noise-free but diverge when noise is increased. It is further shown how optimal solutions to soft-Dice can be recovered by thresholding solutions to cross-entropy with a particular a priori unknown but efficiently computable threshold. Finally, the presented theory is validated experimentally on three different organ segmentation problems from the publicly available TotalSegmentor data set [37] with noise introduced by the model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 142, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 142, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 142, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 342, + 297, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 297, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 297, + 486 + ], + "type": "text", + "content": "Label noise is a well known issue in medical image segmentation and the impact of observer variability has been discussed in various domains for different modalities. This include for radiotherapy [5], for lung nodules on clinical thoracic CT images [1] and for organs in the pelvic region for CT and MR [27]. Examples of work on methods targeting these issues include active learning methods, where experts are iteratively probed for more information [9] and re-weighting schemes where methods dynamically learn to put different weights on different regions depending on certainty [22]. For fairly recent reviews on image segmentation with imperfect data in general, see [14, 34]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 487, + 297, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 297, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 297, + 677 + ], + "type": "text", + "content": "The robustness of medical image segmentation to noise can be studied by training models on data contaminated with varying degrees of noise. In [38], both noise-free and noisy labels are collected and the noise level is determined by the fraction of the training data taken from the noisy data set. Of special importance to this work is segmentation noise provided by Gaussian field deformations. Two works have studied such noise in an experimental setting [10, 35]. In these works, the segmentations in the training data is perturbed according to a low dimensional Gaussian field interpolated to the pixel space. Models are then trained for various noise levels and results obtained are reported. Finally, in [16], a noise model for target delineation in radiotherapy is discussed. The model is based on taking the level set of Gaussian field and efficient methods for sampling is provided." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "The influence label noise has on various loss functions has been studied in several papers. This include the influence it has on calibration and volume bias for predictions generated" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "text", + "content": "with models trained using cross-entropy and soft-Dice [3, 20], re-calibration methods for improving performance of soft-Dice [30] and the relationship between calibration and volume bias [28]. Other experimental work showing that the expected calibration error is lower for cross-entropy than for soft Dice includes [21]. In the above, noise has been analyzed by considering the expected score with respect to some noisy segmentation. An alternative way to incorporate noise is to make use of soft labels. Examples of work using soft labels with soft-Dice include [8, 15]. Examples of work on soft labels with other loss functions include [17, 18, 33]. In [36] a modified version of soft-Dice is introduced. This loss yields optimal solutions that coincide with the soft labels provided for the target structure which is known to not be the case for the classical soft-Dice loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 252, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 556, + 456 + ], + "type": "text", + "content": "Inspired by threshold classifiers studied in [19, 39], optimal segmentations to Accuracy and Dice under label noise was analyzed in [25]. A characterization of optimal segmentations was provided, sharp bounds with respect to volume bias of the optimal segmentations where shown and the volume of the optimizers associated with the metrics was compared. A similar analysis was later provided for the loss function soft-Dice [26]. In particular, a characterization of the optimal solutions and the associated volumes are provided. Moreover, a calibration result showing that a sequence that converge to optimal soft-Dice, when thresholded appropriately, converges to optimal Dice. Experimental work comparing various thresholds for cross-entropy was presented in [4, 12]. The results from these papers experimentally show that using 0.5 as a common threshold leads to sub-optimal segmentations for various loss functions with respect to Dice." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 470, + 399, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 470, + 399, + 483 + ], + "spans": [ + { + "bbox": [ + 314, + 470, + 399, + 483 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 491, + 378, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 378, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 378, + 502 + ], + "type": "text", + "content": "3.1. Notation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "The notation largely follows [25, 26]. Let " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Omega = [0,1]^n\\subset \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " be the unit cube of dimension " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "n\\geq 1" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " be the standard Lebesgue measure associated with " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda (\\Omega) = 1" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " be the space of measurable functions from " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{0,1\\}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " be the space of measurable functions from " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " be the space of bounded measurable functions from " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", where all of the spaces " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " are equipped with their associated Borel " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "-fields and the " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "-norms " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _p" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ". The letter " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "s\\in S" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " is generally used to denote a segmentation, where " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "s(\\omega) = 1" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " indicates foreground of the target structure of interest and " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "s(\\omega) = 0" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " indicates background. The letter " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "m\\in \\mathcal{M}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " is generally used to denote a marginal function, that is, " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "m(\\omega)\\in [0,1]" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " indicates the probability of a noisy label occupying site " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\omega \\in \\Omega" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ". The letter " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "c\\in \\mathcal{M}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " is generally used to denote the soft segmentation that is obtained by optimizing the chosen loss function prior to any thresholding. The letter " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "f\\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " is generally used to denote a logit function usually" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20821" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "associated with a soft segmentation such that " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\sigma \\circ f(\\omega) = c(\\omega)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\omega \\in \\Omega" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " a.e., where " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\sigma(x) \\doteq 1/(1 + e^{-x})" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " is the standard sigmoid function. Composition of functions are denoted by " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "I_A(x)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " is used to denote the indicator function over the set " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", that is " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "I_A(x) = 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "x \\in A" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "I_A(x) = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "x \\notin A" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". The indicator function is used in conjunction with some element " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " to denote thresholding of a soft segmentation. That is, " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "I_{[t,1]} \\circ c" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " assigns 1 to all " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\omega \\in \\Omega" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "c(\\omega) \\geq t" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and 0 to all " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\omega \\in \\Omega" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "c(\\omega) < t" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 187, + 262, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 262, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 262, + 199 + ], + "type": "text", + "content": "3.2. Loss Functions and Evaluation Metrics" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "text", + "content": "Two of the most popular loss functions in the field of medical image segmentation are cross-entropy and soft-Dice. Usually these loss functions are defined over two deterministic arguments, a soft-segmentation (prediction) and a binary segmentation (label). Under the influence of label noise, these loss functions are extended to functions taking one deterministic soft segmentation (prediction) and one random segmentation (label) as arguments. In this work, the soft labeling extension is adopted for the theoretical analysis [8, 15, 17, 18, 33] (see Section 4.4 for more on this). Formally, the loss functions are defined with respect to a marginal function " + }, + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "inline_equation", + "content": "m \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "text", + "content": " which corresponds to the mean of a noisy segmentation " + }, + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "text", + "content": " taking values in " + }, + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "text", + "content": ", that is " + }, + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "inline_equation", + "content": "m(\\omega) = \\mathbb{E}[L(\\omega)], \\omega \\in \\Omega" + }, + { + "bbox": [ + 55, + 205, + 296, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 393 + ], + "type": "text", + "content": "Definition 1. For any " + }, + { + "bbox": [ + 55, + 369, + 296, + 393 + ], + "type": "inline_equation", + "content": "m \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 369, + 296, + 393 + ], + "type": "text", + "content": ", (binary) cross-entropy is given by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 401, + 296, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 401, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 61, + 401, + 296, + 441 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {C E} _ {m} (c) \\doteq - \\int_ {\\Omega} [ m (\\omega) \\log (c (\\omega)) + \\quad c \\in \\mathcal {M}. \\tag {1} \\\\ (1 - m (\\omega)) \\log (1 - c (\\omega)) ] \\lambda (d \\omega), \\\\ \\end{array}", + "image_path": "e9b373781226d35520653af4b876f858942a09f34ec16a5696c0f8dd4365f8ca.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 451, + 269, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 269, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 269, + 463 + ], + "type": "text", + "content": "Definition 2. For any " + }, + { + "bbox": [ + 55, + 451, + 269, + 463 + ], + "type": "inline_equation", + "content": "m \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 451, + 269, + 463 + ], + "type": "text", + "content": ", soft-Dice is given by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 471, + 296, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 471, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 69, + 471, + 296, + 498 + ], + "type": "interline_equation", + "content": "\\mathrm {S D} _ {m} (c) \\doteq 1 - \\frac {2 \\int_ {\\Omega} c (\\omega) m (\\omega) \\lambda (d \\omega)}{\\| c \\| _ {1} + \\| m \\| _ {1}}, \\quad c \\in \\mathcal {M}. \\tag {2}", + "image_path": "83d2f127160951dc50ce285864916b7a0ba0c8cd8043def82225750f922166cd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 590 + ], + "type": "text", + "content": "Arguably, the most classical evaluation metric used is Accuracy. This metric simply measures the fraction of the domain that is correctly labeled. The most popular evaluation metric on the other hand is the Sorensen-Dice coefficient, or Dice for short. To handle label noise, these metrics are defined in the same way as the loss functions, that is, by using soft labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 598, + 270, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 270, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 270, + 611 + ], + "type": "text", + "content": "Definition 3. For any " + }, + { + "bbox": [ + 55, + 598, + 270, + 611 + ], + "type": "inline_equation", + "content": "m \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 598, + 270, + 611 + ], + "type": "text", + "content": ", Accuracy is given by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 80, + 619, + 296, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 619, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 80, + 619, + 296, + 659 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {A} _ {m} (s) \\doteq \\int_ {\\Omega} [ s (\\omega) m (\\omega) + \\quad s \\in \\mathcal {S}. \\tag {3} \\\\ (1 - s (\\omega)) (1 - m (\\omega)) ] \\lambda (d \\omega), \\\\ \\end{array}", + "image_path": "121c1da77f8296cbd27f62e08c5a63f44d6e8547a806847fcc8ddd33e7b5a789.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 668, + 252, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 252, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 252, + 681 + ], + "type": "text", + "content": "Definition 4. For any " + }, + { + "bbox": [ + 55, + 668, + 252, + 681 + ], + "type": "inline_equation", + "content": "m \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 668, + 252, + 681 + ], + "type": "text", + "content": ", Dice is given by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 689, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 689, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 88, + 689, + 296, + 717 + ], + "type": "interline_equation", + "content": "\\mathrm {D} _ {m} (s) \\doteq \\frac {2 \\int_ {\\Omega} s (\\omega) m (\\omega) \\lambda (d \\omega)}{\\| s \\| _ {1} + \\| m \\| _ {1}}, \\quad s \\in \\mathcal {S}. \\tag {4}", + "image_path": "8aef4716a7085cc0de1fa6204206dc7828cc5e9249b3418e610df549262974ee.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "When a label is free from noise, any reasonable loss function or metric is optimized uniquely by the target label. When there is noise present on the other hand, it is no longer clear what the solution should be and loss functions or metrics implicitly assign them. As the consequence, it is important to investigate what the optimal solutions are for a chosen loss or metric. For the case of the cross-entropy loss, it is well known that the optimizers are uniquely given by the marginal function, that is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 335, + 186, + 555, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 186, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 335, + 186, + 555, + 204 + ], + "type": "interline_equation", + "content": "\\mathrm {C E} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {C E} _ {m} \\left(c ^ {\\prime}\\right) \\Longleftrightarrow c (\\omega) = m (\\omega), \\tag {5}", + "image_path": "b7aadabc60e966d9dcfd841ca972d6b6063536d8666819eb295b43131ea771dc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 209, + 555, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 246 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 209, + 555, + 246 + ], + "type": "text", + "content": "-a.e. In other words, minimizing cross-entropy yields estimates of the marginal probability distribution associated with the labels. For soft-Dice it was shown in [26] that" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 250, + 441, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 250, + 441, + 268 + ], + "spans": [ + { + "bbox": [ + 325, + 250, + 441, + 268 + ], + "type": "interline_equation", + "content": "\\mathrm {S D} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {S D} \\left(c ^ {\\prime}\\right) \\Longleftrightarrow", + "image_path": "d0de37b6a62a1d7ac257770a2729cc592716ad443bd78b888728b664a2e08eec.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 332, + 271, + 555, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 271, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 332, + 271, + 555, + 316 + ], + "type": "interline_equation", + "content": "c (\\omega) \\in \\left\\{ \\begin{array}{l l} \\{1 \\} & \\text {i f} m (\\omega) > \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2, \\\\ [ 0, 1 ] & \\text {i f} m (\\omega) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2, \\\\ \\{0 \\} & \\text {i f} m (\\omega) < \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2, \\end{array} \\right. \\tag {6}", + "image_path": "480e6f2c21fb38fd5d5f0e3b792e3ae9157012732bc8bb5ddad1f8ae8e10bd65.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "content": "-a.e. Characterizations of the optimal segmentations for Accuracy and Dice using the setup adopted in this paper can be found in [25]. For Accuracy, the solutions are given by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 357, + 372, + 474, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 372, + 474, + 392 + ], + "spans": [ + { + "bbox": [ + 357, + 372, + 474, + 392 + ], + "type": "interline_equation", + "content": "\\mathrm {A} _ {m} (s) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {A} _ {m} \\left(s ^ {\\prime}\\right) \\iff", + "image_path": "cd9ae50928cbb0114c804b5a915b7b751a5d7aef58ce098fed86aae83faf87ea.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 366, + 395, + 555, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 395, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 366, + 395, + 555, + 439 + ], + "type": "interline_equation", + "content": "s (\\omega) \\in \\left\\{ \\begin{array}{l l} \\{1 \\} & \\text {i f} m (\\omega) > 1 / 2, \\\\ \\{0, 1 \\} & \\text {i f} m (\\omega) = 1 / 2, \\\\ \\{0 \\} & \\text {i f} m (\\omega) < 1 / 2, \\end{array} \\right. \\tag {7}", + "image_path": "1c5a9d555162c7102a684566e6296f1071483509e7605fa8cf7e40b5fff6b5da.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 445, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 445, + 554, + 468 + ], + "type": "text", + "content": "-a.e. Similarly, for Dice the optimal segmentations are given by" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 322, + 474, + 555, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 474, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 322, + 474, + 555, + 495 + ], + "type": "interline_equation", + "content": "\\mathrm {D} _ {m} (s) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) \\Longleftrightarrow \\tag {8}", + "image_path": "12fb422e616851809fa3ca268a59015def432b3ccac8df74582c1806aaa3a94e.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 331, + 498, + 554, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 498, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 331, + 498, + 554, + 542 + ], + "type": "interline_equation", + "content": "s (\\omega) \\in \\left\\{ \\begin{array}{l l} \\{1 \\} & \\text {i f} m (\\omega) > \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} (s ^ {\\prime}) / 2, \\\\ \\{0, 1 \\} & \\text {i f} m (\\omega) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} (s ^ {\\prime}) / 2, \\\\ \\{0 \\} & \\text {i f} m (\\omega) < \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} (s ^ {\\prime}) / 2, \\end{array} \\right. \\tag {9}", + "image_path": "f1e6afe6ca7ecb4129114d4abe017168d45f064d64be91884dd5b5a2ceb9598a.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 547, + 555, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 607 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 547, + 555, + 607 + ], + "type": "text", + "content": "-a.e. The resulting soft segmentation obtained by optimizing a loss function is typically converted to a binary segmentation by thresholding with 0.5. Ideally this resulting segmentation maximizes a desired target metric. Using the above characterizations, for cross-entropy it follows that" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 370, + 612, + 494, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 612, + 494, + 630 + ], + "spans": [ + { + "bbox": [ + 370, + 612, + 494, + 630 + ], + "type": "interline_equation", + "content": "\\mathrm {C E} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {C E} \\left(c ^ {\\prime}\\right) \\Longrightarrow", + "image_path": "40988539ccc2d7c42755bdc7f735860225f5f857bcc120e53b8892a84e7f8090.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 373, + 628, + 555, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 628, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 373, + 628, + 555, + 652 + ], + "type": "interline_equation", + "content": "\\mathrm {A} _ {m} \\left(I _ {[ 1 / 2, 1 ]} \\circ c\\right) = \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {A} \\left(s ^ {\\prime}\\right). \\tag {10}", + "image_path": "b492dd797053b762a858bc387f2f5dfb992554abd7702503ac11cbe2098e7c21.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 658, + 468, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 658, + 468, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 658, + 468, + 670 + ], + "type": "text", + "content": "Similarly, for soft-Dice it follows that" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 372, + 675, + 494, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 675, + 494, + 693 + ], + "spans": [ + { + "bbox": [ + 372, + 675, + 494, + 693 + ], + "type": "interline_equation", + "content": "\\mathrm {S D} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {S D} \\left(c ^ {\\prime}\\right) \\Longrightarrow", + "image_path": "b2934ae2007b234df41cc791d3d010d8a9317a5c7b9c25523cbd5b9ef4a2e142.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 372, + 689, + 555, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 689, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 372, + 689, + 555, + 716 + ], + "type": "interline_equation", + "content": "\\mathrm {D} _ {m} \\left(I _ {[ 1 / 2, 1 ]} \\circ c\\right) = \\sup _ {s ^ {\\prime} \\in S} \\mathrm {D} \\left(s ^ {\\prime}\\right). \\tag {11}", + "image_path": "202180360dfafbca49ea960b6879ef6df20982c803da76053e716d1cdd677018.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20822" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "3.3. Gaussian Fields and Independently Scattered Gaussian Measures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 296, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 296, + 150 + ], + "type": "text", + "content": "Amongst the most central objects in probability theory and statistics are Gaussian variables taking values in Euclidean space. Such a variable, when centered and isotropic, has the probability density function" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 159, + 296, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 159, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 67, + 159, + 296, + 186 + ], + "type": "interline_equation", + "content": "p _ {\\sigma^ {2}} (\\omega) \\dot {=} \\frac {1}{(2 \\pi \\sigma^ {2}) ^ {n / 2}} \\exp \\left\\{- \\frac {\\| \\omega \\| _ {2} ^ {2}}{2 \\sigma^ {2}} \\right\\}, \\quad \\omega \\in \\mathbb {R} ^ {n}, \\tag {12}", + "image_path": "912ae3581036a4827e7615e1dd7368e3341784db6df02121cba3a78c98344e8a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 196, + 297, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 297, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 297, + 256 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 196, + 297, + 256 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 55, + 196, + 297, + 256 + ], + "type": "text", + "content": " denotes the variance in each of the dimensions. Various generalizations of this random variable have been studied extensively in the literature. For the theoretical treatment of the noise model in this work, two such generalizations are introduced. The first is a Gaussian field." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "text", + "content": "Definition 5. A random variable " + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "inline_equation", + "content": "Y(\\omega), \\omega \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "text", + "content": " is a Gaussian field with kernel function " + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "text", + "content": " if any finite collection of elements " + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\{\\omega_1, \\dots, \\omega_M\\} \\subset \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "text", + "content": " satisfies " + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "inline_equation", + "content": "(Y(\\omega_1), \\dots, Y(\\omega_M)) \\sim \\text{normal}(0, \\Sigma)" + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\Sigma_{m,m'} = k(\\omega_m, \\omega_{m'})" + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "inline_equation", + "content": "m, m' \\in \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 55, + 266, + 296, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 335, + 296, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 372 + ], + "type": "text", + "content": "Many types of kernel functions with different properties have been studied in the literature. Amongst the most common is the squared exponential kernel given by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 380, + 296, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 380, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 64, + 380, + 296, + 419 + ], + "type": "interline_equation", + "content": "k _ {a, b} \\left(\\omega , \\omega^ {\\prime}\\right) \\doteq a ^ {2} \\exp \\left(- \\frac {\\| \\omega - \\omega^ {\\prime} \\| _ {2} ^ {2}}{2 b ^ {2}}\\right), \\quad \\omega , \\omega^ {\\prime} \\in \\mathbb {R} ^ {n}. \\tag {13}", + "image_path": "1d6b7d5e46f294c3bdce1472e39cc2f4ff5517ad6f51a290652e083215f2f907.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "content": "An important property of Gaussian fields with such covariance functions are that samples are almost surely continuous. In other words, a randomly sampled function will be continuous with probability one. For a good reference on Gaussian variables, see [11]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": "The second generalization that is introduced is what is referred to as independently scattered Gaussian measures. Such constructions are used to formalize integration over Gaussian fields, similar to how the Ito-integral build on Brownian motion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": "Definition 6. A random variable " + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": " is an independently scattered Gaussian measure on " + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": " with respect to the control measure " + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": ", if for any collection of disjoint sets " + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "inline_equation", + "content": "A_1, \\ldots, A_k" + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": "-measurable sets with finite measure " + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_0" + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "inline_equation", + "content": "W(A_1), \\ldots, W(A_k)" + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": " are independent, and " + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "inline_equation", + "content": "W(A_i)" + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": " has centered Gaussian distribution with variance " + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\lambda(A_i)" + }, + { + "bbox": [ + 55, + 559, + 296, + 632 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 639, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 677 + ], + "type": "text", + "content": "Importantly, for a function " + }, + { + "bbox": [ + 55, + 639, + 296, + 677 + ], + "type": "inline_equation", + "content": "f \\in L^{2}(\\mathbb{R}^{n})" + }, + { + "bbox": [ + 55, + 639, + 296, + 677 + ], + "type": "text", + "content": ", the stochastic integral " + }, + { + "bbox": [ + 55, + 639, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\int_{\\mathbb{R}^n} f(\\omega) W(d\\omega)" + }, + { + "bbox": [ + 55, + 639, + 296, + 677 + ], + "type": "text", + "content": " is well defined and the process " + }, + { + "bbox": [ + 55, + 639, + 296, + 677 + ], + "type": "inline_equation", + "content": "Y = \\{Y(\\omega), \\omega \\in \\mathbb{R}^n\\}" + }, + { + "bbox": [ + 55, + 639, + 296, + 677 + ], + "type": "text", + "content": " given by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 685, + 296, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 685, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 107, + 685, + 296, + 712 + ], + "type": "interline_equation", + "content": "Y (\\omega) \\doteq \\int_ {\\mathbb {R} ^ {n}} f \\left(\\omega - \\omega^ {\\prime}\\right) W \\left(d \\omega^ {\\prime}\\right), \\tag {14}", + "image_path": "ca564737720cb64947fcc7269cfeaffbb2b18d4395863e348cfaa95ae936e326.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "is a centered Gaussian process. Furthermore, this Gaussian process has a covariance kernel given by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 105, + 555, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 105, + 555, + 142 + ], + "spans": [ + { + "bbox": [ + 317, + 105, + 555, + 142 + ], + "type": "interline_equation", + "content": "k \\left(\\omega , \\omega^ {\\prime}\\right) = \\int_ {\\mathbb {R} ^ {n}} f (\\omega - u) f \\left(\\omega^ {\\prime} - u\\right) \\lambda (d u), \\quad \\omega , \\omega^ {\\prime} \\in \\mathbb {R} ^ {n}. \\tag {15}", + "image_path": "5bc0b5a09f6bc67b4ed12ad1b558866542c5a91aa762824c971c7d5c920f2a4a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 152, + 555, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 198 + ], + "type": "text", + "content": "Independently scattered measures are introduced more generally for " + }, + { + "bbox": [ + 313, + 152, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 152, + 555, + 198 + ], + "type": "text", + "content": "-stable distributions in [31, Section 3.3]. The Gaussian case used in this work is simply the special case when " + }, + { + "bbox": [ + 313, + 152, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\alpha = 2" + }, + { + "bbox": [ + 313, + 152, + 555, + 198 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 210, + 367, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 367, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 367, + 224 + ], + "type": "text", + "content": "4. Theory" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 230, + 425, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 425, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 425, + 242 + ], + "type": "text", + "content": "4.1. Label Noise Model" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 354 + ], + "type": "text", + "content": "In contrast to binary classification, where a lot of label noise can be modeled by simple independent flipping of binary variables, label noise in segmentation can in general have very complicated dependency structures. Often labels affected by noise have the appearance of being drawn by someone with shaky hands. For such noise, the correlation between border points will depend on the distance between them. That is, if two points are close to each other, then the correlation will be high and vice versa." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 356, + 554, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 390 + ], + "type": "text", + "content": "One way to model segmentations that look as if they have been drawn by shaky hands is by means of random Gaussian field deformations. Consider a variable" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 390, + 399, + 555, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 399, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 390, + 399, + 555, + 415 + ], + "type": "interline_equation", + "content": "X \\doteq \\left(X _ {1}, \\dots , X _ {n}\\right) ^ {T} \\tag {16}", + "image_path": "4357822f7dadc74d57577eae03202d9e3bbdb7f9930b1917fe8a6aefd8adfe4a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 423, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 472 + ], + "type": "text", + "content": "such that " + }, + { + "bbox": [ + 313, + 423, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\{X_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 423, + 555, + 472 + ], + "type": "text", + "content": " are independent Gaussian fields with kernel function (13). This means that " + }, + { + "bbox": [ + 313, + 423, + 555, + 472 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 423, + 555, + 472 + ], + "type": "text", + "content": " is a random vector field taking values in " + }, + { + "bbox": [ + 313, + 423, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n \\mapsto \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 423, + 555, + 472 + ], + "type": "text", + "content": ". With such a random vector field, a random segmentation can be formed by" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 480, + 555, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 480, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 318, + 480, + 555, + 524 + ], + "type": "interline_equation", + "content": "L (\\omega) \\doteq \\left\\{ \\begin{array}{l l} l (\\omega + X (\\omega)) & \\text {i f} \\omega + X (\\omega) \\in \\Omega , \\\\ 0 & \\text {i f} \\omega + X (\\omega) \\notin \\Omega . \\end{array} \\right., \\quad \\omega \\in \\Omega , \\tag {17}", + "image_path": "05171796159fccc9be9764e7e37467029cd88b38bc2ec8af5960b0780e3ae8f7.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "inline_equation", + "content": "l \\in S" + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": " is a noise-free reference segmentation being randomly deformed. How samples of the random segmentations look depend on what parameters are chosen when specifying the kernel function (13). The parameter " + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": " controls how large the deformations are, where greater " + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": " leads to larger deformations and vice versa. The parameter " + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": " controls how smooth the deformations are, where greater " + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": " leads to smoother deformations and vice versa. Modeling noisy segmentations with (17) is convenient because it allows for the derivation of several theoretical properties. In this work, three such important properties are derived." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": "In Proposition 1, it is shown that the marginal probabilities associated with a noisy segmentation can be computed by convolving the associated noise-free segmentation label with an appropriate Gaussian density." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20823" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": "Proposition 1. Let " + }, + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": " be defined as in (17) and " + }, + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "inline_equation", + "content": "p_{\\sigma^2}" + }, + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": " be defined as in (12), then" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 103, + 297, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 297, + 129 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 297, + 129 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ L (\\omega) ] = \\int_ {\\Omega} l \\left(\\omega^ {\\prime}\\right) p _ {a ^ {2}} \\left(\\omega - \\omega^ {\\prime}\\right) \\lambda \\left(d \\omega^ {\\prime}\\right), \\quad \\omega \\in \\Omega . \\tag {18}", + "image_path": "fff245e08fd63b51851daeb01b6684818581113dbc7affb98894227a98f3ea10.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "content": "In Proposition 2, the marginal representation is used to show that the volume of the noisy segmentation is approximately unbiased in the sense that the expected volume of a noisy segmentation is approximately equal to the volume of the corresponding noise-free segmentation. The reason why this does not hold exactly is that there is some probability for the target structure to deform in such a way that parts of the structure leave the domain " + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 149, + 296, + 281 + ], + "type": "text", + "content": ". However, for most reasonable choices of parameters, when the regions of interest is sufficiently close to the center, the probability of this happening is negligible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "text", + "content": "Proposition 2. Let " + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "text", + "content": " be defined as in (17) and " + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "inline_equation", + "content": "p_{\\sigma^2}" + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "text", + "content": " be defined as in 12, then" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 320, + 296, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 320, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 320, + 296, + 335 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ \\| L \\| _ {1} ] = \\| l \\| _ {1} - \\xi , \\tag {19}", + "image_path": "5fd40b06c62a3608b7bee6887ad930dbe177c57395006bc8f73646841f46969e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 84, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 84, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 84, + 353 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 359, + 296, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 359, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 72, + 359, + 296, + 386 + ], + "type": "interline_equation", + "content": "\\xi = \\int_ {\\Omega} \\int_ {\\mathbb {R} ^ {n} \\backslash \\Omega} l \\left(\\omega^ {\\prime}\\right) p _ {a ^ {2}} \\left(\\omega - \\omega^ {\\prime}\\right) \\lambda (d \\omega) \\lambda \\left(d \\omega^ {\\prime}\\right). \\tag {20}", + "image_path": "e7f978d2bf8e8d8c1a8f29a1e631c60e310e357484b973fc2b0f45da19ff9bdd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 406, + 297, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 297, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 297, + 465 + ], + "type": "text", + "content": "In Proposition 3, it is shown that " + }, + { + "bbox": [ + 55, + 406, + 297, + 465 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 406, + 297, + 465 + ], + "type": "text", + "content": " (16) is equal in distribution to a particular vector of stochastic integrals. In the case " + }, + { + "bbox": [ + 55, + 406, + 297, + 465 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 55, + 406, + 297, + 465 + ], + "type": "text", + "content": ", this could have been done using regular Ito-integrals, but for the case " + }, + { + "bbox": [ + 55, + 406, + 297, + 465 + ], + "type": "inline_equation", + "content": "n > 1" + }, + { + "bbox": [ + 55, + 406, + 297, + 465 + ], + "type": "text", + "content": " a multivariate counter part is needed. For this scattered Gaussian measures are used." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "text", + "content": "Proposition 3. Let " + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "text", + "content": " be as in (16), " + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "inline_equation", + "content": "p_{\\alpha^2}" + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "text", + "content": " be as in (12) and " + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "inline_equation", + "content": "W_1, \\ldots, W_n" + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "text", + "content": " be independent copies of independently scattered Gaussian measures on " + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "text", + "content": " with control measure " + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 472, + 297, + 521 + ], + "type": "text", + "content": ", then it follows that" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 528, + 296, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 528, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 64, + 528, + 296, + 585 + ], + "type": "interline_equation", + "content": "X (\\omega) \\stackrel {d} {=} \\left( \\begin{array}{c} a (2 \\pi b ^ {2}) ^ {n / 4} \\int_ {\\mathbb {R} ^ {n}} p _ {b ^ {2} / 2} \\left(\\omega - \\omega^ {\\prime}\\right) W _ {1} \\left(d w ^ {\\prime}\\right) \\\\ \\vdots \\\\ a (2 \\pi b ^ {2}) ^ {n / 4} \\int_ {\\mathbb {R} ^ {n}} p _ {b ^ {2} / 2} \\left(\\omega - \\omega^ {\\prime}\\right) W _ {n} \\left(d w ^ {\\prime}\\right) \\end{array} \\right). \\tag {21}", + "image_path": "8070d49b2504893719bb039a4eceecd422b6a154a347768926857b212d54a701.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "type": "text", + "content": "The presented propositions provide meaningful insights into the behaviour of the noise model. The importance of Proposition 1 is that it makes the relationship between the noise model parameters and the resulting marginal probabilities clear. Also, it provides an exact method for computing the marginals that otherwise would have to be estimated with sampling. The importance of Proposition 2 is that the noise beside from looking reasonable also numerically behave reasonable in an important manner. Finally, the importance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "of Proposition 3 is that noise for discretized domains can be sampled by smoothing arrays of independently sampled Gaussian variables with an appropriate Gaussian filters. Both operations which scale well with image size. This allows for efficient sampling of noisy segmentations for images composed of millions of voxels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 151, + 537, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 537, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 537, + 164 + ], + "type": "text", + "content": "4.2. Optimal Segmentations Under Label Noise" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 168, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 336 + ], + "type": "text", + "content": "Since the loss functions and metrics investigated in this work only depend on the marginal probabilities and Proposition 1 provides an explicit expression for these marginal probabilities, it is possible to study how noise model parameters affect optimal segmentations. Figure 1 depicts this relationship for a two dimensional segmentation problem with varying noise intensity. In the first row, pixel-wise averages of 5 samples are illustrated. In the second row, the marginal probabilities are illustrated. In the third row, theoretically optimal solutions to Accuracy are shown. Note that by (10), this is equivalently given by the minimizer to cross-entropy thresholded by 0.5. In the forth row, theoretical optimal solutions to Dice are shown. Note that by (11), this is equivalently given by the minimizer to soft-Dice thresholded by 0.5." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 556, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 455 + ], + "type": "text", + "content": "A first observation is that noise affects the optimal segmentations in such a way that the shape of the noise-free reference segmentation is changed. This is an important observation that contradicts the picture that unbiased noise (unbiased in the sense that the expected deformation is the identity function) yield unbiased optimal segmentations. In particular, optimal segmentations are changed in such a way that the corners of the shapes are smoothed and details such as thin structures and holes disappear. Informally, the higher the noise, the more ball shaped the optimal segmentations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 456, + 557, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 557, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 557, + 599 + ], + "type": "text", + "content": "A second observation is that the two metrics yield similar looking segmentations for low noise but increasingly different segmentations as the noise increase. Since cross-entropy with the 0.5-threshold theoretically gives optimal segmentations with respect to Accuracy and soft-Dice with the 0.5-threshold theoretically gives optimal segmentations with respect to Dice, the performance of the two loss functions as measured by Dice is similar in low noise situations but increasingly better for soft-Dice as the noise is increased. Similarly, the performance of the two loss functions as measured by Accuracy is similar in low noise situations but increasingly better for cross-entropy as the noise is increased." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 550, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 550, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 550, + 619 + ], + "type": "text", + "content": "4.3. Cross-entropy With an Alternative Threshold" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 623, + 555, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 555, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 555, + 683 + ], + "type": "text", + "content": "By the characterization of optimizers to cross-entropy (5) and the characterization of optimizers to soft-Dice (6), it follows that an optimizer to soft-Dice can be obtained by first optimizing cross-entropy and then thresholding with the a priori unknown value" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 388, + 691, + 555, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 691, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 388, + 691, + 555, + 711 + ], + "type": "interline_equation", + "content": "t _ {m} \\doteq \\sup _ {s ^ {\\prime} \\in \\mathcal {S}} \\mathrm {D} _ {m} \\left(s ^ {\\prime}\\right) / 2. \\tag {22}", + "image_path": "6b1bed084e6dfd10e88191366277272fd48145a2b0e296818187d2be10ea1f4b.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20824" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 361 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 361 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 361 + ], + "type": "image", + "image_path": "71238ae1e0ffc3ac440274252f170f0eb5622b6369704d29ba4a7f47f4e34172.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 373, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 54, + 373, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 555, + 430 + ], + "type": "text", + "content": "Figure 1. Illustration of the noise model and its effects on a two-dimensional segmentation problem, with the columns showing the effect of different noise strengths. The first row shows the pixel-wise average obtained from five samples. The second row shows the marginal function. The third row shows optimal segmentations associated with Accuracy (7), which in this and most natural cases is unique. The forth row shows optimal segmentations associated with Dice (9), which in this and most natural cases is unique. In all of the reported cases " + }, + { + "bbox": [ + 54, + 373, + 555, + 430 + ], + "type": "inline_equation", + "content": "b = 0.15 / \\sqrt{2}" + }, + { + "bbox": [ + 54, + 373, + 555, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 450, + 298, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 450, + 298, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 450, + 298, + 485 + ], + "type": "text", + "content": "That is, by thresholding an optimizer to cross-entropy with this alternative recursively defined threshold " + }, + { + "bbox": [ + 54, + 450, + 298, + 485 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 54, + 450, + 298, + 485 + ], + "type": "text", + "content": ", it follows that" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 502, + 296, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 502, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 108, + 502, + 296, + 540 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {C E} _ {m} (c) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {C E} \\left(c ^ {\\prime}\\right) \\Longrightarrow \\tag {23} \\\\ \\mathrm {S D} _ {m} \\left(I _ {[ t _ {c}, 1 ]} \\circ c\\right) = \\inf _ {c ^ {\\prime} \\in \\mathcal {M}} \\mathrm {S D} \\left(c ^ {\\prime}\\right). \\\\ \\end{array}", + "image_path": "e91ea5230048d26d63e1bbd20d15e98b4e4f8aacbb27cdbe38d08969847ea29c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": "Methods based on this principle have not been discussed in the medical image segmentation community but are known in the statistical learning literature [19]. In the medical image segmentation community, however, there has been experimental work comparing the performance of different constant thresholds, that is, thresholds not dependent on the marginal probabilities [4, 12]. The results experimentally show that using a 0.5-threshold leads to sub-optimal segmentations for various loss functions when the Dice metric is used for evaluation. Consequently, the results are in line with the theoretical observations above and motivate further investigation of methods based on cross-entropy with the threshold (22)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 449, + 495, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 495, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 495, + 462 + ], + "type": "text", + "content": "4.4. Soft Labeling and Expected Score" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": "One reason to why the threshold (22) has not been discussed in the medical image segmentation domain may be that loss functions and metrics often incorporate noise using expectations rather than soft labels. Formally for some random label " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " taking values in " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[L(\\omega)] = m(\\omega), \\omega \\in \\Omega" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{M}" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": ", this means that " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{CE}_L(c)]" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " is considered instead of " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}_m(c)" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{SD}_L(c)]" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " is considered instead of " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{SD}_m(c)" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": ". Similarly for some " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": ", this means that " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{A}_L(s)]" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " is considered instead of " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_m(s)" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{D}_L(s)]" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " is considered instead of " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{D}_m(s)" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "For cross-entropy and Accuracy, it is easy to verify that the definitions using expectations and the definitions using soft labels are equivalent, that is " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{A}_L(s)] = \\mathrm{A}_m(s), s \\in S" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{CE}_L(c)] = \\mathrm{CE}_m(c), c \\in \\mathcal{M}" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": ". For soft-Dice and Dice, however, this is in general not the case, that is, in general " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{D}_L(s)] \\neq \\mathrm{D}_m(s), s \\in S" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{SD}_L(c)] \\neq \\mathrm{SD}_m(c), c \\in \\mathcal{M}" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": ". Hence, the property (23) does not in general transfer to properties of optimizers to soft-Dice using expectations. Importantly however, that the definitions of soft-Dice and Dice using soft labels and expectations are" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20825" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "equivalent provided that the volume of the noisy label is constant, that is " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{Var}[\\| L\\| _1] = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". For situations when the noise is approximately constant " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{Var}[\\| L\\| _1]\\approx 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", the definitions are approximately equal " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{D}_L(s)]\\approx \\mathrm{D}_m(s),s\\in S" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathrm{SD}_L(c)]\\approx \\mathrm{SD}_m(c),c\\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " [25]. Hence the property (23) holds approximately for the kind of noise studied in this work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 136, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 136, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 136, + 181 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 188, + 108, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 108, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 108, + 199 + ], + "type": "text", + "content": "5.1. Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": "A standard 3D U-net is implemented [6]. The network is composed of blocks of two convolutions with kernel size " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "3^3" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": ", instance normalization and ReLu-activations. The number of filters used is increased for each down-sampling according to 16, 32, 64, 128, 256, where 16 corresponds to the number of filters used in the input layer and 256 corresponds to the number of filters in the bottleneck layer. Convolutions with kernel " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "2^3" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": " and stride 2 are used for down-sampling. Transposed convolutions with kernel " + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "inline_equation", + "content": "2^3" + }, + { + "bbox": [ + 55, + 205, + 296, + 338 + ], + "type": "text", + "content": " stride 2 are used for up-sampling. Details and code can be found in the Supplementary Material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 346, + 119, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 119, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 119, + 358 + ], + "type": "text", + "content": "5.2. Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 364, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 518 + ], + "type": "text", + "content": "Models are trained either with the cross-entropy loss or with the soft-Dice loss. In both cases " + }, + { + "bbox": [ + 55, + 364, + 296, + 518 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 364, + 296, + 518 + ], + "type": "text", + "content": " regularization is used with a weight of " + }, + { + "bbox": [ + 55, + 364, + 296, + 518 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 55, + 364, + 296, + 518 + ], + "type": "text", + "content": ". Adam with learning rate " + }, + { + "bbox": [ + 55, + 364, + 296, + 518 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 364, + 296, + 518 + ], + "type": "text", + "content": " is used for optimization together with a scheduler that divides the learning rate by 5 after 30 epochs of non-improvement. The number of epochs computed is fixed to 100 and the order of the training data is randomly shuffled. Batch sizes of 1 is used and with probability 0.5 the data is augmented. The augmentation is composed of random shifts, random rotations and image blurring with Gaussian filters. The model that performs best on the training data is the model selected. Details and code can be found in the Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 100, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 100, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 100, + 540 + ], + "type": "text", + "content": "5.3.Data" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "The experiments are conducted with respect to the TotalSegmentor data set [37]. This data set contains 1204 CT images with 104 anatomical structures (27 organs, 59 bones, 10 muscles, 8 vessels). To illustrate the effect noise may have on organs with different shape, three different organs are chosen. This includes the right kidney which in general is spherical, the aorta which is tubular and relatively thick and the esophagus which is tubular and relatively thin. For each of the organs, 400 cases is selected and split into 5 folds each containing 80 cases. The images are sub-sampled to half resolution and patches of " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " voxels centered in each of the structures are extracted. In order to simulate label noise the described noise model is used. For each organ data set, four different datasets was created by randomly deforming the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 395, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 395, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 395, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 395, + 148 + ], + "type": "image", + "image_path": "e5b7a1bdd2328e6729d7b379c04e17ac03fdc708fa5d17de501e2601dc41dc16.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 395, + 71, + 473, + 148 + ], + "blocks": [ + { + "bbox": [ + 395, + 71, + 473, + 148 + ], + "lines": [ + { + "bbox": [ + 395, + 71, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 395, + 71, + 473, + 148 + ], + "type": "image", + "image_path": "ef0aaf31ee8630f99f8edeaba658505db80f7d2add6eed19b95b447623758590.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 474, + 71, + 552, + 148 + ], + "blocks": [ + { + "bbox": [ + 474, + 71, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 474, + 71, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 474, + 71, + 552, + 148 + ], + "type": "image", + "image_path": "852e87ac540c4ef801d6fd2481a190a899c31187b5f54ee31a3192694809cbde.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 150, + 394, + 227 + ], + "blocks": [ + { + "bbox": [ + 317, + 150, + 394, + 227 + ], + "lines": [ + { + "bbox": [ + 317, + 150, + 394, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 394, + 227 + ], + "type": "image", + "image_path": "903fb259c84959ff8d6d857f4d978944eb21a46a914551d7fea1551c955ec9bf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 395, + 150, + 473, + 227 + ], + "blocks": [ + { + "bbox": [ + 395, + 150, + 473, + 227 + ], + "lines": [ + { + "bbox": [ + 395, + 150, + 473, + 227 + ], + "spans": [ + { + "bbox": [ + 395, + 150, + 473, + 227 + ], + "type": "image", + "image_path": "0694241640b8c721f29c283c37aa7edbcac75c442a57030825a85477e4837dca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 474, + 150, + 552, + 227 + ], + "blocks": [ + { + "bbox": [ + 474, + 150, + 552, + 227 + ], + "lines": [ + { + "bbox": [ + 474, + 150, + 552, + 227 + ], + "spans": [ + { + "bbox": [ + 474, + 150, + 552, + 227 + ], + "type": "image", + "image_path": "dbd44d53b5d96d07b63c9de096353bb8411d33ff2f6074f888b04e2474b11b2a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 228, + 394, + 306 + ], + "blocks": [ + { + "bbox": [ + 317, + 228, + 394, + 306 + ], + "lines": [ + { + "bbox": [ + 317, + 228, + 394, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 394, + 306 + ], + "type": "image", + "image_path": "57a52df7f11289c6fa0d52824939709e2e64f24ced00ff73d6c99a825ddd649d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 315, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 361 + ], + "type": "text", + "content": "Figure 2. Illustration of the data from the TotalSegmentor data set [37] used in the experiments. The first depicts a kidney, the second row an aorta and the third row an esophagus. The columns correspond to axial, coronal and sagittal views respectively." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 395, + 228, + 473, + 306 + ], + "blocks": [ + { + "bbox": [ + 395, + 228, + 473, + 306 + ], + "lines": [ + { + "bbox": [ + 395, + 228, + 473, + 306 + ], + "spans": [ + { + "bbox": [ + 395, + 228, + 473, + 306 + ], + "type": "image", + "image_path": "3dff81445302c5f8019c4a8b811b8de9ec39bd0394b4cc9f4b1a3687f2e21f2f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 474, + 228, + 552, + 306 + ], + "blocks": [ + { + "bbox": [ + 474, + 228, + 552, + 306 + ], + "lines": [ + { + "bbox": [ + 474, + 228, + 552, + 306 + ], + "spans": [ + { + "bbox": [ + 474, + 228, + 552, + 306 + ], + "type": "image", + "image_path": "79a24752fe1a5b48cd01005710aebd9cec16799502a517bc3fc163b0a82ae353.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 383, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 431 + ], + "type": "text", + "content": "labels with the noise strengths " + }, + { + "bbox": [ + 313, + 383, + 555, + 431 + ], + "type": "inline_equation", + "content": "a = 0.0, 0.1, 0.2, 0.3" + }, + { + "bbox": [ + 313, + 383, + 555, + 431 + ], + "type": "text", + "content": ". Each data point is deformed once, resulting in one sample for each random label. Details and code can be found in the Supplementary Material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 441, + 372, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 372, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 372, + 453 + ], + "type": "text", + "content": "5.4. Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "content": "Results from the experiments are illustrated separately for the three organ segmentation problems as a figure in Figure 3 and listed as a table in the Supplementary Material. For each organ segmentation problem, the average Dice score for the various methods is reported as a function of the parameter controlling the noise strength in the noise model. The average score is obtained by taking the average of the results obtained after 5 folds. The notation " + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}^{(0)}" + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "content": " is used to denote a model trained with cross-entropy and thresholded with 0.5, " + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{SD}^{(0)}" + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "content": " a model trained with cross-entropy and thresholded with 0.5 and " + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}^{(*)}" + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "content": " a model trained with cross-entropy and thresholded with " + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "content": " (22)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "text", + "content": "The first thing to notice is that the difference amongst the average performance of cross-entropy and soft-Dice with the 0.5-threshold, that is, " + }, + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}^{(0)}" + }, + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{SD}^{(0)}" + }, + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "text", + "content": ", increase as the amount of added noise increase. When no noise is added, the performance is indistinguishable. This type of result is expected from the theory and depicted in Figure 1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "The second thing to notice is that cross-entropy with the alternative threshold (22) " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}^{(*)}" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": " performs better than cross-entropy with the 0.5-threshold " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}^{(0)}" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": ". In regions with low" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20826" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 220, + 242 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 220, + 242 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 220, + 242 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 220, + 242 + ], + "type": "image", + "image_path": "4c1789cbdfe4ad40ce452d8b9e3c6758772f75a6ded02f559eb5a0854d441295.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "text", + "content": "Figure 3. To the left are the results from the kidney segmentation problem, in the middle are the results from the aorta segmentation problem and to the right are the results from the esophagus segmentation problem. In each of the figures, the average Dice value obtained from five fold experiments is plotted as a function of the noise level " + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "text", + "content": ". The legend " + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}^{(0)}" + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "text", + "content": " indicate that cross-entropy with the " + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "text", + "content": "-threshold has been used, " + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{SD}^{(0)}" + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "text", + "content": " indicate that soft-Dice with the " + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "text", + "content": "-threshold has been used and " + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}^{(*)}" + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "text", + "content": " indicate that cross-entropy with the threshold described in (22) has been used. In all of the reported cases " + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "inline_equation", + "content": "b = 0.15/\\sqrt{2}" + }, + { + "bbox": [ + 55, + 252, + 555, + 308 + ], + "type": "text", + "content": ". A table of the results can be found in the Supplementary Material." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 71, + 375, + 242 + ], + "blocks": [ + { + "bbox": [ + 220, + 71, + 375, + 242 + ], + "lines": [ + { + "bbox": [ + 220, + 71, + 375, + 242 + ], + "spans": [ + { + "bbox": [ + 220, + 71, + 375, + 242 + ], + "type": "image", + "image_path": "4323af32b5afb573ab02198424bdd0eafc7553ea8a8a1e152eadb2d808ab8471.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 72, + 539, + 242 + ], + "blocks": [ + { + "bbox": [ + 376, + 72, + 539, + 242 + ], + "lines": [ + { + "bbox": [ + 376, + 72, + 539, + 242 + ], + "spans": [ + { + "bbox": [ + 376, + 72, + 539, + 242 + ], + "type": "image", + "image_path": "a629b5e6beb130a8bf7929068fc02c3a867ae3eaee252c727d5a7860bfa4951d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 328, + 297, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 297, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 297, + 377 + ], + "type": "text", + "content": "noise the discrepancy is not noticeable, but as the noise is increased the discrepancy also increases. The performance improvement closes the gap to soft-Dice with the 0.5-threshold " + }, + { + "bbox": [ + 55, + 328, + 297, + 377 + ], + "type": "inline_equation", + "content": "\\mathrm{SD}^0" + }, + { + "bbox": [ + 55, + 328, + 297, + 377 + ], + "type": "text", + "content": " which is expected by the theory (23)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 391, + 126, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 126, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 126, + 403 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 508 + ], + "type": "text", + "content": "The results in this work follow the results presented by [2] in the sense that the soft-Dice loss outperforms the cross-entropy loss with respect to the Dice metric when the 0.5-threshold is used and that there seems to be a trend of bigger discrepancy when performance for both of the loss functions is low. The results also explain the experimental observations of alternative thresholds outperforming the standard 0.5-threshold that was presented in [4, 12]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 509, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 676 + ], + "type": "text", + "content": "In [24], it was proposed that the instability sometimes observed for soft-Dice may be associated with the fact that the loss function is non-convex (nor quasi-convex) when using sigmoid functions to enforce the bounds, that is " + }, + { + "bbox": [ + 55, + 509, + 296, + 676 + ], + "type": "inline_equation", + "content": "f \\mapsto \\mathrm{SD}_m(\\sigma \\circ f), f \\in \\mathcal{F}" + }, + { + "bbox": [ + 55, + 509, + 296, + 676 + ], + "type": "text", + "content": ". This is in contrast to cross-entropy where the corresponding " + }, + { + "bbox": [ + 55, + 509, + 296, + 676 + ], + "type": "inline_equation", + "content": "f \\mapsto \\mathrm{CE}_m(\\sigma \\circ f), f \\in \\mathcal{F}" + }, + { + "bbox": [ + 55, + 509, + 296, + 676 + ], + "type": "text", + "content": " is known to be convex. If cross-entropy with the alternative threshold yields the same performance as soft-Dice, then using it may provide the performance of soft-Dice without sacrificing the stability of cross-entropy. Another potential gain of using cross-entropy over soft-Dice is that marginal estimates are more informative than the approximate binary estimates that soft-Dice yields. These estimates can for instance be used to estimate uncertainty in different regions of a segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "A cost associated with using cross-entropy with the alternative threshold (22), is that the threshold needs to be computed for every prediction. This can be implemented ef" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 328, + 555, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 364 + ], + "type": "text", + "content": "ficiently but requires an extra step that may be non-negligible for sufficiently large images especially when there is a large number of predicted classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 375, + 388, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 375, + 388, + 387 + ], + "spans": [ + { + "bbox": [ + 314, + 375, + 388, + 387 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 395, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 556, + 598 + ], + "type": "text", + "content": "In this work, a realistic approximately volume preserving label noise model has been proposed and methods for explicit computation of marginal probabilities and sampling have been derived. Using the model, it has been shown how 0.5-thresholded optimal solutions to cross-entropy and soft-Dice are affected by various levels of noise, and how the resulting segmentations diverge as the noise is increased. It has further been shown how optimal solutions to soft-Dice can be recovered by thresholding the optimal solutions to cross-entropy with an a priori unknown but efficiently computable threshold generally different from 0.5. Finally, the theoretical observations are verified on data from the TotalSegmentor data set [37]. The presented results give insights into expected behaviour of widely used segmentation methods and suggest that cross-entropy with the alternative threshold yield the benefits of cross-entropy without sacrificing the performance of soft-Dice." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 613, + 556, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 556, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 556, + 672 + ], + "type": "text", + "content": "Limitations. To be able to systematically study the effect of label noise with varying strengths, a noise model is used for experiments. The scope of the experiments is consequently limited to the kind of noise that is produced by the model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 687, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 687, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 687, + 556, + 711 + ], + "type": "text", + "content": "Acknowledgements The authors were supported by Ray-Search Laboratories AB." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20827" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 704 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 167 + ], + "type": "text", + "content": "[1] Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): A Completed Reference Database of Lung Nodules on CT Scans. Medical Physics, 38(2):915–931, 2011. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 167, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 167, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 167, + 296, + 243 + ], + "type": "text", + "content": "[2] Jeroen Bertels, Tom Eelbode, Maxim Berman, Dirk Vandermeulen, Frederik Maes, Raf Bisschops, and Matthew B Blaschko. Optimizing the Dice Score and Jaccard Index for Medical Image Segmentation: Theory and Practice. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 92-100. Springer, 2019. 1, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 243, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 296, + 297 + ], + "type": "text", + "content": "[3] Jeroen Bertels, David Robben, Dirk Vandermeulen, and Paul Suetens. Theoretical Analysis and Experimental Validation of Volume Bias of Soft Dice Optimized Segmentation Maps in the Context of Inherent Uncertainty. Medical Image Analysis, 67:101833, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 297, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 361 + ], + "type": "text", + "content": "[4] Noah Bice, Neil Kirby, Ruiqi Li, Dan Nguyen, Tyler Bahr, Christopher Kabat, Pamela Myers, Niko Papanikolaou, and Mohammad Fakhreddine. A sensitivity analysis of probability maps in deep-learning-based anatomical segmentation. Journal of applied clinical medical physics, 22(8):105-119, 2021. 2, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 362, + 296, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 296, + 404 + ], + "type": "text", + "content": "[5] Pete Bridge, Andrew Fielding, Pamela Rowntree, and Andrew Pullar. Intraobserver Variability: Should We Worry? Journal of Medical Imaging and Radiation Sciences, 47(3):217-220, 2016. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 404, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 296, + 468 + ], + "type": "text", + "content": "[6] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation From Sparse Annotation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 424–432. Springer, 2016. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 469, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 296, + 512 + ], + "type": "text", + "content": "[7] Getao Du, Xu Cao, Jimin Liang, Xueli Chen, and Yonghua Zhan. Medical image segmentation based on u-net: A review. Journal of Imaging Science and Technology, 64:1-12, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 512, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 554 + ], + "type": "text", + "content": "[8] Charley Gros, Andreanne Lemay, and Julien Cohen-Adad. SoftSeg: Advantages of Soft Versus Binary Training for Image Segmentation. Medical Image Analysis, 71:102038, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 555, + 296, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 296, + 597 + ], + "type": "text", + "content": "[9] Yun Gu, Mali Shen, Jie Yang, and Guang-Zhong Yang. Reliable label-efficient learning for biomedical image recognition. IEEE Transactions on Biomedical Engineering, 66(9):2423-2432, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 597, + 296, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 597, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 57, + 597, + 296, + 652 + ], + "type": "text", + "content": "[10] Nicholas Heller, Joshua Dean, and Nikolaos Papanikolopoulos. Imperfect segmentation labels: How much do they matter? In Intravascular Imaging and Computer Assisted Stenting and Large-Scale Annotation of Biomedical Data and Expert Label Synthesis, pages 112-120. Springer, 2018. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 652, + 296, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 652, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 57, + 652, + 296, + 673 + ], + "type": "text", + "content": "[11] Takeyuki Hida and Masuyuki Hitsuda. Gaussian processes, volume 120. American Mathematical Soc., 1993. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 673, + 296, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 673, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 57, + 673, + 296, + 704 + ], + "type": "text", + "content": "[12] Burhan Rashid Hussein, Cédric Meurée, Malo Gaubert, Arthur Masson, Anne Kerbrat, Benoit Combès, and Francesca Galassi. A study on loss functions and decision thresholds" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 707 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "for the segmentation of multiple sclerosis lesions on spinal cord mri. hal preprint hal-03865212, 2022. 2, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 555, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 555, + 137 + ], + "type": "text", + "content": "[13] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 137, + 555, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 137, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 137, + 555, + 180 + ], + "type": "text", + "content": "[14] Davood Karimi, Haoran Dou, Simon K Warfield, and Ali Gholipour. Deep Learning With Noisy Labels: Exploring Techniques and Remedies in Medical Image Analysis. Medical Image Analysis, 65:101759, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 180, + 555, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 180, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 555, + 233 + ], + "type": "text", + "content": "[15] Eytan Kats, Jacob Goldberger, and Hayit Greenspan. Soft Labeling by Distilling Anatomical Knowledge for Improved MS Lesion Segmentation. In 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), pages 1563-1566. IEEE, 2019. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 234, + 555, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 555, + 266 + ], + "type": "text", + "content": "[16] Matthieu Lé, Jan Unkelbach, Nicholas Ayache, and Hervé Delingette. Sampling image segmentations for uncertainty quantification. Medical image analysis, 34:42-51, 2016. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 266, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 555, + 308 + ], + "type": "text", + "content": "[17] Andreanne Lemay, Charley Gros, and Julien Cohen-Adad. Label Fusion and Training Methods for Reliable Representation of Inter-Rater Uncertainty. arXiv preprint arXiv:2202.07550, 2022.2,3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 363 + ], + "type": "text", + "content": "[18] Hang Li, Dong Wei, Shilei Cao, Kai Ma, Liansheng Wang, and Yefeng Zheng. Superpixel-Guided Label Softening for Medical Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 227-237. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 363, + 555, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 555, + 417 + ], + "type": "text", + "content": "[19] Zachary C Lipton, Charles Elkan, and Balakrishnan Naryanaswamy. Optimal Thresholding of Classifiers to Maximize F1 Measure. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 225-239. Springer, 2014. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 417, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 555, + 459 + ], + "type": "text", + "content": "[20] Bingyuan Liu, Jose Dolz, Adrian Galdran, Riadh Kobbi, and Ismail Ben Ayed. Do we really need dice? the hidden region-size biases of segmentation losses. Medical Image Analysis, 91:103015, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 459, + 555, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 555, + 513 + ], + "type": "text", + "content": "[21] Alireza Mehrtash, William M Wells, Clare M Tempany, Purang Abolmaesumi, and Tina Kapur. Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging, 39(12):3868-3878, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 514, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 555, + 567 + ], + "type": "text", + "content": "[22] Zahra Mirikharaji, Yiqi Yan, and Ghassan Hamarneh. Learning to Segment Skin Lesions From Noisy Annotations. In Domain adaptation and representation transfer and medical image learning with less labels and imperfect data, pages 207-215. Springer, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 555, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 632 + ], + "type": "text", + "content": "[23] Guy Nir, Soheil Hor, Davood Karimi, Ladan Fazli, Brian F Skinnider, Peyman Tavassoli, Dmitry Turbin, Carlos F Villamil, Gang Wang, R Storey Wilson, et al. Automatic Grading of Prostate Cancer in Digitized Histopathology Images: Learning From Multiple Experts. Medical Image Analysis, 50:167-180, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 632, + 555, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 632, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 316, + 632, + 555, + 686 + ], + "type": "text", + "content": "[24] Marcus Nordström, Han Bao, Fredrik Lofman, Henrik Hult, Atsuto Maki, and Masashi Sugiyama. Calibrated Surrogate Maximization of Dice. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 269-278. Springer, 2020. 1, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 686, + 555, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 686, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 316, + 686, + 555, + 707 + ], + "type": "text", + "content": "[25] Marcus Nordstrom, Henrik Hult, Fredrik Lofman, and Jonas Söderberg. On image segmentation with noisy labels: Char" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20828" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "acterization and volume properties of the optimal solutions to accuracy and dice. Advances in Neural Information Processing Systems, 35:34321-34333, 2022. 2, 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 137 + ], + "type": "text", + "content": "[26] Marcus Nordstrom, Henrik Hult, Atsuto Maki, and Fredrik Lofman. Noisy image segmentation with soft-dice. arXiv preprint arXiv, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 296, + 201 + ], + "type": "text", + "content": "[27] Tufve Nyholm, Stina Svensson, Sebastian Andersson, Joakim Jonsson, Maja Sohlin, Christian Gustafsson, Elisabeth Kjellén, Karin Söderström, Per Albertsson, Lennart Blomqvist, et al. MR and CT Data With Multiobserver Delineations of Organs in the Pelvic Area—Part of the Gold Atlas Project. Medical Physics, 45(3):1295-1300, 2018. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 201, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 201, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 296, + 267 + ], + "type": "text", + "content": "[28] Teodora Popordanoska, Jeroen Bertels, Dirk Vandermeulen, Frederik Maes, and Matthew B Blaschko. On the Relationship Between Calibrated Predictors and Unbiased Volume Estimation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 678-688. Springer, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 267, + 296, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 296, + 320 + ], + "type": "text", + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional Networks for Biomedical Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241. Springer, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 374 + ], + "type": "text", + "content": "[30] Axel-Jan Rousseau, Thijs Becker, Jeroen Bertels, Matthew B Blaschko, and Dirk Valkenborg. Post Training Uncertainty Calibration of Deep Networks for Medical Image Segmentation. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1052-1056. IEEE, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 374, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 296, + 407 + ], + "type": "text", + "content": "[31] Gennady Samoradnitsky and Murad S. Taqqu. Stable non-Gaussian random processes: stochastic models with infinite variance. Chapman and Hall, 1994. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 407, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 296, + 449 + ], + "type": "text", + "content": "[32] Nahian Siddique, Sidike Paheding, Colin P Elkin, and Vijay Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. *IEEE Access*, 9:82031-82057, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 296, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 482 + ], + "type": "text", + "content": "[33] Joao Lourenço Silva and Arlindo L Oliveira. Using Soft Labels to Model Uncertainty in Medical Image Segmentation. arXiv preprint arXiv:2109.12622, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 296, + 534 + ], + "type": "text", + "content": "[34] Nima Tajbakhsh, Laura Jeyaseelan, Qian Li, Jeffrey N Chiang, Zhihao Wu, and Xiaowei Ding. Embracing Imperfect Datasets: A Review of Deep Learning Solutions for Medical Image Segmentation. Medical Image Analysis, 63:101693, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 296, + 610 + ], + "type": "text", + "content": "[35] Eugene Vorontsov and Samuel Kadoury. Label noise in segmentation networks: mitigation must deal with bias. In Deep Generative Models, and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 1, pages 251-258. Springer, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 296, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 296, + 665 + ], + "type": "text", + "content": "[36] Zifu Wang, Teodora Popordanoska, Jeroen Bertels, Robin Lemmens, and Matthew B Blaschko. Dice semimetric losses: Optimizing the dice score with soft labels. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 475-485. Springer, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 665, + 296, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 296, + 708 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 296, + 708 + ], + "type": "text", + "content": "[37] Jakob Wasserthal, Manfred Meyer, Hanns-Christian Breit, Joshy Cyriac, Shan Yang, and Martin Segeroth. Totalsegmentator: robust segmentation of 104 anatomical structures in ct images. arXiv preprint arXiv:2208.05868, 2022. 2, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 182 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "text", + "content": "[38] Shaode Yu, Mingli Chen, Erlei Zhang, Junjie Wu, Hang Yu, Zi Yang, Lin Ma, Xuejun Gu, and Weiguo Lu. Robustness Study of Noisy Annotation in Deep Learning Based Medical Image Segmentation. Physics in Medicine & Biology, 65(17):175007, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 127, + 555, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 127, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 127, + 555, + 182 + ], + "type": "text", + "content": "[39] Ming-Jie Zhao, Narayanan Edakunni, Adam Pocock, and Gavin Brown. Beyond Fano's Inequality: Bounds on the Optimal F-score, BER, and Cost-Sensitive Risk and Their Implications. The Journal of Machine Learning Research, 14(1):1033-1090, 2013. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20829" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/5856b279-35bc-44f9-b231-826fe1b3d735_content_list.json b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/5856b279-35bc-44f9-b231-826fe1b3d735_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e551552141914211069dac0061ea269079d12d5c --- /dev/null +++ b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/5856b279-35bc-44f9-b231-826fe1b3d735_content_list.json @@ -0,0 +1,1653 @@ +[ + { + "type": "text", + "text": "The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion", + "text_level": 1, + "bbox": [ + 158, + 130, + 810, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changan Chen* Juze Zhang* Shrinidhi K. Lakshmikanth* Yusu Fang Ruizhi Shao Gordon Wetzstein Li Fei-Fei Ehsan Adeli Stanford University", + "bbox": [ + 138, + 203, + 821, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eaf9dc24282cd37a53e4b1fb4c2c063f920cd82ca925261a42c5e27b1630f62c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 270, + 875, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c57ac4e75b795d65ef0e599448f3f4f9ea3874ddf4b74cbe91bc32fac8e57b3.jpg", + "image_caption": [ + "Figure 1. We introduce a language-model-based motion understanding and generation framework that takes in any of the audio/motion/text modalities and outputs the desired target modality. Coupled with our generative pre-training strategy, our model demonstrates competitive performance on an array of tasks, showing promising signs toward unified verbal and non-verbal language of human motions." + ], + "image_footnote": [], + "bbox": [ + 78, + 415, + 880, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 606, + 313, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human communication is inherently multimodal, involving a combination of verbal and non-verbal cues such as speech, facial expressions, and body gestures. Modeling these behaviors is essential for understanding human interaction and for creating virtual characters that can communicate naturally in applications like games, films, and virtual reality. However, existing motion generation models are typically limited to specific input modalities—either speech, text, or motion data—and cannot fully leverage the diversity of available data. In this paper, we propose a novel framework that unifies verbal and non-verbal language using multimodal language models for human motion understanding and generation. This model is flexible in taking text, speech, and motion or any combination of them as input. Coupled with our novel pre-training strategy, our model not only achieves state-of-the-art perfor", + "bbox": [ + 75, + 638, + 472, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mance on co-speech gesture generation but also requires much less data for training. Our model also unlocks an array of novel tasks such as editable gesture generation and emotion prediction from motion. We believe unifying the verbal and non-verbal language of human motion is essential for real-world applications, and language models offer a powerful approach to achieving this goal. Project page: languageofmotion.github.io.", + "bbox": [ + 500, + 608, + 892, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 755, + 630, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human communication is multimodal. We use spoken and body language, including hand gestures, facial expressions, body postures and emotional expressions to interact with each other effectively. For example, people use linguistic cues along with body language, including hand gestures, facial expressions, overall body posture, and even emotional expressions to interact with the environment effectively. Modeling these multimodal behaviors is essen", + "bbox": [ + 496, + 780, + 890, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*indicates equal contribution", + "bbox": [ + 94, + 887, + 250, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6200", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tial for understanding and generating human motion, enabling a wide range of applications for virtual characters in games, movies, and virtual reality—areas that have recently received substantial attention.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing work has been focused on modeling human motion from different modalities, such as speech [43, 48, 70], text [23, 64, 75], egocentric vision [30, 34], or the surrounding environment [5, 25, 63, 74, 77]. These models only take specific modalities as input, whose performance is thus limited to the data available for their downstream task. For example, co-speech gesture generation work typically trains speaker-dependent models [3, 19, 43, 70], which requires high-quality speech-motion capture of a person. While gesture style varies from person to person, many gestures are shared across people as well as non-speech-driven motion, such as walking or waving hands. Existing work has yet to leverage motion priors from all forms of motion data.", + "bbox": [ + 75, + 152, + 470, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One of the promising ways to unify different tasks is multimodal language models, where a single language model can take different modalities as input and output target modalities. These models have shown promising results in a wide range of multimodal tasks, such as visual question answering [2, 33, 42], audio understanding and generation [6, 15, 73], and text-to-motion generation [9, 10, 61, 75, 82]. While language models have been widely applied, it has not been explored in the speech-text-motion generation setting.", + "bbox": [ + 75, + 351, + 468, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We argue language models play a crucial role in unifying the verbal and non-verbal language of human motion for three reasons: 1) language models naturally connect different modalities, 2) speech is highly semantic, and tasks like modeling laughter in response to a joke require strong semantic reasoning capabilities and 3) language models are equipped with strong semantic understanding from extensive pre-training.", + "bbox": [ + 75, + 503, + 468, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards this goal, we propose a novel multimodal language model for expressional motion generation and understanding (see Fig. 1). To leverage language models to model motion, we first tokenize motion separately for different body parts (face, hand, upper-body, lower-body). Such compositionality has shown to be more beneficial to model the expressive human expressions [43, 48]. Along with off-the-shelf tokenizers for text and speech [28], we can represent any given modality inputs as a sequence of tokens, which are consumed by language models. To train the language models, we design a two-stage training pipeline. The model is first pre-trained to align various modalities with compositional body motion alignment and audio-text alignment. After pre-training, we compile downstream tasks into instructions and train the model on these instructions to allow the model to follow various task instructions.", + "bbox": [ + 75, + 627, + 468, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first validate our model on the BEATv2 co-speech gesture generation benchmark [43] and show that our model", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "strongly outperforms state-of-the-art models. We then conduct a thorough evaluation to demonstrate the effectiveness of our pre-training tasks. We also show that our pre-training strategy is more powerful when under severe data scarcity. While never seeing speech-motion data during pre-training, our model reaches competitive performance with a relatively small amount of data for a novel speaker, showing remarkable generalization. By performing post-training on both speech-motion and text-motion tasks, we show that our model not only follows audio and text prompts but also unlocks novel tasks such as predicting emotion from motion data. Please watch the Supp. video for the qualitative examples. To the best of our knowledge, this is the first work to build multimodal language models to unify the verbal and non-verbal language of 3D human motions.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 333, + 640, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Speech-Driven Motion Generation", + "text_level": 1, + "bbox": [ + 500, + 358, + 799, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human communication is multimodal and we use our speech, facial expressions, and body gestures to communicate with each other. Given this complimentary nature, recent work [1, 31] explores cross-modal generation of human motion from speech in different forms. These models are trained and evaluated on specific upper body joints, full body joints, and even facial expressions. Recent work in co-speech gesture generation often utilizes generative models to create gestures from audio conditions [8, 12, 43, 69]. Other work also explores the possibility of generating the listener's motion [46, 47, 59]. Another area of research focuses on generating speech-driven co-speech facial expressions, with notable works including ViCO [85] and CodeTalker [67]. However, these works are limited in that they only take speech as input and do not utilize other forms of motion data, making it challenging to follow both speech and textual cues. To tackle this, we propose to unify input/output modalities with a language model framework.", + "bbox": [ + 496, + 382, + 890, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Text to Motion Generation", + "text_level": 1, + "bbox": [ + 500, + 665, + 740, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Humans communicate with spoken language and nonverbal means such as emotions and interactions with surrounding environment [25, 63], among other cues. Recent work has explored generating human motion from text descriptions [4, 13, 16-18, 24, 26, 44, 52, 56, 62, 66, 76, 76, 83]. Some work attempts to generate human motion using diffusion models [11, 57, 58, 72, 78, 78, 79, 86] while other work exploring language models for generating human motion [9, 23, 38, 61, 64, 64, 75, 80, 82]. While these works have shown promising results in generating human motion from text instructions, they fall short in capturing the underlying meaning of the motion language itself. This limitation makes it challenging to develop a model capable of generating human motion from both verbal and non-verbal", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6201", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b8b427208a05acceec378bf9f1d63a6374446dca4ec6d0c438c914c991c791dd.jpg", + "image_caption": [ + "Figure 2. Method overview. We employ modality-specific tokenizers to process various input modalities. Specifically, we train a compositional body motion VQ-VAE to tokenize face, hands, upper body, and lower body motions into discrete tokens, combining these modality-specific vocabularies (audio and text) into a unified multimodal vocabulary. During training, mixed tokens from different modalities are used as input, and the output is generated through an encoder-decoder language model. The mixed tokens are fed into the transformer encoder, while the decoder predicts the probability distribution of the next token in an autoregressive manner at each step." + ], + "image_footnote": [], + "bbox": [ + 127, + 88, + 844, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "language. In this work, we propose a novel framework to capture patterns in body language and subtle expressive gestures inherently present in human communication.", + "bbox": [ + 75, + 416, + 470, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Multimodal Language Models", + "text_level": 1, + "bbox": [ + 76, + 470, + 344, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent years have witnessed the rise of language models [7, 14, 53, 54, 65, 84], primarily leveraging transformer architectures [60] that process text tokens as input and generate text tokens. Building upon these advancements, substantial efforts have expanded into multimodal language models capable of handling various types of input and output, with notable examples including BLIP-2 [32], LLaVA [42], and VideoChat [35]. Furthermore, the scope of multimodal language models (MM-LLMs) has broadened to include modality-specific outputs, as demonstrated by models like GILL [27] and SpeechGPT [73]. Efforts such as LLaVA [42] and AudioGPT [22] are advancing towards seamless any-to-any modality conversion, with the goal of emulating human-like cognitive abilities in multimodal contexts. Inspired by this line of work, we propose a new framework aimed at unifying verbal and non-verbal language within language models. Our framework takes text, speech, and motion data as input and generates human motion or text as output, further exploring the potential synergy between different tasks and modalities to enhance the performance of human motion generation.", + "bbox": [ + 75, + 494, + 470, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multimodal Language Model for Motion Generation and Understanding", + "text_level": 1, + "bbox": [ + 76, + 827, + 468, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present a multimodal language model for motion generation and understanding, which is illustrated", + "bbox": [ + 75, + 869, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in Fig. 2. We first describe the tokenization of different modalities (Sec. 3.2), then we introduce our generative pretraining for modality alignment (Sec. 3.3), and finally, we detail post-training for instructions following (Sec. 3.4).", + "bbox": [ + 496, + 415, + 893, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 489, + 638, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the neutral SMPL-X [51] body model including FLAME [37] face model. This model is parameterized by per-person body shape $\\beta \\in \\mathbb{R}^{T\\times 300}$ , 55 joint pose $\\mathbf{g}\\in \\mathbb{R}^{T\\times 55\\times 3}$ , facial expression $\\psi \\in \\mathbb{R}^{T\\times 100}$ , and global body translation $\\gamma \\in \\mathbb{R}^{T\\times 3}$ , where $T$ is the frame number.", + "bbox": [ + 496, + 513, + 892, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Tokenization", + "text_level": 1, + "bbox": [ + 500, + 603, + 635, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order for our modelt take various modalities as input (audio, tex,t and motion), we first tokenize different modalities with modality-specific tokenizers, and then combine them into a multimodal vocabulary.", + "bbox": [ + 496, + 627, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compositional body motion tokenization. Following the motion representation approach in EMAGE [43], we divide the body into four parts with 6D rotation representation: 9 joints forming the lower-body $\\mathbf{g}_l\\in \\mathbb{R}^{T\\times 54}$ , 13 joints forming the upper-body $\\mathbf{g}_u\\in \\mathbb{R}^{T\\times 78}$ , 30 joints forming the hands $\\mathbf{g}_h\\in \\mathbb{R}^{T\\times 180}$ and 1-joint along with 100 expression parameters representing the face $\\mathbf{g}_f\\in \\mathbb{R}^{T\\times 106}$ . Collectively, the motion space is represented as $G = \\{\\mathbf{g}_f,\\mathbf{g}_h,\\mathbf{g}_u,\\mathbf{g}_l\\}$ . Notably, we avoid using the commonly adopted HumanML3D representation [16] (H3D-Format) in text-to-motion tasks, as it predominantly focuses on skeletal movement, emphasizing swinging motions while overlooking twisting rotations of body parts—an essential aspect for effectively conveying body language. With this", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6202", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "compositional representation, we train four separate VQVAEs tockenize the body pose for each part. Each VQVAE encoder $\\mathcal{E}$ applies a four-layer temporal convolutional network (TCN) to extract continuous latent motion features $\\mathbf{z}^{1:T} = \\mathcal{E}(\\mathbf{g}^{1:T})$ . This encoded representation $\\mathbf{z}^{1:T}$ is quantized using:", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} ^ {t} = \\mathcal {Q} \\left(\\mathbf {z} ^ {t}\\right) := \\arg \\min _ {\\mathbf {q} ^ {k} \\in Q} \\| \\mathbf {z} ^ {t} - \\mathbf {q} ^ {k} \\| ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 189, + 468, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{q}^t$ is the discrete code in the codebook representing the encoded $\\mathbf{z}^t$ . Collectively, the quantized motion latent space is $Q = \\{\\mathbf{q}_f,\\mathbf{q}_h,\\mathbf{q}_u,\\mathbf{q}_l\\}$ . Each VQ-VAE decoder $\\mathcal{D}$ decodes the quantized motion $\\mathbf{q}^t$ back into the motion space $\\hat{\\mathbf{g}}^{1:T} = \\mathcal{D}(\\mathbf{q}^{1:T})$ and applies the following reconstruction losses:", + "bbox": [ + 75, + 223, + 468, + 314 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c}} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\mathcal {L} _ {\\text {v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\\\ \\mathcal {L} _ {\\mathrm {a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) + \\mathcal {L} _ {\\mathrm {m r e c}} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\tag {2} \\\\ \\mathcal {L} _ {\\mathrm {m v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\mathcal {L} _ {\\mathrm {m a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) + \\\\ \\mathcal {L} _ {\\mathrm {c o m m}} (\\mathbf {g}, \\mathbf {q}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 319, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{g}}^{\\prime}$ represent the reconstructed motion, $\\hat{\\mathbf{g}}^{\\prime}$ and $\\mathbf{g}^{\\prime}$ represent the velocity of $\\hat{\\mathbf{g}}$ and $\\mathbf{g}$ , while $\\hat{\\mathbf{g}}^{\\prime\\prime}$ and $\\mathbf{g}^{\\prime\\prime}$ represent their acceleration. For lower-body, upper-body and hands VQ-VAEs, pose reconstruction loss $\\mathcal{L}_{rec}$ is a Geodesic loss. For face VQ-VAE, $\\mathcal{L}_{rec}$ is $\\ell_{2}$ loss. Pose velocity/acceleration losses $\\mathcal{L}_{vel}$ and $\\mathcal{L}_{acc}$ are $\\ell_{1}$ losses. Mesh reconstruction loss $\\mathcal{L}_{mrec}$ is $\\ell_{2}$ loss. Mesh velocity/acceleration losses $\\mathcal{L}_{mvel}$ and $\\mathcal{L}_{macc}$ are $\\ell_{1}$ losses. Codebook commitment loss $\\mathcal{L}_{comm}$ is $\\ell_{2}$ loss. Vertices of the SMPLX-2020 mesh computed from the pose $\\mathbf{g}$ and $\\hat{\\mathbf{g}}$ are used to compute mesh losses.", + "bbox": [ + 75, + 401, + 468, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Speech tokenization. Similar to the motion modality, speech data is also continuous by nature. To facilitate speech training within a language model, we used HuBERT [20] to represent audio streams as discrete tokens. In this work, audio input is sampled at $16\\mathrm{kHz}$ , resulting in $\\mathbf{a} \\in \\mathbb{R}^{T \\times s}$ , where $s$ represents the audio frame rate after quantization. HuBERT further downsamples audio by a factor of 320, resulting in $s = 50$ . This frame rate, compared to the typical motion frame rate of 30 fps, provides an acceptable input token length for language models. The resulting audio token space is noted as $A = \\{\\mathbf{a}\\}$ .", + "bbox": [ + 75, + 568, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text tokenization. Following previous work [23, 54], we use SentencePiece [28] to tokenize text inputs and outputs into WordPiece tokens [29, 55] for the language model, with a vocabulary of 32,000 wordpieces inherited from the T5 [54] language model, which can be represented as $W = \\{\\mathbf{w}\\}$ . This vocabulary enables the model to process a fixed set of predetermined languages. Additionally, we extend the vocabulary with several multimodal tokens to support multi-modal inputs.", + "bbox": [ + 75, + 734, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal vocabulary. Altogether, we have a combined token space defined as $M := Q \\cup A \\cup W \\cup C =$", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/734e320712aa901c2edc4795e23f077389f2bbaf23dc81d851bb4d8260ea2ca5.jpg", + "image_caption": [ + "Figure 3. Illustration of pre-training. We pre-train our language model by translating one modality to another using paired data." + ], + "image_footnote": [], + "bbox": [ + 506, + 97, + 885, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\{\\mathbf{q}_f,\\mathbf{q}_u,\\mathbf{q}_h,\\mathbf{q}_l,\\mathbf{a},\\mathbf{w}\\}$ . Each modality-specific tokenizer outputs its modality-specific vocabulary. To build a unified language model that can process these different modalities, we need to combine these vocabularies into a joint vocabulary. Since the language model is pre-trained with the text modality, we choose to extend the original text vocabulary $V_{t} = \\{v_{t}^{i}\\}_{i = 1}^{K_{t}}$ with vocabularies from other modalities, including audio $V_{a} = \\{v_{a}^{i}\\}_{i = 1}^{K_{a}}$ , face $V_{f} = \\{v_{f}^{i}\\}_{i = 1}^{K_{f}}$ , hands $V_{h} = \\{v_{h}^{i}\\}_{i = 1}^{K_{h}}$ , upper body $V_{u} = \\{v_{u}^{i}\\}_{i = 1}^{K_{u}}$ , and lower body $V_{l} = \\{v_{l}^{i}\\}_{i = 1}^{K_{l}}$ , following previous work [23]. In particular, the motion vocabulary is defined as a combination of four body-part vocabularies: $V_{m} = \\{v_{f}^{i},v_{h}^{i},v_{u}^{i},v_{l}^{i}\\}_{i = 1}^{K_{m}}$ . Additionally, each modality-specific vocabulary includes special tokens for boundary recognition, such as $$ and $$ to indicate the start and end of an audio sequence. As a result, all modalities can be represented in a unified format with one joint multimodal vocabulary $V = \\{V_{t},V_{a},V_{f},V_{h},V_{u},V_{l}\\}$ .", + "bbox": [ + 498, + 223, + 890, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pre-training for Modality Alignment", + "text_level": 1, + "bbox": [ + 500, + 513, + 816, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing motion generation models rely heavily on paired data to train downstream tasks. Yet, collecting high-quality paired motion data is both costly and time consuming while there exists a large amount of unpaired data of each modality that can be explored. Inspired by this, we introduce our generative pre-training strategy, as shown in Fig. 3. More specifically, we implement two types of modality alignment during the pre-training stage: compositional motion alignment and audio-text alignment that are detailed below.", + "bbox": [ + 496, + 537, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compositional body motion alignment. Our body motion is inherently compositional, i.e., different body parts move in accordance. For example, when we are happy, our faces express smiles and our gestures tend to become more positive. The correlation between different body part motions is universal, transcending cultural boundaries. This shared prior forms the basis of our approach. To explore this correspondence, we consider two types of motion alignment tasks: spatial and temporal.", + "bbox": [ + 496, + 672, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial. To model the correlation between these different body parts, we train the model to take in a randomly selected combination of body parts (e.g., upper or upper + face) and predict another randomly selected combination of other body parts (e.g., lower or lower + hand). This helps our model learn the spatial relations between body parts.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6203", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Below is one example template that defines task prompts, conditions, and answers. The model takes both prompts and conditions as input and is expected to output the answer.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Prompts: Translate upper to lower body.", + "bbox": [ + 93, + 145, + 395, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditions: Upper Body Tokens $V_{\\text{condition}} = \\{v_u^i \\in V_u \\mid i \\in \\{\\text{sequence token index}\\} \\}$", + "bbox": [ + 93, + 161, + 452, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Answer: Lower Body Tokens $V_{\\text{Answer}} = \\{v_l^i \\in V_l \\mid i \\in \\{\\text{sequence token index}\\} \\}$", + "bbox": [ + 93, + 191, + 450, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal. Predicting how motion changes as a function of time is also an important self-supervision, which enables the model to capture the temporal evolution of motion. We model this by randomly masking off certain motion frames to help the model learn the temporal priors of motion.", + "bbox": [ + 76, + 233, + 468, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Prompts: Translate mask to unmasked motion.", + "bbox": [ + 93, + 318, + 436, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditions: Masked Tokens $V_{\\text{condition}} = \\{v_m^i \\in V_m \\mid i \\in \\{\\text{masked sequence token index}\\} \\}$ .", + "bbox": [ + 93, + 333, + 450, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Answer: Unmasked Motion Tokens $V_{\\text{Answer}} = \\{v_m^i \\in V \\mid i \\in \\{\\text{unmasked sequence token index}\\} \\}$", + "bbox": [ + 93, + 364, + 450, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Audio-text alignment. In addition to the motion modality, we also design translation tasks between audio and text modalities, leveraging the abundance of available data. These tasks follow the format of \"predicting modality Y from modality X\". For example, \"predicting text from audio\" should help the model's performance in \"predicting motion from audio\" by mapping the audio embeddings into the well-pre-trained text embedding space.", + "bbox": [ + 76, + 405, + 468, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Post-training with Instruction Following", + "text_level": 1, + "bbox": [ + 76, + 535, + 421, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After pre-training, the model gains an understanding of the underlying grammar and syntax within motion modality's vocabulary and good alignment between audio and text modalities. We then fine-tune the model with paired data on downstream tasks such as co-speech gesture generation or text-to-motion generation. To enable the model to perform desired downstream tasks while following natural human instructions, we construct a multi-task instruction-following template by formatting several key tasks such as audio-to-motion, text-to-motion, and emotion-to-motion into instructions. Specifically, for each task, we compose dozens of different instruction templates, resulting in more than one thousand different tasks, each having a unique instruction prompt. An example of our instruction template is shown below. See Supp. for more examples.", + "bbox": [ + 76, + 556, + 468, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Prompts: Based on $<$ AudioPLACEHOLDER $>$ , generate a full-body movement sequence involving face, hands, upper body, and lower body that matches the audio's rhythm.", + "bbox": [ + 91, + 792, + 452, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditions: Audio Tokens $V_{\\text{condition}} = \\{v_a^i \\in V_a \\mid i \\in \\{\\text{audio sequence token index}\\} \\}$", + "bbox": [ + 93, + 854, + 450, + 886 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Answer: Unmasked Motion Tokens $V_{\\text{Answer}} = \\{v_m^i \\in V \\mid i \\in \\{\\text{motion sequence token index}\\} \\}$", + "bbox": [ + 514, + 90, + 875, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Language Model Training Details", + "text_level": 1, + "bbox": [ + 500, + 143, + 794, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model leverages 220M pre-trained Flan-T5-Base model [54] with an encoder-decoder transformer structure to address the conditional generation task. Through our shared multimodal vocabulary $V$ , every input modality is represented as \"text\" tokens, allowing us to fully leverage the original T5 model for each conditional generation task. Specifically, with constructed modality latent codebook indices—such as index 8 of the upper body codebook—the upper body can be formatted as \"< upper_8 >\". Thus, the input can be converted into a sequence of tokens $S_{i} = \\{s_{i}^{k}\\}_{k=1}^{L}$ , where $s_{i} \\in V$ and $L$ represents the input length. Similarly, the model outputs a sequence of tokens $S_{o} = \\{s_{o}^{k}\\}_{k=1}^{L}$ , with a fixed input/output token length and $s_{o} \\in V$ .", + "bbox": [ + 498, + 166, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our model is encoder-decoder architecture, we set a maximum input length of 512. We specify modalities with start and stop tokens. Following the original T5 implementation, the sequence of tokens is sent to the encoder, and the decoder then performs next-token predictions in an auto-regressive manner at each step. The training objective can be formulated as follows:", + "bbox": [ + 498, + 364, + 890, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L M} = - \\sum_ {k = 0} ^ {L _ {t} - 1} \\log p _ {\\theta} \\left(s _ {t} ^ {k} \\mid s _ {t} ^ {< k}, s _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 476, + 890, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_t$ represents each token within the sequence, serving as the index in our unified vocabulary $V$ . Through next-token prediction, our model learns the underlying distribution of each modality, enabling the accurate and meaningful generation of target \"words\". For both pre-training and post-training, we finetune the model's entire weights instead of performing low-rank adaptation (LORA [21]) since our goal is to maximally align each modality.", + "bbox": [ + 498, + 527, + 890, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 661, + 632, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first evaluate our model on the co-speech gesture generation benchmark, then investigate the generalization enabled by generative multimodal pre-training, demonstrate our model's capability in following both audio and text prompts, and lastly our novel ability to predict emotion from motion.", + "bbox": [ + 498, + 686, + 890, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Co-Speech Gesture Generation", + "text_level": 1, + "bbox": [ + 500, + 786, + 772, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate our model's audio-to-motion generation ability, we choose to benchmark on co-speech gesture generation on the BEATv2 dataset [43], where the goal is to generate body gesture motion for a given speech of a speaker. Existing co-speech generation work typically models speaker-dependent gestures [43, 47, 70]. During the pre-training", + "bbox": [ + 498, + 809, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6204", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/71aaa66e2095b6032610e564134064e9d229d006d9df70a02f1048ff0fadfeb3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FGD ↓BC↑Diversity↑Condition Signal
DisCo [40]9.4176.4399.912audio
CaMN [41]6.6446.76910.86audio, text, facial
DiffStyleGesture [68]8.8117.24111.49audio, style
Habibie et al. [19]9.0407.7168.213audio, text
TalkSHOW [70]6.2096.94713.47audio
SynTalker [8]6.4137.97112.721audio, text
EMAGE [43]5.5127.72413.06audio, text
Ours w/o language pre-training7.4706.14814.162audio
Ours w/o multimodal pre-training5.4087.74214.418audio
Ours5.3017.78015.167audio
", + "bbox": [ + 192, + 88, + 777, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/568c3ffd2f1850f9d27e80617b905fce4075e211de655dc8f550e1c8bfdd65a8.jpg", + "table_caption": [ + "Table 1. Co-speech gesture generation results on BEATv2 benchmark. We report FGD $\\times 10^{-1}$ , BC $\\times 10^{-1}$ , Diversity, which measures realism, speech-motion synchronization, and diversity respectively. Our model outperforms state-of-the-art methods on this benchmark." + ], + "table_footnote": [], + "table_body": "
FGD ↓BC↑Diversity↑
W/o pre-training5.5017.72114.281
W/o A2T5.4437.72114.499
W/o spatial6.3367.38114.173
W/o temporal6.8007.34113.810
W/o motion7.7767.34414.640
Ours5.3017.78015.165
", + "bbox": [ + 102, + 321, + 444, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablations of pre-training.", + "bbox": [ + 169, + 454, + 375, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "stage, our model utilizes large-scale unpaired data in a self-supervised setup, drawing from two primary datasets, BEATv2 and Librispeech [49]. Together, these datasets provide approximately 1,000 hours of audio-to-text data and 60 hours of motion data. During the post-training, to ensure a fair comparison with baselines, we adopt the same evaluation protocol as [43], i.e., training and testing on speaker-2 and using their motion tokenizer. For pre-training, we ensure the model does not see any audio-to-motion data. Following prior work [43], we adopt Frechet Gesture Distance (FGD) [71] to evaluate the realism of the body gestures, Beat Correlation (BC) [36] to assess speech-motion synchrony and Diversity [31], which is calculated with the $\\ell_1$ distance between multiple body gesture clips.", + "bbox": [ + 75, + 488, + 468, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in Table 1. Compared with the state-of-the-art methods on this benchmark, our model achieves better performance across all metrics, indicating that our model generates more realistic, and diverse motion that is synchronized with the speech. Existing work often supplies additional signals to the model to boost the model's performance such as the text transcribed from the speech [19, 41, 43] or onset/amplitudes [43], partially due to the lack of semantic understanding of speech. In our model, since we use a pre-trained language model, the model naturally has a strong semantic understanding, allowing our model to show competitive performance without heavy reliance on hand-crafting features. If we use randomly ini", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tialized model weights, we can see that the model's performance drops drastically, indicating that language pretraining is vital for co-speech gesture generation. If we remove our multimodal pre-training stage, the model's performance also deteriorates, showing that our model benefits from the generative pre-training.", + "bbox": [ + 496, + 324, + 890, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further understand our model's performance, we show some qualitative results of our model in Fig. 4. We can see that our model generates gestures that are synchronized with the speech. See Supp. video for more examples.", + "bbox": [ + 496, + 415, + 890, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Effect of Generative Pre-training", + "text_level": 1, + "bbox": [ + 498, + 484, + 790, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generating gesture motion for a new speaker requires collecting high-quality motion data, typically from motion capture systems. Collecting such data is time-consuming. In this section, we first validate the importance of each pretraining task and then investigate whether our generative pre-training leads to better generalization on new speakers and thus reduces the amount of data required for training.", + "bbox": [ + 496, + 508, + 890, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Validating pre-training tasks. To understand how different pre-training objectives contribute to the performance, we ablate the audio-to-text alignment task (\"w/o A2T\"), the spatial body motion alignment task (\"w/o spatial\"), the temporal body motion alignment task (\"w/o temporal\") and the whole body alignment task (\"w/o motion\").", + "bbox": [ + 496, + 613, + 890, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in Table 1. \"w/o A2T\" lowers the model's performance, indicating that aligning the audio embedding space with text helps with semantic understanding and also the downstream gesture generation task. Removing either spatial motion prediction, temporal motion prediction or them altogether hurts the performance, showing that learning spatial-temporal motion priors in the pre-training stage is important for the downstream tasks.", + "bbox": [ + 496, + 704, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect on the training data. We hypothesize that our pre-training strategy captures strong multimodal correlation and motion priors, which could reduce the reliance on the amount of paired data for downstream tasks. To validate this hypothesis, we follow the setting in Sec. 4.1 and limit", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6205", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ae7cad739c0c87b2e6019e90ddeee7bf5951e08bb3828426da5bea469cb3619.jpg", + "image_caption": [ + "Figure 4. Qualitative example on co-speech gesture generation. Given a speech, we visualize the ground truth 3D motion accompanying the audio, the motion generated by the baseline EMAGE [43], SynTalker [8] and our method. Our model generates more diverse and expressive motion compared to the baseline, especially when the speaker emphasizes on certain words such as \"tired\" and \"because\"." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/409481e1a5cf52e072b5a289cbf0370dfd29c4057d1609813fba0f341de2db29.jpg", + "image_caption": [ + "Figure 5. Generation performance vs. the amount of post-training data. Our model learns a stronger motion prior from pre-training and thus shows much better under data scarcity." + ], + "image_footnote": [], + "bbox": [ + 89, + 555, + 455, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the amount of training data available to the model during the pre-training stage. Note that the model has never seen audio2motion data during pre-training. We set the amount of the data to $\\frac{1}{2^n}$ , $n \\in [1 \\dots 5]$ . We train our full model, our model without pre-training and EMAGE to convergence under each setting and evaluate on the same test set.", + "bbox": [ + 75, + 794, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are shown in Figure 5. We can see that our", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "full model starts with much lower FGD compared with the model without pre-training even when only using 1/32 of the paired training. As expected, as the amount of paired fine-tuning data increases, the performance reduces but our full model always outperforms the w/o pre-training ablation and EMAGE, showing that our model benefits pretraining and shows greater generalization under extreme data scarcity.", + "bbox": [ + 496, + 549, + 890, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Unifying Audio-to-Motion and Text-to-Motion for Editable Generation", + "text_level": 1, + "bbox": [ + 498, + 680, + 890, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By taking a language model-driven approach, our model is capable of following both audio and text prompts. We first train the motion tokenizer on both BEATv2 and AMASS [45] datasets since the range of motion in these two datasets is very different. We use the same tasks for pre-training. For post-training, we combine Audio2Motion and Text2Motion with various instructions, in which text-to-motion with HumanML3D [16] text annotations. See Supp. for details.", + "bbox": [ + 496, + 719, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By training on both text-to-motion and audio-to-motion data, our model supports joint audio-text prompts, enabling what we call editable gesture generation. This approach", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6206", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d6516e931f15e775591029844f872e3e2a387e25e957510fac33d1929c8e541.jpg", + "image_caption": [ + "Figure 6. Editable gesture generation. We prompt the language with text and audio information and it outputs motions that are both expressional gesture motion as well as general movement motion." + ], + "image_footnote": [], + "bbox": [ + 334, + 92, + 609, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17bd6f7f4152f5814b06b4976a577a72c3b435d66a9f14b1273789d7013af234.jpg", + "image_caption": [ + "Can you generate the lower body movement of someone sitting in a chair and generate the upper body movement according to the speech" + ], + "image_footnote": [], + "bbox": [ + 81, + 130, + 313, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1473874f67c949e1d81bc002c43fc50180be8aa7aaa77b441fad865cee215086.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 123, + 607, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d631a4a6ff6337bbdd1461b5ee428cbd0d97dd5701cd1e0ba1a2b3fddb7ab6cc.jpg", + "image_caption": [ + "Can you generate the lower body movement of a person walking forward and returning backward and generate the upper body movement according to the speech $\\leftarrow +\\uparrow$" + ], + "image_footnote": [], + "bbox": [ + 609, + 125, + 885, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What emotion is conveyed by the movements in the body motion?", + "bbox": [ + 119, + 320, + 410, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c66b72b5f6e047372197f600d470db08ca8c1c7e619013eccfb630b50f8de487.jpg", + "image_caption": [ + "MotionGPT", + "Ours", + "Figure 7. Qualitative example of emotion prediction." + ], + "image_footnote": [], + "bbox": [ + 153, + 351, + 398, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A person dances while moving their hips.", + "bbox": [ + 133, + 431, + 395, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The person expresses the emotion of \"contempt\"", + "bbox": [ + 130, + 474, + 431, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "facilitates the generation of synergistic full-body motions conditioned on both speech and flexibly chosen prompts. For instance, the model can generate the motion of a person walking while talking. In this work, we demonstrate this capability by prompting the model separately for specific body part motions and then combining them seamlessly. Combining conversation gestures with daily motions is extremely useful for applications such as gaming or VR. We show several qualitative examples in Fig. 6. We can see that the model can generate human motion that follows both audio and text prompts, showing the emergent capabilities of our model. See Supp. video for more examples.", + "bbox": [ + 75, + 532, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Predicting Emotion from Motion", + "text_level": 1, + "bbox": [ + 76, + 724, + 366, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model's flexibility in the input/output modality also unlocks an array of new tasks such as translation between different body parts or modalities. In this section, we propose a novel task that predicts emotion from motion.", + "bbox": [ + 75, + 750, + 468, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reading someone's body language, i.e., predicting emotion from motion is important for applications such as mental health or psychiatry, however, existing audio2motion or motion2text do not have this capability. We extract the emotion labels (neutral, anger, happiness, fear, disgust, sadness, contempt, and surprise) on BEATv2 and convert them", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d0ddc1f33cd418309f4c1164f87e925edeb276e0351ced2e456d6bde03f6f210.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Bleu@1↑Rouge Cider↑BertScore↑
GT10010099.9
Random2.454.440.19
MotionGPT1.6810.672.31
Ours14.7126.6716.94
", + "bbox": [ + 503, + 310, + 864, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Motion to emotion. We prompt our model to predict emotion given a motion sequence.", + "bbox": [ + 500, + 398, + 890, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "into instructions for training. To be compatible with arbitrary language output from MotionGPT, we evaluate the model performance by measuring the BLEU [50], Rouge Cider [39], and BERTScore [81] between the prediction and the ground truth, which measures the semantic distances between texts. See more details in Supp.", + "bbox": [ + 498, + 435, + 890, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are shown in Table 3. MotionGPT entirely fails this task with a performance similar to a random baseline because it was only trained to caption general motion rather than subtle gesture movement and body language. Our model outperforms the random and MotionGPT by a large margin, showing our model's ability to predict the emotion from motion. We also show one qualitative example in Fig. 7.", + "bbox": [ + 496, + 527, + 892, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 662, + 612, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel multimodal language model to unify verbal and non-verbal language with a novel pre-training objective. Our model not only shows state-of-the-art performance on co-speech gestures but also unlocks an array of novel tasks.", + "bbox": [ + 496, + 688, + 890, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While promising, the model sometimes fails to produce coherent motion potentially due to discrete motion tokenization. Moving forward, we believe incorporating continuous tokenization is an important step to improve the quality of the generated motion.", + "bbox": [ + 496, + 763, + 890, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We believe unifying verbal and non-verbal language of human motion generation and understanding is crucial for real-world applications, and language models provide a powerful framework to approach that goal.", + "bbox": [ + 496, + 840, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6207", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This project was partially funded by NIH grant R01AG089169 and UST. The authors would also like to thank Georgios Pavlakos for his valuable discussion, Chaitanya Patel, Jingyan Zhang, and Bin Li for their feedback on the paper.", + "bbox": [ + 76, + 90, + 472, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 179, + 171, + 194 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In ECCV, pages 248–265. Springer, 2020. 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 2", + "[3] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 2", + "[4] Nikos Athanasiou, Alpár Ceske, Markos Diomataris, Michael J Black, and Gül Varol. Motionfix: Text-driven 3d human motion editing. arXiv preprint arXiv:2408.00712, 2024. 2", + "[5] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In CVPR, pages 15935-15946, 2022. 2", + "[6] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: a language modeling approach to audio generation, 2023. 2", + "[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 3", + "[8] Bohong Chen, Yumeng Li, Yao-Xiang Ding, Tianjia Shao, and Kun Zhou. Enabling synergistic full-body control in prompt-based co-speech motion generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6774-6783, 2024. 2, 6, 7", + "[9] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos, 2024. 2" + ], + "bbox": [ + 86, + 204, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2", + "[11] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2", + "[12] Kiran Chhatre, Radek Daněček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J. Black, and Timo Bolkart. AMUSE: Emotional speech-driven 3D body animation via disentangled latent diffusion. In CVPR, pages 1942-1953, 2024. 2", + "[13] Seunggeun Chi, Hyung-gun Chi, Hengbo Ma, Nakul Agarwal, Faizan Siddiqui, Karthik Ramani, and Kwonjoon Lee. M2d2m: Multi-motion generation from text with discrete diffusion models. arXiv preprint arXiv:2407.14502, 2024. 2", + "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 3", + "[15] Yuan Gong, Hongyin Luo, Alexander H Liu, Leonid Karlinsky, and James Glass. Listen, think, and understand. arXiv preprint arXiv:2305.10790, 2023. 2", + "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, pages 5152-5161, 2022. 2, 3, 7", + "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597. Springer, 2022.", + "[18] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, pages 1900-1910, 2024. 2", + "[19] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video, 2021. 2, 6", + "[20] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29: 3451-3460, 2021. 4", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5", + "[22] Rongjie Huang, Mingze Li, Dongchao Yang, Jiatong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu, Zhiqing Hong, Jiawei Huang, Jinglin Liu, et al. Audiogpt: Understanding and generating speech, music, sound, and talking head. In AAAI, pages 23802-23804, 2024. 3", + "[23] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2, 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6208", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Biao Jiang, Xin Chen, Chi Zhang, Fukun Yin, Zhuoyuan Li, Gang Yu, and Jiayuan Fan. Motionchain: Conversational motion controllers via multimodal prompts. In ECCV, pages 54-74. Springer, 2025. 2", + "[25] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1737-1747, 2024. 2", + "[26] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, pages 2151-2162, 2023. 2", + "[27] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[28] T Kudo. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 2, 4", + "[29] Taku Kudo. Subword regularization: Improving neural network translation models with multiple subword candidates. arXiv preprint arXiv:1804.10959, 2018. 4", + "[30] Gen Li, Kaifeng Zhao, Siwei Zhang, Xiaozhong Lyu, Mihai Dusmanu, Yan Zhang, Marc Pollefeys, and Siyu Tang. EgoGen: An Egocentric Synthetic Data Generator. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[31] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 2, 6", + "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 3", + "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2", + "[34] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2", + "[35] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3", + "[36] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++, 2021. 6", + "[37] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Han Liang, Jiacheng Bao, Ruichi Zhang, Sihan Ren, Yuecheng Xu, Sibei Yang, Xin Chen, Jingyi Yu, and Lan Xu. Omg: Towards open-vocabulary motion generation via mixture of controllers. In CVPR, pages 482-493, 2024. 2", + "[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 8", + "[40] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3764-3773, 2022. 6", + "[41] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In ECCV, 2022. 6", + "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 3", + "[43] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J. Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In CVPR, 2024. 2, 3, 5, 6, 7", + "[44] Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-world text-to-motion generation. arXiv preprint arXiv:2312.14828, 2023. 2", + "[45] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 7", + "[46] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In CVPR, 2022. 2", + "[47] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In ICCV, 2023. 2, 5", + "[48] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations, 2024. 2", + "[49] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 6", + "[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 8", + "[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6209", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2", + "[53] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit Bermano, Eric Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. In Computer Graphics Forum, page e15063. Wiley Online Library, 2024. 3", + "[54] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 3, 4, 5", + "[55] Rico Sennrich. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 4", + "[56] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2", + "[57] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H. Bermano. Human motion diffusion model. In ICLR, 2022. 2", + "[58] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[59] Minh Tran, Di Chang, Maksim Siniukov, and Mohammad Soleymani. Dyadic interaction modeling for social behavior generation. In ECCV, 2024. 2", + "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 3", + "[61] Yuan Wang, Di Huang, Yaqi Zhang, Wanli Ouyang, Jile Jiao, Xuetao Feng, Yan Zhou, Pengfei Wan, Shixiang Tang, and Dan Xu. Motionpt-2: A general-purpose motion-language model for motion generation and understanding. arXiv, 2024. 2", + "[62] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 2", + "[63] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say, interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[64] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chi-Keung Tang. Motionllm: Multimodal motion-language learning with large language models, 2024. 2", + "[65] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024.3", + "[66] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "human motion generation. arXiv preprint arXiv:2310.08580, 2023. 2", + "[67] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In CVPR, pages 12780-12790, 2023. 2", + "[68] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffuse-style gesture: Stylized audio-driven co-speech gesture generation with diffusion models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 5860–5868. International Joint Conferences on Artificial Intelligence Organization, 2023. 6", + "[69] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2", + "[70] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J. Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 2, 5, 6", + "[71] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics, 39(6), 2020. 6", + "[72] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, pages 16010-16021, 2023. 2", + "[73] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities, 2023. 2, 3", + "[74] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Neuraldome: A neural modeling pipeline on multi-view human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8834-8845, 2023. 2", + "[75] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2", + "[76] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023. 2", + "[77] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, pages 516-526, 2024. 2", + "[78] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6210", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In ICCV, pages 364-373, 2023. 2", + "[80] Mingyuan Zhang, Daisheng Jin, Chenyang Gu, Fangzhou Hong, Zhongang Cai, Jingfang Huang, Chongzhi Zhang, Xinying Guo, Lei Yang, Ying He, et al. Large motion model for unified multi-modal motion generation. arXiv, 2024. 2", + "[81] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 8", + "[82] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, 2024. 2", + "[83] Zhikai Zhang, Yitang Li, Haofeng Huang, Mingxian Lin, and Li Yi. Freemotion: Mocap-free human motion synthesis with multimodal large language models. In ECCV, pages 403-421. Springer, 2025. 2", + "[84] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2024. 3", + "[85] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Responsive listening head generation: A benchmark dataset and baseline. In ECCV, 2022. 2", + "[86] Wenyang Zhou, Zhiyang Dou, Zeyu Cao, Zhouyingcheng Liao, Jingbo Wang, Wenjia Wang, Yuan Liu, Taku Komura, Wenping Wang, and Lingjie Liu. Emdm: Efficient motion diffusion model for fast and high-quality motion generation. In ECCV, pages 18-38. Springer, 2025. 2" + ], + "bbox": [ + 78, + 90, + 468, + 556 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6211", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/5856b279-35bc-44f9-b231-826fe1b3d735_model.json b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/5856b279-35bc-44f9-b231-826fe1b3d735_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7eee9dd4dc46f4fc7e1264bc9aff4db9fe432ab3 --- /dev/null +++ b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/5856b279-35bc-44f9-b231-826fe1b3d735_model.json @@ -0,0 +1,2589 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.131, + 0.812, + 0.177 + ], + "angle": 0, + "content": "The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.204, + 0.823, + 0.258 + ], + "angle": 0, + "content": "Changan Chen* Juze Zhang* Shrinidhi K. Lakshmikanth* Yusu Fang Ruizhi Shao Gordon Wetzstein Li Fei-Fei Ehsan Adeli Stanford University" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.271, + 0.877, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.416, + 0.882, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.545, + 0.893, + 0.586 + ], + "angle": 0, + "content": "Figure 1. We introduce a language-model-based motion understanding and generation framework that takes in any of the audio/motion/text modalities and outputs the desired target modality. Coupled with our generative pre-training strategy, our model demonstrates competitive performance on an array of tasks, showing promising signs toward unified verbal and non-verbal language of human motions." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.607, + 0.314, + 0.623 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.473, + 0.881 + ], + "angle": 0, + "content": "Human communication is inherently multimodal, involving a combination of verbal and non-verbal cues such as speech, facial expressions, and body gestures. Modeling these behaviors is essential for understanding human interaction and for creating virtual characters that can communicate naturally in applications like games, films, and virtual reality. However, existing motion generation models are typically limited to specific input modalities—either speech, text, or motion data—and cannot fully leverage the diversity of available data. In this paper, we propose a novel framework that unifies verbal and non-verbal language using multimodal language models for human motion understanding and generation. This model is flexible in taking text, speech, and motion or any combination of them as input. Coupled with our novel pre-training strategy, our model not only achieves state-of-the-art perfor" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.609, + 0.893, + 0.73 + ], + "angle": 0, + "content": "mance on co-speech gesture generation but also requires much less data for training. Our model also unlocks an array of novel tasks such as editable gesture generation and emotion prediction from motion. We believe unifying the verbal and non-verbal language of human motion is essential for real-world applications, and language models offer a powerful approach to achieving this goal. Project page: languageofmotion.github.io." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.756, + 0.631, + 0.771 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Human communication is multimodal. We use spoken and body language, including hand gestures, facial expressions, body postures and emotional expressions to interact with each other effectively. For example, people use linguistic cues along with body language, including hand gestures, facial expressions, overall body posture, and even emotional expressions to interact with the environment effectively. Modeling these multimodal behaviors is essen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.251, + 0.902 + ], + "angle": 0, + "content": "*indicates equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6200" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "tial for understanding and generating human motion, enabling a wide range of applications for virtual characters in games, movies, and virtual reality—areas that have recently received substantial attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.471, + 0.35 + ], + "angle": 0, + "content": "Existing work has been focused on modeling human motion from different modalities, such as speech [43, 48, 70], text [23, 64, 75], egocentric vision [30, 34], or the surrounding environment [5, 25, 63, 74, 77]. These models only take specific modalities as input, whose performance is thus limited to the data available for their downstream task. For example, co-speech gesture generation work typically trains speaker-dependent models [3, 19, 43, 70], which requires high-quality speech-motion capture of a person. While gesture style varies from person to person, many gestures are shared across people as well as non-speech-driven motion, such as walking or waving hands. Existing work has yet to leverage motion priors from all forms of motion data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.47, + 0.502 + ], + "angle": 0, + "content": "One of the promising ways to unify different tasks is multimodal language models, where a single language model can take different modalities as input and output target modalities. These models have shown promising results in a wide range of multimodal tasks, such as visual question answering [2, 33, 42], audio understanding and generation [6, 15, 73], and text-to-motion generation [9, 10, 61, 75, 82]. While language models have been widely applied, it has not been explored in the speech-text-motion generation setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.626 + ], + "angle": 0, + "content": "We argue language models play a crucial role in unifying the verbal and non-verbal language of human motion for three reasons: 1) language models naturally connect different modalities, 2) speech is highly semantic, and tasks like modeling laughter in response to a joke require strong semantic reasoning capabilities and 3) language models are equipped with strong semantic understanding from extensive pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Towards this goal, we propose a novel multimodal language model for expressional motion generation and understanding (see Fig. 1). To leverage language models to model motion, we first tokenize motion separately for different body parts (face, hand, upper-body, lower-body). Such compositionality has shown to be more beneficial to model the expressive human expressions [43, 48]. Along with off-the-shelf tokenizers for text and speech [28], we can represent any given modality inputs as a sequence of tokens, which are consumed by language models. To train the language models, we design a two-stage training pipeline. The model is first pre-trained to align various modalities with compositional body motion alignment and audio-text alignment. After pre-training, we compile downstream tasks into instructions and train the model on these instructions to allow the model to follow various task instructions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We first validate our model on the BEATv2 co-speech gesture generation benchmark [43] and show that our model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "strongly outperforms state-of-the-art models. We then conduct a thorough evaluation to demonstrate the effectiveness of our pre-training tasks. We also show that our pre-training strategy is more powerful when under severe data scarcity. While never seeing speech-motion data during pre-training, our model reaches competitive performance with a relatively small amount of data for a novel speaker, showing remarkable generalization. By performing post-training on both speech-motion and text-motion tasks, we show that our model not only follows audio and text prompts but also unlocks novel tasks such as predicting emotion from motion data. Please watch the Supp. video for the qualitative examples. To the best of our knowledge, this is the first work to build multimodal language models to unify the verbal and non-verbal language of 3D human motions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.334, + 0.642, + 0.349 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.359, + 0.8, + 0.374 + ], + "angle": 0, + "content": "2.1. Speech-Driven Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Human communication is multimodal and we use our speech, facial expressions, and body gestures to communicate with each other. Given this complimentary nature, recent work [1, 31] explores cross-modal generation of human motion from speech in different forms. These models are trained and evaluated on specific upper body joints, full body joints, and even facial expressions. Recent work in co-speech gesture generation often utilizes generative models to create gestures from audio conditions [8, 12, 43, 69]. Other work also explores the possibility of generating the listener's motion [46, 47, 59]. Another area of research focuses on generating speech-driven co-speech facial expressions, with notable works including ViCO [85] and CodeTalker [67]. However, these works are limited in that they only take speech as input and do not utilize other forms of motion data, making it challenging to follow both speech and textual cues. To tackle this, we propose to unify input/output modalities with a language model framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.741, + 0.681 + ], + "angle": 0, + "content": "2.2. Text to Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Humans communicate with spoken language and nonverbal means such as emotions and interactions with surrounding environment [25, 63], among other cues. Recent work has explored generating human motion from text descriptions [4, 13, 16-18, 24, 26, 44, 52, 56, 62, 66, 76, 76, 83]. Some work attempts to generate human motion using diffusion models [11, 57, 58, 72, 78, 78, 79, 86] while other work exploring language models for generating human motion [9, 23, 38, 61, 64, 64, 75, 80, 82]. While these works have shown promising results in generating human motion from text instructions, they fall short in capturing the underlying meaning of the motion language itself. This limitation makes it challenging to develop a model capable of generating human motion from both verbal and non-verbal" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6201" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.089, + 0.845, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.334, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 2. Method overview. We employ modality-specific tokenizers to process various input modalities. Specifically, we train a compositional body motion VQ-VAE to tokenize face, hands, upper body, and lower body motions into discrete tokens, combining these modality-specific vocabularies (audio and text) into a unified multimodal vocabulary. During training, mixed tokens from different modalities are used as input, and the output is generated through an encoder-decoder language model. The mixed tokens are fed into the transformer encoder, while the decoder predicts the probability distribution of the next token in an autoregressive manner at each step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.471, + 0.462 + ], + "angle": 0, + "content": "language. In this work, we propose a novel framework to capture patterns in body language and subtle expressive gestures inherently present in human communication." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.345, + 0.488 + ], + "angle": 0, + "content": "2.3. Multimodal Language Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.471, + 0.814 + ], + "angle": 0, + "content": "Recent years have witnessed the rise of language models [7, 14, 53, 54, 65, 84], primarily leveraging transformer architectures [60] that process text tokens as input and generate text tokens. Building upon these advancements, substantial efforts have expanded into multimodal language models capable of handling various types of input and output, with notable examples including BLIP-2 [32], LLaVA [42], and VideoChat [35]. Furthermore, the scope of multimodal language models (MM-LLMs) has broadened to include modality-specific outputs, as demonstrated by models like GILL [27] and SpeechGPT [73]. Efforts such as LLaVA [42] and AudioGPT [22] are advancing towards seamless any-to-any modality conversion, with the goal of emulating human-like cognitive abilities in multimodal contexts. Inspired by this line of work, we propose a new framework aimed at unifying verbal and non-verbal language within language models. Our framework takes text, speech, and motion data as input and generates human motion or text as output, further exploring the potential synergy between different tasks and modalities to enhance the performance of human motion generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.828, + 0.47, + 0.863 + ], + "angle": 0, + "content": "3. Multimodal Language Model for Motion Generation and Understanding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this section, we present a multimodal language model for motion generation and understanding, which is illustrated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.894, + 0.478 + ], + "angle": 0, + "content": "in Fig. 2. We first describe the tokenization of different modalities (Sec. 3.2), then we introduce our generative pretraining for modality alignment (Sec. 3.3), and finally, we detail post-training for instructions following (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.49, + 0.64, + 0.505 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.893, + 0.591 + ], + "angle": 0, + "content": "We use the neutral SMPL-X [51] body model including FLAME [37] face model. This model is parameterized by per-person body shape \\(\\beta \\in \\mathbb{R}^{T\\times 300}\\), 55 joint pose \\(\\mathbf{g}\\in \\mathbb{R}^{T\\times 55\\times 3}\\), facial expression \\(\\psi \\in \\mathbb{R}^{T\\times 100}\\), and global body translation \\(\\gamma \\in \\mathbb{R}^{T\\times 3}\\), where \\(T\\) is the frame number." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.637, + 0.618 + ], + "angle": 0, + "content": "3.2. Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.688 + ], + "angle": 0, + "content": "In order for our modelt take various modalities as input (audio, tex,t and motion), we first tokenize different modalities with modality-specific tokenizers, and then combine them into a multimodal vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Compositional body motion tokenization. Following the motion representation approach in EMAGE [43], we divide the body into four parts with 6D rotation representation: 9 joints forming the lower-body \\(\\mathbf{g}_l\\in \\mathbb{R}^{T\\times 54}\\), 13 joints forming the upper-body \\(\\mathbf{g}_u\\in \\mathbb{R}^{T\\times 78}\\), 30 joints forming the hands \\(\\mathbf{g}_h\\in \\mathbb{R}^{T\\times 180}\\) and 1-joint along with 100 expression parameters representing the face \\(\\mathbf{g}_f\\in \\mathbb{R}^{T\\times 106}\\). Collectively, the motion space is represented as \\(G = \\{\\mathbf{g}_f,\\mathbf{g}_h,\\mathbf{g}_u,\\mathbf{g}_l\\}\\). Notably, we avoid using the commonly adopted HumanML3D representation [16] (H3D-Format) in text-to-motion tasks, as it predominantly focuses on skeletal movement, emphasizing swinging motions while overlooking twisting rotations of body parts—an essential aspect for effectively conveying body language. With this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6202" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "compositional representation, we train four separate VQVAEs tockenize the body pose for each part. Each VQVAE encoder \\(\\mathcal{E}\\) applies a four-layer temporal convolutional network (TCN) to extract continuous latent motion features \\(\\mathbf{z}^{1:T} = \\mathcal{E}(\\mathbf{g}^{1:T})\\). This encoded representation \\(\\mathbf{z}^{1:T}\\) is quantized using:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.19, + 0.47, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} ^ {t} = \\mathcal {Q} \\left(\\mathbf {z} ^ {t}\\right) := \\arg \\min _ {\\mathbf {q} ^ {k} \\in Q} \\| \\mathbf {z} ^ {t} - \\mathbf {q} ^ {k} \\| ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.315 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}^t\\) is the discrete code in the codebook representing the encoded \\(\\mathbf{z}^t\\). Collectively, the quantized motion latent space is \\(Q = \\{\\mathbf{q}_f,\\mathbf{q}_h,\\mathbf{q}_u,\\mathbf{q}_l\\}\\). Each VQ-VAE decoder \\(\\mathcal{D}\\) decodes the quantized motion \\(\\mathbf{q}^t\\) back into the motion space \\(\\hat{\\mathbf{g}}^{1:T} = \\mathcal{D}(\\mathbf{q}^{1:T})\\) and applies the following reconstruction losses:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.32, + 0.469, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c}} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\mathcal {L} _ {\\text {v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\\\ \\mathcal {L} _ {\\mathrm {a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) + \\mathcal {L} _ {\\mathrm {m r e c}} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\tag {2} \\\\ \\mathcal {L} _ {\\mathrm {m v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\mathcal {L} _ {\\mathrm {m a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) + \\\\ \\mathcal {L} _ {\\mathrm {c o m m}} (\\mathbf {g}, \\mathbf {q}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.569 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{g}}^{\\prime}\\) represent the reconstructed motion, \\(\\hat{\\mathbf{g}}^{\\prime}\\) and \\(\\mathbf{g}^{\\prime}\\) represent the velocity of \\(\\hat{\\mathbf{g}}\\) and \\(\\mathbf{g}\\), while \\(\\hat{\\mathbf{g}}^{\\prime\\prime}\\) and \\(\\mathbf{g}^{\\prime\\prime}\\) represent their acceleration. For lower-body, upper-body and hands VQ-VAEs, pose reconstruction loss \\(\\mathcal{L}_{rec}\\) is a Geodesic loss. For face VQ-VAE, \\(\\mathcal{L}_{rec}\\) is \\(\\ell_{2}\\) loss. Pose velocity/acceleration losses \\(\\mathcal{L}_{vel}\\) and \\(\\mathcal{L}_{acc}\\) are \\(\\ell_{1}\\) losses. Mesh reconstruction loss \\(\\mathcal{L}_{mrec}\\) is \\(\\ell_{2}\\) loss. Mesh velocity/acceleration losses \\(\\mathcal{L}_{mvel}\\) and \\(\\mathcal{L}_{macc}\\) are \\(\\ell_{1}\\) losses. Codebook commitment loss \\(\\mathcal{L}_{comm}\\) is \\(\\ell_{2}\\) loss. Vertices of the SMPLX-2020 mesh computed from the pose \\(\\mathbf{g}\\) and \\(\\hat{\\mathbf{g}}\\) are used to compute mesh losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Speech tokenization. Similar to the motion modality, speech data is also continuous by nature. To facilitate speech training within a language model, we used HuBERT [20] to represent audio streams as discrete tokens. In this work, audio input is sampled at \\(16\\mathrm{kHz}\\), resulting in \\(\\mathbf{a} \\in \\mathbb{R}^{T \\times s}\\), where \\(s\\) represents the audio frame rate after quantization. HuBERT further downsamples audio by a factor of 320, resulting in \\(s = 50\\). This frame rate, compared to the typical motion frame rate of 30 fps, provides an acceptable input token length for language models. The resulting audio token space is noted as \\(A = \\{\\mathbf{a}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Text tokenization. Following previous work [23, 54], we use SentencePiece [28] to tokenize text inputs and outputs into WordPiece tokens [29, 55] for the language model, with a vocabulary of 32,000 wordpieces inherited from the T5 [54] language model, which can be represented as \\( W = \\{\\mathbf{w}\\} \\). This vocabulary enables the model to process a fixed set of predetermined languages. Additionally, we extend the vocabulary with several multimodal tokens to support multi-modal inputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Multimodal vocabulary. Altogether, we have a combined token space defined as \\( M := Q \\cup A \\cup W \\cup C = \\)" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.098, + 0.887, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.179, + 0.892, + 0.207 + ], + "angle": 0, + "content": "Figure 3. Illustration of pre-training. We pre-train our language model by translating one modality to another using paired data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.224, + 0.892, + 0.506 + ], + "angle": 0, + "content": "\\(\\{\\mathbf{q}_f,\\mathbf{q}_u,\\mathbf{q}_h,\\mathbf{q}_l,\\mathbf{a},\\mathbf{w}\\}\\). Each modality-specific tokenizer outputs its modality-specific vocabulary. To build a unified language model that can process these different modalities, we need to combine these vocabularies into a joint vocabulary. Since the language model is pre-trained with the text modality, we choose to extend the original text vocabulary \\(V_{t} = \\{v_{t}^{i}\\}_{i = 1}^{K_{t}}\\) with vocabularies from other modalities, including audio \\(V_{a} = \\{v_{a}^{i}\\}_{i = 1}^{K_{a}}\\), face \\(V_{f} = \\{v_{f}^{i}\\}_{i = 1}^{K_{f}}\\), hands \\(V_{h} = \\{v_{h}^{i}\\}_{i = 1}^{K_{h}}\\), upper body \\(V_{u} = \\{v_{u}^{i}\\}_{i = 1}^{K_{u}}\\), and lower body \\(V_{l} = \\{v_{l}^{i}\\}_{i = 1}^{K_{l}}\\), following previous work [23]. In particular, the motion vocabulary is defined as a combination of four body-part vocabularies: \\(V_{m} = \\{v_{f}^{i},v_{h}^{i},v_{u}^{i},v_{l}^{i}\\}_{i = 1}^{K_{m}}\\). Additionally, each modality-specific vocabulary includes special tokens for boundary recognition, such as \\(\\) and \\(\\) to indicate the start and end of an audio sequence. As a result, all modalities can be represented in a unified format with one joint multimodal vocabulary \\(V = \\{V_{t},V_{a},V_{f},V_{h},V_{u},V_{l}\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.818, + 0.531 + ], + "angle": 0, + "content": "3.3. Pre-training for Modality Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Existing motion generation models rely heavily on paired data to train downstream tasks. Yet, collecting high-quality paired motion data is both costly and time consuming while there exists a large amount of unpaired data of each modality that can be explored. Inspired by this, we introduce our generative pre-training strategy, as shown in Fig. 3. More specifically, we implement two types of modality alignment during the pre-training stage: compositional motion alignment and audio-text alignment that are detailed below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Compositional body motion alignment. Our body motion is inherently compositional, i.e., different body parts move in accordance. For example, when we are happy, our faces express smiles and our gestures tend to become more positive. The correlation between different body part motions is universal, transcending cultural boundaries. This shared prior forms the basis of our approach. To explore this correspondence, we consider two types of motion alignment tasks: spatial and temporal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Spatial. To model the correlation between these different body parts, we train the model to take in a randomly selected combination of body parts (e.g., upper or upper + face) and predict another randomly selected combination of other body parts (e.g., lower or lower + hand). This helps our model learn the spatial relations between body parts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6203" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Below is one example template that defines task prompts, conditions, and answers. The model takes both prompts and conditions as input and is expected to output the answer." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.146, + 0.397, + 0.162 + ], + "angle": 0, + "content": "Task Prompts: Translate upper to lower body." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.162, + 0.453, + 0.192 + ], + "angle": 0, + "content": "Conditions: Upper Body Tokens \\( V_{\\text{condition}} = \\{v_u^i \\in V_u \\mid i \\in \\{\\text{sequence token index}\\} \\} \\)" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.192, + 0.452, + 0.224 + ], + "angle": 0, + "content": "Answer: Lower Body Tokens \\( V_{\\text{Answer}} = \\{v_l^i \\in V_l \\mid i \\in \\{\\text{sequence token index}\\} \\} \\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.234, + 0.469, + 0.309 + ], + "angle": 0, + "content": "Temporal. Predicting how motion changes as a function of time is also an important self-supervision, which enables the model to capture the temporal evolution of motion. We model this by randomly masking off certain motion frames to help the model learn the temporal priors of motion." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.319, + 0.437, + 0.334 + ], + "angle": 0, + "content": "Task Prompts: Translate mask to unmasked motion." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.334, + 0.451, + 0.365 + ], + "angle": 0, + "content": "Conditions: Masked Tokens \\( V_{\\text{condition}} = \\{v_m^i \\in V_m \\mid i \\in \\{\\text{masked sequence token index}\\} \\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.365, + 0.451, + 0.396 + ], + "angle": 0, + "content": "Answer: Unmasked Motion Tokens \\( V_{\\text{Answer}} = \\{v_m^i \\in V \\mid i \\in \\{\\text{unmasked sequence token index}\\} \\} \\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.406, + 0.469, + 0.528 + ], + "angle": 0, + "content": "Audio-text alignment. In addition to the motion modality, we also design translation tasks between audio and text modalities, leveraging the abundance of available data. These tasks follow the format of \"predicting modality Y from modality X\". For example, \"predicting text from audio\" should help the model's performance in \"predicting motion from audio\" by mapping the audio embeddings into the well-pre-trained text embedding space." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.536, + 0.423, + 0.551 + ], + "angle": 0, + "content": "3.4. Post-training with Instruction Following" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.47, + 0.786 + ], + "angle": 0, + "content": "After pre-training, the model gains an understanding of the underlying grammar and syntax within motion modality's vocabulary and good alignment between audio and text modalities. We then fine-tune the model with paired data on downstream tasks such as co-speech gesture generation or text-to-motion generation. To enable the model to perform desired downstream tasks while following natural human instructions, we construct a multi-task instruction-following template by formatting several key tasks such as audio-to-motion, text-to-motion, and emotion-to-motion into instructions. Specifically, for each task, we compose dozens of different instruction templates, resulting in more than one thousand different tasks, each having a unique instruction prompt. An example of our instruction template is shown below. See Supp. for more examples." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.794, + 0.454, + 0.854 + ], + "angle": 0, + "content": "Task Prompts: Based on \\(<\\) AudioPLACEHOLDER \\(>\\), generate a full-body movement sequence involving face, hands, upper body, and lower body that matches the audio's rhythm." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.855, + 0.452, + 0.887 + ], + "angle": 0, + "content": "Conditions: Audio Tokens \\( V_{\\text{condition}} = \\{v_a^i \\in V_a \\mid i \\in \\{\\text{audio sequence token index}\\} \\} \\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.091, + 0.876, + 0.123 + ], + "angle": 0, + "content": "Answer: Unmasked Motion Tokens \\( V_{\\text{Answer}} = \\{v_m^i \\in V \\mid i \\in \\{\\text{motion sequence token index}\\} \\} \\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.144, + 0.795, + 0.161 + ], + "angle": 0, + "content": "3.5. Language Model Training Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.167, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Our model leverages 220M pre-trained Flan-T5-Base model [54] with an encoder-decoder transformer structure to address the conditional generation task. Through our shared multimodal vocabulary \\(V\\), every input modality is represented as \"text\" tokens, allowing us to fully leverage the original T5 model for each conditional generation task. Specifically, with constructed modality latent codebook indices—such as index 8 of the upper body codebook—the upper body can be formatted as \"< upper_8 >\". Thus, the input can be converted into a sequence of tokens \\(S_{i} = \\{s_{i}^{k}\\}_{k=1}^{L}\\), where \\(s_{i} \\in V\\) and \\(L\\) represents the input length. Similarly, the model outputs a sequence of tokens \\(S_{o} = \\{s_{o}^{k}\\}_{k=1}^{L}\\), with a fixed input/output token length and \\(s_{o} \\in V\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Since our model is encoder-decoder architecture, we set a maximum input length of 512. We specify modalities with start and stop tokens. Following the original T5 implementation, the sequence of tokens is sent to the encoder, and the decoder then performs next-token predictions in an auto-regressive manner at each step. The training objective can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.477, + 0.891, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L M} = - \\sum_ {k = 0} ^ {L _ {t} - 1} \\log p _ {\\theta} \\left(s _ {t} ^ {k} \\mid s _ {t} ^ {< k}, s _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.528, + 0.892, + 0.65 + ], + "angle": 0, + "content": "where \\( s_t \\) represents each token within the sequence, serving as the index in our unified vocabulary \\( V \\). Through next-token prediction, our model learns the underlying distribution of each modality, enabling the accurate and meaningful generation of target \"words\". For both pre-training and post-training, we finetune the model's entire weights instead of performing low-rank adaptation (LORA [21]) since our goal is to maximally align each modality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.662, + 0.633, + 0.679 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.688, + 0.892, + 0.777 + ], + "angle": 0, + "content": "In this section, we first evaluate our model on the co-speech gesture generation benchmark, then investigate the generalization enabled by generative multimodal pre-training, demonstrate our model's capability in following both audio and text prompts, and lastly our novel ability to predict emotion from motion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.774, + 0.803 + ], + "angle": 0, + "content": "4.1. Co-Speech Gesture Generation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To evaluate our model's audio-to-motion generation ability, we choose to benchmark on co-speech gesture generation on the BEATv2 dataset [43], where the goal is to generate body gesture motion for a given speech of a speaker. Existing co-speech generation work typically models speaker-dependent gestures [43, 47, 70]. During the pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6204" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.194, + 0.089, + 0.778, + 0.278 + ], + "angle": 0, + "content": "
FGD ↓BC↑Diversity↑Condition Signal
DisCo [40]9.4176.4399.912audio
CaMN [41]6.6446.76910.86audio, text, facial
DiffStyleGesture [68]8.8117.24111.49audio, style
Habibie et al. [19]9.0407.7168.213audio, text
TalkSHOW [70]6.2096.94713.47audio
SynTalker [8]6.4137.97112.721audio, text
EMAGE [43]5.5127.72413.06audio, text
Ours w/o language pre-training7.4706.14814.162audio
Ours w/o multimodal pre-training5.4087.74214.418audio
Ours5.3017.78015.167audio
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.288, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Table 1. Co-speech gesture generation results on BEATv2 benchmark. We report FGD \\(\\times 10^{-1}\\), BC \\(\\times 10^{-1}\\), Diversity, which measures realism, speech-motion synchronization, and diversity respectively. Our model outperforms state-of-the-art methods on this benchmark." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.323, + 0.445, + 0.445 + ], + "angle": 0, + "content": "
FGD ↓BC↑Diversity↑
W/o pre-training5.5017.72114.281
W/o A2T5.4437.72114.499
W/o spatial6.3367.38114.173
W/o temporal6.8007.34113.810
W/o motion7.7767.34414.640
Ours5.3017.78015.165
" + }, + { + "type": "table_caption", + "bbox": [ + 0.17, + 0.455, + 0.376, + 0.47 + ], + "angle": 0, + "content": "Table 2. Ablations of pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.7 + ], + "angle": 0, + "content": "stage, our model utilizes large-scale unpaired data in a self-supervised setup, drawing from two primary datasets, BEATv2 and Librispeech [49]. Together, these datasets provide approximately 1,000 hours of audio-to-text data and 60 hours of motion data. During the post-training, to ensure a fair comparison with baselines, we adopt the same evaluation protocol as [43], i.e., training and testing on speaker-2 and using their motion tokenizer. For pre-training, we ensure the model does not see any audio-to-motion data. Following prior work [43], we adopt Frechet Gesture Distance (FGD) [71] to evaluate the realism of the body gestures, Beat Correlation (BC) [36] to assess speech-motion synchrony and Diversity [31], which is calculated with the \\(\\ell_1\\) distance between multiple body gesture clips." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The results are shown in Table 1. Compared with the state-of-the-art methods on this benchmark, our model achieves better performance across all metrics, indicating that our model generates more realistic, and diverse motion that is synchronized with the speech. Existing work often supplies additional signals to the model to boost the model's performance such as the text transcribed from the speech [19, 41, 43] or onset/amplitudes [43], partially due to the lack of semantic understanding of speech. In our model, since we use a pre-trained language model, the model naturally has a strong semantic understanding, allowing our model to show competitive performance without heavy reliance on hand-crafting features. If we use randomly ini" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.415 + ], + "angle": 0, + "content": "tialized model weights, we can see that the model's performance drops drastically, indicating that language pretraining is vital for co-speech gesture generation. If we remove our multimodal pre-training stage, the model's performance also deteriorates, showing that our model benefits from the generative pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.477 + ], + "angle": 0, + "content": "To further understand our model's performance, we show some qualitative results of our model in Fig. 4. We can see that our model generates gestures that are synchronized with the speech. See Supp. video for more examples." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.485, + 0.791, + 0.502 + ], + "angle": 0, + "content": "4.2. Effect of Generative Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Generating gesture motion for a new speaker requires collecting high-quality motion data, typically from motion capture systems. Collecting such data is time-consuming. In this section, we first validate the importance of each pretraining task and then investigate whether our generative pre-training leads to better generalization on new speakers and thus reduces the amount of data required for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Validating pre-training tasks. To understand how different pre-training objectives contribute to the performance, we ablate the audio-to-text alignment task (\"w/o A2T\"), the spatial body motion alignment task (\"w/o spatial\"), the temporal body motion alignment task (\"w/o temporal\") and the whole body alignment task (\"w/o motion\")." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.825 + ], + "angle": 0, + "content": "The results are shown in Table 1. \"w/o A2T\" lowers the model's performance, indicating that aligning the audio embedding space with text helps with semantic understanding and also the downstream gesture generation task. Removing either spatial motion prediction, temporal motion prediction or them altogether hurts the performance, showing that learning spatial-temporal motion priors in the pre-training stage is important for the downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect on the training data. We hypothesize that our pre-training strategy captures strong multimodal correlation and motion priors, which could reduce the reliance on the amount of paired data for downstream tasks. To validate this hypothesis, we follow the setting in Sec. 4.1 and limit" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6205" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.495, + 0.893, + 0.539 + ], + "angle": 0, + "content": "Figure 4. Qualitative example on co-speech gesture generation. Given a speech, we visualize the ground truth 3D motion accompanying the audio, the motion generated by the baseline EMAGE [43], SynTalker [8] and our method. Our model generates more diverse and expressive motion compared to the baseline, especially when the speaker emphasizes on certain words such as \"tired\" and \"because\"." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.556, + 0.456, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Figure 5. Generation performance vs. the amount of post-training data. Our model learns a stronger motion prior from pre-training and thus shows much better under data scarcity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.884 + ], + "angle": 0, + "content": "the amount of training data available to the model during the pre-training stage. Note that the model has never seen audio2motion data during pre-training. We set the amount of the data to \\(\\frac{1}{2^n}\\), \\(n \\in [1 \\dots 5]\\). We train our full model, our model without pre-training and EMAGE to convergence under each setting and evaluate on the same test set." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The results are shown in Figure 5. We can see that our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.673 + ], + "angle": 0, + "content": "full model starts with much lower FGD compared with the model without pre-training even when only using 1/32 of the paired training. As expected, as the amount of paired fine-tuning data increases, the performance reduces but our full model always outperforms the w/o pre-training ablation and EMAGE, showing that our model benefits pretraining and shows greater generalization under extreme data scarcity." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.681, + 0.891, + 0.711 + ], + "angle": 0, + "content": "4.3. Unifying Audio-to-Motion and Text-to-Motion for Editable Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.854 + ], + "angle": 0, + "content": "By taking a language model-driven approach, our model is capable of following both audio and text prompts. We first train the motion tokenizer on both BEATv2 and AMASS [45] datasets since the range of motion in these two datasets is very different. We use the same tasks for pre-training. For post-training, we combine Audio2Motion and Text2Motion with various instructions, in which text-to-motion with HumanML3D [16] text annotations. See Supp. for details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "By training on both text-to-motion and audio-to-motion data, our model supports joint audio-text prompts, enabling what we call editable gesture generation. This approach" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6206" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.093, + 0.316, + 0.123 + ], + "angle": 0, + "content": "Can you generate the lower body movement of someone sitting in a chair and generate the upper body movement according to the speech" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.093, + 0.61, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.093, + 0.863, + 0.123 + ], + "angle": 0, + "content": "Can you generate the lower body movement of a person walking forward and returning backward and generate the upper body movement according to the speech \\(\\leftarrow +\\uparrow\\)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.131, + 0.314, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.124, + 0.609, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.125, + 0.887, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.893, + 0.302 + ], + "angle": 0, + "content": "Figure 6. Editable gesture generation. We prompt the language with text and audio information and it outputs motions that are both expressional gesture motion as well as general movement motion." + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.321, + 0.411, + 0.35 + ], + "angle": 0, + "content": "What emotion is conveyed by the movements in the body motion?" + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.352, + 0.4, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.416, + 0.191, + 0.426 + ], + "angle": 0, + "content": "MotionGPT" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.432, + 0.396, + 0.446 + ], + "angle": 0, + "content": "A person dances while moving their hips." + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.46, + 0.159, + 0.47 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.476, + 0.432, + 0.491 + ], + "angle": 0, + "content": "The person expresses the emotion of \"contempt\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.502, + 0.43, + 0.516 + ], + "angle": 0, + "content": "Figure 7. Qualitative example of emotion prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.716 + ], + "angle": 0, + "content": "facilitates the generation of synergistic full-body motions conditioned on both speech and flexibly chosen prompts. For instance, the model can generate the motion of a person walking while talking. In this work, we demonstrate this capability by prompting the model separately for specific body part motions and then combining them seamlessly. Combining conversation gestures with daily motions is extremely useful for applications such as gaming or VR. We show several qualitative examples in Fig. 6. We can see that the model can generate human motion that follows both audio and text prompts, showing the emergent capabilities of our model. See Supp. video for more examples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.367, + 0.743 + ], + "angle": 0, + "content": "4.4. Predicting Emotion from Motion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Our model's flexibility in the input/output modality also unlocks an array of new tasks such as translation between different body parts or modalities. In this section, we propose a novel task that predicts emotion from motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Reading someone's body language, i.e., predicting emotion from motion is important for applications such as mental health or psychiatry, however, existing audio2motion or motion2text do not have this capability. We extract the emotion labels (neutral, anger, happiness, fear, disgust, sadness, contempt, and surprise) on BEATv2 and convert them" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.311, + 0.866, + 0.397 + ], + "angle": 0, + "content": "
Bleu@1↑Rouge Cider↑BertScore↑
GT10010099.9
Random2.454.440.19
MotionGPT1.6810.672.31
Ours14.7126.6716.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.399, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Table 3. Motion to emotion. We prompt our model to predict emotion given a motion sequence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.436, + 0.892, + 0.527 + ], + "angle": 0, + "content": "into instructions for training. To be compatible with arbitrary language output from MotionGPT, we evaluate the model performance by measuring the BLEU [50], Rouge Cider [39], and BERTScore [81] between the prediction and the ground truth, which measures the semantic distances between texts. See more details in Supp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.893, + 0.649 + ], + "angle": 0, + "content": "The results are shown in Table 3. MotionGPT entirely fails this task with a performance similar to a random baseline because it was only trained to caption general motion rather than subtle gesture movement and body language. Our model outperforms the random and MotionGPT by a large margin, showing our model's ability to predict the emotion from motion. We also show one qualitative example in Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.614, + 0.679 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.764 + ], + "angle": 0, + "content": "In this work, we propose a novel multimodal language model to unify verbal and non-verbal language with a novel pre-training objective. Our model not only shows state-of-the-art performance on co-speech gestures but also unlocks an array of novel tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.84 + ], + "angle": 0, + "content": "While promising, the model sometimes fails to produce coherent motion potentially due to discrete motion tokenization. Moving forward, we believe incorporating continuous tokenization is an important step to improve the quality of the generated motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We believe unifying verbal and non-verbal language of human motion generation and understanding is crucial for real-world applications, and language models provide a powerful framework to approach that goal." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6207" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "Acknowledgments: This project was partially funded by NIH grant R01AG089169 and UST. The authors would also like to thank Georgios Pavlakos for his valuable discussion, Chaitanya Patel, Jingyan Zhang, and Bin Li for their feedback on the paper." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.18, + 0.173, + 0.195 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[1] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In ECCV, pages 248–265. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.472, + 0.412 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.413, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[3] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.456, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[4] Nikos Athanasiou, Alpár Ceske, Markos Diomataris, Michael J Black, and Gül Varol. Motionfix: Text-driven 3d human motion editing. arXiv preprint arXiv:2408.00712, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.47, + 0.567 + ], + "angle": 0, + "content": "[5] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In CVPR, pages 15935-15946, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.568, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[6] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: a language modeling approach to audio generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.638, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.776, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[8] Bohong Chen, Yumeng Li, Yao-Xiang Ding, Tianjia Shao, and Kun Zhou. Enabling synergistic full-body control in prompt-based co-speech motion generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6774-6783, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[9] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.205, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[10] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[11] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[12] Kiran Chhatre, Radek Daněček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J. Black, and Timo Bolkart. AMUSE: Emotional speech-driven 3D body animation via disentangled latent diffusion. In CVPR, pages 1942-1953, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[13] Seunggeun Chi, Hyung-gun Chi, Hengbo Ma, Nakul Agarwal, Faizan Siddiqui, Karthik Ramani, and Kwonjoon Lee. M2d2m: Multi-motion generation from text with discrete diffusion models. arXiv preprint arXiv:2407.14502, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[15] Yuan Gong, Hongyin Luo, Alexander H Liu, Leonid Karlinsky, and James Glass. Listen, think, and understand. arXiv preprint arXiv:2305.10790, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, pages 5152-5161, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[18] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, pages 1900-1910, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[19] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[20] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29: 3451-3460, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[22] Rongjie Huang, Mingze Li, Dongchao Yang, Jiatong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu, Zhiqing Hong, Jiawei Huang, Jinglin Liu, et al. Audiogpt: Understanding and generating speech, music, sound, and talking head. In AAAI, pages 23802-23804, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[23] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6208" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[24] Biao Jiang, Xin Chen, Chi Zhang, Fukun Yin, Zhuoyuan Li, Gang Yu, and Jiayuan Fan. Motionchain: Conversational motion controllers via multimodal prompts. In ECCV, pages 54-74. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[25] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1737-1747, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[26] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, pages 2151-2162, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.333 + ], + "angle": 0, + "content": "[27] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.376 + ], + "angle": 0, + "content": "[28] T Kudo. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.418 + ], + "angle": 0, + "content": "[29] Taku Kudo. Subword regularization: Improving neural network translation models with multiple subword candidates. arXiv preprint arXiv:1804.10959, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[30] Gen Li, Kaifeng Zhao, Siwei Zhang, Xiaozhong Lyu, Mihai Dusmanu, Yan Zhang, Marc Pollefeys, and Siyu Tang. EgoGen: An Egocentric Synthetic Data Generator. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[31] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.471, + 0.644 + ], + "angle": 0, + "content": "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.471, + 0.7 + ], + "angle": 0, + "content": "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.471, + 0.758 + ], + "angle": 0, + "content": "[34] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[35] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[36] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[37] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[38] Han Liang, Jiacheng Bao, Ruichi Zhang, Sihan Ren, Yuecheng Xu, Sibei Yang, Xin Chen, Jingyi Yu, and Lan Xu. Omg: Towards open-vocabulary motion generation via mixture of controllers. In CVPR, pages 482-493, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[40] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3764-3773, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[41] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[43] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J. Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In CVPR, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[44] Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-world text-to-motion generation. arXiv preprint arXiv:2312.14828, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[45] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[46] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[47] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In ICCV, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[48] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[49] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6209" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[53] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit Bermano, Eric Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. In Computer Graphics Forum, page e15063. Wiley Online Library, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[54] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[55] Rico Sennrich. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[56] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[57] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H. Bermano. Human motion diffusion model. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[58] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.511 + ], + "angle": 0, + "content": "[59] Minh Tran, Di Chang, Maksim Siniukov, and Mohammad Soleymani. Dyadic interaction modeling for social behavior generation. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.567 + ], + "angle": 0, + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[61] Yuan Wang, Di Huang, Yaqi Zhang, Wanli Ouyang, Jile Jiao, Xuetao Feng, Yan Zhou, Pengfei Wan, Shixiang Tang, and Dan Xu. Motionpt-2: A general-purpose motion-language model for motion generation and understanding. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[62] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.762 + ], + "angle": 0, + "content": "[63] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say, interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[64] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chi-Keung Tang. Motionllm: Multimodal motion-language learning with large language models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[65] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[66] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "human motion generation. arXiv preprint arXiv:2310.08580, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[67] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In CVPR, pages 12780-12790, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[68] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffuse-style gesture: Stylized audio-driven co-speech gesture generation with diffusion models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 5860–5868. International Joint Conferences on Artificial Intelligence Organization, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[69] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[70] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J. Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[71] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics, 39(6), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[72] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, pages 16010-16021, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[73] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[74] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Neuraldome: A neural modeling pipeline on multi-view human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8834-8845, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[75] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[76] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[77] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, pages 516-526, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[78] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6210" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[79] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In ICCV, pages 364-373, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[80] Mingyuan Zhang, Daisheng Jin, Chenyang Gu, Fangzhou Hong, Zhongang Cai, Jingfang Huang, Chongzhi Zhang, Xinying Guo, Lei Yang, Ying He, et al. Large motion model for unified multi-modal motion generation. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[81] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[82] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[83] Zhikai Zhang, Yitang Li, Haofeng Huang, Mingxian Lin, and Li Yi. Freemotion: Mocap-free human motion synthesis with multimodal large language models. In ECCV, pages 403-421. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[84] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.446, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[85] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Responsive listening head generation: A benchmark dataset and baseline. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.489, + 0.468, + 0.558 + ], + "angle": 0, + "content": "[86] Wenyang Zhou, Zhiyang Dou, Zeyu Cao, Zhouyingcheng Liao, Jingbo Wang, Wenjia Wang, Yuan Liu, Taku Komura, Wenping Wang, and Lingjie Liu. Emdm: Efficient motion diffusion model for fast and high-quality motion generation. In ECCV, pages 18-38. Springer, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "6211" + } + ] +] \ No newline at end of file diff --git a/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/5856b279-35bc-44f9-b231-826fe1b3d735_origin.pdf b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/5856b279-35bc-44f9-b231-826fe1b3d735_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0ad12815c62ddc8deb407be4c14f9bd2c8160826 --- /dev/null +++ b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/5856b279-35bc-44f9-b231-826fe1b3d735_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08531f79587220ec15373855a7d2c684ea967abca0f46d4b398f6b7e583264a1 +size 1189153 diff --git a/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/full.md b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a11371b670a9aa3ccab1d72d3d566d0d451f670f --- /dev/null +++ b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/full.md @@ -0,0 +1,336 @@ +# The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion + +Changan Chen* Juze Zhang* Shrinidhi K. Lakshmikanth* Yusu Fang Ruizhi Shao Gordon Wetzstein Li Fei-Fei Ehsan Adeli Stanford University + +![](images/eaf9dc24282cd37a53e4b1fb4c2c063f920cd82ca925261a42c5e27b1630f62c.jpg) + +![](images/1c57ac4e75b795d65ef0e599448f3f4f9ea3874ddf4b74cbe91bc32fac8e57b3.jpg) +Figure 1. We introduce a language-model-based motion understanding and generation framework that takes in any of the audio/motion/text modalities and outputs the desired target modality. Coupled with our generative pre-training strategy, our model demonstrates competitive performance on an array of tasks, showing promising signs toward unified verbal and non-verbal language of human motions. + +# Abstract + +Human communication is inherently multimodal, involving a combination of verbal and non-verbal cues such as speech, facial expressions, and body gestures. Modeling these behaviors is essential for understanding human interaction and for creating virtual characters that can communicate naturally in applications like games, films, and virtual reality. However, existing motion generation models are typically limited to specific input modalities—either speech, text, or motion data—and cannot fully leverage the diversity of available data. In this paper, we propose a novel framework that unifies verbal and non-verbal language using multimodal language models for human motion understanding and generation. This model is flexible in taking text, speech, and motion or any combination of them as input. Coupled with our novel pre-training strategy, our model not only achieves state-of-the-art perfor + +mance on co-speech gesture generation but also requires much less data for training. Our model also unlocks an array of novel tasks such as editable gesture generation and emotion prediction from motion. We believe unifying the verbal and non-verbal language of human motion is essential for real-world applications, and language models offer a powerful approach to achieving this goal. Project page: languageofmotion.github.io. + +# 1. Introduction + +Human communication is multimodal. We use spoken and body language, including hand gestures, facial expressions, body postures and emotional expressions to interact with each other effectively. For example, people use linguistic cues along with body language, including hand gestures, facial expressions, overall body posture, and even emotional expressions to interact with the environment effectively. Modeling these multimodal behaviors is essen + +tial for understanding and generating human motion, enabling a wide range of applications for virtual characters in games, movies, and virtual reality—areas that have recently received substantial attention. + +Existing work has been focused on modeling human motion from different modalities, such as speech [43, 48, 70], text [23, 64, 75], egocentric vision [30, 34], or the surrounding environment [5, 25, 63, 74, 77]. These models only take specific modalities as input, whose performance is thus limited to the data available for their downstream task. For example, co-speech gesture generation work typically trains speaker-dependent models [3, 19, 43, 70], which requires high-quality speech-motion capture of a person. While gesture style varies from person to person, many gestures are shared across people as well as non-speech-driven motion, such as walking or waving hands. Existing work has yet to leverage motion priors from all forms of motion data. + +One of the promising ways to unify different tasks is multimodal language models, where a single language model can take different modalities as input and output target modalities. These models have shown promising results in a wide range of multimodal tasks, such as visual question answering [2, 33, 42], audio understanding and generation [6, 15, 73], and text-to-motion generation [9, 10, 61, 75, 82]. While language models have been widely applied, it has not been explored in the speech-text-motion generation setting. + +We argue language models play a crucial role in unifying the verbal and non-verbal language of human motion for three reasons: 1) language models naturally connect different modalities, 2) speech is highly semantic, and tasks like modeling laughter in response to a joke require strong semantic reasoning capabilities and 3) language models are equipped with strong semantic understanding from extensive pre-training. + +Towards this goal, we propose a novel multimodal language model for expressional motion generation and understanding (see Fig. 1). To leverage language models to model motion, we first tokenize motion separately for different body parts (face, hand, upper-body, lower-body). Such compositionality has shown to be more beneficial to model the expressive human expressions [43, 48]. Along with off-the-shelf tokenizers for text and speech [28], we can represent any given modality inputs as a sequence of tokens, which are consumed by language models. To train the language models, we design a two-stage training pipeline. The model is first pre-trained to align various modalities with compositional body motion alignment and audio-text alignment. After pre-training, we compile downstream tasks into instructions and train the model on these instructions to allow the model to follow various task instructions. + +We first validate our model on the BEATv2 co-speech gesture generation benchmark [43] and show that our model + +strongly outperforms state-of-the-art models. We then conduct a thorough evaluation to demonstrate the effectiveness of our pre-training tasks. We also show that our pre-training strategy is more powerful when under severe data scarcity. While never seeing speech-motion data during pre-training, our model reaches competitive performance with a relatively small amount of data for a novel speaker, showing remarkable generalization. By performing post-training on both speech-motion and text-motion tasks, we show that our model not only follows audio and text prompts but also unlocks novel tasks such as predicting emotion from motion data. Please watch the Supp. video for the qualitative examples. To the best of our knowledge, this is the first work to build multimodal language models to unify the verbal and non-verbal language of 3D human motions. + +# 2. Related Work + +# 2.1. Speech-Driven Motion Generation + +Human communication is multimodal and we use our speech, facial expressions, and body gestures to communicate with each other. Given this complimentary nature, recent work [1, 31] explores cross-modal generation of human motion from speech in different forms. These models are trained and evaluated on specific upper body joints, full body joints, and even facial expressions. Recent work in co-speech gesture generation often utilizes generative models to create gestures from audio conditions [8, 12, 43, 69]. Other work also explores the possibility of generating the listener's motion [46, 47, 59]. Another area of research focuses on generating speech-driven co-speech facial expressions, with notable works including ViCO [85] and CodeTalker [67]. However, these works are limited in that they only take speech as input and do not utilize other forms of motion data, making it challenging to follow both speech and textual cues. To tackle this, we propose to unify input/output modalities with a language model framework. + +# 2.2. Text to Motion Generation + +Humans communicate with spoken language and nonverbal means such as emotions and interactions with surrounding environment [25, 63], among other cues. Recent work has explored generating human motion from text descriptions [4, 13, 16-18, 24, 26, 44, 52, 56, 62, 66, 76, 76, 83]. Some work attempts to generate human motion using diffusion models [11, 57, 58, 72, 78, 78, 79, 86] while other work exploring language models for generating human motion [9, 23, 38, 61, 64, 64, 75, 80, 82]. While these works have shown promising results in generating human motion from text instructions, they fall short in capturing the underlying meaning of the motion language itself. This limitation makes it challenging to develop a model capable of generating human motion from both verbal and non-verbal + +![](images/b8b427208a05acceec378bf9f1d63a6374446dca4ec6d0c438c914c991c791dd.jpg) +Figure 2. Method overview. We employ modality-specific tokenizers to process various input modalities. Specifically, we train a compositional body motion VQ-VAE to tokenize face, hands, upper body, and lower body motions into discrete tokens, combining these modality-specific vocabularies (audio and text) into a unified multimodal vocabulary. During training, mixed tokens from different modalities are used as input, and the output is generated through an encoder-decoder language model. The mixed tokens are fed into the transformer encoder, while the decoder predicts the probability distribution of the next token in an autoregressive manner at each step. + +language. In this work, we propose a novel framework to capture patterns in body language and subtle expressive gestures inherently present in human communication. + +# 2.3. Multimodal Language Models + +Recent years have witnessed the rise of language models [7, 14, 53, 54, 65, 84], primarily leveraging transformer architectures [60] that process text tokens as input and generate text tokens. Building upon these advancements, substantial efforts have expanded into multimodal language models capable of handling various types of input and output, with notable examples including BLIP-2 [32], LLaVA [42], and VideoChat [35]. Furthermore, the scope of multimodal language models (MM-LLMs) has broadened to include modality-specific outputs, as demonstrated by models like GILL [27] and SpeechGPT [73]. Efforts such as LLaVA [42] and AudioGPT [22] are advancing towards seamless any-to-any modality conversion, with the goal of emulating human-like cognitive abilities in multimodal contexts. Inspired by this line of work, we propose a new framework aimed at unifying verbal and non-verbal language within language models. Our framework takes text, speech, and motion data as input and generates human motion or text as output, further exploring the potential synergy between different tasks and modalities to enhance the performance of human motion generation. + +# 3. Multimodal Language Model for Motion Generation and Understanding + +In this section, we present a multimodal language model for motion generation and understanding, which is illustrated + +in Fig. 2. We first describe the tokenization of different modalities (Sec. 3.2), then we introduce our generative pretraining for modality alignment (Sec. 3.3), and finally, we detail post-training for instructions following (Sec. 3.4). + +# 3.1. Preliminaries + +We use the neutral SMPL-X [51] body model including FLAME [37] face model. This model is parameterized by per-person body shape $\beta \in \mathbb{R}^{T\times 300}$ , 55 joint pose $\mathbf{g}\in \mathbb{R}^{T\times 55\times 3}$ , facial expression $\psi \in \mathbb{R}^{T\times 100}$ , and global body translation $\gamma \in \mathbb{R}^{T\times 3}$ , where $T$ is the frame number. + +# 3.2. Tokenization + +In order for our modelt take various modalities as input (audio, tex,t and motion), we first tokenize different modalities with modality-specific tokenizers, and then combine them into a multimodal vocabulary. + +Compositional body motion tokenization. Following the motion representation approach in EMAGE [43], we divide the body into four parts with 6D rotation representation: 9 joints forming the lower-body $\mathbf{g}_l\in \mathbb{R}^{T\times 54}$ , 13 joints forming the upper-body $\mathbf{g}_u\in \mathbb{R}^{T\times 78}$ , 30 joints forming the hands $\mathbf{g}_h\in \mathbb{R}^{T\times 180}$ and 1-joint along with 100 expression parameters representing the face $\mathbf{g}_f\in \mathbb{R}^{T\times 106}$ . Collectively, the motion space is represented as $G = \{\mathbf{g}_f,\mathbf{g}_h,\mathbf{g}_u,\mathbf{g}_l\}$ . Notably, we avoid using the commonly adopted HumanML3D representation [16] (H3D-Format) in text-to-motion tasks, as it predominantly focuses on skeletal movement, emphasizing swinging motions while overlooking twisting rotations of body parts—an essential aspect for effectively conveying body language. With this + +compositional representation, we train four separate VQVAEs tockenize the body pose for each part. Each VQVAE encoder $\mathcal{E}$ applies a four-layer temporal convolutional network (TCN) to extract continuous latent motion features $\mathbf{z}^{1:T} = \mathcal{E}(\mathbf{g}^{1:T})$ . This encoded representation $\mathbf{z}^{1:T}$ is quantized using: + +$$ +\mathbf {q} ^ {t} = \mathcal {Q} \left(\mathbf {z} ^ {t}\right) := \arg \min _ {\mathbf {q} ^ {k} \in Q} \| \mathbf {z} ^ {t} - \mathbf {q} ^ {k} \| ^ {2} \tag {1} +$$ + +where $\mathbf{q}^t$ is the discrete code in the codebook representing the encoded $\mathbf{z}^t$ . Collectively, the quantized motion latent space is $Q = \{\mathbf{q}_f,\mathbf{q}_h,\mathbf{q}_u,\mathbf{q}_l\}$ . Each VQ-VAE decoder $\mathcal{D}$ decodes the quantized motion $\mathbf{q}^t$ back into the motion space $\hat{\mathbf{g}}^{1:T} = \mathcal{D}(\mathbf{q}^{1:T})$ and applies the following reconstruction losses: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {r e c}} (\mathbf {g}, \hat {\mathbf {g}}) + \mathcal {L} _ {\text {v e l}} \left(\mathbf {g} ^ {\prime}, \hat {\mathbf {g}} ^ {\prime}\right) + \\ \mathcal {L} _ {\mathrm {a c c}} \left(\mathbf {g} ^ {\prime \prime}, \hat {\mathbf {g}} ^ {\prime \prime}\right) + \mathcal {L} _ {\mathrm {m r e c}} (\mathbf {g}, \hat {\mathbf {g}}) + \tag {2} \\ \mathcal {L} _ {\mathrm {m v e l}} \left(\mathbf {g} ^ {\prime}, \hat {\mathbf {g}} ^ {\prime}\right) + \mathcal {L} _ {\mathrm {m a c c}} \left(\mathbf {g} ^ {\prime \prime}, \hat {\mathbf {g}} ^ {\prime \prime}\right) + \\ \mathcal {L} _ {\mathrm {c o m m}} (\mathbf {g}, \mathbf {q}), \\ \end{array} +$$ + +where $\hat{\mathbf{g}}^{\prime}$ represent the reconstructed motion, $\hat{\mathbf{g}}^{\prime}$ and $\mathbf{g}^{\prime}$ represent the velocity of $\hat{\mathbf{g}}$ and $\mathbf{g}$ , while $\hat{\mathbf{g}}^{\prime\prime}$ and $\mathbf{g}^{\prime\prime}$ represent their acceleration. For lower-body, upper-body and hands VQ-VAEs, pose reconstruction loss $\mathcal{L}_{rec}$ is a Geodesic loss. For face VQ-VAE, $\mathcal{L}_{rec}$ is $\ell_{2}$ loss. Pose velocity/acceleration losses $\mathcal{L}_{vel}$ and $\mathcal{L}_{acc}$ are $\ell_{1}$ losses. Mesh reconstruction loss $\mathcal{L}_{mrec}$ is $\ell_{2}$ loss. Mesh velocity/acceleration losses $\mathcal{L}_{mvel}$ and $\mathcal{L}_{macc}$ are $\ell_{1}$ losses. Codebook commitment loss $\mathcal{L}_{comm}$ is $\ell_{2}$ loss. Vertices of the SMPLX-2020 mesh computed from the pose $\mathbf{g}$ and $\hat{\mathbf{g}}$ are used to compute mesh losses. + +Speech tokenization. Similar to the motion modality, speech data is also continuous by nature. To facilitate speech training within a language model, we used HuBERT [20] to represent audio streams as discrete tokens. In this work, audio input is sampled at $16\mathrm{kHz}$ , resulting in $\mathbf{a} \in \mathbb{R}^{T \times s}$ , where $s$ represents the audio frame rate after quantization. HuBERT further downsamples audio by a factor of 320, resulting in $s = 50$ . This frame rate, compared to the typical motion frame rate of 30 fps, provides an acceptable input token length for language models. The resulting audio token space is noted as $A = \{\mathbf{a}\}$ . + +Text tokenization. Following previous work [23, 54], we use SentencePiece [28] to tokenize text inputs and outputs into WordPiece tokens [29, 55] for the language model, with a vocabulary of 32,000 wordpieces inherited from the T5 [54] language model, which can be represented as $W = \{\mathbf{w}\}$ . This vocabulary enables the model to process a fixed set of predetermined languages. Additionally, we extend the vocabulary with several multimodal tokens to support multi-modal inputs. + +Multimodal vocabulary. Altogether, we have a combined token space defined as $M := Q \cup A \cup W \cup C =$ + +![](images/734e320712aa901c2edc4795e23f077389f2bbaf23dc81d851bb4d8260ea2ca5.jpg) +Figure 3. Illustration of pre-training. We pre-train our language model by translating one modality to another using paired data. + +$\{\mathbf{q}_f,\mathbf{q}_u,\mathbf{q}_h,\mathbf{q}_l,\mathbf{a},\mathbf{w}\}$ . Each modality-specific tokenizer outputs its modality-specific vocabulary. To build a unified language model that can process these different modalities, we need to combine these vocabularies into a joint vocabulary. Since the language model is pre-trained with the text modality, we choose to extend the original text vocabulary $V_{t} = \{v_{t}^{i}\}_{i = 1}^{K_{t}}$ with vocabularies from other modalities, including audio $V_{a} = \{v_{a}^{i}\}_{i = 1}^{K_{a}}$ , face $V_{f} = \{v_{f}^{i}\}_{i = 1}^{K_{f}}$ , hands $V_{h} = \{v_{h}^{i}\}_{i = 1}^{K_{h}}$ , upper body $V_{u} = \{v_{u}^{i}\}_{i = 1}^{K_{u}}$ , and lower body $V_{l} = \{v_{l}^{i}\}_{i = 1}^{K_{l}}$ , following previous work [23]. In particular, the motion vocabulary is defined as a combination of four body-part vocabularies: $V_{m} = \{v_{f}^{i},v_{h}^{i},v_{u}^{i},v_{l}^{i}\}_{i = 1}^{K_{m}}$ . Additionally, each modality-specific vocabulary includes special tokens for boundary recognition, such as $$ and $$ to indicate the start and end of an audio sequence. As a result, all modalities can be represented in a unified format with one joint multimodal vocabulary $V = \{V_{t},V_{a},V_{f},V_{h},V_{u},V_{l}\}$ . + +# 3.3. Pre-training for Modality Alignment + +Existing motion generation models rely heavily on paired data to train downstream tasks. Yet, collecting high-quality paired motion data is both costly and time consuming while there exists a large amount of unpaired data of each modality that can be explored. Inspired by this, we introduce our generative pre-training strategy, as shown in Fig. 3. More specifically, we implement two types of modality alignment during the pre-training stage: compositional motion alignment and audio-text alignment that are detailed below. + +Compositional body motion alignment. Our body motion is inherently compositional, i.e., different body parts move in accordance. For example, when we are happy, our faces express smiles and our gestures tend to become more positive. The correlation between different body part motions is universal, transcending cultural boundaries. This shared prior forms the basis of our approach. To explore this correspondence, we consider two types of motion alignment tasks: spatial and temporal. + +Spatial. To model the correlation between these different body parts, we train the model to take in a randomly selected combination of body parts (e.g., upper or upper + face) and predict another randomly selected combination of other body parts (e.g., lower or lower + hand). This helps our model learn the spatial relations between body parts. + +Below is one example template that defines task prompts, conditions, and answers. The model takes both prompts and conditions as input and is expected to output the answer. + +Task Prompts: Translate upper to lower body. + +Conditions: Upper Body Tokens $V_{\text{condition}} = \{v_u^i \in V_u \mid i \in \{\text{sequence token index}\} \}$ + +Answer: Lower Body Tokens $V_{\text{Answer}} = \{v_l^i \in V_l \mid i \in \{\text{sequence token index}\} \}$ + +Temporal. Predicting how motion changes as a function of time is also an important self-supervision, which enables the model to capture the temporal evolution of motion. We model this by randomly masking off certain motion frames to help the model learn the temporal priors of motion. + +Task Prompts: Translate mask to unmasked motion. + +Conditions: Masked Tokens $V_{\text{condition}} = \{v_m^i \in V_m \mid i \in \{\text{masked sequence token index}\} \}$ . + +Answer: Unmasked Motion Tokens $V_{\text{Answer}} = \{v_m^i \in V \mid i \in \{\text{unmasked sequence token index}\} \}$ + +Audio-text alignment. In addition to the motion modality, we also design translation tasks between audio and text modalities, leveraging the abundance of available data. These tasks follow the format of "predicting modality Y from modality X". For example, "predicting text from audio" should help the model's performance in "predicting motion from audio" by mapping the audio embeddings into the well-pre-trained text embedding space. + +# 3.4. Post-training with Instruction Following + +After pre-training, the model gains an understanding of the underlying grammar and syntax within motion modality's vocabulary and good alignment between audio and text modalities. We then fine-tune the model with paired data on downstream tasks such as co-speech gesture generation or text-to-motion generation. To enable the model to perform desired downstream tasks while following natural human instructions, we construct a multi-task instruction-following template by formatting several key tasks such as audio-to-motion, text-to-motion, and emotion-to-motion into instructions. Specifically, for each task, we compose dozens of different instruction templates, resulting in more than one thousand different tasks, each having a unique instruction prompt. An example of our instruction template is shown below. See Supp. for more examples. + +Task Prompts: Based on $<$ AudioPLACEHOLDER $>$ , generate a full-body movement sequence involving face, hands, upper body, and lower body that matches the audio's rhythm. + +Conditions: Audio Tokens $V_{\text{condition}} = \{v_a^i \in V_a \mid i \in \{\text{audio sequence token index}\} \}$ + +Answer: Unmasked Motion Tokens $V_{\text{Answer}} = \{v_m^i \in V \mid i \in \{\text{motion sequence token index}\} \}$ + +# 3.5. Language Model Training Details + +Our model leverages 220M pre-trained Flan-T5-Base model [54] with an encoder-decoder transformer structure to address the conditional generation task. Through our shared multimodal vocabulary $V$ , every input modality is represented as "text" tokens, allowing us to fully leverage the original T5 model for each conditional generation task. Specifically, with constructed modality latent codebook indices—such as index 8 of the upper body codebook—the upper body can be formatted as "< upper_8 >". Thus, the input can be converted into a sequence of tokens $S_{i} = \{s_{i}^{k}\}_{k=1}^{L}$ , where $s_{i} \in V$ and $L$ represents the input length. Similarly, the model outputs a sequence of tokens $S_{o} = \{s_{o}^{k}\}_{k=1}^{L}$ , with a fixed input/output token length and $s_{o} \in V$ . + +Since our model is encoder-decoder architecture, we set a maximum input length of 512. We specify modalities with start and stop tokens. Following the original T5 implementation, the sequence of tokens is sent to the encoder, and the decoder then performs next-token predictions in an auto-regressive manner at each step. The training objective can be formulated as follows: + +$$ +\mathcal {L} _ {L M} = - \sum_ {k = 0} ^ {L _ {t} - 1} \log p _ {\theta} \left(s _ {t} ^ {k} \mid s _ {t} ^ {< k}, s _ {i}\right), \tag {3} +$$ + +where $s_t$ represents each token within the sequence, serving as the index in our unified vocabulary $V$ . Through next-token prediction, our model learns the underlying distribution of each modality, enabling the accurate and meaningful generation of target "words". For both pre-training and post-training, we finetune the model's entire weights instead of performing low-rank adaptation (LORA [21]) since our goal is to maximally align each modality. + +# 4. Experiments + +In this section, we first evaluate our model on the co-speech gesture generation benchmark, then investigate the generalization enabled by generative multimodal pre-training, demonstrate our model's capability in following both audio and text prompts, and lastly our novel ability to predict emotion from motion. + +# 4.1. Co-Speech Gesture Generation + +To evaluate our model's audio-to-motion generation ability, we choose to benchmark on co-speech gesture generation on the BEATv2 dataset [43], where the goal is to generate body gesture motion for a given speech of a speaker. Existing co-speech generation work typically models speaker-dependent gestures [43, 47, 70]. During the pre-training + +
FGD ↓BC↑Diversity↑Condition Signal
DisCo [40]9.4176.4399.912audio
CaMN [41]6.6446.76910.86audio, text, facial
DiffStyleGesture [68]8.8117.24111.49audio, style
Habibie et al. [19]9.0407.7168.213audio, text
TalkSHOW [70]6.2096.94713.47audio
SynTalker [8]6.4137.97112.721audio, text
EMAGE [43]5.5127.72413.06audio, text
Ours w/o language pre-training7.4706.14814.162audio
Ours w/o multimodal pre-training5.4087.74214.418audio
Ours5.3017.78015.167audio
+ +Table 1. Co-speech gesture generation results on BEATv2 benchmark. We report FGD $\times 10^{-1}$ , BC $\times 10^{-1}$ , Diversity, which measures realism, speech-motion synchronization, and diversity respectively. Our model outperforms state-of-the-art methods on this benchmark. + +
FGD ↓BC↑Diversity↑
W/o pre-training5.5017.72114.281
W/o A2T5.4437.72114.499
W/o spatial6.3367.38114.173
W/o temporal6.8007.34113.810
W/o motion7.7767.34414.640
Ours5.3017.78015.165
+ +Table 2. Ablations of pre-training. + +stage, our model utilizes large-scale unpaired data in a self-supervised setup, drawing from two primary datasets, BEATv2 and Librispeech [49]. Together, these datasets provide approximately 1,000 hours of audio-to-text data and 60 hours of motion data. During the post-training, to ensure a fair comparison with baselines, we adopt the same evaluation protocol as [43], i.e., training and testing on speaker-2 and using their motion tokenizer. For pre-training, we ensure the model does not see any audio-to-motion data. Following prior work [43], we adopt Frechet Gesture Distance (FGD) [71] to evaluate the realism of the body gestures, Beat Correlation (BC) [36] to assess speech-motion synchrony and Diversity [31], which is calculated with the $\ell_1$ distance between multiple body gesture clips. + +The results are shown in Table 1. Compared with the state-of-the-art methods on this benchmark, our model achieves better performance across all metrics, indicating that our model generates more realistic, and diverse motion that is synchronized with the speech. Existing work often supplies additional signals to the model to boost the model's performance such as the text transcribed from the speech [19, 41, 43] or onset/amplitudes [43], partially due to the lack of semantic understanding of speech. In our model, since we use a pre-trained language model, the model naturally has a strong semantic understanding, allowing our model to show competitive performance without heavy reliance on hand-crafting features. If we use randomly ini + +tialized model weights, we can see that the model's performance drops drastically, indicating that language pretraining is vital for co-speech gesture generation. If we remove our multimodal pre-training stage, the model's performance also deteriorates, showing that our model benefits from the generative pre-training. + +To further understand our model's performance, we show some qualitative results of our model in Fig. 4. We can see that our model generates gestures that are synchronized with the speech. See Supp. video for more examples. + +# 4.2. Effect of Generative Pre-training + +Generating gesture motion for a new speaker requires collecting high-quality motion data, typically from motion capture systems. Collecting such data is time-consuming. In this section, we first validate the importance of each pretraining task and then investigate whether our generative pre-training leads to better generalization on new speakers and thus reduces the amount of data required for training. + +Validating pre-training tasks. To understand how different pre-training objectives contribute to the performance, we ablate the audio-to-text alignment task ("w/o A2T"), the spatial body motion alignment task ("w/o spatial"), the temporal body motion alignment task ("w/o temporal") and the whole body alignment task ("w/o motion"). + +The results are shown in Table 1. "w/o A2T" lowers the model's performance, indicating that aligning the audio embedding space with text helps with semantic understanding and also the downstream gesture generation task. Removing either spatial motion prediction, temporal motion prediction or them altogether hurts the performance, showing that learning spatial-temporal motion priors in the pre-training stage is important for the downstream tasks. + +Effect on the training data. We hypothesize that our pre-training strategy captures strong multimodal correlation and motion priors, which could reduce the reliance on the amount of paired data for downstream tasks. To validate this hypothesis, we follow the setting in Sec. 4.1 and limit + +![](images/3ae7cad739c0c87b2e6019e90ddeee7bf5951e08bb3828426da5bea469cb3619.jpg) +Figure 4. Qualitative example on co-speech gesture generation. Given a speech, we visualize the ground truth 3D motion accompanying the audio, the motion generated by the baseline EMAGE [43], SynTalker [8] and our method. Our model generates more diverse and expressive motion compared to the baseline, especially when the speaker emphasizes on certain words such as "tired" and "because". + +![](images/409481e1a5cf52e072b5a289cbf0370dfd29c4057d1609813fba0f341de2db29.jpg) +Figure 5. Generation performance vs. the amount of post-training data. Our model learns a stronger motion prior from pre-training and thus shows much better under data scarcity. + +the amount of training data available to the model during the pre-training stage. Note that the model has never seen audio2motion data during pre-training. We set the amount of the data to $\frac{1}{2^n}$ , $n \in [1 \dots 5]$ . We train our full model, our model without pre-training and EMAGE to convergence under each setting and evaluate on the same test set. + +The results are shown in Figure 5. We can see that our + +full model starts with much lower FGD compared with the model without pre-training even when only using 1/32 of the paired training. As expected, as the amount of paired fine-tuning data increases, the performance reduces but our full model always outperforms the w/o pre-training ablation and EMAGE, showing that our model benefits pretraining and shows greater generalization under extreme data scarcity. + +# 4.3. Unifying Audio-to-Motion and Text-to-Motion for Editable Generation + +By taking a language model-driven approach, our model is capable of following both audio and text prompts. We first train the motion tokenizer on both BEATv2 and AMASS [45] datasets since the range of motion in these two datasets is very different. We use the same tasks for pre-training. For post-training, we combine Audio2Motion and Text2Motion with various instructions, in which text-to-motion with HumanML3D [16] text annotations. See Supp. for details. + +By training on both text-to-motion and audio-to-motion data, our model supports joint audio-text prompts, enabling what we call editable gesture generation. This approach + +![](images/8d6516e931f15e775591029844f872e3e2a387e25e957510fac33d1929c8e541.jpg) +Figure 6. Editable gesture generation. We prompt the language with text and audio information and it outputs motions that are both expressional gesture motion as well as general movement motion. + +![](images/17bd6f7f4152f5814b06b4976a577a72c3b435d66a9f14b1273789d7013af234.jpg) +Can you generate the lower body movement of someone sitting in a chair and generate the upper body movement according to the speech + +![](images/1473874f67c949e1d81bc002c43fc50180be8aa7aaa77b441fad865cee215086.jpg) + +![](images/d631a4a6ff6337bbdd1461b5ee428cbd0d97dd5701cd1e0ba1a2b3fddb7ab6cc.jpg) +Can you generate the lower body movement of a person walking forward and returning backward and generate the upper body movement according to the speech $\leftarrow +\uparrow$ + +What emotion is conveyed by the movements in the body motion? + +![](images/c66b72b5f6e047372197f600d470db08ca8c1c7e619013eccfb630b50f8de487.jpg) +MotionGPT +Ours +Figure 7. Qualitative example of emotion prediction. + +A person dances while moving their hips. + +The person expresses the emotion of "contempt" + +facilitates the generation of synergistic full-body motions conditioned on both speech and flexibly chosen prompts. For instance, the model can generate the motion of a person walking while talking. In this work, we demonstrate this capability by prompting the model separately for specific body part motions and then combining them seamlessly. Combining conversation gestures with daily motions is extremely useful for applications such as gaming or VR. We show several qualitative examples in Fig. 6. We can see that the model can generate human motion that follows both audio and text prompts, showing the emergent capabilities of our model. See Supp. video for more examples. + +# 4.4. Predicting Emotion from Motion + +Our model's flexibility in the input/output modality also unlocks an array of new tasks such as translation between different body parts or modalities. In this section, we propose a novel task that predicts emotion from motion. + +Reading someone's body language, i.e., predicting emotion from motion is important for applications such as mental health or psychiatry, however, existing audio2motion or motion2text do not have this capability. We extract the emotion labels (neutral, anger, happiness, fear, disgust, sadness, contempt, and surprise) on BEATv2 and convert them + +
Bleu@1↑Rouge Cider↑BertScore↑
GT10010099.9
Random2.454.440.19
MotionGPT1.6810.672.31
Ours14.7126.6716.94
+ +Table 3. Motion to emotion. We prompt our model to predict emotion given a motion sequence. + +into instructions for training. To be compatible with arbitrary language output from MotionGPT, we evaluate the model performance by measuring the BLEU [50], Rouge Cider [39], and BERTScore [81] between the prediction and the ground truth, which measures the semantic distances between texts. See more details in Supp. + +The results are shown in Table 3. MotionGPT entirely fails this task with a performance similar to a random baseline because it was only trained to caption general motion rather than subtle gesture movement and body language. Our model outperforms the random and MotionGPT by a large margin, showing our model's ability to predict the emotion from motion. We also show one qualitative example in Fig. 7. + +# 5. Discussion + +In this work, we propose a novel multimodal language model to unify verbal and non-verbal language with a novel pre-training objective. Our model not only shows state-of-the-art performance on co-speech gestures but also unlocks an array of novel tasks. + +While promising, the model sometimes fails to produce coherent motion potentially due to discrete motion tokenization. Moving forward, we believe incorporating continuous tokenization is an important step to improve the quality of the generated motion. + +We believe unifying verbal and non-verbal language of human motion generation and understanding is crucial for real-world applications, and language models provide a powerful framework to approach that goal. + +Acknowledgments: This project was partially funded by NIH grant R01AG089169 and UST. The authors would also like to thank Georgios Pavlakos for his valuable discussion, Chaitanya Patel, Jingyan Zhang, and Bin Li for their feedback on the paper. + +# References + +[1] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In ECCV, pages 248–265. Springer, 2020. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 2 +[3] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 2 +[4] Nikos Athanasiou, Alpár Ceske, Markos Diomataris, Michael J Black, and Gül Varol. Motionfix: Text-driven 3d human motion editing. arXiv preprint arXiv:2408.00712, 2024. 2 +[5] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In CVPR, pages 15935-15946, 2022. 2 +[6] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: a language modeling approach to audio generation, 2023. 2 +[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 3 +[8] Bohong Chen, Yumeng Li, Yao-Xiang Ding, Tianjia Shao, and Kun Zhou. Enabling synergistic full-body control in prompt-based co-speech motion generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6774-6783, 2024. 2, 6, 7 +[9] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos, 2024. 2 + +[10] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2 +[11] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2 +[12] Kiran Chhatre, Radek Daněček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J. Black, and Timo Bolkart. AMUSE: Emotional speech-driven 3D body animation via disentangled latent diffusion. In CVPR, pages 1942-1953, 2024. 2 +[13] Seunggeun Chi, Hyung-gun Chi, Hengbo Ma, Nakul Agarwal, Faizan Siddiqui, Karthik Ramani, and Kwonjoon Lee. M2d2m: Multi-motion generation from text with discrete diffusion models. arXiv preprint arXiv:2407.14502, 2024. 2 +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 3 +[15] Yuan Gong, Hongyin Luo, Alexander H Liu, Leonid Karlinsky, and James Glass. Listen, think, and understand. arXiv preprint arXiv:2305.10790, 2023. 2 +[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, pages 5152-5161, 2022. 2, 3, 7 +[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597. Springer, 2022. +[18] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, pages 1900-1910, 2024. 2 +[19] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video, 2021. 2, 6 +[20] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29: 3451-3460, 2021. 4 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5 +[22] Rongjie Huang, Mingze Li, Dongchao Yang, Jiatong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu, Zhiqing Hong, Jiawei Huang, Jinglin Liu, et al. Audiogpt: Understanding and generating speech, music, sound, and talking head. In AAAI, pages 23802-23804, 2024. 3 +[23] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2, 4 + +[24] Biao Jiang, Xin Chen, Chi Zhang, Fukun Yin, Zhuoyuan Li, Gang Yu, and Jiayuan Fan. Motionchain: Conversational motion controllers via multimodal prompts. In ECCV, pages 54-74. Springer, 2025. 2 +[25] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1737-1747, 2024. 2 +[26] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, pages 2151-2162, 2023. 2 +[27] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[28] T Kudo. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 2, 4 +[29] Taku Kudo. Subword regularization: Improving neural network translation models with multiple subword candidates. arXiv preprint arXiv:1804.10959, 2018. 4 +[30] Gen Li, Kaifeng Zhao, Siwei Zhang, Xiaozhong Lyu, Mihai Dusmanu, Yan Zhang, Marc Pollefeys, and Siyu Tang. EgoGen: An Egocentric Synthetic Data Generator. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[31] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 2, 6 +[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 3 +[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2 +[34] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2 +[35] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3 +[36] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++, 2021. 6 +[37] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3 + +[38] Han Liang, Jiacheng Bao, Ruichi Zhang, Sihan Ren, Yuecheng Xu, Sibei Yang, Xin Chen, Jingyi Yu, and Lan Xu. Omg: Towards open-vocabulary motion generation via mixture of controllers. In CVPR, pages 482-493, 2024. 2 +[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 8 +[40] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3764-3773, 2022. 6 +[41] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In ECCV, 2022. 6 +[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 3 +[43] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J. Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In CVPR, 2024. 2, 3, 5, 6, 7 +[44] Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-world text-to-motion generation. arXiv preprint arXiv:2312.14828, 2023. 2 +[45] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 7 +[46] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In CVPR, 2022. 2 +[47] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In ICCV, 2023. 2, 5 +[48] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations, 2024. 2 +[49] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 6 +[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 8 +[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3 + +[52] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2 +[53] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit Bermano, Eric Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. In Computer Graphics Forum, page e15063. Wiley Online Library, 2024. 3 +[54] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 3, 4, 5 +[55] Rico Sennrich. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 4 +[56] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2 +[57] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H. Bermano. Human motion diffusion model. In ICLR, 2022. 2 +[58] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2 +[59] Minh Tran, Di Chang, Maksim Siniukov, and Mohammad Soleymani. Dyadic interaction modeling for social behavior generation. In ECCV, 2024. 2 +[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 3 +[61] Yuan Wang, Di Huang, Yaqi Zhang, Wanli Ouyang, Jile Jiao, Xuetao Feng, Yan Zhou, Pengfei Wan, Shixiang Tang, and Dan Xu. Motionpt-2: A general-purpose motion-language model for motion generation and understanding. arXiv, 2024. 2 +[62] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 2 +[63] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say, interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[64] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chi-Keung Tang. Motionllm: Multimodal motion-language learning with large language models, 2024. 2 +[65] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024.3 +[66] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for + +human motion generation. arXiv preprint arXiv:2310.08580, 2023. 2 +[67] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In CVPR, pages 12780-12790, 2023. 2 +[68] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffuse-style gesture: Stylized audio-driven co-speech gesture generation with diffusion models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 5860–5868. International Joint Conferences on Artificial Intelligence Organization, 2023. 6 +[69] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2 +[70] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J. Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 2, 5, 6 +[71] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics, 39(6), 2020. 6 +[72] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, pages 16010-16021, 2023. 2 +[73] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities, 2023. 2, 3 +[74] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Neuraldome: A neural modeling pipeline on multi-view human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8834-8845, 2023. 2 +[75] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2 +[76] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023. 2 +[77] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, pages 516-526, 2024. 2 +[78] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 2 + +[79] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In ICCV, pages 364-373, 2023. 2 +[80] Mingyuan Zhang, Daisheng Jin, Chenyang Gu, Fangzhou Hong, Zhongang Cai, Jingfang Huang, Chongzhi Zhang, Xinying Guo, Lei Yang, Ying He, et al. Large motion model for unified multi-modal motion generation. arXiv, 2024. 2 +[81] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 8 +[82] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, 2024. 2 +[83] Zhikai Zhang, Yitang Li, Haofeng Huang, Mingxian Lin, and Li Yi. Freemotion: Mocap-free human motion synthesis with multimodal large language models. In ECCV, pages 403-421. Springer, 2025. 2 +[84] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2024. 3 +[85] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Responsive listening head generation: A benchmark dataset and baseline. In ECCV, 2022. 2 +[86] Wenyang Zhou, Zhiyang Dou, Zeyu Cao, Zhouyingcheng Liao, Jingbo Wang, Wenjia Wang, Yuan Liu, Taku Komura, Wenping Wang, and Lingjie Liu. Emdm: Efficient motion diffusion model for fast and high-quality motion generation. In ECCV, pages 18-38. Springer, 2025. 2 \ No newline at end of file diff --git a/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/images.zip b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..07cec653bd6a9ee336f33f1ff1de326e4b10a504 --- /dev/null +++ b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e584bc053a2e80e71d5978ab21c5f5a810e26e872cab1f663d00d81077396363 +size 495690 diff --git a/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/layout.json b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d8b37499e42dc6f120591c66292006fe07f778c --- /dev/null +++ b/2025/The Language of Motion_ Unifying Verbal and Non-verbal Language of 3D Human Motion/layout.json @@ -0,0 +1,9408 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 496, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 496, + 140 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 496, + 140 + ], + "type": "text", + "content": "The Language of Motion: Unifying Verbal and Non-verbal Language of 3D Human Motion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 161, + 503, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 161, + 503, + 204 + ], + "spans": [ + { + "bbox": [ + 85, + 161, + 503, + 204 + ], + "type": "text", + "content": "Changan Chen* Juze Zhang* Shrinidhi K. Lakshmikanth* Yusu Fang Ruizhi Shao Gordon Wetzstein Li Fei-Fei Ehsan Adeli Stanford University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 214, + 536, + 328 + ], + "blocks": [ + { + "bbox": [ + 60, + 214, + 536, + 328 + ], + "lines": [ + { + "bbox": [ + 60, + 214, + 536, + 328 + ], + "spans": [ + { + "bbox": [ + 60, + 214, + 536, + 328 + ], + "type": "image", + "image_path": "eaf9dc24282cd37a53e4b1fb4c2c063f920cd82ca925261a42c5e27b1630f62c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 329, + 539, + 424 + ], + "blocks": [ + { + "bbox": [ + 48, + 329, + 539, + 424 + ], + "lines": [ + { + "bbox": [ + 48, + 329, + 539, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 329, + 539, + 424 + ], + "type": "image", + "image_path": "1c57ac4e75b795d65ef0e599448f3f4f9ea3874ddf4b74cbe91bc32fac8e57b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 431, + 546, + 464 + ], + "lines": [ + { + "bbox": [ + 46, + 431, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 546, + 464 + ], + "type": "text", + "content": "Figure 1. We introduce a language-model-based motion understanding and generation framework that takes in any of the audio/motion/text modalities and outputs the desired target modality. Coupled with our generative pre-training strategy, our model demonstrates competitive performance on an array of tasks, showing promising signs toward unified verbal and non-verbal language of human motions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 480, + 192, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 480, + 192, + 493 + ], + "spans": [ + { + "bbox": [ + 143, + 480, + 192, + 493 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 506, + 289, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 697 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 697 + ], + "type": "text", + "content": "Human communication is inherently multimodal, involving a combination of verbal and non-verbal cues such as speech, facial expressions, and body gestures. Modeling these behaviors is essential for understanding human interaction and for creating virtual characters that can communicate naturally in applications like games, films, and virtual reality. However, existing motion generation models are typically limited to specific input modalities—either speech, text, or motion data—and cannot fully leverage the diversity of available data. In this paper, we propose a novel framework that unifies verbal and non-verbal language using multimodal language models for human motion understanding and generation. This model is flexible in taking text, speech, and motion or any combination of them as input. Coupled with our novel pre-training strategy, our model not only achieves state-of-the-art perfor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 482, + 546, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 482, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 482, + 546, + 578 + ], + "type": "text", + "content": "mance on co-speech gesture generation but also requires much less data for training. Our model also unlocks an array of novel tasks such as editable gesture generation and emotion prediction from motion. We believe unifying the verbal and non-verbal language of human motion is essential for real-world applications, and language models offer a powerful approach to achieving this goal. Project page: languageofmotion.github.io." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 598, + 386, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 386, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 386, + 610 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "content": "Human communication is multimodal. We use spoken and body language, including hand gestures, facial expressions, body postures and emotional expressions to interact with each other effectively. For example, people use linguistic cues along with body language, including hand gestures, facial expressions, overall body posture, and even emotional expressions to interact with the environment effectively. Modeling these multimodal behaviors is essen" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 153, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 153, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 153, + 714 + ], + "type": "text", + "content": "*indicates equal contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6200" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "tial for understanding and generating human motion, enabling a wide range of applications for virtual characters in games, movies, and virtual reality—areas that have recently received substantial attention." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 277 + ], + "type": "text", + "content": "Existing work has been focused on modeling human motion from different modalities, such as speech [43, 48, 70], text [23, 64, 75], egocentric vision [30, 34], or the surrounding environment [5, 25, 63, 74, 77]. These models only take specific modalities as input, whose performance is thus limited to the data available for their downstream task. For example, co-speech gesture generation work typically trains speaker-dependent models [3, 19, 43, 70], which requires high-quality speech-motion capture of a person. While gesture style varies from person to person, many gestures are shared across people as well as non-speech-driven motion, such as walking or waving hands. Existing work has yet to leverage motion priors from all forms of motion data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 278, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 397 + ], + "type": "text", + "content": "One of the promising ways to unify different tasks is multimodal language models, where a single language model can take different modalities as input and output target modalities. These models have shown promising results in a wide range of multimodal tasks, such as visual question answering [2, 33, 42], audio understanding and generation [6, 15, 73], and text-to-motion generation [9, 10, 61, 75, 82]. While language models have been widely applied, it has not been explored in the speech-text-motion generation setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 495 + ], + "type": "text", + "content": "We argue language models play a crucial role in unifying the verbal and non-verbal language of human motion for three reasons: 1) language models naturally connect different modalities, 2) speech is highly semantic, and tasks like modeling laughter in response to a joke require strong semantic reasoning capabilities and 3) language models are equipped with strong semantic understanding from extensive pre-training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 687 + ], + "type": "text", + "content": "Towards this goal, we propose a novel multimodal language model for expressional motion generation and understanding (see Fig. 1). To leverage language models to model motion, we first tokenize motion separately for different body parts (face, hand, upper-body, lower-body). Such compositionality has shown to be more beneficial to model the expressive human expressions [43, 48]. Along with off-the-shelf tokenizers for text and speech [28], we can represent any given modality inputs as a sequence of tokens, which are consumed by language models. To train the language models, we design a two-stage training pipeline. The model is first pre-trained to align various modalities with compositional body motion alignment and audio-text alignment. After pre-training, we compile downstream tasks into instructions and train the model on these instructions to allow the model to follow various task instructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We first validate our model on the BEATv2 co-speech gesture generation benchmark [43] and show that our model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "strongly outperforms state-of-the-art models. We then conduct a thorough evaluation to demonstrate the effectiveness of our pre-training tasks. We also show that our pre-training strategy is more powerful when under severe data scarcity. While never seeing speech-motion data during pre-training, our model reaches competitive performance with a relatively small amount of data for a novel speaker, showing remarkable generalization. By performing post-training on both speech-motion and text-motion tasks, we show that our model not only follows audio and text prompts but also unlocks novel tasks such as predicting emotion from motion data. Please watch the Supp. video for the qualitative examples. To the best of our knowledge, this is the first work to build multimodal language models to unify the verbal and non-verbal language of 3D human motions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 264, + 392, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 392, + 276 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 392, + 276 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 284, + 489, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 489, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 489, + 296 + ], + "type": "text", + "content": "2.1. Speech-Driven Motion Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 303, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 518 + ], + "type": "text", + "content": "Human communication is multimodal and we use our speech, facial expressions, and body gestures to communicate with each other. Given this complimentary nature, recent work [1, 31] explores cross-modal generation of human motion from speech in different forms. These models are trained and evaluated on specific upper body joints, full body joints, and even facial expressions. Recent work in co-speech gesture generation often utilizes generative models to create gestures from audio conditions [8, 12, 43, 69]. Other work also explores the possibility of generating the listener's motion [46, 47, 59]. Another area of research focuses on generating speech-driven co-speech facial expressions, with notable works including ViCO [85] and CodeTalker [67]. However, these works are limited in that they only take speech as input and do not utilize other forms of motion data, making it challenging to follow both speech and textual cues. To tackle this, we propose to unify input/output modalities with a language model framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 527, + 453, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 453, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 453, + 539 + ], + "type": "text", + "content": "2.2. Text to Motion Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Humans communicate with spoken language and nonverbal means such as emotions and interactions with surrounding environment [25, 63], among other cues. Recent work has explored generating human motion from text descriptions [4, 13, 16-18, 24, 26, 44, 52, 56, 62, 66, 76, 76, 83]. Some work attempts to generate human motion using diffusion models [11, 57, 58, 72, 78, 78, 79, 86] while other work exploring language models for generating human motion [9, 23, 38, 61, 64, 64, 75, 80, 82]. While these works have shown promising results in generating human motion from text instructions, they fall short in capturing the underlying meaning of the motion language itself. This limitation makes it challenging to develop a model capable of generating human motion from both verbal and non-verbal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6201" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 70, + 517, + 259 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 517, + 259 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 517, + 259 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 517, + 259 + ], + "type": "image", + "image_path": "b8b427208a05acceec378bf9f1d63a6374446dca4ec6d0c438c914c991c791dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 264, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 547, + 319 + ], + "type": "text", + "content": "Figure 2. Method overview. We employ modality-specific tokenizers to process various input modalities. Specifically, we train a compositional body motion VQ-VAE to tokenize face, hands, upper body, and lower body motions into discrete tokens, combining these modality-specific vocabularies (audio and text) into a unified multimodal vocabulary. During training, mixed tokens from different modalities are used as input, and the output is generated through an encoder-decoder language model. The mixed tokens are fed into the transformer encoder, while the decoder predicts the probability distribution of the next token in an autoregressive manner at each step." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 330, + 288, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 288, + 365 + ], + "type": "text", + "content": "language. In this work, we propose a novel framework to capture patterns in body language and subtle expressive gestures inherently present in human communication." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 373, + 211, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 211, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 211, + 386 + ], + "type": "text", + "content": "2.3. Multimodal Language Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 392, + 288, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 288, + 644 + ], + "type": "text", + "content": "Recent years have witnessed the rise of language models [7, 14, 53, 54, 65, 84], primarily leveraging transformer architectures [60] that process text tokens as input and generate text tokens. Building upon these advancements, substantial efforts have expanded into multimodal language models capable of handling various types of input and output, with notable examples including BLIP-2 [32], LLaVA [42], and VideoChat [35]. Furthermore, the scope of multimodal language models (MM-LLMs) has broadened to include modality-specific outputs, as demonstrated by models like GILL [27] and SpeechGPT [73]. Efforts such as LLaVA [42] and AudioGPT [22] are advancing towards seamless any-to-any modality conversion, with the goal of emulating human-like cognitive abilities in multimodal contexts. Inspired by this line of work, we propose a new framework aimed at unifying verbal and non-verbal language within language models. Our framework takes text, speech, and motion data as input and generates human motion or text as output, further exploring the potential synergy between different tasks and modalities to enhance the performance of human motion generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 655, + 287, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 655, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 655, + 287, + 683 + ], + "type": "text", + "content": "3. Multimodal Language Model for Motion Generation and Understanding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "content": "In this section, we present a multimodal language model for motion generation and understanding, which is illustrated" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 329, + 547, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 378 + ], + "type": "text", + "content": "in Fig. 2. We first describe the tokenization of different modalities (Sec. 3.2), then we introduce our generative pretraining for modality alignment (Sec. 3.3), and finally, we detail post-training for instructions following (Sec. 3.4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 388, + 391, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 391, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 391, + 399 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "text", + "content": "We use the neutral SMPL-X [51] body model including FLAME [37] face model. This model is parameterized by per-person body shape " + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{T\\times 300}" + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "text", + "content": ", 55 joint pose " + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{g}\\in \\mathbb{R}^{T\\times 55\\times 3}" + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "text", + "content": ", facial expression " + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "inline_equation", + "content": "\\psi \\in \\mathbb{R}^{T\\times 100}" + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "text", + "content": ", and global body translation " + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{T\\times 3}" + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 407, + 546, + 468 + ], + "type": "text", + "content": " is the frame number." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 478, + 389, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 389, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 389, + 489 + ], + "type": "text", + "content": "3.2. Tokenization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": "In order for our modelt take various modalities as input (audio, tex,t and motion), we first tokenize different modalities with modality-specific tokenizers, and then combine them into a multimodal vocabulary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "Compositional body motion tokenization. Following the motion representation approach in EMAGE [43], we divide the body into four parts with 6D rotation representation: 9 joints forming the lower-body " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_l\\in \\mathbb{R}^{T\\times 54}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", 13 joints forming the upper-body " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_u\\in \\mathbb{R}^{T\\times 78}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", 30 joints forming the hands " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_h\\in \\mathbb{R}^{T\\times 180}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " and 1-joint along with 100 expression parameters representing the face " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_f\\in \\mathbb{R}^{T\\times 106}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ". Collectively, the motion space is represented as " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "G = \\{\\mathbf{g}_f,\\mathbf{g}_h,\\mathbf{g}_u,\\mathbf{g}_l\\}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ". Notably, we avoid using the commonly adopted HumanML3D representation [16] (H3D-Format) in text-to-motion tasks, as it predominantly focuses on skeletal movement, emphasizing swinging motions while overlooking twisting rotations of body parts—an essential aspect for effectively conveying body language. With this" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6202" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "compositional representation, we train four separate VQVAEs tockenize the body pose for each part. Each VQVAE encoder " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " applies a four-layer temporal convolutional network (TCN) to extract continuous latent motion features " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{1:T} = \\mathcal{E}(\\mathbf{g}^{1:T})" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ". This encoded representation " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{1:T}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is quantized using:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 150, + 287, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 150, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 91, + 150, + 287, + 171 + ], + "type": "interline_equation", + "content": "\\mathbf {q} ^ {t} = \\mathcal {Q} \\left(\\mathbf {z} ^ {t}\\right) := \\arg \\min _ {\\mathbf {q} ^ {k} \\in Q} \\| \\mathbf {z} ^ {t} - \\mathbf {q} ^ {k} \\| ^ {2} \\tag {1}", + "image_path": "7107970f03631f42cd2931509cfdbbc53a56ff4ec7e0c35beb8fee11486bfc4e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^t" + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "text", + "content": " is the discrete code in the codebook representing the encoded " + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^t" + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "text", + "content": ". Collectively, the quantized motion latent space is " + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "inline_equation", + "content": "Q = \\{\\mathbf{q}_f,\\mathbf{q}_h,\\mathbf{q}_u,\\mathbf{q}_l\\}" + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "text", + "content": ". Each VQ-VAE decoder " + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "text", + "content": " decodes the quantized motion " + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^t" + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "text", + "content": " back into the motion space " + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{g}}^{1:T} = \\mathcal{D}(\\mathbf{q}^{1:T})" + }, + { + "bbox": [ + 46, + 177, + 287, + 249 + ], + "type": "text", + "content": " and applies the following reconstruction losses:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 253, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 253, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 84, + 253, + 287, + 312 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c}} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\mathcal {L} _ {\\text {v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\\\ \\mathcal {L} _ {\\mathrm {a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) + \\mathcal {L} _ {\\mathrm {m r e c}} (\\mathbf {g}, \\hat {\\mathbf {g}}) + \\tag {2} \\\\ \\mathcal {L} _ {\\mathrm {m v e l}} \\left(\\mathbf {g} ^ {\\prime}, \\hat {\\mathbf {g}} ^ {\\prime}\\right) + \\mathcal {L} _ {\\mathrm {m a c c}} \\left(\\mathbf {g} ^ {\\prime \\prime}, \\hat {\\mathbf {g}} ^ {\\prime \\prime}\\right) + \\\\ \\mathcal {L} _ {\\mathrm {c o m m}} (\\mathbf {g}, \\mathbf {q}), \\\\ \\end{array}", + "image_path": "3bd2e3ffe1908b6f069b015b4098e12e1a8c1bdeae76b31fd175587b36a84ad7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{g}}^{\\prime}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " represent the reconstructed motion, " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{g}}^{\\prime}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\prime}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " represent the velocity of " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{g}}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{g}}^{\\prime\\prime}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{g}^{\\prime\\prime}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " represent their acceleration. For lower-body, upper-body and hands VQ-VAEs, pose reconstruction loss " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " is a Geodesic loss. For face VQ-VAE, " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " loss. Pose velocity/acceleration losses " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vel}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{acc}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\ell_{1}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " losses. Mesh reconstruction loss " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mrec}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " loss. Mesh velocity/acceleration losses " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mvel}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{macc}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\ell_{1}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " losses. Codebook commitment loss " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{comm}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " loss. Vertices of the SMPLX-2020 mesh computed from the pose " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{g}}" + }, + { + "bbox": [ + 46, + 318, + 287, + 450 + ], + "type": "text", + "content": " are used to compute mesh losses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "content": "Speech tokenization. Similar to the motion modality, speech data is also continuous by nature. To facilitate speech training within a language model, we used HuBERT [20] to represent audio streams as discrete tokens. In this work, audio input is sampled at " + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "inline_equation", + "content": "16\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{T \\times s}" + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "content": " represents the audio frame rate after quantization. HuBERT further downsamples audio by a factor of 320, resulting in " + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "inline_equation", + "content": "s = 50" + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "content": ". This frame rate, compared to the typical motion frame rate of 30 fps, provides an acceptable input token length for language models. The resulting audio token space is noted as " + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "inline_equation", + "content": "A = \\{\\mathbf{a}\\}" + }, + { + "bbox": [ + 46, + 450, + 287, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": "Text tokenization. Following previous work [23, 54], we use SentencePiece [28] to tokenize text inputs and outputs into WordPiece tokens [29, 55] for the language model, with a vocabulary of 32,000 wordpieces inherited from the T5 [54] language model, which can be represented as " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "W = \\{\\mathbf{w}\\}" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": ". This vocabulary enables the model to process a fixed set of predetermined languages. Additionally, we extend the vocabulary with several multimodal tokens to support multi-modal inputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Multimodal vocabulary. Altogether, we have a combined token space defined as " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "M := Q \\cup A \\cup W \\cup C =" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 77, + 542, + 133 + ], + "blocks": [ + { + "bbox": [ + 310, + 77, + 542, + 133 + ], + "lines": [ + { + "bbox": [ + 310, + 77, + 542, + 133 + ], + "spans": [ + { + "bbox": [ + 310, + 77, + 542, + 133 + ], + "type": "image", + "image_path": "734e320712aa901c2edc4795e23f077389f2bbaf23dc81d851bb4d8260ea2ca5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 141, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 305, + 141, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 141, + 545, + 163 + ], + "type": "text", + "content": "Figure 3. Illustration of pre-training. We pre-train our language model by translating one modality to another using paired data." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q}_f,\\mathbf{q}_u,\\mathbf{q}_h,\\mathbf{q}_l,\\mathbf{a},\\mathbf{w}\\}" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": ". Each modality-specific tokenizer outputs its modality-specific vocabulary. To build a unified language model that can process these different modalities, we need to combine these vocabularies into a joint vocabulary. Since the language model is pre-trained with the text modality, we choose to extend the original text vocabulary " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "V_{t} = \\{v_{t}^{i}\\}_{i = 1}^{K_{t}}" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": " with vocabularies from other modalities, including audio " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "V_{a} = \\{v_{a}^{i}\\}_{i = 1}^{K_{a}}" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": ", face " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "V_{f} = \\{v_{f}^{i}\\}_{i = 1}^{K_{f}}" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": ", hands " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "V_{h} = \\{v_{h}^{i}\\}_{i = 1}^{K_{h}}" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": ", upper body " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "V_{u} = \\{v_{u}^{i}\\}_{i = 1}^{K_{u}}" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": ", and lower body " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "V_{l} = \\{v_{l}^{i}\\}_{i = 1}^{K_{l}}" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": ", following previous work [23]. In particular, the motion vocabulary is defined as a combination of four body-part vocabularies: " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "V_{m} = \\{v_{f}^{i},v_{h}^{i},v_{u}^{i},v_{l}^{i}\\}_{i = 1}^{K_{m}}" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": ". Additionally, each modality-specific vocabulary includes special tokens for boundary recognition, such as " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": " to indicate the start and end of an audio sequence. As a result, all modalities can be represented in a unified format with one joint multimodal vocabulary " + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "inline_equation", + "content": "V = \\{V_{t},V_{a},V_{f},V_{h},V_{u},V_{l}\\}" + }, + { + "bbox": [ + 305, + 177, + 545, + 400 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 407, + 500, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 500, + 420 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 500, + 420 + ], + "type": "text", + "content": "3.3. Pre-training for Modality Alignment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": "Existing motion generation models rely heavily on paired data to train downstream tasks. Yet, collecting high-quality paired motion data is both costly and time consuming while there exists a large amount of unpaired data of each modality that can be explored. Inspired by this, we introduce our generative pre-training strategy, as shown in Fig. 3. More specifically, we implement two types of modality alignment during the pre-training stage: compositional motion alignment and audio-text alignment that are detailed below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 641 + ], + "type": "text", + "content": "Compositional body motion alignment. Our body motion is inherently compositional, i.e., different body parts move in accordance. For example, when we are happy, our faces express smiles and our gestures tend to become more positive. The correlation between different body part motions is universal, transcending cultural boundaries. This shared prior forms the basis of our approach. To explore this correspondence, we consider two types of motion alignment tasks: spatial and temporal." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Spatial. To model the correlation between these different body parts, we train the model to take in a randomly selected combination of body parts (e.g., upper or upper + face) and predict another randomly selected combination of other body parts (e.g., lower or lower + hand). This helps our model learn the spatial relations between body parts." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6203" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "Below is one example template that defines task prompts, conditions, and answers. The model takes both prompts and conditions as input and is expected to output the answer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 115, + 242, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 115, + 242, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 115, + 242, + 128 + ], + "type": "text", + "content": "Task Prompts: Translate upper to lower body." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 128, + 277, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 128, + 277, + 152 + ], + "spans": [ + { + "bbox": [ + 57, + 128, + 277, + 152 + ], + "type": "text", + "content": "Conditions: Upper Body Tokens " + }, + { + "bbox": [ + 57, + 128, + 277, + 152 + ], + "type": "inline_equation", + "content": "V_{\\text{condition}} = \\{v_u^i \\in V_u \\mid i \\in \\{\\text{sequence token index}\\} \\}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 152, + 276, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 276, + 177 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 276, + 177 + ], + "type": "text", + "content": "Answer: Lower Body Tokens " + }, + { + "bbox": [ + 57, + 152, + 276, + 177 + ], + "type": "inline_equation", + "content": "V_{\\text{Answer}} = \\{v_l^i \\in V_l \\mid i \\in \\{\\text{sequence token index}\\} \\}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 185, + 287, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 185, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 185, + 287, + 244 + ], + "type": "text", + "content": "Temporal. Predicting how motion changes as a function of time is also an important self-supervision, which enables the model to capture the temporal evolution of motion. We model this by randomly masking off certain motion frames to help the model learn the temporal priors of motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 252, + 267, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 252, + 267, + 264 + ], + "spans": [ + { + "bbox": [ + 57, + 252, + 267, + 264 + ], + "type": "text", + "content": "Task Prompts: Translate mask to unmasked motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 264, + 276, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 264, + 276, + 289 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 276, + 289 + ], + "type": "text", + "content": "Conditions: Masked Tokens " + }, + { + "bbox": [ + 57, + 264, + 276, + 289 + ], + "type": "inline_equation", + "content": "V_{\\text{condition}} = \\{v_m^i \\in V_m \\mid i \\in \\{\\text{masked sequence token index}\\} \\}" + }, + { + "bbox": [ + 57, + 264, + 276, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 289, + 276, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 289, + 276, + 313 + ], + "spans": [ + { + "bbox": [ + 57, + 289, + 276, + 313 + ], + "type": "text", + "content": "Answer: Unmasked Motion Tokens " + }, + { + "bbox": [ + 57, + 289, + 276, + 313 + ], + "type": "inline_equation", + "content": "V_{\\text{Answer}} = \\{v_m^i \\in V \\mid i \\in \\{\\text{unmasked sequence token index}\\} \\}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 321, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 321, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 287, + 418 + ], + "type": "text", + "content": "Audio-text alignment. In addition to the motion modality, we also design translation tasks between audio and text modalities, leveraging the abundance of available data. These tasks follow the format of \"predicting modality Y from modality X\". For example, \"predicting text from audio\" should help the model's performance in \"predicting motion from audio\" by mapping the audio embeddings into the well-pre-trained text embedding space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 424, + 258, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 258, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 258, + 436 + ], + "type": "text", + "content": "3.4. Post-training with Instruction Following" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 441, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 622 + ], + "type": "text", + "content": "After pre-training, the model gains an understanding of the underlying grammar and syntax within motion modality's vocabulary and good alignment between audio and text modalities. We then fine-tune the model with paired data on downstream tasks such as co-speech gesture generation or text-to-motion generation. To enable the model to perform desired downstream tasks while following natural human instructions, we construct a multi-task instruction-following template by formatting several key tasks such as audio-to-motion, text-to-motion, and emotion-to-motion into instructions. Specifically, for each task, we compose dozens of different instruction templates, resulting in more than one thousand different tasks, each having a unique instruction prompt. An example of our instruction template is shown below. See Supp. for more examples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 628, + 277, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 277, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 277, + 676 + ], + "type": "text", + "content": "Task Prompts: Based on " + }, + { + "bbox": [ + 56, + 628, + 277, + 676 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 56, + 628, + 277, + 676 + ], + "type": "text", + "content": " AudioPLACEHOLDER " + }, + { + "bbox": [ + 56, + 628, + 277, + 676 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 56, + 628, + 277, + 676 + ], + "type": "text", + "content": ", generate a full-body movement sequence involving face, hands, upper body, and lower body that matches the audio's rhythm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 677, + 276, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 677, + 276, + 702 + ], + "spans": [ + { + "bbox": [ + 57, + 677, + 276, + 702 + ], + "type": "text", + "content": "Conditions: Audio Tokens " + }, + { + "bbox": [ + 57, + 677, + 276, + 702 + ], + "type": "inline_equation", + "content": "V_{\\text{condition}} = \\{v_a^i \\in V_a \\mid i \\in \\{\\text{audio sequence token index}\\} \\}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 72, + 536, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 536, + 97 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 536, + 97 + ], + "type": "text", + "content": "Answer: Unmasked Motion Tokens " + }, + { + "bbox": [ + 315, + 72, + 536, + 97 + ], + "type": "inline_equation", + "content": "V_{\\text{Answer}} = \\{v_m^i \\in V \\mid i \\in \\{\\text{motion sequence token index}\\} \\}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 114, + 486, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 114, + 486, + 127 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 486, + 127 + ], + "type": "text", + "content": "3.5. Language Model Training Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "text", + "content": "Our model leverages 220M pre-trained Flan-T5-Base model [54] with an encoder-decoder transformer structure to address the conditional generation task. Through our shared multimodal vocabulary " + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "text", + "content": ", every input modality is represented as \"text\" tokens, allowing us to fully leverage the original T5 model for each conditional generation task. Specifically, with constructed modality latent codebook indices—such as index 8 of the upper body codebook—the upper body can be formatted as \"< upper_8 >\". Thus, the input can be converted into a sequence of tokens " + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "inline_equation", + "content": "S_{i} = \\{s_{i}^{k}\\}_{k=1}^{L}" + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "inline_equation", + "content": "s_{i} \\in V" + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "text", + "content": " represents the input length. Similarly, the model outputs a sequence of tokens " + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "inline_equation", + "content": "S_{o} = \\{s_{o}^{k}\\}_{k=1}^{L}" + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "text", + "content": ", with a fixed input/output token length and " + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "inline_equation", + "content": "s_{o} \\in V" + }, + { + "bbox": [ + 305, + 132, + 545, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 289, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 370 + ], + "type": "text", + "content": "Since our model is encoder-decoder architecture, we set a maximum input length of 512. We specify modalities with start and stop tokens. Following the original T5 implementation, the sequence of tokens is sent to the encoder, and the decoder then performs next-token predictions in an auto-regressive manner at each step. The training objective can be formulated as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 354, + 377, + 545, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 377, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 354, + 377, + 545, + 410 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L M} = - \\sum_ {k = 0} ^ {L _ {t} - 1} \\log p _ {\\theta} \\left(s _ {t} ^ {k} \\mid s _ {t} ^ {< k}, s _ {i}\\right), \\tag {3}", + "image_path": "06e55a0467101b989af770795128a98d8841ae63b6c049a8e943105bd68552bf.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 418, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 305, + 418, + 545, + 514 + ], + "type": "text", + "content": " represents each token within the sequence, serving as the index in our unified vocabulary " + }, + { + "bbox": [ + 305, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 418, + 545, + 514 + ], + "type": "text", + "content": ". Through next-token prediction, our model learns the underlying distribution of each modality, enabling the accurate and meaningful generation of target \"words\". For both pre-training and post-training, we finetune the model's entire weights instead of performing low-rank adaptation (LORA [21]) since our goal is to maximally align each modality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 524, + 387, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 524, + 387, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 524, + 387, + 537 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 544, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 615 + ], + "type": "text", + "content": "In this section, we first evaluate our model on the co-speech gesture generation benchmark, then investigate the generalization enabled by generative multimodal pre-training, demonstrate our model's capability in following both audio and text prompts, and lastly our novel ability to predict emotion from motion." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 623, + 473, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 473, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 473, + 635 + ], + "type": "text", + "content": "4.1. Co-Speech Gesture Generation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": "To evaluate our model's audio-to-motion generation ability, we choose to benchmark on co-speech gesture generation on the BEATv2 dataset [43], where the goal is to generate body gesture motion for a given speech of a speaker. Existing co-speech generation work typically models speaker-dependent gestures [43, 47, 70]. During the pre-training" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6204" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 70, + 476, + 220 + ], + "blocks": [ + { + "bbox": [ + 118, + 70, + 476, + 220 + ], + "lines": [ + { + "bbox": [ + 118, + 70, + 476, + 220 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 476, + 220 + ], + "type": "table", + "html": "
FGD ↓BC↑Diversity↑Condition Signal
DisCo [40]9.4176.4399.912audio
CaMN [41]6.6446.76910.86audio, text, facial
DiffStyleGesture [68]8.8117.24111.49audio, style
Habibie et al. [19]9.0407.7168.213audio, text
TalkSHOW [70]6.2096.94713.47audio
SynTalker [8]6.4137.97112.721audio, text
EMAGE [43]5.5127.72413.06audio, text
Ours w/o language pre-training7.4706.14814.162audio
Ours w/o multimodal pre-training5.4087.74214.418audio
Ours5.3017.78015.167audio
", + "image_path": "71aaa66e2095b6032610e564134064e9d229d006d9df70a02f1048ff0fadfeb3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 255, + 272, + 352 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "type": "text", + "content": "Table 1. Co-speech gesture generation results on BEATv2 benchmark. We report FGD " + }, + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\times 10^{-1}" + }, + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "type": "text", + "content": ", BC " + }, + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\times 10^{-1}" + }, + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "type": "text", + "content": ", Diversity, which measures realism, speech-motion synchronization, and diversity respectively. Our model outperforms state-of-the-art methods on this benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 255, + 272, + 352 + ], + "lines": [ + { + "bbox": [ + 63, + 255, + 272, + 352 + ], + "spans": [ + { + "bbox": [ + 63, + 255, + 272, + 352 + ], + "type": "table", + "html": "
FGD ↓BC↑Diversity↑
W/o pre-training5.5017.72114.281
W/o A2T5.4437.72114.499
W/o spatial6.3367.38114.173
W/o temporal6.8007.34113.810
W/o motion7.7767.34414.640
Ours5.3017.78015.165
", + "image_path": "568c3ffd2f1850f9d27e80617b905fce4075e211de655dc8f550e1c8bfdd65a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 360, + 230, + 372 + ], + "lines": [ + { + "bbox": [ + 104, + 360, + 230, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 360, + 230, + 372 + ], + "type": "text", + "content": "Table 2. Ablations of pre-training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 387, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 554 + ], + "type": "text", + "content": "stage, our model utilizes large-scale unpaired data in a self-supervised setup, drawing from two primary datasets, BEATv2 and Librispeech [49]. Together, these datasets provide approximately 1,000 hours of audio-to-text data and 60 hours of motion data. During the post-training, to ensure a fair comparison with baselines, we adopt the same evaluation protocol as [43], i.e., training and testing on speaker-2 and using their motion tokenizer. For pre-training, we ensure the model does not see any audio-to-motion data. Following prior work [43], we adopt Frechet Gesture Distance (FGD) [71] to evaluate the realism of the body gestures, Beat Correlation (BC) [36] to assess speech-motion synchrony and Diversity [31], which is calculated with the " + }, + { + "bbox": [ + 46, + 387, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 387, + 287, + 554 + ], + "type": "text", + "content": " distance between multiple body gesture clips." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "The results are shown in Table 1. Compared with the state-of-the-art methods on this benchmark, our model achieves better performance across all metrics, indicating that our model generates more realistic, and diverse motion that is synchronized with the speech. Existing work often supplies additional signals to the model to boost the model's performance such as the text transcribed from the speech [19, 41, 43] or onset/amplitudes [43], partially due to the lack of semantic understanding of speech. In our model, since we use a pre-trained language model, the model naturally has a strong semantic understanding, allowing our model to show competitive performance without heavy reliance on hand-crafting features. If we use randomly ini" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 257, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 328 + ], + "type": "text", + "content": "tialized model weights, we can see that the model's performance drops drastically, indicating that language pretraining is vital for co-speech gesture generation. If we remove our multimodal pre-training stage, the model's performance also deteriorates, showing that our model benefits from the generative pre-training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 377 + ], + "type": "text", + "content": "To further understand our model's performance, we show some qualitative results of our model in Fig. 4. We can see that our model generates gestures that are synchronized with the speech. See Supp. video for more examples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 384, + 484, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 484, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 484, + 397 + ], + "type": "text", + "content": "4.2. Effect of Generative Pre-training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 403, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 486 + ], + "type": "text", + "content": "Generating gesture motion for a new speaker requires collecting high-quality motion data, typically from motion capture systems. Collecting such data is time-consuming. In this section, we first validate the importance of each pretraining task and then investigate whether our generative pre-training leads to better generalization on new speakers and thus reduces the amount of data required for training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "text", + "content": "Validating pre-training tasks. To understand how different pre-training objectives contribute to the performance, we ablate the audio-to-text alignment task (\"w/o A2T\"), the spatial body motion alignment task (\"w/o spatial\"), the temporal body motion alignment task (\"w/o temporal\") and the whole body alignment task (\"w/o motion\")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 653 + ], + "type": "text", + "content": "The results are shown in Table 1. \"w/o A2T\" lowers the model's performance, indicating that aligning the audio embedding space with text helps with semantic understanding and also the downstream gesture generation task. Removing either spatial motion prediction, temporal motion prediction or them altogether hurts the performance, showing that learning spatial-temporal motion priors in the pre-training stage is important for the downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Effect on the training data. We hypothesize that our pre-training strategy captures strong multimodal correlation and motion priors, which could reduce the reliance on the amount of paired data for downstream tasks. To validate this hypothesis, we follow the setting in Sec. 4.1 and limit" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6205" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 387 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 387 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 387 + ], + "type": "image", + "image_path": "3ae7cad739c0c87b2e6019e90ddeee7bf5951e08bb3828426da5bea469cb3619.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 392, + 546, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 546, + 426 + ], + "type": "text", + "content": "Figure 4. Qualitative example on co-speech gesture generation. Given a speech, we visualize the ground truth 3D motion accompanying the audio, the motion generated by the baseline EMAGE [43], SynTalker [8] and our method. Our model generates more diverse and expressive motion compared to the baseline, especially when the speaker emphasizes on certain words such as \"tired\" and \"because\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 440, + 279, + 577 + ], + "blocks": [ + { + "bbox": [ + 55, + 440, + 279, + 577 + ], + "lines": [ + { + "bbox": [ + 55, + 440, + 279, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 279, + 577 + ], + "type": "image", + "image_path": "409481e1a5cf52e072b5a289cbf0370dfd29c4057d1609813fba0f341de2db29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 582, + 287, + 616 + ], + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 616 + ], + "type": "text", + "content": "Figure 5. Generation performance vs. the amount of post-training data. Our model learns a stronger motion prior from pre-training and thus shows much better under data scarcity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "text", + "content": "the amount of training data available to the model during the pre-training stage. Note that the model has never seen audio2motion data during pre-training. We set the amount of the data to " + }, + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\frac{1}{2^n}" + }, + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "inline_equation", + "content": "n \\in [1 \\dots 5]" + }, + { + "bbox": [ + 46, + 629, + 287, + 700 + ], + "type": "text", + "content": ". We train our full model, our model without pre-training and EMAGE to convergence under each setting and evaluate on the same test set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "The results are shown in Figure 5. We can see that our" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 435, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 533 + ], + "type": "text", + "content": "full model starts with much lower FGD compared with the model without pre-training even when only using 1/32 of the paired training. As expected, as the amount of paired fine-tuning data increases, the performance reduces but our full model always outperforms the w/o pre-training ablation and EMAGE, showing that our model benefits pretraining and shows greater generalization under extreme data scarcity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "type": "text", + "content": "4.3. Unifying Audio-to-Motion and Text-to-Motion for Editable Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 676 + ], + "type": "text", + "content": "By taking a language model-driven approach, our model is capable of following both audio and text prompts. We first train the motion tokenizer on both BEATv2 and AMASS [45] datasets since the range of motion in these two datasets is very different. We use the same tasks for pre-training. For post-training, we combine Audio2Motion and Text2Motion with various instructions, in which text-to-motion with HumanML3D [16] text annotations. See Supp. for details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "By training on both text-to-motion and audio-to-motion data, our model supports joint audio-text prompts, enabling what we call editable gesture generation. This approach" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6206" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 205, + 73, + 373, + 212 + ], + "blocks": [ + { + "bbox": [ + 205, + 73, + 373, + 212 + ], + "lines": [ + { + "bbox": [ + 205, + 73, + 373, + 212 + ], + "spans": [ + { + "bbox": [ + 205, + 73, + 373, + 212 + ], + "type": "image", + "image_path": "8d6516e931f15e775591029844f872e3e2a387e25e957510fac33d1929c8e541.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 546, + 239 + ], + "type": "text", + "content": "Figure 6. Editable gesture generation. We prompt the language with text and audio information and it outputs motions that are both expressional gesture motion as well as general movement motion." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 103, + 192, + 207 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 193, + 97 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 193, + 97 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 193, + 97 + ], + "type": "text", + "content": "Can you generate the lower body movement of someone sitting in a chair and generate the upper body movement according to the speech" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 103, + 192, + 207 + ], + "lines": [ + { + "bbox": [ + 50, + 103, + 192, + 207 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 192, + 207 + ], + "type": "image", + "image_path": "17bd6f7f4152f5814b06b4976a577a72c3b435d66a9f14b1273789d7013af234.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 204, + 98, + 372, + 212 + ], + "blocks": [ + { + "bbox": [ + 204, + 98, + 372, + 212 + ], + "lines": [ + { + "bbox": [ + 204, + 98, + 372, + 212 + ], + "spans": [ + { + "bbox": [ + 204, + 98, + 372, + 212 + ], + "type": "image", + "image_path": "1473874f67c949e1d81bc002c43fc50180be8aa7aaa77b441fad865cee215086.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 373, + 99, + 542, + 212 + ], + "blocks": [ + { + "bbox": [ + 391, + 73, + 528, + 97 + ], + "lines": [ + { + "bbox": [ + 391, + 73, + 528, + 97 + ], + "spans": [ + { + "bbox": [ + 391, + 73, + 528, + 97 + ], + "type": "text", + "content": "Can you generate the lower body movement of a person walking forward and returning backward and generate the upper body movement according to the speech " + }, + { + "bbox": [ + 391, + 73, + 528, + 97 + ], + "type": "inline_equation", + "content": "\\leftarrow +\\uparrow" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 373, + 99, + 542, + 212 + ], + "lines": [ + { + "bbox": [ + 373, + 99, + 542, + 212 + ], + "spans": [ + { + "bbox": [ + 373, + 99, + 542, + 212 + ], + "type": "image", + "image_path": "d631a4a6ff6337bbdd1461b5ee428cbd0d97dd5701cd1e0ba1a2b3fddb7ab6cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 254, + 251, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 254, + 251, + 277 + ], + "spans": [ + { + "bbox": [ + 73, + 254, + 251, + 277 + ], + "type": "text", + "content": "What emotion is conveyed by the movements in the body motion?" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 94, + 278, + 244, + 322 + ], + "blocks": [ + { + "bbox": [ + 94, + 278, + 244, + 322 + ], + "lines": [ + { + "bbox": [ + 94, + 278, + 244, + 322 + ], + "spans": [ + { + "bbox": [ + 94, + 278, + 244, + 322 + ], + "type": "image", + "image_path": "c66b72b5f6e047372197f600d470db08ca8c1c7e619013eccfb630b50f8de487.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 329, + 116, + 337 + ], + "lines": [ + { + "bbox": [ + 80, + 329, + 116, + 337 + ], + "spans": [ + { + "bbox": [ + 80, + 329, + 116, + 337 + ], + "type": "text", + "content": "MotionGPT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 364, + 97, + 372 + ], + "lines": [ + { + "bbox": [ + 80, + 364, + 97, + 372 + ], + "spans": [ + { + "bbox": [ + 80, + 364, + 97, + 372 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 397, + 263, + 408 + ], + "lines": [ + { + "bbox": [ + 70, + 397, + 263, + 408 + ], + "spans": [ + { + "bbox": [ + 70, + 397, + 263, + 408 + ], + "type": "text", + "content": "Figure 7. Qualitative example of emotion prediction." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 82, + 342, + 242, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 342, + 242, + 353 + ], + "spans": [ + { + "bbox": [ + 82, + 342, + 242, + 353 + ], + "type": "text", + "content": "A person dances while moving their hips." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 80, + 376, + 264, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 376, + 264, + 388 + ], + "spans": [ + { + "bbox": [ + 80, + 376, + 264, + 388 + ], + "type": "text", + "content": "The person expresses the emotion of \"contempt\"" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 422, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 567 + ], + "type": "text", + "content": "facilitates the generation of synergistic full-body motions conditioned on both speech and flexibly chosen prompts. For instance, the model can generate the motion of a person walking while talking. In this work, we demonstrate this capability by prompting the model separately for specific body part motions and then combining them seamlessly. Combining conversation gestures with daily motions is extremely useful for applications such as gaming or VR. We show several qualitative examples in Fig. 6. We can see that the model can generate human motion that follows both audio and text prompts, showing the emergent capabilities of our model. See Supp. video for more examples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 574, + 224, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 224, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 224, + 588 + ], + "type": "text", + "content": "4.4. Predicting Emotion from Motion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": "Our model's flexibility in the input/output modality also unlocks an array of new tasks such as translation between different body parts or modalities. In this section, we propose a novel task that predicts emotion from motion." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Reading someone's body language, i.e., predicting emotion from motion is important for applications such as mental health or psychiatry, however, existing audio2motion or motion2text do not have this capability. We extract the emotion labels (neutral, anger, happiness, fear, disgust, sadness, contempt, and surprise) on BEATv2 and convert them" + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 308, + 246, + 529, + 314 + ], + "blocks": [ + { + "bbox": [ + 308, + 246, + 529, + 314 + ], + "lines": [ + { + "bbox": [ + 308, + 246, + 529, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 529, + 314 + ], + "type": "table", + "html": "
Bleu@1↑Rouge Cider↑BertScore↑
GT10010099.9
Random2.454.440.19
MotionGPT1.6810.672.31
Ours14.7126.6716.94
", + "image_path": "d0ddc1f33cd418309f4c1164f87e925edeb276e0351ced2e456d6bde03f6f210.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 316, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 306, + 316, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 545, + 338 + ], + "type": "text", + "content": "Table 3. Motion to emotion. We prompt our model to predict emotion given a motion sequence." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 345, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 345, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 545, + 417 + ], + "type": "text", + "content": "into instructions for training. To be compatible with arbitrary language output from MotionGPT, we evaluate the model performance by measuring the BLEU [50], Rouge Cider [39], and BERTScore [81] between the prediction and the ground truth, which measures the semantic distances between texts. See more details in Supp." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 514 + ], + "type": "text", + "content": "The results are shown in Table 3. MotionGPT entirely fails this task with a performance similar to a random baseline because it was only trained to caption general motion rather than subtle gesture movement and body language. Our model outperforms the random and MotionGPT by a large margin, showing our model's ability to predict the emotion from motion. We also show one qualitative example in Fig. 7." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 525, + 375, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 375, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 375, + 537 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 605 + ], + "type": "text", + "content": "In this work, we propose a novel multimodal language model to unify verbal and non-verbal language with a novel pre-training objective. Our model not only shows state-of-the-art performance on co-speech gestures but also unlocks an array of novel tasks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": "While promising, the model sometimes fails to produce coherent motion potentially due to discrete motion tokenization. Moving forward, we believe incorporating continuous tokenization is an important step to improve the quality of the generated motion." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 666, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 546, + 714 + ], + "type": "text", + "content": "We believe unifying verbal and non-verbal language of human motion generation and understanding is crucial for real-world applications, and language models provide a powerful framework to approach that goal." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6207" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": "Acknowledgments: This project was partially funded by NIH grant R01AG089169 and UST. The authors would also like to thank Georgios Pavlakos for his valuable discussion, Chaitanya Patel, Jingyan Zhang, and Bin Li for their feedback on the paper." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 142, + 105, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 142, + 105, + 154 + ], + "spans": [ + { + "bbox": [ + 48, + 142, + 105, + 154 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 162, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 162, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 287, + 205 + ], + "type": "text", + "content": "[1] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In ECCV, pages 248–265. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 206, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 288, + 326 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 327, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 287, + 360 + ], + "type": "text", + "content": "[3] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 361, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 287, + 403 + ], + "type": "text", + "content": "[4] Nikos Athanasiou, Alpár Ceske, Markos Diomataris, Michael J Black, and Gül Varol. Motionfix: Text-driven 3d human motion editing. arXiv preprint arXiv:2408.00712, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 404, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 449 + ], + "type": "text", + "content": "[5] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In CVPR, pages 15935-15946, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 449, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 287, + 503 + ], + "type": "text", + "content": "[6] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: a language modeling approach to audio generation, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 505, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 505, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 53, + 505, + 287, + 613 + ], + "type": "text", + "content": "[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 614, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 287, + 669 + ], + "type": "text", + "content": "[8] Bohong Chen, Yumeng Li, Yao-Xiang Ding, Tianjia Shao, and Kun Zhou. Enabling synergistic full-body control in prompt-based co-speech motion generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6774-6783, 2024. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 712 + ], + "type": "text", + "content": "[9] Ling-Hao Chen, Shunlin Lu, Ailing Zeng, Hao Zhang, Benyou Wang, Ruimao Zhang, and Lei Zhang. Motionllm: Understanding human behaviors from human motions and videos, 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[10] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "text", + "content": "[11] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[12] Kiran Chhatre, Radek Daněček, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J. Black, and Timo Bolkart. AMUSE: Emotional speech-driven 3D body animation via disentangled latent diffusion. In CVPR, pages 1942-1953, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[13] Seunggeun Chi, Hyung-gun Chi, Hengbo Ma, Nakul Agarwal, Faizan Siddiqui, Karthik Ramani, and Kwonjoon Lee. M2d2m: Multi-motion generation from text with discrete diffusion models. arXiv preprint arXiv:2407.14502, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "text", + "content": "[15] Yuan Gong, Hongyin Luo, Alexander H Liu, Leonid Karlinsky, and James Glass. Listen, think, and understand. arXiv preprint arXiv:2305.10790, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 332, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 376 + ], + "type": "text", + "content": "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, pages 5152-5161, 2022. 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 377, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 421 + ], + "type": "text", + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597. Springer, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "text", + "content": "[18] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In CVPR, pages 1900-1910, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "text", + "content": "[19] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video, 2021. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 513, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 577 + ], + "type": "text", + "content": "[20] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29: 3451-3460, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[22] Rongjie Huang, Mingze Li, Dongchao Yang, Jiatong Shi, Xuankai Chang, Zhenhui Ye, Yuning Wu, Zhiqing Hong, Jiawei Huang, Jinglin Liu, et al. Audiogpt: Understanding and generating speech, music, sound, and talking head. In AAAI, pages 23802-23804, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[23] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2, 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6208" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[24] Biao Jiang, Xin Chen, Chi Zhang, Fukun Yin, Zhuoyuan Li, Gang Yu, and Jiayuan Fan. Motionchain: Conversational motion controllers via multimodal prompts. In ECCV, pages 54-74. Springer, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "text", + "content": "[25] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1737-1747, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "text", + "content": "[26] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, pages 2151-2162, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 263 + ], + "type": "text", + "content": "[27] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 297 + ], + "type": "text", + "content": "[28] T Kudo. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "type": "text", + "content": "[29] Taku Kudo. Subword regularization: Improving neural network translation models with multiple subword candidates. arXiv preprint arXiv:1804.10959, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "type": "text", + "content": "[30] Gen Li, Kaifeng Zhao, Siwei Zhang, Xiaozhong Lyu, Mihai Dusmanu, Yan Zhang, Marc Pollefeys, and Siyu Tang. EgoGen: An Egocentric Synthetic Data Generator. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "type": "text", + "content": "[31] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 288, + 510 + ], + "type": "text", + "content": "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 554 + ], + "type": "text", + "content": "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 600 + ], + "type": "text", + "content": "[34] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 644 + ], + "type": "text", + "content": "[35] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 678 + ], + "type": "text", + "content": "[36] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[37] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[38] Han Liang, Jiacheng Bao, Ruichi Zhang, Sihan Ren, Yuecheng Xu, Sibei Yang, Xin Chen, Jingyi Yu, and Lan Xu. Omg: Towards open-vocabulary motion generation via mixture of controllers. In CVPR, pages 482-493, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "text", + "content": "[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 151, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 215 + ], + "type": "text", + "content": "[40] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3764-3773, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 260 + ], + "type": "text", + "content": "[41] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In ECCV, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 282 + ], + "type": "text", + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 283, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 338 + ], + "type": "text", + "content": "[43] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J. Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In CVPR, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 381 + ], + "type": "text", + "content": "[44] Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-world text-to-motion generation. arXiv preprint arXiv:2312.14828, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 425 + ], + "type": "text", + "content": "[45] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "text", + "content": "[46] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 503 + ], + "type": "text", + "content": "[47] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In ICCV, 2023. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "text", + "content": "[48] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "type": "text", + "content": "[49] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 603, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 656 + ], + "type": "text", + "content": "[50] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6209" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[52] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 183 + ], + "type": "text", + "content": "[53] Ryan Po, Wang Yifan, Vladislav Golyanik, Kfir Aberman, Jonathan T Barron, Amit Bermano, Eric Chan, Tali Dekel, Aleksander Holynski, Angjoo Kanazawa, et al. State of the art on diffusion models for visual computing. In Computer Graphics Forum, page e15063. Wiley Online Library, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "text", + "content": "[54] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 3, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "text", + "content": "[55] Rico Sennrich. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 294 + ], + "type": "text", + "content": "[56] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 327 + ], + "type": "text", + "content": "[57] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H. Bermano. Human motion diffusion model. In ICLR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "text", + "content": "[58] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 404 + ], + "type": "text", + "content": "[59] Minh Tran, Di Chang, Maksim Siniukov, and Mohammad Soleymani. Dyadic interaction modeling for social behavior generation. In ECCV, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "type": "text", + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 502 + ], + "type": "text", + "content": "[61] Yuan Wang, Di Huang, Yaqi Zhang, Wanli Ouyang, Jile Jiao, Xuetao Feng, Yan Zhou, Pengfei Wan, Shixiang Tang, and Dan Xu. Motionpt-2: A general-purpose motion-language model for motion generation and understanding. arXiv, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "text", + "content": "[62] Zhenzhi Wang, Jingbo Wang, Dahua Lin, and Bo Dai. Intercontrol: Generate human motion interactions by controlling every joint. arXiv preprint arXiv:2311.15864, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 288, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 603 + ], + "type": "text", + "content": "[63] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say, interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 635 + ], + "type": "text", + "content": "[64] Qi Wu, Yubo Zhao, Yifan Wang, Yu-Wing Tai, and Chi-Keung Tang. Motionllm: Multimodal motion-language learning with large language models, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "text", + "content": "[65] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[66] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "human motion generation. arXiv preprint arXiv:2310.08580, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[67] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In CVPR, pages 12780-12790, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 218 + ], + "type": "text", + "content": "[68] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffuse-style gesture: Stylized audio-driven co-speech gesture generation with diffusion models. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 5860–5868. International Joint Conferences on Artificial Intelligence Organization, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[69] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[70] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J. Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "text", + "content": "[71] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics, 39(6), 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 399 + ], + "type": "text", + "content": "[72] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, pages 16010-16021, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[73] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 510 + ], + "type": "text", + "content": "[74] Juze Zhang, Haimin Luo, Hongdi Yang, Xinru Xu, Qianyang Wu, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Neuraldome: A neural modeling pipeline on multi-view human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8834-8845, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "text", + "content": "[75] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "type": "text", + "content": "[76] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[77] Juze Zhang, Jingyan Zhang, Zining Song, Zhanhe Shi, Chengfeng Zhao, Ye Shi, Jingyi Yu, Lan Xu, and Jingya Wang. Hoi-m^3: Capture multiple humans and objects interaction within contextual environment. In CVPR, pages 516-526, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[78] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6210" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 441 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[79] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In ICCV, pages 364-373, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 162 + ], + "type": "text", + "content": "[80] Mingyuan Zhang, Daisheng Jin, Chenyang Gu, Fangzhou Hong, Zhongang Cai, Jingfang Huang, Chongzhi Zhang, Xinying Guo, Lei Yang, Ying He, et al. Large motion model for unified multi-modal motion generation. arXiv, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "text", + "content": "[81] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "type": "text", + "content": "[82] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 285 + ], + "type": "text", + "content": "[83] Zhikai Zhang, Yitang Li, Haofeng Huang, Mingxian Lin, and Li Yi. Freemotion: Mocap-free human motion synthesis with multimodal large language models. In ECCV, pages 403-421. Springer, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 286, + 352 + ], + "type": "text", + "content": "[84] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 353, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 353, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 49, + 353, + 286, + 385 + ], + "type": "text", + "content": "[85] Mohan Zhou, Yalong Bai, Wei Zhang, Ting Yao, Tiejun Zhao, and Tao Mei. Responsive listening head generation: A benchmark dataset and baseline. In ECCV, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 387, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 387, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 286, + 441 + ], + "type": "text", + "content": "[86] Wenyang Zhou, Zhiyang Dou, Zeyu Cao, Zhouyingcheng Liao, Jingbo Wang, Wenjia Wang, Yuan Liu, Taku Komura, Wenping Wang, and Lingjie Liu. Emdm: Efficient motion diffusion model for fast and high-quality motion generation. In ECCV, pages 18-38. Springer, 2025. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6211" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_content_list.json b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d5f3a400ec6d7723b444b83519207b8f05a3cf31 --- /dev/null +++ b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_content_list.json @@ -0,0 +1,1162 @@ +[ + { + "type": "text", + "text": "The PanAf-FGBG Dataset:", + "text_level": 1, + "bbox": [ + 359, + 104, + 638, + 125 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition", + "text_level": 1, + "bbox": [ + 109, + 128, + 888, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Otto Brookes $^{1,2\\dagger*}$ , Maksim Kukushkin $^{3,4*}$ , Majid Mirmehdi $^{1}$ , Colleen Stephens $^{5}$ , Paula Dieguez $^{5}$ , Thurston C. Hicks $^{6}$ , Sorrel Jones $^{5}$ , Kevin Lee $^{5}$ , Maureen S. McCarthy $^{5}$ , Amelia Meier $^{5}$ , Emmanuelle Normand $^{2}$ , Erin G. Wessling $^{7}$ , Roman M. Wittig $^{5,8}$ , Kevin Langergraber $^{9}$ , Klaus Zuberbühler $^{10}$ , Lukas Boesch $^{2}$ , Thomas Schmid $^{3,11}$ , Mimi Arandjelovic $^{5}$ , Hjalmar Kuhl $^{5,12}$ , Tilo Burghardt $^{1}$", + "bbox": [ + 114, + 154, + 879, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Bristol, 2Wild Chimpanzee Foundation, 3Martin Luther University Halle-Wittenberg, 4Leipzig University, 5Max Planck Institute for Evolutionary Anthropology, 6University of Warsaw, 7Harvard University, 8University of Lyon, 9Arizona State University, 10University of St Andrews, 11Lancaster University in Leipzig, 12Senckenberg Museum of Natural History", + "bbox": [ + 166, + 253, + 826, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 358, + 326, + 373 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computer vision analysis of camera trap video footage is essential for wildlife conservation, as captured behaviours offer some of the earliest indicators of changes in population health. Recently, several high-impact animal behaviour datasets and methods have been introduced to encourage their use; however, the role of behaviour-correlated background information and its significant effect on out-of-distribution generalisation remain unexplored. In response, we present the PanAf-FGBG dataset, featuring 21 hours of wild chimpanzee behaviours, recorded at 389 individual camera locations. Uniquely, it pairs every video with a chimpanzee (referred to as a foreground video) with a corresponding background video (with no chimpanzee) from the same camera location. We present two views of the dataset: one with overlapping camera locations and one with disjoint locations. This setup enables, for the first time, direct evaluation of in-distribution and out-of-distribution conditions, and for the impact of backgrounds on behaviour recognition models to be quantified. All clips have rich behavioural annotations and metadata, including unique camera IDs. Additionally, we establish several baselines and present a highly effective latent-space normalisation technique that boosts out-of-distribution performance by $+5.42\\%$ mAP for convolutional and $+3.75\\%$ mAP for transformer-based models. Finally, we provide an in-depth analysis on the role of backgrounds in out-of-distribution behaviour recognition, including the so far unexplored impact of background durations. (i.e., the count of background frames within foreground videos). The dataset is available at https://obrookes.github.io/panaf-fbgb.github.io/.", + "bbox": [ + 86, + 392, + 485, + 845 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5f14db0bf10262b794a2615c63ddc6abd774994f8b9277f8062d80850ed8d50.jpg", + "image_caption": [ + "Figure 1. Conceptual Overview. The PanAf-FGBG dataset comprises $>20$ hours of paired and richly annotated foreground-background camera trap videos of wild chimpanzees. The dataset unlocks systematic analyses of the impact of background information on wildlife behaviour recognition. We provide baselines, quantify the background impact on performance, and demonstrate that utilising background information in latent space (see blue data point) can significantly improve recognition performance in this challenging domain." + ], + "image_footnote": [], + "bbox": [ + 516, + 358, + 906, + 722 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author, *Equal technical contribution", + "bbox": [ + 114, + 886, + 401, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5433", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning models have shown impressive results on action recognition tasks [57], particularly when both training and test data share the same distribution. However, their performance declines significantly on out-of-distribution (OOD) data [15, 27]. In practice, distribution shifts are common in nearly all settings [43], but are particularly unforgiving when computer vision is applied to uncontrolled, complex or natural environments [27].", + "bbox": [ + 89, + 114, + 480, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the human action recognition domain, poor OOD performance is well-known to originate from shortcut learning [18] of static cues or backgrounds [24, 29, 30]. This leads models to rely heavily on background information instead of the behaviour of interest solely [2], and several studies have shown that the majority of action recognition performance in this domain can be attributed to background learning [12, 22, 49]. Despite the known impact of backgrounds on action recognition and a critical need to advance model performance in wildlife applications [28, 42, 48], reliance on static cues for animal behaviour recognition has received limited attention. This is largely due to a lack of datasets explicitly designed to study its effects.", + "bbox": [ + 89, + 236, + 482, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response, we introduce the PanAf-FGBG dataset, which comprises real-world foreground-background video pairs (see Fig. 1) extracted from six African countries, 14 national parks, and over 350 individual camera locations across many habitats and location types (see Fig. 3). Here, we define foreground videos as those containing a chimpanzee, and background videos as those that do not. This allows us to address a key limitation in existing studies, where the background's role is typically assessed by creating background-only versions of datasets through synthetic modifications (e.g., masking or in-painting) that remove the actor [11, 12, 23]. While such synthetic inputs often serve as useful proxies, the actual effect of synthesis on training is largely unknown [39, 47].", + "bbox": [ + 89, + 431, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, our dataset introduces paired camera locations and enables a truly realistic disentangling of background effects under in- and out-of-distribution conditions: disjoint experiments with mutually exclusive camera locations for training and test splits, and overlapping experiments with shared locations. This focus is particularly important in the wildlife domain, where the transferability and generalisation of camera trap analysis to new locations is critical for effective operation. While studies have shown the over-reliance of action recognition models on backgrounds [22, 37, 49], we find that the effect of background duration (i.e., the count of background frames within foreground videos with no actor) is largely unexplored. In camera trap video footage like in many other CCTV or triggered camera settings, actors may only appear momentarily, leaving behind footage of only the background [14, 38] within foreground videos. Since this effect alters learning given", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "video-level behaviour annotations, we quantify its effect on several action recognition architectures.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows: (i) we release PanAf-FGBG with its rich annotations and meta-data, providing a large wildlife dataset suitable for explicitly investigating the impact of backgrounds on behaviour recognition; (ii) we show models trained on background-only videos achieve strong animal behaviour recognition performance ( $\\sim 65\\%$ of baseline performance), corroborating findings in the human domain; (iii) we conduct the first evaluation of background duration in videos, revealing its impact on action recognition models and highlighting key differences between convolutional and attention-based networks; and (iv) we introduce a simple yet powerful latent-space background neutralisation that yields notable performance gains, particularly in OOD scenarios for convolutional $(+5.42\\%)$ mAP) and transformer-based approaches $(+3.75\\%)$ mAP).", + "bbox": [ + 511, + 121, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 361, + 653, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Animal Behaviour Datasets. In recent years, several high-impact animal behaviour datasets have been introduced to the CV community, often accompanied by new methods [4, 6-8, 10, 26, 33, 40, 46]. Although they all centre on animal behaviour understanding, they each have a different focus. The large-scale datasets Animal Kingdom [40] and MammalNet [10] concentrate on behavioural understanding across species, each including videos of $\\sim 800$ different species. They provide atomic action and high-level behaviour annotations, respectively, and support, in addition to action recognition, the related tasks of action localisation [10, 40] and pose estimation [40]. The smaller ChimpAct dataset [36] provides longitudinal data that follows a zoo-housed group of 20 chimpanzees with a focus on a single male juvenile. It also contains key-point pose information, and fine-grained spatio-temporal behaviour labels. While valuable, these datasets lack in-situ footage from ecological sensors like camera traps and drones, meaning animals are not observed in their natural environment. This may result in unrepresentative behaviour distributions and limit their suitability for studying wild behaviour [9, 13].", + "bbox": [ + 511, + 387, + 903, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, KABR [26] and BaboonLand [17] comprise drone footage of free-living Kenyan wildlife, with fine-grained spatio-temporal behaviour annotations, focussing on individual and group-level behaviour, respectively. The LoTE dataset [33] provides camera trap video footage of 11 endangered species from a 12-year longitudinal study; it is accompanied by environmental information (i.e., weather conditions and habitat) and highlights the importance of temporal and spatial continuity for conservation efforts. The dataset most comparable to our own is the PanAf20k dataset [8] and its related PanAf500-based works [46, 52, 53]. PanAf20k comprises camera trap videos of great apes accompanied by multi-label behaviour annotations.", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5434", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our dataset orthogonally complements this growing body of work dedicated to understanding animal behaviour by providing the first dataset specifically designed to evaluate backgrounds and out-of-distribution generalisation for behaviour recognition. It is unique in providing foreground-background pairs, and its configurations for in- and out-of-distribution evaluation are defined identically to the well-known iWildCam [5] and Wilds dataset [27] for species classification. Similar to LotE [33], we provide a detailed description of the background, however we also include a unique camera ID (i.e., geospatial location) to allow footage taken from the same camera to be linked. Similar to [8], our data is gathered from 14 national parks, spanning 6 different countries, although unlike ours they do not include metadata to allow videos to be linked to specific countries, research sites, locations, or habitats. Similarly, it is more diverse than [33] with its footage from a single national park.", + "bbox": [ + 89, + 90, + 483, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Background Impact. The effect of backgrounds or static bias is well-studied within computer vision [29, 54], where it has shown to be necessary for models to perform well [51]. Action recognition models are particularly vulnerable to shortcut learning. These models often rely on correlated, but distinct information beyond the intended action, such as object [20, 56], scene [37], and actor biases [31]. For example, several studies have reported that strong performance on action recognition tasks can be achieved without the actor [22, 37, 49] and that approximately $70\\%$ of model performance is attributable to static cues (i.e., scene or background) [12]. This results in biased representations and ultimately impairs out-of-distribution generalization performance [30].", + "bbox": [ + 89, + 351, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Background Bias Datasets. The effect of the background has often been modelled using synthetically generated data that removes either the foreground or the background. [11, 12, 19, 29]. For instance, Ilic et al. [24] present a synthetic version of the UCF101 dataset, animating spatial noise using image motion extracted by an optical flow estimator, Li et al. [29] replace the backgrounds of IID HMDB51 with sinusoidal stripe images, and Chung et al [12] adopt an in-painting model to remove actors from the video. Datasets that include fine-grained spatio-temporal action annotations also remove static cues implicitly by isolating the region of the video where the action is occurring.", + "bbox": [ + 89, + 565, + 483, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mitigating Background Bias. Prominent attempts to mitigate this bias include dataset resampling methods, which down-weight biased samples [30, 31], adversarial training regimes that rely on scene labels and human masks [11], and augmentation techniques that mix or swap the backgrounds of different videos during training [16, 29, 50]. There are also several implicit approaches, such as frame selection, that aim to utilise only frames where activity is occurring [55]. Our dataset, introduced next, instead introduces foreground-background pairing in", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "formation with close proximity in time of capture, which describes the environment truly independently and in relation to the presence of animals showcasing behaviour.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 511, + 148, + 602, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset Overview and Statistics. PanAf-FGBG comprises footage gathered under ethical oversight as part of the PanAf Programme: The Cultured Chimpanzee [1]. It contains 21 hours of camera trap footage of individual chimpanzees in tropical Africa. Its footage is collected from 389 individual camera locations across 14 national parks in 6 African countries. In total, we provide 5,070 video pairs, each 15 seconds in duration. All videos are accompanied by a multi-label behaviour annotation and scene metadata, plus the unique feature of foreground-background video pairing (see Fig. 1). Class and habitat distributions of the data are given in Fig. 2 and Fig. 3, respectively. Our dataset follows a long-tailed distribution [34, 41] common in animal behaviour datasets [8, 10, 33, 36, 40]. The three of the most commonly occurring classes are observed in $>60\\%$ of videos, while the rarest are observed in fewer than $< 3\\%$ .", + "bbox": [ + 511, + 169, + 906, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Collection. The data collection strategy was designed specifically to maximise the chance of filming the terrestrial activity of apes and we provide only a brief description here (for full details we refer the reader to [1]). Data was collected from 14 different research sites, where an average of 29 movement-triggered Bushnell cameras were installed per site. Grids comprising 20 to $961 \\times 1$ km cells were established and one camera was installed per grid cell. Footage was recorded at 30 frames-per-second and a resolution of $1280 \\times 720$ . The cameras were visited every 1-3 months for maintenance and to download the recorded footage throughout the study periods.", + "bbox": [ + 511, + 412, + 908, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ethical Considerations. To protect the location of the apes from threats (e.g., poaching) we use monikers for the names of the research sites and assign a unique camera ID to each video, hiding the geospatial location of the cameras.", + "bbox": [ + 511, + 592, + 908, + 654 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/69e55246557040f4101ce07536dce39f1e1a9b9b869b1e47ce1baeae86eec24b.jpg", + "image_caption": [ + "Figure 2. Distribution of Behaviour. Proportion of behaviours in the dataset where smaller segments are colour coordinated." + ], + "image_footnote": [], + "bbox": [ + 531, + 676, + 890, + 857 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5435", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/044ebbe43aa5fbd39b131aa99a39f6bd76127a9a1f2ce571bfa8a12ac2ccea3b.jpg", + "image_caption": [ + "Figure 3. Distribution of Habitat and Location Type. The pie chart gives the proportion of videos taken at each of the three main habitats: forest, marshes and savannah. The histogram shows the total number of videos extracted from each location type." + ], + "image_footnote": [], + "bbox": [ + 96, + 97, + 475, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Annotation. The behavioural annotations were provided by users on the community science platform Chimp&See [3]. Annotators were presented with a choice of behaviour classification categories, chosen specifically by experts for their ecological importance. Behaviours not listed in the classification categories were also permitted. These annotations were then extracted and expertly grouped into 14 co-occurring classes, which form the multi-label behavioural annotations presented here. The annotations follow a multi-hot binary format that indicate the presence of one or more behaviours. It should also be noted that behaviours are not assigned to individual apes or temporally localised within each video. To ensure annotation quality and consistency a video was only deemed to be analysed when either three volunteers marked the video as blank, or unanimous agreement between seven volunteers was observed, or 15 volunteers annotated the video. Habitat type was noted during camera setup.", + "bbox": [ + 88, + 492, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Foreground-Background Pair Selection. Pairs are formed by selecting the temporally closest available foreground and background videos to ensure similarity. Complete standardisation statistics can be found in Sup. Mat.", + "bbox": [ + 89, + 763, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overlapping & Disjoint Configurations. We present both a disjoint view of dataset, where camera locations in the training and test splits are mutually exclusive, denoted as $\\mathcal{D}^{disjoint} = \\{d_i \\mid d_i \\cap d_j = \\emptyset, \\forall i \\neq j\\}$ following [27]. A second, overlapping view is also provided, where camera", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "locations are shared between training and test splits, represented as $\\mathcal{D}^{overlap} = \\{d_i \\mid d_i \\cap d_j \\neq \\emptyset$ for some $i \\neq j\\}$ . Note that $d_i$ refers to sets of camera locations associated with videos in each split. Fig. 4 provides an overview of both configurations. We ensure that the class distribution in each configuration remains approximately the same for full comparability. Unique camera identifiers are available for each video to allow researchers interested in this problem to create additional configurations, should they wish to.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 250, + 651, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train multi-label classification models on both views of the dataset: $\\mathcal{D}^{\\mathrm{overlap}}$ and $\\mathcal{D}^{\\mathrm{disjoint}}$ using various combinations of foreground $(\\mathcal{F})$ , background $(\\mathcal{B})$ and synthetic background $(\\tilde{\\mathcal{B}})$ videos. The latter were created by generating a segmentation mask using SAM2 [44] followed by mean pixel value filling, similar to [11, 12, 23]. Both views comprise foreground-background video pairs $\\{(v_{i}^{\\mathcal{F}}, v_{i}^{\\mathcal{B}})\\}_{i=1}^{N}$ , where $v_{i}^{\\mathcal{F}} \\in \\mathcal{F}$ and $v_{i}^{\\mathcal{B}} \\in \\mathcal{B}$ , with $N$ representing the total number of video pairs. Each foreground video $v_{i}^{\\mathcal{F}}$ is associated with a video-level label vector $y_{i} = (y_{i1}, y_{i2}, \\ldots, y_{ic})$ , where each $y_{ij} \\in \\{0,1\\}$ indicates the presence (1) or absence (0) of the $j$ -th class in video $i$ , and $C$ is the total number of classes.", + "bbox": [ + 511, + 279, + 908, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 500, + 645, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 527, + 596, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All experiments were performed using the ResNet-50 [21] and MViT-V2 [32] architectures, which were initialised with feature extractors pre-trained on Kinetics-400 [25]. For ResNet-50, both 2D and 3D convolution versions are tested. After initial experiments to ensure optimal optimiser-network pairings for the task, ResNet and MViT-V2 models were fine-tuned for 300 epochs using the SGD [45] and AdamW [35] optimisers, respectively. Both models followed a linear warm-up schedule with subsequent cosine annealing. For ResNet models, the learning rate was increased from $1\\times 10^{-5}$ to $1\\times 10^{-4}$ over 20 epochs, while it was increased from $1\\times 10^{-6}$ to $1\\times 10^{-4}$ over 30 epochs for MViT-V2 models. In both cases, momentum of 0.9 was utilised. All computation was performed on $4\\times$ NVIDIA H200 GPUs using a distributed batch size of 336. During training, video tensors with a spatio-temporal resolution of $16\\times 256\\times 256$ were employed. Frames were sampled uniformly at a rate of $1/24$ , avoiding dense sampling strategies which are shown to exacerbate shortcut learning in action recognition models [12]. For MViT-V2 models, a patch size of $16\\times 16$ pixels was utilised. Evaluation was performed using micro and mean Average Precision and are referred to as uAP and mAP, respectively.", + "bbox": [ + 511, + 551, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5436", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/02d4d3d3405d48170cb7f18b37bf5f1e938866b7e0a4d83fcc726664f2caae48.jpg", + "image_caption": [ + "Figure 4. Overlapping & Disjoint Dataset Configurations. Six foreground-background video pairs are shown to visualise overlapping (left) and disjoint (right) dataset configurations. On the left, train and test videos taken from the same camera are selected, emphasizing that the locations are not mutually exclusive, with backgrounds containing several cues that are highly correlated with the displayed behaviours (i.e., algae fishing, a form of tool use). Train and test videos on the right are extracted from mutually exclusive or disjoint camera locations. They are extracted from forest and savannah, respectively, and highlight the challenge of generalising to new locations." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Background Reliance", + "text_level": 1, + "bbox": [ + 89, + 491, + 290, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Overview. First, we examined the reliance of animal behaviour recognition on the background by training model architectures on the foreground $\\mathcal{F}$ , background $\\mathcal{B}$ , and synthetic background videos $\\tilde{\\mathcal{B}}$ only. We refer to models trained on $\\mathcal{F}$ and $\\mathcal{B}$ (or $\\tilde{\\mathcal{B}}$ ) as FG-only and BG-only models, respectively, as shown earlier in Fig. 1. To separate background information from class-distribution intrinsic information, a dummy classifier with no access to visual information was also evaluated. It would effectively return predictions based on the class distribution information alone. Since our dataset consists of unique foreground-background pairs, it was possible to train BG-only models by retaining the original labels associated with the foreground video of the pairing. In all cases, models were tested on $\\mathcal{F}$ for comparability. Note that this makes the task more difficult for BG-only models, since animals are only observed at test time and not during training. All experiments are conducted on both $D^{overlap}$ and $D^{disjoint}$ .", + "bbox": [ + 88, + 518, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dual-Stream Fusion Model $(\\mathcal{F} + \\mathcal{B})$ . To clarify whether BG-only models and FG-only models, each trained separately, can be fused late to orthogonally contribute towards performance we also implemented a dual-stream model. To do this, the independently optimised weights of the FG-only and BG-only models trained for 300 epochs were loaded. During training the weights were frozen and video pairs", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "were processed to extract late features $z^{F}$ and $z^{B}$ . For ResNet models, features were extracted from the last convolutional layer and in MViT-V2 models an average of the tokens from the last hidden state were taken. Late features were then concatenated $\\tilde{z} = [z^{F}, z^{B}]$ and final class predictions were produced by a trainable multi-layer perceptron.", + "bbox": [ + 511, + 491, + 906, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Result 1 - Backgrounds are a Predictor of Animal Behaviour. Table 1 shows that BG-only models achieve strong performance when compared to their respective FG-only counterparts. To illustrate this, the ratio of background-toForeground performance $\\frac{AP^B}{AP^F}$ published in [12], showing how close a model can get to FG performance with just BG cues, is reported for each architecture in rows 6, 11, and 16. Background-to foreground performance ratios are consistently high across all models, never dropping below 0.70 and 0.65 for uAP and mAP, respectively, on either dataset. This indicates that much of the behaviour recognition performance is indeed achievable by just utilisation of background information - the environment is a strong predictor of behaviour, even out-of-distribution. Thus, features of the environment are already behaviour predictors (trees for climbing, paths for travel, termite mound for feeding etc.). This result is further compounded by the fact that BG-only models outperform the non-visual classifier (row 1) by a significant margin, particularly on $D^{overlap}$ . On $D^{disjoint}$ this observation still holds, although by somewhat", + "bbox": [ + 511, + 595, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5437", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6758e7e6dccf12eb4e99011380a09492f11fe2acadf6f56b4deb13195f48518b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelData\\(D^{overlap}\\)\\(D^{disjoint}\\)
uAPmAPuAPmAP
1Dummy-19.1514.7319.6114.87
2\\(\\bar{\\mathcal{F}}\\)64.1541.2844.6718.08
32D R50\\(\\mathcal{B}\\)57.1330.0345.8018.27
4\\(\\tilde{\\mathcal{B}}\\)57.8730.8248.2820.39
5\\(\\mathcal{F}+\\mathcal{B}\\)67.7444.9945.2818.84
60.890.741.020.98
7\\(\\bar{\\mathcal{F}}\\)72.7949.3257.7430.89
83D R50\\(\\mathcal{B}\\)58.9231.7248.4021.38
9\\(\\tilde{\\mathcal{B}}\\)59.6032.8649.2921.64
10\\(\\mathcal{F}+\\mathcal{B}\\)76.4153.5765.6639.17
110.800.660.830.69
12\\(\\bar{\\mathcal{F}}\\)75.4451.5470.5035.29
13MViT-V2\\(\\mathcal{B}\\)61.9032.7450.0324.00
14\\(\\tilde{\\mathcal{B}}\\)62.1533.5852.1324.77
15\\(\\mathcal{F}+\\mathcal{B}\\)72.8253.4968.0038.09
160.820.650.700.68
", + "bbox": [ + 107, + 88, + 472, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantifying Background Reliance. Performance comparison of models trained on $\\mathcal{F}$ , $\\mathcal{B}$ , and $\\tilde{\\mathcal{B}}$ videos. Dual-stream models are indicated by $\\mathcal{F} + \\mathcal{B}$ . The absolute highest performance is indicated in bold whereas best performance using only a single-stream is underlined. The performance based on class distribution only is given by a dummy classifier (see row 1). The background-toForeground performance ratio $(\\frac{AP^B}{AP^F})$ is also provided (see rows 6, 11, and 16). Performance is reported on $\\mathcal{D}_{\\text{overlap}}$ and $\\mathcal{D}_{\\text{disjoint}}$ where Fig. 1 visualises an mAP scatter plot between the two for 3D R50.", + "bbox": [ + 89, + 340, + 486, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "smaller margins. Note that for the weakest model (2D R50) background mAP is even on-par or above the foreground baseline (see rows 2-3 in the final column). Together, these results illustrate powerfully, that across architectures background information alone is employable as a strong animal behaviour predictor.", + "bbox": [ + 89, + 521, + 482, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Result 2 - Multi-scale Vision Transformers rely less on backgrounds. On the overlapping dataset, the 3D-R50 and MViT-V2 models (rows 7-16) display significantly better performance when compared with the 2D-R50 (rows 2-6), and perform similarly to each other with respect to uAP and mAP. However, the MViT-V2 model widely outperforms the 3D R50 on the disjoint dataset, suggesting that it does not rely as heavily on background features for performance. In support of this, while all models experience a large decrease in performance when moving from the overlapping to the disjoint dataset (generalisation gap), the smallest impact is on the MViT-V2. These differences between architectures are highlighted most clearly by observing uAP on the disjoint dataset. Here, the background only ratio for the MViT-V2 model is $\\sim 30\\%$ and $\\sim 13\\%$ lower than the 2D-R50 and 3D-R50, respectively. MViT-V2 achieves the best overall performance and exhibits the lowest background-toForeground performance ratio on both dataset configurations, corroborating similar findings from", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the still image domain that show transformers are more invariant to changes in the background [39].", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Result 3 - Analysing Background Information Contributions. The dual 3D-R50 model achieves significant performance gains over its FG-only baseline, particularly on $\\mathcal{D}^{disjoint}$ , where increases of $7.92\\%$ and $8.28\\%$ in uAP and mAP, respectively, can be observed. Smaller, but still significant improvements are also observed on $\\mathcal{D}^{overlap}$ ( $+3.62\\%$ uAP and $+4.25\\%$ mAP). Interestingly, the performance of the dual-stream 3D-R50 exceeds that of MViT-v2, including its dual-stream counterpart, on nearly all metrics except for disjoint uAP. The dual-stream MViT-V2 model, on the other hand, shows improved mAP but a lower uAP when compared with the baseline model, a trend consistent across both dataset configurations. This analysis indicates that BG-only performance in the wildlife domain is not a subset of the FG-only model – which of course also has access to background information. Instead, explicitly learned (and frozen) background features can aid the FG-only baseline in models such as ResNet via late fusion. Finally, all BG-only models trained on $\\tilde{\\mathcal{B}}$ seem to outperform those trained on $\\mathcal{B}$ . However, we will show later in Sec. 5.4 that non-synthetic background videos dominate synthetic counterparts when explicitly utilised for performance improvements above the foreground baselines.", + "bbox": [ + 511, + 123, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Background Duration", + "text_level": 1, + "bbox": [ + 513, + 483, + 720, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Overview. Noting that foreground videos may contain some frames with no animals, effectively supplying background only information for these frames, we aim at quantifying the effect of this on model performance (see Fig. 5). Practically, this is achieved by concatenating increasing numbers of background frames to our foreground videos. To do this, we sampled a set of $\\lambda T$ frames from $v^{B}$ , where $\\lambda$ is a predefined factor controlling the proportion of background frames used. The sampled background frames are denoted as $V^{\\mathcal{B}} = \\{V_{i_0}^{\\mathcal{B}}, V_{i_1}^{\\mathcal{B}}, \\dots, V_{i_{\\lambda T - 1}}^{\\mathcal{B}}\\}$ , with each index $i_j$ selected from $\\{0, \\dots, T - 1\\}$ . The foreground frames $V^{\\mathcal{F}} = \\{F_0^{\\mathcal{F}}, F_1^{\\mathcal{F}}, \\dots, F_{T - 1}^{\\mathcal{F}}\\}$ were then concatenated with $V^{\\mathcal{B}}$ , resulting in a combined video sequence $V^{\\mathrm{concat}} = \\{F_0^{\\mathcal{F}}, F_1^{\\mathcal{F}}, \\dots, F_{T - 1}^{\\mathcal{F}}, F_{i_0}^{\\mathcal{B}}, F_{i_1}^{\\mathcal{B}}, \\dots, F_{i_{\\lambda T - 1}}^{\\mathcal{B}}\\}$ of length $T + \\lambda T$ . From this concatenated sequence, a starting frame $F_{\\mathrm{start}}$ was sampled at random from the indices $\\{0, \\dots, \\lambda T\\}$ , ensuring that $T$ consecutive frames could be sampled without exceeding the sequence bounds. The final sampled frame sequence $S = \\{F_{\\mathrm{start}}, F_{\\mathrm{start + 1}}, \\dots, F_{\\mathrm{start + T - 1}}\\}$ was then used as input for training while retaining the original label $y$ from the foreground video $v^{\\mathcal{F}}$ . Background frames are appended because motion-triggered cameras are more likely to capture activity first, resulting in a more natural ordering.", + "bbox": [ + 511, + 506, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysing the Continuum Between Foreground and Background Videos. Fig. 5 shows the performance com", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5438", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1928c8b96a524535c5ff3fd1bd4dc1d253149a4a184dbaf9bfbc3513d74c4943.jpg", + "image_caption": [ + "$\\mathcal{D}_{overlap}$" + ], + "image_footnote": [], + "bbox": [ + 91, + 107, + 287, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca755ab717380fb21982308ca388d20690ed7146a490171c4deace5f5a0b0160.jpg", + "image_caption": [ + "$\\mathcal{D}_{disjoint}$" + ], + "image_footnote": [], + "bbox": [ + 285, + 107, + 480, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/704097270d42955ac054993dcc057dc994d23deb2be4a9b90c5729dcd00d8c16.jpg", + "image_caption": [ + "Figure 5. Effect of Increasing Background Duration on Performance. Comparison of 2D R50, 3D R50 and MViT-V2 model performance as a function of the simulated background duration $(\\lambda)$ . The 3D-R50 is more sensitive to increasing background durations than MViT-V2. MViT-V2 is relatively robust to increasing background durations on the overlapping dataset, although it becomes highly sensitive when evaluated on OOD data." + ], + "image_footnote": [], + "bbox": [ + 91, + 220, + 287, + 334 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37f9a6acb610c604010d16e9bac2cec7f82ab8c14e47d98df28834911cba570e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 220, + 480, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "parison between the 2D R50, 3D R50 and MViT-V2 models as background duration increases in videos $(\\lambda)$ . As expected, Fig. 5 confirms that as $\\lambda$ increases model performance declines towards background model performance (see Tab 1). On $D^{overlap}$ , the 3D-R50 and MViT-V2 models exhibit a sharp decline before beginning to converge. This occurs at $\\lambda = 32$ and $\\lambda = 64$ for the 3D-R50 and MViT-V2 models, respectively. Additionally, as shown by the interval between $\\lambda = 0$ and $\\lambda = 32$ , the uAP and mAP of 3D-R50 falls by $16.88\\%$ and $14.32\\%$ , respectively, while falling by only $7.23\\%$ and $9.13\\%$ for the MViT-V2 model. Earlier convergence to near BG-only performance and a more rapid rate of decline, show that 3D-R50 is more sensitive to increases in background durations than MViT-V2. Despite starting from lower performance, the 2D R50 model exhibits a more gradual, near-linear decline, indicating less sensitivity to changes in $\\lambda$ . While the behaviour of the 2D-R50 and 3D-R50 models remain relatively consistent on $D^{disjoint}$ , the behaviour of MViT-V2 changes significantly. Its uAP declines more steeply, falling from $70.05\\%$ to $48.29\\%$ (a decrease of $22.21\\%$ ) between $\\lambda = 0$ and $\\lambda = 32$ , where its performance converges. It is also more sensitive to smaller increases in background durations, falling from $70.50\\%$ to $64.08\\%$ (a decrease of $6.42\\%$ ) between $\\lambda = 0$ and $\\lambda = 1.0$ , where a decrease of only $0.38\\%$ is experienced on $D^{overlap}$ over the same interval. At higher values of $\\lambda$ (i.e., $>40$ ) the performance drops below that of the 3D-R50 (although only once", + "bbox": [ + 89, + 477, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the background-only performance has been surpassed). Although uAP falls more steeply, mAP declines more slowly and performance above that of the BG-only model is maintained even at the highest values of $\\lambda$ . This suggests that, although the MViT-V2 achieves higher baseline performance on the disjoint dataset, it is more sensitive to increases in background frames when evaluating uAP; a scenario where the model's advantage is drowned out by background information.", + "bbox": [ + 511, + 90, + 905, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Background Bias Mitigation", + "text_level": 1, + "bbox": [ + 511, + 239, + 767, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Augmenting Performance via Explicit Background Neutralisation. In this final experimental section, we justify the need for paired, real-world background samples, as uniquely provided by our dataset. We show that these sample pairings are able to significantly improve model performance. To show this, we analyse the impact of background subtraction operations in both input and embedding spaces across each of the model architectures and dataset configurations. Two background-subtracted views of our dataset were created for input space experiments. This was done by performing pixel-wise subtraction of $\\mathcal{B}$ and $\\tilde{\\mathcal{B}}$ from $\\mathcal{F}$ . To perform algebraic background subtraction in latent space, we extracted the feature vectors of the foreground video pairs. As described in Sec. 5.2, we achieve this by processing the video tensors up to the last convolutional layer and averaging the tokens from the last hidden state in the ResNet and MViT-V2 models, respectively. The background-subtracted embedding is then computed as", + "bbox": [ + 511, + 263, + 906, + 536 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/005b94d4eed3c5553ed3cdf803597d0099fc8bb2cf4e216c16501ad2ca7cba9a.jpg", + "image_caption": [ + "foreground x fx 3x 224x 224" + ], + "image_footnote": [], + "bbox": [ + 516, + 560, + 712, + 650 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da1cfbb492290cd4d4c63f0fe109f7ae51c23439a47baba6bccd1a8b419fb509.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 561, + 900, + 650 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa6d66bad7d7a778dbfa4ee378fbe25848f1cbc285034806f09f32d57a393ac6.jpg", + "image_caption": [ + "background (b x f x 3 x 224 x 224)", + "Figure 6. Latent Space Background Compensation. The proposed operation uses $\\alpha$ to perform a weighted subtraction of background features (computed via the dataset pairing of a real-world, background video) from foreground features in latent space. The dotted line between the encoders indicates shared weights (i.e., they are the same model)." + ], + "image_footnote": [], + "bbox": [ + 517, + 684, + 903, + 801 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5439", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$z^{\\mathcal{F} - \\mathcal{B}} = z^{\\mathcal{F}} - (1 - \\alpha)\\cdot z^{\\mathcal{B}}$ where $\\alpha$ is a coefficient used to control the relative magnitude of the algebraic subtraction operation as schematically shown in Fig. 6. In practice, the modulation parameter $\\alpha \\in [0,1]$ was ablated to increase linearly or exponentially over the course of training and results are reported for both. In all cases $\\alpha$ was initialized at zero and increased to one by the end of training.", + "bbox": [ + 89, + 89, + 480, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/afbece896aeee9e550c686a5e5901ae6f66e60fb0ddf0f71f1e883ba8f48eb05.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Quantifying Performance Improvements of Background Subtraction. Comparison of the 2D R50, 3D R50, and MViT-V2 models on both dataset configurations. Performance is reported when training on $\\mathcal{F} - \\mathcal{B}$ or $\\mathcal{F} - \\tilde{\\mathcal{B}}$ videos in input (light green) and embedding space (light blue). First and second highest scores are indicated for each model, while the highest overall scores are marked by an asterisks $(\\ast)$ . Note the baseline is trained on $\\mathcal{F}$ , as reported in Tab. 1. $\\mathcal{F} + \\mathcal{B}$ provides an alternative baseline but is reliant on background videos at test time. Results confirm the superiority of real-world background videos over synthetic data, and of latent space neutralisation over input space background subtraction." + ], + "table_body": "
ModelSubtraction\\(D^{overlap}\\)\\(D^{disjoint}\\)
uAPmAPuAPmAP
2D R50Baseline64.1541.2844.6718.08
\\(\\mathcal{F}+\\mathcal{B}\\)67.7444.9945.2818.84
\\(\\mathcal{F}-\\mathcal{B}\\)65.9642.1254.5326.65
\\(\\mathcal{F}-\\mathcal{B}\\)72.2147.1253.7325.84
None56.2138.8427.2417.94
Linear53.7839.0129.7517.17
Exponential54.0537.1125.8017.73
3D R50Baseline72.7949.3257.7430.89
\\(\\mathcal{F}+\\mathcal{B}\\)76.4153.5765.6639.17
\\(\\mathcal{F}-\\mathcal{B}\\)66.5142.5653.5827.98
\\(\\mathcal{F}-\\mathcal{B}\\)71.9346.3653.9225.05
None72.3051.3963.1434.85
Linear77.3152.8266.6936.31
Exponential73.5251.2464.5836.36
Baseline75.4451.5470.5035.29
\\(\\mathcal{F}+\\mathcal{B}\\)72.8253.4968.0038.09
MViT-V2\\(\\mathcal{F}-\\mathcal{B}\\)66.8442.2757.2128.80
\\(\\mathcal{F}-\\mathcal{B}\\)74.9049.4061.2231.51
None74.2344.1168.5635.14
Linear79.32*53.83*71.64*39.04
Exponential77.7050.9371.2839.25*
", + "bbox": [ + 94, + 207, + 480, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results and Performance Improvements. Tab. 2 shows that for the basic 2D-R50 model, background subtraction in input space is highly effective. When using $\\mathcal{B}$ , we observe substantial performance improvements over the baseline, with percentage increases of $+8.06\\%$ and $5.84\\%$ in uAP and mAP on $D^{overlap}$ , and $+9.06\\%$ and $+7.76\\%$ in uAP and mAP on $D^{disjoint}$ . Using $\\tilde{\\mathcal{B}}$ also yields improvements, though these gains are notably smaller. Background subtraction in embedding space does not provide any benefit to the 2D-R50. Although input space background subtraction is effective for the 2D-R50, it does not improve performance for the 3D-R50 and MViT-V2 models. Neither original nor synthetic backgrounds provide an advantage over", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "their respective baselines, indicating that input space subtraction may not suit spatio-temporal architectures.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conversely, background subtraction in embedding space provides significant performance improvements for the 3D-R50 and MViT-V2 models, particularly on $D^{disjoint}$ . For the 3D-R50, embedding space subtraction with linear scheduling results in a $+8.95\\%$ and $+5.42\\%$ increase in uAP and mAP, respectively. Exponential scheduling also yields improvements over the baseline. The improvements are lower than linear scheduling with respect to uAP $(+6.84\\%)$ but marginally better for mAP $(+5.47\\%)$ . MViT-V2 also benefits from embedding space subtraction on $D^{disjoint}$ : linear scheduling improves uAP and mAP by $+1.14\\%$ and $+3.75\\%$ , respectively, whereas exponential scheduling provides a $+0.78\\%$ increase in uAP and a $+3.96\\%$ increase in mAP. These findings suggest that embedding space subtraction is effective for both 3D-R50 and MViT-V2 models on $D^{disjoint}$ , with linear and exponential scheduling favouring uAP and mAP performance, respectively. Embedding space subtraction also enhances performance on $D^{overlap}$ , though to a lesser extent than on $D^{disjoint}$ . For the 3D-R50 model, applying embedding space subtraction with linear scheduling yields an improvement of $4.52\\%$ and $3.50\\%$ in uAP and mAP, respectively, while exponential scheduling results in smaller gains of $0.73\\%$ and $1.92\\%$ . The MViT-V2 model follows a similar trend: linear scheduling increases uAP and mAP by $3.88\\%$ and $2.29\\%$ , respectively, whereas exponential scheduling improves uAP by $2.26\\%$ but slightly reduces mAP by $-0.61\\%$ . These results suggest that algebraic embedding space subtraction is beneficial on $D^{overlap}$ , although the performance gains are relatively small compared to $D^{disjoint}$ , and that linear scheduling generally provides better results than exponential scheduling.", + "bbox": [ + 511, + 123, + 903, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 614, + 633, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced PanAf-FGBG, the first wildlife dataset designed to assess backgrounds and out-of-distribution generalization in behavior recognition. Our results demonstrate that real-world background information is a strong predictor of animal behavior and can significantly enhance recognition performance. While there are some limitations — such as its restriction to statically installed camera trap videos, the additional data and compute required for latent space background subtraction, and the potential for unnatural artifacts in background concatenation — the dataset remains a valuable resource. PanAf-FGBG enables direct evaluation under out-of-distribution conditions, and its rich metadata supports additional tasks and experimental setups. We hope it will help with understanding and improving model performance in this challenging domain and ultimately benefit endangered and charismatic species like the one featured in this paper.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5440", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank the Pan African Programme: 'The Cultured Chimpanzee' team and its collaborators for allowing the use of their data for this paper. We thank Amelie Pettrich, Antonio Buzharevski, Eva Martinez Garcia, Ivana Kirchmair, Sebastian Schutte, Linda Gerlach and Fabina Haas. We also thank management and support staff across all sites; specifically Yasmin Moebius, Geoffrey Muhanguzi, Martha Robbins, Henk Eshuis, Sergio Marrocoli and John Hart. Thanks to the team at https://www.chimpandsee.org particularly Briana Harder, Anja Landsmann, Laura K. Lynn, Zuzana Macháčková, Heidi Pfund, Kristeena Sigler and Jane Wideness. The work that allowed for the collection of the dataset was funded by the Max Planck Society, Max Planck Society Innovation Fund, and Heinz L. Krekeler. In this respect we would like to thank: Ministre des Eaux et Forets, Ministère de l'Enseignement supérieur et de la Recherche scientifique in Côte d'Ivoire; Institut Congolais pour la Conservation de la Nature, Ministère de la Recherche Scientifique in Democratic Republic of Congo; Forestry Development Authority in Liberia; Direction Des Eaux Et Forets, Chasses Et Conservation Des Sols in Senegal; Makerere University Biological Field Station, Uganda National Council for Science and Technology, Uganda Wildlife Authority, National Forestry Authority in Uganda; National Institute for Forestry Development and Protected Area Management, Ministry of Agriculture and Forests, Ministry of Fisheries and Environment in Equatorial Guinea. This work was supported by the UKRI CDT in Interactive AI (grant EP/S022937/1). This work was in part supported by the US National Science Foundation Awards No. 2118240 \"HDR Institute: Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning\" and Award No. 2330423 and Natural Sciences and Engineering Research Council of Canada under Award No. 585136 for the \"AI and Biodiversity Change (ABC) Global Center\".", + "bbox": [ + 89, + 117, + 485, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 667, + 187, + 684 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pan African Programme: The Cultured Chimpanzee, 1999. Online; accessed 29 January 2014. 3", + "[2] Analysing animal behaviour in wildlife videos using face detection and tracking. IEEE Proceedings-Vision, Image and Signal Processing, 153(3):305-312, 2006. 2", + "[3] Mimi Arandjelovic, Colleen R Stephens, Maureen S McCarthy, Paula Dieguez, Emmie K Kalan, Nuria Maldonado, Christophe Boesch, and Hjalmar S Kuehl. Chimp&see: An online citizen science platform for large-scale, remote video camera trap annotation of chimpanzee behaviour, demography and individual identification. PeerJ Preprints, 2016. 4", + "[4] Max Bain, Arsha Nagrani, Daniel Schofield, Sophie Berdugo, Joana Bessa, Jake Owen, Kimberly J Hockings, Tetsuro Matsuzawa, Misato Hayashi, Dora Biro, et al. Au" + ], + "bbox": [ + 99, + 694, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tomated audiovisual behavior recognition in wild primates. Science advances, 7(46):eabi4883, 2021. 2", + "[5] Sara Beery, Arushi Agarwal, Elijah Cole, and Vighnesh Birodkar. The iwildcam 2021 competition dataset. arXiv preprint arXiv:2105.03494, 2021. 3", + "[6] Otto Brookes, Majid Mirmehdi, Hjalmar S. Kuhl, and Tilo Burghardt. Triple-stream deep metric learning of great ape behavioural actions. In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 294-302, 2023. 2", + "[7] Otto Brookes, Majid Mirmehdi, Hjalmar Kuhl, and Tilo Burghardt. Chimpvlm: Ethogram-enhanced chimpanzee behaviour recognition. 2024.", + "[8] Otto Brookes, Majid Mirmethdi, Colleen Stephens, Samuel Angedakin, Katherine Corogenes, Dervla Dowd, Paula Dieguez, Thurston C Hicks, Sorrel Jones, Kevin Lee, et al. Panaf20k: a large video dataset for wild ape detection and behaviour recognition. International Journal of Computer Vision, pages 1-17, 2024. 2, 3", + "[9] Jackie Chappell and Susannah KS Thorpe. The role of great ape behavioral ecology in one health: Implications for captive welfare and re-habilitation success. American journal of primatology, 84(4-5):e23328, 2022. 2", + "[10] Jun Chen, Ming Hu, Darren J Coker, Michael L Berumen, Blair Costelloe, Sara Beery, Anna Rohrbach, and Mohamed Elhoseiny. Mammalnet: A large-scale video benchmark for mammal recognition and behavior understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13052-13061, 2023. 2, 3", + "[11] Jinwoo Choi, Chen Gao, Joseph CE Messou, and Jia-Bin Huang. Why can't i dance in the mall? learning to mitigate scene bias in action recognition. Advances in Neural Information Processing Systems, 32, 2019. 2, 3, 4", + "[12] Jihoon Chung, Yu Wu, and Olga Russakovsky. Enabling detailed action recognition evaluation through video dataset augmentation. Advances in Neural Information Processing Systems, 35:39020-39033, 2022. 2, 3, 4, 5", + "[13] Fay E Clark. Great ape cognition and captive care: Can cognitive challenges enhance well-being? Applied Animal Behaviour Science, 135(1-2):1-12, 2011. 2", + "[14] Fagner Cunha, Eulanda M dos Santos, Raimundo Barreto, and Juan G Colonna. Filtering empty camera trap images in embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2438-2446, 2021. 2", + "[15] Alexander D'Amour, Katherine Heller, Dan Moldovan, Ben Adlam, Babak Alipanahi, Alex Beutel, Christina Chen, Jonathan Deaton, Jacob Eisenstein, Matthew D Hoffman, et al. Underspecification presents challenges for credibility in modern machine learning. Journal of Machine Learning Research, 23(226):1-61, 2022. 2", + "[16] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 3" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5441", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Isla Duporge, Maksim Kholiavchenko, Roi Harel, Dan Rubenstein, Meg Crofoot, Tanya Berger-Wolf, Stephen Lee, Scott Wolf, Julie Barreau, Jenna Kline, et al. Baboonland dataset: Tracking primates in the wild and automating behaviour recognition from drone videos. arXiv preprint arXiv:2405.17698, 2024. 2", + "[18] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2", + "[19] Rohit Girdhar and Deva Ramanan. Cater: A diagnostic dataset for compositional actions and temporal reasoning. arXiv preprint arXiv:1910.04744, 2019. 3", + "[20] Abhinav Gupta and Larry S Davis. Objects in action: An approach for combining action understanding and object perception. pages 1-8. IEEE, 2007. 3", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[22] Yun He, Soma Shirakabe, Yutaka Satoh, and Hirokatsu Kataoka. Human action recognition without human. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 11-17. Springer, 2016. 2, 3", + "[23] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In Proceedings of the European conference on computer vision (ECCV), pages 771-787, 2018. 2, 4", + "[24] Filip Ilic, Thomas Pock, and Richard P Wildes. Is appearance free action recognition possible? In European Conference on Computer Vision, pages 156-173. Springer, 2022. 2, 3", + "[25] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 4", + "[26] Maksim Kholiavchenko, Jenna Kline, Maksim Kukushkin, Otto Brookes, Sam Stevens, Isla Duporge, Alec Sheets, Reshma R Babu, Namrata Banerji, Elizabeth Campolongo, et al. Deep dive into kabr: a dataset for understanding ungulate behavior from in-situ drone video. Multimedia Tools and Applications, pages 1-20, 2024. 2", + "[27] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 2, 3, 4", + "[28] Hjalmar S Kuhl and Tilo Burghardt. Animal biometrics: quantifying and detecting phenotypic appearance. Trends in ecology & evolution, 28(7):432-441, 2013. 2", + "[29] Haoxin Li, Yuan Liu, Hanwang Zhang, and Boyang Li. Mitigating and evaluating static bias of action representations in the background and the foreground. In Proceedings of the" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 19911-19923, 2023. 2, 3", + "[30] Yi Li and Nuno Vasconcelos. Repair: Removing representation bias by dataset resampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9572-9581, 2019. 2, 3", + "[31] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 3", + "[32] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4804-4814, 2022. 4", + "[33] Dan Liu, Jin Hou, Shaoli Huang, Jing Liu, Yuxin He, Bochuan Zheng, Jifeng Ning, and Jingdong Zhang. Lote-animal: A long time-span dataset for endangered animal behavior understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20064-20075, 2023. 2, 3", + "[34] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 3", + "[35] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 4", + "[36] Xiaoxuan Ma, Stephan Kaufhold, Jiajun Su, Wentao Zhu, Jack Terwilliger, Andres Meza, Yixin Zhu, Federico Rossano, and Yizhou Wang. Chimpact: A longitudinal dataset for understanding chimpanzee behaviors. Advances in Neural Information Processing Systems, 36:27501-27531, 2023. 2, 3", + "[37] Marcin Marszalek, Ivan Laptev, and Cordelia Schmid. Actions in context. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 2929-2936. IEEE, 2009. 2, 3", + "[38] Zhongqi Miao, Ziwei Liu, Kaitlyn M Gaynor, Meredith S Palmer, Stella X Yu, and Wayne M Getz. Iterative human and automated identification of wildlife images. Nature Machine Intelligence, 3(10):885-895, 2021. 2", + "[39] Mazda Moayeri, Phillip Pope, Yogesh Balaji, and Soheil Feizi. A comprehensive study of image classification model sensitivity to foregrounds, backgrounds, and visual attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19087-19097, 2022. 2, 6", + "[40] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19023-19034, 2022. 2, 3", + "[41] Toby Perrett, Saptarshi Sinha, Tilo Burghardt, Majid Mirmethdi, and Dima Damen. Use your head: Improving long-tail video recognition. In Proceedings of the IEEE/CVF con" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5442", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on computer vision and pattern recognition, pages 2415-2425, 2023. 3", + "[42] Laura J Pollock, Justin Kitzes, Sara Beery, Kaitlyn M Gaynor, Marta A Jarzyna, Oisin Mac Aodha, Bernd Meyer, David Rolnick, Graham W Taylor, Devis Tuia, et al. Harnessing artificial intelligence to fill global shortfalls in biodiversity knowledge. Nature Reviews Biodiversity, pages 1-17, 2025. 2", + "[43] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2022. 2", + "[44] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4", + "[45] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 4", + "[46] Faizaan Sakib and Tilo Burghardt. Visual recognition of great ape behaviours in the wild. In Workshop on the Visual Observation and Analysis of Vertebrate and Insect Behaviour, 2020. 2", + "[47] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9661-9669, 2021. 2", + "[48] Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank Van Langevelde, Tilo Burghardt, et al. Perspectives in machine learning for wildlife conservation. Nature communications, 13(1):1-15, 2022. 2", + "[49] Tuan-Hung Vu, Catherine Olsson, Ivan Laptev, Aude Oliva, and Josef Sivic. Predicting actions from static scenes. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 421-436. Springer, 2014. 2, 3", + "[50] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11804-11813, 2021. 3", + "[51] Kai Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. arXiv preprint arXiv:2006.09994, 2020.3", + "[52] Xinyu Yang, Majid Mirmehdi, and Tilo Burghardt. Great ape detection in challenging jungle camera trap footage via attention-based spatial and temporal feature blending. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 2", + "[53] Xinyu Yang, Tilo Burghardt, and Majid Mirmehdi. Dynamic curriculum learning for great ape detection in the wild. In-" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ternational Journal of Computer Vision, 131(5):1163-1181, 2023. 2", + "[54] Manlin Zhang, Jinpeng Wang, and Andy J Ma. Suppressing static visual cues via normalizing flows for self-supervised video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3300-3308, 2022. 3", + "[55] Mingjun Zhao, Yakun Yu, Xiaoli Wang, Lei Yang, and Di Niu. Search-map-search: a frame selection paradigm for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10627-10636, 2023. 3", + "[56] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 3", + "[57] Yi Zhu, Xinyu Li, Chunhui Liu, Mohammadreza Zolfaghari, Yuanjun Xiong, Chongruo Wu, Zhi Zhang, Joseph Tighe, R Manmatha, and Mu Li. A comprehensive study of deep video action recognition. arXiv preprint arXiv:2012.06567, 2020. 2" + ], + "bbox": [ + 516, + 92, + 906, + 386 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5443", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_model.json b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..43d2dd280e28bf6393dfc6d6a1961bdd77a97e47 --- /dev/null +++ b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_model.json @@ -0,0 +1,1839 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.36, + 0.106, + 0.639, + 0.125 + ], + "angle": 0, + "content": "The PanAf-FGBG Dataset:" + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.129, + 0.89, + 0.15 + ], + "angle": 0, + "content": "Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.155, + 0.88, + 0.251 + ], + "angle": 0, + "content": "Otto Brookes\\(^{1,2\\dagger*}\\), Maksim Kukushkin\\(^{3,4*}\\), Majid Mirmehdi\\(^{1}\\), Colleen Stephens\\(^{5}\\), Paula Dieguez\\(^{5}\\), Thurston C. Hicks\\(^{6}\\), Sorrel Jones\\(^{5}\\), Kevin Lee\\(^{5}\\), Maureen S. McCarthy\\(^{5}\\), Amelia Meier\\(^{5}\\), Emmanuelle Normand\\(^{2}\\), Erin G. Wessling\\(^{7}\\), Roman M. Wittig\\(^{5,8}\\), Kevin Langergraber\\(^{9}\\), Klaus Zuberbühler\\(^{10}\\), Lukas Boesch\\(^{2}\\), Thomas Schmid\\(^{3,11}\\), Mimi Arandjelovic\\(^{5}\\), Hjalmar Kuhl\\(^{5,12}\\), Tilo Burghardt\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.254, + 0.828, + 0.325 + ], + "angle": 0, + "content": "1University of Bristol, 2Wild Chimpanzee Foundation, 3Martin Luther University Halle-Wittenberg, 4Leipzig University, 5Max Planck Institute for Evolutionary Anthropology, 6University of Warsaw, 7Harvard University, 8University of Lyon, 9Arizona State University, 10University of St Andrews, 11Lancaster University in Leipzig, 12Senckenberg Museum of Natural History" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.359, + 0.327, + 0.374 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.393, + 0.486, + 0.846 + ], + "angle": 0, + "content": "Computer vision analysis of camera trap video footage is essential for wildlife conservation, as captured behaviours offer some of the earliest indicators of changes in population health. Recently, several high-impact animal behaviour datasets and methods have been introduced to encourage their use; however, the role of behaviour-correlated background information and its significant effect on out-of-distribution generalisation remain unexplored. In response, we present the PanAf-FGBG dataset, featuring 21 hours of wild chimpanzee behaviours, recorded at 389 individual camera locations. Uniquely, it pairs every video with a chimpanzee (referred to as a foreground video) with a corresponding background video (with no chimpanzee) from the same camera location. We present two views of the dataset: one with overlapping camera locations and one with disjoint locations. This setup enables, for the first time, direct evaluation of in-distribution and out-of-distribution conditions, and for the impact of backgrounds on behaviour recognition models to be quantified. All clips have rich behavioural annotations and metadata, including unique camera IDs. Additionally, we establish several baselines and present a highly effective latent-space normalisation technique that boosts out-of-distribution performance by \\(+5.42\\%\\) mAP for convolutional and \\(+3.75\\%\\) mAP for transformer-based models. Finally, we provide an in-depth analysis on the role of backgrounds in out-of-distribution behaviour recognition, including the so far unexplored impact of background durations. (i.e., the count of background frames within foreground videos). The dataset is available at https://obrookes.github.io/panaf-fbgb.github.io/." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.359, + 0.907, + 0.723 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.734, + 0.909, + 0.86 + ], + "angle": 0, + "content": "Figure 1. Conceptual Overview. The PanAf-FGBG dataset comprises \\(>20\\) hours of paired and richly annotated foreground-background camera trap videos of wild chimpanzees. The dataset unlocks systematic analyses of the impact of background information on wildlife behaviour recognition. We provide baselines, quantify the background impact on performance, and demonstrate that utilising background information in latent space (see blue data point) can significantly improve recognition performance in this challenging domain." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.402, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author, *Equal technical contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5433" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.107 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.482, + 0.236 + ], + "angle": 0, + "content": "Deep learning models have shown impressive results on action recognition tasks [57], particularly when both training and test data share the same distribution. However, their performance declines significantly on out-of-distribution (OOD) data [15, 27]. In practice, distribution shifts are common in nearly all settings [43], but are particularly unforgiving when computer vision is applied to uncontrolled, complex or natural environments [27]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.483, + 0.432 + ], + "angle": 0, + "content": "In the human action recognition domain, poor OOD performance is well-known to originate from shortcut learning [18] of static cues or backgrounds [24, 29, 30]. This leads models to rely heavily on background information instead of the behaviour of interest solely [2], and several studies have shown that the majority of action recognition performance in this domain can be attributed to background learning [12, 22, 49]. Despite the known impact of backgrounds on action recognition and a critical need to advance model performance in wildlife applications [28, 42, 48], reliance on static cues for animal behaviour recognition has received limited attention. This is largely due to a lack of datasets explicitly designed to study its effects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.644 + ], + "angle": 0, + "content": "In response, we introduce the PanAf-FGBG dataset, which comprises real-world foreground-background video pairs (see Fig. 1) extracted from six African countries, 14 national parks, and over 350 individual camera locations across many habitats and location types (see Fig. 3). Here, we define foreground videos as those containing a chimpanzee, and background videos as those that do not. This allows us to address a key limitation in existing studies, where the background's role is typically assessed by creating background-only versions of datasets through synthetic modifications (e.g., masking or in-painting) that remove the actor [11, 12, 23]. While such synthetic inputs often serve as useful proxies, the actual effect of synthesis on training is largely unknown [39, 47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In contrast, our dataset introduces paired camera locations and enables a truly realistic disentangling of background effects under in- and out-of-distribution conditions: disjoint experiments with mutually exclusive camera locations for training and test splits, and overlapping experiments with shared locations. This focus is particularly important in the wildlife domain, where the transferability and generalisation of camera trap analysis to new locations is critical for effective operation. While studies have shown the over-reliance of action recognition models on backgrounds [22, 37, 49], we find that the effect of background duration (i.e., the count of background frames within foreground videos with no actor) is largely unexplored. In camera trap video footage like in many other CCTV or triggered camera settings, actors may only appear momentarily, leaving behind footage of only the background [14, 38] within foreground videos. Since this effect alters learning given" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "video-level behaviour annotations, we quantify its effect on several action recognition architectures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Our contributions are as follows: (i) we release PanAf-FGBG with its rich annotations and meta-data, providing a large wildlife dataset suitable for explicitly investigating the impact of backgrounds on behaviour recognition; (ii) we show models trained on background-only videos achieve strong animal behaviour recognition performance (\\(\\sim 65\\%\\) of baseline performance), corroborating findings in the human domain; (iii) we conduct the first evaluation of background duration in videos, revealing its impact on action recognition models and highlighting key differences between convolutional and attention-based networks; and (iv) we introduce a simple yet powerful latent-space background neutralisation that yields notable performance gains, particularly in OOD scenarios for convolutional \\((+5.42\\%)\\) mAP) and transformer-based approaches \\((+3.75\\%)\\) mAP)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.655, + 0.377 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.704 + ], + "angle": 0, + "content": "Animal Behaviour Datasets. In recent years, several high-impact animal behaviour datasets have been introduced to the CV community, often accompanied by new methods [4, 6-8, 10, 26, 33, 40, 46]. Although they all centre on animal behaviour understanding, they each have a different focus. The large-scale datasets Animal Kingdom [40] and MammalNet [10] concentrate on behavioural understanding across species, each including videos of \\(\\sim 800\\) different species. They provide atomic action and high-level behaviour annotations, respectively, and support, in addition to action recognition, the related tasks of action localisation [10, 40] and pose estimation [40]. The smaller ChimpAct dataset [36] provides longitudinal data that follows a zoo-housed group of 20 chimpanzees with a focus on a single male juvenile. It also contains key-point pose information, and fine-grained spatio-temporal behaviour labels. While valuable, these datasets lack in-situ footage from ecological sensors like camera traps and drones, meaning animals are not observed in their natural environment. This may result in unrepresentative behaviour distributions and limit their suitability for studying wild behaviour [9, 13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In contrast, KABR [26] and BaboonLand [17] comprise drone footage of free-living Kenyan wildlife, with fine-grained spatio-temporal behaviour annotations, focussing on individual and group-level behaviour, respectively. The LoTE dataset [33] provides camera trap video footage of 11 endangered species from a 12-year longitudinal study; it is accompanied by environmental information (i.e., weather conditions and habitat) and highlights the importance of temporal and spatial continuity for conservation efforts. The dataset most comparable to our own is the PanAf20k dataset [8] and its related PanAf500-based works [46, 52, 53]. PanAf20k comprises camera trap videos of great apes accompanied by multi-label behaviour annotations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5434" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.349 + ], + "angle": 0, + "content": "Our dataset orthogonally complements this growing body of work dedicated to understanding animal behaviour by providing the first dataset specifically designed to evaluate backgrounds and out-of-distribution generalisation for behaviour recognition. It is unique in providing foreground-background pairs, and its configurations for in- and out-of-distribution evaluation are defined identically to the well-known iWildCam [5] and Wilds dataset [27] for species classification. Similar to LotE [33], we provide a detailed description of the background, however we also include a unique camera ID (i.e., geospatial location) to allow footage taken from the same camera to be linked. Similar to [8], our data is gathered from 14 national parks, spanning 6 different countries, although unlike ours they do not include metadata to allow videos to be linked to specific countries, research sites, locations, or habitats. Similarly, it is more diverse than [33] with its footage from a single national park." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.563 + ], + "angle": 0, + "content": "Background Impact. The effect of backgrounds or static bias is well-studied within computer vision [29, 54], where it has shown to be necessary for models to perform well [51]. Action recognition models are particularly vulnerable to shortcut learning. These models often rely on correlated, but distinct information beyond the intended action, such as object [20, 56], scene [37], and actor biases [31]. For example, several studies have reported that strong performance on action recognition tasks can be achieved without the actor [22, 37, 49] and that approximately \\(70\\%\\) of model performance is attributable to static cues (i.e., scene or background) [12]. This results in biased representations and ultimately impairs out-of-distribution generalization performance [30]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.748 + ], + "angle": 0, + "content": "Background Bias Datasets. The effect of the background has often been modelled using synthetically generated data that removes either the foreground or the background. [11, 12, 19, 29]. For instance, Ilic et al. [24] present a synthetic version of the UCF101 dataset, animating spatial noise using image motion extracted by an optical flow estimator, Li et al. [29] replace the backgrounds of IID HMDB51 with sinusoidal stripe images, and Chung et al [12] adopt an in-painting model to remove actors from the video. Datasets that include fine-grained spatio-temporal action annotations also remove static cues implicitly by isolating the region of the video where the action is occurring." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Mitigating Background Bias. Prominent attempts to mitigate this bias include dataset resampling methods, which down-weight biased samples [30, 31], adversarial training regimes that rely on scene labels and human masks [11], and augmentation techniques that mix or swap the backgrounds of different videos during training [16, 29, 50]. There are also several implicit approaches, such as frame selection, that aim to utilise only frames where activity is occurring [55]. Our dataset, introduced next, instead introduces foreground-background pairing in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "formation with close proximity in time of capture, which describes the environment truly independently and in relation to the presence of animals showcasing behaviour." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.15, + 0.603, + 0.166 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Dataset Overview and Statistics. PanAf-FGBG comprises footage gathered under ethical oversight as part of the PanAf Programme: The Cultured Chimpanzee [1]. It contains 21 hours of camera trap footage of individual chimpanzees in tropical Africa. Its footage is collected from 389 individual camera locations across 14 national parks in 6 African countries. In total, we provide 5,070 video pairs, each 15 seconds in duration. All videos are accompanied by a multi-label behaviour annotation and scene metadata, plus the unique feature of foreground-background video pairing (see Fig. 1). Class and habitat distributions of the data are given in Fig. 2 and Fig. 3, respectively. Our dataset follows a long-tailed distribution [34, 41] common in animal behaviour datasets [8, 10, 33, 36, 40]. The three of the most commonly occurring classes are observed in \\(>60\\%\\) of videos, while the rarest are observed in fewer than \\(< 3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.909, + 0.593 + ], + "angle": 0, + "content": "Data Collection. The data collection strategy was designed specifically to maximise the chance of filming the terrestrial activity of apes and we provide only a brief description here (for full details we refer the reader to [1]). Data was collected from 14 different research sites, where an average of 29 movement-triggered Bushnell cameras were installed per site. Grids comprising 20 to \\(961 \\times 1\\) km cells were established and one camera was installed per grid cell. Footage was recorded at 30 frames-per-second and a resolution of \\(1280 \\times 720\\). The cameras were visited every 1-3 months for maintenance and to download the recorded footage throughout the study periods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.655 + ], + "angle": 0, + "content": "Ethical Considerations. To protect the location of the apes from threats (e.g., poaching) we use monikers for the names of the research sites and assign a unique camera ID to each video, hiding the geospatial location of the cameras." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.677, + 0.891, + 0.858 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.869, + 0.908, + 0.898 + ], + "angle": 0, + "content": "Figure 2. Distribution of Behaviour. Proportion of behaviours in the dataset where smaller segments are colour coordinated." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5435" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.098, + 0.476, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.409, + 0.486, + 0.466 + ], + "angle": 0, + "content": "Figure 3. Distribution of Habitat and Location Type. The pie chart gives the proportion of videos taken at each of the three main habitats: forest, marshes and savannah. The histogram shows the total number of videos extracted from each location type." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Data Annotation. The behavioural annotations were provided by users on the community science platform Chimp&See [3]. Annotators were presented with a choice of behaviour classification categories, chosen specifically by experts for their ecological importance. Behaviours not listed in the classification categories were also permitted. These annotations were then extracted and expertly grouped into 14 co-occurring classes, which form the multi-label behavioural annotations presented here. The annotations follow a multi-hot binary format that indicate the presence of one or more behaviours. It should also be noted that behaviours are not assigned to individual apes or temporally localised within each video. To ensure annotation quality and consistency a video was only deemed to be analysed when either three volunteers marked the video as blank, or unanimous agreement between seven volunteers was observed, or 15 volunteers annotated the video. Habitat type was noted during camera setup." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Foreground-Background Pair Selection. Pairs are formed by selecting the temporally closest available foreground and background videos to ensure similarity. Complete standardisation statistics can be found in Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Overlapping & Disjoint Configurations. We present both a disjoint view of dataset, where camera locations in the training and test splits are mutually exclusive, denoted as \\(\\mathcal{D}^{disjoint} = \\{d_i \\mid d_i \\cap d_j = \\emptyset, \\forall i \\neq j\\}\\) following [27]. A second, overlapping view is also provided, where camera" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "locations are shared between training and test splits, represented as \\(\\mathcal{D}^{overlap} = \\{d_i \\mid d_i \\cap d_j \\neq \\emptyset\\) for some \\(i \\neq j\\}\\). Note that \\(d_i\\) refers to sets of camera locations associated with videos in each split. Fig. 4 provides an overview of both configurations. We ensure that the class distribution in each configuration remains approximately the same for full comparability. Unique camera identifiers are available for each video to allow researchers interested in this problem to create additional configurations, should they wish to." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.653, + 0.267 + ], + "angle": 0, + "content": "4. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.909, + 0.476 + ], + "angle": 0, + "content": "We train multi-label classification models on both views of the dataset: \\(\\mathcal{D}^{\\mathrm{overlap}}\\) and \\(\\mathcal{D}^{\\mathrm{disjoint}}\\) using various combinations of foreground \\((\\mathcal{F})\\), background \\((\\mathcal{B})\\) and synthetic background \\((\\tilde{\\mathcal{B}})\\) videos. The latter were created by generating a segmentation mask using SAM2 [44] followed by mean pixel value filling, similar to [11, 12, 23]. Both views comprise foreground-background video pairs \\(\\{(v_{i}^{\\mathcal{F}}, v_{i}^{\\mathcal{B}})\\}_{i=1}^{N}\\), where \\(v_{i}^{\\mathcal{F}} \\in \\mathcal{F}\\) and \\(v_{i}^{\\mathcal{B}} \\in \\mathcal{B}\\), with \\(N\\) representing the total number of video pairs. Each foreground video \\(v_{i}^{\\mathcal{F}}\\) is associated with a video-level label vector \\(y_{i} = (y_{i1}, y_{i2}, \\ldots, y_{ic})\\), where each \\(y_{ij} \\in \\{0,1\\}\\) indicates the presence (1) or absence (0) of the \\(j\\)-th class in video \\(i\\), and \\(C\\) is the total number of classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.646, + 0.518 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.528, + 0.597, + 0.545 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.903 + ], + "angle": 0, + "content": "All experiments were performed using the ResNet-50 [21] and MViT-V2 [32] architectures, which were initialised with feature extractors pre-trained on Kinetics-400 [25]. For ResNet-50, both 2D and 3D convolution versions are tested. After initial experiments to ensure optimal optimiser-network pairings for the task, ResNet and MViT-V2 models were fine-tuned for 300 epochs using the SGD [45] and AdamW [35] optimisers, respectively. Both models followed a linear warm-up schedule with subsequent cosine annealing. For ResNet models, the learning rate was increased from \\(1\\times 10^{-5}\\) to \\(1\\times 10^{-4}\\) over 20 epochs, while it was increased from \\(1\\times 10^{-6}\\) to \\(1\\times 10^{-4}\\) over 30 epochs for MViT-V2 models. In both cases, momentum of 0.9 was utilised. All computation was performed on \\(4\\times\\) NVIDIA H200 GPUs using a distributed batch size of 336. During training, video tensors with a spatio-temporal resolution of \\(16\\times 256\\times 256\\) were employed. Frames were sampled uniformly at a rate of \\(1/24\\), avoiding dense sampling strategies which are shown to exacerbate shortcut learning in action recognition models [12]. For MViT-V2 models, a patch size of \\(16\\times 16\\) pixels was utilised. Evaluation was performed using micro and mean Average Precision and are referred to as uAP and mAP, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5436" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.905, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.908, + 0.466 + ], + "angle": 0, + "content": "Figure 4. Overlapping & Disjoint Dataset Configurations. Six foreground-background video pairs are shown to visualise overlapping (left) and disjoint (right) dataset configurations. On the left, train and test videos taken from the same camera are selected, emphasizing that the locations are not mutually exclusive, with backgrounds containing several cues that are highly correlated with the displayed behaviours (i.e., algae fishing, a form of tool use). Train and test videos on the right are extracted from mutually exclusive or disjoint camera locations. They are extracted from forest and savannah, respectively, and highlight the challenge of generalising to new locations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.492, + 0.292, + 0.508 + ], + "angle": 0, + "content": "5.2. Background Reliance" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.519, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Experimental Overview. First, we examined the reliance of animal behaviour recognition on the background by training model architectures on the foreground \\(\\mathcal{F}\\), background \\(\\mathcal{B}\\), and synthetic background videos \\(\\tilde{\\mathcal{B}}\\) only. We refer to models trained on \\(\\mathcal{F}\\) and \\(\\mathcal{B}\\) (or \\(\\tilde{\\mathcal{B}}\\)) as FG-only and BG-only models, respectively, as shown earlier in Fig. 1. To separate background information from class-distribution intrinsic information, a dummy classifier with no access to visual information was also evaluated. It would effectively return predictions based on the class distribution information alone. Since our dataset consists of unique foreground-background pairs, it was possible to train BG-only models by retaining the original labels associated with the foreground video of the pairing. In all cases, models were tested on \\(\\mathcal{F}\\) for comparability. Note that this makes the task more difficult for BG-only models, since animals are only observed at test time and not during training. All experiments are conducted on both \\(D^{overlap}\\) and \\(D^{disjoint}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Dual-Stream Fusion Model \\((\\mathcal{F} + \\mathcal{B})\\). To clarify whether BG-only models and FG-only models, each trained separately, can be fused late to orthogonally contribute towards performance we also implemented a dual-stream model. To do this, the independently optimised weights of the FG-only and BG-only models trained for 300 epochs were loaded. During training the weights were frozen and video pairs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.584 + ], + "angle": 0, + "content": "were processed to extract late features \\( z^{F} \\) and \\( z^{B} \\). For ResNet models, features were extracted from the last convolutional layer and in MViT-V2 models an average of the tokens from the last hidden state were taken. Late features were then concatenated \\( \\tilde{z} = [z^{F}, z^{B}] \\) and final class predictions were produced by a trainable multi-layer perceptron." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Result 1 - Backgrounds are a Predictor of Animal Behaviour. Table 1 shows that BG-only models achieve strong performance when compared to their respective FG-only counterparts. To illustrate this, the ratio of background-toForeground performance \\(\\frac{AP^B}{AP^F}\\) published in [12], showing how close a model can get to FG performance with just BG cues, is reported for each architecture in rows 6, 11, and 16. Background-to foreground performance ratios are consistently high across all models, never dropping below 0.70 and 0.65 for uAP and mAP, respectively, on either dataset. This indicates that much of the behaviour recognition performance is indeed achievable by just utilisation of background information - the environment is a strong predictor of behaviour, even out-of-distribution. Thus, features of the environment are already behaviour predictors (trees for climbing, paths for travel, termite mound for feeding etc.). This result is further compounded by the fact that BG-only models outperform the non-visual classifier (row 1) by a significant margin, particularly on \\(D^{overlap}\\). On \\(D^{disjoint}\\) this observation still holds, although by somewhat" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5437" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.089, + 0.473, + 0.332 + ], + "angle": 0, + "content": "
ModelData\\(D^{overlap}\\)\\(D^{disjoint}\\)
uAPmAPuAPmAP
1Dummy-19.1514.7319.6114.87
2\\(\\bar{\\mathcal{F}}\\)64.1541.2844.6718.08
32D R50\\(\\mathcal{B}\\)57.1330.0345.8018.27
4\\(\\tilde{\\mathcal{B}}\\)57.8730.8248.2820.39
5\\(\\mathcal{F}+\\mathcal{B}\\)67.7444.9945.2818.84
60.890.741.020.98
7\\(\\bar{\\mathcal{F}}\\)72.7949.3257.7430.89
83D R50\\(\\mathcal{B}\\)58.9231.7248.4021.38
9\\(\\tilde{\\mathcal{B}}\\)59.6032.8649.2921.64
10\\(\\mathcal{F}+\\mathcal{B}\\)76.4153.5765.6639.17
110.800.660.830.69
12\\(\\bar{\\mathcal{F}}\\)75.4451.5470.5035.29
13MViT-V2\\(\\mathcal{B}\\)61.9032.7450.0324.00
14\\(\\tilde{\\mathcal{B}}\\)62.1533.5852.1324.77
15\\(\\mathcal{F}+\\mathcal{B}\\)72.8253.4968.0038.09
160.820.650.700.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.342, + 0.487, + 0.496 + ], + "angle": 0, + "content": "Table 1. Quantifying Background Reliance. Performance comparison of models trained on \\(\\mathcal{F}\\), \\(\\mathcal{B}\\), and \\(\\tilde{\\mathcal{B}}\\) videos. Dual-stream models are indicated by \\(\\mathcal{F} + \\mathcal{B}\\). The absolute highest performance is indicated in bold whereas best performance using only a single-stream is underlined. The performance based on class distribution only is given by a dummy classifier (see row 1). The background-toForeground performance ratio \\((\\frac{AP^B}{AP^F})\\) is also provided (see rows 6, 11, and 16). Performance is reported on \\(\\mathcal{D}_{\\text{overlap}}\\) and \\(\\mathcal{D}_{\\text{disjoint}}\\) where Fig. 1 visualises an mAP scatter plot between the two for 3D R50." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.613 + ], + "angle": 0, + "content": "smaller margins. Note that for the weakest model (2D R50) background mAP is even on-par or above the foreground baseline (see rows 2-3 in the final column). Together, these results illustrate powerfully, that across architectures background information alone is employable as a strong animal behaviour predictor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Result 2 - Multi-scale Vision Transformers rely less on backgrounds. On the overlapping dataset, the 3D-R50 and MViT-V2 models (rows 7-16) display significantly better performance when compared with the 2D-R50 (rows 2-6), and perform similarly to each other with respect to uAP and mAP. However, the MViT-V2 model widely outperforms the 3D R50 on the disjoint dataset, suggesting that it does not rely as heavily on background features for performance. In support of this, while all models experience a large decrease in performance when moving from the overlapping to the disjoint dataset (generalisation gap), the smallest impact is on the MViT-V2. These differences between architectures are highlighted most clearly by observing uAP on the disjoint dataset. Here, the background only ratio for the MViT-V2 model is \\(\\sim 30\\%\\) and \\(\\sim 13\\%\\) lower than the 2D-R50 and 3D-R50, respectively. MViT-V2 achieves the best overall performance and exhibits the lowest background-toForeground performance ratio on both dataset configurations, corroborating similar findings from" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "the still image domain that show transformers are more invariant to changes in the background [39]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.124, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Result 3 - Analysing Background Information Contributions. The dual 3D-R50 model achieves significant performance gains over its FG-only baseline, particularly on \\( \\mathcal{D}^{disjoint} \\), where increases of \\( 7.92\\% \\) and \\( 8.28\\% \\) in uAP and mAP, respectively, can be observed. Smaller, but still significant improvements are also observed on \\( \\mathcal{D}^{overlap} \\) (\\(+3.62\\%\\) uAP and \\( +4.25\\% \\) mAP). Interestingly, the performance of the dual-stream 3D-R50 exceeds that of MViT-v2, including its dual-stream counterpart, on nearly all metrics except for disjoint uAP. The dual-stream MViT-V2 model, on the other hand, shows improved mAP but a lower uAP when compared with the baseline model, a trend consistent across both dataset configurations. This analysis indicates that BG-only performance in the wildlife domain is not a subset of the FG-only model – which of course also has access to background information. Instead, explicitly learned (and frozen) background features can aid the FG-only baseline in models such as ResNet via late fusion. Finally, all BG-only models trained on \\( \\tilde{\\mathcal{B}} \\) seem to outperform those trained on \\( \\mathcal{B} \\). However, we will show later in Sec. 5.4 that non-synthetic background videos dominate synthetic counterparts when explicitly utilised for performance improvements above the foreground baselines." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.484, + 0.722, + 0.5 + ], + "angle": 0, + "content": "5.3. Background Duration" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.507, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Experimental Overview. Noting that foreground videos may contain some frames with no animals, effectively supplying background only information for these frames, we aim at quantifying the effect of this on model performance (see Fig. 5). Practically, this is achieved by concatenating increasing numbers of background frames to our foreground videos. To do this, we sampled a set of \\(\\lambda T\\) frames from \\(v^{B}\\), where \\(\\lambda\\) is a predefined factor controlling the proportion of background frames used. The sampled background frames are denoted as \\(V^{\\mathcal{B}} = \\{V_{i_0}^{\\mathcal{B}}, V_{i_1}^{\\mathcal{B}}, \\dots, V_{i_{\\lambda T - 1}}^{\\mathcal{B}}\\}\\), with each index \\(i_j\\) selected from \\(\\{0, \\dots, T - 1\\}\\). The foreground frames \\(V^{\\mathcal{F}} = \\{F_0^{\\mathcal{F}}, F_1^{\\mathcal{F}}, \\dots, F_{T - 1}^{\\mathcal{F}}\\}\\) were then concatenated with \\(V^{\\mathcal{B}}\\), resulting in a combined video sequence \\(V^{\\mathrm{concat}} = \\{F_0^{\\mathcal{F}}, F_1^{\\mathcal{F}}, \\dots, F_{T - 1}^{\\mathcal{F}}, F_{i_0}^{\\mathcal{B}}, F_{i_1}^{\\mathcal{B}}, \\dots, F_{i_{\\lambda T - 1}}^{\\mathcal{B}}\\}\\) of length \\(T + \\lambda T\\). From this concatenated sequence, a starting frame \\(F_{\\mathrm{start}}\\) was sampled at random from the indices \\(\\{0, \\dots, \\lambda T\\}\\), ensuring that \\(T\\) consecutive frames could be sampled without exceeding the sequence bounds. The final sampled frame sequence \\(S = \\{F_{\\mathrm{start}}, F_{\\mathrm{start + 1}}, \\dots, F_{\\mathrm{start + T - 1}}\\}\\) was then used as input for training while retaining the original label \\(y\\) from the foreground video \\(v^{\\mathcal{F}}\\). Background frames are appended because motion-triggered cameras are more likely to capture activity first, resulting in a more natural ordering." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Analysing the Continuum Between Foreground and Background Videos. Fig. 5 shows the performance com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5438" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.093, + 0.234, + 0.107 + ], + "angle": 0, + "content": "\\(\\mathcal{D}_{overlap}\\)" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.108, + 0.288, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.093, + 0.422, + 0.107 + ], + "angle": 0, + "content": "\\(\\mathcal{D}_{disjoint}\\)" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.108, + 0.482, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.221, + 0.288, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.221, + 0.482, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.347, + 0.483, + 0.445 + ], + "angle": 0, + "content": "Figure 5. Effect of Increasing Background Duration on Performance. Comparison of 2D R50, 3D R50 and MViT-V2 model performance as a function of the simulated background duration \\((\\lambda)\\). The 3D-R50 is more sensitive to increasing background durations than MViT-V2. MViT-V2 is relatively robust to increasing background durations on the overlapping dataset, although it becomes highly sensitive when evaluated on OOD data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.485, + 0.902 + ], + "angle": 0, + "content": "parison between the 2D R50, 3D R50 and MViT-V2 models as background duration increases in videos \\((\\lambda)\\). As expected, Fig. 5 confirms that as \\(\\lambda\\) increases model performance declines towards background model performance (see Tab 1). On \\(D^{overlap}\\), the 3D-R50 and MViT-V2 models exhibit a sharp decline before beginning to converge. This occurs at \\(\\lambda = 32\\) and \\(\\lambda = 64\\) for the 3D-R50 and MViT-V2 models, respectively. Additionally, as shown by the interval between \\(\\lambda = 0\\) and \\(\\lambda = 32\\), the uAP and mAP of 3D-R50 falls by \\(16.88\\%\\) and \\(14.32\\%\\), respectively, while falling by only \\(7.23\\%\\) and \\(9.13\\%\\) for the MViT-V2 model. Earlier convergence to near BG-only performance and a more rapid rate of decline, show that 3D-R50 is more sensitive to increases in background durations than MViT-V2. Despite starting from lower performance, the 2D R50 model exhibits a more gradual, near-linear decline, indicating less sensitivity to changes in \\(\\lambda\\). While the behaviour of the 2D-R50 and 3D-R50 models remain relatively consistent on \\(D^{disjoint}\\), the behaviour of MViT-V2 changes significantly. Its uAP declines more steeply, falling from \\(70.05\\%\\) to \\(48.29\\%\\) (a decrease of \\(22.21\\%\\)) between \\(\\lambda = 0\\) and \\(\\lambda = 32\\), where its performance converges. It is also more sensitive to smaller increases in background durations, falling from \\(70.50\\%\\) to \\(64.08\\%\\) (a decrease of \\(6.42\\%\\)) between \\(\\lambda = 0\\) and \\(\\lambda = 1.0\\), where a decrease of only \\(0.38\\%\\) is experienced on \\(D^{overlap}\\) over the same interval. At higher values of \\(\\lambda\\) (i.e., \\(>40\\)) the performance drops below that of the 3D-R50 (although only once" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.227 + ], + "angle": 0, + "content": "the background-only performance has been surpassed). Although uAP falls more steeply, mAP declines more slowly and performance above that of the BG-only model is maintained even at the highest values of \\(\\lambda\\). This suggests that, although the MViT-V2 achieves higher baseline performance on the disjoint dataset, it is more sensitive to increases in background frames when evaluating uAP; a scenario where the model's advantage is drowned out by background information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.241, + 0.768, + 0.258 + ], + "angle": 0, + "content": "5.4. Background Bias Mitigation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.264, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Augmenting Performance via Explicit Background Neutralisation. In this final experimental section, we justify the need for paired, real-world background samples, as uniquely provided by our dataset. We show that these sample pairings are able to significantly improve model performance. To show this, we analyse the impact of background subtraction operations in both input and embedding spaces across each of the model architectures and dataset configurations. Two background-subtracted views of our dataset were created for input space experiments. This was done by performing pixel-wise subtraction of \\(\\mathcal{B}\\) and \\(\\tilde{\\mathcal{B}}\\) from \\(\\mathcal{F}\\). To perform algebraic background subtraction in latent space, we extracted the feature vectors of the foreground video pairs. As described in Sec. 5.2, we achieve this by processing the video tensors up to the last convolutional layer and averaging the tokens from the last hidden state in the ResNet and MViT-V2 models, respectively. The background-subtracted embedding is then computed as" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.561, + 0.714, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.653, + 0.68, + 0.682 + ], + "angle": 0, + "content": "foreground x fx 3x 224x 224" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.562, + 0.901, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.654, + 0.885, + 0.684 + ], + "angle": 0, + "content": "background (b x f x 3 x 224 x 224)" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.685, + 0.905, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.814, + 0.906, + 0.897 + ], + "angle": 0, + "content": "Figure 6. Latent Space Background Compensation. The proposed operation uses \\(\\alpha\\) to perform a weighted subtraction of background features (computed via the dataset pairing of a real-world, background video) from foreground features in latent space. The dotted line between the encoders indicates shared weights (i.e., they are the same model)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5439" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.198 + ], + "angle": 0, + "content": "\\(z^{\\mathcal{F} - \\mathcal{B}} = z^{\\mathcal{F}} - (1 - \\alpha)\\cdot z^{\\mathcal{B}}\\) where \\(\\alpha\\) is a coefficient used to control the relative magnitude of the algebraic subtraction operation as schematically shown in Fig. 6. In practice, the modulation parameter \\(\\alpha \\in [0,1]\\) was ablated to increase linearly or exponentially over the course of training and results are reported for both. In all cases \\(\\alpha\\) was initialized at zero and increased to one by the end of training." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.208, + 0.481, + 0.508 + ], + "angle": 0, + "content": "
ModelSubtraction\\(D^{overlap}\\)\\(D^{disjoint}\\)
uAPmAPuAPmAP
2D R50Baseline64.1541.2844.6718.08
\\(\\mathcal{F}+\\mathcal{B}\\)67.7444.9945.2818.84
\\(\\mathcal{F}-\\mathcal{B}\\)65.9642.1254.5326.65
\\(\\mathcal{F}-\\mathcal{B}\\)72.2147.1253.7325.84
None56.2138.8427.2417.94
Linear53.7839.0129.7517.17
Exponential54.0537.1125.8017.73
3D R50Baseline72.7949.3257.7430.89
\\(\\mathcal{F}+\\mathcal{B}\\)76.4153.5765.6639.17
\\(\\mathcal{F}-\\mathcal{B}\\)66.5142.5653.5827.98
\\(\\mathcal{F}-\\mathcal{B}\\)71.9346.3653.9225.05
None72.3051.3963.1434.85
Linear77.3152.8266.6936.31
Exponential73.5251.2464.5836.36
Baseline75.4451.5470.5035.29
\\(\\mathcal{F}+\\mathcal{B}\\)72.8253.4968.0038.09
MViT-V2\\(\\mathcal{F}-\\mathcal{B}\\)66.8442.2757.2128.80
\\(\\mathcal{F}-\\mathcal{B}\\)74.9049.4061.2231.51
None74.2344.1168.5635.14
Linear79.32*53.83*71.64*39.04
Exponential77.7050.9371.2839.25*
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.691 + ], + "angle": 0, + "content": "Table 2. Quantifying Performance Improvements of Background Subtraction. Comparison of the 2D R50, 3D R50, and MViT-V2 models on both dataset configurations. Performance is reported when training on \\(\\mathcal{F} - \\mathcal{B}\\) or \\(\\mathcal{F} - \\tilde{\\mathcal{B}}\\) videos in input (light green) and embedding space (light blue). First and second highest scores are indicated for each model, while the highest overall scores are marked by an asterisks \\((\\ast)\\). Note the baseline is trained on \\(\\mathcal{F}\\), as reported in Tab. 1. \\(\\mathcal{F} + \\mathcal{B}\\) provides an alternative baseline but is reliant on background videos at test time. Results confirm the superiority of real-world background videos over synthetic data, and of latent space neutralisation over input space background subtraction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results and Performance Improvements. Tab. 2 shows that for the basic 2D-R50 model, background subtraction in input space is highly effective. When using \\(\\mathcal{B}\\), we observe substantial performance improvements over the baseline, with percentage increases of \\(+8.06\\%\\) and \\(5.84\\%\\) in uAP and mAP on \\(D^{overlap}\\), and \\(+9.06\\%\\) and \\(+7.76\\%\\) in uAP and mAP on \\(D^{disjoint}\\). Using \\(\\tilde{\\mathcal{B}}\\) also yields improvements, though these gains are notably smaller. Background subtraction in embedding space does not provide any benefit to the 2D-R50. Although input space background subtraction is effective for the 2D-R50, it does not improve performance for the 3D-R50 and MViT-V2 models. Neither original nor synthetic backgrounds provide an advantage over" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "their respective baselines, indicating that input space subtraction may not suit spatio-temporal architectures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.905, + 0.593 + ], + "angle": 0, + "content": "Conversely, background subtraction in embedding space provides significant performance improvements for the 3D-R50 and MViT-V2 models, particularly on \\( D^{disjoint} \\). For the 3D-R50, embedding space subtraction with linear scheduling results in a \\( +8.95\\% \\) and \\( +5.42\\% \\) increase in uAP and mAP, respectively. Exponential scheduling also yields improvements over the baseline. The improvements are lower than linear scheduling with respect to uAP \\( (+6.84\\%) \\) but marginally better for mAP \\( (+5.47\\%) \\). MViT-V2 also benefits from embedding space subtraction on \\( D^{disjoint} \\): linear scheduling improves uAP and mAP by \\( +1.14\\% \\) and \\( +3.75\\% \\), respectively, whereas exponential scheduling provides a \\( +0.78\\% \\) increase in uAP and a \\( +3.96\\% \\) increase in mAP. These findings suggest that embedding space subtraction is effective for both 3D-R50 and MViT-V2 models on \\( D^{disjoint} \\), with linear and exponential scheduling favouring uAP and mAP performance, respectively. Embedding space subtraction also enhances performance on \\( D^{overlap} \\), though to a lesser extent than on \\( D^{disjoint} \\). For the 3D-R50 model, applying embedding space subtraction with linear scheduling yields an improvement of \\( 4.52\\% \\) and \\( 3.50\\% \\) in uAP and mAP, respectively, while exponential scheduling results in smaller gains of \\( 0.73\\% \\) and \\( 1.92\\% \\). The MViT-V2 model follows a similar trend: linear scheduling increases uAP and mAP by \\( 3.88\\% \\) and \\( 2.29\\% \\), respectively, whereas exponential scheduling improves uAP by \\( 2.26\\% \\) but slightly reduces mAP by \\( -0.61\\% \\). These results suggest that algebraic embedding space subtraction is beneficial on \\( D^{overlap} \\), although the performance gains are relatively small compared to \\( D^{disjoint} \\), and that linear scheduling generally provides better results than exponential scheduling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.616, + 0.634, + 0.632 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduced PanAf-FGBG, the first wildlife dataset designed to assess backgrounds and out-of-distribution generalization in behavior recognition. Our results demonstrate that real-world background information is a strong predictor of animal behavior and can significantly enhance recognition performance. While there are some limitations — such as its restriction to statically installed camera trap videos, the additional data and compute required for latent space background subtraction, and the potential for unnatural artifacts in background concatenation — the dataset remains a valuable resource. PanAf-FGBG enables direct evaluation under out-of-distribution conditions, and its rich metadata supports additional tasks and experimental setups. We hope it will help with understanding and improving model performance in this challenging domain and ultimately benefit endangered and charismatic species like the one featured in this paper." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5440" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.118, + 0.486, + 0.65 + ], + "angle": 0, + "content": "We thank the Pan African Programme: 'The Cultured Chimpanzee' team and its collaborators for allowing the use of their data for this paper. We thank Amelie Pettrich, Antonio Buzharevski, Eva Martinez Garcia, Ivana Kirchmair, Sebastian Schutte, Linda Gerlach and Fabina Haas. We also thank management and support staff across all sites; specifically Yasmin Moebius, Geoffrey Muhanguzi, Martha Robbins, Henk Eshuis, Sergio Marrocoli and John Hart. Thanks to the team at https://www.chimpandsee.org particularly Briana Harder, Anja Landsmann, Laura K. Lynn, Zuzana Macháčková, Heidi Pfund, Kristeena Sigler and Jane Wideness. The work that allowed for the collection of the dataset was funded by the Max Planck Society, Max Planck Society Innovation Fund, and Heinz L. Krekeler. In this respect we would like to thank: Ministre des Eaux et Forets, Ministère de l'Enseignement supérieur et de la Recherche scientifique in Côte d'Ivoire; Institut Congolais pour la Conservation de la Nature, Ministère de la Recherche Scientifique in Democratic Republic of Congo; Forestry Development Authority in Liberia; Direction Des Eaux Et Forets, Chasses Et Conservation Des Sols in Senegal; Makerere University Biological Field Station, Uganda National Council for Science and Technology, Uganda Wildlife Authority, National Forestry Authority in Uganda; National Institute for Forestry Development and Protected Area Management, Ministry of Agriculture and Forests, Ministry of Fisheries and Environment in Equatorial Guinea. This work was supported by the UKRI CDT in Interactive AI (grant EP/S022937/1). This work was in part supported by the US National Science Foundation Awards No. 2118240 \"HDR Institute: Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning\" and Award No. 2330423 and Natural Sciences and Engineering Research Council of Canada under Award No. 585136 for the \"AI and Biodiversity Change (ABC) Global Center\"." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.669, + 0.188, + 0.685 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.695, + 0.483, + 0.722 + ], + "angle": 0, + "content": "[1] Pan African Programme: The Cultured Chimpanzee, 1999. Online; accessed 29 January 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.727, + 0.484, + 0.768 + ], + "angle": 0, + "content": "[2] Analysing animal behaviour in wildlife videos using face detection and tracking. IEEE Proceedings-Vision, Image and Signal Processing, 153(3):305-312, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.772, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[3] Mimi Arandjelovic, Colleen R Stephens, Maureen S McCarthy, Paula Dieguez, Emmie K Kalan, Nuria Maldonado, Christophe Boesch, and Hjalmar S Kuehl. Chimp&see: An online citizen science platform for large-scale, remote video camera trap annotation of chimpanzee behaviour, demography and individual identification. PeerJ Preprints, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Daniel Schofield, Sophie Berdugo, Joana Bessa, Jake Owen, Kimberly J Hockings, Tetsuro Matsuzawa, Misato Hayashi, Dora Biro, et al. Au" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.695, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tomated audiovisual behavior recognition in wild primates. Science advances, 7(46):eabi4883, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[5] Sara Beery, Arushi Agarwal, Elijah Cole, and Vighnesh Birodkar. The iwildcam 2021 competition dataset. arXiv preprint arXiv:2105.03494, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.163, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[6] Otto Brookes, Majid Mirmehdi, Hjalmar S. Kuhl, and Tilo Burghardt. Triple-stream deep metric learning of great ape behavioural actions. In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 294-302, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[7] Otto Brookes, Majid Mirmehdi, Hjalmar Kuhl, and Tilo Burghardt. Chimpvlm: Ethogram-enhanced chimpanzee behaviour recognition. 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.289, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[8] Otto Brookes, Majid Mirmethdi, Colleen Stephens, Samuel Angedakin, Katherine Corogenes, Dervla Dowd, Paula Dieguez, Thurston C Hicks, Sorrel Jones, Kevin Lee, et al. Panaf20k: a large video dataset for wild ape detection and behaviour recognition. International Journal of Computer Vision, pages 1-17, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.372, + 0.907, + 0.427 + ], + "angle": 0, + "content": "[9] Jackie Chappell and Susannah KS Thorpe. The role of great ape behavioral ecology in one health: Implications for captive welfare and re-habilitation success. American journal of primatology, 84(4-5):e23328, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[10] Jun Chen, Ming Hu, Darren J Coker, Michael L Berumen, Blair Costelloe, Sara Beery, Anna Rohrbach, and Mohamed Elhoseiny. Mammalnet: A large-scale video benchmark for mammal recognition and behavior understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13052-13061, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[11] Jinwoo Choi, Chen Gao, Joseph CE Messou, and Jia-Bin Huang. Why can't i dance in the mall? learning to mitigate scene bias in action recognition. Advances in Neural Information Processing Systems, 32, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[12] Jihoon Chung, Yu Wu, and Olga Russakovsky. Enabling detailed action recognition evaluation through video dataset augmentation. Advances in Neural Information Processing Systems, 35:39020-39033, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[13] Fay E Clark. Great ape cognition and captive care: Can cognitive challenges enhance well-being? Applied Animal Behaviour Science, 135(1-2):1-12, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[14] Fagner Cunha, Eulanda M dos Santos, Raimundo Barreto, and Juan G Colonna. Filtering empty camera trap images in embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2438-2446, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[15] Alexander D'Amour, Katherine Heller, Dan Moldovan, Ben Adlam, Babak Alipanahi, Alex Beutel, Christina Chen, Jonathan Deaton, Jacob Eisenstein, Matthew D Hoffman, et al. Underspecification presents challenges for credibility in modern machine learning. Journal of Machine Learning Research, 23(226):1-61, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[16] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5441" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[17] Isla Duporge, Maksim Kholiavchenko, Roi Harel, Dan Rubenstein, Meg Crofoot, Tanya Berger-Wolf, Stephen Lee, Scott Wolf, Julie Barreau, Jenna Kline, et al. Baboonland dataset: Tracking primates in the wild and automating behaviour recognition from drone videos. arXiv preprint arXiv:2405.17698, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[18] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[19] Rohit Girdhar and Deva Ramanan. Cater: A diagnostic dataset for compositional actions and temporal reasoning. arXiv preprint arXiv:1910.04744, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[20] Abhinav Gupta and Larry S Davis. Objects in action: An approach for combining action understanding and object perception. pages 1-8. IEEE, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[22] Yun He, Soma Shirakabe, Yutaka Satoh, and Hirokatsu Kataoka. Human action recognition without human. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 11-17. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[23] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In Proceedings of the European conference on computer vision (ECCV), pages 771-787, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[24] Filip Ilic, Thomas Pock, and Richard P Wildes. Is appearance free action recognition possible? In European Conference on Computer Vision, pages 156-173. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[25] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[26] Maksim Kholiavchenko, Jenna Kline, Maksim Kukushkin, Otto Brookes, Sam Stevens, Isla Duporge, Alec Sheets, Reshma R Babu, Namrata Banerji, Elizabeth Campolongo, et al. Deep dive into kabr: a dataset for understanding ungulate behavior from in-situ drone video. Multimedia Tools and Applications, pages 1-20, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[27] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[28] Hjalmar S Kuhl and Tilo Burghardt. Animal biometrics: quantifying and detecting phenotypic appearance. Trends in ecology & evolution, 28(7):432-441, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[29] Haoxin Li, Yuan Liu, Hanwang Zhang, and Boyang Li. Mitigating and evaluating static bias of action representations in the background and the foreground. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 19911-19923, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[30] Yi Li and Nuno Vasconcelos. Repair: Removing representation bias by dataset resampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9572-9581, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[31] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[32] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4804-4814, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[33] Dan Liu, Jin Hou, Shaoli Huang, Jing Liu, Yuxin He, Bochuan Zheng, Jifeng Ning, and Jingdong Zhang. Lote-animal: A long time-span dataset for endangered animal behavior understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20064-20075, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[34] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[35] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[36] Xiaoxuan Ma, Stephan Kaufhold, Jiajun Su, Wentao Zhu, Jack Terwilliger, Andres Meza, Yixin Zhu, Federico Rossano, and Yizhou Wang. Chimpact: A longitudinal dataset for understanding chimpanzee behaviors. Advances in Neural Information Processing Systems, 36:27501-27531, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[37] Marcin Marszalek, Ivan Laptev, and Cordelia Schmid. Actions in context. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 2929-2936. IEEE, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[38] Zhongqi Miao, Ziwei Liu, Kaitlyn M Gaynor, Meredith S Palmer, Stella X Yu, and Wayne M Getz. Iterative human and automated identification of wildlife images. Nature Machine Intelligence, 3(10):885-895, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[39] Mazda Moayeri, Phillip Pope, Yogesh Balaji, and Soheil Feizi. A comprehensive study of image classification model sensitivity to foregrounds, backgrounds, and visual attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19087-19097, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[40] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19023-19034, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[41] Toby Perrett, Saptarshi Sinha, Tilo Burghardt, Majid Mirmethdi, and Dima Damen. Use your head: Improving long-tail video recognition. In Proceedings of the IEEE/CVF con" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5442" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ference on computer vision and pattern recognition, pages 2415-2425, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[42] Laura J Pollock, Justin Kitzes, Sara Beery, Kaitlyn M Gaynor, Marta A Jarzyna, Oisin Mac Aodha, Bernd Meyer, David Rolnick, Graham W Taylor, Devis Tuia, et al. Harnessing artificial intelligence to fill global shortfalls in biodiversity knowledge. Nature Reviews Biodiversity, pages 1-17, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[43] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[44] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[45] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[46] Faizaan Sakib and Tilo Burghardt. Visual recognition of great ape behaviours in the wild. In Workshop on the Visual Observation and Analysis of Vertebrate and Insect Behaviour, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.49 + ], + "angle": 0, + "content": "[47] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9661-9669, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[48] Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank Van Langevelde, Tilo Burghardt, et al. Perspectives in machine learning for wildlife conservation. Nature communications, 13(1):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[49] Tuan-Hung Vu, Catherine Olsson, Ivan Laptev, Aude Oliva, and Josef Sivic. Predicting actions from static scenes. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 421-436. Springer, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[50] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11804-11813, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[51] Kai Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. arXiv preprint arXiv:2006.09994, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[52] Xinyu Yang, Majid Mirmehdi, and Tilo Burghardt. Great ape detection in challenging jungle camera trap footage via attention-based spatial and temporal feature blending. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[53] Xinyu Yang, Tilo Burghardt, and Majid Mirmehdi. Dynamic curriculum learning for great ape detection in the wild. In-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "ternational Journal of Computer Vision, 131(5):1163-1181, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[54] Manlin Zhang, Jinpeng Wang, and Andy J Ma. Suppressing static visual cues via normalizing flows for self-supervised video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3300-3308, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[55] Mingjun Zhao, Yakun Yu, Xiaoli Wang, Lei Yang, and Di Niu. Search-map-search: a frame selection paradigm for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10627-10636, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[56] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[57] Yi Zhu, Xinyu Li, Chunhui Liu, Mohammadreza Zolfaghari, Yuanjun Xiong, Chongruo Wu, Zhi Zhang, Joseph Tighe, R Manmatha, and Mu Li. A comprehensive study of deep video action recognition. arXiv preprint arXiv:2012.06567, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5443" + } + ] +] \ No newline at end of file diff --git a/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_origin.pdf b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d064ff40a95ac56f7842df38b3d95d4d1b1f1a91 --- /dev/null +++ b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/2e0201f3-4718-4a61-af3c-ec2f92f1137d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5df6c73301b15808a92135cdffec776b8cd315a8c5d9df23b38b5e9ea4c2857d +size 5349143 diff --git a/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/full.md b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5464d3e0a81de285807e54eee3523e23aed7e536 --- /dev/null +++ b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/full.md @@ -0,0 +1,224 @@ +# The PanAf-FGBG Dataset: + +# Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition + +Otto Brookes $^{1,2\dagger*}$ , Maksim Kukushkin $^{3,4*}$ , Majid Mirmehdi $^{1}$ , Colleen Stephens $^{5}$ , Paula Dieguez $^{5}$ , Thurston C. Hicks $^{6}$ , Sorrel Jones $^{5}$ , Kevin Lee $^{5}$ , Maureen S. McCarthy $^{5}$ , Amelia Meier $^{5}$ , Emmanuelle Normand $^{2}$ , Erin G. Wessling $^{7}$ , Roman M. Wittig $^{5,8}$ , Kevin Langergraber $^{9}$ , Klaus Zuberbühler $^{10}$ , Lukas Boesch $^{2}$ , Thomas Schmid $^{3,11}$ , Mimi Arandjelovic $^{5}$ , Hjalmar Kuhl $^{5,12}$ , Tilo Burghardt $^{1}$ + +1University of Bristol, 2Wild Chimpanzee Foundation, 3Martin Luther University Halle-Wittenberg, 4Leipzig University, 5Max Planck Institute for Evolutionary Anthropology, 6University of Warsaw, 7Harvard University, 8University of Lyon, 9Arizona State University, 10University of St Andrews, 11Lancaster University in Leipzig, 12Senckenberg Museum of Natural History + +# Abstract + +Computer vision analysis of camera trap video footage is essential for wildlife conservation, as captured behaviours offer some of the earliest indicators of changes in population health. Recently, several high-impact animal behaviour datasets and methods have been introduced to encourage their use; however, the role of behaviour-correlated background information and its significant effect on out-of-distribution generalisation remain unexplored. In response, we present the PanAf-FGBG dataset, featuring 21 hours of wild chimpanzee behaviours, recorded at 389 individual camera locations. Uniquely, it pairs every video with a chimpanzee (referred to as a foreground video) with a corresponding background video (with no chimpanzee) from the same camera location. We present two views of the dataset: one with overlapping camera locations and one with disjoint locations. This setup enables, for the first time, direct evaluation of in-distribution and out-of-distribution conditions, and for the impact of backgrounds on behaviour recognition models to be quantified. All clips have rich behavioural annotations and metadata, including unique camera IDs. Additionally, we establish several baselines and present a highly effective latent-space normalisation technique that boosts out-of-distribution performance by $+5.42\%$ mAP for convolutional and $+3.75\%$ mAP for transformer-based models. Finally, we provide an in-depth analysis on the role of backgrounds in out-of-distribution behaviour recognition, including the so far unexplored impact of background durations. (i.e., the count of background frames within foreground videos). The dataset is available at https://obrookes.github.io/panaf-fbgb.github.io/. + +![](images/d5f14db0bf10262b794a2615c63ddc6abd774994f8b9277f8062d80850ed8d50.jpg) +Figure 1. Conceptual Overview. The PanAf-FGBG dataset comprises $>20$ hours of paired and richly annotated foreground-background camera trap videos of wild chimpanzees. The dataset unlocks systematic analyses of the impact of background information on wildlife behaviour recognition. We provide baselines, quantify the background impact on performance, and demonstrate that utilising background information in latent space (see blue data point) can significantly improve recognition performance in this challenging domain. + +# 1. Introduction + +Deep learning models have shown impressive results on action recognition tasks [57], particularly when both training and test data share the same distribution. However, their performance declines significantly on out-of-distribution (OOD) data [15, 27]. In practice, distribution shifts are common in nearly all settings [43], but are particularly unforgiving when computer vision is applied to uncontrolled, complex or natural environments [27]. + +In the human action recognition domain, poor OOD performance is well-known to originate from shortcut learning [18] of static cues or backgrounds [24, 29, 30]. This leads models to rely heavily on background information instead of the behaviour of interest solely [2], and several studies have shown that the majority of action recognition performance in this domain can be attributed to background learning [12, 22, 49]. Despite the known impact of backgrounds on action recognition and a critical need to advance model performance in wildlife applications [28, 42, 48], reliance on static cues for animal behaviour recognition has received limited attention. This is largely due to a lack of datasets explicitly designed to study its effects. + +In response, we introduce the PanAf-FGBG dataset, which comprises real-world foreground-background video pairs (see Fig. 1) extracted from six African countries, 14 national parks, and over 350 individual camera locations across many habitats and location types (see Fig. 3). Here, we define foreground videos as those containing a chimpanzee, and background videos as those that do not. This allows us to address a key limitation in existing studies, where the background's role is typically assessed by creating background-only versions of datasets through synthetic modifications (e.g., masking or in-painting) that remove the actor [11, 12, 23]. While such synthetic inputs often serve as useful proxies, the actual effect of synthesis on training is largely unknown [39, 47]. + +In contrast, our dataset introduces paired camera locations and enables a truly realistic disentangling of background effects under in- and out-of-distribution conditions: disjoint experiments with mutually exclusive camera locations for training and test splits, and overlapping experiments with shared locations. This focus is particularly important in the wildlife domain, where the transferability and generalisation of camera trap analysis to new locations is critical for effective operation. While studies have shown the over-reliance of action recognition models on backgrounds [22, 37, 49], we find that the effect of background duration (i.e., the count of background frames within foreground videos with no actor) is largely unexplored. In camera trap video footage like in many other CCTV or triggered camera settings, actors may only appear momentarily, leaving behind footage of only the background [14, 38] within foreground videos. Since this effect alters learning given + +video-level behaviour annotations, we quantify its effect on several action recognition architectures. + +Our contributions are as follows: (i) we release PanAf-FGBG with its rich annotations and meta-data, providing a large wildlife dataset suitable for explicitly investigating the impact of backgrounds on behaviour recognition; (ii) we show models trained on background-only videos achieve strong animal behaviour recognition performance ( $\sim 65\%$ of baseline performance), corroborating findings in the human domain; (iii) we conduct the first evaluation of background duration in videos, revealing its impact on action recognition models and highlighting key differences between convolutional and attention-based networks; and (iv) we introduce a simple yet powerful latent-space background neutralisation that yields notable performance gains, particularly in OOD scenarios for convolutional $(+5.42\%)$ mAP) and transformer-based approaches $(+3.75\%)$ mAP). + +# 2. Related Work + +Animal Behaviour Datasets. In recent years, several high-impact animal behaviour datasets have been introduced to the CV community, often accompanied by new methods [4, 6-8, 10, 26, 33, 40, 46]. Although they all centre on animal behaviour understanding, they each have a different focus. The large-scale datasets Animal Kingdom [40] and MammalNet [10] concentrate on behavioural understanding across species, each including videos of $\sim 800$ different species. They provide atomic action and high-level behaviour annotations, respectively, and support, in addition to action recognition, the related tasks of action localisation [10, 40] and pose estimation [40]. The smaller ChimpAct dataset [36] provides longitudinal data that follows a zoo-housed group of 20 chimpanzees with a focus on a single male juvenile. It also contains key-point pose information, and fine-grained spatio-temporal behaviour labels. While valuable, these datasets lack in-situ footage from ecological sensors like camera traps and drones, meaning animals are not observed in their natural environment. This may result in unrepresentative behaviour distributions and limit their suitability for studying wild behaviour [9, 13]. + +In contrast, KABR [26] and BaboonLand [17] comprise drone footage of free-living Kenyan wildlife, with fine-grained spatio-temporal behaviour annotations, focussing on individual and group-level behaviour, respectively. The LoTE dataset [33] provides camera trap video footage of 11 endangered species from a 12-year longitudinal study; it is accompanied by environmental information (i.e., weather conditions and habitat) and highlights the importance of temporal and spatial continuity for conservation efforts. The dataset most comparable to our own is the PanAf20k dataset [8] and its related PanAf500-based works [46, 52, 53]. PanAf20k comprises camera trap videos of great apes accompanied by multi-label behaviour annotations. + +Our dataset orthogonally complements this growing body of work dedicated to understanding animal behaviour by providing the first dataset specifically designed to evaluate backgrounds and out-of-distribution generalisation for behaviour recognition. It is unique in providing foreground-background pairs, and its configurations for in- and out-of-distribution evaluation are defined identically to the well-known iWildCam [5] and Wilds dataset [27] for species classification. Similar to LotE [33], we provide a detailed description of the background, however we also include a unique camera ID (i.e., geospatial location) to allow footage taken from the same camera to be linked. Similar to [8], our data is gathered from 14 national parks, spanning 6 different countries, although unlike ours they do not include metadata to allow videos to be linked to specific countries, research sites, locations, or habitats. Similarly, it is more diverse than [33] with its footage from a single national park. + +Background Impact. The effect of backgrounds or static bias is well-studied within computer vision [29, 54], where it has shown to be necessary for models to perform well [51]. Action recognition models are particularly vulnerable to shortcut learning. These models often rely on correlated, but distinct information beyond the intended action, such as object [20, 56], scene [37], and actor biases [31]. For example, several studies have reported that strong performance on action recognition tasks can be achieved without the actor [22, 37, 49] and that approximately $70\%$ of model performance is attributable to static cues (i.e., scene or background) [12]. This results in biased representations and ultimately impairs out-of-distribution generalization performance [30]. + +Background Bias Datasets. The effect of the background has often been modelled using synthetically generated data that removes either the foreground or the background. [11, 12, 19, 29]. For instance, Ilic et al. [24] present a synthetic version of the UCF101 dataset, animating spatial noise using image motion extracted by an optical flow estimator, Li et al. [29] replace the backgrounds of IID HMDB51 with sinusoidal stripe images, and Chung et al [12] adopt an in-painting model to remove actors from the video. Datasets that include fine-grained spatio-temporal action annotations also remove static cues implicitly by isolating the region of the video where the action is occurring. + +Mitigating Background Bias. Prominent attempts to mitigate this bias include dataset resampling methods, which down-weight biased samples [30, 31], adversarial training regimes that rely on scene labels and human masks [11], and augmentation techniques that mix or swap the backgrounds of different videos during training [16, 29, 50]. There are also several implicit approaches, such as frame selection, that aim to utilise only frames where activity is occurring [55]. Our dataset, introduced next, instead introduces foreground-background pairing in + +formation with close proximity in time of capture, which describes the environment truly independently and in relation to the presence of animals showcasing behaviour. + +# 3. Dataset + +Dataset Overview and Statistics. PanAf-FGBG comprises footage gathered under ethical oversight as part of the PanAf Programme: The Cultured Chimpanzee [1]. It contains 21 hours of camera trap footage of individual chimpanzees in tropical Africa. Its footage is collected from 389 individual camera locations across 14 national parks in 6 African countries. In total, we provide 5,070 video pairs, each 15 seconds in duration. All videos are accompanied by a multi-label behaviour annotation and scene metadata, plus the unique feature of foreground-background video pairing (see Fig. 1). Class and habitat distributions of the data are given in Fig. 2 and Fig. 3, respectively. Our dataset follows a long-tailed distribution [34, 41] common in animal behaviour datasets [8, 10, 33, 36, 40]. The three of the most commonly occurring classes are observed in $>60\%$ of videos, while the rarest are observed in fewer than $< 3\%$ . + +Data Collection. The data collection strategy was designed specifically to maximise the chance of filming the terrestrial activity of apes and we provide only a brief description here (for full details we refer the reader to [1]). Data was collected from 14 different research sites, where an average of 29 movement-triggered Bushnell cameras were installed per site. Grids comprising 20 to $961 \times 1$ km cells were established and one camera was installed per grid cell. Footage was recorded at 30 frames-per-second and a resolution of $1280 \times 720$ . The cameras were visited every 1-3 months for maintenance and to download the recorded footage throughout the study periods. + +Ethical Considerations. To protect the location of the apes from threats (e.g., poaching) we use monikers for the names of the research sites and assign a unique camera ID to each video, hiding the geospatial location of the cameras. + +![](images/69e55246557040f4101ce07536dce39f1e1a9b9b869b1e47ce1baeae86eec24b.jpg) +Figure 2. Distribution of Behaviour. Proportion of behaviours in the dataset where smaller segments are colour coordinated. + +![](images/044ebbe43aa5fbd39b131aa99a39f6bd76127a9a1f2ce571bfa8a12ac2ccea3b.jpg) +Figure 3. Distribution of Habitat and Location Type. The pie chart gives the proportion of videos taken at each of the three main habitats: forest, marshes and savannah. The histogram shows the total number of videos extracted from each location type. + +Data Annotation. The behavioural annotations were provided by users on the community science platform Chimp&See [3]. Annotators were presented with a choice of behaviour classification categories, chosen specifically by experts for their ecological importance. Behaviours not listed in the classification categories were also permitted. These annotations were then extracted and expertly grouped into 14 co-occurring classes, which form the multi-label behavioural annotations presented here. The annotations follow a multi-hot binary format that indicate the presence of one or more behaviours. It should also be noted that behaviours are not assigned to individual apes or temporally localised within each video. To ensure annotation quality and consistency a video was only deemed to be analysed when either three volunteers marked the video as blank, or unanimous agreement between seven volunteers was observed, or 15 volunteers annotated the video. Habitat type was noted during camera setup. + +Foreground-Background Pair Selection. Pairs are formed by selecting the temporally closest available foreground and background videos to ensure similarity. Complete standardisation statistics can be found in Sup. Mat. + +Overlapping & Disjoint Configurations. We present both a disjoint view of dataset, where camera locations in the training and test splits are mutually exclusive, denoted as $\mathcal{D}^{disjoint} = \{d_i \mid d_i \cap d_j = \emptyset, \forall i \neq j\}$ following [27]. A second, overlapping view is also provided, where camera + +locations are shared between training and test splits, represented as $\mathcal{D}^{overlap} = \{d_i \mid d_i \cap d_j \neq \emptyset$ for some $i \neq j\}$ . Note that $d_i$ refers to sets of camera locations associated with videos in each split. Fig. 4 provides an overview of both configurations. We ensure that the class distribution in each configuration remains approximately the same for full comparability. Unique camera identifiers are available for each video to allow researchers interested in this problem to create additional configurations, should they wish to. + +# 4. Preliminaries + +We train multi-label classification models on both views of the dataset: $\mathcal{D}^{\mathrm{overlap}}$ and $\mathcal{D}^{\mathrm{disjoint}}$ using various combinations of foreground $(\mathcal{F})$ , background $(\mathcal{B})$ and synthetic background $(\tilde{\mathcal{B}})$ videos. The latter were created by generating a segmentation mask using SAM2 [44] followed by mean pixel value filling, similar to [11, 12, 23]. Both views comprise foreground-background video pairs $\{(v_{i}^{\mathcal{F}}, v_{i}^{\mathcal{B}})\}_{i=1}^{N}$ , where $v_{i}^{\mathcal{F}} \in \mathcal{F}$ and $v_{i}^{\mathcal{B}} \in \mathcal{B}$ , with $N$ representing the total number of video pairs. Each foreground video $v_{i}^{\mathcal{F}}$ is associated with a video-level label vector $y_{i} = (y_{i1}, y_{i2}, \ldots, y_{ic})$ , where each $y_{ij} \in \{0,1\}$ indicates the presence (1) or absence (0) of the $j$ -th class in video $i$ , and $C$ is the total number of classes. + +# 5. Experiments + +# 5.1. Setup + +All experiments were performed using the ResNet-50 [21] and MViT-V2 [32] architectures, which were initialised with feature extractors pre-trained on Kinetics-400 [25]. For ResNet-50, both 2D and 3D convolution versions are tested. After initial experiments to ensure optimal optimiser-network pairings for the task, ResNet and MViT-V2 models were fine-tuned for 300 epochs using the SGD [45] and AdamW [35] optimisers, respectively. Both models followed a linear warm-up schedule with subsequent cosine annealing. For ResNet models, the learning rate was increased from $1\times 10^{-5}$ to $1\times 10^{-4}$ over 20 epochs, while it was increased from $1\times 10^{-6}$ to $1\times 10^{-4}$ over 30 epochs for MViT-V2 models. In both cases, momentum of 0.9 was utilised. All computation was performed on $4\times$ NVIDIA H200 GPUs using a distributed batch size of 336. During training, video tensors with a spatio-temporal resolution of $16\times 256\times 256$ were employed. Frames were sampled uniformly at a rate of $1/24$ , avoiding dense sampling strategies which are shown to exacerbate shortcut learning in action recognition models [12]. For MViT-V2 models, a patch size of $16\times 16$ pixels was utilised. Evaluation was performed using micro and mean Average Precision and are referred to as uAP and mAP, respectively. + +![](images/02d4d3d3405d48170cb7f18b37bf5f1e938866b7e0a4d83fcc726664f2caae48.jpg) +Figure 4. Overlapping & Disjoint Dataset Configurations. Six foreground-background video pairs are shown to visualise overlapping (left) and disjoint (right) dataset configurations. On the left, train and test videos taken from the same camera are selected, emphasizing that the locations are not mutually exclusive, with backgrounds containing several cues that are highly correlated with the displayed behaviours (i.e., algae fishing, a form of tool use). Train and test videos on the right are extracted from mutually exclusive or disjoint camera locations. They are extracted from forest and savannah, respectively, and highlight the challenge of generalising to new locations. + +# 5.2. Background Reliance + +Experimental Overview. First, we examined the reliance of animal behaviour recognition on the background by training model architectures on the foreground $\mathcal{F}$ , background $\mathcal{B}$ , and synthetic background videos $\tilde{\mathcal{B}}$ only. We refer to models trained on $\mathcal{F}$ and $\mathcal{B}$ (or $\tilde{\mathcal{B}}$ ) as FG-only and BG-only models, respectively, as shown earlier in Fig. 1. To separate background information from class-distribution intrinsic information, a dummy classifier with no access to visual information was also evaluated. It would effectively return predictions based on the class distribution information alone. Since our dataset consists of unique foreground-background pairs, it was possible to train BG-only models by retaining the original labels associated with the foreground video of the pairing. In all cases, models were tested on $\mathcal{F}$ for comparability. Note that this makes the task more difficult for BG-only models, since animals are only observed at test time and not during training. All experiments are conducted on both $D^{overlap}$ and $D^{disjoint}$ . + +Dual-Stream Fusion Model $(\mathcal{F} + \mathcal{B})$ . To clarify whether BG-only models and FG-only models, each trained separately, can be fused late to orthogonally contribute towards performance we also implemented a dual-stream model. To do this, the independently optimised weights of the FG-only and BG-only models trained for 300 epochs were loaded. During training the weights were frozen and video pairs + +were processed to extract late features $z^{F}$ and $z^{B}$ . For ResNet models, features were extracted from the last convolutional layer and in MViT-V2 models an average of the tokens from the last hidden state were taken. Late features were then concatenated $\tilde{z} = [z^{F}, z^{B}]$ and final class predictions were produced by a trainable multi-layer perceptron. + +Result 1 - Backgrounds are a Predictor of Animal Behaviour. Table 1 shows that BG-only models achieve strong performance when compared to their respective FG-only counterparts. To illustrate this, the ratio of background-toForeground performance $\frac{AP^B}{AP^F}$ published in [12], showing how close a model can get to FG performance with just BG cues, is reported for each architecture in rows 6, 11, and 16. Background-to foreground performance ratios are consistently high across all models, never dropping below 0.70 and 0.65 for uAP and mAP, respectively, on either dataset. This indicates that much of the behaviour recognition performance is indeed achievable by just utilisation of background information - the environment is a strong predictor of behaviour, even out-of-distribution. Thus, features of the environment are already behaviour predictors (trees for climbing, paths for travel, termite mound for feeding etc.). This result is further compounded by the fact that BG-only models outperform the non-visual classifier (row 1) by a significant margin, particularly on $D^{overlap}$ . On $D^{disjoint}$ this observation still holds, although by somewhat + +
ModelData\(D^{overlap}\)\(D^{disjoint}\)
uAPmAPuAPmAP
1Dummy-19.1514.7319.6114.87
2\(\bar{\mathcal{F}}\)64.1541.2844.6718.08
32D R50\(\mathcal{B}\)57.1330.0345.8018.27
4\(\tilde{\mathcal{B}}\)57.8730.8248.2820.39
5\(\mathcal{F}+\mathcal{B}\)67.7444.9945.2818.84
60.890.741.020.98
7\(\bar{\mathcal{F}}\)72.7949.3257.7430.89
83D R50\(\mathcal{B}\)58.9231.7248.4021.38
9\(\tilde{\mathcal{B}}\)59.6032.8649.2921.64
10\(\mathcal{F}+\mathcal{B}\)76.4153.5765.6639.17
110.800.660.830.69
12\(\bar{\mathcal{F}}\)75.4451.5470.5035.29
13MViT-V2\(\mathcal{B}\)61.9032.7450.0324.00
14\(\tilde{\mathcal{B}}\)62.1533.5852.1324.77
15\(\mathcal{F}+\mathcal{B}\)72.8253.4968.0038.09
160.820.650.700.68
+ +Table 1. Quantifying Background Reliance. Performance comparison of models trained on $\mathcal{F}$ , $\mathcal{B}$ , and $\tilde{\mathcal{B}}$ videos. Dual-stream models are indicated by $\mathcal{F} + \mathcal{B}$ . The absolute highest performance is indicated in bold whereas best performance using only a single-stream is underlined. The performance based on class distribution only is given by a dummy classifier (see row 1). The background-toForeground performance ratio $(\frac{AP^B}{AP^F})$ is also provided (see rows 6, 11, and 16). Performance is reported on $\mathcal{D}_{\text{overlap}}$ and $\mathcal{D}_{\text{disjoint}}$ where Fig. 1 visualises an mAP scatter plot between the two for 3D R50. + +smaller margins. Note that for the weakest model (2D R50) background mAP is even on-par or above the foreground baseline (see rows 2-3 in the final column). Together, these results illustrate powerfully, that across architectures background information alone is employable as a strong animal behaviour predictor. + +Result 2 - Multi-scale Vision Transformers rely less on backgrounds. On the overlapping dataset, the 3D-R50 and MViT-V2 models (rows 7-16) display significantly better performance when compared with the 2D-R50 (rows 2-6), and perform similarly to each other with respect to uAP and mAP. However, the MViT-V2 model widely outperforms the 3D R50 on the disjoint dataset, suggesting that it does not rely as heavily on background features for performance. In support of this, while all models experience a large decrease in performance when moving from the overlapping to the disjoint dataset (generalisation gap), the smallest impact is on the MViT-V2. These differences between architectures are highlighted most clearly by observing uAP on the disjoint dataset. Here, the background only ratio for the MViT-V2 model is $\sim 30\%$ and $\sim 13\%$ lower than the 2D-R50 and 3D-R50, respectively. MViT-V2 achieves the best overall performance and exhibits the lowest background-toForeground performance ratio on both dataset configurations, corroborating similar findings from + +the still image domain that show transformers are more invariant to changes in the background [39]. + +Result 3 - Analysing Background Information Contributions. The dual 3D-R50 model achieves significant performance gains over its FG-only baseline, particularly on $\mathcal{D}^{disjoint}$ , where increases of $7.92\%$ and $8.28\%$ in uAP and mAP, respectively, can be observed. Smaller, but still significant improvements are also observed on $\mathcal{D}^{overlap}$ ( $+3.62\%$ uAP and $+4.25\%$ mAP). Interestingly, the performance of the dual-stream 3D-R50 exceeds that of MViT-v2, including its dual-stream counterpart, on nearly all metrics except for disjoint uAP. The dual-stream MViT-V2 model, on the other hand, shows improved mAP but a lower uAP when compared with the baseline model, a trend consistent across both dataset configurations. This analysis indicates that BG-only performance in the wildlife domain is not a subset of the FG-only model – which of course also has access to background information. Instead, explicitly learned (and frozen) background features can aid the FG-only baseline in models such as ResNet via late fusion. Finally, all BG-only models trained on $\tilde{\mathcal{B}}$ seem to outperform those trained on $\mathcal{B}$ . However, we will show later in Sec. 5.4 that non-synthetic background videos dominate synthetic counterparts when explicitly utilised for performance improvements above the foreground baselines. + +# 5.3. Background Duration + +Experimental Overview. Noting that foreground videos may contain some frames with no animals, effectively supplying background only information for these frames, we aim at quantifying the effect of this on model performance (see Fig. 5). Practically, this is achieved by concatenating increasing numbers of background frames to our foreground videos. To do this, we sampled a set of $\lambda T$ frames from $v^{B}$ , where $\lambda$ is a predefined factor controlling the proportion of background frames used. The sampled background frames are denoted as $V^{\mathcal{B}} = \{V_{i_0}^{\mathcal{B}}, V_{i_1}^{\mathcal{B}}, \dots, V_{i_{\lambda T - 1}}^{\mathcal{B}}\}$ , with each index $i_j$ selected from $\{0, \dots, T - 1\}$ . The foreground frames $V^{\mathcal{F}} = \{F_0^{\mathcal{F}}, F_1^{\mathcal{F}}, \dots, F_{T - 1}^{\mathcal{F}}\}$ were then concatenated with $V^{\mathcal{B}}$ , resulting in a combined video sequence $V^{\mathrm{concat}} = \{F_0^{\mathcal{F}}, F_1^{\mathcal{F}}, \dots, F_{T - 1}^{\mathcal{F}}, F_{i_0}^{\mathcal{B}}, F_{i_1}^{\mathcal{B}}, \dots, F_{i_{\lambda T - 1}}^{\mathcal{B}}\}$ of length $T + \lambda T$ . From this concatenated sequence, a starting frame $F_{\mathrm{start}}$ was sampled at random from the indices $\{0, \dots, \lambda T\}$ , ensuring that $T$ consecutive frames could be sampled without exceeding the sequence bounds. The final sampled frame sequence $S = \{F_{\mathrm{start}}, F_{\mathrm{start + 1}}, \dots, F_{\mathrm{start + T - 1}}\}$ was then used as input for training while retaining the original label $y$ from the foreground video $v^{\mathcal{F}}$ . Background frames are appended because motion-triggered cameras are more likely to capture activity first, resulting in a more natural ordering. + +Analysing the Continuum Between Foreground and Background Videos. Fig. 5 shows the performance com + +![](images/1928c8b96a524535c5ff3fd1bd4dc1d253149a4a184dbaf9bfbc3513d74c4943.jpg) +$\mathcal{D}_{overlap}$ + +![](images/ca755ab717380fb21982308ca388d20690ed7146a490171c4deace5f5a0b0160.jpg) +$\mathcal{D}_{disjoint}$ + +![](images/704097270d42955ac054993dcc057dc994d23deb2be4a9b90c5729dcd00d8c16.jpg) +Figure 5. Effect of Increasing Background Duration on Performance. Comparison of 2D R50, 3D R50 and MViT-V2 model performance as a function of the simulated background duration $(\lambda)$ . The 3D-R50 is more sensitive to increasing background durations than MViT-V2. MViT-V2 is relatively robust to increasing background durations on the overlapping dataset, although it becomes highly sensitive when evaluated on OOD data. + +![](images/37f9a6acb610c604010d16e9bac2cec7f82ab8c14e47d98df28834911cba570e.jpg) + +parison between the 2D R50, 3D R50 and MViT-V2 models as background duration increases in videos $(\lambda)$ . As expected, Fig. 5 confirms that as $\lambda$ increases model performance declines towards background model performance (see Tab 1). On $D^{overlap}$ , the 3D-R50 and MViT-V2 models exhibit a sharp decline before beginning to converge. This occurs at $\lambda = 32$ and $\lambda = 64$ for the 3D-R50 and MViT-V2 models, respectively. Additionally, as shown by the interval between $\lambda = 0$ and $\lambda = 32$ , the uAP and mAP of 3D-R50 falls by $16.88\%$ and $14.32\%$ , respectively, while falling by only $7.23\%$ and $9.13\%$ for the MViT-V2 model. Earlier convergence to near BG-only performance and a more rapid rate of decline, show that 3D-R50 is more sensitive to increases in background durations than MViT-V2. Despite starting from lower performance, the 2D R50 model exhibits a more gradual, near-linear decline, indicating less sensitivity to changes in $\lambda$ . While the behaviour of the 2D-R50 and 3D-R50 models remain relatively consistent on $D^{disjoint}$ , the behaviour of MViT-V2 changes significantly. Its uAP declines more steeply, falling from $70.05\%$ to $48.29\%$ (a decrease of $22.21\%$ ) between $\lambda = 0$ and $\lambda = 32$ , where its performance converges. It is also more sensitive to smaller increases in background durations, falling from $70.50\%$ to $64.08\%$ (a decrease of $6.42\%$ ) between $\lambda = 0$ and $\lambda = 1.0$ , where a decrease of only $0.38\%$ is experienced on $D^{overlap}$ over the same interval. At higher values of $\lambda$ (i.e., $>40$ ) the performance drops below that of the 3D-R50 (although only once + +the background-only performance has been surpassed). Although uAP falls more steeply, mAP declines more slowly and performance above that of the BG-only model is maintained even at the highest values of $\lambda$ . This suggests that, although the MViT-V2 achieves higher baseline performance on the disjoint dataset, it is more sensitive to increases in background frames when evaluating uAP; a scenario where the model's advantage is drowned out by background information. + +# 5.4. Background Bias Mitigation + +Augmenting Performance via Explicit Background Neutralisation. In this final experimental section, we justify the need for paired, real-world background samples, as uniquely provided by our dataset. We show that these sample pairings are able to significantly improve model performance. To show this, we analyse the impact of background subtraction operations in both input and embedding spaces across each of the model architectures and dataset configurations. Two background-subtracted views of our dataset were created for input space experiments. This was done by performing pixel-wise subtraction of $\mathcal{B}$ and $\tilde{\mathcal{B}}$ from $\mathcal{F}$ . To perform algebraic background subtraction in latent space, we extracted the feature vectors of the foreground video pairs. As described in Sec. 5.2, we achieve this by processing the video tensors up to the last convolutional layer and averaging the tokens from the last hidden state in the ResNet and MViT-V2 models, respectively. The background-subtracted embedding is then computed as + +![](images/005b94d4eed3c5553ed3cdf803597d0099fc8bb2cf4e216c16501ad2ca7cba9a.jpg) +foreground x fx 3x 224x 224 + +![](images/da1cfbb492290cd4d4c63f0fe109f7ae51c23439a47baba6bccd1a8b419fb509.jpg) + +![](images/aa6d66bad7d7a778dbfa4ee378fbe25848f1cbc285034806f09f32d57a393ac6.jpg) +background (b x f x 3 x 224 x 224) +Figure 6. Latent Space Background Compensation. The proposed operation uses $\alpha$ to perform a weighted subtraction of background features (computed via the dataset pairing of a real-world, background video) from foreground features in latent space. The dotted line between the encoders indicates shared weights (i.e., they are the same model). + +$z^{\mathcal{F} - \mathcal{B}} = z^{\mathcal{F}} - (1 - \alpha)\cdot z^{\mathcal{B}}$ where $\alpha$ is a coefficient used to control the relative magnitude of the algebraic subtraction operation as schematically shown in Fig. 6. In practice, the modulation parameter $\alpha \in [0,1]$ was ablated to increase linearly or exponentially over the course of training and results are reported for both. In all cases $\alpha$ was initialized at zero and increased to one by the end of training. + +
ModelSubtraction\(D^{overlap}\)\(D^{disjoint}\)
uAPmAPuAPmAP
2D R50Baseline64.1541.2844.6718.08
\(\mathcal{F}+\mathcal{B}\)67.7444.9945.2818.84
\(\mathcal{F}-\mathcal{B}\)65.9642.1254.5326.65
\(\mathcal{F}-\mathcal{B}\)72.2147.1253.7325.84
None56.2138.8427.2417.94
Linear53.7839.0129.7517.17
Exponential54.0537.1125.8017.73
3D R50Baseline72.7949.3257.7430.89
\(\mathcal{F}+\mathcal{B}\)76.4153.5765.6639.17
\(\mathcal{F}-\mathcal{B}\)66.5142.5653.5827.98
\(\mathcal{F}-\mathcal{B}\)71.9346.3653.9225.05
None72.3051.3963.1434.85
Linear77.3152.8266.6936.31
Exponential73.5251.2464.5836.36
Baseline75.4451.5470.5035.29
\(\mathcal{F}+\mathcal{B}\)72.8253.4968.0038.09
MViT-V2\(\mathcal{F}-\mathcal{B}\)66.8442.2757.2128.80
\(\mathcal{F}-\mathcal{B}\)74.9049.4061.2231.51
None74.2344.1168.5635.14
Linear79.32*53.83*71.64*39.04
Exponential77.7050.9371.2839.25*
+ +Table 2. Quantifying Performance Improvements of Background Subtraction. Comparison of the 2D R50, 3D R50, and MViT-V2 models on both dataset configurations. Performance is reported when training on $\mathcal{F} - \mathcal{B}$ or $\mathcal{F} - \tilde{\mathcal{B}}$ videos in input (light green) and embedding space (light blue). First and second highest scores are indicated for each model, while the highest overall scores are marked by an asterisks $(\ast)$ . Note the baseline is trained on $\mathcal{F}$ , as reported in Tab. 1. $\mathcal{F} + \mathcal{B}$ provides an alternative baseline but is reliant on background videos at test time. Results confirm the superiority of real-world background videos over synthetic data, and of latent space neutralisation over input space background subtraction. + +Results and Performance Improvements. Tab. 2 shows that for the basic 2D-R50 model, background subtraction in input space is highly effective. When using $\mathcal{B}$ , we observe substantial performance improvements over the baseline, with percentage increases of $+8.06\%$ and $5.84\%$ in uAP and mAP on $D^{overlap}$ , and $+9.06\%$ and $+7.76\%$ in uAP and mAP on $D^{disjoint}$ . Using $\tilde{\mathcal{B}}$ also yields improvements, though these gains are notably smaller. Background subtraction in embedding space does not provide any benefit to the 2D-R50. Although input space background subtraction is effective for the 2D-R50, it does not improve performance for the 3D-R50 and MViT-V2 models. Neither original nor synthetic backgrounds provide an advantage over + +their respective baselines, indicating that input space subtraction may not suit spatio-temporal architectures. + +Conversely, background subtraction in embedding space provides significant performance improvements for the 3D-R50 and MViT-V2 models, particularly on $D^{disjoint}$ . For the 3D-R50, embedding space subtraction with linear scheduling results in a $+8.95\%$ and $+5.42\%$ increase in uAP and mAP, respectively. Exponential scheduling also yields improvements over the baseline. The improvements are lower than linear scheduling with respect to uAP $(+6.84\%)$ but marginally better for mAP $(+5.47\%)$ . MViT-V2 also benefits from embedding space subtraction on $D^{disjoint}$ : linear scheduling improves uAP and mAP by $+1.14\%$ and $+3.75\%$ , respectively, whereas exponential scheduling provides a $+0.78\%$ increase in uAP and a $+3.96\%$ increase in mAP. These findings suggest that embedding space subtraction is effective for both 3D-R50 and MViT-V2 models on $D^{disjoint}$ , with linear and exponential scheduling favouring uAP and mAP performance, respectively. Embedding space subtraction also enhances performance on $D^{overlap}$ , though to a lesser extent than on $D^{disjoint}$ . For the 3D-R50 model, applying embedding space subtraction with linear scheduling yields an improvement of $4.52\%$ and $3.50\%$ in uAP and mAP, respectively, while exponential scheduling results in smaller gains of $0.73\%$ and $1.92\%$ . The MViT-V2 model follows a similar trend: linear scheduling increases uAP and mAP by $3.88\%$ and $2.29\%$ , respectively, whereas exponential scheduling improves uAP by $2.26\%$ but slightly reduces mAP by $-0.61\%$ . These results suggest that algebraic embedding space subtraction is beneficial on $D^{overlap}$ , although the performance gains are relatively small compared to $D^{disjoint}$ , and that linear scheduling generally provides better results than exponential scheduling. + +# 6. Conclusion + +In this work, we introduced PanAf-FGBG, the first wildlife dataset designed to assess backgrounds and out-of-distribution generalization in behavior recognition. Our results demonstrate that real-world background information is a strong predictor of animal behavior and can significantly enhance recognition performance. While there are some limitations — such as its restriction to statically installed camera trap videos, the additional data and compute required for latent space background subtraction, and the potential for unnatural artifacts in background concatenation — the dataset remains a valuable resource. PanAf-FGBG enables direct evaluation under out-of-distribution conditions, and its rich metadata supports additional tasks and experimental setups. We hope it will help with understanding and improving model performance in this challenging domain and ultimately benefit endangered and charismatic species like the one featured in this paper. + +# Acknowledgements + +We thank the Pan African Programme: 'The Cultured Chimpanzee' team and its collaborators for allowing the use of their data for this paper. We thank Amelie Pettrich, Antonio Buzharevski, Eva Martinez Garcia, Ivana Kirchmair, Sebastian Schutte, Linda Gerlach and Fabina Haas. We also thank management and support staff across all sites; specifically Yasmin Moebius, Geoffrey Muhanguzi, Martha Robbins, Henk Eshuis, Sergio Marrocoli and John Hart. Thanks to the team at https://www.chimpandsee.org particularly Briana Harder, Anja Landsmann, Laura K. Lynn, Zuzana Macháčková, Heidi Pfund, Kristeena Sigler and Jane Wideness. The work that allowed for the collection of the dataset was funded by the Max Planck Society, Max Planck Society Innovation Fund, and Heinz L. Krekeler. In this respect we would like to thank: Ministre des Eaux et Forets, Ministère de l'Enseignement supérieur et de la Recherche scientifique in Côte d'Ivoire; Institut Congolais pour la Conservation de la Nature, Ministère de la Recherche Scientifique in Democratic Republic of Congo; Forestry Development Authority in Liberia; Direction Des Eaux Et Forets, Chasses Et Conservation Des Sols in Senegal; Makerere University Biological Field Station, Uganda National Council for Science and Technology, Uganda Wildlife Authority, National Forestry Authority in Uganda; National Institute for Forestry Development and Protected Area Management, Ministry of Agriculture and Forests, Ministry of Fisheries and Environment in Equatorial Guinea. This work was supported by the UKRI CDT in Interactive AI (grant EP/S022937/1). This work was in part supported by the US National Science Foundation Awards No. 2118240 "HDR Institute: Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning" and Award No. 2330423 and Natural Sciences and Engineering Research Council of Canada under Award No. 585136 for the "AI and Biodiversity Change (ABC) Global Center". + +# References + +[1] Pan African Programme: The Cultured Chimpanzee, 1999. Online; accessed 29 January 2014. 3 +[2] Analysing animal behaviour in wildlife videos using face detection and tracking. IEEE Proceedings-Vision, Image and Signal Processing, 153(3):305-312, 2006. 2 +[3] Mimi Arandjelovic, Colleen R Stephens, Maureen S McCarthy, Paula Dieguez, Emmie K Kalan, Nuria Maldonado, Christophe Boesch, and Hjalmar S Kuehl. Chimp&see: An online citizen science platform for large-scale, remote video camera trap annotation of chimpanzee behaviour, demography and individual identification. PeerJ Preprints, 2016. 4 +[4] Max Bain, Arsha Nagrani, Daniel Schofield, Sophie Berdugo, Joana Bessa, Jake Owen, Kimberly J Hockings, Tetsuro Matsuzawa, Misato Hayashi, Dora Biro, et al. Au + +tomated audiovisual behavior recognition in wild primates. Science advances, 7(46):eabi4883, 2021. 2 +[5] Sara Beery, Arushi Agarwal, Elijah Cole, and Vighnesh Birodkar. The iwildcam 2021 competition dataset. arXiv preprint arXiv:2105.03494, 2021. 3 +[6] Otto Brookes, Majid Mirmehdi, Hjalmar S. Kuhl, and Tilo Burghardt. Triple-stream deep metric learning of great ape behavioural actions. In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 294-302, 2023. 2 +[7] Otto Brookes, Majid Mirmehdi, Hjalmar Kuhl, and Tilo Burghardt. Chimpvlm: Ethogram-enhanced chimpanzee behaviour recognition. 2024. +[8] Otto Brookes, Majid Mirmethdi, Colleen Stephens, Samuel Angedakin, Katherine Corogenes, Dervla Dowd, Paula Dieguez, Thurston C Hicks, Sorrel Jones, Kevin Lee, et al. Panaf20k: a large video dataset for wild ape detection and behaviour recognition. International Journal of Computer Vision, pages 1-17, 2024. 2, 3 +[9] Jackie Chappell and Susannah KS Thorpe. The role of great ape behavioral ecology in one health: Implications for captive welfare and re-habilitation success. American journal of primatology, 84(4-5):e23328, 2022. 2 +[10] Jun Chen, Ming Hu, Darren J Coker, Michael L Berumen, Blair Costelloe, Sara Beery, Anna Rohrbach, and Mohamed Elhoseiny. Mammalnet: A large-scale video benchmark for mammal recognition and behavior understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13052-13061, 2023. 2, 3 +[11] Jinwoo Choi, Chen Gao, Joseph CE Messou, and Jia-Bin Huang. Why can't i dance in the mall? learning to mitigate scene bias in action recognition. Advances in Neural Information Processing Systems, 32, 2019. 2, 3, 4 +[12] Jihoon Chung, Yu Wu, and Olga Russakovsky. Enabling detailed action recognition evaluation through video dataset augmentation. Advances in Neural Information Processing Systems, 35:39020-39033, 2022. 2, 3, 4, 5 +[13] Fay E Clark. Great ape cognition and captive care: Can cognitive challenges enhance well-being? Applied Animal Behaviour Science, 135(1-2):1-12, 2011. 2 +[14] Fagner Cunha, Eulanda M dos Santos, Raimundo Barreto, and Juan G Colonna. Filtering empty camera trap images in embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2438-2446, 2021. 2 +[15] Alexander D'Amour, Katherine Heller, Dan Moldovan, Ben Adlam, Babak Alipanahi, Alex Beutel, Christina Chen, Jonathan Deaton, Jacob Eisenstein, Matthew D Hoffman, et al. Underspecification presents challenges for credibility in modern machine learning. Journal of Machine Learning Research, 23(226):1-61, 2022. 2 +[16] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 3 + +[17] Isla Duporge, Maksim Kholiavchenko, Roi Harel, Dan Rubenstein, Meg Crofoot, Tanya Berger-Wolf, Stephen Lee, Scott Wolf, Julie Barreau, Jenna Kline, et al. Baboonland dataset: Tracking primates in the wild and automating behaviour recognition from drone videos. arXiv preprint arXiv:2405.17698, 2024. 2 +[18] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2 +[19] Rohit Girdhar and Deva Ramanan. Cater: A diagnostic dataset for compositional actions and temporal reasoning. arXiv preprint arXiv:1910.04744, 2019. 3 +[20] Abhinav Gupta and Larry S Davis. Objects in action: An approach for combining action understanding and object perception. pages 1-8. IEEE, 2007. 3 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[22] Yun He, Soma Shirakabe, Yutaka Satoh, and Hirokatsu Kataoka. Human action recognition without human. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 11-17. Springer, 2016. 2, 3 +[23] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In Proceedings of the European conference on computer vision (ECCV), pages 771-787, 2018. 2, 4 +[24] Filip Ilic, Thomas Pock, and Richard P Wildes. Is appearance free action recognition possible? In European Conference on Computer Vision, pages 156-173. Springer, 2022. 2, 3 +[25] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 4 +[26] Maksim Kholiavchenko, Jenna Kline, Maksim Kukushkin, Otto Brookes, Sam Stevens, Isla Duporge, Alec Sheets, Reshma R Babu, Namrata Banerji, Elizabeth Campolongo, et al. Deep dive into kabr: a dataset for understanding ungulate behavior from in-situ drone video. Multimedia Tools and Applications, pages 1-20, 2024. 2 +[27] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 2, 3, 4 +[28] Hjalmar S Kuhl and Tilo Burghardt. Animal biometrics: quantifying and detecting phenotypic appearance. Trends in ecology & evolution, 28(7):432-441, 2013. 2 +[29] Haoxin Li, Yuan Liu, Hanwang Zhang, and Boyang Li. Mitigating and evaluating static bias of action representations in the background and the foreground. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 19911-19923, 2023. 2, 3 +[30] Yi Li and Nuno Vasconcelos. Repair: Removing representation bias by dataset resampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9572-9581, 2019. 2, 3 +[31] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 3 +[32] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4804-4814, 2022. 4 +[33] Dan Liu, Jin Hou, Shaoli Huang, Jing Liu, Yuxin He, Bochuan Zheng, Jifeng Ning, and Jingdong Zhang. Lote-animal: A long time-span dataset for endangered animal behavior understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20064-20075, 2023. 2, 3 +[34] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 3 +[35] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 4 +[36] Xiaoxuan Ma, Stephan Kaufhold, Jiajun Su, Wentao Zhu, Jack Terwilliger, Andres Meza, Yixin Zhu, Federico Rossano, and Yizhou Wang. Chimpact: A longitudinal dataset for understanding chimpanzee behaviors. Advances in Neural Information Processing Systems, 36:27501-27531, 2023. 2, 3 +[37] Marcin Marszalek, Ivan Laptev, and Cordelia Schmid. Actions in context. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 2929-2936. IEEE, 2009. 2, 3 +[38] Zhongqi Miao, Ziwei Liu, Kaitlyn M Gaynor, Meredith S Palmer, Stella X Yu, and Wayne M Getz. Iterative human and automated identification of wildlife images. Nature Machine Intelligence, 3(10):885-895, 2021. 2 +[39] Mazda Moayeri, Phillip Pope, Yogesh Balaji, and Soheil Feizi. A comprehensive study of image classification model sensitivity to foregrounds, backgrounds, and visual attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19087-19097, 2022. 2, 6 +[40] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19023-19034, 2022. 2, 3 +[41] Toby Perrett, Saptarshi Sinha, Tilo Burghardt, Majid Mirmethdi, and Dima Damen. Use your head: Improving long-tail video recognition. In Proceedings of the IEEE/CVF con + +ference on computer vision and pattern recognition, pages 2415-2425, 2023. 3 +[42] Laura J Pollock, Justin Kitzes, Sara Beery, Kaitlyn M Gaynor, Marta A Jarzyna, Oisin Mac Aodha, Bernd Meyer, David Rolnick, Graham W Taylor, Devis Tuia, et al. Harnessing artificial intelligence to fill global shortfalls in biodiversity knowledge. Nature Reviews Biodiversity, pages 1-17, 2025. 2 +[43] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2022. 2 +[44] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4 +[45] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 4 +[46] Faizaan Sakib and Tilo Burghardt. Visual recognition of great ape behaviours in the wild. In Workshop on the Visual Observation and Analysis of Vertebrate and Insect Behaviour, 2020. 2 +[47] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9661-9669, 2021. 2 +[48] Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank Van Langevelde, Tilo Burghardt, et al. Perspectives in machine learning for wildlife conservation. Nature communications, 13(1):1-15, 2022. 2 +[49] Tuan-Hung Vu, Catherine Olsson, Ivan Laptev, Aude Oliva, and Josef Sivic. Predicting actions from static scenes. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 421-436. Springer, 2014. 2, 3 +[50] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11804-11813, 2021. 3 +[51] Kai Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. arXiv preprint arXiv:2006.09994, 2020.3 +[52] Xinyu Yang, Majid Mirmehdi, and Tilo Burghardt. Great ape detection in challenging jungle camera trap footage via attention-based spatial and temporal feature blending. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 2 +[53] Xinyu Yang, Tilo Burghardt, and Majid Mirmehdi. Dynamic curriculum learning for great ape detection in the wild. In- + +ternational Journal of Computer Vision, 131(5):1163-1181, 2023. 2 +[54] Manlin Zhang, Jinpeng Wang, and Andy J Ma. Suppressing static visual cues via normalizing flows for self-supervised video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3300-3308, 2022. 3 +[55] Mingjun Zhao, Yakun Yu, Xiaoli Wang, Lei Yang, and Di Niu. Search-map-search: a frame selection paradigm for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10627-10636, 2023. 3 +[56] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 3 +[57] Yi Zhu, Xinyu Li, Chunhui Liu, Mohammadreza Zolfaghari, Yuanjun Xiong, Chongruo Wu, Zhi Zhang, Joseph Tighe, R Manmatha, and Mu Li. A comprehensive study of deep video action recognition. arXiv preprint arXiv:2012.06567, 2020. 2 \ No newline at end of file diff --git a/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/images.zip b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2a656c23a0441f2210c16b18b99413708c8013ba --- /dev/null +++ b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b69cce6cec5d0808325650b8e3711e86b5dd1c2146c2a114ce904d29c6a2d720 +size 657079 diff --git a/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/layout.json b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b6e757a85142d0d035b3d418047cfcd8f80cdb58 --- /dev/null +++ b/2025/The PanAf-FGBG Dataset_ Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition/layout.json @@ -0,0 +1,9216 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 220, + 83, + 391, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 83, + 391, + 99 + ], + "spans": [ + { + "bbox": [ + 220, + 83, + 391, + 99 + ], + "type": "text", + "content": "The PanAf-FGBG Dataset:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 102, + 544, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 544, + 118 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 544, + 118 + ], + "type": "text", + "content": "Understanding the Impact of Backgrounds in Wildlife Behaviour Recognition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "spans": [ + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": "Otto Brookes" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger*}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Maksim Kukushkin" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{3,4*}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Majid Mirmehdi" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Colleen Stephens" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Paula Dieguez" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Thurston C. Hicks" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Sorrel Jones" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Kevin Lee" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Maureen S. McCarthy" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Amelia Meier" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Emmanuelle Normand" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Erin G. Wessling" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Roman M. Wittig" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{5,8}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Kevin Langergraber" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Klaus Zuberbühler" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Lukas Boesch" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Thomas Schmid" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{3,11}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Mimi Arandjelovic" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Hjalmar Kuhl" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{5,12}" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "text", + "content": ", Tilo Burghardt" + }, + { + "bbox": [ + 70, + 122, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 201, + 506, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 201, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 102, + 201, + 506, + 257 + ], + "type": "text", + "content": "1University of Bristol, 2Wild Chimpanzee Foundation, 3Martin Luther University Halle-Wittenberg, 4Leipzig University, 5Max Planck Institute for Evolutionary Anthropology, 6University of Warsaw, 7Harvard University, 8University of Lyon, 9Arizona State University, 10University of St Andrews, 11Lancaster University in Leipzig, 12Senckenberg Museum of Natural History" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 284, + 200, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 284, + 200, + 296 + ], + "spans": [ + { + "bbox": [ + 151, + 284, + 200, + 296 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 311, + 297, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 311, + 297, + 670 + ], + "spans": [ + { + "bbox": [ + 53, + 311, + 297, + 670 + ], + "type": "text", + "content": "Computer vision analysis of camera trap video footage is essential for wildlife conservation, as captured behaviours offer some of the earliest indicators of changes in population health. Recently, several high-impact animal behaviour datasets and methods have been introduced to encourage their use; however, the role of behaviour-correlated background information and its significant effect on out-of-distribution generalisation remain unexplored. In response, we present the PanAf-FGBG dataset, featuring 21 hours of wild chimpanzee behaviours, recorded at 389 individual camera locations. Uniquely, it pairs every video with a chimpanzee (referred to as a foreground video) with a corresponding background video (with no chimpanzee) from the same camera location. We present two views of the dataset: one with overlapping camera locations and one with disjoint locations. This setup enables, for the first time, direct evaluation of in-distribution and out-of-distribution conditions, and for the impact of backgrounds on behaviour recognition models to be quantified. All clips have rich behavioural annotations and metadata, including unique camera IDs. Additionally, we establish several baselines and present a highly effective latent-space normalisation technique that boosts out-of-distribution performance by " + }, + { + "bbox": [ + 53, + 311, + 297, + 670 + ], + "type": "inline_equation", + "content": "+5.42\\%" + }, + { + "bbox": [ + 53, + 311, + 297, + 670 + ], + "type": "text", + "content": " mAP for convolutional and " + }, + { + "bbox": [ + 53, + 311, + 297, + 670 + ], + "type": "inline_equation", + "content": "+3.75\\%" + }, + { + "bbox": [ + 53, + 311, + 297, + 670 + ], + "type": "text", + "content": " mAP for transformer-based models. Finally, we provide an in-depth analysis on the role of backgrounds in out-of-distribution behaviour recognition, including the so far unexplored impact of background durations. (i.e., the count of background frames within foreground videos). The dataset is available at https://obrookes.github.io/panaf-fbgb.github.io/." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 284, + 555, + 572 + ], + "blocks": [ + { + "bbox": [ + 316, + 284, + 555, + 572 + ], + "lines": [ + { + "bbox": [ + 316, + 284, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 555, + 572 + ], + "type": "image", + "image_path": "d5f14db0bf10262b794a2615c63ddc6abd774994f8b9277f8062d80850ed8d50.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 581, + 556, + 681 + ], + "lines": [ + { + "bbox": [ + 313, + 581, + 556, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 556, + 681 + ], + "type": "text", + "content": "Figure 1. Conceptual Overview. The PanAf-FGBG dataset comprises " + }, + { + "bbox": [ + 313, + 581, + 556, + 681 + ], + "type": "inline_equation", + "content": ">20" + }, + { + "bbox": [ + 313, + 581, + 556, + 681 + ], + "type": "text", + "content": " hours of paired and richly annotated foreground-background camera trap videos of wild chimpanzees. The dataset unlocks systematic analyses of the impact of background information on wildlife behaviour recognition. We provide baselines, quantify the background impact on performance, and demonstrate that utilising background information in latent space (see blue data point) can significantly improve recognition performance in this challenging domain." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 246, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 246, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 246, + 712 + ], + "type": "text", + "content": "†Corresponding author, *Equal technical contribution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5433" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 186 + ], + "type": "text", + "content": "Deep learning models have shown impressive results on action recognition tasks [57], particularly when both training and test data share the same distribution. However, their performance declines significantly on out-of-distribution (OOD) data [15, 27]. In practice, distribution shifts are common in nearly all settings [43], but are particularly unforgiving when computer vision is applied to uncontrolled, complex or natural environments [27]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 187, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 295, + 342 + ], + "type": "text", + "content": "In the human action recognition domain, poor OOD performance is well-known to originate from shortcut learning [18] of static cues or backgrounds [24, 29, 30]. This leads models to rely heavily on background information instead of the behaviour of interest solely [2], and several studies have shown that the majority of action recognition performance in this domain can be attributed to background learning [12, 22, 49]. Despite the known impact of backgrounds on action recognition and a critical need to advance model performance in wildlife applications [28, 42, 48], reliance on static cues for animal behaviour recognition has received limited attention. This is largely due to a lack of datasets explicitly designed to study its effects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 510 + ], + "type": "text", + "content": "In response, we introduce the PanAf-FGBG dataset, which comprises real-world foreground-background video pairs (see Fig. 1) extracted from six African countries, 14 national parks, and over 350 individual camera locations across many habitats and location types (see Fig. 3). Here, we define foreground videos as those containing a chimpanzee, and background videos as those that do not. This allows us to address a key limitation in existing studies, where the background's role is typically assessed by creating background-only versions of datasets through synthetic modifications (e.g., masking or in-painting) that remove the actor [11, 12, 23]. While such synthetic inputs often serve as useful proxies, the actual effect of synthesis on training is largely unknown [39, 47]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "In contrast, our dataset introduces paired camera locations and enables a truly realistic disentangling of background effects under in- and out-of-distribution conditions: disjoint experiments with mutually exclusive camera locations for training and test splits, and overlapping experiments with shared locations. This focus is particularly important in the wildlife domain, where the transferability and generalisation of camera trap analysis to new locations is critical for effective operation. While studies have shown the over-reliance of action recognition models on backgrounds [22, 37, 49], we find that the effect of background duration (i.e., the count of background frames within foreground videos with no actor) is largely unexplored. In camera trap video footage like in many other CCTV or triggered camera settings, actors may only appear momentarily, leaving behind footage of only the background [14, 38] within foreground videos. Since this effect alters learning given" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "video-level behaviour annotations, we quantify its effect on several action recognition architectures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 96, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 276 + ], + "type": "text", + "content": "Our contributions are as follows: (i) we release PanAf-FGBG with its rich annotations and meta-data, providing a large wildlife dataset suitable for explicitly investigating the impact of backgrounds on behaviour recognition; (ii) we show models trained on background-only videos achieve strong animal behaviour recognition performance (" + }, + { + "bbox": [ + 313, + 96, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\sim 65\\%" + }, + { + "bbox": [ + 313, + 96, + 553, + 276 + ], + "type": "text", + "content": " of baseline performance), corroborating findings in the human domain; (iii) we conduct the first evaluation of background duration in videos, revealing its impact on action recognition models and highlighting key differences between convolutional and attention-based networks; and (iv) we introduce a simple yet powerful latent-space background neutralisation that yields notable performance gains, particularly in OOD scenarios for convolutional " + }, + { + "bbox": [ + 313, + 96, + 553, + 276 + ], + "type": "inline_equation", + "content": "(+5.42\\%)" + }, + { + "bbox": [ + 313, + 96, + 553, + 276 + ], + "type": "text", + "content": " mAP) and transformer-based approaches " + }, + { + "bbox": [ + 313, + 96, + 553, + 276 + ], + "type": "inline_equation", + "content": "(+3.75\\%)" + }, + { + "bbox": [ + 313, + 96, + 553, + 276 + ], + "type": "text", + "content": " mAP)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 286, + 400, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 400, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 400, + 298 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 307, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 557 + ], + "type": "text", + "content": "Animal Behaviour Datasets. In recent years, several high-impact animal behaviour datasets have been introduced to the CV community, often accompanied by new methods [4, 6-8, 10, 26, 33, 40, 46]. Although they all centre on animal behaviour understanding, they each have a different focus. The large-scale datasets Animal Kingdom [40] and MammalNet [10] concentrate on behavioural understanding across species, each including videos of " + }, + { + "bbox": [ + 313, + 307, + 553, + 557 + ], + "type": "inline_equation", + "content": "\\sim 800" + }, + { + "bbox": [ + 313, + 307, + 553, + 557 + ], + "type": "text", + "content": " different species. They provide atomic action and high-level behaviour annotations, respectively, and support, in addition to action recognition, the related tasks of action localisation [10, 40] and pose estimation [40]. The smaller ChimpAct dataset [36] provides longitudinal data that follows a zoo-housed group of 20 chimpanzees with a focus on a single male juvenile. It also contains key-point pose information, and fine-grained spatio-temporal behaviour labels. While valuable, these datasets lack in-situ footage from ecological sensors like camera traps and drones, meaning animals are not observed in their natural environment. This may result in unrepresentative behaviour distributions and limit their suitability for studying wild behaviour [9, 13]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "In contrast, KABR [26] and BaboonLand [17] comprise drone footage of free-living Kenyan wildlife, with fine-grained spatio-temporal behaviour annotations, focussing on individual and group-level behaviour, respectively. The LoTE dataset [33] provides camera trap video footage of 11 endangered species from a 12-year longitudinal study; it is accompanied by environmental information (i.e., weather conditions and habitat) and highlights the importance of temporal and spatial continuity for conservation efforts. The dataset most comparable to our own is the PanAf20k dataset [8] and its related PanAf500-based works [46, 52, 53]. PanAf20k comprises camera trap videos of great apes accompanied by multi-label behaviour annotations." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5434" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 276 + ], + "type": "text", + "content": "Our dataset orthogonally complements this growing body of work dedicated to understanding animal behaviour by providing the first dataset specifically designed to evaluate backgrounds and out-of-distribution generalisation for behaviour recognition. It is unique in providing foreground-background pairs, and its configurations for in- and out-of-distribution evaluation are defined identically to the well-known iWildCam [5] and Wilds dataset [27] for species classification. Similar to LotE [33], we provide a detailed description of the background, however we also include a unique camera ID (i.e., geospatial location) to allow footage taken from the same camera to be linked. Similar to [8], our data is gathered from 14 national parks, spanning 6 different countries, although unlike ours they do not include metadata to allow videos to be linked to specific countries, research sites, locations, or habitats. Similarly, it is more diverse than [33] with its footage from a single national park." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 445 + ], + "type": "text", + "content": "Background Impact. The effect of backgrounds or static bias is well-studied within computer vision [29, 54], where it has shown to be necessary for models to perform well [51]. Action recognition models are particularly vulnerable to shortcut learning. These models often rely on correlated, but distinct information beyond the intended action, such as object [20, 56], scene [37], and actor biases [31]. For example, several studies have reported that strong performance on action recognition tasks can be achieved without the actor [22, 37, 49] and that approximately " + }, + { + "bbox": [ + 55, + 278, + 296, + 445 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 278, + 296, + 445 + ], + "type": "text", + "content": " of model performance is attributable to static cues (i.e., scene or background) [12]. This results in biased representations and ultimately impairs out-of-distribution generalization performance [30]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 448, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 592 + ], + "type": "text", + "content": "Background Bias Datasets. The effect of the background has often been modelled using synthetically generated data that removes either the foreground or the background. [11, 12, 19, 29]. For instance, Ilic et al. [24] present a synthetic version of the UCF101 dataset, animating spatial noise using image motion extracted by an optical flow estimator, Li et al. [29] replace the backgrounds of IID HMDB51 with sinusoidal stripe images, and Chung et al [12] adopt an in-painting model to remove actors from the video. Datasets that include fine-grained spatio-temporal action annotations also remove static cues implicitly by isolating the region of the video where the action is occurring." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Mitigating Background Bias. Prominent attempts to mitigate this bias include dataset resampling methods, which down-weight biased samples [30, 31], adversarial training regimes that rely on scene labels and human masks [11], and augmentation techniques that mix or swap the backgrounds of different videos during training [16, 29, 50]. There are also several implicit approaches, such as frame selection, that aim to utilise only frames where activity is occurring [55]. Our dataset, introduced next, instead introduces foreground-background pairing in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "formation with close proximity in time of capture, which describes the environment truly independently and in relation to the presence of animals showcasing behaviour." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 118, + 369, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 369, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 369, + 131 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 134, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 555, + 326 + ], + "type": "text", + "content": "Dataset Overview and Statistics. PanAf-FGBG comprises footage gathered under ethical oversight as part of the PanAf Programme: The Cultured Chimpanzee [1]. It contains 21 hours of camera trap footage of individual chimpanzees in tropical Africa. Its footage is collected from 389 individual camera locations across 14 national parks in 6 African countries. In total, we provide 5,070 video pairs, each 15 seconds in duration. All videos are accompanied by a multi-label behaviour annotation and scene metadata, plus the unique feature of foreground-background video pairing (see Fig. 1). Class and habitat distributions of the data are given in Fig. 2 and Fig. 3, respectively. Our dataset follows a long-tailed distribution [34, 41] common in animal behaviour datasets [8, 10, 33, 36, 40]. The three of the most commonly occurring classes are observed in " + }, + { + "bbox": [ + 313, + 134, + 555, + 326 + ], + "type": "inline_equation", + "content": ">60\\%" + }, + { + "bbox": [ + 313, + 134, + 555, + 326 + ], + "type": "text", + "content": " of videos, while the rarest are observed in fewer than " + }, + { + "bbox": [ + 313, + 134, + 555, + 326 + ], + "type": "inline_equation", + "content": "< 3\\%" + }, + { + "bbox": [ + 313, + 134, + 555, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 327, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 469 + ], + "type": "text", + "content": "Data Collection. The data collection strategy was designed specifically to maximise the chance of filming the terrestrial activity of apes and we provide only a brief description here (for full details we refer the reader to [1]). Data was collected from 14 different research sites, where an average of 29 movement-triggered Bushnell cameras were installed per site. Grids comprising 20 to " + }, + { + "bbox": [ + 313, + 327, + 556, + 469 + ], + "type": "inline_equation", + "content": "961 \\times 1" + }, + { + "bbox": [ + 313, + 327, + 556, + 469 + ], + "type": "text", + "content": " km cells were established and one camera was installed per grid cell. Footage was recorded at 30 frames-per-second and a resolution of " + }, + { + "bbox": [ + 313, + 327, + 556, + 469 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 313, + 327, + 556, + 469 + ], + "type": "text", + "content": ". The cameras were visited every 1-3 months for maintenance and to download the recorded footage throughout the study periods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 469, + 556, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 518 + ], + "type": "text", + "content": "Ethical Considerations. To protect the location of the apes from threats (e.g., poaching) we use monikers for the names of the research sites and assign a unique camera ID to each video, hiding the geospatial location of the cameras." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 325, + 536, + 545, + 679 + ], + "blocks": [ + { + "bbox": [ + 325, + 536, + 545, + 679 + ], + "lines": [ + { + "bbox": [ + 325, + 536, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 325, + 536, + 545, + 679 + ], + "type": "image", + "image_path": "69e55246557040f4101ce07536dce39f1e1a9b9b869b1e47ce1baeae86eec24b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 688, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 688, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 555, + 711 + ], + "type": "text", + "content": "Figure 2. Distribution of Behaviour. Proportion of behaviours in the dataset where smaller segments are colour coordinated." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5435" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 77, + 291, + 312 + ], + "blocks": [ + { + "bbox": [ + 59, + 77, + 291, + 312 + ], + "lines": [ + { + "bbox": [ + 59, + 77, + 291, + 312 + ], + "spans": [ + { + "bbox": [ + 59, + 77, + 291, + 312 + ], + "type": "image", + "image_path": "044ebbe43aa5fbd39b131aa99a39f6bd76127a9a1f2ce571bfa8a12ac2ccea3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 323, + 297, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 297, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 297, + 369 + ], + "type": "text", + "content": "Figure 3. Distribution of Habitat and Location Type. The pie chart gives the proportion of videos taken at each of the three main habitats: forest, marshes and savannah. The histogram shows the total number of videos extracted from each location type." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 390, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 295, + 605 + ], + "type": "text", + "content": "Data Annotation. The behavioural annotations were provided by users on the community science platform Chimp&See [3]. Annotators were presented with a choice of behaviour classification categories, chosen specifically by experts for their ecological importance. Behaviours not listed in the classification categories were also permitted. These annotations were then extracted and expertly grouped into 14 co-occurring classes, which form the multi-label behavioural annotations presented here. The annotations follow a multi-hot binary format that indicate the presence of one or more behaviours. It should also be noted that behaviours are not assigned to individual apes or temporally localised within each video. To ensure annotation quality and consistency a video was only deemed to be analysed when either three volunteers marked the video as blank, or unanimous agreement between seven volunteers was observed, or 15 volunteers annotated the video. Habitat type was noted during camera setup." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": "Foreground-Background Pair Selection. Pairs are formed by selecting the temporally closest available foreground and background videos to ensure similarity. Complete standardisation statistics can be found in Sup. Mat." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Overlapping & Disjoint Configurations. We present both a disjoint view of dataset, where camera locations in the training and test splits are mutually exclusive, denoted as " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{disjoint} = \\{d_i \\mid d_i \\cap d_j = \\emptyset, \\forall i \\neq j\\}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " following [27]. A second, overlapping view is also provided, where camera" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "locations are shared between training and test splits, represented as " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{overlap} = \\{d_i \\mid d_i \\cap d_j \\neq \\emptyset" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " for some " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "i \\neq j\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " refers to sets of camera locations associated with videos in each split. Fig. 4 provides an overview of both configurations. We ensure that the class distribution in each configuration remains approximately the same for full comparability. Unique camera identifiers are available for each video to allow researchers interested in this problem to create additional configurations, should they wish to." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 198, + 399, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 399, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 399, + 211 + ], + "type": "text", + "content": "4. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": "We train multi-label classification models on both views of the dataset: " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{overlap}}" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{disjoint}}" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": " using various combinations of foreground " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": ", background " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "(\\mathcal{B})" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": " and synthetic background " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathcal{B}})" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": " videos. The latter were created by generating a segmentation mask using SAM2 [44] followed by mean pixel value filling, similar to [11, 12, 23]. Both views comprise foreground-background video pairs " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "\\{(v_{i}^{\\mathcal{F}}, v_{i}^{\\mathcal{B}})\\}_{i=1}^{N}" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "v_{i}^{\\mathcal{F}} \\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "v_{i}^{\\mathcal{B}} \\in \\mathcal{B}" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": " representing the total number of video pairs. Each foreground video " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "v_{i}^{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": " is associated with a video-level label vector " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "y_{i} = (y_{i1}, y_{i2}, \\ldots, y_{ic})" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "y_{ij} \\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": " indicates the presence (1) or absence (0) of the " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": "-th class in video " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 221, + 556, + 376 + ], + "type": "text", + "content": " is the total number of classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 396, + 395, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 395, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 395, + 410 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 418, + 365, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 365, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 365, + 431 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": "All experiments were performed using the ResNet-50 [21] and MViT-V2 [32] architectures, which were initialised with feature extractors pre-trained on Kinetics-400 [25]. For ResNet-50, both 2D and 3D convolution versions are tested. After initial experiments to ensure optimal optimiser-network pairings for the task, ResNet and MViT-V2 models were fine-tuned for 300 epochs using the SGD [45] and AdamW [35] optimisers, respectively. Both models followed a linear warm-up schedule with subsequent cosine annealing. For ResNet models, the learning rate was increased from " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "1\\times 10^{-5}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "1\\times 10^{-4}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " over 20 epochs, while it was increased from " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "1\\times 10^{-6}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "1\\times 10^{-4}" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " over 30 epochs for MViT-V2 models. In both cases, momentum of 0.9 was utilised. All computation was performed on " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " NVIDIA H200 GPUs using a distributed batch size of 336. During training, video tensors with a spatio-temporal resolution of " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "16\\times 256\\times 256" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " were employed. Frames were sampled uniformly at a rate of " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "1/24" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": ", avoiding dense sampling strategies which are shown to exacerbate shortcut learning in action recognition models [12]. For MViT-V2 models, a patch size of " + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "inline_equation", + "content": "16\\times 16" + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": " pixels was utilised. Evaluation was performed using micro and mean Average Precision and are referred to as uAP and mAP, respectively." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5436" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 305 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 305 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 305 + ], + "type": "image", + "image_path": "02d4d3d3405d48170cb7f18b37bf5f1e938866b7e0a4d83fcc726664f2caae48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 313, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 369 + ], + "type": "text", + "content": "Figure 4. Overlapping & Disjoint Dataset Configurations. Six foreground-background video pairs are shown to visualise overlapping (left) and disjoint (right) dataset configurations. On the left, train and test videos taken from the same camera are selected, emphasizing that the locations are not mutually exclusive, with backgrounds containing several cues that are highly correlated with the displayed behaviours (i.e., algae fishing, a form of tool use). Train and test videos on the right are extracted from mutually exclusive or disjoint camera locations. They are extracted from forest and savannah, respectively, and highlight the challenge of generalising to new locations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 389, + 178, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 178, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 178, + 402 + ], + "type": "text", + "content": "5.2. Background Reliance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": "Experimental Overview. First, we examined the reliance of animal behaviour recognition on the background by training model architectures on the foreground " + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": ", background " + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": ", and synthetic background videos " + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}}" + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": " only. We refer to models trained on " + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}}" + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": ") as FG-only and BG-only models, respectively, as shown earlier in Fig. 1. To separate background information from class-distribution intrinsic information, a dummy classifier with no access to visual information was also evaluated. It would effectively return predictions based on the class distribution information alone. Since our dataset consists of unique foreground-background pairs, it was possible to train BG-only models by retaining the original labels associated with the foreground video of the pairing. In all cases, models were tested on " + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": " for comparability. Note that this makes the task more difficult for BG-only models, since animals are only observed at test time and not during training. All experiments are conducted on both " + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "inline_equation", + "content": "D^{overlap}" + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "inline_equation", + "content": "D^{disjoint}" + }, + { + "bbox": [ + 54, + 411, + 295, + 625 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Dual-Stream Fusion Model " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{F} + \\mathcal{B})" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": ". To clarify whether BG-only models and FG-only models, each trained separately, can be fused late to orthogonally contribute towards performance we also implemented a dual-stream model. To do this, the independently optimised weights of the FG-only and BG-only models trained for 300 epochs were loaded. During training the weights were frozen and video pairs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 389, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 462 + ], + "type": "text", + "content": "were processed to extract late features " + }, + { + "bbox": [ + 313, + 389, + 555, + 462 + ], + "type": "inline_equation", + "content": "z^{F}" + }, + { + "bbox": [ + 313, + 389, + 555, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 389, + 555, + 462 + ], + "type": "inline_equation", + "content": "z^{B}" + }, + { + "bbox": [ + 313, + 389, + 555, + 462 + ], + "type": "text", + "content": ". For ResNet models, features were extracted from the last convolutional layer and in MViT-V2 models an average of the tokens from the last hidden state were taken. Late features were then concatenated " + }, + { + "bbox": [ + 313, + 389, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\tilde{z} = [z^{F}, z^{B}]" + }, + { + "bbox": [ + 313, + 389, + 555, + 462 + ], + "type": "text", + "content": " and final class predictions were produced by a trainable multi-layer perceptron." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": "Result 1 - Backgrounds are a Predictor of Animal Behaviour. Table 1 shows that BG-only models achieve strong performance when compared to their respective FG-only counterparts. To illustrate this, the ratio of background-toForeground performance " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\frac{AP^B}{AP^F}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " published in [12], showing how close a model can get to FG performance with just BG cues, is reported for each architecture in rows 6, 11, and 16. Background-to foreground performance ratios are consistently high across all models, never dropping below 0.70 and 0.65 for uAP and mAP, respectively, on either dataset. This indicates that much of the behaviour recognition performance is indeed achievable by just utilisation of background information - the environment is a strong predictor of behaviour, even out-of-distribution. Thus, features of the environment are already behaviour predictors (trees for climbing, paths for travel, termite mound for feeding etc.). This result is further compounded by the fact that BG-only models outperform the non-visual classifier (row 1) by a significant margin, particularly on " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "D^{overlap}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": ". On " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "D^{disjoint}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " this observation still holds, although by somewhat" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5437" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 289, + 262 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 289, + 262 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 289, + 262 + ], + "type": "table", + "html": "
ModelData\\(D^{overlap}\\)\\(D^{disjoint}\\)
uAPmAPuAPmAP
1Dummy-19.1514.7319.6114.87
2\\(\\bar{\\mathcal{F}}\\)64.1541.2844.6718.08
32D R50\\(\\mathcal{B}\\)57.1330.0345.8018.27
4\\(\\tilde{\\mathcal{B}}\\)57.8730.8248.2820.39
5\\(\\mathcal{F}+\\mathcal{B}\\)67.7444.9945.2818.84
60.890.741.020.98
7\\(\\bar{\\mathcal{F}}\\)72.7949.3257.7430.89
83D R50\\(\\mathcal{B}\\)58.9231.7248.4021.38
9\\(\\tilde{\\mathcal{B}}\\)59.6032.8649.2921.64
10\\(\\mathcal{F}+\\mathcal{B}\\)76.4153.5765.6639.17
110.800.660.830.69
12\\(\\bar{\\mathcal{F}}\\)75.4451.5470.5035.29
13MViT-V2\\(\\mathcal{B}\\)61.9032.7450.0324.00
14\\(\\tilde{\\mathcal{B}}\\)62.1533.5852.1324.77
15\\(\\mathcal{F}+\\mathcal{B}\\)72.8253.4968.0038.09
160.820.650.700.68
", + "image_path": "6758e7e6dccf12eb4e99011380a09492f11fe2acadf6f56b4deb13195f48518b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "text", + "content": "Table 1. Quantifying Background Reliance. Performance comparison of models trained on " + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "text", + "content": " videos. Dual-stream models are indicated by " + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{F} + \\mathcal{B}" + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "text", + "content": ". The absolute highest performance is indicated in bold whereas best performance using only a single-stream is underlined. The performance based on class distribution only is given by a dummy classifier (see row 1). The background-toForeground performance ratio " + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "inline_equation", + "content": "(\\frac{AP^B}{AP^F})" + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "text", + "content": " is also provided (see rows 6, 11, and 16). Performance is reported on " + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{overlap}}" + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{disjoint}}" + }, + { + "bbox": [ + 55, + 270, + 298, + 392 + ], + "type": "text", + "content": " where Fig. 1 visualises an mAP scatter plot between the two for 3D R50." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "text", + "content": "smaller margins. Note that for the weakest model (2D R50) background mAP is even on-par or above the foreground baseline (see rows 2-3 in the final column). Together, these results illustrate powerfully, that across architectures background information alone is employable as a strong animal behaviour predictor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Result 2 - Multi-scale Vision Transformers rely less on backgrounds. On the overlapping dataset, the 3D-R50 and MViT-V2 models (rows 7-16) display significantly better performance when compared with the 2D-R50 (rows 2-6), and perform similarly to each other with respect to uAP and mAP. However, the MViT-V2 model widely outperforms the 3D R50 on the disjoint dataset, suggesting that it does not rely as heavily on background features for performance. In support of this, while all models experience a large decrease in performance when moving from the overlapping to the disjoint dataset (generalisation gap), the smallest impact is on the MViT-V2. These differences between architectures are highlighted most clearly by observing uAP on the disjoint dataset. Here, the background only ratio for the MViT-V2 model is " + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\sim 30\\%" + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\sim 13\\%" + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": " lower than the 2D-R50 and 3D-R50, respectively. MViT-V2 achieves the best overall performance and exhibits the lowest background-toForeground performance ratio on both dataset configurations, corroborating similar findings from" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "the still image domain that show transformers are more invariant to changes in the background [39]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "content": "Result 3 - Analysing Background Information Contributions. The dual 3D-R50 model achieves significant performance gains over its FG-only baseline, particularly on " + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{disjoint}" + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "content": ", where increases of " + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "inline_equation", + "content": "7.92\\%" + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "inline_equation", + "content": "8.28\\%" + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "content": " in uAP and mAP, respectively, can be observed. Smaller, but still significant improvements are also observed on " + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{overlap}" + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "inline_equation", + "content": "+3.62\\%" + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "content": " uAP and " + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "inline_equation", + "content": "+4.25\\%" + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "content": " mAP). Interestingly, the performance of the dual-stream 3D-R50 exceeds that of MViT-v2, including its dual-stream counterpart, on nearly all metrics except for disjoint uAP. The dual-stream MViT-V2 model, on the other hand, shows improved mAP but a lower uAP when compared with the baseline model, a trend consistent across both dataset configurations. This analysis indicates that BG-only performance in the wildlife domain is not a subset of the FG-only model – which of course also has access to background information. Instead, explicitly learned (and frozen) background features can aid the FG-only baseline in models such as ResNet via late fusion. Finally, all BG-only models trained on " + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "content": " seem to outperform those trained on " + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 98, + 555, + 373 + ], + "type": "text", + "content": ". However, we will show later in Sec. 5.4 that non-synthetic background videos dominate synthetic counterparts when explicitly utilised for performance improvements above the foreground baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 383, + 441, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 383, + 441, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 383, + 441, + 396 + ], + "type": "text", + "content": "5.3. Background Duration" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": "Experimental Overview. Noting that foreground videos may contain some frames with no animals, effectively supplying background only information for these frames, we aim at quantifying the effect of this on model performance (see Fig. 5). Practically, this is achieved by concatenating increasing numbers of background frames to our foreground videos. To do this, we sampled a set of " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\lambda T" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": " frames from " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "v^{B}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": " is a predefined factor controlling the proportion of background frames used. The sampled background frames are denoted as " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "V^{\\mathcal{B}} = \\{V_{i_0}^{\\mathcal{B}}, V_{i_1}^{\\mathcal{B}}, \\dots, V_{i_{\\lambda T - 1}}^{\\mathcal{B}}\\}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": ", with each index " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "i_j" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": " selected from " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\{0, \\dots, T - 1\\}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": ". The foreground frames " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "V^{\\mathcal{F}} = \\{F_0^{\\mathcal{F}}, F_1^{\\mathcal{F}}, \\dots, F_{T - 1}^{\\mathcal{F}}\\}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": " were then concatenated with " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "V^{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": ", resulting in a combined video sequence " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{concat}} = \\{F_0^{\\mathcal{F}}, F_1^{\\mathcal{F}}, \\dots, F_{T - 1}^{\\mathcal{F}}, F_{i_0}^{\\mathcal{B}}, F_{i_1}^{\\mathcal{B}}, \\dots, F_{i_{\\lambda T - 1}}^{\\mathcal{B}}\\}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "T + \\lambda T" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": ". From this concatenated sequence, a starting frame " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{start}}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": " was sampled at random from the indices " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\{0, \\dots, \\lambda T\\}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": ", ensuring that " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": " consecutive frames could be sampled without exceeding the sequence bounds. The final sampled frame sequence " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "S = \\{F_{\\mathrm{start}}, F_{\\mathrm{start + 1}}, \\dots, F_{\\mathrm{start + T - 1}}\\}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": " was then used as input for training while retaining the original label " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": " from the foreground video " + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "inline_equation", + "content": "v^{\\mathcal{F}}" + }, + { + "bbox": [ + 313, + 401, + 555, + 689 + ], + "type": "text", + "content": ". Background frames are appended because motion-triggered cameras are more likely to capture activity first, resulting in a more natural ordering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Analysing the Continuum Between Foreground and Background Videos. Fig. 5 shows the performance com" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5438" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 85, + 176, + 175 + ], + "blocks": [ + { + "bbox": [ + 110, + 73, + 143, + 84 + ], + "lines": [ + { + "bbox": [ + 110, + 73, + 143, + 84 + ], + "spans": [ + { + "bbox": [ + 110, + 73, + 143, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{overlap}" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 85, + 176, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 85, + 176, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 85, + 176, + 175 + ], + "type": "image", + "image_path": "1928c8b96a524535c5ff3fd1bd4dc1d253149a4a184dbaf9bfbc3513d74c4943.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 175, + 85, + 294, + 174 + ], + "blocks": [ + { + "bbox": [ + 223, + 73, + 258, + 84 + ], + "lines": [ + { + "bbox": [ + 223, + 73, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 223, + 73, + 258, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{disjoint}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 85, + 294, + 174 + ], + "lines": [ + { + "bbox": [ + 175, + 85, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 175, + 85, + 294, + 174 + ], + "type": "image", + "image_path": "ca755ab717380fb21982308ca388d20690ed7146a490171c4deace5f5a0b0160.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 175, + 176, + 265 + ], + "blocks": [ + { + "bbox": [ + 56, + 175, + 176, + 265 + ], + "lines": [ + { + "bbox": [ + 56, + 175, + 176, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 176, + 265 + ], + "type": "image", + "image_path": "704097270d42955ac054993dcc057dc994d23deb2be4a9b90c5729dcd00d8c16.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 274, + 295, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 352 + ], + "type": "text", + "content": "Figure 5. Effect of Increasing Background Duration on Performance. Comparison of 2D R50, 3D R50 and MViT-V2 model performance as a function of the simulated background duration " + }, + { + "bbox": [ + 55, + 274, + 295, + 352 + ], + "type": "inline_equation", + "content": "(\\lambda)" + }, + { + "bbox": [ + 55, + 274, + 295, + 352 + ], + "type": "text", + "content": ". The 3D-R50 is more sensitive to increasing background durations than MViT-V2. MViT-V2 is relatively robust to increasing background durations on the overlapping dataset, although it becomes highly sensitive when evaluated on OOD data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 175, + 294, + 265 + ], + "blocks": [ + { + "bbox": [ + 175, + 175, + 294, + 265 + ], + "lines": [ + { + "bbox": [ + 175, + 175, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 175, + 175, + 294, + 265 + ], + "type": "image", + "image_path": "37f9a6acb610c604010d16e9bac2cec7f82ab8c14e47d98df28834911cba570e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": "parison between the 2D R50, 3D R50 and MViT-V2 models as background duration increases in videos " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\lambda)" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ". As expected, Fig. 5 confirms that as " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " increases model performance declines towards background model performance (see Tab 1). On " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "D^{overlap}" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", the 3D-R50 and MViT-V2 models exhibit a sharp decline before beginning to converge. This occurs at " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 32" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 64" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " for the 3D-R50 and MViT-V2 models, respectively. Additionally, as shown by the interval between " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 32" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", the uAP and mAP of 3D-R50 falls by " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "16.88\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "14.32\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", respectively, while falling by only " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "7.23\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "9.13\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " for the MViT-V2 model. Earlier convergence to near BG-only performance and a more rapid rate of decline, show that 3D-R50 is more sensitive to increases in background durations than MViT-V2. Despite starting from lower performance, the 2D R50 model exhibits a more gradual, near-linear decline, indicating less sensitivity to changes in " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ". While the behaviour of the 2D-R50 and 3D-R50 models remain relatively consistent on " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "D^{disjoint}" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", the behaviour of MViT-V2 changes significantly. Its uAP declines more steeply, falling from " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "70.05\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "48.29\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " (a decrease of " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "22.21\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ") between " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 32" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", where its performance converges. It is also more sensitive to smaller increases in background durations, falling from " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "70.50\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "64.08\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " (a decrease of " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "6.42\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ") between " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ", where a decrease of only " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "0.38\\%" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " is experienced on " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "D^{overlap}" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " over the same interval. At higher values of " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "inline_equation", + "content": ">40" + }, + { + "bbox": [ + 55, + 378, + 296, + 714 + ], + "type": "text", + "content": ") the performance drops below that of the 3D-R50 (although only once" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "type": "text", + "content": "the background-only performance has been surpassed). Although uAP falls more steeply, mAP declines more slowly and performance above that of the BG-only model is maintained even at the highest values of " + }, + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 554, + 179 + ], + "type": "text", + "content": ". This suggests that, although the MViT-V2 achieves higher baseline performance on the disjoint dataset, it is more sensitive to increases in background frames when evaluating uAP; a scenario where the model's advantage is drowned out by background information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 190, + 470, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 470, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 470, + 204 + ], + "type": "text", + "content": "5.4. Background Bias Mitigation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 209, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 425 + ], + "type": "text", + "content": "Augmenting Performance via Explicit Background Neutralisation. In this final experimental section, we justify the need for paired, real-world background samples, as uniquely provided by our dataset. We show that these sample pairings are able to significantly improve model performance. To show this, we analyse the impact of background subtraction operations in both input and embedding spaces across each of the model architectures and dataset configurations. Two background-subtracted views of our dataset were created for input space experiments. This was done by performing pixel-wise subtraction of " + }, + { + "bbox": [ + 313, + 209, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 209, + 555, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 209, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 209, + 555, + 425 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 209, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 209, + 555, + 425 + ], + "type": "text", + "content": ". To perform algebraic background subtraction in latent space, we extracted the feature vectors of the foreground video pairs. As described in Sec. 5.2, we achieve this by processing the video tensors up to the last convolutional layer and averaging the tokens from the last hidden state in the ResNet and MViT-V2 models, respectively. The background-subtracted embedding is then computed as" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 444, + 436, + 515 + ], + "blocks": [ + { + "bbox": [ + 316, + 444, + 436, + 515 + ], + "lines": [ + { + "bbox": [ + 316, + 444, + 436, + 515 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 436, + 515 + ], + "type": "image", + "image_path": "005b94d4eed3c5553ed3cdf803597d0099fc8bb2cf4e216c16501ad2ca7cba9a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 517, + 416, + 540 + ], + "lines": [ + { + "bbox": [ + 345, + 517, + 416, + 540 + ], + "spans": [ + { + "bbox": [ + 345, + 517, + 416, + 540 + ], + "type": "text", + "content": "foreground x fx 3x 224x 224" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 439, + 445, + 551, + 515 + ], + "blocks": [ + { + "bbox": [ + 439, + 445, + 551, + 515 + ], + "lines": [ + { + "bbox": [ + 439, + 445, + 551, + 515 + ], + "spans": [ + { + "bbox": [ + 439, + 445, + 551, + 515 + ], + "type": "image", + "image_path": "da1cfbb492290cd4d4c63f0fe109f7ae51c23439a47baba6bccd1a8b419fb509.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 542, + 553, + 635 + ], + "blocks": [ + { + "bbox": [ + 451, + 517, + 541, + 541 + ], + "lines": [ + { + "bbox": [ + 451, + 517, + 541, + 541 + ], + "spans": [ + { + "bbox": [ + 451, + 517, + 541, + 541 + ], + "type": "text", + "content": "background (b x f x 3 x 224 x 224)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 542, + 553, + 635 + ], + "lines": [ + { + "bbox": [ + 317, + 542, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 542, + 553, + 635 + ], + "type": "image", + "image_path": "aa6d66bad7d7a778dbfa4ee378fbe25848f1cbc285034806f09f32d57a393ac6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 644, + 554, + 710 + ], + "lines": [ + { + "bbox": [ + 313, + 644, + 554, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 554, + 710 + ], + "type": "text", + "content": "Figure 6. Latent Space Background Compensation. The proposed operation uses " + }, + { + "bbox": [ + 313, + 644, + 554, + 710 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 644, + 554, + 710 + ], + "type": "text", + "content": " to perform a weighted subtraction of background features (computed via the dataset pairing of a real-world, background video) from foreground features in latent space. The dotted line between the encoders indicates shared weights (i.e., they are the same model)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5439" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "type": "inline_equation", + "content": "z^{\\mathcal{F} - \\mathcal{B}} = z^{\\mathcal{F}} - (1 - \\alpha)\\cdot z^{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "type": "text", + "content": " is a coefficient used to control the relative magnitude of the algebraic subtraction operation as schematically shown in Fig. 6. In practice, the modulation parameter " + }, + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "type": "text", + "content": " was ablated to increase linearly or exponentially over the course of training and results are reported for both. In all cases " + }, + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 71, + 294, + 156 + ], + "type": "text", + "content": " was initialized at zero and increased to one by the end of training." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 164, + 294, + 402 + ], + "blocks": [ + { + "bbox": [ + 58, + 164, + 294, + 402 + ], + "lines": [ + { + "bbox": [ + 58, + 164, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 58, + 164, + 294, + 402 + ], + "type": "table", + "html": "
ModelSubtraction\\(D^{overlap}\\)\\(D^{disjoint}\\)
uAPmAPuAPmAP
2D R50Baseline64.1541.2844.6718.08
\\(\\mathcal{F}+\\mathcal{B}\\)67.7444.9945.2818.84
\\(\\mathcal{F}-\\mathcal{B}\\)65.9642.1254.5326.65
\\(\\mathcal{F}-\\mathcal{B}\\)72.2147.1253.7325.84
None56.2138.8427.2417.94
Linear53.7839.0129.7517.17
Exponential54.0537.1125.8017.73
3D R50Baseline72.7949.3257.7430.89
\\(\\mathcal{F}+\\mathcal{B}\\)76.4153.5765.6639.17
\\(\\mathcal{F}-\\mathcal{B}\\)66.5142.5653.5827.98
\\(\\mathcal{F}-\\mathcal{B}\\)71.9346.3653.9225.05
None72.3051.3963.1434.85
Linear77.3152.8266.6936.31
Exponential73.5251.2464.5836.36
Baseline75.4451.5470.5035.29
\\(\\mathcal{F}+\\mathcal{B}\\)72.8253.4968.0038.09
MViT-V2\\(\\mathcal{F}-\\mathcal{B}\\)66.8442.2757.2128.80
\\(\\mathcal{F}-\\mathcal{B}\\)74.9049.4061.2231.51
None74.2344.1168.5635.14
Linear79.32*53.83*71.64*39.04
Exponential77.7050.9371.2839.25*
", + "image_path": "afbece896aeee9e550c686a5e5901ae6f66e60fb0ddf0f71f1e883ba8f48eb05.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "text", + "content": "Table 2. Quantifying Performance Improvements of Background Subtraction. Comparison of the 2D R50, 3D R50, and MViT-V2 models on both dataset configurations. Performance is reported when training on " + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{F} - \\mathcal{B}" + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{F} - \\tilde{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "text", + "content": " videos in input (light green) and embedding space (light blue). First and second highest scores are indicated for each model, while the highest overall scores are marked by an asterisks " + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "inline_equation", + "content": "(\\ast)" + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "text", + "content": ". Note the baseline is trained on " + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "text", + "content": ", as reported in Tab. 1. " + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{F} + \\mathcal{B}" + }, + { + "bbox": [ + 55, + 412, + 295, + 547 + ], + "type": "text", + "content": " provides an alternative baseline but is reliant on background videos at test time. Results confirm the superiority of real-world background videos over synthetic data, and of latent space neutralisation over input space background subtraction." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Results and Performance Improvements. Tab. 2 shows that for the basic 2D-R50 model, background subtraction in input space is highly effective. When using " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": ", we observe substantial performance improvements over the baseline, with percentage increases of " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "+8.06\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "5.84\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " in uAP and mAP on " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "D^{overlap}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "+9.06\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "+7.76\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " in uAP and mAP on " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "D^{disjoint}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": ". Using " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " also yields improvements, though these gains are notably smaller. Background subtraction in embedding space does not provide any benefit to the 2D-R50. Although input space background subtraction is effective for the 2D-R50, it does not improve performance for the 3D-R50 and MViT-V2 models. Neither original nor synthetic backgrounds provide an advantage over" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "their respective baselines, indicating that input space subtraction may not suit spatio-temporal architectures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": "Conversely, background subtraction in embedding space provides significant performance improvements for the 3D-R50 and MViT-V2 models, particularly on " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "D^{disjoint}" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ". For the 3D-R50, embedding space subtraction with linear scheduling results in a " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "+8.95\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "+5.42\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " increase in uAP and mAP, respectively. Exponential scheduling also yields improvements over the baseline. The improvements are lower than linear scheduling with respect to uAP " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "(+6.84\\%)" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " but marginally better for mAP " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "(+5.47\\%)" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ". MViT-V2 also benefits from embedding space subtraction on " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "D^{disjoint}" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ": linear scheduling improves uAP and mAP by " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "+1.14\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "+3.75\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ", respectively, whereas exponential scheduling provides a " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "+0.78\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " increase in uAP and a " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "+3.96\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " increase in mAP. These findings suggest that embedding space subtraction is effective for both 3D-R50 and MViT-V2 models on " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "D^{disjoint}" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ", with linear and exponential scheduling favouring uAP and mAP performance, respectively. Embedding space subtraction also enhances performance on " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "D^{overlap}" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ", though to a lesser extent than on " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "D^{disjoint}" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ". For the 3D-R50 model, applying embedding space subtraction with linear scheduling yields an improvement of " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "4.52\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "3.50\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " in uAP and mAP, respectively, while exponential scheduling results in smaller gains of " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "0.73\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "1.92\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ". The MViT-V2 model follows a similar trend: linear scheduling increases uAP and mAP by " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "3.88\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "2.29\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ", respectively, whereas exponential scheduling improves uAP by " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "2.26\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": " but slightly reduces mAP by " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "-0.61\\%" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ". These results suggest that algebraic embedding space subtraction is beneficial on " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "D^{overlap}" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ", although the performance gains are relatively small compared to " + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "inline_equation", + "content": "D^{disjoint}" + }, + { + "bbox": [ + 313, + 98, + 553, + 469 + ], + "type": "text", + "content": ", and that linear scheduling generally provides better results than exponential scheduling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 487, + 388, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 388, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 388, + 500 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduced PanAf-FGBG, the first wildlife dataset designed to assess backgrounds and out-of-distribution generalization in behavior recognition. Our results demonstrate that real-world background information is a strong predictor of animal behavior and can significantly enhance recognition performance. While there are some limitations — such as its restriction to statically installed camera trap videos, the additional data and compute required for latent space background subtraction, and the potential for unnatural artifacts in background concatenation — the dataset remains a valuable resource. PanAf-FGBG enables direct evaluation under out-of-distribution conditions, and its rich metadata supports additional tasks and experimental setups. We hope it will help with understanding and improving model performance in this challenging domain and ultimately benefit endangered and charismatic species like the one featured in this paper." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5440" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 93, + 297, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 93, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 93, + 297, + 514 + ], + "type": "text", + "content": "We thank the Pan African Programme: 'The Cultured Chimpanzee' team and its collaborators for allowing the use of their data for this paper. We thank Amelie Pettrich, Antonio Buzharevski, Eva Martinez Garcia, Ivana Kirchmair, Sebastian Schutte, Linda Gerlach and Fabina Haas. We also thank management and support staff across all sites; specifically Yasmin Moebius, Geoffrey Muhanguzi, Martha Robbins, Henk Eshuis, Sergio Marrocoli and John Hart. Thanks to the team at https://www.chimpandsee.org particularly Briana Harder, Anja Landsmann, Laura K. Lynn, Zuzana Macháčková, Heidi Pfund, Kristeena Sigler and Jane Wideness. The work that allowed for the collection of the dataset was funded by the Max Planck Society, Max Planck Society Innovation Fund, and Heinz L. Krekeler. In this respect we would like to thank: Ministre des Eaux et Forets, Ministère de l'Enseignement supérieur et de la Recherche scientifique in Côte d'Ivoire; Institut Congolais pour la Conservation de la Nature, Ministère de la Recherche Scientifique in Democratic Republic of Congo; Forestry Development Authority in Liberia; Direction Des Eaux Et Forets, Chasses Et Conservation Des Sols in Senegal; Makerere University Biological Field Station, Uganda National Council for Science and Technology, Uganda Wildlife Authority, National Forestry Authority in Uganda; National Institute for Forestry Development and Protected Area Management, Ministry of Agriculture and Forests, Ministry of Fisheries and Environment in Equatorial Guinea. This work was supported by the UKRI CDT in Interactive AI (grant EP/S022937/1). This work was in part supported by the US National Science Foundation Awards No. 2118240 \"HDR Institute: Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning\" and Award No. 2330423 and Natural Sciences and Engineering Research Council of Canada under Award No. 585136 for the \"AI and Biodiversity Change (ABC) Global Center\"." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 529, + 115, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 529, + 115, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 115, + 542 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 550, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 61, + 550, + 295, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 550, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 61, + 550, + 295, + 571 + ], + "type": "text", + "content": "[1] Pan African Programme: The Cultured Chimpanzee, 1999. Online; accessed 29 January 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 575, + 296, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 575, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 61, + 575, + 296, + 608 + ], + "type": "text", + "content": "[2] Analysing animal behaviour in wildlife videos using face detection and tracking. IEEE Proceedings-Vision, Image and Signal Processing, 153(3):305-312, 2006. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 611, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 611, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 61, + 611, + 296, + 677 + ], + "type": "text", + "content": "[3] Mimi Arandjelovic, Colleen R Stephens, Maureen S McCarthy, Paula Dieguez, Emmie K Kalan, Nuria Maldonado, Christophe Boesch, and Hjalmar S Kuehl. Chimp&see: An online citizen science platform for large-scale, remote video camera trap annotation of chimpanzee behaviour, demography and individual identification. PeerJ Preprints, 2016. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 680, + 296, + 714 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Daniel Schofield, Sophie Berdugo, Joana Bessa, Jake Owen, Kimberly J Hockings, Tetsuro Matsuzawa, Misato Hayashi, Dora Biro, et al. Au" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tomated audiovisual behavior recognition in wild primates. Science advances, 7(46):eabi4883, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 319, + 96, + 553, + 128 + ], + "type": "text", + "content": "[5] Sara Beery, Arushi Agarwal, Elijah Cole, and Vighnesh Birodkar. The iwildcam 2021 competition dataset. arXiv preprint arXiv:2105.03494, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 319, + 129, + 553, + 194 + ], + "type": "text", + "content": "[6] Otto Brookes, Majid Mirmehdi, Hjalmar S. Kuhl, and Tilo Burghardt. Triple-stream deep metric learning of great ape behavioural actions. In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 294-302, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 319, + 195, + 553, + 227 + ], + "type": "text", + "content": "[7] Otto Brookes, Majid Mirmehdi, Hjalmar Kuhl, and Tilo Burghardt. Chimpvlm: Ethogram-enhanced chimpanzee behaviour recognition. 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 228, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 228, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 319, + 228, + 553, + 293 + ], + "type": "text", + "content": "[8] Otto Brookes, Majid Mirmethdi, Colleen Stephens, Samuel Angedakin, Katherine Corogenes, Dervla Dowd, Paula Dieguez, Thurston C Hicks, Sorrel Jones, Kevin Lee, et al. Panaf20k: a large video dataset for wild ape detection and behaviour recognition. International Journal of Computer Vision, pages 1-17, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 294, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 294, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 319, + 294, + 555, + 338 + ], + "type": "text", + "content": "[9] Jackie Chappell and Susannah KS Thorpe. The role of great ape behavioral ecology in one health: Implications for captive welfare and re-habilitation success. American journal of primatology, 84(4-5):e23328, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 338, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 404 + ], + "type": "text", + "content": "[10] Jun Chen, Ming Hu, Darren J Coker, Michael L Berumen, Blair Costelloe, Sara Beery, Anna Rohrbach, and Mohamed Elhoseiny. Mammalnet: A large-scale video benchmark for mammal recognition and behavior understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13052-13061, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 404, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 449 + ], + "type": "text", + "content": "[11] Jinwoo Choi, Chen Gao, Joseph CE Messou, and Jia-Bin Huang. Why can't i dance in the mall? learning to mitigate scene bias in action recognition. Advances in Neural Information Processing Systems, 32, 2019. 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[12] Jihoon Chung, Yu Wu, and Olga Russakovsky. Enabling detailed action recognition evaluation through video dataset augmentation. Advances in Neural Information Processing Systems, 35:39020-39033, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "text", + "content": "[13] Fay E Clark. Great ape cognition and captive care: Can cognitive challenges enhance well-being? Applied Animal Behaviour Science, 135(1-2):1-12, 2011. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 581 + ], + "type": "text", + "content": "[14] Fagner Cunha, Eulanda M dos Santos, Raimundo Barreto, and Juan G Colonna. Filtering empty camera trap images in embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2438-2446, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 582, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 647 + ], + "type": "text", + "content": "[15] Alexander D'Amour, Katherine Heller, Dan Moldovan, Ben Adlam, Babak Alipanahi, Alex Beutel, Christina Chen, Jonathan Deaton, Jacob Eisenstein, Matthew D Hoffman, et al. Underspecification presents challenges for credibility in modern machine learning. Journal of Machine Learning Research, 23(226):1-61, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 647, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 714 + ], + "type": "text", + "content": "[16] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5441" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "text", + "content": "[17] Isla Duporge, Maksim Kholiavchenko, Roi Harel, Dan Rubenstein, Meg Crofoot, Tanya Berger-Wolf, Stephen Lee, Scott Wolf, Julie Barreau, Jenna Kline, et al. Baboonland dataset: Tracking primates in the wild and automating behaviour recognition from drone videos. arXiv preprint arXiv:2405.17698, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "text", + "content": "[18] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "text", + "content": "[19] Rohit Girdhar and Deva Ramanan. Cater: A diagnostic dataset for compositional actions and temporal reasoning. arXiv preprint arXiv:1910.04744, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "text", + "content": "[20] Abhinav Gupta and Larry S Davis. Objects in action: An approach for combining action understanding and object perception. pages 1-8. IEEE, 2007. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 354 + ], + "type": "text", + "content": "[22] Yun He, Soma Shirakabe, Yutaka Satoh, and Hirokatsu Kataoka. Human action recognition without human. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 11-17. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "text", + "content": "[23] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In Proceedings of the European conference on computer vision (ECCV), pages 771-787, 2018. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[24] Filip Ilic, Thomas Pock, and Richard P Wildes. Is appearance free action recognition possible? In European Conference on Computer Vision, pages 156-173. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[25] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "text", + "content": "[26] Maksim Kholiavchenko, Jenna Kline, Maksim Kukushkin, Otto Brookes, Sam Stevens, Isla Duporge, Alec Sheets, Reshma R Babu, Namrata Banerji, Elizabeth Campolongo, et al. Deep dive into kabr: a dataset for understanding ungulate behavior from in-situ drone video. Multimedia Tools and Applications, pages 1-20, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 644 + ], + "type": "text", + "content": "[27] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribution shifts. In International conference on machine learning, pages 5637-5664. PMLR, 2021. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "type": "text", + "content": "[28] Hjalmar S Kuhl and Tilo Burghardt. Animal biometrics: quantifying and detecting phenotypic appearance. Trends in ecology & evolution, 28(7):432-441, 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[29] Haoxin Li, Yuan Liu, Hanwang Zhang, and Boyang Li. Mitigating and evaluating static bias of action representations in the background and the foreground. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 19911-19923, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "text", + "content": "[30] Yi Li and Nuno Vasconcelos. Repair: Removing representation bias by dataset resampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9572-9581, 2019. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[31] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 252 + ], + "type": "text", + "content": "[32] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4804-4814, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "type": "text", + "content": "[33] Dan Liu, Jin Hou, Shaoli Huang, Jing Liu, Yuxin He, Bochuan Zheng, Jifeng Ning, and Jingdong Zhang. Lote-animal: A long time-span dataset for endangered animal behavior understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20064-20075, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 376 + ], + "type": "text", + "content": "[34] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 399 + ], + "type": "text", + "content": "[35] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 464 + ], + "type": "text", + "content": "[36] Xiaoxuan Ma, Stephan Kaufhold, Jiajun Su, Wentao Zhu, Jack Terwilliger, Andres Meza, Yixin Zhu, Federico Rossano, and Yizhou Wang. Chimpact: A longitudinal dataset for understanding chimpanzee behaviors. Advances in Neural Information Processing Systems, 36:27501-27531, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "text", + "content": "[37] Marcin Marszalek, Ivan Laptev, and Cordelia Schmid. Actions in context. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 2929-2936. IEEE, 2009. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "text", + "content": "[38] Zhongqi Miao, Ziwei Liu, Kaitlyn M Gaynor, Meredith S Palmer, Stella X Yu, and Wayne M Getz. Iterative human and automated identification of wildlife images. Nature Machine Intelligence, 3(10):885-895, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 622 + ], + "type": "text", + "content": "[39] Mazda Moayeri, Phillip Pope, Yogesh Balaji, and Soheil Feizi. A comprehensive study of image classification model sensitivity to foregrounds, backgrounds, and visual attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19087-19097, 2022. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[40] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19023-19034, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[41] Toby Perrett, Saptarshi Sinha, Tilo Burghardt, Majid Mirmethdi, and Dima Damen. Use your head: Improving long-tail video recognition. In Proceedings of the IEEE/CVF con" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5442" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "text", + "content": "ference on computer vision and pattern recognition, pages 2415-2425, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "text", + "content": "[42] Laura J Pollock, Justin Kitzes, Sara Beery, Kaitlyn M Gaynor, Marta A Jarzyna, Oisin Mac Aodha, Bernd Meyer, David Rolnick, Graham W Taylor, Devis Tuia, et al. Harnessing artificial intelligence to fill global shortfalls in biodiversity knowledge. Nature Reviews Biodiversity, pages 1-17, 2025. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "text", + "content": "[43] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[44] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 286 + ], + "type": "text", + "content": "[45] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "text", + "content": "[46] Faizaan Sakib and Tilo Burghardt. Visual recognition of great ape behaviours in the wild. In Workshop on the Visual Observation and Analysis of Vertebrate and Insect Behaviour, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 388 + ], + "type": "text", + "content": "[47] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9661-9669, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 294, + 453 + ], + "type": "text", + "content": "[48] Devis Tuia, Benjamin Kellenberger, Sara Beery, Blair R Costelloe, Silvia Zuffi, Benjamin Risse, Alexander Mathis, Mackenzie W Mathis, Frank Van Langevelde, Tilo Burghardt, et al. Perspectives in machine learning for wildlife conservation. Nature communications, 13(1):1-15, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "text", + "content": "[49] Tuan-Hung Vu, Catherine Olsson, Ivan Laptev, Aude Oliva, and Josef Sivic. Predicting actions from static scenes. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 421-436. Springer, 2014. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 513, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 588 + ], + "type": "text", + "content": "[50] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11804-11813, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[51] Kai Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. arXiv preprint arXiv:2006.09994, 2020.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 690 + ], + "type": "text", + "content": "[52] Xinyu Yang, Majid Mirmehdi, and Tilo Burghardt. Great ape detection in challenging jungle camera trap footage via attention-based spatial and temporal feature blending. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[53] Xinyu Yang, Tilo Burghardt, and Majid Mirmehdi. Dynamic curriculum learning for great ape detection in the wild. In-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 306 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "ternational Journal of Computer Vision, 131(5):1163-1181, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[54] Manlin Zhang, Jinpeng Wang, and Andy J Ma. Suppressing static visual cues via normalizing flows for self-supervised video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3300-3308, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[55] Mingjun Zhao, Yakun Yu, Xiaoli Wang, Lei Yang, and Di Niu. Search-map-search: a frame selection paradigm for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10627-10636, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[56] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "text", + "content": "[57] Yi Zhu, Xinyu Li, Chunhui Liu, Mohammadreza Zolfaghari, Yuanjun Xiong, Chongruo Wu, Zhi Zhang, Joseph Tighe, R Manmatha, and Mu Li. A comprehensive study of deep video action recognition. arXiv preprint arXiv:2012.06567, 2020. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5443" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_content_list.json b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac73c6cfd4a5bfa6167a59a52d85723cd32014fb --- /dev/null +++ b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_content_list.json @@ -0,0 +1,1743 @@ +[ + { + "type": "text", + "text": "The Photographer's Eye: Teaching Multimodal Large Language Models to See and Critique like Photographers", + "text_level": 1, + "bbox": [ + 143, + 130, + 854, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daiqing Qi $^{1,2}$ , Handong Zhao $^{2}$ , Jing Shi $^{2}$ , Simon Jenni $^{2}$ , Yifei Fan $^{2}$ , Franck Dernoncourt $^{2}$ , Scott Cohen $^{2}$ , Sheng Li $^{1}$", + "bbox": [ + 233, + 224, + 764, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Virginia, 2Adobe", + "bbox": [ + 370, + 267, + 622, + 287 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/376fbcb19caa8ab0088e9d2bb1b09a71bba274980cecd3fc779986b6c99d6c1d.jpg", + "image_caption": [ + "User: Are there any weaknesses in this photograph?", + "Q-Instruct: The image features a beautiful sunset with the sun's rays shining through the trees.", + "The scene captures the vibrant colors of the sunset, creating a serene and picturesque atmosphere. The hillside is covered in lush green trees, providing a sense of tranquility and natural beauty.", + "The combination of the sunset and the verdant landscape creates a captivating and visually pleasing view.", + "AesExpert: This image looks very clear. The colors in this image are rich and vibrant, with a good color combination, making the picture appear more vivid and bright. The lighting in this image is strong, creating a warm and soft atmosphere, with a strong contrast of dark and light. The composition of this image is a high horizon composition, highlighting the texture in the lower half of the picture. Therefore, this image looks quite beautiful.", + "PhotoEye: The composition of this image could be improved by straightening the horizon to create a more balanced and visually appealing frame. Additionally, the presence of lens flare may be distracting and detract from the appearance of the image. To address this, the photographer could consider using a lens hood or adjusting their shooting angle to minimize the effect of the flare. Furthermore, the tilted horizon could be corrected to create a more stable and visually appealing atmosphere. However, the amount of the image could be more nuanced, with a more subtle transition between highlights and shadows to create a more natural and inviting atmosphere." + ], + "image_footnote": [], + "bbox": [ + 109, + 300, + 272, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02db6404065d57764b2fc7920d13e00e39a93ead4e48068b6d013b22cdbaa684.jpg", + "image_caption": [ + "User: Are there any distractors in this image?", + "Q-Instruct: No", + "GPT-4o: 1. The Chain and Metal Poles: These elements draw attention because they are close to the subjects and create visual lines in the foreground, potentially competing for focus with the main subjects (the two individual). 2. Darker Darker Colors: The person in the yellow jacket stands out significantly against the darker colors of the other person and the background...", + "PhotoEye: The photo effectively captures a romantic moment between the couple, but there are some distracting elements that take away from the intimacy of the scene." + ], + "image_footnote": [], + "bbox": [ + 406, + 301, + 578, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f92f23541fd85ab48a0e33fe14d3f3d4f7dab68a0d12d3cc96f33c40b5ddf225.jpg", + "image_caption": [ + "Figure 1. Examples on our model (PhotoEye), existing MLLMs tailored for low-level vision or aesthetics, and GPT-4o (2024-08-06). The left example and the middle top example highlight a notable limitation of existing open-source MLLMs for low-level vision and aesthetics: insufficient coverage of visual aesthetics. When they fail to identify issues, they either provide positive aspects or claim there are no issues, significantly limiting their usefulness in real-world scenarios. The middle bottom example shows existing MLLMs' lack of expertise in photography: lowering exposure properly can instead enhance colors of objects, so B is correct. While other models, including GPT-4o, made mistakes, our model is correct. The right example reveals another clear limitation of existing open-source MLLMs: their vision encoders are insensitive to low-level vision and aesthetics. In a series of increasingly overexposed photos (2-nd, 3-rd, and 4-th), PhotoEye's vision modules, more attuned to low-level and aesthetic features, identified overexposure by the 2-nd photo, while other models recognized it only when the photos were severely overexposed (i.e., 3-rd and 4-th images). High-quality aesthetic content is highlighted." + ], + "image_footnote": [], + "bbox": [ + 406, + 388, + 576, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f4831d2d2e3ef4c1197dc7996076480887e1ef4b691e27df186c9cac7019498c.jpg", + "image_caption": [ + "AesExpert:", + "User: Please describe the exposure of the photo" + ], + "image_footnote": [], + "bbox": [ + 733, + 319, + 821, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a27dad46dd577d6b92bb09da86b378cb0dfffb3400dcd3c75bb393a615e976f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 321, + 883, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/009a912907b8339633276cc95293fac3f41842ba7130fd8ba72eebf648ba85c3.jpg", + "image_caption": [ + "The lines and wires in the background are a notable distraction, drawing the viewer's attention away from the couple and disrupting the visual flow. The image of the photographer could consider removing or minimizing these distractions to allow the viewer to focus on the tender moment between the two characters. This distraction would not be the most flattering choice, which could also detract from the overall impact of the image." + ], + "image_footnote": [], + "bbox": [ + 733, + 398, + 821, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/accdd638a95643858ceacbcc77ad4eebdf864974a50e7d594bd0f6de2a498871.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 407, + 883, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 638, + 326, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"While editing directly from life, photographers have found it too difficult to see simultaneously both the blue and the sky.\"", + "bbox": [ + 114, + 662, + 459, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "John Szarkowski, William Eggleston's Guide", + "bbox": [ + 140, + 709, + 447, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photographer and curator, Szarkowski insightfully revealed one of the notable gaps between general and aesthetic visual understanding: while the former focuses on identifying the factual element in an image (sky), the latter transcends such object identification, viewing it instead as an aesthetic component—a pure color block (blue). Such fundamental distinctions between general (detection, localization, etc.) and aesthetic (color, lighting, composition, etc.)", + "bbox": [ + 89, + 734, + 483, + 856 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "visual understanding present a significant challenge for Multimodal Large Language Models (MLLMs). Although some recent works have made initial explorations, they are often limited to general and basic aesthetic commonsense. As a result, they frequently fall short in real-world scenarios (Fig. 1), which require extensive expertise—including photographic techniques, photo pre/post-processing knowledge, and more, to provide a detailed analysis and description. To fundamentally enhance the aesthetics understanding of MLLMs, we first introduce a novel dataset, PhotoCritique, derived from extensive discussions among professional photographers and enthusiasts, and characterized by the large scale, expertise, and diversity. Then, to better learn visual aesthetics from PhotoCritique, we further propose a novel model, PhotoEye, featuring a language-guided multi-view vision fusion mechanism to understand image aesthetics from multiple perspectives. Finally, we present a", + "bbox": [ + 511, + 641, + 906, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Part of this work was completed during Daiqing's internship at Adobe. \n1William Eggleston is an American photographer, widely recognized as a pioneering figure in color photography.", + "bbox": [ + 89, + 862, + 482, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24807", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "novel benchmark, PhotoBench, a comprehensive and professional benchmark for aesthetic visual understanding. On existing benchmarks and PhotoBench, our model demonstrates clear advantages over existing models.", + "bbox": [ + 88, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 166, + 222, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs) [16, 19] have succeeded in general visual tasks but still struggle with a critical ability in human visual perception: aesthetic visual understanding, which is desired in various applications like image recommendation, editing, or generation. The challenge lies in the fundamental distinction between general and aesthetic visual understanding: while the former focuses on factual elements like identification of objects, the latter requires nuanced perception of aesthetic elements such as chromatic properties, or compositional principles. We attribute the shortcomings of MLLMs in aesthetic understanding to two primary factors: dataset and model.", + "bbox": [ + 91, + 191, + 482, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dataset: Current datasets such as Q-Instruct [29] and AesExpert [9], made initial explorations in low-level vision or visual aesthetics with MLLM but suffer from limitations in scale, expertise, and diversity. With around 20K images and 58K-88K image annotations, these datasets are insufficient for comprehensive aesthetic understanding, showing clear deficiencies in real-world scenarios, where images vary widely in subjects, scenes, styles, environments, and techniques, requiring larger datasets to ensure coverage. Additionally, their annotations come from small (typically tens of annotators), fixed groups of annotators, limiting diversity and expertise. Even if the annotators are reportedly trained, their expertise remains limited compared to professional photographers and enthusiasts.", + "bbox": [ + 91, + 373, + 482, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model: When adapting to aesthetic visual understanding, existing approaches [8, 10, 30, 38] typically finetune existing MLLMs like LLaVA [16] on customized datasets, ignoring the unique requirements of visual features for aesthetic visual understanding. Comprehensive understanding of image aesthetics requires multi-view information: for example, high-level features are essential for interpreting narrative and storytelling, while detailed spatial information is critical for understanding composition. Similarly, low-level details are required for perception of color and lighting, etc.", + "bbox": [ + 91, + 584, + 482, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges, we first introduce the PhotoCritique dataset, featuring (1) large scale: containing over 450K images with detailed aesthetic descriptions, and around 20 times the size of existing datasets [8, 30]; and (2) expertise and diversity: the aesthetic descriptions are derived from extensive discussions among large groups of professional photographers and enthusiasts online, rather than a fixed small number of annotators as in existing works [8, 30]. Subsequently, given the notable different focuses in general and aesthetic visual understanding, we further introduce PhotoEye, an MLLM tailored for aes", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "thetic visual understanding. PhotoEye features a language-guided multi-view vision fusion mechanism to effectively learn aesthetic knowledge from PhotoCritique, enabling it to understand image aesthetics from multiple perspectives for various aesthetics-related questions.", + "bbox": [ + 511, + 90, + 903, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, to provide a comprehensive evaluation of our model and existing MLLMs on aesthetic visual understanding, we introduce our new benchmark PhotoBench. PhotoBench draws its questions from in-depth discussions among professional photographers and enthusiasts, covering 284 sub-topics in photography, making it notably more professional and diverse than existing works. We summarize our contributions as follows:", + "bbox": [ + 511, + 167, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce PhotoCritique, the first large-scale, professional, and diverse dataset for aesthetic visual understanding, consisting of 450K images with 2.63M instruction-tuning pairs, sourced from extensive discussions among professional photographers and enthusiasts.", + "- We propose PhotoEye, an MLLM specifically designed for aesthetic understanding, featuring a language-guided multi-view vision fusion mechanism.", + "- We develop PhotoBench, a professional benchmark for aesthetic visual understanding, offering 284 photography sub-topics and insights from photography professionals.", + "- We conduct a comprehensive study on aesthetic visual understanding with our model and existing MLLMs, revealing the limitations of existing MLLMs and demonstrating how we overcome them through extensive experiments." + ], + "bbox": [ + 514, + 287, + 903, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 527, + 661, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Large Language Models. MLLMs [3, 16] are evolving rapidly recently, demonstrating advanced capabilities in image-text dialogue through enhanced alignment and fine-tuning. Further studies [2, 5, 6, 28, 33] have strengthened MLLMs by increasing the data scale or improving the backbone model. Furthermore, with customized data, a branch of works [4, 21, 33, 35] extend MLLMs' capabilities to referring and grounding. However, despite these advancements in general visual tasks, the aesthetic visual understanding of MLLMs remains largely underexplored.", + "bbox": [ + 511, + 551, + 903, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Understanding Image Aesthetics with MLLM. Traditional Image Aesthetics Assessment (IAA) [17, 26, 32] is often formalized as a regression task, where the model learns to predict the score of the aesthetic quality of images. One notable limitation is the lack of interpretability, which is critical in practice because we also want to understand why the image is good or bad. Recent advances in MLLM make it possible: a few recent works [8-10, 29-31, 38] explore interpretable aesthetic understanding with MLLMs. However, they often struggle with the limited data scale or lack diverse, professionally curated aesthetic descriptions.", + "bbox": [ + 511, + 704, + 903, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Improving Visual Perception in MLLM. Recently, several studies [13, 22-24, 28, 41] have investigated back", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24808", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c3872fe1fa86838fb543bf9ab75a521ee72837624fc4842fa92472056d04890.jpg", + "image_caption": [ + "Figure 2. An example from the dpchallenge platform. 3 of the total 64 comments for the photo are presented for illustration." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 187, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "#1: Very good one, even if the unsharp mask is too strong.", + "bbox": [ + 191, + 94, + 429, + 103 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "#2: This is gorgeous. The use of parallel lines and the vanishing point really leads us into this photo. And the mixture of warm and cool colors is very pleasing.", + "bbox": [ + 191, + 107, + 460, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "#3: This appearance is particularly prominent in the lower left foreground foliage which is also quite bright and thus accentuates the over-sharpened look. Perhaps it is jpeg artifacting, if I look in the upper branches of the tree I can see a cross-hatch pattern...", + "bbox": [ + 191, + 133, + 467, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bone vision modules for MLLM. While SPHINX [14] simply concatenates features from different encoders, Cambrian [28] and MoVA [41] merge different visual features into a single one. However, the fusion process in MoVA is based on the CLIP feature, which may introduce bias, and the feature extraction (prior to fusion) lacks language-based control. Cambrian [28] employs query-based attention, however, visual features from different encoders are simply averaged during fusion, and the queries are independent of language instructions. To address these limitations, we propose a new MLLM, PhotoEye, featuring a language-guided multi-view vision fusion mechanism to effectively learn aesthetic knowledge.", + "bbox": [ + 89, + 223, + 483, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Learning with Photography Enthusiasts and Professionals", + "text_level": 1, + "bbox": [ + 89, + 431, + 483, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One key reason existing MLLMs struggle with aesthetic visual understanding in real-world scenarios is the training data limitations: (1) Data Scale: Current datasets [8, 30] ( $\\sim$ 20K images) lack sufficient coverage of photography genres, styles, themes, etc. High-quality aesthetic descriptions are also limited, with typically 57K-88K detailed descriptions in existing works [8, 30, 38], while our dataset provides over 450K images and detailed descriptions. (2) Description Quality: Most datasets [8, 30] rely on a small, fixed group of annotators, limiting expertise and diversity. In contrast, our PhotoCritique dataset draws insights from hundreds of thousands of photographers and enthusiasts, yielding more professional, nuanced feedback.", + "bbox": [ + 89, + 476, + 483, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Mining from Digital Photography Platforms", + "text_level": 1, + "bbox": [ + 89, + 681, + 465, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the limitations, we avoided using annotators and instead drew insights from a broad range of online communities. The Digital Photography Challenge (DPC) is an online platform, where users submit their works across various themes and genres and receive constructive feedback from other photographers or enthusiasts (Fig. 2). The wealth of high-quality discussions paired with images on DPC provides valuable resources for MLLMs to learn visual aesthetics. However, directly using this data for supervised fine-tuning is impractical due to the significant noise present in the raw comments. To address this issue, we developed strategies to analyze, summarize, and refine raw comments into a single, well-crafted critique for each image. First,", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8d88079903468e2a223e789861137172bef6d9ba6b8c2d061c792d13f0e925ec.jpg", + "image_caption": [ + "Figure 3. Data Generation Pipeline. The generation of one critique from a group of comments of one image is shown as an example." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 906, + 132 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0834d1657549b669aec94bf4d587975ea2cde3a5a575d62cc7a8cd9f95500234.jpg", + "image_caption": [ + "PhotoCritique (Aesthetic Conversation)" + ], + "image_footnote": [], + "bbox": [ + 517, + 181, + 903, + 304 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8a023aaaa6b9700fe04c29538e13bb179a2d248321b44a7f5ba045965a70e81a.jpg", + "image_caption": [ + "Figure 4. Top: Examples of PhotoCritique. Bottom: Comparison with examples in Q-Instruct [30]." + ], + "image_footnote": [], + "bbox": [ + 517, + 316, + 903, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we need to collect raw data from dpchallenge.com online. Since existing works [37] already collected raw data prior to 2022 from the website, we can skip the data scraping process and focus on data processing. Given a post from the DPC platform (comprising one uploaded image and a series of user comments, shown in Fig. 2), we design a multi-step pipeline to generate a single unified photo critique using prompt-based methods from comments only without the image. Fig. 3 briefly illustrates the pipeline with the example of generating a single critique from a set of comments of one image. We first instruct the LLM to analyze and summarize the raw comments, Then the LLM is asked to integrate the comments into a cohesive and unified photo critique. While LLMs possess strong reasoning and comprehension abilities, the immediate output critique is sometimes sparse in information or misleading. Therefore, we designed a filtering strategy to eliminate low-quality data with a smaller LLM for improved efficiency. During the filtering stage, generated photo critiques are input into the smaller LLM, and prompt-based strategies are used to detect whether the generated photo critique conveys enough aesthetics-related information. Only critiques that pass this filtering stage are accepted. We term this portion of data as aesthetic description. Details available in Appendix A.", + "bbox": [ + 511, + 425, + 906, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From General Photo Critiques to Diverse Instruction-tuning Pairs. In the previous step, we generated photo critiques from DPC's raw comments, which provide overall aesthetic reviews of the given images. In this step, we further utilize LLMs to generate multiple high-quality, diverse conversations addressing various aspects of image aesthetics, such as lighting, composition, color, emotion, narrative,", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24809", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/de3ee9b250f023bfc33849c916f8623d1d6f3474b3632a414e5d8a94b0ee39cf.jpg", + "image_caption": [ + "Figure 5. Comparison of description length Distributions of aesthetic comments in PhotoCritique and Q-Instruct. Scale of PhotoCritique is Blue and that of Q-Instruct is Red." + ], + "image_footnote": [], + "bbox": [ + 91, + 84, + 480, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c75d4c252b3011307a47f3da654080497969977cf477c3c0c85766d5652d154.jpg", + "image_caption": [ + "Figure 6. Top: Distribution of Top 40 categories. Bottom: Examples of photos from a subset of categories." + ], + "image_footnote": [], + "bbox": [ + 96, + 214, + 478, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and so on, as well as extensive photographic techniques and pre/post-processing skills. Another critical reason for this step is that raw comments are sometimes too lengthy, which can lead to certain aesthetic aspects being overlooked in the LLM-generated critique. By specifically prompting about these aesthetic aspects, the LLM can effectively summarize essential information related to each aspect from the lengthy raw comments. Similarly, after generating all the question-answer pairs, we apply prompt-based filtering strategies to reject less informative pairs, accepting only the high-quality samples. We term this portion of data as aesthetic conversation. Additionally, we also generate MCQs from photo critiques with LLMs. Specifically, each critique is input into the LLM with prompts to generate five MCQs based on the ground truth provided in it. We term this portion of data as aesthetic VQA. Details available in Appendix A.", + "bbox": [ + 88, + 391, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. PhotoCritique", + "text_level": 1, + "bbox": [ + 89, + 643, + 236, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. Combining all above, we build PhotoCritique, the first large-scale dataset for expert-level image aesthetics understanding. PhotoCritique consists of three parts: aesthetic description, aesthetic conversation, and aesthetic VQA. Examples of each are shown in Fig. 4 (Top). The three parts contain 450K, 1.9M, and 250K samples, respectively. Born from a large volume of shared images and discussions by photography enthusiasts and professionals on online photography forums, PhotoCritique covers a wide range of over 70 photographic categories, encompassing diverse scenes and subjects. Fig. 6 shows the distribution of the top 40 categories along with several image examples. The extensive scale and diversity enhance the model's effectiveness in real-world scenarios.", + "bbox": [ + 89, + 667, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparison with Existing Datasets. PhotoCritique rep", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7518d21a9be5782301f3d42f4dfd102c12cce8e972889b2c65d3c034131f6667.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Image#Annotations#Training Samples#Avg. Length
UNIAA57K57K57K32.8
AesExpert*21K88K--
Q-Instruct18K58K66K46.4
PhotoCritique450K450K2.4M65.2
", + "bbox": [ + 527, + 88, + 897, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Statistics of samples attributed to detailed aesthetic descriptions and conversations across different datasets. Datasets with * are currently not available.", + "bbox": [ + 511, + 154, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resents a fundamental improvement over existing works in both quantity and quality. Data Scale: we compares PhotoCritique with existing datasets by showing statistics for data related to detailed aesthetic descriptions and conversations in Tab. 1. This includes the number of source images, the annotations for these images, and the final count of instruction-tuning pairs generated from these annotations using LLMs. Additionally, we calculate the average length of aesthetic comments (including both descriptions and conversations) in the training data. Tab. 1 focuses on training data statistics for aesthetic comments, as these are essential in practical applications where users need detailed feedback on why their photos are strong or weak and how to improve them. In contrast, Multiple Choice Questions (MCQ) are relatively limited in this regard. Besides, aesthetic comments better reflect the information density of the dataset than MCQs, as generating MCQs from a detailed comment is easy and can result in numerous MCQs of varying quality depending on the input prompt. Therefore, the number of MCQs serves only as a relatively rough indicator of the information density, although the number of MCQs in PhotoCritique is also the largest. As shown in Tab. 1, PhotoCritique demonstrates a notable improvement across all metrics compared to existing datasets. Data Quality: PhotoCritique is more professional, and diverse. Unlike most existing works [8, 30], which rely on a small group (usually a few dozen) of annotators to label data, our dataset is sourced from images and discussion from over 107,000 photography enthusiasts and professionals on online photography forums. Fig. 2 provides an intuitive comparison between our dataset and Q-Instruct, with examples of Q-Instruct in its original paper. While Q-Instruct primarily focuses on basic low-level visual facts, PhotoCritique not only addresses low-level details but also considers how these elements impact image aesthetics and ways to improve them. Additionally, we include discussions on photography techniques (e.g., the first example in the aesthetic description covers flash usage and lighting) and the overall impression conveyed by the photo. The comprehensiveness of aesthetic comments in PhotoCritique is also reflected in the average length: Fig. 5 shows the distribution of aesthetic comment lengths in PhotoCritique and Q-Instruct, with PhotoCritique containing a large number of detailed descriptions. Tab. 1 also shows that the average length of aesthetic comments in PhotoCritique is notably greater than that of existing datasets. More examples in Appendix A.", + "bbox": [ + 511, + 205, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24810", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bdf846442ce09e879f78d4584365a6d600adcfe3e5c39fc00d8557d33f9d6927.jpg", + "image_caption": [ + "Figure 7. Framework of PhotoEye. Given an input photo and a language instruction, visual features from different visual encoders are fused in our vision fusion with language guidance, then fed into the LLM with language embeddings to generate LLM response." + ], + "image_footnote": [], + "bbox": [ + 122, + 80, + 877, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Improving Visual Aesthetics Perception", + "text_level": 1, + "bbox": [ + 89, + 330, + 444, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. See like Photographers", + "text_level": 1, + "bbox": [ + 89, + 357, + 305, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to limitations in existing datasets, Open-source MLLMs face another notable challenge: the limited sensitivity of the vision encoder to aesthetic elements. Most existing works [8, 30] use CLIP as the vision encoder, which, however, is pre-trained with high-level image-text alignment in general domains and is relatively less effective at capturing aesthetic elements. Fig. 1 (right) illustrates this issue, where Q-Instruct and AesExpert only report 'overexposure' when the image is severely overexposed. In Fig. 8, we quantitatively validate it by calculating the discriminability (defined as the average distance between these embeddings) of aesthetic features from different vision encoders. Details are available in Appendix B. It shows that CLIP is less effective in discriminating towards aesthetics-related features as they are more densely clustered. To handle the diversity of aesthetics understanding, from high-level elements like emotion and storytelling to low-level features such as lighting, color, and composition, we propose PhotoEye, a MLLM with a language-guided multi-view vision fusor, which enhances visual perception by fusing multiple vision encoders pre-trained on complementary tasks.", + "bbox": [ + 91, + 381, + 483, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. PhotoEye", + "text_level": 1, + "bbox": [ + 89, + 710, + 202, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. PhotoEye is shown in Fig. 7. It consists of a multi-view vision fusor and a backbone LLM. Given an image and an aesthetics-related language instruction, visual features of the image from different vision encoders are fused into a single visual feature by the multi-view vision fusor, conditioned on both the image and the language instruction. Specifically, the multi-view vision fusor consists of $L$ fusion blocks. In each $l$ -th block (or layer), a group of visual features $\\{\\mathbf{X}^i \\in \\mathbb{R}^{C \\times H \\times W}\\}_{i=1}^N$ from $N$ different vision encoders are fused with a learnable query $\\mathbf{Q}^l \\in \\mathbb{R}^{C \\times H \\times W}$ . Denote the output of the $l$ -th block as $\\mathbf{O}^l$ ,", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6752d6467ca07f2f89cc6a6cc8d6fe775a94b8553597801358fccb8203658917.jpg", + "image_caption": [ + "Figure 8. Comparison of aesthetics-related visual feature discriminability of ours and CLIP [15] in existing works [8, 30, 38]." + ], + "image_footnote": [], + "bbox": [ + 535, + 330, + 888, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{Q}^{l + 1} = \\mathbf{O}^l$ , if $l\\neq 0$ . The first learnable query is generated by the language-guided query generator.", + "bbox": [ + 511, + 458, + 906, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Language-guided Query Generation. At the first layer, $\\mathbf{Q}^1$ is generated by the language-guided query generator (bottom left in Fig. 7). Specifically, we maintain a group of $M$ learnable queries $\\{\\mathbf{Q}^i\\in \\mathbb{R}^{C\\times H\\times W}\\}_{i = 1}^M$ . Given a specific language instruction, it is encoded into a text embedding $\\mathbf{T}$ , which serves as the query to dynamically generate a single query from all $M$ learnable queries with attention mechanism: $\\mathbf{Q}^1 = \\mathrm{ATTN}(\\mathbf{W}_q^{\\mathrm{T}}\\mathbf{T},\\mathbf{W}_k^{\\mathrm{T}}\\mathbf{Q}_{\\mathrm{M}},\\mathbf{W}_v^{\\mathrm{T}}\\mathbf{Q}_{\\mathrm{M}})$ where $\\mathbf{Q}_{\\mathrm{M}}$ is the concatenation of all $M$ queries, and $\\mathbf{W}_q^{\\mathrm{T}}$ , $\\mathbf{W}_k^{\\mathrm{T}}$ , $\\mathbf{W}_v^{\\mathrm{T}}$ are the projection matrixes of the query, key and value, respectively. Note that a pre-trained BERT is introduced as the text encoder, where the [CLS] token from the output is used as the text token $\\mathbf{T}$ . We avoid the CLIP text encoder to prevent bias towards the CLIP visual features in the later multimodal gating process.", + "bbox": [ + 511, + 496, + 908, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, for each language instruction that may focus on a specific aesthetic concept, a tailored query is generated from all $M$ learnable queries, which can best extract visual features related to the aesthetic concept from each vision encoder in the fusion block during inference.", + "bbox": [ + 511, + 727, + 908, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Extraction with Learnable Query. At $l$ -th layer, given a query $\\mathbf{Q}^l$ and a group of visual features $\\{\\mathbf{X}_i\\}_{i=1}^N$ from $N$ different vision encoders, each visual feature is first interpolated to the same shape: $\\mathbf{X}_n^* = \\text{Interpolate}(\\mathbf{X}_n)$ , where $\\mathbf{X}_n^* \\in \\mathbb{R}^{C \\times H \\times W}$ . Then, each of the extracted visual features from the corresponding vision encoder is given by", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24811", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{F}_n^l = \\mathrm{ATTN}(\\mathbf{W}_q^n\\mathbf{Q}^l,\\mathbf{W}_k^n\\mathbf{X}_n^*,\\mathbf{W}_v^n\\mathbf{X}_n^*)$ , where $W_{q}^{n},W_{k}^{n}$ $W_{v}^{n}$ are the projection matrixes of the query, key and value.", + "bbox": [ + 89, + 89, + 480, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multimodal Gating Network. The extracted visual features $\\{\\mathbf{F}_i^l\\}_{i = 1}^N$ are then fused into a single visual feature in the multimodal gating network. The multimodal gating network assigns different weights to each extracted visual feature conditioned on both image and language and fuses them into one single feature based on their importance. The weights of visual features are directly predicted by an MLP. Specifically, given the language embedding $\\mathbf{T}$ and a set of visual features $\\{\\mathbf{F}_i^l\\}_{i = 1}^N$ , weights are given by: $\\boldsymbol {w}^{l} = \\mathrm{MLP}(\\mathbf{T};\\mathrm{Pooling}(\\{\\mathbf{F}_{i}^{l}\\}_{i = 1}^{N}))$ , where $\\boldsymbol {w}\\in \\mathbb{R}^{N\\times 1}$ , and; means concatenation. The fused single feature is obtained by $\\hat{\\mathbf{F}}^l = \\sum_{i = 1}^N\\boldsymbol {w}_i^l\\cdot \\mathbf{F}_i^l$ . Then, $\\hat{\\mathbf{F}}^l$ is further fed forward to a transformer block with residual links. The output of the vision fusor is input to the $(l + 1)$ -th layer if $l < L$ ; otherwise, it is fed to an MLP and input to the LLM.", + "bbox": [ + 89, + 127, + 482, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark. Compared to existing works, PhotoEye differs from MoVA [41] in removing its potential bias towards CLIP features by fusing visual features with a group of learnable queries. Moreover, a language-guided query generation mechanism is introduced to extract different features towards different language instructions. When compared to Cambrian-1 [28], PhotoEye conditions both feature extraction and fusion on language, allowing it to capture more refined, instruction-aware aesthetic visual features. In contrast, Cambrian-1 is independent of instructions, potentially resulting in relatively coarser visual features.", + "bbox": [ + 89, + 354, + 482, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Benchmarking Aesthetic Visual Understanding with Professionals", + "text_level": 1, + "bbox": [ + 89, + 537, + 482, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Section 3, we highlighted a major limitation of existing datasets: annotations are typically provided by a few dozen annotators, which limits the diversity and expertise of the annotations. Even with reported training, these annotators lack the expertise of the many experienced photographers and enthusiasts. Existing benchmarks, such as Q-Bench [29] and AesBench [9], also suffer from this issue, relying on similar methods to obtain annotations.", + "bbox": [ + 89, + 582, + 482, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address the limitation and cover a wider range of scenarios, we directly draw insights from extensive photos and discussions on Reddit's PhotoCritique, a vibrant photography community. This approach allows us to generate comprehensive, diverse, and expert-level multiple-choice questions to evaluate the aesthetic visual understanding of MLLMs. Specifically, we do not generate MCQs from annotator-provided annotations; instead, we derive MCQs from photo reviews in Reddit's PhotoCritique. We refer to our new benchmark as PhotoBench. Since RPCD [18] has already collected raw data from PhotoCritique posts between 2009 and 2022, we directly design pipelines that use an LLM (GPT-4-turbo) to generate MCQs.", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. From Insights to Questions", + "text_level": 1, + "bbox": [ + 511, + 90, + 758, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The entire question-generation process consists of three steps: (1) summarization, (2) question generation, and (3) multi-stage filtering. (1) Summarization. Since the raw comments from the forum are quite noisy, empirically we find directly inputting the raw comments corresponding to an image to the LLM and prompting it to generate multiple-choice questions based on these comments does not yield ideal results. To address this issue, we first use the LLM to summarize the noisy raw comments for each image, producing a single, clean, comprehensive, and detailed photo critique. This process aligns with the pipeline shown in Fig. 3. (2) Question Generation. For each image, we input its corresponding photo critique into the LLM and instruct it to generate five multiple-choice questions grounded on the provided critique. (3) Multi-stage Filtering. This step directly impacts the quality of PhotoBench. We first selected the top 5,000 most detailed photo critiques from the initial generation step and instructed the LLM (using GPT-4 Turbo, as we empirically found it slightly outperforms GPT-4o) to generate 25,000 questions. The first filtering stage is LLM-based, primarily aimed at removing MCQs with low visual dependency—that is, questions that could likely be answered correctly by reading only the question and options without viewing the image. We show an example in Fig. 9 (1). Our method is intuitive yet effective: we provide only the question and options to the LLM (GPT-4o) without the image and filter out questions it answers correctly. This step eliminated about $70\\%$ of the items. We then instructed GPT-4o to score the remaining questions based on aesthetics relevance, visual dependency, and expertise. We selected the top 1,500 questions based on their average score. Details available in Appendix C.", + "bbox": [ + 511, + 113, + 906, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. PhotoBench", + "text_level": 1, + "bbox": [ + 511, + 604, + 643, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show detailed information of PhotoBench in Fig. 9. Compared to existing works [9, 29, 38], Photobench gains more advantages in diversity and expertise. Due to potentially limited expertise and diversity among annotators [9, 29], MCQs from existing works are often too simple. We can have a straightforward comparison in Fig. 10.", + "bbox": [ + 511, + 628, + 903, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While questions from existing works are relatively straightforward and focus on factual elements, our questions are more challenging and practical, requiring deeper aesthetics understanding and covering topics directly related to photography in practice, such as camera settings, post-processing, and photographic techniques (topics marked in blue in Fig 9. Examples are shown in Fig 9 (3,6). In Fig. 9 (3), to make the rainbow stand out more, the exposure should be reduced (slight underexposure can instead enhance color saturation). In Fig. 9 (6), we can observe that the shadows are not purely black but rather \"grayish\" (zoom in for a better view)—a common post-processing", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24812", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/52c1ce3fec77e245fa46bdda2854dfe7d7d993b788c8609c577215e43401d782.jpg", + "image_caption": [ + "What adjustments can be made to the photo to make the image cleaner and more organized? A. Add more objects in the foreground to make the image more complex and engaging B. Remove the fence to create a cleaner composition in the photo C. Increase the saturation to make it more vibrant. D. Lower the exposure Correct Answer: B Rejected for low visual dependence" + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 176, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01941c10d1270f344c95cf1ba430e77d0f1cbd8921b485aa3a0291f12a2176fc.jpg", + "image_caption": [ + "How might the orientation of the photograph better emphasize the structure of the ceiling in the church?" + ], + "image_footnote": [], + "bbox": [ + 116, + 148, + 197, + 217 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A. Landscape orientation B. Planting orientation", + "B. Portrait orientation C. Diagonal orientation", + "D. No change needed" + ], + "bbox": [ + 202, + 176, + 279, + 198 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Correct Answer: B Topic: Composition", + "bbox": [ + 202, + 207, + 380, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c34cf37499c9aafd89e79e6aed7b55d69c11c58b9c872569ce99d89ce7d4a4b8.jpg", + "image_caption": [ + "(2)", + "Which of the following aspects in the photo might be perceived as slightly jarring, requiring adjustment?" + ], + "image_footnote": [], + "bbox": [ + 390, + 85, + 509, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ccfa218b89c40aef55f107f765aefc27b26286541235dff2bfd262adf2c08af6.jpg", + "image_caption": [ + "(5)" + ], + "image_footnote": [], + "bbox": [ + 390, + 157, + 509, + 217 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A. The brightness gradient of the fog \nB. The difference between the brightness gradient", + "bbox": [ + 516, + 111, + 653, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B. The contrast between C. The depth of field used", + "bbox": [ + 516, + 119, + 594, + 127 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "D. The hue of the morning light", + "bbox": [ + 516, + 127, + 609, + 135 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Correct Answer: A Topic: Tone", + "bbox": [ + 516, + 136, + 647, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6bfb902ebc8ecbe07eec5623ce69e3026bb83f5362ed0508506f4645171ab68b.jpg", + "image_caption": [ + "How could the image be optimized to enhance the appeal of the rainbow?", + "(3)", + "What is the characterization of the processing style of the photo's shadows?" + ], + "image_footnote": [], + "bbox": [ + 661, + 85, + 714, + 147 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A. Increase the exposure time", + "bbox": [ + 720, + 111, + 812, + 117 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "B. Decrease the ISO setting C. Add more artificial lighting", + "D. Use a flash" + ], + "bbox": [ + 720, + 118, + 808, + 132 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Correct Answer: B", + "bbox": [ + 720, + 138, + 777, + 145 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Topic: Exposure", + "bbox": [ + 810, + 138, + 862, + 145 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c3192dda3e12859d68ad7d41f1d2ff56551b85c3730016dc43a09bebac1d388.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 217, + 385, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/676f2ac2afd7c1f12593d6229edca88b8fb5369edbba75195f18afe765ad9ba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 224, + 470, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e78541d024020c68ef12d8ca705443a1a85258fe3b15bd1ba630588ab939656.jpg", + "image_caption": [ + "A. Naturally soft with subtle transitions B. High contrast with distinct black and white tones C. Heavy-handed with raised blacks and crushed shadows.", + "D. Minimal processing with a high dynamic range Correct Answer: C Topic: Editing", + "(6)" + ], + "image_footnote": [], + "bbox": [ + 473, + 224, + 879, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "technique achieved by raising the shadow curve to enhance the photo's texture or create a film-like aesthetic. They suggest that PhotoBench includes more professional and detailed questions, closely aligned with practical applications such as photo post-processing and editing.", + "bbox": [ + 89, + 386, + 482, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 477, + 220, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We build PhotoEye with Vicuna-v1.5-7B as the LLM backbone, following existing works [8, 30] for fair comparison. We also use minimal data from the general domain, including LLaVA-Pretrain [15] for pretraining and LLaVA-665K [15] for instruction finetuning. We mix our data with LLaVA-665K during the finetuning stage. We use a batch size of 128 and a learning rate of $1e - 3$ for pre-training, and $2e - 5$ for finetuning. The training takes around 96 hours with 8 A100 GPUs with ZeRO2. For vision encoders, we use: CLIP-ViT-L/14 [25], DINOv2-giant [20] CoDETR-ViT-L [40] and SAM-ViT-H [11]. Details are available in the Appendix D.", + "bbox": [ + 89, + 502, + 483, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Results", + "text_level": 1, + "bbox": [ + 89, + 695, + 183, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report results of PhotoEye on existing benchmarks. Although PhotoCritique is not specifically designed for low-level vision, but PhotoEye still achieves competing results on Q-Bench when finetuned with it [29]. PhotoEye clearly outperforms existing open-source models, including both general-purpose models and those specifically fine-tuned on low-level vision and image aesthetics datasets, such as Q-Instruct [30] and AesExpert [8]. Compared to closed-source models, our lightweight approach also demonstrates competitive performance, slightly surpassing GPT-4o.", + "bbox": [ + 89, + 719, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 3, we compare our model with most competing baselines on PhotoBench, including UNIAA [38], AesEx", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae1e0dc934debf50ad3eeffca4d3bbabb554ec71cc4e0057c8008ea37b96c742.jpg", + "image_caption": [ + "Figure 9. We show one rejected example, several accepted examples, topic distribution, and some sub-topics of a subset of topics of questions in PhotoBench.", + "What aspect of the image is primarily used to enhance the aesthetic perception? A. The high level of clarity and detailed texture of the petals and stamens (Correct) B. The use of a monotone color palette C. The inclusion of multiple subjects in the background" + ], + "image_footnote": [], + "bbox": [ + 542, + 385, + 598, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1450f11a2aa286962dde5dcd968df0c7388e427692a20edcc4814fa7af4a3356.jpg", + "image_caption": [ + "What atmosphere can you feel in objects and colors? A. Joy & vitality B. Mystery & serenity (Correct) C. Sadness & despair D. Sweet & romance", + "Figure 10. Examples in related works from their original papers." + ], + "image_footnote": [], + "bbox": [ + 542, + 417, + 598, + 455 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/545906481829a67a1f95ebb83749827897e65581bdc68eb1132ad1353adc7a61.jpg", + "image_caption": [ + "Q-Instruct" + ], + "image_footnote": [], + "bbox": [ + 759, + 419, + 792, + 454 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3eaa214257140aa590bdf6105d8ff7ad65cdae5e67645faab15060726ece97d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 421, + 872, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pert [8] and Q-Instruct [30], which specially focus on low-level vision and aesthetics. Note that when reporting results in Tab. 3, we merge similar topics into a single category. Categories may overlap - for instance, a question about editing exposure would be counted in both Post-Processing and Exposure. PhotoEye notably outperforms baselines. As discussed in Section 5, PhotoBench shows substantially higher expertise and diversity than existing benchmarks, demanding stronger aesthetic understanding capabilities from models. Therefore, PhotoEye, which learns from extensive discussions by photography professionals and enthusiasts, holds a clear advantage.", + "bbox": [ + 511, + 483, + 905, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations. We show the effectiveness of our proposed vision fusor and dataset PhotoCritique in Tab. 4. Both contribute to the model. More ablations including qualitative results are available in Appendix.", + "bbox": [ + 511, + 667, + 905, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation. Providing aesthetic recommendations in conversations is actually the primary need in real-world scenarios. In addition to examples in Fig 1, We show examples in Appendix E. with diverse images and contexts.", + "bbox": [ + 511, + 732, + 905, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. See through PhotoEye", + "text_level": 1, + "bbox": [ + 511, + 801, + 720, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As multiple visual features from different vision encoders are fused in PhotoEye, we are particularly interested in their roles under different cases and how they contribute to our model. Fig. 11 presents the normalized weights of each vision encoder in PhotoEye. We first randomly col", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24813", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d9e85eb70d07b7e9bc6f18d7429eabc07e988b82f23a694ea59111a57a30b3b6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelLLM BackboneQuestion TypesQuadrants of Low-level ConcernsOverall↑
Yes-or-No↑What↑How↑Distortion↑Other↑In-context Distortion↑In-context Other↑
Random guess-50.0027.8633.3137.8938.4838.2835.8237.80
GPT-4o (2024-08-06)Proprietary Model76.7274.3367.7470.0373.3773.6877.9573.04
Qwen-VL-Max [2]Proprietary Model75.6079.4366.0973.3974.0871.0076.9273.63
Qwen-VL-Plus [2]Proprietary Model73.7769.4753.8866.2165.7263.8168.7566.04
Gemini-ProProprietary Model68.8073.7462.3466.3071.3463.9173.0968.16
Emu2-Chat [27]LLaMA-2-33B71.8167.2556.1864.7863.1963.4872.2465.28
mPLUG-Owl2 [34]LLaMA-2-7B72.1857.9656.1956.6869.2153.2972.6561.61
LLAMA-Adapter-V2 [7]LLaMA-2-7B66.1859.2952.1357.3956.2563.1664.9059.46
IDEFICS-Instruct [1]LLaMA-2-7B56.1844.6944.0242.8054.1744.7456.3348.70
SPHINX [14]LLaMA-274.1868.8162.0763.6271.7666.1276.3368.56
InternLM-XComposer-VL [36]InternLM69.4565.2760.8561.6770.1456.9175.1065.35
Qwen-VL [2]QwenLM63.0958.1956.3950.5862.7357.8973.8859.40
Otter-v1 [12]MPT-7B57.0940.7139.5542.2249.3144.0852.6546.35
LLAVA-v1.5 [15]Vicuna-v1.5-7B66.3658.1950.5149.4265.7454.6170.6158.66
InstructBLIP [6]Vicuna-v1.5-7B71.6452.6543.8148.6462.5055.5964.9056.72
Shikra [4]Vicuna-v1.5-7B65.6447.3549.0948.8359.4950.0064.0854.65
MiniGPT-4 [39]Vicuna-v1.5-13B55.8250.2240.3742.0248.3851.9761.2249.03
AesExpert [8]Vicuna-v1.5-7B73.2764.3853.7570.0373.3873.6877.9664.15
Q-Instruct [30]Vicuna-v1.5-7B76.1866.3757.6165.1867.5973.0671.5367.09
PhotoEye (Ours)Vicuna-v1.5-7B80.0176.1067.0274.3274.5977.3081.2274.50
", + "bbox": [ + 106, + 88, + 890, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4388aea944e22507467632debf86fd65eb86b09901a5a22c3b5e004912520b85.jpg", + "table_caption": [ + "Table 2. Model Performance on Q-Bench (LLVisionQA-dev). PhotoEye has a clear advantage. Best open-source model results in **bold**. Proprietary model results are quoted from Q-Bench [29]." + ], + "table_footnote": [], + "table_body": "
ModelLLM BackboneComposition↑Equipments↑Contrast↑Techniques↑Color and Tone↑Lighting↑
GPT-4o (2024-08-06)Proprietary Model68.1160.2459.0962.7563.9763.74
LLaVA-v1.6-34B [15]Hermes-2-Yi-34B47.6952.5656.8249.3365.4453.80
UNIAA [38]Vicuna-v1.5-7B29.8938.6145.4539.5640.1929.24
AesExpert [8]Vicuna-v1.5-7B51.4960.2456.8253.9961.0364.33
Q-Instruct [30]Vicuna-v1.5-7B31.0234.7740.9131.3754.3844.44
PhotoEye (Ours)Vicuna-v1.5-7B68.3261.3972.9765.6976.0077.78
ModelLLM BackboneExposure↑Post-Processing↑Aperture and Focus↑Storytelling↑Sharpness and Clarity↑Overall↑
GPT-4o (2024-08-06)Proprietary Model54.4858.5759.8769.4461.9064.12
LLaVA-v1.6-34B [15]Hermes-2-Yi-34B56.5553.3352.8766.6728.5755.68
UNIAA [38]Vicuna-v1.5-7B25.5238.5740.1330.5623.8136.87
AesExpert [8]Vicuna-v1.5-7B46.2161.9052.8772.9733.3260.01
Q-Instruct [30]Vicuna-v1.5-7B46.2146.1938.8555.5619.0543.86
PhotoEye (Ours)Vicuna-v1.5-7B69.6680.9570.7075.0061.9073.92
", + "bbox": [ + 112, + 343, + 883, + 491 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c38fd10401b6ab3780323d8706eba504165fef3f3446a0899fae2c8b23403a4e.jpg", + "table_caption": [ + "Table 3. Model Performance on PhotoBench Metrics. Best results in bold." + ], + "table_footnote": [], + "table_body": "
MethodQ-BenchPhotoEye
PhotoEye (full method)74.5073.92
w/o Multi-view Vision Fusor70.0868.83
w/o Multi-view Vision Fusor and dataset58.0433.74
", + "bbox": [ + 119, + 518, + 452, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lect 100 samples from PhotoCritique on compositional aesthetics (e.g., photographic composition, distraction, framing, etc.) and general aesthetics (e.g., storytelling, lighting, etc), respectively, as well as 100 samples for general visual tasks from LLaVA [15] training data. Then we study the averaged weight distribution (normalized) of encoders on different data. We have an interesting observation: (1) aesthetic features are extracted in the first layers, and general features are extracted in the last layers. Fig. 11 shows CLIP and DINO features are dominant regardless of domains in the last layer, which is intuitive as the recognition of objects lays the foundation in both domains. In contrast, aesthetics-related features are primarily extracted from the first layer, especially in compositional problems, where the localization or arrangement of objects are critical. In this case, CoDETR (pre-trained with detection objectives) is dominant. Tab. 5 validated it, where we only activated one encoder during inference to detect their effectiveness in photographic composition.", + "bbox": [ + 86, + 590, + 485, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Case Study. We show how different choices of vision en", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8355cce4e3db4dd5bc8c9f5950cea519ec72325f0128b006b7e9e217254484d.jpg", + "image_caption": [ + "Figure 11. Weights of vision encoders across different layers." + ], + "image_footnote": [], + "bbox": [ + 514, + 518, + 903, + 560 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/621fabbdab22ce41168d41465354b64ec60f70f8bedd6e0fddf98a866fb6126d.jpg", + "table_caption": [ + "Table 4. Ablation Results." + ], + "table_footnote": [], + "table_body": "
TopicCLIP-ViT-L/14DINOv2-giantCoDETR-ViT-LSAM-ViT-H
Composition62.6161.6164.6960.24
", + "bbox": [ + 535, + 588, + 880, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. PhotoEye on composition with different encoders.", + "bbox": [ + 532, + 619, + 883, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "coders affect model output in real-world conversation scenarios. We provide interesting examples in Appendix F.", + "bbox": [ + 511, + 641, + 905, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning of Aesthetic Concepts. Fig. 8 demonstrates how PhotoEye effectively learns discriminative aesthetics-related visual representations. In Appendix B, we conduct a deeper investigation into the learning of aesthetic concepts at feature level, which lays the foundation of high-quality aesthetic response from LLM.", + "bbox": [ + 511, + 676, + 906, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 782, + 633, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present PhotoCritique dataset, and PhotoEye, an MLLM with language-guided multi-view vision fusion for aesthetic visual understanding, along with the new benchmark PhotoBench. Extensive experiments show PhotoEye's clear advantage over existing models, advancing aesthetic visual perception in MLLMs.", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24814", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by gifts from Adobe Research and the National Science Foundation under Grants IIS-2316306 and CNS-2330215.", + "bbox": [ + 89, + 113, + 483, + 156 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 181, + 187, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 2022. 8", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 8", + "[3] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2", + "[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 8", + "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 2", + "[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 8", + "[7] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2311.07575, 2023. 8", + "[8] Yipo Huang, Xiangfei Sheng, Zhichao Yang, Quan Yuan, Zhichao Duan, Pengfei Chen, Leida Li, Weisi Lin, and Guangming Shi. AesExpert: Towards multi-modality foundation model for image aesthetics perception. arXiv preprint arXiv:2404.09624, 2024. 2, 3, 4, 5, 7, 8", + "[9] Yipo Huang, Quan Yuan, Xiangfei Sheng, Zhichao Yang, Haoning Wu, Pengfei Chen, Yuzhe Yang, Leida Li, and Weisi Lin. AesBench: An expert benchmark for multimodal large language models on image aesthetics perception. arXiv preprint arXiv:2401.08276, 2024. 2, 6", + "[10] Zhipeng Huang, Zhizheng Zhang, Yiting Lu, Zheng-Jun Zha, Zhibo Chen, and Baining Guo. Visualcritic: Making Imm's perceive visual quality like humans. arXiv preprint arXiv:2403.12806, 2024. 2" + ], + "bbox": [ + 93, + 205, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 7", + "[12] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023.8", + "[13] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Jiaming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. SPHINX: The Joint Mixing of Weights, Tasks, and Visual Embeddings for Multi-modal Large Language Models. arXiv preprint arXiv:2311.07575, 2023. 2", + "[14] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Ji-aming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2024. 3, 8", + "[15] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5, 7, 8", + "[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[17] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415. IEEE, 2012. 2", + "[18] Daniel Vera Nieto, Luigi Celona, and Clara Fernandez-Labrador. Understanding aesthetics with language: A photo critique dataset for aesthetic assessment. In NeurIPS Track on Datasets and Benchmarks, 2022. 6", + "[19] OpenAI. Gpt-4 technical report. 2023. 2", + "[20] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7", + "[21] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2", + "[22] Daiqing Qi, Handong Zhao, and Sheng Li. Easy regional contrastive learning of expressive fashion representations. Advances in Neural Information Processing Systems, 37: 20480-20509, 2024. 2", + "[23] Daiqing Qi, Handong Zhao, Zijun Wei, and Sheng Li. Tag-grounded visual instruction tuning with retrieval augmentation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 2008-2026, 2024.", + "[24] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation: A training-free approach. In European Conference on Computer Vision, pages 285-300. Springer, 2024. 2" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24815", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7", + "[26] Jian Ren, Xiaohui Shen, Zhe Lin, Radomir Mech, and David J. Foran. Personalized image aesthetics. In The IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[27] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2024. 8", + "[28] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 2, 3, 6", + "[29] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, and Weisi Lin. Q-BENCH: A benchmark for general-purpose foundation models on low-level vision. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 2, 6, 7, 8", + "[30] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, Geng Xue, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-Instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 5, 7, 8", + "[31] Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, Xiaohong Liu, Guangtao Zhai, Shiqi Wang, and Weisi Lin. Towards open-ended visual quality comparison. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2", + "[32] Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aesthetics assessment with rich attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19861-19869, 2022. 2", + "[33] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2", + "[34] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 8", + "[35] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2", + "[36] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 8", + "[37] Zhipeng Zhong, Fei Zhou, and Guoping Qiu. Aesthetically relevant image captioning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 3", + "[38] Zhaokun Zhou, Yiwei Su, Amin Zheng, Qiulin Wang, Rui Chen, Li Yuan, and Di Zhang. UNIAA: A unified multimodal image aesthetic assessment baseline and benchmark. arXiv preprint arXiv:2404.09619, 2024. 2, 3, 5, 6, 7, 8", + "[39] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8", + "[40] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6758, 2023. 7", + "[41] Zhuofan Zong, Bingqi Ma, Dazhong Shen, Guanglu Song, Hao Shao, Dongzhi Jiang, Hongsheng Li, and Yu Liu. Mova: Adapting mixture of vision experts to multimodal context. In Proceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 6" + ], + "bbox": [ + 516, + 92, + 906, + 500 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24816", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_model.json b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d950a0d8897fa371dd202a27233fdd10b6f48a2d --- /dev/null +++ b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_model.json @@ -0,0 +1,2574 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.856, + 0.2 + ], + "angle": 0, + "content": "The Photographer's Eye: Teaching Multimodal Large Language Models to See and Critique like Photographers" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.226, + 0.765, + 0.262 + ], + "angle": 0, + "content": "Daiqing Qi\\(^{1,2}\\), Handong Zhao\\(^{2}\\), Jing Shi\\(^{2}\\), Simon Jenni\\(^{2}\\), Yifei Fan\\(^{2}\\), Franck Dernoncourt\\(^{2}\\), Scott Cohen\\(^{2}\\), Sheng Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.268, + 0.624, + 0.288 + ], + "angle": 0, + "content": "1University of Virginia, 2Adobe" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.301, + 0.273, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.386, + 0.231, + 0.398 + ], + "angle": 0, + "content": "User: Are there any weaknesses in this photograph?" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.406, + 0.263, + 0.424 + ], + "angle": 0, + "content": "Q-Instruct: The image features a beautiful sunset with the sun's rays shining through the trees." + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.427, + 0.267, + 0.45 + ], + "angle": 0, + "content": "The scene captures the vibrant colors of the sunset, creating a serene and picturesque atmosphere. The hillside is covered in lush green trees, providing a sense of tranquility and natural beauty." + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.453, + 0.253, + 0.471 + ], + "angle": 0, + "content": "The combination of the sunset and the verdant landscape creates a captivating and visually pleasing view." + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.304, + 0.397, + 0.375 + ], + "angle": 0, + "content": "AesExpert: This image looks very clear. The colors in this image are rich and vibrant, with a good color combination, making the picture appear more vivid and bright. The lighting in this image is strong, creating a warm and soft atmosphere, with a strong contrast of dark and light. The composition of this image is a high horizon composition, highlighting the texture in the lower half of the picture. Therefore, this image looks quite beautiful." + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.378, + 0.398, + 0.476 + ], + "angle": 0, + "content": "PhotoEye: The composition of this image could be improved by straightening the horizon to create a more balanced and visually appealing frame. Additionally, the presence of lens flare may be distracting and detract from the appearance of the image. To address this, the photographer could consider using a lens hood or adjusting their shooting angle to minimize the effect of the flare. Furthermore, the tilted horizon could be corrected to create a more stable and visually appealing atmosphere. However, the amount of the image could be more nuanced, with a more subtle transition between highlights and shadows to create a more natural and inviting atmosphere." + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.302, + 0.579, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.389, + 0.577, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.306, + 0.718, + 0.314 + ], + "angle": 0, + "content": "User: Are there any distractors in this image?" + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.319, + 0.644, + 0.327 + ], + "angle": 0, + "content": "Q-Instruct: No" + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.318, + 0.696, + 0.326 + ], + "angle": 0, + "content": "AesExpert:" + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.306, + 0.884, + 0.314 + ], + "angle": 0, + "content": "User: Please describe the exposure of the photo" + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.333, + 0.726, + 0.383 + ], + "angle": 0, + "content": "GPT-4o: 1. The Chain and Metal Poles: These elements draw attention because they are close to the subjects and create visual lines in the foreground, potentially competing for focus with the main subjects (the two individual). 2. Darker Darker Colors: The person in the yellow jacket stands out significantly against the darker colors of the other person and the background..." + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.39, + 0.726, + 0.414 + ], + "angle": 0, + "content": "PhotoEye: The photo effectively captures a romantic moment between the couple, but there are some distracting elements that take away from the intimacy of the scene." + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.417, + 0.728, + 0.473 + ], + "angle": 0, + "content": "The lines and wires in the background are a notable distraction, drawing the viewer's attention away from the couple and disrupting the visual flow. The image of the photographer could consider removing or minimizing these distractions to allow the viewer to focus on the tender moment between the two characters. This distraction would not be the most flattering choice, which could also detract from the overall impact of the image." + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.32, + 0.822, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.322, + 0.885, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.4, + 0.822, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.409, + 0.885, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.486, + 0.88, + 0.626 + ], + "angle": 0, + "content": "Figure 1. Examples on our model (PhotoEye), existing MLLMs tailored for low-level vision or aesthetics, and GPT-4o (2024-08-06). The left example and the middle top example highlight a notable limitation of existing open-source MLLMs for low-level vision and aesthetics: insufficient coverage of visual aesthetics. When they fail to identify issues, they either provide positive aspects or claim there are no issues, significantly limiting their usefulness in real-world scenarios. The middle bottom example shows existing MLLMs' lack of expertise in photography: lowering exposure properly can instead enhance colors of objects, so B is correct. While other models, including GPT-4o, made mistakes, our model is correct. The right example reveals another clear limitation of existing open-source MLLMs: their vision encoders are insensitive to low-level vision and aesthetics. In a series of increasingly overexposed photos (2-nd, 3-rd, and 4-th), PhotoEye's vision modules, more attuned to low-level and aesthetic features, identified overexposure by the 2-nd photo, while other models recognized it only when the photos were severely overexposed (i.e., 3-rd and 4-th images). High-quality aesthetic content is highlighted." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.64, + 0.327, + 0.656 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.664, + 0.46, + 0.709 + ], + "angle": 0, + "content": "\"While editing directly from life, photographers have found it too difficult to see simultaneously both the blue and the sky.\"" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.71, + 0.449, + 0.726 + ], + "angle": 0, + "content": "John Szarkowski, William Eggleston's Guide" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.857 + ], + "angle": 0, + "content": "Photographer and curator, Szarkowski insightfully revealed one of the notable gaps between general and aesthetic visual understanding: while the former focuses on identifying the factual element in an image (sky), the latter transcends such object identification, viewing it instead as an aesthetic component—a pure color block (blue). Such fundamental distinctions between general (detection, localization, etc.) and aesthetic (color, lighting, composition, etc.)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Part of this work was completed during Daiqing's internship at Adobe. \n1William Eggleston is an American photographer, widely recognized as a pioneering figure in color photography." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.908, + 0.899 + ], + "angle": 0, + "content": "visual understanding present a significant challenge for Multimodal Large Language Models (MLLMs). Although some recent works have made initial explorations, they are often limited to general and basic aesthetic commonsense. As a result, they frequently fall short in real-world scenarios (Fig. 1), which require extensive expertise—including photographic techniques, photo pre/post-processing knowledge, and more, to provide a detailed analysis and description. To fundamentally enhance the aesthetics understanding of MLLMs, we first introduce a novel dataset, PhotoCritique, derived from extensive discussions among professional photographers and enthusiasts, and characterized by the large scale, expertise, and diversity. Then, to better learn visual aesthetics from PhotoCritique, we further propose a novel model, PhotoEye, featuring a language-guided multi-view vision fusion mechanism to understand image aesthetics from multiple perspectives. Finally, we present a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24807" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "novel benchmark, PhotoBench, a comprehensive and professional benchmark for aesthetic visual understanding. On existing benchmarks and PhotoBench, our model demonstrates clear advantages over existing models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.167, + 0.223, + 0.182 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.373 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs) [16, 19] have succeeded in general visual tasks but still struggle with a critical ability in human visual perception: aesthetic visual understanding, which is desired in various applications like image recommendation, editing, or generation. The challenge lies in the fundamental distinction between general and aesthetic visual understanding: while the former focuses on factual elements like identification of objects, the latter requires nuanced perception of aesthetic elements such as chromatic properties, or compositional principles. We attribute the shortcomings of MLLMs in aesthetic understanding to two primary factors: dataset and model." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.374, + 0.483, + 0.584 + ], + "angle": 0, + "content": "Dataset: Current datasets such as Q-Instruct [29] and AesExpert [9], made initial explorations in low-level vision or visual aesthetics with MLLM but suffer from limitations in scale, expertise, and diversity. With around 20K images and 58K-88K image annotations, these datasets are insufficient for comprehensive aesthetic understanding, showing clear deficiencies in real-world scenarios, where images vary widely in subjects, scenes, styles, environments, and techniques, requiring larger datasets to ensure coverage. Additionally, their annotations come from small (typically tens of annotators), fixed groups of annotators, limiting diversity and expertise. Even if the annotators are reportedly trained, their expertise remains limited compared to professional photographers and enthusiasts." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.585, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Model: When adapting to aesthetic visual understanding, existing approaches [8, 10, 30, 38] typically finetune existing MLLMs like LLaVA [16] on customized datasets, ignoring the unique requirements of visual features for aesthetic visual understanding. Comprehensive understanding of image aesthetics requires multi-view information: for example, high-level features are essential for interpreting narrative and storytelling, while detailed spatial information is critical for understanding composition. Similarly, low-level details are required for perception of color and lighting, etc." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the challenges, we first introduce the PhotoCritique dataset, featuring (1) large scale: containing over 450K images with detailed aesthetic descriptions, and around 20 times the size of existing datasets [8, 30]; and (2) expertise and diversity: the aesthetic descriptions are derived from extensive discussions among large groups of professional photographers and enthusiasts online, rather than a fixed small number of annotators as in existing works [8, 30]. Subsequently, given the notable different focuses in general and aesthetic visual understanding, we further introduce PhotoEye, an MLLM tailored for aes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "thetic visual understanding. PhotoEye features a language-guided multi-view vision fusion mechanism to effectively learn aesthetic knowledge from PhotoCritique, enabling it to understand image aesthetics from multiple perspectives for various aesthetics-related questions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Finally, to provide a comprehensive evaluation of our model and existing MLLMs on aesthetic visual understanding, we introduce our new benchmark PhotoBench. PhotoBench draws its questions from in-depth discussions among professional photographers and enthusiasts, covering 284 sub-topics in photography, making it notably more professional and diverse than existing works. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.289, + 0.905, + 0.365 + ], + "angle": 0, + "content": "- We introduce PhotoCritique, the first large-scale, professional, and diverse dataset for aesthetic visual understanding, consisting of 450K images with 2.63M instruction-tuning pairs, sourced from extensive discussions among professional photographers and enthusiasts." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.366, + 0.905, + 0.409 + ], + "angle": 0, + "content": "- We propose PhotoEye, an MLLM specifically designed for aesthetic understanding, featuring a language-guided multi-view vision fusion mechanism." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.411, + 0.905, + 0.456 + ], + "angle": 0, + "content": "- We develop PhotoBench, a professional benchmark for aesthetic visual understanding, offering 284 photography sub-topics and insights from photography professionals." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.457, + 0.905, + 0.516 + ], + "angle": 0, + "content": "- We conduct a comprehensive study on aesthetic visual understanding with our model and existing MLLMs, revealing the limitations of existing MLLMs and demonstrating how we overcome them through extensive experiments." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.289, + 0.905, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.529, + 0.663, + 0.544 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.905, + 0.704 + ], + "angle": 0, + "content": "Multimodal Large Language Models. MLLMs [3, 16] are evolving rapidly recently, demonstrating advanced capabilities in image-text dialogue through enhanced alignment and fine-tuning. Further studies [2, 5, 6, 28, 33] have strengthened MLLMs by increasing the data scale or improving the backbone model. Furthermore, with customized data, a branch of works [4, 21, 33, 35] extend MLLMs' capabilities to referring and grounding. However, despite these advancements in general visual tasks, the aesthetic visual understanding of MLLMs remains largely underexplored." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.871 + ], + "angle": 0, + "content": "Understanding Image Aesthetics with MLLM. Traditional Image Aesthetics Assessment (IAA) [17, 26, 32] is often formalized as a regression task, where the model learns to predict the score of the aesthetic quality of images. One notable limitation is the lack of interpretability, which is critical in practice because we also want to understand why the image is good or bad. Recent advances in MLLM make it possible: a few recent works [8-10, 29-31, 38] explore interpretable aesthetic understanding with MLLMs. However, they often struggle with the limited data scale or lack diverse, professionally curated aesthetic descriptions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Improving Visual Perception in MLLM. Recently, several studies [13, 22-24, 28, 41] have investigated back" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.091, + 0.189, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.095, + 0.431, + 0.104 + ], + "angle": 0, + "content": "#1: Very good one, even if the unsharp mask is too strong." + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.108, + 0.461, + 0.132 + ], + "angle": 0, + "content": "#2: This is gorgeous. The use of parallel lines and the vanishing point really leads us into this photo. And the mixture of warm and cool colors is very pleasing." + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.135, + 0.468, + 0.166 + ], + "angle": 0, + "content": "#3: This appearance is particularly prominent in the lower left foreground foliage which is also quite bright and thus accentuates the over-sharpened look. Perhaps it is jpeg artifacting, if I look in the upper branches of the tree I can see a cross-hatch pattern..." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.214 + ], + "angle": 0, + "content": "Figure 2. An example from the dpchallenge platform. 3 of the total 64 comments for the photo are presented for illustration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.224, + 0.484, + 0.421 + ], + "angle": 0, + "content": "bone vision modules for MLLM. While SPHINX [14] simply concatenates features from different encoders, Cambrian [28] and MoVA [41] merge different visual features into a single one. However, the fusion process in MoVA is based on the CLIP feature, which may introduce bias, and the feature extraction (prior to fusion) lacks language-based control. Cambrian [28] employs query-based attention, however, visual features from different encoders are simply averaged during fusion, and the queries are independent of language instructions. To address these limitations, we propose a new MLLM, PhotoEye, featuring a language-guided multi-view vision fusion mechanism to effectively learn aesthetic knowledge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.433, + 0.484, + 0.467 + ], + "angle": 0, + "content": "3. Learning with Photography Enthusiasts and Professionals" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.673 + ], + "angle": 0, + "content": "One key reason existing MLLMs struggle with aesthetic visual understanding in real-world scenarios is the training data limitations: (1) Data Scale: Current datasets [8, 30] (\\(\\sim\\)20K images) lack sufficient coverage of photography genres, styles, themes, etc. High-quality aesthetic descriptions are also limited, with typically 57K-88K detailed descriptions in existing works [8, 30, 38], while our dataset provides over 450K images and detailed descriptions. (2) Description Quality: Most datasets [8, 30] rely on a small, fixed group of annotators, limiting expertise and diversity. In contrast, our PhotoCritique dataset draws insights from hundreds of thousands of photographers and enthusiasts, yielding more professional, nuanced feedback." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.466, + 0.699 + ], + "angle": 0, + "content": "3.1. Mining from Digital Photography Platforms" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To address the limitations, we avoided using annotators and instead drew insights from a broad range of online communities. The Digital Photography Challenge (DPC) is an online platform, where users submit their works across various themes and genres and receive constructive feedback from other photographers or enthusiasts (Fig. 2). The wealth of high-quality discussions paired with images on DPC provides valuable resources for MLLMs to learn visual aesthetics. However, directly using this data for supervised fine-tuning is impractical due to the significant noise present in the raw comments. To address this issue, we developed strategies to analyze, summarize, and refine raw comments into a single, well-crafted critique for each image. First," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.907, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.147, + 0.907, + 0.174 + ], + "angle": 0, + "content": "Figure 3. Data Generation Pipeline. The generation of one critique from a group of comments of one image is shown as an example." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.182, + 0.905, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.305, + 0.639, + 0.312 + ], + "angle": 0, + "content": "PhotoCritique (Aesthetic Conversation)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.317, + 0.904, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.388, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Figure 4. Top: Examples of PhotoCritique. Bottom: Comparison with examples in Q-Instruct [30]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.789 + ], + "angle": 0, + "content": "we need to collect raw data from dpchallenge.com online. Since existing works [37] already collected raw data prior to 2022 from the website, we can skip the data scraping process and focus on data processing. Given a post from the DPC platform (comprising one uploaded image and a series of user comments, shown in Fig. 2), we design a multi-step pipeline to generate a single unified photo critique using prompt-based methods from comments only without the image. Fig. 3 briefly illustrates the pipeline with the example of generating a single critique from a set of comments of one image. We first instruct the LLM to analyze and summarize the raw comments, Then the LLM is asked to integrate the comments into a cohesive and unified photo critique. While LLMs possess strong reasoning and comprehension abilities, the immediate output critique is sometimes sparse in information or misleading. Therefore, we designed a filtering strategy to eliminate low-quality data with a smaller LLM for improved efficiency. During the filtering stage, generated photo critiques are input into the smaller LLM, and prompt-based strategies are used to detect whether the generated photo critique conveys enough aesthetics-related information. Only critiques that pass this filtering stage are accepted. We term this portion of data as aesthetic description. Details available in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "From General Photo Critiques to Diverse Instruction-tuning Pairs. In the previous step, we generated photo critiques from DPC's raw comments, which provide overall aesthetic reviews of the given images. In this step, we further utilize LLMs to generate multiple high-quality, diverse conversations addressing various aspects of image aesthetics, such as lighting, composition, color, emotion, narrative," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24809" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.085, + 0.482, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.165, + 0.483, + 0.207 + ], + "angle": 0, + "content": "Figure 5. Comparison of description length Distributions of aesthetic comments in PhotoCritique and Q-Instruct. Scale of PhotoCritique is Blue and that of Q-Instruct is Red." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.215, + 0.48, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.361, + 0.483, + 0.389 + ], + "angle": 0, + "content": "Figure 6. Top: Distribution of Top 40 categories. Bottom: Examples of photos from a subset of categories." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.392, + 0.483, + 0.634 + ], + "angle": 0, + "content": "and so on, as well as extensive photographic techniques and pre/post-processing skills. Another critical reason for this step is that raw comments are sometimes too lengthy, which can lead to certain aesthetic aspects being overlooked in the LLM-generated critique. By specifically prompting about these aesthetic aspects, the LLM can effectively summarize essential information related to each aspect from the lengthy raw comments. Similarly, after generating all the question-answer pairs, we apply prompt-based filtering strategies to reject less informative pairs, accepting only the high-quality samples. We term this portion of data as aesthetic conversation. Additionally, we also generate MCQs from photo critiques with LLMs. Specifically, each critique is input into the LLM with prompts to generate five MCQs based on the ground truth provided in it. We term this portion of data as aesthetic VQA. Details available in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.645, + 0.238, + 0.661 + ], + "angle": 0, + "content": "3.2. PhotoCritique" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Overview. Combining all above, we build PhotoCritique, the first large-scale dataset for expert-level image aesthetics understanding. PhotoCritique consists of three parts: aesthetic description, aesthetic conversation, and aesthetic VQA. Examples of each are shown in Fig. 4 (Top). The three parts contain 450K, 1.9M, and 250K samples, respectively. Born from a large volume of shared images and discussions by photography enthusiasts and professionals on online photography forums, PhotoCritique covers a wide range of over 70 photographic categories, encompassing diverse scenes and subjects. Fig. 6 shows the distribution of the top 40 categories along with several image examples. The extensive scale and diversity enhance the model's effectiveness in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparison with Existing Datasets. PhotoCritique rep" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.089, + 0.898, + 0.145 + ], + "angle": 0, + "content": "
Dataset#Image#Annotations#Training Samples#Avg. Length
UNIAA57K57K57K32.8
AesExpert*21K88K--
Q-Instruct18K58K66K46.4
PhotoCritique450K450K2.4M65.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.155, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Table 1. Statistics of samples attributed to detailed aesthetic descriptions and conversations across different datasets. Datasets with * are currently not available." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.206, + 0.907, + 0.903 + ], + "angle": 0, + "content": "resents a fundamental improvement over existing works in both quantity and quality. Data Scale: we compares PhotoCritique with existing datasets by showing statistics for data related to detailed aesthetic descriptions and conversations in Tab. 1. This includes the number of source images, the annotations for these images, and the final count of instruction-tuning pairs generated from these annotations using LLMs. Additionally, we calculate the average length of aesthetic comments (including both descriptions and conversations) in the training data. Tab. 1 focuses on training data statistics for aesthetic comments, as these are essential in practical applications where users need detailed feedback on why their photos are strong or weak and how to improve them. In contrast, Multiple Choice Questions (MCQ) are relatively limited in this regard. Besides, aesthetic comments better reflect the information density of the dataset than MCQs, as generating MCQs from a detailed comment is easy and can result in numerous MCQs of varying quality depending on the input prompt. Therefore, the number of MCQs serves only as a relatively rough indicator of the information density, although the number of MCQs in PhotoCritique is also the largest. As shown in Tab. 1, PhotoCritique demonstrates a notable improvement across all metrics compared to existing datasets. Data Quality: PhotoCritique is more professional, and diverse. Unlike most existing works [8, 30], which rely on a small group (usually a few dozen) of annotators to label data, our dataset is sourced from images and discussion from over 107,000 photography enthusiasts and professionals on online photography forums. Fig. 2 provides an intuitive comparison between our dataset and Q-Instruct, with examples of Q-Instruct in its original paper. While Q-Instruct primarily focuses on basic low-level visual facts, PhotoCritique not only addresses low-level details but also considers how these elements impact image aesthetics and ways to improve them. Additionally, we include discussions on photography techniques (e.g., the first example in the aesthetic description covers flash usage and lighting) and the overall impression conveyed by the photo. The comprehensiveness of aesthetic comments in PhotoCritique is also reflected in the average length: Fig. 5 shows the distribution of aesthetic comment lengths in PhotoCritique and Q-Instruct, with PhotoCritique containing a large number of detailed descriptions. Tab. 1 also shows that the average length of aesthetic comments in PhotoCritique is notably greater than that of existing datasets. More examples in Appendix A." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24810" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.082, + 0.878, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.286, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 7. Framework of PhotoEye. Given an input photo and a language instruction, visual features from different visual encoders are fused in our vision fusion with language guidance, then fed into the LLM with language embeddings to generate LLM response." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.332, + 0.446, + 0.35 + ], + "angle": 0, + "content": "4. Improving Visual Aesthetics Perception" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.358, + 0.306, + 0.374 + ], + "angle": 0, + "content": "4.1. See like Photographers" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.382, + 0.485, + 0.699 + ], + "angle": 0, + "content": "In addition to limitations in existing datasets, Open-source MLLMs face another notable challenge: the limited sensitivity of the vision encoder to aesthetic elements. Most existing works [8, 30] use CLIP as the vision encoder, which, however, is pre-trained with high-level image-text alignment in general domains and is relatively less effective at capturing aesthetic elements. Fig. 1 (right) illustrates this issue, where Q-Instruct and AesExpert only report 'overexposure' when the image is severely overexposed. In Fig. 8, we quantitatively validate it by calculating the discriminability (defined as the average distance between these embeddings) of aesthetic features from different vision encoders. Details are available in Appendix B. It shows that CLIP is less effective in discriminating towards aesthetics-related features as they are more densely clustered. To handle the diversity of aesthetics understanding, from high-level elements like emotion and storytelling to low-level features such as lighting, color, and composition, we propose PhotoEye, a MLLM with a language-guided multi-view vision fusor, which enhances visual perception by fusing multiple vision encoders pre-trained on complementary tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.203, + 0.727 + ], + "angle": 0, + "content": "4.2. PhotoEye" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Architecture. PhotoEye is shown in Fig. 7. It consists of a multi-view vision fusor and a backbone LLM. Given an image and an aesthetics-related language instruction, visual features of the image from different vision encoders are fused into a single visual feature by the multi-view vision fusor, conditioned on both the image and the language instruction. Specifically, the multi-view vision fusor consists of \\(L\\) fusion blocks. In each \\(l\\)-th block (or layer), a group of visual features \\(\\{\\mathbf{X}^i \\in \\mathbb{R}^{C \\times H \\times W}\\}_{i=1}^N\\) from \\(N\\) different vision encoders are fused with a learnable query \\(\\mathbf{Q}^l \\in \\mathbb{R}^{C \\times H \\times W}\\). Denote the output of the \\(l\\)-th block as \\(\\mathbf{O}^l\\)," + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.332, + 0.889, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.424, + 0.907, + 0.452 + ], + "angle": 0, + "content": "Figure 8. Comparison of aesthetics-related visual feature discriminability of ours and CLIP [15] in existing works [8, 30, 38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.492 + ], + "angle": 0, + "content": "\\(\\mathbf{Q}^{l + 1} = \\mathbf{O}^l\\) , if \\(l\\neq 0\\) . The first learnable query is generated by the language-guided query generator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.909, + 0.727 + ], + "angle": 0, + "content": "Language-guided Query Generation. At the first layer, \\(\\mathbf{Q}^1\\) is generated by the language-guided query generator (bottom left in Fig. 7). Specifically, we maintain a group of \\(M\\) learnable queries \\(\\{\\mathbf{Q}^i\\in \\mathbb{R}^{C\\times H\\times W}\\}_{i = 1}^M\\). Given a specific language instruction, it is encoded into a text embedding \\(\\mathbf{T}\\), which serves as the query to dynamically generate a single query from all \\(M\\) learnable queries with attention mechanism: \\(\\mathbf{Q}^1 = \\mathrm{ATTN}(\\mathbf{W}_q^{\\mathrm{T}}\\mathbf{T},\\mathbf{W}_k^{\\mathrm{T}}\\mathbf{Q}_{\\mathrm{M}},\\mathbf{W}_v^{\\mathrm{T}}\\mathbf{Q}_{\\mathrm{M}})\\) where \\(\\mathbf{Q}_{\\mathrm{M}}\\) is the concatenation of all \\(M\\) queries, and \\(\\mathbf{W}_q^{\\mathrm{T}}\\), \\(\\mathbf{W}_k^{\\mathrm{T}}\\), \\(\\mathbf{W}_v^{\\mathrm{T}}\\) are the projection matrixes of the query, key and value, respectively. Note that a pre-trained BERT is introduced as the text encoder, where the [CLS] token from the output is used as the text token \\(\\mathbf{T}\\). We avoid the CLIP text encoder to prevent bias towards the CLIP visual features in the later multimodal gating process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.909, + 0.804 + ], + "angle": 0, + "content": "In this way, for each language instruction that may focus on a specific aesthetic concept, a tailored query is generated from all \\( M \\) learnable queries, which can best extract visual features related to the aesthetic concept from each vision encoder in the fusion block during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Feature Extraction with Learnable Query. At \\(l\\)-th layer, given a query \\(\\mathbf{Q}^l\\) and a group of visual features \\(\\{\\mathbf{X}_i\\}_{i=1}^N\\) from \\(N\\) different vision encoders, each visual feature is first interpolated to the same shape: \\(\\mathbf{X}_n^* = \\text{Interpolate}(\\mathbf{X}_n)\\), where \\(\\mathbf{X}_n^* \\in \\mathbb{R}^{C \\times H \\times W}\\). Then, each of the extracted visual features from the corresponding vision encoder is given by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24811" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.09, + 0.482, + 0.123 + ], + "angle": 0, + "content": "\\(\\mathbf{F}_n^l = \\mathrm{ATTN}(\\mathbf{W}_q^n\\mathbf{Q}^l,\\mathbf{W}_k^n\\mathbf{X}_n^*,\\mathbf{W}_v^n\\mathbf{X}_n^*)\\) , where \\(W_{q}^{n},W_{k}^{n}\\) \\(W_{v}^{n}\\) are the projection matrixes of the query, key and value." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.128, + 0.483, + 0.355 + ], + "angle": 0, + "content": "Multimodal Gating Network. The extracted visual features \\(\\{\\mathbf{F}_i^l\\}_{i = 1}^N\\) are then fused into a single visual feature in the multimodal gating network. The multimodal gating network assigns different weights to each extracted visual feature conditioned on both image and language and fuses them into one single feature based on their importance. The weights of visual features are directly predicted by an MLP. Specifically, given the language embedding \\(\\mathbf{T}\\) and a set of visual features \\(\\{\\mathbf{F}_i^l\\}_{i = 1}^N\\), weights are given by: \\(\\boldsymbol {w}^{l} = \\mathrm{MLP}(\\mathbf{T};\\mathrm{Pooling}(\\{\\mathbf{F}_{i}^{l}\\}_{i = 1}^{N}))\\), where \\(\\boldsymbol {w}\\in \\mathbb{R}^{N\\times 1}\\), and; means concatenation. The fused single feature is obtained by \\(\\hat{\\mathbf{F}}^l = \\sum_{i = 1}^N\\boldsymbol {w}_i^l\\cdot \\mathbf{F}_i^l\\). Then, \\(\\hat{\\mathbf{F}}^l\\) is further fed forward to a transformer block with residual links. The output of the vision fusor is input to the \\((l + 1)\\)-th layer if \\(l < L\\); otherwise, it is fed to an MLP and input to the LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.522 + ], + "angle": 0, + "content": "Remark. Compared to existing works, PhotoEye differs from MoVA [41] in removing its potential bias towards CLIP features by fusing visual features with a group of learnable queries. Moreover, a language-guided query generation mechanism is introduced to extract different features towards different language instructions. When compared to Cambrian-1 [28], PhotoEye conditions both feature extraction and fusion on language, allowing it to capture more refined, instruction-aware aesthetic visual features. In contrast, Cambrian-1 is independent of instructions, potentially resulting in relatively coarser visual features." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.573 + ], + "angle": 0, + "content": "5. Benchmarking Aesthetic Visual Understanding with Professionals" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.703 + ], + "angle": 0, + "content": "In Section 3, we highlighted a major limitation of existing datasets: annotations are typically provided by a few dozen annotators, which limits the diversity and expertise of the annotations. Even with reported training, these annotators lack the expertise of the many experienced photographers and enthusiasts. Existing benchmarks, such as Q-Bench [29] and AesBench [9], also suffer from this issue, relying on similar methods to obtain annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the limitation and cover a wider range of scenarios, we directly draw insights from extensive photos and discussions on Reddit's PhotoCritique, a vibrant photography community. This approach allows us to generate comprehensive, diverse, and expert-level multiple-choice questions to evaluate the aesthetic visual understanding of MLLMs. Specifically, we do not generate MCQs from annotator-provided annotations; instead, we derive MCQs from photo reviews in Reddit's PhotoCritique. We refer to our new benchmark as PhotoBench. Since RPCD [18] has already collected raw data from PhotoCritique posts between 2009 and 2022, we directly design pipelines that use an LLM (GPT-4-turbo) to generate MCQs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.759, + 0.108 + ], + "angle": 0, + "content": "5.1. From Insights to Questions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.907, + 0.598 + ], + "angle": 0, + "content": "The entire question-generation process consists of three steps: (1) summarization, (2) question generation, and (3) multi-stage filtering. (1) Summarization. Since the raw comments from the forum are quite noisy, empirically we find directly inputting the raw comments corresponding to an image to the LLM and prompting it to generate multiple-choice questions based on these comments does not yield ideal results. To address this issue, we first use the LLM to summarize the noisy raw comments for each image, producing a single, clean, comprehensive, and detailed photo critique. This process aligns with the pipeline shown in Fig. 3. (2) Question Generation. For each image, we input its corresponding photo critique into the LLM and instruct it to generate five multiple-choice questions grounded on the provided critique. (3) Multi-stage Filtering. This step directly impacts the quality of PhotoBench. We first selected the top 5,000 most detailed photo critiques from the initial generation step and instructed the LLM (using GPT-4 Turbo, as we empirically found it slightly outperforms GPT-4o) to generate 25,000 questions. The first filtering stage is LLM-based, primarily aimed at removing MCQs with low visual dependency—that is, questions that could likely be answered correctly by reading only the question and options without viewing the image. We show an example in Fig. 9 (1). Our method is intuitive yet effective: we provide only the question and options to the LLM (GPT-4o) without the image and filter out questions it answers correctly. This step eliminated about \\(70\\%\\) of the items. We then instructed GPT-4o to score the remaining questions based on aesthetics relevance, visual dependency, and expertise. We selected the top 1,500 questions based on their average score. Details available in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.644, + 0.621 + ], + "angle": 0, + "content": "5.2. PhotoBench" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.72 + ], + "angle": 0, + "content": "We show detailed information of PhotoBench in Fig. 9. Compared to existing works [9, 29, 38], Photobench gains more advantages in diversity and expertise. Due to potentially limited expertise and diversity among annotators [9, 29], MCQs from existing works are often too simple. We can have a straightforward comparison in Fig. 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "While questions from existing works are relatively straightforward and focus on factual elements, our questions are more challenging and practical, requiring deeper aesthetics understanding and covering topics directly related to photography in practice, such as camera settings, post-processing, and photographic techniques (topics marked in blue in Fig 9. Examples are shown in Fig 9 (3,6). In Fig. 9 (3), to make the rainbow stand out more, the exposure should be reduced (slight underexposure can instead enhance color saturation). In Fig. 9 (6), we can observe that the shadows are not purely black but rather \"grayish\" (zoom in for a better view)—a common post-processing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.087, + 0.178, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.089, + 0.387, + 0.147 + ], + "angle": 0, + "content": "What adjustments can be made to the photo to make the image cleaner and more organized? A. Add more objects in the foreground to make the image more complex and engaging B. Remove the fence to create a cleaner composition in the photo C. Increase the saturation to make it more vibrant. D. Lower the exposure Correct Answer: B Rejected for low visual dependence" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.149, + 0.199, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.158, + 0.373, + 0.17 + ], + "angle": 0, + "content": "How might the orientation of the photograph better emphasize the structure of the ceiling in the church?" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.177, + 0.281, + 0.185 + ], + "angle": 0, + "content": "A. Landscape orientation B. Planting orientation" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.185, + 0.273, + 0.192 + ], + "angle": 0, + "content": "B. Portrait orientation C. Diagonal orientation" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.192, + 0.272, + 0.199 + ], + "angle": 0, + "content": "D. No change needed" + }, + { + "type": "list", + "bbox": [ + 0.203, + 0.177, + 0.281, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.208, + 0.382, + 0.216 + ], + "angle": 0, + "content": "Correct Answer: B Topic: Composition" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.087, + 0.51, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.15, + 0.402, + 0.156 + ], + "angle": 0, + "content": "(2)" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.158, + 0.511, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.089, + 0.655, + 0.108 + ], + "angle": 0, + "content": "Which of the following aspects in the photo might be perceived as slightly jarring, requiring adjustment?" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.112, + 0.655, + 0.12 + ], + "angle": 0, + "content": "A. The brightness gradient of the fog \nB. The difference between the brightness gradient" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.12, + 0.596, + 0.128 + ], + "angle": 0, + "content": "B. The contrast between C. The depth of field used" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.128, + 0.61, + 0.136 + ], + "angle": 0, + "content": "D. The hue of the morning light" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.137, + 0.648, + 0.145 + ], + "angle": 0, + "content": "Correct Answer: A Topic: Tone" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.087, + 0.715, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.09, + 0.869, + 0.102 + ], + "angle": 0, + "content": "How could the image be optimized to enhance the appeal of the rainbow?" + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.112, + 0.813, + 0.118 + ], + "angle": 0, + "content": "A. Increase the exposure time" + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.119, + 0.809, + 0.126 + ], + "angle": 0, + "content": "B. Decrease the ISO setting C. Add more artificial lighting" + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.126, + 0.765, + 0.133 + ], + "angle": 0, + "content": "D. Use a flash" + }, + { + "type": "list", + "bbox": [ + 0.721, + 0.119, + 0.809, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.14, + 0.779, + 0.146 + ], + "angle": 0, + "content": "Correct Answer: B" + }, + { + "type": "text", + "bbox": [ + 0.812, + 0.139, + 0.863, + 0.146 + ], + "angle": 0, + "content": "Topic: Exposure" + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.149, + 0.683, + 0.155 + ], + "angle": 0, + "content": "(3)" + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.157, + 0.875, + 0.17 + ], + "angle": 0, + "content": "What is the characterization of the processing style of the photo's shadows?" + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.176, + 0.877, + 0.199 + ], + "angle": 0, + "content": "A. Naturally soft with subtle transitions B. High contrast with distinct black and white tones C. Heavy-handed with raised blacks and crushed shadows." + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.2, + 0.868, + 0.216 + ], + "angle": 0, + "content": "D. Minimal processing with a high dynamic range Correct Answer: C Topic: Editing" + }, + { + "type": "list", + "bbox": [ + 0.661, + 0.176, + 0.877, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.218, + 0.673, + 0.224 + ], + "angle": 0, + "content": "(6)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.218, + 0.387, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.218, + 0.4, + 0.225 + ], + "angle": 0, + "content": "(5)" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.226, + 0.472, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.226, + 0.88, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.354, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Figure 9. We show one rejected example, several accepted examples, topic distribution, and some sub-topics of a subset of topics of questions in PhotoBench." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.463 + ], + "angle": 0, + "content": "technique achieved by raising the shadow curve to enhance the photo's texture or create a film-like aesthetic. They suggest that PhotoBench includes more professional and detailed questions, closely aligned with practical applications such as photo post-processing and editing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.222, + 0.495 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.685 + ], + "angle": 0, + "content": "Implementation Details. We build PhotoEye with Vicuna-v1.5-7B as the LLM backbone, following existing works [8, 30] for fair comparison. We also use minimal data from the general domain, including LLaVA-Pretrain [15] for pretraining and LLaVA-665K [15] for instruction finetuning. We mix our data with LLaVA-665K during the finetuning stage. We use a batch size of 128 and a learning rate of \\(1e - 3\\) for pre-training, and \\(2e - 5\\) for finetuning. The training takes around 96 hours with 8 A100 GPUs with ZeRO2. For vision encoders, we use: CLIP-ViT-L/14 [25], DINOv2-giant [20] CoDETR-ViT-L [40] and SAM-ViT-H [11]. Details are available in the Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.184, + 0.71 + ], + "angle": 0, + "content": "6.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.87 + ], + "angle": 0, + "content": "We report results of PhotoEye on existing benchmarks. Although PhotoCritique is not specifically designed for low-level vision, but PhotoEye still achieves competing results on Q-Bench when finetuned with it [29]. PhotoEye clearly outperforms existing open-source models, including both general-purpose models and those specifically fine-tuned on low-level vision and image aesthetics datasets, such as Q-Instruct [30] and AesExpert [8]. Compared to closed-source models, our lightweight approach also demonstrates competitive performance, slightly surpassing GPT-4o." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Tab. 3, we compare our model with most competing baselines on PhotoBench, including UNIAA [38], AesEx" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.386, + 0.599, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.389, + 0.872, + 0.414 + ], + "angle": 0, + "content": "What aspect of the image is primarily used to enhance the aesthetic perception? A. The high level of clarity and detailed texture of the petals and stamens (Correct) B. The use of a monotone color palette C. The inclusion of multiple subjects in the background" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.419, + 0.599, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.423, + 0.754, + 0.453 + ], + "angle": 0, + "content": "What atmosphere can you feel in objects and colors? A. Joy & vitality B. Mystery & serenity (Correct) C. Sadness & despair D. Sweet & romance" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.42, + 0.794, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.454, + 0.793, + 0.46 + ], + "angle": 0, + "content": "Q-Instruct" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.422, + 0.874, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.466, + 0.903, + 0.48 + ], + "angle": 0, + "content": "Figure 10. Examples in related works from their original papers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.906, + 0.665 + ], + "angle": 0, + "content": "pert [8] and Q-Instruct [30], which specially focus on low-level vision and aesthetics. Note that when reporting results in Tab. 3, we merge similar topics into a single category. Categories may overlap - for instance, a question about editing exposure would be counted in both Post-Processing and Exposure. PhotoEye notably outperforms baselines. As discussed in Section 5, PhotoBench shows substantially higher expertise and diversity than existing benchmarks, demanding stronger aesthetic understanding capabilities from models. Therefore, PhotoEye, which learns from extensive discussions by photography professionals and enthusiasts, holds a clear advantage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.906, + 0.729 + ], + "angle": 0, + "content": "Ablations. We show the effectiveness of our proposed vision fusor and dataset PhotoCritique in Tab. 4. Both contribute to the model. More ablations including qualitative results are available in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.906, + 0.793 + ], + "angle": 0, + "content": "Qualitative Evaluation. Providing aesthetic recommendations in conversations is actually the primary need in real-world scenarios. In addition to examples in Fig 1, We show examples in Appendix E. with diverse images and contexts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.721, + 0.818 + ], + "angle": 0, + "content": "6.2. See through PhotoEye" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As multiple visual features from different vision encoders are fused in PhotoEye, we are particularly interested in their roles under different cases and how they contribute to our model. Fig. 11 presents the normalized weights of each vision encoder in PhotoEye. We first randomly col" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24813" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.891, + 0.307 + ], + "angle": 0, + "content": "
ModelLLM BackboneQuestion TypesQuadrants of Low-level ConcernsOverall↑
Yes-or-No↑What↑How↑Distortion↑Other↑In-context Distortion↑In-context Other↑
Random guess-50.0027.8633.3137.8938.4838.2835.8237.80
GPT-4o (2024-08-06)Proprietary Model76.7274.3367.7470.0373.3773.6877.9573.04
Qwen-VL-Max [2]Proprietary Model75.6079.4366.0973.3974.0871.0076.9273.63
Qwen-VL-Plus [2]Proprietary Model73.7769.4753.8866.2165.7263.8168.7566.04
Gemini-ProProprietary Model68.8073.7462.3466.3071.3463.9173.0968.16
Emu2-Chat [27]LLaMA-2-33B71.8167.2556.1864.7863.1963.4872.2465.28
mPLUG-Owl2 [34]LLaMA-2-7B72.1857.9656.1956.6869.2153.2972.6561.61
LLAMA-Adapter-V2 [7]LLaMA-2-7B66.1859.2952.1357.3956.2563.1664.9059.46
IDEFICS-Instruct [1]LLaMA-2-7B56.1844.6944.0242.8054.1744.7456.3348.70
SPHINX [14]LLaMA-274.1868.8162.0763.6271.7666.1276.3368.56
InternLM-XComposer-VL [36]InternLM69.4565.2760.8561.6770.1456.9175.1065.35
Qwen-VL [2]QwenLM63.0958.1956.3950.5862.7357.8973.8859.40
Otter-v1 [12]MPT-7B57.0940.7139.5542.2249.3144.0852.6546.35
LLAVA-v1.5 [15]Vicuna-v1.5-7B66.3658.1950.5149.4265.7454.6170.6158.66
InstructBLIP [6]Vicuna-v1.5-7B71.6452.6543.8148.6462.5055.5964.9056.72
Shikra [4]Vicuna-v1.5-7B65.6447.3549.0948.8359.4950.0064.0854.65
MiniGPT-4 [39]Vicuna-v1.5-13B55.8250.2240.3742.0248.3851.9761.2249.03
AesExpert [8]Vicuna-v1.5-7B73.2764.3853.7570.0373.3873.6877.9664.15
Q-Instruct [30]Vicuna-v1.5-7B76.1866.3757.6165.1867.5973.0671.5367.09
PhotoEye (Ours)Vicuna-v1.5-7B80.0176.1067.0274.3274.5977.3081.2274.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.31, + 0.904, + 0.339 + ], + "angle": 0, + "content": "Table 2. Model Performance on Q-Bench (LLVisionQA-dev). PhotoEye has a clear advantage. Best open-source model results in **bold**. Proprietary model results are quoted from Q-Bench [29]." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.344, + 0.885, + 0.492 + ], + "angle": 0, + "content": "
ModelLLM BackboneComposition↑Equipments↑Contrast↑Techniques↑Color and Tone↑Lighting↑
GPT-4o (2024-08-06)Proprietary Model68.1160.2459.0962.7563.9763.74
LLaVA-v1.6-34B [15]Hermes-2-Yi-34B47.6952.5656.8249.3365.4453.80
UNIAA [38]Vicuna-v1.5-7B29.8938.6145.4539.5640.1929.24
AesExpert [8]Vicuna-v1.5-7B51.4960.2456.8253.9961.0364.33
Q-Instruct [30]Vicuna-v1.5-7B31.0234.7740.9131.3754.3844.44
PhotoEye (Ours)Vicuna-v1.5-7B68.3261.3972.9765.6976.0077.78
ModelLLM BackboneExposure↑Post-Processing↑Aperture and Focus↑Storytelling↑Sharpness and Clarity↑Overall↑
GPT-4o (2024-08-06)Proprietary Model54.4858.5759.8769.4461.9064.12
LLaVA-v1.6-34B [15]Hermes-2-Yi-34B56.5553.3352.8766.6728.5755.68
UNIAA [38]Vicuna-v1.5-7B25.5238.5740.1330.5623.8136.87
AesExpert [8]Vicuna-v1.5-7B46.2161.9052.8772.9733.3260.01
Q-Instruct [30]Vicuna-v1.5-7B46.2146.1938.8555.5619.0543.86
PhotoEye (Ours)Vicuna-v1.5-7B69.6680.9570.7075.0061.9073.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.277, + 0.495, + 0.719, + 0.509 + ], + "angle": 0, + "content": "Table 3. Model Performance on PhotoBench Metrics. Best results in bold." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.519, + 0.453, + 0.563 + ], + "angle": 0, + "content": "
MethodQ-BenchPhotoEye
PhotoEye (full method)74.5073.92
w/o Multi-view Vision Fusor70.0868.83
w/o Multi-view Vision Fusor and dataset58.0433.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.208, + 0.566, + 0.366, + 0.578 + ], + "angle": 0, + "content": "Table 4. Ablation Results." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.591, + 0.486, + 0.879 + ], + "angle": 0, + "content": "lect 100 samples from PhotoCritique on compositional aesthetics (e.g., photographic composition, distraction, framing, etc.) and general aesthetics (e.g., storytelling, lighting, etc), respectively, as well as 100 samples for general visual tasks from LLaVA [15] training data. Then we study the averaged weight distribution (normalized) of encoders on different data. We have an interesting observation: (1) aesthetic features are extracted in the first layers, and general features are extracted in the last layers. Fig. 11 shows CLIP and DINO features are dominant regardless of domains in the last layer, which is intuitive as the recognition of objects lays the foundation in both domains. In contrast, aesthetics-related features are primarily extracted from the first layer, especially in compositional problems, where the localization or arrangement of objects are critical. In this case, CoDETR (pre-trained with detection objectives) is dominant. Tab. 5 validated it, where we only activated one encoder during inference to detect their effectiveness in photographic composition." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Case Study. We show how different choices of vision en" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.519, + 0.905, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.566, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Figure 11. Weights of vision encoders across different layers." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.589, + 0.882, + 0.618 + ], + "angle": 0, + "content": "
TopicCLIP-ViT-L/14DINOv2-giantCoDETR-ViT-LSAM-ViT-H
Composition62.6161.6164.6960.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.62, + 0.884, + 0.634 + ], + "angle": 0, + "content": "Table 5. PhotoEye on composition with different encoders." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.673 + ], + "angle": 0, + "content": "coders affect model output in real-world conversation scenarios. We provide interesting examples in Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.907, + 0.768 + ], + "angle": 0, + "content": "Learning of Aesthetic Concepts. Fig. 8 demonstrates how PhotoEye effectively learns discriminative aesthetics-related visual representations. In Appendix B, we conduct a deeper investigation into the learning of aesthetic concepts at feature level, which lays the foundation of high-quality aesthetic response from LLM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.784, + 0.634, + 0.8 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We present PhotoCritique dataset, and PhotoEye, an MLLM with language-guided multi-view vision fusion for aesthetic visual understanding, along with the new benchmark PhotoBench. Extensive experiments show PhotoEye's clear advantage over existing models, advancing aesthetic visual perception in MLLMs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24814" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.157 + ], + "angle": 0, + "content": "This work was partially supported by gifts from Adobe Research and the National Science Foundation under Grants IIS-2316306 and CNS-2330215." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.182, + 0.188, + 0.198 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.278, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[3] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.484, + 0.568 + ], + "angle": 0, + "content": "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.569, + 0.484, + 0.637 + ], + "angle": 0, + "content": "[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.484, + 0.707 + ], + "angle": 0, + "content": "[7] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2311.07575, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.708, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[8] Yipo Huang, Xiangfei Sheng, Zhichao Yang, Quan Yuan, Zhichao Duan, Pengfei Chen, Leida Li, Weisi Lin, and Guangming Shi. AesExpert: Towards multi-modality foundation model for image aesthetics perception. arXiv preprint arXiv:2404.09624, 2024. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.777, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[9] Yipo Huang, Quan Yuan, Xiangfei Sheng, Zhichao Yang, Haoning Wu, Pengfei Chen, Yuzhe Yang, Leida Li, and Weisi Lin. AesBench: An expert benchmark for multimodal large language models on image aesthetics perception. arXiv preprint arXiv:2401.08276, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[10] Zhipeng Huang, Zhizheng Zhang, Yiting Lu, Zheng-Jun Zha, Zhibo Chen, and Baining Guo. Visualcritic: Making Imm's perceive visual quality like humans. arXiv preprint arXiv:2403.12806, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.207, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[12] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[13] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Jiaming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. SPHINX: The Joint Mixing of Weights, Tasks, and Visual Embeddings for Multi-modal Large Language Models. arXiv preprint arXiv:2311.07575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[14] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Ji-aming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[15] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.469 + ], + "angle": 0, + "content": "[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[17] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[18] Daniel Vera Nieto, Luigi Celona, and Clara Fernandez-Labrador. Understanding aesthetics with language: A photo critique dataset for aesthetic assessment. In NeurIPS Track on Datasets and Benchmarks, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.794, + 0.595 + ], + "angle": 0, + "content": "[19] OpenAI. Gpt-4 technical report. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.906, + 0.664 + ], + "angle": 0, + "content": "[20] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.72 + ], + "angle": 0, + "content": "[21] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[22] Daiqing Qi, Handong Zhao, and Sheng Li. Easy regional contrastive learning of expressive fashion representations. Advances in Neural Information Processing Systems, 37: 20480-20509, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[23] Daiqing Qi, Handong Zhao, Zijun Wei, and Sheng Li. Tag-grounded visual instruction tuning with retrieval augmentation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 2008-2026, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[24] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation: A training-free approach. In European Conference on Computer Vision, pages 285-300. Springer, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24815" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[26] Jian Ren, Xiaohui Shen, Zhe Lin, Radomir Mech, and David J. Foran. Personalized image aesthetics. In The IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[27] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[28] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[29] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, and Weisi Lin. Q-BENCH: A benchmark for general-purpose foundation models on low-level vision. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[30] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, Geng Xue, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-Instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[31] Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, Xiaohong Liu, Guangtao Zhai, Shiqi Wang, and Weisi Lin. Towards open-ended visual quality comparison. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[32] Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aesthetics assessment with rich attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19861-19869, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[33] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[34] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[36] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[37] Zhipeng Zhong, Fei Zhou, and Guoping Qiu. Aesthetically relevant image captioning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[38] Zhaokun Zhou, Yiwei Su, Amin Zheng, Qiulin Wang, Rui Chen, Li Yuan, and Di Zhang. UNIAA: A unified multimodal image aesthetic assessment baseline and benchmark. arXiv preprint arXiv:2404.09619, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[39] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[40] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6758, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[41] Zhuofan Zong, Bingqi Ma, Dazhong Shen, Guanglu Song, Hao Shao, Dongzhi Jiang, Hongsheng Li, and Yu Liu. Mova: Adapting mixture of vision experts to multimodal context. In Proceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24816" + } + ] +] \ No newline at end of file diff --git a/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_origin.pdf b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..13ae16de0ec0658ba2731db991a08c1fd49eae2c --- /dev/null +++ b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/e9b6e548-ef9a-4478-aeda-c759b34b1d36_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adc0b6c927a6a8592cc96f56e1380e9f3218d77fe546f183e65719841db6df42 +size 7973070 diff --git a/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/full.md b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..878685ed87ca9030c9d3dcb20645ab5c9eebc7ca --- /dev/null +++ b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/full.md @@ -0,0 +1,339 @@ +# The Photographer's Eye: Teaching Multimodal Large Language Models to See and Critique like Photographers + +Daiqing Qi $^{1,2}$ , Handong Zhao $^{2}$ , Jing Shi $^{2}$ , Simon Jenni $^{2}$ , Yifei Fan $^{2}$ , Franck Dernoncourt $^{2}$ , Scott Cohen $^{2}$ , Sheng Li $^{1}$ + +1University of Virginia, 2Adobe + +![](images/376fbcb19caa8ab0088e9d2bb1b09a71bba274980cecd3fc779986b6c99d6c1d.jpg) +User: Are there any weaknesses in this photograph? +Q-Instruct: The image features a beautiful sunset with the sun's rays shining through the trees. +The scene captures the vibrant colors of the sunset, creating a serene and picturesque atmosphere. The hillside is covered in lush green trees, providing a sense of tranquility and natural beauty. +The combination of the sunset and the verdant landscape creates a captivating and visually pleasing view. +AesExpert: This image looks very clear. The colors in this image are rich and vibrant, with a good color combination, making the picture appear more vivid and bright. The lighting in this image is strong, creating a warm and soft atmosphere, with a strong contrast of dark and light. The composition of this image is a high horizon composition, highlighting the texture in the lower half of the picture. Therefore, this image looks quite beautiful. +PhotoEye: The composition of this image could be improved by straightening the horizon to create a more balanced and visually appealing frame. Additionally, the presence of lens flare may be distracting and detract from the appearance of the image. To address this, the photographer could consider using a lens hood or adjusting their shooting angle to minimize the effect of the flare. Furthermore, the tilted horizon could be corrected to create a more stable and visually appealing atmosphere. However, the amount of the image could be more nuanced, with a more subtle transition between highlights and shadows to create a more natural and inviting atmosphere. + +![](images/02db6404065d57764b2fc7920d13e00e39a93ead4e48068b6d013b22cdbaa684.jpg) +User: Are there any distractors in this image? +Q-Instruct: No +GPT-4o: 1. The Chain and Metal Poles: These elements draw attention because they are close to the subjects and create visual lines in the foreground, potentially competing for focus with the main subjects (the two individual). 2. Darker Darker Colors: The person in the yellow jacket stands out significantly against the darker colors of the other person and the background... +PhotoEye: The photo effectively captures a romantic moment between the couple, but there are some distracting elements that take away from the intimacy of the scene. + +![](images/f92f23541fd85ab48a0e33fe14d3f3d4f7dab68a0d12d3cc96f33c40b5ddf225.jpg) +Figure 1. Examples on our model (PhotoEye), existing MLLMs tailored for low-level vision or aesthetics, and GPT-4o (2024-08-06). The left example and the middle top example highlight a notable limitation of existing open-source MLLMs for low-level vision and aesthetics: insufficient coverage of visual aesthetics. When they fail to identify issues, they either provide positive aspects or claim there are no issues, significantly limiting their usefulness in real-world scenarios. The middle bottom example shows existing MLLMs' lack of expertise in photography: lowering exposure properly can instead enhance colors of objects, so B is correct. While other models, including GPT-4o, made mistakes, our model is correct. The right example reveals another clear limitation of existing open-source MLLMs: their vision encoders are insensitive to low-level vision and aesthetics. In a series of increasingly overexposed photos (2-nd, 3-rd, and 4-th), PhotoEye's vision modules, more attuned to low-level and aesthetic features, identified overexposure by the 2-nd photo, while other models recognized it only when the photos were severely overexposed (i.e., 3-rd and 4-th images). High-quality aesthetic content is highlighted. + +![](images/f4831d2d2e3ef4c1197dc7996076480887e1ef4b691e27df186c9cac7019498c.jpg) +AesExpert: +User: Please describe the exposure of the photo + +![](images/a27dad46dd577d6b92bb09da86b378cb0dfffb3400dcd3c75bb393a615e976f3.jpg) + +![](images/009a912907b8339633276cc95293fac3f41842ba7130fd8ba72eebf648ba85c3.jpg) +The lines and wires in the background are a notable distraction, drawing the viewer's attention away from the couple and disrupting the visual flow. The image of the photographer could consider removing or minimizing these distractions to allow the viewer to focus on the tender moment between the two characters. This distraction would not be the most flattering choice, which could also detract from the overall impact of the image. + +![](images/accdd638a95643858ceacbcc77ad4eebdf864974a50e7d594bd0f6de2a498871.jpg) + +# Abstract + +"While editing directly from life, photographers have found it too difficult to see simultaneously both the blue and the sky." + +John Szarkowski, William Eggleston's Guide + +Photographer and curator, Szarkowski insightfully revealed one of the notable gaps between general and aesthetic visual understanding: while the former focuses on identifying the factual element in an image (sky), the latter transcends such object identification, viewing it instead as an aesthetic component—a pure color block (blue). Such fundamental distinctions between general (detection, localization, etc.) and aesthetic (color, lighting, composition, etc.) + +visual understanding present a significant challenge for Multimodal Large Language Models (MLLMs). Although some recent works have made initial explorations, they are often limited to general and basic aesthetic commonsense. As a result, they frequently fall short in real-world scenarios (Fig. 1), which require extensive expertise—including photographic techniques, photo pre/post-processing knowledge, and more, to provide a detailed analysis and description. To fundamentally enhance the aesthetics understanding of MLLMs, we first introduce a novel dataset, PhotoCritique, derived from extensive discussions among professional photographers and enthusiasts, and characterized by the large scale, expertise, and diversity. Then, to better learn visual aesthetics from PhotoCritique, we further propose a novel model, PhotoEye, featuring a language-guided multi-view vision fusion mechanism to understand image aesthetics from multiple perspectives. Finally, we present a + +novel benchmark, PhotoBench, a comprehensive and professional benchmark for aesthetic visual understanding. On existing benchmarks and PhotoBench, our model demonstrates clear advantages over existing models. + +# 1. Introduction + +Multimodal Large Language Models (MLLMs) [16, 19] have succeeded in general visual tasks but still struggle with a critical ability in human visual perception: aesthetic visual understanding, which is desired in various applications like image recommendation, editing, or generation. The challenge lies in the fundamental distinction between general and aesthetic visual understanding: while the former focuses on factual elements like identification of objects, the latter requires nuanced perception of aesthetic elements such as chromatic properties, or compositional principles. We attribute the shortcomings of MLLMs in aesthetic understanding to two primary factors: dataset and model. + +Dataset: Current datasets such as Q-Instruct [29] and AesExpert [9], made initial explorations in low-level vision or visual aesthetics with MLLM but suffer from limitations in scale, expertise, and diversity. With around 20K images and 58K-88K image annotations, these datasets are insufficient for comprehensive aesthetic understanding, showing clear deficiencies in real-world scenarios, where images vary widely in subjects, scenes, styles, environments, and techniques, requiring larger datasets to ensure coverage. Additionally, their annotations come from small (typically tens of annotators), fixed groups of annotators, limiting diversity and expertise. Even if the annotators are reportedly trained, their expertise remains limited compared to professional photographers and enthusiasts. + +Model: When adapting to aesthetic visual understanding, existing approaches [8, 10, 30, 38] typically finetune existing MLLMs like LLaVA [16] on customized datasets, ignoring the unique requirements of visual features for aesthetic visual understanding. Comprehensive understanding of image aesthetics requires multi-view information: for example, high-level features are essential for interpreting narrative and storytelling, while detailed spatial information is critical for understanding composition. Similarly, low-level details are required for perception of color and lighting, etc. + +To address the challenges, we first introduce the PhotoCritique dataset, featuring (1) large scale: containing over 450K images with detailed aesthetic descriptions, and around 20 times the size of existing datasets [8, 30]; and (2) expertise and diversity: the aesthetic descriptions are derived from extensive discussions among large groups of professional photographers and enthusiasts online, rather than a fixed small number of annotators as in existing works [8, 30]. Subsequently, given the notable different focuses in general and aesthetic visual understanding, we further introduce PhotoEye, an MLLM tailored for aes + +thetic visual understanding. PhotoEye features a language-guided multi-view vision fusion mechanism to effectively learn aesthetic knowledge from PhotoCritique, enabling it to understand image aesthetics from multiple perspectives for various aesthetics-related questions. + +Finally, to provide a comprehensive evaluation of our model and existing MLLMs on aesthetic visual understanding, we introduce our new benchmark PhotoBench. PhotoBench draws its questions from in-depth discussions among professional photographers and enthusiasts, covering 284 sub-topics in photography, making it notably more professional and diverse than existing works. We summarize our contributions as follows: + +- We introduce PhotoCritique, the first large-scale, professional, and diverse dataset for aesthetic visual understanding, consisting of 450K images with 2.63M instruction-tuning pairs, sourced from extensive discussions among professional photographers and enthusiasts. +- We propose PhotoEye, an MLLM specifically designed for aesthetic understanding, featuring a language-guided multi-view vision fusion mechanism. +- We develop PhotoBench, a professional benchmark for aesthetic visual understanding, offering 284 photography sub-topics and insights from photography professionals. +- We conduct a comprehensive study on aesthetic visual understanding with our model and existing MLLMs, revealing the limitations of existing MLLMs and demonstrating how we overcome them through extensive experiments. + +# 2. Related Works + +Multimodal Large Language Models. MLLMs [3, 16] are evolving rapidly recently, demonstrating advanced capabilities in image-text dialogue through enhanced alignment and fine-tuning. Further studies [2, 5, 6, 28, 33] have strengthened MLLMs by increasing the data scale or improving the backbone model. Furthermore, with customized data, a branch of works [4, 21, 33, 35] extend MLLMs' capabilities to referring and grounding. However, despite these advancements in general visual tasks, the aesthetic visual understanding of MLLMs remains largely underexplored. + +Understanding Image Aesthetics with MLLM. Traditional Image Aesthetics Assessment (IAA) [17, 26, 32] is often formalized as a regression task, where the model learns to predict the score of the aesthetic quality of images. One notable limitation is the lack of interpretability, which is critical in practice because we also want to understand why the image is good or bad. Recent advances in MLLM make it possible: a few recent works [8-10, 29-31, 38] explore interpretable aesthetic understanding with MLLMs. However, they often struggle with the limited data scale or lack diverse, professionally curated aesthetic descriptions. + +Improving Visual Perception in MLLM. Recently, several studies [13, 22-24, 28, 41] have investigated back + +![](images/1c3872fe1fa86838fb543bf9ab75a521ee72837624fc4842fa92472056d04890.jpg) +Figure 2. An example from the dpchallenge platform. 3 of the total 64 comments for the photo are presented for illustration. + +#1: Very good one, even if the unsharp mask is too strong. + +#2: This is gorgeous. The use of parallel lines and the vanishing point really leads us into this photo. And the mixture of warm and cool colors is very pleasing. + +#3: This appearance is particularly prominent in the lower left foreground foliage which is also quite bright and thus accentuates the over-sharpened look. Perhaps it is jpeg artifacting, if I look in the upper branches of the tree I can see a cross-hatch pattern... + +bone vision modules for MLLM. While SPHINX [14] simply concatenates features from different encoders, Cambrian [28] and MoVA [41] merge different visual features into a single one. However, the fusion process in MoVA is based on the CLIP feature, which may introduce bias, and the feature extraction (prior to fusion) lacks language-based control. Cambrian [28] employs query-based attention, however, visual features from different encoders are simply averaged during fusion, and the queries are independent of language instructions. To address these limitations, we propose a new MLLM, PhotoEye, featuring a language-guided multi-view vision fusion mechanism to effectively learn aesthetic knowledge. + +# 3. Learning with Photography Enthusiasts and Professionals + +One key reason existing MLLMs struggle with aesthetic visual understanding in real-world scenarios is the training data limitations: (1) Data Scale: Current datasets [8, 30] ( $\sim$ 20K images) lack sufficient coverage of photography genres, styles, themes, etc. High-quality aesthetic descriptions are also limited, with typically 57K-88K detailed descriptions in existing works [8, 30, 38], while our dataset provides over 450K images and detailed descriptions. (2) Description Quality: Most datasets [8, 30] rely on a small, fixed group of annotators, limiting expertise and diversity. In contrast, our PhotoCritique dataset draws insights from hundreds of thousands of photographers and enthusiasts, yielding more professional, nuanced feedback. + +# 3.1. Mining from Digital Photography Platforms + +To address the limitations, we avoided using annotators and instead drew insights from a broad range of online communities. The Digital Photography Challenge (DPC) is an online platform, where users submit their works across various themes and genres and receive constructive feedback from other photographers or enthusiasts (Fig. 2). The wealth of high-quality discussions paired with images on DPC provides valuable resources for MLLMs to learn visual aesthetics. However, directly using this data for supervised fine-tuning is impractical due to the significant noise present in the raw comments. To address this issue, we developed strategies to analyze, summarize, and refine raw comments into a single, well-crafted critique for each image. First, + +![](images/8d88079903468e2a223e789861137172bef6d9ba6b8c2d061c792d13f0e925ec.jpg) +Figure 3. Data Generation Pipeline. The generation of one critique from a group of comments of one image is shown as an example. + +![](images/0834d1657549b669aec94bf4d587975ea2cde3a5a575d62cc7a8cd9f95500234.jpg) +PhotoCritique (Aesthetic Conversation) + +![](images/8a023aaaa6b9700fe04c29538e13bb179a2d248321b44a7f5ba045965a70e81a.jpg) +Figure 4. Top: Examples of PhotoCritique. Bottom: Comparison with examples in Q-Instruct [30]. + +we need to collect raw data from dpchallenge.com online. Since existing works [37] already collected raw data prior to 2022 from the website, we can skip the data scraping process and focus on data processing. Given a post from the DPC platform (comprising one uploaded image and a series of user comments, shown in Fig. 2), we design a multi-step pipeline to generate a single unified photo critique using prompt-based methods from comments only without the image. Fig. 3 briefly illustrates the pipeline with the example of generating a single critique from a set of comments of one image. We first instruct the LLM to analyze and summarize the raw comments, Then the LLM is asked to integrate the comments into a cohesive and unified photo critique. While LLMs possess strong reasoning and comprehension abilities, the immediate output critique is sometimes sparse in information or misleading. Therefore, we designed a filtering strategy to eliminate low-quality data with a smaller LLM for improved efficiency. During the filtering stage, generated photo critiques are input into the smaller LLM, and prompt-based strategies are used to detect whether the generated photo critique conveys enough aesthetics-related information. Only critiques that pass this filtering stage are accepted. We term this portion of data as aesthetic description. Details available in Appendix A. + +From General Photo Critiques to Diverse Instruction-tuning Pairs. In the previous step, we generated photo critiques from DPC's raw comments, which provide overall aesthetic reviews of the given images. In this step, we further utilize LLMs to generate multiple high-quality, diverse conversations addressing various aspects of image aesthetics, such as lighting, composition, color, emotion, narrative, + +![](images/de3ee9b250f023bfc33849c916f8623d1d6f3474b3632a414e5d8a94b0ee39cf.jpg) +Figure 5. Comparison of description length Distributions of aesthetic comments in PhotoCritique and Q-Instruct. Scale of PhotoCritique is Blue and that of Q-Instruct is Red. + +![](images/3c75d4c252b3011307a47f3da654080497969977cf477c3c0c85766d5652d154.jpg) +Figure 6. Top: Distribution of Top 40 categories. Bottom: Examples of photos from a subset of categories. + +and so on, as well as extensive photographic techniques and pre/post-processing skills. Another critical reason for this step is that raw comments are sometimes too lengthy, which can lead to certain aesthetic aspects being overlooked in the LLM-generated critique. By specifically prompting about these aesthetic aspects, the LLM can effectively summarize essential information related to each aspect from the lengthy raw comments. Similarly, after generating all the question-answer pairs, we apply prompt-based filtering strategies to reject less informative pairs, accepting only the high-quality samples. We term this portion of data as aesthetic conversation. Additionally, we also generate MCQs from photo critiques with LLMs. Specifically, each critique is input into the LLM with prompts to generate five MCQs based on the ground truth provided in it. We term this portion of data as aesthetic VQA. Details available in Appendix A. + +# 3.2. PhotoCritique + +Overview. Combining all above, we build PhotoCritique, the first large-scale dataset for expert-level image aesthetics understanding. PhotoCritique consists of three parts: aesthetic description, aesthetic conversation, and aesthetic VQA. Examples of each are shown in Fig. 4 (Top). The three parts contain 450K, 1.9M, and 250K samples, respectively. Born from a large volume of shared images and discussions by photography enthusiasts and professionals on online photography forums, PhotoCritique covers a wide range of over 70 photographic categories, encompassing diverse scenes and subjects. Fig. 6 shows the distribution of the top 40 categories along with several image examples. The extensive scale and diversity enhance the model's effectiveness in real-world scenarios. + +Comparison with Existing Datasets. PhotoCritique rep + +
Dataset#Image#Annotations#Training Samples#Avg. Length
UNIAA57K57K57K32.8
AesExpert*21K88K--
Q-Instruct18K58K66K46.4
PhotoCritique450K450K2.4M65.2
+ +Table 1. Statistics of samples attributed to detailed aesthetic descriptions and conversations across different datasets. Datasets with * are currently not available. + +resents a fundamental improvement over existing works in both quantity and quality. Data Scale: we compares PhotoCritique with existing datasets by showing statistics for data related to detailed aesthetic descriptions and conversations in Tab. 1. This includes the number of source images, the annotations for these images, and the final count of instruction-tuning pairs generated from these annotations using LLMs. Additionally, we calculate the average length of aesthetic comments (including both descriptions and conversations) in the training data. Tab. 1 focuses on training data statistics for aesthetic comments, as these are essential in practical applications where users need detailed feedback on why their photos are strong or weak and how to improve them. In contrast, Multiple Choice Questions (MCQ) are relatively limited in this regard. Besides, aesthetic comments better reflect the information density of the dataset than MCQs, as generating MCQs from a detailed comment is easy and can result in numerous MCQs of varying quality depending on the input prompt. Therefore, the number of MCQs serves only as a relatively rough indicator of the information density, although the number of MCQs in PhotoCritique is also the largest. As shown in Tab. 1, PhotoCritique demonstrates a notable improvement across all metrics compared to existing datasets. Data Quality: PhotoCritique is more professional, and diverse. Unlike most existing works [8, 30], which rely on a small group (usually a few dozen) of annotators to label data, our dataset is sourced from images and discussion from over 107,000 photography enthusiasts and professionals on online photography forums. Fig. 2 provides an intuitive comparison between our dataset and Q-Instruct, with examples of Q-Instruct in its original paper. While Q-Instruct primarily focuses on basic low-level visual facts, PhotoCritique not only addresses low-level details but also considers how these elements impact image aesthetics and ways to improve them. Additionally, we include discussions on photography techniques (e.g., the first example in the aesthetic description covers flash usage and lighting) and the overall impression conveyed by the photo. The comprehensiveness of aesthetic comments in PhotoCritique is also reflected in the average length: Fig. 5 shows the distribution of aesthetic comment lengths in PhotoCritique and Q-Instruct, with PhotoCritique containing a large number of detailed descriptions. Tab. 1 also shows that the average length of aesthetic comments in PhotoCritique is notably greater than that of existing datasets. More examples in Appendix A. + +![](images/bdf846442ce09e879f78d4584365a6d600adcfe3e5c39fc00d8557d33f9d6927.jpg) +Figure 7. Framework of PhotoEye. Given an input photo and a language instruction, visual features from different visual encoders are fused in our vision fusion with language guidance, then fed into the LLM with language embeddings to generate LLM response. + +# 4. Improving Visual Aesthetics Perception + +# 4.1. See like Photographers + +In addition to limitations in existing datasets, Open-source MLLMs face another notable challenge: the limited sensitivity of the vision encoder to aesthetic elements. Most existing works [8, 30] use CLIP as the vision encoder, which, however, is pre-trained with high-level image-text alignment in general domains and is relatively less effective at capturing aesthetic elements. Fig. 1 (right) illustrates this issue, where Q-Instruct and AesExpert only report 'overexposure' when the image is severely overexposed. In Fig. 8, we quantitatively validate it by calculating the discriminability (defined as the average distance between these embeddings) of aesthetic features from different vision encoders. Details are available in Appendix B. It shows that CLIP is less effective in discriminating towards aesthetics-related features as they are more densely clustered. To handle the diversity of aesthetics understanding, from high-level elements like emotion and storytelling to low-level features such as lighting, color, and composition, we propose PhotoEye, a MLLM with a language-guided multi-view vision fusor, which enhances visual perception by fusing multiple vision encoders pre-trained on complementary tasks. + +# 4.2. PhotoEye + +Architecture. PhotoEye is shown in Fig. 7. It consists of a multi-view vision fusor and a backbone LLM. Given an image and an aesthetics-related language instruction, visual features of the image from different vision encoders are fused into a single visual feature by the multi-view vision fusor, conditioned on both the image and the language instruction. Specifically, the multi-view vision fusor consists of $L$ fusion blocks. In each $l$ -th block (or layer), a group of visual features $\{\mathbf{X}^i \in \mathbb{R}^{C \times H \times W}\}_{i=1}^N$ from $N$ different vision encoders are fused with a learnable query $\mathbf{Q}^l \in \mathbb{R}^{C \times H \times W}$ . Denote the output of the $l$ -th block as $\mathbf{O}^l$ , + +![](images/6752d6467ca07f2f89cc6a6cc8d6fe775a94b8553597801358fccb8203658917.jpg) +Figure 8. Comparison of aesthetics-related visual feature discriminability of ours and CLIP [15] in existing works [8, 30, 38]. + +$\mathbf{Q}^{l + 1} = \mathbf{O}^l$ , if $l\neq 0$ . The first learnable query is generated by the language-guided query generator. + +Language-guided Query Generation. At the first layer, $\mathbf{Q}^1$ is generated by the language-guided query generator (bottom left in Fig. 7). Specifically, we maintain a group of $M$ learnable queries $\{\mathbf{Q}^i\in \mathbb{R}^{C\times H\times W}\}_{i = 1}^M$ . Given a specific language instruction, it is encoded into a text embedding $\mathbf{T}$ , which serves as the query to dynamically generate a single query from all $M$ learnable queries with attention mechanism: $\mathbf{Q}^1 = \mathrm{ATTN}(\mathbf{W}_q^{\mathrm{T}}\mathbf{T},\mathbf{W}_k^{\mathrm{T}}\mathbf{Q}_{\mathrm{M}},\mathbf{W}_v^{\mathrm{T}}\mathbf{Q}_{\mathrm{M}})$ where $\mathbf{Q}_{\mathrm{M}}$ is the concatenation of all $M$ queries, and $\mathbf{W}_q^{\mathrm{T}}$ , $\mathbf{W}_k^{\mathrm{T}}$ , $\mathbf{W}_v^{\mathrm{T}}$ are the projection matrixes of the query, key and value, respectively. Note that a pre-trained BERT is introduced as the text encoder, where the [CLS] token from the output is used as the text token $\mathbf{T}$ . We avoid the CLIP text encoder to prevent bias towards the CLIP visual features in the later multimodal gating process. + +In this way, for each language instruction that may focus on a specific aesthetic concept, a tailored query is generated from all $M$ learnable queries, which can best extract visual features related to the aesthetic concept from each vision encoder in the fusion block during inference. + +Feature Extraction with Learnable Query. At $l$ -th layer, given a query $\mathbf{Q}^l$ and a group of visual features $\{\mathbf{X}_i\}_{i=1}^N$ from $N$ different vision encoders, each visual feature is first interpolated to the same shape: $\mathbf{X}_n^* = \text{Interpolate}(\mathbf{X}_n)$ , where $\mathbf{X}_n^* \in \mathbb{R}^{C \times H \times W}$ . Then, each of the extracted visual features from the corresponding vision encoder is given by + +$\mathbf{F}_n^l = \mathrm{ATTN}(\mathbf{W}_q^n\mathbf{Q}^l,\mathbf{W}_k^n\mathbf{X}_n^*,\mathbf{W}_v^n\mathbf{X}_n^*)$ , where $W_{q}^{n},W_{k}^{n}$ $W_{v}^{n}$ are the projection matrixes of the query, key and value. + +Multimodal Gating Network. The extracted visual features $\{\mathbf{F}_i^l\}_{i = 1}^N$ are then fused into a single visual feature in the multimodal gating network. The multimodal gating network assigns different weights to each extracted visual feature conditioned on both image and language and fuses them into one single feature based on their importance. The weights of visual features are directly predicted by an MLP. Specifically, given the language embedding $\mathbf{T}$ and a set of visual features $\{\mathbf{F}_i^l\}_{i = 1}^N$ , weights are given by: $\boldsymbol {w}^{l} = \mathrm{MLP}(\mathbf{T};\mathrm{Pooling}(\{\mathbf{F}_{i}^{l}\}_{i = 1}^{N}))$ , where $\boldsymbol {w}\in \mathbb{R}^{N\times 1}$ , and; means concatenation. The fused single feature is obtained by $\hat{\mathbf{F}}^l = \sum_{i = 1}^N\boldsymbol {w}_i^l\cdot \mathbf{F}_i^l$ . Then, $\hat{\mathbf{F}}^l$ is further fed forward to a transformer block with residual links. The output of the vision fusor is input to the $(l + 1)$ -th layer if $l < L$ ; otherwise, it is fed to an MLP and input to the LLM. + +Remark. Compared to existing works, PhotoEye differs from MoVA [41] in removing its potential bias towards CLIP features by fusing visual features with a group of learnable queries. Moreover, a language-guided query generation mechanism is introduced to extract different features towards different language instructions. When compared to Cambrian-1 [28], PhotoEye conditions both feature extraction and fusion on language, allowing it to capture more refined, instruction-aware aesthetic visual features. In contrast, Cambrian-1 is independent of instructions, potentially resulting in relatively coarser visual features. + +# 5. Benchmarking Aesthetic Visual Understanding with Professionals + +In Section 3, we highlighted a major limitation of existing datasets: annotations are typically provided by a few dozen annotators, which limits the diversity and expertise of the annotations. Even with reported training, these annotators lack the expertise of the many experienced photographers and enthusiasts. Existing benchmarks, such as Q-Bench [29] and AesBench [9], also suffer from this issue, relying on similar methods to obtain annotations. + +To address the limitation and cover a wider range of scenarios, we directly draw insights from extensive photos and discussions on Reddit's PhotoCritique, a vibrant photography community. This approach allows us to generate comprehensive, diverse, and expert-level multiple-choice questions to evaluate the aesthetic visual understanding of MLLMs. Specifically, we do not generate MCQs from annotator-provided annotations; instead, we derive MCQs from photo reviews in Reddit's PhotoCritique. We refer to our new benchmark as PhotoBench. Since RPCD [18] has already collected raw data from PhotoCritique posts between 2009 and 2022, we directly design pipelines that use an LLM (GPT-4-turbo) to generate MCQs. + +# 5.1. From Insights to Questions + +The entire question-generation process consists of three steps: (1) summarization, (2) question generation, and (3) multi-stage filtering. (1) Summarization. Since the raw comments from the forum are quite noisy, empirically we find directly inputting the raw comments corresponding to an image to the LLM and prompting it to generate multiple-choice questions based on these comments does not yield ideal results. To address this issue, we first use the LLM to summarize the noisy raw comments for each image, producing a single, clean, comprehensive, and detailed photo critique. This process aligns with the pipeline shown in Fig. 3. (2) Question Generation. For each image, we input its corresponding photo critique into the LLM and instruct it to generate five multiple-choice questions grounded on the provided critique. (3) Multi-stage Filtering. This step directly impacts the quality of PhotoBench. We first selected the top 5,000 most detailed photo critiques from the initial generation step and instructed the LLM (using GPT-4 Turbo, as we empirically found it slightly outperforms GPT-4o) to generate 25,000 questions. The first filtering stage is LLM-based, primarily aimed at removing MCQs with low visual dependency—that is, questions that could likely be answered correctly by reading only the question and options without viewing the image. We show an example in Fig. 9 (1). Our method is intuitive yet effective: we provide only the question and options to the LLM (GPT-4o) without the image and filter out questions it answers correctly. This step eliminated about $70\%$ of the items. We then instructed GPT-4o to score the remaining questions based on aesthetics relevance, visual dependency, and expertise. We selected the top 1,500 questions based on their average score. Details available in Appendix C. + +# 5.2. PhotoBench + +We show detailed information of PhotoBench in Fig. 9. Compared to existing works [9, 29, 38], Photobench gains more advantages in diversity and expertise. Due to potentially limited expertise and diversity among annotators [9, 29], MCQs from existing works are often too simple. We can have a straightforward comparison in Fig. 10. + +While questions from existing works are relatively straightforward and focus on factual elements, our questions are more challenging and practical, requiring deeper aesthetics understanding and covering topics directly related to photography in practice, such as camera settings, post-processing, and photographic techniques (topics marked in blue in Fig 9. Examples are shown in Fig 9 (3,6). In Fig. 9 (3), to make the rainbow stand out more, the exposure should be reduced (slight underexposure can instead enhance color saturation). In Fig. 9 (6), we can observe that the shadows are not purely black but rather "grayish" (zoom in for a better view)—a common post-processing + +![](images/52c1ce3fec77e245fa46bdda2854dfe7d7d993b788c8609c577215e43401d782.jpg) +What adjustments can be made to the photo to make the image cleaner and more organized? A. Add more objects in the foreground to make the image more complex and engaging B. Remove the fence to create a cleaner composition in the photo C. Increase the saturation to make it more vibrant. D. Lower the exposure Correct Answer: B Rejected for low visual dependence + +![](images/01941c10d1270f344c95cf1ba430e77d0f1cbd8921b485aa3a0291f12a2176fc.jpg) +How might the orientation of the photograph better emphasize the structure of the ceiling in the church? + +A. Landscape orientation B. Planting orientation +B. Portrait orientation C. Diagonal orientation +D. No change needed + +Correct Answer: B Topic: Composition + +![](images/c34cf37499c9aafd89e79e6aed7b55d69c11c58b9c872569ce99d89ce7d4a4b8.jpg) +(2) +Which of the following aspects in the photo might be perceived as slightly jarring, requiring adjustment? + +![](images/ccfa218b89c40aef55f107f765aefc27b26286541235dff2bfd262adf2c08af6.jpg) +(5) + +A. The brightness gradient of the fog +B. The difference between the brightness gradient + +B. The contrast between C. The depth of field used + +D. The hue of the morning light + +Correct Answer: A Topic: Tone + +![](images/6bfb902ebc8ecbe07eec5623ce69e3026bb83f5362ed0508506f4645171ab68b.jpg) +How could the image be optimized to enhance the appeal of the rainbow? +(3) +What is the characterization of the processing style of the photo's shadows? + +A. Increase the exposure time + +B. Decrease the ISO setting C. Add more artificial lighting +D. Use a flash + +Correct Answer: B + +Topic: Exposure + +![](images/4c3192dda3e12859d68ad7d41f1d2ff56551b85c3730016dc43a09bebac1d388.jpg) + +![](images/676f2ac2afd7c1f12593d6229edca88b8fb5369edbba75195f18afe765ad9ba0.jpg) + +![](images/6e78541d024020c68ef12d8ca705443a1a85258fe3b15bd1ba630588ab939656.jpg) +A. Naturally soft with subtle transitions B. High contrast with distinct black and white tones C. Heavy-handed with raised blacks and crushed shadows. +D. Minimal processing with a high dynamic range Correct Answer: C Topic: Editing +(6) + +technique achieved by raising the shadow curve to enhance the photo's texture or create a film-like aesthetic. They suggest that PhotoBench includes more professional and detailed questions, closely aligned with practical applications such as photo post-processing and editing. + +# 6. Experiments + +Implementation Details. We build PhotoEye with Vicuna-v1.5-7B as the LLM backbone, following existing works [8, 30] for fair comparison. We also use minimal data from the general domain, including LLaVA-Pretrain [15] for pretraining and LLaVA-665K [15] for instruction finetuning. We mix our data with LLaVA-665K during the finetuning stage. We use a batch size of 128 and a learning rate of $1e - 3$ for pre-training, and $2e - 5$ for finetuning. The training takes around 96 hours with 8 A100 GPUs with ZeRO2. For vision encoders, we use: CLIP-ViT-L/14 [25], DINOv2-giant [20] CoDETR-ViT-L [40] and SAM-ViT-H [11]. Details are available in the Appendix D. + +# 6.1. Results + +We report results of PhotoEye on existing benchmarks. Although PhotoCritique is not specifically designed for low-level vision, but PhotoEye still achieves competing results on Q-Bench when finetuned with it [29]. PhotoEye clearly outperforms existing open-source models, including both general-purpose models and those specifically fine-tuned on low-level vision and image aesthetics datasets, such as Q-Instruct [30] and AesExpert [8]. Compared to closed-source models, our lightweight approach also demonstrates competitive performance, slightly surpassing GPT-4o. + +In Tab. 3, we compare our model with most competing baselines on PhotoBench, including UNIAA [38], AesEx + +![](images/ae1e0dc934debf50ad3eeffca4d3bbabb554ec71cc4e0057c8008ea37b96c742.jpg) +Figure 9. We show one rejected example, several accepted examples, topic distribution, and some sub-topics of a subset of topics of questions in PhotoBench. +What aspect of the image is primarily used to enhance the aesthetic perception? A. The high level of clarity and detailed texture of the petals and stamens (Correct) B. The use of a monotone color palette C. The inclusion of multiple subjects in the background + +![](images/1450f11a2aa286962dde5dcd968df0c7388e427692a20edcc4814fa7af4a3356.jpg) +What atmosphere can you feel in objects and colors? A. Joy & vitality B. Mystery & serenity (Correct) C. Sadness & despair D. Sweet & romance +Figure 10. Examples in related works from their original papers. + +![](images/545906481829a67a1f95ebb83749827897e65581bdc68eb1132ad1353adc7a61.jpg) +Q-Instruct + +![](images/3eaa214257140aa590bdf6105d8ff7ad65cdae5e67645faab15060726ece97d0.jpg) + +pert [8] and Q-Instruct [30], which specially focus on low-level vision and aesthetics. Note that when reporting results in Tab. 3, we merge similar topics into a single category. Categories may overlap - for instance, a question about editing exposure would be counted in both Post-Processing and Exposure. PhotoEye notably outperforms baselines. As discussed in Section 5, PhotoBench shows substantially higher expertise and diversity than existing benchmarks, demanding stronger aesthetic understanding capabilities from models. Therefore, PhotoEye, which learns from extensive discussions by photography professionals and enthusiasts, holds a clear advantage. + +Ablations. We show the effectiveness of our proposed vision fusor and dataset PhotoCritique in Tab. 4. Both contribute to the model. More ablations including qualitative results are available in Appendix. + +Qualitative Evaluation. Providing aesthetic recommendations in conversations is actually the primary need in real-world scenarios. In addition to examples in Fig 1, We show examples in Appendix E. with diverse images and contexts. + +# 6.2. See through PhotoEye + +As multiple visual features from different vision encoders are fused in PhotoEye, we are particularly interested in their roles under different cases and how they contribute to our model. Fig. 11 presents the normalized weights of each vision encoder in PhotoEye. We first randomly col + +
ModelLLM BackboneQuestion TypesQuadrants of Low-level ConcernsOverall↑
Yes-or-No↑What↑How↑Distortion↑Other↑In-context Distortion↑In-context Other↑
Random guess-50.0027.8633.3137.8938.4838.2835.8237.80
GPT-4o (2024-08-06)Proprietary Model76.7274.3367.7470.0373.3773.6877.9573.04
Qwen-VL-Max [2]Proprietary Model75.6079.4366.0973.3974.0871.0076.9273.63
Qwen-VL-Plus [2]Proprietary Model73.7769.4753.8866.2165.7263.8168.7566.04
Gemini-ProProprietary Model68.8073.7462.3466.3071.3463.9173.0968.16
Emu2-Chat [27]LLaMA-2-33B71.8167.2556.1864.7863.1963.4872.2465.28
mPLUG-Owl2 [34]LLaMA-2-7B72.1857.9656.1956.6869.2153.2972.6561.61
LLAMA-Adapter-V2 [7]LLaMA-2-7B66.1859.2952.1357.3956.2563.1664.9059.46
IDEFICS-Instruct [1]LLaMA-2-7B56.1844.6944.0242.8054.1744.7456.3348.70
SPHINX [14]LLaMA-274.1868.8162.0763.6271.7666.1276.3368.56
InternLM-XComposer-VL [36]InternLM69.4565.2760.8561.6770.1456.9175.1065.35
Qwen-VL [2]QwenLM63.0958.1956.3950.5862.7357.8973.8859.40
Otter-v1 [12]MPT-7B57.0940.7139.5542.2249.3144.0852.6546.35
LLAVA-v1.5 [15]Vicuna-v1.5-7B66.3658.1950.5149.4265.7454.6170.6158.66
InstructBLIP [6]Vicuna-v1.5-7B71.6452.6543.8148.6462.5055.5964.9056.72
Shikra [4]Vicuna-v1.5-7B65.6447.3549.0948.8359.4950.0064.0854.65
MiniGPT-4 [39]Vicuna-v1.5-13B55.8250.2240.3742.0248.3851.9761.2249.03
AesExpert [8]Vicuna-v1.5-7B73.2764.3853.7570.0373.3873.6877.9664.15
Q-Instruct [30]Vicuna-v1.5-7B76.1866.3757.6165.1867.5973.0671.5367.09
PhotoEye (Ours)Vicuna-v1.5-7B80.0176.1067.0274.3274.5977.3081.2274.50
+ +Table 2. Model Performance on Q-Bench (LLVisionQA-dev). PhotoEye has a clear advantage. Best open-source model results in **bold**. Proprietary model results are quoted from Q-Bench [29]. + +
ModelLLM BackboneComposition↑Equipments↑Contrast↑Techniques↑Color and Tone↑Lighting↑
GPT-4o (2024-08-06)Proprietary Model68.1160.2459.0962.7563.9763.74
LLaVA-v1.6-34B [15]Hermes-2-Yi-34B47.6952.5656.8249.3365.4453.80
UNIAA [38]Vicuna-v1.5-7B29.8938.6145.4539.5640.1929.24
AesExpert [8]Vicuna-v1.5-7B51.4960.2456.8253.9961.0364.33
Q-Instruct [30]Vicuna-v1.5-7B31.0234.7740.9131.3754.3844.44
PhotoEye (Ours)Vicuna-v1.5-7B68.3261.3972.9765.6976.0077.78
ModelLLM BackboneExposure↑Post-Processing↑Aperture and Focus↑Storytelling↑Sharpness and Clarity↑Overall↑
GPT-4o (2024-08-06)Proprietary Model54.4858.5759.8769.4461.9064.12
LLaVA-v1.6-34B [15]Hermes-2-Yi-34B56.5553.3352.8766.6728.5755.68
UNIAA [38]Vicuna-v1.5-7B25.5238.5740.1330.5623.8136.87
AesExpert [8]Vicuna-v1.5-7B46.2161.9052.8772.9733.3260.01
Q-Instruct [30]Vicuna-v1.5-7B46.2146.1938.8555.5619.0543.86
PhotoEye (Ours)Vicuna-v1.5-7B69.6680.9570.7075.0061.9073.92
+ +Table 3. Model Performance on PhotoBench Metrics. Best results in bold. + +
MethodQ-BenchPhotoEye
PhotoEye (full method)74.5073.92
w/o Multi-view Vision Fusor70.0868.83
w/o Multi-view Vision Fusor and dataset58.0433.74
+ +lect 100 samples from PhotoCritique on compositional aesthetics (e.g., photographic composition, distraction, framing, etc.) and general aesthetics (e.g., storytelling, lighting, etc), respectively, as well as 100 samples for general visual tasks from LLaVA [15] training data. Then we study the averaged weight distribution (normalized) of encoders on different data. We have an interesting observation: (1) aesthetic features are extracted in the first layers, and general features are extracted in the last layers. Fig. 11 shows CLIP and DINO features are dominant regardless of domains in the last layer, which is intuitive as the recognition of objects lays the foundation in both domains. In contrast, aesthetics-related features are primarily extracted from the first layer, especially in compositional problems, where the localization or arrangement of objects are critical. In this case, CoDETR (pre-trained with detection objectives) is dominant. Tab. 5 validated it, where we only activated one encoder during inference to detect their effectiveness in photographic composition. + +Case Study. We show how different choices of vision en + +![](images/d8355cce4e3db4dd5bc8c9f5950cea519ec72325f0128b006b7e9e217254484d.jpg) +Figure 11. Weights of vision encoders across different layers. + +Table 4. Ablation Results. + +
TopicCLIP-ViT-L/14DINOv2-giantCoDETR-ViT-LSAM-ViT-H
Composition62.6161.6164.6960.24
+ +Table 5. PhotoEye on composition with different encoders. + +coders affect model output in real-world conversation scenarios. We provide interesting examples in Appendix F. + +Learning of Aesthetic Concepts. Fig. 8 demonstrates how PhotoEye effectively learns discriminative aesthetics-related visual representations. In Appendix B, we conduct a deeper investigation into the learning of aesthetic concepts at feature level, which lays the foundation of high-quality aesthetic response from LLM. + +# 7. Conclusion + +We present PhotoCritique dataset, and PhotoEye, an MLLM with language-guided multi-view vision fusion for aesthetic visual understanding, along with the new benchmark PhotoBench. Extensive experiments show PhotoEye's clear advantage over existing models, advancing aesthetic visual perception in MLLMs. + +# Acknowledgment + +This work was partially supported by gifts from Adobe Research and the National Science Foundation under Grants IIS-2316306 and CNS-2330215. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 2022. 8 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 8 +[3] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2 +[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 8 +[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 2 +[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 8 +[7] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2311.07575, 2023. 8 +[8] Yipo Huang, Xiangfei Sheng, Zhichao Yang, Quan Yuan, Zhichao Duan, Pengfei Chen, Leida Li, Weisi Lin, and Guangming Shi. AesExpert: Towards multi-modality foundation model for image aesthetics perception. arXiv preprint arXiv:2404.09624, 2024. 2, 3, 4, 5, 7, 8 +[9] Yipo Huang, Quan Yuan, Xiangfei Sheng, Zhichao Yang, Haoning Wu, Pengfei Chen, Yuzhe Yang, Leida Li, and Weisi Lin. AesBench: An expert benchmark for multimodal large language models on image aesthetics perception. arXiv preprint arXiv:2401.08276, 2024. 2, 6 +[10] Zhipeng Huang, Zhizheng Zhang, Yiting Lu, Zheng-Jun Zha, Zhibo Chen, and Baining Guo. Visualcritic: Making Imm's perceive visual quality like humans. arXiv preprint arXiv:2403.12806, 2024. 2 + +[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 7 +[12] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023.8 +[13] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Jiaming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. SPHINX: The Joint Mixing of Weights, Tasks, and Visual Embeddings for Multi-modal Large Language Models. arXiv preprint arXiv:2311.07575, 2023. 2 +[14] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Ji-aming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2024. 3, 8 +[15] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5, 7, 8 +[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[17] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415. IEEE, 2012. 2 +[18] Daniel Vera Nieto, Luigi Celona, and Clara Fernandez-Labrador. Understanding aesthetics with language: A photo critique dataset for aesthetic assessment. In NeurIPS Track on Datasets and Benchmarks, 2022. 6 +[19] OpenAI. Gpt-4 technical report. 2023. 2 +[20] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7 +[21] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2 +[22] Daiqing Qi, Handong Zhao, and Sheng Li. Easy regional contrastive learning of expressive fashion representations. Advances in Neural Information Processing Systems, 37: 20480-20509, 2024. 2 +[23] Daiqing Qi, Handong Zhao, Zijun Wei, and Sheng Li. Tag-grounded visual instruction tuning with retrieval augmentation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 2008-2026, 2024. +[24] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation: A training-free approach. In European Conference on Computer Vision, pages 285-300. Springer, 2024. 2 + +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[26] Jian Ren, Xiaohui Shen, Zhe Lin, Radomir Mech, and David J. Foran. Personalized image aesthetics. In The IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[27] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2024. 8 +[28] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 2, 3, 6 +[29] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, and Weisi Lin. Q-BENCH: A benchmark for general-purpose foundation models on low-level vision. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 2, 6, 7, 8 +[30] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, Geng Xue, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-Instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 5, 7, 8 +[31] Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, Xiaohong Liu, Guangtao Zhai, Shiqi Wang, and Weisi Lin. Towards open-ended visual quality comparison. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2 +[32] Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aesthetics assessment with rich attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19861-19869, 2022. 2 +[33] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2 +[34] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 8 +[35] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and + +Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2 +[36] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 8 +[37] Zhipeng Zhong, Fei Zhou, and Guoping Qiu. Aesthetically relevant image captioning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 3 +[38] Zhaokun Zhou, Yiwei Su, Amin Zheng, Qiulin Wang, Rui Chen, Li Yuan, and Di Zhang. UNIAA: A unified multimodal image aesthetic assessment baseline and benchmark. arXiv preprint arXiv:2404.09619, 2024. 2, 3, 5, 6, 7, 8 +[39] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8 +[40] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6758, 2023. 7 +[41] Zhuofan Zong, Bingqi Ma, Dazhong Shen, Guanglu Song, Hao Shao, Dongzhi Jiang, Hongsheng Li, and Yu Liu. Mova: Adapting mixture of vision experts to multimodal context. In Proceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 6 \ No newline at end of file diff --git a/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/images.zip b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..039e84c9963d030df25b35c60e1f77e2a30b849c --- /dev/null +++ b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d46ffbe7cc4b6b613ae839689682fa7e5140ff90a67e745265d67748fb7c5a1 +size 609344 diff --git a/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/layout.json b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..76b23293036fa8779ac61220f2f87f461ed71dfc --- /dev/null +++ b/2025/The Photographer's Eye_ Teaching Multimodal Large Language Models to See, and Critique Like Photographers/layout.json @@ -0,0 +1,9007 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 523, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 523, + 158 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 523, + 158 + ], + "type": "text", + "content": "The Photographer's Eye: Teaching Multimodal Large Language Models to See and Critique like Photographers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "spans": [ + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "text", + "content": "Daiqing Qi" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "text", + "content": ", Handong Zhao" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "text", + "content": ", Jing Shi" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "text", + "content": ", Simon Jenni" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "text", + "content": ", Yifei Fan" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "text", + "content": ", Franck Dernoncourt" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "text", + "content": ", Scott Cohen" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "text", + "content": ", Sheng Li" + }, + { + "bbox": [ + 143, + 178, + 468, + 207 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 227, + 212, + 381, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 212, + 381, + 228 + ], + "spans": [ + { + "bbox": [ + 227, + 212, + 381, + 228 + ], + "type": "text", + "content": "1University of Virginia, 2Adobe" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 238, + 167, + 304 + ], + "blocks": [ + { + "bbox": [ + 67, + 238, + 167, + 304 + ], + "lines": [ + { + "bbox": [ + 67, + 238, + 167, + 304 + ], + "spans": [ + { + "bbox": [ + 67, + 238, + 167, + 304 + ], + "type": "image", + "image_path": "376fbcb19caa8ab0088e9d2bb1b09a71bba274980cecd3fc779986b6c99d6c1d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 305, + 141, + 315 + ], + "lines": [ + { + "bbox": [ + 69, + 305, + 141, + 315 + ], + "spans": [ + { + "bbox": [ + 69, + 305, + 141, + 315 + ], + "type": "text", + "content": "User: Are there any weaknesses in this photograph?" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 321, + 160, + 335 + ], + "lines": [ + { + "bbox": [ + 69, + 321, + 160, + 335 + ], + "spans": [ + { + "bbox": [ + 69, + 321, + 160, + 335 + ], + "type": "text", + "content": "Q-Instruct: The image features a beautiful sunset with the sun's rays shining through the trees." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 338, + 163, + 356 + ], + "lines": [ + { + "bbox": [ + 70, + 338, + 163, + 356 + ], + "spans": [ + { + "bbox": [ + 70, + 338, + 163, + 356 + ], + "type": "text", + "content": "The scene captures the vibrant colors of the sunset, creating a serene and picturesque atmosphere. The hillside is covered in lush green trees, providing a sense of tranquility and natural beauty." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 358, + 154, + 373 + ], + "lines": [ + { + "bbox": [ + 70, + 358, + 154, + 373 + ], + "spans": [ + { + "bbox": [ + 70, + 358, + 154, + 373 + ], + "type": "text", + "content": "The combination of the sunset and the verdant landscape creates a captivating and visually pleasing view." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 240, + 242, + 297 + ], + "lines": [ + { + "bbox": [ + 170, + 240, + 242, + 297 + ], + "spans": [ + { + "bbox": [ + 170, + 240, + 242, + 297 + ], + "type": "text", + "content": "AesExpert: This image looks very clear. The colors in this image are rich and vibrant, with a good color combination, making the picture appear more vivid and bright. The lighting in this image is strong, creating a warm and soft atmosphere, with a strong contrast of dark and light. The composition of this image is a high horizon composition, highlighting the texture in the lower half of the picture. Therefore, this image looks quite beautiful." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 299, + 243, + 376 + ], + "lines": [ + { + "bbox": [ + 170, + 299, + 243, + 376 + ], + "spans": [ + { + "bbox": [ + 170, + 299, + 243, + 376 + ], + "type": "text", + "content": "PhotoEye: The composition of this image could be improved by straightening the horizon to create a more balanced and visually appealing frame. Additionally, the presence of lens flare may be distracting and detract from the appearance of the image. To address this, the photographer could consider using a lens hood or adjusting their shooting angle to minimize the effect of the flare. Furthermore, the tilted horizon could be corrected to create a more stable and visually appealing atmosphere. However, the amount of the image could be more nuanced, with a more subtle transition between highlights and shadows to create a more natural and inviting atmosphere." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 249, + 239, + 354, + 305 + ], + "blocks": [ + { + "bbox": [ + 249, + 239, + 354, + 305 + ], + "lines": [ + { + "bbox": [ + 249, + 239, + 354, + 305 + ], + "spans": [ + { + "bbox": [ + 249, + 239, + 354, + 305 + ], + "type": "image", + "image_path": "02db6404065d57764b2fc7920d13e00e39a93ead4e48068b6d013b22cdbaa684.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 242, + 439, + 248 + ], + "lines": [ + { + "bbox": [ + 357, + 242, + 439, + 248 + ], + "spans": [ + { + "bbox": [ + 357, + 242, + 439, + 248 + ], + "type": "text", + "content": "User: Are there any distractors in this image?" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 252, + 394, + 258 + ], + "lines": [ + { + "bbox": [ + 358, + 252, + 394, + 258 + ], + "spans": [ + { + "bbox": [ + 358, + 252, + 394, + 258 + ], + "type": "text", + "content": "Q-Instruct: No" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 263, + 444, + 303 + ], + "lines": [ + { + "bbox": [ + 358, + 263, + 444, + 303 + ], + "spans": [ + { + "bbox": [ + 358, + 263, + 444, + 303 + ], + "type": "text", + "content": "GPT-4o: 1. The Chain and Metal Poles: These elements draw attention because they are close to the subjects and create visual lines in the foreground, potentially competing for focus with the main subjects (the two individual). 2. Darker Darker Colors: The person in the yellow jacket stands out significantly against the darker colors of the other person and the background..." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 308, + 444, + 327 + ], + "lines": [ + { + "bbox": [ + 358, + 308, + 444, + 327 + ], + "spans": [ + { + "bbox": [ + 358, + 308, + 444, + 327 + ], + "type": "text", + "content": "PhotoEye: The photo effectively captures a romantic moment between the couple, but there are some distracting elements that take away from the intimacy of the scene." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 249, + 308, + 353, + 376 + ], + "blocks": [ + { + "bbox": [ + 249, + 308, + 353, + 376 + ], + "lines": [ + { + "bbox": [ + 249, + 308, + 353, + 376 + ], + "spans": [ + { + "bbox": [ + 249, + 308, + 353, + 376 + ], + "type": "image", + "image_path": "f92f23541fd85ab48a0e33fe14d3f3d4f7dab68a0d12d3cc96f33c40b5ddf225.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 384, + 538, + 495 + ], + "lines": [ + { + "bbox": [ + 72, + 384, + 538, + 495 + ], + "spans": [ + { + "bbox": [ + 72, + 384, + 538, + 495 + ], + "type": "text", + "content": "Figure 1. Examples on our model (PhotoEye), existing MLLMs tailored for low-level vision or aesthetics, and GPT-4o (2024-08-06). The left example and the middle top example highlight a notable limitation of existing open-source MLLMs for low-level vision and aesthetics: insufficient coverage of visual aesthetics. When they fail to identify issues, they either provide positive aspects or claim there are no issues, significantly limiting their usefulness in real-world scenarios. The middle bottom example shows existing MLLMs' lack of expertise in photography: lowering exposure properly can instead enhance colors of objects, so B is correct. While other models, including GPT-4o, made mistakes, our model is correct. The right example reveals another clear limitation of existing open-source MLLMs: their vision encoders are insensitive to low-level vision and aesthetics. In a series of increasingly overexposed photos (2-nd, 3-rd, and 4-th), PhotoEye's vision modules, more attuned to low-level and aesthetic features, identified overexposure by the 2-nd photo, while other models recognized it only when the photos were severely overexposed (i.e., 3-rd and 4-th images). High-quality aesthetic content is highlighted." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 449, + 253, + 503, + 314 + ], + "blocks": [ + { + "bbox": [ + 403, + 251, + 425, + 258 + ], + "lines": [ + { + "bbox": [ + 403, + 251, + 425, + 258 + ], + "spans": [ + { + "bbox": [ + 403, + 251, + 425, + 258 + ], + "type": "text", + "content": "AesExpert:" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 242, + 541, + 248 + ], + "lines": [ + { + "bbox": [ + 451, + 242, + 541, + 248 + ], + "spans": [ + { + "bbox": [ + 451, + 242, + 541, + 248 + ], + "type": "text", + "content": "User: Please describe the exposure of the photo" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 253, + 503, + 314 + ], + "lines": [ + { + "bbox": [ + 449, + 253, + 503, + 314 + ], + "spans": [ + { + "bbox": [ + 449, + 253, + 503, + 314 + ], + "type": "image", + "image_path": "f4831d2d2e3ef4c1197dc7996076480887e1ef4b691e27df186c9cac7019498c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 504, + 255, + 541, + 314 + ], + "blocks": [ + { + "bbox": [ + 504, + 255, + 541, + 314 + ], + "lines": [ + { + "bbox": [ + 504, + 255, + 541, + 314 + ], + "spans": [ + { + "bbox": [ + 504, + 255, + 541, + 314 + ], + "type": "image", + "image_path": "a27dad46dd577d6b92bb09da86b378cb0dfffb3400dcd3c75bb393a615e976f3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 449, + 316, + 503, + 376 + ], + "blocks": [ + { + "bbox": [ + 358, + 330, + 445, + 374 + ], + "lines": [ + { + "bbox": [ + 358, + 330, + 445, + 374 + ], + "spans": [ + { + "bbox": [ + 358, + 330, + 445, + 374 + ], + "type": "text", + "content": "The lines and wires in the background are a notable distraction, drawing the viewer's attention away from the couple and disrupting the visual flow. The image of the photographer could consider removing or minimizing these distractions to allow the viewer to focus on the tender moment between the two characters. This distraction would not be the most flattering choice, which could also detract from the overall impact of the image." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 316, + 503, + 376 + ], + "lines": [ + { + "bbox": [ + 449, + 316, + 503, + 376 + ], + "spans": [ + { + "bbox": [ + 449, + 316, + 503, + 376 + ], + "type": "image", + "image_path": "009a912907b8339633276cc95293fac3f41842ba7130fd8ba72eebf648ba85c3.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 504, + 323, + 541, + 376 + ], + "blocks": [ + { + "bbox": [ + 504, + 323, + 541, + 376 + ], + "lines": [ + { + "bbox": [ + 504, + 323, + 541, + 376 + ], + "spans": [ + { + "bbox": [ + 504, + 323, + 541, + 376 + ], + "type": "image", + "image_path": "accdd638a95643858ceacbcc77ad4eebdf864974a50e7d594bd0f6de2a498871.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "spans": [ + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 70, + 525, + 281, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 525, + 281, + 561 + ], + "spans": [ + { + "bbox": [ + 70, + 525, + 281, + 561 + ], + "type": "text", + "content": "\"While editing directly from life, photographers have found it too difficult to see simultaneously both the blue and the sky.\"" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 86, + 562, + 274, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 562, + 274, + 574 + ], + "spans": [ + { + "bbox": [ + 86, + 562, + 274, + 574 + ], + "type": "text", + "content": "John Szarkowski, William Eggleston's Guide" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 582, + 296, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 678 + ], + "type": "text", + "content": "Photographer and curator, Szarkowski insightfully revealed one of the notable gaps between general and aesthetic visual understanding: while the former focuses on identifying the factual element in an image (sky), the latter transcends such object identification, viewing it instead as an aesthetic component—a pure color block (blue). Such fundamental distinctions between general (detection, localization, etc.) and aesthetic (color, lighting, composition, etc.)" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 508, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 712 + ], + "type": "text", + "content": "visual understanding present a significant challenge for Multimodal Large Language Models (MLLMs). Although some recent works have made initial explorations, they are often limited to general and basic aesthetic commonsense. As a result, they frequently fall short in real-world scenarios (Fig. 1), which require extensive expertise—including photographic techniques, photo pre/post-processing knowledge, and more, to provide a detailed analysis and description. To fundamentally enhance the aesthetics understanding of MLLMs, we first introduce a novel dataset, PhotoCritique, derived from extensive discussions among professional photographers and enthusiasts, and characterized by the large scale, expertise, and diversity. Then, to better learn visual aesthetics from PhotoCritique, we further propose a novel model, PhotoEye, featuring a language-guided multi-view vision fusion mechanism to understand image aesthetics from multiple perspectives. Finally, we present a" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 715 + ], + "type": "text", + "content": "Part of this work was completed during Daiqing's internship at Adobe. \n1William Eggleston is an American photographer, widely recognized as a pioneering figure in color photography." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24807" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "content": "novel benchmark, PhotoBench, a comprehensive and professional benchmark for aesthetic visual understanding. On existing benchmarks and PhotoBench, our model demonstrates clear advantages over existing models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 132, + 136, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 132, + 136, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 136, + 144 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 295 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs) [16, 19] have succeeded in general visual tasks but still struggle with a critical ability in human visual perception: aesthetic visual understanding, which is desired in various applications like image recommendation, editing, or generation. The challenge lies in the fundamental distinction between general and aesthetic visual understanding: while the former focuses on factual elements like identification of objects, the latter requires nuanced perception of aesthetic elements such as chromatic properties, or compositional principles. We attribute the shortcomings of MLLMs in aesthetic understanding to two primary factors: dataset and model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 296, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 296, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 296, + 295, + 462 + ], + "type": "text", + "content": "Dataset: Current datasets such as Q-Instruct [29] and AesExpert [9], made initial explorations in low-level vision or visual aesthetics with MLLM but suffer from limitations in scale, expertise, and diversity. With around 20K images and 58K-88K image annotations, these datasets are insufficient for comprehensive aesthetic understanding, showing clear deficiencies in real-world scenarios, where images vary widely in subjects, scenes, styles, environments, and techniques, requiring larger datasets to ensure coverage. Additionally, their annotations come from small (typically tens of annotators), fixed groups of annotators, limiting diversity and expertise. Even if the annotators are reportedly trained, their expertise remains limited compared to professional photographers and enthusiasts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 463, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 295, + 582 + ], + "type": "text", + "content": "Model: When adapting to aesthetic visual understanding, existing approaches [8, 10, 30, 38] typically finetune existing MLLMs like LLaVA [16] on customized datasets, ignoring the unique requirements of visual features for aesthetic visual understanding. Comprehensive understanding of image aesthetics requires multi-view information: for example, high-level features are essential for interpreting narrative and storytelling, while detailed spatial information is critical for understanding composition. Similarly, low-level details are required for perception of color and lighting, etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "To address the challenges, we first introduce the PhotoCritique dataset, featuring (1) large scale: containing over 450K images with detailed aesthetic descriptions, and around 20 times the size of existing datasets [8, 30]; and (2) expertise and diversity: the aesthetic descriptions are derived from extensive discussions among large groups of professional photographers and enthusiasts online, rather than a fixed small number of annotators as in existing works [8, 30]. Subsequently, given the notable different focuses in general and aesthetic visual understanding, we further introduce PhotoEye, an MLLM tailored for aes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": "thetic visual understanding. PhotoEye features a language-guided multi-view vision fusion mechanism to effectively learn aesthetic knowledge from PhotoCritique, enabling it to understand image aesthetics from multiple perspectives for various aesthetics-related questions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 227 + ], + "type": "text", + "content": "Finally, to provide a comprehensive evaluation of our model and existing MLLMs on aesthetic visual understanding, we introduce our new benchmark PhotoBench. PhotoBench draws its questions from in-depth discussions among professional photographers and enthusiasts, covering 284 sub-topics in photography, making it notably more professional and diverse than existing works. We summarize our contributions as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 228, + 553, + 408 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 315, + 228, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 228, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 315, + 228, + 553, + 289 + ], + "type": "text", + "content": "- We introduce PhotoCritique, the first large-scale, professional, and diverse dataset for aesthetic visual understanding, consisting of 450K images with 2.63M instruction-tuning pairs, sourced from extensive discussions among professional photographers and enthusiasts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 289, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 289, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 315, + 289, + 553, + 323 + ], + "type": "text", + "content": "- We propose PhotoEye, an MLLM specifically designed for aesthetic understanding, featuring a language-guided multi-view vision fusion mechanism." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 325, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 325, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 315, + 325, + 553, + 361 + ], + "type": "text", + "content": "- We develop PhotoBench, a professional benchmark for aesthetic visual understanding, offering 284 photography sub-topics and insights from photography professionals." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 361, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 553, + 408 + ], + "type": "text", + "content": "- We conduct a comprehensive study on aesthetic visual understanding with our model and existing MLLMs, revealing the limitations of existing MLLMs and demonstrating how we overcome them through extensive experiments." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 418, + 405, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 405, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 405, + 430 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 557 + ], + "type": "text", + "content": "Multimodal Large Language Models. MLLMs [3, 16] are evolving rapidly recently, demonstrating advanced capabilities in image-text dialogue through enhanced alignment and fine-tuning. Further studies [2, 5, 6, 28, 33] have strengthened MLLMs by increasing the data scale or improving the backbone model. Furthermore, with customized data, a branch of works [4, 21, 33, 35] extend MLLMs' capabilities to referring and grounding. However, despite these advancements in general visual tasks, the aesthetic visual understanding of MLLMs remains largely underexplored." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 689 + ], + "type": "text", + "content": "Understanding Image Aesthetics with MLLM. Traditional Image Aesthetics Assessment (IAA) [17, 26, 32] is often formalized as a regression task, where the model learns to predict the score of the aesthetic quality of images. One notable limitation is the lack of interpretability, which is critical in practice because we also want to understand why the image is good or bad. Recent advances in MLLM make it possible: a few recent works [8-10, 29-31, 38] explore interpretable aesthetic understanding with MLLMs. However, they often struggle with the limited data scale or lack diverse, professionally curated aesthetic descriptions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Improving Visual Perception in MLLM. Recently, several studies [13, 22-24, 28, 41] have investigated back" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24808" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 115, + 133 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 115, + 133 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 115, + 133 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 115, + 133 + ], + "type": "image", + "image_path": "1c3872fe1fa86838fb543bf9ab75a521ee72837624fc4842fa92472056d04890.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 147, + 295, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 169 + ], + "type": "text", + "content": "Figure 2. An example from the dpchallenge platform. 3 of the total 64 comments for the photo are presented for illustration." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 75, + 263, + 82 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 75, + 263, + 82 + ], + "spans": [ + { + "bbox": [ + 117, + 75, + 263, + 82 + ], + "type": "text", + "content": "#1: Very good one, even if the unsharp mask is too strong." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 85, + 282, + 104 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 85, + 282, + 104 + ], + "spans": [ + { + "bbox": [ + 117, + 85, + 282, + 104 + ], + "type": "text", + "content": "#2: This is gorgeous. The use of parallel lines and the vanishing point really leads us into this photo. And the mixture of warm and cool colors is very pleasing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 106, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 106, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 117, + 106, + 286, + 131 + ], + "type": "text", + "content": "#3: This appearance is particularly prominent in the lower left foreground foliage which is also quite bright and thus accentuates the over-sharpened look. Perhaps it is jpeg artifacting, if I look in the upper branches of the tree I can see a cross-hatch pattern..." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 177, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 296, + 333 + ], + "type": "text", + "content": "bone vision modules for MLLM. While SPHINX [14] simply concatenates features from different encoders, Cambrian [28] and MoVA [41] merge different visual features into a single one. However, the fusion process in MoVA is based on the CLIP feature, which may introduce bias, and the feature extraction (prior to fusion) lacks language-based control. Cambrian [28] employs query-based attention, however, visual features from different encoders are simply averaged during fusion, and the queries are independent of language instructions. To address these limitations, we propose a new MLLM, PhotoEye, featuring a language-guided multi-view vision fusion mechanism to effectively learn aesthetic knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 342, + 296, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 369 + ], + "type": "text", + "content": "3. Learning with Photography Enthusiasts and Professionals" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 377, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 533 + ], + "type": "text", + "content": "One key reason existing MLLMs struggle with aesthetic visual understanding in real-world scenarios is the training data limitations: (1) Data Scale: Current datasets [8, 30] (" + }, + { + "bbox": [ + 55, + 377, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 55, + 377, + 296, + 533 + ], + "type": "text", + "content": "20K images) lack sufficient coverage of photography genres, styles, themes, etc. High-quality aesthetic descriptions are also limited, with typically 57K-88K detailed descriptions in existing works [8, 30, 38], while our dataset provides over 450K images and detailed descriptions. (2) Description Quality: Most datasets [8, 30] rely on a small, fixed group of annotators, limiting expertise and diversity. In contrast, our PhotoCritique dataset draws insights from hundreds of thousands of photographers and enthusiasts, yielding more professional, nuanced feedback." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 285, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 285, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 285, + 553 + ], + "type": "text", + "content": "3.1. Mining from Digital Photography Platforms" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "To address the limitations, we avoided using annotators and instead drew insights from a broad range of online communities. The Digital Photography Challenge (DPC) is an online platform, where users submit their works across various themes and genres and receive constructive feedback from other photographers or enthusiasts (Fig. 2). The wealth of high-quality discussions paired with images on DPC provides valuable resources for MLLMs to learn visual aesthetics. However, directly using this data for supervised fine-tuning is impractical due to the significant noise present in the raw comments. To address this issue, we developed strategies to analyze, summarize, and refine raw comments into a single, well-crafted critique for each image. First," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 555, + 105 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 105 + ], + "type": "image", + "image_path": "8d88079903468e2a223e789861137172bef6d9ba6b8c2d061c792d13f0e925ec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 116, + 555, + 137 + ], + "lines": [ + { + "bbox": [ + 313, + 116, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 555, + 137 + ], + "type": "text", + "content": "Figure 3. Data Generation Pipeline. The generation of one critique from a group of comments of one image is shown as an example." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 144, + 553, + 241 + ], + "blocks": [ + { + "bbox": [ + 317, + 144, + 553, + 241 + ], + "lines": [ + { + "bbox": [ + 317, + 144, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 144, + 553, + 241 + ], + "type": "image", + "image_path": "0834d1657549b669aec94bf4d587975ea2cde3a5a575d62cc7a8cd9f95500234.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 241, + 391, + 247 + ], + "lines": [ + { + "bbox": [ + 317, + 241, + 391, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 241, + 391, + 247 + ], + "type": "text", + "content": "PhotoCritique (Aesthetic Conversation)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 251, + 553, + 294 + ], + "blocks": [ + { + "bbox": [ + 317, + 251, + 553, + 294 + ], + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 294 + ], + "type": "image", + "image_path": "8a023aaaa6b9700fe04c29538e13bb179a2d248321b44a7f5ba045965a70e81a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 307, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 329 + ], + "type": "text", + "content": "Figure 4. Top: Examples of PhotoCritique. Bottom: Comparison with examples in Q-Instruct [30]." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 337, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 624 + ], + "type": "text", + "content": "we need to collect raw data from dpchallenge.com online. Since existing works [37] already collected raw data prior to 2022 from the website, we can skip the data scraping process and focus on data processing. Given a post from the DPC platform (comprising one uploaded image and a series of user comments, shown in Fig. 2), we design a multi-step pipeline to generate a single unified photo critique using prompt-based methods from comments only without the image. Fig. 3 briefly illustrates the pipeline with the example of generating a single critique from a set of comments of one image. We first instruct the LLM to analyze and summarize the raw comments, Then the LLM is asked to integrate the comments into a cohesive and unified photo critique. While LLMs possess strong reasoning and comprehension abilities, the immediate output critique is sometimes sparse in information or misleading. Therefore, we designed a filtering strategy to eliminate low-quality data with a smaller LLM for improved efficiency. During the filtering stage, generated photo critiques are input into the smaller LLM, and prompt-based strategies are used to detect whether the generated photo critique conveys enough aesthetics-related information. Only critiques that pass this filtering stage are accepted. We term this portion of data as aesthetic description. Details available in Appendix A." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "From General Photo Critiques to Diverse Instruction-tuning Pairs. In the previous step, we generated photo critiques from DPC's raw comments, which provide overall aesthetic reviews of the given images. In this step, we further utilize LLMs to generate multiple high-quality, diverse conversations addressing various aspects of image aesthetics, such as lighting, composition, color, emotion, narrative," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24809" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 67, + 294, + 125 + ], + "blocks": [ + { + "bbox": [ + 56, + 67, + 294, + 125 + ], + "lines": [ + { + "bbox": [ + 56, + 67, + 294, + 125 + ], + "spans": [ + { + "bbox": [ + 56, + 67, + 294, + 125 + ], + "type": "image", + "image_path": "de3ee9b250f023bfc33849c916f8623d1d6f3474b3632a414e5d8a94b0ee39cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 130, + 295, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 130, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 295, + 163 + ], + "type": "text", + "content": "Figure 5. Comparison of description length Distributions of aesthetic comments in PhotoCritique and Q-Instruct. Scale of PhotoCritique is Blue and that of Q-Instruct is Red." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 170, + 293, + 276 + ], + "blocks": [ + { + "bbox": [ + 59, + 170, + 293, + 276 + ], + "lines": [ + { + "bbox": [ + 59, + 170, + 293, + 276 + ], + "spans": [ + { + "bbox": [ + 59, + 170, + 293, + 276 + ], + "type": "image", + "image_path": "3c75d4c252b3011307a47f3da654080497969977cf477c3c0c85766d5652d154.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 285, + 295, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 285, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 295, + 308 + ], + "type": "text", + "content": "Figure 6. Top: Distribution of Top 40 categories. Bottom: Examples of photos from a subset of categories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 310, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 310, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 295, + 502 + ], + "type": "text", + "content": "and so on, as well as extensive photographic techniques and pre/post-processing skills. Another critical reason for this step is that raw comments are sometimes too lengthy, which can lead to certain aesthetic aspects being overlooked in the LLM-generated critique. By specifically prompting about these aesthetic aspects, the LLM can effectively summarize essential information related to each aspect from the lengthy raw comments. Similarly, after generating all the question-answer pairs, we apply prompt-based filtering strategies to reject less informative pairs, accepting only the high-quality samples. We term this portion of data as aesthetic conversation. Additionally, we also generate MCQs from photo critiques with LLMs. Specifically, each critique is input into the LLM with prompts to generate five MCQs based on the ground truth provided in it. We term this portion of data as aesthetic VQA. Details available in Appendix A." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 145, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 145, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 145, + 523 + ], + "type": "text", + "content": "3.2. PhotoCritique" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "text", + "content": "Overview. Combining all above, we build PhotoCritique, the first large-scale dataset for expert-level image aesthetics understanding. PhotoCritique consists of three parts: aesthetic description, aesthetic conversation, and aesthetic VQA. Examples of each are shown in Fig. 4 (Top). The three parts contain 450K, 1.9M, and 250K samples, respectively. Born from a large volume of shared images and discussions by photography enthusiasts and professionals on online photography forums, PhotoCritique covers a wide range of over 70 photographic categories, encompassing diverse scenes and subjects. Fig. 6 shows the distribution of the top 40 categories along with several image examples. The extensive scale and diversity enhance the model's effectiveness in real-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Comparison with Existing Datasets. PhotoCritique rep" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 323, + 70, + 549, + 114 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 549, + 114 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 549, + 114 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 549, + 114 + ], + "type": "table", + "html": "
Dataset#Image#Annotations#Training Samples#Avg. Length
UNIAA57K57K57K32.8
AesExpert*21K88K--
Q-Instruct18K58K66K46.4
PhotoCritique450K450K2.4M65.2
", + "image_path": "7518d21a9be5782301f3d42f4dfd102c12cce8e972889b2c65d3c034131f6667.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 122, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 313, + 122, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 553, + 156 + ], + "type": "text", + "content": "Table 1. Statistics of samples attributed to detailed aesthetic descriptions and conversations across different datasets. Datasets with * are currently not available." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 163, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 715 + ], + "type": "text", + "content": "resents a fundamental improvement over existing works in both quantity and quality. Data Scale: we compares PhotoCritique with existing datasets by showing statistics for data related to detailed aesthetic descriptions and conversations in Tab. 1. This includes the number of source images, the annotations for these images, and the final count of instruction-tuning pairs generated from these annotations using LLMs. Additionally, we calculate the average length of aesthetic comments (including both descriptions and conversations) in the training data. Tab. 1 focuses on training data statistics for aesthetic comments, as these are essential in practical applications where users need detailed feedback on why their photos are strong or weak and how to improve them. In contrast, Multiple Choice Questions (MCQ) are relatively limited in this regard. Besides, aesthetic comments better reflect the information density of the dataset than MCQs, as generating MCQs from a detailed comment is easy and can result in numerous MCQs of varying quality depending on the input prompt. Therefore, the number of MCQs serves only as a relatively rough indicator of the information density, although the number of MCQs in PhotoCritique is also the largest. As shown in Tab. 1, PhotoCritique demonstrates a notable improvement across all metrics compared to existing datasets. Data Quality: PhotoCritique is more professional, and diverse. Unlike most existing works [8, 30], which rely on a small group (usually a few dozen) of annotators to label data, our dataset is sourced from images and discussion from over 107,000 photography enthusiasts and professionals on online photography forums. Fig. 2 provides an intuitive comparison between our dataset and Q-Instruct, with examples of Q-Instruct in its original paper. While Q-Instruct primarily focuses on basic low-level visual facts, PhotoCritique not only addresses low-level details but also considers how these elements impact image aesthetics and ways to improve them. Additionally, we include discussions on photography techniques (e.g., the first example in the aesthetic description covers flash usage and lighting) and the overall impression conveyed by the photo. The comprehensiveness of aesthetic comments in PhotoCritique is also reflected in the average length: Fig. 5 shows the distribution of aesthetic comment lengths in PhotoCritique and Q-Instruct, with PhotoCritique containing a large number of detailed descriptions. Tab. 1 also shows that the average length of aesthetic comments in PhotoCritique is notably greater than that of existing datasets. More examples in Appendix A." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24810" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 64, + 537, + 218 + ], + "blocks": [ + { + "bbox": [ + 75, + 64, + 537, + 218 + ], + "lines": [ + { + "bbox": [ + 75, + 64, + 537, + 218 + ], + "spans": [ + { + "bbox": [ + 75, + 64, + 537, + 218 + ], + "type": "image", + "image_path": "bdf846442ce09e879f78d4584365a6d600adcfe3e5c39fc00d8557d33f9d6927.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 226, + 555, + 251 + ], + "type": "text", + "content": "Figure 7. Framework of PhotoEye. Given an input photo and a language instruction, visual features from different visual encoders are fused in our vision fusion with language guidance, then fed into the LLM with language embeddings to generate LLM response." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 272, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 272, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 272, + 277 + ], + "type": "text", + "content": "4. Improving Visual Aesthetics Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 187, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 187, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 187, + 296 + ], + "type": "text", + "content": "4.1. See like Photographers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 302, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 296, + 553 + ], + "type": "text", + "content": "In addition to limitations in existing datasets, Open-source MLLMs face another notable challenge: the limited sensitivity of the vision encoder to aesthetic elements. Most existing works [8, 30] use CLIP as the vision encoder, which, however, is pre-trained with high-level image-text alignment in general domains and is relatively less effective at capturing aesthetic elements. Fig. 1 (right) illustrates this issue, where Q-Instruct and AesExpert only report 'overexposure' when the image is severely overexposed. In Fig. 8, we quantitatively validate it by calculating the discriminability (defined as the average distance between these embeddings) of aesthetic features from different vision encoders. Details are available in Appendix B. It shows that CLIP is less effective in discriminating towards aesthetics-related features as they are more densely clustered. To handle the diversity of aesthetics understanding, from high-level elements like emotion and storytelling to low-level features such as lighting, color, and composition, we propose PhotoEye, a MLLM with a language-guided multi-view vision fusor, which enhances visual perception by fusing multiple vision encoders pre-trained on complementary tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 124, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 124, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 124, + 575 + ], + "type": "text", + "content": "4.2. PhotoEye" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Architecture. PhotoEye is shown in Fig. 7. It consists of a multi-view vision fusor and a backbone LLM. Given an image and an aesthetics-related language instruction, visual features of the image from different vision encoders are fused into a single visual feature by the multi-view vision fusor, conditioned on both the image and the language instruction. Specifically, the multi-view vision fusor consists of " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " fusion blocks. In each " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "-th block (or layer), a group of visual features " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}^i \\in \\mathbb{R}^{C \\times H \\times W}\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " different vision encoders are fused with a learnable query " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^l \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". Denote the output of the " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "-th block as " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^l" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 262, + 544, + 327 + ], + "blocks": [ + { + "bbox": [ + 328, + 262, + 544, + 327 + ], + "lines": [ + { + "bbox": [ + 328, + 262, + 544, + 327 + ], + "spans": [ + { + "bbox": [ + 328, + 262, + 544, + 327 + ], + "type": "image", + "image_path": "6752d6467ca07f2f89cc6a6cc8d6fe775a94b8553597801358fccb8203658917.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 335, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 357 + ], + "type": "text", + "content": "Figure 8. Comparison of aesthetics-related visual feature discriminability of ours and CLIP [15] in existing works [8, 30, 38]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 363, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{l + 1} = \\mathbf{O}^l" + }, + { + "bbox": [ + 313, + 363, + 555, + 389 + ], + "type": "text", + "content": " , if " + }, + { + "bbox": [ + 313, + 363, + 555, + 389 + ], + "type": "inline_equation", + "content": "l\\neq 0" + }, + { + "bbox": [ + 313, + 363, + 555, + 389 + ], + "type": "text", + "content": " . The first learnable query is generated by the language-guided query generator." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": "Language-guided Query Generation. At the first layer, " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^1" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": " is generated by the language-guided query generator (bottom left in Fig. 7). Specifically, we maintain a group of " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": " learnable queries " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q}^i\\in \\mathbb{R}^{C\\times H\\times W}\\}_{i = 1}^M" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": ". Given a specific language instruction, it is encoded into a text embedding " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": ", which serves as the query to dynamically generate a single query from all " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": " learnable queries with attention mechanism: " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^1 = \\mathrm{ATTN}(\\mathbf{W}_q^{\\mathrm{T}}\\mathbf{T},\\mathbf{W}_k^{\\mathrm{T}}\\mathbf{Q}_{\\mathrm{M}},\\mathbf{W}_v^{\\mathrm{T}}\\mathbf{Q}_{\\mathrm{M}})" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{M}}" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": " is the concatenation of all " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": " queries, and " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": " are the projection matrixes of the query, key and value, respectively. Note that a pre-trained BERT is introduced as the text encoder, where the [CLS] token from the output is used as the text token " + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 393, + 556, + 575 + ], + "type": "text", + "content": ". We avoid the CLIP text encoder to prevent bias towards the CLIP visual features in the later multimodal gating process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 576, + 556, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 556, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 556, + 636 + ], + "type": "text", + "content": "In this way, for each language instruction that may focus on a specific aesthetic concept, a tailored query is generated from all " + }, + { + "bbox": [ + 313, + 576, + 556, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 576, + 556, + 636 + ], + "type": "text", + "content": " learnable queries, which can best extract visual features related to the aesthetic concept from each vision encoder in the fusion block during inference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Feature Extraction with Learnable Query. At " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "-th layer, given a query " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^l" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " and a group of visual features " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " different vision encoders, each visual feature is first interpolated to the same shape: " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^* = \\text{Interpolate}(\\mathbf{X}_n)" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_n^* \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ". Then, each of the extracted visual features from the corresponding vision encoder is given by" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24811" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_n^l = \\mathrm{ATTN}(\\mathbf{W}_q^n\\mathbf{Q}^l,\\mathbf{W}_k^n\\mathbf{X}_n^*,\\mathbf{W}_v^n\\mathbf{X}_n^*)" + }, + { + "bbox": [ + 55, + 71, + 294, + 97 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 71, + 294, + 97 + ], + "type": "inline_equation", + "content": "W_{q}^{n},W_{k}^{n}" + }, + { + "bbox": [ + 55, + 71, + 294, + 97 + ], + "type": "inline_equation", + "content": "W_{v}^{n}" + }, + { + "bbox": [ + 55, + 71, + 294, + 97 + ], + "type": "text", + "content": " are the projection matrixes of the query, key and value." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": "Multimodal Gating Network. The extracted visual features " + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_i^l\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": " are then fused into a single visual feature in the multimodal gating network. The multimodal gating network assigns different weights to each extracted visual feature conditioned on both image and language and fuses them into one single feature based on their importance. The weights of visual features are directly predicted by an MLP. Specifically, given the language embedding " + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": " and a set of visual features " + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_i^l\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": ", weights are given by: " + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\boldsymbol {w}^{l} = \\mathrm{MLP}(\\mathbf{T};\\mathrm{Pooling}(\\{\\mathbf{F}_{i}^{l}\\}_{i = 1}^{N}))" + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\boldsymbol {w}\\in \\mathbb{R}^{N\\times 1}" + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": ", and; means concatenation. The fused single feature is obtained by " + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^l = \\sum_{i = 1}^N\\boldsymbol {w}_i^l\\cdot \\mathbf{F}_i^l" + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^l" + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": " is further fed forward to a transformer block with residual links. The output of the vision fusor is input to the " + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "inline_equation", + "content": "(l + 1)" + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": "-th layer if " + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "inline_equation", + "content": "l < L" + }, + { + "bbox": [ + 55, + 101, + 295, + 281 + ], + "type": "text", + "content": "; otherwise, it is fed to an MLP and input to the LLM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 413 + ], + "type": "text", + "content": "Remark. Compared to existing works, PhotoEye differs from MoVA [41] in removing its potential bias towards CLIP features by fusing visual features with a group of learnable queries. Moreover, a language-guided query generation mechanism is introduced to extract different features towards different language instructions. When compared to Cambrian-1 [28], PhotoEye conditions both feature extraction and fusion on language, allowing it to capture more refined, instruction-aware aesthetic visual features. In contrast, Cambrian-1 is independent of instructions, potentially resulting in relatively coarser visual features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 453 + ], + "type": "text", + "content": "5. Benchmarking Aesthetic Visual Understanding with Professionals" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 556 + ], + "type": "text", + "content": "In Section 3, we highlighted a major limitation of existing datasets: annotations are typically provided by a few dozen annotators, which limits the diversity and expertise of the annotations. Even with reported training, these annotators lack the expertise of the many experienced photographers and enthusiasts. Existing benchmarks, such as Q-Bench [29] and AesBench [9], also suffer from this issue, relying on similar methods to obtain annotations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "To address the limitation and cover a wider range of scenarios, we directly draw insights from extensive photos and discussions on Reddit's PhotoCritique, a vibrant photography community. This approach allows us to generate comprehensive, diverse, and expert-level multiple-choice questions to evaluate the aesthetic visual understanding of MLLMs. Specifically, we do not generate MCQs from annotator-provided annotations; instead, we derive MCQs from photo reviews in Reddit's PhotoCritique. We refer to our new benchmark as PhotoBench. Since RPCD [18] has already collected raw data from PhotoCritique posts between 2009 and 2022, we directly design pipelines that use an LLM (GPT-4-turbo) to generate MCQs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 464, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 464, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 464, + 85 + ], + "type": "text", + "content": "5.1. From Insights to Questions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 90, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 473 + ], + "type": "text", + "content": "The entire question-generation process consists of three steps: (1) summarization, (2) question generation, and (3) multi-stage filtering. (1) Summarization. Since the raw comments from the forum are quite noisy, empirically we find directly inputting the raw comments corresponding to an image to the LLM and prompting it to generate multiple-choice questions based on these comments does not yield ideal results. To address this issue, we first use the LLM to summarize the noisy raw comments for each image, producing a single, clean, comprehensive, and detailed photo critique. This process aligns with the pipeline shown in Fig. 3. (2) Question Generation. For each image, we input its corresponding photo critique into the LLM and instruct it to generate five multiple-choice questions grounded on the provided critique. (3) Multi-stage Filtering. This step directly impacts the quality of PhotoBench. We first selected the top 5,000 most detailed photo critiques from the initial generation step and instructed the LLM (using GPT-4 Turbo, as we empirically found it slightly outperforms GPT-4o) to generate 25,000 questions. The first filtering stage is LLM-based, primarily aimed at removing MCQs with low visual dependency—that is, questions that could likely be answered correctly by reading only the question and options without viewing the image. We show an example in Fig. 9 (1). Our method is intuitive yet effective: we provide only the question and options to the LLM (GPT-4o) without the image and filter out questions it answers correctly. This step eliminated about " + }, + { + "bbox": [ + 313, + 90, + 555, + 473 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 90, + 555, + 473 + ], + "type": "text", + "content": " of the items. We then instructed GPT-4o to score the remaining questions based on aesthetics relevance, visual dependency, and expertise. We selected the top 1,500 questions based on their average score. Details available in Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 394, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 394, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 394, + 491 + ], + "type": "text", + "content": "5.2. PhotoBench" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 553, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 570 + ], + "type": "text", + "content": "We show detailed information of PhotoBench in Fig. 9. Compared to existing works [9, 29, 38], Photobench gains more advantages in diversity and expertise. Due to potentially limited expertise and diversity among annotators [9, 29], MCQs from existing works are often too simple. We can have a straightforward comparison in Fig. 10." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "While questions from existing works are relatively straightforward and focus on factual elements, our questions are more challenging and practical, requiring deeper aesthetics understanding and covering topics directly related to photography in practice, such as camera settings, post-processing, and photographic techniques (topics marked in blue in Fig 9. Examples are shown in Fig 9 (3,6). In Fig. 9 (3), to make the rainbow stand out more, the exposure should be reduced (slight underexposure can instead enhance color saturation). In Fig. 9 (6), we can observe that the shadows are not purely black but rather \"grayish\" (zoom in for a better view)—a common post-processing" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24812" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 108, + 118 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 108, + 118 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 108, + 118 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 108, + 118 + ], + "type": "image", + "image_path": "52c1ce3fec77e245fa46bdda2854dfe7d7d993b788c8609c577215e43401d782.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 70, + 236, + 116 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 236, + 116 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 236, + 116 + ], + "type": "text", + "content": "What adjustments can be made to the photo to make the image cleaner and more organized? A. Add more objects in the foreground to make the image more complex and engaging B. Remove the fence to create a cleaner composition in the photo C. Increase the saturation to make it more vibrant. D. Lower the exposure Correct Answer: B Rejected for low visual dependence" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 118, + 121, + 172 + ], + "blocks": [ + { + "bbox": [ + 71, + 118, + 121, + 172 + ], + "lines": [ + { + "bbox": [ + 71, + 118, + 121, + 172 + ], + "spans": [ + { + "bbox": [ + 71, + 118, + 121, + 172 + ], + "type": "image", + "image_path": "01941c10d1270f344c95cf1ba430e77d0f1cbd8921b485aa3a0291f12a2176fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 125, + 228, + 134 + ], + "lines": [ + { + "bbox": [ + 123, + 125, + 228, + 134 + ], + "spans": [ + { + "bbox": [ + 123, + 125, + 228, + 134 + ], + "type": "text", + "content": "How might the orientation of the photograph better emphasize the structure of the ceiling in the church?" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 140, + 171, + 157 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 124, + 140, + 171, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 140, + 171, + 146 + ], + "spans": [ + { + "bbox": [ + 124, + 140, + 171, + 146 + ], + "type": "text", + "content": "A. Landscape orientation B. Planting orientation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 146, + 167, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 146, + 167, + 152 + ], + "spans": [ + { + "bbox": [ + 124, + 146, + 167, + 152 + ], + "type": "text", + "content": "B. Portrait orientation C. Diagonal orientation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 152, + 166, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 152, + 166, + 157 + ], + "spans": [ + { + "bbox": [ + 124, + 152, + 166, + 157 + ], + "type": "text", + "content": "D. No change needed" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 124, + 164, + 233, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 164, + 233, + 171 + ], + "spans": [ + { + "bbox": [ + 124, + 164, + 233, + 171 + ], + "type": "text", + "content": "Correct Answer: B Topic: Composition" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 239, + 68, + 312, + 118 + ], + "blocks": [ + { + "bbox": [ + 239, + 68, + 312, + 118 + ], + "lines": [ + { + "bbox": [ + 239, + 68, + 312, + 118 + ], + "spans": [ + { + "bbox": [ + 239, + 68, + 312, + 118 + ], + "type": "image", + "image_path": "c34cf37499c9aafd89e79e6aed7b55d69c11c58b9c872569ce99d89ce7d4a4b8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 118, + 246, + 123 + ], + "lines": [ + { + "bbox": [ + 239, + 118, + 246, + 123 + ], + "spans": [ + { + "bbox": [ + 239, + 118, + 246, + 123 + ], + "type": "text", + "content": "(2)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 70, + 400, + 85 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 400, + 85 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 400, + 85 + ], + "type": "text", + "content": "Which of the following aspects in the photo might be perceived as slightly jarring, requiring adjustment?" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 239, + 125, + 312, + 172 + ], + "blocks": [ + { + "bbox": [ + 239, + 125, + 312, + 172 + ], + "lines": [ + { + "bbox": [ + 239, + 125, + 312, + 172 + ], + "spans": [ + { + "bbox": [ + 239, + 125, + 312, + 172 + ], + "type": "image", + "image_path": "ccfa218b89c40aef55f107f765aefc27b26286541235dff2bfd262adf2c08af6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 172, + 244, + 178 + ], + "lines": [ + { + "bbox": [ + 238, + 172, + 244, + 178 + ], + "spans": [ + { + "bbox": [ + 238, + 172, + 244, + 178 + ], + "type": "text", + "content": "(5)" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 88, + 400, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 88, + 400, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 88, + 400, + 95 + ], + "type": "text", + "content": "A. The brightness gradient of the fog \nB. The difference between the brightness gradient" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 95, + 364, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 364, + 101 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 364, + 101 + ], + "type": "text", + "content": "B. The contrast between C. The depth of field used" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 101, + 373, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 101, + 373, + 107 + ], + "spans": [ + { + "bbox": [ + 316, + 101, + 373, + 107 + ], + "type": "text", + "content": "D. The hue of the morning light" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 396, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 396, + 114 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 396, + 114 + ], + "type": "text", + "content": "Correct Answer: A Topic: Tone" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 405, + 68, + 437, + 117 + ], + "blocks": [ + { + "bbox": [ + 405, + 68, + 437, + 117 + ], + "lines": [ + { + "bbox": [ + 405, + 68, + 437, + 117 + ], + "spans": [ + { + "bbox": [ + 405, + 68, + 437, + 117 + ], + "type": "image", + "image_path": "6bfb902ebc8ecbe07eec5623ce69e3026bb83f5362ed0508506f4645171ab68b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 71, + 531, + 80 + ], + "lines": [ + { + "bbox": [ + 440, + 71, + 531, + 80 + ], + "spans": [ + { + "bbox": [ + 440, + 71, + 531, + 80 + ], + "type": "text", + "content": "How could the image be optimized to enhance the appeal of the rainbow?" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 118, + 417, + 122 + ], + "lines": [ + { + "bbox": [ + 405, + 118, + 417, + 122 + ], + "spans": [ + { + "bbox": [ + 405, + 118, + 417, + 122 + ], + "type": "text", + "content": "(3)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 404, + 124, + 535, + 134 + ], + "lines": [ + { + "bbox": [ + 404, + 124, + 535, + 134 + ], + "spans": [ + { + "bbox": [ + 404, + 124, + 535, + 134 + ], + "type": "text", + "content": "What is the characterization of the processing style of the photo's shadows?" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 441, + 88, + 497, + 93 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 88, + 497, + 93 + ], + "spans": [ + { + "bbox": [ + 441, + 88, + 497, + 93 + ], + "type": "text", + "content": "A. Increase the exposure time" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 441, + 94, + 495, + 105 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 441, + 94, + 495, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 94, + 495, + 99 + ], + "spans": [ + { + "bbox": [ + 441, + 94, + 495, + 99 + ], + "type": "text", + "content": "B. Decrease the ISO setting C. Add more artificial lighting" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 441, + 99, + 468, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 99, + 468, + 105 + ], + "spans": [ + { + "bbox": [ + 441, + 99, + 468, + 105 + ], + "type": "text", + "content": "D. Use a flash" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 441, + 110, + 476, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 110, + 476, + 115 + ], + "spans": [ + { + "bbox": [ + 441, + 110, + 476, + 115 + ], + "type": "text", + "content": "Correct Answer: B" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 496, + 110, + 528, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 496, + 110, + 528, + 115 + ], + "spans": [ + { + "bbox": [ + 496, + 110, + 528, + 115 + ], + "type": "text", + "content": "Topic: Exposure" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 71, + 172, + 236, + 269 + ], + "blocks": [ + { + "bbox": [ + 71, + 172, + 236, + 269 + ], + "lines": [ + { + "bbox": [ + 71, + 172, + 236, + 269 + ], + "spans": [ + { + "bbox": [ + 71, + 172, + 236, + 269 + ], + "type": "image", + "image_path": "4c3192dda3e12859d68ad7d41f1d2ff56551b85c3730016dc43a09bebac1d388.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 238, + 178, + 288, + 269 + ], + "blocks": [ + { + "bbox": [ + 238, + 178, + 288, + 269 + ], + "lines": [ + { + "bbox": [ + 238, + 178, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 238, + 178, + 288, + 269 + ], + "type": "image", + "image_path": "676f2ac2afd7c1f12593d6229edca88b8fb5369edbba75195f18afe765ad9ba0.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 290, + 178, + 538, + 270 + ], + "blocks": [ + { + "bbox": [ + 404, + 139, + 536, + 157 + ], + "lines": [ + { + "bbox": [ + 404, + 139, + 536, + 157 + ], + "spans": [ + { + "bbox": [ + 404, + 139, + 536, + 157 + ], + "type": "text", + "content": "A. Naturally soft with subtle transitions B. High contrast with distinct black and white tones C. Heavy-handed with raised blacks and crushed shadows." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 404, + 158, + 531, + 171 + ], + "lines": [ + { + "bbox": [ + 404, + 158, + 531, + 171 + ], + "spans": [ + { + "bbox": [ + 404, + 158, + 531, + 171 + ], + "type": "text", + "content": "D. Minimal processing with a high dynamic range Correct Answer: C Topic: Editing" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 172, + 411, + 177 + ], + "lines": [ + { + "bbox": [ + 405, + 172, + 411, + 177 + ], + "spans": [ + { + "bbox": [ + 405, + 172, + 411, + 177 + ], + "type": "text", + "content": "(6)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 290, + 178, + 538, + 270 + ], + "lines": [ + { + "bbox": [ + 290, + 178, + 538, + 270 + ], + "spans": [ + { + "bbox": [ + 290, + 178, + 538, + 270 + ], + "type": "image", + "image_path": "6e78541d024020c68ef12d8ca705443a1a85258fe3b15bd1ba630588ab939656.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 306, + 295, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 366 + ], + "type": "text", + "content": "technique achieved by raising the shadow curve to enhance the photo's texture or create a film-like aesthetic. They suggest that PhotoBench includes more professional and detailed questions, closely aligned with practical applications such as photo post-processing and editing." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 378, + 135, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 135, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 135, + 392 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "content": "Implementation Details. We build PhotoEye with Vicuna-v1.5-7B as the LLM backbone, following existing works [8, 30] for fair comparison. We also use minimal data from the general domain, including LLaVA-Pretrain [15] for pretraining and LLaVA-665K [15] for instruction finetuning. We mix our data with LLaVA-665K during the finetuning stage. We use a batch size of 128 and a learning rate of " + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "content": " for pre-training, and " + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "inline_equation", + "content": "2e - 5" + }, + { + "bbox": [ + 55, + 398, + 296, + 542 + ], + "type": "text", + "content": " for finetuning. The training takes around 96 hours with 8 A100 GPUs with ZeRO2. For vision encoders, we use: CLIP-ViT-L/14 [25], DINOv2-giant [20] CoDETR-ViT-L [40] and SAM-ViT-H [11]. Details are available in the Appendix D." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 55, + 551, + 112, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 112, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 112, + 562 + ], + "type": "text", + "content": "6.1. Results" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "content": "We report results of PhotoEye on existing benchmarks. Although PhotoCritique is not specifically designed for low-level vision, but PhotoEye still achieves competing results on Q-Bench when finetuned with it [29]. PhotoEye clearly outperforms existing open-source models, including both general-purpose models and those specifically fine-tuned on low-level vision and image aesthetics datasets, such as Q-Instruct [30] and AesExpert [8]. Compared to closed-source models, our lightweight approach also demonstrates competitive performance, slightly surpassing GPT-4o." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In Tab. 3, we compare our model with most competing baselines on PhotoBench, including UNIAA [38], AesEx" + } + ] + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 332, + 305, + 366, + 330 + ], + "blocks": [ + { + "bbox": [ + 55, + 280, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 280, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 555, + 301 + ], + "type": "text", + "content": "Figure 9. We show one rejected example, several accepted examples, topic distribution, and some sub-topics of a subset of topics of questions in PhotoBench." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 305, + 366, + 330 + ], + "lines": [ + { + "bbox": [ + 332, + 305, + 366, + 330 + ], + "spans": [ + { + "bbox": [ + 332, + 305, + 366, + 330 + ], + "type": "image", + "image_path": "ae1e0dc934debf50ad3eeffca4d3bbabb554ec71cc4e0057c8008ea37b96c742.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 308, + 533, + 327 + ], + "lines": [ + { + "bbox": [ + 369, + 308, + 533, + 327 + ], + "spans": [ + { + "bbox": [ + 369, + 308, + 533, + 327 + ], + "type": "text", + "content": "What aspect of the image is primarily used to enhance the aesthetic perception? A. The high level of clarity and detailed texture of the petals and stamens (Correct) B. The use of a monotone color palette C. The inclusion of multiple subjects in the background" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 332, + 331, + 366, + 361 + ], + "blocks": [ + { + "bbox": [ + 332, + 331, + 366, + 361 + ], + "lines": [ + { + "bbox": [ + 332, + 331, + 366, + 361 + ], + "spans": [ + { + "bbox": [ + 332, + 331, + 366, + 361 + ], + "type": "image", + "image_path": "1450f11a2aa286962dde5dcd968df0c7388e427692a20edcc4814fa7af4a3356.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 335, + 461, + 358 + ], + "lines": [ + { + "bbox": [ + 353, + 335, + 461, + 358 + ], + "spans": [ + { + "bbox": [ + 353, + 335, + 461, + 358 + ], + "type": "text", + "content": "What atmosphere can you feel in objects and colors? A. Joy & vitality B. Mystery & serenity (Correct) C. Sadness & despair D. Sweet & romance" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 369, + 552, + 380 + ], + "lines": [ + { + "bbox": [ + 315, + 369, + 552, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 369, + 552, + 380 + ], + "type": "text", + "content": "Figure 10. Examples in related works from their original papers." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 465, + 332, + 485, + 360 + ], + "blocks": [ + { + "bbox": [ + 465, + 332, + 485, + 360 + ], + "lines": [ + { + "bbox": [ + 465, + 332, + 485, + 360 + ], + "spans": [ + { + "bbox": [ + 465, + 332, + 485, + 360 + ], + "type": "image", + "image_path": "545906481829a67a1f95ebb83749827897e65581bdc68eb1132ad1353adc7a61.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 359, + 485, + 364 + ], + "lines": [ + { + "bbox": [ + 465, + 359, + 485, + 364 + ], + "spans": [ + { + "bbox": [ + 465, + 359, + 485, + 364 + ], + "type": "text", + "content": "Q-Instruct" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 485, + 334, + 534, + 358 + ], + "blocks": [ + { + "bbox": [ + 485, + 334, + 534, + 358 + ], + "lines": [ + { + "bbox": [ + 485, + 334, + 534, + 358 + ], + "spans": [ + { + "bbox": [ + 485, + 334, + 534, + 358 + ], + "type": "image", + "image_path": "3eaa214257140aa590bdf6105d8ff7ad65cdae5e67645faab15060726ece97d0.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 383, + 554, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 554, + 526 + ], + "type": "text", + "content": "pert [8] and Q-Instruct [30], which specially focus on low-level vision and aesthetics. Note that when reporting results in Tab. 3, we merge similar topics into a single category. Categories may overlap - for instance, a question about editing exposure would be counted in both Post-Processing and Exposure. PhotoEye notably outperforms baselines. As discussed in Section 5, PhotoBench shows substantially higher expertise and diversity than existing benchmarks, demanding stronger aesthetic understanding capabilities from models. Therefore, PhotoEye, which learns from extensive discussions by photography professionals and enthusiasts, holds a clear advantage." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 529, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 577 + ], + "type": "text", + "content": "Ablations. We show the effectiveness of our proposed vision fusor and dataset PhotoCritique in Tab. 4. Both contribute to the model. More ablations including qualitative results are available in Appendix." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 580, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 554, + 628 + ], + "type": "text", + "content": "Qualitative Evaluation. Providing aesthetic recommendations in conversations is actually the primary need in real-world scenarios. In addition to examples in Fig 1, We show examples in Appendix E. with diverse images and contexts." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 635, + 441, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 441, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 441, + 647 + ], + "type": "text", + "content": "6.2. See through PhotoEye" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "As multiple visual features from different vision encoders are fused in PhotoEye, we are particularly interested in their roles under different cases and how they contribute to our model. Fig. 11 presents the normalized weights of each vision encoder in PhotoEye. We first randomly col" + } + ] + } + ], + "index": 54 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24813" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 545, + 243 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 545, + 243 + ], + "type": "table", + "html": "
ModelLLM BackboneQuestion TypesQuadrants of Low-level ConcernsOverall↑
Yes-or-No↑What↑How↑Distortion↑Other↑In-context Distortion↑In-context Other↑
Random guess-50.0027.8633.3137.8938.4838.2835.8237.80
GPT-4o (2024-08-06)Proprietary Model76.7274.3367.7470.0373.3773.6877.9573.04
Qwen-VL-Max [2]Proprietary Model75.6079.4366.0973.3974.0871.0076.9273.63
Qwen-VL-Plus [2]Proprietary Model73.7769.4753.8866.2165.7263.8168.7566.04
Gemini-ProProprietary Model68.8073.7462.3466.3071.3463.9173.0968.16
Emu2-Chat [27]LLaMA-2-33B71.8167.2556.1864.7863.1963.4872.2465.28
mPLUG-Owl2 [34]LLaMA-2-7B72.1857.9656.1956.6869.2153.2972.6561.61
LLAMA-Adapter-V2 [7]LLaMA-2-7B66.1859.2952.1357.3956.2563.1664.9059.46
IDEFICS-Instruct [1]LLaMA-2-7B56.1844.6944.0242.8054.1744.7456.3348.70
SPHINX [14]LLaMA-274.1868.8162.0763.6271.7666.1276.3368.56
InternLM-XComposer-VL [36]InternLM69.4565.2760.8561.6770.1456.9175.1065.35
Qwen-VL [2]QwenLM63.0958.1956.3950.5862.7357.8973.8859.40
Otter-v1 [12]MPT-7B57.0940.7139.5542.2249.3144.0852.6546.35
LLAVA-v1.5 [15]Vicuna-v1.5-7B66.3658.1950.5149.4265.7454.6170.6158.66
InstructBLIP [6]Vicuna-v1.5-7B71.6452.6543.8148.6462.5055.5964.9056.72
Shikra [4]Vicuna-v1.5-7B65.6447.3549.0948.8359.4950.0064.0854.65
MiniGPT-4 [39]Vicuna-v1.5-13B55.8250.2240.3742.0248.3851.9761.2249.03
AesExpert [8]Vicuna-v1.5-7B73.2764.3853.7570.0373.3873.6877.9664.15
Q-Instruct [30]Vicuna-v1.5-7B76.1866.3757.6165.1867.5973.0671.5367.09
PhotoEye (Ours)Vicuna-v1.5-7B80.0176.1067.0274.3274.5977.3081.2274.50
", + "image_path": "d9e85eb70d07b7e9bc6f18d7429eabc07e988b82f23a694ea59111a57a30b3b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 272, + 541, + 389 + ], + "blocks": [ + { + "bbox": [ + 55, + 245, + 553, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 245, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 553, + 268 + ], + "type": "text", + "content": "Table 2. Model Performance on Q-Bench (LLVisionQA-dev). PhotoEye has a clear advantage. Best open-source model results in **bold**. Proprietary model results are quoted from Q-Bench [29]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 272, + 541, + 389 + ], + "lines": [ + { + "bbox": [ + 69, + 272, + 541, + 389 + ], + "spans": [ + { + "bbox": [ + 69, + 272, + 541, + 389 + ], + "type": "table", + "html": "
ModelLLM BackboneComposition↑Equipments↑Contrast↑Techniques↑Color and Tone↑Lighting↑
GPT-4o (2024-08-06)Proprietary Model68.1160.2459.0962.7563.9763.74
LLaVA-v1.6-34B [15]Hermes-2-Yi-34B47.6952.5656.8249.3365.4453.80
UNIAA [38]Vicuna-v1.5-7B29.8938.6145.4539.5640.1929.24
AesExpert [8]Vicuna-v1.5-7B51.4960.2456.8253.9961.0364.33
Q-Instruct [30]Vicuna-v1.5-7B31.0234.7740.9131.3754.3844.44
PhotoEye (Ours)Vicuna-v1.5-7B68.3261.3972.9765.6976.0077.78
ModelLLM BackboneExposure↑Post-Processing↑Aperture and Focus↑Storytelling↑Sharpness and Clarity↑Overall↑
GPT-4o (2024-08-06)Proprietary Model54.4858.5759.8769.4461.9064.12
LLaVA-v1.6-34B [15]Hermes-2-Yi-34B56.5553.3352.8766.6728.5755.68
UNIAA [38]Vicuna-v1.5-7B25.5238.5740.1330.5623.8136.87
AesExpert [8]Vicuna-v1.5-7B46.2161.9052.8772.9733.3260.01
Q-Instruct [30]Vicuna-v1.5-7B46.2146.1938.8555.5619.0543.86
PhotoEye (Ours)Vicuna-v1.5-7B69.6680.9570.7075.0061.9073.92
", + "image_path": "4388aea944e22507467632debf86fd65eb86b09901a5a22c3b5e004912520b85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 73, + 411, + 277, + 445 + ], + "blocks": [ + { + "bbox": [ + 169, + 392, + 440, + 403 + ], + "lines": [ + { + "bbox": [ + 169, + 392, + 440, + 403 + ], + "spans": [ + { + "bbox": [ + 169, + 392, + 440, + 403 + ], + "type": "text", + "content": "Table 3. Model Performance on PhotoBench Metrics. Best results in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 411, + 277, + 445 + ], + "lines": [ + { + "bbox": [ + 73, + 411, + 277, + 445 + ], + "spans": [ + { + "bbox": [ + 73, + 411, + 277, + 445 + ], + "type": "table", + "html": "
MethodQ-BenchPhotoEye
PhotoEye (full method)74.5073.92
w/o Multi-view Vision Fusor70.0868.83
w/o Multi-view Vision Fusor and dataset58.0433.74
", + "image_path": "c38fd10401b6ab3780323d8706eba504165fef3f3446a0899fae2c8b23403a4e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 468, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 297, + 696 + ], + "type": "text", + "content": "lect 100 samples from PhotoCritique on compositional aesthetics (e.g., photographic composition, distraction, framing, etc.) and general aesthetics (e.g., storytelling, lighting, etc), respectively, as well as 100 samples for general visual tasks from LLaVA [15] training data. Then we study the averaged weight distribution (normalized) of encoders on different data. We have an interesting observation: (1) aesthetic features are extracted in the first layers, and general features are extracted in the last layers. Fig. 11 shows CLIP and DINO features are dominant regardless of domains in the last layer, which is intuitive as the recognition of objects lays the foundation in both domains. In contrast, aesthetics-related features are primarily extracted from the first layer, especially in compositional problems, where the localization or arrangement of objects are critical. In this case, CoDETR (pre-trained with detection objectives) is dominant. Tab. 5 validated it, where we only activated one encoder during inference to detect their effectiveness in photographic composition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Case Study. We show how different choices of vision en" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 411, + 553, + 444 + ], + "blocks": [ + { + "bbox": [ + 315, + 411, + 553, + 444 + ], + "lines": [ + { + "bbox": [ + 315, + 411, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 315, + 411, + 553, + 444 + ], + "type": "image", + "image_path": "d8355cce4e3db4dd5bc8c9f5950cea519ec72325f0128b006b7e9e217254484d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 448, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 322, + 448, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 322, + 448, + 545, + 460 + ], + "type": "text", + "content": "Figure 11. Weights of vision encoders across different layers." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 328, + 466, + 539, + 489 + ], + "blocks": [ + { + "bbox": [ + 127, + 448, + 223, + 457 + ], + "lines": [ + { + "bbox": [ + 127, + 448, + 223, + 457 + ], + "spans": [ + { + "bbox": [ + 127, + 448, + 223, + 457 + ], + "type": "text", + "content": "Table 4. Ablation Results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 466, + 539, + 489 + ], + "lines": [ + { + "bbox": [ + 328, + 466, + 539, + 489 + ], + "spans": [ + { + "bbox": [ + 328, + 466, + 539, + 489 + ], + "type": "table", + "html": "
TopicCLIP-ViT-L/14DINOv2-giantCoDETR-ViT-LSAM-ViT-H
Composition62.6161.6164.6960.24
", + "image_path": "621fabbdab22ce41168d41465354b64ec60f70f8bedd6e0fddf98a866fb6126d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 491, + 541, + 502 + ], + "lines": [ + { + "bbox": [ + 326, + 491, + 541, + 502 + ], + "spans": [ + { + "bbox": [ + 326, + 491, + 541, + 502 + ], + "type": "text", + "content": "Table 5. PhotoEye on composition with different encoders." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "text", + "content": "coders affect model output in real-world conversation scenarios. We provide interesting examples in Appendix F." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 536, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 608 + ], + "type": "text", + "content": "Learning of Aesthetic Concepts. Fig. 8 demonstrates how PhotoEye effectively learns discriminative aesthetics-related visual representations. In Appendix B, we conduct a deeper investigation into the learning of aesthetic concepts at feature level, which lays the foundation of high-quality aesthetic response from LLM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 620, + 388, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 388, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 388, + 633 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "We present PhotoCritique dataset, and PhotoEye, an MLLM with language-guided multi-view vision fusion for aesthetic visual understanding, along with the new benchmark PhotoBench. Extensive experiments show PhotoEye's clear advantage over existing models, advancing aesthetic visual perception in MLLMs." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24814" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 124 + ], + "type": "text", + "content": "This work was partially supported by gifts from Adobe Research and the National Science Foundation under Grants IIS-2316306 and CNS-2330215." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 163, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 163, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 296, + 219 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 220, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 296, + 273 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 296, + 339 + ], + "type": "text", + "content": "[3] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 340, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 295, + 384 + ], + "type": "text", + "content": "[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 384, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 449 + ], + "type": "text", + "content": "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 450, + 296, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 450, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 62, + 450, + 296, + 504 + ], + "type": "text", + "content": "[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "text", + "content": "[7] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2311.07575, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 560, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 560, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 560, + 296, + 613 + ], + "type": "text", + "content": "[8] Yipo Huang, Xiangfei Sheng, Zhichao Yang, Quan Yuan, Zhichao Duan, Pengfei Chen, Leida Li, Weisi Lin, and Guangming Shi. AesExpert: Towards multi-modality foundation model for image aesthetics perception. arXiv preprint arXiv:2404.09624, 2024. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "type": "text", + "content": "[9] Yipo Huang, Quan Yuan, Xiangfei Sheng, Zhichao Yang, Haoning Wu, Pengfei Chen, Yuzhe Yang, Leida Li, and Weisi Lin. AesBench: An expert benchmark for multimodal large language models on image aesthetics perception. arXiv preprint arXiv:2401.08276, 2024. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[10] Zhipeng Huang, Zhizheng Zhang, Yiting Lu, Zheng-Jun Zha, Zhibo Chen, and Baining Guo. Visualcritic: Making Imm's perceive visual quality like humans. arXiv preprint arXiv:2403.12806, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[12] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023.8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 554, + 237 + ], + "type": "text", + "content": "[13] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Jiaming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. SPHINX: The Joint Mixing of Weights, Tasks, and Visual Embeddings for Multi-modal Large Language Models. arXiv preprint arXiv:2311.07575, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 238, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 554, + 304 + ], + "type": "text", + "content": "[14] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, Ji-aming Han, Siyuan Huang, Yichi Zhang, Xuming He, Hongsheng Li, and Yu Qiao. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2024. 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 305, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 338 + ], + "type": "text", + "content": "[15] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 5, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 554, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 371 + ], + "type": "text", + "content": "[16] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "text", + "content": "[17] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415. IEEE, 2012. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "text", + "content": "[18] Daniel Vera Nieto, Luigi Celona, and Clara Fernandez-Labrador. Understanding aesthetics with language: A photo critique dataset for aesthetic assessment. In NeurIPS Track on Datasets and Benchmarks, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 485, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 485, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 485, + 471 + ], + "type": "text", + "content": "[19] OpenAI. Gpt-4 technical report. 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 472, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 554, + 525 + ], + "type": "text", + "content": "[20] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 554, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 570 + ], + "type": "text", + "content": "[21] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 571, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 554, + 613 + ], + "type": "text", + "content": "[22] Daiqing Qi, Handong Zhao, and Sheng Li. Easy regional contrastive learning of expressive fashion representations. Advances in Neural Information Processing Systems, 37: 20480-20509, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 668 + ], + "type": "text", + "content": "[23] Daiqing Qi, Handong Zhao, Zijun Wei, and Sheng Li. Tag-grounded visual instruction tuning with retrieval augmentation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 2008-2026, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 714 + ], + "type": "text", + "content": "[24] Daiqing Qi, Handong Zhao, Aidong Zhang, and Sheng Li. Generalizing to unseen domains via text-guided augmentation: A training-free approach. In European Conference on Computer Vision, pages 285-300. Springer, 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24815" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[26] Jian Ren, Xiaohui Shen, Zhe Lin, Radomir Mech, and David J. Foran. Personalized image aesthetics. In The IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "text", + "content": "[27] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 243, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 306 + ], + "type": "text", + "content": "[28] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 309, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 374 + ], + "type": "text", + "content": "[29] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Chunyi Li, Wenxiu Sun, Qiong Yan, Guangtao Zhai, and Weisi Lin. Q-BENCH: A benchmark for general-purpose foundation models on low-level vision. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 453 + ], + "type": "text", + "content": "[30] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, Geng Xue, Wenxiu Sun, Qiong Yan, and Weisi Lin. Q-Instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 455, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 520 + ], + "type": "text", + "content": "[31] Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, Xiaohong Liu, Guangtao Zhai, Shiqi Wang, and Weisi Lin. Towards open-ended visual quality comparison. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 522, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 577 + ], + "type": "text", + "content": "[32] Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aesthetics assessment with rich attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19861-19869, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 579, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 632 + ], + "type": "text", + "content": "[33] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[34] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[35] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 396 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "text", + "content": "[36] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "text", + "content": "[37] Zhipeng Zhong, Fei Zhou, and Guoping Qiu. Aesthetically relevant image captioning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "text", + "content": "[38] Zhaokun Zhou, Yiwei Su, Amin Zheng, Qiulin Wang, Rui Chen, Li Yuan, and Di Zhang. UNIAA: A unified multimodal image aesthetic assessment baseline and benchmark. arXiv preprint arXiv:2404.09619, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "text", + "content": "[39] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 297, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 341 + ], + "type": "text", + "content": "[40] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6758, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "text", + "content": "[41] Zhuofan Zong, Bingqi Ma, Dazhong Shen, Guanglu Song, Hao Shao, Dongzhi Jiang, Hongsheng Li, and Yu Liu. Mova: Adapting mixture of vision experts to multimodal context. In Proceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 6" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24816" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_content_list.json b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d2bf9f1120ccd7cb2edc54e0afc2f6694cce128d --- /dev/null +++ b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_content_list.json @@ -0,0 +1,2318 @@ +[ + { + "type": "text", + "text": "The Power of Context: How Multimodality Improves Image Super-Resolution", + "text_level": 1, + "bbox": [ + 106, + 114, + 890, + 138 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kangfu Mei\\*1,2, Hossein Talebi', Mojtaba Ardakani', Vishal M. Patel?, Peyman Milanfar', Mauricio Delbracio' 1 Google, 2 Johns Hopkins University Project Page: https://mmsr.kfmei.com/", + "bbox": [ + 267, + 164, + 725, + 237 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ddbe435e00ca0c7626f7fa2886ba73af9bd9e8569079aa84632bc6bfe743901.jpg", + "image_caption": [ + "Figure 1. Our Multimodal Super-Resolution (MMSR) method leverages the rich context of multimodal guidance, including image captions, depth maps, semantic segmentation maps, and edges inferred from LR. MMSR surpasses state-of-the-art methods by producing more realistic results and suppressing artifacts that, while plausible, are inconsistent with the information present in the LR input." + ], + "image_footnote": [], + "bbox": [ + 94, + 243, + 923, + 410 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 256, + 470, + 316, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-image super-resolution (SISR) remains challenging due to the inherent difficulty of recovering fine-grained details and preserving perceptual quality from low-resolution inputs. Existing methods often rely on limited image priors, leading to suboptimal results. We propose a novel approach that leverages the rich contextual information available in multiple modalities – including depth, segmentation, edges, and text prompts—to learn a powerful generative prior for SISR within a diffusion model framework. We introduce a flexible network architecture that effectively fuses multimodal information, accommodating an arbitrary number of input modalities without requiring significant modifications to the diffusion process. Crucially, we mitigate hallucinations, often introduced by text prompts, by using spatial information from other modalities to guide regional text-based conditioning. Each modality's guidance strength can also be controlled independently, allowing steering outputs toward different directions, such as increasing bokeh through depth or adjusting object prominence via segmentation. Extensive experiments demonstrate that our model surpasses state-of-the-art generative SISR methods, achieving superior visual quality and fidelity.", + "bbox": [ + 122, + 492, + 452, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 823, + 220, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single image super-resolution (SISR) aims to generate high-resolution images from low-resolution inputs while preserving", + "bbox": [ + 89, + 848, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing semantic identity and texture details. While it is not essential to treat SISR as a regression problem, past methods [17, 29, 37, 62] typically use a deep neural network to learn a direct mapping from low-resolution images to high-resolution images. Even though these regression-based methods achieve good scores on paired metrics like PSNR and SSIM, they have failed to produce results with high quality comparable to natural images, a task at which recent generative models have excelled [5, 12, 14]. The advent of powerful generative models, such as autoregressive models [10, 51] and diffusion models [47, 53, 54], has revolutionized image generation tasks, including text-to-image synthesis. This has inspired recent efforts to leverage these pre-trained generative models for downstream tasks like SISR [6, 41, 53]. For instance, very recent works [50, 71] achieve super-resolution by leveraging emerging vision-language models (e.g., Gemini [56], LLaVA [36], ChatGPT-4 [1]) and pretrained text-to-image models to first generate captions from low-resolution images and then use these captions as prompts to generate high-resolution images.", + "bbox": [ + 511, + 470, + 908, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While providing rich textual descriptions can significantly enhance the quality of generated images [4, 19, 34], relying solely on text for SISR poses challenges. Recent works [7, 26, 35, 77] have shown that text prompts cannot represent spatial relationships. This implies that textual information, such as texture descriptions, can only be applied to the whole image. Figure 1 provides a representative example. Previous text-based super-resolution methods [65, 68] use a 'lion'", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during an internship at Google.", + "bbox": [ + 107, + 886, + 388, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23141", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "caption, which results in a furry tongue. However, while lions have fur, their tongues do not grow hair. Can we leverage spatial cues from depth and segmentation data to improve the learned prior and enhance the quality of SISR?", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fortunately, we can directly extract additional modalities from the low-resolution image by using various pretrained cross-modal prediction models [11, 66]. In this paper, we introduce a new diffusion model architecture dedicated to multimodal super-resolution, which is conceptually illustrated in Figure 2. By integrating multiple modalities into a single diffusion model, our method overcomes the challenges of recovering fine-grained details and preserving perceptual quality. Specifically, we propose conditioning diffusion models on modalities including text captions, semantic segmentation maps, depth maps, and edges to implicitly align text captions for correctly prompting different regions.", + "bbox": [ + 88, + 152, + 482, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the proposed multimodal diffusion model on the SISR task. Our effective multimodal architecture achieves better realism in SISR results than the best text-driven method and largely eliminates hallucinations that do not match the input. Moreover, we find that the multimodal SISR enables a new controlling feature, where we can explicitly adjust the weights of each modality to steer the generated results in different directions.", + "bbox": [ + 89, + 335, + 483, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 455, + 413, + 470 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We demonstrate the effectiveness of token-wise encoding for seamlessly injecting multiple modalities into pretrained text-to-image diffusion models without architectural modifications or significant model size overhead.", + "- We propose a novel multimodal latent connector that efficiently fuses information from different modalities, maintaining linear time complexity with respect to the number of modalities.", + "- We introduce a new multimodal classifier-free guidance technique that enhances realism at higher guidance rates while mitigating excessive hallucinations and fake details.", + "- Our method enables adjustment of the influence of each modality, allowing for fine-grained manipulation of SISR results while preserving realism and quality." + ], + "bbox": [ + 91, + 476, + 483, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 708, + 232, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Prior Powered Super-resolution. Recent advances in image restoration leverage the power of foundational models to enhance the quality of degraded images [9, 13, 20, 33, 38, 39, 46, 50, 57, 64, 69, 72, 76]. This trend is fueled by the capacity of pretrained generative models to capture natural image statistics and transfer this knowledge to the super-resolution task, enabling photorealistic image generation. Early works in this domain include LDM-SR [53] and StableSR [59] for single image superresolution. Beyond this, methods like InstructPix2Pix [6] utilize instructions for image editing, while ControlNet [74]", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/25b71099370a5a1779372762ce898e29b44e412e97821444b001b59b9f22fb81.jpg", + "image_caption": [ + "Figure 2. Language models struggle to accurately represent spatial information, leading to coarse and imprecise image superresolution. To overcome this limitation, we incorporate additional spatial modalities like depth maps and semantic segmentation maps. These modalities provide detailed spatial context, allowing our model to implicitly align language descriptions with individual pixels through a transformer network. This enriched understanding of the image significantly enhances the realism of our super-resolution results and minimizes distortion." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 903, + 188 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and IP-Adapter [70] facilitate cross-modality image translation. These approaches highlight a common observation: the quality of results in downstream tasks is strongly correlated with the quality of the pretrained models [40].", + "bbox": [ + 511, + 342, + 906, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works focus on leveraging powerful text-to-image diffusion models, employing text prompts to fully exploit the learned prior. PASD [68] uses image content descriptions; SPIRE [49] and PromptIP [48] use degradation descriptions; and SeeSR [65] employs a combination of context and degradation descriptions. Compared to earlier methods that directly fine-tuned diffusion models on super-resolution data, these text-prompt-driven approaches not only achieve superior realism but also enable multi-faceted outputs by conditioning on different prompts [20]. This capability enriches the typically single-output super-resolution task.", + "bbox": [ + 511, + 405, + 908, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While text prompts offer advantages, they can be ambiguous in representing spatial relationships [4, 7, 19, 34]. Our proposed MMSR framework addresses this by integrating multimodal inputs within a novel architecture. We also demonstrate native super-resolution effects control with this new architecture, enabling control of both overall realism and the effect of each modality.", + "bbox": [ + 511, + 573, + 908, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language Understanding and Generation. Recent large vision-language models (e.g., Gemini [56], LLaVA [36], GPT-4o [1]) excel in tasks like image captioning [61], but translating visual information into diverse modalities for image generation remains challenging. While models like 4M [2, 43] effectively extract high-level visual information (depth, normals, semantics, etc.), leveraging these for generation is a promising direction. Recent exploration includes GLIGEN [30] generate images from bounding boxes, and ControlNet [74] generate images from depth and other modalities. More recent works [21, 22] explore using discrete hidden-feature tokens to guide generation. In this paper, we introduce a new approach for multimodal control with discrete vision tokens, effectively integrating depth maps,", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23142", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6520790bb98cff0df4b5f817ac4cb7189a2075aa9ecdbcfe5e30fd81417f3ab1.jpg", + "image_caption": [ + "Figure 3. This diagram illustrates our multimodal super-resolution pipeline. Starting with a low-resolution (LR) image, we extract modalities like depth and semantic segmentation maps. These modalities are encoded into tokens and transformed into multimodal latent tokens $(m)$ . Our diffusion model uses these tokens and the LR input to generate a high-resolution (SR) output. A multimodal classifier-free guidance $(m$ -cfg) refines the SR image for enhanced quality." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "semantic segmentation maps, and text prompts for superior performance. Experiments on real-world super-resolution demonstrate the effectiveness of our approach.", + "bbox": [ + 89, + 367, + 482, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 429, + 181, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single-image super-resolution aims to recover a high-resolution image $\\mathbf{x}$ from its low-resolution counterpart $\\mathbf{x}_{LR}$ . This is an ill-posed problem, often leading to generative models that produce \"hallucinated\" details – plausible yet inconsistent with the input. To mitigate this, we introduce auxiliary information, such as depth $(m_{\\mathrm{dep}})$ , semantic segmentation $(m_{\\mathrm{seg}})$ , and edge maps $(m_{\\mathrm{edg}})$ , collectively denoted as $m$ . By conditioning the generative process on both the low-resolution image and these auxiliary modalities, we propose a new distribution $p(\\mathbf{x}|\\mathbf{x}_{LR}, m)$ with reduced uncertainty compared to the original distribution $p(\\mathbf{x}|\\mathbf{x}_{LR})$ .", + "bbox": [ + 88, + 455, + 482, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This reduction in uncertainty can be understood through the lens of information theory. The auxiliary modality $m$ provides additional information about the high-resolution image $\\mathbf{x}$ that is not present in the low-resolution input $\\mathbf{x}_{LR}$ . Since (conditional) mutual information is non-negative:", + "bbox": [ + 89, + 623, + 483, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (\\mathbf {x}; m | \\mathbf {x} _ {L R}) = H (p (\\mathbf {x} | \\mathbf {x} _ {L R})) - H (p (\\mathbf {x} | \\mathbf {x} _ {L R}, m)), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 704, + 480, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $H(\\cdot)$ denotes entropy. Consequently, the entropy of the conditional distribution with the auxiliary modality is:", + "bbox": [ + 89, + 724, + 482, + 753 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH \\left(p \\left(\\mathbf {x} \\mid \\mathbf {x} _ {L R}\\right)\\right) \\geq H \\left(p \\left(\\mathbf {x} \\mid \\mathbf {x} _ {L R}, m\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 758, + 480, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The same motivation for reducing uncertainty is also shared by recent diffusion guidance works [23, 44]. While low-entropy sampling does not guarantee sharper images, it often leads to outputs with better visual quality compared to standard sampling. Motivated by this observation, we introduce a diffusion model [24] to learn the multimodal distribution $p(\\mathbf{x}|\\mathbf{x}_{LR},m)$ , effectively incorporating auxiliary information to enhance SISR quality.", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ideally, the auxiliary modalities should provide information complementary to the low-resolution input. In practice, the auxiliary modalities are derived from the high-resolution image during training, ensuring informative conditioning. While we use modalities derived from the low-resolution input during inference, we demonstrate that this still leads to superior performance, including higher-quality details and improved fidelity to the input.", + "bbox": [ + 511, + 366, + 906, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Unified Multimodal Diffusion Conditioning", + "text_level": 1, + "bbox": [ + 511, + 503, + 880, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a new diffusion network architecture, illustrated in Figure 3, for simultaneously conditioning on multiple modalities. Unlike recent methods like ControlNet [74] and IP-Adapter [70], which duplicate network components for each modality and incur significant computational overhead, our approach leverages a pretrained VQGAN image tokenizer [18]. This allows us to encode diverse modalities into a unified token representation for conditioning the diffusion model, without introducing additional model parameters or modifying the diffusion network itself. These tokens are concatenated with the text prompt embedding and used for cross-attention within the diffusion model. To efficiently process this long token sequence, we introduce a lightweight multimodal connector. This connector employs a dedicated architecture to achieve linear complexity for cross-attention, significantly reducing the computational burden. In what follows we present the implementation details.", + "bbox": [ + 511, + 529, + 906, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token-wise Multimodal Encoding. While VQGAN [18] has proven effective for cross-modal encoding in image understanding and generation [2, 3, 43], its optimal application for image super-resolution remains an open question. Unlike tasks focused on understanding or generation, superresolution requires tokenization to not only capture semantic information but also preserve pixel-wise details crucial for", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23143", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/778161cd7d44af9759e8652466760e5134190732f4932803dcd9db98f7c50262.jpg", + "image_caption": [ + "Figure 4. Using discrete multimodal tokens leads to superior reconstruction of modalities compared to continuous tokens." + ], + "image_footnote": [], + "bbox": [ + 96, + 97, + 194, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/78c20180567971698edb70f60abe6d3b0781aa0e177ac5af7b515c19e18469e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 97, + 290, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de0f768b1e15772c2c9fa84ae6c9b57066674ba2bc5c515a541ea8909b7d25d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 97, + 388, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/777b723c7a4b9b2cb2ca8e6a86951d08dc2044a8d2821b619e02dc04e9837df9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 97, + 490, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "accurate reconstruction, such as spatial relationships within a depth map. Inspired by recent discussions on using discrete or continuous tokenization for autoregressive image generation [55], we investigate the impact of quantization on multimodal super-resolution. Specifically, we analyze how different quantization strategies affect the reconstruction quality of individual modalities and the final generated image. To this end, we utilize a VQGAN model pre-trained on a large-scale multimodal image dataset.", + "bbox": [ + 89, + 271, + 483, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 4 compares the reconstruction quality of multi-modal tokens before and after quantization. As the figure illustrates, discrete tokens (post-quantization) better preserve individual modality information, while continuous tokens introduce noticeable artifacts. Consequently, we employ discrete tokens for our super-resolution experiments.", + "bbox": [ + 89, + 407, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our multimodal encoding process utilizes the encoder and quantizer of the pre-trained VQGAN. Each $256 \\times 256$ input modality is encoded into a $16 \\times 16$ multimodal token sequence with a feature dimension of 256. These tokens are then quantized using a codebook of size 1024. To facilitate concatenation with text embeddings, we pad the feature dimension of the multimodal tokens to 1024, resulting in a $(256 \\times 3 + 77) \\times 1024$ multimodal token sequence condition.", + "bbox": [ + 89, + 500, + 483, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal Latent Connector. While cross-attention provides a flexible mechanism for conditioning diffusion models on multimodal data, its quadratic complexity with respect to the number of condition tokens introduces a significant computational burden. To address this, we introduce the Multimodal Latent Connector (MMLC), inspired by recent advances in efficient attention architectures [25, 28, 60].", + "bbox": [ + 89, + 628, + 483, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, the MMLC employs a transformer architecture to efficiently process the multimodal token sequence. The transformer receives two inputs: a randomly initialized sequence of learnable latent tokens, and the multimodal input sequence. The output is a token sequence of the same length as the latent token sequence, which serves as conditioning for the diffusion model. Therefore, the diffusion model conditions on fixed-length latent tokens (128 in our experiments), which are significantly shorter than the original multimodal token sequence input. The MMLC distills the essential information from the longer multimodal", + "bbox": [ + 89, + 734, + 485, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "token sequence into the shorter multimodal latent token sequence through cross-attention. Following cross-attention, several self-attention blocks further process the latent token sequence, allowing the model to fully integrate the distilled information.", + "bbox": [ + 511, + 90, + 906, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach significantly reduces the computational cost of cross-attention in the diffusion model. Standard self-attention operates on the full multimodal token sequence $(K, V \\in \\mathbb{R}^{M \\times D})$ , resulting in a time complexity of $\\mathcal{O}(M^2)$ where $M$ is the length of the sequence and $D$ is the dimensionality of each token. In contrast, the MMLC uses a cross-attention between the latent token sequence (of size $N \\times D$ ) and the multimodal token sequence which reduces this to $\\mathcal{O}(MN)$ . Here $N$ is the length of the latent sequence and $N \\ll M$ . This linear complexity with respect to the multimodal sequence length enables efficient processing of high-dimensional multimodal data, effectively capturing essential information for super-resolution.", + "bbox": [ + 511, + 167, + 908, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Flexible Multimodal Input. To enhance flexibility and robustness, we enable our method to handle scenarios where certain input modalities are unreliable or unavailable. We adopt a learnable embedding approach inspired by DALL-E 2 [52]. Specifically, we introduce a special learnable token, $m_{\\emptyset}$ , optimized alongside the diffusion model and MMLC to represent the absence of a modality. During training, we independently randomly replace each modality with 256 $m_{\\emptyset}$ tokens with a probability of 0.1. This encourages the model to learn robust representations that can effectively handle missing information. During inference, any missing modality is represented by a sequence of 256 $m_{\\emptyset}$ tokens. This approach allows for flexible multimodal input, enabling the model to generate high-quality images even with limited or no auxiliary modalities. As demonstrated in our experiments, this strategy significantly improves performance when dealing with input that includes fewer modalities than were available during training.", + "bbox": [ + 511, + 369, + 908, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multimodal Guidance and Control", + "text_level": 1, + "bbox": [ + 511, + 654, + 816, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building upon the success of guidance techniques [15, 23] in improving sample quality across various image generation tasks [45, 47, 52], recent diffusion-based SISR methods have begun incorporating prompt tuning to enhance superresolution results [59, 65, 71]. These methods improve the result by using negative prompts with Classifier-free Guidance (CFG) [23]. which can be expressed as:", + "bbox": [ + 511, + 678, + 908, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} (\\mathbf {z} _ {t}, c) = (1 + w) \\epsilon (\\mathbf {z} _ {t}, c, \\text {p o s}) - w \\epsilon (\\mathbf {z} _ {t}, c, \\text {n e g}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 789, + 903, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\epsilon} (\\mathbf{z}_t,c)$ represents the guided denoising process, $\\pmb {\\epsilon}(\\cdot)$ denotes the diffusion model, $\\mathbf{z}_t$ denotes the noisy image latent, $c = \\{\\mathbf{x}_{\\mathrm{LR}},t,\\dots \\}$ denotes the conditioning inputs (including the low-resolution image $\\mathbf{x}_{\\mathrm{LR}}$ , timestep $t$ , and other parameters), $w$ is the guidance scale, and pos and neg are the positive and negative prompts. While increasing $w$", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23144", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1060f0d42e54684ab8ed1a13af25cf12e53f001bec4124c2a59cf616fa1af652.jpg", + "image_caption": [ + "LR", + "Inputs" + ], + "image_footnote": [ + "The image shows the words \"CIPAL PARK\" prominently displayed on a metal structure. Below the main sign are several smaller screens displaying baseball ..." + ], + "bbox": [ + 99, + 97, + 209, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a90369b5705d808dfc40a768960412bd9d15a09df82ff239e2e478506dc03c9.jpg", + "image_caption": [ + "Depth", + "Figure 5. MMSR super-resolution results on real-world images compared with state-of-the-art methods. Zoom in to appreciate the details." + ], + "image_footnote": [], + "bbox": [ + 99, + 176, + 207, + 238 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4f545a9c5c14865e7203b02e458015917361c3541f2e77876d07f61852e4ebaa.jpg", + "image_caption": [ + "Caption", + "Edge" + ], + "image_footnote": [], + "bbox": [ + 218, + 176, + 321, + 238 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/98ac589cd0485996e53efff4146db3c8145996c837b2dcc0901d82d5e7835835.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 106, + 442, + 166 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/60693ef1d446c7837706e7127d90bd34adbd3cc50e5c8ed824aff947a3e5faa3.jpg", + "image_caption": [ + "Patch 1", + "Patch 2" + ], + "image_footnote": [], + "bbox": [ + 336, + 176, + 442, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd66632d97f0b46f0416f073458b1c016ec0d590cd1b92b0f764a1cec363d657.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 106, + 558, + 166 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/50814422c0a98a2ecd1e24cbe3b9b3fea6d9b8139b02ed309b50bdaee2490113.jpg", + "image_caption": [ + "PASD (Zoomed)", + "PASD (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 452, + 176, + 558, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/833ec91aea8fe8e1ecf98e34e60597041ba53bdd973183c3ef4407daed954497.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 106, + 671, + 166 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc7ac3e89e425fc07566a04ef5f0a8f7d0b5a4132ecf15da3698999e99445db0.jpg", + "image_caption": [ + "SeeSR (Zoomed)", + "SeeSR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 566, + 176, + 671, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/036d27b098b7638971fe0d417f508518019c019493bee1f94420996ebae5892e.jpg", + "image_caption": [ + "uts" + ], + "image_footnote": [], + "bbox": [ + 681, + 106, + 785, + 166 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e598cb327a5a8912fa75cf7a203ec72f9a1d9fc7d83922afdd84ca584579758f.jpg", + "image_caption": [ + "SUPIR (Zoomed)", + "SUPIR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 681, + 176, + 785, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9be8fecc9e1b9c9d878a2c4007a38d68df0f1a16ed942cc510bb87522c7ffcdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 106, + 898, + 166 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/84ca9552e002008f2d6748c92be9e1993d5c344e850a9b4117a81f280a89b647.jpg", + "image_caption": [ + "MMSR (Zoomed)", + "MMSR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 794, + 176, + 898, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "often leads to sharper and more detailed outputs, it can also exacerbate hallucination, resulting in details inconsistent with the low-resolution input. Such artifacts are widely reported but are difficult to suppress, even with recent efforts like balancing the training data [71].", + "bbox": [ + 89, + 279, + 483, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Classifier-free Guidance. To mitigate the issue of excessive hallucination often associated with high guidance scales in CFG, we propose a novel multimodal guidance strategy. We argue that the artifacts come from the weak guidance in the negative prompting process. Instead of simply relying on text prompts for guidance, we leverage the rich information encoded in the multimodal latent tokens to strengthen both positive and negative promptings. Specifically, we condition both the postive and negative generation processes on the multimodal latent token sequence, denoted as $m$ . This leads to the following multimodal CFG:", + "bbox": [ + 89, + 363, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {\\epsilon}} (\\mathbf {z} _ {t}, c, m) = (1 + w) \\boldsymbol {\\epsilon} (\\mathbf {z} _ {t}, c, \\text {p o s}, m) - w \\boldsymbol {\\epsilon} (\\mathbf {z} _ {t}, c, \\text {n e g}, m). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 534, + 480, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sec. 4.1 shows that this change in negative generation helps to achieve a better trade-off between perceptual quality and identity preservation, better maintaining the semantic content of the low-resolution input in the upscaled output, compared with the standard CFG in previous text-based methods.", + "bbox": [ + 89, + 565, + 483, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scaling Single-modal Guidance. Multimodal CFG effectively controls the overall influence of the prompts but does not offer control over each modality individually. To address this limitation, we introduce a mechanism to selectively amplify or suppress the contribution of specific modalities. Specifically, we modify the attention temperature $\\delta$ of MMLC during cross-conditioning, where $\\delta$ scales the attention maps before applying the softmax operation:", + "bbox": [ + 89, + 648, + 483, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\delta}\\right) V. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 773, + 482, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This parameter controls the sensitivity of the attention mechanism to differences between the feature sequence $Q$ and the multimodal token sequences $K$ and $V$ . A smaller temperature $\\delta$ typically leads to a stronger conditioning effect on the attention mechanism. In the standard scaled dot-product attention, the temperature $\\delta$ is empirically set to the square", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "root of the key dimension $\\sqrt{d_k}$ . We observe that scaling the standard temperature $\\sqrt{d_k}$ (where $d_{k}$ is the key dimension) within a range of [0.4, 10] can produce high-quality results, with varying levels of fidelity to the input modalities. By scaling the temperature during sampling, we achieve precise control over the super-resolution process, enabling manageable manipulation of the output with fine-grained control.", + "bbox": [ + 511, + 279, + 906, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. MMSR Implementation", + "text_level": 1, + "bbox": [ + 511, + 414, + 733, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 illustrates our multimodal guided super-resolution pipeline. During inference, we first extract four modalities from the low-resolution (LR) image: a text caption generated by Gemini Flash [56], a depth map estimated by Depth Anything [66], a semantic segmentation mask produced by Mask2Former [11], and edge information extracted with a Canny edge detector. The depth map, segmentation mask, and edge information are encoded into token sequences using a pretrained VQGAN, while the text caption is processed by a pretrained CLIP encoder to obtain a text embedding.", + "bbox": [ + 511, + 439, + 906, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following a similar conditioning strategy to InstructPix2Pix [6], the LR image is concatenated with a noisy latent vector sampled from the diffusion model, providing additional conditioning.", + "bbox": [ + 511, + 592, + 906, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 670, + 643, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. Our super-resolution model is initialized with the weights of a pretrained text-to-image model, with the same architecture and size as Stable Diffusion v2 [53]. The super-resolution training dataset consists of randomly degraded (using RealESRGAN degradation [63]) high-resolution images from the combined LSDIR and DIV2K datasets, with corresponding $512 \\times 512$ high-resolution images as ground truth. We set the batch size to 1024 and the learning rate to 1e-4, which we empirically found maximized compute efficiency on TPUv5 pods. During testing, all benchmarks use a model checkpoint finetuned for 160k iterations. We use 50-step DDIM sampling, consistent with previous methods. We use a guidance rate of 4 as the default.", + "bbox": [ + 511, + 704, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23145", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dcea7f995f1d0693ee265c99b3e67182fb67e9be509651809d332ed6a681da02.jpg", + "image_caption": [ + "LR" + ], + "image_footnote": [], + "bbox": [ + 94, + 94, + 207, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2201a59f269c441e560bdc563e45adaba72cc1d1a9e2c2487d2b548c9344e1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 94, + 326, + 166 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d3e1a10217413595a67c657828064a0813cdf6d54f1f4b3a8d5e4b1fffe1473.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 102, + 444, + 166 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9082679ea62fd50d13c372b0a732deecf47e369e1f6a7fc940d5401c2d5c58c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 95, + 566, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/30601950487f29035e5a733103e139de74abf7cc3ca9ecfcd45f0f0675c2c4c8.jpg", + "image_caption": [ + "SeeSR" + ], + "image_footnote": [], + "bbox": [ + 573, + 95, + 681, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0727b30a5320d57fb4837559716ee716c86593f2b70db06aff85e81349a4be6.jpg", + "image_caption": [ + "MMSR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 689, + 95, + 800, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f2dc4043b5d4560bff38ed12e3939838b5219334924ca5638deab2e07140ef09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 95, + 916, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/314ca8417f6cf02d36615c6c51eaffc24ff25179e5e80a9bddffdbd22e5ad915.jpg", + "image_caption": [ + "LR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 101, + 175, + 210, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e08e88b3aed45a685a8207378e87b2cd3b40555df28cd538535a690a724c1ab0.jpg", + "image_caption": [ + "Segmentation" + ], + "image_footnote": [], + "bbox": [ + 218, + 167, + 325, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6eec81efe8334cc29d08fab95388c76579c21ccfc5b1949eee8e7a884ffeccd.jpg", + "image_caption": [ + "Edge" + ], + "image_footnote": [], + "bbox": [ + 336, + 167, + 444, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9431c23601e9346875331e0173db4a29fe1e2230252b61d94d312ab895255693.jpg", + "image_caption": [ + "PASD (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 460, + 167, + 568, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b27a4e98c9c4c1f065417b6eb2c34006071329e85877897a74d1fa23e22351bb.jpg", + "image_caption": [ + "SeeSR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 576, + 176, + 684, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dadc586c4b6dfd2ecff2efbd796b7b4f6b91485a73f4a815915d83f083606e5a.jpg", + "image_caption": [ + "MMSR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 692, + 176, + 800, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20ecb5709aa00efd3f458ea02d8003c04b4c1f8d0b14696c674c34b7aadb74aa.jpg", + "image_caption": [ + "HR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 807, + 176, + 918, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ac961b0687907b4b5c6f3badd1135758699326bf84bdd42424fa845c7a31af8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 258, + 205, + 320 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7a01931086f2a99efa50eb5055bee659c22964f2c021acdcb88d17729a368260.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 261, + 320, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18fd6a59706527cf34e84f4ed1c87763d37f1bd38d15442ed10e4fe157afab09.jpg", + "image_caption": [ + "LR", + "LR (Zoomed)", + "Figure 6. Super-resolution results on common benchmarks, comparing MMSR with state-of-the-art techniques. Zoom in for detail." + ], + "image_footnote": [], + "bbox": [ + 102, + 332, + 210, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0f79b5b70da690e3b618e85314c809136e3ef9d713f429faa6a42d2faf4d66c.jpg", + "image_caption": [ + "Caption", + "Segmentation" + ], + "image_footnote": [], + "bbox": [ + 218, + 332, + 325, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0232628f8066fc3e8940a5bd13aebe43ae8e8f74f9fb7f70d2457af568c7fef0.jpg", + "image_caption": [ + "Depth" + ], + "image_footnote": [], + "bbox": [ + 336, + 260, + 444, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1305a8bd471d9adc23ac4117102a82cc0c1d2fe5045a525f04806ac30952251c.jpg", + "image_caption": [ + "Edge" + ], + "image_footnote": [], + "bbox": [ + 336, + 332, + 444, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/26fa23a053fe8ce905d48a7ea7472a7491062891c556dadeb1991ec48acc6ca2.jpg", + "image_caption": [ + "PASD" + ], + "image_footnote": [], + "bbox": [ + 455, + 258, + 563, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0de75abeba3895053529084962b17030e1fd04042b939bdc0f8b0f449392b2f0.jpg", + "image_caption": [ + "PASD (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 460, + 332, + 566, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8550f03dcfca4951c439bb8d26ab1eb4188baf0db4e442fb3b6fe018bcb430e.jpg", + "image_caption": [ + "SeeSR" + ], + "image_footnote": [], + "bbox": [ + 573, + 260, + 679, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc5b05e6cb23d8f6e80e9f93a594fffbf55c5c41182f3cc5f73a3d0b1d67ed89.jpg", + "image_caption": [ + "SeeSR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 578, + 332, + 683, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d7a388792b499d21119f39ab174717ae6f7606991b8d5f0b677236e6f9ef594.jpg", + "image_caption": [ + "MMSR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 689, + 258, + 797, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f0fb3998679a2fcdaf15d33903d67b7a2f6555ad7d3a71100d4cdf76feb2ddc.jpg", + "image_caption": [ + "MMSR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 694, + 332, + 800, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/248b4414d2e211f23b8c182caebea5783d57b9ec4f14bc34cb59cf96cdd34a30.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 805, + 260, + 913, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d22995bef713bb026ddaec73958546b868c45cd7f738edb0f7e789b982d62076.jpg", + "image_caption": [ + "HR (Zoomed)" + ], + "image_footnote": [], + "bbox": [ + 810, + 332, + 916, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Details. We use reference-based metrics like LPIPS [75] and DISTS [16], and non-reference-based metrics like NIQE [42], MANIQA [67], MUSIQ [27], and CLIPIQA [58], as these have been shown to align well with humans' aesthetic preferences. Our compared baselines include recent diffusion-based super-resolution methods [33, 53, 59, 65, 68, 72] and the representative methods [8, 31, 32, 63, 73]. We collected their results either by running their official code or from published results. In a few cases where the multimodal prediction failed (two flat images and one pencil drawing from DRealSR), we manually replaced the prediction with $m_{\\emptyset}$ (see supplemental material).", + "bbox": [ + 89, + 438, + 483, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 633, + 284, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the quantitative result comparisons of ours with the other baselines, compared on the synthetic benchmark and real-world super-resolution benchmarks. Several facts are worth noting: (1) Our method achieves the best LPIPS, DISTS, NIQE, and FID scores on the DIV2K-Val and RealSR dataset, significantly outperforming previous state-of-the-art methods. This superiority demonstrates that our method can generate more perceptually identical details from the guidance of multimodal context. (2) Our method also achieves the best performance in the non-reference visual quality metrics including NIQE, MANIQA, CLIPIQA, and MUSIQ scores. This advantage demonstrates our method can fully utilize the photorealistic priors encapsulated in the pretrained diffusion model. Overall, these observations fully demonstrate our advantages in using multimodal guidance compared with the past single-modality methods.", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effects of Multimodal CFG Guidance. Table 3 demonstrates the effects of our proposed multimodal CFG on the 1MP DIV2K-Val benchmark, which replaces the negative score guided by empty language token $\\epsilon (\\mathbf{z}_t,c,\\mathrm{neg})$ , shortened as $cfg$ , with multimodal guided negative score $(\\epsilon (\\mathbf{z}_t,c,m,\\mathrm{neg}))$ , shortened as $m$ -cfg. In order to exclude the influence of architectural differences, we also compare our method with the empty multimodal token $\\epsilon (\\mathbf{z}_t,c,m_{\\emptyset},\\mathrm{neg})$ , shortened as $m_{\\emptyset}$ -cfg. Note that the positive score for all three compared methods is the same $\\epsilon (\\mathbf{z}_t,c,m,\\mathrm{pos})$ . The results show that our method mitigates the degradation reflected in the LPIPS and NIQE scores when using guidance rates of 10 and 14 in both CFG and $m_{\\emptyset}$ -cfg, while maintaining comparable performance. The visual results in Table 3 further demonstrate our superiority, where ours suppresses the color change and incorrect texture of high CFG guidance rate, leading to higher CLIPIQA score.", + "bbox": [ + 511, + 438, + 906, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effects of Latent Connector. To demonstrate the effectiveness of the proposed multimodal latent connector, we conducted an ablation study by training a new multimodal super-resolution model without the latent connector, which directly uses the longer multimodal token sequence as input. The quantitative performance is shown in Table 2. Our $w$ . MMLC outperforms the $w/o$ . MMLC in all reference and non-reference metrics. The visual results in Figure 7 show that the model without the MMLC is more likely to exacerbate hallucination. Its result shows bokeh on the branches but a clear trunk, even though the depth input clearly shows they are at the same depth.", + "bbox": [ + 511, + 700, + 908, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Contributions of Each Modality. Figure 8 shows the con", + "bbox": [ + 513, + 885, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23146", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/284306d0f5bfd49cc475d0a36a18f0969eba73bf80c46cbc0f059ae65c492264.jpg", + "table_caption": [ + "DIV2K-Val-3k 512 × 512" + ], + "table_footnote": [], + "table_body": "
MethodsPSNRSSIMLPIPS ↓DISTS ↓NIQE ↓FID ↓MUSIQCLIPQA
BSRGAN21.870.55390.41360.27374.761564.2859.110.5183
R-ESRGAN21.940.57360.38680.26014.920953.4658.640.5424
LDL21.520.56900.39950.26885.024958.9457.900.5313
DASR21.720.55360.42660.26884.859667.2254.220.5241
FeMASR20.850.51630.39730.24284.572653.7058.100.5597
LDM21.260.52390.41540.25006.466741.9356.520.5695
StableSR20.840.48870.40550.25424.655136.5762.950.6486
ResShift21.750.54220.42840.26066.973155.7758.230.5948
PASD20.770.49580.44100.25384.832840.7766.850.6799
DiffBIR20.940.49380.42700.24714.721140.4265.230.6664
SeeSR21.190.53860.38430.22574.927531.9368.330.6946
MMSR21.740.56930.37070.20714.253229.3570.060.7164
", + "bbox": [ + 93, + 102, + 478, + 276 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43ae4967a5a146122d817c94c19509b6a24ef641a96be37267bf2bfafd94efb9.jpg", + "table_caption": [ + "DIV2K-Val-100 1024 × 1024" + ], + "table_footnote": [], + "table_body": "
MethodsPSNRSSIMLPIPS ↓DISTS ↓NIQE ↓MANIQAMUSIQCLIPiQA
R-ESRGAN21.770.58130.36240.19903.65730.404647.540.5358
StableSR20.070.39470.50970.24273.62600.411365.390.6938
PASD21.270.53690.34730.17533.63210.470869.690.6914
SUPIR20.650.53500.38490.18143.64580.405165.880.5697
SeeSR21.310.55780.32730.16204.02150.543969.790.6941
MMSR21.870.55650.28100.14923.42430.488572.310.7294
", + "bbox": [ + 93, + 295, + 478, + 388 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9e7647af8b83debf2660cb6b847a406cc761f9caa902c7cab7994bf21cb98bdf.jpg", + "table_caption": [ + "RealSR $512\\times 512$" + ], + "table_footnote": [], + "table_body": "
MethodsPSNRSSIMLPIPS ↓LIQENIMAMANIQAMUSIQCLIPiQA
R-ESRGAN25.690.76160.27273.35744.65480.548760.180.4449
StableSR24.700.70850.30183.61064.81500.622165.780.6178
PASD24.290.66300.34353.57494.85540.649368.690.6590
SUPIR22.970.62980.37503.56824.57570.574561.490.6434
SeeSR25.180.72160.30094.13604.91930.644269.770.6612
MMSR24.830.70030.29524.34685.10940.657871.330.6717
", + "bbox": [ + 94, + 409, + 478, + 503 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fc5235a0103e9ac2eda9a46859369918f64ddb6f0007d175a344d43d80c0c2f2.jpg", + "table_caption": [ + "DrealSR $512\\times 512$" + ], + "table_footnote": [], + "table_body": "
MethodsPSNRSSIMLPIPS ↓LIQENIMAMANIQAMUSIQCLIPiQA
R-ESRGAN28.640.80530.28472.92554.32580.490754.180.4422
StableSR26.710.72240.32843.24254.48610.559458.510.6357
PASD27.000.70840.39313.59084.66180.585064.810.6773
SUPIR24.610.61230.42943.47104.38150.538157.320.6758
SeeSR26.750.74050.31744.12704.69420.605265.090.6908
MMSR27.280.74560.32494.50235.05580.630168.930.6999
", + "bbox": [ + 93, + 523, + 478, + 619 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e5763c9b9a44445386cf61255538423942c5f15e1ebf90ba55804c2630eaf88d.jpg", + "table_caption": [ + "Table 1. Severely degraded, low-resolution (LR) images can produce inaccurate multi-modal information, manifesting as distorted edges, misidentified objects, and other artifacts." + ], + "table_footnote": [], + "table_body": "
MUSIQNIQE ↓DISTS ↓LPIPS ↓Throughput
w/o. MMLC69.693.48450.17810.39293.48 img/s
w. MMLC72.313.42430.14920.28103.32 img/s
", + "bbox": [ + 96, + 670, + 478, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tributions of each modality by masking out input modalities during testing. Benefiting from the improved flexibility of using multimodal token $m_{\\emptyset}$ , discussed in Sec. 3.1, our method is robust in scenarios with only fewer modalities input during testing. We would like to note several observations: (1) Our default multimodal setting achieves the best trade-off between the non-reference metric MUSIQ and the reference-based metric DISTS. (2) Depth information primarily enhances perceptual quality (MUSIQ), while other modalities", + "bbox": [ + 88, + 763, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5b4480ca13d2918bc60ce4201738b49b7081d15f8dfbc9233b979a4dafb0ac13.jpg", + "table_caption": [ + "Table 2. Ablation of the impact of the Multimodal Latent Connector module (MMLC) on DIV2K-Val-100 1024p." + ], + "table_footnote": [], + "table_body": "
guidance21014
cfg0.32390.44910.5064
m0cfg0.28150.48030.5493
mcfg0.28100.34710.3772
", + "bbox": [ + 517, + 90, + 702, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/471c9862b8304157c561b833c29b15e6089a7e5a5c408dedcab0796b0df51e6b.jpg", + "image_caption": [ + "Segment, Depth, LR, and Edge Input", + "(c) Visual comaprisons" + ], + "image_footnote": [], + "bbox": [ + 519, + 162, + 746, + 294 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/29503b5997778ecba991ea2ce959b8856f9888661a7e6dedf8b2ef0425773d04.jpg", + "table_caption": [ + "(a) LPIPS score" + ], + "table_footnote": [], + "table_body": "
guidance21014
cfg3.5774.61795.1886
m0cfg3.62615.10815.9175
mcfg3.46793.74193.9815
", + "bbox": [ + 715, + 90, + 898, + 148 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) NIQE score", + "bbox": [ + 767, + 152, + 841, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bccfcd896c218f8402f694cedaf30180c9c27eca1780b17e5b6e09f80050ba58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 162, + 898, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2fb16ce560d277d9ac9b41456fffe629f39327c38a80ec517e4f01c701203c6f.jpg", + "image_caption": [ + "$cfg = 2.0$ CIPIQA $= 0.4481$", + "$cfg = 14.0$ CLIPIQA=0.7091" + ], + "image_footnote": [], + "bbox": [ + 753, + 198, + 897, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/708285030746072e83429d216889c81bdb5b81d53200da6954bf84f9cbc778cd.jpg", + "image_caption": [ + "$m\\text{-} cfg = 2.0$ CLIPIQA $= 0.4122$" + ], + "image_footnote": [], + "bbox": [ + 753, + 233, + 897, + 260 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3e3ac5369266df04fbfa8f1c4a2e899b27cbd942094bacd84651fda47de91f7.jpg", + "image_caption": [ + "$m\\text{-} cfg = 14.0$ CIPIQA $= 0.7352$" + ], + "image_footnote": [], + "bbox": [ + 753, + 268, + 897, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b3fd4014db8a5eed536a6d3664f059a3b4d5b8b072239c3d026eebe95122258.jpg", + "image_caption": [ + "Table 3. Ablation of guidance on DIV2K-Val-100 1024p. Our multimodal CFG ( $m$ -cfg) mitigates artifacts often present when using high guidance rates. This leads to an improved balance between visual fidelity and preservation of key identifying features.", + "LR Input" + ], + "image_footnote": [], + "bbox": [ + 519, + 375, + 705, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df974042424819b6ad44cf618c72580032d8eb49c9ef08783297b2ec2dcd6bbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 375, + 900, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f0106977ab5347565c3243ae29405907dea68c6754fc3e9609da218cc954142.jpg", + "image_caption": [ + "w/o. MMLC Result" + ], + "image_footnote": [], + "bbox": [ + 519, + 446, + 707, + 503 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/309df6ac578aebc7bfeb3d9dec1b1a33e32eed7d1583cc359e332b6824252d03.jpg", + "image_caption": [ + "Depth Input", + "w. MMLC Result" + ], + "image_footnote": [], + "bbox": [ + 710, + 445, + 898, + 503 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c8ec3f47a4344137af101359c05d1c3e92b3928e330b4235acd7a6487d4878f.jpg", + "image_caption": [ + "Figure 7. Visual results when (not) using the MMLC module.", + "Figure 8. Ablation of modalities on DIV2K-Val-100 1024p. Incorporating different modalities enhances various aspects of the super-resolution results compared to the text-guided baseline." + ], + "image_footnote": [], + "bbox": [ + 521, + 539, + 736, + 608 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/888c95ca94a1bed5fd4b8d9bfd6c82278c314b5ff5e46703b21d5ca3a658b13b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 540, + 898, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "contribute more significantly to preserving identity (DISTS). These results highlight the diverse contributions of each modality and emphasize the advantage of our multimodal approach in effectively combining their strengths for optimal super-resolution performance.", + "bbox": [ + 511, + 669, + 906, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 755, + 694, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figures 5 and 6 provide a visual comparison of our method against state-of-the-art approaches, namely SeeSR [65], PASD [68], and SUPIR [71]. Our method demonstrates several key advantages. Enhanced Realism: In the first example of Figure 6, our method produces fewer artifacts and remains more faithful to the high-resolution image compared to other methods. This highlights our ability to maintain realism without introducing spurious details. Robustness to", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23147", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49fae6f99adc81fda8523d7826166484f2a755733a2bff7320b497d1b41a8d87.jpg", + "image_caption": [ + "(a). Results of PASD" + ], + "image_footnote": [], + "bbox": [ + 102, + 114, + 215, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15eae51b990b4a0e7675878d8ff7068bccd05013eec4bdc650e1631fdd4ce5cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 195, + 215, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7096bc195c9c7475167467d3aeac064e261c4d18d7e9020afcfe39a81eb67cc.jpg", + "image_caption": [ + "Figure 9. Our method allows for fine-grained control over super-resolution results by adjusting the influence of each input modality. For example, reducing the edge temperature enhances edge sharpness (first row). Lowering the segmentation temperature emphasizes distinct features, such as the star pattern on the flag (second row). Decreasing the depth temperature accentuates depth-of-field effects, like the bokeh between the foreground and background (third row). In contrast, PASD [68] exhibits limited control over such fine-grained details." + ], + "image_footnote": [], + "bbox": [ + 102, + 277, + 215, + 353 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1ebf466d0fd63f3a8eba5df203f90bbb9347e8277e28f420eef76d2169eb899.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 114, + 341, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83f73c6f98f89b437c7cd845654d1b1d76a1817e26aacf7b5dca378364529fa0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 195, + 341, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a156ad46d9cc4447636837d62f43b894a65c2f72ec2fd21182ac6f619aa9631.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 277, + 341, + 352 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03ebe657ad6d0f362bbbd98d85061116b860a8606687cd6cd31f8622e9c6df9a.jpg", + "image_caption": [ + "(b). Results of MMSR under different modality-specific controls" + ], + "image_footnote": [], + "bbox": [ + 357, + 117, + 893, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f23dbfb7c3bf62e05058b955e276d1e727aa1abddf451f78a4671f1bea85c32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 199, + 893, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3cd652ef8a05a0e179a7f609e7a39646cc403a280653dcd404923b8d1bb84cf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 282, + 893, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Challenging Conditions: The second example in Figure 6, a long-range shot affected by noise and turbulence, shows our method's ability to generate clear human details where other methods struggle. This is attributed to the effective use of semantic segmentation for accurate human localization, further enriched by the textual caption. These examples illustrate the effectiveness of our multimodal approach in generating high-quality images across diverse scenarios. See supplemental material for additional results and analysis.", + "bbox": [ + 89, + 450, + 483, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Controllability Comparisons", + "text_level": 1, + "bbox": [ + 89, + 603, + 346, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While recent works leverage text prompts for controlling super-resolution [65, 68, 71], these often lack fine-grained control and can yield inconsistent results. Our multimodal approach introduces modality-specific temperature weights to scale attention scores, enabling precise manipulation of SR outputs by amplifying or diminishing the influence of each modality (e.g., depth, segmentation). Figure 9 contrasts our approach with PASD [68], which relies solely on text prompts. Specifically, decreasing the temperature of the edge modality changes the richness of details in our result, while changing PASD's prompt with \"more edge\" doesn't change their result. Decreasing the temperature of semantic segmentation map and depth map also lead to manageable changes such as more visible star pattern and stronger bokeh. In contrast, directly add corresponding prompt in PASD's prompt barely changes its result. Moreover, our approach exhibits smooth transitions in image characteristics as temperatures are adjusted, providing interpretable control and insights into", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the role of each modality.", + "bbox": [ + 511, + 450, + 683, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 479, + 632, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduces a novel diffusion-based framework for image super-resolution that seamlessly integrates diverse modalities—including text descriptions, depth maps, edges, and segmentation maps—within a single, unified model. By leveraging pretrained text-to-image models, our method achieves enhanced realism and accurate reconstruction, outperforming existing text-guided super-resolution methods both qualitatively and quantitatively. Furthermore, a learnable multimodal token and modality-specific contribution controls provide fine-grained control over the super-resolution process, enabling adjustable perception-distortion tradeoffs and robust performance even with imperfect or missing modalities in most cases.", + "bbox": [ + 511, + 505, + 906, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. While adding multimodal information significantly enhances SR performance, it introduces computational overhead. For instance, using Gemini Flash for image captioning results in a throughput of 0.34 images per second, which is slower than depth at $1.99\\mathrm{img / s}$ , semantic segmentation at $2.09\\mathrm{img / s}$ , and DDIM sampling at $0.54\\mathrm{img / s}$ . However, cross-modal predictions can be parallelized, and our method achieves speeds comparable to other text-driven SR models [68, 71]. Future work will explore optimizing the vision-language component for faster inference and investigate more robust modules for extracting modality-specific information, potentially enhancing performance even with noisy or incomplete inputs.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23148", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Vishal M. Patel was supported by NSF CAREER award 2045489. We are grateful to Keren Ye and Mo Zhou for their valuable feedback. We also extend our gratitude to Shlomi Fruchter, Kevin Murphy, Mohammad Babaeizadeh, and Han Zhang for their instrumental contributions in facilitating the initial implementation of the latent diffusion models.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2", + "[2] Roman Bachmann, Oğuzhan Fatih Kar, David Mizrahi, Ali Garjani, Mingfei Gao, David Griffiths, Jiaming Hu, Afshin Dehghan, and Amir Zamir. 4M-21: An any-to-any vision model for tens of tasks and modalities. NeurIPS, 2024. 2, 3", + "[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, pages 22861-22872, 2024. 3", + "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2023.1, 2", + "[5] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In CVPR, pages 6228-6237, 2018. 1", + "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 2, 5", + "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, pages 14455-14465, 2024. 1, 2", + "[8] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Real-world blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 6", + "[9] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25857-25867, 2024. 2", + "[10] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1", + "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 5" + ], + "bbox": [ + 93, + 246, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. arXiv preprint arXiv:2310.01110, 2023. 1", + "[13] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. In *Forty-first International Conference on Machine Learning*, 2024. 2", + "[14] Maurizio Delbracio and Peyman Milanfar. Inversion by direct iteration: An alternative to denoising diffusion for image restoration. Transactions on Machine Learning Research, 2023. Featured Certification. 1", + "[15] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, 2021. 4", + "[16] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In Proceedings of the 29th ACM International Conference on multimedia, pages 2483-2491, 2021. 6", + "[17] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 2015. 1", + "[18] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3", + "[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2", + "[20] Kanchana Vaishnavi Gandikota and Paramanand Chandramouli. Text-guided explorable image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25900-25911, 2024. 2", + "[21] Jiatao Gu, Ying Shen, Shuangfei Zhai, Yizhe Zhang, Navdeep Jaitly, and Joshua M Susskind. Kaleido diffusion: Improving conditional diffusion models with autoregressive latent modeling. Advances in neural information processing systems, 2024. 2", + "[22] Jiatao Gu, Yuyang Wang, Yizhe Zhang, Qihang Zhang, Dinghuai Zhang, Navdeep Jaitly, Josh Susskind, and Shuangfei Zhai. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024. 2", + "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 4", + "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3", + "[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23149", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 4", + "[26] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's \"up\" with vision-language models? investigating their struggle with spatial reasoning. In EMNLP, pages 9161–9175, 2023. 1", + "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6", + "[28] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning, pages 3744-3753. PMLR, 2019. 4", + "[29] Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image super-resolution. In Proceedings of the European conference on computer vision (ECCV), pages 517-532, 2018. 1", + "[30] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2", + "[31] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 6", + "[32] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision. Springer, 2022. 6", + "[33] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diffbir: Towards blind image restoration with generative diffusion prior. ArXiv preprint, 2023. 2, 6", + "[34] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 1, 2", + "[35] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 1", + "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2", + "[37] Kangfu Mei, Vishal M Patel, and Rui Huang. Deep semantic statistics matching (d2sm) denoising network. In European Conference on Computer Vision, pages 384-400. Springer, 2022. 1", + "[38] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image generation. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 9048-9058, 2024. 2", + "[39] Kangfu Mei, Luis Figueroa, Zhe Lin, Zhihong Ding, Scott Cohen, and Vishal M Patel. Latent feature-guided diffusion models for shadow removal. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4313-4322, 2024. 2", + "[40] Kangfu Mei, Zhengzhong Tu, Mauricio Delbracio, Hossein Talebi, Vishal M Patel, and Peyman Milanfar. Bigger is not always better: Scaling properties of latent diffusion models. arXiv preprint arXiv:2404.01367, 2024. 2", + "[41] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022.", + "[42] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6", + "[43] David Mizrahi, Roman Bachmann, Oğuzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4M: Massively multimodal masked modeling. NeurIPS, 2023. 2, 3", + "[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3", + "[45] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4", + "[46] Mehdi Noroozi, Isma Hadji, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. You only need one step: Fast super-resolution with stable diffusion via scale distillation. In European Conference on Computer Vision, 2024. 2", + "[47] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 4", + "[48] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[49] Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi. Spire: Semantic prompt-driven image restoration. In European Conference on Computer Vision, 2024. 2", + "[50] Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou. Xpsr: Cross-modal priors for diffusion-based image super-resolution. In European Conference on Computer Vision, pages 285-303. Springer, 2024. 1, 2", + "[51] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23150", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1", + "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 4", + "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 2, 5, 6", + "[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[55] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 4", + "[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 5", + "[57] Li-Yuan Tsao, Yi-Chen Lo, Chia-Che Chang, Hao-Wei Chen, Roy Tseng, Chien Feng, and Chun-Yi Lee. Boosting flow-based generative super-resolution models via learned prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26005–26015, 2024. 2", + "[58] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6", + "[59] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 2, 4, 6", + "[60] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 4", + "[61] Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, and Shanshan Zhao. Caption anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677, 2023. 2", + "[62] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 1", + "[63] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Realesrgan: Training real-world blind super-resolution with pure synthetic data supplementary material. Computer Vision Foundation open access, 2022. 5, 6", + "[64] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2", + "[65] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 1, 2, 4, 6, 7, 8", + "[66] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 5", + "[67] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6", + "[68] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, 2024. 1, 2, 6, 7, 8", + "[69] Zhixiong Yang, Jingyuan Xia, Shengxi Li, Xinghua Huang, Shuanghui Zhang, Zhen Liu, Yaowen Fu, and Yongxiang Liu. A dynamic kernel prior model for unsupervised blind image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26046-26056, 2024. 2", + "[70] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3", + "[71] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 1, 4, 5, 7, 8", + "[72] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super-resolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 2, 6", + "[73] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 6", + "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3", + "[75] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23151", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. In European Conference on Computer Vision, 2024. 2", + "[77] Zheyuan Zhang, Fengyuan Hu, Jayjun Lee, Freda Shi, Parisa Kordjamshidi, Joyce Chai, and Ziqiao Ma. Do vision-language models represent space and how? evaluating spatial frame of reference under ambiguities, 2024. 1" + ], + "bbox": [ + 91, + 92, + 485, + 191 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "23152", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_model.json b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_model.json new file mode 100644 index 0000000000000000000000000000000000000000..47268e11eebe0c522534a5fc81ecd892b6981257 --- /dev/null +++ b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_model.json @@ -0,0 +1,3645 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.116, + 0.892, + 0.139 + ], + "angle": 0, + "content": "The Power of Context: How Multimodality Improves Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.165, + 0.726, + 0.238 + ], + "angle": 0, + "content": "Kangfu Mei\\*1,2, Hossein Talebi', Mojtaba Ardakani', Vishal M. Patel?, Peyman Milanfar', Mauricio Delbracio' 1 Google, 2 Johns Hopkins University Project Page: https://mmsr.kfmei.com/" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.244, + 0.924, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 1. Our Multimodal Super-Resolution (MMSR) method leverages the rich context of multimodal guidance, including image captions, depth maps, semantic segmentation maps, and edges inferred from LR. MMSR surpasses state-of-the-art methods by producing more realistic results and suppressing artifacts that, while plausible, are inconsistent with the information present in the LR input." + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.472, + 0.317, + 0.485 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.493, + 0.454, + 0.813 + ], + "angle": 0, + "content": "Single-image super-resolution (SISR) remains challenging due to the inherent difficulty of recovering fine-grained details and preserving perceptual quality from low-resolution inputs. Existing methods often rely on limited image priors, leading to suboptimal results. We propose a novel approach that leverages the rich contextual information available in multiple modalities – including depth, segmentation, edges, and text prompts—to learn a powerful generative prior for SISR within a diffusion model framework. We introduce a flexible network architecture that effectively fuses multimodal information, accommodating an arbitrary number of input modalities without requiring significant modifications to the diffusion process. Crucially, we mitigate hallucinations, often introduced by text prompts, by using spatial information from other modalities to guide regional text-based conditioning. Each modality's guidance strength can also be controlled independently, allowing steering outputs toward different directions, such as increasing bokeh through depth or adjusting object prominence via segmentation. Extensive experiments demonstrate that our model surpasses state-of-the-art generative SISR methods, achieving superior visual quality and fidelity." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.824, + 0.222, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.486, + 0.88 + ], + "angle": 0, + "content": "Single image super-resolution (SISR) aims to generate high-resolution images from low-resolution inputs while preserving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.909, + 0.773 + ], + "angle": 0, + "content": "ing semantic identity and texture details. While it is not essential to treat SISR as a regression problem, past methods [17, 29, 37, 62] typically use a deep neural network to learn a direct mapping from low-resolution images to high-resolution images. Even though these regression-based methods achieve good scores on paired metrics like PSNR and SSIM, they have failed to produce results with high quality comparable to natural images, a task at which recent generative models have excelled [5, 12, 14]. The advent of powerful generative models, such as autoregressive models [10, 51] and diffusion models [47, 53, 54], has revolutionized image generation tasks, including text-to-image synthesis. This has inspired recent efforts to leverage these pre-trained generative models for downstream tasks like SISR [6, 41, 53]. For instance, very recent works [50, 71] achieve super-resolution by leveraging emerging vision-language models (e.g., Gemini [56], LLaVA [36], ChatGPT-4 [1]) and pretrained text-to-image models to first generate captions from low-resolution images and then use these captions as prompts to generate high-resolution images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "While providing rich textual descriptions can significantly enhance the quality of generated images [4, 19, 34], relying solely on text for SISR poses challenges. Recent works [7, 26, 35, 77] have shown that text prompts cannot represent spatial relationships. This implies that textual information, such as texture descriptions, can only be applied to the whole image. Figure 1 provides a representative example. Previous text-based super-resolution methods [65, 68] use a 'lion'" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.39, + 0.901 + ], + "angle": 0, + "content": "*This work was done during an internship at Google." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23141" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "caption, which results in a furry tongue. However, while lions have fur, their tongues do not grow hair. Can we leverage spatial cues from depth and segmentation data to improve the learned prior and enhance the quality of SISR?" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.153, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Fortunately, we can directly extract additional modalities from the low-resolution image by using various pretrained cross-modal prediction models [11, 66]. In this paper, we introduce a new diffusion model architecture dedicated to multimodal super-resolution, which is conceptually illustrated in Figure 2. By integrating multiple modalities into a single diffusion model, our method overcomes the challenges of recovering fine-grained details and preserving perceptual quality. Specifically, we propose conditioning diffusion models on modalities including text captions, semantic segmentation maps, depth maps, and edges to implicitly align text captions for correctly prompting different regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.456 + ], + "angle": 0, + "content": "We demonstrate the proposed multimodal diffusion model on the SISR task. Our effective multimodal architecture achieves better realism in SISR results than the best text-driven method and largely eliminates hallucinations that do not match the input. Moreover, we find that the multimodal SISR enables a new controlling feature, where we can explicitly adjust the weights of each modality to steer the generated results in different directions." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.457, + 0.414, + 0.472 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.477, + 0.483, + 0.538 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of token-wise encoding for seamlessly injecting multiple modalities into pretrained text-to-image diffusion models without architectural modifications or significant model size overhead." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.538, + 0.484, + 0.597 + ], + "angle": 0, + "content": "- We propose a novel multimodal latent connector that efficiently fuses information from different modalities, maintaining linear time complexity with respect to the number of modalities." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.598, + 0.483, + 0.643 + ], + "angle": 0, + "content": "- We introduce a new multimodal classifier-free guidance technique that enhances realism at higher guidance rates while mitigating excessive hallucinations and fake details." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.643, + 0.483, + 0.689 + ], + "angle": 0, + "content": "- Our method enables adjustment of the influence of each modality, allowing for fine-grained manipulation of SISR results while preserving realism and quality." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.477, + 0.484, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.709, + 0.233, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Generative Prior Powered Super-resolution. Recent advances in image restoration leverage the power of foundational models to enhance the quality of degraded images [9, 13, 20, 33, 38, 39, 46, 50, 57, 64, 69, 72, 76]. This trend is fueled by the capacity of pretrained generative models to capture natural image statistics and transfer this knowledge to the super-resolution task, enabling photorealistic image generation. Early works in this domain include LDM-SR [53] and StableSR [59] for single image superresolution. Beyond this, methods like InstructPix2Pix [6] utilize instructions for image editing, while ControlNet [74]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.905, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.195, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Language models struggle to accurately represent spatial information, leading to coarse and imprecise image superresolution. To overcome this limitation, we incorporate additional spatial modalities like depth maps and semantic segmentation maps. These modalities provide detailed spatial context, allowing our model to implicitly align language descriptions with individual pixels through a transformer network. This enriched understanding of the image significantly enhances the realism of our super-resolution results and minimizes distortion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.404 + ], + "angle": 0, + "content": "and IP-Adapter [70] facilitate cross-modality image translation. These approaches highlight a common observation: the quality of results in downstream tasks is strongly correlated with the quality of the pretrained models [40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.909, + 0.572 + ], + "angle": 0, + "content": "Recent works focus on leveraging powerful text-to-image diffusion models, employing text prompts to fully exploit the learned prior. PASD [68] uses image content descriptions; SPIRE [49] and PromptIP [48] use degradation descriptions; and SeeSR [65] employs a combination of context and degradation descriptions. Compared to earlier methods that directly fine-tuned diffusion models on super-resolution data, these text-prompt-driven approaches not only achieve superior realism but also enable multi-faceted outputs by conditioning on different prompts [20]. This capability enriches the typically single-output super-resolution task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.909, + 0.68 + ], + "angle": 0, + "content": "While text prompts offer advantages, they can be ambiguous in representing spatial relationships [4, 7, 19, 34]. Our proposed MMSR framework addresses this by integrating multimodal inputs within a novel architecture. We also demonstrate native super-resolution effects control with this new architecture, enabling control of both overall realism and the effect of each modality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Vision-language Understanding and Generation. Recent large vision-language models (e.g., Gemini [56], LLaVA [36], GPT-4o [1]) excel in tasks like image captioning [61], but translating visual information into diverse modalities for image generation remains challenging. While models like 4M [2, 43] effectively extract high-level visual information (depth, normals, semantics, etc.), leveraging these for generation is a promising direction. Recent exploration includes GLIGEN [30] generate images from bounding boxes, and ControlNet [74] generate images from depth and other modalities. More recent works [21, 22] explore using discrete hidden-feature tokens to guide generation. In this paper, we introduce a new approach for multimodal control with discrete vision tokens, effectively integrating depth maps," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23142" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.904, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 3. This diagram illustrates our multimodal super-resolution pipeline. Starting with a low-resolution (LR) image, we extract modalities like depth and semantic segmentation maps. These modalities are encoded into tokens and transformed into multimodal latent tokens \\((m)\\). Our diffusion model uses these tokens and the LR input to generate a high-resolution (SR) output. A multimodal classifier-free guidance \\((m\\)-cfg) refines the SR image for enhanced quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.413 + ], + "angle": 0, + "content": "semantic segmentation maps, and text prompts for superior performance. Experiments on real-world super-resolution demonstrate the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.43, + 0.182, + 0.445 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.457, + 0.483, + 0.623 + ], + "angle": 0, + "content": "Single-image super-resolution aims to recover a high-resolution image \\(\\mathbf{x}\\) from its low-resolution counterpart \\(\\mathbf{x}_{LR}\\). This is an ill-posed problem, often leading to generative models that produce \"hallucinated\" details – plausible yet inconsistent with the input. To mitigate this, we introduce auxiliary information, such as depth \\((m_{\\mathrm{dep}})\\), semantic segmentation \\((m_{\\mathrm{seg}})\\), and edge maps \\((m_{\\mathrm{edg}})\\), collectively denoted as \\(m\\). By conditioning the generative process on both the low-resolution image and these auxiliary modalities, we propose a new distribution \\(p(\\mathbf{x}|\\mathbf{x}_{LR}, m)\\) with reduced uncertainty compared to the original distribution \\(p(\\mathbf{x}|\\mathbf{x}_{LR})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.7 + ], + "angle": 0, + "content": "This reduction in uncertainty can be understood through the lens of information theory. The auxiliary modality \\( m \\) provides additional information about the high-resolution image \\( \\mathbf{x} \\) that is not present in the low-resolution input \\( \\mathbf{x}_{LR} \\). Since (conditional) mutual information is non-negative:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.705, + 0.482, + 0.721 + ], + "angle": 0, + "content": "\\[\nI (\\mathbf {x}; m | \\mathbf {x} _ {L R}) = H (p (\\mathbf {x} | \\mathbf {x} _ {L R})) - H (p (\\mathbf {x} | \\mathbf {x} _ {L R}, m)), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.725, + 0.483, + 0.755 + ], + "angle": 0, + "content": "where \\( H(\\cdot) \\) denotes entropy. Consequently, the entropy of the conditional distribution with the auxiliary modality is:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.76, + 0.482, + 0.776 + ], + "angle": 0, + "content": "\\[\nH \\left(p \\left(\\mathbf {x} \\mid \\mathbf {x} _ {L R}\\right)\\right) \\geq H \\left(p \\left(\\mathbf {x} \\mid \\mathbf {x} _ {L R}, m\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The same motivation for reducing uncertainty is also shared by recent diffusion guidance works [23, 44]. While low-entropy sampling does not guarantee sharper images, it often leads to outputs with better visual quality compared to standard sampling. Motivated by this observation, we introduce a diffusion model [24] to learn the multimodal distribution \\( p(\\mathbf{x}|\\mathbf{x}_{LR},m) \\), effectively incorporating auxiliary information to enhance SISR quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Ideally, the auxiliary modalities should provide information complementary to the low-resolution input. In practice, the auxiliary modalities are derived from the high-resolution image during training, ensuring informative conditioning. While we use modalities derived from the low-resolution input during inference, we demonstrate that this still leads to superior performance, including higher-quality details and improved fidelity to the input." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.504, + 0.882, + 0.521 + ], + "angle": 0, + "content": "3.1. Unified Multimodal Diffusion Conditioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.786 + ], + "angle": 0, + "content": "We introduce a new diffusion network architecture, illustrated in Figure 3, for simultaneously conditioning on multiple modalities. Unlike recent methods like ControlNet [74] and IP-Adapter [70], which duplicate network components for each modality and incur significant computational overhead, our approach leverages a pretrained VQGAN image tokenizer [18]. This allows us to encode diverse modalities into a unified token representation for conditioning the diffusion model, without introducing additional model parameters or modifying the diffusion network itself. These tokens are concatenated with the text prompt embedding and used for cross-attention within the diffusion model. To efficiently process this long token sequence, we introduce a lightweight multimodal connector. This connector employs a dedicated architecture to achieve linear complexity for cross-attention, significantly reducing the computational burden. In what follows we present the implementation details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Token-wise Multimodal Encoding. While VQGAN [18] has proven effective for cross-modal encoding in image understanding and generation [2, 3, 43], its optimal application for image super-resolution remains an open question. Unlike tasks focused on understanding or generation, superresolution requires tokenization to not only capture semantic information but also preserve pixel-wise details crucial for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23143" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.098, + 0.195, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.098, + 0.292, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.098, + 0.39, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.098, + 0.491, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.23, + 0.485, + 0.258 + ], + "angle": 0, + "content": "Figure 4. Using discrete multimodal tokens leads to superior reconstruction of modalities compared to continuous tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.484, + 0.408 + ], + "angle": 0, + "content": "accurate reconstruction, such as spatial relationships within a depth map. Inspired by recent discussions on using discrete or continuous tokenization for autoregressive image generation [55], we investigate the impact of quantization on multimodal super-resolution. Specifically, we analyze how different quantization strategies affect the reconstruction quality of individual modalities and the final generated image. To this end, we utilize a VQGAN model pre-trained on a large-scale multimodal image dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.5 + ], + "angle": 0, + "content": "Figure 4 compares the reconstruction quality of multi-modal tokens before and after quantization. As the figure illustrates, discrete tokens (post-quantization) better preserve individual modality information, while continuous tokens introduce noticeable artifacts. Consequently, we employ discrete tokens for our super-resolution experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.622 + ], + "angle": 0, + "content": "Our multimodal encoding process utilizes the encoder and quantizer of the pre-trained VQGAN. Each \\(256 \\times 256\\) input modality is encoded into a \\(16 \\times 16\\) multimodal token sequence with a feature dimension of 256. These tokens are then quantized using a codebook of size 1024. To facilitate concatenation with text embeddings, we pad the feature dimension of the multimodal tokens to 1024, resulting in a \\((256 \\times 3 + 77) \\times 1024\\) multimodal token sequence condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Multimodal Latent Connector. While cross-attention provides a flexible mechanism for conditioning diffusion models on multimodal data, its quadratic complexity with respect to the number of condition tokens introduces a significant computational burden. To address this, we introduce the Multimodal Latent Connector (MMLC), inspired by recent advances in efficient attention architectures [25, 28, 60]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Figure 3, the MMLC employs a transformer architecture to efficiently process the multimodal token sequence. The transformer receives two inputs: a randomly initialized sequence of learnable latent tokens, and the multimodal input sequence. The output is a token sequence of the same length as the latent token sequence, which serves as conditioning for the diffusion model. Therefore, the diffusion model conditions on fixed-length latent tokens (128 in our experiments), which are significantly shorter than the original multimodal token sequence input. The MMLC distills the essential information from the longer multimodal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.166 + ], + "angle": 0, + "content": "token sequence into the shorter multimodal latent token sequence through cross-attention. Following cross-attention, several self-attention blocks further process the latent token sequence, allowing the model to fully integrate the distilled information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.909, + 0.365 + ], + "angle": 0, + "content": "This approach significantly reduces the computational cost of cross-attention in the diffusion model. Standard self-attention operates on the full multimodal token sequence \\((K, V \\in \\mathbb{R}^{M \\times D})\\), resulting in a time complexity of \\(\\mathcal{O}(M^2)\\) where \\(M\\) is the length of the sequence and \\(D\\) is the dimensionality of each token. In contrast, the MMLC uses a cross-attention between the latent token sequence (of size \\(N \\times D\\)) and the multimodal token sequence which reduces this to \\(\\mathcal{O}(MN)\\). Here \\(N\\) is the length of the latent sequence and \\(N \\ll M\\). This linear complexity with respect to the multimodal sequence length enables efficient processing of high-dimensional multimodal data, effectively capturing essential information for super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.909, + 0.644 + ], + "angle": 0, + "content": "Flexible Multimodal Input. To enhance flexibility and robustness, we enable our method to handle scenarios where certain input modalities are unreliable or unavailable. We adopt a learnable embedding approach inspired by DALL-E 2 [52]. Specifically, we introduce a special learnable token, \\( m_{\\emptyset} \\), optimized alongside the diffusion model and MMLC to represent the absence of a modality. During training, we independently randomly replace each modality with 256 \\( m_{\\emptyset} \\) tokens with a probability of 0.1. This encourages the model to learn robust representations that can effectively handle missing information. During inference, any missing modality is represented by a sequence of 256 \\( m_{\\emptyset} \\) tokens. This approach allows for flexible multimodal input, enabling the model to generate high-quality images even with limited or no auxiliary modalities. As demonstrated in our experiments, this strategy significantly improves performance when dealing with input that includes fewer modalities than were available during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.818, + 0.67 + ], + "angle": 0, + "content": "3.2. Multimodal Guidance and Control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.909, + 0.786 + ], + "angle": 0, + "content": "Building upon the success of guidance techniques [15, 23] in improving sample quality across various image generation tasks [45, 47, 52], recent diffusion-based SISR methods have begun incorporating prompt tuning to enhance superresolution results [59, 65, 71]. These methods improve the result by using negative prompts with Classifier-free Guidance (CFG) [23]. which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.79, + 0.905, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} (\\mathbf {z} _ {t}, c) = (1 + w) \\epsilon (\\mathbf {z} _ {t}, c, \\text {p o s}) - w \\epsilon (\\mathbf {z} _ {t}, c, \\text {n e g}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tilde{\\epsilon} (\\mathbf{z}_t,c)\\) represents the guided denoising process, \\(\\pmb {\\epsilon}(\\cdot)\\) denotes the diffusion model, \\(\\mathbf{z}_t\\) denotes the noisy image latent, \\(c = \\{\\mathbf{x}_{\\mathrm{LR}},t,\\dots \\}\\) denotes the conditioning inputs (including the low-resolution image \\(\\mathbf{x}_{\\mathrm{LR}}\\), timestep \\(t\\), and other parameters), \\(w\\) is the guidance scale, and pos and neg are the positive and negative prompts. While increasing \\(w\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23144" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.098, + 0.21, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.169, + 0.163, + 0.177 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.178, + 0.208, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.239, + 0.169, + 0.248 + ], + "angle": 0, + "content": "Depth" + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.098, + 0.287, + 0.107 + ], + "angle": 0, + "content": "Inputs" + }, + { + "type": "image_footnote", + "bbox": [ + 0.227, + 0.109, + 0.318, + 0.166 + ], + "angle": 0, + "content": "The image shows the words \"CIPAL PARK\" prominently displayed on a metal structure. Below the main sign are several smaller screens displaying baseball ..." + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.169, + 0.288, + 0.177 + ], + "angle": 0, + "content": "Caption" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.178, + 0.323, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.24, + 0.283, + 0.248 + ], + "angle": 0, + "content": "Edge" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.107, + 0.443, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.169, + 0.403, + 0.177 + ], + "angle": 0, + "content": "Patch 1" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.178, + 0.443, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.239, + 0.406, + 0.248 + ], + "angle": 0, + "content": "Patch 2" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.107, + 0.56, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.169, + 0.541, + 0.177 + ], + "angle": 0, + "content": "PASD (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.178, + 0.56, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.239, + 0.54, + 0.248 + ], + "angle": 0, + "content": "PASD (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.107, + 0.673, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.169, + 0.654, + 0.177 + ], + "angle": 0, + "content": "SeeSR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.178, + 0.673, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.239, + 0.655, + 0.248 + ], + "angle": 0, + "content": "SeeSR (Zoomed)" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.098, + 0.693, + 0.106 + ], + "angle": 0, + "content": "uts" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.107, + 0.787, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.169, + 0.768, + 0.177 + ], + "angle": 0, + "content": "SUPIR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.178, + 0.787, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.239, + 0.769, + 0.248 + ], + "angle": 0, + "content": "SUPIR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.107, + 0.899, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.169, + 0.883, + 0.177 + ], + "angle": 0, + "content": "MMSR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.178, + 0.9, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.239, + 0.884, + 0.248 + ], + "angle": 0, + "content": "MMSR (Zoomed)" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.254, + 0.904, + 0.269 + ], + "angle": 0, + "content": "Figure 5. MMSR super-resolution results on real-world images compared with state-of-the-art methods. Zoom in to appreciate the details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.356 + ], + "angle": 0, + "content": "often leads to sharper and more detailed outputs, it can also exacerbate hallucination, resulting in details inconsistent with the low-resolution input. Such artifacts are widely reported but are difficult to suppress, even with recent efforts like balancing the training data [71]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.53 + ], + "angle": 0, + "content": "Multimodal Classifier-free Guidance. To mitigate the issue of excessive hallucination often associated with high guidance scales in CFG, we propose a novel multimodal guidance strategy. We argue that the artifacts come from the weak guidance in the negative prompting process. Instead of simply relying on text prompts for guidance, we leverage the rich information encoded in the multimodal latent tokens to strengthen both positive and negative promptings. Specifically, we condition both the postive and negative generation processes on the multimodal latent token sequence, denoted as \\( m \\). This leads to the following multimodal CFG:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.535, + 0.482, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {\\epsilon}} (\\mathbf {z} _ {t}, c, m) = (1 + w) \\boldsymbol {\\epsilon} (\\mathbf {z} _ {t}, c, \\text {p o s}, m) - w \\boldsymbol {\\epsilon} (\\mathbf {z} _ {t}, c, \\text {n e g}, m). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Sec. 4.1 shows that this change in negative generation helps to achieve a better trade-off between perceptual quality and identity preservation, better maintaining the semantic content of the low-resolution input in the upscaled output, compared with the standard CFG in previous text-based methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.485, + 0.771 + ], + "angle": 0, + "content": "Scaling Single-modal Guidance. Multimodal CFG effectively controls the overall influence of the prompts but does not offer control over each modality individually. To address this limitation, we introduce a mechanism to selectively amplify or suppress the contribution of specific modalities. Specifically, we modify the attention temperature \\(\\delta\\) of MMLC during cross-conditioning, where \\(\\delta\\) scales the attention maps before applying the softmax operation:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.774, + 0.483, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\delta}\\right) V. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This parameter controls the sensitivity of the attention mechanism to differences between the feature sequence \\( Q \\) and the multimodal token sequences \\( K \\) and \\( V \\). A smaller temperature \\( \\delta \\) typically leads to a stronger conditioning effect on the attention mechanism. In the standard scaled dot-product attention, the temperature \\( \\delta \\) is empirically set to the square" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.4 + ], + "angle": 0, + "content": "root of the key dimension \\(\\sqrt{d_k}\\). We observe that scaling the standard temperature \\(\\sqrt{d_k}\\) (where \\(d_{k}\\) is the key dimension) within a range of [0.4, 10] can produce high-quality results, with varying levels of fidelity to the input modalities. By scaling the temperature during sampling, we achieve precise control over the super-resolution process, enabling manageable manipulation of the output with fine-grained control." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.415, + 0.734, + 0.431 + ], + "angle": 0, + "content": "3.3. MMSR Implementation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.591 + ], + "angle": 0, + "content": "Figure 3 illustrates our multimodal guided super-resolution pipeline. During inference, we first extract four modalities from the low-resolution (LR) image: a text caption generated by Gemini Flash [56], a depth map estimated by Depth Anything [66], a semantic segmentation mask produced by Mask2Former [11], and edge information extracted with a Canny edge detector. The depth map, segmentation mask, and edge information are encoded into token sequences using a pretrained VQGAN, while the text caption is processed by a pretrained CLIP encoder to obtain a text embedding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.907, + 0.654 + ], + "angle": 0, + "content": "Following a similar conditioning strategy to InstructPix2Pix [6], the LR image is concatenated with a noisy latent vector sampled from the diffusion model, providing additional conditioning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.671, + 0.645, + 0.689 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Training Details. Our super-resolution model is initialized with the weights of a pretrained text-to-image model, with the same architecture and size as Stable Diffusion v2 [53]. The super-resolution training dataset consists of randomly degraded (using RealESRGAN degradation [63]) high-resolution images from the combined LSDIR and DIV2K datasets, with corresponding \\(512 \\times 512\\) high-resolution images as ground truth. We set the batch size to 1024 and the learning rate to 1e-4, which we empirically found maximized compute efficiency on TPUv5 pods. During testing, all benchmarks use a model checkpoint finetuned for 160k iterations. We use 50-step DDIM sampling, consistent with previous methods. We use a guidance rate of 4 as the default." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23145" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.095, + 0.209, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.169, + 0.163, + 0.176 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.095, + 0.328, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.103, + 0.445, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.454, + 0.096, + 0.568, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.096, + 0.682, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.096, + 0.801, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.096, + 0.918, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.176, + 0.211, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.241, + 0.187, + 0.25 + ], + "angle": 0, + "content": "LR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.169, + 0.326, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.242, + 0.306, + 0.25 + ], + "angle": 0, + "content": "Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.169, + 0.445, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.242, + 0.406, + 0.25 + ], + "angle": 0, + "content": "Edge" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.169, + 0.569, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.242, + 0.552, + 0.25 + ], + "angle": 0, + "content": "PASD (Zoomed)" + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.169, + 0.645, + 0.176 + ], + "angle": 0, + "content": "SeeSR" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.177, + 0.685, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.242, + 0.668, + 0.25 + ], + "angle": 0, + "content": "SeeSR (Zoomed)" + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.168, + 0.78, + 0.176 + ], + "angle": 0, + "content": "MMSR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.177, + 0.802, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.242, + 0.787, + 0.25 + ], + "angle": 0, + "content": "MMSR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.177, + 0.919, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.242, + 0.895, + 0.25 + ], + "angle": 0, + "content": "HR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.259, + 0.207, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.324, + 0.163, + 0.331 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.262, + 0.321, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.324, + 0.293, + 0.332 + ], + "angle": 0, + "content": "Caption" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.333, + 0.211, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.396, + 0.187, + 0.405 + ], + "angle": 0, + "content": "LR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.333, + 0.326, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.396, + 0.306, + 0.405 + ], + "angle": 0, + "content": "Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.261, + 0.445, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.324, + 0.408, + 0.332 + ], + "angle": 0, + "content": "Depth" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.333, + 0.445, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.396, + 0.405, + 0.405 + ], + "angle": 0, + "content": "Edge" + }, + { + "type": "image", + "bbox": [ + 0.457, + 0.259, + 0.564, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.324, + 0.527, + 0.331 + ], + "angle": 0, + "content": "PASD" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.333, + 0.567, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.396, + 0.552, + 0.405 + ], + "angle": 0, + "content": "PASD (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.261, + 0.681, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.324, + 0.645, + 0.331 + ], + "angle": 0, + "content": "SeeSR" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.333, + 0.684, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.396, + 0.669, + 0.405 + ], + "angle": 0, + "content": "SeeSR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.259, + 0.798, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.323, + 0.78, + 0.332 + ], + "angle": 0, + "content": "MMSR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.333, + 0.802, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.396, + 0.788, + 0.405 + ], + "angle": 0, + "content": "MMSR (Zoomed)" + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.261, + 0.915, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.854, + 0.324, + 0.871, + 0.331 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.333, + 0.918, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.396, + 0.896, + 0.405 + ], + "angle": 0, + "content": "HR (Zoomed)" + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.413, + 0.884, + 0.427 + ], + "angle": 0, + "content": "Figure 6. Super-resolution results on common benchmarks, comparing MMSR with state-of-the-art techniques. Zoom in for detail." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Evaluation Details. We use reference-based metrics like LPIPS [75] and DISTS [16], and non-reference-based metrics like NIQE [42], MANIQA [67], MUSIQ [27], and CLIPIQA [58], as these have been shown to align well with humans' aesthetic preferences. Our compared baselines include recent diffusion-based super-resolution methods [33, 53, 59, 65, 68, 72] and the representative methods [8, 31, 32, 63, 73]. We collected their results either by running their official code or from published results. In a few cases where the multimodal prediction failed (two flat images and one pencil drawing from DRealSR), we manually replaced the prediction with \\( m_{\\emptyset} \\) (see supplemental material)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.634, + 0.285, + 0.649 + ], + "angle": 0, + "content": "4.1. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Table 1 shows the quantitative result comparisons of ours with the other baselines, compared on the synthetic benchmark and real-world super-resolution benchmarks. Several facts are worth noting: (1) Our method achieves the best LPIPS, DISTS, NIQE, and FID scores on the DIV2K-Val and RealSR dataset, significantly outperforming previous state-of-the-art methods. This superiority demonstrates that our method can generate more perceptually identical details from the guidance of multimodal context. (2) Our method also achieves the best performance in the non-reference visual quality metrics including NIQE, MANIQA, CLIPIQA, and MUSIQ scores. This advantage demonstrates our method can fully utilize the photorealistic priors encapsulated in the pretrained diffusion model. Overall, these observations fully demonstrate our advantages in using multimodal guidance compared with the past single-modality methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.908, + 0.695 + ], + "angle": 0, + "content": "Effects of Multimodal CFG Guidance. Table 3 demonstrates the effects of our proposed multimodal CFG on the 1MP DIV2K-Val benchmark, which replaces the negative score guided by empty language token \\(\\epsilon (\\mathbf{z}_t,c,\\mathrm{neg})\\), shortened as \\(cfg\\), with multimodal guided negative score \\((\\epsilon (\\mathbf{z}_t,c,m,\\mathrm{neg}))\\), shortened as \\(m\\) -cfg. In order to exclude the influence of architectural differences, we also compare our method with the empty multimodal token \\(\\epsilon (\\mathbf{z}_t,c,m_{\\emptyset},\\mathrm{neg})\\), shortened as \\(m_{\\emptyset}\\) -cfg. Note that the positive score for all three compared methods is the same \\(\\epsilon (\\mathbf{z}_t,c,m,\\mathrm{pos})\\). The results show that our method mitigates the degradation reflected in the LPIPS and NIQE scores when using guidance rates of 10 and 14 in both CFG and \\(m_{\\emptyset}\\) -cfg, while maintaining comparable performance. The visual results in Table 3 further demonstrate our superiority, where ours suppresses the color change and incorrect texture of high CFG guidance rate, leading to higher CLIPIQA score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.909, + 0.883 + ], + "angle": 0, + "content": "Effects of Latent Connector. To demonstrate the effectiveness of the proposed multimodal latent connector, we conducted an ablation study by training a new multimodal super-resolution model without the latent connector, which directly uses the longer multimodal token sequence as input. The quantitative performance is shown in Table 2. Our \\( w \\). MMLC outperforms the \\( w/o \\). MMLC in all reference and non-reference metrics. The visual results in Figure 7 show that the model without the MMLC is more likely to exacerbate hallucination. Its result shows bokeh on the branches but a clear trunk, even though the depth input clearly shows they are at the same depth." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Contributions of Each Modality. Figure 8 shows the con" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23146" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.091, + 0.277, + 0.101 + ], + "angle": 0, + "content": "DIV2K-Val-3k 512 × 512" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.103, + 0.48, + 0.277 + ], + "angle": 0, + "content": "
MethodsPSNRSSIMLPIPS ↓DISTS ↓NIQE ↓FID ↓MUSIQCLIPQA
BSRGAN21.870.55390.41360.27374.761564.2859.110.5183
R-ESRGAN21.940.57360.38680.26014.920953.4658.640.5424
LDL21.520.56900.39950.26885.024958.9457.900.5313
DASR21.720.55360.42660.26884.859667.2254.220.5241
FeMASR20.850.51630.39730.24284.572653.7058.100.5597
LDM21.260.52390.41540.25006.466741.9356.520.5695
StableSR20.840.48870.40550.25424.655136.5762.950.6486
ResShift21.750.54220.42840.26066.973155.7758.230.5948
PASD20.770.49580.44100.25384.832840.7766.850.6799
DiffBIR20.940.49380.42700.24714.721140.4265.230.6664
SeeSR21.190.53860.38430.22574.927531.9368.330.6946
MMSR21.740.56930.37070.20714.253229.3570.060.7164
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.286, + 0.284, + 0.295 + ], + "angle": 0, + "content": "DIV2K-Val-100 1024 × 1024" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.296, + 0.48, + 0.389 + ], + "angle": 0, + "content": "
MethodsPSNRSSIMLPIPS ↓DISTS ↓NIQE ↓MANIQAMUSIQCLIPiQA
R-ESRGAN21.770.58130.36240.19903.65730.404647.540.5358
StableSR20.070.39470.50970.24273.62600.411365.390.6938
PASD21.270.53690.34730.17533.63210.470869.690.6914
SUPIR20.650.53500.38490.18143.64580.405165.880.5697
SeeSR21.310.55780.32730.16204.02150.543969.790.6941
MMSR21.870.55650.28100.14923.42430.488572.310.7294
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.398, + 0.242, + 0.408 + ], + "angle": 0, + "content": "RealSR \\(512\\times 512\\)" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.41, + 0.48, + 0.505 + ], + "angle": 0, + "content": "
MethodsPSNRSSIMLPIPS ↓LIQENIMAMANIQAMUSIQCLIPiQA
R-ESRGAN25.690.76160.27273.35744.65480.548760.180.4449
StableSR24.700.70850.30183.61064.81500.622165.780.6178
PASD24.290.66300.34353.57494.85540.649368.690.6590
SUPIR22.970.62980.37503.56824.57570.574561.490.6434
SeeSR25.180.72160.30094.13604.91930.644269.770.6612
MMSR24.830.70030.29524.34685.10940.657871.330.6717
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.513, + 0.248, + 0.522 + ], + "angle": 0, + "content": "DrealSR \\(512\\times 512\\)" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.524, + 0.48, + 0.621 + ], + "angle": 0, + "content": "
MethodsPSNRSSIMLPIPS ↓LIQENIMAMANIQAMUSIQCLIPiQA
R-ESRGAN28.640.80530.28472.92554.32580.490754.180.4422
StableSR26.710.72240.32843.24254.48610.559458.510.6357
PASD27.000.70840.39313.59084.66180.585064.810.6773
SUPIR24.610.61230.42943.47104.38150.538157.320.6758
SeeSR26.750.74050.31744.12704.69420.605265.090.6908
MMSR27.280.74560.32494.50235.05580.630168.930.6999
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.668 + ], + "angle": 0, + "content": "Table 1. Severely degraded, low-resolution (LR) images can produce inaccurate multi-modal information, manifesting as distorted edges, misidentified objects, and other artifacts." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.671, + 0.48, + 0.716 + ], + "angle": 0, + "content": "
MUSIQNIQE ↓DISTS ↓LPIPS ↓Throughput
w/o. MMLC69.693.48450.17810.39293.48 img/s
w. MMLC72.313.42430.14920.28103.32 img/s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.721, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Table 2. Ablation of the impact of the Multimodal Latent Connector module (MMLC) on DIV2K-Val-100 1024p." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.486, + 0.902 + ], + "angle": 0, + "content": "tributions of each modality by masking out input modalities during testing. Benefiting from the improved flexibility of using multimodal token \\( m_{\\emptyset} \\), discussed in Sec. 3.1, our method is robust in scenarios with only fewer modalities input during testing. We would like to note several observations: (1) Our default multimodal setting achieves the best trade-off between the non-reference metric MUSIQ and the reference-based metric DISTS. (2) Depth information primarily enhances perceptual quality (MUSIQ), while other modalities" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.091, + 0.704, + 0.15 + ], + "angle": 0, + "content": "
guidance21014
cfg0.32390.44910.5064
m0cfg0.28150.48030.5493
mcfg0.28100.34710.3772
" + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.153, + 0.649, + 0.163 + ], + "angle": 0, + "content": "(a) LPIPS score" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.164, + 0.748, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.295, + 0.72, + 0.306 + ], + "angle": 0, + "content": "Segment, Depth, LR, and Edge Input" + }, + { + "type": "table", + "bbox": [ + 0.716, + 0.091, + 0.899, + 0.15 + ], + "angle": 0, + "content": "
guidance21014
cfg3.5774.61795.1886
m0cfg3.62615.10815.9175
mcfg3.46793.74193.9815
" + }, + { + "type": "table_caption", + "bbox": [ + 0.769, + 0.153, + 0.842, + 0.163 + ], + "angle": 0, + "content": "(b) NIQE score" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.164, + 0.899, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.191, + 0.884, + 0.199 + ], + "angle": 0, + "content": "\\(cfg = 2.0\\) CIPIQA \\(= 0.4481\\)" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.199, + 0.898, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.226, + 0.886, + 0.234 + ], + "angle": 0, + "content": "\\(cfg = 14.0\\) CLIPIQA=0.7091" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.234, + 0.898, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.261, + 0.893, + 0.269 + ], + "angle": 0, + "content": "\\(m\\text{-} cfg = 2.0\\) CLIPIQA \\(= 0.4122\\)" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.269, + 0.898, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.295, + 0.895, + 0.305 + ], + "angle": 0, + "content": "\\(m\\text{-} cfg = 14.0\\) CIPIQA \\(= 0.7352\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.308, + 0.765, + 0.318 + ], + "angle": 0, + "content": "(c) Visual comaprisons" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.318, + 0.906, + 0.373 + ], + "angle": 0, + "content": "Table 3. Ablation of guidance on DIV2K-Val-100 1024p. Our multimodal CFG (\\(m\\)-cfg) mitigates artifacts often present when using high guidance rates. This leads to an improved balance between visual fidelity and preservation of key identifying features." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.376, + 0.707, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.436, + 0.638, + 0.446 + ], + "angle": 0, + "content": "LR Input" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.376, + 0.901, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.436, + 0.835, + 0.446 + ], + "angle": 0, + "content": "Depth Input" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.447, + 0.708, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.505, + 0.663, + 0.516 + ], + "angle": 0, + "content": "w/o. MMLC Result" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.446, + 0.899, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.506, + 0.848, + 0.515 + ], + "angle": 0, + "content": "w. MMLC Result" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.519, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Figure 7. Visual results when (not) using the MMLC module." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.54, + 0.738, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.541, + 0.899, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.613, + 0.906, + 0.656 + ], + "angle": 0, + "content": "Figure 8. Ablation of modalities on DIV2K-Val-100 1024p. Incorporating different modalities enhances various aspects of the super-resolution results compared to the text-guided baseline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.907, + 0.745 + ], + "angle": 0, + "content": "contribute more significantly to preserving identity (DISTS). These results highlight the diverse contributions of each modality and emphasize the advantage of our multimodal approach in effectively combining their strengths for optimal super-resolution performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.696, + 0.772 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Figures 5 and 6 provide a visual comparison of our method against state-of-the-art approaches, namely SeeSR [65], PASD [68], and SUPIR [71]. Our method demonstrates several key advantages. Enhanced Realism: In the first example of Figure 6, our method produces fewer artifacts and remains more faithful to the high-resolution image compared to other methods. This highlights our ability to maintain realism without introducing spurious details. Robustness to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23147" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.098, + 0.203, + 0.108 + ], + "angle": 0, + "content": "(a). Results of PASD" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.115, + 0.216, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.196, + 0.216, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.279, + 0.216, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.098, + 0.721, + 0.108 + ], + "angle": 0, + "content": "(b). Results of MMSR under different modality-specific controls" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.115, + 0.342, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.196, + 0.342, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.278, + 0.342, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.118, + 0.895, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.2, + 0.895, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.283, + 0.895, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.369, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 9. Our method allows for fine-grained control over super-resolution results by adjusting the influence of each input modality. For example, reducing the edge temperature enhances edge sharpness (first row). Lowering the segmentation temperature emphasizes distinct features, such as the star pattern on the flag (second row). Decreasing the depth temperature accentuates depth-of-field effects, like the bokeh between the foreground and background (third row). In contrast, PASD [68] exhibits limited control over such fine-grained details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Challenging Conditions: The second example in Figure 6, a long-range shot affected by noise and turbulence, shows our method's ability to generate clear human details where other methods struggle. This is attributed to the effective use of semantic segmentation for accurate human localization, further enriched by the textual caption. These examples illustrate the effectiveness of our multimodal approach in generating high-quality images across diverse scenarios. See supplemental material for additional results and analysis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.347, + 0.619 + ], + "angle": 0, + "content": "4.3. Controllability Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "While recent works leverage text prompts for controlling super-resolution [65, 68, 71], these often lack fine-grained control and can yield inconsistent results. Our multimodal approach introduces modality-specific temperature weights to scale attention scores, enabling precise manipulation of SR outputs by amplifying or diminishing the influence of each modality (e.g., depth, segmentation). Figure 9 contrasts our approach with PASD [68], which relies solely on text prompts. Specifically, decreasing the temperature of the edge modality changes the richness of details in our result, while changing PASD's prompt with \"more edge\" doesn't change their result. Decreasing the temperature of semantic segmentation map and depth map also lead to manageable changes such as more visible star pattern and stronger bokeh. In contrast, directly add corresponding prompt in PASD's prompt barely changes its result. Moreover, our approach exhibits smooth transitions in image characteristics as temperatures are adjusted, providing interpretable control and insights into" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.451, + 0.684, + 0.466 + ], + "angle": 0, + "content": "the role of each modality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.633, + 0.495 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.908, + 0.701 + ], + "angle": 0, + "content": "This work introduces a novel diffusion-based framework for image super-resolution that seamlessly integrates diverse modalities—including text descriptions, depth maps, edges, and segmentation maps—within a single, unified model. By leveraging pretrained text-to-image models, our method achieves enhanced realism and accurate reconstruction, outperforming existing text-guided super-resolution methods both qualitatively and quantitatively. Furthermore, a learnable multimodal token and modality-specific contribution controls provide fine-grained control over the super-resolution process, enabling adjustable perception-distortion tradeoffs and robust performance even with imperfect or missing modalities in most cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitations and Future Work. While adding multimodal information significantly enhances SR performance, it introduces computational overhead. For instance, using Gemini Flash for image captioning results in a throughput of 0.34 images per second, which is slower than depth at \\(1.99\\mathrm{img / s}\\), semantic segmentation at \\(2.09\\mathrm{img / s}\\), and DDIM sampling at \\(0.54\\mathrm{img / s}\\). However, cross-modal predictions can be parallelized, and our method achieves speeds comparable to other text-driven SR models [68, 71]. Future work will explore optimizing the vision-language component for faster inference and investigate more robust modules for extracting modality-specific information, potentially enhancing performance even with noisy or incomplete inputs." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23148" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "Vishal M. Patel was supported by NSF CAREER award 2045489. We are grateful to Keren Ye and Mo Zhou for their valuable feedback. We also extend our gratitude to Shlomi Fruchter, Kevin Murphy, Mohammad Babaeizadeh, and Han Zhang for their instrumental contributions in facilitating the initial implementation of the latent diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.304, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[2] Roman Bachmann, Oğuzhan Fatih Kar, David Mizrahi, Ali Garjani, Mingfei Gao, David Griffiths, Jiaming Hu, Afshin Dehghan, and Amir Zamir. 4M-21: An any-to-any vision model for tens of tasks and modalities. NeurIPS, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.484, + 0.416 + ], + "angle": 0, + "content": "[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, pages 22861-22872, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2023.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[5] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In CVPR, pages 6228-6237, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.484, + 0.617 + ], + "angle": 0, + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, pages 14455-14465, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[8] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Real-world blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.484, + 0.772 + ], + "angle": 0, + "content": "[9] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25857-25867, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[10] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.247, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "[12] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. arXiv preprint arXiv:2310.01110, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.908, + 0.189 + ], + "angle": 0, + "content": "[13] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.907, + 0.244 + ], + "angle": 0, + "content": "[14] Maurizio Delbracio and Peyman Milanfar. Inversion by direct iteration: An alternative to denoising diffusion for image restoration. Transactions on Machine Learning Research, 2023. Featured Certification. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.907, + 0.314 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.908, + 0.383 + ], + "angle": 0, + "content": "[16] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In Proceedings of the 29th ACM International Conference on multimedia, pages 2483-2491, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.907, + 0.44 + ], + "angle": 0, + "content": "[17] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.496 + ], + "angle": 0, + "content": "[18] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.908, + 0.566 + ], + "angle": 0, + "content": "[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.908, + 0.621 + ], + "angle": 0, + "content": "[20] Kanchana Vaishnavi Gandikota and Paramanand Chandramouli. Text-guided explorable image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25900-25911, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.908, + 0.69 + ], + "angle": 0, + "content": "[21] Jiatao Gu, Ying Shen, Shuangfei Zhai, Yizhe Zhang, Navdeep Jaitly, and Joshua M Susskind. Kaleido diffusion: Improving conditional diffusion models with autoregressive latent modeling. Advances in neural information processing systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.908, + 0.76 + ], + "angle": 0, + "content": "[22] Jiatao Gu, Yuyang Wang, Yizhe Zhang, Qihang Zhang, Dinghuai Zhang, Navdeep Jaitly, Josh Susskind, and Shuangfei Zhai. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.908, + 0.871 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23149" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.486, + 0.133 + ], + "angle": 0, + "content": "perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.486, + 0.19 + ], + "angle": 0, + "content": "[26] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's \"up\" with vision-language models? investigating their struggle with spatial reasoning. In EMNLP, pages 9161–9175, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.485, + 0.248 + ], + "angle": 0, + "content": "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.485, + 0.319 + ], + "angle": 0, + "content": "[28] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning, pages 3744-3753. PMLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.485, + 0.376 + ], + "angle": 0, + "content": "[29] Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image super-resolution. In Proceedings of the European conference on computer vision (ECCV), pages 517-532, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.485, + 0.446 + ], + "angle": 0, + "content": "[30] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.485, + 0.516 + ], + "angle": 0, + "content": "[31] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.485, + 0.574 + ], + "angle": 0, + "content": "[32] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.485, + 0.631 + ], + "angle": 0, + "content": "[33] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diffbir: Towards blind image restoration with generative diffusion prior. ArXiv preprint, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.485, + 0.701 + ], + "angle": 0, + "content": "[34] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.485, + 0.744 + ], + "angle": 0, + "content": "[35] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.485, + 0.788 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.485, + 0.843 + ], + "angle": 0, + "content": "[37] Kangfu Mei, Vishal M Patel, and Rui Huang. Deep semantic statistics matching (d2sm) denoising network. In European Conference on Computer Vision, pages 384-400. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[38] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image generation. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 9048-9058, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.188 + ], + "angle": 0, + "content": "[39] Kangfu Mei, Luis Figueroa, Zhe Lin, Zhihong Ding, Scott Cohen, and Vishal M Patel. Latent feature-guided diffusion models for shadow removal. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4313-4322, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[40] Kangfu Mei, Zhengzhong Tu, Mauricio Delbracio, Hossein Talebi, Vishal M Patel, and Peyman Milanfar. Bigger is not always better: Scaling properties of latent diffusion models. arXiv preprint arXiv:2404.01367, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.907, + 0.327 + ], + "angle": 0, + "content": "[41] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.907, + 0.371 + ], + "angle": 0, + "content": "[42] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.907, + 0.425 + ], + "angle": 0, + "content": "[43] David Mizrahi, Roman Bachmann, Oğuzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4M: Massively multimodal masked modeling. NeurIPS, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.907, + 0.481 + ], + "angle": 0, + "content": "[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[45] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.622 + ], + "angle": 0, + "content": "[46] Mehdi Noroozi, Isma Hadji, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. You only need one step: Fast super-resolution with stable diffusion via scale distillation. In European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.907, + 0.691 + ], + "angle": 0, + "content": "[47] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.907, + 0.746 + ], + "angle": 0, + "content": "[48] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.907, + 0.802 + ], + "angle": 0, + "content": "[49] Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi. Spire: Semantic prompt-driven image restoration. In European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[50] Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou. Xpsr: Cross-modal priors for diffusion-based image super-resolution. In European Conference on Computer Vision, pages 285-303. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[51] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23150" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.486, + 0.189 + ], + "angle": 0, + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.485, + 0.259 + ], + "angle": 0, + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.484, + 0.411 + ], + "angle": 0, + "content": "[55] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.486, + 0.552 + ], + "angle": 0, + "content": "[57] Li-Yuan Tsao, Yi-Chen Lo, Chia-Che Chang, Hao-Wei Chen, Roy Tseng, Chien Feng, and Chun-Yi Lee. Boosting flow-based generative super-resolution models via learned prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26005–26015, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.593 + ], + "angle": 0, + "content": "[58] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[59] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[60] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[61] Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, and Shanshan Zhao. Caption anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[62] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[63] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Realesrgan: Training real-world blind super-resolution with pure synthetic data supplementary material. Computer Vision Foundation open access, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[64] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.149 + ], + "angle": 0, + "content": "Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[65] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.908, + 0.288 + ], + "angle": 0, + "content": "[66] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[67] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[68] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.907, + 0.511 + ], + "angle": 0, + "content": "[69] Zhixiong Yang, Jingyuan Xia, Shengxi Li, Xinghua Huang, Shuanghui Zhang, Zhen Liu, Yaowen Fu, and Yongxiang Liu. A dynamic kernel prior model for unsupervised blind image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26046-26056, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.908, + 0.566 + ], + "angle": 0, + "content": "[70] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.566, + 0.908, + 0.65 + ], + "angle": 0, + "content": "[71] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 1, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.908, + 0.705 + ], + "angle": 0, + "content": "[72] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super-resolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.908, + 0.774 + ], + "angle": 0, + "content": "[73] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.907, + 0.832 + ], + "angle": 0, + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[75] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23151" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.486, + 0.135 + ], + "angle": 0, + "content": "[76] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. In European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.136, + 0.486, + 0.192 + ], + "angle": 0, + "content": "[77] Zheyuan Zhang, Fengyuan Hu, Jayjun Lee, Freda Shi, Parisa Kordjamshidi, Joyce Chai, and Ziqiao Ma. Do vision-language models represent space and how? evaluating spatial frame of reference under ambiguities, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.093, + 0.486, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23152" + } + ] +] \ No newline at end of file diff --git a/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_origin.pdf b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82b874973116b032eb38157e8cb053afc9743a7a --- /dev/null +++ b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/10472ffe-820d-4b89-beac-a0259c3bd355_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51f642c4a1170b0bc5b6da7432cafb26af15afab9135533d97cff3013cee7309 +size 5270209 diff --git a/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/full.md b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..982ff47838d568124cc41c2b10f4e5d6919c7f48 --- /dev/null +++ b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/full.md @@ -0,0 +1,481 @@ +# The Power of Context: How Multimodality Improves Image Super-Resolution + +Kangfu Mei\*1,2, Hossein Talebi', Mojtaba Ardakani', Vishal M. Patel?, Peyman Milanfar', Mauricio Delbracio' 1 Google, 2 Johns Hopkins University Project Page: https://mmsr.kfmei.com/ + +![](images/1ddbe435e00ca0c7626f7fa2886ba73af9bd9e8569079aa84632bc6bfe743901.jpg) +Figure 1. Our Multimodal Super-Resolution (MMSR) method leverages the rich context of multimodal guidance, including image captions, depth maps, semantic segmentation maps, and edges inferred from LR. MMSR surpasses state-of-the-art methods by producing more realistic results and suppressing artifacts that, while plausible, are inconsistent with the information present in the LR input. + +# Abstract + +Single-image super-resolution (SISR) remains challenging due to the inherent difficulty of recovering fine-grained details and preserving perceptual quality from low-resolution inputs. Existing methods often rely on limited image priors, leading to suboptimal results. We propose a novel approach that leverages the rich contextual information available in multiple modalities – including depth, segmentation, edges, and text prompts—to learn a powerful generative prior for SISR within a diffusion model framework. We introduce a flexible network architecture that effectively fuses multimodal information, accommodating an arbitrary number of input modalities without requiring significant modifications to the diffusion process. Crucially, we mitigate hallucinations, often introduced by text prompts, by using spatial information from other modalities to guide regional text-based conditioning. Each modality's guidance strength can also be controlled independently, allowing steering outputs toward different directions, such as increasing bokeh through depth or adjusting object prominence via segmentation. Extensive experiments demonstrate that our model surpasses state-of-the-art generative SISR methods, achieving superior visual quality and fidelity. + +# 1. Introduction + +Single image super-resolution (SISR) aims to generate high-resolution images from low-resolution inputs while preserving + +ing semantic identity and texture details. While it is not essential to treat SISR as a regression problem, past methods [17, 29, 37, 62] typically use a deep neural network to learn a direct mapping from low-resolution images to high-resolution images. Even though these regression-based methods achieve good scores on paired metrics like PSNR and SSIM, they have failed to produce results with high quality comparable to natural images, a task at which recent generative models have excelled [5, 12, 14]. The advent of powerful generative models, such as autoregressive models [10, 51] and diffusion models [47, 53, 54], has revolutionized image generation tasks, including text-to-image synthesis. This has inspired recent efforts to leverage these pre-trained generative models for downstream tasks like SISR [6, 41, 53]. For instance, very recent works [50, 71] achieve super-resolution by leveraging emerging vision-language models (e.g., Gemini [56], LLaVA [36], ChatGPT-4 [1]) and pretrained text-to-image models to first generate captions from low-resolution images and then use these captions as prompts to generate high-resolution images. + +While providing rich textual descriptions can significantly enhance the quality of generated images [4, 19, 34], relying solely on text for SISR poses challenges. Recent works [7, 26, 35, 77] have shown that text prompts cannot represent spatial relationships. This implies that textual information, such as texture descriptions, can only be applied to the whole image. Figure 1 provides a representative example. Previous text-based super-resolution methods [65, 68] use a 'lion' + +caption, which results in a furry tongue. However, while lions have fur, their tongues do not grow hair. Can we leverage spatial cues from depth and segmentation data to improve the learned prior and enhance the quality of SISR? + +Fortunately, we can directly extract additional modalities from the low-resolution image by using various pretrained cross-modal prediction models [11, 66]. In this paper, we introduce a new diffusion model architecture dedicated to multimodal super-resolution, which is conceptually illustrated in Figure 2. By integrating multiple modalities into a single diffusion model, our method overcomes the challenges of recovering fine-grained details and preserving perceptual quality. Specifically, we propose conditioning diffusion models on modalities including text captions, semantic segmentation maps, depth maps, and edges to implicitly align text captions for correctly prompting different regions. + +We demonstrate the proposed multimodal diffusion model on the SISR task. Our effective multimodal architecture achieves better realism in SISR results than the best text-driven method and largely eliminates hallucinations that do not match the input. Moreover, we find that the multimodal SISR enables a new controlling feature, where we can explicitly adjust the weights of each modality to steer the generated results in different directions. + +Our contributions are summarized as follows: + +- We demonstrate the effectiveness of token-wise encoding for seamlessly injecting multiple modalities into pretrained text-to-image diffusion models without architectural modifications or significant model size overhead. +- We propose a novel multimodal latent connector that efficiently fuses information from different modalities, maintaining linear time complexity with respect to the number of modalities. +- We introduce a new multimodal classifier-free guidance technique that enhances realism at higher guidance rates while mitigating excessive hallucinations and fake details. +- Our method enables adjustment of the influence of each modality, allowing for fine-grained manipulation of SISR results while preserving realism and quality. + +# 2. Related Work + +Generative Prior Powered Super-resolution. Recent advances in image restoration leverage the power of foundational models to enhance the quality of degraded images [9, 13, 20, 33, 38, 39, 46, 50, 57, 64, 69, 72, 76]. This trend is fueled by the capacity of pretrained generative models to capture natural image statistics and transfer this knowledge to the super-resolution task, enabling photorealistic image generation. Early works in this domain include LDM-SR [53] and StableSR [59] for single image superresolution. Beyond this, methods like InstructPix2Pix [6] utilize instructions for image editing, while ControlNet [74] + +![](images/25b71099370a5a1779372762ce898e29b44e412e97821444b001b59b9f22fb81.jpg) +Figure 2. Language models struggle to accurately represent spatial information, leading to coarse and imprecise image superresolution. To overcome this limitation, we incorporate additional spatial modalities like depth maps and semantic segmentation maps. These modalities provide detailed spatial context, allowing our model to implicitly align language descriptions with individual pixels through a transformer network. This enriched understanding of the image significantly enhances the realism of our super-resolution results and minimizes distortion. + +and IP-Adapter [70] facilitate cross-modality image translation. These approaches highlight a common observation: the quality of results in downstream tasks is strongly correlated with the quality of the pretrained models [40]. + +Recent works focus on leveraging powerful text-to-image diffusion models, employing text prompts to fully exploit the learned prior. PASD [68] uses image content descriptions; SPIRE [49] and PromptIP [48] use degradation descriptions; and SeeSR [65] employs a combination of context and degradation descriptions. Compared to earlier methods that directly fine-tuned diffusion models on super-resolution data, these text-prompt-driven approaches not only achieve superior realism but also enable multi-faceted outputs by conditioning on different prompts [20]. This capability enriches the typically single-output super-resolution task. + +While text prompts offer advantages, they can be ambiguous in representing spatial relationships [4, 7, 19, 34]. Our proposed MMSR framework addresses this by integrating multimodal inputs within a novel architecture. We also demonstrate native super-resolution effects control with this new architecture, enabling control of both overall realism and the effect of each modality. + +Vision-language Understanding and Generation. Recent large vision-language models (e.g., Gemini [56], LLaVA [36], GPT-4o [1]) excel in tasks like image captioning [61], but translating visual information into diverse modalities for image generation remains challenging. While models like 4M [2, 43] effectively extract high-level visual information (depth, normals, semantics, etc.), leveraging these for generation is a promising direction. Recent exploration includes GLIGEN [30] generate images from bounding boxes, and ControlNet [74] generate images from depth and other modalities. More recent works [21, 22] explore using discrete hidden-feature tokens to guide generation. In this paper, we introduce a new approach for multimodal control with discrete vision tokens, effectively integrating depth maps, + +![](images/6520790bb98cff0df4b5f817ac4cb7189a2075aa9ecdbcfe5e30fd81417f3ab1.jpg) +Figure 3. This diagram illustrates our multimodal super-resolution pipeline. Starting with a low-resolution (LR) image, we extract modalities like depth and semantic segmentation maps. These modalities are encoded into tokens and transformed into multimodal latent tokens $(m)$ . Our diffusion model uses these tokens and the LR input to generate a high-resolution (SR) output. A multimodal classifier-free guidance $(m$ -cfg) refines the SR image for enhanced quality. + +semantic segmentation maps, and text prompts for superior performance. Experiments on real-world super-resolution demonstrate the effectiveness of our approach. + +# 3. Method + +Single-image super-resolution aims to recover a high-resolution image $\mathbf{x}$ from its low-resolution counterpart $\mathbf{x}_{LR}$ . This is an ill-posed problem, often leading to generative models that produce "hallucinated" details – plausible yet inconsistent with the input. To mitigate this, we introduce auxiliary information, such as depth $(m_{\mathrm{dep}})$ , semantic segmentation $(m_{\mathrm{seg}})$ , and edge maps $(m_{\mathrm{edg}})$ , collectively denoted as $m$ . By conditioning the generative process on both the low-resolution image and these auxiliary modalities, we propose a new distribution $p(\mathbf{x}|\mathbf{x}_{LR}, m)$ with reduced uncertainty compared to the original distribution $p(\mathbf{x}|\mathbf{x}_{LR})$ . + +This reduction in uncertainty can be understood through the lens of information theory. The auxiliary modality $m$ provides additional information about the high-resolution image $\mathbf{x}$ that is not present in the low-resolution input $\mathbf{x}_{LR}$ . Since (conditional) mutual information is non-negative: + +$$ +I (\mathbf {x}; m | \mathbf {x} _ {L R}) = H (p (\mathbf {x} | \mathbf {x} _ {L R})) - H (p (\mathbf {x} | \mathbf {x} _ {L R}, m)), \tag {1} +$$ + +where $H(\cdot)$ denotes entropy. Consequently, the entropy of the conditional distribution with the auxiliary modality is: + +$$ +H \left(p \left(\mathbf {x} \mid \mathbf {x} _ {L R}\right)\right) \geq H \left(p \left(\mathbf {x} \mid \mathbf {x} _ {L R}, m\right)\right). \tag {2} +$$ + +The same motivation for reducing uncertainty is also shared by recent diffusion guidance works [23, 44]. While low-entropy sampling does not guarantee sharper images, it often leads to outputs with better visual quality compared to standard sampling. Motivated by this observation, we introduce a diffusion model [24] to learn the multimodal distribution $p(\mathbf{x}|\mathbf{x}_{LR},m)$ , effectively incorporating auxiliary information to enhance SISR quality. + +Ideally, the auxiliary modalities should provide information complementary to the low-resolution input. In practice, the auxiliary modalities are derived from the high-resolution image during training, ensuring informative conditioning. While we use modalities derived from the low-resolution input during inference, we demonstrate that this still leads to superior performance, including higher-quality details and improved fidelity to the input. + +# 3.1. Unified Multimodal Diffusion Conditioning + +We introduce a new diffusion network architecture, illustrated in Figure 3, for simultaneously conditioning on multiple modalities. Unlike recent methods like ControlNet [74] and IP-Adapter [70], which duplicate network components for each modality and incur significant computational overhead, our approach leverages a pretrained VQGAN image tokenizer [18]. This allows us to encode diverse modalities into a unified token representation for conditioning the diffusion model, without introducing additional model parameters or modifying the diffusion network itself. These tokens are concatenated with the text prompt embedding and used for cross-attention within the diffusion model. To efficiently process this long token sequence, we introduce a lightweight multimodal connector. This connector employs a dedicated architecture to achieve linear complexity for cross-attention, significantly reducing the computational burden. In what follows we present the implementation details. + +Token-wise Multimodal Encoding. While VQGAN [18] has proven effective for cross-modal encoding in image understanding and generation [2, 3, 43], its optimal application for image super-resolution remains an open question. Unlike tasks focused on understanding or generation, superresolution requires tokenization to not only capture semantic information but also preserve pixel-wise details crucial for + +![](images/778161cd7d44af9759e8652466760e5134190732f4932803dcd9db98f7c50262.jpg) +Figure 4. Using discrete multimodal tokens leads to superior reconstruction of modalities compared to continuous tokens. + +![](images/78c20180567971698edb70f60abe6d3b0781aa0e177ac5af7b515c19e18469e9.jpg) + +![](images/de0f768b1e15772c2c9fa84ae6c9b57066674ba2bc5c515a541ea8909b7d25d6.jpg) + +![](images/777b723c7a4b9b2cb2ca8e6a86951d08dc2044a8d2821b619e02dc04e9837df9.jpg) + +accurate reconstruction, such as spatial relationships within a depth map. Inspired by recent discussions on using discrete or continuous tokenization for autoregressive image generation [55], we investigate the impact of quantization on multimodal super-resolution. Specifically, we analyze how different quantization strategies affect the reconstruction quality of individual modalities and the final generated image. To this end, we utilize a VQGAN model pre-trained on a large-scale multimodal image dataset. + +Figure 4 compares the reconstruction quality of multi-modal tokens before and after quantization. As the figure illustrates, discrete tokens (post-quantization) better preserve individual modality information, while continuous tokens introduce noticeable artifacts. Consequently, we employ discrete tokens for our super-resolution experiments. + +Our multimodal encoding process utilizes the encoder and quantizer of the pre-trained VQGAN. Each $256 \times 256$ input modality is encoded into a $16 \times 16$ multimodal token sequence with a feature dimension of 256. These tokens are then quantized using a codebook of size 1024. To facilitate concatenation with text embeddings, we pad the feature dimension of the multimodal tokens to 1024, resulting in a $(256 \times 3 + 77) \times 1024$ multimodal token sequence condition. + +Multimodal Latent Connector. While cross-attention provides a flexible mechanism for conditioning diffusion models on multimodal data, its quadratic complexity with respect to the number of condition tokens introduces a significant computational burden. To address this, we introduce the Multimodal Latent Connector (MMLC), inspired by recent advances in efficient attention architectures [25, 28, 60]. + +As illustrated in Figure 3, the MMLC employs a transformer architecture to efficiently process the multimodal token sequence. The transformer receives two inputs: a randomly initialized sequence of learnable latent tokens, and the multimodal input sequence. The output is a token sequence of the same length as the latent token sequence, which serves as conditioning for the diffusion model. Therefore, the diffusion model conditions on fixed-length latent tokens (128 in our experiments), which are significantly shorter than the original multimodal token sequence input. The MMLC distills the essential information from the longer multimodal + +token sequence into the shorter multimodal latent token sequence through cross-attention. Following cross-attention, several self-attention blocks further process the latent token sequence, allowing the model to fully integrate the distilled information. + +This approach significantly reduces the computational cost of cross-attention in the diffusion model. Standard self-attention operates on the full multimodal token sequence $(K, V \in \mathbb{R}^{M \times D})$ , resulting in a time complexity of $\mathcal{O}(M^2)$ where $M$ is the length of the sequence and $D$ is the dimensionality of each token. In contrast, the MMLC uses a cross-attention between the latent token sequence (of size $N \times D$ ) and the multimodal token sequence which reduces this to $\mathcal{O}(MN)$ . Here $N$ is the length of the latent sequence and $N \ll M$ . This linear complexity with respect to the multimodal sequence length enables efficient processing of high-dimensional multimodal data, effectively capturing essential information for super-resolution. + +Flexible Multimodal Input. To enhance flexibility and robustness, we enable our method to handle scenarios where certain input modalities are unreliable or unavailable. We adopt a learnable embedding approach inspired by DALL-E 2 [52]. Specifically, we introduce a special learnable token, $m_{\emptyset}$ , optimized alongside the diffusion model and MMLC to represent the absence of a modality. During training, we independently randomly replace each modality with 256 $m_{\emptyset}$ tokens with a probability of 0.1. This encourages the model to learn robust representations that can effectively handle missing information. During inference, any missing modality is represented by a sequence of 256 $m_{\emptyset}$ tokens. This approach allows for flexible multimodal input, enabling the model to generate high-quality images even with limited or no auxiliary modalities. As demonstrated in our experiments, this strategy significantly improves performance when dealing with input that includes fewer modalities than were available during training. + +# 3.2. Multimodal Guidance and Control + +Building upon the success of guidance techniques [15, 23] in improving sample quality across various image generation tasks [45, 47, 52], recent diffusion-based SISR methods have begun incorporating prompt tuning to enhance superresolution results [59, 65, 71]. These methods improve the result by using negative prompts with Classifier-free Guidance (CFG) [23]. which can be expressed as: + +$$ +\tilde {\epsilon} (\mathbf {z} _ {t}, c) = (1 + w) \epsilon (\mathbf {z} _ {t}, c, \text {p o s}) - w \epsilon (\mathbf {z} _ {t}, c, \text {n e g}), \tag {3} +$$ + +where $\tilde{\epsilon} (\mathbf{z}_t,c)$ represents the guided denoising process, $\pmb {\epsilon}(\cdot)$ denotes the diffusion model, $\mathbf{z}_t$ denotes the noisy image latent, $c = \{\mathbf{x}_{\mathrm{LR}},t,\dots \}$ denotes the conditioning inputs (including the low-resolution image $\mathbf{x}_{\mathrm{LR}}$ , timestep $t$ , and other parameters), $w$ is the guidance scale, and pos and neg are the positive and negative prompts. While increasing $w$ + +LR +Inputs +![](images/1060f0d42e54684ab8ed1a13af25cf12e53f001bec4124c2a59cf616fa1af652.jpg) +The image shows the words "CIPAL PARK" prominently displayed on a metal structure. Below the main sign are several smaller screens displaying baseball ... + +![](images/4a90369b5705d808dfc40a768960412bd9d15a09df82ff239e2e478506dc03c9.jpg) +Depth +Figure 5. MMSR super-resolution results on real-world images compared with state-of-the-art methods. Zoom in to appreciate the details. + +![](images/4f545a9c5c14865e7203b02e458015917361c3541f2e77876d07f61852e4ebaa.jpg) +Caption +Edge + +![](images/98ac589cd0485996e53efff4146db3c8145996c837b2dcc0901d82d5e7835835.jpg) + +![](images/60693ef1d446c7837706e7127d90bd34adbd3cc50e5c8ed824aff947a3e5faa3.jpg) +Patch 1 +Patch 2 + +![](images/dd66632d97f0b46f0416f073458b1c016ec0d590cd1b92b0f764a1cec363d657.jpg) + +![](images/50814422c0a98a2ecd1e24cbe3b9b3fea6d9b8139b02ed309b50bdaee2490113.jpg) +PASD (Zoomed) +PASD (Zoomed) + +![](images/833ec91aea8fe8e1ecf98e34e60597041ba53bdd973183c3ef4407daed954497.jpg) + +![](images/dc7ac3e89e425fc07566a04ef5f0a8f7d0b5a4132ecf15da3698999e99445db0.jpg) +SeeSR (Zoomed) +SeeSR (Zoomed) + +![](images/036d27b098b7638971fe0d417f508518019c019493bee1f94420996ebae5892e.jpg) +uts + +![](images/e598cb327a5a8912fa75cf7a203ec72f9a1d9fc7d83922afdd84ca584579758f.jpg) +SUPIR (Zoomed) +SUPIR (Zoomed) + +![](images/9be8fecc9e1b9c9d878a2c4007a38d68df0f1a16ed942cc510bb87522c7ffcdd.jpg) + +![](images/84ca9552e002008f2d6748c92be9e1993d5c344e850a9b4117a81f280a89b647.jpg) +MMSR (Zoomed) +MMSR (Zoomed) + +often leads to sharper and more detailed outputs, it can also exacerbate hallucination, resulting in details inconsistent with the low-resolution input. Such artifacts are widely reported but are difficult to suppress, even with recent efforts like balancing the training data [71]. + +Multimodal Classifier-free Guidance. To mitigate the issue of excessive hallucination often associated with high guidance scales in CFG, we propose a novel multimodal guidance strategy. We argue that the artifacts come from the weak guidance in the negative prompting process. Instead of simply relying on text prompts for guidance, we leverage the rich information encoded in the multimodal latent tokens to strengthen both positive and negative promptings. Specifically, we condition both the postive and negative generation processes on the multimodal latent token sequence, denoted as $m$ . This leads to the following multimodal CFG: + +$$ +\tilde {\boldsymbol {\epsilon}} (\mathbf {z} _ {t}, c, m) = (1 + w) \boldsymbol {\epsilon} (\mathbf {z} _ {t}, c, \text {p o s}, m) - w \boldsymbol {\epsilon} (\mathbf {z} _ {t}, c, \text {n e g}, m). \tag {4} +$$ + +Sec. 4.1 shows that this change in negative generation helps to achieve a better trade-off between perceptual quality and identity preservation, better maintaining the semantic content of the low-resolution input in the upscaled output, compared with the standard CFG in previous text-based methods. + +Scaling Single-modal Guidance. Multimodal CFG effectively controls the overall influence of the prompts but does not offer control over each modality individually. To address this limitation, we introduce a mechanism to selectively amplify or suppress the contribution of specific modalities. Specifically, we modify the attention temperature $\delta$ of MMLC during cross-conditioning, where $\delta$ scales the attention maps before applying the softmax operation: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\delta}\right) V. \tag {5} +$$ + +This parameter controls the sensitivity of the attention mechanism to differences between the feature sequence $Q$ and the multimodal token sequences $K$ and $V$ . A smaller temperature $\delta$ typically leads to a stronger conditioning effect on the attention mechanism. In the standard scaled dot-product attention, the temperature $\delta$ is empirically set to the square + +root of the key dimension $\sqrt{d_k}$ . We observe that scaling the standard temperature $\sqrt{d_k}$ (where $d_{k}$ is the key dimension) within a range of [0.4, 10] can produce high-quality results, with varying levels of fidelity to the input modalities. By scaling the temperature during sampling, we achieve precise control over the super-resolution process, enabling manageable manipulation of the output with fine-grained control. + +# 3.3. MMSR Implementation + +Figure 3 illustrates our multimodal guided super-resolution pipeline. During inference, we first extract four modalities from the low-resolution (LR) image: a text caption generated by Gemini Flash [56], a depth map estimated by Depth Anything [66], a semantic segmentation mask produced by Mask2Former [11], and edge information extracted with a Canny edge detector. The depth map, segmentation mask, and edge information are encoded into token sequences using a pretrained VQGAN, while the text caption is processed by a pretrained CLIP encoder to obtain a text embedding. + +Following a similar conditioning strategy to InstructPix2Pix [6], the LR image is concatenated with a noisy latent vector sampled from the diffusion model, providing additional conditioning. + +# 4. Experiments + +Training Details. Our super-resolution model is initialized with the weights of a pretrained text-to-image model, with the same architecture and size as Stable Diffusion v2 [53]. The super-resolution training dataset consists of randomly degraded (using RealESRGAN degradation [63]) high-resolution images from the combined LSDIR and DIV2K datasets, with corresponding $512 \times 512$ high-resolution images as ground truth. We set the batch size to 1024 and the learning rate to 1e-4, which we empirically found maximized compute efficiency on TPUv5 pods. During testing, all benchmarks use a model checkpoint finetuned for 160k iterations. We use 50-step DDIM sampling, consistent with previous methods. We use a guidance rate of 4 as the default. + +![](images/dcea7f995f1d0693ee265c99b3e67182fb67e9be509651809d332ed6a681da02.jpg) +LR + +![](images/b2201a59f269c441e560bdc563e45adaba72cc1d1a9e2c2487d2b548c9344e1d.jpg) + +![](images/9d3e1a10217413595a67c657828064a0813cdf6d54f1f4b3a8d5e4b1fffe1473.jpg) + +![](images/9082679ea62fd50d13c372b0a732deecf47e369e1f6a7fc940d5401c2d5c58c8.jpg) + +![](images/30601950487f29035e5a733103e139de74abf7cc3ca9ecfcd45f0f0675c2c4c8.jpg) +SeeSR + +![](images/d0727b30a5320d57fb4837559716ee716c86593f2b70db06aff85e81349a4be6.jpg) +MMSR (Ours) + +![](images/f2dc4043b5d4560bff38ed12e3939838b5219334924ca5638deab2e07140ef09.jpg) + +![](images/314ca8417f6cf02d36615c6c51eaffc24ff25179e5e80a9bddffdbd22e5ad915.jpg) +LR (Zoomed) + +![](images/e08e88b3aed45a685a8207378e87b2cd3b40555df28cd538535a690a724c1ab0.jpg) +Segmentation + +![](images/f6eec81efe8334cc29d08fab95388c76579c21ccfc5b1949eee8e7a884ffeccd.jpg) +Edge + +![](images/9431c23601e9346875331e0173db4a29fe1e2230252b61d94d312ab895255693.jpg) +PASD (Zoomed) + +![](images/b27a4e98c9c4c1f065417b6eb2c34006071329e85877897a74d1fa23e22351bb.jpg) +SeeSR (Zoomed) + +![](images/dadc586c4b6dfd2ecff2efbd796b7b4f6b91485a73f4a815915d83f083606e5a.jpg) +MMSR (Zoomed) + +![](images/20ecb5709aa00efd3f458ea02d8003c04b4c1f8d0b14696c674c34b7aadb74aa.jpg) +HR (Zoomed) + +![](images/5ac961b0687907b4b5c6f3badd1135758699326bf84bdd42424fa845c7a31af8.jpg) + +![](images/7a01931086f2a99efa50eb5055bee659c22964f2c021acdcb88d17729a368260.jpg) + +![](images/18fd6a59706527cf34e84f4ed1c87763d37f1bd38d15442ed10e4fe157afab09.jpg) +LR +LR (Zoomed) +Figure 6. Super-resolution results on common benchmarks, comparing MMSR with state-of-the-art techniques. Zoom in for detail. + +![](images/c0f79b5b70da690e3b618e85314c809136e3ef9d713f429faa6a42d2faf4d66c.jpg) +Caption +Segmentation + +![](images/0232628f8066fc3e8940a5bd13aebe43ae8e8f74f9fb7f70d2457af568c7fef0.jpg) +Depth + +![](images/1305a8bd471d9adc23ac4117102a82cc0c1d2fe5045a525f04806ac30952251c.jpg) +Edge + +![](images/26fa23a053fe8ce905d48a7ea7472a7491062891c556dadeb1991ec48acc6ca2.jpg) +PASD + +![](images/0de75abeba3895053529084962b17030e1fd04042b939bdc0f8b0f449392b2f0.jpg) +PASD (Zoomed) + +![](images/e8550f03dcfca4951c439bb8d26ab1eb4188baf0db4e442fb3b6fe018bcb430e.jpg) +SeeSR + +![](images/bc5b05e6cb23d8f6e80e9f93a594fffbf55c5c41182f3cc5f73a3d0b1d67ed89.jpg) +SeeSR (Zoomed) + +![](images/2d7a388792b499d21119f39ab174717ae6f7606991b8d5f0b677236e6f9ef594.jpg) +MMSR (Ours) + +![](images/4f0fb3998679a2fcdaf15d33903d67b7a2f6555ad7d3a71100d4cdf76feb2ddc.jpg) +MMSR (Zoomed) + +![](images/248b4414d2e211f23b8c182caebea5783d57b9ec4f14bc34cb59cf96cdd34a30.jpg) +HR + +![](images/d22995bef713bb026ddaec73958546b868c45cd7f738edb0f7e789b982d62076.jpg) +HR (Zoomed) + +Evaluation Details. We use reference-based metrics like LPIPS [75] and DISTS [16], and non-reference-based metrics like NIQE [42], MANIQA [67], MUSIQ [27], and CLIPIQA [58], as these have been shown to align well with humans' aesthetic preferences. Our compared baselines include recent diffusion-based super-resolution methods [33, 53, 59, 65, 68, 72] and the representative methods [8, 31, 32, 63, 73]. We collected their results either by running their official code or from published results. In a few cases where the multimodal prediction failed (two flat images and one pencil drawing from DRealSR), we manually replaced the prediction with $m_{\emptyset}$ (see supplemental material). + +# 4.1. Quantitative Results + +Table 1 shows the quantitative result comparisons of ours with the other baselines, compared on the synthetic benchmark and real-world super-resolution benchmarks. Several facts are worth noting: (1) Our method achieves the best LPIPS, DISTS, NIQE, and FID scores on the DIV2K-Val and RealSR dataset, significantly outperforming previous state-of-the-art methods. This superiority demonstrates that our method can generate more perceptually identical details from the guidance of multimodal context. (2) Our method also achieves the best performance in the non-reference visual quality metrics including NIQE, MANIQA, CLIPIQA, and MUSIQ scores. This advantage demonstrates our method can fully utilize the photorealistic priors encapsulated in the pretrained diffusion model. Overall, these observations fully demonstrate our advantages in using multimodal guidance compared with the past single-modality methods. + +Effects of Multimodal CFG Guidance. Table 3 demonstrates the effects of our proposed multimodal CFG on the 1MP DIV2K-Val benchmark, which replaces the negative score guided by empty language token $\epsilon (\mathbf{z}_t,c,\mathrm{neg})$ , shortened as $cfg$ , with multimodal guided negative score $(\epsilon (\mathbf{z}_t,c,m,\mathrm{neg}))$ , shortened as $m$ -cfg. In order to exclude the influence of architectural differences, we also compare our method with the empty multimodal token $\epsilon (\mathbf{z}_t,c,m_{\emptyset},\mathrm{neg})$ , shortened as $m_{\emptyset}$ -cfg. Note that the positive score for all three compared methods is the same $\epsilon (\mathbf{z}_t,c,m,\mathrm{pos})$ . The results show that our method mitigates the degradation reflected in the LPIPS and NIQE scores when using guidance rates of 10 and 14 in both CFG and $m_{\emptyset}$ -cfg, while maintaining comparable performance. The visual results in Table 3 further demonstrate our superiority, where ours suppresses the color change and incorrect texture of high CFG guidance rate, leading to higher CLIPIQA score. + +Effects of Latent Connector. To demonstrate the effectiveness of the proposed multimodal latent connector, we conducted an ablation study by training a new multimodal super-resolution model without the latent connector, which directly uses the longer multimodal token sequence as input. The quantitative performance is shown in Table 2. Our $w$ . MMLC outperforms the $w/o$ . MMLC in all reference and non-reference metrics. The visual results in Figure 7 show that the model without the MMLC is more likely to exacerbate hallucination. Its result shows bokeh on the branches but a clear trunk, even though the depth input clearly shows they are at the same depth. + +Contributions of Each Modality. Figure 8 shows the con + +DIV2K-Val-3k 512 × 512 + +
MethodsPSNRSSIMLPIPS ↓DISTS ↓NIQE ↓FID ↓MUSIQCLIPQA
BSRGAN21.870.55390.41360.27374.761564.2859.110.5183
R-ESRGAN21.940.57360.38680.26014.920953.4658.640.5424
LDL21.520.56900.39950.26885.024958.9457.900.5313
DASR21.720.55360.42660.26884.859667.2254.220.5241
FeMASR20.850.51630.39730.24284.572653.7058.100.5597
LDM21.260.52390.41540.25006.466741.9356.520.5695
StableSR20.840.48870.40550.25424.655136.5762.950.6486
ResShift21.750.54220.42840.26066.973155.7758.230.5948
PASD20.770.49580.44100.25384.832840.7766.850.6799
DiffBIR20.940.49380.42700.24714.721140.4265.230.6664
SeeSR21.190.53860.38430.22574.927531.9368.330.6946
MMSR21.740.56930.37070.20714.253229.3570.060.7164
+ +DIV2K-Val-100 1024 × 1024 + +
MethodsPSNRSSIMLPIPS ↓DISTS ↓NIQE ↓MANIQAMUSIQCLIPiQA
R-ESRGAN21.770.58130.36240.19903.65730.404647.540.5358
StableSR20.070.39470.50970.24273.62600.411365.390.6938
PASD21.270.53690.34730.17533.63210.470869.690.6914
SUPIR20.650.53500.38490.18143.64580.405165.880.5697
SeeSR21.310.55780.32730.16204.02150.543969.790.6941
MMSR21.870.55650.28100.14923.42430.488572.310.7294
+ +RealSR $512\times 512$ + +
MethodsPSNRSSIMLPIPS ↓LIQENIMAMANIQAMUSIQCLIPiQA
R-ESRGAN25.690.76160.27273.35744.65480.548760.180.4449
StableSR24.700.70850.30183.61064.81500.622165.780.6178
PASD24.290.66300.34353.57494.85540.649368.690.6590
SUPIR22.970.62980.37503.56824.57570.574561.490.6434
SeeSR25.180.72160.30094.13604.91930.644269.770.6612
MMSR24.830.70030.29524.34685.10940.657871.330.6717
+ +DrealSR $512\times 512$ + +
MethodsPSNRSSIMLPIPS ↓LIQENIMAMANIQAMUSIQCLIPiQA
R-ESRGAN28.640.80530.28472.92554.32580.490754.180.4422
StableSR26.710.72240.32843.24254.48610.559458.510.6357
PASD27.000.70840.39313.59084.66180.585064.810.6773
SUPIR24.610.61230.42943.47104.38150.538157.320.6758
SeeSR26.750.74050.31744.12704.69420.605265.090.6908
MMSR27.280.74560.32494.50235.05580.630168.930.6999
+ +Table 1. Severely degraded, low-resolution (LR) images can produce inaccurate multi-modal information, manifesting as distorted edges, misidentified objects, and other artifacts. + +
MUSIQNIQE ↓DISTS ↓LPIPS ↓Throughput
w/o. MMLC69.693.48450.17810.39293.48 img/s
w. MMLC72.313.42430.14920.28103.32 img/s
+ +tributions of each modality by masking out input modalities during testing. Benefiting from the improved flexibility of using multimodal token $m_{\emptyset}$ , discussed in Sec. 3.1, our method is robust in scenarios with only fewer modalities input during testing. We would like to note several observations: (1) Our default multimodal setting achieves the best trade-off between the non-reference metric MUSIQ and the reference-based metric DISTS. (2) Depth information primarily enhances perceptual quality (MUSIQ), while other modalities + +Table 2. Ablation of the impact of the Multimodal Latent Connector module (MMLC) on DIV2K-Val-100 1024p. + +
guidance21014
cfg0.32390.44910.5064
m0cfg0.28150.48030.5493
mcfg0.28100.34710.3772
+ +![](images/471c9862b8304157c561b833c29b15e6089a7e5a5c408dedcab0796b0df51e6b.jpg) +Segment, Depth, LR, and Edge Input +(c) Visual comaprisons + +(a) LPIPS score + +
guidance21014
cfg3.5774.61795.1886
m0cfg3.62615.10815.9175
mcfg3.46793.74193.9815
+ +(b) NIQE score + +![](images/bccfcd896c218f8402f694cedaf30180c9c27eca1780b17e5b6e09f80050ba58.jpg) + +![](images/2fb16ce560d277d9ac9b41456fffe629f39327c38a80ec517e4f01c701203c6f.jpg) +$cfg = 2.0$ CIPIQA $= 0.4481$ +$cfg = 14.0$ CLIPIQA=0.7091 + +![](images/708285030746072e83429d216889c81bdb5b81d53200da6954bf84f9cbc778cd.jpg) +$m\text{-} cfg = 2.0$ CLIPIQA $= 0.4122$ + +![](images/d3e3ac5369266df04fbfa8f1c4a2e899b27cbd942094bacd84651fda47de91f7.jpg) +$m\text{-} cfg = 14.0$ CIPIQA $= 0.7352$ + +![](images/3b3fd4014db8a5eed536a6d3664f059a3b4d5b8b072239c3d026eebe95122258.jpg) +Table 3. Ablation of guidance on DIV2K-Val-100 1024p. Our multimodal CFG ( $m$ -cfg) mitigates artifacts often present when using high guidance rates. This leads to an improved balance between visual fidelity and preservation of key identifying features. +LR Input + +![](images/df974042424819b6ad44cf618c72580032d8eb49c9ef08783297b2ec2dcd6bbc.jpg) + +![](images/5f0106977ab5347565c3243ae29405907dea68c6754fc3e9609da218cc954142.jpg) +w/o. MMLC Result + +![](images/309df6ac578aebc7bfeb3d9dec1b1a33e32eed7d1583cc359e332b6824252d03.jpg) +Depth Input +w. MMLC Result + +![](images/4c8ec3f47a4344137af101359c05d1c3e92b3928e330b4235acd7a6487d4878f.jpg) +Figure 7. Visual results when (not) using the MMLC module. +Figure 8. Ablation of modalities on DIV2K-Val-100 1024p. Incorporating different modalities enhances various aspects of the super-resolution results compared to the text-guided baseline. + +![](images/888c95ca94a1bed5fd4b8d9bfd6c82278c314b5ff5e46703b21d5ca3a658b13b.jpg) + +contribute more significantly to preserving identity (DISTS). These results highlight the diverse contributions of each modality and emphasize the advantage of our multimodal approach in effectively combining their strengths for optimal super-resolution performance. + +# 4.2. Qualitative Results + +Figures 5 and 6 provide a visual comparison of our method against state-of-the-art approaches, namely SeeSR [65], PASD [68], and SUPIR [71]. Our method demonstrates several key advantages. Enhanced Realism: In the first example of Figure 6, our method produces fewer artifacts and remains more faithful to the high-resolution image compared to other methods. This highlights our ability to maintain realism without introducing spurious details. Robustness to + +![](images/49fae6f99adc81fda8523d7826166484f2a755733a2bff7320b497d1b41a8d87.jpg) +(a). Results of PASD + +![](images/15eae51b990b4a0e7675878d8ff7068bccd05013eec4bdc650e1631fdd4ce5cc.jpg) + +![](images/d7096bc195c9c7475167467d3aeac064e261c4d18d7e9020afcfe39a81eb67cc.jpg) +Figure 9. Our method allows for fine-grained control over super-resolution results by adjusting the influence of each input modality. For example, reducing the edge temperature enhances edge sharpness (first row). Lowering the segmentation temperature emphasizes distinct features, such as the star pattern on the flag (second row). Decreasing the depth temperature accentuates depth-of-field effects, like the bokeh between the foreground and background (third row). In contrast, PASD [68] exhibits limited control over such fine-grained details. + +![](images/f1ebf466d0fd63f3a8eba5df203f90bbb9347e8277e28f420eef76d2169eb899.jpg) + +![](images/83f73c6f98f89b437c7cd845654d1b1d76a1817e26aacf7b5dca378364529fa0.jpg) + +![](images/8a156ad46d9cc4447636837d62f43b894a65c2f72ec2fd21182ac6f619aa9631.jpg) + +![](images/03ebe657ad6d0f362bbbd98d85061116b860a8606687cd6cd31f8622e9c6df9a.jpg) +(b). Results of MMSR under different modality-specific controls + +![](images/7f23dbfb7c3bf62e05058b955e276d1e727aa1abddf451f78a4671f1bea85c32.jpg) + +![](images/3cd652ef8a05a0e179a7f609e7a39646cc403a280653dcd404923b8d1bb84cf9.jpg) + +Challenging Conditions: The second example in Figure 6, a long-range shot affected by noise and turbulence, shows our method's ability to generate clear human details where other methods struggle. This is attributed to the effective use of semantic segmentation for accurate human localization, further enriched by the textual caption. These examples illustrate the effectiveness of our multimodal approach in generating high-quality images across diverse scenarios. See supplemental material for additional results and analysis. + +# 4.3. Controllability Comparisons + +While recent works leverage text prompts for controlling super-resolution [65, 68, 71], these often lack fine-grained control and can yield inconsistent results. Our multimodal approach introduces modality-specific temperature weights to scale attention scores, enabling precise manipulation of SR outputs by amplifying or diminishing the influence of each modality (e.g., depth, segmentation). Figure 9 contrasts our approach with PASD [68], which relies solely on text prompts. Specifically, decreasing the temperature of the edge modality changes the richness of details in our result, while changing PASD's prompt with "more edge" doesn't change their result. Decreasing the temperature of semantic segmentation map and depth map also lead to manageable changes such as more visible star pattern and stronger bokeh. In contrast, directly add corresponding prompt in PASD's prompt barely changes its result. Moreover, our approach exhibits smooth transitions in image characteristics as temperatures are adjusted, providing interpretable control and insights into + +the role of each modality. + +# 5. Conclusion + +This work introduces a novel diffusion-based framework for image super-resolution that seamlessly integrates diverse modalities—including text descriptions, depth maps, edges, and segmentation maps—within a single, unified model. By leveraging pretrained text-to-image models, our method achieves enhanced realism and accurate reconstruction, outperforming existing text-guided super-resolution methods both qualitatively and quantitatively. Furthermore, a learnable multimodal token and modality-specific contribution controls provide fine-grained control over the super-resolution process, enabling adjustable perception-distortion tradeoffs and robust performance even with imperfect or missing modalities in most cases. + +Limitations and Future Work. While adding multimodal information significantly enhances SR performance, it introduces computational overhead. For instance, using Gemini Flash for image captioning results in a throughput of 0.34 images per second, which is slower than depth at $1.99\mathrm{img / s}$ , semantic segmentation at $2.09\mathrm{img / s}$ , and DDIM sampling at $0.54\mathrm{img / s}$ . However, cross-modal predictions can be parallelized, and our method achieves speeds comparable to other text-driven SR models [68, 71]. Future work will explore optimizing the vision-language component for faster inference and investigate more robust modules for extracting modality-specific information, potentially enhancing performance even with noisy or incomplete inputs. + +# 6. Acknowledgments + +Vishal M. Patel was supported by NSF CAREER award 2045489. We are grateful to Keren Ye and Mo Zhou for their valuable feedback. We also extend our gratitude to Shlomi Fruchter, Kevin Murphy, Mohammad Babaeizadeh, and Han Zhang for their instrumental contributions in facilitating the initial implementation of the latent diffusion models. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2 +[2] Roman Bachmann, Oğuzhan Fatih Kar, David Mizrahi, Ali Garjani, Mingfei Gao, David Griffiths, Jiaming Hu, Afshin Dehghan, and Amir Zamir. 4M-21: An any-to-any vision model for tens of tasks and modalities. NeurIPS, 2024. 2, 3 +[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, pages 22861-22872, 2024. 3 +[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2023.1, 2 +[5] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In CVPR, pages 6228-6237, 2018. 1 +[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 2, 5 +[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, pages 14455-14465, 2024. 1, 2 +[8] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Real-world blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 6 +[9] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25857-25867, 2024. 2 +[10] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1 +[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 5 + +[12] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. arXiv preprint arXiv:2310.01110, 2023. 1 +[13] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. In *Forty-first International Conference on Machine Learning*, 2024. 2 +[14] Maurizio Delbracio and Peyman Milanfar. Inversion by direct iteration: An alternative to denoising diffusion for image restoration. Transactions on Machine Learning Research, 2023. Featured Certification. 1 +[15] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, 2021. 4 +[16] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In Proceedings of the 29th ACM International Conference on multimedia, pages 2483-2491, 2021. 6 +[17] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 2015. 1 +[18] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3 +[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2 +[20] Kanchana Vaishnavi Gandikota and Paramanand Chandramouli. Text-guided explorable image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25900-25911, 2024. 2 +[21] Jiatao Gu, Ying Shen, Shuangfei Zhai, Yizhe Zhang, Navdeep Jaitly, and Joshua M Susskind. Kaleido diffusion: Improving conditional diffusion models with autoregressive latent modeling. Advances in neural information processing systems, 2024. 2 +[22] Jiatao Gu, Yuyang Wang, Yizhe Zhang, Qihang Zhang, Dinghuai Zhang, Navdeep Jaitly, Josh Susskind, and Shuangfei Zhai. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024. 2 +[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 4 +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3 +[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General + +perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 4 +[26] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's "up" with vision-language models? investigating their struggle with spatial reasoning. In EMNLP, pages 9161–9175, 2023. 1 +[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6 +[28] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning, pages 3744-3753. PMLR, 2019. 4 +[29] Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image super-resolution. In Proceedings of the European conference on computer vision (ECCV), pages 517-532, 2018. 1 +[30] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2 +[31] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 6 +[32] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision. Springer, 2022. 6 +[33] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diffbir: Towards blind image restoration with generative diffusion prior. ArXiv preprint, 2023. 2, 6 +[34] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 1, 2 +[35] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 1 +[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2 +[37] Kangfu Mei, Vishal M Patel, and Rui Huang. Deep semantic statistics matching (d2sm) denoising network. In European Conference on Computer Vision, pages 384-400. Springer, 2022. 1 +[38] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image generation. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 9048-9058, 2024. 2 +[39] Kangfu Mei, Luis Figueroa, Zhe Lin, Zhihong Ding, Scott Cohen, and Vishal M Patel. Latent feature-guided diffusion models for shadow removal. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4313-4322, 2024. 2 +[40] Kangfu Mei, Zhengzhong Tu, Mauricio Delbracio, Hossein Talebi, Vishal M Patel, and Peyman Milanfar. Bigger is not always better: Scaling properties of latent diffusion models. arXiv preprint arXiv:2404.01367, 2024. 2 +[41] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. +[42] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6 +[43] David Mizrahi, Roman Bachmann, Oğuzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4M: Massively multimodal masked modeling. NeurIPS, 2023. 2, 3 +[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3 +[45] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4 +[46] Mehdi Noroozi, Isma Hadji, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. You only need one step: Fast super-resolution with stable diffusion via scale distillation. In European Conference on Computer Vision, 2024. 2 +[47] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 4 +[48] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2 +[49] Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi. Spire: Semantic prompt-driven image restoration. In European Conference on Computer Vision, 2024. 2 +[50] Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou. Xpsr: Cross-modal priors for diffusion-based image super-resolution. In European Conference on Computer Vision, pages 285-303. Springer, 2024. 1, 2 +[51] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. + +Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1 +[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 4 +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 2, 5, 6 +[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[55] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 4 +[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 5 +[57] Li-Yuan Tsao, Yi-Chen Lo, Chia-Che Chang, Hao-Wei Chen, Roy Tseng, Chien Feng, and Chun-Yi Lee. Boosting flow-based generative super-resolution models via learned prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26005–26015, 2024. 2 +[58] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6 +[59] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 2, 4, 6 +[60] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 4 +[61] Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, and Shanshan Zhao. Caption anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677, 2023. 2 +[62] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 1 +[63] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Realesrgan: Training real-world blind super-resolution with pure synthetic data supplementary material. Computer Vision Foundation open access, 2022. 5, 6 +[64] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C + +Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2 +[65] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 1, 2, 4, 6, 7, 8 +[66] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 5 +[67] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6 +[68] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, 2024. 1, 2, 6, 7, 8 +[69] Zhixiong Yang, Jingyuan Xia, Shengxi Li, Xinghua Huang, Shuanghui Zhang, Zhen Liu, Yaowen Fu, and Yongxiang Liu. A dynamic kernel prior model for unsupervised blind image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26046-26056, 2024. 2 +[70] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3 +[71] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 1, 4, 5, 7, 8 +[72] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super-resolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 2, 6 +[73] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 6 +[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3 +[75] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 + +[76] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. In European Conference on Computer Vision, 2024. 2 +[77] Zheyuan Zhang, Fengyuan Hu, Jayjun Lee, Freda Shi, Parisa Kordjamshidi, Joyce Chai, and Ziqiao Ma. Do vision-language models represent space and how? evaluating spatial frame of reference under ambiguities, 2024. 1 \ No newline at end of file diff --git a/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/images.zip b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5c251b8bcb218f204c3138c745f08a2d4f9d250b --- /dev/null +++ b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cc0f983c1368cb095cd95777670e6561b16868882b4c09f4cda5d006c7c6f32 +size 856465 diff --git a/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/layout.json b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..15a31e5b06d0e2b40bcd7968bc5d57f1b4be437b --- /dev/null +++ b/2025/The Power of Context_ How Multimodality Improves Image Super-Resolution/layout.json @@ -0,0 +1,13135 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 91, + 545, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 91, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 65, + 91, + 545, + 110 + ], + "type": "text", + "content": "The Power of Context: How Multimodality Improves Image Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 130, + 444, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 130, + 444, + 188 + ], + "spans": [ + { + "bbox": [ + 164, + 130, + 444, + 188 + ], + "type": "text", + "content": "Kangfu Mei\\*1,2, Hossein Talebi', Mojtaba Ardakani', Vishal M. Patel?, Peyman Milanfar', Mauricio Delbracio' 1 Google, 2 Johns Hopkins University Project Page: https://mmsr.kfmei.com/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 193, + 565, + 325 + ], + "blocks": [ + { + "bbox": [ + 58, + 193, + 565, + 325 + ], + "lines": [ + { + "bbox": [ + 58, + 193, + 565, + 325 + ], + "spans": [ + { + "bbox": [ + 58, + 193, + 565, + 325 + ], + "type": "image", + "image_path": "1ddbe435e00ca0c7626f7fa2886ba73af9bd9e8569079aa84632bc6bfe743901.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 555, + 361 + ], + "type": "text", + "content": "Figure 1. Our Multimodal Super-Resolution (MMSR) method leverages the rich context of multimodal guidance, including image captions, depth maps, semantic segmentation maps, and edges inferred from LR. MMSR surpasses state-of-the-art methods by producing more realistic results and suppressing artifacts that, while plausible, are inconsistent with the information present in the LR input." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 373, + 194, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 373, + 194, + 384 + ], + "spans": [ + { + "bbox": [ + 157, + 373, + 194, + 384 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 390, + 277, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 390, + 277, + 643 + ], + "spans": [ + { + "bbox": [ + 75, + 390, + 277, + 643 + ], + "type": "text", + "content": "Single-image super-resolution (SISR) remains challenging due to the inherent difficulty of recovering fine-grained details and preserving perceptual quality from low-resolution inputs. Existing methods often rely on limited image priors, leading to suboptimal results. We propose a novel approach that leverages the rich contextual information available in multiple modalities – including depth, segmentation, edges, and text prompts—to learn a powerful generative prior for SISR within a diffusion model framework. We introduce a flexible network architecture that effectively fuses multimodal information, accommodating an arbitrary number of input modalities without requiring significant modifications to the diffusion process. Crucially, we mitigate hallucinations, often introduced by text prompts, by using spatial information from other modalities to guide regional text-based conditioning. Each modality's guidance strength can also be controlled independently, allowing steering outputs toward different directions, such as increasing bokeh through depth or adjusting object prominence via segmentation. Extensive experiments demonstrate that our model surpasses state-of-the-art generative SISR methods, achieving superior visual quality and fidelity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 297, + 696 + ], + "type": "text", + "content": "Single image super-resolution (SISR) aims to generate high-resolution images from low-resolution inputs while preserving" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 612 + ], + "type": "text", + "content": "ing semantic identity and texture details. While it is not essential to treat SISR as a regression problem, past methods [17, 29, 37, 62] typically use a deep neural network to learn a direct mapping from low-resolution images to high-resolution images. Even though these regression-based methods achieve good scores on paired metrics like PSNR and SSIM, they have failed to produce results with high quality comparable to natural images, a task at which recent generative models have excelled [5, 12, 14]. The advent of powerful generative models, such as autoregressive models [10, 51] and diffusion models [47, 53, 54], has revolutionized image generation tasks, including text-to-image synthesis. This has inspired recent efforts to leverage these pre-trained generative models for downstream tasks like SISR [6, 41, 53]. For instance, very recent works [50, 71] achieve super-resolution by leveraging emerging vision-language models (e.g., Gemini [56], LLaVA [36], ChatGPT-4 [1]) and pretrained text-to-image models to first generate captions from low-resolution images and then use these captions as prompts to generate high-resolution images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "While providing rich textual descriptions can significantly enhance the quality of generated images [4, 19, 34], relying solely on text for SISR poses challenges. Recent works [7, 26, 35, 77] have shown that text prompts cannot represent spatial relationships. This implies that textual information, such as texture descriptions, can only be applied to the whole image. Figure 1 provides a representative example. Previous text-based super-resolution methods [65, 68] use a 'lion'" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 238, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 238, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 238, + 713 + ], + "type": "text", + "content": "*This work was done during an internship at Google." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23141" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "caption, which results in a furry tongue. However, while lions have fur, their tongues do not grow hair. Can we leverage spatial cues from depth and segmentation data to improve the learned prior and enhance the quality of SISR?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 295, + 264 + ], + "type": "text", + "content": "Fortunately, we can directly extract additional modalities from the low-resolution image by using various pretrained cross-modal prediction models [11, 66]. In this paper, we introduce a new diffusion model architecture dedicated to multimodal super-resolution, which is conceptually illustrated in Figure 2. By integrating multiple modalities into a single diffusion model, our method overcomes the challenges of recovering fine-grained details and preserving perceptual quality. Specifically, we propose conditioning diffusion models on modalities including text captions, semantic segmentation maps, depth maps, and edges to implicitly align text captions for correctly prompting different regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 361 + ], + "type": "text", + "content": "We demonstrate the proposed multimodal diffusion model on the SISR task. Our effective multimodal architecture achieves better realism in SISR results than the best text-driven method and largely eliminates hallucinations that do not match the input. Moreover, we find that the multimodal SISR enables a new controlling feature, where we can explicitly adjust the weights of each modality to steer the generated results in different directions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 361, + 253, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 361, + 253, + 373 + ], + "spans": [ + { + "bbox": [ + 67, + 361, + 253, + 373 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 377, + 296, + 545 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 377, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 426 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of token-wise encoding for seamlessly injecting multiple modalities into pretrained text-to-image diffusion models without architectural modifications or significant model size overhead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 426, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 472 + ], + "type": "text", + "content": "- We propose a novel multimodal latent connector that efficiently fuses information from different modalities, maintaining linear time complexity with respect to the number of modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 473, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 473, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 473, + 295, + 509 + ], + "type": "text", + "content": "- We introduce a new multimodal classifier-free guidance technique that enhances realism at higher guidance rates while mitigating excessive hallucinations and fake details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 509, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 295, + 545 + ], + "type": "text", + "content": "- Our method enables adjustment of the influence of each modality, allowing for fine-grained manipulation of SISR results while preserving realism and quality." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 561, + 142, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 142, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 142, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Generative Prior Powered Super-resolution. Recent advances in image restoration leverage the power of foundational models to enhance the quality of degraded images [9, 13, 20, 33, 38, 39, 46, 50, 57, 64, 69, 72, 76]. This trend is fueled by the capacity of pretrained generative models to capture natural image statistics and transfer this knowledge to the super-resolution task, enabling photorealistic image generation. Early works in this domain include LDM-SR [53] and StableSR [59] for single image superresolution. Beyond this, methods like InstructPix2Pix [6] utilize instructions for image editing, while ControlNet [74]" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 553, + 149 + ], + "type": "image", + "image_path": "25b71099370a5a1779372762ce898e29b44e412e97821444b001b59b9f22fb81.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 154, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 251 + ], + "type": "text", + "content": "Figure 2. Language models struggle to accurately represent spatial information, leading to coarse and imprecise image superresolution. To overcome this limitation, we incorporate additional spatial modalities like depth maps and semantic segmentation maps. These modalities provide detailed spatial context, allowing our model to implicitly align language descriptions with individual pixels through a transformer network. This enriched understanding of the image significantly enhances the realism of our super-resolution results and minimizes distortion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 271, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 319 + ], + "type": "text", + "content": "and IP-Adapter [70] facilitate cross-modality image translation. These approaches highlight a common observation: the quality of results in downstream tasks is strongly correlated with the quality of the pretrained models [40]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 321, + 556, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 556, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 556, + 453 + ], + "type": "text", + "content": "Recent works focus on leveraging powerful text-to-image diffusion models, employing text prompts to fully exploit the learned prior. PASD [68] uses image content descriptions; SPIRE [49] and PromptIP [48] use degradation descriptions; and SeeSR [65] employs a combination of context and degradation descriptions. Compared to earlier methods that directly fine-tuned diffusion models on super-resolution data, these text-prompt-driven approaches not only achieve superior realism but also enable multi-faceted outputs by conditioning on different prompts [20]. This capability enriches the typically single-output super-resolution task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 454, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 556, + 538 + ], + "type": "text", + "content": "While text prompts offer advantages, they can be ambiguous in representing spatial relationships [4, 7, 19, 34]. Our proposed MMSR framework addresses this by integrating multimodal inputs within a novel architecture. We also demonstrate native super-resolution effects control with this new architecture, enabling control of both overall realism and the effect of each modality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Vision-language Understanding and Generation. Recent large vision-language models (e.g., Gemini [56], LLaVA [36], GPT-4o [1]) excel in tasks like image captioning [61], but translating visual information into diverse modalities for image generation remains challenging. While models like 4M [2, 43] effectively extract high-level visual information (depth, normals, semantics, etc.), leveraging these for generation is a promising direction. Recent exploration includes GLIGEN [30] generate images from bounding boxes, and ControlNet [74] generate images from depth and other modalities. More recent works [21, 22] explore using discrete hidden-feature tokens to guide generation. In this paper, we introduce a new approach for multimodal control with discrete vision tokens, effectively integrating depth maps," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23142" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 553, + 228 + ], + "type": "image", + "image_path": "6520790bb98cff0df4b5f817ac4cb7189a2075aa9ecdbcfe5e30fd81417f3ab1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 282 + ], + "type": "text", + "content": "Figure 3. This diagram illustrates our multimodal super-resolution pipeline. Starting with a low-resolution (LR) image, we extract modalities like depth and semantic segmentation maps. These modalities are encoded into tokens and transformed into multimodal latent tokens " + }, + { + "bbox": [ + 55, + 237, + 555, + 282 + ], + "type": "inline_equation", + "content": "(m)" + }, + { + "bbox": [ + 55, + 237, + 555, + 282 + ], + "type": "text", + "content": ". Our diffusion model uses these tokens and the LR input to generate a high-resolution (SR) output. A multimodal classifier-free guidance " + }, + { + "bbox": [ + 55, + 237, + 555, + 282 + ], + "type": "inline_equation", + "content": "(m" + }, + { + "bbox": [ + 55, + 237, + 555, + 282 + ], + "type": "text", + "content": "-cfg) refines the SR image for enhanced quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "type": "text", + "content": "semantic segmentation maps, and text prompts for superior performance. Experiments on real-world super-resolution demonstrate the effectiveness of our approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 111, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 111, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 111, + 352 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "content": "Single-image super-resolution aims to recover a high-resolution image " + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "content": " from its low-resolution counterpart " + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{LR}" + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "content": ". This is an ill-posed problem, often leading to generative models that produce \"hallucinated\" details – plausible yet inconsistent with the input. To mitigate this, we introduce auxiliary information, such as depth " + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "inline_equation", + "content": "(m_{\\mathrm{dep}})" + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "content": ", semantic segmentation " + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "inline_equation", + "content": "(m_{\\mathrm{seg}})" + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "content": ", and edge maps " + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "inline_equation", + "content": "(m_{\\mathrm{edg}})" + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "content": ", collectively denoted as " + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "content": ". By conditioning the generative process on both the low-resolution image and these auxiliary modalities, we propose a new distribution " + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{x}_{LR}, m)" + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "content": " with reduced uncertainty compared to the original distribution " + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{x}_{LR})" + }, + { + "bbox": [ + 54, + 361, + 295, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 494, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 554 + ], + "type": "text", + "content": "This reduction in uncertainty can be understood through the lens of information theory. The auxiliary modality " + }, + { + "bbox": [ + 55, + 494, + 296, + 554 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 494, + 296, + 554 + ], + "type": "text", + "content": " provides additional information about the high-resolution image " + }, + { + "bbox": [ + 55, + 494, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 494, + 296, + 554 + ], + "type": "text", + "content": " that is not present in the low-resolution input " + }, + { + "bbox": [ + 55, + 494, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{LR}" + }, + { + "bbox": [ + 55, + 494, + 296, + 554 + ], + "type": "text", + "content": ". Since (conditional) mutual information is non-negative:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 558, + 294, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 558, + 294, + 571 + ], + "spans": [ + { + "bbox": [ + 64, + 558, + 294, + 571 + ], + "type": "interline_equation", + "content": "I (\\mathbf {x}; m | \\mathbf {x} _ {L R}) = H (p (\\mathbf {x} | \\mathbf {x} _ {L R})) - H (p (\\mathbf {x} | \\mathbf {x} _ {L R}, m)), \\tag {1}", + "image_path": "e8672e14487575a381da1daaf105a1d0ec72c9f2947cfe139cee8de7cd856256.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 574, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 574, + 295, + 597 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 55, + 574, + 295, + 597 + ], + "type": "text", + "content": " denotes entropy. Consequently, the entropy of the conditional distribution with the auxiliary modality is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 601, + 294, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 601, + 294, + 614 + ], + "spans": [ + { + "bbox": [ + 102, + 601, + 294, + 614 + ], + "type": "interline_equation", + "content": "H \\left(p \\left(\\mathbf {x} \\mid \\mathbf {x} _ {L R}\\right)\\right) \\geq H \\left(p \\left(\\mathbf {x} \\mid \\mathbf {x} _ {L R}, m\\right)\\right). \\tag {2}", + "image_path": "6ed066f27021815310a41516238881542babe9146c1ac8afa70d93cbc43c9e75.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "The same motivation for reducing uncertainty is also shared by recent diffusion guidance works [23, 44]. While low-entropy sampling does not guarantee sharper images, it often leads to outputs with better visual quality compared to standard sampling. Motivated by this observation, we introduce a diffusion model [24] to learn the multimodal distribution " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{x}_{LR},m)" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": ", effectively incorporating auxiliary information to enhance SISR quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 290, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 387 + ], + "type": "text", + "content": "Ideally, the auxiliary modalities should provide information complementary to the low-resolution input. In practice, the auxiliary modalities are derived from the high-resolution image during training, ensuring informative conditioning. While we use modalities derived from the low-resolution input during inference, we demonstrate that this still leads to superior performance, including higher-quality details and improved fidelity to the input." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 539, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 539, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 539, + 412 + ], + "type": "text", + "content": "3.1. Unified Multimodal Diffusion Conditioning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 419, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 622 + ], + "type": "text", + "content": "We introduce a new diffusion network architecture, illustrated in Figure 3, for simultaneously conditioning on multiple modalities. Unlike recent methods like ControlNet [74] and IP-Adapter [70], which duplicate network components for each modality and incur significant computational overhead, our approach leverages a pretrained VQGAN image tokenizer [18]. This allows us to encode diverse modalities into a unified token representation for conditioning the diffusion model, without introducing additional model parameters or modifying the diffusion network itself. These tokens are concatenated with the text prompt embedding and used for cross-attention within the diffusion model. To efficiently process this long token sequence, we introduce a lightweight multimodal connector. This connector employs a dedicated architecture to achieve linear complexity for cross-attention, significantly reducing the computational burden. In what follows we present the implementation details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "Token-wise Multimodal Encoding. While VQGAN [18] has proven effective for cross-modal encoding in image understanding and generation [2, 3, 43], its optimal application for image super-resolution remains an open question. Unlike tasks focused on understanding or generation, superresolution requires tokenization to not only capture semantic information but also preserve pixel-wise details crucial for" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23143" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 77, + 119, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 77, + 119, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 77, + 119, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 77, + 119, + 178 + ], + "type": "image", + "image_path": "778161cd7d44af9759e8652466760e5134190732f4932803dcd9db98f7c50262.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 296, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 204 + ], + "type": "text", + "content": "Figure 4. Using discrete multimodal tokens leads to superior reconstruction of modalities compared to continuous tokens." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 77, + 178, + 177 + ], + "blocks": [ + { + "bbox": [ + 119, + 77, + 178, + 177 + ], + "lines": [ + { + "bbox": [ + 119, + 77, + 178, + 177 + ], + "spans": [ + { + "bbox": [ + 119, + 77, + 178, + 177 + ], + "type": "image", + "image_path": "78c20180567971698edb70f60abe6d3b0781aa0e177ac5af7b515c19e18469e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 179, + 77, + 238, + 177 + ], + "blocks": [ + { + "bbox": [ + 179, + 77, + 238, + 177 + ], + "lines": [ + { + "bbox": [ + 179, + 77, + 238, + 177 + ], + "spans": [ + { + "bbox": [ + 179, + 77, + 238, + 177 + ], + "type": "image", + "image_path": "de0f768b1e15772c2c9fa84ae6c9b57066674ba2bc5c515a541ea8909b7d25d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 77, + 300, + 177 + ], + "blocks": [ + { + "bbox": [ + 242, + 77, + 300, + 177 + ], + "lines": [ + { + "bbox": [ + 242, + 77, + 300, + 177 + ], + "spans": [ + { + "bbox": [ + 242, + 77, + 300, + 177 + ], + "type": "image", + "image_path": "777b723c7a4b9b2cb2ca8e6a86951d08dc2044a8d2821b619e02dc04e9837df9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 323 + ], + "type": "text", + "content": "accurate reconstruction, such as spatial relationships within a depth map. Inspired by recent discussions on using discrete or continuous tokenization for autoregressive image generation [55], we investigate the impact of quantization on multimodal super-resolution. Specifically, we analyze how different quantization strategies affect the reconstruction quality of individual modalities and the final generated image. To this end, we utilize a VQGAN model pre-trained on a large-scale multimodal image dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 323, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 396 + ], + "type": "text", + "content": "Figure 4 compares the reconstruction quality of multi-modal tokens before and after quantization. As the figure illustrates, discrete tokens (post-quantization) better preserve individual modality information, while continuous tokens introduce noticeable artifacts. Consequently, we employ discrete tokens for our super-resolution experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": "Our multimodal encoding process utilizes the encoder and quantizer of the pre-trained VQGAN. Each " + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": " input modality is encoded into a " + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": " multimodal token sequence with a feature dimension of 256. These tokens are then quantized using a codebook of size 1024. To facilitate concatenation with text embeddings, we pad the feature dimension of the multimodal tokens to 1024, resulting in a " + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "inline_equation", + "content": "(256 \\times 3 + 77) \\times 1024" + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": " multimodal token sequence condition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": "Multimodal Latent Connector. While cross-attention provides a flexible mechanism for conditioning diffusion models on multimodal data, its quadratic complexity with respect to the number of condition tokens introduces a significant computational burden. To address this, we introduce the Multimodal Latent Connector (MMLC), inspired by recent advances in efficient attention architectures [25, 28, 60]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": "As illustrated in Figure 3, the MMLC employs a transformer architecture to efficiently process the multimodal token sequence. The transformer receives two inputs: a randomly initialized sequence of learnable latent tokens, and the multimodal input sequence. The output is a token sequence of the same length as the latent token sequence, which serves as conditioning for the diffusion model. Therefore, the diffusion model conditions on fixed-length latent tokens (128 in our experiments), which are significantly shorter than the original multimodal token sequence input. The MMLC distills the essential information from the longer multimodal" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 131 + ], + "type": "text", + "content": "token sequence into the shorter multimodal latent token sequence through cross-attention. Following cross-attention, several self-attention blocks further process the latent token sequence, allowing the model to fully integrate the distilled information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "content": "This approach significantly reduces the computational cost of cross-attention in the diffusion model. Standard self-attention operates on the full multimodal token sequence " + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "inline_equation", + "content": "(K, V \\in \\mathbb{R}^{M \\times D})" + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "content": ", resulting in a time complexity of " + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(M^2)" + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "content": " is the length of the sequence and " + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "content": " is the dimensionality of each token. In contrast, the MMLC uses a cross-attention between the latent token sequence (of size " + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "inline_equation", + "content": "N \\times D" + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "content": ") and the multimodal token sequence which reduces this to " + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(MN)" + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "content": " is the length of the latent sequence and " + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "inline_equation", + "content": "N \\ll M" + }, + { + "bbox": [ + 313, + 133, + 556, + 289 + ], + "type": "text", + "content": ". This linear complexity with respect to the multimodal sequence length enables efficient processing of high-dimensional multimodal data, effectively capturing essential information for super-resolution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 293, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 510 + ], + "type": "text", + "content": "Flexible Multimodal Input. To enhance flexibility and robustness, we enable our method to handle scenarios where certain input modalities are unreliable or unavailable. We adopt a learnable embedding approach inspired by DALL-E 2 [52]. Specifically, we introduce a special learnable token, " + }, + { + "bbox": [ + 313, + 293, + 556, + 510 + ], + "type": "inline_equation", + "content": "m_{\\emptyset}" + }, + { + "bbox": [ + 313, + 293, + 556, + 510 + ], + "type": "text", + "content": ", optimized alongside the diffusion model and MMLC to represent the absence of a modality. During training, we independently randomly replace each modality with 256 " + }, + { + "bbox": [ + 313, + 293, + 556, + 510 + ], + "type": "inline_equation", + "content": "m_{\\emptyset}" + }, + { + "bbox": [ + 313, + 293, + 556, + 510 + ], + "type": "text", + "content": " tokens with a probability of 0.1. This encourages the model to learn robust representations that can effectively handle missing information. During inference, any missing modality is represented by a sequence of 256 " + }, + { + "bbox": [ + 313, + 293, + 556, + 510 + ], + "type": "inline_equation", + "content": "m_{\\emptyset}" + }, + { + "bbox": [ + 313, + 293, + 556, + 510 + ], + "type": "text", + "content": " tokens. This approach allows for flexible multimodal input, enabling the model to generate high-quality images even with limited or no auxiliary modalities. As demonstrated in our experiments, this strategy significantly improves performance when dealing with input that includes fewer modalities than were available during training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 518, + 500, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 500, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 500, + 530 + ], + "type": "text", + "content": "3.2. Multimodal Guidance and Control" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 537, + 556, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 556, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 556, + 622 + ], + "type": "text", + "content": "Building upon the success of guidance techniques [15, 23] in improving sample quality across various image generation tasks [45, 47, 52], recent diffusion-based SISR methods have begun incorporating prompt tuning to enhance superresolution results [59, 65, 71]. These methods improve the result by using negative prompts with Classifier-free Guidance (CFG) [23]. which can be expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 625, + 553, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 625, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 325, + 625, + 553, + 639 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} (\\mathbf {z} _ {t}, c) = (1 + w) \\epsilon (\\mathbf {z} _ {t}, c, \\text {p o s}) - w \\epsilon (\\mathbf {z} _ {t}, c, \\text {n e g}), \\tag {3}", + "image_path": "f6427c5a96935412eb3bc79cf3f796b3523a33e3809b941a378a76d95d3962a0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon} (\\mathbf{z}_t,c)" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " represents the guided denoising process, " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\pmb {\\epsilon}(\\cdot)" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " denotes the diffusion model, " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " denotes the noisy image latent, " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "c = \\{\\mathbf{x}_{\\mathrm{LR}},t,\\dots \\}" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " denotes the conditioning inputs (including the low-resolution image " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{LR}}" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ", timestep " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ", and other parameters), " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " is the guidance scale, and pos and neg are the positive and negative prompts. While increasing " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "w" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23144" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 77, + 128, + 133 + ], + "blocks": [ + { + "bbox": [ + 61, + 77, + 128, + 133 + ], + "lines": [ + { + "bbox": [ + 61, + 77, + 128, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 77, + 128, + 133 + ], + "type": "image", + "image_path": "1060f0d42e54684ab8ed1a13af25cf12e53f001bec4124c2a59cf616fa1af652.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 133, + 99, + 140 + ], + "lines": [ + { + "bbox": [ + 89, + 133, + 99, + 140 + ], + "spans": [ + { + "bbox": [ + 89, + 133, + 99, + 140 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 77, + 175, + 84 + ], + "lines": [ + { + "bbox": [ + 157, + 77, + 175, + 84 + ], + "spans": [ + { + "bbox": [ + 157, + 77, + 175, + 84 + ], + "type": "text", + "content": "Inputs" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 86, + 194, + 131 + ], + "lines": [ + { + "bbox": [ + 138, + 86, + 194, + 131 + ], + "spans": [ + { + "bbox": [ + 138, + 86, + 194, + 131 + ], + "type": "text", + "content": "The image shows the words \"CIPAL PARK\" prominently displayed on a metal structure. Below the main sign are several smaller screens displaying baseball ..." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 140, + 127, + 189 + ], + "blocks": [ + { + "bbox": [ + 61, + 140, + 127, + 189 + ], + "lines": [ + { + "bbox": [ + 61, + 140, + 127, + 189 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 127, + 189 + ], + "type": "image", + "image_path": "4a90369b5705d808dfc40a768960412bd9d15a09df82ff239e2e478506dc03c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 189, + 103, + 196 + ], + "lines": [ + { + "bbox": [ + 86, + 189, + 103, + 196 + ], + "spans": [ + { + "bbox": [ + 86, + 189, + 103, + 196 + ], + "type": "text", + "content": "Depth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 201, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 553, + 213 + ], + "type": "text", + "content": "Figure 5. MMSR super-resolution results on real-world images compared with state-of-the-art methods. Zoom in to appreciate the details." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 140, + 197, + 189 + ], + "blocks": [ + { + "bbox": [ + 154, + 133, + 176, + 140 + ], + "lines": [ + { + "bbox": [ + 154, + 133, + 176, + 140 + ], + "spans": [ + { + "bbox": [ + 154, + 133, + 176, + 140 + ], + "type": "text", + "content": "Caption" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 140, + 197, + 189 + ], + "lines": [ + { + "bbox": [ + 134, + 140, + 197, + 189 + ], + "spans": [ + { + "bbox": [ + 134, + 140, + 197, + 189 + ], + "type": "image", + "image_path": "4f545a9c5c14865e7203b02e458015917361c3541f2e77876d07f61852e4ebaa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 190, + 173, + 196 + ], + "lines": [ + { + "bbox": [ + 158, + 190, + 173, + 196 + ], + "spans": [ + { + "bbox": [ + 158, + 190, + 173, + 196 + ], + "type": "text", + "content": "Edge" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 206, + 84, + 271, + 132 + ], + "blocks": [ + { + "bbox": [ + 206, + 84, + 271, + 132 + ], + "lines": [ + { + "bbox": [ + 206, + 84, + 271, + 132 + ], + "spans": [ + { + "bbox": [ + 206, + 84, + 271, + 132 + ], + "type": "image", + "image_path": "98ac589cd0485996e53efff4146db3c8145996c837b2dcc0901d82d5e7835835.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 206, + 140, + 271, + 188 + ], + "blocks": [ + { + "bbox": [ + 227, + 133, + 246, + 140 + ], + "lines": [ + { + "bbox": [ + 227, + 133, + 246, + 140 + ], + "spans": [ + { + "bbox": [ + 227, + 133, + 246, + 140 + ], + "type": "text", + "content": "Patch 1" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 206, + 140, + 271, + 188 + ], + "lines": [ + { + "bbox": [ + 206, + 140, + 271, + 188 + ], + "spans": [ + { + "bbox": [ + 206, + 140, + 271, + 188 + ], + "type": "image", + "image_path": "60693ef1d446c7837706e7127d90bd34adbd3cc50e5c8ed824aff947a3e5faa3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 189, + 248, + 196 + ], + "lines": [ + { + "bbox": [ + 228, + 189, + 248, + 196 + ], + "spans": [ + { + "bbox": [ + 228, + 189, + 248, + 196 + ], + "type": "text", + "content": "Patch 2" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 276, + 84, + 342, + 132 + ], + "blocks": [ + { + "bbox": [ + 276, + 84, + 342, + 132 + ], + "lines": [ + { + "bbox": [ + 276, + 84, + 342, + 132 + ], + "spans": [ + { + "bbox": [ + 276, + 84, + 342, + 132 + ], + "type": "image", + "image_path": "dd66632d97f0b46f0416f073458b1c016ec0d590cd1b92b0f764a1cec363d657.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 277, + 140, + 342, + 188 + ], + "blocks": [ + { + "bbox": [ + 287, + 133, + 331, + 140 + ], + "lines": [ + { + "bbox": [ + 287, + 133, + 331, + 140 + ], + "spans": [ + { + "bbox": [ + 287, + 133, + 331, + 140 + ], + "type": "text", + "content": "PASD (Zoomed)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 277, + 140, + 342, + 188 + ], + "lines": [ + { + "bbox": [ + 277, + 140, + 342, + 188 + ], + "spans": [ + { + "bbox": [ + 277, + 140, + 342, + 188 + ], + "type": "image", + "image_path": "50814422c0a98a2ecd1e24cbe3b9b3fea6d9b8139b02ed309b50bdaee2490113.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 189, + 330, + 196 + ], + "lines": [ + { + "bbox": [ + 286, + 189, + 330, + 196 + ], + "spans": [ + { + "bbox": [ + 286, + 189, + 330, + 196 + ], + "type": "text", + "content": "PASD (Zoomed)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 347, + 84, + 411, + 132 + ], + "blocks": [ + { + "bbox": [ + 347, + 84, + 411, + 132 + ], + "lines": [ + { + "bbox": [ + 347, + 84, + 411, + 132 + ], + "spans": [ + { + "bbox": [ + 347, + 84, + 411, + 132 + ], + "type": "image", + "image_path": "833ec91aea8fe8e1ecf98e34e60597041ba53bdd973183c3ef4407daed954497.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 347, + 140, + 411, + 188 + ], + "blocks": [ + { + "bbox": [ + 355, + 133, + 400, + 140 + ], + "lines": [ + { + "bbox": [ + 355, + 133, + 400, + 140 + ], + "spans": [ + { + "bbox": [ + 355, + 133, + 400, + 140 + ], + "type": "text", + "content": "SeeSR (Zoomed)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 140, + 411, + 188 + ], + "lines": [ + { + "bbox": [ + 347, + 140, + 411, + 188 + ], + "spans": [ + { + "bbox": [ + 347, + 140, + 411, + 188 + ], + "type": "image", + "image_path": "dc7ac3e89e425fc07566a04ef5f0a8f7d0b5a4132ecf15da3698999e99445db0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 189, + 400, + 196 + ], + "lines": [ + { + "bbox": [ + 355, + 189, + 400, + 196 + ], + "spans": [ + { + "bbox": [ + 355, + 189, + 400, + 196 + ], + "type": "text", + "content": "SeeSR (Zoomed)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 417, + 84, + 481, + 132 + ], + "blocks": [ + { + "bbox": [ + 417, + 77, + 424, + 83 + ], + "lines": [ + { + "bbox": [ + 417, + 77, + 424, + 83 + ], + "spans": [ + { + "bbox": [ + 417, + 77, + 424, + 83 + ], + "type": "text", + "content": "uts" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 84, + 481, + 132 + ], + "lines": [ + { + "bbox": [ + 417, + 84, + 481, + 132 + ], + "spans": [ + { + "bbox": [ + 417, + 84, + 481, + 132 + ], + "type": "image", + "image_path": "036d27b098b7638971fe0d417f508518019c019493bee1f94420996ebae5892e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 417, + 140, + 481, + 188 + ], + "blocks": [ + { + "bbox": [ + 424, + 133, + 470, + 140 + ], + "lines": [ + { + "bbox": [ + 424, + 133, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 424, + 133, + 470, + 140 + ], + "type": "text", + "content": "SUPIR (Zoomed)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 140, + 481, + 188 + ], + "lines": [ + { + "bbox": [ + 417, + 140, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 417, + 140, + 481, + 188 + ], + "type": "image", + "image_path": "e598cb327a5a8912fa75cf7a203ec72f9a1d9fc7d83922afdd84ca584579758f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 189, + 470, + 196 + ], + "lines": [ + { + "bbox": [ + 424, + 189, + 470, + 196 + ], + "spans": [ + { + "bbox": [ + 424, + 189, + 470, + 196 + ], + "type": "text", + "content": "SUPIR (Zoomed)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 486, + 84, + 550, + 132 + ], + "blocks": [ + { + "bbox": [ + 486, + 84, + 550, + 132 + ], + "lines": [ + { + "bbox": [ + 486, + 84, + 550, + 132 + ], + "spans": [ + { + "bbox": [ + 486, + 84, + 550, + 132 + ], + "type": "image", + "image_path": "9be8fecc9e1b9c9d878a2c4007a38d68df0f1a16ed942cc510bb87522c7ffcdd.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 486, + 140, + 550, + 188 + ], + "blocks": [ + { + "bbox": [ + 493, + 133, + 540, + 140 + ], + "lines": [ + { + "bbox": [ + 493, + 133, + 540, + 140 + ], + "spans": [ + { + "bbox": [ + 493, + 133, + 540, + 140 + ], + "type": "text", + "content": "MMSR (Zoomed)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 486, + 140, + 550, + 188 + ], + "lines": [ + { + "bbox": [ + 486, + 140, + 550, + 188 + ], + "spans": [ + { + "bbox": [ + 486, + 140, + 550, + 188 + ], + "type": "image", + "image_path": "84ca9552e002008f2d6748c92be9e1993d5c344e850a9b4117a81f280a89b647.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 189, + 541, + 196 + ], + "lines": [ + { + "bbox": [ + 493, + 189, + 541, + 196 + ], + "spans": [ + { + "bbox": [ + 493, + 189, + 541, + 196 + ], + "type": "text", + "content": "MMSR (Zoomed)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 221, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 281 + ], + "type": "text", + "content": "often leads to sharper and more detailed outputs, it can also exacerbate hallucination, resulting in details inconsistent with the low-resolution input. Such artifacts are widely reported but are difficult to suppress, even with recent efforts like balancing the training data [71]." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 288, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 419 + ], + "type": "text", + "content": "Multimodal Classifier-free Guidance. To mitigate the issue of excessive hallucination often associated with high guidance scales in CFG, we propose a novel multimodal guidance strategy. We argue that the artifacts come from the weak guidance in the negative prompting process. Instead of simply relying on text prompts for guidance, we leverage the rich information encoded in the multimodal latent tokens to strengthen both positive and negative promptings. Specifically, we condition both the postive and negative generation processes on the multimodal latent token sequence, denoted as " + }, + { + "bbox": [ + 55, + 288, + 296, + 419 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 288, + 296, + 419 + ], + "type": "text", + "content": ". This leads to the following multimodal CFG:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 58, + 423, + 294, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 423, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 58, + 423, + 294, + 447 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {\\epsilon}} (\\mathbf {z} _ {t}, c, m) = (1 + w) \\boldsymbol {\\epsilon} (\\mathbf {z} _ {t}, c, \\text {p o s}, m) - w \\boldsymbol {\\epsilon} (\\mathbf {z} _ {t}, c, \\text {n e g}, m). \\tag {4}", + "image_path": "df1b16bfcdd6b0fd25fafc6a4f27e4f26e7a3491969fdd5aa9087f7a0521b000.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 448, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 507 + ], + "type": "text", + "content": "Sec. 4.1 shows that this change in negative generation helps to achieve a better trade-off between perceptual quality and identity preservation, better maintaining the semantic content of the low-resolution input in the upscaled output, compared with the standard CFG in previous text-based methods." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 514, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 610 + ], + "type": "text", + "content": "Scaling Single-modal Guidance. Multimodal CFG effectively controls the overall influence of the prompts but does not offer control over each modality individually. To address this limitation, we introduce a mechanism to selectively amplify or suppress the contribution of specific modalities. Specifically, we modify the attention temperature " + }, + { + "bbox": [ + 55, + 514, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 514, + 296, + 610 + ], + "type": "text", + "content": " of MMLC during cross-conditioning, where " + }, + { + "bbox": [ + 55, + 514, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 514, + 296, + 610 + ], + "type": "text", + "content": " scales the attention maps before applying the softmax operation:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 82, + 613, + 295, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 613, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 82, + 613, + 295, + 640 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\delta}\\right) V. \\tag {5}", + "image_path": "1e7f765e03b12df6e64d8f73e850d927f818a80b0afe73a5419d32a51ac09caf.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "This parameter controls the sensitivity of the attention mechanism to differences between the feature sequence " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and the multimodal token sequences " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". A smaller temperature " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " typically leads to a stronger conditioning effect on the attention mechanism. In the standard scaled dot-product attention, the temperature " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " is empirically set to the square" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 221, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 316 + ], + "type": "text", + "content": "root of the key dimension " + }, + { + "bbox": [ + 313, + 221, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\sqrt{d_k}" + }, + { + "bbox": [ + 313, + 221, + 555, + 316 + ], + "type": "text", + "content": ". We observe that scaling the standard temperature " + }, + { + "bbox": [ + 313, + 221, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\sqrt{d_k}" + }, + { + "bbox": [ + 313, + 221, + 555, + 316 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 313, + 221, + 555, + 316 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 313, + 221, + 555, + 316 + ], + "type": "text", + "content": " is the key dimension) within a range of [0.4, 10] can produce high-quality results, with varying levels of fidelity to the input modalities. By scaling the temperature during sampling, we achieve precise control over the super-resolution process, enabling manageable manipulation of the output with fine-grained control." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 328, + 449, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 449, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 449, + 341 + ], + "type": "text", + "content": "3.3. MMSR Implementation" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 348, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 468 + ], + "type": "text", + "content": "Figure 3 illustrates our multimodal guided super-resolution pipeline. During inference, we first extract four modalities from the low-resolution (LR) image: a text caption generated by Gemini Flash [56], a depth map estimated by Depth Anything [66], a semantic segmentation mask produced by Mask2Former [11], and edge information extracted with a Canny edge detector. The depth map, segmentation mask, and edge information are encoded into token sequences using a pretrained VQGAN, while the text caption is processed by a pretrained CLIP encoder to obtain a text embedding." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 469, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 555, + 517 + ], + "type": "text", + "content": "Following a similar conditioning strategy to InstructPix2Pix [6], the LR image is concatenated with a noisy latent vector sampled from the diffusion model, providing additional conditioning." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 531, + 394, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 394, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 394, + 545 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": "Training Details. Our super-resolution model is initialized with the weights of a pretrained text-to-image model, with the same architecture and size as Stable Diffusion v2 [53]. The super-resolution training dataset consists of randomly degraded (using RealESRGAN degradation [63]) high-resolution images from the combined LSDIR and DIV2K datasets, with corresponding " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " high-resolution images as ground truth. We set the batch size to 1024 and the learning rate to 1e-4, which we empirically found maximized compute efficiency on TPUv5 pods. During testing, all benchmarks use a model checkpoint finetuned for 160k iterations. We use 50-step DDIM sampling, consistent with previous methods. We use a guidance rate of 4 as the default." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23145" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 75, + 127, + 133 + ], + "blocks": [ + { + "bbox": [ + 58, + 75, + 127, + 133 + ], + "lines": [ + { + "bbox": [ + 58, + 75, + 127, + 133 + ], + "spans": [ + { + "bbox": [ + 58, + 75, + 127, + 133 + ], + "type": "image", + "image_path": "dcea7f995f1d0693ee265c99b3e67182fb67e9be509651809d332ed6a681da02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 133, + 99, + 139 + ], + "lines": [ + { + "bbox": [ + 89, + 133, + 99, + 139 + ], + "spans": [ + { + "bbox": [ + 89, + 133, + 99, + 139 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 75, + 200, + 132 + ], + "blocks": [ + { + "bbox": [ + 129, + 75, + 200, + 132 + ], + "lines": [ + { + "bbox": [ + 129, + 75, + 200, + 132 + ], + "spans": [ + { + "bbox": [ + 129, + 75, + 200, + 132 + ], + "type": "image", + "image_path": "b2201a59f269c441e560bdc563e45adaba72cc1d1a9e2c2487d2b548c9344e1d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 206, + 81, + 272, + 132 + ], + "blocks": [ + { + "bbox": [ + 206, + 81, + 272, + 132 + ], + "lines": [ + { + "bbox": [ + 206, + 81, + 272, + 132 + ], + "spans": [ + { + "bbox": [ + 206, + 81, + 272, + 132 + ], + "type": "image", + "image_path": "9d3e1a10217413595a67c657828064a0813cdf6d54f1f4b3a8d5e4b1fffe1473.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 277, + 76, + 347, + 133 + ], + "blocks": [ + { + "bbox": [ + 277, + 76, + 347, + 133 + ], + "lines": [ + { + "bbox": [ + 277, + 76, + 347, + 133 + ], + "spans": [ + { + "bbox": [ + 277, + 76, + 347, + 133 + ], + "type": "image", + "image_path": "9082679ea62fd50d13c372b0a732deecf47e369e1f6a7fc940d5401c2d5c58c8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 351, + 76, + 417, + 133 + ], + "blocks": [ + { + "bbox": [ + 351, + 76, + 417, + 133 + ], + "lines": [ + { + "bbox": [ + 351, + 76, + 417, + 133 + ], + "spans": [ + { + "bbox": [ + 351, + 76, + 417, + 133 + ], + "type": "image", + "image_path": "30601950487f29035e5a733103e139de74abf7cc3ca9ecfcd45f0f0675c2c4c8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 133, + 394, + 139 + ], + "lines": [ + { + "bbox": [ + 375, + 133, + 394, + 139 + ], + "spans": [ + { + "bbox": [ + 375, + 133, + 394, + 139 + ], + "type": "text", + "content": "SeeSR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 422, + 76, + 490, + 133 + ], + "blocks": [ + { + "bbox": [ + 422, + 76, + 490, + 133 + ], + "lines": [ + { + "bbox": [ + 422, + 76, + 490, + 133 + ], + "spans": [ + { + "bbox": [ + 422, + 76, + 490, + 133 + ], + "type": "image", + "image_path": "d0727b30a5320d57fb4837559716ee716c86593f2b70db06aff85e81349a4be6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 133, + 477, + 139 + ], + "lines": [ + { + "bbox": [ + 435, + 133, + 477, + 139 + ], + "spans": [ + { + "bbox": [ + 435, + 133, + 477, + 139 + ], + "type": "text", + "content": "MMSR (Ours)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 493, + 76, + 561, + 133 + ], + "blocks": [ + { + "bbox": [ + 493, + 76, + 561, + 133 + ], + "lines": [ + { + "bbox": [ + 493, + 76, + 561, + 133 + ], + "spans": [ + { + "bbox": [ + 493, + 76, + 561, + 133 + ], + "type": "image", + "image_path": "f2dc4043b5d4560bff38ed12e3939838b5219334924ca5638deab2e07140ef09.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 62, + 139, + 129, + 190 + ], + "blocks": [ + { + "bbox": [ + 62, + 139, + 129, + 190 + ], + "lines": [ + { + "bbox": [ + 62, + 139, + 129, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 139, + 129, + 190 + ], + "type": "image", + "image_path": "314ca8417f6cf02d36615c6c51eaffc24ff25179e5e80a9bddffdbd22e5ad915.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 190, + 114, + 198 + ], + "lines": [ + { + "bbox": [ + 75, + 190, + 114, + 198 + ], + "spans": [ + { + "bbox": [ + 75, + 190, + 114, + 198 + ], + "type": "text", + "content": "LR (Zoomed)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 134, + 133, + 199, + 190 + ], + "blocks": [ + { + "bbox": [ + 134, + 133, + 199, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 133, + 199, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 133, + 199, + 190 + ], + "type": "image", + "image_path": "e08e88b3aed45a685a8207378e87b2cd3b40555df28cd538535a690a724c1ab0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 191, + 187, + 198 + ], + "lines": [ + { + "bbox": [ + 148, + 191, + 187, + 198 + ], + "spans": [ + { + "bbox": [ + 148, + 191, + 187, + 198 + ], + "type": "text", + "content": "Segmentation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 206, + 133, + 272, + 189 + ], + "blocks": [ + { + "bbox": [ + 206, + 133, + 272, + 189 + ], + "lines": [ + { + "bbox": [ + 206, + 133, + 272, + 189 + ], + "spans": [ + { + "bbox": [ + 206, + 133, + 272, + 189 + ], + "type": "image", + "image_path": "f6eec81efe8334cc29d08fab95388c76579c21ccfc5b1949eee8e7a884ffeccd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 191, + 248, + 198 + ], + "lines": [ + { + "bbox": [ + 233, + 191, + 248, + 198 + ], + "spans": [ + { + "bbox": [ + 233, + 191, + 248, + 198 + ], + "type": "text", + "content": "Edge" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 282, + 133, + 348, + 190 + ], + "blocks": [ + { + "bbox": [ + 282, + 133, + 348, + 190 + ], + "lines": [ + { + "bbox": [ + 282, + 133, + 348, + 190 + ], + "spans": [ + { + "bbox": [ + 282, + 133, + 348, + 190 + ], + "type": "image", + "image_path": "9431c23601e9346875331e0173db4a29fe1e2230252b61d94d312ab895255693.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 191, + 337, + 198 + ], + "lines": [ + { + "bbox": [ + 290, + 191, + 337, + 198 + ], + "spans": [ + { + "bbox": [ + 290, + 191, + 337, + 198 + ], + "type": "text", + "content": "PASD (Zoomed)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 353, + 140, + 419, + 189 + ], + "blocks": [ + { + "bbox": [ + 353, + 140, + 419, + 189 + ], + "lines": [ + { + "bbox": [ + 353, + 140, + 419, + 189 + ], + "spans": [ + { + "bbox": [ + 353, + 140, + 419, + 189 + ], + "type": "image", + "image_path": "b27a4e98c9c4c1f065417b6eb2c34006071329e85877897a74d1fa23e22351bb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 191, + 408, + 198 + ], + "lines": [ + { + "bbox": [ + 361, + 191, + 408, + 198 + ], + "spans": [ + { + "bbox": [ + 361, + 191, + 408, + 198 + ], + "type": "text", + "content": "SeeSR (Zoomed)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 424, + 140, + 490, + 189 + ], + "blocks": [ + { + "bbox": [ + 424, + 140, + 490, + 189 + ], + "lines": [ + { + "bbox": [ + 424, + 140, + 490, + 189 + ], + "spans": [ + { + "bbox": [ + 424, + 140, + 490, + 189 + ], + "type": "image", + "image_path": "dadc586c4b6dfd2ecff2efbd796b7b4f6b91485a73f4a815915d83f083606e5a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 191, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 430, + 191, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 430, + 191, + 481, + 198 + ], + "type": "text", + "content": "MMSR (Zoomed)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 494, + 140, + 562, + 189 + ], + "blocks": [ + { + "bbox": [ + 494, + 140, + 562, + 189 + ], + "lines": [ + { + "bbox": [ + 494, + 140, + 562, + 189 + ], + "spans": [ + { + "bbox": [ + 494, + 140, + 562, + 189 + ], + "type": "image", + "image_path": "20ecb5709aa00efd3f458ea02d8003c04b4c1f8d0b14696c674c34b7aadb74aa.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 191, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 507, + 191, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 507, + 191, + 547, + 198 + ], + "type": "text", + "content": "HR (Zoomed)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 61, + 205, + 126, + 254 + ], + "blocks": [ + { + "bbox": [ + 61, + 205, + 126, + 254 + ], + "lines": [ + { + "bbox": [ + 61, + 205, + 126, + 254 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 126, + 254 + ], + "type": "image", + "image_path": "5ac961b0687907b4b5c6f3badd1135758699326bf84bdd42424fa845c7a31af8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 135, + 207, + 196, + 252 + ], + "blocks": [ + { + "bbox": [ + 135, + 207, + 196, + 252 + ], + "lines": [ + { + "bbox": [ + 135, + 207, + 196, + 252 + ], + "spans": [ + { + "bbox": [ + 135, + 207, + 196, + 252 + ], + "type": "image", + "image_path": "7a01931086f2a99efa50eb5055bee659c22964f2c021acdcb88d17729a368260.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 63, + 263, + 129, + 312 + ], + "blocks": [ + { + "bbox": [ + 89, + 256, + 99, + 262 + ], + "lines": [ + { + "bbox": [ + 89, + 256, + 99, + 262 + ], + "spans": [ + { + "bbox": [ + 89, + 256, + 99, + 262 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 263, + 129, + 312 + ], + "lines": [ + { + "bbox": [ + 63, + 263, + 129, + 312 + ], + "spans": [ + { + "bbox": [ + 63, + 263, + 129, + 312 + ], + "type": "image", + "image_path": "18fd6a59706527cf34e84f4ed1c87763d37f1bd38d15442ed10e4fe157afab09.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 313, + 114, + 320 + ], + "lines": [ + { + "bbox": [ + 75, + 313, + 114, + 320 + ], + "spans": [ + { + "bbox": [ + 75, + 313, + 114, + 320 + ], + "type": "text", + "content": "LR (Zoomed)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 327, + 541, + 338 + ], + "lines": [ + { + "bbox": [ + 67, + 327, + 541, + 338 + ], + "spans": [ + { + "bbox": [ + 67, + 327, + 541, + 338 + ], + "type": "text", + "content": "Figure 6. Super-resolution results on common benchmarks, comparing MMSR with state-of-the-art techniques. Zoom in for detail." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 134, + 263, + 199, + 312 + ], + "blocks": [ + { + "bbox": [ + 156, + 256, + 179, + 262 + ], + "lines": [ + { + "bbox": [ + 156, + 256, + 179, + 262 + ], + "spans": [ + { + "bbox": [ + 156, + 256, + 179, + 262 + ], + "type": "text", + "content": "Caption" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 263, + 199, + 312 + ], + "lines": [ + { + "bbox": [ + 134, + 263, + 199, + 312 + ], + "spans": [ + { + "bbox": [ + 134, + 263, + 199, + 312 + ], + "type": "image", + "image_path": "c0f79b5b70da690e3b618e85314c809136e3ef9d713f429faa6a42d2faf4d66c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 313, + 187, + 320 + ], + "lines": [ + { + "bbox": [ + 148, + 313, + 187, + 320 + ], + "spans": [ + { + "bbox": [ + 148, + 313, + 187, + 320 + ], + "type": "text", + "content": "Segmentation" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 206, + 206, + 272, + 255 + ], + "blocks": [ + { + "bbox": [ + 206, + 206, + 272, + 255 + ], + "lines": [ + { + "bbox": [ + 206, + 206, + 272, + 255 + ], + "spans": [ + { + "bbox": [ + 206, + 206, + 272, + 255 + ], + "type": "image", + "image_path": "0232628f8066fc3e8940a5bd13aebe43ae8e8f74f9fb7f70d2457af568c7fef0.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 256, + 249, + 262 + ], + "lines": [ + { + "bbox": [ + 231, + 256, + 249, + 262 + ], + "spans": [ + { + "bbox": [ + 231, + 256, + 249, + 262 + ], + "type": "text", + "content": "Depth" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 206, + 263, + 272, + 312 + ], + "blocks": [ + { + "bbox": [ + 206, + 263, + 272, + 312 + ], + "lines": [ + { + "bbox": [ + 206, + 263, + 272, + 312 + ], + "spans": [ + { + "bbox": [ + 206, + 263, + 272, + 312 + ], + "type": "image", + "image_path": "1305a8bd471d9adc23ac4117102a82cc0c1d2fe5045a525f04806ac30952251c.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 313, + 247, + 320 + ], + "lines": [ + { + "bbox": [ + 233, + 313, + 247, + 320 + ], + "spans": [ + { + "bbox": [ + 233, + 313, + 247, + 320 + ], + "type": "text", + "content": "Edge" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 279, + 205, + 345, + 255 + ], + "blocks": [ + { + "bbox": [ + 279, + 205, + 345, + 255 + ], + "lines": [ + { + "bbox": [ + 279, + 205, + 345, + 255 + ], + "spans": [ + { + "bbox": [ + 279, + 205, + 345, + 255 + ], + "type": "image", + "image_path": "26fa23a053fe8ce905d48a7ea7472a7491062891c556dadeb1991ec48acc6ca2.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 256, + 322, + 262 + ], + "lines": [ + { + "bbox": [ + 304, + 256, + 322, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 322, + 262 + ], + "type": "text", + "content": "PASD" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 282, + 263, + 347, + 312 + ], + "blocks": [ + { + "bbox": [ + 282, + 263, + 347, + 312 + ], + "lines": [ + { + "bbox": [ + 282, + 263, + 347, + 312 + ], + "spans": [ + { + "bbox": [ + 282, + 263, + 347, + 312 + ], + "type": "image", + "image_path": "0de75abeba3895053529084962b17030e1fd04042b939bdc0f8b0f449392b2f0.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 313, + 337, + 320 + ], + "lines": [ + { + "bbox": [ + 290, + 313, + 337, + 320 + ], + "spans": [ + { + "bbox": [ + 290, + 313, + 337, + 320 + ], + "type": "text", + "content": "PASD (Zoomed)" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 351, + 206, + 416, + 255 + ], + "blocks": [ + { + "bbox": [ + 351, + 206, + 416, + 255 + ], + "lines": [ + { + "bbox": [ + 351, + 206, + 416, + 255 + ], + "spans": [ + { + "bbox": [ + 351, + 206, + 416, + 255 + ], + "type": "image", + "image_path": "e8550f03dcfca4951c439bb8d26ab1eb4188baf0db4e442fb3b6fe018bcb430e.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 256, + 394, + 262 + ], + "lines": [ + { + "bbox": [ + 375, + 256, + 394, + 262 + ], + "spans": [ + { + "bbox": [ + 375, + 256, + 394, + 262 + ], + "type": "text", + "content": "SeeSR" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 354, + 263, + 418, + 312 + ], + "blocks": [ + { + "bbox": [ + 354, + 263, + 418, + 312 + ], + "lines": [ + { + "bbox": [ + 354, + 263, + 418, + 312 + ], + "spans": [ + { + "bbox": [ + 354, + 263, + 418, + 312 + ], + "type": "image", + "image_path": "bc5b05e6cb23d8f6e80e9f93a594fffbf55c5c41182f3cc5f73a3d0b1d67ed89.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 313, + 409, + 320 + ], + "lines": [ + { + "bbox": [ + 360, + 313, + 409, + 320 + ], + "spans": [ + { + "bbox": [ + 360, + 313, + 409, + 320 + ], + "type": "text", + "content": "SeeSR (Zoomed)" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 422, + 205, + 488, + 255 + ], + "blocks": [ + { + "bbox": [ + 422, + 205, + 488, + 255 + ], + "lines": [ + { + "bbox": [ + 422, + 205, + 488, + 255 + ], + "spans": [ + { + "bbox": [ + 422, + 205, + 488, + 255 + ], + "type": "image", + "image_path": "2d7a388792b499d21119f39ab174717ae6f7606991b8d5f0b677236e6f9ef594.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 255, + 477, + 262 + ], + "lines": [ + { + "bbox": [ + 435, + 255, + 477, + 262 + ], + "spans": [ + { + "bbox": [ + 435, + 255, + 477, + 262 + ], + "type": "text", + "content": "MMSR (Ours)" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 425, + 263, + 490, + 312 + ], + "blocks": [ + { + "bbox": [ + 425, + 263, + 490, + 312 + ], + "lines": [ + { + "bbox": [ + 425, + 263, + 490, + 312 + ], + "spans": [ + { + "bbox": [ + 425, + 263, + 490, + 312 + ], + "type": "image", + "image_path": "4f0fb3998679a2fcdaf15d33903d67b7a2f6555ad7d3a71100d4cdf76feb2ddc.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 313, + 482, + 320 + ], + "lines": [ + { + "bbox": [ + 430, + 313, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 430, + 313, + 482, + 320 + ], + "type": "text", + "content": "MMSR (Zoomed)" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 493, + 206, + 559, + 255 + ], + "blocks": [ + { + "bbox": [ + 493, + 206, + 559, + 255 + ], + "lines": [ + { + "bbox": [ + 493, + 206, + 559, + 255 + ], + "spans": [ + { + "bbox": [ + 493, + 206, + 559, + 255 + ], + "type": "image", + "image_path": "248b4414d2e211f23b8c182caebea5783d57b9ec4f14bc34cb59cf96cdd34a30.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 522, + 256, + 533, + 262 + ], + "lines": [ + { + "bbox": [ + 522, + 256, + 533, + 262 + ], + "spans": [ + { + "bbox": [ + 522, + 256, + 533, + 262 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 496, + 263, + 561, + 312 + ], + "blocks": [ + { + "bbox": [ + 496, + 263, + 561, + 312 + ], + "lines": [ + { + "bbox": [ + 496, + 263, + 561, + 312 + ], + "spans": [ + { + "bbox": [ + 496, + 263, + 561, + 312 + ], + "type": "image", + "image_path": "d22995bef713bb026ddaec73958546b868c45cd7f738edb0f7e789b982d62076.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 313, + 548, + 320 + ], + "lines": [ + { + "bbox": [ + 507, + 313, + 548, + 320 + ], + "spans": [ + { + "bbox": [ + 507, + 313, + 548, + 320 + ], + "type": "text", + "content": "HR (Zoomed)" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "bbox": [ + 55, + 347, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 491 + ], + "type": "text", + "content": "Evaluation Details. We use reference-based metrics like LPIPS [75] and DISTS [16], and non-reference-based metrics like NIQE [42], MANIQA [67], MUSIQ [27], and CLIPIQA [58], as these have been shown to align well with humans' aesthetic preferences. Our compared baselines include recent diffusion-based super-resolution methods [33, 53, 59, 65, 68, 72] and the representative methods [8, 31, 32, 63, 73]. We collected their results either by running their official code or from published results. In a few cases where the multimodal prediction failed (two flat images and one pencil drawing from DRealSR), we manually replaced the prediction with " + }, + { + "bbox": [ + 55, + 347, + 296, + 491 + ], + "type": "inline_equation", + "content": "m_{\\emptyset}" + }, + { + "bbox": [ + 55, + 347, + 296, + 491 + ], + "type": "text", + "content": " (see supplemental material)." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 55, + 502, + 174, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 174, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 174, + 514 + ], + "type": "text", + "content": "4.1. Quantitative Results" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "Table 1 shows the quantitative result comparisons of ours with the other baselines, compared on the synthetic benchmark and real-world super-resolution benchmarks. Several facts are worth noting: (1) Our method achieves the best LPIPS, DISTS, NIQE, and FID scores on the DIV2K-Val and RealSR dataset, significantly outperforming previous state-of-the-art methods. This superiority demonstrates that our method can generate more perceptually identical details from the guidance of multimodal context. (2) Our method also achieves the best performance in the non-reference visual quality metrics including NIQE, MANIQA, CLIPIQA, and MUSIQ scores. This advantage demonstrates our method can fully utilize the photorealistic priors encapsulated in the pretrained diffusion model. Overall, these observations fully demonstrate our advantages in using multimodal guidance compared with the past single-modality methods." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "content": "Effects of Multimodal CFG Guidance. Table 3 demonstrates the effects of our proposed multimodal CFG on the 1MP DIV2K-Val benchmark, which replaces the negative score guided by empty language token " + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "inline_equation", + "content": "\\epsilon (\\mathbf{z}_t,c,\\mathrm{neg})" + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "content": ", shortened as " + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "inline_equation", + "content": "cfg" + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "content": ", with multimodal guided negative score " + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "inline_equation", + "content": "(\\epsilon (\\mathbf{z}_t,c,m,\\mathrm{neg}))" + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "content": ", shortened as " + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "content": " -cfg. In order to exclude the influence of architectural differences, we also compare our method with the empty multimodal token " + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "inline_equation", + "content": "\\epsilon (\\mathbf{z}_t,c,m_{\\emptyset},\\mathrm{neg})" + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "content": ", shortened as " + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "inline_equation", + "content": "m_{\\emptyset}" + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "content": " -cfg. Note that the positive score for all three compared methods is the same " + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "inline_equation", + "content": "\\epsilon (\\mathbf{z}_t,c,m,\\mathrm{pos})" + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "content": ". The results show that our method mitigates the degradation reflected in the LPIPS and NIQE scores when using guidance rates of 10 and 14 in both CFG and " + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "inline_equation", + "content": "m_{\\emptyset}" + }, + { + "bbox": [ + 313, + 347, + 555, + 550 + ], + "type": "text", + "content": " -cfg, while maintaining comparable performance. The visual results in Table 3 further demonstrate our superiority, where ours suppresses the color change and incorrect texture of high CFG guidance rate, leading to higher CLIPIQA score." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "text", + "content": "Effects of Latent Connector. To demonstrate the effectiveness of the proposed multimodal latent connector, we conducted an ablation study by training a new multimodal super-resolution model without the latent connector, which directly uses the longer multimodal token sequence as input. The quantitative performance is shown in Table 2. Our " + }, + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "text", + "content": ". MMLC outperforms the " + }, + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "text", + "content": ". MMLC in all reference and non-reference metrics. The visual results in Figure 7 show that the model without the MMLC is more likely to exacerbate hallucination. Its result shows bokeh on the branches but a clear trunk, even though the depth input clearly shows they are at the same depth." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 314, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 555, + 713 + ], + "type": "text", + "content": "Contributions of Each Modality. Figure 8 shows the con" + } + ] + } + ], + "index": 58 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23146" + } + ] + } + ], + "index": 59 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 81, + 293, + 219 + ], + "blocks": [ + { + "bbox": [ + 95, + 72, + 169, + 79 + ], + "lines": [ + { + "bbox": [ + 95, + 72, + 169, + 79 + ], + "spans": [ + { + "bbox": [ + 95, + 72, + 169, + 79 + ], + "type": "text", + "content": "DIV2K-Val-3k 512 × 512" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 81, + 293, + 219 + ], + "lines": [ + { + "bbox": [ + 57, + 81, + 293, + 219 + ], + "spans": [ + { + "bbox": [ + 57, + 81, + 293, + 219 + ], + "type": "table", + "html": "
MethodsPSNRSSIMLPIPS ↓DISTS ↓NIQE ↓FID ↓MUSIQCLIPQA
BSRGAN21.870.55390.41360.27374.761564.2859.110.5183
R-ESRGAN21.940.57360.38680.26014.920953.4658.640.5424
LDL21.520.56900.39950.26885.024958.9457.900.5313
DASR21.720.55360.42660.26884.859667.2254.220.5241
FeMASR20.850.51630.39730.24284.572653.7058.100.5597
LDM21.260.52390.41540.25006.466741.9356.520.5695
StableSR20.840.48870.40550.25424.655136.5762.950.6486
ResShift21.750.54220.42840.26066.973155.7758.230.5948
PASD20.770.49580.44100.25384.832840.7766.850.6799
DiffBIR20.940.49380.42700.24714.721140.4265.230.6664
SeeSR21.190.53860.38430.22574.927531.9368.330.6946
MMSR21.740.56930.37070.20714.253229.3570.060.7164
", + "image_path": "284306d0f5bfd49cc475d0a36a18f0969eba73bf80c46cbc0f059ae65c492264.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 234, + 293, + 308 + ], + "blocks": [ + { + "bbox": [ + 94, + 226, + 173, + 233 + ], + "lines": [ + { + "bbox": [ + 94, + 226, + 173, + 233 + ], + "spans": [ + { + "bbox": [ + 94, + 226, + 173, + 233 + ], + "type": "text", + "content": "DIV2K-Val-100 1024 × 1024" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 234, + 293, + 308 + ], + "lines": [ + { + "bbox": [ + 57, + 234, + 293, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 234, + 293, + 308 + ], + "type": "table", + "html": "
MethodsPSNRSSIMLPIPS ↓DISTS ↓NIQE ↓MANIQAMUSIQCLIPiQA
R-ESRGAN21.770.58130.36240.19903.65730.404647.540.5358
StableSR20.070.39470.50970.24273.62600.411365.390.6938
PASD21.270.53690.34730.17533.63210.470869.690.6914
SUPIR20.650.53500.38490.18143.64580.405165.880.5697
SeeSR21.310.55780.32730.16204.02150.543969.790.6941
MMSR21.870.55650.28100.14923.42430.488572.310.7294
", + "image_path": "43ae4967a5a146122d817c94c19509b6a24ef641a96be37267bf2bfafd94efb9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 324, + 293, + 399 + ], + "blocks": [ + { + "bbox": [ + 95, + 315, + 148, + 323 + ], + "lines": [ + { + "bbox": [ + 95, + 315, + 148, + 323 + ], + "spans": [ + { + "bbox": [ + 95, + 315, + 148, + 323 + ], + "type": "text", + "content": "RealSR " + }, + { + "bbox": [ + 95, + 315, + 148, + 323 + ], + "type": "inline_equation", + "content": "512\\times 512" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 324, + 293, + 399 + ], + "lines": [ + { + "bbox": [ + 58, + 324, + 293, + 399 + ], + "spans": [ + { + "bbox": [ + 58, + 324, + 293, + 399 + ], + "type": "table", + "html": "
MethodsPSNRSSIMLPIPS ↓LIQENIMAMANIQAMUSIQCLIPiQA
R-ESRGAN25.690.76160.27273.35744.65480.548760.180.4449
StableSR24.700.70850.30183.61064.81500.622165.780.6178
PASD24.290.66300.34353.57494.85540.649368.690.6590
SUPIR22.970.62980.37503.56824.57570.574561.490.6434
SeeSR25.180.72160.30094.13604.91930.644269.770.6612
MMSR24.830.70030.29524.34685.10940.657871.330.6717
", + "image_path": "9e7647af8b83debf2660cb6b847a406cc761f9caa902c7cab7994bf21cb98bdf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 57, + 415, + 293, + 491 + ], + "blocks": [ + { + "bbox": [ + 95, + 406, + 151, + 413 + ], + "lines": [ + { + "bbox": [ + 95, + 406, + 151, + 413 + ], + "spans": [ + { + "bbox": [ + 95, + 406, + 151, + 413 + ], + "type": "text", + "content": "DrealSR " + }, + { + "bbox": [ + 95, + 406, + 151, + 413 + ], + "type": "inline_equation", + "content": "512\\times 512" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 415, + 293, + 491 + ], + "lines": [ + { + "bbox": [ + 57, + 415, + 293, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 293, + 491 + ], + "type": "table", + "html": "
MethodsPSNRSSIMLPIPS ↓LIQENIMAMANIQAMUSIQCLIPiQA
R-ESRGAN28.640.80530.28472.92554.32580.490754.180.4422
StableSR26.710.72240.32843.24254.48610.559458.510.6357
PASD27.000.70840.39313.59084.66180.585064.810.6773
SUPIR24.610.61230.42943.47104.38150.538157.320.6758
SeeSR26.750.74050.31744.12704.69420.605265.090.6908
MMSR27.280.74560.32494.50235.05580.630168.930.6999
", + "image_path": "fc5235a0103e9ac2eda9a46859369918f64ddb6f0007d175a344d43d80c0c2f2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 59, + 531, + 293, + 567 + ], + "blocks": [ + { + "bbox": [ + 55, + 495, + 296, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 529 + ], + "type": "text", + "content": "Table 1. Severely degraded, low-resolution (LR) images can produce inaccurate multi-modal information, manifesting as distorted edges, misidentified objects, and other artifacts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 531, + 293, + 567 + ], + "lines": [ + { + "bbox": [ + 59, + 531, + 293, + 567 + ], + "spans": [ + { + "bbox": [ + 59, + 531, + 293, + 567 + ], + "type": "table", + "html": "
MUSIQNIQE ↓DISTS ↓LPIPS ↓Throughput
w/o. MMLC69.693.48450.17810.39293.48 img/s
w. MMLC72.313.42430.14920.28103.32 img/s
", + "image_path": "e5763c9b9a44445386cf61255538423942c5f15e1ebf90ba55804c2630eaf88d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 605, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 297, + 714 + ], + "type": "text", + "content": "tributions of each modality by masking out input modalities during testing. Benefiting from the improved flexibility of using multimodal token " + }, + { + "bbox": [ + 54, + 605, + 297, + 714 + ], + "type": "inline_equation", + "content": "m_{\\emptyset}" + }, + { + "bbox": [ + 54, + 605, + 297, + 714 + ], + "type": "text", + "content": ", discussed in Sec. 3.1, our method is robust in scenarios with only fewer modalities input during testing. We would like to note several observations: (1) Our default multimodal setting achieves the best trade-off between the non-reference metric MUSIQ and the reference-based metric DISTS. (2) Depth information primarily enhances perceptual quality (MUSIQ), while other modalities" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 72, + 430, + 118 + ], + "blocks": [ + { + "bbox": [ + 55, + 571, + 295, + 593 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 593 + ], + "type": "text", + "content": "Table 2. Ablation of the impact of the Multimodal Latent Connector module (MMLC) on DIV2K-Val-100 1024p." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 72, + 430, + 118 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 430, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 430, + 118 + ], + "type": "table", + "html": "
guidance21014
cfg0.32390.44910.5064
m0cfg0.28150.48030.5493
mcfg0.28100.34710.3772
", + "image_path": "5b4480ca13d2918bc60ce4201738b49b7081d15f8dfbc9233b979a4dafb0ac13.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 129, + 457, + 233 + ], + "blocks": [ + { + "bbox": [ + 318, + 129, + 457, + 233 + ], + "lines": [ + { + "bbox": [ + 318, + 129, + 457, + 233 + ], + "spans": [ + { + "bbox": [ + 318, + 129, + 457, + 233 + ], + "type": "image", + "image_path": "471c9862b8304157c561b833c29b15e6089a7e5a5c408dedcab0796b0df51e6b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 233, + 440, + 242 + ], + "lines": [ + { + "bbox": [ + 339, + 233, + 440, + 242 + ], + "spans": [ + { + "bbox": [ + 339, + 233, + 440, + 242 + ], + "type": "text", + "content": "Segment, Depth, LR, and Edge Input" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 402, + 243, + 468, + 251 + ], + "lines": [ + { + "bbox": [ + 402, + 243, + 468, + 251 + ], + "spans": [ + { + "bbox": [ + 402, + 243, + 468, + 251 + ], + "type": "text", + "content": "(c) Visual comaprisons" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 438, + 72, + 550, + 118 + ], + "blocks": [ + { + "bbox": [ + 351, + 121, + 397, + 129 + ], + "lines": [ + { + "bbox": [ + 351, + 121, + 397, + 129 + ], + "spans": [ + { + "bbox": [ + 351, + 121, + 397, + 129 + ], + "type": "text", + "content": "(a) LPIPS score" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 438, + 72, + 550, + 118 + ], + "lines": [ + { + "bbox": [ + 438, + 72, + 550, + 118 + ], + "spans": [ + { + "bbox": [ + 438, + 72, + 550, + 118 + ], + "type": "table", + "html": "
guidance21014
cfg3.5774.61795.1886
m0cfg3.62615.10815.9175
mcfg3.46793.74193.9815
", + "image_path": "29503b5997778ecba991ea2ce959b8856f9888661a7e6dedf8b2ef0425773d04.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 470, + 121, + 515, + 129 + ], + "lines": [ + { + "bbox": [ + 470, + 121, + 515, + 129 + ], + "spans": [ + { + "bbox": [ + 470, + 121, + 515, + 129 + ], + "type": "text", + "content": "(b) NIQE score" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 462, + 129, + 550, + 150 + ], + "blocks": [ + { + "bbox": [ + 462, + 129, + 550, + 150 + ], + "lines": [ + { + "bbox": [ + 462, + 129, + 550, + 150 + ], + "spans": [ + { + "bbox": [ + 462, + 129, + 550, + 150 + ], + "type": "image", + "image_path": "bccfcd896c218f8402f694cedaf30180c9c27eca1780b17e5b6e09f80050ba58.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 461, + 157, + 549, + 178 + ], + "blocks": [ + { + "bbox": [ + 470, + 151, + 541, + 157 + ], + "lines": [ + { + "bbox": [ + 470, + 151, + 541, + 157 + ], + "spans": [ + { + "bbox": [ + 470, + 151, + 541, + 157 + ], + "type": "inline_equation", + "content": "cfg = 2.0" + }, + { + "bbox": [ + 470, + 151, + 541, + 157 + ], + "type": "text", + "content": " CIPIQA " + }, + { + "bbox": [ + 470, + 151, + 541, + 157 + ], + "type": "inline_equation", + "content": "= 0.4481" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 461, + 157, + 549, + 178 + ], + "lines": [ + { + "bbox": [ + 461, + 157, + 549, + 178 + ], + "spans": [ + { + "bbox": [ + 461, + 157, + 549, + 178 + ], + "type": "image", + "image_path": "2fb16ce560d277d9ac9b41456fffe629f39327c38a80ec517e4f01c701203c6f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 178, + 542, + 185 + ], + "lines": [ + { + "bbox": [ + 468, + 178, + 542, + 185 + ], + "spans": [ + { + "bbox": [ + 468, + 178, + 542, + 185 + ], + "type": "inline_equation", + "content": "cfg = 14.0" + }, + { + "bbox": [ + 468, + 178, + 542, + 185 + ], + "type": "text", + "content": " CLIPIQA=0.7091" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 461, + 185, + 549, + 206 + ], + "blocks": [ + { + "bbox": [ + 461, + 185, + 549, + 206 + ], + "lines": [ + { + "bbox": [ + 461, + 185, + 549, + 206 + ], + "spans": [ + { + "bbox": [ + 461, + 185, + 549, + 206 + ], + "type": "image", + "image_path": "708285030746072e83429d216889c81bdb5b81d53200da6954bf84f9cbc778cd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 206, + 546, + 213 + ], + "lines": [ + { + "bbox": [ + 465, + 206, + 546, + 213 + ], + "spans": [ + { + "bbox": [ + 465, + 206, + 546, + 213 + ], + "type": "inline_equation", + "content": "m\\text{-} cfg = 2.0" + }, + { + "bbox": [ + 465, + 206, + 546, + 213 + ], + "type": "text", + "content": " CLIPIQA " + }, + { + "bbox": [ + 465, + 206, + 546, + 213 + ], + "type": "inline_equation", + "content": "= 0.4122" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 461, + 213, + 549, + 233 + ], + "blocks": [ + { + "bbox": [ + 461, + 213, + 549, + 233 + ], + "lines": [ + { + "bbox": [ + 461, + 213, + 549, + 233 + ], + "spans": [ + { + "bbox": [ + 461, + 213, + 549, + 233 + ], + "type": "image", + "image_path": "d3e3ac5369266df04fbfa8f1c4a2e899b27cbd942094bacd84651fda47de91f7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 233, + 547, + 241 + ], + "lines": [ + { + "bbox": [ + 463, + 233, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 463, + 233, + 547, + 241 + ], + "type": "inline_equation", + "content": "m\\text{-} cfg = 14.0" + }, + { + "bbox": [ + 463, + 233, + 547, + 241 + ], + "type": "text", + "content": " CIPIQA " + }, + { + "bbox": [ + 463, + 233, + 547, + 241 + ], + "type": "inline_equation", + "content": "= 0.7352" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 318, + 297, + 432, + 344 + ], + "blocks": [ + { + "bbox": [ + 313, + 251, + 554, + 295 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 295 + ], + "type": "text", + "content": "Table 3. Ablation of guidance on DIV2K-Val-100 1024p. Our multimodal CFG (" + }, + { + "bbox": [ + 313, + 251, + 554, + 295 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 251, + 554, + 295 + ], + "type": "text", + "content": "-cfg) mitigates artifacts often present when using high guidance rates. This leads to an improved balance between visual fidelity and preservation of key identifying features." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 297, + 432, + 344 + ], + "lines": [ + { + "bbox": [ + 318, + 297, + 432, + 344 + ], + "spans": [ + { + "bbox": [ + 318, + 297, + 432, + 344 + ], + "type": "image", + "image_path": "3b3fd4014db8a5eed536a6d3664f059a3b4d5b8b072239c3d026eebe95122258.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 345, + 390, + 353 + ], + "lines": [ + { + "bbox": [ + 361, + 345, + 390, + 353 + ], + "spans": [ + { + "bbox": [ + 361, + 345, + 390, + 353 + ], + "type": "text", + "content": "LR Input" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 436, + 297, + 551, + 344 + ], + "blocks": [ + { + "bbox": [ + 436, + 297, + 551, + 344 + ], + "lines": [ + { + "bbox": [ + 436, + 297, + 551, + 344 + ], + "spans": [ + { + "bbox": [ + 436, + 297, + 551, + 344 + ], + "type": "image", + "image_path": "df974042424819b6ad44cf618c72580032d8eb49c9ef08783297b2ec2dcd6bbc.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 318, + 354, + 433, + 399 + ], + "blocks": [ + { + "bbox": [ + 318, + 354, + 433, + 399 + ], + "lines": [ + { + "bbox": [ + 318, + 354, + 433, + 399 + ], + "spans": [ + { + "bbox": [ + 318, + 354, + 433, + 399 + ], + "type": "image", + "image_path": "5f0106977ab5347565c3243ae29405907dea68c6754fc3e9609da218cc954142.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 399, + 405, + 408 + ], + "lines": [ + { + "bbox": [ + 345, + 399, + 405, + 408 + ], + "spans": [ + { + "bbox": [ + 345, + 399, + 405, + 408 + ], + "type": "text", + "content": "w/o. MMLC Result" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 435, + 353, + 550, + 399 + ], + "blocks": [ + { + "bbox": [ + 473, + 345, + 511, + 353 + ], + "lines": [ + { + "bbox": [ + 473, + 345, + 511, + 353 + ], + "spans": [ + { + "bbox": [ + 473, + 345, + 511, + 353 + ], + "type": "text", + "content": "Depth Input" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 353, + 550, + 399 + ], + "lines": [ + { + "bbox": [ + 435, + 353, + 550, + 399 + ], + "spans": [ + { + "bbox": [ + 435, + 353, + 550, + 399 + ], + "type": "image", + "image_path": "309df6ac578aebc7bfeb3d9dec1b1a33e32eed7d1583cc359e332b6824252d03.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 400, + 518, + 407 + ], + "lines": [ + { + "bbox": [ + 465, + 400, + 518, + 407 + ], + "spans": [ + { + "bbox": [ + 465, + 400, + 518, + 407 + ], + "type": "text", + "content": "w. MMLC Result" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 319, + 427, + 451, + 482 + ], + "blocks": [ + { + "bbox": [ + 321, + 411, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 321, + 411, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 321, + 411, + 545, + 422 + ], + "type": "text", + "content": "Figure 7. Visual results when (not) using the MMLC module." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 427, + 451, + 482 + ], + "lines": [ + { + "bbox": [ + 319, + 427, + 451, + 482 + ], + "spans": [ + { + "bbox": [ + 319, + 427, + 451, + 482 + ], + "type": "image", + "image_path": "4c8ec3f47a4344137af101359c05d1c3e92b3928e330b4235acd7a6487d4878f.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 485, + 554, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 519 + ], + "type": "text", + "content": "Figure 8. Ablation of modalities on DIV2K-Val-100 1024p. Incorporating different modalities enhances various aspects of the super-resolution results compared to the text-guided baseline." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 458, + 428, + 550, + 482 + ], + "blocks": [ + { + "bbox": [ + 458, + 428, + 550, + 482 + ], + "lines": [ + { + "bbox": [ + 458, + 428, + 550, + 482 + ], + "spans": [ + { + "bbox": [ + 458, + 428, + 550, + 482 + ], + "type": "image", + "image_path": "888c95ca94a1bed5fd4b8d9bfd6c82278c314b5ff5e46703b21d5ca3a658b13b.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "text", + "content": "contribute more significantly to preserving identity (DISTS). These results highlight the diverse contributions of each modality and emphasize the advantage of our multimodal approach in effectively combining their strengths for optimal super-resolution performance." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 598, + 425, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 425, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 425, + 611 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Figures 5 and 6 provide a visual comparison of our method against state-of-the-art approaches, namely SeeSR [65], PASD [68], and SUPIR [71]. Our method demonstrates several key advantages. Enhanced Realism: In the first example of Figure 6, our method produces fewer artifacts and remains more faithful to the high-resolution image compared to other methods. This highlights our ability to maintain realism without introducing spurious details. Robustness to" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23147" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 91, + 132, + 149 + ], + "blocks": [ + { + "bbox": [ + 69, + 77, + 124, + 85 + ], + "lines": [ + { + "bbox": [ + 69, + 77, + 124, + 85 + ], + "spans": [ + { + "bbox": [ + 69, + 77, + 124, + 85 + ], + "type": "text", + "content": "(a). Results of PASD" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 91, + 132, + 149 + ], + "lines": [ + { + "bbox": [ + 63, + 91, + 132, + 149 + ], + "spans": [ + { + "bbox": [ + 63, + 91, + 132, + 149 + ], + "type": "image", + "image_path": "49fae6f99adc81fda8523d7826166484f2a755733a2bff7320b497d1b41a8d87.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 155, + 132, + 215 + ], + "blocks": [ + { + "bbox": [ + 63, + 155, + 132, + 215 + ], + "lines": [ + { + "bbox": [ + 63, + 155, + 132, + 215 + ], + "spans": [ + { + "bbox": [ + 63, + 155, + 132, + 215 + ], + "type": "image", + "image_path": "15eae51b990b4a0e7675878d8ff7068bccd05013eec4bdc650e1631fdd4ce5cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 220, + 132, + 280 + ], + "blocks": [ + { + "bbox": [ + 63, + 220, + 132, + 280 + ], + "lines": [ + { + "bbox": [ + 63, + 220, + 132, + 280 + ], + "spans": [ + { + "bbox": [ + 63, + 220, + 132, + 280 + ], + "type": "image", + "image_path": "d7096bc195c9c7475167467d3aeac064e261c4d18d7e9020afcfe39a81eb67cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 292, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 555, + 336 + ], + "type": "text", + "content": "Figure 9. Our method allows for fine-grained control over super-resolution results by adjusting the influence of each input modality. For example, reducing the edge temperature enhances edge sharpness (first row). Lowering the segmentation temperature emphasizes distinct features, such as the star pattern on the flag (second row). Decreasing the depth temperature accentuates depth-of-field effects, like the bokeh between the foreground and background (third row). In contrast, PASD [68] exhibits limited control over such fine-grained details." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 91, + 209, + 150 + ], + "blocks": [ + { + "bbox": [ + 140, + 91, + 209, + 150 + ], + "lines": [ + { + "bbox": [ + 140, + 91, + 209, + 150 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 209, + 150 + ], + "type": "image", + "image_path": "f1ebf466d0fd63f3a8eba5df203f90bbb9347e8277e28f420eef76d2169eb899.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 140, + 155, + 209, + 215 + ], + "blocks": [ + { + "bbox": [ + 140, + 155, + 209, + 215 + ], + "lines": [ + { + "bbox": [ + 140, + 155, + 209, + 215 + ], + "spans": [ + { + "bbox": [ + 140, + 155, + 209, + 215 + ], + "type": "image", + "image_path": "83f73c6f98f89b437c7cd845654d1b1d76a1817e26aacf7b5dca378364529fa0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 140, + 220, + 209, + 279 + ], + "blocks": [ + { + "bbox": [ + 140, + 220, + 209, + 279 + ], + "lines": [ + { + "bbox": [ + 140, + 220, + 209, + 279 + ], + "spans": [ + { + "bbox": [ + 140, + 220, + 209, + 279 + ], + "type": "image", + "image_path": "8a156ad46d9cc4447636837d62f43b894a65c2f72ec2fd21182ac6f619aa9631.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 219, + 93, + 547, + 150 + ], + "blocks": [ + { + "bbox": [ + 280, + 77, + 441, + 85 + ], + "lines": [ + { + "bbox": [ + 280, + 77, + 441, + 85 + ], + "spans": [ + { + "bbox": [ + 280, + 77, + 441, + 85 + ], + "type": "text", + "content": "(b). Results of MMSR under different modality-specific controls" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 93, + 547, + 150 + ], + "lines": [ + { + "bbox": [ + 219, + 93, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 219, + 93, + 547, + 150 + ], + "type": "image", + "image_path": "03ebe657ad6d0f362bbbd98d85061116b860a8606687cd6cd31f8622e9c6df9a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 219, + 158, + 547, + 215 + ], + "blocks": [ + { + "bbox": [ + 219, + 158, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 219, + 158, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 219, + 158, + 547, + 215 + ], + "type": "image", + "image_path": "7f23dbfb7c3bf62e05058b955e276d1e727aa1abddf451f78a4671f1bea85c32.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 219, + 224, + 547, + 279 + ], + "blocks": [ + { + "bbox": [ + 219, + 224, + 547, + 279 + ], + "lines": [ + { + "bbox": [ + 219, + 224, + 547, + 279 + ], + "spans": [ + { + "bbox": [ + 219, + 224, + 547, + 279 + ], + "type": "image", + "image_path": "3cd652ef8a05a0e179a7f609e7a39646cc403a280653dcd404923b8d1bb84cf9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 357, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 464 + ], + "type": "text", + "content": "Challenging Conditions: The second example in Figure 6, a long-range shot affected by noise and turbulence, shows our method's ability to generate clear human details where other methods struggle. This is attributed to the effective use of semantic segmentation for accurate human localization, further enriched by the textual caption. These examples illustrate the effectiveness of our multimodal approach in generating high-quality images across diverse scenarios. See supplemental material for additional results and analysis." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 478, + 212, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 212, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 212, + 490 + ], + "type": "text", + "content": "4.3. Controllability Comparisons" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "While recent works leverage text prompts for controlling super-resolution [65, 68, 71], these often lack fine-grained control and can yield inconsistent results. Our multimodal approach introduces modality-specific temperature weights to scale attention scores, enabling precise manipulation of SR outputs by amplifying or diminishing the influence of each modality (e.g., depth, segmentation). Figure 9 contrasts our approach with PASD [68], which relies solely on text prompts. Specifically, decreasing the temperature of the edge modality changes the richness of details in our result, while changing PASD's prompt with \"more edge\" doesn't change their result. Decreasing the temperature of semantic segmentation map and depth map also lead to manageable changes such as more visible star pattern and stronger bokeh. In contrast, directly add corresponding prompt in PASD's prompt barely changes its result. Moreover, our approach exhibits smooth transitions in image characteristics as temperatures are adjusted, providing interpretable control and insights into" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 357, + 418, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 418, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 418, + 369 + ], + "type": "text", + "content": "the role of each modality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 380, + 387, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 387, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 387, + 392 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 400, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 555 + ], + "type": "text", + "content": "This work introduces a novel diffusion-based framework for image super-resolution that seamlessly integrates diverse modalities—including text descriptions, depth maps, edges, and segmentation maps—within a single, unified model. By leveraging pretrained text-to-image models, our method achieves enhanced realism and accurate reconstruction, outperforming existing text-guided super-resolution methods both qualitatively and quantitatively. Furthermore, a learnable multimodal token and modality-specific contribution controls provide fine-grained control over the super-resolution process, enabling adjustable perception-distortion tradeoffs and robust performance even with imperfect or missing modalities in most cases." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Limitations and Future Work. While adding multimodal information significantly enhances SR performance, it introduces computational overhead. For instance, using Gemini Flash for image captioning results in a throughput of 0.34 images per second, which is slower than depth at " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.99\\mathrm{img / s}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", semantic segmentation at " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "2.09\\mathrm{img / s}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", and DDIM sampling at " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "0.54\\mathrm{img / s}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ". However, cross-modal predictions can be parallelized, and our method achieves speeds comparable to other text-driven SR models [68, 71]. Future work will explore optimizing the vision-language component for faster inference and investigate more robust modules for extracting modality-specific information, potentially enhancing performance even with noisy or incomplete inputs." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23148" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "Vishal M. Patel was supported by NSF CAREER award 2045489. We are grateful to Keren Ye and Mo Zhou for their valuable feedback. We also extend our gratitude to Shlomi Fruchter, Kevin Murphy, Mohammad Babaeizadeh, and Han Zhang for their instrumental contributions in facilitating the initial implementation of the latent diffusion models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 195, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 296, + 239 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "text", + "content": "[2] Roman Bachmann, Oğuzhan Fatih Kar, David Mizrahi, Ali Garjani, Mingfei Gao, David Griffiths, Jiaming Hu, Afshin Dehghan, and Amir Zamir. 4M-21: An any-to-any vision model for tens of tasks and modalities. NeurIPS, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 286, + 296, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 296, + 329 + ], + "type": "text", + "content": "[3] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan L Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. In CVPR, pages 22861-22872, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 296, + 385 + ], + "type": "text", + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2023.1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 294, + 409 + ], + "type": "text", + "content": "[5] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In CVPR, pages 6228-6237, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 411, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 442 + ], + "type": "text", + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 445, + 296, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 488 + ], + "type": "text", + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, pages 14455-14465, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 491, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 544 + ], + "type": "text", + "content": "[8] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Real-world blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 546, + 296, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 296, + 611 + ], + "type": "text", + "content": "[9] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25857-25867, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 657 + ], + "type": "text", + "content": "[10] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[11] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[12] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. arXiv preprint arXiv:2310.01110, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 149 + ], + "type": "text", + "content": "[13] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 151, + 555, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 555, + 193 + ], + "type": "text", + "content": "[14] Maurizio Delbracio and Peyman Milanfar. Inversion by direct iteration: An alternative to denoising diffusion for image restoration. Transactions on Machine Learning Research, 2023. Featured Certification. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 555, + 248 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 555, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 303 + ], + "type": "text", + "content": "[16] Keyan Ding, Yi Liu, Xueyi Zou, Shiqi Wang, and Kede Ma. Locally adaptive structure and texture similarity for image quality assessment. In Proceedings of the 29th ACM International Conference on multimedia, pages 2483-2491, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 555, + 348 + ], + "type": "text", + "content": "[17] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 2015. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "text", + "content": "[18] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 394, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 555, + 448 + ], + "type": "text", + "content": "[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 449, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 555, + 491 + ], + "type": "text", + "content": "[20] Kanchana Vaishnavi Gandikota and Paramanand Chandramouli. Text-guided explorable image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25900-25911, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 493, + 555, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 555, + 546 + ], + "type": "text", + "content": "[21] Jiatao Gu, Ying Shen, Shuangfei Zhai, Yizhe Zhang, Navdeep Jaitly, and Joshua M Susskind. Kaleido diffusion: Improving conditional diffusion models with autoregressive latent modeling. Advances in neural information processing systems, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 548, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 555, + 601 + ], + "type": "text", + "content": "[22] Jiatao Gu, Yuyang Wang, Yizhe Zhang, Qihang Zhang, Dinghuai Zhang, Navdeep Jaitly, Josh Susskind, and Shuangfei Zhai. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "text", + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 637, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 555, + 689 + ], + "type": "text", + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23149" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 297, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 297, + 105 + ], + "type": "text", + "content": "perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 297, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 297, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 297, + 150 + ], + "type": "text", + "content": "[26] Anita Kamath, Jack Hessel, and Kai-Wei Chang. What's \"up\" with vision-language models? investigating their struggle with spatial reasoning. In EMNLP, pages 9161–9175, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "text", + "content": "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 252 + ], + "type": "text", + "content": "[28] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning, pages 3744-3753. PMLR, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "text", + "content": "[29] Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image super-resolution. In Proceedings of the European conference on computer vision (ECCV), pages 517-532, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 296, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 296, + 353 + ], + "type": "text", + "content": "[30] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 408 + ], + "type": "text", + "content": "[31] Jie Liang, Hui Zeng, and Lei Zhang. Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 454 + ], + "type": "text", + "content": "[32] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision. Springer, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 499 + ], + "type": "text", + "content": "[33] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diffbir: Towards blind image restoration with generative diffusion prior. ArXiv preprint, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "text", + "content": "[34] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 556, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 589 + ], + "type": "text", + "content": "[35] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 624 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "text", + "content": "[37] Kangfu Mei, Vishal M Patel, and Rui Huang. Deep semantic statistics matching (d2sm) denoising network. In European Conference on Computer Vision, pages 384-400. Springer, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[38] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image generation. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 9048-9058, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 148 + ], + "type": "text", + "content": "[39] Kangfu Mei, Luis Figueroa, Zhe Lin, Zhihong Ding, Scott Cohen, and Vishal M Patel. Latent feature-guided diffusion models for shadow removal. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4313-4322, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 555, + 194 + ], + "type": "text", + "content": "[40] Kangfu Mei, Zhengzhong Tu, Mauricio Delbracio, Hossein Talebi, Vishal M Patel, and Peyman Milanfar. Bigger is not always better: Scaling properties of latent diffusion models. arXiv preprint arXiv:2404.01367, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 555, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 555, + 258 + ], + "type": "text", + "content": "[41] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 293 + ], + "type": "text", + "content": "[42] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 555, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 336 + ], + "type": "text", + "content": "[43] David Mizrahi, Roman Bachmann, Oğuzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4M: Massively multimodal masked modeling. NeurIPS, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 555, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 380 + ], + "type": "text", + "content": "[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 555, + 448 + ], + "type": "text", + "content": "[45] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 555, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 492 + ], + "type": "text", + "content": "[46] Mehdi Noroozi, Isma Hadji, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. You only need one step: Fast super-resolution with stable diffusion via scale distillation. In European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 494, + 555, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 555, + 547 + ], + "type": "text", + "content": "[47] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 555, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 590 + ], + "type": "text", + "content": "[48] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "text", + "content": "[49] Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi. Spire: Semantic prompt-driven image restoration. In European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 689 + ], + "type": "text", + "content": "[50] Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou. Xpsr: Cross-modal priors for diffusion-based image super-resolution. In European Conference on Computer Vision, pages 285-303. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[51] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23150" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 297, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 297, + 105 + ], + "type": "text", + "content": "Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 297, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 297, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 297, + 149 + ], + "type": "text", + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "text", + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 270 + ], + "type": "text", + "content": "[54] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 296, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 296, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 296, + 325 + ], + "type": "text", + "content": "[55] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 381 + ], + "type": "text", + "content": "[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 297, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 297, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 297, + 437 + ], + "type": "text", + "content": "[57] Li-Yuan Tsao, Yi-Chen Lo, Chia-Che Chang, Hao-Wei Chen, Roy Tseng, Chien Feng, and Chun-Yi Lee. Boosting flow-based generative super-resolution models via learned prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26005–26015, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 296, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 469 + ], + "type": "text", + "content": "[58] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "text", + "content": "[59] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 2, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 296, + 548 + ], + "type": "text", + "content": "[60] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "type": "text", + "content": "[61] Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, and Shanshan Zhao. Caption anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 647 + ], + "type": "text", + "content": "[62] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "type": "text", + "content": "[63] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Realesrgan: Training real-world blind super-resolution with pure synthetic data supplementary material. Computer Vision Foundation open access, 2022. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[64] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 118 + ], + "type": "text", + "content": "Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 172 + ], + "type": "text", + "content": "[65] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 172, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 555, + 228 + ], + "type": "text", + "content": "[66] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "text", + "content": "[67] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 293, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 554, + 338 + ], + "type": "text", + "content": "[68] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 555, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 404 + ], + "type": "text", + "content": "[69] Zhixiong Yang, Jingyuan Xia, Shengxi Li, Xinghua Huang, Shuanghui Zhang, Zhen Liu, Yaowen Fu, and Yongxiang Liu. A dynamic kernel prior model for unsupervised blind image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26046-26056, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 555, + 448 + ], + "type": "text", + "content": "[70] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 448, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 448, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 448, + 555, + 514 + ], + "type": "text", + "content": "[71] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 1, 4, 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 555, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 555, + 558 + ], + "type": "text", + "content": "[72] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image super-resolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "text", + "content": "[73] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 658 + ], + "type": "text", + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[75] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23151" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 152 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 106 + ], + "type": "text", + "content": "[76] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. In European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 297, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 297, + 152 + ], + "type": "text", + "content": "[77] Zheyuan Zhang, Fengyuan Hu, Jayjun Lee, Freda Shi, Parisa Kordjamshidi, Joyce Chai, and Ziqiao Ma. Do vision-language models represent space and how? evaluating spatial frame of reference under ambiguities, 2024. 1" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23152" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_content_list.json b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1a3d4348a01507c5107523f20facb1ac64c03596 --- /dev/null +++ b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_content_list.json @@ -0,0 +1,2109 @@ +[ + { + "type": "text", + "text": "The Scene Language: Representing Scenes with Programs, Words, and Embeddings", + "text_level": 1, + "bbox": [ + 282, + 130, + 715, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunzhi Zhang $^{1}$ Zizhang Li $^{1}$ Matt Zhou $^{2}$ Shangzhe Wu $^{1}$ Jiajun Wu $^{1}$ $^{1}$ Stanford University $^{2}$ UC Berkeley", + "bbox": [ + 194, + 202, + 802, + 247 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91213d4fe6a537581af84bb20e44f900ddcb888c526fac06413df4a3706a9ef2.jpg", + "image_caption": [ + "Figure 1. Structured Scene Generation and Editing Using the Scene Language. We develop a scene representation for 3D scene generation and editing tasks. Given textual scene descriptions, the representation can be inferred by a pre-trained large language model, rendered in 3D, and edited following language instructions. The representation contains a program consisting of semantic-aware functions bound to words, providing high interpretability and an intuitive scene-editing interface, and embeddings enabling editing with fine controls, e.g., transferring the style of $>$ from a user's image to the output by updating $$ which controls global attributes of the scene. Project page: https://ai.stanford.edu/~yzzhang/projects/scene-language/." + ], + "image_footnote": [], + "bbox": [ + 101, + 257, + 406, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76cf9657f56e00cc283db7ce71606672ad889c072439cd87448fabf06070e054.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 257, + 880, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 594, + 326, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce the Scene Language, a visual scene representation that concisely and precisely describes the structure, semantics, and identity of visual scenes. It represents a scene with three key components: a program that specifies the hierarchical and relational structure of entities in the scene, words in natural language that summarize the semantic class of each entity, and embeddings that capture the visual identity of each entity. This representation can be inferred from pre-trained language models via a training-free inference technique, given text or image inputs. The resulting scene can be rendered into images using traditional, neural, or hybrid graphics renderers. Together, this forms an automated system for high-quality 3D and 4D scene generation. Compared with existing representations like scene graphs, our proposed Scene Language generates complex scenes with higher fidelity, while explicitly modeling the scene structures to enable precise control and editing.", + "bbox": [ + 86, + 626, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 594, + 643, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How do you describe a scene? Imagine that you just traveled to Easter Island and would like to explain to Alice the wondrous scene of Ahu Akivi: \"There are seven moai in a row, facing the same direction.\" \"What is a moai?\" Alice asked. \"A moai is a stone human figure without legs, but each also looks slightly different.\" At this point, you realize it seems difficult to precisely explain the scene using natural language alone.", + "bbox": [ + 511, + 625, + 906, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In fact, this example highlights a complete scene representation requires at least three types of complementary information: structural knowledge, which is about the joint distribution of multiple instances, like \"seven moai in a row, facing the same direction,\" most naturally described as programs; category-level semantics, which may be shared across instances, often described in words, such as \"moai\"; and instance-level intrinsics, tied to the identity of each specific object or part, such as its geometry, color, and texture, which is hard to describe but easy to recognize.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24625", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "bind \"chessboard\" (lambda (embedding ...) (union (call \"board\" ...) (call \"chess pieces\" ...)))) \n(bind \"board\" (lambda (embedding ...) (define (create-squares i) ...) (union-loop 64 create-squares))) \n(bind \"chess pieces\" (lambda (embedding ... ) (define (create-pawns i) (let* ((white-pawn (call \"pawn\" ...) (black-pawn (call \"pawn\" ...) (white-pose (translate (list (* i 0.125) 0.02 0.125)))) (black-pose (translate (list (* i 0.125) 0.02 0.75))) (white-pawn-t (transform white-pawn white-pose)) (black-pawn-t (transform black-pawn black-pose))) (union white-pawn-t black-pawn-t))) (define pieces-order (((\"rook\" \"knight\" \"bishop\" \"queen\" \"king\" \"bishop\" \"knight\" \"rook\")) (define (create-other-pieces i) (let* ((piece-order (list-ref pieces-order i)) (white-piece (call piece-order ...)) (black-piece (call piece-order ...)) (white-pose (translate (list (* i 0.125) 0.02 0)) (black-pose (translate (list (* i 0.125) 0.02 0.875))) (white-piece-t (transform white-piece white-pose)) (black-piece-t (transform black-piece black-pose))) (union white-piece-t black-piece-t))) (union (union-loop 8 create-pawns) (union-loop 8 create-other-pieces)))", + "guess_lang": "lisp", + "bbox": [ + 99, + 89, + 589, + 496 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a8e9bc56400491082bb426f615365567f7f2a4fa8bd9bd40e31057e3889ab17.jpg", + "image_caption": [ + "Figure 2. Overview. A Scene Language represents a scene with three components: a program consisting of entity functions, a set of words (e.g., pawn) denoting the semantic class of the entity functions, and a list of embeddings (e.g., ) capturing the identity of each entity in the scene. Each entity function is bound with an entity class name given by a word, and maps an input embedding to an output entity of that class. Executing the program evaluates entity functions to compute the full set of entities in the scene. The resulting computation graph, illustrated on the right, captures the dependency structure among entities (indicated by arrows). The program shown is converted from our text-conditioned inference method's output, with details included in Appendix F.2.1; it is written in Lisp-style syntax for brevity and is implemented with Python in practice (Sec. 3.2)." + ], + "image_footnote": [], + "bbox": [ + 599, + 88, + 903, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modern AI techniques provide natural grounding for each of the three modalities, while also falling short of capturing all: in-context learning of pre-trained language models (LMs) enables the inference of domain-specific programs [3]; LMs themselves capture rich semantic information based on words in natural language; and embeddings obtained via techniques like textual inversion [6] or low-rank adaptation [9] best capture object identity. However, none of these existing representations alone is sufficient for scene generation and editing.", + "bbox": [ + 88, + 611, + 482, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce the Scene Language, a representation that integrates the three modalities—programs, words, and embeddings—to precisely and concisely describe the structure, semantics, and identity of visual scenes. In the Scene Language, a program specifies a computation process that defines the organization of a collection of entities in the scene, including extrinsics like poses and structural regularity like repetitions. Each entity is associated with a word referring to its semantic group, as well as an embedding describing", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "its instance-specific attributes.", + "bbox": [ + 511, + 611, + 714, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to the representation itself, we propose a training-free inference module using a pre-trained LM as a backbone to infer the Scene Language from texts and images. When provided with a domain-specific language (DSL) for scenes, LMs decompose the task of complex scene generation into simpler tasks of scene component generation by predicting their corresponding modular functions. We also discuss possible neural, traditional, and hybrid graphics engines that render images based on the representation. Together, the Scene Language, the inference module, and the renderer lead to a system for high-quality, detailed 3D and 4D scene generation and editing.", + "bbox": [ + 511, + 627, + 906, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include", + "bbox": [ + 531, + 810, + 787, + 823 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A scene representation, the Scene Language, capturing structure, semantics, and identity of visual scenes using programs, words, and embeddings.", + "2. A training-free method that infers the representation from texts and/or images using pre-trained LMs." + ], + "bbox": [ + 511, + 825, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24626", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3. A generic rendering module that renders the Scene Language into an image.", + "4. Empirical results on text- and image-prompted scene generation and editing tasks." + ], + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 164, + 230, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual scene representations are arguably the most fundamental problem in computer vision; thus, for sure, we may not enumerate all related work. As our Scene Language comprises programs, words, and embeddings, we organize our discussion accordingly into three categories: scene representations using program-based representations (Sec. 2.1), semantic graph-based representations (Sec. 2.2), and a pre-trained generative model's latent space (Sec. 2.3).", + "bbox": [ + 89, + 189, + 483, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Representing Scenes as Programs", + "text_level": 1, + "bbox": [ + 89, + 319, + 383, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Programs can specify not only the relations among scene components mentioned in Sec. 2.2, but also structural patterns such as hierarchy and repetitions, making them suitable as explicit descriptions of scene structures. Prior works have proposed to use programs in the form of sequences of execution commands as object-centric representations, followed by neural executors that render the programs into 3D shapes [5, 30, 36]. In comparison, ShapeAssembly [17] introduces higher-level functions with semantically meaningful function names, e.g., \"chair\" and \"back\", to its program representation. Both ShapeAssembly and ours adopt the design principle of function abstraction, which results in clearly stated hierarchy relation among components and better program editability. However, ShapeAssembly uses cuboids as the shape representation and does not model appearance, while ours allows for more precise geometry and appearance modeling using expressive neural embeddings.", + "bbox": [ + 88, + 340, + 482, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All representations above require 3D datasets for training. More recently, with the advance of language models (LMs), several methods [10, 33, 34, 37, 39, 43] have proposed to use zero-shot LM inference for generating programs that will be rendered into scenes. These methods operate on top of program syntax from specific graphics renderers such as Blender [4], and they do not permit parameters in high-dimensional embedding spaces unlike ours.", + "bbox": [ + 89, + 598, + 482, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Representing Scenes with Semantic Graphs", + "text_level": 1, + "bbox": [ + 89, + 727, + 459, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior semantic scene representations often adopt a graph to encode semantic scene components, such as objects and parts. In particular, [11, 38] propose to employ a parse graph of context-free grammar, using terminal nodes to correspond to objects and their attributes, to represent a scene. Both works employ an analysis-by-synthesis approach to infer the representation from images that heavily rely on domain-specific priors. Alternative representations include scene graph [7, 15, 16], where each node in a graph corresponds to an object and an edge corresponds to a pairwise", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/856fe8cbee4c4db56eeec65b3a52f0bbb778a57473222e65e0279d1ff32a259a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Sec. 3.1Sec. 3.2Definition
Expressions
ΨtransformtransformTransform an entity
ΨunionunionCompose entities
fw:z,γ→h<entity FUNC>Mapping embeddings to entity
fw(z,γ)(call <word>)Entity function application
Data Types
wWordDescribing semantics
tMatrixEntity pose
zEmbeddingSpecifying entity identity
γList [Embedding]Specifying descendant entities' identities
hEntityAn entity
sEntityThe represented scene
", + "bbox": [ + 516, + 89, + 903, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Summary of Notations in Secs. 3.1 and 3.2.", + "bbox": [ + 547, + 263, + 870, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "relation, and StructureNet [22], which focuses on an object-centric setting and uses nodes for object parts. While these representations preserve the high-level semantics of scenes or objects, they leave out low-level precision; thus, geometric, textural, or relational details that cannot be fully specified by language or hand-crafted rules are often ignored. We address this issue via the inclusion of embeddings.", + "bbox": [ + 511, + 290, + 903, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Representing Scenes with Generative Model Latents", + "text_level": 1, + "bbox": [ + 511, + 409, + 903, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The latent space of visual generative models can serve as a representation space for visual scenes. Such latent space can effectively capture the exact visual content of scenes, including geometry and appearance details, and can be either directly inferred, e.g., in variational inference [19] and model inversion [45]. More recently, text-to-image diffusion models have shown remarkable results in image synthesis. This class of models offers several candidate representation spaces including the space of textual embeddings [6], low-rank network weights [9], full model weights [29], or noise vectors in the diffusion process [8, 24, 32]. However, such representations typically do not offer interpretable semantics or explicitly encode hierarchical scene structures unlike ours.", + "bbox": [ + 511, + 448, + 906, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Scene Language", + "text_level": 1, + "bbox": [ + 511, + 676, + 710, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to design a visual scene representation that encodes the structure, semantics, and visual content of scenes. Towards this goal, we propose the Scene Language, which represents a scene with three components: a program that encodes scene structure by specifying the existence and relations of scene components, which we will refer to as entities; words in natural language that denote the semantic group of each entity in the scene; and neural embeddings that pertain the low-level visual details and identities of the entities by permitting an expressive input parameter space. In the following, we will first give a formal definition of the representation (Sec. 3.1), and then introduce a domain-specific language (DSL) (Sec. 3.2) as its realization.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24627", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/886bb5c93805d49be452cb8203f429239c72eadd2c3abbae32ccf42ee8a61330.jpg", + "image_caption": [ + "Figure 3. Rendering. Given a Scene Language in (a), a program interpreter executes the program to obtain a data object in (b). A graphics renderer first reparameterizes the data object from (b) into the renderer-specific parameter space, and then executes the rendering operation $\\mathcal{R}$ to obtain the final image output in (d)." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 898, + 166 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1dc539938ff93fbf8fed8ab0afa7d16cc8219345c552ee4b184c3052808b120c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RendererExamples
Rendering Operation RParameters from Θg_reparam
Primitive-basedLight transportShape and BRDFLM inference
Asset-basedRay tracingAsset metadataLM inference
SDS-basedGaussian splattingGaussian parametersSDS optimization
T2I modelFeed-forward passText embeddings in ZCLIP encoding
", + "bbox": [ + 93, + 218, + 478, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Graphics Renderers Examples.", + "bbox": [ + 160, + 304, + 413, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Formal Definition", + "text_level": 1, + "bbox": [ + 89, + 325, + 264, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Scene Language for a scene $s$ , denoted as $\\Phi(s)$ , is formally defined as follows:", + "bbox": [ + 89, + 347, + 482, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi (s) := (W, P, Z). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 387, + 482, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $W$ is a collection of phrases in natural language, referred to as words, e.g., $W = \\{\\text{board}, \\text{pawn}, \\dots\\}$ as illustrated in Fig. 2, and $P := \\{f_w\\}_{w \\in W}$ is a program consisting of a set of entity functions $f_w$ indexed by $w \\in W$ . Each entity function $f_w$ defines a class of entities in the scene; it is uniquely identified by the associated $w \\in W$ , which concisely summarizes the semantic meaning of the defined class. Each of $f_w$ maps neural embeddings $(z, \\gamma)$ to a specific entity $h$ in the scene, where $z$ specifies the attributes and identity of $h$ , like a specific color of a pawn, and $\\gamma$ specifies identities of its sub-entities. Hence, the complete Scene Language $\\Phi(s)$ of a particular scene $s$ also contains an ordered set of neural embeddings $Z := [z_1, z_2, \\dots, z_J]$ corresponding to the $J$ specific entities $[h_1, h_2, \\dots, h_J]$ from scene $s$ , where $h_j \\in J$ is computed by applying an entity function on inputs $z_j \\in Z, \\gamma_j \\subset Z$ .", + "bbox": [ + 88, + 412, + 483, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Crucially, program $P$ captures scene structures in three aspects. First, each entity function $f_{w} \\in P$ transforms and composes multiple sub-entities (e.g., 64 squares) into a new, more complex entity (e.g., a board), naturally reflecting the hierarchical, part-whole relations in the scene, as the arrows in Fig. 2 highlight. Second, multiple entities $h_{j}$ 's may belong to the same semantic class $w$ (e.g., square), and can thus be represented by reusing the same entity function $f_{w}$ with distinct embeddings $z_{j}$ 's. Finally, each entity function also captures the precise spatial layout of the sub-entities by specifying their relative poses during the composition, such as 64 squares forming an $8 \\times 8$ grid.", + "bbox": [ + 88, + 655, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The following expands on how functions from $P$ are defined and the program execution procedure to compute the represented scene $s$ . Notations are summarized in Tab. 1.", + "bbox": [ + 89, + 835, + 482, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Entity Function Definitions. An entity function $f_{w}$ outputs", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an entity $h$ of semantic class $w$ . The function $f_w$ takes two inputs: an embedding $z$ that specifies $h$ 's identity, as well as an ordered set $\\gamma$ containing embeddings of all descendant entities of $h$ . Denote $\\{h^{(i)}\\}_{i=1}^N$ as the $N$ sub-entities of $h$ , where $N$ is a constant determined by $f_w$ . Then we have $\\gamma = \\{z^{(i)}, \\gamma^{(i)}\\}_{i=1}^N$ , where $z^{(i)}$ and $\\gamma^{(i)}$ are the embeddings of $h^{(i)}$ and of the descendant entities of $h^{(i)}$ , respectively. Thus, $f_w(z, \\gamma)$ is recursively defined as", + "bbox": [ + 511, + 220, + 906, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh = f _ {w} (z, \\gamma) := \\Psi_ {\\text {u n i o n}} \\left(\\left\\{\\Psi_ {\\text {t r a n s f o r m}} \\left(h ^ {(i)}, t ^ {(i)} (z)\\right) \\right\\} _ {i = 1} ^ {N}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 353, + 903, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h^{(i)} = f_{w^{(i)}}(z^{(i)},\\gamma^{(i)})$ is a sub-entity of $h$ and $t^{(i)}(z)$ specifies its pose.", + "bbox": [ + 511, + 398, + 903, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $f_{w}$ specifies the computation procedure to obtain output $h$ from $N$ sub-entities $\\{h^{(i)}\\}_{i = 1}^{N}$ via two operations: $\\Psi_{\\mathrm{transform}}$ applies an input-dependent pose $t^{(i)}(z)$ to a subentity $h^{(i)}$ , transforming it from its canonical frame to the world frame of $h$ , and $\\Psi_{\\mathrm{union}}$ composes multiple sub-entities into one single entity. Each sub-entity $h^{(i)}$ is computed by recursively applying an entity function $f_{w^{(i)}}$ also defined using Eq. (2). For instance, let $f_{w}$ denote the entity function that produces the board in Fig. 2 (namely, $w = \\mathrm{board}$ ). This function $f_{w}$ composes 64 sub-entities $\\{h^{(i)}\\}_{i = 1}^{64}$ of the same class square, which are in turn obtained by executing the same entity function $f_{w^{(i)}} = f_{\\mathrm{square}}$ with different embeddings $z^{(i)}$ and $\\gamma^{(i)}$ .", + "bbox": [ + 511, + 430, + 906, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Program Execution. To obtain a scene $s$ from the Scene Language $\\Phi(s) = (W, P, Z)$ , a program executor identifies a root entity function $f_{w_1}$ from $P$ that is not dependent by any other function (e.g., $w_1 =$ chessboard from Fig. 2), and evaluates this root function using the first embedding $z_1 \\in Z$ and the rest as its descendants, $\\gamma_1 := [z_2, \\dots, z_J] \\subset Z$ , to obtain $s = f_{w_1}(z_1, \\gamma_1)$ . Evaluating $f_{w_1}(z_1, \\gamma_1)$ expands the computation recursively to its children functions $h_j = f_{w_j}(z_j, \\gamma_j)$ as defined in Eq. (2), obtaining a full sequence of all the entities $h_j$ of the scene, where $j = 1, 2, \\dots, J, z_j \\in Z, w_j \\in W$ . The order in $Z$ corresponds to the depth-first-search of the computation graph starting from $z_1$ . An example computation graph is visualized on the right of Fig. 2.", + "bbox": [ + 511, + 628, + 908, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Scene Language as Programming Language", + "text_level": 1, + "bbox": [ + 511, + 848, + 885, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now concretize the definition in Sec. 3.1 with a domain-specific language (DSL) specified in Appendix B.1. We in-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24628", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "produce macro operations union for $\\Psi_{\\text{union}}$ , union-loop which calls union on entities evaluated in a for-loop, and transform for $\\Psi_{\\text{transform}}$ . Together with call which applies function calls, these four macro operations fully define entity functions in the DSL. Entity functions are identified with the associated words in the DSL via two special forms: bind, which binds entity function $f_w$ to word $w$ , and retrieve, which retrieves $f_w$ given $w$ and applies $f_w$ on actual embedding parameters. If $w$ is never bound to a function, it corresponds to an entity function with no sub-entities (i.e., $N = 0$ in Eq. (2)), in which case call returns a primitive entity with no children.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The data type of an entity $h$ (Eq. (2)) is denoted as Entity. It stores two data fields Word and Embedding, describing the entity's semantic group and identity respectively, and stores each children entity together with its pose in the frame of $h$ . In particular, Embedding captures the visual details and requires a highly expressive representation, e.g., neural embeddings. We employ the textual embedding space of OpenCLIP-ViT/H [13] for parameterization, denoted as $\\mathcal{Z}_{\\mathrm{CLIP}}$ . It offers the advantage that embeddings can be either encoded directly from natural language or inferred from images with Textual Inversion [6]. Tab. 1 summarizes the expressions and data types in accordance with the notations from Sec. 3.1.", + "bbox": [ + 89, + 273, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Rendering", + "text_level": 1, + "bbox": [ + 89, + 487, + 205, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Applying the proposed scene representation to image generation tasks requires rendering a Scene Language $\\Phi(s)$ into images. To do so, first, the program interpreter evaluates $\\Phi(s)$ to obtain a data object of type Entity. Afterward, a graphics renderer maps the Entity data object to its rendering parameter space and renders it into a final image.", + "bbox": [ + 89, + 513, + 482, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Renderer Specifications.** We define the specification of a graphics renderer, a module in the proposed representation, as follows. A graphics renderer is determined by primitive parameter space $\\Theta$ and a rendering operation $\\mathcal{R} : \\mathcal{P}(\\Theta \\times \\mathcal{T}) \\to \\mathcal{I}$ , where $\\mathcal{T}$ is the space of 3D affine transformations representing poses, $\\mathcal{P}$ denotes all possible subsets, and $\\mathcal{I}$ is the space of rendered images. In order to determine a mapping from program execution outputs of type Entity (Fig. 3b) to the admissible input domain of rendering operation $\\mathcal{R}$ (Fig. 3c), we assume access to a reparameterization function $g_{\\text{reparam}}$ that maps from Tuple[Word, Embedding] to $\\Theta$ , and compute a primitive's pose $t \\in \\mathcal{T}$ by multiplying all Matrix values along the path from the primitive to the scene (root) in the entity hierarchy, analogous to computing limb poses in a kinematic tree.", + "bbox": [ + 89, + 609, + 482, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Renderer Instantiations.** An example renderer instantiation is with Score Distillation Sampling (SDS) [25] guidance, where $\\Theta$ is a differentiable 3D representation, and we specify $g_{\\mathrm{reparam}}$ as follows. Recall that an entity is as-", + "bbox": [ + 89, + 839, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sociated with a Word value, e.g., moai for one statue from Figs. 1 and 3, and an Embedding value, e.g., $$ . For each primitive entity (i.e., an entity with no children), given these two value fields together with the fields of its ancestors, we use a manually specified language template $c$ , or a $$ moai, in the style of $$ , 3D model in this example, to embed them into $z = g_{\\mathrm{CLIP}}(c) \\in \\mathcal{Z}_{\\mathrm{CLIP}}$ where $g_{\\mathrm{CLIP}}$ is the pre-trained CLIP text encoder. Then, $g_{\\mathrm{reparam}}: \\mathcal{Z}_{\\mathrm{CLIP}} \\to \\Theta$ corresponds to the SDS-guided optimization to find a solution in $\\Theta$ that aligns with the input condition $z \\in \\mathcal{Z}_{\\mathrm{CLIP}}$ . Output 3D scenes can be personalized by editing embeddings, e.g., $$ which controls the global style in Fig. 1.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the underlying 3D representation, we use 3D Gaussian Splatting [18] where images are rendered by splatting a set of 3D Gaussians onto the image plane; other differentiable 3D representations such as neural fields will also be suitable. We base our implementation on GALA3D [44], and use MVDream [31] and a depth-conditioned ControlNet [40] for guidance.", + "bbox": [ + 511, + 287, + 903, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We refer to the renderer above as the Gaussians renderer. Other possible renderers include primitive-based renderers, such as Mitsuba [14] with graphics primitives of cubes, spheres, and cylinders; asset-based game engines, such as MineCraft [23]; and feed-forward inference of layout-conditioned text-to-image (T2I) diffusion models, such as MIGC [42], which achieves 2D bounding box conditioning by controlling attention layers from Stable Diffusion [28]). Tab. 2 shows a summary; details are in Appendix B.2.", + "bbox": [ + 511, + 393, + 903, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Inference via Pre-Trained Language Models", + "text_level": 1, + "bbox": [ + 511, + 542, + 903, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a training-free method to infer the representation $\\Phi(s) = (W, P, Z)$ from text or image descriptions of a scene $s$ . As explained below, we first prompt a pre-trained language model (LM) to generate the non-neural components $(W, P)$ and then obtain neural embeddings $(Z)$ from texts via the CLIP text encoder or from images with a pretrained text-to-image diffusion model.", + "bbox": [ + 511, + 568, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LMs have shown remarkable capability in code generation with common programming languages such as Python. In our implementation, we prompt LMs to generate Python scripts. We prompt the LM with the input condition, i.e., a scene description in texts or an image; a Python script of helper functions converted from the DSL in Sec. 3.2; and an example script using the helper functions. We use Claude 3.5 Sonnet [1] for all experiments for our method and LM-dependent baselines. Full LM prompts are in Appendix F.1.", + "bbox": [ + 511, + 674, + 905, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Function arguments in the LM-generated scripts, which are numeric values or string tokens, are converted to embeddings from $\\mathcal{Z}_{\\mathrm{CLIP}}$ (Sec. 3.2) using language templates and the CLIP text encoder $g_{\\mathrm{CLIP}}$ . For example, in the raw LM output, function calls for white pieces in Fig. 2 have input attribute {\"color\": (.9, .9, .9)}}, and we prompt LM", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24629", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e27eb4cab15f145112aa840512c15b6c229eddc8217d5dc5227acbb3467ede56.jpg", + "image_caption": [ + "Ours (Two Views)" + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 254, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4af557c8345bc77887bca7c4a53291a0ea45a466639a8173dc63102e93cd1cf8.jpg", + "image_caption": [ + "MVDream" + ], + "image_footnote": [], + "bbox": [ + 277, + 95, + 346, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08e1d50ea05369dce00e693a6569a3526b2c2d929149ba051e5a307926eba983.jpg", + "image_caption": [ + "GraphDreamer" + ], + "image_footnote": [], + "bbox": [ + 359, + 109, + 457, + 185 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/69df59f0bd149f44a98e6ba74d3c0f69ebaabe3c8b10358f48bba40aaab9639c.jpg", + "image_caption": [ + "Ours (Two Views)" + ], + "image_footnote": [], + "bbox": [ + 114, + 213, + 169, + 299 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e447ef0eb88c727a2fb6beb46bff05b1108c8cc4607f1361eb8b443f9beca79.jpg", + "image_caption": [ + "\"An 8-layer shelf\"" + ], + "image_footnote": [], + "bbox": [ + 192, + 213, + 243, + 303 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17dcbc9225b699fde1748a2d9cac5a1e34e40a64b2abff6dcd6d9f63f2b7717a.jpg", + "image_caption": [ + "MVDream" + ], + "image_footnote": [], + "bbox": [ + 264, + 224, + 356, + 305 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d961f9ae50dd92a01ebdaec81324b0ea69724b0c7e6d431a0f12720b69cb1814.jpg", + "image_caption": [ + "GraphDreamer", + "\"15 coke cans stacking in a pyramid\"", + "Figure 4. Text-Prompted Scene Generation. Ours produce more accurate outputs than using no intermediate representation (MV-Dream) or scene graph (GraphDreamer)." + ], + "image_footnote": [], + "bbox": [ + 364, + 223, + 460, + 306 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/53990c41bb795ad949f8eb246e2bc67b8311d6bc6b8311663585dca80642a19d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsNumeric PromptsGeneric Prompts
Alignment (↑)CLIP (↑)Count (↑)Alignment (↑)CLIP (↑)
GraphDreamer [7]3.56±7.380.297±0.0850.1115.75±20.390.230±0.097
MVDream [31]10.79±12.830.312±0.0460.1135.13±31.900.328±0.031
Ours85.65±13.710.351±0.0511.052.13±26.830.328±0.030
", + "bbox": [ + 91, + 390, + 480, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to describe the color value as a word, and feed the word into $g_{\\mathrm{CLIP}}$ to compute $< z68 >$ . For image-conditioned tasks, for each primitive entity in the execution output of $P$ , we first use GroundingSAM [20, 27] to segment out the region defined by the word associated with the entity. We then use Textual Inversion [6] to optimize an embedding to reconstruct the cropped image with the diffusion model training objective. The full process is deferred to Appendix C.2.", + "bbox": [ + 89, + 547, + 483, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 684, + 223, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Text-Prompted Generation and Editing", + "text_level": 1, + "bbox": [ + 89, + 708, + 429, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. To evaluate our representation in text-prompted 3D generation task, we compare our inference pipeline with 3D scene generation methods using alternative intermediate representations, e.g., scene graph. In particular, we compare with GraphDreamer [7] as an exemplar approach, which generates scene graphs from texts via LM prompting and then synthesizes scenes conditioned on the graphs via SDS guidance. We further ablate the role of structural representation in this task by comparing ours with the backbone of our SDS-based renderer, MVDream [31], as a direct scene generation approach. Details are in Appendix C.3.", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3a0dbd181d380a934c8da1ddd1c14df554101b4cc75b27b8bf985ee8c1656ad.jpg", + "image_caption": [ + "rnnnne nnnnne" + ], + "image_footnote": [], + "bbox": [ + 542, + 95, + 630, + 125 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0cad2d9bdb96aea4db1e662d978c3cbade744f8ba1811b90ee12e0c9a922329f.jpg", + "image_caption": [ + "Gaussians" + ], + "image_footnote": [], + "bbox": [ + 542, + 132, + 630, + 185 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf75a7d2b2bd68f2e3c7bf77b9f3991a652d194a65c58e9514a1f17cd5a83443.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 95, + 720, + 125 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62f64907d5772a0d50b7faa55a7591ff8d4868ceb72f11c899de8cb273cf4cbd.jpg", + "image_caption": [ + "Minecraft" + ], + "image_footnote": [], + "bbox": [ + 640, + 130, + 720, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model of", + "bbox": [ + 647, + 193, + 694, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a tennis court\"", + "bbox": [ + 586, + 205, + 674, + 217 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53a0ac878901b617508688e452da050daec5f8b222c8d07ac2909851d6e6ee60.jpg", + "image_caption": [ + "nne nnnnne", + "Mitsuba", + "\"City skyline at sunset\"", + "Figure 5. Renderings Across Graphics Renderers. Different renderers produce renderings that adhere to the same representation and are therefore visually aligned, while each exhibits a different imaging style. Text inputs are shown at the subfigure bottoms." + ], + "image_footnote": [], + "bbox": [ + 545, + 223, + 629, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ebaba16252602202b24c74d24aaf78a0e00d640fb58f4d0c1c45f2c05a7ed21.jpg", + "image_caption": [ + "\"A detailed model of a tennis court\"", + "T2I" + ], + "image_footnote": [], + "bbox": [ + 638, + 222, + 722, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e13064e231a80ebe12101c498c7a1a270393d4a64891166b836853c08b57ea2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 92, + 812, + 131 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f37b8c6405b561fb401bb003fd77b7d99ceaf19c55195633f0008a2eedfd28b.jpg", + "image_caption": [ + "Gaussians" + ], + "image_footnote": [], + "bbox": [ + 730, + 133, + 812, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/171ca8f7d2774261606cbe2fce8a34b241888d6ff61285a48ed75b01c7331def.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 90, + 900, + 132 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7d94cd6445d432d89f038748b81666e3686adcc049597e1f597bac16faab3ec.jpg", + "image_caption": [ + "Minecraft" + ], + "image_footnote": [], + "bbox": [ + 821, + 133, + 895, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gaussians", + "bbox": [ + 740, + 181, + 800, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20989f2745383b633cb4f968f4a060e28ba844e04dc620d60fb36de9eeec3f4f.jpg", + "image_caption": [ + "\"An interior scene of a university lecture hall\"", + "Mitsuba" + ], + "image_footnote": [], + "bbox": [ + 738, + 223, + 777, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df7a2c73d654451176e8738b0c791f7d0eb1161195b6b7dd4339c600201e23c1.jpg", + "image_caption": [ + "T2I", + "\"Marionette\"" + ], + "image_footnote": [], + "bbox": [ + 816, + 223, + 898, + 268 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8ccc19108a6f0016e798592103d08467d02044b1345d1546aadae33fb07a11c9.jpg", + "table_caption": [ + "Table 3. Evaluation for text-prompted 3D generation. We report the following metrics averaged across 9 numeric and 8 generic prompts: the percentages of user preferences for prompt alignment, CLIP similarity scores, and counting accuracy (numeric prompts only, 0 for inaccurate and 1 for accurate). Here $\\pm$ denotes standard deviation across scenes." + ], + "table_footnote": [], + "table_body": "
MethodsCLIP [13] (↑)Dynamic [12] (↑)
4D-fy [2]0.3520.2%
Ours0.3415.9%
", + "bbox": [ + 516, + 371, + 728, + 417 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d948eaa94260b88cb357dac0e745e1750781dd2eee9a3fcbe13b5e44c252c759.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLPIPS [41] (↓)
GraphDreamer [7]0.811
Ours0.681
", + "bbox": [ + 733, + 372, + 903, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Evaluation. Left: For text-prompted 4D generation, 4Dfy achieves higher prompt alignment measured with CLIP similarity but has small motion. Right: For image-prompted scene generation, our results are better aligned with input image prompts compared to GraphDreamer.", + "bbox": [ + 511, + 421, + 905, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We conduct a user study where 103 users are asked to choose one of the three videos, showing 360-renderings from ours and two baselines in a randomized order, that aligns the best with text prompts. Following [7], for each scene, we compute the cosine similarity of CLIP [13, 26] embeddings between a rendering and the text prompt and report the average over 120 viewpoints. We further report whether outputs have the correct object counts. Details are deferred to Appendix C.4.", + "bbox": [ + 511, + 502, + 905, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. Tab. 3 shows that our method achieves a more favorable prompt alignment than the baselines as measured in the user study and CLIP similarity, and has a clear advantage in counting accuracy.", + "bbox": [ + 511, + 642, + 905, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4 and Appendix D.1 show qualitative results of ours using the SDS-based renderer. Compared to the direct 3D scene generation method MVDream, our approach is compositional and adheres more closely to input prompts in scenes involving multiple objects. Compared to a scene graph representation, where entity relations are restricted to be between two objects and are bottlenecked by the coarseness of natural language descriptions, e.g., \"aligned in a row\", a program-based representation offers more flexible and precise specifications for relations, e.g., the particular coke can arrangement in Fig. 4. This brings the practical benefit of offloading the burden of generating scenes involving complex entity relations from the T2I model (used", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24630", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f49a812c2ecb43a64bde2247baa32f8b5545f3592c0b9a38d8cb63747a3176b9.jpg", + "image_caption": [ + "Ours (w/ tracking)" + ], + "image_footnote": [ + "\"Carousel with a small canopy\"" + ], + "bbox": [ + 94, + 94, + 207, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c20f40760c1fbe7feef3c9407018a4359bf2a91d7bc20f6e1f781eb8c37e55bb.jpg", + "image_caption": [ + "Frame 1" + ], + "image_footnote": [ + "\"Solar system model\"" + ], + "bbox": [ + 209, + 94, + 303, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c350ffa2e0ebbfebe499673b10061231a397dc4e190c8adfd9282befa846aa5.jpg", + "image_caption": [ + "1 Fran" + ], + "image_footnote": [], + "bbox": [ + 305, + 94, + 403, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/afbf2b027051b852cb9ba9f63bbaff50eb9545b7f1cd4a85d1b31e92560f88d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 118, + 542, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc8404f3abf0418a4da2d4220b055787b6237181129b3ed92ce364883bfeeac1.jpg", + "image_caption": [ + "Ours (w/ tracking)", + "4D-fy" + ], + "image_footnote": [ + "\"A toy wind turbine\"" + ], + "bbox": [ + 547, + 126, + 689, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7910274de014f94814f54f0f04fc564cb16421ef97fb1ef2c9f05e13c522cc53.jpg", + "image_caption": [ + "Frame 1 Frame 2", + "Ours Frame 1 Frame 2 (w/ tracking) 4D-fy" + ], + "image_footnote": [], + "bbox": [ + 694, + 95, + 890, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a787646a9aaaa304bb65bf72cbbe71162ed1e6b6c87319fa4b6843b27344f8e8.jpg", + "image_caption": [ + "Figure 6. Text-Prompted 4D Scene Generation. The proposed representation may also capture scene dynamics and be applied for synthesizing 4D scenes. It explicitly represents the temporal correspondence of an entity in a dynamic scene. Each colored trajectory denotes tracking of a temporally moving point. Compared to 4D-fy, our method produces more salient motion.", + "Initial (Mitsuba)" + ], + "image_footnote": [ + "\"Bramante Staircase," + ], + "bbox": [ + 116, + 295, + 223, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e469e9debb3146c85c5714ef38ed26a94b415c7ae091a9902459f73936ba649c.jpg", + "image_caption": [ + "Initial (Gaussians)" + ], + "image_footnote": [], + "bbox": [ + 277, + 296, + 377, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/420fa821f36b296c62832668367680129116b5b873c2fbc42610d2023388549c.jpg", + "image_caption": [ + "Initial (Mitsuba)" + ], + "image_footnote": [ + "\"An 8-layer 3-color Jenga set at game start\"" + ], + "bbox": [ + 117, + 455, + 225, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7494e9cb6314093d089f31e4001fdb8dbc6729ece29f9a9f093764aa84fa2726.jpg", + "image_caption": [ + "Initial (Gaussians)" + ], + "image_footnote": [ + "\"Remove 2 blocks from second top layer and the tower should not fall\"" + ], + "bbox": [ + 279, + 455, + 383, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a12ecbca67a42acbab6150493ac5546c4f5a492eaaff780007d09fc1b3a17559.jpg", + "image_caption": [ + "Initial (Mitsuba)" + ], + "image_footnote": [ + "\"A fractal tree-like structure\"" + ], + "bbox": [ + 106, + 604, + 235, + 695 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3233401808ce1fe1083368d3b2ed9e4b3fedcf21a68455a3f6f3b3204703799e.jpg", + "image_caption": [ + "Initial (Gaussians)" + ], + "image_footnote": [], + "bbox": [ + 264, + 604, + 392, + 694 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/610909dad8a134a295bd73cb6e62fd5d10926697104846242a3b77f7868e23a2.jpg", + "image_caption": [ + "Scene Language" + ], + "image_footnote": [ + "\"Shrink staircase radius by $80\\%$" + ], + "bbox": [ + 419, + 296, + 584, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfff393f4519e1ea777c9af8969be3cc683240a5bccea7b6b28a8b59e8dead52.jpg", + "image_caption": [ + "Edited (Mitsuba)" + ], + "image_footnote": [], + "bbox": [ + 647, + 295, + 681, + 409 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8306f9060792f238a687a5bc153c62003f76daaeec9fb37322978af119114bd6.jpg", + "image_caption": [ + "Edited (Gaussians)" + ], + "image_footnote": [], + "bbox": [ + 803, + 295, + 839, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43d44208cd4c2899eaf287343dddd7127d1e00006e4e749859b7d03246b7bbbd.jpg", + "image_caption": [ + "(\"Jenga layer\" (lambda(is-horizontal colors remove-blocks)))" + ], + "image_footnote": [], + "bbox": [ + 424, + 454, + 578, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15e6717a6f34304ee89a4456267eb9d2183c7c252d2df8941f6474e8490f6c73.jpg", + "image_caption": [ + "Scene Language", + "Edited (Mitsuba)" + ], + "image_footnote": [], + "bbox": [ + 614, + 455, + 722, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/577780964a6c7f543880c084b480fe9d1fb2a0207168620598a031e91d056514.jpg", + "image_caption": [ + "Edited (Gaussians)" + ], + "image_footnote": [], + "bbox": [ + 774, + 455, + 879, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9ec9d0a675e0861d0dc26ec7198490cacfb05b2712be6cebaed6a5d9be83a5f.jpg", + "image_caption": [ + "union left-branch right-branch)", + "Scene Language", + "Figure 7. Scene Editing with Language Instructions. Our representation is editing-friendly. Bottoms of each row show initial scene descriptions and editing instructions in the format of user text prompts. We prompt an LM to infer the initial Scene Language as well as the edits (shown with texts in red), and show image renderings with two renderers." + ], + "image_footnote": [ + "\"Make branching structure to be trinary and 3D\"" + ], + "bbox": [ + 429, + 642, + 542, + 710 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee0dc5c0de5a5971fd3edb07b78f9de8f6365f884e8932c76844a4a293ed31b0.jpg", + "image_caption": [ + "Edited (Mitsuba)" + ], + "image_footnote": [], + "bbox": [ + 622, + 609, + 699, + 690 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d14d9edc0360e884bedfe29ddbbb652449c9d3978f5fa9f45216776f8300e5d7.jpg", + "image_caption": [ + "Edited (Gaussians)" + ], + "image_footnote": [], + "bbox": [ + 781, + 609, + 857, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for SDS guidance in both ours and GraphDreamer) towards LM, leading to accurate and detailed generation results.", + "bbox": [ + 89, + 801, + 480, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Across Graphics Renderers. Fig. 5 shows that our representation can be rendered with different renderers (Sec. 4), producing aligned outputs in different visual domains.", + "bbox": [ + 89, + 835, + 482, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4D Generation. The inference method from Sec. 5 can be", + "bbox": [ + 89, + 885, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "applied to 4D scene generation. The 4D scene representation is identical to Eq. (1), except that there is an additional 4D entity function in the program $P$ .", + "bbox": [ + 511, + 801, + 905, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare with 4D-fy [2] as an exemplar text-to-4D generation method. We measure the prompt alignment of output dynamic renderings using CLIP similarity between", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24631", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee2fa90caf16d39cd01114c508865773a30df9a29d082d0207d3eb25816d7e28.jpg", + "image_caption": [ + "Figure 8. Scene Editing with Image Instructions. This figure shows style transfer results under the same setup as in Fig. 1. Replacing embeddings of target entities in a Scene Language with a new embedding specified by a user image achieves the editing effect." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 908, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f2cb0b9931875f3420055327cf07469e908616130d0ec185f2f602efc9df62a4.jpg", + "image_caption": [ + "Ours (Gaussians)" + ], + "image_footnote": [], + "bbox": [ + 94, + 253, + 187, + 308 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d0ff618feb55adc6b7440bfa611a08f3f04418c1b652837e999352585f4e75c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 258, + 274, + 304 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82e01fea7ee36cf75c2f365ecbf2039bfa9af69070e5447f60c682e5c5bf09cb.jpg", + "image_caption": [ + "Ours (Mitsuba)" + ], + "image_footnote": [], + "bbox": [ + 305, + 258, + 372, + 304 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/554887df68e7ad17a2fa6c4436f9e403fc9e530523f6d0eb3c7bbe3d2553425d.jpg", + "image_caption": [ + "GraphDreamer" + ], + "image_footnote": [], + "bbox": [ + 398, + 253, + 470, + 310 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9037c5d4e535ced3a3e62ee70bdb935fcd0250f3045d676182baaa100591ca5.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 94, + 323, + 187, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/026dc445a1bcdfe81d8f38f050465c7e08a37a8b293e8172abcbeeaa9af88dd2.jpg", + "image_caption": [ + "Input Image", + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 106, + 392, + 176, + 433 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b9c69da6988cc6880d39835d2b1d181b7130f140e9686040476ada7e6076ce0f.jpg", + "image_caption": [ + "Ours (Gaussians)" + ], + "image_footnote": [], + "bbox": [ + 191, + 339, + 285, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7b557b2f0107c4970eeaa80279f0ee5aca402c9d72e650ba029d7fb95357dcd.jpg", + "image_caption": [ + "Ours (Mitsuba)", + "GraphDreamer" + ], + "image_footnote": [], + "bbox": [ + 292, + 339, + 387, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8eca0563e95db2002d7dc86d56ba18b7f95f9c233482e98ec050e60d3f1c8b10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 335, + 459, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c47dda6ab1505373b3e2088bdb2466a99391f2e2531ccf8d396f21ac2384345.jpg", + "image_caption": [ + "Ours (Gaussians)" + ], + "image_footnote": [], + "bbox": [ + 207, + 395, + 276, + 435 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84a15ea6fcd20e7c50254e09d4b49d26b45a4cd70b9e5d69b0afc090e6339078.jpg", + "image_caption": [ + "Ours (Mitsuba)" + ], + "image_footnote": [], + "bbox": [ + 303, + 397, + 372, + 438 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/462c165752163db06354d52aa4e89439ba906ed2691b38e27f64e0ff2e8b1be8.jpg", + "image_caption": [ + "GraphDreamer", + "Figure 9. Image-Prompted Scene Generation. Compared to scene graphs from GraphDreamer, programs from our representation encode additional scene structure, e.g., repetitions, and specify geometric entity relations more precisely. Embeddings from ours further enable visual identity preservation in the outputs." + ], + "image_footnote": [], + "bbox": [ + 400, + 402, + 468, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rendered frames and input texts, and the motion magnitude using dynamic degrees [12] computed using RAFT [35] optical flow estimation over neighboring frames. 4D-fy results have a higher per-frame alignment score than ours but present small dynamic motions. See Fig. 6, the webpage, and Appendix D.2 for further results.", + "bbox": [ + 89, + 532, + 482, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scene Editing. Scenes synthesized from our proposed representation can further be edited by prompting LM with its previously generated script and an editing instruction in natural language instruction. Results are shown in Fig. 7. Our representation provides an interpretable and intuitive interface for scene editing, as functions have explicit semantic meanings associated with words, and function reuse significantly improves program readability. Furthermore, since the structure of programs reflects the structure of scenes, editing program parameters leads to precise changes in the scenes while preserving the original structure, e.g., the circular arrangement of staircases in Fig. 7. Desirable editing effects involving multiple primitives, or all staircases in this example, can be effectively achieved via only minor changes in the program space. Finally, the program structure itself, such as the function header in the Jenga set example, can be adjusted for editing to achieve localized edits that only affect relevant parts of the scene. Further results", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are shown in Appendix D.3.", + "bbox": [ + 511, + 253, + 702, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Image-Prompted Generation and Editing", + "text_level": 1, + "bbox": [ + 511, + 276, + 866, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Texts provide an easy user interface for scene generation as shown in Sec. 6.1, and images allow users to specify further intentions that could be hard to describe with natural language alone. Integrating neural embeddings increases the expressivity of the Scene Language and allows it to encode visual details from image prompts, and as a result, enhances our method's controllability in image-prompted scene generation and editing tasks.", + "bbox": [ + 511, + 297, + 903, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generation. Fig. 9 shows results on the task of generating 3D scenes consistent with the structure and content from a user-provided image, termed as the \"inverse semantics\" task in GraphDreamer [7]. We report the LPIPS [41] score, i.e., the AlexNet [21] feature dissimilarity, between an input image and an output rendering, averaged over 3 scenes with 20 viewpoints each in Tab. 4. Embeddings allow our representation to encode visual details from inputs, such as the Coke can texture in Fig. 9, which is not captured in scene graphs from GraphDreamer using per-object natural language descriptions. This is reflected quantitatively in Tab. 4.", + "bbox": [ + 511, + 422, + 905, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Editing. Our representation applies to image-prompted editing tasks such as style transfer (Fig. 1). Edits can target partial or full scenes as shown in Fig. 8(a) and (b), respectively, by updating embeddings for corresponding entities.", + "bbox": [ + 511, + 592, + 905, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced a visual scene representation, termed the Scene Language, which encodes three key aspects of visual scenes: scene structure, such as hierarchy and repetition, specified via programs; semantics of individual components succinctly summarized via words; and identities of each component precisely captured via neural embeddings. We formalize the representation as a programming language defined using a DSL. We show that the Scene Language can be efficiently inferred from both text and image inputs leveraging pre-trained language models. Once the program is executed, the resulting scene can be rendered into images using a variety of graphics renderers. Our Scene Language produces 3D and 4D scenes with high fidelity and enables easy and precise editing.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24632", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is in part supported by AFOSR YIP FA9550-23-1-0127, ONR N00014-23-1-2355, ONR YIP N00014-24-1-2117, ONR MURI N00014-22-1-2740, and NSF RI #2211258 and #2338203. YZ is in part supported by the Stanford Interdisciplinary Graduate Fellowship.", + "bbox": [ + 89, + 90, + 485, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 195, + 187, + 210 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anthropic. The Claude 3 model family: Opus, Sonnet, Haiku, 2024. 5", + "[2] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B. Lindell. 4D-fly: Text-to-4D generation using hybrid score distillation sampling. CVPR, 2024. 6, 7, 16", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2", + "[4] Blender Online Community. Blender, 1994. 3", + "[5] Boyang Deng, Sumith Kulal, Zhengyang Dong, Congyue Deng, Yonglong Tian, and Jiajun Wu. Unsupervised learning of shape programs with repeatable implicit parts. In NeurIPS, 2022. 3", + "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 3, 5, 6, 14", + "[7] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf. GraphDreamer: Compositional 3D scene synthesis from scene graphs. In CVPR, 2024. 3, 6, 8, 14", + "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3", + "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3", + "[10] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. SceneCraft: An LLM agent for synthesizing 3D scene as Blender code. In ICLR Workshop, 2024. 3", + "[11] Siyuan Huang, Siyuan Qi, Yixin Zhu, Yinxue Xiao, Yuanlu Xu, and Song-Chun Zhu. Holistic 3D scene parsing and reconstruction from a single RGB image. In ECCV, 2018. 3", + "[12] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 6, 8, 16", + "[13] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, 2021. 5, 6, 16" + ], + "bbox": [ + 93, + 220, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. 5", + "[15] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, 2015. 3, 16", + "[16] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, 2018. 3", + "[17] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy Mitra, and Daniel Ritchie. ShapeAssembly: Learning to generate programs for 3D shape structure synthesis. ACM TOG, 39(6):1-20, 2020. 3", + "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 5, 11", + "[19] Diederik P Kingma. Auto-encoding variational bayes. In ICLR, 2014. 3", + "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6", + "[21] Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014.8", + "[22] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas Guibas. StructureNet: Hierarchical graph networks for 3D shape generation. ACM TOG, 38 (6):1-19, 2019. 3", + "[23] Mojang Studios. Minecraft, 2009. 5", + "[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 3", + "[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 5, 12", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 13", + "[27] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded SAM: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 6, 14", + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 5, 13", + "[29] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3", + "[30] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In CVPR, 2018. 3, 16" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24633", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. In ICLR, 2024. 5, 6, 14", + "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3", + "[33] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3D-GPT: Procedural 3D modeling with large language models. arXiv preprint arXiv:2310.12945, 2023. 3", + "[34] Hou In Ivan Tam, Hou In Derek Pun, Austin T Wang, Angel X Chang, and Manolis Savva. SceneMotifCoder: Example-driven visual program learning for generating 3D object arrangements. arXiv preprint arXiv:2408.02211, 2024.3", + "[35] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 8", + "[36] Yonglong Tian, Andrew Luo, Xingyuan Sun, Kevin Ellis, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Learning to infer and execute 3D shape programs. In ICLR, 2019. 3, 16", + "[37] Yutaro Yamada, Khyathi Chandu, Yuchen Lin, Jack Hessel, Ilker Yildirim, and Yejin Choi. L3GO: Language agents with chain-of-3D-thoughts for generating unconventional objects. In ICLR Workshop, 2024. 3", + "[38] Alan Yuille and Daniel Kersten. Vision as Bayesian inference: analysis by synthesis? Trends in Cognitive Sciences, 10(7):301-308, 2006. 3", + "[39] Chi Zhang, Penglin Cai, Yuhui Fu, Haoqi Yuan, and Zongqing Lu. Creative agents: Empowering agents with imagination for creative tasks. arXiv preprint arXiv:2312.02519, 2023. 3", + "[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 5", + "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 8", + "[42] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. MIGC: Multi-instance generation controller for text-to-image synthesis. In CVPR, 2024. 5, 13", + "[43] Mengqi Zhou, Jun Hou, Chuanchen Luo, Yuxi Wang, Zhaoxiang Zhang, and Junran Peng. SceneX: Procedural controllable large-scale scene generation via large-language models. arXiv preprint arXiv:2403.15698, 2024. 3", + "[44] Xiaoyu Zhou, Xingjian Ran, Yajiao Xiong, Jinlin He, Zhiwei Lin, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. GALA3D: Towards text-to-3D complex scene generation via layout-guided generative Gaussian splatting. In ICML, 2024. 5, 12", + "[45] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In ECCV, 2016. 3" + ], + "bbox": [ + 91, + 92, + 480, + 829 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24634", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_model.json b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4a2f082f628db8dd891348e1cee75148e9a2931c --- /dev/null +++ b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_model.json @@ -0,0 +1,3168 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.284, + 0.131, + 0.717, + 0.177 + ], + "angle": 0, + "content": "The Scene Language: Representing Scenes with Programs, Words, and Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.203, + 0.803, + 0.248 + ], + "angle": 0, + "content": "Yunzhi Zhang\\(^{1}\\) Zizhang Li\\(^{1}\\) Matt Zhou\\(^{2}\\) Shangzhe Wu\\(^{1}\\) Jiajun Wu\\(^{1}\\) \n\\(^{1}\\)Stanford University \\(^{2}\\)UC Berkeley" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.258, + 0.408, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.258, + 0.882, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.501, + 0.907, + 0.586 + ], + "angle": 0, + "content": "Figure 1. Structured Scene Generation and Editing Using the Scene Language. We develop a scene representation for 3D scene generation and editing tasks. Given textual scene descriptions, the representation can be inferred by a pre-trained large language model, rendered in 3D, and edited following language instructions. The representation contains a program consisting of semantic-aware functions bound to words, providing high interpretability and an intuitive scene-editing interface, and embeddings enabling editing with fine controls, e.g., transferring the style of \\(>\\) from a user's image to the output by updating \\(\\) which controls global attributes of the scene. Project page: https://ai.stanford.edu/~yzzhang/projects/scene-language/." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.595, + 0.327, + 0.61 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.627, + 0.486, + 0.884 + ], + "angle": 0, + "content": "We introduce the Scene Language, a visual scene representation that concisely and precisely describes the structure, semantics, and identity of visual scenes. It represents a scene with three key components: a program that specifies the hierarchical and relational structure of entities in the scene, words in natural language that summarize the semantic class of each entity, and embeddings that capture the visual identity of each entity. This representation can be inferred from pre-trained language models via a training-free inference technique, given text or image inputs. The resulting scene can be rendered into images using traditional, neural, or hybrid graphics renderers. Together, this forms an automated system for high-quality 3D and 4D scene generation. Compared with existing representations like scene graphs, our proposed Scene Language generates complex scenes with higher fidelity, while explicitly modeling the scene structures to enable precise control and editing." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.595, + 0.645, + 0.61 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.746 + ], + "angle": 0, + "content": "How do you describe a scene? Imagine that you just traveled to Easter Island and would like to explain to Alice the wondrous scene of Ahu Akivi: \"There are seven moai in a row, facing the same direction.\" \"What is a moai?\" Alice asked. \"A moai is a stone human figure without legs, but each also looks slightly different.\" At this point, you realize it seems difficult to precisely explain the scene using natural language alone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In fact, this example highlights a complete scene representation requires at least three types of complementary information: structural knowledge, which is about the joint distribution of multiple instances, like \"seven moai in a row, facing the same direction,\" most naturally described as programs; category-level semantics, which may be shared across instances, often described in words, such as \"moai\"; and instance-level intrinsics, tied to the identity of each specific object or part, such as its geometry, color, and texture, which is hard to describe but easy to recognize." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24625" + } + ], + [ + { + "type": "code", + "bbox": [ + 0.1, + 0.09, + 0.59, + 0.497 + ], + "angle": 0, + "content": "bind \"chessboard\" (lambda (embedding ...) (union (call \"board\" ...) (call \"chess pieces\" ...)))) \n(bind \"board\" (lambda (embedding ...) (define (create-squares i) ...) (union-loop 64 create-squares))) \n(bind \"chess pieces\" (lambda (embedding ... ) (define (create-pawns i) (let* ((white-pawn (call \"pawn\" ...) (black-pawn (call \"pawn\" ...) (white-pose (translate (list (* i 0.125) 0.02 0.125)))) (black-pose (translate (list (* i 0.125) 0.02 0.75))) (white-pawn-t (transform white-pawn white-pose)) (black-pawn-t (transform black-pawn black-pose))) (union white-pawn-t black-pawn-t))) (define pieces-order (((\"rook\" \"knight\" \"bishop\" \"queen\" \"king\" \"bishop\" \"knight\" \"rook\")) (define (create-other-pieces i) (let* ((piece-order (list-ref pieces-order i)) (white-piece (call piece-order ...)) (black-piece (call piece-order ...)) (white-pose (translate (list (* i 0.125) 0.02 0)) (black-pose (translate (list (* i 0.125) 0.02 0.875))) (white-piece-t (transform white-piece white-pose)) (black-piece-t (transform black-piece black-pose))) (union white-piece-t black-piece-t))) (union (union-loop 8 create-pawns) (union-loop 8 create-other-pieces)))" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.089, + 0.905, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.506, + 0.907, + 0.604 + ], + "angle": 0, + "content": "Figure 2. Overview. A Scene Language represents a scene with three components: a program consisting of entity functions, a set of words (e.g., pawn) denoting the semantic class of the entity functions, and a list of embeddings (e.g., ) capturing the identity of each entity in the scene. Each entity function is bound with an entity class name given by a word, and maps an input embedding to an output entity of that class. Executing the program evaluates entity functions to compute the full set of entities in the scene. The resulting computation graph, illustrated on the right, captures the dependency structure among entities (indicated by arrows). The program shown is converted from our text-conditioned inference method's output, with details included in Appendix F.2.1; it is written in Lisp-style syntax for brevity and is implemented with Python in practice (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.612, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Modern AI techniques provide natural grounding for each of the three modalities, while also falling short of capturing all: in-context learning of pre-trained language models (LMs) enables the inference of domain-specific programs [3]; LMs themselves capture rich semantic information based on words in natural language; and embeddings obtained via techniques like textual inversion [6] or low-rank adaptation [9] best capture object identity. However, none of these existing representations alone is sufficient for scene generation and editing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We introduce the Scene Language, a representation that integrates the three modalities—programs, words, and embeddings—to precisely and concisely describe the structure, semantics, and identity of visual scenes. In the Scene Language, a program specifies a computation process that defines the organization of a collection of entities in the scene, including extrinsics like poses and structural regularity like repetitions. Each entity is associated with a word referring to its semantic group, as well as an embedding describing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.612, + 0.715, + 0.626 + ], + "angle": 0, + "content": "its instance-specific attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.809 + ], + "angle": 0, + "content": "In addition to the representation itself, we propose a training-free inference module using a pre-trained LM as a backbone to infer the Scene Language from texts and images. When provided with a domain-specific language (DSL) for scenes, LMs decompose the task of complex scene generation into simpler tasks of scene component generation by predicting their corresponding modular functions. We also discuss possible neural, traditional, and hybrid graphics engines that render images based on the representation. Together, the Scene Language, the inference module, and the renderer lead to a system for high-quality, detailed 3D and 4D scene generation and editing." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.811, + 0.788, + 0.824 + ], + "angle": 0, + "content": "In summary, our contributions include" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.826, + 0.905, + 0.871 + ], + "angle": 0, + "content": "1. A scene representation, the Scene Language, capturing structure, semantics, and identity of visual scenes using programs, words, and embeddings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "2. A training-free method that infers the representation from texts and/or images using pre-trained LMs." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24626" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "3. A generic rendering module that renders the Scene Language into an image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.483, + 0.152 + ], + "angle": 0, + "content": "4. Empirical results on text- and image-prompted scene generation and editing tasks." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.165, + 0.232, + 0.181 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.19, + 0.484, + 0.311 + ], + "angle": 0, + "content": "Visual scene representations are arguably the most fundamental problem in computer vision; thus, for sure, we may not enumerate all related work. As our Scene Language comprises programs, words, and embeddings, we organize our discussion accordingly into three categories: scene representations using program-based representations (Sec. 2.1), semantic graph-based representations (Sec. 2.2), and a pre-trained generative model's latent space (Sec. 2.3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.32, + 0.384, + 0.336 + ], + "angle": 0, + "content": "2.1. Representing Scenes as Programs" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Programs can specify not only the relations among scene components mentioned in Sec. 2.2, but also structural patterns such as hierarchy and repetitions, making them suitable as explicit descriptions of scene structures. Prior works have proposed to use programs in the form of sequences of execution commands as object-centric representations, followed by neural executors that render the programs into 3D shapes [5, 30, 36]. In comparison, ShapeAssembly [17] introduces higher-level functions with semantically meaningful function names, e.g., \"chair\" and \"back\", to its program representation. Both ShapeAssembly and ours adopt the design principle of function abstraction, which results in clearly stated hierarchy relation among components and better program editability. However, ShapeAssembly uses cuboids as the shape representation and does not model appearance, while ours allows for more precise geometry and appearance modeling using expressive neural embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.72 + ], + "angle": 0, + "content": "All representations above require 3D datasets for training. More recently, with the advance of language models (LMs), several methods [10, 33, 34, 37, 39, 43] have proposed to use zero-shot LM inference for generating programs that will be rendered into scenes. These methods operate on top of program syntax from specific graphics renderers such as Blender [4], and they do not permit parameters in high-dimensional embedding spaces unlike ours." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.46, + 0.744 + ], + "angle": 0, + "content": "2.2. Representing Scenes with Semantic Graphs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Prior semantic scene representations often adopt a graph to encode semantic scene components, such as objects and parts. In particular, [11, 38] propose to employ a parse graph of context-free grammar, using terminal nodes to correspond to objects and their attributes, to represent a scene. Both works employ an analysis-by-synthesis approach to infer the representation from images that heavily rely on domain-specific priors. Alternative representations include scene graph [7, 15, 16], where each node in a graph corresponds to an object and an edge corresponds to a pairwise" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.262 + ], + "angle": 0, + "content": "
Sec. 3.1Sec. 3.2Definition
Expressions
ΨtransformtransformTransform an entity
ΨunionunionCompose entities
fw:z,γ→h<entity FUNC>Mapping embeddings to entity
fw(z,γ)(call <word>)Entity function application
Data Types
wWordDescribing semantics
tMatrixEntity pose
zEmbeddingSpecifying entity identity
γList [Embedding]Specifying descendant entities' identities
hEntityAn entity
sEntityThe represented scene
" + }, + { + "type": "table_caption", + "bbox": [ + 0.548, + 0.265, + 0.871, + 0.278 + ], + "angle": 0, + "content": "Table 1. Summary of Notations in Secs. 3.1 and 3.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.905, + 0.397 + ], + "angle": 0, + "content": "relation, and StructureNet [22], which focuses on an object-centric setting and uses nodes for object parts. While these representations preserve the high-level semantics of scenes or objects, they leave out low-level precision; thus, geometric, textural, or relational details that cannot be fully specified by language or hand-crafted rules are often ignored. We address this issue via the inclusion of embeddings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.905, + 0.44 + ], + "angle": 0, + "content": "2.3. Representing Scenes with Generative Model Latents" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.907, + 0.66 + ], + "angle": 0, + "content": "The latent space of visual generative models can serve as a representation space for visual scenes. Such latent space can effectively capture the exact visual content of scenes, including geometry and appearance details, and can be either directly inferred, e.g., in variational inference [19] and model inversion [45]. More recently, text-to-image diffusion models have shown remarkable results in image synthesis. This class of models offers several candidate representation spaces including the space of textual embeddings [6], low-rank network weights [9], full model weights [29], or noise vectors in the diffusion process [8, 24, 32]. However, such representations typically do not offer interpretable semantics or explicitly encode hierarchical scene structures unlike ours." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.712, + 0.696 + ], + "angle": 0, + "content": "3. The Scene Language" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We aim to design a visual scene representation that encodes the structure, semantics, and visual content of scenes. Towards this goal, we propose the Scene Language, which represents a scene with three components: a program that encodes scene structure by specifying the existence and relations of scene components, which we will refer to as entities; words in natural language that denote the semantic group of each entity in the scene; and neural embeddings that pertain the low-level visual details and identities of the entities by permitting an expressive input parameter space. In the following, we will first give a formal definition of the representation (Sec. 3.1), and then introduce a domain-specific language (DSL) (Sec. 3.2) as its realization." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24627" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.9, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.17, + 0.908, + 0.213 + ], + "angle": 0, + "content": "Figure 3. Rendering. Given a Scene Language in (a), a program interpreter executes the program to obtain a data object in (b). A graphics renderer first reparameterizes the data object from (b) into the renderer-specific parameter space, and then executes the rendering operation \\(\\mathcal{R}\\) to obtain the final image output in (d)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.219, + 0.48, + 0.304 + ], + "angle": 0, + "content": "
RendererExamples
Rendering Operation RParameters from Θg_reparam
Primitive-basedLight transportShape and BRDFLM inference
Asset-basedRay tracingAsset metadataLM inference
SDS-basedGaussian splattingGaussian parametersSDS optimization
T2I modelFeed-forward passText embeddings in ZCLIP encoding
" + }, + { + "type": "table_caption", + "bbox": [ + 0.161, + 0.305, + 0.414, + 0.319 + ], + "angle": 0, + "content": "Table 2. Graphics Renderers Examples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.326, + 0.266, + 0.34 + ], + "angle": 0, + "content": "3.1. Formal Definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.379 + ], + "angle": 0, + "content": "The Scene Language for a scene \\(s\\), denoted as \\(\\Phi(s)\\), is formally defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.388, + 0.483, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\Phi (s) := (W, P, Z). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.414, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Here, \\(W\\) is a collection of phrases in natural language, referred to as words, e.g., \\(W = \\{\\text{board}, \\text{pawn}, \\dots\\}\\) as illustrated in Fig. 2, and \\(P := \\{f_w\\}_{w \\in W}\\) is a program consisting of a set of entity functions \\(f_w\\) indexed by \\(w \\in W\\). Each entity function \\(f_w\\) defines a class of entities in the scene; it is uniquely identified by the associated \\(w \\in W\\), which concisely summarizes the semantic meaning of the defined class. Each of \\(f_w\\) maps neural embeddings \\((z, \\gamma)\\) to a specific entity \\(h\\) in the scene, where \\(z\\) specifies the attributes and identity of \\(h\\), like a specific color of a pawn, and \\(\\gamma\\) specifies identities of its sub-entities. Hence, the complete Scene Language \\(\\Phi(s)\\) of a particular scene \\(s\\) also contains an ordered set of neural embeddings \\(Z := [z_1, z_2, \\dots, z_J]\\) corresponding to the \\(J\\) specific entities \\([h_1, h_2, \\dots, h_J]\\) from scene \\(s\\), where \\(h_j \\in J\\) is computed by applying an entity function on inputs \\(z_j \\in Z, \\gamma_j \\subset Z\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Crucially, program \\( P \\) captures scene structures in three aspects. First, each entity function \\( f_{w} \\in P \\) transforms and composes multiple sub-entities (e.g., 64 squares) into a new, more complex entity (e.g., a board), naturally reflecting the hierarchical, part-whole relations in the scene, as the arrows in Fig. 2 highlight. Second, multiple entities \\( h_{j} \\)'s may belong to the same semantic class \\( w \\) (e.g., square), and can thus be represented by reusing the same entity function \\( f_{w} \\) with distinct embeddings \\( z_{j} \\)'s. Finally, each entity function also captures the precise spatial layout of the sub-entities by specifying their relative poses during the composition, such as 64 squares forming an \\( 8 \\times 8 \\) grid." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.837, + 0.483, + 0.883 + ], + "angle": 0, + "content": "The following expands on how functions from \\( P \\) are defined and the program execution procedure to compute the represented scene \\( s \\). Notations are summarized in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Entity Function Definitions. An entity function \\( f_{w} \\) outputs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.907, + 0.35 + ], + "angle": 0, + "content": "an entity \\( h \\) of semantic class \\( w \\). The function \\( f_w \\) takes two inputs: an embedding \\( z \\) that specifies \\( h \\)'s identity, as well as an ordered set \\( \\gamma \\) containing embeddings of all descendant entities of \\( h \\). Denote \\( \\{h^{(i)}\\}_{i=1}^N \\) as the \\( N \\) sub-entities of \\( h \\), where \\( N \\) is a constant determined by \\( f_w \\). Then we have \\( \\gamma = \\{z^{(i)}, \\gamma^{(i)}\\}_{i=1}^N \\), where \\( z^{(i)} \\) and \\( \\gamma^{(i)} \\) are the embeddings of \\( h^{(i)} \\) and of the descendant entities of \\( h^{(i)} \\), respectively. Thus, \\( f_w(z, \\gamma) \\) is recursively defined as" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.354, + 0.905, + 0.399 + ], + "angle": 0, + "content": "\\[\nh = f _ {w} (z, \\gamma) := \\Psi_ {\\text {u n i o n}} \\left(\\left\\{\\Psi_ {\\text {t r a n s f o r m}} \\left(h ^ {(i)}, t ^ {(i)} (z)\\right) \\right\\} _ {i = 1} ^ {N}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.905, + 0.431 + ], + "angle": 0, + "content": "where \\(h^{(i)} = f_{w^{(i)}}(z^{(i)},\\gamma^{(i)})\\) is a sub-entity of \\(h\\) and \\(t^{(i)}(z)\\) specifies its pose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Here, \\(f_{w}\\) specifies the computation procedure to obtain output \\(h\\) from \\(N\\) sub-entities \\(\\{h^{(i)}\\}_{i = 1}^{N}\\) via two operations: \\(\\Psi_{\\mathrm{transform}}\\) applies an input-dependent pose \\(t^{(i)}(z)\\) to a subentity \\(h^{(i)}\\), transforming it from its canonical frame to the world frame of \\(h\\), and \\(\\Psi_{\\mathrm{union}}\\) composes multiple sub-entities into one single entity. Each sub-entity \\(h^{(i)}\\) is computed by recursively applying an entity function \\(f_{w^{(i)}}\\) also defined using Eq. (2). For instance, let \\(f_{w}\\) denote the entity function that produces the board in Fig. 2 (namely, \\(w = \\mathrm{board}\\)). This function \\(f_{w}\\) composes 64 sub-entities \\(\\{h^{(i)}\\}_{i = 1}^{64}\\) of the same class square, which are in turn obtained by executing the same entity function \\(f_{w^{(i)}} = f_{\\mathrm{square}}\\) with different embeddings \\(z^{(i)}\\) and \\(\\gamma^{(i)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.844 + ], + "angle": 0, + "content": "Program Execution. To obtain a scene \\( s \\) from the Scene Language \\( \\Phi(s) = (W, P, Z) \\), a program executor identifies a root entity function \\( f_{w_1} \\) from \\( P \\) that is not dependent by any other function (e.g., \\( w_1 = \\) chessboard from Fig. 2), and evaluates this root function using the first embedding \\( z_1 \\in Z \\) and the rest as its descendants, \\( \\gamma_1 := [z_2, \\dots, z_J] \\subset Z \\), to obtain \\( s = f_{w_1}(z_1, \\gamma_1) \\). Evaluating \\( f_{w_1}(z_1, \\gamma_1) \\) expands the computation recursively to its children functions \\( h_j = f_{w_j}(z_j, \\gamma_j) \\) as defined in Eq. (2), obtaining a full sequence of all the entities \\( h_j \\) of the scene, where \\( j = 1, 2, \\dots, J, z_j \\in Z, w_j \\in W \\). The order in \\( Z \\) corresponds to the depth-first-search of the computation graph starting from \\( z_1 \\). An example computation graph is visualized on the right of Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.887, + 0.866 + ], + "angle": 0, + "content": "3.2. Scene Language as Programming Language" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "We now concretize the definition in Sec. 3.1 with a domain-specific language (DSL) specified in Appendix B.1. We in-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24628" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "produce macro operations union for \\(\\Psi_{\\text{union}}\\), union-loop which calls union on entities evaluated in a for-loop, and transform for \\(\\Psi_{\\text{transform}}\\). Together with call which applies function calls, these four macro operations fully define entity functions in the DSL. Entity functions are identified with the associated words in the DSL via two special forms: bind, which binds entity function \\(f_w\\) to word \\(w\\), and retrieve, which retrieves \\(f_w\\) given \\(w\\) and applies \\(f_w\\) on actual embedding parameters. If \\(w\\) is never bound to a function, it corresponds to an entity function with no sub-entities (i.e., \\(N = 0\\) in Eq. (2)), in which case call returns a primitive entity with no children." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.471 + ], + "angle": 0, + "content": "The data type of an entity \\( h \\) (Eq. (2)) is denoted as Entity. It stores two data fields Word and Embedding, describing the entity's semantic group and identity respectively, and stores each children entity together with its pose in the frame of \\( h \\). In particular, Embedding captures the visual details and requires a highly expressive representation, e.g., neural embeddings. We employ the textual embedding space of OpenCLIP-ViT/H [13] for parameterization, denoted as \\( \\mathcal{Z}_{\\mathrm{CLIP}} \\). It offers the advantage that embeddings can be either encoded directly from natural language or inferred from images with Textual Inversion [6]. Tab. 1 summarizes the expressions and data types in accordance with the notations from Sec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.488, + 0.206, + 0.505 + ], + "angle": 0, + "content": "4. Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.483, + 0.605 + ], + "angle": 0, + "content": "Applying the proposed scene representation to image generation tasks requires rendering a Scene Language \\(\\Phi(s)\\) into images. To do so, first, the program interpreter evaluates \\(\\Phi(s)\\) to obtain a data object of type Entity. Afterward, a graphics renderer maps the Entity data object to its rendering parameter space and renders it into a final image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.837 + ], + "angle": 0, + "content": "**Renderer Specifications.** We define the specification of a graphics renderer, a module in the proposed representation, as follows. A graphics renderer is determined by primitive parameter space \\(\\Theta\\) and a rendering operation \\(\\mathcal{R} : \\mathcal{P}(\\Theta \\times \\mathcal{T}) \\to \\mathcal{I}\\), where \\(\\mathcal{T}\\) is the space of 3D affine transformations representing poses, \\(\\mathcal{P}\\) denotes all possible subsets, and \\(\\mathcal{I}\\) is the space of rendered images. In order to determine a mapping from program execution outputs of type Entity (Fig. 3b) to the admissible input domain of rendering operation \\(\\mathcal{R}\\) (Fig. 3c), we assume access to a reparameterization function \\(g_{\\text{reparam}}\\) that maps from Tuple[Word, Embedding] to \\(\\Theta\\), and compute a primitive's pose \\(t \\in \\mathcal{T}\\) by multiplying all Matrix values along the path from the primitive to the scene (root) in the entity hierarchy, analogous to computing limb poses in a kinematic tree." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.903 + ], + "angle": 0, + "content": "**Renderer Instantiations.** An example renderer instantiation is with Score Distillation Sampling (SDS) [25] guidance, where \\(\\Theta\\) is a differentiable 3D representation, and we specify \\(g_{\\mathrm{reparam}}\\) as follows. Recall that an entity is as-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "sociated with a Word value, e.g., moai for one statue from Figs. 1 and 3, and an Embedding value, e.g., \\(\\). For each primitive entity (i.e., an entity with no children), given these two value fields together with the fields of its ancestors, we use a manually specified language template \\(c\\), or a \\(\\) moai, in the style of \\(\\), 3D model in this example, to embed them into \\(z = g_{\\mathrm{CLIP}}(c) \\in \\mathcal{Z}_{\\mathrm{CLIP}}\\) where \\(g_{\\mathrm{CLIP}}\\) is the pre-trained CLIP text encoder. Then, \\(g_{\\mathrm{reparam}}: \\mathcal{Z}_{\\mathrm{CLIP}} \\to \\Theta\\) corresponds to the SDS-guided optimization to find a solution in \\(\\Theta\\) that aligns with the input condition \\(z \\in \\mathcal{Z}_{\\mathrm{CLIP}}\\). Output 3D scenes can be personalized by editing embeddings, e.g., \\(\\) which controls the global style in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.393 + ], + "angle": 0, + "content": "For the underlying 3D representation, we use 3D Gaussian Splatting [18] where images are rendered by splatting a set of 3D Gaussians onto the image plane; other differentiable 3D representations such as neural fields will also be suitable. We base our implementation on GALA3D [44], and use MVDream [31] and a depth-conditioned ControlNet [40] for guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.53 + ], + "angle": 0, + "content": "We refer to the renderer above as the Gaussians renderer. Other possible renderers include primitive-based renderers, such as Mitsuba [14] with graphics primitives of cubes, spheres, and cylinders; asset-based game engines, such as MineCraft [23]; and feed-forward inference of layout-conditioned text-to-image (T2I) diffusion models, such as MIGC [42], which achieves 2D bounding box conditioning by controlling attention layers from Stable Diffusion [28]). Tab. 2 shows a summary; details are in Appendix B.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.905, + 0.56 + ], + "angle": 0, + "content": "5. Inference via Pre-Trained Language Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.674 + ], + "angle": 0, + "content": "We introduce a training-free method to infer the representation \\(\\Phi(s) = (W, P, Z)\\) from text or image descriptions of a scene \\(s\\). As explained below, we first prompt a pre-trained language model (LM) to generate the non-neural components \\((W, P)\\) and then obtain neural embeddings \\((Z)\\) from texts via the CLIP text encoder or from images with a pretrained text-to-image diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.811 + ], + "angle": 0, + "content": "LMs have shown remarkable capability in code generation with common programming languages such as Python. In our implementation, we prompt LMs to generate Python scripts. We prompt the LM with the input condition, i.e., a scene description in texts or an image; a Python script of helper functions converted from the DSL in Sec. 3.2; and an example script using the helper functions. We use Claude 3.5 Sonnet [1] for all experiments for our method and LM-dependent baselines. Full LM prompts are in Appendix F.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Function arguments in the LM-generated scripts, which are numeric values or string tokens, are converted to embeddings from \\(\\mathcal{Z}_{\\mathrm{CLIP}}\\) (Sec. 3.2) using language templates and the CLIP text encoder \\(g_{\\mathrm{CLIP}}\\). For example, in the raw LM output, function calls for white pieces in Fig. 2 have input attribute {\"color\": (.9, .9, .9)}}, and we prompt LM" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24629" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.092, + 0.255, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.187, + 0.24, + 0.2 + ], + "angle": 0, + "content": "Ours (Two Views)" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.097, + 0.348, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.188, + 0.345, + 0.2 + ], + "angle": 0, + "content": "MVDream" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.111, + 0.459, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.188, + 0.46, + 0.2 + ], + "angle": 0, + "content": "GraphDreamer" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.202, + 0.34, + 0.216 + ], + "angle": 0, + "content": "\"An 8-layer shelf\"" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.214, + 0.171, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.309, + 0.24, + 0.322 + ], + "angle": 0, + "content": "Ours (Two Views)" + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.214, + 0.244, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.226, + 0.357, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.308, + 0.344, + 0.321 + ], + "angle": 0, + "content": "MVDream" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.224, + 0.461, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.309, + 0.459, + 0.322 + ], + "angle": 0, + "content": "GraphDreamer" + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.324, + 0.407, + 0.338 + ], + "angle": 0, + "content": "\"15 coke cans stacking in a pyramid\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.34, + 0.483, + 0.382 + ], + "angle": 0, + "content": "Figure 4. Text-Prompted Scene Generation. Ours produce more accurate outputs than using no intermediate representation (MV-Dream) or scene graph (GraphDreamer)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.452 + ], + "angle": 0, + "content": "
MethodsNumeric PromptsGeneric Prompts
Alignment (↑)CLIP (↑)Count (↑)Alignment (↑)CLIP (↑)
GraphDreamer [7]3.56±7.380.297±0.0850.1115.75±20.390.230±0.097
MVDream [31]10.79±12.830.312±0.0460.1135.13±31.900.328±0.031
Ours85.65±13.710.351±0.0511.052.13±26.830.328±0.030
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.537 + ], + "angle": 0, + "content": "Table 3. Evaluation for text-prompted 3D generation. We report the following metrics averaged across 9 numeric and 8 generic prompts: the percentages of user preferences for prompt alignment, CLIP similarity scores, and counting accuracy (numeric prompts only, 0 for inaccurate and 1 for accurate). Here \\(\\pm\\) denotes standard deviation across scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.671 + ], + "angle": 0, + "content": "to describe the color value as a word, and feed the word into \\( g_{\\mathrm{CLIP}} \\) to compute \\( < z68 > \\). For image-conditioned tasks, for each primitive entity in the execution output of \\( P \\), we first use GroundingSAM [20, 27] to segment out the region defined by the word associated with the entity. We then use Textual Inversion [6] to optimize an embedding to reconstruct the cropped image with the diffusion model training objective. The full process is deferred to Appendix C.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.224, + 0.701 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.709, + 0.43, + 0.726 + ], + "angle": 0, + "content": "6.1. Text-Prompted Generation and Editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Baselines. To evaluate our representation in text-prompted 3D generation task, we compare our inference pipeline with 3D scene generation methods using alternative intermediate representations, e.g., scene graph. In particular, we compare with GraphDreamer [7] as an exemplar approach, which generates scene graphs from texts via LM prompting and then synthesizes scenes conditioned on the graphs via SDS guidance. We further ablate the role of structural representation in this task by comparing ours with the backbone of our SDS-based renderer, MVDream [31], as a direct scene generation approach. Details are in Appendix C.3." + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.097, + 0.541, + 0.176 + ], + "angle": 0, + "content": "rnnnne nnnnne" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.096, + 0.632, + 0.125 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.133, + 0.632, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.182, + 0.62, + 0.193 + ], + "angle": 0, + "content": "Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.194, + 0.695, + 0.219 + ], + "angle": 0, + "content": "\"A detailed model of a tennis court\"" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.096, + 0.721, + 0.125 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.131, + 0.722, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.182, + 0.71, + 0.193 + ], + "angle": 0, + "content": "Minecraft" + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.194, + 0.696, + 0.205 + ], + "angle": 0, + "content": "model of" + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.207, + 0.676, + 0.218 + ], + "angle": 0, + "content": "a tennis court\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.223, + 0.541, + 0.256 + ], + "angle": 0, + "content": "nne nnnnne" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.224, + 0.63, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.274, + 0.612, + 0.285 + ], + "angle": 0, + "content": "Mitsuba" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.223, + 0.723, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.274, + 0.692, + 0.285 + ], + "angle": 0, + "content": "T2I" + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.287, + 0.706, + 0.3 + ], + "angle": 0, + "content": "\"City skyline at sunset\"" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.093, + 0.813, + 0.132 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.134, + 0.813, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.182, + 0.801, + 0.193 + ], + "angle": 0, + "content": "Gaussians" + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.092, + 0.901, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.135, + 0.897, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.182, + 0.802, + 0.193 + ], + "angle": 0, + "content": "Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.182, + 0.892, + 0.193 + ], + "angle": 0, + "content": "Minecraft" + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.194, + 0.886, + 0.22 + ], + "angle": 0, + "content": "\"An interior scene of a university lecture hall\"" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.224, + 0.779, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.274, + 0.795, + 0.285 + ], + "angle": 0, + "content": "Mitsuba" + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.224, + 0.9, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.846, + 0.274, + 0.869, + 0.285 + ], + "angle": 0, + "content": "T2I" + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.287, + 0.855, + 0.299 + ], + "angle": 0, + "content": "\"Marionette\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.304, + 0.906, + 0.36 + ], + "angle": 0, + "content": "Figure 5. Renderings Across Graphics Renderers. Different renderers produce renderings that adhere to the same representation and are therefore visually aligned, while each exhibits a different imaging style. Text inputs are shown at the subfigure bottoms." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.372, + 0.73, + 0.419 + ], + "angle": 0, + "content": "
MethodsCLIP [13] (↑)Dynamic [12] (↑)
4D-fy [2]0.3520.2%
Ours0.3415.9%
" + }, + { + "type": "table", + "bbox": [ + 0.735, + 0.373, + 0.904, + 0.418 + ], + "angle": 0, + "content": "
MethodsLPIPS [41] (↓)
GraphDreamer [7]0.811
Ours0.681
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.422, + 0.906, + 0.492 + ], + "angle": 0, + "content": "Table 4. Evaluation. Left: For text-prompted 4D generation, 4Dfy achieves higher prompt alignment measured with CLIP similarity but has small motion. Right: For image-prompted scene generation, our results are better aligned with input image prompts compared to GraphDreamer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.906, + 0.639 + ], + "angle": 0, + "content": "Metrics. We conduct a user study where 103 users are asked to choose one of the three videos, showing 360-renderings from ours and two baselines in a randomized order, that aligns the best with text prompts. Following [7], for each scene, we compute the cosine similarity of CLIP [13, 26] embeddings between a rendering and the text prompt and report the average over 120 viewpoints. We further report whether outputs have the correct object counts. Details are deferred to Appendix C.4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Results. Tab. 3 shows that our method achieves a more favorable prompt alignment than the baselines as measured in the user study and CLIP similarity, and has a clear advantage in counting accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 and Appendix D.1 show qualitative results of ours using the SDS-based renderer. Compared to the direct 3D scene generation method MVDream, our approach is compositional and adheres more closely to input prompts in scenes involving multiple objects. Compared to a scene graph representation, where entity relations are restricted to be between two objects and are bottlenecked by the coarseness of natural language descriptions, e.g., \"aligned in a row\", a program-based representation offers more flexible and precise specifications for relations, e.g., the particular coke can arrangement in Fig. 4. This brings the practical benefit of offloading the burden of generating scenes involving complex entity relations from the T2I model (used" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24630" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.095, + 0.208, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.191, + 0.19, + 0.22 + ], + "angle": 0, + "content": "Ours (w/ tracking)" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.095, + 0.304, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.184, + 0.29, + 0.195 + ], + "angle": 0, + "content": "Frame 1" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.095, + 0.404, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.185, + 0.357, + 0.195 + ], + "angle": 0, + "content": "1 Fran" + }, + { + "type": "image_footnote", + "bbox": [ + 0.144, + 0.227, + 0.348, + 0.242 + ], + "angle": 0, + "content": "\"Carousel with a small canopy\"" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.119, + 0.543, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.191, + 0.522, + 0.221 + ], + "angle": 0, + "content": "Ours (w/ tracking)" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.127, + 0.691, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.183, + 0.686, + 0.196 + ], + "angle": 0, + "content": "Frame 1 Frame 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.198, + 0.64, + 0.213 + ], + "angle": 0, + "content": "4D-fy" + }, + { + "type": "image_footnote", + "bbox": [ + 0.489, + 0.227, + 0.63, + 0.241 + ], + "angle": 0, + "content": "\"Solar system model\"" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.096, + 0.892, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.184, + 0.893, + 0.224 + ], + "angle": 0, + "content": "Ours Frame 1 Frame 2 (w/ tracking) 4D-fy" + }, + { + "type": "image_footnote", + "bbox": [ + 0.725, + 0.227, + 0.861, + 0.242 + ], + "angle": 0, + "content": "\"A toy wind turbine\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.907, + 0.288 + ], + "angle": 0, + "content": "Figure 6. Text-Prompted 4D Scene Generation. The proposed representation may also capture scene dynamics and be applied for synthesizing 4D scenes. It explicitly represents the temporal correspondence of an entity in a dynamic scene. Each colored trajectory denotes tracking of a temporally moving point. Compared to 4D-fy, our method produces more salient motion." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.296, + 0.225, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.416, + 0.225, + 0.43 + ], + "angle": 0, + "content": "Initial (Mitsuba)" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.297, + 0.379, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.416, + 0.388, + 0.43 + ], + "angle": 0, + "content": "Initial (Gaussians)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.115, + 0.433, + 0.256, + 0.447 + ], + "angle": 0, + "content": "\"Bramante Staircase," + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.456, + 0.226, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.565, + 0.225, + 0.58 + ], + "angle": 0, + "content": "Initial (Mitsuba)" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.456, + 0.385, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.565, + 0.389, + 0.58 + ], + "angle": 0, + "content": "Initial (Gaussians)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.103, + 0.582, + 0.393, + 0.597 + ], + "angle": 0, + "content": "\"An 8-layer 3-color Jenga set at game start\"" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.605, + 0.236, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.715, + 0.225, + 0.73 + ], + "angle": 0, + "content": "Initial (Mitsuba)" + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.605, + 0.393, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.715, + 0.389, + 0.73 + ], + "angle": 0, + "content": "Initial (Gaussians)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.152, + 0.732, + 0.346, + 0.746 + ], + "angle": 0, + "content": "\"A fractal tree-like structure\"" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.297, + 0.585, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.417, + 0.558, + 0.431 + ], + "angle": 0, + "content": "Scene Language" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.296, + 0.683, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.416, + 0.721, + 0.43 + ], + "angle": 0, + "content": "Edited (Mitsuba)" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.296, + 0.841, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.416, + 0.886, + 0.429 + ], + "angle": 0, + "content": "Edited (Gaussians)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.55, + 0.433, + 0.769, + 0.448 + ], + "angle": 0, + "content": "\"Shrink staircase radius by \\(80\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.455, + 0.579, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.511, + 0.569, + 0.562 + ], + "angle": 0, + "content": "(\"Jenga layer\" (lambda(is-horizontal colors remove-blocks)))" + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.566, + 0.558, + 0.581 + ], + "angle": 0, + "content": "Scene Language" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.456, + 0.723, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.566, + 0.721, + 0.58 + ], + "angle": 0, + "content": "Edited (Mitsuba)" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.456, + 0.88, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.566, + 0.886, + 0.58 + ], + "angle": 0, + "content": "Edited (Gaussians)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.425, + 0.582, + 0.893, + 0.597 + ], + "angle": 0, + "content": "\"Remove 2 blocks from second top layer and the tower should not fall\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.602, + 0.543, + 0.64 + ], + "angle": 0, + "content": "union left-branch right-branch)" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.643, + 0.544, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.716, + 0.558, + 0.73 + ], + "angle": 0, + "content": "Scene Language" + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.611, + 0.7, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.715, + 0.721, + 0.73 + ], + "angle": 0, + "content": "Edited (Mitsuba)" + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.61, + 0.859, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.715, + 0.887, + 0.73 + ], + "angle": 0, + "content": "Edited (Gaussians)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.499, + 0.732, + 0.82, + 0.747 + ], + "angle": 0, + "content": "\"Make branching structure to be trinary and 3D\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.753, + 0.907, + 0.796 + ], + "angle": 0, + "content": "Figure 7. Scene Editing with Language Instructions. Our representation is editing-friendly. Bottoms of each row show initial scene descriptions and editing instructions in the format of user text prompts. We prompt an LM to infer the initial Scene Language as well as the edits (shown with texts in red), and show image renderings with two renderers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.482, + 0.832 + ], + "angle": 0, + "content": "for SDS guidance in both ours and GraphDreamer) towards LM, leading to accurate and detailed generation results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.836, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Across Graphics Renderers. Fig. 5 shows that our representation can be rendered with different renderers (Sec. 4), producing aligned outputs in different visual domains." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.9 + ], + "angle": 0, + "content": "4D Generation. The inference method from Sec. 5 can be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.906, + 0.848 + ], + "angle": 0, + "content": "applied to 4D scene generation. The 4D scene representation is identical to Eq. (1), except that there is an additional 4D entity function in the program \\( P \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We compare with 4D-fy [2] as an exemplar text-to-4D generation method. We measure the prompt alignment of output dynamic renderings using CLIP similarity between" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24631" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.91, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.217, + 0.908, + 0.247 + ], + "angle": 0, + "content": "Figure 8. Scene Editing with Image Instructions. This figure shows style transfer results under the same setup as in Fig. 1. Replacing embeddings of target entities in a Scene Language with a new embedding specified by a user image achieves the editing effect." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.254, + 0.188, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.312, + 0.177, + 0.324 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.259, + 0.276, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.311, + 0.29, + 0.323 + ], + "angle": 0, + "content": "Ours (Gaussians)" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.259, + 0.374, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.311, + 0.384, + 0.323 + ], + "angle": 0, + "content": "Ours (Mitsuba)" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.254, + 0.472, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.312, + 0.478, + 0.324 + ], + "angle": 0, + "content": "GraphDreamer" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.324, + 0.188, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.381, + 0.177, + 0.393 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.393, + 0.177, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.34, + 0.287, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.381, + 0.29, + 0.393 + ], + "angle": 0, + "content": "Ours (Gaussians)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.34, + 0.388, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.381, + 0.384, + 0.392 + ], + "angle": 0, + "content": "Ours (Mitsuba)" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.336, + 0.46, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.381, + 0.477, + 0.393 + ], + "angle": 0, + "content": "GraphDreamer" + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.442, + 0.177, + 0.454 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.396, + 0.277, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.442, + 0.29, + 0.454 + ], + "angle": 0, + "content": "Ours (Gaussians)" + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.398, + 0.374, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.442, + 0.384, + 0.454 + ], + "angle": 0, + "content": "Ours (Mitsuba)" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.404, + 0.47, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.442, + 0.478, + 0.454 + ], + "angle": 0, + "content": "GraphDreamer" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.528 + ], + "angle": 0, + "content": "Figure 9. Image-Prompted Scene Generation. Compared to scene graphs from GraphDreamer, programs from our representation encode additional scene structure, e.g., repetitions, and specify geometric entity relations more precisely. Embeddings from ours further enable visual identity preservation in the outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.483, + 0.625 + ], + "angle": 0, + "content": "rendered frames and input texts, and the motion magnitude using dynamic degrees [12] computed using RAFT [35] optical flow estimation over neighboring frames. 4D-fy results have a higher per-frame alignment score than ours but present small dynamic motions. See Fig. 6, the webpage, and Appendix D.2 for further results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Scene Editing. Scenes synthesized from our proposed representation can further be edited by prompting LM with its previously generated script and an editing instruction in natural language instruction. Results are shown in Fig. 7. Our representation provides an interpretable and intuitive interface for scene editing, as functions have explicit semantic meanings associated with words, and function reuse significantly improves program readability. Furthermore, since the structure of programs reflects the structure of scenes, editing program parameters leads to precise changes in the scenes while preserving the original structure, e.g., the circular arrangement of staircases in Fig. 7. Desirable editing effects involving multiple primitives, or all staircases in this example, can be effectively achieved via only minor changes in the program space. Finally, the program structure itself, such as the function header in the Jenga set example, can be adjusted for editing to achieve localized edits that only affect relevant parts of the scene. Further results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.255, + 0.704, + 0.27 + ], + "angle": 0, + "content": "are shown in Appendix D.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.277, + 0.867, + 0.293 + ], + "angle": 0, + "content": "6.2. Image-Prompted Generation and Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.905, + 0.42 + ], + "angle": 0, + "content": "Texts provide an easy user interface for scene generation as shown in Sec. 6.1, and images allow users to specify further intentions that could be hard to describe with natural language alone. Integrating neural embeddings increases the expressivity of the Scene Language and allows it to encode visual details from image prompts, and as a result, enhances our method's controllability in image-prompted scene generation and editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.906, + 0.59 + ], + "angle": 0, + "content": "Generation. Fig. 9 shows results on the task of generating 3D scenes consistent with the structure and content from a user-provided image, termed as the \"inverse semantics\" task in GraphDreamer [7]. We report the LPIPS [41] score, i.e., the AlexNet [21] feature dissimilarity, between an input image and an output rendering, averaged over 3 scenes with 20 viewpoints each in Tab. 4. Embeddings allow our representation to encode visual details from inputs, such as the Coke can texture in Fig. 9, which is not captured in scene graphs from GraphDreamer using per-object natural language descriptions. This is reflected quantitatively in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Editing. Our representation applies to image-prompted editing tasks such as style transfer (Fig. 1). Edits can target partial or full scenes as shown in Fig. 8(a) and (b), respectively, by updating embeddings for corresponding entities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We have introduced a visual scene representation, termed the Scene Language, which encodes three key aspects of visual scenes: scene structure, such as hierarchy and repetition, specified via programs; semantics of individual components succinctly summarized via words; and identities of each component precisely captured via neural embeddings. We formalize the representation as a programming language defined using a DSL. We show that the Scene Language can be efficiently inferred from both text and image inputs leveraging pre-trained language models. Once the program is executed, the resulting scene can be rendered into images using a variety of graphics renderers. Our Scene Language produces 3D and 4D scenes with high fidelity and enables easy and precise editing." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24632" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.184 + ], + "angle": 0, + "content": "Acknowledgments. This work is in part supported by AFOSR YIP FA9550-23-1-0127, ONR N00014-23-1-2355, ONR YIP N00014-24-1-2117, ONR MURI N00014-22-1-2740, and NSF RI #2211258 and #2338203. YZ is in part supported by the Stanford Interdisciplinary Graduate Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.196, + 0.188, + 0.212 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.221, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[1] Anthropic. The Claude 3 model family: Opus, Sonnet, Haiku, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.251, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[2] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B. Lindell. 4D-fly: Text-to-4D generation using hybrid score distillation sampling. CVPR, 2024. 6, 7, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.321, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.4, + 0.391 + ], + "angle": 0, + "content": "[4] Blender Online Community. Blender, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.393, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[5] Boyang Deng, Sumith Kulal, Zhengyang Dong, Congyue Deng, Yonglong Tian, and Jiajun Wu. Unsupervised learning of shape programs with repeatable implicit parts. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.449, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 3, 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[7] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf. GraphDreamer: Compositional 3D scene synthesis from scene graphs. In CVPR, 2024. 3, 6, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.606, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.662, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[10] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. SceneCraft: An LLM agent for synthesizing 3D scene as Blender code. In ICLR Workshop, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[11] Siyuan Huang, Siyuan Qi, Yixin Zhu, Yinxue Xiao, Yuanlu Xu, and Song-Chun Zhu. Holistic 3D scene parsing and reconstruction from a single RGB image. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[12] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 6, 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, 2021. 5, 6, 16" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.221, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[15] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, 2015. 3, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[16] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[17] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy Mitra, and Daniel Ritchie. ShapeAssembly: Learning to generate programs for 3D shape structure synthesis. ACM TOG, 39(6):1-20, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[19] Diederik P Kingma. Auto-encoding variational bayes. In ICLR, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[22] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas Guibas. StructureNet: Hierarchical graph networks for 3D shape generation. ACM TOG, 38 (6):1-19, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.765, + 0.533 + ], + "angle": 0, + "content": "[23] Mojang Studios. Minecraft, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.535, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[27] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded SAM: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[29] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In CVPR, 2018. 3, 16" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24633" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. In ICLR, 2024. 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.163 + ], + "angle": 0, + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[33] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3D-GPT: Procedural 3D modeling with large language models. arXiv preprint arXiv:2310.12945, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[34] Hou In Ivan Tam, Hou In Derek Pun, Austin T Wang, Angel X Chang, and Manolis Savva. SceneMotifCoder: Example-driven visual program learning for generating 3D object arrangements. arXiv preprint arXiv:2408.02211, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[35] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[36] Yonglong Tian, Andrew Luo, Xingyuan Sun, Kevin Ellis, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Learning to infer and execute 3D shape programs. In ICLR, 2019. 3, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[37] Yutaro Yamada, Khyathi Chandu, Yuchen Lin, Jack Hessel, Ilker Yildirim, and Yejin Choi. L3GO: Language agents with chain-of-3D-thoughts for generating unconventional objects. In ICLR Workshop, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[38] Alan Yuille and Daniel Kersten. Vision as Bayesian inference: analysis by synthesis? Trends in Cognitive Sciences, 10(7):301-308, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[39] Chi Zhang, Penglin Cai, Yuhui Fu, Haoqi Yuan, and Zongqing Lu. Creative agents: Empowering agents with imagination for creative tasks. arXiv preprint arXiv:2312.02519, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[42] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. MIGC: Multi-instance generation controller for text-to-image synthesis. In CVPR, 2024. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.717 + ], + "angle": 0, + "content": "[43] Mengqi Zhou, Jun Hou, Chuanchen Luo, Yuxi Wang, Zhaoxiang Zhang, and Junran Peng. SceneX: Procedural controllable large-scale scene generation via large-language models. arXiv preprint arXiv:2403.15698, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[44] Xiaoyu Zhou, Xingjian Ran, Yajiao Xiong, Jinlin He, Zhiwei Lin, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. GALA3D: Towards text-to-3D complex scene generation via layout-guided generative Gaussian splatting. In ICML, 2024. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[45] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In ECCV, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24634" + } + ] +] \ No newline at end of file diff --git a/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_origin.pdf b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7cd4404fed7c7e5e27a81f4633041965b95e7f41 --- /dev/null +++ b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/dc337ba4-1560-4ebb-a36f-4fd2fd0e7e1e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:292f107a6e19456942eb37911ffc3f7b89d950c632d60a2888bd27ef08ec307b +size 5269414 diff --git a/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/full.md b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c32ca079a3e806de40ab97d94cfeec30c6425de0 --- /dev/null +++ b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/full.md @@ -0,0 +1,425 @@ +# The Scene Language: Representing Scenes with Programs, Words, and Embeddings + +Yunzhi Zhang $^{1}$ Zizhang Li $^{1}$ Matt Zhou $^{2}$ Shangzhe Wu $^{1}$ Jiajun Wu $^{1}$ $^{1}$ Stanford University $^{2}$ UC Berkeley + +![](images/91213d4fe6a537581af84bb20e44f900ddcb888c526fac06413df4a3706a9ef2.jpg) +Figure 1. Structured Scene Generation and Editing Using the Scene Language. We develop a scene representation for 3D scene generation and editing tasks. Given textual scene descriptions, the representation can be inferred by a pre-trained large language model, rendered in 3D, and edited following language instructions. The representation contains a program consisting of semantic-aware functions bound to words, providing high interpretability and an intuitive scene-editing interface, and embeddings enabling editing with fine controls, e.g., transferring the style of $>$ from a user's image to the output by updating $$ which controls global attributes of the scene. Project page: https://ai.stanford.edu/~yzzhang/projects/scene-language/. + +![](images/76cf9657f56e00cc283db7ce71606672ad889c072439cd87448fabf06070e054.jpg) + +# Abstract + +We introduce the Scene Language, a visual scene representation that concisely and precisely describes the structure, semantics, and identity of visual scenes. It represents a scene with three key components: a program that specifies the hierarchical and relational structure of entities in the scene, words in natural language that summarize the semantic class of each entity, and embeddings that capture the visual identity of each entity. This representation can be inferred from pre-trained language models via a training-free inference technique, given text or image inputs. The resulting scene can be rendered into images using traditional, neural, or hybrid graphics renderers. Together, this forms an automated system for high-quality 3D and 4D scene generation. Compared with existing representations like scene graphs, our proposed Scene Language generates complex scenes with higher fidelity, while explicitly modeling the scene structures to enable precise control and editing. + +# 1. Introduction + +How do you describe a scene? Imagine that you just traveled to Easter Island and would like to explain to Alice the wondrous scene of Ahu Akivi: "There are seven moai in a row, facing the same direction." "What is a moai?" Alice asked. "A moai is a stone human figure without legs, but each also looks slightly different." At this point, you realize it seems difficult to precisely explain the scene using natural language alone. + +In fact, this example highlights a complete scene representation requires at least three types of complementary information: structural knowledge, which is about the joint distribution of multiple instances, like "seven moai in a row, facing the same direction," most naturally described as programs; category-level semantics, which may be shared across instances, often described in words, such as "moai"; and instance-level intrinsics, tied to the identity of each specific object or part, such as its geometry, color, and texture, which is hard to describe but easy to recognize. + +```lisp +bind "chessboard" (lambda (embedding ...) (union (call "board" ...) (call "chess pieces" ...)))) +(bind "board" (lambda (embedding ...) (define (create-squares i) ...) (union-loop 64 create-squares))) +(bind "chess pieces" (lambda (embedding ... ) (define (create-pawns i) (let* ((white-pawn (call "pawn" ...) (black-pawn (call "pawn" ...) (white-pose (translate (list (* i 0.125) 0.02 0.125)))) (black-pose (translate (list (* i 0.125) 0.02 0.75))) (white-pawn-t (transform white-pawn white-pose)) (black-pawn-t (transform black-pawn black-pose))) (union white-pawn-t black-pawn-t))) (define pieces-order ((("rook" "knight" "bishop" "queen" "king" "bishop" "knight" "rook")) (define (create-other-pieces i) (let* ((piece-order (list-ref pieces-order i)) (white-piece (call piece-order ...)) (black-piece (call piece-order ...)) (white-pose (translate (list (* i 0.125) 0.02 0)) (black-pose (translate (list (* i 0.125) 0.02 0.875))) (white-piece-t (transform white-piece white-pose)) (black-piece-t (transform black-piece black-pose))) (union white-piece-t black-piece-t))) (union (union-loop 8 create-pawns) (union-loop 8 create-other-pieces))) +``` + +![](images/2a8e9bc56400491082bb426f615365567f7f2a4fa8bd9bd40e31057e3889ab17.jpg) +Figure 2. Overview. A Scene Language represents a scene with three components: a program consisting of entity functions, a set of words (e.g., pawn) denoting the semantic class of the entity functions, and a list of embeddings (e.g., ) capturing the identity of each entity in the scene. Each entity function is bound with an entity class name given by a word, and maps an input embedding to an output entity of that class. Executing the program evaluates entity functions to compute the full set of entities in the scene. The resulting computation graph, illustrated on the right, captures the dependency structure among entities (indicated by arrows). The program shown is converted from our text-conditioned inference method's output, with details included in Appendix F.2.1; it is written in Lisp-style syntax for brevity and is implemented with Python in practice (Sec. 3.2). + +Modern AI techniques provide natural grounding for each of the three modalities, while also falling short of capturing all: in-context learning of pre-trained language models (LMs) enables the inference of domain-specific programs [3]; LMs themselves capture rich semantic information based on words in natural language; and embeddings obtained via techniques like textual inversion [6] or low-rank adaptation [9] best capture object identity. However, none of these existing representations alone is sufficient for scene generation and editing. + +We introduce the Scene Language, a representation that integrates the three modalities—programs, words, and embeddings—to precisely and concisely describe the structure, semantics, and identity of visual scenes. In the Scene Language, a program specifies a computation process that defines the organization of a collection of entities in the scene, including extrinsics like poses and structural regularity like repetitions. Each entity is associated with a word referring to its semantic group, as well as an embedding describing + +its instance-specific attributes. + +In addition to the representation itself, we propose a training-free inference module using a pre-trained LM as a backbone to infer the Scene Language from texts and images. When provided with a domain-specific language (DSL) for scenes, LMs decompose the task of complex scene generation into simpler tasks of scene component generation by predicting their corresponding modular functions. We also discuss possible neural, traditional, and hybrid graphics engines that render images based on the representation. Together, the Scene Language, the inference module, and the renderer lead to a system for high-quality, detailed 3D and 4D scene generation and editing. + +In summary, our contributions include + +1. A scene representation, the Scene Language, capturing structure, semantics, and identity of visual scenes using programs, words, and embeddings. +2. A training-free method that infers the representation from texts and/or images using pre-trained LMs. + +3. A generic rendering module that renders the Scene Language into an image. +4. Empirical results on text- and image-prompted scene generation and editing tasks. + +# 2. Related Work + +Visual scene representations are arguably the most fundamental problem in computer vision; thus, for sure, we may not enumerate all related work. As our Scene Language comprises programs, words, and embeddings, we organize our discussion accordingly into three categories: scene representations using program-based representations (Sec. 2.1), semantic graph-based representations (Sec. 2.2), and a pre-trained generative model's latent space (Sec. 2.3). + +# 2.1. Representing Scenes as Programs + +Programs can specify not only the relations among scene components mentioned in Sec. 2.2, but also structural patterns such as hierarchy and repetitions, making them suitable as explicit descriptions of scene structures. Prior works have proposed to use programs in the form of sequences of execution commands as object-centric representations, followed by neural executors that render the programs into 3D shapes [5, 30, 36]. In comparison, ShapeAssembly [17] introduces higher-level functions with semantically meaningful function names, e.g., "chair" and "back", to its program representation. Both ShapeAssembly and ours adopt the design principle of function abstraction, which results in clearly stated hierarchy relation among components and better program editability. However, ShapeAssembly uses cuboids as the shape representation and does not model appearance, while ours allows for more precise geometry and appearance modeling using expressive neural embeddings. + +All representations above require 3D datasets for training. More recently, with the advance of language models (LMs), several methods [10, 33, 34, 37, 39, 43] have proposed to use zero-shot LM inference for generating programs that will be rendered into scenes. These methods operate on top of program syntax from specific graphics renderers such as Blender [4], and they do not permit parameters in high-dimensional embedding spaces unlike ours. + +# 2.2. Representing Scenes with Semantic Graphs + +Prior semantic scene representations often adopt a graph to encode semantic scene components, such as objects and parts. In particular, [11, 38] propose to employ a parse graph of context-free grammar, using terminal nodes to correspond to objects and their attributes, to represent a scene. Both works employ an analysis-by-synthesis approach to infer the representation from images that heavily rely on domain-specific priors. Alternative representations include scene graph [7, 15, 16], where each node in a graph corresponds to an object and an edge corresponds to a pairwise + +
Sec. 3.1Sec. 3.2Definition
Expressions
ΨtransformtransformTransform an entity
ΨunionunionCompose entities
fw:z,γ→h<entity FUNC>Mapping embeddings to entity
fw(z,γ)(call <word>)Entity function application
Data Types
wWordDescribing semantics
tMatrixEntity pose
zEmbeddingSpecifying entity identity
γList [Embedding]Specifying descendant entities' identities
hEntityAn entity
sEntityThe represented scene
+ +Table 1. Summary of Notations in Secs. 3.1 and 3.2. + +relation, and StructureNet [22], which focuses on an object-centric setting and uses nodes for object parts. While these representations preserve the high-level semantics of scenes or objects, they leave out low-level precision; thus, geometric, textural, or relational details that cannot be fully specified by language or hand-crafted rules are often ignored. We address this issue via the inclusion of embeddings. + +# 2.3. Representing Scenes with Generative Model Latents + +The latent space of visual generative models can serve as a representation space for visual scenes. Such latent space can effectively capture the exact visual content of scenes, including geometry and appearance details, and can be either directly inferred, e.g., in variational inference [19] and model inversion [45]. More recently, text-to-image diffusion models have shown remarkable results in image synthesis. This class of models offers several candidate representation spaces including the space of textual embeddings [6], low-rank network weights [9], full model weights [29], or noise vectors in the diffusion process [8, 24, 32]. However, such representations typically do not offer interpretable semantics or explicitly encode hierarchical scene structures unlike ours. + +# 3. The Scene Language + +We aim to design a visual scene representation that encodes the structure, semantics, and visual content of scenes. Towards this goal, we propose the Scene Language, which represents a scene with three components: a program that encodes scene structure by specifying the existence and relations of scene components, which we will refer to as entities; words in natural language that denote the semantic group of each entity in the scene; and neural embeddings that pertain the low-level visual details and identities of the entities by permitting an expressive input parameter space. In the following, we will first give a formal definition of the representation (Sec. 3.1), and then introduce a domain-specific language (DSL) (Sec. 3.2) as its realization. + +![](images/886bb5c93805d49be452cb8203f429239c72eadd2c3abbae32ccf42ee8a61330.jpg) +Figure 3. Rendering. Given a Scene Language in (a), a program interpreter executes the program to obtain a data object in (b). A graphics renderer first reparameterizes the data object from (b) into the renderer-specific parameter space, and then executes the rendering operation $\mathcal{R}$ to obtain the final image output in (d). + +
RendererExamples
Rendering Operation RParameters from Θg_reparam
Primitive-basedLight transportShape and BRDFLM inference
Asset-basedRay tracingAsset metadataLM inference
SDS-basedGaussian splattingGaussian parametersSDS optimization
T2I modelFeed-forward passText embeddings in ZCLIP encoding
+ +Table 2. Graphics Renderers Examples. + +# 3.1. Formal Definition + +The Scene Language for a scene $s$ , denoted as $\Phi(s)$ , is formally defined as follows: + +$$ +\Phi (s) := (W, P, Z). \tag {1} +$$ + +Here, $W$ is a collection of phrases in natural language, referred to as words, e.g., $W = \{\text{board}, \text{pawn}, \dots\}$ as illustrated in Fig. 2, and $P := \{f_w\}_{w \in W}$ is a program consisting of a set of entity functions $f_w$ indexed by $w \in W$ . Each entity function $f_w$ defines a class of entities in the scene; it is uniquely identified by the associated $w \in W$ , which concisely summarizes the semantic meaning of the defined class. Each of $f_w$ maps neural embeddings $(z, \gamma)$ to a specific entity $h$ in the scene, where $z$ specifies the attributes and identity of $h$ , like a specific color of a pawn, and $\gamma$ specifies identities of its sub-entities. Hence, the complete Scene Language $\Phi(s)$ of a particular scene $s$ also contains an ordered set of neural embeddings $Z := [z_1, z_2, \dots, z_J]$ corresponding to the $J$ specific entities $[h_1, h_2, \dots, h_J]$ from scene $s$ , where $h_j \in J$ is computed by applying an entity function on inputs $z_j \in Z, \gamma_j \subset Z$ . + +Crucially, program $P$ captures scene structures in three aspects. First, each entity function $f_{w} \in P$ transforms and composes multiple sub-entities (e.g., 64 squares) into a new, more complex entity (e.g., a board), naturally reflecting the hierarchical, part-whole relations in the scene, as the arrows in Fig. 2 highlight. Second, multiple entities $h_{j}$ 's may belong to the same semantic class $w$ (e.g., square), and can thus be represented by reusing the same entity function $f_{w}$ with distinct embeddings $z_{j}$ 's. Finally, each entity function also captures the precise spatial layout of the sub-entities by specifying their relative poses during the composition, such as 64 squares forming an $8 \times 8$ grid. + +The following expands on how functions from $P$ are defined and the program execution procedure to compute the represented scene $s$ . Notations are summarized in Tab. 1. + +Entity Function Definitions. An entity function $f_{w}$ outputs + +an entity $h$ of semantic class $w$ . The function $f_w$ takes two inputs: an embedding $z$ that specifies $h$ 's identity, as well as an ordered set $\gamma$ containing embeddings of all descendant entities of $h$ . Denote $\{h^{(i)}\}_{i=1}^N$ as the $N$ sub-entities of $h$ , where $N$ is a constant determined by $f_w$ . Then we have $\gamma = \{z^{(i)}, \gamma^{(i)}\}_{i=1}^N$ , where $z^{(i)}$ and $\gamma^{(i)}$ are the embeddings of $h^{(i)}$ and of the descendant entities of $h^{(i)}$ , respectively. Thus, $f_w(z, \gamma)$ is recursively defined as + +$$ +h = f _ {w} (z, \gamma) := \Psi_ {\text {u n i o n}} \left(\left\{\Psi_ {\text {t r a n s f o r m}} \left(h ^ {(i)}, t ^ {(i)} (z)\right) \right\} _ {i = 1} ^ {N}\right), \tag {2} +$$ + +where $h^{(i)} = f_{w^{(i)}}(z^{(i)},\gamma^{(i)})$ is a sub-entity of $h$ and $t^{(i)}(z)$ specifies its pose. + +Here, $f_{w}$ specifies the computation procedure to obtain output $h$ from $N$ sub-entities $\{h^{(i)}\}_{i = 1}^{N}$ via two operations: $\Psi_{\mathrm{transform}}$ applies an input-dependent pose $t^{(i)}(z)$ to a subentity $h^{(i)}$ , transforming it from its canonical frame to the world frame of $h$ , and $\Psi_{\mathrm{union}}$ composes multiple sub-entities into one single entity. Each sub-entity $h^{(i)}$ is computed by recursively applying an entity function $f_{w^{(i)}}$ also defined using Eq. (2). For instance, let $f_{w}$ denote the entity function that produces the board in Fig. 2 (namely, $w = \mathrm{board}$ ). This function $f_{w}$ composes 64 sub-entities $\{h^{(i)}\}_{i = 1}^{64}$ of the same class square, which are in turn obtained by executing the same entity function $f_{w^{(i)}} = f_{\mathrm{square}}$ with different embeddings $z^{(i)}$ and $\gamma^{(i)}$ . + +Program Execution. To obtain a scene $s$ from the Scene Language $\Phi(s) = (W, P, Z)$ , a program executor identifies a root entity function $f_{w_1}$ from $P$ that is not dependent by any other function (e.g., $w_1 =$ chessboard from Fig. 2), and evaluates this root function using the first embedding $z_1 \in Z$ and the rest as its descendants, $\gamma_1 := [z_2, \dots, z_J] \subset Z$ , to obtain $s = f_{w_1}(z_1, \gamma_1)$ . Evaluating $f_{w_1}(z_1, \gamma_1)$ expands the computation recursively to its children functions $h_j = f_{w_j}(z_j, \gamma_j)$ as defined in Eq. (2), obtaining a full sequence of all the entities $h_j$ of the scene, where $j = 1, 2, \dots, J, z_j \in Z, w_j \in W$ . The order in $Z$ corresponds to the depth-first-search of the computation graph starting from $z_1$ . An example computation graph is visualized on the right of Fig. 2. + +# 3.2. Scene Language as Programming Language + +We now concretize the definition in Sec. 3.1 with a domain-specific language (DSL) specified in Appendix B.1. We in- + +produce macro operations union for $\Psi_{\text{union}}$ , union-loop which calls union on entities evaluated in a for-loop, and transform for $\Psi_{\text{transform}}$ . Together with call which applies function calls, these four macro operations fully define entity functions in the DSL. Entity functions are identified with the associated words in the DSL via two special forms: bind, which binds entity function $f_w$ to word $w$ , and retrieve, which retrieves $f_w$ given $w$ and applies $f_w$ on actual embedding parameters. If $w$ is never bound to a function, it corresponds to an entity function with no sub-entities (i.e., $N = 0$ in Eq. (2)), in which case call returns a primitive entity with no children. + +The data type of an entity $h$ (Eq. (2)) is denoted as Entity. It stores two data fields Word and Embedding, describing the entity's semantic group and identity respectively, and stores each children entity together with its pose in the frame of $h$ . In particular, Embedding captures the visual details and requires a highly expressive representation, e.g., neural embeddings. We employ the textual embedding space of OpenCLIP-ViT/H [13] for parameterization, denoted as $\mathcal{Z}_{\mathrm{CLIP}}$ . It offers the advantage that embeddings can be either encoded directly from natural language or inferred from images with Textual Inversion [6]. Tab. 1 summarizes the expressions and data types in accordance with the notations from Sec. 3.1. + +# 4. Rendering + +Applying the proposed scene representation to image generation tasks requires rendering a Scene Language $\Phi(s)$ into images. To do so, first, the program interpreter evaluates $\Phi(s)$ to obtain a data object of type Entity. Afterward, a graphics renderer maps the Entity data object to its rendering parameter space and renders it into a final image. + +**Renderer Specifications.** We define the specification of a graphics renderer, a module in the proposed representation, as follows. A graphics renderer is determined by primitive parameter space $\Theta$ and a rendering operation $\mathcal{R} : \mathcal{P}(\Theta \times \mathcal{T}) \to \mathcal{I}$ , where $\mathcal{T}$ is the space of 3D affine transformations representing poses, $\mathcal{P}$ denotes all possible subsets, and $\mathcal{I}$ is the space of rendered images. In order to determine a mapping from program execution outputs of type Entity (Fig. 3b) to the admissible input domain of rendering operation $\mathcal{R}$ (Fig. 3c), we assume access to a reparameterization function $g_{\text{reparam}}$ that maps from Tuple[Word, Embedding] to $\Theta$ , and compute a primitive's pose $t \in \mathcal{T}$ by multiplying all Matrix values along the path from the primitive to the scene (root) in the entity hierarchy, analogous to computing limb poses in a kinematic tree. + +**Renderer Instantiations.** An example renderer instantiation is with Score Distillation Sampling (SDS) [25] guidance, where $\Theta$ is a differentiable 3D representation, and we specify $g_{\mathrm{reparam}}$ as follows. Recall that an entity is as- + +sociated with a Word value, e.g., moai for one statue from Figs. 1 and 3, and an Embedding value, e.g., $$ . For each primitive entity (i.e., an entity with no children), given these two value fields together with the fields of its ancestors, we use a manually specified language template $c$ , or a $$ moai, in the style of $$ , 3D model in this example, to embed them into $z = g_{\mathrm{CLIP}}(c) \in \mathcal{Z}_{\mathrm{CLIP}}$ where $g_{\mathrm{CLIP}}$ is the pre-trained CLIP text encoder. Then, $g_{\mathrm{reparam}}: \mathcal{Z}_{\mathrm{CLIP}} \to \Theta$ corresponds to the SDS-guided optimization to find a solution in $\Theta$ that aligns with the input condition $z \in \mathcal{Z}_{\mathrm{CLIP}}$ . Output 3D scenes can be personalized by editing embeddings, e.g., $$ which controls the global style in Fig. 1. + +For the underlying 3D representation, we use 3D Gaussian Splatting [18] where images are rendered by splatting a set of 3D Gaussians onto the image plane; other differentiable 3D representations such as neural fields will also be suitable. We base our implementation on GALA3D [44], and use MVDream [31] and a depth-conditioned ControlNet [40] for guidance. + +We refer to the renderer above as the Gaussians renderer. Other possible renderers include primitive-based renderers, such as Mitsuba [14] with graphics primitives of cubes, spheres, and cylinders; asset-based game engines, such as MineCraft [23]; and feed-forward inference of layout-conditioned text-to-image (T2I) diffusion models, such as MIGC [42], which achieves 2D bounding box conditioning by controlling attention layers from Stable Diffusion [28]). Tab. 2 shows a summary; details are in Appendix B.2. + +# 5. Inference via Pre-Trained Language Models + +We introduce a training-free method to infer the representation $\Phi(s) = (W, P, Z)$ from text or image descriptions of a scene $s$ . As explained below, we first prompt a pre-trained language model (LM) to generate the non-neural components $(W, P)$ and then obtain neural embeddings $(Z)$ from texts via the CLIP text encoder or from images with a pretrained text-to-image diffusion model. + +LMs have shown remarkable capability in code generation with common programming languages such as Python. In our implementation, we prompt LMs to generate Python scripts. We prompt the LM with the input condition, i.e., a scene description in texts or an image; a Python script of helper functions converted from the DSL in Sec. 3.2; and an example script using the helper functions. We use Claude 3.5 Sonnet [1] for all experiments for our method and LM-dependent baselines. Full LM prompts are in Appendix F.1. + +Function arguments in the LM-generated scripts, which are numeric values or string tokens, are converted to embeddings from $\mathcal{Z}_{\mathrm{CLIP}}$ (Sec. 3.2) using language templates and the CLIP text encoder $g_{\mathrm{CLIP}}$ . For example, in the raw LM output, function calls for white pieces in Fig. 2 have input attribute {"color": (.9, .9, .9)}}, and we prompt LM + +![](images/e27eb4cab15f145112aa840512c15b6c229eddc8217d5dc5227acbb3467ede56.jpg) +Ours (Two Views) + +![](images/4af557c8345bc77887bca7c4a53291a0ea45a466639a8173dc63102e93cd1cf8.jpg) +MVDream + +![](images/08e1d50ea05369dce00e693a6569a3526b2c2d929149ba051e5a307926eba983.jpg) +GraphDreamer + +![](images/69df59f0bd149f44a98e6ba74d3c0f69ebaabe3c8b10358f48bba40aaab9639c.jpg) +Ours (Two Views) + +![](images/1e447ef0eb88c727a2fb6beb46bff05b1108c8cc4607f1361eb8b443f9beca79.jpg) +"An 8-layer shelf" + +![](images/17dcbc9225b699fde1748a2d9cac5a1e34e40a64b2abff6dcd6d9f63f2b7717a.jpg) +MVDream + +![](images/d961f9ae50dd92a01ebdaec81324b0ea69724b0c7e6d431a0f12720b69cb1814.jpg) +GraphDreamer +"15 coke cans stacking in a pyramid" +Figure 4. Text-Prompted Scene Generation. Ours produce more accurate outputs than using no intermediate representation (MV-Dream) or scene graph (GraphDreamer). + +
MethodsNumeric PromptsGeneric Prompts
Alignment (↑)CLIP (↑)Count (↑)Alignment (↑)CLIP (↑)
GraphDreamer [7]3.56±7.380.297±0.0850.1115.75±20.390.230±0.097
MVDream [31]10.79±12.830.312±0.0460.1135.13±31.900.328±0.031
Ours85.65±13.710.351±0.0511.052.13±26.830.328±0.030
+ +to describe the color value as a word, and feed the word into $g_{\mathrm{CLIP}}$ to compute $< z68 >$ . For image-conditioned tasks, for each primitive entity in the execution output of $P$ , we first use GroundingSAM [20, 27] to segment out the region defined by the word associated with the entity. We then use Textual Inversion [6] to optimize an embedding to reconstruct the cropped image with the diffusion model training objective. The full process is deferred to Appendix C.2. + +# 6. Experiments + +# 6.1. Text-Prompted Generation and Editing + +Baselines. To evaluate our representation in text-prompted 3D generation task, we compare our inference pipeline with 3D scene generation methods using alternative intermediate representations, e.g., scene graph. In particular, we compare with GraphDreamer [7] as an exemplar approach, which generates scene graphs from texts via LM prompting and then synthesizes scenes conditioned on the graphs via SDS guidance. We further ablate the role of structural representation in this task by comparing ours with the backbone of our SDS-based renderer, MVDream [31], as a direct scene generation approach. Details are in Appendix C.3. + +![](images/f3a0dbd181d380a934c8da1ddd1c14df554101b4cc75b27b8bf985ee8c1656ad.jpg) +rnnnne nnnnne + +![](images/0cad2d9bdb96aea4db1e662d978c3cbade744f8ba1811b90ee12e0c9a922329f.jpg) +Gaussians + +![](images/bf75a7d2b2bd68f2e3c7bf77b9f3991a652d194a65c58e9514a1f17cd5a83443.jpg) + +![](images/62f64907d5772a0d50b7faa55a7591ff8d4868ceb72f11c899de8cb273cf4cbd.jpg) +Minecraft + +model of + +a tennis court" + +![](images/53a0ac878901b617508688e452da050daec5f8b222c8d07ac2909851d6e6ee60.jpg) +nne nnnnne +Mitsuba +"City skyline at sunset" +Figure 5. Renderings Across Graphics Renderers. Different renderers produce renderings that adhere to the same representation and are therefore visually aligned, while each exhibits a different imaging style. Text inputs are shown at the subfigure bottoms. + +![](images/1ebaba16252602202b24c74d24aaf78a0e00d640fb58f4d0c1c45f2c05a7ed21.jpg) +"A detailed model of a tennis court" +T2I + +![](images/e13064e231a80ebe12101c498c7a1a270393d4a64891166b836853c08b57ea2a.jpg) + +![](images/1f37b8c6405b561fb401bb003fd77b7d99ceaf19c55195633f0008a2eedfd28b.jpg) +Gaussians + +![](images/171ca8f7d2774261606cbe2fce8a34b241888d6ff61285a48ed75b01c7331def.jpg) + +![](images/d7d94cd6445d432d89f038748b81666e3686adcc049597e1f597bac16faab3ec.jpg) +Minecraft + +Gaussians + +![](images/20989f2745383b633cb4f968f4a060e28ba844e04dc620d60fb36de9eeec3f4f.jpg) +"An interior scene of a university lecture hall" +Mitsuba + +![](images/df7a2c73d654451176e8738b0c791f7d0eb1161195b6b7dd4339c600201e23c1.jpg) +T2I +"Marionette" + +Table 3. Evaluation for text-prompted 3D generation. We report the following metrics averaged across 9 numeric and 8 generic prompts: the percentages of user preferences for prompt alignment, CLIP similarity scores, and counting accuracy (numeric prompts only, 0 for inaccurate and 1 for accurate). Here $\pm$ denotes standard deviation across scenes. + +
MethodsCLIP [13] (↑)Dynamic [12] (↑)
4D-fy [2]0.3520.2%
Ours0.3415.9%
+ +
MethodsLPIPS [41] (↓)
GraphDreamer [7]0.811
Ours0.681
+ +Table 4. Evaluation. Left: For text-prompted 4D generation, 4Dfy achieves higher prompt alignment measured with CLIP similarity but has small motion. Right: For image-prompted scene generation, our results are better aligned with input image prompts compared to GraphDreamer. + +Metrics. We conduct a user study where 103 users are asked to choose one of the three videos, showing 360-renderings from ours and two baselines in a randomized order, that aligns the best with text prompts. Following [7], for each scene, we compute the cosine similarity of CLIP [13, 26] embeddings between a rendering and the text prompt and report the average over 120 viewpoints. We further report whether outputs have the correct object counts. Details are deferred to Appendix C.4. + +Results. Tab. 3 shows that our method achieves a more favorable prompt alignment than the baselines as measured in the user study and CLIP similarity, and has a clear advantage in counting accuracy. + +Fig. 4 and Appendix D.1 show qualitative results of ours using the SDS-based renderer. Compared to the direct 3D scene generation method MVDream, our approach is compositional and adheres more closely to input prompts in scenes involving multiple objects. Compared to a scene graph representation, where entity relations are restricted to be between two objects and are bottlenecked by the coarseness of natural language descriptions, e.g., "aligned in a row", a program-based representation offers more flexible and precise specifications for relations, e.g., the particular coke can arrangement in Fig. 4. This brings the practical benefit of offloading the burden of generating scenes involving complex entity relations from the T2I model (used + +Ours (w/ tracking) +![](images/f49a812c2ecb43a64bde2247baa32f8b5545f3592c0b9a38d8cb63747a3176b9.jpg) +"Carousel with a small canopy" + +Frame 1 +![](images/c20f40760c1fbe7feef3c9407018a4359bf2a91d7bc20f6e1f781eb8c37e55bb.jpg) +"Solar system model" + +![](images/3c350ffa2e0ebbfebe499673b10061231a397dc4e190c8adfd9282befa846aa5.jpg) +1 Fran + +![](images/afbf2b027051b852cb9ba9f63bbaff50eb9545b7f1cd4a85d1b31e92560f88d6.jpg) + +Ours (w/ tracking) +4D-fy +![](images/cc8404f3abf0418a4da2d4220b055787b6237181129b3ed92ce364883bfeeac1.jpg) +"A toy wind turbine" + +![](images/7910274de014f94814f54f0f04fc564cb16421ef97fb1ef2c9f05e13c522cc53.jpg) +Frame 1 Frame 2 +Ours Frame 1 Frame 2 (w/ tracking) 4D-fy + +Figure 6. Text-Prompted 4D Scene Generation. The proposed representation may also capture scene dynamics and be applied for synthesizing 4D scenes. It explicitly represents the temporal correspondence of an entity in a dynamic scene. Each colored trajectory denotes tracking of a temporally moving point. Compared to 4D-fy, our method produces more salient motion. +Initial (Mitsuba) +![](images/a787646a9aaaa304bb65bf72cbbe71162ed1e6b6c87319fa4b6843b27344f8e8.jpg) +"Bramante Staircase, + +![](images/e469e9debb3146c85c5714ef38ed26a94b415c7ae091a9902459f73936ba649c.jpg) +Initial (Gaussians) + +Initial (Mitsuba) +![](images/420fa821f36b296c62832668367680129116b5b873c2fbc42610d2023388549c.jpg) +"An 8-layer 3-color Jenga set at game start" + +Initial (Gaussians) +![](images/7494e9cb6314093d089f31e4001fdb8dbc6729ece29f9a9f093764aa84fa2726.jpg) +"Remove 2 blocks from second top layer and the tower should not fall" + +Initial (Mitsuba) +![](images/a12ecbca67a42acbab6150493ac5546c4f5a492eaaff780007d09fc1b3a17559.jpg) +"A fractal tree-like structure" + +![](images/3233401808ce1fe1083368d3b2ed9e4b3fedcf21a68455a3f6f3b3204703799e.jpg) +Initial (Gaussians) + +Scene Language +![](images/610909dad8a134a295bd73cb6e62fd5d10926697104846242a3b77f7868e23a2.jpg) +"Shrink staircase radius by $80\%$ + +![](images/cfff393f4519e1ea777c9af8969be3cc683240a5bccea7b6b28a8b59e8dead52.jpg) +Edited (Mitsuba) + +![](images/8306f9060792f238a687a5bc153c62003f76daaeec9fb37322978af119114bd6.jpg) +Edited (Gaussians) + +![](images/43d44208cd4c2899eaf287343dddd7127d1e00006e4e749859b7d03246b7bbbd.jpg) +("Jenga layer" (lambda(is-horizontal colors remove-blocks))) + +![](images/15e6717a6f34304ee89a4456267eb9d2183c7c252d2df8941f6474e8490f6c73.jpg) +Scene Language +Edited (Mitsuba) + +![](images/577780964a6c7f543880c084b480fe9d1fb2a0207168620598a031e91d056514.jpg) +Edited (Gaussians) + +union left-branch right-branch) +Scene Language +Figure 7. Scene Editing with Language Instructions. Our representation is editing-friendly. Bottoms of each row show initial scene descriptions and editing instructions in the format of user text prompts. We prompt an LM to infer the initial Scene Language as well as the edits (shown with texts in red), and show image renderings with two renderers. +![](images/d9ec9d0a675e0861d0dc26ec7198490cacfb05b2712be6cebaed6a5d9be83a5f.jpg) +"Make branching structure to be trinary and 3D" + +![](images/ee0dc5c0de5a5971fd3edb07b78f9de8f6365f884e8932c76844a4a293ed31b0.jpg) +Edited (Mitsuba) + +![](images/d14d9edc0360e884bedfe29ddbbb652449c9d3978f5fa9f45216776f8300e5d7.jpg) +Edited (Gaussians) + +for SDS guidance in both ours and GraphDreamer) towards LM, leading to accurate and detailed generation results. + +Across Graphics Renderers. Fig. 5 shows that our representation can be rendered with different renderers (Sec. 4), producing aligned outputs in different visual domains. + +4D Generation. The inference method from Sec. 5 can be + +applied to 4D scene generation. The 4D scene representation is identical to Eq. (1), except that there is an additional 4D entity function in the program $P$ . + +We compare with 4D-fy [2] as an exemplar text-to-4D generation method. We measure the prompt alignment of output dynamic renderings using CLIP similarity between + +![](images/ee2fa90caf16d39cd01114c508865773a30df9a29d082d0207d3eb25816d7e28.jpg) +Figure 8. Scene Editing with Image Instructions. This figure shows style transfer results under the same setup as in Fig. 1. Replacing embeddings of target entities in a Scene Language with a new embedding specified by a user image achieves the editing effect. + +![](images/f2cb0b9931875f3420055327cf07469e908616130d0ec185f2f602efc9df62a4.jpg) +Ours (Gaussians) + +![](images/1d0ff618feb55adc6b7440bfa611a08f3f04418c1b652837e999352585f4e75c.jpg) + +![](images/82e01fea7ee36cf75c2f365ecbf2039bfa9af69070e5447f60c682e5c5bf09cb.jpg) +Ours (Mitsuba) + +![](images/554887df68e7ad17a2fa6c4436f9e403fc9e530523f6d0eb3c7bbe3d2553425d.jpg) +GraphDreamer + +![](images/e9037c5d4e535ced3a3e62ee70bdb935fcd0250f3045d676182baaa100591ca5.jpg) +Input Image + +![](images/026dc445a1bcdfe81d8f38f050465c7e08a37a8b293e8172abcbeeaa9af88dd2.jpg) +Input Image +Input Image + +![](images/b9c69da6988cc6880d39835d2b1d181b7130f140e9686040476ada7e6076ce0f.jpg) +Ours (Gaussians) + +![](images/b7b557b2f0107c4970eeaa80279f0ee5aca402c9d72e650ba029d7fb95357dcd.jpg) +Ours (Mitsuba) +GraphDreamer + +![](images/8eca0563e95db2002d7dc86d56ba18b7f95f9c233482e98ec050e60d3f1c8b10.jpg) + +![](images/5c47dda6ab1505373b3e2088bdb2466a99391f2e2531ccf8d396f21ac2384345.jpg) +Ours (Gaussians) + +![](images/84a15ea6fcd20e7c50254e09d4b49d26b45a4cd70b9e5d69b0afc090e6339078.jpg) +Ours (Mitsuba) + +![](images/462c165752163db06354d52aa4e89439ba906ed2691b38e27f64e0ff2e8b1be8.jpg) +GraphDreamer +Figure 9. Image-Prompted Scene Generation. Compared to scene graphs from GraphDreamer, programs from our representation encode additional scene structure, e.g., repetitions, and specify geometric entity relations more precisely. Embeddings from ours further enable visual identity preservation in the outputs. + +rendered frames and input texts, and the motion magnitude using dynamic degrees [12] computed using RAFT [35] optical flow estimation over neighboring frames. 4D-fy results have a higher per-frame alignment score than ours but present small dynamic motions. See Fig. 6, the webpage, and Appendix D.2 for further results. + +Scene Editing. Scenes synthesized from our proposed representation can further be edited by prompting LM with its previously generated script and an editing instruction in natural language instruction. Results are shown in Fig. 7. Our representation provides an interpretable and intuitive interface for scene editing, as functions have explicit semantic meanings associated with words, and function reuse significantly improves program readability. Furthermore, since the structure of programs reflects the structure of scenes, editing program parameters leads to precise changes in the scenes while preserving the original structure, e.g., the circular arrangement of staircases in Fig. 7. Desirable editing effects involving multiple primitives, or all staircases in this example, can be effectively achieved via only minor changes in the program space. Finally, the program structure itself, such as the function header in the Jenga set example, can be adjusted for editing to achieve localized edits that only affect relevant parts of the scene. Further results + +are shown in Appendix D.3. + +# 6.2. Image-Prompted Generation and Editing + +Texts provide an easy user interface for scene generation as shown in Sec. 6.1, and images allow users to specify further intentions that could be hard to describe with natural language alone. Integrating neural embeddings increases the expressivity of the Scene Language and allows it to encode visual details from image prompts, and as a result, enhances our method's controllability in image-prompted scene generation and editing tasks. + +Generation. Fig. 9 shows results on the task of generating 3D scenes consistent with the structure and content from a user-provided image, termed as the "inverse semantics" task in GraphDreamer [7]. We report the LPIPS [41] score, i.e., the AlexNet [21] feature dissimilarity, between an input image and an output rendering, averaged over 3 scenes with 20 viewpoints each in Tab. 4. Embeddings allow our representation to encode visual details from inputs, such as the Coke can texture in Fig. 9, which is not captured in scene graphs from GraphDreamer using per-object natural language descriptions. This is reflected quantitatively in Tab. 4. + +Editing. Our representation applies to image-prompted editing tasks such as style transfer (Fig. 1). Edits can target partial or full scenes as shown in Fig. 8(a) and (b), respectively, by updating embeddings for corresponding entities. + +# 7. Conclusion + +We have introduced a visual scene representation, termed the Scene Language, which encodes three key aspects of visual scenes: scene structure, such as hierarchy and repetition, specified via programs; semantics of individual components succinctly summarized via words; and identities of each component precisely captured via neural embeddings. We formalize the representation as a programming language defined using a DSL. We show that the Scene Language can be efficiently inferred from both text and image inputs leveraging pre-trained language models. Once the program is executed, the resulting scene can be rendered into images using a variety of graphics renderers. Our Scene Language produces 3D and 4D scenes with high fidelity and enables easy and precise editing. + +Acknowledgments. This work is in part supported by AFOSR YIP FA9550-23-1-0127, ONR N00014-23-1-2355, ONR YIP N00014-24-1-2117, ONR MURI N00014-22-1-2740, and NSF RI #2211258 and #2338203. YZ is in part supported by the Stanford Interdisciplinary Graduate Fellowship. + +# References + +[1] Anthropic. The Claude 3 model family: Opus, Sonnet, Haiku, 2024. 5 +[2] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B. Lindell. 4D-fly: Text-to-4D generation using hybrid score distillation sampling. CVPR, 2024. 6, 7, 16 +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2 +[4] Blender Online Community. Blender, 1994. 3 +[5] Boyang Deng, Sumith Kulal, Zhengyang Dong, Congyue Deng, Yonglong Tian, and Jiajun Wu. Unsupervised learning of shape programs with repeatable implicit parts. In NeurIPS, 2022. 3 +[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 3, 5, 6, 14 +[7] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf. GraphDreamer: Compositional 3D scene synthesis from scene graphs. In CVPR, 2024. 3, 6, 8, 14 +[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3 +[10] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. SceneCraft: An LLM agent for synthesizing 3D scene as Blender code. In ICLR Workshop, 2024. 3 +[11] Siyuan Huang, Siyuan Qi, Yixin Zhu, Yinxue Xiao, Yuanlu Xu, and Song-Chun Zhu. Holistic 3D scene parsing and reconstruction from a single RGB image. In ECCV, 2018. 3 +[12] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 6, 8, 16 +[13] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, 2021. 5, 6, 16 + +[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. 5 +[15] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, 2015. 3, 16 +[16] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, 2018. 3 +[17] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy Mitra, and Daniel Ritchie. ShapeAssembly: Learning to generate programs for 3D shape structure synthesis. ACM TOG, 39(6):1-20, 2020. 3 +[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 5, 11 +[19] Diederik P Kingma. Auto-encoding variational bayes. In ICLR, 2014. 3 +[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6 +[21] Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014.8 +[22] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas Guibas. StructureNet: Hierarchical graph networks for 3D shape generation. ACM TOG, 38 (6):1-19, 2019. 3 +[23] Mojang Studios. Minecraft, 2009. 5 +[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 3 +[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 5, 12 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 13 +[27] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded SAM: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 6, 14 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 5, 13 +[29] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3 +[30] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In CVPR, 2018. 3, 16 + +[31] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. In ICLR, 2024. 5, 6, 14 +[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3 +[33] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3D-GPT: Procedural 3D modeling with large language models. arXiv preprint arXiv:2310.12945, 2023. 3 +[34] Hou In Ivan Tam, Hou In Derek Pun, Austin T Wang, Angel X Chang, and Manolis Savva. SceneMotifCoder: Example-driven visual program learning for generating 3D object arrangements. arXiv preprint arXiv:2408.02211, 2024.3 +[35] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 8 +[36] Yonglong Tian, Andrew Luo, Xingyuan Sun, Kevin Ellis, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Learning to infer and execute 3D shape programs. In ICLR, 2019. 3, 16 +[37] Yutaro Yamada, Khyathi Chandu, Yuchen Lin, Jack Hessel, Ilker Yildirim, and Yejin Choi. L3GO: Language agents with chain-of-3D-thoughts for generating unconventional objects. In ICLR Workshop, 2024. 3 +[38] Alan Yuille and Daniel Kersten. Vision as Bayesian inference: analysis by synthesis? Trends in Cognitive Sciences, 10(7):301-308, 2006. 3 +[39] Chi Zhang, Penglin Cai, Yuhui Fu, Haoqi Yuan, and Zongqing Lu. Creative agents: Empowering agents with imagination for creative tasks. arXiv preprint arXiv:2312.02519, 2023. 3 +[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 5 +[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 8 +[42] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. MIGC: Multi-instance generation controller for text-to-image synthesis. In CVPR, 2024. 5, 13 +[43] Mengqi Zhou, Jun Hou, Chuanchen Luo, Yuxi Wang, Zhaoxiang Zhang, and Junran Peng. SceneX: Procedural controllable large-scale scene generation via large-language models. arXiv preprint arXiv:2403.15698, 2024. 3 +[44] Xiaoyu Zhou, Xingjian Ran, Yajiao Xiong, Jinlin He, Zhiwei Lin, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. GALA3D: Towards text-to-3D complex scene generation via layout-guided generative Gaussian splatting. In ICML, 2024. 5, 12 +[45] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In ECCV, 2016. 3 \ No newline at end of file diff --git a/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/images.zip b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d8bf7a1289602da766500fdc7f41381e49e0d5a6 --- /dev/null +++ b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19023f90011227fc8cca286fa7367a1a6edb3ac7500da696e3c5cabd58f78ab8 +size 579128 diff --git a/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/layout.json b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6e50a24ac2bda48147d6b0451afb7cf430dc3610 --- /dev/null +++ b/2025/The Scene Language_ Representing Scenes with Programs, Words, and Embeddings/layout.json @@ -0,0 +1,13064 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 103, + 438, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 103, + 438, + 140 + ], + "spans": [ + { + "bbox": [ + 173, + 103, + 438, + 140 + ], + "type": "text", + "content": "The Scene Language: Representing Scenes with Programs, Words, and Embeddings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "spans": [ + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "text", + "content": "Yunzhi Zhang" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "text", + "content": " Zizhang Li" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "text", + "content": " Matt Zhou" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "text", + "content": " Shangzhe Wu" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "text", + "content": " Jiajun Wu" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "text", + "content": "Stanford University " + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 160, + 491, + 196 + ], + "type": "text", + "content": "UC Berkeley" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 204, + 249, + 393 + ], + "blocks": [ + { + "bbox": [ + 62, + 204, + 249, + 393 + ], + "lines": [ + { + "bbox": [ + 62, + 204, + 249, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 249, + 393 + ], + "type": "image", + "image_path": "91213d4fe6a537581af84bb20e44f900ddcb888c526fac06413df4a3706a9ef2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 396, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 54, + 396, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. Structured Scene Generation and Editing Using the Scene Language. We develop a scene representation for 3D scene generation and editing tasks. Given textual scene descriptions, the representation can be inferred by a pre-trained large language model, rendered in 3D, and edited following language instructions. The representation contains a program consisting of semantic-aware functions bound to words, providing high interpretability and an intuitive scene-editing interface, and embeddings enabling editing with fine controls, e.g., transferring the style of " + }, + { + "bbox": [ + 54, + 396, + 555, + 464 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 54, + 396, + 555, + 464 + ], + "type": "text", + "content": " from a user's image to the output by updating " + }, + { + "bbox": [ + 54, + 396, + 555, + 464 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 54, + 396, + 555, + 464 + ], + "type": "text", + "content": " which controls global attributes of the scene. Project page: https://ai.stanford.edu/~yzzhang/projects/scene-language/." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 204, + 539, + 393 + ], + "blocks": [ + { + "bbox": [ + 256, + 204, + 539, + 393 + ], + "lines": [ + { + "bbox": [ + 256, + 204, + 539, + 393 + ], + "spans": [ + { + "bbox": [ + 256, + 204, + 539, + 393 + ], + "type": "image", + "image_path": "76cf9657f56e00cc283db7ce71606672ad889c072439cd87448fabf06070e054.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 471, + 200, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 471, + 200, + 483 + ], + "spans": [ + { + "bbox": [ + 152, + 471, + 200, + 483 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 496, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 496, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 53, + 496, + 297, + 700 + ], + "type": "text", + "content": "We introduce the Scene Language, a visual scene representation that concisely and precisely describes the structure, semantics, and identity of visual scenes. It represents a scene with three key components: a program that specifies the hierarchical and relational structure of entities in the scene, words in natural language that summarize the semantic class of each entity, and embeddings that capture the visual identity of each entity. This representation can be inferred from pre-trained language models via a training-free inference technique, given text or image inputs. The resulting scene can be rendered into images using traditional, neural, or hybrid graphics renderers. Together, this forms an automated system for high-quality 3D and 4D scene generation. Compared with existing representations like scene graphs, our proposed Scene Language generates complex scenes with higher fidelity, while explicitly modeling the scene structures to enable precise control and editing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 471, + 394, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 471, + 394, + 483 + ], + "spans": [ + { + "bbox": [ + 315, + 471, + 394, + 483 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 590 + ], + "type": "text", + "content": "How do you describe a scene? Imagine that you just traveled to Easter Island and would like to explain to Alice the wondrous scene of Ahu Akivi: \"There are seven moai in a row, facing the same direction.\" \"What is a moai?\" Alice asked. \"A moai is a stone human figure without legs, but each also looks slightly different.\" At this point, you realize it seems difficult to precisely explain the scene using natural language alone." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In fact, this example highlights a complete scene representation requires at least three types of complementary information: structural knowledge, which is about the joint distribution of multiple instances, like \"seven moai in a row, facing the same direction,\" most naturally described as programs; category-level semantics, which may be shared across instances, often described in words, such as \"moai\"; and instance-level intrinsics, tied to the identity of each specific object or part, such as its geometry, color, and texture, which is hard to describe but easy to recognize." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24625" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 61, + 71, + 361, + 393 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 361, + 393 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 361, + 393 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 361, + 393 + ], + "type": "text", + "content": "bind \"chessboard\" (lambda (embedding ...) (union (call \"board\" ...) (call \"chess pieces\" ...)))) \n(bind \"board\" (lambda (embedding ...) (define (create-squares i) ...) (union-loop 64 create-squares))) \n(bind \"chess pieces\" (lambda (embedding ... ) (define (create-pawns i) (let* ((white-pawn (call \"pawn\" ...) (black-pawn (call \"pawn\" ...) (white-pose (translate (list (* i 0.125) 0.02 0.125)))) (black-pose (translate (list (* i 0.125) 0.02 0.75))) (white-pawn-t (transform white-pawn white-pose)) (black-pawn-t (transform black-pawn black-pose))) (union white-pawn-t black-pawn-t))) (define pieces-order (((\"rook\" \"knight\" \"bishop\" \"queen\" \"king\" \"bishop\" \"knight\" \"rook\")) (define (create-other-pieces i) (let* ((piece-order (list-ref pieces-order i)) (white-piece (call piece-order ...)) (black-piece (call piece-order ...)) (white-pose (translate (list (* i 0.125) 0.02 0)) (black-pose (translate (list (* i 0.125) 0.02 0.875))) (white-piece-t (transform white-piece white-pose)) (black-piece-t (transform black-piece black-pose))) (union white-piece-t black-piece-t))) (union (union-loop 8 create-pawns) (union-loop 8 create-other-pieces)))" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "code", + "guess_lang": "lisp" + }, + { + "type": "image", + "bbox": [ + 367, + 70, + 553, + 395 + ], + "blocks": [ + { + "bbox": [ + 367, + 70, + 553, + 395 + ], + "lines": [ + { + "bbox": [ + 367, + 70, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 367, + 70, + 553, + 395 + ], + "type": "image", + "image_path": "2a8e9bc56400491082bb426f615365567f7f2a4fa8bd9bd40e31057e3889ab17.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 400, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 54, + 400, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 555, + 478 + ], + "type": "text", + "content": "Figure 2. Overview. A Scene Language represents a scene with three components: a program consisting of entity functions, a set of words (e.g., pawn) denoting the semantic class of the entity functions, and a list of embeddings (e.g., ) capturing the identity of each entity in the scene. Each entity function is bound with an entity class name given by a word, and maps an input embedding to an output entity of that class. Executing the program evaluates entity functions to compute the full set of entities in the scene. The resulting computation graph, illustrated on the right, captures the dependency structure among entities (indicated by arrows). The program shown is converted from our text-conditioned inference method's output, with details included in Appendix F.2.1; it is written in Lisp-style syntax for brevity and is implemented with Python in practice (Sec. 3.2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 484, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 484, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 54, + 484, + 295, + 604 + ], + "type": "text", + "content": "Modern AI techniques provide natural grounding for each of the three modalities, while also falling short of capturing all: in-context learning of pre-trained language models (LMs) enables the inference of domain-specific programs [3]; LMs themselves capture rich semantic information based on words in natural language; and embeddings obtained via techniques like textual inversion [6] or low-rank adaptation [9] best capture object identity. However, none of these existing representations alone is sufficient for scene generation and editing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "We introduce the Scene Language, a representation that integrates the three modalities—programs, words, and embeddings—to precisely and concisely describe the structure, semantics, and identity of visual scenes. In the Scene Language, a program specifies a computation process that defines the organization of a collection of entities in the scene, including extrinsics like poses and structural regularity like repetitions. Each entity is associated with a word referring to its semantic group, as well as an embedding describing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 484, + 437, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 437, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 437, + 495 + ], + "type": "text", + "content": "its instance-specific attributes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 497, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 640 + ], + "type": "text", + "content": "In addition to the representation itself, we propose a training-free inference module using a pre-trained LM as a backbone to infer the Scene Language from texts and images. When provided with a domain-specific language (DSL) for scenes, LMs decompose the task of complex scene generation into simpler tasks of scene component generation by predicting their corresponding modular functions. We also discuss possible neural, traditional, and hybrid graphics engines that render images based on the representation. Together, the Scene Language, the inference module, and the renderer lead to a system for high-quality, detailed 3D and 4D scene generation and editing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 325, + 642, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 642, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 325, + 642, + 482, + 652 + ], + "type": "text", + "content": "In summary, our contributions include" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 654, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 654, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 314, + 654, + 553, + 689 + ], + "type": "text", + "content": "1. A scene representation, the Scene Language, capturing structure, semantics, and identity of visual scenes using programs, words, and embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "2. A training-free method that infers the representation from texts and/or images using pre-trained LMs." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24626" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "3. A generic rendering module that renders the Scene Language into an image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 120 + ], + "type": "text", + "content": "4. Empirical results on text- and image-prompted scene generation and editing tasks." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 130, + 141, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 130, + 141, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 141, + 143 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 150, + 296, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 150, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 296, + 246 + ], + "type": "text", + "content": "Visual scene representations are arguably the most fundamental problem in computer vision; thus, for sure, we may not enumerate all related work. As our Scene Language comprises programs, words, and embeddings, we organize our discussion accordingly into three categories: scene representations using program-based representations (Sec. 2.1), semantic graph-based representations (Sec. 2.2), and a pre-trained generative model's latent space (Sec. 2.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 253, + 235, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 235, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 235, + 266 + ], + "type": "text", + "content": "2.1. Representing Scenes as Programs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 270, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 295, + 474 + ], + "type": "text", + "content": "Programs can specify not only the relations among scene components mentioned in Sec. 2.2, but also structural patterns such as hierarchy and repetitions, making them suitable as explicit descriptions of scene structures. Prior works have proposed to use programs in the form of sequences of execution commands as object-centric representations, followed by neural executors that render the programs into 3D shapes [5, 30, 36]. In comparison, ShapeAssembly [17] introduces higher-level functions with semantically meaningful function names, e.g., \"chair\" and \"back\", to its program representation. Both ShapeAssembly and ours adopt the design principle of function abstraction, which results in clearly stated hierarchy relation among components and better program editability. However, ShapeAssembly uses cuboids as the shape representation and does not model appearance, while ours allows for more precise geometry and appearance modeling using expressive neural embeddings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 570 + ], + "type": "text", + "content": "All representations above require 3D datasets for training. More recently, with the advance of language models (LMs), several methods [10, 33, 34, 37, 39, 43] have proposed to use zero-shot LM inference for generating programs that will be rendered into scenes. These methods operate on top of program syntax from specific graphics renderers such as Blender [4], and they do not permit parameters in high-dimensional embedding spaces unlike ours." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 576, + 281, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 281, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 281, + 589 + ], + "type": "text", + "content": "2.2. Representing Scenes with Semantic Graphs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Prior semantic scene representations often adopt a graph to encode semantic scene components, such as objects and parts. In particular, [11, 38] propose to employ a parse graph of context-free grammar, using terminal nodes to correspond to objects and their attributes, to represent a scene. Both works employ an analysis-by-synthesis approach to infer the representation from images that heavily rely on domain-specific priors. Alternative representations include scene graph [7, 15, 16], where each node in a graph corresponds to an object and an edge corresponds to a pairwise" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 553, + 207 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 207 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 207 + ], + "type": "table", + "html": "
Sec. 3.1Sec. 3.2Definition
Expressions
ΨtransformtransformTransform an entity
ΨunionunionCompose entities
fw:z,γ→h<entity FUNC>Mapping embeddings to entity
fw(z,γ)(call <word>)Entity function application
Data Types
wWordDescribing semantics
tMatrixEntity pose
zEmbeddingSpecifying entity identity
γList [Embedding]Specifying descendant entities' identities
hEntityAn entity
sEntityThe represented scene
", + "image_path": "856fe8cbee4c4db56eeec65b3a52f0bbb778a57473222e65e0279d1ff32a259a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 335, + 209, + 533, + 220 + ], + "lines": [ + { + "bbox": [ + 335, + 209, + 533, + 220 + ], + "spans": [ + { + "bbox": [ + 335, + 209, + 533, + 220 + ], + "type": "text", + "content": "Table 1. Summary of Notations in Secs. 3.1 and 3.2." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 230, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 553, + 314 + ], + "type": "text", + "content": "relation, and StructureNet [22], which focuses on an object-centric setting and uses nodes for object parts. While these representations preserve the high-level semantics of scenes or objects, they leave out low-level precision; thus, geometric, textural, or relational details that cannot be fully specified by language or hand-crafted rules are often ignored. We address this issue via the inclusion of embeddings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 324, + 553, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 348 + ], + "type": "text", + "content": "2.3. Representing Scenes with Generative Model Latents" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 355, + 555, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 522 + ], + "type": "text", + "content": "The latent space of visual generative models can serve as a representation space for visual scenes. Such latent space can effectively capture the exact visual content of scenes, including geometry and appearance details, and can be either directly inferred, e.g., in variational inference [19] and model inversion [45]. More recently, text-to-image diffusion models have shown remarkable results in image synthesis. This class of models offers several candidate representation spaces including the space of textual embeddings [6], low-rank network weights [9], full model weights [29], or noise vectors in the diffusion process [8, 24, 32]. However, such representations typically do not offer interpretable semantics or explicitly encode hierarchical scene structures unlike ours." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 536, + 435, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 435, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 435, + 551 + ], + "type": "text", + "content": "3. The Scene Language" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "We aim to design a visual scene representation that encodes the structure, semantics, and visual content of scenes. Towards this goal, we propose the Scene Language, which represents a scene with three components: a program that encodes scene structure by specifying the existence and relations of scene components, which we will refer to as entities; words in natural language that denote the semantic group of each entity in the scene; and neural embeddings that pertain the low-level visual details and identities of the entities by permitting an expressive input parameter space. In the following, we will first give a formal definition of the representation (Sec. 3.1), and then introduce a domain-specific language (DSL) (Sec. 3.2) as its realization." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24627" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 550, + 132 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 550, + 132 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 550, + 132 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 550, + 132 + ], + "type": "image", + "image_path": "886bb5c93805d49be452cb8203f429239c72eadd2c3abbae32ccf42ee8a61330.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 134, + 555, + 168 + ], + "lines": [ + { + "bbox": [ + 55, + 134, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 555, + 168 + ], + "type": "text", + "content": "Figure 3. Rendering. Given a Scene Language in (a), a program interpreter executes the program to obtain a data object in (b). A graphics renderer first reparameterizes the data object from (b) into the renderer-specific parameter space, and then executes the rendering operation " + }, + { + "bbox": [ + 55, + 134, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 134, + 555, + 168 + ], + "type": "text", + "content": " to obtain the final image output in (d)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 173, + 293, + 240 + ], + "blocks": [ + { + "bbox": [ + 57, + 173, + 293, + 240 + ], + "lines": [ + { + "bbox": [ + 57, + 173, + 293, + 240 + ], + "spans": [ + { + "bbox": [ + 57, + 173, + 293, + 240 + ], + "type": "table", + "html": "
RendererExamples
Rendering Operation RParameters from Θg_reparam
Primitive-basedLight transportShape and BRDFLM inference
Asset-basedRay tracingAsset metadataLM inference
SDS-basedGaussian splattingGaussian parametersSDS optimization
T2I modelFeed-forward passText embeddings in ZCLIP encoding
", + "image_path": "1dc539938ff93fbf8fed8ab0afa7d16cc8219345c552ee4b184c3052808b120c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 98, + 241, + 253, + 252 + ], + "lines": [ + { + "bbox": [ + 98, + 241, + 253, + 252 + ], + "spans": [ + { + "bbox": [ + 98, + 241, + 253, + 252 + ], + "type": "text", + "content": "Table 2. Graphics Renderers Examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 258, + 162, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 162, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 162, + 269 + ], + "type": "text", + "content": "3.1. Formal Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "type": "text", + "content": "The Scene Language for a scene " + }, + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\Phi(s)" + }, + { + "bbox": [ + 55, + 275, + 295, + 300 + ], + "type": "text", + "content": ", is formally defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 307, + 295, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 307, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 134, + 307, + 295, + 321 + ], + "type": "interline_equation", + "content": "\\Phi (s) := (W, P, Z). \\tag {1}", + "image_path": "71509372b7939af174286530397f2262716facc99de615cf08b358391811c653.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " is a collection of phrases in natural language, referred to as words, e.g., " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "W = \\{\\text{board}, \\text{pawn}, \\dots\\}" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " as illustrated in Fig. 2, and " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "P := \\{f_w\\}_{w \\in W}" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " is a program consisting of a set of entity functions " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " indexed by " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "w \\in W" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": ". Each entity function " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " defines a class of entities in the scene; it is uniquely identified by the associated " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "w \\in W" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": ", which concisely summarizes the semantic meaning of the defined class. Each of " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " maps neural embeddings " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "(z, \\gamma)" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " to a specific entity " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " in the scene, where " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " specifies the attributes and identity of " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": ", like a specific color of a pawn, and " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " specifies identities of its sub-entities. Hence, the complete Scene Language " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\Phi(s)" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " of a particular scene " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " also contains an ordered set of neural embeddings " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "Z := [z_1, z_2, \\dots, z_J]" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " specific entities " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "[h_1, h_2, \\dots, h_J]" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " from scene " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "h_j \\in J" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": " is computed by applying an entity function on inputs " + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "inline_equation", + "content": "z_j \\in Z, \\gamma_j \\subset Z" + }, + { + "bbox": [ + 54, + 327, + 296, + 519 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "text", + "content": "Crucially, program " + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "text", + "content": " captures scene structures in three aspects. First, each entity function " + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "inline_equation", + "content": "f_{w} \\in P" + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "text", + "content": " transforms and composes multiple sub-entities (e.g., 64 squares) into a new, more complex entity (e.g., a board), naturally reflecting the hierarchical, part-whole relations in the scene, as the arrows in Fig. 2 highlight. Second, multiple entities " + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "inline_equation", + "content": "h_{j}" + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "text", + "content": "'s may belong to the same semantic class " + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "text", + "content": " (e.g., square), and can thus be represented by reusing the same entity function " + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "inline_equation", + "content": "f_{w}" + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "text", + "content": " with distinct embeddings " + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "inline_equation", + "content": "z_{j}" + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "text", + "content": "'s. Finally, each entity function also captures the precise spatial layout of the sub-entities by specifying their relative poses during the composition, such as 64 squares forming an " + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 54, + 519, + 295, + 662 + ], + "type": "text", + "content": " grid." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 662, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 295, + 699 + ], + "type": "text", + "content": "The following expands on how functions from " + }, + { + "bbox": [ + 55, + 662, + 295, + 699 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 662, + 295, + 699 + ], + "type": "text", + "content": " are defined and the program execution procedure to compute the represented scene " + }, + { + "bbox": [ + 55, + 662, + 295, + 699 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 662, + 295, + 699 + ], + "type": "text", + "content": ". Notations are summarized in Tab. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Entity Function Definitions. An entity function " + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{w}" + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": " outputs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": "an entity " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " of semantic class " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " takes two inputs: an embedding " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " that specifies " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": "'s identity, as well as an ordered set " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " containing embeddings of all descendant entities of " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": ". Denote " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\{h^{(i)}\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " sub-entities of " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " is a constant determined by " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": ". Then we have " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\gamma = \\{z^{(i)}, \\gamma^{(i)}\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "z^{(i)}" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\gamma^{(i)}" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " are the embeddings of " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "h^{(i)}" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " and of the descendant entities of " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "h^{(i)}" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": ", respectively. Thus, " + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "inline_equation", + "content": "f_w(z, \\gamma)" + }, + { + "bbox": [ + 313, + 175, + 555, + 277 + ], + "type": "text", + "content": " is recursively defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 280, + 553, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 316 + ], + "type": "interline_equation", + "content": "h = f _ {w} (z, \\gamma) := \\Psi_ {\\text {u n i o n}} \\left(\\left\\{\\Psi_ {\\text {t r a n s f o r m}} \\left(h ^ {(i)}, t ^ {(i)} (z)\\right) \\right\\} _ {i = 1} ^ {N}\\right), \\tag {2}", + "image_path": "9c52c75d1158ec89d6941a0b1eb3fe1559526e259f1276a7dff7178032964865.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 316, + 553, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 316, + 553, + 341 + ], + "type": "inline_equation", + "content": "h^{(i)} = f_{w^{(i)}}(z^{(i)},\\gamma^{(i)})" + }, + { + "bbox": [ + 313, + 316, + 553, + 341 + ], + "type": "text", + "content": " is a sub-entity of " + }, + { + "bbox": [ + 313, + 316, + 553, + 341 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 316, + 553, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 316, + 553, + 341 + ], + "type": "inline_equation", + "content": "t^{(i)}(z)" + }, + { + "bbox": [ + 313, + 316, + 553, + 341 + ], + "type": "text", + "content": " specifies its pose." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "f_{w}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " specifies the computation procedure to obtain output " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " sub-entities " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "\\{h^{(i)}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " via two operations: " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{transform}}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " applies an input-dependent pose " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "t^{(i)}(z)" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " to a subentity " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "h^{(i)}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": ", transforming it from its canonical frame to the world frame of " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{union}}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " composes multiple sub-entities into one single entity. Each sub-entity " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "h^{(i)}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " is computed by recursively applying an entity function " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "f_{w^{(i)}}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " also defined using Eq. (2). For instance, let " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "f_{w}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " denote the entity function that produces the board in Fig. 2 (namely, " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "w = \\mathrm{board}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": "). This function " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "f_{w}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " composes 64 sub-entities " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "\\{h^{(i)}\\}_{i = 1}^{64}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " of the same class square, which are in turn obtained by executing the same entity function " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "f_{w^{(i)}} = f_{\\mathrm{square}}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " with different embeddings " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "z^{(i)}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "inline_equation", + "content": "\\gamma^{(i)}" + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": "Program Execution. To obtain a scene " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": " from the Scene Language " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "\\Phi(s) = (W, P, Z)" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": ", a program executor identifies a root entity function " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "f_{w_1}" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": " that is not dependent by any other function (e.g., " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "w_1 =" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": " chessboard from Fig. 2), and evaluates this root function using the first embedding " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "z_1 \\in Z" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": " and the rest as its descendants, " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "\\gamma_1 := [z_2, \\dots, z_J] \\subset Z" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": ", to obtain " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "s = f_{w_1}(z_1, \\gamma_1)" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": ". Evaluating " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "f_{w_1}(z_1, \\gamma_1)" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": " expands the computation recursively to its children functions " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "h_j = f_{w_j}(z_j, \\gamma_j)" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": " as defined in Eq. (2), obtaining a full sequence of all the entities " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": " of the scene, where " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "j = 1, 2, \\dots, J, z_j \\in Z, w_j \\in W" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": ". The order in " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": " corresponds to the depth-first-search of the computation graph starting from " + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "inline_equation", + "content": "z_1" + }, + { + "bbox": [ + 313, + 498, + 556, + 668 + ], + "type": "text", + "content": ". An example computation graph is visualized on the right of Fig. 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 672, + 542, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 542, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 542, + 685 + ], + "type": "text", + "content": "3.2. Scene Language as Programming Language" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "We now concretize the definition in Sec. 3.1 with a domain-specific language (DSL) specified in Appendix B.1. We in-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24628" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "produce macro operations union for " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\Psi_{\\text{union}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": ", union-loop which calls union on entities evaluated in a for-loop, and transform for " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\Psi_{\\text{transform}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": ". Together with call which applies function calls, these four macro operations fully define entity functions in the DSL. Entity functions are identified with the associated words in the DSL via two special forms: bind, which binds entity function " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " to word " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": ", and retrieve, which retrieves " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " and applies " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " on actual embedding parameters. If " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " is never bound to a function, it corresponds to an entity function with no sub-entities (i.e., " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "N = 0" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " in Eq. (2)), in which case call returns a primitive entity with no children." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "text", + "content": "The data type of an entity " + }, + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "text", + "content": " (Eq. (2)) is denoted as Entity. It stores two data fields Word and Embedding, describing the entity's semantic group and identity respectively, and stores each children entity together with its pose in the frame of " + }, + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "text", + "content": ". In particular, Embedding captures the visual details and requires a highly expressive representation, e.g., neural embeddings. We employ the textual embedding space of OpenCLIP-ViT/H [13] for parameterization, denoted as " + }, + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "text", + "content": ". It offers the advantage that embeddings can be either encoded directly from natural language or inferred from images with Textual Inversion [6]. Tab. 1 summarizes the expressions and data types in accordance with the notations from Sec. 3.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 386, + 126, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 126, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 126, + 399 + ], + "type": "text", + "content": "4. Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 479 + ], + "type": "text", + "content": "Applying the proposed scene representation to image generation tasks requires rendering a Scene Language " + }, + { + "bbox": [ + 55, + 407, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\Phi(s)" + }, + { + "bbox": [ + 55, + 407, + 295, + 479 + ], + "type": "text", + "content": " into images. To do so, first, the program interpreter evaluates " + }, + { + "bbox": [ + 55, + 407, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\Phi(s)" + }, + { + "bbox": [ + 55, + 407, + 295, + 479 + ], + "type": "text", + "content": " to obtain a data object of type Entity. Afterward, a graphics renderer maps the Entity data object to its rendering parameter space and renders it into a final image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": "**Renderer Specifications.** We define the specification of a graphics renderer, a module in the proposed representation, as follows. A graphics renderer is determined by primitive parameter space " + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": " and a rendering operation " + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{R} : \\mathcal{P}(\\Theta \\times \\mathcal{T}) \\to \\mathcal{I}" + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": " is the space of 3D affine transformations representing poses, " + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": " denotes all possible subsets, and " + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": " is the space of rendered images. In order to determine a mapping from program execution outputs of type Entity (Fig. 3b) to the admissible input domain of rendering operation " + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": " (Fig. 3c), we assume access to a reparameterization function " + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "inline_equation", + "content": "g_{\\text{reparam}}" + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": " that maps from Tuple[Word, Embedding] to " + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": ", and compute a primitive's pose " + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "inline_equation", + "content": "t \\in \\mathcal{T}" + }, + { + "bbox": [ + 55, + 483, + 295, + 662 + ], + "type": "text", + "content": " by multiplying all Matrix values along the path from the primitive to the scene (root) in the entity hierarchy, analogous to computing limb poses in a kinematic tree." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "**Renderer Instantiations.** An example renderer instantiation is with Score Distillation Sampling (SDS) [25] guidance, where " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": " is a differentiable 3D representation, and we specify " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{reparam}}" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": " as follows. Recall that an entity is as-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "sociated with a Word value, e.g., moai for one statue from Figs. 1 and 3, and an Embedding value, e.g., " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ". For each primitive entity (i.e., an entity with no children), given these two value fields together with the fields of its ancestors, we use a manually specified language template " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ", or a " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " moai, in the style of " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ", 3D model in this example, to embed them into " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "z = g_{\\mathrm{CLIP}}(c) \\in \\mathcal{Z}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " is the pre-trained CLIP text encoder. Then, " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{reparam}}: \\mathcal{Z}_{\\mathrm{CLIP}} \\to \\Theta" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " corresponds to the SDS-guided optimization to find a solution in " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " that aligns with the input condition " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ". Output 3D scenes can be personalized by editing embeddings, e.g., " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " which controls the global style in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "text", + "content": "For the underlying 3D representation, we use 3D Gaussian Splatting [18] where images are rendered by splatting a set of 3D Gaussians onto the image plane; other differentiable 3D representations such as neural fields will also be suitable. We base our implementation on GALA3D [44], and use MVDream [31] and a depth-conditioned ControlNet [40] for guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 419 + ], + "type": "text", + "content": "We refer to the renderer above as the Gaussians renderer. Other possible renderers include primitive-based renderers, such as Mitsuba [14] with graphics primitives of cubes, spheres, and cylinders; asset-based game engines, such as MineCraft [23]; and feed-forward inference of layout-conditioned text-to-image (T2I) diffusion models, such as MIGC [42], which achieves 2D bounding box conditioning by controlling attention layers from Stable Diffusion [28]). Tab. 2 shows a summary; details are in Appendix B.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 553, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 553, + 443 + ], + "type": "text", + "content": "5. Inference via Pre-Trained Language Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "text", + "content": "We introduce a training-free method to infer the representation " + }, + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\Phi(s) = (W, P, Z)" + }, + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "text", + "content": " from text or image descriptions of a scene " + }, + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "text", + "content": ". As explained below, we first prompt a pre-trained language model (LM) to generate the non-neural components " + }, + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "inline_equation", + "content": "(W, P)" + }, + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "text", + "content": " and then obtain neural embeddings " + }, + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "inline_equation", + "content": "(Z)" + }, + { + "bbox": [ + 313, + 450, + 553, + 533 + ], + "type": "text", + "content": " from texts via the CLIP text encoder or from images with a pretrained text-to-image diffusion model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 642 + ], + "type": "text", + "content": "LMs have shown remarkable capability in code generation with common programming languages such as Python. In our implementation, we prompt LMs to generate Python scripts. We prompt the LM with the input condition, i.e., a scene description in texts or an image; a Python script of helper functions converted from the DSL in Sec. 3.2; and an example script using the helper functions. We use Claude 3.5 Sonnet [1] for all experiments for our method and LM-dependent baselines. Full LM prompts are in Appendix F.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "Function arguments in the LM-generated scripts, which are numeric values or string tokens, are converted to embeddings from " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " (Sec. 3.2) using language templates and the CLIP text encoder " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": ". For example, in the raw LM output, function calls for white pieces in Fig. 2 have input attribute {\"color\": (.9, .9, .9)}}, and we prompt LM" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24629" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 156, + 144 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 156, + 144 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 156, + 144 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 156, + 144 + ], + "type": "image", + "image_path": "e27eb4cab15f145112aa840512c15b6c229eddc8217d5dc5227acbb3467ede56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 148, + 146, + 158 + ], + "lines": [ + { + "bbox": [ + 78, + 148, + 146, + 158 + ], + "spans": [ + { + "bbox": [ + 78, + 148, + 146, + 158 + ], + "type": "text", + "content": "Ours (Two Views)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 76, + 212, + 148 + ], + "blocks": [ + { + "bbox": [ + 170, + 76, + 212, + 148 + ], + "lines": [ + { + "bbox": [ + 170, + 76, + 212, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 76, + 212, + 148 + ], + "type": "image", + "image_path": "4af557c8345bc77887bca7c4a53291a0ea45a466639a8173dc63102e93cd1cf8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 148, + 211, + 158 + ], + "lines": [ + { + "bbox": [ + 169, + 148, + 211, + 158 + ], + "spans": [ + { + "bbox": [ + 169, + 148, + 211, + 158 + ], + "type": "text", + "content": "MVDream" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 87, + 280, + 147 + ], + "blocks": [ + { + "bbox": [ + 220, + 87, + 280, + 147 + ], + "lines": [ + { + "bbox": [ + 220, + 87, + 280, + 147 + ], + "spans": [ + { + "bbox": [ + 220, + 87, + 280, + 147 + ], + "type": "image", + "image_path": "08e1d50ea05369dce00e693a6569a3526b2c2d929149ba051e5a307926eba983.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 148, + 281, + 158 + ], + "lines": [ + { + "bbox": [ + 224, + 148, + 281, + 158 + ], + "spans": [ + { + "bbox": [ + 224, + 148, + 281, + 158 + ], + "type": "text", + "content": "GraphDreamer" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 70, + 169, + 104, + 237 + ], + "blocks": [ + { + "bbox": [ + 70, + 169, + 104, + 237 + ], + "lines": [ + { + "bbox": [ + 70, + 169, + 104, + 237 + ], + "spans": [ + { + "bbox": [ + 70, + 169, + 104, + 237 + ], + "type": "image", + "image_path": "69df59f0bd149f44a98e6ba74d3c0f69ebaabe3c8b10358f48bba40aaab9639c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 244, + 146, + 255 + ], + "lines": [ + { + "bbox": [ + 78, + 244, + 146, + 255 + ], + "spans": [ + { + "bbox": [ + 78, + 244, + 146, + 255 + ], + "type": "text", + "content": "Ours (Two Views)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 118, + 169, + 149, + 240 + ], + "blocks": [ + { + "bbox": [ + 138, + 159, + 208, + 171 + ], + "lines": [ + { + "bbox": [ + 138, + 159, + 208, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 208, + 171 + ], + "type": "text", + "content": "\"An 8-layer shelf\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 118, + 169, + 149, + 240 + ], + "lines": [ + { + "bbox": [ + 118, + 169, + 149, + 240 + ], + "spans": [ + { + "bbox": [ + 118, + 169, + 149, + 240 + ], + "type": "image", + "image_path": "1e447ef0eb88c727a2fb6beb46bff05b1108c8cc4607f1361eb8b443f9beca79.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 162, + 178, + 218, + 242 + ], + "blocks": [ + { + "bbox": [ + 162, + 178, + 218, + 242 + ], + "lines": [ + { + "bbox": [ + 162, + 178, + 218, + 242 + ], + "spans": [ + { + "bbox": [ + 162, + 178, + 218, + 242 + ], + "type": "image", + "image_path": "17dcbc9225b699fde1748a2d9cac5a1e34e40a64b2abff6dcd6d9f63f2b7717a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 243, + 210, + 254 + ], + "lines": [ + { + "bbox": [ + 169, + 243, + 210, + 254 + ], + "spans": [ + { + "bbox": [ + 169, + 243, + 210, + 254 + ], + "type": "text", + "content": "MVDream" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 223, + 177, + 282, + 243 + ], + "blocks": [ + { + "bbox": [ + 223, + 177, + 282, + 243 + ], + "lines": [ + { + "bbox": [ + 223, + 177, + 282, + 243 + ], + "spans": [ + { + "bbox": [ + 223, + 177, + 282, + 243 + ], + "type": "image", + "image_path": "d961f9ae50dd92a01ebdaec81324b0ea69724b0c7e6d431a0f12720b69cb1814.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 244, + 280, + 255 + ], + "lines": [ + { + "bbox": [ + 225, + 244, + 280, + 255 + ], + "spans": [ + { + "bbox": [ + 225, + 244, + 280, + 255 + ], + "type": "text", + "content": "GraphDreamer" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 256, + 249, + 267 + ], + "lines": [ + { + "bbox": [ + 111, + 256, + 249, + 267 + ], + "spans": [ + { + "bbox": [ + 111, + 256, + 249, + 267 + ], + "type": "text", + "content": "\"15 coke cans stacking in a pyramid\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 269, + 295, + 302 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 295, + 302 + ], + "type": "text", + "content": "Figure 4. Text-Prompted Scene Generation. Ours produce more accurate outputs than using no intermediate representation (MV-Dream) or scene graph (GraphDreamer)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 56, + 309, + 294, + 357 + ], + "blocks": [ + { + "bbox": [ + 56, + 309, + 294, + 357 + ], + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 357 + ], + "type": "table", + "html": "
MethodsNumeric PromptsGeneric Prompts
Alignment (↑)CLIP (↑)Count (↑)Alignment (↑)CLIP (↑)
GraphDreamer [7]3.56±7.380.297±0.0850.1115.75±20.390.230±0.097
MVDream [31]10.79±12.830.312±0.0460.1135.13±31.900.328±0.031
Ours85.65±13.710.351±0.0511.052.13±26.830.328±0.030
", + "image_path": "53990c41bb795ad949f8eb246e2bc67b8311d6bc6b8311663585dca80642a19d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "text", + "content": "to describe the color value as a word, and feed the word into " + }, + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "text", + "content": " to compute " + }, + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "inline_equation", + "content": "< z68 >" + }, + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "text", + "content": ". For image-conditioned tasks, for each primitive entity in the execution output of " + }, + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "text", + "content": ", we first use GroundingSAM [20, 27] to segment out the region defined by the word associated with the entity. We then use Textual Inversion [6] to optimize an embedding to reconstruct the cropped image with the diffusion model training objective. The full process is deferred to Appendix C.2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 542, + 137, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 137, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 137, + 555 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 561, + 263, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 263, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 263, + 574 + ], + "type": "text", + "content": "6.1. Text-Prompted Generation and Editing" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Baselines. To evaluate our representation in text-prompted 3D generation task, we compare our inference pipeline with 3D scene generation methods using alternative intermediate representations, e.g., scene graph. In particular, we compare with GraphDreamer [7] as an exemplar approach, which generates scene graphs from texts via LM prompting and then synthesizes scenes conditioned on the graphs via SDS guidance. We further ablate the role of structural representation in this task by comparing ours with the backbone of our SDS-based renderer, MVDream [31], as a direct scene generation approach. Details are in Appendix C.3." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 332, + 76, + 386, + 99 + ], + "blocks": [ + { + "bbox": [ + 320, + 76, + 331, + 139 + ], + "lines": [ + { + "bbox": [ + 320, + 76, + 331, + 139 + ], + "spans": [ + { + "bbox": [ + 320, + 76, + 331, + 139 + ], + "type": "text", + "content": "rnnnne nnnnne" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 76, + 386, + 99 + ], + "lines": [ + { + "bbox": [ + 332, + 76, + 386, + 99 + ], + "spans": [ + { + "bbox": [ + 332, + 76, + 386, + 99 + ], + "type": "image", + "image_path": "f3a0dbd181d380a934c8da1ddd1c14df554101b4cc75b27b8bf985ee8c1656ad.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 332, + 105, + 386, + 147 + ], + "blocks": [ + { + "bbox": [ + 332, + 105, + 386, + 147 + ], + "lines": [ + { + "bbox": [ + 332, + 105, + 386, + 147 + ], + "spans": [ + { + "bbox": [ + 332, + 105, + 386, + 147 + ], + "type": "image", + "image_path": "0cad2d9bdb96aea4db1e662d978c3cbade744f8ba1811b90ee12e0c9a922329f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 144, + 379, + 152 + ], + "lines": [ + { + "bbox": [ + 343, + 144, + 379, + 152 + ], + "spans": [ + { + "bbox": [ + 343, + 144, + 379, + 152 + ], + "type": "text", + "content": "Gaussians" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 392, + 76, + 441, + 99 + ], + "blocks": [ + { + "bbox": [ + 392, + 76, + 441, + 99 + ], + "lines": [ + { + "bbox": [ + 392, + 76, + 441, + 99 + ], + "spans": [ + { + "bbox": [ + 392, + 76, + 441, + 99 + ], + "type": "image", + "image_path": "bf75a7d2b2bd68f2e3c7bf77b9f3991a652d194a65c58e9514a1f17cd5a83443.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 392, + 103, + 441, + 148 + ], + "blocks": [ + { + "bbox": [ + 392, + 103, + 441, + 148 + ], + "lines": [ + { + "bbox": [ + 392, + 103, + 441, + 148 + ], + "spans": [ + { + "bbox": [ + 392, + 103, + 441, + 148 + ], + "type": "image", + "image_path": "62f64907d5772a0d50b7faa55a7591ff8d4868ceb72f11c899de8cb273cf4cbd.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 144, + 434, + 152 + ], + "lines": [ + { + "bbox": [ + 397, + 144, + 434, + 152 + ], + "spans": [ + { + "bbox": [ + 397, + 144, + 434, + 152 + ], + "type": "text", + "content": "Minecraft" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 396, + 153, + 425, + 162 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 153, + 425, + 162 + ], + "spans": [ + { + "bbox": [ + 396, + 153, + 425, + 162 + ], + "type": "text", + "content": "model of" + } + ] + } + ], + "index": 30, + "type": "text" + }, + { + "bbox": [ + 359, + 163, + 413, + 172 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 163, + 413, + 172 + ], + "spans": [ + { + "bbox": [ + 359, + 163, + 413, + 172 + ], + "type": "text", + "content": "a tennis court\"" + } + ] + } + ], + "index": 31, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 334, + 177, + 385, + 215 + ], + "blocks": [ + { + "bbox": [ + 320, + 176, + 331, + 202 + ], + "lines": [ + { + "bbox": [ + 320, + 176, + 331, + 202 + ], + "spans": [ + { + "bbox": [ + 320, + 176, + 331, + 202 + ], + "type": "text", + "content": "nne nnnnne" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 177, + 385, + 215 + ], + "lines": [ + { + "bbox": [ + 334, + 177, + 385, + 215 + ], + "spans": [ + { + "bbox": [ + 334, + 177, + 385, + 215 + ], + "type": "image", + "image_path": "53a0ac878901b617508688e452da050daec5f8b222c8d07ac2909851d6e6ee60.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 217, + 374, + 225 + ], + "lines": [ + { + "bbox": [ + 343, + 217, + 374, + 225 + ], + "spans": [ + { + "bbox": [ + 343, + 217, + 374, + 225 + ], + "type": "text", + "content": "Mitsuba" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 227, + 432, + 237 + ], + "lines": [ + { + "bbox": [ + 345, + 227, + 432, + 237 + ], + "spans": [ + { + "bbox": [ + 345, + 227, + 432, + 237 + ], + "type": "text", + "content": "\"City skyline at sunset\"" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 240, + 554, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 240, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 554, + 285 + ], + "type": "text", + "content": "Figure 5. Renderings Across Graphics Renderers. Different renderers produce renderings that adhere to the same representation and are therefore visually aligned, while each exhibits a different imaging style. Text inputs are shown at the subfigure bottoms." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 391, + 176, + 442, + 215 + ], + "blocks": [ + { + "bbox": [ + 347, + 153, + 425, + 173 + ], + "lines": [ + { + "bbox": [ + 347, + 153, + 425, + 173 + ], + "spans": [ + { + "bbox": [ + 347, + 153, + 425, + 173 + ], + "type": "text", + "content": "\"A detailed model of a tennis court\"" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 176, + 442, + 215 + ], + "lines": [ + { + "bbox": [ + 391, + 176, + 442, + 215 + ], + "spans": [ + { + "bbox": [ + 391, + 176, + 442, + 215 + ], + "type": "image", + "image_path": "1ebaba16252602202b24c74d24aaf78a0e00d640fb58f4d0c1c45f2c05a7ed21.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 217, + 423, + 225 + ], + "lines": [ + { + "bbox": [ + 408, + 217, + 423, + 225 + ], + "spans": [ + { + "bbox": [ + 408, + 217, + 423, + 225 + ], + "type": "text", + "content": "T2I" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 444, + 73, + 497, + 104 + ], + "blocks": [ + { + "bbox": [ + 444, + 73, + 497, + 104 + ], + "lines": [ + { + "bbox": [ + 444, + 73, + 497, + 104 + ], + "spans": [ + { + "bbox": [ + 444, + 73, + 497, + 104 + ], + "type": "image", + "image_path": "e13064e231a80ebe12101c498c7a1a270393d4a64891166b836853c08b57ea2a.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 447, + 106, + 497, + 142 + ], + "blocks": [ + { + "bbox": [ + 447, + 106, + 497, + 142 + ], + "lines": [ + { + "bbox": [ + 447, + 106, + 497, + 142 + ], + "spans": [ + { + "bbox": [ + 447, + 106, + 497, + 142 + ], + "type": "image", + "image_path": "1f37b8c6405b561fb401bb003fd77b7d99ceaf19c55195633f0008a2eedfd28b.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 144, + 490, + 152 + ], + "lines": [ + { + "bbox": [ + 452, + 144, + 490, + 152 + ], + "spans": [ + { + "bbox": [ + 452, + 144, + 490, + 152 + ], + "type": "text", + "content": "Gaussians" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 503, + 72, + 551, + 105 + ], + "blocks": [ + { + "bbox": [ + 503, + 72, + 551, + 105 + ], + "lines": [ + { + "bbox": [ + 503, + 72, + 551, + 105 + ], + "spans": [ + { + "bbox": [ + 503, + 72, + 551, + 105 + ], + "type": "image", + "image_path": "171ca8f7d2774261606cbe2fce8a34b241888d6ff61285a48ed75b01c7331def.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 503, + 106, + 548, + 142 + ], + "blocks": [ + { + "bbox": [ + 503, + 106, + 548, + 142 + ], + "lines": [ + { + "bbox": [ + 503, + 106, + 548, + 142 + ], + "spans": [ + { + "bbox": [ + 503, + 106, + 548, + 142 + ], + "type": "image", + "image_path": "d7d94cd6445d432d89f038748b81666e3686adcc049597e1f597bac16faab3ec.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 144, + 545, + 152 + ], + "lines": [ + { + "bbox": [ + 509, + 144, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 509, + 144, + 545, + 152 + ], + "type": "text", + "content": "Minecraft" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "bbox": [ + 453, + 144, + 490, + 152 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 453, + 144, + 490, + 152 + ], + "spans": [ + { + "bbox": [ + 453, + 144, + 490, + 152 + ], + "type": "text", + "content": "Gaussians" + } + ] + } + ], + "index": 43, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 452, + 177, + 476, + 212 + ], + "blocks": [ + { + "bbox": [ + 455, + 153, + 542, + 174 + ], + "lines": [ + { + "bbox": [ + 455, + 153, + 542, + 174 + ], + "spans": [ + { + "bbox": [ + 455, + 153, + 542, + 174 + ], + "type": "text", + "content": "\"An interior scene of a university lecture hall\"" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 177, + 476, + 212 + ], + "lines": [ + { + "bbox": [ + 452, + 177, + 476, + 212 + ], + "spans": [ + { + "bbox": [ + 452, + 177, + 476, + 212 + ], + "type": "image", + "image_path": "20989f2745383b633cb4f968f4a060e28ba844e04dc620d60fb36de9eeec3f4f.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 217, + 486, + 225 + ], + "lines": [ + { + "bbox": [ + 455, + 217, + 486, + 225 + ], + "spans": [ + { + "bbox": [ + 455, + 217, + 486, + 225 + ], + "type": "text", + "content": "Mitsuba" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 500, + 177, + 550, + 213 + ], + "blocks": [ + { + "bbox": [ + 500, + 177, + 550, + 213 + ], + "lines": [ + { + "bbox": [ + 500, + 177, + 550, + 213 + ], + "spans": [ + { + "bbox": [ + 500, + 177, + 550, + 213 + ], + "type": "image", + "image_path": "df7a2c73d654451176e8738b0c791f7d0eb1161195b6b7dd4339c600201e23c1.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 517, + 217, + 531, + 225 + ], + "lines": [ + { + "bbox": [ + 517, + 217, + 531, + 225 + ], + "spans": [ + { + "bbox": [ + 517, + 217, + 531, + 225 + ], + "type": "text", + "content": "T2I" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 227, + 523, + 236 + ], + "lines": [ + { + "bbox": [ + 473, + 227, + 523, + 236 + ], + "spans": [ + { + "bbox": [ + 473, + 227, + 523, + 236 + ], + "type": "text", + "content": "\"Marionette\"" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "table", + "bbox": [ + 316, + 294, + 446, + 331 + ], + "blocks": [ + { + "bbox": [ + 55, + 359, + 296, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 425 + ], + "type": "text", + "content": "Table 3. Evaluation for text-prompted 3D generation. We report the following metrics averaged across 9 numeric and 8 generic prompts: the percentages of user preferences for prompt alignment, CLIP similarity scores, and counting accuracy (numeric prompts only, 0 for inaccurate and 1 for accurate). Here " + }, + { + "bbox": [ + 55, + 359, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 55, + 359, + 296, + 425 + ], + "type": "text", + "content": " denotes standard deviation across scenes." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 294, + 446, + 331 + ], + "lines": [ + { + "bbox": [ + 316, + 294, + 446, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 446, + 331 + ], + "type": "table", + "html": "
MethodsCLIP [13] (↑)Dynamic [12] (↑)
4D-fy [2]0.3520.2%
Ours0.3415.9%
", + "image_path": "8ccc19108a6f0016e798592103d08467d02044b1345d1546aadae33fb07a11c9.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "table_body" + } + ], + "index": 52 + }, + { + "type": "table", + "bbox": [ + 449, + 295, + 553, + 331 + ], + "blocks": [ + { + "bbox": [ + 449, + 295, + 553, + 331 + ], + "lines": [ + { + "bbox": [ + 449, + 295, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 449, + 295, + 553, + 331 + ], + "type": "table", + "html": "
MethodsLPIPS [41] (↓)
GraphDreamer [7]0.811
Ours0.681
", + "image_path": "d948eaa94260b88cb357dac0e745e1750781dd2eee9a3fcbe13b5e44c252c759.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "table_body" + } + ], + "index": 53 + }, + { + "bbox": [ + 313, + 334, + 554, + 389 + ], + "lines": [ + { + "bbox": [ + 313, + 334, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 554, + 389 + ], + "type": "text", + "content": "Table 4. Evaluation. Left: For text-prompted 4D generation, 4Dfy achieves higher prompt alignment measured with CLIP similarity but has small motion. Right: For image-prompted scene generation, our results are better aligned with input image prompts compared to GraphDreamer." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 398, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 554, + 506 + ], + "type": "text", + "content": "Metrics. We conduct a user study where 103 users are asked to choose one of the three videos, showing 360-renderings from ours and two baselines in a randomized order, that aligns the best with text prompts. Following [7], for each scene, we compute the cosine similarity of CLIP [13, 26] embeddings between a rendering and the text prompt and report the average over 120 viewpoints. We further report whether outputs have the correct object counts. Details are deferred to Appendix C.4." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 313, + 509, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 557 + ], + "type": "text", + "content": "Results. Tab. 3 shows that our method achieves a more favorable prompt alignment than the baselines as measured in the user study and CLIP similarity, and has a clear advantage in counting accuracy." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Fig. 4 and Appendix D.1 show qualitative results of ours using the SDS-based renderer. Compared to the direct 3D scene generation method MVDream, our approach is compositional and adheres more closely to input prompts in scenes involving multiple objects. Compared to a scene graph representation, where entity relations are restricted to be between two objects and are bottlenecked by the coarseness of natural language descriptions, e.g., \"aligned in a row\", a program-based representation offers more flexible and precise specifications for relations, e.g., the particular coke can arrangement in Fig. 4. This brings the practical benefit of offloading the burden of generating scenes involving complex entity relations from the T2I model (used" + } + ] + } + ], + "index": 57 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24630" + } + ] + } + ], + "index": 58 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 75, + 127, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 75, + 127, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 75, + 127, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 75, + 127, + 148 + ], + "type": "image", + "image_path": "f49a812c2ecb43a64bde2247baa32f8b5545f3592c0b9a38d8cb63747a3176b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 151, + 116, + 174 + ], + "lines": [ + { + "bbox": [ + 63, + 151, + 116, + 174 + ], + "spans": [ + { + "bbox": [ + 63, + 151, + 116, + 174 + ], + "type": "text", + "content": "Ours (w/ tracking)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 179, + 212, + 191 + ], + "lines": [ + { + "bbox": [ + 88, + 179, + 212, + 191 + ], + "spans": [ + { + "bbox": [ + 88, + 179, + 212, + 191 + ], + "type": "text", + "content": "\"Carousel with a small canopy\"" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 128, + 75, + 186, + 144 + ], + "blocks": [ + { + "bbox": [ + 128, + 75, + 186, + 144 + ], + "lines": [ + { + "bbox": [ + 128, + 75, + 186, + 144 + ], + "spans": [ + { + "bbox": [ + 128, + 75, + 186, + 144 + ], + "type": "image", + "image_path": "c20f40760c1fbe7feef3c9407018a4359bf2a91d7bc20f6e1f781eb8c37e55bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 145, + 177, + 154 + ], + "lines": [ + { + "bbox": [ + 146, + 145, + 177, + 154 + ], + "spans": [ + { + "bbox": [ + 146, + 145, + 177, + 154 + ], + "type": "text", + "content": "Frame 1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 179, + 385, + 190 + ], + "lines": [ + { + "bbox": [ + 299, + 179, + 385, + 190 + ], + "spans": [ + { + "bbox": [ + 299, + 179, + 385, + 190 + ], + "type": "text", + "content": "\"Solar system model\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 187, + 75, + 247, + 145 + ], + "blocks": [ + { + "bbox": [ + 187, + 75, + 247, + 145 + ], + "lines": [ + { + "bbox": [ + 187, + 75, + 247, + 145 + ], + "spans": [ + { + "bbox": [ + 187, + 75, + 247, + 145 + ], + "type": "image", + "image_path": "3c350ffa2e0ebbfebe499673b10061231a397dc4e190c8adfd9282befa846aa5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 146, + 218, + 154 + ], + "lines": [ + { + "bbox": [ + 176, + 146, + 218, + 154 + ], + "spans": [ + { + "bbox": [ + 176, + 146, + 218, + 154 + ], + "type": "text", + "content": "1 Fran" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 255, + 94, + 332, + 125 + ], + "blocks": [ + { + "bbox": [ + 255, + 94, + 332, + 125 + ], + "lines": [ + { + "bbox": [ + 255, + 94, + 332, + 125 + ], + "spans": [ + { + "bbox": [ + 255, + 94, + 332, + 125 + ], + "type": "image", + "image_path": "afbf2b027051b852cb9ba9f63bbaff50eb9545b7f1cd4a85d1b31e92560f88d6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 335, + 100, + 422, + 121 + ], + "blocks": [ + { + "bbox": [ + 266, + 151, + 319, + 175 + ], + "lines": [ + { + "bbox": [ + 266, + 151, + 319, + 175 + ], + "spans": [ + { + "bbox": [ + 266, + 151, + 319, + 175 + ], + "type": "text", + "content": "Ours (w/ tracking)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 100, + 422, + 121 + ], + "lines": [ + { + "bbox": [ + 335, + 100, + 422, + 121 + ], + "spans": [ + { + "bbox": [ + 335, + 100, + 422, + 121 + ], + "type": "image", + "image_path": "cc8404f3abf0418a4da2d4220b055787b6237181129b3ed92ce364883bfeeac1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 156, + 391, + 168 + ], + "lines": [ + { + "bbox": [ + 366, + 156, + 391, + 168 + ], + "spans": [ + { + "bbox": [ + 366, + 156, + 391, + 168 + ], + "type": "text", + "content": "4D-fy" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 179, + 526, + 191 + ], + "lines": [ + { + "bbox": [ + 443, + 179, + 526, + 191 + ], + "spans": [ + { + "bbox": [ + 443, + 179, + 526, + 191 + ], + "type": "text", + "content": "\"A toy wind turbine\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 425, + 76, + 545, + 144 + ], + "blocks": [ + { + "bbox": [ + 342, + 144, + 419, + 155 + ], + "lines": [ + { + "bbox": [ + 342, + 144, + 419, + 155 + ], + "spans": [ + { + "bbox": [ + 342, + 144, + 419, + 155 + ], + "type": "text", + "content": "Frame 1 Frame 2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 76, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 425, + 76, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 425, + 76, + 545, + 144 + ], + "type": "image", + "image_path": "7910274de014f94814f54f0f04fc564cb16421ef97fb1ef2c9f05e13c522cc53.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 145, + 546, + 177 + ], + "lines": [ + { + "bbox": [ + 427, + 145, + 546, + 177 + ], + "spans": [ + { + "bbox": [ + 427, + 145, + 546, + 177 + ], + "type": "text", + "content": "Ours Frame 1 Frame 2 (w/ tracking) 4D-fy" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 71, + 234, + 137, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 194, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 228 + ], + "type": "text", + "content": "Figure 6. Text-Prompted 4D Scene Generation. The proposed representation may also capture scene dynamics and be applied for synthesizing 4D scenes. It explicitly represents the temporal correspondence of an entity in a dynamic scene. Each colored trajectory denotes tracking of a temporally moving point. Compared to 4D-fy, our method produces more salient motion." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 234, + 137, + 327 + ], + "lines": [ + { + "bbox": [ + 71, + 234, + 137, + 327 + ], + "spans": [ + { + "bbox": [ + 71, + 234, + 137, + 327 + ], + "type": "image", + "image_path": "a787646a9aaaa304bb65bf72cbbe71162ed1e6b6c87319fa4b6843b27344f8e8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 329, + 137, + 340 + ], + "lines": [ + { + "bbox": [ + 69, + 329, + 137, + 340 + ], + "spans": [ + { + "bbox": [ + 69, + 329, + 137, + 340 + ], + "type": "text", + "content": "Initial (Mitsuba)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 342, + 156, + 354 + ], + "lines": [ + { + "bbox": [ + 70, + 342, + 156, + 354 + ], + "spans": [ + { + "bbox": [ + 70, + 342, + 156, + 354 + ], + "type": "text", + "content": "\"Bramante Staircase," + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 170, + 235, + 231, + 323 + ], + "blocks": [ + { + "bbox": [ + 170, + 235, + 231, + 323 + ], + "lines": [ + { + "bbox": [ + 170, + 235, + 231, + 323 + ], + "spans": [ + { + "bbox": [ + 170, + 235, + 231, + 323 + ], + "type": "image", + "image_path": "e469e9debb3146c85c5714ef38ed26a94b415c7ae091a9902459f73936ba649c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 329, + 237, + 340 + ], + "lines": [ + { + "bbox": [ + 162, + 329, + 237, + 340 + ], + "spans": [ + { + "bbox": [ + 162, + 329, + 237, + 340 + ], + "type": "text", + "content": "Initial (Gaussians)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 72, + 361, + 138, + 446 + ], + "blocks": [ + { + "bbox": [ + 72, + 361, + 138, + 446 + ], + "lines": [ + { + "bbox": [ + 72, + 361, + 138, + 446 + ], + "spans": [ + { + "bbox": [ + 72, + 361, + 138, + 446 + ], + "type": "image", + "image_path": "420fa821f36b296c62832668367680129116b5b873c2fbc42610d2023388549c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 447, + 137, + 459 + ], + "lines": [ + { + "bbox": [ + 70, + 447, + 137, + 459 + ], + "spans": [ + { + "bbox": [ + 70, + 447, + 137, + 459 + ], + "type": "text", + "content": "Initial (Mitsuba)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 460, + 240, + 472 + ], + "lines": [ + { + "bbox": [ + 63, + 460, + 240, + 472 + ], + "spans": [ + { + "bbox": [ + 63, + 460, + 240, + 472 + ], + "type": "text", + "content": "\"An 8-layer 3-color Jenga set at game start\"" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 171, + 361, + 235, + 446 + ], + "blocks": [ + { + "bbox": [ + 171, + 361, + 235, + 446 + ], + "lines": [ + { + "bbox": [ + 171, + 361, + 235, + 446 + ], + "spans": [ + { + "bbox": [ + 171, + 361, + 235, + 446 + ], + "type": "image", + "image_path": "7494e9cb6314093d089f31e4001fdb8dbc6729ece29f9a9f093764aa84fa2726.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 447, + 238, + 459 + ], + "lines": [ + { + "bbox": [ + 162, + 447, + 238, + 459 + ], + "spans": [ + { + "bbox": [ + 162, + 447, + 238, + 459 + ], + "type": "text", + "content": "Initial (Gaussians)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 260, + 460, + 546, + 472 + ], + "lines": [ + { + "bbox": [ + 260, + 460, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 260, + 460, + 546, + 472 + ], + "type": "text", + "content": "\"Remove 2 blocks from second top layer and the tower should not fall\"" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 65, + 479, + 144, + 551 + ], + "blocks": [ + { + "bbox": [ + 65, + 479, + 144, + 551 + ], + "lines": [ + { + "bbox": [ + 65, + 479, + 144, + 551 + ], + "spans": [ + { + "bbox": [ + 65, + 479, + 144, + 551 + ], + "type": "image", + "image_path": "a12ecbca67a42acbab6150493ac5546c4f5a492eaaff780007d09fc1b3a17559.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 566, + 137, + 578 + ], + "lines": [ + { + "bbox": [ + 70, + 566, + 137, + 578 + ], + "spans": [ + { + "bbox": [ + 70, + 566, + 137, + 578 + ], + "type": "text", + "content": "Initial (Mitsuba)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 579, + 211, + 590 + ], + "lines": [ + { + "bbox": [ + 93, + 579, + 211, + 590 + ], + "spans": [ + { + "bbox": [ + 93, + 579, + 211, + 590 + ], + "type": "text", + "content": "\"A fractal tree-like structure\"" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 162, + 479, + 240, + 550 + ], + "blocks": [ + { + "bbox": [ + 162, + 479, + 240, + 550 + ], + "lines": [ + { + "bbox": [ + 162, + 479, + 240, + 550 + ], + "spans": [ + { + "bbox": [ + 162, + 479, + 240, + 550 + ], + "type": "image", + "image_path": "3233401808ce1fe1083368d3b2ed9e4b3fedcf21a68455a3f6f3b3204703799e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 566, + 238, + 578 + ], + "lines": [ + { + "bbox": [ + 162, + 566, + 238, + 578 + ], + "spans": [ + { + "bbox": [ + 162, + 566, + 238, + 578 + ], + "type": "text", + "content": "Initial (Gaussians)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 257, + 235, + 358, + 327 + ], + "blocks": [ + { + "bbox": [ + 257, + 235, + 358, + 327 + ], + "lines": [ + { + "bbox": [ + 257, + 235, + 358, + 327 + ], + "spans": [ + { + "bbox": [ + 257, + 235, + 358, + 327 + ], + "type": "image", + "image_path": "610909dad8a134a295bd73cb6e62fd5d10926697104846242a3b77f7868e23a2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 330, + 341, + 341 + ], + "lines": [ + { + "bbox": [ + 273, + 330, + 341, + 341 + ], + "spans": [ + { + "bbox": [ + 273, + 330, + 341, + 341 + ], + "type": "text", + "content": "Scene Language" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 336, + 342, + 470, + 354 + ], + "lines": [ + { + "bbox": [ + 336, + 342, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 336, + 342, + 470, + 354 + ], + "type": "text", + "content": "\"Shrink staircase radius by " + }, + { + "bbox": [ + 336, + 342, + 470, + 354 + ], + "type": "inline_equation", + "content": "80\\%" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 396, + 234, + 417, + 324 + ], + "blocks": [ + { + "bbox": [ + 396, + 234, + 417, + 324 + ], + "lines": [ + { + "bbox": [ + 396, + 234, + 417, + 324 + ], + "spans": [ + { + "bbox": [ + 396, + 234, + 417, + 324 + ], + "type": "image", + "image_path": "cfff393f4519e1ea777c9af8969be3cc683240a5bccea7b6b28a8b59e8dead52.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 329, + 441, + 340 + ], + "lines": [ + { + "bbox": [ + 372, + 329, + 441, + 340 + ], + "spans": [ + { + "bbox": [ + 372, + 329, + 441, + 340 + ], + "type": "text", + "content": "Edited (Mitsuba)" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 492, + 234, + 514, + 323 + ], + "blocks": [ + { + "bbox": [ + 492, + 234, + 514, + 323 + ], + "lines": [ + { + "bbox": [ + 492, + 234, + 514, + 323 + ], + "spans": [ + { + "bbox": [ + 492, + 234, + 514, + 323 + ], + "type": "image", + "image_path": "8306f9060792f238a687a5bc153c62003f76daaeec9fb37322978af119114bd6.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 329, + 542, + 339 + ], + "lines": [ + { + "bbox": [ + 465, + 329, + 542, + 339 + ], + "spans": [ + { + "bbox": [ + 465, + 329, + 542, + 339 + ], + "type": "text", + "content": "Edited (Gaussians)" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 260, + 360, + 354, + 401 + ], + "blocks": [ + { + "bbox": [ + 260, + 360, + 354, + 401 + ], + "lines": [ + { + "bbox": [ + 260, + 360, + 354, + 401 + ], + "spans": [ + { + "bbox": [ + 260, + 360, + 354, + 401 + ], + "type": "image", + "image_path": "43d44208cd4c2899eaf287343dddd7127d1e00006e4e749859b7d03246b7bbbd.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 404, + 348, + 445 + ], + "lines": [ + { + "bbox": [ + 263, + 404, + 348, + 445 + ], + "spans": [ + { + "bbox": [ + 263, + 404, + 348, + 445 + ], + "type": "text", + "content": "(\"Jenga layer\" (lambda(is-horizontal colors remove-blocks)))" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 376, + 361, + 442, + 446 + ], + "blocks": [ + { + "bbox": [ + 273, + 448, + 341, + 460 + ], + "lines": [ + { + "bbox": [ + 273, + 448, + 341, + 460 + ], + "spans": [ + { + "bbox": [ + 273, + 448, + 341, + 460 + ], + "type": "text", + "content": "Scene Language" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 361, + 442, + 446 + ], + "lines": [ + { + "bbox": [ + 376, + 361, + 442, + 446 + ], + "spans": [ + { + "bbox": [ + 376, + 361, + 442, + 446 + ], + "type": "image", + "image_path": "15e6717a6f34304ee89a4456267eb9d2183c7c252d2df8941f6474e8490f6c73.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 448, + 441, + 459 + ], + "lines": [ + { + "bbox": [ + 372, + 448, + 441, + 459 + ], + "spans": [ + { + "bbox": [ + 372, + 448, + 441, + 459 + ], + "type": "text", + "content": "Edited (Mitsuba)" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 474, + 361, + 538, + 446 + ], + "blocks": [ + { + "bbox": [ + 474, + 361, + 538, + 446 + ], + "lines": [ + { + "bbox": [ + 474, + 361, + 538, + 446 + ], + "spans": [ + { + "bbox": [ + 474, + 361, + 538, + 446 + ], + "type": "image", + "image_path": "577780964a6c7f543880c084b480fe9d1fb2a0207168620598a031e91d056514.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 448, + 542, + 459 + ], + "lines": [ + { + "bbox": [ + 465, + 448, + 542, + 459 + ], + "spans": [ + { + "bbox": [ + 465, + 448, + 542, + 459 + ], + "type": "text", + "content": "Edited (Gaussians)" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 263, + 509, + 332, + 563 + ], + "blocks": [ + { + "bbox": [ + 262, + 476, + 332, + 506 + ], + "lines": [ + { + "bbox": [ + 262, + 476, + 332, + 506 + ], + "spans": [ + { + "bbox": [ + 262, + 476, + 332, + 506 + ], + "type": "text", + "content": "union left-branch right-branch)" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 509, + 332, + 563 + ], + "lines": [ + { + "bbox": [ + 263, + 509, + 332, + 563 + ], + "spans": [ + { + "bbox": [ + 263, + 509, + 332, + 563 + ], + "type": "image", + "image_path": "d9ec9d0a675e0861d0dc26ec7198490cacfb05b2712be6cebaed6a5d9be83a5f.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 567, + 341, + 578 + ], + "lines": [ + { + "bbox": [ + 273, + 567, + 341, + 578 + ], + "spans": [ + { + "bbox": [ + 273, + 567, + 341, + 578 + ], + "type": "text", + "content": "Scene Language" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 579, + 501, + 591 + ], + "lines": [ + { + "bbox": [ + 305, + 579, + 501, + 591 + ], + "spans": [ + { + "bbox": [ + 305, + 579, + 501, + 591 + ], + "type": "text", + "content": "\"Make branching structure to be trinary and 3D\"" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 596, + 555, + 630 + ], + "lines": [ + { + "bbox": [ + 55, + 596, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 555, + 630 + ], + "type": "text", + "content": "Figure 7. Scene Editing with Language Instructions. Our representation is editing-friendly. Bottoms of each row show initial scene descriptions and editing instructions in the format of user text prompts. We prompt an LM to infer the initial Scene Language as well as the edits (shown with texts in red), and show image renderings with two renderers." + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 381, + 483, + 428, + 547 + ], + "blocks": [ + { + "bbox": [ + 381, + 483, + 428, + 547 + ], + "lines": [ + { + "bbox": [ + 381, + 483, + 428, + 547 + ], + "spans": [ + { + "bbox": [ + 381, + 483, + 428, + 547 + ], + "type": "image", + "image_path": "ee0dc5c0de5a5971fd3edb07b78f9de8f6365f884e8932c76844a4a293ed31b0.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 566, + 441, + 578 + ], + "lines": [ + { + "bbox": [ + 372, + 566, + 441, + 578 + ], + "spans": [ + { + "bbox": [ + 372, + 566, + 441, + 578 + ], + "type": "text", + "content": "Edited (Mitsuba)" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 478, + 483, + 525, + 546 + ], + "blocks": [ + { + "bbox": [ + 478, + 483, + 525, + 546 + ], + "lines": [ + { + "bbox": [ + 478, + 483, + 525, + 546 + ], + "spans": [ + { + "bbox": [ + 478, + 483, + 525, + 546 + ], + "type": "image", + "image_path": "d14d9edc0360e884bedfe29ddbbb652449c9d3978f5fa9f45216776f8300e5d7.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 566, + 542, + 578 + ], + "lines": [ + { + "bbox": [ + 465, + 566, + 542, + 578 + ], + "spans": [ + { + "bbox": [ + 465, + 566, + 542, + 578 + ], + "type": "text", + "content": "Edited (Gaussians)" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "bbox": [ + 55, + 635, + 294, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 294, + 658 + ], + "type": "text", + "content": "for SDS guidance in both ours and GraphDreamer) towards LM, leading to accurate and detailed generation results." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 55, + 662, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 295, + 698 + ], + "type": "text", + "content": "Across Graphics Renderers. Fig. 5 shows that our representation can be rendered with different renderers (Sec. 4), producing aligned outputs in different visual domains." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "type": "text", + "content": "4D Generation. The inference method from Sec. 5 can be" + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 313, + 635, + 554, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 554, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 554, + 671 + ], + "type": "text", + "content": "applied to 4D scene generation. The 4D scene representation is identical to Eq. (1), except that there is an additional 4D entity function in the program " + }, + { + "bbox": [ + 313, + 635, + 554, + 671 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 635, + 554, + 671 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We compare with 4D-fy [2] as an exemplar text-to-4D generation method. We measure the prompt alignment of output dynamic renderings using CLIP similarity between" + } + ] + } + ], + "index": 60 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24631" + } + ] + } + ], + "index": 61 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 556, + 170 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 556, + 170 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 556, + 170 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 556, + 170 + ], + "type": "image", + "image_path": "ee2fa90caf16d39cd01114c508865773a30df9a29d082d0207d3eb25816d7e28.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 171, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 555, + 195 + ], + "type": "text", + "content": "Figure 8. Scene Editing with Image Instructions. This figure shows style transfer results under the same setup as in Fig. 1. Replacing embeddings of target entities in a Scene Language with a new embedding specified by a user image achieves the editing effect." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 201, + 115, + 244 + ], + "blocks": [ + { + "bbox": [ + 58, + 201, + 115, + 244 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 115, + 244 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 115, + 244 + ], + "type": "image", + "image_path": "f2cb0b9931875f3420055327cf07469e908616130d0ec185f2f602efc9df62a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 246, + 177, + 255 + ], + "lines": [ + { + "bbox": [ + 116, + 246, + 177, + 255 + ], + "spans": [ + { + "bbox": [ + 116, + 246, + 177, + 255 + ], + "type": "text", + "content": "Ours (Gaussians)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 126, + 205, + 168, + 241 + ], + "blocks": [ + { + "bbox": [ + 126, + 205, + 168, + 241 + ], + "lines": [ + { + "bbox": [ + 126, + 205, + 168, + 241 + ], + "spans": [ + { + "bbox": [ + 126, + 205, + 168, + 241 + ], + "type": "image", + "image_path": "1d0ff618feb55adc6b7440bfa611a08f3f04418c1b652837e999352585f4e75c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 187, + 205, + 228, + 241 + ], + "blocks": [ + { + "bbox": [ + 187, + 205, + 228, + 241 + ], + "lines": [ + { + "bbox": [ + 187, + 205, + 228, + 241 + ], + "spans": [ + { + "bbox": [ + 187, + 205, + 228, + 241 + ], + "type": "image", + "image_path": "82e01fea7ee36cf75c2f365ecbf2039bfa9af69070e5447f60c682e5c5bf09cb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 246, + 235, + 255 + ], + "lines": [ + { + "bbox": [ + 181, + 246, + 235, + 255 + ], + "spans": [ + { + "bbox": [ + 181, + 246, + 235, + 255 + ], + "type": "text", + "content": "Ours (Mitsuba)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 244, + 201, + 288, + 246 + ], + "blocks": [ + { + "bbox": [ + 244, + 201, + 288, + 246 + ], + "lines": [ + { + "bbox": [ + 244, + 201, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 244, + 201, + 288, + 246 + ], + "type": "image", + "image_path": "554887df68e7ad17a2fa6c4436f9e403fc9e530523f6d0eb3c7bbe3d2553425d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 247, + 292, + 256 + ], + "lines": [ + { + "bbox": [ + 238, + 247, + 292, + 256 + ], + "spans": [ + { + "bbox": [ + 238, + 247, + 292, + 256 + ], + "type": "text", + "content": "GraphDreamer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 58, + 256, + 115, + 300 + ], + "blocks": [ + { + "bbox": [ + 64, + 247, + 108, + 256 + ], + "lines": [ + { + "bbox": [ + 64, + 247, + 108, + 256 + ], + "spans": [ + { + "bbox": [ + 64, + 247, + 108, + 256 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 256, + 115, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 256, + 115, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 256, + 115, + 300 + ], + "type": "image", + "image_path": "e9037c5d4e535ced3a3e62ee70bdb935fcd0250f3045d676182baaa100591ca5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 65, + 311, + 108, + 343 + ], + "blocks": [ + { + "bbox": [ + 64, + 301, + 108, + 311 + ], + "lines": [ + { + "bbox": [ + 64, + 301, + 108, + 311 + ], + "spans": [ + { + "bbox": [ + 64, + 301, + 108, + 311 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 311, + 108, + 343 + ], + "lines": [ + { + "bbox": [ + 65, + 311, + 108, + 343 + ], + "spans": [ + { + "bbox": [ + 65, + 311, + 108, + 343 + ], + "type": "image", + "image_path": "026dc445a1bcdfe81d8f38f050465c7e08a37a8b293e8172abcbeeaa9af88dd2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 350, + 108, + 359 + ], + "lines": [ + { + "bbox": [ + 64, + 350, + 108, + 359 + ], + "spans": [ + { + "bbox": [ + 64, + 350, + 108, + 359 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 117, + 269, + 175, + 300 + ], + "blocks": [ + { + "bbox": [ + 117, + 269, + 175, + 300 + ], + "lines": [ + { + "bbox": [ + 117, + 269, + 175, + 300 + ], + "spans": [ + { + "bbox": [ + 117, + 269, + 175, + 300 + ], + "type": "image", + "image_path": "b9c69da6988cc6880d39835d2b1d181b7130f140e9686040476ada7e6076ce0f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 301, + 177, + 311 + ], + "lines": [ + { + "bbox": [ + 116, + 301, + 177, + 311 + ], + "spans": [ + { + "bbox": [ + 116, + 301, + 177, + 311 + ], + "type": "text", + "content": "Ours (Gaussians)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 179, + 269, + 237, + 300 + ], + "blocks": [ + { + "bbox": [ + 179, + 269, + 237, + 300 + ], + "lines": [ + { + "bbox": [ + 179, + 269, + 237, + 300 + ], + "spans": [ + { + "bbox": [ + 179, + 269, + 237, + 300 + ], + "type": "image", + "image_path": "b7b557b2f0107c4970eeaa80279f0ee5aca402c9d72e650ba029d7fb95357dcd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 301, + 235, + 310 + ], + "lines": [ + { + "bbox": [ + 181, + 301, + 235, + 310 + ], + "spans": [ + { + "bbox": [ + 181, + 301, + 235, + 310 + ], + "type": "text", + "content": "Ours (Mitsuba)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 301, + 291, + 311 + ], + "lines": [ + { + "bbox": [ + 238, + 301, + 291, + 311 + ], + "spans": [ + { + "bbox": [ + 238, + 301, + 291, + 311 + ], + "type": "text", + "content": "GraphDreamer" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 247, + 266, + 281, + 294 + ], + "blocks": [ + { + "bbox": [ + 247, + 266, + 281, + 294 + ], + "lines": [ + { + "bbox": [ + 247, + 266, + 281, + 294 + ], + "spans": [ + { + "bbox": [ + 247, + 266, + 281, + 294 + ], + "type": "image", + "image_path": "8eca0563e95db2002d7dc86d56ba18b7f95f9c233482e98ec050e60d3f1c8b10.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 127, + 313, + 169, + 345 + ], + "blocks": [ + { + "bbox": [ + 127, + 313, + 169, + 345 + ], + "lines": [ + { + "bbox": [ + 127, + 313, + 169, + 345 + ], + "spans": [ + { + "bbox": [ + 127, + 313, + 169, + 345 + ], + "type": "image", + "image_path": "5c47dda6ab1505373b3e2088bdb2466a99391f2e2531ccf8d396f21ac2384345.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 350, + 177, + 359 + ], + "lines": [ + { + "bbox": [ + 116, + 350, + 177, + 359 + ], + "spans": [ + { + "bbox": [ + 116, + 350, + 177, + 359 + ], + "type": "text", + "content": "Ours (Gaussians)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 186, + 315, + 228, + 347 + ], + "blocks": [ + { + "bbox": [ + 186, + 315, + 228, + 347 + ], + "lines": [ + { + "bbox": [ + 186, + 315, + 228, + 347 + ], + "spans": [ + { + "bbox": [ + 186, + 315, + 228, + 347 + ], + "type": "image", + "image_path": "84a15ea6fcd20e7c50254e09d4b49d26b45a4cd70b9e5d69b0afc090e6339078.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 350, + 235, + 359 + ], + "lines": [ + { + "bbox": [ + 181, + 350, + 235, + 359 + ], + "spans": [ + { + "bbox": [ + 181, + 350, + 235, + 359 + ], + "type": "text", + "content": "Ours (Mitsuba)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 245, + 319, + 287, + 350 + ], + "blocks": [ + { + "bbox": [ + 245, + 319, + 287, + 350 + ], + "lines": [ + { + "bbox": [ + 245, + 319, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 245, + 319, + 287, + 350 + ], + "type": "image", + "image_path": "462c165752163db06354d52aa4e89439ba906ed2691b38e27f64e0ff2e8b1be8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 350, + 292, + 359 + ], + "lines": [ + { + "bbox": [ + 239, + 350, + 292, + 359 + ], + "spans": [ + { + "bbox": [ + 239, + 350, + 292, + 359 + ], + "type": "text", + "content": "GraphDreamer" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 361, + 295, + 418 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 418 + ], + "type": "text", + "content": "Figure 9. Image-Prompted Scene Generation. Compared to scene graphs from GraphDreamer, programs from our representation encode additional scene structure, e.g., repetitions, and specify geometric entity relations more precisely. Embeddings from ours further enable visual identity preservation in the outputs." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 422, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 495 + ], + "type": "text", + "content": "rendered frames and input texts, and the motion magnitude using dynamic degrees [12] computed using RAFT [35] optical flow estimation over neighboring frames. 4D-fy results have a higher per-frame alignment score than ours but present small dynamic motions. See Fig. 6, the webpage, and Appendix D.2 for further results." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "Scene Editing. Scenes synthesized from our proposed representation can further be edited by prompting LM with its previously generated script and an editing instruction in natural language instruction. Results are shown in Fig. 7. Our representation provides an interpretable and intuitive interface for scene editing, as functions have explicit semantic meanings associated with words, and function reuse significantly improves program readability. Furthermore, since the structure of programs reflects the structure of scenes, editing program parameters leads to precise changes in the scenes while preserving the original structure, e.g., the circular arrangement of staircases in Fig. 7. Desirable editing effects involving multiple primitives, or all staircases in this example, can be effectively achieved via only minor changes in the program space. Finally, the program structure itself, such as the function header in the Jenga set example, can be adjusted for editing to achieve localized edits that only affect relevant parts of the scene. Further results" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 201, + 430, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 430, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 430, + 213 + ], + "type": "text", + "content": "are shown in Appendix D.3." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 219, + 530, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 530, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 530, + 232 + ], + "type": "text", + "content": "6.2. Image-Prompted Generation and Editing" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 236, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 332 + ], + "type": "text", + "content": "Texts provide an easy user interface for scene generation as shown in Sec. 6.1, and images allow users to specify further intentions that could be hard to describe with natural language alone. Integrating neural embeddings increases the expressivity of the Scene Language and allows it to encode visual details from image prompts, and as a result, enhances our method's controllability in image-prompted scene generation and editing tasks." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "content": "Generation. Fig. 9 shows results on the task of generating 3D scenes consistent with the structure and content from a user-provided image, termed as the \"inverse semantics\" task in GraphDreamer [7]. We report the LPIPS [41] score, i.e., the AlexNet [21] feature dissimilarity, between an input image and an output rendering, averaged over 3 scenes with 20 viewpoints each in Tab. 4. Embeddings allow our representation to encode visual details from inputs, such as the Coke can texture in Fig. 9, which is not captured in scene graphs from GraphDreamer using per-object natural language descriptions. This is reflected quantitatively in Tab. 4." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 469, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 518 + ], + "type": "text", + "content": "Editing. Our representation applies to image-prompted editing tasks such as style transfer (Fig. 1). Edits can target partial or full scenes as shown in Fig. 8(a) and (b), respectively, by updating embeddings for corresponding entities." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "We have introduced a visual scene representation, termed the Scene Language, which encodes three key aspects of visual scenes: scene structure, such as hierarchy and repetition, specified via programs; semantics of individual components succinctly summarized via words; and identities of each component precisely captured via neural embeddings. We formalize the representation as a programming language defined using a DSL. We show that the Scene Language can be efficiently inferred from both text and image inputs leveraging pre-trained language models. Once the program is executed, the resulting scene can be rendered into images using a variety of graphics renderers. Our Scene Language produces 3D and 4D scenes with high fidelity and enables easy and precise editing." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24632" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "type": "text", + "content": "Acknowledgments. This work is in part supported by AFOSR YIP FA9550-23-1-0127, ONR N00014-23-1-2355, ONR YIP N00014-24-1-2117, ONR MURI N00014-22-1-2740, and NSF RI #2211258 and #2338203. YZ is in part supported by the Stanford Interdisciplinary Graduate Fellowship." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 175, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 175, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 294, + 196 + ], + "type": "text", + "content": "[1] Anthropic. The Claude 3 model family: Opus, Sonnet, Haiku, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 252 + ], + "type": "text", + "content": "[2] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B. Lindell. 4D-fly: Text-to-4D generation using hybrid score distillation sampling. CVPR, 2024. 6, 7, 16" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 295, + 297 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 244, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 244, + 309 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 244, + 309 + ], + "type": "text", + "content": "[4] Blender Online Community. Blender, 1994. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 311, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 311, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 311, + 295, + 354 + ], + "type": "text", + "content": "[5] Boyang Deng, Sumith Kulal, Zhengyang Dong, Congyue Deng, Yonglong Tian, and Jiajun Wu. Unsupervised learning of shape programs with repeatable implicit parts. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 355, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 355, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 355, + 295, + 408 + ], + "type": "text", + "content": "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 2, 3, 5, 6, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 411, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 295, + 454 + ], + "type": "text", + "content": "[7] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf. GraphDreamer: Compositional 3D scene synthesis from scene graphs. In CVPR, 2024. 3, 6, 8, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 456, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 295, + 478 + ], + "type": "text", + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 479, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 295, + 521 + ], + "type": "text", + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 524, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 524, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 524, + 295, + 567 + ], + "type": "text", + "content": "[10] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. SceneCraft: An LLM agent for synthesizing 3D scene as Blender code. In ICLR Workshop, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "text", + "content": "[11] Siyuan Huang, Siyuan Qi, Yixin Zhu, Yinxue Xiao, Yuanlu Xu, and Song-Chun Zhu. Holistic 3D scene parsing and reconstruction from a single RGB image. In ECCV, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "text", + "content": "[12] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 6, 8, 16" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[13] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, 2021. 5, 6, 16" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[14] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[15] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR, 2015. 3, 16" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 174 + ], + "type": "text", + "content": "[16] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "text", + "content": "[17] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy Mitra, and Daniel Ritchie. ShapeAssembly: Learning to generate programs for 3D shape structure synthesis. ACM TOG, 39(6):1-20, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 220, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 261 + ], + "type": "text", + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):1-14, 2023. 5, 11" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 285 + ], + "type": "text", + "content": "[19] Diederik P Kingma. Auto-encoding variational bayes. In ICLR, 2014. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 332, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 363 + ], + "type": "text", + "content": "[21] Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014.8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[22] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas Guibas. StructureNet: Hierarchical graph networks for 3D shape generation. ACM TOG, 38 (6):1-19, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 468, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 468, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 468, + 422 + ], + "type": "text", + "content": "[23] Mojang Studios. Minecraft, 2009. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 423, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 423, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 553, + 455 + ], + "type": "text", + "content": "[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 487 + ], + "type": "text", + "content": "[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In ICLR, 2023. 5, 12" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 13" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 546, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 600 + ], + "type": "text", + "content": "[27] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded SAM: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 6, 14" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 5, 13" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 679 + ], + "type": "text", + "content": "[29] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[30] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In CVPR, 2018. 3, 16" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24633" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 657 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[31] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. In ICLR, 2024. 5, 6, 14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 129 + ], + "type": "text", + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "text", + "content": "[33] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3D-GPT: Procedural 3D modeling with large language models. arXiv preprint arXiv:2310.12945, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[34] Hou In Ivan Tam, Hou In Derek Pun, Austin T Wang, Angel X Chang, and Manolis Savva. SceneMotifCoder: Example-driven visual program learning for generating 3D object arrangements. arXiv preprint arXiv:2408.02211, 2024.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 252 + ], + "type": "text", + "content": "[35] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[36] Yonglong Tian, Andrew Luo, Xingyuan Sun, Kevin Ellis, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Learning to infer and execute 3D shape programs. In ICLR, 2019. 3, 16" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[37] Yutaro Yamada, Khyathi Chandu, Yuchen Lin, Jack Hessel, Ilker Yildirim, and Yejin Choi. L3GO: Language agents with chain-of-3D-thoughts for generating unconventional objects. In ICLR Workshop, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "type": "text", + "content": "[38] Alan Yuille and Daniel Kersten. Vision as Bayesian inference: analysis by synthesis? Trends in Cognitive Sciences, 10(7):301-308, 2006. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "text", + "content": "[39] Chi Zhang, Penglin Cai, Yuhui Fu, Haoqi Yuan, and Zongqing Lu. Creative agents: Empowering agents with imagination for creative tasks. arXiv preprint arXiv:2312.02519, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "text", + "content": "[40] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[41] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "text", + "content": "[42] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. MIGC: Multi-instance generation controller for text-to-image synthesis. In CVPR, 2024. 5, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[43] Mengqi Zhou, Jun Hou, Chuanchen Luo, Yuxi Wang, Zhaoxiang Zhang, and Junran Peng. SceneX: Procedural controllable large-scale scene generation via large-language models. arXiv preprint arXiv:2403.15698, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[44] Xiaoyu Zhou, Xingjian Ran, Yajiao Xiong, Jinlin He, Zhiwei Lin, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. GALA3D: Towards text-to-3D complex scene generation via layout-guided generative Gaussian splatting. In ICML, 2024. 5, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "text", + "content": "[45] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In ECCV, 2016. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24634" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_content_list.json b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2cedde78584326902c602d06a1c0643065a19746 --- /dev/null +++ b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_content_list.json @@ -0,0 +1,1500 @@ +[ + { + "type": "text", + "text": "Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems", + "text_level": 1, + "bbox": [ + 112, + 128, + 885, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Song Xia $^{1}$ , Yi Yu $^{1}$ , Wenhan Yang $^{2*}$ , Meiwen Ding $^{1}$ , Zhuo Chen $^{2}$ , Ling-Yu Duan $^{2,3}$ , Alex C. Kot $^{1}$ , Xudong Jiang $^{1}$", + "bbox": [ + 250, + 202, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ROSE Lab, Nanyang Technological University, $^{2}$ Pengcheng Laboratory, $^{3}$ Peking University {xias0002,yuyi0010,ding0159,eackot, exdjiang}@ntu.edu.sg, yangwh@pcl.ac.cn, lingyu@pku.edu.cn", + "bbox": [ + 102, + 241, + 906, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 311, + 326, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By locally encoding raw data into intermediate features, collaborative inference enables end users to leverage powerful deep learning models without exposure of sensitive raw data to cloud servers. However, recent studies have revealed that these intermediate features may not sufficiently preserve privacy, as information can be leaked and raw data can be reconstructed via model inversion attacks (MIAs). Obfuscation-based methods, such as noise corruption, adversarial representation learning, and information filters, enhance the inversion robustness by obfuscating the task-irrelevant redundancy empirically. However, methods for quantifying such redundancy remain elusive, and the explicit mathematical relation between this redundancy minimization and inversion robustness enhancement has not yet been established. To address that, this work first theoretically proves that the conditional entropy of inputs given intermediate features provides a guaranteed lower bound on the reconstruction mean square error (MSE) under any MIA. Then, we derive a differentiable and solvable measure for bounding this conditional entropy based on the Gaussian mixture estimation and propose a conditional entropy maximization (CEM) algorithm to enhance the inversion robustness. Experimental results on four datasets demonstrate the effectiveness and adaptability of our proposed CEM; without compromising feature utility and computing efficiency, plugging the proposed CEM into obfuscation-based defense mechanisms consistently boosts their inversion robustness, achieving average gains ranging from $12.9\\%$ to $48.2\\%$ . Code is available at https://github.com/xiasong0501/CEM.", + "bbox": [ + 89, + 344, + 485, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 797, + 223, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs), trained on extensive datasets, have demonstrated outstanding performance across a growing spectrum of complex applications [22, 29, 41]. However, the increasing reliance on deploying these powerful mod", + "bbox": [ + 89, + 816, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/419dcb8ae0d9385ca64f1e1271bfdf676d2a7ba9509930e6afbf563485cc5f4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 311, + 885, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Privacy protection for collaborative inference via CEM. els on cloud platforms, such as ChatGPT-4 [2], introduces significant privacy and security concerns, as users may upload data containing sensitive information to cloud servers. Collaborative inference [48, 50, 52] offers a solution by partitioning the deep learning model across edge devices and cloud servers, where computations on the initial shallow layers are performed locally on the user's device, and only the extracted intermediate features are transmitted to the cloud for subsequent processing. This allows end users to utilize powerful neural networks with minimal exposure of their raw inputs, and hence enhances data privacy. However, recent works [4, 5, 14, 15, 21, 24, 31, 39, 44, 47, 49, 59, 65, 68, 69] have revealed that those seemingly minor signals in these intermediate features still contain substantial sensitive information. As shown in Figure 1, the MIAs can steal those unprotected features to accurately reconstruct the raw inputs.", + "bbox": [ + 511, + 474, + 906, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing defense against MIAs can be broadly categorized into cryptography-based [23, 37, 42] and obfuscation-based methods [7, 11, 17, 19, 63]. Cryptography-based methods, such as homomorphic encryption [10, 20] and secure multi-party computation [37, 54], provide robust theoretical guarantees against MIAs by computing over encrypted data. However, the inherent computational overhead poses substantial challenges for their scalability on large-scale datasets [36]. Obfuscation-based defense aims to obfuscate the task-irrelevant redundancy by learning a privacy-preserving feature encoder [3, 6, 18, 19, 27, 28, 35, 60].", + "bbox": [ + 509, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8753", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Those approaches primarily rely on empirical heuristics, such as assuming a proxy inversion adversary, to estimate task-irrelevant redundancy during encoder optimization. However, a rigorous quantification for evaluating such redundancy remains absent. Existing works [40, 61] have indicated that such empirical measure is not fully reliable, rendering it insufficient for fully exploring the inversion robustness of the trained feature encoder. Some methods [30, 36, 43, 56] employ information-theoretic frameworks to constrain the redundancy. However, none of them establishes a formal mathematical relationship between information redundancy and robustness against the worst-case inversion adversary, leaving a gap in fully understanding the interplay between redundancy minimization and robustness enhancement.", + "bbox": [ + 89, + 90, + 483, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work aims to establish a systematic quantification approach to measure the task-irrelevant yet privacy-critical redundancy within the intermediate features. Furthermore, we endeavor to establish a theoretical relationship between the quantified redundancy and the worst-case model inversion robustness, thereby providing a tractable approach to enhance the inversion robustness of existing models against MIAs. We first demonstrate that the conditional entropy of the input $x$ given intermediate feature $z$ is strongly correlated with the information leakage, which guarantees a theoretical lower bound on the reconstruction MSE between the original and reconstructed inputs under any inversion adversary. Moreover, a differentiable and tractable measure is developed for bounding this conditional entropy based on Gaussian mixture estimation. Utilizing this differentiable measure, we propose a versatile conditional entropy maximization (CEM) algorithm that can be seamlessly plugged into existing empirical obfuscation-based methods, as shown in Figure 1, to consistently enhance their robustness against MIAs. The contributions of our work can be summarized as:", + "bbox": [ + 91, + 303, + 483, + 603 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We make the first effort in establishing a theoretical relationship between the conditional entropy of inputs given intermediate features and the worst-case MIA robustness. Additionally, we derive a differentiable and tractable measure for quantifying this conditional entropy.", + "- Building upon these theoretical insights, we propose a versatile CEM algorithm that can be seamlessly plugged into existing obfuscation defense to enhance its effectiveness in defending against MIAs.", + "- We conduct extensive experiments across four datasets to empirically validate the effectiveness and adaptability of the proposed CEM algorithm. Our findings demonstrate that integrating CEM with existing obfuscation-based defenses consistently yields substantial gains in inversion robustness, without sacrificing feature utility or incurring additional computational overhead." + ], + "bbox": [ + 89, + 604, + 482, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 852, + 232, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Inversion Attacks (MIAs). MIAs represent a serious privacy threat, wherein adversaries reconstruct users'", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "private input data by exploiting the information in model parameters or the redundancy present in intermediate outputs. The inversion adversaries can leverage a generative model [39, 40, 55, 64, 66, 67], such as the generative adversarial network (GAN) [12], or a DNN-based decoder [28, 46, 62] to learn the underlying mapping between intermediate features and original inputs, thereby uncovering the hidden inversion patterns. MIAs can be executed under a variety of conditions. Early research on MIAs primarily focuses on the white-box setting where the adversary has full access to the models along with the training data [9, 16, 58, 67]. However, recent work indicated that only by accessing the intermediate features and with little prior knowledge of the data [11, 31, 32, 69], the adversary can launch strong MIAs.", + "bbox": [ + 511, + 90, + 906, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Obfuscation-based MIAs defense mechanisms. The obfuscation-based defense methods protect the input privacy by obfuscating the redundant information in the intermediate features. These methods typically adopt strategies such as perturbing network weights [1] or intermediate features [18, 35] via noise corruption, purifying intermediate features through frequency domain filtering [33, 34, 57] or sparse coding [6, 19], and training inversion-robust encoders via adversarial representation learning [3, 27, 28, 60]. Those methods generally incur no extra computational overhead during inference, thus serving as a practical and efficient solution for protecting data privacy against MIAs. Although these approaches offer practical effectiveness in defending against MIAs, they primarily measure such redundancy by some empirical heuristics without rigorous quantification [28, 40, 61], leading to a sub-optimal trade-off between feature robustness and utility. Furthermore, a formal mathematical relationship between redundancy and inversion robustness has not been established, leaving a critical gap in comprehensively understanding the interplay between redundancy minimization and robustness enhancement.", + "bbox": [ + 511, + 316, + 908, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 640, + 648, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 657, + 653, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inversion threats on collaborative inference systems: We consider MIAs on collaborative inference systems, where deep learning models are split into a lightweight encoder $\\mathcal{F}_e$ deployed locally and a decoder $\\mathcal{F}_d$ deployed in the cloud. The end-users first encode their raw input $\\pmb{x}$ into the intermediate features $\\pmb{z} = \\mathcal{F}_e(\\pmb{x})$ locally, and then upload them to the cloud for prediction. This process is known to be susceptible to MIAs, as the intermediate features $\\pmb{z}$ is considered as a direct representation of the input [8, 51].", + "bbox": [ + 511, + 674, + 906, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The attack model: We consider the scenarios where the inversion adversary $\\mathcal{A}$ can steal those user-shared intermediate features $z$ to reconstruct raw inputs. Such threats may arise from the presence of an untrustworthy cloud server or unauthorized access to data traffic during transmission. To evaluate the worst-case inversion robustness, we assume the", + "bbox": [ + 511, + 810, + 908, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8754", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adversary possesses white-box access to the feature encoder $\\mathcal{F}_e$ and has full prior knowledge of the input data space $\\mathcal{X}$ . The MIA process is formulated as $\\hat{\\boldsymbol{x}} = \\mathcal{A}(\\boldsymbol{z},\\mathcal{F}_e,\\mathcal{X})$ . This information inversion can be achieved by utilizing the DNN-based decoder [13, 28] or the generative models [39, 64, 67] to learn the potential mapping from $\\boldsymbol{z}\\rightarrow \\boldsymbol{x}$ .", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Theoretical Insights on Worst-case Robustness Assume that the information leakage is quantified by the MSE between the original $\\mathbf{x}$ and the reconstructed $\\hat{\\mathbf{x}}$ , i.e., $\\| \\hat{\\mathbf{x}} - \\mathbf{x} \\|_2^2 / d$ , where $d$ is the input dimensionality.", + "text_level": 1, + "bbox": [ + 89, + 191, + 483, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1 (Minimal reconstruction MSE $\\xi$ ). In the worst-case scenarios, where the adversary $\\mathcal{A}(z,\\mathcal{X},\\mathcal{F}_e)$ precisely estimates the posterior probability $\\mathbb{P}(\\pmb{x}|\\pmb{z})$ based on the extensive data prior $\\mathcal{X}$ and the white-box access to the feature encoder $\\mathcal{F}_{\\mathrm{e}}$ , the expectation of the minimal reconstruction MSE $\\xi$ over the whole dataset satisfies that:", + "bbox": [ + 88, + 262, + 483, + 353 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\xi \\cdot d = \\mathbb {E} _ {\\mathcal {Z}} \\mathbb {E} _ {\\mathcal {X}} \\left[ \\| \\boldsymbol {x} - \\mathbb {E} [ \\boldsymbol {x} | \\boldsymbol {z} ] \\| _ {2} ^ {2} | \\boldsymbol {z} \\right] = \\mathbb {E} _ {\\mathcal {Z}} \\left[ T r (C o v (\\boldsymbol {x} | \\boldsymbol {z}) \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 356, + 483, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$Cov(\\pmb{x}|\\pmb{z})$ is the covariance matrix of $\\pmb{x}$ conditioned on $\\pmb{z}$ , and $Tr$ is the trace operator. Eq. 1 is established based on the fact that the minimized MSE is given by the expectation of the posterior probability of $\\pmb{x}$ given $\\pmb{z}$ .", + "bbox": [ + 89, + 386, + 483, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1 (Lower bound on the minimal reconstruction MSE $\\xi$ ). Let $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ denote the conditional entropy of the input $\\boldsymbol{x}$ given the intermediate feature $\\boldsymbol{z}$ . The minimal reconstruction MSE $\\xi$ is bounded by: $\\xi \\geq \\frac{1}{(2\\pi e)}\\exp\\left(\\frac{2\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})}{d}\\right)$ .", + "bbox": [ + 89, + 454, + 483, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1 provides a lower bound on the minimal reconstruction MSE $\\xi$ in terms of the conditional entropy $\\mathcal{H}(\\pmb{x}|\\pmb{z})$ , which serves as a robust measure for the worst-case robustness against MIAs. The reconstruction MSE $\\xi$ and the conditional entropy are highly correlated. Physically, a higher $\\mathcal{H}(\\pmb{x}|\\pmb{z})$ means a more uncertain estimation of $\\pmb{x}$ given $\\pmb{z}$ , resulting in a larger estimated error $\\xi$ . Thus, indicated by Theorem 1, a straightforward way to enhance the inversion robustness is to maximize the $\\mathcal{H}(\\pmb{x}|\\pmb{z})$ during training. The proof of Theorem 1 is in the supplementary materials S.1.", + "bbox": [ + 89, + 523, + 483, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, deriving a closed-form expression between $z$ and $x$ is exceedingly difficult due to the inherent intractability of neural networks, especially when applied to large-scale datasets. This renders the direct calculation of $\\mathcal{H}(x|z)$ computationally prohibitive, making its maximization via backpropagation to eliminate the information redundancy during training exceedingly intractable. To solve this, a differentiable and computationally efficient lower bound on the conditional entropy $\\mathcal{H}(x|z)$ is introduced in the next section to facilitate its maximization during training.", + "bbox": [ + 89, + 676, + 483, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Differentiable Bound on Conditional Entropy Proposition 2 (The conditional entropy under uncertain encoding). Without loss of the generality, we consider that the encoding process $x \\rightarrow z$ consists of two components:", + "text_level": 1, + "bbox": [ + 89, + 838, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a deterministic mapping: $\\pmb{x} \\rightarrow \\hat{\\pmb{z}}$ and a stochastic process: $z = \\hat{z} + \\varepsilon$ , where $\\varepsilon$ is a random noise. $\\mathcal{H}(\\pmb{x}|\\pmb{z})$ satisfies:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} (\\boldsymbol {x} | \\boldsymbol {z}) = \\mathcal {H} (\\boldsymbol {x}) - \\mathcal {H} (\\boldsymbol {z}) + \\mathcal {H} (\\boldsymbol {z} | \\hat {\\boldsymbol {z}}) = \\mathcal {H} (\\boldsymbol {x}) - \\mathcal {I} (\\boldsymbol {z}; \\hat {\\boldsymbol {z}}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 126, + 906, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 2 is derived based on the definition of joint entropy, where $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z}) = \\mathcal{H}(\\boldsymbol{x}) + \\mathcal{H}(\\boldsymbol{z}|\\boldsymbol{x}) - \\mathcal{H}(\\boldsymbol{z})$ . The equation $\\mathcal{H}(z|\\boldsymbol{x}) = \\mathcal{H}(z|\\hat{\\boldsymbol{z}})$ holds due to the causal dependency from $\\boldsymbol{x} \\to \\hat{\\boldsymbol{z}} \\to \\boldsymbol{z}$ and $\\boldsymbol{x} \\to \\hat{\\boldsymbol{z}}$ is deterministic. Since $\\mathcal{H}(\\boldsymbol{x})$ remains constant for a given data distribution, the primary challenge in maximizing the conditional entropy $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ indicated by Eq. 2 reduces to deriving a tractable and differentiable measure for $\\mathcal{I}(\\boldsymbol{z};\\hat{\\boldsymbol{z}})$ .", + "bbox": [ + 511, + 146, + 906, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider the scenario where the task has $n$ discrete targets $\\mathbf{y} = \\{y_{1},\\dots,y_{n}\\}$ . Following previous statements, we consider that the encoding process can be separated into a deterministic mapping: $\\pmb{x} \\rightarrow \\hat{\\pmb{z}}$ and a noise corruption process: $\\pmb{z} = \\hat{\\pmb{z}} + \\varepsilon$ , where $\\varepsilon \\sim \\mathcal{N}(0,\\Sigma_p)$ is a Gaussian noise. We assume that the distribution of the encoded feature $\\hat{\\pmb{z}}$ can be effectively estimated by a $k$ -component Gaussian mixture distribution denoted as $\\hat{\\pmb{z}} \\sim \\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i,\\Sigma_i)$ , leveraging the proven capability of Gaussian mixtures to closely approximate complex natural data distributions [45].", + "bbox": [ + 511, + 267, + 908, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 2 (Differentiable lower bound on $\\mathcal{H}(\\pmb{x}|\\pmb{z})$ ). Given $\\hat{\\pmb{z}}$ follows the Gaussian mixture distribution and $\\varepsilon$ is a Gaussian noise. The encoded feature $\\pmb{z} = \\hat{\\pmb{z}} + \\varepsilon$ also follows a Gaussian mixture distribution with $\\pmb{z} \\sim \\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i, \\Sigma_i + \\Sigma_p)$ . Consequently, the conditional entropy $\\mathcal{H}(\\pmb{x}|\\pmb{z})$ is bounded by:", + "bbox": [ + 511, + 424, + 908, + 515 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} (\\boldsymbol {x} | \\boldsymbol {z}) \\geq \\mathcal {H} (\\boldsymbol {x}) - \\sum_ {i = 1} ^ {k} \\pi_ {i} \\left(- \\log \\left(\\pi_ {i}\\right) + \\frac {1}{2} \\log \\left(\\frac {\\left| \\Sigma_ {i} + \\Sigma_ {p} \\right|}{\\left| \\Sigma_ {p} \\right|}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 515, + 906, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 2 offers an effective and efficient way to bound the conditional entropy $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ , as the parameters $\\pi_{i}$ and $\\Sigma_{i}$ are all tractable and differentiable with respect to $\\hat{\\boldsymbol{z}}$ . Given the extracted $\\hat{\\boldsymbol{z}}$ after deterministic mapping and the Gaussian noise $\\varepsilon$ , one can easily derive the parameters $\\pi_{i}$ and $\\Sigma_{i}$ using a Gaussian mixture model. This facilitates the maximization of $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ via gradient backpropagation utilizing Eq. 3, thereby improving the worst-case inversion robustness of the trained model. The proof of the Theorem 2 can be found in the supplementary material S.2.", + "bbox": [ + 511, + 555, + 908, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Practical Insights on the Utility and Robustness Trade-off", + "text_level": 1, + "bbox": [ + 511, + 714, + 905, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature utility: Building on the previous discussion, the encoded feature $\\mathbf{z}$ follows a Gaussian mixture distribution with $\\mathbf{z} \\sim \\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i, \\Sigma_i + \\Sigma_p)$ . In general, we assume that each Gaussian component or cluster belongs to one specific target. Ideally, we have $k = n$ to accomplish the task. Due to the causal dependency from $\\mathbf{x} \\rightarrow \\hat{\\mathbf{z}} \\rightarrow \\mathbf{z}$ , the utility of the intermediate feature $\\mathbf{z}$ is intrinsically linked to its correlation with preceding states. Thereby, the utility of $\\mathbf{z}$ is positively correlated with the expected posterior probability that $\\hat{\\mathbf{z}}$ belonging to its original cluster $j$ given $\\mathbf{z}$ , which is:", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8755", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f325142237c0eb01188cdd6872f31f53202c7197b12a8d414626e2a49b6e5dc5.jpg", + "image_caption": [ + "Figure 2. The versatile conditional entropy maximization algorithm for collaborative learning systems with split encoder and decoder." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 867, + 300 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathcal {Z}} \\left[ \\mathbb {P} \\left(\\left(\\hat {\\boldsymbol {z}} \\in j \\mid \\boldsymbol {z}\\right)\\right) \\right] = \\mathbb {E} _ {\\mathcal {Z}} \\left[ \\frac {\\pi_ {j} \\mathcal {N} \\left(\\boldsymbol {z} ; \\mu_ {j} , \\Sigma_ {j} + \\Sigma_ {p}\\right)}{\\sum_ {i = 1} ^ {k} \\pi_ {i} \\mathcal {N} \\left(\\boldsymbol {z} ; \\mu_ {i} , \\Sigma_ {i} + \\Sigma_ {p}\\right)} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 321, + 483, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining Eq. 3 and Eq. 4, we can get that the utility and robustness of the feature $\\mathbf{z}$ are dependent on the $\\mathbb{E}_{\\mathcal{Z}}[\\mathbb{P}((\\hat{\\mathbf{z}} = j|\\mathbf{z}))]$ and the $\\mathcal{I}(z;\\hat{z})$ respectively. These two items are empirically closely linked, leading to a trade-off between utility and robustness. Additionally, the following practical insights can be derived:", + "bbox": [ + 89, + 368, + 483, + 459 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Physical interpretation of $\\mu_{i}$ and $\\Sigma_{i}$ : Physically, the mean $\\mu_{i}$ generally captures dominant or representative characteristics of a cluster, while the covariance reflects the degree of variability or noise within the cluster. Consistent with this interpretation, our analysis reveals that $\\mu_{i}$ is crucial for evaluating the feature utility. In contrast, the covariance $\\Sigma_{i}$ is closely linked to the task-irrelevant redundancy in $\\hat{z}$ , which diminishes feature utility and increases susceptibility to MIAs.", + "- Impact of noise corruption on inversion robustness: The noise corruption introduced by $\\Sigma_{p}$ perturbs both the task-important and redundant information, thus balancing the trade-off between utility and robustness. When there is a distinctive difference between the dominant features, such as $\\min_{i \\neq j} \\| \\mu_{i} - \\mu_{j} \\|_{2}^{2} \\gg Tr(\\Sigma_{p}) \\gg Tr(\\Sigma_{i} + \\Sigma_{j})$ , adding the noise greatly corrupts the redundancy while bringing a small influence on the utility.", + "- Ideal situation: For the ideal situation, we have $\\Sigma_{i} = \\Sigma_{p} = 0$ , $k = n$ and, $\\pi_{i} = \\mathcal{P}(y_{i})$ . Thereby, we can easily get $\\mathbb{E}_{\\mathcal{Z}}[\\mathbb{P}((\\hat{z} = j|z))] = 1$ and $\\mathcal{H}(\\boldsymbol{z}) = \\mathcal{H}(\\boldsymbol{y})$ , indicating that the feature $\\boldsymbol{z}$ contains only the necessary information for the task, without any redundancy." + ], + "bbox": [ + 89, + 460, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, for a lightweight encoder deployed on edge devices, it is exceedingly challenging to diminish all redundancies while preserving only the essential features. An alternative approach for developing a robust feature encoder is to jointly optimize both the covariance matrices $\\Sigma_{i}$ and the distance between the $\\mu_{j}$ among different clusters, leading to $\\min_{i\\neq j}\\| \\mu_i - \\mu_j\\| _2^2\\gg Tr(\\Sigma_p)\\gg Tr(\\Sigma_i + \\Sigma_j)$ . Thus,", + "bbox": [ + 89, + 792, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Conditional entropy maximization algorithm", + "text_level": 1, + "bbox": [ + 514, + 325, + 901, + 340 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: local encoder $\\mathcal{F}_e$ , cloud decoder $\\mathcal{F}_d$ , image dataset $\\mathcal{D}$ of size $N$ , training epochs $m$ , conditional entropy loss $\\mathcal{L}_C$ with weight factor $\\lambda$ , Gaussian noise $\\varepsilon$ , existing defense methods $\\mathcal{M}$ with loss $\\mathcal{L}_D$ , Gaussian mixture model GMM, number of mixture clusters $k$ .", + "2: Initialize $\\mathcal{F}_e, \\mathcal{F}_d$", + "3: for epoch from 1 to $m$ do", + "4: $\\pmb{z} \\gets \\text{concatenate}\\{\\mathcal{F}_e(\\pmb{x}) + \\varepsilon, \\forall \\pmb{x} \\in \\mathcal{D}\\}$", + "5: Estimate the distribution of $\\mathbf{z}$ as a $k$ component Gaussian mixture: $\\sum_{j=1}^{k} \\pi_j \\mathcal{N}(\\mu_j, \\Sigma_j + \\Sigma_p)$ using GMM( $\\mathbf{z}, k$ )", + "6: for image batch $(\\pmb{x}_i, \\pmb{y}_i)$ in the dataset $\\mathcal{D}$ do", + "7: $\\pmb{z}_i = \\mathcal{F}_e(\\pmb{x}_i) + \\varepsilon_i$", + "8: $\\pmb{y}_i' = \\mathcal{F}_d(\\pmb{z}_i)$", + "9: Assign $\\mathbf{z}_i$ to the nearest cluster", + "10: for cluster $j\\gets 1$ to $k$ do", + "11: Update the weight $\\pi_{j}$ by Eq. 5", + "12: Update the covariance matrix by Eq. 6", + "13: Calculate $\\mathcal{L}_D$ by $\\mathcal{M}(\\pmb{x}_i, \\pmb{y}_i, \\pmb{z}_i, \\pmb{y}_i')$", + "14: Calculate $\\mathcal{L}_C$ by Eq. 7.", + "15: Optimize $\\min_{\\mathcal{F}_e,\\mathcal{F}_d}\\left(\\mathcal{L}_D + \\lambda *\\mathcal{L}_C\\right)$", + "16: Output: Trained encoder $\\mathcal{F}_e$ and decoder $\\mathcal{F}_d$" + ], + "bbox": [ + 516, + 344, + 906, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "incorporating appropriate noise can effectively diminish the residual redundancies, achieving the maximization of the lower bound on $\\mathcal{H}(\\pmb{x}|\\pmb{z})$ while ensuring prediction accuracy.", + "bbox": [ + 511, + 660, + 906, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. The Versatile Conditional Entropy Maximization Algorithm for Collaborative Inference", + "text_level": 1, + "bbox": [ + 511, + 723, + 906, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building upon the derived theoretical insights on the differentiable lower bound of the conditional entropy $\\mathcal{H}(\\pmb{x}|\\pmb{z})$ and the practical insights on the utility and robustness trade-off, this section introduces a versatile conditional entropy maximization (CEM) algorithm. The CEM algorithm is designed to enhance the robustness of existing obfuscation-based defense mechanisms against MIAs in collaborative inference systems. For a given defense mechanism, denoted as $\\mathcal{M}$ , the proposed CEM algorithm begins by assessing its worst-case", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8756", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "robustness. This is achieved through a Gaussian mixture estimation process on the feature $z$ to determine the parameters $\\mu_{i}$ and $\\Sigma_{i}$ . Subsequently, during each training batch, the algorithm maximizes the lower bound of $\\mathcal{H}(x|z)$ specified in Theorem 2 to strengthen robustness against MIAs. The details of the proposed CEM algorithm are presented in Figure 2 and Algorithm 1.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Key techniques: Let $\\mathcal{F}_e$ denote the local feature encoder and $\\mathcal{F}_d$ denote the decoder deployed in the cloud. By estimating the distribution of noisy representation $z$ with a $k$ -component Gaussian mixture $\\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i, \\Sigma_i + \\Sigma_p)$ , the lower bound of the conditional entropy $\\mathcal{H}(x|z)$ can be derived based on Eq. 3, utilizing the covariance matrices $\\Sigma_i$ and the weights of the mixture components $\\pi_i$ . Since both of them are differentiable and solvable to the feature representation $z$ , this lower bound of the conditional entropy $\\mathcal{H}(x|z)$ can be effectively maximized via the gradient backpropagation. As described in Algorithm 1, each training iteration begins by fitting the distribution of $z$ with a Gaussian mixture model (GMM). This involves optimizing the component weights $\\pi$ , means $\\mu$ , and covariance matrices $\\Sigma$ to minimize the estimation error. For each data batch, the extracted feature representations are assigned to the mixture component $j$ corresponding to the nearest mean $\\mu_j$ , with the mixture parameters updated accordingly. The component weights $\\pi_j$ are updated by:", + "bbox": [ + 91, + 198, + 486, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {j} = \\frac {\\pi_ {j} \\left(N - N _ {b a t c h}\\right) + n _ {j}}{N}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 489, + 483, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the total length of the dataset and $N_{\\text{batch}}$ is the batch size. $n_j$ is the number of feature representations that are assigned to the mixture component $j$ . The covariance matrices are updated by:", + "bbox": [ + 89, + 523, + 483, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {j} = \\left(1 - \\frac {n _ {j}}{\\pi_ {j} N}\\right) \\Sigma_ {j} + \\frac {n _ {j}}{\\pi_ {j} N} \\Delta \\Sigma_ {j}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 593, + 483, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta \\Sigma_{j}$ is the calculated covariance based on newly assigned feature representations. The updated parameters are then utilized to compute the conditional entropy loss $\\mathcal{L}_C$ , defined as:", + "bbox": [ + 89, + 628, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} = \\sum_ {i = j} ^ {k} \\pi_ {j} \\left(- \\log \\left(\\pi_ {j}\\right) + \\frac {1}{2} \\log \\left(\\frac {\\left| \\Sigma_ {j} + \\Sigma_ {p} \\right|}{\\left| \\Sigma_ {p} \\right|}\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 688, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This loss function is then utilized as an effective indicator of worst-case inversion robustness and is minimized through gradient-based optimization.", + "bbox": [ + 89, + 734, + 483, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The combination with other defense strategies: As shown in Figure 2, existing obfuscation-based defense methods, such as noise corruption [18, 51], adversarial representation learning [7, 28], and feature purification [6, 19, 53] are collectively denoted as $\\mathcal{M}$ , and an auxiliary training loss $\\mathcal{L}_D$ is used to encapsulate their impact in the backward optimization process. For approaches that involve architectural modifications, such as introducing information filtering", + "bbox": [ + 89, + 780, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ab3e19b7dbc88aeb220e6675f89f57d829d21d5e527db683046544a4928e26ac.jpg", + "image_caption": [ + "Figure 3. Reconstruction MSE Vs. $\\mathcal{H}(\\pmb{x}|\\pmb{z})$ on the CIFAR10." + ], + "image_footnote": [], + "bbox": [ + 555, + 92, + 857, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "layers or adversarial reconstruction models, we adhere to their defined configurations to incorporate our method. The proposed CEM algorithm can be seamlessly plugged into other defense mechanisms by embedding Gaussian mixture estimation within the training process and optimizing the combined loss function:", + "bbox": [ + 511, + 292, + 905, + 381 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {D} + \\lambda * \\mathcal {L} _ {C}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 388, + 906, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a weight factor to balance the optimization. The utilization of the proposed CEM algorithm allows for an effective evaluation of worst-case inversion robustness of the feature representation throughout training and further optimizes this robustness via gradient backward propagation.", + "bbox": [ + 511, + 409, + 906, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditional entropy $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ and reconstruction MSE by MIAs in practical scenarios: Theorem 1 indicates an exponential relationship between the lower bound on minimal MSE $\\xi$ and the conditional entropy $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ . We thereby empirically evaluate this relationship on the CIFAR10 dataset by training multiple models with varying levels of measured $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ and assessing their inversion robustness by assuming a proxy inversion adversary. Following the experimental setting in [28], we utilize a VGG11 model with the first 2 convolutional layers as encoder $\\mathcal{F}_e$ and the rest of layers as the decoder $\\mathcal{F}_d$ . We fix the $\\lambda = 16$ and train different models using different Gaussian noise with variance varying from 0.01 to 0.3, thus leading to varying $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ . The relative conditional entropy $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ is calculated by Eq. 3 (we use 'relative' due to $\\mathcal{H}(\\boldsymbol{x})$ is constant and its calculation over the whole data space is intractable. So a proper value is selected to replace it). We report the average reconstruction MSE by training a DNN-based MIA decoder $\\mathcal{A}(\\boldsymbol{z},\\mathcal{X},\\mathcal{F}_e)$ with the same architecture as [28] and allow its full access to the training data and encoder $\\mathcal{F}_e$ . The analytic results are shown in Figure 3, revealing a strong exponential correlation between the $\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})$ and the MIA reconstruction MSE.", + "bbox": [ + 511, + 484, + 908, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 829, + 718, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 849, + 720, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We investigate the inversion robustness of collaborative inference models implemented on three general object classifi-", + "bbox": [ + 511, + 869, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8757", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dcf961df057d6cf756fc31e1b785c542b2a9fc512518b5ab7aad8436dc52e728.jpg", + "table_caption": [ + "Table 1. Comparative results on CIFAR10 dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the VGG11 model before and after the integration with the proposed CEM algorithm. The last 2 rows report the average performance of all methods, comparing results with and without incorporating the CEM algorithm." + ], + "table_footnote": [], + "table_body": "
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense91.860.00130.00140.00140.0016
Bottleneck90.870.00360.00410.00390.0045
Bottleneck+CEM90.690.00540.00580.00760.0080
DistCorr [53]89.520.00740.00820.00790.0088
DistCorr+CEM89.800.00900.00930.00940.0096
Dropout [15]87.750.00980.01040.00990.0111
Dropout+CEM87.530.01290.01340.01320.0142
PATROL [7]89.580.02450.02930.02570.0307
PATROL+CEM89.670.03040.03350.03130.0347
ResSFL [28]89.680.01460.02400.01580.0243
ResSFL+CEM90.110.02010.02510.02290.0273
Noise_Nopeek [51]87.190.01520.01590.01560.0165
Noise_Nopeek+CEM87.080.01740.01740.01730.0178
Noise_ARL [18]87.780.02900.03360.03040.0342
Noise_ARL+CEM87.620.03300.03550.03410.0363
Average w/o CEM88.910.01480.01790.01560.0185
Average w/ CEM88.920.01830.02000.01940.0211
", + "bbox": [ + 91, + 172, + 478, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cation datasets: CIFAR-10 [25], CIFAR-100 [25], TinyImageNet [26], and a face recognition dataset FaceScrub [38]. For the CIFAR-10, CIFAR-100, and FaceScrub datasets, we use the VGG11 as the basic model and for the TinyImageNet dataset, we use the ResNet-20 as the basic model.", + "bbox": [ + 89, + 422, + 483, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Collaborative inference system: To simulate the collaborative inference, we split the VGG11 model by allocating the first two convolutional layers as the local encoder and assigning the remaining layers as the cloud decoder. For the ResNet-20 model, the first four convolutional years are allocated as the local encoder, with the remaining blocks functioning as the cloud-based decoder. This partitioning yields a computational distribution where the encoder is responsible for approximately $10\\%$ of the total computation and the decoder handles the remaining $90\\%$ . Unless otherwise specified, we maintain the same partitioning scheme across all evaluated methods for a fair comparison. We utilize the Stochastic Gradient Descent (SGD) optimizer to jointly optimize both the encoder and decoder with an initial learning rate of 0.05 and an appropriate learning rate decay. The VGG11 model is trained over 240 epochs and the ResNet20 model is trained over 120 epochs.", + "bbox": [ + 88, + 503, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Obfuscation defense methods: The proposed CEM can be flexibly integrated into the model training process to enhance the worst-case inversion robustness by introducing an auxiliary Gaussian mixture estimation to maximize the conditional entropy. We evaluate the effectiveness of the proposed CEM in enhancing several existing defense methods, including the information pruning-based methods called DistCorr [53] and Dropout [15]; adversarial representation learning-based method called ResSFL [28] and", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PATROL [7]; and noise corruption-based method called Noise_Nopeek [51] and Noise_ARL [18]. As observed in [28], the integration of bottleneck layers substantially enhances inversion robustness. To ensure a fair and rigorous comparison, we incorporate a bottleneck layer with 8 channels for both pre- and post-processing of extracted feature representations across all evaluated methods. For PATROL [7] that introduces additional layers within the encoder and maintains the inference efficiency through network pruning techniques, we adjust the pruning ratio to ensure the encoder in PATROL remains the same size as other methods. To analyze the effect of the proposed CEM algorithm, we evaluate the performance of all methods both before and after integrating them with the CEM algorithm. If not otherwise stated, we set hyperparameters $\\lambda = 16$ and utilize the isotropic Gaussian noise $\\varepsilon$ with a standard deviation of 0.025. We set the number of Gaussian mixture components $k = 3n$ due to the limited number of training data and the imperfect representation ability of the shallow encoder.", + "bbox": [ + 511, + 90, + 906, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Threat model: To evaluate the worst-case inversion robustness, we consider the white-box scenarios where the MIAs are trained with full access to the collaborative inference model and training dataset. We split the training and testing data strictly following the default setting, e.g. 50,000 images for training and 10,000 for testing on the CIFAR10 dataset. Two types of inversion attacks: the DNN-based decoding method [28] and the generative-based inversion method [67] are utilized as the MIAs. The detailed architecture and training mechanism of those threat models can be found in the supplementary material S.3.", + "bbox": [ + 511, + 378, + 908, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics: We report the prediction accuracy and the reconstruction MSE to evaluate the utility and inversion robustness of the intermediate feature. We also report other metrics such as PSNR and SSIM to evaluate the robustness and present the experimental results in the supplementary material S.4. We consider the information leakage on training and inference data, where we report the MSE on reconstructing the training and testing dataset by MIAs. To further explore the intrinsic trade-off between the feature utility and inversion robustness for different defense mechanisms, we present the accuracy-MSE curve by varying the defense hyperparameters, such as the noise strength and CEM loss weight factor $\\lambda$ .", + "bbox": [ + 511, + 547, + 908, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results on Different Datasets", + "text_level": 1, + "bbox": [ + 511, + 758, + 772, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on CIFAR10: The performance of different defense methods on the CIFAR10 dataset is presented in Table 1. We provide a comparative analysis of each method, both before and after integrating the proposed CEM algorithm. The results indicate that the features, extracted by the lightweight encoder without defense mechanisms, exhibit significant redundancy, which can be exploited by MIAs to reconstruct high-fidelity inputs. Information pruning methods, such", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8758", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3bfd592597ef3a362f8564bdbd9b994bdcdddefc583ab3cbfca485773aba692f.jpg", + "table_caption": [ + "Table 2. Comparative results on TinyImageNet dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the ResNet-20 model with and without integrating the proposed CEM." + ], + "table_footnote": [], + "table_body": "
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense53.730.00250.00220.00200.0021
Bottleneck52.770.01070.01030.00910.0092
Bottleneck+CEM52.250.01400.01360.01230.0125
DistCorr [53]51.790.01480.01450.01260.0136
DistCorr+CEM51.780.01950.01900.01670.0168
Dropout [15]50.720.01650.01630.01480.0151
Dropout+CEM50.750.01880.01860.01670.0172
PATROL [7]51.750.01870.01870.01680.0176
PATROL+CEM51.640.02110.02090.01850.0189
ResSFL [28]51.990.01730.01720.01570.0161
ResSFL+CEM52.070.01970.01940.01800.0173
Noise_Nopeek [51]51.630.01610.01570.01470.0152
Noise_Nopeek+CEM52.010.01830.01800.01660.0167
Noise_ARL [18]51.030.02290.02240.02040.0205
Noise_ARL+CEM50.850.02810.02710.02510.0253
Average w/o CEM51.660.01670.01640.01480.0153
Average w/ CEM51.620.01990.01950.01770.0178
", + "bbox": [ + 91, + 148, + 485, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as DistCorr [53] and dropout [15] that are based on some empirical observations, inevitably remove task-specific information when eliminating the redundancy, leading to a degradation of the prediction accuracy. Methods based on adversarial representation learning (ARL), such as ResSFL, PATROL, and Noise_ARL, offer a better trade-off in feature utility and robustness, achieving substantial robustness gain with a moderate accuracy drop. By incorporating a Gaussian mixture estimation process that is independent of the prediction process, the proposed CEM algorithm effectively increases the lower bound of the reconstruction MSE via maximizing the conditional entropy. This positions the CEM algorithm as a highly adaptable framework for assessing and enhancing the inversion robustness of a wide range of collaborative inference models. The results in Table 1 show that the integration of the CEM algorithm consistently enhances all defense methods, yielding an average increase in reconstruction MSE of $24.0\\%$ for training data and $12.9\\%$ for inference data, without compromising prediction accuracy.", + "bbox": [ + 91, + 398, + 483, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on TinyImageNet: The performance of different defense methods on the TinyImageNet dataset is presented in Table 2, where we provide a comparative analysis of each method, both before and after integrating the proposed CEM algorithm. The results indicate again that the integration of the CEM algorithm brings substantial inversion robustness gain. Integrating the CEM algorithm yields an average increase in the reconstruction MSE of $19.4\\%$ for training data and $17.7\\%$ for inference data, without compromising prediction accuracy. Meanwhile, the combination of ARL-based methods such as PATROL and Noise_ARL with CEM achieves the best utility and robustness trade-off bringing a great robustness gain with only a marginal accuracy drop.", + "bbox": [ + 89, + 688, + 482, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Facescrub: The performance of different de", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/05962d005a0c79debaaedd7df486fb814f8d1dbeec8e5f3c4ff9ec77270508dc.jpg", + "table_caption": [ + "Table 3. Comparative results on Facescrub dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the VGG11 model with and without integrating the proposed CEM." + ], + "table_footnote": [], + "table_body": "
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense86.690.00120.00110.00120.0014
Bottleneck85.120.00250.00250.00260.0027
Bottleneck+CEM85.000.00360.00350.00380.0037
DistCorr [53]83.420.00380.00380.00410.0041
DistCorr+CEM83.780.00480.00470.00690.0074
Dropout [15]79.300.00520.00520.00540.0057
Dropout+CEM79.190.00740.00740.00760.0081
PATROL [7]79.180.00990.01180.01140.0137
PATROL+CEM79.880.01660.01850.01840.0192
ResSFL [28]79.600.00940.01120.01110.0142
ResSFL+CEM79.540.01280.01480.01430.0161
Noise_Nopeek [51]82.060.00520.00520.00530.0056
Noise_Nopeek+CEM81.960.00760.00750.00780.0090
Noise_ARL [18]80.140.01220.01550.01320.0158
Noise_ARL+CEM80.330.01820.02110.02120.0231
Average w/o CEM81.260.00690.00790.00760.088
Average w/ CEM81.380.01010.01110.01140.0123
", + "bbox": [ + 516, + 148, + 908, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fense methods on the Facescrub dataset is presented in Table 3. Unlike their performance on general object classification datasets, models in this task demonstrate increased vulnerability to MIAs. The evaluated defense methods exhibit a notable accuracy reduction as a trade-off for increased inversion robustness. This may caused by the strong correlation between facial features and identity information. In this dataset, integrating the CEM algorithm with existing defense methods brings a more pronounced improvement in inversion robustness. Specifically, the incorporation of the CEM algorithm yields an average increase in inversion robustness of $48.2\\%$ for training data and $40.1\\%$ for inference data while maintaining prediction accuracy. Meanwhile, the combination of ARL-based methods also achieves the best utility and robustness trade-off, which leads to a great robustness gain with only a marginal accuracy drop.", + "bbox": [ + 511, + 396, + 908, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on CIFAR100: The results on the CIFAR-100 dataset are provided in the supplementary materials S.5.", + "bbox": [ + 511, + 637, + 906, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualized results: We present the visualized MIA reconstruction on the TinyImageNet dataset in Figure 5. The results illustrate that incorporating the proposed CEM algorithm provides enhanced user input protection, resulting in decreased inverse reconstruction fidelity.", + "bbox": [ + 511, + 667, + 908, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study: Effect of Hyperparameters", + "text_level": 1, + "bbox": [ + 511, + 748, + 883, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To rigorously evaluate the impact of different hyperparameters, such as the noise strength and CEM weight factor $\\lambda$ , on the performance of the CEM algorithm, we conduct experiments under various settings of hyperparameters to demonstrate its robustness, versatility, and adaptability in diverse scenarios.", + "bbox": [ + 511, + 763, + 906, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effect of noise strength: As discussed in Subsection 3.4, the intensity of additive Gaussian noise is pivotal in balancing a good trade-off between utility and robust-", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8759", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c122f6ba0ad8ab67eeaddabbe8cec56cf08d48e38a83b0b2b47ef614e43982f.jpg", + "image_caption": [ + "(a) Noise variance varying from 0.01 to 0.5" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 352, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97704faa14adae42b14152fc3ce6cc7195237730e7d26a01571c738b485b67b3.jpg", + "image_caption": [ + "(b) comparing results of different value of $\\lambda$" + ], + "image_footnote": [], + "bbox": [ + 366, + 88, + 625, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1aee86309c41569810318595932f57a91956d88710e7f9362d74af4eae0e5bf1.jpg", + "image_caption": [ + "(c) comparing results of different partitions" + ], + "image_footnote": [], + "bbox": [ + 638, + 88, + 897, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ness. We thereby evaluate the performance of the CEM algorithm under various noise strengths. We utilize the Noise_Nopeek [51] and Noise_ARL [18] as the basic defense strategies $\\mathcal{M}$ . In Figure 4a, we present the prediction accuracy Vs. reconstruction MSE curves on the CIFAR-10 dataset by varying the noise variance from 0.001 to 0.5. Multiple models were trained to analyze the impact comprehensively. The results demonstrate that the noise variance significantly influences the trade-off between feature utility and robustness. The proposed CEM algorithm exhibits robust performance across the whole range of noise intensities. Integrating the CEM algorithm consistently achieves a more favorable balance between utility and robustness.", + "bbox": [ + 88, + 292, + 485, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of $\\lambda$ : In Equation 8, the weight factor $\\lambda$ is important in adjusting the conditional entropy maximization during the joint optimization. To investigate the influence of $\\lambda$ , we use the Noise_ARL method [18], which achieves a comparatively high performance as our baseline strategy $\\mathcal{M}$ . We then evaluate its performance when integrated with the CEM algorithm, using $\\lambda$ values of 0, 2, 8, 16, and 32. We present the prediction accuracy and reconstruction MSE curve in Figure 4b, by varying the noise variance from 0.001 to 0.5. The results demonstrate that increasing $\\lambda$ from 0 to 16 markedly enhances the robustness and utility trade-off.", + "bbox": [ + 88, + 491, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different partitioning mechanisms: Encoders with more layers extract more task-specific features from input data, thereby providing improved robustness against MIAs. To illustrate the effectiveness of the proposed CEM under different levels of redundancy, we evaluate its performance across multiple partitioning schemes. Specifically, we use the VGG11 architecture as the backbone, partitioned at the 1st, 2nd, and 3rd convolutional layers (denoted as $c_{1}$ , $c_{2}$ , and $c_{3}$ ), and assess the performance of CEM in combination with Noise_ARL [18]. We present the prediction accuracy Vs. reconstruction MSE curve in Figure 4c using the same strategy. The results demonstrate a significant performance advantage with deeper partitioning: specifically, partitioning at the third convolutional layer achieves substantial inversion robustness while maintaining prediction accuracy without degradation. Furthermore, the CEM algorithm exhibits high", + "bbox": [ + 89, + 659, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9f407e4ca8090589dde4511f48ba7269e5a1746ed78ca576258213cd945020e.jpg", + "image_caption": [ + "Figure 4. The effect of different hyperparameters on the performance of the proposed CEM algorithm. The effect of noise strength, $\\lambda$ , and partitioning schemes are demonstrated by the utility-robustness trade-off curve.", + "Figure 5. The visualized result. The last three rows are inputs reconstructed by MIA." + ], + "image_footnote": [], + "bbox": [ + 527, + 292, + 893, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "adaptability, consistently enhancing the robustness-utility trade-off across different partitioning schemes.", + "bbox": [ + 511, + 460, + 903, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of the efficiency: The computational complexity on the local encoder and the cloud server across various methods is presented in the supplementary material S.6. The results confirm that integrating the proposed CEM does not introduce any additional computational overhead or latency.", + "bbox": [ + 511, + 492, + 905, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 575, + 633, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work addresses problems of inversion robustness in collaborative inference systems, where MIAs can exploit subtle signals within these intermediate features to reconstruct high-fidelity inputs. Existing obfuscation defenses seek to protect privacy by empirically eliminating feature redundancy. However, the precise quantification of such redundancy is lacking, and rigorous mathematical analysis is needed to elucidate the relation between redundancy minimization and enhanced inversion robustness. To solve that, we prove that the conditional entropy of inputs given intermediate features provides a lower bound on the reconstruction MSE. Based on that, we derive a differentiable lower bound on this conditional entropy using Gaussian mixture estimation, making it amenable for efficient optimization through backpropagation. Building on this, we propose a versatile conditional entropy maximization (CEM) algorithm that can be easily plugged into existing methods. Comprehensive Experiments demonstrate that the proposed CEM significantly and consistently boosts robustness while maintaining feature utility and computational efficiency.", + "bbox": [ + 509, + 598, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8760", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was carried out at the Rapid-Rich Object Search (ROSE) Lab, School of Electrical & Electronic Engineering, Nanyang Technological University (NTU), Singapore. This research is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative, the Basic and Frontier Research Project of PCL, the Major Key Project of PCL, Guangdong Basic and Applied Basic Research Foundation under Grant 2024A1515010454, the Program of Beijing Municipal Science and Technology Commission Foundation (No.Z241100003524010), and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore and Infocomm Media Development Authority.", + "bbox": [ + 89, + 114, + 485, + 375 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 388, + 187, + 404 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sharif Abuadbba, Kyuyeon Kim, Minki Kim, Chandra Thapa, Seyit A Camtepe, Yansong Gao, Hyoungshick Kim, and Surya Nepal. Can we use split learning on 1d cnn models for privacy preserving training? In Proceedings of the 15th ACM Asia conference on computer and communications security, 2020. 2", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[3] Martin Bertran, Natalia Martinez, Afroditi Papadaki, Qiang Qiu, Miguel Rodrigues, Galen Reeves, and Guillermo Sapiro. Adversarily learned representations for information obfuscation and inference. In ICML, 2019. 1, 2", + "[4] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), 2023. 1", + "[5] Sayanton V Dibbo. Sok: Model inversion attack landscape: Taxonomy, challenges, and future roadmap. In 2023 IEEE 36th Computer Security Foundations Symposium (CSF), 2023. 1", + "[6] Sayanton V Dibbo, Adam Breuer, Juston Moore, and Michael Teti. Improving robustness to model inversion attacks via sparse coding architectures. ECCV, 2024. 1, 2, 5", + "[7] Shiwei Ding, Lan Zhang, Miao Pan, and Xiaoyong Yuan. Patrol: Privacy-oriented pruning for collaborative inference against model inversion attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 5, 6, 7", + "[8] Hao Fang, Bin Chen, Xuan Wang, Zhi Wang, and Shu-Tao Xia. Gidf: A generative gradient inversion method with feature domain optimization. In ICCV, 2023. 2" + ], + "bbox": [ + 99, + 414, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, 2015.", + "[10] Craig Gentry and Shai Halevi. Implementing gentry's fully-homomorphic encryption scheme. In Annual international conference on the theory and applications of cryptographic techniques, 2011. 1", + "[11] Xueluan Gong, Ziyao Wang, Shuaike Li, Yanjiao Chen, and Qian Wang. A gan-based defense framework against model inversion attacks. IEEE Transactions on Information Forensics and Security, 2023. 1, 2", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2", + "[13] Gyojin Han, Jaehyun Choi, Haeil Lee, and Junmo Kim. Reinforcement learning-based black-box model inversion attacks. In ICCV, 2023. 3", + "[14] Zecheng He, Tianwei Zhang, and Ruby B Lee. Model inversion attacks against collaborative inference. In Proceedings of the 35th Annual Computer Security Applications Conference, 2019. 1", + "[15] Zecheng He, Tianwei Zhang, and Ruby B Lee. Attacking and protecting data privacy in edge-cloud collaborative inference systems. IEEE Internet of Things Journal, 2020. 1, 6, 7", + "[16] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, 2017. 2", + "[17] Sy-Tuyen Ho, Koh Jun Hao, Keshigeyan Chandrasegaran, Ngoc-Bao Nguyen, and Ngai-Man Cheung. Model inversion robustness: Can transfer learning help? In CVPR, 2024. 1", + "[18] Jonghu Jeong, Minyong Cho, Philipp Benz, and Tae-hoon Kim. Noisy adversarial representation learning for effective and efficient image obfuscation. In Uncertainty in Artificial Intelligence, 2023. 1, 2, 5, 6, 7, 8", + "[19] Shuaifan Jin, He Wang, Zhibo Wang, Feng Xiao, Jiahui Hu, Yuan He, Wenwen Zhang, Zhongjie Ba, Weijie Fang, Shuhong Yuan, et al. {FaceObfuscator}: Defending deep learning-based privacy attacks with gradient descent-resistant features in face recognition. In 33rd USENIX Security Symposium (USENIX Security 24), 2024. 1, 2, 5", + "[20] Chiraag Juvekar, Vinod Vaikuntanathan, and Anantha Chandrakasan. {GAZELLE}: A low latency framework for secure neural network inference. In 27th USENIX security symposium (USENIX security 18), 2018. 1", + "[21] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1", + "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1", + "[23] Brian Knott, Shobha Venkataraman, Awni Hannun, Shubho Sengupta, Mark Ibrahim, and Laurens van der Maaten." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8761", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Cryptography: Secure multi-party computation meets machine learning. NeurIPS, 2021. 1", + "[24] Chenqi Kong, Anwei Luo, Shiqi Wang, Haoliang Li, Anderson Rocha, and Alex C Kot. Pixel-inconsistency modeling for image manipulation localization. IEEE TPAMI, 2025. 1", + "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. 2015. 6", + "[27] Ang Li, Jiayi Guo, Huanrui Yang, and Yiran Chen. Deepobfuscator: Adversarial training framework for privacy-preserving image classification. arXiv preprint arXiv:1909.04126, 2019. 1, 2", + "[28] Jingtao Li, Adnan Siraj Rakin, Xing Chen, Zhezhi He, Deliang Fan, and Chaitali Chakrabarti. Ressfl: A resistance transfer framework for defending model inversion attack in split federated learning. In CVPR, 2022. 1, 2, 3, 5, 6, 7", + "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1", + "[30] Kiwan Maeng, Chuan Guo, Sanjay Kariyappa, and G Edward Suh. Bounding the invertibility of privacy-preserving instance encoding using fisher information. NeurIPS, 2024. 2", + "[31] Shagufta Mehnaz, Sayanton V Dibbo, Roberta De Viti, Ehsanul Kabir, Björn B Brandenburg, Stefan Mangard, Ninghui Li, Elisa Bertino, Michael Backes, Emiliano De Cristofaro, et al. Are your sensitive attributes private? novel model inversion attribute inference attacks on classification models. In 31st USENIX Security Symposium (USENIX Security 22), 2022. 1, 2", + "[32] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In 2019 IEEE symposium on security and privacy (SP), 2019. 2", + "[33] Yuxi Mi, Yuge Huang, Jiazhen Ji, Minyi Zhao, Jiaxiang Wu, Xingkun Xu, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using random frequency components. In ICCV, 2023. 2", + "[34] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In CVPR, 2024. 2", + "[35] Fatemehsadat Mireshghallah, Mohammadkazem Taram, Prakash Ramrakhyani, Ali Jalali, Dean Tullsen, and Hadi Esmaeilzadeh. Shredder: Learning noise distributions to protect inference privacy. In Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, 2020. 1, 2", + "[36] Fatemehsadat Mireshghallah, Mohammadkazem Taram, Ali Jalali, Ahmed Taha Taha Elthakeb, Dean Tullsen, and Hadi Esmaeilzadeh. Not all features are equal: Discovering essential features for preserving prediction privacy. In Proceedings of the Web Conference 2021, 2021. 1, 2", + "[37] Pratyush Mishra, Ryan Lehmkuhl, Akshayaram Srinivasan, Wenting Zheng, and Raluca Ada Popa. Delphi: A cryptographic inference service for neural networks. In 29th USENIX Security Symposium (USENIX Security 20), 2020. 1" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Hong-Wei Ng and Stefan Winkler. A data-driven approach to cleaning large face datasets. In ICIP, 2014. 6", + "[39] Bao-Ngoc Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Man Cheung. Label-only model inversion attacks via knowledge transfer. NeurIPS, 2024. 1, 2, 3", + "[40] Ngoc-Bao Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Cheung. Re-thinking model inversion attacks against deep neural networks. In CVPR, 2023. 2", + "[41] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 1", + "[42] Olga Ohrimenko, Felix Schuster, Cédric Fournet, Aastha Mehta, Sebastian Nowozin, Kapil Vaswani, and Manuel Costa. Oblivious {Multi-Party} machine learning on trusted processors. In 25th USENIX Security Symposium (USENIX Security 16), 2016. 1", + "[43] Xiong Peng, Feng Liu, Jingfeng Zhang, Long Lan, Junjie Ye, Tongliang Liu, and Bo Han. Bilateral dependency optimization: Defending against model-inversion attacks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022. 2", + "[44] Yixiang Qiu, Hao Fang, Hongyao Yu, Bin Chen, MeiKang Qiu, and Shu-Tao Xia. A closer look at gan priors: Exploiting intermediate features for enhanced model inversion attacks. In ECCV, 2024. 1", + "[45] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 2009. 3", + "[46] Ahmed Salem, Apratim Bhattacharya, Michael Backes, Mario Fritz, and Yang Zhang. {Updates-Leak}: Data set inference and reconstruction attacks in online learning. In 29th USENIX security symposium (USENIX Security 20), 2020. 2", + "[47] Sunandini Sanyal, Sravanti Addepalli, and R Venkatesh Babu. Towards data-free model stealing in a hard label setting. In CVPR, 2022. 1", + "[48] Nir Shlezinger, Erez Farhan, Hai Morgenstern, and Yonina C Eldar. Collaborative inference via ensembles on the edge. In ICASSP, 2021. 1", + "[49] Lukas Struppek, Dominik Hintersdorf, Antonio De Almeida Correia, Antonia Adler, and Kristian Kersting. Plug & play attacks: Towards robust and flexible model inversion attacks. In ICML, 2022. 1", + "[50] Chandra Thapa, Pathum Chamikara Mahawaga Arachchige, Seyit Camtepe, and Lichao Sun. Splitfed: When federated learning meets split learning. In AAAI, 2022. 1", + "[51] Tom Titcombe, Adam J Hall, Pavlos Papadopoulos, and Daniele Romanini. Practical defences against model inversion attacks for split neural networks. arXiv preprint arXiv:2104.05743, 2021. 2, 5, 6, 7, 8", + "[52] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8762", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Praneeth Vepakomma, Abhishek Singh, Otkrist Gupta, and Ramesh Raskar. Nopek: Information leakage reduction to share activations in distributed deep learning. In 2020 International Conference on Data Mining Workshops (ICDMW), 2020. 5, 6, 7", + "[54] Sameer Wagh, Shruti Tople, Fabrice Benhamouda, Eyal Kushilevitz, Prateek Mittal, and Tal Rabin. Falcon: Honest-majority maliciously secure framework for private deep learning. Proceedings on Privacy Enhancing Technologies, 2021. 1", + "[55] Kuan-Chieh Wang, Yan Fu, Ke Li, Ashish Khisti, Richard Zemel, and Alireza Makhzani. Variational model inversion attacks. NeurIPS, 2021. 2", + "[56] Tianhao Wang, Yuheng Zhang, and Ruoxi Jia. Improving robustness to model inversion attacks via mutual information regularization. In AAAI, 2021. 2", + "[57] Yinggui Wang, Jian Liu, Man Luo, Le Yang, and Li Wang. Privacy-preserving face recognition in the frequency domain. In AAAI, 2022. 2", + "[58] Zhibo Wang, Mengkai Song, Zhifei Zhang, Yang Song, Qian Wang, and Hairong Qi. Beyond inferring class representatives: User-level privacy leakage from federated learning. In IEEE INFOCOM 2019-IEEE conference on computer communications, 2019. 2", + "[59] Song Xia, Wenhan Yang, Yi Yu, Xun Lin, Henghui Ding, Lingyu Duan, and Xudong Jiang. Transferable adversarial attacks on sam and its downstream models. In NeurIPS, 2024. 1", + "[60] Taihong Xiao, Yi-Hsuan Tsai, Kihyuk Sohn, Manmohan Chandraker, and Ming-Hsuan Yang. Adversarial learning of privacy-preserving and task-oriented representations. In AAAI, 2020. 1, 2", + "[61] Mengda Yang, Ziang Li, Juan Wang, Hongxin Hu, Ao Ren, Xiaoyang Xu, and Wenzhe Yi. Measuring data reconstruction defenses in collaborative inference systems. NeurIPS, 2022. 2", + "[62] Ziqi Yang, Jiyi Zhang, Ee-Chien Chang, and Zhenkai Liang. Neural network inversion in adversarial setting via background knowledge alignment. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 2019. 2", + "[63] Ziqi Yang, Lijin Wang, Da Yang, Jie Wan, Ziming Zhao, Ee-Chien Chang, Fan Zhang, and Kui Ren. Purifier: Defending data inference attacks via transforming confidence scores. In AAAI, 2023. 1", + "[64] Yupeng Yin, Xianglong Zhang, Huanle Zhang, Feng Li, Yue Yu, Xiuzhen Cheng, and Pengfei Hu. Ginver: Generative model inversion attacks against collaborative inference. In Proceedings of the ACM Web Conference 2023, 2023. 2, 3", + "[65] Yi Yu, Song Xia, Xun Lin, Wenhan Yang, Shijian Lu, Yappeng Tan, and Alex Kot. Backdoor attacks against no-reference image quality assessment models via a scalable trigger. In AAAI, 2025. 1", + "[66] Xiaojian Yuan, Kejiang Chen, Jie Zhang, Weiming Zhang, Nenghai Yu, and Yang Zhang. Pseudo label-guided model inversion attack via conditional generative adversarial network. In AAAI, 2023. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Yuheng Zhang, Ruoxi Jia, Hengzhi Pei, Wenxiao Wang, Bo Li, and Dawn Song. The secret revealer: Generative model-inversion attacks against deep neural networks. In CVPR, 2020. 2, 3, 6, 7", + "[68] Xuejun Zhao, Wencan Zhang, Xiaokui Xiao, and Brian Lim. Exploiting explanations for model inversion attacks. In ICCV, 2021. 1", + "[69] Wei Zong, Yang-Wai Chow, Willy Susilo, Joonsang Baek, Jongkil Kim, and Seyit Camtepe. Ipremover: A generative model inversion attack against deep neural network fingerprinting and watermarking. In AAAI, 2024. 1, 2" + ], + "bbox": [ + 514, + 90, + 906, + 247 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8763", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_model.json b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_model.json new file mode 100644 index 0000000000000000000000000000000000000000..075ae7ecbdd0dc55102471a615bbe3edf3e5235e --- /dev/null +++ b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.13, + 0.887, + 0.177 + ], + "angle": 0, + "content": "Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.203, + 0.761, + 0.24 + ], + "angle": 0, + "content": "Song Xia\\(^{1}\\), Yi Yu\\(^{1}\\), Wenhan Yang\\(^{2*}\\), Meiwen Ding\\(^{1}\\), Zhuo Chen\\(^{2}\\), Ling-Yu Duan\\(^{2,3}\\), Alex C. Kot\\(^{1}\\), Xudong Jiang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.242, + 0.908, + 0.279 + ], + "angle": 0, + "content": "\\(^{1}\\)ROSE Lab, Nanyang Technological University, \\(^{2}\\)Pengcheng Laboratory, \\(^{3}\\)Peking University {xias0002,yuyi0010,ding0159,eackot, exdjiang}@ntu.edu.sg, yangwh@pcl.ac.cn, lingyu@pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.313, + 0.327, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.486, + 0.784 + ], + "angle": 0, + "content": "By locally encoding raw data into intermediate features, collaborative inference enables end users to leverage powerful deep learning models without exposure of sensitive raw data to cloud servers. However, recent studies have revealed that these intermediate features may not sufficiently preserve privacy, as information can be leaked and raw data can be reconstructed via model inversion attacks (MIAs). Obfuscation-based methods, such as noise corruption, adversarial representation learning, and information filters, enhance the inversion robustness by obfuscating the task-irrelevant redundancy empirically. However, methods for quantifying such redundancy remain elusive, and the explicit mathematical relation between this redundancy minimization and inversion robustness enhancement has not yet been established. To address that, this work first theoretically proves that the conditional entropy of inputs given intermediate features provides a guaranteed lower bound on the reconstruction mean square error (MSE) under any MIA. Then, we derive a differentiable and solvable measure for bounding this conditional entropy based on the Gaussian mixture estimation and propose a conditional entropy maximization (CEM) algorithm to enhance the inversion robustness. Experimental results on four datasets demonstrate the effectiveness and adaptability of our proposed CEM; without compromising feature utility and computing efficiency, plugging the proposed CEM into obfuscation-based defense mechanisms consistently boosts their inversion robustness, achieving average gains ranging from \\(12.9\\%\\) to \\(48.2\\%\\). Code is available at https://github.com/xiasong0501/CEM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.798, + 0.224, + 0.813 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Deep neural networks (DNNs), trained on extensive datasets, have demonstrated outstanding performance across a growing spectrum of complex applications [22, 29, 41]. However, the increasing reliance on deploying these powerful mod" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.312, + 0.887, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.475, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Figure 1. Privacy protection for collaborative inference via CEM. els on cloud platforms, such as ChatGPT-4 [2], introduces significant privacy and security concerns, as users may upload data containing sensitive information to cloud servers. Collaborative inference [48, 50, 52] offers a solution by partitioning the deep learning model across edge devices and cloud servers, where computations on the initial shallow layers are performed locally on the user's device, and only the extracted intermediate features are transmitted to the cloud for subsequent processing. This allows end users to utilize powerful neural networks with minimal exposure of their raw inputs, and hence enhances data privacy. However, recent works [4, 5, 14, 15, 21, 24, 31, 39, 44, 47, 49, 59, 65, 68, 69] have revealed that those seemingly minor signals in these intermediate features still contain substantial sensitive information. As shown in Figure 1, the MIAs can steal those unprotected features to accurately reconstruct the raw inputs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Existing defense against MIAs can be broadly categorized into cryptography-based [23, 37, 42] and obfuscation-based methods [7, 11, 17, 19, 63]. Cryptography-based methods, such as homomorphic encryption [10, 20] and secure multi-party computation [37, 54], provide robust theoretical guarantees against MIAs by computing over encrypted data. However, the inherent computational overhead poses substantial challenges for their scalability on large-scale datasets [36]. Obfuscation-based defense aims to obfuscate the task-irrelevant redundancy by learning a privacy-preserving feature encoder [3, 6, 18, 19, 27, 28, 35, 60]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.239, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8753" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.302 + ], + "angle": 0, + "content": "Those approaches primarily rely on empirical heuristics, such as assuming a proxy inversion adversary, to estimate task-irrelevant redundancy during encoder optimization. However, a rigorous quantification for evaluating such redundancy remains absent. Existing works [40, 61] have indicated that such empirical measure is not fully reliable, rendering it insufficient for fully exploring the inversion robustness of the trained feature encoder. Some methods [30, 36, 43, 56] employ information-theoretic frameworks to constrain the redundancy. However, none of them establishes a formal mathematical relationship between information redundancy and robustness against the worst-case inversion adversary, leaving a gap in fully understanding the interplay between redundancy minimization and robustness enhancement." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.304, + 0.484, + 0.604 + ], + "angle": 0, + "content": "This work aims to establish a systematic quantification approach to measure the task-irrelevant yet privacy-critical redundancy within the intermediate features. Furthermore, we endeavor to establish a theoretical relationship between the quantified redundancy and the worst-case model inversion robustness, thereby providing a tractable approach to enhance the inversion robustness of existing models against MIAs. We first demonstrate that the conditional entropy of the input \\( x \\) given intermediate feature \\( z \\) is strongly correlated with the information leakage, which guarantees a theoretical lower bound on the reconstruction MSE between the original and reconstructed inputs under any inversion adversary. Moreover, a differentiable and tractable measure is developed for bounding this conditional entropy based on Gaussian mixture estimation. Utilizing this differentiable measure, we propose a versatile conditional entropy maximization (CEM) algorithm that can be seamlessly plugged into existing empirical obfuscation-based methods, as shown in Figure 1, to consistently enhance their robustness against MIAs. The contributions of our work can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.606, + 0.483, + 0.68 + ], + "angle": 0, + "content": "- We make the first effort in establishing a theoretical relationship between the conditional entropy of inputs given intermediate features and the worst-case MIA robustness. Additionally, we derive a differentiable and tractable measure for quantifying this conditional entropy." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.681, + 0.483, + 0.74 + ], + "angle": 0, + "content": "- Building upon these theoretical insights, we propose a versatile CEM algorithm that can be seamlessly plugged into existing obfuscation defense to enhance its effectiveness in defending against MIAs." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.741, + 0.483, + 0.845 + ], + "angle": 0, + "content": "- We conduct extensive experiments across four datasets to empirically validate the effectiveness and adaptability of the proposed CEM algorithm. Our findings demonstrate that integrating CEM with existing obfuscation-based defenses consistently yields substantial gains in inversion robustness, without sacrificing feature utility or incurring additional computational overhead." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.606, + 0.483, + 0.845 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.853, + 0.233, + 0.868 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Model Inversion Attacks (MIAs). MIAs represent a serious privacy threat, wherein adversaries reconstruct users'" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.318 + ], + "angle": 0, + "content": "private input data by exploiting the information in model parameters or the redundancy present in intermediate outputs. The inversion adversaries can leverage a generative model [39, 40, 55, 64, 66, 67], such as the generative adversarial network (GAN) [12], or a DNN-based decoder [28, 46, 62] to learn the underlying mapping between intermediate features and original inputs, thereby uncovering the hidden inversion patterns. MIAs can be executed under a variety of conditions. Early research on MIAs primarily focuses on the white-box setting where the adversary has full access to the models along with the training data [9, 16, 58, 67]. However, recent work indicated that only by accessing the intermediate features and with little prior knowledge of the data [11, 31, 32, 69], the adversary can launch strong MIAs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.909, + 0.635 + ], + "angle": 0, + "content": "Obfuscation-based MIAs defense mechanisms. The obfuscation-based defense methods protect the input privacy by obfuscating the redundant information in the intermediate features. These methods typically adopt strategies such as perturbing network weights [1] or intermediate features [18, 35] via noise corruption, purifying intermediate features through frequency domain filtering [33, 34, 57] or sparse coding [6, 19], and training inversion-robust encoders via adversarial representation learning [3, 27, 28, 60]. Those methods generally incur no extra computational overhead during inference, thus serving as a practical and efficient solution for protecting data privacy against MIAs. Although these approaches offer practical effectiveness in defending against MIAs, they primarily measure such redundancy by some empirical heuristics without rigorous quantification [28, 40, 61], leading to a sub-optimal trade-off between feature robustness and utility. Furthermore, a formal mathematical relationship between redundancy and inversion robustness has not been established, leaving a critical gap in comprehensively understanding the interplay between redundancy minimization and robustness enhancement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.649, + 0.657 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.659, + 0.655, + 0.673 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Inversion threats on collaborative inference systems: We consider MIAs on collaborative inference systems, where deep learning models are split into a lightweight encoder \\(\\mathcal{F}_e\\) deployed locally and a decoder \\(\\mathcal{F}_d\\) deployed in the cloud. The end-users first encode their raw input \\(\\pmb{x}\\) into the intermediate features \\(\\pmb{z} = \\mathcal{F}_e(\\pmb{x})\\) locally, and then upload them to the cloud for prediction. This process is known to be susceptible to MIAs, as the intermediate features \\(\\pmb{z}\\) is considered as a direct representation of the input [8, 51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.9 + ], + "angle": 0, + "content": "The attack model: We consider the scenarios where the inversion adversary \\(\\mathcal{A}\\) can steal those user-shared intermediate features \\(z\\) to reconstruct raw inputs. Such threats may arise from the presence of an untrustworthy cloud server or unauthorized access to data traffic during transmission. To evaluate the worst-case inversion robustness, we assume the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8754" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "adversary possesses white-box access to the feature encoder \\(\\mathcal{F}_e\\) and has full prior knowledge of the input data space \\(\\mathcal{X}\\). The MIA process is formulated as \\(\\hat{\\boldsymbol{x}} = \\mathcal{A}(\\boldsymbol{z},\\mathcal{F}_e,\\mathcal{X})\\). This information inversion can be achieved by utilizing the DNN-based decoder [13, 28] or the generative models [39, 64, 67] to learn the potential mapping from \\(\\boldsymbol{z}\\rightarrow \\boldsymbol{x}\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.193, + 0.484, + 0.258 + ], + "angle": 0, + "content": "3.2. Theoretical Insights on Worst-case Robustness Assume that the information leakage is quantified by the MSE between the original \\( \\mathbf{x} \\) and the reconstructed \\( \\hat{\\mathbf{x}} \\), i.e., \\( \\| \\hat{\\mathbf{x}} - \\mathbf{x} \\|_2^2 / d \\), where \\( d \\) is the input dimensionality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.263, + 0.485, + 0.354 + ], + "angle": 0, + "content": "Proposition 1 (Minimal reconstruction MSE \\(\\xi\\)). In the worst-case scenarios, where the adversary \\(\\mathcal{A}(z,\\mathcal{X},\\mathcal{F}_e)\\) precisely estimates the posterior probability \\(\\mathbb{P}(\\pmb{x}|\\pmb{z})\\) based on the extensive data prior \\(\\mathcal{X}\\) and the white-box access to the feature encoder \\(\\mathcal{F}_{\\mathrm{e}}\\), the expectation of the minimal reconstruction MSE \\(\\xi\\) over the whole dataset satisfies that:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.357, + 0.484, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\xi \\cdot d = \\mathbb {E} _ {\\mathcal {Z}} \\mathbb {E} _ {\\mathcal {X}} \\left[ \\| \\boldsymbol {x} - \\mathbb {E} [ \\boldsymbol {x} | \\boldsymbol {z} ] \\| _ {2} ^ {2} | \\boldsymbol {z} \\right] = \\mathbb {E} _ {\\mathcal {Z}} \\left[ T r (C o v (\\boldsymbol {x} | \\boldsymbol {z}) \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.485, + 0.448 + ], + "angle": 0, + "content": "\\(Cov(\\pmb{x}|\\pmb{z})\\) is the covariance matrix of \\(\\pmb{x}\\) conditioned on \\(\\pmb{z}\\), and \\(Tr\\) is the trace operator. Eq. 1 is established based on the fact that the minimized MSE is given by the expectation of the posterior probability of \\(\\pmb{x}\\) given \\(\\pmb{z}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.52 + ], + "angle": 0, + "content": "Theorem 1 (Lower bound on the minimal reconstruction MSE \\(\\xi\\)). Let \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\) denote the conditional entropy of the input \\(\\boldsymbol{x}\\) given the intermediate feature \\(\\boldsymbol{z}\\). The minimal reconstruction MSE \\(\\xi\\) is bounded by: \\(\\xi \\geq \\frac{1}{(2\\pi e)}\\exp\\left(\\frac{2\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})}{d}\\right)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.677 + ], + "angle": 0, + "content": "Theorem 1 provides a lower bound on the minimal reconstruction MSE \\(\\xi\\) in terms of the conditional entropy \\(\\mathcal{H}(\\pmb{x}|\\pmb{z})\\), which serves as a robust measure for the worst-case robustness against MIAs. The reconstruction MSE \\(\\xi\\) and the conditional entropy are highly correlated. Physically, a higher \\(\\mathcal{H}(\\pmb{x}|\\pmb{z})\\) means a more uncertain estimation of \\(\\pmb{x}\\) given \\(\\pmb{z}\\), resulting in a larger estimated error \\(\\xi\\). Thus, indicated by Theorem 1, a straightforward way to enhance the inversion robustness is to maximize the \\(\\mathcal{H}(\\pmb{x}|\\pmb{z})\\) during training. The proof of Theorem 1 is in the supplementary materials S.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.484, + 0.83 + ], + "angle": 0, + "content": "However, deriving a closed-form expression between \\( z \\) and \\( x \\) is exceedingly difficult due to the inherent intractability of neural networks, especially when applied to large-scale datasets. This renders the direct calculation of \\( \\mathcal{H}(x|z) \\) computationally prohibitive, making its maximization via backpropagation to eliminate the information redundancy during training exceedingly intractable. To solve this, a differentiable and computationally efficient lower bound on the conditional entropy \\( \\mathcal{H}(x|z) \\) is introduced in the next section to facilitate its maximization during training." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.839, + 0.484, + 0.903 + ], + "angle": 0, + "content": "3.3. Differentiable Bound on Conditional Entropy Proposition 2 (The conditional entropy under uncertain encoding). Without loss of the generality, we consider that the encoding process \\( x \\rightarrow z \\) consists of two components:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "a deterministic mapping: \\( \\pmb{x} \\rightarrow \\hat{\\pmb{z}} \\) and a stochastic process: \\( z = \\hat{z} + \\varepsilon \\), where \\( \\varepsilon \\) is a random noise. \\( \\mathcal{H}(\\pmb{x}|\\pmb{z}) \\) satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.127, + 0.907, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} (\\boldsymbol {x} | \\boldsymbol {z}) = \\mathcal {H} (\\boldsymbol {x}) - \\mathcal {H} (\\boldsymbol {z}) + \\mathcal {H} (\\boldsymbol {z} | \\hat {\\boldsymbol {z}}) = \\mathcal {H} (\\boldsymbol {x}) - \\mathcal {I} (\\boldsymbol {z}; \\hat {\\boldsymbol {z}}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.147, + 0.907, + 0.267 + ], + "angle": 0, + "content": "Proposition 2 is derived based on the definition of joint entropy, where \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z}) = \\mathcal{H}(\\boldsymbol{x}) + \\mathcal{H}(\\boldsymbol{z}|\\boldsymbol{x}) - \\mathcal{H}(\\boldsymbol{z})\\). The equation \\(\\mathcal{H}(z|\\boldsymbol{x}) = \\mathcal{H}(z|\\hat{\\boldsymbol{z}})\\) holds due to the causal dependency from \\(\\boldsymbol{x} \\to \\hat{\\boldsymbol{z}} \\to \\boldsymbol{z}\\) and \\(\\boldsymbol{x} \\to \\hat{\\boldsymbol{z}}\\) is deterministic. Since \\(\\mathcal{H}(\\boldsymbol{x})\\) remains constant for a given data distribution, the primary challenge in maximizing the conditional entropy \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\) indicated by Eq. 2 reduces to deriving a tractable and differentiable measure for \\(\\mathcal{I}(\\boldsymbol{z};\\hat{\\boldsymbol{z}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.909, + 0.42 + ], + "angle": 0, + "content": "Consider the scenario where the task has \\( n \\) discrete targets \\( \\mathbf{y} = \\{y_{1},\\dots,y_{n}\\} \\). Following previous statements, we consider that the encoding process can be separated into a deterministic mapping: \\( \\pmb{x} \\rightarrow \\hat{\\pmb{z}} \\) and a noise corruption process: \\( \\pmb{z} = \\hat{\\pmb{z}} + \\varepsilon \\), where \\( \\varepsilon \\sim \\mathcal{N}(0,\\Sigma_p) \\) is a Gaussian noise. We assume that the distribution of the encoded feature \\( \\hat{\\pmb{z}} \\) can be effectively estimated by a \\( k \\)-component Gaussian mixture distribution denoted as \\( \\hat{\\pmb{z}} \\sim \\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i,\\Sigma_i) \\), leveraging the proven capability of Gaussian mixtures to closely approximate complex natural data distributions [45]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.425, + 0.909, + 0.516 + ], + "angle": 0, + "content": "Theorem 2 (Differentiable lower bound on \\(\\mathcal{H}(\\pmb{x}|\\pmb{z})\\)). Given \\(\\hat{\\pmb{z}}\\) follows the Gaussian mixture distribution and \\(\\varepsilon\\) is a Gaussian noise. The encoded feature \\(\\pmb{z} = \\hat{\\pmb{z}} + \\varepsilon\\) also follows a Gaussian mixture distribution with \\(\\pmb{z} \\sim \\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i, \\Sigma_i + \\Sigma_p)\\). Consequently, the conditional entropy \\(\\mathcal{H}(\\pmb{x}|\\pmb{z})\\) is bounded by:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.516, + 0.907, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} (\\boldsymbol {x} | \\boldsymbol {z}) \\geq \\mathcal {H} (\\boldsymbol {x}) - \\sum_ {i = 1} ^ {k} \\pi_ {i} \\left(- \\log \\left(\\pi_ {i}\\right) + \\frac {1}{2} \\log \\left(\\frac {\\left| \\Sigma_ {i} + \\Sigma_ {p} \\right|}{\\left| \\Sigma_ {p} \\right|}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.909, + 0.707 + ], + "angle": 0, + "content": "Theorem 2 offers an effective and efficient way to bound the conditional entropy \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\), as the parameters \\(\\pi_{i}\\) and \\(\\Sigma_{i}\\) are all tractable and differentiable with respect to \\(\\hat{\\boldsymbol{z}}\\). Given the extracted \\(\\hat{\\boldsymbol{z}}\\) after deterministic mapping and the Gaussian noise \\(\\varepsilon\\), one can easily derive the parameters \\(\\pi_{i}\\) and \\(\\Sigma_{i}\\) using a Gaussian mixture model. This facilitates the maximization of \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\) via gradient backpropagation utilizing Eq. 3, thereby improving the worst-case inversion robustness of the trained model. The proof of the Theorem 2 can be found in the supplementary material S.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.715, + 0.906, + 0.745 + ], + "angle": 0, + "content": "3.4. Practical Insights on the Utility and Robustness Trade-off" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Feature utility: Building on the previous discussion, the encoded feature \\( \\mathbf{z} \\) follows a Gaussian mixture distribution with \\( \\mathbf{z} \\sim \\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i, \\Sigma_i + \\Sigma_p) \\). In general, we assume that each Gaussian component or cluster belongs to one specific target. Ideally, we have \\( k = n \\) to accomplish the task. Due to the causal dependency from \\( \\mathbf{x} \\rightarrow \\hat{\\mathbf{z}} \\rightarrow \\mathbf{z} \\), the utility of the intermediate feature \\( \\mathbf{z} \\) is intrinsically linked to its correlation with preceding states. Thereby, the utility of \\( \\mathbf{z} \\) is positively correlated with the expected posterior probability that \\( \\hat{\\mathbf{z}} \\) belonging to its original cluster \\( j \\) given \\( \\mathbf{z} \\), which is:" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8755" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.868, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.3, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Figure 2. The versatile conditional entropy maximization algorithm for collaborative learning systems with split encoder and decoder." + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.322, + 0.484, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathcal {Z}} \\left[ \\mathbb {P} \\left(\\left(\\hat {\\boldsymbol {z}} \\in j \\mid \\boldsymbol {z}\\right)\\right) \\right] = \\mathbb {E} _ {\\mathcal {Z}} \\left[ \\frac {\\pi_ {j} \\mathcal {N} \\left(\\boldsymbol {z} ; \\mu_ {j} , \\Sigma_ {j} + \\Sigma_ {p}\\right)}{\\sum_ {i = 1} ^ {k} \\pi_ {i} \\mathcal {N} \\left(\\boldsymbol {z} ; \\mu_ {i} , \\Sigma_ {i} + \\Sigma_ {p}\\right)} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.46 + ], + "angle": 0, + "content": "Combining Eq. 3 and Eq. 4, we can get that the utility and robustness of the feature \\( \\mathbf{z} \\) are dependent on the \\( \\mathbb{E}_{\\mathcal{Z}}[\\mathbb{P}((\\hat{\\mathbf{z}} = j|\\mathbf{z}))] \\) and the \\( \\mathcal{I}(z;\\hat{z}) \\) respectively. These two items are empirically closely linked, leading to a trade-off between utility and robustness. Additionally, the following practical insights can be derived:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.461, + 0.484, + 0.596 + ], + "angle": 0, + "content": "- Physical interpretation of \\(\\mu_{i}\\) and \\(\\Sigma_{i}\\): Physically, the mean \\(\\mu_{i}\\) generally captures dominant or representative characteristics of a cluster, while the covariance reflects the degree of variability or noise within the cluster. Consistent with this interpretation, our analysis reveals that \\(\\mu_{i}\\) is crucial for evaluating the feature utility. In contrast, the covariance \\(\\Sigma_{i}\\) is closely linked to the task-irrelevant redundancy in \\(\\hat{z}\\), which diminishes feature utility and increases susceptibility to MIAs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.597, + 0.484, + 0.717 + ], + "angle": 0, + "content": "- Impact of noise corruption on inversion robustness: The noise corruption introduced by \\(\\Sigma_{p}\\) perturbs both the task-important and redundant information, thus balancing the trade-off between utility and robustness. When there is a distinctive difference between the dominant features, such as \\(\\min_{i \\neq j} \\| \\mu_{i} - \\mu_{j} \\|_{2}^{2} \\gg Tr(\\Sigma_{p}) \\gg Tr(\\Sigma_{i} + \\Sigma_{j})\\), adding the noise greatly corrupts the redundancy while bringing a small influence on the utility." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.718, + 0.484, + 0.793 + ], + "angle": 0, + "content": "- Ideal situation: For the ideal situation, we have \\(\\Sigma_{i} = \\Sigma_{p} = 0\\), \\(k = n\\) and, \\(\\pi_{i} = \\mathcal{P}(y_{i})\\). Thereby, we can easily get \\(\\mathbb{E}_{\\mathcal{Z}}[\\mathbb{P}((\\hat{z} = j|z))] = 1\\) and \\(\\mathcal{H}(\\boldsymbol{z}) = \\mathcal{H}(\\boldsymbol{y})\\), indicating that the feature \\(\\boldsymbol{z}\\) contains only the necessary information for the task, without any redundancy." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.461, + 0.484, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.484, + 0.904 + ], + "angle": 0, + "content": "However, for a lightweight encoder deployed on edge devices, it is exceedingly challenging to diminish all redundancies while preserving only the essential features. An alternative approach for developing a robust feature encoder is to jointly optimize both the covariance matrices \\(\\Sigma_{i}\\) and the distance between the \\(\\mu_{j}\\) among different clusters, leading to \\(\\min_{i\\neq j}\\| \\mu_i - \\mu_j\\| _2^2\\gg Tr(\\Sigma_p)\\gg Tr(\\Sigma_i + \\Sigma_j)\\). Thus," + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.326, + 0.902, + 0.342 + ], + "angle": 0, + "content": "Algorithm 1 Conditional entropy maximization algorithm" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.345, + 0.907, + 0.413 + ], + "angle": 0, + "content": "1: Input: local encoder \\(\\mathcal{F}_e\\), cloud decoder \\(\\mathcal{F}_d\\), image dataset \\(\\mathcal{D}\\) of size \\(N\\), training epochs \\(m\\), conditional entropy loss \\(\\mathcal{L}_C\\) with weight factor \\(\\lambda\\), Gaussian noise \\(\\varepsilon\\), existing defense methods \\(\\mathcal{M}\\) with loss \\(\\mathcal{L}_D\\), Gaussian mixture model GMM, number of mixture clusters \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.415, + 0.646, + 0.428 + ], + "angle": 0, + "content": "2: Initialize \\(\\mathcal{F}_e, \\mathcal{F}_d\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.429, + 0.695, + 0.441 + ], + "angle": 0, + "content": "3: for epoch from 1 to \\( m \\) do" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.442, + 0.801, + 0.456 + ], + "angle": 0, + "content": "4: \\( \\pmb{z} \\gets \\text{concatenate}\\{\\mathcal{F}_e(\\pmb{x}) + \\varepsilon, \\forall \\pmb{x} \\in \\mathcal{D}\\} \\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.457, + 0.905, + 0.485 + ], + "angle": 0, + "content": "5: Estimate the distribution of \\( \\mathbf{z} \\) as a \\( k \\) component Gaussian mixture: \\( \\sum_{j=1}^{k} \\pi_j \\mathcal{N}(\\mu_j, \\Sigma_j + \\Sigma_p) \\) using GMM(\\( \\mathbf{z}, k \\))" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.485, + 0.826, + 0.497 + ], + "angle": 0, + "content": "6: for image batch \\((\\pmb{x}_i, \\pmb{y}_i)\\) in the dataset \\(\\mathcal{D}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.498, + 0.696, + 0.511 + ], + "angle": 0, + "content": "7: \\(\\pmb{z}_i = \\mathcal{F}_e(\\pmb{x}_i) + \\varepsilon_i\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.512, + 0.663, + 0.526 + ], + "angle": 0, + "content": "8: \\( \\pmb{y}_i' = \\mathcal{F}_d(\\pmb{z}_i) \\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.527, + 0.771, + 0.539 + ], + "angle": 0, + "content": "9: Assign \\( \\mathbf{z}_i \\) to the nearest cluster" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.54, + 0.736, + 0.553 + ], + "angle": 0, + "content": "10: for cluster \\(j\\gets 1\\) to \\(k\\) do" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.554, + 0.789, + 0.567 + ], + "angle": 0, + "content": "11: Update the weight \\(\\pi_{j}\\) by Eq. 5" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.568, + 0.836, + 0.581 + ], + "angle": 0, + "content": "12: Update the covariance matrix by Eq. 6" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.583, + 0.796, + 0.597 + ], + "angle": 0, + "content": "13: Calculate \\(\\mathcal{L}_D\\) by \\(\\mathcal{M}(\\pmb{x}_i, \\pmb{y}_i, \\pmb{z}_i, \\pmb{y}_i')\\)" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.598, + 0.724, + 0.611 + ], + "angle": 0, + "content": "14: Calculate \\(\\mathcal{L}_C\\) by Eq. 7." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.612, + 0.804, + 0.626 + ], + "angle": 0, + "content": "15: Optimize \\(\\min_{\\mathcal{F}_e,\\mathcal{F}_d}\\left(\\mathcal{L}_D + \\lambda *\\mathcal{L}_C\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.628, + 0.812, + 0.641 + ], + "angle": 0, + "content": "16: Output: Trained encoder \\(\\mathcal{F}_e\\) and decoder \\(\\mathcal{F}_d\\)" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.345, + 0.907, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.908, + 0.708 + ], + "angle": 0, + "content": "incorporating appropriate noise can effectively diminish the residual redundancies, achieving the maximization of the lower bound on \\(\\mathcal{H}(\\pmb{x}|\\pmb{z})\\) while ensuring prediction accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.908, + 0.759 + ], + "angle": 0, + "content": "4. The Versatile Conditional Entropy Maximization Algorithm for Collaborative Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Building upon the derived theoretical insights on the differentiable lower bound of the conditional entropy \\(\\mathcal{H}(\\pmb{x}|\\pmb{z})\\) and the practical insights on the utility and robustness trade-off, this section introduces a versatile conditional entropy maximization (CEM) algorithm. The CEM algorithm is designed to enhance the robustness of existing obfuscation-based defense mechanisms against MIAs in collaborative inference systems. For a given defense mechanism, denoted as \\(\\mathcal{M}\\), the proposed CEM algorithm begins by assessing its worst-case" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8756" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "robustness. This is achieved through a Gaussian mixture estimation process on the feature \\( z \\) to determine the parameters \\( \\mu_{i} \\) and \\( \\Sigma_{i} \\). Subsequently, during each training batch, the algorithm maximizes the lower bound of \\( \\mathcal{H}(x|z) \\) specified in Theorem 2 to strengthen robustness against MIAs. The details of the proposed CEM algorithm are presented in Figure 2 and Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.199, + 0.487, + 0.486 + ], + "angle": 0, + "content": "Key techniques: Let \\(\\mathcal{F}_e\\) denote the local feature encoder and \\(\\mathcal{F}_d\\) denote the decoder deployed in the cloud. By estimating the distribution of noisy representation \\(z\\) with a \\(k\\)-component Gaussian mixture \\(\\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i, \\Sigma_i + \\Sigma_p)\\), the lower bound of the conditional entropy \\(\\mathcal{H}(x|z)\\) can be derived based on Eq. 3, utilizing the covariance matrices \\(\\Sigma_i\\) and the weights of the mixture components \\(\\pi_i\\). Since both of them are differentiable and solvable to the feature representation \\(z\\), this lower bound of the conditional entropy \\(\\mathcal{H}(x|z)\\) can be effectively maximized via the gradient backpropagation. As described in Algorithm 1, each training iteration begins by fitting the distribution of \\(z\\) with a Gaussian mixture model (GMM). This involves optimizing the component weights \\(\\pi\\), means \\(\\mu\\), and covariance matrices \\(\\Sigma\\) to minimize the estimation error. For each data batch, the extracted feature representations are assigned to the mixture component \\(j\\) corresponding to the nearest mean \\(\\mu_j\\), with the mixture parameters updated accordingly. The component weights \\(\\pi_j\\) are updated by:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.491, + 0.484, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\pi_ {j} = \\frac {\\pi_ {j} \\left(N - N _ {b a t c h}\\right) + n _ {j}}{N}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.587 + ], + "angle": 0, + "content": "where \\(N\\) is the total length of the dataset and \\(N_{\\text{batch}}\\) is the batch size. \\(n_j\\) is the number of feature representations that are assigned to the mixture component \\(j\\). The covariance matrices are updated by:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.594, + 0.484, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {j} = \\left(1 - \\frac {n _ {j}}{\\pi_ {j} N}\\right) \\Sigma_ {j} + \\frac {n _ {j}}{\\pi_ {j} N} \\Delta \\Sigma_ {j}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.485, + 0.687 + ], + "angle": 0, + "content": "where \\(\\Delta \\Sigma_{j}\\) is the calculated covariance based on newly assigned feature representations. The updated parameters are then utilized to compute the conditional entropy loss \\(\\mathcal{L}_C\\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.689, + 0.484, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} = \\sum_ {i = j} ^ {k} \\pi_ {j} \\left(- \\log \\left(\\pi_ {j}\\right) + \\frac {1}{2} \\log \\left(\\frac {\\left| \\Sigma_ {j} + \\Sigma_ {p} \\right|}{\\left| \\Sigma_ {p} \\right|}\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.779 + ], + "angle": 0, + "content": "This loss function is then utilized as an effective indicator of worst-case inversion robustness and is minimized through gradient-based optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The combination with other defense strategies: As shown in Figure 2, existing obfuscation-based defense methods, such as noise corruption [18, 51], adversarial representation learning [7, 28], and feature purification [6, 19, 53] are collectively denoted as \\(\\mathcal{M}\\), and an auxiliary training loss \\(\\mathcal{L}_D\\) is used to encapsulate their impact in the backward optimization process. For approaches that involve architectural modifications, such as introducing information filtering" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.093, + 0.858, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.275, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Figure 3. Reconstruction MSE Vs. \\(\\mathcal{H}(\\pmb{x}|\\pmb{z})\\) on the CIFAR10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.906, + 0.382 + ], + "angle": 0, + "content": "layers or adversarial reconstruction models, we adhere to their defined configurations to incorporate our method. The proposed CEM algorithm can be seamlessly plugged into other defense mechanisms by embedding Gaussian mixture estimation within the training process and optimizing the combined loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.389, + 0.907, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {D} + \\lambda * \\mathcal {L} _ {C}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.485 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a weight factor to balance the optimization. The utilization of the proposed CEM algorithm allows for an effective evaluation of worst-case inversion robustness of the feature representation throughout training and further optimizes this robustness via gradient backward propagation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.486, + 0.909, + 0.818 + ], + "angle": 0, + "content": "Conditional entropy \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\) and reconstruction MSE by MIAs in practical scenarios: Theorem 1 indicates an exponential relationship between the lower bound on minimal MSE \\(\\xi\\) and the conditional entropy \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\). We thereby empirically evaluate this relationship on the CIFAR10 dataset by training multiple models with varying levels of measured \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\) and assessing their inversion robustness by assuming a proxy inversion adversary. Following the experimental setting in [28], we utilize a VGG11 model with the first 2 convolutional layers as encoder \\(\\mathcal{F}_e\\) and the rest of layers as the decoder \\(\\mathcal{F}_d\\). We fix the \\(\\lambda = 16\\) and train different models using different Gaussian noise with variance varying from 0.01 to 0.3, thus leading to varying \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\). The relative conditional entropy \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\) is calculated by Eq. 3 (we use 'relative' due to \\(\\mathcal{H}(\\boldsymbol{x})\\) is constant and its calculation over the whole data space is intractable. So a proper value is selected to replace it). We report the average reconstruction MSE by training a DNN-based MIA decoder \\(\\mathcal{A}(\\boldsymbol{z},\\mathcal{X},\\mathcal{F}_e)\\) with the same architecture as [28] and allow its full access to the training data and encoder \\(\\mathcal{F}_e\\). The analytic results are shown in Figure 3, revealing a strong exponential correlation between the \\(\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})\\) and the MIA reconstruction MSE." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.72, + 0.847 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.851, + 0.722, + 0.867 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We investigate the inversion robustness of collaborative inference models implemented on three general object classifi-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8757" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.173 + ], + "angle": 0, + "content": "Table 1. Comparative results on CIFAR10 dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the VGG11 model before and after the integration with the proposed CEM algorithm. The last 2 rows report the average performance of all methods, comparing results with and without incorporating the CEM algorithm." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.174, + 0.48, + 0.41 + ], + "angle": 0, + "content": "
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense91.860.00130.00140.00140.0016
Bottleneck90.870.00360.00410.00390.0045
Bottleneck+CEM90.690.00540.00580.00760.0080
DistCorr [53]89.520.00740.00820.00790.0088
DistCorr+CEM89.800.00900.00930.00940.0096
Dropout [15]87.750.00980.01040.00990.0111
Dropout+CEM87.530.01290.01340.01320.0142
PATROL [7]89.580.02450.02930.02570.0307
PATROL+CEM89.670.03040.03350.03130.0347
ResSFL [28]89.680.01460.02400.01580.0243
ResSFL+CEM90.110.02010.02510.02290.0273
Noise_Nopeek [51]87.190.01520.01590.01560.0165
Noise_Nopeek+CEM87.080.01740.01740.01730.0178
Noise_ARL [18]87.780.02900.03360.03040.0342
Noise_ARL+CEM87.620.03300.03550.03410.0363
Average w/o CEM88.910.01480.01790.01560.0185
Average w/ CEM88.920.01830.02000.01940.0211
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.485, + 0.501 + ], + "angle": 0, + "content": "cation datasets: CIFAR-10 [25], CIFAR-100 [25], TinyImageNet [26], and a face recognition dataset FaceScrub [38]. For the CIFAR-10, CIFAR-100, and FaceScrub datasets, we use the VGG11 as the basic model and for the TinyImageNet dataset, we use the ResNet-20 as the basic model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Collaborative inference system: To simulate the collaborative inference, we split the VGG11 model by allocating the first two convolutional layers as the local encoder and assigning the remaining layers as the cloud decoder. For the ResNet-20 model, the first four convolutional years are allocated as the local encoder, with the remaining blocks functioning as the cloud-based decoder. This partitioning yields a computational distribution where the encoder is responsible for approximately \\(10\\%\\) of the total computation and the decoder handles the remaining \\(90\\%\\). Unless otherwise specified, we maintain the same partitioning scheme across all evaluated methods for a fair comparison. We utilize the Stochastic Gradient Descent (SGD) optimizer to jointly optimize both the encoder and decoder with an initial learning rate of 0.05 and an appropriate learning rate decay. The VGG11 model is trained over 240 epochs and the ResNet20 model is trained over 120 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Obfuscation defense methods: The proposed CEM can be flexibly integrated into the model training process to enhance the worst-case inversion robustness by introducing an auxiliary Gaussian mixture estimation to maximize the conditional entropy. We evaluate the effectiveness of the proposed CEM in enhancing several existing defense methods, including the information pruning-based methods called DistCorr [53] and Dropout [15]; adversarial representation learning-based method called ResSFL [28] and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.379 + ], + "angle": 0, + "content": "PATROL [7]; and noise corruption-based method called Noise_Nopeek [51] and Noise_ARL [18]. As observed in [28], the integration of bottleneck layers substantially enhances inversion robustness. To ensure a fair and rigorous comparison, we incorporate a bottleneck layer with 8 channels for both pre- and post-processing of extracted feature representations across all evaluated methods. For PATROL [7] that introduces additional layers within the encoder and maintains the inference efficiency through network pruning techniques, we adjust the pruning ratio to ensure the encoder in PATROL remains the same size as other methods. To analyze the effect of the proposed CEM algorithm, we evaluate the performance of all methods both before and after integrating them with the CEM algorithm. If not otherwise stated, we set hyperparameters \\(\\lambda = 16\\) and utilize the isotropic Gaussian noise \\(\\varepsilon\\) with a standard deviation of 0.025. We set the number of Gaussian mixture components \\(k = 3n\\) due to the limited number of training data and the imperfect representation ability of the shallow encoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.909, + 0.547 + ], + "angle": 0, + "content": "Threat model: To evaluate the worst-case inversion robustness, we consider the white-box scenarios where the MIAs are trained with full access to the collaborative inference model and training dataset. We split the training and testing data strictly following the default setting, e.g. 50,000 images for training and 10,000 for testing on the CIFAR10 dataset. Two types of inversion attacks: the DNN-based decoding method [28] and the generative-based inversion method [67] are utilized as the MIAs. The detailed architecture and training mechanism of those threat models can be found in the supplementary material S.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.744 + ], + "angle": 0, + "content": "Evaluation metrics: We report the prediction accuracy and the reconstruction MSE to evaluate the utility and inversion robustness of the intermediate feature. We also report other metrics such as PSNR and SSIM to evaluate the robustness and present the experimental results in the supplementary material S.4. We consider the information leakage on training and inference data, where we report the MSE on reconstructing the training and testing dataset by MIAs. To further explore the intrinsic trade-off between the feature utility and inversion robustness for different defense mechanisms, we present the accuracy-MSE curve by varying the defense hyperparameters, such as the noise strength and CEM loss weight factor \\(\\lambda\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.759, + 0.774, + 0.773 + ], + "angle": 0, + "content": "5.2. Results on Different Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results on CIFAR10: The performance of different defense methods on the CIFAR10 dataset is presented in Table 1. We provide a comparative analysis of each method, both before and after integrating the proposed CEM algorithm. The results indicate that the features, extracted by the lightweight encoder without defense mechanisms, exhibit significant redundancy, which can be exploited by MIAs to reconstruct high-fidelity inputs. Information pruning methods, such" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8758" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.146 + ], + "angle": 0, + "content": "Table 2. Comparative results on TinyImageNet dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the ResNet-20 model with and without integrating the proposed CEM." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.15, + 0.486, + 0.391 + ], + "angle": 0, + "content": "
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense53.730.00250.00220.00200.0021
Bottleneck52.770.01070.01030.00910.0092
Bottleneck+CEM52.250.01400.01360.01230.0125
DistCorr [53]51.790.01480.01450.01260.0136
DistCorr+CEM51.780.01950.01900.01670.0168
Dropout [15]50.720.01650.01630.01480.0151
Dropout+CEM50.750.01880.01860.01670.0172
PATROL [7]51.750.01870.01870.01680.0176
PATROL+CEM51.640.02110.02090.01850.0189
ResSFL [28]51.990.01730.01720.01570.0161
ResSFL+CEM52.070.01970.01940.01800.0173
Noise_Nopeek [51]51.630.01610.01570.01470.0152
Noise_Nopeek+CEM52.010.01830.01800.01660.0167
Noise_ARL [18]51.030.02290.02240.02040.0205
Noise_ARL+CEM50.850.02810.02710.02510.0253
Average w/o CEM51.660.01670.01640.01480.0153
Average w/ CEM51.620.01990.01950.01770.0178
" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.4, + 0.485, + 0.687 + ], + "angle": 0, + "content": "as DistCorr [53] and dropout [15] that are based on some empirical observations, inevitably remove task-specific information when eliminating the redundancy, leading to a degradation of the prediction accuracy. Methods based on adversarial representation learning (ARL), such as ResSFL, PATROL, and Noise_ARL, offer a better trade-off in feature utility and robustness, achieving substantial robustness gain with a moderate accuracy drop. By incorporating a Gaussian mixture estimation process that is independent of the prediction process, the proposed CEM algorithm effectively increases the lower bound of the reconstruction MSE via maximizing the conditional entropy. This positions the CEM algorithm as a highly adaptable framework for assessing and enhancing the inversion robustness of a wide range of collaborative inference models. The results in Table 1 show that the integration of the CEM algorithm consistently enhances all defense methods, yielding an average increase in reconstruction MSE of \\(24.0\\%\\) for training data and \\(12.9\\%\\) for inference data, without compromising prediction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Results on TinyImageNet: The performance of different defense methods on the TinyImageNet dataset is presented in Table 2, where we provide a comparative analysis of each method, both before and after integrating the proposed CEM algorithm. The results indicate again that the integration of the CEM algorithm brings substantial inversion robustness gain. Integrating the CEM algorithm yields an average increase in the reconstruction MSE of \\(19.4\\%\\) for training data and \\(17.7\\%\\) for inference data, without compromising prediction accuracy. Meanwhile, the combination of ARL-based methods such as PATROL and Noise_ARL with CEM achieves the best utility and robustness trade-off bringing a great robustness gain with only a marginal accuracy drop." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Results on Facescrub: The performance of different de" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 3. Comparative results on Facescrub dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the VGG11 model with and without integrating the proposed CEM." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.15, + 0.909, + 0.39 + ], + "angle": 0, + "content": "
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense86.690.00120.00110.00120.0014
Bottleneck85.120.00250.00250.00260.0027
Bottleneck+CEM85.000.00360.00350.00380.0037
DistCorr [53]83.420.00380.00380.00410.0041
DistCorr+CEM83.780.00480.00470.00690.0074
Dropout [15]79.300.00520.00520.00540.0057
Dropout+CEM79.190.00740.00740.00760.0081
PATROL [7]79.180.00990.01180.01140.0137
PATROL+CEM79.880.01660.01850.01840.0192
ResSFL [28]79.600.00940.01120.01110.0142
ResSFL+CEM79.540.01280.01480.01430.0161
Noise_Nopeek [51]82.060.00520.00520.00530.0056
Noise_Nopeek+CEM81.960.00760.00750.00780.0090
Noise_ARL [18]80.140.01220.01550.01320.0158
Noise_ARL+CEM80.330.01820.02110.02120.0231
Average w/o CEM81.260.00690.00790.00760.088
Average w/ CEM81.380.01010.01110.01140.0123
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.909, + 0.638 + ], + "angle": 0, + "content": "fense methods on the Facescrub dataset is presented in Table 3. Unlike their performance on general object classification datasets, models in this task demonstrate increased vulnerability to MIAs. The evaluated defense methods exhibit a notable accuracy reduction as a trade-off for increased inversion robustness. This may caused by the strong correlation between facial features and identity information. In this dataset, integrating the CEM algorithm with existing defense methods brings a more pronounced improvement in inversion robustness. Specifically, the incorporation of the CEM algorithm yields an average increase in inversion robustness of \\(48.2\\%\\) for training data and \\(40.1\\%\\) for inference data while maintaining prediction accuracy. Meanwhile, the combination of ARL-based methods also achieves the best utility and robustness trade-off, which leads to a great robustness gain with only a marginal accuracy drop." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.638, + 0.907, + 0.668 + ], + "angle": 0, + "content": "Results on CIFAR100: The results on the CIFAR-100 dataset are provided in the supplementary materials S.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.909, + 0.744 + ], + "angle": 0, + "content": "Visualized results: We present the visualized MIA reconstruction on the TinyImageNet dataset in Figure 5. The results illustrate that incorporating the proposed CEM algorithm provides enhanced user input protection, resulting in decreased inverse reconstruction fidelity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.885, + 0.765 + ], + "angle": 0, + "content": "5.3. Ablation Study: Effect of Hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.854 + ], + "angle": 0, + "content": "To rigorously evaluate the impact of different hyperparameters, such as the noise strength and CEM weight factor \\(\\lambda\\), on the performance of the CEM algorithm, we conduct experiments under various settings of hyperparameters to demonstrate its robustness, versatility, and adaptability in diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The effect of noise strength: As discussed in Subsection 3.4, the intensity of additive Gaussian noise is pivotal in balancing a good trade-off between utility and robust-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8759" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.353, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.243, + 0.338, + 0.256 + ], + "angle": 0, + "content": "(a) Noise variance varying from 0.01 to 0.5" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.089, + 0.627, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.244, + 0.614, + 0.256 + ], + "angle": 0, + "content": "(b) comparing results of different value of \\(\\lambda\\)" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.089, + 0.898, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.244, + 0.884, + 0.256 + ], + "angle": 0, + "content": "(c) comparing results of different partitions" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.905, + 0.286 + ], + "angle": 0, + "content": "Figure 4. The effect of different hyperparameters on the performance of the proposed CEM algorithm. The effect of noise strength, \\(\\lambda\\), and partitioning schemes are demonstrated by the utility-robustness trade-off curve." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.294, + 0.486, + 0.49 + ], + "angle": 0, + "content": "ness. We thereby evaluate the performance of the CEM algorithm under various noise strengths. We utilize the Noise_Nopeek [51] and Noise_ARL [18] as the basic defense strategies \\(\\mathcal{M}\\). In Figure 4a, we present the prediction accuracy Vs. reconstruction MSE curves on the CIFAR-10 dataset by varying the noise variance from 0.001 to 0.5. Multiple models were trained to analyze the impact comprehensively. The results demonstrate that the noise variance significantly influences the trade-off between feature utility and robustness. The proposed CEM algorithm exhibits robust performance across the whole range of noise intensities. Integrating the CEM algorithm consistently achieves a more favorable balance between utility and robustness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.485, + 0.658 + ], + "angle": 0, + "content": "The effect of \\(\\lambda\\): In Equation 8, the weight factor \\(\\lambda\\) is important in adjusting the conditional entropy maximization during the joint optimization. To investigate the influence of \\(\\lambda\\), we use the Noise_ARL method [18], which achieves a comparatively high performance as our baseline strategy \\(\\mathcal{M}\\). We then evaluate its performance when integrated with the CEM algorithm, using \\(\\lambda\\) values of 0, 2, 8, 16, and 32. We present the prediction accuracy and reconstruction MSE curve in Figure 4b, by varying the noise variance from 0.001 to 0.5. The results demonstrate that increasing \\(\\lambda\\) from 0 to 16 markedly enhances the robustness and utility trade-off." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Different partitioning mechanisms: Encoders with more layers extract more task-specific features from input data, thereby providing improved robustness against MIAs. To illustrate the effectiveness of the proposed CEM under different levels of redundancy, we evaluate its performance across multiple partitioning schemes. Specifically, we use the VGG11 architecture as the backbone, partitioned at the 1st, 2nd, and 3rd convolutional layers (denoted as \\( c_{1} \\), \\( c_{2} \\), and \\( c_{3} \\)), and assess the performance of CEM in combination with Noise_ARL [18]. We present the prediction accuracy Vs. reconstruction MSE curve in Figure 4c using the same strategy. The results demonstrate a significant performance advantage with deeper partitioning: specifically, partitioning at the third convolutional layer achieves substantial inversion robustness while maintaining prediction accuracy without degradation. Furthermore, the CEM algorithm exhibits high" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.293, + 0.895, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.459 + ], + "angle": 0, + "content": "Figure 5. The visualized result. The last three rows are inputs reconstructed by MIA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.905, + 0.492 + ], + "angle": 0, + "content": "adaptability, consistently enhancing the robustness-utility trade-off across different partitioning schemes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.569 + ], + "angle": 0, + "content": "Analysis of the efficiency: The computational complexity on the local encoder and the cloud server across various methods is presented in the supplementary material S.6. The results confirm that integrating the proposed CEM does not introduce any additional computational overhead or latency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.634, + 0.592 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This work addresses problems of inversion robustness in collaborative inference systems, where MIAs can exploit subtle signals within these intermediate features to reconstruct high-fidelity inputs. Existing obfuscation defenses seek to protect privacy by empirically eliminating feature redundancy. However, the precise quantification of such redundancy is lacking, and rigorous mathematical analysis is needed to elucidate the relation between redundancy minimization and enhanced inversion robustness. To solve that, we prove that the conditional entropy of inputs given intermediate features provides a lower bound on the reconstruction MSE. Based on that, we derive a differentiable lower bound on this conditional entropy using Gaussian mixture estimation, making it amenable for efficient optimization through backpropagation. Building on this, we propose a versatile conditional entropy maximization (CEM) algorithm that can be easily plugged into existing methods. Comprehensive Experiments demonstrate that the proposed CEM significantly and consistently boosts robustness while maintaining feature utility and computational efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8760" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.375 + ], + "angle": 0, + "content": "This work was carried out at the Rapid-Rich Object Search (ROSE) Lab, School of Electrical & Electronic Engineering, Nanyang Technological University (NTU), Singapore. This research is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative, the Basic and Frontier Research Project of PCL, the Major Key Project of PCL, Guangdong Basic and Applied Basic Research Foundation under Grant 2024A1515010454, the Program of Beijing Municipal Science and Technology Commission Foundation (No.Z241100003524010), and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore and Infocomm Media Development Authority." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.389, + 0.188, + 0.405 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.415, + 0.484, + 0.497 + ], + "angle": 0, + "content": "[1] Sharif Abuadbba, Kyuyeon Kim, Minki Kim, Chandra Thapa, Seyit A Camtepe, Yansong Gao, Hyoungshick Kim, and Surya Nepal. Can we use split learning on 1d cnn models for privacy preserving training? In Proceedings of the 15th ACM Asia conference on computer and communications security, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.5, + 0.484, + 0.556 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[3] Martin Bertran, Natalia Martinez, Afroditi Papadaki, Qiang Qiu, Miguel Rodrigues, Galen Reeves, and Guillermo Sapiro. Adversarily learned representations for information obfuscation and inference. In ICML, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[4] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.687, + 0.484, + 0.741 + ], + "angle": 0, + "content": "[5] Sayanton V Dibbo. Sok: Model inversion attack landscape: Taxonomy, challenges, and future roadmap. In 2023 IEEE 36th Computer Security Foundations Symposium (CSF), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.744, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[6] Sayanton V Dibbo, Adam Breuer, Juston Moore, and Michael Teti. Improving robustness to model inversion attacks via sparse coding architectures. ECCV, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[7] Shiwei Ding, Lan Zhang, Miao Pan, and Xiaoyong Yuan. Patrol: Privacy-oriented pruning for collaborative inference against model inversion attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[8] Hao Fang, Bin Chen, Xuan Wang, Zhi Wang, and Shu-Tao Xia. Gidf: A generative gradient inversion method with feature domain optimization. In ICCV, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.415, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[9] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[10] Craig Gentry and Shai Halevi. Implementing gentry's fully-homomorphic encryption scheme. In Annual international conference on the theory and applications of cryptographic techniques, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.907, + 0.273 + ], + "angle": 0, + "content": "[11] Xueluan Gong, Ziyao Wang, Shuaike Li, Yanjiao Chen, and Qian Wang. A gan-based defense framework against model inversion attacks. IEEE Transactions on Information Forensics and Security, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.907, + 0.327 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[13] Gyojin Han, Jaehyun Choi, Haeil Lee, and Junmo Kim. Reinforcement learning-based black-box model inversion attacks. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.907, + 0.425 + ], + "angle": 0, + "content": "[14] Zecheng He, Tianwei Zhang, and Ruby B Lee. Model inversion attacks against collaborative inference. In Proceedings of the 35th Annual Computer Security Applications Conference, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.427, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[15] Zecheng He, Tianwei Zhang, and Ruby B Lee. Attacking and protecting data privacy in edge-cloud collaborative inference systems. IEEE Internet of Things Journal, 2020. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.907, + 0.537 + ], + "angle": 0, + "content": "[16] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.907, + 0.58 + ], + "angle": 0, + "content": "[17] Sy-Tuyen Ho, Koh Jun Hao, Keshigeyan Chandrasegaran, Ngoc-Bao Nguyen, and Ngai-Man Cheung. Model inversion robustness: Can transfer learning help? In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[18] Jonghu Jeong, Minyong Cho, Philipp Benz, and Tae-hoon Kim. Noisy adversarial representation learning for effective and efficient image obfuscation. In Uncertainty in Artificial Intelligence, 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.907, + 0.719 + ], + "angle": 0, + "content": "[19] Shuaifan Jin, He Wang, Zhibo Wang, Feng Xiao, Jiahui Hu, Yuan He, Wenwen Zhang, Zhongjie Ba, Weijie Fang, Shuhong Yuan, et al. {FaceObfuscator}: Defending deep learning-based privacy attacks with gradient descent-resistant features in face recognition. In 33rd USENIX Security Symposium (USENIX Security 24), 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.907, + 0.774 + ], + "angle": 0, + "content": "[20] Chiraag Juvekar, Vinod Vaikuntanathan, and Anantha Chandrakasan. {GAZELLE}: A low latency framework for secure neural network inference. In 27th USENIX security symposium (USENIX security 18), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.907, + 0.817 + ], + "angle": 0, + "content": "[21] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[23] Brian Knott, Shobha Venkataraman, Awni Hannun, Shubho Sengupta, Mark Ibrahim, and Laurens van der Maaten." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8761" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Cryptography: Secure multi-party computation meets machine learning. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.484, + 0.162 + ], + "angle": 0, + "content": "[24] Chenqi Kong, Anwei Luo, Shiqi Wang, Haoliang Li, Anderson Rocha, and Alex C Kot. Pixel-inconsistency modeling for image manipulation localization. IEEE TPAMI, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.163, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.191, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[27] Ang Li, Jiayi Guo, Huanrui Yang, and Yiran Chen. Deepobfuscator: Adversarial training framework for privacy-preserving image classification. arXiv preprint arXiv:1909.04126, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[28] Jingtao Li, Adnan Siraj Rakin, Xing Chen, Zhezhi He, Deliang Fan, and Chaitali Chakrabarti. Ressfl: A resistance transfer framework for defending model inversion attack in split federated learning. In CVPR, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[30] Kiwan Maeng, Chuan Guo, Sanjay Kariyappa, and G Edward Suh. Bounding the invertibility of privacy-preserving instance encoding using fisher information. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[31] Shagufta Mehnaz, Sayanton V Dibbo, Roberta De Viti, Ehsanul Kabir, Björn B Brandenburg, Stefan Mangard, Ninghui Li, Elisa Bertino, Michael Backes, Emiliano De Cristofaro, et al. Are your sensitive attributes private? novel model inversion attribute inference attacks on classification models. In 31st USENIX Security Symposium (USENIX Security 22), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[32] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In 2019 IEEE symposium on security and privacy (SP), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[33] Yuxi Mi, Yuge Huang, Jiazhen Ji, Minyi Zhao, Jiaxiang Wu, Xingkun Xu, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using random frequency components. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[34] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.776 + ], + "angle": 0, + "content": "[35] Fatemehsadat Mireshghallah, Mohammadkazem Taram, Prakash Ramrakhyani, Ali Jalali, Dean Tullsen, and Hadi Esmaeilzadeh. Shredder: Learning noise distributions to protect inference privacy. In Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[36] Fatemehsadat Mireshghallah, Mohammadkazem Taram, Ali Jalali, Ahmed Taha Taha Elthakeb, Dean Tullsen, and Hadi Esmaeilzadeh. Not all features are equal: Discovering essential features for preserving prediction privacy. In Proceedings of the Web Conference 2021, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[37] Pratyush Mishra, Ryan Lehmkuhl, Akshayaram Srinivasan, Wenting Zheng, and Raluca Ada Popa. Delphi: A cryptographic inference service for neural networks. In 29th USENIX Security Symposium (USENIX Security 20), 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[38] Hong-Wei Ng and Stefan Winkler. A data-driven approach to cleaning large face datasets. In ICIP, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.178 + ], + "angle": 0, + "content": "[39] Bao-Ngoc Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Man Cheung. Label-only model inversion attacks via knowledge transfer. NeurIPS, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.907, + 0.234 + ], + "angle": 0, + "content": "[40] Ngoc-Bao Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Cheung. Re-thinking model inversion attacks against deep neural networks. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.907, + 0.306 + ], + "angle": 0, + "content": "[41] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.308, + 0.907, + 0.379 + ], + "angle": 0, + "content": "[42] Olga Ohrimenko, Felix Schuster, Cédric Fournet, Aastha Mehta, Sebastian Nowozin, Kapil Vaswani, and Manuel Costa. Oblivious {Multi-Party} machine learning on trusted processors. In 25th USENIX Security Symposium (USENIX Security 16), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.381, + 0.907, + 0.45 + ], + "angle": 0, + "content": "[43] Xiong Peng, Feng Liu, Jingfeng Zhang, Long Lan, Junjie Ye, Tongliang Liu, and Bo Han. Bilateral dependency optimization: Defending against model-inversion attacks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.452, + 0.906, + 0.507 + ], + "angle": 0, + "content": "[44] Yixiang Qiu, Hao Fang, Hongyao Yu, Bin Chen, MeiKang Qiu, and Shu-Tao Xia. A closer look at gan priors: Exploiting intermediate features for enhanced model inversion attacks. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.509, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[45] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.906, + 0.596 + ], + "angle": 0, + "content": "[46] Ahmed Salem, Apratim Bhattacharya, Michael Backes, Mario Fritz, and Yang Zhang. {Updates-Leak}: Data set inference and reconstruction attacks in online learning. In 29th USENIX security symposium (USENIX Security 20), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.598, + 0.906, + 0.639 + ], + "angle": 0, + "content": "[47] Sunandini Sanyal, Sravanti Addepalli, and R Venkatesh Babu. Towards data-free model stealing in a hard label setting. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.641, + 0.906, + 0.683 + ], + "angle": 0, + "content": "[48] Nir Shlezinger, Erez Farhan, Hai Morgenstern, and Yonina C Eldar. Collaborative inference via ensembles on the edge. In ICASSP, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.685, + 0.906, + 0.741 + ], + "angle": 0, + "content": "[49] Lukas Struppek, Dominik Hintersdorf, Antonio De Almeida Correia, Antonia Adler, and Kristian Kersting. Plug & play attacks: Towards robust and flexible model inversion attacks. In ICML, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[50] Chandra Thapa, Pathum Chamikara Mahawaga Arachchige, Seyit Camtepe, and Lichao Sun. Splitfed: When federated learning meets split learning. In AAAI, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[51] Tom Titcombe, Adam J Hall, Pavlos Papadopoulos, and Daniele Romanini. Practical defences against model inversion attacks for split neural networks. arXiv preprint arXiv:2104.05743, 2021. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[52] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8762" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[53] Praneeth Vepakomma, Abhishek Singh, Otkrist Gupta, and Ramesh Raskar. Nopek: Information leakage reduction to share activations in distributed deep learning. In 2020 International Conference on Data Mining Workshops (ICDMW), 2020. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.233 + ], + "angle": 0, + "content": "[54] Sameer Wagh, Shruti Tople, Fabrice Benhamouda, Eyal Kushilevitz, Prateek Mittal, and Tal Rabin. Falcon: Honest-majority maliciously secure framework for private deep learning. Proceedings on Privacy Enhancing Technologies, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[55] Kuan-Chieh Wang, Yan Fu, Ke Li, Ashish Khisti, Richard Zemel, and Alireza Makhzani. Variational model inversion attacks. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[56] Tianhao Wang, Yuheng Zhang, and Ruoxi Jia. Improving robustness to model inversion attacks via mutual information regularization. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[57] Yinggui Wang, Jian Liu, Man Luo, Le Yang, and Li Wang. Privacy-preserving face recognition in the frequency domain. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[58] Zhibo Wang, Mengkai Song, Zhifei Zhang, Yang Song, Qian Wang, and Hairong Qi. Beyond inferring class representatives: User-level privacy leakage from federated learning. In IEEE INFOCOM 2019-IEEE conference on computer communications, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[59] Song Xia, Wenhan Yang, Yi Yu, Xun Lin, Henghui Ding, Lingyu Duan, and Xudong Jiang. Transferable adversarial attacks on sam and its downstream models. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[60] Taihong Xiao, Yi-Hsuan Tsai, Kihyuk Sohn, Manmohan Chandraker, and Ming-Hsuan Yang. Adversarial learning of privacy-preserving and task-oriented representations. In AAAI, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.601 + ], + "angle": 0, + "content": "[61] Mengda Yang, Ziang Li, Juan Wang, Hongxin Hu, Ao Ren, Xiaoyang Xu, and Wenzhe Yi. Measuring data reconstruction defenses in collaborative inference systems. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[62] Ziqi Yang, Jiyi Zhang, Ee-Chien Chang, and Zhenkai Liang. Neural network inversion in adversarial setting via background knowledge alignment. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[63] Ziqi Yang, Lijin Wang, Da Yang, Jie Wan, Ziming Zhao, Ee-Chien Chang, Fan Zhang, and Kui Ren. Purifier: Defending data inference attacks via transforming confidence scores. In AAAI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[64] Yupeng Yin, Xianglong Zhang, Huanle Zhang, Feng Li, Yue Yu, Xiuzhen Cheng, and Pengfei Hu. Ginver: Generative model inversion attacks against collaborative inference. In Proceedings of the ACM Web Conference 2023, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[65] Yi Yu, Song Xia, Xun Lin, Wenhan Yang, Shijian Lu, Yappeng Tan, and Alex Kot. Backdoor attacks against no-reference image quality assessment models via a scalable trigger. In AAAI, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[66] Xiaojian Yuan, Kejiang Chen, Jie Zhang, Weiming Zhang, Nenghai Yu, and Yang Zhang. Pseudo label-guided model inversion attack via conditional generative adversarial network. In AAAI, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[67] Yuheng Zhang, Ruoxi Jia, Hengzhi Pei, Wenxiao Wang, Bo Li, and Dawn Song. The secret revealer: Generative model-inversion attacks against deep neural networks. In CVPR, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[68] Xuejun Zhao, Wencan Zhang, Xiaokui Xiao, and Brian Lim. Exploiting explanations for model inversion attacks. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.908, + 0.248 + ], + "angle": 0, + "content": "[69] Wei Zong, Yang-Wai Chow, Willy Susilo, Joonsang Baek, Jongkil Kim, and Seyit Camtepe. Ipremover: A generative model inversion attack against deep neural network fingerprinting and watermarking. In AAAI, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8763" + } + ] +] \ No newline at end of file diff --git a/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_origin.pdf b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a0078398728c3d75bae89fe6caeb6877efada74 --- /dev/null +++ b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/64dad82b-15cb-48dd-921a-3abe9eba9f91_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99228a9948145c5760ac4ce3166eb66f7f927d976a06eee338560543370d6af2 +size 2658697 diff --git a/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/full.md b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/full.md new file mode 100644 index 0000000000000000000000000000000000000000..57a7c7c720c29a10dba9719fc09bdf9fae37429e --- /dev/null +++ b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/full.md @@ -0,0 +1,323 @@ +# Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems + +Song Xia $^{1}$ , Yi Yu $^{1}$ , Wenhan Yang $^{2*}$ , Meiwen Ding $^{1}$ , Zhuo Chen $^{2}$ , Ling-Yu Duan $^{2,3}$ , Alex C. Kot $^{1}$ , Xudong Jiang $^{1}$ + +$^{1}$ ROSE Lab, Nanyang Technological University, $^{2}$ Pengcheng Laboratory, $^{3}$ Peking University {xias0002,yuyi0010,ding0159,eackot, exdjiang}@ntu.edu.sg, yangwh@pcl.ac.cn, lingyu@pku.edu.cn + +# Abstract + +By locally encoding raw data into intermediate features, collaborative inference enables end users to leverage powerful deep learning models without exposure of sensitive raw data to cloud servers. However, recent studies have revealed that these intermediate features may not sufficiently preserve privacy, as information can be leaked and raw data can be reconstructed via model inversion attacks (MIAs). Obfuscation-based methods, such as noise corruption, adversarial representation learning, and information filters, enhance the inversion robustness by obfuscating the task-irrelevant redundancy empirically. However, methods for quantifying such redundancy remain elusive, and the explicit mathematical relation between this redundancy minimization and inversion robustness enhancement has not yet been established. To address that, this work first theoretically proves that the conditional entropy of inputs given intermediate features provides a guaranteed lower bound on the reconstruction mean square error (MSE) under any MIA. Then, we derive a differentiable and solvable measure for bounding this conditional entropy based on the Gaussian mixture estimation and propose a conditional entropy maximization (CEM) algorithm to enhance the inversion robustness. Experimental results on four datasets demonstrate the effectiveness and adaptability of our proposed CEM; without compromising feature utility and computing efficiency, plugging the proposed CEM into obfuscation-based defense mechanisms consistently boosts their inversion robustness, achieving average gains ranging from $12.9\%$ to $48.2\%$ . Code is available at https://github.com/xiasong0501/CEM. + +# 1. Introduction + +Deep neural networks (DNNs), trained on extensive datasets, have demonstrated outstanding performance across a growing spectrum of complex applications [22, 29, 41]. However, the increasing reliance on deploying these powerful mod + +![](images/419dcb8ae0d9385ca64f1e1271bfdf676d2a7ba9509930e6afbf563485cc5f4e.jpg) + +Figure 1. Privacy protection for collaborative inference via CEM. els on cloud platforms, such as ChatGPT-4 [2], introduces significant privacy and security concerns, as users may upload data containing sensitive information to cloud servers. Collaborative inference [48, 50, 52] offers a solution by partitioning the deep learning model across edge devices and cloud servers, where computations on the initial shallow layers are performed locally on the user's device, and only the extracted intermediate features are transmitted to the cloud for subsequent processing. This allows end users to utilize powerful neural networks with minimal exposure of their raw inputs, and hence enhances data privacy. However, recent works [4, 5, 14, 15, 21, 24, 31, 39, 44, 47, 49, 59, 65, 68, 69] have revealed that those seemingly minor signals in these intermediate features still contain substantial sensitive information. As shown in Figure 1, the MIAs can steal those unprotected features to accurately reconstruct the raw inputs. + +Existing defense against MIAs can be broadly categorized into cryptography-based [23, 37, 42] and obfuscation-based methods [7, 11, 17, 19, 63]. Cryptography-based methods, such as homomorphic encryption [10, 20] and secure multi-party computation [37, 54], provide robust theoretical guarantees against MIAs by computing over encrypted data. However, the inherent computational overhead poses substantial challenges for their scalability on large-scale datasets [36]. Obfuscation-based defense aims to obfuscate the task-irrelevant redundancy by learning a privacy-preserving feature encoder [3, 6, 18, 19, 27, 28, 35, 60]. + +Those approaches primarily rely on empirical heuristics, such as assuming a proxy inversion adversary, to estimate task-irrelevant redundancy during encoder optimization. However, a rigorous quantification for evaluating such redundancy remains absent. Existing works [40, 61] have indicated that such empirical measure is not fully reliable, rendering it insufficient for fully exploring the inversion robustness of the trained feature encoder. Some methods [30, 36, 43, 56] employ information-theoretic frameworks to constrain the redundancy. However, none of them establishes a formal mathematical relationship between information redundancy and robustness against the worst-case inversion adversary, leaving a gap in fully understanding the interplay between redundancy minimization and robustness enhancement. + +This work aims to establish a systematic quantification approach to measure the task-irrelevant yet privacy-critical redundancy within the intermediate features. Furthermore, we endeavor to establish a theoretical relationship between the quantified redundancy and the worst-case model inversion robustness, thereby providing a tractable approach to enhance the inversion robustness of existing models against MIAs. We first demonstrate that the conditional entropy of the input $x$ given intermediate feature $z$ is strongly correlated with the information leakage, which guarantees a theoretical lower bound on the reconstruction MSE between the original and reconstructed inputs under any inversion adversary. Moreover, a differentiable and tractable measure is developed for bounding this conditional entropy based on Gaussian mixture estimation. Utilizing this differentiable measure, we propose a versatile conditional entropy maximization (CEM) algorithm that can be seamlessly plugged into existing empirical obfuscation-based methods, as shown in Figure 1, to consistently enhance their robustness against MIAs. The contributions of our work can be summarized as: + +- We make the first effort in establishing a theoretical relationship between the conditional entropy of inputs given intermediate features and the worst-case MIA robustness. Additionally, we derive a differentiable and tractable measure for quantifying this conditional entropy. +- Building upon these theoretical insights, we propose a versatile CEM algorithm that can be seamlessly plugged into existing obfuscation defense to enhance its effectiveness in defending against MIAs. +- We conduct extensive experiments across four datasets to empirically validate the effectiveness and adaptability of the proposed CEM algorithm. Our findings demonstrate that integrating CEM with existing obfuscation-based defenses consistently yields substantial gains in inversion robustness, without sacrificing feature utility or incurring additional computational overhead. + +# 2. Related Work + +Model Inversion Attacks (MIAs). MIAs represent a serious privacy threat, wherein adversaries reconstruct users' + +private input data by exploiting the information in model parameters or the redundancy present in intermediate outputs. The inversion adversaries can leverage a generative model [39, 40, 55, 64, 66, 67], such as the generative adversarial network (GAN) [12], or a DNN-based decoder [28, 46, 62] to learn the underlying mapping between intermediate features and original inputs, thereby uncovering the hidden inversion patterns. MIAs can be executed under a variety of conditions. Early research on MIAs primarily focuses on the white-box setting where the adversary has full access to the models along with the training data [9, 16, 58, 67]. However, recent work indicated that only by accessing the intermediate features and with little prior knowledge of the data [11, 31, 32, 69], the adversary can launch strong MIAs. + +Obfuscation-based MIAs defense mechanisms. The obfuscation-based defense methods protect the input privacy by obfuscating the redundant information in the intermediate features. These methods typically adopt strategies such as perturbing network weights [1] or intermediate features [18, 35] via noise corruption, purifying intermediate features through frequency domain filtering [33, 34, 57] or sparse coding [6, 19], and training inversion-robust encoders via adversarial representation learning [3, 27, 28, 60]. Those methods generally incur no extra computational overhead during inference, thus serving as a practical and efficient solution for protecting data privacy against MIAs. Although these approaches offer practical effectiveness in defending against MIAs, they primarily measure such redundancy by some empirical heuristics without rigorous quantification [28, 40, 61], leading to a sub-optimal trade-off between feature robustness and utility. Furthermore, a formal mathematical relationship between redundancy and inversion robustness has not been established, leaving a critical gap in comprehensively understanding the interplay between redundancy minimization and robustness enhancement. + +# 3. Methodology + +# 3.1. Preliminaries + +Inversion threats on collaborative inference systems: We consider MIAs on collaborative inference systems, where deep learning models are split into a lightweight encoder $\mathcal{F}_e$ deployed locally and a decoder $\mathcal{F}_d$ deployed in the cloud. The end-users first encode their raw input $\pmb{x}$ into the intermediate features $\pmb{z} = \mathcal{F}_e(\pmb{x})$ locally, and then upload them to the cloud for prediction. This process is known to be susceptible to MIAs, as the intermediate features $\pmb{z}$ is considered as a direct representation of the input [8, 51]. + +The attack model: We consider the scenarios where the inversion adversary $\mathcal{A}$ can steal those user-shared intermediate features $z$ to reconstruct raw inputs. Such threats may arise from the presence of an untrustworthy cloud server or unauthorized access to data traffic during transmission. To evaluate the worst-case inversion robustness, we assume the + +adversary possesses white-box access to the feature encoder $\mathcal{F}_e$ and has full prior knowledge of the input data space $\mathcal{X}$ . The MIA process is formulated as $\hat{\boldsymbol{x}} = \mathcal{A}(\boldsymbol{z},\mathcal{F}_e,\mathcal{X})$ . This information inversion can be achieved by utilizing the DNN-based decoder [13, 28] or the generative models [39, 64, 67] to learn the potential mapping from $\boldsymbol{z}\rightarrow \boldsymbol{x}$ . + +# 3.2. Theoretical Insights on Worst-case Robustness Assume that the information leakage is quantified by the MSE between the original $\mathbf{x}$ and the reconstructed $\hat{\mathbf{x}}$ , i.e., $\| \hat{\mathbf{x}} - \mathbf{x} \|_2^2 / d$ , where $d$ is the input dimensionality. + +Proposition 1 (Minimal reconstruction MSE $\xi$ ). In the worst-case scenarios, where the adversary $\mathcal{A}(z,\mathcal{X},\mathcal{F}_e)$ precisely estimates the posterior probability $\mathbb{P}(\pmb{x}|\pmb{z})$ based on the extensive data prior $\mathcal{X}$ and the white-box access to the feature encoder $\mathcal{F}_{\mathrm{e}}$ , the expectation of the minimal reconstruction MSE $\xi$ over the whole dataset satisfies that: + +$$ +\xi \cdot d = \mathbb {E} _ {\mathcal {Z}} \mathbb {E} _ {\mathcal {X}} \left[ \| \boldsymbol {x} - \mathbb {E} [ \boldsymbol {x} | \boldsymbol {z} ] \| _ {2} ^ {2} | \boldsymbol {z} \right] = \mathbb {E} _ {\mathcal {Z}} \left[ T r (C o v (\boldsymbol {x} | \boldsymbol {z}) \right]. \tag {1} +$$ + +$Cov(\pmb{x}|\pmb{z})$ is the covariance matrix of $\pmb{x}$ conditioned on $\pmb{z}$ , and $Tr$ is the trace operator. Eq. 1 is established based on the fact that the minimized MSE is given by the expectation of the posterior probability of $\pmb{x}$ given $\pmb{z}$ . + +Theorem 1 (Lower bound on the minimal reconstruction MSE $\xi$ ). Let $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ denote the conditional entropy of the input $\boldsymbol{x}$ given the intermediate feature $\boldsymbol{z}$ . The minimal reconstruction MSE $\xi$ is bounded by: $\xi \geq \frac{1}{(2\pi e)}\exp\left(\frac{2\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})}{d}\right)$ . + +Theorem 1 provides a lower bound on the minimal reconstruction MSE $\xi$ in terms of the conditional entropy $\mathcal{H}(\pmb{x}|\pmb{z})$ , which serves as a robust measure for the worst-case robustness against MIAs. The reconstruction MSE $\xi$ and the conditional entropy are highly correlated. Physically, a higher $\mathcal{H}(\pmb{x}|\pmb{z})$ means a more uncertain estimation of $\pmb{x}$ given $\pmb{z}$ , resulting in a larger estimated error $\xi$ . Thus, indicated by Theorem 1, a straightforward way to enhance the inversion robustness is to maximize the $\mathcal{H}(\pmb{x}|\pmb{z})$ during training. The proof of Theorem 1 is in the supplementary materials S.1. + +However, deriving a closed-form expression between $z$ and $x$ is exceedingly difficult due to the inherent intractability of neural networks, especially when applied to large-scale datasets. This renders the direct calculation of $\mathcal{H}(x|z)$ computationally prohibitive, making its maximization via backpropagation to eliminate the information redundancy during training exceedingly intractable. To solve this, a differentiable and computationally efficient lower bound on the conditional entropy $\mathcal{H}(x|z)$ is introduced in the next section to facilitate its maximization during training. + +# 3.3. Differentiable Bound on Conditional Entropy Proposition 2 (The conditional entropy under uncertain encoding). Without loss of the generality, we consider that the encoding process $x \rightarrow z$ consists of two components: + +a deterministic mapping: $\pmb{x} \rightarrow \hat{\pmb{z}}$ and a stochastic process: $z = \hat{z} + \varepsilon$ , where $\varepsilon$ is a random noise. $\mathcal{H}(\pmb{x}|\pmb{z})$ satisfies: + +$$ +\mathcal {H} (\boldsymbol {x} | \boldsymbol {z}) = \mathcal {H} (\boldsymbol {x}) - \mathcal {H} (\boldsymbol {z}) + \mathcal {H} (\boldsymbol {z} | \hat {\boldsymbol {z}}) = \mathcal {H} (\boldsymbol {x}) - \mathcal {I} (\boldsymbol {z}; \hat {\boldsymbol {z}}). \tag {2} +$$ + +Proposition 2 is derived based on the definition of joint entropy, where $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z}) = \mathcal{H}(\boldsymbol{x}) + \mathcal{H}(\boldsymbol{z}|\boldsymbol{x}) - \mathcal{H}(\boldsymbol{z})$ . The equation $\mathcal{H}(z|\boldsymbol{x}) = \mathcal{H}(z|\hat{\boldsymbol{z}})$ holds due to the causal dependency from $\boldsymbol{x} \to \hat{\boldsymbol{z}} \to \boldsymbol{z}$ and $\boldsymbol{x} \to \hat{\boldsymbol{z}}$ is deterministic. Since $\mathcal{H}(\boldsymbol{x})$ remains constant for a given data distribution, the primary challenge in maximizing the conditional entropy $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ indicated by Eq. 2 reduces to deriving a tractable and differentiable measure for $\mathcal{I}(\boldsymbol{z};\hat{\boldsymbol{z}})$ . + +Consider the scenario where the task has $n$ discrete targets $\mathbf{y} = \{y_{1},\dots,y_{n}\}$ . Following previous statements, we consider that the encoding process can be separated into a deterministic mapping: $\pmb{x} \rightarrow \hat{\pmb{z}}$ and a noise corruption process: $\pmb{z} = \hat{\pmb{z}} + \varepsilon$ , where $\varepsilon \sim \mathcal{N}(0,\Sigma_p)$ is a Gaussian noise. We assume that the distribution of the encoded feature $\hat{\pmb{z}}$ can be effectively estimated by a $k$ -component Gaussian mixture distribution denoted as $\hat{\pmb{z}} \sim \sum_{i=1}^{k} \pi_i \mathcal{N}(\mu_i,\Sigma_i)$ , leveraging the proven capability of Gaussian mixtures to closely approximate complex natural data distributions [45]. + +Theorem 2 (Differentiable lower bound on $\mathcal{H}(\pmb{x}|\pmb{z})$ ). Given $\hat{\pmb{z}}$ follows the Gaussian mixture distribution and $\varepsilon$ is a Gaussian noise. The encoded feature $\pmb{z} = \hat{\pmb{z}} + \varepsilon$ also follows a Gaussian mixture distribution with $\pmb{z} \sim \sum_{i=1}^{k} \pi_i \mathcal{N}(\mu_i, \Sigma_i + \Sigma_p)$ . Consequently, the conditional entropy $\mathcal{H}(\pmb{x}|\pmb{z})$ is bounded by: + +$$ +\mathcal {H} (\boldsymbol {x} | \boldsymbol {z}) \geq \mathcal {H} (\boldsymbol {x}) - \sum_ {i = 1} ^ {k} \pi_ {i} \left(- \log \left(\pi_ {i}\right) + \frac {1}{2} \log \left(\frac {\left| \Sigma_ {i} + \Sigma_ {p} \right|}{\left| \Sigma_ {p} \right|}\right)\right). \tag {3} +$$ + +Theorem 2 offers an effective and efficient way to bound the conditional entropy $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ , as the parameters $\pi_{i}$ and $\Sigma_{i}$ are all tractable and differentiable with respect to $\hat{\boldsymbol{z}}$ . Given the extracted $\hat{\boldsymbol{z}}$ after deterministic mapping and the Gaussian noise $\varepsilon$ , one can easily derive the parameters $\pi_{i}$ and $\Sigma_{i}$ using a Gaussian mixture model. This facilitates the maximization of $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ via gradient backpropagation utilizing Eq. 3, thereby improving the worst-case inversion robustness of the trained model. The proof of the Theorem 2 can be found in the supplementary material S.2. + +# 3.4. Practical Insights on the Utility and Robustness Trade-off + +Feature utility: Building on the previous discussion, the encoded feature $\mathbf{z}$ follows a Gaussian mixture distribution with $\mathbf{z} \sim \sum_{i=1}^{k} \pi_i \mathcal{N}(\mu_i, \Sigma_i + \Sigma_p)$ . In general, we assume that each Gaussian component or cluster belongs to one specific target. Ideally, we have $k = n$ to accomplish the task. Due to the causal dependency from $\mathbf{x} \rightarrow \hat{\mathbf{z}} \rightarrow \mathbf{z}$ , the utility of the intermediate feature $\mathbf{z}$ is intrinsically linked to its correlation with preceding states. Thereby, the utility of $\mathbf{z}$ is positively correlated with the expected posterior probability that $\hat{\mathbf{z}}$ belonging to its original cluster $j$ given $\mathbf{z}$ , which is: + +![](images/f325142237c0eb01188cdd6872f31f53202c7197b12a8d414626e2a49b6e5dc5.jpg) +Figure 2. The versatile conditional entropy maximization algorithm for collaborative learning systems with split encoder and decoder. + +$$ +\mathbb {E} _ {\mathcal {Z}} \left[ \mathbb {P} \left(\left(\hat {\boldsymbol {z}} \in j \mid \boldsymbol {z}\right)\right) \right] = \mathbb {E} _ {\mathcal {Z}} \left[ \frac {\pi_ {j} \mathcal {N} \left(\boldsymbol {z} ; \mu_ {j} , \Sigma_ {j} + \Sigma_ {p}\right)}{\sum_ {i = 1} ^ {k} \pi_ {i} \mathcal {N} \left(\boldsymbol {z} ; \mu_ {i} , \Sigma_ {i} + \Sigma_ {p}\right)} \right]. \tag {4} +$$ + +Combining Eq. 3 and Eq. 4, we can get that the utility and robustness of the feature $\mathbf{z}$ are dependent on the $\mathbb{E}_{\mathcal{Z}}[\mathbb{P}((\hat{\mathbf{z}} = j|\mathbf{z}))]$ and the $\mathcal{I}(z;\hat{z})$ respectively. These two items are empirically closely linked, leading to a trade-off between utility and robustness. Additionally, the following practical insights can be derived: + +- Physical interpretation of $\mu_{i}$ and $\Sigma_{i}$ : Physically, the mean $\mu_{i}$ generally captures dominant or representative characteristics of a cluster, while the covariance reflects the degree of variability or noise within the cluster. Consistent with this interpretation, our analysis reveals that $\mu_{i}$ is crucial for evaluating the feature utility. In contrast, the covariance $\Sigma_{i}$ is closely linked to the task-irrelevant redundancy in $\hat{z}$ , which diminishes feature utility and increases susceptibility to MIAs. +- Impact of noise corruption on inversion robustness: The noise corruption introduced by $\Sigma_{p}$ perturbs both the task-important and redundant information, thus balancing the trade-off between utility and robustness. When there is a distinctive difference between the dominant features, such as $\min_{i \neq j} \| \mu_{i} - \mu_{j} \|_{2}^{2} \gg Tr(\Sigma_{p}) \gg Tr(\Sigma_{i} + \Sigma_{j})$ , adding the noise greatly corrupts the redundancy while bringing a small influence on the utility. +- Ideal situation: For the ideal situation, we have $\Sigma_{i} = \Sigma_{p} = 0$ , $k = n$ and, $\pi_{i} = \mathcal{P}(y_{i})$ . Thereby, we can easily get $\mathbb{E}_{\mathcal{Z}}[\mathbb{P}((\hat{z} = j|z))] = 1$ and $\mathcal{H}(\boldsymbol{z}) = \mathcal{H}(\boldsymbol{y})$ , indicating that the feature $\boldsymbol{z}$ contains only the necessary information for the task, without any redundancy. + +However, for a lightweight encoder deployed on edge devices, it is exceedingly challenging to diminish all redundancies while preserving only the essential features. An alternative approach for developing a robust feature encoder is to jointly optimize both the covariance matrices $\Sigma_{i}$ and the distance between the $\mu_{j}$ among different clusters, leading to $\min_{i\neq j}\| \mu_i - \mu_j\| _2^2\gg Tr(\Sigma_p)\gg Tr(\Sigma_i + \Sigma_j)$ . Thus, + +# Algorithm 1 Conditional entropy maximization algorithm + +1: Input: local encoder $\mathcal{F}_e$ , cloud decoder $\mathcal{F}_d$ , image dataset $\mathcal{D}$ of size $N$ , training epochs $m$ , conditional entropy loss $\mathcal{L}_C$ with weight factor $\lambda$ , Gaussian noise $\varepsilon$ , existing defense methods $\mathcal{M}$ with loss $\mathcal{L}_D$ , Gaussian mixture model GMM, number of mixture clusters $k$ . +2: Initialize $\mathcal{F}_e, \mathcal{F}_d$ +3: for epoch from 1 to $m$ do +4: $\pmb{z} \gets \text{concatenate}\{\mathcal{F}_e(\pmb{x}) + \varepsilon, \forall \pmb{x} \in \mathcal{D}\}$ +5: Estimate the distribution of $\mathbf{z}$ as a $k$ component Gaussian mixture: $\sum_{j=1}^{k} \pi_j \mathcal{N}(\mu_j, \Sigma_j + \Sigma_p)$ using GMM( $\mathbf{z}, k$ ) +6: for image batch $(\pmb{x}_i, \pmb{y}_i)$ in the dataset $\mathcal{D}$ do +7: $\pmb{z}_i = \mathcal{F}_e(\pmb{x}_i) + \varepsilon_i$ +8: $\pmb{y}_i' = \mathcal{F}_d(\pmb{z}_i)$ +9: Assign $\mathbf{z}_i$ to the nearest cluster +10: for cluster $j\gets 1$ to $k$ do +11: Update the weight $\pi_{j}$ by Eq. 5 +12: Update the covariance matrix by Eq. 6 +13: Calculate $\mathcal{L}_D$ by $\mathcal{M}(\pmb{x}_i, \pmb{y}_i, \pmb{z}_i, \pmb{y}_i')$ +14: Calculate $\mathcal{L}_C$ by Eq. 7. +15: Optimize $\min_{\mathcal{F}_e,\mathcal{F}_d}\left(\mathcal{L}_D + \lambda *\mathcal{L}_C\right)$ +16: Output: Trained encoder $\mathcal{F}_e$ and decoder $\mathcal{F}_d$ + +incorporating appropriate noise can effectively diminish the residual redundancies, achieving the maximization of the lower bound on $\mathcal{H}(\pmb{x}|\pmb{z})$ while ensuring prediction accuracy. + +# 4. The Versatile Conditional Entropy Maximization Algorithm for Collaborative Inference + +Building upon the derived theoretical insights on the differentiable lower bound of the conditional entropy $\mathcal{H}(\pmb{x}|\pmb{z})$ and the practical insights on the utility and robustness trade-off, this section introduces a versatile conditional entropy maximization (CEM) algorithm. The CEM algorithm is designed to enhance the robustness of existing obfuscation-based defense mechanisms against MIAs in collaborative inference systems. For a given defense mechanism, denoted as $\mathcal{M}$ , the proposed CEM algorithm begins by assessing its worst-case + +robustness. This is achieved through a Gaussian mixture estimation process on the feature $z$ to determine the parameters $\mu_{i}$ and $\Sigma_{i}$ . Subsequently, during each training batch, the algorithm maximizes the lower bound of $\mathcal{H}(x|z)$ specified in Theorem 2 to strengthen robustness against MIAs. The details of the proposed CEM algorithm are presented in Figure 2 and Algorithm 1. + +Key techniques: Let $\mathcal{F}_e$ denote the local feature encoder and $\mathcal{F}_d$ denote the decoder deployed in the cloud. By estimating the distribution of noisy representation $z$ with a $k$ -component Gaussian mixture $\sum_{i=1}^{k} \pi_i \mathcal{N}(\mu_i, \Sigma_i + \Sigma_p)$ , the lower bound of the conditional entropy $\mathcal{H}(x|z)$ can be derived based on Eq. 3, utilizing the covariance matrices $\Sigma_i$ and the weights of the mixture components $\pi_i$ . Since both of them are differentiable and solvable to the feature representation $z$ , this lower bound of the conditional entropy $\mathcal{H}(x|z)$ can be effectively maximized via the gradient backpropagation. As described in Algorithm 1, each training iteration begins by fitting the distribution of $z$ with a Gaussian mixture model (GMM). This involves optimizing the component weights $\pi$ , means $\mu$ , and covariance matrices $\Sigma$ to minimize the estimation error. For each data batch, the extracted feature representations are assigned to the mixture component $j$ corresponding to the nearest mean $\mu_j$ , with the mixture parameters updated accordingly. The component weights $\pi_j$ are updated by: + +$$ +\pi_ {j} = \frac {\pi_ {j} \left(N - N _ {b a t c h}\right) + n _ {j}}{N}, \tag {5} +$$ + +where $N$ is the total length of the dataset and $N_{\text{batch}}$ is the batch size. $n_j$ is the number of feature representations that are assigned to the mixture component $j$ . The covariance matrices are updated by: + +$$ +\Sigma_ {j} = \left(1 - \frac {n _ {j}}{\pi_ {j} N}\right) \Sigma_ {j} + \frac {n _ {j}}{\pi_ {j} N} \Delta \Sigma_ {j}, \tag {6} +$$ + +where $\Delta \Sigma_{j}$ is the calculated covariance based on newly assigned feature representations. The updated parameters are then utilized to compute the conditional entropy loss $\mathcal{L}_C$ , defined as: + +$$ +\mathcal {L} _ {C} = \sum_ {i = j} ^ {k} \pi_ {j} \left(- \log \left(\pi_ {j}\right) + \frac {1}{2} \log \left(\frac {\left| \Sigma_ {j} + \Sigma_ {p} \right|}{\left| \Sigma_ {p} \right|}\right)\right). \tag {7} +$$ + +This loss function is then utilized as an effective indicator of worst-case inversion robustness and is minimized through gradient-based optimization. + +The combination with other defense strategies: As shown in Figure 2, existing obfuscation-based defense methods, such as noise corruption [18, 51], adversarial representation learning [7, 28], and feature purification [6, 19, 53] are collectively denoted as $\mathcal{M}$ , and an auxiliary training loss $\mathcal{L}_D$ is used to encapsulate their impact in the backward optimization process. For approaches that involve architectural modifications, such as introducing information filtering + +![](images/ab3e19b7dbc88aeb220e6675f89f57d829d21d5e527db683046544a4928e26ac.jpg) +Figure 3. Reconstruction MSE Vs. $\mathcal{H}(\pmb{x}|\pmb{z})$ on the CIFAR10. + +layers or adversarial reconstruction models, we adhere to their defined configurations to incorporate our method. The proposed CEM algorithm can be seamlessly plugged into other defense mechanisms by embedding Gaussian mixture estimation within the training process and optimizing the combined loss function: + +$$ +\mathcal {L} = \mathcal {L} _ {D} + \lambda * \mathcal {L} _ {C}, \tag {8} +$$ + +where $\lambda$ is a weight factor to balance the optimization. The utilization of the proposed CEM algorithm allows for an effective evaluation of worst-case inversion robustness of the feature representation throughout training and further optimizes this robustness via gradient backward propagation. + +Conditional entropy $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ and reconstruction MSE by MIAs in practical scenarios: Theorem 1 indicates an exponential relationship between the lower bound on minimal MSE $\xi$ and the conditional entropy $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ . We thereby empirically evaluate this relationship on the CIFAR10 dataset by training multiple models with varying levels of measured $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ and assessing their inversion robustness by assuming a proxy inversion adversary. Following the experimental setting in [28], we utilize a VGG11 model with the first 2 convolutional layers as encoder $\mathcal{F}_e$ and the rest of layers as the decoder $\mathcal{F}_d$ . We fix the $\lambda = 16$ and train different models using different Gaussian noise with variance varying from 0.01 to 0.3, thus leading to varying $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ . The relative conditional entropy $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ is calculated by Eq. 3 (we use 'relative' due to $\mathcal{H}(\boldsymbol{x})$ is constant and its calculation over the whole data space is intractable. So a proper value is selected to replace it). We report the average reconstruction MSE by training a DNN-based MIA decoder $\mathcal{A}(\boldsymbol{z},\mathcal{X},\mathcal{F}_e)$ with the same architecture as [28] and allow its full access to the training data and encoder $\mathcal{F}_e$ . The analytic results are shown in Figure 3, revealing a strong exponential correlation between the $\mathcal{H}(\boldsymbol{x}|\boldsymbol{z})$ and the MIA reconstruction MSE. + +# 5. Experimental Results + +# 5.1. Experimental Settings + +We investigate the inversion robustness of collaborative inference models implemented on three general object classifi- + +Table 1. Comparative results on CIFAR10 dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the VGG11 model before and after the integration with the proposed CEM algorithm. The last 2 rows report the average performance of all methods, comparing results with and without incorporating the CEM algorithm. + +
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense91.860.00130.00140.00140.0016
Bottleneck90.870.00360.00410.00390.0045
Bottleneck+CEM90.690.00540.00580.00760.0080
DistCorr [53]89.520.00740.00820.00790.0088
DistCorr+CEM89.800.00900.00930.00940.0096
Dropout [15]87.750.00980.01040.00990.0111
Dropout+CEM87.530.01290.01340.01320.0142
PATROL [7]89.580.02450.02930.02570.0307
PATROL+CEM89.670.03040.03350.03130.0347
ResSFL [28]89.680.01460.02400.01580.0243
ResSFL+CEM90.110.02010.02510.02290.0273
Noise_Nopeek [51]87.190.01520.01590.01560.0165
Noise_Nopeek+CEM87.080.01740.01740.01730.0178
Noise_ARL [18]87.780.02900.03360.03040.0342
Noise_ARL+CEM87.620.03300.03550.03410.0363
Average w/o CEM88.910.01480.01790.01560.0185
Average w/ CEM88.920.01830.02000.01940.0211
+ +cation datasets: CIFAR-10 [25], CIFAR-100 [25], TinyImageNet [26], and a face recognition dataset FaceScrub [38]. For the CIFAR-10, CIFAR-100, and FaceScrub datasets, we use the VGG11 as the basic model and for the TinyImageNet dataset, we use the ResNet-20 as the basic model. + +Collaborative inference system: To simulate the collaborative inference, we split the VGG11 model by allocating the first two convolutional layers as the local encoder and assigning the remaining layers as the cloud decoder. For the ResNet-20 model, the first four convolutional years are allocated as the local encoder, with the remaining blocks functioning as the cloud-based decoder. This partitioning yields a computational distribution where the encoder is responsible for approximately $10\%$ of the total computation and the decoder handles the remaining $90\%$ . Unless otherwise specified, we maintain the same partitioning scheme across all evaluated methods for a fair comparison. We utilize the Stochastic Gradient Descent (SGD) optimizer to jointly optimize both the encoder and decoder with an initial learning rate of 0.05 and an appropriate learning rate decay. The VGG11 model is trained over 240 epochs and the ResNet20 model is trained over 120 epochs. + +Obfuscation defense methods: The proposed CEM can be flexibly integrated into the model training process to enhance the worst-case inversion robustness by introducing an auxiliary Gaussian mixture estimation to maximize the conditional entropy. We evaluate the effectiveness of the proposed CEM in enhancing several existing defense methods, including the information pruning-based methods called DistCorr [53] and Dropout [15]; adversarial representation learning-based method called ResSFL [28] and + +PATROL [7]; and noise corruption-based method called Noise_Nopeek [51] and Noise_ARL [18]. As observed in [28], the integration of bottleneck layers substantially enhances inversion robustness. To ensure a fair and rigorous comparison, we incorporate a bottleneck layer with 8 channels for both pre- and post-processing of extracted feature representations across all evaluated methods. For PATROL [7] that introduces additional layers within the encoder and maintains the inference efficiency through network pruning techniques, we adjust the pruning ratio to ensure the encoder in PATROL remains the same size as other methods. To analyze the effect of the proposed CEM algorithm, we evaluate the performance of all methods both before and after integrating them with the CEM algorithm. If not otherwise stated, we set hyperparameters $\lambda = 16$ and utilize the isotropic Gaussian noise $\varepsilon$ with a standard deviation of 0.025. We set the number of Gaussian mixture components $k = 3n$ due to the limited number of training data and the imperfect representation ability of the shallow encoder. + +Threat model: To evaluate the worst-case inversion robustness, we consider the white-box scenarios where the MIAs are trained with full access to the collaborative inference model and training dataset. We split the training and testing data strictly following the default setting, e.g. 50,000 images for training and 10,000 for testing on the CIFAR10 dataset. Two types of inversion attacks: the DNN-based decoding method [28] and the generative-based inversion method [67] are utilized as the MIAs. The detailed architecture and training mechanism of those threat models can be found in the supplementary material S.3. + +Evaluation metrics: We report the prediction accuracy and the reconstruction MSE to evaluate the utility and inversion robustness of the intermediate feature. We also report other metrics such as PSNR and SSIM to evaluate the robustness and present the experimental results in the supplementary material S.4. We consider the information leakage on training and inference data, where we report the MSE on reconstructing the training and testing dataset by MIAs. To further explore the intrinsic trade-off between the feature utility and inversion robustness for different defense mechanisms, we present the accuracy-MSE curve by varying the defense hyperparameters, such as the noise strength and CEM loss weight factor $\lambda$ . + +# 5.2. Results on Different Datasets + +Results on CIFAR10: The performance of different defense methods on the CIFAR10 dataset is presented in Table 1. We provide a comparative analysis of each method, both before and after integrating the proposed CEM algorithm. The results indicate that the features, extracted by the lightweight encoder without defense mechanisms, exhibit significant redundancy, which can be exploited by MIAs to reconstruct high-fidelity inputs. Information pruning methods, such + +Table 2. Comparative results on TinyImageNet dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the ResNet-20 model with and without integrating the proposed CEM. + +
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense53.730.00250.00220.00200.0021
Bottleneck52.770.01070.01030.00910.0092
Bottleneck+CEM52.250.01400.01360.01230.0125
DistCorr [53]51.790.01480.01450.01260.0136
DistCorr+CEM51.780.01950.01900.01670.0168
Dropout [15]50.720.01650.01630.01480.0151
Dropout+CEM50.750.01880.01860.01670.0172
PATROL [7]51.750.01870.01870.01680.0176
PATROL+CEM51.640.02110.02090.01850.0189
ResSFL [28]51.990.01730.01720.01570.0161
ResSFL+CEM52.070.01970.01940.01800.0173
Noise_Nopeek [51]51.630.01610.01570.01470.0152
Noise_Nopeek+CEM52.010.01830.01800.01660.0167
Noise_ARL [18]51.030.02290.02240.02040.0205
Noise_ARL+CEM50.850.02810.02710.02510.0253
Average w/o CEM51.660.01670.01640.01480.0153
Average w/ CEM51.620.01990.01950.01770.0178
+ +as DistCorr [53] and dropout [15] that are based on some empirical observations, inevitably remove task-specific information when eliminating the redundancy, leading to a degradation of the prediction accuracy. Methods based on adversarial representation learning (ARL), such as ResSFL, PATROL, and Noise_ARL, offer a better trade-off in feature utility and robustness, achieving substantial robustness gain with a moderate accuracy drop. By incorporating a Gaussian mixture estimation process that is independent of the prediction process, the proposed CEM algorithm effectively increases the lower bound of the reconstruction MSE via maximizing the conditional entropy. This positions the CEM algorithm as a highly adaptable framework for assessing and enhancing the inversion robustness of a wide range of collaborative inference models. The results in Table 1 show that the integration of the CEM algorithm consistently enhances all defense methods, yielding an average increase in reconstruction MSE of $24.0\%$ for training data and $12.9\%$ for inference data, without compromising prediction accuracy. + +Results on TinyImageNet: The performance of different defense methods on the TinyImageNet dataset is presented in Table 2, where we provide a comparative analysis of each method, both before and after integrating the proposed CEM algorithm. The results indicate again that the integration of the CEM algorithm brings substantial inversion robustness gain. Integrating the CEM algorithm yields an average increase in the reconstruction MSE of $19.4\%$ for training data and $17.7\%$ for inference data, without compromising prediction accuracy. Meanwhile, the combination of ARL-based methods such as PATROL and Noise_ARL with CEM achieves the best utility and robustness trade-off bringing a great robustness gain with only a marginal accuracy drop. + +Results on Facescrub: The performance of different de + +Table 3. Comparative results on Facescrub dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the VGG11 model with and without integrating the proposed CEM. + +
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense86.690.00120.00110.00120.0014
Bottleneck85.120.00250.00250.00260.0027
Bottleneck+CEM85.000.00360.00350.00380.0037
DistCorr [53]83.420.00380.00380.00410.0041
DistCorr+CEM83.780.00480.00470.00690.0074
Dropout [15]79.300.00520.00520.00540.0057
Dropout+CEM79.190.00740.00740.00760.0081
PATROL [7]79.180.00990.01180.01140.0137
PATROL+CEM79.880.01660.01850.01840.0192
ResSFL [28]79.600.00940.01120.01110.0142
ResSFL+CEM79.540.01280.01480.01430.0161
Noise_Nopeek [51]82.060.00520.00520.00530.0056
Noise_Nopeek+CEM81.960.00760.00750.00780.0090
Noise_ARL [18]80.140.01220.01550.01320.0158
Noise_ARL+CEM80.330.01820.02110.02120.0231
Average w/o CEM81.260.00690.00790.00760.088
Average w/ CEM81.380.01010.01110.01140.0123
+ +fense methods on the Facescrub dataset is presented in Table 3. Unlike their performance on general object classification datasets, models in this task demonstrate increased vulnerability to MIAs. The evaluated defense methods exhibit a notable accuracy reduction as a trade-off for increased inversion robustness. This may caused by the strong correlation between facial features and identity information. In this dataset, integrating the CEM algorithm with existing defense methods brings a more pronounced improvement in inversion robustness. Specifically, the incorporation of the CEM algorithm yields an average increase in inversion robustness of $48.2\%$ for training data and $40.1\%$ for inference data while maintaining prediction accuracy. Meanwhile, the combination of ARL-based methods also achieves the best utility and robustness trade-off, which leads to a great robustness gain with only a marginal accuracy drop. + +Results on CIFAR100: The results on the CIFAR-100 dataset are provided in the supplementary materials S.5. + +Visualized results: We present the visualized MIA reconstruction on the TinyImageNet dataset in Figure 5. The results illustrate that incorporating the proposed CEM algorithm provides enhanced user input protection, resulting in decreased inverse reconstruction fidelity. + +# 5.3. Ablation Study: Effect of Hyperparameters + +To rigorously evaluate the impact of different hyperparameters, such as the noise strength and CEM weight factor $\lambda$ , on the performance of the CEM algorithm, we conduct experiments under various settings of hyperparameters to demonstrate its robustness, versatility, and adaptability in diverse scenarios. + +The effect of noise strength: As discussed in Subsection 3.4, the intensity of additive Gaussian noise is pivotal in balancing a good trade-off between utility and robust- + +![](images/1c122f6ba0ad8ab67eeaddabbe8cec56cf08d48e38a83b0b2b47ef614e43982f.jpg) +(a) Noise variance varying from 0.01 to 0.5 + +![](images/97704faa14adae42b14152fc3ce6cc7195237730e7d26a01571c738b485b67b3.jpg) +(b) comparing results of different value of $\lambda$ + +![](images/1aee86309c41569810318595932f57a91956d88710e7f9362d74af4eae0e5bf1.jpg) +(c) comparing results of different partitions + +ness. We thereby evaluate the performance of the CEM algorithm under various noise strengths. We utilize the Noise_Nopeek [51] and Noise_ARL [18] as the basic defense strategies $\mathcal{M}$ . In Figure 4a, we present the prediction accuracy Vs. reconstruction MSE curves on the CIFAR-10 dataset by varying the noise variance from 0.001 to 0.5. Multiple models were trained to analyze the impact comprehensively. The results demonstrate that the noise variance significantly influences the trade-off between feature utility and robustness. The proposed CEM algorithm exhibits robust performance across the whole range of noise intensities. Integrating the CEM algorithm consistently achieves a more favorable balance between utility and robustness. + +The effect of $\lambda$ : In Equation 8, the weight factor $\lambda$ is important in adjusting the conditional entropy maximization during the joint optimization. To investigate the influence of $\lambda$ , we use the Noise_ARL method [18], which achieves a comparatively high performance as our baseline strategy $\mathcal{M}$ . We then evaluate its performance when integrated with the CEM algorithm, using $\lambda$ values of 0, 2, 8, 16, and 32. We present the prediction accuracy and reconstruction MSE curve in Figure 4b, by varying the noise variance from 0.001 to 0.5. The results demonstrate that increasing $\lambda$ from 0 to 16 markedly enhances the robustness and utility trade-off. + +Different partitioning mechanisms: Encoders with more layers extract more task-specific features from input data, thereby providing improved robustness against MIAs. To illustrate the effectiveness of the proposed CEM under different levels of redundancy, we evaluate its performance across multiple partitioning schemes. Specifically, we use the VGG11 architecture as the backbone, partitioned at the 1st, 2nd, and 3rd convolutional layers (denoted as $c_{1}$ , $c_{2}$ , and $c_{3}$ ), and assess the performance of CEM in combination with Noise_ARL [18]. We present the prediction accuracy Vs. reconstruction MSE curve in Figure 4c using the same strategy. The results demonstrate a significant performance advantage with deeper partitioning: specifically, partitioning at the third convolutional layer achieves substantial inversion robustness while maintaining prediction accuracy without degradation. Furthermore, the CEM algorithm exhibits high + +![](images/a9f407e4ca8090589dde4511f48ba7269e5a1746ed78ca576258213cd945020e.jpg) +Figure 4. The effect of different hyperparameters on the performance of the proposed CEM algorithm. The effect of noise strength, $\lambda$ , and partitioning schemes are demonstrated by the utility-robustness trade-off curve. +Figure 5. The visualized result. The last three rows are inputs reconstructed by MIA. + +adaptability, consistently enhancing the robustness-utility trade-off across different partitioning schemes. + +Analysis of the efficiency: The computational complexity on the local encoder and the cloud server across various methods is presented in the supplementary material S.6. The results confirm that integrating the proposed CEM does not introduce any additional computational overhead or latency. + +# 6. Conclusion + +This work addresses problems of inversion robustness in collaborative inference systems, where MIAs can exploit subtle signals within these intermediate features to reconstruct high-fidelity inputs. Existing obfuscation defenses seek to protect privacy by empirically eliminating feature redundancy. However, the precise quantification of such redundancy is lacking, and rigorous mathematical analysis is needed to elucidate the relation between redundancy minimization and enhanced inversion robustness. To solve that, we prove that the conditional entropy of inputs given intermediate features provides a lower bound on the reconstruction MSE. Based on that, we derive a differentiable lower bound on this conditional entropy using Gaussian mixture estimation, making it amenable for efficient optimization through backpropagation. Building on this, we propose a versatile conditional entropy maximization (CEM) algorithm that can be easily plugged into existing methods. Comprehensive Experiments demonstrate that the proposed CEM significantly and consistently boosts robustness while maintaining feature utility and computational efficiency. + +# Acknowledgement + +This work was carried out at the Rapid-Rich Object Search (ROSE) Lab, School of Electrical & Electronic Engineering, Nanyang Technological University (NTU), Singapore. This research is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative, the Basic and Frontier Research Project of PCL, the Major Key Project of PCL, Guangdong Basic and Applied Basic Research Foundation under Grant 2024A1515010454, the Program of Beijing Municipal Science and Technology Commission Foundation (No.Z241100003524010), and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore and Infocomm Media Development Authority. + +# References + +[1] Sharif Abuadbba, Kyuyeon Kim, Minki Kim, Chandra Thapa, Seyit A Camtepe, Yansong Gao, Hyoungshick Kim, and Surya Nepal. Can we use split learning on 1d cnn models for privacy preserving training? In Proceedings of the 15th ACM Asia conference on computer and communications security, 2020. 2 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[3] Martin Bertran, Natalia Martinez, Afroditi Papadaki, Qiang Qiu, Miguel Rodrigues, Galen Reeves, and Guillermo Sapiro. Adversarily learned representations for information obfuscation and inference. In ICML, 2019. 1, 2 +[4] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), 2023. 1 +[5] Sayanton V Dibbo. Sok: Model inversion attack landscape: Taxonomy, challenges, and future roadmap. In 2023 IEEE 36th Computer Security Foundations Symposium (CSF), 2023. 1 +[6] Sayanton V Dibbo, Adam Breuer, Juston Moore, and Michael Teti. Improving robustness to model inversion attacks via sparse coding architectures. ECCV, 2024. 1, 2, 5 +[7] Shiwei Ding, Lan Zhang, Miao Pan, and Xiaoyong Yuan. Patrol: Privacy-oriented pruning for collaborative inference against model inversion attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 5, 6, 7 +[8] Hao Fang, Bin Chen, Xuan Wang, Zhi Wang, and Shu-Tao Xia. Gidf: A generative gradient inversion method with feature domain optimization. In ICCV, 2023. 2 + +[9] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, 2015. +[10] Craig Gentry and Shai Halevi. Implementing gentry's fully-homomorphic encryption scheme. In Annual international conference on the theory and applications of cryptographic techniques, 2011. 1 +[11] Xueluan Gong, Ziyao Wang, Shuaike Li, Yanjiao Chen, and Qian Wang. A gan-based defense framework against model inversion attacks. IEEE Transactions on Information Forensics and Security, 2023. 1, 2 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2 +[13] Gyojin Han, Jaehyun Choi, Haeil Lee, and Junmo Kim. Reinforcement learning-based black-box model inversion attacks. In ICCV, 2023. 3 +[14] Zecheng He, Tianwei Zhang, and Ruby B Lee. Model inversion attacks against collaborative inference. In Proceedings of the 35th Annual Computer Security Applications Conference, 2019. 1 +[15] Zecheng He, Tianwei Zhang, and Ruby B Lee. Attacking and protecting data privacy in edge-cloud collaborative inference systems. IEEE Internet of Things Journal, 2020. 1, 6, 7 +[16] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, 2017. 2 +[17] Sy-Tuyen Ho, Koh Jun Hao, Keshigeyan Chandrasegaran, Ngoc-Bao Nguyen, and Ngai-Man Cheung. Model inversion robustness: Can transfer learning help? In CVPR, 2024. 1 +[18] Jonghu Jeong, Minyong Cho, Philipp Benz, and Tae-hoon Kim. Noisy adversarial representation learning for effective and efficient image obfuscation. In Uncertainty in Artificial Intelligence, 2023. 1, 2, 5, 6, 7, 8 +[19] Shuaifan Jin, He Wang, Zhibo Wang, Feng Xiao, Jiahui Hu, Yuan He, Wenwen Zhang, Zhongjie Ba, Weijie Fang, Shuhong Yuan, et al. {FaceObfuscator}: Defending deep learning-based privacy attacks with gradient descent-resistant features in face recognition. In 33rd USENIX Security Symposium (USENIX Security 24), 2024. 1, 2, 5 +[20] Chiraag Juvekar, Vinod Vaikuntanathan, and Anantha Chandrakasan. {GAZELLE}: A low latency framework for secure neural network inference. In 27th USENIX security symposium (USENIX security 18), 2018. 1 +[21] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1 +[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1 +[23] Brian Knott, Shobha Venkataraman, Awni Hannun, Shubho Sengupta, Mark Ibrahim, and Laurens van der Maaten. + +Cryptography: Secure multi-party computation meets machine learning. NeurIPS, 2021. 1 +[24] Chenqi Kong, Anwei Luo, Shiqi Wang, Haoliang Li, Anderson Rocha, and Alex C Kot. Pixel-inconsistency modeling for image manipulation localization. IEEE TPAMI, 2025. 1 +[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. 2015. 6 +[27] Ang Li, Jiayi Guo, Huanrui Yang, and Yiran Chen. Deepobfuscator: Adversarial training framework for privacy-preserving image classification. arXiv preprint arXiv:1909.04126, 2019. 1, 2 +[28] Jingtao Li, Adnan Siraj Rakin, Xing Chen, Zhezhi He, Deliang Fan, and Chaitali Chakrabarti. Ressfl: A resistance transfer framework for defending model inversion attack in split federated learning. In CVPR, 2022. 1, 2, 3, 5, 6, 7 +[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1 +[30] Kiwan Maeng, Chuan Guo, Sanjay Kariyappa, and G Edward Suh. Bounding the invertibility of privacy-preserving instance encoding using fisher information. NeurIPS, 2024. 2 +[31] Shagufta Mehnaz, Sayanton V Dibbo, Roberta De Viti, Ehsanul Kabir, Björn B Brandenburg, Stefan Mangard, Ninghui Li, Elisa Bertino, Michael Backes, Emiliano De Cristofaro, et al. Are your sensitive attributes private? novel model inversion attribute inference attacks on classification models. In 31st USENIX Security Symposium (USENIX Security 22), 2022. 1, 2 +[32] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In 2019 IEEE symposium on security and privacy (SP), 2019. 2 +[33] Yuxi Mi, Yuge Huang, Jiazhen Ji, Minyi Zhao, Jiaxiang Wu, Xingkun Xu, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using random frequency components. In ICCV, 2023. 2 +[34] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In CVPR, 2024. 2 +[35] Fatemehsadat Mireshghallah, Mohammadkazem Taram, Prakash Ramrakhyani, Ali Jalali, Dean Tullsen, and Hadi Esmaeilzadeh. Shredder: Learning noise distributions to protect inference privacy. In Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, 2020. 1, 2 +[36] Fatemehsadat Mireshghallah, Mohammadkazem Taram, Ali Jalali, Ahmed Taha Taha Elthakeb, Dean Tullsen, and Hadi Esmaeilzadeh. Not all features are equal: Discovering essential features for preserving prediction privacy. In Proceedings of the Web Conference 2021, 2021. 1, 2 +[37] Pratyush Mishra, Ryan Lehmkuhl, Akshayaram Srinivasan, Wenting Zheng, and Raluca Ada Popa. Delphi: A cryptographic inference service for neural networks. In 29th USENIX Security Symposium (USENIX Security 20), 2020. 1 + +[38] Hong-Wei Ng and Stefan Winkler. A data-driven approach to cleaning large face datasets. In ICIP, 2014. 6 +[39] Bao-Ngoc Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Man Cheung. Label-only model inversion attacks via knowledge transfer. NeurIPS, 2024. 1, 2, 3 +[40] Ngoc-Bao Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Cheung. Re-thinking model inversion attacks against deep neural networks. In CVPR, 2023. 2 +[41] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 1 +[42] Olga Ohrimenko, Felix Schuster, Cédric Fournet, Aastha Mehta, Sebastian Nowozin, Kapil Vaswani, and Manuel Costa. Oblivious {Multi-Party} machine learning on trusted processors. In 25th USENIX Security Symposium (USENIX Security 16), 2016. 1 +[43] Xiong Peng, Feng Liu, Jingfeng Zhang, Long Lan, Junjie Ye, Tongliang Liu, and Bo Han. Bilateral dependency optimization: Defending against model-inversion attacks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022. 2 +[44] Yixiang Qiu, Hao Fang, Hongyao Yu, Bin Chen, MeiKang Qiu, and Shu-Tao Xia. A closer look at gan priors: Exploiting intermediate features for enhanced model inversion attacks. In ECCV, 2024. 1 +[45] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 2009. 3 +[46] Ahmed Salem, Apratim Bhattacharya, Michael Backes, Mario Fritz, and Yang Zhang. {Updates-Leak}: Data set inference and reconstruction attacks in online learning. In 29th USENIX security symposium (USENIX Security 20), 2020. 2 +[47] Sunandini Sanyal, Sravanti Addepalli, and R Venkatesh Babu. Towards data-free model stealing in a hard label setting. In CVPR, 2022. 1 +[48] Nir Shlezinger, Erez Farhan, Hai Morgenstern, and Yonina C Eldar. Collaborative inference via ensembles on the edge. In ICASSP, 2021. 1 +[49] Lukas Struppek, Dominik Hintersdorf, Antonio De Almeida Correia, Antonia Adler, and Kristian Kersting. Plug & play attacks: Towards robust and flexible model inversion attacks. In ICML, 2022. 1 +[50] Chandra Thapa, Pathum Chamikara Mahawaga Arachchige, Seyit Camtepe, and Lichao Sun. Splitfed: When federated learning meets split learning. In AAAI, 2022. 1 +[51] Tom Titcombe, Adam J Hall, Pavlos Papadopoulos, and Daniele Romanini. Practical defences against model inversion attacks for split neural networks. arXiv preprint arXiv:2104.05743, 2021. 2, 5, 6, 7, 8 +[52] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1 + +[53] Praneeth Vepakomma, Abhishek Singh, Otkrist Gupta, and Ramesh Raskar. Nopek: Information leakage reduction to share activations in distributed deep learning. In 2020 International Conference on Data Mining Workshops (ICDMW), 2020. 5, 6, 7 +[54] Sameer Wagh, Shruti Tople, Fabrice Benhamouda, Eyal Kushilevitz, Prateek Mittal, and Tal Rabin. Falcon: Honest-majority maliciously secure framework for private deep learning. Proceedings on Privacy Enhancing Technologies, 2021. 1 +[55] Kuan-Chieh Wang, Yan Fu, Ke Li, Ashish Khisti, Richard Zemel, and Alireza Makhzani. Variational model inversion attacks. NeurIPS, 2021. 2 +[56] Tianhao Wang, Yuheng Zhang, and Ruoxi Jia. Improving robustness to model inversion attacks via mutual information regularization. In AAAI, 2021. 2 +[57] Yinggui Wang, Jian Liu, Man Luo, Le Yang, and Li Wang. Privacy-preserving face recognition in the frequency domain. In AAAI, 2022. 2 +[58] Zhibo Wang, Mengkai Song, Zhifei Zhang, Yang Song, Qian Wang, and Hairong Qi. Beyond inferring class representatives: User-level privacy leakage from federated learning. In IEEE INFOCOM 2019-IEEE conference on computer communications, 2019. 2 +[59] Song Xia, Wenhan Yang, Yi Yu, Xun Lin, Henghui Ding, Lingyu Duan, and Xudong Jiang. Transferable adversarial attacks on sam and its downstream models. In NeurIPS, 2024. 1 +[60] Taihong Xiao, Yi-Hsuan Tsai, Kihyuk Sohn, Manmohan Chandraker, and Ming-Hsuan Yang. Adversarial learning of privacy-preserving and task-oriented representations. In AAAI, 2020. 1, 2 +[61] Mengda Yang, Ziang Li, Juan Wang, Hongxin Hu, Ao Ren, Xiaoyang Xu, and Wenzhe Yi. Measuring data reconstruction defenses in collaborative inference systems. NeurIPS, 2022. 2 +[62] Ziqi Yang, Jiyi Zhang, Ee-Chien Chang, and Zhenkai Liang. Neural network inversion in adversarial setting via background knowledge alignment. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 2019. 2 +[63] Ziqi Yang, Lijin Wang, Da Yang, Jie Wan, Ziming Zhao, Ee-Chien Chang, Fan Zhang, and Kui Ren. Purifier: Defending data inference attacks via transforming confidence scores. In AAAI, 2023. 1 +[64] Yupeng Yin, Xianglong Zhang, Huanle Zhang, Feng Li, Yue Yu, Xiuzhen Cheng, and Pengfei Hu. Ginver: Generative model inversion attacks against collaborative inference. In Proceedings of the ACM Web Conference 2023, 2023. 2, 3 +[65] Yi Yu, Song Xia, Xun Lin, Wenhan Yang, Shijian Lu, Yappeng Tan, and Alex Kot. Backdoor attacks against no-reference image quality assessment models via a scalable trigger. In AAAI, 2025. 1 +[66] Xiaojian Yuan, Kejiang Chen, Jie Zhang, Weiming Zhang, Nenghai Yu, and Yang Zhang. Pseudo label-guided model inversion attack via conditional generative adversarial network. In AAAI, 2023. 2 + +[67] Yuheng Zhang, Ruoxi Jia, Hengzhi Pei, Wenxiao Wang, Bo Li, and Dawn Song. The secret revealer: Generative model-inversion attacks against deep neural networks. In CVPR, 2020. 2, 3, 6, 7 +[68] Xuejun Zhao, Wencan Zhang, Xiaokui Xiao, and Brian Lim. Exploiting explanations for model inversion attacks. In ICCV, 2021. 1 +[69] Wei Zong, Yang-Wai Chow, Willy Susilo, Joonsang Baek, Jongkil Kim, and Seyit Camtepe. Ipremover: A generative model inversion attack against deep neural network fingerprinting and watermarking. In AAAI, 2024. 1, 2 \ No newline at end of file diff --git a/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/images.zip b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5ce56da39338024cd10b1d14705e26532f936e34 --- /dev/null +++ b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a9fa56067e98f02df20cb063ae9bb3afb2f65373d653ac90b3952721a1398ca +size 488294 diff --git a/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/layout.json b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a1eb0580907a5530c933141941400793229b791e --- /dev/null +++ b/2025/Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems/layout.json @@ -0,0 +1,11863 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 542, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 542, + 140 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 542, + 140 + ], + "type": "text", + "content": "Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "text", + "content": "Song Xia" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Yi Yu" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Wenhan Yang" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Meiwen Ding" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Zhuo Chen" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Ling-Yu Duan" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Alex C. Kot" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "text", + "content": ", Xudong Jiang" + }, + { + "bbox": [ + 153, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 191, + 555, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 191, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 63, + 191, + 555, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 191, + 555, + 220 + ], + "type": "text", + "content": "ROSE Lab, Nanyang Technological University, " + }, + { + "bbox": [ + 63, + 191, + 555, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 191, + 555, + 220 + ], + "type": "text", + "content": "Pengcheng Laboratory, " + }, + { + "bbox": [ + 63, + 191, + 555, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 191, + 555, + 220 + ], + "type": "text", + "content": "Peking University {xias0002,yuyi0010,ding0159,eackot, exdjiang}@ntu.edu.sg, yangwh@pcl.ac.cn, lingyu@pku.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "spans": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 273, + 297, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 297, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 297, + 620 + ], + "type": "text", + "content": "By locally encoding raw data into intermediate features, collaborative inference enables end users to leverage powerful deep learning models without exposure of sensitive raw data to cloud servers. However, recent studies have revealed that these intermediate features may not sufficiently preserve privacy, as information can be leaked and raw data can be reconstructed via model inversion attacks (MIAs). Obfuscation-based methods, such as noise corruption, adversarial representation learning, and information filters, enhance the inversion robustness by obfuscating the task-irrelevant redundancy empirically. However, methods for quantifying such redundancy remain elusive, and the explicit mathematical relation between this redundancy minimization and inversion robustness enhancement has not yet been established. To address that, this work first theoretically proves that the conditional entropy of inputs given intermediate features provides a guaranteed lower bound on the reconstruction mean square error (MSE) under any MIA. Then, we derive a differentiable and solvable measure for bounding this conditional entropy based on the Gaussian mixture estimation and propose a conditional entropy maximization (CEM) algorithm to enhance the inversion robustness. Experimental results on four datasets demonstrate the effectiveness and adaptability of our proposed CEM; without compromising feature utility and computing efficiency, plugging the proposed CEM into obfuscation-based defense mechanisms consistently boosts their inversion robustness, achieving average gains ranging from " + }, + { + "bbox": [ + 55, + 273, + 297, + 620 + ], + "type": "inline_equation", + "content": "12.9\\%" + }, + { + "bbox": [ + 55, + 273, + 297, + 620 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 273, + 297, + 620 + ], + "type": "inline_equation", + "content": "48.2\\%" + }, + { + "bbox": [ + 55, + 273, + 297, + 620 + ], + "type": "text", + "content": ". Code is available at https://github.com/xiasong0501/CEM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 632, + 137, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 137, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 137, + 643 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 297, + 696 + ], + "type": "text", + "content": "Deep neural networks (DNNs), trained on extensive datasets, have demonstrated outstanding performance across a growing spectrum of complex applications [22, 29, 41]. However, the increasing reliance on deploying these powerful mod" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 247, + 542, + 373 + ], + "blocks": [ + { + "bbox": [ + 328, + 247, + 542, + 373 + ], + "lines": [ + { + "bbox": [ + 328, + 247, + 542, + 373 + ], + "spans": [ + { + "bbox": [ + 328, + 247, + 542, + 373 + ], + "type": "image", + "image_path": "419dcb8ae0d9385ca64f1e1271bfdf676d2a7ba9509930e6afbf563485cc5f4e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 582 + ], + "type": "text", + "content": "Figure 1. Privacy protection for collaborative inference via CEM. els on cloud platforms, such as ChatGPT-4 [2], introduces significant privacy and security concerns, as users may upload data containing sensitive information to cloud servers. Collaborative inference [48, 50, 52] offers a solution by partitioning the deep learning model across edge devices and cloud servers, where computations on the initial shallow layers are performed locally on the user's device, and only the extracted intermediate features are transmitted to the cloud for subsequent processing. This allows end users to utilize powerful neural networks with minimal exposure of their raw inputs, and hence enhances data privacy. However, recent works [4, 5, 14, 15, 21, 24, 31, 39, 44, 47, 49, 59, 65, 68, 69] have revealed that those seemingly minor signals in these intermediate features still contain substantial sensitive information. As shown in Figure 1, the MIAs can steal those unprotected features to accurately reconstruct the raw inputs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 557, + 715 + ], + "type": "text", + "content": "Existing defense against MIAs can be broadly categorized into cryptography-based [23, 37, 42] and obfuscation-based methods [7, 11, 17, 19, 63]. Cryptography-based methods, such as homomorphic encryption [10, 20] and secure multi-party computation [37, 54], provide robust theoretical guarantees against MIAs by computing over encrypted data. However, the inherent computational overhead poses substantial challenges for their scalability on large-scale datasets [36]. Obfuscation-based defense aims to obfuscate the task-irrelevant redundancy by learning a privacy-preserving feature encoder [3, 6, 18, 19, 27, 28, 35, 60]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8753" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "type": "text", + "content": "Those approaches primarily rely on empirical heuristics, such as assuming a proxy inversion adversary, to estimate task-irrelevant redundancy during encoder optimization. However, a rigorous quantification for evaluating such redundancy remains absent. Existing works [40, 61] have indicated that such empirical measure is not fully reliable, rendering it insufficient for fully exploring the inversion robustness of the trained feature encoder. Some methods [30, 36, 43, 56] employ information-theoretic frameworks to constrain the redundancy. However, none of them establishes a formal mathematical relationship between information redundancy and robustness against the worst-case inversion adversary, leaving a gap in fully understanding the interplay between redundancy minimization and robustness enhancement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 240, + 296, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 296, + 478 + ], + "type": "text", + "content": "This work aims to establish a systematic quantification approach to measure the task-irrelevant yet privacy-critical redundancy within the intermediate features. Furthermore, we endeavor to establish a theoretical relationship between the quantified redundancy and the worst-case model inversion robustness, thereby providing a tractable approach to enhance the inversion robustness of existing models against MIAs. We first demonstrate that the conditional entropy of the input " + }, + { + "bbox": [ + 56, + 240, + 296, + 478 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 56, + 240, + 296, + 478 + ], + "type": "text", + "content": " given intermediate feature " + }, + { + "bbox": [ + 56, + 240, + 296, + 478 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 56, + 240, + 296, + 478 + ], + "type": "text", + "content": " is strongly correlated with the information leakage, which guarantees a theoretical lower bound on the reconstruction MSE between the original and reconstructed inputs under any inversion adversary. Moreover, a differentiable and tractable measure is developed for bounding this conditional entropy based on Gaussian mixture estimation. Utilizing this differentiable measure, we propose a versatile conditional entropy maximization (CEM) algorithm that can be seamlessly plugged into existing empirical obfuscation-based methods, as shown in Figure 1, to consistently enhance their robustness against MIAs. The contributions of our work can be summarized as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 479, + 295, + 669 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 479, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 538 + ], + "type": "text", + "content": "- We make the first effort in establishing a theoretical relationship between the conditional entropy of inputs given intermediate features and the worst-case MIA robustness. Additionally, we derive a differentiable and tractable measure for quantifying this conditional entropy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 539, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 539, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 539, + 295, + 586 + ], + "type": "text", + "content": "- Building upon these theoretical insights, we propose a versatile CEM algorithm that can be seamlessly plugged into existing obfuscation defense to enhance its effectiveness in defending against MIAs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 586, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 586, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 586, + 295, + 669 + ], + "type": "text", + "content": "- We conduct extensive experiments across four datasets to empirically validate the effectiveness and adaptability of the proposed CEM algorithm. Our findings demonstrate that integrating CEM with existing obfuscation-based defenses consistently yields substantial gains in inversion robustness, without sacrificing feature utility or incurring additional computational overhead." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 675, + 142, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 142, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 142, + 687 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Model Inversion Attacks (MIAs). MIAs represent a serious privacy threat, wherein adversaries reconstruct users'" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "text", + "content": "private input data by exploiting the information in model parameters or the redundancy present in intermediate outputs. The inversion adversaries can leverage a generative model [39, 40, 55, 64, 66, 67], such as the generative adversarial network (GAN) [12], or a DNN-based decoder [28, 46, 62] to learn the underlying mapping between intermediate features and original inputs, thereby uncovering the hidden inversion patterns. MIAs can be executed under a variety of conditions. Early research on MIAs primarily focuses on the white-box setting where the adversary has full access to the models along with the training data [9, 16, 58, 67]. However, recent work indicated that only by accessing the intermediate features and with little prior knowledge of the data [11, 31, 32, 69], the adversary can launch strong MIAs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 556, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 556, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 556, + 502 + ], + "type": "text", + "content": "Obfuscation-based MIAs defense mechanisms. The obfuscation-based defense methods protect the input privacy by obfuscating the redundant information in the intermediate features. These methods typically adopt strategies such as perturbing network weights [1] or intermediate features [18, 35] via noise corruption, purifying intermediate features through frequency domain filtering [33, 34, 57] or sparse coding [6, 19], and training inversion-robust encoders via adversarial representation learning [3, 27, 28, 60]. Those methods generally incur no extra computational overhead during inference, thus serving as a practical and efficient solution for protecting data privacy against MIAs. Although these approaches offer practical effectiveness in defending against MIAs, they primarily measure such redundancy by some empirical heuristics without rigorous quantification [28, 40, 61], leading to a sub-optimal trade-off between feature robustness and utility. Furthermore, a formal mathematical relationship between redundancy and inversion robustness has not been established, leaving a critical gap in comprehensively understanding the interplay between redundancy minimization and robustness enhancement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 507, + 397, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 397, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 397, + 520 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 400, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 400, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 400, + 533 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "text", + "content": "Inversion threats on collaborative inference systems: We consider MIAs on collaborative inference systems, where deep learning models are split into a lightweight encoder " + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e" + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "text", + "content": " deployed locally and a decoder " + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d" + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "text", + "content": " deployed in the cloud. The end-users first encode their raw input " + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "text", + "content": " into the intermediate features " + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{z} = \\mathcal{F}_e(\\pmb{x})" + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "text", + "content": " locally, and then upload them to the cloud for prediction. This process is known to be susceptible to MIAs, as the intermediate features " + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 313, + 534, + 555, + 641 + ], + "type": "text", + "content": " is considered as a direct representation of the input [8, 51]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": "The attack model: We consider the scenarios where the inversion adversary " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " can steal those user-shared intermediate features " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " to reconstruct raw inputs. Such threats may arise from the presence of an untrustworthy cloud server or unauthorized access to data traffic during transmission. To evaluate the worst-case inversion robustness, we assume the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8754" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "adversary possesses white-box access to the feature encoder " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " and has full prior knowledge of the input data space " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ". The MIA process is formulated as " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{x}} = \\mathcal{A}(\\boldsymbol{z},\\mathcal{F}_e,\\mathcal{X})" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ". This information inversion can be achieved by utilizing the DNN-based decoder [13, 28] or the generative models [39, 64, 67] to learn the potential mapping from " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}\\rightarrow \\boldsymbol{x}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "text", + "content": "3.2. Theoretical Insights on Worst-case Robustness Assume that the information leakage is quantified by the MSE between the original " + }, + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "text", + "content": " and the reconstructed " + }, + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\| \\hat{\\mathbf{x}} - \\mathbf{x} \\|_2^2 / d" + }, + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 152, + 296, + 204 + ], + "type": "text", + "content": " is the input dimensionality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "text", + "content": "Proposition 1 (Minimal reconstruction MSE " + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "text", + "content": "). In the worst-case scenarios, where the adversary " + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(z,\\mathcal{X},\\mathcal{F}_e)" + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "text", + "content": " precisely estimates the posterior probability " + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "text", + "content": " based on the extensive data prior " + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "text", + "content": " and the white-box access to the feature encoder " + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{e}}" + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "text", + "content": ", the expectation of the minimal reconstruction MSE " + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 54, + 208, + 296, + 280 + ], + "type": "text", + "content": " over the whole dataset satisfies that:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 282, + 296, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 282, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 296, + 297 + ], + "type": "interline_equation", + "content": "\\xi \\cdot d = \\mathbb {E} _ {\\mathcal {Z}} \\mathbb {E} _ {\\mathcal {X}} \\left[ \\| \\boldsymbol {x} - \\mathbb {E} [ \\boldsymbol {x} | \\boldsymbol {z} ] \\| _ {2} ^ {2} | \\boldsymbol {z} \\right] = \\mathbb {E} _ {\\mathcal {Z}} \\left[ T r (C o v (\\boldsymbol {x} | \\boldsymbol {z}) \\right]. \\tag {1}", + "image_path": "03d2aca7bdeebf64595984f8ac48458920be733171617ceb6275a4cc2d78d1ec.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "inline_equation", + "content": "Cov(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "text", + "content": " is the covariance matrix of " + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "inline_equation", + "content": "Tr" + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "text", + "content": " is the trace operator. Eq. 1 is established based on the fact that the minimized MSE is given by the expectation of the posterior probability of " + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "text", + "content": "Theorem 1 (Lower bound on the minimal reconstruction MSE " + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "text", + "content": "). Let " + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "text", + "content": " denote the conditional entropy of the input " + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "text", + "content": " given the intermediate feature " + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}" + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "text", + "content": ". The minimal reconstruction MSE " + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "text", + "content": " is bounded by: " + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\xi \\geq \\frac{1}{(2\\pi e)}\\exp\\left(\\frac{2\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})}{d}\\right)" + }, + { + "bbox": [ + 55, + 360, + 296, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": "Theorem 1 provides a lower bound on the minimal reconstruction MSE " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " in terms of the conditional entropy " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": ", which serves as a robust measure for the worst-case robustness against MIAs. The reconstruction MSE " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " and the conditional entropy are highly correlated. Physically, a higher " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " means a more uncertain estimation of " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": ", resulting in a larger estimated error " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": ". Thus, indicated by Theorem 1, a straightforward way to enhance the inversion robustness is to maximize the " + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 55, + 415, + 296, + 536 + ], + "type": "text", + "content": " during training. The proof of Theorem 1 is in the supplementary materials S.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "text", + "content": "However, deriving a closed-form expression between " + }, + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "text", + "content": " is exceedingly difficult due to the inherent intractability of neural networks, especially when applied to large-scale datasets. This renders the direct calculation of " + }, + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(x|z)" + }, + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "text", + "content": " computationally prohibitive, making its maximization via backpropagation to eliminate the information redundancy during training exceedingly intractable. To solve this, a differentiable and computationally efficient lower bound on the conditional entropy " + }, + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(x|z)" + }, + { + "bbox": [ + 55, + 536, + 296, + 657 + ], + "type": "text", + "content": " is introduced in the next section to facilitate its maximization during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 664, + 296, + 715 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 296, + 715 + ], + "type": "text", + "content": "3.3. Differentiable Bound on Conditional Entropy Proposition 2 (The conditional entropy under uncertain encoding). Without loss of the generality, we consider that the encoding process " + }, + { + "bbox": [ + 55, + 664, + 296, + 715 + ], + "type": "inline_equation", + "content": "x \\rightarrow z" + }, + { + "bbox": [ + 55, + 664, + 296, + 715 + ], + "type": "text", + "content": " consists of two components:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "a deterministic mapping: " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\rightarrow \\hat{\\pmb{z}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " and a stochastic process: " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "z = \\hat{z} + \\varepsilon" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " is a random noise. " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " satisfies:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 323, + 100, + 555, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 100, + 555, + 113 + ], + "spans": [ + { + "bbox": [ + 323, + 100, + 555, + 113 + ], + "type": "interline_equation", + "content": "\\mathcal {H} (\\boldsymbol {x} | \\boldsymbol {z}) = \\mathcal {H} (\\boldsymbol {x}) - \\mathcal {H} (\\boldsymbol {z}) + \\mathcal {H} (\\boldsymbol {z} | \\hat {\\boldsymbol {z}}) = \\mathcal {H} (\\boldsymbol {x}) - \\mathcal {I} (\\boldsymbol {z}; \\hat {\\boldsymbol {z}}). \\tag {2}", + "image_path": "2735dc06ef345a9b2fe9f24a3806cf7a8b0d15e85819ea09b48cbc877e99b5d3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "text", + "content": "Proposition 2 is derived based on the definition of joint entropy, where " + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z}) = \\mathcal{H}(\\boldsymbol{x}) + \\mathcal{H}(\\boldsymbol{z}|\\boldsymbol{x}) - \\mathcal{H}(\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "text", + "content": ". The equation " + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(z|\\boldsymbol{x}) = \\mathcal{H}(z|\\hat{\\boldsymbol{z}})" + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "text", + "content": " holds due to the causal dependency from " + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} \\to \\hat{\\boldsymbol{z}} \\to \\boldsymbol{z}" + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} \\to \\hat{\\boldsymbol{z}}" + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "text", + "content": " is deterministic. Since " + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "text", + "content": " remains constant for a given data distribution, the primary challenge in maximizing the conditional entropy " + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "text", + "content": " indicated by Eq. 2 reduces to deriving a tractable and differentiable measure for " + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\boldsymbol{z};\\hat{\\boldsymbol{z}})" + }, + { + "bbox": [ + 313, + 116, + 555, + 211 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "content": "Consider the scenario where the task has " + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "content": " discrete targets " + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\{y_{1},\\dots,y_{n}\\}" + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "content": ". Following previous statements, we consider that the encoding process can be separated into a deterministic mapping: " + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\rightarrow \\hat{\\pmb{z}}" + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "content": " and a noise corruption process: " + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "inline_equation", + "content": "\\pmb{z} = \\hat{\\pmb{z}} + \\varepsilon" + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "inline_equation", + "content": "\\varepsilon \\sim \\mathcal{N}(0,\\Sigma_p)" + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "content": " is a Gaussian noise. We assume that the distribution of the encoded feature " + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}" + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "content": " can be effectively estimated by a " + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "content": "-component Gaussian mixture distribution denoted as " + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}} \\sim \\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i,\\Sigma_i)" + }, + { + "bbox": [ + 313, + 212, + 556, + 332 + ], + "type": "text", + "content": ", leveraging the proven capability of Gaussian mixtures to closely approximate complex natural data distributions [45]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "text", + "content": "Theorem 2 (Differentiable lower bound on " + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "text", + "content": "). Given " + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}}" + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "text", + "content": " follows the Gaussian mixture distribution and " + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "text", + "content": " is a Gaussian noise. The encoded feature " + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "inline_equation", + "content": "\\pmb{z} = \\hat{\\pmb{z}} + \\varepsilon" + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "text", + "content": " also follows a Gaussian mixture distribution with " + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\sim \\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i, \\Sigma_i + \\Sigma_p)" + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "text", + "content": ". Consequently, the conditional entropy " + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 313, + 336, + 556, + 408 + ], + "type": "text", + "content": " is bounded by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 408, + 555, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 408, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 318, + 408, + 555, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {H} (\\boldsymbol {x} | \\boldsymbol {z}) \\geq \\mathcal {H} (\\boldsymbol {x}) - \\sum_ {i = 1} ^ {k} \\pi_ {i} \\left(- \\log \\left(\\pi_ {i}\\right) + \\frac {1}{2} \\log \\left(\\frac {\\left| \\Sigma_ {i} + \\Sigma_ {p} \\right|}{\\left| \\Sigma_ {p} \\right|}\\right)\\right). \\tag {3}", + "image_path": "bbbc7fc9a6eda3cdb6c167bf94d6e4f9389ba21f120561dc3ee43b0ca93a7461.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": "Theorem 2 offers an effective and efficient way to bound the conditional entropy " + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": ", as the parameters " + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\pi_{i}" + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": " are all tractable and differentiable with respect to " + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{z}}" + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": ". Given the extracted " + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{z}}" + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": " after deterministic mapping and the Gaussian noise " + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": ", one can easily derive the parameters " + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\pi_{i}" + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": " using a Gaussian mixture model. This facilitates the maximization of " + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 440, + 556, + 559 + ], + "type": "text", + "content": " via gradient backpropagation utilizing Eq. 3, thereby improving the worst-case inversion robustness of the trained model. The proof of the Theorem 2 can be found in the supplementary material S.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 566, + 554, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 590 + ], + "type": "text", + "content": "3.4. Practical Insights on the Utility and Robustness Trade-off" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Feature utility: Building on the previous discussion, the encoded feature " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " follows a Gaussian mixture distribution with " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i, \\Sigma_i + \\Sigma_p)" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": ". In general, we assume that each Gaussian component or cluster belongs to one specific target. Ideally, we have " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "k = n" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " to accomplish the task. Due to the causal dependency from " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\rightarrow \\hat{\\mathbf{z}} \\rightarrow \\mathbf{z}" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": ", the utility of the intermediate feature " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " is intrinsically linked to its correlation with preceding states. Thereby, the utility of " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " is positively correlated with the expected posterior probability that " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " belonging to its original cluster " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": ", which is:" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8755" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 531, + 238 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 531, + 238 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 531, + 238 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 531, + 238 + ], + "type": "image", + "image_path": "f325142237c0eb01188cdd6872f31f53202c7197b12a8d414626e2a49b6e5dc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 237, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 62, + 237, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 237, + 545, + 248 + ], + "type": "text", + "content": "Figure 2. The versatile conditional entropy maximization algorithm for collaborative learning systems with split encoder and decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 255, + 296, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 255, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 255, + 296, + 285 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathcal {Z}} \\left[ \\mathbb {P} \\left(\\left(\\hat {\\boldsymbol {z}} \\in j \\mid \\boldsymbol {z}\\right)\\right) \\right] = \\mathbb {E} _ {\\mathcal {Z}} \\left[ \\frac {\\pi_ {j} \\mathcal {N} \\left(\\boldsymbol {z} ; \\mu_ {j} , \\Sigma_ {j} + \\Sigma_ {p}\\right)}{\\sum_ {i = 1} ^ {k} \\pi_ {i} \\mathcal {N} \\left(\\boldsymbol {z} ; \\mu_ {i} , \\Sigma_ {i} + \\Sigma_ {p}\\right)} \\right]. \\tag {4}", + "image_path": "57c28f881313c127efe17064cd6880e9da581057e42c8c2641aef11e4917e2eb.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "text", + "content": "Combining Eq. 3 and Eq. 4, we can get that the utility and robustness of the feature " + }, + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "text", + "content": " are dependent on the " + }, + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathcal{Z}}[\\mathbb{P}((\\hat{\\mathbf{z}} = j|\\mathbf{z}))]" + }, + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(z;\\hat{z})" + }, + { + "bbox": [ + 55, + 292, + 296, + 364 + ], + "type": "text", + "content": " respectively. These two items are empirically closely linked, leading to a trade-off between utility and robustness. Additionally, the following practical insights can be derived:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 296, + 628 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "text", + "content": "- Physical interpretation of " + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "text", + "content": ": Physically, the mean " + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "text", + "content": " generally captures dominant or representative characteristics of a cluster, while the covariance reflects the degree of variability or noise within the cluster. Consistent with this interpretation, our analysis reveals that " + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "text", + "content": " is crucial for evaluating the feature utility. In contrast, the covariance " + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "text", + "content": " is closely linked to the task-irrelevant redundancy in " + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 55, + 365, + 296, + 472 + ], + "type": "text", + "content": ", which diminishes feature utility and increases susceptibility to MIAs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 567 + ], + "type": "text", + "content": "- Impact of noise corruption on inversion robustness: The noise corruption introduced by " + }, + { + "bbox": [ + 55, + 472, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\Sigma_{p}" + }, + { + "bbox": [ + 55, + 472, + 296, + 567 + ], + "type": "text", + "content": " perturbs both the task-important and redundant information, thus balancing the trade-off between utility and robustness. When there is a distinctive difference between the dominant features, such as " + }, + { + "bbox": [ + 55, + 472, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\min_{i \\neq j} \\| \\mu_{i} - \\mu_{j} \\|_{2}^{2} \\gg Tr(\\Sigma_{p}) \\gg Tr(\\Sigma_{i} + \\Sigma_{j})" + }, + { + "bbox": [ + 55, + 472, + 296, + 567 + ], + "type": "text", + "content": ", adding the noise greatly corrupts the redundancy while bringing a small influence on the utility." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "content": "- Ideal situation: For the ideal situation, we have " + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\Sigma_{i} = \\Sigma_{p} = 0" + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "inline_equation", + "content": "k = n" + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "content": " and, " + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\pi_{i} = \\mathcal{P}(y_{i})" + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "content": ". Thereby, we can easily get " + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathcal{Z}}[\\mathbb{P}((\\hat{z} = j|z))] = 1" + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{z}) = \\mathcal{H}(\\boldsymbol{y})" + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "content": ", indicating that the feature " + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}" + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "content": " contains only the necessary information for the task, without any redundancy." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": "However, for a lightweight encoder deployed on edge devices, it is exceedingly challenging to diminish all redundancies while preserving only the essential features. An alternative approach for developing a robust feature encoder is to jointly optimize both the covariance matrices " + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": " and the distance between the " + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu_{j}" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": " among different clusters, leading to " + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\min_{i\\neq j}\\| \\mu_i - \\mu_j\\| _2^2\\gg Tr(\\Sigma_p)\\gg Tr(\\Sigma_i + \\Sigma_j)" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": ". Thus," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 258, + 552, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 258, + 552, + 270 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 552, + 270 + ], + "type": "text", + "content": "Algorithm 1 Conditional entropy maximization algorithm" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 273, + 555, + 507 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": "1: Input: local encoder " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": ", cloud decoder " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": ", image dataset " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": ", training epochs " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": ", conditional entropy loss " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": " with weight factor " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": ", Gaussian noise " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": ", existing defense methods " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": " with loss " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": ", Gaussian mixture model GMM, number of mixture clusters " + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 318, + 273, + 555, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 328, + 395, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 328, + 395, + 338 + ], + "spans": [ + { + "bbox": [ + 319, + 328, + 395, + 338 + ], + "type": "text", + "content": "2: Initialize " + }, + { + "bbox": [ + 319, + 328, + 395, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e, \\mathcal{F}_d" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 339, + 425, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 339, + 425, + 349 + ], + "spans": [ + { + "bbox": [ + 320, + 339, + 425, + 349 + ], + "type": "text", + "content": "3: for epoch from 1 to " + }, + { + "bbox": [ + 320, + 339, + 425, + 349 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 320, + 339, + 425, + 349 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 350, + 490, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 350, + 490, + 361 + ], + "spans": [ + { + "bbox": [ + 320, + 350, + 490, + 361 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 320, + 350, + 490, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\gets \\text{concatenate}\\{\\mathcal{F}_e(\\pmb{x}) + \\varepsilon, \\forall \\pmb{x} \\in \\mathcal{D}\\}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "text", + "content": "5: Estimate the distribution of " + }, + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "text", + "content": " as a " + }, + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "text", + "content": " component Gaussian mixture: " + }, + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\sum_{j=1}^{k} \\pi_j \\mathcal{N}(\\mu_j, \\Sigma_j + \\Sigma_p)" + }, + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "text", + "content": " using GMM(" + }, + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{z}, k" + }, + { + "bbox": [ + 320, + 361, + 553, + 384 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 384, + 505, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 384, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 320, + 384, + 505, + 393 + ], + "type": "text", + "content": "6: for image batch " + }, + { + "bbox": [ + 320, + 384, + 505, + 393 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_i, \\pmb{y}_i)" + }, + { + "bbox": [ + 320, + 384, + 505, + 393 + ], + "type": "text", + "content": " in the dataset " + }, + { + "bbox": [ + 320, + 384, + 505, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 320, + 384, + 505, + 393 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 394, + 425, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 394, + 425, + 404 + ], + "spans": [ + { + "bbox": [ + 320, + 394, + 425, + 404 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 320, + 394, + 425, + 404 + ], + "type": "inline_equation", + "content": "\\pmb{z}_i = \\mathcal{F}_e(\\pmb{x}_i) + \\varepsilon_i" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 405, + 405, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 405, + 405, + 416 + ], + "spans": [ + { + "bbox": [ + 320, + 405, + 405, + 416 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 320, + 405, + 405, + 416 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i' = \\mathcal{F}_d(\\pmb{z}_i)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 417, + 471, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 417, + 471, + 426 + ], + "spans": [ + { + "bbox": [ + 320, + 417, + 471, + 426 + ], + "type": "text", + "content": "9: Assign " + }, + { + "bbox": [ + 320, + 417, + 471, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 320, + 417, + 471, + 426 + ], + "type": "text", + "content": " to the nearest cluster" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 427, + 450, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 450, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 450, + 437 + ], + "type": "text", + "content": "10: for cluster " + }, + { + "bbox": [ + 317, + 427, + 450, + 437 + ], + "type": "inline_equation", + "content": "j\\gets 1" + }, + { + "bbox": [ + 317, + 427, + 450, + 437 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 317, + 427, + 450, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 317, + 427, + 450, + 437 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 438, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 482, + 449 + ], + "type": "text", + "content": "11: Update the weight " + }, + { + "bbox": [ + 317, + 438, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\pi_{j}" + }, + { + "bbox": [ + 317, + 438, + 482, + 449 + ], + "type": "text", + "content": " by Eq. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 449, + 511, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 511, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 511, + 460 + ], + "type": "text", + "content": "12: Update the covariance matrix by Eq. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 461, + 487, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 461, + 487, + 472 + ], + "spans": [ + { + "bbox": [ + 317, + 461, + 487, + 472 + ], + "type": "text", + "content": "13: Calculate " + }, + { + "bbox": [ + 317, + 461, + 487, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 317, + 461, + 487, + 472 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 317, + 461, + 487, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\pmb{x}_i, \\pmb{y}_i, \\pmb{z}_i, \\pmb{y}_i')" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 473, + 443, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 473, + 443, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 473, + 443, + 483 + ], + "type": "text", + "content": "14: Calculate " + }, + { + "bbox": [ + 317, + 473, + 443, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 317, + 473, + 443, + 483 + ], + "type": "text", + "content": " by Eq. 7." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 484, + 492, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 484, + 492, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 484, + 492, + 495 + ], + "type": "text", + "content": "15: Optimize " + }, + { + "bbox": [ + 317, + 484, + 492, + 495 + ], + "type": "inline_equation", + "content": "\\min_{\\mathcal{F}_e,\\mathcal{F}_d}\\left(\\mathcal{L}_D + \\lambda *\\mathcal{L}_C\\right)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 497, + 496, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 496, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 496, + 507 + ], + "type": "text", + "content": "16: Output: Trained encoder " + }, + { + "bbox": [ + 316, + 497, + 496, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e" + }, + { + "bbox": [ + 316, + 497, + 496, + 507 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 316, + 497, + 496, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 523, + 555, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 560 + ], + "type": "text", + "content": "incorporating appropriate noise can effectively diminish the residual redundancies, achieving the maximization of the lower bound on " + }, + { + "bbox": [ + 313, + 523, + 555, + 560 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 313, + 523, + 555, + 560 + ], + "type": "text", + "content": " while ensuring prediction accuracy." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 573, + 555, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 555, + 601 + ], + "type": "text", + "content": "4. The Versatile Conditional Entropy Maximization Algorithm for Collaborative Inference" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Building upon the derived theoretical insights on the differentiable lower bound of the conditional entropy " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " and the practical insights on the utility and robustness trade-off, this section introduces a versatile conditional entropy maximization (CEM) algorithm. The CEM algorithm is designed to enhance the robustness of existing obfuscation-based defense mechanisms against MIAs in collaborative inference systems. For a given defense mechanism, denoted as " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", the proposed CEM algorithm begins by assessing its worst-case" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8756" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "robustness. This is achieved through a Gaussian mixture estimation process on the feature " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " to determine the parameters " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". Subsequently, during each training batch, the algorithm maximizes the lower bound of " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(x|z)" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " specified in Theorem 2 to strengthen robustness against MIAs. The details of the proposed CEM algorithm are presented in Figure 2 and Algorithm 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": "Key techniques: Let " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " denote the local feature encoder and " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " denote the decoder deployed in the cloud. By estimating the distribution of noisy representation " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": "-component Gaussian mixture " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{k} \\pi_i \\mathcal{N}(\\mu_i, \\Sigma_i + \\Sigma_p)" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": ", the lower bound of the conditional entropy " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(x|z)" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " can be derived based on Eq. 3, utilizing the covariance matrices " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\Sigma_i" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " and the weights of the mixture components " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\pi_i" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": ". Since both of them are differentiable and solvable to the feature representation " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": ", this lower bound of the conditional entropy " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(x|z)" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " can be effectively maximized via the gradient backpropagation. As described in Algorithm 1, each training iteration begins by fitting the distribution of " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " with a Gaussian mixture model (GMM). This involves optimizing the component weights " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": ", means " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": ", and covariance matrices " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " to minimize the estimation error. For each data batch, the extracted feature representations are assigned to the mixture component " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " corresponding to the nearest mean " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\mu_j" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": ", with the mixture parameters updated accordingly. The component weights " + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "inline_equation", + "content": "\\pi_j" + }, + { + "bbox": [ + 56, + 157, + 298, + 384 + ], + "type": "text", + "content": " are updated by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 119, + 388, + 296, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 388, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 119, + 388, + 296, + 410 + ], + "type": "interline_equation", + "content": "\\pi_ {j} = \\frac {\\pi_ {j} \\left(N - N _ {b a t c h}\\right) + n _ {j}}{N}, \\tag {5}", + "image_path": "3e50b95358182ff3a619d3f75ce3c7bae8c390701fcf49ef8be40549e7cb566d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "text", + "content": " is the total length of the dataset and " + }, + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "inline_equation", + "content": "N_{\\text{batch}}" + }, + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "text", + "content": " is the batch size. " + }, + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "inline_equation", + "content": "n_j" + }, + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "text", + "content": " is the number of feature representations that are assigned to the mixture component " + }, + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 415, + 296, + 464 + ], + "type": "text", + "content": ". The covariance matrices are updated by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 470, + 296, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 470, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 296, + 495 + ], + "type": "interline_equation", + "content": "\\Sigma_ {j} = \\left(1 - \\frac {n _ {j}}{\\pi_ {j} N}\\right) \\Sigma_ {j} + \\frac {n _ {j}}{\\pi_ {j} N} \\Delta \\Sigma_ {j}, \\tag {6}", + "image_path": "26b655b70cb90f3579ffc6c1a4a3fa0cf90e5bb60fb3cf2cba1caaa8001a656f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\Delta \\Sigma_{j}" + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "text", + "content": " is the calculated covariance based on newly assigned feature representations. The updated parameters are then utilized to compute the conditional entropy loss " + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 55, + 498, + 296, + 544 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 545, + 296, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 545, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 77, + 545, + 296, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} = \\sum_ {i = j} ^ {k} \\pi_ {j} \\left(- \\log \\left(\\pi_ {j}\\right) + \\frac {1}{2} \\log \\left(\\frac {\\left| \\Sigma_ {j} + \\Sigma_ {p} \\right|}{\\left| \\Sigma_ {p} \\right|}\\right)\\right). \\tag {7}", + "image_path": "b0f3265e033d43581c9789a86c51647242011b2d511fbcf88db979ce1f58354c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 616 + ], + "type": "text", + "content": "This loss function is then utilized as an effective indicator of worst-case inversion robustness and is minimized through gradient-based optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "The combination with other defense strategies: As shown in Figure 2, existing obfuscation-based defense methods, such as noise corruption [18, 51], adversarial representation learning [7, 28], and feature purification [6, 19, 53] are collectively denoted as " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": ", and an auxiliary training loss " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " is used to encapsulate their impact in the backward optimization process. For approaches that involve architectural modifications, such as introducing information filtering" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 340, + 73, + 525, + 217 + ], + "blocks": [ + { + "bbox": [ + 340, + 73, + 525, + 217 + ], + "lines": [ + { + "bbox": [ + 340, + 73, + 525, + 217 + ], + "spans": [ + { + "bbox": [ + 340, + 73, + 525, + 217 + ], + "type": "image", + "image_path": "ab3e19b7dbc88aeb220e6675f89f57d829d21d5e527db683046544a4928e26ac.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 217, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 322, + 217, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 322, + 217, + 545, + 228 + ], + "type": "text", + "content": "Figure 3. Reconstruction MSE Vs. " + }, + { + "bbox": [ + 322, + 217, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\pmb{x}|\\pmb{z})" + }, + { + "bbox": [ + 322, + 217, + 545, + 228 + ], + "type": "text", + "content": " on the CIFAR10." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 554, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 554, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 554, + 302 + ], + "type": "text", + "content": "layers or adversarial reconstruction models, we adhere to their defined configurations to incorporate our method. The proposed CEM algorithm can be seamlessly plugged into other defense mechanisms by embedding Gaussian mixture estimation within the training process and optimizing the combined loss function:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 397, + 308, + 555, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 308, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 397, + 308, + 555, + 319 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {D} + \\lambda * \\mathcal {L} _ {C}, \\tag {8}", + "image_path": "d214b517be3f910ceb690177034ff4907c8432840eaec11270f51e45387fe20d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 324, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 324, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 324, + 555, + 384 + ], + "type": "text", + "content": " is a weight factor to balance the optimization. The utilization of the proposed CEM algorithm allows for an effective evaluation of worst-case inversion robustness of the feature representation throughout training and further optimizes this robustness via gradient backward propagation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": "Conditional entropy " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": " and reconstruction MSE by MIAs in practical scenarios: Theorem 1 indicates an exponential relationship between the lower bound on minimal MSE " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": " and the conditional entropy " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": ". We thereby empirically evaluate this relationship on the CIFAR10 dataset by training multiple models with varying levels of measured " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": " and assessing their inversion robustness by assuming a proxy inversion adversary. Following the experimental setting in [28], we utilize a VGG11 model with the first 2 convolutional layers as encoder " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": " and the rest of layers as the decoder " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": ". We fix the " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\lambda = 16" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": " and train different models using different Gaussian noise with variance varying from 0.01 to 0.3, thus leading to varying " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": ". The relative conditional entropy " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": " is calculated by Eq. 3 (we use 'relative' due to " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": " is constant and its calculation over the whole data space is intractable. So a proper value is selected to replace it). We report the average reconstruction MSE by training a DNN-based MIA decoder " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\boldsymbol{z},\\mathcal{X},\\mathcal{F}_e)" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": " with the same architecture as [28] and allow its full access to the training data and encoder " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": ". The analytic results are shown in Figure 3, revealing a strong exponential correlation between the " + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\boldsymbol{x}|\\boldsymbol{z})" + }, + { + "bbox": [ + 313, + 384, + 556, + 647 + ], + "type": "text", + "content": " and the MIA reconstruction MSE." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 657, + 440, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 440, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 440, + 670 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 673, + 441, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 441, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 441, + 686 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "content": "We investigate the inversion robustness of collaborative inference models implemented on three general object classifi-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8757" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 137, + 293, + 324 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 137 + ], + "type": "text", + "content": "Table 1. Comparative results on CIFAR10 dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the VGG11 model before and after the integration with the proposed CEM algorithm. The last 2 rows report the average performance of all methods, comparing results with and without incorporating the CEM algorithm." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 137, + 293, + 324 + ], + "lines": [ + { + "bbox": [ + 56, + 137, + 293, + 324 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 293, + 324 + ], + "type": "table", + "html": "
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense91.860.00130.00140.00140.0016
Bottleneck90.870.00360.00410.00390.0045
Bottleneck+CEM90.690.00540.00580.00760.0080
DistCorr [53]89.520.00740.00820.00790.0088
DistCorr+CEM89.800.00900.00930.00940.0096
Dropout [15]87.750.00980.01040.00990.0111
Dropout+CEM87.530.01290.01340.01320.0142
PATROL [7]89.580.02450.02930.02570.0307
PATROL+CEM89.670.03040.03350.03130.0347
ResSFL [28]89.680.01460.02400.01580.0243
ResSFL+CEM90.110.02010.02510.02290.0273
Noise_Nopeek [51]87.190.01520.01590.01560.0165
Noise_Nopeek+CEM87.080.01740.01740.01730.0178
Noise_ARL [18]87.780.02900.03360.03040.0342
Noise_ARL+CEM87.620.03300.03550.03410.0363
Average w/o CEM88.910.01480.01790.01560.0185
Average w/ CEM88.920.01830.02000.01940.0211
", + "image_path": "dcf961df057d6cf756fc31e1b785c542b2a9fc512518b5ab7aad8436dc52e728.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 335, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 396 + ], + "type": "text", + "content": "cation datasets: CIFAR-10 [25], CIFAR-100 [25], TinyImageNet [26], and a face recognition dataset FaceScrub [38]. For the CIFAR-10, CIFAR-100, and FaceScrub datasets, we use the VGG11 as the basic model and for the TinyImageNet dataset, we use the ResNet-20 as the basic model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 399, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 295, + 602 + ], + "type": "text", + "content": "Collaborative inference system: To simulate the collaborative inference, we split the VGG11 model by allocating the first two convolutional layers as the local encoder and assigning the remaining layers as the cloud decoder. For the ResNet-20 model, the first four convolutional years are allocated as the local encoder, with the remaining blocks functioning as the cloud-based decoder. This partitioning yields a computational distribution where the encoder is responsible for approximately " + }, + { + "bbox": [ + 54, + 399, + 295, + 602 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 399, + 295, + 602 + ], + "type": "text", + "content": " of the total computation and the decoder handles the remaining " + }, + { + "bbox": [ + 54, + 399, + 295, + 602 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 54, + 399, + 295, + 602 + ], + "type": "text", + "content": ". Unless otherwise specified, we maintain the same partitioning scheme across all evaluated methods for a fair comparison. We utilize the Stochastic Gradient Descent (SGD) optimizer to jointly optimize both the encoder and decoder with an initial learning rate of 0.05 and an appropriate learning rate decay. The VGG11 model is trained over 240 epochs and the ResNet20 model is trained over 120 epochs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Obfuscation defense methods: The proposed CEM can be flexibly integrated into the model training process to enhance the worst-case inversion robustness by introducing an auxiliary Gaussian mixture estimation to maximize the conditional entropy. We evaluate the effectiveness of the proposed CEM in enhancing several existing defense methods, including the information pruning-based methods called DistCorr [53] and Dropout [15]; adversarial representation learning-based method called ResSFL [28] and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": "PATROL [7]; and noise corruption-based method called Noise_Nopeek [51] and Noise_ARL [18]. As observed in [28], the integration of bottleneck layers substantially enhances inversion robustness. To ensure a fair and rigorous comparison, we incorporate a bottleneck layer with 8 channels for both pre- and post-processing of extracted feature representations across all evaluated methods. For PATROL [7] that introduces additional layers within the encoder and maintains the inference efficiency through network pruning techniques, we adjust the pruning ratio to ensure the encoder in PATROL remains the same size as other methods. To analyze the effect of the proposed CEM algorithm, we evaluate the performance of all methods both before and after integrating them with the CEM algorithm. If not otherwise stated, we set hyperparameters " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\lambda = 16" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " and utilize the isotropic Gaussian noise " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " with a standard deviation of 0.025. We set the number of Gaussian mixture components " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "k = 3n" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " due to the limited number of training data and the imperfect representation ability of the shallow encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 300, + 556, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 556, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 556, + 433 + ], + "type": "text", + "content": "Threat model: To evaluate the worst-case inversion robustness, we consider the white-box scenarios where the MIAs are trained with full access to the collaborative inference model and training dataset. We split the training and testing data strictly following the default setting, e.g. 50,000 images for training and 10,000 for testing on the CIFAR10 dataset. Two types of inversion attacks: the DNN-based decoding method [28] and the generative-based inversion method [67] are utilized as the MIAs. The detailed architecture and training mechanism of those threat models can be found in the supplementary material S.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 434, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 589 + ], + "type": "text", + "content": "Evaluation metrics: We report the prediction accuracy and the reconstruction MSE to evaluate the utility and inversion robustness of the intermediate feature. We also report other metrics such as PSNR and SSIM to evaluate the robustness and present the experimental results in the supplementary material S.4. We consider the information leakage on training and inference data, where we report the MSE on reconstructing the training and testing dataset by MIAs. To further explore the intrinsic trade-off between the feature utility and inversion robustness for different defense mechanisms, we present the accuracy-MSE curve by varying the defense hyperparameters, such as the noise strength and CEM loss weight factor " + }, + { + "bbox": [ + 313, + 434, + 556, + 589 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 434, + 556, + 589 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 601, + 473, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 473, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 473, + 612 + ], + "type": "text", + "content": "5.2. Results on Different Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Results on CIFAR10: The performance of different defense methods on the CIFAR10 dataset is presented in Table 1. We provide a comparative analysis of each method, both before and after integrating the proposed CEM algorithm. The results indicate that the features, extracted by the lightweight encoder without defense mechanisms, exhibit significant redundancy, which can be exploited by MIAs to reconstruct high-fidelity inputs. Information pruning methods, such" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8758" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 118, + 297, + 309 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "text", + "content": "Table 2. Comparative results on TinyImageNet dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the ResNet-20 model with and without integrating the proposed CEM." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 118, + 297, + 309 + ], + "lines": [ + { + "bbox": [ + 56, + 118, + 297, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 297, + 309 + ], + "type": "table", + "html": "
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense53.730.00250.00220.00200.0021
Bottleneck52.770.01070.01030.00910.0092
Bottleneck+CEM52.250.01400.01360.01230.0125
DistCorr [53]51.790.01480.01450.01260.0136
DistCorr+CEM51.780.01950.01900.01670.0168
Dropout [15]50.720.01650.01630.01480.0151
Dropout+CEM50.750.01880.01860.01670.0172
PATROL [7]51.750.01870.01870.01680.0176
PATROL+CEM51.640.02110.02090.01850.0189
ResSFL [28]51.990.01730.01720.01570.0161
ResSFL+CEM52.070.01970.01940.01800.0173
Noise_Nopeek [51]51.630.01610.01570.01470.0152
Noise_Nopeek+CEM52.010.01830.01800.01660.0167
Noise_ARL [18]51.030.02290.02240.02040.0205
Noise_ARL+CEM50.850.02810.02710.02510.0253
Average w/o CEM51.660.01670.01640.01480.0153
Average w/ CEM51.620.01990.01950.01770.0178
", + "image_path": "3bfd592597ef3a362f8564bdbd9b994bdcdddefc583ab3cbfca485773aba692f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 316, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 544 + ], + "type": "text", + "content": "as DistCorr [53] and dropout [15] that are based on some empirical observations, inevitably remove task-specific information when eliminating the redundancy, leading to a degradation of the prediction accuracy. Methods based on adversarial representation learning (ARL), such as ResSFL, PATROL, and Noise_ARL, offer a better trade-off in feature utility and robustness, achieving substantial robustness gain with a moderate accuracy drop. By incorporating a Gaussian mixture estimation process that is independent of the prediction process, the proposed CEM algorithm effectively increases the lower bound of the reconstruction MSE via maximizing the conditional entropy. This positions the CEM algorithm as a highly adaptable framework for assessing and enhancing the inversion robustness of a wide range of collaborative inference models. The results in Table 1 show that the integration of the CEM algorithm consistently enhances all defense methods, yielding an average increase in reconstruction MSE of " + }, + { + "bbox": [ + 56, + 316, + 296, + 544 + ], + "type": "inline_equation", + "content": "24.0\\%" + }, + { + "bbox": [ + 56, + 316, + 296, + 544 + ], + "type": "text", + "content": " for training data and " + }, + { + "bbox": [ + 56, + 316, + 296, + 544 + ], + "type": "inline_equation", + "content": "12.9\\%" + }, + { + "bbox": [ + 56, + 316, + 296, + 544 + ], + "type": "text", + "content": " for inference data, without compromising prediction accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "content": "Results on TinyImageNet: The performance of different defense methods on the TinyImageNet dataset is presented in Table 2, where we provide a comparative analysis of each method, both before and after integrating the proposed CEM algorithm. The results indicate again that the integration of the CEM algorithm brings substantial inversion robustness gain. Integrating the CEM algorithm yields an average increase in the reconstruction MSE of " + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "inline_equation", + "content": "19.4\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "content": " for training data and " + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "inline_equation", + "content": "17.7\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "content": " for inference data, without compromising prediction accuracy. Meanwhile, the combination of ARL-based methods such as PATROL and Noise_ARL with CEM achieves the best utility and robustness trade-off bringing a great robustness gain with only a marginal accuracy drop." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "Results on Facescrub: The performance of different de" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 118, + 556, + 308 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 3. Comparative results on Facescrub dataset: we present the prediction accuracy and reconstruction MSE of different defense methods using the VGG11 model with and without integrating the proposed CEM." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 118, + 556, + 308 + ], + "lines": [ + { + "bbox": [ + 316, + 118, + 556, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 556, + 308 + ], + "type": "table", + "html": "
MethodsAcc.↑Dec.-based MSE [28]GAN-based MSE [67]
Train↑Infer↑Train↑Infer↑
No-defense86.690.00120.00110.00120.0014
Bottleneck85.120.00250.00250.00260.0027
Bottleneck+CEM85.000.00360.00350.00380.0037
DistCorr [53]83.420.00380.00380.00410.0041
DistCorr+CEM83.780.00480.00470.00690.0074
Dropout [15]79.300.00520.00520.00540.0057
Dropout+CEM79.190.00740.00740.00760.0081
PATROL [7]79.180.00990.01180.01140.0137
PATROL+CEM79.880.01660.01850.01840.0192
ResSFL [28]79.600.00940.01120.01110.0142
ResSFL+CEM79.540.01280.01480.01430.0161
Noise_Nopeek [51]82.060.00520.00520.00530.0056
Noise_Nopeek+CEM81.960.00760.00750.00780.0090
Noise_ARL [18]80.140.01220.01550.01320.0158
Noise_ARL+CEM80.330.01820.02110.02120.0231
Average w/o CEM81.260.00690.00790.00760.088
Average w/ CEM81.380.01010.01110.01140.0123
", + "image_path": "05962d005a0c79debaaedd7df486fb814f8d1dbeec8e5f3c4ff9ec77270508dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 314, + 556, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 556, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 556, + 505 + ], + "type": "text", + "content": "fense methods on the Facescrub dataset is presented in Table 3. Unlike their performance on general object classification datasets, models in this task demonstrate increased vulnerability to MIAs. The evaluated defense methods exhibit a notable accuracy reduction as a trade-off for increased inversion robustness. This may caused by the strong correlation between facial features and identity information. In this dataset, integrating the CEM algorithm with existing defense methods brings a more pronounced improvement in inversion robustness. Specifically, the incorporation of the CEM algorithm yields an average increase in inversion robustness of " + }, + { + "bbox": [ + 313, + 314, + 556, + 505 + ], + "type": "inline_equation", + "content": "48.2\\%" + }, + { + "bbox": [ + 313, + 314, + 556, + 505 + ], + "type": "text", + "content": " for training data and " + }, + { + "bbox": [ + 313, + 314, + 556, + 505 + ], + "type": "inline_equation", + "content": "40.1\\%" + }, + { + "bbox": [ + 313, + 314, + 556, + 505 + ], + "type": "text", + "content": " for inference data while maintaining prediction accuracy. Meanwhile, the combination of ARL-based methods also achieves the best utility and robustness trade-off, which leads to a great robustness gain with only a marginal accuracy drop." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 505, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 529 + ], + "type": "text", + "content": "Results on CIFAR100: The results on the CIFAR-100 dataset are provided in the supplementary materials S.5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 529, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 589 + ], + "type": "text", + "content": "Visualized results: We present the visualized MIA reconstruction on the TinyImageNet dataset in Figure 5. The results illustrate that incorporating the proposed CEM algorithm provides enhanced user input protection, resulting in decreased inverse reconstruction fidelity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 593, + 541, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 541, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 541, + 605 + ], + "type": "text", + "content": "5.3. Ablation Study: Effect of Hyperparameters" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "text", + "content": "To rigorously evaluate the impact of different hyperparameters, such as the noise strength and CEM weight factor " + }, + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 605, + 555, + 676 + ], + "type": "text", + "content": ", on the performance of the CEM algorithm, we conduct experiments under various settings of hyperparameters to demonstrate its robustness, versatility, and adaptability in diverse scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "The effect of noise strength: As discussed in Subsection 3.4, the intensity of additive Gaussian noise is pivotal in balancing a good trade-off between utility and robust-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8759" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 216, + 189 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 216, + 189 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 216, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 216, + 189 + ], + "type": "image", + "image_path": "1c122f6ba0ad8ab67eeaddabbe8cec56cf08d48e38a83b0b2b47ef614e43982f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 192, + 206, + 202 + ], + "lines": [ + { + "bbox": [ + 66, + 192, + 206, + 202 + ], + "spans": [ + { + "bbox": [ + 66, + 192, + 206, + 202 + ], + "type": "text", + "content": "(a) Noise variance varying from 0.01 to 0.5" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 70, + 383, + 189 + ], + "blocks": [ + { + "bbox": [ + 224, + 70, + 383, + 189 + ], + "lines": [ + { + "bbox": [ + 224, + 70, + 383, + 189 + ], + "spans": [ + { + "bbox": [ + 224, + 70, + 383, + 189 + ], + "type": "image", + "image_path": "97704faa14adae42b14152fc3ce6cc7195237730e7d26a01571c738b485b67b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 193, + 375, + 202 + ], + "lines": [ + { + "bbox": [ + 233, + 193, + 375, + 202 + ], + "spans": [ + { + "bbox": [ + 233, + 193, + 375, + 202 + ], + "type": "text", + "content": "(b) comparing results of different value of " + }, + { + "bbox": [ + 233, + 193, + 375, + 202 + ], + "type": "inline_equation", + "content": "\\lambda" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 549, + 189 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 549, + 189 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 549, + 189 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 549, + 189 + ], + "type": "image", + "image_path": "1aee86309c41569810318595932f57a91956d88710e7f9362d74af4eae0e5bf1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 193, + 541, + 202 + ], + "lines": [ + { + "bbox": [ + 401, + 193, + 541, + 202 + ], + "spans": [ + { + "bbox": [ + 401, + 193, + 541, + 202 + ], + "type": "text", + "content": "(c) comparing results of different partitions" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 232, + 297, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 232, + 297, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 297, + 388 + ], + "type": "text", + "content": "ness. We thereby evaluate the performance of the CEM algorithm under various noise strengths. We utilize the Noise_Nopeek [51] and Noise_ARL [18] as the basic defense strategies " + }, + { + "bbox": [ + 54, + 232, + 297, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 54, + 232, + 297, + 388 + ], + "type": "text", + "content": ". In Figure 4a, we present the prediction accuracy Vs. reconstruction MSE curves on the CIFAR-10 dataset by varying the noise variance from 0.001 to 0.5. Multiple models were trained to analyze the impact comprehensively. The results demonstrate that the noise variance significantly influences the trade-off between feature utility and robustness. The proposed CEM algorithm exhibits robust performance across the whole range of noise intensities. Integrating the CEM algorithm consistently achieves a more favorable balance between utility and robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "text", + "content": "The effect of " + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "text", + "content": ": In Equation 8, the weight factor " + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "text", + "content": " is important in adjusting the conditional entropy maximization during the joint optimization. To investigate the influence of " + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "text", + "content": ", we use the Noise_ARL method [18], which achieves a comparatively high performance as our baseline strategy " + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "text", + "content": ". We then evaluate its performance when integrated with the CEM algorithm, using " + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "text", + "content": " values of 0, 2, 8, 16, and 32. We present the prediction accuracy and reconstruction MSE curve in Figure 4b, by varying the noise variance from 0.001 to 0.5. The results demonstrate that increasing " + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 389, + 296, + 521 + ], + "type": "text", + "content": " from 0 to 16 markedly enhances the robustness and utility trade-off." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": "Different partitioning mechanisms: Encoders with more layers extract more task-specific features from input data, thereby providing improved robustness against MIAs. To illustrate the effectiveness of the proposed CEM under different levels of redundancy, we evaluate its performance across multiple partitioning schemes. Specifically, we use the VGG11 architecture as the backbone, partitioned at the 1st, 2nd, and 3rd convolutional layers (denoted as " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "c_{2}" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "c_{3}" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": "), and assess the performance of CEM in combination with Noise_ARL [18]. We present the prediction accuracy Vs. reconstruction MSE curve in Figure 4c using the same strategy. The results demonstrate a significant performance advantage with deeper partitioning: specifically, partitioning at the third convolutional layer achieves substantial inversion robustness while maintaining prediction accuracy without degradation. Furthermore, the CEM algorithm exhibits high" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 232, + 547, + 342 + ], + "blocks": [ + { + "bbox": [ + 55, + 204, + 553, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 553, + 226 + ], + "type": "text", + "content": "Figure 4. The effect of different hyperparameters on the performance of the proposed CEM algorithm. The effect of noise strength, " + }, + { + "bbox": [ + 55, + 204, + 553, + 226 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 204, + 553, + 226 + ], + "type": "text", + "content": ", and partitioning schemes are demonstrated by the utility-robustness trade-off curve." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 232, + 547, + 342 + ], + "lines": [ + { + "bbox": [ + 323, + 232, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 323, + 232, + 547, + 342 + ], + "type": "image", + "image_path": "a9f407e4ca8090589dde4511f48ba7269e5a1746ed78ca576258213cd945020e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 342, + 553, + 363 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 363 + ], + "type": "text", + "content": "Figure 5. The visualized result. The last three rows are inputs reconstructed by MIA." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 389 + ], + "type": "text", + "content": "adaptability, consistently enhancing the robustness-utility trade-off across different partitioning schemes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 390, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 450 + ], + "type": "text", + "content": "Analysis of the efficiency: The computational complexity on the local encoder and the cloud server across various methods is presented in the supplementary material S.6. The results confirm that integrating the proposed CEM does not introduce any additional computational overhead or latency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 456, + 388, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 388, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 388, + 468 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 474, + 556, + 715 + ], + "type": "text", + "content": "This work addresses problems of inversion robustness in collaborative inference systems, where MIAs can exploit subtle signals within these intermediate features to reconstruct high-fidelity inputs. Existing obfuscation defenses seek to protect privacy by empirically eliminating feature redundancy. However, the precise quantification of such redundancy is lacking, and rigorous mathematical analysis is needed to elucidate the relation between redundancy minimization and enhanced inversion robustness. To solve that, we prove that the conditional entropy of inputs given intermediate features provides a lower bound on the reconstruction MSE. Based on that, we derive a differentiable lower bound on this conditional entropy using Gaussian mixture estimation, making it amenable for efficient optimization through backpropagation. Building on this, we propose a versatile conditional entropy maximization (CEM) algorithm that can be easily plugged into existing methods. Comprehensive Experiments demonstrate that the proposed CEM significantly and consistently boosts robustness while maintaining feature utility and computational efficiency." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8760" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 297 + ], + "type": "text", + "content": "This work was carried out at the Rapid-Rich Object Search (ROSE) Lab, School of Electrical & Electronic Engineering, Nanyang Technological University (NTU), Singapore. This research is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative, the Basic and Frontier Research Project of PCL, the Major Key Project of PCL, Guangdong Basic and Applied Basic Research Foundation under Grant 2024A1515010454, the Program of Beijing Municipal Science and Technology Commission Foundation (No.Z241100003524010), and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore and Infocomm Media Development Authority." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 308, + 115, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 115, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 115, + 320 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 328, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 328, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 328, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 61, + 328, + 296, + 393 + ], + "type": "text", + "content": "[1] Sharif Abuadbba, Kyuyeon Kim, Minki Kim, Chandra Thapa, Seyit A Camtepe, Yansong Gao, Hyoungshick Kim, and Surya Nepal. Can we use split learning on 1d cnn models for privacy preserving training? In Proceedings of the 15th ACM Asia conference on computer and communications security, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 396, + 296, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 296, + 440 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "type": "text", + "content": "[3] Martin Bertran, Natalia Martinez, Afroditi Papadaki, Qiang Qiu, Miguel Rodrigues, Galen Reeves, and Guillermo Sapiro. Adversarily learned representations for information obfuscation and inference. In ICML, 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 487, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 542 + ], + "type": "text", + "content": "[4] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 544, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 586 + ], + "type": "text", + "content": "[5] Sayanton V Dibbo. Sok: Model inversion attack landscape: Taxonomy, challenges, and future roadmap. In 2023 IEEE 36th Computer Security Foundations Symposium (CSF), 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 589, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 589, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 589, + 296, + 622 + ], + "type": "text", + "content": "[6] Sayanton V Dibbo, Adam Breuer, Juston Moore, and Michael Teti. Improving robustness to model inversion attacks via sparse coding architectures. ECCV, 2024. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "text", + "content": "[7] Shiwei Ding, Lan Zhang, Miao Pan, and Xiaoyong Yuan. Patrol: Privacy-oriented pruning for collaborative inference against model inversion attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "text", + "content": "[8] Hao Fang, Bin Chen, Xuan Wang, Zhi Wang, and Shu-Tao Xia. Gidf: A generative gradient inversion method with feature domain optimization. In ICCV, 2023. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "type": "text", + "content": "[9] Matt Fredrikson, Somesh Jha, and Thomas Ristenpart. Model inversion attacks that exploit confidence information and basic countermeasures. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security, 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "text", + "content": "[10] Craig Gentry and Shai Halevi. Implementing gentry's fully-homomorphic encryption scheme. In Annual international conference on the theory and applications of cryptographic techniques, 2011. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 173, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 216 + ], + "type": "text", + "content": "[11] Xueluan Gong, Ziyao Wang, Shuaike Li, Yanjiao Chen, and Qian Wang. A gan-based defense framework against model inversion attacks. IEEE Transactions on Information Forensics and Security, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 217, + 555, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 258 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 261, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 293 + ], + "type": "text", + "content": "[13] Gyojin Han, Jaehyun Choi, Haeil Lee, and Junmo Kim. Reinforcement learning-based black-box model inversion attacks. In ICCV, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 294, + 555, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 555, + 336 + ], + "type": "text", + "content": "[14] Zecheng He, Tianwei Zhang, and Ruby B Lee. Model inversion attacks against collaborative inference. In Proceedings of the 35th Annual Computer Security Applications Conference, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 338, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 554, + 370 + ], + "type": "text", + "content": "[15] Zecheng He, Tianwei Zhang, and Ruby B Lee. Attacking and protecting data privacy in edge-cloud collaborative inference systems. IEEE Internet of Things Journal, 2020. 1, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 372, + 555, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 555, + 425 + ], + "type": "text", + "content": "[16] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 426, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 555, + 459 + ], + "type": "text", + "content": "[17] Sy-Tuyen Ho, Koh Jun Hao, Keshigeyan Chandrasegaran, Ngoc-Bao Nguyen, and Ngai-Man Cheung. Model inversion robustness: Can transfer learning help? In CVPR, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 502 + ], + "type": "text", + "content": "[18] Jonghu Jeong, Minyong Cho, Philipp Benz, and Tae-hoon Kim. Noisy adversarial representation learning for effective and efficient image obfuscation. In Uncertainty in Artificial Intelligence, 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 504, + 555, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 555, + 569 + ], + "type": "text", + "content": "[19] Shuaifan Jin, He Wang, Zhibo Wang, Feng Xiao, Jiahui Hu, Yuan He, Wenwen Zhang, Zhongjie Ba, Weijie Fang, Shuhong Yuan, et al. {FaceObfuscator}: Defending deep learning-based privacy attacks with gradient descent-resistant features in face recognition. In 33rd USENIX Security Symposium (USENIX Security 24), 2024. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 570, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 555, + 613 + ], + "type": "text", + "content": "[20] Chiraag Juvekar, Vinod Vaikuntanathan, and Anantha Chandrakasan. {GAZELLE}: A low latency framework for secure neural network inference. In 27th USENIX security symposium (USENIX security 18), 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 614, + 555, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 555, + 647 + ], + "type": "text", + "content": "[21] Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In CVPR, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 555, + 690 + ], + "type": "text", + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[23] Brian Knott, Shobha Venkataraman, Awni Hannun, Shubho Sengupta, Mark Ibrahim, and Laurens van der Maaten." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8761" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "Cryptography: Secure multi-party computation meets machine learning. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 296, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 296, + 128 + ], + "type": "text", + "content": "[24] Chenqi Kong, Anwei Luo, Shiqi Wang, Haoliang Li, Anderson Rocha, and Alex C Kot. Pixel-inconsistency modeling for image manipulation localization. IEEE TPAMI, 2025. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 129, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 129, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 57, + 129, + 294, + 150 + ], + "type": "text", + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 151, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 151, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 57, + 151, + 294, + 172 + ], + "type": "text", + "content": "[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 173, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 216 + ], + "type": "text", + "content": "[27] Ang Li, Jiayi Guo, Huanrui Yang, and Yiran Chen. Deepobfuscator: Adversarial training framework for privacy-preserving image classification. arXiv preprint arXiv:1909.04126, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 217, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 261 + ], + "type": "text", + "content": "[28] Jingtao Li, Adnan Siraj Rakin, Xing Chen, Zhezhi He, Deliang Fan, and Chaitali Chakrabarti. Ressfl: A resistance transfer framework for defending model inversion attack in split federated learning. In CVPR, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "text", + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 305, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 338 + ], + "type": "text", + "content": "[30] Kiwan Maeng, Chuan Guo, Sanjay Kariyappa, and G Edward Suh. Bounding the invertibility of privacy-preserving instance encoding using fisher information. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 338, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 415 + ], + "type": "text", + "content": "[31] Shagufta Mehnaz, Sayanton V Dibbo, Roberta De Viti, Ehsanul Kabir, Björn B Brandenburg, Stefan Mangard, Ninghui Li, Elisa Bertino, Michael Backes, Emiliano De Cristofaro, et al. Are your sensitive attributes private? novel model inversion attribute inference attacks on classification models. In 31st USENIX Security Symposium (USENIX Security 22), 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "text", + "content": "[32] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In 2019 IEEE symposium on security and privacy (SP), 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 460, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 502 + ], + "type": "text", + "content": "[33] Yuxi Mi, Yuge Huang, Jiazhen Ji, Minyi Zhao, Jiaxiang Wu, Xingkun Xu, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using random frequency components. In ICCV, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 503, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 296, + 547 + ], + "type": "text", + "content": "[34] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In CVPR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 296, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 614 + ], + "type": "text", + "content": "[35] Fatemehsadat Mireshghallah, Mohammadkazem Taram, Prakash Ramrakhyani, Ali Jalali, Dean Tullsen, and Hadi Esmaeilzadeh. Shredder: Learning noise distributions to protect inference privacy. In Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 668 + ], + "type": "text", + "content": "[36] Fatemehsadat Mireshghallah, Mohammadkazem Taram, Ali Jalali, Ahmed Taha Taha Elthakeb, Dean Tullsen, and Hadi Esmaeilzadeh. Not all features are equal: Discovering essential features for preserving prediction privacy. In Proceedings of the Web Conference 2021, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[37] Pratyush Mishra, Ryan Lehmkuhl, Akshayaram Srinivasan, Wenting Zheng, and Raluca Ada Popa. Delphi: A cryptographic inference service for neural networks. In 29th USENIX Security Symposium (USENIX Security 20), 2020. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[38] Hong-Wei Ng and Stefan Winkler. A data-driven approach to cleaning large face datasets. In ICIP, 2014. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[39] Bao-Ngoc Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Man Cheung. Label-only model inversion attacks via knowledge transfer. NeurIPS, 2024. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 142, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 555, + 185 + ], + "type": "text", + "content": "[40] Ngoc-Bao Nguyen, Keshigeyan Chandrasegaran, Milad Abdollahzadeh, and Ngai-Man Cheung. Re-thinking model inversion attacks against deep neural networks. In CVPR, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 555, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 242 + ], + "type": "text", + "content": "[41] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 555, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 300 + ], + "type": "text", + "content": "[42] Olga Ohrimenko, Felix Schuster, Cédric Fournet, Aastha Mehta, Sebastian Nowozin, Kapil Vaswani, and Manuel Costa. Oblivious {Multi-Party} machine learning on trusted processors. In 25th USENIX Security Symposium (USENIX Security 16), 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 301, + 555, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 301, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 316, + 301, + 555, + 356 + ], + "type": "text", + "content": "[43] Xiong Peng, Feng Liu, Jingfeng Zhang, Long Lan, Junjie Ye, Tongliang Liu, and Bo Han. Bilateral dependency optimization: Defending against model-inversion attacks. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 357, + 554, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 357, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 316, + 357, + 554, + 401 + ], + "type": "text", + "content": "[44] Yixiang Qiu, Hao Fang, Hongyao Yu, Bin Chen, MeiKang Qiu, and Shu-Tao Xia. A closer look at gan priors: Exploiting intermediate features for enhanced model inversion attacks. In ECCV, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 403, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 555, + 426 + ], + "type": "text", + "content": "[45] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 2009. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 427, + 554, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 554, + 472 + ], + "type": "text", + "content": "[46] Ahmed Salem, Apratim Bhattacharya, Michael Backes, Mario Fritz, and Yang Zhang. {Updates-Leak}: Data set inference and reconstruction attacks in online learning. In 29th USENIX security symposium (USENIX Security 20), 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 473, + 554, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 473, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 473, + 554, + 506 + ], + "type": "text", + "content": "[47] Sunandini Sanyal, Sravanti Addepalli, and R Venkatesh Babu. Towards data-free model stealing in a hard label setting. In CVPR, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 507, + 554, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 507, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 507, + 554, + 540 + ], + "type": "text", + "content": "[48] Nir Shlezinger, Erez Farhan, Hai Morgenstern, and Yonina C Eldar. Collaborative inference via ensembles on the edge. In ICASSP, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 542, + 554, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 542, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 542, + 554, + 586 + ], + "type": "text", + "content": "[49] Lukas Struppek, Dominik Hintersdorf, Antonio De Almeida Correia, Antonia Adler, and Kristian Kersting. Plug & play attacks: Towards robust and flexible model inversion attacks. In ICML, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 588, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 554, + 622 + ], + "type": "text", + "content": "[50] Chandra Thapa, Pathum Chamikara Mahawaga Arachchige, Seyit Camtepe, and Lichao Sun. Splitfed: When federated learning meets split learning. In AAAI, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "text", + "content": "[51] Tom Titcombe, Adam J Hall, Pavlos Papadopoulos, and Daniele Romanini. Practical defences against model inversion attacks for split neural networks. arXiv preprint arXiv:2104.05743, 2021. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "text", + "content": "[52] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8762" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[53] Praneeth Vepakomma, Abhishek Singh, Otkrist Gupta, and Ramesh Raskar. Nopek: Information leakage reduction to share activations in distributed deep learning. In 2020 International Conference on Data Mining Workshops (ICDMW), 2020. 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 184 + ], + "type": "text", + "content": "[54] Sameer Wagh, Shruti Tople, Fabrice Benhamouda, Eyal Kushilevitz, Prateek Mittal, and Tal Rabin. Falcon: Honest-majority maliciously secure framework for private deep learning. Proceedings on Privacy Enhancing Technologies, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 217 + ], + "type": "text", + "content": "[55] Kuan-Chieh Wang, Yan Fu, Ke Li, Ashish Khisti, Richard Zemel, and Alireza Makhzani. Variational model inversion attacks. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 252 + ], + "type": "text", + "content": "[56] Tianhao Wang, Yuheng Zhang, and Ruoxi Jia. Improving robustness to model inversion attacks via mutual information regularization. In AAAI, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 285 + ], + "type": "text", + "content": "[57] Yinggui Wang, Jian Liu, Man Luo, Le Yang, and Li Wang. Privacy-preserving face recognition in the frequency domain. In AAAI, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 296, + 342 + ], + "type": "text", + "content": "[58] Zhibo Wang, Mengkai Song, Zhifei Zhang, Yang Song, Qian Wang, and Hairong Qi. Beyond inferring class representatives: User-level privacy leakage from federated learning. In IEEE INFOCOM 2019-IEEE conference on computer communications, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 386 + ], + "type": "text", + "content": "[59] Song Xia, Wenhan Yang, Yi Yu, Xun Lin, Henghui Ding, Lingyu Duan, and Xudong Jiang. Transferable adversarial attacks on sam and its downstream models. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 431 + ], + "type": "text", + "content": "[60] Taihong Xiao, Yi-Hsuan Tsai, Kihyuk Sohn, Manmohan Chandraker, and Ming-Hsuan Yang. Adversarial learning of privacy-preserving and task-oriented representations. In AAAI, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 475 + ], + "type": "text", + "content": "[61] Mengda Yang, Ziang Li, Juan Wang, Hongxin Hu, Ao Ren, Xiaoyang Xu, and Wenzhe Yi. Measuring data reconstruction defenses in collaborative inference systems. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 296, + 533 + ], + "type": "text", + "content": "[62] Ziqi Yang, Jiyi Zhang, Ee-Chien Chang, and Zhenkai Liang. Neural network inversion in adversarial setting via background knowledge alignment. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 296, + 578 + ], + "type": "text", + "content": "[63] Ziqi Yang, Lijin Wang, Da Yang, Jie Wan, Ziming Zhao, Ee-Chien Chang, Fan Zhang, and Kui Ren. Purifier: Defending data inference attacks via transforming confidence scores. In AAAI, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 296, + 623 + ], + "type": "text", + "content": "[64] Yupeng Yin, Xianglong Zhang, Huanle Zhang, Feng Li, Yue Yu, Xiuzhen Cheng, and Pengfei Hu. Ginver: Generative model inversion attacks against collaborative inference. In Proceedings of the ACM Web Conference 2023, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "type": "text", + "content": "[65] Yi Yu, Song Xia, Xun Lin, Wenhan Yang, Shijian Lu, Yappeng Tan, and Alex Kot. Backdoor attacks against no-reference image quality assessment models via a scalable trigger. In AAAI, 2025. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[66] Xiaojian Yuan, Kejiang Chen, Jie Zhang, Weiming Zhang, Nenghai Yu, and Yang Zhang. Pseudo label-guided model inversion attack via conditional generative adversarial network. In AAAI, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 196 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 116 + ], + "type": "text", + "content": "[67] Yuheng Zhang, Ruoxi Jia, Hengzhi Pei, Wenxiao Wang, Bo Li, and Dawn Song. The secret revealer: Generative model-inversion attacks against deep neural networks. In CVPR, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 150 + ], + "type": "text", + "content": "[68] Xuejun Zhao, Wencan Zhang, Xiaokui Xiao, and Brian Lim. Exploiting explanations for model inversion attacks. In ICCV, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "text", + "content": "[69] Wei Zong, Yang-Wai Chow, Willy Susilo, Joonsang Baek, Jongkil Kim, and Seyit Camtepe. Ipremover: A generative model inversion attack against deep neural network fingerprinting and watermarking. In AAAI, 2024. 1, 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8763" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_content_list.json b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6e4b22aee9d16aa498724caf60c9c9222bc49b65 --- /dev/null +++ b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_content_list.json @@ -0,0 +1,1566 @@ +[ + { + "type": "text", + "text": "Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels", + "text_level": 1, + "bbox": [ + 96, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Quanjiang Li, Tingjin Luo*, Jiahui Liao National University of Defense Technology, China", + "bbox": [ + 297, + 203, + 699, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liquanjiang@nudt.edu.cn,tingjinluo@hotmail.com,liaojiahui19@nudt.edu.cn", + "bbox": [ + 179, + 241, + 812, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Incomplete features and label noise in multi-view multi-label data significantly undermine the reliability and performance, motivating researchers to explore the mechanism of representation and information recovery. However, learning for such dual deficiencies is crucial but rarely studied. In this paper, we propose a theory-inspired Deep Multi-View Multi-Label Learning method with Incomplete Views and Noisy Labels named DMMIvNL to address these problems. Specifically, to promote the synthesis of task-relevant shared information and preserve the distinctiveness of individual features from limited views, we have developed a feature extraction modular based on the information bottleneck theory, and formulated its theoretical upper bound into its objective. Meanwhile, we theoretically prove that minimizing the volume of the transition matrix ensures the statistical consistency with classifier training. Besides, a cycle-consistency estimation principle is proposed in the volume minimization network to improve the recognition stability of multi-label noise. Moreover, leveraging inherent real semantics information and label correlations are employed as model regularization to reduce the risk of excessive noise fitting. Finally, extensive experimental results validate the effectiveness and robustness of our DMMIvNL.", + "bbox": [ + 86, + 323, + 485, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 700, + 220, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advances in feature extraction has shifted data analysis towards multi-view frameworks [27], while growing data attributes drive multi-label learning to offer rich label semantics [13]. Therefore, Multi-view multi-label classification (MvMLC) has emerged to address the intricate demands of processing diverse features and numerous targets. Traditional methods [30, 46] assume the given data has complete views and clean labels, which is violated in practice [41]. On one hand, heterogeneous data from diverse sources frequently encounters missing views due to acquire", + "bbox": [ + 89, + 724, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sition complexities and storage limitations [29, 47]. On the other hand, real-world collections often contain label noise as manual annotation can introduce incorrect labels beyond human knowledge [34]. Therefore, insufficient features and noisy labels severely reduce the classification performance. Regarding MvMLC, many solutions have been proposed to address missing views. iMvWL [37] used weighted matrix factorization to learn a shared subspace and AIMNet [24] leveraged a cross-view attention mechanism to recover missing views. Moreover, two types of label deficiencies have been considered in MvMLC. The problem of Partial Multi-Label (PML) [40] involves both relevant and irrelevant labels, while the problem of Multi-Label Learning with Missing Labels (MLML)[36] occurs when some relevant labels are missing. TFMDD [38] addressed PML with a two-stage manifold disambiguation model and LMVCAT [23] tackled MLML by employing label-guided graphs and category-aware transformers. While existing methods address each annotation problem separately, both issues tend to occur simultaneously in practice. Thus, the challenge of multi-label noise demands substantial attention, where each category suffers from both over-labeled and missing labels.", + "bbox": [ + 511, + 292, + 906, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing research has not yet tackled the combined issues of missing views and multi-label noise, which are prevalent in real-world applications like incomplete and subjective medical diagnostics [31]. In fact, overcoming the simultaneous challenges requires refined representation extraction under missing views and effective classifier leaning with multi-label noise. The core of multi-view learning lies in extracting shared information across different views while preserving unique features of each view [33]. DICNet [22], DIMC [39] and SIP [26] concentrated solely on shared subspace and neglected distinctive characteristics of individual views. MDDRL [16] and MTD [25] attempted to extract both shared and specific representations with minimal interactions. However, their reliance on geometric orthogonality or distance metrics restricted them to capture only linear patterns, overlooking the complex nature of inter-view relations. Therefore, constraining feature semantics at the intrinsic information level is crucial for obtaining high-level", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20706", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representations with precise distinction. Most approaches for handling multi-label noise utilize strategies like early stopping [14] and cyclical learning rates [3]. These methods often lack stability and require dataset-specific adjustments, making them insufficiently adaptive. ENTMLC [19] employed label correlations to identify the noise transition matrix for recovering clean labels. However, the decoupling of transition matrix acquisition and classifier training causes error accumulation, limiting the practical utility. To circumvent the uncertainty inherent in those strategies and the inefficiency of stepwise noise learning, a statistically consistent estimation framework should be developed, which would allow for direct training the optimal classifier from noisy labels and achieving strong robustness.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these problems, we propose a theory-inspired method with deep networks named DMMIvNL. Specifically, a mutual information model is developed to enhance synergy of shared features, suppress interactions from specific features, and impose orthogonality between shared and specific features of each view. The upper bound of the model is derived through the variational autoencoder framework, with instance-level contrastive operations applied to simulate diverse joint and marginal distributions. Moreover, the ideal classifier is trained on noisy data by minimizing the volume of the noise transition matrix and the process is proven to converge to a unique and globally optimal solution under the sufficiently scattered assumption. The difficulty in guaranteeing the assumption and relying exclusively on noisy class-posterior estimation promotes the design of a cycle-consistency estimation method, which integrates the volume maximization of clean class-posterior. Finally, the real semantics embedded in noisy data and hidden label correlation information are employed as model regularization to reduce the risk of noise overfitting. Our main contributions are summarized as follows:", + "bbox": [ + 91, + 304, + 483, + 619 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- This is the first multi-view multi-label learning framework capable of handling incomplete views and noisy labels, including both over-labeled and missing labels.", + "- Our mutual information based model extracts semantically distinct representations from limited views, with its upper bound established through variational derivation.", + "- We theoretically prove the statistical consistency and joint convergence between minimizing the noise transition matrix volume for each category and classifier training. Besides, a cycle-consistency loss framework is designed to enhance transition matrix estimation." + ], + "bbox": [ + 89, + 622, + 482, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Methodology", + "text_level": 1, + "bbox": [ + 89, + 805, + 225, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Problem Definition", + "text_level": 1, + "bbox": [ + 89, + 830, + 272, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The multi-view dataset $\\{\\mathbf{X},\\mathbf{Y}\\}$ contains $N$ samples, where $\\mathbf{X}_i$ consists of $V$ views with dimension $d_v$ , i.e., $\\mathbf{X}_i = \\{\\pmb{x}_i^{(v)}\\in \\mathbb{R}^{d_v}\\}_{v = 1}^V$ . The label matrix $\\mathbf{Y}\\in \\{0,1\\}^{N\\times C}$ corre", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sponds to $C$ categories, with $Y_{i,j} = 1$ if sample $i$ belongs to class $j$ , and $Y_{i,j} = 0$ otherwise. $\\mathbf{W} \\in \\mathbb{R}^{N \\times V}$ is an indicator matrix, where $W_{i,j} = 1$ indicates the $j$ -th view is available for sample $i$ , and $W_{i,j} = 0$ means missing. Since precise annotation is hard to obtain, the training dataset in practice is $\\{\\mathbf{X}, \\bar{\\mathbf{Y}}\\}$ , where $\\bar{\\mathbf{Y}}$ denotes noisy labels. Thus, we denote the clean and noisy labels of the $j$ -th class as $\\mathbf{Y}^j$ and $\\bar{\\mathbf{Y}}^j$ , related by a noise transition matrix $\\mathbf{T}^j \\in [0,1]^{2 \\times 2}$ , $j \\in [C]$ . When only noisy samples are available and without further constraints, instance-dependent transition matrix becomes non-identifiable [42]. Therefore, we employ a general assumption that the noise transition matrix is class-dependent and instance-independent, i.e., $P(\\bar{\\mathbf{Y}}^j = k \\mid \\mathbf{Y}^j = i, \\mathbf{X} = x) = P(\\bar{\\mathbf{Y}}^j = k \\mid \\mathbf{Y}^j = i) = T_{i,k}^j$ .", + "bbox": [ + 511, + 90, + 906, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Feature Extraction via Information Bottleneck", + "text_level": 1, + "bbox": [ + 511, + 316, + 906, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deriving effective representations from raw data aids in reducing noise and redundancy [17]. Consequently, both task-relevant shared semantics and view-invariant characteristic features should be explored to enhance information processing efficiency. Let $z^v$ and $s^v$ denote the $D$ -dimensional shared and specific representation extracted from each view and the goal is to ensure consistency of the shared features across multiple views and preserve the unique contributions of each view. Considering different interaction patterns between representations from $x^{v^*}$ and $x^v$ , mutual information $I(z^{v^*}|x^{v^*};z^v|x^v)$ should be maximized to capture only universally shared information, while $I(s^{v^*}|x^{v^*};s^v|x^v)$ should be minimized to preserve only individual attributes. Besides, the shared and specific features should be as orthogonal as possible to ensure absolute distinction in information content [35], i.e., $I(z^v |x^v;s^v |x^v) = 0$ . Thus, the constrained optimization problem based on the information bottleneck theory [8] can be formulated as", + "bbox": [ + 511, + 339, + 908, + 611 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\max \\frac {1}{V} \\sum_ {v = 1} ^ {V} \\sum_ {v ^ {*} \\neq v} ^ {V} \\left(I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) - I \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 625, + 906, + 667 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad \\frac {1}{V} \\sum_ {v = 1} ^ {V} I \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) = 0. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 671, + 903, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By applying the Lagrange multiplier method, the problem (1) can be equally converted as", + "bbox": [ + 511, + 726, + 903, + 756 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max \\frac {1}{V} \\sum_ {v = 1} ^ {V} \\sum_ {v ^ {*} \\neq v} ^ {V} \\left(I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) - I \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right)\\right) \\\\ - \\frac {\\beta}{V} \\sum_ {v = 1} ^ {V} I \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 768, + 906, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\beta \\geq 0$ is the Lagrange multiplier to balance the efficiency and compactness of information [26]. For the first", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20707", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4e7cf2f6120729b621481246e5f816945ca06cbe7ca32b7ec89ba18c2e7120f8.jpg", + "image_caption": [ + "Figure 1. The main framework of our proposed DMMIvNL. Different shapes signify different samples." + ], + "image_footnote": [], + "bbox": [ + 178, + 90, + 820, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "term, we can obtain the following lower bound:", + "bbox": [ + 89, + 439, + 406, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\\\ = \\iint p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) \\log \\frac {p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v} , \\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right)}{p \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right)} d \\boldsymbol {z} ^ {v ^ {*}} d \\boldsymbol {z} ^ {v} \\\\ \\geq \\int \\int p (\\boldsymbol {z} ^ {v} | \\boldsymbol {z} ^ {v ^ {*}}) p (\\boldsymbol {z} ^ {v ^ {*}} | \\boldsymbol {x} ^ {v ^ {*}}) \\log p (\\boldsymbol {z} ^ {v} | \\boldsymbol {z} ^ {v ^ {*}}) d \\boldsymbol {z} ^ {v ^ {*}} d \\boldsymbol {z} ^ {v}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 460, + 503, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p(\\cdot)$ denotes the probability density function. Since computing $p(z^v |z^{v^*})$ is intractable, we approximate it using a stochastic variational distribution $q(z^{v}|z^{v^{*}})$ . Then we can obtain a new variational lower bound for maximization:", + "bbox": [ + 89, + 559, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\\\ \\geq \\int \\int p (\\boldsymbol {z} ^ {v} | \\boldsymbol {x} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}} | \\boldsymbol {x} ^ {v ^ {*}}) \\log q (\\boldsymbol {z} ^ {v} | \\boldsymbol {z} ^ {v ^ {*}}) d \\boldsymbol {z} ^ {v ^ {*}} d \\boldsymbol {z} ^ {v} \\\\ = D _ {K L} ^ {\\dagger} \\left(p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) \\| q \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {z} ^ {v ^ {*}}\\right)\\right) = D _ {K L} ^ {\\dagger} \\left(\\boldsymbol {z} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}}\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 625, + 482, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D_{KL}(\\cdot ||\\cdot)$ means the Kullback-Leibler divergence and $D_{KL}^{\\dagger}(\\cdot ||\\cdot)$ is a pseudo-symbol for notational consistency. For the middle term, the upper bound can be obtained by introducing another variational distribution $q(s^{v}|s^{v^{*}})$ :", + "bbox": [ + 89, + 714, + 483, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\\\ \\leq D _ {K L} \\left(p \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}, \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\| p \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) q \\left(\\boldsymbol {s} ^ {v} \\mid \\boldsymbol {s} ^ {v ^ {*}}\\right)\\right) \\tag {4} \\\\ = D _ {K L} ^ {\\left(\\mathbf {s} ^ {v}, \\mathbf {s} ^ {v ^ {*}}\\right)}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 785, + 482, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, the upper bound for the last term is", + "bbox": [ + 89, + 858, + 392, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\leq D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {z} ^ {v}\\right)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 881, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Combining Eqs. (3), (4) and (5), our optimization objective becomes minimizing the upper bound of the problem (2):", + "bbox": [ + 511, + 439, + 906, + 469 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I B} = \\frac {1}{V} \\sum_ {v = 1} ^ {V} \\sum_ {v ^ {*} \\neq v} ^ {V} \\left(- D _ {K L} ^ {\\dagger \\left(\\boldsymbol {z} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}}\\right)} + D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {s} ^ {v ^ {*}}\\right)}\\right) + \\frac {\\beta}{V} \\sum_ {v = 1} ^ {V} D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {z} ^ {v}\\right)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 473, + 937, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning [21] is used to compute various distributions. Specifically, representations are treated as probability vectors over $D$ classes via a Softmax activation function, and the joint distribution matrix is obtained by", + "bbox": [ + 511, + 530, + 906, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {P} ^ {(v, v ^ {*})} = \\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {z} _ {i} ^ {(v)}\\right) ^ {T} \\boldsymbol {z} _ {i} ^ {(v ^ {*})}, \\boldsymbol {Q} ^ {(v, v ^ {*})} = \\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {s} _ {i} ^ {(v)}\\right) ^ {T} \\boldsymbol {s} _ {i} ^ {(v ^ {*})}.\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 595, + 911, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the strong interaction between $z^v$ and $z^{v^*}$ , the variational distribution $q(z^v |z^{v^*})$ can be estimated by the obtained conditional distribution. Given the extremely weak correlation between $s^v$ and $s^{v^*}$ , as well as $z^v$ and $s^v$ , the distribution $q(s^v /s^{v^*})$ and $q(z^{v} / s^{v})$ can be approximated by the corresponding marginal distribution. Subsequently, each term in Eq. (6) can be expressed as", + "bbox": [ + 511, + 643, + 906, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} D _ {K L} ^ {\\dagger \\left(\\boldsymbol {z} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}}\\right)} = \\sum_ {d = 1} ^ {D} \\sum_ {d ^ {\\prime} = 1} ^ {D} \\boldsymbol {P} _ {d, d ^ {\\prime}} ^ {(v, v ^ {*})} \\log \\frac {\\boldsymbol {P} _ {d , d ^ {\\prime}} ^ {(v , v ^ {*})}}{\\boldsymbol {P} _ {d} ^ {(v)}} \\\\ D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {s} ^ {v ^ {*}}\\right)} = \\sum_ {d = 1} ^ {D} \\sum_ {d ^ {\\prime} = 1} ^ {D} \\boldsymbol {Q} _ {d, d ^ {\\prime}} ^ {(v, v ^ {*})} \\log \\frac {\\boldsymbol {Q} _ {d , d ^ {\\prime}} ^ {(v , v ^ {*})}}{\\left(\\boldsymbol {Q} _ {d} ^ {(v)}\\right) ^ {\\alpha} \\boldsymbol {Q} _ {d ^ {\\prime}} ^ {(v ^ {*})}} \\\\ D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {z} ^ {v}\\right)} = \\sum_ {d = 1} ^ {D} \\sum_ {d ^ {\\prime} = 1} ^ {D} \\boldsymbol {M} _ {d, d ^ {\\prime}} ^ {\\left(z ^ {v}, s ^ {v}\\right)} \\log \\frac {\\boldsymbol {M} _ {d , d ^ {\\prime}} ^ {\\left(z ^ {v} , s ^ {v}\\right)}}{\\left(\\boldsymbol {M} _ {d} ^ {\\left(z ^ {v}\\right)}\\right) ^ {\\alpha} \\boldsymbol {M} _ {d ^ {\\prime}} ^ {\\left(s ^ {v}\\right)}}, \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 753, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20708", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the marginal distributions $P_{d}^{(v)}$ and $P_{d^{\\prime}}^{(v^{*})}$ can be obtained by summing over the $d$ -th row and the $d^{\\prime}$ -th column of the joint distribution matrix $P^{(v,v^{*})}$ , with other symbols defined as the same. In this conversion, we let $q(\\boldsymbol{s}^v /\\boldsymbol{s}^{v^*}) = (Q^{(v)})^\\alpha$ and $q(z^{v} / s^{v}) = (M^{(z^{v})})^{\\alpha}$ , where $\\alpha$ is a balance factor to preserve crucial information and ensure model stability [16]. By introducing the mask matrix $\\pmb{W}$ to mitigate the impact of unavailable instances, we can perform cross-view fusion to acquire the information-enriched shared and specific representation:", + "bbox": [ + 89, + 88, + 483, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {Z}} _ {i,:} = \\sum_ {v = 1} ^ {V} \\frac {\\boldsymbol {z} _ {i , :} ^ {(v)} \\boldsymbol {W} _ {i , v}}{\\sum_ {v} \\boldsymbol {W} _ {i , v}}, \\quad \\bar {\\boldsymbol {S}} _ {i,:} = \\sum_ {v = 1} ^ {V} \\frac {\\boldsymbol {s} _ {i , :} ^ {(v)} \\boldsymbol {W} _ {i , v}}{\\sum_ {v} \\boldsymbol {W} _ {i , v}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 252, + 482, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Departing from traditional addition or concatenation, we conduct feature interaction by enhancing shared information with proprietary data [25]:", + "bbox": [ + 89, + 304, + 483, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {X}} _ {i, j} = \\theta_ {s} \\left(\\bar {\\boldsymbol {Z}} _ {i, j}\\right) \\cdot \\bar {\\boldsymbol {S}} _ {i, j}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 358, + 482, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{s}$ is the sigmoid activation function and $\\hat{\\pmb{X}}\\in \\mathbb{R}^{N\\times D}$ is the final fusion representation.", + "bbox": [ + 89, + 388, + 482, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Identify Multi-Label Noise by Volume Minimization Network", + "text_level": 1, + "bbox": [ + 89, + 429, + 483, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The success of deep learning relies on large-scale accurately labeled data, with performance significantly dropping under noisy data [12], making noise identification crucial for network training [2]. In multi-class tasks, many statistically consistent methods based on the transition matrix have been developed [28, 32], which guarantees the classifier learned from noisy data can converge to the optimal classifier under the clean risk. The anchor-point assumption [28] is commonly utilized to determine the transition matrix, which postulates the existence of samples that belong to each class with probability one. However, implementing this approach in multi-label learning is challenging due to the difficulty of pinpointing suitable anchor samples for each label. Motivated by [20], we capitalize on the geometric properties of the transition matrix to seek a breakthrough. The transition matrix for the $j$ -th label is defined as", + "bbox": [ + 89, + 465, + 483, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} ^ {j} = \\left( \\begin{array}{l l} P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 0 \\mid \\boldsymbol {Y} ^ {j} = 0\\right) & P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 0 \\mid \\boldsymbol {Y} ^ {j} = 1\\right) \\\\ P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 1 \\mid \\boldsymbol {Y} ^ {j} = 0\\right) & P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 1 \\mid \\boldsymbol {Y} ^ {j} = 1\\right) \\end{array} \\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 715, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an instance $\\hat{\\pmb{x}}$ , the noisy class-posterior for the $j$ -th label is $P(\\bar{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}}) = [P(\\bar{Y}^j = 0|\\mathbf{X} = \\hat{\\pmb{x}}), P(\\bar{Y}^j = 1|\\mathbf{X} = \\hat{\\pmb{x}})]^T$ and the clean class-posterior is $P(\\mathbf{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}}) = [P(\\mathbf{Y}^j = 0|\\mathbf{X} = \\hat{\\pmb{x}}), P(\\mathbf{Y}^j = 1|\\mathbf{X} = \\hat{\\pmb{x}})]^T$ . Thus, the relationship $P(\\bar{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}}) = \\pmb{T}^j P(\\mathbf{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}})$ holds. Since $\\sum_{i=0}^{1} P(\\mathbf{Y} = i|\\mathbf{X} = \\hat{\\pmb{x}}) = 1$ , $P(\\bar{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}})$ is then a convex combination of the columns of $\\pmb{T}^j$ , which implies that the noisy class-posterior is enclosed within the simplex $\\mathrm{Sim}\\{\\pmb{T}^j\\}$ . To reduce noise identification errors, it", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is essential to recover $\\operatorname{Sim}\\{\\pmb{T}^j\\}$ such that it tightly encompasses all sample posterior probability points, thereby making volume minimization of $\\operatorname{Sim}\\{\\pmb{T}^j\\}$ the primary goal. For synchronized learning, we define a trainable $\\hat{T}^j \\in [0,1]^{2 \\times 2}$ with $\\sum_{i=1}^{2} \\hat{T}^j{}_{i,k} = 1$ , and construct a multi-label classifier $h_\\theta$ with parameters $\\theta$ . The optimization of $\\hat{T}^j$ and $h_\\theta(\\hat{x})$ involves two objectives: the discrepancy between $\\hat{T}^j(h_\\theta(\\hat{x})_j, 1 - h_\\theta(\\hat{x})_j)^T$ and $P(\\bar{Y}^j | X = \\hat{x})$ and the volume of each simplex spanned by the columns of $\\hat{T}^j$ . Therefore, the following model to jointly learn the transition matrix and classifier is proposed:", + "bbox": [ + 511, + 90, + 906, + 267 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\min _ {\\hat {\\boldsymbol {T}} ^ {j}} \\sum_ {j = 1} ^ {C} \\operatorname {v o l} \\left(\\hat {\\boldsymbol {T}} ^ {j}\\right) \\\\ s. t. \\hat {\\boldsymbol {T}} ^ {j} \\left[ \\begin{array}{c} h _ {\\theta} (\\hat {\\boldsymbol {x}}) _ {j} \\\\ 1 - h _ {\\theta} (\\hat {\\boldsymbol {x}}) _ {j} \\end{array} \\right] = P \\left(\\bar {\\boldsymbol {Y}} ^ {j} \\mid X = \\hat {\\boldsymbol {x}}\\right), \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 273, + 906, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{vol}(\\hat{T}^j)$ denotes a measure to the volume of $\\operatorname{Sim}\\{\\hat{T}^j\\}$ . In the absence of anchor points, the class-posterior probability should incorporate substantial information to uniquely determine the minimum-volume simplex, with true sample probability points of each class being distinguishable. Therefore, the sufficiently scattered assumption [9] is introduced to limit the occurrence of anomalous clean class-posterior and aid in theoretically establishing that addressing problem (11) will drive $\\hat{T}^j$ to converge to $\\pmb{T}^j$ and $h_\\theta (\\hat{\\pmb{x}})$ to accurately approximate $P(\\mathbf{Y}|\\mathbf{X} = \\hat{\\mathbf{x}})$ :", + "bbox": [ + 511, + 362, + 906, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. If each $P(\\mathbf{Y}^j | \\mathbf{X})$ is sufficiently scattered, then $\\mathbf{T}_{\\star}^{j} = \\mathbf{T}^{j}$ and $(h_{\\theta_{\\star}}(\\hat{\\mathbf{x}})_j, 1 - h_{\\theta_{\\star}}(\\hat{\\mathbf{x}})_j)^T = P(\\mathbf{Y}^j | \\mathbf{X} = \\hat{\\mathbf{x}})$ ( $j = 1, 2, \\dots, C$ ) must hold, where $(\\mathbf{T}_{\\star}^{1}, \\mathbf{T}_{\\star}^{2}, \\dots, \\mathbf{T}_{\\star}^{C}, \\theta_{\\star})$ are the optimal solutions of the problem (11).", + "bbox": [ + 511, + 523, + 905, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4. Cycle-Consistency Estimation and Model Regularization", + "text_level": 1, + "bbox": [ + 511, + 590, + 903, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In multi-label learning, imbalanced label distributions [44] and inter-class correlations [7] can compress class-posterior probability points and limit their scatter. Besides, the volume approximation of $T^j$ in the problem (11) is highly sensitive to the noisy class-posterior, making it suffer from large estimation errors in noisy environments [18]. The failure of the sufficiently scattered assumption, along with the excessive reliance on the noisy class-posterior, compromises the effectiveness of the problem (11). To enhance transition matrix estimation, we develop an adaptive loss framework [4] that incorporates the volume maximization of the clean class-posterior, which leverages additional information to mitigate bias in single-target decision-making. Based on the equation $P(\\bar{Y}^j | \\mathbf{X}) = T^j P(\\mathbf{Y}^j | \\mathbf{X})$ , where $P(\\bar{Y}^j | \\mathbf{X})$ can be derived from the noisy data, minimizing the volume of $T^j$ is equivalent to imposing a volume constraint on $P(\\mathbf{Y}^j | \\mathbf{X})$ . Through the matrix $\\tilde{T}^j$ (an approximate inverse of $T^j$ ), we project the simplex formed in the", + "bbox": [ + 511, + 628, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20709", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noisy space onto the clean class-posterior space to obtain $P^{\\prime}(\\mathbf{Y}|\\mathbf{X})$ , which will make up the largest volume. Then we design a cycle-consistency optimization strategy to enhance both the volume constraint on $\\pmb{T}^{j}$ and maximization of $P(\\mathbf{Y}|\\mathbf{X})$ , in which the estimated noisy class-posterior matches the provided noisy label $\\bar{\\mathbf{Y}}^j$ , the backward estimated clean class-posterior $P^{\\prime}(\\mathbf{Y}|\\mathbf{X})$ aligns with the predicted clean-class posterior and the forward-backward consistency term $\\tilde{T}^j T^j P(\\mathbf{Y}|\\mathbf{X})$ coheres with $P(\\mathbf{Y}|\\mathbf{X})$ .", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, we employ forward estimation to simultaneously optimize the transition matrix $T^{j}$ and classifier $h_{\\theta}$ . By leveraging the cross-entropy loss and reporting the output as $F^{ij} = (h_{\\theta}(\\hat{\\boldsymbol{x}}_i)_j, 1 - h_{\\theta}(\\hat{\\boldsymbol{x}}_i)_j)^T$ , we aim to minimize the approximation error between the inferred noisy class-posterior probability $T^{j}F^{ij}$ and the given noisy label $\\bar{Y}^{j}$ :", + "bbox": [ + 89, + 227, + 480, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {1} (\\theta , \\boldsymbol {T} ^ {1}, \\dots , \\boldsymbol {T} ^ {C}) = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} (\\bar {Y} _ {i, j} \\log (\\boldsymbol {T} _ {1} ^ {j} \\boldsymbol {F} ^ {i j}) \\\\ + \\left(1 - \\bar {Y} _ {i, j}\\right) \\log \\left(1 - \\mathbf {T} _ {2} ^ {j} \\mathbf {F} ^ {i j}\\right)). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 330, + 467, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Secondly, the Sigmoid function is employed to convert the one-hot label into the noisy class-posterior probability, which is then normalized along the label dimension to capture the soft labels $\\bar{\\mathbf{Y}}^{sn}$ . Denoting the outcome $(\\bar{\\mathbf{Y}}_{i,j}^{sn}, 1 - \\bar{\\mathbf{Y}}_{i,j}^{sn})^T$ as $\\bar{\\mathbf{F}}^{ij}$ , we utilize $\\tilde{T}^j$ to project $\\bar{\\mathbf{F}}^{ij}$ onto the clean class-posterior space. Thus, we minimize the following cross-entropy loss between the backward estimated clean class-posterior $\\tilde{T}^j \\bar{\\mathbf{F}}^{ij}$ and the true class-posterior probability estimation $h_\\theta(\\hat{\\boldsymbol{x}}_i)_j$ :", + "bbox": [ + 89, + 405, + 483, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {2} (\\theta , \\tilde {\\boldsymbol {T}} ^ {1}, \\dots , \\tilde {\\boldsymbol {T}} ^ {C}) = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} (h _ {\\theta} (\\hat {\\boldsymbol {x}} _ {i}) _ {j} \\log (\\tilde {\\boldsymbol {T}} _ {1} ^ {j} \\bar {\\boldsymbol {F}} ^ {i j}) \\\\ + \\left(1 - h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j}\\right) \\log \\left(1 - \\tilde {\\boldsymbol {T}} _ {2} ^ {j} \\bar {\\boldsymbol {F}} ^ {i j}\\right)). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 555, + 482, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At this stage, the transition matrix can be derived from both the forward and backward ways by optimizing $\\mathcal{L}_1$ and $\\mathcal{L}_2$ . Thirdly, we should establish the cycle-consistency loss by minimizing the difference between $\\tilde{T}^j T^j F^{ij}$ and $h_\\theta (\\hat{x}_i)_j$ to reinforce the invertible relationship of $T^j$ and $\\tilde{T}^j$ :", + "bbox": [ + 89, + 628, + 482, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {3} (\\theta) = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} \\left(h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j} \\log \\left(\\tilde {\\boldsymbol {T}} _ {1} ^ {j} \\boldsymbol {T} ^ {j} \\boldsymbol {F} ^ {i j}\\right) \\right. \\tag {12} \\\\ + \\left(1 - h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j}\\right) \\log \\left(1 - \\bar {\\boldsymbol {T}} _ {2} ^ {j} \\boldsymbol {T} ^ {j} \\boldsymbol {F} ^ {i j}\\right)), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 720, + 480, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we only optimize the network parameters $\\theta$ while keeping $T^j$ and $\\tilde{T}^j$ constant in order to maintain their consistency and focus on learning an optimal classifier. Therefore, the loss for identifying the transition matrix and learning the clean class-posterior is $\\mathcal{L}_{IB} = \\mathcal{L}_1 + \\mathcal{L}_2 + \\mathcal{L}_3$ .", + "bbox": [ + 89, + 794, + 480, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate overfitting on noisy data, we design model regularization by leveraging the partial fidelity of noisy la", + "bbox": [ + 89, + 869, + 480, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bels and label correlations. We allow the classifier to directly access raw labels and obtain the classification loss:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {c} = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} \\left(\\bar {Y} _ {i, j} \\log \\left(h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j}\\right)\\right) \\tag {13} \\\\ + \\left(1 - \\bar {Y} _ {i, j}\\right) \\log \\left(1 - h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j})\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 133, + 903, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The construction of $\\mathcal{L}_c$ is motivated by two key considerations. On one hand, it is essential to harness the underlying true information embedded in the noisy data to improve the inefficient classifier. On the other hand, strengthening the model memory with original noise helps minimize the risk of misclassifying clean samples as noisy ones. Since label correlation is a crucial factor for multi-label learning, we minimize the discrepancy in the label relationships reflected by the estimated clean class posterior and the classifier output to reduce classification bias. By defining the clean class-posterior matrix of the positive labels as $P^c$ , with elements given by $P_{i,j}^{c} = \\tilde{T}_{1}^{j}\\bar{F}^{ij}$ , we can build the following correlation-aligned loss driven by label co-occurrence [43]:", + "bbox": [ + 511, + 203, + 903, + 400 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\frac {1}{C ^ {2}} \\left\\| \\left(\\boldsymbol {P} ^ {c}\\right) ^ {T} \\boldsymbol {P} ^ {c} - \\left(h _ {\\theta} (\\hat {\\boldsymbol {X}})\\right) ^ {T} h _ {\\theta} (\\hat {\\boldsymbol {X}}) \\right\\| _ {F} ^ {2}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 411, + 903, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, our model regularization can be formulated as $\\mathcal{L}_R = \\mathcal{L}_c + \\mathcal{L}_r$ . The total training loss of our model is", + "bbox": [ + 511, + 455, + 903, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {I B} + \\lambda_ {1} \\mathcal {L} _ {I D} + \\lambda_ {2} \\mathcal {L} _ {R}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 497, + 903, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are penalty coefficients to balance the effects of different losses when handling noisy labels.", + "bbox": [ + 511, + 523, + 903, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 511, + 566, + 643, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 511, + 590, + 712, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. Following [23, 45], our experiments incorporate six widely recognized multi-view multi-label datasets, i.e., Yeast [11], Corel 5k [5], VOC 2007 [6], Espgame [1], IAPR TC-12 [10], and MIR FLICKR [15]. Similar to [25], we employ six diverse metrics to comprehensively evaluate each method, i.e., Hamming Loss (HL), Ranking Loss (RL), OneError (OE), Coverage (Cov), Average Precision (AP), and Area Under Curve (AUC). To facilitate performance comparison, we present 1-HL, 1-RL, 1-OE, and 1-Cov values in our report, with higher values consistently indicating better performance.", + "bbox": [ + 511, + 613, + 903, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Methods. To validate the effectiveness of DMMIvNL, we perform a comparative analysis against eight state-of-the-art methods i.e., iMVWL, TM3L, AlM-Net, DICNet, DIMC, MTD, SIP, ENTMLC. Since the first seven methods are designed to handle incomplete multiview multi-label problems, we train them directly on noisy data without modifications. As ENTMLC specializes in addressing multi-label noise, we tackle missing views by fill", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20710", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c078ef559af90c5d5711894517a8926cc26cfb15b7d5ed4cc3774a349d6b956a.jpg", + "table_caption": [ + "Table 1. Experimental results of nine methods on all datasets with PER=50% and ρ = 30%. The best result on each row is bolded and the second-best result is underlined. The average ranking on the six metrics is shown as Ave.R." + ], + "table_footnote": [], + "table_body": "
DataMETDICNetDIMCMTDSIPAIMNetiMVWLTM3LNTMLCSDMMIvNL
Yeast1-HL71.54±1.8174.94±1.5676.40±0.3371.24±1.4973.03±1.7966.94±0.4763.89±0.2969.88±0.1677.64±0.47
1-RL73.81±0.9174.82±2.3377.08±0.7571.30±1.6773.95±0.4372.83±1.0565.72±0.3771.40±3.1480.18±0.37
1-OE71.86±3.7470.95±5.3775.33±0.4861.94±6.7470.66±2.7963.91±2.2956.49±0.7773.28±0.4373.69±1.22
1-Cov41.68±2.4844.12±3.1045.53±1.9539.04±2.1641.19±0.3141.50±1.4534.93±0.7335.55±1.3750.77±0.84
AP65.61±1.3466.91±2.0968.87±0.8261.78±2.2565.55±1.1863.74±0.9256.90±0.3066.69±0.6072.08±0.34
AUC75.33±0.7376.34±2.3578.36±0.5473.42±1.5375.46±0.4574.85±0.9767.92±0.2873.33±2.4481.03±0.55
Ave.R4.503.331.837.174.836.338.837.001.17
Corel 5k1-HL96.37±0.6896.08±0.9998.65±0.0098.59±0.0498.66±0.0079.41±0.0597.23±0.6798.57±0.1798.68±0.02
1-RL60.14±0.8461.00±0.8767.26±0.5965.32±1.4163.37±0.6564.53±0.5552.01±0.3965.92±1.2372.27±0.57
1-OE6.30±0.895.44±1.0521.22±1.0612.88±4.2824.00±0.7310.29±2.4620.18±0.332.30±1.6725.20±0.78
1-Cov32.29±0.8133.31±0.9637.53±0.8836.69±2.0341.62±0.4736.09±1.1233.03±0.0235.19±0.9644.35±1.10
AP6.72±0.406.80±0.2915.22±0.6510.54±1.8618.49±0.2110.18±0.914.48±0.098.25±0.9919.18±0.45
AUC60.30±0.8551.74±1.3067.63±0.5965.69±1.5470.37±0.3764.81±0.7658.99±0.1165.97±1.2772.42±0.47
Ave.R7.677.672.834.332.676.007.335.501.00
VOC 20071-HL89.48±1.7089.72±1.3192.85±0.0592.00±0.1792.23±0.3274.95±0.2084.92±0.2092.66±0.0292.67±0.06
1-RL64.72±0.8965.70±0.7666.04±1.1063.69±1.7263.56±0.9267.03±0.8156.30±0.5866.03±1.9572.07±0.84
1-OE26.14±3.4131.28±3.1939.21±0.2131.84±4.2135.22±4.8337.47±1.5326.53±0.7139.35±1.8337.16±1.90
1-Cov58.13±0.7858.76±1.2059.06±0.9356.94±1.5256.55±0.8959.19±0.8153.99±0.3958.02±1.9765.64±1.21
AP34.45±1.3837.52±1.1440.18±0.6536.57±2.0337.16±2.1739.83±0.1827.49±0.3541.41±0.6843.86±1.15
AUC67.92±0.8068.84±1.0068.94±0.8766.52±1.4165.81±0.4969.76±0.6862.29±0.3968.55±2.0874.65±1.01
Ave.R6.835.172.506.506.503.678.673.501.67
Espauge1-HL97.88±0.1297.50±0.5698.40±0.0098.26±0.0198.41±0.0079.07±0.0494.75±4.9697.78±0.1898.25±0.01
1-RL62.26±0.3263.18±0.5063.94±0.3062.14±0.8464.67±0.6662.57±0.9854.05±0.6466.66±0.6069.79±0.47
1-OE14.49±1.2015.75±1.0814.97±0.7013.63±0.9118.75±0.2410.63±0.5417.27±7.211.48±1.1931.77±1.51
1-Cov29.73±0.3130.41±0.4933.06±0.3031.95±0.6634.65±0.6030.14±1.0125.73±0.1733.37±0.4335.80±0.55
AP9.83±0.3510.78±0.3211.28±0.1310.19±0.4713.38±0.299.66±0.555.82±0.096.77±0.7819.13±0.58
AUC62.15±0.3263.35±0.5664.12±0.2362.20±0.9066.28±0.7062.65±0.9156.03±0.0666.64±0.7969.86±0.54
Ave.R6.675.173.675.832.177.177.835.001.50
IAPR TC-121-HL97.54±0.0497.67±0.1898.00±0.0097.93±0.0398.00±0.0079.41±0.0288.22±1.2497.18±0.1498.04±0.02
1-RL61.86±0.3265.48±2.4566.90±0.1465.17±0.9169.20±0.3370.78±0.5755.79±0.9169.68±0.7372.66±0.48
1-OE14.86±0.6418.27±1.6829.51±0.8127.42±5.4433.89±4.6723.45±4.1615.02±9.561.49±1.3934.33±1.37
1-Cov25.98±0.4328.56±1.9529.24±0.3128.12±1.1631.06±0.5034.23±1.0823.23±0.2933.67±0.4635.40±0.53
AP9.18±0.2412.07±1.5715.10±0.1013.42±0.7817.37±0.8914.88±0.636.25±0.077.32±0.4319.66±0.07
AUC62.12±0.2465.52±2.3866.95±0.0965.38±0.6769.48±0.3170.87±0.6057.53±0.1169.75±0.8872.56±0.50
Ave.R7.505.833.835.673.174.008.505.501.00
MIR FLICKR1-HL86.63±0.2585.90±0.8787.46±0.0786.24±0.3187.12±0.1973.17±0.1079.90±0.1386.24±0.9487.61±0.09
1-RL75.11±0.4376.91±1.6176.23±0.2073.11±0.5074.97±0.9575.97±0.8860.07±0.1979.74±1.6882.34±0.75
1-OE44.98±0.7643.88±1.7636.58±0.7335.51±1.1935.78±1.0736.97±2.9642.44±0.3738.67±6.7261.34±0.99
1-Cov46.06±0.5948.98±2.7949.39±0.2346.30±0.4947.18±3.1348.15±2.1435.08±0.0953.12±4.1256.94±0.98
AP42.55±0.4244.11±0.7940.80±0.2038.50±0.1840.71±0.2242.18±0.5629.72±0.3445.69±2.7354.89±0.94
AUC73.96±0.4175.67±1.9175.42±0.1772.27±0.5873.97±1.3874.62±0.7262.80±0.1479.16±1.2881.14±0.93
Ave.R5.173.834.337.506.175.838.003.171.00
", + "bbox": [ + 96, + 126, + 906, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing them with average imputation, followed by feature concatenation. All parameters of compared methods are configured as the recommended values in their original codes.", + "bbox": [ + 89, + 704, + 480, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Each dataset is divided into training, validation and test sets in the ratio of 7:1:2. To simulate the partial view setting, a certain proportion of instances defined by the partial example ratio (PER) are randomly selected as unavailable instances on each view. Besides, we ensure each sample contains at least one complete view to avoid invalid cases. In the multi-label noise configuration, we set a uniform noise rate for each label, i.e., $\\rho$ of positive and negative labels are flipped to their opposite. All results are obtained from ten independent runs, with final outcomes", + "bbox": [ + 88, + 750, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "reported as mean values alongside standard deviations. Our model is implemented by PyTorch on one NVIDIA GeForce RTX 4090 GPU of 24GB memory.", + "bbox": [ + 511, + 704, + 906, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Experimental Results.", + "text_level": 1, + "bbox": [ + 511, + 757, + 718, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance Evaluation. To evaluate the capabilities of DMMIvNL in addressing data deficiencies, we conduct extensive comparative experiments focusing on two critical aspects, i.e., label noise and view missing. Besides, we assign a broad range of noise rates and missing proportions to demonstrate the versatility of our approach. For label noise, $\\rho$ is considered at levels of $\\{10\\%, 15\\%, 20\\%, 25\\%, 30\\%, 35\\%, 40\\}$ ,", + "bbox": [ + 509, + 779, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20711", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/52e333e3da7bc03b518954980d5716f2be29d7c1523910ae13d73ab259d57d40.jpg", + "image_caption": [ + "(a) Yeast" + ], + "image_footnote": [], + "bbox": [ + 135, + 111, + 370, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/161217ce1dd5492f730d2e346f8e5773c0fb12d3ea577345c201016e5bdcd951.jpg", + "image_caption": [ + "(b) Corel 5k" + ], + "image_footnote": [], + "bbox": [ + 377, + 112, + 614, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51dd9063fea18a18c94ceac0bdc1183280175dd99f78cef399314d9babdc3788.jpg", + "image_caption": [ + "(c) VOC 2007" + ], + "image_footnote": [], + "bbox": [ + 622, + 109, + 859, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5087eda2c27a671dfb81526747cb660bf098eba72dc48e63ae0c1c4921bcec58.jpg", + "image_caption": [ + "(d) ESP Game" + ], + "image_footnote": [], + "bbox": [ + 138, + 303, + 375, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b43bd37c2d906ae295d75fe5d75cbea65e4fbd37e423d9bfd6ff05ac29e24369.jpg", + "image_caption": [ + "(e) IAPR TC-12" + ], + "image_footnote": [], + "bbox": [ + 379, + 304, + 614, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc5f98b1f92e024bb2699244a9a74ea23de57d560e685b12131311257aee1669.jpg", + "image_caption": [ + "(f) MIR FlickR" + ], + "image_footnote": [], + "bbox": [ + 624, + 305, + 859, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3bf4e66f985ac688fb16cd191a1ebb13d852341fc3387040ef1eb6d45c73e89.jpg", + "image_caption": [ + "Figure 2. AUC comparisons on six datasets with noise rate $\\rho$ varying from $10\\%$ to $40\\%$ while PER=50%.", + "(a) Espgame", + "Figure 3. Parameter analysis of the trade-off parameters $\\lambda_{1}$ and $\\lambda_{2}$ on Espgame and IAPR TC-12." + ], + "image_footnote": [], + "bbox": [ + 91, + 546, + 285, + 675 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/744832ee31c6c5145dbd2301fb4869c6d23ce005eeab8129955e177eb4198031.jpg", + "image_caption": [ + "(b) IAPR TC-12" + ], + "image_footnote": [], + "bbox": [ + 285, + 546, + 480, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "while for view missing, PER encompasses values of $\\{10\\%, 30\\%, 50\\%, 70\\}$ . Tab. 1 displays all metrics with $\\rho$ fixed at $30\\%$ and PER fixed at $50\\%$ , Fig. 2 illustrates the variation in AUC as $\\rho$ varies from $10\\%$ to $40\\%$ , and Fig. 4 shows the metrics distribution when $\\mathrm{PER} = 10\\%$ and $70\\%$ . Additional results are included in the Appendix.", + "bbox": [ + 89, + 763, + 482, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding the label noise, we can find that 1) Our method outperforms all compared approaches, particularly distinguished in the most representative AP metric. For", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "instance, our method achieves a remarkable AP improvement of up to $43\\%$ on Espgame in Tab. 1. 2) Compared to NTMLCS for handling label noise, our approach reveals substantial advantages. Besides, in terms of the performance on VOC 2007, DMMIvNL exhibits greater stability than the stepwise estimation way of NTMLCS, which is achieved through the cycle-consistency framework that simultaneously determines the transition matrix and classifier. Additionally, our approach surpasses methods developed for incomplete MvMLC. For example, Fig. 2 illustrates that our DMMIvNL reliably improves performance even against the top-performing MvMLC methods MTD and AIMNet. 3) As $\\rho$ increases from $10\\%$ to $40\\%$ , DMMIvNL consistently outperforms the other eight methods. Moreover, many methods lose reliability at elevated noise levels, such as NTMLCS on Yeast, TM3L on MIR FLICKR and DIMC on Corel 5k, while our approach continues to learn a competent classifier and deliver robust performance.", + "bbox": [ + 511, + 544, + 906, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As for the view missing, we have the following insights: 1) Our method exhibits outstanding performance in handling missing views. As illustrated in the radar chart of Fig. 4, DMMIvNL consistently occupies the outermost circle, reflecting impressive results across all metrics. 2)", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "DICNet DIMC MTD SIP AIMNet iMVWL TM3L NTMLCS DMMIvNL", + "bbox": [ + 174, + 89, + 823, + 102 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20712", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39fcff99320936a896395f5de3a52a15c222116229d7ade810422e84abe36cc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 89, + 823, + 102 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/293accfec9fb56fb79f030aebdbc6a027c20742bebeb9c6a11ca4d0b1d88a6ca.jpg", + "image_caption": [ + "(a) Yeast, PER=10%" + ], + "image_footnote": [], + "bbox": [ + 91, + 107, + 294, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea37a13ebdea01c15b7321498646c40da7b46ba9611e5f104186ce58bafd1e6c.jpg", + "image_caption": [ + "(b) VOC 2007, PER=10%" + ], + "image_footnote": [], + "bbox": [ + 297, + 109, + 496, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/feb5d9a7f31e34a9bba5915ba2e66ee01618a4995f9f8288a5e294272e756e46.jpg", + "image_caption": [ + "(c) Espgame, PER=10%" + ], + "image_footnote": [], + "bbox": [ + 500, + 109, + 700, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ddf951c0cc525b648deb557bd42932c0c55e56a04caa64fbffa1afa59dbf6bbf.jpg", + "image_caption": [ + "(d) MIR FLICKR, PER=10%" + ], + "image_footnote": [], + "bbox": [ + 704, + 109, + 903, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c651e7fd7c34e877e663208ef30fe9df9ac9e06ce05204bd05fa435f8af88e2.jpg", + "image_caption": [ + "(e) Yeast, PER=70%" + ], + "image_footnote": [], + "bbox": [ + 91, + 257, + 294, + 377 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00030c4cb484d5b79e2bb2a13d20468ba8f41817a2735acb5da56978d2837a79.jpg", + "image_caption": [ + "(f) VOC 2007, PER=70%" + ], + "image_footnote": [], + "bbox": [ + 297, + 258, + 496, + 377 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65b3b5d287cdb538e84f9d8362c540fbda24e25c5c1247da342ba70dec3c5222.jpg", + "image_caption": [ + "(g) Espgame, PER=70%" + ], + "image_footnote": [], + "bbox": [ + 500, + 258, + 700, + 377 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5da81f1db5f85b0a15fc6edd59c01367941ca7b3bacd2756172a459b9080ba56.jpg", + "image_caption": [ + "(h) MIR FLICKR,PER=70%", + "Figure 4. Experimental results of nine methods on four datasets with PER=10% and 70% while $\\rho = 30\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 704, + 258, + 903, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As PER changes from $10\\%$ to $70\\%$ , DMMIvNL continues to achieve top performance. Notably, while DIMC experiences a substantial decline at $70\\%$ PER, our method remains efficient. It indicates that DMMIvNL delivers strong robustness across varying levels of feature absence. 3) Our method makes remarkable breakthroughs over deep learning based methods for missing views, like SIP, MTD, and DICNet. For example, on MIR FLICKR, DMMIvNL improves 1-OE by nearly $50\\%$ compared to SIP at $50\\%$ PER. This improvement is primarily due to the focus on extracting the shared semantics and specific features through information separation, an aspect overlooked by other methods.", + "bbox": [ + 88, + 450, + 483, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study and Parameter Sensitivity. The ablation experiments are conducted to analyze the effect of the three crucial aspects of DMMIvNL, i.e., representation extraction based on the information bottleneck theory $(\\mathcal{L}_{IB})$ , cycle-consistency estimation for noise identification and classifier learning $(\\mathcal{L}_{ID})$ , and model regularization to combat overfitting $(\\mathcal{L}_r)$ . Note that loss $\\mathcal{L}_c$ is retained since it is essential for classifier improvement during early training. The first three rows of Tab. 2 indicate that $\\mathcal{L}_{IB}$ , $\\mathcal{L}_{ID}$ and $\\mathcal{L}_r$ are integral components of DMMIvNL. Besides, the comparison between the empirical loss $\\mathcal{L}_{vol}$ from problem (11) and $\\mathcal{L}_{ID}$ confirms that the cycle-consistency framework enhances the stability of transition matrix estimation. To study the impact of $\\lambda_1$ and $\\lambda_2$ , we present the results under distinct parameter settings in Fig. 3. As shown in the heatmap, the performance of DMMIvNL exhibits variability and improves as $\\lambda_1$ gets closer to 0.001 and $\\lambda_2$ approaches 1.", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1987b0ddaafc71a4d841655aa18f4fc3f329901d9b3e4e8103677248b3231412.jpg", + "table_caption": [ + "Table 2. Ablation study on Yeast, Corel 5k and VOC 2007 with PER=50% and ρ = 30%. ‘√’ and ‘X’ represent the used and not used corresponding item, respectively." + ], + "table_footnote": [], + "table_body": "
LvolLIBLIDLrEspgameCorel 5kVOC 2007
APAUCAPAUCAPAUC
XX✓✓0.17500.68460.17130.69720.41450.7366
X✓X✓0.18380.71130.18500.70890.43050.7367
X✓✓X0.18340.71310.18000.70960.43590.7148
✓X✓✓0.14990.67320.15430.62840.36120.6715
X✓✓✓0.19340.73070.19970.72840.46220.7566
", + "bbox": [ + 516, + 502, + 903, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 638, + 633, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address the dual challenges of view missing and multi-label noise, we propose DMMIvNL in this paper. Built upon the information bottleneck theory, a mutual information model is established to constrain information exchange across diverse representations, along with a derived theoretical upper bound. Besides, we prove that minimizing the volume of the transition matrix for each category consistently leads to the optimal classifier. Moreover, stability in noise recognition is further improved by introducing a cycle-consistency estimation framework, while raw fidelity information and inherent label correlations are used to reduce overfitting. Finally, extensive experiments demonstrate the superiority of DMMIvNL. In future work, we will leverage the prior knowledge embedded in LLM to rectify inaccurate labels and tackle the issue of class imbalance.", + "bbox": [ + 509, + 672, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20713", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the National Natural Science Foundation of China [Grant No. 62376281], and the Key NSF of China under Grant No. 62136005.", + "bbox": [ + 89, + 114, + 482, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 175, + 186, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Luis Von Ahn and Laura Dabbish. Labeling images with a computer game. In SIGCHI Conference on Human Factors in Computing Systems, page 319-326, 2004. 5", + "[2] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. In ICML, pages 825-836, 2021. 4", + "[3] Jia-Yao Chen, Shao-Yuan Li, Sheng-Jun Huang, and Chm. Unm: A universal approach for noisen, songcan and wang, lei and xie, ming-kun. IEEE Transactions on Knowledge and Data Engineering, 36:4968-4980, 2024. 2", + "[4] De Cheng, Yixiong Ning, Nannan Wang, Xinbo Gao, Heng Yang, Yuxuan Du, Bo Han, and Tongliang Liu. Class-dependent label-noise learning with cycle-consistency regularization. pages 11104-11116, 2022. 4", + "[5] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In ECCV, pages 97-112, 2002. 5", + "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2010. 5", + "[7] Yuling Fan, Jinghua Liu, Jianeng Tang, Peizhong Liu, Yaojin Lin, and Yongzhao Du. Learning correlation information for multi-label feature selection. Pattern Recognition, 145: 109899, 2024. 4", + "[8] Marco Federici, Anjan Dutta, Patrick Forre, Nate Kushman, and Zeynep Akata. Learning robust representations via multi-view information bottleneck. In ICLR, 2020. 2", + "[9] Xiao Fu, Kejun Huang, and Nicholas D Sidiropoulos. On identifiability of nonnegative matrix factorization. IEEE Signal Processing Letters, 25(3):328-332, 2018. 4", + "[10] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International Workshop onto Image, pages 1–11, 2006. 5", + "[11] Cordelia Guillaumin. Multimodal semi-supervised learning for image classification. In CVPR, pages 902-909, 2010. 5", + "[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, I Tsang, and Masashi Sugiyama. Robust training of deep neural networks with extremely noisy labels. In NeurIPS, page 4, 2020. 4", + "[13] Pingting Hao, Weiping Ding, Wanfu Gao, and Jialong He. Exploring view-specific label relationships for multi-view multi-label feature selection. Information Sciences, 681: 121215, 2024. 1", + "[14] Huaxi Huang, Hui Kang, Sheng Liu, Olivier Salvado, Thierry Rakotoarivelo, Dadong Wang, and Tongliang Liu." + ], + "bbox": [ + 93, + 200, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Paddles: Phase-amplitude spectrum disentangled early stopping for learning with noisy labels. In ICCV, pages 16719-16730, 2023. 2", + "[15] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In ICMR, pages 39-43, 2008. 5", + "[16] Xiaodong Jia, Xiao-Yuan Jing, Xiaoke Zhu, Songcan Chen, Bo Du, Ziyun Cai, Zhenyu He, and Dong Yue. Semi-supervised multi-view deep discriminant representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2496-2509, 2020. 1, 4", + "[17] Zhangqi Jiang, Tingjin Luo, and Xinyan Liang. Deep incomplete multi-view learning network with insufficient label information. In AAAI, pages 12919-12927, 2024. 2", + "[18] Rihuan Ke. Deep variation prior: Joint image denoising and noise variance estimation without clean data. IEEE Transactions on Image Processing, 33:2908-2923, 2024. 4", + "[19] Shikun Li, Xiaobo Xia, Hansong Zhang, Yibing Zhan, Shiming Ge, and Tongliang Liu. Estimating noise transition matrix with label correlations for noisy multi-label learning. In NeurIPS, pages 24184-24198, 2022. 2", + "[20] Xuefeng Li, Tongliang Liu, Bo Han, Gang Niu, and Masashi Sugiyama. Provably end-to-end label-noise learning without anchor points. In ICML, pages 6403-6413, 2021. 4", + "[21] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4447-4461, 2022. 3", + "[22] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In AAAI, pages 8807-8815, 2023. 1", + "[23] Chengliang Liu, Jie Wen, Xiaoling Luo, and Yong Xu. Incomplete multi-view multi-label learning via label-guided masked view-and category-aware transformers. In AAAI, pages 8816-8824, 2023. 1, 5", + "[24] Chengliang Liu, Jinlong Jia, Jie Wen, Yabo Liu, Xiaoling Luo, Chao Huang, and Yong Xu. Attention-induced embedding imputation for incomplete multi-view partial multi-label classification. In AAAI, pages 13864-13872, 2024. 1", + "[25] Chengliang Liu, Jie Wen, Yabo Liu, Chao Huang, Zhihao Wu, Xiaoling Luo, and Yong Xu. Masked two-channel decoupling framework for incomplete multi-view weak multi-label learning. In NeurIPS, 2024. 1, 4, 5", + "[26] Chengliang Liu, Gehui Xu, Jie Wen, Yabo Liu, Chao Huang, and Yong Xu. Partial multi-view multi-label classification via semantic invariance learning and prototype modeling. In ICML, 2024. 1, 2", + "[27] Chengliang Liu, Jie Wen, Yong Xu, Bob Zhang, Liqiang Nie, and Min Zhang. Reliable representation learning for incomplete multi-view missing multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2025. 1", + "[28] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3):447-461, 2015. 4" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20714", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Tingjin Luo, Yang Yang, Dongyun Yi, and Jieping Ye. Robust discriminative feature learning with calibrated data reconstruction and sparse low-rank model. Applied Intelligence, 54(3):2867-2880, 2024. 1", + "[30] Yong Luo, Dacheng Tao, Chang Xu, Chao Xu, Hong Liu, and Yonggang Wen. Multiview vector-valued manifold regularization for multilabel image classification. IEEE Transactions on Neural Networks and Learning Systems, 24(5): 709-722, 2013. 1", + "[31] Yubin Park and Joyce C Ho. Tackling overfitting in boosting for noisy healthcare data. IEEE Transactions on Knowledge and Data Engineering, 33(7):2995-3006, 2019. 1", + "[32] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, pages 1944-1952, 2017. 4", + "[33] Yalan Qin, Hanzhou Wu, Xinpeng Zhang, and Guorui Feng. Semi-supervised structured subspace learning for multi-view clustering. IEEE Transactions on Image Processing, 31:1-14, 2021. 1", + "[34] Alisa Smirnova, Jie Yang, Dingqi Yang, and Philippe CudreMauroux. Nessy: A neuro-symbolic system for label noise reduction. IEEE Transactions on Knowledge and Data Engineering, 35(8):8300-8311, 2022. 1", + "[35] Shiding Sun, Bo Wang, and Yingjie Tian. Decoupled representation for multi-view learning. Pattern Recognition, 151: 110377, 2024. 2", + "[36] Yu-Yin Sun, Yin Zhang, and Zhi-Hua Zhou. Multi-label learning with weak label. In AAAI, pages 593-598, 2010. 1", + "[37] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In IJCAI, pages 2703-2709, 2018. 1", + "[38] Ziyi Wang and Yitian Xu. A two-stage multi-view partial multi-label learning for enhanced disambiguation. Knowledge-Based Systems, 293:111680, 2024. 1", + "[39] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems, 35(8):11396-11408, 2023. 1", + "[40] Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning with noisy label identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3676-3687, 2021. 1", + "[41] Wulin Xie, Xiaohuan Lu, Yadong Liu, Jiang Long, Bob Zhang, Shuping Zhao, and Jie Wen. Uncertainty-aware pseudo-labeling and dual graph driven network for incomplete multi-view multi-label classification. In ACMMM, pages 6656-6665, 2024. 1", + "[42] Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. Instance-dependent label-noise learning under a structural causal model. In NeurIPS, 2021. 2", + "[43] Yitong Yao, Jing Zhang, Peng Zhang, and Yueheng Sun. A dual-branch learning model with gradient-balanced loss for long-tailed multi-label text classification. ACM Transactions on Information Systems, 42(2):1-24, 2023. 5" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Selim F Yilmaz, E Batuhan Kaynak, Aykut Koç, Hamdi Dibeklioglu, and Suleyman Serdar Kozat. Multi-label sentiment analysis on 100 languages with dynamic weighting for label imbalance. IEEE Transactions on Neural Networks and Learning Systems, 34:331-343, 2021. 4", + "[45] Dawei Zhao, Qingwei Gao, Yixiang Lu, and Dong Sun. Two-step multi-view and multi-label learning with missing label via subspace learning. Applied Soft Computing, 102:107120, 2021. 5", + "[46] Pengfei Zhu, Qi Hu, Qinghua Hu, Changqing Zhang, and Zhizhao Feng. Multi-view label embedding. Pattern Recognition, 84:126-135, 2018. 1", + "[47] Wenzhang Zhuge, Tingjin Luo, Ruidong Fan, Hong Tao, Chenping Hou, and Dongyun Yi. Absent multiview semisupervised classification. IEEE Transactions on Cybernetics, 54(3):1708-1721, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 316 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20715", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_model.json b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_model.json new file mode 100644 index 0000000000000000000000000000000000000000..720ed327ddd8dc7eb9570a59f1680b2d535877f5 --- /dev/null +++ b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.204, + 0.7, + 0.24 + ], + "angle": 0, + "content": "Quanjiang Li, Tingjin Luo*, Jiahui Liao National University of Defense Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.242, + 0.813, + 0.257 + ], + "angle": 0, + "content": "liquanjiang@nudt.edu.cn,tingjinluo@hotmail.com,liaojiahui19@nudt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.324, + 0.486, + 0.673 + ], + "angle": 0, + "content": "Incomplete features and label noise in multi-view multi-label data significantly undermine the reliability and performance, motivating researchers to explore the mechanism of representation and information recovery. However, learning for such dual deficiencies is crucial but rarely studied. In this paper, we propose a theory-inspired Deep Multi-View Multi-Label Learning method with Incomplete Views and Noisy Labels named DMMIvNL to address these problems. Specifically, to promote the synthesis of task-relevant shared information and preserve the distinctiveness of individual features from limited views, we have developed a feature extraction modular based on the information bottleneck theory, and formulated its theoretical upper bound into its objective. Meanwhile, we theoretically prove that minimizing the volume of the transition matrix ensures the statistical consistency with classifier training. Besides, a cycle-consistency estimation principle is proposed in the volume minimization network to improve the recognition stability of multi-label noise. Moreover, leveraging inherent real semantics information and label correlations are employed as model regularization to reduce the risk of excessive noise fitting. Finally, extensive experimental results validate the effectiveness and robustness of our DMMIvNL." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.701, + 0.222, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Advances in feature extraction has shifted data analysis towards multi-view frameworks [27], while growing data attributes drive multi-label learning to offer rich label semantics [13]. Therefore, Multi-view multi-label classification (MvMLC) has emerged to address the intricate demands of processing diverse features and numerous targets. Traditional methods [30, 46] assume the given data has complete views and clean labels, which is violated in practice [41]. On one hand, heterogeneous data from diverse sources frequently encounters missing views due to acquire" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.627 + ], + "angle": 0, + "content": "sition complexities and storage limitations [29, 47]. On the other hand, real-world collections often contain label noise as manual annotation can introduce incorrect labels beyond human knowledge [34]. Therefore, insufficient features and noisy labels severely reduce the classification performance. Regarding MvMLC, many solutions have been proposed to address missing views. iMvWL [37] used weighted matrix factorization to learn a shared subspace and AIMNet [24] leveraged a cross-view attention mechanism to recover missing views. Moreover, two types of label deficiencies have been considered in MvMLC. The problem of Partial Multi-Label (PML) [40] involves both relevant and irrelevant labels, while the problem of Multi-Label Learning with Missing Labels (MLML)[36] occurs when some relevant labels are missing. TFMDD [38] addressed PML with a two-stage manifold disambiguation model and LMVCAT [23] tackled MLML by employing label-guided graphs and category-aware transformers. While existing methods address each annotation problem separately, both issues tend to occur simultaneously in practice. Thus, the challenge of multi-label noise demands substantial attention, where each category suffers from both over-labeled and missing labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Existing research has not yet tackled the combined issues of missing views and multi-label noise, which are prevalent in real-world applications like incomplete and subjective medical diagnostics [31]. In fact, overcoming the simultaneous challenges requires refined representation extraction under missing views and effective classifier leaning with multi-label noise. The core of multi-view learning lies in extracting shared information across different views while preserving unique features of each view [33]. DICNet [22], DIMC [39] and SIP [26] concentrated solely on shared subspace and neglected distinctive characteristics of individual views. MDDRL [16] and MTD [25] attempted to extract both shared and specific representations with minimal interactions. However, their reliance on geometric orthogonality or distance metrics restricted them to capture only linear patterns, overlooking the complex nature of inter-view relations. Therefore, constraining feature semantics at the intrinsic information level is crucial for obtaining high-level" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.238, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20706" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "representations with precise distinction. Most approaches for handling multi-label noise utilize strategies like early stopping [14] and cyclical learning rates [3]. These methods often lack stability and require dataset-specific adjustments, making them insufficiently adaptive. ENTMLC [19] employed label correlations to identify the noise transition matrix for recovering clean labels. However, the decoupling of transition matrix acquisition and classifier training causes error accumulation, limiting the practical utility. To circumvent the uncertainty inherent in those strategies and the inefficiency of stepwise noise learning, a statistically consistent estimation framework should be developed, which would allow for direct training the optimal classifier from noisy labels and achieving strong robustness." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.305, + 0.484, + 0.621 + ], + "angle": 0, + "content": "To tackle these problems, we propose a theory-inspired method with deep networks named DMMIvNL. Specifically, a mutual information model is developed to enhance synergy of shared features, suppress interactions from specific features, and impose orthogonality between shared and specific features of each view. The upper bound of the model is derived through the variational autoencoder framework, with instance-level contrastive operations applied to simulate diverse joint and marginal distributions. Moreover, the ideal classifier is trained on noisy data by minimizing the volume of the noise transition matrix and the process is proven to converge to a unique and globally optimal solution under the sufficiently scattered assumption. The difficulty in guaranteeing the assumption and relying exclusively on noisy class-posterior estimation promotes the design of a cycle-consistency estimation method, which integrates the volume maximization of clean class-posterior. Finally, the real semantics embedded in noisy data and hidden label correlation information are employed as model regularization to reduce the risk of noise overfitting. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.667 + ], + "angle": 0, + "content": "- This is the first multi-view multi-label learning framework capable of handling incomplete views and noisy labels, including both over-labeled and missing labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.713 + ], + "angle": 0, + "content": "- Our mutual information based model extracts semantically distinct representations from limited views, with its upper bound established through variational derivation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.483, + 0.787 + ], + "angle": 0, + "content": "- We theoretically prove the statistical consistency and joint convergence between minimizing the noise transition matrix volume for each category and classifier training. Besides, a cycle-consistency loss framework is designed to enhance transition matrix estimation." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.806, + 0.226, + 0.824 + ], + "angle": 0, + "content": "2. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.274, + 0.848 + ], + "angle": 0, + "content": "2.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The multi-view dataset \\(\\{\\mathbf{X},\\mathbf{Y}\\}\\) contains \\(N\\) samples, where \\(\\mathbf{X}_i\\) consists of \\(V\\) views with dimension \\(d_v\\), i.e., \\(\\mathbf{X}_i = \\{\\pmb{x}_i^{(v)}\\in \\mathbb{R}^{d_v}\\}_{v = 1}^V\\). The label matrix \\(\\mathbf{Y}\\in \\{0,1\\}^{N\\times C}\\) corre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.306 + ], + "angle": 0, + "content": "sponds to \\(C\\) categories, with \\(Y_{i,j} = 1\\) if sample \\(i\\) belongs to class \\(j\\), and \\(Y_{i,j} = 0\\) otherwise. \\(\\mathbf{W} \\in \\mathbb{R}^{N \\times V}\\) is an indicator matrix, where \\(W_{i,j} = 1\\) indicates the \\(j\\)-th view is available for sample \\(i\\), and \\(W_{i,j} = 0\\) means missing. Since precise annotation is hard to obtain, the training dataset in practice is \\(\\{\\mathbf{X}, \\bar{\\mathbf{Y}}\\}\\), where \\(\\bar{\\mathbf{Y}}\\) denotes noisy labels. Thus, we denote the clean and noisy labels of the \\(j\\)-th class as \\(\\mathbf{Y}^j\\) and \\(\\bar{\\mathbf{Y}}^j\\), related by a noise transition matrix \\(\\mathbf{T}^j \\in [0,1]^{2 \\times 2}\\), \\(j \\in [C]\\). When only noisy samples are available and without further constraints, instance-dependent transition matrix becomes non-identifiable [42]. Therefore, we employ a general assumption that the noise transition matrix is class-dependent and instance-independent, i.e., \\(P(\\bar{\\mathbf{Y}}^j = k \\mid \\mathbf{Y}^j = i, \\mathbf{X} = x) = P(\\bar{\\mathbf{Y}}^j = k \\mid \\mathbf{Y}^j = i) = T_{i,k}^j\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.317, + 0.907, + 0.332 + ], + "angle": 0, + "content": "2.2. Feature Extraction via Information Bottleneck" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.909, + 0.612 + ], + "angle": 0, + "content": "Deriving effective representations from raw data aids in reducing noise and redundancy [17]. Consequently, both task-relevant shared semantics and view-invariant characteristic features should be explored to enhance information processing efficiency. Let \\( z^v \\) and \\( s^v \\) denote the \\( D \\)-dimensional shared and specific representation extracted from each view and the goal is to ensure consistency of the shared features across multiple views and preserve the unique contributions of each view. Considering different interaction patterns between representations from \\( x^{v^*} \\) and \\( x^v \\), mutual information \\( I(z^{v^*}|x^{v^*};z^v|x^v) \\) should be maximized to capture only universally shared information, while \\( I(s^{v^*}|x^{v^*};s^v|x^v) \\) should be minimized to preserve only individual attributes. Besides, the shared and specific features should be as orthogonal as possible to ensure absolute distinction in information content [35], i.e., \\( I(z^v |x^v;s^v |x^v) = 0 \\). Thus, the constrained optimization problem based on the information bottleneck theory [8] can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.626, + 0.907, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\max \\frac {1}{V} \\sum_ {v = 1} ^ {V} \\sum_ {v ^ {*} \\neq v} ^ {V} \\left(I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) - I \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right)\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.672, + 0.905, + 0.726 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad \\frac {1}{V} \\sum_ {v = 1} ^ {V} I \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) = 0. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.727, + 0.905, + 0.757 + ], + "angle": 0, + "content": "By applying the Lagrange multiplier method, the problem (1) can be equally converted as" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.77, + 0.907, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max \\frac {1}{V} \\sum_ {v = 1} ^ {V} \\sum_ {v ^ {*} \\neq v} ^ {V} \\left(I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) - I \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right)\\right) \\\\ - \\frac {\\beta}{V} \\sum_ {v = 1} ^ {V} I \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\beta \\geq 0\\) is the Lagrange multiplier to balance the efficiency and compactness of information [26]. For the first" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20707" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.091, + 0.821, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.398, + 0.804, + 0.415 + ], + "angle": 0, + "content": "Figure 1. The main framework of our proposed DMMIvNL. Different shapes signify different samples." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.44, + 0.407, + 0.455 + ], + "angle": 0, + "content": "term, we can obtain the following lower bound:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.462, + 0.504, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\\\ = \\iint p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) \\log \\frac {p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v} , \\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right)}{p \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right)} d \\boldsymbol {z} ^ {v ^ {*}} d \\boldsymbol {z} ^ {v} \\\\ \\geq \\int \\int p (\\boldsymbol {z} ^ {v} | \\boldsymbol {z} ^ {v ^ {*}}) p (\\boldsymbol {z} ^ {v ^ {*}} | \\boldsymbol {x} ^ {v ^ {*}}) \\log p (\\boldsymbol {z} ^ {v} | \\boldsymbol {z} ^ {v ^ {*}}) d \\boldsymbol {z} ^ {v ^ {*}} d \\boldsymbol {z} ^ {v}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.62 + ], + "angle": 0, + "content": "where \\(p(\\cdot)\\) denotes the probability density function. Since computing \\(p(z^v |z^{v^*})\\) is intractable, we approximate it using a stochastic variational distribution \\(q(z^{v}|z^{v^{*}})\\). Then we can obtain a new variational lower bound for maximization:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.625, + 0.483, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\\\ \\geq \\int \\int p (\\boldsymbol {z} ^ {v} | \\boldsymbol {x} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}} | \\boldsymbol {x} ^ {v ^ {*}}) \\log q (\\boldsymbol {z} ^ {v} | \\boldsymbol {z} ^ {v ^ {*}}) d \\boldsymbol {z} ^ {v ^ {*}} d \\boldsymbol {z} ^ {v} \\\\ = D _ {K L} ^ {\\dagger} \\left(p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) \\| q \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {z} ^ {v ^ {*}}\\right)\\right) = D _ {K L} ^ {\\dagger} \\left(\\boldsymbol {z} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}}\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.484, + 0.779 + ], + "angle": 0, + "content": "where \\(D_{KL}(\\cdot ||\\cdot)\\) means the Kullback-Leibler divergence and \\(D_{KL}^{\\dagger}(\\cdot ||\\cdot)\\) is a pseudo-symbol for notational consistency. For the middle term, the upper bound can be obtained by introducing another variational distribution \\(q(s^{v}|s^{v^{*}})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.786, + 0.483, + 0.852 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\\\ \\leq D _ {K L} \\left(p \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}, \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\| p \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) q \\left(\\boldsymbol {s} ^ {v} \\mid \\boldsymbol {s} ^ {v ^ {*}}\\right)\\right) \\tag {4} \\\\ = D _ {K L} ^ {\\left(\\mathbf {s} ^ {v}, \\mathbf {s} ^ {v ^ {*}}\\right)}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.859, + 0.393, + 0.875 + ], + "angle": 0, + "content": "Similarly, the upper bound for the last term is" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.882, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nI \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\leq D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {z} ^ {v}\\right)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Combining Eqs. (3), (4) and (5), our optimization objective becomes minimizing the upper bound of the problem (2):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.474, + 0.938, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I B} = \\frac {1}{V} \\sum_ {v = 1} ^ {V} \\sum_ {v ^ {*} \\neq v} ^ {V} \\left(- D _ {K L} ^ {\\dagger \\left(\\boldsymbol {z} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}}\\right)} + D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {s} ^ {v ^ {*}}\\right)}\\right) + \\frac {\\beta}{V} \\sum_ {v = 1} ^ {V} D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {z} ^ {v}\\right)}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Contrastive learning [21] is used to compute various distributions. Specifically, representations are treated as probability vectors over \\(D\\) classes via a Softmax activation function, and the joint distribution matrix is obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.597, + 0.912, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {P} ^ {(v, v ^ {*})} = \\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {z} _ {i} ^ {(v)}\\right) ^ {T} \\boldsymbol {z} _ {i} ^ {(v ^ {*})}, \\boldsymbol {Q} ^ {(v, v ^ {*})} = \\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {s} _ {i} ^ {(v)}\\right) ^ {T} \\boldsymbol {s} _ {i} ^ {(v ^ {*})}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.751 + ], + "angle": 0, + "content": "Due to the strong interaction between \\( z^v \\) and \\( z^{v^*} \\), the variational distribution \\( q(z^v |z^{v^*}) \\) can be estimated by the obtained conditional distribution. Given the extremely weak correlation between \\( s^v \\) and \\( s^{v^*} \\), as well as \\( z^v \\) and \\( s^v \\), the distribution \\( q(s^v /s^{v^*}) \\) and \\( q(z^{v} / s^{v}) \\) can be approximated by the corresponding marginal distribution. Subsequently, each term in Eq. (6) can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.755, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} D _ {K L} ^ {\\dagger \\left(\\boldsymbol {z} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}}\\right)} = \\sum_ {d = 1} ^ {D} \\sum_ {d ^ {\\prime} = 1} ^ {D} \\boldsymbol {P} _ {d, d ^ {\\prime}} ^ {(v, v ^ {*})} \\log \\frac {\\boldsymbol {P} _ {d , d ^ {\\prime}} ^ {(v , v ^ {*})}}{\\boldsymbol {P} _ {d} ^ {(v)}} \\\\ D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {s} ^ {v ^ {*}}\\right)} = \\sum_ {d = 1} ^ {D} \\sum_ {d ^ {\\prime} = 1} ^ {D} \\boldsymbol {Q} _ {d, d ^ {\\prime}} ^ {(v, v ^ {*})} \\log \\frac {\\boldsymbol {Q} _ {d , d ^ {\\prime}} ^ {(v , v ^ {*})}}{\\left(\\boldsymbol {Q} _ {d} ^ {(v)}\\right) ^ {\\alpha} \\boldsymbol {Q} _ {d ^ {\\prime}} ^ {(v ^ {*})}} \\\\ D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {z} ^ {v}\\right)} = \\sum_ {d = 1} ^ {D} \\sum_ {d ^ {\\prime} = 1} ^ {D} \\boldsymbol {M} _ {d, d ^ {\\prime}} ^ {\\left(z ^ {v}, s ^ {v}\\right)} \\log \\frac {\\boldsymbol {M} _ {d , d ^ {\\prime}} ^ {\\left(z ^ {v} , s ^ {v}\\right)}}{\\left(\\boldsymbol {M} _ {d} ^ {\\left(z ^ {v}\\right)}\\right) ^ {\\alpha} \\boldsymbol {M} _ {d ^ {\\prime}} ^ {\\left(s ^ {v}\\right)}}, \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20708" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.485, + 0.244 + ], + "angle": 0, + "content": "where the marginal distributions \\(P_{d}^{(v)}\\) and \\(P_{d^{\\prime}}^{(v^{*})}\\) can be obtained by summing over the \\(d\\)-th row and the \\(d^{\\prime}\\)-th column of the joint distribution matrix \\(P^{(v,v^{*})}\\), with other symbols defined as the same. In this conversion, we let \\(q(\\boldsymbol{s}^v /\\boldsymbol{s}^{v^*}) = (Q^{(v)})^\\alpha\\) and \\(q(z^{v} / s^{v}) = (M^{(z^{v})})^{\\alpha}\\), where \\(\\alpha\\) is a balance factor to preserve crucial information and ensure model stability [16]. By introducing the mask matrix \\(\\pmb{W}\\) to mitigate the impact of unavailable instances, we can perform cross-view fusion to acquire the information-enriched shared and specific representation:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.253, + 0.483, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {Z}} _ {i,:} = \\sum_ {v = 1} ^ {V} \\frac {\\boldsymbol {z} _ {i , :} ^ {(v)} \\boldsymbol {W} _ {i , v}}{\\sum_ {v} \\boldsymbol {W} _ {i , v}}, \\quad \\bar {\\boldsymbol {S}} _ {i,:} = \\sum_ {v = 1} ^ {V} \\frac {\\boldsymbol {s} _ {i , :} ^ {(v)} \\boldsymbol {W} _ {i , v}}{\\sum_ {v} \\boldsymbol {W} _ {i , v}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.351 + ], + "angle": 0, + "content": "Departing from traditional addition or concatenation, we conduct feature interaction by enhancing shared information with proprietary data [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.359, + 0.483, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {X}} _ {i, j} = \\theta_ {s} \\left(\\bar {\\boldsymbol {Z}} _ {i, j}\\right) \\cdot \\bar {\\boldsymbol {S}} _ {i, j}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.483, + 0.421 + ], + "angle": 0, + "content": "where \\(\\theta_{s}\\) is the sigmoid activation function and \\(\\hat{\\pmb{X}}\\in \\mathbb{R}^{N\\times D}\\) is the final fusion representation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.43, + 0.484, + 0.46 + ], + "angle": 0, + "content": "2.3. Identify Multi-Label Noise by Volume Minimization Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.709 + ], + "angle": 0, + "content": "The success of deep learning relies on large-scale accurately labeled data, with performance significantly dropping under noisy data [12], making noise identification crucial for network training [2]. In multi-class tasks, many statistically consistent methods based on the transition matrix have been developed [28, 32], which guarantees the classifier learned from noisy data can converge to the optimal classifier under the clean risk. The anchor-point assumption [28] is commonly utilized to determine the transition matrix, which postulates the existence of samples that belong to each class with probability one. However, implementing this approach in multi-label learning is challenging due to the difficulty of pinpointing suitable anchor samples for each label. Motivated by [20], we capitalize on the geometric properties of the transition matrix to seek a breakthrough. The transition matrix for the \\(j\\)-th label is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.717, + 0.483, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} ^ {j} = \\left( \\begin{array}{l l} P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 0 \\mid \\boldsymbol {Y} ^ {j} = 0\\right) & P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 0 \\mid \\boldsymbol {Y} ^ {j} = 1\\right) \\\\ P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 1 \\mid \\boldsymbol {Y} ^ {j} = 0\\right) & P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 1 \\mid \\boldsymbol {Y} ^ {j} = 1\\right) \\end{array} \\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Given an instance \\(\\hat{\\pmb{x}}\\), the noisy class-posterior for the \\(j\\)-th label is \\(P(\\bar{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}}) = [P(\\bar{Y}^j = 0|\\mathbf{X} = \\hat{\\pmb{x}}), P(\\bar{Y}^j = 1|\\mathbf{X} = \\hat{\\pmb{x}})]^T\\) and the clean class-posterior is \\(P(\\mathbf{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}}) = [P(\\mathbf{Y}^j = 0|\\mathbf{X} = \\hat{\\pmb{x}}), P(\\mathbf{Y}^j = 1|\\mathbf{X} = \\hat{\\pmb{x}})]^T\\). Thus, the relationship \\(P(\\bar{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}}) = \\pmb{T}^j P(\\mathbf{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}})\\) holds. Since \\(\\sum_{i=0}^{1} P(\\mathbf{Y} = i|\\mathbf{X} = \\hat{\\pmb{x}}) = 1\\), \\(P(\\bar{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}})\\) is then a convex combination of the columns of \\(\\pmb{T}^j\\), which implies that the noisy class-posterior is enclosed within the simplex \\(\\mathrm{Sim}\\{\\pmb{T}^j\\}\\). To reduce noise identification errors, it" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.268 + ], + "angle": 0, + "content": "is essential to recover \\(\\operatorname{Sim}\\{\\pmb{T}^j\\}\\) such that it tightly encompasses all sample posterior probability points, thereby making volume minimization of \\(\\operatorname{Sim}\\{\\pmb{T}^j\\}\\) the primary goal. For synchronized learning, we define a trainable \\(\\hat{T}^j \\in [0,1]^{2 \\times 2}\\) with \\(\\sum_{i=1}^{2} \\hat{T}^j{}_{i,k} = 1\\), and construct a multi-label classifier \\(h_\\theta\\) with parameters \\(\\theta\\). The optimization of \\(\\hat{T}^j\\) and \\(h_\\theta(\\hat{x})\\) involves two objectives: the discrepancy between \\(\\hat{T}^j(h_\\theta(\\hat{x})_j, 1 - h_\\theta(\\hat{x})_j)^T\\) and \\(P(\\bar{Y}^j | X = \\hat{x})\\) and the volume of each simplex spanned by the columns of \\(\\hat{T}^j\\). Therefore, the following model to jointly learn the transition matrix and classifier is proposed:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.274, + 0.907, + 0.355 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\min _ {\\hat {\\boldsymbol {T}} ^ {j}} \\sum_ {j = 1} ^ {C} \\operatorname {v o l} \\left(\\hat {\\boldsymbol {T}} ^ {j}\\right) \\\\ s. t. \\hat {\\boldsymbol {T}} ^ {j} \\left[ \\begin{array}{c} h _ {\\theta} (\\hat {\\boldsymbol {x}}) _ {j} \\\\ 1 - h _ {\\theta} (\\hat {\\boldsymbol {x}}) _ {j} \\end{array} \\right] = P \\left(\\bar {\\boldsymbol {Y}} ^ {j} \\mid X = \\hat {\\boldsymbol {x}}\\right), \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.518 + ], + "angle": 0, + "content": "where \\(\\mathrm{vol}(\\hat{T}^j)\\) denotes a measure to the volume of \\(\\operatorname{Sim}\\{\\hat{T}^j\\}\\). In the absence of anchor points, the class-posterior probability should incorporate substantial information to uniquely determine the minimum-volume simplex, with true sample probability points of each class being distinguishable. Therefore, the sufficiently scattered assumption [9] is introduced to limit the occurrence of anomalous clean class-posterior and aid in theoretically establishing that addressing problem (11) will drive \\(\\hat{T}^j\\) to converge to \\(\\pmb{T}^j\\) and \\(h_\\theta (\\hat{\\pmb{x}})\\) to accurately approximate \\(P(\\mathbf{Y}|\\mathbf{X} = \\hat{\\mathbf{x}})\\):" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.906, + 0.585 + ], + "angle": 0, + "content": "Theorem 1. If each \\( P(\\mathbf{Y}^j | \\mathbf{X}) \\) is sufficiently scattered, then \\( \\mathbf{T}_{\\star}^{j} = \\mathbf{T}^{j} \\) and \\( (h_{\\theta_{\\star}}(\\hat{\\mathbf{x}})_j, 1 - h_{\\theta_{\\star}}(\\hat{\\mathbf{x}})_j)^T = P(\\mathbf{Y}^j | \\mathbf{X} = \\hat{\\mathbf{x}}) \\) (\\( j = 1, 2, \\dots, C \\)) must hold, where \\( (\\mathbf{T}_{\\star}^{1}, \\mathbf{T}_{\\star}^{2}, \\dots, \\mathbf{T}_{\\star}^{C}, \\theta_{\\star}) \\) are the optimal solutions of the problem (11)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.905, + 0.622 + ], + "angle": 0, + "content": "2.4. Cycle-Consistency Estimation and Model Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In multi-label learning, imbalanced label distributions [44] and inter-class correlations [7] can compress class-posterior probability points and limit their scatter. Besides, the volume approximation of \\( T^j \\) in the problem (11) is highly sensitive to the noisy class-posterior, making it suffer from large estimation errors in noisy environments [18]. The failure of the sufficiently scattered assumption, along with the excessive reliance on the noisy class-posterior, compromises the effectiveness of the problem (11). To enhance transition matrix estimation, we develop an adaptive loss framework [4] that incorporates the volume maximization of the clean class-posterior, which leverages additional information to mitigate bias in single-target decision-making. Based on the equation \\( P(\\bar{Y}^j | \\mathbf{X}) = T^j P(\\mathbf{Y}^j | \\mathbf{X}) \\), where \\( P(\\bar{Y}^j | \\mathbf{X}) \\) can be derived from the noisy data, minimizing the volume of \\( T^j \\) is equivalent to imposing a volume constraint on \\( P(\\mathbf{Y}^j | \\mathbf{X}) \\). Through the matrix \\( \\tilde{T}^j \\) (an approximate inverse of \\( T^j \\)), we project the simplex formed in the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20709" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "noisy space onto the clean class-posterior space to obtain \\(P^{\\prime}(\\mathbf{Y}|\\mathbf{X})\\), which will make up the largest volume. Then we design a cycle-consistency optimization strategy to enhance both the volume constraint on \\(\\pmb{T}^{j}\\) and maximization of \\(P(\\mathbf{Y}|\\mathbf{X})\\), in which the estimated noisy class-posterior matches the provided noisy label \\(\\bar{\\mathbf{Y}}^j\\), the backward estimated clean class-posterior \\(P^{\\prime}(\\mathbf{Y}|\\mathbf{X})\\) aligns with the predicted clean-class posterior and the forward-backward consistency term \\(\\tilde{T}^j T^j P(\\mathbf{Y}|\\mathbf{X})\\) coheres with \\(P(\\mathbf{Y}|\\mathbf{X})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Firstly, we employ forward estimation to simultaneously optimize the transition matrix \\( T^{j} \\) and classifier \\( h_{\\theta} \\). By leveraging the cross-entropy loss and reporting the output as \\( F^{ij} = (h_{\\theta}(\\hat{\\boldsymbol{x}}_i)_j, 1 - h_{\\theta}(\\hat{\\boldsymbol{x}}_i)_j)^T \\), we aim to minimize the approximation error between the inferred noisy class-posterior probability \\( T^{j}F^{ij} \\) and the given noisy label \\( \\bar{Y}^{j} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.332, + 0.468, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {1} (\\theta , \\boldsymbol {T} ^ {1}, \\dots , \\boldsymbol {T} ^ {C}) = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} (\\bar {Y} _ {i, j} \\log (\\boldsymbol {T} _ {1} ^ {j} \\boldsymbol {F} ^ {i j}) \\\\ + \\left(1 - \\bar {Y} _ {i, j}\\right) \\log \\left(1 - \\mathbf {T} _ {2} ^ {j} \\mathbf {F} ^ {i j}\\right)). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.543 + ], + "angle": 0, + "content": "Secondly, the Sigmoid function is employed to convert the one-hot label into the noisy class-posterior probability, which is then normalized along the label dimension to capture the soft labels \\(\\bar{\\mathbf{Y}}^{sn}\\). Denoting the outcome \\((\\bar{\\mathbf{Y}}_{i,j}^{sn}, 1 - \\bar{\\mathbf{Y}}_{i,j}^{sn})^T\\) as \\(\\bar{\\mathbf{F}}^{ij}\\), we utilize \\(\\tilde{T}^j\\) to project \\(\\bar{\\mathbf{F}}^{ij}\\) onto the clean class-posterior space. Thus, we minimize the following cross-entropy loss between the backward estimated clean class-posterior \\(\\tilde{T}^j \\bar{\\mathbf{F}}^{ij}\\) and the true class-posterior probability estimation \\(h_\\theta(\\hat{\\boldsymbol{x}}_i)_j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.556, + 0.483, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {2} (\\theta , \\tilde {\\boldsymbol {T}} ^ {1}, \\dots , \\tilde {\\boldsymbol {T}} ^ {C}) = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} (h _ {\\theta} (\\hat {\\boldsymbol {x}} _ {i}) _ {j} \\log (\\tilde {\\boldsymbol {T}} _ {1} ^ {j} \\bar {\\boldsymbol {F}} ^ {i j}) \\\\ + \\left(1 - h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j}\\right) \\log \\left(1 - \\tilde {\\boldsymbol {T}} _ {2} ^ {j} \\bar {\\boldsymbol {F}} ^ {i j}\\right)). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.707 + ], + "angle": 0, + "content": "At this stage, the transition matrix can be derived from both the forward and backward ways by optimizing \\(\\mathcal{L}_1\\) and \\(\\mathcal{L}_2\\). Thirdly, we should establish the cycle-consistency loss by minimizing the difference between \\(\\tilde{T}^j T^j F^{ij}\\) and \\(h_\\theta (\\hat{x}_i)_j\\) to reinforce the invertible relationship of \\(T^j\\) and \\(\\tilde{T}^j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.721, + 0.482, + 0.783 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {3} (\\theta) = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} \\left(h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j} \\log \\left(\\tilde {\\boldsymbol {T}} _ {1} ^ {j} \\boldsymbol {T} ^ {j} \\boldsymbol {F} ^ {i j}\\right) \\right. \\tag {12} \\\\ + \\left(1 - h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j}\\right) \\log \\left(1 - \\bar {\\boldsymbol {T}} _ {2} ^ {j} \\boldsymbol {T} ^ {j} \\boldsymbol {F} ^ {i j}\\right)), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.87 + ], + "angle": 0, + "content": "where we only optimize the network parameters \\(\\theta\\) while keeping \\(T^j\\) and \\(\\tilde{T}^j\\) constant in order to maintain their consistency and focus on learning an optimal classifier. Therefore, the loss for identifying the transition matrix and learning the clean class-posterior is \\(\\mathcal{L}_{IB} = \\mathcal{L}_1 + \\mathcal{L}_2 + \\mathcal{L}_3\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To mitigate overfitting on noisy data, we design model regularization by leveraging the partial fidelity of noisy la" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "bels and label correlations. We allow the classifier to directly access raw labels and obtain the classification loss:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.134, + 0.905, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {c} = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} \\left(\\bar {Y} _ {i, j} \\log \\left(h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j}\\right)\\right) \\tag {13} \\\\ + \\left(1 - \\bar {Y} _ {i, j}\\right) \\log \\left(1 - h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j})\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.204, + 0.905, + 0.401 + ], + "angle": 0, + "content": "The construction of \\(\\mathcal{L}_c\\) is motivated by two key considerations. On one hand, it is essential to harness the underlying true information embedded in the noisy data to improve the inefficient classifier. On the other hand, strengthening the model memory with original noise helps minimize the risk of misclassifying clean samples as noisy ones. Since label correlation is a crucial factor for multi-label learning, we minimize the discrepancy in the label relationships reflected by the estimated clean class posterior and the classifier output to reduce classification bias. By defining the clean class-posterior matrix of the positive labels as \\(P^c\\), with elements given by \\(P_{i,j}^{c} = \\tilde{T}_{1}^{j}\\bar{F}^{ij}\\), we can build the following correlation-aligned loss driven by label co-occurrence [43]:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.412, + 0.904, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\frac {1}{C ^ {2}} \\left\\| \\left(\\boldsymbol {P} ^ {c}\\right) ^ {T} \\boldsymbol {P} ^ {c} - \\left(h _ {\\theta} (\\hat {\\boldsymbol {X}})\\right) ^ {T} h _ {\\theta} (\\hat {\\boldsymbol {X}}) \\right\\| _ {F} ^ {2}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.456, + 0.905, + 0.486 + ], + "angle": 0, + "content": "Thus, our model regularization can be formulated as \\(\\mathcal{L}_R = \\mathcal{L}_c + \\mathcal{L}_r\\). The total training loss of our model is" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.498, + 0.904, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {I B} + \\lambda_ {1} \\mathcal {L} _ {I D} + \\lambda_ {2} \\mathcal {L} _ {R}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.904, + 0.555 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are penalty coefficients to balance the effects of different losses when handling noisy labels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.568, + 0.645, + 0.584 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.714, + 0.608 + ], + "angle": 0, + "content": "3.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.78 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. Following [23, 45], our experiments incorporate six widely recognized multi-view multi-label datasets, i.e., Yeast [11], Corel 5k [5], VOC 2007 [6], Espgame [1], IAPR TC-12 [10], and MIR FLICKR [15]. Similar to [25], we employ six diverse metrics to comprehensively evaluate each method, i.e., Hamming Loss (HL), Ranking Loss (RL), OneError (OE), Coverage (Cov), Average Precision (AP), and Area Under Curve (AUC). To facilitate performance comparison, we present 1-HL, 1-RL, 1-OE, and 1-Cov values in our report, with higher values consistently indicating better performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparison Methods. To validate the effectiveness of DMMIvNL, we perform a comparative analysis against eight state-of-the-art methods i.e., iMVWL, TM3L, AlM-Net, DICNet, DIMC, MTD, SIP, ENTMLC. Since the first seven methods are designed to handle incomplete multiview multi-label problems, we train them directly on noisy data without modifications. As ENTMLC specializes in addressing multi-label noise, we tackle missing views by fill" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20710" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. Experimental results of nine methods on all datasets with PER=50% and ρ = 30%. The best result on each row is bolded and the second-best result is underlined. The average ranking on the six metrics is shown as Ave.R." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.127, + 0.907, + 0.681 + ], + "angle": 0, + "content": "
DataMETDICNetDIMCMTDSIPAIMNetiMVWLTM3LNTMLCSDMMIvNL
Yeast1-HL71.54±1.8174.94±1.5676.40±0.3371.24±1.4973.03±1.7966.94±0.4763.89±0.2969.88±0.1677.64±0.47
1-RL73.81±0.9174.82±2.3377.08±0.7571.30±1.6773.95±0.4372.83±1.0565.72±0.3771.40±3.1480.18±0.37
1-OE71.86±3.7470.95±5.3775.33±0.4861.94±6.7470.66±2.7963.91±2.2956.49±0.7773.28±0.4373.69±1.22
1-Cov41.68±2.4844.12±3.1045.53±1.9539.04±2.1641.19±0.3141.50±1.4534.93±0.7335.55±1.3750.77±0.84
AP65.61±1.3466.91±2.0968.87±0.8261.78±2.2565.55±1.1863.74±0.9256.90±0.3066.69±0.6072.08±0.34
AUC75.33±0.7376.34±2.3578.36±0.5473.42±1.5375.46±0.4574.85±0.9767.92±0.2873.33±2.4481.03±0.55
Ave.R4.503.331.837.174.836.338.837.001.17
Corel 5k1-HL96.37±0.6896.08±0.9998.65±0.0098.59±0.0498.66±0.0079.41±0.0597.23±0.6798.57±0.1798.68±0.02
1-RL60.14±0.8461.00±0.8767.26±0.5965.32±1.4163.37±0.6564.53±0.5552.01±0.3965.92±1.2372.27±0.57
1-OE6.30±0.895.44±1.0521.22±1.0612.88±4.2824.00±0.7310.29±2.4620.18±0.332.30±1.6725.20±0.78
1-Cov32.29±0.8133.31±0.9637.53±0.8836.69±2.0341.62±0.4736.09±1.1233.03±0.0235.19±0.9644.35±1.10
AP6.72±0.406.80±0.2915.22±0.6510.54±1.8618.49±0.2110.18±0.914.48±0.098.25±0.9919.18±0.45
AUC60.30±0.8551.74±1.3067.63±0.5965.69±1.5470.37±0.3764.81±0.7658.99±0.1165.97±1.2772.42±0.47
Ave.R7.677.672.834.332.676.007.335.501.00
VOC 20071-HL89.48±1.7089.72±1.3192.85±0.0592.00±0.1792.23±0.3274.95±0.2084.92±0.2092.66±0.0292.67±0.06
1-RL64.72±0.8965.70±0.7666.04±1.1063.69±1.7263.56±0.9267.03±0.8156.30±0.5866.03±1.9572.07±0.84
1-OE26.14±3.4131.28±3.1939.21±0.2131.84±4.2135.22±4.8337.47±1.5326.53±0.7139.35±1.8337.16±1.90
1-Cov58.13±0.7858.76±1.2059.06±0.9356.94±1.5256.55±0.8959.19±0.8153.99±0.3958.02±1.9765.64±1.21
AP34.45±1.3837.52±1.1440.18±0.6536.57±2.0337.16±2.1739.83±0.1827.49±0.3541.41±0.6843.86±1.15
AUC67.92±0.8068.84±1.0068.94±0.8766.52±1.4165.81±0.4969.76±0.6862.29±0.3968.55±2.0874.65±1.01
Ave.R6.835.172.506.506.503.678.673.501.67
Espauge1-HL97.88±0.1297.50±0.5698.40±0.0098.26±0.0198.41±0.0079.07±0.0494.75±4.9697.78±0.1898.25±0.01
1-RL62.26±0.3263.18±0.5063.94±0.3062.14±0.8464.67±0.6662.57±0.9854.05±0.6466.66±0.6069.79±0.47
1-OE14.49±1.2015.75±1.0814.97±0.7013.63±0.9118.75±0.2410.63±0.5417.27±7.211.48±1.1931.77±1.51
1-Cov29.73±0.3130.41±0.4933.06±0.3031.95±0.6634.65±0.6030.14±1.0125.73±0.1733.37±0.4335.80±0.55
AP9.83±0.3510.78±0.3211.28±0.1310.19±0.4713.38±0.299.66±0.555.82±0.096.77±0.7819.13±0.58
AUC62.15±0.3263.35±0.5664.12±0.2362.20±0.9066.28±0.7062.65±0.9156.03±0.0666.64±0.7969.86±0.54
Ave.R6.675.173.675.832.177.177.835.001.50
IAPR TC-121-HL97.54±0.0497.67±0.1898.00±0.0097.93±0.0398.00±0.0079.41±0.0288.22±1.2497.18±0.1498.04±0.02
1-RL61.86±0.3265.48±2.4566.90±0.1465.17±0.9169.20±0.3370.78±0.5755.79±0.9169.68±0.7372.66±0.48
1-OE14.86±0.6418.27±1.6829.51±0.8127.42±5.4433.89±4.6723.45±4.1615.02±9.561.49±1.3934.33±1.37
1-Cov25.98±0.4328.56±1.9529.24±0.3128.12±1.1631.06±0.5034.23±1.0823.23±0.2933.67±0.4635.40±0.53
AP9.18±0.2412.07±1.5715.10±0.1013.42±0.7817.37±0.8914.88±0.636.25±0.077.32±0.4319.66±0.07
AUC62.12±0.2465.52±2.3866.95±0.0965.38±0.6769.48±0.3170.87±0.6057.53±0.1169.75±0.8872.56±0.50
Ave.R7.505.833.835.673.174.008.505.501.00
MIR FLICKR1-HL86.63±0.2585.90±0.8787.46±0.0786.24±0.3187.12±0.1973.17±0.1079.90±0.1386.24±0.9487.61±0.09
1-RL75.11±0.4376.91±1.6176.23±0.2073.11±0.5074.97±0.9575.97±0.8860.07±0.1979.74±1.6882.34±0.75
1-OE44.98±0.7643.88±1.7636.58±0.7335.51±1.1935.78±1.0736.97±2.9642.44±0.3738.67±6.7261.34±0.99
1-Cov46.06±0.5948.98±2.7949.39±0.2346.30±0.4947.18±3.1348.15±2.1435.08±0.0953.12±4.1256.94±0.98
AP42.55±0.4244.11±0.7940.80±0.2038.50±0.1840.71±0.2242.18±0.5629.72±0.3445.69±2.7354.89±0.94
AUC73.96±0.4175.67±1.9175.42±0.1772.27±0.5873.97±1.3874.62±0.7262.80±0.1479.16±1.2881.14±0.93
Ave.R5.173.834.337.506.175.838.003.171.00
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.482, + 0.748 + ], + "angle": 0, + "content": "ing them with average imputation, followed by feature concatenation. All parameters of compared methods are configured as the recommended values in their original codes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Each dataset is divided into training, validation and test sets in the ratio of 7:1:2. To simulate the partial view setting, a certain proportion of instances defined by the partial example ratio (PER) are randomly selected as unavailable instances on each view. Besides, we ensure each sample contains at least one complete view to avoid invalid cases. In the multi-label noise configuration, we set a uniform noise rate for each label, i.e., \\(\\rho\\) of positive and negative labels are flipped to their opposite. All results are obtained from ten independent runs, with final outcomes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.75 + ], + "angle": 0, + "content": "reported as mean values alongside standard deviations. Our model is implemented by PyTorch on one NVIDIA GeForce RTX 4090 GPU of 24GB memory." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.72, + 0.774 + ], + "angle": 0, + "content": "3.2. Experimental Results." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Performance Evaluation. To evaluate the capabilities of DMMIvNL in addressing data deficiencies, we conduct extensive comparative experiments focusing on two critical aspects, i.e., label noise and view missing. Besides, we assign a broad range of noise rates and missing proportions to demonstrate the versatility of our approach. For label noise, \\(\\rho\\) is considered at levels of \\(\\{10\\%, 15\\%, 20\\%, 25\\%, 30\\%, 35\\%, 40\\}\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20711" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.175, + 0.09, + 0.825, + 0.103 + ], + "angle": 0, + "content": "DICNet DIMC MTD SIP AIMNet iMVWL TM3L NTMLCS DMMIvNL" + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.112, + 0.372, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.282, + 0.28, + 0.294 + ], + "angle": 0, + "content": "(a) Yeast" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.113, + 0.616, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.281, + 0.531, + 0.294 + ], + "angle": 0, + "content": "(b) Corel 5k" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.111, + 0.861, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.282, + 0.779, + 0.294 + ], + "angle": 0, + "content": "(c) VOC 2007" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.304, + 0.377, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.476, + 0.296, + 0.487 + ], + "angle": 0, + "content": "(d) ESP Game" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.305, + 0.616, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.476, + 0.542, + 0.487 + ], + "angle": 0, + "content": "(e) IAPR TC-12" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.306, + 0.861, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.475, + 0.786, + 0.487 + ], + "angle": 0, + "content": "(f) MIR FlickR" + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.504, + 0.812, + 0.519 + ], + "angle": 0, + "content": "Figure 2. AUC comparisons on six datasets with noise rate \\(\\rho\\) varying from \\(10\\%\\) to \\(40\\%\\) while PER=50%." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.547, + 0.287, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.681, + 0.226, + 0.693 + ], + "angle": 0, + "content": "(a) Espgame" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.547, + 0.482, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.681, + 0.43, + 0.692 + ], + "angle": 0, + "content": "(b) IAPR TC-12" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.71, + 0.483, + 0.739 + ], + "angle": 0, + "content": "Figure 3. Parameter analysis of the trade-off parameters \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) on Espgame and IAPR TC-12." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.856 + ], + "angle": 0, + "content": "while for view missing, PER encompasses values of \\(\\{10\\%, 30\\%, 50\\%, 70\\}\\). Tab. 1 displays all metrics with \\(\\rho\\) fixed at \\(30\\%\\) and PER fixed at \\(50\\%\\), Fig. 2 illustrates the variation in AUC as \\(\\rho\\) varies from \\(10\\%\\) to \\(40\\%\\), and Fig. 4 shows the metrics distribution when \\(\\mathrm{PER} = 10\\%\\) and \\(70\\%\\). Additional results are included in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Regarding the label noise, we can find that 1) Our method outperforms all compared approaches, particularly distinguished in the most representative AP metric. For" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.818 + ], + "angle": 0, + "content": "instance, our method achieves a remarkable AP improvement of up to \\(43\\%\\) on Espgame in Tab. 1. 2) Compared to NTMLCS for handling label noise, our approach reveals substantial advantages. Besides, in terms of the performance on VOC 2007, DMMIvNL exhibits greater stability than the stepwise estimation way of NTMLCS, which is achieved through the cycle-consistency framework that simultaneously determines the transition matrix and classifier. Additionally, our approach surpasses methods developed for incomplete MvMLC. For example, Fig. 2 illustrates that our DMMIvNL reliably improves performance even against the top-performing MvMLC methods MTD and AIMNet. 3) As \\(\\rho\\) increases from \\(10\\%\\) to \\(40\\%\\), DMMIvNL consistently outperforms the other eight methods. Moreover, many methods lose reliability at elevated noise levels, such as NTMLCS on Yeast, TM3L on MIR FLICKR and DIMC on Corel 5k, while our approach continues to learn a competent classifier and deliver robust performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As for the view missing, we have the following insights: 1) Our method exhibits outstanding performance in handling missing views. As illustrated in the radar chart of Fig. 4, DMMIvNL consistently occupies the outermost circle, reflecting impressive results across all metrics. 2)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20712" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.173, + 0.09, + 0.824, + 0.103 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.108, + 0.295, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.234, + 0.25, + 0.247 + ], + "angle": 0, + "content": "(a) Yeast, PER=10%" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.111, + 0.498, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.234, + 0.468, + 0.247 + ], + "angle": 0, + "content": "(b) VOC 2007, PER=10%" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.111, + 0.702, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.234, + 0.667, + 0.247 + ], + "angle": 0, + "content": "(c) Espgame, PER=10%" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.111, + 0.905, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.234, + 0.883, + 0.247 + ], + "angle": 0, + "content": "(d) MIR FLICKR, PER=10%" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.258, + 0.295, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.382, + 0.25, + 0.394 + ], + "angle": 0, + "content": "(e) Yeast, PER=70%" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.259, + 0.498, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.382, + 0.466, + 0.394 + ], + "angle": 0, + "content": "(f) VOC 2007, PER=70%" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.259, + 0.702, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.382, + 0.666, + 0.395 + ], + "angle": 0, + "content": "(g) Espgame, PER=70%" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.259, + 0.904, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.382, + 0.882, + 0.394 + ], + "angle": 0, + "content": "(h) MIR FLICKR,PER=70%" + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.411, + 0.816, + 0.427 + ], + "angle": 0, + "content": "Figure 4. Experimental results of nine methods on four datasets with PER=10% and 70% while \\(\\rho = 30\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.484, + 0.634 + ], + "angle": 0, + "content": "As PER changes from \\(10\\%\\) to \\(70\\%\\), DMMIvNL continues to achieve top performance. Notably, while DIMC experiences a substantial decline at \\(70\\%\\) PER, our method remains efficient. It indicates that DMMIvNL delivers strong robustness across varying levels of feature absence. 3) Our method makes remarkable breakthroughs over deep learning based methods for missing views, like SIP, MTD, and DICNet. For example, on MIR FLICKR, DMMIvNL improves 1-OE by nearly \\(50\\%\\) compared to SIP at \\(50\\%\\) PER. This improvement is primarily due to the focus on extracting the shared semantics and specific features through information separation, an aspect overlooked by other methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ablation Study and Parameter Sensitivity. The ablation experiments are conducted to analyze the effect of the three crucial aspects of DMMIvNL, i.e., representation extraction based on the information bottleneck theory \\((\\mathcal{L}_{IB})\\), cycle-consistency estimation for noise identification and classifier learning \\((\\mathcal{L}_{ID})\\), and model regularization to combat overfitting \\((\\mathcal{L}_r)\\). Note that loss \\(\\mathcal{L}_c\\) is retained since it is essential for classifier improvement during early training. The first three rows of Tab. 2 indicate that \\(\\mathcal{L}_{IB}\\), \\(\\mathcal{L}_{ID}\\) and \\(\\mathcal{L}_r\\) are integral components of DMMIvNL. Besides, the comparison between the empirical loss \\(\\mathcal{L}_{vol}\\) from problem (11) and \\(\\mathcal{L}_{ID}\\) confirms that the cycle-consistency framework enhances the stability of transition matrix estimation. To study the impact of \\(\\lambda_1\\) and \\(\\lambda_2\\), we present the results under distinct parameter settings in Fig. 3. As shown in the heatmap, the performance of DMMIvNL exhibits variability and improves as \\(\\lambda_1\\) gets closer to 0.001 and \\(\\lambda_2\\) approaches 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.45, + 0.907, + 0.494 + ], + "angle": 0, + "content": "Table 2. Ablation study on Yeast, Corel 5k and VOC 2007 with PER=50% and ρ = 30%. ‘√’ and ‘X’ represent the used and not used corresponding item, respectively." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.598 + ], + "angle": 0, + "content": "
LvolLIBLIDLrEspgameCorel 5kVOC 2007
APAUCAPAUCAPAUC
XX✓✓0.17500.68460.17130.69720.41450.7366
X✓X✓0.18380.71130.18500.70890.43050.7367
X✓✓X0.18340.71310.18000.70960.43590.7148
✓X✓✓0.14990.67320.15430.62840.36120.6715
X✓✓✓0.19340.73070.19970.72840.46220.7566
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.64, + 0.634, + 0.656 + ], + "angle": 0, + "content": "4. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.674, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To address the dual challenges of view missing and multi-label noise, we propose DMMIvNL in this paper. Built upon the information bottleneck theory, a mutual information model is established to constrain information exchange across diverse representations, along with a derived theoretical upper bound. Besides, we prove that minimizing the volume of the transition matrix for each category consistently leads to the optimal classifier. Moreover, stability in noise recognition is further improved by introducing a cycle-consistency estimation framework, while raw fidelity information and inherent label correlations are used to reduce overfitting. Finally, extensive experiments demonstrate the superiority of DMMIvNL. In future work, we will leverage the prior knowledge embedded in LLM to rectify inaccurate labels and tackle the issue of class imbalance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20713" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.161 + ], + "angle": 0, + "content": "This work was partially supported by the National Natural Science Foundation of China [Grant No. 62376281], and the Key NSF of China under Grant No. 62136005." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.176, + 0.187, + 0.192 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.483, + 0.244 + ], + "angle": 0, + "content": "[1] Luis Von Ahn and Laura Dabbish. Labeling images with a computer game. In SIGCHI Conference on Human Factors in Computing Systems, page 319-326, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.245, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[2] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. In ICML, pages 825-836, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[3] Jia-Yao Chen, Shao-Yuan Li, Sheng-Jun Huang, and Chm. Unm: A universal approach for noisen, songcan and wang, lei and xie, ming-kun. IEEE Transactions on Knowledge and Data Engineering, 36:4968-4980, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.414 + ], + "angle": 0, + "content": "[4] De Cheng, Yixiong Ning, Nannan Wang, Xinbo Gao, Heng Yang, Yuxuan Du, Bo Han, and Tongliang Liu. Class-dependent label-noise learning with cycle-consistency regularization. pages 11104-11116, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[5] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In ECCV, pages 97-112, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.527 + ], + "angle": 0, + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[7] Yuling Fan, Jinghua Liu, Jianeng Tang, Peizhong Liu, Yaojin Lin, and Yongzhao Du. Learning correlation information for multi-label feature selection. Pattern Recognition, 145: 109899, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[8] Marco Federici, Anjan Dutta, Patrick Forre, Nate Kushman, and Zeynep Akata. Learning robust representations via multi-view information bottleneck. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[9] Xiao Fu, Kejun Huang, and Nicholas D Sidiropoulos. On identifiability of nonnegative matrix factorization. IEEE Signal Processing Letters, 25(3):328-332, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[10] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International Workshop onto Image, pages 1–11, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[11] Cordelia Guillaumin. Multimodal semi-supervised learning for image classification. In CVPR, pages 902-909, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, I Tsang, and Masashi Sugiyama. Robust training of deep neural networks with extremely noisy labels. In NeurIPS, page 4, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[13] Pingting Hao, Weiping Ding, Wanfu Gao, and Jialong He. Exploring view-specific label relationships for multi-view multi-label feature selection. Information Sciences, 681: 121215, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[14] Huaxi Huang, Hui Kang, Sheng Liu, Olivier Salvado, Thierry Rakotoarivelo, Dadong Wang, and Tongliang Liu." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.201, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "Paddles: Phase-amplitude spectrum disentangled early stopping for learning with noisy labels. In ICCV, pages 16719-16730, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[15] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In ICMR, pages 39-43, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[16] Xiaodong Jia, Xiao-Yuan Jing, Xiaoke Zhu, Songcan Chen, Bo Du, Ziyun Cai, Zhenyu He, and Dong Yue. Semi-supervised multi-view deep discriminant representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2496-2509, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[17] Zhangqi Jiang, Tingjin Luo, and Xinyan Liang. Deep incomplete multi-view learning network with insufficient label information. In AAAI, pages 12919-12927, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[18] Rihuan Ke. Deep variation prior: Joint image denoising and noise variance estimation without clean data. IEEE Transactions on Image Processing, 33:2908-2923, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[19] Shikun Li, Xiaobo Xia, Hansong Zhang, Yibing Zhan, Shiming Ge, and Tongliang Liu. Estimating noise transition matrix with label correlations for noisy multi-label learning. In NeurIPS, pages 24184-24198, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[20] Xuefeng Li, Tongliang Liu, Bo Han, Gang Niu, and Masashi Sugiyama. Provably end-to-end label-noise learning without anchor points. In ICML, pages 6403-6413, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[21] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4447-4461, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[22] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In AAAI, pages 8807-8815, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[23] Chengliang Liu, Jie Wen, Xiaoling Luo, and Yong Xu. Incomplete multi-view multi-label learning via label-guided masked view-and category-aware transformers. In AAAI, pages 8816-8824, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.66 + ], + "angle": 0, + "content": "[24] Chengliang Liu, Jinlong Jia, Jie Wen, Yabo Liu, Xiaoling Luo, Chao Huang, and Yong Xu. Attention-induced embedding imputation for incomplete multi-view partial multi-label classification. In AAAI, pages 13864-13872, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[25] Chengliang Liu, Jie Wen, Yabo Liu, Chao Huang, Zhihao Wu, Xiaoling Luo, and Yong Xu. Masked two-channel decoupling framework for incomplete multi-view weak multi-label learning. In NeurIPS, 2024. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[26] Chengliang Liu, Gehui Xu, Jie Wen, Yabo Liu, Chao Huang, and Yong Xu. Partial multi-view multi-label classification via semantic invariance learning and prototype modeling. In ICML, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[27] Chengliang Liu, Jie Wen, Yong Xu, Bob Zhang, Liqiang Nie, and Min Zhang. Reliable representation learning for incomplete multi-view missing multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3):447-461, 2015. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20714" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[29] Tingjin Luo, Yang Yang, Dongyun Yi, and Jieping Ye. Robust discriminative feature learning with calibrated data reconstruction and sparse low-rank model. Applied Intelligence, 54(3):2867-2880, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[30] Yong Luo, Dacheng Tao, Chang Xu, Chao Xu, Hong Liu, and Yonggang Wen. Multiview vector-valued manifold regularization for multilabel image classification. IEEE Transactions on Neural Networks and Learning Systems, 24(5): 709-722, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[31] Yubin Park and Joyce C Ho. Tackling overfitting in boosting for noisy healthcare data. IEEE Transactions on Knowledge and Data Engineering, 33(7):2995-3006, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[32] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, pages 1944-1952, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[33] Yalan Qin, Hanzhou Wu, Xinpeng Zhang, and Guorui Feng. Semi-supervised structured subspace learning for multi-view clustering. IEEE Transactions on Image Processing, 31:1-14, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[34] Alisa Smirnova, Jie Yang, Dingqi Yang, and Philippe CudreMauroux. Nessy: A neuro-symbolic system for label noise reduction. IEEE Transactions on Knowledge and Data Engineering, 35(8):8300-8311, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[35] Shiding Sun, Bo Wang, and Yingjie Tian. Decoupled representation for multi-view learning. Pattern Recognition, 151: 110377, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[36] Yu-Yin Sun, Yin Zhang, and Zhi-Hua Zhou. Multi-label learning with weak label. In AAAI, pages 593-598, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[37] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In IJCAI, pages 2703-2709, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[38] Ziyi Wang and Yitian Xu. A two-stage multi-view partial multi-label learning for enhanced disambiguation. Knowledge-Based Systems, 293:111680, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[39] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems, 35(8):11396-11408, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[40] Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning with noisy label identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3676-3687, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[41] Wulin Xie, Xiaohuan Lu, Yadong Liu, Jiang Long, Bob Zhang, Shuping Zhao, and Jie Wen. Uncertainty-aware pseudo-labeling and dual graph driven network for incomplete multi-view multi-label classification. In ACMMM, pages 6656-6665, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[42] Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. Instance-dependent label-noise learning under a structural causal model. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[43] Yitong Yao, Jing Zhang, Peng Zhang, and Yueheng Sun. A dual-branch learning model with gradient-balanced loss for long-tailed multi-label text classification. ACM Transactions on Information Systems, 42(2):1-24, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[44] Selim F Yilmaz, E Batuhan Kaynak, Aykut Koç, Hamdi Dibeklioglu, and Suleyman Serdar Kozat. Multi-label sentiment analysis on 100 languages with dynamic weighting for label imbalance. IEEE Transactions on Neural Networks and Learning Systems, 34:331-343, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[45] Dawei Zhao, Qingwei Gao, Yixiang Lu, and Dong Sun. Two-step multi-view and multi-label learning with missing label via subspace learning. Applied Soft Computing, 102:107120, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[46] Pengfei Zhu, Qi Hu, Qinghua Hu, Changqing Zhang, and Zhizhao Feng. Multi-view label embedding. Pattern Recognition, 84:126-135, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[47] Wenzhang Zhuge, Tingjin Luo, Ruidong Fan, Hong Tao, Chenping Hou, and Dongyun Yi. Absent multiview semisupervised classification. IEEE Transactions on Cybernetics, 54(3):1708-1721, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20715" + } + ] +] \ No newline at end of file diff --git a/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_origin.pdf b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a398c0f76eefab306a4533b4ed84c88f53499a1d --- /dev/null +++ b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/fa88376e-5594-45d4-abf8-c70bbaaceb86_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42a4ac1662f06b34321356cb041c93edf7f200e2aca2b27d68928bcfcb270e60 +size 1878973 diff --git a/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/full.md b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7106ea1a7de690e19d277cef1eb4459298b7c6d8 --- /dev/null +++ b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/full.md @@ -0,0 +1,333 @@ +# Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels + +Quanjiang Li, Tingjin Luo*, Jiahui Liao National University of Defense Technology, China + +liquanjiang@nudt.edu.cn,tingjinluo@hotmail.com,liaojiahui19@nudt.edu.cn + +# Abstract + +Incomplete features and label noise in multi-view multi-label data significantly undermine the reliability and performance, motivating researchers to explore the mechanism of representation and information recovery. However, learning for such dual deficiencies is crucial but rarely studied. In this paper, we propose a theory-inspired Deep Multi-View Multi-Label Learning method with Incomplete Views and Noisy Labels named DMMIvNL to address these problems. Specifically, to promote the synthesis of task-relevant shared information and preserve the distinctiveness of individual features from limited views, we have developed a feature extraction modular based on the information bottleneck theory, and formulated its theoretical upper bound into its objective. Meanwhile, we theoretically prove that minimizing the volume of the transition matrix ensures the statistical consistency with classifier training. Besides, a cycle-consistency estimation principle is proposed in the volume minimization network to improve the recognition stability of multi-label noise. Moreover, leveraging inherent real semantics information and label correlations are employed as model regularization to reduce the risk of excessive noise fitting. Finally, extensive experimental results validate the effectiveness and robustness of our DMMIvNL. + +# 1. Introduction + +Advances in feature extraction has shifted data analysis towards multi-view frameworks [27], while growing data attributes drive multi-label learning to offer rich label semantics [13]. Therefore, Multi-view multi-label classification (MvMLC) has emerged to address the intricate demands of processing diverse features and numerous targets. Traditional methods [30, 46] assume the given data has complete views and clean labels, which is violated in practice [41]. On one hand, heterogeneous data from diverse sources frequently encounters missing views due to acquire + +sition complexities and storage limitations [29, 47]. On the other hand, real-world collections often contain label noise as manual annotation can introduce incorrect labels beyond human knowledge [34]. Therefore, insufficient features and noisy labels severely reduce the classification performance. Regarding MvMLC, many solutions have been proposed to address missing views. iMvWL [37] used weighted matrix factorization to learn a shared subspace and AIMNet [24] leveraged a cross-view attention mechanism to recover missing views. Moreover, two types of label deficiencies have been considered in MvMLC. The problem of Partial Multi-Label (PML) [40] involves both relevant and irrelevant labels, while the problem of Multi-Label Learning with Missing Labels (MLML)[36] occurs when some relevant labels are missing. TFMDD [38] addressed PML with a two-stage manifold disambiguation model and LMVCAT [23] tackled MLML by employing label-guided graphs and category-aware transformers. While existing methods address each annotation problem separately, both issues tend to occur simultaneously in practice. Thus, the challenge of multi-label noise demands substantial attention, where each category suffers from both over-labeled and missing labels. + +Existing research has not yet tackled the combined issues of missing views and multi-label noise, which are prevalent in real-world applications like incomplete and subjective medical diagnostics [31]. In fact, overcoming the simultaneous challenges requires refined representation extraction under missing views and effective classifier leaning with multi-label noise. The core of multi-view learning lies in extracting shared information across different views while preserving unique features of each view [33]. DICNet [22], DIMC [39] and SIP [26] concentrated solely on shared subspace and neglected distinctive characteristics of individual views. MDDRL [16] and MTD [25] attempted to extract both shared and specific representations with minimal interactions. However, their reliance on geometric orthogonality or distance metrics restricted them to capture only linear patterns, overlooking the complex nature of inter-view relations. Therefore, constraining feature semantics at the intrinsic information level is crucial for obtaining high-level + +representations with precise distinction. Most approaches for handling multi-label noise utilize strategies like early stopping [14] and cyclical learning rates [3]. These methods often lack stability and require dataset-specific adjustments, making them insufficiently adaptive. ENTMLC [19] employed label correlations to identify the noise transition matrix for recovering clean labels. However, the decoupling of transition matrix acquisition and classifier training causes error accumulation, limiting the practical utility. To circumvent the uncertainty inherent in those strategies and the inefficiency of stepwise noise learning, a statistically consistent estimation framework should be developed, which would allow for direct training the optimal classifier from noisy labels and achieving strong robustness. + +To tackle these problems, we propose a theory-inspired method with deep networks named DMMIvNL. Specifically, a mutual information model is developed to enhance synergy of shared features, suppress interactions from specific features, and impose orthogonality between shared and specific features of each view. The upper bound of the model is derived through the variational autoencoder framework, with instance-level contrastive operations applied to simulate diverse joint and marginal distributions. Moreover, the ideal classifier is trained on noisy data by minimizing the volume of the noise transition matrix and the process is proven to converge to a unique and globally optimal solution under the sufficiently scattered assumption. The difficulty in guaranteeing the assumption and relying exclusively on noisy class-posterior estimation promotes the design of a cycle-consistency estimation method, which integrates the volume maximization of clean class-posterior. Finally, the real semantics embedded in noisy data and hidden label correlation information are employed as model regularization to reduce the risk of noise overfitting. Our main contributions are summarized as follows: + +- This is the first multi-view multi-label learning framework capable of handling incomplete views and noisy labels, including both over-labeled and missing labels. +- Our mutual information based model extracts semantically distinct representations from limited views, with its upper bound established through variational derivation. +- We theoretically prove the statistical consistency and joint convergence between minimizing the noise transition matrix volume for each category and classifier training. Besides, a cycle-consistency loss framework is designed to enhance transition matrix estimation. + +# 2. Methodology + +# 2.1. Problem Definition + +The multi-view dataset $\{\mathbf{X},\mathbf{Y}\}$ contains $N$ samples, where $\mathbf{X}_i$ consists of $V$ views with dimension $d_v$ , i.e., $\mathbf{X}_i = \{\pmb{x}_i^{(v)}\in \mathbb{R}^{d_v}\}_{v = 1}^V$ . The label matrix $\mathbf{Y}\in \{0,1\}^{N\times C}$ corre + +sponds to $C$ categories, with $Y_{i,j} = 1$ if sample $i$ belongs to class $j$ , and $Y_{i,j} = 0$ otherwise. $\mathbf{W} \in \mathbb{R}^{N \times V}$ is an indicator matrix, where $W_{i,j} = 1$ indicates the $j$ -th view is available for sample $i$ , and $W_{i,j} = 0$ means missing. Since precise annotation is hard to obtain, the training dataset in practice is $\{\mathbf{X}, \bar{\mathbf{Y}}\}$ , where $\bar{\mathbf{Y}}$ denotes noisy labels. Thus, we denote the clean and noisy labels of the $j$ -th class as $\mathbf{Y}^j$ and $\bar{\mathbf{Y}}^j$ , related by a noise transition matrix $\mathbf{T}^j \in [0,1]^{2 \times 2}$ , $j \in [C]$ . When only noisy samples are available and without further constraints, instance-dependent transition matrix becomes non-identifiable [42]. Therefore, we employ a general assumption that the noise transition matrix is class-dependent and instance-independent, i.e., $P(\bar{\mathbf{Y}}^j = k \mid \mathbf{Y}^j = i, \mathbf{X} = x) = P(\bar{\mathbf{Y}}^j = k \mid \mathbf{Y}^j = i) = T_{i,k}^j$ . + +# 2.2. Feature Extraction via Information Bottleneck + +Deriving effective representations from raw data aids in reducing noise and redundancy [17]. Consequently, both task-relevant shared semantics and view-invariant characteristic features should be explored to enhance information processing efficiency. Let $z^v$ and $s^v$ denote the $D$ -dimensional shared and specific representation extracted from each view and the goal is to ensure consistency of the shared features across multiple views and preserve the unique contributions of each view. Considering different interaction patterns between representations from $x^{v^*}$ and $x^v$ , mutual information $I(z^{v^*}|x^{v^*};z^v|x^v)$ should be maximized to capture only universally shared information, while $I(s^{v^*}|x^{v^*};s^v|x^v)$ should be minimized to preserve only individual attributes. Besides, the shared and specific features should be as orthogonal as possible to ensure absolute distinction in information content [35], i.e., $I(z^v |x^v;s^v |x^v) = 0$ . Thus, the constrained optimization problem based on the information bottleneck theory [8] can be formulated as + +$$ +\max \frac {1}{V} \sum_ {v = 1} ^ {V} \sum_ {v ^ {*} \neq v} ^ {V} \left(I \left(\boldsymbol {z} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}; \boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}\right) - I \left(\boldsymbol {s} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}; \boldsymbol {s} ^ {v} \mid \boldsymbol {x} ^ {v}\right)\right) +$$ + +$$ +s. t. \quad \frac {1}{V} \sum_ {v = 1} ^ {V} I \left(\boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}; \boldsymbol {s} ^ {v} \mid \boldsymbol {x} ^ {v}\right) = 0. \tag {1} +$$ + +By applying the Lagrange multiplier method, the problem (1) can be equally converted as + +$$ +\begin{array}{l} \max \frac {1}{V} \sum_ {v = 1} ^ {V} \sum_ {v ^ {*} \neq v} ^ {V} \left(I \left(\boldsymbol {z} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}; \boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}\right) - I \left(\boldsymbol {s} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}; \boldsymbol {s} ^ {v} \mid \boldsymbol {x} ^ {v}\right)\right) \\ - \frac {\beta}{V} \sum_ {v = 1} ^ {V} I \left(\boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}; \boldsymbol {s} ^ {v} \mid \boldsymbol {x} ^ {v}\right), \tag {2} \\ \end{array} +$$ + +where $\beta \geq 0$ is the Lagrange multiplier to balance the efficiency and compactness of information [26]. For the first + +![](images/4e7cf2f6120729b621481246e5f816945ca06cbe7ca32b7ec89ba18c2e7120f8.jpg) +Figure 1. The main framework of our proposed DMMIvNL. Different shapes signify different samples. + +term, we can obtain the following lower bound: + +$$ +\begin{array}{l} I \left(\boldsymbol {z} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}; \boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}\right) \\ = \iint p \left(\boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}, \boldsymbol {z} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}\right) \log \frac {p \left(\boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v} , \boldsymbol {z} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}\right)}{p \left(\boldsymbol {z} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}\right) p \left(\boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}\right)} d \boldsymbol {z} ^ {v ^ {*}} d \boldsymbol {z} ^ {v} \\ \geq \int \int p (\boldsymbol {z} ^ {v} | \boldsymbol {z} ^ {v ^ {*}}) p (\boldsymbol {z} ^ {v ^ {*}} | \boldsymbol {x} ^ {v ^ {*}}) \log p (\boldsymbol {z} ^ {v} | \boldsymbol {z} ^ {v ^ {*}}) d \boldsymbol {z} ^ {v ^ {*}} d \boldsymbol {z} ^ {v}, \\ \end{array} +$$ + +where $p(\cdot)$ denotes the probability density function. Since computing $p(z^v |z^{v^*})$ is intractable, we approximate it using a stochastic variational distribution $q(z^{v}|z^{v^{*}})$ . Then we can obtain a new variational lower bound for maximization: + +$$ +\begin{array}{l} I \left(\boldsymbol {z} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}; \boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}\right) \\ \geq \int \int p (\boldsymbol {z} ^ {v} | \boldsymbol {x} ^ {v}, \boldsymbol {z} ^ {v ^ {*}} | \boldsymbol {x} ^ {v ^ {*}}) \log q (\boldsymbol {z} ^ {v} | \boldsymbol {z} ^ {v ^ {*}}) d \boldsymbol {z} ^ {v ^ {*}} d \boldsymbol {z} ^ {v} \\ = D _ {K L} ^ {\dagger} \left(p \left(\boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}, \boldsymbol {z} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}\right) \| q \left(\boldsymbol {z} ^ {v} \mid \boldsymbol {z} ^ {v ^ {*}}\right)\right) = D _ {K L} ^ {\dagger} \left(\boldsymbol {z} ^ {v}, \boldsymbol {z} ^ {v ^ {*}}\right), \tag {3} \\ \end{array} +$$ + +where $D_{KL}(\cdot ||\cdot)$ means the Kullback-Leibler divergence and $D_{KL}^{\dagger}(\cdot ||\cdot)$ is a pseudo-symbol for notational consistency. For the middle term, the upper bound can be obtained by introducing another variational distribution $q(s^{v}|s^{v^{*}})$ : + +$$ +\begin{array}{l} I \left(\boldsymbol {s} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}; \boldsymbol {s} ^ {v} \mid \boldsymbol {x} ^ {v}\right) \\ \leq D _ {K L} \left(p \left(\boldsymbol {s} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}, \boldsymbol {s} ^ {v} \mid \boldsymbol {x} ^ {v}\right) \| p \left(\boldsymbol {s} ^ {v ^ {*}} \mid \boldsymbol {x} ^ {v ^ {*}}\right) q \left(\boldsymbol {s} ^ {v} \mid \boldsymbol {s} ^ {v ^ {*}}\right)\right) \tag {4} \\ = D _ {K L} ^ {\left(\mathbf {s} ^ {v}, \mathbf {s} ^ {v ^ {*}}\right)}. \\ \end{array} +$$ + +Similarly, the upper bound for the last term is + +$$ +I \left(\boldsymbol {z} ^ {v} \mid \boldsymbol {x} ^ {v}; \boldsymbol {s} ^ {v} \mid \boldsymbol {x} ^ {v}\right) \leq D _ {K L} ^ {\left(\boldsymbol {s} ^ {v}, \boldsymbol {z} ^ {v}\right)}. \tag {5} +$$ + +Combining Eqs. (3), (4) and (5), our optimization objective becomes minimizing the upper bound of the problem (2): + +$$ +\mathcal {L} _ {I B} = \frac {1}{V} \sum_ {v = 1} ^ {V} \sum_ {v ^ {*} \neq v} ^ {V} \left(- D _ {K L} ^ {\dagger \left(\boldsymbol {z} ^ {v}, \boldsymbol {z} ^ {v ^ {*}}\right)} + D _ {K L} ^ {\left(\boldsymbol {s} ^ {v}, \boldsymbol {s} ^ {v ^ {*}}\right)}\right) + \frac {\beta}{V} \sum_ {v = 1} ^ {V} D _ {K L} ^ {\left(\boldsymbol {s} ^ {v}, \boldsymbol {z} ^ {v}\right)}. \tag {6} +$$ + +Contrastive learning [21] is used to compute various distributions. Specifically, representations are treated as probability vectors over $D$ classes via a Softmax activation function, and the joint distribution matrix is obtained by + +$$ +\boldsymbol {P} ^ {(v, v ^ {*})} = \sum_ {i = 1} ^ {N} \left(\boldsymbol {z} _ {i} ^ {(v)}\right) ^ {T} \boldsymbol {z} _ {i} ^ {(v ^ {*})}, \boldsymbol {Q} ^ {(v, v ^ {*})} = \sum_ {i = 1} ^ {N} \left(\boldsymbol {s} _ {i} ^ {(v)}\right) ^ {T} \boldsymbol {s} _ {i} ^ {(v ^ {*})}. +$$ + +Due to the strong interaction between $z^v$ and $z^{v^*}$ , the variational distribution $q(z^v |z^{v^*})$ can be estimated by the obtained conditional distribution. Given the extremely weak correlation between $s^v$ and $s^{v^*}$ , as well as $z^v$ and $s^v$ , the distribution $q(s^v /s^{v^*})$ and $q(z^{v} / s^{v})$ can be approximated by the corresponding marginal distribution. Subsequently, each term in Eq. (6) can be expressed as + +$$ +\left\{ \begin{array}{l} D _ {K L} ^ {\dagger \left(\boldsymbol {z} ^ {v}, \boldsymbol {z} ^ {v ^ {*}}\right)} = \sum_ {d = 1} ^ {D} \sum_ {d ^ {\prime} = 1} ^ {D} \boldsymbol {P} _ {d, d ^ {\prime}} ^ {(v, v ^ {*})} \log \frac {\boldsymbol {P} _ {d , d ^ {\prime}} ^ {(v , v ^ {*})}}{\boldsymbol {P} _ {d} ^ {(v)}} \\ D _ {K L} ^ {\left(\boldsymbol {s} ^ {v}, \boldsymbol {s} ^ {v ^ {*}}\right)} = \sum_ {d = 1} ^ {D} \sum_ {d ^ {\prime} = 1} ^ {D} \boldsymbol {Q} _ {d, d ^ {\prime}} ^ {(v, v ^ {*})} \log \frac {\boldsymbol {Q} _ {d , d ^ {\prime}} ^ {(v , v ^ {*})}}{\left(\boldsymbol {Q} _ {d} ^ {(v)}\right) ^ {\alpha} \boldsymbol {Q} _ {d ^ {\prime}} ^ {(v ^ {*})}} \\ D _ {K L} ^ {\left(\boldsymbol {s} ^ {v}, \boldsymbol {z} ^ {v}\right)} = \sum_ {d = 1} ^ {D} \sum_ {d ^ {\prime} = 1} ^ {D} \boldsymbol {M} _ {d, d ^ {\prime}} ^ {\left(z ^ {v}, s ^ {v}\right)} \log \frac {\boldsymbol {M} _ {d , d ^ {\prime}} ^ {\left(z ^ {v} , s ^ {v}\right)}}{\left(\boldsymbol {M} _ {d} ^ {\left(z ^ {v}\right)}\right) ^ {\alpha} \boldsymbol {M} _ {d ^ {\prime}} ^ {\left(s ^ {v}\right)}}, \end{array} \right. \tag {7} +$$ + +where the marginal distributions $P_{d}^{(v)}$ and $P_{d^{\prime}}^{(v^{*})}$ can be obtained by summing over the $d$ -th row and the $d^{\prime}$ -th column of the joint distribution matrix $P^{(v,v^{*})}$ , with other symbols defined as the same. In this conversion, we let $q(\boldsymbol{s}^v /\boldsymbol{s}^{v^*}) = (Q^{(v)})^\alpha$ and $q(z^{v} / s^{v}) = (M^{(z^{v})})^{\alpha}$ , where $\alpha$ is a balance factor to preserve crucial information and ensure model stability [16]. By introducing the mask matrix $\pmb{W}$ to mitigate the impact of unavailable instances, we can perform cross-view fusion to acquire the information-enriched shared and specific representation: + +$$ +\bar {\boldsymbol {Z}} _ {i,:} = \sum_ {v = 1} ^ {V} \frac {\boldsymbol {z} _ {i , :} ^ {(v)} \boldsymbol {W} _ {i , v}}{\sum_ {v} \boldsymbol {W} _ {i , v}}, \quad \bar {\boldsymbol {S}} _ {i,:} = \sum_ {v = 1} ^ {V} \frac {\boldsymbol {s} _ {i , :} ^ {(v)} \boldsymbol {W} _ {i , v}}{\sum_ {v} \boldsymbol {W} _ {i , v}}. \tag {8} +$$ + +Departing from traditional addition or concatenation, we conduct feature interaction by enhancing shared information with proprietary data [25]: + +$$ +\hat {\boldsymbol {X}} _ {i, j} = \theta_ {s} \left(\bar {\boldsymbol {Z}} _ {i, j}\right) \cdot \bar {\boldsymbol {S}} _ {i, j}, \tag {9} +$$ + +where $\theta_{s}$ is the sigmoid activation function and $\hat{\pmb{X}}\in \mathbb{R}^{N\times D}$ is the final fusion representation. + +# 2.3. Identify Multi-Label Noise by Volume Minimization Network + +The success of deep learning relies on large-scale accurately labeled data, with performance significantly dropping under noisy data [12], making noise identification crucial for network training [2]. In multi-class tasks, many statistically consistent methods based on the transition matrix have been developed [28, 32], which guarantees the classifier learned from noisy data can converge to the optimal classifier under the clean risk. The anchor-point assumption [28] is commonly utilized to determine the transition matrix, which postulates the existence of samples that belong to each class with probability one. However, implementing this approach in multi-label learning is challenging due to the difficulty of pinpointing suitable anchor samples for each label. Motivated by [20], we capitalize on the geometric properties of the transition matrix to seek a breakthrough. The transition matrix for the $j$ -th label is defined as + +$$ +\boldsymbol {T} ^ {j} = \left( \begin{array}{l l} P \left(\bar {\boldsymbol {Y}} ^ {j} = 0 \mid \boldsymbol {Y} ^ {j} = 0\right) & P \left(\bar {\boldsymbol {Y}} ^ {j} = 0 \mid \boldsymbol {Y} ^ {j} = 1\right) \\ P \left(\bar {\boldsymbol {Y}} ^ {j} = 1 \mid \boldsymbol {Y} ^ {j} = 0\right) & P \left(\bar {\boldsymbol {Y}} ^ {j} = 1 \mid \boldsymbol {Y} ^ {j} = 1\right) \end{array} \right). \tag {10} +$$ + +Given an instance $\hat{\pmb{x}}$ , the noisy class-posterior for the $j$ -th label is $P(\bar{Y}^j|\mathbf{X} = \hat{\pmb{x}}) = [P(\bar{Y}^j = 0|\mathbf{X} = \hat{\pmb{x}}), P(\bar{Y}^j = 1|\mathbf{X} = \hat{\pmb{x}})]^T$ and the clean class-posterior is $P(\mathbf{Y}^j|\mathbf{X} = \hat{\pmb{x}}) = [P(\mathbf{Y}^j = 0|\mathbf{X} = \hat{\pmb{x}}), P(\mathbf{Y}^j = 1|\mathbf{X} = \hat{\pmb{x}})]^T$ . Thus, the relationship $P(\bar{Y}^j|\mathbf{X} = \hat{\pmb{x}}) = \pmb{T}^j P(\mathbf{Y}^j|\mathbf{X} = \hat{\pmb{x}})$ holds. Since $\sum_{i=0}^{1} P(\mathbf{Y} = i|\mathbf{X} = \hat{\pmb{x}}) = 1$ , $P(\bar{Y}^j|\mathbf{X} = \hat{\pmb{x}})$ is then a convex combination of the columns of $\pmb{T}^j$ , which implies that the noisy class-posterior is enclosed within the simplex $\mathrm{Sim}\{\pmb{T}^j\}$ . To reduce noise identification errors, it + +is essential to recover $\operatorname{Sim}\{\pmb{T}^j\}$ such that it tightly encompasses all sample posterior probability points, thereby making volume minimization of $\operatorname{Sim}\{\pmb{T}^j\}$ the primary goal. For synchronized learning, we define a trainable $\hat{T}^j \in [0,1]^{2 \times 2}$ with $\sum_{i=1}^{2} \hat{T}^j{}_{i,k} = 1$ , and construct a multi-label classifier $h_\theta$ with parameters $\theta$ . The optimization of $\hat{T}^j$ and $h_\theta(\hat{x})$ involves two objectives: the discrepancy between $\hat{T}^j(h_\theta(\hat{x})_j, 1 - h_\theta(\hat{x})_j)^T$ and $P(\bar{Y}^j | X = \hat{x})$ and the volume of each simplex spanned by the columns of $\hat{T}^j$ . Therefore, the following model to jointly learn the transition matrix and classifier is proposed: + +$$ +\left\{ \begin{array}{l} \min _ {\hat {\boldsymbol {T}} ^ {j}} \sum_ {j = 1} ^ {C} \operatorname {v o l} \left(\hat {\boldsymbol {T}} ^ {j}\right) \\ s. t. \hat {\boldsymbol {T}} ^ {j} \left[ \begin{array}{c} h _ {\theta} (\hat {\boldsymbol {x}}) _ {j} \\ 1 - h _ {\theta} (\hat {\boldsymbol {x}}) _ {j} \end{array} \right] = P \left(\bar {\boldsymbol {Y}} ^ {j} \mid X = \hat {\boldsymbol {x}}\right), \end{array} \right. \tag {11} +$$ + +where $\mathrm{vol}(\hat{T}^j)$ denotes a measure to the volume of $\operatorname{Sim}\{\hat{T}^j\}$ . In the absence of anchor points, the class-posterior probability should incorporate substantial information to uniquely determine the minimum-volume simplex, with true sample probability points of each class being distinguishable. Therefore, the sufficiently scattered assumption [9] is introduced to limit the occurrence of anomalous clean class-posterior and aid in theoretically establishing that addressing problem (11) will drive $\hat{T}^j$ to converge to $\pmb{T}^j$ and $h_\theta (\hat{\pmb{x}})$ to accurately approximate $P(\mathbf{Y}|\mathbf{X} = \hat{\mathbf{x}})$ : + +Theorem 1. If each $P(\mathbf{Y}^j | \mathbf{X})$ is sufficiently scattered, then $\mathbf{T}_{\star}^{j} = \mathbf{T}^{j}$ and $(h_{\theta_{\star}}(\hat{\mathbf{x}})_j, 1 - h_{\theta_{\star}}(\hat{\mathbf{x}})_j)^T = P(\mathbf{Y}^j | \mathbf{X} = \hat{\mathbf{x}})$ ( $j = 1, 2, \dots, C$ ) must hold, where $(\mathbf{T}_{\star}^{1}, \mathbf{T}_{\star}^{2}, \dots, \mathbf{T}_{\star}^{C}, \theta_{\star})$ are the optimal solutions of the problem (11). + +# 2.4. Cycle-Consistency Estimation and Model Regularization + +In multi-label learning, imbalanced label distributions [44] and inter-class correlations [7] can compress class-posterior probability points and limit their scatter. Besides, the volume approximation of $T^j$ in the problem (11) is highly sensitive to the noisy class-posterior, making it suffer from large estimation errors in noisy environments [18]. The failure of the sufficiently scattered assumption, along with the excessive reliance on the noisy class-posterior, compromises the effectiveness of the problem (11). To enhance transition matrix estimation, we develop an adaptive loss framework [4] that incorporates the volume maximization of the clean class-posterior, which leverages additional information to mitigate bias in single-target decision-making. Based on the equation $P(\bar{Y}^j | \mathbf{X}) = T^j P(\mathbf{Y}^j | \mathbf{X})$ , where $P(\bar{Y}^j | \mathbf{X})$ can be derived from the noisy data, minimizing the volume of $T^j$ is equivalent to imposing a volume constraint on $P(\mathbf{Y}^j | \mathbf{X})$ . Through the matrix $\tilde{T}^j$ (an approximate inverse of $T^j$ ), we project the simplex formed in the + +noisy space onto the clean class-posterior space to obtain $P^{\prime}(\mathbf{Y}|\mathbf{X})$ , which will make up the largest volume. Then we design a cycle-consistency optimization strategy to enhance both the volume constraint on $\pmb{T}^{j}$ and maximization of $P(\mathbf{Y}|\mathbf{X})$ , in which the estimated noisy class-posterior matches the provided noisy label $\bar{\mathbf{Y}}^j$ , the backward estimated clean class-posterior $P^{\prime}(\mathbf{Y}|\mathbf{X})$ aligns with the predicted clean-class posterior and the forward-backward consistency term $\tilde{T}^j T^j P(\mathbf{Y}|\mathbf{X})$ coheres with $P(\mathbf{Y}|\mathbf{X})$ . + +Firstly, we employ forward estimation to simultaneously optimize the transition matrix $T^{j}$ and classifier $h_{\theta}$ . By leveraging the cross-entropy loss and reporting the output as $F^{ij} = (h_{\theta}(\hat{\boldsymbol{x}}_i)_j, 1 - h_{\theta}(\hat{\boldsymbol{x}}_i)_j)^T$ , we aim to minimize the approximation error between the inferred noisy class-posterior probability $T^{j}F^{ij}$ and the given noisy label $\bar{Y}^{j}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {1} (\theta , \boldsymbol {T} ^ {1}, \dots , \boldsymbol {T} ^ {C}) = - \frac {1}{C N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {C} (\bar {Y} _ {i, j} \log (\boldsymbol {T} _ {1} ^ {j} \boldsymbol {F} ^ {i j}) \\ + \left(1 - \bar {Y} _ {i, j}\right) \log \left(1 - \mathbf {T} _ {2} ^ {j} \mathbf {F} ^ {i j}\right)). \\ \end{array} +$$ + +Secondly, the Sigmoid function is employed to convert the one-hot label into the noisy class-posterior probability, which is then normalized along the label dimension to capture the soft labels $\bar{\mathbf{Y}}^{sn}$ . Denoting the outcome $(\bar{\mathbf{Y}}_{i,j}^{sn}, 1 - \bar{\mathbf{Y}}_{i,j}^{sn})^T$ as $\bar{\mathbf{F}}^{ij}$ , we utilize $\tilde{T}^j$ to project $\bar{\mathbf{F}}^{ij}$ onto the clean class-posterior space. Thus, we minimize the following cross-entropy loss between the backward estimated clean class-posterior $\tilde{T}^j \bar{\mathbf{F}}^{ij}$ and the true class-posterior probability estimation $h_\theta(\hat{\boldsymbol{x}}_i)_j$ : + +$$ +\begin{array}{l} \mathcal {L} _ {2} (\theta , \tilde {\boldsymbol {T}} ^ {1}, \dots , \tilde {\boldsymbol {T}} ^ {C}) = - \frac {1}{C N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {C} (h _ {\theta} (\hat {\boldsymbol {x}} _ {i}) _ {j} \log (\tilde {\boldsymbol {T}} _ {1} ^ {j} \bar {\boldsymbol {F}} ^ {i j}) \\ + \left(1 - h _ {\theta} \left(\hat {\boldsymbol {x}} _ {i}\right) _ {j}\right) \log \left(1 - \tilde {\boldsymbol {T}} _ {2} ^ {j} \bar {\boldsymbol {F}} ^ {i j}\right)). \\ \end{array} +$$ + +At this stage, the transition matrix can be derived from both the forward and backward ways by optimizing $\mathcal{L}_1$ and $\mathcal{L}_2$ . Thirdly, we should establish the cycle-consistency loss by minimizing the difference between $\tilde{T}^j T^j F^{ij}$ and $h_\theta (\hat{x}_i)_j$ to reinforce the invertible relationship of $T^j$ and $\tilde{T}^j$ : + +$$ +\begin{array}{l} \mathcal {L} _ {3} (\theta) = - \frac {1}{C N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {C} \left(h _ {\theta} \left(\hat {\boldsymbol {x}} _ {i}\right) _ {j} \log \left(\tilde {\boldsymbol {T}} _ {1} ^ {j} \boldsymbol {T} ^ {j} \boldsymbol {F} ^ {i j}\right) \right. \tag {12} \\ + \left(1 - h _ {\theta} \left(\hat {\boldsymbol {x}} _ {i}\right) _ {j}\right) \log \left(1 - \bar {\boldsymbol {T}} _ {2} ^ {j} \boldsymbol {T} ^ {j} \boldsymbol {F} ^ {i j}\right)), \\ \end{array} +$$ + +where we only optimize the network parameters $\theta$ while keeping $T^j$ and $\tilde{T}^j$ constant in order to maintain their consistency and focus on learning an optimal classifier. Therefore, the loss for identifying the transition matrix and learning the clean class-posterior is $\mathcal{L}_{IB} = \mathcal{L}_1 + \mathcal{L}_2 + \mathcal{L}_3$ . + +To mitigate overfitting on noisy data, we design model regularization by leveraging the partial fidelity of noisy la + +bels and label correlations. We allow the classifier to directly access raw labels and obtain the classification loss: + +$$ +\begin{array}{l} \mathcal {L} _ {c} = - \frac {1}{C N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {C} \left(\bar {Y} _ {i, j} \log \left(h _ {\theta} \left(\hat {\boldsymbol {x}} _ {i}\right) _ {j}\right)\right) \tag {13} \\ + \left(1 - \bar {Y} _ {i, j}\right) \log \left(1 - h _ {\theta} \left(\hat {\boldsymbol {x}} _ {i}\right) _ {j})\right). \\ \end{array} +$$ + +The construction of $\mathcal{L}_c$ is motivated by two key considerations. On one hand, it is essential to harness the underlying true information embedded in the noisy data to improve the inefficient classifier. On the other hand, strengthening the model memory with original noise helps minimize the risk of misclassifying clean samples as noisy ones. Since label correlation is a crucial factor for multi-label learning, we minimize the discrepancy in the label relationships reflected by the estimated clean class posterior and the classifier output to reduce classification bias. By defining the clean class-posterior matrix of the positive labels as $P^c$ , with elements given by $P_{i,j}^{c} = \tilde{T}_{1}^{j}\bar{F}^{ij}$ , we can build the following correlation-aligned loss driven by label co-occurrence [43]: + +$$ +\mathcal {L} _ {r} = \frac {1}{C ^ {2}} \left\| \left(\boldsymbol {P} ^ {c}\right) ^ {T} \boldsymbol {P} ^ {c} - \left(h _ {\theta} (\hat {\boldsymbol {X}})\right) ^ {T} h _ {\theta} (\hat {\boldsymbol {X}}) \right\| _ {F} ^ {2}. \tag {14} +$$ + +Thus, our model regularization can be formulated as $\mathcal{L}_R = \mathcal{L}_c + \mathcal{L}_r$ . The total training loss of our model is + +$$ +\mathcal {L} = \mathcal {L} _ {I B} + \lambda_ {1} \mathcal {L} _ {I D} + \lambda_ {2} \mathcal {L} _ {R}, \tag {15} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are penalty coefficients to balance the effects of different losses when handling noisy labels. + +# 3. Experiments + +# 3.1. Experimental Setting + +Datasets and Evaluation Metrics. Following [23, 45], our experiments incorporate six widely recognized multi-view multi-label datasets, i.e., Yeast [11], Corel 5k [5], VOC 2007 [6], Espgame [1], IAPR TC-12 [10], and MIR FLICKR [15]. Similar to [25], we employ six diverse metrics to comprehensively evaluate each method, i.e., Hamming Loss (HL), Ranking Loss (RL), OneError (OE), Coverage (Cov), Average Precision (AP), and Area Under Curve (AUC). To facilitate performance comparison, we present 1-HL, 1-RL, 1-OE, and 1-Cov values in our report, with higher values consistently indicating better performance. + +Comparison Methods. To validate the effectiveness of DMMIvNL, we perform a comparative analysis against eight state-of-the-art methods i.e., iMVWL, TM3L, AlM-Net, DICNet, DIMC, MTD, SIP, ENTMLC. Since the first seven methods are designed to handle incomplete multiview multi-label problems, we train them directly on noisy data without modifications. As ENTMLC specializes in addressing multi-label noise, we tackle missing views by fill + +Table 1. Experimental results of nine methods on all datasets with PER=50% and ρ = 30%. The best result on each row is bolded and the second-best result is underlined. The average ranking on the six metrics is shown as Ave.R. + +
DataMETDICNetDIMCMTDSIPAIMNetiMVWLTM3LNTMLCSDMMIvNL
Yeast1-HL71.54±1.8174.94±1.5676.40±0.3371.24±1.4973.03±1.7966.94±0.4763.89±0.2969.88±0.1677.64±0.47
1-RL73.81±0.9174.82±2.3377.08±0.7571.30±1.6773.95±0.4372.83±1.0565.72±0.3771.40±3.1480.18±0.37
1-OE71.86±3.7470.95±5.3775.33±0.4861.94±6.7470.66±2.7963.91±2.2956.49±0.7773.28±0.4373.69±1.22
1-Cov41.68±2.4844.12±3.1045.53±1.9539.04±2.1641.19±0.3141.50±1.4534.93±0.7335.55±1.3750.77±0.84
AP65.61±1.3466.91±2.0968.87±0.8261.78±2.2565.55±1.1863.74±0.9256.90±0.3066.69±0.6072.08±0.34
AUC75.33±0.7376.34±2.3578.36±0.5473.42±1.5375.46±0.4574.85±0.9767.92±0.2873.33±2.4481.03±0.55
Ave.R4.503.331.837.174.836.338.837.001.17
Corel 5k1-HL96.37±0.6896.08±0.9998.65±0.0098.59±0.0498.66±0.0079.41±0.0597.23±0.6798.57±0.1798.68±0.02
1-RL60.14±0.8461.00±0.8767.26±0.5965.32±1.4163.37±0.6564.53±0.5552.01±0.3965.92±1.2372.27±0.57
1-OE6.30±0.895.44±1.0521.22±1.0612.88±4.2824.00±0.7310.29±2.4620.18±0.332.30±1.6725.20±0.78
1-Cov32.29±0.8133.31±0.9637.53±0.8836.69±2.0341.62±0.4736.09±1.1233.03±0.0235.19±0.9644.35±1.10
AP6.72±0.406.80±0.2915.22±0.6510.54±1.8618.49±0.2110.18±0.914.48±0.098.25±0.9919.18±0.45
AUC60.30±0.8551.74±1.3067.63±0.5965.69±1.5470.37±0.3764.81±0.7658.99±0.1165.97±1.2772.42±0.47
Ave.R7.677.672.834.332.676.007.335.501.00
VOC 20071-HL89.48±1.7089.72±1.3192.85±0.0592.00±0.1792.23±0.3274.95±0.2084.92±0.2092.66±0.0292.67±0.06
1-RL64.72±0.8965.70±0.7666.04±1.1063.69±1.7263.56±0.9267.03±0.8156.30±0.5866.03±1.9572.07±0.84
1-OE26.14±3.4131.28±3.1939.21±0.2131.84±4.2135.22±4.8337.47±1.5326.53±0.7139.35±1.8337.16±1.90
1-Cov58.13±0.7858.76±1.2059.06±0.9356.94±1.5256.55±0.8959.19±0.8153.99±0.3958.02±1.9765.64±1.21
AP34.45±1.3837.52±1.1440.18±0.6536.57±2.0337.16±2.1739.83±0.1827.49±0.3541.41±0.6843.86±1.15
AUC67.92±0.8068.84±1.0068.94±0.8766.52±1.4165.81±0.4969.76±0.6862.29±0.3968.55±2.0874.65±1.01
Ave.R6.835.172.506.506.503.678.673.501.67
Espauge1-HL97.88±0.1297.50±0.5698.40±0.0098.26±0.0198.41±0.0079.07±0.0494.75±4.9697.78±0.1898.25±0.01
1-RL62.26±0.3263.18±0.5063.94±0.3062.14±0.8464.67±0.6662.57±0.9854.05±0.6466.66±0.6069.79±0.47
1-OE14.49±1.2015.75±1.0814.97±0.7013.63±0.9118.75±0.2410.63±0.5417.27±7.211.48±1.1931.77±1.51
1-Cov29.73±0.3130.41±0.4933.06±0.3031.95±0.6634.65±0.6030.14±1.0125.73±0.1733.37±0.4335.80±0.55
AP9.83±0.3510.78±0.3211.28±0.1310.19±0.4713.38±0.299.66±0.555.82±0.096.77±0.7819.13±0.58
AUC62.15±0.3263.35±0.5664.12±0.2362.20±0.9066.28±0.7062.65±0.9156.03±0.0666.64±0.7969.86±0.54
Ave.R6.675.173.675.832.177.177.835.001.50
IAPR TC-121-HL97.54±0.0497.67±0.1898.00±0.0097.93±0.0398.00±0.0079.41±0.0288.22±1.2497.18±0.1498.04±0.02
1-RL61.86±0.3265.48±2.4566.90±0.1465.17±0.9169.20±0.3370.78±0.5755.79±0.9169.68±0.7372.66±0.48
1-OE14.86±0.6418.27±1.6829.51±0.8127.42±5.4433.89±4.6723.45±4.1615.02±9.561.49±1.3934.33±1.37
1-Cov25.98±0.4328.56±1.9529.24±0.3128.12±1.1631.06±0.5034.23±1.0823.23±0.2933.67±0.4635.40±0.53
AP9.18±0.2412.07±1.5715.10±0.1013.42±0.7817.37±0.8914.88±0.636.25±0.077.32±0.4319.66±0.07
AUC62.12±0.2465.52±2.3866.95±0.0965.38±0.6769.48±0.3170.87±0.6057.53±0.1169.75±0.8872.56±0.50
Ave.R7.505.833.835.673.174.008.505.501.00
MIR FLICKR1-HL86.63±0.2585.90±0.8787.46±0.0786.24±0.3187.12±0.1973.17±0.1079.90±0.1386.24±0.9487.61±0.09
1-RL75.11±0.4376.91±1.6176.23±0.2073.11±0.5074.97±0.9575.97±0.8860.07±0.1979.74±1.6882.34±0.75
1-OE44.98±0.7643.88±1.7636.58±0.7335.51±1.1935.78±1.0736.97±2.9642.44±0.3738.67±6.7261.34±0.99
1-Cov46.06±0.5948.98±2.7949.39±0.2346.30±0.4947.18±3.1348.15±2.1435.08±0.0953.12±4.1256.94±0.98
AP42.55±0.4244.11±0.7940.80±0.2038.50±0.1840.71±0.2242.18±0.5629.72±0.3445.69±2.7354.89±0.94
AUC73.96±0.4175.67±1.9175.42±0.1772.27±0.5873.97±1.3874.62±0.7262.80±0.1479.16±1.2881.14±0.93
Ave.R5.173.834.337.506.175.838.003.171.00
+ +ing them with average imputation, followed by feature concatenation. All parameters of compared methods are configured as the recommended values in their original codes. + +Implementation Details. Each dataset is divided into training, validation and test sets in the ratio of 7:1:2. To simulate the partial view setting, a certain proportion of instances defined by the partial example ratio (PER) are randomly selected as unavailable instances on each view. Besides, we ensure each sample contains at least one complete view to avoid invalid cases. In the multi-label noise configuration, we set a uniform noise rate for each label, i.e., $\rho$ of positive and negative labels are flipped to their opposite. All results are obtained from ten independent runs, with final outcomes + +reported as mean values alongside standard deviations. Our model is implemented by PyTorch on one NVIDIA GeForce RTX 4090 GPU of 24GB memory. + +# 3.2. Experimental Results. + +Performance Evaluation. To evaluate the capabilities of DMMIvNL in addressing data deficiencies, we conduct extensive comparative experiments focusing on two critical aspects, i.e., label noise and view missing. Besides, we assign a broad range of noise rates and missing proportions to demonstrate the versatility of our approach. For label noise, $\rho$ is considered at levels of $\{10\%, 15\%, 20\%, 25\%, 30\%, 35\%, 40\}$ , + +![](images/52e333e3da7bc03b518954980d5716f2be29d7c1523910ae13d73ab259d57d40.jpg) +(a) Yeast + +![](images/161217ce1dd5492f730d2e346f8e5773c0fb12d3ea577345c201016e5bdcd951.jpg) +(b) Corel 5k + +![](images/51dd9063fea18a18c94ceac0bdc1183280175dd99f78cef399314d9babdc3788.jpg) +(c) VOC 2007 + +![](images/5087eda2c27a671dfb81526747cb660bf098eba72dc48e63ae0c1c4921bcec58.jpg) +(d) ESP Game + +![](images/b43bd37c2d906ae295d75fe5d75cbea65e4fbd37e423d9bfd6ff05ac29e24369.jpg) +(e) IAPR TC-12 + +![](images/dc5f98b1f92e024bb2699244a9a74ea23de57d560e685b12131311257aee1669.jpg) +(f) MIR FlickR + +![](images/a3bf4e66f985ac688fb16cd191a1ebb13d852341fc3387040ef1eb6d45c73e89.jpg) +Figure 2. AUC comparisons on six datasets with noise rate $\rho$ varying from $10\%$ to $40\%$ while PER=50%. +(a) Espgame +Figure 3. Parameter analysis of the trade-off parameters $\lambda_{1}$ and $\lambda_{2}$ on Espgame and IAPR TC-12. + +![](images/744832ee31c6c5145dbd2301fb4869c6d23ce005eeab8129955e177eb4198031.jpg) +(b) IAPR TC-12 + +while for view missing, PER encompasses values of $\{10\%, 30\%, 50\%, 70\}$ . Tab. 1 displays all metrics with $\rho$ fixed at $30\%$ and PER fixed at $50\%$ , Fig. 2 illustrates the variation in AUC as $\rho$ varies from $10\%$ to $40\%$ , and Fig. 4 shows the metrics distribution when $\mathrm{PER} = 10\%$ and $70\%$ . Additional results are included in the Appendix. + +Regarding the label noise, we can find that 1) Our method outperforms all compared approaches, particularly distinguished in the most representative AP metric. For + +instance, our method achieves a remarkable AP improvement of up to $43\%$ on Espgame in Tab. 1. 2) Compared to NTMLCS for handling label noise, our approach reveals substantial advantages. Besides, in terms of the performance on VOC 2007, DMMIvNL exhibits greater stability than the stepwise estimation way of NTMLCS, which is achieved through the cycle-consistency framework that simultaneously determines the transition matrix and classifier. Additionally, our approach surpasses methods developed for incomplete MvMLC. For example, Fig. 2 illustrates that our DMMIvNL reliably improves performance even against the top-performing MvMLC methods MTD and AIMNet. 3) As $\rho$ increases from $10\%$ to $40\%$ , DMMIvNL consistently outperforms the other eight methods. Moreover, many methods lose reliability at elevated noise levels, such as NTMLCS on Yeast, TM3L on MIR FLICKR and DIMC on Corel 5k, while our approach continues to learn a competent classifier and deliver robust performance. + +As for the view missing, we have the following insights: 1) Our method exhibits outstanding performance in handling missing views. As illustrated in the radar chart of Fig. 4, DMMIvNL consistently occupies the outermost circle, reflecting impressive results across all metrics. 2) + +![](images/39fcff99320936a896395f5de3a52a15c222116229d7ade810422e84abe36cc5.jpg) + +![](images/293accfec9fb56fb79f030aebdbc6a027c20742bebeb9c6a11ca4d0b1d88a6ca.jpg) +(a) Yeast, PER=10% + +![](images/ea37a13ebdea01c15b7321498646c40da7b46ba9611e5f104186ce58bafd1e6c.jpg) +(b) VOC 2007, PER=10% + +![](images/feb5d9a7f31e34a9bba5915ba2e66ee01618a4995f9f8288a5e294272e756e46.jpg) +(c) Espgame, PER=10% + +![](images/ddf951c0cc525b648deb557bd42932c0c55e56a04caa64fbffa1afa59dbf6bbf.jpg) +(d) MIR FLICKR, PER=10% + +![](images/4c651e7fd7c34e877e663208ef30fe9df9ac9e06ce05204bd05fa435f8af88e2.jpg) +(e) Yeast, PER=70% + +![](images/00030c4cb484d5b79e2bb2a13d20468ba8f41817a2735acb5da56978d2837a79.jpg) +(f) VOC 2007, PER=70% + +![](images/65b3b5d287cdb538e84f9d8362c540fbda24e25c5c1247da342ba70dec3c5222.jpg) +(g) Espgame, PER=70% + +![](images/5da81f1db5f85b0a15fc6edd59c01367941ca7b3bacd2756172a459b9080ba56.jpg) +(h) MIR FLICKR,PER=70% +Figure 4. Experimental results of nine methods on four datasets with PER=10% and 70% while $\rho = 30\%$ . + +As PER changes from $10\%$ to $70\%$ , DMMIvNL continues to achieve top performance. Notably, while DIMC experiences a substantial decline at $70\%$ PER, our method remains efficient. It indicates that DMMIvNL delivers strong robustness across varying levels of feature absence. 3) Our method makes remarkable breakthroughs over deep learning based methods for missing views, like SIP, MTD, and DICNet. For example, on MIR FLICKR, DMMIvNL improves 1-OE by nearly $50\%$ compared to SIP at $50\%$ PER. This improvement is primarily due to the focus on extracting the shared semantics and specific features through information separation, an aspect overlooked by other methods. + +Ablation Study and Parameter Sensitivity. The ablation experiments are conducted to analyze the effect of the three crucial aspects of DMMIvNL, i.e., representation extraction based on the information bottleneck theory $(\mathcal{L}_{IB})$ , cycle-consistency estimation for noise identification and classifier learning $(\mathcal{L}_{ID})$ , and model regularization to combat overfitting $(\mathcal{L}_r)$ . Note that loss $\mathcal{L}_c$ is retained since it is essential for classifier improvement during early training. The first three rows of Tab. 2 indicate that $\mathcal{L}_{IB}$ , $\mathcal{L}_{ID}$ and $\mathcal{L}_r$ are integral components of DMMIvNL. Besides, the comparison between the empirical loss $\mathcal{L}_{vol}$ from problem (11) and $\mathcal{L}_{ID}$ confirms that the cycle-consistency framework enhances the stability of transition matrix estimation. To study the impact of $\lambda_1$ and $\lambda_2$ , we present the results under distinct parameter settings in Fig. 3. As shown in the heatmap, the performance of DMMIvNL exhibits variability and improves as $\lambda_1$ gets closer to 0.001 and $\lambda_2$ approaches 1. + +Table 2. Ablation study on Yeast, Corel 5k and VOC 2007 with PER=50% and ρ = 30%. ‘√’ and ‘X’ represent the used and not used corresponding item, respectively. + +
LvolLIBLIDLrEspgameCorel 5kVOC 2007
APAUCAPAUCAPAUC
XX✓✓0.17500.68460.17130.69720.41450.7366
X✓X✓0.18380.71130.18500.70890.43050.7367
X✓✓X0.18340.71310.18000.70960.43590.7148
✓X✓✓0.14990.67320.15430.62840.36120.6715
X✓✓✓0.19340.73070.19970.72840.46220.7566
+ +# 4. Conclusion + +To address the dual challenges of view missing and multi-label noise, we propose DMMIvNL in this paper. Built upon the information bottleneck theory, a mutual information model is established to constrain information exchange across diverse representations, along with a derived theoretical upper bound. Besides, we prove that minimizing the volume of the transition matrix for each category consistently leads to the optimal classifier. Moreover, stability in noise recognition is further improved by introducing a cycle-consistency estimation framework, while raw fidelity information and inherent label correlations are used to reduce overfitting. Finally, extensive experiments demonstrate the superiority of DMMIvNL. In future work, we will leverage the prior knowledge embedded in LLM to rectify inaccurate labels and tackle the issue of class imbalance. + +# Acknowledgment + +This work was partially supported by the National Natural Science Foundation of China [Grant No. 62376281], and the Key NSF of China under Grant No. 62136005. + +# References + +[1] Luis Von Ahn and Laura Dabbish. Labeling images with a computer game. In SIGCHI Conference on Human Factors in Computing Systems, page 319-326, 2004. 5 +[2] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. In ICML, pages 825-836, 2021. 4 +[3] Jia-Yao Chen, Shao-Yuan Li, Sheng-Jun Huang, and Chm. Unm: A universal approach for noisen, songcan and wang, lei and xie, ming-kun. IEEE Transactions on Knowledge and Data Engineering, 36:4968-4980, 2024. 2 +[4] De Cheng, Yixiong Ning, Nannan Wang, Xinbo Gao, Heng Yang, Yuxuan Du, Bo Han, and Tongliang Liu. Class-dependent label-noise learning with cycle-consistency regularization. pages 11104-11116, 2022. 4 +[5] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In ECCV, pages 97-112, 2002. 5 +[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2010. 5 +[7] Yuling Fan, Jinghua Liu, Jianeng Tang, Peizhong Liu, Yaojin Lin, and Yongzhao Du. Learning correlation information for multi-label feature selection. Pattern Recognition, 145: 109899, 2024. 4 +[8] Marco Federici, Anjan Dutta, Patrick Forre, Nate Kushman, and Zeynep Akata. Learning robust representations via multi-view information bottleneck. In ICLR, 2020. 2 +[9] Xiao Fu, Kejun Huang, and Nicholas D Sidiropoulos. On identifiability of nonnegative matrix factorization. IEEE Signal Processing Letters, 25(3):328-332, 2018. 4 +[10] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International Workshop onto Image, pages 1–11, 2006. 5 +[11] Cordelia Guillaumin. Multimodal semi-supervised learning for image classification. In CVPR, pages 902-909, 2010. 5 +[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, I Tsang, and Masashi Sugiyama. Robust training of deep neural networks with extremely noisy labels. In NeurIPS, page 4, 2020. 4 +[13] Pingting Hao, Weiping Ding, Wanfu Gao, and Jialong He. Exploring view-specific label relationships for multi-view multi-label feature selection. Information Sciences, 681: 121215, 2024. 1 +[14] Huaxi Huang, Hui Kang, Sheng Liu, Olivier Salvado, Thierry Rakotoarivelo, Dadong Wang, and Tongliang Liu. + +Paddles: Phase-amplitude spectrum disentangled early stopping for learning with noisy labels. In ICCV, pages 16719-16730, 2023. 2 +[15] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In ICMR, pages 39-43, 2008. 5 +[16] Xiaodong Jia, Xiao-Yuan Jing, Xiaoke Zhu, Songcan Chen, Bo Du, Ziyun Cai, Zhenyu He, and Dong Yue. Semi-supervised multi-view deep discriminant representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2496-2509, 2020. 1, 4 +[17] Zhangqi Jiang, Tingjin Luo, and Xinyan Liang. Deep incomplete multi-view learning network with insufficient label information. In AAAI, pages 12919-12927, 2024. 2 +[18] Rihuan Ke. Deep variation prior: Joint image denoising and noise variance estimation without clean data. IEEE Transactions on Image Processing, 33:2908-2923, 2024. 4 +[19] Shikun Li, Xiaobo Xia, Hansong Zhang, Yibing Zhan, Shiming Ge, and Tongliang Liu. Estimating noise transition matrix with label correlations for noisy multi-label learning. In NeurIPS, pages 24184-24198, 2022. 2 +[20] Xuefeng Li, Tongliang Liu, Bo Han, Gang Niu, and Masashi Sugiyama. Provably end-to-end label-noise learning without anchor points. In ICML, pages 6403-6413, 2021. 4 +[21] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4447-4461, 2022. 3 +[22] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In AAAI, pages 8807-8815, 2023. 1 +[23] Chengliang Liu, Jie Wen, Xiaoling Luo, and Yong Xu. Incomplete multi-view multi-label learning via label-guided masked view-and category-aware transformers. In AAAI, pages 8816-8824, 2023. 1, 5 +[24] Chengliang Liu, Jinlong Jia, Jie Wen, Yabo Liu, Xiaoling Luo, Chao Huang, and Yong Xu. Attention-induced embedding imputation for incomplete multi-view partial multi-label classification. In AAAI, pages 13864-13872, 2024. 1 +[25] Chengliang Liu, Jie Wen, Yabo Liu, Chao Huang, Zhihao Wu, Xiaoling Luo, and Yong Xu. Masked two-channel decoupling framework for incomplete multi-view weak multi-label learning. In NeurIPS, 2024. 1, 4, 5 +[26] Chengliang Liu, Gehui Xu, Jie Wen, Yabo Liu, Chao Huang, and Yong Xu. Partial multi-view multi-label classification via semantic invariance learning and prototype modeling. In ICML, 2024. 1, 2 +[27] Chengliang Liu, Jie Wen, Yong Xu, Bob Zhang, Liqiang Nie, and Min Zhang. Reliable representation learning for incomplete multi-view missing multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2025. 1 +[28] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3):447-461, 2015. 4 + +[29] Tingjin Luo, Yang Yang, Dongyun Yi, and Jieping Ye. Robust discriminative feature learning with calibrated data reconstruction and sparse low-rank model. Applied Intelligence, 54(3):2867-2880, 2024. 1 +[30] Yong Luo, Dacheng Tao, Chang Xu, Chao Xu, Hong Liu, and Yonggang Wen. Multiview vector-valued manifold regularization for multilabel image classification. IEEE Transactions on Neural Networks and Learning Systems, 24(5): 709-722, 2013. 1 +[31] Yubin Park and Joyce C Ho. Tackling overfitting in boosting for noisy healthcare data. IEEE Transactions on Knowledge and Data Engineering, 33(7):2995-3006, 2019. 1 +[32] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, pages 1944-1952, 2017. 4 +[33] Yalan Qin, Hanzhou Wu, Xinpeng Zhang, and Guorui Feng. Semi-supervised structured subspace learning for multi-view clustering. IEEE Transactions on Image Processing, 31:1-14, 2021. 1 +[34] Alisa Smirnova, Jie Yang, Dingqi Yang, and Philippe CudreMauroux. Nessy: A neuro-symbolic system for label noise reduction. IEEE Transactions on Knowledge and Data Engineering, 35(8):8300-8311, 2022. 1 +[35] Shiding Sun, Bo Wang, and Yingjie Tian. Decoupled representation for multi-view learning. Pattern Recognition, 151: 110377, 2024. 2 +[36] Yu-Yin Sun, Yin Zhang, and Zhi-Hua Zhou. Multi-label learning with weak label. In AAAI, pages 593-598, 2010. 1 +[37] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In IJCAI, pages 2703-2709, 2018. 1 +[38] Ziyi Wang and Yitian Xu. A two-stage multi-view partial multi-label learning for enhanced disambiguation. Knowledge-Based Systems, 293:111680, 2024. 1 +[39] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems, 35(8):11396-11408, 2023. 1 +[40] Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning with noisy label identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3676-3687, 2021. 1 +[41] Wulin Xie, Xiaohuan Lu, Yadong Liu, Jiang Long, Bob Zhang, Shuping Zhao, and Jie Wen. Uncertainty-aware pseudo-labeling and dual graph driven network for incomplete multi-view multi-label classification. In ACMMM, pages 6656-6665, 2024. 1 +[42] Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. Instance-dependent label-noise learning under a structural causal model. In NeurIPS, 2021. 2 +[43] Yitong Yao, Jing Zhang, Peng Zhang, and Yueheng Sun. A dual-branch learning model with gradient-balanced loss for long-tailed multi-label text classification. ACM Transactions on Information Systems, 42(2):1-24, 2023. 5 + +[44] Selim F Yilmaz, E Batuhan Kaynak, Aykut Koç, Hamdi Dibeklioglu, and Suleyman Serdar Kozat. Multi-label sentiment analysis on 100 languages with dynamic weighting for label imbalance. IEEE Transactions on Neural Networks and Learning Systems, 34:331-343, 2021. 4 +[45] Dawei Zhao, Qingwei Gao, Yixiang Lu, and Dong Sun. Two-step multi-view and multi-label learning with missing label via subspace learning. Applied Soft Computing, 102:107120, 2021. 5 +[46] Pengfei Zhu, Qi Hu, Qinghua Hu, Changqing Zhang, and Zhizhao Feng. Multi-view label embedding. Pattern Recognition, 84:126-135, 2018. 1 +[47] Wenzhang Zhuge, Tingjin Luo, Ruidong Fan, Hong Tao, Chenping Hou, and Dongyun Yi. Absent multiview semisupervised classification. IEEE Transactions on Cybernetics, 54(3):1708-1721, 2024. 1 \ No newline at end of file diff --git a/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/images.zip b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..38efe047a49e4d179ca6cf76cd557ff3f60d76f5 --- /dev/null +++ b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d508806d7eeda31974a9884b25f16de2a65ef8505d5290e53376009af39be8d0 +size 1100414 diff --git a/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/layout.json b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..58f08fd7e48fc3a5e6ffe423eb08ff95cee2338a --- /dev/null +++ b/2025/Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels/layout.json @@ -0,0 +1,10350 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "type": "text", + "content": "Theory-Inspired Deep Multi-View Multi-Label Learning with Incomplete Views and Noisy Labels" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 161, + 428, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 161, + 428, + 190 + ], + "spans": [ + { + "bbox": [ + 182, + 161, + 428, + 190 + ], + "type": "text", + "content": "Quanjiang Li, Tingjin Luo*, Jiahui Liao National University of Defense Technology, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 191, + 497, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 191, + 497, + 203 + ], + "spans": [ + { + "bbox": [ + 110, + 191, + 497, + 203 + ], + "type": "text", + "content": "liquanjiang@nudt.edu.cn,tingjinluo@hotmail.com,liaojiahui19@nudt.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 256, + 297, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 297, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 297, + 533 + ], + "type": "text", + "content": "Incomplete features and label noise in multi-view multi-label data significantly undermine the reliability and performance, motivating researchers to explore the mechanism of representation and information recovery. However, learning for such dual deficiencies is crucial but rarely studied. In this paper, we propose a theory-inspired Deep Multi-View Multi-Label Learning method with Incomplete Views and Noisy Labels named DMMIvNL to address these problems. Specifically, to promote the synthesis of task-relevant shared information and preserve the distinctiveness of individual features from limited views, we have developed a feature extraction modular based on the information bottleneck theory, and formulated its theoretical upper bound into its objective. Meanwhile, we theoretically prove that minimizing the volume of the transition matrix ensures the statistical consistency with classifier training. Besides, a cycle-consistency estimation principle is proposed in the volume minimization network to improve the recognition stability of multi-label noise. Moreover, leveraging inherent real semantics information and label correlations are employed as model regularization to reduce the risk of excessive noise fitting. Finally, extensive experimental results validate the effectiveness and robustness of our DMMIvNL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 555, + 135, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 135, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 574, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 696 + ], + "type": "text", + "content": "Advances in feature extraction has shifted data analysis towards multi-view frameworks [27], while growing data attributes drive multi-label learning to offer rich label semantics [13]. Therefore, Multi-view multi-label classification (MvMLC) has emerged to address the intricate demands of processing diverse features and numerous targets. Traditional methods [30, 46] assume the given data has complete views and clean labels, which is violated in practice [41]. On one hand, heterogeneous data from diverse sources frequently encounters missing views due to acquire" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 232, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 496 + ], + "type": "text", + "content": "sition complexities and storage limitations [29, 47]. On the other hand, real-world collections often contain label noise as manual annotation can introduce incorrect labels beyond human knowledge [34]. Therefore, insufficient features and noisy labels severely reduce the classification performance. Regarding MvMLC, many solutions have been proposed to address missing views. iMvWL [37] used weighted matrix factorization to learn a shared subspace and AIMNet [24] leveraged a cross-view attention mechanism to recover missing views. Moreover, two types of label deficiencies have been considered in MvMLC. The problem of Partial Multi-Label (PML) [40] involves both relevant and irrelevant labels, while the problem of Multi-Label Learning with Missing Labels (MLML)[36] occurs when some relevant labels are missing. TFMDD [38] addressed PML with a two-stage manifold disambiguation model and LMVCAT [23] tackled MLML by employing label-guided graphs and category-aware transformers. While existing methods address each annotation problem separately, both issues tend to occur simultaneously in practice. Thus, the challenge of multi-label noise demands substantial attention, where each category suffers from both over-labeled and missing labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "Existing research has not yet tackled the combined issues of missing views and multi-label noise, which are prevalent in real-world applications like incomplete and subjective medical diagnostics [31]. In fact, overcoming the simultaneous challenges requires refined representation extraction under missing views and effective classifier leaning with multi-label noise. The core of multi-view learning lies in extracting shared information across different views while preserving unique features of each view [33]. DICNet [22], DIMC [39] and SIP [26] concentrated solely on shared subspace and neglected distinctive characteristics of individual views. MDDRL [16] and MTD [25] attempted to extract both shared and specific representations with minimal interactions. However, their reliance on geometric orthogonality or distance metrics restricted them to capture only linear patterns, overlooking the complex nature of inter-view relations. Therefore, constraining feature semantics at the intrinsic information level is crucial for obtaining high-level" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20706" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "representations with precise distinction. Most approaches for handling multi-label noise utilize strategies like early stopping [14] and cyclical learning rates [3]. These methods often lack stability and require dataset-specific adjustments, making them insufficiently adaptive. ENTMLC [19] employed label correlations to identify the noise transition matrix for recovering clean labels. However, the decoupling of transition matrix acquisition and classifier training causes error accumulation, limiting the practical utility. To circumvent the uncertainty inherent in those strategies and the inefficiency of stepwise noise learning, a statistically consistent estimation framework should be developed, which would allow for direct training the optimal classifier from noisy labels and achieving strong robustness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 241, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 296, + 491 + ], + "type": "text", + "content": "To tackle these problems, we propose a theory-inspired method with deep networks named DMMIvNL. Specifically, a mutual information model is developed to enhance synergy of shared features, suppress interactions from specific features, and impose orthogonality between shared and specific features of each view. The upper bound of the model is derived through the variational autoencoder framework, with instance-level contrastive operations applied to simulate diverse joint and marginal distributions. Moreover, the ideal classifier is trained on noisy data by minimizing the volume of the noise transition matrix and the process is proven to converge to a unique and globally optimal solution under the sufficiently scattered assumption. The difficulty in guaranteeing the assumption and relying exclusively on noisy class-posterior estimation promotes the design of a cycle-consistency estimation method, which integrates the volume maximization of clean class-posterior. Finally, the real semantics embedded in noisy data and hidden label correlation information are employed as model regularization to reduce the risk of noise overfitting. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 493, + 295, + 623 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 493, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 528 + ], + "type": "text", + "content": "- This is the first multi-view multi-label learning framework capable of handling incomplete views and noisy labels, including both over-labeled and missing labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 529, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 564 + ], + "type": "text", + "content": "- Our mutual information based model extracts semantically distinct representations from limited views, with its upper bound established through variational derivation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 565, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 623 + ], + "type": "text", + "content": "- We theoretically prove the statistical consistency and joint convergence between minimizing the noise transition matrix volume for each category and classifier training. Besides, a cycle-consistency loss framework is designed to enhance transition matrix estimation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 638, + 138, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 138, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 138, + 652 + ], + "type": "text", + "content": "2. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 658, + 167, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 167, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 167, + 671 + ], + "type": "text", + "content": "2.1. Problem Definition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "The multi-view dataset " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X},\\mathbf{Y}\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " samples, where " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " views with dimension " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i = \\{\\pmb{x}_i^{(v)}\\in \\mathbb{R}^{d_v}\\}_{v = 1}^V" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ". The label matrix " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}\\in \\{0,1\\}^{N\\times C}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " corre" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": "sponds to " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": " categories, with " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "Y_{i,j} = 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": " if sample " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": " belongs to class " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "Y_{i,j} = 0" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": " otherwise. " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{N \\times V}" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": " is an indicator matrix, where " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "W_{i,j} = 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": "-th view is available for sample " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "W_{i,j} = 0" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": " means missing. Since precise annotation is hard to obtain, the training dataset in practice is " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}, \\bar{\\mathbf{Y}}\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Y}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": " denotes noisy labels. Thus, we denote the clean and noisy labels of the " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": "-th class as " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^j" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Y}}^j" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": ", related by a noise transition matrix " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^j \\in [0,1]^{2 \\times 2}" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "j \\in [C]" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": ". When only noisy samples are available and without further constraints, instance-dependent transition matrix becomes non-identifiable [42]. Therefore, we employ a general assumption that the noise transition matrix is class-dependent and instance-independent, i.e., " + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "inline_equation", + "content": "P(\\bar{\\mathbf{Y}}^j = k \\mid \\mathbf{Y}^j = i, \\mathbf{X} = x) = P(\\bar{\\mathbf{Y}}^j = k \\mid \\mathbf{Y}^j = i) = T_{i,k}^j" + }, + { + "bbox": [ + 313, + 72, + 555, + 242 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 555, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 262 + ], + "type": "text", + "content": "2.2. Feature Extraction via Information Bottleneck" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "content": "Deriving effective representations from raw data aids in reducing noise and redundancy [17]. Consequently, both task-relevant shared semantics and view-invariant characteristic features should be explored to enhance information processing efficiency. Let " + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "inline_equation", + "content": "z^v" + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "inline_equation", + "content": "s^v" + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "content": "-dimensional shared and specific representation extracted from each view and the goal is to ensure consistency of the shared features across multiple views and preserve the unique contributions of each view. Considering different interaction patterns between representations from " + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "inline_equation", + "content": "x^{v^*}" + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "inline_equation", + "content": "x^v" + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "content": ", mutual information " + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "inline_equation", + "content": "I(z^{v^*}|x^{v^*};z^v|x^v)" + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "content": " should be maximized to capture only universally shared information, while " + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "inline_equation", + "content": "I(s^{v^*}|x^{v^*};s^v|x^v)" + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "content": " should be minimized to preserve only individual attributes. Besides, the shared and specific features should be as orthogonal as possible to ensure absolute distinction in information content [35], i.e., " + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "inline_equation", + "content": "I(z^v |x^v;s^v |x^v) = 0" + }, + { + "bbox": [ + 313, + 269, + 556, + 484 + ], + "type": "text", + "content": ". Thus, the constrained optimization problem based on the information bottleneck theory [8] can be formulated as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 495, + 555, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 495, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 315, + 495, + 555, + 529 + ], + "type": "interline_equation", + "content": "\\max \\frac {1}{V} \\sum_ {v = 1} ^ {V} \\sum_ {v ^ {*} \\neq v} ^ {V} \\left(I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) - I \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right)\\right)", + "image_path": "ffd5692b509064fb29347af0b3be5a21f8b6e699e1c132e1f45cafcd38fcdee6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 532, + 553, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 532, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 315, + 532, + 553, + 574 + ], + "type": "interline_equation", + "content": "s. t. \\quad \\frac {1}{V} \\sum_ {v = 1} ^ {V} I \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) = 0. \\tag {1}", + "image_path": "fca79d72032b2957dc546798891c4cbffa8339b1c648d58813061ac9c1a606ef.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 575, + 553, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 553, + 599 + ], + "type": "text", + "content": "By applying the Lagrange multiplier method, the problem (1) can be equally converted as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 609, + 555, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 609, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 609, + 555, + 689 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max \\frac {1}{V} \\sum_ {v = 1} ^ {V} \\sum_ {v ^ {*} \\neq v} ^ {V} \\left(I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) - I \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right)\\right) \\\\ - \\frac {\\beta}{V} \\sum_ {v = 1} ^ {V} I \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "3d1d7266de5e230b105cba4a553fb4fdc9c4777afd06b934fdb0d39a747e5b8c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\beta \\geq 0" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " is the Lagrange multiplier to balance the efficiency and compactness of information [26]. For the first" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20707" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 72, + 502, + 305 + ], + "blocks": [ + { + "bbox": [ + 109, + 72, + 502, + 305 + ], + "lines": [ + { + "bbox": [ + 109, + 72, + 502, + 305 + ], + "spans": [ + { + "bbox": [ + 109, + 72, + 502, + 305 + ], + "type": "image", + "image_path": "4e7cf2f6120729b621481246e5f816945ca06cbe7ca32b7ec89ba18c2e7120f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 315, + 492, + 328 + ], + "lines": [ + { + "bbox": [ + 117, + 315, + 492, + 328 + ], + "spans": [ + { + "bbox": [ + 117, + 315, + 492, + 328 + ], + "type": "text", + "content": "Figure 1. The main framework of our proposed DMMIvNL. Different shapes signify different samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 348, + 249, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 249, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 249, + 360 + ], + "type": "text", + "content": "term, we can obtain the following lower bound:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 308, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 308, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 308, + 437 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\\\ = \\iint p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) \\log \\frac {p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v} , \\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right)}{p \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right)} d \\boldsymbol {z} ^ {v ^ {*}} d \\boldsymbol {z} ^ {v} \\\\ \\geq \\int \\int p (\\boldsymbol {z} ^ {v} | \\boldsymbol {z} ^ {v ^ {*}}) p (\\boldsymbol {z} ^ {v ^ {*}} | \\boldsymbol {x} ^ {v ^ {*}}) \\log p (\\boldsymbol {z} ^ {v} | \\boldsymbol {z} ^ {v ^ {*}}) d \\boldsymbol {z} ^ {v ^ {*}} d \\boldsymbol {z} ^ {v}, \\\\ \\end{array}", + "image_path": "dcccc4d041afac42472cd99de44c6a63916bfa36b6b9dbd4ba78835ce3a9d10e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "content": " denotes the probability density function. Since computing " + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "inline_equation", + "content": "p(z^v |z^{v^*})" + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "content": " is intractable, we approximate it using a stochastic variational distribution " + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "inline_equation", + "content": "q(z^{v}|z^{v^{*}})" + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "content": ". Then we can obtain a new variational lower bound for maximization:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 495, + 295, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 495, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 495, + 295, + 566 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I \\left(\\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\\\ \\geq \\int \\int p (\\boldsymbol {z} ^ {v} | \\boldsymbol {x} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}} | \\boldsymbol {x} ^ {v ^ {*}}) \\log q (\\boldsymbol {z} ^ {v} | \\boldsymbol {z} ^ {v ^ {*}}) d \\boldsymbol {z} ^ {v ^ {*}} d \\boldsymbol {z} ^ {v} \\\\ = D _ {K L} ^ {\\dagger} \\left(p \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) \\| q \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {z} ^ {v ^ {*}}\\right)\\right) = D _ {K L} ^ {\\dagger} \\left(\\boldsymbol {z} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}}\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "b6afd9982c65be3efa227bd67ba591cbe0d14d232b0c5dcaf577b8a8525411da.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 566, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 296, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 566, + 296, + 616 + ], + "type": "inline_equation", + "content": "D_{KL}(\\cdot ||\\cdot)" + }, + { + "bbox": [ + 55, + 566, + 296, + 616 + ], + "type": "text", + "content": " means the Kullback-Leibler divergence and " + }, + { + "bbox": [ + 55, + 566, + 296, + 616 + ], + "type": "inline_equation", + "content": "D_{KL}^{\\dagger}(\\cdot ||\\cdot)" + }, + { + "bbox": [ + 55, + 566, + 296, + 616 + ], + "type": "text", + "content": " is a pseudo-symbol for notational consistency. For the middle term, the upper bound can be obtained by introducing another variational distribution " + }, + { + "bbox": [ + 55, + 566, + 296, + 616 + ], + "type": "inline_equation", + "content": "q(s^{v}|s^{v^{*}})" + }, + { + "bbox": [ + 55, + 566, + 296, + 616 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 622, + 295, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 622, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 64, + 622, + 295, + 674 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\\\ \\leq D _ {K L} \\left(p \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}, \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\| p \\left(\\boldsymbol {s} ^ {v ^ {*}} \\mid \\boldsymbol {x} ^ {v ^ {*}}\\right) q \\left(\\boldsymbol {s} ^ {v} \\mid \\boldsymbol {s} ^ {v ^ {*}}\\right)\\right) \\tag {4} \\\\ = D _ {K L} ^ {\\left(\\mathbf {s} ^ {v}, \\mathbf {s} ^ {v ^ {*}}\\right)}. \\\\ \\end{array}", + "image_path": "386809c2f90acb48bd98f936e45e2cb4c7894c3d050ef09d57e358977bb30b6b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 680, + 240, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 680, + 240, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 680, + 240, + 693 + ], + "type": "text", + "content": "Similarly, the upper bound for the last term is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 698, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 698, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 115, + 698, + 295, + 715 + ], + "type": "interline_equation", + "content": "I \\left(\\boldsymbol {z} ^ {v} \\mid \\boldsymbol {x} ^ {v}; \\boldsymbol {s} ^ {v} \\mid \\boldsymbol {x} ^ {v}\\right) \\leq D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {z} ^ {v}\\right)}. \\tag {5}", + "image_path": "31b17b0f45e181896826b940e3a30fa8e4e0d420d26404c6895e57e209110588.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 348, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 372 + ], + "type": "text", + "content": "Combining Eqs. (3), (4) and (5), our optimization objective becomes minimizing the upper bound of the problem (2):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 375, + 574, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 375, + 574, + 419 + ], + "spans": [ + { + "bbox": [ + 314, + 375, + 574, + 419 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I B} = \\frac {1}{V} \\sum_ {v = 1} ^ {V} \\sum_ {v ^ {*} \\neq v} ^ {V} \\left(- D _ {K L} ^ {\\dagger \\left(\\boldsymbol {z} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}}\\right)} + D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {s} ^ {v ^ {*}}\\right)}\\right) + \\frac {\\beta}{V} \\sum_ {v = 1} ^ {V} D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {z} ^ {v}\\right)}. \\tag {6}", + "image_path": "e4b9a66a91b0cf30019948d801937b8550b682d6dc615117eb888b1133824328.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 420, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 555, + 468 + ], + "type": "text", + "content": "Contrastive learning [21] is used to compute various distributions. Specifically, representations are treated as probability vectors over " + }, + { + "bbox": [ + 313, + 420, + 555, + 468 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 420, + 555, + 468 + ], + "type": "text", + "content": " classes via a Softmax activation function, and the joint distribution matrix is obtained by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 472, + 558, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 472, + 558, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 472, + 558, + 506 + ], + "type": "interline_equation", + "content": "\\boldsymbol {P} ^ {(v, v ^ {*})} = \\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {z} _ {i} ^ {(v)}\\right) ^ {T} \\boldsymbol {z} _ {i} ^ {(v ^ {*})}, \\boldsymbol {Q} ^ {(v, v ^ {*})} = \\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {s} _ {i} ^ {(v)}\\right) ^ {T} \\boldsymbol {s} _ {i} ^ {(v ^ {*})}.", + "image_path": "d09ea64891034b57f1241e5cdfa699cf04517536e7d46ada5573a7e8930871ed.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": "Due to the strong interaction between " + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "inline_equation", + "content": "z^v" + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "inline_equation", + "content": "z^{v^*}" + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": ", the variational distribution " + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "inline_equation", + "content": "q(z^v |z^{v^*})" + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": " can be estimated by the obtained conditional distribution. Given the extremely weak correlation between " + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "inline_equation", + "content": "s^v" + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "inline_equation", + "content": "s^{v^*}" + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "inline_equation", + "content": "z^v" + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "inline_equation", + "content": "s^v" + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": ", the distribution " + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "inline_equation", + "content": "q(s^v /s^{v^*})" + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "inline_equation", + "content": "q(z^{v} / s^{v})" + }, + { + "bbox": [ + 313, + 510, + 555, + 594 + ], + "type": "text", + "content": " can be approximated by the corresponding marginal distribution. Subsequently, each term in Eq. (6) can be expressed as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 597, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 597, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 324, + 597, + 555, + 714 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} D _ {K L} ^ {\\dagger \\left(\\boldsymbol {z} ^ {v}, \\boldsymbol {z} ^ {v ^ {*}}\\right)} = \\sum_ {d = 1} ^ {D} \\sum_ {d ^ {\\prime} = 1} ^ {D} \\boldsymbol {P} _ {d, d ^ {\\prime}} ^ {(v, v ^ {*})} \\log \\frac {\\boldsymbol {P} _ {d , d ^ {\\prime}} ^ {(v , v ^ {*})}}{\\boldsymbol {P} _ {d} ^ {(v)}} \\\\ D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {s} ^ {v ^ {*}}\\right)} = \\sum_ {d = 1} ^ {D} \\sum_ {d ^ {\\prime} = 1} ^ {D} \\boldsymbol {Q} _ {d, d ^ {\\prime}} ^ {(v, v ^ {*})} \\log \\frac {\\boldsymbol {Q} _ {d , d ^ {\\prime}} ^ {(v , v ^ {*})}}{\\left(\\boldsymbol {Q} _ {d} ^ {(v)}\\right) ^ {\\alpha} \\boldsymbol {Q} _ {d ^ {\\prime}} ^ {(v ^ {*})}} \\\\ D _ {K L} ^ {\\left(\\boldsymbol {s} ^ {v}, \\boldsymbol {z} ^ {v}\\right)} = \\sum_ {d = 1} ^ {D} \\sum_ {d ^ {\\prime} = 1} ^ {D} \\boldsymbol {M} _ {d, d ^ {\\prime}} ^ {\\left(z ^ {v}, s ^ {v}\\right)} \\log \\frac {\\boldsymbol {M} _ {d , d ^ {\\prime}} ^ {\\left(z ^ {v} , s ^ {v}\\right)}}{\\left(\\boldsymbol {M} _ {d} ^ {\\left(z ^ {v}\\right)}\\right) ^ {\\alpha} \\boldsymbol {M} _ {d ^ {\\prime}} ^ {\\left(s ^ {v}\\right)}}, \\end{array} \\right. \\tag {7}", + "image_path": "7ae28e1b8ef21a7956320d864c639a8f4192e03ea6d68c6bdbcc82023b106be4.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20708" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": "where the marginal distributions " + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "inline_equation", + "content": "P_{d}^{(v)}" + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "inline_equation", + "content": "P_{d^{\\prime}}^{(v^{*})}" + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": " can be obtained by summing over the " + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": "-th row and the " + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "inline_equation", + "content": "d^{\\prime}" + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": "-th column of the joint distribution matrix " + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "inline_equation", + "content": "P^{(v,v^{*})}" + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": ", with other symbols defined as the same. In this conversion, we let " + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "inline_equation", + "content": "q(\\boldsymbol{s}^v /\\boldsymbol{s}^{v^*}) = (Q^{(v)})^\\alpha" + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "inline_equation", + "content": "q(z^{v} / s^{v}) = (M^{(z^{v})})^{\\alpha}" + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": " is a balance factor to preserve crucial information and ensure model stability [16]. By introducing the mask matrix " + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\pmb{W}" + }, + { + "bbox": [ + 55, + 70, + 296, + 193 + ], + "type": "text", + "content": " to mitigate the impact of unavailable instances, we can perform cross-view fusion to acquire the information-enriched shared and specific representation:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 200, + 295, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 200, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 80, + 200, + 295, + 233 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {Z}} _ {i,:} = \\sum_ {v = 1} ^ {V} \\frac {\\boldsymbol {z} _ {i , :} ^ {(v)} \\boldsymbol {W} _ {i , v}}{\\sum_ {v} \\boldsymbol {W} _ {i , v}}, \\quad \\bar {\\boldsymbol {S}} _ {i,:} = \\sum_ {v = 1} ^ {V} \\frac {\\boldsymbol {s} _ {i , :} ^ {(v)} \\boldsymbol {W} _ {i , v}}{\\sum_ {v} \\boldsymbol {W} _ {i , v}}. \\tag {8}", + "image_path": "6f8f8294085807d7848577c6161a4fb93507babe02dfc943b69192ba8a72ce1c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 277 + ], + "type": "text", + "content": "Departing from traditional addition or concatenation, we conduct feature interaction by enhancing shared information with proprietary data [25]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 284, + 295, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 284, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 127, + 284, + 295, + 300 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {X}} _ {i, j} = \\theta_ {s} \\left(\\bar {\\boldsymbol {Z}} _ {i, j}\\right) \\cdot \\bar {\\boldsymbol {S}} _ {i, j}, \\tag {9}", + "image_path": "384e01e3e67adfa313c63871260da178108739e54dd240a0801c1e4689fc4de5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 295, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 295, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 308, + 295, + 333 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 55, + 308, + 295, + 333 + ], + "type": "text", + "content": " is the sigmoid activation function and " + }, + { + "bbox": [ + 55, + 308, + 295, + 333 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{X}}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 55, + 308, + 295, + 333 + ], + "type": "text", + "content": " is the final fusion representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 340, + 296, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 364 + ], + "type": "text", + "content": "2.3. Identify Multi-Label Noise by Volume Minimization Network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 369, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 561 + ], + "type": "text", + "content": "The success of deep learning relies on large-scale accurately labeled data, with performance significantly dropping under noisy data [12], making noise identification crucial for network training [2]. In multi-class tasks, many statistically consistent methods based on the transition matrix have been developed [28, 32], which guarantees the classifier learned from noisy data can converge to the optimal classifier under the clean risk. The anchor-point assumption [28] is commonly utilized to determine the transition matrix, which postulates the existence of samples that belong to each class with probability one. However, implementing this approach in multi-label learning is challenging due to the difficulty of pinpointing suitable anchor samples for each label. Motivated by [20], we capitalize on the geometric properties of the transition matrix to seek a breakthrough. The transition matrix for the " + }, + { + "bbox": [ + 55, + 369, + 296, + 561 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 369, + 296, + 561 + ], + "type": "text", + "content": "-th label is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 567, + 295, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 567, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 59, + 567, + 295, + 604 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} ^ {j} = \\left( \\begin{array}{l l} P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 0 \\mid \\boldsymbol {Y} ^ {j} = 0\\right) & P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 0 \\mid \\boldsymbol {Y} ^ {j} = 1\\right) \\\\ P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 1 \\mid \\boldsymbol {Y} ^ {j} = 0\\right) & P \\left(\\bar {\\boldsymbol {Y}} ^ {j} = 1 \\mid \\boldsymbol {Y} ^ {j} = 1\\right) \\end{array} \\right). \\tag {10}", + "image_path": "6266000cb4d82b206e4788ee3103f67606e77d8cd36ab26af20383ba0063c456.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Given an instance " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", the noisy class-posterior for the " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "-th label is " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "P(\\bar{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}}) = [P(\\bar{Y}^j = 0|\\mathbf{X} = \\hat{\\pmb{x}}), P(\\bar{Y}^j = 1|\\mathbf{X} = \\hat{\\pmb{x}})]^T" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " and the clean class-posterior is " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}}) = [P(\\mathbf{Y}^j = 0|\\mathbf{X} = \\hat{\\pmb{x}}), P(\\mathbf{Y}^j = 1|\\mathbf{X} = \\hat{\\pmb{x}})]^T" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". Thus, the relationship " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "P(\\bar{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}}) = \\pmb{T}^j P(\\mathbf{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}})" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " holds. Since " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\sum_{i=0}^{1} P(\\mathbf{Y} = i|\\mathbf{X} = \\hat{\\pmb{x}}) = 1" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "P(\\bar{Y}^j|\\mathbf{X} = \\hat{\\pmb{x}})" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " is then a convex combination of the columns of " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{T}^j" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", which implies that the noisy class-posterior is enclosed within the simplex " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}\\{\\pmb{T}^j\\}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". To reduce noise identification errors, it" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": "is essential to recover " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\operatorname{Sim}\\{\\pmb{T}^j\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": " such that it tightly encompasses all sample posterior probability points, thereby making volume minimization of " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\operatorname{Sim}\\{\\pmb{T}^j\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": " the primary goal. For synchronized learning, we define a trainable " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\hat{T}^j \\in [0,1]^{2 \\times 2}" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{2} \\hat{T}^j{}_{i,k} = 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": ", and construct a multi-label classifier " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": ". The optimization of " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\hat{T}^j" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "h_\\theta(\\hat{x})" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": " involves two objectives: the discrepancy between " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\hat{T}^j(h_\\theta(\\hat{x})_j, 1 - h_\\theta(\\hat{x})_j)^T" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "P(\\bar{Y}^j | X = \\hat{x})" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": " and the volume of each simplex spanned by the columns of " + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\hat{T}^j" + }, + { + "bbox": [ + 313, + 72, + 555, + 212 + ], + "type": "text", + "content": ". Therefore, the following model to jointly learn the transition matrix and classifier is proposed:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 334, + 217, + 555, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 217, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 334, + 217, + 555, + 281 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\min _ {\\hat {\\boldsymbol {T}} ^ {j}} \\sum_ {j = 1} ^ {C} \\operatorname {v o l} \\left(\\hat {\\boldsymbol {T}} ^ {j}\\right) \\\\ s. t. \\hat {\\boldsymbol {T}} ^ {j} \\left[ \\begin{array}{c} h _ {\\theta} (\\hat {\\boldsymbol {x}}) _ {j} \\\\ 1 - h _ {\\theta} (\\hat {\\boldsymbol {x}}) _ {j} \\end{array} \\right] = P \\left(\\bar {\\boldsymbol {Y}} ^ {j} \\mid X = \\hat {\\boldsymbol {x}}\\right), \\end{array} \\right. \\tag {11}", + "image_path": "01a9127692242195bdcdbc5b60003b481940a5f87f5c3e395e489ac99cd59046.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{vol}(\\hat{T}^j)" + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "text", + "content": " denotes a measure to the volume of " + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\operatorname{Sim}\\{\\hat{T}^j\\}" + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "text", + "content": ". In the absence of anchor points, the class-posterior probability should incorporate substantial information to uniquely determine the minimum-volume simplex, with true sample probability points of each class being distinguishable. Therefore, the sufficiently scattered assumption [9] is introduced to limit the occurrence of anomalous clean class-posterior and aid in theoretically establishing that addressing problem (11) will drive " + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\hat{T}^j" + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "text", + "content": " to converge to " + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\pmb{T}^j" + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "inline_equation", + "content": "h_\\theta (\\hat{\\pmb{x}})" + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "text", + "content": " to accurately approximate " + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Y}|\\mathbf{X} = \\hat{\\mathbf{x}})" + }, + { + "bbox": [ + 313, + 287, + 555, + 410 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "text", + "content": "Theorem 1. If each " + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Y}^j | \\mathbf{X})" + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "text", + "content": " is sufficiently scattered, then " + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\star}^{j} = \\mathbf{T}^{j}" + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "inline_equation", + "content": "(h_{\\theta_{\\star}}(\\hat{\\mathbf{x}})_j, 1 - h_{\\theta_{\\star}}(\\hat{\\mathbf{x}})_j)^T = P(\\mathbf{Y}^j | \\mathbf{X} = \\hat{\\mathbf{x}})" + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "inline_equation", + "content": "j = 1, 2, \\dots, C" + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "text", + "content": ") must hold, where " + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "inline_equation", + "content": "(\\mathbf{T}_{\\star}^{1}, \\mathbf{T}_{\\star}^{2}, \\dots, \\mathbf{T}_{\\star}^{C}, \\theta_{\\star})" + }, + { + "bbox": [ + 313, + 415, + 554, + 463 + ], + "type": "text", + "content": " are the optimal solutions of the problem (11)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 468, + 553, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 553, + 492 + ], + "type": "text", + "content": "2.4. Cycle-Consistency Estimation and Model Regularization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "content": "In multi-label learning, imbalanced label distributions [44] and inter-class correlations [7] can compress class-posterior probability points and limit their scatter. Besides, the volume approximation of " + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "inline_equation", + "content": "T^j" + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "content": " in the problem (11) is highly sensitive to the noisy class-posterior, making it suffer from large estimation errors in noisy environments [18]. The failure of the sufficiently scattered assumption, along with the excessive reliance on the noisy class-posterior, compromises the effectiveness of the problem (11). To enhance transition matrix estimation, we develop an adaptive loss framework [4] that incorporates the volume maximization of the clean class-posterior, which leverages additional information to mitigate bias in single-target decision-making. Based on the equation " + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "inline_equation", + "content": "P(\\bar{Y}^j | \\mathbf{X}) = T^j P(\\mathbf{Y}^j | \\mathbf{X})" + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "inline_equation", + "content": "P(\\bar{Y}^j | \\mathbf{X})" + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "content": " can be derived from the noisy data, minimizing the volume of " + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "inline_equation", + "content": "T^j" + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "content": " is equivalent to imposing a volume constraint on " + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Y}^j | \\mathbf{X})" + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "content": ". Through the matrix " + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{T}^j" + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "content": " (an approximate inverse of " + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "inline_equation", + "content": "T^j" + }, + { + "bbox": [ + 313, + 498, + 555, + 714 + ], + "type": "text", + "content": "), we project the simplex formed in the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20709" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "noisy space onto the clean class-posterior space to obtain " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "P^{\\prime}(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": ", which will make up the largest volume. Then we design a cycle-consistency optimization strategy to enhance both the volume constraint on " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{T}^{j}" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " and maximization of " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": ", in which the estimated noisy class-posterior matches the provided noisy label " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Y}}^j" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": ", the backward estimated clean class-posterior " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "P^{\\prime}(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " aligns with the predicted clean-class posterior and the forward-backward consistency term " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\tilde{T}^j T^j P(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " coheres with " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "text", + "content": "Firstly, we employ forward estimation to simultaneously optimize the transition matrix " + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "inline_equation", + "content": "T^{j}" + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "text", + "content": " and classifier " + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "inline_equation", + "content": "h_{\\theta}" + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "text", + "content": ". By leveraging the cross-entropy loss and reporting the output as " + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "inline_equation", + "content": "F^{ij} = (h_{\\theta}(\\hat{\\boldsymbol{x}}_i)_j, 1 - h_{\\theta}(\\hat{\\boldsymbol{x}}_i)_j)^T" + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "text", + "content": ", we aim to minimize the approximation error between the inferred noisy class-posterior probability " + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "inline_equation", + "content": "T^{j}F^{ij}" + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "text", + "content": " and the given noisy label " + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\bar{Y}^{j}" + }, + { + "bbox": [ + 55, + 180, + 294, + 251 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 262, + 286, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 262, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 66, + 262, + 286, + 312 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {1} (\\theta , \\boldsymbol {T} ^ {1}, \\dots , \\boldsymbol {T} ^ {C}) = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} (\\bar {Y} _ {i, j} \\log (\\boldsymbol {T} _ {1} ^ {j} \\boldsymbol {F} ^ {i j}) \\\\ + \\left(1 - \\bar {Y} _ {i, j}\\right) \\log \\left(1 - \\mathbf {T} _ {2} ^ {j} \\mathbf {F} ^ {i j}\\right)). \\\\ \\end{array}", + "image_path": "2b54de236d60d488f93d76c2dad3937da180e014518a4102c45ad11cc4e6b53e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "text", + "content": "Secondly, the Sigmoid function is employed to convert the one-hot label into the noisy class-posterior probability, which is then normalized along the label dimension to capture the soft labels " + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{Y}}^{sn}" + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "text", + "content": ". Denoting the outcome " + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "inline_equation", + "content": "(\\bar{\\mathbf{Y}}_{i,j}^{sn}, 1 - \\bar{\\mathbf{Y}}_{i,j}^{sn})^T" + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}^{ij}" + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "text", + "content": ", we utilize " + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\tilde{T}^j" + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "text", + "content": " to project " + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}^{ij}" + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "text", + "content": " onto the clean class-posterior space. Thus, we minimize the following cross-entropy loss between the backward estimated clean class-posterior " + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\tilde{T}^j \\bar{\\mathbf{F}}^{ij}" + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "text", + "content": " and the true class-posterior probability estimation " + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "inline_equation", + "content": "h_\\theta(\\hat{\\boldsymbol{x}}_i)_j" + }, + { + "bbox": [ + 55, + 321, + 296, + 430 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 440, + 295, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 440, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 58, + 440, + 295, + 490 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {2} (\\theta , \\tilde {\\boldsymbol {T}} ^ {1}, \\dots , \\tilde {\\boldsymbol {T}} ^ {C}) = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} (h _ {\\theta} (\\hat {\\boldsymbol {x}} _ {i}) _ {j} \\log (\\tilde {\\boldsymbol {T}} _ {1} ^ {j} \\bar {\\boldsymbol {F}} ^ {i j}) \\\\ + \\left(1 - h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j}\\right) \\log \\left(1 - \\tilde {\\boldsymbol {T}} _ {2} ^ {j} \\bar {\\boldsymbol {F}} ^ {i j}\\right)). \\\\ \\end{array}", + "image_path": "e70666560002692efe9ed142705b4807af055ac4422f55e380d78fe12c8b1f64.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "text", + "content": "At this stage, the transition matrix can be derived from both the forward and backward ways by optimizing " + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "text", + "content": ". Thirdly, we should establish the cycle-consistency loss by minimizing the difference between " + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{T}^j T^j F^{ij}" + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "inline_equation", + "content": "h_\\theta (\\hat{x}_i)_j" + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "text", + "content": " to reinforce the invertible relationship of " + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "inline_equation", + "content": "T^j" + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{T}^j" + }, + { + "bbox": [ + 55, + 498, + 295, + 559 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 571, + 294, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 571, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 69, + 571, + 294, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {3} (\\theta) = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} \\left(h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j} \\log \\left(\\tilde {\\boldsymbol {T}} _ {1} ^ {j} \\boldsymbol {T} ^ {j} \\boldsymbol {F} ^ {i j}\\right) \\right. \\tag {12} \\\\ + \\left(1 - h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j}\\right) \\log \\left(1 - \\bar {\\boldsymbol {T}} _ {2} ^ {j} \\boldsymbol {T} ^ {j} \\boldsymbol {F} ^ {i j}\\right)), \\\\ \\end{array}", + "image_path": "f6eace68854fd1080c9cd2a0e5755b8f37f03f99b8bd018a95227abab21928b4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "text", + "content": "where we only optimize the network parameters " + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "text", + "content": " while keeping " + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "inline_equation", + "content": "T^j" + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "inline_equation", + "content": "\\tilde{T}^j" + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "text", + "content": " constant in order to maintain their consistency and focus on learning an optimal classifier. Therefore, the loss for identifying the transition matrix and learning the clean class-posterior is " + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IB} = \\mathcal{L}_1 + \\mathcal{L}_2 + \\mathcal{L}_3" + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "To mitigate overfitting on noisy data, we design model regularization by leveraging the partial fidelity of noisy la" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "bels and label correlations. We allow the classifier to directly access raw labels and obtain the classification loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 106, + 553, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 106, + 553, + 154 + ], + "spans": [ + { + "bbox": [ + 354, + 106, + 553, + 154 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {c} = - \\frac {1}{C N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {C} \\left(\\bar {Y} _ {i, j} \\log \\left(h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j}\\right)\\right) \\tag {13} \\\\ + \\left(1 - \\bar {Y} _ {i, j}\\right) \\log \\left(1 - h _ {\\theta} \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) _ {j})\\right). \\\\ \\end{array}", + "image_path": "908b75b2944be45627876cfde6765ef27c16820d3f3f9c90515e9d8cc5cee2f9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 161, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 553, + 317 + ], + "type": "text", + "content": "The construction of " + }, + { + "bbox": [ + 313, + 161, + 553, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 161, + 553, + 317 + ], + "type": "text", + "content": " is motivated by two key considerations. On one hand, it is essential to harness the underlying true information embedded in the noisy data to improve the inefficient classifier. On the other hand, strengthening the model memory with original noise helps minimize the risk of misclassifying clean samples as noisy ones. Since label correlation is a crucial factor for multi-label learning, we minimize the discrepancy in the label relationships reflected by the estimated clean class posterior and the classifier output to reduce classification bias. By defining the clean class-posterior matrix of the positive labels as " + }, + { + "bbox": [ + 313, + 161, + 553, + 317 + ], + "type": "inline_equation", + "content": "P^c" + }, + { + "bbox": [ + 313, + 161, + 553, + 317 + ], + "type": "text", + "content": ", with elements given by " + }, + { + "bbox": [ + 313, + 161, + 553, + 317 + ], + "type": "inline_equation", + "content": "P_{i,j}^{c} = \\tilde{T}_{1}^{j}\\bar{F}^{ij}" + }, + { + "bbox": [ + 313, + 161, + 553, + 317 + ], + "type": "text", + "content": ", we can build the following correlation-aligned loss driven by label co-occurrence [43]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 329, + 326, + 553, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 326, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 329, + 326, + 553, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\frac {1}{C ^ {2}} \\left\\| \\left(\\boldsymbol {P} ^ {c}\\right) ^ {T} \\boldsymbol {P} ^ {c} - \\left(h _ {\\theta} (\\hat {\\boldsymbol {X}})\\right) ^ {T} h _ {\\theta} (\\hat {\\boldsymbol {X}}) \\right\\| _ {F} ^ {2}. \\tag {14}", + "image_path": "a7b9051bdbd5bd9d1a263e7c3e9ce30839eb7319e0432067b3970536bedb55f4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 361, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 384 + ], + "type": "text", + "content": "Thus, our model regularization can be formulated as " + }, + { + "bbox": [ + 313, + 361, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R = \\mathcal{L}_c + \\mathcal{L}_r" + }, + { + "bbox": [ + 313, + 361, + 553, + 384 + ], + "type": "text", + "content": ". The total training loss of our model is" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 394, + 553, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 394, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 376, + 394, + 553, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {I B} + \\lambda_ {1} \\mathcal {L} _ {I D} + \\lambda_ {2} \\mathcal {L} _ {R}, \\tag {15}", + "image_path": "b18d6bb70bf12c04bd9e08eea0a80180868c38354712cb1ec83c05f9a9781bf9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 415, + 553, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 415, + 553, + 439 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 415, + 553, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 553, + 439 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 415, + 553, + 439 + ], + "type": "text", + "content": " are penalty coefficients to balance the effects of different losses when handling noisy labels." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 449, + 394, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 394, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 394, + 462 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 468, + 436, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 436, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 436, + 481 + ], + "type": "text", + "content": "3.1. Experimental Setting" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 486, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 617 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. Following [23, 45], our experiments incorporate six widely recognized multi-view multi-label datasets, i.e., Yeast [11], Corel 5k [5], VOC 2007 [6], Espgame [1], IAPR TC-12 [10], and MIR FLICKR [15]. Similar to [25], we employ six diverse metrics to comprehensively evaluate each method, i.e., Hamming Loss (HL), Ranking Loss (RL), OneError (OE), Coverage (Cov), Average Precision (AP), and Area Under Curve (AUC). To facilitate performance comparison, we present 1-HL, 1-RL, 1-OE, and 1-Cov values in our report, with higher values consistently indicating better performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Comparison Methods. To validate the effectiveness of DMMIvNL, we perform a comparative analysis against eight state-of-the-art methods i.e., iMVWL, TM3L, AlM-Net, DICNet, DIMC, MTD, SIP, ENTMLC. Since the first seven methods are designed to handle incomplete multiview multi-label problems, we train them directly on noisy data without modifications. As ENTMLC specializes in addressing multi-label noise, we tackle missing views by fill" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20710" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 100, + 555, + 539 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Experimental results of nine methods on all datasets with PER=50% and ρ = 30%. The best result on each row is bolded and the second-best result is underlined. The average ranking on the six metrics is shown as Ave.R." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 100, + 555, + 539 + ], + "lines": [ + { + "bbox": [ + 59, + 100, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 59, + 100, + 555, + 539 + ], + "type": "table", + "html": "
DataMETDICNetDIMCMTDSIPAIMNetiMVWLTM3LNTMLCSDMMIvNL
Yeast1-HL71.54±1.8174.94±1.5676.40±0.3371.24±1.4973.03±1.7966.94±0.4763.89±0.2969.88±0.1677.64±0.47
1-RL73.81±0.9174.82±2.3377.08±0.7571.30±1.6773.95±0.4372.83±1.0565.72±0.3771.40±3.1480.18±0.37
1-OE71.86±3.7470.95±5.3775.33±0.4861.94±6.7470.66±2.7963.91±2.2956.49±0.7773.28±0.4373.69±1.22
1-Cov41.68±2.4844.12±3.1045.53±1.9539.04±2.1641.19±0.3141.50±1.4534.93±0.7335.55±1.3750.77±0.84
AP65.61±1.3466.91±2.0968.87±0.8261.78±2.2565.55±1.1863.74±0.9256.90±0.3066.69±0.6072.08±0.34
AUC75.33±0.7376.34±2.3578.36±0.5473.42±1.5375.46±0.4574.85±0.9767.92±0.2873.33±2.4481.03±0.55
Ave.R4.503.331.837.174.836.338.837.001.17
Corel 5k1-HL96.37±0.6896.08±0.9998.65±0.0098.59±0.0498.66±0.0079.41±0.0597.23±0.6798.57±0.1798.68±0.02
1-RL60.14±0.8461.00±0.8767.26±0.5965.32±1.4163.37±0.6564.53±0.5552.01±0.3965.92±1.2372.27±0.57
1-OE6.30±0.895.44±1.0521.22±1.0612.88±4.2824.00±0.7310.29±2.4620.18±0.332.30±1.6725.20±0.78
1-Cov32.29±0.8133.31±0.9637.53±0.8836.69±2.0341.62±0.4736.09±1.1233.03±0.0235.19±0.9644.35±1.10
AP6.72±0.406.80±0.2915.22±0.6510.54±1.8618.49±0.2110.18±0.914.48±0.098.25±0.9919.18±0.45
AUC60.30±0.8551.74±1.3067.63±0.5965.69±1.5470.37±0.3764.81±0.7658.99±0.1165.97±1.2772.42±0.47
Ave.R7.677.672.834.332.676.007.335.501.00
VOC 20071-HL89.48±1.7089.72±1.3192.85±0.0592.00±0.1792.23±0.3274.95±0.2084.92±0.2092.66±0.0292.67±0.06
1-RL64.72±0.8965.70±0.7666.04±1.1063.69±1.7263.56±0.9267.03±0.8156.30±0.5866.03±1.9572.07±0.84
1-OE26.14±3.4131.28±3.1939.21±0.2131.84±4.2135.22±4.8337.47±1.5326.53±0.7139.35±1.8337.16±1.90
1-Cov58.13±0.7858.76±1.2059.06±0.9356.94±1.5256.55±0.8959.19±0.8153.99±0.3958.02±1.9765.64±1.21
AP34.45±1.3837.52±1.1440.18±0.6536.57±2.0337.16±2.1739.83±0.1827.49±0.3541.41±0.6843.86±1.15
AUC67.92±0.8068.84±1.0068.94±0.8766.52±1.4165.81±0.4969.76±0.6862.29±0.3968.55±2.0874.65±1.01
Ave.R6.835.172.506.506.503.678.673.501.67
Espauge1-HL97.88±0.1297.50±0.5698.40±0.0098.26±0.0198.41±0.0079.07±0.0494.75±4.9697.78±0.1898.25±0.01
1-RL62.26±0.3263.18±0.5063.94±0.3062.14±0.8464.67±0.6662.57±0.9854.05±0.6466.66±0.6069.79±0.47
1-OE14.49±1.2015.75±1.0814.97±0.7013.63±0.9118.75±0.2410.63±0.5417.27±7.211.48±1.1931.77±1.51
1-Cov29.73±0.3130.41±0.4933.06±0.3031.95±0.6634.65±0.6030.14±1.0125.73±0.1733.37±0.4335.80±0.55
AP9.83±0.3510.78±0.3211.28±0.1310.19±0.4713.38±0.299.66±0.555.82±0.096.77±0.7819.13±0.58
AUC62.15±0.3263.35±0.5664.12±0.2362.20±0.9066.28±0.7062.65±0.9156.03±0.0666.64±0.7969.86±0.54
Ave.R6.675.173.675.832.177.177.835.001.50
IAPR TC-121-HL97.54±0.0497.67±0.1898.00±0.0097.93±0.0398.00±0.0079.41±0.0288.22±1.2497.18±0.1498.04±0.02
1-RL61.86±0.3265.48±2.4566.90±0.1465.17±0.9169.20±0.3370.78±0.5755.79±0.9169.68±0.7372.66±0.48
1-OE14.86±0.6418.27±1.6829.51±0.8127.42±5.4433.89±4.6723.45±4.1615.02±9.561.49±1.3934.33±1.37
1-Cov25.98±0.4328.56±1.9529.24±0.3128.12±1.1631.06±0.5034.23±1.0823.23±0.2933.67±0.4635.40±0.53
AP9.18±0.2412.07±1.5715.10±0.1013.42±0.7817.37±0.8914.88±0.636.25±0.077.32±0.4319.66±0.07
AUC62.12±0.2465.52±2.3866.95±0.0965.38±0.6769.48±0.3170.87±0.6057.53±0.1169.75±0.8872.56±0.50
Ave.R7.505.833.835.673.174.008.505.501.00
MIR FLICKR1-HL86.63±0.2585.90±0.8787.46±0.0786.24±0.3187.12±0.1973.17±0.1079.90±0.1386.24±0.9487.61±0.09
1-RL75.11±0.4376.91±1.6176.23±0.2073.11±0.5074.97±0.9575.97±0.8860.07±0.1979.74±1.6882.34±0.75
1-OE44.98±0.7643.88±1.7636.58±0.7335.51±1.1935.78±1.0736.97±2.9642.44±0.3738.67±6.7261.34±0.99
1-Cov46.06±0.5948.98±2.7949.39±0.2346.30±0.4947.18±3.1348.15±2.1435.08±0.0953.12±4.1256.94±0.98
AP42.55±0.4244.11±0.7940.80±0.2038.50±0.1840.71±0.2242.18±0.5629.72±0.3445.69±2.7354.89±0.94
AUC73.96±0.4175.67±1.9175.42±0.1772.27±0.5873.97±1.3874.62±0.7262.80±0.1479.16±1.2881.14±0.93
Ave.R5.173.834.337.506.175.838.003.171.00
", + "image_path": "c078ef559af90c5d5711894517a8926cc26cfb15b7d5ed4cc3774a349d6b956a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 558, + 294, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 294, + 592 + ], + "type": "text", + "content": "ing them with average imputation, followed by feature concatenation. All parameters of compared methods are configured as the recommended values in their original codes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 594, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 294, + 713 + ], + "type": "text", + "content": "Implementation Details. Each dataset is divided into training, validation and test sets in the ratio of 7:1:2. To simulate the partial view setting, a certain proportion of instances defined by the partial example ratio (PER) are randomly selected as unavailable instances on each view. Besides, we ensure each sample contains at least one complete view to avoid invalid cases. In the multi-label noise configuration, we set a uniform noise rate for each label, i.e., " + }, + { + "bbox": [ + 54, + 594, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 54, + 594, + 294, + 713 + ], + "type": "text", + "content": " of positive and negative labels are flipped to their opposite. All results are obtained from ten independent runs, with final outcomes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 558, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 594 + ], + "type": "text", + "content": "reported as mean values alongside standard deviations. Our model is implemented by PyTorch on one NVIDIA GeForce RTX 4090 GPU of 24GB memory." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 600, + 440, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 440, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 440, + 613 + ], + "type": "text", + "content": "3.2. Experimental Results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "type": "text", + "content": "Performance Evaluation. To evaluate the capabilities of DMMIvNL in addressing data deficiencies, we conduct extensive comparative experiments focusing on two critical aspects, i.e., label noise and view missing. Besides, we assign a broad range of noise rates and missing proportions to demonstrate the versatility of our approach. For label noise, " + }, + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "type": "text", + "content": " is considered at levels of " + }, + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{10\\%, 15\\%, 20\\%, 25\\%, 30\\%, 35\\%, 40\\}" + }, + { + "bbox": [ + 312, + 617, + 555, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20711" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 88, + 227, + 220 + ], + "blocks": [ + { + "bbox": [ + 83, + 88, + 227, + 220 + ], + "lines": [ + { + "bbox": [ + 83, + 88, + 227, + 220 + ], + "spans": [ + { + "bbox": [ + 83, + 88, + 227, + 220 + ], + "type": "image", + "image_path": "52e333e3da7bc03b518954980d5716f2be29d7c1523910ae13d73ab259d57d40.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 223, + 171, + 232 + ], + "lines": [ + { + "bbox": [ + 139, + 223, + 171, + 232 + ], + "spans": [ + { + "bbox": [ + 139, + 223, + 171, + 232 + ], + "type": "text", + "content": "(a) Yeast" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 231, + 89, + 376, + 219 + ], + "blocks": [ + { + "bbox": [ + 231, + 89, + 376, + 219 + ], + "lines": [ + { + "bbox": [ + 231, + 89, + 376, + 219 + ], + "spans": [ + { + "bbox": [ + 231, + 89, + 376, + 219 + ], + "type": "image", + "image_path": "161217ce1dd5492f730d2e346f8e5773c0fb12d3ea577345c201016e5bdcd951.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 222, + 324, + 232 + ], + "lines": [ + { + "bbox": [ + 281, + 222, + 324, + 232 + ], + "spans": [ + { + "bbox": [ + 281, + 222, + 324, + 232 + ], + "type": "text", + "content": "(b) Corel 5k" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 381, + 87, + 526, + 219 + ], + "blocks": [ + { + "bbox": [ + 381, + 87, + 526, + 219 + ], + "lines": [ + { + "bbox": [ + 381, + 87, + 526, + 219 + ], + "spans": [ + { + "bbox": [ + 381, + 87, + 526, + 219 + ], + "type": "image", + "image_path": "51dd9063fea18a18c94ceac0bdc1183280175dd99f78cef399314d9babdc3788.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 223, + 476, + 232 + ], + "lines": [ + { + "bbox": [ + 427, + 223, + 476, + 232 + ], + "spans": [ + { + "bbox": [ + 427, + 223, + 476, + 232 + ], + "type": "text", + "content": "(c) VOC 2007" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 85, + 240, + 230, + 373 + ], + "blocks": [ + { + "bbox": [ + 85, + 240, + 230, + 373 + ], + "lines": [ + { + "bbox": [ + 85, + 240, + 230, + 373 + ], + "spans": [ + { + "bbox": [ + 85, + 240, + 230, + 373 + ], + "type": "image", + "image_path": "5087eda2c27a671dfb81526747cb660bf098eba72dc48e63ae0c1c4921bcec58.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 376, + 181, + 385 + ], + "lines": [ + { + "bbox": [ + 131, + 376, + 181, + 385 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 181, + 385 + ], + "type": "text", + "content": "(d) ESP Game" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 232, + 241, + 376, + 373 + ], + "blocks": [ + { + "bbox": [ + 232, + 241, + 376, + 373 + ], + "lines": [ + { + "bbox": [ + 232, + 241, + 376, + 373 + ], + "spans": [ + { + "bbox": [ + 232, + 241, + 376, + 373 + ], + "type": "image", + "image_path": "b43bd37c2d906ae295d75fe5d75cbea65e4fbd37e423d9bfd6ff05ac29e24369.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 376, + 331, + 385 + ], + "lines": [ + { + "bbox": [ + 277, + 376, + 331, + 385 + ], + "spans": [ + { + "bbox": [ + 277, + 376, + 331, + 385 + ], + "type": "text", + "content": "(e) IAPR TC-12" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 382, + 242, + 526, + 373 + ], + "blocks": [ + { + "bbox": [ + 382, + 242, + 526, + 373 + ], + "lines": [ + { + "bbox": [ + 382, + 242, + 526, + 373 + ], + "spans": [ + { + "bbox": [ + 382, + 242, + 526, + 373 + ], + "type": "image", + "image_path": "dc5f98b1f92e024bb2699244a9a74ea23de57d560e685b12131311257aee1669.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 376, + 481, + 385 + ], + "lines": [ + { + "bbox": [ + 422, + 376, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 422, + 376, + 481, + 385 + ], + "type": "text", + "content": "(f) MIR FlickR" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 56, + 433, + 175, + 535 + ], + "blocks": [ + { + "bbox": [ + 113, + 399, + 496, + 411 + ], + "lines": [ + { + "bbox": [ + 113, + 399, + 496, + 411 + ], + "spans": [ + { + "bbox": [ + 113, + 399, + 496, + 411 + ], + "type": "text", + "content": "Figure 2. AUC comparisons on six datasets with noise rate " + }, + { + "bbox": [ + 113, + 399, + 496, + 411 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 113, + 399, + 496, + 411 + ], + "type": "text", + "content": " varying from " + }, + { + "bbox": [ + 113, + 399, + 496, + 411 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 113, + 399, + 496, + 411 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 113, + 399, + 496, + 411 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 113, + 399, + 496, + 411 + ], + "type": "text", + "content": " while PER=50%." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 433, + 175, + 535 + ], + "lines": [ + { + "bbox": [ + 56, + 433, + 175, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 175, + 535 + ], + "type": "image", + "image_path": "a3bf4e66f985ac688fb16cd191a1ebb13d852341fc3387040ef1eb6d45c73e89.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 539, + 138, + 548 + ], + "lines": [ + { + "bbox": [ + 95, + 539, + 138, + 548 + ], + "spans": [ + { + "bbox": [ + 95, + 539, + 138, + 548 + ], + "type": "text", + "content": "(a) Espgame" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 562, + 295, + 585 + ], + "lines": [ + { + "bbox": [ + 55, + 562, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 295, + 585 + ], + "type": "text", + "content": "Figure 3. Parameter analysis of the trade-off parameters " + }, + { + "bbox": [ + 55, + 562, + 295, + 585 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 562, + 295, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 562, + 295, + 585 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 562, + 295, + 585 + ], + "type": "text", + "content": " on Espgame and IAPR TC-12." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 175, + 433, + 294, + 535 + ], + "blocks": [ + { + "bbox": [ + 175, + 433, + 294, + 535 + ], + "lines": [ + { + "bbox": [ + 175, + 433, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 175, + 433, + 294, + 535 + ], + "type": "image", + "image_path": "744832ee31c6c5145dbd2301fb4869c6d23ce005eeab8129955e177eb4198031.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 539, + 263, + 548 + ], + "lines": [ + { + "bbox": [ + 208, + 539, + 263, + 548 + ], + "spans": [ + { + "bbox": [ + 208, + 539, + 263, + 548 + ], + "type": "text", + "content": "(b) IAPR TC-12" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": "while for view missing, PER encompasses values of " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\{10\\%, 30\\%, 50\\%, 70\\}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ". Tab. 1 displays all metrics with " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " fixed at " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " and PER fixed at " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ", Fig. 2 illustrates the variation in AUC as " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " varies from " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ", and Fig. 4 shows the metrics distribution when " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{PER} = 10\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ". Additional results are included in the Appendix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Regarding the label noise, we can find that 1) Our method outperforms all compared approaches, particularly distinguished in the most representative AP metric. For" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "text", + "content": "instance, our method achieves a remarkable AP improvement of up to " + }, + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "text", + "content": " on Espgame in Tab. 1. 2) Compared to NTMLCS for handling label noise, our approach reveals substantial advantages. Besides, in terms of the performance on VOC 2007, DMMIvNL exhibits greater stability than the stepwise estimation way of NTMLCS, which is achieved through the cycle-consistency framework that simultaneously determines the transition matrix and classifier. Additionally, our approach surpasses methods developed for incomplete MvMLC. For example, Fig. 2 illustrates that our DMMIvNL reliably improves performance even against the top-performing MvMLC methods MTD and AIMNet. 3) As " + }, + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "text", + "content": " increases from " + }, + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 431, + 555, + 647 + ], + "type": "text", + "content": ", DMMIvNL consistently outperforms the other eight methods. Moreover, many methods lose reliability at elevated noise levels, such as NTMLCS on Yeast, TM3L on MIR FLICKR and DIMC on Corel 5k, while our approach continues to learn a competent classifier and deliver robust performance." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "As for the view missing, we have the following insights: 1) Our method exhibits outstanding performance in handling missing views. As illustrated in the radar chart of Fig. 4, DMMIvNL consistently occupies the outermost circle, reflecting impressive results across all metrics. 2)" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 107, + 71, + 504, + 81 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 71, + 504, + 81 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 504, + 81 + ], + "type": "text", + "content": "DICNet DIMC MTD SIP AIMNet iMVWL TM3L NTMLCS DMMIvNL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20712" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 71, + 504, + 81 + ], + "blocks": [ + { + "bbox": [ + 105, + 71, + 504, + 81 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 504, + 81 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 504, + 81 + ], + "type": "image", + "image_path": "39fcff99320936a896395f5de3a52a15c222116229d7ade810422e84abe36cc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 85, + 180, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 85, + 180, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 85, + 180, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 85, + 180, + 182 + ], + "type": "image", + "image_path": "293accfec9fb56fb79f030aebdbc6a027c20742bebeb9c6a11ca4d0b1d88a6ca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 185, + 153, + 195 + ], + "lines": [ + { + "bbox": [ + 84, + 185, + 153, + 195 + ], + "spans": [ + { + "bbox": [ + 84, + 185, + 153, + 195 + ], + "type": "text", + "content": "(a) Yeast, PER=10%" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 182, + 87, + 304, + 182 + ], + "blocks": [ + { + "bbox": [ + 182, + 87, + 304, + 182 + ], + "lines": [ + { + "bbox": [ + 182, + 87, + 304, + 182 + ], + "spans": [ + { + "bbox": [ + 182, + 87, + 304, + 182 + ], + "type": "image", + "image_path": "ea37a13ebdea01c15b7321498646c40da7b46ba9611e5f104186ce58bafd1e6c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 185, + 286, + 195 + ], + "lines": [ + { + "bbox": [ + 199, + 185, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 199, + 185, + 286, + 195 + ], + "type": "text", + "content": "(b) VOC 2007, PER=10%" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 87, + 429, + 182 + ], + "blocks": [ + { + "bbox": [ + 306, + 87, + 429, + 182 + ], + "lines": [ + { + "bbox": [ + 306, + 87, + 429, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 429, + 182 + ], + "type": "image", + "image_path": "feb5d9a7f31e34a9bba5915ba2e66ee01618a4995f9f8288a5e294272e756e46.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 185, + 408, + 195 + ], + "lines": [ + { + "bbox": [ + 327, + 185, + 408, + 195 + ], + "spans": [ + { + "bbox": [ + 327, + 185, + 408, + 195 + ], + "type": "text", + "content": "(c) Espgame, PER=10%" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 431, + 87, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 431, + 87, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 431, + 87, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 431, + 87, + 553, + 182 + ], + "type": "image", + "image_path": "ddf951c0cc525b648deb557bd42932c0c55e56a04caa64fbffa1afa59dbf6bbf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 185, + 540, + 195 + ], + "lines": [ + { + "bbox": [ + 443, + 185, + 540, + 195 + ], + "spans": [ + { + "bbox": [ + 443, + 185, + 540, + 195 + ], + "type": "text", + "content": "(d) MIR FLICKR, PER=10%" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 56, + 204, + 180, + 299 + ], + "blocks": [ + { + "bbox": [ + 56, + 204, + 180, + 299 + ], + "lines": [ + { + "bbox": [ + 56, + 204, + 180, + 299 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 180, + 299 + ], + "type": "image", + "image_path": "4c651e7fd7c34e877e663208ef30fe9df9ac9e06ce05204bd05fa435f8af88e2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 302, + 153, + 312 + ], + "lines": [ + { + "bbox": [ + 84, + 302, + 153, + 312 + ], + "spans": [ + { + "bbox": [ + 84, + 302, + 153, + 312 + ], + "type": "text", + "content": "(e) Yeast, PER=70%" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 182, + 205, + 304, + 299 + ], + "blocks": [ + { + "bbox": [ + 182, + 205, + 304, + 299 + ], + "lines": [ + { + "bbox": [ + 182, + 205, + 304, + 299 + ], + "spans": [ + { + "bbox": [ + 182, + 205, + 304, + 299 + ], + "type": "image", + "image_path": "00030c4cb484d5b79e2bb2a13d20468ba8f41817a2735acb5da56978d2837a79.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 302, + 285, + 312 + ], + "lines": [ + { + "bbox": [ + 200, + 302, + 285, + 312 + ], + "spans": [ + { + "bbox": [ + 200, + 302, + 285, + 312 + ], + "type": "text", + "content": "(f) VOC 2007, PER=70%" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 205, + 429, + 299 + ], + "blocks": [ + { + "bbox": [ + 306, + 205, + 429, + 299 + ], + "lines": [ + { + "bbox": [ + 306, + 205, + 429, + 299 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 429, + 299 + ], + "type": "image", + "image_path": "65b3b5d287cdb538e84f9d8362c540fbda24e25c5c1247da342ba70dec3c5222.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 302, + 407, + 312 + ], + "lines": [ + { + "bbox": [ + 327, + 302, + 407, + 312 + ], + "spans": [ + { + "bbox": [ + 327, + 302, + 407, + 312 + ], + "type": "text", + "content": "(g) Espgame, PER=70%" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 431, + 205, + 553, + 299 + ], + "blocks": [ + { + "bbox": [ + 431, + 205, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 431, + 205, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 431, + 205, + 553, + 299 + ], + "type": "image", + "image_path": "5da81f1db5f85b0a15fc6edd59c01367941ca7b3bacd2756172a459b9080ba56.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 302, + 539, + 312 + ], + "lines": [ + { + "bbox": [ + 444, + 302, + 539, + 312 + ], + "spans": [ + { + "bbox": [ + 444, + 302, + 539, + 312 + ], + "type": "text", + "content": "(h) MIR FLICKR,PER=70%" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 325, + 499, + 338 + ], + "lines": [ + { + "bbox": [ + 110, + 325, + 499, + 338 + ], + "spans": [ + { + "bbox": [ + 110, + 325, + 499, + 338 + ], + "type": "text", + "content": "Figure 4. Experimental results of nine methods on four datasets with PER=10% and 70% while " + }, + { + "bbox": [ + 110, + 325, + 499, + 338 + ], + "type": "inline_equation", + "content": "\\rho = 30\\%" + }, + { + "bbox": [ + 110, + 325, + 499, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "text", + "content": "As PER changes from " + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "text", + "content": ", DMMIvNL continues to achieve top performance. Notably, while DIMC experiences a substantial decline at " + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "text", + "content": " PER, our method remains efficient. It indicates that DMMIvNL delivers strong robustness across varying levels of feature absence. 3) Our method makes remarkable breakthroughs over deep learning based methods for missing views, like SIP, MTD, and DICNet. For example, on MIR FLICKR, DMMIvNL improves 1-OE by nearly " + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "text", + "content": " compared to SIP at " + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 357, + 296, + 502 + ], + "type": "text", + "content": " PER. This improvement is primarily due to the focus on extracting the shared semantics and specific features through information separation, an aspect overlooked by other methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Ablation Study and Parameter Sensitivity. The ablation experiments are conducted to analyze the effect of the three crucial aspects of DMMIvNL, i.e., representation extraction based on the information bottleneck theory " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{IB})" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", cycle-consistency estimation for noise identification and classifier learning " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{ID})" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", and model regularization to combat overfitting " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_r)" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ". Note that loss " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " is retained since it is essential for classifier improvement during early training. The first three rows of Tab. 2 indicate that " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IB}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ID}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " are integral components of DMMIvNL. Besides, the comparison between the empirical loss " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vol}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " from problem (11) and " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ID}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " confirms that the cycle-consistency framework enhances the stability of transition matrix estimation. To study the impact of " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", we present the results under distinct parameter settings in Fig. 3. As shown in the heatmap, the performance of DMMIvNL exhibits variability and improves as " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " gets closer to 0.001 and " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " approaches 1." + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 316, + 398, + 553, + 473 + ], + "blocks": [ + { + "bbox": [ + 313, + 356, + 555, + 391 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 391 + ], + "type": "text", + "content": "Table 2. Ablation study on Yeast, Corel 5k and VOC 2007 with PER=50% and ρ = 30%. ‘√’ and ‘X’ represent the used and not used corresponding item, respectively." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 398, + 553, + 473 + ], + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 473 + ], + "type": "table", + "html": "
LvolLIBLIDLrEspgameCorel 5kVOC 2007
APAUCAPAUCAPAUC
XX✓✓0.17500.68460.17130.69720.41450.7366
X✓X✓0.18380.71130.18500.70890.43050.7367
X✓✓X0.18340.71310.18000.70960.43590.7148
✓X✓✓0.14990.67320.15430.62840.36120.6715
X✓✓✓0.19340.73070.19970.72840.46220.7566
", + "image_path": "1987b0ddaafc71a4d841655aa18f4fc3f329901d9b3e4e8103677248b3231412.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 506, + 388, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 388, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 388, + 519 + ], + "type": "text", + "content": "4. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "content": "To address the dual challenges of view missing and multi-label noise, we propose DMMIvNL in this paper. Built upon the information bottleneck theory, a mutual information model is established to constrain information exchange across diverse representations, along with a derived theoretical upper bound. Besides, we prove that minimizing the volume of the transition matrix for each category consistently leads to the optimal classifier. Moreover, stability in noise recognition is further improved by introducing a cycle-consistency estimation framework, while raw fidelity information and inherent label correlations are used to reduce overfitting. Finally, extensive experiments demonstrate the superiority of DMMIvNL. In future work, we will leverage the prior knowledge embedded in LLM to rectify inaccurate labels and tackle the issue of class imbalance." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20713" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 127 + ], + "type": "text", + "content": "This work was partially supported by the National Natural Science Foundation of China [Grant No. 62376281], and the Key NSF of China under Grant No. 62136005." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 139, + 114, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 139, + 114, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 139, + 114, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 159, + 295, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 295, + 193 + ], + "type": "text", + "content": "[1] Luis Von Ahn and Laura Dabbish. Labeling images with a computer game. In SIGCHI Conference on Human Factors in Computing Systems, page 319-326, 2004. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 194, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 295, + 237 + ], + "type": "text", + "content": "[2] Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, and Masashi Sugiyama. Confidence scores make instance-dependent label-noise learning possible. In ICML, pages 825-836, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "text", + "content": "[3] Jia-Yao Chen, Shao-Yuan Li, Sheng-Jun Huang, and Chm. Unm: A universal approach for noisen, songcan and wang, lei and xie, ming-kun. IEEE Transactions on Knowledge and Data Engineering, 36:4968-4980, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "text", + "content": "[4] De Cheng, Yixiong Ning, Nannan Wang, Xinbo Gao, Heng Yang, Yuxuan Du, Bo Han, and Tongliang Liu. Class-dependent label-noise learning with cycle-consistency regularization. pages 11104-11116, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 373 + ], + "type": "text", + "content": "[5] Pinar Duygulu, Kobus Barnard, Joao FG de Freitas, and David A Forsyth. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In ECCV, pages 97-112, 2002. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 374, + 295, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 417 + ], + "type": "text", + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2010. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 419, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 295, + 462 + ], + "type": "text", + "content": "[7] Yuling Fan, Jinghua Liu, Jianeng Tang, Peizhong Liu, Yaojin Lin, and Yongzhao Du. Learning correlation information for multi-label feature selection. Pattern Recognition, 145: 109899, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 464, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 295, + 497 + ], + "type": "text", + "content": "[8] Marco Federici, Anjan Dutta, Patrick Forre, Nate Kushman, and Zeynep Akata. Learning robust representations via multi-view information bottleneck. In ICLR, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 498, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 531 + ], + "type": "text", + "content": "[9] Xiao Fu, Kejun Huang, and Nicholas D Sidiropoulos. On identifiability of nonnegative matrix factorization. IEEE Signal Processing Letters, 25(3):328-332, 2018. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 533, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 577 + ], + "type": "text", + "content": "[10] Michael Grubinger, Paul Clough, Henning Müller, and Thomas Deselaers. The iapr tc-12 benchmark: A new evaluation resource for visual information systems. In International Workshop onto Image, pages 1–11, 2006. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 578, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 600 + ], + "type": "text", + "content": "[11] Cordelia Guillaumin. Multimodal semi-supervised learning for image classification. In CVPR, pages 902-909, 2010. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 601, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 645 + ], + "type": "text", + "content": "[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, I Tsang, and Masashi Sugiyama. Robust training of deep neural networks with extremely noisy labels. In NeurIPS, page 4, 2020. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[13] Pingting Hao, Weiping Ding, Wanfu Gao, and Jialong He. Exploring view-specific label relationships for multi-view multi-label feature selection. Information Sciences, 681: 121215, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[14] Huaxi Huang, Hui Kang, Sheng Liu, Olivier Salvado, Thierry Rakotoarivelo, Dadong Wang, and Tongliang Liu." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "text", + "content": "Paddles: Phase-amplitude spectrum disentangled early stopping for learning with noisy labels. In ICCV, pages 16719-16730, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "text", + "content": "[15] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In ICMR, pages 39-43, 2008. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 130, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 185 + ], + "type": "text", + "content": "[16] Xiaodong Jia, Xiao-Yuan Jing, Xiaoke Zhu, Songcan Chen, Bo Du, Ziyun Cai, Zhenyu He, and Dong Yue. Semi-supervised multi-view deep discriminant representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7):2496-2509, 2020. 1, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "text", + "content": "[17] Zhangqi Jiang, Tingjin Luo, and Xinyan Liang. Deep incomplete multi-view learning network with insufficient label information. In AAAI, pages 12919-12927, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 220, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 253 + ], + "type": "text", + "content": "[18] Rihuan Ke. Deep variation prior: Joint image denoising and noise variance estimation without clean data. IEEE Transactions on Image Processing, 33:2908-2923, 2024. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 255, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 297 + ], + "type": "text", + "content": "[19] Shikun Li, Xiaobo Xia, Hansong Zhang, Yibing Zhan, Shiming Ge, and Tongliang Liu. Estimating noise transition matrix with label correlations for noisy multi-label learning. In NeurIPS, pages 24184-24198, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 299, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 332 + ], + "type": "text", + "content": "[20] Xuefeng Li, Tongliang Liu, Bo Han, Gang Niu, and Masashi Sugiyama. Provably end-to-end label-noise learning without anchor points. In ICML, pages 6403-6413, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "type": "text", + "content": "[21] Yijie Lin, Yuanbiao Gou, Xiaotian Liu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Dual contrastive prediction for incomplete multi-view representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4): 4447-4461, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "type": "text", + "content": "[22] Chengliang Liu, Jie Wen, Xiaoling Luo, Chao Huang, Zhihao Wu, and Yong Xu. Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In AAAI, pages 8807-8815, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 434, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 478 + ], + "type": "text", + "content": "[23] Chengliang Liu, Jie Wen, Xiaoling Luo, and Yong Xu. Incomplete multi-view multi-label learning via label-guided masked view-and category-aware transformers. In AAAI, pages 8816-8824, 2023. 1, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 479, + 553, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 522 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 522 + ], + "type": "text", + "content": "[24] Chengliang Liu, Jinlong Jia, Jie Wen, Yabo Liu, Xiaoling Luo, Chao Huang, and Yong Xu. Attention-induced embedding imputation for incomplete multi-view partial multi-label classification. In AAAI, pages 13864-13872, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[25] Chengliang Liu, Jie Wen, Yabo Liu, Chao Huang, Zhihao Wu, Xiaoling Luo, and Yong Xu. Masked two-channel decoupling framework for incomplete multi-view weak multi-label learning. In NeurIPS, 2024. 1, 4, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "text", + "content": "[26] Chengliang Liu, Gehui Xu, Jie Wen, Yabo Liu, Chao Huang, and Yong Xu. Partial multi-view multi-label classification via semantic invariance learning and prototype modeling. In ICML, 2024. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 613, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 669 + ], + "type": "text", + "content": "[27] Chengliang Liu, Jie Wen, Yong Xu, Bob Zhang, Liqiang Nie, and Min Zhang. Reliable representation learning for incomplete multi-view missing multi-label classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2025. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[28] Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3):447-461, 2015. 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20714" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[29] Tingjin Luo, Yang Yang, Dongyun Yi, and Jieping Ye. Robust discriminative feature learning with calibrated data reconstruction and sparse low-rank model. Applied Intelligence, 54(3):2867-2880, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[30] Yong Luo, Dacheng Tao, Chang Xu, Chao Xu, Hong Liu, and Yonggang Wen. Multiview vector-valued manifold regularization for multilabel image classification. IEEE Transactions on Neural Networks and Learning Systems, 24(5): 709-722, 2013. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "text", + "content": "[31] Yubin Park and Joyce C Ho. Tackling overfitting in boosting for noisy healthcare data. IEEE Transactions on Knowledge and Data Engineering, 33(7):2995-3006, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[32] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In CVPR, pages 1944-1952, 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[33] Yalan Qin, Hanzhou Wu, Xinpeng Zhang, and Guorui Feng. Semi-supervised structured subspace learning for multi-view clustering. IEEE Transactions on Image Processing, 31:1-14, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[34] Alisa Smirnova, Jie Yang, Dingqi Yang, and Philippe CudreMauroux. Nessy: A neuro-symbolic system for label noise reduction. IEEE Transactions on Knowledge and Data Engineering, 35(8):8300-8311, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "text", + "content": "[35] Shiding Sun, Bo Wang, and Yingjie Tian. Decoupled representation for multi-view learning. Pattern Recognition, 151: 110377, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 408 + ], + "type": "text", + "content": "[36] Yu-Yin Sun, Yin Zhang, and Zhi-Hua Zhou. Multi-label learning with weak label. In AAAI, pages 593-598, 2010. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[37] Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In IJCAI, pages 2703-2709, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 477 + ], + "type": "text", + "content": "[38] Ziyi Wang and Yitian Xu. A two-stage multi-view partial multi-label learning for enhanced disambiguation. Knowledge-Based Systems, 293:111680, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[39] Jie Wen, Chengliang Liu, Shijie Deng, Yicheng Liu, Lunke Fei, Ke Yan, and Yong Xu. Deep double incomplete multiview multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems, 35(8):11396-11408, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[40] Ming-Kun Xie and Sheng-Jun Huang. Partial multi-label learning with noisy label identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3676-3687, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "text", + "content": "[41] Wulin Xie, Xiaohuan Lu, Yadong Liu, Jiang Long, Bob Zhang, Shuping Zhao, and Jie Wen. Uncertainty-aware pseudo-labeling and dual graph driven network for incomplete multi-view multi-label classification. In ACMMM, pages 6656-6665, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[42] Yu Yao, Tongliang Liu, Mingming Gong, Bo Han, Gang Niu, and Kun Zhang. Instance-dependent label-noise learning under a structural causal model. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[43] Yitong Yao, Jing Zhang, Peng Zhang, and Yueheng Sun. A dual-branch learning model with gradient-balanced loss for long-tailed multi-label text classification. ACM Transactions on Information Systems, 42(2):1-24, 2023. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 251 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[44] Selim F Yilmaz, E Batuhan Kaynak, Aykut Koç, Hamdi Dibeklioglu, and Suleyman Serdar Kozat. Multi-label sentiment analysis on 100 languages with dynamic weighting for label imbalance. IEEE Transactions on Neural Networks and Learning Systems, 34:331-343, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "text", + "content": "[45] Dawei Zhao, Qingwei Gao, Yixiang Lu, and Dong Sun. Two-step multi-view and multi-label learning with missing label via subspace learning. Applied Soft Computing, 102:107120, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "type": "text", + "content": "[46] Pengfei Zhu, Qi Hu, Qinghua Hu, Changqing Zhang, and Zhizhao Feng. Multi-view label embedding. Pattern Recognition, 84:126-135, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[47] Wenzhang Zhuge, Tingjin Luo, Ruidong Fan, Hong Tao, Chenping Hou, and Dongyun Yi. Absent multiview semisupervised classification. IEEE Transactions on Cybernetics, 54(3):1708-1721, 2024. 1" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20715" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_content_list.json b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..129cec5fe8c5ffe0dfbd090d108367220d0e3691 --- /dev/null +++ b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_content_list.json @@ -0,0 +1,1538 @@ +[ + { + "type": "text", + "text": "Thin-Shell-SfT: Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields", + "text_level": 1, + "bbox": [ + 130, + 128, + 867, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Navami Kairanda $^{1}$ Marc Habermann $^{1,2}$ Shanthika Naik $^{3}$ Christian Theobalt $^{1,2}$ Vladislav Golyanik $^{1}$ $^{1}$ MPI for Informatics, SIC ${}^{2}$ VIA Research Center ${}^{3}$ IIT Jodhpur", + "bbox": [ + 99, + 202, + 898, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 287, + 326, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D reconstruction of highly deformable surfaces (e.g. cloths) from monocular RGB videos is a challenging problem, and no solution provides a consistent and accurate recovery of fine-grained surface details. To account for the ill-posed nature of the setting, existing methods use deformation models with statistical, neural, or physical priors. They also predominantly rely on nonadaptive discrete surface representations (e.g. polygonal meshes), perform frame-by-frame optimisation leading to error propagation, and suffer from poor gradients of the mesh-based differentiable renderers. Consequently, fine surface details such as cloth wrinkles are often not recovered with the desired accuracy. In response to these limitations, we propose ThinShell-SfT, a new method for non-rigid 3D tracking that represents a surface as an implicit and continuous spatiotemporal neural field. We incorporate continuous thin shell physics prior based on the Kirchhoff-Love model for spatial regularisation, which starkly contrasts the discretised alternatives of earlier works. Lastly, we leverage 3D Gaussian splatting to differentiably render the surface into image space and optimise the deformations based on analysis-by-synthesis principles. Our Thin-Shell-SfT outperforms prior works qualitatively and quantitatively thanks to our continuous surface formulation in conjunction with a specially tailored simulation prior and surface-induced 3D Gaussians. See our project page at https://4dqv.mpi-inf.mpg.de/ThinShellSfT.", + "bbox": [ + 88, + 320, + 483, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 753, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Non-rigid 3D reconstruction and tracking of general deformable surfaces from a monocular RGB camera is an important, challenging and ill-posed problem that is far from being solved [71]. It has applications in game development, robotics and augmented reality, to name a few areas.", + "bbox": [ + 89, + 779, + 482, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior works on temporally-coherent general surface reconstruction can be grouped into Shape-from-Template (SfT) [3, 8, 33] and Non-Rigid Structure-from-Motion [39,", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50eace23936289e74a9e49973a93b3e7db24dcaf23c830454c35b1f9ab79e834.jpg", + "image_caption": [ + "Figure 1. Our Thin-Shell-SfT approach reconstructs high-fidelity deformable 3D surface geometry with fine-grained wrinkles from a monocular video, while the previous best method [33] struggles. The coloured tracks (\"Ours, Input View\") visualise the 3D Gaussian trajectories over time and across all input video frames." + ], + "image_footnote": [], + "bbox": [ + 519, + 287, + 593, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a76305e924ba4058082f0411934b7a6a215de218ebd3aa7d46894bba644baf3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 287, + 674, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6fd630602f3b6da00bfe490470051ae2a48a35a4a187cf363bd4c1adc1d8017.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 287, + 743, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/68722334cdc192f9dd2e6b40eb6f5ec7dd5442d342b3ef76fdf0ef837aaa21f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 287, + 823, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6075cd515a0aea28f6efbbd407d516bf197f9f0693a801471cb82fe398a94c7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 287, + 903, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "53, 65]; they rely on 2D point tracks across monocular images (NRSfM) [53, 65] or between the input image and the template (SfT) [3, 8]. Recent physics-based SfT approaches [33, 67] demonstrate state-of-the-art results and cause a paradigm shift from geometric- [3, 8, 52] to physics-based constraints and from 2D-point-based [8, 50] to dense photometric loss using differentiable renderers. However, even such approaches as $\\phi$ -SfT [33] do not support fine wrinkles and require multiple hours to reconstruct a limited number of frames ( $\\approx 50$ ) of a single object, which is due to the underlying surface mesh representation, i.e., an explicit and discrete. First, determining the mesh resolution for a specific scene is difficult, e.g. selecting a high resolution could prohibitively increase the memory and computational cost, while a lower resolution might not account for fine-scale deformations. Second, the Finite Element Mesh (FEM)-based differentiable physics simulators [42, 43] lead to inconsistent simulations at different resolutions, preventing the adoption of coarse-to-fine strategies for tracking. Third, simulator-based approaches [33, 67] optimise frame-by-frame and can get stuck in local minima. Moreover, mesh-based differentiable renderers used in recent approaches [33, 67] are nonadaptive and do not support the complex remeshing required for the dynamic details arising in deformable scenes.", + "bbox": [ + 511, + 523, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11373", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome these limitations, our key idea is to replace meshes with an adaptive and continuous surface and deformation representation1. We tightly couple it with differentiable physics to guide the deformations while ensuring photometric consistency with the monocular images through analysis-by-synthesis. We model surfaces as Kirchhoff-Love thin shells [46] and propose a physics-based continuous deformation prior. Our regulariser minimises the internal hyperelastic energy of the tracked surface, ensuring its physical plausibility and providing prior for occlusion handling. Unlike discrete priors operating on mesh vertices [33, 82], our continuous prior updates any point on the surface, enabling physics supervision for fine-grained details such as wrinkles. While similar modelling has been applied to cloth simulation [34], it has not been shown for inverse problems like monocular surface tracking, which involves many unknowns (e.g., material, forces, contacts) and ambiguities between them. In addition to continuous prior, we perform joint space-time optimisation while taking the causality of deformation into account to impose temporal coherence (i.e. the current deformed state can update previous state parameters but not the future ones). As the differentiable renderer, we employ 3D Gaussian Splatting [36], which recently emerged as a prominent technique for radiance field rendering. It integrates straightforwardly with our continuous per-point deformation model and offers high-quality image gradients due to the continuous volumetric radiance field formulation. We leverage differential geometric quantities from the thin shell physics to couple Gaussians to the surface and optimise the parameters of dynamically tracked Gaussians. While there are extensions of Gaussian splatting focussing on dynamic view synthesis [12, 81] or multi-view static surface reconstruction [24, 73], we show how to adapt it for surface tracking from monocular videos with de-facto absent multi-view 3D reconstruction cues [18] such as the $\\phi$ -SfT [33] dataset. The technical contributions of this paper are as follows:", + "bbox": [ + 89, + 90, + 485, + 648 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Thin-Shell-SfT, i.e. a new method for monocular non-rigid 3D surface tracking operating on a continuous adaptive spatiotemporal representation;", + "- A continuous deformation prior based on the principles of the Kirchhoff-Love thin shell theory and application of such a prior in an inverse problem (shape from template);", + "- Adaptation of 3D Gaussian Splating for 3D tracking of highly deformable dynamic surfaces forming folds and fine wrinkles, captured by a static monocular camera." + ], + "bbox": [ + 89, + 651, + 482, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experimental results on the challenging $\\phi$ -SfT benchmark [33] show a significant improvement over the state of the art in terms of reconstruction accuracy; see Fig. 1.", + "bbox": [ + 89, + 789, + 482, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 656, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We review methods for monocular reconstruction of general non-rigid surfaces. They differ in their assumptions about the deformation model, data terms, and the priors; we refer to a recent survey by Tretschk et al. [71] for an in-depth review. Similar to ours, many recent works integrate Gaussians with physics or clothing but in very different contexts such as simulation [14, 79], material estimation [44, 83], and tracking from multi-view video [13, 45, 84].", + "bbox": [ + 511, + 114, + 903, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shape-from-Template (SfT) methods [16, 27, 33, 50, 55, 56, 60, 82] assume a single static shape or template as a prior. A template often corresponds to the first frame of the sequence [3, 33, 50] and, in other cases, to the rigid initialisation [17, 61]. SfT approaches [3, 26, 50, 82] employ a 3D-2D reprojection constraint and deform the template using temporal and geometric soft constraints to encourage physical plausibility. On the other hand, learning-based SfTs [16, 21, 56, 61] encode the template in the network and regress the surface deformations. They are sometimes object-specific with template encoded in the neural network weights [17] and often object-generic [16, 22, 61] when supervised with synthetic datasets. Recently, Kairanda et al. [33] and Stoko et al. [67] explain 2D observations through physics-based simulation of the deformation process and employ differentiable rendering for per-sequence gradient-based optimisation. In contrast to traditional SfT [3, 8, 52], they do not require registrations (the correspondence between the 3D template and the image). Ours falls under this category as we require the template corresponding to the first frame. In contrast to the prior works, ours is the first method to continuously represent the surface and its dynamics with a neural field capable of representing high-frequency signals [66]. We supervise with thin shell physics constraint applied at continuous points in the domain and employ the Gaussian Splatting functionality [36].", + "bbox": [ + 511, + 236, + 903, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-Rigid Structure from Motion (NRSfM) [1, 6, 10, 51, 69] relies on motion cues of 2D point tracks over the input monocular images and outputs per-frame camera-object poses and the 3D shapes. Recent dense NRSfM methods [2, 22, 23, 40, 53, 65, 74] rely on per-pixel multi-frame optical flow or video registration [19]. In contrast, we do not make a restrictive assumption of available 2D point tracks and track highly challenging cloth deformations (fine wrinkles), which goes beyond NRSfM capabilities.", + "bbox": [ + 511, + 628, + 903, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Novel View Synthesis. Radiance field methods [36, 47, 48, 75] learn volumetric scene representation for high-quality novel-view synthesis from multi-view images, which are extended to support dynamic scenes [11, 15, 32, 38, 54, 70, 76, 81]. Though they demonstrate impressive results, most do not show or evaluate geometry reconstruction. Moreover, they require multi-view cues [18, 56]; in contrast, we focus on dynamic scenes captured from a single static camera without such cues.", + "bbox": [ + 511, + 765, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Most prior methods use a discrete number $N$ of points or mesh vertices (including neural ones e.g. [65]) where $N$ needs to be decided beforehand and cannot be changed during optimisation, unlike our continuous formulation, allowing arbitrary queries.", + "bbox": [ + 89, + 849, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11374", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physics-based Priors have been applied in inverse problems [25, 41, 59, 62]. The examples range from 3D human pose estimation [41, 63] and parameter estimation [29, 43, 49, 77] to dense SfT [33, 64, 67]. Recent works [9, 57] extend neural representations to physical parameter inference from videos. Some methods [31, 35, 80] combine differentiable simulation and rendering. In the inverse setting, we are the first to impose continuous physics constraints while prior methods use mesh-based simulators.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 238, + 230, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We review 3D Gaussian Splatting [36] and NeuralCloth-Sim [34], from which we take inspiration for the data term (ensuring consistency of reconstructions with input images) and the prior term (encouraging physical plausibility).", + "bbox": [ + 89, + 263, + 482, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) models a scene as a volumetric radiance field with a dense set of 3D Gaussians, each defined by its position (mean), anisotropic covariance, opacity, and colour. $N_{g}$ Gaussians are represented as", + "bbox": [ + 89, + 324, + 482, + 385 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\left\\{\\mathcal {N} \\left(\\mathbf {x} _ {i}, \\mathbf {R} _ {i}, \\mathbf {S} _ {i}\\right), o _ {i}, \\mathbf {c} _ {i} \\right\\} _ {i = 1} ^ {N _ {g}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 391, + 482, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ denotes the position; $\\mathbf{RSS}^{\\top}\\mathbf{R}^{\\top}$ is the anisotropic covariance parameterised as an ellipsoid with scale $\\mathbf{S}$ and rotation $\\mathbf{R}$ ; $o$ is the opacity, and $\\mathbf{c}$ is the colour of the Gaussian. In 3DGS [36], $\\mathcal{G}$ is optimised through gradient-based training with multi-view image loss. Learning a 3DGS representation from monocular inputs requires substantial adjustments in different contexts [11, 81], and we show its utility in SfT for the first time.", + "bbox": [ + 89, + 416, + 482, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeuralClothSim [34] is a recent quasistatic cloth simulator representing surface deformations as a coordinate-based implicit neural deformation field (NDF). Given a target simulation scenario specified by the initial surface state, material properties and external forces, NeuralClothSim learns the equilibrium deformation field using the laws of the Kirchhoff-Love thin shell theory. Upon convergence, the equilibrium state can be queried continuously and consistently at multiple resolutions. These properties, combined with the memory adaptivity of neural fields, make NeuralClothSim well-suitable for inverse problems like ours.", + "bbox": [ + 89, + 537, + 482, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Kairanda et al. [34] model a cloth quasistatically as an NDF $\\mathbf{u}(\\pmb {\\xi}):\\Omega \\mapsto \\mathbb{R}^3$ defined on its curvilinear coordinate space $\\Omega$ . For a volumetric thin shell such as cloth, the Kirchhoff-Love model [78] offers a reduced kinematic parameterisation of the volume characterised by a 2D midsurface that fully determines the strain components throughout the thickness. Following thin shell assumptions, they decompose the Green strain due to deformation $\\mathbf{u}$ parameterised by a multilayer perceptron (MLP)—into membrane strain $\\varepsilon$ and bending strain $\\kappa$ , measuring the inplane stretching and the change in curvature, respectively. NeuralClothSim then computes the internal hyperelastic energy $\\Psi [\\varepsilon ,\\kappa ;\\Phi ]$ as a functional of the geometric strains and", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the cloth's material properties $\\Phi$ . Under the action of external force $\\mathbf{f}$ , their neural solver finds the equilibrium solution following the principle of minimum potential energy, where the potential $\\int_{\\Omega} \\Psi d\\Omega - \\int_{\\Omega} \\mathbf{f} \\cdot \\mathbf{u} d\\Omega$ is employed as the loss function. Note that the forward model of NeuralCloth-Sim [34] cannot be naively applied or extended (e.g., when coupled with a differentiable renderer, estimating physics parameters with the loss of NeuralClothSim would not converge due to the lack of lower bound) in our inverse setting.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 239, + 604, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose Thin-Shell-SfT, a new template-based method for the fine-grained 3D reconstruction of a deforming surface (such as cloth, paper or metal) seen in a monocular RGB video; see Fig. 2. We aim to reconstruct continuous 3D surfaces $\\{\\mathbf{S}_t\\}_{t\\in [1,\\dots,T]}$ corresponding to the input image sequence $\\{\\mathbf{I}_t\\}_t$ and optional masks $\\{\\mathbf{M}_t\\}_t$ . Similar to previous surface tracking methods [33, 50], we assume a static camera with known calibration and take the surface $\\mathbf{S}_1$ corresponding to the first frame $t = 1$ as a template.", + "bbox": [ + 511, + 263, + 903, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to the discrete representations (e.g. points, meshes) used in previous monocular tracking approaches, our deformation model encodes the surface and its dynamics as continuous and adaptive neural fields (Sec. 4.1). We optimise the neural model by relating input monocular views to the estimated surface states using 3DGS [36], where the Gaussians are initialised and dynamically tracked with parameters induced from the surface deformation (Sec. 4.2). To ensure plausible deformations due to the inherent ambiguity of the monocular setting and in contrast to prior discrete regularisers, we model the surfaces as thin shells imposing the continuous Kirchhoff-Love physics constraints [34] on the neural field, allowing us to reconstruct fine-grained details (Sec. 4.3).", + "bbox": [ + 511, + 401, + 903, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Deformation Model and Parameterisation", + "text_level": 1, + "bbox": [ + 511, + 619, + 870, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deformation Model. As our deformation model, we employ a continuous surface representation and deformation dynamics modelled as coordinated-based neural fields. Consider a surface with initial state $\\mathbf{S}_1\\subset \\mathbb{R}^3$ and 2D parameterisation $\\Omega \\subset \\mathbb{R}^2$ . For any parametric point $\\xi \\coloneqq (\\xi^{1},\\xi^{2})\\in \\Omega$ , we represent its position on the initial surface state with a mapping $\\bar{\\mathbf{x}} (\\pmb {\\xi}): \\Omega \\to \\mathbf{S}_1$ . Furthermore, we encode the time-varying spatial location of $\\pmb{\\xi}$ on the tracked surface $\\mathbf{S}_t$ as $\\mathbf{x}(\\pmb {\\xi},t):\\Omega \\times [1,\\dots,T]\\rightarrow \\mathbf{S}_t$ , with", + "bbox": [ + 511, + 641, + 903, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {x} (\\boldsymbol {\\xi}, t) = \\bar {\\mathbf {x}} (\\boldsymbol {\\xi}) + \\mathbf {u} (\\boldsymbol {\\xi}, t), \\text {w i t h} \\mathbf {u} (\\boldsymbol {\\xi}, 1) = 0 \\text {a n d} \\tag {2} \\\\ \\mathbf {u} (\\boldsymbol {\\xi}, t) = \\lambda \\mathbf {u} (\\boldsymbol {\\xi}, t - 1) + \\mathcal {F} (\\boldsymbol {\\xi}, t), \\forall t > 1, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 784, + 903, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{u}(\\pmb {\\xi},t)$ is the deformation field, $\\lambda >0$ is a scalar value and $\\mathcal{F}(\\pmb {\\xi},t)$ represents the estimated deformation offset. Here, we encourage the conservation of momentum by setting the deformation of the future surface states in the direction of the previous deformations.", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11375", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d7cb6e1e3f11600bf14d5f740e23d65ac60d1f995075729125829d01a4f8e65.jpg", + "image_caption": [ + "Figure 2. Overview of Thin-Shell-SfT. Our deformation model encodes the surface and its dynamics as neural fields. Given the template $\\mathbf{S}_1$ , we first fit a reference field (NRF) from 2D parametric points $\\pmb{\\xi}$ to the initial 3D positions $\\bar{\\mathbf{x}}$ . In the main stage, we optimise the deformation field (NDF) $\\mathbf{u}(\\pmb{\\xi}, t)$ by relating estimated surface states $\\mathbf{S}_t / \\mathcal{G}_t$ to the input monocular views. We induce the dynamically tracked Gaussians to the surface by: (1) Computing their positions $\\mathbf{x}$ as the sum of the initial position $\\bar{\\mathbf{x}}$ and NDF output $\\mathbf{u}$ , (2) Setting their rotations $\\bar{\\mathbf{a}}_i$ as the template's local coordinate system, and (3) Fixing the normal scale $\\epsilon$ , and optimising the colour, opacity and tangential scales $(s_1, s_2)$ using only the template texture. For physical plausibility, we impose continuous Kirchhoff-Love physics constraints." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parameterisation of the Template and Deformations. Next, we present an adaptive parameterisation of the above deformation model that allocates capacity to dynamic deformation details. Assuming that the template is provided as a mesh $\\mathcal{M} \\subset \\mathbf{S}_1$ , we generate the corresponding 2D parametric space $\\mathcal{T} \\subset \\Omega$ using established techniques (e.g. conformal parameterisation [30]). To learn a continuous representation of the initial state, we fit an MLP, which we call neural reference field (NRF) $\\bar{\\mathbf{x}}(\\boldsymbol{\\xi}; \\Upsilon)$ to the template mesh parameterisation $\\mathcal{T} \\mapsto \\mathcal{M}$ . For training NRF, we construct points in the parametric and vertex space with randomly sampled barycentric coordinates and train with $\\ell_1$ geometry loss. Akin to NRF, we regress the spatio-temporal neural deformation field (NDF) $\\mathbf{u}(\\boldsymbol{\\xi}, t; \\Theta)$ using another MLP predicting deformation offsets, $\\mathcal{F}(\\boldsymbol{\\xi}, t; \\Theta): \\Omega \\times [1, \\dots, T] \\to \\mathbb{R}^3$ . Before training, the initial estimate for tracked surface points $\\mathbf{x}(\\boldsymbol{\\xi}, t; \\Theta)$ is the sum of noisy NDF deformations, $\\mathbf{u}(\\boldsymbol{\\xi}, t; \\Theta)$ and the pre-trained NRF $\\bar{\\mathbf{x}}(\\boldsymbol{\\xi}; \\Upsilon)$ as given by Eq. (2). Our key idea is to optimise the NDF weights $\\Theta$ to ensure photometric consistency of the tracked surfaces with input monocular views while minimising the internal energy of the surface states modelled as a thin shell. The NRF pre-fit enables the coupling of Gaussians to the surface and computing surface metrics for the thin shell energy. We must represent high-frequency signals, such as fine folds and wrinkles on the tracked surface. Moreover, our physics loss requires the computation of higher-order derivatives. Hence, we use sine activation [66] in both MLPs.", + "bbox": [ + 89, + 472, + 482, + 893 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference. At testing, $\\mathbf{x}(\\boldsymbol {\\xi},t;\\Theta)$ provides continuous access to the reconstructed surface, where the NRF and NDF networks can be consistently queried at varied spatiotemporal resolutions. Moreover, our surface deformation model provides temporal correspondences. Meshing and texture mapping can be achieved in the parametric domain and then transferred to the tracked surface in 3D using Eq. (2). We next describe how to optimise the deformation field $\\mathbf{u}(\\boldsymbol {\\xi},t;\\Theta)$ that ensures consistency with monocular images and the physical plausibility of reconstructions.", + "bbox": [ + 511, + 472, + 906, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Surface Tracking with Gaussian Splitting", + "text_level": 1, + "bbox": [ + 511, + 635, + 870, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recall that we have an NDF parameterised by $\\Theta$ that outputs the deformation $\\mathbf{u}(\\boldsymbol {\\xi},t;\\Theta)$ at any time step. We seek to optimise the NDF so that the tracked surfaces $\\{\\mathbf{S}_t\\}_{t}$ (evaluated with Eq. (2)) generate images matching with input views $\\{\\mathbf{I}_t\\}_{t}$ . A straightforward approach to encourage 3D-to-2D consistency is to minimise the photometric loss with differentiable rendering. We choose Gaussian Splatting [36] as the differentiable renderer, as it seamlessly integrates with our per-point (Eq. (2)) continuous deformation model. Initialisation of Gaussians. To initialise a set of Gaussians, we sample $N_{g}$ well-distributed points (e.g. Poisson disk sampling [7]) from the surface of the input template mesh $\\mathcal{M}$ . The sampled points include parametric coordinates $\\{\\pmb {\\xi}_i\\}_{i\\in [1,\\dots ,N_g]}$ , their corresponding positions $\\{\\bar{\\mathbf{x}}_i\\}_i$ , and colours $\\{\\mathbf{c}_i\\}_i$ . Next, we will motivate our approach to parameterising Gaussian rotations and scales. Unlike the orig", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11376", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "inal multi-view [36] or dynamic Gaussian methods [12, 81] that leverage multi-view cues [18], we have a single static camera and, consequently, a challenging monocular setup. Thus, learning the anisotropic covariance of 3D Gaussians from all input frames can lead to poor results due to the inherent monocular ambiguities between deformation, texture and appearance. In our datasets, the initial surface state is reasonably (but not exactly) flat and fully visible. With this assumption, we take the initial surface normal as a proxy for the viewing direction and then fix the scale along normal while allowing for optimisation of the scales along the initial surface tangents. More concretely, we set the rotation matrix of $i$ -th Gaussian equal to the local basis vectors, i.e. $\\mathbf{R}_i \\equiv [\\bar{\\mathbf{a}}_1 \\bar{\\mathbf{a}}_2 \\bar{\\mathbf{a}}_3]_i$ (see Sec. 4.3 for computation) of the initial state (NRF) $\\bar{\\mathbf{x}}(\\pmb{\\xi}_i; \\Upsilon)$ and the scale $s_3$ as small $\\epsilon$ along surface normal $\\bar{\\mathbf{a}}_3$ . Finally, the dynamically tracked surface-induced Gaussian $\\{\\mathcal{G}_t\\}_{t}$ can be written as", + "bbox": [ + 89, + 90, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {G} _ {t} = \\left\\{\\mathcal {N} \\left(\\mathbf {x} \\left(\\boldsymbol {\\xi} _ {i}, t; \\Upsilon , \\Theta\\right), \\left[ \\bar {\\mathbf {a}} _ {1} \\left(\\boldsymbol {\\xi} _ {i}; \\Upsilon\\right), \\bar {\\mathbf {a}} _ {2} \\left(\\boldsymbol {\\xi} _ {i}; \\Upsilon\\right), \\right. \\right. \\right. \\\\ \\left. \\bar {\\mathbf {a}} _ {3} (\\boldsymbol {\\xi} _ {i}; \\Upsilon) ], (s _ {1}, s _ {2}, \\epsilon) _ {i}\\right), o _ {i}, \\mathbf {c} _ {i} \\} _ {i = 1} ^ {N _ {g}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 354, + 483, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "following Eq. (1). Note that the deformed positions $\\mathbf{x}(\\pmb{\\xi}_i, t)$ are computed using our deformation model (Eq. (2)) and all other Gaussian parameters are shared across surface states.", + "bbox": [ + 89, + 402, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimisation of NDF and Gaussian Parameters. For optimisation, we use the $\\ell_1$ photometric loss similar to 3DGS [36]. If segmentation masks $\\{\\mathbf{M}_t\\}_t$ are available, an additional silhouette loss is added to primarily speed up training. Thus, our data loss reads as", + "bbox": [ + 89, + 446, + 483, + 522 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {d} \\left(s _ {1}, s _ {2}, o, \\mathbf {c}, \\Theta\\right) = \\mathcal {R} \\left(\\mathcal {G} _ {1}, \\mathbf {I} _ {1}; s _ {1}, s _ {2}, o, \\mathbf {c}\\right) + \\\\ \\sum_ {t = 2} ^ {T} \\mathcal {R} \\left(\\mathcal {G} _ {t}, \\mathbf {I} _ {t}; \\Theta\\right) + \\mathcal {R} \\left(\\tilde {\\mathcal {G}} _ {t}, \\mathbf {M} _ {t}; \\Theta\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 530, + 480, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{R}(x;\\phi)$ is the Gaussian rasterisation loss with inputs $x$ and optimisable parameters $\\phi$ , and $\\tilde{\\mathcal{G}}_t$ are the dynamic Gaussians $\\mathcal{G}_t$ with colour $\\mathbf{c}$ set to mask foreground (e.g. white). Note that splating in the first frame updates shared Gaussian properties, whereas all other frames backpropagate to update the deformation field parameters $\\Theta$ . With the formulation in Eq. (2), our method propagates gradient information from future states to update 3D reconstructions of the earlier frames. This enables global space-time surface optimisation while respecting the causality of the surface dynamics. While temporal consistency is implicit in Eq. (2), other variants of temporal regularisation are possible (see Appendix B). In contrast to previous 3DGS or dynamic Gaussian methods [81], we note that it is important to optimise all frames in each iteration rather than random frame $t \\in [2,\\dots,T]$ , as it can otherwise lead to local minima (incorrect learning of folds). Global optimisation is preferred as our deformation model is a single global MLP for all frames, and it is computationally efficient over random frame selection due to the causality of deformations", + "bbox": [ + 89, + 598, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6cd1f02763f8d9b00c8ffc73ec6fe1b58c57227312075ac724f213eb366024aa.jpg", + "image_caption": [ + "Figure 3. Thin-shell physics prior is a spatial regulariser that minimises the hyperelastic strain energy density due to deformation w.r.t. the known template." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 905, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(Eq. (2)). Moreover, we keep the number of Gaussians fixed (i.e., no adaptive density control) since Gaussian parameters are learned only from the first frame; it is not known beforehand how the deformation details will arise in future states.", + "bbox": [ + 511, + 348, + 906, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Thin Shell Physical Prior", + "text_level": 1, + "bbox": [ + 511, + 420, + 743, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If we optimise the NDF solely using the photometric loss, it could perfectly fit the input images but could result in physically implausible surfaces due to monocular ambiguity. Therefore, we formulate a physics loss applicable to thin shells inspired by the NeuralClothSim approach [34] using the Kirchhoff-Love theory [78]. Since we model the dynamically tracked surface with an MLP, NeuralClothSim enables physics-based prior directly on the continuous surface. However, unlike the forward model of NeuralCloth-Sim, the external forces (such as contacts and the wind) and the material properties generating the deformations are unknown in an inverse setting like ours. For this reason, we assign material parameters to values typical for the surface modelled (e.g., cloth and paper). In contrast to material, it is hard to set reasonable values for forces or boundary conditions as the space of external forces is too large. Nevertheless, intuitively, the image loss effectively takes the role of external force as it guides the motion and deformation of the tracked surface. Thus, we omit the potential energy due to external forces from the optimisation. Finally, with the strain evaluated from the deformation field and the assumed material, our physical prior aims to minimise the internal hyperelastic energy that captures the surface stretching, shearing and bending stiffness; see Fig. 3.", + "bbox": [ + 511, + 444, + 906, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Strain Computation. To evaluate stretching and bending strain at each training iteration, we randomly sample $N_{p}$ parametric points $\\{\\pmb{\\xi}_i\\}_{i=1}^{N_p}$ from the template mesh $\\mathcal{M}$ , and perform differential geometry operations on the initial surface state (i.e., NRF) and the deformed state (i.e., NDF). We next present the computation of the local quantities at", + "bbox": [ + 511, + 806, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11377", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each point $\\xi_{i}$ on the initial state using the pre-trained NRF $\\overline{\\mathbf{x}} (\\pmb {\\xi};\\Upsilon)$ . We use Greek letters for indexing quantities on the parameterised surface (e.g., $\\bar{\\mathbf{a}}_{\\alpha},\\alpha ,\\beta ,\\ldots = 1,2$ ). For notational clarity, we drop the input $\\xi_{i},t$ and network weights $\\Upsilon ,\\Theta$ in all the derived quantities (e.g., $\\bar{\\mathbf{a}}_1\\equiv \\bar{\\mathbf{a}}_1(\\pmb {\\xi}_i;\\Upsilon),\\varepsilon \\equiv \\varepsilon (\\pmb {\\xi}_i,t;\\Upsilon ,\\Theta))$ . In the first step, we extract the local covariant basis: $\\{\\bar{\\mathbf{a}}_{\\alpha}:= \\partial \\bar{\\mathbf{x}} /\\partial \\xi^{\\alpha},\\bar{\\mathbf{a}}_3:= \\bar{\\mathbf{a}}_1\\times \\bar{\\mathbf{a}}_2\\}$ , the set of two vectors tangential to the curvilinear coordinate lines $\\xi^{\\alpha}$ and the local normal. For the Gaussian samples $\\{\\pmb {\\xi}_i\\}_{i = 1}^{N_g}$ , an identical computation is employed for setting the covariant basis as the rotation parameters (as in Sec. 4.2). Next, we evaluate the surface metric tensor $\\bar{a}_{\\alpha \\beta}$ measuring the lengths and the curvature tensor $\\bar{b}_{\\alpha \\beta}$ measuring the curvature of the midsurface. Akin to the covariant tensors (e.g. $\\bar{b}_{\\alpha \\beta}$ ), their contravariant (e.g. $\\bar{b}^{\\alpha \\beta}$ ) and mixed variants (e.g. $\\bar{b}_{\\alpha}^{\\beta}$ ) counterparts are extracted as well.", + "bbox": [ + 89, + 90, + 483, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With these initial state quantities, we next present the strain—due to deformation—as a function of the NDF $\\mathbf{u}(\\boldsymbol {\\xi};\\Theta)$ . The predicted $\\mathbf{u} = \\hat{u}_i\\mathbf{e}_i$ in global Cartesian coordinate system is transformed to contravariant coordinate basis, $\\mathbf{u} = u_{\\alpha}\\bar{\\mathbf{a}}^{\\alpha} + u_{3}\\bar{\\mathbf{a}}^{3}$ for strain calculation [34]. Given the deformation gradient $\\mathbf{u}_{,\\alpha}$ derived from the NDF as", + "bbox": [ + 89, + 335, + 483, + 424 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {, \\alpha} = \\varphi_ {\\alpha \\lambda} \\bar {\\mathbf {a}} ^ {\\lambda} + \\varphi_ {\\alpha 3} \\bar {\\mathbf {a}} ^ {3}, \\text {w i t h} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 426, + 483, + 450 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {\\alpha \\lambda} := u _ {\\lambda} | _ {\\alpha} - \\bar {b} _ {\\alpha \\lambda} u _ {3} \\text {a n d} \\varphi_ {\\alpha 3} := u _ {3, \\alpha} + \\bar {b} _ {\\alpha} ^ {\\lambda} u _ {\\lambda},\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 446, + 437, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we evaluate the non-linear membrane strain $\\varepsilon = [\\varepsilon_{\\alpha \\beta}]$ and bending strain $\\kappa = [\\kappa_{\\alpha \\beta}]$ as", + "bbox": [ + 89, + 470, + 482, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\varepsilon_ {\\alpha \\beta} = \\frac {1}{2} (\\varphi_ {\\alpha \\beta} + \\varphi_ {\\beta \\alpha} + \\varphi_ {\\alpha \\lambda} \\varphi_ {\\beta} ^ {\\lambda} + \\varphi_ {\\alpha 3} \\varphi_ {\\beta 3}),\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 500, + 367, + 523 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa_ {\\alpha \\beta} = - \\varphi_ {\\alpha 3} | _ {\\beta} - \\bar {b} _ {\\beta} ^ {\\lambda} \\varphi_ {\\alpha \\lambda} + \\varphi_ {3} ^ {\\lambda} \\left(\\varphi_ {\\alpha \\lambda} | _ {\\beta} + \\frac {1}{2} \\bar {b} _ {\\alpha \\beta} \\varphi_ {\\lambda 3} - \\bar {b} _ {\\beta \\lambda} \\varphi_ {\\alpha 3}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 526, + 480, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{a}}^i$ denote the local contravariant basis and $\\bar{b}_{\\alpha \\beta}$ , the components of the curvature tensor. In Eq. (6), we use a vertical bar for covariant derivatives, lower comma notation for partial derivatives w.r.t. the curvilinear coordinates $\\xi^{\\alpha}$ (e.g. $u_{\\lambda}|_{\\alpha}$ , and $\\mathbf{u}_{,\\alpha} = \\partial \\mathbf{u} / \\partial \\xi^{\\alpha}$ ), and Einstein summation convention of repeated indices for tensorial operations (e.g., $\\varphi_{\\alpha \\lambda}\\varphi_{\\beta}^{\\lambda} = \\varphi_{\\alpha 1}\\varphi_{\\beta}^{1} + \\varphi_{\\alpha 2}\\varphi_{\\beta}^{2}$ ). Notably, we compute all the aforementioned physical quantities with automatic differentiation. Please see Appendix A for more details.", + "bbox": [ + 89, + 571, + 482, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NDF Optimisation. As material model, we use $\\Phi \\coloneqq \\{\\rho, h, E, \\nu\\}$ , with mass density $\\rho$ and the surface thickness $h$ , and elastic coefficients: Young's modulus $E$ , and Poisson's ratio $\\nu$ . For simplicity and computational efficiency, we use a linear isotropic constitutive model relating strain to stress, thereby computing the in-plane stiffness $D$ , the bending stiffness $B$ and the elastic tensor $\\mathbf{H} = [H^{\\alpha \\beta \\lambda \\delta}]$ on the initial template, which read as", + "bbox": [ + 89, + 708, + 482, + 829 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD := \\frac {E h}{1 - v ^ {2}}, B := \\frac {E h ^ {3}}{1 2 \\left(1 - v ^ {2}\\right)}, \\text {a n d} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 834, + 480, + 872 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {\\alpha \\beta \\lambda \\delta} := \\nu \\bar {a} ^ {\\alpha \\beta} \\bar {a} ^ {\\lambda \\delta} + \\frac {1}{2} (1 - \\nu) (\\bar {a} ^ {\\alpha \\lambda} \\bar {a} ^ {\\beta \\delta} + \\bar {a} ^ {\\alpha \\delta} \\bar {a} ^ {\\beta \\lambda}).\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 869, + 452, + 897 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, our physics loss over the tracked surface states reads as the following:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal{L}_{p}(\\Theta) = \\frac{1}{2N_{p}T}\\sum_{i = 1}^{N_{p}}\\sum_{t = 2}^{T}\\Big(\\underbrace{D\\varepsilon^{\\top}(\\boldsymbol{\\xi}_{i},t;\\Theta)\\mathbf{H}(\\boldsymbol{\\xi}_{i})\\varepsilon(\\boldsymbol{\\xi}_{i},t;\\Theta)}_{\\text{stretching/shearing stiffness}} \\\\ + \\underbrace {B \\boldsymbol {\\kappa} ^ {\\top} \\left(\\boldsymbol {\\xi} _ {i} , t ; \\Theta\\right) \\mathbf {H} \\left(\\boldsymbol {\\xi} _ {i}\\right) \\boldsymbol {\\kappa} \\left(\\boldsymbol {\\xi} _ {i} , t ; \\Theta\\right)} _ {\\text {b e n d i n g s t i f f n e s s}}) \\sqrt {\\bar {a} \\left(\\boldsymbol {\\xi} _ {i}\\right)}, \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 146, + 903, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sqrt{a} := |\\bar{\\mathbf{a}}_1 \\times \\bar{\\mathbf{a}}_2|$ . Note that the term inside $\\mathcal{L}_p$ evaluates per point; therefore, we re-sample at each training iteration to explore the continuous surface domain. Finally, we solve for the optimal NDF weights $\\Theta^*$ by minimising the objective function $\\mathcal{L} = \\lambda_d\\mathcal{L}_d + \\lambda_p\\mathcal{L}_p$ with empirically determined loss weights $\\{\\lambda_d, \\lambda_p\\}$ . The optimisation can be interpreted as an inverse simulation, with data loss guiding deformation as an external force and physics loss modelling the surface's intrinsic behaviour. We use iterative gradient-based optimisation to that end [37]. We choose $N_p \\ll N_g$ as the physics loss involves expensive higher-order derivative computations, whereas the Gaussian rasterisation with samples fixed over training iterations is rather efficient. With resampling at each iteration, the physics loss backpropagates to the continuous surface, offering adaptive, memory-efficient performance compared to mesh-based physics alternatives requiring high resolution.", + "bbox": [ + 511, + 255, + 906, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 513, + 527, + 718, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We implement Thin-Shell-SfT using PyTorch [58] and optimise separate NRF and NDF networks for each sequence. The NRF and NDF, both, employ the SIREN [66] architecture using a frequency of $\\omega = 5$ and $\\omega = 30$ , respectively. Both networks have five hidden layers and 256 units in each layer, and are optimised using Adam [37]. Each network is trained for $\\approx 2 \\cdot 10^{3}$ iterations; first, the template is fitted with NRF, followed by surface tracking with NDF. While NRF training takes up to two minutes, the core part of the method, i.e. the NDF training, typically takes between 30 minutes to one hour until convergence on an NVIDIA A100 GPU. The Gaussian scales along the normal are set to $\\epsilon = 10^{-5}$ . In our experiments, the number of Gaussian samples is $N_{g} \\approx 90k$ , and the number of samples for physics loss is $N_{p} = 100$ . Concerning the physics loss, the linear elastic material properties of the surface are set as follows for all sequences: $E = 5000\\mathrm{Pa}, \\nu = 0.25$ , and $h = 1.2\\mathrm{mm}$ . Moreover, we set the image and physics loss weights to $\\lambda_{d} = 5$ , and $\\lambda_{p} = 1$ , whereas the temporal constraint scalar is set to $\\lambda = 0.4$ . For visualisation of the tracked continuous surface, we generate temporally coherent meshes with ball-pivot meshing [4] of the template Gaussian positions $\\{\\bar{\\mathbf{x}}_i\\}_{i \\in [1,\\dots,N_g]}$ . We refer to", + "bbox": [ + 511, + 551, + 908, + 904 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11378", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9fcbcc0e7e77179623df6da807c22e84c8025480381d4d1b3a7cac8b88fe4bf7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Seq.S1S2S3S4S5S6S7S8S9Avg
N-NRSfM [65]8.2533.62104.677.0272.668.73129.438.0619.8154.69
D-NRSfM [53]17.144.464.4041.3726.9214.0212.499.915.2915.11
Shimada [61]19.6922.1833.5490.3092.7857.6249.2724.4553.1249.22
DDD [82]2.951.693.8025.7310.466.9715.647.6111.7710.87
Ngo [50]2.191.512.1715.9010.723.017.95*failfail5.92*
Stokko [67]6.13.912.514.511.715.16.910.18.69.93
φ-SfT [33]0.792.753.547.606.153.144.732.522.363.93
Ours1.170.552.4†/3.55.5†/5.78.692.513.82.273.003.3†/3.5
", + "bbox": [ + 91, + 88, + 480, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. We quantitatively compare Thin-Shell-SfT to the state of the art on the $\\phi$ -SfT real dataset. The average Chamfer distance is multiplied by $10^4$ for readability. “*” notes that Ngo et al. failed on the last few frames of S7, which we exclude from the error computation. “†” denotes that we report the numbers on the variant of temporal coherency constraint described in Appendix B.", + "bbox": [ + 89, + 176, + 482, + 260 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1c0b95ffb1b50ddf2ea5e719b0da4b9e6c95bfd51f36aeca65d5c7a1cd7658a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 263, + 171, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/339f950a0868abb95aa9562dc7c24748cfb57555b3b2fd2cb31babf5d0e37e35.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 327, + 171, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bffa853f26dc7b92741ee9742b5d151a50287d7c2a5935d288438b01f399fbd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 266, + 246, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6be45416f2de0cfb47e4cdfc09ac534516f4ab4b42a0453f27dbc1fbaa4f7212.jpg", + "image_caption": [ + "Ground-Truth Normals" + ], + "image_footnote": [], + "bbox": [ + 174, + 327, + 246, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bcec40b359c8f506edb14d59d71a08456ba03168851df4b18441ac90eff2695.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 267, + 321, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc6cafb72167351b3bc48653869df7fece3dc174c40fb9eecdc2d5acb5a80269.jpg", + "image_caption": [ + "Normal Map Ours" + ], + "image_footnote": [], + "bbox": [ + 250, + 327, + 321, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b29878780a6d9ec6ed3f95754d64061a715f195a54c872131dc6dbcca493c37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 267, + 398, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bd5bb8e6fd4384978b38c825f24250f410302ceeb96daaa9dc1c2dd4f7dade5.jpg", + "image_caption": [ + "Normal Map $\\phi$ -SfT", + "Figure 4. Comparison of the reconstruction normal maps for ours and $\\phi$ -SfT. We show the cosine normal consistency to ground truth on the right and the normal metrics in Tab. I (Appendix)." + ], + "image_footnote": [], + "bbox": [ + 325, + 327, + 398, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d6853676543f2b457af600ff02fbf67f66eda2c11b183337d7c27c7218271de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 267, + 475, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2864c47de2278b87ac2a8a6b520a671afcea0bac6288323d151966da38f6a658.jpg", + "image_caption": [ + "Normal Error Ours" + ], + "image_footnote": [], + "bbox": [ + 401, + 327, + 475, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Appendices B and D and Figs. II and III for studies on the effect of hyperparameters $\\omega, \\lambda$ and $N_{g}$ .", + "bbox": [ + 89, + 458, + 482, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and Error Metrics. We evaluate on the benchmark $\\phi$ -SfT dataset [33] consisting of nine RGB videos (and reference depths) of highly challenging cloth deformations. As is common with non-rigid reconstruction methods [33, 65], we rigidly align the reconstructions of all compared methods to the ground truth. This is achieved with Procrustes alignment [72] for the first frame, which is refined with rigid ICP [5] for subsequent frames.", + "bbox": [ + 89, + 489, + 482, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Comparison", + "text_level": 1, + "bbox": [ + 89, + 619, + 222, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SfT and NRSfM. We compare Thin-Shell-SfT to different SfT methods, namely $\\phi$ -SfT [33], Yu et al.'s Direct, Dense, Deformable (DDD) [82], Ngo et al.'s Ngo2015 [50] and Shimada et al.'s IsMo-GAN [61], and finally to Stotko et al.'s approach [67]. DDD is provided with the required hierarchy of coarse-to-fine templates, and Ngo2015 and Stotko et al. with the same template as in the $\\phi$ -SfT dataset. Since the dataset is richly textured, the 2D point tracking methods would perform well; therefore, we additionally compare to the NRSfM methods. In particular, we compare to Sidhu et al.'s Neural NRSfM (N-NRSfM) [65] and Parashar et al.'s Diff-NRSfM [53]. The 2D point correspondences required by the NRSfM methods are provided by densely tracking the 2D points across input images with multi-frame subspace flow (MFSF) [19, 20]. The first frame of the sequence is selected as a keyframe for 2D tracking.", + "bbox": [ + 89, + 642, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 5-(left), we show qualitative reconstructions of", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thin-Shell-SfT compared to the prior state of the art (SotA) [33, 53, 67]. We capture fine wrinkles and folds in the deforming sequences that were not addressed by the earlier works. Next, we reconstruct surfaces with severe self-occlusions due to multiple layered folds in the extended versions of the $\\phi$ -SfT sequences [33]. See Fig. 5-(right), where our physics-based model provides a reasonable prior for occluded regions; $\\phi$ -SfT fails here due to prohibitive memory requirements. In Tab. 1, we show a quantitative comparison where we report the Chamfer distance (with pseudoground-truth point clouds) against the state-of-the-art methods over all the $\\phi$ -SfT dataset sequences. We outperform the existing methods on most sequences, often substantially and on average over all sequences. While the default temporal coherency works best on the evaluated dataset, we notice qualitative improvements for two out of the nine sequences with the other variants. In Fig. 4, we further show comparisons between the normal maps of the tracked surface. Additional numerical comparisons with the previous SotA [33] on runtime and normal consistency are in Appendix C and Tab. I. We obtain a lower $0.009(\\ell_2)$ and $0.034$ (cosine) normal error, whereas $\\phi$ -SfT shows $0.013$ and $0.041$ , respectively. One of the primary reasons for the failure of $\\phi$ -SfT [33] is its limitation to low-resolution meshes ( $\\approx 300$ vertices) capturing just coarse deformations. Our NDF offers, similar to scene representation approaches [47, 75], a smooth (cf. global MLP), low-dimensional (cf. a fixed parameter count) deformation space that is adaptive to dynamic details and can represent high-frequency signals (cf. sine activation). Although we discretise the physics loss, we re-sample at each iteration, leading to an adaptive discretisation. In contrast, earlier physics-based methods [33, 67] use fixed discretisation, either due to differentiable simulation not supporting remeshing [33] or a fixed-resolution surrogate model [67]. Our continuous formulation enables unprecedented fine-grained results.", + "bbox": [ + 511, + 90, + 906, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dynamic View Synthesis. Next, we compare with state-of-the-art dynamic view synthesis methods, particularly K-Planes [15] and Deformable Gaussians [81]. For the latter, we initialise the Gaussians with points sampled from the input template instead of the default SfM points. The compared radiance field methods are very effective in novel-view synthesis for monocular videos that provide multiview cues [18]. However, they fail to recover spatiotemporally coherent surface geometry for monocular RGB videos captured with a single static camera; see Fig. 6.", + "bbox": [ + 511, + 636, + 908, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablative Studies", + "text_level": 1, + "bbox": [ + 511, + 800, + 674, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we test the following aspects of our method: contribution of the physics loss, global optimisation and surface-induced Gaussians. Additional ablations, including momentum and mask loss, are presented in the Appendix E.", + "bbox": [ + 511, + 823, + 903, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Physics Prior. While relying solely on image loss yields", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11379", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24d56cec1882690718dd630e7467da6ce2d5ba15b1734cc9155d8110310480ae.jpg", + "image_caption": [ + "Figure 5. Exampary 3D reconstructions. (Left:) Comparisons focusing on high-frequency wrinkles. Thin-Shell-SfT captures the wrinkles best among all compared methods in one of the most challenging examples, outperforming $\\phi$ -SfT [33], Stotko et al. [67] and Diff-NRSfM [53]. (Right:) Our results on the extended $\\phi$ -SfT dataset highlight the excellent tracking in the occluded regions." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b9ea7e84bce812eb006a66806e589ec0870436ae9c0521547b2d998d01deeb0.jpg", + "image_caption": [ + "Figure 6. We compare our reconstructions to dynamic view synthesis methods [15, 81] on the rendered depth maps. The texture details retained in the depth maps imply that the compared methods fail to learn accurate surface (mistake texture for geometry)." + ], + "image_footnote": [], + "bbox": [ + 93, + 303, + 480, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea6b27ca125c9ea7cd03434845fcad8e531ff88b2f70dcc1437419acbd2bea61.jpg", + "image_caption": [ + "Figure 7. Ablations. (a:) Without physics, the surface can severely stretch or shrink. (b:) Random frame optimisation leads to local minima, (c,d:) Not positioning and orienting Gaussians on the surface leads to distortions and poor reconstructions." + ], + "image_footnote": [], + "bbox": [ + 93, + 473, + 482, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "accurate RGB recovery during rendering, it can lead to severely distorted surface tracking. This arises from the inherent monocular ambiguity of our setting as multiple possible deformations in 3D correspond to the same 2D image. As seen in Fig. 7-(a), incorporating Kirchhoff-Love-based thin shell prior is essential for achieving physically plausible and accurate surface reconstruction.", + "bbox": [ + 89, + 686, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Global Optimisation. We test a version of our method employing optimisation with randomly selected frames at each iteration instead of the proposed joint optimisation over all frames. Fig. 7-(b) visualises the ablated results showing poor performance for folds and wrinkles. Without any multi-view cues, as in the case of the used $\\phi$ -SfT dataset, random frame optimisation leads to local minima.", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Surface-induced Gaussians. The specially-tailored initialisation and optimisation of Gaussian parameters (Eqs. (3) and (4)) are crucial for the accurate geometric reconstruction. In the ablated version of our method, we optimise the shared Gaussian parameters, i.e., covariance, opacity, and colour on all input frames (similar to earlier methods [36, 81]) instead of the single template frame as in Eq. (4). Due to the ambiguities between geometric details and appearance over the deforming sequence, this leads to wrongly reconstructed Gaussians, eventually resulting in poor surface reconstructions; see Fig. 7-(c). In addition, fixing the normal scale on the template frame is useful for preventing elongated Gaussians (Fig. 7-(d)).", + "bbox": [ + 511, + 303, + 906, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. II-appendix, we report the above results on the full $\\phi$ -SfT dataset. On average, we obtain Chamfer distances of 34.25, 14.0, 3.75, and 3.46 for no physics, no surface-induced Gaussians, no normal scale, and the full model, respectively. We notice that including physics and surface induction are crucial for accurate tracking.", + "bbox": [ + 511, + 500, + 905, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 603, + 633, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Thin-Shell-SfT, a new approach for dense monocular non-rigid 3D surface tracking relying on new principles, i.e. adaptive neural deformation field, continuous Kirchhoff-Love thin shell prior and surface-induced 3D Gaussian Splating. Experiments on the $\\phi$ -SfT dataset demonstrate that our method accurately reconstructs the challenging fine-grained deformations such as cloth folds and wrinkles. Overall, we significantly improve fine-grained surface tracking using an adaptive deformation model and continuous thin shell physics compared to existing approaches that only support coarse deformations. Thanks to the physics prior, Thin-Shell-SfT is reasonably robust to occlusions, although extreme self-collisions remain challenging. Similarly, tracking textureless surfaces is another possible research direction that would necessitate special handling in future.", + "bbox": [ + 511, + 628, + 906, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was partially supported by the VIA Research Center.", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11380", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Nonrigid structure from motion in trajectory space. In Advances in Neural Information Processing Systems (NeurIPS), 2009. 2", + "[2] Mohammad Dawud Ansari, Vladislav Golyanik, and Didier Stricker. Scalable dense monocular surface reconstruction. In International Conference on 3D Vision (3DV), 2017. 2", + "[3] Adrien Bartoli, Yan Gérard, François Chadebecq, Toby Collins, and Daniel Pizarro. Shape-from-template. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2015. 1, 2", + "[4] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE TVCG, 1999. 6", + "[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 7", + "[6] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Computer Vision and Pattern Recognition (CVPR), 2000. 2", + "[7] Robert Bridson. Fastoisson disk sampling in arbitrary dimensions. SIGGRAPH sketches, 2007. 4", + "[8] David Casillas-Perez, Daniel Pizarro, David Fuentes-Jimenez, Manuel Mazo, and Adrien Bartoli. The isowarp: the template-based visual geometry of isometric surfaces. International Journal of Computer Vision, 2021. 1, 2", + "[9] Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne Vouga, and Christoph Lassner. Virtual elastic objects. In Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[10] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. In Computer Vision and Pattern Recognition (CVPR), 2014. 2", + "[11] Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3", + "[12] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2, 5", + "[13] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splating: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 2", + "[14] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Gaussian splashing: Unified particles for versatile motion synthesis and rendering. Computer Vision and Pattern Recognition (CVPR), 2025. 2", + "[15] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7, 8", + "[16] David Fuentes-Jimenez, Daniel Pizarro, David Casillas-Perez, Toby Collins, and Adrien Bartoli. Texture-generic deep shape-from-template. IEEE Access, 2021. 2", + "[17] David Fuentes-Jimenez, Daniel Pizarro, David Casillas-Pérez, Toby Collins, and Adrien Bartoli. Deep shape-from-template: Single-image quasi-isometric deformable registration and reconstruction. Image and Vision Computing, 127: 104531, 2022. 2", + "[18] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 5, 7", + "[19] Ravi Garg, Anastasios Roussos, and Lourdes Agapito. A variational approach to video registration with subspace constraints. International Journal of Computer Vision (IJCV), 104(3):286-314, 2013. 2, 7", + "[20] Ravi Garg, Anastasios Roussos, and Lourdes Agapito. Source code of Multi-Frame Subspace Flow (MFSF). http://www0.cs.ucl.ac.uk/staff/lagapito/subspace_flow/, 2015.7", + "[21] Vladislav Golyanik, Soshi Shimada, Kiran Varanasi, and Didier Stricker. Hdm-net: Monocular non-rigid 3d reconstruction with learned deformation model. In EuroVR, 2018. 2", + "[22] Vladislav Golyanik, André Jonas, and Didier Stricker. Consolidating segmentwise non-rigid structure from motion. In Machine Vision Applications (MVA), 2019. 2", + "[23] Stella Graßhof and Sami Sebastian Brandt. Tensor-based non-rigid structure from motion. In Winter Conference on Applications of Computer Vision (WACV), 2022. 2", + "[24] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[25] Jingfan Guo, Jie Li, Rahul Narain, and Hyun Soo Park. Inverse simulation: Reconstructing dynamic geometry of clothed humans via optimal control. In Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[26] Marc Habermann, Weipeng Xu, Helge Rhodin, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Nrst: Non-rigid surface tracking from monocular video. German Conference on Pattern Recognition (GCPR), 2018. 2", + "[27] Nazim Haouchine and Stephane Cotin. Template-based monocular 3d recovery of elastic shapes using lagrangian multipliers. In Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[28] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2", + "[29] Florian Hofherr, Lukas Koestler, Florian Bernard, and Daniel Cremers. Neural implicit representations for physical parameter inference from a single video. In Winter Conference on Applications of Computer Vision (WACV), 2023. 3", + "[30] Alec Jacobson, Daniele Panozzo, et al. libigl: A simple $\\mathbf{C}++$ geometry processing library, 2018. https://libigl.github.io/. 4" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11381", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Miguel Jaques, Michael Burke, and Timothy Hospedales. Physics-as-inverse-graphics: Unsupervised physical parameter estimation from video. In International Conference on Learning Representations (ICLR), 2020. 3", + "[32] Erik Johnson, Marc Habermann, Soshi Shimada, Vladislav Golyanik, and Christian Theobalt. Unbiased 4d: Monocular 4d reconstruction with a neural deformation model. In Computer Vision and Pattern Recognition (CVPR) Workshops, 2023. 2", + "[33] Navami Kairanda, Edith Tretschk, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. $\\phi$ -SfT: Shape-from-template with a physics-based deformation model. In Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 7, 8", + "[34] Navami Kairanda, Marc Habermann, Christian Theobalt, and Vladislav Golyanik. Neuralclothsim: Neural deformation fields meet the thin shell theory. Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 5, 6, 1", + "[35] Rama Kandukuri, Jan Achterhold, Michael Moeller, and Joerg Stueckler. Learning to identify physical parameters from video using differentiable physics. In German Conference for Pattern Recognition (GCPR), 2020. 3", + "[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 2, 3, 4, 5, 8", + "[37] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv e-prints, 2017. 6", + "[38] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. European Conference on Computer Vision (ECCV), 2024. 2", + "[39] Suryansh Kumar. Jumping manifolds: Geometry aware dense non-rigid structure from motion. In Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[40] Suryansh Kumar, Anoop Cherian, Yuchao Dai, and Hongdong Li. Scalable dense non-rigid structure-from-motion: A grassmannian perspective. In Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[41] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep Physics-aware Inference of Cloth Deformation for Monocular Human Performance Capture. In International Conference on 3D Vision (3DV), 2021. 3", + "[42] Yifei Li, Tao Du, Kui Wu, Jie Xu, and Wojciech Matusik. Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics, 2022. 1", + "[43] Junbang Liang, Ming Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse problems. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 3", + "[44] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338, 2024. 2", + "[45] Alberta Longhini, Marcel Busching, Bardienus Pieter Duisterhof, Jens Lundell, Jeffrey Ichnowski, Marten Björkman," + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Danica Kragic. Cloth-splatting: 3d cloth state estimation from rgb supervision. In 8th Annual Conference on Robot Learning, 2024. 2", + "[46] Augustus Edward Hough Love. A treatise on the mathematical theory of elasticity. University press, 1927. 2", + "[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 7", + "[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 2", + "[49] J. Krishna Murthy, Miles Macklin, Florian Golemo, Vikram Voleti, Linda Petrini, Martin Weiss, Breandan Considine, Jérôme Parent-Lévesque, Kevin Xie, Kenny Erleben, Liam Paull, Florian Shkurti, Derek Nowrouzezahrai, and Sanja Fidler. gradsim: Differentiable simulation for system identification and visuomotor control. In International Conference on Learning Representations (ICLR), 2021. 3", + "[50] Dat Tien Ngo, Sanghyuk Park, Anne Jorstad, Alberto Crivellaro, Chang D. Yoo, and Pascal Fua. Dense image registration and deformable surface reconstruction in presence of occlusions and minimal texture. In International Conference on Computer Vision (ICCV), 2015. 1, 2, 3, 7", + "[51] Marco Paladini, Alessio Del Bue, Marko Stosic, Marija Dodig, Jo ao Xavier, and Lourdes Agapito. Factorization for non-rigid and articulated structure using metric projections. In Computer Vision and Pattern Recognition (CVPR), 2009.", + "[52] Shaifali Parashar, Daniel Pizarro, Adrien Bartoli, and Toby Collins. As-rigid-as-possible volumetric shape-from-template. In International Conference on Computer Vision (ICCV), 2015. 1, 2", + "[53] Shaifali Parashar, Mathieu Salzmann, and Pascal Fua. Local non-rigid structure-from-motion from diffeomorphic mappings. In Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7, 8", + "[54] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40 (6):1-12, 2021. 2", + "[55] Mathieu Perriollat, Richard Hartley, and Adrien Bartoli. Monocular template-based reconstruction of inextensible surfaces. International Journal of Computer Vision (IJCV), 95(2):124-137, 2011. 2", + "[56] Albert Pumarola, Antonio Agudo, Lorenzo Porzi, Alberto Sanfeliu, Vincent Lepetit, and Francesc Moreno-Noguer. Geometry-Aware Network for Non-Rigid Shape Prediction from a Single View. In Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[57] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11382", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 6", + "[59] Davis Rempe, Leonidas J. Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jimei Yang. Contact and human dynamics from monocular video. In European Conference on Computer Vision (ECCV), 2020. 3", + "[60] Mathieu Salzmann, Julien Pilet, Slobodan Ilic, and Pascal Fua. Surface deformation models for nonrigid 3d shape recovery. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 29(8):1481-1487, 2007. 2", + "[61] Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Didier Stricker. IsMo-GAN: Adversarial learning for monocular non-rigid 3d reconstruction. In Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 2, 7", + "[62] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics, 39(6), 2020. 3", + "[63] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, Patrick Pérez, and Christian Theobalt. Neural monocular 3d human motion capture with physical awareness. ACM Transactions on Graphics, 40(4), 2021. 3", + "[64] Soshi Shimada, Vladislav Golyanik, Patrick Pérez, and Christian Theobalt. Decaf: Monocular deformation capture for face and hand interactions. ACM Transactions on Graphics, 42(6), 2023. 3", + "[65] Vikramjit Sidhu, Edgar Tretschk, Vladislav Golyanik, Antonio Agudo, and Christian Theobalt. Neural dense non-rigid structure from motion with latent space constraints. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 7", + "[66] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 4, 6, 3", + "[67] David Stotko, Nils Wandel, and Reinhard Klein. Physics-guided shape-from-template: Monocular video perception through neural surrogate models. Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 7, 8", + "[68] Guoxing Sun, Rishabh Dabral, Pascal Fua, Christian Theobalt, and Marc Habermann. Metacap: Meta-learning priors from multi-view imagery for sparse-view human performance capture and rendering. In European Conference on Computer Vision (ECCV), 2024. 2", + "[69] Lorenzo Torresani, Aaron Hertzmann, and Chris Bregler. Nonrigid structure-from-motion: Estimating shape and motion with hierarchical priors. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 30(5), 2008. 2", + "[70] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In International Conference on Computer Vision (ICCV), 2021. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Edith Tretschk, Navami Kairanda, Mallikarjun BR, Rishabh Dabral, Adam Kortylewski, Bernhard Egger, Marc Habermann, Pascal Fua, Christian Theobalt, and Vladislav Golyanik. State of the art in dense monocular non-rigid 3d reconstruction. In Computer Graphics Forum (Eurographics State of the Art Reports), 2023. 1, 2", + "[72] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 1991. 7", + "[73] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. Games: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459, 2024. 2", + "[74] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation. In Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[75] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2, 7", + "[76] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. International Conference on Computer Vision (ICCV), 2023. 2", + "[77] Sebastian Weiss, Robert Maier, Daniel Cremers, Rüdiger Westermann, and Nils Thuerey. Correspondence-free material reconstruction using sparse surface constraints. Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[78] Gerald Kempner and Demosthenes Talaslidis. Mechanics of solids and shells. CRC, Boca Raton, 2003. 3, 5", + "[79] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[80] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In International Conference on Computer Vision (ICCV), 2023. 3", + "[81] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 7, 8", + "[82] Rui Yu, Chris Russell, Neill D. F. Campbell, and Lourdes Agapito. Direct, dense, and deformable: Template-based non-rigid 3d reconstruction from rgb video. In International Conference on Computer Vision (ICCV), 2015. 2, 7", + "[83] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y. Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. PhysDreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision (ECCV), 2024. 2", + "[84] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, and Gordon Wetzstein. Physavatar: Learning the physics of dressed 3d avatars from visual observations. In European Conference on Computer Vision (ECCV), 2024. 2" + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11383", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_model.json b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2a3728fcdf946c29d1250434867a89616e94929b --- /dev/null +++ b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.13, + 0.868, + 0.175 + ], + "angle": 0, + "content": "Thin-Shell-SfT: Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.9, + 0.24 + ], + "angle": 0, + "content": "Navami Kairanda\\(^{1}\\) Marc Habermann\\(^{1,2}\\) Shanthika Naik\\(^{3}\\) Christian Theobalt\\(^{1,2}\\) Vladislav Golyanik\\(^{1}\\) \n\\(^{1}\\)MPI for Informatics, SIC \\({}^{2}\\)VIA Research Center \\({}^{3}\\)IIT Jodhpur" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.289, + 0.327, + 0.306 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.321, + 0.485, + 0.73 + ], + "angle": 0, + "content": "3D reconstruction of highly deformable surfaces (e.g. cloths) from monocular RGB videos is a challenging problem, and no solution provides a consistent and accurate recovery of fine-grained surface details. To account for the ill-posed nature of the setting, existing methods use deformation models with statistical, neural, or physical priors. They also predominantly rely on nonadaptive discrete surface representations (e.g. polygonal meshes), perform frame-by-frame optimisation leading to error propagation, and suffer from poor gradients of the mesh-based differentiable renderers. Consequently, fine surface details such as cloth wrinkles are often not recovered with the desired accuracy. In response to these limitations, we propose ThinShell-SfT, a new method for non-rigid 3D tracking that represents a surface as an implicit and continuous spatiotemporal neural field. We incorporate continuous thin shell physics prior based on the Kirchhoff-Love model for spatial regularisation, which starkly contrasts the discretised alternatives of earlier works. Lastly, we leverage 3D Gaussian splatting to differentiably render the surface into image space and optimise the deformations based on analysis-by-synthesis principles. Our Thin-Shell-SfT outperforms prior works qualitatively and quantitatively thanks to our continuous surface formulation in conjunction with a specially tailored simulation prior and surface-induced 3D Gaussians. See our project page at https://4dqv.mpi-inf.mpg.de/ThinShellSfT." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.222, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Non-rigid 3D reconstruction and tracking of general deformable surfaces from a monocular RGB camera is an important, challenging and ill-posed problem that is far from being solved [71]. It has applications in game development, robotics and augmented reality, to name a few areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Prior works on temporally-coherent general surface reconstruction can be grouped into Shape-from-Template (SfT) [3, 8, 33] and Non-Rigid Structure-from-Motion [39," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.289, + 0.594, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.289, + 0.676, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.289, + 0.744, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.289, + 0.824, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.289, + 0.904, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.439, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Figure 1. Our Thin-Shell-SfT approach reconstructs high-fidelity deformable 3D surface geometry with fine-grained wrinkles from a monocular video, while the previous best method [33] struggles. The coloured tracks (\"Ours, Input View\") visualise the 3D Gaussian trajectories over time and across all input video frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.901 + ], + "angle": 0, + "content": "53, 65]; they rely on 2D point tracks across monocular images (NRSfM) [53, 65] or between the input image and the template (SfT) [3, 8]. Recent physics-based SfT approaches [33, 67] demonstrate state-of-the-art results and cause a paradigm shift from geometric- [3, 8, 52] to physics-based constraints and from 2D-point-based [8, 50] to dense photometric loss using differentiable renderers. However, even such approaches as \\(\\phi\\)-SfT [33] do not support fine wrinkles and require multiple hours to reconstruct a limited number of frames (\\(\\approx 50\\)) of a single object, which is due to the underlying surface mesh representation, i.e., an explicit and discrete. First, determining the mesh resolution for a specific scene is difficult, e.g. selecting a high resolution could prohibitively increase the memory and computational cost, while a lower resolution might not account for fine-scale deformations. Second, the Finite Element Mesh (FEM)-based differentiable physics simulators [42, 43] lead to inconsistent simulations at different resolutions, preventing the adoption of coarse-to-fine strategies for tracking. Third, simulator-based approaches [33, 67] optimise frame-by-frame and can get stuck in local minima. Moreover, mesh-based differentiable renderers used in recent approaches [33, 67] are nonadaptive and do not support the complex remeshing required for the dynamic details arising in deformable scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11373" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.65 + ], + "angle": 0, + "content": "To overcome these limitations, our key idea is to replace meshes with an adaptive and continuous surface and deformation representation1. We tightly couple it with differentiable physics to guide the deformations while ensuring photometric consistency with the monocular images through analysis-by-synthesis. We model surfaces as Kirchhoff-Love thin shells [46] and propose a physics-based continuous deformation prior. Our regulariser minimises the internal hyperelastic energy of the tracked surface, ensuring its physical plausibility and providing prior for occlusion handling. Unlike discrete priors operating on mesh vertices [33, 82], our continuous prior updates any point on the surface, enabling physics supervision for fine-grained details such as wrinkles. While similar modelling has been applied to cloth simulation [34], it has not been shown for inverse problems like monocular surface tracking, which involves many unknowns (e.g., material, forces, contacts) and ambiguities between them. In addition to continuous prior, we perform joint space-time optimisation while taking the causality of deformation into account to impose temporal coherence (i.e. the current deformed state can update previous state parameters but not the future ones). As the differentiable renderer, we employ 3D Gaussian Splatting [36], which recently emerged as a prominent technique for radiance field rendering. It integrates straightforwardly with our continuous per-point deformation model and offers high-quality image gradients due to the continuous volumetric radiance field formulation. We leverage differential geometric quantities from the thin shell physics to couple Gaussians to the surface and optimise the parameters of dynamically tracked Gaussians. While there are extensions of Gaussian splatting focussing on dynamic view synthesis [12, 81] or multi-view static surface reconstruction [24, 73], we show how to adapt it for surface tracking from monocular videos with de-facto absent multi-view 3D reconstruction cues [18] such as the \\(\\phi\\)-SfT [33] dataset. The technical contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.652, + 0.483, + 0.697 + ], + "angle": 0, + "content": "- Thin-Shell-SfT, i.e. a new method for monocular non-rigid 3D surface tracking operating on a continuous adaptive spatiotemporal representation;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.698, + 0.483, + 0.742 + ], + "angle": 0, + "content": "- A continuous deformation prior based on the principles of the Kirchhoff-Love thin shell theory and application of such a prior in an inverse problem (shape from template);" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.743, + 0.483, + 0.787 + ], + "angle": 0, + "content": "- Adaptation of 3D Gaussian Splating for 3D tracking of highly deformable dynamic surfaces forming folds and fine wrinkles, captured by a static monocular camera." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.652, + 0.483, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Our experimental results on the challenging \\(\\phi\\)-SfT benchmark [33] show a significant improvement over the state of the art in terms of reconstruction accuracy; see Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.657, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.236 + ], + "angle": 0, + "content": "We review methods for monocular reconstruction of general non-rigid surfaces. They differ in their assumptions about the deformation model, data terms, and the priors; we refer to a recent survey by Tretschk et al. [71] for an in-depth review. Similar to ours, many recent works integrate Gaussians with physics or clothing but in very different contexts such as simulation [14, 79], material estimation [44, 83], and tracking from multi-view video [13, 45, 84]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.237, + 0.905, + 0.629 + ], + "angle": 0, + "content": "Shape-from-Template (SfT) methods [16, 27, 33, 50, 55, 56, 60, 82] assume a single static shape or template as a prior. A template often corresponds to the first frame of the sequence [3, 33, 50] and, in other cases, to the rigid initialisation [17, 61]. SfT approaches [3, 26, 50, 82] employ a 3D-2D reprojection constraint and deform the template using temporal and geometric soft constraints to encourage physical plausibility. On the other hand, learning-based SfTs [16, 21, 56, 61] encode the template in the network and regress the surface deformations. They are sometimes object-specific with template encoded in the neural network weights [17] and often object-generic [16, 22, 61] when supervised with synthetic datasets. Recently, Kairanda et al. [33] and Stoko et al. [67] explain 2D observations through physics-based simulation of the deformation process and employ differentiable rendering for per-sequence gradient-based optimisation. In contrast to traditional SfT [3, 8, 52], they do not require registrations (the correspondence between the 3D template and the image). Ours falls under this category as we require the template corresponding to the first frame. In contrast to the prior works, ours is the first method to continuously represent the surface and its dynamics with a neural field capable of representing high-frequency signals [66]. We supervise with thin shell physics constraint applied at continuous points in the domain and employ the Gaussian Splatting functionality [36]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.905, + 0.765 + ], + "angle": 0, + "content": "Non-Rigid Structure from Motion (NRSfM) [1, 6, 10, 51, 69] relies on motion cues of 2D point tracks over the input monocular images and outputs per-frame camera-object poses and the 3D shapes. Recent dense NRSfM methods [2, 22, 23, 40, 53, 65, 74] rely on per-pixel multi-frame optical flow or video registration [19]. In contrast, we do not make a restrictive assumption of available 2D point tracks and track highly challenging cloth deformations (fine wrinkles), which goes beyond NRSfM capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Dynamic Novel View Synthesis. Radiance field methods [36, 47, 48, 75] learn volumetric scene representation for high-quality novel-view synthesis from multi-view images, which are extended to support dynamic scenes [11, 15, 32, 38, 54, 70, 76, 81]. Though they demonstrate impressive results, most do not show or evaluate geometry reconstruction. Moreover, they require multi-view cues [18, 56]; in contrast, we focus on dynamic scenes captured from a single static camera without such cues." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Most prior methods use a discrete number \\(N\\) of points or mesh vertices (including neural ones e.g. [65]) where \\(N\\) needs to be decided beforehand and cannot be changed during optimisation, unlike our continuous formulation, allowing arbitrary queries." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11374" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "Physics-based Priors have been applied in inverse problems [25, 41, 59, 62]. The examples range from 3D human pose estimation [41, 63] and parameter estimation [29, 43, 49, 77] to dense SfT [33, 64, 67]. Recent works [9, 57] extend neural representations to physical parameter inference from videos. Some methods [31, 35, 80] combine differentiable simulation and rendering. In the inverse setting, we are the first to impose continuous physics constraints while prior methods use mesh-based simulators." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.239, + 0.231, + 0.255 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.483, + 0.325 + ], + "angle": 0, + "content": "We review 3D Gaussian Splatting [36] and NeuralCloth-Sim [34], from which we take inspiration for the data term (ensuring consistency of reconstructions with input images) and the prior term (encouraging physical plausibility)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.483, + 0.386 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) models a scene as a volumetric radiance field with a dense set of 3D Gaussians, each defined by its position (mean), anisotropic covariance, opacity, and colour. \\(N_{g}\\) Gaussians are represented as" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.392, + 0.483, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\left\\{\\mathcal {N} \\left(\\mathbf {x} _ {i}, \\mathbf {R} _ {i}, \\mathbf {S} _ {i}\\right), o _ {i}, \\mathbf {c} _ {i} \\right\\} _ {i = 1} ^ {N _ {g}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.483, + 0.538 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) denotes the position; \\(\\mathbf{RSS}^{\\top}\\mathbf{R}^{\\top}\\) is the anisotropic covariance parameterised as an ellipsoid with scale \\(\\mathbf{S}\\) and rotation \\(\\mathbf{R}\\); \\(o\\) is the opacity, and \\(\\mathbf{c}\\) is the colour of the Gaussian. In 3DGS [36], \\(\\mathcal{G}\\) is optimised through gradient-based training with multi-view image loss. Learning a 3DGS representation from monocular inputs requires substantial adjustments in different contexts [11, 81], and we show its utility in SfT for the first time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.704 + ], + "angle": 0, + "content": "NeuralClothSim [34] is a recent quasistatic cloth simulator representing surface deformations as a coordinate-based implicit neural deformation field (NDF). Given a target simulation scenario specified by the initial surface state, material properties and external forces, NeuralClothSim learns the equilibrium deformation field using the laws of the Kirchhoff-Love thin shell theory. Upon convergence, the equilibrium state can be queried continuously and consistently at multiple resolutions. These properties, combined with the memory adaptivity of neural fields, make NeuralClothSim well-suitable for inverse problems like ours." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Kairanda et al. [34] model a cloth quasistatically as an NDF \\(\\mathbf{u}(\\pmb {\\xi}):\\Omega \\mapsto \\mathbb{R}^3\\) defined on its curvilinear coordinate space \\(\\Omega\\). For a volumetric thin shell such as cloth, the Kirchhoff-Love model [78] offers a reduced kinematic parameterisation of the volume characterised by a 2D midsurface that fully determines the strain components throughout the thickness. Following thin shell assumptions, they decompose the Green strain due to deformation \\(\\mathbf{u}\\) parameterised by a multilayer perceptron (MLP)—into membrane strain \\(\\varepsilon\\) and bending strain \\(\\kappa\\), measuring the inplane stretching and the change in curvature, respectively. NeuralClothSim then computes the internal hyperelastic energy \\(\\Psi [\\varepsilon ,\\kappa ;\\Phi ]\\) as a functional of the geometric strains and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "the cloth's material properties \\(\\Phi\\). Under the action of external force \\(\\mathbf{f}\\), their neural solver finds the equilibrium solution following the principle of minimum potential energy, where the potential \\(\\int_{\\Omega} \\Psi d\\Omega - \\int_{\\Omega} \\mathbf{f} \\cdot \\mathbf{u} d\\Omega\\) is employed as the loss function. Note that the forward model of NeuralCloth-Sim [34] cannot be naively applied or extended (e.g., when coupled with a differentiable renderer, estimating physics parameters with the loss of NeuralClothSim would not converge due to the lack of lower bound) in our inverse setting." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.24, + 0.605, + 0.255 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.905, + 0.401 + ], + "angle": 0, + "content": "We propose Thin-Shell-SfT, a new template-based method for the fine-grained 3D reconstruction of a deforming surface (such as cloth, paper or metal) seen in a monocular RGB video; see Fig. 2. We aim to reconstruct continuous 3D surfaces \\(\\{\\mathbf{S}_t\\}_{t\\in [1,\\dots,T]}\\) corresponding to the input image sequence \\(\\{\\mathbf{I}_t\\}_t\\) and optional masks \\(\\{\\mathbf{M}_t\\}_t\\). Similar to previous surface tracking methods [33, 50], we assume a static camera with known calibration and take the surface \\(\\mathbf{S}_1\\) corresponding to the first frame \\(t = 1\\) as a template." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.905, + 0.612 + ], + "angle": 0, + "content": "In contrast to the discrete representations (e.g. points, meshes) used in previous monocular tracking approaches, our deformation model encodes the surface and its dynamics as continuous and adaptive neural fields (Sec. 4.1). We optimise the neural model by relating input monocular views to the estimated surface states using 3DGS [36], where the Gaussians are initialised and dynamically tracked with parameters induced from the surface deformation (Sec. 4.2). To ensure plausible deformations due to the inherent ambiguity of the monocular setting and in contrast to prior discrete regularisers, we model the surfaces as thin shells imposing the continuous Kirchhoff-Love physics constraints [34] on the neural field, allowing us to reconstruct fine-grained details (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.871, + 0.635 + ], + "angle": 0, + "content": "4.1. Deformation Model and Parameterisation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.905, + 0.779 + ], + "angle": 0, + "content": "Deformation Model. As our deformation model, we employ a continuous surface representation and deformation dynamics modelled as coordinated-based neural fields. Consider a surface with initial state \\(\\mathbf{S}_1\\subset \\mathbb{R}^3\\) and 2D parameterisation \\(\\Omega \\subset \\mathbb{R}^2\\). For any parametric point \\(\\xi \\coloneqq (\\xi^{1},\\xi^{2})\\in \\Omega\\), we represent its position on the initial surface state with a mapping \\(\\bar{\\mathbf{x}} (\\pmb {\\xi}): \\Omega \\to \\mathbf{S}_1\\). Furthermore, we encode the time-varying spatial location of \\(\\pmb{\\xi}\\) on the tracked surface \\(\\mathbf{S}_t\\) as \\(\\mathbf{x}(\\pmb {\\xi},t):\\Omega \\times [1,\\dots,T]\\rightarrow \\mathbf{S}_t\\), with" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.785, + 0.905, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {x} (\\boldsymbol {\\xi}, t) = \\bar {\\mathbf {x}} (\\boldsymbol {\\xi}) + \\mathbf {u} (\\boldsymbol {\\xi}, t), \\text {w i t h} \\mathbf {u} (\\boldsymbol {\\xi}, 1) = 0 \\text {a n d} \\tag {2} \\\\ \\mathbf {u} (\\boldsymbol {\\xi}, t) = \\lambda \\mathbf {u} (\\boldsymbol {\\xi}, t - 1) + \\mathcal {F} (\\boldsymbol {\\xi}, t), \\forall t > 1, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{u}(\\pmb {\\xi},t)\\) is the deformation field, \\(\\lambda >0\\) is a scalar value and \\(\\mathcal{F}(\\pmb {\\xi},t)\\) represents the estimated deformation offset. Here, we encourage the conservation of momentum by setting the deformation of the future surface states in the direction of the previous deformations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11375" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.38, + 0.907, + 0.466 + ], + "angle": 0, + "content": "Figure 2. Overview of Thin-Shell-SfT. Our deformation model encodes the surface and its dynamics as neural fields. Given the template \\(\\mathbf{S}_1\\), we first fit a reference field (NRF) from 2D parametric points \\(\\pmb{\\xi}\\) to the initial 3D positions \\(\\bar{\\mathbf{x}}\\). In the main stage, we optimise the deformation field (NDF) \\(\\mathbf{u}(\\pmb{\\xi}, t)\\) by relating estimated surface states \\(\\mathbf{S}_t / \\mathcal{G}_t\\) to the input monocular views. We induce the dynamically tracked Gaussians to the surface by: (1) Computing their positions \\(\\mathbf{x}\\) as the sum of the initial position \\(\\bar{\\mathbf{x}}\\) and NDF output \\(\\mathbf{u}\\), (2) Setting their rotations \\(\\bar{\\mathbf{a}}_i\\) as the template's local coordinate system, and (3) Fixing the normal scale \\(\\epsilon\\), and optimising the colour, opacity and tangential scales \\((s_1, s_2)\\) using only the template texture. For physical plausibility, we impose continuous Kirchhoff-Love physics constraints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.895 + ], + "angle": 0, + "content": "Parameterisation of the Template and Deformations. Next, we present an adaptive parameterisation of the above deformation model that allocates capacity to dynamic deformation details. Assuming that the template is provided as a mesh \\(\\mathcal{M} \\subset \\mathbf{S}_1\\), we generate the corresponding 2D parametric space \\(\\mathcal{T} \\subset \\Omega\\) using established techniques (e.g. conformal parameterisation [30]). To learn a continuous representation of the initial state, we fit an MLP, which we call neural reference field (NRF) \\(\\bar{\\mathbf{x}}(\\boldsymbol{\\xi}; \\Upsilon)\\) to the template mesh parameterisation \\(\\mathcal{T} \\mapsto \\mathcal{M}\\). For training NRF, we construct points in the parametric and vertex space with randomly sampled barycentric coordinates and train with \\(\\ell_1\\) geometry loss. Akin to NRF, we regress the spatio-temporal neural deformation field (NDF) \\(\\mathbf{u}(\\boldsymbol{\\xi}, t; \\Theta)\\) using another MLP predicting deformation offsets, \\(\\mathcal{F}(\\boldsymbol{\\xi}, t; \\Theta): \\Omega \\times [1, \\dots, T] \\to \\mathbb{R}^3\\). Before training, the initial estimate for tracked surface points \\(\\mathbf{x}(\\boldsymbol{\\xi}, t; \\Theta)\\) is the sum of noisy NDF deformations, \\(\\mathbf{u}(\\boldsymbol{\\xi}, t; \\Theta)\\) and the pre-trained NRF \\(\\bar{\\mathbf{x}}(\\boldsymbol{\\xi}; \\Upsilon)\\) as given by Eq. (2). Our key idea is to optimise the NDF weights \\(\\Theta\\) to ensure photometric consistency of the tracked surfaces with input monocular views while minimising the internal energy of the surface states modelled as a thin shell. The NRF pre-fit enables the coupling of Gaussians to the surface and computing surface metrics for the thin shell energy. We must represent high-frequency signals, such as fine folds and wrinkles on the tracked surface. Moreover, our physics loss requires the computation of higher-order derivatives. Hence, we use sine activation [66] in both MLPs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.624 + ], + "angle": 0, + "content": "Inference. At testing, \\(\\mathbf{x}(\\boldsymbol {\\xi},t;\\Theta)\\) provides continuous access to the reconstructed surface, where the NRF and NDF networks can be consistently queried at varied spatiotemporal resolutions. Moreover, our surface deformation model provides temporal correspondences. Meshing and texture mapping can be achieved in the parametric domain and then transferred to the tracked surface in 3D using Eq. (2). We next describe how to optimise the deformation field \\(\\mathbf{u}(\\boldsymbol {\\xi},t;\\Theta)\\) that ensures consistency with monocular images and the physical plausibility of reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.636, + 0.871, + 0.652 + ], + "angle": 0, + "content": "4.2. Surface Tracking with Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Recall that we have an NDF parameterised by \\(\\Theta\\) that outputs the deformation \\(\\mathbf{u}(\\boldsymbol {\\xi},t;\\Theta)\\) at any time step. We seek to optimise the NDF so that the tracked surfaces \\(\\{\\mathbf{S}_t\\}_{t}\\) (evaluated with Eq. (2)) generate images matching with input views \\(\\{\\mathbf{I}_t\\}_{t}\\). A straightforward approach to encourage 3D-to-2D consistency is to minimise the photometric loss with differentiable rendering. We choose Gaussian Splatting [36] as the differentiable renderer, as it seamlessly integrates with our per-point (Eq. (2)) continuous deformation model. Initialisation of Gaussians. To initialise a set of Gaussians, we sample \\(N_{g}\\) well-distributed points (e.g. Poisson disk sampling [7]) from the surface of the input template mesh \\(\\mathcal{M}\\). The sampled points include parametric coordinates \\(\\{\\pmb {\\xi}_i\\}_{i\\in [1,\\dots ,N_g]}\\), their corresponding positions \\(\\{\\bar{\\mathbf{x}}_i\\}_i\\), and colours \\(\\{\\mathbf{c}_i\\}_i\\). Next, we will motivate our approach to parameterising Gaussian rotations and scales. Unlike the orig" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11376" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.35 + ], + "angle": 0, + "content": "inal multi-view [36] or dynamic Gaussian methods [12, 81] that leverage multi-view cues [18], we have a single static camera and, consequently, a challenging monocular setup. Thus, learning the anisotropic covariance of 3D Gaussians from all input frames can lead to poor results due to the inherent monocular ambiguities between deformation, texture and appearance. In our datasets, the initial surface state is reasonably (but not exactly) flat and fully visible. With this assumption, we take the initial surface normal as a proxy for the viewing direction and then fix the scale along normal while allowing for optimisation of the scales along the initial surface tangents. More concretely, we set the rotation matrix of \\(i\\)-th Gaussian equal to the local basis vectors, i.e. \\(\\mathbf{R}_i \\equiv [\\bar{\\mathbf{a}}_1 \\bar{\\mathbf{a}}_2 \\bar{\\mathbf{a}}_3]_i\\) (see Sec. 4.3 for computation) of the initial state (NRF) \\(\\bar{\\mathbf{x}}(\\pmb{\\xi}_i; \\Upsilon)\\) and the scale \\(s_3\\) as small \\(\\epsilon\\) along surface normal \\(\\bar{\\mathbf{a}}_3\\). Finally, the dynamically tracked surface-induced Gaussian \\(\\{\\mathcal{G}_t\\}_{t}\\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.356, + 0.484, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {G} _ {t} = \\left\\{\\mathcal {N} \\left(\\mathbf {x} \\left(\\boldsymbol {\\xi} _ {i}, t; \\Upsilon , \\Theta\\right), \\left[ \\bar {\\mathbf {a}} _ {1} \\left(\\boldsymbol {\\xi} _ {i}; \\Upsilon\\right), \\bar {\\mathbf {a}} _ {2} \\left(\\boldsymbol {\\xi} _ {i}; \\Upsilon\\right), \\right. \\right. \\right. \\\\ \\left. \\bar {\\mathbf {a}} _ {3} (\\boldsymbol {\\xi} _ {i}; \\Upsilon) ], (s _ {1}, s _ {2}, \\epsilon) _ {i}\\right), o _ {i}, \\mathbf {c} _ {i} \\} _ {i = 1} ^ {N _ {g}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.447 + ], + "angle": 0, + "content": "following Eq. (1). Note that the deformed positions \\(\\mathbf{x}(\\pmb{\\xi}_i, t)\\) are computed using our deformation model (Eq. (2)) and all other Gaussian parameters are shared across surface states." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.523 + ], + "angle": 0, + "content": "Optimisation of NDF and Gaussian Parameters. For optimisation, we use the \\(\\ell_1\\) photometric loss similar to 3DGS [36]. If segmentation masks \\(\\{\\mathbf{M}_t\\}_t\\) are available, an additional silhouette loss is added to primarily speed up training. Thus, our data loss reads as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.531, + 0.482, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {d} \\left(s _ {1}, s _ {2}, o, \\mathbf {c}, \\Theta\\right) = \\mathcal {R} \\left(\\mathcal {G} _ {1}, \\mathbf {I} _ {1}; s _ {1}, s _ {2}, o, \\mathbf {c}\\right) + \\\\ \\sum_ {t = 2} ^ {T} \\mathcal {R} \\left(\\mathcal {G} _ {t}, \\mathbf {I} _ {t}; \\Theta\\right) + \\mathcal {R} \\left(\\tilde {\\mathcal {G}} _ {t}, \\mathbf {M} _ {t}; \\Theta\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}(x;\\phi)\\) is the Gaussian rasterisation loss with inputs \\(x\\) and optimisable parameters \\(\\phi\\), and \\(\\tilde{\\mathcal{G}}_t\\) are the dynamic Gaussians \\(\\mathcal{G}_t\\) with colour \\(\\mathbf{c}\\) set to mask foreground (e.g. white). Note that splating in the first frame updates shared Gaussian properties, whereas all other frames backpropagate to update the deformation field parameters \\(\\Theta\\). With the formulation in Eq. (2), our method propagates gradient information from future states to update 3D reconstructions of the earlier frames. This enables global space-time surface optimisation while respecting the causality of the surface dynamics. While temporal consistency is implicit in Eq. (2), other variants of temporal regularisation are possible (see Appendix B). In contrast to previous 3DGS or dynamic Gaussian methods [81], we note that it is important to optimise all frames in each iteration rather than random frame \\(t \\in [2,\\dots,T]\\), as it can otherwise lead to local minima (incorrect learning of folds). Global optimisation is preferred as our deformation model is a single global MLP for all frames, and it is computationally efficient over random frame selection due to the causality of deformations" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.906, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 3. Thin-shell physics prior is a spatial regulariser that minimises the hyperelastic strain energy density due to deformation w.r.t. the known template." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.907, + 0.409 + ], + "angle": 0, + "content": "(Eq. (2)). Moreover, we keep the number of Gaussians fixed (i.e., no adaptive density control) since Gaussian parameters are learned only from the first frame; it is not known beforehand how the deformation details will arise in future states." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.421, + 0.744, + 0.437 + ], + "angle": 0, + "content": "4.3. Thin Shell Physical Prior" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.807 + ], + "angle": 0, + "content": "If we optimise the NDF solely using the photometric loss, it could perfectly fit the input images but could result in physically implausible surfaces due to monocular ambiguity. Therefore, we formulate a physics loss applicable to thin shells inspired by the NeuralClothSim approach [34] using the Kirchhoff-Love theory [78]. Since we model the dynamically tracked surface with an MLP, NeuralClothSim enables physics-based prior directly on the continuous surface. However, unlike the forward model of NeuralCloth-Sim, the external forces (such as contacts and the wind) and the material properties generating the deformations are unknown in an inverse setting like ours. For this reason, we assign material parameters to values typical for the surface modelled (e.g., cloth and paper). In contrast to material, it is hard to set reasonable values for forces or boundary conditions as the space of external forces is too large. Nevertheless, intuitively, the image loss effectively takes the role of external force as it guides the motion and deformation of the tracked surface. Thus, we omit the potential energy due to external forces from the optimisation. Finally, with the strain evaluated from the deformation field and the assumed material, our physical prior aims to minimise the internal hyperelastic energy that captures the surface stretching, shearing and bending stiffness; see Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Strain Computation. To evaluate stretching and bending strain at each training iteration, we randomly sample \\(N_{p}\\) parametric points \\(\\{\\pmb{\\xi}_i\\}_{i=1}^{N_p}\\) from the template mesh \\(\\mathcal{M}\\), and perform differential geometry operations on the initial surface state (i.e., NRF) and the deformed state (i.e., NDF). We next present the computation of the local quantities at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11377" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.335 + ], + "angle": 0, + "content": "each point \\(\\xi_{i}\\) on the initial state using the pre-trained NRF \\(\\overline{\\mathbf{x}} (\\pmb {\\xi};\\Upsilon)\\). We use Greek letters for indexing quantities on the parameterised surface (e.g., \\(\\bar{\\mathbf{a}}_{\\alpha},\\alpha ,\\beta ,\\ldots = 1,2\\)). For notational clarity, we drop the input \\(\\xi_{i},t\\) and network weights \\(\\Upsilon ,\\Theta\\) in all the derived quantities (e.g., \\(\\bar{\\mathbf{a}}_1\\equiv \\bar{\\mathbf{a}}_1(\\pmb {\\xi}_i;\\Upsilon),\\varepsilon \\equiv \\varepsilon (\\pmb {\\xi}_i,t;\\Upsilon ,\\Theta))\\). In the first step, we extract the local covariant basis: \\(\\{\\bar{\\mathbf{a}}_{\\alpha}:= \\partial \\bar{\\mathbf{x}} /\\partial \\xi^{\\alpha},\\bar{\\mathbf{a}}_3:= \\bar{\\mathbf{a}}_1\\times \\bar{\\mathbf{a}}_2\\}\\), the set of two vectors tangential to the curvilinear coordinate lines \\(\\xi^{\\alpha}\\) and the local normal. For the Gaussian samples \\(\\{\\pmb {\\xi}_i\\}_{i = 1}^{N_g}\\), an identical computation is employed for setting the covariant basis as the rotation parameters (as in Sec. 4.2). Next, we evaluate the surface metric tensor \\(\\bar{a}_{\\alpha \\beta}\\) measuring the lengths and the curvature tensor \\(\\bar{b}_{\\alpha \\beta}\\) measuring the curvature of the midsurface. Akin to the covariant tensors (e.g. \\(\\bar{b}_{\\alpha \\beta}\\)), their contravariant (e.g. \\(\\bar{b}^{\\alpha \\beta}\\)) and mixed variants (e.g. \\(\\bar{b}_{\\alpha}^{\\beta}\\)) counterparts are extracted as well." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.425 + ], + "angle": 0, + "content": "With these initial state quantities, we next present the strain—due to deformation—as a function of the NDF \\(\\mathbf{u}(\\boldsymbol {\\xi};\\Theta)\\). The predicted \\(\\mathbf{u} = \\hat{u}_i\\mathbf{e}_i\\) in global Cartesian coordinate system is transformed to contravariant coordinate basis, \\(\\mathbf{u} = u_{\\alpha}\\bar{\\mathbf{a}}^{\\alpha} + u_{3}\\bar{\\mathbf{a}}^{3}\\) for strain calculation [34]. Given the deformation gradient \\(\\mathbf{u}_{,\\alpha}\\) derived from the NDF as" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.427, + 0.484, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {, \\alpha} = \\varphi_ {\\alpha \\lambda} \\bar {\\mathbf {a}} ^ {\\lambda} + \\varphi_ {\\alpha 3} \\bar {\\mathbf {a}} ^ {3}, \\text {w i t h} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.448, + 0.439, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {\\alpha \\lambda} := u _ {\\lambda} | _ {\\alpha} - \\bar {b} _ {\\alpha \\lambda} u _ {3} \\text {a n d} \\varphi_ {\\alpha 3} := u _ {3, \\alpha} + \\bar {b} _ {\\alpha} ^ {\\lambda} u _ {\\lambda},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.5 + ], + "angle": 0, + "content": "we evaluate the non-linear membrane strain \\(\\varepsilon = [\\varepsilon_{\\alpha \\beta}]\\) and bending strain \\(\\kappa = [\\kappa_{\\alpha \\beta}]\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.5, + 0.369, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\varepsilon_ {\\alpha \\beta} = \\frac {1}{2} (\\varphi_ {\\alpha \\beta} + \\varphi_ {\\beta \\alpha} + \\varphi_ {\\alpha \\lambda} \\varphi_ {\\beta} ^ {\\lambda} + \\varphi_ {\\alpha 3} \\varphi_ {\\beta 3}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.527, + 0.482, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\kappa_ {\\alpha \\beta} = - \\varphi_ {\\alpha 3} | _ {\\beta} - \\bar {b} _ {\\beta} ^ {\\lambda} \\varphi_ {\\alpha \\lambda} + \\varphi_ {3} ^ {\\lambda} \\left(\\varphi_ {\\alpha \\lambda} | _ {\\beta} + \\frac {1}{2} \\bar {b} _ {\\alpha \\beta} \\varphi_ {\\lambda 3} - \\bar {b} _ {\\beta \\lambda} \\varphi_ {\\alpha 3}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.708 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{a}}^i\\) denote the local contravariant basis and \\(\\bar{b}_{\\alpha \\beta}\\), the components of the curvature tensor. In Eq. (6), we use a vertical bar for covariant derivatives, lower comma notation for partial derivatives w.r.t. the curvilinear coordinates \\(\\xi^{\\alpha}\\) (e.g. \\(u_{\\lambda}|_{\\alpha}\\), and \\(\\mathbf{u}_{,\\alpha} = \\partial \\mathbf{u} / \\partial \\xi^{\\alpha}\\)), and Einstein summation convention of repeated indices for tensorial operations (e.g., \\(\\varphi_{\\alpha \\lambda}\\varphi_{\\beta}^{\\lambda} = \\varphi_{\\alpha 1}\\varphi_{\\beta}^{1} + \\varphi_{\\alpha 2}\\varphi_{\\beta}^{2}\\)). Notably, we compute all the aforementioned physical quantities with automatic differentiation. Please see Appendix A for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.483, + 0.83 + ], + "angle": 0, + "content": "NDF Optimisation. As material model, we use \\(\\Phi \\coloneqq \\{\\rho, h, E, \\nu\\}\\), with mass density \\(\\rho\\) and the surface thickness \\(h\\), and elastic coefficients: Young's modulus \\(E\\), and Poisson's ratio \\(\\nu\\). For simplicity and computational efficiency, we use a linear isotropic constitutive model relating strain to stress, thereby computing the in-plane stiffness \\(D\\), the bending stiffness \\(B\\) and the elastic tensor \\(\\mathbf{H} = [H^{\\alpha \\beta \\lambda \\delta}]\\) on the initial template, which read as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.835, + 0.482, + 0.873 + ], + "angle": 0, + "content": "\\[\nD := \\frac {E h}{1 - v ^ {2}}, B := \\frac {E h ^ {3}}{1 2 \\left(1 - v ^ {2}\\right)}, \\text {a n d} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.871, + 0.453, + 0.898 + ], + "angle": 0, + "content": "\\[\nH ^ {\\alpha \\beta \\lambda \\delta} := \\nu \\bar {a} ^ {\\alpha \\beta} \\bar {a} ^ {\\lambda \\delta} + \\frac {1}{2} (1 - \\nu) (\\bar {a} ^ {\\alpha \\lambda} \\bar {a} ^ {\\beta \\delta} + \\bar {a} ^ {\\alpha \\delta} \\bar {a} ^ {\\beta \\lambda}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "Finally, our physics loss over the tracked surface states reads as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.147, + 0.905, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal{L}_{p}(\\Theta) = \\frac{1}{2N_{p}T}\\sum_{i = 1}^{N_{p}}\\sum_{t = 2}^{T}\\Big(\\underbrace{D\\varepsilon^{\\top}(\\boldsymbol{\\xi}_{i},t;\\Theta)\\mathbf{H}(\\boldsymbol{\\xi}_{i})\\varepsilon(\\boldsymbol{\\xi}_{i},t;\\Theta)}_{\\text{stretching/shearing stiffness}} \\\\ + \\underbrace {B \\boldsymbol {\\kappa} ^ {\\top} \\left(\\boldsymbol {\\xi} _ {i} , t ; \\Theta\\right) \\mathbf {H} \\left(\\boldsymbol {\\xi} _ {i}\\right) \\boldsymbol {\\kappa} \\left(\\boldsymbol {\\xi} _ {i} , t ; \\Theta\\right)} _ {\\text {b e n d i n g s t i f f n e s s}}) \\sqrt {\\bar {a} \\left(\\boldsymbol {\\xi} _ {i}\\right)}, \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.907, + 0.515 + ], + "angle": 0, + "content": "where \\(\\sqrt{a} := |\\bar{\\mathbf{a}}_1 \\times \\bar{\\mathbf{a}}_2|\\). Note that the term inside \\(\\mathcal{L}_p\\) evaluates per point; therefore, we re-sample at each training iteration to explore the continuous surface domain. Finally, we solve for the optimal NDF weights \\(\\Theta^*\\) by minimising the objective function \\(\\mathcal{L} = \\lambda_d\\mathcal{L}_d + \\lambda_p\\mathcal{L}_p\\) with empirically determined loss weights \\(\\{\\lambda_d, \\lambda_p\\}\\). The optimisation can be interpreted as an inverse simulation, with data loss guiding deformation as an external force and physics loss modelling the surface's intrinsic behaviour. We use iterative gradient-based optimisation to that end [37]. We choose \\(N_p \\ll N_g\\) as the physics loss involves expensive higher-order derivative computations, whereas the Gaussian rasterisation with samples fixed over training iterations is rather efficient. With resampling at each iteration, the physics loss backpropagates to the continuous surface, offering adaptive, memory-efficient performance compared to mesh-based physics alternatives requiring high resolution." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.528, + 0.72, + 0.545 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.905 + ], + "angle": 0, + "content": "Implementation Details. We implement Thin-Shell-SfT using PyTorch [58] and optimise separate NRF and NDF networks for each sequence. The NRF and NDF, both, employ the SIREN [66] architecture using a frequency of \\(\\omega = 5\\) and \\(\\omega = 30\\), respectively. Both networks have five hidden layers and 256 units in each layer, and are optimised using Adam [37]. Each network is trained for \\(\\approx 2 \\cdot 10^{3}\\) iterations; first, the template is fitted with NRF, followed by surface tracking with NDF. While NRF training takes up to two minutes, the core part of the method, i.e. the NDF training, typically takes between 30 minutes to one hour until convergence on an NVIDIA A100 GPU. The Gaussian scales along the normal are set to \\(\\epsilon = 10^{-5}\\). In our experiments, the number of Gaussian samples is \\(N_{g} \\approx 90k\\), and the number of samples for physics loss is \\(N_{p} = 100\\). Concerning the physics loss, the linear elastic material properties of the surface are set as follows for all sequences: \\(E = 5000\\mathrm{Pa}, \\nu = 0.25\\), and \\(h = 1.2\\mathrm{mm}\\). Moreover, we set the image and physics loss weights to \\(\\lambda_{d} = 5\\), and \\(\\lambda_{p} = 1\\), whereas the temporal constraint scalar is set to \\(\\lambda = 0.4\\). For visualisation of the tracked continuous surface, we generate temporally coherent meshes with ball-pivot meshing [4] of the template Gaussian positions \\(\\{\\bar{\\mathbf{x}}_i\\}_{i \\in [1,\\dots,N_g]}\\). We refer to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11378" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.173 + ], + "angle": 0, + "content": "
Seq.S1S2S3S4S5S6S7S8S9Avg
N-NRSfM [65]8.2533.62104.677.0272.668.73129.438.0619.8154.69
D-NRSfM [53]17.144.464.4041.3726.9214.0212.499.915.2915.11
Shimada [61]19.6922.1833.5490.3092.7857.6249.2724.4553.1249.22
DDD [82]2.951.693.8025.7310.466.9715.647.6111.7710.87
Ngo [50]2.191.512.1715.9010.723.017.95*failfail5.92*
Stokko [67]6.13.912.514.511.715.16.910.18.69.93
φ-SfT [33]0.792.753.547.606.153.144.732.522.363.93
Ours1.170.552.4†/3.55.5†/5.78.692.513.82.273.003.3†/3.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.261 + ], + "angle": 0, + "content": "Table 1. We quantitatively compare Thin-Shell-SfT to the state of the art on the \\(\\phi\\)-SfT real dataset. The average Chamfer distance is multiplied by \\(10^4\\) for readability. “*” notes that Ngo et al. failed on the last few frames of S7, which we exclude from the error computation. “†” denotes that we report the numbers on the variant of temporal coherency constraint described in Appendix B." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.264, + 0.172, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.328, + 0.172, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.395, + 0.162, + 0.406 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.267, + 0.248, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.328, + 0.248, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.391, + 0.239, + 0.409 + ], + "angle": 0, + "content": "Ground-Truth Normals" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.268, + 0.323, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.328, + 0.323, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.391, + 0.311, + 0.409 + ], + "angle": 0, + "content": "Normal Map Ours" + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.268, + 0.4, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.328, + 0.4, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.391, + 0.386, + 0.409 + ], + "angle": 0, + "content": "Normal Map \\(\\phi\\) -SfT" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.268, + 0.476, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.328, + 0.477, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.391, + 0.469, + 0.409 + ], + "angle": 0, + "content": "Normal Error Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.412, + 0.483, + 0.455 + ], + "angle": 0, + "content": "Figure 4. Comparison of the reconstruction normal maps for ours and \\(\\phi\\)-SfT. We show the cosine normal consistency to ground truth on the right and the normal metrics in Tab. I (Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.49 + ], + "angle": 0, + "content": "Appendices B and D and Figs. II and III for studies on the effect of hyperparameters \\(\\omega, \\lambda\\) and \\(N_{g}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Datasets and Error Metrics. We evaluate on the benchmark \\(\\phi\\)-SfT dataset [33] consisting of nine RGB videos (and reference depths) of highly challenging cloth deformations. As is common with non-rigid reconstruction methods [33, 65], we rigidly align the reconstructions of all compared methods to the ground truth. This is achieved with Procrustes alignment [72] for the first frame, which is refined with rigid ICP [5] for subsequent frames." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.223, + 0.637 + ], + "angle": 0, + "content": "5.1. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.885 + ], + "angle": 0, + "content": "SfT and NRSfM. We compare Thin-Shell-SfT to different SfT methods, namely \\(\\phi\\)-SfT [33], Yu et al.'s Direct, Dense, Deformable (DDD) [82], Ngo et al.'s Ngo2015 [50] and Shimada et al.'s IsMo-GAN [61], and finally to Stotko et al.'s approach [67]. DDD is provided with the required hierarchy of coarse-to-fine templates, and Ngo2015 and Stotko et al. with the same template as in the \\(\\phi\\)-SfT dataset. Since the dataset is richly textured, the 2D point tracking methods would perform well; therefore, we additionally compare to the NRSfM methods. In particular, we compare to Sidhu et al.'s Neural NRSfM (N-NRSfM) [65] and Parashar et al.'s Diff-NRSfM [53]. The 2D point correspondences required by the NRSfM methods are provided by densely tracking the 2D points across input images with multi-frame subspace flow (MFSF) [19, 20]. The first frame of the sequence is selected as a keyframe for 2D tracking." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Fig. 5-(left), we show qualitative reconstructions of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Thin-Shell-SfT compared to the prior state of the art (SotA) [33, 53, 67]. We capture fine wrinkles and folds in the deforming sequences that were not addressed by the earlier works. Next, we reconstruct surfaces with severe self-occlusions due to multiple layered folds in the extended versions of the \\(\\phi\\)-SfT sequences [33]. See Fig. 5-(right), where our physics-based model provides a reasonable prior for occluded regions; \\(\\phi\\)-SfT fails here due to prohibitive memory requirements. In Tab. 1, we show a quantitative comparison where we report the Chamfer distance (with pseudoground-truth point clouds) against the state-of-the-art methods over all the \\(\\phi\\)-SfT dataset sequences. We outperform the existing methods on most sequences, often substantially and on average over all sequences. While the default temporal coherency works best on the evaluated dataset, we notice qualitative improvements for two out of the nine sequences with the other variants. In Fig. 4, we further show comparisons between the normal maps of the tracked surface. Additional numerical comparisons with the previous SotA [33] on runtime and normal consistency are in Appendix C and Tab. I. We obtain a lower \\(0.009(\\ell_2)\\) and \\(0.034\\) (cosine) normal error, whereas \\(\\phi\\)-SfT shows \\(0.013\\) and \\(0.041\\), respectively. One of the primary reasons for the failure of \\(\\phi\\)-SfT [33] is its limitation to low-resolution meshes (\\(\\approx 300\\) vertices) capturing just coarse deformations. Our NDF offers, similar to scene representation approaches [47, 75], a smooth (cf. global MLP), low-dimensional (cf. a fixed parameter count) deformation space that is adaptive to dynamic details and can represent high-frequency signals (cf. sine activation). Although we discretise the physics loss, we re-sample at each iteration, leading to an adaptive discretisation. In contrast, earlier physics-based methods [33, 67] use fixed discretisation, either due to differentiable simulation not supporting remeshing [33] or a fixed-resolution surrogate model [67]. Our continuous formulation enables unprecedented fine-grained results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.909, + 0.789 + ], + "angle": 0, + "content": "Dynamic View Synthesis. Next, we compare with state-of-the-art dynamic view synthesis methods, particularly K-Planes [15] and Deformable Gaussians [81]. For the latter, we initialise the Gaussians with points sampled from the input template instead of the default SfM points. The compared radiance field methods are very effective in novel-view synthesis for monocular videos that provide multiview cues [18]. However, they fail to recover spatiotemporally coherent surface geometry for monocular RGB videos captured with a single static camera; see Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.676, + 0.815 + ], + "angle": 0, + "content": "5.2. Ablative Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.905, + 0.885 + ], + "angle": 0, + "content": "Next, we test the following aspects of our method: contribution of the physics loss, global optimisation and surface-induced Gaussians. Additional ablations, including momentum and mask loss, are presented in the Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Physics Prior. While relying solely on image loss yields" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11379" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.907, + 0.302 + ], + "angle": 0, + "content": "Figure 5. Exampary 3D reconstructions. (Left:) Comparisons focusing on high-frequency wrinkles. Thin-Shell-SfT captures the wrinkles best among all compared methods in one of the most challenging examples, outperforming \\(\\phi\\)-SfT [33], Stotko et al. [67] and Diff-NRSfM [53]. (Right:) Our results on the extended \\(\\phi\\)-SfT dataset highlight the excellent tracking in the occluded regions." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.304, + 0.482, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.414, + 0.483, + 0.469 + ], + "angle": 0, + "content": "Figure 6. We compare our reconstructions to dynamic view synthesis methods [15, 81] on the rendered depth maps. The texture details retained in the depth maps imply that the compared methods fail to learn accurate surface (mistake texture for geometry)." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.474, + 0.483, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Figure 7. Ablations. (a:) Without physics, the surface can severely stretch or shrink. (b:) Random frame optimisation leads to local minima, (c,d:) Not positioning and orienting Gaussians on the surface leads to distortions and poor reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.793 + ], + "angle": 0, + "content": "accurate RGB recovery during rendering, it can lead to severely distorted surface tracking. This arises from the inherent monocular ambiguity of our setting as multiple possible deformations in 3D correspond to the same 2D image. As seen in Fig. 7-(a), incorporating Kirchhoff-Love-based thin shell prior is essential for achieving physically plausible and accurate surface reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Global Optimisation. We test a version of our method employing optimisation with randomly selected frames at each iteration instead of the proposed joint optimisation over all frames. Fig. 7-(b) visualises the ablated results showing poor performance for folds and wrinkles. Without any multi-view cues, as in the case of the used \\(\\phi\\)-SfT dataset, random frame optimisation leads to local minima." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Surface-induced Gaussians. The specially-tailored initialisation and optimisation of Gaussian parameters (Eqs. (3) and (4)) are crucial for the accurate geometric reconstruction. In the ablated version of our method, we optimise the shared Gaussian parameters, i.e., covariance, opacity, and colour on all input frames (similar to earlier methods [36, 81]) instead of the single template frame as in Eq. (4). Due to the ambiguities between geometric details and appearance over the deforming sequence, this leads to wrongly reconstructed Gaussians, eventually resulting in poor surface reconstructions; see Fig. 7-(c). In addition, fixing the normal scale on the template frame is useful for preventing elongated Gaussians (Fig. 7-(d))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.591 + ], + "angle": 0, + "content": "In Tab. II-appendix, we report the above results on the full \\(\\phi\\)-SfT dataset. On average, we obtain Chamfer distances of 34.25, 14.0, 3.75, and 3.46 for no physics, no surface-induced Gaussians, no normal scale, and the full model, respectively. We notice that including physics and surface induction are crucial for accurate tracking." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.634, + 0.62 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.871 + ], + "angle": 0, + "content": "We present Thin-Shell-SfT, a new approach for dense monocular non-rigid 3D surface tracking relying on new principles, i.e. adaptive neural deformation field, continuous Kirchhoff-Love thin shell prior and surface-induced 3D Gaussian Splating. Experiments on the \\(\\phi\\)-SfT dataset demonstrate that our method accurately reconstructs the challenging fine-grained deformations such as cloth folds and wrinkles. Overall, we significantly improve fine-grained surface tracking using an adaptive deformation model and continuous thin shell physics compared to existing approaches that only support coarse deformations. Thanks to the physics prior, Thin-Shell-SfT is reasonably robust to occlusions, although extreme self-collisions remain challenging. Similarly, tracking textureless surfaces is another possible research direction that would necessitate special handling in future." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was partially supported by the VIA Research Center." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11380" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Nonrigid structure from motion in trajectory space. In Advances in Neural Information Processing Systems (NeurIPS), 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Mohammad Dawud Ansari, Vladislav Golyanik, and Didier Stricker. Scalable dense monocular surface reconstruction. In International Conference on 3D Vision (3DV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[3] Adrien Bartoli, Yan Gérard, François Chadebecq, Toby Collins, and Daniel Pizarro. Shape-from-template. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[4] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE TVCG, 1999. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[6] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Computer Vision and Pattern Recognition (CVPR), 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[7] Robert Bridson. Fastoisson disk sampling in arbitrary dimensions. SIGGRAPH sketches, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.483, + 0.494 + ], + "angle": 0, + "content": "[8] David Casillas-Perez, Daniel Pizarro, David Fuentes-Jimenez, Manuel Mazo, and Adrien Bartoli. The isowarp: the template-based visual geometry of isometric surfaces. International Journal of Computer Vision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[9] Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne Vouga, and Christoph Lassner. Virtual elastic objects. In Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.552, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[10] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. In Computer Vision and Pattern Recognition (CVPR), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.608, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[11] Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.664, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[12] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[13] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splating: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[14] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Gaussian splashing: Unified particles for versatile motion synthesis and rendering. Computer Vision and Pattern Recognition (CVPR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[16] David Fuentes-Jimenez, Daniel Pizarro, David Casillas-Perez, Toby Collins, and Adrien Bartoli. Texture-generic deep shape-from-template. IEEE Access, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[17] David Fuentes-Jimenez, Daniel Pizarro, David Casillas-Pérez, Toby Collins, and Adrien Bartoli. Deep shape-from-template: Single-image quasi-isometric deformable registration and reconstruction. Image and Vision Computing, 127: 104531, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[18] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[19] Ravi Garg, Anastasios Roussos, and Lourdes Agapito. A variational approach to video registration with subspace constraints. International Journal of Computer Vision (IJCV), 104(3):286-314, 2013. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[20] Ravi Garg, Anastasios Roussos, and Lourdes Agapito. Source code of Multi-Frame Subspace Flow (MFSF). http://www0.cs.ucl.ac.uk/staff/lagapito/subspace_flow/, 2015.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[21] Vladislav Golyanik, Soshi Shimada, Kiran Varanasi, and Didier Stricker. Hdm-net: Monocular non-rigid 3d reconstruction with learned deformation model. In EuroVR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[22] Vladislav Golyanik, André Jonas, and Didier Stricker. Consolidating segmentwise non-rigid structure from motion. In Machine Vision Applications (MVA), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[23] Stella Graßhof and Sami Sebastian Brandt. Tensor-based non-rigid structure from motion. In Winter Conference on Applications of Computer Vision (WACV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[24] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[25] Jingfan Guo, Jie Li, Rahul Narain, and Hyun Soo Park. Inverse simulation: Reconstructing dynamic geometry of clothed humans via optimal control. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[26] Marc Habermann, Weipeng Xu, Helge Rhodin, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Nrst: Non-rigid surface tracking from monocular video. German Conference on Pattern Recognition (GCPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[27] Nazim Haouchine and Stephane Cotin. Template-based monocular 3d recovery of elastic shapes using lagrangian multipliers. In Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[28] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[29] Florian Hofherr, Lukas Koestler, Florian Bernard, and Daniel Cremers. Neural implicit representations for physical parameter inference from a single video. In Winter Conference on Applications of Computer Vision (WACV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Alec Jacobson, Daniele Panozzo, et al. libigl: A simple \\(\\mathbf{C}++\\) geometry processing library, 2018. https://libigl.github.io/. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11381" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Miguel Jaques, Michael Burke, and Timothy Hospedales. Physics-as-inverse-graphics: Unsupervised physical parameter estimation from video. In International Conference on Learning Representations (ICLR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.216 + ], + "angle": 0, + "content": "[32] Erik Johnson, Marc Habermann, Soshi Shimada, Vladislav Golyanik, and Christian Theobalt. Unbiased 4d: Monocular 4d reconstruction with a neural deformation model. In Computer Vision and Pattern Recognition (CVPR) Workshops, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.285 + ], + "angle": 0, + "content": "[33] Navami Kairanda, Edith Tretschk, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. \\(\\phi\\)-SfT: Shape-from-template with a physics-based deformation model. In Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[34] Navami Kairanda, Marc Habermann, Christian Theobalt, and Vladislav Golyanik. Neuralclothsim: Neural deformation fields meet the thin shell theory. Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[35] Rama Kandukuri, Jan Achterhold, Michael Moeller, and Joerg Stueckler. Learning to identify physical parameters from video using differentiable physics. In German Conference for Pattern Recognition (GCPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[37] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv e-prints, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[38] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[39] Suryansh Kumar. Jumping manifolds: Geometry aware dense non-rigid structure from motion. In Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[40] Suryansh Kumar, Anoop Cherian, Yuchao Dai, and Hongdong Li. Scalable dense non-rigid structure-from-motion: A grassmannian perspective. In Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[41] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep Physics-aware Inference of Cloth Deformation for Monocular Human Performance Capture. In International Conference on 3D Vision (3DV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[42] Yifei Li, Tao Du, Kui Wu, Jie Xu, and Wojciech Matusik. Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[43] Junbang Liang, Ming Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse problems. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[44] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Alberta Longhini, Marcel Busching, Bardienus Pieter Duisterhof, Jens Lundell, Jeffrey Ichnowski, Marten Björkman," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "and Danica Kragic. Cloth-splatting: 3d cloth state estimation from rgb supervision. In 8th Annual Conference on Robot Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[46] Augustus Edward Hough Love. A treatise on the mathematical theory of elasticity. University press, 1927. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[49] J. Krishna Murthy, Miles Macklin, Florian Golemo, Vikram Voleti, Linda Petrini, Martin Weiss, Breandan Considine, Jérôme Parent-Lévesque, Kevin Xie, Kenny Erleben, Liam Paull, Florian Shkurti, Derek Nowrouzezahrai, and Sanja Fidler. gradsim: Differentiable simulation for system identification and visuomotor control. In International Conference on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[50] Dat Tien Ngo, Sanghyuk Park, Anne Jorstad, Alberto Crivellaro, Chang D. Yoo, and Pascal Fua. Dense image registration and deformable surface reconstruction in presence of occlusions and minimal texture. In International Conference on Computer Vision (ICCV), 2015. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[51] Marco Paladini, Alessio Del Bue, Marko Stosic, Marija Dodig, Jo ao Xavier, and Lourdes Agapito. Factorization for non-rigid and articulated structure using metric projections. In Computer Vision and Pattern Recognition (CVPR), 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[52] Shaifali Parashar, Daniel Pizarro, Adrien Bartoli, and Toby Collins. As-rigid-as-possible volumetric shape-from-template. In International Conference on Computer Vision (ICCV), 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[53] Shaifali Parashar, Mathieu Salzmann, and Pascal Fua. Local non-rigid structure-from-motion from diffeomorphic mappings. In Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[54] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40 (6):1-12, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[55] Mathieu Perriollat, Richard Hartley, and Adrien Bartoli. Monocular template-based reconstruction of inextensible surfaces. International Journal of Computer Vision (IJCV), 95(2):124-137, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[56] Albert Pumarola, Antonio Agudo, Lorenzo Porzi, Alberto Sanfeliu, Vincent Lepetit, and Francesc Moreno-Noguer. Geometry-Aware Network for Non-Rigid Shape Prediction from a Single View. In Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[57] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11382" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[58] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[59] Davis Rempe, Leonidas J. Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jimei Yang. Contact and human dynamics from monocular video. In European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[60] Mathieu Salzmann, Julien Pilet, Slobodan Ilic, and Pascal Fua. Surface deformation models for nonrigid 3d shape recovery. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 29(8):1481-1487, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[61] Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Didier Stricker. IsMo-GAN: Adversarial learning for monocular non-rigid 3d reconstruction. In Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[62] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics, 39(6), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[63] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, Patrick Pérez, and Christian Theobalt. Neural monocular 3d human motion capture with physical awareness. ACM Transactions on Graphics, 40(4), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[64] Soshi Shimada, Vladislav Golyanik, Patrick Pérez, and Christian Theobalt. Decaf: Monocular deformation capture for face and hand interactions. ACM Transactions on Graphics, 42(6), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[65] Vikramjit Sidhu, Edgar Tretschk, Vladislav Golyanik, Antonio Agudo, and Christian Theobalt. Neural dense non-rigid structure from motion with latent space constraints. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[66] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 4, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[67] David Stotko, Nils Wandel, and Reinhard Klein. Physics-guided shape-from-template: Monocular video perception through neural surrogate models. Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[68] Guoxing Sun, Rishabh Dabral, Pascal Fua, Christian Theobalt, and Marc Habermann. Metacap: Meta-learning priors from multi-view imagery for sparse-view human performance capture and rendering. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[69] Lorenzo Torresani, Aaron Hertzmann, and Chris Bregler. Nonrigid structure-from-motion: Estimating shape and motion with hierarchical priors. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 30(5), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[70] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[71] Edith Tretschk, Navami Kairanda, Mallikarjun BR, Rishabh Dabral, Adam Kortylewski, Bernhard Egger, Marc Habermann, Pascal Fua, Christian Theobalt, and Vladislav Golyanik. State of the art in dense monocular non-rigid 3d reconstruction. In Computer Graphics Forum (Eurographics State of the Art Reports), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.177, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[72] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 1991. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[73] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. Games: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.274, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[74] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[75] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[76] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[77] Sebastian Weiss, Robert Maier, Daniel Cremers, Rüdiger Westermann, and Nils Thuerey. Correspondence-free material reconstruction using sparse surface constraints. Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.497, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[78] Gerald Kempner and Demosthenes Talaslidis. Mechanics of solids and shells. CRC, Boca Raton, 2003. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.525, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[79] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.581, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[80] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.637, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[81] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[82] Rui Yu, Chris Russell, Neill D. F. Campbell, and Lourdes Agapito. Direct, dense, and deformable: Template-based non-rigid 3d reconstruction from rgb video. In International Conference on Computer Vision (ICCV), 2015. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[83] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y. Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. PhysDreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[84] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, and Gordon Wetzstein. Physavatar: Learning the physics of dressed 3d avatars from visual observations. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11383" + } + ] +] \ No newline at end of file diff --git a/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_origin.pdf b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a7121382616602ccd4bba866d2c02175cbde16a5 --- /dev/null +++ b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/513fc02c-a19b-4f3d-a5da-fa87c09303bb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa7ba8a11b3cba82e9bd4ee317b742e720573d789175c94dbbbc86a9450e8f74 +size 3399161 diff --git a/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/full.md b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..265df7ca8ca717128eb8e162daff436569b30f9d --- /dev/null +++ b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/full.md @@ -0,0 +1,331 @@ +# Thin-Shell-SfT: Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields + +Navami Kairanda $^{1}$ Marc Habermann $^{1,2}$ Shanthika Naik $^{3}$ Christian Theobalt $^{1,2}$ Vladislav Golyanik $^{1}$ $^{1}$ MPI for Informatics, SIC ${}^{2}$ VIA Research Center ${}^{3}$ IIT Jodhpur + +# Abstract + +3D reconstruction of highly deformable surfaces (e.g. cloths) from monocular RGB videos is a challenging problem, and no solution provides a consistent and accurate recovery of fine-grained surface details. To account for the ill-posed nature of the setting, existing methods use deformation models with statistical, neural, or physical priors. They also predominantly rely on nonadaptive discrete surface representations (e.g. polygonal meshes), perform frame-by-frame optimisation leading to error propagation, and suffer from poor gradients of the mesh-based differentiable renderers. Consequently, fine surface details such as cloth wrinkles are often not recovered with the desired accuracy. In response to these limitations, we propose ThinShell-SfT, a new method for non-rigid 3D tracking that represents a surface as an implicit and continuous spatiotemporal neural field. We incorporate continuous thin shell physics prior based on the Kirchhoff-Love model for spatial regularisation, which starkly contrasts the discretised alternatives of earlier works. Lastly, we leverage 3D Gaussian splatting to differentiably render the surface into image space and optimise the deformations based on analysis-by-synthesis principles. Our Thin-Shell-SfT outperforms prior works qualitatively and quantitatively thanks to our continuous surface formulation in conjunction with a specially tailored simulation prior and surface-induced 3D Gaussians. See our project page at https://4dqv.mpi-inf.mpg.de/ThinShellSfT. + +# 1. Introduction + +Non-rigid 3D reconstruction and tracking of general deformable surfaces from a monocular RGB camera is an important, challenging and ill-posed problem that is far from being solved [71]. It has applications in game development, robotics and augmented reality, to name a few areas. + +Prior works on temporally-coherent general surface reconstruction can be grouped into Shape-from-Template (SfT) [3, 8, 33] and Non-Rigid Structure-from-Motion [39, + +![](images/50eace23936289e74a9e49973a93b3e7db24dcaf23c830454c35b1f9ab79e834.jpg) +Figure 1. Our Thin-Shell-SfT approach reconstructs high-fidelity deformable 3D surface geometry with fine-grained wrinkles from a monocular video, while the previous best method [33] struggles. The coloured tracks ("Ours, Input View") visualise the 3D Gaussian trajectories over time and across all input video frames. + +![](images/5a76305e924ba4058082f0411934b7a6a215de218ebd3aa7d46894bba644baf3.jpg) + +![](images/e6fd630602f3b6da00bfe490470051ae2a48a35a4a187cf363bd4c1adc1d8017.jpg) + +![](images/68722334cdc192f9dd2e6b40eb6f5ec7dd5442d342b3ef76fdf0ef837aaa21f6.jpg) + +![](images/6075cd515a0aea28f6efbbd407d516bf197f9f0693a801471cb82fe398a94c7b.jpg) + +53, 65]; they rely on 2D point tracks across monocular images (NRSfM) [53, 65] or between the input image and the template (SfT) [3, 8]. Recent physics-based SfT approaches [33, 67] demonstrate state-of-the-art results and cause a paradigm shift from geometric- [3, 8, 52] to physics-based constraints and from 2D-point-based [8, 50] to dense photometric loss using differentiable renderers. However, even such approaches as $\phi$ -SfT [33] do not support fine wrinkles and require multiple hours to reconstruct a limited number of frames ( $\approx 50$ ) of a single object, which is due to the underlying surface mesh representation, i.e., an explicit and discrete. First, determining the mesh resolution for a specific scene is difficult, e.g. selecting a high resolution could prohibitively increase the memory and computational cost, while a lower resolution might not account for fine-scale deformations. Second, the Finite Element Mesh (FEM)-based differentiable physics simulators [42, 43] lead to inconsistent simulations at different resolutions, preventing the adoption of coarse-to-fine strategies for tracking. Third, simulator-based approaches [33, 67] optimise frame-by-frame and can get stuck in local minima. Moreover, mesh-based differentiable renderers used in recent approaches [33, 67] are nonadaptive and do not support the complex remeshing required for the dynamic details arising in deformable scenes. + +To overcome these limitations, our key idea is to replace meshes with an adaptive and continuous surface and deformation representation1. We tightly couple it with differentiable physics to guide the deformations while ensuring photometric consistency with the monocular images through analysis-by-synthesis. We model surfaces as Kirchhoff-Love thin shells [46] and propose a physics-based continuous deformation prior. Our regulariser minimises the internal hyperelastic energy of the tracked surface, ensuring its physical plausibility and providing prior for occlusion handling. Unlike discrete priors operating on mesh vertices [33, 82], our continuous prior updates any point on the surface, enabling physics supervision for fine-grained details such as wrinkles. While similar modelling has been applied to cloth simulation [34], it has not been shown for inverse problems like monocular surface tracking, which involves many unknowns (e.g., material, forces, contacts) and ambiguities between them. In addition to continuous prior, we perform joint space-time optimisation while taking the causality of deformation into account to impose temporal coherence (i.e. the current deformed state can update previous state parameters but not the future ones). As the differentiable renderer, we employ 3D Gaussian Splatting [36], which recently emerged as a prominent technique for radiance field rendering. It integrates straightforwardly with our continuous per-point deformation model and offers high-quality image gradients due to the continuous volumetric radiance field formulation. We leverage differential geometric quantities from the thin shell physics to couple Gaussians to the surface and optimise the parameters of dynamically tracked Gaussians. While there are extensions of Gaussian splatting focussing on dynamic view synthesis [12, 81] or multi-view static surface reconstruction [24, 73], we show how to adapt it for surface tracking from monocular videos with de-facto absent multi-view 3D reconstruction cues [18] such as the $\phi$ -SfT [33] dataset. The technical contributions of this paper are as follows: + +- Thin-Shell-SfT, i.e. a new method for monocular non-rigid 3D surface tracking operating on a continuous adaptive spatiotemporal representation; +- A continuous deformation prior based on the principles of the Kirchhoff-Love thin shell theory and application of such a prior in an inverse problem (shape from template); +- Adaptation of 3D Gaussian Splating for 3D tracking of highly deformable dynamic surfaces forming folds and fine wrinkles, captured by a static monocular camera. + +Our experimental results on the challenging $\phi$ -SfT benchmark [33] show a significant improvement over the state of the art in terms of reconstruction accuracy; see Fig. 1. + +# 2. Related Work + +We review methods for monocular reconstruction of general non-rigid surfaces. They differ in their assumptions about the deformation model, data terms, and the priors; we refer to a recent survey by Tretschk et al. [71] for an in-depth review. Similar to ours, many recent works integrate Gaussians with physics or clothing but in very different contexts such as simulation [14, 79], material estimation [44, 83], and tracking from multi-view video [13, 45, 84]. + +Shape-from-Template (SfT) methods [16, 27, 33, 50, 55, 56, 60, 82] assume a single static shape or template as a prior. A template often corresponds to the first frame of the sequence [3, 33, 50] and, in other cases, to the rigid initialisation [17, 61]. SfT approaches [3, 26, 50, 82] employ a 3D-2D reprojection constraint and deform the template using temporal and geometric soft constraints to encourage physical plausibility. On the other hand, learning-based SfTs [16, 21, 56, 61] encode the template in the network and regress the surface deformations. They are sometimes object-specific with template encoded in the neural network weights [17] and often object-generic [16, 22, 61] when supervised with synthetic datasets. Recently, Kairanda et al. [33] and Stoko et al. [67] explain 2D observations through physics-based simulation of the deformation process and employ differentiable rendering for per-sequence gradient-based optimisation. In contrast to traditional SfT [3, 8, 52], they do not require registrations (the correspondence between the 3D template and the image). Ours falls under this category as we require the template corresponding to the first frame. In contrast to the prior works, ours is the first method to continuously represent the surface and its dynamics with a neural field capable of representing high-frequency signals [66]. We supervise with thin shell physics constraint applied at continuous points in the domain and employ the Gaussian Splatting functionality [36]. + +Non-Rigid Structure from Motion (NRSfM) [1, 6, 10, 51, 69] relies on motion cues of 2D point tracks over the input monocular images and outputs per-frame camera-object poses and the 3D shapes. Recent dense NRSfM methods [2, 22, 23, 40, 53, 65, 74] rely on per-pixel multi-frame optical flow or video registration [19]. In contrast, we do not make a restrictive assumption of available 2D point tracks and track highly challenging cloth deformations (fine wrinkles), which goes beyond NRSfM capabilities. + +Dynamic Novel View Synthesis. Radiance field methods [36, 47, 48, 75] learn volumetric scene representation for high-quality novel-view synthesis from multi-view images, which are extended to support dynamic scenes [11, 15, 32, 38, 54, 70, 76, 81]. Though they demonstrate impressive results, most do not show or evaluate geometry reconstruction. Moreover, they require multi-view cues [18, 56]; in contrast, we focus on dynamic scenes captured from a single static camera without such cues. + +Physics-based Priors have been applied in inverse problems [25, 41, 59, 62]. The examples range from 3D human pose estimation [41, 63] and parameter estimation [29, 43, 49, 77] to dense SfT [33, 64, 67]. Recent works [9, 57] extend neural representations to physical parameter inference from videos. Some methods [31, 35, 80] combine differentiable simulation and rendering. In the inverse setting, we are the first to impose continuous physics constraints while prior methods use mesh-based simulators. + +# 3. Preliminaries + +We review 3D Gaussian Splatting [36] and NeuralCloth-Sim [34], from which we take inspiration for the data term (ensuring consistency of reconstructions with input images) and the prior term (encouraging physical plausibility). + +3D Gaussian Splatting (3DGS) models a scene as a volumetric radiance field with a dense set of 3D Gaussians, each defined by its position (mean), anisotropic covariance, opacity, and colour. $N_{g}$ Gaussians are represented as + +$$ +\mathcal {G} = \left\{\mathcal {N} \left(\mathbf {x} _ {i}, \mathbf {R} _ {i}, \mathbf {S} _ {i}\right), o _ {i}, \mathbf {c} _ {i} \right\} _ {i = 1} ^ {N _ {g}}, \tag {1} +$$ + +where $\mathbf{x}$ denotes the position; $\mathbf{RSS}^{\top}\mathbf{R}^{\top}$ is the anisotropic covariance parameterised as an ellipsoid with scale $\mathbf{S}$ and rotation $\mathbf{R}$ ; $o$ is the opacity, and $\mathbf{c}$ is the colour of the Gaussian. In 3DGS [36], $\mathcal{G}$ is optimised through gradient-based training with multi-view image loss. Learning a 3DGS representation from monocular inputs requires substantial adjustments in different contexts [11, 81], and we show its utility in SfT for the first time. + +NeuralClothSim [34] is a recent quasistatic cloth simulator representing surface deformations as a coordinate-based implicit neural deformation field (NDF). Given a target simulation scenario specified by the initial surface state, material properties and external forces, NeuralClothSim learns the equilibrium deformation field using the laws of the Kirchhoff-Love thin shell theory. Upon convergence, the equilibrium state can be queried continuously and consistently at multiple resolutions. These properties, combined with the memory adaptivity of neural fields, make NeuralClothSim well-suitable for inverse problems like ours. + +Kairanda et al. [34] model a cloth quasistatically as an NDF $\mathbf{u}(\pmb {\xi}):\Omega \mapsto \mathbb{R}^3$ defined on its curvilinear coordinate space $\Omega$ . For a volumetric thin shell such as cloth, the Kirchhoff-Love model [78] offers a reduced kinematic parameterisation of the volume characterised by a 2D midsurface that fully determines the strain components throughout the thickness. Following thin shell assumptions, they decompose the Green strain due to deformation $\mathbf{u}$ parameterised by a multilayer perceptron (MLP)—into membrane strain $\varepsilon$ and bending strain $\kappa$ , measuring the inplane stretching and the change in curvature, respectively. NeuralClothSim then computes the internal hyperelastic energy $\Psi [\varepsilon ,\kappa ;\Phi ]$ as a functional of the geometric strains and + +the cloth's material properties $\Phi$ . Under the action of external force $\mathbf{f}$ , their neural solver finds the equilibrium solution following the principle of minimum potential energy, where the potential $\int_{\Omega} \Psi d\Omega - \int_{\Omega} \mathbf{f} \cdot \mathbf{u} d\Omega$ is employed as the loss function. Note that the forward model of NeuralCloth-Sim [34] cannot be naively applied or extended (e.g., when coupled with a differentiable renderer, estimating physics parameters with the loss of NeuralClothSim would not converge due to the lack of lower bound) in our inverse setting. + +# 4. Method + +We propose Thin-Shell-SfT, a new template-based method for the fine-grained 3D reconstruction of a deforming surface (such as cloth, paper or metal) seen in a monocular RGB video; see Fig. 2. We aim to reconstruct continuous 3D surfaces $\{\mathbf{S}_t\}_{t\in [1,\dots,T]}$ corresponding to the input image sequence $\{\mathbf{I}_t\}_t$ and optional masks $\{\mathbf{M}_t\}_t$ . Similar to previous surface tracking methods [33, 50], we assume a static camera with known calibration and take the surface $\mathbf{S}_1$ corresponding to the first frame $t = 1$ as a template. + +In contrast to the discrete representations (e.g. points, meshes) used in previous monocular tracking approaches, our deformation model encodes the surface and its dynamics as continuous and adaptive neural fields (Sec. 4.1). We optimise the neural model by relating input monocular views to the estimated surface states using 3DGS [36], where the Gaussians are initialised and dynamically tracked with parameters induced from the surface deformation (Sec. 4.2). To ensure plausible deformations due to the inherent ambiguity of the monocular setting and in contrast to prior discrete regularisers, we model the surfaces as thin shells imposing the continuous Kirchhoff-Love physics constraints [34] on the neural field, allowing us to reconstruct fine-grained details (Sec. 4.3). + +# 4.1. Deformation Model and Parameterisation + +Deformation Model. As our deformation model, we employ a continuous surface representation and deformation dynamics modelled as coordinated-based neural fields. Consider a surface with initial state $\mathbf{S}_1\subset \mathbb{R}^3$ and 2D parameterisation $\Omega \subset \mathbb{R}^2$ . For any parametric point $\xi \coloneqq (\xi^{1},\xi^{2})\in \Omega$ , we represent its position on the initial surface state with a mapping $\bar{\mathbf{x}} (\pmb {\xi}): \Omega \to \mathbf{S}_1$ . Furthermore, we encode the time-varying spatial location of $\pmb{\xi}$ on the tracked surface $\mathbf{S}_t$ as $\mathbf{x}(\pmb {\xi},t):\Omega \times [1,\dots,T]\rightarrow \mathbf{S}_t$ , with + +$$ +\begin{array}{l} \mathbf {x} (\boldsymbol {\xi}, t) = \bar {\mathbf {x}} (\boldsymbol {\xi}) + \mathbf {u} (\boldsymbol {\xi}, t), \text {w i t h} \mathbf {u} (\boldsymbol {\xi}, 1) = 0 \text {a n d} \tag {2} \\ \mathbf {u} (\boldsymbol {\xi}, t) = \lambda \mathbf {u} (\boldsymbol {\xi}, t - 1) + \mathcal {F} (\boldsymbol {\xi}, t), \forall t > 1, \\ \end{array} +$$ + +where $\mathbf{u}(\pmb {\xi},t)$ is the deformation field, $\lambda >0$ is a scalar value and $\mathcal{F}(\pmb {\xi},t)$ represents the estimated deformation offset. Here, we encourage the conservation of momentum by setting the deformation of the future surface states in the direction of the previous deformations. + +![](images/0d7cb6e1e3f11600bf14d5f740e23d65ac60d1f995075729125829d01a4f8e65.jpg) +Figure 2. Overview of Thin-Shell-SfT. Our deformation model encodes the surface and its dynamics as neural fields. Given the template $\mathbf{S}_1$ , we first fit a reference field (NRF) from 2D parametric points $\pmb{\xi}$ to the initial 3D positions $\bar{\mathbf{x}}$ . In the main stage, we optimise the deformation field (NDF) $\mathbf{u}(\pmb{\xi}, t)$ by relating estimated surface states $\mathbf{S}_t / \mathcal{G}_t$ to the input monocular views. We induce the dynamically tracked Gaussians to the surface by: (1) Computing their positions $\mathbf{x}$ as the sum of the initial position $\bar{\mathbf{x}}$ and NDF output $\mathbf{u}$ , (2) Setting their rotations $\bar{\mathbf{a}}_i$ as the template's local coordinate system, and (3) Fixing the normal scale $\epsilon$ , and optimising the colour, opacity and tangential scales $(s_1, s_2)$ using only the template texture. For physical plausibility, we impose continuous Kirchhoff-Love physics constraints. + +Parameterisation of the Template and Deformations. Next, we present an adaptive parameterisation of the above deformation model that allocates capacity to dynamic deformation details. Assuming that the template is provided as a mesh $\mathcal{M} \subset \mathbf{S}_1$ , we generate the corresponding 2D parametric space $\mathcal{T} \subset \Omega$ using established techniques (e.g. conformal parameterisation [30]). To learn a continuous representation of the initial state, we fit an MLP, which we call neural reference field (NRF) $\bar{\mathbf{x}}(\boldsymbol{\xi}; \Upsilon)$ to the template mesh parameterisation $\mathcal{T} \mapsto \mathcal{M}$ . For training NRF, we construct points in the parametric and vertex space with randomly sampled barycentric coordinates and train with $\ell_1$ geometry loss. Akin to NRF, we regress the spatio-temporal neural deformation field (NDF) $\mathbf{u}(\boldsymbol{\xi}, t; \Theta)$ using another MLP predicting deformation offsets, $\mathcal{F}(\boldsymbol{\xi}, t; \Theta): \Omega \times [1, \dots, T] \to \mathbb{R}^3$ . Before training, the initial estimate for tracked surface points $\mathbf{x}(\boldsymbol{\xi}, t; \Theta)$ is the sum of noisy NDF deformations, $\mathbf{u}(\boldsymbol{\xi}, t; \Theta)$ and the pre-trained NRF $\bar{\mathbf{x}}(\boldsymbol{\xi}; \Upsilon)$ as given by Eq. (2). Our key idea is to optimise the NDF weights $\Theta$ to ensure photometric consistency of the tracked surfaces with input monocular views while minimising the internal energy of the surface states modelled as a thin shell. The NRF pre-fit enables the coupling of Gaussians to the surface and computing surface metrics for the thin shell energy. We must represent high-frequency signals, such as fine folds and wrinkles on the tracked surface. Moreover, our physics loss requires the computation of higher-order derivatives. Hence, we use sine activation [66] in both MLPs. + +Inference. At testing, $\mathbf{x}(\boldsymbol {\xi},t;\Theta)$ provides continuous access to the reconstructed surface, where the NRF and NDF networks can be consistently queried at varied spatiotemporal resolutions. Moreover, our surface deformation model provides temporal correspondences. Meshing and texture mapping can be achieved in the parametric domain and then transferred to the tracked surface in 3D using Eq. (2). We next describe how to optimise the deformation field $\mathbf{u}(\boldsymbol {\xi},t;\Theta)$ that ensures consistency with monocular images and the physical plausibility of reconstructions. + +# 4.2. Surface Tracking with Gaussian Splitting + +Recall that we have an NDF parameterised by $\Theta$ that outputs the deformation $\mathbf{u}(\boldsymbol {\xi},t;\Theta)$ at any time step. We seek to optimise the NDF so that the tracked surfaces $\{\mathbf{S}_t\}_{t}$ (evaluated with Eq. (2)) generate images matching with input views $\{\mathbf{I}_t\}_{t}$ . A straightforward approach to encourage 3D-to-2D consistency is to minimise the photometric loss with differentiable rendering. We choose Gaussian Splatting [36] as the differentiable renderer, as it seamlessly integrates with our per-point (Eq. (2)) continuous deformation model. Initialisation of Gaussians. To initialise a set of Gaussians, we sample $N_{g}$ well-distributed points (e.g. Poisson disk sampling [7]) from the surface of the input template mesh $\mathcal{M}$ . The sampled points include parametric coordinates $\{\pmb {\xi}_i\}_{i\in [1,\dots ,N_g]}$ , their corresponding positions $\{\bar{\mathbf{x}}_i\}_i$ , and colours $\{\mathbf{c}_i\}_i$ . Next, we will motivate our approach to parameterising Gaussian rotations and scales. Unlike the orig + +inal multi-view [36] or dynamic Gaussian methods [12, 81] that leverage multi-view cues [18], we have a single static camera and, consequently, a challenging monocular setup. Thus, learning the anisotropic covariance of 3D Gaussians from all input frames can lead to poor results due to the inherent monocular ambiguities between deformation, texture and appearance. In our datasets, the initial surface state is reasonably (but not exactly) flat and fully visible. With this assumption, we take the initial surface normal as a proxy for the viewing direction and then fix the scale along normal while allowing for optimisation of the scales along the initial surface tangents. More concretely, we set the rotation matrix of $i$ -th Gaussian equal to the local basis vectors, i.e. $\mathbf{R}_i \equiv [\bar{\mathbf{a}}_1 \bar{\mathbf{a}}_2 \bar{\mathbf{a}}_3]_i$ (see Sec. 4.3 for computation) of the initial state (NRF) $\bar{\mathbf{x}}(\pmb{\xi}_i; \Upsilon)$ and the scale $s_3$ as small $\epsilon$ along surface normal $\bar{\mathbf{a}}_3$ . Finally, the dynamically tracked surface-induced Gaussian $\{\mathcal{G}_t\}_{t}$ can be written as + +$$ +\begin{array}{l} \mathcal {G} _ {t} = \left\{\mathcal {N} \left(\mathbf {x} \left(\boldsymbol {\xi} _ {i}, t; \Upsilon , \Theta\right), \left[ \bar {\mathbf {a}} _ {1} \left(\boldsymbol {\xi} _ {i}; \Upsilon\right), \bar {\mathbf {a}} _ {2} \left(\boldsymbol {\xi} _ {i}; \Upsilon\right), \right. \right. \right. \\ \left. \bar {\mathbf {a}} _ {3} (\boldsymbol {\xi} _ {i}; \Upsilon) ], (s _ {1}, s _ {2}, \epsilon) _ {i}\right), o _ {i}, \mathbf {c} _ {i} \} _ {i = 1} ^ {N _ {g}} \\ \end{array} +$$ + +following Eq. (1). Note that the deformed positions $\mathbf{x}(\pmb{\xi}_i, t)$ are computed using our deformation model (Eq. (2)) and all other Gaussian parameters are shared across surface states. + +Optimisation of NDF and Gaussian Parameters. For optimisation, we use the $\ell_1$ photometric loss similar to 3DGS [36]. If segmentation masks $\{\mathbf{M}_t\}_t$ are available, an additional silhouette loss is added to primarily speed up training. Thus, our data loss reads as + +$$ +\begin{array}{l} \mathcal {L} _ {d} \left(s _ {1}, s _ {2}, o, \mathbf {c}, \Theta\right) = \mathcal {R} \left(\mathcal {G} _ {1}, \mathbf {I} _ {1}; s _ {1}, s _ {2}, o, \mathbf {c}\right) + \\ \sum_ {t = 2} ^ {T} \mathcal {R} \left(\mathcal {G} _ {t}, \mathbf {I} _ {t}; \Theta\right) + \mathcal {R} \left(\tilde {\mathcal {G}} _ {t}, \mathbf {M} _ {t}; \Theta\right), \tag {4} \\ \end{array} +$$ + +where $\mathcal{R}(x;\phi)$ is the Gaussian rasterisation loss with inputs $x$ and optimisable parameters $\phi$ , and $\tilde{\mathcal{G}}_t$ are the dynamic Gaussians $\mathcal{G}_t$ with colour $\mathbf{c}$ set to mask foreground (e.g. white). Note that splating in the first frame updates shared Gaussian properties, whereas all other frames backpropagate to update the deformation field parameters $\Theta$ . With the formulation in Eq. (2), our method propagates gradient information from future states to update 3D reconstructions of the earlier frames. This enables global space-time surface optimisation while respecting the causality of the surface dynamics. While temporal consistency is implicit in Eq. (2), other variants of temporal regularisation are possible (see Appendix B). In contrast to previous 3DGS or dynamic Gaussian methods [81], we note that it is important to optimise all frames in each iteration rather than random frame $t \in [2,\dots,T]$ , as it can otherwise lead to local minima (incorrect learning of folds). Global optimisation is preferred as our deformation model is a single global MLP for all frames, and it is computationally efficient over random frame selection due to the causality of deformations + +![](images/6cd1f02763f8d9b00c8ffc73ec6fe1b58c57227312075ac724f213eb366024aa.jpg) +Figure 3. Thin-shell physics prior is a spatial regulariser that minimises the hyperelastic strain energy density due to deformation w.r.t. the known template. + +(Eq. (2)). Moreover, we keep the number of Gaussians fixed (i.e., no adaptive density control) since Gaussian parameters are learned only from the first frame; it is not known beforehand how the deformation details will arise in future states. + +# 4.3. Thin Shell Physical Prior + +If we optimise the NDF solely using the photometric loss, it could perfectly fit the input images but could result in physically implausible surfaces due to monocular ambiguity. Therefore, we formulate a physics loss applicable to thin shells inspired by the NeuralClothSim approach [34] using the Kirchhoff-Love theory [78]. Since we model the dynamically tracked surface with an MLP, NeuralClothSim enables physics-based prior directly on the continuous surface. However, unlike the forward model of NeuralCloth-Sim, the external forces (such as contacts and the wind) and the material properties generating the deformations are unknown in an inverse setting like ours. For this reason, we assign material parameters to values typical for the surface modelled (e.g., cloth and paper). In contrast to material, it is hard to set reasonable values for forces or boundary conditions as the space of external forces is too large. Nevertheless, intuitively, the image loss effectively takes the role of external force as it guides the motion and deformation of the tracked surface. Thus, we omit the potential energy due to external forces from the optimisation. Finally, with the strain evaluated from the deformation field and the assumed material, our physical prior aims to minimise the internal hyperelastic energy that captures the surface stretching, shearing and bending stiffness; see Fig. 3. + +Strain Computation. To evaluate stretching and bending strain at each training iteration, we randomly sample $N_{p}$ parametric points $\{\pmb{\xi}_i\}_{i=1}^{N_p}$ from the template mesh $\mathcal{M}$ , and perform differential geometry operations on the initial surface state (i.e., NRF) and the deformed state (i.e., NDF). We next present the computation of the local quantities at + +each point $\xi_{i}$ on the initial state using the pre-trained NRF $\overline{\mathbf{x}} (\pmb {\xi};\Upsilon)$ . We use Greek letters for indexing quantities on the parameterised surface (e.g., $\bar{\mathbf{a}}_{\alpha},\alpha ,\beta ,\ldots = 1,2$ ). For notational clarity, we drop the input $\xi_{i},t$ and network weights $\Upsilon ,\Theta$ in all the derived quantities (e.g., $\bar{\mathbf{a}}_1\equiv \bar{\mathbf{a}}_1(\pmb {\xi}_i;\Upsilon),\varepsilon \equiv \varepsilon (\pmb {\xi}_i,t;\Upsilon ,\Theta))$ . In the first step, we extract the local covariant basis: $\{\bar{\mathbf{a}}_{\alpha}:= \partial \bar{\mathbf{x}} /\partial \xi^{\alpha},\bar{\mathbf{a}}_3:= \bar{\mathbf{a}}_1\times \bar{\mathbf{a}}_2\}$ , the set of two vectors tangential to the curvilinear coordinate lines $\xi^{\alpha}$ and the local normal. For the Gaussian samples $\{\pmb {\xi}_i\}_{i = 1}^{N_g}$ , an identical computation is employed for setting the covariant basis as the rotation parameters (as in Sec. 4.2). Next, we evaluate the surface metric tensor $\bar{a}_{\alpha \beta}$ measuring the lengths and the curvature tensor $\bar{b}_{\alpha \beta}$ measuring the curvature of the midsurface. Akin to the covariant tensors (e.g. $\bar{b}_{\alpha \beta}$ ), their contravariant (e.g. $\bar{b}^{\alpha \beta}$ ) and mixed variants (e.g. $\bar{b}_{\alpha}^{\beta}$ ) counterparts are extracted as well. + +With these initial state quantities, we next present the strain—due to deformation—as a function of the NDF $\mathbf{u}(\boldsymbol {\xi};\Theta)$ . The predicted $\mathbf{u} = \hat{u}_i\mathbf{e}_i$ in global Cartesian coordinate system is transformed to contravariant coordinate basis, $\mathbf{u} = u_{\alpha}\bar{\mathbf{a}}^{\alpha} + u_{3}\bar{\mathbf{a}}^{3}$ for strain calculation [34]. Given the deformation gradient $\mathbf{u}_{,\alpha}$ derived from the NDF as + +$$ +\mathbf {u} _ {, \alpha} = \varphi_ {\alpha \lambda} \bar {\mathbf {a}} ^ {\lambda} + \varphi_ {\alpha 3} \bar {\mathbf {a}} ^ {3}, \text {w i t h} \tag {5} +$$ + +$$ +\varphi_ {\alpha \lambda} := u _ {\lambda} | _ {\alpha} - \bar {b} _ {\alpha \lambda} u _ {3} \text {a n d} \varphi_ {\alpha 3} := u _ {3, \alpha} + \bar {b} _ {\alpha} ^ {\lambda} u _ {\lambda}, +$$ + +we evaluate the non-linear membrane strain $\varepsilon = [\varepsilon_{\alpha \beta}]$ and bending strain $\kappa = [\kappa_{\alpha \beta}]$ as + +$$ +\varepsilon_ {\alpha \beta} = \frac {1}{2} (\varphi_ {\alpha \beta} + \varphi_ {\beta \alpha} + \varphi_ {\alpha \lambda} \varphi_ {\beta} ^ {\lambda} + \varphi_ {\alpha 3} \varphi_ {\beta 3}), +$$ + +$$ +\kappa_ {\alpha \beta} = - \varphi_ {\alpha 3} | _ {\beta} - \bar {b} _ {\beta} ^ {\lambda} \varphi_ {\alpha \lambda} + \varphi_ {3} ^ {\lambda} \left(\varphi_ {\alpha \lambda} | _ {\beta} + \frac {1}{2} \bar {b} _ {\alpha \beta} \varphi_ {\lambda 3} - \bar {b} _ {\beta \lambda} \varphi_ {\alpha 3}\right), \tag {6} +$$ + +where $\bar{\mathbf{a}}^i$ denote the local contravariant basis and $\bar{b}_{\alpha \beta}$ , the components of the curvature tensor. In Eq. (6), we use a vertical bar for covariant derivatives, lower comma notation for partial derivatives w.r.t. the curvilinear coordinates $\xi^{\alpha}$ (e.g. $u_{\lambda}|_{\alpha}$ , and $\mathbf{u}_{,\alpha} = \partial \mathbf{u} / \partial \xi^{\alpha}$ ), and Einstein summation convention of repeated indices for tensorial operations (e.g., $\varphi_{\alpha \lambda}\varphi_{\beta}^{\lambda} = \varphi_{\alpha 1}\varphi_{\beta}^{1} + \varphi_{\alpha 2}\varphi_{\beta}^{2}$ ). Notably, we compute all the aforementioned physical quantities with automatic differentiation. Please see Appendix A for more details. + +NDF Optimisation. As material model, we use $\Phi \coloneqq \{\rho, h, E, \nu\}$ , with mass density $\rho$ and the surface thickness $h$ , and elastic coefficients: Young's modulus $E$ , and Poisson's ratio $\nu$ . For simplicity and computational efficiency, we use a linear isotropic constitutive model relating strain to stress, thereby computing the in-plane stiffness $D$ , the bending stiffness $B$ and the elastic tensor $\mathbf{H} = [H^{\alpha \beta \lambda \delta}]$ on the initial template, which read as + +$$ +D := \frac {E h}{1 - v ^ {2}}, B := \frac {E h ^ {3}}{1 2 \left(1 - v ^ {2}\right)}, \text {a n d} \tag {7} +$$ + +$$ +H ^ {\alpha \beta \lambda \delta} := \nu \bar {a} ^ {\alpha \beta} \bar {a} ^ {\lambda \delta} + \frac {1}{2} (1 - \nu) (\bar {a} ^ {\alpha \lambda} \bar {a} ^ {\beta \delta} + \bar {a} ^ {\alpha \delta} \bar {a} ^ {\beta \lambda}). +$$ + +Finally, our physics loss over the tracked surface states reads as the following: + +$$ +\begin{array}{l} \mathcal{L}_{p}(\Theta) = \frac{1}{2N_{p}T}\sum_{i = 1}^{N_{p}}\sum_{t = 2}^{T}\Big(\underbrace{D\varepsilon^{\top}(\boldsymbol{\xi}_{i},t;\Theta)\mathbf{H}(\boldsymbol{\xi}_{i})\varepsilon(\boldsymbol{\xi}_{i},t;\Theta)}_{\text{stretching/shearing stiffness}} \\ + \underbrace {B \boldsymbol {\kappa} ^ {\top} \left(\boldsymbol {\xi} _ {i} , t ; \Theta\right) \mathbf {H} \left(\boldsymbol {\xi} _ {i}\right) \boldsymbol {\kappa} \left(\boldsymbol {\xi} _ {i} , t ; \Theta\right)} _ {\text {b e n d i n g s t i f f n e s s}}) \sqrt {\bar {a} \left(\boldsymbol {\xi} _ {i}\right)}, \tag {8} \\ \end{array} +$$ + +where $\sqrt{a} := |\bar{\mathbf{a}}_1 \times \bar{\mathbf{a}}_2|$ . Note that the term inside $\mathcal{L}_p$ evaluates per point; therefore, we re-sample at each training iteration to explore the continuous surface domain. Finally, we solve for the optimal NDF weights $\Theta^*$ by minimising the objective function $\mathcal{L} = \lambda_d\mathcal{L}_d + \lambda_p\mathcal{L}_p$ with empirically determined loss weights $\{\lambda_d, \lambda_p\}$ . The optimisation can be interpreted as an inverse simulation, with data loss guiding deformation as an external force and physics loss modelling the surface's intrinsic behaviour. We use iterative gradient-based optimisation to that end [37]. We choose $N_p \ll N_g$ as the physics loss involves expensive higher-order derivative computations, whereas the Gaussian rasterisation with samples fixed over training iterations is rather efficient. With resampling at each iteration, the physics loss backpropagates to the continuous surface, offering adaptive, memory-efficient performance compared to mesh-based physics alternatives requiring high resolution. + +# 5. Experimental Results + +Implementation Details. We implement Thin-Shell-SfT using PyTorch [58] and optimise separate NRF and NDF networks for each sequence. The NRF and NDF, both, employ the SIREN [66] architecture using a frequency of $\omega = 5$ and $\omega = 30$ , respectively. Both networks have five hidden layers and 256 units in each layer, and are optimised using Adam [37]. Each network is trained for $\approx 2 \cdot 10^{3}$ iterations; first, the template is fitted with NRF, followed by surface tracking with NDF. While NRF training takes up to two minutes, the core part of the method, i.e. the NDF training, typically takes between 30 minutes to one hour until convergence on an NVIDIA A100 GPU. The Gaussian scales along the normal are set to $\epsilon = 10^{-5}$ . In our experiments, the number of Gaussian samples is $N_{g} \approx 90k$ , and the number of samples for physics loss is $N_{p} = 100$ . Concerning the physics loss, the linear elastic material properties of the surface are set as follows for all sequences: $E = 5000\mathrm{Pa}, \nu = 0.25$ , and $h = 1.2\mathrm{mm}$ . Moreover, we set the image and physics loss weights to $\lambda_{d} = 5$ , and $\lambda_{p} = 1$ , whereas the temporal constraint scalar is set to $\lambda = 0.4$ . For visualisation of the tracked continuous surface, we generate temporally coherent meshes with ball-pivot meshing [4] of the template Gaussian positions $\{\bar{\mathbf{x}}_i\}_{i \in [1,\dots,N_g]}$ . We refer to + +
Seq.S1S2S3S4S5S6S7S8S9Avg
N-NRSfM [65]8.2533.62104.677.0272.668.73129.438.0619.8154.69
D-NRSfM [53]17.144.464.4041.3726.9214.0212.499.915.2915.11
Shimada [61]19.6922.1833.5490.3092.7857.6249.2724.4553.1249.22
DDD [82]2.951.693.8025.7310.466.9715.647.6111.7710.87
Ngo [50]2.191.512.1715.9010.723.017.95*failfail5.92*
Stokko [67]6.13.912.514.511.715.16.910.18.69.93
φ-SfT [33]0.792.753.547.606.153.144.732.522.363.93
Ours1.170.552.4†/3.55.5†/5.78.692.513.82.273.003.3†/3.5
+ +Table 1. We quantitatively compare Thin-Shell-SfT to the state of the art on the $\phi$ -SfT real dataset. The average Chamfer distance is multiplied by $10^4$ for readability. “*” notes that Ngo et al. failed on the last few frames of S7, which we exclude from the error computation. “†” denotes that we report the numbers on the variant of temporal coherency constraint described in Appendix B. + +![](images/d1c0b95ffb1b50ddf2ea5e719b0da4b9e6c95bfd51f36aeca65d5c7a1cd7658a.jpg) + +![](images/339f950a0868abb95aa9562dc7c24748cfb57555b3b2fd2cb31babf5d0e37e35.jpg) +Input Image + +![](images/bffa853f26dc7b92741ee9742b5d151a50287d7c2a5935d288438b01f399fbd0.jpg) + +![](images/6be45416f2de0cfb47e4cdfc09ac534516f4ab4b42a0453f27dbc1fbaa4f7212.jpg) +Ground-Truth Normals + +![](images/0bcec40b359c8f506edb14d59d71a08456ba03168851df4b18441ac90eff2695.jpg) + +![](images/cc6cafb72167351b3bc48653869df7fece3dc174c40fb9eecdc2d5acb5a80269.jpg) +Normal Map Ours + +![](images/1b29878780a6d9ec6ed3f95754d64061a715f195a54c872131dc6dbcca493c37.jpg) + +![](images/0bd5bb8e6fd4384978b38c825f24250f410302ceeb96daaa9dc1c2dd4f7dade5.jpg) +Normal Map $\phi$ -SfT +Figure 4. Comparison of the reconstruction normal maps for ours and $\phi$ -SfT. We show the cosine normal consistency to ground truth on the right and the normal metrics in Tab. I (Appendix). + +![](images/8d6853676543f2b457af600ff02fbf67f66eda2c11b183337d7c27c7218271de.jpg) + +![](images/2864c47de2278b87ac2a8a6b520a671afcea0bac6288323d151966da38f6a658.jpg) +Normal Error Ours + +Appendices B and D and Figs. II and III for studies on the effect of hyperparameters $\omega, \lambda$ and $N_{g}$ . + +Datasets and Error Metrics. We evaluate on the benchmark $\phi$ -SfT dataset [33] consisting of nine RGB videos (and reference depths) of highly challenging cloth deformations. As is common with non-rigid reconstruction methods [33, 65], we rigidly align the reconstructions of all compared methods to the ground truth. This is achieved with Procrustes alignment [72] for the first frame, which is refined with rigid ICP [5] for subsequent frames. + +# 5.1. Comparison + +SfT and NRSfM. We compare Thin-Shell-SfT to different SfT methods, namely $\phi$ -SfT [33], Yu et al.'s Direct, Dense, Deformable (DDD) [82], Ngo et al.'s Ngo2015 [50] and Shimada et al.'s IsMo-GAN [61], and finally to Stotko et al.'s approach [67]. DDD is provided with the required hierarchy of coarse-to-fine templates, and Ngo2015 and Stotko et al. with the same template as in the $\phi$ -SfT dataset. Since the dataset is richly textured, the 2D point tracking methods would perform well; therefore, we additionally compare to the NRSfM methods. In particular, we compare to Sidhu et al.'s Neural NRSfM (N-NRSfM) [65] and Parashar et al.'s Diff-NRSfM [53]. The 2D point correspondences required by the NRSfM methods are provided by densely tracking the 2D points across input images with multi-frame subspace flow (MFSF) [19, 20]. The first frame of the sequence is selected as a keyframe for 2D tracking. + +In Fig. 5-(left), we show qualitative reconstructions of + +Thin-Shell-SfT compared to the prior state of the art (SotA) [33, 53, 67]. We capture fine wrinkles and folds in the deforming sequences that were not addressed by the earlier works. Next, we reconstruct surfaces with severe self-occlusions due to multiple layered folds in the extended versions of the $\phi$ -SfT sequences [33]. See Fig. 5-(right), where our physics-based model provides a reasonable prior for occluded regions; $\phi$ -SfT fails here due to prohibitive memory requirements. In Tab. 1, we show a quantitative comparison where we report the Chamfer distance (with pseudoground-truth point clouds) against the state-of-the-art methods over all the $\phi$ -SfT dataset sequences. We outperform the existing methods on most sequences, often substantially and on average over all sequences. While the default temporal coherency works best on the evaluated dataset, we notice qualitative improvements for two out of the nine sequences with the other variants. In Fig. 4, we further show comparisons between the normal maps of the tracked surface. Additional numerical comparisons with the previous SotA [33] on runtime and normal consistency are in Appendix C and Tab. I. We obtain a lower $0.009(\ell_2)$ and $0.034$ (cosine) normal error, whereas $\phi$ -SfT shows $0.013$ and $0.041$ , respectively. One of the primary reasons for the failure of $\phi$ -SfT [33] is its limitation to low-resolution meshes ( $\approx 300$ vertices) capturing just coarse deformations. Our NDF offers, similar to scene representation approaches [47, 75], a smooth (cf. global MLP), low-dimensional (cf. a fixed parameter count) deformation space that is adaptive to dynamic details and can represent high-frequency signals (cf. sine activation). Although we discretise the physics loss, we re-sample at each iteration, leading to an adaptive discretisation. In contrast, earlier physics-based methods [33, 67] use fixed discretisation, either due to differentiable simulation not supporting remeshing [33] or a fixed-resolution surrogate model [67]. Our continuous formulation enables unprecedented fine-grained results. + +Dynamic View Synthesis. Next, we compare with state-of-the-art dynamic view synthesis methods, particularly K-Planes [15] and Deformable Gaussians [81]. For the latter, we initialise the Gaussians with points sampled from the input template instead of the default SfM points. The compared radiance field methods are very effective in novel-view synthesis for monocular videos that provide multiview cues [18]. However, they fail to recover spatiotemporally coherent surface geometry for monocular RGB videos captured with a single static camera; see Fig. 6. + +# 5.2. Ablative Studies + +Next, we test the following aspects of our method: contribution of the physics loss, global optimisation and surface-induced Gaussians. Additional ablations, including momentum and mask loss, are presented in the Appendix E. + +Physics Prior. While relying solely on image loss yields + +![](images/24d56cec1882690718dd630e7467da6ce2d5ba15b1734cc9155d8110310480ae.jpg) +Figure 5. Exampary 3D reconstructions. (Left:) Comparisons focusing on high-frequency wrinkles. Thin-Shell-SfT captures the wrinkles best among all compared methods in one of the most challenging examples, outperforming $\phi$ -SfT [33], Stotko et al. [67] and Diff-NRSfM [53]. (Right:) Our results on the extended $\phi$ -SfT dataset highlight the excellent tracking in the occluded regions. + +![](images/7b9ea7e84bce812eb006a66806e589ec0870436ae9c0521547b2d998d01deeb0.jpg) +Figure 6. We compare our reconstructions to dynamic view synthesis methods [15, 81] on the rendered depth maps. The texture details retained in the depth maps imply that the compared methods fail to learn accurate surface (mistake texture for geometry). + +![](images/ea6b27ca125c9ea7cd03434845fcad8e531ff88b2f70dcc1437419acbd2bea61.jpg) +Figure 7. Ablations. (a:) Without physics, the surface can severely stretch or shrink. (b:) Random frame optimisation leads to local minima, (c,d:) Not positioning and orienting Gaussians on the surface leads to distortions and poor reconstructions. + +accurate RGB recovery during rendering, it can lead to severely distorted surface tracking. This arises from the inherent monocular ambiguity of our setting as multiple possible deformations in 3D correspond to the same 2D image. As seen in Fig. 7-(a), incorporating Kirchhoff-Love-based thin shell prior is essential for achieving physically plausible and accurate surface reconstruction. + +Global Optimisation. We test a version of our method employing optimisation with randomly selected frames at each iteration instead of the proposed joint optimisation over all frames. Fig. 7-(b) visualises the ablated results showing poor performance for folds and wrinkles. Without any multi-view cues, as in the case of the used $\phi$ -SfT dataset, random frame optimisation leads to local minima. + +Surface-induced Gaussians. The specially-tailored initialisation and optimisation of Gaussian parameters (Eqs. (3) and (4)) are crucial for the accurate geometric reconstruction. In the ablated version of our method, we optimise the shared Gaussian parameters, i.e., covariance, opacity, and colour on all input frames (similar to earlier methods [36, 81]) instead of the single template frame as in Eq. (4). Due to the ambiguities between geometric details and appearance over the deforming sequence, this leads to wrongly reconstructed Gaussians, eventually resulting in poor surface reconstructions; see Fig. 7-(c). In addition, fixing the normal scale on the template frame is useful for preventing elongated Gaussians (Fig. 7-(d)). + +In Tab. II-appendix, we report the above results on the full $\phi$ -SfT dataset. On average, we obtain Chamfer distances of 34.25, 14.0, 3.75, and 3.46 for no physics, no surface-induced Gaussians, no normal scale, and the full model, respectively. We notice that including physics and surface induction are crucial for accurate tracking. + +# 6. Conclusion + +We present Thin-Shell-SfT, a new approach for dense monocular non-rigid 3D surface tracking relying on new principles, i.e. adaptive neural deformation field, continuous Kirchhoff-Love thin shell prior and surface-induced 3D Gaussian Splating. Experiments on the $\phi$ -SfT dataset demonstrate that our method accurately reconstructs the challenging fine-grained deformations such as cloth folds and wrinkles. Overall, we significantly improve fine-grained surface tracking using an adaptive deformation model and continuous thin shell physics compared to existing approaches that only support coarse deformations. Thanks to the physics prior, Thin-Shell-SfT is reasonably robust to occlusions, although extreme self-collisions remain challenging. Similarly, tracking textureless surfaces is another possible research direction that would necessitate special handling in future. + +Acknowledgement. This work was partially supported by the VIA Research Center. + +# References + +[1] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Nonrigid structure from motion in trajectory space. In Advances in Neural Information Processing Systems (NeurIPS), 2009. 2 +[2] Mohammad Dawud Ansari, Vladislav Golyanik, and Didier Stricker. Scalable dense monocular surface reconstruction. In International Conference on 3D Vision (3DV), 2017. 2 +[3] Adrien Bartoli, Yan Gérard, François Chadebecq, Toby Collins, and Daniel Pizarro. Shape-from-template. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2015. 1, 2 +[4] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE TVCG, 1999. 6 +[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 7 +[6] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Computer Vision and Pattern Recognition (CVPR), 2000. 2 +[7] Robert Bridson. Fastoisson disk sampling in arbitrary dimensions. SIGGRAPH sketches, 2007. 4 +[8] David Casillas-Perez, Daniel Pizarro, David Fuentes-Jimenez, Manuel Mazo, and Adrien Bartoli. The isowarp: the template-based visual geometry of isometric surfaces. International Journal of Computer Vision, 2021. 1, 2 +[9] Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne Vouga, and Christoph Lassner. Virtual elastic objects. In Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[10] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. In Computer Vision and Pattern Recognition (CVPR), 2014. 2 +[11] Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3 +[12] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2, 5 +[13] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splating: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 2 +[14] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Gaussian splashing: Unified particles for versatile motion synthesis and rendering. Computer Vision and Pattern Recognition (CVPR), 2025. 2 +[15] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In + +Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7, 8 +[16] David Fuentes-Jimenez, Daniel Pizarro, David Casillas-Perez, Toby Collins, and Adrien Bartoli. Texture-generic deep shape-from-template. IEEE Access, 2021. 2 +[17] David Fuentes-Jimenez, Daniel Pizarro, David Casillas-Pérez, Toby Collins, and Adrien Bartoli. Deep shape-from-template: Single-image quasi-isometric deformable registration and reconstruction. Image and Vision Computing, 127: 104531, 2022. 2 +[18] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 5, 7 +[19] Ravi Garg, Anastasios Roussos, and Lourdes Agapito. A variational approach to video registration with subspace constraints. International Journal of Computer Vision (IJCV), 104(3):286-314, 2013. 2, 7 +[20] Ravi Garg, Anastasios Roussos, and Lourdes Agapito. Source code of Multi-Frame Subspace Flow (MFSF). http://www0.cs.ucl.ac.uk/staff/lagapito/subspace_flow/, 2015.7 +[21] Vladislav Golyanik, Soshi Shimada, Kiran Varanasi, and Didier Stricker. Hdm-net: Monocular non-rigid 3d reconstruction with learned deformation model. In EuroVR, 2018. 2 +[22] Vladislav Golyanik, André Jonas, and Didier Stricker. Consolidating segmentwise non-rigid structure from motion. In Machine Vision Applications (MVA), 2019. 2 +[23] Stella Graßhof and Sami Sebastian Brandt. Tensor-based non-rigid structure from motion. In Winter Conference on Applications of Computer Vision (WACV), 2022. 2 +[24] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[25] Jingfan Guo, Jie Li, Rahul Narain, and Hyun Soo Park. Inverse simulation: Reconstructing dynamic geometry of clothed humans via optimal control. In Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[26] Marc Habermann, Weipeng Xu, Helge Rhodin, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Nrst: Non-rigid surface tracking from monocular video. German Conference on Pattern Recognition (GCPR), 2018. 2 +[27] Nazim Haouchine and Stephane Cotin. Template-based monocular 3d recovery of elastic shapes using lagrangian multipliers. In Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[28] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2 +[29] Florian Hofherr, Lukas Koestler, Florian Bernard, and Daniel Cremers. Neural implicit representations for physical parameter inference from a single video. In Winter Conference on Applications of Computer Vision (WACV), 2023. 3 +[30] Alec Jacobson, Daniele Panozzo, et al. libigl: A simple $\mathbf{C}++$ geometry processing library, 2018. https://libigl.github.io/. 4 + +[31] Miguel Jaques, Michael Burke, and Timothy Hospedales. Physics-as-inverse-graphics: Unsupervised physical parameter estimation from video. In International Conference on Learning Representations (ICLR), 2020. 3 +[32] Erik Johnson, Marc Habermann, Soshi Shimada, Vladislav Golyanik, and Christian Theobalt. Unbiased 4d: Monocular 4d reconstruction with a neural deformation model. In Computer Vision and Pattern Recognition (CVPR) Workshops, 2023. 2 +[33] Navami Kairanda, Edith Tretschk, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. $\phi$ -SfT: Shape-from-template with a physics-based deformation model. In Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 7, 8 +[34] Navami Kairanda, Marc Habermann, Christian Theobalt, and Vladislav Golyanik. Neuralclothsim: Neural deformation fields meet the thin shell theory. Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 5, 6, 1 +[35] Rama Kandukuri, Jan Achterhold, Michael Moeller, and Joerg Stueckler. Learning to identify physical parameters from video using differentiable physics. In German Conference for Pattern Recognition (GCPR), 2020. 3 +[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 2, 3, 4, 5, 8 +[37] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv e-prints, 2017. 6 +[38] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. European Conference on Computer Vision (ECCV), 2024. 2 +[39] Suryansh Kumar. Jumping manifolds: Geometry aware dense non-rigid structure from motion. In Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[40] Suryansh Kumar, Anoop Cherian, Yuchao Dai, and Hongdong Li. Scalable dense non-rigid structure-from-motion: A grassmannian perspective. In Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[41] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep Physics-aware Inference of Cloth Deformation for Monocular Human Performance Capture. In International Conference on 3D Vision (3DV), 2021. 3 +[42] Yifei Li, Tao Du, Kui Wu, Jie Xu, and Wojciech Matusik. Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics, 2022. 1 +[43] Junbang Liang, Ming Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse problems. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 3 +[44] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338, 2024. 2 +[45] Alberta Longhini, Marcel Busching, Bardienus Pieter Duisterhof, Jens Lundell, Jeffrey Ichnowski, Marten Björkman, + +and Danica Kragic. Cloth-splatting: 3d cloth state estimation from rgb supervision. In 8th Annual Conference on Robot Learning, 2024. 2 +[46] Augustus Edward Hough Love. A treatise on the mathematical theory of elasticity. University press, 1927. 2 +[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 7 +[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 2 +[49] J. Krishna Murthy, Miles Macklin, Florian Golemo, Vikram Voleti, Linda Petrini, Martin Weiss, Breandan Considine, Jérôme Parent-Lévesque, Kevin Xie, Kenny Erleben, Liam Paull, Florian Shkurti, Derek Nowrouzezahrai, and Sanja Fidler. gradsim: Differentiable simulation for system identification and visuomotor control. In International Conference on Learning Representations (ICLR), 2021. 3 +[50] Dat Tien Ngo, Sanghyuk Park, Anne Jorstad, Alberto Crivellaro, Chang D. Yoo, and Pascal Fua. Dense image registration and deformable surface reconstruction in presence of occlusions and minimal texture. In International Conference on Computer Vision (ICCV), 2015. 1, 2, 3, 7 +[51] Marco Paladini, Alessio Del Bue, Marko Stosic, Marija Dodig, Jo ao Xavier, and Lourdes Agapito. Factorization for non-rigid and articulated structure using metric projections. In Computer Vision and Pattern Recognition (CVPR), 2009. +[52] Shaifali Parashar, Daniel Pizarro, Adrien Bartoli, and Toby Collins. As-rigid-as-possible volumetric shape-from-template. In International Conference on Computer Vision (ICCV), 2015. 1, 2 +[53] Shaifali Parashar, Mathieu Salzmann, and Pascal Fua. Local non-rigid structure-from-motion from diffeomorphic mappings. In Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7, 8 +[54] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40 (6):1-12, 2021. 2 +[55] Mathieu Perriollat, Richard Hartley, and Adrien Bartoli. Monocular template-based reconstruction of inextensible surfaces. International Journal of Computer Vision (IJCV), 95(2):124-137, 2011. 2 +[56] Albert Pumarola, Antonio Agudo, Lorenzo Porzi, Alberto Sanfeliu, Vincent Lepetit, and Francesc Moreno-Noguer. Geometry-Aware Network for Non-Rigid Shape Prediction from a Single View. In Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[57] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3 + +[58] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 6 +[59] Davis Rempe, Leonidas J. Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jimei Yang. Contact and human dynamics from monocular video. In European Conference on Computer Vision (ECCV), 2020. 3 +[60] Mathieu Salzmann, Julien Pilet, Slobodan Ilic, and Pascal Fua. Surface deformation models for nonrigid 3d shape recovery. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 29(8):1481-1487, 2007. 2 +[61] Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Didier Stricker. IsMo-GAN: Adversarial learning for monocular non-rigid 3d reconstruction. In Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 2, 7 +[62] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics, 39(6), 2020. 3 +[63] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, Patrick Pérez, and Christian Theobalt. Neural monocular 3d human motion capture with physical awareness. ACM Transactions on Graphics, 40(4), 2021. 3 +[64] Soshi Shimada, Vladislav Golyanik, Patrick Pérez, and Christian Theobalt. Decaf: Monocular deformation capture for face and hand interactions. ACM Transactions on Graphics, 42(6), 2023. 3 +[65] Vikramjit Sidhu, Edgar Tretschk, Vladislav Golyanik, Antonio Agudo, and Christian Theobalt. Neural dense non-rigid structure from motion with latent space constraints. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 7 +[66] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 4, 6, 3 +[67] David Stotko, Nils Wandel, and Reinhard Klein. Physics-guided shape-from-template: Monocular video perception through neural surrogate models. Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 7, 8 +[68] Guoxing Sun, Rishabh Dabral, Pascal Fua, Christian Theobalt, and Marc Habermann. Metacap: Meta-learning priors from multi-view imagery for sparse-view human performance capture and rendering. In European Conference on Computer Vision (ECCV), 2024. 2 +[69] Lorenzo Torresani, Aaron Hertzmann, and Chris Bregler. Nonrigid structure-from-motion: Estimating shape and motion with hierarchical priors. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 30(5), 2008. 2 +[70] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In International Conference on Computer Vision (ICCV), 2021. 2 + +[71] Edith Tretschk, Navami Kairanda, Mallikarjun BR, Rishabh Dabral, Adam Kortylewski, Bernhard Egger, Marc Habermann, Pascal Fua, Christian Theobalt, and Vladislav Golyanik. State of the art in dense monocular non-rigid 3d reconstruction. In Computer Graphics Forum (Eurographics State of the Art Reports), 2023. 1, 2 +[72] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 1991. 7 +[73] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. Games: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459, 2024. 2 +[74] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation. In Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[75] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2, 7 +[76] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. International Conference on Computer Vision (ICCV), 2023. 2 +[77] Sebastian Weiss, Robert Maier, Daniel Cremers, Rüdiger Westermann, and Nils Thuerey. Correspondence-free material reconstruction using sparse surface constraints. Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[78] Gerald Kempner and Demosthenes Talaslidis. Mechanics of solids and shells. CRC, Boca Raton, 2003. 3, 5 +[79] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[80] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In International Conference on Computer Vision (ICCV), 2023. 3 +[81] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 7, 8 +[82] Rui Yu, Chris Russell, Neill D. F. Campbell, and Lourdes Agapito. Direct, dense, and deformable: Template-based non-rigid 3d reconstruction from rgb video. In International Conference on Computer Vision (ICCV), 2015. 2, 7 +[83] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y. Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. PhysDreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision (ECCV), 2024. 2 +[84] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, and Gordon Wetzstein. Physavatar: Learning the physics of dressed 3d avatars from visual observations. In European Conference on Computer Vision (ECCV), 2024. 2 \ No newline at end of file diff --git a/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/images.zip b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..129b8f7bae538eb2c6214f4d18a24513951058d0 --- /dev/null +++ b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8b1db7953be5b9de828adb899c04994b01f74f09d74a61c1e7a524de2c497ae +size 497870 diff --git a/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/layout.json b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..42c14c30338f8a3815872a2a7c990b003ef4ab6c --- /dev/null +++ b/2025/Thin-Shell-SfT_ Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields/layout.json @@ -0,0 +1,11167 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 102, + 531, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 102, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 80, + 102, + 531, + 138 + ], + "type": "text", + "content": "Thin-Shell-SfT: Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "text", + "content": "Navami Kairanda" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "text", + "content": " Marc Habermann" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "text", + "content": " Shanthika Naik" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "text", + "content": " Christian Theobalt" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "text", + "content": " Vladislav Golyanik" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "text", + "content": "MPI for Informatics, SIC " + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "text", + "content": "VIA Research Center " + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 61, + 160, + 550, + 190 + ], + "type": "text", + "content": "IIT Jodhpur" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 228, + 200, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 228, + 200, + 242 + ], + "spans": [ + { + "bbox": [ + 151, + 228, + 200, + 242 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 254, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 254, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 54, + 254, + 296, + 578 + ], + "type": "text", + "content": "3D reconstruction of highly deformable surfaces (e.g. cloths) from monocular RGB videos is a challenging problem, and no solution provides a consistent and accurate recovery of fine-grained surface details. To account for the ill-posed nature of the setting, existing methods use deformation models with statistical, neural, or physical priors. They also predominantly rely on nonadaptive discrete surface representations (e.g. polygonal meshes), perform frame-by-frame optimisation leading to error propagation, and suffer from poor gradients of the mesh-based differentiable renderers. Consequently, fine surface details such as cloth wrinkles are often not recovered with the desired accuracy. In response to these limitations, we propose ThinShell-SfT, a new method for non-rigid 3D tracking that represents a surface as an implicit and continuous spatiotemporal neural field. We incorporate continuous thin shell physics prior based on the Kirchhoff-Love model for spatial regularisation, which starkly contrasts the discretised alternatives of earlier works. Lastly, we leverage 3D Gaussian splatting to differentiably render the surface into image space and optimise the deformations based on analysis-by-synthesis principles. Our Thin-Shell-SfT outperforms prior works qualitatively and quantitatively thanks to our continuous surface formulation in conjunction with a specially tailored simulation prior and surface-induced 3D Gaussians. See our project page at https://4dqv.mpi-inf.mpg.de/ThinShellSfT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "Non-rigid 3D reconstruction and tracking of general deformable surfaces from a monocular RGB camera is an important, challenging and ill-posed problem that is far from being solved [71]. It has applications in game development, robotics and augmented reality, to name a few areas." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Prior works on temporally-coherent general surface reconstruction can be grouped into Shape-from-Template (SfT) [3, 8, 33] and Non-Rigid Structure-from-Motion [39," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 228, + 363, + 339 + ], + "blocks": [ + { + "bbox": [ + 318, + 228, + 363, + 339 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 363, + 339 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 363, + 339 + ], + "type": "image", + "image_path": "50eace23936289e74a9e49973a93b3e7db24dcaf23c830454c35b1f9ab79e834.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Our Thin-Shell-SfT approach reconstructs high-fidelity deformable 3D surface geometry with fine-grained wrinkles from a monocular video, while the previous best method [33] struggles. The coloured tracks (\"Ours, Input View\") visualise the 3D Gaussian trajectories over time and across all input video frames." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 366, + 228, + 413, + 339 + ], + "blocks": [ + { + "bbox": [ + 366, + 228, + 413, + 339 + ], + "lines": [ + { + "bbox": [ + 366, + 228, + 413, + 339 + ], + "spans": [ + { + "bbox": [ + 366, + 228, + 413, + 339 + ], + "type": "image", + "image_path": "5a76305e924ba4058082f0411934b7a6a215de218ebd3aa7d46894bba644baf3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 416, + 228, + 455, + 339 + ], + "blocks": [ + { + "bbox": [ + 416, + 228, + 455, + 339 + ], + "lines": [ + { + "bbox": [ + 416, + 228, + 455, + 339 + ], + "spans": [ + { + "bbox": [ + 416, + 228, + 455, + 339 + ], + "type": "image", + "image_path": "e6fd630602f3b6da00bfe490470051ae2a48a35a4a187cf363bd4c1adc1d8017.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 457, + 228, + 504, + 338 + ], + "blocks": [ + { + "bbox": [ + 457, + 228, + 504, + 338 + ], + "lines": [ + { + "bbox": [ + 457, + 228, + 504, + 338 + ], + "spans": [ + { + "bbox": [ + 457, + 228, + 504, + 338 + ], + "type": "image", + "image_path": "68722334cdc192f9dd2e6b40eb6f5ec7dd5442d342b3ef76fdf0ef837aaa21f6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 506, + 228, + 553, + 336 + ], + "blocks": [ + { + "bbox": [ + 506, + 228, + 553, + 336 + ], + "lines": [ + { + "bbox": [ + 506, + 228, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 506, + 228, + 553, + 336 + ], + "type": "image", + "image_path": "6075cd515a0aea28f6efbbd407d516bf197f9f0693a801471cb82fe398a94c7b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 415, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 713 + ], + "type": "text", + "content": "53, 65]; they rely on 2D point tracks across monocular images (NRSfM) [53, 65] or between the input image and the template (SfT) [3, 8]. Recent physics-based SfT approaches [33, 67] demonstrate state-of-the-art results and cause a paradigm shift from geometric- [3, 8, 52] to physics-based constraints and from 2D-point-based [8, 50] to dense photometric loss using differentiable renderers. However, even such approaches as " + }, + { + "bbox": [ + 313, + 415, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 415, + 555, + 713 + ], + "type": "text", + "content": "-SfT [33] do not support fine wrinkles and require multiple hours to reconstruct a limited number of frames (" + }, + { + "bbox": [ + 313, + 415, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\approx 50" + }, + { + "bbox": [ + 313, + 415, + 555, + 713 + ], + "type": "text", + "content": ") of a single object, which is due to the underlying surface mesh representation, i.e., an explicit and discrete. First, determining the mesh resolution for a specific scene is difficult, e.g. selecting a high resolution could prohibitively increase the memory and computational cost, while a lower resolution might not account for fine-scale deformations. Second, the Finite Element Mesh (FEM)-based differentiable physics simulators [42, 43] lead to inconsistent simulations at different resolutions, preventing the adoption of coarse-to-fine strategies for tracking. Third, simulator-based approaches [33, 67] optimise frame-by-frame and can get stuck in local minima. Moreover, mesh-based differentiable renderers used in recent approaches [33, 67] are nonadaptive and do not support the complex remeshing required for the dynamic details arising in deformable scenes." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11373" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 514 + ], + "type": "text", + "content": "To overcome these limitations, our key idea is to replace meshes with an adaptive and continuous surface and deformation representation1. We tightly couple it with differentiable physics to guide the deformations while ensuring photometric consistency with the monocular images through analysis-by-synthesis. We model surfaces as Kirchhoff-Love thin shells [46] and propose a physics-based continuous deformation prior. Our regulariser minimises the internal hyperelastic energy of the tracked surface, ensuring its physical plausibility and providing prior for occlusion handling. Unlike discrete priors operating on mesh vertices [33, 82], our continuous prior updates any point on the surface, enabling physics supervision for fine-grained details such as wrinkles. While similar modelling has been applied to cloth simulation [34], it has not been shown for inverse problems like monocular surface tracking, which involves many unknowns (e.g., material, forces, contacts) and ambiguities between them. In addition to continuous prior, we perform joint space-time optimisation while taking the causality of deformation into account to impose temporal coherence (i.e. the current deformed state can update previous state parameters but not the future ones). As the differentiable renderer, we employ 3D Gaussian Splatting [36], which recently emerged as a prominent technique for radiance field rendering. It integrates straightforwardly with our continuous per-point deformation model and offers high-quality image gradients due to the continuous volumetric radiance field formulation. We leverage differential geometric quantities from the thin shell physics to couple Gaussians to the surface and optimise the parameters of dynamically tracked Gaussians. While there are extensions of Gaussian splatting focussing on dynamic view synthesis [12, 81] or multi-view static surface reconstruction [24, 73], we show how to adapt it for surface tracking from monocular videos with de-facto absent multi-view 3D reconstruction cues [18] such as the " + }, + { + "bbox": [ + 55, + 72, + 297, + 514 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 72, + 297, + 514 + ], + "type": "text", + "content": "-SfT [33] dataset. The technical contributions of this paper are as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 516, + 295, + 623 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 516, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 552 + ], + "type": "text", + "content": "- Thin-Shell-SfT, i.e. a new method for monocular non-rigid 3D surface tracking operating on a continuous adaptive spatiotemporal representation;" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 552, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 587 + ], + "type": "text", + "content": "- A continuous deformation prior based on the principles of the Kirchhoff-Love thin shell theory and application of such a prior in an inverse problem (shape from template);" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 588, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 623 + ], + "type": "text", + "content": "- Adaptation of 3D Gaussian Splating for 3D tracking of highly deformable dynamic surfaces forming folds and fine wrinkles, captured by a static monocular camera." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 625, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 295, + 661 + ], + "type": "text", + "content": "Our experimental results on the challenging " + }, + { + "bbox": [ + 55, + 625, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 625, + 295, + 661 + ], + "type": "text", + "content": "-SfT benchmark [33] show a significant improvement over the state of the art in terms of reconstruction accuracy; see Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 71, + 402, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 402, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 402, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 553, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 186 + ], + "type": "text", + "content": "We review methods for monocular reconstruction of general non-rigid surfaces. They differ in their assumptions about the deformation model, data terms, and the priors; we refer to a recent survey by Tretschk et al. [71] for an in-depth review. Similar to ours, many recent works integrate Gaussians with physics or clothing but in very different contexts such as simulation [14, 79], material estimation [44, 83], and tracking from multi-view video [13, 45, 84]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 187, + 553, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 498 + ], + "type": "text", + "content": "Shape-from-Template (SfT) methods [16, 27, 33, 50, 55, 56, 60, 82] assume a single static shape or template as a prior. A template often corresponds to the first frame of the sequence [3, 33, 50] and, in other cases, to the rigid initialisation [17, 61]. SfT approaches [3, 26, 50, 82] employ a 3D-2D reprojection constraint and deform the template using temporal and geometric soft constraints to encourage physical plausibility. On the other hand, learning-based SfTs [16, 21, 56, 61] encode the template in the network and regress the surface deformations. They are sometimes object-specific with template encoded in the neural network weights [17] and often object-generic [16, 22, 61] when supervised with synthetic datasets. Recently, Kairanda et al. [33] and Stoko et al. [67] explain 2D observations through physics-based simulation of the deformation process and employ differentiable rendering for per-sequence gradient-based optimisation. In contrast to traditional SfT [3, 8, 52], they do not require registrations (the correspondence between the 3D template and the image). Ours falls under this category as we require the template corresponding to the first frame. In contrast to the prior works, ours is the first method to continuously represent the surface and its dynamics with a neural field capable of representing high-frequency signals [66]. We supervise with thin shell physics constraint applied at continuous points in the domain and employ the Gaussian Splatting functionality [36]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 605 + ], + "type": "text", + "content": "Non-Rigid Structure from Motion (NRSfM) [1, 6, 10, 51, 69] relies on motion cues of 2D point tracks over the input monocular images and outputs per-frame camera-object poses and the 3D shapes. Recent dense NRSfM methods [2, 22, 23, 40, 53, 65, 74] rely on per-pixel multi-frame optical flow or video registration [19]. In contrast, we do not make a restrictive assumption of available 2D point tracks and track highly challenging cloth deformations (fine wrinkles), which goes beyond NRSfM capabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": "Dynamic Novel View Synthesis. Radiance field methods [36, 47, 48, 75] learn volumetric scene representation for high-quality novel-view synthesis from multi-view images, which are extended to support dynamic scenes [11, 15, 32, 38, 54, 70, 76, 81]. Though they demonstrate impressive results, most do not show or evaluate geometry reconstruction. Moreover, they require multi-view cues [18, 56]; in contrast, we focus on dynamic scenes captured from a single static camera without such cues." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": "1Most prior methods use a discrete number " + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": " of points or mesh vertices (including neural ones e.g. [65]) where " + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": " needs to be decided beforehand and cannot be changed during optimisation, unlike our continuous formulation, allowing arbitrary queries." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11374" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "Physics-based Priors have been applied in inverse problems [25, 41, 59, 62]. The examples range from 3D human pose estimation [41, 63] and parameter estimation [29, 43, 49, 77] to dense SfT [33, 64, 67]. Recent works [9, 57] extend neural representations to physical parameter inference from videos. Some methods [31, 35, 80] combine differentiable simulation and rendering. In the inverse setting, we are the first to impose continuous physics constraints while prior methods use mesh-based simulators." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 189, + 141, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 141, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 141, + 201 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 209, + 295, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 257 + ], + "type": "text", + "content": "We review 3D Gaussian Splatting [36] and NeuralCloth-Sim [34], from which we take inspiration for the data term (ensuring consistency of reconstructions with input images) and the prior term (encouraging physical plausibility)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 295, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 295, + 305 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) models a scene as a volumetric radiance field with a dense set of 3D Gaussians, each defined by its position (mean), anisotropic covariance, opacity, and colour. " + }, + { + "bbox": [ + 55, + 257, + 295, + 305 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 55, + 257, + 295, + 305 + ], + "type": "text", + "content": " Gaussians are represented as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 310, + 295, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 310, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 109, + 310, + 295, + 326 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\left\\{\\mathcal {N} \\left(\\mathbf {x} _ {i}, \\mathbf {R} _ {i}, \\mathbf {S} _ {i}\\right), o _ {i}, \\mathbf {c} _ {i} \\right\\} _ {i = 1} ^ {N _ {g}}, \\tag {1}", + "image_path": "345542af40bd8cc55328bb40c197708f20e8b57ad6e2d629117fa7fc337df045.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "text", + "content": " denotes the position; " + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{RSS}^{\\top}\\mathbf{R}^{\\top}" + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "text", + "content": " is the anisotropic covariance parameterised as an ellipsoid with scale " + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "text", + "content": " and rotation " + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "text", + "content": " is the opacity, and " + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "text", + "content": " is the colour of the Gaussian. In 3DGS [36], " + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 330, + 295, + 426 + ], + "type": "text", + "content": " is optimised through gradient-based training with multi-view image loss. Learning a 3DGS representation from monocular inputs requires substantial adjustments in different contexts [11, 81], and we show its utility in SfT for the first time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 557 + ], + "type": "text", + "content": "NeuralClothSim [34] is a recent quasistatic cloth simulator representing surface deformations as a coordinate-based implicit neural deformation field (NDF). Given a target simulation scenario specified by the initial surface state, material properties and external forces, NeuralClothSim learns the equilibrium deformation field using the laws of the Kirchhoff-Love thin shell theory. Upon convergence, the equilibrium state can be queried continuously and consistently at multiple resolutions. These properties, combined with the memory adaptivity of neural fields, make NeuralClothSim well-suitable for inverse problems like ours." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Kairanda et al. [34] model a cloth quasistatically as an NDF " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(\\pmb {\\xi}):\\Omega \\mapsto \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " defined on its curvilinear coordinate space " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": ". For a volumetric thin shell such as cloth, the Kirchhoff-Love model [78] offers a reduced kinematic parameterisation of the volume characterised by a 2D midsurface that fully determines the strain components throughout the thickness. Following thin shell assumptions, they decompose the Green strain due to deformation " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " parameterised by a multilayer perceptron (MLP)—into membrane strain " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " and bending strain " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": ", measuring the inplane stretching and the change in curvature, respectively. NeuralClothSim then computes the internal hyperelastic energy " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\Psi [\\varepsilon ,\\kappa ;\\Phi ]" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " as a functional of the geometric strains and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "the cloth's material properties " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ". Under the action of external force " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ", their neural solver finds the equilibrium solution following the principle of minimum potential energy, where the potential " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\int_{\\Omega} \\Psi d\\Omega - \\int_{\\Omega} \\mathbf{f} \\cdot \\mathbf{u} d\\Omega" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " is employed as the loss function. Note that the forward model of NeuralCloth-Sim [34] cannot be naively applied or extended (e.g., when coupled with a differentiable renderer, estimating physics parameters with the loss of NeuralClothSim would not converge due to the lack of lower bound) in our inverse setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 190, + 370, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 190, + 370, + 201 + ], + "spans": [ + { + "bbox": [ + 314, + 190, + 370, + 201 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "text", + "content": "We propose Thin-Shell-SfT, a new template-based method for the fine-grained 3D reconstruction of a deforming surface (such as cloth, paper or metal) seen in a monocular RGB video; see Fig. 2. We aim to reconstruct continuous 3D surfaces " + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_t\\}_{t\\in [1,\\dots,T]}" + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "text", + "content": " corresponding to the input image sequence " + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t\\}_t" + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "text", + "content": " and optional masks " + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_t\\}_t" + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "text", + "content": ". Similar to previous surface tracking methods [33, 50], we assume a static camera with known calibration and take the surface " + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1" + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "text", + "content": " corresponding to the first frame " + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 209, + 553, + 317 + ], + "type": "text", + "content": " as a template." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 318, + 553, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 553, + 484 + ], + "type": "text", + "content": "In contrast to the discrete representations (e.g. points, meshes) used in previous monocular tracking approaches, our deformation model encodes the surface and its dynamics as continuous and adaptive neural fields (Sec. 4.1). We optimise the neural model by relating input monocular views to the estimated surface states using 3DGS [36], where the Gaussians are initialised and dynamically tracked with parameters induced from the surface deformation (Sec. 4.2). To ensure plausible deformations due to the inherent ambiguity of the monocular setting and in contrast to prior discrete regularisers, we model the surfaces as thin shells imposing the continuous Kirchhoff-Love physics constraints [34] on the neural field, allowing us to reconstruct fine-grained details (Sec. 4.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 491, + 533, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 533, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 533, + 502 + ], + "type": "text", + "content": "4.1. Deformation Model and Parameterisation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "text", + "content": "Deformation Model. As our deformation model, we employ a continuous surface representation and deformation dynamics modelled as coordinated-based neural fields. Consider a surface with initial state " + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1\\subset \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "text", + "content": " and 2D parameterisation " + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "text", + "content": ". For any parametric point " + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\xi \\coloneqq (\\xi^{1},\\xi^{2})\\in \\Omega" + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "text", + "content": ", we represent its position on the initial surface state with a mapping " + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}} (\\pmb {\\xi}): \\Omega \\to \\mathbf{S}_1" + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "text", + "content": ". Furthermore, we encode the time-varying spatial location of " + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{\\xi}" + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "text", + "content": " on the tracked surface " + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_t" + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(\\pmb {\\xi},t):\\Omega \\times [1,\\dots,T]\\rightarrow \\mathbf{S}_t" + }, + { + "bbox": [ + 313, + 508, + 553, + 616 + ], + "type": "text", + "content": ", with" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 621, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 621, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 339, + 621, + 553, + 647 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {x} (\\boldsymbol {\\xi}, t) = \\bar {\\mathbf {x}} (\\boldsymbol {\\xi}) + \\mathbf {u} (\\boldsymbol {\\xi}, t), \\text {w i t h} \\mathbf {u} (\\boldsymbol {\\xi}, 1) = 0 \\text {a n d} \\tag {2} \\\\ \\mathbf {u} (\\boldsymbol {\\xi}, t) = \\lambda \\mathbf {u} (\\boldsymbol {\\xi}, t - 1) + \\mathcal {F} (\\boldsymbol {\\xi}, t), \\forall t > 1, \\\\ \\end{array}", + "image_path": "c317e793d48d8889e88795ba9abf7dc3873c4ee42599398c39ec2dff76c5e4b1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(\\pmb {\\xi},t)" + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": " is the deformation field, " + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\lambda >0" + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": " is a scalar value and " + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\pmb {\\xi},t)" + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": " represents the estimated deformation offset. Here, we encourage the conservation of momentum by setting the deformation of the future surface states in the direction of the previous deformations." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11375" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 297 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 297 + ], + "type": "image", + "image_path": "0d7cb6e1e3f11600bf14d5f740e23d65ac60d1f995075729125829d01a4f8e65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": "Figure 2. Overview of Thin-Shell-SfT. Our deformation model encodes the surface and its dynamics as neural fields. Given the template " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_1" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": ", we first fit a reference field (NRF) from 2D parametric points " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{\\xi}" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": " to the initial 3D positions " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": ". In the main stage, we optimise the deformation field (NDF) " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(\\pmb{\\xi}, t)" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": " by relating estimated surface states " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_t / \\mathcal{G}_t" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": " to the input monocular views. We induce the dynamically tracked Gaussians to the surface by: (1) Computing their positions " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": " as the sum of the initial position " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": " and NDF output " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": ", (2) Setting their rotations " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_i" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": " as the template's local coordinate system, and (3) Fixing the normal scale " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": ", and optimising the colour, opacity and tangential scales " + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "inline_equation", + "content": "(s_1, s_2)" + }, + { + "bbox": [ + 55, + 300, + 555, + 369 + ], + "type": "text", + "content": " using only the template texture. For physical plausibility, we impose continuous Kirchhoff-Love physics constraints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": "Parameterisation of the Template and Deformations. Next, we present an adaptive parameterisation of the above deformation model that allocates capacity to dynamic deformation details. Assuming that the template is provided as a mesh " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\subset \\mathbf{S}_1" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": ", we generate the corresponding 2D parametric space " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\subset \\Omega" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": " using established techniques (e.g. conformal parameterisation [30]). To learn a continuous representation of the initial state, we fit an MLP, which we call neural reference field (NRF) " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}(\\boldsymbol{\\xi}; \\Upsilon)" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": " to the template mesh parameterisation " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\mapsto \\mathcal{M}" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": ". For training NRF, we construct points in the parametric and vertex space with randomly sampled barycentric coordinates and train with " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": " geometry loss. Akin to NRF, we regress the spatio-temporal neural deformation field (NDF) " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(\\boldsymbol{\\xi}, t; \\Theta)" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": " using another MLP predicting deformation offsets, " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\boldsymbol{\\xi}, t; \\Theta): \\Omega \\times [1, \\dots, T] \\to \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": ". Before training, the initial estimate for tracked surface points " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(\\boldsymbol{\\xi}, t; \\Theta)" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": " is the sum of noisy NDF deformations, " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(\\boldsymbol{\\xi}, t; \\Theta)" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": " and the pre-trained NRF " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}(\\boldsymbol{\\xi}; \\Upsilon)" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": " as given by Eq. (2). Our key idea is to optimise the NDF weights " + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 374, + 295, + 708 + ], + "type": "text", + "content": " to ensure photometric consistency of the tracked surfaces with input monocular views while minimising the internal energy of the surface states modelled as a thin shell. The NRF pre-fit enables the coupling of Gaussians to the surface and computing surface metrics for the thin shell energy. We must represent high-frequency signals, such as fine folds and wrinkles on the tracked surface. Moreover, our physics loss requires the computation of higher-order derivatives. Hence, we use sine activation [66] in both MLPs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 374, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 494 + ], + "type": "text", + "content": "Inference. At testing, " + }, + { + "bbox": [ + 313, + 374, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(\\boldsymbol {\\xi},t;\\Theta)" + }, + { + "bbox": [ + 313, + 374, + 555, + 494 + ], + "type": "text", + "content": " provides continuous access to the reconstructed surface, where the NRF and NDF networks can be consistently queried at varied spatiotemporal resolutions. Moreover, our surface deformation model provides temporal correspondences. Meshing and texture mapping can be achieved in the parametric domain and then transferred to the tracked surface in 3D using Eq. (2). We next describe how to optimise the deformation field " + }, + { + "bbox": [ + 313, + 374, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(\\boldsymbol {\\xi},t;\\Theta)" + }, + { + "bbox": [ + 313, + 374, + 555, + 494 + ], + "type": "text", + "content": " that ensures consistency with monocular images and the physical plausibility of reconstructions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 503, + 533, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 533, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 533, + 516 + ], + "type": "text", + "content": "4.2. Surface Tracking with Gaussian Splitting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "Recall that we have an NDF parameterised by " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " that outputs the deformation " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(\\boldsymbol {\\xi},t;\\Theta)" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " at any time step. We seek to optimise the NDF so that the tracked surfaces " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_t\\}_{t}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " (evaluated with Eq. (2)) generate images matching with input views " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t\\}_{t}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ". A straightforward approach to encourage 3D-to-2D consistency is to minimise the photometric loss with differentiable rendering. We choose Gaussian Splatting [36] as the differentiable renderer, as it seamlessly integrates with our per-point (Eq. (2)) continuous deformation model. Initialisation of Gaussians. To initialise a set of Gaussians, we sample " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " well-distributed points (e.g. Poisson disk sampling [7]) from the surface of the input template mesh " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ". The sampled points include parametric coordinates " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\pmb {\\xi}_i\\}_{i\\in [1,\\dots ,N_g]}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ", their corresponding positions " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{x}}_i\\}_i" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ", and colours " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_i\\}_i" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ". Next, we will motivate our approach to parameterising Gaussian rotations and scales. Unlike the orig" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11376" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": "inal multi-view [36] or dynamic Gaussian methods [12, 81] that leverage multi-view cues [18], we have a single static camera and, consequently, a challenging monocular setup. Thus, learning the anisotropic covariance of 3D Gaussians from all input frames can lead to poor results due to the inherent monocular ambiguities between deformation, texture and appearance. In our datasets, the initial surface state is reasonably (but not exactly) flat and fully visible. With this assumption, we take the initial surface normal as a proxy for the viewing direction and then fix the scale along normal while allowing for optimisation of the scales along the initial surface tangents. More concretely, we set the rotation matrix of " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": "-th Gaussian equal to the local basis vectors, i.e. " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i \\equiv [\\bar{\\mathbf{a}}_1 \\bar{\\mathbf{a}}_2 \\bar{\\mathbf{a}}_3]_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " (see Sec. 4.3 for computation) of the initial state (NRF) " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}(\\pmb{\\xi}_i; \\Upsilon)" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " and the scale " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "s_3" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " as small " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " along surface normal " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_3" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": ". Finally, the dynamically tracked surface-induced Gaussian " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_t\\}_{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 281, + 296, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 281, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 79, + 281, + 296, + 313 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {G} _ {t} = \\left\\{\\mathcal {N} \\left(\\mathbf {x} \\left(\\boldsymbol {\\xi} _ {i}, t; \\Upsilon , \\Theta\\right), \\left[ \\bar {\\mathbf {a}} _ {1} \\left(\\boldsymbol {\\xi} _ {i}; \\Upsilon\\right), \\bar {\\mathbf {a}} _ {2} \\left(\\boldsymbol {\\xi} _ {i}; \\Upsilon\\right), \\right. \\right. \\right. \\\\ \\left. \\bar {\\mathbf {a}} _ {3} (\\boldsymbol {\\xi} _ {i}; \\Upsilon) ], (s _ {1}, s _ {2}, \\epsilon) _ {i}\\right), o _ {i}, \\mathbf {c} _ {i} \\} _ {i = 1} ^ {N _ {g}} \\\\ \\end{array}", + "image_path": "375b7300aa0ea7a885506af9b694a668eee6a4f46b75329a1b8226508e978c03.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "type": "text", + "content": "following Eq. (1). Note that the deformed positions " + }, + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(\\pmb{\\xi}_i, t)" + }, + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "type": "text", + "content": " are computed using our deformation model (Eq. (2)) and all other Gaussian parameters are shared across surface states." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 414 + ], + "type": "text", + "content": "Optimisation of NDF and Gaussian Parameters. For optimisation, we use the " + }, + { + "bbox": [ + 55, + 354, + 296, + 414 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 55, + 354, + 296, + 414 + ], + "type": "text", + "content": " photometric loss similar to 3DGS [36]. If segmentation masks " + }, + { + "bbox": [ + 55, + 354, + 296, + 414 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_t\\}_t" + }, + { + "bbox": [ + 55, + 354, + 296, + 414 + ], + "type": "text", + "content": " are available, an additional silhouette loss is added to primarily speed up training. Thus, our data loss reads as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 420, + 294, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 420, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 83, + 420, + 294, + 468 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {d} \\left(s _ {1}, s _ {2}, o, \\mathbf {c}, \\Theta\\right) = \\mathcal {R} \\left(\\mathcal {G} _ {1}, \\mathbf {I} _ {1}; s _ {1}, s _ {2}, o, \\mathbf {c}\\right) + \\\\ \\sum_ {t = 2} ^ {T} \\mathcal {R} \\left(\\mathcal {G} _ {t}, \\mathbf {I} _ {t}; \\Theta\\right) + \\mathcal {R} \\left(\\tilde {\\mathcal {G}} _ {t}, \\mathbf {M} _ {t}; \\Theta\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "227e0be7cd2a45290c7ee4cf854e53d188d3dabc1b65610e598e4ac0cb12fac0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(x;\\phi)" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " is the Gaussian rasterisation loss with inputs " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " and optimisable parameters " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{G}}_t" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " are the dynamic Gaussians " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " with colour " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": " set to mask foreground (e.g. white). Note that splating in the first frame updates shared Gaussian properties, whereas all other frames backpropagate to update the deformation field parameters " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ". With the formulation in Eq. (2), our method propagates gradient information from future states to update 3D reconstructions of the earlier frames. This enables global space-time surface optimisation while respecting the causality of the surface dynamics. While temporal consistency is implicit in Eq. (2), other variants of temporal regularisation are possible (see Appendix B). In contrast to previous 3DGS or dynamic Gaussian methods [81], we note that it is important to optimise all frames in each iteration rather than random frame " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "t \\in [2,\\dots,T]" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ", as it can otherwise lead to local minima (incorrect learning of folds). Global optimisation is preferred as our deformation model is a single global MLP for all frames, and it is computationally efficient over random frame selection due to the causality of deformations" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 554, + 230 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 554, + 230 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 554, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 554, + 230 + ], + "type": "image", + "image_path": "6cd1f02763f8d9b00c8ffc73ec6fe1b58c57227312075ac724f213eb366024aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 238, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 271 + ], + "type": "text", + "content": "Figure 3. Thin-shell physics prior is a spatial regulariser that minimises the hyperelastic strain energy density due to deformation w.r.t. the known template." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 276, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 323 + ], + "type": "text", + "content": "(Eq. (2)). Moreover, we keep the number of Gaussians fixed (i.e., no adaptive density control) since Gaussian parameters are learned only from the first frame; it is not known beforehand how the deformation details will arise in future states." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 333, + 455, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 455, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 455, + 346 + ], + "type": "text", + "content": "4.3. Thin Shell Physical Prior" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 352, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 639 + ], + "type": "text", + "content": "If we optimise the NDF solely using the photometric loss, it could perfectly fit the input images but could result in physically implausible surfaces due to monocular ambiguity. Therefore, we formulate a physics loss applicable to thin shells inspired by the NeuralClothSim approach [34] using the Kirchhoff-Love theory [78]. Since we model the dynamically tracked surface with an MLP, NeuralClothSim enables physics-based prior directly on the continuous surface. However, unlike the forward model of NeuralCloth-Sim, the external forces (such as contacts and the wind) and the material properties generating the deformations are unknown in an inverse setting like ours. For this reason, we assign material parameters to values typical for the surface modelled (e.g., cloth and paper). In contrast to material, it is hard to set reasonable values for forces or boundary conditions as the space of external forces is too large. Nevertheless, intuitively, the image loss effectively takes the role of external force as it guides the motion and deformation of the tracked surface. Thus, we omit the potential energy due to external forces from the optimisation. Finally, with the strain evaluated from the deformation field and the assumed material, our physical prior aims to minimise the internal hyperelastic energy that captures the surface stretching, shearing and bending stiffness; see Fig. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "content": "Strain Computation. To evaluate stretching and bending strain at each training iteration, we randomly sample " + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "content": " parametric points " + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\xi}_i\\}_{i=1}^{N_p}" + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "content": " from the template mesh " + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 639, + 556, + 714 + ], + "type": "text", + "content": ", and perform differential geometry operations on the initial surface state (i.e., NRF) and the deformed state (i.e., NDF). We next present the computation of the local quantities at" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11377" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": "each point " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\xi_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": " on the initial state using the pre-trained NRF " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{x}} (\\pmb {\\xi};\\Upsilon)" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": ". We use Greek letters for indexing quantities on the parameterised surface (e.g., " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_{\\alpha},\\alpha ,\\beta ,\\ldots = 1,2" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": "). For notational clarity, we drop the input " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\xi_{i},t" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": " and network weights " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\Upsilon ,\\Theta" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": " in all the derived quantities (e.g., " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_1\\equiv \\bar{\\mathbf{a}}_1(\\pmb {\\xi}_i;\\Upsilon),\\varepsilon \\equiv \\varepsilon (\\pmb {\\xi}_i,t;\\Upsilon ,\\Theta))" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": ". In the first step, we extract the local covariant basis: " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{a}}_{\\alpha}:= \\partial \\bar{\\mathbf{x}} /\\partial \\xi^{\\alpha},\\bar{\\mathbf{a}}_3:= \\bar{\\mathbf{a}}_1\\times \\bar{\\mathbf{a}}_2\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": ", the set of two vectors tangential to the curvilinear coordinate lines " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\xi^{\\alpha}" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": " and the local normal. For the Gaussian samples " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\{\\pmb {\\xi}_i\\}_{i = 1}^{N_g}" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": ", an identical computation is employed for setting the covariant basis as the rotation parameters (as in Sec. 4.2). Next, we evaluate the surface metric tensor " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\bar{a}_{\\alpha \\beta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": " measuring the lengths and the curvature tensor " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\bar{b}_{\\alpha \\beta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": " measuring the curvature of the midsurface. Akin to the covariant tensors (e.g. " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\bar{b}_{\\alpha \\beta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": "), their contravariant (e.g. " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\bar{b}^{\\alpha \\beta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": ") and mixed variants (e.g. " + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\bar{b}_{\\alpha}^{\\beta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 265 + ], + "type": "text", + "content": ") counterparts are extracted as well." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "text", + "content": "With these initial state quantities, we next present the strain—due to deformation—as a function of the NDF " + }, + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(\\boldsymbol {\\xi};\\Theta)" + }, + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "text", + "content": ". The predicted " + }, + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\hat{u}_i\\mathbf{e}_i" + }, + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "text", + "content": " in global Cartesian coordinate system is transformed to contravariant coordinate basis, " + }, + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = u_{\\alpha}\\bar{\\mathbf{a}}^{\\alpha} + u_{3}\\bar{\\mathbf{a}}^{3}" + }, + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "text", + "content": " for strain calculation [34]. Given the deformation gradient " + }, + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{,\\alpha}" + }, + { + "bbox": [ + 55, + 266, + 296, + 336 + ], + "type": "text", + "content": " derived from the NDF as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 338, + 296, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 338, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 74, + 338, + 296, + 357 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {, \\alpha} = \\varphi_ {\\alpha \\lambda} \\bar {\\mathbf {a}} ^ {\\lambda} + \\varphi_ {\\alpha 3} \\bar {\\mathbf {a}} ^ {3}, \\text {w i t h} \\tag {5}", + "image_path": "12e165b553fc1a5e0a45280a130d7281edddb7bddf0af051ee574b90f02a0060.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 354, + 268, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 354, + 268, + 368 + ], + "spans": [ + { + "bbox": [ + 72, + 354, + 268, + 368 + ], + "type": "interline_equation", + "content": "\\varphi_ {\\alpha \\lambda} := u _ {\\lambda} | _ {\\alpha} - \\bar {b} _ {\\alpha \\lambda} u _ {3} \\text {a n d} \\varphi_ {\\alpha 3} := u _ {3, \\alpha} + \\bar {b} _ {\\alpha} ^ {\\lambda} u _ {\\lambda},", + "image_path": "808bd64c800fbc94a08198446a6c6551feecc56ee7f33793c7d6fe320504f58a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "type": "text", + "content": "we evaluate the non-linear membrane strain " + }, + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\varepsilon = [\\varepsilon_{\\alpha \\beta}]" + }, + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "type": "text", + "content": " and bending strain " + }, + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\kappa = [\\kappa_{\\alpha \\beta}]" + }, + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 396, + 225, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 396, + 225, + 415 + ], + "spans": [ + { + "bbox": [ + 59, + 396, + 225, + 415 + ], + "type": "interline_equation", + "content": "\\varepsilon_ {\\alpha \\beta} = \\frac {1}{2} (\\varphi_ {\\alpha \\beta} + \\varphi_ {\\beta \\alpha} + \\varphi_ {\\alpha \\lambda} \\varphi_ {\\beta} ^ {\\lambda} + \\varphi_ {\\alpha 3} \\varphi_ {\\beta 3}),", + "image_path": "b88f1c3ecaf9dc8dd650141476adb4fa74382f5f01ce4fe87ee54bee250781b8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 417, + 294, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 417, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 58, + 417, + 294, + 447 + ], + "type": "interline_equation", + "content": "\\kappa_ {\\alpha \\beta} = - \\varphi_ {\\alpha 3} | _ {\\beta} - \\bar {b} _ {\\beta} ^ {\\lambda} \\varphi_ {\\alpha \\lambda} + \\varphi_ {3} ^ {\\lambda} \\left(\\varphi_ {\\alpha \\lambda} | _ {\\beta} + \\frac {1}{2} \\bar {b} _ {\\alpha \\beta} \\varphi_ {\\lambda 3} - \\bar {b} _ {\\beta \\lambda} \\varphi_ {\\alpha 3}\\right), \\tag {6}", + "image_path": "c748291c6009b19ed2a5815816fa0edb1dcd74164cef239e942f0fa5a7ae15ef.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}^i" + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "text", + "content": " denote the local contravariant basis and " + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "inline_equation", + "content": "\\bar{b}_{\\alpha \\beta}" + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "text", + "content": ", the components of the curvature tensor. In Eq. (6), we use a vertical bar for covariant derivatives, lower comma notation for partial derivatives w.r.t. the curvilinear coordinates " + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "inline_equation", + "content": "\\xi^{\\alpha}" + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "text", + "content": " (e.g. " + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "inline_equation", + "content": "u_{\\lambda}|_{\\alpha}" + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{,\\alpha} = \\partial \\mathbf{u} / \\partial \\xi^{\\alpha}" + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "text", + "content": "), and Einstein summation convention of repeated indices for tensorial operations (e.g., " + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "inline_equation", + "content": "\\varphi_{\\alpha \\lambda}\\varphi_{\\beta}^{\\lambda} = \\varphi_{\\alpha 1}\\varphi_{\\beta}^{1} + \\varphi_{\\alpha 2}\\varphi_{\\beta}^{2}" + }, + { + "bbox": [ + 55, + 453, + 295, + 560 + ], + "type": "text", + "content": "). Notably, we compute all the aforementioned physical quantities with automatic differentiation. Please see Appendix A for more details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "content": "NDF Optimisation. As material model, we use " + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "inline_equation", + "content": "\\Phi \\coloneqq \\{\\rho, h, E, \\nu\\}" + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "content": ", with mass density " + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "content": " and the surface thickness " + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "content": ", and elastic coefficients: Young's modulus " + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "content": ", and Poisson's ratio " + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "content": ". For simplicity and computational efficiency, we use a linear isotropic constitutive model relating strain to stress, thereby computing the in-plane stiffness " + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "content": ", the bending stiffness " + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "content": " and the elastic tensor " + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = [H^{\\alpha \\beta \\lambda \\delta}]" + }, + { + "bbox": [ + 55, + 561, + 295, + 657 + ], + "type": "text", + "content": " on the initial template, which read as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 661, + 294, + 691 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 661, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 83, + 661, + 294, + 691 + ], + "type": "interline_equation", + "content": "D := \\frac {E h}{1 - v ^ {2}}, B := \\frac {E h ^ {3}}{1 2 \\left(1 - v ^ {2}\\right)}, \\text {a n d} \\tag {7}", + "image_path": "01afce4c9d2448b89b69f1102d3caf66cf0dba3414e12d7564932b37bb7ebed0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 689, + 277, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 689, + 277, + 711 + ], + "spans": [ + { + "bbox": [ + 63, + 689, + 277, + 711 + ], + "type": "interline_equation", + "content": "H ^ {\\alpha \\beta \\lambda \\delta} := \\nu \\bar {a} ^ {\\alpha \\beta} \\bar {a} ^ {\\lambda \\delta} + \\frac {1}{2} (1 - \\nu) (\\bar {a} ^ {\\alpha \\lambda} \\bar {a} ^ {\\beta \\delta} + \\bar {a} ^ {\\alpha \\delta} \\bar {a} ^ {\\beta \\lambda}).", + "image_path": "161217978e2261aca88049b4f2ad576075104e79d7cd7a23bb86c0e553ea2b4f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "Finally, our physics loss over the tracked surface states reads as the following:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 116, + 553, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 116, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 326, + 116, + 553, + 193 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal{L}_{p}(\\Theta) = \\frac{1}{2N_{p}T}\\sum_{i = 1}^{N_{p}}\\sum_{t = 2}^{T}\\Big(\\underbrace{D\\varepsilon^{\\top}(\\boldsymbol{\\xi}_{i},t;\\Theta)\\mathbf{H}(\\boldsymbol{\\xi}_{i})\\varepsilon(\\boldsymbol{\\xi}_{i},t;\\Theta)}_{\\text{stretching/shearing stiffness}} \\\\ + \\underbrace {B \\boldsymbol {\\kappa} ^ {\\top} \\left(\\boldsymbol {\\xi} _ {i} , t ; \\Theta\\right) \\mathbf {H} \\left(\\boldsymbol {\\xi} _ {i}\\right) \\boldsymbol {\\kappa} \\left(\\boldsymbol {\\xi} _ {i} , t ; \\Theta\\right)} _ {\\text {b e n d i n g s t i f f n e s s}}) \\sqrt {\\bar {a} \\left(\\boldsymbol {\\xi} _ {i}\\right)}, \\tag {8} \\\\ \\end{array}", + "image_path": "07b05759ca250a7da847bfcd47aad2965bb48878ee116377af3d9cf4e8e9b862.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\sqrt{a} := |\\bar{\\mathbf{a}}_1 \\times \\bar{\\mathbf{a}}_2|" + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "text", + "content": ". Note that the term inside " + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "text", + "content": " evaluates per point; therefore, we re-sample at each training iteration to explore the continuous surface domain. Finally, we solve for the optimal NDF weights " + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\Theta^*" + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "text", + "content": " by minimising the objective function " + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\lambda_d\\mathcal{L}_d + \\lambda_p\\mathcal{L}_p" + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "text", + "content": " with empirically determined loss weights " + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\{\\lambda_d, \\lambda_p\\}" + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "text", + "content": ". The optimisation can be interpreted as an inverse simulation, with data loss guiding deformation as an external force and physics loss modelling the surface's intrinsic behaviour. We use iterative gradient-based optimisation to that end [37]. We choose " + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "inline_equation", + "content": "N_p \\ll N_g" + }, + { + "bbox": [ + 313, + 202, + 555, + 407 + ], + "type": "text", + "content": " as the physics loss involves expensive higher-order derivative computations, whereas the Gaussian rasterisation with samples fixed over training iterations is rather efficient. With resampling at each iteration, the physics loss backpropagates to the continuous surface, offering adaptive, memory-efficient performance compared to mesh-based physics alternatives requiring high resolution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 418, + 440, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 440, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 440, + 431 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": "Implementation Details. We implement Thin-Shell-SfT using PyTorch [58] and optimise separate NRF and NDF networks for each sequence. The NRF and NDF, both, employ the SIREN [66] architecture using a frequency of " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\omega = 5" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\omega = 30" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ", respectively. Both networks have five hidden layers and 256 units in each layer, and are optimised using Adam [37]. Each network is trained for " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\approx 2 \\cdot 10^{3}" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": " iterations; first, the template is fitted with NRF, followed by surface tracking with NDF. While NRF training takes up to two minutes, the core part of the method, i.e. the NDF training, typically takes between 30 minutes to one hour until convergence on an NVIDIA A100 GPU. The Gaussian scales along the normal are set to " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\epsilon = 10^{-5}" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ". In our experiments, the number of Gaussian samples is " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "N_{g} \\approx 90k" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ", and the number of samples for physics loss is " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "N_{p} = 100" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ". Concerning the physics loss, the linear elastic material properties of the surface are set as follows for all sequences: " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "E = 5000\\mathrm{Pa}, \\nu = 0.25" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "h = 1.2\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ". Moreover, we set the image and physics loss weights to " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\lambda_{d} = 5" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\lambda_{p} = 1" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ", whereas the temporal constraint scalar is set to " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\lambda = 0.4" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ". For visualisation of the tracked continuous surface, we generate temporally coherent meshes with ball-pivot meshing [4] of the template Gaussian positions " + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{x}}_i\\}_{i \\in [1,\\dots,N_g]}" + }, + { + "bbox": [ + 313, + 437, + 556, + 716 + ], + "type": "text", + "content": ". We refer to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11378" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 137 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 137 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 137 + ], + "type": "table", + "html": "
Seq.S1S2S3S4S5S6S7S8S9Avg
N-NRSfM [65]8.2533.62104.677.0272.668.73129.438.0619.8154.69
D-NRSfM [53]17.144.464.4041.3726.9214.0212.499.915.2915.11
Shimada [61]19.6922.1833.5490.3092.7857.6249.2724.4553.1249.22
DDD [82]2.951.693.8025.7310.466.9715.647.6111.7710.87
Ngo [50]2.191.512.1715.9010.723.017.95*failfail5.92*
Stokko [67]6.13.912.514.511.715.16.910.18.69.93
φ-SfT [33]0.792.753.547.606.153.144.732.522.363.93
Ours1.170.552.4†/3.55.5†/5.78.692.513.82.273.003.3†/3.5
", + "image_path": "9fcbcc0e7e77179623df6da807c22e84c8025480381d4d1b3a7cac8b88fe4bf7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 295, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 206 + ], + "type": "text", + "content": "Table 1. We quantitatively compare Thin-Shell-SfT to the state of the art on the " + }, + { + "bbox": [ + 55, + 140, + 295, + 206 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 140, + 295, + 206 + ], + "type": "text", + "content": "-SfT real dataset. The average Chamfer distance is multiplied by " + }, + { + "bbox": [ + 55, + 140, + 295, + 206 + ], + "type": "inline_equation", + "content": "10^4" + }, + { + "bbox": [ + 55, + 140, + 295, + 206 + ], + "type": "text", + "content": " for readability. “*” notes that Ngo et al. failed on the last few frames of S7, which we exclude from the error computation. “†” denotes that we report the numbers on the variant of temporal coherency constraint described in Appendix B." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 209, + 105, + 258 + ], + "blocks": [ + { + "bbox": [ + 59, + 209, + 105, + 258 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 105, + 258 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 105, + 258 + ], + "type": "image", + "image_path": "d1c0b95ffb1b50ddf2ea5e719b0da4b9e6c95bfd51f36aeca65d5c7a1cd7658a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 259, + 105, + 310 + ], + "blocks": [ + { + "bbox": [ + 59, + 259, + 105, + 310 + ], + "lines": [ + { + "bbox": [ + 59, + 259, + 105, + 310 + ], + "spans": [ + { + "bbox": [ + 59, + 259, + 105, + 310 + ], + "type": "image", + "image_path": "339f950a0868abb95aa9562dc7c24748cfb57555b3b2fd2cb31babf5d0e37e35.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 312, + 99, + 321 + ], + "lines": [ + { + "bbox": [ + 66, + 312, + 99, + 321 + ], + "spans": [ + { + "bbox": [ + 66, + 312, + 99, + 321 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 107, + 211, + 151, + 258 + ], + "blocks": [ + { + "bbox": [ + 107, + 211, + 151, + 258 + ], + "lines": [ + { + "bbox": [ + 107, + 211, + 151, + 258 + ], + "spans": [ + { + "bbox": [ + 107, + 211, + 151, + 258 + ], + "type": "image", + "image_path": "bffa853f26dc7b92741ee9742b5d151a50287d7c2a5935d288438b01f399fbd0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 107, + 259, + 151, + 308 + ], + "blocks": [ + { + "bbox": [ + 107, + 259, + 151, + 308 + ], + "lines": [ + { + "bbox": [ + 107, + 259, + 151, + 308 + ], + "spans": [ + { + "bbox": [ + 107, + 259, + 151, + 308 + ], + "type": "image", + "image_path": "6be45416f2de0cfb47e4cdfc09ac534516f4ab4b42a0453f27dbc1fbaa4f7212.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 309, + 146, + 323 + ], + "lines": [ + { + "bbox": [ + 108, + 309, + 146, + 323 + ], + "spans": [ + { + "bbox": [ + 108, + 309, + 146, + 323 + ], + "type": "text", + "content": "Ground-Truth Normals" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 153, + 212, + 197, + 258 + ], + "blocks": [ + { + "bbox": [ + 153, + 212, + 197, + 258 + ], + "lines": [ + { + "bbox": [ + 153, + 212, + 197, + 258 + ], + "spans": [ + { + "bbox": [ + 153, + 212, + 197, + 258 + ], + "type": "image", + "image_path": "0bcec40b359c8f506edb14d59d71a08456ba03168851df4b18441ac90eff2695.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 153, + 259, + 197, + 308 + ], + "blocks": [ + { + "bbox": [ + 153, + 259, + 197, + 308 + ], + "lines": [ + { + "bbox": [ + 153, + 259, + 197, + 308 + ], + "spans": [ + { + "bbox": [ + 153, + 259, + 197, + 308 + ], + "type": "image", + "image_path": "cc6cafb72167351b3bc48653869df7fece3dc174c40fb9eecdc2d5acb5a80269.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 309, + 190, + 323 + ], + "lines": [ + { + "bbox": [ + 155, + 309, + 190, + 323 + ], + "spans": [ + { + "bbox": [ + 155, + 309, + 190, + 323 + ], + "type": "text", + "content": "Normal Map Ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 199, + 212, + 244, + 258 + ], + "blocks": [ + { + "bbox": [ + 199, + 212, + 244, + 258 + ], + "lines": [ + { + "bbox": [ + 199, + 212, + 244, + 258 + ], + "spans": [ + { + "bbox": [ + 199, + 212, + 244, + 258 + ], + "type": "image", + "image_path": "1b29878780a6d9ec6ed3f95754d64061a715f195a54c872131dc6dbcca493c37.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 199, + 259, + 244, + 308 + ], + "blocks": [ + { + "bbox": [ + 199, + 259, + 244, + 308 + ], + "lines": [ + { + "bbox": [ + 199, + 259, + 244, + 308 + ], + "spans": [ + { + "bbox": [ + 199, + 259, + 244, + 308 + ], + "type": "image", + "image_path": "0bd5bb8e6fd4384978b38c825f24250f410302ceeb96daaa9dc1c2dd4f7dade5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 309, + 236, + 323 + ], + "lines": [ + { + "bbox": [ + 201, + 309, + 236, + 323 + ], + "spans": [ + { + "bbox": [ + 201, + 309, + 236, + 323 + ], + "type": "text", + "content": "Normal Map " + }, + { + "bbox": [ + 201, + 309, + 236, + 323 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 201, + 309, + 236, + 323 + ], + "type": "text", + "content": " -SfT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 326, + 295, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 360 + ], + "type": "text", + "content": "Figure 4. Comparison of the reconstruction normal maps for ours and " + }, + { + "bbox": [ + 55, + 326, + 295, + 360 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 326, + 295, + 360 + ], + "type": "text", + "content": "-SfT. We show the cosine normal consistency to ground truth on the right and the normal metrics in Tab. I (Appendix)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 246, + 212, + 291, + 258 + ], + "blocks": [ + { + "bbox": [ + 246, + 212, + 291, + 258 + ], + "lines": [ + { + "bbox": [ + 246, + 212, + 291, + 258 + ], + "spans": [ + { + "bbox": [ + 246, + 212, + 291, + 258 + ], + "type": "image", + "image_path": "8d6853676543f2b457af600ff02fbf67f66eda2c11b183337d7c27c7218271de.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 246, + 259, + 291, + 308 + ], + "blocks": [ + { + "bbox": [ + 246, + 259, + 291, + 308 + ], + "lines": [ + { + "bbox": [ + 246, + 259, + 291, + 308 + ], + "spans": [ + { + "bbox": [ + 246, + 259, + 291, + 308 + ], + "type": "image", + "image_path": "2864c47de2278b87ac2a8a6b520a671afcea0bac6288323d151966da38f6a658.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 309, + 287, + 323 + ], + "lines": [ + { + "bbox": [ + 249, + 309, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 249, + 309, + 287, + 323 + ], + "type": "text", + "content": "Normal Error Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 363, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 388 + ], + "type": "text", + "content": "Appendices B and D and Figs. II and III for studies on the effect of hyperparameters " + }, + { + "bbox": [ + 55, + 363, + 295, + 388 + ], + "type": "inline_equation", + "content": "\\omega, \\lambda" + }, + { + "bbox": [ + 55, + 363, + 295, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 363, + 295, + 388 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 55, + 363, + 295, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "text", + "content": "Datasets and Error Metrics. We evaluate on the benchmark " + }, + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "text", + "content": "-SfT dataset [33] consisting of nine RGB videos (and reference depths) of highly challenging cloth deformations. As is common with non-rigid reconstruction methods [33, 65], we rigidly align the reconstructions of all compared methods to the ground truth. This is achieved with Procrustes alignment [72] for the first frame, which is refined with rigid ICP [5] for subsequent frames." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 491, + 136, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 136, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 136, + 504 + ], + "type": "text", + "content": "5.1. Comparison" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "text", + "content": "SfT and NRSfM. We compare Thin-Shell-SfT to different SfT methods, namely " + }, + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "text", + "content": "-SfT [33], Yu et al.'s Direct, Dense, Deformable (DDD) [82], Ngo et al.'s Ngo2015 [50] and Shimada et al.'s IsMo-GAN [61], and finally to Stotko et al.'s approach [67]. DDD is provided with the required hierarchy of coarse-to-fine templates, and Ngo2015 and Stotko et al. with the same template as in the " + }, + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "text", + "content": "-SfT dataset. Since the dataset is richly textured, the 2D point tracking methods would perform well; therefore, we additionally compare to the NRSfM methods. In particular, we compare to Sidhu et al.'s Neural NRSfM (N-NRSfM) [65] and Parashar et al.'s Diff-NRSfM [53]. The 2D point correspondences required by the NRSfM methods are provided by densely tracking the 2D points across input images with multi-frame subspace flow (MFSF) [19, 20]. The first frame of the sequence is selected as a keyframe for 2D tracking." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "In Fig. 5-(left), we show qualitative reconstructions of" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": "Thin-Shell-SfT compared to the prior state of the art (SotA) [33, 53, 67]. We capture fine wrinkles and folds in the deforming sequences that were not addressed by the earlier works. Next, we reconstruct surfaces with severe self-occlusions due to multiple layered folds in the extended versions of the " + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": "-SfT sequences [33]. See Fig. 5-(right), where our physics-based model provides a reasonable prior for occluded regions; " + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": "-SfT fails here due to prohibitive memory requirements. In Tab. 1, we show a quantitative comparison where we report the Chamfer distance (with pseudoground-truth point clouds) against the state-of-the-art methods over all the " + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": "-SfT dataset sequences. We outperform the existing methods on most sequences, often substantially and on average over all sequences. While the default temporal coherency works best on the evaluated dataset, we notice qualitative improvements for two out of the nine sequences with the other variants. In Fig. 4, we further show comparisons between the normal maps of the tracked surface. Additional numerical comparisons with the previous SotA [33] on runtime and normal consistency are in Appendix C and Tab. I. We obtain a lower " + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "0.009(\\ell_2)" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "0.034" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": " (cosine) normal error, whereas " + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": "-SfT shows " + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "0.013" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "0.041" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": ", respectively. One of the primary reasons for the failure of " + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": "-SfT [33] is its limitation to low-resolution meshes (" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\approx 300" + }, + { + "bbox": [ + 313, + 72, + 555, + 503 + ], + "type": "text", + "content": " vertices) capturing just coarse deformations. Our NDF offers, similar to scene representation approaches [47, 75], a smooth (cf. global MLP), low-dimensional (cf. a fixed parameter count) deformation space that is adaptive to dynamic details and can represent high-frequency signals (cf. sine activation). Although we discretise the physics loss, we re-sample at each iteration, leading to an adaptive discretisation. In contrast, earlier physics-based methods [33, 67] use fixed discretisation, either due to differentiable simulation not supporting remeshing [33] or a fixed-resolution surrogate model [67]. Our continuous formulation enables unprecedented fine-grained results." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 504, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 556, + 624 + ], + "type": "text", + "content": "Dynamic View Synthesis. Next, we compare with state-of-the-art dynamic view synthesis methods, particularly K-Planes [15] and Deformable Gaussians [81]. For the latter, we initialise the Gaussians with points sampled from the input template instead of the default SfM points. The compared radiance field methods are very effective in novel-view synthesis for monocular videos that provide multiview cues [18]. However, they fail to recover spatiotemporally coherent surface geometry for monocular RGB videos captured with a single static camera; see Fig. 6." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 634, + 413, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 413, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 413, + 645 + ], + "type": "text", + "content": "5.2. Ablative Studies" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 652, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 553, + 700 + ], + "type": "text", + "content": "Next, we test the following aspects of our method: contribution of the physics loss, global optimisation and surface-induced Gaussians. Additional ablations, including momentum and mask loss, are presented in the Appendix E." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Physics Prior. While relying solely on image loss yields" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11379" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 203 + ], + "type": "image", + "image_path": "24d56cec1882690718dd630e7467da6ce2d5ba15b1734cc9155d8110310480ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "type": "text", + "content": "Figure 5. Exampary 3D reconstructions. (Left:) Comparisons focusing on high-frequency wrinkles. Thin-Shell-SfT captures the wrinkles best among all compared methods in one of the most challenging examples, outperforming " + }, + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "type": "text", + "content": "-SfT [33], Stotko et al. [67] and Diff-NRSfM [53]. (Right:) Our results on the extended " + }, + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 204, + 555, + 239 + ], + "type": "text", + "content": "-SfT dataset highlight the excellent tracking in the occluded regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 240, + 294, + 323 + ], + "blocks": [ + { + "bbox": [ + 57, + 240, + 294, + 323 + ], + "lines": [ + { + "bbox": [ + 57, + 240, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 57, + 240, + 294, + 323 + ], + "type": "image", + "image_path": "7b9ea7e84bce812eb006a66806e589ec0870436ae9c0521547b2d998d01deeb0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 327, + 295, + 371 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 371 + ], + "type": "text", + "content": "Figure 6. We compare our reconstructions to dynamic view synthesis methods [15, 81] on the rendered depth maps. The texture details retained in the depth maps imply that the compared methods fail to learn accurate surface (mistake texture for geometry)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 375, + 295, + 494 + ], + "blocks": [ + { + "bbox": [ + 57, + 375, + 295, + 494 + ], + "lines": [ + { + "bbox": [ + 57, + 375, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 57, + 375, + 295, + 494 + ], + "type": "image", + "image_path": "ea6b27ca125c9ea7cd03434845fcad8e531ff88b2f70dcc1437419acbd2bea61.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 499, + 295, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 544 + ], + "type": "text", + "content": "Figure 7. Ablations. (a:) Without physics, the surface can severely stretch or shrink. (b:) Random frame optimisation leads to local minima, (c,d:) Not positioning and orienting Gaussians on the surface leads to distortions and poor reconstructions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "content": "accurate RGB recovery during rendering, it can lead to severely distorted surface tracking. This arises from the inherent monocular ambiguity of our setting as multiple possible deformations in 3D correspond to the same 2D image. As seen in Fig. 7-(a), incorporating Kirchhoff-Love-based thin shell prior is essential for achieving physically plausible and accurate surface reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Global Optimisation. We test a version of our method employing optimisation with randomly selected frames at each iteration instead of the proposed joint optimisation over all frames. Fig. 7-(b) visualises the ablated results showing poor performance for folds and wrinkles. Without any multi-view cues, as in the case of the used " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "-SfT dataset, random frame optimisation leads to local minima." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 240, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 396 + ], + "type": "text", + "content": "Surface-induced Gaussians. The specially-tailored initialisation and optimisation of Gaussian parameters (Eqs. (3) and (4)) are crucial for the accurate geometric reconstruction. In the ablated version of our method, we optimise the shared Gaussian parameters, i.e., covariance, opacity, and colour on all input frames (similar to earlier methods [36, 81]) instead of the single template frame as in Eq. (4). Due to the ambiguities between geometric details and appearance over the deforming sequence, this leads to wrongly reconstructed Gaussians, eventually resulting in poor surface reconstructions; see Fig. 7-(c). In addition, fixing the normal scale on the template frame is useful for preventing elongated Gaussians (Fig. 7-(d))." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "content": "In Tab. II-appendix, we report the above results on the full " + }, + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "content": "-SfT dataset. On average, we obtain Chamfer distances of 34.25, 14.0, 3.75, and 3.46 for no physics, no surface-induced Gaussians, no normal scale, and the full model, respectively. We notice that including physics and surface induction are crucial for accurate tracking." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": "We present Thin-Shell-SfT, a new approach for dense monocular non-rigid 3D surface tracking relying on new principles, i.e. adaptive neural deformation field, continuous Kirchhoff-Love thin shell prior and surface-induced 3D Gaussian Splating. Experiments on the " + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": "-SfT dataset demonstrate that our method accurately reconstructs the challenging fine-grained deformations such as cloth folds and wrinkles. Overall, we significantly improve fine-grained surface tracking using an adaptive deformation model and continuous thin shell physics compared to existing approaches that only support coarse deformations. Thanks to the physics prior, Thin-Shell-SfT is reasonably robust to occlusions, although extreme self-collisions remain challenging. Similarly, tracking textureless surfaces is another possible research direction that would necessitate special handling in future." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was partially supported by the VIA Research Center." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11380" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Nonrigid structure from motion in trajectory space. In Advances in Neural Information Processing Systems (NeurIPS), 2009. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "text", + "content": "[2] Mohammad Dawud Ansari, Vladislav Golyanik, and Didier Stricker. Scalable dense monocular surface reconstruction. In International Conference on 3D Vision (3DV), 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 212 + ], + "type": "text", + "content": "[3] Adrien Bartoli, Yan Gérard, François Chadebecq, Toby Collins, and Daniel Pizarro. Shape-from-template. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2015. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 246 + ], + "type": "text", + "content": "[4] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE TVCG, 1999. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 280 + ], + "type": "text", + "content": "[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, pages 586-606. Spie, 1992. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 281, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 323 + ], + "type": "text", + "content": "[6] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Computer Vision and Pattern Recognition (CVPR), 2000. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 325, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 346 + ], + "type": "text", + "content": "[7] Robert Bridson. Fastoisson disk sampling in arbitrary dimensions. SIGGRAPH sketches, 2007. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 348, + 295, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 295, + 391 + ], + "type": "text", + "content": "[8] David Casillas-Perez, Daniel Pizarro, David Fuentes-Jimenez, Manuel Mazo, and Adrien Bartoli. The isowarp: the template-based visual geometry of isometric surfaces. International Journal of Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 393, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 295, + 435 + ], + "type": "text", + "content": "[9] Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne Vouga, and Christoph Lassner. Virtual elastic objects. In Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 437, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 437, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 295, + 479 + ], + "type": "text", + "content": "[10] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. In Computer Vision and Pattern Recognition (CVPR), 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 481, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 481, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 57, + 481, + 295, + 524 + ], + "type": "text", + "content": "[11] Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "text", + "content": "[12] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 570, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 623 + ], + "type": "text", + "content": "[13] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splating: Learning metric deformation from 4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "text", + "content": "[14] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, and Yin Yang. Gaussian splashing: Unified particles for versatile motion synthesis and rendering. Computer Vision and Pattern Recognition (CVPR), 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[15] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition (CVPR), 2023. 2, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 129 + ], + "type": "text", + "content": "[16] David Fuentes-Jimenez, Daniel Pizarro, David Casillas-Perez, Toby Collins, and Adrien Bartoli. Texture-generic deep shape-from-template. IEEE Access, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 553, + 184 + ], + "type": "text", + "content": "[17] David Fuentes-Jimenez, Daniel Pizarro, David Casillas-Pérez, Toby Collins, and Adrien Bartoli. Deep shape-from-template: Single-image quasi-isometric deformable registration and reconstruction. Image and Vision Computing, 127: 104531, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "text", + "content": "[18] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "text", + "content": "[19] Ravi Garg, Anastasios Roussos, and Lourdes Agapito. A variational approach to video registration with subspace constraints. International Journal of Computer Vision (IJCV), 104(3):286-314, 2013. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[20] Ravi Garg, Anastasios Roussos, and Lourdes Agapito. Source code of Multi-Frame Subspace Flow (MFSF). http://www0.cs.ucl.ac.uk/staff/lagapito/subspace_flow/, 2015.7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 321, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 353 + ], + "type": "text", + "content": "[21] Vladislav Golyanik, Soshi Shimada, Kiran Varanasi, and Didier Stricker. Hdm-net: Monocular non-rigid 3d reconstruction with learned deformation model. In EuroVR, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 355, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 387 + ], + "type": "text", + "content": "[22] Vladislav Golyanik, André Jonas, and Didier Stricker. Consolidating segmentwise non-rigid structure from motion. In Machine Vision Applications (MVA), 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "text", + "content": "[23] Stella Graßhof and Sami Sebastian Brandt. Tensor-based non-rigid structure from motion. In Winter Conference on Applications of Computer Vision (WACV), 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "text", + "content": "[24] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "text", + "content": "[25] Jingfan Guo, Jie Li, Rahul Narain, and Hyun Soo Park. Inverse simulation: Reconstructing dynamic geometry of clothed humans via optimal control. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "text", + "content": "[26] Marc Habermann, Weipeng Xu, Helge Rhodin, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Nrst: Non-rigid surface tracking from monocular video. German Conference on Pattern Recognition (GCPR), 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "text", + "content": "[27] Nazim Haouchine and Stephane Cotin. Template-based monocular 3d recovery of elastic shapes using lagrangian multipliers. In Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 602, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 645 + ], + "type": "text", + "content": "[28] Tong He, John Collomosse, Hailin Jin, and Stefano Soatto. Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[29] Florian Hofherr, Lukas Koestler, Florian Bernard, and Daniel Cremers. Neural implicit representations for physical parameter inference from a single video. In Winter Conference on Applications of Computer Vision (WACV), 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[30] Alec Jacobson, Daniele Panozzo, et al. libigl: A simple " + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}++" + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": " geometry processing library, 2018. https://libigl.github.io/. 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11381" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[31] Miguel Jaques, Michael Burke, and Timothy Hospedales. Physics-as-inverse-graphics: Unsupervised physical parameter estimation from video. In International Conference on Learning Representations (ICLR), 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[32] Erik Johnson, Marc Habermann, Soshi Shimada, Vladislav Golyanik, and Christian Theobalt. Unbiased 4d: Monocular 4d reconstruction with a neural deformation model. In Computer Vision and Pattern Recognition (CVPR) Workshops, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "type": "text", + "content": "[33] Navami Kairanda, Edith Tretschk, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. " + }, + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "type": "text", + "content": "-SfT: Shape-from-template with a physics-based deformation model. In Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 280 + ], + "type": "text", + "content": "[34] Navami Kairanda, Marc Habermann, Christian Theobalt, and Vladislav Golyanik. Neuralclothsim: Neural deformation fields meet the thin shell theory. Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 5, 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "text", + "content": "[35] Rama Kandukuri, Jan Achterhold, Michael Moeller, and Joerg Stueckler. Learning to identify physical parameters from video using differentiable physics. In German Conference for Pattern Recognition (GCPR), 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[36] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 2023. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 392 + ], + "type": "text", + "content": "[37] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv e-prints, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 436 + ], + "type": "text", + "content": "[38] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[39] Suryansh Kumar. Jumping manifolds: Geometry aware dense non-rigid structure from motion. In Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[40] Suryansh Kumar, Anoop Cherian, Yuchao Dai, and Hongdong Li. Scalable dense non-rigid structure-from-motion: A grassmannian perspective. In Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "text", + "content": "[41] Yue Li, Marc Habermann, Bernhard Thomaszewski, Stelian Coros, Thabo Beeler, and Christian Theobalt. Deep Physics-aware Inference of Cloth Deformation for Monocular Human Performance Capture. In International Conference on 3D Vision (3DV), 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[42] Yifei Li, Tao Du, Kui Wu, Jie Xu, and Wojciech Matusik. Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 644 + ], + "type": "text", + "content": "[43] Junbang Liang, Ming Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse problems. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[44] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[45] Alberta Longhini, Marcel Busching, Bardienus Pieter Duisterhof, Jens Lundell, Jeffrey Ichnowski, Marten Björkman," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "and Danica Kragic. Cloth-splatting: 3d cloth state estimation from rgb supervision. In 8th Annual Conference on Robot Learning, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "type": "text", + "content": "[46] Augustus Edward Hough Love. A treatise on the mathematical theory of elasticity. University press, 1927. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[47] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "text", + "content": "[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 297 + ], + "type": "text", + "content": "[49] J. Krishna Murthy, Miles Macklin, Florian Golemo, Vikram Voleti, Linda Petrini, Martin Weiss, Breandan Considine, Jérôme Parent-Lévesque, Kevin Xie, Kenny Erleben, Liam Paull, Florian Shkurti, Derek Nowrouzezahrai, and Sanja Fidler. gradsim: Differentiable simulation for system identification and visuomotor control. In International Conference on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "text", + "content": "[50] Dat Tien Ngo, Sanghyuk Park, Anne Jorstad, Alberto Crivellaro, Chang D. Yoo, and Pascal Fua. Dense image registration and deformable surface reconstruction in presence of occlusions and minimal texture. In International Conference on Computer Vision (ICCV), 2015. 1, 2, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "type": "text", + "content": "[51] Marco Paladini, Alessio Del Bue, Marko Stosic, Marija Dodig, Jo ao Xavier, and Lourdes Agapito. Factorization for non-rigid and articulated structure using metric projections. In Computer Vision and Pattern Recognition (CVPR), 2009." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[52] Shaifali Parashar, Daniel Pizarro, Adrien Bartoli, and Toby Collins. As-rigid-as-possible volumetric shape-from-template. In International Conference on Computer Vision (ICCV), 2015. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "text", + "content": "[53] Shaifali Parashar, Mathieu Salzmann, and Pascal Fua. Local non-rigid structure-from-motion from diffeomorphic mappings. In Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 566 + ], + "type": "text", + "content": "[54] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40 (6):1-12, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 611 + ], + "type": "text", + "content": "[55] Mathieu Perriollat, Richard Hartley, and Adrien Bartoli. Monocular template-based reconstruction of inextensible surfaces. International Journal of Computer Vision (IJCV), 95(2):124-137, 2011. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "text", + "content": "[56] Albert Pumarola, Antonio Agudo, Lorenzo Porzi, Alberto Sanfeliu, Vincent Lepetit, and Francesc Moreno-Noguer. Geometry-Aware Network for Non-Rigid Shape Prediction from a Single View. In Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[57] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11382" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[58] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[59] Davis Rempe, Leonidas J. Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jimei Yang. Contact and human dynamics from monocular video. In European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "text", + "content": "[60] Mathieu Salzmann, Julien Pilet, Slobodan Ilic, and Pascal Fua. Surface deformation models for nonrigid 3d shape recovery. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 29(8):1481-1487, 2007. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[61] Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Didier Stricker. IsMo-GAN: Adversarial learning for monocular non-rigid 3d reconstruction. In Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[62] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. Physcap: Physically plausible monocular 3d motion capture in real time. ACM Transactions on Graphics, 39(6), 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[63] Soshi Shimada, Vladislav Golyanik, Weipeng Xu, Patrick Pérez, and Christian Theobalt. Neural monocular 3d human motion capture with physical awareness. ACM Transactions on Graphics, 40(4), 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "text", + "content": "[64] Soshi Shimada, Vladislav Golyanik, Patrick Pérez, and Christian Theobalt. Decaf: Monocular deformation capture for face and hand interactions. ACM Transactions on Graphics, 42(6), 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[65] Vikramjit Sidhu, Edgar Tretschk, Vladislav Golyanik, Antonio Agudo, and Christian Theobalt. Neural dense non-rigid structure from motion with latent space constraints. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[66] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 4, 6, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[67] David Stotko, Nils Wandel, and Reinhard Klein. Physics-guided shape-from-template: Monocular video perception through neural surrogate models. Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "text", + "content": "[68] Guoxing Sun, Rishabh Dabral, Pascal Fua, Christian Theobalt, and Marc Habermann. Metacap: Meta-learning priors from multi-view imagery for sparse-view human performance capture and rendering. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 646 + ], + "type": "text", + "content": "[69] Lorenzo Torresani, Aaron Hertzmann, and Chris Bregler. Nonrigid structure-from-motion: Estimating shape and motion with hierarchical priors. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 30(5), 2008. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[70] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 139 + ], + "type": "text", + "content": "[71] Edith Tretschk, Navami Kairanda, Mallikarjun BR, Rishabh Dabral, Adam Kortylewski, Bernhard Egger, Marc Habermann, Pascal Fua, Christian Theobalt, and Vladislav Golyanik. State of the art in dense monocular non-rigid 3d reconstruction. In Computer Graphics Forum (Eurographics State of the Art Reports), 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 315, + 140, + 553, + 172 + ], + "type": "text", + "content": "[72] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 1991. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 315, + 173, + 553, + 215 + ], + "type": "text", + "content": "[73] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. Games: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 553, + 249 + ], + "type": "text", + "content": "[74] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 315, + 250, + 553, + 293 + ], + "type": "text", + "content": "[75] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 315, + 294, + 553, + 348 + ], + "type": "text", + "content": "[76] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 350, + 553, + 392 + ], + "type": "text", + "content": "[77] Sebastian Weiss, Robert Maier, Daniel Cremers, Rüdiger Westermann, and Nils Thuerey. Correspondence-free material reconstruction using sparse surface constraints. Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 393, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 393, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 393, + 553, + 415 + ], + "type": "text", + "content": "[78] Gerald Kempner and Demosthenes Talaslidis. Mechanics of solids and shells. CRC, Boca Raton, 2003. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 415, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 415, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 315, + 415, + 553, + 459 + ], + "type": "text", + "content": "[79] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 315, + 460, + 553, + 503 + ], + "type": "text", + "content": "[80] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 504, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 504, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 315, + 504, + 553, + 548 + ], + "type": "text", + "content": "[81] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 315, + 548, + 553, + 592 + ], + "type": "text", + "content": "[82] Rui Yu, Chris Russell, Neill D. F. Campbell, and Lourdes Agapito. Direct, dense, and deformable: Template-based non-rigid 3d reconstruction from rgb video. In International Conference on Computer Vision (ICCV), 2015. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 315, + 593, + 553, + 646 + ], + "type": "text", + "content": "[83] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y. Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. PhysDreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 553, + 712 + ], + "type": "text", + "content": "[84] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, and Gordon Wetzstein. Physavatar: Learning the physics of dressed 3d avatars from visual observations. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11383" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file